SUPDrv.c@ 28800

最後變更在這個檔案從28800是 28800,由 vboxsync 提交於 15 年前
Automated rebranding to Oracle copyright/license strings via filemuncher
屬性 svn:eol-style 設為 `native` 屬性 svn:keywords 設為 `Author Date Id Revision`
檔案大小: 193.2 KB

行
1	/* $Revision: 28800 $ */
2	/** @file
3	* VBoxDrv - The VirtualBox Support Driver - Common code.
4	*/
5
6	/*
7	* Copyright (C) 2006-2009 Oracle Corporation
8	*
9	* This file is part of VirtualBox Open Source Edition (OSE), as
10	* available from http://www.alldomusa.eu.org. This file is free software;
11	* you can redistribute it and/or modify it under the terms of the GNU
12	* General Public License (GPL) as published by the Free Software
13	* Foundation, in version 2 as it comes in the "COPYING" file of the
14	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16	*
17	* The contents of this file may alternatively be used under the terms
18	* of the Common Development and Distribution License Version 1.0
19	* (CDDL) only, as it comes in the "COPYING.CDDL" file of the
20	* VirtualBox OSE distribution, in which case the provisions of the
21	* CDDL are applicable instead of those of the GPL.
22	*
23	* You may elect to license modified versions of this file under the
24	* terms and conditions of either the GPL or the CDDL or both.
25	*/
26
27	/*******************************************************************************
28	* Header Files *
29	*******************************************************************************/
30	#define LOG_GROUP LOG_GROUP_SUP_DRV
31	#define SUPDRV_AGNOSTIC
32	#include "SUPDrvInternal.h"
33	#ifndef PAGE_SHIFT
34	# include <iprt/param.h>
35	#endif
36	#include <iprt/alloc.h>
37	#include <iprt/cpuset.h>
38	#include <iprt/handletable.h>
39	#include <iprt/mp.h>
40	#include <iprt/power.h>
41	#include <iprt/process.h>
42	#include <iprt/semaphore.h>
43	#include <iprt/spinlock.h>
44	#include <iprt/thread.h>
45	#include <iprt/uuid.h>
46	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
47	# include <iprt/crc32.h>
48	# include <iprt/net.h>
49	# include <iprt/string.h>
50	# include <iprt/rand.h>
51	# include <iprt/path.h>
52	#endif
53
54	#include <VBox/param.h>
55	#include <VBox/log.h>
56	#include <VBox/err.h>
57	#include <VBox/hwacc_svm.h>
58	#include <VBox/hwacc_vmx.h>
59	#include <VBox/x86.h>
60
61	/*
62	* Logging assignments:
63	* Log - useful stuff, like failures.
64	* LogFlow - program flow, except the really noisy bits.
65	* Log2 - Cleanup.
66	* Log3 - Loader flow noise.
67	* Log4 - Call VMMR0 flow noise.
68	* Log5 - Native yet-to-be-defined noise.
69	* Log6 - Native ioctl flow noise.
70	*
71	* Logging requires BUILD_TYPE=debug and possibly changes to the logger
72	* instanciation in log-vbox.c(pp).
73	*/
74
75
76	/*******************************************************************************
77	* Defined Constants And Macros *
78	*******************************************************************************/
79	/** The frequency by which we recalculate the u32UpdateHz and
80	* u32UpdateIntervalNS GIP members. The value must be a power of 2. */
81	#define GIP_UPDATEHZ_RECALC_FREQ 0x800
82
83	/** @def VBOX_SVN_REV
84	* The makefile should define this if it can. */
85	#ifndef VBOX_SVN_REV
86	# define VBOX_SVN_REV 0
87	#endif
88
89
90	/*******************************************************************************
91	* Internal Functions *
92	*******************************************************************************/
93	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser);
94	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser);
95	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession);
96	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType);
97	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq);
98	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq);
99	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq);
100	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq);
101	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq);
102	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx);
103	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt);
104	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage);
105	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage);
106	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt);
107	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt);
108	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq);
109	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq);
110	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt);
111	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt);
112	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
113	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
114	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser);
115	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys, uint64_t u64NanoTS, unsigned uUpdateHz);
116	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip);
117	static void supdrvGipUpdate(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick);
118	static void supdrvGipUpdatePerCpu(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC, unsigned iCpu, uint64_t iTick);
119
120
121	/*******************************************************************************
122	* Global Variables *
123	*******************************************************************************/
124	DECLEXPORT(PSUPGLOBALINFOPAGE) g_pSUPGlobalInfoPage = NULL;
125
126	/**
127	* Array of the R0 SUP API.
128	*/
129	static SUPFUNC g_aFunctions[] =
130	{
131	/* name function */
132	/* Entries with absolute addresses determined at runtime, fixup
133	code makes ugly ASSUMPTIONS about the order here: */
134	{ "SUPR0AbsIs64bit", (void *)0 },
135	{ "SUPR0Abs64bitKernelCS", (void *)0 },
136	{ "SUPR0Abs64bitKernelSS", (void *)0 },
137	{ "SUPR0Abs64bitKernelDS", (void *)0 },
138	{ "SUPR0AbsKernelCS", (void *)0 },
139	{ "SUPR0AbsKernelSS", (void *)0 },
140	{ "SUPR0AbsKernelDS", (void *)0 },
141	{ "SUPR0AbsKernelES", (void *)0 },
142	{ "SUPR0AbsKernelFS", (void *)0 },
143	{ "SUPR0AbsKernelGS", (void *)0 },
144	/* Normal function pointers: */
145	{ "SUPR0ComponentRegisterFactory", (void *)SUPR0ComponentRegisterFactory },
146	{ "SUPR0ComponentDeregisterFactory", (void *)SUPR0ComponentDeregisterFactory },
147	{ "SUPR0ComponentQueryFactory", (void *)SUPR0ComponentQueryFactory },
148	{ "SUPR0ObjRegister", (void *)SUPR0ObjRegister },
149	{ "SUPR0ObjAddRef", (void *)SUPR0ObjAddRef },
150	{ "SUPR0ObjAddRefEx", (void *)SUPR0ObjAddRefEx },
151	{ "SUPR0ObjRelease", (void *)SUPR0ObjRelease },
152	{ "SUPR0ObjVerifyAccess", (void *)SUPR0ObjVerifyAccess },
153	{ "SUPR0LockMem", (void *)SUPR0LockMem },
154	{ "SUPR0UnlockMem", (void *)SUPR0UnlockMem },
155	{ "SUPR0ContAlloc", (void *)SUPR0ContAlloc },
156	{ "SUPR0ContFree", (void *)SUPR0ContFree },
157	{ "SUPR0LowAlloc", (void *)SUPR0LowAlloc },
158	{ "SUPR0LowFree", (void *)SUPR0LowFree },
159	{ "SUPR0MemAlloc", (void *)SUPR0MemAlloc },
160	{ "SUPR0MemGetPhys", (void *)SUPR0MemGetPhys },
161	{ "SUPR0MemFree", (void *)SUPR0MemFree },
162	{ "SUPR0PageAllocEx", (void *)SUPR0PageAllocEx },
163	{ "SUPR0PageFree", (void *)SUPR0PageFree },
164	{ "SUPR0Printf", (void )SUPR0Printf }, /* @todo needs wrapping? */
165	{ "SUPSemEventCreate", (void *)SUPSemEventCreate },
166	{ "SUPSemEventClose", (void *)SUPSemEventClose },
167	{ "SUPSemEventSignal", (void *)SUPSemEventSignal },
168	{ "SUPSemEventWait", (void *)SUPSemEventWait },
169	{ "SUPSemEventWaitNoResume", (void *)SUPSemEventWaitNoResume },
170	{ "SUPSemEventMultiCreate", (void *)SUPSemEventMultiCreate },
171	{ "SUPSemEventMultiClose", (void *)SUPSemEventMultiClose },
172	{ "SUPSemEventMultiSignal", (void *)SUPSemEventMultiSignal },
173	{ "SUPSemEventMultiReset", (void *)SUPSemEventMultiReset },
174	{ "SUPSemEventMultiWait", (void *)SUPSemEventMultiWait },
175	{ "SUPSemEventMultiWaitNoResume", (void *)SUPSemEventMultiWaitNoResume },
176	{ "SUPR0GetPagingMode", (void *)SUPR0GetPagingMode },
177	{ "SUPR0EnableVTx", (void *)SUPR0EnableVTx },
178	{ "SUPGetGIP", (void *)SUPGetGIP },
179	{ "g_pSUPGlobalInfoPage", (void *)&g_pSUPGlobalInfoPage },
180	{ "RTMemAlloc", (void *)RTMemAlloc },
181	{ "RTMemAllocZ", (void *)RTMemAllocZ },
182	{ "RTMemFree", (void *)RTMemFree },
183	/{ "RTMemDup", (void )RTMemDup },
184	{ "RTMemDupEx", (void )RTMemDupEx },/
185	{ "RTMemRealloc", (void *)RTMemRealloc },
186	{ "RTR0MemObjAllocLow", (void *)RTR0MemObjAllocLow },
187	{ "RTR0MemObjAllocPage", (void *)RTR0MemObjAllocPage },
188	{ "RTR0MemObjAllocPhys", (void *)RTR0MemObjAllocPhys },
189	{ "RTR0MemObjAllocPhysEx", (void *)RTR0MemObjAllocPhysEx },
190	{ "RTR0MemObjAllocPhysNC", (void *)RTR0MemObjAllocPhysNC },
191	{ "RTR0MemObjAllocCont", (void *)RTR0MemObjAllocCont },
192	{ "RTR0MemObjEnterPhys", (void *)RTR0MemObjEnterPhys },
193	{ "RTR0MemObjLockUser", (void *)RTR0MemObjLockUser },
194	{ "RTR0MemObjMapKernel", (void *)RTR0MemObjMapKernel },
195	{ "RTR0MemObjMapKernelEx", (void *)RTR0MemObjMapKernelEx },
196	{ "RTR0MemObjMapUser", (void *)RTR0MemObjMapUser },
197	{ "RTR0MemObjProtect", (void *)RTR0MemObjProtect },
198	{ "RTR0MemObjAddress", (void *)RTR0MemObjAddress },
199	{ "RTR0MemObjAddressR3", (void *)RTR0MemObjAddressR3 },
200	{ "RTR0MemObjSize", (void *)RTR0MemObjSize },
201	{ "RTR0MemObjIsMapping", (void *)RTR0MemObjIsMapping },
202	{ "RTR0MemObjGetPagePhysAddr", (void *)RTR0MemObjGetPagePhysAddr },
203	{ "RTR0MemObjFree", (void *)RTR0MemObjFree },
204	{ "RTR0MemUserCopyFrom", (void *)RTR0MemUserCopyFrom },
205	{ "RTR0MemUserCopyTo", (void *)RTR0MemUserCopyTo },
206	{ "RTR0MemUserIsValidAddr", (void *)RTR0MemUserIsValidAddr },
207	{ "RTR0MemKernelIsValidAddr", (void *)RTR0MemKernelIsValidAddr },
208	{ "RTR0MemAreKrnlAndUsrDifferent", (void *)RTR0MemAreKrnlAndUsrDifferent },
209	{ "RTSemMutexCreate", (void *)RTSemMutexCreate },
210	{ "RTSemMutexRequest", (void *)RTSemMutexRequest },
211	{ "RTSemMutexRequestDebug", (void *)RTSemMutexRequestDebug },
212	{ "RTSemMutexRequestNoResume", (void *)RTSemMutexRequestNoResume },
213	{ "RTSemMutexRequestNoResumeDebug", (void *)RTSemMutexRequestNoResumeDebug },
214	{ "RTSemMutexRelease", (void *)RTSemMutexRelease },
215	{ "RTSemMutexDestroy", (void *)RTSemMutexDestroy },
216	{ "RTProcSelf", (void *)RTProcSelf },
217	{ "RTR0ProcHandleSelf", (void *)RTR0ProcHandleSelf },
218	{ "RTSemFastMutexCreate", (void *)RTSemFastMutexCreate },
219	{ "RTSemFastMutexDestroy", (void *)RTSemFastMutexDestroy },
220	{ "RTSemFastMutexRequest", (void *)RTSemFastMutexRequest },
221	{ "RTSemFastMutexRelease", (void *)RTSemFastMutexRelease },
222	{ "RTSemEventCreate", (void *)RTSemEventCreate },
223	{ "RTSemEventSignal", (void *)RTSemEventSignal },
224	{ "RTSemEventWait", (void *)RTSemEventWait },
225	{ "RTSemEventWaitNoResume", (void *)RTSemEventWaitNoResume },
226	{ "RTSemEventDestroy", (void *)RTSemEventDestroy },
227	{ "RTSemEventMultiCreate", (void *)RTSemEventMultiCreate },
228	{ "RTSemEventMultiSignal", (void *)RTSemEventMultiSignal },
229	{ "RTSemEventMultiReset", (void *)RTSemEventMultiReset },
230	{ "RTSemEventMultiWait", (void *)RTSemEventMultiWait },
231	{ "RTSemEventMultiWaitNoResume", (void *)RTSemEventMultiWaitNoResume },
232	{ "RTSemEventMultiDestroy", (void *)RTSemEventMultiDestroy },
233	{ "RTSpinlockCreate", (void *)RTSpinlockCreate },
234	{ "RTSpinlockDestroy", (void *)RTSpinlockDestroy },
235	{ "RTSpinlockAcquire", (void *)RTSpinlockAcquire },
236	{ "RTSpinlockRelease", (void *)RTSpinlockRelease },
237	{ "RTSpinlockAcquireNoInts", (void *)RTSpinlockAcquireNoInts },
238	{ "RTSpinlockReleaseNoInts", (void *)RTSpinlockReleaseNoInts },
239	{ "RTTimeNanoTS", (void *)RTTimeNanoTS },
240	{ "RTTimeMilliTS", (void *)RTTimeMilliTS },
241	{ "RTTimeSystemNanoTS", (void *)RTTimeSystemNanoTS },
242	{ "RTTimeSystemMilliTS", (void *)RTTimeSystemMilliTS },
243	{ "RTThreadNativeSelf", (void *)RTThreadNativeSelf },
244	{ "RTThreadSleep", (void *)RTThreadSleep },
245	{ "RTThreadYield", (void *)RTThreadYield },
246	#if 0 /* Thread APIs, Part 2. */
247	{ "RTThreadSelf", (void *)RTThreadSelf },
248	{ "RTThreadCreate", (void *)RTThreadCreate },
249	{ "RTThreadGetNative", (void *)RTThreadGetNative },
250	{ "RTThreadWait", (void *)RTThreadWait },
251	{ "RTThreadWaitNoResume", (void *)RTThreadWaitNoResume },
252	{ "RTThreadGetName", (void *)RTThreadGetName },
253	{ "RTThreadSelfName", (void *)RTThreadSelfName },
254	{ "RTThreadGetType", (void *)RTThreadGetType },
255	{ "RTThreadUserSignal", (void *)RTThreadUserSignal },
256	{ "RTThreadUserReset", (void *)RTThreadUserReset },
257	{ "RTThreadUserWait", (void *)RTThreadUserWait },
258	{ "RTThreadUserWaitNoResume", (void *)RTThreadUserWaitNoResume },
259	#endif
260	{ "RTThreadPreemptIsEnabled", (void *)RTThreadPreemptIsEnabled },
261	{ "RTThreadPreemptIsPending", (void *)RTThreadPreemptIsPending },
262	{ "RTThreadPreemptIsPendingTrusty", (void *)RTThreadPreemptIsPendingTrusty },
263	{ "RTThreadPreemptIsPossible", (void *)RTThreadPreemptIsPossible },
264	{ "RTThreadPreemptDisable", (void *)RTThreadPreemptDisable },
265	{ "RTThreadPreemptRestore", (void *)RTThreadPreemptRestore },
266	{ "RTThreadIsInInterrupt", (void *)RTThreadIsInInterrupt },
267
268	{ "RTLogDefaultInstance", (void *)RTLogDefaultInstance },
269	{ "RTMpCpuId", (void *)RTMpCpuId },
270	{ "RTMpCpuIdFromSetIndex", (void *)RTMpCpuIdFromSetIndex },
271	{ "RTMpCpuIdToSetIndex", (void *)RTMpCpuIdToSetIndex },
272	{ "RTMpIsCpuPossible", (void *)RTMpIsCpuPossible },
273	{ "RTMpGetCount", (void *)RTMpGetCount },
274	{ "RTMpGetMaxCpuId", (void *)RTMpGetMaxCpuId },
275	{ "RTMpGetOnlineCount", (void *)RTMpGetOnlineCount },
276	{ "RTMpGetOnlineSet", (void *)RTMpGetOnlineSet },
277	{ "RTMpGetSet", (void *)RTMpGetSet },
278	{ "RTMpIsCpuOnline", (void *)RTMpIsCpuOnline },
279	{ "RTMpIsCpuWorkPending", (void *)RTMpIsCpuWorkPending },
280	{ "RTMpOnAll", (void *)RTMpOnAll },
281	{ "RTMpOnOthers", (void *)RTMpOnOthers },
282	{ "RTMpOnSpecific", (void *)RTMpOnSpecific },
283	{ "RTMpPokeCpu", (void *)RTMpPokeCpu },
284	{ "RTPowerNotificationRegister", (void *)RTPowerNotificationRegister },
285	{ "RTPowerNotificationDeregister", (void *)RTPowerNotificationDeregister },
286	{ "RTLogRelDefaultInstance", (void *)RTLogRelDefaultInstance },
287	{ "RTLogSetDefaultInstanceThread", (void *)RTLogSetDefaultInstanceThread },
288	{ "RTLogLoggerExV", (void *)RTLogLoggerExV },
289	{ "RTLogPrintfV", (void *)RTLogPrintfV },
290	{ "RTR0AssertPanicSystem", (void *)RTR0AssertPanicSystem },
291	{ "RTAssertMsg1", (void *)RTAssertMsg1 },
292	{ "RTAssertMsg2V", (void *)RTAssertMsg2V },
293	{ "RTAssertSetQuiet", (void *)RTAssertSetQuiet },
294	{ "RTAssertMayPanic", (void *)RTAssertMayPanic },
295	{ "RTAssertSetMayPanic", (void *)RTAssertSetMayPanic },
296	{ "RTAssertAreQuiet", (void *)RTAssertAreQuiet },
297	};
298
299	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
300	/**
301	* Drag in the rest of IRPT since we share it with the
302	* rest of the kernel modules on darwin.
303	*/
304	PFNRT g_apfnVBoxDrvIPRTDeps[] =
305	{
306	/* VBoxNetFlt */
307	(PFNRT)RTCrc32,
308	(PFNRT)RTErrConvertFromErrno,
309	(PFNRT)RTNetIPv4IsHdrValid,
310	(PFNRT)RTNetIPv4TCPChecksum,
311	(PFNRT)RTNetIPv4UDPChecksum,
312	(PFNRT)RTUuidCompare,
313	(PFNRT)RTUuidCompareStr,
314	(PFNRT)RTUuidFromStr,
315	(PFNRT)RTStrDup,
316	(PFNRT)RTStrFree,
317	/* VBoxNetAdp */
318	(PFNRT)RTRandBytes,
319	/* VBoxUSB */
320	(PFNRT)RTPathStripFilename,
321	NULL
322	};
323	#endif /* RT_OS_DARWIN \|\| RT_OS_SOLARIS \|\| RT_OS_SOLARIS */
324
325
326	/**
327	* Initializes the device extentsion structure.
328	*
329	* @returns IPRT status code.
330	* @param pDevExt The device extension to initialize.
331	* @param cbSession The size of the session structure. The size of
332	* SUPDRVSESSION may be smaller when SUPDRV_AGNOSTIC is
333	* defined because we're skipping the OS specific members
334	* then.
335	*/
336	int VBOXCALL supdrvInitDevExt(PSUPDRVDEVEXT pDevExt, size_t cbSession)
337	{
338	int rc;
339
340	#ifdef SUPDRV_WITH_RELEASE_LOGGER
341	/*
342	* Create the release log.
343	*/
344	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
345	PRTLOGGER pRelLogger;
346	rc = RTLogCreate(&pRelLogger, 0 /* fFlags */, "all",
347	"VBOX_RELEASE_LOG", RT_ELEMENTS(s_apszGroups), s_apszGroups,
348	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER, NULL);
349	if (RT_SUCCESS(rc))
350	RTLogRelSetDefaultInstance(pRelLogger);
351	/** @todo Add native hook for getting logger config parameters and setting
352	* them. On linux we should use the module parameter stuff... */
353	#endif
354
355	/*
356	* Initialize it.
357	*/
358	memset(pDevExt, 0, sizeof(*pDevExt));
359	rc = RTSpinlockCreate(&pDevExt->Spinlock);
360	if (RT_SUCCESS(rc))
361	{
362	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
363	rc = RTSemMutexCreate(&pDevExt->mtxLdr);
364	#else
365	rc = RTSemFastMutexCreate(&pDevExt->mtxLdr);
366	#endif
367	if (RT_SUCCESS(rc))
368	{
369	rc = RTSemFastMutexCreate(&pDevExt->mtxComponentFactory);
370	if (RT_SUCCESS(rc))
371	{
372	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
373	rc = RTSemMutexCreate(&pDevExt->mtxGip);
374	#else
375	rc = RTSemFastMutexCreate(&pDevExt->mtxGip);
376	#endif
377	if (RT_SUCCESS(rc))
378	{
379	rc = supdrvGipCreate(pDevExt);
380	if (RT_SUCCESS(rc))
381	{
382	pDevExt->u32Cookie = BIRD; /** @todo make this random? */
383	pDevExt->cbSession = cbSession;
384
385	/*
386	* Fixup the absolute symbols.
387	*
388	* Because of the table indexing assumptions we'll have a little #ifdef orgy
389	* here rather than distributing this to OS specific files. At least for now.
390	*/
391	#ifdef RT_OS_DARWIN
392	# if ARCH_BITS == 32
393	if (SUPR0GetPagingMode() >= SUPPAGINGMODE_AMD64)
394	{
395	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
396	g_aFunctions[1].pfn = (void )0x80; / SUPR0Abs64bitKernelCS - KERNEL64_CS, seg.h */
397	g_aFunctions[2].pfn = (void )0x88; / SUPR0Abs64bitKernelSS - KERNEL64_SS, seg.h */
398	g_aFunctions[3].pfn = (void )0x88; / SUPR0Abs64bitKernelDS - KERNEL64_SS, seg.h */
399	}
400	else
401	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
402	g_aFunctions[4].pfn = (void )0x08; / SUPR0AbsKernelCS - KERNEL_CS, seg.h */
403	g_aFunctions[5].pfn = (void )0x10; / SUPR0AbsKernelSS - KERNEL_DS, seg.h */
404	g_aFunctions[6].pfn = (void )0x10; / SUPR0AbsKernelDS - KERNEL_DS, seg.h */
405	g_aFunctions[7].pfn = (void )0x10; / SUPR0AbsKernelES - KERNEL_DS, seg.h */
406	g_aFunctions[8].pfn = (void )0x10; / SUPR0AbsKernelFS - KERNEL_DS, seg.h */
407	g_aFunctions[9].pfn = (void )0x48; / SUPR0AbsKernelGS - CPU_DATA_GS, seg.h */
408	# else /* 64-bit darwin: */
409	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
410	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
411	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
412	g_aFunctions[3].pfn = (void )0; / SUPR0Abs64bitKernelDS */
413	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
414	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
415	g_aFunctions[6].pfn = (void )0; / SUPR0AbsKernelDS */
416	g_aFunctions[7].pfn = (void )0; / SUPR0AbsKernelES */
417	g_aFunctions[8].pfn = (void )0; / SUPR0AbsKernelFS */
418	g_aFunctions[9].pfn = (void )0; / SUPR0AbsKernelGS */
419
420	# endif
421	#else /* !RT_OS_DARWIN */
422	# if ARCH_BITS == 64
423	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
424	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
425	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
426	g_aFunctions[3].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0Abs64bitKernelDS */
427	# else
428	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
429	# endif
430	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
431	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
432	g_aFunctions[6].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0AbsKernelDS */
433	g_aFunctions[7].pfn = (void )(uintptr_t)ASMGetES(); / SUPR0AbsKernelES */
434	g_aFunctions[8].pfn = (void )(uintptr_t)ASMGetFS(); / SUPR0AbsKernelFS */
435	g_aFunctions[9].pfn = (void )(uintptr_t)ASMGetGS(); / SUPR0AbsKernelGS */
436	#endif /* !RT_OS_DARWIN */
437	return VINF_SUCCESS;
438	}
439
440	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
441	RTSemMutexDestroy(pDevExt->mtxGip);
442	pDevExt->mtxGip = NIL_RTSEMMUTEX;
443	#else
444	RTSemFastMutexDestroy(pDevExt->mtxGip);
445	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
446	#endif
447	}
448	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
449	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
450	}
451	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
452	RTSemMutexDestroy(pDevExt->mtxLdr);
453	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
454	#else
455	RTSemFastMutexDestroy(pDevExt->mtxLdr);
456	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
457	#endif
458	}
459	RTSpinlockDestroy(pDevExt->Spinlock);
460	pDevExt->Spinlock = NIL_RTSPINLOCK;
461	}
462	#ifdef SUPDRV_WITH_RELEASE_LOGGER
463	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
464	RTLogDestroy(RTLogSetDefaultInstance(NULL));
465	#endif
466
467	return rc;
468	}
469
470
471	/**
472	* Delete the device extension (e.g. cleanup members).
473	*
474	* @param pDevExt The device extension to delete.
475	*/
476	void VBOXCALL supdrvDeleteDevExt(PSUPDRVDEVEXT pDevExt)
477	{
478	PSUPDRVOBJ pObj;
479	PSUPDRVUSAGE pUsage;
480
481	/*
482	* Kill mutexes and spinlocks.
483	*/
484	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
485	RTSemMutexDestroy(pDevExt->mtxGip);
486	pDevExt->mtxGip = NIL_RTSEMMUTEX;
487	#else
488	RTSemFastMutexDestroy(pDevExt->mtxGip);
489	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
490	#endif
491	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
492	RTSemMutexDestroy(pDevExt->mtxLdr);
493	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
494	#else
495	RTSemFastMutexDestroy(pDevExt->mtxLdr);
496	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
497	#endif
498	RTSpinlockDestroy(pDevExt->Spinlock);
499	pDevExt->Spinlock = NIL_RTSPINLOCK;
500	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
501	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
502
503	/*
504	* Free lists.
505	*/
506	/* objects. */
507	pObj = pDevExt->pObjs;
508	Assert(!pObj); /* (can trigger on forced unloads) */
509	pDevExt->pObjs = NULL;
510	while (pObj)
511	{
512	void *pvFree = pObj;
513	pObj = pObj->pNext;
514	RTMemFree(pvFree);
515	}
516
517	/* usage records. */
518	pUsage = pDevExt->pUsageFree;
519	pDevExt->pUsageFree = NULL;
520	while (pUsage)
521	{
522	void *pvFree = pUsage;
523	pUsage = pUsage->pNext;
524	RTMemFree(pvFree);
525	}
526
527	/* kill the GIP. */
528	supdrvGipDestroy(pDevExt);
529
530	#ifdef SUPDRV_WITH_RELEASE_LOGGER
531	/* destroy the loggers. */
532	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
533	RTLogDestroy(RTLogSetDefaultInstance(NULL));
534	#endif
535	}
536
537
538	/**
539	* Create session.
540	*
541	* @returns IPRT status code.
542	* @param pDevExt Device extension.
543	* @param fUser Flag indicating whether this is a user or kernel session.
544	* @param ppSession Where to store the pointer to the session data.
545	*/
546	int VBOXCALL supdrvCreateSession(PSUPDRVDEVEXT pDevExt, bool fUser, PSUPDRVSESSION *ppSession)
547	{
548	/*
549	* Allocate memory for the session data.
550	*/
551	int rc;
552	PSUPDRVSESSION pSession = *ppSession = (PSUPDRVSESSION)RTMemAllocZ(pDevExt->cbSession);
553	if (pSession)
554	{
555	/* Initialize session data. */
556	rc = RTSpinlockCreate(&pSession->Spinlock);
557	if (!rc)
558	{
559	rc = RTHandleTableCreateEx(&pSession->hHandleTable,
560	RTHANDLETABLE_FLAGS_LOCKED \| RTHANDLETABLE_FLAGS_CONTEXT,
561	1 /uBase/, 32768 /cMax/, supdrvSessionObjHandleRetain, pSession);
562	if (RT_SUCCESS(rc))
563	{
564	Assert(pSession->Spinlock != NIL_RTSPINLOCK);
565	pSession->pDevExt = pDevExt;
566	pSession->u32Cookie = BIRD_INV;
567	/*pSession->pLdrUsage = NULL;
568	pSession->pVM = NULL;
569	pSession->pUsage = NULL;
570	pSession->pGip = NULL;
571	pSession->fGipReferenced = false;
572	pSession->Bundle.cUsed = 0; */
573	pSession->Uid = NIL_RTUID;
574	pSession->Gid = NIL_RTGID;
575	if (fUser)
576	{
577	pSession->Process = RTProcSelf();
578	pSession->R0Process = RTR0ProcHandleSelf();
579	}
580	else
581	{
582	pSession->Process = NIL_RTPROCESS;
583	pSession->R0Process = NIL_RTR0PROCESS;
584	}
585
586	LogFlow(("Created session %p initial cookie=%#x\n", pSession, pSession->u32Cookie));
587	return VINF_SUCCESS;
588	}
589
590	RTSpinlockDestroy(pSession->Spinlock);
591	}
592	RTMemFree(pSession);
593	*ppSession = NULL;
594	Log(("Failed to create spinlock, rc=%d!\n", rc));
595	}
596	else
597	rc = VERR_NO_MEMORY;
598
599	return rc;
600	}
601
602
603	/**
604	* Shared code for cleaning up a session.
605	*
606	* @param pDevExt Device extension.
607	* @param pSession Session data.
608	* This data will be freed by this routine.
609	*/
610	void VBOXCALL supdrvCloseSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
611	{
612	/*
613	* Cleanup the session first.
614	*/
615	supdrvCleanupSession(pDevExt, pSession);
616
617	/*
618	* Free the rest of the session stuff.
619	*/
620	RTSpinlockDestroy(pSession->Spinlock);
621	pSession->Spinlock = NIL_RTSPINLOCK;
622	pSession->pDevExt = NULL;
623	RTMemFree(pSession);
624	LogFlow(("supdrvCloseSession: returns\n"));
625	}
626
627
628	/**
629	* Shared code for cleaning up a session (but not quite freeing it).
630	*
631	* This is primarily intended for MAC OS X where we have to clean up the memory
632	* stuff before the file handle is closed.
633	*
634	* @param pDevExt Device extension.
635	* @param pSession Session data.
636	* This data will be freed by this routine.
637	*/
638	void VBOXCALL supdrvCleanupSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
639	{
640	int rc;
641	PSUPDRVBUNDLE pBundle;
642	LogFlow(("supdrvCleanupSession: pSession=%p\n", pSession));
643
644	/*
645	* Remove logger instances related to this session.
646	*/
647	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pSession);
648
649	/*
650	* Destroy the handle table.
651	*/
652	rc = RTHandleTableDestroy(pSession->hHandleTable, supdrvSessionObjHandleDelete, pSession);
653	AssertRC(rc);
654	pSession->hHandleTable = NIL_RTHANDLETABLE;
655
656	/*
657	* Release object references made in this session.
658	* In theory there should be noone racing us in this session.
659	*/
660	Log2(("release objects - start\n"));
661	if (pSession->pUsage)
662	{
663	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
664	PSUPDRVUSAGE pUsage;
665	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
666
667	while ((pUsage = pSession->pUsage) != NULL)
668	{
669	PSUPDRVOBJ pObj = pUsage->pObj;
670	pSession->pUsage = pUsage->pNext;
671
672	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
673	if (pUsage->cUsage < pObj->cUsage)
674	{
675	pObj->cUsage -= pUsage->cUsage;
676	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
677	}
678	else
679	{
680	/* Destroy the object and free the record. */
681	if (pDevExt->pObjs == pObj)
682	pDevExt->pObjs = pObj->pNext;
683	else
684	{
685	PSUPDRVOBJ pObjPrev;
686	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
687	if (pObjPrev->pNext == pObj)
688	{
689	pObjPrev->pNext = pObj->pNext;
690	break;
691	}
692	Assert(pObjPrev);
693	}
694	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
695
696	Log(("supdrvCleanupSession: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
697	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
698	if (pObj->pfnDestructor)
699	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
700	RTMemFree(pObj);
701	}
702
703	/* free it and continue. */
704	RTMemFree(pUsage);
705
706	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
707	}
708
709	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
710	AssertMsg(!pSession->pUsage, ("Some buster reregistered an object during desturction!\n"));
711	}
712	Log2(("release objects - done\n"));
713
714	/*
715	* Release memory allocated in the session.
716	*
717	* We do not serialize this as we assume that the application will
718	* not allocated memory while closing the file handle object.
719	*/
720	Log2(("freeing memory:\n"));
721	pBundle = &pSession->Bundle;
722	while (pBundle)
723	{
724	PSUPDRVBUNDLE pToFree;
725	unsigned i;
726
727	/*
728	* Check and unlock all entries in the bundle.
729	*/
730	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
731	{
732	if (pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ)
733	{
734	Log2(("eType=%d pvR0=%p pvR3=%p cb=%ld\n", pBundle->aMem[i].eType, RTR0MemObjAddress(pBundle->aMem[i].MemObj),
735	(void *)RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3), (long)RTR0MemObjSize(pBundle->aMem[i].MemObj)));
736	if (pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ)
737	{
738	rc = RTR0MemObjFree(pBundle->aMem[i].MapObjR3, false);
739	AssertRC(rc); /** @todo figure out how to handle this. */
740	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
741	}
742	rc = RTR0MemObjFree(pBundle->aMem[i].MemObj, true /* fFreeMappings */);
743	AssertRC(rc); /** @todo figure out how to handle this. */
744	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
745	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
746	}
747	}
748
749	/*
750	* Advance and free previous bundle.
751	*/
752	pToFree = pBundle;
753	pBundle = pBundle->pNext;
754
755	pToFree->pNext = NULL;
756	pToFree->cUsed = 0;
757	if (pToFree != &pSession->Bundle)
758	RTMemFree(pToFree);
759	}
760	Log2(("freeing memory - done\n"));
761
762	/*
763	* Deregister component factories.
764	*/
765	RTSemFastMutexRequest(pDevExt->mtxComponentFactory);
766	Log2(("deregistering component factories:\n"));
767	if (pDevExt->pComponentFactoryHead)
768	{
769	PSUPDRVFACTORYREG pPrev = NULL;
770	PSUPDRVFACTORYREG pCur = pDevExt->pComponentFactoryHead;
771	while (pCur)
772	{
773	if (pCur->pSession == pSession)
774	{
775	/* unlink it */
776	PSUPDRVFACTORYREG pNext = pCur->pNext;
777	if (pPrev)
778	pPrev->pNext = pNext;
779	else
780	pDevExt->pComponentFactoryHead = pNext;
781
782	/* free it */
783	pCur->pNext = NULL;
784	pCur->pSession = NULL;
785	pCur->pFactory = NULL;
786	RTMemFree(pCur);
787
788	/* next */
789	pCur = pNext;
790	}
791	else
792	{
793	/* next */
794	pPrev = pCur;
795	pCur = pCur->pNext;
796	}
797	}
798	}
799	RTSemFastMutexRelease(pDevExt->mtxComponentFactory);
800	Log2(("deregistering component factories - done\n"));
801
802	/*
803	* Loaded images needs to be dereferenced and possibly freed up.
804	*/
805	supdrvLdrLock(pDevExt);
806	Log2(("freeing images:\n"));
807	if (pSession->pLdrUsage)
808	{
809	PSUPDRVLDRUSAGE pUsage = pSession->pLdrUsage;
810	pSession->pLdrUsage = NULL;
811	while (pUsage)
812	{
813	void *pvFree = pUsage;
814	PSUPDRVLDRIMAGE pImage = pUsage->pImage;
815	if (pImage->cUsage > pUsage->cUsage)
816	pImage->cUsage -= pUsage->cUsage;
817	else
818	supdrvLdrFree(pDevExt, pImage);
819	pUsage->pImage = NULL;
820	pUsage = pUsage->pNext;
821	RTMemFree(pvFree);
822	}
823	}
824	supdrvLdrUnlock(pDevExt);
825	Log2(("freeing images - done\n"));
826
827	/*
828	* Unmap the GIP.
829	*/
830	Log2(("umapping GIP:\n"));
831	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
832	{
833	SUPR0GipUnmap(pSession);
834	pSession->fGipReferenced = 0;
835	}
836	Log2(("umapping GIP - done\n"));
837	}
838
839
840	/**
841	* RTHandleTableDestroy callback used by supdrvCleanupSession.
842	*
843	* @returns IPRT status code, see SUPR0ObjAddRef.
844	* @param hHandleTable The handle table handle. Ignored.
845	* @param pvObj The object pointer.
846	* @param pvCtx Context, the handle type. Ignored.
847	* @param pvUser Session pointer.
848	*/
849	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser)
850	{
851	NOREF(pvCtx);
852	NOREF(hHandleTable);
853	return SUPR0ObjAddRefEx(pvObj, (PSUPDRVSESSION)pvUser, true /fNoBlocking/);
854	}
855
856
857	/**
858	* RTHandleTableDestroy callback used by supdrvCleanupSession.
859	*
860	* @param hHandleTable The handle table handle. Ignored.
861	* @param h The handle value. Ignored.
862	* @param pvObj The object pointer.
863	* @param pvCtx Context, the handle type. Ignored.
864	* @param pvUser Session pointer.
865	*/
866	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser)
867	{
868	NOREF(pvCtx);
869	NOREF(h);
870	NOREF(hHandleTable);
871	SUPR0ObjRelease(pvObj, (PSUPDRVSESSION)pvUser);
872	}
873
874
875	/**
876	* Fast path I/O Control worker.
877	*
878	* @returns VBox status code that should be passed down to ring-3 unchanged.
879	* @param uIOCtl Function number.
880	* @param idCpu VMCPU id.
881	* @param pDevExt Device extention.
882	* @param pSession Session data.
883	*/
884	int VBOXCALL supdrvIOCtlFast(uintptr_t uIOCtl, VMCPUID idCpu, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
885	{
886	/*
887	* We check the two prereqs after doing this only to allow the compiler to optimize things better.
888	*/
889	if (RT_LIKELY(pSession->pVM && pDevExt->pfnVMMR0EntryFast))
890	{
891	switch (uIOCtl)
892	{
893	case SUP_IOCTL_FAST_DO_RAW_RUN:
894	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_RAW_RUN);
895	break;
896	case SUP_IOCTL_FAST_DO_HWACC_RUN:
897	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_HWACC_RUN);
898	break;
899	case SUP_IOCTL_FAST_DO_NOP:
900	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_NOP);
901	break;
902	default:
903	return VERR_INTERNAL_ERROR;
904	}
905	return VINF_SUCCESS;
906	}
907	return VERR_INTERNAL_ERROR;
908	}
909
910
911	/**
912	* Helper for supdrvIOCtl. Check if pszStr contains any character of pszChars.
913	* We would use strpbrk here if this function would be contained in the RedHat kABI white
914	* list, see http://www.kerneldrivers.org/RHEL5.
915	*
916	* @returns 1 if pszStr does contain any character of pszChars, 0 otherwise.
917	* @param pszStr String to check
918	* @param pszChars Character set
919	*/
920	static int supdrvCheckInvalidChar(const char pszStr, const char pszChars)
921	{
922	int chCur;
923	while ((chCur = *pszStr++) != '\0')
924	{
925	int ch;
926	const char *psz = pszChars;
927	while ((ch = *psz++) != '\0')
928	if (ch == chCur)
929	return 1;
930
931	}
932	return 0;
933	}
934
935
936	/**
937	* I/O Control worker.
938	*
939	* @returns IPRT status code.
940	* @retval VERR_INVALID_PARAMETER if the request is invalid.
941	*
942	* @param uIOCtl Function number.
943	* @param pDevExt Device extention.
944	* @param pSession Session data.
945	* @param pReqHdr The request header.
946	*/
947	int VBOXCALL supdrvIOCtl(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
948	{
949	/*
950	* Validate the request.
951	*/
952	/* this first check could probably be omitted as its also done by the OS specific code... */
953	if (RT_UNLIKELY( (pReqHdr->fFlags & SUPREQHDR_FLAGS_MAGIC_MASK) != SUPREQHDR_FLAGS_MAGIC
954	\|\| pReqHdr->cbIn < sizeof(*pReqHdr)
955	\|\| pReqHdr->cbOut < sizeof(*pReqHdr)))
956	{
957	OSDBGPRINT(("vboxdrv: Bad ioctl request header; cbIn=%#lx cbOut=%#lx fFlags=%#lx\n",
958	(long)pReqHdr->cbIn, (long)pReqHdr->cbOut, (long)pReqHdr->fFlags));
959	return VERR_INVALID_PARAMETER;
960	}
961	if (RT_UNLIKELY(uIOCtl == SUP_IOCTL_COOKIE))
962	{
963	if (pReqHdr->u32Cookie != SUPCOOKIE_INITIAL_COOKIE)
964	{
965	OSDBGPRINT(("SUP_IOCTL_COOKIE: bad cookie %#lx\n", (long)pReqHdr->u32Cookie));
966	return VERR_INVALID_PARAMETER;
967	}
968	}
969	else if (RT_UNLIKELY( pReqHdr->u32Cookie != pDevExt->u32Cookie
970	\|\| pReqHdr->u32SessionCookie != pSession->u32Cookie))
971	{
972	OSDBGPRINT(("vboxdrv: bad cookie %#lx / %#lx.\n", (long)pReqHdr->u32Cookie, (long)pReqHdr->u32SessionCookie));
973	return VERR_INVALID_PARAMETER;
974	}
975
976	/*
977	* Validation macros
978	*/
979	#define REQ_CHECK_SIZES_EX(Name, cbInExpect, cbOutExpect) \
980	do { \
981	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect) \|\| pReqHdr->cbOut != (cbOutExpect))) \
982	{ \
983	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld. cbOut=%ld expected %ld.\n", \
984	(long)pReq->Hdr.cbIn, (long)(cbInExpect), (long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
985	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
986	} \
987	} while (0)
988
989	#define REQ_CHECK_SIZES(Name) REQ_CHECK_SIZES_EX(Name, Name ## _SIZE_IN, Name ## _SIZE_OUT)
990
991	#define REQ_CHECK_SIZE_IN(Name, cbInExpect) \
992	do { \
993	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect))) \
994	{ \
995	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld.\n", \
996	(long)pReq->Hdr.cbIn, (long)(cbInExpect))); \
997	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
998	} \
999	} while (0)
1000
1001	#define REQ_CHECK_SIZE_OUT(Name, cbOutExpect) \
1002	do { \
1003	if (RT_UNLIKELY(pReqHdr->cbOut != (cbOutExpect))) \
1004	{ \
1005	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbOut=%ld expected %ld.\n", \
1006	(long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
1007	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1008	} \
1009	} while (0)
1010
1011	#define REQ_CHECK_EXPR(Name, expr) \
1012	do { \
1013	if (RT_UNLIKELY(!(expr))) \
1014	{ \
1015	OSDBGPRINT(( #Name ": %s\n", #expr)); \
1016	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1017	} \
1018	} while (0)
1019
1020	#define REQ_CHECK_EXPR_FMT(expr, fmt) \
1021	do { \
1022	if (RT_UNLIKELY(!(expr))) \
1023	{ \
1024	OSDBGPRINT( fmt ); \
1025	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1026	} \
1027	} while (0)
1028
1029
1030	/*
1031	* The switch.
1032	*/
1033	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
1034	{
1035	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
1036	{
1037	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
1038	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
1039	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
1040	{
1041	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
1042	pReq->Hdr.rc = VERR_INVALID_MAGIC;
1043	return 0;
1044	}
1045
1046	#if 0
1047	/*
1048	* Call out to the OS specific code and let it do permission checks on the
1049	* client process.
1050	*/
1051	if (!supdrvOSValidateClientProcess(pDevExt, pSession))
1052	{
1053	pReq->u.Out.u32Cookie = 0xffffffff;
1054	pReq->u.Out.u32SessionCookie = 0xffffffff;
1055	pReq->u.Out.u32SessionVersion = 0xffffffff;
1056	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1057	pReq->u.Out.pSession = NULL;
1058	pReq->u.Out.cFunctions = 0;
1059	pReq->Hdr.rc = VERR_PERMISSION_DENIED;
1060	return 0;
1061	}
1062	#endif
1063
1064	/*
1065	* Match the version.
1066	* The current logic is very simple, match the major interface version.
1067	*/
1068	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
1069	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
1070	{
1071	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1072	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
1073	pReq->u.Out.u32Cookie = 0xffffffff;
1074	pReq->u.Out.u32SessionCookie = 0xffffffff;
1075	pReq->u.Out.u32SessionVersion = 0xffffffff;
1076	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1077	pReq->u.Out.pSession = NULL;
1078	pReq->u.Out.cFunctions = 0;
1079	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1080	return 0;
1081	}
1082
1083	/*
1084	* Fill in return data and be gone.
1085	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
1086	* u32SessionVersion <= u32ReqVersion!
1087	*/
1088	/** @todo Somehow validate the client and negotiate a secure cookie... */
1089	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
1090	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
1091	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
1092	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1093	pReq->u.Out.pSession = pSession;
1094	pReq->u.Out.cFunctions = sizeof(g_aFunctions) / sizeof(g_aFunctions[0]);
1095	pReq->Hdr.rc = VINF_SUCCESS;
1096	return 0;
1097	}
1098
1099	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_QUERY_FUNCS(0)):
1100	{
1101	/* validate */
1102	PSUPQUERYFUNCS pReq = (PSUPQUERYFUNCS)pReqHdr;
1103	REQ_CHECK_SIZES_EX(SUP_IOCTL_QUERY_FUNCS, SUP_IOCTL_QUERY_FUNCS_SIZE_IN, SUP_IOCTL_QUERY_FUNCS_SIZE_OUT(RT_ELEMENTS(g_aFunctions)));
1104
1105	/* execute */
1106	pReq->u.Out.cFunctions = RT_ELEMENTS(g_aFunctions);
1107	memcpy(&pReq->u.Out.aFunctions[0], g_aFunctions, sizeof(g_aFunctions));
1108	pReq->Hdr.rc = VINF_SUCCESS;
1109	return 0;
1110	}
1111
1112	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_LOCK):
1113	{
1114	/* validate */
1115	PSUPPAGELOCK pReq = (PSUPPAGELOCK)pReqHdr;
1116	REQ_CHECK_SIZE_IN(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_IN);
1117	REQ_CHECK_SIZE_OUT(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_OUT(pReq->u.In.cPages));
1118	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.cPages > 0);
1119	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.pvR3 >= PAGE_SIZE);
1120
1121	/* execute */
1122	pReq->Hdr.rc = SUPR0LockMem(pSession, pReq->u.In.pvR3, pReq->u.In.cPages, &pReq->u.Out.aPages[0]);
1123	if (RT_FAILURE(pReq->Hdr.rc))
1124	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1125	return 0;
1126	}
1127
1128	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_UNLOCK):
1129	{
1130	/* validate */
1131	PSUPPAGEUNLOCK pReq = (PSUPPAGEUNLOCK)pReqHdr;
1132	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_UNLOCK);
1133
1134	/* execute */
1135	pReq->Hdr.rc = SUPR0UnlockMem(pSession, pReq->u.In.pvR3);
1136	return 0;
1137	}
1138
1139	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_ALLOC):
1140	{
1141	/* validate */
1142	PSUPCONTALLOC pReq = (PSUPCONTALLOC)pReqHdr;
1143	REQ_CHECK_SIZES(SUP_IOCTL_CONT_ALLOC);
1144
1145	/* execute */
1146	pReq->Hdr.rc = SUPR0ContAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.HCPhys);
1147	if (RT_FAILURE(pReq->Hdr.rc))
1148	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1149	return 0;
1150	}
1151
1152	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_FREE):
1153	{
1154	/* validate */
1155	PSUPCONTFREE pReq = (PSUPCONTFREE)pReqHdr;
1156	REQ_CHECK_SIZES(SUP_IOCTL_CONT_FREE);
1157
1158	/* execute */
1159	pReq->Hdr.rc = SUPR0ContFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1160	return 0;
1161	}
1162
1163	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_OPEN):
1164	{
1165	/* validate */
1166	PSUPLDROPEN pReq = (PSUPLDROPEN)pReqHdr;
1167	REQ_CHECK_SIZES(SUP_IOCTL_LDR_OPEN);
1168	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs > 0);
1169	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs < 16*_1M);
1170	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1171	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1172	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits < pReq->u.In.cbImageWithTabs);
1173	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.szName[0]);
1174	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, memchr(pReq->u.In.szName, '\0', sizeof(pReq->u.In.szName)));
1175	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, !supdrvCheckInvalidChar(pReq->u.In.szName, ";:()[]{}/\\\|&*%#@!~`\"'"));
1176	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, memchr(pReq->u.In.szFilename, '\0', sizeof(pReq->u.In.szFilename)));
1177
1178	/* execute */
1179	pReq->Hdr.rc = supdrvIOCtl_LdrOpen(pDevExt, pSession, pReq);
1180	return 0;
1181	}
1182
1183	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOAD):
1184	{
1185	/* validate */
1186	PSUPLDRLOAD pReq = (PSUPLDRLOAD)pReqHdr;
1187	REQ_CHECK_EXPR(Name, pReq->Hdr.cbIn >= sizeof(*pReq));
1188	REQ_CHECK_SIZES_EX(SUP_IOCTL_LDR_LOAD, SUP_IOCTL_LDR_LOAD_SIZE_IN(pReq->u.In.cbImageWithTabs), SUP_IOCTL_LDR_LOAD_SIZE_OUT);
1189	REQ_CHECK_EXPR(SUP_IOCTL_LDR_LOAD, pReq->u.In.cSymbols <= 16384);
1190	REQ_CHECK_EXPR_FMT( !pReq->u.In.cSymbols
1191	\|\| ( pReq->u.In.offSymbols < pReq->u.In.cbImageWithTabs
1192	&& pReq->u.In.offSymbols + pReq->u.In.cSymbols * sizeof(SUPLDRSYM) <= pReq->u.In.cbImageWithTabs),
1193	("SUP_IOCTL_LDR_LOAD: offSymbols=%#lx cSymbols=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offSymbols,
1194	(long)pReq->u.In.cSymbols, (long)pReq->u.In.cbImageWithTabs));
1195	REQ_CHECK_EXPR_FMT( !pReq->u.In.cbStrTab
1196	\|\| ( pReq->u.In.offStrTab < pReq->u.In.cbImageWithTabs
1197	&& pReq->u.In.offStrTab + pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs
1198	&& pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs),
1199	("SUP_IOCTL_LDR_LOAD: offStrTab=%#lx cbStrTab=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offStrTab,
1200	(long)pReq->u.In.cbStrTab, (long)pReq->u.In.cbImageWithTabs));
1201
1202	if (pReq->u.In.cSymbols)
1203	{
1204	uint32_t i;
1205	PSUPLDRSYM paSyms = (PSUPLDRSYM)&pReq->u.In.abImage[pReq->u.In.offSymbols];
1206	for (i = 0; i < pReq->u.In.cSymbols; i++)
1207	{
1208	REQ_CHECK_EXPR_FMT(paSyms[i].offSymbol < pReq->u.In.cbImageWithTabs,
1209	("SUP_IOCTL_LDR_LOAD: sym #%ld: symb off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offSymbol, (long)pReq->u.In.cbImageWithTabs));
1210	REQ_CHECK_EXPR_FMT(paSyms[i].offName < pReq->u.In.cbStrTab,
1211	("SUP_IOCTL_LDR_LOAD: sym #%ld: name off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1212	REQ_CHECK_EXPR_FMT(memchr(&pReq->u.In.abImage[pReq->u.In.offStrTab + paSyms[i].offName], '\0', pReq->u.In.cbStrTab - paSyms[i].offName),
1213	("SUP_IOCTL_LDR_LOAD: sym #%ld: unterminated name! (%#lx / %#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1214	}
1215	}
1216
1217	/* execute */
1218	pReq->Hdr.rc = supdrvIOCtl_LdrLoad(pDevExt, pSession, pReq);
1219	return 0;
1220	}
1221
1222	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_FREE):
1223	{
1224	/* validate */
1225	PSUPLDRFREE pReq = (PSUPLDRFREE)pReqHdr;
1226	REQ_CHECK_SIZES(SUP_IOCTL_LDR_FREE);
1227
1228	/* execute */
1229	pReq->Hdr.rc = supdrvIOCtl_LdrFree(pDevExt, pSession, pReq);
1230	return 0;
1231	}
1232
1233	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_GET_SYMBOL):
1234	{
1235	/* validate */
1236	PSUPLDRGETSYMBOL pReq = (PSUPLDRGETSYMBOL)pReqHdr;
1237	REQ_CHECK_SIZES(SUP_IOCTL_LDR_GET_SYMBOL);
1238	REQ_CHECK_EXPR(SUP_IOCTL_LDR_GET_SYMBOL, memchr(pReq->u.In.szSymbol, '\0', sizeof(pReq->u.In.szSymbol)));
1239
1240	/* execute */
1241	pReq->Hdr.rc = supdrvIOCtl_LdrGetSymbol(pDevExt, pSession, pReq);
1242	return 0;
1243	}
1244
1245	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0(0)):
1246	{
1247	/* validate */
1248	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1249	Log4(("SUP_IOCTL_CALL_VMMR0: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1250	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1251
1252	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_VMMR0_SIZE(0))
1253	{
1254	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(0), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(0));
1255
1256	/* execute */
1257	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1258	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg, pSession);
1259	else
1260	pReq->Hdr.rc = VERR_WRONG_ORDER;
1261	}
1262	else
1263	{
1264	PSUPVMMR0REQHDR pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1265	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR)),
1266	("SUP_IOCTL_CALL_VMMR0: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR))));
1267	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1268	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(pVMMReq->cbReq));
1269
1270	/* execute */
1271	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1272	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1273	else
1274	pReq->Hdr.rc = VERR_WRONG_ORDER;
1275	}
1276
1277	if ( RT_FAILURE(pReq->Hdr.rc)
1278	&& pReq->Hdr.rc != VERR_INTERRUPTED
1279	&& pReq->Hdr.rc != VERR_TIMEOUT)
1280	Log(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1281	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1282	else
1283	Log4(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1284	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1285	return 0;
1286	}
1287
1288	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_PAGING_MODE):
1289	{
1290	/* validate */
1291	PSUPGETPAGINGMODE pReq = (PSUPGETPAGINGMODE)pReqHdr;
1292	REQ_CHECK_SIZES(SUP_IOCTL_GET_PAGING_MODE);
1293
1294	/* execute */
1295	pReq->Hdr.rc = VINF_SUCCESS;
1296	pReq->u.Out.enmMode = SUPR0GetPagingMode();
1297	return 0;
1298	}
1299
1300	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_ALLOC):
1301	{
1302	/* validate */
1303	PSUPLOWALLOC pReq = (PSUPLOWALLOC)pReqHdr;
1304	REQ_CHECK_EXPR(SUP_IOCTL_LOW_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_LOW_ALLOC_SIZE_IN);
1305	REQ_CHECK_SIZES_EX(SUP_IOCTL_LOW_ALLOC, SUP_IOCTL_LOW_ALLOC_SIZE_IN, SUP_IOCTL_LOW_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1306
1307	/* execute */
1308	pReq->Hdr.rc = SUPR0LowAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1309	if (RT_FAILURE(pReq->Hdr.rc))
1310	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1311	return 0;
1312	}
1313
1314	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_FREE):
1315	{
1316	/* validate */
1317	PSUPLOWFREE pReq = (PSUPLOWFREE)pReqHdr;
1318	REQ_CHECK_SIZES(SUP_IOCTL_LOW_FREE);
1319
1320	/* execute */
1321	pReq->Hdr.rc = SUPR0LowFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1322	return 0;
1323	}
1324
1325	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_MAP):
1326	{
1327	/* validate */
1328	PSUPGIPMAP pReq = (PSUPGIPMAP)pReqHdr;
1329	REQ_CHECK_SIZES(SUP_IOCTL_GIP_MAP);
1330
1331	/* execute */
1332	pReq->Hdr.rc = SUPR0GipMap(pSession, &pReq->u.Out.pGipR3, &pReq->u.Out.HCPhysGip);
1333	if (RT_SUCCESS(pReq->Hdr.rc))
1334	pReq->u.Out.pGipR0 = pDevExt->pGip;
1335	return 0;
1336	}
1337
1338	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_UNMAP):
1339	{
1340	/* validate */
1341	PSUPGIPUNMAP pReq = (PSUPGIPUNMAP)pReqHdr;
1342	REQ_CHECK_SIZES(SUP_IOCTL_GIP_UNMAP);
1343
1344	/* execute */
1345	pReq->Hdr.rc = SUPR0GipUnmap(pSession);
1346	return 0;
1347	}
1348
1349	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SET_VM_FOR_FAST):
1350	{
1351	/* validate */
1352	PSUPSETVMFORFAST pReq = (PSUPSETVMFORFAST)pReqHdr;
1353	REQ_CHECK_SIZES(SUP_IOCTL_SET_VM_FOR_FAST);
1354	REQ_CHECK_EXPR_FMT( !pReq->u.In.pVMR0
1355	\|\| ( VALID_PTR(pReq->u.In.pVMR0)
1356	&& !((uintptr_t)pReq->u.In.pVMR0 & (PAGE_SIZE - 1))),
1357	("SUP_IOCTL_SET_VM_FOR_FAST: pVMR0=%p!\n", pReq->u.In.pVMR0));
1358	/* execute */
1359	pSession->pVM = pReq->u.In.pVMR0;
1360	pReq->Hdr.rc = VINF_SUCCESS;
1361	return 0;
1362	}
1363
1364	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_ALLOC_EX):
1365	{
1366	/* validate */
1367	PSUPPAGEALLOCEX pReq = (PSUPPAGEALLOCEX)pReqHdr;
1368	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC_EX, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_EX_SIZE_IN);
1369	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));
1370	REQ_CHECK_EXPR_FMT(pReq->u.In.fKernelMapping \|\| pReq->u.In.fUserMapping,
1371	("SUP_IOCTL_PAGE_ALLOC_EX: No mapping requested!\n"));
1372	REQ_CHECK_EXPR_FMT(pReq->u.In.fUserMapping,
1373	("SUP_IOCTL_PAGE_ALLOC_EX: Must have user mapping!\n"));
1374	REQ_CHECK_EXPR_FMT(!pReq->u.In.fReserved0 && !pReq->u.In.fReserved1,
1375	("SUP_IOCTL_PAGE_ALLOC_EX: fReserved0=%d fReserved1=%d\n", pReq->u.In.fReserved0, pReq->u.In.fReserved1));
1376
1377	/* execute */
1378	pReq->Hdr.rc = SUPR0PageAllocEx(pSession, pReq->u.In.cPages, 0 /* fFlags */,
1379	pReq->u.In.fUserMapping ? &pReq->u.Out.pvR3 : NULL,
1380	pReq->u.In.fKernelMapping ? &pReq->u.Out.pvR0 : NULL,
1381	&pReq->u.Out.aPages[0]);
1382	if (RT_FAILURE(pReq->Hdr.rc))
1383	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1384	return 0;
1385	}
1386
1387	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_MAP_KERNEL):
1388	{
1389	/* validate */
1390	PSUPPAGEMAPKERNEL pReq = (PSUPPAGEMAPKERNEL)pReqHdr;
1391	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_MAP_KERNEL);
1392	REQ_CHECK_EXPR_FMT(!pReq->u.In.fFlags, ("SUP_IOCTL_PAGE_MAP_KERNEL: fFlags=%#x! MBZ\n", pReq->u.In.fFlags));
1393	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_MAP_KERNEL: offSub=%#x\n", pReq->u.In.offSub));
1394	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1395	("SUP_IOCTL_PAGE_MAP_KERNEL: cbSub=%#x\n", pReq->u.In.cbSub));
1396
1397	/* execute */
1398	pReq->Hdr.rc = SUPR0PageMapKernel(pSession, pReq->u.In.pvR3, pReq->u.In.offSub, pReq->u.In.cbSub,
1399	pReq->u.In.fFlags, &pReq->u.Out.pvR0);
1400	if (RT_FAILURE(pReq->Hdr.rc))
1401	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1402	return 0;
1403	}
1404
1405	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_PROTECT):
1406	{
1407	/* validate */
1408	PSUPPAGEPROTECT pReq = (PSUPPAGEPROTECT)pReqHdr;
1409	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_PROTECT);
1410	REQ_CHECK_EXPR_FMT(!(pReq->u.In.fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)),
1411	("SUP_IOCTL_PAGE_PROTECT: fProt=%#x!\n", pReq->u.In.fProt));
1412	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_PROTECT: offSub=%#x\n", pReq->u.In.offSub));
1413	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1414	("SUP_IOCTL_PAGE_PROTECT: cbSub=%#x\n", pReq->u.In.cbSub));
1415
1416	/* execute */
1417	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);
1418	return 0;
1419	}
1420
1421	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_FREE):
1422	{
1423	/* validate */
1424	PSUPPAGEFREE pReq = (PSUPPAGEFREE)pReqHdr;
1425	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_FREE);
1426
1427	/* execute */
1428	pReq->Hdr.rc = SUPR0PageFree(pSession, pReq->u.In.pvR3);
1429	return 0;
1430	}
1431
1432	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_SERVICE(0)):
1433	{
1434	/* validate */
1435	PSUPCALLSERVICE pReq = (PSUPCALLSERVICE)pReqHdr;
1436	Log4(("SUP_IOCTL_CALL_SERVICE: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1437	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1438
1439	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
1440	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(0), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(0));
1441	else
1442	{
1443	PSUPR0SERVICEREQHDR pSrvReq = (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0];
1444	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR)),
1445	("SUP_IOCTL_CALL_SERVICE: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR))));
1446	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, pSrvReq->u32Magic == SUPR0SERVICEREQHDR_MAGIC);
1447	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(pSrvReq->cbReq), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(pSrvReq->cbReq));
1448	}
1449	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, memchr(pReq->u.In.szName, '\0', sizeof(pReq->u.In.szName)));
1450
1451	/* execute */
1452	pReq->Hdr.rc = supdrvIOCtl_CallServiceModule(pDevExt, pSession, pReq);
1453	return 0;
1454	}
1455
1456	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOGGER_SETTINGS(0)):
1457	{
1458	/* validate */
1459	PSUPLOGGERSETTINGS pReq = (PSUPLOGGERSETTINGS)pReqHdr;
1460	size_t cbStrTab;
1461	REQ_CHECK_SIZE_OUT(SUP_IOCTL_LOGGER_SETTINGS, SUP_IOCTL_LOGGER_SETTINGS_SIZE_OUT);
1462	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->Hdr.cbIn >= SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(1));
1463	cbStrTab = pReq->Hdr.cbIn - SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(0);
1464	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offGroups < cbStrTab);
1465	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offFlags < cbStrTab);
1466	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offDestination < cbStrTab);
1467	REQ_CHECK_EXPR_FMT(pReq->u.In.szStrings[cbStrTab - 1] == '\0',
1468	("SUP_IOCTL_LOGGER_SETTINGS: cbIn=%#x cbStrTab=%#zx LastChar=%d\n",
1469	pReq->Hdr.cbIn, cbStrTab, pReq->u.In.szStrings[cbStrTab - 1]));
1470	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhich <= SUPLOGGERSETTINGS_WHICH_RELEASE);
1471	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhat <= SUPLOGGERSETTINGS_WHAT_DESTROY);
1472
1473	/* execute */
1474	pReq->Hdr.rc = supdrvIOCtl_LoggerSettings(pDevExt, pSession, pReq);
1475	return 0;
1476	}
1477
1478	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_CREATE):
1479	{
1480	/* validate */
1481	PSUPSEMCREATE pReq = (PSUPSEMCREATE)pReqHdr;
1482	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_CREATE, SUP_IOCTL_SEM_CREATE_SIZE_IN, SUP_IOCTL_SEM_CREATE_SIZE_OUT);
1483
1484	/* execute */
1485	switch (pReq->u.In.uType)
1486	{
1487	case SUP_SEM_TYPE_EVENT:
1488	{
1489	SUPSEMEVENT hEvent;
1490	pReq->Hdr.rc = SUPSemEventCreate(pSession, &hEvent);
1491	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEvent;
1492	break;
1493	}
1494
1495	case SUP_SEM_TYPE_EVENT_MULTI:
1496	{
1497	SUPSEMEVENTMULTI hEventMulti;
1498	pReq->Hdr.rc = SUPSemEventMultiCreate(pSession, &hEventMulti);
1499	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEventMulti;
1500	break;
1501	}
1502
1503	default:
1504	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1505	break;
1506	}
1507	return 0;
1508	}
1509
1510	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP):
1511	{
1512	/* validate */
1513	PSUPSEMOP pReq = (PSUPSEMOP)pReqHdr;
1514	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP, SUP_IOCTL_SEM_OP_SIZE_IN, SUP_IOCTL_SEM_OP_SIZE_OUT);
1515
1516	/* execute */
1517	switch (pReq->u.In.uType)
1518	{
1519	case SUP_SEM_TYPE_EVENT:
1520	{
1521	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1522	switch (pReq->u.In.uOp)
1523	{
1524	case SUPSEMOP_WAIT:
1525	pReq->Hdr.rc = SUPSemEventWaitNoResume(pSession, hEvent, pReq->u.In.cMillies);
1526	break;
1527	case SUPSEMOP_SIGNAL:
1528	pReq->Hdr.rc = SUPSemEventSignal(pSession, hEvent);
1529	break;
1530	case SUPSEMOP_CLOSE:
1531	pReq->Hdr.rc = SUPSemEventClose(pSession, hEvent);
1532	break;
1533	case SUPSEMOP_RESET:
1534	default:
1535	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1536	break;
1537	}
1538	break;
1539	}
1540
1541	case SUP_SEM_TYPE_EVENT_MULTI:
1542	{
1543	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1544	switch (pReq->u.In.uOp)
1545	{
1546	case SUPSEMOP_WAIT:
1547	pReq->Hdr.rc = SUPSemEventMultiWaitNoResume(pSession, hEventMulti, pReq->u.In.cMillies);
1548	break;
1549	case SUPSEMOP_SIGNAL:
1550	pReq->Hdr.rc = SUPSemEventMultiSignal(pSession, hEventMulti);
1551	break;
1552	case SUPSEMOP_CLOSE:
1553	pReq->Hdr.rc = SUPSemEventMultiClose(pSession, hEventMulti);
1554	break;
1555	case SUPSEMOP_RESET:
1556	pReq->Hdr.rc = SUPSemEventMultiReset(pSession, hEventMulti);
1557	break;
1558	default:
1559	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1560	break;
1561	}
1562	break;
1563	}
1564
1565	default:
1566	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1567	break;
1568	}
1569	return 0;
1570	}
1571
1572	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
1573	{
1574	/* validate */
1575	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
1576	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
1577	REQ_CHECK_EXPR(SUP_IOCTL_VT_CAPS, pReq->Hdr.cbIn <= SUP_IOCTL_VT_CAPS_SIZE_IN);
1578
1579	/* execute */
1580	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.Caps);
1581	if (RT_FAILURE(pReq->Hdr.rc))
1582	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1583	return 0;
1584	}
1585
1586	default:
1587	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1588	break;
1589	}
1590	return VERR_GENERAL_FAILURE;
1591	}
1592
1593
1594	/**
1595	* Inter-Driver Communcation (IDC) worker.
1596	*
1597	* @returns VBox status code.
1598	* @retval VINF_SUCCESS on success.
1599	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1600	* @retval VERR_NOT_SUPPORTED if the request isn't supported.
1601	*
1602	* @param uReq The request (function) code.
1603	* @param pDevExt Device extention.
1604	* @param pSession Session data.
1605	* @param pReqHdr The request header.
1606	*/
1607	int VBOXCALL supdrvIDC(uintptr_t uReq, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQHDR pReqHdr)
1608	{
1609	/*
1610	* The OS specific code has already validated the pSession
1611	* pointer, and the request size being greater or equal to
1612	* size of the header.
1613	*
1614	* So, just check that pSession is a kernel context session.
1615	*/
1616	if (RT_UNLIKELY( pSession
1617	&& pSession->R0Process != NIL_RTR0PROCESS))
1618	return VERR_INVALID_PARAMETER;
1619
1620	/*
1621	* Validation macro.
1622	*/
1623	#define REQ_CHECK_IDC_SIZE(Name, cbExpect) \
1624	do { \
1625	if (RT_UNLIKELY(pReqHdr->cb != (cbExpect))) \
1626	{ \
1627	OSDBGPRINT(( #Name ": Invalid input/output sizes. cb=%ld expected %ld.\n", \
1628	(long)pReqHdr->cb, (long)(cbExpect))); \
1629	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1630	} \
1631	} while (0)
1632
1633	switch (uReq)
1634	{
1635	case SUPDRV_IDC_REQ_CONNECT:
1636	{
1637	PSUPDRVIDCREQCONNECT pReq = (PSUPDRVIDCREQCONNECT)pReqHdr;
1638	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_CONNECT, sizeof(*pReq));
1639
1640	/*
1641	* Validate the cookie and other input.
1642	*/
1643	if (pReq->Hdr.pSession != NULL)
1644	{
1645	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: pSession=%p expected NULL!\n", pReq->Hdr.pSession));
1646	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1647	}
1648	if (pReq->u.In.u32MagicCookie != SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE)
1649	{
1650	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: u32MagicCookie=%#x expected %#x!\n",
1651	(unsigned)pReq->u.In.u32MagicCookie, (unsigned)SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE));
1652	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1653	}
1654	if ( pReq->u.In.uMinVersion > pReq->u.In.uReqVersion
1655	\|\| (pReq->u.In.uMinVersion & UINT32_C(0xffff0000)) != (pReq->u.In.uReqVersion & UINT32_C(0xffff0000)))
1656	{
1657	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: uMinVersion=%#x uMaxVersion=%#x doesn't match!\n",
1658	pReq->u.In.uMinVersion, pReq->u.In.uReqVersion));
1659	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1660	}
1661
1662	/*
1663	* Match the version.
1664	* The current logic is very simple, match the major interface version.
1665	*/
1666	if ( pReq->u.In.uMinVersion > SUPDRV_IDC_VERSION
1667	\|\| (pReq->u.In.uMinVersion & 0xffff0000) != (SUPDRV_IDC_VERSION & 0xffff0000))
1668	{
1669	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1670	pReq->u.In.uReqVersion, pReq->u.In.uMinVersion, (unsigned)SUPDRV_IDC_VERSION));
1671	pReq->u.Out.pSession = NULL;
1672	pReq->u.Out.uSessionVersion = 0xffffffff;
1673	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1674	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1675	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1676	return VINF_SUCCESS;
1677	}
1678
1679	pReq->u.Out.pSession = NULL;
1680	pReq->u.Out.uSessionVersion = SUPDRV_IDC_VERSION;
1681	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1682	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1683
1684	/*
1685	* On NT we will already have a session associated with the
1686	* client, just like with the SUP_IOCTL_COOKIE request, while
1687	* the other doesn't.
1688	*/
1689	#ifdef RT_OS_WINDOWS
1690	pReq->Hdr.rc = VINF_SUCCESS;
1691	#else
1692	AssertReturn(!pSession, VERR_INTERNAL_ERROR);
1693	pReq->Hdr.rc = supdrvCreateSession(pDevExt, false /* fUser */, &pSession);
1694	if (RT_FAILURE(pReq->Hdr.rc))
1695	{
1696	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: failed to create session, rc=%d\n", pReq->Hdr.rc));
1697	return VINF_SUCCESS;
1698	}
1699	#endif
1700
1701	pReq->u.Out.pSession = pSession;
1702	pReq->Hdr.pSession = pSession;
1703
1704	return VINF_SUCCESS;
1705	}
1706
1707	case SUPDRV_IDC_REQ_DISCONNECT:
1708	{
1709	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_DISCONNECT, sizeof(*pReqHdr));
1710
1711	#ifdef RT_OS_WINDOWS
1712	/* Windows will destroy the session when the file object is destroyed. */
1713	#else
1714	supdrvCloseSession(pDevExt, pSession);
1715	#endif
1716	return pReqHdr->rc = VINF_SUCCESS;
1717	}
1718
1719	case SUPDRV_IDC_REQ_GET_SYMBOL:
1720	{
1721	PSUPDRVIDCREQGETSYM pReq = (PSUPDRVIDCREQGETSYM)pReqHdr;
1722	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_GET_SYMBOL, sizeof(*pReq));
1723
1724	pReq->Hdr.rc = supdrvIDC_LdrGetSymbol(pDevExt, pSession, pReq);
1725	return VINF_SUCCESS;
1726	}
1727
1728	case SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY:
1729	{
1730	PSUPDRVIDCREQCOMPREGFACTORY pReq = (PSUPDRVIDCREQCOMPREGFACTORY)pReqHdr;
1731	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY, sizeof(*pReq));
1732
1733	pReq->Hdr.rc = SUPR0ComponentRegisterFactory(pSession, pReq->u.In.pFactory);
1734	return VINF_SUCCESS;
1735	}
1736
1737	case SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY:
1738	{
1739	PSUPDRVIDCREQCOMPDEREGFACTORY pReq = (PSUPDRVIDCREQCOMPDEREGFACTORY)pReqHdr;
1740	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY, sizeof(*pReq));
1741
1742	pReq->Hdr.rc = SUPR0ComponentDeregisterFactory(pSession, pReq->u.In.pFactory);
1743	return VINF_SUCCESS;
1744	}
1745
1746	default:
1747	Log(("Unknown IDC %#lx\n", (long)uReq));
1748	break;
1749	}
1750
1751	#undef REQ_CHECK_IDC_SIZE
1752	return VERR_NOT_SUPPORTED;
1753	}
1754
1755
1756	/**
1757	* Register a object for reference counting.
1758	* The object is registered with one reference in the specified session.
1759	*
1760	* @returns Unique identifier on success (pointer).
1761	* All future reference must use this identifier.
1762	* @returns NULL on failure.
1763	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
1764	* @param pvUser1 The first user argument.
1765	* @param pvUser2 The second user argument.
1766	*/
1767	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
1768	{
1769	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1770	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1771	PSUPDRVOBJ pObj;
1772	PSUPDRVUSAGE pUsage;
1773
1774	/*
1775	* Validate the input.
1776	*/
1777	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
1778	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
1779	AssertPtrReturn(pfnDestructor, NULL);
1780
1781	/*
1782	* Allocate and initialize the object.
1783	*/
1784	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
1785	if (!pObj)
1786	return NULL;
1787	pObj->u32Magic = SUPDRVOBJ_MAGIC;
1788	pObj->enmType = enmType;
1789	pObj->pNext = NULL;
1790	pObj->cUsage = 1;
1791	pObj->pfnDestructor = pfnDestructor;
1792	pObj->pvUser1 = pvUser1;
1793	pObj->pvUser2 = pvUser2;
1794	pObj->CreatorUid = pSession->Uid;
1795	pObj->CreatorGid = pSession->Gid;
1796	pObj->CreatorProcess= pSession->Process;
1797	supdrvOSObjInitCreator(pObj, pSession);
1798
1799	/*
1800	* Allocate the usage record.
1801	* (We keep freed usage records around to simplify SUPR0ObjAddRefEx().)
1802	*/
1803	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1804
1805	pUsage = pDevExt->pUsageFree;
1806	if (pUsage)
1807	pDevExt->pUsageFree = pUsage->pNext;
1808	else
1809	{
1810	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1811	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
1812	if (!pUsage)
1813	{
1814	RTMemFree(pObj);
1815	return NULL;
1816	}
1817	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1818	}
1819
1820	/*
1821	* Insert the object and create the session usage record.
1822	*/
1823	/* The object. */
1824	pObj->pNext = pDevExt->pObjs;
1825	pDevExt->pObjs = pObj;
1826
1827	/* The session record. */
1828	pUsage->cUsage = 1;
1829	pUsage->pObj = pObj;
1830	pUsage->pNext = pSession->pUsage;
1831	/* Log2(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext)); */
1832	pSession->pUsage = pUsage;
1833
1834	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1835
1836	Log(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
1837	return pObj;
1838	}
1839
1840
1841	/**
1842	* Increment the reference counter for the object associating the reference
1843	* with the specified session.
1844	*
1845	* @returns IPRT status code.
1846	* @param pvObj The identifier returned by SUPR0ObjRegister().
1847	* @param pSession The session which is referencing the object.
1848	*
1849	* @remarks The caller should not own any spinlocks and must carefully protect
1850	* itself against potential race with the destructor so freed memory
1851	* isn't accessed here.
1852	*/
1853	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
1854	{
1855	return SUPR0ObjAddRefEx(pvObj, pSession, false /* fNoBlocking */);
1856	}
1857
1858
1859	/**
1860	* Increment the reference counter for the object associating the reference
1861	* with the specified session.
1862	*
1863	* @returns IPRT status code.
1864	* @retval VERR_TRY_AGAIN if fNoBlocking was set and a new usage record
1865	* couldn't be allocated. (If you see this you're not doing the right
1866	* thing and it won't ever work reliably.)
1867	*
1868	* @param pvObj The identifier returned by SUPR0ObjRegister().
1869	* @param pSession The session which is referencing the object.
1870	* @param fNoBlocking Set if it's not OK to block. Never try to make the
1871	* first reference to an object in a session with this
1872	* argument set.
1873	*
1874	* @remarks The caller should not own any spinlocks and must carefully protect
1875	* itself against potential race with the destructor so freed memory
1876	* isn't accessed here.
1877	*/
1878	SUPR0DECL(int) SUPR0ObjAddRefEx(void *pvObj, PSUPDRVSESSION pSession, bool fNoBlocking)
1879	{
1880	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1881	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1882	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
1883	int rc = VINF_SUCCESS;
1884	PSUPDRVUSAGE pUsagePre;
1885	PSUPDRVUSAGE pUsage;
1886
1887	/*
1888	* Validate the input.
1889	* Be ready for the destruction race (someone might be stuck in the
1890	* destructor waiting a lock we own).
1891	*/
1892	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1893	AssertPtrReturn(pObj, VERR_INVALID_POINTER);
1894	AssertMsgReturn(pObj->u32Magic == SUPDRVOBJ_MAGIC \|\| pObj->u32Magic == SUPDRVOBJ_MAGIC_DEAD,
1895	("Invalid pvObj=%p magic=%#x (expected %#x or %#x)\n", pvObj, pObj->u32Magic, SUPDRVOBJ_MAGIC, SUPDRVOBJ_MAGIC_DEAD),
1896	VERR_INVALID_PARAMETER);
1897
1898	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1899
1900	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
1901	{
1902	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1903
1904	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
1905	return VERR_WRONG_ORDER;
1906	}
1907
1908	/*
1909	* Preallocate the usage record if we can.
1910	*/
1911	pUsagePre = pDevExt->pUsageFree;
1912	if (pUsagePre)
1913	pDevExt->pUsageFree = pUsagePre->pNext;
1914	else if (!fNoBlocking)
1915	{
1916	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1917	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
1918	if (!pUsagePre)
1919	return VERR_NO_MEMORY;
1920
1921	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1922	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
1923	{
1924	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1925
1926	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
1927	return VERR_WRONG_ORDER;
1928	}
1929	}
1930
1931	/*
1932	* Reference the object.
1933	*/
1934	pObj->cUsage++;
1935
1936	/*
1937	* Look for the session record.
1938	*/
1939	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
1940	{
1941	/Log(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
1942	if (pUsage->pObj == pObj)
1943	break;
1944	}
1945	if (pUsage)
1946	pUsage->cUsage++;
1947	else if (pUsagePre)
1948	{
1949	/* create a new session record. */
1950	pUsagePre->cUsage = 1;
1951	pUsagePre->pObj = pObj;
1952	pUsagePre->pNext = pSession->pUsage;
1953	pSession->pUsage = pUsagePre;
1954	/Log(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));/
1955
1956	pUsagePre = NULL;
1957	}
1958	else
1959	{
1960	pObj->cUsage--;
1961	rc = VERR_TRY_AGAIN;
1962	}
1963
1964	/*
1965	* Put any unused usage record into the free list..
1966	*/
1967	if (pUsagePre)
1968	{
1969	pUsagePre->pNext = pDevExt->pUsageFree;
1970	pDevExt->pUsageFree = pUsagePre;
1971	}
1972
1973	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1974
1975	return rc;
1976	}
1977
1978
1979	/**
1980	* Decrement / destroy a reference counter record for an object.
1981	*
1982	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
1983	*
1984	* @returns IPRT status code.
1985	* @retval VINF_SUCCESS if not destroyed.
1986	* @retval VINF_OBJECT_DESTROYED if it's destroyed by this release call.
1987	* @retval VERR_INVALID_PARAMETER if the object isn't valid. Will assert in
1988	* string builds.
1989	*
1990	* @param pvObj The identifier returned by SUPR0ObjRegister().
1991	* @param pSession The session which is referencing the object.
1992	*/
1993	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
1994	{
1995	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1996	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1997	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
1998	int rc = VERR_INVALID_PARAMETER;
1999	PSUPDRVUSAGE pUsage;
2000	PSUPDRVUSAGE pUsagePrev;
2001
2002	/*
2003	* Validate the input.
2004	*/
2005	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2006	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2007	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2008	VERR_INVALID_PARAMETER);
2009
2010	/*
2011	* Acquire the spinlock and look for the usage record.
2012	*/
2013	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2014
2015	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
2016	pUsage;
2017	pUsagePrev = pUsage, pUsage = pUsage->pNext)
2018	{
2019	/Log2(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2020	if (pUsage->pObj == pObj)
2021	{
2022	rc = VINF_SUCCESS;
2023	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
2024	if (pUsage->cUsage > 1)
2025	{
2026	pObj->cUsage--;
2027	pUsage->cUsage--;
2028	}
2029	else
2030	{
2031	/*
2032	* Free the session record.
2033	*/
2034	if (pUsagePrev)
2035	pUsagePrev->pNext = pUsage->pNext;
2036	else
2037	pSession->pUsage = pUsage->pNext;
2038	pUsage->pNext = pDevExt->pUsageFree;
2039	pDevExt->pUsageFree = pUsage;
2040
2041	/* What about the object? */
2042	if (pObj->cUsage > 1)
2043	pObj->cUsage--;
2044	else
2045	{
2046	/*
2047	* Object is to be destroyed, unlink it.
2048	*/
2049	pObj->u32Magic = SUPDRVOBJ_MAGIC_DEAD;
2050	rc = VINF_OBJECT_DESTROYED;
2051	if (pDevExt->pObjs == pObj)
2052	pDevExt->pObjs = pObj->pNext;
2053	else
2054	{
2055	PSUPDRVOBJ pObjPrev;
2056	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
2057	if (pObjPrev->pNext == pObj)
2058	{
2059	pObjPrev->pNext = pObj->pNext;
2060	break;
2061	}
2062	Assert(pObjPrev);
2063	}
2064	}
2065	}
2066	break;
2067	}
2068	}
2069
2070	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2071
2072	/*
2073	* Call the destructor and free the object if required.
2074	*/
2075	if (rc == VINF_OBJECT_DESTROYED)
2076	{
2077	Log(("SUPR0ObjRelease: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
2078	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
2079	if (pObj->pfnDestructor)
2080	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
2081	RTMemFree(pObj);
2082	}
2083
2084	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
2085	return rc;
2086	}
2087
2088
2089	/**
2090	* Verifies that the current process can access the specified object.
2091	*
2092	* @returns The following IPRT status code:
2093	* @retval VINF_SUCCESS if access was granted.
2094	* @retval VERR_PERMISSION_DENIED if denied access.
2095	* @retval VERR_INVALID_PARAMETER if invalid parameter.
2096	*
2097	* @param pvObj The identifier returned by SUPR0ObjRegister().
2098	* @param pSession The session which wishes to access the object.
2099	* @param pszObjName Object string name. This is optional and depends on the object type.
2100	*
2101	* @remark The caller is responsible for making sure the object isn't removed while
2102	* we're inside this function. If uncertain about this, just call AddRef before calling us.
2103	*/
2104	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
2105	{
2106	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2107	int rc;
2108
2109	/*
2110	* Validate the input.
2111	*/
2112	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2113	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2114	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2115	VERR_INVALID_PARAMETER);
2116
2117	/*
2118	* Check access. (returns true if a decision has been made.)
2119	*/
2120	rc = VERR_INTERNAL_ERROR;
2121	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
2122	return rc;
2123
2124	/*
2125	* Default policy is to allow the user to access his own
2126	* stuff but nothing else.
2127	*/
2128	if (pObj->CreatorUid == pSession->Uid)
2129	return VINF_SUCCESS;
2130	return VERR_PERMISSION_DENIED;
2131	}
2132
2133
2134	/**
2135	* Lock pages.
2136	*
2137	* @returns IPRT status code.
2138	* @param pSession Session to which the locked memory should be associated.
2139	* @param pvR3 Start of the memory range to lock.
2140	* This must be page aligned.
2141	* @param cPages Number of pages to lock.
2142	* @param paPages Where to put the physical addresses of locked memory.
2143	*/
2144	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
2145	{
2146	int rc;
2147	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2148	const size_t cb = (size_t)cPages << PAGE_SHIFT;
2149	LogFlow(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
2150
2151	/*
2152	* Verify input.
2153	*/
2154	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2155	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
2156	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
2157	\|\| !pvR3)
2158	{
2159	Log(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
2160	return VERR_INVALID_PARAMETER;
2161	}
2162
2163	/*
2164	* Let IPRT do the job.
2165	*/
2166	Mem.eType = MEMREF_TYPE_LOCKED;
2167	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTMEM_PROT_READ \| RTMEM_PROT_WRITE, RTR0ProcHandleSelf());
2168	if (RT_SUCCESS(rc))
2169	{
2170	uint32_t iPage = cPages;
2171	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
2172	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
2173
2174	while (iPage-- > 0)
2175	{
2176	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2177	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
2178	{
2179	AssertMsgFailed(("iPage=%d\n", iPage));
2180	rc = VERR_INTERNAL_ERROR;
2181	break;
2182	}
2183	}
2184	if (RT_SUCCESS(rc))
2185	rc = supdrvMemAdd(&Mem, pSession);
2186	if (RT_FAILURE(rc))
2187	{
2188	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
2189	AssertRC(rc2);
2190	}
2191	}
2192
2193	return rc;
2194	}
2195
2196
2197	/**
2198	* Unlocks the memory pointed to by pv.
2199	*
2200	* @returns IPRT status code.
2201	* @param pSession Session to which the memory was locked.
2202	* @param pvR3 Memory to unlock.
2203	*/
2204	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2205	{
2206	LogFlow(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2207	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2208	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
2209	}
2210
2211
2212	/**
2213	* Allocates a chunk of page aligned memory with contiguous and fixed physical
2214	* backing.
2215	*
2216	* @returns IPRT status code.
2217	* @param pSession Session data.
2218	* @param cPages Number of pages to allocate.
2219	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
2220	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
2221	* @param pHCPhys Where to put the physical address of allocated memory.
2222	*/
2223	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
2224	{
2225	int rc;
2226	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2227	LogFlow(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
2228
2229	/*
2230	* Validate input.
2231	*/
2232	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2233	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
2234	{
2235	Log(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
2236	pSession, ppvR0, ppvR3, pHCPhys));
2237	return VERR_INVALID_PARAMETER;
2238
2239	}
2240	if (cPages < 1 \|\| cPages >= 256)
2241	{
2242	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2243	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2244	}
2245
2246	/*
2247	* Let IPRT do the job.
2248	*/
2249	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
2250	if (RT_SUCCESS(rc))
2251	{
2252	int rc2;
2253	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2254	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2255	if (RT_SUCCESS(rc))
2256	{
2257	Mem.eType = MEMREF_TYPE_CONT;
2258	rc = supdrvMemAdd(&Mem, pSession);
2259	if (!rc)
2260	{
2261	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2262	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2263	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
2264	return 0;
2265	}
2266
2267	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2268	AssertRC(rc2);
2269	}
2270	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2271	AssertRC(rc2);
2272	}
2273
2274	return rc;
2275	}
2276
2277
2278	/**
2279	* Frees memory allocated using SUPR0ContAlloc().
2280	*
2281	* @returns IPRT status code.
2282	* @param pSession The session to which the memory was allocated.
2283	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2284	*/
2285	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2286	{
2287	LogFlow(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2288	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2289	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
2290	}
2291
2292
2293	/**
2294	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
2295	*
2296	* The memory isn't zeroed.
2297	*
2298	* @returns IPRT status code.
2299	* @param pSession Session data.
2300	* @param cPages Number of pages to allocate.
2301	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
2302	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
2303	* @param paPages Where to put the physical addresses of allocated memory.
2304	*/
2305	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
2306	{
2307	unsigned iPage;
2308	int rc;
2309	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2310	LogFlow(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
2311
2312	/*
2313	* Validate input.
2314	*/
2315	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2316	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
2317	{
2318	Log(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
2319	pSession, ppvR3, ppvR0, paPages));
2320	return VERR_INVALID_PARAMETER;
2321
2322	}
2323	if (cPages < 1 \|\| cPages >= 256)
2324	{
2325	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2326	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2327	}
2328
2329	/*
2330	* Let IPRT do the work.
2331	*/
2332	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
2333	if (RT_SUCCESS(rc))
2334	{
2335	int rc2;
2336	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2337	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2338	if (RT_SUCCESS(rc))
2339	{
2340	Mem.eType = MEMREF_TYPE_LOW;
2341	rc = supdrvMemAdd(&Mem, pSession);
2342	if (!rc)
2343	{
2344	for (iPage = 0; iPage < cPages; iPage++)
2345	{
2346	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2347	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%RHp\n", paPages[iPage]));
2348	}
2349	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2350	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2351	return 0;
2352	}
2353
2354	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2355	AssertRC(rc2);
2356	}
2357
2358	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2359	AssertRC(rc2);
2360	}
2361
2362	return rc;
2363	}
2364
2365
2366	/**
2367	* Frees memory allocated using SUPR0LowAlloc().
2368	*
2369	* @returns IPRT status code.
2370	* @param pSession The session to which the memory was allocated.
2371	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2372	*/
2373	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2374	{
2375	LogFlow(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2376	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2377	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
2378	}
2379
2380
2381
2382	/**
2383	* Allocates a chunk of memory with both R0 and R3 mappings.
2384	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
2385	*
2386	* @returns IPRT status code.
2387	* @param pSession The session to associated the allocation with.
2388	* @param cb Number of bytes to allocate.
2389	* @param ppvR0 Where to store the address of the Ring-0 mapping.
2390	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2391	*/
2392	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
2393	{
2394	int rc;
2395	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2396	LogFlow(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
2397
2398	/*
2399	* Validate input.
2400	*/
2401	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2402	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
2403	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
2404	if (cb < 1 \|\| cb >= _4M)
2405	{
2406	Log(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
2407	return VERR_INVALID_PARAMETER;
2408	}
2409
2410	/*
2411	* Let IPRT do the work.
2412	*/
2413	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
2414	if (RT_SUCCESS(rc))
2415	{
2416	int rc2;
2417	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2418	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2419	if (RT_SUCCESS(rc))
2420	{
2421	Mem.eType = MEMREF_TYPE_MEM;
2422	rc = supdrvMemAdd(&Mem, pSession);
2423	if (!rc)
2424	{
2425	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2426	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2427	return VINF_SUCCESS;
2428	}
2429
2430	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2431	AssertRC(rc2);
2432	}
2433
2434	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2435	AssertRC(rc2);
2436	}
2437
2438	return rc;
2439	}
2440
2441
2442	/**
2443	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
2444	*
2445	* @returns IPRT status code.
2446	* @param pSession The session to which the memory was allocated.
2447	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2448	* @param paPages Where to store the physical addresses.
2449	*/
2450	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
2451	{
2452	PSUPDRVBUNDLE pBundle;
2453	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2454	LogFlow(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
2455
2456	/*
2457	* Validate input.
2458	*/
2459	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2460	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
2461	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
2462
2463	/*
2464	* Search for the address.
2465	*/
2466	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2467	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2468	{
2469	if (pBundle->cUsed > 0)
2470	{
2471	unsigned i;
2472	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2473	{
2474	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
2475	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2476	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2477	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2478	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
2479	)
2480	)
2481	{
2482	const size_t cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2483	size_t iPage;
2484	for (iPage = 0; iPage < cPages; iPage++)
2485	{
2486	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2487	paPages[iPage].uReserved = 0;
2488	}
2489	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2490	return VINF_SUCCESS;
2491	}
2492	}
2493	}
2494	}
2495	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2496	Log(("Failed to find %p!!!\n", (void *)uPtr));
2497	return VERR_INVALID_PARAMETER;
2498	}
2499
2500
2501	/**
2502	* Free memory allocated by SUPR0MemAlloc().
2503	*
2504	* @returns IPRT status code.
2505	* @param pSession The session owning the allocation.
2506	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2507	*/
2508	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2509	{
2510	LogFlow(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2511	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2512	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
2513	}
2514
2515
2516	/**
2517	* Allocates a chunk of memory with a kernel or/and a user mode mapping.
2518	*
2519	* The memory is fixed and it's possible to query the physical addresses using
2520	* SUPR0MemGetPhys().
2521	*
2522	* @returns IPRT status code.
2523	* @param pSession The session to associated the allocation with.
2524	* @param cPages The number of pages to allocate.
2525	* @param fFlags Flags, reserved for the future. Must be zero.
2526	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2527	* NULL if no ring-3 mapping.
2528	* @param ppvR3 Where to store the address of the Ring-0 mapping.
2529	* NULL if no ring-0 mapping.
2530	* @param paPages Where to store the addresses of the pages. Optional.
2531	*/
2532	SUPR0DECL(int) SUPR0PageAllocEx(PSUPDRVSESSION pSession, uint32_t cPages, uint32_t fFlags, PRTR3PTR ppvR3, PRTR0PTR ppvR0, PRTHCPHYS paPages)
2533	{
2534	int rc;
2535	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2536	LogFlow(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
2537
2538	/*
2539	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2540	*/
2541	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2542	AssertPtrNullReturn(ppvR3, VERR_INVALID_POINTER);
2543	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2544	AssertReturn(ppvR3 \|\| ppvR0, VERR_INVALID_PARAMETER);
2545	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2546	if (cPages < 1 \|\| cPages > VBOX_MAX_ALLOC_PAGE_COUNT)
2547	{
2548	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)));
2549	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2550	}
2551
2552	/*
2553	* Let IPRT do the work.
2554	*/
2555	if (ppvR0)
2556	rc = RTR0MemObjAllocPage(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, true /* fExecutable */);
2557	else
2558	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
2559	if (RT_SUCCESS(rc))
2560	{
2561	int rc2;
2562	if (ppvR3)
2563	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2564	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2565	else
2566	Mem.MapObjR3 = NIL_RTR0MEMOBJ;
2567	if (RT_SUCCESS(rc))
2568	{
2569	Mem.eType = MEMREF_TYPE_PAGE;
2570	rc = supdrvMemAdd(&Mem, pSession);
2571	if (!rc)
2572	{
2573	if (ppvR3)
2574	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2575	if (ppvR0)
2576	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2577	if (paPages)
2578	{
2579	uint32_t iPage = cPages;
2580	while (iPage-- > 0)
2581	{
2582	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
2583	Assert(paPages[iPage] != NIL_RTHCPHYS);
2584	}
2585	}
2586	return VINF_SUCCESS;
2587	}
2588
2589	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2590	AssertRC(rc2);
2591	}
2592
2593	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2594	AssertRC(rc2);
2595	}
2596	return rc;
2597	}
2598
2599
2600	/**
2601	* Maps a chunk of memory previously allocated by SUPR0PageAllocEx into kernel
2602	* space.
2603	*
2604	* @returns IPRT status code.
2605	* @param pSession The session to associated the allocation with.
2606	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
2607	* @param offSub Where to start mapping. Must be page aligned.
2608	* @param cbSub How much to map. Must be page aligned.
2609	* @param fFlags Flags, MBZ.
2610	* @param ppvR0 Where to reutrn the address of the ring-0 mapping on
2611	* success.
2612	*/
2613	SUPR0DECL(int) SUPR0PageMapKernel(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t offSub, uint32_t cbSub,
2614	uint32_t fFlags, PRTR0PTR ppvR0)
2615	{
2616	int rc;
2617	PSUPDRVBUNDLE pBundle;
2618	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2619	RTR0MEMOBJ hMemObj = NIL_RTR0MEMOBJ;
2620	LogFlow(("SUPR0PageMapKernel: pSession=%p pvR3=%p offSub=%#x cbSub=%#x\n", pSession, pvR3, offSub, cbSub));
2621
2622	/*
2623	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2624	*/
2625	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2626	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2627	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2628	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2629	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2630	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
2631
2632	/*
2633	* Find the memory object.
2634	*/
2635	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2636	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2637	{
2638	if (pBundle->cUsed > 0)
2639	{
2640	unsigned i;
2641	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2642	{
2643	if ( ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
2644	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2645	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2646	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
2647	\|\| ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED
2648	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2649	&& pBundle->aMem[i].MapObjR3 == NIL_RTR0MEMOBJ
2650	&& RTR0MemObjAddressR3(pBundle->aMem[i].MemObj) == pvR3))
2651	{
2652	hMemObj = pBundle->aMem[i].MemObj;
2653	break;
2654	}
2655	}
2656	}
2657	}
2658	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2659
2660	rc = VERR_INVALID_PARAMETER;
2661	if (hMemObj != NIL_RTR0MEMOBJ)
2662	{
2663	/*
2664	* Do some furter input validations before calling IPRT.
2665	* (Cleanup is done indirectly by telling RTR0MemObjFree to include mappings.)
2666	*/
2667	size_t cbMemObj = RTR0MemObjSize(hMemObj);
2668	if ( offSub < cbMemObj
2669	&& cbSub <= cbMemObj
2670	&& offSub + cbSub <= cbMemObj)
2671	{
2672	RTR0MEMOBJ hMapObj;
2673	rc = RTR0MemObjMapKernelEx(&hMapObj, hMemObj, (void *)-1, 0,
2674	RTMEM_PROT_READ \| RTMEM_PROT_WRITE, offSub, cbSub);
2675	if (RT_SUCCESS(rc))
2676	*ppvR0 = RTR0MemObjAddress(hMapObj);
2677	}
2678	else
2679	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
2680
2681	}
2682	return rc;
2683	}
2684
2685
2686	/**
2687	* Changes the page level protection of one or more pages previously allocated
2688	* by SUPR0PageAllocEx.
2689	*
2690	* @returns IPRT status code.
2691	* @param pSession The session to associated the allocation with.
2692	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
2693	* NIL_RTR3PTR if the ring-3 mapping should be unaffected.
2694	* @param pvR0 The ring-0 address returned by SUPR0PageAllocEx.
2695	* NIL_RTR0PTR if the ring-0 mapping should be unaffected.
2696	* @param offSub Where to start changing. Must be page aligned.
2697	* @param cbSub How much to change. Must be page aligned.
2698	* @param fProt The new page level protection, see RTMEM_PROT_*.
2699	*/
2700	SUPR0DECL(int) SUPR0PageProtect(PSUPDRVSESSION pSession, RTR3PTR pvR3, RTR0PTR pvR0, uint32_t offSub, uint32_t cbSub, uint32_t fProt)
2701	{
2702	int rc;
2703	PSUPDRVBUNDLE pBundle;
2704	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2705	RTR0MEMOBJ hMemObjR0 = NIL_RTR0MEMOBJ;
2706	RTR0MEMOBJ hMemObjR3 = NIL_RTR0MEMOBJ;
2707	LogFlow(("SUPR0PageProtect: pSession=%p pvR3=%p pvR0=%p offSub=%#x cbSub=%#x fProt-%#x\n", pSession, pvR3, pvR0, offSub, cbSub, fProt));
2708
2709	/*
2710	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2711	*/
2712	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2713	AssertReturn(!(fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)), VERR_INVALID_PARAMETER);
2714	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2715	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2716	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
2717
2718	/*
2719	* Find the memory object.
2720	*/
2721	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2722	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2723	{
2724	if (pBundle->cUsed > 0)
2725	{
2726	unsigned i;
2727	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2728	{
2729	if ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
2730	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2731	&& ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2732	\|\| pvR3 == NIL_RTR3PTR)
2733	&& ( pvR0 == NIL_RTR0PTR
2734	\|\| RTR0MemObjAddress(pBundle->aMem[i].MemObj) == pvR0)
2735	&& ( pvR3 == NIL_RTR3PTR
2736	\|\| RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3))
2737	{
2738	if (pvR0 != NIL_RTR0PTR)
2739	hMemObjR0 = pBundle->aMem[i].MemObj;
2740	if (pvR3 != NIL_RTR3PTR)
2741	hMemObjR3 = pBundle->aMem[i].MapObjR3;
2742	break;
2743	}
2744	}
2745	}
2746	}
2747	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2748
2749	rc = VERR_INVALID_PARAMETER;
2750	if ( hMemObjR0 != NIL_RTR0MEMOBJ
2751	\|\| hMemObjR3 != NIL_RTR0MEMOBJ)
2752	{
2753	/*
2754	* Do some furter input validations before calling IPRT.
2755	*/
2756	size_t cbMemObj = hMemObjR0 != NIL_RTR0PTR ? RTR0MemObjSize(hMemObjR0) : RTR0MemObjSize(hMemObjR3);
2757	if ( offSub < cbMemObj
2758	&& cbSub <= cbMemObj
2759	&& offSub + cbSub <= cbMemObj)
2760	{
2761	rc = VINF_SUCCESS;
2762	if (hMemObjR3 != NIL_RTR0PTR)
2763	rc = RTR0MemObjProtect(hMemObjR3, offSub, cbSub, fProt);
2764	if (hMemObjR0 != NIL_RTR0PTR && RT_SUCCESS(rc))
2765	rc = RTR0MemObjProtect(hMemObjR0, offSub, cbSub, fProt);
2766	}
2767	else
2768	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
2769
2770	}
2771	return rc;
2772
2773	}
2774
2775
2776	/**
2777	* Free memory allocated by SUPR0PageAlloc() and SUPR0PageAllocEx().
2778	*
2779	* @returns IPRT status code.
2780	* @param pSession The session owning the allocation.
2781	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc() or
2782	* SUPR0PageAllocEx().
2783	*/
2784	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2785	{
2786	LogFlow(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2787	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2788	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_PAGE);
2789	}
2790
2791
2792	/**
2793	* Gets the paging mode of the current CPU.
2794	*
2795	* @returns Paging mode, SUPPAGEINGMODE_INVALID on error.
2796	*/
2797	SUPR0DECL(SUPPAGINGMODE) SUPR0GetPagingMode(void)
2798	{
2799	SUPPAGINGMODE enmMode;
2800
2801	RTR0UINTREG cr0 = ASMGetCR0();
2802	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
2803	enmMode = SUPPAGINGMODE_INVALID;
2804	else
2805	{
2806	RTR0UINTREG cr4 = ASMGetCR4();
2807	uint32_t fNXEPlusLMA = 0;
2808	if (cr4 & X86_CR4_PAE)
2809	{
2810	uint32_t fAmdFeatures = ASMCpuId_EDX(0x80000001);
2811	if (fAmdFeatures & (X86_CPUID_AMD_FEATURE_EDX_NX \| X86_CPUID_AMD_FEATURE_EDX_LONG_MODE))
2812	{
2813	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
2814	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
2815	fNXEPlusLMA \|= RT_BIT(0);
2816	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
2817	fNXEPlusLMA \|= RT_BIT(1);
2818	}
2819	}
2820
2821	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
2822	{
2823	case 0:
2824	enmMode = SUPPAGINGMODE_32_BIT;
2825	break;
2826
2827	case X86_CR4_PGE:
2828	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
2829	break;
2830
2831	case X86_CR4_PAE:
2832	enmMode = SUPPAGINGMODE_PAE;
2833	break;
2834
2835	case X86_CR4_PAE \| RT_BIT(0):
2836	enmMode = SUPPAGINGMODE_PAE_NX;
2837	break;
2838
2839	case X86_CR4_PAE \| X86_CR4_PGE:
2840	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
2841	break;
2842
2843	case X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
2844	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
2845	break;
2846
2847	case RT_BIT(1) \| X86_CR4_PAE:
2848	enmMode = SUPPAGINGMODE_AMD64;
2849	break;
2850
2851	case RT_BIT(1) \| X86_CR4_PAE \| RT_BIT(0):
2852	enmMode = SUPPAGINGMODE_AMD64_NX;
2853	break;
2854
2855	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
2856	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
2857	break;
2858
2859	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
2860	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
2861	break;
2862
2863	default:
2864	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
2865	enmMode = SUPPAGINGMODE_INVALID;
2866	break;
2867	}
2868	}
2869	return enmMode;
2870	}
2871
2872
2873	/**
2874	* Enables or disabled hardware virtualization extensions using native OS APIs.
2875	*
2876	* @returns VBox status code.
2877	* @retval VINF_SUCCESS on success.
2878	* @retval VERR_NOT_SUPPORTED if not supported by the native OS.
2879	*
2880	* @param fEnable Whether to enable or disable.
2881	*/
2882	SUPR0DECL(int) SUPR0EnableVTx(bool fEnable)
2883	{
2884	#ifdef RT_OS_DARWIN
2885	return supdrvOSEnableVTx(fEnable);
2886	#else
2887	return VERR_NOT_SUPPORTED;
2888	#endif
2889	}
2890
2891
2892	/** @todo document me */
2893	SUPR0DECL(int) SUPR0QueryVTCaps(PSUPDRVSESSION pSession, uint32_t *pfCaps)
2894	{
2895	/*
2896	* Input validation.
2897	*/
2898	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2899	AssertPtrReturn(pfCaps, VERR_INVALID_POINTER);
2900
2901	*pfCaps = 0;
2902
2903	if (ASMHasCpuId())
2904	{
2905	uint32_t u32FeaturesECX;
2906	uint32_t u32Dummy;
2907	uint32_t u32FeaturesEDX;
2908	uint32_t u32VendorEBX, u32VendorECX, u32VendorEDX, u32AMDFeatureEDX, u32AMDFeatureECX;
2909	uint64_t val;
2910
2911	ASMCpuId(0, &u32Dummy, &u32VendorEBX, &u32VendorECX, &u32VendorEDX);
2912	ASMCpuId(1, &u32Dummy, &u32Dummy, &u32FeaturesECX, &u32FeaturesEDX);
2913	/* Query AMD features. */
2914	ASMCpuId(0x80000001, &u32Dummy, &u32Dummy, &u32AMDFeatureECX, &u32AMDFeatureEDX);
2915
2916	if ( u32VendorEBX == X86_CPUID_VENDOR_INTEL_EBX
2917	&& u32VendorECX == X86_CPUID_VENDOR_INTEL_ECX
2918	&& u32VendorEDX == X86_CPUID_VENDOR_INTEL_EDX
2919	)
2920	{
2921	if ( (u32FeaturesECX & X86_CPUID_FEATURE_ECX_VMX)
2922	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
2923	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
2924	)
2925	{
2926	val = ASMRdMsr(MSR_IA32_FEATURE_CONTROL);
2927	/*
2928	* Both the LOCK and VMXON bit must be set; otherwise VMXON will generate a #GP.
2929	* Once the lock bit is set, this MSR can no longer be modified.
2930	*/
2931	if ( (val & (MSR_IA32_FEATURE_CONTROL_VMXON\|MSR_IA32_FEATURE_CONTROL_LOCK))
2932	== (MSR_IA32_FEATURE_CONTROL_VMXON\|MSR_IA32_FEATURE_CONTROL_LOCK) /* enabled and locked */
2933	\|\| !(val & MSR_IA32_FEATURE_CONTROL_LOCK) /* not enabled, but not locked either */
2934	)
2935	{
2936	VMX_CAPABILITY vtCaps;
2937
2938	*pfCaps \|= SUPVTCAPS_VT_X;
2939
2940	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS);
2941	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC_USE_SECONDARY_EXEC_CTRL)
2942	{
2943	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS2);
2944	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC2_EPT)
2945	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
2946	}
2947	return VINF_SUCCESS;
2948	}
2949	return VERR_VMX_MSR_LOCKED_OR_DISABLED;
2950	}
2951	return VERR_VMX_NO_VMX;
2952	}
2953
2954	if ( u32VendorEBX == X86_CPUID_VENDOR_AMD_EBX
2955	&& u32VendorECX == X86_CPUID_VENDOR_AMD_ECX
2956	&& u32VendorEDX == X86_CPUID_VENDOR_AMD_EDX
2957	)
2958	{
2959	if ( (u32AMDFeatureECX & X86_CPUID_AMD_FEATURE_ECX_SVM)
2960	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
2961	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
2962	)
2963	{
2964	/* Check if SVM is disabled */
2965	val = ASMRdMsr(MSR_K8_VM_CR);
2966	if (!(val & MSR_K8_VM_CR_SVM_DISABLE))
2967	{
2968	*pfCaps \|= SUPVTCAPS_AMD_V;
2969
2970	/* Query AMD features. */
2971	ASMCpuId(0x8000000A, &u32Dummy, &u32Dummy, &u32Dummy, &u32FeaturesEDX);
2972
2973	if (u32FeaturesEDX & AMD_CPUID_SVM_FEATURE_EDX_NESTED_PAGING)
2974	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
2975
2976	return VINF_SUCCESS;
2977	}
2978	return VERR_SVM_DISABLED;
2979	}
2980	return VERR_SVM_NO_SVM;
2981	}
2982	}
2983
2984	return VERR_UNSUPPORTED_CPU;
2985	}
2986
2987
2988	/**
2989	* (Re-)initializes the per-cpu structure prior to starting or resuming the GIP
2990	* updating.
2991	*
2992	* @param pGipCpu The per CPU structure for this CPU.
2993	* @param u64NanoTS The current time.
2994	*/
2995	static void supdrvGipReInitCpu(PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
2996	{
2997	pGipCpu->u64TSC = ASMReadTSC() - pGipCpu->u32UpdateIntervalTSC;
2998	pGipCpu->u64NanoTS = u64NanoTS;
2999	}
3000
3001
3002	/**
3003	* Set the current TSC and NanoTS value for the CPU.
3004	*
3005	* @param idCpu The CPU ID. Unused - we have to use the APIC ID.
3006	* @param pvUser1 Pointer to the ring-0 GIP mapping.
3007	* @param pvUser2 Pointer to the variable holding the current time.
3008	*/
3009	static DECLCALLBACK(void) supdrvGipReInitCpuCallback(RTCPUID idCpu, void pvUser1, void pvUser2)
3010	{
3011	PSUPGLOBALINFOPAGE pGip = (PSUPGLOBALINFOPAGE)pvUser1;
3012	unsigned iCpu = ASMGetApicId();
3013
3014	if (RT_LIKELY(iCpu < RT_ELEMENTS(pGip->aCPUs)))
3015	supdrvGipReInitCpu(&pGip->aCPUs[iCpu], (uint64_t )pvUser2);
3016
3017	NOREF(pvUser2);
3018	NOREF(idCpu);
3019	}
3020
3021
3022	/**
3023	* Maps the GIP into userspace and/or get the physical address of the GIP.
3024	*
3025	* @returns IPRT status code.
3026	* @param pSession Session to which the GIP mapping should belong.
3027	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
3028	* @param pHCPhysGip Where to store the physical address. (optional)
3029	*
3030	* @remark There is no reference counting on the mapping, so one call to this function
3031	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
3032	* and remove the session as a GIP user.
3033	*/
3034	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
3035	{
3036	int rc;
3037	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3038	RTR3PTR pGipR3 = NIL_RTR3PTR;
3039	RTHCPHYS HCPhys = NIL_RTHCPHYS;
3040	LogFlow(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
3041
3042	/*
3043	* Validate
3044	*/
3045	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3046	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
3047	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
3048
3049	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3050	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3051	#else
3052	RTSemFastMutexRequest(pDevExt->mtxGip);
3053	#endif
3054	if (pDevExt->pGip)
3055	{
3056	/*
3057	* Map it?
3058	*/
3059	rc = VINF_SUCCESS;
3060	if (ppGipR3)
3061	{
3062	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
3063	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
3064	RTMEM_PROT_READ, RTR0ProcHandleSelf());
3065	if (RT_SUCCESS(rc))
3066	pGipR3 = RTR0MemObjAddressR3(pSession->GipMapObjR3);
3067	}
3068
3069	/*
3070	* Get physical address.
3071	*/
3072	if (pHCPhysGip && RT_SUCCESS(rc))
3073	HCPhys = pDevExt->HCPhysGip;
3074
3075	/*
3076	* Reference globally.
3077	*/
3078	if (!pSession->fGipReferenced && RT_SUCCESS(rc))
3079	{
3080	pSession->fGipReferenced = 1;
3081	pDevExt->cGipUsers++;
3082	if (pDevExt->cGipUsers == 1)
3083	{
3084	PSUPGLOBALINFOPAGE pGipR0 = pDevExt->pGip;
3085	uint64_t u64NanoTS;
3086	unsigned i;
3087
3088	LogFlow(("SUPR0GipMap: Resumes GIP updating\n"));
3089
3090	if (pGipR0->aCPUs[0].u32TransactionId != 2 /* not the first time */)
3091	{
3092	for (i = 0; i < RT_ELEMENTS(pGipR0->aCPUs); i++)
3093	ASMAtomicUoWriteU32(&pGipR0->aCPUs[i].u32TransactionId,
3094	(pGipR0->aCPUs[i].u32TransactionId + GIP_UPDATEHZ_RECALC_FREQ * 2)
3095	& ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
3096	ASMAtomicWriteU64(&pGipR0->u64NanoTSLastUpdateHz, 0);
3097	}
3098
3099	u64NanoTS = RTTimeSystemNanoTS() - pGipR0->u32UpdateIntervalNS;
3100	if ( pGipR0->u32Mode == SUPGIPMODE_SYNC_TSC
3101	\|\| RTMpGetOnlineCount() == 1)
3102	supdrvGipReInitCpu(&pGipR0->aCPUs[0], u64NanoTS);
3103	else
3104	RTMpOnAll(supdrvGipReInitCpuCallback, pGipR0, &u64NanoTS);
3105
3106	rc = RTTimerStart(pDevExt->pGipTimer, 0);
3107	AssertRC(rc); rc = VINF_SUCCESS;
3108	}
3109	}
3110	}
3111	else
3112	{
3113	rc = VERR_GENERAL_FAILURE;
3114	Log(("SUPR0GipMap: GIP is not available!\n"));
3115	}
3116	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3117	RTSemMutexRelease(pDevExt->mtxGip);
3118	#else
3119	RTSemFastMutexRelease(pDevExt->mtxGip);
3120	#endif
3121
3122	/*
3123	* Write returns.
3124	*/
3125	if (pHCPhysGip)
3126	*pHCPhysGip = HCPhys;
3127	if (ppGipR3)
3128	*ppGipR3 = pGipR3;
3129
3130	#ifdef DEBUG_DARWIN_GIP
3131	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3132	#else
3133	LogFlow(( "SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3134	#endif
3135	return rc;
3136	}
3137
3138
3139	/**
3140	* Unmaps any user mapping of the GIP and terminates all GIP access
3141	* from this session.
3142	*
3143	* @returns IPRT status code.
3144	* @param pSession Session to which the GIP mapping should belong.
3145	*/
3146	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
3147	{
3148	int rc = VINF_SUCCESS;
3149	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3150	#ifdef DEBUG_DARWIN_GIP
3151	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
3152	pSession,
3153	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
3154	pSession->GipMapObjR3));
3155	#else
3156	LogFlow(("SUPR0GipUnmap: pSession=%p\n", pSession));
3157	#endif
3158	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3159
3160	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3161	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3162	#else
3163	RTSemFastMutexRequest(pDevExt->mtxGip);
3164	#endif
3165
3166	/*
3167	* Unmap anything?
3168	*/
3169	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
3170	{
3171	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
3172	AssertRC(rc);
3173	if (RT_SUCCESS(rc))
3174	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
3175	}
3176
3177	/*
3178	* Dereference global GIP.
3179	*/
3180	if (pSession->fGipReferenced && !rc)
3181	{
3182	pSession->fGipReferenced = 0;
3183	if ( pDevExt->cGipUsers > 0
3184	&& !--pDevExt->cGipUsers)
3185	{
3186	LogFlow(("SUPR0GipUnmap: Suspends GIP updating\n"));
3187	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = VINF_SUCCESS;
3188	}
3189	}
3190
3191	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3192	RTSemMutexRelease(pDevExt->mtxGip);
3193	#else
3194	RTSemFastMutexRelease(pDevExt->mtxGip);
3195	#endif
3196
3197	return rc;
3198	}
3199
3200
3201	/**
3202	* Gets the GIP pointer.
3203	*
3204	* @returns Pointer to the GIP or NULL.
3205	*/
3206	SUPDECL(PSUPGLOBALINFOPAGE) SUPGetGIP(void)
3207	{
3208	return g_pSUPGlobalInfoPage;
3209	}
3210
3211
3212	/**
3213	* Register a component factory with the support driver.
3214	*
3215	* This is currently restricted to kernel sessions only.
3216	*
3217	* @returns VBox status code.
3218	* @retval VINF_SUCCESS on success.
3219	* @retval VERR_NO_MEMORY if we're out of memory.
3220	* @retval VERR_ALREADY_EXISTS if the factory has already been registered.
3221	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3222	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3223	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3224	*
3225	* @param pSession The SUPDRV session (must be a ring-0 session).
3226	* @param pFactory Pointer to the component factory registration structure.
3227	*
3228	* @remarks This interface is also available via SUPR0IdcComponentRegisterFactory.
3229	*/
3230	SUPR0DECL(int) SUPR0ComponentRegisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3231	{
3232	PSUPDRVFACTORYREG pNewReg;
3233	const char *psz;
3234	int rc;
3235
3236	/*
3237	* Validate parameters.
3238	*/
3239	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3240	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3241	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3242	AssertPtrReturn(pFactory->pfnQueryFactoryInterface, VERR_INVALID_POINTER);
3243	psz = (const char *)memchr(pFactory->szName, '\0', sizeof(pFactory->szName));
3244	AssertReturn(psz, VERR_INVALID_PARAMETER);
3245
3246	/*
3247	* Allocate and initialize a new registration structure.
3248	*/
3249	pNewReg = (PSUPDRVFACTORYREG)RTMemAlloc(sizeof(SUPDRVFACTORYREG));
3250	if (pNewReg)
3251	{
3252	pNewReg->pNext = NULL;
3253	pNewReg->pFactory = pFactory;
3254	pNewReg->pSession = pSession;
3255	pNewReg->cchName = psz - &pFactory->szName[0];
3256
3257	/*
3258	* Add it to the tail of the list after checking for prior registration.
3259	*/
3260	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3261	if (RT_SUCCESS(rc))
3262	{
3263	PSUPDRVFACTORYREG pPrev = NULL;
3264	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3265	while (pCur && pCur->pFactory != pFactory)
3266	{
3267	pPrev = pCur;
3268	pCur = pCur->pNext;
3269	}
3270	if (!pCur)
3271	{
3272	if (pPrev)
3273	pPrev->pNext = pNewReg;
3274	else
3275	pSession->pDevExt->pComponentFactoryHead = pNewReg;
3276	rc = VINF_SUCCESS;
3277	}
3278	else
3279	rc = VERR_ALREADY_EXISTS;
3280
3281	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3282	}
3283
3284	if (RT_FAILURE(rc))
3285	RTMemFree(pNewReg);
3286	}
3287	else
3288	rc = VERR_NO_MEMORY;
3289	return rc;
3290	}
3291
3292
3293	/**
3294	* Deregister a component factory.
3295	*
3296	* @returns VBox status code.
3297	* @retval VINF_SUCCESS on success.
3298	* @retval VERR_NOT_FOUND if the factory wasn't registered.
3299	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3300	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3301	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3302	*
3303	* @param pSession The SUPDRV session (must be a ring-0 session).
3304	* @param pFactory Pointer to the component factory registration structure
3305	* previously passed SUPR0ComponentRegisterFactory().
3306	*
3307	* @remarks This interface is also available via SUPR0IdcComponentDeregisterFactory.
3308	*/
3309	SUPR0DECL(int) SUPR0ComponentDeregisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3310	{
3311	int rc;
3312
3313	/*
3314	* Validate parameters.
3315	*/
3316	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3317	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3318	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3319
3320	/*
3321	* Take the lock and look for the registration record.
3322	*/
3323	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3324	if (RT_SUCCESS(rc))
3325	{
3326	PSUPDRVFACTORYREG pPrev = NULL;
3327	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3328	while (pCur && pCur->pFactory != pFactory)
3329	{
3330	pPrev = pCur;
3331	pCur = pCur->pNext;
3332	}
3333	if (pCur)
3334	{
3335	if (!pPrev)
3336	pSession->pDevExt->pComponentFactoryHead = pCur->pNext;
3337	else
3338	pPrev->pNext = pCur->pNext;
3339
3340	pCur->pNext = NULL;
3341	pCur->pFactory = NULL;
3342	pCur->pSession = NULL;
3343	rc = VINF_SUCCESS;
3344	}
3345	else
3346	rc = VERR_NOT_FOUND;
3347
3348	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3349
3350	RTMemFree(pCur);
3351	}
3352	return rc;
3353	}
3354
3355
3356	/**
3357	* Queries a component factory.
3358	*
3359	* @returns VBox status code.
3360	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3361	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3362	* @retval VERR_SUPDRV_COMPONENT_NOT_FOUND if the component factory wasn't found.
3363	* @retval VERR_SUPDRV_INTERFACE_NOT_SUPPORTED if the interface wasn't supported.
3364	*
3365	* @param pSession The SUPDRV session.
3366	* @param pszName The name of the component factory.
3367	* @param pszInterfaceUuid The UUID of the factory interface (stringified).
3368	* @param ppvFactoryIf Where to store the factory interface.
3369	*/
3370	SUPR0DECL(int) SUPR0ComponentQueryFactory(PSUPDRVSESSION pSession, const char pszName, const char pszInterfaceUuid, void **ppvFactoryIf)
3371	{
3372	const char *pszEnd;
3373	size_t cchName;
3374	int rc;
3375
3376	/*
3377	* Validate parameters.
3378	*/
3379	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3380
3381	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
3382	pszEnd = memchr(pszName, '\0', RT_SIZEOFMEMB(SUPDRVFACTORY, szName));
3383	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3384	cchName = pszEnd - pszName;
3385
3386	AssertPtrReturn(pszInterfaceUuid, VERR_INVALID_POINTER);
3387	pszEnd = memchr(pszInterfaceUuid, '\0', RTUUID_STR_LENGTH);
3388	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3389
3390	AssertPtrReturn(ppvFactoryIf, VERR_INVALID_POINTER);
3391	*ppvFactoryIf = NULL;
3392
3393	/*
3394	* Take the lock and try all factories by this name.
3395	*/
3396	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3397	if (RT_SUCCESS(rc))
3398	{
3399	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3400	rc = VERR_SUPDRV_COMPONENT_NOT_FOUND;
3401	while (pCur)
3402	{
3403	if ( pCur->cchName == cchName
3404	&& !memcmp(pCur->pFactory->szName, pszName, cchName))
3405	{
3406	void *pvFactory = pCur->pFactory->pfnQueryFactoryInterface(pCur->pFactory, pSession, pszInterfaceUuid);
3407	if (pvFactory)
3408	{
3409	*ppvFactoryIf = pvFactory;
3410	rc = VINF_SUCCESS;
3411	break;
3412	}
3413	rc = VERR_SUPDRV_INTERFACE_NOT_SUPPORTED;
3414	}
3415
3416	/* next */
3417	pCur = pCur->pNext;
3418	}
3419
3420	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3421	}
3422	return rc;
3423	}
3424
3425
3426	/**
3427	* Adds a memory object to the session.
3428	*
3429	* @returns IPRT status code.
3430	* @param pMem Memory tracking structure containing the
3431	* information to track.
3432	* @param pSession The session.
3433	*/
3434	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
3435	{
3436	PSUPDRVBUNDLE pBundle;
3437	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3438
3439	/*
3440	* Find free entry and record the allocation.
3441	*/
3442	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3443	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3444	{
3445	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
3446	{
3447	unsigned i;
3448	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3449	{
3450	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
3451	{
3452	pBundle->cUsed++;
3453	pBundle->aMem[i] = *pMem;
3454	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3455	return VINF_SUCCESS;
3456	}
3457	}
3458	AssertFailed(); /* !!this can't be happening!!! */
3459	}
3460	}
3461	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3462
3463	/*
3464	* Need to allocate a new bundle.
3465	* Insert into the last entry in the bundle.
3466	*/
3467	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
3468	if (!pBundle)
3469	return VERR_NO_MEMORY;
3470
3471	/* take last entry. */
3472	pBundle->cUsed++;
3473	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
3474
3475	/* insert into list. */
3476	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3477	pBundle->pNext = pSession->Bundle.pNext;
3478	pSession->Bundle.pNext = pBundle;
3479	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3480
3481	return VINF_SUCCESS;
3482	}
3483
3484
3485	/**
3486	* Releases a memory object referenced by pointer and type.
3487	*
3488	* @returns IPRT status code.
3489	* @param pSession Session data.
3490	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
3491	* @param eType Memory type.
3492	*/
3493	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
3494	{
3495	PSUPDRVBUNDLE pBundle;
3496	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3497
3498	/*
3499	* Validate input.
3500	*/
3501	if (!uPtr)
3502	{
3503	Log(("Illegal address %p\n", (void *)uPtr));
3504	return VERR_INVALID_PARAMETER;
3505	}
3506
3507	/*
3508	* Search for the address.
3509	*/
3510	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3511	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3512	{
3513	if (pBundle->cUsed > 0)
3514	{
3515	unsigned i;
3516	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3517	{
3518	if ( pBundle->aMem[i].eType == eType
3519	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3520	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
3521	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3522	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
3523	)
3524	{
3525	/* Make a copy of it and release it outside the spinlock. */
3526	SUPDRVMEMREF Mem = pBundle->aMem[i];
3527	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
3528	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
3529	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
3530	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3531
3532	if (Mem.MapObjR3 != NIL_RTR0MEMOBJ)
3533	{
3534	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
3535	AssertRC(rc); /** @todo figure out how to handle this. */
3536	}
3537	if (Mem.MemObj != NIL_RTR0MEMOBJ)
3538	{
3539	int rc = RTR0MemObjFree(Mem.MemObj, true /* fFreeMappings */);
3540	AssertRC(rc); /** @todo figure out how to handle this. */
3541	}
3542	return VINF_SUCCESS;
3543	}
3544	}
3545	}
3546	}
3547	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3548	Log(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
3549	return VERR_INVALID_PARAMETER;
3550	}
3551
3552
3553	/**
3554	* Opens an image. If it's the first time it's opened the call must upload
3555	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
3556	*
3557	* This is the 1st step of the loading.
3558	*
3559	* @returns IPRT status code.
3560	* @param pDevExt Device globals.
3561	* @param pSession Session data.
3562	* @param pReq The open request.
3563	*/
3564	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
3565	{
3566	int rc;
3567	PSUPDRVLDRIMAGE pImage;
3568	void *pv;
3569	size_t cchName = strlen(pReq->u.In.szName); /* (caller checked < 32). */
3570	LogFlow(("supdrvIOCtl_LdrOpen: szName=%s cbImageWithTabs=%d\n", pReq->u.In.szName, pReq->u.In.cbImageWithTabs));
3571
3572	/*
3573	* Check if we got an instance of the image already.
3574	*/
3575	supdrvLdrLock(pDevExt);
3576	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
3577	{
3578	if ( pImage->szName[cchName] == '\0'
3579	&& !memcmp(pImage->szName, pReq->u.In.szName, cchName))
3580	{
3581	/** @todo check cbImageBits and cbImageWithTabs here, if they differs that indicates that the images are different. */
3582	pImage->cUsage++;
3583	pReq->u.Out.pvImageBase = pImage->pvImage;
3584	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
3585	pReq->u.Out.fNativeLoader = pImage->fNative;
3586	supdrvLdrAddUsage(pSession, pImage);
3587	supdrvLdrUnlock(pDevExt);
3588	return VINF_SUCCESS;
3589	}
3590	}
3591	/* (not found - add it!) */
3592
3593	/*
3594	* Allocate memory.
3595	*/
3596	pv = RTMemAlloc(RT_OFFSETOF(SUPDRVLDRIMAGE, szName[cchName + 1]));
3597	if (!pv)
3598	{
3599	supdrvLdrUnlock(pDevExt);
3600	Log(("supdrvIOCtl_LdrOpen: RTMemAlloc() failed\n"));
3601	return VERR_NO_MEMORY;
3602	}
3603
3604	/*
3605	* Setup and link in the LDR stuff.
3606	*/
3607	pImage = (PSUPDRVLDRIMAGE)pv;
3608	pImage->pvImage = NULL;
3609	pImage->pvImageAlloc = NULL;
3610	pImage->cbImageWithTabs = pReq->u.In.cbImageWithTabs;
3611	pImage->cbImageBits = pReq->u.In.cbImageBits;
3612	pImage->cSymbols = 0;
3613	pImage->paSymbols = NULL;
3614	pImage->pachStrTab = NULL;
3615	pImage->cbStrTab = 0;
3616	pImage->pfnModuleInit = NULL;
3617	pImage->pfnModuleTerm = NULL;
3618	pImage->pfnServiceReqHandler = NULL;
3619	pImage->uState = SUP_IOCTL_LDR_OPEN;
3620	pImage->cUsage = 1;
3621	memcpy(pImage->szName, pReq->u.In.szName, cchName + 1);
3622
3623	/*
3624	* Try load it using the native loader, if that isn't supported, fall back
3625	* on the older method.
3626	*/
3627	pImage->fNative = true;
3628	rc = supdrvOSLdrOpen(pDevExt, pImage, pReq->u.In.szFilename);
3629	if (rc == VERR_NOT_SUPPORTED)
3630	{
3631	pImage->pvImageAlloc = RTMemExecAlloc(pImage->cbImageBits + 31);
3632	pImage->pvImage = RT_ALIGN_P(pImage->pvImageAlloc, 32);
3633	pImage->fNative = false;
3634	rc = pImage->pvImageAlloc ? VINF_SUCCESS : VERR_NO_MEMORY;
3635	}
3636	if (RT_FAILURE(rc))
3637	{
3638	supdrvLdrUnlock(pDevExt);
3639	RTMemFree(pImage);
3640	Log(("supdrvIOCtl_LdrOpen(%s): failed - %Rrc\n", pReq->u.In.szName, rc));
3641	return rc;
3642	}
3643	Assert(VALID_PTR(pImage->pvImage) \|\| RT_FAILURE(rc));
3644
3645	/*
3646	* Link it.
3647	*/
3648	pImage->pNext = pDevExt->pLdrImages;
3649	pDevExt->pLdrImages = pImage;
3650
3651	supdrvLdrAddUsage(pSession, pImage);
3652
3653	pReq->u.Out.pvImageBase = pImage->pvImage;
3654	pReq->u.Out.fNeedsLoading = true;
3655	pReq->u.Out.fNativeLoader = pImage->fNative;
3656	supdrvLdrUnlock(pDevExt);
3657
3658	#if defined(RT_OS_WINDOWS) && defined(DEBUG)
3659	SUPR0Printf("VBoxDrv: windbg> .reload /f %s=%#p\n", pImage->szName, pImage->pvImage);
3660	#endif
3661	return VINF_SUCCESS;
3662	}
3663
3664
3665	/**
3666	* Worker that validates a pointer to an image entrypoint.
3667	*
3668	* @returns IPRT status code.
3669	* @param pDevExt The device globals.
3670	* @param pImage The loader image.
3671	* @param pv The pointer into the image.
3672	* @param fMayBeNull Whether it may be NULL.
3673	* @param pszWhat What is this entrypoint? (for logging)
3674	* @param pbImageBits The image bits prepared by ring-3.
3675	*
3676	* @remarks Will leave the lock on failure.
3677	*/
3678	static int supdrvLdrValidatePointer(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage, void *pv,
3679	bool fMayBeNull, const uint8_t pbImageBits, const char pszWhat)
3680	{
3681	if (!fMayBeNull \|\| pv)
3682	{
3683	if ((uintptr_t)pv - (uintptr_t)pImage->pvImage >= pImage->cbImageBits)
3684	{
3685	supdrvLdrUnlock(pDevExt);
3686	Log(("Out of range (%p LB %#x): %s=%p\n", pImage->pvImage, pImage->cbImageBits, pszWhat, pv));
3687	return VERR_INVALID_PARAMETER;
3688	}
3689
3690	if (pImage->fNative)
3691	{
3692	int rc = supdrvOSLdrValidatePointer(pDevExt, pImage, pv, pbImageBits);
3693	if (RT_FAILURE(rc))
3694	{
3695	supdrvLdrUnlock(pDevExt);
3696	Log(("Bad entry point address: %s=%p (rc=%Rrc)\n", pszWhat, pv, rc));
3697	return rc;
3698	}
3699	}
3700	}
3701	return VINF_SUCCESS;
3702	}
3703
3704
3705	/**
3706	* Loads the image bits.
3707	*
3708	* This is the 2nd step of the loading.
3709	*
3710	* @returns IPRT status code.
3711	* @param pDevExt Device globals.
3712	* @param pSession Session data.
3713	* @param pReq The request.
3714	*/
3715	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
3716	{
3717	PSUPDRVLDRUSAGE pUsage;
3718	PSUPDRVLDRIMAGE pImage;
3719	int rc;
3720	LogFlow(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImageWithBits=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImageWithTabs));
3721
3722	/*
3723	* Find the ldr image.
3724	*/
3725	supdrvLdrLock(pDevExt);
3726	pUsage = pSession->pLdrUsage;
3727	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3728	pUsage = pUsage->pNext;
3729	if (!pUsage)
3730	{
3731	supdrvLdrUnlock(pDevExt);
3732	Log(("SUP_IOCTL_LDR_LOAD: couldn't find image!\n"));
3733	return VERR_INVALID_HANDLE;
3734	}
3735	pImage = pUsage->pImage;
3736
3737	/*
3738	* Validate input.
3739	*/
3740	if ( pImage->cbImageWithTabs != pReq->u.In.cbImageWithTabs
3741	\|\| pImage->cbImageBits != pReq->u.In.cbImageBits)
3742	{
3743	supdrvLdrUnlock(pDevExt);
3744	Log(("SUP_IOCTL_LDR_LOAD: image size mismatch!! %d(prep) != %d(load) or %d != %d\n",
3745	pImage->cbImageWithTabs, pReq->u.In.cbImageWithTabs, pImage->cbImageBits, pReq->u.In.cbImageBits));
3746	return VERR_INVALID_HANDLE;
3747	}
3748
3749	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
3750	{
3751	unsigned uState = pImage->uState;
3752	supdrvLdrUnlock(pDevExt);
3753	if (uState != SUP_IOCTL_LDR_LOAD)
3754	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
3755	return VERR_ALREADY_LOADED;
3756	}
3757
3758	switch (pReq->u.In.eEPType)
3759	{
3760	case SUPLDRLOADEP_NOTHING:
3761	break;
3762
3763	case SUPLDRLOADEP_VMMR0:
3764	rc = supdrvLdrValidatePointer( pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0, false, pReq->u.In.abImage, "pvVMMR0");
3765	if (RT_SUCCESS(rc))
3766	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt, false, pReq->u.In.abImage, "pvVMMR0EntryInt");
3767	if (RT_SUCCESS(rc))
3768	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, false, pReq->u.In.abImage, "pvVMMR0EntryFast");
3769	if (RT_SUCCESS(rc))
3770	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx, false, pReq->u.In.abImage, "pvVMMR0EntryEx");
3771	if (RT_FAILURE(rc))
3772	return rc;
3773	break;
3774
3775	case SUPLDRLOADEP_SERVICE:
3776	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.Service.pfnServiceReq, false, pReq->u.In.abImage, "pfnServiceReq");
3777	if (RT_FAILURE(rc))
3778	return rc;
3779	if ( pReq->u.In.EP.Service.apvReserved[0] != NIL_RTR0PTR
3780	\|\| pReq->u.In.EP.Service.apvReserved[1] != NIL_RTR0PTR
3781	\|\| pReq->u.In.EP.Service.apvReserved[2] != NIL_RTR0PTR)
3782	{
3783	supdrvLdrUnlock(pDevExt);
3784	Log(("Out of range (%p LB %#x): apvReserved={%p,%p,%p} MBZ!\n",
3785	pImage->pvImage, pReq->u.In.cbImageWithTabs,
3786	pReq->u.In.EP.Service.apvReserved[0],
3787	pReq->u.In.EP.Service.apvReserved[1],
3788	pReq->u.In.EP.Service.apvReserved[2]));
3789	return VERR_INVALID_PARAMETER;
3790	}
3791	break;
3792
3793	default:
3794	supdrvLdrUnlock(pDevExt);
3795	Log(("Invalid eEPType=%d\n", pReq->u.In.eEPType));
3796	return VERR_INVALID_PARAMETER;
3797	}
3798
3799	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleInit, true, pReq->u.In.abImage, "pfnModuleInit");
3800	if (RT_FAILURE(rc))
3801	return rc;
3802	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleTerm, true, pReq->u.In.abImage, "pfnModuleTerm");
3803	if (RT_FAILURE(rc))
3804	return rc;
3805
3806	/*
3807	* Allocate and copy the tables.
3808	* (No need to do try/except as this is a buffered request.)
3809	*/
3810	pImage->cbStrTab = pReq->u.In.cbStrTab;
3811	if (pImage->cbStrTab)
3812	{
3813	pImage->pachStrTab = (char *)RTMemAlloc(pImage->cbStrTab);
3814	if (pImage->pachStrTab)
3815	memcpy(pImage->pachStrTab, &pReq->u.In.abImage[pReq->u.In.offStrTab], pImage->cbStrTab);
3816	else
3817	rc = VERR_NO_MEMORY;
3818	}
3819
3820	pImage->cSymbols = pReq->u.In.cSymbols;
3821	if (RT_SUCCESS(rc) && pImage->cSymbols)
3822	{
3823	size_t cbSymbols = pImage->cSymbols * sizeof(SUPLDRSYM);
3824	pImage->paSymbols = (PSUPLDRSYM)RTMemAlloc(cbSymbols);
3825	if (pImage->paSymbols)
3826	memcpy(pImage->paSymbols, &pReq->u.In.abImage[pReq->u.In.offSymbols], cbSymbols);
3827	else
3828	rc = VERR_NO_MEMORY;
3829	}
3830
3831	/*
3832	* Copy the bits / complete native loading.
3833	*/
3834	if (RT_SUCCESS(rc))
3835	{
3836	pImage->uState = SUP_IOCTL_LDR_LOAD;
3837	pImage->pfnModuleInit = pReq->u.In.pfnModuleInit;
3838	pImage->pfnModuleTerm = pReq->u.In.pfnModuleTerm;
3839
3840	if (pImage->fNative)
3841	rc = supdrvOSLdrLoad(pDevExt, pImage, pReq->u.In.abImage);
3842	else
3843	memcpy(pImage->pvImage, &pReq->u.In.abImage[0], pImage->cbImageBits);
3844	}
3845
3846	/*
3847	* Update any entry points.
3848	*/
3849	if (RT_SUCCESS(rc))
3850	{
3851	switch (pReq->u.In.eEPType)
3852	{
3853	default:
3854	case SUPLDRLOADEP_NOTHING:
3855	rc = VINF_SUCCESS;
3856	break;
3857	case SUPLDRLOADEP_VMMR0:
3858	rc = supdrvLdrSetVMMR0EPs(pDevExt, pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
3859	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
3860	break;
3861	case SUPLDRLOADEP_SERVICE:
3862	pImage->pfnServiceReqHandler = pReq->u.In.EP.Service.pfnServiceReq;
3863	rc = VINF_SUCCESS;
3864	break;
3865	}
3866	}
3867
3868	/*
3869	* On success call the module initialization.
3870	*/
3871	LogFlow(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
3872	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
3873	{
3874	Log(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
3875	rc = pImage->pfnModuleInit();
3876	if (rc && pDevExt->pvVMMR0 == pImage->pvImage)
3877	supdrvLdrUnsetVMMR0EPs(pDevExt);
3878	}
3879
3880	if (RT_FAILURE(rc))
3881	{
3882	pImage->uState = SUP_IOCTL_LDR_OPEN;
3883	pImage->pfnModuleInit = NULL;
3884	pImage->pfnModuleTerm = NULL;
3885	pImage->pfnServiceReqHandler= NULL;
3886	pImage->cbStrTab = 0;
3887	RTMemFree(pImage->pachStrTab);
3888	pImage->pachStrTab = NULL;
3889	RTMemFree(pImage->paSymbols);
3890	pImage->paSymbols = NULL;
3891	pImage->cSymbols = 0;
3892	}
3893
3894	supdrvLdrUnlock(pDevExt);
3895	return rc;
3896	}
3897
3898
3899	/**
3900	* Frees a previously loaded (prep'ed) image.
3901	*
3902	* @returns IPRT status code.
3903	* @param pDevExt Device globals.
3904	* @param pSession Session data.
3905	* @param pReq The request.
3906	*/
3907	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
3908	{
3909	int rc;
3910	PSUPDRVLDRUSAGE pUsagePrev;
3911	PSUPDRVLDRUSAGE pUsage;
3912	PSUPDRVLDRIMAGE pImage;
3913	LogFlow(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
3914
3915	/*
3916	* Find the ldr image.
3917	*/
3918	supdrvLdrLock(pDevExt);
3919	pUsagePrev = NULL;
3920	pUsage = pSession->pLdrUsage;
3921	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3922	{
3923	pUsagePrev = pUsage;
3924	pUsage = pUsage->pNext;
3925	}
3926	if (!pUsage)
3927	{
3928	supdrvLdrUnlock(pDevExt);
3929	Log(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
3930	return VERR_INVALID_HANDLE;
3931	}
3932
3933	/*
3934	* Check if we can remove anything.
3935	*/
3936	rc = VINF_SUCCESS;
3937	pImage = pUsage->pImage;
3938	if (pImage->cUsage <= 1 \|\| pUsage->cUsage <= 1)
3939	{
3940	/*
3941	* Check if there are any objects with destructors in the image, if
3942	* so leave it for the session cleanup routine so we get a chance to
3943	* clean things up in the right order and not leave them all dangling.
3944	*/
3945	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3946	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
3947	if (pImage->cUsage <= 1)
3948	{
3949	PSUPDRVOBJ pObj;
3950	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
3951	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
3952	{
3953	rc = VERR_DANGLING_OBJECTS;
3954	break;
3955	}
3956	}
3957	else
3958	{
3959	PSUPDRVUSAGE pGenUsage;
3960	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
3961	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
3962	{
3963	rc = VERR_DANGLING_OBJECTS;
3964	break;
3965	}
3966	}
3967	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
3968	if (rc == VINF_SUCCESS)
3969	{
3970	/* unlink it */
3971	if (pUsagePrev)
3972	pUsagePrev->pNext = pUsage->pNext;
3973	else
3974	pSession->pLdrUsage = pUsage->pNext;
3975
3976	/* free it */
3977	pUsage->pImage = NULL;
3978	pUsage->pNext = NULL;
3979	RTMemFree(pUsage);
3980
3981	/*
3982	* Dereference the image.
3983	*/
3984	if (pImage->cUsage <= 1)
3985	supdrvLdrFree(pDevExt, pImage);
3986	else
3987	pImage->cUsage--;
3988	}
3989	else
3990	{
3991	Log(("supdrvIOCtl_LdrFree: Dangling objects in %p/%s!\n", pImage->pvImage, pImage->szName));
3992	rc = VINF_SUCCESS; /** @todo BRANCH-2.1: remove this after branching. */
3993	}
3994	}
3995	else
3996	{
3997	/*
3998	* Dereference both image and usage.
3999	*/
4000	pImage->cUsage--;
4001	pUsage->cUsage--;
4002	}
4003
4004	supdrvLdrUnlock(pDevExt);
4005	return rc;
4006	}
4007
4008
4009	/**
4010	* Gets the address of a symbol in an open image.
4011	*
4012	* @returns IPRT status code.
4013	* @param pDevExt Device globals.
4014	* @param pSession Session data.
4015	* @param pReq The request buffer.
4016	*/
4017	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
4018	{
4019	PSUPDRVLDRIMAGE pImage;
4020	PSUPDRVLDRUSAGE pUsage;
4021	uint32_t i;
4022	PSUPLDRSYM paSyms;
4023	const char *pchStrings;
4024	const size_t cbSymbol = strlen(pReq->u.In.szSymbol) + 1;
4025	void *pvSymbol = NULL;
4026	int rc = VERR_GENERAL_FAILURE;
4027	Log3(("supdrvIOCtl_LdrGetSymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
4028
4029	/*
4030	* Find the ldr image.
4031	*/
4032	supdrvLdrLock(pDevExt);
4033	pUsage = pSession->pLdrUsage;
4034	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4035	pUsage = pUsage->pNext;
4036	if (!pUsage)
4037	{
4038	supdrvLdrUnlock(pDevExt);
4039	Log(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
4040	return VERR_INVALID_HANDLE;
4041	}
4042	pImage = pUsage->pImage;
4043	if (pImage->uState != SUP_IOCTL_LDR_LOAD)
4044	{
4045	unsigned uState = pImage->uState;
4046	supdrvLdrUnlock(pDevExt);
4047	Log(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", uState, uState)); NOREF(uState);
4048	return VERR_ALREADY_LOADED;
4049	}
4050
4051	/*
4052	* Search the symbol strings.
4053	*/
4054	pchStrings = pImage->pachStrTab;
4055	paSyms = pImage->paSymbols;
4056	for (i = 0; i < pImage->cSymbols; i++)
4057	{
4058	if ( paSyms[i].offSymbol < pImage->cbImageBits /* paranoia */
4059	&& paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4060	&& !memcmp(pchStrings + paSyms[i].offName, pReq->u.In.szSymbol, cbSymbol))
4061	{
4062	pvSymbol = (uint8_t *)pImage->pvImage + paSyms[i].offSymbol;
4063	rc = VINF_SUCCESS;
4064	break;
4065	}
4066	}
4067	supdrvLdrUnlock(pDevExt);
4068	pReq->u.Out.pvSymbol = pvSymbol;
4069	return rc;
4070	}
4071
4072
4073	/**
4074	* Gets the address of a symbol in an open image or the support driver.
4075	*
4076	* @returns VINF_SUCCESS on success.
4077	* @returns
4078	* @param pDevExt Device globals.
4079	* @param pSession Session data.
4080	* @param pReq The request buffer.
4081	*/
4082	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq)
4083	{
4084	int rc = VINF_SUCCESS;
4085	const char *pszSymbol = pReq->u.In.pszSymbol;
4086	const char *pszModule = pReq->u.In.pszModule;
4087	size_t cbSymbol;
4088	char const *pszEnd;
4089	uint32_t i;
4090
4091	/*
4092	* Input validation.
4093	*/
4094	AssertPtrReturn(pszSymbol, VERR_INVALID_POINTER);
4095	pszEnd = (char *)memchr(pszSymbol, '\0', 512);
4096	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4097	cbSymbol = pszEnd - pszSymbol + 1;
4098
4099	if (pszModule)
4100	{
4101	AssertPtrReturn(pszModule, VERR_INVALID_POINTER);
4102	pszEnd = (char *)memchr(pszModule, '\0', 64);
4103	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4104	}
4105	Log3(("supdrvIDC_LdrGetSymbol: pszModule=%p:{%s} pszSymbol=%p:{%s}\n", pszModule, pszModule, pszSymbol, pszSymbol));
4106
4107
4108	if ( !pszModule
4109	\|\| !strcmp(pszModule, "SupDrv"))
4110	{
4111	/*
4112	* Search the support driver export table.
4113	*/
4114	for (i = 0; i < RT_ELEMENTS(g_aFunctions); i++)
4115	if (!strcmp(g_aFunctions[i].szName, pszSymbol))
4116	{
4117	pReq->u.Out.pfnSymbol = g_aFunctions[i].pfn;
4118	break;
4119	}
4120	}
4121	else
4122	{
4123	/*
4124	* Find the loader image.
4125	*/
4126	PSUPDRVLDRIMAGE pImage;
4127
4128	supdrvLdrLock(pDevExt);
4129
4130	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
4131	if (!strcmp(pImage->szName, pszModule))
4132	break;
4133	if (pImage && pImage->uState == SUP_IOCTL_LDR_LOAD)
4134	{
4135	/*
4136	* Search the symbol strings.
4137	*/
4138	const char *pchStrings = pImage->pachStrTab;
4139	PCSUPLDRSYM paSyms = pImage->paSymbols;
4140	for (i = 0; i < pImage->cSymbols; i++)
4141	{
4142	if ( paSyms[i].offSymbol < pImage->cbImageBits /* paranoia */
4143	&& paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4144	&& !memcmp(pchStrings + paSyms[i].offName, pszSymbol, cbSymbol))
4145	{
4146	/*
4147	* Found it! Calc the symbol address and add a reference to the module.
4148	*/
4149	pReq->u.Out.pfnSymbol = (PFNRT)((uint8_t *)pImage->pvImage + paSyms[i].offSymbol);
4150	rc = supdrvLdrAddUsage(pSession, pImage);
4151	break;
4152	}
4153	}
4154	}
4155	else
4156	rc = pImage ? VERR_WRONG_ORDER : VERR_MODULE_NOT_FOUND;
4157
4158	supdrvLdrUnlock(pDevExt);
4159	}
4160	return rc;
4161	}
4162
4163
4164	/**
4165	* Updates the VMMR0 entry point pointers.
4166	*
4167	* @returns IPRT status code.
4168	* @param pDevExt Device globals.
4169	* @param pSession Session data.
4170	* @param pVMMR0 VMMR0 image handle.
4171	* @param pvVMMR0EntryInt VMMR0EntryInt address.
4172	* @param pvVMMR0EntryFast VMMR0EntryFast address.
4173	* @param pvVMMR0EntryEx VMMR0EntryEx address.
4174	* @remark Caller must own the loader mutex.
4175	*/
4176	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx)
4177	{
4178	int rc = VINF_SUCCESS;
4179	LogFlow(("supdrvLdrSetR0EP pvVMMR0=%p pvVMMR0EntryInt=%p\n", pvVMMR0, pvVMMR0EntryInt));
4180
4181
4182	/*
4183	* Check if not yet set.
4184	*/
4185	if (!pDevExt->pvVMMR0)
4186	{
4187	pDevExt->pvVMMR0 = pvVMMR0;
4188	pDevExt->pfnVMMR0EntryInt = pvVMMR0EntryInt;
4189	pDevExt->pfnVMMR0EntryFast = pvVMMR0EntryFast;
4190	pDevExt->pfnVMMR0EntryEx = pvVMMR0EntryEx;
4191	}
4192	else
4193	{
4194	/*
4195	* Return failure or success depending on whether the values match or not.
4196	*/
4197	if ( pDevExt->pvVMMR0 != pvVMMR0
4198	\|\| (void *)pDevExt->pfnVMMR0EntryInt != pvVMMR0EntryInt
4199	\|\| (void *)pDevExt->pfnVMMR0EntryFast != pvVMMR0EntryFast
4200	\|\| (void *)pDevExt->pfnVMMR0EntryEx != pvVMMR0EntryEx)
4201	{
4202	AssertMsgFailed(("SUP_IOCTL_LDR_SETR0EP: Already set pointing to a different module!\n"));
4203	rc = VERR_INVALID_PARAMETER;
4204	}
4205	}
4206	return rc;
4207	}
4208
4209
4210	/**
4211	* Unsets the VMMR0 entry point installed by supdrvLdrSetR0EP.
4212	*
4213	* @param pDevExt Device globals.
4214	*/
4215	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt)
4216	{
4217	pDevExt->pvVMMR0 = NULL;
4218	pDevExt->pfnVMMR0EntryInt = NULL;
4219	pDevExt->pfnVMMR0EntryFast = NULL;
4220	pDevExt->pfnVMMR0EntryEx = NULL;
4221	}
4222
4223
4224	/**
4225	* Adds a usage reference in the specified session of an image.
4226	*
4227	* Called while owning the loader semaphore.
4228	*
4229	* @returns VINF_SUCCESS on success and VERR_NO_MEMORY on failure.
4230	* @param pSession Session in question.
4231	* @param pImage Image which the session is using.
4232	*/
4233	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage)
4234	{
4235	PSUPDRVLDRUSAGE pUsage;
4236	LogFlow(("supdrvLdrAddUsage: pImage=%p\n", pImage));
4237
4238	/*
4239	* Referenced it already?
4240	*/
4241	pUsage = pSession->pLdrUsage;
4242	while (pUsage)
4243	{
4244	if (pUsage->pImage == pImage)
4245	{
4246	pUsage->cUsage++;
4247	return VINF_SUCCESS;
4248	}
4249	pUsage = pUsage->pNext;
4250	}
4251
4252	/*
4253	* Allocate new usage record.
4254	*/
4255	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
4256	AssertReturn(pUsage, VERR_NO_MEMORY);
4257	pUsage->cUsage = 1;
4258	pUsage->pImage = pImage;
4259	pUsage->pNext = pSession->pLdrUsage;
4260	pSession->pLdrUsage = pUsage;
4261	return VINF_SUCCESS;
4262	}
4263
4264
4265	/**
4266	* Frees a load image.
4267	*
4268	* @param pDevExt Pointer to device extension.
4269	* @param pImage Pointer to the image we're gonna free.
4270	* This image must exit!
4271	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
4272	*/
4273	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
4274	{
4275	PSUPDRVLDRIMAGE pImagePrev;
4276	LogFlow(("supdrvLdrFree: pImage=%p\n", pImage));
4277
4278	/* find it - arg. should've used doubly linked list. */
4279	Assert(pDevExt->pLdrImages);
4280	pImagePrev = NULL;
4281	if (pDevExt->pLdrImages != pImage)
4282	{
4283	pImagePrev = pDevExt->pLdrImages;
4284	while (pImagePrev->pNext != pImage)
4285	pImagePrev = pImagePrev->pNext;
4286	Assert(pImagePrev->pNext == pImage);
4287	}
4288
4289	/* unlink */
4290	if (pImagePrev)
4291	pImagePrev->pNext = pImage->pNext;
4292	else
4293	pDevExt->pLdrImages = pImage->pNext;
4294
4295	/* check if this is VMMR0.r0 unset its entry point pointers. */
4296	if (pDevExt->pvVMMR0 == pImage->pvImage)
4297	supdrvLdrUnsetVMMR0EPs(pDevExt);
4298
4299	/* check for objects with destructors in this image. (Shouldn't happen.) */
4300	if (pDevExt->pObjs)
4301	{
4302	unsigned cObjs = 0;
4303	PSUPDRVOBJ pObj;
4304	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
4305	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
4306	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4307	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4308	{
4309	pObj->pfnDestructor = NULL;
4310	cObjs++;
4311	}
4312	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
4313	if (cObjs)
4314	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
4315	}
4316
4317	/* call termination function if fully loaded. */
4318	if ( pImage->pfnModuleTerm
4319	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
4320	{
4321	LogFlow(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
4322	pImage->pfnModuleTerm();
4323	}
4324
4325	/* do native unload if appropriate. */
4326	if (pImage->fNative)
4327	supdrvOSLdrUnload(pDevExt, pImage);
4328
4329	/* free the image */
4330	pImage->cUsage = 0;
4331	pImage->pNext = 0;
4332	pImage->uState = SUP_IOCTL_LDR_FREE;
4333	RTMemExecFree(pImage->pvImageAlloc);
4334	pImage->pvImageAlloc = NULL;
4335	RTMemFree(pImage->pachStrTab);
4336	pImage->pachStrTab = NULL;
4337	RTMemFree(pImage->paSymbols);
4338	pImage->paSymbols = NULL;
4339	RTMemFree(pImage);
4340	}
4341
4342
4343	/**
4344	* Acquires the loader lock.
4345	*
4346	* @returns IPRT status code.
4347	* @param pDevExt The device extension.
4348	*/
4349	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt)
4350	{
4351	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4352	int rc = RTSemMutexRequest(pDevExt->mtxLdr, RT_INDEFINITE_WAIT);
4353	#else
4354	int rc = RTSemFastMutexRequest(pDevExt->mtxLdr);
4355	#endif
4356	AssertRC(rc);
4357	return rc;
4358	}
4359
4360
4361	/**
4362	* Releases the loader lock.
4363	*
4364	* @returns IPRT status code.
4365	* @param pDevExt The device extension.
4366	*/
4367	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt)
4368	{
4369	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4370	return RTSemMutexRelease(pDevExt->mtxLdr);
4371	#else
4372	return RTSemFastMutexRelease(pDevExt->mtxLdr);
4373	#endif
4374	}
4375
4376
4377	/**
4378	* Implements the service call request.
4379	*
4380	* @returns VBox status code.
4381	* @param pDevExt The device extension.
4382	* @param pSession The calling session.
4383	* @param pReq The request packet, valid.
4384	*/
4385	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq)
4386	{
4387	#if !defined(RT_OS_WINDOWS) \|\| defined(DEBUG)
4388	int rc;
4389
4390	/*
4391	* Find the module first in the module referenced by the calling session.
4392	*/
4393	rc = supdrvLdrLock(pDevExt);
4394	if (RT_SUCCESS(rc))
4395	{
4396	PFNSUPR0SERVICEREQHANDLER pfnServiceReqHandler = NULL;
4397	PSUPDRVLDRUSAGE pUsage;
4398
4399	for (pUsage = pSession->pLdrUsage; pUsage; pUsage = pUsage->pNext)
4400	if ( pUsage->pImage->pfnServiceReqHandler
4401	&& !strcmp(pUsage->pImage->szName, pReq->u.In.szName))
4402	{
4403	pfnServiceReqHandler = pUsage->pImage->pfnServiceReqHandler;
4404	break;
4405	}
4406	supdrvLdrUnlock(pDevExt);
4407
4408	if (pfnServiceReqHandler)
4409	{
4410	/*
4411	* Call it.
4412	*/
4413	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
4414	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, NULL);
4415	else
4416	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0]);
4417	}
4418	else
4419	rc = VERR_SUPDRV_SERVICE_NOT_FOUND;
4420	}
4421
4422	/* log it */
4423	if ( RT_FAILURE(rc)
4424	&& rc != VERR_INTERRUPTED
4425	&& rc != VERR_TIMEOUT)
4426	Log(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4427	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4428	else
4429	Log4(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4430	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4431	return rc;
4432	#else /* RT_OS_WINDOWS && !DEBUG */
4433	return VERR_NOT_IMPLEMENTED;
4434	#endif /* RT_OS_WINDOWS && !DEBUG */
4435	}
4436
4437
4438	/**
4439	* Implements the logger settings request.
4440	*
4441	* @returns VBox status code.
4442	* @param pDevExt The device extension.
4443	* @param pSession The caller's session.
4444	* @param pReq The request.
4445	*/
4446	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq)
4447	{
4448	const char *pszGroup = &pReq->u.In.szStrings[pReq->u.In.offGroups];
4449	const char *pszFlags = &pReq->u.In.szStrings[pReq->u.In.offFlags];
4450	const char *pszDest = &pReq->u.In.szStrings[pReq->u.In.offDestination];
4451	PRTLOGGER pLogger = NULL;
4452	int rc;
4453
4454	/*
4455	* Some further validation.
4456	*/
4457	switch (pReq->u.In.fWhat)
4458	{
4459	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4460	case SUPLOGGERSETTINGS_WHAT_CREATE:
4461	break;
4462
4463	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4464	if (pszGroup \|\| pszFlags \|\| *pszDest)
4465	return VERR_INVALID_PARAMETER;
4466	if (pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_RELEASE)
4467	return VERR_ACCESS_DENIED;
4468	break;
4469
4470	default:
4471	return VERR_INTERNAL_ERROR;
4472	}
4473
4474	/*
4475	* Get the logger.
4476	*/
4477	switch (pReq->u.In.fWhich)
4478	{
4479	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4480	pLogger = RTLogGetDefaultInstance();
4481	break;
4482
4483	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4484	pLogger = RTLogRelDefaultInstance();
4485	break;
4486
4487	default:
4488	return VERR_INTERNAL_ERROR;
4489	}
4490
4491	/*
4492	* Do the job.
4493	*/
4494	switch (pReq->u.In.fWhat)
4495	{
4496	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4497	if (pLogger)
4498	{
4499	rc = RTLogFlags(pLogger, pszFlags);
4500	if (RT_SUCCESS(rc))
4501	rc = RTLogGroupSettings(pLogger, pszGroup);
4502	NOREF(pszDest);
4503	}
4504	else
4505	rc = VERR_NOT_FOUND;
4506	break;
4507
4508	case SUPLOGGERSETTINGS_WHAT_CREATE:
4509	{
4510	if (pLogger)
4511	rc = VERR_ALREADY_EXISTS;
4512	else
4513	{
4514	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
4515
4516	rc = RTLogCreate(&pLogger,
4517	0 /* fFlags */,
4518	pszGroup,
4519	pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_DEBUG
4520	? "VBOX_LOG"
4521	: "VBOX_RELEASE_LOG",
4522	RT_ELEMENTS(s_apszGroups),
4523	s_apszGroups,
4524	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER,
4525	NULL);
4526	if (RT_SUCCESS(rc))
4527	{
4528	rc = RTLogFlags(pLogger, pszFlags);
4529	NOREF(pszDest);
4530	if (RT_SUCCESS(rc))
4531	{
4532	switch (pReq->u.In.fWhich)
4533	{
4534	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4535	pLogger = RTLogSetDefaultInstance(pLogger);
4536	break;
4537	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4538	pLogger = RTLogRelSetDefaultInstance(pLogger);
4539	break;
4540	}
4541	}
4542	RTLogDestroy(pLogger);
4543	}
4544	}
4545	break;
4546	}
4547
4548	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4549	switch (pReq->u.In.fWhich)
4550	{
4551	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4552	pLogger = RTLogSetDefaultInstance(NULL);
4553	break;
4554	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4555	pLogger = RTLogRelSetDefaultInstance(NULL);
4556	break;
4557	}
4558	rc = RTLogDestroy(pLogger);
4559	break;
4560
4561	default:
4562	{
4563	rc = VERR_INTERNAL_ERROR;
4564	break;
4565	}
4566	}
4567
4568	return rc;
4569	}
4570
4571
4572	/**
4573	* Creates the GIP.
4574	*
4575	* @returns VBox status code.
4576	* @param pDevExt Instance data. GIP stuff may be updated.
4577	*/
4578	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
4579	{
4580	PSUPGLOBALINFOPAGE pGip;
4581	RTHCPHYS HCPhysGip;
4582	uint32_t u32SystemResolution;
4583	uint32_t u32Interval;
4584	int rc;
4585
4586	LogFlow(("supdrvGipCreate:\n"));
4587
4588	/* assert order */
4589	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
4590	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
4591	Assert(!pDevExt->pGipTimer);
4592
4593	/*
4594	* Allocate a suitable page with a default kernel mapping.
4595	*/
4596	rc = RTR0MemObjAllocLow(&pDevExt->GipMemObj, PAGE_SIZE, false);
4597	if (RT_FAILURE(rc))
4598	{
4599	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
4600	return rc;
4601	}
4602	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
4603	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
4604
4605	#if 0 /** @todo Disabled this as we didn't used to do it before and causes unnecessary stress on laptops.
4606	* It only applies to Windows and should probably revisited later, if possible made part of the
4607	* timer code (return min granularity in RTTimerGetSystemGranularity and set it in RTTimerStart). */
4608	/*
4609	* Try bump up the system timer resolution.
4610	* The more interrupts the better...
4611	*/
4612	if ( RT_SUCCESS(RTTimerRequestSystemGranularity( 488281 /* 2048 HZ */, &u32SystemResolution))
4613	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 500000 /* 2000 HZ */, &u32SystemResolution))
4614	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
4615	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
4616	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 1953125 /* 512 HZ */, &u32SystemResolution))
4617	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 2000000 /* 500 HZ */, &u32SystemResolution))
4618	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 3906250 /* 256 HZ */, &u32SystemResolution))
4619	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 4000000 /* 250 HZ */, &u32SystemResolution))
4620	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 7812500 /* 128 HZ */, &u32SystemResolution))
4621	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity(10000000 /* 100 HZ */, &u32SystemResolution))
4622	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity(15625000 /* 64 HZ */, &u32SystemResolution))
4623	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity(31250000 /* 32 HZ */, &u32SystemResolution))
4624	)
4625	{
4626	Assert(RTTimerGetSystemGranularity() <= u32SystemResolution);
4627	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
4628	}
4629	#endif
4630
4631	/*
4632	* Find a reasonable update interval and initialize the structure.
4633	*/
4634	u32Interval = u32SystemResolution = RTTimerGetSystemGranularity();
4635	while (u32Interval < 10000000 /* 10 ms */)
4636	u32Interval += u32SystemResolution;
4637
4638	supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), 1000000000 / u32Interval /=Hz/);
4639
4640	/*
4641	* Create the timer.
4642	* If CPU_ALL isn't supported we'll have to fall back to synchronous mode.
4643	*/
4644	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
4645	{
4646	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, RTTIMER_FLAGS_CPU_ALL, supdrvGipAsyncTimer, pDevExt);
4647	if (rc == VERR_NOT_SUPPORTED)
4648	{
4649	OSDBGPRINT(("supdrvGipCreate: omni timer not supported, falling back to synchronous mode\n"));
4650	pGip->u32Mode = SUPGIPMODE_SYNC_TSC;
4651	}
4652	}
4653	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
4654	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0, supdrvGipSyncTimer, pDevExt);
4655	if (RT_SUCCESS(rc))
4656	{
4657	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
4658	rc = RTMpNotificationRegister(supdrvGipMpEvent, pDevExt);
4659	if (RT_SUCCESS(rc))
4660	{
4661	/*
4662	* We're good.
4663	*/
4664	Log(("supdrvGipCreate: %ld ns interval.\n", (long)u32Interval));
4665	g_pSUPGlobalInfoPage = pGip;
4666	return VINF_SUCCESS;
4667	}
4668
4669	OSDBGPRINT(("supdrvGipCreate: failed register MP event notfication. rc=%d\n", rc));
4670	}
4671	else
4672	{
4673	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %ld ns interval. rc=%d\n", (long)u32Interval, rc));
4674	Assert(!pDevExt->pGipTimer);
4675	}
4676	supdrvGipDestroy(pDevExt);
4677	return rc;
4678	}
4679
4680
4681	/**
4682	* Terminates the GIP.
4683	*
4684	* @param pDevExt Instance data. GIP stuff may be updated.
4685	*/
4686	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
4687	{
4688	int rc;
4689	#ifdef DEBUG_DARWIN_GIP
4690	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
4691	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
4692	pDevExt->pGipTimer, pDevExt->GipMemObj));
4693	#endif
4694
4695	/*
4696	* Invalid the GIP data.
4697	*/
4698	if (pDevExt->pGip)
4699	{
4700	supdrvGipTerm(pDevExt->pGip);
4701	pDevExt->pGip = NULL;
4702	}
4703	g_pSUPGlobalInfoPage = NULL;
4704
4705	/*
4706	* Destroy the timer and free the GIP memory object.
4707	*/
4708	if (pDevExt->pGipTimer)
4709	{
4710	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
4711	pDevExt->pGipTimer = NULL;
4712	}
4713
4714	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
4715	{
4716	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
4717	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
4718	}
4719
4720	/*
4721	* Finally, release the system timer resolution request if one succeeded.
4722	*/
4723	if (pDevExt->u32SystemTimerGranularityGrant)
4724	{
4725	rc = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
4726	pDevExt->u32SystemTimerGranularityGrant = 0;
4727	}
4728	}
4729
4730
4731	/**
4732	* Timer callback function sync GIP mode.
4733	* @param pTimer The timer.
4734	* @param pvUser The device extension.
4735	*/
4736	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
4737	{
4738	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
4739	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4740	uint64_t u64TSC = ASMReadTSC();
4741	uint64_t NanoTS = RTTimeSystemNanoTS();
4742
4743	supdrvGipUpdate(pDevExt->pGip, NanoTS, u64TSC, iTick);
4744
4745	ASMSetFlags(fOldFlags);
4746	}
4747
4748
4749	/**
4750	* Timer callback function for async GIP mode.
4751	* @param pTimer The timer.
4752	* @param pvUser The device extension.
4753	*/
4754	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
4755	{
4756	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
4757	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4758	RTCPUID idCpu = RTMpCpuId();
4759	uint64_t u64TSC = ASMReadTSC();
4760	uint64_t NanoTS = RTTimeSystemNanoTS();
4761
4762	/** @todo reset the transaction number and whatnot when iTick == 1. */
4763	if (pDevExt->idGipMaster == idCpu)
4764	supdrvGipUpdate(pDevExt->pGip, NanoTS, u64TSC, iTick);
4765	else
4766	supdrvGipUpdatePerCpu(pDevExt->pGip, NanoTS, u64TSC, ASMGetApicId(), iTick);
4767
4768	ASMSetFlags(fOldFlags);
4769	}
4770
4771
4772	/**
4773	* Multiprocessor event notification callback.
4774	*
4775	* This is used to make sue that the GIP master gets passed on to
4776	* another CPU.
4777	*
4778	* @param enmEvent The event.
4779	* @param idCpu The cpu it applies to.
4780	* @param pvUser Pointer to the device extension.
4781	*/
4782	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser)
4783	{
4784	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4785	if (enmEvent == RTMPEVENT_OFFLINE)
4786	{
4787	RTCPUID idGipMaster;
4788	ASMAtomicReadSize(&pDevExt->idGipMaster, &idGipMaster);
4789	if (idGipMaster == idCpu)
4790	{
4791	/*
4792	* Find a new GIP master.
4793	*/
4794	bool fIgnored;
4795	unsigned i;
4796	RTCPUID idNewGipMaster = NIL_RTCPUID;
4797	RTCPUSET OnlineCpus;
4798	RTMpGetOnlineSet(&OnlineCpus);
4799
4800	for (i = 0; i < RTCPUSET_MAX_CPUS; i++)
4801	{
4802	RTCPUID idCurCpu = RTMpCpuIdFromSetIndex(i);
4803	if ( RTCpuSetIsMember(&OnlineCpus, idCurCpu)
4804	&& idCurCpu != idGipMaster)
4805	{
4806	idNewGipMaster = idCurCpu;
4807	break;
4808	}
4809	}
4810
4811	Log(("supdrvGipMpEvent: Gip master %#lx -> %#lx\n", (long)idGipMaster, (long)idNewGipMaster));
4812	ASMAtomicCmpXchgSize(&pDevExt->idGipMaster, idNewGipMaster, idGipMaster, fIgnored);
4813	NOREF(fIgnored);
4814	}
4815	}
4816	}
4817
4818
4819	/**
4820	* Callback used by supdrvDetermineAsyncTSC to read the TSC on a CPU.
4821	*
4822	* @param idCpu Ignored.
4823	* @param pvUser1 Where to put the TSC.
4824	* @param pvUser2 Ignored.
4825	*/
4826	static DECLCALLBACK(void) supdrvDetermineAsyncTscWorker(RTCPUID idCpu, void pvUser1, void pvUser2)
4827	{
4828	#if 1
4829	ASMAtomicWriteU64((uint64_t volatile *)pvUser1, ASMReadTSC());
4830	#else
4831	(uint64_t )pvUser1 = ASMReadTSC();
4832	#endif
4833	}
4834
4835
4836	/**
4837	* Determine if Async GIP mode is required because of TSC drift.
4838	*
4839	* When using the default/normal timer code it is essential that the time stamp counter
4840	* (TSC) runs never backwards, that is, a read operation to the counter should return
4841	* a bigger value than any previous read operation. This is guaranteed by the latest
4842	* AMD CPUs and by newer Intel CPUs which never enter the C2 state (P4). In any other
4843	* case we have to choose the asynchronous timer mode.
4844	*
4845	* @param poffMin Pointer to the determined difference between different cores.
4846	* @return false if the time stamp counters appear to be synchron, true otherwise.
4847	*/
4848	static bool supdrvDetermineAsyncTsc(uint64_t *poffMin)
4849	{
4850	/*
4851	* Just iterate all the cpus 8 times and make sure that the TSC is
4852	* ever increasing. We don't bother taking TSC rollover into account.
4853	*/
4854	RTCPUSET CpuSet;
4855	int iLastCpu = RTCpuLastIndex(RTMpGetSet(&CpuSet));
4856	int iCpu;
4857	int cLoops = 8;
4858	bool fAsync = false;
4859	int rc = VINF_SUCCESS;
4860	uint64_t offMax = 0;
4861	uint64_t offMin = ~(uint64_t)0;
4862	uint64_t PrevTsc = ASMReadTSC();
4863
4864	while (cLoops-- > 0)
4865	{
4866	for (iCpu = 0; iCpu <= iLastCpu; iCpu++)
4867	{
4868	uint64_t CurTsc;
4869	rc = RTMpOnSpecific(RTMpCpuIdFromSetIndex(iCpu), supdrvDetermineAsyncTscWorker, &CurTsc, NULL);
4870	if (RT_SUCCESS(rc))
4871	{
4872	if (CurTsc <= PrevTsc)
4873	{
4874	fAsync = true;
4875	offMin = offMax = PrevTsc - CurTsc;
4876	Log(("supdrvDetermineAsyncTsc: iCpu=%d cLoops=%d CurTsc=%llx PrevTsc=%llx\n",
4877	iCpu, cLoops, CurTsc, PrevTsc));
4878	break;
4879	}
4880
4881	/* Gather statistics (except the first time). */
4882	if (iCpu != 0 \|\| cLoops != 7)
4883	{
4884	uint64_t off = CurTsc - PrevTsc;
4885	if (off < offMin)
4886	offMin = off;
4887	if (off > offMax)
4888	offMax = off;
4889	Log2(("%d/%d: off=%llx\n", cLoops, iCpu, off));
4890	}
4891
4892	/* Next */
4893	PrevTsc = CurTsc;
4894	}
4895	else if (rc == VERR_NOT_SUPPORTED)
4896	break;
4897	else
4898	AssertMsg(rc == VERR_CPU_NOT_FOUND \|\| rc == VERR_CPU_OFFLINE, ("%d\n", rc));
4899	}
4900
4901	/* broke out of the loop. */
4902	if (iCpu <= iLastCpu)
4903	break;
4904	}
4905
4906	poffMin = offMin; / Almost RTMpOnSpecific profiling. */
4907	Log(("supdrvDetermineAsyncTsc: returns %d; iLastCpu=%d rc=%d offMin=%llx offMax=%llx\n",
4908	fAsync, iLastCpu, rc, offMin, offMax));
4909	#if !defined(RT_OS_SOLARIS) && !defined(RT_OS_OS2) && !defined(RT_OS_WINDOWS)
4910	OSDBGPRINT(("vboxdrv: fAsync=%d offMin=%#lx offMax=%#lx\n", fAsync, (long)offMin, (long)offMax));
4911	#endif
4912	return fAsync;
4913	}
4914
4915
4916	/**
4917	* Determin the GIP TSC mode.
4918	*
4919	* @returns The most suitable TSC mode.
4920	* @param pDevExt Pointer to the device instance data.
4921	*/
4922	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt)
4923	{
4924	/*
4925	* On SMP we're faced with two problems:
4926	* (1) There might be a skew between the CPU, so that cpu0
4927	* returns a TSC that is sligtly different from cpu1.
4928	* (2) Power management (and other things) may cause the TSC
4929	* to run at a non-constant speed, and cause the speed
4930	* to be different on the cpus. This will result in (1).
4931	*
4932	* So, on SMP systems we'll have to select the ASYNC update method
4933	* if there are symphoms of these problems.
4934	*/
4935	if (RTMpGetCount() > 1)
4936	{
4937	uint32_t uEAX, uEBX, uECX, uEDX;
4938	uint64_t u64DiffCoresIgnored;
4939
4940	/* Permit the user and/or the OS specfic bits to force async mode. */
4941	if (supdrvOSGetForcedAsyncTscMode(pDevExt))
4942	return SUPGIPMODE_ASYNC_TSC;
4943
4944	/* Try check for current differences between the cpus. */
4945	if (supdrvDetermineAsyncTsc(&u64DiffCoresIgnored))
4946	return SUPGIPMODE_ASYNC_TSC;
4947
4948	/*
4949	* If the CPU supports power management and is an AMD one we
4950	* won't trust it unless it has the TscInvariant bit is set.
4951	*/
4952	/* Check for "AuthenticAMD" */
4953	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
4954	if ( uEAX >= 1
4955	&& uEBX == X86_CPUID_VENDOR_AMD_EBX
4956	&& uECX == X86_CPUID_VENDOR_AMD_ECX
4957	&& uEDX == X86_CPUID_VENDOR_AMD_EDX)
4958	{
4959	/* Check for APM support and that TscInvariant is cleared. */
4960	ASMCpuId(0x80000000, &uEAX, &uEBX, &uECX, &uEDX);
4961	if (uEAX >= 0x80000007)
4962	{
4963	ASMCpuId(0x80000007, &uEAX, &uEBX, &uECX, &uEDX);
4964	if ( !(uEDX & RT_BIT(8))/* TscInvariant */
4965	&& (uEDX & 0x3e)) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
4966	return SUPGIPMODE_ASYNC_TSC;
4967	}
4968	}
4969	}
4970	return SUPGIPMODE_SYNC_TSC;
4971	}
4972
4973
4974
4975	/**
4976	* Initializes the GIP data.
4977	*
4978	* @param pDevExt Pointer to the device instance data.
4979	* @param pGip Pointer to the read-write kernel mapping of the GIP.
4980	* @param HCPhys The physical address of the GIP.
4981	* @param u64NanoTS The current nanosecond timestamp.
4982	* @param uUpdateHz The update freqence.
4983	*/
4984	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys, uint64_t u64NanoTS, unsigned uUpdateHz)
4985	{
4986	unsigned i;
4987	#ifdef DEBUG_DARWIN_GIP
4988	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz));
4989	#else
4990	LogFlow(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz));
4991	#endif
4992
4993	/*
4994	* Initialize the structure.
4995	*/
4996	memset(pGip, 0, PAGE_SIZE);
4997	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
4998	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
4999	pGip->u32Mode = supdrvGipDeterminTscMode(pDevExt);
5000	pGip->u32UpdateHz = uUpdateHz;
5001	pGip->u32UpdateIntervalNS = 1000000000 / uUpdateHz;
5002	pGip->u64NanoTSLastUpdateHz = u64NanoTS;
5003
5004	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
5005	{
5006	pGip->aCPUs[i].u32TransactionId = 2;
5007	pGip->aCPUs[i].u64NanoTS = u64NanoTS;
5008	pGip->aCPUs[i].u64TSC = ASMReadTSC();
5009
5010	/*
5011	* We don't know the following values until we've executed updates.
5012	* So, we'll just pretend it's a 4 GHz CPU and adjust the history it on
5013	* the 2nd timer callout.
5014	*/
5015	pGip->aCPUs[i].u64CpuHz = _4G + 1; /* tstGIP-2 depends on this. */
5016	pGip->aCPUs[i].u32UpdateIntervalTSC
5017	= pGip->aCPUs[i].au32TSCHistory[0]
5018	= pGip->aCPUs[i].au32TSCHistory[1]
5019	= pGip->aCPUs[i].au32TSCHistory[2]
5020	= pGip->aCPUs[i].au32TSCHistory[3]
5021	= pGip->aCPUs[i].au32TSCHistory[4]
5022	= pGip->aCPUs[i].au32TSCHistory[5]
5023	= pGip->aCPUs[i].au32TSCHistory[6]
5024	= pGip->aCPUs[i].au32TSCHistory[7]
5025	= /pGip->aCPUs[i].u64CpuHz/ _4G / uUpdateHz;
5026	}
5027
5028	/*
5029	* Link it to the device extension.
5030	*/
5031	pDevExt->pGip = pGip;
5032	pDevExt->HCPhysGip = HCPhys;
5033	pDevExt->cGipUsers = 0;
5034	}
5035
5036
5037	/**
5038	* Invalidates the GIP data upon termination.
5039	*
5040	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5041	*/
5042	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
5043	{
5044	unsigned i;
5045	pGip->u32Magic = 0;
5046	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
5047	{
5048	pGip->aCPUs[i].u64NanoTS = 0;
5049	pGip->aCPUs[i].u64TSC = 0;
5050	pGip->aCPUs[i].iTSCHistoryHead = 0;
5051	}
5052	}
5053
5054
5055	/**
5056	* Worker routine for supdrvGipUpdate and supdrvGipUpdatePerCpu that
5057	* updates all the per cpu data except the transaction id.
5058	*
5059	* @param pGip The GIP.
5060	* @param pGipCpu Pointer to the per cpu data.
5061	* @param u64NanoTS The current time stamp.
5062	* @param u64TSC The current TSC.
5063	* @param iTick The current timer tick.
5064	*/
5065	static void supdrvGipDoUpdateCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick)
5066	{
5067	uint64_t u64TSCDelta;
5068	uint32_t u32UpdateIntervalTSC;
5069	uint32_t u32UpdateIntervalTSCSlack;
5070	unsigned iTSCHistoryHead;
5071	uint64_t u64CpuHz;
5072	uint32_t u32TransactionId;
5073
5074	/* Delta between this and the previous update. */
5075	ASMAtomicUoWriteU32(&pGipCpu->u32PrevUpdateIntervalNS, (uint32_t)(u64NanoTS - pGipCpu->u64NanoTS));
5076
5077	/*
5078	* Update the NanoTS.
5079	*/
5080	ASMAtomicXchgU64(&pGipCpu->u64NanoTS, u64NanoTS);
5081
5082	/*
5083	* Calc TSC delta.
5084	*/
5085	/** @todo validate the NanoTS delta, don't trust the OS to call us when it should... */
5086	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
5087	ASMAtomicXchgU64(&pGipCpu->u64TSC, u64TSC);
5088
5089	if (u64TSCDelta >> 32)
5090	{
5091	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
5092	pGipCpu->cErrors++;
5093	}
5094
5095	/*
5096	* On the 2nd and 3rd callout, reset the history with the current TSC
5097	* interval since the values entered by supdrvGipInit are totally off.
5098	* The interval on the 1st callout completely unreliable, the 2nd is a bit
5099	* better, while the 3rd should be most reliable.
5100	*/
5101	u32TransactionId = pGipCpu->u32TransactionId;
5102	if (RT_UNLIKELY( ( u32TransactionId == 5
5103	\|\| u32TransactionId == 7)
5104	&& ( iTick == 2
5105	\|\| iTick == 3) ))
5106	{
5107	unsigned i;
5108	for (i = 0; i < RT_ELEMENTS(pGipCpu->au32TSCHistory); i++)
5109	ASMAtomicUoWriteU32(&pGipCpu->au32TSCHistory[i], (uint32_t)u64TSCDelta);
5110	}
5111
5112	/*
5113	* TSC History.
5114	*/
5115	Assert(RT_ELEMENTS(pGipCpu->au32TSCHistory) == 8);
5116	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
5117	ASMAtomicXchgU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
5118	ASMAtomicXchgU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
5119
5120	/*
5121	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
5122	*/
5123	if (pGip->u32UpdateHz >= 1000)
5124	{
5125	uint32_t u32;
5126	u32 = pGipCpu->au32TSCHistory[0];
5127	u32 += pGipCpu->au32TSCHistory[1];
5128	u32 += pGipCpu->au32TSCHistory[2];
5129	u32 += pGipCpu->au32TSCHistory[3];
5130	u32 >>= 2;
5131	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
5132	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
5133	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
5134	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
5135	u32UpdateIntervalTSC >>= 2;
5136	u32UpdateIntervalTSC += u32;
5137	u32UpdateIntervalTSC >>= 1;
5138
5139	/* Value choosen for a 2GHz Athlon64 running linux 2.6.10/11, . */
5140	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
5141	}
5142	else if (pGip->u32UpdateHz >= 90)
5143	{
5144	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5145	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
5146	u32UpdateIntervalTSC >>= 1;
5147
5148	/* value choosen on a 2GHz thinkpad running windows */
5149	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
5150	}
5151	else
5152	{
5153	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5154
5155	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
5156	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
5157	}
5158	ASMAtomicXchgU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
5159
5160	/*
5161	* CpuHz.
5162	*/
5163	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, pGip->u32UpdateHz);
5164	ASMAtomicXchgU64(&pGipCpu->u64CpuHz, u64CpuHz);
5165	}
5166
5167
5168	/**
5169	* Updates the GIP.
5170	*
5171	* @param pGip Pointer to the GIP.
5172	* @param u64NanoTS The current nanosecond timesamp.
5173	* @param u64TSC The current TSC timesamp.
5174	* @param iTick The current timer tick.
5175	*/
5176	static void supdrvGipUpdate(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick)
5177	{
5178	/*
5179	* Determin the relevant CPU data.
5180	*/
5181	PSUPGIPCPU pGipCpu;
5182	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
5183	pGipCpu = &pGip->aCPUs[0];
5184	else
5185	{
5186	unsigned iCpu = ASMGetApicId();
5187	if (RT_UNLIKELY(iCpu >= RT_ELEMENTS(pGip->aCPUs)))
5188	return;
5189	pGipCpu = &pGip->aCPUs[iCpu];
5190	}
5191
5192	/*
5193	* Start update transaction.
5194	*/
5195	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5196	{
5197	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
5198	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5199	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5200	pGipCpu->cErrors++;
5201	return;
5202	}
5203
5204	/*
5205	* Recalc the update frequency every 0x800th time.
5206	*/
5207	if (!(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
5208	{
5209	if (pGip->u64NanoTSLastUpdateHz)
5210	{
5211	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
5212	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
5213	uint32_t u32UpdateHz = (uint32_t)((UINT64_C(1000000000) * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
5214	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
5215	{
5216	ASMAtomicXchgU32(&pGip->u32UpdateHz, u32UpdateHz);
5217	ASMAtomicXchgU32(&pGip->u32UpdateIntervalNS, 1000000000 / u32UpdateHz);
5218	}
5219	#endif
5220	}
5221	ASMAtomicXchgU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS);
5222	}
5223
5224	/*
5225	* Update the data.
5226	*/
5227	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS, u64TSC, iTick);
5228
5229	/*
5230	* Complete transaction.
5231	*/
5232	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5233	}
5234
5235
5236	/**
5237	* Updates the per cpu GIP data for the calling cpu.
5238	*
5239	* @param pGip Pointer to the GIP.
5240	* @param u64NanoTS The current nanosecond timesamp.
5241	* @param u64TSC The current TSC timesamp.
5242	* @param iCpu The CPU index.
5243	* @param iTick The current timer tick.
5244	*/
5245	static void supdrvGipUpdatePerCpu(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC, unsigned iCpu, uint64_t iTick)
5246	{
5247	PSUPGIPCPU pGipCpu;
5248
5249	if (RT_LIKELY(iCpu < RT_ELEMENTS(pGip->aCPUs)))
5250	{
5251	pGipCpu = &pGip->aCPUs[iCpu];
5252
5253	/*
5254	* Start update transaction.
5255	*/
5256	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5257	{
5258	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5259	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5260	pGipCpu->cErrors++;
5261	return;
5262	}
5263
5264	/*
5265	* Update the data.
5266	*/
5267	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS, u64TSC, iTick);
5268
5269	/*
5270	* Complete transaction.
5271	*/
5272	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5273	}
5274	}
5275

注意: 瀏覽 TracBrowser 來幫助您使用儲存庫瀏覽器

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDrv.c@ 28800

以其他格式下載: