SUPDrv.c@ 29709

Last change on this file since 29709 was 29250, checked in by vboxsync, 15 years ago
iprt/asm.h: split out asm-math.h, don't include asm-.h from asm.h, don't include asm.h from sup.h. Fixed a couple file headers.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 196.5 KB

Line
1	/* $Revision: 29250 $ */
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.virtualbox.org. This file is free software;
11	* you can redistribute it and/or modify it under the terms of the GNU
12	* General Public License (GPL) as published by the Free Software
13	* Foundation, in version 2 as it comes in the "COPYING" file of the
14	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16	*
17	* The contents of this file may alternatively be used under the terms
18	* of the Common Development and Distribution License Version 1.0
19	* (CDDL) only, as it comes in the "COPYING.CDDL" file of the
20	* VirtualBox OSE distribution, in which case the provisions of the
21	* CDDL are applicable instead of those of the GPL.
22	*
23	* You may elect to license modified versions of this file under the
24	* terms and conditions of either the GPL or the CDDL or both.
25	*/
26
27	/*******************************************************************************
28	* 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/asm.h>
37	#include <iprt/asm-amd64-x86.h>
38	#include <iprt/asm-math.h>
39	#include <iprt/cpuset.h>
40	#include <iprt/handletable.h>
41	#include <iprt/mem.h>
42	#include <iprt/mp.h>
43	#include <iprt/power.h>
44	#include <iprt/process.h>
45	#include <iprt/semaphore.h>
46	#include <iprt/spinlock.h>
47	#include <iprt/thread.h>
48	#include <iprt/uuid.h>
49	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD) \|\| defined(RT_OS_WINDOWS)
50	# include <iprt/net.h>
51	# include <iprt/crc32.h>
52	# include <iprt/string.h>
53	# include <iprt/rand.h>
54	# include <iprt/path.h>
55	#endif
56
57	#include <VBox/param.h>
58	#include <VBox/log.h>
59	#include <VBox/err.h>
60	#include <VBox/hwacc_svm.h>
61	#include <VBox/hwacc_vmx.h>
62	#include <VBox/x86.h>
63
64	/*
65	* Logging assignments:
66	* Log - useful stuff, like failures.
67	* LogFlow - program flow, except the really noisy bits.
68	* Log2 - Cleanup.
69	* Log3 - Loader flow noise.
70	* Log4 - Call VMMR0 flow noise.
71	* Log5 - Native yet-to-be-defined noise.
72	* Log6 - Native ioctl flow noise.
73	*
74	* Logging requires BUILD_TYPE=debug and possibly changes to the logger
75	* instanciation in log-vbox.c(pp).
76	*/
77
78
79	/*******************************************************************************
80	* Defined Constants And Macros *
81	*******************************************************************************/
82	/** The frequency by which we recalculate the u32UpdateHz and
83	* u32UpdateIntervalNS GIP members. The value must be a power of 2. */
84	#define GIP_UPDATEHZ_RECALC_FREQ 0x800
85
86	/** @def VBOX_SVN_REV
87	* The makefile should define this if it can. */
88	#ifndef VBOX_SVN_REV
89	# define VBOX_SVN_REV 0
90	#endif
91
92
93	/*******************************************************************************
94	* Internal Functions *
95	*******************************************************************************/
96	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser);
97	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser);
98	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession);
99	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType);
100	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq);
101	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq);
102	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq);
103	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq);
104	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq);
105	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx);
106	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt);
107	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage);
108	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage);
109	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt);
110	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt);
111	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq);
112	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq);
113	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt);
114	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt);
115	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
116	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
117	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser);
118	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys, uint64_t u64NanoTS, unsigned uUpdateHz);
119	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip);
120	static void supdrvGipUpdate(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick);
121	static void supdrvGipUpdatePerCpu(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC, unsigned iCpu, uint64_t iTick);
122
123
124	/*******************************************************************************
125	* Global Variables *
126	*******************************************************************************/
127	DECLEXPORT(PSUPGLOBALINFOPAGE) g_pSUPGlobalInfoPage = NULL;
128
129	/**
130	* Array of the R0 SUP API.
131	*/
132	static SUPFUNC g_aFunctions[] =
133	{
134	/* name function */
135	/* Entries with absolute addresses determined at runtime, fixup
136	code makes ugly ASSUMPTIONS about the order here: */
137	{ "SUPR0AbsIs64bit", (void *)0 },
138	{ "SUPR0Abs64bitKernelCS", (void *)0 },
139	{ "SUPR0Abs64bitKernelSS", (void *)0 },
140	{ "SUPR0Abs64bitKernelDS", (void *)0 },
141	{ "SUPR0AbsKernelCS", (void *)0 },
142	{ "SUPR0AbsKernelSS", (void *)0 },
143	{ "SUPR0AbsKernelDS", (void *)0 },
144	{ "SUPR0AbsKernelES", (void *)0 },
145	{ "SUPR0AbsKernelFS", (void *)0 },
146	{ "SUPR0AbsKernelGS", (void *)0 },
147	/* Normal function pointers: */
148	{ "SUPR0ComponentRegisterFactory", (void *)SUPR0ComponentRegisterFactory },
149	{ "SUPR0ComponentDeregisterFactory", (void *)SUPR0ComponentDeregisterFactory },
150	{ "SUPR0ComponentQueryFactory", (void *)SUPR0ComponentQueryFactory },
151	{ "SUPR0ObjRegister", (void *)SUPR0ObjRegister },
152	{ "SUPR0ObjAddRef", (void *)SUPR0ObjAddRef },
153	{ "SUPR0ObjAddRefEx", (void *)SUPR0ObjAddRefEx },
154	{ "SUPR0ObjRelease", (void *)SUPR0ObjRelease },
155	{ "SUPR0ObjVerifyAccess", (void *)SUPR0ObjVerifyAccess },
156	{ "SUPR0LockMem", (void *)SUPR0LockMem },
157	{ "SUPR0UnlockMem", (void *)SUPR0UnlockMem },
158	{ "SUPR0ContAlloc", (void *)SUPR0ContAlloc },
159	{ "SUPR0ContFree", (void *)SUPR0ContFree },
160	{ "SUPR0LowAlloc", (void *)SUPR0LowAlloc },
161	{ "SUPR0LowFree", (void *)SUPR0LowFree },
162	{ "SUPR0MemAlloc", (void *)SUPR0MemAlloc },
163	{ "SUPR0MemGetPhys", (void *)SUPR0MemGetPhys },
164	{ "SUPR0MemFree", (void *)SUPR0MemFree },
165	{ "SUPR0PageAllocEx", (void *)SUPR0PageAllocEx },
166	{ "SUPR0PageFree", (void *)SUPR0PageFree },
167	{ "SUPR0Printf", (void )SUPR0Printf }, /* @todo needs wrapping? */
168	{ "SUPSemEventCreate", (void *)SUPSemEventCreate },
169	{ "SUPSemEventClose", (void *)SUPSemEventClose },
170	{ "SUPSemEventSignal", (void *)SUPSemEventSignal },
171	{ "SUPSemEventWait", (void *)SUPSemEventWait },
172	{ "SUPSemEventWaitNoResume", (void *)SUPSemEventWaitNoResume },
173	{ "SUPSemEventMultiCreate", (void *)SUPSemEventMultiCreate },
174	{ "SUPSemEventMultiClose", (void *)SUPSemEventMultiClose },
175	{ "SUPSemEventMultiSignal", (void *)SUPSemEventMultiSignal },
176	{ "SUPSemEventMultiReset", (void *)SUPSemEventMultiReset },
177	{ "SUPSemEventMultiWait", (void *)SUPSemEventMultiWait },
178	{ "SUPSemEventMultiWaitNoResume", (void *)SUPSemEventMultiWaitNoResume },
179	{ "SUPR0GetPagingMode", (void *)SUPR0GetPagingMode },
180	{ "SUPR0EnableVTx", (void *)SUPR0EnableVTx },
181	{ "SUPGetGIP", (void *)SUPGetGIP },
182	{ "g_pSUPGlobalInfoPage", (void *)&g_pSUPGlobalInfoPage },
183	{ "RTMemAlloc", (void *)RTMemAlloc },
184	{ "RTMemAllocZ", (void *)RTMemAllocZ },
185	{ "RTMemFree", (void *)RTMemFree },
186	/{ "RTMemDup", (void )RTMemDup },
187	{ "RTMemDupEx", (void )RTMemDupEx },/
188	{ "RTMemRealloc", (void *)RTMemRealloc },
189	{ "RTR0MemObjAllocLow", (void *)RTR0MemObjAllocLow },
190	{ "RTR0MemObjAllocPage", (void *)RTR0MemObjAllocPage },
191	{ "RTR0MemObjAllocPhys", (void *)RTR0MemObjAllocPhys },
192	{ "RTR0MemObjAllocPhysEx", (void *)RTR0MemObjAllocPhysEx },
193	{ "RTR0MemObjAllocPhysNC", (void *)RTR0MemObjAllocPhysNC },
194	{ "RTR0MemObjAllocCont", (void *)RTR0MemObjAllocCont },
195	{ "RTR0MemObjEnterPhys", (void *)RTR0MemObjEnterPhys },
196	{ "RTR0MemObjLockUser", (void *)RTR0MemObjLockUser },
197	{ "RTR0MemObjMapKernel", (void *)RTR0MemObjMapKernel },
198	{ "RTR0MemObjMapKernelEx", (void *)RTR0MemObjMapKernelEx },
199	{ "RTR0MemObjMapUser", (void *)RTR0MemObjMapUser },
200	{ "RTR0MemObjProtect", (void *)RTR0MemObjProtect },
201	{ "RTR0MemObjAddress", (void *)RTR0MemObjAddress },
202	{ "RTR0MemObjAddressR3", (void *)RTR0MemObjAddressR3 },
203	{ "RTR0MemObjSize", (void *)RTR0MemObjSize },
204	{ "RTR0MemObjIsMapping", (void *)RTR0MemObjIsMapping },
205	{ "RTR0MemObjGetPagePhysAddr", (void *)RTR0MemObjGetPagePhysAddr },
206	{ "RTR0MemObjFree", (void *)RTR0MemObjFree },
207	{ "RTR0MemUserCopyFrom", (void *)RTR0MemUserCopyFrom },
208	{ "RTR0MemUserCopyTo", (void *)RTR0MemUserCopyTo },
209	{ "RTR0MemUserIsValidAddr", (void *)RTR0MemUserIsValidAddr },
210	{ "RTR0MemKernelIsValidAddr", (void *)RTR0MemKernelIsValidAddr },
211	{ "RTR0MemAreKrnlAndUsrDifferent", (void *)RTR0MemAreKrnlAndUsrDifferent },
212	{ "RTSemMutexCreate", (void *)RTSemMutexCreate },
213	{ "RTSemMutexRequest", (void *)RTSemMutexRequest },
214	{ "RTSemMutexRequestDebug", (void *)RTSemMutexRequestDebug },
215	{ "RTSemMutexRequestNoResume", (void *)RTSemMutexRequestNoResume },
216	{ "RTSemMutexRequestNoResumeDebug", (void *)RTSemMutexRequestNoResumeDebug },
217	{ "RTSemMutexRelease", (void *)RTSemMutexRelease },
218	{ "RTSemMutexDestroy", (void *)RTSemMutexDestroy },
219	{ "RTProcSelf", (void *)RTProcSelf },
220	{ "RTR0ProcHandleSelf", (void *)RTR0ProcHandleSelf },
221	{ "RTSemFastMutexCreate", (void *)RTSemFastMutexCreate },
222	{ "RTSemFastMutexDestroy", (void *)RTSemFastMutexDestroy },
223	{ "RTSemFastMutexRequest", (void *)RTSemFastMutexRequest },
224	{ "RTSemFastMutexRelease", (void *)RTSemFastMutexRelease },
225	{ "RTSemEventCreate", (void *)RTSemEventCreate },
226	{ "RTSemEventSignal", (void *)RTSemEventSignal },
227	{ "RTSemEventWait", (void *)RTSemEventWait },
228	{ "RTSemEventWaitNoResume", (void *)RTSemEventWaitNoResume },
229	{ "RTSemEventDestroy", (void *)RTSemEventDestroy },
230	{ "RTSemEventMultiCreate", (void *)RTSemEventMultiCreate },
231	{ "RTSemEventMultiSignal", (void *)RTSemEventMultiSignal },
232	{ "RTSemEventMultiReset", (void *)RTSemEventMultiReset },
233	{ "RTSemEventMultiWait", (void *)RTSemEventMultiWait },
234	{ "RTSemEventMultiWaitNoResume", (void *)RTSemEventMultiWaitNoResume },
235	{ "RTSemEventMultiDestroy", (void *)RTSemEventMultiDestroy },
236	{ "RTSpinlockCreate", (void *)RTSpinlockCreate },
237	{ "RTSpinlockDestroy", (void *)RTSpinlockDestroy },
238	{ "RTSpinlockAcquire", (void *)RTSpinlockAcquire },
239	{ "RTSpinlockRelease", (void *)RTSpinlockRelease },
240	{ "RTSpinlockAcquireNoInts", (void *)RTSpinlockAcquireNoInts },
241	{ "RTSpinlockReleaseNoInts", (void *)RTSpinlockReleaseNoInts },
242	{ "RTTimeNanoTS", (void *)RTTimeNanoTS },
243	{ "RTTimeMilliTS", (void *)RTTimeMilliTS },
244	{ "RTTimeSystemNanoTS", (void *)RTTimeSystemNanoTS },
245	{ "RTTimeSystemMilliTS", (void *)RTTimeSystemMilliTS },
246	{ "RTThreadNativeSelf", (void *)RTThreadNativeSelf },
247	{ "RTThreadSleep", (void *)RTThreadSleep },
248	{ "RTThreadYield", (void *)RTThreadYield },
249	#if 0 /* Thread APIs, Part 2. */
250	{ "RTThreadSelf", (void *)RTThreadSelf },
251	{ "RTThreadCreate", (void *)RTThreadCreate },
252	{ "RTThreadGetNative", (void *)RTThreadGetNative },
253	{ "RTThreadWait", (void *)RTThreadWait },
254	{ "RTThreadWaitNoResume", (void *)RTThreadWaitNoResume },
255	{ "RTThreadGetName", (void *)RTThreadGetName },
256	{ "RTThreadSelfName", (void *)RTThreadSelfName },
257	{ "RTThreadGetType", (void *)RTThreadGetType },
258	{ "RTThreadUserSignal", (void *)RTThreadUserSignal },
259	{ "RTThreadUserReset", (void *)RTThreadUserReset },
260	{ "RTThreadUserWait", (void *)RTThreadUserWait },
261	{ "RTThreadUserWaitNoResume", (void *)RTThreadUserWaitNoResume },
262	#endif
263	{ "RTThreadPreemptIsEnabled", (void *)RTThreadPreemptIsEnabled },
264	{ "RTThreadPreemptIsPending", (void *)RTThreadPreemptIsPending },
265	{ "RTThreadPreemptIsPendingTrusty", (void *)RTThreadPreemptIsPendingTrusty },
266	{ "RTThreadPreemptIsPossible", (void *)RTThreadPreemptIsPossible },
267	{ "RTThreadPreemptDisable", (void *)RTThreadPreemptDisable },
268	{ "RTThreadPreemptRestore", (void *)RTThreadPreemptRestore },
269	{ "RTThreadIsInInterrupt", (void *)RTThreadIsInInterrupt },
270
271	{ "RTLogDefaultInstance", (void *)RTLogDefaultInstance },
272	{ "RTMpCpuId", (void *)RTMpCpuId },
273	{ "RTMpCpuIdFromSetIndex", (void *)RTMpCpuIdFromSetIndex },
274	{ "RTMpCpuIdToSetIndex", (void *)RTMpCpuIdToSetIndex },
275	{ "RTMpIsCpuPossible", (void *)RTMpIsCpuPossible },
276	{ "RTMpGetCount", (void *)RTMpGetCount },
277	{ "RTMpGetMaxCpuId", (void *)RTMpGetMaxCpuId },
278	{ "RTMpGetOnlineCount", (void *)RTMpGetOnlineCount },
279	{ "RTMpGetOnlineSet", (void *)RTMpGetOnlineSet },
280	{ "RTMpGetSet", (void *)RTMpGetSet },
281	{ "RTMpIsCpuOnline", (void *)RTMpIsCpuOnline },
282	{ "RTMpIsCpuWorkPending", (void *)RTMpIsCpuWorkPending },
283	{ "RTMpOnAll", (void *)RTMpOnAll },
284	{ "RTMpOnOthers", (void *)RTMpOnOthers },
285	{ "RTMpOnSpecific", (void *)RTMpOnSpecific },
286	{ "RTMpPokeCpu", (void *)RTMpPokeCpu },
287	{ "RTPowerNotificationRegister", (void *)RTPowerNotificationRegister },
288	{ "RTPowerNotificationDeregister", (void *)RTPowerNotificationDeregister },
289	{ "RTLogRelDefaultInstance", (void *)RTLogRelDefaultInstance },
290	{ "RTLogSetDefaultInstanceThread", (void *)RTLogSetDefaultInstanceThread },
291	{ "RTLogLoggerExV", (void *)RTLogLoggerExV },
292	{ "RTLogPrintfV", (void *)RTLogPrintfV },
293	{ "RTR0AssertPanicSystem", (void *)RTR0AssertPanicSystem },
294	{ "RTAssertMsg1", (void *)RTAssertMsg1 },
295	{ "RTAssertMsg2V", (void *)RTAssertMsg2V },
296	{ "RTAssertSetQuiet", (void *)RTAssertSetQuiet },
297	{ "RTAssertMayPanic", (void *)RTAssertMayPanic },
298	{ "RTAssertSetMayPanic", (void *)RTAssertSetMayPanic },
299	{ "RTAssertAreQuiet", (void *)RTAssertAreQuiet },
300	#ifdef RT_OS_WINDOWS
301	/* We need to include more of our runtime to prevent the dynamically linked R0 modules to get too large. */
302	{ "RTNetIPv4AddDataChecksum", (void *)RTNetIPv4AddDataChecksum },
303	{ "RTNetIPv4AddTCPChecksum", (void *)RTNetIPv4AddTCPChecksum },
304	{ "RTNetIPv4AddUDPChecksum", (void *)RTNetIPv4AddUDPChecksum },
305	{ "RTNetIPv4FinalizeChecksum", (void *)RTNetIPv4FinalizeChecksum },
306	{ "RTNetIPv4HdrChecksum", (void *)RTNetIPv4HdrChecksum },
307	{ "RTNetIPv4IsDHCPValid", (void *)RTNetIPv4IsDHCPValid },
308	{ "RTNetIPv4IsHdrValid", (void *)RTNetIPv4IsHdrValid },
309	{ "RTNetIPv4IsTCPSizeValid", (void *)RTNetIPv4IsTCPSizeValid },
310	{ "RTNetIPv4IsTCPValid", (void *)RTNetIPv4IsTCPValid },
311	{ "RTNetIPv4IsUDPSizeValid", (void *)RTNetIPv4IsUDPSizeValid },
312	{ "RTNetIPv4IsUDPValid", (void *)RTNetIPv4IsUDPValid },
313	{ "RTNetIPv4PseudoChecksum", (void *)RTNetIPv4PseudoChecksum },
314	{ "RTNetIPv4PseudoChecksumBits", (void *)RTNetIPv4PseudoChecksumBits },
315	{ "RTNetIPv4TCPChecksum", (void *)RTNetIPv4TCPChecksum },
316	{ "RTNetIPv4UDPChecksum", (void *)RTNetIPv4UDPChecksum },
317	{ "RTNetIPv6PseudoChecksum", (void *)RTNetIPv6PseudoChecksum },
318	{ "RTNetIPv6PseudoChecksumBits", (void *)RTNetIPv6PseudoChecksumBits },
319	{ "RTNetIPv6PseudoChecksumEx", (void *)RTNetIPv6PseudoChecksumEx },
320	{ "RTNetTCPChecksum", (void *)RTNetTCPChecksum },
321	{ "RTNetUDPChecksum", (void *)RTNetUDPChecksum },
322	{ "RTStrFormat", (void *)RTStrFormat },
323	{ "RTStrFormatNumber", (void *)RTStrFormatNumber },
324	{ "RTStrFormatTypeDeregister", (void *)RTStrFormatTypeDeregister },
325	{ "RTStrFormatTypeRegister", (void *)RTStrFormatTypeRegister },
326	{ "RTStrFormatTypeSetUser", (void *)RTStrFormatTypeSetUser },
327	{ "RTStrFormatV", (void *)RTStrFormatV },
328	{ "RTStrPrintf", (void *)RTStrPrintf },
329	{ "RTStrPrintfEx", (void *)RTStrPrintfEx },
330	{ "RTStrPrintfExV", (void *)RTStrPrintfExV },
331	{ "RTStrPrintfV", (void *)RTStrPrintfV },
332	{ "RTCrc32", (void *)RTCrc32 },
333	{ "RTCrc32Finish", (void *)RTCrc32Finish },
334	{ "RTCrc32Process", (void *)RTCrc32Process },
335	{ "RTCrc32Start", (void *)RTCrc32Start },
336	{ "RTHandleTableAllocWithCtx", (void *)RTHandleTableAllocWithCtx },
337	{ "RTHandleTableCreate", (void *)RTHandleTableCreate },
338	{ "RTHandleTableCreateEx", (void *)RTHandleTableCreateEx },
339	{ "RTHandleTableDestroy", (void *)RTHandleTableDestroy },
340	{ "RTHandleTableFreeWithCtx", (void *)RTHandleTableFreeWithCtx },
341	{ "RTHandleTableLookupWithCtx", (void *)RTHandleTableLookupWithCtx },
342	#endif
343	};
344
345	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
346	/**
347	* Drag in the rest of IRPT since we share it with the
348	* rest of the kernel modules on darwin.
349	*/
350	PFNRT g_apfnVBoxDrvIPRTDeps[] =
351	{
352	/* VBoxNetFlt */
353	(PFNRT)RTCrc32,
354	(PFNRT)RTErrConvertFromErrno,
355	(PFNRT)RTNetIPv4IsHdrValid,
356	(PFNRT)RTNetIPv4TCPChecksum,
357	(PFNRT)RTNetIPv4UDPChecksum,
358	(PFNRT)RTUuidCompare,
359	(PFNRT)RTUuidCompareStr,
360	(PFNRT)RTUuidFromStr,
361	(PFNRT)RTStrDup,
362	(PFNRT)RTStrFree,
363	/* VBoxNetAdp */
364	(PFNRT)RTRandBytes,
365	/* VBoxUSB */
366	(PFNRT)RTPathStripFilename,
367	NULL
368	};
369	#endif /* RT_OS_DARWIN \|\| RT_OS_SOLARIS \|\| RT_OS_SOLARIS */
370
371
372	/**
373	* Initializes the device extentsion structure.
374	*
375	* @returns IPRT status code.
376	* @param pDevExt The device extension to initialize.
377	* @param cbSession The size of the session structure. The size of
378	* SUPDRVSESSION may be smaller when SUPDRV_AGNOSTIC is
379	* defined because we're skipping the OS specific members
380	* then.
381	*/
382	int VBOXCALL supdrvInitDevExt(PSUPDRVDEVEXT pDevExt, size_t cbSession)
383	{
384	int rc;
385
386	#ifdef SUPDRV_WITH_RELEASE_LOGGER
387	/*
388	* Create the release log.
389	*/
390	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
391	PRTLOGGER pRelLogger;
392	rc = RTLogCreate(&pRelLogger, 0 /* fFlags */, "all",
393	"VBOX_RELEASE_LOG", RT_ELEMENTS(s_apszGroups), s_apszGroups,
394	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER, NULL);
395	if (RT_SUCCESS(rc))
396	RTLogRelSetDefaultInstance(pRelLogger);
397	/** @todo Add native hook for getting logger config parameters and setting
398	* them. On linux we should use the module parameter stuff... */
399	#endif
400
401	/*
402	* Initialize it.
403	*/
404	memset(pDevExt, 0, sizeof(*pDevExt));
405	rc = RTSpinlockCreate(&pDevExt->Spinlock);
406	if (RT_SUCCESS(rc))
407	{
408	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
409	rc = RTSemMutexCreate(&pDevExt->mtxLdr);
410	#else
411	rc = RTSemFastMutexCreate(&pDevExt->mtxLdr);
412	#endif
413	if (RT_SUCCESS(rc))
414	{
415	rc = RTSemFastMutexCreate(&pDevExt->mtxComponentFactory);
416	if (RT_SUCCESS(rc))
417	{
418	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
419	rc = RTSemMutexCreate(&pDevExt->mtxGip);
420	#else
421	rc = RTSemFastMutexCreate(&pDevExt->mtxGip);
422	#endif
423	if (RT_SUCCESS(rc))
424	{
425	rc = supdrvGipCreate(pDevExt);
426	if (RT_SUCCESS(rc))
427	{
428	pDevExt->u32Cookie = BIRD; /** @todo make this random? */
429	pDevExt->cbSession = cbSession;
430
431	/*
432	* Fixup the absolute symbols.
433	*
434	* Because of the table indexing assumptions we'll have a little #ifdef orgy
435	* here rather than distributing this to OS specific files. At least for now.
436	*/
437	#ifdef RT_OS_DARWIN
438	# if ARCH_BITS == 32
439	if (SUPR0GetPagingMode() >= SUPPAGINGMODE_AMD64)
440	{
441	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
442	g_aFunctions[1].pfn = (void )0x80; / SUPR0Abs64bitKernelCS - KERNEL64_CS, seg.h */
443	g_aFunctions[2].pfn = (void )0x88; / SUPR0Abs64bitKernelSS - KERNEL64_SS, seg.h */
444	g_aFunctions[3].pfn = (void )0x88; / SUPR0Abs64bitKernelDS - KERNEL64_SS, seg.h */
445	}
446	else
447	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
448	g_aFunctions[4].pfn = (void )0x08; / SUPR0AbsKernelCS - KERNEL_CS, seg.h */
449	g_aFunctions[5].pfn = (void )0x10; / SUPR0AbsKernelSS - KERNEL_DS, seg.h */
450	g_aFunctions[6].pfn = (void )0x10; / SUPR0AbsKernelDS - KERNEL_DS, seg.h */
451	g_aFunctions[7].pfn = (void )0x10; / SUPR0AbsKernelES - KERNEL_DS, seg.h */
452	g_aFunctions[8].pfn = (void )0x10; / SUPR0AbsKernelFS - KERNEL_DS, seg.h */
453	g_aFunctions[9].pfn = (void )0x48; / SUPR0AbsKernelGS - CPU_DATA_GS, seg.h */
454	# else /* 64-bit darwin: */
455	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
456	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
457	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
458	g_aFunctions[3].pfn = (void )0; / SUPR0Abs64bitKernelDS */
459	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
460	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
461	g_aFunctions[6].pfn = (void )0; / SUPR0AbsKernelDS */
462	g_aFunctions[7].pfn = (void )0; / SUPR0AbsKernelES */
463	g_aFunctions[8].pfn = (void )0; / SUPR0AbsKernelFS */
464	g_aFunctions[9].pfn = (void )0; / SUPR0AbsKernelGS */
465
466	# endif
467	#else /* !RT_OS_DARWIN */
468	# if ARCH_BITS == 64
469	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
470	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
471	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
472	g_aFunctions[3].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0Abs64bitKernelDS */
473	# else
474	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
475	# endif
476	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
477	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
478	g_aFunctions[6].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0AbsKernelDS */
479	g_aFunctions[7].pfn = (void )(uintptr_t)ASMGetES(); / SUPR0AbsKernelES */
480	g_aFunctions[8].pfn = (void )(uintptr_t)ASMGetFS(); / SUPR0AbsKernelFS */
481	g_aFunctions[9].pfn = (void )(uintptr_t)ASMGetGS(); / SUPR0AbsKernelGS */
482	#endif /* !RT_OS_DARWIN */
483	return VINF_SUCCESS;
484	}
485
486	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
487	RTSemMutexDestroy(pDevExt->mtxGip);
488	pDevExt->mtxGip = NIL_RTSEMMUTEX;
489	#else
490	RTSemFastMutexDestroy(pDevExt->mtxGip);
491	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
492	#endif
493	}
494	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
495	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
496	}
497	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
498	RTSemMutexDestroy(pDevExt->mtxLdr);
499	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
500	#else
501	RTSemFastMutexDestroy(pDevExt->mtxLdr);
502	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
503	#endif
504	}
505	RTSpinlockDestroy(pDevExt->Spinlock);
506	pDevExt->Spinlock = NIL_RTSPINLOCK;
507	}
508	#ifdef SUPDRV_WITH_RELEASE_LOGGER
509	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
510	RTLogDestroy(RTLogSetDefaultInstance(NULL));
511	#endif
512
513	return rc;
514	}
515
516
517	/**
518	* Delete the device extension (e.g. cleanup members).
519	*
520	* @param pDevExt The device extension to delete.
521	*/
522	void VBOXCALL supdrvDeleteDevExt(PSUPDRVDEVEXT pDevExt)
523	{
524	PSUPDRVOBJ pObj;
525	PSUPDRVUSAGE pUsage;
526
527	/*
528	* Kill mutexes and spinlocks.
529	*/
530	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
531	RTSemMutexDestroy(pDevExt->mtxGip);
532	pDevExt->mtxGip = NIL_RTSEMMUTEX;
533	#else
534	RTSemFastMutexDestroy(pDevExt->mtxGip);
535	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
536	#endif
537	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
538	RTSemMutexDestroy(pDevExt->mtxLdr);
539	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
540	#else
541	RTSemFastMutexDestroy(pDevExt->mtxLdr);
542	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
543	#endif
544	RTSpinlockDestroy(pDevExt->Spinlock);
545	pDevExt->Spinlock = NIL_RTSPINLOCK;
546	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
547	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
548
549	/*
550	* Free lists.
551	*/
552	/* objects. */
553	pObj = pDevExt->pObjs;
554	Assert(!pObj); /* (can trigger on forced unloads) */
555	pDevExt->pObjs = NULL;
556	while (pObj)
557	{
558	void *pvFree = pObj;
559	pObj = pObj->pNext;
560	RTMemFree(pvFree);
561	}
562
563	/* usage records. */
564	pUsage = pDevExt->pUsageFree;
565	pDevExt->pUsageFree = NULL;
566	while (pUsage)
567	{
568	void *pvFree = pUsage;
569	pUsage = pUsage->pNext;
570	RTMemFree(pvFree);
571	}
572
573	/* kill the GIP. */
574	supdrvGipDestroy(pDevExt);
575
576	#ifdef SUPDRV_WITH_RELEASE_LOGGER
577	/* destroy the loggers. */
578	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
579	RTLogDestroy(RTLogSetDefaultInstance(NULL));
580	#endif
581	}
582
583
584	/**
585	* Create session.
586	*
587	* @returns IPRT status code.
588	* @param pDevExt Device extension.
589	* @param fUser Flag indicating whether this is a user or kernel session.
590	* @param ppSession Where to store the pointer to the session data.
591	*/
592	int VBOXCALL supdrvCreateSession(PSUPDRVDEVEXT pDevExt, bool fUser, PSUPDRVSESSION *ppSession)
593	{
594	/*
595	* Allocate memory for the session data.
596	*/
597	int rc;
598	PSUPDRVSESSION pSession = *ppSession = (PSUPDRVSESSION)RTMemAllocZ(pDevExt->cbSession);
599	if (pSession)
600	{
601	/* Initialize session data. */
602	rc = RTSpinlockCreate(&pSession->Spinlock);
603	if (!rc)
604	{
605	rc = RTHandleTableCreateEx(&pSession->hHandleTable,
606	RTHANDLETABLE_FLAGS_LOCKED \| RTHANDLETABLE_FLAGS_CONTEXT,
607	1 /uBase/, 32768 /cMax/, supdrvSessionObjHandleRetain, pSession);
608	if (RT_SUCCESS(rc))
609	{
610	Assert(pSession->Spinlock != NIL_RTSPINLOCK);
611	pSession->pDevExt = pDevExt;
612	pSession->u32Cookie = BIRD_INV;
613	/*pSession->pLdrUsage = NULL;
614	pSession->pVM = NULL;
615	pSession->pUsage = NULL;
616	pSession->pGip = NULL;
617	pSession->fGipReferenced = false;
618	pSession->Bundle.cUsed = 0; */
619	pSession->Uid = NIL_RTUID;
620	pSession->Gid = NIL_RTGID;
621	if (fUser)
622	{
623	pSession->Process = RTProcSelf();
624	pSession->R0Process = RTR0ProcHandleSelf();
625	}
626	else
627	{
628	pSession->Process = NIL_RTPROCESS;
629	pSession->R0Process = NIL_RTR0PROCESS;
630	}
631
632	LogFlow(("Created session %p initial cookie=%#x\n", pSession, pSession->u32Cookie));
633	return VINF_SUCCESS;
634	}
635
636	RTSpinlockDestroy(pSession->Spinlock);
637	}
638	RTMemFree(pSession);
639	*ppSession = NULL;
640	Log(("Failed to create spinlock, rc=%d!\n", rc));
641	}
642	else
643	rc = VERR_NO_MEMORY;
644
645	return rc;
646	}
647
648
649	/**
650	* Shared code for cleaning up a session.
651	*
652	* @param pDevExt Device extension.
653	* @param pSession Session data.
654	* This data will be freed by this routine.
655	*/
656	void VBOXCALL supdrvCloseSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
657	{
658	/*
659	* Cleanup the session first.
660	*/
661	supdrvCleanupSession(pDevExt, pSession);
662
663	/*
664	* Free the rest of the session stuff.
665	*/
666	RTSpinlockDestroy(pSession->Spinlock);
667	pSession->Spinlock = NIL_RTSPINLOCK;
668	pSession->pDevExt = NULL;
669	RTMemFree(pSession);
670	LogFlow(("supdrvCloseSession: returns\n"));
671	}
672
673
674	/**
675	* Shared code for cleaning up a session (but not quite freeing it).
676	*
677	* This is primarily intended for MAC OS X where we have to clean up the memory
678	* stuff before the file handle is closed.
679	*
680	* @param pDevExt Device extension.
681	* @param pSession Session data.
682	* This data will be freed by this routine.
683	*/
684	void VBOXCALL supdrvCleanupSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
685	{
686	int rc;
687	PSUPDRVBUNDLE pBundle;
688	LogFlow(("supdrvCleanupSession: pSession=%p\n", pSession));
689
690	/*
691	* Remove logger instances related to this session.
692	*/
693	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pSession);
694
695	/*
696	* Destroy the handle table.
697	*/
698	rc = RTHandleTableDestroy(pSession->hHandleTable, supdrvSessionObjHandleDelete, pSession);
699	AssertRC(rc);
700	pSession->hHandleTable = NIL_RTHANDLETABLE;
701
702	/*
703	* Release object references made in this session.
704	* In theory there should be noone racing us in this session.
705	*/
706	Log2(("release objects - start\n"));
707	if (pSession->pUsage)
708	{
709	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
710	PSUPDRVUSAGE pUsage;
711	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
712
713	while ((pUsage = pSession->pUsage) != NULL)
714	{
715	PSUPDRVOBJ pObj = pUsage->pObj;
716	pSession->pUsage = pUsage->pNext;
717
718	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
719	if (pUsage->cUsage < pObj->cUsage)
720	{
721	pObj->cUsage -= pUsage->cUsage;
722	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
723	}
724	else
725	{
726	/* Destroy the object and free the record. */
727	if (pDevExt->pObjs == pObj)
728	pDevExt->pObjs = pObj->pNext;
729	else
730	{
731	PSUPDRVOBJ pObjPrev;
732	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
733	if (pObjPrev->pNext == pObj)
734	{
735	pObjPrev->pNext = pObj->pNext;
736	break;
737	}
738	Assert(pObjPrev);
739	}
740	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
741
742	Log(("supdrvCleanupSession: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
743	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
744	if (pObj->pfnDestructor)
745	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
746	RTMemFree(pObj);
747	}
748
749	/* free it and continue. */
750	RTMemFree(pUsage);
751
752	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
753	}
754
755	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
756	AssertMsg(!pSession->pUsage, ("Some buster reregistered an object during desturction!\n"));
757	}
758	Log2(("release objects - done\n"));
759
760	/*
761	* Release memory allocated in the session.
762	*
763	* We do not serialize this as we assume that the application will
764	* not allocated memory while closing the file handle object.
765	*/
766	Log2(("freeing memory:\n"));
767	pBundle = &pSession->Bundle;
768	while (pBundle)
769	{
770	PSUPDRVBUNDLE pToFree;
771	unsigned i;
772
773	/*
774	* Check and unlock all entries in the bundle.
775	*/
776	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
777	{
778	if (pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ)
779	{
780	Log2(("eType=%d pvR0=%p pvR3=%p cb=%ld\n", pBundle->aMem[i].eType, RTR0MemObjAddress(pBundle->aMem[i].MemObj),
781	(void *)RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3), (long)RTR0MemObjSize(pBundle->aMem[i].MemObj)));
782	if (pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ)
783	{
784	rc = RTR0MemObjFree(pBundle->aMem[i].MapObjR3, false);
785	AssertRC(rc); /** @todo figure out how to handle this. */
786	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
787	}
788	rc = RTR0MemObjFree(pBundle->aMem[i].MemObj, true /* fFreeMappings */);
789	AssertRC(rc); /** @todo figure out how to handle this. */
790	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
791	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
792	}
793	}
794
795	/*
796	* Advance and free previous bundle.
797	*/
798	pToFree = pBundle;
799	pBundle = pBundle->pNext;
800
801	pToFree->pNext = NULL;
802	pToFree->cUsed = 0;
803	if (pToFree != &pSession->Bundle)
804	RTMemFree(pToFree);
805	}
806	Log2(("freeing memory - done\n"));
807
808	/*
809	* Deregister component factories.
810	*/
811	RTSemFastMutexRequest(pDevExt->mtxComponentFactory);
812	Log2(("deregistering component factories:\n"));
813	if (pDevExt->pComponentFactoryHead)
814	{
815	PSUPDRVFACTORYREG pPrev = NULL;
816	PSUPDRVFACTORYREG pCur = pDevExt->pComponentFactoryHead;
817	while (pCur)
818	{
819	if (pCur->pSession == pSession)
820	{
821	/* unlink it */
822	PSUPDRVFACTORYREG pNext = pCur->pNext;
823	if (pPrev)
824	pPrev->pNext = pNext;
825	else
826	pDevExt->pComponentFactoryHead = pNext;
827
828	/* free it */
829	pCur->pNext = NULL;
830	pCur->pSession = NULL;
831	pCur->pFactory = NULL;
832	RTMemFree(pCur);
833
834	/* next */
835	pCur = pNext;
836	}
837	else
838	{
839	/* next */
840	pPrev = pCur;
841	pCur = pCur->pNext;
842	}
843	}
844	}
845	RTSemFastMutexRelease(pDevExt->mtxComponentFactory);
846	Log2(("deregistering component factories - done\n"));
847
848	/*
849	* Loaded images needs to be dereferenced and possibly freed up.
850	*/
851	supdrvLdrLock(pDevExt);
852	Log2(("freeing images:\n"));
853	if (pSession->pLdrUsage)
854	{
855	PSUPDRVLDRUSAGE pUsage = pSession->pLdrUsage;
856	pSession->pLdrUsage = NULL;
857	while (pUsage)
858	{
859	void *pvFree = pUsage;
860	PSUPDRVLDRIMAGE pImage = pUsage->pImage;
861	if (pImage->cUsage > pUsage->cUsage)
862	pImage->cUsage -= pUsage->cUsage;
863	else
864	supdrvLdrFree(pDevExt, pImage);
865	pUsage->pImage = NULL;
866	pUsage = pUsage->pNext;
867	RTMemFree(pvFree);
868	}
869	}
870	supdrvLdrUnlock(pDevExt);
871	Log2(("freeing images - done\n"));
872
873	/*
874	* Unmap the GIP.
875	*/
876	Log2(("umapping GIP:\n"));
877	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
878	{
879	SUPR0GipUnmap(pSession);
880	pSession->fGipReferenced = 0;
881	}
882	Log2(("umapping GIP - done\n"));
883	}
884
885
886	/**
887	* RTHandleTableDestroy callback used by supdrvCleanupSession.
888	*
889	* @returns IPRT status code, see SUPR0ObjAddRef.
890	* @param hHandleTable The handle table handle. Ignored.
891	* @param pvObj The object pointer.
892	* @param pvCtx Context, the handle type. Ignored.
893	* @param pvUser Session pointer.
894	*/
895	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser)
896	{
897	NOREF(pvCtx);
898	NOREF(hHandleTable);
899	return SUPR0ObjAddRefEx(pvObj, (PSUPDRVSESSION)pvUser, true /fNoBlocking/);
900	}
901
902
903	/**
904	* RTHandleTableDestroy callback used by supdrvCleanupSession.
905	*
906	* @param hHandleTable The handle table handle. Ignored.
907	* @param h The handle value. Ignored.
908	* @param pvObj The object pointer.
909	* @param pvCtx Context, the handle type. Ignored.
910	* @param pvUser Session pointer.
911	*/
912	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser)
913	{
914	NOREF(pvCtx);
915	NOREF(h);
916	NOREF(hHandleTable);
917	SUPR0ObjRelease(pvObj, (PSUPDRVSESSION)pvUser);
918	}
919
920
921	/**
922	* Fast path I/O Control worker.
923	*
924	* @returns VBox status code that should be passed down to ring-3 unchanged.
925	* @param uIOCtl Function number.
926	* @param idCpu VMCPU id.
927	* @param pDevExt Device extention.
928	* @param pSession Session data.
929	*/
930	int VBOXCALL supdrvIOCtlFast(uintptr_t uIOCtl, VMCPUID idCpu, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
931	{
932	/*
933	* We check the two prereqs after doing this only to allow the compiler to optimize things better.
934	*/
935	if (RT_LIKELY(pSession->pVM && pDevExt->pfnVMMR0EntryFast))
936	{
937	switch (uIOCtl)
938	{
939	case SUP_IOCTL_FAST_DO_RAW_RUN:
940	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_RAW_RUN);
941	break;
942	case SUP_IOCTL_FAST_DO_HWACC_RUN:
943	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_HWACC_RUN);
944	break;
945	case SUP_IOCTL_FAST_DO_NOP:
946	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_NOP);
947	break;
948	default:
949	return VERR_INTERNAL_ERROR;
950	}
951	return VINF_SUCCESS;
952	}
953	return VERR_INTERNAL_ERROR;
954	}
955
956
957	/**
958	* Helper for supdrvIOCtl. Check if pszStr contains any character of pszChars.
959	* We would use strpbrk here if this function would be contained in the RedHat kABI white
960	* list, see http://www.kerneldrivers.org/RHEL5.
961	*
962	* @returns 1 if pszStr does contain any character of pszChars, 0 otherwise.
963	* @param pszStr String to check
964	* @param pszChars Character set
965	*/
966	static int supdrvCheckInvalidChar(const char pszStr, const char pszChars)
967	{
968	int chCur;
969	while ((chCur = *pszStr++) != '\0')
970	{
971	int ch;
972	const char *psz = pszChars;
973	while ((ch = *psz++) != '\0')
974	if (ch == chCur)
975	return 1;
976
977	}
978	return 0;
979	}
980
981
982	/**
983	* I/O Control worker.
984	*
985	* @returns IPRT status code.
986	* @retval VERR_INVALID_PARAMETER if the request is invalid.
987	*
988	* @param uIOCtl Function number.
989	* @param pDevExt Device extention.
990	* @param pSession Session data.
991	* @param pReqHdr The request header.
992	*/
993	int VBOXCALL supdrvIOCtl(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
994	{
995	/*
996	* Validate the request.
997	*/
998	/* this first check could probably be omitted as its also done by the OS specific code... */
999	if (RT_UNLIKELY( (pReqHdr->fFlags & SUPREQHDR_FLAGS_MAGIC_MASK) != SUPREQHDR_FLAGS_MAGIC
1000	\|\| pReqHdr->cbIn < sizeof(*pReqHdr)
1001	\|\| pReqHdr->cbOut < sizeof(*pReqHdr)))
1002	{
1003	OSDBGPRINT(("vboxdrv: Bad ioctl request header; cbIn=%#lx cbOut=%#lx fFlags=%#lx\n",
1004	(long)pReqHdr->cbIn, (long)pReqHdr->cbOut, (long)pReqHdr->fFlags));
1005	return VERR_INVALID_PARAMETER;
1006	}
1007	if (RT_UNLIKELY(uIOCtl == SUP_IOCTL_COOKIE))
1008	{
1009	if (pReqHdr->u32Cookie != SUPCOOKIE_INITIAL_COOKIE)
1010	{
1011	OSDBGPRINT(("SUP_IOCTL_COOKIE: bad cookie %#lx\n", (long)pReqHdr->u32Cookie));
1012	return VERR_INVALID_PARAMETER;
1013	}
1014	}
1015	else if (RT_UNLIKELY( pReqHdr->u32Cookie != pDevExt->u32Cookie
1016	\|\| pReqHdr->u32SessionCookie != pSession->u32Cookie))
1017	{
1018	OSDBGPRINT(("vboxdrv: bad cookie %#lx / %#lx.\n", (long)pReqHdr->u32Cookie, (long)pReqHdr->u32SessionCookie));
1019	return VERR_INVALID_PARAMETER;
1020	}
1021
1022	/*
1023	* Validation macros
1024	*/
1025	#define REQ_CHECK_SIZES_EX(Name, cbInExpect, cbOutExpect) \
1026	do { \
1027	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect) \|\| pReqHdr->cbOut != (cbOutExpect))) \
1028	{ \
1029	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld. cbOut=%ld expected %ld.\n", \
1030	(long)pReq->Hdr.cbIn, (long)(cbInExpect), (long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
1031	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1032	} \
1033	} while (0)
1034
1035	#define REQ_CHECK_SIZES(Name) REQ_CHECK_SIZES_EX(Name, Name ## _SIZE_IN, Name ## _SIZE_OUT)
1036
1037	#define REQ_CHECK_SIZE_IN(Name, cbInExpect) \
1038	do { \
1039	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect))) \
1040	{ \
1041	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld.\n", \
1042	(long)pReq->Hdr.cbIn, (long)(cbInExpect))); \
1043	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1044	} \
1045	} while (0)
1046
1047	#define REQ_CHECK_SIZE_OUT(Name, cbOutExpect) \
1048	do { \
1049	if (RT_UNLIKELY(pReqHdr->cbOut != (cbOutExpect))) \
1050	{ \
1051	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbOut=%ld expected %ld.\n", \
1052	(long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
1053	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1054	} \
1055	} while (0)
1056
1057	#define REQ_CHECK_EXPR(Name, expr) \
1058	do { \
1059	if (RT_UNLIKELY(!(expr))) \
1060	{ \
1061	OSDBGPRINT(( #Name ": %s\n", #expr)); \
1062	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1063	} \
1064	} while (0)
1065
1066	#define REQ_CHECK_EXPR_FMT(expr, fmt) \
1067	do { \
1068	if (RT_UNLIKELY(!(expr))) \
1069	{ \
1070	OSDBGPRINT( fmt ); \
1071	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1072	} \
1073	} while (0)
1074
1075
1076	/*
1077	* The switch.
1078	*/
1079	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
1080	{
1081	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
1082	{
1083	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
1084	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
1085	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
1086	{
1087	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
1088	pReq->Hdr.rc = VERR_INVALID_MAGIC;
1089	return 0;
1090	}
1091
1092	#if 0
1093	/*
1094	* Call out to the OS specific code and let it do permission checks on the
1095	* client process.
1096	*/
1097	if (!supdrvOSValidateClientProcess(pDevExt, pSession))
1098	{
1099	pReq->u.Out.u32Cookie = 0xffffffff;
1100	pReq->u.Out.u32SessionCookie = 0xffffffff;
1101	pReq->u.Out.u32SessionVersion = 0xffffffff;
1102	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1103	pReq->u.Out.pSession = NULL;
1104	pReq->u.Out.cFunctions = 0;
1105	pReq->Hdr.rc = VERR_PERMISSION_DENIED;
1106	return 0;
1107	}
1108	#endif
1109
1110	/*
1111	* Match the version.
1112	* The current logic is very simple, match the major interface version.
1113	*/
1114	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
1115	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
1116	{
1117	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1118	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
1119	pReq->u.Out.u32Cookie = 0xffffffff;
1120	pReq->u.Out.u32SessionCookie = 0xffffffff;
1121	pReq->u.Out.u32SessionVersion = 0xffffffff;
1122	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1123	pReq->u.Out.pSession = NULL;
1124	pReq->u.Out.cFunctions = 0;
1125	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1126	return 0;
1127	}
1128
1129	/*
1130	* Fill in return data and be gone.
1131	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
1132	* u32SessionVersion <= u32ReqVersion!
1133	*/
1134	/** @todo Somehow validate the client and negotiate a secure cookie... */
1135	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
1136	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
1137	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
1138	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1139	pReq->u.Out.pSession = pSession;
1140	pReq->u.Out.cFunctions = sizeof(g_aFunctions) / sizeof(g_aFunctions[0]);
1141	pReq->Hdr.rc = VINF_SUCCESS;
1142	return 0;
1143	}
1144
1145	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_QUERY_FUNCS(0)):
1146	{
1147	/* validate */
1148	PSUPQUERYFUNCS pReq = (PSUPQUERYFUNCS)pReqHdr;
1149	REQ_CHECK_SIZES_EX(SUP_IOCTL_QUERY_FUNCS, SUP_IOCTL_QUERY_FUNCS_SIZE_IN, SUP_IOCTL_QUERY_FUNCS_SIZE_OUT(RT_ELEMENTS(g_aFunctions)));
1150
1151	/* execute */
1152	pReq->u.Out.cFunctions = RT_ELEMENTS(g_aFunctions);
1153	memcpy(&pReq->u.Out.aFunctions[0], g_aFunctions, sizeof(g_aFunctions));
1154	pReq->Hdr.rc = VINF_SUCCESS;
1155	return 0;
1156	}
1157
1158	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_LOCK):
1159	{
1160	/* validate */
1161	PSUPPAGELOCK pReq = (PSUPPAGELOCK)pReqHdr;
1162	REQ_CHECK_SIZE_IN(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_IN);
1163	REQ_CHECK_SIZE_OUT(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_OUT(pReq->u.In.cPages));
1164	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.cPages > 0);
1165	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.pvR3 >= PAGE_SIZE);
1166
1167	/* execute */
1168	pReq->Hdr.rc = SUPR0LockMem(pSession, pReq->u.In.pvR3, pReq->u.In.cPages, &pReq->u.Out.aPages[0]);
1169	if (RT_FAILURE(pReq->Hdr.rc))
1170	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1171	return 0;
1172	}
1173
1174	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_UNLOCK):
1175	{
1176	/* validate */
1177	PSUPPAGEUNLOCK pReq = (PSUPPAGEUNLOCK)pReqHdr;
1178	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_UNLOCK);
1179
1180	/* execute */
1181	pReq->Hdr.rc = SUPR0UnlockMem(pSession, pReq->u.In.pvR3);
1182	return 0;
1183	}
1184
1185	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_ALLOC):
1186	{
1187	/* validate */
1188	PSUPCONTALLOC pReq = (PSUPCONTALLOC)pReqHdr;
1189	REQ_CHECK_SIZES(SUP_IOCTL_CONT_ALLOC);
1190
1191	/* execute */
1192	pReq->Hdr.rc = SUPR0ContAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.HCPhys);
1193	if (RT_FAILURE(pReq->Hdr.rc))
1194	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1195	return 0;
1196	}
1197
1198	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_FREE):
1199	{
1200	/* validate */
1201	PSUPCONTFREE pReq = (PSUPCONTFREE)pReqHdr;
1202	REQ_CHECK_SIZES(SUP_IOCTL_CONT_FREE);
1203
1204	/* execute */
1205	pReq->Hdr.rc = SUPR0ContFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1206	return 0;
1207	}
1208
1209	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_OPEN):
1210	{
1211	/* validate */
1212	PSUPLDROPEN pReq = (PSUPLDROPEN)pReqHdr;
1213	REQ_CHECK_SIZES(SUP_IOCTL_LDR_OPEN);
1214	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs > 0);
1215	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs < 16*_1M);
1216	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1217	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1218	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits < pReq->u.In.cbImageWithTabs);
1219	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.szName[0]);
1220	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, memchr(pReq->u.In.szName, '\0', sizeof(pReq->u.In.szName)));
1221	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, !supdrvCheckInvalidChar(pReq->u.In.szName, ";:()[]{}/\\\|&*%#@!~`\"'"));
1222	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, memchr(pReq->u.In.szFilename, '\0', sizeof(pReq->u.In.szFilename)));
1223
1224	/* execute */
1225	pReq->Hdr.rc = supdrvIOCtl_LdrOpen(pDevExt, pSession, pReq);
1226	return 0;
1227	}
1228
1229	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOAD):
1230	{
1231	/* validate */
1232	PSUPLDRLOAD pReq = (PSUPLDRLOAD)pReqHdr;
1233	REQ_CHECK_EXPR(Name, pReq->Hdr.cbIn >= sizeof(*pReq));
1234	REQ_CHECK_SIZES_EX(SUP_IOCTL_LDR_LOAD, SUP_IOCTL_LDR_LOAD_SIZE_IN(pReq->u.In.cbImageWithTabs), SUP_IOCTL_LDR_LOAD_SIZE_OUT);
1235	REQ_CHECK_EXPR(SUP_IOCTL_LDR_LOAD, pReq->u.In.cSymbols <= 16384);
1236	REQ_CHECK_EXPR_FMT( !pReq->u.In.cSymbols
1237	\|\| ( pReq->u.In.offSymbols < pReq->u.In.cbImageWithTabs
1238	&& pReq->u.In.offSymbols + pReq->u.In.cSymbols * sizeof(SUPLDRSYM) <= pReq->u.In.cbImageWithTabs),
1239	("SUP_IOCTL_LDR_LOAD: offSymbols=%#lx cSymbols=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offSymbols,
1240	(long)pReq->u.In.cSymbols, (long)pReq->u.In.cbImageWithTabs));
1241	REQ_CHECK_EXPR_FMT( !pReq->u.In.cbStrTab
1242	\|\| ( pReq->u.In.offStrTab < pReq->u.In.cbImageWithTabs
1243	&& pReq->u.In.offStrTab + pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs
1244	&& pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs),
1245	("SUP_IOCTL_LDR_LOAD: offStrTab=%#lx cbStrTab=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offStrTab,
1246	(long)pReq->u.In.cbStrTab, (long)pReq->u.In.cbImageWithTabs));
1247
1248	if (pReq->u.In.cSymbols)
1249	{
1250	uint32_t i;
1251	PSUPLDRSYM paSyms = (PSUPLDRSYM)&pReq->u.In.abImage[pReq->u.In.offSymbols];
1252	for (i = 0; i < pReq->u.In.cSymbols; i++)
1253	{
1254	REQ_CHECK_EXPR_FMT(paSyms[i].offSymbol < pReq->u.In.cbImageWithTabs,
1255	("SUP_IOCTL_LDR_LOAD: sym #%ld: symb off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offSymbol, (long)pReq->u.In.cbImageWithTabs));
1256	REQ_CHECK_EXPR_FMT(paSyms[i].offName < pReq->u.In.cbStrTab,
1257	("SUP_IOCTL_LDR_LOAD: sym #%ld: name off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1258	REQ_CHECK_EXPR_FMT(memchr(&pReq->u.In.abImage[pReq->u.In.offStrTab + paSyms[i].offName], '\0', pReq->u.In.cbStrTab - paSyms[i].offName),
1259	("SUP_IOCTL_LDR_LOAD: sym #%ld: unterminated name! (%#lx / %#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1260	}
1261	}
1262
1263	/* execute */
1264	pReq->Hdr.rc = supdrvIOCtl_LdrLoad(pDevExt, pSession, pReq);
1265	return 0;
1266	}
1267
1268	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_FREE):
1269	{
1270	/* validate */
1271	PSUPLDRFREE pReq = (PSUPLDRFREE)pReqHdr;
1272	REQ_CHECK_SIZES(SUP_IOCTL_LDR_FREE);
1273
1274	/* execute */
1275	pReq->Hdr.rc = supdrvIOCtl_LdrFree(pDevExt, pSession, pReq);
1276	return 0;
1277	}
1278
1279	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_GET_SYMBOL):
1280	{
1281	/* validate */
1282	PSUPLDRGETSYMBOL pReq = (PSUPLDRGETSYMBOL)pReqHdr;
1283	REQ_CHECK_SIZES(SUP_IOCTL_LDR_GET_SYMBOL);
1284	REQ_CHECK_EXPR(SUP_IOCTL_LDR_GET_SYMBOL, memchr(pReq->u.In.szSymbol, '\0', sizeof(pReq->u.In.szSymbol)));
1285
1286	/* execute */
1287	pReq->Hdr.rc = supdrvIOCtl_LdrGetSymbol(pDevExt, pSession, pReq);
1288	return 0;
1289	}
1290
1291	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0(0)):
1292	{
1293	/* validate */
1294	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1295	Log4(("SUP_IOCTL_CALL_VMMR0: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1296	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1297
1298	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_VMMR0_SIZE(0))
1299	{
1300	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(0), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(0));
1301
1302	/* execute */
1303	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1304	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg, pSession);
1305	else
1306	pReq->Hdr.rc = VERR_WRONG_ORDER;
1307	}
1308	else
1309	{
1310	PSUPVMMR0REQHDR pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1311	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR)),
1312	("SUP_IOCTL_CALL_VMMR0: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR))));
1313	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1314	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(pVMMReq->cbReq));
1315
1316	/* execute */
1317	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1318	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1319	else
1320	pReq->Hdr.rc = VERR_WRONG_ORDER;
1321	}
1322
1323	if ( RT_FAILURE(pReq->Hdr.rc)
1324	&& pReq->Hdr.rc != VERR_INTERRUPTED
1325	&& pReq->Hdr.rc != VERR_TIMEOUT)
1326	Log(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1327	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1328	else
1329	Log4(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1330	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1331	return 0;
1332	}
1333
1334	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_PAGING_MODE):
1335	{
1336	/* validate */
1337	PSUPGETPAGINGMODE pReq = (PSUPGETPAGINGMODE)pReqHdr;
1338	REQ_CHECK_SIZES(SUP_IOCTL_GET_PAGING_MODE);
1339
1340	/* execute */
1341	pReq->Hdr.rc = VINF_SUCCESS;
1342	pReq->u.Out.enmMode = SUPR0GetPagingMode();
1343	return 0;
1344	}
1345
1346	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_ALLOC):
1347	{
1348	/* validate */
1349	PSUPLOWALLOC pReq = (PSUPLOWALLOC)pReqHdr;
1350	REQ_CHECK_EXPR(SUP_IOCTL_LOW_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_LOW_ALLOC_SIZE_IN);
1351	REQ_CHECK_SIZES_EX(SUP_IOCTL_LOW_ALLOC, SUP_IOCTL_LOW_ALLOC_SIZE_IN, SUP_IOCTL_LOW_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1352
1353	/* execute */
1354	pReq->Hdr.rc = SUPR0LowAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1355	if (RT_FAILURE(pReq->Hdr.rc))
1356	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1357	return 0;
1358	}
1359
1360	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_FREE):
1361	{
1362	/* validate */
1363	PSUPLOWFREE pReq = (PSUPLOWFREE)pReqHdr;
1364	REQ_CHECK_SIZES(SUP_IOCTL_LOW_FREE);
1365
1366	/* execute */
1367	pReq->Hdr.rc = SUPR0LowFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1368	return 0;
1369	}
1370
1371	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_MAP):
1372	{
1373	/* validate */
1374	PSUPGIPMAP pReq = (PSUPGIPMAP)pReqHdr;
1375	REQ_CHECK_SIZES(SUP_IOCTL_GIP_MAP);
1376
1377	/* execute */
1378	pReq->Hdr.rc = SUPR0GipMap(pSession, &pReq->u.Out.pGipR3, &pReq->u.Out.HCPhysGip);
1379	if (RT_SUCCESS(pReq->Hdr.rc))
1380	pReq->u.Out.pGipR0 = pDevExt->pGip;
1381	return 0;
1382	}
1383
1384	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_UNMAP):
1385	{
1386	/* validate */
1387	PSUPGIPUNMAP pReq = (PSUPGIPUNMAP)pReqHdr;
1388	REQ_CHECK_SIZES(SUP_IOCTL_GIP_UNMAP);
1389
1390	/* execute */
1391	pReq->Hdr.rc = SUPR0GipUnmap(pSession);
1392	return 0;
1393	}
1394
1395	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SET_VM_FOR_FAST):
1396	{
1397	/* validate */
1398	PSUPSETVMFORFAST pReq = (PSUPSETVMFORFAST)pReqHdr;
1399	REQ_CHECK_SIZES(SUP_IOCTL_SET_VM_FOR_FAST);
1400	REQ_CHECK_EXPR_FMT( !pReq->u.In.pVMR0
1401	\|\| ( VALID_PTR(pReq->u.In.pVMR0)
1402	&& !((uintptr_t)pReq->u.In.pVMR0 & (PAGE_SIZE - 1))),
1403	("SUP_IOCTL_SET_VM_FOR_FAST: pVMR0=%p!\n", pReq->u.In.pVMR0));
1404	/* execute */
1405	pSession->pVM = pReq->u.In.pVMR0;
1406	pReq->Hdr.rc = VINF_SUCCESS;
1407	return 0;
1408	}
1409
1410	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_ALLOC_EX):
1411	{
1412	/* validate */
1413	PSUPPAGEALLOCEX pReq = (PSUPPAGEALLOCEX)pReqHdr;
1414	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC_EX, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_EX_SIZE_IN);
1415	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));
1416	REQ_CHECK_EXPR_FMT(pReq->u.In.fKernelMapping \|\| pReq->u.In.fUserMapping,
1417	("SUP_IOCTL_PAGE_ALLOC_EX: No mapping requested!\n"));
1418	REQ_CHECK_EXPR_FMT(pReq->u.In.fUserMapping,
1419	("SUP_IOCTL_PAGE_ALLOC_EX: Must have user mapping!\n"));
1420	REQ_CHECK_EXPR_FMT(!pReq->u.In.fReserved0 && !pReq->u.In.fReserved1,
1421	("SUP_IOCTL_PAGE_ALLOC_EX: fReserved0=%d fReserved1=%d\n", pReq->u.In.fReserved0, pReq->u.In.fReserved1));
1422
1423	/* execute */
1424	pReq->Hdr.rc = SUPR0PageAllocEx(pSession, pReq->u.In.cPages, 0 /* fFlags */,
1425	pReq->u.In.fUserMapping ? &pReq->u.Out.pvR3 : NULL,
1426	pReq->u.In.fKernelMapping ? &pReq->u.Out.pvR0 : NULL,
1427	&pReq->u.Out.aPages[0]);
1428	if (RT_FAILURE(pReq->Hdr.rc))
1429	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1430	return 0;
1431	}
1432
1433	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_MAP_KERNEL):
1434	{
1435	/* validate */
1436	PSUPPAGEMAPKERNEL pReq = (PSUPPAGEMAPKERNEL)pReqHdr;
1437	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_MAP_KERNEL);
1438	REQ_CHECK_EXPR_FMT(!pReq->u.In.fFlags, ("SUP_IOCTL_PAGE_MAP_KERNEL: fFlags=%#x! MBZ\n", pReq->u.In.fFlags));
1439	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_MAP_KERNEL: offSub=%#x\n", pReq->u.In.offSub));
1440	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1441	("SUP_IOCTL_PAGE_MAP_KERNEL: cbSub=%#x\n", pReq->u.In.cbSub));
1442
1443	/* execute */
1444	pReq->Hdr.rc = SUPR0PageMapKernel(pSession, pReq->u.In.pvR3, pReq->u.In.offSub, pReq->u.In.cbSub,
1445	pReq->u.In.fFlags, &pReq->u.Out.pvR0);
1446	if (RT_FAILURE(pReq->Hdr.rc))
1447	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1448	return 0;
1449	}
1450
1451	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_PROTECT):
1452	{
1453	/* validate */
1454	PSUPPAGEPROTECT pReq = (PSUPPAGEPROTECT)pReqHdr;
1455	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_PROTECT);
1456	REQ_CHECK_EXPR_FMT(!(pReq->u.In.fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)),
1457	("SUP_IOCTL_PAGE_PROTECT: fProt=%#x!\n", pReq->u.In.fProt));
1458	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_PROTECT: offSub=%#x\n", pReq->u.In.offSub));
1459	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1460	("SUP_IOCTL_PAGE_PROTECT: cbSub=%#x\n", pReq->u.In.cbSub));
1461
1462	/* execute */
1463	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);
1464	return 0;
1465	}
1466
1467	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_FREE):
1468	{
1469	/* validate */
1470	PSUPPAGEFREE pReq = (PSUPPAGEFREE)pReqHdr;
1471	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_FREE);
1472
1473	/* execute */
1474	pReq->Hdr.rc = SUPR0PageFree(pSession, pReq->u.In.pvR3);
1475	return 0;
1476	}
1477
1478	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_SERVICE(0)):
1479	{
1480	/* validate */
1481	PSUPCALLSERVICE pReq = (PSUPCALLSERVICE)pReqHdr;
1482	Log4(("SUP_IOCTL_CALL_SERVICE: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1483	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1484
1485	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
1486	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(0), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(0));
1487	else
1488	{
1489	PSUPR0SERVICEREQHDR pSrvReq = (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0];
1490	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR)),
1491	("SUP_IOCTL_CALL_SERVICE: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR))));
1492	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, pSrvReq->u32Magic == SUPR0SERVICEREQHDR_MAGIC);
1493	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(pSrvReq->cbReq), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(pSrvReq->cbReq));
1494	}
1495	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, memchr(pReq->u.In.szName, '\0', sizeof(pReq->u.In.szName)));
1496
1497	/* execute */
1498	pReq->Hdr.rc = supdrvIOCtl_CallServiceModule(pDevExt, pSession, pReq);
1499	return 0;
1500	}
1501
1502	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOGGER_SETTINGS(0)):
1503	{
1504	/* validate */
1505	PSUPLOGGERSETTINGS pReq = (PSUPLOGGERSETTINGS)pReqHdr;
1506	size_t cbStrTab;
1507	REQ_CHECK_SIZE_OUT(SUP_IOCTL_LOGGER_SETTINGS, SUP_IOCTL_LOGGER_SETTINGS_SIZE_OUT);
1508	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->Hdr.cbIn >= SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(1));
1509	cbStrTab = pReq->Hdr.cbIn - SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(0);
1510	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offGroups < cbStrTab);
1511	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offFlags < cbStrTab);
1512	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offDestination < cbStrTab);
1513	REQ_CHECK_EXPR_FMT(pReq->u.In.szStrings[cbStrTab - 1] == '\0',
1514	("SUP_IOCTL_LOGGER_SETTINGS: cbIn=%#x cbStrTab=%#zx LastChar=%d\n",
1515	pReq->Hdr.cbIn, cbStrTab, pReq->u.In.szStrings[cbStrTab - 1]));
1516	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhich <= SUPLOGGERSETTINGS_WHICH_RELEASE);
1517	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhat <= SUPLOGGERSETTINGS_WHAT_DESTROY);
1518
1519	/* execute */
1520	pReq->Hdr.rc = supdrvIOCtl_LoggerSettings(pDevExt, pSession, pReq);
1521	return 0;
1522	}
1523
1524	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_CREATE):
1525	{
1526	/* validate */
1527	PSUPSEMCREATE pReq = (PSUPSEMCREATE)pReqHdr;
1528	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_CREATE, SUP_IOCTL_SEM_CREATE_SIZE_IN, SUP_IOCTL_SEM_CREATE_SIZE_OUT);
1529
1530	/* execute */
1531	switch (pReq->u.In.uType)
1532	{
1533	case SUP_SEM_TYPE_EVENT:
1534	{
1535	SUPSEMEVENT hEvent;
1536	pReq->Hdr.rc = SUPSemEventCreate(pSession, &hEvent);
1537	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEvent;
1538	break;
1539	}
1540
1541	case SUP_SEM_TYPE_EVENT_MULTI:
1542	{
1543	SUPSEMEVENTMULTI hEventMulti;
1544	pReq->Hdr.rc = SUPSemEventMultiCreate(pSession, &hEventMulti);
1545	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEventMulti;
1546	break;
1547	}
1548
1549	default:
1550	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1551	break;
1552	}
1553	return 0;
1554	}
1555
1556	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP):
1557	{
1558	/* validate */
1559	PSUPSEMOP pReq = (PSUPSEMOP)pReqHdr;
1560	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP, SUP_IOCTL_SEM_OP_SIZE_IN, SUP_IOCTL_SEM_OP_SIZE_OUT);
1561
1562	/* execute */
1563	switch (pReq->u.In.uType)
1564	{
1565	case SUP_SEM_TYPE_EVENT:
1566	{
1567	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1568	switch (pReq->u.In.uOp)
1569	{
1570	case SUPSEMOP_WAIT:
1571	pReq->Hdr.rc = SUPSemEventWaitNoResume(pSession, hEvent, pReq->u.In.cMillies);
1572	break;
1573	case SUPSEMOP_SIGNAL:
1574	pReq->Hdr.rc = SUPSemEventSignal(pSession, hEvent);
1575	break;
1576	case SUPSEMOP_CLOSE:
1577	pReq->Hdr.rc = SUPSemEventClose(pSession, hEvent);
1578	break;
1579	case SUPSEMOP_RESET:
1580	default:
1581	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1582	break;
1583	}
1584	break;
1585	}
1586
1587	case SUP_SEM_TYPE_EVENT_MULTI:
1588	{
1589	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1590	switch (pReq->u.In.uOp)
1591	{
1592	case SUPSEMOP_WAIT:
1593	pReq->Hdr.rc = SUPSemEventMultiWaitNoResume(pSession, hEventMulti, pReq->u.In.cMillies);
1594	break;
1595	case SUPSEMOP_SIGNAL:
1596	pReq->Hdr.rc = SUPSemEventMultiSignal(pSession, hEventMulti);
1597	break;
1598	case SUPSEMOP_CLOSE:
1599	pReq->Hdr.rc = SUPSemEventMultiClose(pSession, hEventMulti);
1600	break;
1601	case SUPSEMOP_RESET:
1602	pReq->Hdr.rc = SUPSemEventMultiReset(pSession, hEventMulti);
1603	break;
1604	default:
1605	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1606	break;
1607	}
1608	break;
1609	}
1610
1611	default:
1612	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1613	break;
1614	}
1615	return 0;
1616	}
1617
1618	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
1619	{
1620	/* validate */
1621	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
1622	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
1623	REQ_CHECK_EXPR(SUP_IOCTL_VT_CAPS, pReq->Hdr.cbIn <= SUP_IOCTL_VT_CAPS_SIZE_IN);
1624
1625	/* execute */
1626	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.Caps);
1627	if (RT_FAILURE(pReq->Hdr.rc))
1628	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1629	return 0;
1630	}
1631
1632	default:
1633	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1634	break;
1635	}
1636	return VERR_GENERAL_FAILURE;
1637	}
1638
1639
1640	/**
1641	* Inter-Driver Communcation (IDC) worker.
1642	*
1643	* @returns VBox status code.
1644	* @retval VINF_SUCCESS on success.
1645	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1646	* @retval VERR_NOT_SUPPORTED if the request isn't supported.
1647	*
1648	* @param uReq The request (function) code.
1649	* @param pDevExt Device extention.
1650	* @param pSession Session data.
1651	* @param pReqHdr The request header.
1652	*/
1653	int VBOXCALL supdrvIDC(uintptr_t uReq, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQHDR pReqHdr)
1654	{
1655	/*
1656	* The OS specific code has already validated the pSession
1657	* pointer, and the request size being greater or equal to
1658	* size of the header.
1659	*
1660	* So, just check that pSession is a kernel context session.
1661	*/
1662	if (RT_UNLIKELY( pSession
1663	&& pSession->R0Process != NIL_RTR0PROCESS))
1664	return VERR_INVALID_PARAMETER;
1665
1666	/*
1667	* Validation macro.
1668	*/
1669	#define REQ_CHECK_IDC_SIZE(Name, cbExpect) \
1670	do { \
1671	if (RT_UNLIKELY(pReqHdr->cb != (cbExpect))) \
1672	{ \
1673	OSDBGPRINT(( #Name ": Invalid input/output sizes. cb=%ld expected %ld.\n", \
1674	(long)pReqHdr->cb, (long)(cbExpect))); \
1675	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1676	} \
1677	} while (0)
1678
1679	switch (uReq)
1680	{
1681	case SUPDRV_IDC_REQ_CONNECT:
1682	{
1683	PSUPDRVIDCREQCONNECT pReq = (PSUPDRVIDCREQCONNECT)pReqHdr;
1684	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_CONNECT, sizeof(*pReq));
1685
1686	/*
1687	* Validate the cookie and other input.
1688	*/
1689	if (pReq->Hdr.pSession != NULL)
1690	{
1691	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: pSession=%p expected NULL!\n", pReq->Hdr.pSession));
1692	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1693	}
1694	if (pReq->u.In.u32MagicCookie != SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE)
1695	{
1696	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: u32MagicCookie=%#x expected %#x!\n",
1697	(unsigned)pReq->u.In.u32MagicCookie, (unsigned)SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE));
1698	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1699	}
1700	if ( pReq->u.In.uMinVersion > pReq->u.In.uReqVersion
1701	\|\| (pReq->u.In.uMinVersion & UINT32_C(0xffff0000)) != (pReq->u.In.uReqVersion & UINT32_C(0xffff0000)))
1702	{
1703	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: uMinVersion=%#x uMaxVersion=%#x doesn't match!\n",
1704	pReq->u.In.uMinVersion, pReq->u.In.uReqVersion));
1705	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1706	}
1707
1708	/*
1709	* Match the version.
1710	* The current logic is very simple, match the major interface version.
1711	*/
1712	if ( pReq->u.In.uMinVersion > SUPDRV_IDC_VERSION
1713	\|\| (pReq->u.In.uMinVersion & 0xffff0000) != (SUPDRV_IDC_VERSION & 0xffff0000))
1714	{
1715	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1716	pReq->u.In.uReqVersion, pReq->u.In.uMinVersion, (unsigned)SUPDRV_IDC_VERSION));
1717	pReq->u.Out.pSession = NULL;
1718	pReq->u.Out.uSessionVersion = 0xffffffff;
1719	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1720	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1721	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1722	return VINF_SUCCESS;
1723	}
1724
1725	pReq->u.Out.pSession = NULL;
1726	pReq->u.Out.uSessionVersion = SUPDRV_IDC_VERSION;
1727	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1728	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1729
1730	/*
1731	* On NT we will already have a session associated with the
1732	* client, just like with the SUP_IOCTL_COOKIE request, while
1733	* the other doesn't.
1734	*/
1735	#ifdef RT_OS_WINDOWS
1736	pReq->Hdr.rc = VINF_SUCCESS;
1737	#else
1738	AssertReturn(!pSession, VERR_INTERNAL_ERROR);
1739	pReq->Hdr.rc = supdrvCreateSession(pDevExt, false /* fUser */, &pSession);
1740	if (RT_FAILURE(pReq->Hdr.rc))
1741	{
1742	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: failed to create session, rc=%d\n", pReq->Hdr.rc));
1743	return VINF_SUCCESS;
1744	}
1745	#endif
1746
1747	pReq->u.Out.pSession = pSession;
1748	pReq->Hdr.pSession = pSession;
1749
1750	return VINF_SUCCESS;
1751	}
1752
1753	case SUPDRV_IDC_REQ_DISCONNECT:
1754	{
1755	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_DISCONNECT, sizeof(*pReqHdr));
1756
1757	#ifdef RT_OS_WINDOWS
1758	/* Windows will destroy the session when the file object is destroyed. */
1759	#else
1760	supdrvCloseSession(pDevExt, pSession);
1761	#endif
1762	return pReqHdr->rc = VINF_SUCCESS;
1763	}
1764
1765	case SUPDRV_IDC_REQ_GET_SYMBOL:
1766	{
1767	PSUPDRVIDCREQGETSYM pReq = (PSUPDRVIDCREQGETSYM)pReqHdr;
1768	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_GET_SYMBOL, sizeof(*pReq));
1769
1770	pReq->Hdr.rc = supdrvIDC_LdrGetSymbol(pDevExt, pSession, pReq);
1771	return VINF_SUCCESS;
1772	}
1773
1774	case SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY:
1775	{
1776	PSUPDRVIDCREQCOMPREGFACTORY pReq = (PSUPDRVIDCREQCOMPREGFACTORY)pReqHdr;
1777	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY, sizeof(*pReq));
1778
1779	pReq->Hdr.rc = SUPR0ComponentRegisterFactory(pSession, pReq->u.In.pFactory);
1780	return VINF_SUCCESS;
1781	}
1782
1783	case SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY:
1784	{
1785	PSUPDRVIDCREQCOMPDEREGFACTORY pReq = (PSUPDRVIDCREQCOMPDEREGFACTORY)pReqHdr;
1786	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY, sizeof(*pReq));
1787
1788	pReq->Hdr.rc = SUPR0ComponentDeregisterFactory(pSession, pReq->u.In.pFactory);
1789	return VINF_SUCCESS;
1790	}
1791
1792	default:
1793	Log(("Unknown IDC %#lx\n", (long)uReq));
1794	break;
1795	}
1796
1797	#undef REQ_CHECK_IDC_SIZE
1798	return VERR_NOT_SUPPORTED;
1799	}
1800
1801
1802	/**
1803	* Register a object for reference counting.
1804	* The object is registered with one reference in the specified session.
1805	*
1806	* @returns Unique identifier on success (pointer).
1807	* All future reference must use this identifier.
1808	* @returns NULL on failure.
1809	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
1810	* @param pvUser1 The first user argument.
1811	* @param pvUser2 The second user argument.
1812	*/
1813	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
1814	{
1815	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1816	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1817	PSUPDRVOBJ pObj;
1818	PSUPDRVUSAGE pUsage;
1819
1820	/*
1821	* Validate the input.
1822	*/
1823	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
1824	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
1825	AssertPtrReturn(pfnDestructor, NULL);
1826
1827	/*
1828	* Allocate and initialize the object.
1829	*/
1830	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
1831	if (!pObj)
1832	return NULL;
1833	pObj->u32Magic = SUPDRVOBJ_MAGIC;
1834	pObj->enmType = enmType;
1835	pObj->pNext = NULL;
1836	pObj->cUsage = 1;
1837	pObj->pfnDestructor = pfnDestructor;
1838	pObj->pvUser1 = pvUser1;
1839	pObj->pvUser2 = pvUser2;
1840	pObj->CreatorUid = pSession->Uid;
1841	pObj->CreatorGid = pSession->Gid;
1842	pObj->CreatorProcess= pSession->Process;
1843	supdrvOSObjInitCreator(pObj, pSession);
1844
1845	/*
1846	* Allocate the usage record.
1847	* (We keep freed usage records around to simplify SUPR0ObjAddRefEx().)
1848	*/
1849	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1850
1851	pUsage = pDevExt->pUsageFree;
1852	if (pUsage)
1853	pDevExt->pUsageFree = pUsage->pNext;
1854	else
1855	{
1856	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1857	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
1858	if (!pUsage)
1859	{
1860	RTMemFree(pObj);
1861	return NULL;
1862	}
1863	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1864	}
1865
1866	/*
1867	* Insert the object and create the session usage record.
1868	*/
1869	/* The object. */
1870	pObj->pNext = pDevExt->pObjs;
1871	pDevExt->pObjs = pObj;
1872
1873	/* The session record. */
1874	pUsage->cUsage = 1;
1875	pUsage->pObj = pObj;
1876	pUsage->pNext = pSession->pUsage;
1877	/* Log2(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext)); */
1878	pSession->pUsage = pUsage;
1879
1880	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1881
1882	Log(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
1883	return pObj;
1884	}
1885
1886
1887	/**
1888	* Increment the reference counter for the object associating the reference
1889	* with the specified session.
1890	*
1891	* @returns IPRT status code.
1892	* @param pvObj The identifier returned by SUPR0ObjRegister().
1893	* @param pSession The session which is referencing the object.
1894	*
1895	* @remarks The caller should not own any spinlocks and must carefully protect
1896	* itself against potential race with the destructor so freed memory
1897	* isn't accessed here.
1898	*/
1899	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
1900	{
1901	return SUPR0ObjAddRefEx(pvObj, pSession, false /* fNoBlocking */);
1902	}
1903
1904
1905	/**
1906	* Increment the reference counter for the object associating the reference
1907	* with the specified session.
1908	*
1909	* @returns IPRT status code.
1910	* @retval VERR_TRY_AGAIN if fNoBlocking was set and a new usage record
1911	* couldn't be allocated. (If you see this you're not doing the right
1912	* thing and it won't ever work reliably.)
1913	*
1914	* @param pvObj The identifier returned by SUPR0ObjRegister().
1915	* @param pSession The session which is referencing the object.
1916	* @param fNoBlocking Set if it's not OK to block. Never try to make the
1917	* first reference to an object in a session with this
1918	* argument set.
1919	*
1920	* @remarks The caller should not own any spinlocks and must carefully protect
1921	* itself against potential race with the destructor so freed memory
1922	* isn't accessed here.
1923	*/
1924	SUPR0DECL(int) SUPR0ObjAddRefEx(void *pvObj, PSUPDRVSESSION pSession, bool fNoBlocking)
1925	{
1926	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1927	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1928	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
1929	int rc = VINF_SUCCESS;
1930	PSUPDRVUSAGE pUsagePre;
1931	PSUPDRVUSAGE pUsage;
1932
1933	/*
1934	* Validate the input.
1935	* Be ready for the destruction race (someone might be stuck in the
1936	* destructor waiting a lock we own).
1937	*/
1938	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1939	AssertPtrReturn(pObj, VERR_INVALID_POINTER);
1940	AssertMsgReturn(pObj->u32Magic == SUPDRVOBJ_MAGIC \|\| pObj->u32Magic == SUPDRVOBJ_MAGIC_DEAD,
1941	("Invalid pvObj=%p magic=%#x (expected %#x or %#x)\n", pvObj, pObj->u32Magic, SUPDRVOBJ_MAGIC, SUPDRVOBJ_MAGIC_DEAD),
1942	VERR_INVALID_PARAMETER);
1943
1944	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1945
1946	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
1947	{
1948	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1949
1950	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
1951	return VERR_WRONG_ORDER;
1952	}
1953
1954	/*
1955	* Preallocate the usage record if we can.
1956	*/
1957	pUsagePre = pDevExt->pUsageFree;
1958	if (pUsagePre)
1959	pDevExt->pUsageFree = pUsagePre->pNext;
1960	else if (!fNoBlocking)
1961	{
1962	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1963	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
1964	if (!pUsagePre)
1965	return VERR_NO_MEMORY;
1966
1967	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1968	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
1969	{
1970	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1971
1972	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
1973	return VERR_WRONG_ORDER;
1974	}
1975	}
1976
1977	/*
1978	* Reference the object.
1979	*/
1980	pObj->cUsage++;
1981
1982	/*
1983	* Look for the session record.
1984	*/
1985	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
1986	{
1987	/Log(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
1988	if (pUsage->pObj == pObj)
1989	break;
1990	}
1991	if (pUsage)
1992	pUsage->cUsage++;
1993	else if (pUsagePre)
1994	{
1995	/* create a new session record. */
1996	pUsagePre->cUsage = 1;
1997	pUsagePre->pObj = pObj;
1998	pUsagePre->pNext = pSession->pUsage;
1999	pSession->pUsage = pUsagePre;
2000	/Log(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));/
2001
2002	pUsagePre = NULL;
2003	}
2004	else
2005	{
2006	pObj->cUsage--;
2007	rc = VERR_TRY_AGAIN;
2008	}
2009
2010	/*
2011	* Put any unused usage record into the free list..
2012	*/
2013	if (pUsagePre)
2014	{
2015	pUsagePre->pNext = pDevExt->pUsageFree;
2016	pDevExt->pUsageFree = pUsagePre;
2017	}
2018
2019	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2020
2021	return rc;
2022	}
2023
2024
2025	/**
2026	* Decrement / destroy a reference counter record for an object.
2027	*
2028	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
2029	*
2030	* @returns IPRT status code.
2031	* @retval VINF_SUCCESS if not destroyed.
2032	* @retval VINF_OBJECT_DESTROYED if it's destroyed by this release call.
2033	* @retval VERR_INVALID_PARAMETER if the object isn't valid. Will assert in
2034	* string builds.
2035	*
2036	* @param pvObj The identifier returned by SUPR0ObjRegister().
2037	* @param pSession The session which is referencing the object.
2038	*/
2039	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
2040	{
2041	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2042	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2043	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2044	int rc = VERR_INVALID_PARAMETER;
2045	PSUPDRVUSAGE pUsage;
2046	PSUPDRVUSAGE pUsagePrev;
2047
2048	/*
2049	* Validate the input.
2050	*/
2051	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2052	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2053	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2054	VERR_INVALID_PARAMETER);
2055
2056	/*
2057	* Acquire the spinlock and look for the usage record.
2058	*/
2059	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2060
2061	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
2062	pUsage;
2063	pUsagePrev = pUsage, pUsage = pUsage->pNext)
2064	{
2065	/Log2(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2066	if (pUsage->pObj == pObj)
2067	{
2068	rc = VINF_SUCCESS;
2069	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
2070	if (pUsage->cUsage > 1)
2071	{
2072	pObj->cUsage--;
2073	pUsage->cUsage--;
2074	}
2075	else
2076	{
2077	/*
2078	* Free the session record.
2079	*/
2080	if (pUsagePrev)
2081	pUsagePrev->pNext = pUsage->pNext;
2082	else
2083	pSession->pUsage = pUsage->pNext;
2084	pUsage->pNext = pDevExt->pUsageFree;
2085	pDevExt->pUsageFree = pUsage;
2086
2087	/* What about the object? */
2088	if (pObj->cUsage > 1)
2089	pObj->cUsage--;
2090	else
2091	{
2092	/*
2093	* Object is to be destroyed, unlink it.
2094	*/
2095	pObj->u32Magic = SUPDRVOBJ_MAGIC_DEAD;
2096	rc = VINF_OBJECT_DESTROYED;
2097	if (pDevExt->pObjs == pObj)
2098	pDevExt->pObjs = pObj->pNext;
2099	else
2100	{
2101	PSUPDRVOBJ pObjPrev;
2102	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
2103	if (pObjPrev->pNext == pObj)
2104	{
2105	pObjPrev->pNext = pObj->pNext;
2106	break;
2107	}
2108	Assert(pObjPrev);
2109	}
2110	}
2111	}
2112	break;
2113	}
2114	}
2115
2116	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2117
2118	/*
2119	* Call the destructor and free the object if required.
2120	*/
2121	if (rc == VINF_OBJECT_DESTROYED)
2122	{
2123	Log(("SUPR0ObjRelease: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
2124	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
2125	if (pObj->pfnDestructor)
2126	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
2127	RTMemFree(pObj);
2128	}
2129
2130	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
2131	return rc;
2132	}
2133
2134
2135	/**
2136	* Verifies that the current process can access the specified object.
2137	*
2138	* @returns The following IPRT status code:
2139	* @retval VINF_SUCCESS if access was granted.
2140	* @retval VERR_PERMISSION_DENIED if denied access.
2141	* @retval VERR_INVALID_PARAMETER if invalid parameter.
2142	*
2143	* @param pvObj The identifier returned by SUPR0ObjRegister().
2144	* @param pSession The session which wishes to access the object.
2145	* @param pszObjName Object string name. This is optional and depends on the object type.
2146	*
2147	* @remark The caller is responsible for making sure the object isn't removed while
2148	* we're inside this function. If uncertain about this, just call AddRef before calling us.
2149	*/
2150	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
2151	{
2152	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2153	int rc;
2154
2155	/*
2156	* Validate the input.
2157	*/
2158	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2159	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2160	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2161	VERR_INVALID_PARAMETER);
2162
2163	/*
2164	* Check access. (returns true if a decision has been made.)
2165	*/
2166	rc = VERR_INTERNAL_ERROR;
2167	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
2168	return rc;
2169
2170	/*
2171	* Default policy is to allow the user to access his own
2172	* stuff but nothing else.
2173	*/
2174	if (pObj->CreatorUid == pSession->Uid)
2175	return VINF_SUCCESS;
2176	return VERR_PERMISSION_DENIED;
2177	}
2178
2179
2180	/**
2181	* Lock pages.
2182	*
2183	* @returns IPRT status code.
2184	* @param pSession Session to which the locked memory should be associated.
2185	* @param pvR3 Start of the memory range to lock.
2186	* This must be page aligned.
2187	* @param cPages Number of pages to lock.
2188	* @param paPages Where to put the physical addresses of locked memory.
2189	*/
2190	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
2191	{
2192	int rc;
2193	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2194	const size_t cb = (size_t)cPages << PAGE_SHIFT;
2195	LogFlow(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
2196
2197	/*
2198	* Verify input.
2199	*/
2200	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2201	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
2202	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
2203	\|\| !pvR3)
2204	{
2205	Log(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
2206	return VERR_INVALID_PARAMETER;
2207	}
2208
2209	/*
2210	* Let IPRT do the job.
2211	*/
2212	Mem.eType = MEMREF_TYPE_LOCKED;
2213	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTMEM_PROT_READ \| RTMEM_PROT_WRITE, RTR0ProcHandleSelf());
2214	if (RT_SUCCESS(rc))
2215	{
2216	uint32_t iPage = cPages;
2217	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
2218	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
2219
2220	while (iPage-- > 0)
2221	{
2222	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2223	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
2224	{
2225	AssertMsgFailed(("iPage=%d\n", iPage));
2226	rc = VERR_INTERNAL_ERROR;
2227	break;
2228	}
2229	}
2230	if (RT_SUCCESS(rc))
2231	rc = supdrvMemAdd(&Mem, pSession);
2232	if (RT_FAILURE(rc))
2233	{
2234	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
2235	AssertRC(rc2);
2236	}
2237	}
2238
2239	return rc;
2240	}
2241
2242
2243	/**
2244	* Unlocks the memory pointed to by pv.
2245	*
2246	* @returns IPRT status code.
2247	* @param pSession Session to which the memory was locked.
2248	* @param pvR3 Memory to unlock.
2249	*/
2250	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2251	{
2252	LogFlow(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2253	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2254	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
2255	}
2256
2257
2258	/**
2259	* Allocates a chunk of page aligned memory with contiguous and fixed physical
2260	* backing.
2261	*
2262	* @returns IPRT status code.
2263	* @param pSession Session data.
2264	* @param cPages Number of pages to allocate.
2265	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
2266	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
2267	* @param pHCPhys Where to put the physical address of allocated memory.
2268	*/
2269	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
2270	{
2271	int rc;
2272	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2273	LogFlow(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
2274
2275	/*
2276	* Validate input.
2277	*/
2278	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2279	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
2280	{
2281	Log(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
2282	pSession, ppvR0, ppvR3, pHCPhys));
2283	return VERR_INVALID_PARAMETER;
2284
2285	}
2286	if (cPages < 1 \|\| cPages >= 256)
2287	{
2288	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2289	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2290	}
2291
2292	/*
2293	* Let IPRT do the job.
2294	*/
2295	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
2296	if (RT_SUCCESS(rc))
2297	{
2298	int rc2;
2299	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2300	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2301	if (RT_SUCCESS(rc))
2302	{
2303	Mem.eType = MEMREF_TYPE_CONT;
2304	rc = supdrvMemAdd(&Mem, pSession);
2305	if (!rc)
2306	{
2307	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2308	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2309	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
2310	return 0;
2311	}
2312
2313	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2314	AssertRC(rc2);
2315	}
2316	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2317	AssertRC(rc2);
2318	}
2319
2320	return rc;
2321	}
2322
2323
2324	/**
2325	* Frees memory allocated using SUPR0ContAlloc().
2326	*
2327	* @returns IPRT status code.
2328	* @param pSession The session to which the memory was allocated.
2329	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2330	*/
2331	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2332	{
2333	LogFlow(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2334	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2335	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
2336	}
2337
2338
2339	/**
2340	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
2341	*
2342	* The memory isn't zeroed.
2343	*
2344	* @returns IPRT status code.
2345	* @param pSession Session data.
2346	* @param cPages Number of pages to allocate.
2347	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
2348	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
2349	* @param paPages Where to put the physical addresses of allocated memory.
2350	*/
2351	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
2352	{
2353	unsigned iPage;
2354	int rc;
2355	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2356	LogFlow(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
2357
2358	/*
2359	* Validate input.
2360	*/
2361	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2362	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
2363	{
2364	Log(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
2365	pSession, ppvR3, ppvR0, paPages));
2366	return VERR_INVALID_PARAMETER;
2367
2368	}
2369	if (cPages < 1 \|\| cPages >= 256)
2370	{
2371	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2372	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2373	}
2374
2375	/*
2376	* Let IPRT do the work.
2377	*/
2378	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
2379	if (RT_SUCCESS(rc))
2380	{
2381	int rc2;
2382	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2383	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2384	if (RT_SUCCESS(rc))
2385	{
2386	Mem.eType = MEMREF_TYPE_LOW;
2387	rc = supdrvMemAdd(&Mem, pSession);
2388	if (!rc)
2389	{
2390	for (iPage = 0; iPage < cPages; iPage++)
2391	{
2392	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2393	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%RHp\n", paPages[iPage]));
2394	}
2395	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2396	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2397	return 0;
2398	}
2399
2400	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2401	AssertRC(rc2);
2402	}
2403
2404	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2405	AssertRC(rc2);
2406	}
2407
2408	return rc;
2409	}
2410
2411
2412	/**
2413	* Frees memory allocated using SUPR0LowAlloc().
2414	*
2415	* @returns IPRT status code.
2416	* @param pSession The session to which the memory was allocated.
2417	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2418	*/
2419	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2420	{
2421	LogFlow(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2422	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2423	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
2424	}
2425
2426
2427
2428	/**
2429	* Allocates a chunk of memory with both R0 and R3 mappings.
2430	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
2431	*
2432	* @returns IPRT status code.
2433	* @param pSession The session to associated the allocation with.
2434	* @param cb Number of bytes to allocate.
2435	* @param ppvR0 Where to store the address of the Ring-0 mapping.
2436	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2437	*/
2438	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
2439	{
2440	int rc;
2441	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2442	LogFlow(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
2443
2444	/*
2445	* Validate input.
2446	*/
2447	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2448	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
2449	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
2450	if (cb < 1 \|\| cb >= _4M)
2451	{
2452	Log(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
2453	return VERR_INVALID_PARAMETER;
2454	}
2455
2456	/*
2457	* Let IPRT do the work.
2458	*/
2459	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
2460	if (RT_SUCCESS(rc))
2461	{
2462	int rc2;
2463	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2464	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2465	if (RT_SUCCESS(rc))
2466	{
2467	Mem.eType = MEMREF_TYPE_MEM;
2468	rc = supdrvMemAdd(&Mem, pSession);
2469	if (!rc)
2470	{
2471	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2472	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2473	return VINF_SUCCESS;
2474	}
2475
2476	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2477	AssertRC(rc2);
2478	}
2479
2480	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2481	AssertRC(rc2);
2482	}
2483
2484	return rc;
2485	}
2486
2487
2488	/**
2489	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
2490	*
2491	* @returns IPRT status code.
2492	* @param pSession The session to which the memory was allocated.
2493	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2494	* @param paPages Where to store the physical addresses.
2495	*/
2496	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
2497	{
2498	PSUPDRVBUNDLE pBundle;
2499	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2500	LogFlow(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
2501
2502	/*
2503	* Validate input.
2504	*/
2505	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2506	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
2507	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
2508
2509	/*
2510	* Search for the address.
2511	*/
2512	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2513	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2514	{
2515	if (pBundle->cUsed > 0)
2516	{
2517	unsigned i;
2518	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2519	{
2520	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
2521	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2522	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2523	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2524	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
2525	)
2526	)
2527	{
2528	const size_t cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2529	size_t iPage;
2530	for (iPage = 0; iPage < cPages; iPage++)
2531	{
2532	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2533	paPages[iPage].uReserved = 0;
2534	}
2535	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2536	return VINF_SUCCESS;
2537	}
2538	}
2539	}
2540	}
2541	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2542	Log(("Failed to find %p!!!\n", (void *)uPtr));
2543	return VERR_INVALID_PARAMETER;
2544	}
2545
2546
2547	/**
2548	* Free memory allocated by SUPR0MemAlloc().
2549	*
2550	* @returns IPRT status code.
2551	* @param pSession The session owning the allocation.
2552	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2553	*/
2554	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2555	{
2556	LogFlow(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2557	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2558	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
2559	}
2560
2561
2562	/**
2563	* Allocates a chunk of memory with a kernel or/and a user mode mapping.
2564	*
2565	* The memory is fixed and it's possible to query the physical addresses using
2566	* SUPR0MemGetPhys().
2567	*
2568	* @returns IPRT status code.
2569	* @param pSession The session to associated the allocation with.
2570	* @param cPages The number of pages to allocate.
2571	* @param fFlags Flags, reserved for the future. Must be zero.
2572	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2573	* NULL if no ring-3 mapping.
2574	* @param ppvR3 Where to store the address of the Ring-0 mapping.
2575	* NULL if no ring-0 mapping.
2576	* @param paPages Where to store the addresses of the pages. Optional.
2577	*/
2578	SUPR0DECL(int) SUPR0PageAllocEx(PSUPDRVSESSION pSession, uint32_t cPages, uint32_t fFlags, PRTR3PTR ppvR3, PRTR0PTR ppvR0, PRTHCPHYS paPages)
2579	{
2580	int rc;
2581	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2582	LogFlow(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
2583
2584	/*
2585	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2586	*/
2587	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2588	AssertPtrNullReturn(ppvR3, VERR_INVALID_POINTER);
2589	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2590	AssertReturn(ppvR3 \|\| ppvR0, VERR_INVALID_PARAMETER);
2591	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2592	if (cPages < 1 \|\| cPages > VBOX_MAX_ALLOC_PAGE_COUNT)
2593	{
2594	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)));
2595	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2596	}
2597
2598	/*
2599	* Let IPRT do the work.
2600	*/
2601	if (ppvR0)
2602	rc = RTR0MemObjAllocPage(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, true /* fExecutable */);
2603	else
2604	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
2605	if (RT_SUCCESS(rc))
2606	{
2607	int rc2;
2608	if (ppvR3)
2609	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2610	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2611	else
2612	Mem.MapObjR3 = NIL_RTR0MEMOBJ;
2613	if (RT_SUCCESS(rc))
2614	{
2615	Mem.eType = MEMREF_TYPE_PAGE;
2616	rc = supdrvMemAdd(&Mem, pSession);
2617	if (!rc)
2618	{
2619	if (ppvR3)
2620	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2621	if (ppvR0)
2622	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2623	if (paPages)
2624	{
2625	uint32_t iPage = cPages;
2626	while (iPage-- > 0)
2627	{
2628	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
2629	Assert(paPages[iPage] != NIL_RTHCPHYS);
2630	}
2631	}
2632	return VINF_SUCCESS;
2633	}
2634
2635	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2636	AssertRC(rc2);
2637	}
2638
2639	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2640	AssertRC(rc2);
2641	}
2642	return rc;
2643	}
2644
2645
2646	/**
2647	* Maps a chunk of memory previously allocated by SUPR0PageAllocEx into kernel
2648	* space.
2649	*
2650	* @returns IPRT status code.
2651	* @param pSession The session to associated the allocation with.
2652	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
2653	* @param offSub Where to start mapping. Must be page aligned.
2654	* @param cbSub How much to map. Must be page aligned.
2655	* @param fFlags Flags, MBZ.
2656	* @param ppvR0 Where to reutrn the address of the ring-0 mapping on
2657	* success.
2658	*/
2659	SUPR0DECL(int) SUPR0PageMapKernel(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t offSub, uint32_t cbSub,
2660	uint32_t fFlags, PRTR0PTR ppvR0)
2661	{
2662	int rc;
2663	PSUPDRVBUNDLE pBundle;
2664	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2665	RTR0MEMOBJ hMemObj = NIL_RTR0MEMOBJ;
2666	LogFlow(("SUPR0PageMapKernel: pSession=%p pvR3=%p offSub=%#x cbSub=%#x\n", pSession, pvR3, offSub, cbSub));
2667
2668	/*
2669	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2670	*/
2671	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2672	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2673	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2674	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2675	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2676	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
2677
2678	/*
2679	* Find the memory object.
2680	*/
2681	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2682	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2683	{
2684	if (pBundle->cUsed > 0)
2685	{
2686	unsigned i;
2687	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2688	{
2689	if ( ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
2690	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2691	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2692	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
2693	\|\| ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED
2694	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2695	&& pBundle->aMem[i].MapObjR3 == NIL_RTR0MEMOBJ
2696	&& RTR0MemObjAddressR3(pBundle->aMem[i].MemObj) == pvR3))
2697	{
2698	hMemObj = pBundle->aMem[i].MemObj;
2699	break;
2700	}
2701	}
2702	}
2703	}
2704	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2705
2706	rc = VERR_INVALID_PARAMETER;
2707	if (hMemObj != NIL_RTR0MEMOBJ)
2708	{
2709	/*
2710	* Do some furter input validations before calling IPRT.
2711	* (Cleanup is done indirectly by telling RTR0MemObjFree to include mappings.)
2712	*/
2713	size_t cbMemObj = RTR0MemObjSize(hMemObj);
2714	if ( offSub < cbMemObj
2715	&& cbSub <= cbMemObj
2716	&& offSub + cbSub <= cbMemObj)
2717	{
2718	RTR0MEMOBJ hMapObj;
2719	rc = RTR0MemObjMapKernelEx(&hMapObj, hMemObj, (void *)-1, 0,
2720	RTMEM_PROT_READ \| RTMEM_PROT_WRITE, offSub, cbSub);
2721	if (RT_SUCCESS(rc))
2722	*ppvR0 = RTR0MemObjAddress(hMapObj);
2723	}
2724	else
2725	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
2726
2727	}
2728	return rc;
2729	}
2730
2731
2732	/**
2733	* Changes the page level protection of one or more pages previously allocated
2734	* by SUPR0PageAllocEx.
2735	*
2736	* @returns IPRT status code.
2737	* @param pSession The session to associated the allocation with.
2738	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
2739	* NIL_RTR3PTR if the ring-3 mapping should be unaffected.
2740	* @param pvR0 The ring-0 address returned by SUPR0PageAllocEx.
2741	* NIL_RTR0PTR if the ring-0 mapping should be unaffected.
2742	* @param offSub Where to start changing. Must be page aligned.
2743	* @param cbSub How much to change. Must be page aligned.
2744	* @param fProt The new page level protection, see RTMEM_PROT_*.
2745	*/
2746	SUPR0DECL(int) SUPR0PageProtect(PSUPDRVSESSION pSession, RTR3PTR pvR3, RTR0PTR pvR0, uint32_t offSub, uint32_t cbSub, uint32_t fProt)
2747	{
2748	int rc;
2749	PSUPDRVBUNDLE pBundle;
2750	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2751	RTR0MEMOBJ hMemObjR0 = NIL_RTR0MEMOBJ;
2752	RTR0MEMOBJ hMemObjR3 = NIL_RTR0MEMOBJ;
2753	LogFlow(("SUPR0PageProtect: pSession=%p pvR3=%p pvR0=%p offSub=%#x cbSub=%#x fProt-%#x\n", pSession, pvR3, pvR0, offSub, cbSub, fProt));
2754
2755	/*
2756	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2757	*/
2758	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2759	AssertReturn(!(fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)), VERR_INVALID_PARAMETER);
2760	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2761	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2762	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
2763
2764	/*
2765	* Find the memory object.
2766	*/
2767	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2768	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2769	{
2770	if (pBundle->cUsed > 0)
2771	{
2772	unsigned i;
2773	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2774	{
2775	if ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
2776	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2777	&& ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2778	\|\| pvR3 == NIL_RTR3PTR)
2779	&& ( pvR0 == NIL_RTR0PTR
2780	\|\| RTR0MemObjAddress(pBundle->aMem[i].MemObj) == pvR0)
2781	&& ( pvR3 == NIL_RTR3PTR
2782	\|\| RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3))
2783	{
2784	if (pvR0 != NIL_RTR0PTR)
2785	hMemObjR0 = pBundle->aMem[i].MemObj;
2786	if (pvR3 != NIL_RTR3PTR)
2787	hMemObjR3 = pBundle->aMem[i].MapObjR3;
2788	break;
2789	}
2790	}
2791	}
2792	}
2793	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2794
2795	rc = VERR_INVALID_PARAMETER;
2796	if ( hMemObjR0 != NIL_RTR0MEMOBJ
2797	\|\| hMemObjR3 != NIL_RTR0MEMOBJ)
2798	{
2799	/*
2800	* Do some furter input validations before calling IPRT.
2801	*/
2802	size_t cbMemObj = hMemObjR0 != NIL_RTR0PTR ? RTR0MemObjSize(hMemObjR0) : RTR0MemObjSize(hMemObjR3);
2803	if ( offSub < cbMemObj
2804	&& cbSub <= cbMemObj
2805	&& offSub + cbSub <= cbMemObj)
2806	{
2807	rc = VINF_SUCCESS;
2808	if (hMemObjR3 != NIL_RTR0PTR)
2809	rc = RTR0MemObjProtect(hMemObjR3, offSub, cbSub, fProt);
2810	if (hMemObjR0 != NIL_RTR0PTR && RT_SUCCESS(rc))
2811	rc = RTR0MemObjProtect(hMemObjR0, offSub, cbSub, fProt);
2812	}
2813	else
2814	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
2815
2816	}
2817	return rc;
2818
2819	}
2820
2821
2822	/**
2823	* Free memory allocated by SUPR0PageAlloc() and SUPR0PageAllocEx().
2824	*
2825	* @returns IPRT status code.
2826	* @param pSession The session owning the allocation.
2827	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc() or
2828	* SUPR0PageAllocEx().
2829	*/
2830	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2831	{
2832	LogFlow(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2833	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2834	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_PAGE);
2835	}
2836
2837
2838	/**
2839	* Gets the paging mode of the current CPU.
2840	*
2841	* @returns Paging mode, SUPPAGEINGMODE_INVALID on error.
2842	*/
2843	SUPR0DECL(SUPPAGINGMODE) SUPR0GetPagingMode(void)
2844	{
2845	SUPPAGINGMODE enmMode;
2846
2847	RTR0UINTREG cr0 = ASMGetCR0();
2848	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
2849	enmMode = SUPPAGINGMODE_INVALID;
2850	else
2851	{
2852	RTR0UINTREG cr4 = ASMGetCR4();
2853	uint32_t fNXEPlusLMA = 0;
2854	if (cr4 & X86_CR4_PAE)
2855	{
2856	uint32_t fAmdFeatures = ASMCpuId_EDX(0x80000001);
2857	if (fAmdFeatures & (X86_CPUID_AMD_FEATURE_EDX_NX \| X86_CPUID_AMD_FEATURE_EDX_LONG_MODE))
2858	{
2859	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
2860	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
2861	fNXEPlusLMA \|= RT_BIT(0);
2862	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
2863	fNXEPlusLMA \|= RT_BIT(1);
2864	}
2865	}
2866
2867	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
2868	{
2869	case 0:
2870	enmMode = SUPPAGINGMODE_32_BIT;
2871	break;
2872
2873	case X86_CR4_PGE:
2874	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
2875	break;
2876
2877	case X86_CR4_PAE:
2878	enmMode = SUPPAGINGMODE_PAE;
2879	break;
2880
2881	case X86_CR4_PAE \| RT_BIT(0):
2882	enmMode = SUPPAGINGMODE_PAE_NX;
2883	break;
2884
2885	case X86_CR4_PAE \| X86_CR4_PGE:
2886	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
2887	break;
2888
2889	case X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
2890	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
2891	break;
2892
2893	case RT_BIT(1) \| X86_CR4_PAE:
2894	enmMode = SUPPAGINGMODE_AMD64;
2895	break;
2896
2897	case RT_BIT(1) \| X86_CR4_PAE \| RT_BIT(0):
2898	enmMode = SUPPAGINGMODE_AMD64_NX;
2899	break;
2900
2901	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
2902	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
2903	break;
2904
2905	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
2906	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
2907	break;
2908
2909	default:
2910	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
2911	enmMode = SUPPAGINGMODE_INVALID;
2912	break;
2913	}
2914	}
2915	return enmMode;
2916	}
2917
2918
2919	/**
2920	* Enables or disabled hardware virtualization extensions using native OS APIs.
2921	*
2922	* @returns VBox status code.
2923	* @retval VINF_SUCCESS on success.
2924	* @retval VERR_NOT_SUPPORTED if not supported by the native OS.
2925	*
2926	* @param fEnable Whether to enable or disable.
2927	*/
2928	SUPR0DECL(int) SUPR0EnableVTx(bool fEnable)
2929	{
2930	#ifdef RT_OS_DARWIN
2931	return supdrvOSEnableVTx(fEnable);
2932	#else
2933	return VERR_NOT_SUPPORTED;
2934	#endif
2935	}
2936
2937
2938	/** @todo document me */
2939	SUPR0DECL(int) SUPR0QueryVTCaps(PSUPDRVSESSION pSession, uint32_t *pfCaps)
2940	{
2941	/*
2942	* Input validation.
2943	*/
2944	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2945	AssertPtrReturn(pfCaps, VERR_INVALID_POINTER);
2946
2947	*pfCaps = 0;
2948
2949	if (ASMHasCpuId())
2950	{
2951	uint32_t u32FeaturesECX;
2952	uint32_t u32Dummy;
2953	uint32_t u32FeaturesEDX;
2954	uint32_t u32VendorEBX, u32VendorECX, u32VendorEDX, u32AMDFeatureEDX, u32AMDFeatureECX;
2955	uint64_t val;
2956
2957	ASMCpuId(0, &u32Dummy, &u32VendorEBX, &u32VendorECX, &u32VendorEDX);
2958	ASMCpuId(1, &u32Dummy, &u32Dummy, &u32FeaturesECX, &u32FeaturesEDX);
2959	/* Query AMD features. */
2960	ASMCpuId(0x80000001, &u32Dummy, &u32Dummy, &u32AMDFeatureECX, &u32AMDFeatureEDX);
2961
2962	if ( u32VendorEBX == X86_CPUID_VENDOR_INTEL_EBX
2963	&& u32VendorECX == X86_CPUID_VENDOR_INTEL_ECX
2964	&& u32VendorEDX == X86_CPUID_VENDOR_INTEL_EDX
2965	)
2966	{
2967	if ( (u32FeaturesECX & X86_CPUID_FEATURE_ECX_VMX)
2968	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
2969	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
2970	)
2971	{
2972	val = ASMRdMsr(MSR_IA32_FEATURE_CONTROL);
2973	/*
2974	* Both the LOCK and VMXON bit must be set; otherwise VMXON will generate a #GP.
2975	* Once the lock bit is set, this MSR can no longer be modified.
2976	*/
2977	if ( (val & (MSR_IA32_FEATURE_CONTROL_VMXON\|MSR_IA32_FEATURE_CONTROL_LOCK))
2978	== (MSR_IA32_FEATURE_CONTROL_VMXON\|MSR_IA32_FEATURE_CONTROL_LOCK) /* enabled and locked */
2979	\|\| !(val & MSR_IA32_FEATURE_CONTROL_LOCK) /* not enabled, but not locked either */
2980	)
2981	{
2982	VMX_CAPABILITY vtCaps;
2983
2984	*pfCaps \|= SUPVTCAPS_VT_X;
2985
2986	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS);
2987	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC_USE_SECONDARY_EXEC_CTRL)
2988	{
2989	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS2);
2990	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC2_EPT)
2991	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
2992	}
2993	return VINF_SUCCESS;
2994	}
2995	return VERR_VMX_MSR_LOCKED_OR_DISABLED;
2996	}
2997	return VERR_VMX_NO_VMX;
2998	}
2999
3000	if ( u32VendorEBX == X86_CPUID_VENDOR_AMD_EBX
3001	&& u32VendorECX == X86_CPUID_VENDOR_AMD_ECX
3002	&& u32VendorEDX == X86_CPUID_VENDOR_AMD_EDX
3003	)
3004	{
3005	if ( (u32AMDFeatureECX & X86_CPUID_AMD_FEATURE_ECX_SVM)
3006	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3007	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3008	)
3009	{
3010	/* Check if SVM is disabled */
3011	val = ASMRdMsr(MSR_K8_VM_CR);
3012	if (!(val & MSR_K8_VM_CR_SVM_DISABLE))
3013	{
3014	*pfCaps \|= SUPVTCAPS_AMD_V;
3015
3016	/* Query AMD features. */
3017	ASMCpuId(0x8000000A, &u32Dummy, &u32Dummy, &u32Dummy, &u32FeaturesEDX);
3018
3019	if (u32FeaturesEDX & AMD_CPUID_SVM_FEATURE_EDX_NESTED_PAGING)
3020	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3021
3022	return VINF_SUCCESS;
3023	}
3024	return VERR_SVM_DISABLED;
3025	}
3026	return VERR_SVM_NO_SVM;
3027	}
3028	}
3029
3030	return VERR_UNSUPPORTED_CPU;
3031	}
3032
3033
3034	/**
3035	* (Re-)initializes the per-cpu structure prior to starting or resuming the GIP
3036	* updating.
3037	*
3038	* @param pGipCpu The per CPU structure for this CPU.
3039	* @param u64NanoTS The current time.
3040	*/
3041	static void supdrvGipReInitCpu(PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
3042	{
3043	pGipCpu->u64TSC = ASMReadTSC() - pGipCpu->u32UpdateIntervalTSC;
3044	pGipCpu->u64NanoTS = u64NanoTS;
3045	}
3046
3047
3048	/**
3049	* Set the current TSC and NanoTS value for the CPU.
3050	*
3051	* @param idCpu The CPU ID. Unused - we have to use the APIC ID.
3052	* @param pvUser1 Pointer to the ring-0 GIP mapping.
3053	* @param pvUser2 Pointer to the variable holding the current time.
3054	*/
3055	static DECLCALLBACK(void) supdrvGipReInitCpuCallback(RTCPUID idCpu, void pvUser1, void pvUser2)
3056	{
3057	PSUPGLOBALINFOPAGE pGip = (PSUPGLOBALINFOPAGE)pvUser1;
3058	unsigned iCpu = ASMGetApicId();
3059
3060	if (RT_LIKELY(iCpu < RT_ELEMENTS(pGip->aCPUs)))
3061	supdrvGipReInitCpu(&pGip->aCPUs[iCpu], (uint64_t )pvUser2);
3062
3063	NOREF(pvUser2);
3064	NOREF(idCpu);
3065	}
3066
3067
3068	/**
3069	* Maps the GIP into userspace and/or get the physical address of the GIP.
3070	*
3071	* @returns IPRT status code.
3072	* @param pSession Session to which the GIP mapping should belong.
3073	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
3074	* @param pHCPhysGip Where to store the physical address. (optional)
3075	*
3076	* @remark There is no reference counting on the mapping, so one call to this function
3077	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
3078	* and remove the session as a GIP user.
3079	*/
3080	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
3081	{
3082	int rc;
3083	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3084	RTR3PTR pGipR3 = NIL_RTR3PTR;
3085	RTHCPHYS HCPhys = NIL_RTHCPHYS;
3086	LogFlow(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
3087
3088	/*
3089	* Validate
3090	*/
3091	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3092	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
3093	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
3094
3095	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3096	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3097	#else
3098	RTSemFastMutexRequest(pDevExt->mtxGip);
3099	#endif
3100	if (pDevExt->pGip)
3101	{
3102	/*
3103	* Map it?
3104	*/
3105	rc = VINF_SUCCESS;
3106	if (ppGipR3)
3107	{
3108	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
3109	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
3110	RTMEM_PROT_READ, RTR0ProcHandleSelf());
3111	if (RT_SUCCESS(rc))
3112	pGipR3 = RTR0MemObjAddressR3(pSession->GipMapObjR3);
3113	}
3114
3115	/*
3116	* Get physical address.
3117	*/
3118	if (pHCPhysGip && RT_SUCCESS(rc))
3119	HCPhys = pDevExt->HCPhysGip;
3120
3121	/*
3122	* Reference globally.
3123	*/
3124	if (!pSession->fGipReferenced && RT_SUCCESS(rc))
3125	{
3126	pSession->fGipReferenced = 1;
3127	pDevExt->cGipUsers++;
3128	if (pDevExt->cGipUsers == 1)
3129	{
3130	PSUPGLOBALINFOPAGE pGipR0 = pDevExt->pGip;
3131	uint64_t u64NanoTS;
3132	unsigned i;
3133
3134	LogFlow(("SUPR0GipMap: Resumes GIP updating\n"));
3135
3136	if (pGipR0->aCPUs[0].u32TransactionId != 2 /* not the first time */)
3137	{
3138	for (i = 0; i < RT_ELEMENTS(pGipR0->aCPUs); i++)
3139	ASMAtomicUoWriteU32(&pGipR0->aCPUs[i].u32TransactionId,
3140	(pGipR0->aCPUs[i].u32TransactionId + GIP_UPDATEHZ_RECALC_FREQ * 2)
3141	& ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
3142	ASMAtomicWriteU64(&pGipR0->u64NanoTSLastUpdateHz, 0);
3143	}
3144
3145	u64NanoTS = RTTimeSystemNanoTS() - pGipR0->u32UpdateIntervalNS;
3146	if ( pGipR0->u32Mode == SUPGIPMODE_SYNC_TSC
3147	\|\| RTMpGetOnlineCount() == 1)
3148	supdrvGipReInitCpu(&pGipR0->aCPUs[0], u64NanoTS);
3149	else
3150	RTMpOnAll(supdrvGipReInitCpuCallback, pGipR0, &u64NanoTS);
3151
3152	rc = RTTimerStart(pDevExt->pGipTimer, 0);
3153	AssertRC(rc); rc = VINF_SUCCESS;
3154	}
3155	}
3156	}
3157	else
3158	{
3159	rc = VERR_GENERAL_FAILURE;
3160	Log(("SUPR0GipMap: GIP is not available!\n"));
3161	}
3162	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3163	RTSemMutexRelease(pDevExt->mtxGip);
3164	#else
3165	RTSemFastMutexRelease(pDevExt->mtxGip);
3166	#endif
3167
3168	/*
3169	* Write returns.
3170	*/
3171	if (pHCPhysGip)
3172	*pHCPhysGip = HCPhys;
3173	if (ppGipR3)
3174	*ppGipR3 = pGipR3;
3175
3176	#ifdef DEBUG_DARWIN_GIP
3177	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3178	#else
3179	LogFlow(( "SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3180	#endif
3181	return rc;
3182	}
3183
3184
3185	/**
3186	* Unmaps any user mapping of the GIP and terminates all GIP access
3187	* from this session.
3188	*
3189	* @returns IPRT status code.
3190	* @param pSession Session to which the GIP mapping should belong.
3191	*/
3192	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
3193	{
3194	int rc = VINF_SUCCESS;
3195	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3196	#ifdef DEBUG_DARWIN_GIP
3197	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
3198	pSession,
3199	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
3200	pSession->GipMapObjR3));
3201	#else
3202	LogFlow(("SUPR0GipUnmap: pSession=%p\n", pSession));
3203	#endif
3204	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3205
3206	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3207	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3208	#else
3209	RTSemFastMutexRequest(pDevExt->mtxGip);
3210	#endif
3211
3212	/*
3213	* Unmap anything?
3214	*/
3215	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
3216	{
3217	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
3218	AssertRC(rc);
3219	if (RT_SUCCESS(rc))
3220	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
3221	}
3222
3223	/*
3224	* Dereference global GIP.
3225	*/
3226	if (pSession->fGipReferenced && !rc)
3227	{
3228	pSession->fGipReferenced = 0;
3229	if ( pDevExt->cGipUsers > 0
3230	&& !--pDevExt->cGipUsers)
3231	{
3232	LogFlow(("SUPR0GipUnmap: Suspends GIP updating\n"));
3233	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = VINF_SUCCESS;
3234	}
3235	}
3236
3237	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3238	RTSemMutexRelease(pDevExt->mtxGip);
3239	#else
3240	RTSemFastMutexRelease(pDevExt->mtxGip);
3241	#endif
3242
3243	return rc;
3244	}
3245
3246
3247	/**
3248	* Gets the GIP pointer.
3249	*
3250	* @returns Pointer to the GIP or NULL.
3251	*/
3252	SUPDECL(PSUPGLOBALINFOPAGE) SUPGetGIP(void)
3253	{
3254	return g_pSUPGlobalInfoPage;
3255	}
3256
3257
3258	/**
3259	* Register a component factory with the support driver.
3260	*
3261	* This is currently restricted to kernel sessions only.
3262	*
3263	* @returns VBox status code.
3264	* @retval VINF_SUCCESS on success.
3265	* @retval VERR_NO_MEMORY if we're out of memory.
3266	* @retval VERR_ALREADY_EXISTS if the factory has already been registered.
3267	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3268	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3269	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3270	*
3271	* @param pSession The SUPDRV session (must be a ring-0 session).
3272	* @param pFactory Pointer to the component factory registration structure.
3273	*
3274	* @remarks This interface is also available via SUPR0IdcComponentRegisterFactory.
3275	*/
3276	SUPR0DECL(int) SUPR0ComponentRegisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3277	{
3278	PSUPDRVFACTORYREG pNewReg;
3279	const char *psz;
3280	int rc;
3281
3282	/*
3283	* Validate parameters.
3284	*/
3285	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3286	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3287	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3288	AssertPtrReturn(pFactory->pfnQueryFactoryInterface, VERR_INVALID_POINTER);
3289	psz = (const char *)memchr(pFactory->szName, '\0', sizeof(pFactory->szName));
3290	AssertReturn(psz, VERR_INVALID_PARAMETER);
3291
3292	/*
3293	* Allocate and initialize a new registration structure.
3294	*/
3295	pNewReg = (PSUPDRVFACTORYREG)RTMemAlloc(sizeof(SUPDRVFACTORYREG));
3296	if (pNewReg)
3297	{
3298	pNewReg->pNext = NULL;
3299	pNewReg->pFactory = pFactory;
3300	pNewReg->pSession = pSession;
3301	pNewReg->cchName = psz - &pFactory->szName[0];
3302
3303	/*
3304	* Add it to the tail of the list after checking for prior registration.
3305	*/
3306	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3307	if (RT_SUCCESS(rc))
3308	{
3309	PSUPDRVFACTORYREG pPrev = NULL;
3310	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3311	while (pCur && pCur->pFactory != pFactory)
3312	{
3313	pPrev = pCur;
3314	pCur = pCur->pNext;
3315	}
3316	if (!pCur)
3317	{
3318	if (pPrev)
3319	pPrev->pNext = pNewReg;
3320	else
3321	pSession->pDevExt->pComponentFactoryHead = pNewReg;
3322	rc = VINF_SUCCESS;
3323	}
3324	else
3325	rc = VERR_ALREADY_EXISTS;
3326
3327	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3328	}
3329
3330	if (RT_FAILURE(rc))
3331	RTMemFree(pNewReg);
3332	}
3333	else
3334	rc = VERR_NO_MEMORY;
3335	return rc;
3336	}
3337
3338
3339	/**
3340	* Deregister a component factory.
3341	*
3342	* @returns VBox status code.
3343	* @retval VINF_SUCCESS on success.
3344	* @retval VERR_NOT_FOUND if the factory wasn't registered.
3345	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3346	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3347	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3348	*
3349	* @param pSession The SUPDRV session (must be a ring-0 session).
3350	* @param pFactory Pointer to the component factory registration structure
3351	* previously passed SUPR0ComponentRegisterFactory().
3352	*
3353	* @remarks This interface is also available via SUPR0IdcComponentDeregisterFactory.
3354	*/
3355	SUPR0DECL(int) SUPR0ComponentDeregisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3356	{
3357	int rc;
3358
3359	/*
3360	* Validate parameters.
3361	*/
3362	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3363	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3364	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3365
3366	/*
3367	* Take the lock and look for the registration record.
3368	*/
3369	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3370	if (RT_SUCCESS(rc))
3371	{
3372	PSUPDRVFACTORYREG pPrev = NULL;
3373	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3374	while (pCur && pCur->pFactory != pFactory)
3375	{
3376	pPrev = pCur;
3377	pCur = pCur->pNext;
3378	}
3379	if (pCur)
3380	{
3381	if (!pPrev)
3382	pSession->pDevExt->pComponentFactoryHead = pCur->pNext;
3383	else
3384	pPrev->pNext = pCur->pNext;
3385
3386	pCur->pNext = NULL;
3387	pCur->pFactory = NULL;
3388	pCur->pSession = NULL;
3389	rc = VINF_SUCCESS;
3390	}
3391	else
3392	rc = VERR_NOT_FOUND;
3393
3394	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3395
3396	RTMemFree(pCur);
3397	}
3398	return rc;
3399	}
3400
3401
3402	/**
3403	* Queries a component factory.
3404	*
3405	* @returns VBox status code.
3406	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3407	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3408	* @retval VERR_SUPDRV_COMPONENT_NOT_FOUND if the component factory wasn't found.
3409	* @retval VERR_SUPDRV_INTERFACE_NOT_SUPPORTED if the interface wasn't supported.
3410	*
3411	* @param pSession The SUPDRV session.
3412	* @param pszName The name of the component factory.
3413	* @param pszInterfaceUuid The UUID of the factory interface (stringified).
3414	* @param ppvFactoryIf Where to store the factory interface.
3415	*/
3416	SUPR0DECL(int) SUPR0ComponentQueryFactory(PSUPDRVSESSION pSession, const char pszName, const char pszInterfaceUuid, void **ppvFactoryIf)
3417	{
3418	const char *pszEnd;
3419	size_t cchName;
3420	int rc;
3421
3422	/*
3423	* Validate parameters.
3424	*/
3425	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3426
3427	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
3428	pszEnd = memchr(pszName, '\0', RT_SIZEOFMEMB(SUPDRVFACTORY, szName));
3429	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3430	cchName = pszEnd - pszName;
3431
3432	AssertPtrReturn(pszInterfaceUuid, VERR_INVALID_POINTER);
3433	pszEnd = memchr(pszInterfaceUuid, '\0', RTUUID_STR_LENGTH);
3434	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3435
3436	AssertPtrReturn(ppvFactoryIf, VERR_INVALID_POINTER);
3437	*ppvFactoryIf = NULL;
3438
3439	/*
3440	* Take the lock and try all factories by this name.
3441	*/
3442	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3443	if (RT_SUCCESS(rc))
3444	{
3445	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3446	rc = VERR_SUPDRV_COMPONENT_NOT_FOUND;
3447	while (pCur)
3448	{
3449	if ( pCur->cchName == cchName
3450	&& !memcmp(pCur->pFactory->szName, pszName, cchName))
3451	{
3452	void *pvFactory = pCur->pFactory->pfnQueryFactoryInterface(pCur->pFactory, pSession, pszInterfaceUuid);
3453	if (pvFactory)
3454	{
3455	*ppvFactoryIf = pvFactory;
3456	rc = VINF_SUCCESS;
3457	break;
3458	}
3459	rc = VERR_SUPDRV_INTERFACE_NOT_SUPPORTED;
3460	}
3461
3462	/* next */
3463	pCur = pCur->pNext;
3464	}
3465
3466	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3467	}
3468	return rc;
3469	}
3470
3471
3472	/**
3473	* Adds a memory object to the session.
3474	*
3475	* @returns IPRT status code.
3476	* @param pMem Memory tracking structure containing the
3477	* information to track.
3478	* @param pSession The session.
3479	*/
3480	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
3481	{
3482	PSUPDRVBUNDLE pBundle;
3483	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3484
3485	/*
3486	* Find free entry and record the allocation.
3487	*/
3488	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3489	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3490	{
3491	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
3492	{
3493	unsigned i;
3494	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3495	{
3496	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
3497	{
3498	pBundle->cUsed++;
3499	pBundle->aMem[i] = *pMem;
3500	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3501	return VINF_SUCCESS;
3502	}
3503	}
3504	AssertFailed(); /* !!this can't be happening!!! */
3505	}
3506	}
3507	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3508
3509	/*
3510	* Need to allocate a new bundle.
3511	* Insert into the last entry in the bundle.
3512	*/
3513	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
3514	if (!pBundle)
3515	return VERR_NO_MEMORY;
3516
3517	/* take last entry. */
3518	pBundle->cUsed++;
3519	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
3520
3521	/* insert into list. */
3522	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3523	pBundle->pNext = pSession->Bundle.pNext;
3524	pSession->Bundle.pNext = pBundle;
3525	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3526
3527	return VINF_SUCCESS;
3528	}
3529
3530
3531	/**
3532	* Releases a memory object referenced by pointer and type.
3533	*
3534	* @returns IPRT status code.
3535	* @param pSession Session data.
3536	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
3537	* @param eType Memory type.
3538	*/
3539	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
3540	{
3541	PSUPDRVBUNDLE pBundle;
3542	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3543
3544	/*
3545	* Validate input.
3546	*/
3547	if (!uPtr)
3548	{
3549	Log(("Illegal address %p\n", (void *)uPtr));
3550	return VERR_INVALID_PARAMETER;
3551	}
3552
3553	/*
3554	* Search for the address.
3555	*/
3556	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3557	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3558	{
3559	if (pBundle->cUsed > 0)
3560	{
3561	unsigned i;
3562	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3563	{
3564	if ( pBundle->aMem[i].eType == eType
3565	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3566	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
3567	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3568	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
3569	)
3570	{
3571	/* Make a copy of it and release it outside the spinlock. */
3572	SUPDRVMEMREF Mem = pBundle->aMem[i];
3573	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
3574	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
3575	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
3576	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3577
3578	if (Mem.MapObjR3 != NIL_RTR0MEMOBJ)
3579	{
3580	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
3581	AssertRC(rc); /** @todo figure out how to handle this. */
3582	}
3583	if (Mem.MemObj != NIL_RTR0MEMOBJ)
3584	{
3585	int rc = RTR0MemObjFree(Mem.MemObj, true /* fFreeMappings */);
3586	AssertRC(rc); /** @todo figure out how to handle this. */
3587	}
3588	return VINF_SUCCESS;
3589	}
3590	}
3591	}
3592	}
3593	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3594	Log(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
3595	return VERR_INVALID_PARAMETER;
3596	}
3597
3598
3599	/**
3600	* Opens an image. If it's the first time it's opened the call must upload
3601	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
3602	*
3603	* This is the 1st step of the loading.
3604	*
3605	* @returns IPRT status code.
3606	* @param pDevExt Device globals.
3607	* @param pSession Session data.
3608	* @param pReq The open request.
3609	*/
3610	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
3611	{
3612	int rc;
3613	PSUPDRVLDRIMAGE pImage;
3614	void *pv;
3615	size_t cchName = strlen(pReq->u.In.szName); /* (caller checked < 32). */
3616	LogFlow(("supdrvIOCtl_LdrOpen: szName=%s cbImageWithTabs=%d\n", pReq->u.In.szName, pReq->u.In.cbImageWithTabs));
3617
3618	/*
3619	* Check if we got an instance of the image already.
3620	*/
3621	supdrvLdrLock(pDevExt);
3622	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
3623	{
3624	if ( pImage->szName[cchName] == '\0'
3625	&& !memcmp(pImage->szName, pReq->u.In.szName, cchName))
3626	{
3627	/** @todo check cbImageBits and cbImageWithTabs here, if they differs that indicates that the images are different. */
3628	pImage->cUsage++;
3629	pReq->u.Out.pvImageBase = pImage->pvImage;
3630	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
3631	pReq->u.Out.fNativeLoader = pImage->fNative;
3632	supdrvLdrAddUsage(pSession, pImage);
3633	supdrvLdrUnlock(pDevExt);
3634	return VINF_SUCCESS;
3635	}
3636	}
3637	/* (not found - add it!) */
3638
3639	/*
3640	* Allocate memory.
3641	*/
3642	pv = RTMemAlloc(RT_OFFSETOF(SUPDRVLDRIMAGE, szName[cchName + 1]));
3643	if (!pv)
3644	{
3645	supdrvLdrUnlock(pDevExt);
3646	Log(("supdrvIOCtl_LdrOpen: RTMemAlloc() failed\n"));
3647	return VERR_NO_MEMORY;
3648	}
3649
3650	/*
3651	* Setup and link in the LDR stuff.
3652	*/
3653	pImage = (PSUPDRVLDRIMAGE)pv;
3654	pImage->pvImage = NULL;
3655	pImage->pvImageAlloc = NULL;
3656	pImage->cbImageWithTabs = pReq->u.In.cbImageWithTabs;
3657	pImage->cbImageBits = pReq->u.In.cbImageBits;
3658	pImage->cSymbols = 0;
3659	pImage->paSymbols = NULL;
3660	pImage->pachStrTab = NULL;
3661	pImage->cbStrTab = 0;
3662	pImage->pfnModuleInit = NULL;
3663	pImage->pfnModuleTerm = NULL;
3664	pImage->pfnServiceReqHandler = NULL;
3665	pImage->uState = SUP_IOCTL_LDR_OPEN;
3666	pImage->cUsage = 1;
3667	memcpy(pImage->szName, pReq->u.In.szName, cchName + 1);
3668
3669	/*
3670	* Try load it using the native loader, if that isn't supported, fall back
3671	* on the older method.
3672	*/
3673	pImage->fNative = true;
3674	rc = supdrvOSLdrOpen(pDevExt, pImage, pReq->u.In.szFilename);
3675	if (rc == VERR_NOT_SUPPORTED)
3676	{
3677	pImage->pvImageAlloc = RTMemExecAlloc(pImage->cbImageBits + 31);
3678	pImage->pvImage = RT_ALIGN_P(pImage->pvImageAlloc, 32);
3679	pImage->fNative = false;
3680	rc = pImage->pvImageAlloc ? VINF_SUCCESS : VERR_NO_MEMORY;
3681	}
3682	if (RT_FAILURE(rc))
3683	{
3684	supdrvLdrUnlock(pDevExt);
3685	RTMemFree(pImage);
3686	Log(("supdrvIOCtl_LdrOpen(%s): failed - %Rrc\n", pReq->u.In.szName, rc));
3687	return rc;
3688	}
3689	Assert(VALID_PTR(pImage->pvImage) \|\| RT_FAILURE(rc));
3690
3691	/*
3692	* Link it.
3693	*/
3694	pImage->pNext = pDevExt->pLdrImages;
3695	pDevExt->pLdrImages = pImage;
3696
3697	supdrvLdrAddUsage(pSession, pImage);
3698
3699	pReq->u.Out.pvImageBase = pImage->pvImage;
3700	pReq->u.Out.fNeedsLoading = true;
3701	pReq->u.Out.fNativeLoader = pImage->fNative;
3702	supdrvLdrUnlock(pDevExt);
3703
3704	#if defined(RT_OS_WINDOWS) && defined(DEBUG)
3705	SUPR0Printf("VBoxDrv: windbg> .reload /f %s=%#p\n", pImage->szName, pImage->pvImage);
3706	#endif
3707	return VINF_SUCCESS;
3708	}
3709
3710
3711	/**
3712	* Worker that validates a pointer to an image entrypoint.
3713	*
3714	* @returns IPRT status code.
3715	* @param pDevExt The device globals.
3716	* @param pImage The loader image.
3717	* @param pv The pointer into the image.
3718	* @param fMayBeNull Whether it may be NULL.
3719	* @param pszWhat What is this entrypoint? (for logging)
3720	* @param pbImageBits The image bits prepared by ring-3.
3721	*
3722	* @remarks Will leave the lock on failure.
3723	*/
3724	static int supdrvLdrValidatePointer(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage, void *pv,
3725	bool fMayBeNull, const uint8_t pbImageBits, const char pszWhat)
3726	{
3727	if (!fMayBeNull \|\| pv)
3728	{
3729	if ((uintptr_t)pv - (uintptr_t)pImage->pvImage >= pImage->cbImageBits)
3730	{
3731	supdrvLdrUnlock(pDevExt);
3732	Log(("Out of range (%p LB %#x): %s=%p\n", pImage->pvImage, pImage->cbImageBits, pszWhat, pv));
3733	return VERR_INVALID_PARAMETER;
3734	}
3735
3736	if (pImage->fNative)
3737	{
3738	int rc = supdrvOSLdrValidatePointer(pDevExt, pImage, pv, pbImageBits);
3739	if (RT_FAILURE(rc))
3740	{
3741	supdrvLdrUnlock(pDevExt);
3742	Log(("Bad entry point address: %s=%p (rc=%Rrc)\n", pszWhat, pv, rc));
3743	return rc;
3744	}
3745	}
3746	}
3747	return VINF_SUCCESS;
3748	}
3749
3750
3751	/**
3752	* Loads the image bits.
3753	*
3754	* This is the 2nd step of the loading.
3755	*
3756	* @returns IPRT status code.
3757	* @param pDevExt Device globals.
3758	* @param pSession Session data.
3759	* @param pReq The request.
3760	*/
3761	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
3762	{
3763	PSUPDRVLDRUSAGE pUsage;
3764	PSUPDRVLDRIMAGE pImage;
3765	int rc;
3766	LogFlow(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImageWithBits=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImageWithTabs));
3767
3768	/*
3769	* Find the ldr image.
3770	*/
3771	supdrvLdrLock(pDevExt);
3772	pUsage = pSession->pLdrUsage;
3773	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3774	pUsage = pUsage->pNext;
3775	if (!pUsage)
3776	{
3777	supdrvLdrUnlock(pDevExt);
3778	Log(("SUP_IOCTL_LDR_LOAD: couldn't find image!\n"));
3779	return VERR_INVALID_HANDLE;
3780	}
3781	pImage = pUsage->pImage;
3782
3783	/*
3784	* Validate input.
3785	*/
3786	if ( pImage->cbImageWithTabs != pReq->u.In.cbImageWithTabs
3787	\|\| pImage->cbImageBits != pReq->u.In.cbImageBits)
3788	{
3789	supdrvLdrUnlock(pDevExt);
3790	Log(("SUP_IOCTL_LDR_LOAD: image size mismatch!! %d(prep) != %d(load) or %d != %d\n",
3791	pImage->cbImageWithTabs, pReq->u.In.cbImageWithTabs, pImage->cbImageBits, pReq->u.In.cbImageBits));
3792	return VERR_INVALID_HANDLE;
3793	}
3794
3795	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
3796	{
3797	unsigned uState = pImage->uState;
3798	supdrvLdrUnlock(pDevExt);
3799	if (uState != SUP_IOCTL_LDR_LOAD)
3800	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
3801	return VERR_ALREADY_LOADED;
3802	}
3803
3804	switch (pReq->u.In.eEPType)
3805	{
3806	case SUPLDRLOADEP_NOTHING:
3807	break;
3808
3809	case SUPLDRLOADEP_VMMR0:
3810	rc = supdrvLdrValidatePointer( pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0, false, pReq->u.In.abImage, "pvVMMR0");
3811	if (RT_SUCCESS(rc))
3812	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt, false, pReq->u.In.abImage, "pvVMMR0EntryInt");
3813	if (RT_SUCCESS(rc))
3814	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, false, pReq->u.In.abImage, "pvVMMR0EntryFast");
3815	if (RT_SUCCESS(rc))
3816	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx, false, pReq->u.In.abImage, "pvVMMR0EntryEx");
3817	if (RT_FAILURE(rc))
3818	return rc;
3819	break;
3820
3821	case SUPLDRLOADEP_SERVICE:
3822	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.Service.pfnServiceReq, false, pReq->u.In.abImage, "pfnServiceReq");
3823	if (RT_FAILURE(rc))
3824	return rc;
3825	if ( pReq->u.In.EP.Service.apvReserved[0] != NIL_RTR0PTR
3826	\|\| pReq->u.In.EP.Service.apvReserved[1] != NIL_RTR0PTR
3827	\|\| pReq->u.In.EP.Service.apvReserved[2] != NIL_RTR0PTR)
3828	{
3829	supdrvLdrUnlock(pDevExt);
3830	Log(("Out of range (%p LB %#x): apvReserved={%p,%p,%p} MBZ!\n",
3831	pImage->pvImage, pReq->u.In.cbImageWithTabs,
3832	pReq->u.In.EP.Service.apvReserved[0],
3833	pReq->u.In.EP.Service.apvReserved[1],
3834	pReq->u.In.EP.Service.apvReserved[2]));
3835	return VERR_INVALID_PARAMETER;
3836	}
3837	break;
3838
3839	default:
3840	supdrvLdrUnlock(pDevExt);
3841	Log(("Invalid eEPType=%d\n", pReq->u.In.eEPType));
3842	return VERR_INVALID_PARAMETER;
3843	}
3844
3845	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleInit, true, pReq->u.In.abImage, "pfnModuleInit");
3846	if (RT_FAILURE(rc))
3847	return rc;
3848	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleTerm, true, pReq->u.In.abImage, "pfnModuleTerm");
3849	if (RT_FAILURE(rc))
3850	return rc;
3851
3852	/*
3853	* Allocate and copy the tables.
3854	* (No need to do try/except as this is a buffered request.)
3855	*/
3856	pImage->cbStrTab = pReq->u.In.cbStrTab;
3857	if (pImage->cbStrTab)
3858	{
3859	pImage->pachStrTab = (char *)RTMemAlloc(pImage->cbStrTab);
3860	if (pImage->pachStrTab)
3861	memcpy(pImage->pachStrTab, &pReq->u.In.abImage[pReq->u.In.offStrTab], pImage->cbStrTab);
3862	else
3863	rc = VERR_NO_MEMORY;
3864	}
3865
3866	pImage->cSymbols = pReq->u.In.cSymbols;
3867	if (RT_SUCCESS(rc) && pImage->cSymbols)
3868	{
3869	size_t cbSymbols = pImage->cSymbols * sizeof(SUPLDRSYM);
3870	pImage->paSymbols = (PSUPLDRSYM)RTMemAlloc(cbSymbols);
3871	if (pImage->paSymbols)
3872	memcpy(pImage->paSymbols, &pReq->u.In.abImage[pReq->u.In.offSymbols], cbSymbols);
3873	else
3874	rc = VERR_NO_MEMORY;
3875	}
3876
3877	/*
3878	* Copy the bits / complete native loading.
3879	*/
3880	if (RT_SUCCESS(rc))
3881	{
3882	pImage->uState = SUP_IOCTL_LDR_LOAD;
3883	pImage->pfnModuleInit = pReq->u.In.pfnModuleInit;
3884	pImage->pfnModuleTerm = pReq->u.In.pfnModuleTerm;
3885
3886	if (pImage->fNative)
3887	rc = supdrvOSLdrLoad(pDevExt, pImage, pReq->u.In.abImage);
3888	else
3889	memcpy(pImage->pvImage, &pReq->u.In.abImage[0], pImage->cbImageBits);
3890	}
3891
3892	/*
3893	* Update any entry points.
3894	*/
3895	if (RT_SUCCESS(rc))
3896	{
3897	switch (pReq->u.In.eEPType)
3898	{
3899	default:
3900	case SUPLDRLOADEP_NOTHING:
3901	rc = VINF_SUCCESS;
3902	break;
3903	case SUPLDRLOADEP_VMMR0:
3904	rc = supdrvLdrSetVMMR0EPs(pDevExt, pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
3905	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
3906	break;
3907	case SUPLDRLOADEP_SERVICE:
3908	pImage->pfnServiceReqHandler = pReq->u.In.EP.Service.pfnServiceReq;
3909	rc = VINF_SUCCESS;
3910	break;
3911	}
3912	}
3913
3914	/*
3915	* On success call the module initialization.
3916	*/
3917	LogFlow(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
3918	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
3919	{
3920	Log(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
3921	rc = pImage->pfnModuleInit();
3922	if (rc && pDevExt->pvVMMR0 == pImage->pvImage)
3923	supdrvLdrUnsetVMMR0EPs(pDevExt);
3924	}
3925
3926	if (RT_FAILURE(rc))
3927	{
3928	pImage->uState = SUP_IOCTL_LDR_OPEN;
3929	pImage->pfnModuleInit = NULL;
3930	pImage->pfnModuleTerm = NULL;
3931	pImage->pfnServiceReqHandler= NULL;
3932	pImage->cbStrTab = 0;
3933	RTMemFree(pImage->pachStrTab);
3934	pImage->pachStrTab = NULL;
3935	RTMemFree(pImage->paSymbols);
3936	pImage->paSymbols = NULL;
3937	pImage->cSymbols = 0;
3938	}
3939
3940	supdrvLdrUnlock(pDevExt);
3941	return rc;
3942	}
3943
3944
3945	/**
3946	* Frees a previously loaded (prep'ed) image.
3947	*
3948	* @returns IPRT status code.
3949	* @param pDevExt Device globals.
3950	* @param pSession Session data.
3951	* @param pReq The request.
3952	*/
3953	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
3954	{
3955	int rc;
3956	PSUPDRVLDRUSAGE pUsagePrev;
3957	PSUPDRVLDRUSAGE pUsage;
3958	PSUPDRVLDRIMAGE pImage;
3959	LogFlow(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
3960
3961	/*
3962	* Find the ldr image.
3963	*/
3964	supdrvLdrLock(pDevExt);
3965	pUsagePrev = NULL;
3966	pUsage = pSession->pLdrUsage;
3967	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3968	{
3969	pUsagePrev = pUsage;
3970	pUsage = pUsage->pNext;
3971	}
3972	if (!pUsage)
3973	{
3974	supdrvLdrUnlock(pDevExt);
3975	Log(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
3976	return VERR_INVALID_HANDLE;
3977	}
3978
3979	/*
3980	* Check if we can remove anything.
3981	*/
3982	rc = VINF_SUCCESS;
3983	pImage = pUsage->pImage;
3984	if (pImage->cUsage <= 1 \|\| pUsage->cUsage <= 1)
3985	{
3986	/*
3987	* Check if there are any objects with destructors in the image, if
3988	* so leave it for the session cleanup routine so we get a chance to
3989	* clean things up in the right order and not leave them all dangling.
3990	*/
3991	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3992	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
3993	if (pImage->cUsage <= 1)
3994	{
3995	PSUPDRVOBJ pObj;
3996	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
3997	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
3998	{
3999	rc = VERR_DANGLING_OBJECTS;
4000	break;
4001	}
4002	}
4003	else
4004	{
4005	PSUPDRVUSAGE pGenUsage;
4006	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
4007	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4008	{
4009	rc = VERR_DANGLING_OBJECTS;
4010	break;
4011	}
4012	}
4013	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
4014	if (rc == VINF_SUCCESS)
4015	{
4016	/* unlink it */
4017	if (pUsagePrev)
4018	pUsagePrev->pNext = pUsage->pNext;
4019	else
4020	pSession->pLdrUsage = pUsage->pNext;
4021
4022	/* free it */
4023	pUsage->pImage = NULL;
4024	pUsage->pNext = NULL;
4025	RTMemFree(pUsage);
4026
4027	/*
4028	* Dereference the image.
4029	*/
4030	if (pImage->cUsage <= 1)
4031	supdrvLdrFree(pDevExt, pImage);
4032	else
4033	pImage->cUsage--;
4034	}
4035	else
4036	{
4037	Log(("supdrvIOCtl_LdrFree: Dangling objects in %p/%s!\n", pImage->pvImage, pImage->szName));
4038	rc = VINF_SUCCESS; /** @todo BRANCH-2.1: remove this after branching. */
4039	}
4040	}
4041	else
4042	{
4043	/*
4044	* Dereference both image and usage.
4045	*/
4046	pImage->cUsage--;
4047	pUsage->cUsage--;
4048	}
4049
4050	supdrvLdrUnlock(pDevExt);
4051	return rc;
4052	}
4053
4054
4055	/**
4056	* Gets the address of a symbol in an open image.
4057	*
4058	* @returns IPRT status code.
4059	* @param pDevExt Device globals.
4060	* @param pSession Session data.
4061	* @param pReq The request buffer.
4062	*/
4063	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
4064	{
4065	PSUPDRVLDRIMAGE pImage;
4066	PSUPDRVLDRUSAGE pUsage;
4067	uint32_t i;
4068	PSUPLDRSYM paSyms;
4069	const char *pchStrings;
4070	const size_t cbSymbol = strlen(pReq->u.In.szSymbol) + 1;
4071	void *pvSymbol = NULL;
4072	int rc = VERR_GENERAL_FAILURE;
4073	Log3(("supdrvIOCtl_LdrGetSymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
4074
4075	/*
4076	* Find the ldr image.
4077	*/
4078	supdrvLdrLock(pDevExt);
4079	pUsage = pSession->pLdrUsage;
4080	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4081	pUsage = pUsage->pNext;
4082	if (!pUsage)
4083	{
4084	supdrvLdrUnlock(pDevExt);
4085	Log(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
4086	return VERR_INVALID_HANDLE;
4087	}
4088	pImage = pUsage->pImage;
4089	if (pImage->uState != SUP_IOCTL_LDR_LOAD)
4090	{
4091	unsigned uState = pImage->uState;
4092	supdrvLdrUnlock(pDevExt);
4093	Log(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", uState, uState)); NOREF(uState);
4094	return VERR_ALREADY_LOADED;
4095	}
4096
4097	/*
4098	* Search the symbol strings.
4099	*/
4100	pchStrings = pImage->pachStrTab;
4101	paSyms = pImage->paSymbols;
4102	for (i = 0; i < pImage->cSymbols; i++)
4103	{
4104	if ( paSyms[i].offSymbol < pImage->cbImageBits /* paranoia */
4105	&& paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4106	&& !memcmp(pchStrings + paSyms[i].offName, pReq->u.In.szSymbol, cbSymbol))
4107	{
4108	pvSymbol = (uint8_t *)pImage->pvImage + paSyms[i].offSymbol;
4109	rc = VINF_SUCCESS;
4110	break;
4111	}
4112	}
4113	supdrvLdrUnlock(pDevExt);
4114	pReq->u.Out.pvSymbol = pvSymbol;
4115	return rc;
4116	}
4117
4118
4119	/**
4120	* Gets the address of a symbol in an open image or the support driver.
4121	*
4122	* @returns VINF_SUCCESS on success.
4123	* @returns
4124	* @param pDevExt Device globals.
4125	* @param pSession Session data.
4126	* @param pReq The request buffer.
4127	*/
4128	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq)
4129	{
4130	int rc = VINF_SUCCESS;
4131	const char *pszSymbol = pReq->u.In.pszSymbol;
4132	const char *pszModule = pReq->u.In.pszModule;
4133	size_t cbSymbol;
4134	char const *pszEnd;
4135	uint32_t i;
4136
4137	/*
4138	* Input validation.
4139	*/
4140	AssertPtrReturn(pszSymbol, VERR_INVALID_POINTER);
4141	pszEnd = (char *)memchr(pszSymbol, '\0', 512);
4142	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4143	cbSymbol = pszEnd - pszSymbol + 1;
4144
4145	if (pszModule)
4146	{
4147	AssertPtrReturn(pszModule, VERR_INVALID_POINTER);
4148	pszEnd = (char *)memchr(pszModule, '\0', 64);
4149	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4150	}
4151	Log3(("supdrvIDC_LdrGetSymbol: pszModule=%p:{%s} pszSymbol=%p:{%s}\n", pszModule, pszModule, pszSymbol, pszSymbol));
4152
4153
4154	if ( !pszModule
4155	\|\| !strcmp(pszModule, "SupDrv"))
4156	{
4157	/*
4158	* Search the support driver export table.
4159	*/
4160	for (i = 0; i < RT_ELEMENTS(g_aFunctions); i++)
4161	if (!strcmp(g_aFunctions[i].szName, pszSymbol))
4162	{
4163	pReq->u.Out.pfnSymbol = g_aFunctions[i].pfn;
4164	break;
4165	}
4166	}
4167	else
4168	{
4169	/*
4170	* Find the loader image.
4171	*/
4172	PSUPDRVLDRIMAGE pImage;
4173
4174	supdrvLdrLock(pDevExt);
4175
4176	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
4177	if (!strcmp(pImage->szName, pszModule))
4178	break;
4179	if (pImage && pImage->uState == SUP_IOCTL_LDR_LOAD)
4180	{
4181	/*
4182	* Search the symbol strings.
4183	*/
4184	const char *pchStrings = pImage->pachStrTab;
4185	PCSUPLDRSYM paSyms = pImage->paSymbols;
4186	for (i = 0; i < pImage->cSymbols; i++)
4187	{
4188	if ( paSyms[i].offSymbol < pImage->cbImageBits /* paranoia */
4189	&& paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4190	&& !memcmp(pchStrings + paSyms[i].offName, pszSymbol, cbSymbol))
4191	{
4192	/*
4193	* Found it! Calc the symbol address and add a reference to the module.
4194	*/
4195	pReq->u.Out.pfnSymbol = (PFNRT)((uint8_t *)pImage->pvImage + paSyms[i].offSymbol);
4196	rc = supdrvLdrAddUsage(pSession, pImage);
4197	break;
4198	}
4199	}
4200	}
4201	else
4202	rc = pImage ? VERR_WRONG_ORDER : VERR_MODULE_NOT_FOUND;
4203
4204	supdrvLdrUnlock(pDevExt);
4205	}
4206	return rc;
4207	}
4208
4209
4210	/**
4211	* Updates the VMMR0 entry point pointers.
4212	*
4213	* @returns IPRT status code.
4214	* @param pDevExt Device globals.
4215	* @param pSession Session data.
4216	* @param pVMMR0 VMMR0 image handle.
4217	* @param pvVMMR0EntryInt VMMR0EntryInt address.
4218	* @param pvVMMR0EntryFast VMMR0EntryFast address.
4219	* @param pvVMMR0EntryEx VMMR0EntryEx address.
4220	* @remark Caller must own the loader mutex.
4221	*/
4222	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx)
4223	{
4224	int rc = VINF_SUCCESS;
4225	LogFlow(("supdrvLdrSetR0EP pvVMMR0=%p pvVMMR0EntryInt=%p\n", pvVMMR0, pvVMMR0EntryInt));
4226
4227
4228	/*
4229	* Check if not yet set.
4230	*/
4231	if (!pDevExt->pvVMMR0)
4232	{
4233	pDevExt->pvVMMR0 = pvVMMR0;
4234	pDevExt->pfnVMMR0EntryInt = pvVMMR0EntryInt;
4235	pDevExt->pfnVMMR0EntryFast = pvVMMR0EntryFast;
4236	pDevExt->pfnVMMR0EntryEx = pvVMMR0EntryEx;
4237	}
4238	else
4239	{
4240	/*
4241	* Return failure or success depending on whether the values match or not.
4242	*/
4243	if ( pDevExt->pvVMMR0 != pvVMMR0
4244	\|\| (void *)pDevExt->pfnVMMR0EntryInt != pvVMMR0EntryInt
4245	\|\| (void *)pDevExt->pfnVMMR0EntryFast != pvVMMR0EntryFast
4246	\|\| (void *)pDevExt->pfnVMMR0EntryEx != pvVMMR0EntryEx)
4247	{
4248	AssertMsgFailed(("SUP_IOCTL_LDR_SETR0EP: Already set pointing to a different module!\n"));
4249	rc = VERR_INVALID_PARAMETER;
4250	}
4251	}
4252	return rc;
4253	}
4254
4255
4256	/**
4257	* Unsets the VMMR0 entry point installed by supdrvLdrSetR0EP.
4258	*
4259	* @param pDevExt Device globals.
4260	*/
4261	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt)
4262	{
4263	pDevExt->pvVMMR0 = NULL;
4264	pDevExt->pfnVMMR0EntryInt = NULL;
4265	pDevExt->pfnVMMR0EntryFast = NULL;
4266	pDevExt->pfnVMMR0EntryEx = NULL;
4267	}
4268
4269
4270	/**
4271	* Adds a usage reference in the specified session of an image.
4272	*
4273	* Called while owning the loader semaphore.
4274	*
4275	* @returns VINF_SUCCESS on success and VERR_NO_MEMORY on failure.
4276	* @param pSession Session in question.
4277	* @param pImage Image which the session is using.
4278	*/
4279	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage)
4280	{
4281	PSUPDRVLDRUSAGE pUsage;
4282	LogFlow(("supdrvLdrAddUsage: pImage=%p\n", pImage));
4283
4284	/*
4285	* Referenced it already?
4286	*/
4287	pUsage = pSession->pLdrUsage;
4288	while (pUsage)
4289	{
4290	if (pUsage->pImage == pImage)
4291	{
4292	pUsage->cUsage++;
4293	return VINF_SUCCESS;
4294	}
4295	pUsage = pUsage->pNext;
4296	}
4297
4298	/*
4299	* Allocate new usage record.
4300	*/
4301	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
4302	AssertReturn(pUsage, VERR_NO_MEMORY);
4303	pUsage->cUsage = 1;
4304	pUsage->pImage = pImage;
4305	pUsage->pNext = pSession->pLdrUsage;
4306	pSession->pLdrUsage = pUsage;
4307	return VINF_SUCCESS;
4308	}
4309
4310
4311	/**
4312	* Frees a load image.
4313	*
4314	* @param pDevExt Pointer to device extension.
4315	* @param pImage Pointer to the image we're gonna free.
4316	* This image must exit!
4317	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
4318	*/
4319	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
4320	{
4321	PSUPDRVLDRIMAGE pImagePrev;
4322	LogFlow(("supdrvLdrFree: pImage=%p\n", pImage));
4323
4324	/* find it - arg. should've used doubly linked list. */
4325	Assert(pDevExt->pLdrImages);
4326	pImagePrev = NULL;
4327	if (pDevExt->pLdrImages != pImage)
4328	{
4329	pImagePrev = pDevExt->pLdrImages;
4330	while (pImagePrev->pNext != pImage)
4331	pImagePrev = pImagePrev->pNext;
4332	Assert(pImagePrev->pNext == pImage);
4333	}
4334
4335	/* unlink */
4336	if (pImagePrev)
4337	pImagePrev->pNext = pImage->pNext;
4338	else
4339	pDevExt->pLdrImages = pImage->pNext;
4340
4341	/* check if this is VMMR0.r0 unset its entry point pointers. */
4342	if (pDevExt->pvVMMR0 == pImage->pvImage)
4343	supdrvLdrUnsetVMMR0EPs(pDevExt);
4344
4345	/* check for objects with destructors in this image. (Shouldn't happen.) */
4346	if (pDevExt->pObjs)
4347	{
4348	unsigned cObjs = 0;
4349	PSUPDRVOBJ pObj;
4350	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
4351	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
4352	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4353	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4354	{
4355	pObj->pfnDestructor = NULL;
4356	cObjs++;
4357	}
4358	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
4359	if (cObjs)
4360	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
4361	}
4362
4363	/* call termination function if fully loaded. */
4364	if ( pImage->pfnModuleTerm
4365	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
4366	{
4367	LogFlow(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
4368	pImage->pfnModuleTerm();
4369	}
4370
4371	/* do native unload if appropriate. */
4372	if (pImage->fNative)
4373	supdrvOSLdrUnload(pDevExt, pImage);
4374
4375	/* free the image */
4376	pImage->cUsage = 0;
4377	pImage->pNext = 0;
4378	pImage->uState = SUP_IOCTL_LDR_FREE;
4379	RTMemExecFree(pImage->pvImageAlloc);
4380	pImage->pvImageAlloc = NULL;
4381	RTMemFree(pImage->pachStrTab);
4382	pImage->pachStrTab = NULL;
4383	RTMemFree(pImage->paSymbols);
4384	pImage->paSymbols = NULL;
4385	RTMemFree(pImage);
4386	}
4387
4388
4389	/**
4390	* Acquires the loader lock.
4391	*
4392	* @returns IPRT status code.
4393	* @param pDevExt The device extension.
4394	*/
4395	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt)
4396	{
4397	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4398	int rc = RTSemMutexRequest(pDevExt->mtxLdr, RT_INDEFINITE_WAIT);
4399	#else
4400	int rc = RTSemFastMutexRequest(pDevExt->mtxLdr);
4401	#endif
4402	AssertRC(rc);
4403	return rc;
4404	}
4405
4406
4407	/**
4408	* Releases the loader lock.
4409	*
4410	* @returns IPRT status code.
4411	* @param pDevExt The device extension.
4412	*/
4413	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt)
4414	{
4415	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4416	return RTSemMutexRelease(pDevExt->mtxLdr);
4417	#else
4418	return RTSemFastMutexRelease(pDevExt->mtxLdr);
4419	#endif
4420	}
4421
4422
4423	/**
4424	* Implements the service call request.
4425	*
4426	* @returns VBox status code.
4427	* @param pDevExt The device extension.
4428	* @param pSession The calling session.
4429	* @param pReq The request packet, valid.
4430	*/
4431	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq)
4432	{
4433	#if !defined(RT_OS_WINDOWS) \|\| defined(DEBUG)
4434	int rc;
4435
4436	/*
4437	* Find the module first in the module referenced by the calling session.
4438	*/
4439	rc = supdrvLdrLock(pDevExt);
4440	if (RT_SUCCESS(rc))
4441	{
4442	PFNSUPR0SERVICEREQHANDLER pfnServiceReqHandler = NULL;
4443	PSUPDRVLDRUSAGE pUsage;
4444
4445	for (pUsage = pSession->pLdrUsage; pUsage; pUsage = pUsage->pNext)
4446	if ( pUsage->pImage->pfnServiceReqHandler
4447	&& !strcmp(pUsage->pImage->szName, pReq->u.In.szName))
4448	{
4449	pfnServiceReqHandler = pUsage->pImage->pfnServiceReqHandler;
4450	break;
4451	}
4452	supdrvLdrUnlock(pDevExt);
4453
4454	if (pfnServiceReqHandler)
4455	{
4456	/*
4457	* Call it.
4458	*/
4459	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
4460	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, NULL);
4461	else
4462	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0]);
4463	}
4464	else
4465	rc = VERR_SUPDRV_SERVICE_NOT_FOUND;
4466	}
4467
4468	/* log it */
4469	if ( RT_FAILURE(rc)
4470	&& rc != VERR_INTERRUPTED
4471	&& rc != VERR_TIMEOUT)
4472	Log(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4473	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4474	else
4475	Log4(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4476	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4477	return rc;
4478	#else /* RT_OS_WINDOWS && !DEBUG */
4479	return VERR_NOT_IMPLEMENTED;
4480	#endif /* RT_OS_WINDOWS && !DEBUG */
4481	}
4482
4483
4484	/**
4485	* Implements the logger settings request.
4486	*
4487	* @returns VBox status code.
4488	* @param pDevExt The device extension.
4489	* @param pSession The caller's session.
4490	* @param pReq The request.
4491	*/
4492	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq)
4493	{
4494	const char *pszGroup = &pReq->u.In.szStrings[pReq->u.In.offGroups];
4495	const char *pszFlags = &pReq->u.In.szStrings[pReq->u.In.offFlags];
4496	const char *pszDest = &pReq->u.In.szStrings[pReq->u.In.offDestination];
4497	PRTLOGGER pLogger = NULL;
4498	int rc;
4499
4500	/*
4501	* Some further validation.
4502	*/
4503	switch (pReq->u.In.fWhat)
4504	{
4505	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4506	case SUPLOGGERSETTINGS_WHAT_CREATE:
4507	break;
4508
4509	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4510	if (pszGroup \|\| pszFlags \|\| *pszDest)
4511	return VERR_INVALID_PARAMETER;
4512	if (pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_RELEASE)
4513	return VERR_ACCESS_DENIED;
4514	break;
4515
4516	default:
4517	return VERR_INTERNAL_ERROR;
4518	}
4519
4520	/*
4521	* Get the logger.
4522	*/
4523	switch (pReq->u.In.fWhich)
4524	{
4525	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4526	pLogger = RTLogGetDefaultInstance();
4527	break;
4528
4529	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4530	pLogger = RTLogRelDefaultInstance();
4531	break;
4532
4533	default:
4534	return VERR_INTERNAL_ERROR;
4535	}
4536
4537	/*
4538	* Do the job.
4539	*/
4540	switch (pReq->u.In.fWhat)
4541	{
4542	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4543	if (pLogger)
4544	{
4545	rc = RTLogFlags(pLogger, pszFlags);
4546	if (RT_SUCCESS(rc))
4547	rc = RTLogGroupSettings(pLogger, pszGroup);
4548	NOREF(pszDest);
4549	}
4550	else
4551	rc = VERR_NOT_FOUND;
4552	break;
4553
4554	case SUPLOGGERSETTINGS_WHAT_CREATE:
4555	{
4556	if (pLogger)
4557	rc = VERR_ALREADY_EXISTS;
4558	else
4559	{
4560	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
4561
4562	rc = RTLogCreate(&pLogger,
4563	0 /* fFlags */,
4564	pszGroup,
4565	pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_DEBUG
4566	? "VBOX_LOG"
4567	: "VBOX_RELEASE_LOG",
4568	RT_ELEMENTS(s_apszGroups),
4569	s_apszGroups,
4570	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER,
4571	NULL);
4572	if (RT_SUCCESS(rc))
4573	{
4574	rc = RTLogFlags(pLogger, pszFlags);
4575	NOREF(pszDest);
4576	if (RT_SUCCESS(rc))
4577	{
4578	switch (pReq->u.In.fWhich)
4579	{
4580	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4581	pLogger = RTLogSetDefaultInstance(pLogger);
4582	break;
4583	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4584	pLogger = RTLogRelSetDefaultInstance(pLogger);
4585	break;
4586	}
4587	}
4588	RTLogDestroy(pLogger);
4589	}
4590	}
4591	break;
4592	}
4593
4594	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4595	switch (pReq->u.In.fWhich)
4596	{
4597	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4598	pLogger = RTLogSetDefaultInstance(NULL);
4599	break;
4600	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4601	pLogger = RTLogRelSetDefaultInstance(NULL);
4602	break;
4603	}
4604	rc = RTLogDestroy(pLogger);
4605	break;
4606
4607	default:
4608	{
4609	rc = VERR_INTERNAL_ERROR;
4610	break;
4611	}
4612	}
4613
4614	return rc;
4615	}
4616
4617
4618	/**
4619	* Creates the GIP.
4620	*
4621	* @returns VBox status code.
4622	* @param pDevExt Instance data. GIP stuff may be updated.
4623	*/
4624	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
4625	{
4626	PSUPGLOBALINFOPAGE pGip;
4627	RTHCPHYS HCPhysGip;
4628	uint32_t u32SystemResolution;
4629	uint32_t u32Interval;
4630	int rc;
4631
4632	LogFlow(("supdrvGipCreate:\n"));
4633
4634	/* assert order */
4635	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
4636	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
4637	Assert(!pDevExt->pGipTimer);
4638
4639	/*
4640	* Allocate a suitable page with a default kernel mapping.
4641	*/
4642	rc = RTR0MemObjAllocLow(&pDevExt->GipMemObj, PAGE_SIZE, false);
4643	if (RT_FAILURE(rc))
4644	{
4645	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
4646	return rc;
4647	}
4648	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
4649	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
4650
4651	#if 0 /** @todo Disabled this as we didn't used to do it before and causes unnecessary stress on laptops.
4652	* It only applies to Windows and should probably revisited later, if possible made part of the
4653	* timer code (return min granularity in RTTimerGetSystemGranularity and set it in RTTimerStart). */
4654	/*
4655	* Try bump up the system timer resolution.
4656	* The more interrupts the better...
4657	*/
4658	if ( RT_SUCCESS(RTTimerRequestSystemGranularity( 488281 /* 2048 HZ */, &u32SystemResolution))
4659	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 500000 /* 2000 HZ */, &u32SystemResolution))
4660	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
4661	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
4662	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 1953125 /* 512 HZ */, &u32SystemResolution))
4663	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 2000000 /* 500 HZ */, &u32SystemResolution))
4664	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 3906250 /* 256 HZ */, &u32SystemResolution))
4665	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 4000000 /* 250 HZ */, &u32SystemResolution))
4666	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 7812500 /* 128 HZ */, &u32SystemResolution))
4667	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity(10000000 /* 100 HZ */, &u32SystemResolution))
4668	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity(15625000 /* 64 HZ */, &u32SystemResolution))
4669	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity(31250000 /* 32 HZ */, &u32SystemResolution))
4670	)
4671	{
4672	Assert(RTTimerGetSystemGranularity() <= u32SystemResolution);
4673	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
4674	}
4675	#endif
4676
4677	/*
4678	* Find a reasonable update interval and initialize the structure.
4679	*/
4680	u32Interval = u32SystemResolution = RTTimerGetSystemGranularity();
4681	while (u32Interval < 10000000 /* 10 ms */)
4682	u32Interval += u32SystemResolution;
4683
4684	supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), 1000000000 / u32Interval /=Hz/);
4685
4686	/*
4687	* Create the timer.
4688	* If CPU_ALL isn't supported we'll have to fall back to synchronous mode.
4689	*/
4690	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
4691	{
4692	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, RTTIMER_FLAGS_CPU_ALL, supdrvGipAsyncTimer, pDevExt);
4693	if (rc == VERR_NOT_SUPPORTED)
4694	{
4695	OSDBGPRINT(("supdrvGipCreate: omni timer not supported, falling back to synchronous mode\n"));
4696	pGip->u32Mode = SUPGIPMODE_SYNC_TSC;
4697	}
4698	}
4699	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
4700	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0, supdrvGipSyncTimer, pDevExt);
4701	if (RT_SUCCESS(rc))
4702	{
4703	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
4704	rc = RTMpNotificationRegister(supdrvGipMpEvent, pDevExt);
4705	if (RT_SUCCESS(rc))
4706	{
4707	/*
4708	* We're good.
4709	*/
4710	Log(("supdrvGipCreate: %ld ns interval.\n", (long)u32Interval));
4711	g_pSUPGlobalInfoPage = pGip;
4712	return VINF_SUCCESS;
4713	}
4714
4715	OSDBGPRINT(("supdrvGipCreate: failed register MP event notfication. rc=%d\n", rc));
4716	}
4717	else
4718	{
4719	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %ld ns interval. rc=%d\n", (long)u32Interval, rc));
4720	Assert(!pDevExt->pGipTimer);
4721	}
4722	supdrvGipDestroy(pDevExt);
4723	return rc;
4724	}
4725
4726
4727	/**
4728	* Terminates the GIP.
4729	*
4730	* @param pDevExt Instance data. GIP stuff may be updated.
4731	*/
4732	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
4733	{
4734	int rc;
4735	#ifdef DEBUG_DARWIN_GIP
4736	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
4737	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
4738	pDevExt->pGipTimer, pDevExt->GipMemObj));
4739	#endif
4740
4741	/*
4742	* Invalid the GIP data.
4743	*/
4744	if (pDevExt->pGip)
4745	{
4746	supdrvGipTerm(pDevExt->pGip);
4747	pDevExt->pGip = NULL;
4748	}
4749	g_pSUPGlobalInfoPage = NULL;
4750
4751	/*
4752	* Destroy the timer and free the GIP memory object.
4753	*/
4754	if (pDevExt->pGipTimer)
4755	{
4756	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
4757	pDevExt->pGipTimer = NULL;
4758	}
4759
4760	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
4761	{
4762	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
4763	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
4764	}
4765
4766	/*
4767	* Finally, release the system timer resolution request if one succeeded.
4768	*/
4769	if (pDevExt->u32SystemTimerGranularityGrant)
4770	{
4771	rc = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
4772	pDevExt->u32SystemTimerGranularityGrant = 0;
4773	}
4774	}
4775
4776
4777	/**
4778	* Timer callback function sync GIP mode.
4779	* @param pTimer The timer.
4780	* @param pvUser The device extension.
4781	*/
4782	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
4783	{
4784	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
4785	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4786	uint64_t u64TSC = ASMReadTSC();
4787	uint64_t NanoTS = RTTimeSystemNanoTS();
4788
4789	supdrvGipUpdate(pDevExt->pGip, NanoTS, u64TSC, iTick);
4790
4791	ASMSetFlags(fOldFlags);
4792	}
4793
4794
4795	/**
4796	* Timer callback function for async GIP mode.
4797	* @param pTimer The timer.
4798	* @param pvUser The device extension.
4799	*/
4800	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
4801	{
4802	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
4803	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4804	RTCPUID idCpu = RTMpCpuId();
4805	uint64_t u64TSC = ASMReadTSC();
4806	uint64_t NanoTS = RTTimeSystemNanoTS();
4807
4808	/** @todo reset the transaction number and whatnot when iTick == 1. */
4809	if (pDevExt->idGipMaster == idCpu)
4810	supdrvGipUpdate(pDevExt->pGip, NanoTS, u64TSC, iTick);
4811	else
4812	supdrvGipUpdatePerCpu(pDevExt->pGip, NanoTS, u64TSC, ASMGetApicId(), iTick);
4813
4814	ASMSetFlags(fOldFlags);
4815	}
4816
4817
4818	/**
4819	* Multiprocessor event notification callback.
4820	*
4821	* This is used to make sue that the GIP master gets passed on to
4822	* another CPU.
4823	*
4824	* @param enmEvent The event.
4825	* @param idCpu The cpu it applies to.
4826	* @param pvUser Pointer to the device extension.
4827	*/
4828	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser)
4829	{
4830	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4831	if (enmEvent == RTMPEVENT_OFFLINE)
4832	{
4833	RTCPUID idGipMaster;
4834	ASMAtomicReadSize(&pDevExt->idGipMaster, &idGipMaster);
4835	if (idGipMaster == idCpu)
4836	{
4837	/*
4838	* Find a new GIP master.
4839	*/
4840	bool fIgnored;
4841	unsigned i;
4842	RTCPUID idNewGipMaster = NIL_RTCPUID;
4843	RTCPUSET OnlineCpus;
4844	RTMpGetOnlineSet(&OnlineCpus);
4845
4846	for (i = 0; i < RTCPUSET_MAX_CPUS; i++)
4847	{
4848	RTCPUID idCurCpu = RTMpCpuIdFromSetIndex(i);
4849	if ( RTCpuSetIsMember(&OnlineCpus, idCurCpu)
4850	&& idCurCpu != idGipMaster)
4851	{
4852	idNewGipMaster = idCurCpu;
4853	break;
4854	}
4855	}
4856
4857	Log(("supdrvGipMpEvent: Gip master %#lx -> %#lx\n", (long)idGipMaster, (long)idNewGipMaster));
4858	ASMAtomicCmpXchgSize(&pDevExt->idGipMaster, idNewGipMaster, idGipMaster, fIgnored);
4859	NOREF(fIgnored);
4860	}
4861	}
4862	}
4863
4864
4865	/**
4866	* Callback used by supdrvDetermineAsyncTSC to read the TSC on a CPU.
4867	*
4868	* @param idCpu Ignored.
4869	* @param pvUser1 Where to put the TSC.
4870	* @param pvUser2 Ignored.
4871	*/
4872	static DECLCALLBACK(void) supdrvDetermineAsyncTscWorker(RTCPUID idCpu, void pvUser1, void pvUser2)
4873	{
4874	#if 1
4875	ASMAtomicWriteU64((uint64_t volatile *)pvUser1, ASMReadTSC());
4876	#else
4877	(uint64_t )pvUser1 = ASMReadTSC();
4878	#endif
4879	}
4880
4881
4882	/**
4883	* Determine if Async GIP mode is required because of TSC drift.
4884	*
4885	* When using the default/normal timer code it is essential that the time stamp counter
4886	* (TSC) runs never backwards, that is, a read operation to the counter should return
4887	* a bigger value than any previous read operation. This is guaranteed by the latest
4888	* AMD CPUs and by newer Intel CPUs which never enter the C2 state (P4). In any other
4889	* case we have to choose the asynchronous timer mode.
4890	*
4891	* @param poffMin Pointer to the determined difference between different cores.
4892	* @return false if the time stamp counters appear to be synchron, true otherwise.
4893	*/
4894	static bool supdrvDetermineAsyncTsc(uint64_t *poffMin)
4895	{
4896	/*
4897	* Just iterate all the cpus 8 times and make sure that the TSC is
4898	* ever increasing. We don't bother taking TSC rollover into account.
4899	*/
4900	RTCPUSET CpuSet;
4901	int iLastCpu = RTCpuLastIndex(RTMpGetSet(&CpuSet));
4902	int iCpu;
4903	int cLoops = 8;
4904	bool fAsync = false;
4905	int rc = VINF_SUCCESS;
4906	uint64_t offMax = 0;
4907	uint64_t offMin = ~(uint64_t)0;
4908	uint64_t PrevTsc = ASMReadTSC();
4909
4910	while (cLoops-- > 0)
4911	{
4912	for (iCpu = 0; iCpu <= iLastCpu; iCpu++)
4913	{
4914	uint64_t CurTsc;
4915	rc = RTMpOnSpecific(RTMpCpuIdFromSetIndex(iCpu), supdrvDetermineAsyncTscWorker, &CurTsc, NULL);
4916	if (RT_SUCCESS(rc))
4917	{
4918	if (CurTsc <= PrevTsc)
4919	{
4920	fAsync = true;
4921	offMin = offMax = PrevTsc - CurTsc;
4922	Log(("supdrvDetermineAsyncTsc: iCpu=%d cLoops=%d CurTsc=%llx PrevTsc=%llx\n",
4923	iCpu, cLoops, CurTsc, PrevTsc));
4924	break;
4925	}
4926
4927	/* Gather statistics (except the first time). */
4928	if (iCpu != 0 \|\| cLoops != 7)
4929	{
4930	uint64_t off = CurTsc - PrevTsc;
4931	if (off < offMin)
4932	offMin = off;
4933	if (off > offMax)
4934	offMax = off;
4935	Log2(("%d/%d: off=%llx\n", cLoops, iCpu, off));
4936	}
4937
4938	/* Next */
4939	PrevTsc = CurTsc;
4940	}
4941	else if (rc == VERR_NOT_SUPPORTED)
4942	break;
4943	else
4944	AssertMsg(rc == VERR_CPU_NOT_FOUND \|\| rc == VERR_CPU_OFFLINE, ("%d\n", rc));
4945	}
4946
4947	/* broke out of the loop. */
4948	if (iCpu <= iLastCpu)
4949	break;
4950	}
4951
4952	poffMin = offMin; / Almost RTMpOnSpecific profiling. */
4953	Log(("supdrvDetermineAsyncTsc: returns %d; iLastCpu=%d rc=%d offMin=%llx offMax=%llx\n",
4954	fAsync, iLastCpu, rc, offMin, offMax));
4955	#if !defined(RT_OS_SOLARIS) && !defined(RT_OS_OS2) && !defined(RT_OS_WINDOWS)
4956	OSDBGPRINT(("vboxdrv: fAsync=%d offMin=%#lx offMax=%#lx\n", fAsync, (long)offMin, (long)offMax));
4957	#endif
4958	return fAsync;
4959	}
4960
4961
4962	/**
4963	* Determin the GIP TSC mode.
4964	*
4965	* @returns The most suitable TSC mode.
4966	* @param pDevExt Pointer to the device instance data.
4967	*/
4968	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt)
4969	{
4970	/*
4971	* On SMP we're faced with two problems:
4972	* (1) There might be a skew between the CPU, so that cpu0
4973	* returns a TSC that is sligtly different from cpu1.
4974	* (2) Power management (and other things) may cause the TSC
4975	* to run at a non-constant speed, and cause the speed
4976	* to be different on the cpus. This will result in (1).
4977	*
4978	* So, on SMP systems we'll have to select the ASYNC update method
4979	* if there are symphoms of these problems.
4980	*/
4981	if (RTMpGetCount() > 1)
4982	{
4983	uint32_t uEAX, uEBX, uECX, uEDX;
4984	uint64_t u64DiffCoresIgnored;
4985
4986	/* Permit the user and/or the OS specfic bits to force async mode. */
4987	if (supdrvOSGetForcedAsyncTscMode(pDevExt))
4988	return SUPGIPMODE_ASYNC_TSC;
4989
4990	/* Try check for current differences between the cpus. */
4991	if (supdrvDetermineAsyncTsc(&u64DiffCoresIgnored))
4992	return SUPGIPMODE_ASYNC_TSC;
4993
4994	/*
4995	* If the CPU supports power management and is an AMD one we
4996	* won't trust it unless it has the TscInvariant bit is set.
4997	*/
4998	/* Check for "AuthenticAMD" */
4999	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
5000	if ( uEAX >= 1
5001	&& uEBX == X86_CPUID_VENDOR_AMD_EBX
5002	&& uECX == X86_CPUID_VENDOR_AMD_ECX
5003	&& uEDX == X86_CPUID_VENDOR_AMD_EDX)
5004	{
5005	/* Check for APM support and that TscInvariant is cleared. */
5006	ASMCpuId(0x80000000, &uEAX, &uEBX, &uECX, &uEDX);
5007	if (uEAX >= 0x80000007)
5008	{
5009	ASMCpuId(0x80000007, &uEAX, &uEBX, &uECX, &uEDX);
5010	if ( !(uEDX & RT_BIT(8))/* TscInvariant */
5011	&& (uEDX & 0x3e)) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
5012	return SUPGIPMODE_ASYNC_TSC;
5013	}
5014	}
5015	}
5016	return SUPGIPMODE_SYNC_TSC;
5017	}
5018
5019
5020
5021	/**
5022	* Initializes the GIP data.
5023	*
5024	* @param pDevExt Pointer to the device instance data.
5025	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5026	* @param HCPhys The physical address of the GIP.
5027	* @param u64NanoTS The current nanosecond timestamp.
5028	* @param uUpdateHz The update freqence.
5029	*/
5030	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys, uint64_t u64NanoTS, unsigned uUpdateHz)
5031	{
5032	unsigned i;
5033	#ifdef DEBUG_DARWIN_GIP
5034	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz));
5035	#else
5036	LogFlow(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz));
5037	#endif
5038
5039	/*
5040	* Initialize the structure.
5041	*/
5042	memset(pGip, 0, PAGE_SIZE);
5043	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
5044	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
5045	pGip->u32Mode = supdrvGipDeterminTscMode(pDevExt);
5046	pGip->u32UpdateHz = uUpdateHz;
5047	pGip->u32UpdateIntervalNS = 1000000000 / uUpdateHz;
5048	pGip->u64NanoTSLastUpdateHz = u64NanoTS;
5049
5050	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
5051	{
5052	pGip->aCPUs[i].u32TransactionId = 2;
5053	pGip->aCPUs[i].u64NanoTS = u64NanoTS;
5054	pGip->aCPUs[i].u64TSC = ASMReadTSC();
5055
5056	/*
5057	* We don't know the following values until we've executed updates.
5058	* So, we'll just pretend it's a 4 GHz CPU and adjust the history it on
5059	* the 2nd timer callout.
5060	*/
5061	pGip->aCPUs[i].u64CpuHz = _4G + 1; /* tstGIP-2 depends on this. */
5062	pGip->aCPUs[i].u32UpdateIntervalTSC
5063	= pGip->aCPUs[i].au32TSCHistory[0]
5064	= pGip->aCPUs[i].au32TSCHistory[1]
5065	= pGip->aCPUs[i].au32TSCHistory[2]
5066	= pGip->aCPUs[i].au32TSCHistory[3]
5067	= pGip->aCPUs[i].au32TSCHistory[4]
5068	= pGip->aCPUs[i].au32TSCHistory[5]
5069	= pGip->aCPUs[i].au32TSCHistory[6]
5070	= pGip->aCPUs[i].au32TSCHistory[7]
5071	= /pGip->aCPUs[i].u64CpuHz/ _4G / uUpdateHz;
5072	}
5073
5074	/*
5075	* Link it to the device extension.
5076	*/
5077	pDevExt->pGip = pGip;
5078	pDevExt->HCPhysGip = HCPhys;
5079	pDevExt->cGipUsers = 0;
5080	}
5081
5082
5083	/**
5084	* Invalidates the GIP data upon termination.
5085	*
5086	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5087	*/
5088	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
5089	{
5090	unsigned i;
5091	pGip->u32Magic = 0;
5092	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
5093	{
5094	pGip->aCPUs[i].u64NanoTS = 0;
5095	pGip->aCPUs[i].u64TSC = 0;
5096	pGip->aCPUs[i].iTSCHistoryHead = 0;
5097	}
5098	}
5099
5100
5101	/**
5102	* Worker routine for supdrvGipUpdate and supdrvGipUpdatePerCpu that
5103	* updates all the per cpu data except the transaction id.
5104	*
5105	* @param pGip The GIP.
5106	* @param pGipCpu Pointer to the per cpu data.
5107	* @param u64NanoTS The current time stamp.
5108	* @param u64TSC The current TSC.
5109	* @param iTick The current timer tick.
5110	*/
5111	static void supdrvGipDoUpdateCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick)
5112	{
5113	uint64_t u64TSCDelta;
5114	uint32_t u32UpdateIntervalTSC;
5115	uint32_t u32UpdateIntervalTSCSlack;
5116	unsigned iTSCHistoryHead;
5117	uint64_t u64CpuHz;
5118	uint32_t u32TransactionId;
5119
5120	/* Delta between this and the previous update. */
5121	ASMAtomicUoWriteU32(&pGipCpu->u32PrevUpdateIntervalNS, (uint32_t)(u64NanoTS - pGipCpu->u64NanoTS));
5122
5123	/*
5124	* Update the NanoTS.
5125	*/
5126	ASMAtomicXchgU64(&pGipCpu->u64NanoTS, u64NanoTS);
5127
5128	/*
5129	* Calc TSC delta.
5130	*/
5131	/** @todo validate the NanoTS delta, don't trust the OS to call us when it should... */
5132	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
5133	ASMAtomicXchgU64(&pGipCpu->u64TSC, u64TSC);
5134
5135	if (u64TSCDelta >> 32)
5136	{
5137	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
5138	pGipCpu->cErrors++;
5139	}
5140
5141	/*
5142	* On the 2nd and 3rd callout, reset the history with the current TSC
5143	* interval since the values entered by supdrvGipInit are totally off.
5144	* The interval on the 1st callout completely unreliable, the 2nd is a bit
5145	* better, while the 3rd should be most reliable.
5146	*/
5147	u32TransactionId = pGipCpu->u32TransactionId;
5148	if (RT_UNLIKELY( ( u32TransactionId == 5
5149	\|\| u32TransactionId == 7)
5150	&& ( iTick == 2
5151	\|\| iTick == 3) ))
5152	{
5153	unsigned i;
5154	for (i = 0; i < RT_ELEMENTS(pGipCpu->au32TSCHistory); i++)
5155	ASMAtomicUoWriteU32(&pGipCpu->au32TSCHistory[i], (uint32_t)u64TSCDelta);
5156	}
5157
5158	/*
5159	* TSC History.
5160	*/
5161	Assert(RT_ELEMENTS(pGipCpu->au32TSCHistory) == 8);
5162	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
5163	ASMAtomicXchgU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
5164	ASMAtomicXchgU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
5165
5166	/*
5167	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
5168	*/
5169	if (pGip->u32UpdateHz >= 1000)
5170	{
5171	uint32_t u32;
5172	u32 = pGipCpu->au32TSCHistory[0];
5173	u32 += pGipCpu->au32TSCHistory[1];
5174	u32 += pGipCpu->au32TSCHistory[2];
5175	u32 += pGipCpu->au32TSCHistory[3];
5176	u32 >>= 2;
5177	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
5178	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
5179	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
5180	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
5181	u32UpdateIntervalTSC >>= 2;
5182	u32UpdateIntervalTSC += u32;
5183	u32UpdateIntervalTSC >>= 1;
5184
5185	/* Value choosen for a 2GHz Athlon64 running linux 2.6.10/11, . */
5186	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
5187	}
5188	else if (pGip->u32UpdateHz >= 90)
5189	{
5190	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5191	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
5192	u32UpdateIntervalTSC >>= 1;
5193
5194	/* value choosen on a 2GHz thinkpad running windows */
5195	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
5196	}
5197	else
5198	{
5199	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5200
5201	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
5202	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
5203	}
5204	ASMAtomicXchgU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
5205
5206	/*
5207	* CpuHz.
5208	*/
5209	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, pGip->u32UpdateHz);
5210	ASMAtomicXchgU64(&pGipCpu->u64CpuHz, u64CpuHz);
5211	}
5212
5213
5214	/**
5215	* Updates the GIP.
5216	*
5217	* @param pGip Pointer to the GIP.
5218	* @param u64NanoTS The current nanosecond timesamp.
5219	* @param u64TSC The current TSC timesamp.
5220	* @param iTick The current timer tick.
5221	*/
5222	static void supdrvGipUpdate(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick)
5223	{
5224	/*
5225	* Determin the relevant CPU data.
5226	*/
5227	PSUPGIPCPU pGipCpu;
5228	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
5229	pGipCpu = &pGip->aCPUs[0];
5230	else
5231	{
5232	unsigned iCpu = ASMGetApicId();
5233	if (RT_UNLIKELY(iCpu >= RT_ELEMENTS(pGip->aCPUs)))
5234	return;
5235	pGipCpu = &pGip->aCPUs[iCpu];
5236	}
5237
5238	/*
5239	* Start update transaction.
5240	*/
5241	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5242	{
5243	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
5244	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5245	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5246	pGipCpu->cErrors++;
5247	return;
5248	}
5249
5250	/*
5251	* Recalc the update frequency every 0x800th time.
5252	*/
5253	if (!(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
5254	{
5255	if (pGip->u64NanoTSLastUpdateHz)
5256	{
5257	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
5258	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
5259	uint32_t u32UpdateHz = (uint32_t)((UINT64_C(1000000000) * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
5260	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
5261	{
5262	ASMAtomicXchgU32(&pGip->u32UpdateHz, u32UpdateHz);
5263	ASMAtomicXchgU32(&pGip->u32UpdateIntervalNS, 1000000000 / u32UpdateHz);
5264	}
5265	#endif
5266	}
5267	ASMAtomicXchgU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS);
5268	}
5269
5270	/*
5271	* Update the data.
5272	*/
5273	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS, u64TSC, iTick);
5274
5275	/*
5276	* Complete transaction.
5277	*/
5278	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5279	}
5280
5281
5282	/**
5283	* Updates the per cpu GIP data for the calling cpu.
5284	*
5285	* @param pGip Pointer to the GIP.
5286	* @param u64NanoTS The current nanosecond timesamp.
5287	* @param u64TSC The current TSC timesamp.
5288	* @param iCpu The CPU index.
5289	* @param iTick The current timer tick.
5290	*/
5291	static void supdrvGipUpdatePerCpu(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC, unsigned iCpu, uint64_t iTick)
5292	{
5293	PSUPGIPCPU pGipCpu;
5294
5295	if (RT_LIKELY(iCpu < RT_ELEMENTS(pGip->aCPUs)))
5296	{
5297	pGipCpu = &pGip->aCPUs[iCpu];
5298
5299	/*
5300	* Start update transaction.
5301	*/
5302	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5303	{
5304	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5305	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5306	pGipCpu->cErrors++;
5307	return;
5308	}
5309
5310	/*
5311	* Update the data.
5312	*/
5313	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS, u64TSC, iTick);
5314
5315	/*
5316	* Complete transaction.
5317	*/
5318	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5319	}
5320	}
5321

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

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDrv.c@ 29709

Download in other formats: