SUPDrv.c@ 36254

Last change on this file since 36254 was 36254, checked in by vboxsync, 14 years ago
GIP,++: Lots of CPUs (disabled).
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 214.2 KB

Line
1	/* $Revision: 36254 $ */
2	/** @file
3	* VBoxDrv - The VirtualBox Support Driver - Common code.
4	*/
5
6	/*
7	* Copyright (C) 2006-2010 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	#include <iprt/net.h>
50	#include <iprt/crc.h>
51	#include <iprt/string.h>
52	#include <iprt/timer.h>
53	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
54	# include <iprt/rand.h>
55	# include <iprt/path.h>
56	#endif
57
58	#include <VBox/param.h>
59	#include <VBox/log.h>
60	#include <VBox/err.h>
61	#include <VBox/vmm/hwacc_svm.h>
62	#include <VBox/vmm/hwacc_vmx.h>
63	#include <VBox/x86.h>
64
65	#ifdef VBOX_WITH_DTRACE
66	# include "SUPDrv-dtrace.h"
67	#else
68	# define VBOXDRV_SUPDRV_SESSION_CREATE(pvSession, fUser) do { } while (0)
69	# define VBOXDRV_SUPDRV_SESSION_CLOSE(pvSession) do { } while (0)
70	# define VBOXDRV_SUPDRV_IOCCLOSE(pvSession) do { } while (0)
71	# define VBOXDRV_SUPDRV_IOCTL_ENTRY(pvSession, uIOCtl, pvReqHdr) do { } while (0)
72	# define VBOXDRV_SUPDRV_IOCTL_RETURN(pvSession, uIOCtl, pvReqHdr, rcRet, rcReq) do { } while (0)
73	#endif
74
75	/*
76	* Logging assignments:
77	* Log - useful stuff, like failures.
78	* LogFlow - program flow, except the really noisy bits.
79	* Log2 - Cleanup.
80	* Log3 - Loader flow noise.
81	* Log4 - Call VMMR0 flow noise.
82	* Log5 - Native yet-to-be-defined noise.
83	* Log6 - Native ioctl flow noise.
84	*
85	* Logging requires BUILD_TYPE=debug and possibly changes to the logger
86	* instantiation in log-vbox.c(pp).
87	*/
88
89
90	/*******************************************************************************
91	* Defined Constants And Macros *
92	*******************************************************************************/
93	/** The frequency by which we recalculate the u32UpdateHz and
94	* u32UpdateIntervalNS GIP members. The value must be a power of 2. */
95	#define GIP_UPDATEHZ_RECALC_FREQ 0x800
96
97	/** @def VBOX_SVN_REV
98	* The makefile should define this if it can. */
99	#ifndef VBOX_SVN_REV
100	# define VBOX_SVN_REV 0
101	#endif
102
103	#if 0 /* Don't start the GIP timers. Useful when debugging the IPRT timer code. */
104	# define DO_NOT_START_GIP
105	#endif
106
107
108	/*******************************************************************************
109	* Internal Functions *
110	*******************************************************************************/
111	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser);
112	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser);
113	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession);
114	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType);
115	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq);
116	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq);
117	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq);
118	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq);
119	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq);
120	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx);
121	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt);
122	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage);
123	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage);
124	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt);
125	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt);
126	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq);
127	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq);
128	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt);
129	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt);
130	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
131	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
132	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser);
133	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
134	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned cCpus);
135	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2);
136	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip);
137	static void supdrvGipUpdate(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick);
138	static void supdrvGipUpdatePerCpu(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC,
139	RTCPUID idCpu, uint8_t idApic, uint64_t iTick);
140
141
142	/*******************************************************************************
143	* Global Variables *
144	*******************************************************************************/
145	DECLEXPORT(PSUPGLOBALINFOPAGE) g_pSUPGlobalInfoPage = NULL;
146
147	/**
148	* Array of the R0 SUP API.
149	*/
150	static SUPFUNC g_aFunctions[] =
151	{
152	/* name function */
153	/* Entries with absolute addresses determined at runtime, fixup
154	code makes ugly ASSUMPTIONS about the order here: */
155	{ "SUPR0AbsIs64bit", (void *)0 },
156	{ "SUPR0Abs64bitKernelCS", (void *)0 },
157	{ "SUPR0Abs64bitKernelSS", (void *)0 },
158	{ "SUPR0Abs64bitKernelDS", (void *)0 },
159	{ "SUPR0AbsKernelCS", (void *)0 },
160	{ "SUPR0AbsKernelSS", (void *)0 },
161	{ "SUPR0AbsKernelDS", (void *)0 },
162	{ "SUPR0AbsKernelES", (void *)0 },
163	{ "SUPR0AbsKernelFS", (void *)0 },
164	{ "SUPR0AbsKernelGS", (void *)0 },
165	/* Normal function pointers: */
166	{ "SUPR0ComponentRegisterFactory", (void *)SUPR0ComponentRegisterFactory },
167	{ "SUPR0ComponentDeregisterFactory", (void *)SUPR0ComponentDeregisterFactory },
168	{ "SUPR0ComponentQueryFactory", (void *)SUPR0ComponentQueryFactory },
169	{ "SUPR0ObjRegister", (void *)SUPR0ObjRegister },
170	{ "SUPR0ObjAddRef", (void *)SUPR0ObjAddRef },
171	{ "SUPR0ObjAddRefEx", (void *)SUPR0ObjAddRefEx },
172	{ "SUPR0ObjRelease", (void *)SUPR0ObjRelease },
173	{ "SUPR0ObjVerifyAccess", (void *)SUPR0ObjVerifyAccess },
174	{ "SUPR0LockMem", (void *)SUPR0LockMem },
175	{ "SUPR0UnlockMem", (void *)SUPR0UnlockMem },
176	{ "SUPR0ContAlloc", (void *)SUPR0ContAlloc },
177	{ "SUPR0ContFree", (void *)SUPR0ContFree },
178	{ "SUPR0LowAlloc", (void *)SUPR0LowAlloc },
179	{ "SUPR0LowFree", (void *)SUPR0LowFree },
180	{ "SUPR0MemAlloc", (void *)SUPR0MemAlloc },
181	{ "SUPR0MemGetPhys", (void *)SUPR0MemGetPhys },
182	{ "SUPR0MemFree", (void *)SUPR0MemFree },
183	{ "SUPR0PageAllocEx", (void *)SUPR0PageAllocEx },
184	{ "SUPR0PageFree", (void *)SUPR0PageFree },
185	{ "SUPR0Printf", (void )SUPR0Printf }, /* @todo needs wrapping? */
186	{ "SUPSemEventCreate", (void *)SUPSemEventCreate },
187	{ "SUPSemEventClose", (void *)SUPSemEventClose },
188	{ "SUPSemEventSignal", (void *)SUPSemEventSignal },
189	{ "SUPSemEventWait", (void *)SUPSemEventWait },
190	{ "SUPSemEventWaitNoResume", (void *)SUPSemEventWaitNoResume },
191	{ "SUPSemEventWaitNsAbsIntr", (void *)SUPSemEventWaitNsAbsIntr },
192	{ "SUPSemEventWaitNsRelIntr", (void *)SUPSemEventWaitNsRelIntr },
193	{ "SUPSemEventGetResolution", (void *)SUPSemEventGetResolution },
194	{ "SUPSemEventMultiCreate", (void *)SUPSemEventMultiCreate },
195	{ "SUPSemEventMultiClose", (void *)SUPSemEventMultiClose },
196	{ "SUPSemEventMultiSignal", (void *)SUPSemEventMultiSignal },
197	{ "SUPSemEventMultiReset", (void *)SUPSemEventMultiReset },
198	{ "SUPSemEventMultiWait", (void *)SUPSemEventMultiWait },
199	{ "SUPSemEventMultiWaitNoResume", (void *)SUPSemEventMultiWaitNoResume },
200	{ "SUPSemEventMultiWaitNsAbsIntr", (void *)SUPSemEventMultiWaitNsAbsIntr },
201	{ "SUPSemEventMultiWaitNsRelIntr", (void *)SUPSemEventMultiWaitNsRelIntr },
202	{ "SUPSemEventMultiGetResolution", (void *)SUPSemEventMultiGetResolution },
203	{ "SUPR0GetPagingMode", (void *)SUPR0GetPagingMode },
204	{ "SUPR0EnableVTx", (void *)SUPR0EnableVTx },
205	{ "SUPGetGIP", (void *)SUPGetGIP },
206	{ "g_pSUPGlobalInfoPage", (void *)&g_pSUPGlobalInfoPage },
207	{ "RTMemAllocTag", (void *)RTMemAllocTag },
208	{ "RTMemAllocZTag", (void *)RTMemAllocZTag },
209	{ "RTMemAllocVarTag", (void *)RTMemAllocVarTag },
210	{ "RTMemAllocZVarTag", (void *)RTMemAllocZVarTag },
211	{ "RTMemFree", (void *)RTMemFree },
212	{ "RTMemDupTag", (void *)RTMemDupTag },
213	{ "RTMemDupExTag", (void *)RTMemDupExTag },
214	{ "RTMemReallocTag", (void *)RTMemReallocTag },
215	{ "RTR0MemObjAllocLowTag", (void *)RTR0MemObjAllocLowTag },
216	{ "RTR0MemObjAllocPageTag", (void *)RTR0MemObjAllocPageTag },
217	{ "RTR0MemObjAllocPhysTag", (void *)RTR0MemObjAllocPhysTag },
218	{ "RTR0MemObjAllocPhysExTag", (void *)RTR0MemObjAllocPhysExTag },
219	{ "RTR0MemObjAllocPhysNCTag", (void *)RTR0MemObjAllocPhysNCTag },
220	{ "RTR0MemObjAllocContTag", (void *)RTR0MemObjAllocContTag },
221	{ "RTR0MemObjEnterPhysTag", (void *)RTR0MemObjEnterPhysTag },
222	{ "RTR0MemObjLockUserTag", (void *)RTR0MemObjLockUserTag },
223	{ "RTR0MemObjMapKernelTag", (void *)RTR0MemObjMapKernelTag },
224	{ "RTR0MemObjMapKernelExTag", (void *)RTR0MemObjMapKernelExTag },
225	{ "RTR0MemObjMapUserTag", (void *)RTR0MemObjMapUserTag },
226	{ "RTR0MemObjProtect", (void *)RTR0MemObjProtect },
227	{ "RTR0MemObjAddress", (void *)RTR0MemObjAddress },
228	{ "RTR0MemObjAddressR3", (void *)RTR0MemObjAddressR3 },
229	{ "RTR0MemObjSize", (void *)RTR0MemObjSize },
230	{ "RTR0MemObjIsMapping", (void *)RTR0MemObjIsMapping },
231	{ "RTR0MemObjGetPagePhysAddr", (void *)RTR0MemObjGetPagePhysAddr },
232	{ "RTR0MemObjFree", (void *)RTR0MemObjFree },
233	{ "RTR0MemUserCopyFrom", (void *)RTR0MemUserCopyFrom },
234	{ "RTR0MemUserCopyTo", (void *)RTR0MemUserCopyTo },
235	{ "RTR0MemUserIsValidAddr", (void *)RTR0MemUserIsValidAddr },
236	{ "RTR0MemKernelIsValidAddr", (void *)RTR0MemKernelIsValidAddr },
237	{ "RTR0MemAreKrnlAndUsrDifferent", (void *)RTR0MemAreKrnlAndUsrDifferent },
238	{ "RTSemMutexCreate", (void *)RTSemMutexCreate },
239	{ "RTSemMutexRequest", (void *)RTSemMutexRequest },
240	{ "RTSemMutexRequestDebug", (void *)RTSemMutexRequestDebug },
241	{ "RTSemMutexRequestNoResume", (void *)RTSemMutexRequestNoResume },
242	{ "RTSemMutexRequestNoResumeDebug", (void *)RTSemMutexRequestNoResumeDebug },
243	{ "RTSemMutexRelease", (void *)RTSemMutexRelease },
244	{ "RTSemMutexDestroy", (void *)RTSemMutexDestroy },
245	{ "RTProcSelf", (void *)RTProcSelf },
246	{ "RTR0ProcHandleSelf", (void *)RTR0ProcHandleSelf },
247	{ "RTSemFastMutexCreate", (void *)RTSemFastMutexCreate },
248	{ "RTSemFastMutexDestroy", (void *)RTSemFastMutexDestroy },
249	{ "RTSemFastMutexRequest", (void *)RTSemFastMutexRequest },
250	{ "RTSemFastMutexRelease", (void *)RTSemFastMutexRelease },
251	{ "RTSemEventCreate", (void *)RTSemEventCreate },
252	{ "RTSemEventSignal", (void *)RTSemEventSignal },
253	{ "RTSemEventWait", (void *)RTSemEventWait },
254	{ "RTSemEventWaitNoResume", (void *)RTSemEventWaitNoResume },
255	{ "RTSemEventWaitEx", (void *)RTSemEventWaitEx },
256	{ "RTSemEventWaitExDebug", (void *)RTSemEventWaitExDebug },
257	{ "RTSemEventGetResolution", (void *)RTSemEventGetResolution },
258	{ "RTSemEventDestroy", (void *)RTSemEventDestroy },
259	{ "RTSemEventMultiCreate", (void *)RTSemEventMultiCreate },
260	{ "RTSemEventMultiSignal", (void *)RTSemEventMultiSignal },
261	{ "RTSemEventMultiReset", (void *)RTSemEventMultiReset },
262	{ "RTSemEventMultiWait", (void *)RTSemEventMultiWait },
263	{ "RTSemEventMultiWaitNoResume", (void *)RTSemEventMultiWaitNoResume },
264	{ "RTSemEventMultiWaitEx", (void *)RTSemEventMultiWaitEx },
265	{ "RTSemEventMultiWaitExDebug", (void *)RTSemEventMultiWaitExDebug },
266	{ "RTSemEventMultiGetResolution", (void *)RTSemEventMultiGetResolution },
267	{ "RTSemEventMultiDestroy", (void *)RTSemEventMultiDestroy },
268	{ "RTSpinlockCreate", (void *)RTSpinlockCreate },
269	{ "RTSpinlockDestroy", (void *)RTSpinlockDestroy },
270	{ "RTSpinlockAcquire", (void *)RTSpinlockAcquire },
271	{ "RTSpinlockRelease", (void *)RTSpinlockRelease },
272	{ "RTSpinlockAcquireNoInts", (void *)RTSpinlockAcquireNoInts },
273	{ "RTSpinlockReleaseNoInts", (void *)RTSpinlockReleaseNoInts },
274	{ "RTTimeNanoTS", (void *)RTTimeNanoTS },
275	{ "RTTimeMilliTS", (void *)RTTimeMilliTS },
276	{ "RTTimeSystemNanoTS", (void *)RTTimeSystemNanoTS },
277	{ "RTTimeSystemMilliTS", (void *)RTTimeSystemMilliTS },
278	{ "RTThreadNativeSelf", (void *)RTThreadNativeSelf },
279	{ "RTThreadSleep", (void *)RTThreadSleep },
280	{ "RTThreadYield", (void *)RTThreadYield },
281	#if 0 /* Thread APIs, Part 2. */
282	{ "RTThreadSelf", (void *)RTThreadSelf },
283	{ "RTThreadCreate", (void *)RTThreadCreate },
284	{ "RTThreadGetNative", (void *)RTThreadGetNative },
285	{ "RTThreadWait", (void *)RTThreadWait },
286	{ "RTThreadWaitNoResume", (void *)RTThreadWaitNoResume },
287	{ "RTThreadGetName", (void *)RTThreadGetName },
288	{ "RTThreadSelfName", (void *)RTThreadSelfName },
289	{ "RTThreadGetType", (void *)RTThreadGetType },
290	{ "RTThreadUserSignal", (void *)RTThreadUserSignal },
291	{ "RTThreadUserReset", (void *)RTThreadUserReset },
292	{ "RTThreadUserWait", (void *)RTThreadUserWait },
293	{ "RTThreadUserWaitNoResume", (void *)RTThreadUserWaitNoResume },
294	#else
295	/**
296	* @todo: remove me, once above code enabled.
297	* We need RTThreadCreate/RTThreadWait in the PCI driver.
298	*/
299	{ "RTThreadCreate", (void *)RTThreadCreate },
300	{ "RTThreadWait", (void *)RTThreadWait },
301	#endif
302	{ "RTThreadPreemptIsEnabled", (void *)RTThreadPreemptIsEnabled },
303	{ "RTThreadPreemptIsPending", (void *)RTThreadPreemptIsPending },
304	{ "RTThreadPreemptIsPendingTrusty", (void *)RTThreadPreemptIsPendingTrusty },
305	{ "RTThreadPreemptIsPossible", (void *)RTThreadPreemptIsPossible },
306	{ "RTThreadPreemptDisable", (void *)RTThreadPreemptDisable },
307	{ "RTThreadPreemptRestore", (void *)RTThreadPreemptRestore },
308	{ "RTThreadIsInInterrupt", (void *)RTThreadIsInInterrupt },
309	{ "RTTimerCreate", (void *)RTTimerCreate },
310	{ "RTTimerCreateEx", (void *)RTTimerCreateEx },
311	{ "RTTimerDestroy", (void *)RTTimerDestroy },
312	{ "RTTimerStart", (void *)RTTimerStart },
313	{ "RTTimerStop", (void *)RTTimerStop },
314	{ "RTTimerChangeInterval", (void *)RTTimerChangeInterval },
315	{ "RTTimerGetSystemGranularity", (void *)RTTimerGetSystemGranularity },
316	{ "RTTimerRequestSystemGranularity", (void *)RTTimerRequestSystemGranularity },
317	{ "RTTimerReleaseSystemGranularity", (void *)RTTimerReleaseSystemGranularity },
318	{ "RTTimerCanDoHighResolution", (void *)RTTimerCanDoHighResolution },
319
320	{ "RTLogDefaultInstance", (void *)RTLogDefaultInstance },
321	{ "RTMpCpuId", (void *)RTMpCpuId },
322	{ "RTMpCpuIdFromSetIndex", (void *)RTMpCpuIdFromSetIndex },
323	{ "RTMpCpuIdToSetIndex", (void *)RTMpCpuIdToSetIndex },
324	{ "RTMpGetArraySize", (void *)RTMpGetArraySize },
325	{ "RTMpIsCpuPossible", (void *)RTMpIsCpuPossible },
326	{ "RTMpGetCount", (void *)RTMpGetCount },
327	{ "RTMpGetMaxCpuId", (void *)RTMpGetMaxCpuId },
328	{ "RTMpGetOnlineCount", (void *)RTMpGetOnlineCount },
329	{ "RTMpGetOnlineSet", (void *)RTMpGetOnlineSet },
330	{ "RTMpGetSet", (void *)RTMpGetSet },
331	{ "RTMpIsCpuOnline", (void *)RTMpIsCpuOnline },
332	{ "RTMpIsCpuWorkPending", (void *)RTMpIsCpuWorkPending },
333	{ "RTMpOnAll", (void *)RTMpOnAll },
334	{ "RTMpOnOthers", (void *)RTMpOnOthers },
335	{ "RTMpOnSpecific", (void *)RTMpOnSpecific },
336	{ "RTMpPokeCpu", (void *)RTMpPokeCpu },
337	{ "RTPowerNotificationRegister", (void *)RTPowerNotificationRegister },
338	{ "RTPowerNotificationDeregister", (void *)RTPowerNotificationDeregister },
339	{ "RTLogRelDefaultInstance", (void *)RTLogRelDefaultInstance },
340	{ "RTLogSetDefaultInstanceThread", (void *)RTLogSetDefaultInstanceThread },
341	{ "RTLogLoggerExV", (void *)RTLogLoggerExV },
342	{ "RTLogPrintfV", (void *)RTLogPrintfV },
343	{ "RTR0AssertPanicSystem", (void *)RTR0AssertPanicSystem },
344	{ "RTAssertMsg1", (void *)RTAssertMsg1 },
345	{ "RTAssertMsg2V", (void *)RTAssertMsg2V },
346	{ "RTAssertSetQuiet", (void *)RTAssertSetQuiet },
347	{ "RTAssertMayPanic", (void *)RTAssertMayPanic },
348	{ "RTAssertSetMayPanic", (void *)RTAssertSetMayPanic },
349	{ "RTAssertAreQuiet", (void *)RTAssertAreQuiet },
350	{ "RTStrFormat", (void *)RTStrFormat },
351	{ "RTStrFormatNumber", (void *)RTStrFormatNumber },
352	{ "RTStrFormatTypeDeregister", (void *)RTStrFormatTypeDeregister },
353	{ "RTStrFormatTypeRegister", (void *)RTStrFormatTypeRegister },
354	{ "RTStrFormatTypeSetUser", (void *)RTStrFormatTypeSetUser },
355	{ "RTStrFormatV", (void *)RTStrFormatV },
356	{ "RTStrPrintf", (void *)RTStrPrintf },
357	{ "RTStrPrintfEx", (void *)RTStrPrintfEx },
358	{ "RTStrPrintfExV", (void *)RTStrPrintfExV },
359	{ "RTStrPrintfV", (void *)RTStrPrintfV },
360	{ "RTHandleTableAllocWithCtx", (void *)RTHandleTableAllocWithCtx },
361	{ "RTHandleTableCreate", (void *)RTHandleTableCreate },
362	{ "RTHandleTableCreateEx", (void *)RTHandleTableCreateEx },
363	{ "RTHandleTableDestroy", (void *)RTHandleTableDestroy },
364	{ "RTHandleTableFreeWithCtx", (void *)RTHandleTableFreeWithCtx },
365	{ "RTHandleTableLookupWithCtx", (void *)RTHandleTableLookupWithCtx },
366	{ "RTNetIPv4AddDataChecksum", (void *)RTNetIPv4AddDataChecksum },
367	{ "RTNetIPv4AddTCPChecksum", (void *)RTNetIPv4AddTCPChecksum },
368	{ "RTNetIPv4AddUDPChecksum", (void *)RTNetIPv4AddUDPChecksum },
369	{ "RTNetIPv4FinalizeChecksum", (void *)RTNetIPv4FinalizeChecksum },
370	{ "RTNetIPv4HdrChecksum", (void *)RTNetIPv4HdrChecksum },
371	{ "RTNetIPv4IsDHCPValid", (void *)RTNetIPv4IsDHCPValid },
372	{ "RTNetIPv4IsHdrValid", (void *)RTNetIPv4IsHdrValid },
373	{ "RTNetIPv4IsTCPSizeValid", (void *)RTNetIPv4IsTCPSizeValid },
374	{ "RTNetIPv4IsTCPValid", (void *)RTNetIPv4IsTCPValid },
375	{ "RTNetIPv4IsUDPSizeValid", (void *)RTNetIPv4IsUDPSizeValid },
376	{ "RTNetIPv4IsUDPValid", (void *)RTNetIPv4IsUDPValid },
377	{ "RTNetIPv4PseudoChecksum", (void *)RTNetIPv4PseudoChecksum },
378	{ "RTNetIPv4PseudoChecksumBits", (void *)RTNetIPv4PseudoChecksumBits },
379	{ "RTNetIPv4TCPChecksum", (void *)RTNetIPv4TCPChecksum },
380	{ "RTNetIPv4UDPChecksum", (void *)RTNetIPv4UDPChecksum },
381	{ "RTNetIPv6PseudoChecksum", (void *)RTNetIPv6PseudoChecksum },
382	{ "RTNetIPv6PseudoChecksumBits", (void *)RTNetIPv6PseudoChecksumBits },
383	{ "RTNetIPv6PseudoChecksumEx", (void *)RTNetIPv6PseudoChecksumEx },
384	{ "RTNetTCPChecksum", (void *)RTNetTCPChecksum },
385	{ "RTNetUDPChecksum", (void *)RTNetUDPChecksum },
386	{ "RTCrc32", (void *)RTCrc32 },
387	{ "RTCrc32Finish", (void *)RTCrc32Finish },
388	{ "RTCrc32Process", (void *)RTCrc32Process },
389	{ "RTCrc32Start", (void *)RTCrc32Start },
390	};
391
392	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
393	/**
394	* Drag in the rest of IRPT since we share it with the
395	* rest of the kernel modules on darwin.
396	*/
397	PFNRT g_apfnVBoxDrvIPRTDeps[] =
398	{
399	/* VBoxNetFlt */
400	(PFNRT)RTErrConvertFromErrno,
401	(PFNRT)RTUuidCompare,
402	(PFNRT)RTUuidCompareStr,
403	(PFNRT)RTUuidFromStr,
404	(PFNRT)RTStrDupTag,
405	(PFNRT)RTStrFree,
406	/* VBoxNetAdp */
407	(PFNRT)RTRandBytes,
408	/* VBoxUSB */
409	(PFNRT)RTPathStripFilename,
410	NULL
411	};
412	#endif /* RT_OS_DARWIN \|\| RT_OS_SOLARIS \|\| RT_OS_SOLARIS */
413
414
415	/**
416	* Initializes the device extentsion structure.
417	*
418	* @returns IPRT status code.
419	* @param pDevExt The device extension to initialize.
420	* @param cbSession The size of the session structure. The size of
421	* SUPDRVSESSION may be smaller when SUPDRV_AGNOSTIC is
422	* defined because we're skipping the OS specific members
423	* then.
424	*/
425	int VBOXCALL supdrvInitDevExt(PSUPDRVDEVEXT pDevExt, size_t cbSession)
426	{
427	int rc;
428
429	#ifdef SUPDRV_WITH_RELEASE_LOGGER
430	/*
431	* Create the release log.
432	*/
433	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
434	PRTLOGGER pRelLogger;
435	rc = RTLogCreate(&pRelLogger, 0 /* fFlags */, "all",
436	"VBOX_RELEASE_LOG", RT_ELEMENTS(s_apszGroups), s_apszGroups,
437	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER, NULL);
438	if (RT_SUCCESS(rc))
439	RTLogRelSetDefaultInstance(pRelLogger);
440	/** @todo Add native hook for getting logger config parameters and setting
441	* them. On linux we should use the module parameter stuff... */
442	#endif
443
444	/*
445	* Initialize it.
446	*/
447	memset(pDevExt, 0, sizeof(*pDevExt));
448	rc = RTSpinlockCreate(&pDevExt->Spinlock);
449	if (RT_SUCCESS(rc))
450	{
451	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
452	rc = RTSemMutexCreate(&pDevExt->mtxLdr);
453	#else
454	rc = RTSemFastMutexCreate(&pDevExt->mtxLdr);
455	#endif
456	if (RT_SUCCESS(rc))
457	{
458	rc = RTSemFastMutexCreate(&pDevExt->mtxComponentFactory);
459	if (RT_SUCCESS(rc))
460	{
461	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
462	rc = RTSemMutexCreate(&pDevExt->mtxGip);
463	#else
464	rc = RTSemFastMutexCreate(&pDevExt->mtxGip);
465	#endif
466	if (RT_SUCCESS(rc))
467	{
468	rc = supdrvGipCreate(pDevExt);
469	if (RT_SUCCESS(rc))
470	{
471	pDevExt->u32Cookie = BIRD; /** @todo make this random? */
472	pDevExt->cbSession = (uint32_t)cbSession;
473
474	/*
475	* Fixup the absolute symbols.
476	*
477	* Because of the table indexing assumptions we'll have a little #ifdef orgy
478	* here rather than distributing this to OS specific files. At least for now.
479	*/
480	#ifdef RT_OS_DARWIN
481	# if ARCH_BITS == 32
482	if (SUPR0GetPagingMode() >= SUPPAGINGMODE_AMD64)
483	{
484	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
485	g_aFunctions[1].pfn = (void )0x80; / SUPR0Abs64bitKernelCS - KERNEL64_CS, seg.h */
486	g_aFunctions[2].pfn = (void )0x88; / SUPR0Abs64bitKernelSS - KERNEL64_SS, seg.h */
487	g_aFunctions[3].pfn = (void )0x88; / SUPR0Abs64bitKernelDS - KERNEL64_SS, seg.h */
488	}
489	else
490	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
491	g_aFunctions[4].pfn = (void )0x08; / SUPR0AbsKernelCS - KERNEL_CS, seg.h */
492	g_aFunctions[5].pfn = (void )0x10; / SUPR0AbsKernelSS - KERNEL_DS, seg.h */
493	g_aFunctions[6].pfn = (void )0x10; / SUPR0AbsKernelDS - KERNEL_DS, seg.h */
494	g_aFunctions[7].pfn = (void )0x10; / SUPR0AbsKernelES - KERNEL_DS, seg.h */
495	g_aFunctions[8].pfn = (void )0x10; / SUPR0AbsKernelFS - KERNEL_DS, seg.h */
496	g_aFunctions[9].pfn = (void )0x48; / SUPR0AbsKernelGS - CPU_DATA_GS, seg.h */
497	# else /* 64-bit darwin: */
498	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
499	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
500	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
501	g_aFunctions[3].pfn = (void )0; / SUPR0Abs64bitKernelDS */
502	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
503	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
504	g_aFunctions[6].pfn = (void )0; / SUPR0AbsKernelDS */
505	g_aFunctions[7].pfn = (void )0; / SUPR0AbsKernelES */
506	g_aFunctions[8].pfn = (void )0; / SUPR0AbsKernelFS */
507	g_aFunctions[9].pfn = (void )0; / SUPR0AbsKernelGS */
508
509	# endif
510	#else /* !RT_OS_DARWIN */
511	# if ARCH_BITS == 64
512	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
513	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
514	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
515	g_aFunctions[3].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0Abs64bitKernelDS */
516	# else
517	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
518	# endif
519	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
520	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
521	g_aFunctions[6].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0AbsKernelDS */
522	g_aFunctions[7].pfn = (void )(uintptr_t)ASMGetES(); / SUPR0AbsKernelES */
523	g_aFunctions[8].pfn = (void )(uintptr_t)ASMGetFS(); / SUPR0AbsKernelFS */
524	g_aFunctions[9].pfn = (void )(uintptr_t)ASMGetGS(); / SUPR0AbsKernelGS */
525	#endif /* !RT_OS_DARWIN */
526	return VINF_SUCCESS;
527	}
528
529	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
530	RTSemMutexDestroy(pDevExt->mtxGip);
531	pDevExt->mtxGip = NIL_RTSEMMUTEX;
532	#else
533	RTSemFastMutexDestroy(pDevExt->mtxGip);
534	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
535	#endif
536	}
537	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
538	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
539	}
540	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
541	RTSemMutexDestroy(pDevExt->mtxLdr);
542	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
543	#else
544	RTSemFastMutexDestroy(pDevExt->mtxLdr);
545	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
546	#endif
547	}
548	RTSpinlockDestroy(pDevExt->Spinlock);
549	pDevExt->Spinlock = NIL_RTSPINLOCK;
550	}
551	#ifdef SUPDRV_WITH_RELEASE_LOGGER
552	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
553	RTLogDestroy(RTLogSetDefaultInstance(NULL));
554	#endif
555
556	return rc;
557	}
558
559
560	/**
561	* Delete the device extension (e.g. cleanup members).
562	*
563	* @param pDevExt The device extension to delete.
564	*/
565	void VBOXCALL supdrvDeleteDevExt(PSUPDRVDEVEXT pDevExt)
566	{
567	PSUPDRVOBJ pObj;
568	PSUPDRVUSAGE pUsage;
569
570	/*
571	* Kill mutexes and spinlocks.
572	*/
573	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
574	RTSemMutexDestroy(pDevExt->mtxGip);
575	pDevExt->mtxGip = NIL_RTSEMMUTEX;
576	#else
577	RTSemFastMutexDestroy(pDevExt->mtxGip);
578	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
579	#endif
580	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
581	RTSemMutexDestroy(pDevExt->mtxLdr);
582	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
583	#else
584	RTSemFastMutexDestroy(pDevExt->mtxLdr);
585	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
586	#endif
587	RTSpinlockDestroy(pDevExt->Spinlock);
588	pDevExt->Spinlock = NIL_RTSPINLOCK;
589	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
590	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
591
592	/*
593	* Free lists.
594	*/
595	/* objects. */
596	pObj = pDevExt->pObjs;
597	Assert(!pObj); /* (can trigger on forced unloads) */
598	pDevExt->pObjs = NULL;
599	while (pObj)
600	{
601	void *pvFree = pObj;
602	pObj = pObj->pNext;
603	RTMemFree(pvFree);
604	}
605
606	/* usage records. */
607	pUsage = pDevExt->pUsageFree;
608	pDevExt->pUsageFree = NULL;
609	while (pUsage)
610	{
611	void *pvFree = pUsage;
612	pUsage = pUsage->pNext;
613	RTMemFree(pvFree);
614	}
615
616	/* kill the GIP. */
617	supdrvGipDestroy(pDevExt);
618
619	#ifdef SUPDRV_WITH_RELEASE_LOGGER
620	/* destroy the loggers. */
621	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
622	RTLogDestroy(RTLogSetDefaultInstance(NULL));
623	#endif
624	}
625
626
627	/**
628	* Create session.
629	*
630	* @returns IPRT status code.
631	* @param pDevExt Device extension.
632	* @param fUser Flag indicating whether this is a user or kernel session.
633	* @param ppSession Where to store the pointer to the session data.
634	*/
635	int VBOXCALL supdrvCreateSession(PSUPDRVDEVEXT pDevExt, bool fUser, PSUPDRVSESSION *ppSession)
636	{
637	/*
638	* Allocate memory for the session data.
639	*/
640	int rc;
641	PSUPDRVSESSION pSession = *ppSession = (PSUPDRVSESSION)RTMemAllocZ(pDevExt->cbSession);
642	if (pSession)
643	{
644	/* Initialize session data. */
645	rc = RTSpinlockCreate(&pSession->Spinlock);
646	if (!rc)
647	{
648	rc = RTHandleTableCreateEx(&pSession->hHandleTable,
649	RTHANDLETABLE_FLAGS_LOCKED \| RTHANDLETABLE_FLAGS_CONTEXT,
650	1 /uBase/, 32768 /cMax/, supdrvSessionObjHandleRetain, pSession);
651	if (RT_SUCCESS(rc))
652	{
653	Assert(pSession->Spinlock != NIL_RTSPINLOCK);
654	pSession->pDevExt = pDevExt;
655	pSession->u32Cookie = BIRD_INV;
656	/*pSession->pLdrUsage = NULL;
657	pSession->pVM = NULL;
658	pSession->pUsage = NULL;
659	pSession->pGip = NULL;
660	pSession->fGipReferenced = false;
661	pSession->Bundle.cUsed = 0; */
662	pSession->Uid = NIL_RTUID;
663	pSession->Gid = NIL_RTGID;
664	if (fUser)
665	{
666	pSession->Process = RTProcSelf();
667	pSession->R0Process = RTR0ProcHandleSelf();
668	}
669	else
670	{
671	pSession->Process = NIL_RTPROCESS;
672	pSession->R0Process = NIL_RTR0PROCESS;
673	}
674
675	VBOXDRV_SUPDRV_SESSION_CREATE(pSession, fUser);
676	LogFlow(("Created session %p initial cookie=%#x\n", pSession, pSession->u32Cookie));
677	return VINF_SUCCESS;
678	}
679
680	RTSpinlockDestroy(pSession->Spinlock);
681	}
682	RTMemFree(pSession);
683	*ppSession = NULL;
684	Log(("Failed to create spinlock, rc=%d!\n", rc));
685	}
686	else
687	rc = VERR_NO_MEMORY;
688
689	return rc;
690	}
691
692
693	/**
694	* Shared code for cleaning up a session.
695	*
696	* @param pDevExt Device extension.
697	* @param pSession Session data.
698	* This data will be freed by this routine.
699	*/
700	void VBOXCALL supdrvCloseSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
701	{
702	VBOXDRV_SUPDRV_SESSION_CLOSE(pSession);
703
704	/*
705	* Cleanup the session first.
706	*/
707	supdrvCleanupSession(pDevExt, pSession);
708
709	/*
710	* Free the rest of the session stuff.
711	*/
712	RTSpinlockDestroy(pSession->Spinlock);
713	pSession->Spinlock = NIL_RTSPINLOCK;
714	pSession->pDevExt = NULL;
715	RTMemFree(pSession);
716	LogFlow(("supdrvCloseSession: returns\n"));
717	}
718
719
720	/**
721	* Shared code for cleaning up a session (but not quite freeing it).
722	*
723	* This is primarily intended for MAC OS X where we have to clean up the memory
724	* stuff before the file handle is closed.
725	*
726	* @param pDevExt Device extension.
727	* @param pSession Session data.
728	* This data will be freed by this routine.
729	*/
730	void VBOXCALL supdrvCleanupSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
731	{
732	int rc;
733	PSUPDRVBUNDLE pBundle;
734	LogFlow(("supdrvCleanupSession: pSession=%p\n", pSession));
735
736	/*
737	* Remove logger instances related to this session.
738	*/
739	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pSession);
740
741	/*
742	* Destroy the handle table.
743	*/
744	rc = RTHandleTableDestroy(pSession->hHandleTable, supdrvSessionObjHandleDelete, pSession);
745	AssertRC(rc);
746	pSession->hHandleTable = NIL_RTHANDLETABLE;
747
748	/*
749	* Release object references made in this session.
750	* In theory there should be noone racing us in this session.
751	*/
752	Log2(("release objects - start\n"));
753	if (pSession->pUsage)
754	{
755	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
756	PSUPDRVUSAGE pUsage;
757	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
758
759	while ((pUsage = pSession->pUsage) != NULL)
760	{
761	PSUPDRVOBJ pObj = pUsage->pObj;
762	pSession->pUsage = pUsage->pNext;
763
764	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
765	if (pUsage->cUsage < pObj->cUsage)
766	{
767	pObj->cUsage -= pUsage->cUsage;
768	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
769	}
770	else
771	{
772	/* Destroy the object and free the record. */
773	if (pDevExt->pObjs == pObj)
774	pDevExt->pObjs = pObj->pNext;
775	else
776	{
777	PSUPDRVOBJ pObjPrev;
778	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
779	if (pObjPrev->pNext == pObj)
780	{
781	pObjPrev->pNext = pObj->pNext;
782	break;
783	}
784	Assert(pObjPrev);
785	}
786	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
787
788	Log(("supdrvCleanupSession: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
789	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
790	if (pObj->pfnDestructor)
791	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
792	RTMemFree(pObj);
793	}
794
795	/* free it and continue. */
796	RTMemFree(pUsage);
797
798	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
799	}
800
801	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
802	AssertMsg(!pSession->pUsage, ("Some buster reregistered an object during desturction!\n"));
803	}
804	Log2(("release objects - done\n"));
805
806	/*
807	* Release memory allocated in the session.
808	*
809	* We do not serialize this as we assume that the application will
810	* not allocated memory while closing the file handle object.
811	*/
812	Log2(("freeing memory:\n"));
813	pBundle = &pSession->Bundle;
814	while (pBundle)
815	{
816	PSUPDRVBUNDLE pToFree;
817	unsigned i;
818
819	/*
820	* Check and unlock all entries in the bundle.
821	*/
822	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
823	{
824	if (pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ)
825	{
826	Log2(("eType=%d pvR0=%p pvR3=%p cb=%ld\n", pBundle->aMem[i].eType, RTR0MemObjAddress(pBundle->aMem[i].MemObj),
827	(void *)RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3), (long)RTR0MemObjSize(pBundle->aMem[i].MemObj)));
828	if (pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ)
829	{
830	rc = RTR0MemObjFree(pBundle->aMem[i].MapObjR3, false);
831	AssertRC(rc); /** @todo figure out how to handle this. */
832	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
833	}
834	rc = RTR0MemObjFree(pBundle->aMem[i].MemObj, true /* fFreeMappings */);
835	AssertRC(rc); /** @todo figure out how to handle this. */
836	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
837	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
838	}
839	}
840
841	/*
842	* Advance and free previous bundle.
843	*/
844	pToFree = pBundle;
845	pBundle = pBundle->pNext;
846
847	pToFree->pNext = NULL;
848	pToFree->cUsed = 0;
849	if (pToFree != &pSession->Bundle)
850	RTMemFree(pToFree);
851	}
852	Log2(("freeing memory - done\n"));
853
854	/*
855	* Deregister component factories.
856	*/
857	RTSemFastMutexRequest(pDevExt->mtxComponentFactory);
858	Log2(("deregistering component factories:\n"));
859	if (pDevExt->pComponentFactoryHead)
860	{
861	PSUPDRVFACTORYREG pPrev = NULL;
862	PSUPDRVFACTORYREG pCur = pDevExt->pComponentFactoryHead;
863	while (pCur)
864	{
865	if (pCur->pSession == pSession)
866	{
867	/* unlink it */
868	PSUPDRVFACTORYREG pNext = pCur->pNext;
869	if (pPrev)
870	pPrev->pNext = pNext;
871	else
872	pDevExt->pComponentFactoryHead = pNext;
873
874	/* free it */
875	pCur->pNext = NULL;
876	pCur->pSession = NULL;
877	pCur->pFactory = NULL;
878	RTMemFree(pCur);
879
880	/* next */
881	pCur = pNext;
882	}
883	else
884	{
885	/* next */
886	pPrev = pCur;
887	pCur = pCur->pNext;
888	}
889	}
890	}
891	RTSemFastMutexRelease(pDevExt->mtxComponentFactory);
892	Log2(("deregistering component factories - done\n"));
893
894	/*
895	* Loaded images needs to be dereferenced and possibly freed up.
896	*/
897	supdrvLdrLock(pDevExt);
898	Log2(("freeing images:\n"));
899	if (pSession->pLdrUsage)
900	{
901	PSUPDRVLDRUSAGE pUsage = pSession->pLdrUsage;
902	pSession->pLdrUsage = NULL;
903	while (pUsage)
904	{
905	void *pvFree = pUsage;
906	PSUPDRVLDRIMAGE pImage = pUsage->pImage;
907	if (pImage->cUsage > pUsage->cUsage)
908	pImage->cUsage -= pUsage->cUsage;
909	else
910	supdrvLdrFree(pDevExt, pImage);
911	pUsage->pImage = NULL;
912	pUsage = pUsage->pNext;
913	RTMemFree(pvFree);
914	}
915	}
916	supdrvLdrUnlock(pDevExt);
917	Log2(("freeing images - done\n"));
918
919	/*
920	* Unmap the GIP.
921	*/
922	Log2(("umapping GIP:\n"));
923	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
924	{
925	SUPR0GipUnmap(pSession);
926	pSession->fGipReferenced = 0;
927	}
928	Log2(("umapping GIP - done\n"));
929	}
930
931
932	/**
933	* RTHandleTableDestroy callback used by supdrvCleanupSession.
934	*
935	* @returns IPRT status code, see SUPR0ObjAddRef.
936	* @param hHandleTable The handle table handle. Ignored.
937	* @param pvObj The object pointer.
938	* @param pvCtx Context, the handle type. Ignored.
939	* @param pvUser Session pointer.
940	*/
941	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser)
942	{
943	NOREF(pvCtx);
944	NOREF(hHandleTable);
945	return SUPR0ObjAddRefEx(pvObj, (PSUPDRVSESSION)pvUser, true /fNoBlocking/);
946	}
947
948
949	/**
950	* RTHandleTableDestroy callback used by supdrvCleanupSession.
951	*
952	* @param hHandleTable The handle table handle. Ignored.
953	* @param h The handle value. Ignored.
954	* @param pvObj The object pointer.
955	* @param pvCtx Context, the handle type. Ignored.
956	* @param pvUser Session pointer.
957	*/
958	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser)
959	{
960	NOREF(pvCtx);
961	NOREF(h);
962	NOREF(hHandleTable);
963	SUPR0ObjRelease(pvObj, (PSUPDRVSESSION)pvUser);
964	}
965
966
967	/**
968	* Fast path I/O Control worker.
969	*
970	* @returns VBox status code that should be passed down to ring-3 unchanged.
971	* @param uIOCtl Function number.
972	* @param idCpu VMCPU id.
973	* @param pDevExt Device extention.
974	* @param pSession Session data.
975	*/
976	int VBOXCALL supdrvIOCtlFast(uintptr_t uIOCtl, VMCPUID idCpu, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
977	{
978	/*
979	* We check the two prereqs after doing this only to allow the compiler to optimize things better.
980	*/
981	if (RT_LIKELY(pSession->pVM && pDevExt->pfnVMMR0EntryFast))
982	{
983	switch (uIOCtl)
984	{
985	case SUP_IOCTL_FAST_DO_RAW_RUN:
986	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_RAW_RUN);
987	break;
988	case SUP_IOCTL_FAST_DO_HWACC_RUN:
989	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_HWACC_RUN);
990	break;
991	case SUP_IOCTL_FAST_DO_NOP:
992	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_NOP);
993	break;
994	default:
995	return VERR_INTERNAL_ERROR;
996	}
997	return VINF_SUCCESS;
998	}
999	return VERR_INTERNAL_ERROR;
1000	}
1001
1002
1003	/**
1004	* Helper for supdrvIOCtl. Check if pszStr contains any character of pszChars.
1005	* We would use strpbrk here if this function would be contained in the RedHat kABI white
1006	* list, see http://www.kerneldrivers.org/RHEL5.
1007	*
1008	* @returns 1 if pszStr does contain any character of pszChars, 0 otherwise.
1009	* @param pszStr String to check
1010	* @param pszChars Character set
1011	*/
1012	static int supdrvCheckInvalidChar(const char pszStr, const char pszChars)
1013	{
1014	int chCur;
1015	while ((chCur = *pszStr++) != '\0')
1016	{
1017	int ch;
1018	const char *psz = pszChars;
1019	while ((ch = *psz++) != '\0')
1020	if (ch == chCur)
1021	return 1;
1022
1023	}
1024	return 0;
1025	}
1026
1027
1028	/**
1029	* I/O Control worker.
1030	*
1031	* @returns IPRT status code.
1032	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1033	*
1034	* @param uIOCtl Function number.
1035	* @param pDevExt Device extention.
1036	* @param pSession Session data.
1037	* @param pReqHdr The request header.
1038	*/
1039	int VBOXCALL supdrvIOCtl(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1040	{
1041	VBOXDRV_SUPDRV_IOCTL_ENTRY(pSession, uIOCtl, pReqHdr);
1042
1043	/*
1044	* Validate the request.
1045	*/
1046	/* this first check could probably be omitted as its also done by the OS specific code... */
1047	if (RT_UNLIKELY( (pReqHdr->fFlags & SUPREQHDR_FLAGS_MAGIC_MASK) != SUPREQHDR_FLAGS_MAGIC
1048	\|\| pReqHdr->cbIn < sizeof(*pReqHdr)
1049	\|\| pReqHdr->cbOut < sizeof(*pReqHdr)))
1050	{
1051	OSDBGPRINT(("vboxdrv: Bad ioctl request header; cbIn=%#lx cbOut=%#lx fFlags=%#lx\n",
1052	(long)pReqHdr->cbIn, (long)pReqHdr->cbOut, (long)pReqHdr->fFlags));
1053	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1054	return VERR_INVALID_PARAMETER;
1055	}
1056	if (RT_UNLIKELY(uIOCtl == SUP_IOCTL_COOKIE))
1057	{
1058	if (pReqHdr->u32Cookie != SUPCOOKIE_INITIAL_COOKIE)
1059	{
1060	OSDBGPRINT(("SUP_IOCTL_COOKIE: bad cookie %#lx\n", (long)pReqHdr->u32Cookie));
1061	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1062	return VERR_INVALID_PARAMETER;
1063	}
1064	}
1065	else if (RT_UNLIKELY( pReqHdr->u32Cookie != pDevExt->u32Cookie
1066	\|\| pReqHdr->u32SessionCookie != pSession->u32Cookie))
1067	{
1068	OSDBGPRINT(("vboxdrv: bad cookie %#lx / %#lx.\n", (long)pReqHdr->u32Cookie, (long)pReqHdr->u32SessionCookie));
1069	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1070	return VERR_INVALID_PARAMETER;
1071	}
1072
1073	/*
1074	* Validation macros
1075	*/
1076	#define REQ_CHECK_SIZES_EX(Name, cbInExpect, cbOutExpect) \
1077	do { \
1078	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect) \|\| pReqHdr->cbOut != (cbOutExpect))) \
1079	{ \
1080	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld. cbOut=%ld expected %ld.\n", \
1081	(long)pReq->Hdr.cbIn, (long)(cbInExpect), (long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
1082	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1083	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1084	} \
1085	} while (0)
1086
1087	#define REQ_CHECK_SIZES(Name) REQ_CHECK_SIZES_EX(Name, Name ## _SIZE_IN, Name ## _SIZE_OUT)
1088
1089	#define REQ_CHECK_SIZE_IN(Name, cbInExpect) \
1090	do { \
1091	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect))) \
1092	{ \
1093	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld.\n", \
1094	(long)pReq->Hdr.cbIn, (long)(cbInExpect))); \
1095	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1096	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1097	} \
1098	} while (0)
1099
1100	#define REQ_CHECK_SIZE_OUT(Name, cbOutExpect) \
1101	do { \
1102	if (RT_UNLIKELY(pReqHdr->cbOut != (cbOutExpect))) \
1103	{ \
1104	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbOut=%ld expected %ld.\n", \
1105	(long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
1106	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1107	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1108	} \
1109	} while (0)
1110
1111	#define REQ_CHECK_EXPR(Name, expr) \
1112	do { \
1113	if (RT_UNLIKELY(!(expr))) \
1114	{ \
1115	OSDBGPRINT(( #Name ": %s\n", #expr)); \
1116	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1117	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1118	} \
1119	} while (0)
1120
1121	#define REQ_CHECK_EXPR_FMT(expr, fmt) \
1122	do { \
1123	if (RT_UNLIKELY(!(expr))) \
1124	{ \
1125	OSDBGPRINT( fmt ); \
1126	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1127	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1128	} \
1129	} while (0)
1130
1131
1132	/*
1133	* The switch.
1134	*/
1135	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
1136	{
1137	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
1138	{
1139	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
1140	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
1141	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
1142	{
1143	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
1144	pReq->Hdr.rc = VERR_INVALID_MAGIC;
1145	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, VERR_INVALID_MAGIC);
1146	return 0;
1147	}
1148
1149	#if 0
1150	/*
1151	* Call out to the OS specific code and let it do permission checks on the
1152	* client process.
1153	*/
1154	if (!supdrvOSValidateClientProcess(pDevExt, pSession))
1155	{
1156	pReq->u.Out.u32Cookie = 0xffffffff;
1157	pReq->u.Out.u32SessionCookie = 0xffffffff;
1158	pReq->u.Out.u32SessionVersion = 0xffffffff;
1159	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1160	pReq->u.Out.pSession = NULL;
1161	pReq->u.Out.cFunctions = 0;
1162	pReq->Hdr.rc = VERR_PERMISSION_DENIED;
1163	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, VERR_PERMISSION_DENIED);
1164	return 0;
1165	}
1166	#endif
1167
1168	/*
1169	* Match the version.
1170	* The current logic is very simple, match the major interface version.
1171	*/
1172	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
1173	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
1174	{
1175	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1176	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
1177	pReq->u.Out.u32Cookie = 0xffffffff;
1178	pReq->u.Out.u32SessionCookie = 0xffffffff;
1179	pReq->u.Out.u32SessionVersion = 0xffffffff;
1180	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1181	pReq->u.Out.pSession = NULL;
1182	pReq->u.Out.cFunctions = 0;
1183	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1184	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1185	return 0;
1186	}
1187
1188	/*
1189	* Fill in return data and be gone.
1190	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
1191	* u32SessionVersion <= u32ReqVersion!
1192	*/
1193	/** @todo Somehow validate the client and negotiate a secure cookie... */
1194	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
1195	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
1196	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
1197	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1198	pReq->u.Out.pSession = pSession;
1199	pReq->u.Out.cFunctions = sizeof(g_aFunctions) / sizeof(g_aFunctions[0]);
1200	pReq->Hdr.rc = VINF_SUCCESS;
1201	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1202	return 0;
1203	}
1204
1205	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_QUERY_FUNCS(0)):
1206	{
1207	/* validate */
1208	PSUPQUERYFUNCS pReq = (PSUPQUERYFUNCS)pReqHdr;
1209	REQ_CHECK_SIZES_EX(SUP_IOCTL_QUERY_FUNCS, SUP_IOCTL_QUERY_FUNCS_SIZE_IN, SUP_IOCTL_QUERY_FUNCS_SIZE_OUT(RT_ELEMENTS(g_aFunctions)));
1210
1211	/* execute */
1212	pReq->u.Out.cFunctions = RT_ELEMENTS(g_aFunctions);
1213	memcpy(&pReq->u.Out.aFunctions[0], g_aFunctions, sizeof(g_aFunctions));
1214	pReq->Hdr.rc = VINF_SUCCESS;
1215	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1216	return 0;
1217	}
1218
1219	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_LOCK):
1220	{
1221	/* validate */
1222	PSUPPAGELOCK pReq = (PSUPPAGELOCK)pReqHdr;
1223	REQ_CHECK_SIZE_IN(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_IN);
1224	REQ_CHECK_SIZE_OUT(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_OUT(pReq->u.In.cPages));
1225	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.cPages > 0);
1226	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.pvR3 >= PAGE_SIZE);
1227
1228	/* execute */
1229	pReq->Hdr.rc = SUPR0LockMem(pSession, pReq->u.In.pvR3, pReq->u.In.cPages, &pReq->u.Out.aPages[0]);
1230	if (RT_FAILURE(pReq->Hdr.rc))
1231	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1232	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1233	return 0;
1234	}
1235
1236	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_UNLOCK):
1237	{
1238	/* validate */
1239	PSUPPAGEUNLOCK pReq = (PSUPPAGEUNLOCK)pReqHdr;
1240	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_UNLOCK);
1241
1242	/* execute */
1243	pReq->Hdr.rc = SUPR0UnlockMem(pSession, pReq->u.In.pvR3);
1244	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1245	return 0;
1246	}
1247
1248	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_ALLOC):
1249	{
1250	/* validate */
1251	PSUPCONTALLOC pReq = (PSUPCONTALLOC)pReqHdr;
1252	REQ_CHECK_SIZES(SUP_IOCTL_CONT_ALLOC);
1253
1254	/* execute */
1255	pReq->Hdr.rc = SUPR0ContAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.HCPhys);
1256	if (RT_FAILURE(pReq->Hdr.rc))
1257	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1258	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1259	return 0;
1260	}
1261
1262	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_FREE):
1263	{
1264	/* validate */
1265	PSUPCONTFREE pReq = (PSUPCONTFREE)pReqHdr;
1266	REQ_CHECK_SIZES(SUP_IOCTL_CONT_FREE);
1267
1268	/* execute */
1269	pReq->Hdr.rc = SUPR0ContFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1270	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1271	return 0;
1272	}
1273
1274	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_OPEN):
1275	{
1276	/* validate */
1277	PSUPLDROPEN pReq = (PSUPLDROPEN)pReqHdr;
1278	REQ_CHECK_SIZES(SUP_IOCTL_LDR_OPEN);
1279	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs > 0);
1280	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs < 16*_1M);
1281	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1282	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1283	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits < pReq->u.In.cbImageWithTabs);
1284	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.szName[0]);
1285	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1286	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, !supdrvCheckInvalidChar(pReq->u.In.szName, ";:()[]{}/\\\|&*%#@!~`\"'"));
1287	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szFilename, sizeof(pReq->u.In.szFilename)));
1288
1289	/* execute */
1290	pReq->Hdr.rc = supdrvIOCtl_LdrOpen(pDevExt, pSession, pReq);
1291	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1292	return 0;
1293	}
1294
1295	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOAD):
1296	{
1297	/* validate */
1298	PSUPLDRLOAD pReq = (PSUPLDRLOAD)pReqHdr;
1299	REQ_CHECK_EXPR(Name, pReq->Hdr.cbIn >= sizeof(*pReq));
1300	REQ_CHECK_SIZES_EX(SUP_IOCTL_LDR_LOAD, SUP_IOCTL_LDR_LOAD_SIZE_IN(pReq->u.In.cbImageWithTabs), SUP_IOCTL_LDR_LOAD_SIZE_OUT);
1301	REQ_CHECK_EXPR(SUP_IOCTL_LDR_LOAD, pReq->u.In.cSymbols <= 16384);
1302	REQ_CHECK_EXPR_FMT( !pReq->u.In.cSymbols
1303	\|\| ( pReq->u.In.offSymbols < pReq->u.In.cbImageWithTabs
1304	&& pReq->u.In.offSymbols + pReq->u.In.cSymbols * sizeof(SUPLDRSYM) <= pReq->u.In.cbImageWithTabs),
1305	("SUP_IOCTL_LDR_LOAD: offSymbols=%#lx cSymbols=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offSymbols,
1306	(long)pReq->u.In.cSymbols, (long)pReq->u.In.cbImageWithTabs));
1307	REQ_CHECK_EXPR_FMT( !pReq->u.In.cbStrTab
1308	\|\| ( pReq->u.In.offStrTab < pReq->u.In.cbImageWithTabs
1309	&& pReq->u.In.offStrTab + pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs
1310	&& pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs),
1311	("SUP_IOCTL_LDR_LOAD: offStrTab=%#lx cbStrTab=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offStrTab,
1312	(long)pReq->u.In.cbStrTab, (long)pReq->u.In.cbImageWithTabs));
1313
1314	if (pReq->u.In.cSymbols)
1315	{
1316	uint32_t i;
1317	PSUPLDRSYM paSyms = (PSUPLDRSYM)&pReq->u.In.abImage[pReq->u.In.offSymbols];
1318	for (i = 0; i < pReq->u.In.cSymbols; i++)
1319	{
1320	REQ_CHECK_EXPR_FMT(paSyms[i].offSymbol < pReq->u.In.cbImageWithTabs,
1321	("SUP_IOCTL_LDR_LOAD: sym #%ld: symb off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offSymbol, (long)pReq->u.In.cbImageWithTabs));
1322	REQ_CHECK_EXPR_FMT(paSyms[i].offName < pReq->u.In.cbStrTab,
1323	("SUP_IOCTL_LDR_LOAD: sym #%ld: name off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1324	REQ_CHECK_EXPR_FMT(RTStrEnd((char const *)&pReq->u.In.abImage[pReq->u.In.offStrTab + paSyms[i].offName],
1325	pReq->u.In.cbStrTab - paSyms[i].offName),
1326	("SUP_IOCTL_LDR_LOAD: sym #%ld: unterminated name! (%#lx / %#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1327	}
1328	}
1329
1330	/* execute */
1331	pReq->Hdr.rc = supdrvIOCtl_LdrLoad(pDevExt, pSession, pReq);
1332	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1333	return 0;
1334	}
1335
1336	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_FREE):
1337	{
1338	/* validate */
1339	PSUPLDRFREE pReq = (PSUPLDRFREE)pReqHdr;
1340	REQ_CHECK_SIZES(SUP_IOCTL_LDR_FREE);
1341
1342	/* execute */
1343	pReq->Hdr.rc = supdrvIOCtl_LdrFree(pDevExt, pSession, pReq);
1344	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1345	return 0;
1346	}
1347
1348	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_GET_SYMBOL):
1349	{
1350	/* validate */
1351	PSUPLDRGETSYMBOL pReq = (PSUPLDRGETSYMBOL)pReqHdr;
1352	REQ_CHECK_SIZES(SUP_IOCTL_LDR_GET_SYMBOL);
1353	REQ_CHECK_EXPR(SUP_IOCTL_LDR_GET_SYMBOL, RTStrEnd(pReq->u.In.szSymbol, sizeof(pReq->u.In.szSymbol)));
1354
1355	/* execute */
1356	pReq->Hdr.rc = supdrvIOCtl_LdrGetSymbol(pDevExt, pSession, pReq);
1357	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1358	return 0;
1359	}
1360
1361	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0(0)):
1362	{
1363	/* validate */
1364	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1365	Log4(("SUP_IOCTL_CALL_VMMR0: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1366	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1367
1368	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_VMMR0_SIZE(0))
1369	{
1370	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(0), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(0));
1371
1372	/* execute */
1373	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1374	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg, pSession);
1375	else
1376	pReq->Hdr.rc = VERR_WRONG_ORDER;
1377	}
1378	else
1379	{
1380	PSUPVMMR0REQHDR pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1381	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR)),
1382	("SUP_IOCTL_CALL_VMMR0: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR))));
1383	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1384	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(pVMMReq->cbReq));
1385
1386	/* execute */
1387	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1388	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1389	else
1390	pReq->Hdr.rc = VERR_WRONG_ORDER;
1391	}
1392
1393	if ( RT_FAILURE(pReq->Hdr.rc)
1394	&& pReq->Hdr.rc != VERR_INTERRUPTED
1395	&& pReq->Hdr.rc != VERR_TIMEOUT)
1396	Log(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1397	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1398	else
1399	Log4(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1400	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1401	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1402	return 0;
1403	}
1404
1405	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0_BIG):
1406	{
1407	/* validate */
1408	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1409	PSUPVMMR0REQHDR pVMMReq;
1410	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1411	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1412
1413	pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1414	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR)),
1415	("SUP_IOCTL_CALL_VMMR0_BIG: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR))));
1416	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0_BIG, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1417	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0_BIG, SUP_IOCTL_CALL_VMMR0_BIG_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_BIG_SIZE_OUT(pVMMReq->cbReq));
1418
1419	/* execute */
1420	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1421	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1422	else
1423	pReq->Hdr.rc = VERR_WRONG_ORDER;
1424
1425	if ( RT_FAILURE(pReq->Hdr.rc)
1426	&& pReq->Hdr.rc != VERR_INTERRUPTED
1427	&& pReq->Hdr.rc != VERR_TIMEOUT)
1428	Log(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1429	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1430	else
1431	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1432	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1433	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1434	return 0;
1435	}
1436
1437	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_PAGING_MODE):
1438	{
1439	/* validate */
1440	PSUPGETPAGINGMODE pReq = (PSUPGETPAGINGMODE)pReqHdr;
1441	REQ_CHECK_SIZES(SUP_IOCTL_GET_PAGING_MODE);
1442
1443	/* execute */
1444	pReq->Hdr.rc = VINF_SUCCESS;
1445	pReq->u.Out.enmMode = SUPR0GetPagingMode();
1446	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1447	return 0;
1448	}
1449
1450	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_ALLOC):
1451	{
1452	/* validate */
1453	PSUPLOWALLOC pReq = (PSUPLOWALLOC)pReqHdr;
1454	REQ_CHECK_EXPR(SUP_IOCTL_LOW_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_LOW_ALLOC_SIZE_IN);
1455	REQ_CHECK_SIZES_EX(SUP_IOCTL_LOW_ALLOC, SUP_IOCTL_LOW_ALLOC_SIZE_IN, SUP_IOCTL_LOW_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1456
1457	/* execute */
1458	pReq->Hdr.rc = SUPR0LowAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1459	if (RT_FAILURE(pReq->Hdr.rc))
1460	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1461	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1462	return 0;
1463	}
1464
1465	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_FREE):
1466	{
1467	/* validate */
1468	PSUPLOWFREE pReq = (PSUPLOWFREE)pReqHdr;
1469	REQ_CHECK_SIZES(SUP_IOCTL_LOW_FREE);
1470
1471	/* execute */
1472	pReq->Hdr.rc = SUPR0LowFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1473	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1474	return 0;
1475	}
1476
1477	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_MAP):
1478	{
1479	/* validate */
1480	PSUPGIPMAP pReq = (PSUPGIPMAP)pReqHdr;
1481	REQ_CHECK_SIZES(SUP_IOCTL_GIP_MAP);
1482
1483	/* execute */
1484	pReq->Hdr.rc = SUPR0GipMap(pSession, &pReq->u.Out.pGipR3, &pReq->u.Out.HCPhysGip);
1485	if (RT_SUCCESS(pReq->Hdr.rc))
1486	pReq->u.Out.pGipR0 = pDevExt->pGip;
1487	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1488	return 0;
1489	}
1490
1491	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_UNMAP):
1492	{
1493	/* validate */
1494	PSUPGIPUNMAP pReq = (PSUPGIPUNMAP)pReqHdr;
1495	REQ_CHECK_SIZES(SUP_IOCTL_GIP_UNMAP);
1496
1497	/* execute */
1498	pReq->Hdr.rc = SUPR0GipUnmap(pSession);
1499	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1500	return 0;
1501	}
1502
1503	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SET_VM_FOR_FAST):
1504	{
1505	/* validate */
1506	PSUPSETVMFORFAST pReq = (PSUPSETVMFORFAST)pReqHdr;
1507	REQ_CHECK_SIZES(SUP_IOCTL_SET_VM_FOR_FAST);
1508	REQ_CHECK_EXPR_FMT( !pReq->u.In.pVMR0
1509	\|\| ( VALID_PTR(pReq->u.In.pVMR0)
1510	&& !((uintptr_t)pReq->u.In.pVMR0 & (PAGE_SIZE - 1))),
1511	("SUP_IOCTL_SET_VM_FOR_FAST: pVMR0=%p!\n", pReq->u.In.pVMR0));
1512	/* execute */
1513	pSession->pVM = pReq->u.In.pVMR0;
1514	pReq->Hdr.rc = VINF_SUCCESS;
1515	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1516	return 0;
1517	}
1518
1519	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_ALLOC_EX):
1520	{
1521	/* validate */
1522	PSUPPAGEALLOCEX pReq = (PSUPPAGEALLOCEX)pReqHdr;
1523	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC_EX, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_EX_SIZE_IN);
1524	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));
1525	REQ_CHECK_EXPR_FMT(pReq->u.In.fKernelMapping \|\| pReq->u.In.fUserMapping,
1526	("SUP_IOCTL_PAGE_ALLOC_EX: No mapping requested!\n"));
1527	REQ_CHECK_EXPR_FMT(pReq->u.In.fUserMapping,
1528	("SUP_IOCTL_PAGE_ALLOC_EX: Must have user mapping!\n"));
1529	REQ_CHECK_EXPR_FMT(!pReq->u.In.fReserved0 && !pReq->u.In.fReserved1,
1530	("SUP_IOCTL_PAGE_ALLOC_EX: fReserved0=%d fReserved1=%d\n", pReq->u.In.fReserved0, pReq->u.In.fReserved1));
1531
1532	/* execute */
1533	pReq->Hdr.rc = SUPR0PageAllocEx(pSession, pReq->u.In.cPages, 0 /* fFlags */,
1534	pReq->u.In.fUserMapping ? &pReq->u.Out.pvR3 : NULL,
1535	pReq->u.In.fKernelMapping ? &pReq->u.Out.pvR0 : NULL,
1536	&pReq->u.Out.aPages[0]);
1537	if (RT_FAILURE(pReq->Hdr.rc))
1538	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1539	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1540	return 0;
1541	}
1542
1543	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_MAP_KERNEL):
1544	{
1545	/* validate */
1546	PSUPPAGEMAPKERNEL pReq = (PSUPPAGEMAPKERNEL)pReqHdr;
1547	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_MAP_KERNEL);
1548	REQ_CHECK_EXPR_FMT(!pReq->u.In.fFlags, ("SUP_IOCTL_PAGE_MAP_KERNEL: fFlags=%#x! MBZ\n", pReq->u.In.fFlags));
1549	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_MAP_KERNEL: offSub=%#x\n", pReq->u.In.offSub));
1550	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1551	("SUP_IOCTL_PAGE_MAP_KERNEL: cbSub=%#x\n", pReq->u.In.cbSub));
1552
1553	/* execute */
1554	pReq->Hdr.rc = SUPR0PageMapKernel(pSession, pReq->u.In.pvR3, pReq->u.In.offSub, pReq->u.In.cbSub,
1555	pReq->u.In.fFlags, &pReq->u.Out.pvR0);
1556	if (RT_FAILURE(pReq->Hdr.rc))
1557	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1558	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1559	return 0;
1560	}
1561
1562	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_PROTECT):
1563	{
1564	/* validate */
1565	PSUPPAGEPROTECT pReq = (PSUPPAGEPROTECT)pReqHdr;
1566	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_PROTECT);
1567	REQ_CHECK_EXPR_FMT(!(pReq->u.In.fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)),
1568	("SUP_IOCTL_PAGE_PROTECT: fProt=%#x!\n", pReq->u.In.fProt));
1569	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_PROTECT: offSub=%#x\n", pReq->u.In.offSub));
1570	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1571	("SUP_IOCTL_PAGE_PROTECT: cbSub=%#x\n", pReq->u.In.cbSub));
1572
1573	/* execute */
1574	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);
1575	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1576	return 0;
1577	}
1578
1579	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_FREE):
1580	{
1581	/* validate */
1582	PSUPPAGEFREE pReq = (PSUPPAGEFREE)pReqHdr;
1583	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_FREE);
1584
1585	/* execute */
1586	pReq->Hdr.rc = SUPR0PageFree(pSession, pReq->u.In.pvR3);
1587	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1588	return 0;
1589	}
1590
1591	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_SERVICE(0)):
1592	{
1593	/* validate */
1594	PSUPCALLSERVICE pReq = (PSUPCALLSERVICE)pReqHdr;
1595	Log4(("SUP_IOCTL_CALL_SERVICE: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1596	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1597
1598	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
1599	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(0), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(0));
1600	else
1601	{
1602	PSUPR0SERVICEREQHDR pSrvReq = (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0];
1603	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR)),
1604	("SUP_IOCTL_CALL_SERVICE: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR))));
1605	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, pSrvReq->u32Magic == SUPR0SERVICEREQHDR_MAGIC);
1606	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(pSrvReq->cbReq), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(pSrvReq->cbReq));
1607	}
1608	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1609
1610	/* execute */
1611	pReq->Hdr.rc = supdrvIOCtl_CallServiceModule(pDevExt, pSession, pReq);
1612	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1613	return 0;
1614	}
1615
1616	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOGGER_SETTINGS(0)):
1617	{
1618	/* validate */
1619	PSUPLOGGERSETTINGS pReq = (PSUPLOGGERSETTINGS)pReqHdr;
1620	size_t cbStrTab;
1621	REQ_CHECK_SIZE_OUT(SUP_IOCTL_LOGGER_SETTINGS, SUP_IOCTL_LOGGER_SETTINGS_SIZE_OUT);
1622	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->Hdr.cbIn >= SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(1));
1623	cbStrTab = pReq->Hdr.cbIn - SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(0);
1624	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offGroups < cbStrTab);
1625	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offFlags < cbStrTab);
1626	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offDestination < cbStrTab);
1627	REQ_CHECK_EXPR_FMT(pReq->u.In.szStrings[cbStrTab - 1] == '\0',
1628	("SUP_IOCTL_LOGGER_SETTINGS: cbIn=%#x cbStrTab=%#zx LastChar=%d\n",
1629	pReq->Hdr.cbIn, cbStrTab, pReq->u.In.szStrings[cbStrTab - 1]));
1630	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhich <= SUPLOGGERSETTINGS_WHICH_RELEASE);
1631	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhat <= SUPLOGGERSETTINGS_WHAT_DESTROY);
1632
1633	/* execute */
1634	pReq->Hdr.rc = supdrvIOCtl_LoggerSettings(pDevExt, pSession, pReq);
1635	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1636	return 0;
1637	}
1638
1639	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP2):
1640	{
1641	/* validate */
1642	PSUPSEMOP2 pReq = (PSUPSEMOP2)pReqHdr;
1643	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP2, SUP_IOCTL_SEM_OP2_SIZE_IN, SUP_IOCTL_SEM_OP2_SIZE_OUT);
1644	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP2, pReq->u.In.uReserved == 0);
1645
1646	/* execute */
1647	switch (pReq->u.In.uType)
1648	{
1649	case SUP_SEM_TYPE_EVENT:
1650	{
1651	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1652	switch (pReq->u.In.uOp)
1653	{
1654	case SUPSEMOP2_WAIT_MS_REL:
1655	pReq->Hdr.rc = SUPSemEventWaitNoResume(pSession, hEvent, pReq->u.In.uArg.cRelMsTimeout);
1656	break;
1657	case SUPSEMOP2_WAIT_NS_ABS:
1658	pReq->Hdr.rc = SUPSemEventWaitNsAbsIntr(pSession, hEvent, pReq->u.In.uArg.uAbsNsTimeout);
1659	break;
1660	case SUPSEMOP2_WAIT_NS_REL:
1661	pReq->Hdr.rc = SUPSemEventWaitNsRelIntr(pSession, hEvent, pReq->u.In.uArg.cRelNsTimeout);
1662	break;
1663	case SUPSEMOP2_SIGNAL:
1664	pReq->Hdr.rc = SUPSemEventSignal(pSession, hEvent);
1665	break;
1666	case SUPSEMOP2_CLOSE:
1667	pReq->Hdr.rc = SUPSemEventClose(pSession, hEvent);
1668	break;
1669	case SUPSEMOP2_RESET:
1670	default:
1671	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1672	break;
1673	}
1674	break;
1675	}
1676
1677	case SUP_SEM_TYPE_EVENT_MULTI:
1678	{
1679	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1680	switch (pReq->u.In.uOp)
1681	{
1682	case SUPSEMOP2_WAIT_MS_REL:
1683	pReq->Hdr.rc = SUPSemEventMultiWaitNoResume(pSession, hEventMulti, pReq->u.In.uArg.cRelMsTimeout);
1684	break;
1685	case SUPSEMOP2_WAIT_NS_ABS:
1686	pReq->Hdr.rc = SUPSemEventMultiWaitNsAbsIntr(pSession, hEventMulti, pReq->u.In.uArg.uAbsNsTimeout);
1687	break;
1688	case SUPSEMOP2_WAIT_NS_REL:
1689	pReq->Hdr.rc = SUPSemEventMultiWaitNsRelIntr(pSession, hEventMulti, pReq->u.In.uArg.cRelNsTimeout);
1690	break;
1691	case SUPSEMOP2_SIGNAL:
1692	pReq->Hdr.rc = SUPSemEventMultiSignal(pSession, hEventMulti);
1693	break;
1694	case SUPSEMOP2_CLOSE:
1695	pReq->Hdr.rc = SUPSemEventMultiClose(pSession, hEventMulti);
1696	break;
1697	case SUPSEMOP2_RESET:
1698	pReq->Hdr.rc = SUPSemEventMultiReset(pSession, hEventMulti);
1699	break;
1700	default:
1701	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1702	break;
1703	}
1704	break;
1705	}
1706
1707	default:
1708	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1709	break;
1710	}
1711	return 0;
1712	}
1713
1714	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP3):
1715	{
1716	/* validate */
1717	PSUPSEMOP3 pReq = (PSUPSEMOP3)pReqHdr;
1718	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP3, SUP_IOCTL_SEM_OP3_SIZE_IN, SUP_IOCTL_SEM_OP3_SIZE_OUT);
1719	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, pReq->u.In.u32Reserved == 0 && pReq->u.In.u64Reserved == 0);
1720
1721	/* execute */
1722	switch (pReq->u.In.uType)
1723	{
1724	case SUP_SEM_TYPE_EVENT:
1725	{
1726	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1727	switch (pReq->u.In.uOp)
1728	{
1729	case SUPSEMOP3_CREATE:
1730	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1731	pReq->Hdr.rc = SUPSemEventCreate(pSession, &hEvent);
1732	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEvent;
1733	break;
1734	case SUPSEMOP3_GET_RESOLUTION:
1735	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1736	pReq->Hdr.rc = VINF_SUCCESS;
1737	pReq->Hdr.cbOut = sizeof(*pReq);
1738	pReq->u.Out.cNsResolution = SUPSemEventGetResolution(pSession);
1739	break;
1740	default:
1741	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1742	break;
1743	}
1744	break;
1745	}
1746
1747	case SUP_SEM_TYPE_EVENT_MULTI:
1748	{
1749	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1750	switch (pReq->u.In.uOp)
1751	{
1752	case SUPSEMOP3_CREATE:
1753	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1754	pReq->Hdr.rc = SUPSemEventMultiCreate(pSession, &hEventMulti);
1755	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEventMulti;
1756	break;
1757	case SUPSEMOP3_GET_RESOLUTION:
1758	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1759	pReq->Hdr.rc = VINF_SUCCESS;
1760	pReq->u.Out.cNsResolution = SUPSemEventMultiGetResolution(pSession);
1761	break;
1762	default:
1763	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1764	break;
1765	}
1766	break;
1767	}
1768
1769	default:
1770	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1771	break;
1772	}
1773	return 0;
1774	}
1775
1776	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
1777	{
1778	/* validate */
1779	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
1780	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
1781	REQ_CHECK_EXPR(SUP_IOCTL_VT_CAPS, pReq->Hdr.cbIn <= SUP_IOCTL_VT_CAPS_SIZE_IN);
1782
1783	/* execute */
1784	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.Caps);
1785	if (RT_FAILURE(pReq->Hdr.rc))
1786	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1787	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1788	return 0;
1789	}
1790
1791	default:
1792	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1793	break;
1794	}
1795	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_GENERAL_FAILURE, VERR_GENERAL_FAILURE);
1796	return VERR_GENERAL_FAILURE;
1797	}
1798
1799
1800	/**
1801	* Inter-Driver Communication (IDC) worker.
1802	*
1803	* @returns VBox status code.
1804	* @retval VINF_SUCCESS on success.
1805	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1806	* @retval VERR_NOT_SUPPORTED if the request isn't supported.
1807	*
1808	* @param uReq The request (function) code.
1809	* @param pDevExt Device extention.
1810	* @param pSession Session data.
1811	* @param pReqHdr The request header.
1812	*/
1813	int VBOXCALL supdrvIDC(uintptr_t uReq, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQHDR pReqHdr)
1814	{
1815	/*
1816	* The OS specific code has already validated the pSession
1817	* pointer, and the request size being greater or equal to
1818	* size of the header.
1819	*
1820	* So, just check that pSession is a kernel context session.
1821	*/
1822	if (RT_UNLIKELY( pSession
1823	&& pSession->R0Process != NIL_RTR0PROCESS))
1824	return VERR_INVALID_PARAMETER;
1825
1826	/*
1827	* Validation macro.
1828	*/
1829	#define REQ_CHECK_IDC_SIZE(Name, cbExpect) \
1830	do { \
1831	if (RT_UNLIKELY(pReqHdr->cb != (cbExpect))) \
1832	{ \
1833	OSDBGPRINT(( #Name ": Invalid input/output sizes. cb=%ld expected %ld.\n", \
1834	(long)pReqHdr->cb, (long)(cbExpect))); \
1835	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1836	} \
1837	} while (0)
1838
1839	switch (uReq)
1840	{
1841	case SUPDRV_IDC_REQ_CONNECT:
1842	{
1843	PSUPDRVIDCREQCONNECT pReq = (PSUPDRVIDCREQCONNECT)pReqHdr;
1844	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_CONNECT, sizeof(*pReq));
1845
1846	/*
1847	* Validate the cookie and other input.
1848	*/
1849	if (pReq->Hdr.pSession != NULL)
1850	{
1851	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: pSession=%p expected NULL!\n", pReq->Hdr.pSession));
1852	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1853	}
1854	if (pReq->u.In.u32MagicCookie != SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE)
1855	{
1856	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: u32MagicCookie=%#x expected %#x!\n",
1857	(unsigned)pReq->u.In.u32MagicCookie, (unsigned)SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE));
1858	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1859	}
1860	if ( pReq->u.In.uMinVersion > pReq->u.In.uReqVersion
1861	\|\| (pReq->u.In.uMinVersion & UINT32_C(0xffff0000)) != (pReq->u.In.uReqVersion & UINT32_C(0xffff0000)))
1862	{
1863	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: uMinVersion=%#x uMaxVersion=%#x doesn't match!\n",
1864	pReq->u.In.uMinVersion, pReq->u.In.uReqVersion));
1865	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1866	}
1867
1868	/*
1869	* Match the version.
1870	* The current logic is very simple, match the major interface version.
1871	*/
1872	if ( pReq->u.In.uMinVersion > SUPDRV_IDC_VERSION
1873	\|\| (pReq->u.In.uMinVersion & 0xffff0000) != (SUPDRV_IDC_VERSION & 0xffff0000))
1874	{
1875	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1876	pReq->u.In.uReqVersion, pReq->u.In.uMinVersion, (unsigned)SUPDRV_IDC_VERSION));
1877	pReq->u.Out.pSession = NULL;
1878	pReq->u.Out.uSessionVersion = 0xffffffff;
1879	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1880	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1881	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1882	return VINF_SUCCESS;
1883	}
1884
1885	pReq->u.Out.pSession = NULL;
1886	pReq->u.Out.uSessionVersion = SUPDRV_IDC_VERSION;
1887	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1888	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1889
1890	/*
1891	* On NT we will already have a session associated with the
1892	* client, just like with the SUP_IOCTL_COOKIE request, while
1893	* the other doesn't.
1894	*/
1895	#ifdef RT_OS_WINDOWS
1896	pReq->Hdr.rc = VINF_SUCCESS;
1897	#else
1898	AssertReturn(!pSession, VERR_INTERNAL_ERROR);
1899	pReq->Hdr.rc = supdrvCreateSession(pDevExt, false /* fUser */, &pSession);
1900	if (RT_FAILURE(pReq->Hdr.rc))
1901	{
1902	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: failed to create session, rc=%d\n", pReq->Hdr.rc));
1903	return VINF_SUCCESS;
1904	}
1905	#endif
1906
1907	pReq->u.Out.pSession = pSession;
1908	pReq->Hdr.pSession = pSession;
1909
1910	return VINF_SUCCESS;
1911	}
1912
1913	case SUPDRV_IDC_REQ_DISCONNECT:
1914	{
1915	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_DISCONNECT, sizeof(*pReqHdr));
1916
1917	#ifdef RT_OS_WINDOWS
1918	/* Windows will destroy the session when the file object is destroyed. */
1919	#else
1920	supdrvCloseSession(pDevExt, pSession);
1921	#endif
1922	return pReqHdr->rc = VINF_SUCCESS;
1923	}
1924
1925	case SUPDRV_IDC_REQ_GET_SYMBOL:
1926	{
1927	PSUPDRVIDCREQGETSYM pReq = (PSUPDRVIDCREQGETSYM)pReqHdr;
1928	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_GET_SYMBOL, sizeof(*pReq));
1929
1930	pReq->Hdr.rc = supdrvIDC_LdrGetSymbol(pDevExt, pSession, pReq);
1931	return VINF_SUCCESS;
1932	}
1933
1934	case SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY:
1935	{
1936	PSUPDRVIDCREQCOMPREGFACTORY pReq = (PSUPDRVIDCREQCOMPREGFACTORY)pReqHdr;
1937	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY, sizeof(*pReq));
1938
1939	pReq->Hdr.rc = SUPR0ComponentRegisterFactory(pSession, pReq->u.In.pFactory);
1940	return VINF_SUCCESS;
1941	}
1942
1943	case SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY:
1944	{
1945	PSUPDRVIDCREQCOMPDEREGFACTORY pReq = (PSUPDRVIDCREQCOMPDEREGFACTORY)pReqHdr;
1946	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY, sizeof(*pReq));
1947
1948	pReq->Hdr.rc = SUPR0ComponentDeregisterFactory(pSession, pReq->u.In.pFactory);
1949	return VINF_SUCCESS;
1950	}
1951
1952	default:
1953	Log(("Unknown IDC %#lx\n", (long)uReq));
1954	break;
1955	}
1956
1957	#undef REQ_CHECK_IDC_SIZE
1958	return VERR_NOT_SUPPORTED;
1959	}
1960
1961
1962	/**
1963	* Register a object for reference counting.
1964	* The object is registered with one reference in the specified session.
1965	*
1966	* @returns Unique identifier on success (pointer).
1967	* All future reference must use this identifier.
1968	* @returns NULL on failure.
1969	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
1970	* @param pvUser1 The first user argument.
1971	* @param pvUser2 The second user argument.
1972	*/
1973	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
1974	{
1975	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1976	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1977	PSUPDRVOBJ pObj;
1978	PSUPDRVUSAGE pUsage;
1979
1980	/*
1981	* Validate the input.
1982	*/
1983	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
1984	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
1985	AssertPtrReturn(pfnDestructor, NULL);
1986
1987	/*
1988	* Allocate and initialize the object.
1989	*/
1990	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
1991	if (!pObj)
1992	return NULL;
1993	pObj->u32Magic = SUPDRVOBJ_MAGIC;
1994	pObj->enmType = enmType;
1995	pObj->pNext = NULL;
1996	pObj->cUsage = 1;
1997	pObj->pfnDestructor = pfnDestructor;
1998	pObj->pvUser1 = pvUser1;
1999	pObj->pvUser2 = pvUser2;
2000	pObj->CreatorUid = pSession->Uid;
2001	pObj->CreatorGid = pSession->Gid;
2002	pObj->CreatorProcess= pSession->Process;
2003	supdrvOSObjInitCreator(pObj, pSession);
2004
2005	/*
2006	* Allocate the usage record.
2007	* (We keep freed usage records around to simplify SUPR0ObjAddRefEx().)
2008	*/
2009	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2010
2011	pUsage = pDevExt->pUsageFree;
2012	if (pUsage)
2013	pDevExt->pUsageFree = pUsage->pNext;
2014	else
2015	{
2016	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2017	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
2018	if (!pUsage)
2019	{
2020	RTMemFree(pObj);
2021	return NULL;
2022	}
2023	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2024	}
2025
2026	/*
2027	* Insert the object and create the session usage record.
2028	*/
2029	/* The object. */
2030	pObj->pNext = pDevExt->pObjs;
2031	pDevExt->pObjs = pObj;
2032
2033	/* The session record. */
2034	pUsage->cUsage = 1;
2035	pUsage->pObj = pObj;
2036	pUsage->pNext = pSession->pUsage;
2037	/* Log2(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext)); */
2038	pSession->pUsage = pUsage;
2039
2040	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2041
2042	Log(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
2043	return pObj;
2044	}
2045
2046
2047	/**
2048	* Increment the reference counter for the object associating the reference
2049	* with the specified session.
2050	*
2051	* @returns IPRT status code.
2052	* @param pvObj The identifier returned by SUPR0ObjRegister().
2053	* @param pSession The session which is referencing the object.
2054	*
2055	* @remarks The caller should not own any spinlocks and must carefully protect
2056	* itself against potential race with the destructor so freed memory
2057	* isn't accessed here.
2058	*/
2059	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
2060	{
2061	return SUPR0ObjAddRefEx(pvObj, pSession, false /* fNoBlocking */);
2062	}
2063
2064
2065	/**
2066	* Increment the reference counter for the object associating the reference
2067	* with the specified session.
2068	*
2069	* @returns IPRT status code.
2070	* @retval VERR_TRY_AGAIN if fNoBlocking was set and a new usage record
2071	* couldn't be allocated. (If you see this you're not doing the right
2072	* thing and it won't ever work reliably.)
2073	*
2074	* @param pvObj The identifier returned by SUPR0ObjRegister().
2075	* @param pSession The session which is referencing the object.
2076	* @param fNoBlocking Set if it's not OK to block. Never try to make the
2077	* first reference to an object in a session with this
2078	* argument set.
2079	*
2080	* @remarks The caller should not own any spinlocks and must carefully protect
2081	* itself against potential race with the destructor so freed memory
2082	* isn't accessed here.
2083	*/
2084	SUPR0DECL(int) SUPR0ObjAddRefEx(void *pvObj, PSUPDRVSESSION pSession, bool fNoBlocking)
2085	{
2086	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2087	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2088	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2089	int rc = VINF_SUCCESS;
2090	PSUPDRVUSAGE pUsagePre;
2091	PSUPDRVUSAGE pUsage;
2092
2093	/*
2094	* Validate the input.
2095	* Be ready for the destruction race (someone might be stuck in the
2096	* destructor waiting a lock we own).
2097	*/
2098	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2099	AssertPtrReturn(pObj, VERR_INVALID_POINTER);
2100	AssertMsgReturn(pObj->u32Magic == SUPDRVOBJ_MAGIC \|\| pObj->u32Magic == SUPDRVOBJ_MAGIC_DEAD,
2101	("Invalid pvObj=%p magic=%#x (expected %#x or %#x)\n", pvObj, pObj->u32Magic, SUPDRVOBJ_MAGIC, SUPDRVOBJ_MAGIC_DEAD),
2102	VERR_INVALID_PARAMETER);
2103
2104	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2105
2106	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2107	{
2108	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2109
2110	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2111	return VERR_WRONG_ORDER;
2112	}
2113
2114	/*
2115	* Preallocate the usage record if we can.
2116	*/
2117	pUsagePre = pDevExt->pUsageFree;
2118	if (pUsagePre)
2119	pDevExt->pUsageFree = pUsagePre->pNext;
2120	else if (!fNoBlocking)
2121	{
2122	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2123	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
2124	if (!pUsagePre)
2125	return VERR_NO_MEMORY;
2126
2127	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2128	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2129	{
2130	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2131
2132	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2133	return VERR_WRONG_ORDER;
2134	}
2135	}
2136
2137	/*
2138	* Reference the object.
2139	*/
2140	pObj->cUsage++;
2141
2142	/*
2143	* Look for the session record.
2144	*/
2145	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
2146	{
2147	/Log(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2148	if (pUsage->pObj == pObj)
2149	break;
2150	}
2151	if (pUsage)
2152	pUsage->cUsage++;
2153	else if (pUsagePre)
2154	{
2155	/* create a new session record. */
2156	pUsagePre->cUsage = 1;
2157	pUsagePre->pObj = pObj;
2158	pUsagePre->pNext = pSession->pUsage;
2159	pSession->pUsage = pUsagePre;
2160	/Log(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));/
2161
2162	pUsagePre = NULL;
2163	}
2164	else
2165	{
2166	pObj->cUsage--;
2167	rc = VERR_TRY_AGAIN;
2168	}
2169
2170	/*
2171	* Put any unused usage record into the free list..
2172	*/
2173	if (pUsagePre)
2174	{
2175	pUsagePre->pNext = pDevExt->pUsageFree;
2176	pDevExt->pUsageFree = pUsagePre;
2177	}
2178
2179	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2180
2181	return rc;
2182	}
2183
2184
2185	/**
2186	* Decrement / destroy a reference counter record for an object.
2187	*
2188	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
2189	*
2190	* @returns IPRT status code.
2191	* @retval VINF_SUCCESS if not destroyed.
2192	* @retval VINF_OBJECT_DESTROYED if it's destroyed by this release call.
2193	* @retval VERR_INVALID_PARAMETER if the object isn't valid. Will assert in
2194	* string builds.
2195	*
2196	* @param pvObj The identifier returned by SUPR0ObjRegister().
2197	* @param pSession The session which is referencing the object.
2198	*/
2199	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
2200	{
2201	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2202	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2203	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2204	int rc = VERR_INVALID_PARAMETER;
2205	PSUPDRVUSAGE pUsage;
2206	PSUPDRVUSAGE pUsagePrev;
2207
2208	/*
2209	* Validate the input.
2210	*/
2211	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2212	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2213	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2214	VERR_INVALID_PARAMETER);
2215
2216	/*
2217	* Acquire the spinlock and look for the usage record.
2218	*/
2219	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2220
2221	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
2222	pUsage;
2223	pUsagePrev = pUsage, pUsage = pUsage->pNext)
2224	{
2225	/Log2(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2226	if (pUsage->pObj == pObj)
2227	{
2228	rc = VINF_SUCCESS;
2229	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
2230	if (pUsage->cUsage > 1)
2231	{
2232	pObj->cUsage--;
2233	pUsage->cUsage--;
2234	}
2235	else
2236	{
2237	/*
2238	* Free the session record.
2239	*/
2240	if (pUsagePrev)
2241	pUsagePrev->pNext = pUsage->pNext;
2242	else
2243	pSession->pUsage = pUsage->pNext;
2244	pUsage->pNext = pDevExt->pUsageFree;
2245	pDevExt->pUsageFree = pUsage;
2246
2247	/* What about the object? */
2248	if (pObj->cUsage > 1)
2249	pObj->cUsage--;
2250	else
2251	{
2252	/*
2253	* Object is to be destroyed, unlink it.
2254	*/
2255	pObj->u32Magic = SUPDRVOBJ_MAGIC_DEAD;
2256	rc = VINF_OBJECT_DESTROYED;
2257	if (pDevExt->pObjs == pObj)
2258	pDevExt->pObjs = pObj->pNext;
2259	else
2260	{
2261	PSUPDRVOBJ pObjPrev;
2262	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
2263	if (pObjPrev->pNext == pObj)
2264	{
2265	pObjPrev->pNext = pObj->pNext;
2266	break;
2267	}
2268	Assert(pObjPrev);
2269	}
2270	}
2271	}
2272	break;
2273	}
2274	}
2275
2276	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2277
2278	/*
2279	* Call the destructor and free the object if required.
2280	*/
2281	if (rc == VINF_OBJECT_DESTROYED)
2282	{
2283	Log(("SUPR0ObjRelease: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
2284	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
2285	if (pObj->pfnDestructor)
2286	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
2287	RTMemFree(pObj);
2288	}
2289
2290	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
2291	return rc;
2292	}
2293
2294
2295	/**
2296	* Verifies that the current process can access the specified object.
2297	*
2298	* @returns The following IPRT status code:
2299	* @retval VINF_SUCCESS if access was granted.
2300	* @retval VERR_PERMISSION_DENIED if denied access.
2301	* @retval VERR_INVALID_PARAMETER if invalid parameter.
2302	*
2303	* @param pvObj The identifier returned by SUPR0ObjRegister().
2304	* @param pSession The session which wishes to access the object.
2305	* @param pszObjName Object string name. This is optional and depends on the object type.
2306	*
2307	* @remark The caller is responsible for making sure the object isn't removed while
2308	* we're inside this function. If uncertain about this, just call AddRef before calling us.
2309	*/
2310	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
2311	{
2312	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2313	int rc;
2314
2315	/*
2316	* Validate the input.
2317	*/
2318	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2319	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2320	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2321	VERR_INVALID_PARAMETER);
2322
2323	/*
2324	* Check access. (returns true if a decision has been made.)
2325	*/
2326	rc = VERR_INTERNAL_ERROR;
2327	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
2328	return rc;
2329
2330	/*
2331	* Default policy is to allow the user to access his own
2332	* stuff but nothing else.
2333	*/
2334	if (pObj->CreatorUid == pSession->Uid)
2335	return VINF_SUCCESS;
2336	return VERR_PERMISSION_DENIED;
2337	}
2338
2339
2340	/**
2341	* Lock pages.
2342	*
2343	* @returns IPRT status code.
2344	* @param pSession Session to which the locked memory should be associated.
2345	* @param pvR3 Start of the memory range to lock.
2346	* This must be page aligned.
2347	* @param cPages Number of pages to lock.
2348	* @param paPages Where to put the physical addresses of locked memory.
2349	*/
2350	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
2351	{
2352	int rc;
2353	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2354	const size_t cb = (size_t)cPages << PAGE_SHIFT;
2355	LogFlow(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
2356
2357	/*
2358	* Verify input.
2359	*/
2360	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2361	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
2362	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
2363	\|\| !pvR3)
2364	{
2365	Log(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
2366	return VERR_INVALID_PARAMETER;
2367	}
2368
2369	/*
2370	* Let IPRT do the job.
2371	*/
2372	Mem.eType = MEMREF_TYPE_LOCKED;
2373	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTMEM_PROT_READ \| RTMEM_PROT_WRITE, RTR0ProcHandleSelf());
2374	if (RT_SUCCESS(rc))
2375	{
2376	uint32_t iPage = cPages;
2377	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
2378	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
2379
2380	while (iPage-- > 0)
2381	{
2382	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2383	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
2384	{
2385	AssertMsgFailed(("iPage=%d\n", iPage));
2386	rc = VERR_INTERNAL_ERROR;
2387	break;
2388	}
2389	}
2390	if (RT_SUCCESS(rc))
2391	rc = supdrvMemAdd(&Mem, pSession);
2392	if (RT_FAILURE(rc))
2393	{
2394	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
2395	AssertRC(rc2);
2396	}
2397	}
2398
2399	return rc;
2400	}
2401
2402
2403	/**
2404	* Unlocks the memory pointed to by pv.
2405	*
2406	* @returns IPRT status code.
2407	* @param pSession Session to which the memory was locked.
2408	* @param pvR3 Memory to unlock.
2409	*/
2410	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2411	{
2412	LogFlow(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2413	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2414	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
2415	}
2416
2417
2418	/**
2419	* Allocates a chunk of page aligned memory with contiguous and fixed physical
2420	* backing.
2421	*
2422	* @returns IPRT status code.
2423	* @param pSession Session data.
2424	* @param cPages Number of pages to allocate.
2425	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
2426	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
2427	* @param pHCPhys Where to put the physical address of allocated memory.
2428	*/
2429	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
2430	{
2431	int rc;
2432	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2433	LogFlow(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
2434
2435	/*
2436	* Validate input.
2437	*/
2438	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2439	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
2440	{
2441	Log(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
2442	pSession, ppvR0, ppvR3, pHCPhys));
2443	return VERR_INVALID_PARAMETER;
2444
2445	}
2446	if (cPages < 1 \|\| cPages >= 256)
2447	{
2448	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2449	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2450	}
2451
2452	/*
2453	* Let IPRT do the job.
2454	*/
2455	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
2456	if (RT_SUCCESS(rc))
2457	{
2458	int rc2;
2459	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2460	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2461	if (RT_SUCCESS(rc))
2462	{
2463	Mem.eType = MEMREF_TYPE_CONT;
2464	rc = supdrvMemAdd(&Mem, pSession);
2465	if (!rc)
2466	{
2467	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2468	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2469	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
2470	return 0;
2471	}
2472
2473	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2474	AssertRC(rc2);
2475	}
2476	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2477	AssertRC(rc2);
2478	}
2479
2480	return rc;
2481	}
2482
2483
2484	/**
2485	* Frees memory allocated using SUPR0ContAlloc().
2486	*
2487	* @returns IPRT status code.
2488	* @param pSession The session to which the memory was allocated.
2489	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2490	*/
2491	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2492	{
2493	LogFlow(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2494	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2495	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
2496	}
2497
2498
2499	/**
2500	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
2501	*
2502	* The memory isn't zeroed.
2503	*
2504	* @returns IPRT status code.
2505	* @param pSession Session data.
2506	* @param cPages Number of pages to allocate.
2507	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
2508	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
2509	* @param paPages Where to put the physical addresses of allocated memory.
2510	*/
2511	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
2512	{
2513	unsigned iPage;
2514	int rc;
2515	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2516	LogFlow(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
2517
2518	/*
2519	* Validate input.
2520	*/
2521	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2522	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
2523	{
2524	Log(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
2525	pSession, ppvR3, ppvR0, paPages));
2526	return VERR_INVALID_PARAMETER;
2527
2528	}
2529	if (cPages < 1 \|\| cPages >= 256)
2530	{
2531	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2532	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2533	}
2534
2535	/*
2536	* Let IPRT do the work.
2537	*/
2538	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
2539	if (RT_SUCCESS(rc))
2540	{
2541	int rc2;
2542	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2543	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2544	if (RT_SUCCESS(rc))
2545	{
2546	Mem.eType = MEMREF_TYPE_LOW;
2547	rc = supdrvMemAdd(&Mem, pSession);
2548	if (!rc)
2549	{
2550	for (iPage = 0; iPage < cPages; iPage++)
2551	{
2552	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2553	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%RHp\n", paPages[iPage]));
2554	}
2555	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2556	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2557	return 0;
2558	}
2559
2560	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2561	AssertRC(rc2);
2562	}
2563
2564	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2565	AssertRC(rc2);
2566	}
2567
2568	return rc;
2569	}
2570
2571
2572	/**
2573	* Frees memory allocated using SUPR0LowAlloc().
2574	*
2575	* @returns IPRT status code.
2576	* @param pSession The session to which the memory was allocated.
2577	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2578	*/
2579	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2580	{
2581	LogFlow(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2582	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2583	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
2584	}
2585
2586
2587
2588	/**
2589	* Allocates a chunk of memory with both R0 and R3 mappings.
2590	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
2591	*
2592	* @returns IPRT status code.
2593	* @param pSession The session to associated the allocation with.
2594	* @param cb Number of bytes to allocate.
2595	* @param ppvR0 Where to store the address of the Ring-0 mapping.
2596	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2597	*/
2598	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
2599	{
2600	int rc;
2601	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2602	LogFlow(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
2603
2604	/*
2605	* Validate input.
2606	*/
2607	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2608	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
2609	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
2610	if (cb < 1 \|\| cb >= _4M)
2611	{
2612	Log(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
2613	return VERR_INVALID_PARAMETER;
2614	}
2615
2616	/*
2617	* Let IPRT do the work.
2618	*/
2619	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
2620	if (RT_SUCCESS(rc))
2621	{
2622	int rc2;
2623	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2624	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2625	if (RT_SUCCESS(rc))
2626	{
2627	Mem.eType = MEMREF_TYPE_MEM;
2628	rc = supdrvMemAdd(&Mem, pSession);
2629	if (!rc)
2630	{
2631	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2632	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2633	return VINF_SUCCESS;
2634	}
2635
2636	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2637	AssertRC(rc2);
2638	}
2639
2640	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2641	AssertRC(rc2);
2642	}
2643
2644	return rc;
2645	}
2646
2647
2648	/**
2649	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
2650	*
2651	* @returns IPRT status code.
2652	* @param pSession The session to which the memory was allocated.
2653	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2654	* @param paPages Where to store the physical addresses.
2655	*/
2656	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
2657	{
2658	PSUPDRVBUNDLE pBundle;
2659	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2660	LogFlow(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
2661
2662	/*
2663	* Validate input.
2664	*/
2665	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2666	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
2667	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
2668
2669	/*
2670	* Search for the address.
2671	*/
2672	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2673	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2674	{
2675	if (pBundle->cUsed > 0)
2676	{
2677	unsigned i;
2678	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2679	{
2680	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
2681	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2682	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2683	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2684	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
2685	)
2686	)
2687	{
2688	const size_t cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2689	size_t iPage;
2690	for (iPage = 0; iPage < cPages; iPage++)
2691	{
2692	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2693	paPages[iPage].uReserved = 0;
2694	}
2695	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2696	return VINF_SUCCESS;
2697	}
2698	}
2699	}
2700	}
2701	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2702	Log(("Failed to find %p!!!\n", (void *)uPtr));
2703	return VERR_INVALID_PARAMETER;
2704	}
2705
2706
2707	/**
2708	* Free memory allocated by SUPR0MemAlloc().
2709	*
2710	* @returns IPRT status code.
2711	* @param pSession The session owning the allocation.
2712	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2713	*/
2714	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2715	{
2716	LogFlow(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2717	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2718	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
2719	}
2720
2721
2722	/**
2723	* Allocates a chunk of memory with a kernel or/and a user mode mapping.
2724	*
2725	* The memory is fixed and it's possible to query the physical addresses using
2726	* SUPR0MemGetPhys().
2727	*
2728	* @returns IPRT status code.
2729	* @param pSession The session to associated the allocation with.
2730	* @param cPages The number of pages to allocate.
2731	* @param fFlags Flags, reserved for the future. Must be zero.
2732	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2733	* NULL if no ring-3 mapping.
2734	* @param ppvR3 Where to store the address of the Ring-0 mapping.
2735	* NULL if no ring-0 mapping.
2736	* @param paPages Where to store the addresses of the pages. Optional.
2737	*/
2738	SUPR0DECL(int) SUPR0PageAllocEx(PSUPDRVSESSION pSession, uint32_t cPages, uint32_t fFlags, PRTR3PTR ppvR3, PRTR0PTR ppvR0, PRTHCPHYS paPages)
2739	{
2740	int rc;
2741	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2742	LogFlow(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
2743
2744	/*
2745	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2746	*/
2747	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2748	AssertPtrNullReturn(ppvR3, VERR_INVALID_POINTER);
2749	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2750	AssertReturn(ppvR3 \|\| ppvR0, VERR_INVALID_PARAMETER);
2751	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2752	if (cPages < 1 \|\| cPages > VBOX_MAX_ALLOC_PAGE_COUNT)
2753	{
2754	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)));
2755	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2756	}
2757
2758	/*
2759	* Let IPRT do the work.
2760	*/
2761	if (ppvR0)
2762	rc = RTR0MemObjAllocPage(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, true /* fExecutable */);
2763	else
2764	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
2765	if (RT_SUCCESS(rc))
2766	{
2767	int rc2;
2768	if (ppvR3)
2769	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2770	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2771	else
2772	Mem.MapObjR3 = NIL_RTR0MEMOBJ;
2773	if (RT_SUCCESS(rc))
2774	{
2775	Mem.eType = MEMREF_TYPE_PAGE;
2776	rc = supdrvMemAdd(&Mem, pSession);
2777	if (!rc)
2778	{
2779	if (ppvR3)
2780	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2781	if (ppvR0)
2782	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2783	if (paPages)
2784	{
2785	uint32_t iPage = cPages;
2786	while (iPage-- > 0)
2787	{
2788	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
2789	Assert(paPages[iPage] != NIL_RTHCPHYS);
2790	}
2791	}
2792	return VINF_SUCCESS;
2793	}
2794
2795	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2796	AssertRC(rc2);
2797	}
2798
2799	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2800	AssertRC(rc2);
2801	}
2802	return rc;
2803	}
2804
2805
2806	/**
2807	* Maps a chunk of memory previously allocated by SUPR0PageAllocEx into kernel
2808	* space.
2809	*
2810	* @returns IPRT status code.
2811	* @param pSession The session to associated the allocation with.
2812	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
2813	* @param offSub Where to start mapping. Must be page aligned.
2814	* @param cbSub How much to map. Must be page aligned.
2815	* @param fFlags Flags, MBZ.
2816	* @param ppvR0 Where to return the address of the ring-0 mapping on
2817	* success.
2818	*/
2819	SUPR0DECL(int) SUPR0PageMapKernel(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t offSub, uint32_t cbSub,
2820	uint32_t fFlags, PRTR0PTR ppvR0)
2821	{
2822	int rc;
2823	PSUPDRVBUNDLE pBundle;
2824	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2825	RTR0MEMOBJ hMemObj = NIL_RTR0MEMOBJ;
2826	LogFlow(("SUPR0PageMapKernel: pSession=%p pvR3=%p offSub=%#x cbSub=%#x\n", pSession, pvR3, offSub, cbSub));
2827
2828	/*
2829	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2830	*/
2831	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2832	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2833	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2834	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2835	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2836	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
2837
2838	/*
2839	* Find the memory object.
2840	*/
2841	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2842	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2843	{
2844	if (pBundle->cUsed > 0)
2845	{
2846	unsigned i;
2847	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2848	{
2849	if ( ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
2850	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2851	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2852	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
2853	\|\| ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED
2854	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2855	&& pBundle->aMem[i].MapObjR3 == NIL_RTR0MEMOBJ
2856	&& RTR0MemObjAddressR3(pBundle->aMem[i].MemObj) == pvR3))
2857	{
2858	hMemObj = pBundle->aMem[i].MemObj;
2859	break;
2860	}
2861	}
2862	}
2863	}
2864	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2865
2866	rc = VERR_INVALID_PARAMETER;
2867	if (hMemObj != NIL_RTR0MEMOBJ)
2868	{
2869	/*
2870	* Do some further input validations before calling IPRT.
2871	* (Cleanup is done indirectly by telling RTR0MemObjFree to include mappings.)
2872	*/
2873	size_t cbMemObj = RTR0MemObjSize(hMemObj);
2874	if ( offSub < cbMemObj
2875	&& cbSub <= cbMemObj
2876	&& offSub + cbSub <= cbMemObj)
2877	{
2878	RTR0MEMOBJ hMapObj;
2879	rc = RTR0MemObjMapKernelEx(&hMapObj, hMemObj, (void *)-1, 0,
2880	RTMEM_PROT_READ \| RTMEM_PROT_WRITE, offSub, cbSub);
2881	if (RT_SUCCESS(rc))
2882	*ppvR0 = RTR0MemObjAddress(hMapObj);
2883	}
2884	else
2885	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
2886
2887	}
2888	return rc;
2889	}
2890
2891
2892	/**
2893	* Changes the page level protection of one or more pages previously allocated
2894	* by SUPR0PageAllocEx.
2895	*
2896	* @returns IPRT status code.
2897	* @param pSession The session to associated the allocation with.
2898	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
2899	* NIL_RTR3PTR if the ring-3 mapping should be unaffected.
2900	* @param pvR0 The ring-0 address returned by SUPR0PageAllocEx.
2901	* NIL_RTR0PTR if the ring-0 mapping should be unaffected.
2902	* @param offSub Where to start changing. Must be page aligned.
2903	* @param cbSub How much to change. Must be page aligned.
2904	* @param fProt The new page level protection, see RTMEM_PROT_*.
2905	*/
2906	SUPR0DECL(int) SUPR0PageProtect(PSUPDRVSESSION pSession, RTR3PTR pvR3, RTR0PTR pvR0, uint32_t offSub, uint32_t cbSub, uint32_t fProt)
2907	{
2908	int rc;
2909	PSUPDRVBUNDLE pBundle;
2910	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2911	RTR0MEMOBJ hMemObjR0 = NIL_RTR0MEMOBJ;
2912	RTR0MEMOBJ hMemObjR3 = NIL_RTR0MEMOBJ;
2913	LogFlow(("SUPR0PageProtect: pSession=%p pvR3=%p pvR0=%p offSub=%#x cbSub=%#x fProt-%#x\n", pSession, pvR3, pvR0, offSub, cbSub, fProt));
2914
2915	/*
2916	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2917	*/
2918	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2919	AssertReturn(!(fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)), VERR_INVALID_PARAMETER);
2920	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2921	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2922	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
2923
2924	/*
2925	* Find the memory object.
2926	*/
2927	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2928	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2929	{
2930	if (pBundle->cUsed > 0)
2931	{
2932	unsigned i;
2933	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2934	{
2935	if ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
2936	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2937	&& ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2938	\|\| pvR3 == NIL_RTR3PTR)
2939	&& ( pvR0 == NIL_RTR0PTR
2940	\|\| RTR0MemObjAddress(pBundle->aMem[i].MemObj) == pvR0)
2941	&& ( pvR3 == NIL_RTR3PTR
2942	\|\| RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3))
2943	{
2944	if (pvR0 != NIL_RTR0PTR)
2945	hMemObjR0 = pBundle->aMem[i].MemObj;
2946	if (pvR3 != NIL_RTR3PTR)
2947	hMemObjR3 = pBundle->aMem[i].MapObjR3;
2948	break;
2949	}
2950	}
2951	}
2952	}
2953	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2954
2955	rc = VERR_INVALID_PARAMETER;
2956	if ( hMemObjR0 != NIL_RTR0MEMOBJ
2957	\|\| hMemObjR3 != NIL_RTR0MEMOBJ)
2958	{
2959	/*
2960	* Do some further input validations before calling IPRT.
2961	*/
2962	size_t cbMemObj = hMemObjR0 != NIL_RTR0PTR ? RTR0MemObjSize(hMemObjR0) : RTR0MemObjSize(hMemObjR3);
2963	if ( offSub < cbMemObj
2964	&& cbSub <= cbMemObj
2965	&& offSub + cbSub <= cbMemObj)
2966	{
2967	rc = VINF_SUCCESS;
2968	if (hMemObjR3 != NIL_RTR0PTR)
2969	rc = RTR0MemObjProtect(hMemObjR3, offSub, cbSub, fProt);
2970	if (hMemObjR0 != NIL_RTR0PTR && RT_SUCCESS(rc))
2971	rc = RTR0MemObjProtect(hMemObjR0, offSub, cbSub, fProt);
2972	}
2973	else
2974	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
2975
2976	}
2977	return rc;
2978
2979	}
2980
2981
2982	/**
2983	* Free memory allocated by SUPR0PageAlloc() and SUPR0PageAllocEx().
2984	*
2985	* @returns IPRT status code.
2986	* @param pSession The session owning the allocation.
2987	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc() or
2988	* SUPR0PageAllocEx().
2989	*/
2990	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2991	{
2992	LogFlow(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2993	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2994	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_PAGE);
2995	}
2996
2997
2998	/**
2999	* Gets the paging mode of the current CPU.
3000	*
3001	* @returns Paging mode, SUPPAGEINGMODE_INVALID on error.
3002	*/
3003	SUPR0DECL(SUPPAGINGMODE) SUPR0GetPagingMode(void)
3004	{
3005	SUPPAGINGMODE enmMode;
3006
3007	RTR0UINTREG cr0 = ASMGetCR0();
3008	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
3009	enmMode = SUPPAGINGMODE_INVALID;
3010	else
3011	{
3012	RTR0UINTREG cr4 = ASMGetCR4();
3013	uint32_t fNXEPlusLMA = 0;
3014	if (cr4 & X86_CR4_PAE)
3015	{
3016	uint32_t fAmdFeatures = ASMCpuId_EDX(0x80000001);
3017	if (fAmdFeatures & (X86_CPUID_AMD_FEATURE_EDX_NX \| X86_CPUID_AMD_FEATURE_EDX_LONG_MODE))
3018	{
3019	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
3020	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
3021	fNXEPlusLMA \|= RT_BIT(0);
3022	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
3023	fNXEPlusLMA \|= RT_BIT(1);
3024	}
3025	}
3026
3027	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
3028	{
3029	case 0:
3030	enmMode = SUPPAGINGMODE_32_BIT;
3031	break;
3032
3033	case X86_CR4_PGE:
3034	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
3035	break;
3036
3037	case X86_CR4_PAE:
3038	enmMode = SUPPAGINGMODE_PAE;
3039	break;
3040
3041	case X86_CR4_PAE \| RT_BIT(0):
3042	enmMode = SUPPAGINGMODE_PAE_NX;
3043	break;
3044
3045	case X86_CR4_PAE \| X86_CR4_PGE:
3046	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3047	break;
3048
3049	case X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3050	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3051	break;
3052
3053	case RT_BIT(1) \| X86_CR4_PAE:
3054	enmMode = SUPPAGINGMODE_AMD64;
3055	break;
3056
3057	case RT_BIT(1) \| X86_CR4_PAE \| RT_BIT(0):
3058	enmMode = SUPPAGINGMODE_AMD64_NX;
3059	break;
3060
3061	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
3062	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
3063	break;
3064
3065	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3066	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
3067	break;
3068
3069	default:
3070	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
3071	enmMode = SUPPAGINGMODE_INVALID;
3072	break;
3073	}
3074	}
3075	return enmMode;
3076	}
3077
3078
3079	/**
3080	* Enables or disabled hardware virtualization extensions using native OS APIs.
3081	*
3082	* @returns VBox status code.
3083	* @retval VINF_SUCCESS on success.
3084	* @retval VERR_NOT_SUPPORTED if not supported by the native OS.
3085	*
3086	* @param fEnable Whether to enable or disable.
3087	*/
3088	SUPR0DECL(int) SUPR0EnableVTx(bool fEnable)
3089	{
3090	#ifdef RT_OS_DARWIN
3091	return supdrvOSEnableVTx(fEnable);
3092	#else
3093	return VERR_NOT_SUPPORTED;
3094	#endif
3095	}
3096
3097
3098	/** @todo document me */
3099	SUPR0DECL(int) SUPR0QueryVTCaps(PSUPDRVSESSION pSession, uint32_t *pfCaps)
3100	{
3101	/*
3102	* Input validation.
3103	*/
3104	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3105	AssertPtrReturn(pfCaps, VERR_INVALID_POINTER);
3106
3107	*pfCaps = 0;
3108
3109	if (ASMHasCpuId())
3110	{
3111	uint32_t u32FeaturesECX;
3112	uint32_t u32Dummy;
3113	uint32_t u32FeaturesEDX;
3114	uint32_t u32VendorEBX, u32VendorECX, u32VendorEDX, u32AMDFeatureEDX, u32AMDFeatureECX;
3115	uint64_t val;
3116
3117	ASMCpuId(0, &u32Dummy, &u32VendorEBX, &u32VendorECX, &u32VendorEDX);
3118	ASMCpuId(1, &u32Dummy, &u32Dummy, &u32FeaturesECX, &u32FeaturesEDX);
3119	/* Query AMD features. */
3120	ASMCpuId(0x80000001, &u32Dummy, &u32Dummy, &u32AMDFeatureECX, &u32AMDFeatureEDX);
3121
3122	if ( u32VendorEBX == X86_CPUID_VENDOR_INTEL_EBX
3123	&& u32VendorECX == X86_CPUID_VENDOR_INTEL_ECX
3124	&& u32VendorEDX == X86_CPUID_VENDOR_INTEL_EDX
3125	)
3126	{
3127	if ( (u32FeaturesECX & X86_CPUID_FEATURE_ECX_VMX)
3128	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3129	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3130	)
3131	{
3132	val = ASMRdMsr(MSR_IA32_FEATURE_CONTROL);
3133	/*
3134	* Both the LOCK and VMXON bit must be set; otherwise VMXON will generate a #GP.
3135	* Once the lock bit is set, this MSR can no longer be modified.
3136	*/
3137	if ( (val & (MSR_IA32_FEATURE_CONTROL_VMXON\|MSR_IA32_FEATURE_CONTROL_LOCK))
3138	== (MSR_IA32_FEATURE_CONTROL_VMXON\|MSR_IA32_FEATURE_CONTROL_LOCK) /* enabled and locked */
3139	\|\| !(val & MSR_IA32_FEATURE_CONTROL_LOCK) /* not enabled, but not locked either */
3140	)
3141	{
3142	VMX_CAPABILITY vtCaps;
3143
3144	*pfCaps \|= SUPVTCAPS_VT_X;
3145
3146	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS);
3147	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC_USE_SECONDARY_EXEC_CTRL)
3148	{
3149	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS2);
3150	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC2_EPT)
3151	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3152	}
3153	return VINF_SUCCESS;
3154	}
3155	return VERR_VMX_MSR_LOCKED_OR_DISABLED;
3156	}
3157	return VERR_VMX_NO_VMX;
3158	}
3159
3160	if ( u32VendorEBX == X86_CPUID_VENDOR_AMD_EBX
3161	&& u32VendorECX == X86_CPUID_VENDOR_AMD_ECX
3162	&& u32VendorEDX == X86_CPUID_VENDOR_AMD_EDX
3163	)
3164	{
3165	if ( (u32AMDFeatureECX & X86_CPUID_AMD_FEATURE_ECX_SVM)
3166	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3167	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3168	)
3169	{
3170	/* Check if SVM is disabled */
3171	val = ASMRdMsr(MSR_K8_VM_CR);
3172	if (!(val & MSR_K8_VM_CR_SVM_DISABLE))
3173	{
3174	*pfCaps \|= SUPVTCAPS_AMD_V;
3175
3176	/* Query AMD features. */
3177	ASMCpuId(0x8000000A, &u32Dummy, &u32Dummy, &u32Dummy, &u32FeaturesEDX);
3178
3179	if (u32FeaturesEDX & AMD_CPUID_SVM_FEATURE_EDX_NESTED_PAGING)
3180	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3181
3182	return VINF_SUCCESS;
3183	}
3184	return VERR_SVM_DISABLED;
3185	}
3186	return VERR_SVM_NO_SVM;
3187	}
3188	}
3189
3190	return VERR_UNSUPPORTED_CPU;
3191	}
3192
3193
3194	/**
3195	* (Re-)initializes the per-cpu structure prior to starting or resuming the GIP
3196	* updating.
3197	*
3198	* @param pGipCpu The per CPU structure for this CPU.
3199	* @param u64NanoTS The current time.
3200	*/
3201	static void supdrvGipReInitCpu(PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
3202	{
3203	pGipCpu->u64TSC = ASMReadTSC() - pGipCpu->u32UpdateIntervalTSC;
3204	pGipCpu->u64NanoTS = u64NanoTS;
3205	}
3206
3207
3208	/**
3209	* Set the current TSC and NanoTS value for the CPU.
3210	*
3211	* @param idCpu The CPU ID. Unused - we have to use the APIC ID.
3212	* @param pvUser1 Pointer to the ring-0 GIP mapping.
3213	* @param pvUser2 Pointer to the variable holding the current time.
3214	*/
3215	static DECLCALLBACK(void) supdrvGipReInitCpuCallback(RTCPUID idCpu, void pvUser1, void pvUser2)
3216	{
3217	PSUPGLOBALINFOPAGE pGip = (PSUPGLOBALINFOPAGE)pvUser1;
3218	#ifdef SUP_WITH_LOTS_OF_CPUS
3219	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
3220	#else
3221	unsigned iCpu = ASMGetApicId();
3222	#endif
3223
3224	#ifdef SUP_WITH_LOTS_OF_CPUS
3225	if (RT_LIKELY(iCpu < pGip->cCpus && pGip->aCPUs[iCpu].idCpu == idCpu))
3226	#else
3227	if (RT_LIKELY(iCpu < RT_ELEMENTS(pGip->aCPUs)))
3228	#endif
3229	supdrvGipReInitCpu(&pGip->aCPUs[iCpu], (uint64_t )pvUser2);
3230
3231	NOREF(pvUser2);
3232	NOREF(idCpu);
3233	}
3234
3235
3236	/**
3237	* Maps the GIP into userspace and/or get the physical address of the GIP.
3238	*
3239	* @returns IPRT status code.
3240	* @param pSession Session to which the GIP mapping should belong.
3241	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
3242	* @param pHCPhysGip Where to store the physical address. (optional)
3243	*
3244	* @remark There is no reference counting on the mapping, so one call to this function
3245	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
3246	* and remove the session as a GIP user.
3247	*/
3248	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
3249	{
3250	int rc;
3251	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3252	RTR3PTR pGipR3 = NIL_RTR3PTR;
3253	RTHCPHYS HCPhys = NIL_RTHCPHYS;
3254	LogFlow(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
3255
3256	/*
3257	* Validate
3258	*/
3259	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3260	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
3261	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
3262
3263	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3264	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3265	#else
3266	RTSemFastMutexRequest(pDevExt->mtxGip);
3267	#endif
3268	if (pDevExt->pGip)
3269	{
3270	/*
3271	* Map it?
3272	*/
3273	rc = VINF_SUCCESS;
3274	if (ppGipR3)
3275	{
3276	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
3277	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
3278	RTMEM_PROT_READ, RTR0ProcHandleSelf());
3279	if (RT_SUCCESS(rc))
3280	pGipR3 = RTR0MemObjAddressR3(pSession->GipMapObjR3);
3281	}
3282
3283	/*
3284	* Get physical address.
3285	*/
3286	if (pHCPhysGip && RT_SUCCESS(rc))
3287	HCPhys = pDevExt->HCPhysGip;
3288
3289	/*
3290	* Reference globally.
3291	*/
3292	if (!pSession->fGipReferenced && RT_SUCCESS(rc))
3293	{
3294	pSession->fGipReferenced = 1;
3295	pDevExt->cGipUsers++;
3296	if (pDevExt->cGipUsers == 1)
3297	{
3298	PSUPGLOBALINFOPAGE pGipR0 = pDevExt->pGip;
3299	uint64_t u64NanoTS;
3300	uint32_t u32SystemResolution;
3301	unsigned i;
3302
3303	LogFlow(("SUPR0GipMap: Resumes GIP updating\n"));
3304
3305	/*
3306	* Try bump up the system timer resolution.
3307	* The more interrupts the better...
3308	*/
3309	if ( RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
3310	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
3311	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1953125 /* 512 HZ */, &u32SystemResolution))
3312	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 2000000 /* 500 HZ */, &u32SystemResolution))
3313	)
3314	{
3315	Assert(RTTimerGetSystemGranularity() <= u32SystemResolution);
3316	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
3317	}
3318
3319	if (pGipR0->aCPUs[0].u32TransactionId != 2 /* not the first time */)
3320	{
3321	for (i = 0; i < RT_ELEMENTS(pGipR0->aCPUs); i++)
3322	ASMAtomicUoWriteU32(&pGipR0->aCPUs[i].u32TransactionId,
3323	(pGipR0->aCPUs[i].u32TransactionId + GIP_UPDATEHZ_RECALC_FREQ * 2)
3324	& ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
3325	ASMAtomicWriteU64(&pGipR0->u64NanoTSLastUpdateHz, 0);
3326	}
3327
3328	u64NanoTS = RTTimeSystemNanoTS() - pGipR0->u32UpdateIntervalNS;
3329	if ( pGipR0->u32Mode == SUPGIPMODE_SYNC_TSC
3330	\|\| RTMpGetOnlineCount() == 1)
3331	supdrvGipReInitCpu(&pGipR0->aCPUs[0], u64NanoTS);
3332	else
3333	RTMpOnAll(supdrvGipReInitCpuCallback, pGipR0, &u64NanoTS);
3334
3335	#ifndef DO_NOT_START_GIP
3336	rc = RTTimerStart(pDevExt->pGipTimer, 0); AssertRC(rc);
3337	#endif
3338	rc = VINF_SUCCESS;
3339	}
3340	}
3341	}
3342	else
3343	{
3344	rc = VERR_GENERAL_FAILURE;
3345	Log(("SUPR0GipMap: GIP is not available!\n"));
3346	}
3347	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3348	RTSemMutexRelease(pDevExt->mtxGip);
3349	#else
3350	RTSemFastMutexRelease(pDevExt->mtxGip);
3351	#endif
3352
3353	/*
3354	* Write returns.
3355	*/
3356	if (pHCPhysGip)
3357	*pHCPhysGip = HCPhys;
3358	if (ppGipR3)
3359	*ppGipR3 = pGipR3;
3360
3361	#ifdef DEBUG_DARWIN_GIP
3362	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3363	#else
3364	LogFlow(( "SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3365	#endif
3366	return rc;
3367	}
3368
3369
3370	/**
3371	* Unmaps any user mapping of the GIP and terminates all GIP access
3372	* from this session.
3373	*
3374	* @returns IPRT status code.
3375	* @param pSession Session to which the GIP mapping should belong.
3376	*/
3377	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
3378	{
3379	int rc = VINF_SUCCESS;
3380	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3381	#ifdef DEBUG_DARWIN_GIP
3382	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
3383	pSession,
3384	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
3385	pSession->GipMapObjR3));
3386	#else
3387	LogFlow(("SUPR0GipUnmap: pSession=%p\n", pSession));
3388	#endif
3389	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3390
3391	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3392	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3393	#else
3394	RTSemFastMutexRequest(pDevExt->mtxGip);
3395	#endif
3396
3397	/*
3398	* Unmap anything?
3399	*/
3400	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
3401	{
3402	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
3403	AssertRC(rc);
3404	if (RT_SUCCESS(rc))
3405	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
3406	}
3407
3408	/*
3409	* Dereference global GIP.
3410	*/
3411	if (pSession->fGipReferenced && !rc)
3412	{
3413	pSession->fGipReferenced = 0;
3414	if ( pDevExt->cGipUsers > 0
3415	&& !--pDevExt->cGipUsers)
3416	{
3417	LogFlow(("SUPR0GipUnmap: Suspends GIP updating\n"));
3418	#ifndef DO_NOT_START_GIP
3419	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = VINF_SUCCESS;
3420	#endif
3421
3422	if (pDevExt->u32SystemTimerGranularityGrant)
3423	{
3424	int rc2 = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant);
3425	AssertRC(rc2);
3426	pDevExt->u32SystemTimerGranularityGrant = 0;
3427	}
3428	}
3429	}
3430
3431	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3432	RTSemMutexRelease(pDevExt->mtxGip);
3433	#else
3434	RTSemFastMutexRelease(pDevExt->mtxGip);
3435	#endif
3436
3437	return rc;
3438	}
3439
3440
3441	/**
3442	* Gets the GIP pointer.
3443	*
3444	* @returns Pointer to the GIP or NULL.
3445	*/
3446	SUPDECL(PSUPGLOBALINFOPAGE) SUPGetGIP(void)
3447	{
3448	return g_pSUPGlobalInfoPage;
3449	}
3450
3451
3452	/**
3453	* Register a component factory with the support driver.
3454	*
3455	* This is currently restricted to kernel sessions only.
3456	*
3457	* @returns VBox status code.
3458	* @retval VINF_SUCCESS on success.
3459	* @retval VERR_NO_MEMORY if we're out of memory.
3460	* @retval VERR_ALREADY_EXISTS if the factory has already been registered.
3461	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3462	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3463	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3464	*
3465	* @param pSession The SUPDRV session (must be a ring-0 session).
3466	* @param pFactory Pointer to the component factory registration structure.
3467	*
3468	* @remarks This interface is also available via SUPR0IdcComponentRegisterFactory.
3469	*/
3470	SUPR0DECL(int) SUPR0ComponentRegisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3471	{
3472	PSUPDRVFACTORYREG pNewReg;
3473	const char *psz;
3474	int rc;
3475
3476	/*
3477	* Validate parameters.
3478	*/
3479	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3480	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3481	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3482	AssertPtrReturn(pFactory->pfnQueryFactoryInterface, VERR_INVALID_POINTER);
3483	psz = RTStrEnd(pFactory->szName, sizeof(pFactory->szName));
3484	AssertReturn(psz, VERR_INVALID_PARAMETER);
3485
3486	/*
3487	* Allocate and initialize a new registration structure.
3488	*/
3489	pNewReg = (PSUPDRVFACTORYREG)RTMemAlloc(sizeof(SUPDRVFACTORYREG));
3490	if (pNewReg)
3491	{
3492	pNewReg->pNext = NULL;
3493	pNewReg->pFactory = pFactory;
3494	pNewReg->pSession = pSession;
3495	pNewReg->cchName = psz - &pFactory->szName[0];
3496
3497	/*
3498	* Add it to the tail of the list after checking for prior registration.
3499	*/
3500	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3501	if (RT_SUCCESS(rc))
3502	{
3503	PSUPDRVFACTORYREG pPrev = NULL;
3504	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3505	while (pCur && pCur->pFactory != pFactory)
3506	{
3507	pPrev = pCur;
3508	pCur = pCur->pNext;
3509	}
3510	if (!pCur)
3511	{
3512	if (pPrev)
3513	pPrev->pNext = pNewReg;
3514	else
3515	pSession->pDevExt->pComponentFactoryHead = pNewReg;
3516	rc = VINF_SUCCESS;
3517	}
3518	else
3519	rc = VERR_ALREADY_EXISTS;
3520
3521	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3522	}
3523
3524	if (RT_FAILURE(rc))
3525	RTMemFree(pNewReg);
3526	}
3527	else
3528	rc = VERR_NO_MEMORY;
3529	return rc;
3530	}
3531
3532
3533	/**
3534	* Deregister a component factory.
3535	*
3536	* @returns VBox status code.
3537	* @retval VINF_SUCCESS on success.
3538	* @retval VERR_NOT_FOUND if the factory wasn't registered.
3539	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3540	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3541	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3542	*
3543	* @param pSession The SUPDRV session (must be a ring-0 session).
3544	* @param pFactory Pointer to the component factory registration structure
3545	* previously passed SUPR0ComponentRegisterFactory().
3546	*
3547	* @remarks This interface is also available via SUPR0IdcComponentDeregisterFactory.
3548	*/
3549	SUPR0DECL(int) SUPR0ComponentDeregisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3550	{
3551	int rc;
3552
3553	/*
3554	* Validate parameters.
3555	*/
3556	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3557	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3558	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3559
3560	/*
3561	* Take the lock and look for the registration record.
3562	*/
3563	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3564	if (RT_SUCCESS(rc))
3565	{
3566	PSUPDRVFACTORYREG pPrev = NULL;
3567	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3568	while (pCur && pCur->pFactory != pFactory)
3569	{
3570	pPrev = pCur;
3571	pCur = pCur->pNext;
3572	}
3573	if (pCur)
3574	{
3575	if (!pPrev)
3576	pSession->pDevExt->pComponentFactoryHead = pCur->pNext;
3577	else
3578	pPrev->pNext = pCur->pNext;
3579
3580	pCur->pNext = NULL;
3581	pCur->pFactory = NULL;
3582	pCur->pSession = NULL;
3583	rc = VINF_SUCCESS;
3584	}
3585	else
3586	rc = VERR_NOT_FOUND;
3587
3588	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3589
3590	RTMemFree(pCur);
3591	}
3592	return rc;
3593	}
3594
3595
3596	/**
3597	* Queries a component factory.
3598	*
3599	* @returns VBox status code.
3600	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3601	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3602	* @retval VERR_SUPDRV_COMPONENT_NOT_FOUND if the component factory wasn't found.
3603	* @retval VERR_SUPDRV_INTERFACE_NOT_SUPPORTED if the interface wasn't supported.
3604	*
3605	* @param pSession The SUPDRV session.
3606	* @param pszName The name of the component factory.
3607	* @param pszInterfaceUuid The UUID of the factory interface (stringified).
3608	* @param ppvFactoryIf Where to store the factory interface.
3609	*/
3610	SUPR0DECL(int) SUPR0ComponentQueryFactory(PSUPDRVSESSION pSession, const char pszName, const char pszInterfaceUuid, void **ppvFactoryIf)
3611	{
3612	const char *pszEnd;
3613	size_t cchName;
3614	int rc;
3615
3616	/*
3617	* Validate parameters.
3618	*/
3619	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3620
3621	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
3622	pszEnd = RTStrEnd(pszName, RT_SIZEOFMEMB(SUPDRVFACTORY, szName));
3623	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3624	cchName = pszEnd - pszName;
3625
3626	AssertPtrReturn(pszInterfaceUuid, VERR_INVALID_POINTER);
3627	pszEnd = RTStrEnd(pszInterfaceUuid, RTUUID_STR_LENGTH);
3628	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3629
3630	AssertPtrReturn(ppvFactoryIf, VERR_INVALID_POINTER);
3631	*ppvFactoryIf = NULL;
3632
3633	/*
3634	* Take the lock and try all factories by this name.
3635	*/
3636	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3637	if (RT_SUCCESS(rc))
3638	{
3639	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3640	rc = VERR_SUPDRV_COMPONENT_NOT_FOUND;
3641	while (pCur)
3642	{
3643	if ( pCur->cchName == cchName
3644	&& !memcmp(pCur->pFactory->szName, pszName, cchName))
3645	{
3646	void *pvFactory = pCur->pFactory->pfnQueryFactoryInterface(pCur->pFactory, pSession, pszInterfaceUuid);
3647	if (pvFactory)
3648	{
3649	*ppvFactoryIf = pvFactory;
3650	rc = VINF_SUCCESS;
3651	break;
3652	}
3653	rc = VERR_SUPDRV_INTERFACE_NOT_SUPPORTED;
3654	}
3655
3656	/* next */
3657	pCur = pCur->pNext;
3658	}
3659
3660	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3661	}
3662	return rc;
3663	}
3664
3665
3666	/**
3667	* Adds a memory object to the session.
3668	*
3669	* @returns IPRT status code.
3670	* @param pMem Memory tracking structure containing the
3671	* information to track.
3672	* @param pSession The session.
3673	*/
3674	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
3675	{
3676	PSUPDRVBUNDLE pBundle;
3677	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3678
3679	/*
3680	* Find free entry and record the allocation.
3681	*/
3682	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3683	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3684	{
3685	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
3686	{
3687	unsigned i;
3688	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3689	{
3690	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
3691	{
3692	pBundle->cUsed++;
3693	pBundle->aMem[i] = *pMem;
3694	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3695	return VINF_SUCCESS;
3696	}
3697	}
3698	AssertFailed(); /* !!this can't be happening!!! */
3699	}
3700	}
3701	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3702
3703	/*
3704	* Need to allocate a new bundle.
3705	* Insert into the last entry in the bundle.
3706	*/
3707	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
3708	if (!pBundle)
3709	return VERR_NO_MEMORY;
3710
3711	/* take last entry. */
3712	pBundle->cUsed++;
3713	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
3714
3715	/* insert into list. */
3716	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3717	pBundle->pNext = pSession->Bundle.pNext;
3718	pSession->Bundle.pNext = pBundle;
3719	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3720
3721	return VINF_SUCCESS;
3722	}
3723
3724
3725	/**
3726	* Releases a memory object referenced by pointer and type.
3727	*
3728	* @returns IPRT status code.
3729	* @param pSession Session data.
3730	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
3731	* @param eType Memory type.
3732	*/
3733	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
3734	{
3735	PSUPDRVBUNDLE pBundle;
3736	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3737
3738	/*
3739	* Validate input.
3740	*/
3741	if (!uPtr)
3742	{
3743	Log(("Illegal address %p\n", (void *)uPtr));
3744	return VERR_INVALID_PARAMETER;
3745	}
3746
3747	/*
3748	* Search for the address.
3749	*/
3750	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3751	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3752	{
3753	if (pBundle->cUsed > 0)
3754	{
3755	unsigned i;
3756	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3757	{
3758	if ( pBundle->aMem[i].eType == eType
3759	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3760	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
3761	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3762	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
3763	)
3764	{
3765	/* Make a copy of it and release it outside the spinlock. */
3766	SUPDRVMEMREF Mem = pBundle->aMem[i];
3767	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
3768	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
3769	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
3770	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3771
3772	if (Mem.MapObjR3 != NIL_RTR0MEMOBJ)
3773	{
3774	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
3775	AssertRC(rc); /** @todo figure out how to handle this. */
3776	}
3777	if (Mem.MemObj != NIL_RTR0MEMOBJ)
3778	{
3779	int rc = RTR0MemObjFree(Mem.MemObj, true /* fFreeMappings */);
3780	AssertRC(rc); /** @todo figure out how to handle this. */
3781	}
3782	return VINF_SUCCESS;
3783	}
3784	}
3785	}
3786	}
3787	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3788	Log(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
3789	return VERR_INVALID_PARAMETER;
3790	}
3791
3792
3793	/**
3794	* Opens an image. If it's the first time it's opened the call must upload
3795	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
3796	*
3797	* This is the 1st step of the loading.
3798	*
3799	* @returns IPRT status code.
3800	* @param pDevExt Device globals.
3801	* @param pSession Session data.
3802	* @param pReq The open request.
3803	*/
3804	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
3805	{
3806	int rc;
3807	PSUPDRVLDRIMAGE pImage;
3808	void *pv;
3809	size_t cchName = strlen(pReq->u.In.szName); /* (caller checked < 32). */
3810	LogFlow(("supdrvIOCtl_LdrOpen: szName=%s cbImageWithTabs=%d\n", pReq->u.In.szName, pReq->u.In.cbImageWithTabs));
3811
3812	/*
3813	* Check if we got an instance of the image already.
3814	*/
3815	supdrvLdrLock(pDevExt);
3816	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
3817	{
3818	if ( pImage->szName[cchName] == '\0'
3819	&& !memcmp(pImage->szName, pReq->u.In.szName, cchName))
3820	{
3821	/** @todo check cbImageBits and cbImageWithTabs here, if they differs that indicates that the images are different. */
3822	pImage->cUsage++;
3823	pReq->u.Out.pvImageBase = pImage->pvImage;
3824	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
3825	pReq->u.Out.fNativeLoader = pImage->fNative;
3826	supdrvLdrAddUsage(pSession, pImage);
3827	supdrvLdrUnlock(pDevExt);
3828	return VINF_SUCCESS;
3829	}
3830	}
3831	/* (not found - add it!) */
3832
3833	/*
3834	* Allocate memory.
3835	*/
3836	pv = RTMemAlloc(RT_OFFSETOF(SUPDRVLDRIMAGE, szName[cchName + 1]));
3837	if (!pv)
3838	{
3839	supdrvLdrUnlock(pDevExt);
3840	Log(("supdrvIOCtl_LdrOpen: RTMemAlloc() failed\n"));
3841	return /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_2;
3842	}
3843
3844	/*
3845	* Setup and link in the LDR stuff.
3846	*/
3847	pImage = (PSUPDRVLDRIMAGE)pv;
3848	pImage->pvImage = NULL;
3849	pImage->pvImageAlloc = NULL;
3850	pImage->cbImageWithTabs = pReq->u.In.cbImageWithTabs;
3851	pImage->cbImageBits = pReq->u.In.cbImageBits;
3852	pImage->cSymbols = 0;
3853	pImage->paSymbols = NULL;
3854	pImage->pachStrTab = NULL;
3855	pImage->cbStrTab = 0;
3856	pImage->pfnModuleInit = NULL;
3857	pImage->pfnModuleTerm = NULL;
3858	pImage->pfnServiceReqHandler = NULL;
3859	pImage->uState = SUP_IOCTL_LDR_OPEN;
3860	pImage->cUsage = 1;
3861	memcpy(pImage->szName, pReq->u.In.szName, cchName + 1);
3862
3863	/*
3864	* Try load it using the native loader, if that isn't supported, fall back
3865	* on the older method.
3866	*/
3867	pImage->fNative = true;
3868	rc = supdrvOSLdrOpen(pDevExt, pImage, pReq->u.In.szFilename);
3869	if (rc == VERR_NOT_SUPPORTED)
3870	{
3871	pImage->pvImageAlloc = RTMemExecAlloc(pImage->cbImageBits + 31);
3872	pImage->pvImage = RT_ALIGN_P(pImage->pvImageAlloc, 32);
3873	pImage->fNative = false;
3874	rc = pImage->pvImageAlloc ? VINF_SUCCESS : VERR_NO_EXEC_MEMORY;
3875	}
3876	if (RT_FAILURE(rc))
3877	{
3878	supdrvLdrUnlock(pDevExt);
3879	RTMemFree(pImage);
3880	Log(("supdrvIOCtl_LdrOpen(%s): failed - %Rrc\n", pReq->u.In.szName, rc));
3881	return rc;
3882	}
3883	Assert(VALID_PTR(pImage->pvImage) \|\| RT_FAILURE(rc));
3884
3885	/*
3886	* Link it.
3887	*/
3888	pImage->pNext = pDevExt->pLdrImages;
3889	pDevExt->pLdrImages = pImage;
3890
3891	supdrvLdrAddUsage(pSession, pImage);
3892
3893	pReq->u.Out.pvImageBase = pImage->pvImage;
3894	pReq->u.Out.fNeedsLoading = true;
3895	pReq->u.Out.fNativeLoader = pImage->fNative;
3896	supdrvLdrUnlock(pDevExt);
3897
3898	#if defined(RT_OS_WINDOWS) && defined(DEBUG)
3899	SUPR0Printf("VBoxDrv: windbg> .reload /f %s=%#p\n", pImage->szName, pImage->pvImage);
3900	#endif
3901	return VINF_SUCCESS;
3902	}
3903
3904
3905	/**
3906	* Worker that validates a pointer to an image entrypoint.
3907	*
3908	* @returns IPRT status code.
3909	* @param pDevExt The device globals.
3910	* @param pImage The loader image.
3911	* @param pv The pointer into the image.
3912	* @param fMayBeNull Whether it may be NULL.
3913	* @param pszWhat What is this entrypoint? (for logging)
3914	* @param pbImageBits The image bits prepared by ring-3.
3915	*
3916	* @remarks Will leave the lock on failure.
3917	*/
3918	static int supdrvLdrValidatePointer(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage, void *pv,
3919	bool fMayBeNull, const uint8_t pbImageBits, const char pszWhat)
3920	{
3921	if (!fMayBeNull \|\| pv)
3922	{
3923	if ((uintptr_t)pv - (uintptr_t)pImage->pvImage >= pImage->cbImageBits)
3924	{
3925	supdrvLdrUnlock(pDevExt);
3926	Log(("Out of range (%p LB %#x): %s=%p\n", pImage->pvImage, pImage->cbImageBits, pszWhat, pv));
3927	return VERR_INVALID_PARAMETER;
3928	}
3929
3930	if (pImage->fNative)
3931	{
3932	int rc = supdrvOSLdrValidatePointer(pDevExt, pImage, pv, pbImageBits);
3933	if (RT_FAILURE(rc))
3934	{
3935	supdrvLdrUnlock(pDevExt);
3936	Log(("Bad entry point address: %s=%p (rc=%Rrc)\n", pszWhat, pv, rc));
3937	return rc;
3938	}
3939	}
3940	}
3941	return VINF_SUCCESS;
3942	}
3943
3944
3945	/**
3946	* Loads the image bits.
3947	*
3948	* This is the 2nd step of the loading.
3949	*
3950	* @returns IPRT status code.
3951	* @param pDevExt Device globals.
3952	* @param pSession Session data.
3953	* @param pReq The request.
3954	*/
3955	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
3956	{
3957	PSUPDRVLDRUSAGE pUsage;
3958	PSUPDRVLDRIMAGE pImage;
3959	int rc;
3960	LogFlow(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImageWithBits=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImageWithTabs));
3961
3962	/*
3963	* Find the ldr image.
3964	*/
3965	supdrvLdrLock(pDevExt);
3966	pUsage = pSession->pLdrUsage;
3967	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3968	pUsage = pUsage->pNext;
3969	if (!pUsage)
3970	{
3971	supdrvLdrUnlock(pDevExt);
3972	Log(("SUP_IOCTL_LDR_LOAD: couldn't find image!\n"));
3973	return VERR_INVALID_HANDLE;
3974	}
3975	pImage = pUsage->pImage;
3976
3977	/*
3978	* Validate input.
3979	*/
3980	if ( pImage->cbImageWithTabs != pReq->u.In.cbImageWithTabs
3981	\|\| pImage->cbImageBits != pReq->u.In.cbImageBits)
3982	{
3983	supdrvLdrUnlock(pDevExt);
3984	Log(("SUP_IOCTL_LDR_LOAD: image size mismatch!! %d(prep) != %d(load) or %d != %d\n",
3985	pImage->cbImageWithTabs, pReq->u.In.cbImageWithTabs, pImage->cbImageBits, pReq->u.In.cbImageBits));
3986	return VERR_INVALID_HANDLE;
3987	}
3988
3989	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
3990	{
3991	unsigned uState = pImage->uState;
3992	supdrvLdrUnlock(pDevExt);
3993	if (uState != SUP_IOCTL_LDR_LOAD)
3994	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
3995	return VERR_ALREADY_LOADED;
3996	}
3997
3998	switch (pReq->u.In.eEPType)
3999	{
4000	case SUPLDRLOADEP_NOTHING:
4001	break;
4002
4003	case SUPLDRLOADEP_VMMR0:
4004	rc = supdrvLdrValidatePointer( pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0, false, pReq->u.In.abImage, "pvVMMR0");
4005	if (RT_SUCCESS(rc))
4006	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt, false, pReq->u.In.abImage, "pvVMMR0EntryInt");
4007	if (RT_SUCCESS(rc))
4008	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, false, pReq->u.In.abImage, "pvVMMR0EntryFast");
4009	if (RT_SUCCESS(rc))
4010	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx, false, pReq->u.In.abImage, "pvVMMR0EntryEx");
4011	if (RT_FAILURE(rc))
4012	return rc;
4013	break;
4014
4015	case SUPLDRLOADEP_SERVICE:
4016	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.Service.pfnServiceReq, false, pReq->u.In.abImage, "pfnServiceReq");
4017	if (RT_FAILURE(rc))
4018	return rc;
4019	if ( pReq->u.In.EP.Service.apvReserved[0] != NIL_RTR0PTR
4020	\|\| pReq->u.In.EP.Service.apvReserved[1] != NIL_RTR0PTR
4021	\|\| pReq->u.In.EP.Service.apvReserved[2] != NIL_RTR0PTR)
4022	{
4023	supdrvLdrUnlock(pDevExt);
4024	Log(("Out of range (%p LB %#x): apvReserved={%p,%p,%p} MBZ!\n",
4025	pImage->pvImage, pReq->u.In.cbImageWithTabs,
4026	pReq->u.In.EP.Service.apvReserved[0],
4027	pReq->u.In.EP.Service.apvReserved[1],
4028	pReq->u.In.EP.Service.apvReserved[2]));
4029	return VERR_INVALID_PARAMETER;
4030	}
4031	break;
4032
4033	default:
4034	supdrvLdrUnlock(pDevExt);
4035	Log(("Invalid eEPType=%d\n", pReq->u.In.eEPType));
4036	return VERR_INVALID_PARAMETER;
4037	}
4038
4039	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleInit, true, pReq->u.In.abImage, "pfnModuleInit");
4040	if (RT_FAILURE(rc))
4041	return rc;
4042	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleTerm, true, pReq->u.In.abImage, "pfnModuleTerm");
4043	if (RT_FAILURE(rc))
4044	return rc;
4045
4046	/*
4047	* Allocate and copy the tables.
4048	* (No need to do try/except as this is a buffered request.)
4049	*/
4050	pImage->cbStrTab = pReq->u.In.cbStrTab;
4051	if (pImage->cbStrTab)
4052	{
4053	pImage->pachStrTab = (char *)RTMemAlloc(pImage->cbStrTab);
4054	if (pImage->pachStrTab)
4055	memcpy(pImage->pachStrTab, &pReq->u.In.abImage[pReq->u.In.offStrTab], pImage->cbStrTab);
4056	else
4057	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_3;
4058	}
4059
4060	pImage->cSymbols = pReq->u.In.cSymbols;
4061	if (RT_SUCCESS(rc) && pImage->cSymbols)
4062	{
4063	size_t cbSymbols = pImage->cSymbols * sizeof(SUPLDRSYM);
4064	pImage->paSymbols = (PSUPLDRSYM)RTMemAlloc(cbSymbols);
4065	if (pImage->paSymbols)
4066	memcpy(pImage->paSymbols, &pReq->u.In.abImage[pReq->u.In.offSymbols], cbSymbols);
4067	else
4068	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_4;
4069	}
4070
4071	/*
4072	* Copy the bits / complete native loading.
4073	*/
4074	if (RT_SUCCESS(rc))
4075	{
4076	pImage->uState = SUP_IOCTL_LDR_LOAD;
4077	pImage->pfnModuleInit = pReq->u.In.pfnModuleInit;
4078	pImage->pfnModuleTerm = pReq->u.In.pfnModuleTerm;
4079
4080	if (pImage->fNative)
4081	rc = supdrvOSLdrLoad(pDevExt, pImage, pReq->u.In.abImage);
4082	else
4083	memcpy(pImage->pvImage, &pReq->u.In.abImage[0], pImage->cbImageBits);
4084	}
4085
4086	/*
4087	* Update any entry points.
4088	*/
4089	if (RT_SUCCESS(rc))
4090	{
4091	switch (pReq->u.In.eEPType)
4092	{
4093	default:
4094	case SUPLDRLOADEP_NOTHING:
4095	rc = VINF_SUCCESS;
4096	break;
4097	case SUPLDRLOADEP_VMMR0:
4098	rc = supdrvLdrSetVMMR0EPs(pDevExt, pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
4099	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
4100	break;
4101	case SUPLDRLOADEP_SERVICE:
4102	pImage->pfnServiceReqHandler = pReq->u.In.EP.Service.pfnServiceReq;
4103	rc = VINF_SUCCESS;
4104	break;
4105	}
4106	}
4107
4108	/*
4109	* On success call the module initialization.
4110	*/
4111	LogFlow(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
4112	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
4113	{
4114	Log(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
4115	rc = pImage->pfnModuleInit();
4116	if (rc && pDevExt->pvVMMR0 == pImage->pvImage)
4117	supdrvLdrUnsetVMMR0EPs(pDevExt);
4118	}
4119
4120	if (RT_FAILURE(rc))
4121	{
4122	pImage->uState = SUP_IOCTL_LDR_OPEN;
4123	pImage->pfnModuleInit = NULL;
4124	pImage->pfnModuleTerm = NULL;
4125	pImage->pfnServiceReqHandler= NULL;
4126	pImage->cbStrTab = 0;
4127	RTMemFree(pImage->pachStrTab);
4128	pImage->pachStrTab = NULL;
4129	RTMemFree(pImage->paSymbols);
4130	pImage->paSymbols = NULL;
4131	pImage->cSymbols = 0;
4132	}
4133
4134	supdrvLdrUnlock(pDevExt);
4135	return rc;
4136	}
4137
4138
4139	/**
4140	* Frees a previously loaded (prep'ed) image.
4141	*
4142	* @returns IPRT status code.
4143	* @param pDevExt Device globals.
4144	* @param pSession Session data.
4145	* @param pReq The request.
4146	*/
4147	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
4148	{
4149	int rc;
4150	PSUPDRVLDRUSAGE pUsagePrev;
4151	PSUPDRVLDRUSAGE pUsage;
4152	PSUPDRVLDRIMAGE pImage;
4153	LogFlow(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
4154
4155	/*
4156	* Find the ldr image.
4157	*/
4158	supdrvLdrLock(pDevExt);
4159	pUsagePrev = NULL;
4160	pUsage = pSession->pLdrUsage;
4161	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4162	{
4163	pUsagePrev = pUsage;
4164	pUsage = pUsage->pNext;
4165	}
4166	if (!pUsage)
4167	{
4168	supdrvLdrUnlock(pDevExt);
4169	Log(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
4170	return VERR_INVALID_HANDLE;
4171	}
4172
4173	/*
4174	* Check if we can remove anything.
4175	*/
4176	rc = VINF_SUCCESS;
4177	pImage = pUsage->pImage;
4178	if (pImage->cUsage <= 1 \|\| pUsage->cUsage <= 1)
4179	{
4180	/*
4181	* Check if there are any objects with destructors in the image, if
4182	* so leave it for the session cleanup routine so we get a chance to
4183	* clean things up in the right order and not leave them all dangling.
4184	*/
4185	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
4186	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
4187	if (pImage->cUsage <= 1)
4188	{
4189	PSUPDRVOBJ pObj;
4190	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4191	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4192	{
4193	rc = VERR_DANGLING_OBJECTS;
4194	break;
4195	}
4196	}
4197	else
4198	{
4199	PSUPDRVUSAGE pGenUsage;
4200	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
4201	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4202	{
4203	rc = VERR_DANGLING_OBJECTS;
4204	break;
4205	}
4206	}
4207	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
4208	if (rc == VINF_SUCCESS)
4209	{
4210	/* unlink it */
4211	if (pUsagePrev)
4212	pUsagePrev->pNext = pUsage->pNext;
4213	else
4214	pSession->pLdrUsage = pUsage->pNext;
4215
4216	/* free it */
4217	pUsage->pImage = NULL;
4218	pUsage->pNext = NULL;
4219	RTMemFree(pUsage);
4220
4221	/*
4222	* Dereference the image.
4223	*/
4224	if (pImage->cUsage <= 1)
4225	supdrvLdrFree(pDevExt, pImage);
4226	else
4227	pImage->cUsage--;
4228	}
4229	else
4230	{
4231	Log(("supdrvIOCtl_LdrFree: Dangling objects in %p/%s!\n", pImage->pvImage, pImage->szName));
4232	rc = VINF_SUCCESS; /** @todo BRANCH-2.1: remove this after branching. */
4233	}
4234	}
4235	else
4236	{
4237	/*
4238	* Dereference both image and usage.
4239	*/
4240	pImage->cUsage--;
4241	pUsage->cUsage--;
4242	}
4243
4244	supdrvLdrUnlock(pDevExt);
4245	return rc;
4246	}
4247
4248
4249	/**
4250	* Gets the address of a symbol in an open image.
4251	*
4252	* @returns IPRT status code.
4253	* @param pDevExt Device globals.
4254	* @param pSession Session data.
4255	* @param pReq The request buffer.
4256	*/
4257	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
4258	{
4259	PSUPDRVLDRIMAGE pImage;
4260	PSUPDRVLDRUSAGE pUsage;
4261	uint32_t i;
4262	PSUPLDRSYM paSyms;
4263	const char *pchStrings;
4264	const size_t cbSymbol = strlen(pReq->u.In.szSymbol) + 1;
4265	void *pvSymbol = NULL;
4266	int rc = VERR_GENERAL_FAILURE;
4267	Log3(("supdrvIOCtl_LdrGetSymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
4268
4269	/*
4270	* Find the ldr image.
4271	*/
4272	supdrvLdrLock(pDevExt);
4273	pUsage = pSession->pLdrUsage;
4274	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4275	pUsage = pUsage->pNext;
4276	if (!pUsage)
4277	{
4278	supdrvLdrUnlock(pDevExt);
4279	Log(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
4280	return VERR_INVALID_HANDLE;
4281	}
4282	pImage = pUsage->pImage;
4283	if (pImage->uState != SUP_IOCTL_LDR_LOAD)
4284	{
4285	unsigned uState = pImage->uState;
4286	supdrvLdrUnlock(pDevExt);
4287	Log(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", uState, uState)); NOREF(uState);
4288	return VERR_ALREADY_LOADED;
4289	}
4290
4291	/*
4292	* Search the symbol strings.
4293	*/
4294	pchStrings = pImage->pachStrTab;
4295	paSyms = pImage->paSymbols;
4296	for (i = 0; i < pImage->cSymbols; i++)
4297	{
4298	if ( paSyms[i].offSymbol < pImage->cbImageBits /* paranoia */
4299	&& paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4300	&& !memcmp(pchStrings + paSyms[i].offName, pReq->u.In.szSymbol, cbSymbol))
4301	{
4302	pvSymbol = (uint8_t *)pImage->pvImage + paSyms[i].offSymbol;
4303	rc = VINF_SUCCESS;
4304	break;
4305	}
4306	}
4307	supdrvLdrUnlock(pDevExt);
4308	pReq->u.Out.pvSymbol = pvSymbol;
4309	return rc;
4310	}
4311
4312
4313	/**
4314	* Gets the address of a symbol in an open image or the support driver.
4315	*
4316	* @returns VINF_SUCCESS on success.
4317	* @returns
4318	* @param pDevExt Device globals.
4319	* @param pSession Session data.
4320	* @param pReq The request buffer.
4321	*/
4322	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq)
4323	{
4324	int rc = VINF_SUCCESS;
4325	const char *pszSymbol = pReq->u.In.pszSymbol;
4326	const char *pszModule = pReq->u.In.pszModule;
4327	size_t cbSymbol;
4328	char const *pszEnd;
4329	uint32_t i;
4330
4331	/*
4332	* Input validation.
4333	*/
4334	AssertPtrReturn(pszSymbol, VERR_INVALID_POINTER);
4335	pszEnd = RTStrEnd(pszSymbol, 512);
4336	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4337	cbSymbol = pszEnd - pszSymbol + 1;
4338
4339	if (pszModule)
4340	{
4341	AssertPtrReturn(pszModule, VERR_INVALID_POINTER);
4342	pszEnd = RTStrEnd(pszModule, 64);
4343	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4344	}
4345	Log3(("supdrvIDC_LdrGetSymbol: pszModule=%p:{%s} pszSymbol=%p:{%s}\n", pszModule, pszModule, pszSymbol, pszSymbol));
4346
4347
4348	if ( !pszModule
4349	\|\| !strcmp(pszModule, "SupDrv"))
4350	{
4351	/*
4352	* Search the support driver export table.
4353	*/
4354	for (i = 0; i < RT_ELEMENTS(g_aFunctions); i++)
4355	if (!strcmp(g_aFunctions[i].szName, pszSymbol))
4356	{
4357	pReq->u.Out.pfnSymbol = g_aFunctions[i].pfn;
4358	break;
4359	}
4360	}
4361	else
4362	{
4363	/*
4364	* Find the loader image.
4365	*/
4366	PSUPDRVLDRIMAGE pImage;
4367
4368	supdrvLdrLock(pDevExt);
4369
4370	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
4371	if (!strcmp(pImage->szName, pszModule))
4372	break;
4373	if (pImage && pImage->uState == SUP_IOCTL_LDR_LOAD)
4374	{
4375	/*
4376	* Search the symbol strings.
4377	*/
4378	const char *pchStrings = pImage->pachStrTab;
4379	PCSUPLDRSYM paSyms = pImage->paSymbols;
4380	for (i = 0; i < pImage->cSymbols; i++)
4381	{
4382	if ( paSyms[i].offSymbol < pImage->cbImageBits /* paranoia */
4383	&& paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4384	&& !memcmp(pchStrings + paSyms[i].offName, pszSymbol, cbSymbol))
4385	{
4386	/*
4387	* Found it! Calc the symbol address and add a reference to the module.
4388	*/
4389	pReq->u.Out.pfnSymbol = (PFNRT)((uint8_t *)pImage->pvImage + paSyms[i].offSymbol);
4390	rc = supdrvLdrAddUsage(pSession, pImage);
4391	break;
4392	}
4393	}
4394	}
4395	else
4396	rc = pImage ? VERR_WRONG_ORDER : VERR_MODULE_NOT_FOUND;
4397
4398	supdrvLdrUnlock(pDevExt);
4399	}
4400	return rc;
4401	}
4402
4403
4404	/**
4405	* Updates the VMMR0 entry point pointers.
4406	*
4407	* @returns IPRT status code.
4408	* @param pDevExt Device globals.
4409	* @param pSession Session data.
4410	* @param pVMMR0 VMMR0 image handle.
4411	* @param pvVMMR0EntryInt VMMR0EntryInt address.
4412	* @param pvVMMR0EntryFast VMMR0EntryFast address.
4413	* @param pvVMMR0EntryEx VMMR0EntryEx address.
4414	* @remark Caller must own the loader mutex.
4415	*/
4416	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx)
4417	{
4418	int rc = VINF_SUCCESS;
4419	LogFlow(("supdrvLdrSetR0EP pvVMMR0=%p pvVMMR0EntryInt=%p\n", pvVMMR0, pvVMMR0EntryInt));
4420
4421
4422	/*
4423	* Check if not yet set.
4424	*/
4425	if (!pDevExt->pvVMMR0)
4426	{
4427	pDevExt->pvVMMR0 = pvVMMR0;
4428	pDevExt->pfnVMMR0EntryInt = pvVMMR0EntryInt;
4429	pDevExt->pfnVMMR0EntryFast = pvVMMR0EntryFast;
4430	pDevExt->pfnVMMR0EntryEx = pvVMMR0EntryEx;
4431	}
4432	else
4433	{
4434	/*
4435	* Return failure or success depending on whether the values match or not.
4436	*/
4437	if ( pDevExt->pvVMMR0 != pvVMMR0
4438	\|\| (void *)pDevExt->pfnVMMR0EntryInt != pvVMMR0EntryInt
4439	\|\| (void *)pDevExt->pfnVMMR0EntryFast != pvVMMR0EntryFast
4440	\|\| (void *)pDevExt->pfnVMMR0EntryEx != pvVMMR0EntryEx)
4441	{
4442	AssertMsgFailed(("SUP_IOCTL_LDR_SETR0EP: Already set pointing to a different module!\n"));
4443	rc = VERR_INVALID_PARAMETER;
4444	}
4445	}
4446	return rc;
4447	}
4448
4449
4450	/**
4451	* Unsets the VMMR0 entry point installed by supdrvLdrSetR0EP.
4452	*
4453	* @param pDevExt Device globals.
4454	*/
4455	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt)
4456	{
4457	pDevExt->pvVMMR0 = NULL;
4458	pDevExt->pfnVMMR0EntryInt = NULL;
4459	pDevExt->pfnVMMR0EntryFast = NULL;
4460	pDevExt->pfnVMMR0EntryEx = NULL;
4461	}
4462
4463
4464	/**
4465	* Adds a usage reference in the specified session of an image.
4466	*
4467	* Called while owning the loader semaphore.
4468	*
4469	* @returns VINF_SUCCESS on success and VERR_NO_MEMORY on failure.
4470	* @param pSession Session in question.
4471	* @param pImage Image which the session is using.
4472	*/
4473	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage)
4474	{
4475	PSUPDRVLDRUSAGE pUsage;
4476	LogFlow(("supdrvLdrAddUsage: pImage=%p\n", pImage));
4477
4478	/*
4479	* Referenced it already?
4480	*/
4481	pUsage = pSession->pLdrUsage;
4482	while (pUsage)
4483	{
4484	if (pUsage->pImage == pImage)
4485	{
4486	pUsage->cUsage++;
4487	return VINF_SUCCESS;
4488	}
4489	pUsage = pUsage->pNext;
4490	}
4491
4492	/*
4493	* Allocate new usage record.
4494	*/
4495	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
4496	AssertReturn(pUsage, /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_5);
4497	pUsage->cUsage = 1;
4498	pUsage->pImage = pImage;
4499	pUsage->pNext = pSession->pLdrUsage;
4500	pSession->pLdrUsage = pUsage;
4501	return VINF_SUCCESS;
4502	}
4503
4504
4505	/**
4506	* Frees a load image.
4507	*
4508	* @param pDevExt Pointer to device extension.
4509	* @param pImage Pointer to the image we're gonna free.
4510	* This image must exit!
4511	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
4512	*/
4513	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
4514	{
4515	PSUPDRVLDRIMAGE pImagePrev;
4516	LogFlow(("supdrvLdrFree: pImage=%p\n", pImage));
4517
4518	/* find it - arg. should've used doubly linked list. */
4519	Assert(pDevExt->pLdrImages);
4520	pImagePrev = NULL;
4521	if (pDevExt->pLdrImages != pImage)
4522	{
4523	pImagePrev = pDevExt->pLdrImages;
4524	while (pImagePrev->pNext != pImage)
4525	pImagePrev = pImagePrev->pNext;
4526	Assert(pImagePrev->pNext == pImage);
4527	}
4528
4529	/* unlink */
4530	if (pImagePrev)
4531	pImagePrev->pNext = pImage->pNext;
4532	else
4533	pDevExt->pLdrImages = pImage->pNext;
4534
4535	/* check if this is VMMR0.r0 unset its entry point pointers. */
4536	if (pDevExt->pvVMMR0 == pImage->pvImage)
4537	supdrvLdrUnsetVMMR0EPs(pDevExt);
4538
4539	/* check for objects with destructors in this image. (Shouldn't happen.) */
4540	if (pDevExt->pObjs)
4541	{
4542	unsigned cObjs = 0;
4543	PSUPDRVOBJ pObj;
4544	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
4545	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
4546	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4547	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4548	{
4549	pObj->pfnDestructor = NULL;
4550	cObjs++;
4551	}
4552	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
4553	if (cObjs)
4554	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
4555	}
4556
4557	/* call termination function if fully loaded. */
4558	if ( pImage->pfnModuleTerm
4559	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
4560	{
4561	LogFlow(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
4562	pImage->pfnModuleTerm();
4563	}
4564
4565	/* do native unload if appropriate. */
4566	if (pImage->fNative)
4567	supdrvOSLdrUnload(pDevExt, pImage);
4568
4569	/* free the image */
4570	pImage->cUsage = 0;
4571	pImage->pNext = 0;
4572	pImage->uState = SUP_IOCTL_LDR_FREE;
4573	RTMemExecFree(pImage->pvImageAlloc, pImage->cbImageBits + 31);
4574	pImage->pvImageAlloc = NULL;
4575	RTMemFree(pImage->pachStrTab);
4576	pImage->pachStrTab = NULL;
4577	RTMemFree(pImage->paSymbols);
4578	pImage->paSymbols = NULL;
4579	RTMemFree(pImage);
4580	}
4581
4582
4583	/**
4584	* Acquires the loader lock.
4585	*
4586	* @returns IPRT status code.
4587	* @param pDevExt The device extension.
4588	*/
4589	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt)
4590	{
4591	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4592	int rc = RTSemMutexRequest(pDevExt->mtxLdr, RT_INDEFINITE_WAIT);
4593	#else
4594	int rc = RTSemFastMutexRequest(pDevExt->mtxLdr);
4595	#endif
4596	AssertRC(rc);
4597	return rc;
4598	}
4599
4600
4601	/**
4602	* Releases the loader lock.
4603	*
4604	* @returns IPRT status code.
4605	* @param pDevExt The device extension.
4606	*/
4607	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt)
4608	{
4609	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4610	return RTSemMutexRelease(pDevExt->mtxLdr);
4611	#else
4612	return RTSemFastMutexRelease(pDevExt->mtxLdr);
4613	#endif
4614	}
4615
4616
4617	/**
4618	* Implements the service call request.
4619	*
4620	* @returns VBox status code.
4621	* @param pDevExt The device extension.
4622	* @param pSession The calling session.
4623	* @param pReq The request packet, valid.
4624	*/
4625	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq)
4626	{
4627	#if !defined(RT_OS_WINDOWS) \|\| defined(DEBUG)
4628	int rc;
4629
4630	/*
4631	* Find the module first in the module referenced by the calling session.
4632	*/
4633	rc = supdrvLdrLock(pDevExt);
4634	if (RT_SUCCESS(rc))
4635	{
4636	PFNSUPR0SERVICEREQHANDLER pfnServiceReqHandler = NULL;
4637	PSUPDRVLDRUSAGE pUsage;
4638
4639	for (pUsage = pSession->pLdrUsage; pUsage; pUsage = pUsage->pNext)
4640	if ( pUsage->pImage->pfnServiceReqHandler
4641	&& !strcmp(pUsage->pImage->szName, pReq->u.In.szName))
4642	{
4643	pfnServiceReqHandler = pUsage->pImage->pfnServiceReqHandler;
4644	break;
4645	}
4646	supdrvLdrUnlock(pDevExt);
4647
4648	if (pfnServiceReqHandler)
4649	{
4650	/*
4651	* Call it.
4652	*/
4653	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
4654	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, NULL);
4655	else
4656	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0]);
4657	}
4658	else
4659	rc = VERR_SUPDRV_SERVICE_NOT_FOUND;
4660	}
4661
4662	/* log it */
4663	if ( RT_FAILURE(rc)
4664	&& rc != VERR_INTERRUPTED
4665	&& rc != VERR_TIMEOUT)
4666	Log(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4667	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4668	else
4669	Log4(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4670	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4671	return rc;
4672	#else /* RT_OS_WINDOWS && !DEBUG */
4673	return VERR_NOT_IMPLEMENTED;
4674	#endif /* RT_OS_WINDOWS && !DEBUG */
4675	}
4676
4677
4678	/**
4679	* Implements the logger settings request.
4680	*
4681	* @returns VBox status code.
4682	* @param pDevExt The device extension.
4683	* @param pSession The caller's session.
4684	* @param pReq The request.
4685	*/
4686	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq)
4687	{
4688	const char *pszGroup = &pReq->u.In.szStrings[pReq->u.In.offGroups];
4689	const char *pszFlags = &pReq->u.In.szStrings[pReq->u.In.offFlags];
4690	const char *pszDest = &pReq->u.In.szStrings[pReq->u.In.offDestination];
4691	PRTLOGGER pLogger = NULL;
4692	int rc;
4693
4694	/*
4695	* Some further validation.
4696	*/
4697	switch (pReq->u.In.fWhat)
4698	{
4699	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4700	case SUPLOGGERSETTINGS_WHAT_CREATE:
4701	break;
4702
4703	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4704	if (pszGroup \|\| pszFlags \|\| *pszDest)
4705	return VERR_INVALID_PARAMETER;
4706	if (pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_RELEASE)
4707	return VERR_ACCESS_DENIED;
4708	break;
4709
4710	default:
4711	return VERR_INTERNAL_ERROR;
4712	}
4713
4714	/*
4715	* Get the logger.
4716	*/
4717	switch (pReq->u.In.fWhich)
4718	{
4719	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4720	pLogger = RTLogGetDefaultInstance();
4721	break;
4722
4723	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4724	pLogger = RTLogRelDefaultInstance();
4725	break;
4726
4727	default:
4728	return VERR_INTERNAL_ERROR;
4729	}
4730
4731	/*
4732	* Do the job.
4733	*/
4734	switch (pReq->u.In.fWhat)
4735	{
4736	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4737	if (pLogger)
4738	{
4739	rc = RTLogFlags(pLogger, pszFlags);
4740	if (RT_SUCCESS(rc))
4741	rc = RTLogGroupSettings(pLogger, pszGroup);
4742	NOREF(pszDest);
4743	}
4744	else
4745	rc = VERR_NOT_FOUND;
4746	break;
4747
4748	case SUPLOGGERSETTINGS_WHAT_CREATE:
4749	{
4750	if (pLogger)
4751	rc = VERR_ALREADY_EXISTS;
4752	else
4753	{
4754	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
4755
4756	rc = RTLogCreate(&pLogger,
4757	0 /* fFlags */,
4758	pszGroup,
4759	pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_DEBUG
4760	? "VBOX_LOG"
4761	: "VBOX_RELEASE_LOG",
4762	RT_ELEMENTS(s_apszGroups),
4763	s_apszGroups,
4764	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER,
4765	NULL);
4766	if (RT_SUCCESS(rc))
4767	{
4768	rc = RTLogFlags(pLogger, pszFlags);
4769	NOREF(pszDest);
4770	if (RT_SUCCESS(rc))
4771	{
4772	switch (pReq->u.In.fWhich)
4773	{
4774	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4775	pLogger = RTLogSetDefaultInstance(pLogger);
4776	break;
4777	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4778	pLogger = RTLogRelSetDefaultInstance(pLogger);
4779	break;
4780	}
4781	}
4782	RTLogDestroy(pLogger);
4783	}
4784	}
4785	break;
4786	}
4787
4788	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4789	switch (pReq->u.In.fWhich)
4790	{
4791	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4792	pLogger = RTLogSetDefaultInstance(NULL);
4793	break;
4794	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4795	pLogger = RTLogRelSetDefaultInstance(NULL);
4796	break;
4797	}
4798	rc = RTLogDestroy(pLogger);
4799	break;
4800
4801	default:
4802	{
4803	rc = VERR_INTERNAL_ERROR;
4804	break;
4805	}
4806	}
4807
4808	return rc;
4809	}
4810
4811
4812	/**
4813	* Creates the GIP.
4814	*
4815	* @returns VBox status code.
4816	* @param pDevExt Instance data. GIP stuff may be updated.
4817	*/
4818	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
4819	{
4820	PSUPGLOBALINFOPAGE pGip;
4821	RTHCPHYS HCPhysGip;
4822	uint32_t u32SystemResolution;
4823	uint32_t u32Interval;
4824	unsigned cCpus;
4825	int rc;
4826
4827
4828	LogFlow(("supdrvGipCreate:\n"));
4829
4830	/* assert order */
4831	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
4832	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
4833	Assert(!pDevExt->pGipTimer);
4834
4835	/*
4836	* Check the CPU count.
4837	*/
4838	cCpus = RTMpGetArraySize();
4839	#ifdef SUP_WITH_LOTS_OF_CPUS
4840	if ( cCpus > RTCPUSET_MAX_CPUS
4841	\|\| cCpus > 256 /uint8_t is used for the mappings/)
4842	#else
4843	if (cCpus > RT_ELEMENTS(pGip->aCPUs))
4844	#endif
4845	{
4846	#ifdef SUP_WITH_LOTS_OF_CPUS
4847	SUPR0Printf("VBoxDrv: Too many CPUs (%u) for the GIP (max %u)\n", cCpus, RT_MIN(RTCPUSET_MAX_CPUS, 256));
4848	#else
4849	SUPR0Printf("VBoxDrv: Too many CPUs (%u) for the GIP (max %u)\n", cCpus, RT_MIN(RT_ELEMENTS(pGip->aCPUs), 256));
4850	#endif
4851	return VERR_TOO_MANY_CPUS;
4852	}
4853
4854	/*
4855	* Allocate a contiguous set of pages with a default kernel mapping.
4856	*/
4857	#ifdef SUP_WITH_LOTS_OF_CPUS
4858	rc = RTR0MemObjAllocCont(&pDevExt->GipMemObj, RT_UOFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), false /fExecutable/);
4859	#else
4860	cCpus = RT_ELEMENTS(pGip->aCPUs);
4861	rc = RTR0MemObjAllocLow(&pDevExt->GipMemObj, PAGE_SIZE, false /fExecutable/);
4862	#endif
4863	if (RT_FAILURE(rc))
4864	{
4865	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
4866	return rc;
4867	}
4868	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
4869	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
4870
4871	/*
4872	* Find a reasonable update interval and initialize the structure.
4873	*/
4874	u32Interval = u32SystemResolution = RTTimerGetSystemGranularity();
4875	while (u32Interval < 10000000 /* 10 ms */)
4876	u32Interval += u32SystemResolution;
4877
4878	supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), 1000000000 / u32Interval /=Hz/, cCpus);
4879
4880	/*
4881	* Create the timer.
4882	* If CPU_ALL isn't supported we'll have to fall back to synchronous mode.
4883	*/
4884	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
4885	{
4886	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, RTTIMER_FLAGS_CPU_ALL, supdrvGipAsyncTimer, pDevExt);
4887	if (rc == VERR_NOT_SUPPORTED)
4888	{
4889	OSDBGPRINT(("supdrvGipCreate: omni timer not supported, falling back to synchronous mode\n"));
4890	pGip->u32Mode = SUPGIPMODE_SYNC_TSC;
4891	}
4892	}
4893	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
4894	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0, supdrvGipSyncTimer, pDevExt);
4895	if (RT_SUCCESS(rc))
4896	{
4897	rc = RTMpNotificationRegister(supdrvGipMpEvent, pDevExt);
4898	if (RT_SUCCESS(rc))
4899	{
4900	rc = RTMpOnAll(supdrvGipInitOnCpu, pDevExt, pGip);
4901	if (RT_SUCCESS(rc))
4902	{
4903	/*
4904	* We're good.
4905	*/
4906	Log(("supdrvGipCreate: %u ns interval.\n", u32Interval));
4907	g_pSUPGlobalInfoPage = pGip;
4908	return VINF_SUCCESS;
4909	}
4910
4911	OSDBGPRINT(("supdrvGipCreate: RTMpOnAll failed with rc=%Rrc\n", rc));
4912	RTMpNotificationDeregister(supdrvGipMpEvent, pDevExt);
4913
4914	}
4915	else
4916	OSDBGPRINT(("supdrvGipCreate: failed to register MP event notfication. rc=%Rrc\n", rc));
4917	}
4918	else
4919	{
4920	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %u ns interval. rc=%Rrc\n", u32Interval, rc));
4921	Assert(!pDevExt->pGipTimer);
4922	}
4923	supdrvGipDestroy(pDevExt);
4924	return rc;
4925	}
4926
4927
4928	/**
4929	* Terminates the GIP.
4930	*
4931	* @param pDevExt Instance data. GIP stuff may be updated.
4932	*/
4933	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
4934	{
4935	int rc;
4936	#ifdef DEBUG_DARWIN_GIP
4937	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
4938	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
4939	pDevExt->pGipTimer, pDevExt->GipMemObj));
4940	#endif
4941
4942	/*
4943	* Invalid the GIP data.
4944	*/
4945	if (pDevExt->pGip)
4946	{
4947	supdrvGipTerm(pDevExt->pGip);
4948	pDevExt->pGip = NULL;
4949	}
4950	g_pSUPGlobalInfoPage = NULL;
4951
4952	/*
4953	* Destroy the timer and free the GIP memory object.
4954	*/
4955	if (pDevExt->pGipTimer)
4956	{
4957	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
4958	pDevExt->pGipTimer = NULL;
4959	}
4960
4961	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
4962	{
4963	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
4964	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
4965	}
4966
4967	/*
4968	* Finally, make sure we've release the system timer resolution request
4969	* if one actually succeeded and is still pending.
4970	*/
4971	if (pDevExt->u32SystemTimerGranularityGrant)
4972	{
4973	rc = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
4974	pDevExt->u32SystemTimerGranularityGrant = 0;
4975	}
4976	}
4977
4978
4979	/**
4980	* Timer callback function sync GIP mode.
4981	* @param pTimer The timer.
4982	* @param pvUser The device extension.
4983	*/
4984	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
4985	{
4986	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
4987	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4988	uint64_t u64TSC = ASMReadTSC();
4989	uint64_t NanoTS = RTTimeSystemNanoTS();
4990
4991	supdrvGipUpdate(pDevExt->pGip, NanoTS, u64TSC, NIL_RTCPUID, iTick);
4992
4993	ASMSetFlags(fOldFlags);
4994	}
4995
4996
4997	/**
4998	* Timer callback function for async GIP mode.
4999	* @param pTimer The timer.
5000	* @param pvUser The device extension.
5001	*/
5002	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
5003	{
5004	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
5005	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5006	RTCPUID idCpu = RTMpCpuId();
5007	uint64_t u64TSC = ASMReadTSC();
5008	uint64_t NanoTS = RTTimeSystemNanoTS();
5009
5010	/** @todo reset the transaction number and whatnot when iTick == 1. */
5011	if (pDevExt->idGipMaster == idCpu)
5012	supdrvGipUpdate(pDevExt->pGip, NanoTS, u64TSC, idCpu, iTick);
5013	else
5014	supdrvGipUpdatePerCpu(pDevExt->pGip, NanoTS, u64TSC, idCpu, ASMGetApicId(), iTick);
5015
5016	ASMSetFlags(fOldFlags);
5017	}
5018
5019
5020	/**
5021	* The calling CPU should be accounted as online, update GIP accordingly.
5022	*
5023	* This is used by supdrvGipMpEvent as well as the supdrvGipCreate.
5024	*
5025	* @param pGip The GIP.
5026	* @param idCpu The CPU ID.
5027	*/
5028	static void supdrvGipMpEventOnline(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
5029	{
5030	#ifdef SUP_WITH_LOTS_OF_CPUS
5031	int iCpuSet;
5032	uint8_t idApic;
5033	uint32_t i;
5034
5035	Assert(idCpu == RTMpCpuId());
5036	Assert(pGip->cPossibleCpus == RTMpGetCount());
5037
5038	/*
5039	* Update the globals.
5040	*/
5041	ASMAtomicWriteU32(&pGip->cPresentCpus, RTMpGetPresentCount());
5042	ASMAtomicWriteU32(&pGip->cOnlineCpus, RTMpGetOnlineCount());
5043	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5044	if (iCpuSet >= 0)
5045	{
5046	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5047	RTCpuSetAddByIndex(&pGip->OnlineCpuSet, iCpuSet);
5048	RTCpuSetAddByIndex(&pGip->PresentCpuSet, iCpuSet);
5049	}
5050
5051	/*
5052	* Find our entry, or allocate one if not found.
5053	* ASSUMES that CPU IDs are constant.
5054	*/
5055	for (i = 0; i < pGip->cCpus; i++)
5056	if (pGip->aCPUs[i].idCpu == idCpu)
5057	break;
5058
5059	if (i >= pGip->cCpus)
5060	for (i = 0; i < pGip->cCpus; i++)
5061	{
5062	bool fRc;
5063	ASMAtomicCmpXchgSize(&pGip->aCPUs[i].idCpu, idCpu, NIL_RTCPUID, fRc);
5064	if (fRc)
5065	break;
5066	}
5067
5068	AssertReturnVoid(i < pGip->cCpus);
5069
5070	/*
5071	* Update the entry.
5072	*/
5073	idApic = ASMGetApicId();
5074	ASMAtomicUoWriteU16(&pGip->aCPUs[i].idApic, idApic);
5075	ASMAtomicUoWriteS16(&pGip->aCPUs[i].iCpuSet, (int16_t)iCpuSet);
5076	ASMAtomicUoWriteSize(&pGip->aCPUs[i].idCpu, idCpu);
5077	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_ONLINE);
5078
5079	/*
5080	* Update the APIC ID and CPU set index mappings.
5081	*/
5082	ASMAtomicWriteU16(&pGip->aiCpuFromApicId[idApic], i);
5083	ASMAtomicWriteU16(&pGip->aiCpuFromCpuSetIdx[iCpuSet], i);
5084	#endif
5085	}
5086
5087
5088	/**
5089	* The CPU should be accounted as offline, update the GIP accordingly.
5090	*
5091	* This is used by supdrvGipMpEvent.
5092	*
5093	* @param pGip The GIP.
5094	* @param idCpu The CPU ID.
5095	*/
5096	static void supdrvGipMpEventOffline(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
5097	{
5098	#ifdef SUP_WITH_LOTS_OF_CPUS
5099	int iCpuSet;
5100	unsigned i;
5101
5102	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5103	AssertReturnVoid(iCpuSet >= 0);
5104
5105	i = pGip->aiCpuFromCpuSetIdx[iCpuSet];
5106	AssertReturnVoid(i < pGip->cCpus);
5107	AssertReturnVoid(pGip->aCPUs[i].idCpu == idCpu);
5108
5109	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5110	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_OFFLINE);
5111	RTCpuSetDelByIndex(&pGip->OnlineCpuSet, iCpuSet);
5112	#endif
5113	}
5114
5115
5116	/**
5117	* Multiprocessor event notification callback.
5118	*
5119	* This is used to make sure that the GIP master gets passed on to
5120	* another CPU. It also updates the associated CPU data.
5121	*
5122	* @param enmEvent The event.
5123	* @param idCpu The cpu it applies to.
5124	* @param pvUser Pointer to the device extension.
5125	*/
5126	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser)
5127	{
5128	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5129	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5130
5131	/*
5132	* Update the GIP CPU data.
5133	*/
5134	if (pGip)
5135	{
5136	switch (enmEvent)
5137	{
5138	case RTMPEVENT_ONLINE:
5139	supdrvGipMpEventOnline(pGip, idCpu);
5140	break;
5141	case RTMPEVENT_OFFLINE:
5142	supdrvGipMpEventOffline(pGip, idCpu);
5143	break;
5144
5145	}
5146	}
5147
5148	/*
5149	* Make sure there is a master GIP.
5150	*/
5151	if (enmEvent == RTMPEVENT_OFFLINE)
5152	{
5153	RTCPUID idGipMaster;
5154	ASMAtomicReadSize(&pDevExt->idGipMaster, &idGipMaster);
5155	if (idGipMaster == idCpu)
5156	{
5157	/*
5158	* Find a new GIP master.
5159	*/
5160	bool fIgnored;
5161	unsigned i;
5162	RTCPUID idNewGipMaster = NIL_RTCPUID;
5163	RTCPUSET OnlineCpus;
5164	RTMpGetOnlineSet(&OnlineCpus);
5165
5166	for (i = 0; i < RTCPUSET_MAX_CPUS; i++)
5167	{
5168	RTCPUID idCurCpu = RTMpCpuIdFromSetIndex(i);
5169	if ( RTCpuSetIsMember(&OnlineCpus, idCurCpu)
5170	&& idCurCpu != idGipMaster)
5171	{
5172	idNewGipMaster = idCurCpu;
5173	break;
5174	}
5175	}
5176
5177	Log(("supdrvGipMpEvent: Gip master %#lx -> %#lx\n", (long)idGipMaster, (long)idNewGipMaster));
5178	ASMAtomicCmpXchgSize(&pDevExt->idGipMaster, idNewGipMaster, idGipMaster, fIgnored);
5179	NOREF(fIgnored);
5180	}
5181	}
5182	}
5183
5184
5185	/**
5186	* Callback used by supdrvDetermineAsyncTSC to read the TSC on a CPU.
5187	*
5188	* @param idCpu Ignored.
5189	* @param pvUser1 Where to put the TSC.
5190	* @param pvUser2 Ignored.
5191	*/
5192	static DECLCALLBACK(void) supdrvDetermineAsyncTscWorker(RTCPUID idCpu, void pvUser1, void pvUser2)
5193	{
5194	#if 1
5195	ASMAtomicWriteU64((uint64_t volatile *)pvUser1, ASMReadTSC());
5196	#else
5197	(uint64_t )pvUser1 = ASMReadTSC();
5198	#endif
5199	}
5200
5201
5202	/**
5203	* Determine if Async GIP mode is required because of TSC drift.
5204	*
5205	* When using the default/normal timer code it is essential that the time stamp counter
5206	* (TSC) runs never backwards, that is, a read operation to the counter should return
5207	* a bigger value than any previous read operation. This is guaranteed by the latest
5208	* AMD CPUs and by newer Intel CPUs which never enter the C2 state (P4). In any other
5209	* case we have to choose the asynchronous timer mode.
5210	*
5211	* @param poffMin Pointer to the determined difference between different cores.
5212	* @return false if the time stamp counters appear to be synchronized, true otherwise.
5213	*/
5214	static bool supdrvDetermineAsyncTsc(uint64_t *poffMin)
5215	{
5216	/*
5217	* Just iterate all the cpus 8 times and make sure that the TSC is
5218	* ever increasing. We don't bother taking TSC rollover into account.
5219	*/
5220	int iEndCpu = RTMpGetArraySize();
5221	int iCpu;
5222	int cLoops = 8;
5223	bool fAsync = false;
5224	int rc = VINF_SUCCESS;
5225	uint64_t offMax = 0;
5226	uint64_t offMin = ~(uint64_t)0;
5227	uint64_t PrevTsc = ASMReadTSC();
5228
5229	while (cLoops-- > 0)
5230	{
5231	for (iCpu = 0; iCpu < iEndCpu; iCpu++)
5232	{
5233	uint64_t CurTsc;
5234	rc = RTMpOnSpecific(RTMpCpuIdFromSetIndex(iCpu), supdrvDetermineAsyncTscWorker, &CurTsc, NULL);
5235	if (RT_SUCCESS(rc))
5236	{
5237	if (CurTsc <= PrevTsc)
5238	{
5239	fAsync = true;
5240	offMin = offMax = PrevTsc - CurTsc;
5241	Log(("supdrvDetermineAsyncTsc: iCpu=%d cLoops=%d CurTsc=%llx PrevTsc=%llx\n",
5242	iCpu, cLoops, CurTsc, PrevTsc));
5243	break;
5244	}
5245
5246	/* Gather statistics (except the first time). */
5247	if (iCpu != 0 \|\| cLoops != 7)
5248	{
5249	uint64_t off = CurTsc - PrevTsc;
5250	if (off < offMin)
5251	offMin = off;
5252	if (off > offMax)
5253	offMax = off;
5254	Log2(("%d/%d: off=%llx\n", cLoops, iCpu, off));
5255	}
5256
5257	/* Next */
5258	PrevTsc = CurTsc;
5259	}
5260	else if (rc == VERR_NOT_SUPPORTED)
5261	break;
5262	else
5263	AssertMsg(rc == VERR_CPU_NOT_FOUND \|\| rc == VERR_CPU_OFFLINE, ("%d\n", rc));
5264	}
5265
5266	/* broke out of the loop. */
5267	if (iCpu < iEndCpu)
5268	break;
5269	}
5270
5271	poffMin = offMin; / Almost RTMpOnSpecific profiling. */
5272	Log(("supdrvDetermineAsyncTsc: returns %d; iEndCpu=%d rc=%d offMin=%llx offMax=%llx\n",
5273	fAsync, iEndCpu, rc, offMin, offMax));
5274	#if !defined(RT_OS_SOLARIS) && !defined(RT_OS_OS2) && !defined(RT_OS_WINDOWS)
5275	OSDBGPRINT(("vboxdrv: fAsync=%d offMin=%#lx offMax=%#lx\n", fAsync, (long)offMin, (long)offMax));
5276	#endif
5277	return fAsync;
5278	}
5279
5280
5281	/**
5282	* Determine the GIP TSC mode.
5283	*
5284	* @returns The most suitable TSC mode.
5285	* @param pDevExt Pointer to the device instance data.
5286	*/
5287	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt)
5288	{
5289	/*
5290	* On SMP we're faced with two problems:
5291	* (1) There might be a skew between the CPU, so that cpu0
5292	* returns a TSC that is slightly different from cpu1.
5293	* (2) Power management (and other things) may cause the TSC
5294	* to run at a non-constant speed, and cause the speed
5295	* to be different on the cpus. This will result in (1).
5296	*
5297	* So, on SMP systems we'll have to select the ASYNC update method
5298	* if there are symptoms of these problems.
5299	*/
5300	if (RTMpGetCount() > 1)
5301	{
5302	uint32_t uEAX, uEBX, uECX, uEDX;
5303	uint64_t u64DiffCoresIgnored;
5304
5305	/* Permit the user and/or the OS specific bits to force async mode. */
5306	if (supdrvOSGetForcedAsyncTscMode(pDevExt))
5307	return SUPGIPMODE_ASYNC_TSC;
5308
5309	/* Try check for current differences between the cpus. */
5310	if (supdrvDetermineAsyncTsc(&u64DiffCoresIgnored))
5311	return SUPGIPMODE_ASYNC_TSC;
5312
5313	/*
5314	* If the CPU supports power management and is an AMD one we
5315	* won't trust it unless it has the TscInvariant bit is set.
5316	*/
5317	/* Check for "AuthenticAMD" */
5318	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
5319	if ( uEAX >= 1
5320	&& uEBX == X86_CPUID_VENDOR_AMD_EBX
5321	&& uECX == X86_CPUID_VENDOR_AMD_ECX
5322	&& uEDX == X86_CPUID_VENDOR_AMD_EDX)
5323	{
5324	/* Check for APM support and that TscInvariant is cleared. */
5325	ASMCpuId(0x80000000, &uEAX, &uEBX, &uECX, &uEDX);
5326	if (uEAX >= 0x80000007)
5327	{
5328	ASMCpuId(0x80000007, &uEAX, &uEBX, &uECX, &uEDX);
5329	if ( !(uEDX & RT_BIT(8))/* TscInvariant */
5330	&& (uEDX & 0x3e)) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
5331	return SUPGIPMODE_ASYNC_TSC;
5332	}
5333	}
5334	}
5335	return SUPGIPMODE_SYNC_TSC;
5336	}
5337
5338
5339
5340	/**
5341	* Initializes the GIP data.
5342	*
5343	* @param pDevExt Pointer to the device instance data.
5344	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5345	* @param HCPhys The physical address of the GIP.
5346	* @param u64NanoTS The current nanosecond timestamp.
5347	* @param uUpdateHz The update frequency.
5348	* @param cCpus The CPU count.
5349	*/
5350	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
5351	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned cCpus)
5352	{
5353	size_t const cbGip = RT_ALIGN_Z(RT_OFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), PAGE_SIZE);
5354	unsigned i;
5355	#ifdef DEBUG_DARWIN_GIP
5356	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5357	#else
5358	LogFlow(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5359	#endif
5360
5361	/*
5362	* Initialize the structure.
5363	*/
5364	memset(pGip, 0, cbGip);
5365	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
5366	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
5367	pGip->u32Mode = supdrvGipDeterminTscMode(pDevExt);
5368	#ifdef SUP_WITH_LOTS_OF_CPUS
5369	pGip->cCpus = (uint16_t)cCpus;
5370	pGip->cPages = (uint16_t)(cbGip / PAGE_SIZE);
5371	#endif
5372	pGip->u32UpdateHz = uUpdateHz;
5373	pGip->u32UpdateIntervalNS = 1000000000 / uUpdateHz;
5374	pGip->u64NanoTSLastUpdateHz = u64NanoTS;
5375	#ifdef SUP_WITH_LOTS_OF_CPUS
5376	RTCpuSetEmpty(&pGip->OnlineCpuSet);
5377	pGip->cOnlineCpus = RTMpGetOnlineCount();
5378	pGip->cPresentCpus = RTMpGetPresentCount();
5379	RTCpuSetEmpty(&pGip->PresentCpuSet);
5380	RTMpGetSet(&pGip->PossibleCpuSet);
5381	pGip->cPossibleCpus = RTMpGetCount();
5382	pGip->idCpuMax = RTMpGetMaxCpuId();
5383	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromApicId); i++)
5384	pGip->aiCpuFromApicId[i] = 0;
5385	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx); i++)
5386	pGip->aiCpuFromCpuSetIdx[i] = UINT16_MAX;
5387	#endif
5388
5389	for (i = 0; i < cCpus; i++)
5390	{
5391	pGip->aCPUs[i].u32TransactionId = 2;
5392	pGip->aCPUs[i].u64NanoTS = u64NanoTS;
5393	pGip->aCPUs[i].u64TSC = ASMReadTSC();
5394
5395	#ifdef SUP_WITH_LOTS_OF_CPUS
5396	pGip->aCPUs[i].enmState = SUPGIPCPUSTATE_INVALID;
5397	pGip->aCPUs[i].idCpu = NIL_RTCPUID;
5398	pGip->aCPUs[i].iCpuSet = -1;
5399	pGip->aCPUs[i].idApic = UINT8_MAX;
5400	#endif
5401
5402	/*
5403	* We don't know the following values until we've executed updates.
5404	* So, we'll just pretend it's a 4 GHz CPU and adjust the history it on
5405	* the 2nd timer callout.
5406	*/
5407	pGip->aCPUs[i].u64CpuHz = _4G + 1; /* tstGIP-2 depends on this. */
5408	pGip->aCPUs[i].u32UpdateIntervalTSC
5409	= pGip->aCPUs[i].au32TSCHistory[0]
5410	= pGip->aCPUs[i].au32TSCHistory[1]
5411	= pGip->aCPUs[i].au32TSCHistory[2]
5412	= pGip->aCPUs[i].au32TSCHistory[3]
5413	= pGip->aCPUs[i].au32TSCHistory[4]
5414	= pGip->aCPUs[i].au32TSCHistory[5]
5415	= pGip->aCPUs[i].au32TSCHistory[6]
5416	= pGip->aCPUs[i].au32TSCHistory[7]
5417	= /pGip->aCPUs[i].u64CpuHz/ (uint32_t)(_4G / uUpdateHz);
5418	}
5419
5420	/*
5421	* Link it to the device extension.
5422	*/
5423	pDevExt->pGip = pGip;
5424	pDevExt->HCPhysGip = HCPhys;
5425	pDevExt->cGipUsers = 0;
5426	}
5427
5428
5429	/**
5430	* On CPU initialization callback for RTMpOnAll.
5431	*
5432	* @param idCpu The CPU ID.
5433	* @param pvUser1 The device extension.
5434	* @param pvUser2 The GIP.
5435	*/
5436	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2)
5437	{
5438	/* This is good enough, even though it will update some of the globals a
5439	bit to much. */
5440	supdrvGipMpEventOnline((PSUPGLOBALINFOPAGE)pvUser2, idCpu);
5441	}
5442
5443
5444	/**
5445	* Invalidates the GIP data upon termination.
5446	*
5447	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5448	*/
5449	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
5450	{
5451	unsigned i;
5452	pGip->u32Magic = 0;
5453	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
5454	{
5455	pGip->aCPUs[i].u64NanoTS = 0;
5456	pGip->aCPUs[i].u64TSC = 0;
5457	pGip->aCPUs[i].iTSCHistoryHead = 0;
5458	}
5459	}
5460
5461
5462	/**
5463	* Worker routine for supdrvGipUpdate and supdrvGipUpdatePerCpu that
5464	* updates all the per cpu data except the transaction id.
5465	*
5466	* @param pGip The GIP.
5467	* @param pGipCpu Pointer to the per cpu data.
5468	* @param u64NanoTS The current time stamp.
5469	* @param u64TSC The current TSC.
5470	* @param iTick The current timer tick.
5471	*/
5472	static void supdrvGipDoUpdateCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick)
5473	{
5474	uint64_t u64TSCDelta;
5475	uint32_t u32UpdateIntervalTSC;
5476	uint32_t u32UpdateIntervalTSCSlack;
5477	unsigned iTSCHistoryHead;
5478	uint64_t u64CpuHz;
5479	uint32_t u32TransactionId;
5480
5481	/* Delta between this and the previous update. */
5482	ASMAtomicUoWriteU32(&pGipCpu->u32PrevUpdateIntervalNS, (uint32_t)(u64NanoTS - pGipCpu->u64NanoTS));
5483
5484	/*
5485	* Update the NanoTS.
5486	*/
5487	ASMAtomicWriteU64(&pGipCpu->u64NanoTS, u64NanoTS);
5488
5489	/*
5490	* Calc TSC delta.
5491	*/
5492	/** @todo validate the NanoTS delta, don't trust the OS to call us when it should... */
5493	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
5494	ASMAtomicWriteU64(&pGipCpu->u64TSC, u64TSC);
5495
5496	if (u64TSCDelta >> 32)
5497	{
5498	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
5499	pGipCpu->cErrors++;
5500	}
5501
5502	/*
5503	* On the 2nd and 3rd callout, reset the history with the current TSC
5504	* interval since the values entered by supdrvGipInit are totally off.
5505	* The interval on the 1st callout completely unreliable, the 2nd is a bit
5506	* better, while the 3rd should be most reliable.
5507	*/
5508	u32TransactionId = pGipCpu->u32TransactionId;
5509	if (RT_UNLIKELY( ( u32TransactionId == 5
5510	\|\| u32TransactionId == 7)
5511	&& ( iTick == 2
5512	\|\| iTick == 3) ))
5513	{
5514	unsigned i;
5515	for (i = 0; i < RT_ELEMENTS(pGipCpu->au32TSCHistory); i++)
5516	ASMAtomicUoWriteU32(&pGipCpu->au32TSCHistory[i], (uint32_t)u64TSCDelta);
5517	}
5518
5519	/*
5520	* TSC History.
5521	*/
5522	Assert(RT_ELEMENTS(pGipCpu->au32TSCHistory) == 8);
5523	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
5524	ASMAtomicWriteU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
5525	ASMAtomicWriteU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
5526
5527	/*
5528	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
5529	*/
5530	if (pGip->u32UpdateHz >= 1000)
5531	{
5532	uint32_t u32;
5533	u32 = pGipCpu->au32TSCHistory[0];
5534	u32 += pGipCpu->au32TSCHistory[1];
5535	u32 += pGipCpu->au32TSCHistory[2];
5536	u32 += pGipCpu->au32TSCHistory[3];
5537	u32 >>= 2;
5538	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
5539	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
5540	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
5541	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
5542	u32UpdateIntervalTSC >>= 2;
5543	u32UpdateIntervalTSC += u32;
5544	u32UpdateIntervalTSC >>= 1;
5545
5546	/* Value chosen for a 2GHz Athlon64 running linux 2.6.10/11, . */
5547	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
5548	}
5549	else if (pGip->u32UpdateHz >= 90)
5550	{
5551	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5552	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
5553	u32UpdateIntervalTSC >>= 1;
5554
5555	/* value chosen on a 2GHz thinkpad running windows */
5556	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
5557	}
5558	else
5559	{
5560	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5561
5562	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
5563	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
5564	}
5565	ASMAtomicWriteU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
5566
5567	/*
5568	* CpuHz.
5569	*/
5570	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, pGip->u32UpdateHz);
5571	ASMAtomicWriteU64(&pGipCpu->u64CpuHz, u64CpuHz);
5572	}
5573
5574
5575	/**
5576	* Updates the GIP.
5577	*
5578	* @param pGip Pointer to the GIP.
5579	* @param u64NanoTS The current nanosecond timesamp.
5580	* @param u64TSC The current TSC timesamp.
5581	* @param idCpu The CPU ID.
5582	* @param iTick The current timer tick.
5583	*/
5584	static void supdrvGipUpdate(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick)
5585	{
5586	/*
5587	* Determine the relevant CPU data.
5588	*/
5589	PSUPGIPCPU pGipCpu;
5590	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
5591	pGipCpu = &pGip->aCPUs[0];
5592	else
5593	{
5594	#ifdef SUP_WITH_LOTS_OF_CPUS
5595	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
5596	if (RT_UNLIKELY(iCpu >= pGip->cCpus))
5597	return;
5598	pGipCpu = &pGip->aCPUs[iCpu];
5599	if (RT_UNLIKELY(pGipCpu->idCpu != idCpu))
5600	return;
5601	#else
5602	unsigned iCpu = ASMGetApicId();
5603	if (RT_UNLIKELY(iCpu >= RT_ELEMENTS(pGip->aCPUs)))
5604	return;
5605	pGipCpu = &pGip->aCPUs[iCpu];
5606	#endif
5607	}
5608
5609	/*
5610	* Start update transaction.
5611	*/
5612	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5613	{
5614	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
5615	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5616	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5617	pGipCpu->cErrors++;
5618	return;
5619	}
5620
5621	/*
5622	* Recalc the update frequency every 0x800th time.
5623	*/
5624	if (!(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
5625	{
5626	if (pGip->u64NanoTSLastUpdateHz)
5627	{
5628	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
5629	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
5630	uint32_t u32UpdateHz = (uint32_t)((UINT64_C(1000000000) * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
5631	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
5632	{
5633	ASMAtomicWriteU32(&pGip->u32UpdateHz, u32UpdateHz);
5634	ASMAtomicWriteU32(&pGip->u32UpdateIntervalNS, 1000000000 / u32UpdateHz);
5635	}
5636	#endif
5637	}
5638	ASMAtomicWriteU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS);
5639	}
5640
5641	/*
5642	* Update the data.
5643	*/
5644	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS, u64TSC, iTick);
5645
5646	/*
5647	* Complete transaction.
5648	*/
5649	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5650	}
5651
5652
5653	/**
5654	* Updates the per cpu GIP data for the calling cpu.
5655	*
5656	* @param pGip Pointer to the GIP.
5657	* @param u64NanoTS The current nanosecond timesamp.
5658	* @param u64TSC The current TSC timesamp.
5659	* @param idCpu The CPU ID.
5660	* @param idApic The APIC id for the CPU index.
5661	* @param iTick The current timer tick.
5662	*/
5663	static void supdrvGipUpdatePerCpu(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC,
5664	RTCPUID idCpu, uint8_t idApic, uint64_t iTick)
5665	{
5666	#ifdef SUP_WITH_LOTS_OF_CPUS
5667	unsigned iCpu = pGip->aiCpuFromApicId[idApic];
5668
5669	if (RT_LIKELY(iCpu < pGip->cCpus))
5670	#else
5671	unsigned iCpu = idApic;
5672
5673	if (RT_LIKELY(iCpu < RT_ELEMENTS(pGip->aCPUs)))
5674	#endif
5675	{
5676	PSUPGIPCPU pGipCpu = &pGip->aCPUs[iCpu];
5677	#ifdef SUP_WITH_LOTS_OF_CPUS
5678	if (pGipCpu->idCpu == idCpu)
5679	#endif
5680	{
5681
5682	/*
5683	* Start update transaction.
5684	*/
5685	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5686	{
5687	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5688	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5689	pGipCpu->cErrors++;
5690	return;
5691	}
5692
5693	/*
5694	* Update the data.
5695	*/
5696	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS, u64TSC, iTick);
5697
5698	/*
5699	* Complete transaction.
5700	*/
5701	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5702	}
5703	}
5704	}

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

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDrv.c@ 36254

Download in other formats: