SUPDrv.c@ 36191

Last change on this file since 36191 was 35346, checked in by vboxsync, 14 years ago
VMM reorg: Moving the public include files from include/VBox to include/VBox/vmm.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 204.8 KB

Line
1	/* $Revision: 35346 $ */
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, uint64_t u64NanoTS, unsigned uUpdateHz);
134	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip);
135	static void supdrvGipUpdate(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick);
136	static void supdrvGipUpdatePerCpu(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC, unsigned iCpu, uint64_t iTick);
137
138
139	/*******************************************************************************
140	* Global Variables *
141	*******************************************************************************/
142	DECLEXPORT(PSUPGLOBALINFOPAGE) g_pSUPGlobalInfoPage = NULL;
143
144	/**
145	* Array of the R0 SUP API.
146	*/
147	static SUPFUNC g_aFunctions[] =
148	{
149	/* name function */
150	/* Entries with absolute addresses determined at runtime, fixup
151	code makes ugly ASSUMPTIONS about the order here: */
152	{ "SUPR0AbsIs64bit", (void *)0 },
153	{ "SUPR0Abs64bitKernelCS", (void *)0 },
154	{ "SUPR0Abs64bitKernelSS", (void *)0 },
155	{ "SUPR0Abs64bitKernelDS", (void *)0 },
156	{ "SUPR0AbsKernelCS", (void *)0 },
157	{ "SUPR0AbsKernelSS", (void *)0 },
158	{ "SUPR0AbsKernelDS", (void *)0 },
159	{ "SUPR0AbsKernelES", (void *)0 },
160	{ "SUPR0AbsKernelFS", (void *)0 },
161	{ "SUPR0AbsKernelGS", (void *)0 },
162	/* Normal function pointers: */
163	{ "SUPR0ComponentRegisterFactory", (void *)SUPR0ComponentRegisterFactory },
164	{ "SUPR0ComponentDeregisterFactory", (void *)SUPR0ComponentDeregisterFactory },
165	{ "SUPR0ComponentQueryFactory", (void *)SUPR0ComponentQueryFactory },
166	{ "SUPR0ObjRegister", (void *)SUPR0ObjRegister },
167	{ "SUPR0ObjAddRef", (void *)SUPR0ObjAddRef },
168	{ "SUPR0ObjAddRefEx", (void *)SUPR0ObjAddRefEx },
169	{ "SUPR0ObjRelease", (void *)SUPR0ObjRelease },
170	{ "SUPR0ObjVerifyAccess", (void *)SUPR0ObjVerifyAccess },
171	{ "SUPR0LockMem", (void *)SUPR0LockMem },
172	{ "SUPR0UnlockMem", (void *)SUPR0UnlockMem },
173	{ "SUPR0ContAlloc", (void *)SUPR0ContAlloc },
174	{ "SUPR0ContFree", (void *)SUPR0ContFree },
175	{ "SUPR0LowAlloc", (void *)SUPR0LowAlloc },
176	{ "SUPR0LowFree", (void *)SUPR0LowFree },
177	{ "SUPR0MemAlloc", (void *)SUPR0MemAlloc },
178	{ "SUPR0MemGetPhys", (void *)SUPR0MemGetPhys },
179	{ "SUPR0MemFree", (void *)SUPR0MemFree },
180	{ "SUPR0PageAllocEx", (void *)SUPR0PageAllocEx },
181	{ "SUPR0PageFree", (void *)SUPR0PageFree },
182	{ "SUPR0Printf", (void )SUPR0Printf }, /* @todo needs wrapping? */
183	{ "SUPSemEventCreate", (void *)SUPSemEventCreate },
184	{ "SUPSemEventClose", (void *)SUPSemEventClose },
185	{ "SUPSemEventSignal", (void *)SUPSemEventSignal },
186	{ "SUPSemEventWait", (void *)SUPSemEventWait },
187	{ "SUPSemEventWaitNoResume", (void *)SUPSemEventWaitNoResume },
188	{ "SUPSemEventWaitNsAbsIntr", (void *)SUPSemEventWaitNsAbsIntr },
189	{ "SUPSemEventWaitNsRelIntr", (void *)SUPSemEventWaitNsRelIntr },
190	{ "SUPSemEventGetResolution", (void *)SUPSemEventGetResolution },
191	{ "SUPSemEventMultiCreate", (void *)SUPSemEventMultiCreate },
192	{ "SUPSemEventMultiClose", (void *)SUPSemEventMultiClose },
193	{ "SUPSemEventMultiSignal", (void *)SUPSemEventMultiSignal },
194	{ "SUPSemEventMultiReset", (void *)SUPSemEventMultiReset },
195	{ "SUPSemEventMultiWait", (void *)SUPSemEventMultiWait },
196	{ "SUPSemEventMultiWaitNoResume", (void *)SUPSemEventMultiWaitNoResume },
197	{ "SUPSemEventMultiWaitNsAbsIntr", (void *)SUPSemEventMultiWaitNsAbsIntr },
198	{ "SUPSemEventMultiWaitNsRelIntr", (void *)SUPSemEventMultiWaitNsRelIntr },
199	{ "SUPSemEventMultiGetResolution", (void *)SUPSemEventMultiGetResolution },
200	{ "SUPR0GetPagingMode", (void *)SUPR0GetPagingMode },
201	{ "SUPR0EnableVTx", (void *)SUPR0EnableVTx },
202	{ "SUPGetGIP", (void *)SUPGetGIP },
203	{ "g_pSUPGlobalInfoPage", (void *)&g_pSUPGlobalInfoPage },
204	{ "RTMemAllocTag", (void *)RTMemAllocTag },
205	{ "RTMemAllocZTag", (void *)RTMemAllocZTag },
206	{ "RTMemAllocVarTag", (void *)RTMemAllocVarTag },
207	{ "RTMemAllocZVarTag", (void *)RTMemAllocZVarTag },
208	{ "RTMemFree", (void *)RTMemFree },
209	{ "RTMemDupTag", (void *)RTMemDupTag },
210	{ "RTMemDupExTag", (void *)RTMemDupExTag },
211	{ "RTMemReallocTag", (void *)RTMemReallocTag },
212	{ "RTR0MemObjAllocLowTag", (void *)RTR0MemObjAllocLowTag },
213	{ "RTR0MemObjAllocPageTag", (void *)RTR0MemObjAllocPageTag },
214	{ "RTR0MemObjAllocPhysTag", (void *)RTR0MemObjAllocPhysTag },
215	{ "RTR0MemObjAllocPhysExTag", (void *)RTR0MemObjAllocPhysExTag },
216	{ "RTR0MemObjAllocPhysNCTag", (void *)RTR0MemObjAllocPhysNCTag },
217	{ "RTR0MemObjAllocContTag", (void *)RTR0MemObjAllocContTag },
218	{ "RTR0MemObjEnterPhysTag", (void *)RTR0MemObjEnterPhysTag },
219	{ "RTR0MemObjLockUserTag", (void *)RTR0MemObjLockUserTag },
220	{ "RTR0MemObjMapKernelTag", (void *)RTR0MemObjMapKernelTag },
221	{ "RTR0MemObjMapKernelExTag", (void *)RTR0MemObjMapKernelExTag },
222	{ "RTR0MemObjMapUserTag", (void *)RTR0MemObjMapUserTag },
223	{ "RTR0MemObjProtect", (void *)RTR0MemObjProtect },
224	{ "RTR0MemObjAddress", (void *)RTR0MemObjAddress },
225	{ "RTR0MemObjAddressR3", (void *)RTR0MemObjAddressR3 },
226	{ "RTR0MemObjSize", (void *)RTR0MemObjSize },
227	{ "RTR0MemObjIsMapping", (void *)RTR0MemObjIsMapping },
228	{ "RTR0MemObjGetPagePhysAddr", (void *)RTR0MemObjGetPagePhysAddr },
229	{ "RTR0MemObjFree", (void *)RTR0MemObjFree },
230	{ "RTR0MemUserCopyFrom", (void *)RTR0MemUserCopyFrom },
231	{ "RTR0MemUserCopyTo", (void *)RTR0MemUserCopyTo },
232	{ "RTR0MemUserIsValidAddr", (void *)RTR0MemUserIsValidAddr },
233	{ "RTR0MemKernelIsValidAddr", (void *)RTR0MemKernelIsValidAddr },
234	{ "RTR0MemAreKrnlAndUsrDifferent", (void *)RTR0MemAreKrnlAndUsrDifferent },
235	{ "RTSemMutexCreate", (void *)RTSemMutexCreate },
236	{ "RTSemMutexRequest", (void *)RTSemMutexRequest },
237	{ "RTSemMutexRequestDebug", (void *)RTSemMutexRequestDebug },
238	{ "RTSemMutexRequestNoResume", (void *)RTSemMutexRequestNoResume },
239	{ "RTSemMutexRequestNoResumeDebug", (void *)RTSemMutexRequestNoResumeDebug },
240	{ "RTSemMutexRelease", (void *)RTSemMutexRelease },
241	{ "RTSemMutexDestroy", (void *)RTSemMutexDestroy },
242	{ "RTProcSelf", (void *)RTProcSelf },
243	{ "RTR0ProcHandleSelf", (void *)RTR0ProcHandleSelf },
244	{ "RTSemFastMutexCreate", (void *)RTSemFastMutexCreate },
245	{ "RTSemFastMutexDestroy", (void *)RTSemFastMutexDestroy },
246	{ "RTSemFastMutexRequest", (void *)RTSemFastMutexRequest },
247	{ "RTSemFastMutexRelease", (void *)RTSemFastMutexRelease },
248	{ "RTSemEventCreate", (void *)RTSemEventCreate },
249	{ "RTSemEventSignal", (void *)RTSemEventSignal },
250	{ "RTSemEventWait", (void *)RTSemEventWait },
251	{ "RTSemEventWaitNoResume", (void *)RTSemEventWaitNoResume },
252	{ "RTSemEventWaitEx", (void *)RTSemEventWaitEx },
253	{ "RTSemEventWaitExDebug", (void *)RTSemEventWaitExDebug },
254	{ "RTSemEventGetResolution", (void *)RTSemEventGetResolution },
255	{ "RTSemEventDestroy", (void *)RTSemEventDestroy },
256	{ "RTSemEventMultiCreate", (void *)RTSemEventMultiCreate },
257	{ "RTSemEventMultiSignal", (void *)RTSemEventMultiSignal },
258	{ "RTSemEventMultiReset", (void *)RTSemEventMultiReset },
259	{ "RTSemEventMultiWait", (void *)RTSemEventMultiWait },
260	{ "RTSemEventMultiWaitNoResume", (void *)RTSemEventMultiWaitNoResume },
261	{ "RTSemEventMultiWaitEx", (void *)RTSemEventMultiWaitEx },
262	{ "RTSemEventMultiWaitExDebug", (void *)RTSemEventMultiWaitExDebug },
263	{ "RTSemEventMultiGetResolution", (void *)RTSemEventMultiGetResolution },
264	{ "RTSemEventMultiDestroy", (void *)RTSemEventMultiDestroy },
265	{ "RTSpinlockCreate", (void *)RTSpinlockCreate },
266	{ "RTSpinlockDestroy", (void *)RTSpinlockDestroy },
267	{ "RTSpinlockAcquire", (void *)RTSpinlockAcquire },
268	{ "RTSpinlockRelease", (void *)RTSpinlockRelease },
269	{ "RTSpinlockAcquireNoInts", (void *)RTSpinlockAcquireNoInts },
270	{ "RTSpinlockReleaseNoInts", (void *)RTSpinlockReleaseNoInts },
271	{ "RTTimeNanoTS", (void *)RTTimeNanoTS },
272	{ "RTTimeMilliTS", (void *)RTTimeMilliTS },
273	{ "RTTimeSystemNanoTS", (void *)RTTimeSystemNanoTS },
274	{ "RTTimeSystemMilliTS", (void *)RTTimeSystemMilliTS },
275	{ "RTThreadNativeSelf", (void *)RTThreadNativeSelf },
276	{ "RTThreadSleep", (void *)RTThreadSleep },
277	{ "RTThreadYield", (void *)RTThreadYield },
278	#if 0 /* Thread APIs, Part 2. */
279	{ "RTThreadSelf", (void *)RTThreadSelf },
280	{ "RTThreadCreate", (void *)RTThreadCreate },
281	{ "RTThreadGetNative", (void *)RTThreadGetNative },
282	{ "RTThreadWait", (void *)RTThreadWait },
283	{ "RTThreadWaitNoResume", (void *)RTThreadWaitNoResume },
284	{ "RTThreadGetName", (void *)RTThreadGetName },
285	{ "RTThreadSelfName", (void *)RTThreadSelfName },
286	{ "RTThreadGetType", (void *)RTThreadGetType },
287	{ "RTThreadUserSignal", (void *)RTThreadUserSignal },
288	{ "RTThreadUserReset", (void *)RTThreadUserReset },
289	{ "RTThreadUserWait", (void *)RTThreadUserWait },
290	{ "RTThreadUserWaitNoResume", (void *)RTThreadUserWaitNoResume },
291	#endif
292	{ "RTThreadPreemptIsEnabled", (void *)RTThreadPreemptIsEnabled },
293	{ "RTThreadPreemptIsPending", (void *)RTThreadPreemptIsPending },
294	{ "RTThreadPreemptIsPendingTrusty", (void *)RTThreadPreemptIsPendingTrusty },
295	{ "RTThreadPreemptIsPossible", (void *)RTThreadPreemptIsPossible },
296	{ "RTThreadPreemptDisable", (void *)RTThreadPreemptDisable },
297	{ "RTThreadPreemptRestore", (void *)RTThreadPreemptRestore },
298	{ "RTThreadIsInInterrupt", (void *)RTThreadIsInInterrupt },
299	{ "RTTimerCreate", (void *)RTTimerCreate },
300	{ "RTTimerCreateEx", (void *)RTTimerCreateEx },
301	{ "RTTimerDestroy", (void *)RTTimerDestroy },
302	{ "RTTimerStart", (void *)RTTimerStart },
303	{ "RTTimerStop", (void *)RTTimerStop },
304	{ "RTTimerChangeInterval", (void *)RTTimerChangeInterval },
305	{ "RTTimerGetSystemGranularity", (void *)RTTimerGetSystemGranularity },
306	{ "RTTimerRequestSystemGranularity", (void *)RTTimerRequestSystemGranularity },
307	{ "RTTimerReleaseSystemGranularity", (void *)RTTimerReleaseSystemGranularity },
308	{ "RTTimerCanDoHighResolution", (void *)RTTimerCanDoHighResolution },
309
310	{ "RTLogDefaultInstance", (void *)RTLogDefaultInstance },
311	{ "RTMpCpuId", (void *)RTMpCpuId },
312	{ "RTMpCpuIdFromSetIndex", (void *)RTMpCpuIdFromSetIndex },
313	{ "RTMpCpuIdToSetIndex", (void *)RTMpCpuIdToSetIndex },
314	{ "RTMpGetArraySize", (void *)RTMpGetArraySize },
315	{ "RTMpIsCpuPossible", (void *)RTMpIsCpuPossible },
316	{ "RTMpGetCount", (void *)RTMpGetCount },
317	{ "RTMpGetMaxCpuId", (void *)RTMpGetMaxCpuId },
318	{ "RTMpGetOnlineCount", (void *)RTMpGetOnlineCount },
319	{ "RTMpGetOnlineSet", (void *)RTMpGetOnlineSet },
320	{ "RTMpGetSet", (void *)RTMpGetSet },
321	{ "RTMpIsCpuOnline", (void *)RTMpIsCpuOnline },
322	{ "RTMpIsCpuWorkPending", (void *)RTMpIsCpuWorkPending },
323	{ "RTMpOnAll", (void *)RTMpOnAll },
324	{ "RTMpOnOthers", (void *)RTMpOnOthers },
325	{ "RTMpOnSpecific", (void *)RTMpOnSpecific },
326	{ "RTMpPokeCpu", (void *)RTMpPokeCpu },
327	{ "RTPowerNotificationRegister", (void *)RTPowerNotificationRegister },
328	{ "RTPowerNotificationDeregister", (void *)RTPowerNotificationDeregister },
329	{ "RTLogRelDefaultInstance", (void *)RTLogRelDefaultInstance },
330	{ "RTLogSetDefaultInstanceThread", (void *)RTLogSetDefaultInstanceThread },
331	{ "RTLogLoggerExV", (void *)RTLogLoggerExV },
332	{ "RTLogPrintfV", (void *)RTLogPrintfV },
333	{ "RTR0AssertPanicSystem", (void *)RTR0AssertPanicSystem },
334	{ "RTAssertMsg1", (void *)RTAssertMsg1 },
335	{ "RTAssertMsg2V", (void *)RTAssertMsg2V },
336	{ "RTAssertSetQuiet", (void *)RTAssertSetQuiet },
337	{ "RTAssertMayPanic", (void *)RTAssertMayPanic },
338	{ "RTAssertSetMayPanic", (void *)RTAssertSetMayPanic },
339	{ "RTAssertAreQuiet", (void *)RTAssertAreQuiet },
340	{ "RTStrFormat", (void *)RTStrFormat },
341	{ "RTStrFormatNumber", (void *)RTStrFormatNumber },
342	{ "RTStrFormatTypeDeregister", (void *)RTStrFormatTypeDeregister },
343	{ "RTStrFormatTypeRegister", (void *)RTStrFormatTypeRegister },
344	{ "RTStrFormatTypeSetUser", (void *)RTStrFormatTypeSetUser },
345	{ "RTStrFormatV", (void *)RTStrFormatV },
346	{ "RTStrPrintf", (void *)RTStrPrintf },
347	{ "RTStrPrintfEx", (void *)RTStrPrintfEx },
348	{ "RTStrPrintfExV", (void *)RTStrPrintfExV },
349	{ "RTStrPrintfV", (void *)RTStrPrintfV },
350	{ "RTHandleTableAllocWithCtx", (void *)RTHandleTableAllocWithCtx },
351	{ "RTHandleTableCreate", (void *)RTHandleTableCreate },
352	{ "RTHandleTableCreateEx", (void *)RTHandleTableCreateEx },
353	{ "RTHandleTableDestroy", (void *)RTHandleTableDestroy },
354	{ "RTHandleTableFreeWithCtx", (void *)RTHandleTableFreeWithCtx },
355	{ "RTHandleTableLookupWithCtx", (void *)RTHandleTableLookupWithCtx },
356	{ "RTNetIPv4AddDataChecksum", (void *)RTNetIPv4AddDataChecksum },
357	{ "RTNetIPv4AddTCPChecksum", (void *)RTNetIPv4AddTCPChecksum },
358	{ "RTNetIPv4AddUDPChecksum", (void *)RTNetIPv4AddUDPChecksum },
359	{ "RTNetIPv4FinalizeChecksum", (void *)RTNetIPv4FinalizeChecksum },
360	{ "RTNetIPv4HdrChecksum", (void *)RTNetIPv4HdrChecksum },
361	{ "RTNetIPv4IsDHCPValid", (void *)RTNetIPv4IsDHCPValid },
362	{ "RTNetIPv4IsHdrValid", (void *)RTNetIPv4IsHdrValid },
363	{ "RTNetIPv4IsTCPSizeValid", (void *)RTNetIPv4IsTCPSizeValid },
364	{ "RTNetIPv4IsTCPValid", (void *)RTNetIPv4IsTCPValid },
365	{ "RTNetIPv4IsUDPSizeValid", (void *)RTNetIPv4IsUDPSizeValid },
366	{ "RTNetIPv4IsUDPValid", (void *)RTNetIPv4IsUDPValid },
367	{ "RTNetIPv4PseudoChecksum", (void *)RTNetIPv4PseudoChecksum },
368	{ "RTNetIPv4PseudoChecksumBits", (void *)RTNetIPv4PseudoChecksumBits },
369	{ "RTNetIPv4TCPChecksum", (void *)RTNetIPv4TCPChecksum },
370	{ "RTNetIPv4UDPChecksum", (void *)RTNetIPv4UDPChecksum },
371	{ "RTNetIPv6PseudoChecksum", (void *)RTNetIPv6PseudoChecksum },
372	{ "RTNetIPv6PseudoChecksumBits", (void *)RTNetIPv6PseudoChecksumBits },
373	{ "RTNetIPv6PseudoChecksumEx", (void *)RTNetIPv6PseudoChecksumEx },
374	{ "RTNetTCPChecksum", (void *)RTNetTCPChecksum },
375	{ "RTNetUDPChecksum", (void *)RTNetUDPChecksum },
376	{ "RTCrc32", (void *)RTCrc32 },
377	{ "RTCrc32Finish", (void *)RTCrc32Finish },
378	{ "RTCrc32Process", (void *)RTCrc32Process },
379	{ "RTCrc32Start", (void *)RTCrc32Start },
380	};
381
382	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
383	/**
384	* Drag in the rest of IRPT since we share it with the
385	* rest of the kernel modules on darwin.
386	*/
387	PFNRT g_apfnVBoxDrvIPRTDeps[] =
388	{
389	/* VBoxNetFlt */
390	(PFNRT)RTErrConvertFromErrno,
391	(PFNRT)RTUuidCompare,
392	(PFNRT)RTUuidCompareStr,
393	(PFNRT)RTUuidFromStr,
394	(PFNRT)RTStrDupTag,
395	(PFNRT)RTStrFree,
396	/* VBoxNetAdp */
397	(PFNRT)RTRandBytes,
398	/* VBoxUSB */
399	(PFNRT)RTPathStripFilename,
400	NULL
401	};
402	#endif /* RT_OS_DARWIN \|\| RT_OS_SOLARIS \|\| RT_OS_SOLARIS */
403
404
405	/**
406	* Initializes the device extentsion structure.
407	*
408	* @returns IPRT status code.
409	* @param pDevExt The device extension to initialize.
410	* @param cbSession The size of the session structure. The size of
411	* SUPDRVSESSION may be smaller when SUPDRV_AGNOSTIC is
412	* defined because we're skipping the OS specific members
413	* then.
414	*/
415	int VBOXCALL supdrvInitDevExt(PSUPDRVDEVEXT pDevExt, size_t cbSession)
416	{
417	int rc;
418
419	#ifdef SUPDRV_WITH_RELEASE_LOGGER
420	/*
421	* Create the release log.
422	*/
423	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
424	PRTLOGGER pRelLogger;
425	rc = RTLogCreate(&pRelLogger, 0 /* fFlags */, "all",
426	"VBOX_RELEASE_LOG", RT_ELEMENTS(s_apszGroups), s_apszGroups,
427	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER, NULL);
428	if (RT_SUCCESS(rc))
429	RTLogRelSetDefaultInstance(pRelLogger);
430	/** @todo Add native hook for getting logger config parameters and setting
431	* them. On linux we should use the module parameter stuff... */
432	#endif
433
434	/*
435	* Initialize it.
436	*/
437	memset(pDevExt, 0, sizeof(*pDevExt));
438	rc = RTSpinlockCreate(&pDevExt->Spinlock);
439	if (RT_SUCCESS(rc))
440	{
441	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
442	rc = RTSemMutexCreate(&pDevExt->mtxLdr);
443	#else
444	rc = RTSemFastMutexCreate(&pDevExt->mtxLdr);
445	#endif
446	if (RT_SUCCESS(rc))
447	{
448	rc = RTSemFastMutexCreate(&pDevExt->mtxComponentFactory);
449	if (RT_SUCCESS(rc))
450	{
451	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
452	rc = RTSemMutexCreate(&pDevExt->mtxGip);
453	#else
454	rc = RTSemFastMutexCreate(&pDevExt->mtxGip);
455	#endif
456	if (RT_SUCCESS(rc))
457	{
458	rc = supdrvGipCreate(pDevExt);
459	if (RT_SUCCESS(rc))
460	{
461	pDevExt->u32Cookie = BIRD; /** @todo make this random? */
462	pDevExt->cbSession = (uint32_t)cbSession;
463
464	/*
465	* Fixup the absolute symbols.
466	*
467	* Because of the table indexing assumptions we'll have a little #ifdef orgy
468	* here rather than distributing this to OS specific files. At least for now.
469	*/
470	#ifdef RT_OS_DARWIN
471	# if ARCH_BITS == 32
472	if (SUPR0GetPagingMode() >= SUPPAGINGMODE_AMD64)
473	{
474	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
475	g_aFunctions[1].pfn = (void )0x80; / SUPR0Abs64bitKernelCS - KERNEL64_CS, seg.h */
476	g_aFunctions[2].pfn = (void )0x88; / SUPR0Abs64bitKernelSS - KERNEL64_SS, seg.h */
477	g_aFunctions[3].pfn = (void )0x88; / SUPR0Abs64bitKernelDS - KERNEL64_SS, seg.h */
478	}
479	else
480	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
481	g_aFunctions[4].pfn = (void )0x08; / SUPR0AbsKernelCS - KERNEL_CS, seg.h */
482	g_aFunctions[5].pfn = (void )0x10; / SUPR0AbsKernelSS - KERNEL_DS, seg.h */
483	g_aFunctions[6].pfn = (void )0x10; / SUPR0AbsKernelDS - KERNEL_DS, seg.h */
484	g_aFunctions[7].pfn = (void )0x10; / SUPR0AbsKernelES - KERNEL_DS, seg.h */
485	g_aFunctions[8].pfn = (void )0x10; / SUPR0AbsKernelFS - KERNEL_DS, seg.h */
486	g_aFunctions[9].pfn = (void )0x48; / SUPR0AbsKernelGS - CPU_DATA_GS, seg.h */
487	# else /* 64-bit darwin: */
488	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
489	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
490	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
491	g_aFunctions[3].pfn = (void )0; / SUPR0Abs64bitKernelDS */
492	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
493	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
494	g_aFunctions[6].pfn = (void )0; / SUPR0AbsKernelDS */
495	g_aFunctions[7].pfn = (void )0; / SUPR0AbsKernelES */
496	g_aFunctions[8].pfn = (void )0; / SUPR0AbsKernelFS */
497	g_aFunctions[9].pfn = (void )0; / SUPR0AbsKernelGS */
498
499	# endif
500	#else /* !RT_OS_DARWIN */
501	# if ARCH_BITS == 64
502	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
503	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
504	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
505	g_aFunctions[3].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0Abs64bitKernelDS */
506	# else
507	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
508	# endif
509	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
510	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
511	g_aFunctions[6].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0AbsKernelDS */
512	g_aFunctions[7].pfn = (void )(uintptr_t)ASMGetES(); / SUPR0AbsKernelES */
513	g_aFunctions[8].pfn = (void )(uintptr_t)ASMGetFS(); / SUPR0AbsKernelFS */
514	g_aFunctions[9].pfn = (void )(uintptr_t)ASMGetGS(); / SUPR0AbsKernelGS */
515	#endif /* !RT_OS_DARWIN */
516	return VINF_SUCCESS;
517	}
518
519	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
520	RTSemMutexDestroy(pDevExt->mtxGip);
521	pDevExt->mtxGip = NIL_RTSEMMUTEX;
522	#else
523	RTSemFastMutexDestroy(pDevExt->mtxGip);
524	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
525	#endif
526	}
527	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
528	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
529	}
530	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
531	RTSemMutexDestroy(pDevExt->mtxLdr);
532	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
533	#else
534	RTSemFastMutexDestroy(pDevExt->mtxLdr);
535	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
536	#endif
537	}
538	RTSpinlockDestroy(pDevExt->Spinlock);
539	pDevExt->Spinlock = NIL_RTSPINLOCK;
540	}
541	#ifdef SUPDRV_WITH_RELEASE_LOGGER
542	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
543	RTLogDestroy(RTLogSetDefaultInstance(NULL));
544	#endif
545
546	return rc;
547	}
548
549
550	/**
551	* Delete the device extension (e.g. cleanup members).
552	*
553	* @param pDevExt The device extension to delete.
554	*/
555	void VBOXCALL supdrvDeleteDevExt(PSUPDRVDEVEXT pDevExt)
556	{
557	PSUPDRVOBJ pObj;
558	PSUPDRVUSAGE pUsage;
559
560	/*
561	* Kill mutexes and spinlocks.
562	*/
563	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
564	RTSemMutexDestroy(pDevExt->mtxGip);
565	pDevExt->mtxGip = NIL_RTSEMMUTEX;
566	#else
567	RTSemFastMutexDestroy(pDevExt->mtxGip);
568	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
569	#endif
570	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
571	RTSemMutexDestroy(pDevExt->mtxLdr);
572	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
573	#else
574	RTSemFastMutexDestroy(pDevExt->mtxLdr);
575	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
576	#endif
577	RTSpinlockDestroy(pDevExt->Spinlock);
578	pDevExt->Spinlock = NIL_RTSPINLOCK;
579	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
580	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
581
582	/*
583	* Free lists.
584	*/
585	/* objects. */
586	pObj = pDevExt->pObjs;
587	Assert(!pObj); /* (can trigger on forced unloads) */
588	pDevExt->pObjs = NULL;
589	while (pObj)
590	{
591	void *pvFree = pObj;
592	pObj = pObj->pNext;
593	RTMemFree(pvFree);
594	}
595
596	/* usage records. */
597	pUsage = pDevExt->pUsageFree;
598	pDevExt->pUsageFree = NULL;
599	while (pUsage)
600	{
601	void *pvFree = pUsage;
602	pUsage = pUsage->pNext;
603	RTMemFree(pvFree);
604	}
605
606	/* kill the GIP. */
607	supdrvGipDestroy(pDevExt);
608
609	#ifdef SUPDRV_WITH_RELEASE_LOGGER
610	/* destroy the loggers. */
611	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
612	RTLogDestroy(RTLogSetDefaultInstance(NULL));
613	#endif
614	}
615
616
617	/**
618	* Create session.
619	*
620	* @returns IPRT status code.
621	* @param pDevExt Device extension.
622	* @param fUser Flag indicating whether this is a user or kernel session.
623	* @param ppSession Where to store the pointer to the session data.
624	*/
625	int VBOXCALL supdrvCreateSession(PSUPDRVDEVEXT pDevExt, bool fUser, PSUPDRVSESSION *ppSession)
626	{
627	/*
628	* Allocate memory for the session data.
629	*/
630	int rc;
631	PSUPDRVSESSION pSession = *ppSession = (PSUPDRVSESSION)RTMemAllocZ(pDevExt->cbSession);
632	if (pSession)
633	{
634	/* Initialize session data. */
635	rc = RTSpinlockCreate(&pSession->Spinlock);
636	if (!rc)
637	{
638	rc = RTHandleTableCreateEx(&pSession->hHandleTable,
639	RTHANDLETABLE_FLAGS_LOCKED \| RTHANDLETABLE_FLAGS_CONTEXT,
640	1 /uBase/, 32768 /cMax/, supdrvSessionObjHandleRetain, pSession);
641	if (RT_SUCCESS(rc))
642	{
643	Assert(pSession->Spinlock != NIL_RTSPINLOCK);
644	pSession->pDevExt = pDevExt;
645	pSession->u32Cookie = BIRD_INV;
646	/*pSession->pLdrUsage = NULL;
647	pSession->pVM = NULL;
648	pSession->pUsage = NULL;
649	pSession->pGip = NULL;
650	pSession->fGipReferenced = false;
651	pSession->Bundle.cUsed = 0; */
652	pSession->Uid = NIL_RTUID;
653	pSession->Gid = NIL_RTGID;
654	if (fUser)
655	{
656	pSession->Process = RTProcSelf();
657	pSession->R0Process = RTR0ProcHandleSelf();
658	}
659	else
660	{
661	pSession->Process = NIL_RTPROCESS;
662	pSession->R0Process = NIL_RTR0PROCESS;
663	}
664
665	VBOXDRV_SUPDRV_SESSION_CREATE(pSession, fUser);
666	LogFlow(("Created session %p initial cookie=%#x\n", pSession, pSession->u32Cookie));
667	return VINF_SUCCESS;
668	}
669
670	RTSpinlockDestroy(pSession->Spinlock);
671	}
672	RTMemFree(pSession);
673	*ppSession = NULL;
674	Log(("Failed to create spinlock, rc=%d!\n", rc));
675	}
676	else
677	rc = VERR_NO_MEMORY;
678
679	return rc;
680	}
681
682
683	/**
684	* Shared code for cleaning up a session.
685	*
686	* @param pDevExt Device extension.
687	* @param pSession Session data.
688	* This data will be freed by this routine.
689	*/
690	void VBOXCALL supdrvCloseSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
691	{
692	VBOXDRV_SUPDRV_SESSION_CLOSE(pSession);
693
694	/*
695	* Cleanup the session first.
696	*/
697	supdrvCleanupSession(pDevExt, pSession);
698
699	/*
700	* Free the rest of the session stuff.
701	*/
702	RTSpinlockDestroy(pSession->Spinlock);
703	pSession->Spinlock = NIL_RTSPINLOCK;
704	pSession->pDevExt = NULL;
705	RTMemFree(pSession);
706	LogFlow(("supdrvCloseSession: returns\n"));
707	}
708
709
710	/**
711	* Shared code for cleaning up a session (but not quite freeing it).
712	*
713	* This is primarily intended for MAC OS X where we have to clean up the memory
714	* stuff before the file handle is closed.
715	*
716	* @param pDevExt Device extension.
717	* @param pSession Session data.
718	* This data will be freed by this routine.
719	*/
720	void VBOXCALL supdrvCleanupSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
721	{
722	int rc;
723	PSUPDRVBUNDLE pBundle;
724	LogFlow(("supdrvCleanupSession: pSession=%p\n", pSession));
725
726	/*
727	* Remove logger instances related to this session.
728	*/
729	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pSession);
730
731	/*
732	* Destroy the handle table.
733	*/
734	rc = RTHandleTableDestroy(pSession->hHandleTable, supdrvSessionObjHandleDelete, pSession);
735	AssertRC(rc);
736	pSession->hHandleTable = NIL_RTHANDLETABLE;
737
738	/*
739	* Release object references made in this session.
740	* In theory there should be noone racing us in this session.
741	*/
742	Log2(("release objects - start\n"));
743	if (pSession->pUsage)
744	{
745	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
746	PSUPDRVUSAGE pUsage;
747	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
748
749	while ((pUsage = pSession->pUsage) != NULL)
750	{
751	PSUPDRVOBJ pObj = pUsage->pObj;
752	pSession->pUsage = pUsage->pNext;
753
754	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
755	if (pUsage->cUsage < pObj->cUsage)
756	{
757	pObj->cUsage -= pUsage->cUsage;
758	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
759	}
760	else
761	{
762	/* Destroy the object and free the record. */
763	if (pDevExt->pObjs == pObj)
764	pDevExt->pObjs = pObj->pNext;
765	else
766	{
767	PSUPDRVOBJ pObjPrev;
768	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
769	if (pObjPrev->pNext == pObj)
770	{
771	pObjPrev->pNext = pObj->pNext;
772	break;
773	}
774	Assert(pObjPrev);
775	}
776	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
777
778	Log(("supdrvCleanupSession: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
779	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
780	if (pObj->pfnDestructor)
781	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
782	RTMemFree(pObj);
783	}
784
785	/* free it and continue. */
786	RTMemFree(pUsage);
787
788	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
789	}
790
791	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
792	AssertMsg(!pSession->pUsage, ("Some buster reregistered an object during desturction!\n"));
793	}
794	Log2(("release objects - done\n"));
795
796	/*
797	* Release memory allocated in the session.
798	*
799	* We do not serialize this as we assume that the application will
800	* not allocated memory while closing the file handle object.
801	*/
802	Log2(("freeing memory:\n"));
803	pBundle = &pSession->Bundle;
804	while (pBundle)
805	{
806	PSUPDRVBUNDLE pToFree;
807	unsigned i;
808
809	/*
810	* Check and unlock all entries in the bundle.
811	*/
812	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
813	{
814	if (pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ)
815	{
816	Log2(("eType=%d pvR0=%p pvR3=%p cb=%ld\n", pBundle->aMem[i].eType, RTR0MemObjAddress(pBundle->aMem[i].MemObj),
817	(void *)RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3), (long)RTR0MemObjSize(pBundle->aMem[i].MemObj)));
818	if (pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ)
819	{
820	rc = RTR0MemObjFree(pBundle->aMem[i].MapObjR3, false);
821	AssertRC(rc); /** @todo figure out how to handle this. */
822	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
823	}
824	rc = RTR0MemObjFree(pBundle->aMem[i].MemObj, true /* fFreeMappings */);
825	AssertRC(rc); /** @todo figure out how to handle this. */
826	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
827	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
828	}
829	}
830
831	/*
832	* Advance and free previous bundle.
833	*/
834	pToFree = pBundle;
835	pBundle = pBundle->pNext;
836
837	pToFree->pNext = NULL;
838	pToFree->cUsed = 0;
839	if (pToFree != &pSession->Bundle)
840	RTMemFree(pToFree);
841	}
842	Log2(("freeing memory - done\n"));
843
844	/*
845	* Deregister component factories.
846	*/
847	RTSemFastMutexRequest(pDevExt->mtxComponentFactory);
848	Log2(("deregistering component factories:\n"));
849	if (pDevExt->pComponentFactoryHead)
850	{
851	PSUPDRVFACTORYREG pPrev = NULL;
852	PSUPDRVFACTORYREG pCur = pDevExt->pComponentFactoryHead;
853	while (pCur)
854	{
855	if (pCur->pSession == pSession)
856	{
857	/* unlink it */
858	PSUPDRVFACTORYREG pNext = pCur->pNext;
859	if (pPrev)
860	pPrev->pNext = pNext;
861	else
862	pDevExt->pComponentFactoryHead = pNext;
863
864	/* free it */
865	pCur->pNext = NULL;
866	pCur->pSession = NULL;
867	pCur->pFactory = NULL;
868	RTMemFree(pCur);
869
870	/* next */
871	pCur = pNext;
872	}
873	else
874	{
875	/* next */
876	pPrev = pCur;
877	pCur = pCur->pNext;
878	}
879	}
880	}
881	RTSemFastMutexRelease(pDevExt->mtxComponentFactory);
882	Log2(("deregistering component factories - done\n"));
883
884	/*
885	* Loaded images needs to be dereferenced and possibly freed up.
886	*/
887	supdrvLdrLock(pDevExt);
888	Log2(("freeing images:\n"));
889	if (pSession->pLdrUsage)
890	{
891	PSUPDRVLDRUSAGE pUsage = pSession->pLdrUsage;
892	pSession->pLdrUsage = NULL;
893	while (pUsage)
894	{
895	void *pvFree = pUsage;
896	PSUPDRVLDRIMAGE pImage = pUsage->pImage;
897	if (pImage->cUsage > pUsage->cUsage)
898	pImage->cUsage -= pUsage->cUsage;
899	else
900	supdrvLdrFree(pDevExt, pImage);
901	pUsage->pImage = NULL;
902	pUsage = pUsage->pNext;
903	RTMemFree(pvFree);
904	}
905	}
906	supdrvLdrUnlock(pDevExt);
907	Log2(("freeing images - done\n"));
908
909	/*
910	* Unmap the GIP.
911	*/
912	Log2(("umapping GIP:\n"));
913	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
914	{
915	SUPR0GipUnmap(pSession);
916	pSession->fGipReferenced = 0;
917	}
918	Log2(("umapping GIP - done\n"));
919	}
920
921
922	/**
923	* RTHandleTableDestroy callback used by supdrvCleanupSession.
924	*
925	* @returns IPRT status code, see SUPR0ObjAddRef.
926	* @param hHandleTable The handle table handle. Ignored.
927	* @param pvObj The object pointer.
928	* @param pvCtx Context, the handle type. Ignored.
929	* @param pvUser Session pointer.
930	*/
931	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser)
932	{
933	NOREF(pvCtx);
934	NOREF(hHandleTable);
935	return SUPR0ObjAddRefEx(pvObj, (PSUPDRVSESSION)pvUser, true /fNoBlocking/);
936	}
937
938
939	/**
940	* RTHandleTableDestroy callback used by supdrvCleanupSession.
941	*
942	* @param hHandleTable The handle table handle. Ignored.
943	* @param h The handle value. Ignored.
944	* @param pvObj The object pointer.
945	* @param pvCtx Context, the handle type. Ignored.
946	* @param pvUser Session pointer.
947	*/
948	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser)
949	{
950	NOREF(pvCtx);
951	NOREF(h);
952	NOREF(hHandleTable);
953	SUPR0ObjRelease(pvObj, (PSUPDRVSESSION)pvUser);
954	}
955
956
957	/**
958	* Fast path I/O Control worker.
959	*
960	* @returns VBox status code that should be passed down to ring-3 unchanged.
961	* @param uIOCtl Function number.
962	* @param idCpu VMCPU id.
963	* @param pDevExt Device extention.
964	* @param pSession Session data.
965	*/
966	int VBOXCALL supdrvIOCtlFast(uintptr_t uIOCtl, VMCPUID idCpu, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
967	{
968	/*
969	* We check the two prereqs after doing this only to allow the compiler to optimize things better.
970	*/
971	if (RT_LIKELY(pSession->pVM && pDevExt->pfnVMMR0EntryFast))
972	{
973	switch (uIOCtl)
974	{
975	case SUP_IOCTL_FAST_DO_RAW_RUN:
976	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_RAW_RUN);
977	break;
978	case SUP_IOCTL_FAST_DO_HWACC_RUN:
979	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_HWACC_RUN);
980	break;
981	case SUP_IOCTL_FAST_DO_NOP:
982	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_NOP);
983	break;
984	default:
985	return VERR_INTERNAL_ERROR;
986	}
987	return VINF_SUCCESS;
988	}
989	return VERR_INTERNAL_ERROR;
990	}
991
992
993	/**
994	* Helper for supdrvIOCtl. Check if pszStr contains any character of pszChars.
995	* We would use strpbrk here if this function would be contained in the RedHat kABI white
996	* list, see http://www.kerneldrivers.org/RHEL5.
997	*
998	* @returns 1 if pszStr does contain any character of pszChars, 0 otherwise.
999	* @param pszStr String to check
1000	* @param pszChars Character set
1001	*/
1002	static int supdrvCheckInvalidChar(const char pszStr, const char pszChars)
1003	{
1004	int chCur;
1005	while ((chCur = *pszStr++) != '\0')
1006	{
1007	int ch;
1008	const char *psz = pszChars;
1009	while ((ch = *psz++) != '\0')
1010	if (ch == chCur)
1011	return 1;
1012
1013	}
1014	return 0;
1015	}
1016
1017
1018	/**
1019	* I/O Control worker.
1020	*
1021	* @returns IPRT status code.
1022	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1023	*
1024	* @param uIOCtl Function number.
1025	* @param pDevExt Device extention.
1026	* @param pSession Session data.
1027	* @param pReqHdr The request header.
1028	*/
1029	int VBOXCALL supdrvIOCtl(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1030	{
1031	VBOXDRV_SUPDRV_IOCTL_ENTRY(pSession, uIOCtl, pReqHdr);
1032
1033	/*
1034	* Validate the request.
1035	*/
1036	/* this first check could probably be omitted as its also done by the OS specific code... */
1037	if (RT_UNLIKELY( (pReqHdr->fFlags & SUPREQHDR_FLAGS_MAGIC_MASK) != SUPREQHDR_FLAGS_MAGIC
1038	\|\| pReqHdr->cbIn < sizeof(*pReqHdr)
1039	\|\| pReqHdr->cbOut < sizeof(*pReqHdr)))
1040	{
1041	OSDBGPRINT(("vboxdrv: Bad ioctl request header; cbIn=%#lx cbOut=%#lx fFlags=%#lx\n",
1042	(long)pReqHdr->cbIn, (long)pReqHdr->cbOut, (long)pReqHdr->fFlags));
1043	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1044	return VERR_INVALID_PARAMETER;
1045	}
1046	if (RT_UNLIKELY(uIOCtl == SUP_IOCTL_COOKIE))
1047	{
1048	if (pReqHdr->u32Cookie != SUPCOOKIE_INITIAL_COOKIE)
1049	{
1050	OSDBGPRINT(("SUP_IOCTL_COOKIE: bad cookie %#lx\n", (long)pReqHdr->u32Cookie));
1051	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1052	return VERR_INVALID_PARAMETER;
1053	}
1054	}
1055	else if (RT_UNLIKELY( pReqHdr->u32Cookie != pDevExt->u32Cookie
1056	\|\| pReqHdr->u32SessionCookie != pSession->u32Cookie))
1057	{
1058	OSDBGPRINT(("vboxdrv: bad cookie %#lx / %#lx.\n", (long)pReqHdr->u32Cookie, (long)pReqHdr->u32SessionCookie));
1059	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1060	return VERR_INVALID_PARAMETER;
1061	}
1062
1063	/*
1064	* Validation macros
1065	*/
1066	#define REQ_CHECK_SIZES_EX(Name, cbInExpect, cbOutExpect) \
1067	do { \
1068	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect) \|\| pReqHdr->cbOut != (cbOutExpect))) \
1069	{ \
1070	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld. cbOut=%ld expected %ld.\n", \
1071	(long)pReq->Hdr.cbIn, (long)(cbInExpect), (long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
1072	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1073	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1074	} \
1075	} while (0)
1076
1077	#define REQ_CHECK_SIZES(Name) REQ_CHECK_SIZES_EX(Name, Name ## _SIZE_IN, Name ## _SIZE_OUT)
1078
1079	#define REQ_CHECK_SIZE_IN(Name, cbInExpect) \
1080	do { \
1081	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect))) \
1082	{ \
1083	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld.\n", \
1084	(long)pReq->Hdr.cbIn, (long)(cbInExpect))); \
1085	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1086	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1087	} \
1088	} while (0)
1089
1090	#define REQ_CHECK_SIZE_OUT(Name, cbOutExpect) \
1091	do { \
1092	if (RT_UNLIKELY(pReqHdr->cbOut != (cbOutExpect))) \
1093	{ \
1094	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbOut=%ld expected %ld.\n", \
1095	(long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
1096	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1097	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1098	} \
1099	} while (0)
1100
1101	#define REQ_CHECK_EXPR(Name, expr) \
1102	do { \
1103	if (RT_UNLIKELY(!(expr))) \
1104	{ \
1105	OSDBGPRINT(( #Name ": %s\n", #expr)); \
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_FMT(expr, fmt) \
1112	do { \
1113	if (RT_UNLIKELY(!(expr))) \
1114	{ \
1115	OSDBGPRINT( fmt ); \
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
1122	/*
1123	* The switch.
1124	*/
1125	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
1126	{
1127	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
1128	{
1129	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
1130	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
1131	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
1132	{
1133	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
1134	pReq->Hdr.rc = VERR_INVALID_MAGIC;
1135	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, VERR_INVALID_MAGIC);
1136	return 0;
1137	}
1138
1139	#if 0
1140	/*
1141	* Call out to the OS specific code and let it do permission checks on the
1142	* client process.
1143	*/
1144	if (!supdrvOSValidateClientProcess(pDevExt, pSession))
1145	{
1146	pReq->u.Out.u32Cookie = 0xffffffff;
1147	pReq->u.Out.u32SessionCookie = 0xffffffff;
1148	pReq->u.Out.u32SessionVersion = 0xffffffff;
1149	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1150	pReq->u.Out.pSession = NULL;
1151	pReq->u.Out.cFunctions = 0;
1152	pReq->Hdr.rc = VERR_PERMISSION_DENIED;
1153	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, VERR_PERMISSION_DENIED);
1154	return 0;
1155	}
1156	#endif
1157
1158	/*
1159	* Match the version.
1160	* The current logic is very simple, match the major interface version.
1161	*/
1162	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
1163	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
1164	{
1165	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1166	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
1167	pReq->u.Out.u32Cookie = 0xffffffff;
1168	pReq->u.Out.u32SessionCookie = 0xffffffff;
1169	pReq->u.Out.u32SessionVersion = 0xffffffff;
1170	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1171	pReq->u.Out.pSession = NULL;
1172	pReq->u.Out.cFunctions = 0;
1173	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1174	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1175	return 0;
1176	}
1177
1178	/*
1179	* Fill in return data and be gone.
1180	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
1181	* u32SessionVersion <= u32ReqVersion!
1182	*/
1183	/** @todo Somehow validate the client and negotiate a secure cookie... */
1184	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
1185	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
1186	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
1187	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1188	pReq->u.Out.pSession = pSession;
1189	pReq->u.Out.cFunctions = sizeof(g_aFunctions) / sizeof(g_aFunctions[0]);
1190	pReq->Hdr.rc = VINF_SUCCESS;
1191	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1192	return 0;
1193	}
1194
1195	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_QUERY_FUNCS(0)):
1196	{
1197	/* validate */
1198	PSUPQUERYFUNCS pReq = (PSUPQUERYFUNCS)pReqHdr;
1199	REQ_CHECK_SIZES_EX(SUP_IOCTL_QUERY_FUNCS, SUP_IOCTL_QUERY_FUNCS_SIZE_IN, SUP_IOCTL_QUERY_FUNCS_SIZE_OUT(RT_ELEMENTS(g_aFunctions)));
1200
1201	/* execute */
1202	pReq->u.Out.cFunctions = RT_ELEMENTS(g_aFunctions);
1203	memcpy(&pReq->u.Out.aFunctions[0], g_aFunctions, sizeof(g_aFunctions));
1204	pReq->Hdr.rc = VINF_SUCCESS;
1205	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1206	return 0;
1207	}
1208
1209	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_LOCK):
1210	{
1211	/* validate */
1212	PSUPPAGELOCK pReq = (PSUPPAGELOCK)pReqHdr;
1213	REQ_CHECK_SIZE_IN(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_IN);
1214	REQ_CHECK_SIZE_OUT(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_OUT(pReq->u.In.cPages));
1215	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.cPages > 0);
1216	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.pvR3 >= PAGE_SIZE);
1217
1218	/* execute */
1219	pReq->Hdr.rc = SUPR0LockMem(pSession, pReq->u.In.pvR3, pReq->u.In.cPages, &pReq->u.Out.aPages[0]);
1220	if (RT_FAILURE(pReq->Hdr.rc))
1221	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1222	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1223	return 0;
1224	}
1225
1226	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_UNLOCK):
1227	{
1228	/* validate */
1229	PSUPPAGEUNLOCK pReq = (PSUPPAGEUNLOCK)pReqHdr;
1230	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_UNLOCK);
1231
1232	/* execute */
1233	pReq->Hdr.rc = SUPR0UnlockMem(pSession, pReq->u.In.pvR3);
1234	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1235	return 0;
1236	}
1237
1238	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_ALLOC):
1239	{
1240	/* validate */
1241	PSUPCONTALLOC pReq = (PSUPCONTALLOC)pReqHdr;
1242	REQ_CHECK_SIZES(SUP_IOCTL_CONT_ALLOC);
1243
1244	/* execute */
1245	pReq->Hdr.rc = SUPR0ContAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.HCPhys);
1246	if (RT_FAILURE(pReq->Hdr.rc))
1247	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1248	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1249	return 0;
1250	}
1251
1252	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_FREE):
1253	{
1254	/* validate */
1255	PSUPCONTFREE pReq = (PSUPCONTFREE)pReqHdr;
1256	REQ_CHECK_SIZES(SUP_IOCTL_CONT_FREE);
1257
1258	/* execute */
1259	pReq->Hdr.rc = SUPR0ContFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1260	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1261	return 0;
1262	}
1263
1264	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_OPEN):
1265	{
1266	/* validate */
1267	PSUPLDROPEN pReq = (PSUPLDROPEN)pReqHdr;
1268	REQ_CHECK_SIZES(SUP_IOCTL_LDR_OPEN);
1269	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs > 0);
1270	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs < 16*_1M);
1271	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1272	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1273	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits < pReq->u.In.cbImageWithTabs);
1274	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.szName[0]);
1275	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1276	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, !supdrvCheckInvalidChar(pReq->u.In.szName, ";:()[]{}/\\\|&*%#@!~`\"'"));
1277	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szFilename, sizeof(pReq->u.In.szFilename)));
1278
1279	/* execute */
1280	pReq->Hdr.rc = supdrvIOCtl_LdrOpen(pDevExt, pSession, pReq);
1281	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1282	return 0;
1283	}
1284
1285	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOAD):
1286	{
1287	/* validate */
1288	PSUPLDRLOAD pReq = (PSUPLDRLOAD)pReqHdr;
1289	REQ_CHECK_EXPR(Name, pReq->Hdr.cbIn >= sizeof(*pReq));
1290	REQ_CHECK_SIZES_EX(SUP_IOCTL_LDR_LOAD, SUP_IOCTL_LDR_LOAD_SIZE_IN(pReq->u.In.cbImageWithTabs), SUP_IOCTL_LDR_LOAD_SIZE_OUT);
1291	REQ_CHECK_EXPR(SUP_IOCTL_LDR_LOAD, pReq->u.In.cSymbols <= 16384);
1292	REQ_CHECK_EXPR_FMT( !pReq->u.In.cSymbols
1293	\|\| ( pReq->u.In.offSymbols < pReq->u.In.cbImageWithTabs
1294	&& pReq->u.In.offSymbols + pReq->u.In.cSymbols * sizeof(SUPLDRSYM) <= pReq->u.In.cbImageWithTabs),
1295	("SUP_IOCTL_LDR_LOAD: offSymbols=%#lx cSymbols=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offSymbols,
1296	(long)pReq->u.In.cSymbols, (long)pReq->u.In.cbImageWithTabs));
1297	REQ_CHECK_EXPR_FMT( !pReq->u.In.cbStrTab
1298	\|\| ( pReq->u.In.offStrTab < pReq->u.In.cbImageWithTabs
1299	&& pReq->u.In.offStrTab + pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs
1300	&& pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs),
1301	("SUP_IOCTL_LDR_LOAD: offStrTab=%#lx cbStrTab=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offStrTab,
1302	(long)pReq->u.In.cbStrTab, (long)pReq->u.In.cbImageWithTabs));
1303
1304	if (pReq->u.In.cSymbols)
1305	{
1306	uint32_t i;
1307	PSUPLDRSYM paSyms = (PSUPLDRSYM)&pReq->u.In.abImage[pReq->u.In.offSymbols];
1308	for (i = 0; i < pReq->u.In.cSymbols; i++)
1309	{
1310	REQ_CHECK_EXPR_FMT(paSyms[i].offSymbol < pReq->u.In.cbImageWithTabs,
1311	("SUP_IOCTL_LDR_LOAD: sym #%ld: symb off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offSymbol, (long)pReq->u.In.cbImageWithTabs));
1312	REQ_CHECK_EXPR_FMT(paSyms[i].offName < pReq->u.In.cbStrTab,
1313	("SUP_IOCTL_LDR_LOAD: sym #%ld: name off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1314	REQ_CHECK_EXPR_FMT(RTStrEnd((char const *)&pReq->u.In.abImage[pReq->u.In.offStrTab + paSyms[i].offName],
1315	pReq->u.In.cbStrTab - paSyms[i].offName),
1316	("SUP_IOCTL_LDR_LOAD: sym #%ld: unterminated name! (%#lx / %#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1317	}
1318	}
1319
1320	/* execute */
1321	pReq->Hdr.rc = supdrvIOCtl_LdrLoad(pDevExt, pSession, pReq);
1322	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1323	return 0;
1324	}
1325
1326	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_FREE):
1327	{
1328	/* validate */
1329	PSUPLDRFREE pReq = (PSUPLDRFREE)pReqHdr;
1330	REQ_CHECK_SIZES(SUP_IOCTL_LDR_FREE);
1331
1332	/* execute */
1333	pReq->Hdr.rc = supdrvIOCtl_LdrFree(pDevExt, pSession, pReq);
1334	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1335	return 0;
1336	}
1337
1338	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_GET_SYMBOL):
1339	{
1340	/* validate */
1341	PSUPLDRGETSYMBOL pReq = (PSUPLDRGETSYMBOL)pReqHdr;
1342	REQ_CHECK_SIZES(SUP_IOCTL_LDR_GET_SYMBOL);
1343	REQ_CHECK_EXPR(SUP_IOCTL_LDR_GET_SYMBOL, RTStrEnd(pReq->u.In.szSymbol, sizeof(pReq->u.In.szSymbol)));
1344
1345	/* execute */
1346	pReq->Hdr.rc = supdrvIOCtl_LdrGetSymbol(pDevExt, pSession, pReq);
1347	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1348	return 0;
1349	}
1350
1351	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0(0)):
1352	{
1353	/* validate */
1354	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1355	Log4(("SUP_IOCTL_CALL_VMMR0: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1356	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1357
1358	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_VMMR0_SIZE(0))
1359	{
1360	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(0), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(0));
1361
1362	/* execute */
1363	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1364	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg, pSession);
1365	else
1366	pReq->Hdr.rc = VERR_WRONG_ORDER;
1367	}
1368	else
1369	{
1370	PSUPVMMR0REQHDR pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1371	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR)),
1372	("SUP_IOCTL_CALL_VMMR0: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR))));
1373	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1374	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(pVMMReq->cbReq));
1375
1376	/* execute */
1377	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1378	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1379	else
1380	pReq->Hdr.rc = VERR_WRONG_ORDER;
1381	}
1382
1383	if ( RT_FAILURE(pReq->Hdr.rc)
1384	&& pReq->Hdr.rc != VERR_INTERRUPTED
1385	&& pReq->Hdr.rc != VERR_TIMEOUT)
1386	Log(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1387	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1388	else
1389	Log4(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1390	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1391	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1392	return 0;
1393	}
1394
1395	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_PAGING_MODE):
1396	{
1397	/* validate */
1398	PSUPGETPAGINGMODE pReq = (PSUPGETPAGINGMODE)pReqHdr;
1399	REQ_CHECK_SIZES(SUP_IOCTL_GET_PAGING_MODE);
1400
1401	/* execute */
1402	pReq->Hdr.rc = VINF_SUCCESS;
1403	pReq->u.Out.enmMode = SUPR0GetPagingMode();
1404	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1405	return 0;
1406	}
1407
1408	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_ALLOC):
1409	{
1410	/* validate */
1411	PSUPLOWALLOC pReq = (PSUPLOWALLOC)pReqHdr;
1412	REQ_CHECK_EXPR(SUP_IOCTL_LOW_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_LOW_ALLOC_SIZE_IN);
1413	REQ_CHECK_SIZES_EX(SUP_IOCTL_LOW_ALLOC, SUP_IOCTL_LOW_ALLOC_SIZE_IN, SUP_IOCTL_LOW_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1414
1415	/* execute */
1416	pReq->Hdr.rc = SUPR0LowAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1417	if (RT_FAILURE(pReq->Hdr.rc))
1418	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1419	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1420	return 0;
1421	}
1422
1423	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_FREE):
1424	{
1425	/* validate */
1426	PSUPLOWFREE pReq = (PSUPLOWFREE)pReqHdr;
1427	REQ_CHECK_SIZES(SUP_IOCTL_LOW_FREE);
1428
1429	/* execute */
1430	pReq->Hdr.rc = SUPR0LowFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1431	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1432	return 0;
1433	}
1434
1435	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_MAP):
1436	{
1437	/* validate */
1438	PSUPGIPMAP pReq = (PSUPGIPMAP)pReqHdr;
1439	REQ_CHECK_SIZES(SUP_IOCTL_GIP_MAP);
1440
1441	/* execute */
1442	pReq->Hdr.rc = SUPR0GipMap(pSession, &pReq->u.Out.pGipR3, &pReq->u.Out.HCPhysGip);
1443	if (RT_SUCCESS(pReq->Hdr.rc))
1444	pReq->u.Out.pGipR0 = pDevExt->pGip;
1445	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1446	return 0;
1447	}
1448
1449	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_UNMAP):
1450	{
1451	/* validate */
1452	PSUPGIPUNMAP pReq = (PSUPGIPUNMAP)pReqHdr;
1453	REQ_CHECK_SIZES(SUP_IOCTL_GIP_UNMAP);
1454
1455	/* execute */
1456	pReq->Hdr.rc = SUPR0GipUnmap(pSession);
1457	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1458	return 0;
1459	}
1460
1461	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SET_VM_FOR_FAST):
1462	{
1463	/* validate */
1464	PSUPSETVMFORFAST pReq = (PSUPSETVMFORFAST)pReqHdr;
1465	REQ_CHECK_SIZES(SUP_IOCTL_SET_VM_FOR_FAST);
1466	REQ_CHECK_EXPR_FMT( !pReq->u.In.pVMR0
1467	\|\| ( VALID_PTR(pReq->u.In.pVMR0)
1468	&& !((uintptr_t)pReq->u.In.pVMR0 & (PAGE_SIZE - 1))),
1469	("SUP_IOCTL_SET_VM_FOR_FAST: pVMR0=%p!\n", pReq->u.In.pVMR0));
1470	/* execute */
1471	pSession->pVM = pReq->u.In.pVMR0;
1472	pReq->Hdr.rc = VINF_SUCCESS;
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_PAGE_ALLOC_EX):
1478	{
1479	/* validate */
1480	PSUPPAGEALLOCEX pReq = (PSUPPAGEALLOCEX)pReqHdr;
1481	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC_EX, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_EX_SIZE_IN);
1482	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));
1483	REQ_CHECK_EXPR_FMT(pReq->u.In.fKernelMapping \|\| pReq->u.In.fUserMapping,
1484	("SUP_IOCTL_PAGE_ALLOC_EX: No mapping requested!\n"));
1485	REQ_CHECK_EXPR_FMT(pReq->u.In.fUserMapping,
1486	("SUP_IOCTL_PAGE_ALLOC_EX: Must have user mapping!\n"));
1487	REQ_CHECK_EXPR_FMT(!pReq->u.In.fReserved0 && !pReq->u.In.fReserved1,
1488	("SUP_IOCTL_PAGE_ALLOC_EX: fReserved0=%d fReserved1=%d\n", pReq->u.In.fReserved0, pReq->u.In.fReserved1));
1489
1490	/* execute */
1491	pReq->Hdr.rc = SUPR0PageAllocEx(pSession, pReq->u.In.cPages, 0 /* fFlags */,
1492	pReq->u.In.fUserMapping ? &pReq->u.Out.pvR3 : NULL,
1493	pReq->u.In.fKernelMapping ? &pReq->u.Out.pvR0 : NULL,
1494	&pReq->u.Out.aPages[0]);
1495	if (RT_FAILURE(pReq->Hdr.rc))
1496	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1497	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1498	return 0;
1499	}
1500
1501	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_MAP_KERNEL):
1502	{
1503	/* validate */
1504	PSUPPAGEMAPKERNEL pReq = (PSUPPAGEMAPKERNEL)pReqHdr;
1505	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_MAP_KERNEL);
1506	REQ_CHECK_EXPR_FMT(!pReq->u.In.fFlags, ("SUP_IOCTL_PAGE_MAP_KERNEL: fFlags=%#x! MBZ\n", pReq->u.In.fFlags));
1507	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_MAP_KERNEL: offSub=%#x\n", pReq->u.In.offSub));
1508	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1509	("SUP_IOCTL_PAGE_MAP_KERNEL: cbSub=%#x\n", pReq->u.In.cbSub));
1510
1511	/* execute */
1512	pReq->Hdr.rc = SUPR0PageMapKernel(pSession, pReq->u.In.pvR3, pReq->u.In.offSub, pReq->u.In.cbSub,
1513	pReq->u.In.fFlags, &pReq->u.Out.pvR0);
1514	if (RT_FAILURE(pReq->Hdr.rc))
1515	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1516	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1517	return 0;
1518	}
1519
1520	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_PROTECT):
1521	{
1522	/* validate */
1523	PSUPPAGEPROTECT pReq = (PSUPPAGEPROTECT)pReqHdr;
1524	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_PROTECT);
1525	REQ_CHECK_EXPR_FMT(!(pReq->u.In.fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)),
1526	("SUP_IOCTL_PAGE_PROTECT: fProt=%#x!\n", pReq->u.In.fProt));
1527	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_PROTECT: offSub=%#x\n", pReq->u.In.offSub));
1528	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1529	("SUP_IOCTL_PAGE_PROTECT: cbSub=%#x\n", pReq->u.In.cbSub));
1530
1531	/* execute */
1532	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);
1533	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1534	return 0;
1535	}
1536
1537	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_FREE):
1538	{
1539	/* validate */
1540	PSUPPAGEFREE pReq = (PSUPPAGEFREE)pReqHdr;
1541	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_FREE);
1542
1543	/* execute */
1544	pReq->Hdr.rc = SUPR0PageFree(pSession, pReq->u.In.pvR3);
1545	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1546	return 0;
1547	}
1548
1549	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_SERVICE(0)):
1550	{
1551	/* validate */
1552	PSUPCALLSERVICE pReq = (PSUPCALLSERVICE)pReqHdr;
1553	Log4(("SUP_IOCTL_CALL_SERVICE: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1554	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1555
1556	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
1557	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(0), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(0));
1558	else
1559	{
1560	PSUPR0SERVICEREQHDR pSrvReq = (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0];
1561	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR)),
1562	("SUP_IOCTL_CALL_SERVICE: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR))));
1563	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, pSrvReq->u32Magic == SUPR0SERVICEREQHDR_MAGIC);
1564	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(pSrvReq->cbReq), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(pSrvReq->cbReq));
1565	}
1566	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1567
1568	/* execute */
1569	pReq->Hdr.rc = supdrvIOCtl_CallServiceModule(pDevExt, pSession, pReq);
1570	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1571	return 0;
1572	}
1573
1574	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOGGER_SETTINGS(0)):
1575	{
1576	/* validate */
1577	PSUPLOGGERSETTINGS pReq = (PSUPLOGGERSETTINGS)pReqHdr;
1578	size_t cbStrTab;
1579	REQ_CHECK_SIZE_OUT(SUP_IOCTL_LOGGER_SETTINGS, SUP_IOCTL_LOGGER_SETTINGS_SIZE_OUT);
1580	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->Hdr.cbIn >= SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(1));
1581	cbStrTab = pReq->Hdr.cbIn - SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(0);
1582	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offGroups < cbStrTab);
1583	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offFlags < cbStrTab);
1584	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offDestination < cbStrTab);
1585	REQ_CHECK_EXPR_FMT(pReq->u.In.szStrings[cbStrTab - 1] == '\0',
1586	("SUP_IOCTL_LOGGER_SETTINGS: cbIn=%#x cbStrTab=%#zx LastChar=%d\n",
1587	pReq->Hdr.cbIn, cbStrTab, pReq->u.In.szStrings[cbStrTab - 1]));
1588	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhich <= SUPLOGGERSETTINGS_WHICH_RELEASE);
1589	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhat <= SUPLOGGERSETTINGS_WHAT_DESTROY);
1590
1591	/* execute */
1592	pReq->Hdr.rc = supdrvIOCtl_LoggerSettings(pDevExt, pSession, pReq);
1593	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1594	return 0;
1595	}
1596
1597	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP2):
1598	{
1599	/* validate */
1600	PSUPSEMOP2 pReq = (PSUPSEMOP2)pReqHdr;
1601	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP2, SUP_IOCTL_SEM_OP2_SIZE_IN, SUP_IOCTL_SEM_OP2_SIZE_OUT);
1602	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP2, pReq->u.In.uReserved == 0);
1603
1604	/* execute */
1605	switch (pReq->u.In.uType)
1606	{
1607	case SUP_SEM_TYPE_EVENT:
1608	{
1609	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1610	switch (pReq->u.In.uOp)
1611	{
1612	case SUPSEMOP2_WAIT_MS_REL:
1613	pReq->Hdr.rc = SUPSemEventWaitNoResume(pSession, hEvent, pReq->u.In.uArg.cRelMsTimeout);
1614	break;
1615	case SUPSEMOP2_WAIT_NS_ABS:
1616	pReq->Hdr.rc = SUPSemEventWaitNsAbsIntr(pSession, hEvent, pReq->u.In.uArg.uAbsNsTimeout);
1617	break;
1618	case SUPSEMOP2_WAIT_NS_REL:
1619	pReq->Hdr.rc = SUPSemEventWaitNsRelIntr(pSession, hEvent, pReq->u.In.uArg.cRelNsTimeout);
1620	break;
1621	case SUPSEMOP2_SIGNAL:
1622	pReq->Hdr.rc = SUPSemEventSignal(pSession, hEvent);
1623	break;
1624	case SUPSEMOP2_CLOSE:
1625	pReq->Hdr.rc = SUPSemEventClose(pSession, hEvent);
1626	break;
1627	case SUPSEMOP2_RESET:
1628	default:
1629	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1630	break;
1631	}
1632	break;
1633	}
1634
1635	case SUP_SEM_TYPE_EVENT_MULTI:
1636	{
1637	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1638	switch (pReq->u.In.uOp)
1639	{
1640	case SUPSEMOP2_WAIT_MS_REL:
1641	pReq->Hdr.rc = SUPSemEventMultiWaitNoResume(pSession, hEventMulti, pReq->u.In.uArg.cRelMsTimeout);
1642	break;
1643	case SUPSEMOP2_WAIT_NS_ABS:
1644	pReq->Hdr.rc = SUPSemEventMultiWaitNsAbsIntr(pSession, hEventMulti, pReq->u.In.uArg.uAbsNsTimeout);
1645	break;
1646	case SUPSEMOP2_WAIT_NS_REL:
1647	pReq->Hdr.rc = SUPSemEventMultiWaitNsRelIntr(pSession, hEventMulti, pReq->u.In.uArg.cRelNsTimeout);
1648	break;
1649	case SUPSEMOP2_SIGNAL:
1650	pReq->Hdr.rc = SUPSemEventMultiSignal(pSession, hEventMulti);
1651	break;
1652	case SUPSEMOP2_CLOSE:
1653	pReq->Hdr.rc = SUPSemEventMultiClose(pSession, hEventMulti);
1654	break;
1655	case SUPSEMOP2_RESET:
1656	pReq->Hdr.rc = SUPSemEventMultiReset(pSession, hEventMulti);
1657	break;
1658	default:
1659	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1660	break;
1661	}
1662	break;
1663	}
1664
1665	default:
1666	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1667	break;
1668	}
1669	return 0;
1670	}
1671
1672	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP3):
1673	{
1674	/* validate */
1675	PSUPSEMOP3 pReq = (PSUPSEMOP3)pReqHdr;
1676	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP3, SUP_IOCTL_SEM_OP3_SIZE_IN, SUP_IOCTL_SEM_OP3_SIZE_OUT);
1677	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, pReq->u.In.u32Reserved == 0 && pReq->u.In.u64Reserved == 0);
1678
1679	/* execute */
1680	switch (pReq->u.In.uType)
1681	{
1682	case SUP_SEM_TYPE_EVENT:
1683	{
1684	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1685	switch (pReq->u.In.uOp)
1686	{
1687	case SUPSEMOP3_CREATE:
1688	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1689	pReq->Hdr.rc = SUPSemEventCreate(pSession, &hEvent);
1690	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEvent;
1691	break;
1692	case SUPSEMOP3_GET_RESOLUTION:
1693	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1694	pReq->Hdr.rc = VINF_SUCCESS;
1695	pReq->Hdr.cbOut = sizeof(*pReq);
1696	pReq->u.Out.cNsResolution = SUPSemEventGetResolution(pSession);
1697	break;
1698	default:
1699	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1700	break;
1701	}
1702	break;
1703	}
1704
1705	case SUP_SEM_TYPE_EVENT_MULTI:
1706	{
1707	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1708	switch (pReq->u.In.uOp)
1709	{
1710	case SUPSEMOP3_CREATE:
1711	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1712	pReq->Hdr.rc = SUPSemEventMultiCreate(pSession, &hEventMulti);
1713	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEventMulti;
1714	break;
1715	case SUPSEMOP3_GET_RESOLUTION:
1716	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1717	pReq->Hdr.rc = VINF_SUCCESS;
1718	pReq->u.Out.cNsResolution = SUPSemEventMultiGetResolution(pSession);
1719	break;
1720	default:
1721	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1722	break;
1723	}
1724	break;
1725	}
1726
1727	default:
1728	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1729	break;
1730	}
1731	return 0;
1732	}
1733
1734	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
1735	{
1736	/* validate */
1737	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
1738	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
1739	REQ_CHECK_EXPR(SUP_IOCTL_VT_CAPS, pReq->Hdr.cbIn <= SUP_IOCTL_VT_CAPS_SIZE_IN);
1740
1741	/* execute */
1742	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.Caps);
1743	if (RT_FAILURE(pReq->Hdr.rc))
1744	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1745	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1746	return 0;
1747	}
1748
1749	default:
1750	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1751	break;
1752	}
1753	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_GENERAL_FAILURE, VERR_GENERAL_FAILURE);
1754	return VERR_GENERAL_FAILURE;
1755	}
1756
1757
1758	/**
1759	* Inter-Driver Communication (IDC) worker.
1760	*
1761	* @returns VBox status code.
1762	* @retval VINF_SUCCESS on success.
1763	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1764	* @retval VERR_NOT_SUPPORTED if the request isn't supported.
1765	*
1766	* @param uReq The request (function) code.
1767	* @param pDevExt Device extention.
1768	* @param pSession Session data.
1769	* @param pReqHdr The request header.
1770	*/
1771	int VBOXCALL supdrvIDC(uintptr_t uReq, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQHDR pReqHdr)
1772	{
1773	/*
1774	* The OS specific code has already validated the pSession
1775	* pointer, and the request size being greater or equal to
1776	* size of the header.
1777	*
1778	* So, just check that pSession is a kernel context session.
1779	*/
1780	if (RT_UNLIKELY( pSession
1781	&& pSession->R0Process != NIL_RTR0PROCESS))
1782	return VERR_INVALID_PARAMETER;
1783
1784	/*
1785	* Validation macro.
1786	*/
1787	#define REQ_CHECK_IDC_SIZE(Name, cbExpect) \
1788	do { \
1789	if (RT_UNLIKELY(pReqHdr->cb != (cbExpect))) \
1790	{ \
1791	OSDBGPRINT(( #Name ": Invalid input/output sizes. cb=%ld expected %ld.\n", \
1792	(long)pReqHdr->cb, (long)(cbExpect))); \
1793	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1794	} \
1795	} while (0)
1796
1797	switch (uReq)
1798	{
1799	case SUPDRV_IDC_REQ_CONNECT:
1800	{
1801	PSUPDRVIDCREQCONNECT pReq = (PSUPDRVIDCREQCONNECT)pReqHdr;
1802	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_CONNECT, sizeof(*pReq));
1803
1804	/*
1805	* Validate the cookie and other input.
1806	*/
1807	if (pReq->Hdr.pSession != NULL)
1808	{
1809	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: pSession=%p expected NULL!\n", pReq->Hdr.pSession));
1810	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1811	}
1812	if (pReq->u.In.u32MagicCookie != SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE)
1813	{
1814	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: u32MagicCookie=%#x expected %#x!\n",
1815	(unsigned)pReq->u.In.u32MagicCookie, (unsigned)SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE));
1816	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1817	}
1818	if ( pReq->u.In.uMinVersion > pReq->u.In.uReqVersion
1819	\|\| (pReq->u.In.uMinVersion & UINT32_C(0xffff0000)) != (pReq->u.In.uReqVersion & UINT32_C(0xffff0000)))
1820	{
1821	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: uMinVersion=%#x uMaxVersion=%#x doesn't match!\n",
1822	pReq->u.In.uMinVersion, pReq->u.In.uReqVersion));
1823	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1824	}
1825
1826	/*
1827	* Match the version.
1828	* The current logic is very simple, match the major interface version.
1829	*/
1830	if ( pReq->u.In.uMinVersion > SUPDRV_IDC_VERSION
1831	\|\| (pReq->u.In.uMinVersion & 0xffff0000) != (SUPDRV_IDC_VERSION & 0xffff0000))
1832	{
1833	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1834	pReq->u.In.uReqVersion, pReq->u.In.uMinVersion, (unsigned)SUPDRV_IDC_VERSION));
1835	pReq->u.Out.pSession = NULL;
1836	pReq->u.Out.uSessionVersion = 0xffffffff;
1837	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1838	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1839	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1840	return VINF_SUCCESS;
1841	}
1842
1843	pReq->u.Out.pSession = NULL;
1844	pReq->u.Out.uSessionVersion = SUPDRV_IDC_VERSION;
1845	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1846	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1847
1848	/*
1849	* On NT we will already have a session associated with the
1850	* client, just like with the SUP_IOCTL_COOKIE request, while
1851	* the other doesn't.
1852	*/
1853	#ifdef RT_OS_WINDOWS
1854	pReq->Hdr.rc = VINF_SUCCESS;
1855	#else
1856	AssertReturn(!pSession, VERR_INTERNAL_ERROR);
1857	pReq->Hdr.rc = supdrvCreateSession(pDevExt, false /* fUser */, &pSession);
1858	if (RT_FAILURE(pReq->Hdr.rc))
1859	{
1860	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: failed to create session, rc=%d\n", pReq->Hdr.rc));
1861	return VINF_SUCCESS;
1862	}
1863	#endif
1864
1865	pReq->u.Out.pSession = pSession;
1866	pReq->Hdr.pSession = pSession;
1867
1868	return VINF_SUCCESS;
1869	}
1870
1871	case SUPDRV_IDC_REQ_DISCONNECT:
1872	{
1873	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_DISCONNECT, sizeof(*pReqHdr));
1874
1875	#ifdef RT_OS_WINDOWS
1876	/* Windows will destroy the session when the file object is destroyed. */
1877	#else
1878	supdrvCloseSession(pDevExt, pSession);
1879	#endif
1880	return pReqHdr->rc = VINF_SUCCESS;
1881	}
1882
1883	case SUPDRV_IDC_REQ_GET_SYMBOL:
1884	{
1885	PSUPDRVIDCREQGETSYM pReq = (PSUPDRVIDCREQGETSYM)pReqHdr;
1886	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_GET_SYMBOL, sizeof(*pReq));
1887
1888	pReq->Hdr.rc = supdrvIDC_LdrGetSymbol(pDevExt, pSession, pReq);
1889	return VINF_SUCCESS;
1890	}
1891
1892	case SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY:
1893	{
1894	PSUPDRVIDCREQCOMPREGFACTORY pReq = (PSUPDRVIDCREQCOMPREGFACTORY)pReqHdr;
1895	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY, sizeof(*pReq));
1896
1897	pReq->Hdr.rc = SUPR0ComponentRegisterFactory(pSession, pReq->u.In.pFactory);
1898	return VINF_SUCCESS;
1899	}
1900
1901	case SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY:
1902	{
1903	PSUPDRVIDCREQCOMPDEREGFACTORY pReq = (PSUPDRVIDCREQCOMPDEREGFACTORY)pReqHdr;
1904	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY, sizeof(*pReq));
1905
1906	pReq->Hdr.rc = SUPR0ComponentDeregisterFactory(pSession, pReq->u.In.pFactory);
1907	return VINF_SUCCESS;
1908	}
1909
1910	default:
1911	Log(("Unknown IDC %#lx\n", (long)uReq));
1912	break;
1913	}
1914
1915	#undef REQ_CHECK_IDC_SIZE
1916	return VERR_NOT_SUPPORTED;
1917	}
1918
1919
1920	/**
1921	* Register a object for reference counting.
1922	* The object is registered with one reference in the specified session.
1923	*
1924	* @returns Unique identifier on success (pointer).
1925	* All future reference must use this identifier.
1926	* @returns NULL on failure.
1927	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
1928	* @param pvUser1 The first user argument.
1929	* @param pvUser2 The second user argument.
1930	*/
1931	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
1932	{
1933	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1934	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1935	PSUPDRVOBJ pObj;
1936	PSUPDRVUSAGE pUsage;
1937
1938	/*
1939	* Validate the input.
1940	*/
1941	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
1942	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
1943	AssertPtrReturn(pfnDestructor, NULL);
1944
1945	/*
1946	* Allocate and initialize the object.
1947	*/
1948	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
1949	if (!pObj)
1950	return NULL;
1951	pObj->u32Magic = SUPDRVOBJ_MAGIC;
1952	pObj->enmType = enmType;
1953	pObj->pNext = NULL;
1954	pObj->cUsage = 1;
1955	pObj->pfnDestructor = pfnDestructor;
1956	pObj->pvUser1 = pvUser1;
1957	pObj->pvUser2 = pvUser2;
1958	pObj->CreatorUid = pSession->Uid;
1959	pObj->CreatorGid = pSession->Gid;
1960	pObj->CreatorProcess= pSession->Process;
1961	supdrvOSObjInitCreator(pObj, pSession);
1962
1963	/*
1964	* Allocate the usage record.
1965	* (We keep freed usage records around to simplify SUPR0ObjAddRefEx().)
1966	*/
1967	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1968
1969	pUsage = pDevExt->pUsageFree;
1970	if (pUsage)
1971	pDevExt->pUsageFree = pUsage->pNext;
1972	else
1973	{
1974	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1975	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
1976	if (!pUsage)
1977	{
1978	RTMemFree(pObj);
1979	return NULL;
1980	}
1981	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1982	}
1983
1984	/*
1985	* Insert the object and create the session usage record.
1986	*/
1987	/* The object. */
1988	pObj->pNext = pDevExt->pObjs;
1989	pDevExt->pObjs = pObj;
1990
1991	/* The session record. */
1992	pUsage->cUsage = 1;
1993	pUsage->pObj = pObj;
1994	pUsage->pNext = pSession->pUsage;
1995	/* Log2(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext)); */
1996	pSession->pUsage = pUsage;
1997
1998	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1999
2000	Log(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
2001	return pObj;
2002	}
2003
2004
2005	/**
2006	* Increment the reference counter for the object associating the reference
2007	* with the specified session.
2008	*
2009	* @returns IPRT status code.
2010	* @param pvObj The identifier returned by SUPR0ObjRegister().
2011	* @param pSession The session which is referencing the object.
2012	*
2013	* @remarks The caller should not own any spinlocks and must carefully protect
2014	* itself against potential race with the destructor so freed memory
2015	* isn't accessed here.
2016	*/
2017	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
2018	{
2019	return SUPR0ObjAddRefEx(pvObj, pSession, false /* fNoBlocking */);
2020	}
2021
2022
2023	/**
2024	* Increment the reference counter for the object associating the reference
2025	* with the specified session.
2026	*
2027	* @returns IPRT status code.
2028	* @retval VERR_TRY_AGAIN if fNoBlocking was set and a new usage record
2029	* couldn't be allocated. (If you see this you're not doing the right
2030	* thing and it won't ever work reliably.)
2031	*
2032	* @param pvObj The identifier returned by SUPR0ObjRegister().
2033	* @param pSession The session which is referencing the object.
2034	* @param fNoBlocking Set if it's not OK to block. Never try to make the
2035	* first reference to an object in a session with this
2036	* argument set.
2037	*
2038	* @remarks The caller should not own any spinlocks and must carefully protect
2039	* itself against potential race with the destructor so freed memory
2040	* isn't accessed here.
2041	*/
2042	SUPR0DECL(int) SUPR0ObjAddRefEx(void *pvObj, PSUPDRVSESSION pSession, bool fNoBlocking)
2043	{
2044	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2045	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2046	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2047	int rc = VINF_SUCCESS;
2048	PSUPDRVUSAGE pUsagePre;
2049	PSUPDRVUSAGE pUsage;
2050
2051	/*
2052	* Validate the input.
2053	* Be ready for the destruction race (someone might be stuck in the
2054	* destructor waiting a lock we own).
2055	*/
2056	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2057	AssertPtrReturn(pObj, VERR_INVALID_POINTER);
2058	AssertMsgReturn(pObj->u32Magic == SUPDRVOBJ_MAGIC \|\| pObj->u32Magic == SUPDRVOBJ_MAGIC_DEAD,
2059	("Invalid pvObj=%p magic=%#x (expected %#x or %#x)\n", pvObj, pObj->u32Magic, SUPDRVOBJ_MAGIC, SUPDRVOBJ_MAGIC_DEAD),
2060	VERR_INVALID_PARAMETER);
2061
2062	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2063
2064	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2065	{
2066	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2067
2068	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2069	return VERR_WRONG_ORDER;
2070	}
2071
2072	/*
2073	* Preallocate the usage record if we can.
2074	*/
2075	pUsagePre = pDevExt->pUsageFree;
2076	if (pUsagePre)
2077	pDevExt->pUsageFree = pUsagePre->pNext;
2078	else if (!fNoBlocking)
2079	{
2080	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2081	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
2082	if (!pUsagePre)
2083	return VERR_NO_MEMORY;
2084
2085	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2086	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2087	{
2088	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2089
2090	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2091	return VERR_WRONG_ORDER;
2092	}
2093	}
2094
2095	/*
2096	* Reference the object.
2097	*/
2098	pObj->cUsage++;
2099
2100	/*
2101	* Look for the session record.
2102	*/
2103	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
2104	{
2105	/Log(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2106	if (pUsage->pObj == pObj)
2107	break;
2108	}
2109	if (pUsage)
2110	pUsage->cUsage++;
2111	else if (pUsagePre)
2112	{
2113	/* create a new session record. */
2114	pUsagePre->cUsage = 1;
2115	pUsagePre->pObj = pObj;
2116	pUsagePre->pNext = pSession->pUsage;
2117	pSession->pUsage = pUsagePre;
2118	/Log(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));/
2119
2120	pUsagePre = NULL;
2121	}
2122	else
2123	{
2124	pObj->cUsage--;
2125	rc = VERR_TRY_AGAIN;
2126	}
2127
2128	/*
2129	* Put any unused usage record into the free list..
2130	*/
2131	if (pUsagePre)
2132	{
2133	pUsagePre->pNext = pDevExt->pUsageFree;
2134	pDevExt->pUsageFree = pUsagePre;
2135	}
2136
2137	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2138
2139	return rc;
2140	}
2141
2142
2143	/**
2144	* Decrement / destroy a reference counter record for an object.
2145	*
2146	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
2147	*
2148	* @returns IPRT status code.
2149	* @retval VINF_SUCCESS if not destroyed.
2150	* @retval VINF_OBJECT_DESTROYED if it's destroyed by this release call.
2151	* @retval VERR_INVALID_PARAMETER if the object isn't valid. Will assert in
2152	* string builds.
2153	*
2154	* @param pvObj The identifier returned by SUPR0ObjRegister().
2155	* @param pSession The session which is referencing the object.
2156	*/
2157	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
2158	{
2159	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2160	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2161	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2162	int rc = VERR_INVALID_PARAMETER;
2163	PSUPDRVUSAGE pUsage;
2164	PSUPDRVUSAGE pUsagePrev;
2165
2166	/*
2167	* Validate the input.
2168	*/
2169	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2170	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2171	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2172	VERR_INVALID_PARAMETER);
2173
2174	/*
2175	* Acquire the spinlock and look for the usage record.
2176	*/
2177	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2178
2179	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
2180	pUsage;
2181	pUsagePrev = pUsage, pUsage = pUsage->pNext)
2182	{
2183	/Log2(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2184	if (pUsage->pObj == pObj)
2185	{
2186	rc = VINF_SUCCESS;
2187	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
2188	if (pUsage->cUsage > 1)
2189	{
2190	pObj->cUsage--;
2191	pUsage->cUsage--;
2192	}
2193	else
2194	{
2195	/*
2196	* Free the session record.
2197	*/
2198	if (pUsagePrev)
2199	pUsagePrev->pNext = pUsage->pNext;
2200	else
2201	pSession->pUsage = pUsage->pNext;
2202	pUsage->pNext = pDevExt->pUsageFree;
2203	pDevExt->pUsageFree = pUsage;
2204
2205	/* What about the object? */
2206	if (pObj->cUsage > 1)
2207	pObj->cUsage--;
2208	else
2209	{
2210	/*
2211	* Object is to be destroyed, unlink it.
2212	*/
2213	pObj->u32Magic = SUPDRVOBJ_MAGIC_DEAD;
2214	rc = VINF_OBJECT_DESTROYED;
2215	if (pDevExt->pObjs == pObj)
2216	pDevExt->pObjs = pObj->pNext;
2217	else
2218	{
2219	PSUPDRVOBJ pObjPrev;
2220	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
2221	if (pObjPrev->pNext == pObj)
2222	{
2223	pObjPrev->pNext = pObj->pNext;
2224	break;
2225	}
2226	Assert(pObjPrev);
2227	}
2228	}
2229	}
2230	break;
2231	}
2232	}
2233
2234	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2235
2236	/*
2237	* Call the destructor and free the object if required.
2238	*/
2239	if (rc == VINF_OBJECT_DESTROYED)
2240	{
2241	Log(("SUPR0ObjRelease: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
2242	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
2243	if (pObj->pfnDestructor)
2244	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
2245	RTMemFree(pObj);
2246	}
2247
2248	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
2249	return rc;
2250	}
2251
2252
2253	/**
2254	* Verifies that the current process can access the specified object.
2255	*
2256	* @returns The following IPRT status code:
2257	* @retval VINF_SUCCESS if access was granted.
2258	* @retval VERR_PERMISSION_DENIED if denied access.
2259	* @retval VERR_INVALID_PARAMETER if invalid parameter.
2260	*
2261	* @param pvObj The identifier returned by SUPR0ObjRegister().
2262	* @param pSession The session which wishes to access the object.
2263	* @param pszObjName Object string name. This is optional and depends on the object type.
2264	*
2265	* @remark The caller is responsible for making sure the object isn't removed while
2266	* we're inside this function. If uncertain about this, just call AddRef before calling us.
2267	*/
2268	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
2269	{
2270	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2271	int rc;
2272
2273	/*
2274	* Validate the input.
2275	*/
2276	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2277	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2278	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2279	VERR_INVALID_PARAMETER);
2280
2281	/*
2282	* Check access. (returns true if a decision has been made.)
2283	*/
2284	rc = VERR_INTERNAL_ERROR;
2285	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
2286	return rc;
2287
2288	/*
2289	* Default policy is to allow the user to access his own
2290	* stuff but nothing else.
2291	*/
2292	if (pObj->CreatorUid == pSession->Uid)
2293	return VINF_SUCCESS;
2294	return VERR_PERMISSION_DENIED;
2295	}
2296
2297
2298	/**
2299	* Lock pages.
2300	*
2301	* @returns IPRT status code.
2302	* @param pSession Session to which the locked memory should be associated.
2303	* @param pvR3 Start of the memory range to lock.
2304	* This must be page aligned.
2305	* @param cPages Number of pages to lock.
2306	* @param paPages Where to put the physical addresses of locked memory.
2307	*/
2308	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
2309	{
2310	int rc;
2311	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2312	const size_t cb = (size_t)cPages << PAGE_SHIFT;
2313	LogFlow(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
2314
2315	/*
2316	* Verify input.
2317	*/
2318	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2319	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
2320	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
2321	\|\| !pvR3)
2322	{
2323	Log(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
2324	return VERR_INVALID_PARAMETER;
2325	}
2326
2327	/*
2328	* Let IPRT do the job.
2329	*/
2330	Mem.eType = MEMREF_TYPE_LOCKED;
2331	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTMEM_PROT_READ \| RTMEM_PROT_WRITE, RTR0ProcHandleSelf());
2332	if (RT_SUCCESS(rc))
2333	{
2334	uint32_t iPage = cPages;
2335	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
2336	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
2337
2338	while (iPage-- > 0)
2339	{
2340	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2341	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
2342	{
2343	AssertMsgFailed(("iPage=%d\n", iPage));
2344	rc = VERR_INTERNAL_ERROR;
2345	break;
2346	}
2347	}
2348	if (RT_SUCCESS(rc))
2349	rc = supdrvMemAdd(&Mem, pSession);
2350	if (RT_FAILURE(rc))
2351	{
2352	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
2353	AssertRC(rc2);
2354	}
2355	}
2356
2357	return rc;
2358	}
2359
2360
2361	/**
2362	* Unlocks the memory pointed to by pv.
2363	*
2364	* @returns IPRT status code.
2365	* @param pSession Session to which the memory was locked.
2366	* @param pvR3 Memory to unlock.
2367	*/
2368	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2369	{
2370	LogFlow(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2371	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2372	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
2373	}
2374
2375
2376	/**
2377	* Allocates a chunk of page aligned memory with contiguous and fixed physical
2378	* backing.
2379	*
2380	* @returns IPRT status code.
2381	* @param pSession Session data.
2382	* @param cPages Number of pages to allocate.
2383	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
2384	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
2385	* @param pHCPhys Where to put the physical address of allocated memory.
2386	*/
2387	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
2388	{
2389	int rc;
2390	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2391	LogFlow(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
2392
2393	/*
2394	* Validate input.
2395	*/
2396	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2397	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
2398	{
2399	Log(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
2400	pSession, ppvR0, ppvR3, pHCPhys));
2401	return VERR_INVALID_PARAMETER;
2402
2403	}
2404	if (cPages < 1 \|\| cPages >= 256)
2405	{
2406	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2407	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2408	}
2409
2410	/*
2411	* Let IPRT do the job.
2412	*/
2413	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
2414	if (RT_SUCCESS(rc))
2415	{
2416	int rc2;
2417	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2418	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2419	if (RT_SUCCESS(rc))
2420	{
2421	Mem.eType = MEMREF_TYPE_CONT;
2422	rc = supdrvMemAdd(&Mem, pSession);
2423	if (!rc)
2424	{
2425	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2426	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2427	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
2428	return 0;
2429	}
2430
2431	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2432	AssertRC(rc2);
2433	}
2434	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2435	AssertRC(rc2);
2436	}
2437
2438	return rc;
2439	}
2440
2441
2442	/**
2443	* Frees memory allocated using SUPR0ContAlloc().
2444	*
2445	* @returns IPRT status code.
2446	* @param pSession The session to which the memory was allocated.
2447	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2448	*/
2449	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2450	{
2451	LogFlow(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2452	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2453	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
2454	}
2455
2456
2457	/**
2458	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
2459	*
2460	* The memory isn't zeroed.
2461	*
2462	* @returns IPRT status code.
2463	* @param pSession Session data.
2464	* @param cPages Number of pages to allocate.
2465	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
2466	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
2467	* @param paPages Where to put the physical addresses of allocated memory.
2468	*/
2469	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
2470	{
2471	unsigned iPage;
2472	int rc;
2473	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2474	LogFlow(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
2475
2476	/*
2477	* Validate input.
2478	*/
2479	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2480	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
2481	{
2482	Log(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
2483	pSession, ppvR3, ppvR0, paPages));
2484	return VERR_INVALID_PARAMETER;
2485
2486	}
2487	if (cPages < 1 \|\| cPages >= 256)
2488	{
2489	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2490	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2491	}
2492
2493	/*
2494	* Let IPRT do the work.
2495	*/
2496	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
2497	if (RT_SUCCESS(rc))
2498	{
2499	int rc2;
2500	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2501	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2502	if (RT_SUCCESS(rc))
2503	{
2504	Mem.eType = MEMREF_TYPE_LOW;
2505	rc = supdrvMemAdd(&Mem, pSession);
2506	if (!rc)
2507	{
2508	for (iPage = 0; iPage < cPages; iPage++)
2509	{
2510	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2511	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%RHp\n", paPages[iPage]));
2512	}
2513	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2514	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2515	return 0;
2516	}
2517
2518	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2519	AssertRC(rc2);
2520	}
2521
2522	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2523	AssertRC(rc2);
2524	}
2525
2526	return rc;
2527	}
2528
2529
2530	/**
2531	* Frees memory allocated using SUPR0LowAlloc().
2532	*
2533	* @returns IPRT status code.
2534	* @param pSession The session to which the memory was allocated.
2535	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2536	*/
2537	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2538	{
2539	LogFlow(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2540	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2541	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
2542	}
2543
2544
2545
2546	/**
2547	* Allocates a chunk of memory with both R0 and R3 mappings.
2548	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
2549	*
2550	* @returns IPRT status code.
2551	* @param pSession The session to associated the allocation with.
2552	* @param cb Number of bytes to allocate.
2553	* @param ppvR0 Where to store the address of the Ring-0 mapping.
2554	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2555	*/
2556	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
2557	{
2558	int rc;
2559	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2560	LogFlow(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
2561
2562	/*
2563	* Validate input.
2564	*/
2565	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2566	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
2567	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
2568	if (cb < 1 \|\| cb >= _4M)
2569	{
2570	Log(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
2571	return VERR_INVALID_PARAMETER;
2572	}
2573
2574	/*
2575	* Let IPRT do the work.
2576	*/
2577	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
2578	if (RT_SUCCESS(rc))
2579	{
2580	int rc2;
2581	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2582	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2583	if (RT_SUCCESS(rc))
2584	{
2585	Mem.eType = MEMREF_TYPE_MEM;
2586	rc = supdrvMemAdd(&Mem, pSession);
2587	if (!rc)
2588	{
2589	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2590	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2591	return VINF_SUCCESS;
2592	}
2593
2594	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2595	AssertRC(rc2);
2596	}
2597
2598	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2599	AssertRC(rc2);
2600	}
2601
2602	return rc;
2603	}
2604
2605
2606	/**
2607	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
2608	*
2609	* @returns IPRT status code.
2610	* @param pSession The session to which the memory was allocated.
2611	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2612	* @param paPages Where to store the physical addresses.
2613	*/
2614	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
2615	{
2616	PSUPDRVBUNDLE pBundle;
2617	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2618	LogFlow(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
2619
2620	/*
2621	* Validate input.
2622	*/
2623	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2624	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
2625	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
2626
2627	/*
2628	* Search for the address.
2629	*/
2630	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2631	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2632	{
2633	if (pBundle->cUsed > 0)
2634	{
2635	unsigned i;
2636	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2637	{
2638	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
2639	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2640	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2641	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2642	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
2643	)
2644	)
2645	{
2646	const size_t cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2647	size_t iPage;
2648	for (iPage = 0; iPage < cPages; iPage++)
2649	{
2650	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2651	paPages[iPage].uReserved = 0;
2652	}
2653	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2654	return VINF_SUCCESS;
2655	}
2656	}
2657	}
2658	}
2659	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2660	Log(("Failed to find %p!!!\n", (void *)uPtr));
2661	return VERR_INVALID_PARAMETER;
2662	}
2663
2664
2665	/**
2666	* Free memory allocated by SUPR0MemAlloc().
2667	*
2668	* @returns IPRT status code.
2669	* @param pSession The session owning the allocation.
2670	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2671	*/
2672	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2673	{
2674	LogFlow(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2675	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2676	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
2677	}
2678
2679
2680	/**
2681	* Allocates a chunk of memory with a kernel or/and a user mode mapping.
2682	*
2683	* The memory is fixed and it's possible to query the physical addresses using
2684	* SUPR0MemGetPhys().
2685	*
2686	* @returns IPRT status code.
2687	* @param pSession The session to associated the allocation with.
2688	* @param cPages The number of pages to allocate.
2689	* @param fFlags Flags, reserved for the future. Must be zero.
2690	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2691	* NULL if no ring-3 mapping.
2692	* @param ppvR3 Where to store the address of the Ring-0 mapping.
2693	* NULL if no ring-0 mapping.
2694	* @param paPages Where to store the addresses of the pages. Optional.
2695	*/
2696	SUPR0DECL(int) SUPR0PageAllocEx(PSUPDRVSESSION pSession, uint32_t cPages, uint32_t fFlags, PRTR3PTR ppvR3, PRTR0PTR ppvR0, PRTHCPHYS paPages)
2697	{
2698	int rc;
2699	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2700	LogFlow(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
2701
2702	/*
2703	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2704	*/
2705	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2706	AssertPtrNullReturn(ppvR3, VERR_INVALID_POINTER);
2707	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2708	AssertReturn(ppvR3 \|\| ppvR0, VERR_INVALID_PARAMETER);
2709	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2710	if (cPages < 1 \|\| cPages > VBOX_MAX_ALLOC_PAGE_COUNT)
2711	{
2712	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)));
2713	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2714	}
2715
2716	/*
2717	* Let IPRT do the work.
2718	*/
2719	if (ppvR0)
2720	rc = RTR0MemObjAllocPage(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, true /* fExecutable */);
2721	else
2722	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
2723	if (RT_SUCCESS(rc))
2724	{
2725	int rc2;
2726	if (ppvR3)
2727	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2728	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2729	else
2730	Mem.MapObjR3 = NIL_RTR0MEMOBJ;
2731	if (RT_SUCCESS(rc))
2732	{
2733	Mem.eType = MEMREF_TYPE_PAGE;
2734	rc = supdrvMemAdd(&Mem, pSession);
2735	if (!rc)
2736	{
2737	if (ppvR3)
2738	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2739	if (ppvR0)
2740	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2741	if (paPages)
2742	{
2743	uint32_t iPage = cPages;
2744	while (iPage-- > 0)
2745	{
2746	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
2747	Assert(paPages[iPage] != NIL_RTHCPHYS);
2748	}
2749	}
2750	return VINF_SUCCESS;
2751	}
2752
2753	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2754	AssertRC(rc2);
2755	}
2756
2757	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2758	AssertRC(rc2);
2759	}
2760	return rc;
2761	}
2762
2763
2764	/**
2765	* Maps a chunk of memory previously allocated by SUPR0PageAllocEx into kernel
2766	* space.
2767	*
2768	* @returns IPRT status code.
2769	* @param pSession The session to associated the allocation with.
2770	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
2771	* @param offSub Where to start mapping. Must be page aligned.
2772	* @param cbSub How much to map. Must be page aligned.
2773	* @param fFlags Flags, MBZ.
2774	* @param ppvR0 Where to return the address of the ring-0 mapping on
2775	* success.
2776	*/
2777	SUPR0DECL(int) SUPR0PageMapKernel(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t offSub, uint32_t cbSub,
2778	uint32_t fFlags, PRTR0PTR ppvR0)
2779	{
2780	int rc;
2781	PSUPDRVBUNDLE pBundle;
2782	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2783	RTR0MEMOBJ hMemObj = NIL_RTR0MEMOBJ;
2784	LogFlow(("SUPR0PageMapKernel: pSession=%p pvR3=%p offSub=%#x cbSub=%#x\n", pSession, pvR3, offSub, cbSub));
2785
2786	/*
2787	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2788	*/
2789	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2790	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2791	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2792	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2793	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2794	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
2795
2796	/*
2797	* Find the memory object.
2798	*/
2799	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2800	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2801	{
2802	if (pBundle->cUsed > 0)
2803	{
2804	unsigned i;
2805	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2806	{
2807	if ( ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
2808	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2809	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2810	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
2811	\|\| ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED
2812	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2813	&& pBundle->aMem[i].MapObjR3 == NIL_RTR0MEMOBJ
2814	&& RTR0MemObjAddressR3(pBundle->aMem[i].MemObj) == pvR3))
2815	{
2816	hMemObj = pBundle->aMem[i].MemObj;
2817	break;
2818	}
2819	}
2820	}
2821	}
2822	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2823
2824	rc = VERR_INVALID_PARAMETER;
2825	if (hMemObj != NIL_RTR0MEMOBJ)
2826	{
2827	/*
2828	* Do some further input validations before calling IPRT.
2829	* (Cleanup is done indirectly by telling RTR0MemObjFree to include mappings.)
2830	*/
2831	size_t cbMemObj = RTR0MemObjSize(hMemObj);
2832	if ( offSub < cbMemObj
2833	&& cbSub <= cbMemObj
2834	&& offSub + cbSub <= cbMemObj)
2835	{
2836	RTR0MEMOBJ hMapObj;
2837	rc = RTR0MemObjMapKernelEx(&hMapObj, hMemObj, (void *)-1, 0,
2838	RTMEM_PROT_READ \| RTMEM_PROT_WRITE, offSub, cbSub);
2839	if (RT_SUCCESS(rc))
2840	*ppvR0 = RTR0MemObjAddress(hMapObj);
2841	}
2842	else
2843	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
2844
2845	}
2846	return rc;
2847	}
2848
2849
2850	/**
2851	* Changes the page level protection of one or more pages previously allocated
2852	* by SUPR0PageAllocEx.
2853	*
2854	* @returns IPRT status code.
2855	* @param pSession The session to associated the allocation with.
2856	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
2857	* NIL_RTR3PTR if the ring-3 mapping should be unaffected.
2858	* @param pvR0 The ring-0 address returned by SUPR0PageAllocEx.
2859	* NIL_RTR0PTR if the ring-0 mapping should be unaffected.
2860	* @param offSub Where to start changing. Must be page aligned.
2861	* @param cbSub How much to change. Must be page aligned.
2862	* @param fProt The new page level protection, see RTMEM_PROT_*.
2863	*/
2864	SUPR0DECL(int) SUPR0PageProtect(PSUPDRVSESSION pSession, RTR3PTR pvR3, RTR0PTR pvR0, uint32_t offSub, uint32_t cbSub, uint32_t fProt)
2865	{
2866	int rc;
2867	PSUPDRVBUNDLE pBundle;
2868	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2869	RTR0MEMOBJ hMemObjR0 = NIL_RTR0MEMOBJ;
2870	RTR0MEMOBJ hMemObjR3 = NIL_RTR0MEMOBJ;
2871	LogFlow(("SUPR0PageProtect: pSession=%p pvR3=%p pvR0=%p offSub=%#x cbSub=%#x fProt-%#x\n", pSession, pvR3, pvR0, offSub, cbSub, fProt));
2872
2873	/*
2874	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2875	*/
2876	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2877	AssertReturn(!(fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)), VERR_INVALID_PARAMETER);
2878	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2879	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2880	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
2881
2882	/*
2883	* Find the memory object.
2884	*/
2885	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2886	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2887	{
2888	if (pBundle->cUsed > 0)
2889	{
2890	unsigned i;
2891	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2892	{
2893	if ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
2894	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2895	&& ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2896	\|\| pvR3 == NIL_RTR3PTR)
2897	&& ( pvR0 == NIL_RTR0PTR
2898	\|\| RTR0MemObjAddress(pBundle->aMem[i].MemObj) == pvR0)
2899	&& ( pvR3 == NIL_RTR3PTR
2900	\|\| RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3))
2901	{
2902	if (pvR0 != NIL_RTR0PTR)
2903	hMemObjR0 = pBundle->aMem[i].MemObj;
2904	if (pvR3 != NIL_RTR3PTR)
2905	hMemObjR3 = pBundle->aMem[i].MapObjR3;
2906	break;
2907	}
2908	}
2909	}
2910	}
2911	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2912
2913	rc = VERR_INVALID_PARAMETER;
2914	if ( hMemObjR0 != NIL_RTR0MEMOBJ
2915	\|\| hMemObjR3 != NIL_RTR0MEMOBJ)
2916	{
2917	/*
2918	* Do some further input validations before calling IPRT.
2919	*/
2920	size_t cbMemObj = hMemObjR0 != NIL_RTR0PTR ? RTR0MemObjSize(hMemObjR0) : RTR0MemObjSize(hMemObjR3);
2921	if ( offSub < cbMemObj
2922	&& cbSub <= cbMemObj
2923	&& offSub + cbSub <= cbMemObj)
2924	{
2925	rc = VINF_SUCCESS;
2926	if (hMemObjR3 != NIL_RTR0PTR)
2927	rc = RTR0MemObjProtect(hMemObjR3, offSub, cbSub, fProt);
2928	if (hMemObjR0 != NIL_RTR0PTR && RT_SUCCESS(rc))
2929	rc = RTR0MemObjProtect(hMemObjR0, offSub, cbSub, fProt);
2930	}
2931	else
2932	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
2933
2934	}
2935	return rc;
2936
2937	}
2938
2939
2940	/**
2941	* Free memory allocated by SUPR0PageAlloc() and SUPR0PageAllocEx().
2942	*
2943	* @returns IPRT status code.
2944	* @param pSession The session owning the allocation.
2945	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc() or
2946	* SUPR0PageAllocEx().
2947	*/
2948	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2949	{
2950	LogFlow(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2951	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2952	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_PAGE);
2953	}
2954
2955
2956	/**
2957	* Gets the paging mode of the current CPU.
2958	*
2959	* @returns Paging mode, SUPPAGEINGMODE_INVALID on error.
2960	*/
2961	SUPR0DECL(SUPPAGINGMODE) SUPR0GetPagingMode(void)
2962	{
2963	SUPPAGINGMODE enmMode;
2964
2965	RTR0UINTREG cr0 = ASMGetCR0();
2966	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
2967	enmMode = SUPPAGINGMODE_INVALID;
2968	else
2969	{
2970	RTR0UINTREG cr4 = ASMGetCR4();
2971	uint32_t fNXEPlusLMA = 0;
2972	if (cr4 & X86_CR4_PAE)
2973	{
2974	uint32_t fAmdFeatures = ASMCpuId_EDX(0x80000001);
2975	if (fAmdFeatures & (X86_CPUID_AMD_FEATURE_EDX_NX \| X86_CPUID_AMD_FEATURE_EDX_LONG_MODE))
2976	{
2977	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
2978	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
2979	fNXEPlusLMA \|= RT_BIT(0);
2980	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
2981	fNXEPlusLMA \|= RT_BIT(1);
2982	}
2983	}
2984
2985	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
2986	{
2987	case 0:
2988	enmMode = SUPPAGINGMODE_32_BIT;
2989	break;
2990
2991	case X86_CR4_PGE:
2992	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
2993	break;
2994
2995	case X86_CR4_PAE:
2996	enmMode = SUPPAGINGMODE_PAE;
2997	break;
2998
2999	case X86_CR4_PAE \| RT_BIT(0):
3000	enmMode = SUPPAGINGMODE_PAE_NX;
3001	break;
3002
3003	case X86_CR4_PAE \| X86_CR4_PGE:
3004	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3005	break;
3006
3007	case X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3008	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3009	break;
3010
3011	case RT_BIT(1) \| X86_CR4_PAE:
3012	enmMode = SUPPAGINGMODE_AMD64;
3013	break;
3014
3015	case RT_BIT(1) \| X86_CR4_PAE \| RT_BIT(0):
3016	enmMode = SUPPAGINGMODE_AMD64_NX;
3017	break;
3018
3019	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
3020	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
3021	break;
3022
3023	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3024	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
3025	break;
3026
3027	default:
3028	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
3029	enmMode = SUPPAGINGMODE_INVALID;
3030	break;
3031	}
3032	}
3033	return enmMode;
3034	}
3035
3036
3037	/**
3038	* Enables or disabled hardware virtualization extensions using native OS APIs.
3039	*
3040	* @returns VBox status code.
3041	* @retval VINF_SUCCESS on success.
3042	* @retval VERR_NOT_SUPPORTED if not supported by the native OS.
3043	*
3044	* @param fEnable Whether to enable or disable.
3045	*/
3046	SUPR0DECL(int) SUPR0EnableVTx(bool fEnable)
3047	{
3048	#ifdef RT_OS_DARWIN
3049	return supdrvOSEnableVTx(fEnable);
3050	#else
3051	return VERR_NOT_SUPPORTED;
3052	#endif
3053	}
3054
3055
3056	/** @todo document me */
3057	SUPR0DECL(int) SUPR0QueryVTCaps(PSUPDRVSESSION pSession, uint32_t *pfCaps)
3058	{
3059	/*
3060	* Input validation.
3061	*/
3062	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3063	AssertPtrReturn(pfCaps, VERR_INVALID_POINTER);
3064
3065	*pfCaps = 0;
3066
3067	if (ASMHasCpuId())
3068	{
3069	uint32_t u32FeaturesECX;
3070	uint32_t u32Dummy;
3071	uint32_t u32FeaturesEDX;
3072	uint32_t u32VendorEBX, u32VendorECX, u32VendorEDX, u32AMDFeatureEDX, u32AMDFeatureECX;
3073	uint64_t val;
3074
3075	ASMCpuId(0, &u32Dummy, &u32VendorEBX, &u32VendorECX, &u32VendorEDX);
3076	ASMCpuId(1, &u32Dummy, &u32Dummy, &u32FeaturesECX, &u32FeaturesEDX);
3077	/* Query AMD features. */
3078	ASMCpuId(0x80000001, &u32Dummy, &u32Dummy, &u32AMDFeatureECX, &u32AMDFeatureEDX);
3079
3080	if ( u32VendorEBX == X86_CPUID_VENDOR_INTEL_EBX
3081	&& u32VendorECX == X86_CPUID_VENDOR_INTEL_ECX
3082	&& u32VendorEDX == X86_CPUID_VENDOR_INTEL_EDX
3083	)
3084	{
3085	if ( (u32FeaturesECX & X86_CPUID_FEATURE_ECX_VMX)
3086	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3087	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3088	)
3089	{
3090	val = ASMRdMsr(MSR_IA32_FEATURE_CONTROL);
3091	/*
3092	* Both the LOCK and VMXON bit must be set; otherwise VMXON will generate a #GP.
3093	* Once the lock bit is set, this MSR can no longer be modified.
3094	*/
3095	if ( (val & (MSR_IA32_FEATURE_CONTROL_VMXON\|MSR_IA32_FEATURE_CONTROL_LOCK))
3096	== (MSR_IA32_FEATURE_CONTROL_VMXON\|MSR_IA32_FEATURE_CONTROL_LOCK) /* enabled and locked */
3097	\|\| !(val & MSR_IA32_FEATURE_CONTROL_LOCK) /* not enabled, but not locked either */
3098	)
3099	{
3100	VMX_CAPABILITY vtCaps;
3101
3102	*pfCaps \|= SUPVTCAPS_VT_X;
3103
3104	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS);
3105	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC_USE_SECONDARY_EXEC_CTRL)
3106	{
3107	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS2);
3108	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC2_EPT)
3109	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3110	}
3111	return VINF_SUCCESS;
3112	}
3113	return VERR_VMX_MSR_LOCKED_OR_DISABLED;
3114	}
3115	return VERR_VMX_NO_VMX;
3116	}
3117
3118	if ( u32VendorEBX == X86_CPUID_VENDOR_AMD_EBX
3119	&& u32VendorECX == X86_CPUID_VENDOR_AMD_ECX
3120	&& u32VendorEDX == X86_CPUID_VENDOR_AMD_EDX
3121	)
3122	{
3123	if ( (u32AMDFeatureECX & X86_CPUID_AMD_FEATURE_ECX_SVM)
3124	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3125	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3126	)
3127	{
3128	/* Check if SVM is disabled */
3129	val = ASMRdMsr(MSR_K8_VM_CR);
3130	if (!(val & MSR_K8_VM_CR_SVM_DISABLE))
3131	{
3132	*pfCaps \|= SUPVTCAPS_AMD_V;
3133
3134	/* Query AMD features. */
3135	ASMCpuId(0x8000000A, &u32Dummy, &u32Dummy, &u32Dummy, &u32FeaturesEDX);
3136
3137	if (u32FeaturesEDX & AMD_CPUID_SVM_FEATURE_EDX_NESTED_PAGING)
3138	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3139
3140	return VINF_SUCCESS;
3141	}
3142	return VERR_SVM_DISABLED;
3143	}
3144	return VERR_SVM_NO_SVM;
3145	}
3146	}
3147
3148	return VERR_UNSUPPORTED_CPU;
3149	}
3150
3151
3152	/**
3153	* (Re-)initializes the per-cpu structure prior to starting or resuming the GIP
3154	* updating.
3155	*
3156	* @param pGipCpu The per CPU structure for this CPU.
3157	* @param u64NanoTS The current time.
3158	*/
3159	static void supdrvGipReInitCpu(PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
3160	{
3161	pGipCpu->u64TSC = ASMReadTSC() - pGipCpu->u32UpdateIntervalTSC;
3162	pGipCpu->u64NanoTS = u64NanoTS;
3163	}
3164
3165
3166	/**
3167	* Set the current TSC and NanoTS value for the CPU.
3168	*
3169	* @param idCpu The CPU ID. Unused - we have to use the APIC ID.
3170	* @param pvUser1 Pointer to the ring-0 GIP mapping.
3171	* @param pvUser2 Pointer to the variable holding the current time.
3172	*/
3173	static DECLCALLBACK(void) supdrvGipReInitCpuCallback(RTCPUID idCpu, void pvUser1, void pvUser2)
3174	{
3175	PSUPGLOBALINFOPAGE pGip = (PSUPGLOBALINFOPAGE)pvUser1;
3176	unsigned iCpu = ASMGetApicId();
3177
3178	if (RT_LIKELY(iCpu < RT_ELEMENTS(pGip->aCPUs)))
3179	supdrvGipReInitCpu(&pGip->aCPUs[iCpu], (uint64_t )pvUser2);
3180
3181	NOREF(pvUser2);
3182	NOREF(idCpu);
3183	}
3184
3185
3186	/**
3187	* Maps the GIP into userspace and/or get the physical address of the GIP.
3188	*
3189	* @returns IPRT status code.
3190	* @param pSession Session to which the GIP mapping should belong.
3191	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
3192	* @param pHCPhysGip Where to store the physical address. (optional)
3193	*
3194	* @remark There is no reference counting on the mapping, so one call to this function
3195	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
3196	* and remove the session as a GIP user.
3197	*/
3198	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
3199	{
3200	int rc;
3201	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3202	RTR3PTR pGipR3 = NIL_RTR3PTR;
3203	RTHCPHYS HCPhys = NIL_RTHCPHYS;
3204	LogFlow(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
3205
3206	/*
3207	* Validate
3208	*/
3209	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3210	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
3211	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
3212
3213	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3214	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3215	#else
3216	RTSemFastMutexRequest(pDevExt->mtxGip);
3217	#endif
3218	if (pDevExt->pGip)
3219	{
3220	/*
3221	* Map it?
3222	*/
3223	rc = VINF_SUCCESS;
3224	if (ppGipR3)
3225	{
3226	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
3227	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
3228	RTMEM_PROT_READ, RTR0ProcHandleSelf());
3229	if (RT_SUCCESS(rc))
3230	pGipR3 = RTR0MemObjAddressR3(pSession->GipMapObjR3);
3231	}
3232
3233	/*
3234	* Get physical address.
3235	*/
3236	if (pHCPhysGip && RT_SUCCESS(rc))
3237	HCPhys = pDevExt->HCPhysGip;
3238
3239	/*
3240	* Reference globally.
3241	*/
3242	if (!pSession->fGipReferenced && RT_SUCCESS(rc))
3243	{
3244	pSession->fGipReferenced = 1;
3245	pDevExt->cGipUsers++;
3246	if (pDevExt->cGipUsers == 1)
3247	{
3248	PSUPGLOBALINFOPAGE pGipR0 = pDevExt->pGip;
3249	uint64_t u64NanoTS;
3250	uint32_t u32SystemResolution;
3251	unsigned i;
3252
3253	LogFlow(("SUPR0GipMap: Resumes GIP updating\n"));
3254
3255	/*
3256	* Try bump up the system timer resolution.
3257	* The more interrupts the better...
3258	*/
3259	if ( RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
3260	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
3261	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1953125 /* 512 HZ */, &u32SystemResolution))
3262	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 2000000 /* 500 HZ */, &u32SystemResolution))
3263	)
3264	{
3265	Assert(RTTimerGetSystemGranularity() <= u32SystemResolution);
3266	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
3267	}
3268
3269	if (pGipR0->aCPUs[0].u32TransactionId != 2 /* not the first time */)
3270	{
3271	for (i = 0; i < RT_ELEMENTS(pGipR0->aCPUs); i++)
3272	ASMAtomicUoWriteU32(&pGipR0->aCPUs[i].u32TransactionId,
3273	(pGipR0->aCPUs[i].u32TransactionId + GIP_UPDATEHZ_RECALC_FREQ * 2)
3274	& ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
3275	ASMAtomicWriteU64(&pGipR0->u64NanoTSLastUpdateHz, 0);
3276	}
3277
3278	u64NanoTS = RTTimeSystemNanoTS() - pGipR0->u32UpdateIntervalNS;
3279	if ( pGipR0->u32Mode == SUPGIPMODE_SYNC_TSC
3280	\|\| RTMpGetOnlineCount() == 1)
3281	supdrvGipReInitCpu(&pGipR0->aCPUs[0], u64NanoTS);
3282	else
3283	RTMpOnAll(supdrvGipReInitCpuCallback, pGipR0, &u64NanoTS);
3284
3285	#ifndef DO_NOT_START_GIP
3286	rc = RTTimerStart(pDevExt->pGipTimer, 0); AssertRC(rc);
3287	#endif
3288	rc = VINF_SUCCESS;
3289	}
3290	}
3291	}
3292	else
3293	{
3294	rc = VERR_GENERAL_FAILURE;
3295	Log(("SUPR0GipMap: GIP is not available!\n"));
3296	}
3297	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3298	RTSemMutexRelease(pDevExt->mtxGip);
3299	#else
3300	RTSemFastMutexRelease(pDevExt->mtxGip);
3301	#endif
3302
3303	/*
3304	* Write returns.
3305	*/
3306	if (pHCPhysGip)
3307	*pHCPhysGip = HCPhys;
3308	if (ppGipR3)
3309	*ppGipR3 = pGipR3;
3310
3311	#ifdef DEBUG_DARWIN_GIP
3312	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3313	#else
3314	LogFlow(( "SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3315	#endif
3316	return rc;
3317	}
3318
3319
3320	/**
3321	* Unmaps any user mapping of the GIP and terminates all GIP access
3322	* from this session.
3323	*
3324	* @returns IPRT status code.
3325	* @param pSession Session to which the GIP mapping should belong.
3326	*/
3327	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
3328	{
3329	int rc = VINF_SUCCESS;
3330	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3331	#ifdef DEBUG_DARWIN_GIP
3332	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
3333	pSession,
3334	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
3335	pSession->GipMapObjR3));
3336	#else
3337	LogFlow(("SUPR0GipUnmap: pSession=%p\n", pSession));
3338	#endif
3339	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3340
3341	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3342	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3343	#else
3344	RTSemFastMutexRequest(pDevExt->mtxGip);
3345	#endif
3346
3347	/*
3348	* Unmap anything?
3349	*/
3350	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
3351	{
3352	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
3353	AssertRC(rc);
3354	if (RT_SUCCESS(rc))
3355	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
3356	}
3357
3358	/*
3359	* Dereference global GIP.
3360	*/
3361	if (pSession->fGipReferenced && !rc)
3362	{
3363	pSession->fGipReferenced = 0;
3364	if ( pDevExt->cGipUsers > 0
3365	&& !--pDevExt->cGipUsers)
3366	{
3367	LogFlow(("SUPR0GipUnmap: Suspends GIP updating\n"));
3368	#ifndef DO_NOT_START_GIP
3369	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = VINF_SUCCESS;
3370	#endif
3371
3372	if (pDevExt->u32SystemTimerGranularityGrant)
3373	{
3374	int rc2 = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant);
3375	AssertRC(rc2);
3376	pDevExt->u32SystemTimerGranularityGrant = 0;
3377	}
3378	}
3379	}
3380
3381	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3382	RTSemMutexRelease(pDevExt->mtxGip);
3383	#else
3384	RTSemFastMutexRelease(pDevExt->mtxGip);
3385	#endif
3386
3387	return rc;
3388	}
3389
3390
3391	/**
3392	* Gets the GIP pointer.
3393	*
3394	* @returns Pointer to the GIP or NULL.
3395	*/
3396	SUPDECL(PSUPGLOBALINFOPAGE) SUPGetGIP(void)
3397	{
3398	return g_pSUPGlobalInfoPage;
3399	}
3400
3401
3402	/**
3403	* Register a component factory with the support driver.
3404	*
3405	* This is currently restricted to kernel sessions only.
3406	*
3407	* @returns VBox status code.
3408	* @retval VINF_SUCCESS on success.
3409	* @retval VERR_NO_MEMORY if we're out of memory.
3410	* @retval VERR_ALREADY_EXISTS if the factory has already been registered.
3411	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3412	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3413	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3414	*
3415	* @param pSession The SUPDRV session (must be a ring-0 session).
3416	* @param pFactory Pointer to the component factory registration structure.
3417	*
3418	* @remarks This interface is also available via SUPR0IdcComponentRegisterFactory.
3419	*/
3420	SUPR0DECL(int) SUPR0ComponentRegisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3421	{
3422	PSUPDRVFACTORYREG pNewReg;
3423	const char *psz;
3424	int rc;
3425
3426	/*
3427	* Validate parameters.
3428	*/
3429	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3430	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3431	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3432	AssertPtrReturn(pFactory->pfnQueryFactoryInterface, VERR_INVALID_POINTER);
3433	psz = RTStrEnd(pFactory->szName, sizeof(pFactory->szName));
3434	AssertReturn(psz, VERR_INVALID_PARAMETER);
3435
3436	/*
3437	* Allocate and initialize a new registration structure.
3438	*/
3439	pNewReg = (PSUPDRVFACTORYREG)RTMemAlloc(sizeof(SUPDRVFACTORYREG));
3440	if (pNewReg)
3441	{
3442	pNewReg->pNext = NULL;
3443	pNewReg->pFactory = pFactory;
3444	pNewReg->pSession = pSession;
3445	pNewReg->cchName = psz - &pFactory->szName[0];
3446
3447	/*
3448	* Add it to the tail of the list after checking for prior registration.
3449	*/
3450	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3451	if (RT_SUCCESS(rc))
3452	{
3453	PSUPDRVFACTORYREG pPrev = NULL;
3454	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3455	while (pCur && pCur->pFactory != pFactory)
3456	{
3457	pPrev = pCur;
3458	pCur = pCur->pNext;
3459	}
3460	if (!pCur)
3461	{
3462	if (pPrev)
3463	pPrev->pNext = pNewReg;
3464	else
3465	pSession->pDevExt->pComponentFactoryHead = pNewReg;
3466	rc = VINF_SUCCESS;
3467	}
3468	else
3469	rc = VERR_ALREADY_EXISTS;
3470
3471	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3472	}
3473
3474	if (RT_FAILURE(rc))
3475	RTMemFree(pNewReg);
3476	}
3477	else
3478	rc = VERR_NO_MEMORY;
3479	return rc;
3480	}
3481
3482
3483	/**
3484	* Deregister a component factory.
3485	*
3486	* @returns VBox status code.
3487	* @retval VINF_SUCCESS on success.
3488	* @retval VERR_NOT_FOUND if the factory wasn't registered.
3489	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3490	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3491	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3492	*
3493	* @param pSession The SUPDRV session (must be a ring-0 session).
3494	* @param pFactory Pointer to the component factory registration structure
3495	* previously passed SUPR0ComponentRegisterFactory().
3496	*
3497	* @remarks This interface is also available via SUPR0IdcComponentDeregisterFactory.
3498	*/
3499	SUPR0DECL(int) SUPR0ComponentDeregisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3500	{
3501	int rc;
3502
3503	/*
3504	* Validate parameters.
3505	*/
3506	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3507	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3508	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3509
3510	/*
3511	* Take the lock and look for the registration record.
3512	*/
3513	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3514	if (RT_SUCCESS(rc))
3515	{
3516	PSUPDRVFACTORYREG pPrev = NULL;
3517	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3518	while (pCur && pCur->pFactory != pFactory)
3519	{
3520	pPrev = pCur;
3521	pCur = pCur->pNext;
3522	}
3523	if (pCur)
3524	{
3525	if (!pPrev)
3526	pSession->pDevExt->pComponentFactoryHead = pCur->pNext;
3527	else
3528	pPrev->pNext = pCur->pNext;
3529
3530	pCur->pNext = NULL;
3531	pCur->pFactory = NULL;
3532	pCur->pSession = NULL;
3533	rc = VINF_SUCCESS;
3534	}
3535	else
3536	rc = VERR_NOT_FOUND;
3537
3538	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3539
3540	RTMemFree(pCur);
3541	}
3542	return rc;
3543	}
3544
3545
3546	/**
3547	* Queries a component factory.
3548	*
3549	* @returns VBox status code.
3550	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3551	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3552	* @retval VERR_SUPDRV_COMPONENT_NOT_FOUND if the component factory wasn't found.
3553	* @retval VERR_SUPDRV_INTERFACE_NOT_SUPPORTED if the interface wasn't supported.
3554	*
3555	* @param pSession The SUPDRV session.
3556	* @param pszName The name of the component factory.
3557	* @param pszInterfaceUuid The UUID of the factory interface (stringified).
3558	* @param ppvFactoryIf Where to store the factory interface.
3559	*/
3560	SUPR0DECL(int) SUPR0ComponentQueryFactory(PSUPDRVSESSION pSession, const char pszName, const char pszInterfaceUuid, void **ppvFactoryIf)
3561	{
3562	const char *pszEnd;
3563	size_t cchName;
3564	int rc;
3565
3566	/*
3567	* Validate parameters.
3568	*/
3569	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3570
3571	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
3572	pszEnd = RTStrEnd(pszName, RT_SIZEOFMEMB(SUPDRVFACTORY, szName));
3573	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3574	cchName = pszEnd - pszName;
3575
3576	AssertPtrReturn(pszInterfaceUuid, VERR_INVALID_POINTER);
3577	pszEnd = RTStrEnd(pszInterfaceUuid, RTUUID_STR_LENGTH);
3578	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3579
3580	AssertPtrReturn(ppvFactoryIf, VERR_INVALID_POINTER);
3581	*ppvFactoryIf = NULL;
3582
3583	/*
3584	* Take the lock and try all factories by this name.
3585	*/
3586	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3587	if (RT_SUCCESS(rc))
3588	{
3589	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3590	rc = VERR_SUPDRV_COMPONENT_NOT_FOUND;
3591	while (pCur)
3592	{
3593	if ( pCur->cchName == cchName
3594	&& !memcmp(pCur->pFactory->szName, pszName, cchName))
3595	{
3596	void *pvFactory = pCur->pFactory->pfnQueryFactoryInterface(pCur->pFactory, pSession, pszInterfaceUuid);
3597	if (pvFactory)
3598	{
3599	*ppvFactoryIf = pvFactory;
3600	rc = VINF_SUCCESS;
3601	break;
3602	}
3603	rc = VERR_SUPDRV_INTERFACE_NOT_SUPPORTED;
3604	}
3605
3606	/* next */
3607	pCur = pCur->pNext;
3608	}
3609
3610	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3611	}
3612	return rc;
3613	}
3614
3615
3616	/**
3617	* Adds a memory object to the session.
3618	*
3619	* @returns IPRT status code.
3620	* @param pMem Memory tracking structure containing the
3621	* information to track.
3622	* @param pSession The session.
3623	*/
3624	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
3625	{
3626	PSUPDRVBUNDLE pBundle;
3627	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3628
3629	/*
3630	* Find free entry and record the allocation.
3631	*/
3632	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3633	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3634	{
3635	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
3636	{
3637	unsigned i;
3638	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3639	{
3640	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
3641	{
3642	pBundle->cUsed++;
3643	pBundle->aMem[i] = *pMem;
3644	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3645	return VINF_SUCCESS;
3646	}
3647	}
3648	AssertFailed(); /* !!this can't be happening!!! */
3649	}
3650	}
3651	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3652
3653	/*
3654	* Need to allocate a new bundle.
3655	* Insert into the last entry in the bundle.
3656	*/
3657	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
3658	if (!pBundle)
3659	return VERR_NO_MEMORY;
3660
3661	/* take last entry. */
3662	pBundle->cUsed++;
3663	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
3664
3665	/* insert into list. */
3666	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3667	pBundle->pNext = pSession->Bundle.pNext;
3668	pSession->Bundle.pNext = pBundle;
3669	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3670
3671	return VINF_SUCCESS;
3672	}
3673
3674
3675	/**
3676	* Releases a memory object referenced by pointer and type.
3677	*
3678	* @returns IPRT status code.
3679	* @param pSession Session data.
3680	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
3681	* @param eType Memory type.
3682	*/
3683	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
3684	{
3685	PSUPDRVBUNDLE pBundle;
3686	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3687
3688	/*
3689	* Validate input.
3690	*/
3691	if (!uPtr)
3692	{
3693	Log(("Illegal address %p\n", (void *)uPtr));
3694	return VERR_INVALID_PARAMETER;
3695	}
3696
3697	/*
3698	* Search for the address.
3699	*/
3700	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3701	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3702	{
3703	if (pBundle->cUsed > 0)
3704	{
3705	unsigned i;
3706	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3707	{
3708	if ( pBundle->aMem[i].eType == eType
3709	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3710	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
3711	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3712	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
3713	)
3714	{
3715	/* Make a copy of it and release it outside the spinlock. */
3716	SUPDRVMEMREF Mem = pBundle->aMem[i];
3717	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
3718	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
3719	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
3720	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3721
3722	if (Mem.MapObjR3 != NIL_RTR0MEMOBJ)
3723	{
3724	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
3725	AssertRC(rc); /** @todo figure out how to handle this. */
3726	}
3727	if (Mem.MemObj != NIL_RTR0MEMOBJ)
3728	{
3729	int rc = RTR0MemObjFree(Mem.MemObj, true /* fFreeMappings */);
3730	AssertRC(rc); /** @todo figure out how to handle this. */
3731	}
3732	return VINF_SUCCESS;
3733	}
3734	}
3735	}
3736	}
3737	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3738	Log(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
3739	return VERR_INVALID_PARAMETER;
3740	}
3741
3742
3743	/**
3744	* Opens an image. If it's the first time it's opened the call must upload
3745	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
3746	*
3747	* This is the 1st step of the loading.
3748	*
3749	* @returns IPRT status code.
3750	* @param pDevExt Device globals.
3751	* @param pSession Session data.
3752	* @param pReq The open request.
3753	*/
3754	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
3755	{
3756	int rc;
3757	PSUPDRVLDRIMAGE pImage;
3758	void *pv;
3759	size_t cchName = strlen(pReq->u.In.szName); /* (caller checked < 32). */
3760	LogFlow(("supdrvIOCtl_LdrOpen: szName=%s cbImageWithTabs=%d\n", pReq->u.In.szName, pReq->u.In.cbImageWithTabs));
3761
3762	/*
3763	* Check if we got an instance of the image already.
3764	*/
3765	supdrvLdrLock(pDevExt);
3766	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
3767	{
3768	if ( pImage->szName[cchName] == '\0'
3769	&& !memcmp(pImage->szName, pReq->u.In.szName, cchName))
3770	{
3771	/** @todo check cbImageBits and cbImageWithTabs here, if they differs that indicates that the images are different. */
3772	pImage->cUsage++;
3773	pReq->u.Out.pvImageBase = pImage->pvImage;
3774	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
3775	pReq->u.Out.fNativeLoader = pImage->fNative;
3776	supdrvLdrAddUsage(pSession, pImage);
3777	supdrvLdrUnlock(pDevExt);
3778	return VINF_SUCCESS;
3779	}
3780	}
3781	/* (not found - add it!) */
3782
3783	/*
3784	* Allocate memory.
3785	*/
3786	pv = RTMemAlloc(RT_OFFSETOF(SUPDRVLDRIMAGE, szName[cchName + 1]));
3787	if (!pv)
3788	{
3789	supdrvLdrUnlock(pDevExt);
3790	Log(("supdrvIOCtl_LdrOpen: RTMemAlloc() failed\n"));
3791	return /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_2;
3792	}
3793
3794	/*
3795	* Setup and link in the LDR stuff.
3796	*/
3797	pImage = (PSUPDRVLDRIMAGE)pv;
3798	pImage->pvImage = NULL;
3799	pImage->pvImageAlloc = NULL;
3800	pImage->cbImageWithTabs = pReq->u.In.cbImageWithTabs;
3801	pImage->cbImageBits = pReq->u.In.cbImageBits;
3802	pImage->cSymbols = 0;
3803	pImage->paSymbols = NULL;
3804	pImage->pachStrTab = NULL;
3805	pImage->cbStrTab = 0;
3806	pImage->pfnModuleInit = NULL;
3807	pImage->pfnModuleTerm = NULL;
3808	pImage->pfnServiceReqHandler = NULL;
3809	pImage->uState = SUP_IOCTL_LDR_OPEN;
3810	pImage->cUsage = 1;
3811	memcpy(pImage->szName, pReq->u.In.szName, cchName + 1);
3812
3813	/*
3814	* Try load it using the native loader, if that isn't supported, fall back
3815	* on the older method.
3816	*/
3817	pImage->fNative = true;
3818	rc = supdrvOSLdrOpen(pDevExt, pImage, pReq->u.In.szFilename);
3819	if (rc == VERR_NOT_SUPPORTED)
3820	{
3821	pImage->pvImageAlloc = RTMemExecAlloc(pImage->cbImageBits + 31);
3822	pImage->pvImage = RT_ALIGN_P(pImage->pvImageAlloc, 32);
3823	pImage->fNative = false;
3824	rc = pImage->pvImageAlloc ? VINF_SUCCESS : VERR_NO_EXEC_MEMORY;
3825	}
3826	if (RT_FAILURE(rc))
3827	{
3828	supdrvLdrUnlock(pDevExt);
3829	RTMemFree(pImage);
3830	Log(("supdrvIOCtl_LdrOpen(%s): failed - %Rrc\n", pReq->u.In.szName, rc));
3831	return rc;
3832	}
3833	Assert(VALID_PTR(pImage->pvImage) \|\| RT_FAILURE(rc));
3834
3835	/*
3836	* Link it.
3837	*/
3838	pImage->pNext = pDevExt->pLdrImages;
3839	pDevExt->pLdrImages = pImage;
3840
3841	supdrvLdrAddUsage(pSession, pImage);
3842
3843	pReq->u.Out.pvImageBase = pImage->pvImage;
3844	pReq->u.Out.fNeedsLoading = true;
3845	pReq->u.Out.fNativeLoader = pImage->fNative;
3846	supdrvLdrUnlock(pDevExt);
3847
3848	#if defined(RT_OS_WINDOWS) && defined(DEBUG)
3849	SUPR0Printf("VBoxDrv: windbg> .reload /f %s=%#p\n", pImage->szName, pImage->pvImage);
3850	#endif
3851	return VINF_SUCCESS;
3852	}
3853
3854
3855	/**
3856	* Worker that validates a pointer to an image entrypoint.
3857	*
3858	* @returns IPRT status code.
3859	* @param pDevExt The device globals.
3860	* @param pImage The loader image.
3861	* @param pv The pointer into the image.
3862	* @param fMayBeNull Whether it may be NULL.
3863	* @param pszWhat What is this entrypoint? (for logging)
3864	* @param pbImageBits The image bits prepared by ring-3.
3865	*
3866	* @remarks Will leave the lock on failure.
3867	*/
3868	static int supdrvLdrValidatePointer(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage, void *pv,
3869	bool fMayBeNull, const uint8_t pbImageBits, const char pszWhat)
3870	{
3871	if (!fMayBeNull \|\| pv)
3872	{
3873	if ((uintptr_t)pv - (uintptr_t)pImage->pvImage >= pImage->cbImageBits)
3874	{
3875	supdrvLdrUnlock(pDevExt);
3876	Log(("Out of range (%p LB %#x): %s=%p\n", pImage->pvImage, pImage->cbImageBits, pszWhat, pv));
3877	return VERR_INVALID_PARAMETER;
3878	}
3879
3880	if (pImage->fNative)
3881	{
3882	int rc = supdrvOSLdrValidatePointer(pDevExt, pImage, pv, pbImageBits);
3883	if (RT_FAILURE(rc))
3884	{
3885	supdrvLdrUnlock(pDevExt);
3886	Log(("Bad entry point address: %s=%p (rc=%Rrc)\n", pszWhat, pv, rc));
3887	return rc;
3888	}
3889	}
3890	}
3891	return VINF_SUCCESS;
3892	}
3893
3894
3895	/**
3896	* Loads the image bits.
3897	*
3898	* This is the 2nd step of the loading.
3899	*
3900	* @returns IPRT status code.
3901	* @param pDevExt Device globals.
3902	* @param pSession Session data.
3903	* @param pReq The request.
3904	*/
3905	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
3906	{
3907	PSUPDRVLDRUSAGE pUsage;
3908	PSUPDRVLDRIMAGE pImage;
3909	int rc;
3910	LogFlow(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImageWithBits=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImageWithTabs));
3911
3912	/*
3913	* Find the ldr image.
3914	*/
3915	supdrvLdrLock(pDevExt);
3916	pUsage = pSession->pLdrUsage;
3917	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3918	pUsage = pUsage->pNext;
3919	if (!pUsage)
3920	{
3921	supdrvLdrUnlock(pDevExt);
3922	Log(("SUP_IOCTL_LDR_LOAD: couldn't find image!\n"));
3923	return VERR_INVALID_HANDLE;
3924	}
3925	pImage = pUsage->pImage;
3926
3927	/*
3928	* Validate input.
3929	*/
3930	if ( pImage->cbImageWithTabs != pReq->u.In.cbImageWithTabs
3931	\|\| pImage->cbImageBits != pReq->u.In.cbImageBits)
3932	{
3933	supdrvLdrUnlock(pDevExt);
3934	Log(("SUP_IOCTL_LDR_LOAD: image size mismatch!! %d(prep) != %d(load) or %d != %d\n",
3935	pImage->cbImageWithTabs, pReq->u.In.cbImageWithTabs, pImage->cbImageBits, pReq->u.In.cbImageBits));
3936	return VERR_INVALID_HANDLE;
3937	}
3938
3939	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
3940	{
3941	unsigned uState = pImage->uState;
3942	supdrvLdrUnlock(pDevExt);
3943	if (uState != SUP_IOCTL_LDR_LOAD)
3944	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
3945	return VERR_ALREADY_LOADED;
3946	}
3947
3948	switch (pReq->u.In.eEPType)
3949	{
3950	case SUPLDRLOADEP_NOTHING:
3951	break;
3952
3953	case SUPLDRLOADEP_VMMR0:
3954	rc = supdrvLdrValidatePointer( pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0, false, pReq->u.In.abImage, "pvVMMR0");
3955	if (RT_SUCCESS(rc))
3956	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt, false, pReq->u.In.abImage, "pvVMMR0EntryInt");
3957	if (RT_SUCCESS(rc))
3958	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, false, pReq->u.In.abImage, "pvVMMR0EntryFast");
3959	if (RT_SUCCESS(rc))
3960	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx, false, pReq->u.In.abImage, "pvVMMR0EntryEx");
3961	if (RT_FAILURE(rc))
3962	return rc;
3963	break;
3964
3965	case SUPLDRLOADEP_SERVICE:
3966	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.Service.pfnServiceReq, false, pReq->u.In.abImage, "pfnServiceReq");
3967	if (RT_FAILURE(rc))
3968	return rc;
3969	if ( pReq->u.In.EP.Service.apvReserved[0] != NIL_RTR0PTR
3970	\|\| pReq->u.In.EP.Service.apvReserved[1] != NIL_RTR0PTR
3971	\|\| pReq->u.In.EP.Service.apvReserved[2] != NIL_RTR0PTR)
3972	{
3973	supdrvLdrUnlock(pDevExt);
3974	Log(("Out of range (%p LB %#x): apvReserved={%p,%p,%p} MBZ!\n",
3975	pImage->pvImage, pReq->u.In.cbImageWithTabs,
3976	pReq->u.In.EP.Service.apvReserved[0],
3977	pReq->u.In.EP.Service.apvReserved[1],
3978	pReq->u.In.EP.Service.apvReserved[2]));
3979	return VERR_INVALID_PARAMETER;
3980	}
3981	break;
3982
3983	default:
3984	supdrvLdrUnlock(pDevExt);
3985	Log(("Invalid eEPType=%d\n", pReq->u.In.eEPType));
3986	return VERR_INVALID_PARAMETER;
3987	}
3988
3989	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleInit, true, pReq->u.In.abImage, "pfnModuleInit");
3990	if (RT_FAILURE(rc))
3991	return rc;
3992	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleTerm, true, pReq->u.In.abImage, "pfnModuleTerm");
3993	if (RT_FAILURE(rc))
3994	return rc;
3995
3996	/*
3997	* Allocate and copy the tables.
3998	* (No need to do try/except as this is a buffered request.)
3999	*/
4000	pImage->cbStrTab = pReq->u.In.cbStrTab;
4001	if (pImage->cbStrTab)
4002	{
4003	pImage->pachStrTab = (char *)RTMemAlloc(pImage->cbStrTab);
4004	if (pImage->pachStrTab)
4005	memcpy(pImage->pachStrTab, &pReq->u.In.abImage[pReq->u.In.offStrTab], pImage->cbStrTab);
4006	else
4007	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_3;
4008	}
4009
4010	pImage->cSymbols = pReq->u.In.cSymbols;
4011	if (RT_SUCCESS(rc) && pImage->cSymbols)
4012	{
4013	size_t cbSymbols = pImage->cSymbols * sizeof(SUPLDRSYM);
4014	pImage->paSymbols = (PSUPLDRSYM)RTMemAlloc(cbSymbols);
4015	if (pImage->paSymbols)
4016	memcpy(pImage->paSymbols, &pReq->u.In.abImage[pReq->u.In.offSymbols], cbSymbols);
4017	else
4018	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_4;
4019	}
4020
4021	/*
4022	* Copy the bits / complete native loading.
4023	*/
4024	if (RT_SUCCESS(rc))
4025	{
4026	pImage->uState = SUP_IOCTL_LDR_LOAD;
4027	pImage->pfnModuleInit = pReq->u.In.pfnModuleInit;
4028	pImage->pfnModuleTerm = pReq->u.In.pfnModuleTerm;
4029
4030	if (pImage->fNative)
4031	rc = supdrvOSLdrLoad(pDevExt, pImage, pReq->u.In.abImage);
4032	else
4033	memcpy(pImage->pvImage, &pReq->u.In.abImage[0], pImage->cbImageBits);
4034	}
4035
4036	/*
4037	* Update any entry points.
4038	*/
4039	if (RT_SUCCESS(rc))
4040	{
4041	switch (pReq->u.In.eEPType)
4042	{
4043	default:
4044	case SUPLDRLOADEP_NOTHING:
4045	rc = VINF_SUCCESS;
4046	break;
4047	case SUPLDRLOADEP_VMMR0:
4048	rc = supdrvLdrSetVMMR0EPs(pDevExt, pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
4049	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
4050	break;
4051	case SUPLDRLOADEP_SERVICE:
4052	pImage->pfnServiceReqHandler = pReq->u.In.EP.Service.pfnServiceReq;
4053	rc = VINF_SUCCESS;
4054	break;
4055	}
4056	}
4057
4058	/*
4059	* On success call the module initialization.
4060	*/
4061	LogFlow(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
4062	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
4063	{
4064	Log(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
4065	rc = pImage->pfnModuleInit();
4066	if (rc && pDevExt->pvVMMR0 == pImage->pvImage)
4067	supdrvLdrUnsetVMMR0EPs(pDevExt);
4068	}
4069
4070	if (RT_FAILURE(rc))
4071	{
4072	pImage->uState = SUP_IOCTL_LDR_OPEN;
4073	pImage->pfnModuleInit = NULL;
4074	pImage->pfnModuleTerm = NULL;
4075	pImage->pfnServiceReqHandler= NULL;
4076	pImage->cbStrTab = 0;
4077	RTMemFree(pImage->pachStrTab);
4078	pImage->pachStrTab = NULL;
4079	RTMemFree(pImage->paSymbols);
4080	pImage->paSymbols = NULL;
4081	pImage->cSymbols = 0;
4082	}
4083
4084	supdrvLdrUnlock(pDevExt);
4085	return rc;
4086	}
4087
4088
4089	/**
4090	* Frees a previously loaded (prep'ed) image.
4091	*
4092	* @returns IPRT status code.
4093	* @param pDevExt Device globals.
4094	* @param pSession Session data.
4095	* @param pReq The request.
4096	*/
4097	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
4098	{
4099	int rc;
4100	PSUPDRVLDRUSAGE pUsagePrev;
4101	PSUPDRVLDRUSAGE pUsage;
4102	PSUPDRVLDRIMAGE pImage;
4103	LogFlow(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
4104
4105	/*
4106	* Find the ldr image.
4107	*/
4108	supdrvLdrLock(pDevExt);
4109	pUsagePrev = NULL;
4110	pUsage = pSession->pLdrUsage;
4111	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4112	{
4113	pUsagePrev = pUsage;
4114	pUsage = pUsage->pNext;
4115	}
4116	if (!pUsage)
4117	{
4118	supdrvLdrUnlock(pDevExt);
4119	Log(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
4120	return VERR_INVALID_HANDLE;
4121	}
4122
4123	/*
4124	* Check if we can remove anything.
4125	*/
4126	rc = VINF_SUCCESS;
4127	pImage = pUsage->pImage;
4128	if (pImage->cUsage <= 1 \|\| pUsage->cUsage <= 1)
4129	{
4130	/*
4131	* Check if there are any objects with destructors in the image, if
4132	* so leave it for the session cleanup routine so we get a chance to
4133	* clean things up in the right order and not leave them all dangling.
4134	*/
4135	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
4136	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
4137	if (pImage->cUsage <= 1)
4138	{
4139	PSUPDRVOBJ pObj;
4140	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4141	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4142	{
4143	rc = VERR_DANGLING_OBJECTS;
4144	break;
4145	}
4146	}
4147	else
4148	{
4149	PSUPDRVUSAGE pGenUsage;
4150	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
4151	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4152	{
4153	rc = VERR_DANGLING_OBJECTS;
4154	break;
4155	}
4156	}
4157	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
4158	if (rc == VINF_SUCCESS)
4159	{
4160	/* unlink it */
4161	if (pUsagePrev)
4162	pUsagePrev->pNext = pUsage->pNext;
4163	else
4164	pSession->pLdrUsage = pUsage->pNext;
4165
4166	/* free it */
4167	pUsage->pImage = NULL;
4168	pUsage->pNext = NULL;
4169	RTMemFree(pUsage);
4170
4171	/*
4172	* Dereference the image.
4173	*/
4174	if (pImage->cUsage <= 1)
4175	supdrvLdrFree(pDevExt, pImage);
4176	else
4177	pImage->cUsage--;
4178	}
4179	else
4180	{
4181	Log(("supdrvIOCtl_LdrFree: Dangling objects in %p/%s!\n", pImage->pvImage, pImage->szName));
4182	rc = VINF_SUCCESS; /** @todo BRANCH-2.1: remove this after branching. */
4183	}
4184	}
4185	else
4186	{
4187	/*
4188	* Dereference both image and usage.
4189	*/
4190	pImage->cUsage--;
4191	pUsage->cUsage--;
4192	}
4193
4194	supdrvLdrUnlock(pDevExt);
4195	return rc;
4196	}
4197
4198
4199	/**
4200	* Gets the address of a symbol in an open image.
4201	*
4202	* @returns IPRT status code.
4203	* @param pDevExt Device globals.
4204	* @param pSession Session data.
4205	* @param pReq The request buffer.
4206	*/
4207	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
4208	{
4209	PSUPDRVLDRIMAGE pImage;
4210	PSUPDRVLDRUSAGE pUsage;
4211	uint32_t i;
4212	PSUPLDRSYM paSyms;
4213	const char *pchStrings;
4214	const size_t cbSymbol = strlen(pReq->u.In.szSymbol) + 1;
4215	void *pvSymbol = NULL;
4216	int rc = VERR_GENERAL_FAILURE;
4217	Log3(("supdrvIOCtl_LdrGetSymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
4218
4219	/*
4220	* Find the ldr image.
4221	*/
4222	supdrvLdrLock(pDevExt);
4223	pUsage = pSession->pLdrUsage;
4224	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4225	pUsage = pUsage->pNext;
4226	if (!pUsage)
4227	{
4228	supdrvLdrUnlock(pDevExt);
4229	Log(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
4230	return VERR_INVALID_HANDLE;
4231	}
4232	pImage = pUsage->pImage;
4233	if (pImage->uState != SUP_IOCTL_LDR_LOAD)
4234	{
4235	unsigned uState = pImage->uState;
4236	supdrvLdrUnlock(pDevExt);
4237	Log(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", uState, uState)); NOREF(uState);
4238	return VERR_ALREADY_LOADED;
4239	}
4240
4241	/*
4242	* Search the symbol strings.
4243	*/
4244	pchStrings = pImage->pachStrTab;
4245	paSyms = pImage->paSymbols;
4246	for (i = 0; i < pImage->cSymbols; i++)
4247	{
4248	if ( paSyms[i].offSymbol < pImage->cbImageBits /* paranoia */
4249	&& paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4250	&& !memcmp(pchStrings + paSyms[i].offName, pReq->u.In.szSymbol, cbSymbol))
4251	{
4252	pvSymbol = (uint8_t *)pImage->pvImage + paSyms[i].offSymbol;
4253	rc = VINF_SUCCESS;
4254	break;
4255	}
4256	}
4257	supdrvLdrUnlock(pDevExt);
4258	pReq->u.Out.pvSymbol = pvSymbol;
4259	return rc;
4260	}
4261
4262
4263	/**
4264	* Gets the address of a symbol in an open image or the support driver.
4265	*
4266	* @returns VINF_SUCCESS on success.
4267	* @returns
4268	* @param pDevExt Device globals.
4269	* @param pSession Session data.
4270	* @param pReq The request buffer.
4271	*/
4272	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq)
4273	{
4274	int rc = VINF_SUCCESS;
4275	const char *pszSymbol = pReq->u.In.pszSymbol;
4276	const char *pszModule = pReq->u.In.pszModule;
4277	size_t cbSymbol;
4278	char const *pszEnd;
4279	uint32_t i;
4280
4281	/*
4282	* Input validation.
4283	*/
4284	AssertPtrReturn(pszSymbol, VERR_INVALID_POINTER);
4285	pszEnd = RTStrEnd(pszSymbol, 512);
4286	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4287	cbSymbol = pszEnd - pszSymbol + 1;
4288
4289	if (pszModule)
4290	{
4291	AssertPtrReturn(pszModule, VERR_INVALID_POINTER);
4292	pszEnd = RTStrEnd(pszModule, 64);
4293	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4294	}
4295	Log3(("supdrvIDC_LdrGetSymbol: pszModule=%p:{%s} pszSymbol=%p:{%s}\n", pszModule, pszModule, pszSymbol, pszSymbol));
4296
4297
4298	if ( !pszModule
4299	\|\| !strcmp(pszModule, "SupDrv"))
4300	{
4301	/*
4302	* Search the support driver export table.
4303	*/
4304	for (i = 0; i < RT_ELEMENTS(g_aFunctions); i++)
4305	if (!strcmp(g_aFunctions[i].szName, pszSymbol))
4306	{
4307	pReq->u.Out.pfnSymbol = g_aFunctions[i].pfn;
4308	break;
4309	}
4310	}
4311	else
4312	{
4313	/*
4314	* Find the loader image.
4315	*/
4316	PSUPDRVLDRIMAGE pImage;
4317
4318	supdrvLdrLock(pDevExt);
4319
4320	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
4321	if (!strcmp(pImage->szName, pszModule))
4322	break;
4323	if (pImage && pImage->uState == SUP_IOCTL_LDR_LOAD)
4324	{
4325	/*
4326	* Search the symbol strings.
4327	*/
4328	const char *pchStrings = pImage->pachStrTab;
4329	PCSUPLDRSYM paSyms = pImage->paSymbols;
4330	for (i = 0; i < pImage->cSymbols; i++)
4331	{
4332	if ( paSyms[i].offSymbol < pImage->cbImageBits /* paranoia */
4333	&& paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4334	&& !memcmp(pchStrings + paSyms[i].offName, pszSymbol, cbSymbol))
4335	{
4336	/*
4337	* Found it! Calc the symbol address and add a reference to the module.
4338	*/
4339	pReq->u.Out.pfnSymbol = (PFNRT)((uint8_t *)pImage->pvImage + paSyms[i].offSymbol);
4340	rc = supdrvLdrAddUsage(pSession, pImage);
4341	break;
4342	}
4343	}
4344	}
4345	else
4346	rc = pImage ? VERR_WRONG_ORDER : VERR_MODULE_NOT_FOUND;
4347
4348	supdrvLdrUnlock(pDevExt);
4349	}
4350	return rc;
4351	}
4352
4353
4354	/**
4355	* Updates the VMMR0 entry point pointers.
4356	*
4357	* @returns IPRT status code.
4358	* @param pDevExt Device globals.
4359	* @param pSession Session data.
4360	* @param pVMMR0 VMMR0 image handle.
4361	* @param pvVMMR0EntryInt VMMR0EntryInt address.
4362	* @param pvVMMR0EntryFast VMMR0EntryFast address.
4363	* @param pvVMMR0EntryEx VMMR0EntryEx address.
4364	* @remark Caller must own the loader mutex.
4365	*/
4366	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx)
4367	{
4368	int rc = VINF_SUCCESS;
4369	LogFlow(("supdrvLdrSetR0EP pvVMMR0=%p pvVMMR0EntryInt=%p\n", pvVMMR0, pvVMMR0EntryInt));
4370
4371
4372	/*
4373	* Check if not yet set.
4374	*/
4375	if (!pDevExt->pvVMMR0)
4376	{
4377	pDevExt->pvVMMR0 = pvVMMR0;
4378	pDevExt->pfnVMMR0EntryInt = pvVMMR0EntryInt;
4379	pDevExt->pfnVMMR0EntryFast = pvVMMR0EntryFast;
4380	pDevExt->pfnVMMR0EntryEx = pvVMMR0EntryEx;
4381	}
4382	else
4383	{
4384	/*
4385	* Return failure or success depending on whether the values match or not.
4386	*/
4387	if ( pDevExt->pvVMMR0 != pvVMMR0
4388	\|\| (void *)pDevExt->pfnVMMR0EntryInt != pvVMMR0EntryInt
4389	\|\| (void *)pDevExt->pfnVMMR0EntryFast != pvVMMR0EntryFast
4390	\|\| (void *)pDevExt->pfnVMMR0EntryEx != pvVMMR0EntryEx)
4391	{
4392	AssertMsgFailed(("SUP_IOCTL_LDR_SETR0EP: Already set pointing to a different module!\n"));
4393	rc = VERR_INVALID_PARAMETER;
4394	}
4395	}
4396	return rc;
4397	}
4398
4399
4400	/**
4401	* Unsets the VMMR0 entry point installed by supdrvLdrSetR0EP.
4402	*
4403	* @param pDevExt Device globals.
4404	*/
4405	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt)
4406	{
4407	pDevExt->pvVMMR0 = NULL;
4408	pDevExt->pfnVMMR0EntryInt = NULL;
4409	pDevExt->pfnVMMR0EntryFast = NULL;
4410	pDevExt->pfnVMMR0EntryEx = NULL;
4411	}
4412
4413
4414	/**
4415	* Adds a usage reference in the specified session of an image.
4416	*
4417	* Called while owning the loader semaphore.
4418	*
4419	* @returns VINF_SUCCESS on success and VERR_NO_MEMORY on failure.
4420	* @param pSession Session in question.
4421	* @param pImage Image which the session is using.
4422	*/
4423	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage)
4424	{
4425	PSUPDRVLDRUSAGE pUsage;
4426	LogFlow(("supdrvLdrAddUsage: pImage=%p\n", pImage));
4427
4428	/*
4429	* Referenced it already?
4430	*/
4431	pUsage = pSession->pLdrUsage;
4432	while (pUsage)
4433	{
4434	if (pUsage->pImage == pImage)
4435	{
4436	pUsage->cUsage++;
4437	return VINF_SUCCESS;
4438	}
4439	pUsage = pUsage->pNext;
4440	}
4441
4442	/*
4443	* Allocate new usage record.
4444	*/
4445	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
4446	AssertReturn(pUsage, /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_5);
4447	pUsage->cUsage = 1;
4448	pUsage->pImage = pImage;
4449	pUsage->pNext = pSession->pLdrUsage;
4450	pSession->pLdrUsage = pUsage;
4451	return VINF_SUCCESS;
4452	}
4453
4454
4455	/**
4456	* Frees a load image.
4457	*
4458	* @param pDevExt Pointer to device extension.
4459	* @param pImage Pointer to the image we're gonna free.
4460	* This image must exit!
4461	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
4462	*/
4463	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
4464	{
4465	PSUPDRVLDRIMAGE pImagePrev;
4466	LogFlow(("supdrvLdrFree: pImage=%p\n", pImage));
4467
4468	/* find it - arg. should've used doubly linked list. */
4469	Assert(pDevExt->pLdrImages);
4470	pImagePrev = NULL;
4471	if (pDevExt->pLdrImages != pImage)
4472	{
4473	pImagePrev = pDevExt->pLdrImages;
4474	while (pImagePrev->pNext != pImage)
4475	pImagePrev = pImagePrev->pNext;
4476	Assert(pImagePrev->pNext == pImage);
4477	}
4478
4479	/* unlink */
4480	if (pImagePrev)
4481	pImagePrev->pNext = pImage->pNext;
4482	else
4483	pDevExt->pLdrImages = pImage->pNext;
4484
4485	/* check if this is VMMR0.r0 unset its entry point pointers. */
4486	if (pDevExt->pvVMMR0 == pImage->pvImage)
4487	supdrvLdrUnsetVMMR0EPs(pDevExt);
4488
4489	/* check for objects with destructors in this image. (Shouldn't happen.) */
4490	if (pDevExt->pObjs)
4491	{
4492	unsigned cObjs = 0;
4493	PSUPDRVOBJ pObj;
4494	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
4495	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
4496	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4497	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4498	{
4499	pObj->pfnDestructor = NULL;
4500	cObjs++;
4501	}
4502	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
4503	if (cObjs)
4504	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
4505	}
4506
4507	/* call termination function if fully loaded. */
4508	if ( pImage->pfnModuleTerm
4509	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
4510	{
4511	LogFlow(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
4512	pImage->pfnModuleTerm();
4513	}
4514
4515	/* do native unload if appropriate. */
4516	if (pImage->fNative)
4517	supdrvOSLdrUnload(pDevExt, pImage);
4518
4519	/* free the image */
4520	pImage->cUsage = 0;
4521	pImage->pNext = 0;
4522	pImage->uState = SUP_IOCTL_LDR_FREE;
4523	RTMemExecFree(pImage->pvImageAlloc, pImage->cbImageBits + 31);
4524	pImage->pvImageAlloc = NULL;
4525	RTMemFree(pImage->pachStrTab);
4526	pImage->pachStrTab = NULL;
4527	RTMemFree(pImage->paSymbols);
4528	pImage->paSymbols = NULL;
4529	RTMemFree(pImage);
4530	}
4531
4532
4533	/**
4534	* Acquires the loader lock.
4535	*
4536	* @returns IPRT status code.
4537	* @param pDevExt The device extension.
4538	*/
4539	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt)
4540	{
4541	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4542	int rc = RTSemMutexRequest(pDevExt->mtxLdr, RT_INDEFINITE_WAIT);
4543	#else
4544	int rc = RTSemFastMutexRequest(pDevExt->mtxLdr);
4545	#endif
4546	AssertRC(rc);
4547	return rc;
4548	}
4549
4550
4551	/**
4552	* Releases the loader lock.
4553	*
4554	* @returns IPRT status code.
4555	* @param pDevExt The device extension.
4556	*/
4557	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt)
4558	{
4559	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4560	return RTSemMutexRelease(pDevExt->mtxLdr);
4561	#else
4562	return RTSemFastMutexRelease(pDevExt->mtxLdr);
4563	#endif
4564	}
4565
4566
4567	/**
4568	* Implements the service call request.
4569	*
4570	* @returns VBox status code.
4571	* @param pDevExt The device extension.
4572	* @param pSession The calling session.
4573	* @param pReq The request packet, valid.
4574	*/
4575	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq)
4576	{
4577	#if !defined(RT_OS_WINDOWS) \|\| defined(DEBUG)
4578	int rc;
4579
4580	/*
4581	* Find the module first in the module referenced by the calling session.
4582	*/
4583	rc = supdrvLdrLock(pDevExt);
4584	if (RT_SUCCESS(rc))
4585	{
4586	PFNSUPR0SERVICEREQHANDLER pfnServiceReqHandler = NULL;
4587	PSUPDRVLDRUSAGE pUsage;
4588
4589	for (pUsage = pSession->pLdrUsage; pUsage; pUsage = pUsage->pNext)
4590	if ( pUsage->pImage->pfnServiceReqHandler
4591	&& !strcmp(pUsage->pImage->szName, pReq->u.In.szName))
4592	{
4593	pfnServiceReqHandler = pUsage->pImage->pfnServiceReqHandler;
4594	break;
4595	}
4596	supdrvLdrUnlock(pDevExt);
4597
4598	if (pfnServiceReqHandler)
4599	{
4600	/*
4601	* Call it.
4602	*/
4603	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
4604	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, NULL);
4605	else
4606	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0]);
4607	}
4608	else
4609	rc = VERR_SUPDRV_SERVICE_NOT_FOUND;
4610	}
4611
4612	/* log it */
4613	if ( RT_FAILURE(rc)
4614	&& rc != VERR_INTERRUPTED
4615	&& rc != VERR_TIMEOUT)
4616	Log(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4617	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4618	else
4619	Log4(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4620	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4621	return rc;
4622	#else /* RT_OS_WINDOWS && !DEBUG */
4623	return VERR_NOT_IMPLEMENTED;
4624	#endif /* RT_OS_WINDOWS && !DEBUG */
4625	}
4626
4627
4628	/**
4629	* Implements the logger settings request.
4630	*
4631	* @returns VBox status code.
4632	* @param pDevExt The device extension.
4633	* @param pSession The caller's session.
4634	* @param pReq The request.
4635	*/
4636	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq)
4637	{
4638	const char *pszGroup = &pReq->u.In.szStrings[pReq->u.In.offGroups];
4639	const char *pszFlags = &pReq->u.In.szStrings[pReq->u.In.offFlags];
4640	const char *pszDest = &pReq->u.In.szStrings[pReq->u.In.offDestination];
4641	PRTLOGGER pLogger = NULL;
4642	int rc;
4643
4644	/*
4645	* Some further validation.
4646	*/
4647	switch (pReq->u.In.fWhat)
4648	{
4649	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4650	case SUPLOGGERSETTINGS_WHAT_CREATE:
4651	break;
4652
4653	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4654	if (pszGroup \|\| pszFlags \|\| *pszDest)
4655	return VERR_INVALID_PARAMETER;
4656	if (pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_RELEASE)
4657	return VERR_ACCESS_DENIED;
4658	break;
4659
4660	default:
4661	return VERR_INTERNAL_ERROR;
4662	}
4663
4664	/*
4665	* Get the logger.
4666	*/
4667	switch (pReq->u.In.fWhich)
4668	{
4669	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4670	pLogger = RTLogGetDefaultInstance();
4671	break;
4672
4673	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4674	pLogger = RTLogRelDefaultInstance();
4675	break;
4676
4677	default:
4678	return VERR_INTERNAL_ERROR;
4679	}
4680
4681	/*
4682	* Do the job.
4683	*/
4684	switch (pReq->u.In.fWhat)
4685	{
4686	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4687	if (pLogger)
4688	{
4689	rc = RTLogFlags(pLogger, pszFlags);
4690	if (RT_SUCCESS(rc))
4691	rc = RTLogGroupSettings(pLogger, pszGroup);
4692	NOREF(pszDest);
4693	}
4694	else
4695	rc = VERR_NOT_FOUND;
4696	break;
4697
4698	case SUPLOGGERSETTINGS_WHAT_CREATE:
4699	{
4700	if (pLogger)
4701	rc = VERR_ALREADY_EXISTS;
4702	else
4703	{
4704	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
4705
4706	rc = RTLogCreate(&pLogger,
4707	0 /* fFlags */,
4708	pszGroup,
4709	pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_DEBUG
4710	? "VBOX_LOG"
4711	: "VBOX_RELEASE_LOG",
4712	RT_ELEMENTS(s_apszGroups),
4713	s_apszGroups,
4714	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER,
4715	NULL);
4716	if (RT_SUCCESS(rc))
4717	{
4718	rc = RTLogFlags(pLogger, pszFlags);
4719	NOREF(pszDest);
4720	if (RT_SUCCESS(rc))
4721	{
4722	switch (pReq->u.In.fWhich)
4723	{
4724	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4725	pLogger = RTLogSetDefaultInstance(pLogger);
4726	break;
4727	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4728	pLogger = RTLogRelSetDefaultInstance(pLogger);
4729	break;
4730	}
4731	}
4732	RTLogDestroy(pLogger);
4733	}
4734	}
4735	break;
4736	}
4737
4738	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4739	switch (pReq->u.In.fWhich)
4740	{
4741	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4742	pLogger = RTLogSetDefaultInstance(NULL);
4743	break;
4744	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4745	pLogger = RTLogRelSetDefaultInstance(NULL);
4746	break;
4747	}
4748	rc = RTLogDestroy(pLogger);
4749	break;
4750
4751	default:
4752	{
4753	rc = VERR_INTERNAL_ERROR;
4754	break;
4755	}
4756	}
4757
4758	return rc;
4759	}
4760
4761
4762	/**
4763	* Creates the GIP.
4764	*
4765	* @returns VBox status code.
4766	* @param pDevExt Instance data. GIP stuff may be updated.
4767	*/
4768	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
4769	{
4770	PSUPGLOBALINFOPAGE pGip;
4771	RTHCPHYS HCPhysGip;
4772	uint32_t u32SystemResolution;
4773	uint32_t u32Interval;
4774	int rc;
4775
4776	LogFlow(("supdrvGipCreate:\n"));
4777
4778	/* assert order */
4779	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
4780	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
4781	Assert(!pDevExt->pGipTimer);
4782
4783	/*
4784	* Allocate a suitable page with a default kernel mapping.
4785	*/
4786	rc = RTR0MemObjAllocLow(&pDevExt->GipMemObj, PAGE_SIZE, false);
4787	if (RT_FAILURE(rc))
4788	{
4789	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
4790	return rc;
4791	}
4792	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
4793	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
4794
4795	/*
4796	* Find a reasonable update interval and initialize the structure.
4797	*/
4798	u32Interval = u32SystemResolution = RTTimerGetSystemGranularity();
4799	while (u32Interval < 10000000 /* 10 ms */)
4800	u32Interval += u32SystemResolution;
4801
4802	supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), 1000000000 / u32Interval /=Hz/);
4803
4804	/*
4805	* Create the timer.
4806	* If CPU_ALL isn't supported we'll have to fall back to synchronous mode.
4807	*/
4808	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
4809	{
4810	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, RTTIMER_FLAGS_CPU_ALL, supdrvGipAsyncTimer, pDevExt);
4811	if (rc == VERR_NOT_SUPPORTED)
4812	{
4813	OSDBGPRINT(("supdrvGipCreate: omni timer not supported, falling back to synchronous mode\n"));
4814	pGip->u32Mode = SUPGIPMODE_SYNC_TSC;
4815	}
4816	}
4817	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
4818	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0, supdrvGipSyncTimer, pDevExt);
4819	if (RT_SUCCESS(rc))
4820	{
4821	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
4822	rc = RTMpNotificationRegister(supdrvGipMpEvent, pDevExt);
4823	if (RT_SUCCESS(rc))
4824	{
4825	/*
4826	* We're good.
4827	*/
4828	Log(("supdrvGipCreate: %ld ns interval.\n", (long)u32Interval));
4829	g_pSUPGlobalInfoPage = pGip;
4830	return VINF_SUCCESS;
4831	}
4832
4833	OSDBGPRINT(("supdrvGipCreate: failed register MP event notfication. rc=%d\n", rc));
4834	}
4835	else
4836	{
4837	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %ld ns interval. rc=%d\n", (long)u32Interval, rc));
4838	Assert(!pDevExt->pGipTimer);
4839	}
4840	supdrvGipDestroy(pDevExt);
4841	return rc;
4842	}
4843
4844
4845	/**
4846	* Terminates the GIP.
4847	*
4848	* @param pDevExt Instance data. GIP stuff may be updated.
4849	*/
4850	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
4851	{
4852	int rc;
4853	#ifdef DEBUG_DARWIN_GIP
4854	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
4855	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
4856	pDevExt->pGipTimer, pDevExt->GipMemObj));
4857	#endif
4858
4859	/*
4860	* Invalid the GIP data.
4861	*/
4862	if (pDevExt->pGip)
4863	{
4864	supdrvGipTerm(pDevExt->pGip);
4865	pDevExt->pGip = NULL;
4866	}
4867	g_pSUPGlobalInfoPage = NULL;
4868
4869	/*
4870	* Destroy the timer and free the GIP memory object.
4871	*/
4872	if (pDevExt->pGipTimer)
4873	{
4874	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
4875	pDevExt->pGipTimer = NULL;
4876	}
4877
4878	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
4879	{
4880	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
4881	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
4882	}
4883
4884	/*
4885	* Finally, make sure we've release the system timer resolution request
4886	* if one actually succeeded and is still pending.
4887	*/
4888	if (pDevExt->u32SystemTimerGranularityGrant)
4889	{
4890	rc = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
4891	pDevExt->u32SystemTimerGranularityGrant = 0;
4892	}
4893	}
4894
4895
4896	/**
4897	* Timer callback function sync GIP mode.
4898	* @param pTimer The timer.
4899	* @param pvUser The device extension.
4900	*/
4901	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
4902	{
4903	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
4904	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4905	uint64_t u64TSC = ASMReadTSC();
4906	uint64_t NanoTS = RTTimeSystemNanoTS();
4907
4908	supdrvGipUpdate(pDevExt->pGip, NanoTS, u64TSC, iTick);
4909
4910	ASMSetFlags(fOldFlags);
4911	}
4912
4913
4914	/**
4915	* Timer callback function for async GIP mode.
4916	* @param pTimer The timer.
4917	* @param pvUser The device extension.
4918	*/
4919	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
4920	{
4921	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
4922	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4923	RTCPUID idCpu = RTMpCpuId();
4924	uint64_t u64TSC = ASMReadTSC();
4925	uint64_t NanoTS = RTTimeSystemNanoTS();
4926
4927	/** @todo reset the transaction number and whatnot when iTick == 1. */
4928	if (pDevExt->idGipMaster == idCpu)
4929	supdrvGipUpdate(pDevExt->pGip, NanoTS, u64TSC, iTick);
4930	else
4931	supdrvGipUpdatePerCpu(pDevExt->pGip, NanoTS, u64TSC, ASMGetApicId(), iTick);
4932
4933	ASMSetFlags(fOldFlags);
4934	}
4935
4936
4937	/**
4938	* Multiprocessor event notification callback.
4939	*
4940	* This is used to make sue that the GIP master gets passed on to
4941	* another CPU.
4942	*
4943	* @param enmEvent The event.
4944	* @param idCpu The cpu it applies to.
4945	* @param pvUser Pointer to the device extension.
4946	*/
4947	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser)
4948	{
4949	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4950	if (enmEvent == RTMPEVENT_OFFLINE)
4951	{
4952	RTCPUID idGipMaster;
4953	ASMAtomicReadSize(&pDevExt->idGipMaster, &idGipMaster);
4954	if (idGipMaster == idCpu)
4955	{
4956	/*
4957	* Find a new GIP master.
4958	*/
4959	bool fIgnored;
4960	unsigned i;
4961	RTCPUID idNewGipMaster = NIL_RTCPUID;
4962	RTCPUSET OnlineCpus;
4963	RTMpGetOnlineSet(&OnlineCpus);
4964
4965	for (i = 0; i < RTCPUSET_MAX_CPUS; i++)
4966	{
4967	RTCPUID idCurCpu = RTMpCpuIdFromSetIndex(i);
4968	if ( RTCpuSetIsMember(&OnlineCpus, idCurCpu)
4969	&& idCurCpu != idGipMaster)
4970	{
4971	idNewGipMaster = idCurCpu;
4972	break;
4973	}
4974	}
4975
4976	Log(("supdrvGipMpEvent: Gip master %#lx -> %#lx\n", (long)idGipMaster, (long)idNewGipMaster));
4977	ASMAtomicCmpXchgSize(&pDevExt->idGipMaster, idNewGipMaster, idGipMaster, fIgnored);
4978	NOREF(fIgnored);
4979	}
4980	}
4981	}
4982
4983
4984	/**
4985	* Callback used by supdrvDetermineAsyncTSC to read the TSC on a CPU.
4986	*
4987	* @param idCpu Ignored.
4988	* @param pvUser1 Where to put the TSC.
4989	* @param pvUser2 Ignored.
4990	*/
4991	static DECLCALLBACK(void) supdrvDetermineAsyncTscWorker(RTCPUID idCpu, void pvUser1, void pvUser2)
4992	{
4993	#if 1
4994	ASMAtomicWriteU64((uint64_t volatile *)pvUser1, ASMReadTSC());
4995	#else
4996	(uint64_t )pvUser1 = ASMReadTSC();
4997	#endif
4998	}
4999
5000
5001	/**
5002	* Determine if Async GIP mode is required because of TSC drift.
5003	*
5004	* When using the default/normal timer code it is essential that the time stamp counter
5005	* (TSC) runs never backwards, that is, a read operation to the counter should return
5006	* a bigger value than any previous read operation. This is guaranteed by the latest
5007	* AMD CPUs and by newer Intel CPUs which never enter the C2 state (P4). In any other
5008	* case we have to choose the asynchronous timer mode.
5009	*
5010	* @param poffMin Pointer to the determined difference between different cores.
5011	* @return false if the time stamp counters appear to be synchronized, true otherwise.
5012	*/
5013	static bool supdrvDetermineAsyncTsc(uint64_t *poffMin)
5014	{
5015	/*
5016	* Just iterate all the cpus 8 times and make sure that the TSC is
5017	* ever increasing. We don't bother taking TSC rollover into account.
5018	*/
5019	int iEndCpu = RTMpGetArraySize();
5020	int iCpu;
5021	int cLoops = 8;
5022	bool fAsync = false;
5023	int rc = VINF_SUCCESS;
5024	uint64_t offMax = 0;
5025	uint64_t offMin = ~(uint64_t)0;
5026	uint64_t PrevTsc = ASMReadTSC();
5027
5028	while (cLoops-- > 0)
5029	{
5030	for (iCpu = 0; iCpu < iEndCpu; iCpu++)
5031	{
5032	uint64_t CurTsc;
5033	rc = RTMpOnSpecific(RTMpCpuIdFromSetIndex(iCpu), supdrvDetermineAsyncTscWorker, &CurTsc, NULL);
5034	if (RT_SUCCESS(rc))
5035	{
5036	if (CurTsc <= PrevTsc)
5037	{
5038	fAsync = true;
5039	offMin = offMax = PrevTsc - CurTsc;
5040	Log(("supdrvDetermineAsyncTsc: iCpu=%d cLoops=%d CurTsc=%llx PrevTsc=%llx\n",
5041	iCpu, cLoops, CurTsc, PrevTsc));
5042	break;
5043	}
5044
5045	/* Gather statistics (except the first time). */
5046	if (iCpu != 0 \|\| cLoops != 7)
5047	{
5048	uint64_t off = CurTsc - PrevTsc;
5049	if (off < offMin)
5050	offMin = off;
5051	if (off > offMax)
5052	offMax = off;
5053	Log2(("%d/%d: off=%llx\n", cLoops, iCpu, off));
5054	}
5055
5056	/* Next */
5057	PrevTsc = CurTsc;
5058	}
5059	else if (rc == VERR_NOT_SUPPORTED)
5060	break;
5061	else
5062	AssertMsg(rc == VERR_CPU_NOT_FOUND \|\| rc == VERR_CPU_OFFLINE, ("%d\n", rc));
5063	}
5064
5065	/* broke out of the loop. */
5066	if (iCpu < iEndCpu)
5067	break;
5068	}
5069
5070	poffMin = offMin; / Almost RTMpOnSpecific profiling. */
5071	Log(("supdrvDetermineAsyncTsc: returns %d; iEndCpu=%d rc=%d offMin=%llx offMax=%llx\n",
5072	fAsync, iEndCpu, rc, offMin, offMax));
5073	#if !defined(RT_OS_SOLARIS) && !defined(RT_OS_OS2) && !defined(RT_OS_WINDOWS)
5074	OSDBGPRINT(("vboxdrv: fAsync=%d offMin=%#lx offMax=%#lx\n", fAsync, (long)offMin, (long)offMax));
5075	#endif
5076	return fAsync;
5077	}
5078
5079
5080	/**
5081	* Determine the GIP TSC mode.
5082	*
5083	* @returns The most suitable TSC mode.
5084	* @param pDevExt Pointer to the device instance data.
5085	*/
5086	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt)
5087	{
5088	/*
5089	* On SMP we're faced with two problems:
5090	* (1) There might be a skew between the CPU, so that cpu0
5091	* returns a TSC that is slightly different from cpu1.
5092	* (2) Power management (and other things) may cause the TSC
5093	* to run at a non-constant speed, and cause the speed
5094	* to be different on the cpus. This will result in (1).
5095	*
5096	* So, on SMP systems we'll have to select the ASYNC update method
5097	* if there are symptoms of these problems.
5098	*/
5099	if (RTMpGetCount() > 1)
5100	{
5101	uint32_t uEAX, uEBX, uECX, uEDX;
5102	uint64_t u64DiffCoresIgnored;
5103
5104	/* Permit the user and/or the OS specific bits to force async mode. */
5105	if (supdrvOSGetForcedAsyncTscMode(pDevExt))
5106	return SUPGIPMODE_ASYNC_TSC;
5107
5108	/* Try check for current differences between the cpus. */
5109	if (supdrvDetermineAsyncTsc(&u64DiffCoresIgnored))
5110	return SUPGIPMODE_ASYNC_TSC;
5111
5112	/*
5113	* If the CPU supports power management and is an AMD one we
5114	* won't trust it unless it has the TscInvariant bit is set.
5115	*/
5116	/* Check for "AuthenticAMD" */
5117	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
5118	if ( uEAX >= 1
5119	&& uEBX == X86_CPUID_VENDOR_AMD_EBX
5120	&& uECX == X86_CPUID_VENDOR_AMD_ECX
5121	&& uEDX == X86_CPUID_VENDOR_AMD_EDX)
5122	{
5123	/* Check for APM support and that TscInvariant is cleared. */
5124	ASMCpuId(0x80000000, &uEAX, &uEBX, &uECX, &uEDX);
5125	if (uEAX >= 0x80000007)
5126	{
5127	ASMCpuId(0x80000007, &uEAX, &uEBX, &uECX, &uEDX);
5128	if ( !(uEDX & RT_BIT(8))/* TscInvariant */
5129	&& (uEDX & 0x3e)) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
5130	return SUPGIPMODE_ASYNC_TSC;
5131	}
5132	}
5133	}
5134	return SUPGIPMODE_SYNC_TSC;
5135	}
5136
5137
5138
5139	/**
5140	* Initializes the GIP data.
5141	*
5142	* @param pDevExt Pointer to the device instance data.
5143	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5144	* @param HCPhys The physical address of the GIP.
5145	* @param u64NanoTS The current nanosecond timestamp.
5146	* @param uUpdateHz The update frequency.
5147	*/
5148	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys, uint64_t u64NanoTS, unsigned uUpdateHz)
5149	{
5150	unsigned i;
5151	#ifdef DEBUG_DARWIN_GIP
5152	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz));
5153	#else
5154	LogFlow(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz));
5155	#endif
5156
5157	/*
5158	* Initialize the structure.
5159	*/
5160	memset(pGip, 0, PAGE_SIZE);
5161	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
5162	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
5163	pGip->u32Mode = supdrvGipDeterminTscMode(pDevExt);
5164	pGip->u32UpdateHz = uUpdateHz;
5165	pGip->u32UpdateIntervalNS = 1000000000 / uUpdateHz;
5166	pGip->u64NanoTSLastUpdateHz = u64NanoTS;
5167
5168	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
5169	{
5170	pGip->aCPUs[i].u32TransactionId = 2;
5171	pGip->aCPUs[i].u64NanoTS = u64NanoTS;
5172	pGip->aCPUs[i].u64TSC = ASMReadTSC();
5173
5174	/*
5175	* We don't know the following values until we've executed updates.
5176	* So, we'll just pretend it's a 4 GHz CPU and adjust the history it on
5177	* the 2nd timer callout.
5178	*/
5179	pGip->aCPUs[i].u64CpuHz = _4G + 1; /* tstGIP-2 depends on this. */
5180	pGip->aCPUs[i].u32UpdateIntervalTSC
5181	= pGip->aCPUs[i].au32TSCHistory[0]
5182	= pGip->aCPUs[i].au32TSCHistory[1]
5183	= pGip->aCPUs[i].au32TSCHistory[2]
5184	= pGip->aCPUs[i].au32TSCHistory[3]
5185	= pGip->aCPUs[i].au32TSCHistory[4]
5186	= pGip->aCPUs[i].au32TSCHistory[5]
5187	= pGip->aCPUs[i].au32TSCHistory[6]
5188	= pGip->aCPUs[i].au32TSCHistory[7]
5189	= /pGip->aCPUs[i].u64CpuHz/ (uint32_t)(_4G / uUpdateHz);
5190	}
5191
5192	/*
5193	* Link it to the device extension.
5194	*/
5195	pDevExt->pGip = pGip;
5196	pDevExt->HCPhysGip = HCPhys;
5197	pDevExt->cGipUsers = 0;
5198	}
5199
5200
5201	/**
5202	* Invalidates the GIP data upon termination.
5203	*
5204	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5205	*/
5206	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
5207	{
5208	unsigned i;
5209	pGip->u32Magic = 0;
5210	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
5211	{
5212	pGip->aCPUs[i].u64NanoTS = 0;
5213	pGip->aCPUs[i].u64TSC = 0;
5214	pGip->aCPUs[i].iTSCHistoryHead = 0;
5215	}
5216	}
5217
5218
5219	/**
5220	* Worker routine for supdrvGipUpdate and supdrvGipUpdatePerCpu that
5221	* updates all the per cpu data except the transaction id.
5222	*
5223	* @param pGip The GIP.
5224	* @param pGipCpu Pointer to the per cpu data.
5225	* @param u64NanoTS The current time stamp.
5226	* @param u64TSC The current TSC.
5227	* @param iTick The current timer tick.
5228	*/
5229	static void supdrvGipDoUpdateCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick)
5230	{
5231	uint64_t u64TSCDelta;
5232	uint32_t u32UpdateIntervalTSC;
5233	uint32_t u32UpdateIntervalTSCSlack;
5234	unsigned iTSCHistoryHead;
5235	uint64_t u64CpuHz;
5236	uint32_t u32TransactionId;
5237
5238	/* Delta between this and the previous update. */
5239	ASMAtomicUoWriteU32(&pGipCpu->u32PrevUpdateIntervalNS, (uint32_t)(u64NanoTS - pGipCpu->u64NanoTS));
5240
5241	/*
5242	* Update the NanoTS.
5243	*/
5244	ASMAtomicXchgU64(&pGipCpu->u64NanoTS, u64NanoTS);
5245
5246	/*
5247	* Calc TSC delta.
5248	*/
5249	/** @todo validate the NanoTS delta, don't trust the OS to call us when it should... */
5250	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
5251	ASMAtomicXchgU64(&pGipCpu->u64TSC, u64TSC);
5252
5253	if (u64TSCDelta >> 32)
5254	{
5255	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
5256	pGipCpu->cErrors++;
5257	}
5258
5259	/*
5260	* On the 2nd and 3rd callout, reset the history with the current TSC
5261	* interval since the values entered by supdrvGipInit are totally off.
5262	* The interval on the 1st callout completely unreliable, the 2nd is a bit
5263	* better, while the 3rd should be most reliable.
5264	*/
5265	u32TransactionId = pGipCpu->u32TransactionId;
5266	if (RT_UNLIKELY( ( u32TransactionId == 5
5267	\|\| u32TransactionId == 7)
5268	&& ( iTick == 2
5269	\|\| iTick == 3) ))
5270	{
5271	unsigned i;
5272	for (i = 0; i < RT_ELEMENTS(pGipCpu->au32TSCHistory); i++)
5273	ASMAtomicUoWriteU32(&pGipCpu->au32TSCHistory[i], (uint32_t)u64TSCDelta);
5274	}
5275
5276	/*
5277	* TSC History.
5278	*/
5279	Assert(RT_ELEMENTS(pGipCpu->au32TSCHistory) == 8);
5280	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
5281	ASMAtomicXchgU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
5282	ASMAtomicXchgU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
5283
5284	/*
5285	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
5286	*/
5287	if (pGip->u32UpdateHz >= 1000)
5288	{
5289	uint32_t u32;
5290	u32 = pGipCpu->au32TSCHistory[0];
5291	u32 += pGipCpu->au32TSCHistory[1];
5292	u32 += pGipCpu->au32TSCHistory[2];
5293	u32 += pGipCpu->au32TSCHistory[3];
5294	u32 >>= 2;
5295	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
5296	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
5297	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
5298	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
5299	u32UpdateIntervalTSC >>= 2;
5300	u32UpdateIntervalTSC += u32;
5301	u32UpdateIntervalTSC >>= 1;
5302
5303	/* Value chosen for a 2GHz Athlon64 running linux 2.6.10/11, . */
5304	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
5305	}
5306	else if (pGip->u32UpdateHz >= 90)
5307	{
5308	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5309	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
5310	u32UpdateIntervalTSC >>= 1;
5311
5312	/* value chosen on a 2GHz thinkpad running windows */
5313	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
5314	}
5315	else
5316	{
5317	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5318
5319	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
5320	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
5321	}
5322	ASMAtomicXchgU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
5323
5324	/*
5325	* CpuHz.
5326	*/
5327	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, pGip->u32UpdateHz);
5328	ASMAtomicXchgU64(&pGipCpu->u64CpuHz, u64CpuHz);
5329	}
5330
5331
5332	/**
5333	* Updates the GIP.
5334	*
5335	* @param pGip Pointer to the GIP.
5336	* @param u64NanoTS The current nanosecond timesamp.
5337	* @param u64TSC The current TSC timesamp.
5338	* @param iTick The current timer tick.
5339	*/
5340	static void supdrvGipUpdate(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick)
5341	{
5342	/*
5343	* Determine the relevant CPU data.
5344	*/
5345	PSUPGIPCPU pGipCpu;
5346	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
5347	pGipCpu = &pGip->aCPUs[0];
5348	else
5349	{
5350	unsigned iCpu = ASMGetApicId();
5351	if (RT_UNLIKELY(iCpu >= RT_ELEMENTS(pGip->aCPUs)))
5352	return;
5353	pGipCpu = &pGip->aCPUs[iCpu];
5354	}
5355
5356	/*
5357	* Start update transaction.
5358	*/
5359	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5360	{
5361	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
5362	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5363	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5364	pGipCpu->cErrors++;
5365	return;
5366	}
5367
5368	/*
5369	* Recalc the update frequency every 0x800th time.
5370	*/
5371	if (!(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
5372	{
5373	if (pGip->u64NanoTSLastUpdateHz)
5374	{
5375	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
5376	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
5377	uint32_t u32UpdateHz = (uint32_t)((UINT64_C(1000000000) * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
5378	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
5379	{
5380	ASMAtomicXchgU32(&pGip->u32UpdateHz, u32UpdateHz);
5381	ASMAtomicXchgU32(&pGip->u32UpdateIntervalNS, 1000000000 / u32UpdateHz);
5382	}
5383	#endif
5384	}
5385	ASMAtomicXchgU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS);
5386	}
5387
5388	/*
5389	* Update the data.
5390	*/
5391	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS, u64TSC, iTick);
5392
5393	/*
5394	* Complete transaction.
5395	*/
5396	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5397	}
5398
5399
5400	/**
5401	* Updates the per cpu GIP data for the calling cpu.
5402	*
5403	* @param pGip Pointer to the GIP.
5404	* @param u64NanoTS The current nanosecond timesamp.
5405	* @param u64TSC The current TSC timesamp.
5406	* @param iCpu The CPU index.
5407	* @param iTick The current timer tick.
5408	*/
5409	static void supdrvGipUpdatePerCpu(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC, unsigned iCpu, uint64_t iTick)
5410	{
5411	PSUPGIPCPU pGipCpu;
5412
5413	if (RT_LIKELY(iCpu < RT_ELEMENTS(pGip->aCPUs)))
5414	{
5415	pGipCpu = &pGip->aCPUs[iCpu];
5416
5417	/*
5418	* Start update transaction.
5419	*/
5420	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5421	{
5422	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5423	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5424	pGipCpu->cErrors++;
5425	return;
5426	}
5427
5428	/*
5429	* Update the data.
5430	*/
5431	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS, u64TSC, iTick);
5432
5433	/*
5434	* Complete transaction.
5435	*/
5436	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5437	}
5438	}
5439

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

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDrv.c@ 36191

Download in other formats: