SUPDrv.c@ 34484

Last change on this file since 34484 was 34321, checked in by vboxsync, 14 years ago
SUPDrv.c: Set the windows timer granularity to 1024, 1000, 512 or 500 HZ (higher is prefered) while VMs are running (tied to GIP actually).
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 208.6 KB

Line
1	/* $Revision: 34321 $ */
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/hwacc_svm.h>
62	#include <VBox/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_CREATE):
1598	{
1599	/* validate */
1600	PSUPSEMCREATE pReq = (PSUPSEMCREATE)pReqHdr;
1601	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_CREATE, SUP_IOCTL_SEM_CREATE_SIZE_IN, SUP_IOCTL_SEM_CREATE_SIZE_OUT);
1602
1603	/* execute */
1604	switch (pReq->u.In.uType)
1605	{
1606	case SUP_SEM_TYPE_EVENT:
1607	{
1608	SUPSEMEVENT hEvent;
1609	pReq->Hdr.rc = SUPSemEventCreate(pSession, &hEvent);
1610	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEvent;
1611	break;
1612	}
1613
1614	case SUP_SEM_TYPE_EVENT_MULTI:
1615	{
1616	SUPSEMEVENTMULTI hEventMulti;
1617	pReq->Hdr.rc = SUPSemEventMultiCreate(pSession, &hEventMulti);
1618	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEventMulti;
1619	break;
1620	}
1621
1622	default:
1623	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1624	break;
1625	}
1626	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1627	return 0;
1628	}
1629
1630	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP):
1631	{
1632	/* validate */
1633	PSUPSEMOP pReq = (PSUPSEMOP)pReqHdr;
1634	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP, SUP_IOCTL_SEM_OP_SIZE_IN, SUP_IOCTL_SEM_OP_SIZE_OUT);
1635
1636	/* execute */
1637	switch (pReq->u.In.uType)
1638	{
1639	case SUP_SEM_TYPE_EVENT:
1640	{
1641	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1642	switch (pReq->u.In.uOp)
1643	{
1644	case SUPSEMOP_WAIT:
1645	pReq->Hdr.rc = SUPSemEventWaitNoResume(pSession, hEvent, pReq->u.In.cMillies);
1646	break;
1647	case SUPSEMOP_SIGNAL:
1648	pReq->Hdr.rc = SUPSemEventSignal(pSession, hEvent);
1649	break;
1650	case SUPSEMOP_CLOSE:
1651	pReq->Hdr.rc = SUPSemEventClose(pSession, hEvent);
1652	break;
1653	case SUPSEMOP_RESET:
1654	default:
1655	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1656	break;
1657	}
1658	break;
1659	}
1660
1661	case SUP_SEM_TYPE_EVENT_MULTI:
1662	{
1663	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1664	switch (pReq->u.In.uOp)
1665	{
1666	case SUPSEMOP_WAIT:
1667	pReq->Hdr.rc = SUPSemEventMultiWaitNoResume(pSession, hEventMulti, pReq->u.In.cMillies);
1668	break;
1669	case SUPSEMOP_SIGNAL:
1670	pReq->Hdr.rc = SUPSemEventMultiSignal(pSession, hEventMulti);
1671	break;
1672	case SUPSEMOP_CLOSE:
1673	pReq->Hdr.rc = SUPSemEventMultiClose(pSession, hEventMulti);
1674	break;
1675	case SUPSEMOP_RESET:
1676	pReq->Hdr.rc = SUPSemEventMultiReset(pSession, hEventMulti);
1677	break;
1678	default:
1679	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1680	break;
1681	}
1682	break;
1683	}
1684
1685	default:
1686	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1687	break;
1688	}
1689	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1690	return 0;
1691	}
1692
1693	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP2):
1694	{
1695	/* validate */
1696	PSUPSEMOP2 pReq = (PSUPSEMOP2)pReqHdr;
1697	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP2, SUP_IOCTL_SEM_OP2_SIZE_IN, SUP_IOCTL_SEM_OP2_SIZE_OUT);
1698	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP2, pReq->u.In.uReserved == 0);
1699
1700	/* execute */
1701	switch (pReq->u.In.uType)
1702	{
1703	case SUP_SEM_TYPE_EVENT:
1704	{
1705	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1706	switch (pReq->u.In.uOp)
1707	{
1708	case SUPSEMOP2_WAIT_MS_REL:
1709	pReq->Hdr.rc = SUPSemEventWaitNoResume(pSession, hEvent, pReq->u.In.uArg.cRelMsTimeout);
1710	break;
1711	case SUPSEMOP2_WAIT_NS_ABS:
1712	pReq->Hdr.rc = SUPSemEventWaitNsAbsIntr(pSession, hEvent, pReq->u.In.uArg.uAbsNsTimeout);
1713	break;
1714	case SUPSEMOP2_WAIT_NS_REL:
1715	pReq->Hdr.rc = SUPSemEventWaitNsRelIntr(pSession, hEvent, pReq->u.In.uArg.cRelNsTimeout);
1716	break;
1717	case SUPSEMOP2_SIGNAL:
1718	pReq->Hdr.rc = SUPSemEventSignal(pSession, hEvent);
1719	break;
1720	case SUPSEMOP2_CLOSE:
1721	pReq->Hdr.rc = SUPSemEventClose(pSession, hEvent);
1722	break;
1723	case SUPSEMOP2_RESET:
1724	default:
1725	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1726	break;
1727	}
1728	break;
1729	}
1730
1731	case SUP_SEM_TYPE_EVENT_MULTI:
1732	{
1733	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1734	switch (pReq->u.In.uOp)
1735	{
1736	case SUPSEMOP2_WAIT_MS_REL:
1737	pReq->Hdr.rc = SUPSemEventMultiWaitNoResume(pSession, hEventMulti, pReq->u.In.uArg.cRelMsTimeout);
1738	break;
1739	case SUPSEMOP2_WAIT_NS_ABS:
1740	pReq->Hdr.rc = SUPSemEventMultiWaitNsAbsIntr(pSession, hEventMulti, pReq->u.In.uArg.uAbsNsTimeout);
1741	break;
1742	case SUPSEMOP2_WAIT_NS_REL:
1743	pReq->Hdr.rc = SUPSemEventMultiWaitNsRelIntr(pSession, hEventMulti, pReq->u.In.uArg.cRelNsTimeout);
1744	break;
1745	case SUPSEMOP2_SIGNAL:
1746	pReq->Hdr.rc = SUPSemEventMultiSignal(pSession, hEventMulti);
1747	break;
1748	case SUPSEMOP2_CLOSE:
1749	pReq->Hdr.rc = SUPSemEventMultiClose(pSession, hEventMulti);
1750	break;
1751	case SUPSEMOP2_RESET:
1752	pReq->Hdr.rc = SUPSemEventMultiReset(pSession, hEventMulti);
1753	break;
1754	default:
1755	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1756	break;
1757	}
1758	break;
1759	}
1760
1761	default:
1762	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1763	break;
1764	}
1765	return 0;
1766	}
1767
1768	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP3):
1769	{
1770	/* validate */
1771	PSUPSEMOP3 pReq = (PSUPSEMOP3)pReqHdr;
1772	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP3, SUP_IOCTL_SEM_OP3_SIZE_IN, SUP_IOCTL_SEM_OP3_SIZE_OUT);
1773	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, pReq->u.In.u32Reserved == 0 && pReq->u.In.u64Reserved == 0);
1774
1775	/* execute */
1776	switch (pReq->u.In.uType)
1777	{
1778	case SUP_SEM_TYPE_EVENT:
1779	{
1780	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1781	switch (pReq->u.In.uOp)
1782	{
1783	case SUPSEMOP3_CREATE:
1784	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1785	pReq->Hdr.rc = SUPSemEventCreate(pSession, &hEvent);
1786	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEvent;
1787	break;
1788	case SUPSEMOP3_GET_RESOLUTION:
1789	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1790	pReq->Hdr.rc = VINF_SUCCESS;
1791	pReq->Hdr.cbOut = sizeof(*pReq);
1792	pReq->u.Out.cNsResolution = SUPSemEventGetResolution(pSession);
1793	break;
1794	default:
1795	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1796	break;
1797	}
1798	break;
1799	}
1800
1801	case SUP_SEM_TYPE_EVENT_MULTI:
1802	{
1803	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1804	switch (pReq->u.In.uOp)
1805	{
1806	case SUPSEMOP3_CREATE:
1807	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1808	pReq->Hdr.rc = SUPSemEventMultiCreate(pSession, &hEventMulti);
1809	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEventMulti;
1810	break;
1811	case SUPSEMOP3_GET_RESOLUTION:
1812	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1813	pReq->Hdr.rc = VINF_SUCCESS;
1814	pReq->u.Out.cNsResolution = SUPSemEventMultiGetResolution(pSession);
1815	break;
1816	default:
1817	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1818	break;
1819	}
1820	break;
1821	}
1822
1823	default:
1824	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1825	break;
1826	}
1827	return 0;
1828	}
1829
1830	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
1831	{
1832	/* validate */
1833	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
1834	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
1835	REQ_CHECK_EXPR(SUP_IOCTL_VT_CAPS, pReq->Hdr.cbIn <= SUP_IOCTL_VT_CAPS_SIZE_IN);
1836
1837	/* execute */
1838	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.Caps);
1839	if (RT_FAILURE(pReq->Hdr.rc))
1840	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1841	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1842	return 0;
1843	}
1844
1845	default:
1846	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1847	break;
1848	}
1849	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_GENERAL_FAILURE, VERR_GENERAL_FAILURE);
1850	return VERR_GENERAL_FAILURE;
1851	}
1852
1853
1854	/**
1855	* Inter-Driver Communication (IDC) worker.
1856	*
1857	* @returns VBox status code.
1858	* @retval VINF_SUCCESS on success.
1859	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1860	* @retval VERR_NOT_SUPPORTED if the request isn't supported.
1861	*
1862	* @param uReq The request (function) code.
1863	* @param pDevExt Device extention.
1864	* @param pSession Session data.
1865	* @param pReqHdr The request header.
1866	*/
1867	int VBOXCALL supdrvIDC(uintptr_t uReq, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQHDR pReqHdr)
1868	{
1869	/*
1870	* The OS specific code has already validated the pSession
1871	* pointer, and the request size being greater or equal to
1872	* size of the header.
1873	*
1874	* So, just check that pSession is a kernel context session.
1875	*/
1876	if (RT_UNLIKELY( pSession
1877	&& pSession->R0Process != NIL_RTR0PROCESS))
1878	return VERR_INVALID_PARAMETER;
1879
1880	/*
1881	* Validation macro.
1882	*/
1883	#define REQ_CHECK_IDC_SIZE(Name, cbExpect) \
1884	do { \
1885	if (RT_UNLIKELY(pReqHdr->cb != (cbExpect))) \
1886	{ \
1887	OSDBGPRINT(( #Name ": Invalid input/output sizes. cb=%ld expected %ld.\n", \
1888	(long)pReqHdr->cb, (long)(cbExpect))); \
1889	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1890	} \
1891	} while (0)
1892
1893	switch (uReq)
1894	{
1895	case SUPDRV_IDC_REQ_CONNECT:
1896	{
1897	PSUPDRVIDCREQCONNECT pReq = (PSUPDRVIDCREQCONNECT)pReqHdr;
1898	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_CONNECT, sizeof(*pReq));
1899
1900	/*
1901	* Validate the cookie and other input.
1902	*/
1903	if (pReq->Hdr.pSession != NULL)
1904	{
1905	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: pSession=%p expected NULL!\n", pReq->Hdr.pSession));
1906	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1907	}
1908	if (pReq->u.In.u32MagicCookie != SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE)
1909	{
1910	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: u32MagicCookie=%#x expected %#x!\n",
1911	(unsigned)pReq->u.In.u32MagicCookie, (unsigned)SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE));
1912	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1913	}
1914	if ( pReq->u.In.uMinVersion > pReq->u.In.uReqVersion
1915	\|\| (pReq->u.In.uMinVersion & UINT32_C(0xffff0000)) != (pReq->u.In.uReqVersion & UINT32_C(0xffff0000)))
1916	{
1917	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: uMinVersion=%#x uMaxVersion=%#x doesn't match!\n",
1918	pReq->u.In.uMinVersion, pReq->u.In.uReqVersion));
1919	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1920	}
1921
1922	/*
1923	* Match the version.
1924	* The current logic is very simple, match the major interface version.
1925	*/
1926	if ( pReq->u.In.uMinVersion > SUPDRV_IDC_VERSION
1927	\|\| (pReq->u.In.uMinVersion & 0xffff0000) != (SUPDRV_IDC_VERSION & 0xffff0000))
1928	{
1929	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1930	pReq->u.In.uReqVersion, pReq->u.In.uMinVersion, (unsigned)SUPDRV_IDC_VERSION));
1931	pReq->u.Out.pSession = NULL;
1932	pReq->u.Out.uSessionVersion = 0xffffffff;
1933	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1934	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1935	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1936	return VINF_SUCCESS;
1937	}
1938
1939	pReq->u.Out.pSession = NULL;
1940	pReq->u.Out.uSessionVersion = SUPDRV_IDC_VERSION;
1941	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1942	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1943
1944	/*
1945	* On NT we will already have a session associated with the
1946	* client, just like with the SUP_IOCTL_COOKIE request, while
1947	* the other doesn't.
1948	*/
1949	#ifdef RT_OS_WINDOWS
1950	pReq->Hdr.rc = VINF_SUCCESS;
1951	#else
1952	AssertReturn(!pSession, VERR_INTERNAL_ERROR);
1953	pReq->Hdr.rc = supdrvCreateSession(pDevExt, false /* fUser */, &pSession);
1954	if (RT_FAILURE(pReq->Hdr.rc))
1955	{
1956	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: failed to create session, rc=%d\n", pReq->Hdr.rc));
1957	return VINF_SUCCESS;
1958	}
1959	#endif
1960
1961	pReq->u.Out.pSession = pSession;
1962	pReq->Hdr.pSession = pSession;
1963
1964	return VINF_SUCCESS;
1965	}
1966
1967	case SUPDRV_IDC_REQ_DISCONNECT:
1968	{
1969	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_DISCONNECT, sizeof(*pReqHdr));
1970
1971	#ifdef RT_OS_WINDOWS
1972	/* Windows will destroy the session when the file object is destroyed. */
1973	#else
1974	supdrvCloseSession(pDevExt, pSession);
1975	#endif
1976	return pReqHdr->rc = VINF_SUCCESS;
1977	}
1978
1979	case SUPDRV_IDC_REQ_GET_SYMBOL:
1980	{
1981	PSUPDRVIDCREQGETSYM pReq = (PSUPDRVIDCREQGETSYM)pReqHdr;
1982	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_GET_SYMBOL, sizeof(*pReq));
1983
1984	pReq->Hdr.rc = supdrvIDC_LdrGetSymbol(pDevExt, pSession, pReq);
1985	return VINF_SUCCESS;
1986	}
1987
1988	case SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY:
1989	{
1990	PSUPDRVIDCREQCOMPREGFACTORY pReq = (PSUPDRVIDCREQCOMPREGFACTORY)pReqHdr;
1991	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY, sizeof(*pReq));
1992
1993	pReq->Hdr.rc = SUPR0ComponentRegisterFactory(pSession, pReq->u.In.pFactory);
1994	return VINF_SUCCESS;
1995	}
1996
1997	case SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY:
1998	{
1999	PSUPDRVIDCREQCOMPDEREGFACTORY pReq = (PSUPDRVIDCREQCOMPDEREGFACTORY)pReqHdr;
2000	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY, sizeof(*pReq));
2001
2002	pReq->Hdr.rc = SUPR0ComponentDeregisterFactory(pSession, pReq->u.In.pFactory);
2003	return VINF_SUCCESS;
2004	}
2005
2006	default:
2007	Log(("Unknown IDC %#lx\n", (long)uReq));
2008	break;
2009	}
2010
2011	#undef REQ_CHECK_IDC_SIZE
2012	return VERR_NOT_SUPPORTED;
2013	}
2014
2015
2016	/**
2017	* Register a object for reference counting.
2018	* The object is registered with one reference in the specified session.
2019	*
2020	* @returns Unique identifier on success (pointer).
2021	* All future reference must use this identifier.
2022	* @returns NULL on failure.
2023	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
2024	* @param pvUser1 The first user argument.
2025	* @param pvUser2 The second user argument.
2026	*/
2027	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
2028	{
2029	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2030	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2031	PSUPDRVOBJ pObj;
2032	PSUPDRVUSAGE pUsage;
2033
2034	/*
2035	* Validate the input.
2036	*/
2037	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
2038	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
2039	AssertPtrReturn(pfnDestructor, NULL);
2040
2041	/*
2042	* Allocate and initialize the object.
2043	*/
2044	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
2045	if (!pObj)
2046	return NULL;
2047	pObj->u32Magic = SUPDRVOBJ_MAGIC;
2048	pObj->enmType = enmType;
2049	pObj->pNext = NULL;
2050	pObj->cUsage = 1;
2051	pObj->pfnDestructor = pfnDestructor;
2052	pObj->pvUser1 = pvUser1;
2053	pObj->pvUser2 = pvUser2;
2054	pObj->CreatorUid = pSession->Uid;
2055	pObj->CreatorGid = pSession->Gid;
2056	pObj->CreatorProcess= pSession->Process;
2057	supdrvOSObjInitCreator(pObj, pSession);
2058
2059	/*
2060	* Allocate the usage record.
2061	* (We keep freed usage records around to simplify SUPR0ObjAddRefEx().)
2062	*/
2063	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2064
2065	pUsage = pDevExt->pUsageFree;
2066	if (pUsage)
2067	pDevExt->pUsageFree = pUsage->pNext;
2068	else
2069	{
2070	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2071	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
2072	if (!pUsage)
2073	{
2074	RTMemFree(pObj);
2075	return NULL;
2076	}
2077	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2078	}
2079
2080	/*
2081	* Insert the object and create the session usage record.
2082	*/
2083	/* The object. */
2084	pObj->pNext = pDevExt->pObjs;
2085	pDevExt->pObjs = pObj;
2086
2087	/* The session record. */
2088	pUsage->cUsage = 1;
2089	pUsage->pObj = pObj;
2090	pUsage->pNext = pSession->pUsage;
2091	/* Log2(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext)); */
2092	pSession->pUsage = pUsage;
2093
2094	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2095
2096	Log(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
2097	return pObj;
2098	}
2099
2100
2101	/**
2102	* Increment the reference counter for the object associating the reference
2103	* with the specified session.
2104	*
2105	* @returns IPRT status code.
2106	* @param pvObj The identifier returned by SUPR0ObjRegister().
2107	* @param pSession The session which is referencing the object.
2108	*
2109	* @remarks The caller should not own any spinlocks and must carefully protect
2110	* itself against potential race with the destructor so freed memory
2111	* isn't accessed here.
2112	*/
2113	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
2114	{
2115	return SUPR0ObjAddRefEx(pvObj, pSession, false /* fNoBlocking */);
2116	}
2117
2118
2119	/**
2120	* Increment the reference counter for the object associating the reference
2121	* with the specified session.
2122	*
2123	* @returns IPRT status code.
2124	* @retval VERR_TRY_AGAIN if fNoBlocking was set and a new usage record
2125	* couldn't be allocated. (If you see this you're not doing the right
2126	* thing and it won't ever work reliably.)
2127	*
2128	* @param pvObj The identifier returned by SUPR0ObjRegister().
2129	* @param pSession The session which is referencing the object.
2130	* @param fNoBlocking Set if it's not OK to block. Never try to make the
2131	* first reference to an object in a session with this
2132	* argument set.
2133	*
2134	* @remarks The caller should not own any spinlocks and must carefully protect
2135	* itself against potential race with the destructor so freed memory
2136	* isn't accessed here.
2137	*/
2138	SUPR0DECL(int) SUPR0ObjAddRefEx(void *pvObj, PSUPDRVSESSION pSession, bool fNoBlocking)
2139	{
2140	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2141	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2142	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2143	int rc = VINF_SUCCESS;
2144	PSUPDRVUSAGE pUsagePre;
2145	PSUPDRVUSAGE pUsage;
2146
2147	/*
2148	* Validate the input.
2149	* Be ready for the destruction race (someone might be stuck in the
2150	* destructor waiting a lock we own).
2151	*/
2152	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2153	AssertPtrReturn(pObj, VERR_INVALID_POINTER);
2154	AssertMsgReturn(pObj->u32Magic == SUPDRVOBJ_MAGIC \|\| pObj->u32Magic == SUPDRVOBJ_MAGIC_DEAD,
2155	("Invalid pvObj=%p magic=%#x (expected %#x or %#x)\n", pvObj, pObj->u32Magic, SUPDRVOBJ_MAGIC, SUPDRVOBJ_MAGIC_DEAD),
2156	VERR_INVALID_PARAMETER);
2157
2158	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2159
2160	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2161	{
2162	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2163
2164	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2165	return VERR_WRONG_ORDER;
2166	}
2167
2168	/*
2169	* Preallocate the usage record if we can.
2170	*/
2171	pUsagePre = pDevExt->pUsageFree;
2172	if (pUsagePre)
2173	pDevExt->pUsageFree = pUsagePre->pNext;
2174	else if (!fNoBlocking)
2175	{
2176	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2177	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
2178	if (!pUsagePre)
2179	return VERR_NO_MEMORY;
2180
2181	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2182	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2183	{
2184	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2185
2186	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2187	return VERR_WRONG_ORDER;
2188	}
2189	}
2190
2191	/*
2192	* Reference the object.
2193	*/
2194	pObj->cUsage++;
2195
2196	/*
2197	* Look for the session record.
2198	*/
2199	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
2200	{
2201	/Log(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2202	if (pUsage->pObj == pObj)
2203	break;
2204	}
2205	if (pUsage)
2206	pUsage->cUsage++;
2207	else if (pUsagePre)
2208	{
2209	/* create a new session record. */
2210	pUsagePre->cUsage = 1;
2211	pUsagePre->pObj = pObj;
2212	pUsagePre->pNext = pSession->pUsage;
2213	pSession->pUsage = pUsagePre;
2214	/Log(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));/
2215
2216	pUsagePre = NULL;
2217	}
2218	else
2219	{
2220	pObj->cUsage--;
2221	rc = VERR_TRY_AGAIN;
2222	}
2223
2224	/*
2225	* Put any unused usage record into the free list..
2226	*/
2227	if (pUsagePre)
2228	{
2229	pUsagePre->pNext = pDevExt->pUsageFree;
2230	pDevExt->pUsageFree = pUsagePre;
2231	}
2232
2233	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2234
2235	return rc;
2236	}
2237
2238
2239	/**
2240	* Decrement / destroy a reference counter record for an object.
2241	*
2242	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
2243	*
2244	* @returns IPRT status code.
2245	* @retval VINF_SUCCESS if not destroyed.
2246	* @retval VINF_OBJECT_DESTROYED if it's destroyed by this release call.
2247	* @retval VERR_INVALID_PARAMETER if the object isn't valid. Will assert in
2248	* string builds.
2249	*
2250	* @param pvObj The identifier returned by SUPR0ObjRegister().
2251	* @param pSession The session which is referencing the object.
2252	*/
2253	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
2254	{
2255	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2256	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2257	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2258	int rc = VERR_INVALID_PARAMETER;
2259	PSUPDRVUSAGE pUsage;
2260	PSUPDRVUSAGE pUsagePrev;
2261
2262	/*
2263	* Validate the input.
2264	*/
2265	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2266	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2267	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2268	VERR_INVALID_PARAMETER);
2269
2270	/*
2271	* Acquire the spinlock and look for the usage record.
2272	*/
2273	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2274
2275	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
2276	pUsage;
2277	pUsagePrev = pUsage, pUsage = pUsage->pNext)
2278	{
2279	/Log2(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2280	if (pUsage->pObj == pObj)
2281	{
2282	rc = VINF_SUCCESS;
2283	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
2284	if (pUsage->cUsage > 1)
2285	{
2286	pObj->cUsage--;
2287	pUsage->cUsage--;
2288	}
2289	else
2290	{
2291	/*
2292	* Free the session record.
2293	*/
2294	if (pUsagePrev)
2295	pUsagePrev->pNext = pUsage->pNext;
2296	else
2297	pSession->pUsage = pUsage->pNext;
2298	pUsage->pNext = pDevExt->pUsageFree;
2299	pDevExt->pUsageFree = pUsage;
2300
2301	/* What about the object? */
2302	if (pObj->cUsage > 1)
2303	pObj->cUsage--;
2304	else
2305	{
2306	/*
2307	* Object is to be destroyed, unlink it.
2308	*/
2309	pObj->u32Magic = SUPDRVOBJ_MAGIC_DEAD;
2310	rc = VINF_OBJECT_DESTROYED;
2311	if (pDevExt->pObjs == pObj)
2312	pDevExt->pObjs = pObj->pNext;
2313	else
2314	{
2315	PSUPDRVOBJ pObjPrev;
2316	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
2317	if (pObjPrev->pNext == pObj)
2318	{
2319	pObjPrev->pNext = pObj->pNext;
2320	break;
2321	}
2322	Assert(pObjPrev);
2323	}
2324	}
2325	}
2326	break;
2327	}
2328	}
2329
2330	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2331
2332	/*
2333	* Call the destructor and free the object if required.
2334	*/
2335	if (rc == VINF_OBJECT_DESTROYED)
2336	{
2337	Log(("SUPR0ObjRelease: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
2338	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
2339	if (pObj->pfnDestructor)
2340	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
2341	RTMemFree(pObj);
2342	}
2343
2344	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
2345	return rc;
2346	}
2347
2348
2349	/**
2350	* Verifies that the current process can access the specified object.
2351	*
2352	* @returns The following IPRT status code:
2353	* @retval VINF_SUCCESS if access was granted.
2354	* @retval VERR_PERMISSION_DENIED if denied access.
2355	* @retval VERR_INVALID_PARAMETER if invalid parameter.
2356	*
2357	* @param pvObj The identifier returned by SUPR0ObjRegister().
2358	* @param pSession The session which wishes to access the object.
2359	* @param pszObjName Object string name. This is optional and depends on the object type.
2360	*
2361	* @remark The caller is responsible for making sure the object isn't removed while
2362	* we're inside this function. If uncertain about this, just call AddRef before calling us.
2363	*/
2364	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
2365	{
2366	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2367	int rc;
2368
2369	/*
2370	* Validate the input.
2371	*/
2372	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2373	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2374	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2375	VERR_INVALID_PARAMETER);
2376
2377	/*
2378	* Check access. (returns true if a decision has been made.)
2379	*/
2380	rc = VERR_INTERNAL_ERROR;
2381	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
2382	return rc;
2383
2384	/*
2385	* Default policy is to allow the user to access his own
2386	* stuff but nothing else.
2387	*/
2388	if (pObj->CreatorUid == pSession->Uid)
2389	return VINF_SUCCESS;
2390	return VERR_PERMISSION_DENIED;
2391	}
2392
2393
2394	/**
2395	* Lock pages.
2396	*
2397	* @returns IPRT status code.
2398	* @param pSession Session to which the locked memory should be associated.
2399	* @param pvR3 Start of the memory range to lock.
2400	* This must be page aligned.
2401	* @param cPages Number of pages to lock.
2402	* @param paPages Where to put the physical addresses of locked memory.
2403	*/
2404	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
2405	{
2406	int rc;
2407	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2408	const size_t cb = (size_t)cPages << PAGE_SHIFT;
2409	LogFlow(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
2410
2411	/*
2412	* Verify input.
2413	*/
2414	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2415	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
2416	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
2417	\|\| !pvR3)
2418	{
2419	Log(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
2420	return VERR_INVALID_PARAMETER;
2421	}
2422
2423	/*
2424	* Let IPRT do the job.
2425	*/
2426	Mem.eType = MEMREF_TYPE_LOCKED;
2427	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTMEM_PROT_READ \| RTMEM_PROT_WRITE, RTR0ProcHandleSelf());
2428	if (RT_SUCCESS(rc))
2429	{
2430	uint32_t iPage = cPages;
2431	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
2432	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
2433
2434	while (iPage-- > 0)
2435	{
2436	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2437	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
2438	{
2439	AssertMsgFailed(("iPage=%d\n", iPage));
2440	rc = VERR_INTERNAL_ERROR;
2441	break;
2442	}
2443	}
2444	if (RT_SUCCESS(rc))
2445	rc = supdrvMemAdd(&Mem, pSession);
2446	if (RT_FAILURE(rc))
2447	{
2448	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
2449	AssertRC(rc2);
2450	}
2451	}
2452
2453	return rc;
2454	}
2455
2456
2457	/**
2458	* Unlocks the memory pointed to by pv.
2459	*
2460	* @returns IPRT status code.
2461	* @param pSession Session to which the memory was locked.
2462	* @param pvR3 Memory to unlock.
2463	*/
2464	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2465	{
2466	LogFlow(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2467	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2468	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
2469	}
2470
2471
2472	/**
2473	* Allocates a chunk of page aligned memory with contiguous and fixed physical
2474	* backing.
2475	*
2476	* @returns IPRT status code.
2477	* @param pSession Session data.
2478	* @param cPages Number of pages to allocate.
2479	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
2480	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
2481	* @param pHCPhys Where to put the physical address of allocated memory.
2482	*/
2483	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
2484	{
2485	int rc;
2486	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2487	LogFlow(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
2488
2489	/*
2490	* Validate input.
2491	*/
2492	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2493	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
2494	{
2495	Log(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
2496	pSession, ppvR0, ppvR3, pHCPhys));
2497	return VERR_INVALID_PARAMETER;
2498
2499	}
2500	if (cPages < 1 \|\| cPages >= 256)
2501	{
2502	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2503	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2504	}
2505
2506	/*
2507	* Let IPRT do the job.
2508	*/
2509	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
2510	if (RT_SUCCESS(rc))
2511	{
2512	int rc2;
2513	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2514	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2515	if (RT_SUCCESS(rc))
2516	{
2517	Mem.eType = MEMREF_TYPE_CONT;
2518	rc = supdrvMemAdd(&Mem, pSession);
2519	if (!rc)
2520	{
2521	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2522	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2523	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
2524	return 0;
2525	}
2526
2527	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2528	AssertRC(rc2);
2529	}
2530	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2531	AssertRC(rc2);
2532	}
2533
2534	return rc;
2535	}
2536
2537
2538	/**
2539	* Frees memory allocated using SUPR0ContAlloc().
2540	*
2541	* @returns IPRT status code.
2542	* @param pSession The session to which the memory was allocated.
2543	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2544	*/
2545	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2546	{
2547	LogFlow(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2548	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2549	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
2550	}
2551
2552
2553	/**
2554	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
2555	*
2556	* The memory isn't zeroed.
2557	*
2558	* @returns IPRT status code.
2559	* @param pSession Session data.
2560	* @param cPages Number of pages to allocate.
2561	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
2562	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
2563	* @param paPages Where to put the physical addresses of allocated memory.
2564	*/
2565	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
2566	{
2567	unsigned iPage;
2568	int rc;
2569	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2570	LogFlow(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
2571
2572	/*
2573	* Validate input.
2574	*/
2575	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2576	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
2577	{
2578	Log(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
2579	pSession, ppvR3, ppvR0, paPages));
2580	return VERR_INVALID_PARAMETER;
2581
2582	}
2583	if (cPages < 1 \|\| cPages >= 256)
2584	{
2585	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2586	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2587	}
2588
2589	/*
2590	* Let IPRT do the work.
2591	*/
2592	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
2593	if (RT_SUCCESS(rc))
2594	{
2595	int rc2;
2596	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2597	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2598	if (RT_SUCCESS(rc))
2599	{
2600	Mem.eType = MEMREF_TYPE_LOW;
2601	rc = supdrvMemAdd(&Mem, pSession);
2602	if (!rc)
2603	{
2604	for (iPage = 0; iPage < cPages; iPage++)
2605	{
2606	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2607	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%RHp\n", paPages[iPage]));
2608	}
2609	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2610	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2611	return 0;
2612	}
2613
2614	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2615	AssertRC(rc2);
2616	}
2617
2618	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2619	AssertRC(rc2);
2620	}
2621
2622	return rc;
2623	}
2624
2625
2626	/**
2627	* Frees memory allocated using SUPR0LowAlloc().
2628	*
2629	* @returns IPRT status code.
2630	* @param pSession The session to which the memory was allocated.
2631	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2632	*/
2633	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2634	{
2635	LogFlow(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2636	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2637	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
2638	}
2639
2640
2641
2642	/**
2643	* Allocates a chunk of memory with both R0 and R3 mappings.
2644	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
2645	*
2646	* @returns IPRT status code.
2647	* @param pSession The session to associated the allocation with.
2648	* @param cb Number of bytes to allocate.
2649	* @param ppvR0 Where to store the address of the Ring-0 mapping.
2650	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2651	*/
2652	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
2653	{
2654	int rc;
2655	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2656	LogFlow(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
2657
2658	/*
2659	* Validate input.
2660	*/
2661	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2662	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
2663	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
2664	if (cb < 1 \|\| cb >= _4M)
2665	{
2666	Log(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
2667	return VERR_INVALID_PARAMETER;
2668	}
2669
2670	/*
2671	* Let IPRT do the work.
2672	*/
2673	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
2674	if (RT_SUCCESS(rc))
2675	{
2676	int rc2;
2677	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2678	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2679	if (RT_SUCCESS(rc))
2680	{
2681	Mem.eType = MEMREF_TYPE_MEM;
2682	rc = supdrvMemAdd(&Mem, pSession);
2683	if (!rc)
2684	{
2685	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2686	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2687	return VINF_SUCCESS;
2688	}
2689
2690	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2691	AssertRC(rc2);
2692	}
2693
2694	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2695	AssertRC(rc2);
2696	}
2697
2698	return rc;
2699	}
2700
2701
2702	/**
2703	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
2704	*
2705	* @returns IPRT status code.
2706	* @param pSession The session to which the memory was allocated.
2707	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2708	* @param paPages Where to store the physical addresses.
2709	*/
2710	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
2711	{
2712	PSUPDRVBUNDLE pBundle;
2713	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2714	LogFlow(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
2715
2716	/*
2717	* Validate input.
2718	*/
2719	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2720	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
2721	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
2722
2723	/*
2724	* Search for the address.
2725	*/
2726	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2727	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2728	{
2729	if (pBundle->cUsed > 0)
2730	{
2731	unsigned i;
2732	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2733	{
2734	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
2735	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2736	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2737	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2738	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
2739	)
2740	)
2741	{
2742	const size_t cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2743	size_t iPage;
2744	for (iPage = 0; iPage < cPages; iPage++)
2745	{
2746	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2747	paPages[iPage].uReserved = 0;
2748	}
2749	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2750	return VINF_SUCCESS;
2751	}
2752	}
2753	}
2754	}
2755	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2756	Log(("Failed to find %p!!!\n", (void *)uPtr));
2757	return VERR_INVALID_PARAMETER;
2758	}
2759
2760
2761	/**
2762	* Free memory allocated by SUPR0MemAlloc().
2763	*
2764	* @returns IPRT status code.
2765	* @param pSession The session owning the allocation.
2766	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2767	*/
2768	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2769	{
2770	LogFlow(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2771	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2772	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
2773	}
2774
2775
2776	/**
2777	* Allocates a chunk of memory with a kernel or/and a user mode mapping.
2778	*
2779	* The memory is fixed and it's possible to query the physical addresses using
2780	* SUPR0MemGetPhys().
2781	*
2782	* @returns IPRT status code.
2783	* @param pSession The session to associated the allocation with.
2784	* @param cPages The number of pages to allocate.
2785	* @param fFlags Flags, reserved for the future. Must be zero.
2786	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2787	* NULL if no ring-3 mapping.
2788	* @param ppvR3 Where to store the address of the Ring-0 mapping.
2789	* NULL if no ring-0 mapping.
2790	* @param paPages Where to store the addresses of the pages. Optional.
2791	*/
2792	SUPR0DECL(int) SUPR0PageAllocEx(PSUPDRVSESSION pSession, uint32_t cPages, uint32_t fFlags, PRTR3PTR ppvR3, PRTR0PTR ppvR0, PRTHCPHYS paPages)
2793	{
2794	int rc;
2795	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2796	LogFlow(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
2797
2798	/*
2799	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2800	*/
2801	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2802	AssertPtrNullReturn(ppvR3, VERR_INVALID_POINTER);
2803	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2804	AssertReturn(ppvR3 \|\| ppvR0, VERR_INVALID_PARAMETER);
2805	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2806	if (cPages < 1 \|\| cPages > VBOX_MAX_ALLOC_PAGE_COUNT)
2807	{
2808	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)));
2809	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2810	}
2811
2812	/*
2813	* Let IPRT do the work.
2814	*/
2815	if (ppvR0)
2816	rc = RTR0MemObjAllocPage(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, true /* fExecutable */);
2817	else
2818	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
2819	if (RT_SUCCESS(rc))
2820	{
2821	int rc2;
2822	if (ppvR3)
2823	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2824	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2825	else
2826	Mem.MapObjR3 = NIL_RTR0MEMOBJ;
2827	if (RT_SUCCESS(rc))
2828	{
2829	Mem.eType = MEMREF_TYPE_PAGE;
2830	rc = supdrvMemAdd(&Mem, pSession);
2831	if (!rc)
2832	{
2833	if (ppvR3)
2834	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2835	if (ppvR0)
2836	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2837	if (paPages)
2838	{
2839	uint32_t iPage = cPages;
2840	while (iPage-- > 0)
2841	{
2842	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
2843	Assert(paPages[iPage] != NIL_RTHCPHYS);
2844	}
2845	}
2846	return VINF_SUCCESS;
2847	}
2848
2849	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2850	AssertRC(rc2);
2851	}
2852
2853	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2854	AssertRC(rc2);
2855	}
2856	return rc;
2857	}
2858
2859
2860	/**
2861	* Maps a chunk of memory previously allocated by SUPR0PageAllocEx into kernel
2862	* space.
2863	*
2864	* @returns IPRT status code.
2865	* @param pSession The session to associated the allocation with.
2866	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
2867	* @param offSub Where to start mapping. Must be page aligned.
2868	* @param cbSub How much to map. Must be page aligned.
2869	* @param fFlags Flags, MBZ.
2870	* @param ppvR0 Where to return the address of the ring-0 mapping on
2871	* success.
2872	*/
2873	SUPR0DECL(int) SUPR0PageMapKernel(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t offSub, uint32_t cbSub,
2874	uint32_t fFlags, PRTR0PTR ppvR0)
2875	{
2876	int rc;
2877	PSUPDRVBUNDLE pBundle;
2878	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2879	RTR0MEMOBJ hMemObj = NIL_RTR0MEMOBJ;
2880	LogFlow(("SUPR0PageMapKernel: pSession=%p pvR3=%p offSub=%#x cbSub=%#x\n", pSession, pvR3, offSub, cbSub));
2881
2882	/*
2883	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2884	*/
2885	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2886	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2887	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2888	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2889	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2890	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
2891
2892	/*
2893	* Find the memory object.
2894	*/
2895	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2896	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2897	{
2898	if (pBundle->cUsed > 0)
2899	{
2900	unsigned i;
2901	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2902	{
2903	if ( ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
2904	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2905	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2906	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
2907	\|\| ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED
2908	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2909	&& pBundle->aMem[i].MapObjR3 == NIL_RTR0MEMOBJ
2910	&& RTR0MemObjAddressR3(pBundle->aMem[i].MemObj) == pvR3))
2911	{
2912	hMemObj = pBundle->aMem[i].MemObj;
2913	break;
2914	}
2915	}
2916	}
2917	}
2918	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2919
2920	rc = VERR_INVALID_PARAMETER;
2921	if (hMemObj != NIL_RTR0MEMOBJ)
2922	{
2923	/*
2924	* Do some further input validations before calling IPRT.
2925	* (Cleanup is done indirectly by telling RTR0MemObjFree to include mappings.)
2926	*/
2927	size_t cbMemObj = RTR0MemObjSize(hMemObj);
2928	if ( offSub < cbMemObj
2929	&& cbSub <= cbMemObj
2930	&& offSub + cbSub <= cbMemObj)
2931	{
2932	RTR0MEMOBJ hMapObj;
2933	rc = RTR0MemObjMapKernelEx(&hMapObj, hMemObj, (void *)-1, 0,
2934	RTMEM_PROT_READ \| RTMEM_PROT_WRITE, offSub, cbSub);
2935	if (RT_SUCCESS(rc))
2936	*ppvR0 = RTR0MemObjAddress(hMapObj);
2937	}
2938	else
2939	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
2940
2941	}
2942	return rc;
2943	}
2944
2945
2946	/**
2947	* Changes the page level protection of one or more pages previously allocated
2948	* by SUPR0PageAllocEx.
2949	*
2950	* @returns IPRT status code.
2951	* @param pSession The session to associated the allocation with.
2952	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
2953	* NIL_RTR3PTR if the ring-3 mapping should be unaffected.
2954	* @param pvR0 The ring-0 address returned by SUPR0PageAllocEx.
2955	* NIL_RTR0PTR if the ring-0 mapping should be unaffected.
2956	* @param offSub Where to start changing. Must be page aligned.
2957	* @param cbSub How much to change. Must be page aligned.
2958	* @param fProt The new page level protection, see RTMEM_PROT_*.
2959	*/
2960	SUPR0DECL(int) SUPR0PageProtect(PSUPDRVSESSION pSession, RTR3PTR pvR3, RTR0PTR pvR0, uint32_t offSub, uint32_t cbSub, uint32_t fProt)
2961	{
2962	int rc;
2963	PSUPDRVBUNDLE pBundle;
2964	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2965	RTR0MEMOBJ hMemObjR0 = NIL_RTR0MEMOBJ;
2966	RTR0MEMOBJ hMemObjR3 = NIL_RTR0MEMOBJ;
2967	LogFlow(("SUPR0PageProtect: pSession=%p pvR3=%p pvR0=%p offSub=%#x cbSub=%#x fProt-%#x\n", pSession, pvR3, pvR0, offSub, cbSub, fProt));
2968
2969	/*
2970	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2971	*/
2972	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2973	AssertReturn(!(fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)), VERR_INVALID_PARAMETER);
2974	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2975	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2976	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
2977
2978	/*
2979	* Find the memory object.
2980	*/
2981	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2982	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2983	{
2984	if (pBundle->cUsed > 0)
2985	{
2986	unsigned i;
2987	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2988	{
2989	if ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
2990	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2991	&& ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2992	\|\| pvR3 == NIL_RTR3PTR)
2993	&& ( pvR0 == NIL_RTR0PTR
2994	\|\| RTR0MemObjAddress(pBundle->aMem[i].MemObj) == pvR0)
2995	&& ( pvR3 == NIL_RTR3PTR
2996	\|\| RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3))
2997	{
2998	if (pvR0 != NIL_RTR0PTR)
2999	hMemObjR0 = pBundle->aMem[i].MemObj;
3000	if (pvR3 != NIL_RTR3PTR)
3001	hMemObjR3 = pBundle->aMem[i].MapObjR3;
3002	break;
3003	}
3004	}
3005	}
3006	}
3007	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3008
3009	rc = VERR_INVALID_PARAMETER;
3010	if ( hMemObjR0 != NIL_RTR0MEMOBJ
3011	\|\| hMemObjR3 != NIL_RTR0MEMOBJ)
3012	{
3013	/*
3014	* Do some further input validations before calling IPRT.
3015	*/
3016	size_t cbMemObj = hMemObjR0 != NIL_RTR0PTR ? RTR0MemObjSize(hMemObjR0) : RTR0MemObjSize(hMemObjR3);
3017	if ( offSub < cbMemObj
3018	&& cbSub <= cbMemObj
3019	&& offSub + cbSub <= cbMemObj)
3020	{
3021	rc = VINF_SUCCESS;
3022	if (hMemObjR3 != NIL_RTR0PTR)
3023	rc = RTR0MemObjProtect(hMemObjR3, offSub, cbSub, fProt);
3024	if (hMemObjR0 != NIL_RTR0PTR && RT_SUCCESS(rc))
3025	rc = RTR0MemObjProtect(hMemObjR0, offSub, cbSub, fProt);
3026	}
3027	else
3028	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
3029
3030	}
3031	return rc;
3032
3033	}
3034
3035
3036	/**
3037	* Free memory allocated by SUPR0PageAlloc() and SUPR0PageAllocEx().
3038	*
3039	* @returns IPRT status code.
3040	* @param pSession The session owning the allocation.
3041	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc() or
3042	* SUPR0PageAllocEx().
3043	*/
3044	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
3045	{
3046	LogFlow(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
3047	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3048	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_PAGE);
3049	}
3050
3051
3052	/**
3053	* Gets the paging mode of the current CPU.
3054	*
3055	* @returns Paging mode, SUPPAGEINGMODE_INVALID on error.
3056	*/
3057	SUPR0DECL(SUPPAGINGMODE) SUPR0GetPagingMode(void)
3058	{
3059	SUPPAGINGMODE enmMode;
3060
3061	RTR0UINTREG cr0 = ASMGetCR0();
3062	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
3063	enmMode = SUPPAGINGMODE_INVALID;
3064	else
3065	{
3066	RTR0UINTREG cr4 = ASMGetCR4();
3067	uint32_t fNXEPlusLMA = 0;
3068	if (cr4 & X86_CR4_PAE)
3069	{
3070	uint32_t fAmdFeatures = ASMCpuId_EDX(0x80000001);
3071	if (fAmdFeatures & (X86_CPUID_AMD_FEATURE_EDX_NX \| X86_CPUID_AMD_FEATURE_EDX_LONG_MODE))
3072	{
3073	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
3074	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
3075	fNXEPlusLMA \|= RT_BIT(0);
3076	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
3077	fNXEPlusLMA \|= RT_BIT(1);
3078	}
3079	}
3080
3081	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
3082	{
3083	case 0:
3084	enmMode = SUPPAGINGMODE_32_BIT;
3085	break;
3086
3087	case X86_CR4_PGE:
3088	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
3089	break;
3090
3091	case X86_CR4_PAE:
3092	enmMode = SUPPAGINGMODE_PAE;
3093	break;
3094
3095	case X86_CR4_PAE \| RT_BIT(0):
3096	enmMode = SUPPAGINGMODE_PAE_NX;
3097	break;
3098
3099	case X86_CR4_PAE \| X86_CR4_PGE:
3100	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3101	break;
3102
3103	case X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3104	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3105	break;
3106
3107	case RT_BIT(1) \| X86_CR4_PAE:
3108	enmMode = SUPPAGINGMODE_AMD64;
3109	break;
3110
3111	case RT_BIT(1) \| X86_CR4_PAE \| RT_BIT(0):
3112	enmMode = SUPPAGINGMODE_AMD64_NX;
3113	break;
3114
3115	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
3116	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
3117	break;
3118
3119	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3120	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
3121	break;
3122
3123	default:
3124	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
3125	enmMode = SUPPAGINGMODE_INVALID;
3126	break;
3127	}
3128	}
3129	return enmMode;
3130	}
3131
3132
3133	/**
3134	* Enables or disabled hardware virtualization extensions using native OS APIs.
3135	*
3136	* @returns VBox status code.
3137	* @retval VINF_SUCCESS on success.
3138	* @retval VERR_NOT_SUPPORTED if not supported by the native OS.
3139	*
3140	* @param fEnable Whether to enable or disable.
3141	*/
3142	SUPR0DECL(int) SUPR0EnableVTx(bool fEnable)
3143	{
3144	#ifdef RT_OS_DARWIN
3145	return supdrvOSEnableVTx(fEnable);
3146	#else
3147	return VERR_NOT_SUPPORTED;
3148	#endif
3149	}
3150
3151
3152	/** @todo document me */
3153	SUPR0DECL(int) SUPR0QueryVTCaps(PSUPDRVSESSION pSession, uint32_t *pfCaps)
3154	{
3155	/*
3156	* Input validation.
3157	*/
3158	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3159	AssertPtrReturn(pfCaps, VERR_INVALID_POINTER);
3160
3161	*pfCaps = 0;
3162
3163	if (ASMHasCpuId())
3164	{
3165	uint32_t u32FeaturesECX;
3166	uint32_t u32Dummy;
3167	uint32_t u32FeaturesEDX;
3168	uint32_t u32VendorEBX, u32VendorECX, u32VendorEDX, u32AMDFeatureEDX, u32AMDFeatureECX;
3169	uint64_t val;
3170
3171	ASMCpuId(0, &u32Dummy, &u32VendorEBX, &u32VendorECX, &u32VendorEDX);
3172	ASMCpuId(1, &u32Dummy, &u32Dummy, &u32FeaturesECX, &u32FeaturesEDX);
3173	/* Query AMD features. */
3174	ASMCpuId(0x80000001, &u32Dummy, &u32Dummy, &u32AMDFeatureECX, &u32AMDFeatureEDX);
3175
3176	if ( u32VendorEBX == X86_CPUID_VENDOR_INTEL_EBX
3177	&& u32VendorECX == X86_CPUID_VENDOR_INTEL_ECX
3178	&& u32VendorEDX == X86_CPUID_VENDOR_INTEL_EDX
3179	)
3180	{
3181	if ( (u32FeaturesECX & X86_CPUID_FEATURE_ECX_VMX)
3182	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3183	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3184	)
3185	{
3186	val = ASMRdMsr(MSR_IA32_FEATURE_CONTROL);
3187	/*
3188	* Both the LOCK and VMXON bit must be set; otherwise VMXON will generate a #GP.
3189	* Once the lock bit is set, this MSR can no longer be modified.
3190	*/
3191	if ( (val & (MSR_IA32_FEATURE_CONTROL_VMXON\|MSR_IA32_FEATURE_CONTROL_LOCK))
3192	== (MSR_IA32_FEATURE_CONTROL_VMXON\|MSR_IA32_FEATURE_CONTROL_LOCK) /* enabled and locked */
3193	\|\| !(val & MSR_IA32_FEATURE_CONTROL_LOCK) /* not enabled, but not locked either */
3194	)
3195	{
3196	VMX_CAPABILITY vtCaps;
3197
3198	*pfCaps \|= SUPVTCAPS_VT_X;
3199
3200	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS);
3201	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC_USE_SECONDARY_EXEC_CTRL)
3202	{
3203	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS2);
3204	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC2_EPT)
3205	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3206	}
3207	return VINF_SUCCESS;
3208	}
3209	return VERR_VMX_MSR_LOCKED_OR_DISABLED;
3210	}
3211	return VERR_VMX_NO_VMX;
3212	}
3213
3214	if ( u32VendorEBX == X86_CPUID_VENDOR_AMD_EBX
3215	&& u32VendorECX == X86_CPUID_VENDOR_AMD_ECX
3216	&& u32VendorEDX == X86_CPUID_VENDOR_AMD_EDX
3217	)
3218	{
3219	if ( (u32AMDFeatureECX & X86_CPUID_AMD_FEATURE_ECX_SVM)
3220	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3221	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3222	)
3223	{
3224	/* Check if SVM is disabled */
3225	val = ASMRdMsr(MSR_K8_VM_CR);
3226	if (!(val & MSR_K8_VM_CR_SVM_DISABLE))
3227	{
3228	*pfCaps \|= SUPVTCAPS_AMD_V;
3229
3230	/* Query AMD features. */
3231	ASMCpuId(0x8000000A, &u32Dummy, &u32Dummy, &u32Dummy, &u32FeaturesEDX);
3232
3233	if (u32FeaturesEDX & AMD_CPUID_SVM_FEATURE_EDX_NESTED_PAGING)
3234	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3235
3236	return VINF_SUCCESS;
3237	}
3238	return VERR_SVM_DISABLED;
3239	}
3240	return VERR_SVM_NO_SVM;
3241	}
3242	}
3243
3244	return VERR_UNSUPPORTED_CPU;
3245	}
3246
3247
3248	/**
3249	* (Re-)initializes the per-cpu structure prior to starting or resuming the GIP
3250	* updating.
3251	*
3252	* @param pGipCpu The per CPU structure for this CPU.
3253	* @param u64NanoTS The current time.
3254	*/
3255	static void supdrvGipReInitCpu(PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
3256	{
3257	pGipCpu->u64TSC = ASMReadTSC() - pGipCpu->u32UpdateIntervalTSC;
3258	pGipCpu->u64NanoTS = u64NanoTS;
3259	}
3260
3261
3262	/**
3263	* Set the current TSC and NanoTS value for the CPU.
3264	*
3265	* @param idCpu The CPU ID. Unused - we have to use the APIC ID.
3266	* @param pvUser1 Pointer to the ring-0 GIP mapping.
3267	* @param pvUser2 Pointer to the variable holding the current time.
3268	*/
3269	static DECLCALLBACK(void) supdrvGipReInitCpuCallback(RTCPUID idCpu, void pvUser1, void pvUser2)
3270	{
3271	PSUPGLOBALINFOPAGE pGip = (PSUPGLOBALINFOPAGE)pvUser1;
3272	unsigned iCpu = ASMGetApicId();
3273
3274	if (RT_LIKELY(iCpu < RT_ELEMENTS(pGip->aCPUs)))
3275	supdrvGipReInitCpu(&pGip->aCPUs[iCpu], (uint64_t )pvUser2);
3276
3277	NOREF(pvUser2);
3278	NOREF(idCpu);
3279	}
3280
3281
3282	/**
3283	* Maps the GIP into userspace and/or get the physical address of the GIP.
3284	*
3285	* @returns IPRT status code.
3286	* @param pSession Session to which the GIP mapping should belong.
3287	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
3288	* @param pHCPhysGip Where to store the physical address. (optional)
3289	*
3290	* @remark There is no reference counting on the mapping, so one call to this function
3291	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
3292	* and remove the session as a GIP user.
3293	*/
3294	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
3295	{
3296	int rc;
3297	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3298	RTR3PTR pGipR3 = NIL_RTR3PTR;
3299	RTHCPHYS HCPhys = NIL_RTHCPHYS;
3300	LogFlow(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
3301
3302	/*
3303	* Validate
3304	*/
3305	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3306	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
3307	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
3308
3309	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3310	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3311	#else
3312	RTSemFastMutexRequest(pDevExt->mtxGip);
3313	#endif
3314	if (pDevExt->pGip)
3315	{
3316	/*
3317	* Map it?
3318	*/
3319	rc = VINF_SUCCESS;
3320	if (ppGipR3)
3321	{
3322	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
3323	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
3324	RTMEM_PROT_READ, RTR0ProcHandleSelf());
3325	if (RT_SUCCESS(rc))
3326	pGipR3 = RTR0MemObjAddressR3(pSession->GipMapObjR3);
3327	}
3328
3329	/*
3330	* Get physical address.
3331	*/
3332	if (pHCPhysGip && RT_SUCCESS(rc))
3333	HCPhys = pDevExt->HCPhysGip;
3334
3335	/*
3336	* Reference globally.
3337	*/
3338	if (!pSession->fGipReferenced && RT_SUCCESS(rc))
3339	{
3340	pSession->fGipReferenced = 1;
3341	pDevExt->cGipUsers++;
3342	if (pDevExt->cGipUsers == 1)
3343	{
3344	PSUPGLOBALINFOPAGE pGipR0 = pDevExt->pGip;
3345	uint64_t u64NanoTS;
3346	uint32_t u32SystemResolution;
3347	unsigned i;
3348
3349	LogFlow(("SUPR0GipMap: Resumes GIP updating\n"));
3350
3351	/*
3352	* Try bump up the system timer resolution.
3353	* The more interrupts the better...
3354	*/
3355	if ( RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
3356	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
3357	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1953125 /* 512 HZ */, &u32SystemResolution))
3358	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 2000000 /* 500 HZ */, &u32SystemResolution))
3359	)
3360	{
3361	Assert(RTTimerGetSystemGranularity() <= u32SystemResolution);
3362	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
3363	}
3364
3365	if (pGipR0->aCPUs[0].u32TransactionId != 2 /* not the first time */)
3366	{
3367	for (i = 0; i < RT_ELEMENTS(pGipR0->aCPUs); i++)
3368	ASMAtomicUoWriteU32(&pGipR0->aCPUs[i].u32TransactionId,
3369	(pGipR0->aCPUs[i].u32TransactionId + GIP_UPDATEHZ_RECALC_FREQ * 2)
3370	& ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
3371	ASMAtomicWriteU64(&pGipR0->u64NanoTSLastUpdateHz, 0);
3372	}
3373
3374	u64NanoTS = RTTimeSystemNanoTS() - pGipR0->u32UpdateIntervalNS;
3375	if ( pGipR0->u32Mode == SUPGIPMODE_SYNC_TSC
3376	\|\| RTMpGetOnlineCount() == 1)
3377	supdrvGipReInitCpu(&pGipR0->aCPUs[0], u64NanoTS);
3378	else
3379	RTMpOnAll(supdrvGipReInitCpuCallback, pGipR0, &u64NanoTS);
3380
3381	#ifndef DO_NOT_START_GIP
3382	rc = RTTimerStart(pDevExt->pGipTimer, 0); AssertRC(rc);
3383	#endif
3384	rc = VINF_SUCCESS;
3385	}
3386	}
3387	}
3388	else
3389	{
3390	rc = VERR_GENERAL_FAILURE;
3391	Log(("SUPR0GipMap: GIP is not available!\n"));
3392	}
3393	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3394	RTSemMutexRelease(pDevExt->mtxGip);
3395	#else
3396	RTSemFastMutexRelease(pDevExt->mtxGip);
3397	#endif
3398
3399	/*
3400	* Write returns.
3401	*/
3402	if (pHCPhysGip)
3403	*pHCPhysGip = HCPhys;
3404	if (ppGipR3)
3405	*ppGipR3 = pGipR3;
3406
3407	#ifdef DEBUG_DARWIN_GIP
3408	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3409	#else
3410	LogFlow(( "SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3411	#endif
3412	return rc;
3413	}
3414
3415
3416	/**
3417	* Unmaps any user mapping of the GIP and terminates all GIP access
3418	* from this session.
3419	*
3420	* @returns IPRT status code.
3421	* @param pSession Session to which the GIP mapping should belong.
3422	*/
3423	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
3424	{
3425	int rc = VINF_SUCCESS;
3426	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3427	#ifdef DEBUG_DARWIN_GIP
3428	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
3429	pSession,
3430	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
3431	pSession->GipMapObjR3));
3432	#else
3433	LogFlow(("SUPR0GipUnmap: pSession=%p\n", pSession));
3434	#endif
3435	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3436
3437	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3438	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3439	#else
3440	RTSemFastMutexRequest(pDevExt->mtxGip);
3441	#endif
3442
3443	/*
3444	* Unmap anything?
3445	*/
3446	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
3447	{
3448	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
3449	AssertRC(rc);
3450	if (RT_SUCCESS(rc))
3451	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
3452	}
3453
3454	/*
3455	* Dereference global GIP.
3456	*/
3457	if (pSession->fGipReferenced && !rc)
3458	{
3459	pSession->fGipReferenced = 0;
3460	if ( pDevExt->cGipUsers > 0
3461	&& !--pDevExt->cGipUsers)
3462	{
3463	LogFlow(("SUPR0GipUnmap: Suspends GIP updating\n"));
3464	#ifndef DO_NOT_START_GIP
3465	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = VINF_SUCCESS;
3466	#endif
3467
3468	if (pDevExt->u32SystemTimerGranularityGrant)
3469	{
3470	int rc2 = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
3471	pDevExt->u32SystemTimerGranularityGrant = 0;
3472	}
3473	}
3474	}
3475
3476	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3477	RTSemMutexRelease(pDevExt->mtxGip);
3478	#else
3479	RTSemFastMutexRelease(pDevExt->mtxGip);
3480	#endif
3481
3482	return rc;
3483	}
3484
3485
3486	/**
3487	* Gets the GIP pointer.
3488	*
3489	* @returns Pointer to the GIP or NULL.
3490	*/
3491	SUPDECL(PSUPGLOBALINFOPAGE) SUPGetGIP(void)
3492	{
3493	return g_pSUPGlobalInfoPage;
3494	}
3495
3496
3497	/**
3498	* Register a component factory with the support driver.
3499	*
3500	* This is currently restricted to kernel sessions only.
3501	*
3502	* @returns VBox status code.
3503	* @retval VINF_SUCCESS on success.
3504	* @retval VERR_NO_MEMORY if we're out of memory.
3505	* @retval VERR_ALREADY_EXISTS if the factory has already been registered.
3506	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3507	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3508	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3509	*
3510	* @param pSession The SUPDRV session (must be a ring-0 session).
3511	* @param pFactory Pointer to the component factory registration structure.
3512	*
3513	* @remarks This interface is also available via SUPR0IdcComponentRegisterFactory.
3514	*/
3515	SUPR0DECL(int) SUPR0ComponentRegisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3516	{
3517	PSUPDRVFACTORYREG pNewReg;
3518	const char *psz;
3519	int rc;
3520
3521	/*
3522	* Validate parameters.
3523	*/
3524	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3525	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3526	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3527	AssertPtrReturn(pFactory->pfnQueryFactoryInterface, VERR_INVALID_POINTER);
3528	psz = RTStrEnd(pFactory->szName, sizeof(pFactory->szName));
3529	AssertReturn(psz, VERR_INVALID_PARAMETER);
3530
3531	/*
3532	* Allocate and initialize a new registration structure.
3533	*/
3534	pNewReg = (PSUPDRVFACTORYREG)RTMemAlloc(sizeof(SUPDRVFACTORYREG));
3535	if (pNewReg)
3536	{
3537	pNewReg->pNext = NULL;
3538	pNewReg->pFactory = pFactory;
3539	pNewReg->pSession = pSession;
3540	pNewReg->cchName = psz - &pFactory->szName[0];
3541
3542	/*
3543	* Add it to the tail of the list after checking for prior registration.
3544	*/
3545	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3546	if (RT_SUCCESS(rc))
3547	{
3548	PSUPDRVFACTORYREG pPrev = NULL;
3549	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3550	while (pCur && pCur->pFactory != pFactory)
3551	{
3552	pPrev = pCur;
3553	pCur = pCur->pNext;
3554	}
3555	if (!pCur)
3556	{
3557	if (pPrev)
3558	pPrev->pNext = pNewReg;
3559	else
3560	pSession->pDevExt->pComponentFactoryHead = pNewReg;
3561	rc = VINF_SUCCESS;
3562	}
3563	else
3564	rc = VERR_ALREADY_EXISTS;
3565
3566	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3567	}
3568
3569	if (RT_FAILURE(rc))
3570	RTMemFree(pNewReg);
3571	}
3572	else
3573	rc = VERR_NO_MEMORY;
3574	return rc;
3575	}
3576
3577
3578	/**
3579	* Deregister a component factory.
3580	*
3581	* @returns VBox status code.
3582	* @retval VINF_SUCCESS on success.
3583	* @retval VERR_NOT_FOUND if the factory wasn't registered.
3584	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3585	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3586	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3587	*
3588	* @param pSession The SUPDRV session (must be a ring-0 session).
3589	* @param pFactory Pointer to the component factory registration structure
3590	* previously passed SUPR0ComponentRegisterFactory().
3591	*
3592	* @remarks This interface is also available via SUPR0IdcComponentDeregisterFactory.
3593	*/
3594	SUPR0DECL(int) SUPR0ComponentDeregisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3595	{
3596	int rc;
3597
3598	/*
3599	* Validate parameters.
3600	*/
3601	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3602	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3603	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3604
3605	/*
3606	* Take the lock and look for the registration record.
3607	*/
3608	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3609	if (RT_SUCCESS(rc))
3610	{
3611	PSUPDRVFACTORYREG pPrev = NULL;
3612	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3613	while (pCur && pCur->pFactory != pFactory)
3614	{
3615	pPrev = pCur;
3616	pCur = pCur->pNext;
3617	}
3618	if (pCur)
3619	{
3620	if (!pPrev)
3621	pSession->pDevExt->pComponentFactoryHead = pCur->pNext;
3622	else
3623	pPrev->pNext = pCur->pNext;
3624
3625	pCur->pNext = NULL;
3626	pCur->pFactory = NULL;
3627	pCur->pSession = NULL;
3628	rc = VINF_SUCCESS;
3629	}
3630	else
3631	rc = VERR_NOT_FOUND;
3632
3633	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3634
3635	RTMemFree(pCur);
3636	}
3637	return rc;
3638	}
3639
3640
3641	/**
3642	* Queries a component factory.
3643	*
3644	* @returns VBox status code.
3645	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3646	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3647	* @retval VERR_SUPDRV_COMPONENT_NOT_FOUND if the component factory wasn't found.
3648	* @retval VERR_SUPDRV_INTERFACE_NOT_SUPPORTED if the interface wasn't supported.
3649	*
3650	* @param pSession The SUPDRV session.
3651	* @param pszName The name of the component factory.
3652	* @param pszInterfaceUuid The UUID of the factory interface (stringified).
3653	* @param ppvFactoryIf Where to store the factory interface.
3654	*/
3655	SUPR0DECL(int) SUPR0ComponentQueryFactory(PSUPDRVSESSION pSession, const char pszName, const char pszInterfaceUuid, void **ppvFactoryIf)
3656	{
3657	const char *pszEnd;
3658	size_t cchName;
3659	int rc;
3660
3661	/*
3662	* Validate parameters.
3663	*/
3664	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3665
3666	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
3667	pszEnd = RTStrEnd(pszName, RT_SIZEOFMEMB(SUPDRVFACTORY, szName));
3668	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3669	cchName = pszEnd - pszName;
3670
3671	AssertPtrReturn(pszInterfaceUuid, VERR_INVALID_POINTER);
3672	pszEnd = RTStrEnd(pszInterfaceUuid, RTUUID_STR_LENGTH);
3673	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3674
3675	AssertPtrReturn(ppvFactoryIf, VERR_INVALID_POINTER);
3676	*ppvFactoryIf = NULL;
3677
3678	/*
3679	* Take the lock and try all factories by this name.
3680	*/
3681	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3682	if (RT_SUCCESS(rc))
3683	{
3684	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3685	rc = VERR_SUPDRV_COMPONENT_NOT_FOUND;
3686	while (pCur)
3687	{
3688	if ( pCur->cchName == cchName
3689	&& !memcmp(pCur->pFactory->szName, pszName, cchName))
3690	{
3691	void *pvFactory = pCur->pFactory->pfnQueryFactoryInterface(pCur->pFactory, pSession, pszInterfaceUuid);
3692	if (pvFactory)
3693	{
3694	*ppvFactoryIf = pvFactory;
3695	rc = VINF_SUCCESS;
3696	break;
3697	}
3698	rc = VERR_SUPDRV_INTERFACE_NOT_SUPPORTED;
3699	}
3700
3701	/* next */
3702	pCur = pCur->pNext;
3703	}
3704
3705	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3706	}
3707	return rc;
3708	}
3709
3710
3711	/**
3712	* Adds a memory object to the session.
3713	*
3714	* @returns IPRT status code.
3715	* @param pMem Memory tracking structure containing the
3716	* information to track.
3717	* @param pSession The session.
3718	*/
3719	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
3720	{
3721	PSUPDRVBUNDLE pBundle;
3722	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3723
3724	/*
3725	* Find free entry and record the allocation.
3726	*/
3727	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3728	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3729	{
3730	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
3731	{
3732	unsigned i;
3733	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3734	{
3735	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
3736	{
3737	pBundle->cUsed++;
3738	pBundle->aMem[i] = *pMem;
3739	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3740	return VINF_SUCCESS;
3741	}
3742	}
3743	AssertFailed(); /* !!this can't be happening!!! */
3744	}
3745	}
3746	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3747
3748	/*
3749	* Need to allocate a new bundle.
3750	* Insert into the last entry in the bundle.
3751	*/
3752	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
3753	if (!pBundle)
3754	return VERR_NO_MEMORY;
3755
3756	/* take last entry. */
3757	pBundle->cUsed++;
3758	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
3759
3760	/* insert into list. */
3761	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3762	pBundle->pNext = pSession->Bundle.pNext;
3763	pSession->Bundle.pNext = pBundle;
3764	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3765
3766	return VINF_SUCCESS;
3767	}
3768
3769
3770	/**
3771	* Releases a memory object referenced by pointer and type.
3772	*
3773	* @returns IPRT status code.
3774	* @param pSession Session data.
3775	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
3776	* @param eType Memory type.
3777	*/
3778	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
3779	{
3780	PSUPDRVBUNDLE pBundle;
3781	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3782
3783	/*
3784	* Validate input.
3785	*/
3786	if (!uPtr)
3787	{
3788	Log(("Illegal address %p\n", (void *)uPtr));
3789	return VERR_INVALID_PARAMETER;
3790	}
3791
3792	/*
3793	* Search for the address.
3794	*/
3795	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3796	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3797	{
3798	if (pBundle->cUsed > 0)
3799	{
3800	unsigned i;
3801	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3802	{
3803	if ( pBundle->aMem[i].eType == eType
3804	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3805	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
3806	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3807	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
3808	)
3809	{
3810	/* Make a copy of it and release it outside the spinlock. */
3811	SUPDRVMEMREF Mem = pBundle->aMem[i];
3812	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
3813	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
3814	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
3815	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3816
3817	if (Mem.MapObjR3 != NIL_RTR0MEMOBJ)
3818	{
3819	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
3820	AssertRC(rc); /** @todo figure out how to handle this. */
3821	}
3822	if (Mem.MemObj != NIL_RTR0MEMOBJ)
3823	{
3824	int rc = RTR0MemObjFree(Mem.MemObj, true /* fFreeMappings */);
3825	AssertRC(rc); /** @todo figure out how to handle this. */
3826	}
3827	return VINF_SUCCESS;
3828	}
3829	}
3830	}
3831	}
3832	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3833	Log(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
3834	return VERR_INVALID_PARAMETER;
3835	}
3836
3837
3838	/**
3839	* Opens an image. If it's the first time it's opened the call must upload
3840	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
3841	*
3842	* This is the 1st step of the loading.
3843	*
3844	* @returns IPRT status code.
3845	* @param pDevExt Device globals.
3846	* @param pSession Session data.
3847	* @param pReq The open request.
3848	*/
3849	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
3850	{
3851	int rc;
3852	PSUPDRVLDRIMAGE pImage;
3853	void *pv;
3854	size_t cchName = strlen(pReq->u.In.szName); /* (caller checked < 32). */
3855	LogFlow(("supdrvIOCtl_LdrOpen: szName=%s cbImageWithTabs=%d\n", pReq->u.In.szName, pReq->u.In.cbImageWithTabs));
3856
3857	/*
3858	* Check if we got an instance of the image already.
3859	*/
3860	supdrvLdrLock(pDevExt);
3861	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
3862	{
3863	if ( pImage->szName[cchName] == '\0'
3864	&& !memcmp(pImage->szName, pReq->u.In.szName, cchName))
3865	{
3866	/** @todo check cbImageBits and cbImageWithTabs here, if they differs that indicates that the images are different. */
3867	pImage->cUsage++;
3868	pReq->u.Out.pvImageBase = pImage->pvImage;
3869	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
3870	pReq->u.Out.fNativeLoader = pImage->fNative;
3871	supdrvLdrAddUsage(pSession, pImage);
3872	supdrvLdrUnlock(pDevExt);
3873	return VINF_SUCCESS;
3874	}
3875	}
3876	/* (not found - add it!) */
3877
3878	/*
3879	* Allocate memory.
3880	*/
3881	pv = RTMemAlloc(RT_OFFSETOF(SUPDRVLDRIMAGE, szName[cchName + 1]));
3882	if (!pv)
3883	{
3884	supdrvLdrUnlock(pDevExt);
3885	Log(("supdrvIOCtl_LdrOpen: RTMemAlloc() failed\n"));
3886	return /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_2;
3887	}
3888
3889	/*
3890	* Setup and link in the LDR stuff.
3891	*/
3892	pImage = (PSUPDRVLDRIMAGE)pv;
3893	pImage->pvImage = NULL;
3894	pImage->pvImageAlloc = NULL;
3895	pImage->cbImageWithTabs = pReq->u.In.cbImageWithTabs;
3896	pImage->cbImageBits = pReq->u.In.cbImageBits;
3897	pImage->cSymbols = 0;
3898	pImage->paSymbols = NULL;
3899	pImage->pachStrTab = NULL;
3900	pImage->cbStrTab = 0;
3901	pImage->pfnModuleInit = NULL;
3902	pImage->pfnModuleTerm = NULL;
3903	pImage->pfnServiceReqHandler = NULL;
3904	pImage->uState = SUP_IOCTL_LDR_OPEN;
3905	pImage->cUsage = 1;
3906	memcpy(pImage->szName, pReq->u.In.szName, cchName + 1);
3907
3908	/*
3909	* Try load it using the native loader, if that isn't supported, fall back
3910	* on the older method.
3911	*/
3912	pImage->fNative = true;
3913	rc = supdrvOSLdrOpen(pDevExt, pImage, pReq->u.In.szFilename);
3914	if (rc == VERR_NOT_SUPPORTED)
3915	{
3916	pImage->pvImageAlloc = RTMemExecAlloc(pImage->cbImageBits + 31);
3917	pImage->pvImage = RT_ALIGN_P(pImage->pvImageAlloc, 32);
3918	pImage->fNative = false;
3919	rc = pImage->pvImageAlloc ? VINF_SUCCESS : VERR_NO_EXEC_MEMORY;
3920	}
3921	if (RT_FAILURE(rc))
3922	{
3923	supdrvLdrUnlock(pDevExt);
3924	RTMemFree(pImage);
3925	Log(("supdrvIOCtl_LdrOpen(%s): failed - %Rrc\n", pReq->u.In.szName, rc));
3926	return rc;
3927	}
3928	Assert(VALID_PTR(pImage->pvImage) \|\| RT_FAILURE(rc));
3929
3930	/*
3931	* Link it.
3932	*/
3933	pImage->pNext = pDevExt->pLdrImages;
3934	pDevExt->pLdrImages = pImage;
3935
3936	supdrvLdrAddUsage(pSession, pImage);
3937
3938	pReq->u.Out.pvImageBase = pImage->pvImage;
3939	pReq->u.Out.fNeedsLoading = true;
3940	pReq->u.Out.fNativeLoader = pImage->fNative;
3941	supdrvLdrUnlock(pDevExt);
3942
3943	#if defined(RT_OS_WINDOWS) && defined(DEBUG)
3944	SUPR0Printf("VBoxDrv: windbg> .reload /f %s=%#p\n", pImage->szName, pImage->pvImage);
3945	#endif
3946	return VINF_SUCCESS;
3947	}
3948
3949
3950	/**
3951	* Worker that validates a pointer to an image entrypoint.
3952	*
3953	* @returns IPRT status code.
3954	* @param pDevExt The device globals.
3955	* @param pImage The loader image.
3956	* @param pv The pointer into the image.
3957	* @param fMayBeNull Whether it may be NULL.
3958	* @param pszWhat What is this entrypoint? (for logging)
3959	* @param pbImageBits The image bits prepared by ring-3.
3960	*
3961	* @remarks Will leave the lock on failure.
3962	*/
3963	static int supdrvLdrValidatePointer(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage, void *pv,
3964	bool fMayBeNull, const uint8_t pbImageBits, const char pszWhat)
3965	{
3966	if (!fMayBeNull \|\| pv)
3967	{
3968	if ((uintptr_t)pv - (uintptr_t)pImage->pvImage >= pImage->cbImageBits)
3969	{
3970	supdrvLdrUnlock(pDevExt);
3971	Log(("Out of range (%p LB %#x): %s=%p\n", pImage->pvImage, pImage->cbImageBits, pszWhat, pv));
3972	return VERR_INVALID_PARAMETER;
3973	}
3974
3975	if (pImage->fNative)
3976	{
3977	int rc = supdrvOSLdrValidatePointer(pDevExt, pImage, pv, pbImageBits);
3978	if (RT_FAILURE(rc))
3979	{
3980	supdrvLdrUnlock(pDevExt);
3981	Log(("Bad entry point address: %s=%p (rc=%Rrc)\n", pszWhat, pv, rc));
3982	return rc;
3983	}
3984	}
3985	}
3986	return VINF_SUCCESS;
3987	}
3988
3989
3990	/**
3991	* Loads the image bits.
3992	*
3993	* This is the 2nd step of the loading.
3994	*
3995	* @returns IPRT status code.
3996	* @param pDevExt Device globals.
3997	* @param pSession Session data.
3998	* @param pReq The request.
3999	*/
4000	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
4001	{
4002	PSUPDRVLDRUSAGE pUsage;
4003	PSUPDRVLDRIMAGE pImage;
4004	int rc;
4005	LogFlow(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImageWithBits=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImageWithTabs));
4006
4007	/*
4008	* Find the ldr image.
4009	*/
4010	supdrvLdrLock(pDevExt);
4011	pUsage = pSession->pLdrUsage;
4012	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4013	pUsage = pUsage->pNext;
4014	if (!pUsage)
4015	{
4016	supdrvLdrUnlock(pDevExt);
4017	Log(("SUP_IOCTL_LDR_LOAD: couldn't find image!\n"));
4018	return VERR_INVALID_HANDLE;
4019	}
4020	pImage = pUsage->pImage;
4021
4022	/*
4023	* Validate input.
4024	*/
4025	if ( pImage->cbImageWithTabs != pReq->u.In.cbImageWithTabs
4026	\|\| pImage->cbImageBits != pReq->u.In.cbImageBits)
4027	{
4028	supdrvLdrUnlock(pDevExt);
4029	Log(("SUP_IOCTL_LDR_LOAD: image size mismatch!! %d(prep) != %d(load) or %d != %d\n",
4030	pImage->cbImageWithTabs, pReq->u.In.cbImageWithTabs, pImage->cbImageBits, pReq->u.In.cbImageBits));
4031	return VERR_INVALID_HANDLE;
4032	}
4033
4034	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
4035	{
4036	unsigned uState = pImage->uState;
4037	supdrvLdrUnlock(pDevExt);
4038	if (uState != SUP_IOCTL_LDR_LOAD)
4039	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
4040	return VERR_ALREADY_LOADED;
4041	}
4042
4043	switch (pReq->u.In.eEPType)
4044	{
4045	case SUPLDRLOADEP_NOTHING:
4046	break;
4047
4048	case SUPLDRLOADEP_VMMR0:
4049	rc = supdrvLdrValidatePointer( pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0, false, pReq->u.In.abImage, "pvVMMR0");
4050	if (RT_SUCCESS(rc))
4051	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt, false, pReq->u.In.abImage, "pvVMMR0EntryInt");
4052	if (RT_SUCCESS(rc))
4053	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, false, pReq->u.In.abImage, "pvVMMR0EntryFast");
4054	if (RT_SUCCESS(rc))
4055	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx, false, pReq->u.In.abImage, "pvVMMR0EntryEx");
4056	if (RT_FAILURE(rc))
4057	return rc;
4058	break;
4059
4060	case SUPLDRLOADEP_SERVICE:
4061	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.Service.pfnServiceReq, false, pReq->u.In.abImage, "pfnServiceReq");
4062	if (RT_FAILURE(rc))
4063	return rc;
4064	if ( pReq->u.In.EP.Service.apvReserved[0] != NIL_RTR0PTR
4065	\|\| pReq->u.In.EP.Service.apvReserved[1] != NIL_RTR0PTR
4066	\|\| pReq->u.In.EP.Service.apvReserved[2] != NIL_RTR0PTR)
4067	{
4068	supdrvLdrUnlock(pDevExt);
4069	Log(("Out of range (%p LB %#x): apvReserved={%p,%p,%p} MBZ!\n",
4070	pImage->pvImage, pReq->u.In.cbImageWithTabs,
4071	pReq->u.In.EP.Service.apvReserved[0],
4072	pReq->u.In.EP.Service.apvReserved[1],
4073	pReq->u.In.EP.Service.apvReserved[2]));
4074	return VERR_INVALID_PARAMETER;
4075	}
4076	break;
4077
4078	default:
4079	supdrvLdrUnlock(pDevExt);
4080	Log(("Invalid eEPType=%d\n", pReq->u.In.eEPType));
4081	return VERR_INVALID_PARAMETER;
4082	}
4083
4084	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleInit, true, pReq->u.In.abImage, "pfnModuleInit");
4085	if (RT_FAILURE(rc))
4086	return rc;
4087	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleTerm, true, pReq->u.In.abImage, "pfnModuleTerm");
4088	if (RT_FAILURE(rc))
4089	return rc;
4090
4091	/*
4092	* Allocate and copy the tables.
4093	* (No need to do try/except as this is a buffered request.)
4094	*/
4095	pImage->cbStrTab = pReq->u.In.cbStrTab;
4096	if (pImage->cbStrTab)
4097	{
4098	pImage->pachStrTab = (char *)RTMemAlloc(pImage->cbStrTab);
4099	if (pImage->pachStrTab)
4100	memcpy(pImage->pachStrTab, &pReq->u.In.abImage[pReq->u.In.offStrTab], pImage->cbStrTab);
4101	else
4102	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_3;
4103	}
4104
4105	pImage->cSymbols = pReq->u.In.cSymbols;
4106	if (RT_SUCCESS(rc) && pImage->cSymbols)
4107	{
4108	size_t cbSymbols = pImage->cSymbols * sizeof(SUPLDRSYM);
4109	pImage->paSymbols = (PSUPLDRSYM)RTMemAlloc(cbSymbols);
4110	if (pImage->paSymbols)
4111	memcpy(pImage->paSymbols, &pReq->u.In.abImage[pReq->u.In.offSymbols], cbSymbols);
4112	else
4113	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_4;
4114	}
4115
4116	/*
4117	* Copy the bits / complete native loading.
4118	*/
4119	if (RT_SUCCESS(rc))
4120	{
4121	pImage->uState = SUP_IOCTL_LDR_LOAD;
4122	pImage->pfnModuleInit = pReq->u.In.pfnModuleInit;
4123	pImage->pfnModuleTerm = pReq->u.In.pfnModuleTerm;
4124
4125	if (pImage->fNative)
4126	rc = supdrvOSLdrLoad(pDevExt, pImage, pReq->u.In.abImage);
4127	else
4128	memcpy(pImage->pvImage, &pReq->u.In.abImage[0], pImage->cbImageBits);
4129	}
4130
4131	/*
4132	* Update any entry points.
4133	*/
4134	if (RT_SUCCESS(rc))
4135	{
4136	switch (pReq->u.In.eEPType)
4137	{
4138	default:
4139	case SUPLDRLOADEP_NOTHING:
4140	rc = VINF_SUCCESS;
4141	break;
4142	case SUPLDRLOADEP_VMMR0:
4143	rc = supdrvLdrSetVMMR0EPs(pDevExt, pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
4144	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
4145	break;
4146	case SUPLDRLOADEP_SERVICE:
4147	pImage->pfnServiceReqHandler = pReq->u.In.EP.Service.pfnServiceReq;
4148	rc = VINF_SUCCESS;
4149	break;
4150	}
4151	}
4152
4153	/*
4154	* On success call the module initialization.
4155	*/
4156	LogFlow(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
4157	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
4158	{
4159	Log(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
4160	rc = pImage->pfnModuleInit();
4161	if (rc && pDevExt->pvVMMR0 == pImage->pvImage)
4162	supdrvLdrUnsetVMMR0EPs(pDevExt);
4163	}
4164
4165	if (RT_FAILURE(rc))
4166	{
4167	pImage->uState = SUP_IOCTL_LDR_OPEN;
4168	pImage->pfnModuleInit = NULL;
4169	pImage->pfnModuleTerm = NULL;
4170	pImage->pfnServiceReqHandler= NULL;
4171	pImage->cbStrTab = 0;
4172	RTMemFree(pImage->pachStrTab);
4173	pImage->pachStrTab = NULL;
4174	RTMemFree(pImage->paSymbols);
4175	pImage->paSymbols = NULL;
4176	pImage->cSymbols = 0;
4177	}
4178
4179	supdrvLdrUnlock(pDevExt);
4180	return rc;
4181	}
4182
4183
4184	/**
4185	* Frees a previously loaded (prep'ed) image.
4186	*
4187	* @returns IPRT status code.
4188	* @param pDevExt Device globals.
4189	* @param pSession Session data.
4190	* @param pReq The request.
4191	*/
4192	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
4193	{
4194	int rc;
4195	PSUPDRVLDRUSAGE pUsagePrev;
4196	PSUPDRVLDRUSAGE pUsage;
4197	PSUPDRVLDRIMAGE pImage;
4198	LogFlow(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
4199
4200	/*
4201	* Find the ldr image.
4202	*/
4203	supdrvLdrLock(pDevExt);
4204	pUsagePrev = NULL;
4205	pUsage = pSession->pLdrUsage;
4206	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4207	{
4208	pUsagePrev = pUsage;
4209	pUsage = pUsage->pNext;
4210	}
4211	if (!pUsage)
4212	{
4213	supdrvLdrUnlock(pDevExt);
4214	Log(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
4215	return VERR_INVALID_HANDLE;
4216	}
4217
4218	/*
4219	* Check if we can remove anything.
4220	*/
4221	rc = VINF_SUCCESS;
4222	pImage = pUsage->pImage;
4223	if (pImage->cUsage <= 1 \|\| pUsage->cUsage <= 1)
4224	{
4225	/*
4226	* Check if there are any objects with destructors in the image, if
4227	* so leave it for the session cleanup routine so we get a chance to
4228	* clean things up in the right order and not leave them all dangling.
4229	*/
4230	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
4231	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
4232	if (pImage->cUsage <= 1)
4233	{
4234	PSUPDRVOBJ pObj;
4235	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4236	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4237	{
4238	rc = VERR_DANGLING_OBJECTS;
4239	break;
4240	}
4241	}
4242	else
4243	{
4244	PSUPDRVUSAGE pGenUsage;
4245	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
4246	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4247	{
4248	rc = VERR_DANGLING_OBJECTS;
4249	break;
4250	}
4251	}
4252	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
4253	if (rc == VINF_SUCCESS)
4254	{
4255	/* unlink it */
4256	if (pUsagePrev)
4257	pUsagePrev->pNext = pUsage->pNext;
4258	else
4259	pSession->pLdrUsage = pUsage->pNext;
4260
4261	/* free it */
4262	pUsage->pImage = NULL;
4263	pUsage->pNext = NULL;
4264	RTMemFree(pUsage);
4265
4266	/*
4267	* Dereference the image.
4268	*/
4269	if (pImage->cUsage <= 1)
4270	supdrvLdrFree(pDevExt, pImage);
4271	else
4272	pImage->cUsage--;
4273	}
4274	else
4275	{
4276	Log(("supdrvIOCtl_LdrFree: Dangling objects in %p/%s!\n", pImage->pvImage, pImage->szName));
4277	rc = VINF_SUCCESS; /** @todo BRANCH-2.1: remove this after branching. */
4278	}
4279	}
4280	else
4281	{
4282	/*
4283	* Dereference both image and usage.
4284	*/
4285	pImage->cUsage--;
4286	pUsage->cUsage--;
4287	}
4288
4289	supdrvLdrUnlock(pDevExt);
4290	return rc;
4291	}
4292
4293
4294	/**
4295	* Gets the address of a symbol in an open image.
4296	*
4297	* @returns IPRT status code.
4298	* @param pDevExt Device globals.
4299	* @param pSession Session data.
4300	* @param pReq The request buffer.
4301	*/
4302	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
4303	{
4304	PSUPDRVLDRIMAGE pImage;
4305	PSUPDRVLDRUSAGE pUsage;
4306	uint32_t i;
4307	PSUPLDRSYM paSyms;
4308	const char *pchStrings;
4309	const size_t cbSymbol = strlen(pReq->u.In.szSymbol) + 1;
4310	void *pvSymbol = NULL;
4311	int rc = VERR_GENERAL_FAILURE;
4312	Log3(("supdrvIOCtl_LdrGetSymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
4313
4314	/*
4315	* Find the ldr image.
4316	*/
4317	supdrvLdrLock(pDevExt);
4318	pUsage = pSession->pLdrUsage;
4319	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4320	pUsage = pUsage->pNext;
4321	if (!pUsage)
4322	{
4323	supdrvLdrUnlock(pDevExt);
4324	Log(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
4325	return VERR_INVALID_HANDLE;
4326	}
4327	pImage = pUsage->pImage;
4328	if (pImage->uState != SUP_IOCTL_LDR_LOAD)
4329	{
4330	unsigned uState = pImage->uState;
4331	supdrvLdrUnlock(pDevExt);
4332	Log(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", uState, uState)); NOREF(uState);
4333	return VERR_ALREADY_LOADED;
4334	}
4335
4336	/*
4337	* Search the symbol strings.
4338	*/
4339	pchStrings = pImage->pachStrTab;
4340	paSyms = pImage->paSymbols;
4341	for (i = 0; i < pImage->cSymbols; i++)
4342	{
4343	if ( paSyms[i].offSymbol < pImage->cbImageBits /* paranoia */
4344	&& paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4345	&& !memcmp(pchStrings + paSyms[i].offName, pReq->u.In.szSymbol, cbSymbol))
4346	{
4347	pvSymbol = (uint8_t *)pImage->pvImage + paSyms[i].offSymbol;
4348	rc = VINF_SUCCESS;
4349	break;
4350	}
4351	}
4352	supdrvLdrUnlock(pDevExt);
4353	pReq->u.Out.pvSymbol = pvSymbol;
4354	return rc;
4355	}
4356
4357
4358	/**
4359	* Gets the address of a symbol in an open image or the support driver.
4360	*
4361	* @returns VINF_SUCCESS on success.
4362	* @returns
4363	* @param pDevExt Device globals.
4364	* @param pSession Session data.
4365	* @param pReq The request buffer.
4366	*/
4367	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq)
4368	{
4369	int rc = VINF_SUCCESS;
4370	const char *pszSymbol = pReq->u.In.pszSymbol;
4371	const char *pszModule = pReq->u.In.pszModule;
4372	size_t cbSymbol;
4373	char const *pszEnd;
4374	uint32_t i;
4375
4376	/*
4377	* Input validation.
4378	*/
4379	AssertPtrReturn(pszSymbol, VERR_INVALID_POINTER);
4380	pszEnd = RTStrEnd(pszSymbol, 512);
4381	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4382	cbSymbol = pszEnd - pszSymbol + 1;
4383
4384	if (pszModule)
4385	{
4386	AssertPtrReturn(pszModule, VERR_INVALID_POINTER);
4387	pszEnd = RTStrEnd(pszModule, 64);
4388	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4389	}
4390	Log3(("supdrvIDC_LdrGetSymbol: pszModule=%p:{%s} pszSymbol=%p:{%s}\n", pszModule, pszModule, pszSymbol, pszSymbol));
4391
4392
4393	if ( !pszModule
4394	\|\| !strcmp(pszModule, "SupDrv"))
4395	{
4396	/*
4397	* Search the support driver export table.
4398	*/
4399	for (i = 0; i < RT_ELEMENTS(g_aFunctions); i++)
4400	if (!strcmp(g_aFunctions[i].szName, pszSymbol))
4401	{
4402	pReq->u.Out.pfnSymbol = g_aFunctions[i].pfn;
4403	break;
4404	}
4405	}
4406	else
4407	{
4408	/*
4409	* Find the loader image.
4410	*/
4411	PSUPDRVLDRIMAGE pImage;
4412
4413	supdrvLdrLock(pDevExt);
4414
4415	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
4416	if (!strcmp(pImage->szName, pszModule))
4417	break;
4418	if (pImage && pImage->uState == SUP_IOCTL_LDR_LOAD)
4419	{
4420	/*
4421	* Search the symbol strings.
4422	*/
4423	const char *pchStrings = pImage->pachStrTab;
4424	PCSUPLDRSYM paSyms = pImage->paSymbols;
4425	for (i = 0; i < pImage->cSymbols; i++)
4426	{
4427	if ( paSyms[i].offSymbol < pImage->cbImageBits /* paranoia */
4428	&& paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4429	&& !memcmp(pchStrings + paSyms[i].offName, pszSymbol, cbSymbol))
4430	{
4431	/*
4432	* Found it! Calc the symbol address and add a reference to the module.
4433	*/
4434	pReq->u.Out.pfnSymbol = (PFNRT)((uint8_t *)pImage->pvImage + paSyms[i].offSymbol);
4435	rc = supdrvLdrAddUsage(pSession, pImage);
4436	break;
4437	}
4438	}
4439	}
4440	else
4441	rc = pImage ? VERR_WRONG_ORDER : VERR_MODULE_NOT_FOUND;
4442
4443	supdrvLdrUnlock(pDevExt);
4444	}
4445	return rc;
4446	}
4447
4448
4449	/**
4450	* Updates the VMMR0 entry point pointers.
4451	*
4452	* @returns IPRT status code.
4453	* @param pDevExt Device globals.
4454	* @param pSession Session data.
4455	* @param pVMMR0 VMMR0 image handle.
4456	* @param pvVMMR0EntryInt VMMR0EntryInt address.
4457	* @param pvVMMR0EntryFast VMMR0EntryFast address.
4458	* @param pvVMMR0EntryEx VMMR0EntryEx address.
4459	* @remark Caller must own the loader mutex.
4460	*/
4461	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx)
4462	{
4463	int rc = VINF_SUCCESS;
4464	LogFlow(("supdrvLdrSetR0EP pvVMMR0=%p pvVMMR0EntryInt=%p\n", pvVMMR0, pvVMMR0EntryInt));
4465
4466
4467	/*
4468	* Check if not yet set.
4469	*/
4470	if (!pDevExt->pvVMMR0)
4471	{
4472	pDevExt->pvVMMR0 = pvVMMR0;
4473	pDevExt->pfnVMMR0EntryInt = pvVMMR0EntryInt;
4474	pDevExt->pfnVMMR0EntryFast = pvVMMR0EntryFast;
4475	pDevExt->pfnVMMR0EntryEx = pvVMMR0EntryEx;
4476	}
4477	else
4478	{
4479	/*
4480	* Return failure or success depending on whether the values match or not.
4481	*/
4482	if ( pDevExt->pvVMMR0 != pvVMMR0
4483	\|\| (void *)pDevExt->pfnVMMR0EntryInt != pvVMMR0EntryInt
4484	\|\| (void *)pDevExt->pfnVMMR0EntryFast != pvVMMR0EntryFast
4485	\|\| (void *)pDevExt->pfnVMMR0EntryEx != pvVMMR0EntryEx)
4486	{
4487	AssertMsgFailed(("SUP_IOCTL_LDR_SETR0EP: Already set pointing to a different module!\n"));
4488	rc = VERR_INVALID_PARAMETER;
4489	}
4490	}
4491	return rc;
4492	}
4493
4494
4495	/**
4496	* Unsets the VMMR0 entry point installed by supdrvLdrSetR0EP.
4497	*
4498	* @param pDevExt Device globals.
4499	*/
4500	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt)
4501	{
4502	pDevExt->pvVMMR0 = NULL;
4503	pDevExt->pfnVMMR0EntryInt = NULL;
4504	pDevExt->pfnVMMR0EntryFast = NULL;
4505	pDevExt->pfnVMMR0EntryEx = NULL;
4506	}
4507
4508
4509	/**
4510	* Adds a usage reference in the specified session of an image.
4511	*
4512	* Called while owning the loader semaphore.
4513	*
4514	* @returns VINF_SUCCESS on success and VERR_NO_MEMORY on failure.
4515	* @param pSession Session in question.
4516	* @param pImage Image which the session is using.
4517	*/
4518	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage)
4519	{
4520	PSUPDRVLDRUSAGE pUsage;
4521	LogFlow(("supdrvLdrAddUsage: pImage=%p\n", pImage));
4522
4523	/*
4524	* Referenced it already?
4525	*/
4526	pUsage = pSession->pLdrUsage;
4527	while (pUsage)
4528	{
4529	if (pUsage->pImage == pImage)
4530	{
4531	pUsage->cUsage++;
4532	return VINF_SUCCESS;
4533	}
4534	pUsage = pUsage->pNext;
4535	}
4536
4537	/*
4538	* Allocate new usage record.
4539	*/
4540	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
4541	AssertReturn(pUsage, /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_5);
4542	pUsage->cUsage = 1;
4543	pUsage->pImage = pImage;
4544	pUsage->pNext = pSession->pLdrUsage;
4545	pSession->pLdrUsage = pUsage;
4546	return VINF_SUCCESS;
4547	}
4548
4549
4550	/**
4551	* Frees a load image.
4552	*
4553	* @param pDevExt Pointer to device extension.
4554	* @param pImage Pointer to the image we're gonna free.
4555	* This image must exit!
4556	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
4557	*/
4558	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
4559	{
4560	PSUPDRVLDRIMAGE pImagePrev;
4561	LogFlow(("supdrvLdrFree: pImage=%p\n", pImage));
4562
4563	/* find it - arg. should've used doubly linked list. */
4564	Assert(pDevExt->pLdrImages);
4565	pImagePrev = NULL;
4566	if (pDevExt->pLdrImages != pImage)
4567	{
4568	pImagePrev = pDevExt->pLdrImages;
4569	while (pImagePrev->pNext != pImage)
4570	pImagePrev = pImagePrev->pNext;
4571	Assert(pImagePrev->pNext == pImage);
4572	}
4573
4574	/* unlink */
4575	if (pImagePrev)
4576	pImagePrev->pNext = pImage->pNext;
4577	else
4578	pDevExt->pLdrImages = pImage->pNext;
4579
4580	/* check if this is VMMR0.r0 unset its entry point pointers. */
4581	if (pDevExt->pvVMMR0 == pImage->pvImage)
4582	supdrvLdrUnsetVMMR0EPs(pDevExt);
4583
4584	/* check for objects with destructors in this image. (Shouldn't happen.) */
4585	if (pDevExt->pObjs)
4586	{
4587	unsigned cObjs = 0;
4588	PSUPDRVOBJ pObj;
4589	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
4590	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
4591	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4592	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4593	{
4594	pObj->pfnDestructor = NULL;
4595	cObjs++;
4596	}
4597	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
4598	if (cObjs)
4599	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
4600	}
4601
4602	/* call termination function if fully loaded. */
4603	if ( pImage->pfnModuleTerm
4604	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
4605	{
4606	LogFlow(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
4607	pImage->pfnModuleTerm();
4608	}
4609
4610	/* do native unload if appropriate. */
4611	if (pImage->fNative)
4612	supdrvOSLdrUnload(pDevExt, pImage);
4613
4614	/* free the image */
4615	pImage->cUsage = 0;
4616	pImage->pNext = 0;
4617	pImage->uState = SUP_IOCTL_LDR_FREE;
4618	RTMemExecFree(pImage->pvImageAlloc, pImage->cbImageBits + 31);
4619	pImage->pvImageAlloc = NULL;
4620	RTMemFree(pImage->pachStrTab);
4621	pImage->pachStrTab = NULL;
4622	RTMemFree(pImage->paSymbols);
4623	pImage->paSymbols = NULL;
4624	RTMemFree(pImage);
4625	}
4626
4627
4628	/**
4629	* Acquires the loader lock.
4630	*
4631	* @returns IPRT status code.
4632	* @param pDevExt The device extension.
4633	*/
4634	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt)
4635	{
4636	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4637	int rc = RTSemMutexRequest(pDevExt->mtxLdr, RT_INDEFINITE_WAIT);
4638	#else
4639	int rc = RTSemFastMutexRequest(pDevExt->mtxLdr);
4640	#endif
4641	AssertRC(rc);
4642	return rc;
4643	}
4644
4645
4646	/**
4647	* Releases the loader lock.
4648	*
4649	* @returns IPRT status code.
4650	* @param pDevExt The device extension.
4651	*/
4652	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt)
4653	{
4654	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4655	return RTSemMutexRelease(pDevExt->mtxLdr);
4656	#else
4657	return RTSemFastMutexRelease(pDevExt->mtxLdr);
4658	#endif
4659	}
4660
4661
4662	/**
4663	* Implements the service call request.
4664	*
4665	* @returns VBox status code.
4666	* @param pDevExt The device extension.
4667	* @param pSession The calling session.
4668	* @param pReq The request packet, valid.
4669	*/
4670	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq)
4671	{
4672	#if !defined(RT_OS_WINDOWS) \|\| defined(DEBUG)
4673	int rc;
4674
4675	/*
4676	* Find the module first in the module referenced by the calling session.
4677	*/
4678	rc = supdrvLdrLock(pDevExt);
4679	if (RT_SUCCESS(rc))
4680	{
4681	PFNSUPR0SERVICEREQHANDLER pfnServiceReqHandler = NULL;
4682	PSUPDRVLDRUSAGE pUsage;
4683
4684	for (pUsage = pSession->pLdrUsage; pUsage; pUsage = pUsage->pNext)
4685	if ( pUsage->pImage->pfnServiceReqHandler
4686	&& !strcmp(pUsage->pImage->szName, pReq->u.In.szName))
4687	{
4688	pfnServiceReqHandler = pUsage->pImage->pfnServiceReqHandler;
4689	break;
4690	}
4691	supdrvLdrUnlock(pDevExt);
4692
4693	if (pfnServiceReqHandler)
4694	{
4695	/*
4696	* Call it.
4697	*/
4698	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
4699	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, NULL);
4700	else
4701	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0]);
4702	}
4703	else
4704	rc = VERR_SUPDRV_SERVICE_NOT_FOUND;
4705	}
4706
4707	/* log it */
4708	if ( RT_FAILURE(rc)
4709	&& rc != VERR_INTERRUPTED
4710	&& rc != VERR_TIMEOUT)
4711	Log(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4712	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4713	else
4714	Log4(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4715	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4716	return rc;
4717	#else /* RT_OS_WINDOWS && !DEBUG */
4718	return VERR_NOT_IMPLEMENTED;
4719	#endif /* RT_OS_WINDOWS && !DEBUG */
4720	}
4721
4722
4723	/**
4724	* Implements the logger settings request.
4725	*
4726	* @returns VBox status code.
4727	* @param pDevExt The device extension.
4728	* @param pSession The caller's session.
4729	* @param pReq The request.
4730	*/
4731	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq)
4732	{
4733	const char *pszGroup = &pReq->u.In.szStrings[pReq->u.In.offGroups];
4734	const char *pszFlags = &pReq->u.In.szStrings[pReq->u.In.offFlags];
4735	const char *pszDest = &pReq->u.In.szStrings[pReq->u.In.offDestination];
4736	PRTLOGGER pLogger = NULL;
4737	int rc;
4738
4739	/*
4740	* Some further validation.
4741	*/
4742	switch (pReq->u.In.fWhat)
4743	{
4744	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4745	case SUPLOGGERSETTINGS_WHAT_CREATE:
4746	break;
4747
4748	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4749	if (pszGroup \|\| pszFlags \|\| *pszDest)
4750	return VERR_INVALID_PARAMETER;
4751	if (pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_RELEASE)
4752	return VERR_ACCESS_DENIED;
4753	break;
4754
4755	default:
4756	return VERR_INTERNAL_ERROR;
4757	}
4758
4759	/*
4760	* Get the logger.
4761	*/
4762	switch (pReq->u.In.fWhich)
4763	{
4764	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4765	pLogger = RTLogGetDefaultInstance();
4766	break;
4767
4768	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4769	pLogger = RTLogRelDefaultInstance();
4770	break;
4771
4772	default:
4773	return VERR_INTERNAL_ERROR;
4774	}
4775
4776	/*
4777	* Do the job.
4778	*/
4779	switch (pReq->u.In.fWhat)
4780	{
4781	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4782	if (pLogger)
4783	{
4784	rc = RTLogFlags(pLogger, pszFlags);
4785	if (RT_SUCCESS(rc))
4786	rc = RTLogGroupSettings(pLogger, pszGroup);
4787	NOREF(pszDest);
4788	}
4789	else
4790	rc = VERR_NOT_FOUND;
4791	break;
4792
4793	case SUPLOGGERSETTINGS_WHAT_CREATE:
4794	{
4795	if (pLogger)
4796	rc = VERR_ALREADY_EXISTS;
4797	else
4798	{
4799	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
4800
4801	rc = RTLogCreate(&pLogger,
4802	0 /* fFlags */,
4803	pszGroup,
4804	pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_DEBUG
4805	? "VBOX_LOG"
4806	: "VBOX_RELEASE_LOG",
4807	RT_ELEMENTS(s_apszGroups),
4808	s_apszGroups,
4809	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER,
4810	NULL);
4811	if (RT_SUCCESS(rc))
4812	{
4813	rc = RTLogFlags(pLogger, pszFlags);
4814	NOREF(pszDest);
4815	if (RT_SUCCESS(rc))
4816	{
4817	switch (pReq->u.In.fWhich)
4818	{
4819	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4820	pLogger = RTLogSetDefaultInstance(pLogger);
4821	break;
4822	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4823	pLogger = RTLogRelSetDefaultInstance(pLogger);
4824	break;
4825	}
4826	}
4827	RTLogDestroy(pLogger);
4828	}
4829	}
4830	break;
4831	}
4832
4833	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4834	switch (pReq->u.In.fWhich)
4835	{
4836	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4837	pLogger = RTLogSetDefaultInstance(NULL);
4838	break;
4839	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4840	pLogger = RTLogRelSetDefaultInstance(NULL);
4841	break;
4842	}
4843	rc = RTLogDestroy(pLogger);
4844	break;
4845
4846	default:
4847	{
4848	rc = VERR_INTERNAL_ERROR;
4849	break;
4850	}
4851	}
4852
4853	return rc;
4854	}
4855
4856
4857	/**
4858	* Creates the GIP.
4859	*
4860	* @returns VBox status code.
4861	* @param pDevExt Instance data. GIP stuff may be updated.
4862	*/
4863	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
4864	{
4865	PSUPGLOBALINFOPAGE pGip;
4866	RTHCPHYS HCPhysGip;
4867	uint32_t u32SystemResolution;
4868	uint32_t u32Interval;
4869	int rc;
4870
4871	LogFlow(("supdrvGipCreate:\n"));
4872
4873	/* assert order */
4874	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
4875	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
4876	Assert(!pDevExt->pGipTimer);
4877
4878	/*
4879	* Allocate a suitable page with a default kernel mapping.
4880	*/
4881	rc = RTR0MemObjAllocLow(&pDevExt->GipMemObj, PAGE_SIZE, false);
4882	if (RT_FAILURE(rc))
4883	{
4884	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
4885	return rc;
4886	}
4887	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
4888	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
4889
4890	/*
4891	* Find a reasonable update interval and initialize the structure.
4892	*/
4893	u32Interval = u32SystemResolution = RTTimerGetSystemGranularity();
4894	while (u32Interval < 10000000 /* 10 ms */)
4895	u32Interval += u32SystemResolution;
4896
4897	supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), 1000000000 / u32Interval /=Hz/);
4898
4899	/*
4900	* Create the timer.
4901	* If CPU_ALL isn't supported we'll have to fall back to synchronous mode.
4902	*/
4903	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
4904	{
4905	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, RTTIMER_FLAGS_CPU_ALL, supdrvGipAsyncTimer, pDevExt);
4906	if (rc == VERR_NOT_SUPPORTED)
4907	{
4908	OSDBGPRINT(("supdrvGipCreate: omni timer not supported, falling back to synchronous mode\n"));
4909	pGip->u32Mode = SUPGIPMODE_SYNC_TSC;
4910	}
4911	}
4912	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
4913	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0, supdrvGipSyncTimer, pDevExt);
4914	if (RT_SUCCESS(rc))
4915	{
4916	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
4917	rc = RTMpNotificationRegister(supdrvGipMpEvent, pDevExt);
4918	if (RT_SUCCESS(rc))
4919	{
4920	/*
4921	* We're good.
4922	*/
4923	Log(("supdrvGipCreate: %ld ns interval.\n", (long)u32Interval));
4924	g_pSUPGlobalInfoPage = pGip;
4925	return VINF_SUCCESS;
4926	}
4927
4928	OSDBGPRINT(("supdrvGipCreate: failed register MP event notfication. rc=%d\n", rc));
4929	}
4930	else
4931	{
4932	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %ld ns interval. rc=%d\n", (long)u32Interval, rc));
4933	Assert(!pDevExt->pGipTimer);
4934	}
4935	supdrvGipDestroy(pDevExt);
4936	return rc;
4937	}
4938
4939
4940	/**
4941	* Terminates the GIP.
4942	*
4943	* @param pDevExt Instance data. GIP stuff may be updated.
4944	*/
4945	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
4946	{
4947	int rc;
4948	#ifdef DEBUG_DARWIN_GIP
4949	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
4950	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
4951	pDevExt->pGipTimer, pDevExt->GipMemObj));
4952	#endif
4953
4954	/*
4955	* Invalid the GIP data.
4956	*/
4957	if (pDevExt->pGip)
4958	{
4959	supdrvGipTerm(pDevExt->pGip);
4960	pDevExt->pGip = NULL;
4961	}
4962	g_pSUPGlobalInfoPage = NULL;
4963
4964	/*
4965	* Destroy the timer and free the GIP memory object.
4966	*/
4967	if (pDevExt->pGipTimer)
4968	{
4969	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
4970	pDevExt->pGipTimer = NULL;
4971	}
4972
4973	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
4974	{
4975	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
4976	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
4977	}
4978
4979	/*
4980	* Finally, make sure we've release the system timer resolution request
4981	* if one actually succeeded and is still pending.
4982	*/
4983	if (pDevExt->u32SystemTimerGranularityGrant)
4984	{
4985	rc = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
4986	pDevExt->u32SystemTimerGranularityGrant = 0;
4987	}
4988	}
4989
4990
4991	/**
4992	* Timer callback function sync GIP mode.
4993	* @param pTimer The timer.
4994	* @param pvUser The device extension.
4995	*/
4996	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
4997	{
4998	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
4999	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5000	uint64_t u64TSC = ASMReadTSC();
5001	uint64_t NanoTS = RTTimeSystemNanoTS();
5002
5003	supdrvGipUpdate(pDevExt->pGip, NanoTS, u64TSC, iTick);
5004
5005	ASMSetFlags(fOldFlags);
5006	}
5007
5008
5009	/**
5010	* Timer callback function for async GIP mode.
5011	* @param pTimer The timer.
5012	* @param pvUser The device extension.
5013	*/
5014	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
5015	{
5016	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
5017	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5018	RTCPUID idCpu = RTMpCpuId();
5019	uint64_t u64TSC = ASMReadTSC();
5020	uint64_t NanoTS = RTTimeSystemNanoTS();
5021
5022	/** @todo reset the transaction number and whatnot when iTick == 1. */
5023	if (pDevExt->idGipMaster == idCpu)
5024	supdrvGipUpdate(pDevExt->pGip, NanoTS, u64TSC, iTick);
5025	else
5026	supdrvGipUpdatePerCpu(pDevExt->pGip, NanoTS, u64TSC, ASMGetApicId(), iTick);
5027
5028	ASMSetFlags(fOldFlags);
5029	}
5030
5031
5032	/**
5033	* Multiprocessor event notification callback.
5034	*
5035	* This is used to make sue that the GIP master gets passed on to
5036	* another CPU.
5037	*
5038	* @param enmEvent The event.
5039	* @param idCpu The cpu it applies to.
5040	* @param pvUser Pointer to the device extension.
5041	*/
5042	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser)
5043	{
5044	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5045	if (enmEvent == RTMPEVENT_OFFLINE)
5046	{
5047	RTCPUID idGipMaster;
5048	ASMAtomicReadSize(&pDevExt->idGipMaster, &idGipMaster);
5049	if (idGipMaster == idCpu)
5050	{
5051	/*
5052	* Find a new GIP master.
5053	*/
5054	bool fIgnored;
5055	unsigned i;
5056	RTCPUID idNewGipMaster = NIL_RTCPUID;
5057	RTCPUSET OnlineCpus;
5058	RTMpGetOnlineSet(&OnlineCpus);
5059
5060	for (i = 0; i < RTCPUSET_MAX_CPUS; i++)
5061	{
5062	RTCPUID idCurCpu = RTMpCpuIdFromSetIndex(i);
5063	if ( RTCpuSetIsMember(&OnlineCpus, idCurCpu)
5064	&& idCurCpu != idGipMaster)
5065	{
5066	idNewGipMaster = idCurCpu;
5067	break;
5068	}
5069	}
5070
5071	Log(("supdrvGipMpEvent: Gip master %#lx -> %#lx\n", (long)idGipMaster, (long)idNewGipMaster));
5072	ASMAtomicCmpXchgSize(&pDevExt->idGipMaster, idNewGipMaster, idGipMaster, fIgnored);
5073	NOREF(fIgnored);
5074	}
5075	}
5076	}
5077
5078
5079	/**
5080	* Callback used by supdrvDetermineAsyncTSC to read the TSC on a CPU.
5081	*
5082	* @param idCpu Ignored.
5083	* @param pvUser1 Where to put the TSC.
5084	* @param pvUser2 Ignored.
5085	*/
5086	static DECLCALLBACK(void) supdrvDetermineAsyncTscWorker(RTCPUID idCpu, void pvUser1, void pvUser2)
5087	{
5088	#if 1
5089	ASMAtomicWriteU64((uint64_t volatile *)pvUser1, ASMReadTSC());
5090	#else
5091	(uint64_t )pvUser1 = ASMReadTSC();
5092	#endif
5093	}
5094
5095
5096	/**
5097	* Determine if Async GIP mode is required because of TSC drift.
5098	*
5099	* When using the default/normal timer code it is essential that the time stamp counter
5100	* (TSC) runs never backwards, that is, a read operation to the counter should return
5101	* a bigger value than any previous read operation. This is guaranteed by the latest
5102	* AMD CPUs and by newer Intel CPUs which never enter the C2 state (P4). In any other
5103	* case we have to choose the asynchronous timer mode.
5104	*
5105	* @param poffMin Pointer to the determined difference between different cores.
5106	* @return false if the time stamp counters appear to be synchronized, true otherwise.
5107	*/
5108	static bool supdrvDetermineAsyncTsc(uint64_t *poffMin)
5109	{
5110	/*
5111	* Just iterate all the cpus 8 times and make sure that the TSC is
5112	* ever increasing. We don't bother taking TSC rollover into account.
5113	*/
5114	int iEndCpu = RTMpGetArraySize();
5115	int iCpu;
5116	int cLoops = 8;
5117	bool fAsync = false;
5118	int rc = VINF_SUCCESS;
5119	uint64_t offMax = 0;
5120	uint64_t offMin = ~(uint64_t)0;
5121	uint64_t PrevTsc = ASMReadTSC();
5122
5123	while (cLoops-- > 0)
5124	{
5125	for (iCpu = 0; iCpu < iEndCpu; iCpu++)
5126	{
5127	uint64_t CurTsc;
5128	rc = RTMpOnSpecific(RTMpCpuIdFromSetIndex(iCpu), supdrvDetermineAsyncTscWorker, &CurTsc, NULL);
5129	if (RT_SUCCESS(rc))
5130	{
5131	if (CurTsc <= PrevTsc)
5132	{
5133	fAsync = true;
5134	offMin = offMax = PrevTsc - CurTsc;
5135	Log(("supdrvDetermineAsyncTsc: iCpu=%d cLoops=%d CurTsc=%llx PrevTsc=%llx\n",
5136	iCpu, cLoops, CurTsc, PrevTsc));
5137	break;
5138	}
5139
5140	/* Gather statistics (except the first time). */
5141	if (iCpu != 0 \|\| cLoops != 7)
5142	{
5143	uint64_t off = CurTsc - PrevTsc;
5144	if (off < offMin)
5145	offMin = off;
5146	if (off > offMax)
5147	offMax = off;
5148	Log2(("%d/%d: off=%llx\n", cLoops, iCpu, off));
5149	}
5150
5151	/* Next */
5152	PrevTsc = CurTsc;
5153	}
5154	else if (rc == VERR_NOT_SUPPORTED)
5155	break;
5156	else
5157	AssertMsg(rc == VERR_CPU_NOT_FOUND \|\| rc == VERR_CPU_OFFLINE, ("%d\n", rc));
5158	}
5159
5160	/* broke out of the loop. */
5161	if (iCpu < iEndCpu)
5162	break;
5163	}
5164
5165	poffMin = offMin; / Almost RTMpOnSpecific profiling. */
5166	Log(("supdrvDetermineAsyncTsc: returns %d; iEndCpu=%d rc=%d offMin=%llx offMax=%llx\n",
5167	fAsync, iEndCpu, rc, offMin, offMax));
5168	#if !defined(RT_OS_SOLARIS) && !defined(RT_OS_OS2) && !defined(RT_OS_WINDOWS)
5169	OSDBGPRINT(("vboxdrv: fAsync=%d offMin=%#lx offMax=%#lx\n", fAsync, (long)offMin, (long)offMax));
5170	#endif
5171	return fAsync;
5172	}
5173
5174
5175	/**
5176	* Determine the GIP TSC mode.
5177	*
5178	* @returns The most suitable TSC mode.
5179	* @param pDevExt Pointer to the device instance data.
5180	*/
5181	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt)
5182	{
5183	/*
5184	* On SMP we're faced with two problems:
5185	* (1) There might be a skew between the CPU, so that cpu0
5186	* returns a TSC that is slightly different from cpu1.
5187	* (2) Power management (and other things) may cause the TSC
5188	* to run at a non-constant speed, and cause the speed
5189	* to be different on the cpus. This will result in (1).
5190	*
5191	* So, on SMP systems we'll have to select the ASYNC update method
5192	* if there are symptoms of these problems.
5193	*/
5194	if (RTMpGetCount() > 1)
5195	{
5196	uint32_t uEAX, uEBX, uECX, uEDX;
5197	uint64_t u64DiffCoresIgnored;
5198
5199	/* Permit the user and/or the OS specific bits to force async mode. */
5200	if (supdrvOSGetForcedAsyncTscMode(pDevExt))
5201	return SUPGIPMODE_ASYNC_TSC;
5202
5203	/* Try check for current differences between the cpus. */
5204	if (supdrvDetermineAsyncTsc(&u64DiffCoresIgnored))
5205	return SUPGIPMODE_ASYNC_TSC;
5206
5207	/*
5208	* If the CPU supports power management and is an AMD one we
5209	* won't trust it unless it has the TscInvariant bit is set.
5210	*/
5211	/* Check for "AuthenticAMD" */
5212	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
5213	if ( uEAX >= 1
5214	&& uEBX == X86_CPUID_VENDOR_AMD_EBX
5215	&& uECX == X86_CPUID_VENDOR_AMD_ECX
5216	&& uEDX == X86_CPUID_VENDOR_AMD_EDX)
5217	{
5218	/* Check for APM support and that TscInvariant is cleared. */
5219	ASMCpuId(0x80000000, &uEAX, &uEBX, &uECX, &uEDX);
5220	if (uEAX >= 0x80000007)
5221	{
5222	ASMCpuId(0x80000007, &uEAX, &uEBX, &uECX, &uEDX);
5223	if ( !(uEDX & RT_BIT(8))/* TscInvariant */
5224	&& (uEDX & 0x3e)) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
5225	return SUPGIPMODE_ASYNC_TSC;
5226	}
5227	}
5228	}
5229	return SUPGIPMODE_SYNC_TSC;
5230	}
5231
5232
5233
5234	/**
5235	* Initializes the GIP data.
5236	*
5237	* @param pDevExt Pointer to the device instance data.
5238	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5239	* @param HCPhys The physical address of the GIP.
5240	* @param u64NanoTS The current nanosecond timestamp.
5241	* @param uUpdateHz The update frequency.
5242	*/
5243	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys, uint64_t u64NanoTS, unsigned uUpdateHz)
5244	{
5245	unsigned i;
5246	#ifdef DEBUG_DARWIN_GIP
5247	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz));
5248	#else
5249	LogFlow(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz));
5250	#endif
5251
5252	/*
5253	* Initialize the structure.
5254	*/
5255	memset(pGip, 0, PAGE_SIZE);
5256	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
5257	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
5258	pGip->u32Mode = supdrvGipDeterminTscMode(pDevExt);
5259	pGip->u32UpdateHz = uUpdateHz;
5260	pGip->u32UpdateIntervalNS = 1000000000 / uUpdateHz;
5261	pGip->u64NanoTSLastUpdateHz = u64NanoTS;
5262
5263	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
5264	{
5265	pGip->aCPUs[i].u32TransactionId = 2;
5266	pGip->aCPUs[i].u64NanoTS = u64NanoTS;
5267	pGip->aCPUs[i].u64TSC = ASMReadTSC();
5268
5269	/*
5270	* We don't know the following values until we've executed updates.
5271	* So, we'll just pretend it's a 4 GHz CPU and adjust the history it on
5272	* the 2nd timer callout.
5273	*/
5274	pGip->aCPUs[i].u64CpuHz = _4G + 1; /* tstGIP-2 depends on this. */
5275	pGip->aCPUs[i].u32UpdateIntervalTSC
5276	= pGip->aCPUs[i].au32TSCHistory[0]
5277	= pGip->aCPUs[i].au32TSCHistory[1]
5278	= pGip->aCPUs[i].au32TSCHistory[2]
5279	= pGip->aCPUs[i].au32TSCHistory[3]
5280	= pGip->aCPUs[i].au32TSCHistory[4]
5281	= pGip->aCPUs[i].au32TSCHistory[5]
5282	= pGip->aCPUs[i].au32TSCHistory[6]
5283	= pGip->aCPUs[i].au32TSCHistory[7]
5284	= /pGip->aCPUs[i].u64CpuHz/ (uint32_t)(_4G / uUpdateHz);
5285	}
5286
5287	/*
5288	* Link it to the device extension.
5289	*/
5290	pDevExt->pGip = pGip;
5291	pDevExt->HCPhysGip = HCPhys;
5292	pDevExt->cGipUsers = 0;
5293	}
5294
5295
5296	/**
5297	* Invalidates the GIP data upon termination.
5298	*
5299	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5300	*/
5301	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
5302	{
5303	unsigned i;
5304	pGip->u32Magic = 0;
5305	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
5306	{
5307	pGip->aCPUs[i].u64NanoTS = 0;
5308	pGip->aCPUs[i].u64TSC = 0;
5309	pGip->aCPUs[i].iTSCHistoryHead = 0;
5310	}
5311	}
5312
5313
5314	/**
5315	* Worker routine for supdrvGipUpdate and supdrvGipUpdatePerCpu that
5316	* updates all the per cpu data except the transaction id.
5317	*
5318	* @param pGip The GIP.
5319	* @param pGipCpu Pointer to the per cpu data.
5320	* @param u64NanoTS The current time stamp.
5321	* @param u64TSC The current TSC.
5322	* @param iTick The current timer tick.
5323	*/
5324	static void supdrvGipDoUpdateCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick)
5325	{
5326	uint64_t u64TSCDelta;
5327	uint32_t u32UpdateIntervalTSC;
5328	uint32_t u32UpdateIntervalTSCSlack;
5329	unsigned iTSCHistoryHead;
5330	uint64_t u64CpuHz;
5331	uint32_t u32TransactionId;
5332
5333	/* Delta between this and the previous update. */
5334	ASMAtomicUoWriteU32(&pGipCpu->u32PrevUpdateIntervalNS, (uint32_t)(u64NanoTS - pGipCpu->u64NanoTS));
5335
5336	/*
5337	* Update the NanoTS.
5338	*/
5339	ASMAtomicXchgU64(&pGipCpu->u64NanoTS, u64NanoTS);
5340
5341	/*
5342	* Calc TSC delta.
5343	*/
5344	/** @todo validate the NanoTS delta, don't trust the OS to call us when it should... */
5345	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
5346	ASMAtomicXchgU64(&pGipCpu->u64TSC, u64TSC);
5347
5348	if (u64TSCDelta >> 32)
5349	{
5350	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
5351	pGipCpu->cErrors++;
5352	}
5353
5354	/*
5355	* On the 2nd and 3rd callout, reset the history with the current TSC
5356	* interval since the values entered by supdrvGipInit are totally off.
5357	* The interval on the 1st callout completely unreliable, the 2nd is a bit
5358	* better, while the 3rd should be most reliable.
5359	*/
5360	u32TransactionId = pGipCpu->u32TransactionId;
5361	if (RT_UNLIKELY( ( u32TransactionId == 5
5362	\|\| u32TransactionId == 7)
5363	&& ( iTick == 2
5364	\|\| iTick == 3) ))
5365	{
5366	unsigned i;
5367	for (i = 0; i < RT_ELEMENTS(pGipCpu->au32TSCHistory); i++)
5368	ASMAtomicUoWriteU32(&pGipCpu->au32TSCHistory[i], (uint32_t)u64TSCDelta);
5369	}
5370
5371	/*
5372	* TSC History.
5373	*/
5374	Assert(RT_ELEMENTS(pGipCpu->au32TSCHistory) == 8);
5375	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
5376	ASMAtomicXchgU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
5377	ASMAtomicXchgU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
5378
5379	/*
5380	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
5381	*/
5382	if (pGip->u32UpdateHz >= 1000)
5383	{
5384	uint32_t u32;
5385	u32 = pGipCpu->au32TSCHistory[0];
5386	u32 += pGipCpu->au32TSCHistory[1];
5387	u32 += pGipCpu->au32TSCHistory[2];
5388	u32 += pGipCpu->au32TSCHistory[3];
5389	u32 >>= 2;
5390	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
5391	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
5392	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
5393	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
5394	u32UpdateIntervalTSC >>= 2;
5395	u32UpdateIntervalTSC += u32;
5396	u32UpdateIntervalTSC >>= 1;
5397
5398	/* Value chosen for a 2GHz Athlon64 running linux 2.6.10/11, . */
5399	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
5400	}
5401	else if (pGip->u32UpdateHz >= 90)
5402	{
5403	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5404	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
5405	u32UpdateIntervalTSC >>= 1;
5406
5407	/* value chosen on a 2GHz thinkpad running windows */
5408	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
5409	}
5410	else
5411	{
5412	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5413
5414	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
5415	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
5416	}
5417	ASMAtomicXchgU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
5418
5419	/*
5420	* CpuHz.
5421	*/
5422	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, pGip->u32UpdateHz);
5423	ASMAtomicXchgU64(&pGipCpu->u64CpuHz, u64CpuHz);
5424	}
5425
5426
5427	/**
5428	* Updates the GIP.
5429	*
5430	* @param pGip Pointer to the GIP.
5431	* @param u64NanoTS The current nanosecond timesamp.
5432	* @param u64TSC The current TSC timesamp.
5433	* @param iTick The current timer tick.
5434	*/
5435	static void supdrvGipUpdate(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick)
5436	{
5437	/*
5438	* Determine the relevant CPU data.
5439	*/
5440	PSUPGIPCPU pGipCpu;
5441	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
5442	pGipCpu = &pGip->aCPUs[0];
5443	else
5444	{
5445	unsigned iCpu = ASMGetApicId();
5446	if (RT_UNLIKELY(iCpu >= RT_ELEMENTS(pGip->aCPUs)))
5447	return;
5448	pGipCpu = &pGip->aCPUs[iCpu];
5449	}
5450
5451	/*
5452	* Start update transaction.
5453	*/
5454	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5455	{
5456	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
5457	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5458	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5459	pGipCpu->cErrors++;
5460	return;
5461	}
5462
5463	/*
5464	* Recalc the update frequency every 0x800th time.
5465	*/
5466	if (!(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
5467	{
5468	if (pGip->u64NanoTSLastUpdateHz)
5469	{
5470	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
5471	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
5472	uint32_t u32UpdateHz = (uint32_t)((UINT64_C(1000000000) * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
5473	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
5474	{
5475	ASMAtomicXchgU32(&pGip->u32UpdateHz, u32UpdateHz);
5476	ASMAtomicXchgU32(&pGip->u32UpdateIntervalNS, 1000000000 / u32UpdateHz);
5477	}
5478	#endif
5479	}
5480	ASMAtomicXchgU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS);
5481	}
5482
5483	/*
5484	* Update the data.
5485	*/
5486	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS, u64TSC, iTick);
5487
5488	/*
5489	* Complete transaction.
5490	*/
5491	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5492	}
5493
5494
5495	/**
5496	* Updates the per cpu GIP data for the calling cpu.
5497	*
5498	* @param pGip Pointer to the GIP.
5499	* @param u64NanoTS The current nanosecond timesamp.
5500	* @param u64TSC The current TSC timesamp.
5501	* @param iCpu The CPU index.
5502	* @param iTick The current timer tick.
5503	*/
5504	static void supdrvGipUpdatePerCpu(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC, unsigned iCpu, uint64_t iTick)
5505	{
5506	PSUPGIPCPU pGipCpu;
5507
5508	if (RT_LIKELY(iCpu < RT_ELEMENTS(pGip->aCPUs)))
5509	{
5510	pGipCpu = &pGip->aCPUs[iCpu];
5511
5512	/*
5513	* Start update transaction.
5514	*/
5515	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5516	{
5517	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5518	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5519	pGipCpu->cErrors++;
5520	return;
5521	}
5522
5523	/*
5524	* Update the data.
5525	*/
5526	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS, u64TSC, iTick);
5527
5528	/*
5529	* Complete transaction.
5530	*/
5531	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5532	}
5533	}
5534

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

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDrv.c@ 34484

Download in other formats: