SUPDrv.c@ 37049

Last change on this file since 37049 was 36730, checked in by vboxsync, 14 years ago
Solaris/NetFltBow: support for VNIC templates and VLANs.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 212.8 KB

Line
1	/* $Revision: 36730 $ */
2	/** @file
3	* VBoxDrv - The VirtualBox Support Driver - Common code.
4	*/
5
6	/*
7	* Copyright (C) 2006-2011 Oracle Corporation
8	*
9	* This file is part of VirtualBox Open Source Edition (OSE), as
10	* available from http://www.virtualbox.org. This file is free software;
11	* you can redistribute it and/or modify it under the terms of the GNU
12	* General Public License (GPL) as published by the Free Software
13	* Foundation, in version 2 as it comes in the "COPYING" file of the
14	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16	*
17	* The contents of this file may alternatively be used under the terms
18	* of the Common Development and Distribution License Version 1.0
19	* (CDDL) only, as it comes in the "COPYING.CDDL" file of the
20	* VirtualBox OSE distribution, in which case the provisions of the
21	* CDDL are applicable instead of those of the GPL.
22	*
23	* You may elect to license modified versions of this file under the
24	* terms and conditions of either the GPL or the CDDL or both.
25	*/
26
27	/*******************************************************************************
28	* Header Files *
29	*******************************************************************************/
30	#define LOG_GROUP LOG_GROUP_SUP_DRV
31	#define SUPDRV_AGNOSTIC
32	#include "SUPDrvInternal.h"
33	#ifndef PAGE_SHIFT
34	# include <iprt/param.h>
35	#endif
36	#include <iprt/asm.h>
37	#include <iprt/asm-amd64-x86.h>
38	#include <iprt/asm-math.h>
39	#include <iprt/cpuset.h>
40	#include <iprt/handletable.h>
41	#include <iprt/mem.h>
42	#include <iprt/mp.h>
43	#include <iprt/power.h>
44	#include <iprt/process.h>
45	#include <iprt/semaphore.h>
46	#include <iprt/spinlock.h>
47	#include <iprt/thread.h>
48	#include <iprt/uuid.h>
49	#include <iprt/net.h>
50	#include <iprt/crc.h>
51	#include <iprt/string.h>
52	#include <iprt/timer.h>
53	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
54	# include <iprt/rand.h>
55	# include <iprt/path.h>
56	#endif
57
58	#include <VBox/param.h>
59	#include <VBox/log.h>
60	#include <VBox/err.h>
61	#include <VBox/vmm/hwacc_svm.h>
62	#include <VBox/vmm/hwacc_vmx.h>
63	#include <VBox/x86.h>
64
65	#ifdef VBOX_WITH_DTRACE
66	# include "SUPDrv-dtrace.h"
67	#else
68	# define VBOXDRV_SUPDRV_SESSION_CREATE(pvSession, fUser) do { } while (0)
69	# define VBOXDRV_SUPDRV_SESSION_CLOSE(pvSession) do { } while (0)
70	# define VBOXDRV_SUPDRV_IOCCLOSE(pvSession) do { } while (0)
71	# define VBOXDRV_SUPDRV_IOCTL_ENTRY(pvSession, uIOCtl, pvReqHdr) do { } while (0)
72	# define VBOXDRV_SUPDRV_IOCTL_RETURN(pvSession, uIOCtl, pvReqHdr, rcRet, rcReq) do { } while (0)
73	#endif
74
75	/*
76	* Logging assignments:
77	* Log - useful stuff, like failures.
78	* LogFlow - program flow, except the really noisy bits.
79	* Log2 - Cleanup.
80	* Log3 - Loader flow noise.
81	* Log4 - Call VMMR0 flow noise.
82	* Log5 - Native yet-to-be-defined noise.
83	* Log6 - Native ioctl flow noise.
84	*
85	* Logging requires BUILD_TYPE=debug and possibly changes to the logger
86	* instantiation in log-vbox.c(pp).
87	*/
88
89
90	/*******************************************************************************
91	* Defined Constants And Macros *
92	*******************************************************************************/
93	/** The frequency by which we recalculate the u32UpdateHz and
94	* u32UpdateIntervalNS GIP members. The value must be a power of 2. */
95	#define GIP_UPDATEHZ_RECALC_FREQ 0x800
96
97	/** @def VBOX_SVN_REV
98	* The makefile should define this if it can. */
99	#ifndef VBOX_SVN_REV
100	# define VBOX_SVN_REV 0
101	#endif
102
103	#if 0 /* Don't start the GIP timers. Useful when debugging the IPRT timer code. */
104	# define DO_NOT_START_GIP
105	#endif
106
107
108	/*******************************************************************************
109	* Internal Functions *
110	*******************************************************************************/
111	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser);
112	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser);
113	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession);
114	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType);
115	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq);
116	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq);
117	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq);
118	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq);
119	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq);
120	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx);
121	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt);
122	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage);
123	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage);
124	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt);
125	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt);
126	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq);
127	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq);
128	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt);
129	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt);
130	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
131	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
132	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser);
133	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
134	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned cCpus);
135	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2);
136	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip);
137	static void supdrvGipUpdate(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick);
138	static void supdrvGipUpdatePerCpu(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC,
139	RTCPUID idCpu, uint8_t idApic, uint64_t iTick);
140
141
142	/*******************************************************************************
143	* Global Variables *
144	*******************************************************************************/
145	DECLEXPORT(PSUPGLOBALINFOPAGE) g_pSUPGlobalInfoPage = NULL;
146
147	/**
148	* Array of the R0 SUP API.
149	*/
150	static SUPFUNC g_aFunctions[] =
151	{
152	/* name function */
153	/* Entries with absolute addresses determined at runtime, fixup
154	code makes ugly ASSUMPTIONS about the order here: */
155	{ "SUPR0AbsIs64bit", (void *)0 },
156	{ "SUPR0Abs64bitKernelCS", (void *)0 },
157	{ "SUPR0Abs64bitKernelSS", (void *)0 },
158	{ "SUPR0Abs64bitKernelDS", (void *)0 },
159	{ "SUPR0AbsKernelCS", (void *)0 },
160	{ "SUPR0AbsKernelSS", (void *)0 },
161	{ "SUPR0AbsKernelDS", (void *)0 },
162	{ "SUPR0AbsKernelES", (void *)0 },
163	{ "SUPR0AbsKernelFS", (void *)0 },
164	{ "SUPR0AbsKernelGS", (void *)0 },
165	/* Normal function pointers: */
166	{ "SUPR0ComponentRegisterFactory", (void *)SUPR0ComponentRegisterFactory },
167	{ "SUPR0ComponentDeregisterFactory", (void *)SUPR0ComponentDeregisterFactory },
168	{ "SUPR0ComponentQueryFactory", (void *)SUPR0ComponentQueryFactory },
169	{ "SUPR0ObjRegister", (void *)SUPR0ObjRegister },
170	{ "SUPR0ObjAddRef", (void *)SUPR0ObjAddRef },
171	{ "SUPR0ObjAddRefEx", (void *)SUPR0ObjAddRefEx },
172	{ "SUPR0ObjRelease", (void *)SUPR0ObjRelease },
173	{ "SUPR0ObjVerifyAccess", (void *)SUPR0ObjVerifyAccess },
174	{ "SUPR0LockMem", (void *)SUPR0LockMem },
175	{ "SUPR0UnlockMem", (void *)SUPR0UnlockMem },
176	{ "SUPR0ContAlloc", (void *)SUPR0ContAlloc },
177	{ "SUPR0ContFree", (void *)SUPR0ContFree },
178	{ "SUPR0LowAlloc", (void *)SUPR0LowAlloc },
179	{ "SUPR0LowFree", (void *)SUPR0LowFree },
180	{ "SUPR0MemAlloc", (void *)SUPR0MemAlloc },
181	{ "SUPR0MemGetPhys", (void *)SUPR0MemGetPhys },
182	{ "SUPR0MemFree", (void *)SUPR0MemFree },
183	{ "SUPR0PageAllocEx", (void *)SUPR0PageAllocEx },
184	{ "SUPR0PageFree", (void *)SUPR0PageFree },
185	{ "SUPR0Printf", (void )SUPR0Printf }, /* @todo needs wrapping? */
186	{ "SUPSemEventCreate", (void *)SUPSemEventCreate },
187	{ "SUPSemEventClose", (void *)SUPSemEventClose },
188	{ "SUPSemEventSignal", (void *)SUPSemEventSignal },
189	{ "SUPSemEventWait", (void *)SUPSemEventWait },
190	{ "SUPSemEventWaitNoResume", (void *)SUPSemEventWaitNoResume },
191	{ "SUPSemEventWaitNsAbsIntr", (void *)SUPSemEventWaitNsAbsIntr },
192	{ "SUPSemEventWaitNsRelIntr", (void *)SUPSemEventWaitNsRelIntr },
193	{ "SUPSemEventGetResolution", (void *)SUPSemEventGetResolution },
194	{ "SUPSemEventMultiCreate", (void *)SUPSemEventMultiCreate },
195	{ "SUPSemEventMultiClose", (void *)SUPSemEventMultiClose },
196	{ "SUPSemEventMultiSignal", (void *)SUPSemEventMultiSignal },
197	{ "SUPSemEventMultiReset", (void *)SUPSemEventMultiReset },
198	{ "SUPSemEventMultiWait", (void *)SUPSemEventMultiWait },
199	{ "SUPSemEventMultiWaitNoResume", (void *)SUPSemEventMultiWaitNoResume },
200	{ "SUPSemEventMultiWaitNsAbsIntr", (void *)SUPSemEventMultiWaitNsAbsIntr },
201	{ "SUPSemEventMultiWaitNsRelIntr", (void *)SUPSemEventMultiWaitNsRelIntr },
202	{ "SUPSemEventMultiGetResolution", (void *)SUPSemEventMultiGetResolution },
203	{ "SUPR0GetPagingMode", (void *)SUPR0GetPagingMode },
204	{ "SUPR0EnableVTx", (void *)SUPR0EnableVTx },
205	{ "SUPGetGIP", (void *)SUPGetGIP },
206	{ "g_pSUPGlobalInfoPage", (void *)&g_pSUPGlobalInfoPage },
207	{ "RTMemAllocTag", (void *)RTMemAllocTag },
208	{ "RTMemAllocZTag", (void *)RTMemAllocZTag },
209	{ "RTMemAllocVarTag", (void *)RTMemAllocVarTag },
210	{ "RTMemAllocZVarTag", (void *)RTMemAllocZVarTag },
211	{ "RTMemFree", (void *)RTMemFree },
212	{ "RTMemDupTag", (void *)RTMemDupTag },
213	{ "RTMemDupExTag", (void *)RTMemDupExTag },
214	{ "RTMemReallocTag", (void *)RTMemReallocTag },
215	{ "RTR0MemObjAllocLowTag", (void *)RTR0MemObjAllocLowTag },
216	{ "RTR0MemObjAllocPageTag", (void *)RTR0MemObjAllocPageTag },
217	{ "RTR0MemObjAllocPhysTag", (void *)RTR0MemObjAllocPhysTag },
218	{ "RTR0MemObjAllocPhysExTag", (void *)RTR0MemObjAllocPhysExTag },
219	{ "RTR0MemObjAllocPhysNCTag", (void *)RTR0MemObjAllocPhysNCTag },
220	{ "RTR0MemObjAllocContTag", (void *)RTR0MemObjAllocContTag },
221	{ "RTR0MemObjEnterPhysTag", (void *)RTR0MemObjEnterPhysTag },
222	{ "RTR0MemObjLockUserTag", (void *)RTR0MemObjLockUserTag },
223	{ "RTR0MemObjMapKernelTag", (void *)RTR0MemObjMapKernelTag },
224	{ "RTR0MemObjMapKernelExTag", (void *)RTR0MemObjMapKernelExTag },
225	{ "RTR0MemObjMapUserTag", (void *)RTR0MemObjMapUserTag },
226	{ "RTR0MemObjProtect", (void *)RTR0MemObjProtect },
227	{ "RTR0MemObjAddress", (void *)RTR0MemObjAddress },
228	{ "RTR0MemObjAddressR3", (void *)RTR0MemObjAddressR3 },
229	{ "RTR0MemObjSize", (void *)RTR0MemObjSize },
230	{ "RTR0MemObjIsMapping", (void *)RTR0MemObjIsMapping },
231	{ "RTR0MemObjGetPagePhysAddr", (void *)RTR0MemObjGetPagePhysAddr },
232	{ "RTR0MemObjFree", (void *)RTR0MemObjFree },
233	{ "RTR0MemUserCopyFrom", (void *)RTR0MemUserCopyFrom },
234	{ "RTR0MemUserCopyTo", (void *)RTR0MemUserCopyTo },
235	{ "RTR0MemUserIsValidAddr", (void *)RTR0MemUserIsValidAddr },
236	{ "RTR0MemKernelIsValidAddr", (void *)RTR0MemKernelIsValidAddr },
237	{ "RTR0MemAreKrnlAndUsrDifferent", (void *)RTR0MemAreKrnlAndUsrDifferent },
238	{ "RTSemMutexCreate", (void *)RTSemMutexCreate },
239	{ "RTSemMutexRequest", (void *)RTSemMutexRequest },
240	{ "RTSemMutexRequestDebug", (void *)RTSemMutexRequestDebug },
241	{ "RTSemMutexRequestNoResume", (void *)RTSemMutexRequestNoResume },
242	{ "RTSemMutexRequestNoResumeDebug", (void *)RTSemMutexRequestNoResumeDebug },
243	{ "RTSemMutexRelease", (void *)RTSemMutexRelease },
244	{ "RTSemMutexDestroy", (void *)RTSemMutexDestroy },
245	{ "RTProcSelf", (void *)RTProcSelf },
246	{ "RTR0ProcHandleSelf", (void *)RTR0ProcHandleSelf },
247	{ "RTSemFastMutexCreate", (void *)RTSemFastMutexCreate },
248	{ "RTSemFastMutexDestroy", (void *)RTSemFastMutexDestroy },
249	{ "RTSemFastMutexRequest", (void *)RTSemFastMutexRequest },
250	{ "RTSemFastMutexRelease", (void *)RTSemFastMutexRelease },
251	{ "RTSemEventCreate", (void *)RTSemEventCreate },
252	{ "RTSemEventSignal", (void *)RTSemEventSignal },
253	{ "RTSemEventWait", (void *)RTSemEventWait },
254	{ "RTSemEventWaitNoResume", (void *)RTSemEventWaitNoResume },
255	{ "RTSemEventWaitEx", (void *)RTSemEventWaitEx },
256	{ "RTSemEventWaitExDebug", (void *)RTSemEventWaitExDebug },
257	{ "RTSemEventGetResolution", (void *)RTSemEventGetResolution },
258	{ "RTSemEventDestroy", (void *)RTSemEventDestroy },
259	{ "RTSemEventMultiCreate", (void *)RTSemEventMultiCreate },
260	{ "RTSemEventMultiSignal", (void *)RTSemEventMultiSignal },
261	{ "RTSemEventMultiReset", (void *)RTSemEventMultiReset },
262	{ "RTSemEventMultiWait", (void *)RTSemEventMultiWait },
263	{ "RTSemEventMultiWaitNoResume", (void *)RTSemEventMultiWaitNoResume },
264	{ "RTSemEventMultiWaitEx", (void *)RTSemEventMultiWaitEx },
265	{ "RTSemEventMultiWaitExDebug", (void *)RTSemEventMultiWaitExDebug },
266	{ "RTSemEventMultiGetResolution", (void *)RTSemEventMultiGetResolution },
267	{ "RTSemEventMultiDestroy", (void *)RTSemEventMultiDestroy },
268	{ "RTSpinlockCreate", (void *)RTSpinlockCreate },
269	{ "RTSpinlockDestroy", (void *)RTSpinlockDestroy },
270	{ "RTSpinlockAcquire", (void *)RTSpinlockAcquire },
271	{ "RTSpinlockRelease", (void *)RTSpinlockRelease },
272	{ "RTSpinlockAcquireNoInts", (void *)RTSpinlockAcquireNoInts },
273	{ "RTSpinlockReleaseNoInts", (void *)RTSpinlockReleaseNoInts },
274	{ "RTTimeNanoTS", (void *)RTTimeNanoTS },
275	{ "RTTimeMilliTS", (void *)RTTimeMilliTS },
276	{ "RTTimeSystemNanoTS", (void *)RTTimeSystemNanoTS },
277	{ "RTTimeSystemMilliTS", (void *)RTTimeSystemMilliTS },
278	{ "RTThreadNativeSelf", (void *)RTThreadNativeSelf },
279	{ "RTThreadSleep", (void *)RTThreadSleep },
280	{ "RTThreadYield", (void *)RTThreadYield },
281	{ "RTThreadSelf", (void *)RTThreadSelf },
282	{ "RTThreadCreate", (void *)RTThreadCreate },
283	{ "RTThreadGetNative", (void *)RTThreadGetNative },
284	{ "RTThreadWait", (void *)RTThreadWait },
285	{ "RTThreadWaitNoResume", (void *)RTThreadWaitNoResume },
286	{ "RTThreadGetName", (void *)RTThreadGetName },
287	{ "RTThreadSelfName", (void *)RTThreadSelfName },
288	{ "RTThreadGetType", (void *)RTThreadGetType },
289	{ "RTThreadUserSignal", (void *)RTThreadUserSignal },
290	{ "RTThreadUserReset", (void *)RTThreadUserReset },
291	{ "RTThreadUserWait", (void *)RTThreadUserWait },
292	{ "RTThreadUserWaitNoResume", (void *)RTThreadUserWaitNoResume },
293	{ "RTThreadPreemptIsEnabled", (void *)RTThreadPreemptIsEnabled },
294	{ "RTThreadPreemptIsPending", (void *)RTThreadPreemptIsPending },
295	{ "RTThreadPreemptIsPendingTrusty", (void *)RTThreadPreemptIsPendingTrusty },
296	{ "RTThreadPreemptIsPossible", (void *)RTThreadPreemptIsPossible },
297	{ "RTThreadPreemptDisable", (void *)RTThreadPreemptDisable },
298	{ "RTThreadPreemptRestore", (void *)RTThreadPreemptRestore },
299	{ "RTThreadIsInInterrupt", (void *)RTThreadIsInInterrupt },
300	{ "RTTimerCreate", (void *)RTTimerCreate },
301	{ "RTTimerCreateEx", (void *)RTTimerCreateEx },
302	{ "RTTimerDestroy", (void *)RTTimerDestroy },
303	{ "RTTimerStart", (void *)RTTimerStart },
304	{ "RTTimerStop", (void *)RTTimerStop },
305	{ "RTTimerChangeInterval", (void *)RTTimerChangeInterval },
306	{ "RTTimerGetSystemGranularity", (void *)RTTimerGetSystemGranularity },
307	{ "RTTimerRequestSystemGranularity", (void *)RTTimerRequestSystemGranularity },
308	{ "RTTimerReleaseSystemGranularity", (void *)RTTimerReleaseSystemGranularity },
309	{ "RTTimerCanDoHighResolution", (void *)RTTimerCanDoHighResolution },
310
311	{ "RTLogDefaultInstance", (void *)RTLogDefaultInstance },
312	{ "RTMpCpuId", (void *)RTMpCpuId },
313	{ "RTMpCpuIdFromSetIndex", (void *)RTMpCpuIdFromSetIndex },
314	{ "RTMpCpuIdToSetIndex", (void *)RTMpCpuIdToSetIndex },
315	{ "RTMpGetArraySize", (void *)RTMpGetArraySize },
316	{ "RTMpIsCpuPossible", (void *)RTMpIsCpuPossible },
317	{ "RTMpGetCount", (void *)RTMpGetCount },
318	{ "RTMpGetMaxCpuId", (void *)RTMpGetMaxCpuId },
319	{ "RTMpGetOnlineCount", (void *)RTMpGetOnlineCount },
320	{ "RTMpGetOnlineSet", (void *)RTMpGetOnlineSet },
321	{ "RTMpGetSet", (void *)RTMpGetSet },
322	{ "RTMpIsCpuOnline", (void *)RTMpIsCpuOnline },
323	{ "RTMpIsCpuWorkPending", (void *)RTMpIsCpuWorkPending },
324	{ "RTMpOnAll", (void *)RTMpOnAll },
325	{ "RTMpOnOthers", (void *)RTMpOnOthers },
326	{ "RTMpOnSpecific", (void *)RTMpOnSpecific },
327	{ "RTMpPokeCpu", (void *)RTMpPokeCpu },
328	{ "RTPowerNotificationRegister", (void *)RTPowerNotificationRegister },
329	{ "RTPowerNotificationDeregister", (void *)RTPowerNotificationDeregister },
330	{ "RTLogRelDefaultInstance", (void *)RTLogRelDefaultInstance },
331	{ "RTLogSetDefaultInstanceThread", (void *)RTLogSetDefaultInstanceThread },
332	{ "RTLogLoggerExV", (void *)RTLogLoggerExV },
333	{ "RTLogPrintfV", (void *)RTLogPrintfV },
334	{ "RTR0AssertPanicSystem", (void *)RTR0AssertPanicSystem },
335	{ "RTAssertMsg1", (void *)RTAssertMsg1 },
336	{ "RTAssertMsg2V", (void *)RTAssertMsg2V },
337	{ "RTAssertSetQuiet", (void *)RTAssertSetQuiet },
338	{ "RTAssertMayPanic", (void *)RTAssertMayPanic },
339	{ "RTAssertSetMayPanic", (void *)RTAssertSetMayPanic },
340	{ "RTAssertAreQuiet", (void *)RTAssertAreQuiet },
341	{ "RTStrFormat", (void *)RTStrFormat },
342	{ "RTStrFormatNumber", (void *)RTStrFormatNumber },
343	{ "RTStrFormatTypeDeregister", (void *)RTStrFormatTypeDeregister },
344	{ "RTStrFormatTypeRegister", (void *)RTStrFormatTypeRegister },
345	{ "RTStrFormatTypeSetUser", (void *)RTStrFormatTypeSetUser },
346	{ "RTStrFormatV", (void *)RTStrFormatV },
347	{ "RTStrPrintf", (void *)RTStrPrintf },
348	{ "RTStrPrintfEx", (void *)RTStrPrintfEx },
349	{ "RTStrPrintfExV", (void *)RTStrPrintfExV },
350	{ "RTStrPrintfV", (void *)RTStrPrintfV },
351	{ "RTHandleTableAllocWithCtx", (void *)RTHandleTableAllocWithCtx },
352	{ "RTHandleTableCreate", (void *)RTHandleTableCreate },
353	{ "RTHandleTableCreateEx", (void *)RTHandleTableCreateEx },
354	{ "RTHandleTableDestroy", (void *)RTHandleTableDestroy },
355	{ "RTHandleTableFreeWithCtx", (void *)RTHandleTableFreeWithCtx },
356	{ "RTHandleTableLookupWithCtx", (void *)RTHandleTableLookupWithCtx },
357	{ "RTNetIPv4AddDataChecksum", (void *)RTNetIPv4AddDataChecksum },
358	{ "RTNetIPv4AddTCPChecksum", (void *)RTNetIPv4AddTCPChecksum },
359	{ "RTNetIPv4AddUDPChecksum", (void *)RTNetIPv4AddUDPChecksum },
360	{ "RTNetIPv4FinalizeChecksum", (void *)RTNetIPv4FinalizeChecksum },
361	{ "RTNetIPv4HdrChecksum", (void *)RTNetIPv4HdrChecksum },
362	{ "RTNetIPv4IsDHCPValid", (void *)RTNetIPv4IsDHCPValid },
363	{ "RTNetIPv4IsHdrValid", (void *)RTNetIPv4IsHdrValid },
364	{ "RTNetIPv4IsTCPSizeValid", (void *)RTNetIPv4IsTCPSizeValid },
365	{ "RTNetIPv4IsTCPValid", (void *)RTNetIPv4IsTCPValid },
366	{ "RTNetIPv4IsUDPSizeValid", (void *)RTNetIPv4IsUDPSizeValid },
367	{ "RTNetIPv4IsUDPValid", (void *)RTNetIPv4IsUDPValid },
368	{ "RTNetIPv4PseudoChecksum", (void *)RTNetIPv4PseudoChecksum },
369	{ "RTNetIPv4PseudoChecksumBits", (void *)RTNetIPv4PseudoChecksumBits },
370	{ "RTNetIPv4TCPChecksum", (void *)RTNetIPv4TCPChecksum },
371	{ "RTNetIPv4UDPChecksum", (void *)RTNetIPv4UDPChecksum },
372	{ "RTNetIPv6PseudoChecksum", (void *)RTNetIPv6PseudoChecksum },
373	{ "RTNetIPv6PseudoChecksumBits", (void *)RTNetIPv6PseudoChecksumBits },
374	{ "RTNetIPv6PseudoChecksumEx", (void *)RTNetIPv6PseudoChecksumEx },
375	{ "RTNetTCPChecksum", (void *)RTNetTCPChecksum },
376	{ "RTNetUDPChecksum", (void *)RTNetUDPChecksum },
377	{ "RTCrc32", (void *)RTCrc32 },
378	{ "RTCrc32Finish", (void *)RTCrc32Finish },
379	{ "RTCrc32Process", (void *)RTCrc32Process },
380	{ "RTCrc32Start", (void *)RTCrc32Start },
381	};
382
383	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
384	/**
385	* Drag in the rest of IRPT since we share it with the
386	* rest of the kernel modules on darwin.
387	*/
388	PFNRT g_apfnVBoxDrvIPRTDeps[] =
389	{
390	/* VBoxNetFlt */
391	(PFNRT)RTErrConvertFromErrno,
392	(PFNRT)RTUuidCompare,
393	(PFNRT)RTUuidCompareStr,
394	(PFNRT)RTUuidFromStr,
395	(PFNRT)RTStrDupTag,
396	(PFNRT)RTStrFree,
397	(PFNRT)RTStrCopy,
398	(PFNRT)RTStrNCmp,
399	/* VBoxNetAdp */
400	(PFNRT)RTRandBytes,
401	/* VBoxUSB */
402	(PFNRT)RTPathStripFilename,
403	NULL
404	};
405	#endif /* RT_OS_DARWIN \|\| RT_OS_SOLARIS \|\| RT_OS_SOLARIS */
406
407
408	/**
409	* Initializes the device extentsion structure.
410	*
411	* @returns IPRT status code.
412	* @param pDevExt The device extension to initialize.
413	* @param cbSession The size of the session structure. The size of
414	* SUPDRVSESSION may be smaller when SUPDRV_AGNOSTIC is
415	* defined because we're skipping the OS specific members
416	* then.
417	*/
418	int VBOXCALL supdrvInitDevExt(PSUPDRVDEVEXT pDevExt, size_t cbSession)
419	{
420	int rc;
421
422	#ifdef SUPDRV_WITH_RELEASE_LOGGER
423	/*
424	* Create the release log.
425	*/
426	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
427	PRTLOGGER pRelLogger;
428	rc = RTLogCreate(&pRelLogger, 0 /* fFlags */, "all",
429	"VBOX_RELEASE_LOG", RT_ELEMENTS(s_apszGroups), s_apszGroups, RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER, NULL);
430	if (RT_SUCCESS(rc))
431	RTLogRelSetDefaultInstance(pRelLogger);
432	/** @todo Add native hook for getting logger config parameters and setting
433	* them. On linux we should use the module parameter stuff... */
434	#endif
435
436	/*
437	* Initialize it.
438	*/
439	memset(pDevExt, 0, sizeof(*pDevExt));
440	rc = RTSpinlockCreate(&pDevExt->Spinlock);
441	if (RT_SUCCESS(rc))
442	{
443	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
444	rc = RTSemMutexCreate(&pDevExt->mtxLdr);
445	#else
446	rc = RTSemFastMutexCreate(&pDevExt->mtxLdr);
447	#endif
448	if (RT_SUCCESS(rc))
449	{
450	rc = RTSemFastMutexCreate(&pDevExt->mtxComponentFactory);
451	if (RT_SUCCESS(rc))
452	{
453	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
454	rc = RTSemMutexCreate(&pDevExt->mtxGip);
455	#else
456	rc = RTSemFastMutexCreate(&pDevExt->mtxGip);
457	#endif
458	if (RT_SUCCESS(rc))
459	{
460	rc = supdrvGipCreate(pDevExt);
461	if (RT_SUCCESS(rc))
462	{
463	pDevExt->u32Cookie = BIRD; /** @todo make this random? */
464	pDevExt->cbSession = (uint32_t)cbSession;
465
466	/*
467	* Fixup the absolute symbols.
468	*
469	* Because of the table indexing assumptions we'll have a little #ifdef orgy
470	* here rather than distributing this to OS specific files. At least for now.
471	*/
472	#ifdef RT_OS_DARWIN
473	# if ARCH_BITS == 32
474	if (SUPR0GetPagingMode() >= SUPPAGINGMODE_AMD64)
475	{
476	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
477	g_aFunctions[1].pfn = (void )0x80; / SUPR0Abs64bitKernelCS - KERNEL64_CS, seg.h */
478	g_aFunctions[2].pfn = (void )0x88; / SUPR0Abs64bitKernelSS - KERNEL64_SS, seg.h */
479	g_aFunctions[3].pfn = (void )0x88; / SUPR0Abs64bitKernelDS - KERNEL64_SS, seg.h */
480	}
481	else
482	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
483	g_aFunctions[4].pfn = (void )0x08; / SUPR0AbsKernelCS - KERNEL_CS, seg.h */
484	g_aFunctions[5].pfn = (void )0x10; / SUPR0AbsKernelSS - KERNEL_DS, seg.h */
485	g_aFunctions[6].pfn = (void )0x10; / SUPR0AbsKernelDS - KERNEL_DS, seg.h */
486	g_aFunctions[7].pfn = (void )0x10; / SUPR0AbsKernelES - KERNEL_DS, seg.h */
487	g_aFunctions[8].pfn = (void )0x10; / SUPR0AbsKernelFS - KERNEL_DS, seg.h */
488	g_aFunctions[9].pfn = (void )0x48; / SUPR0AbsKernelGS - CPU_DATA_GS, seg.h */
489	# else /* 64-bit darwin: */
490	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
491	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
492	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
493	g_aFunctions[3].pfn = (void )0; / SUPR0Abs64bitKernelDS */
494	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
495	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
496	g_aFunctions[6].pfn = (void )0; / SUPR0AbsKernelDS */
497	g_aFunctions[7].pfn = (void )0; / SUPR0AbsKernelES */
498	g_aFunctions[8].pfn = (void )0; / SUPR0AbsKernelFS */
499	g_aFunctions[9].pfn = (void )0; / SUPR0AbsKernelGS */
500
501	# endif
502	#else /* !RT_OS_DARWIN */
503	# if ARCH_BITS == 64
504	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
505	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
506	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
507	g_aFunctions[3].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0Abs64bitKernelDS */
508	# else
509	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
510	# endif
511	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
512	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
513	g_aFunctions[6].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0AbsKernelDS */
514	g_aFunctions[7].pfn = (void )(uintptr_t)ASMGetES(); / SUPR0AbsKernelES */
515	g_aFunctions[8].pfn = (void )(uintptr_t)ASMGetFS(); / SUPR0AbsKernelFS */
516	g_aFunctions[9].pfn = (void )(uintptr_t)ASMGetGS(); / SUPR0AbsKernelGS */
517	#endif /* !RT_OS_DARWIN */
518	return VINF_SUCCESS;
519	}
520
521	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
522	RTSemMutexDestroy(pDevExt->mtxGip);
523	pDevExt->mtxGip = NIL_RTSEMMUTEX;
524	#else
525	RTSemFastMutexDestroy(pDevExt->mtxGip);
526	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
527	#endif
528	}
529	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
530	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
531	}
532	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
533	RTSemMutexDestroy(pDevExt->mtxLdr);
534	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
535	#else
536	RTSemFastMutexDestroy(pDevExt->mtxLdr);
537	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
538	#endif
539	}
540	RTSpinlockDestroy(pDevExt->Spinlock);
541	pDevExt->Spinlock = NIL_RTSPINLOCK;
542	}
543	#ifdef SUPDRV_WITH_RELEASE_LOGGER
544	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
545	RTLogDestroy(RTLogSetDefaultInstance(NULL));
546	#endif
547
548	return rc;
549	}
550
551
552	/**
553	* Delete the device extension (e.g. cleanup members).
554	*
555	* @param pDevExt The device extension to delete.
556	*/
557	void VBOXCALL supdrvDeleteDevExt(PSUPDRVDEVEXT pDevExt)
558	{
559	PSUPDRVOBJ pObj;
560	PSUPDRVUSAGE pUsage;
561
562	/*
563	* Kill mutexes and spinlocks.
564	*/
565	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
566	RTSemMutexDestroy(pDevExt->mtxGip);
567	pDevExt->mtxGip = NIL_RTSEMMUTEX;
568	#else
569	RTSemFastMutexDestroy(pDevExt->mtxGip);
570	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
571	#endif
572	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
573	RTSemMutexDestroy(pDevExt->mtxLdr);
574	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
575	#else
576	RTSemFastMutexDestroy(pDevExt->mtxLdr);
577	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
578	#endif
579	RTSpinlockDestroy(pDevExt->Spinlock);
580	pDevExt->Spinlock = NIL_RTSPINLOCK;
581	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
582	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
583
584	/*
585	* Free lists.
586	*/
587	/* objects. */
588	pObj = pDevExt->pObjs;
589	Assert(!pObj); /* (can trigger on forced unloads) */
590	pDevExt->pObjs = NULL;
591	while (pObj)
592	{
593	void *pvFree = pObj;
594	pObj = pObj->pNext;
595	RTMemFree(pvFree);
596	}
597
598	/* usage records. */
599	pUsage = pDevExt->pUsageFree;
600	pDevExt->pUsageFree = NULL;
601	while (pUsage)
602	{
603	void *pvFree = pUsage;
604	pUsage = pUsage->pNext;
605	RTMemFree(pvFree);
606	}
607
608	/* kill the GIP. */
609	supdrvGipDestroy(pDevExt);
610
611	#ifdef SUPDRV_WITH_RELEASE_LOGGER
612	/* destroy the loggers. */
613	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
614	RTLogDestroy(RTLogSetDefaultInstance(NULL));
615	#endif
616	}
617
618
619	/**
620	* Create session.
621	*
622	* @returns IPRT status code.
623	* @param pDevExt Device extension.
624	* @param fUser Flag indicating whether this is a user or kernel session.
625	* @param ppSession Where to store the pointer to the session data.
626	*/
627	int VBOXCALL supdrvCreateSession(PSUPDRVDEVEXT pDevExt, bool fUser, PSUPDRVSESSION *ppSession)
628	{
629	/*
630	* Allocate memory for the session data.
631	*/
632	int rc;
633	PSUPDRVSESSION pSession = *ppSession = (PSUPDRVSESSION)RTMemAllocZ(pDevExt->cbSession);
634	if (pSession)
635	{
636	/* Initialize session data. */
637	rc = RTSpinlockCreate(&pSession->Spinlock);
638	if (!rc)
639	{
640	rc = RTHandleTableCreateEx(&pSession->hHandleTable,
641	RTHANDLETABLE_FLAGS_LOCKED \| RTHANDLETABLE_FLAGS_CONTEXT,
642	1 /uBase/, 32768 /cMax/, supdrvSessionObjHandleRetain, pSession);
643	if (RT_SUCCESS(rc))
644	{
645	Assert(pSession->Spinlock != NIL_RTSPINLOCK);
646	pSession->pDevExt = pDevExt;
647	pSession->u32Cookie = BIRD_INV;
648	/*pSession->pLdrUsage = NULL;
649	pSession->pVM = NULL;
650	pSession->pUsage = NULL;
651	pSession->pGip = NULL;
652	pSession->fGipReferenced = false;
653	pSession->Bundle.cUsed = 0; */
654	pSession->Uid = NIL_RTUID;
655	pSession->Gid = NIL_RTGID;
656	if (fUser)
657	{
658	pSession->Process = RTProcSelf();
659	pSession->R0Process = RTR0ProcHandleSelf();
660	}
661	else
662	{
663	pSession->Process = NIL_RTPROCESS;
664	pSession->R0Process = NIL_RTR0PROCESS;
665	}
666
667	VBOXDRV_SUPDRV_SESSION_CREATE(pSession, fUser);
668	LogFlow(("Created session %p initial cookie=%#x\n", pSession, pSession->u32Cookie));
669	return VINF_SUCCESS;
670	}
671
672	RTSpinlockDestroy(pSession->Spinlock);
673	}
674	RTMemFree(pSession);
675	*ppSession = NULL;
676	Log(("Failed to create spinlock, rc=%d!\n", rc));
677	}
678	else
679	rc = VERR_NO_MEMORY;
680
681	return rc;
682	}
683
684
685	/**
686	* Shared code for cleaning up a session.
687	*
688	* @param pDevExt Device extension.
689	* @param pSession Session data.
690	* This data will be freed by this routine.
691	*/
692	void VBOXCALL supdrvCloseSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
693	{
694	VBOXDRV_SUPDRV_SESSION_CLOSE(pSession);
695
696	/*
697	* Cleanup the session first.
698	*/
699	supdrvCleanupSession(pDevExt, pSession);
700
701	/*
702	* Free the rest of the session stuff.
703	*/
704	RTSpinlockDestroy(pSession->Spinlock);
705	pSession->Spinlock = NIL_RTSPINLOCK;
706	pSession->pDevExt = NULL;
707	RTMemFree(pSession);
708	LogFlow(("supdrvCloseSession: returns\n"));
709	}
710
711
712	/**
713	* Shared code for cleaning up a session (but not quite freeing it).
714	*
715	* This is primarily intended for MAC OS X where we have to clean up the memory
716	* stuff before the file handle is closed.
717	*
718	* @param pDevExt Device extension.
719	* @param pSession Session data.
720	* This data will be freed by this routine.
721	*/
722	void VBOXCALL supdrvCleanupSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
723	{
724	int rc;
725	PSUPDRVBUNDLE pBundle;
726	LogFlow(("supdrvCleanupSession: pSession=%p\n", pSession));
727
728	/*
729	* Remove logger instances related to this session.
730	*/
731	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pSession);
732
733	/*
734	* Destroy the handle table.
735	*/
736	rc = RTHandleTableDestroy(pSession->hHandleTable, supdrvSessionObjHandleDelete, pSession);
737	AssertRC(rc);
738	pSession->hHandleTable = NIL_RTHANDLETABLE;
739
740	/*
741	* Release object references made in this session.
742	* In theory there should be noone racing us in this session.
743	*/
744	Log2(("release objects - start\n"));
745	if (pSession->pUsage)
746	{
747	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
748	PSUPDRVUSAGE pUsage;
749	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
750
751	while ((pUsage = pSession->pUsage) != NULL)
752	{
753	PSUPDRVOBJ pObj = pUsage->pObj;
754	pSession->pUsage = pUsage->pNext;
755
756	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
757	if (pUsage->cUsage < pObj->cUsage)
758	{
759	pObj->cUsage -= pUsage->cUsage;
760	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
761	}
762	else
763	{
764	/* Destroy the object and free the record. */
765	if (pDevExt->pObjs == pObj)
766	pDevExt->pObjs = pObj->pNext;
767	else
768	{
769	PSUPDRVOBJ pObjPrev;
770	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
771	if (pObjPrev->pNext == pObj)
772	{
773	pObjPrev->pNext = pObj->pNext;
774	break;
775	}
776	Assert(pObjPrev);
777	}
778	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
779
780	Log(("supdrvCleanupSession: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
781	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
782	if (pObj->pfnDestructor)
783	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
784	RTMemFree(pObj);
785	}
786
787	/* free it and continue. */
788	RTMemFree(pUsage);
789
790	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
791	}
792
793	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
794	AssertMsg(!pSession->pUsage, ("Some buster reregistered an object during desturction!\n"));
795	}
796	Log2(("release objects - done\n"));
797
798	/*
799	* Release memory allocated in the session.
800	*
801	* We do not serialize this as we assume that the application will
802	* not allocated memory while closing the file handle object.
803	*/
804	Log2(("freeing memory:\n"));
805	pBundle = &pSession->Bundle;
806	while (pBundle)
807	{
808	PSUPDRVBUNDLE pToFree;
809	unsigned i;
810
811	/*
812	* Check and unlock all entries in the bundle.
813	*/
814	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
815	{
816	if (pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ)
817	{
818	Log2(("eType=%d pvR0=%p pvR3=%p cb=%ld\n", pBundle->aMem[i].eType, RTR0MemObjAddress(pBundle->aMem[i].MemObj),
819	(void *)RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3), (long)RTR0MemObjSize(pBundle->aMem[i].MemObj)));
820	if (pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ)
821	{
822	rc = RTR0MemObjFree(pBundle->aMem[i].MapObjR3, false);
823	AssertRC(rc); /** @todo figure out how to handle this. */
824	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
825	}
826	rc = RTR0MemObjFree(pBundle->aMem[i].MemObj, true /* fFreeMappings */);
827	AssertRC(rc); /** @todo figure out how to handle this. */
828	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
829	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
830	}
831	}
832
833	/*
834	* Advance and free previous bundle.
835	*/
836	pToFree = pBundle;
837	pBundle = pBundle->pNext;
838
839	pToFree->pNext = NULL;
840	pToFree->cUsed = 0;
841	if (pToFree != &pSession->Bundle)
842	RTMemFree(pToFree);
843	}
844	Log2(("freeing memory - done\n"));
845
846	/*
847	* Deregister component factories.
848	*/
849	RTSemFastMutexRequest(pDevExt->mtxComponentFactory);
850	Log2(("deregistering component factories:\n"));
851	if (pDevExt->pComponentFactoryHead)
852	{
853	PSUPDRVFACTORYREG pPrev = NULL;
854	PSUPDRVFACTORYREG pCur = pDevExt->pComponentFactoryHead;
855	while (pCur)
856	{
857	if (pCur->pSession == pSession)
858	{
859	/* unlink it */
860	PSUPDRVFACTORYREG pNext = pCur->pNext;
861	if (pPrev)
862	pPrev->pNext = pNext;
863	else
864	pDevExt->pComponentFactoryHead = pNext;
865
866	/* free it */
867	pCur->pNext = NULL;
868	pCur->pSession = NULL;
869	pCur->pFactory = NULL;
870	RTMemFree(pCur);
871
872	/* next */
873	pCur = pNext;
874	}
875	else
876	{
877	/* next */
878	pPrev = pCur;
879	pCur = pCur->pNext;
880	}
881	}
882	}
883	RTSemFastMutexRelease(pDevExt->mtxComponentFactory);
884	Log2(("deregistering component factories - done\n"));
885
886	/*
887	* Loaded images needs to be dereferenced and possibly freed up.
888	*/
889	supdrvLdrLock(pDevExt);
890	Log2(("freeing images:\n"));
891	if (pSession->pLdrUsage)
892	{
893	PSUPDRVLDRUSAGE pUsage = pSession->pLdrUsage;
894	pSession->pLdrUsage = NULL;
895	while (pUsage)
896	{
897	void *pvFree = pUsage;
898	PSUPDRVLDRIMAGE pImage = pUsage->pImage;
899	if (pImage->cUsage > pUsage->cUsage)
900	pImage->cUsage -= pUsage->cUsage;
901	else
902	supdrvLdrFree(pDevExt, pImage);
903	pUsage->pImage = NULL;
904	pUsage = pUsage->pNext;
905	RTMemFree(pvFree);
906	}
907	}
908	supdrvLdrUnlock(pDevExt);
909	Log2(("freeing images - done\n"));
910
911	/*
912	* Unmap the GIP.
913	*/
914	Log2(("umapping GIP:\n"));
915	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
916	{
917	SUPR0GipUnmap(pSession);
918	pSession->fGipReferenced = 0;
919	}
920	Log2(("umapping GIP - done\n"));
921	}
922
923
924	/**
925	* RTHandleTableDestroy callback used by supdrvCleanupSession.
926	*
927	* @returns IPRT status code, see SUPR0ObjAddRef.
928	* @param hHandleTable The handle table handle. Ignored.
929	* @param pvObj The object pointer.
930	* @param pvCtx Context, the handle type. Ignored.
931	* @param pvUser Session pointer.
932	*/
933	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser)
934	{
935	NOREF(pvCtx);
936	NOREF(hHandleTable);
937	return SUPR0ObjAddRefEx(pvObj, (PSUPDRVSESSION)pvUser, true /fNoBlocking/);
938	}
939
940
941	/**
942	* RTHandleTableDestroy callback used by supdrvCleanupSession.
943	*
944	* @param hHandleTable The handle table handle. Ignored.
945	* @param h The handle value. Ignored.
946	* @param pvObj The object pointer.
947	* @param pvCtx Context, the handle type. Ignored.
948	* @param pvUser Session pointer.
949	*/
950	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser)
951	{
952	NOREF(pvCtx);
953	NOREF(h);
954	NOREF(hHandleTable);
955	SUPR0ObjRelease(pvObj, (PSUPDRVSESSION)pvUser);
956	}
957
958
959	/**
960	* Fast path I/O Control worker.
961	*
962	* @returns VBox status code that should be passed down to ring-3 unchanged.
963	* @param uIOCtl Function number.
964	* @param idCpu VMCPU id.
965	* @param pDevExt Device extention.
966	* @param pSession Session data.
967	*/
968	int VBOXCALL supdrvIOCtlFast(uintptr_t uIOCtl, VMCPUID idCpu, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
969	{
970	/*
971	* We check the two prereqs after doing this only to allow the compiler to optimize things better.
972	*/
973	if (RT_LIKELY(pSession->pVM && pDevExt->pfnVMMR0EntryFast))
974	{
975	switch (uIOCtl)
976	{
977	case SUP_IOCTL_FAST_DO_RAW_RUN:
978	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_RAW_RUN);
979	break;
980	case SUP_IOCTL_FAST_DO_HWACC_RUN:
981	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_HWACC_RUN);
982	break;
983	case SUP_IOCTL_FAST_DO_NOP:
984	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_NOP);
985	break;
986	default:
987	return VERR_INTERNAL_ERROR;
988	}
989	return VINF_SUCCESS;
990	}
991	return VERR_INTERNAL_ERROR;
992	}
993
994
995	/**
996	* Helper for supdrvIOCtl. Check if pszStr contains any character of pszChars.
997	* We would use strpbrk here if this function would be contained in the RedHat kABI white
998	* list, see http://www.kerneldrivers.org/RHEL5.
999	*
1000	* @returns 1 if pszStr does contain any character of pszChars, 0 otherwise.
1001	* @param pszStr String to check
1002	* @param pszChars Character set
1003	*/
1004	static int supdrvCheckInvalidChar(const char pszStr, const char pszChars)
1005	{
1006	int chCur;
1007	while ((chCur = *pszStr++) != '\0')
1008	{
1009	int ch;
1010	const char *psz = pszChars;
1011	while ((ch = *psz++) != '\0')
1012	if (ch == chCur)
1013	return 1;
1014
1015	}
1016	return 0;
1017	}
1018
1019
1020	/**
1021	* I/O Control worker.
1022	*
1023	* @returns IPRT status code.
1024	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1025	*
1026	* @param uIOCtl Function number.
1027	* @param pDevExt Device extention.
1028	* @param pSession Session data.
1029	* @param pReqHdr The request header.
1030	*/
1031	int VBOXCALL supdrvIOCtl(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1032	{
1033	VBOXDRV_SUPDRV_IOCTL_ENTRY(pSession, uIOCtl, pReqHdr);
1034
1035	/*
1036	* Validate the request.
1037	*/
1038	/* this first check could probably be omitted as its also done by the OS specific code... */
1039	if (RT_UNLIKELY( (pReqHdr->fFlags & SUPREQHDR_FLAGS_MAGIC_MASK) != SUPREQHDR_FLAGS_MAGIC
1040	\|\| pReqHdr->cbIn < sizeof(*pReqHdr)
1041	\|\| pReqHdr->cbOut < sizeof(*pReqHdr)))
1042	{
1043	OSDBGPRINT(("vboxdrv: Bad ioctl request header; cbIn=%#lx cbOut=%#lx fFlags=%#lx\n",
1044	(long)pReqHdr->cbIn, (long)pReqHdr->cbOut, (long)pReqHdr->fFlags));
1045	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1046	return VERR_INVALID_PARAMETER;
1047	}
1048	if (RT_UNLIKELY(uIOCtl == SUP_IOCTL_COOKIE))
1049	{
1050	if (pReqHdr->u32Cookie != SUPCOOKIE_INITIAL_COOKIE)
1051	{
1052	OSDBGPRINT(("SUP_IOCTL_COOKIE: bad cookie %#lx\n", (long)pReqHdr->u32Cookie));
1053	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1054	return VERR_INVALID_PARAMETER;
1055	}
1056	}
1057	else if (RT_UNLIKELY( pReqHdr->u32Cookie != pDevExt->u32Cookie
1058	\|\| pReqHdr->u32SessionCookie != pSession->u32Cookie))
1059	{
1060	OSDBGPRINT(("vboxdrv: bad cookie %#lx / %#lx.\n", (long)pReqHdr->u32Cookie, (long)pReqHdr->u32SessionCookie));
1061	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1062	return VERR_INVALID_PARAMETER;
1063	}
1064
1065	/*
1066	* Validation macros
1067	*/
1068	#define REQ_CHECK_SIZES_EX(Name, cbInExpect, cbOutExpect) \
1069	do { \
1070	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect) \|\| pReqHdr->cbOut != (cbOutExpect))) \
1071	{ \
1072	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld. cbOut=%ld expected %ld.\n", \
1073	(long)pReq->Hdr.cbIn, (long)(cbInExpect), (long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
1074	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1075	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1076	} \
1077	} while (0)
1078
1079	#define REQ_CHECK_SIZES(Name) REQ_CHECK_SIZES_EX(Name, Name ## _SIZE_IN, Name ## _SIZE_OUT)
1080
1081	#define REQ_CHECK_SIZE_IN(Name, cbInExpect) \
1082	do { \
1083	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect))) \
1084	{ \
1085	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld.\n", \
1086	(long)pReq->Hdr.cbIn, (long)(cbInExpect))); \
1087	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1088	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1089	} \
1090	} while (0)
1091
1092	#define REQ_CHECK_SIZE_OUT(Name, cbOutExpect) \
1093	do { \
1094	if (RT_UNLIKELY(pReqHdr->cbOut != (cbOutExpect))) \
1095	{ \
1096	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbOut=%ld expected %ld.\n", \
1097	(long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
1098	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1099	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1100	} \
1101	} while (0)
1102
1103	#define REQ_CHECK_EXPR(Name, expr) \
1104	do { \
1105	if (RT_UNLIKELY(!(expr))) \
1106	{ \
1107	OSDBGPRINT(( #Name ": %s\n", #expr)); \
1108	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1109	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1110	} \
1111	} while (0)
1112
1113	#define REQ_CHECK_EXPR_FMT(expr, fmt) \
1114	do { \
1115	if (RT_UNLIKELY(!(expr))) \
1116	{ \
1117	OSDBGPRINT( fmt ); \
1118	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1119	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1120	} \
1121	} while (0)
1122
1123
1124	/*
1125	* The switch.
1126	*/
1127	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
1128	{
1129	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
1130	{
1131	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
1132	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
1133	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
1134	{
1135	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
1136	pReq->Hdr.rc = VERR_INVALID_MAGIC;
1137	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, VERR_INVALID_MAGIC);
1138	return 0;
1139	}
1140
1141	#if 0
1142	/*
1143	* Call out to the OS specific code and let it do permission checks on the
1144	* client process.
1145	*/
1146	if (!supdrvOSValidateClientProcess(pDevExt, pSession))
1147	{
1148	pReq->u.Out.u32Cookie = 0xffffffff;
1149	pReq->u.Out.u32SessionCookie = 0xffffffff;
1150	pReq->u.Out.u32SessionVersion = 0xffffffff;
1151	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1152	pReq->u.Out.pSession = NULL;
1153	pReq->u.Out.cFunctions = 0;
1154	pReq->Hdr.rc = VERR_PERMISSION_DENIED;
1155	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, VERR_PERMISSION_DENIED);
1156	return 0;
1157	}
1158	#endif
1159
1160	/*
1161	* Match the version.
1162	* The current logic is very simple, match the major interface version.
1163	*/
1164	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
1165	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
1166	{
1167	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1168	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
1169	pReq->u.Out.u32Cookie = 0xffffffff;
1170	pReq->u.Out.u32SessionCookie = 0xffffffff;
1171	pReq->u.Out.u32SessionVersion = 0xffffffff;
1172	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1173	pReq->u.Out.pSession = NULL;
1174	pReq->u.Out.cFunctions = 0;
1175	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1176	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1177	return 0;
1178	}
1179
1180	/*
1181	* Fill in return data and be gone.
1182	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
1183	* u32SessionVersion <= u32ReqVersion!
1184	*/
1185	/** @todo Somehow validate the client and negotiate a secure cookie... */
1186	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
1187	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
1188	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
1189	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1190	pReq->u.Out.pSession = pSession;
1191	pReq->u.Out.cFunctions = sizeof(g_aFunctions) / sizeof(g_aFunctions[0]);
1192	pReq->Hdr.rc = VINF_SUCCESS;
1193	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1194	return 0;
1195	}
1196
1197	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_QUERY_FUNCS(0)):
1198	{
1199	/* validate */
1200	PSUPQUERYFUNCS pReq = (PSUPQUERYFUNCS)pReqHdr;
1201	REQ_CHECK_SIZES_EX(SUP_IOCTL_QUERY_FUNCS, SUP_IOCTL_QUERY_FUNCS_SIZE_IN, SUP_IOCTL_QUERY_FUNCS_SIZE_OUT(RT_ELEMENTS(g_aFunctions)));
1202
1203	/* execute */
1204	pReq->u.Out.cFunctions = RT_ELEMENTS(g_aFunctions);
1205	memcpy(&pReq->u.Out.aFunctions[0], g_aFunctions, sizeof(g_aFunctions));
1206	pReq->Hdr.rc = VINF_SUCCESS;
1207	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1208	return 0;
1209	}
1210
1211	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_LOCK):
1212	{
1213	/* validate */
1214	PSUPPAGELOCK pReq = (PSUPPAGELOCK)pReqHdr;
1215	REQ_CHECK_SIZE_IN(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_IN);
1216	REQ_CHECK_SIZE_OUT(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_OUT(pReq->u.In.cPages));
1217	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.cPages > 0);
1218	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.pvR3 >= PAGE_SIZE);
1219
1220	/* execute */
1221	pReq->Hdr.rc = SUPR0LockMem(pSession, pReq->u.In.pvR3, pReq->u.In.cPages, &pReq->u.Out.aPages[0]);
1222	if (RT_FAILURE(pReq->Hdr.rc))
1223	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1224	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1225	return 0;
1226	}
1227
1228	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_UNLOCK):
1229	{
1230	/* validate */
1231	PSUPPAGEUNLOCK pReq = (PSUPPAGEUNLOCK)pReqHdr;
1232	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_UNLOCK);
1233
1234	/* execute */
1235	pReq->Hdr.rc = SUPR0UnlockMem(pSession, pReq->u.In.pvR3);
1236	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1237	return 0;
1238	}
1239
1240	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_ALLOC):
1241	{
1242	/* validate */
1243	PSUPCONTALLOC pReq = (PSUPCONTALLOC)pReqHdr;
1244	REQ_CHECK_SIZES(SUP_IOCTL_CONT_ALLOC);
1245
1246	/* execute */
1247	pReq->Hdr.rc = SUPR0ContAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.HCPhys);
1248	if (RT_FAILURE(pReq->Hdr.rc))
1249	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1250	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1251	return 0;
1252	}
1253
1254	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_FREE):
1255	{
1256	/* validate */
1257	PSUPCONTFREE pReq = (PSUPCONTFREE)pReqHdr;
1258	REQ_CHECK_SIZES(SUP_IOCTL_CONT_FREE);
1259
1260	/* execute */
1261	pReq->Hdr.rc = SUPR0ContFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1262	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1263	return 0;
1264	}
1265
1266	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_OPEN):
1267	{
1268	/* validate */
1269	PSUPLDROPEN pReq = (PSUPLDROPEN)pReqHdr;
1270	REQ_CHECK_SIZES(SUP_IOCTL_LDR_OPEN);
1271	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs > 0);
1272	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs < 16*_1M);
1273	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1274	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1275	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits < pReq->u.In.cbImageWithTabs);
1276	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.szName[0]);
1277	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1278	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, !supdrvCheckInvalidChar(pReq->u.In.szName, ";:()[]{}/\\\|&*%#@!~`\"'"));
1279	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szFilename, sizeof(pReq->u.In.szFilename)));
1280
1281	/* execute */
1282	pReq->Hdr.rc = supdrvIOCtl_LdrOpen(pDevExt, pSession, pReq);
1283	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1284	return 0;
1285	}
1286
1287	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOAD):
1288	{
1289	/* validate */
1290	PSUPLDRLOAD pReq = (PSUPLDRLOAD)pReqHdr;
1291	REQ_CHECK_EXPR(Name, pReq->Hdr.cbIn >= sizeof(*pReq));
1292	REQ_CHECK_SIZES_EX(SUP_IOCTL_LDR_LOAD, SUP_IOCTL_LDR_LOAD_SIZE_IN(pReq->u.In.cbImageWithTabs), SUP_IOCTL_LDR_LOAD_SIZE_OUT);
1293	REQ_CHECK_EXPR(SUP_IOCTL_LDR_LOAD, pReq->u.In.cSymbols <= 16384);
1294	REQ_CHECK_EXPR_FMT( !pReq->u.In.cSymbols
1295	\|\| ( pReq->u.In.offSymbols < pReq->u.In.cbImageWithTabs
1296	&& pReq->u.In.offSymbols + pReq->u.In.cSymbols * sizeof(SUPLDRSYM) <= pReq->u.In.cbImageWithTabs),
1297	("SUP_IOCTL_LDR_LOAD: offSymbols=%#lx cSymbols=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offSymbols,
1298	(long)pReq->u.In.cSymbols, (long)pReq->u.In.cbImageWithTabs));
1299	REQ_CHECK_EXPR_FMT( !pReq->u.In.cbStrTab
1300	\|\| ( pReq->u.In.offStrTab < pReq->u.In.cbImageWithTabs
1301	&& pReq->u.In.offStrTab + pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs
1302	&& pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs),
1303	("SUP_IOCTL_LDR_LOAD: offStrTab=%#lx cbStrTab=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offStrTab,
1304	(long)pReq->u.In.cbStrTab, (long)pReq->u.In.cbImageWithTabs));
1305
1306	if (pReq->u.In.cSymbols)
1307	{
1308	uint32_t i;
1309	PSUPLDRSYM paSyms = (PSUPLDRSYM)&pReq->u.In.abImage[pReq->u.In.offSymbols];
1310	for (i = 0; i < pReq->u.In.cSymbols; i++)
1311	{
1312	REQ_CHECK_EXPR_FMT(paSyms[i].offSymbol < pReq->u.In.cbImageWithTabs,
1313	("SUP_IOCTL_LDR_LOAD: sym #%ld: symb off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offSymbol, (long)pReq->u.In.cbImageWithTabs));
1314	REQ_CHECK_EXPR_FMT(paSyms[i].offName < pReq->u.In.cbStrTab,
1315	("SUP_IOCTL_LDR_LOAD: sym #%ld: name off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1316	REQ_CHECK_EXPR_FMT(RTStrEnd((char const *)&pReq->u.In.abImage[pReq->u.In.offStrTab + paSyms[i].offName],
1317	pReq->u.In.cbStrTab - paSyms[i].offName),
1318	("SUP_IOCTL_LDR_LOAD: sym #%ld: unterminated name! (%#lx / %#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1319	}
1320	}
1321
1322	/* execute */
1323	pReq->Hdr.rc = supdrvIOCtl_LdrLoad(pDevExt, pSession, pReq);
1324	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1325	return 0;
1326	}
1327
1328	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_FREE):
1329	{
1330	/* validate */
1331	PSUPLDRFREE pReq = (PSUPLDRFREE)pReqHdr;
1332	REQ_CHECK_SIZES(SUP_IOCTL_LDR_FREE);
1333
1334	/* execute */
1335	pReq->Hdr.rc = supdrvIOCtl_LdrFree(pDevExt, pSession, pReq);
1336	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1337	return 0;
1338	}
1339
1340	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_GET_SYMBOL):
1341	{
1342	/* validate */
1343	PSUPLDRGETSYMBOL pReq = (PSUPLDRGETSYMBOL)pReqHdr;
1344	REQ_CHECK_SIZES(SUP_IOCTL_LDR_GET_SYMBOL);
1345	REQ_CHECK_EXPR(SUP_IOCTL_LDR_GET_SYMBOL, RTStrEnd(pReq->u.In.szSymbol, sizeof(pReq->u.In.szSymbol)));
1346
1347	/* execute */
1348	pReq->Hdr.rc = supdrvIOCtl_LdrGetSymbol(pDevExt, pSession, pReq);
1349	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1350	return 0;
1351	}
1352
1353	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0(0)):
1354	{
1355	/* validate */
1356	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1357	Log4(("SUP_IOCTL_CALL_VMMR0: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1358	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1359
1360	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_VMMR0_SIZE(0))
1361	{
1362	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(0), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(0));
1363
1364	/* execute */
1365	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1366	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg, pSession);
1367	else
1368	pReq->Hdr.rc = VERR_WRONG_ORDER;
1369	}
1370	else
1371	{
1372	PSUPVMMR0REQHDR pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1373	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR)),
1374	("SUP_IOCTL_CALL_VMMR0: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR))));
1375	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1376	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(pVMMReq->cbReq));
1377
1378	/* execute */
1379	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1380	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1381	else
1382	pReq->Hdr.rc = VERR_WRONG_ORDER;
1383	}
1384
1385	if ( RT_FAILURE(pReq->Hdr.rc)
1386	&& pReq->Hdr.rc != VERR_INTERRUPTED
1387	&& pReq->Hdr.rc != VERR_TIMEOUT)
1388	Log(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1389	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1390	else
1391	Log4(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1392	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1393	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1394	return 0;
1395	}
1396
1397	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0_BIG):
1398	{
1399	/* validate */
1400	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1401	PSUPVMMR0REQHDR pVMMReq;
1402	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1403	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1404
1405	pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1406	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR)),
1407	("SUP_IOCTL_CALL_VMMR0_BIG: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR))));
1408	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0_BIG, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1409	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0_BIG, SUP_IOCTL_CALL_VMMR0_BIG_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_BIG_SIZE_OUT(pVMMReq->cbReq));
1410
1411	/* execute */
1412	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1413	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1414	else
1415	pReq->Hdr.rc = VERR_WRONG_ORDER;
1416
1417	if ( RT_FAILURE(pReq->Hdr.rc)
1418	&& pReq->Hdr.rc != VERR_INTERRUPTED
1419	&& pReq->Hdr.rc != VERR_TIMEOUT)
1420	Log(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1421	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1422	else
1423	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1424	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1425	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1426	return 0;
1427	}
1428
1429	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_PAGING_MODE):
1430	{
1431	/* validate */
1432	PSUPGETPAGINGMODE pReq = (PSUPGETPAGINGMODE)pReqHdr;
1433	REQ_CHECK_SIZES(SUP_IOCTL_GET_PAGING_MODE);
1434
1435	/* execute */
1436	pReq->Hdr.rc = VINF_SUCCESS;
1437	pReq->u.Out.enmMode = SUPR0GetPagingMode();
1438	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1439	return 0;
1440	}
1441
1442	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_ALLOC):
1443	{
1444	/* validate */
1445	PSUPLOWALLOC pReq = (PSUPLOWALLOC)pReqHdr;
1446	REQ_CHECK_EXPR(SUP_IOCTL_LOW_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_LOW_ALLOC_SIZE_IN);
1447	REQ_CHECK_SIZES_EX(SUP_IOCTL_LOW_ALLOC, SUP_IOCTL_LOW_ALLOC_SIZE_IN, SUP_IOCTL_LOW_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1448
1449	/* execute */
1450	pReq->Hdr.rc = SUPR0LowAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1451	if (RT_FAILURE(pReq->Hdr.rc))
1452	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1453	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1454	return 0;
1455	}
1456
1457	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_FREE):
1458	{
1459	/* validate */
1460	PSUPLOWFREE pReq = (PSUPLOWFREE)pReqHdr;
1461	REQ_CHECK_SIZES(SUP_IOCTL_LOW_FREE);
1462
1463	/* execute */
1464	pReq->Hdr.rc = SUPR0LowFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1465	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1466	return 0;
1467	}
1468
1469	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_MAP):
1470	{
1471	/* validate */
1472	PSUPGIPMAP pReq = (PSUPGIPMAP)pReqHdr;
1473	REQ_CHECK_SIZES(SUP_IOCTL_GIP_MAP);
1474
1475	/* execute */
1476	pReq->Hdr.rc = SUPR0GipMap(pSession, &pReq->u.Out.pGipR3, &pReq->u.Out.HCPhysGip);
1477	if (RT_SUCCESS(pReq->Hdr.rc))
1478	pReq->u.Out.pGipR0 = pDevExt->pGip;
1479	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1480	return 0;
1481	}
1482
1483	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_UNMAP):
1484	{
1485	/* validate */
1486	PSUPGIPUNMAP pReq = (PSUPGIPUNMAP)pReqHdr;
1487	REQ_CHECK_SIZES(SUP_IOCTL_GIP_UNMAP);
1488
1489	/* execute */
1490	pReq->Hdr.rc = SUPR0GipUnmap(pSession);
1491	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1492	return 0;
1493	}
1494
1495	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SET_VM_FOR_FAST):
1496	{
1497	/* validate */
1498	PSUPSETVMFORFAST pReq = (PSUPSETVMFORFAST)pReqHdr;
1499	REQ_CHECK_SIZES(SUP_IOCTL_SET_VM_FOR_FAST);
1500	REQ_CHECK_EXPR_FMT( !pReq->u.In.pVMR0
1501	\|\| ( VALID_PTR(pReq->u.In.pVMR0)
1502	&& !((uintptr_t)pReq->u.In.pVMR0 & (PAGE_SIZE - 1))),
1503	("SUP_IOCTL_SET_VM_FOR_FAST: pVMR0=%p!\n", pReq->u.In.pVMR0));
1504	/* execute */
1505	pSession->pVM = pReq->u.In.pVMR0;
1506	pReq->Hdr.rc = VINF_SUCCESS;
1507	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1508	return 0;
1509	}
1510
1511	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_ALLOC_EX):
1512	{
1513	/* validate */
1514	PSUPPAGEALLOCEX pReq = (PSUPPAGEALLOCEX)pReqHdr;
1515	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC_EX, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_EX_SIZE_IN);
1516	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));
1517	REQ_CHECK_EXPR_FMT(pReq->u.In.fKernelMapping \|\| pReq->u.In.fUserMapping,
1518	("SUP_IOCTL_PAGE_ALLOC_EX: No mapping requested!\n"));
1519	REQ_CHECK_EXPR_FMT(pReq->u.In.fUserMapping,
1520	("SUP_IOCTL_PAGE_ALLOC_EX: Must have user mapping!\n"));
1521	REQ_CHECK_EXPR_FMT(!pReq->u.In.fReserved0 && !pReq->u.In.fReserved1,
1522	("SUP_IOCTL_PAGE_ALLOC_EX: fReserved0=%d fReserved1=%d\n", pReq->u.In.fReserved0, pReq->u.In.fReserved1));
1523
1524	/* execute */
1525	pReq->Hdr.rc = SUPR0PageAllocEx(pSession, pReq->u.In.cPages, 0 /* fFlags */,
1526	pReq->u.In.fUserMapping ? &pReq->u.Out.pvR3 : NULL,
1527	pReq->u.In.fKernelMapping ? &pReq->u.Out.pvR0 : NULL,
1528	&pReq->u.Out.aPages[0]);
1529	if (RT_FAILURE(pReq->Hdr.rc))
1530	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1531	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1532	return 0;
1533	}
1534
1535	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_MAP_KERNEL):
1536	{
1537	/* validate */
1538	PSUPPAGEMAPKERNEL pReq = (PSUPPAGEMAPKERNEL)pReqHdr;
1539	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_MAP_KERNEL);
1540	REQ_CHECK_EXPR_FMT(!pReq->u.In.fFlags, ("SUP_IOCTL_PAGE_MAP_KERNEL: fFlags=%#x! MBZ\n", pReq->u.In.fFlags));
1541	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_MAP_KERNEL: offSub=%#x\n", pReq->u.In.offSub));
1542	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1543	("SUP_IOCTL_PAGE_MAP_KERNEL: cbSub=%#x\n", pReq->u.In.cbSub));
1544
1545	/* execute */
1546	pReq->Hdr.rc = SUPR0PageMapKernel(pSession, pReq->u.In.pvR3, pReq->u.In.offSub, pReq->u.In.cbSub,
1547	pReq->u.In.fFlags, &pReq->u.Out.pvR0);
1548	if (RT_FAILURE(pReq->Hdr.rc))
1549	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1550	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1551	return 0;
1552	}
1553
1554	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_PROTECT):
1555	{
1556	/* validate */
1557	PSUPPAGEPROTECT pReq = (PSUPPAGEPROTECT)pReqHdr;
1558	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_PROTECT);
1559	REQ_CHECK_EXPR_FMT(!(pReq->u.In.fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)),
1560	("SUP_IOCTL_PAGE_PROTECT: fProt=%#x!\n", pReq->u.In.fProt));
1561	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_PROTECT: offSub=%#x\n", pReq->u.In.offSub));
1562	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1563	("SUP_IOCTL_PAGE_PROTECT: cbSub=%#x\n", pReq->u.In.cbSub));
1564
1565	/* execute */
1566	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);
1567	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1568	return 0;
1569	}
1570
1571	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_FREE):
1572	{
1573	/* validate */
1574	PSUPPAGEFREE pReq = (PSUPPAGEFREE)pReqHdr;
1575	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_FREE);
1576
1577	/* execute */
1578	pReq->Hdr.rc = SUPR0PageFree(pSession, pReq->u.In.pvR3);
1579	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1580	return 0;
1581	}
1582
1583	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_SERVICE(0)):
1584	{
1585	/* validate */
1586	PSUPCALLSERVICE pReq = (PSUPCALLSERVICE)pReqHdr;
1587	Log4(("SUP_IOCTL_CALL_SERVICE: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1588	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1589
1590	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
1591	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(0), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(0));
1592	else
1593	{
1594	PSUPR0SERVICEREQHDR pSrvReq = (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0];
1595	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR)),
1596	("SUP_IOCTL_CALL_SERVICE: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR))));
1597	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, pSrvReq->u32Magic == SUPR0SERVICEREQHDR_MAGIC);
1598	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(pSrvReq->cbReq), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(pSrvReq->cbReq));
1599	}
1600	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1601
1602	/* execute */
1603	pReq->Hdr.rc = supdrvIOCtl_CallServiceModule(pDevExt, pSession, pReq);
1604	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1605	return 0;
1606	}
1607
1608	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOGGER_SETTINGS(0)):
1609	{
1610	/* validate */
1611	PSUPLOGGERSETTINGS pReq = (PSUPLOGGERSETTINGS)pReqHdr;
1612	size_t cbStrTab;
1613	REQ_CHECK_SIZE_OUT(SUP_IOCTL_LOGGER_SETTINGS, SUP_IOCTL_LOGGER_SETTINGS_SIZE_OUT);
1614	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->Hdr.cbIn >= SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(1));
1615	cbStrTab = pReq->Hdr.cbIn - SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(0);
1616	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offGroups < cbStrTab);
1617	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offFlags < cbStrTab);
1618	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offDestination < cbStrTab);
1619	REQ_CHECK_EXPR_FMT(pReq->u.In.szStrings[cbStrTab - 1] == '\0',
1620	("SUP_IOCTL_LOGGER_SETTINGS: cbIn=%#x cbStrTab=%#zx LastChar=%d\n",
1621	pReq->Hdr.cbIn, cbStrTab, pReq->u.In.szStrings[cbStrTab - 1]));
1622	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhich <= SUPLOGGERSETTINGS_WHICH_RELEASE);
1623	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhat <= SUPLOGGERSETTINGS_WHAT_DESTROY);
1624
1625	/* execute */
1626	pReq->Hdr.rc = supdrvIOCtl_LoggerSettings(pDevExt, pSession, pReq);
1627	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1628	return 0;
1629	}
1630
1631	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP2):
1632	{
1633	/* validate */
1634	PSUPSEMOP2 pReq = (PSUPSEMOP2)pReqHdr;
1635	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP2, SUP_IOCTL_SEM_OP2_SIZE_IN, SUP_IOCTL_SEM_OP2_SIZE_OUT);
1636	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP2, pReq->u.In.uReserved == 0);
1637
1638	/* execute */
1639	switch (pReq->u.In.uType)
1640	{
1641	case SUP_SEM_TYPE_EVENT:
1642	{
1643	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1644	switch (pReq->u.In.uOp)
1645	{
1646	case SUPSEMOP2_WAIT_MS_REL:
1647	pReq->Hdr.rc = SUPSemEventWaitNoResume(pSession, hEvent, pReq->u.In.uArg.cRelMsTimeout);
1648	break;
1649	case SUPSEMOP2_WAIT_NS_ABS:
1650	pReq->Hdr.rc = SUPSemEventWaitNsAbsIntr(pSession, hEvent, pReq->u.In.uArg.uAbsNsTimeout);
1651	break;
1652	case SUPSEMOP2_WAIT_NS_REL:
1653	pReq->Hdr.rc = SUPSemEventWaitNsRelIntr(pSession, hEvent, pReq->u.In.uArg.cRelNsTimeout);
1654	break;
1655	case SUPSEMOP2_SIGNAL:
1656	pReq->Hdr.rc = SUPSemEventSignal(pSession, hEvent);
1657	break;
1658	case SUPSEMOP2_CLOSE:
1659	pReq->Hdr.rc = SUPSemEventClose(pSession, hEvent);
1660	break;
1661	case SUPSEMOP2_RESET:
1662	default:
1663	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1664	break;
1665	}
1666	break;
1667	}
1668
1669	case SUP_SEM_TYPE_EVENT_MULTI:
1670	{
1671	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1672	switch (pReq->u.In.uOp)
1673	{
1674	case SUPSEMOP2_WAIT_MS_REL:
1675	pReq->Hdr.rc = SUPSemEventMultiWaitNoResume(pSession, hEventMulti, pReq->u.In.uArg.cRelMsTimeout);
1676	break;
1677	case SUPSEMOP2_WAIT_NS_ABS:
1678	pReq->Hdr.rc = SUPSemEventMultiWaitNsAbsIntr(pSession, hEventMulti, pReq->u.In.uArg.uAbsNsTimeout);
1679	break;
1680	case SUPSEMOP2_WAIT_NS_REL:
1681	pReq->Hdr.rc = SUPSemEventMultiWaitNsRelIntr(pSession, hEventMulti, pReq->u.In.uArg.cRelNsTimeout);
1682	break;
1683	case SUPSEMOP2_SIGNAL:
1684	pReq->Hdr.rc = SUPSemEventMultiSignal(pSession, hEventMulti);
1685	break;
1686	case SUPSEMOP2_CLOSE:
1687	pReq->Hdr.rc = SUPSemEventMultiClose(pSession, hEventMulti);
1688	break;
1689	case SUPSEMOP2_RESET:
1690	pReq->Hdr.rc = SUPSemEventMultiReset(pSession, hEventMulti);
1691	break;
1692	default:
1693	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1694	break;
1695	}
1696	break;
1697	}
1698
1699	default:
1700	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1701	break;
1702	}
1703	return 0;
1704	}
1705
1706	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP3):
1707	{
1708	/* validate */
1709	PSUPSEMOP3 pReq = (PSUPSEMOP3)pReqHdr;
1710	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP3, SUP_IOCTL_SEM_OP3_SIZE_IN, SUP_IOCTL_SEM_OP3_SIZE_OUT);
1711	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, pReq->u.In.u32Reserved == 0 && pReq->u.In.u64Reserved == 0);
1712
1713	/* execute */
1714	switch (pReq->u.In.uType)
1715	{
1716	case SUP_SEM_TYPE_EVENT:
1717	{
1718	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1719	switch (pReq->u.In.uOp)
1720	{
1721	case SUPSEMOP3_CREATE:
1722	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1723	pReq->Hdr.rc = SUPSemEventCreate(pSession, &hEvent);
1724	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEvent;
1725	break;
1726	case SUPSEMOP3_GET_RESOLUTION:
1727	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1728	pReq->Hdr.rc = VINF_SUCCESS;
1729	pReq->Hdr.cbOut = sizeof(*pReq);
1730	pReq->u.Out.cNsResolution = SUPSemEventGetResolution(pSession);
1731	break;
1732	default:
1733	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1734	break;
1735	}
1736	break;
1737	}
1738
1739	case SUP_SEM_TYPE_EVENT_MULTI:
1740	{
1741	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1742	switch (pReq->u.In.uOp)
1743	{
1744	case SUPSEMOP3_CREATE:
1745	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1746	pReq->Hdr.rc = SUPSemEventMultiCreate(pSession, &hEventMulti);
1747	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEventMulti;
1748	break;
1749	case SUPSEMOP3_GET_RESOLUTION:
1750	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1751	pReq->Hdr.rc = VINF_SUCCESS;
1752	pReq->u.Out.cNsResolution = SUPSemEventMultiGetResolution(pSession);
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_VT_CAPS):
1769	{
1770	/* validate */
1771	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
1772	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
1773	REQ_CHECK_EXPR(SUP_IOCTL_VT_CAPS, pReq->Hdr.cbIn <= SUP_IOCTL_VT_CAPS_SIZE_IN);
1774
1775	/* execute */
1776	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.Caps);
1777	if (RT_FAILURE(pReq->Hdr.rc))
1778	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1779	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1780	return 0;
1781	}
1782
1783	default:
1784	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1785	break;
1786	}
1787	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_GENERAL_FAILURE, VERR_GENERAL_FAILURE);
1788	return VERR_GENERAL_FAILURE;
1789	}
1790
1791
1792	/**
1793	* Inter-Driver Communication (IDC) worker.
1794	*
1795	* @returns VBox status code.
1796	* @retval VINF_SUCCESS on success.
1797	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1798	* @retval VERR_NOT_SUPPORTED if the request isn't supported.
1799	*
1800	* @param uReq The request (function) code.
1801	* @param pDevExt Device extention.
1802	* @param pSession Session data.
1803	* @param pReqHdr The request header.
1804	*/
1805	int VBOXCALL supdrvIDC(uintptr_t uReq, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQHDR pReqHdr)
1806	{
1807	/*
1808	* The OS specific code has already validated the pSession
1809	* pointer, and the request size being greater or equal to
1810	* size of the header.
1811	*
1812	* So, just check that pSession is a kernel context session.
1813	*/
1814	if (RT_UNLIKELY( pSession
1815	&& pSession->R0Process != NIL_RTR0PROCESS))
1816	return VERR_INVALID_PARAMETER;
1817
1818	/*
1819	* Validation macro.
1820	*/
1821	#define REQ_CHECK_IDC_SIZE(Name, cbExpect) \
1822	do { \
1823	if (RT_UNLIKELY(pReqHdr->cb != (cbExpect))) \
1824	{ \
1825	OSDBGPRINT(( #Name ": Invalid input/output sizes. cb=%ld expected %ld.\n", \
1826	(long)pReqHdr->cb, (long)(cbExpect))); \
1827	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1828	} \
1829	} while (0)
1830
1831	switch (uReq)
1832	{
1833	case SUPDRV_IDC_REQ_CONNECT:
1834	{
1835	PSUPDRVIDCREQCONNECT pReq = (PSUPDRVIDCREQCONNECT)pReqHdr;
1836	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_CONNECT, sizeof(*pReq));
1837
1838	/*
1839	* Validate the cookie and other input.
1840	*/
1841	if (pReq->Hdr.pSession != NULL)
1842	{
1843	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: pSession=%p expected NULL!\n", pReq->Hdr.pSession));
1844	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1845	}
1846	if (pReq->u.In.u32MagicCookie != SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE)
1847	{
1848	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: u32MagicCookie=%#x expected %#x!\n",
1849	(unsigned)pReq->u.In.u32MagicCookie, (unsigned)SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE));
1850	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1851	}
1852	if ( pReq->u.In.uMinVersion > pReq->u.In.uReqVersion
1853	\|\| (pReq->u.In.uMinVersion & UINT32_C(0xffff0000)) != (pReq->u.In.uReqVersion & UINT32_C(0xffff0000)))
1854	{
1855	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: uMinVersion=%#x uMaxVersion=%#x doesn't match!\n",
1856	pReq->u.In.uMinVersion, pReq->u.In.uReqVersion));
1857	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1858	}
1859
1860	/*
1861	* Match the version.
1862	* The current logic is very simple, match the major interface version.
1863	*/
1864	if ( pReq->u.In.uMinVersion > SUPDRV_IDC_VERSION
1865	\|\| (pReq->u.In.uMinVersion & 0xffff0000) != (SUPDRV_IDC_VERSION & 0xffff0000))
1866	{
1867	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1868	pReq->u.In.uReqVersion, pReq->u.In.uMinVersion, (unsigned)SUPDRV_IDC_VERSION));
1869	pReq->u.Out.pSession = NULL;
1870	pReq->u.Out.uSessionVersion = 0xffffffff;
1871	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1872	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1873	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1874	return VINF_SUCCESS;
1875	}
1876
1877	pReq->u.Out.pSession = NULL;
1878	pReq->u.Out.uSessionVersion = SUPDRV_IDC_VERSION;
1879	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1880	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1881
1882	/*
1883	* On NT we will already have a session associated with the
1884	* client, just like with the SUP_IOCTL_COOKIE request, while
1885	* the other doesn't.
1886	*/
1887	#ifdef RT_OS_WINDOWS
1888	pReq->Hdr.rc = VINF_SUCCESS;
1889	#else
1890	AssertReturn(!pSession, VERR_INTERNAL_ERROR);
1891	pReq->Hdr.rc = supdrvCreateSession(pDevExt, false /* fUser */, &pSession);
1892	if (RT_FAILURE(pReq->Hdr.rc))
1893	{
1894	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: failed to create session, rc=%d\n", pReq->Hdr.rc));
1895	return VINF_SUCCESS;
1896	}
1897	#endif
1898
1899	pReq->u.Out.pSession = pSession;
1900	pReq->Hdr.pSession = pSession;
1901
1902	return VINF_SUCCESS;
1903	}
1904
1905	case SUPDRV_IDC_REQ_DISCONNECT:
1906	{
1907	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_DISCONNECT, sizeof(*pReqHdr));
1908
1909	#ifdef RT_OS_WINDOWS
1910	/* Windows will destroy the session when the file object is destroyed. */
1911	#else
1912	supdrvCloseSession(pDevExt, pSession);
1913	#endif
1914	return pReqHdr->rc = VINF_SUCCESS;
1915	}
1916
1917	case SUPDRV_IDC_REQ_GET_SYMBOL:
1918	{
1919	PSUPDRVIDCREQGETSYM pReq = (PSUPDRVIDCREQGETSYM)pReqHdr;
1920	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_GET_SYMBOL, sizeof(*pReq));
1921
1922	pReq->Hdr.rc = supdrvIDC_LdrGetSymbol(pDevExt, pSession, pReq);
1923	return VINF_SUCCESS;
1924	}
1925
1926	case SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY:
1927	{
1928	PSUPDRVIDCREQCOMPREGFACTORY pReq = (PSUPDRVIDCREQCOMPREGFACTORY)pReqHdr;
1929	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY, sizeof(*pReq));
1930
1931	pReq->Hdr.rc = SUPR0ComponentRegisterFactory(pSession, pReq->u.In.pFactory);
1932	return VINF_SUCCESS;
1933	}
1934
1935	case SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY:
1936	{
1937	PSUPDRVIDCREQCOMPDEREGFACTORY pReq = (PSUPDRVIDCREQCOMPDEREGFACTORY)pReqHdr;
1938	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY, sizeof(*pReq));
1939
1940	pReq->Hdr.rc = SUPR0ComponentDeregisterFactory(pSession, pReq->u.In.pFactory);
1941	return VINF_SUCCESS;
1942	}
1943
1944	default:
1945	Log(("Unknown IDC %#lx\n", (long)uReq));
1946	break;
1947	}
1948
1949	#undef REQ_CHECK_IDC_SIZE
1950	return VERR_NOT_SUPPORTED;
1951	}
1952
1953
1954	/**
1955	* Register a object for reference counting.
1956	* The object is registered with one reference in the specified session.
1957	*
1958	* @returns Unique identifier on success (pointer).
1959	* All future reference must use this identifier.
1960	* @returns NULL on failure.
1961	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
1962	* @param pvUser1 The first user argument.
1963	* @param pvUser2 The second user argument.
1964	*/
1965	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
1966	{
1967	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1968	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1969	PSUPDRVOBJ pObj;
1970	PSUPDRVUSAGE pUsage;
1971
1972	/*
1973	* Validate the input.
1974	*/
1975	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
1976	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
1977	AssertPtrReturn(pfnDestructor, NULL);
1978
1979	/*
1980	* Allocate and initialize the object.
1981	*/
1982	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
1983	if (!pObj)
1984	return NULL;
1985	pObj->u32Magic = SUPDRVOBJ_MAGIC;
1986	pObj->enmType = enmType;
1987	pObj->pNext = NULL;
1988	pObj->cUsage = 1;
1989	pObj->pfnDestructor = pfnDestructor;
1990	pObj->pvUser1 = pvUser1;
1991	pObj->pvUser2 = pvUser2;
1992	pObj->CreatorUid = pSession->Uid;
1993	pObj->CreatorGid = pSession->Gid;
1994	pObj->CreatorProcess= pSession->Process;
1995	supdrvOSObjInitCreator(pObj, pSession);
1996
1997	/*
1998	* Allocate the usage record.
1999	* (We keep freed usage records around to simplify SUPR0ObjAddRefEx().)
2000	*/
2001	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2002
2003	pUsage = pDevExt->pUsageFree;
2004	if (pUsage)
2005	pDevExt->pUsageFree = pUsage->pNext;
2006	else
2007	{
2008	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2009	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
2010	if (!pUsage)
2011	{
2012	RTMemFree(pObj);
2013	return NULL;
2014	}
2015	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2016	}
2017
2018	/*
2019	* Insert the object and create the session usage record.
2020	*/
2021	/* The object. */
2022	pObj->pNext = pDevExt->pObjs;
2023	pDevExt->pObjs = pObj;
2024
2025	/* The session record. */
2026	pUsage->cUsage = 1;
2027	pUsage->pObj = pObj;
2028	pUsage->pNext = pSession->pUsage;
2029	/* Log2(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext)); */
2030	pSession->pUsage = pUsage;
2031
2032	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2033
2034	Log(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
2035	return pObj;
2036	}
2037
2038
2039	/**
2040	* Increment the reference counter for the object associating the reference
2041	* with the specified session.
2042	*
2043	* @returns IPRT status code.
2044	* @param pvObj The identifier returned by SUPR0ObjRegister().
2045	* @param pSession The session which is referencing the object.
2046	*
2047	* @remarks The caller should not own any spinlocks and must carefully protect
2048	* itself against potential race with the destructor so freed memory
2049	* isn't accessed here.
2050	*/
2051	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
2052	{
2053	return SUPR0ObjAddRefEx(pvObj, pSession, false /* fNoBlocking */);
2054	}
2055
2056
2057	/**
2058	* Increment the reference counter for the object associating the reference
2059	* with the specified session.
2060	*
2061	* @returns IPRT status code.
2062	* @retval VERR_TRY_AGAIN if fNoBlocking was set and a new usage record
2063	* couldn't be allocated. (If you see this you're not doing the right
2064	* thing and it won't ever work reliably.)
2065	*
2066	* @param pvObj The identifier returned by SUPR0ObjRegister().
2067	* @param pSession The session which is referencing the object.
2068	* @param fNoBlocking Set if it's not OK to block. Never try to make the
2069	* first reference to an object in a session with this
2070	* argument set.
2071	*
2072	* @remarks The caller should not own any spinlocks and must carefully protect
2073	* itself against potential race with the destructor so freed memory
2074	* isn't accessed here.
2075	*/
2076	SUPR0DECL(int) SUPR0ObjAddRefEx(void *pvObj, PSUPDRVSESSION pSession, bool fNoBlocking)
2077	{
2078	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2079	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2080	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2081	int rc = VINF_SUCCESS;
2082	PSUPDRVUSAGE pUsagePre;
2083	PSUPDRVUSAGE pUsage;
2084
2085	/*
2086	* Validate the input.
2087	* Be ready for the destruction race (someone might be stuck in the
2088	* destructor waiting a lock we own).
2089	*/
2090	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2091	AssertPtrReturn(pObj, VERR_INVALID_POINTER);
2092	AssertMsgReturn(pObj->u32Magic == SUPDRVOBJ_MAGIC \|\| pObj->u32Magic == SUPDRVOBJ_MAGIC_DEAD,
2093	("Invalid pvObj=%p magic=%#x (expected %#x or %#x)\n", pvObj, pObj->u32Magic, SUPDRVOBJ_MAGIC, SUPDRVOBJ_MAGIC_DEAD),
2094	VERR_INVALID_PARAMETER);
2095
2096	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2097
2098	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2099	{
2100	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2101
2102	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2103	return VERR_WRONG_ORDER;
2104	}
2105
2106	/*
2107	* Preallocate the usage record if we can.
2108	*/
2109	pUsagePre = pDevExt->pUsageFree;
2110	if (pUsagePre)
2111	pDevExt->pUsageFree = pUsagePre->pNext;
2112	else if (!fNoBlocking)
2113	{
2114	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2115	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
2116	if (!pUsagePre)
2117	return VERR_NO_MEMORY;
2118
2119	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2120	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2121	{
2122	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2123
2124	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2125	return VERR_WRONG_ORDER;
2126	}
2127	}
2128
2129	/*
2130	* Reference the object.
2131	*/
2132	pObj->cUsage++;
2133
2134	/*
2135	* Look for the session record.
2136	*/
2137	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
2138	{
2139	/Log(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2140	if (pUsage->pObj == pObj)
2141	break;
2142	}
2143	if (pUsage)
2144	pUsage->cUsage++;
2145	else if (pUsagePre)
2146	{
2147	/* create a new session record. */
2148	pUsagePre->cUsage = 1;
2149	pUsagePre->pObj = pObj;
2150	pUsagePre->pNext = pSession->pUsage;
2151	pSession->pUsage = pUsagePre;
2152	/Log(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));/
2153
2154	pUsagePre = NULL;
2155	}
2156	else
2157	{
2158	pObj->cUsage--;
2159	rc = VERR_TRY_AGAIN;
2160	}
2161
2162	/*
2163	* Put any unused usage record into the free list..
2164	*/
2165	if (pUsagePre)
2166	{
2167	pUsagePre->pNext = pDevExt->pUsageFree;
2168	pDevExt->pUsageFree = pUsagePre;
2169	}
2170
2171	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2172
2173	return rc;
2174	}
2175
2176
2177	/**
2178	* Decrement / destroy a reference counter record for an object.
2179	*
2180	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
2181	*
2182	* @returns IPRT status code.
2183	* @retval VINF_SUCCESS if not destroyed.
2184	* @retval VINF_OBJECT_DESTROYED if it's destroyed by this release call.
2185	* @retval VERR_INVALID_PARAMETER if the object isn't valid. Will assert in
2186	* string builds.
2187	*
2188	* @param pvObj The identifier returned by SUPR0ObjRegister().
2189	* @param pSession The session which is referencing the object.
2190	*/
2191	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
2192	{
2193	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2194	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2195	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2196	int rc = VERR_INVALID_PARAMETER;
2197	PSUPDRVUSAGE pUsage;
2198	PSUPDRVUSAGE pUsagePrev;
2199
2200	/*
2201	* Validate the input.
2202	*/
2203	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2204	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2205	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2206	VERR_INVALID_PARAMETER);
2207
2208	/*
2209	* Acquire the spinlock and look for the usage record.
2210	*/
2211	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2212
2213	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
2214	pUsage;
2215	pUsagePrev = pUsage, pUsage = pUsage->pNext)
2216	{
2217	/Log2(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2218	if (pUsage->pObj == pObj)
2219	{
2220	rc = VINF_SUCCESS;
2221	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
2222	if (pUsage->cUsage > 1)
2223	{
2224	pObj->cUsage--;
2225	pUsage->cUsage--;
2226	}
2227	else
2228	{
2229	/*
2230	* Free the session record.
2231	*/
2232	if (pUsagePrev)
2233	pUsagePrev->pNext = pUsage->pNext;
2234	else
2235	pSession->pUsage = pUsage->pNext;
2236	pUsage->pNext = pDevExt->pUsageFree;
2237	pDevExt->pUsageFree = pUsage;
2238
2239	/* What about the object? */
2240	if (pObj->cUsage > 1)
2241	pObj->cUsage--;
2242	else
2243	{
2244	/*
2245	* Object is to be destroyed, unlink it.
2246	*/
2247	pObj->u32Magic = SUPDRVOBJ_MAGIC_DEAD;
2248	rc = VINF_OBJECT_DESTROYED;
2249	if (pDevExt->pObjs == pObj)
2250	pDevExt->pObjs = pObj->pNext;
2251	else
2252	{
2253	PSUPDRVOBJ pObjPrev;
2254	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
2255	if (pObjPrev->pNext == pObj)
2256	{
2257	pObjPrev->pNext = pObj->pNext;
2258	break;
2259	}
2260	Assert(pObjPrev);
2261	}
2262	}
2263	}
2264	break;
2265	}
2266	}
2267
2268	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2269
2270	/*
2271	* Call the destructor and free the object if required.
2272	*/
2273	if (rc == VINF_OBJECT_DESTROYED)
2274	{
2275	Log(("SUPR0ObjRelease: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
2276	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
2277	if (pObj->pfnDestructor)
2278	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
2279	RTMemFree(pObj);
2280	}
2281
2282	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
2283	return rc;
2284	}
2285
2286
2287	/**
2288	* Verifies that the current process can access the specified object.
2289	*
2290	* @returns The following IPRT status code:
2291	* @retval VINF_SUCCESS if access was granted.
2292	* @retval VERR_PERMISSION_DENIED if denied access.
2293	* @retval VERR_INVALID_PARAMETER if invalid parameter.
2294	*
2295	* @param pvObj The identifier returned by SUPR0ObjRegister().
2296	* @param pSession The session which wishes to access the object.
2297	* @param pszObjName Object string name. This is optional and depends on the object type.
2298	*
2299	* @remark The caller is responsible for making sure the object isn't removed while
2300	* we're inside this function. If uncertain about this, just call AddRef before calling us.
2301	*/
2302	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
2303	{
2304	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2305	int rc;
2306
2307	/*
2308	* Validate the input.
2309	*/
2310	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2311	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2312	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2313	VERR_INVALID_PARAMETER);
2314
2315	/*
2316	* Check access. (returns true if a decision has been made.)
2317	*/
2318	rc = VERR_INTERNAL_ERROR;
2319	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
2320	return rc;
2321
2322	/*
2323	* Default policy is to allow the user to access his own
2324	* stuff but nothing else.
2325	*/
2326	if (pObj->CreatorUid == pSession->Uid)
2327	return VINF_SUCCESS;
2328	return VERR_PERMISSION_DENIED;
2329	}
2330
2331
2332	/**
2333	* Lock pages.
2334	*
2335	* @returns IPRT status code.
2336	* @param pSession Session to which the locked memory should be associated.
2337	* @param pvR3 Start of the memory range to lock.
2338	* This must be page aligned.
2339	* @param cPages Number of pages to lock.
2340	* @param paPages Where to put the physical addresses of locked memory.
2341	*/
2342	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
2343	{
2344	int rc;
2345	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2346	const size_t cb = (size_t)cPages << PAGE_SHIFT;
2347	LogFlow(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
2348
2349	/*
2350	* Verify input.
2351	*/
2352	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2353	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
2354	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
2355	\|\| !pvR3)
2356	{
2357	Log(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
2358	return VERR_INVALID_PARAMETER;
2359	}
2360
2361	/*
2362	* Let IPRT do the job.
2363	*/
2364	Mem.eType = MEMREF_TYPE_LOCKED;
2365	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTMEM_PROT_READ \| RTMEM_PROT_WRITE, RTR0ProcHandleSelf());
2366	if (RT_SUCCESS(rc))
2367	{
2368	uint32_t iPage = cPages;
2369	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
2370	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
2371
2372	while (iPage-- > 0)
2373	{
2374	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2375	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
2376	{
2377	AssertMsgFailed(("iPage=%d\n", iPage));
2378	rc = VERR_INTERNAL_ERROR;
2379	break;
2380	}
2381	}
2382	if (RT_SUCCESS(rc))
2383	rc = supdrvMemAdd(&Mem, pSession);
2384	if (RT_FAILURE(rc))
2385	{
2386	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
2387	AssertRC(rc2);
2388	}
2389	}
2390
2391	return rc;
2392	}
2393
2394
2395	/**
2396	* Unlocks the memory pointed to by pv.
2397	*
2398	* @returns IPRT status code.
2399	* @param pSession Session to which the memory was locked.
2400	* @param pvR3 Memory to unlock.
2401	*/
2402	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2403	{
2404	LogFlow(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2405	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2406	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
2407	}
2408
2409
2410	/**
2411	* Allocates a chunk of page aligned memory with contiguous and fixed physical
2412	* backing.
2413	*
2414	* @returns IPRT status code.
2415	* @param pSession Session data.
2416	* @param cPages Number of pages to allocate.
2417	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
2418	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
2419	* @param pHCPhys Where to put the physical address of allocated memory.
2420	*/
2421	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
2422	{
2423	int rc;
2424	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2425	LogFlow(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
2426
2427	/*
2428	* Validate input.
2429	*/
2430	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2431	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
2432	{
2433	Log(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
2434	pSession, ppvR0, ppvR3, pHCPhys));
2435	return VERR_INVALID_PARAMETER;
2436
2437	}
2438	if (cPages < 1 \|\| cPages >= 256)
2439	{
2440	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2441	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2442	}
2443
2444	/*
2445	* Let IPRT do the job.
2446	*/
2447	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
2448	if (RT_SUCCESS(rc))
2449	{
2450	int rc2;
2451	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2452	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2453	if (RT_SUCCESS(rc))
2454	{
2455	Mem.eType = MEMREF_TYPE_CONT;
2456	rc = supdrvMemAdd(&Mem, pSession);
2457	if (!rc)
2458	{
2459	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2460	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2461	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
2462	return 0;
2463	}
2464
2465	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2466	AssertRC(rc2);
2467	}
2468	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2469	AssertRC(rc2);
2470	}
2471
2472	return rc;
2473	}
2474
2475
2476	/**
2477	* Frees memory allocated using SUPR0ContAlloc().
2478	*
2479	* @returns IPRT status code.
2480	* @param pSession The session to which the memory was allocated.
2481	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2482	*/
2483	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2484	{
2485	LogFlow(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2486	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2487	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
2488	}
2489
2490
2491	/**
2492	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
2493	*
2494	* The memory isn't zeroed.
2495	*
2496	* @returns IPRT status code.
2497	* @param pSession Session data.
2498	* @param cPages Number of pages to allocate.
2499	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
2500	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
2501	* @param paPages Where to put the physical addresses of allocated memory.
2502	*/
2503	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
2504	{
2505	unsigned iPage;
2506	int rc;
2507	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2508	LogFlow(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
2509
2510	/*
2511	* Validate input.
2512	*/
2513	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2514	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
2515	{
2516	Log(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
2517	pSession, ppvR3, ppvR0, paPages));
2518	return VERR_INVALID_PARAMETER;
2519
2520	}
2521	if (cPages < 1 \|\| cPages >= 256)
2522	{
2523	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2524	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2525	}
2526
2527	/*
2528	* Let IPRT do the work.
2529	*/
2530	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
2531	if (RT_SUCCESS(rc))
2532	{
2533	int rc2;
2534	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2535	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2536	if (RT_SUCCESS(rc))
2537	{
2538	Mem.eType = MEMREF_TYPE_LOW;
2539	rc = supdrvMemAdd(&Mem, pSession);
2540	if (!rc)
2541	{
2542	for (iPage = 0; iPage < cPages; iPage++)
2543	{
2544	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2545	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%RHp\n", paPages[iPage]));
2546	}
2547	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2548	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2549	return 0;
2550	}
2551
2552	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2553	AssertRC(rc2);
2554	}
2555
2556	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2557	AssertRC(rc2);
2558	}
2559
2560	return rc;
2561	}
2562
2563
2564	/**
2565	* Frees memory allocated using SUPR0LowAlloc().
2566	*
2567	* @returns IPRT status code.
2568	* @param pSession The session to which the memory was allocated.
2569	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2570	*/
2571	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2572	{
2573	LogFlow(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2574	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2575	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
2576	}
2577
2578
2579
2580	/**
2581	* Allocates a chunk of memory with both R0 and R3 mappings.
2582	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
2583	*
2584	* @returns IPRT status code.
2585	* @param pSession The session to associated the allocation with.
2586	* @param cb Number of bytes to allocate.
2587	* @param ppvR0 Where to store the address of the Ring-0 mapping.
2588	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2589	*/
2590	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
2591	{
2592	int rc;
2593	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2594	LogFlow(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
2595
2596	/*
2597	* Validate input.
2598	*/
2599	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2600	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
2601	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
2602	if (cb < 1 \|\| cb >= _4M)
2603	{
2604	Log(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
2605	return VERR_INVALID_PARAMETER;
2606	}
2607
2608	/*
2609	* Let IPRT do the work.
2610	*/
2611	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
2612	if (RT_SUCCESS(rc))
2613	{
2614	int rc2;
2615	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2616	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2617	if (RT_SUCCESS(rc))
2618	{
2619	Mem.eType = MEMREF_TYPE_MEM;
2620	rc = supdrvMemAdd(&Mem, pSession);
2621	if (!rc)
2622	{
2623	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2624	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2625	return VINF_SUCCESS;
2626	}
2627
2628	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2629	AssertRC(rc2);
2630	}
2631
2632	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2633	AssertRC(rc2);
2634	}
2635
2636	return rc;
2637	}
2638
2639
2640	/**
2641	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
2642	*
2643	* @returns IPRT status code.
2644	* @param pSession The session to which the memory was allocated.
2645	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2646	* @param paPages Where to store the physical addresses.
2647	*/
2648	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
2649	{
2650	PSUPDRVBUNDLE pBundle;
2651	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2652	LogFlow(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
2653
2654	/*
2655	* Validate input.
2656	*/
2657	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2658	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
2659	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
2660
2661	/*
2662	* Search for the address.
2663	*/
2664	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2665	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2666	{
2667	if (pBundle->cUsed > 0)
2668	{
2669	unsigned i;
2670	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2671	{
2672	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
2673	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2674	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2675	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2676	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
2677	)
2678	)
2679	{
2680	const size_t cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2681	size_t iPage;
2682	for (iPage = 0; iPage < cPages; iPage++)
2683	{
2684	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2685	paPages[iPage].uReserved = 0;
2686	}
2687	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2688	return VINF_SUCCESS;
2689	}
2690	}
2691	}
2692	}
2693	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2694	Log(("Failed to find %p!!!\n", (void *)uPtr));
2695	return VERR_INVALID_PARAMETER;
2696	}
2697
2698
2699	/**
2700	* Free memory allocated by SUPR0MemAlloc().
2701	*
2702	* @returns IPRT status code.
2703	* @param pSession The session owning the allocation.
2704	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2705	*/
2706	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2707	{
2708	LogFlow(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2709	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2710	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
2711	}
2712
2713
2714	/**
2715	* Allocates a chunk of memory with a kernel or/and a user mode mapping.
2716	*
2717	* The memory is fixed and it's possible to query the physical addresses using
2718	* SUPR0MemGetPhys().
2719	*
2720	* @returns IPRT status code.
2721	* @param pSession The session to associated the allocation with.
2722	* @param cPages The number of pages to allocate.
2723	* @param fFlags Flags, reserved for the future. Must be zero.
2724	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2725	* NULL if no ring-3 mapping.
2726	* @param ppvR3 Where to store the address of the Ring-0 mapping.
2727	* NULL if no ring-0 mapping.
2728	* @param paPages Where to store the addresses of the pages. Optional.
2729	*/
2730	SUPR0DECL(int) SUPR0PageAllocEx(PSUPDRVSESSION pSession, uint32_t cPages, uint32_t fFlags, PRTR3PTR ppvR3, PRTR0PTR ppvR0, PRTHCPHYS paPages)
2731	{
2732	int rc;
2733	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2734	LogFlow(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
2735
2736	/*
2737	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2738	*/
2739	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2740	AssertPtrNullReturn(ppvR3, VERR_INVALID_POINTER);
2741	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2742	AssertReturn(ppvR3 \|\| ppvR0, VERR_INVALID_PARAMETER);
2743	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2744	if (cPages < 1 \|\| cPages > VBOX_MAX_ALLOC_PAGE_COUNT)
2745	{
2746	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)));
2747	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2748	}
2749
2750	/*
2751	* Let IPRT do the work.
2752	*/
2753	if (ppvR0)
2754	rc = RTR0MemObjAllocPage(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, true /* fExecutable */);
2755	else
2756	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
2757	if (RT_SUCCESS(rc))
2758	{
2759	int rc2;
2760	if (ppvR3)
2761	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2762	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2763	else
2764	Mem.MapObjR3 = NIL_RTR0MEMOBJ;
2765	if (RT_SUCCESS(rc))
2766	{
2767	Mem.eType = MEMREF_TYPE_PAGE;
2768	rc = supdrvMemAdd(&Mem, pSession);
2769	if (!rc)
2770	{
2771	if (ppvR3)
2772	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2773	if (ppvR0)
2774	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2775	if (paPages)
2776	{
2777	uint32_t iPage = cPages;
2778	while (iPage-- > 0)
2779	{
2780	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
2781	Assert(paPages[iPage] != NIL_RTHCPHYS);
2782	}
2783	}
2784	return VINF_SUCCESS;
2785	}
2786
2787	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2788	AssertRC(rc2);
2789	}
2790
2791	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2792	AssertRC(rc2);
2793	}
2794	return rc;
2795	}
2796
2797
2798	/**
2799	* Maps a chunk of memory previously allocated by SUPR0PageAllocEx into kernel
2800	* space.
2801	*
2802	* @returns IPRT status code.
2803	* @param pSession The session to associated the allocation with.
2804	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
2805	* @param offSub Where to start mapping. Must be page aligned.
2806	* @param cbSub How much to map. Must be page aligned.
2807	* @param fFlags Flags, MBZ.
2808	* @param ppvR0 Where to return the address of the ring-0 mapping on
2809	* success.
2810	*/
2811	SUPR0DECL(int) SUPR0PageMapKernel(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t offSub, uint32_t cbSub,
2812	uint32_t fFlags, PRTR0PTR ppvR0)
2813	{
2814	int rc;
2815	PSUPDRVBUNDLE pBundle;
2816	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2817	RTR0MEMOBJ hMemObj = NIL_RTR0MEMOBJ;
2818	LogFlow(("SUPR0PageMapKernel: pSession=%p pvR3=%p offSub=%#x cbSub=%#x\n", pSession, pvR3, offSub, cbSub));
2819
2820	/*
2821	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2822	*/
2823	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2824	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2825	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2826	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2827	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2828	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
2829
2830	/*
2831	* Find the memory object.
2832	*/
2833	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2834	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2835	{
2836	if (pBundle->cUsed > 0)
2837	{
2838	unsigned i;
2839	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2840	{
2841	if ( ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
2842	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2843	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2844	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
2845	\|\| ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED
2846	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2847	&& pBundle->aMem[i].MapObjR3 == NIL_RTR0MEMOBJ
2848	&& RTR0MemObjAddressR3(pBundle->aMem[i].MemObj) == pvR3))
2849	{
2850	hMemObj = pBundle->aMem[i].MemObj;
2851	break;
2852	}
2853	}
2854	}
2855	}
2856	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2857
2858	rc = VERR_INVALID_PARAMETER;
2859	if (hMemObj != NIL_RTR0MEMOBJ)
2860	{
2861	/*
2862	* Do some further input validations before calling IPRT.
2863	* (Cleanup is done indirectly by telling RTR0MemObjFree to include mappings.)
2864	*/
2865	size_t cbMemObj = RTR0MemObjSize(hMemObj);
2866	if ( offSub < cbMemObj
2867	&& cbSub <= cbMemObj
2868	&& offSub + cbSub <= cbMemObj)
2869	{
2870	RTR0MEMOBJ hMapObj;
2871	rc = RTR0MemObjMapKernelEx(&hMapObj, hMemObj, (void *)-1, 0,
2872	RTMEM_PROT_READ \| RTMEM_PROT_WRITE, offSub, cbSub);
2873	if (RT_SUCCESS(rc))
2874	*ppvR0 = RTR0MemObjAddress(hMapObj);
2875	}
2876	else
2877	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
2878
2879	}
2880	return rc;
2881	}
2882
2883
2884	/**
2885	* Changes the page level protection of one or more pages previously allocated
2886	* by SUPR0PageAllocEx.
2887	*
2888	* @returns IPRT status code.
2889	* @param pSession The session to associated the allocation with.
2890	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
2891	* NIL_RTR3PTR if the ring-3 mapping should be unaffected.
2892	* @param pvR0 The ring-0 address returned by SUPR0PageAllocEx.
2893	* NIL_RTR0PTR if the ring-0 mapping should be unaffected.
2894	* @param offSub Where to start changing. Must be page aligned.
2895	* @param cbSub How much to change. Must be page aligned.
2896	* @param fProt The new page level protection, see RTMEM_PROT_*.
2897	*/
2898	SUPR0DECL(int) SUPR0PageProtect(PSUPDRVSESSION pSession, RTR3PTR pvR3, RTR0PTR pvR0, uint32_t offSub, uint32_t cbSub, uint32_t fProt)
2899	{
2900	int rc;
2901	PSUPDRVBUNDLE pBundle;
2902	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2903	RTR0MEMOBJ hMemObjR0 = NIL_RTR0MEMOBJ;
2904	RTR0MEMOBJ hMemObjR3 = NIL_RTR0MEMOBJ;
2905	LogFlow(("SUPR0PageProtect: pSession=%p pvR3=%p pvR0=%p offSub=%#x cbSub=%#x fProt-%#x\n", pSession, pvR3, pvR0, offSub, cbSub, fProt));
2906
2907	/*
2908	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2909	*/
2910	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2911	AssertReturn(!(fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)), VERR_INVALID_PARAMETER);
2912	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2913	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2914	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
2915
2916	/*
2917	* Find the memory object.
2918	*/
2919	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2920	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2921	{
2922	if (pBundle->cUsed > 0)
2923	{
2924	unsigned i;
2925	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2926	{
2927	if ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
2928	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2929	&& ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2930	\|\| pvR3 == NIL_RTR3PTR)
2931	&& ( pvR0 == NIL_RTR0PTR
2932	\|\| RTR0MemObjAddress(pBundle->aMem[i].MemObj) == pvR0)
2933	&& ( pvR3 == NIL_RTR3PTR
2934	\|\| RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3))
2935	{
2936	if (pvR0 != NIL_RTR0PTR)
2937	hMemObjR0 = pBundle->aMem[i].MemObj;
2938	if (pvR3 != NIL_RTR3PTR)
2939	hMemObjR3 = pBundle->aMem[i].MapObjR3;
2940	break;
2941	}
2942	}
2943	}
2944	}
2945	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2946
2947	rc = VERR_INVALID_PARAMETER;
2948	if ( hMemObjR0 != NIL_RTR0MEMOBJ
2949	\|\| hMemObjR3 != NIL_RTR0MEMOBJ)
2950	{
2951	/*
2952	* Do some further input validations before calling IPRT.
2953	*/
2954	size_t cbMemObj = hMemObjR0 != NIL_RTR0PTR ? RTR0MemObjSize(hMemObjR0) : RTR0MemObjSize(hMemObjR3);
2955	if ( offSub < cbMemObj
2956	&& cbSub <= cbMemObj
2957	&& offSub + cbSub <= cbMemObj)
2958	{
2959	rc = VINF_SUCCESS;
2960	if (hMemObjR3 != NIL_RTR0PTR)
2961	rc = RTR0MemObjProtect(hMemObjR3, offSub, cbSub, fProt);
2962	if (hMemObjR0 != NIL_RTR0PTR && RT_SUCCESS(rc))
2963	rc = RTR0MemObjProtect(hMemObjR0, offSub, cbSub, fProt);
2964	}
2965	else
2966	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
2967
2968	}
2969	return rc;
2970
2971	}
2972
2973
2974	/**
2975	* Free memory allocated by SUPR0PageAlloc() and SUPR0PageAllocEx().
2976	*
2977	* @returns IPRT status code.
2978	* @param pSession The session owning the allocation.
2979	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc() or
2980	* SUPR0PageAllocEx().
2981	*/
2982	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2983	{
2984	LogFlow(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2985	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2986	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_PAGE);
2987	}
2988
2989
2990	/**
2991	* Gets the paging mode of the current CPU.
2992	*
2993	* @returns Paging mode, SUPPAGEINGMODE_INVALID on error.
2994	*/
2995	SUPR0DECL(SUPPAGINGMODE) SUPR0GetPagingMode(void)
2996	{
2997	SUPPAGINGMODE enmMode;
2998
2999	RTR0UINTREG cr0 = ASMGetCR0();
3000	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
3001	enmMode = SUPPAGINGMODE_INVALID;
3002	else
3003	{
3004	RTR0UINTREG cr4 = ASMGetCR4();
3005	uint32_t fNXEPlusLMA = 0;
3006	if (cr4 & X86_CR4_PAE)
3007	{
3008	uint32_t fAmdFeatures = ASMCpuId_EDX(0x80000001);
3009	if (fAmdFeatures & (X86_CPUID_AMD_FEATURE_EDX_NX \| X86_CPUID_AMD_FEATURE_EDX_LONG_MODE))
3010	{
3011	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
3012	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
3013	fNXEPlusLMA \|= RT_BIT(0);
3014	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
3015	fNXEPlusLMA \|= RT_BIT(1);
3016	}
3017	}
3018
3019	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
3020	{
3021	case 0:
3022	enmMode = SUPPAGINGMODE_32_BIT;
3023	break;
3024
3025	case X86_CR4_PGE:
3026	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
3027	break;
3028
3029	case X86_CR4_PAE:
3030	enmMode = SUPPAGINGMODE_PAE;
3031	break;
3032
3033	case X86_CR4_PAE \| RT_BIT(0):
3034	enmMode = SUPPAGINGMODE_PAE_NX;
3035	break;
3036
3037	case X86_CR4_PAE \| X86_CR4_PGE:
3038	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3039	break;
3040
3041	case X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3042	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3043	break;
3044
3045	case RT_BIT(1) \| X86_CR4_PAE:
3046	enmMode = SUPPAGINGMODE_AMD64;
3047	break;
3048
3049	case RT_BIT(1) \| X86_CR4_PAE \| RT_BIT(0):
3050	enmMode = SUPPAGINGMODE_AMD64_NX;
3051	break;
3052
3053	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
3054	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
3055	break;
3056
3057	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3058	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
3059	break;
3060
3061	default:
3062	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
3063	enmMode = SUPPAGINGMODE_INVALID;
3064	break;
3065	}
3066	}
3067	return enmMode;
3068	}
3069
3070
3071	/**
3072	* Enables or disabled hardware virtualization extensions using native OS APIs.
3073	*
3074	* @returns VBox status code.
3075	* @retval VINF_SUCCESS on success.
3076	* @retval VERR_NOT_SUPPORTED if not supported by the native OS.
3077	*
3078	* @param fEnable Whether to enable or disable.
3079	*/
3080	SUPR0DECL(int) SUPR0EnableVTx(bool fEnable)
3081	{
3082	#ifdef RT_OS_DARWIN
3083	return supdrvOSEnableVTx(fEnable);
3084	#else
3085	return VERR_NOT_SUPPORTED;
3086	#endif
3087	}
3088
3089
3090	/** @todo document me */
3091	SUPR0DECL(int) SUPR0QueryVTCaps(PSUPDRVSESSION pSession, uint32_t *pfCaps)
3092	{
3093	/*
3094	* Input validation.
3095	*/
3096	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3097	AssertPtrReturn(pfCaps, VERR_INVALID_POINTER);
3098
3099	*pfCaps = 0;
3100
3101	if (ASMHasCpuId())
3102	{
3103	uint32_t u32FeaturesECX;
3104	uint32_t u32Dummy;
3105	uint32_t u32FeaturesEDX;
3106	uint32_t u32VendorEBX, u32VendorECX, u32VendorEDX, u32AMDFeatureEDX, u32AMDFeatureECX;
3107	uint64_t val;
3108
3109	ASMCpuId(0, &u32Dummy, &u32VendorEBX, &u32VendorECX, &u32VendorEDX);
3110	ASMCpuId(1, &u32Dummy, &u32Dummy, &u32FeaturesECX, &u32FeaturesEDX);
3111	/* Query AMD features. */
3112	ASMCpuId(0x80000001, &u32Dummy, &u32Dummy, &u32AMDFeatureECX, &u32AMDFeatureEDX);
3113
3114	if ( u32VendorEBX == X86_CPUID_VENDOR_INTEL_EBX
3115	&& u32VendorECX == X86_CPUID_VENDOR_INTEL_ECX
3116	&& u32VendorEDX == X86_CPUID_VENDOR_INTEL_EDX
3117	)
3118	{
3119	if ( (u32FeaturesECX & X86_CPUID_FEATURE_ECX_VMX)
3120	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3121	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3122	)
3123	{
3124	val = ASMRdMsr(MSR_IA32_FEATURE_CONTROL);
3125	/*
3126	* Both the LOCK and VMXON bit must be set; otherwise VMXON will generate a #GP.
3127	* Once the lock bit is set, this MSR can no longer be modified.
3128	*/
3129	if ( (val & (MSR_IA32_FEATURE_CONTROL_VMXON\|MSR_IA32_FEATURE_CONTROL_LOCK))
3130	== (MSR_IA32_FEATURE_CONTROL_VMXON\|MSR_IA32_FEATURE_CONTROL_LOCK) /* enabled and locked */
3131	\|\| !(val & MSR_IA32_FEATURE_CONTROL_LOCK) /* not enabled, but not locked either */
3132	)
3133	{
3134	VMX_CAPABILITY vtCaps;
3135
3136	*pfCaps \|= SUPVTCAPS_VT_X;
3137
3138	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS);
3139	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC_USE_SECONDARY_EXEC_CTRL)
3140	{
3141	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS2);
3142	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC2_EPT)
3143	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3144	}
3145	return VINF_SUCCESS;
3146	}
3147	return VERR_VMX_MSR_LOCKED_OR_DISABLED;
3148	}
3149	return VERR_VMX_NO_VMX;
3150	}
3151
3152	if ( u32VendorEBX == X86_CPUID_VENDOR_AMD_EBX
3153	&& u32VendorECX == X86_CPUID_VENDOR_AMD_ECX
3154	&& u32VendorEDX == X86_CPUID_VENDOR_AMD_EDX
3155	)
3156	{
3157	if ( (u32AMDFeatureECX & X86_CPUID_AMD_FEATURE_ECX_SVM)
3158	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3159	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3160	)
3161	{
3162	/* Check if SVM is disabled */
3163	val = ASMRdMsr(MSR_K8_VM_CR);
3164	if (!(val & MSR_K8_VM_CR_SVM_DISABLE))
3165	{
3166	*pfCaps \|= SUPVTCAPS_AMD_V;
3167
3168	/* Query AMD features. */
3169	ASMCpuId(0x8000000A, &u32Dummy, &u32Dummy, &u32Dummy, &u32FeaturesEDX);
3170
3171	if (u32FeaturesEDX & AMD_CPUID_SVM_FEATURE_EDX_NESTED_PAGING)
3172	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3173
3174	return VINF_SUCCESS;
3175	}
3176	return VERR_SVM_DISABLED;
3177	}
3178	return VERR_SVM_NO_SVM;
3179	}
3180	}
3181
3182	return VERR_UNSUPPORTED_CPU;
3183	}
3184
3185
3186	/**
3187	* (Re-)initializes the per-cpu structure prior to starting or resuming the GIP
3188	* updating.
3189	*
3190	* @param pGipCpu The per CPU structure for this CPU.
3191	* @param u64NanoTS The current time.
3192	*/
3193	static void supdrvGipReInitCpu(PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
3194	{
3195	pGipCpu->u64TSC = ASMReadTSC() - pGipCpu->u32UpdateIntervalTSC;
3196	pGipCpu->u64NanoTS = u64NanoTS;
3197	}
3198
3199
3200	/**
3201	* Set the current TSC and NanoTS value for the CPU.
3202	*
3203	* @param idCpu The CPU ID. Unused - we have to use the APIC ID.
3204	* @param pvUser1 Pointer to the ring-0 GIP mapping.
3205	* @param pvUser2 Pointer to the variable holding the current time.
3206	*/
3207	static DECLCALLBACK(void) supdrvGipReInitCpuCallback(RTCPUID idCpu, void pvUser1, void pvUser2)
3208	{
3209	PSUPGLOBALINFOPAGE pGip = (PSUPGLOBALINFOPAGE)pvUser1;
3210	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
3211
3212	if (RT_LIKELY(iCpu < pGip->cCpus && pGip->aCPUs[iCpu].idCpu == idCpu))
3213	supdrvGipReInitCpu(&pGip->aCPUs[iCpu], (uint64_t )pvUser2);
3214
3215	NOREF(pvUser2);
3216	NOREF(idCpu);
3217	}
3218
3219
3220	/**
3221	* Maps the GIP into userspace and/or get the physical address of the GIP.
3222	*
3223	* @returns IPRT status code.
3224	* @param pSession Session to which the GIP mapping should belong.
3225	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
3226	* @param pHCPhysGip Where to store the physical address. (optional)
3227	*
3228	* @remark There is no reference counting on the mapping, so one call to this function
3229	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
3230	* and remove the session as a GIP user.
3231	*/
3232	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
3233	{
3234	int rc;
3235	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3236	RTR3PTR pGipR3 = NIL_RTR3PTR;
3237	RTHCPHYS HCPhys = NIL_RTHCPHYS;
3238	LogFlow(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
3239
3240	/*
3241	* Validate
3242	*/
3243	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3244	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
3245	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
3246
3247	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3248	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3249	#else
3250	RTSemFastMutexRequest(pDevExt->mtxGip);
3251	#endif
3252	if (pDevExt->pGip)
3253	{
3254	/*
3255	* Map it?
3256	*/
3257	rc = VINF_SUCCESS;
3258	if (ppGipR3)
3259	{
3260	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
3261	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
3262	RTMEM_PROT_READ, RTR0ProcHandleSelf());
3263	if (RT_SUCCESS(rc))
3264	pGipR3 = RTR0MemObjAddressR3(pSession->GipMapObjR3);
3265	}
3266
3267	/*
3268	* Get physical address.
3269	*/
3270	if (pHCPhysGip && RT_SUCCESS(rc))
3271	HCPhys = pDevExt->HCPhysGip;
3272
3273	/*
3274	* Reference globally.
3275	*/
3276	if (!pSession->fGipReferenced && RT_SUCCESS(rc))
3277	{
3278	pSession->fGipReferenced = 1;
3279	pDevExt->cGipUsers++;
3280	if (pDevExt->cGipUsers == 1)
3281	{
3282	PSUPGLOBALINFOPAGE pGipR0 = pDevExt->pGip;
3283	uint64_t u64NanoTS;
3284	uint32_t u32SystemResolution;
3285	unsigned i;
3286
3287	LogFlow(("SUPR0GipMap: Resumes GIP updating\n"));
3288
3289	/*
3290	* Try bump up the system timer resolution.
3291	* The more interrupts the better...
3292	*/
3293	if ( RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
3294	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
3295	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1953125 /* 512 HZ */, &u32SystemResolution))
3296	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 2000000 /* 500 HZ */, &u32SystemResolution))
3297	)
3298	{
3299	Assert(RTTimerGetSystemGranularity() <= u32SystemResolution);
3300	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
3301	}
3302
3303	if (pGipR0->aCPUs[0].u32TransactionId != 2 /* not the first time */)
3304	{
3305	for (i = 0; i < RT_ELEMENTS(pGipR0->aCPUs); i++)
3306	ASMAtomicUoWriteU32(&pGipR0->aCPUs[i].u32TransactionId,
3307	(pGipR0->aCPUs[i].u32TransactionId + GIP_UPDATEHZ_RECALC_FREQ * 2)
3308	& ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
3309	ASMAtomicWriteU64(&pGipR0->u64NanoTSLastUpdateHz, 0);
3310	}
3311
3312	u64NanoTS = RTTimeSystemNanoTS() - pGipR0->u32UpdateIntervalNS;
3313	if ( pGipR0->u32Mode == SUPGIPMODE_SYNC_TSC
3314	\|\| RTMpGetOnlineCount() == 1)
3315	supdrvGipReInitCpu(&pGipR0->aCPUs[0], u64NanoTS);
3316	else
3317	RTMpOnAll(supdrvGipReInitCpuCallback, pGipR0, &u64NanoTS);
3318
3319	#ifndef DO_NOT_START_GIP
3320	rc = RTTimerStart(pDevExt->pGipTimer, 0); AssertRC(rc);
3321	#endif
3322	rc = VINF_SUCCESS;
3323	}
3324	}
3325	}
3326	else
3327	{
3328	rc = VERR_GENERAL_FAILURE;
3329	Log(("SUPR0GipMap: GIP is not available!\n"));
3330	}
3331	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3332	RTSemMutexRelease(pDevExt->mtxGip);
3333	#else
3334	RTSemFastMutexRelease(pDevExt->mtxGip);
3335	#endif
3336
3337	/*
3338	* Write returns.
3339	*/
3340	if (pHCPhysGip)
3341	*pHCPhysGip = HCPhys;
3342	if (ppGipR3)
3343	*ppGipR3 = pGipR3;
3344
3345	#ifdef DEBUG_DARWIN_GIP
3346	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3347	#else
3348	LogFlow(( "SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3349	#endif
3350	return rc;
3351	}
3352
3353
3354	/**
3355	* Unmaps any user mapping of the GIP and terminates all GIP access
3356	* from this session.
3357	*
3358	* @returns IPRT status code.
3359	* @param pSession Session to which the GIP mapping should belong.
3360	*/
3361	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
3362	{
3363	int rc = VINF_SUCCESS;
3364	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3365	#ifdef DEBUG_DARWIN_GIP
3366	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
3367	pSession,
3368	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
3369	pSession->GipMapObjR3));
3370	#else
3371	LogFlow(("SUPR0GipUnmap: pSession=%p\n", pSession));
3372	#endif
3373	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3374
3375	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3376	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3377	#else
3378	RTSemFastMutexRequest(pDevExt->mtxGip);
3379	#endif
3380
3381	/*
3382	* Unmap anything?
3383	*/
3384	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
3385	{
3386	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
3387	AssertRC(rc);
3388	if (RT_SUCCESS(rc))
3389	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
3390	}
3391
3392	/*
3393	* Dereference global GIP.
3394	*/
3395	if (pSession->fGipReferenced && !rc)
3396	{
3397	pSession->fGipReferenced = 0;
3398	if ( pDevExt->cGipUsers > 0
3399	&& !--pDevExt->cGipUsers)
3400	{
3401	LogFlow(("SUPR0GipUnmap: Suspends GIP updating\n"));
3402	#ifndef DO_NOT_START_GIP
3403	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = VINF_SUCCESS;
3404	#endif
3405
3406	if (pDevExt->u32SystemTimerGranularityGrant)
3407	{
3408	int rc2 = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant);
3409	AssertRC(rc2);
3410	pDevExt->u32SystemTimerGranularityGrant = 0;
3411	}
3412	}
3413	}
3414
3415	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3416	RTSemMutexRelease(pDevExt->mtxGip);
3417	#else
3418	RTSemFastMutexRelease(pDevExt->mtxGip);
3419	#endif
3420
3421	return rc;
3422	}
3423
3424
3425	/**
3426	* Gets the GIP pointer.
3427	*
3428	* @returns Pointer to the GIP or NULL.
3429	*/
3430	SUPDECL(PSUPGLOBALINFOPAGE) SUPGetGIP(void)
3431	{
3432	return g_pSUPGlobalInfoPage;
3433	}
3434
3435
3436	/**
3437	* Register a component factory with the support driver.
3438	*
3439	* This is currently restricted to kernel sessions only.
3440	*
3441	* @returns VBox status code.
3442	* @retval VINF_SUCCESS on success.
3443	* @retval VERR_NO_MEMORY if we're out of memory.
3444	* @retval VERR_ALREADY_EXISTS if the factory has already been registered.
3445	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3446	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3447	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3448	*
3449	* @param pSession The SUPDRV session (must be a ring-0 session).
3450	* @param pFactory Pointer to the component factory registration structure.
3451	*
3452	* @remarks This interface is also available via SUPR0IdcComponentRegisterFactory.
3453	*/
3454	SUPR0DECL(int) SUPR0ComponentRegisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3455	{
3456	PSUPDRVFACTORYREG pNewReg;
3457	const char *psz;
3458	int rc;
3459
3460	/*
3461	* Validate parameters.
3462	*/
3463	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3464	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3465	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3466	AssertPtrReturn(pFactory->pfnQueryFactoryInterface, VERR_INVALID_POINTER);
3467	psz = RTStrEnd(pFactory->szName, sizeof(pFactory->szName));
3468	AssertReturn(psz, VERR_INVALID_PARAMETER);
3469
3470	/*
3471	* Allocate and initialize a new registration structure.
3472	*/
3473	pNewReg = (PSUPDRVFACTORYREG)RTMemAlloc(sizeof(SUPDRVFACTORYREG));
3474	if (pNewReg)
3475	{
3476	pNewReg->pNext = NULL;
3477	pNewReg->pFactory = pFactory;
3478	pNewReg->pSession = pSession;
3479	pNewReg->cchName = psz - &pFactory->szName[0];
3480
3481	/*
3482	* Add it to the tail of the list after checking for prior registration.
3483	*/
3484	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3485	if (RT_SUCCESS(rc))
3486	{
3487	PSUPDRVFACTORYREG pPrev = NULL;
3488	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3489	while (pCur && pCur->pFactory != pFactory)
3490	{
3491	pPrev = pCur;
3492	pCur = pCur->pNext;
3493	}
3494	if (!pCur)
3495	{
3496	if (pPrev)
3497	pPrev->pNext = pNewReg;
3498	else
3499	pSession->pDevExt->pComponentFactoryHead = pNewReg;
3500	rc = VINF_SUCCESS;
3501	}
3502	else
3503	rc = VERR_ALREADY_EXISTS;
3504
3505	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3506	}
3507
3508	if (RT_FAILURE(rc))
3509	RTMemFree(pNewReg);
3510	}
3511	else
3512	rc = VERR_NO_MEMORY;
3513	return rc;
3514	}
3515
3516
3517	/**
3518	* Deregister a component factory.
3519	*
3520	* @returns VBox status code.
3521	* @retval VINF_SUCCESS on success.
3522	* @retval VERR_NOT_FOUND if the factory wasn't registered.
3523	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3524	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3525	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3526	*
3527	* @param pSession The SUPDRV session (must be a ring-0 session).
3528	* @param pFactory Pointer to the component factory registration structure
3529	* previously passed SUPR0ComponentRegisterFactory().
3530	*
3531	* @remarks This interface is also available via SUPR0IdcComponentDeregisterFactory.
3532	*/
3533	SUPR0DECL(int) SUPR0ComponentDeregisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3534	{
3535	int rc;
3536
3537	/*
3538	* Validate parameters.
3539	*/
3540	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3541	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3542	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3543
3544	/*
3545	* Take the lock and look for the registration record.
3546	*/
3547	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3548	if (RT_SUCCESS(rc))
3549	{
3550	PSUPDRVFACTORYREG pPrev = NULL;
3551	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3552	while (pCur && pCur->pFactory != pFactory)
3553	{
3554	pPrev = pCur;
3555	pCur = pCur->pNext;
3556	}
3557	if (pCur)
3558	{
3559	if (!pPrev)
3560	pSession->pDevExt->pComponentFactoryHead = pCur->pNext;
3561	else
3562	pPrev->pNext = pCur->pNext;
3563
3564	pCur->pNext = NULL;
3565	pCur->pFactory = NULL;
3566	pCur->pSession = NULL;
3567	rc = VINF_SUCCESS;
3568	}
3569	else
3570	rc = VERR_NOT_FOUND;
3571
3572	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3573
3574	RTMemFree(pCur);
3575	}
3576	return rc;
3577	}
3578
3579
3580	/**
3581	* Queries a component factory.
3582	*
3583	* @returns VBox status code.
3584	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3585	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3586	* @retval VERR_SUPDRV_COMPONENT_NOT_FOUND if the component factory wasn't found.
3587	* @retval VERR_SUPDRV_INTERFACE_NOT_SUPPORTED if the interface wasn't supported.
3588	*
3589	* @param pSession The SUPDRV session.
3590	* @param pszName The name of the component factory.
3591	* @param pszInterfaceUuid The UUID of the factory interface (stringified).
3592	* @param ppvFactoryIf Where to store the factory interface.
3593	*/
3594	SUPR0DECL(int) SUPR0ComponentQueryFactory(PSUPDRVSESSION pSession, const char pszName, const char pszInterfaceUuid, void **ppvFactoryIf)
3595	{
3596	const char *pszEnd;
3597	size_t cchName;
3598	int rc;
3599
3600	/*
3601	* Validate parameters.
3602	*/
3603	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3604
3605	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
3606	pszEnd = RTStrEnd(pszName, RT_SIZEOFMEMB(SUPDRVFACTORY, szName));
3607	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3608	cchName = pszEnd - pszName;
3609
3610	AssertPtrReturn(pszInterfaceUuid, VERR_INVALID_POINTER);
3611	pszEnd = RTStrEnd(pszInterfaceUuid, RTUUID_STR_LENGTH);
3612	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3613
3614	AssertPtrReturn(ppvFactoryIf, VERR_INVALID_POINTER);
3615	*ppvFactoryIf = NULL;
3616
3617	/*
3618	* Take the lock and try all factories by this name.
3619	*/
3620	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3621	if (RT_SUCCESS(rc))
3622	{
3623	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3624	rc = VERR_SUPDRV_COMPONENT_NOT_FOUND;
3625	while (pCur)
3626	{
3627	if ( pCur->cchName == cchName
3628	&& !memcmp(pCur->pFactory->szName, pszName, cchName))
3629	{
3630	void *pvFactory = pCur->pFactory->pfnQueryFactoryInterface(pCur->pFactory, pSession, pszInterfaceUuid);
3631	if (pvFactory)
3632	{
3633	*ppvFactoryIf = pvFactory;
3634	rc = VINF_SUCCESS;
3635	break;
3636	}
3637	rc = VERR_SUPDRV_INTERFACE_NOT_SUPPORTED;
3638	}
3639
3640	/* next */
3641	pCur = pCur->pNext;
3642	}
3643
3644	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3645	}
3646	return rc;
3647	}
3648
3649
3650	/**
3651	* Adds a memory object to the session.
3652	*
3653	* @returns IPRT status code.
3654	* @param pMem Memory tracking structure containing the
3655	* information to track.
3656	* @param pSession The session.
3657	*/
3658	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
3659	{
3660	PSUPDRVBUNDLE pBundle;
3661	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3662
3663	/*
3664	* Find free entry and record the allocation.
3665	*/
3666	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3667	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3668	{
3669	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
3670	{
3671	unsigned i;
3672	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3673	{
3674	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
3675	{
3676	pBundle->cUsed++;
3677	pBundle->aMem[i] = *pMem;
3678	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3679	return VINF_SUCCESS;
3680	}
3681	}
3682	AssertFailed(); /* !!this can't be happening!!! */
3683	}
3684	}
3685	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3686
3687	/*
3688	* Need to allocate a new bundle.
3689	* Insert into the last entry in the bundle.
3690	*/
3691	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
3692	if (!pBundle)
3693	return VERR_NO_MEMORY;
3694
3695	/* take last entry. */
3696	pBundle->cUsed++;
3697	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
3698
3699	/* insert into list. */
3700	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3701	pBundle->pNext = pSession->Bundle.pNext;
3702	pSession->Bundle.pNext = pBundle;
3703	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3704
3705	return VINF_SUCCESS;
3706	}
3707
3708
3709	/**
3710	* Releases a memory object referenced by pointer and type.
3711	*
3712	* @returns IPRT status code.
3713	* @param pSession Session data.
3714	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
3715	* @param eType Memory type.
3716	*/
3717	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
3718	{
3719	PSUPDRVBUNDLE pBundle;
3720	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3721
3722	/*
3723	* Validate input.
3724	*/
3725	if (!uPtr)
3726	{
3727	Log(("Illegal address %p\n", (void *)uPtr));
3728	return VERR_INVALID_PARAMETER;
3729	}
3730
3731	/*
3732	* Search for the address.
3733	*/
3734	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3735	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3736	{
3737	if (pBundle->cUsed > 0)
3738	{
3739	unsigned i;
3740	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3741	{
3742	if ( pBundle->aMem[i].eType == eType
3743	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3744	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
3745	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3746	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
3747	)
3748	{
3749	/* Make a copy of it and release it outside the spinlock. */
3750	SUPDRVMEMREF Mem = pBundle->aMem[i];
3751	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
3752	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
3753	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
3754	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3755
3756	if (Mem.MapObjR3 != NIL_RTR0MEMOBJ)
3757	{
3758	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
3759	AssertRC(rc); /** @todo figure out how to handle this. */
3760	}
3761	if (Mem.MemObj != NIL_RTR0MEMOBJ)
3762	{
3763	int rc = RTR0MemObjFree(Mem.MemObj, true /* fFreeMappings */);
3764	AssertRC(rc); /** @todo figure out how to handle this. */
3765	}
3766	return VINF_SUCCESS;
3767	}
3768	}
3769	}
3770	}
3771	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3772	Log(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
3773	return VERR_INVALID_PARAMETER;
3774	}
3775
3776
3777	/**
3778	* Opens an image. If it's the first time it's opened the call must upload
3779	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
3780	*
3781	* This is the 1st step of the loading.
3782	*
3783	* @returns IPRT status code.
3784	* @param pDevExt Device globals.
3785	* @param pSession Session data.
3786	* @param pReq The open request.
3787	*/
3788	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
3789	{
3790	int rc;
3791	PSUPDRVLDRIMAGE pImage;
3792	void *pv;
3793	size_t cchName = strlen(pReq->u.In.szName); /* (caller checked < 32). */
3794	LogFlow(("supdrvIOCtl_LdrOpen: szName=%s cbImageWithTabs=%d\n", pReq->u.In.szName, pReq->u.In.cbImageWithTabs));
3795
3796	/*
3797	* Check if we got an instance of the image already.
3798	*/
3799	supdrvLdrLock(pDevExt);
3800	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
3801	{
3802	if ( pImage->szName[cchName] == '\0'
3803	&& !memcmp(pImage->szName, pReq->u.In.szName, cchName))
3804	{
3805	/** @todo check cbImageBits and cbImageWithTabs here, if they differs that indicates that the images are different. */
3806	pImage->cUsage++;
3807	pReq->u.Out.pvImageBase = pImage->pvImage;
3808	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
3809	pReq->u.Out.fNativeLoader = pImage->fNative;
3810	supdrvLdrAddUsage(pSession, pImage);
3811	supdrvLdrUnlock(pDevExt);
3812	return VINF_SUCCESS;
3813	}
3814	}
3815	/* (not found - add it!) */
3816
3817	/*
3818	* Allocate memory.
3819	*/
3820	pv = RTMemAlloc(RT_OFFSETOF(SUPDRVLDRIMAGE, szName[cchName + 1]));
3821	if (!pv)
3822	{
3823	supdrvLdrUnlock(pDevExt);
3824	Log(("supdrvIOCtl_LdrOpen: RTMemAlloc() failed\n"));
3825	return /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_2;
3826	}
3827
3828	/*
3829	* Setup and link in the LDR stuff.
3830	*/
3831	pImage = (PSUPDRVLDRIMAGE)pv;
3832	pImage->pvImage = NULL;
3833	pImage->pvImageAlloc = NULL;
3834	pImage->cbImageWithTabs = pReq->u.In.cbImageWithTabs;
3835	pImage->cbImageBits = pReq->u.In.cbImageBits;
3836	pImage->cSymbols = 0;
3837	pImage->paSymbols = NULL;
3838	pImage->pachStrTab = NULL;
3839	pImage->cbStrTab = 0;
3840	pImage->pfnModuleInit = NULL;
3841	pImage->pfnModuleTerm = NULL;
3842	pImage->pfnServiceReqHandler = NULL;
3843	pImage->uState = SUP_IOCTL_LDR_OPEN;
3844	pImage->cUsage = 1;
3845	memcpy(pImage->szName, pReq->u.In.szName, cchName + 1);
3846
3847	/*
3848	* Try load it using the native loader, if that isn't supported, fall back
3849	* on the older method.
3850	*/
3851	pImage->fNative = true;
3852	rc = supdrvOSLdrOpen(pDevExt, pImage, pReq->u.In.szFilename);
3853	if (rc == VERR_NOT_SUPPORTED)
3854	{
3855	pImage->pvImageAlloc = RTMemExecAlloc(pImage->cbImageBits + 31);
3856	pImage->pvImage = RT_ALIGN_P(pImage->pvImageAlloc, 32);
3857	pImage->fNative = false;
3858	rc = pImage->pvImageAlloc ? VINF_SUCCESS : VERR_NO_EXEC_MEMORY;
3859	}
3860	if (RT_FAILURE(rc))
3861	{
3862	supdrvLdrUnlock(pDevExt);
3863	RTMemFree(pImage);
3864	Log(("supdrvIOCtl_LdrOpen(%s): failed - %Rrc\n", pReq->u.In.szName, rc));
3865	return rc;
3866	}
3867	Assert(VALID_PTR(pImage->pvImage) \|\| RT_FAILURE(rc));
3868
3869	/*
3870	* Link it.
3871	*/
3872	pImage->pNext = pDevExt->pLdrImages;
3873	pDevExt->pLdrImages = pImage;
3874
3875	supdrvLdrAddUsage(pSession, pImage);
3876
3877	pReq->u.Out.pvImageBase = pImage->pvImage;
3878	pReq->u.Out.fNeedsLoading = true;
3879	pReq->u.Out.fNativeLoader = pImage->fNative;
3880	supdrvLdrUnlock(pDevExt);
3881
3882	#if defined(RT_OS_WINDOWS) && defined(DEBUG)
3883	SUPR0Printf("VBoxDrv: windbg> .reload /f %s=%#p\n", pImage->szName, pImage->pvImage);
3884	#endif
3885	return VINF_SUCCESS;
3886	}
3887
3888
3889	/**
3890	* Worker that validates a pointer to an image entrypoint.
3891	*
3892	* @returns IPRT status code.
3893	* @param pDevExt The device globals.
3894	* @param pImage The loader image.
3895	* @param pv The pointer into the image.
3896	* @param fMayBeNull Whether it may be NULL.
3897	* @param pszWhat What is this entrypoint? (for logging)
3898	* @param pbImageBits The image bits prepared by ring-3.
3899	*
3900	* @remarks Will leave the lock on failure.
3901	*/
3902	static int supdrvLdrValidatePointer(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage, void *pv,
3903	bool fMayBeNull, const uint8_t pbImageBits, const char pszWhat)
3904	{
3905	if (!fMayBeNull \|\| pv)
3906	{
3907	if ((uintptr_t)pv - (uintptr_t)pImage->pvImage >= pImage->cbImageBits)
3908	{
3909	supdrvLdrUnlock(pDevExt);
3910	Log(("Out of range (%p LB %#x): %s=%p\n", pImage->pvImage, pImage->cbImageBits, pszWhat, pv));
3911	return VERR_INVALID_PARAMETER;
3912	}
3913
3914	if (pImage->fNative)
3915	{
3916	int rc = supdrvOSLdrValidatePointer(pDevExt, pImage, pv, pbImageBits);
3917	if (RT_FAILURE(rc))
3918	{
3919	supdrvLdrUnlock(pDevExt);
3920	Log(("Bad entry point address: %s=%p (rc=%Rrc)\n", pszWhat, pv, rc));
3921	return rc;
3922	}
3923	}
3924	}
3925	return VINF_SUCCESS;
3926	}
3927
3928
3929	/**
3930	* Loads the image bits.
3931	*
3932	* This is the 2nd step of the loading.
3933	*
3934	* @returns IPRT status code.
3935	* @param pDevExt Device globals.
3936	* @param pSession Session data.
3937	* @param pReq The request.
3938	*/
3939	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
3940	{
3941	PSUPDRVLDRUSAGE pUsage;
3942	PSUPDRVLDRIMAGE pImage;
3943	int rc;
3944	LogFlow(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImageWithBits=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImageWithTabs));
3945
3946	/*
3947	* Find the ldr image.
3948	*/
3949	supdrvLdrLock(pDevExt);
3950	pUsage = pSession->pLdrUsage;
3951	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3952	pUsage = pUsage->pNext;
3953	if (!pUsage)
3954	{
3955	supdrvLdrUnlock(pDevExt);
3956	Log(("SUP_IOCTL_LDR_LOAD: couldn't find image!\n"));
3957	return VERR_INVALID_HANDLE;
3958	}
3959	pImage = pUsage->pImage;
3960
3961	/*
3962	* Validate input.
3963	*/
3964	if ( pImage->cbImageWithTabs != pReq->u.In.cbImageWithTabs
3965	\|\| pImage->cbImageBits != pReq->u.In.cbImageBits)
3966	{
3967	supdrvLdrUnlock(pDevExt);
3968	Log(("SUP_IOCTL_LDR_LOAD: image size mismatch!! %d(prep) != %d(load) or %d != %d\n",
3969	pImage->cbImageWithTabs, pReq->u.In.cbImageWithTabs, pImage->cbImageBits, pReq->u.In.cbImageBits));
3970	return VERR_INVALID_HANDLE;
3971	}
3972
3973	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
3974	{
3975	unsigned uState = pImage->uState;
3976	supdrvLdrUnlock(pDevExt);
3977	if (uState != SUP_IOCTL_LDR_LOAD)
3978	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
3979	return VERR_ALREADY_LOADED;
3980	}
3981
3982	switch (pReq->u.In.eEPType)
3983	{
3984	case SUPLDRLOADEP_NOTHING:
3985	break;
3986
3987	case SUPLDRLOADEP_VMMR0:
3988	rc = supdrvLdrValidatePointer( pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0, false, pReq->u.In.abImage, "pvVMMR0");
3989	if (RT_SUCCESS(rc))
3990	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt, false, pReq->u.In.abImage, "pvVMMR0EntryInt");
3991	if (RT_SUCCESS(rc))
3992	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, false, pReq->u.In.abImage, "pvVMMR0EntryFast");
3993	if (RT_SUCCESS(rc))
3994	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx, false, pReq->u.In.abImage, "pvVMMR0EntryEx");
3995	if (RT_FAILURE(rc))
3996	return rc;
3997	break;
3998
3999	case SUPLDRLOADEP_SERVICE:
4000	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.Service.pfnServiceReq, false, pReq->u.In.abImage, "pfnServiceReq");
4001	if (RT_FAILURE(rc))
4002	return rc;
4003	if ( pReq->u.In.EP.Service.apvReserved[0] != NIL_RTR0PTR
4004	\|\| pReq->u.In.EP.Service.apvReserved[1] != NIL_RTR0PTR
4005	\|\| pReq->u.In.EP.Service.apvReserved[2] != NIL_RTR0PTR)
4006	{
4007	supdrvLdrUnlock(pDevExt);
4008	Log(("Out of range (%p LB %#x): apvReserved={%p,%p,%p} MBZ!\n",
4009	pImage->pvImage, pReq->u.In.cbImageWithTabs,
4010	pReq->u.In.EP.Service.apvReserved[0],
4011	pReq->u.In.EP.Service.apvReserved[1],
4012	pReq->u.In.EP.Service.apvReserved[2]));
4013	return VERR_INVALID_PARAMETER;
4014	}
4015	break;
4016
4017	default:
4018	supdrvLdrUnlock(pDevExt);
4019	Log(("Invalid eEPType=%d\n", pReq->u.In.eEPType));
4020	return VERR_INVALID_PARAMETER;
4021	}
4022
4023	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleInit, true, pReq->u.In.abImage, "pfnModuleInit");
4024	if (RT_FAILURE(rc))
4025	return rc;
4026	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleTerm, true, pReq->u.In.abImage, "pfnModuleTerm");
4027	if (RT_FAILURE(rc))
4028	return rc;
4029
4030	/*
4031	* Allocate and copy the tables.
4032	* (No need to do try/except as this is a buffered request.)
4033	*/
4034	pImage->cbStrTab = pReq->u.In.cbStrTab;
4035	if (pImage->cbStrTab)
4036	{
4037	pImage->pachStrTab = (char *)RTMemAlloc(pImage->cbStrTab);
4038	if (pImage->pachStrTab)
4039	memcpy(pImage->pachStrTab, &pReq->u.In.abImage[pReq->u.In.offStrTab], pImage->cbStrTab);
4040	else
4041	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_3;
4042	}
4043
4044	pImage->cSymbols = pReq->u.In.cSymbols;
4045	if (RT_SUCCESS(rc) && pImage->cSymbols)
4046	{
4047	size_t cbSymbols = pImage->cSymbols * sizeof(SUPLDRSYM);
4048	pImage->paSymbols = (PSUPLDRSYM)RTMemAlloc(cbSymbols);
4049	if (pImage->paSymbols)
4050	memcpy(pImage->paSymbols, &pReq->u.In.abImage[pReq->u.In.offSymbols], cbSymbols);
4051	else
4052	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_4;
4053	}
4054
4055	/*
4056	* Copy the bits / complete native loading.
4057	*/
4058	if (RT_SUCCESS(rc))
4059	{
4060	pImage->uState = SUP_IOCTL_LDR_LOAD;
4061	pImage->pfnModuleInit = pReq->u.In.pfnModuleInit;
4062	pImage->pfnModuleTerm = pReq->u.In.pfnModuleTerm;
4063
4064	if (pImage->fNative)
4065	rc = supdrvOSLdrLoad(pDevExt, pImage, pReq->u.In.abImage);
4066	else
4067	memcpy(pImage->pvImage, &pReq->u.In.abImage[0], pImage->cbImageBits);
4068	}
4069
4070	/*
4071	* Update any entry points.
4072	*/
4073	if (RT_SUCCESS(rc))
4074	{
4075	switch (pReq->u.In.eEPType)
4076	{
4077	default:
4078	case SUPLDRLOADEP_NOTHING:
4079	rc = VINF_SUCCESS;
4080	break;
4081	case SUPLDRLOADEP_VMMR0:
4082	rc = supdrvLdrSetVMMR0EPs(pDevExt, pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
4083	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
4084	break;
4085	case SUPLDRLOADEP_SERVICE:
4086	pImage->pfnServiceReqHandler = pReq->u.In.EP.Service.pfnServiceReq;
4087	rc = VINF_SUCCESS;
4088	break;
4089	}
4090	}
4091
4092	/*
4093	* On success call the module initialization.
4094	*/
4095	LogFlow(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
4096	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
4097	{
4098	Log(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
4099	rc = pImage->pfnModuleInit();
4100	if (rc && pDevExt->pvVMMR0 == pImage->pvImage)
4101	supdrvLdrUnsetVMMR0EPs(pDevExt);
4102	}
4103
4104	if (RT_FAILURE(rc))
4105	{
4106	pImage->uState = SUP_IOCTL_LDR_OPEN;
4107	pImage->pfnModuleInit = NULL;
4108	pImage->pfnModuleTerm = NULL;
4109	pImage->pfnServiceReqHandler= NULL;
4110	pImage->cbStrTab = 0;
4111	RTMemFree(pImage->pachStrTab);
4112	pImage->pachStrTab = NULL;
4113	RTMemFree(pImage->paSymbols);
4114	pImage->paSymbols = NULL;
4115	pImage->cSymbols = 0;
4116	}
4117
4118	supdrvLdrUnlock(pDevExt);
4119	return rc;
4120	}
4121
4122
4123	/**
4124	* Frees a previously loaded (prep'ed) image.
4125	*
4126	* @returns IPRT status code.
4127	* @param pDevExt Device globals.
4128	* @param pSession Session data.
4129	* @param pReq The request.
4130	*/
4131	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
4132	{
4133	int rc;
4134	PSUPDRVLDRUSAGE pUsagePrev;
4135	PSUPDRVLDRUSAGE pUsage;
4136	PSUPDRVLDRIMAGE pImage;
4137	LogFlow(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
4138
4139	/*
4140	* Find the ldr image.
4141	*/
4142	supdrvLdrLock(pDevExt);
4143	pUsagePrev = NULL;
4144	pUsage = pSession->pLdrUsage;
4145	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4146	{
4147	pUsagePrev = pUsage;
4148	pUsage = pUsage->pNext;
4149	}
4150	if (!pUsage)
4151	{
4152	supdrvLdrUnlock(pDevExt);
4153	Log(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
4154	return VERR_INVALID_HANDLE;
4155	}
4156
4157	/*
4158	* Check if we can remove anything.
4159	*/
4160	rc = VINF_SUCCESS;
4161	pImage = pUsage->pImage;
4162	if (pImage->cUsage <= 1 \|\| pUsage->cUsage <= 1)
4163	{
4164	/*
4165	* Check if there are any objects with destructors in the image, if
4166	* so leave it for the session cleanup routine so we get a chance to
4167	* clean things up in the right order and not leave them all dangling.
4168	*/
4169	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
4170	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
4171	if (pImage->cUsage <= 1)
4172	{
4173	PSUPDRVOBJ pObj;
4174	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4175	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4176	{
4177	rc = VERR_DANGLING_OBJECTS;
4178	break;
4179	}
4180	}
4181	else
4182	{
4183	PSUPDRVUSAGE pGenUsage;
4184	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
4185	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4186	{
4187	rc = VERR_DANGLING_OBJECTS;
4188	break;
4189	}
4190	}
4191	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
4192	if (rc == VINF_SUCCESS)
4193	{
4194	/* unlink it */
4195	if (pUsagePrev)
4196	pUsagePrev->pNext = pUsage->pNext;
4197	else
4198	pSession->pLdrUsage = pUsage->pNext;
4199
4200	/* free it */
4201	pUsage->pImage = NULL;
4202	pUsage->pNext = NULL;
4203	RTMemFree(pUsage);
4204
4205	/*
4206	* Dereference the image.
4207	*/
4208	if (pImage->cUsage <= 1)
4209	supdrvLdrFree(pDevExt, pImage);
4210	else
4211	pImage->cUsage--;
4212	}
4213	else
4214	{
4215	Log(("supdrvIOCtl_LdrFree: Dangling objects in %p/%s!\n", pImage->pvImage, pImage->szName));
4216	rc = VINF_SUCCESS; /** @todo BRANCH-2.1: remove this after branching. */
4217	}
4218	}
4219	else
4220	{
4221	/*
4222	* Dereference both image and usage.
4223	*/
4224	pImage->cUsage--;
4225	pUsage->cUsage--;
4226	}
4227
4228	supdrvLdrUnlock(pDevExt);
4229	return rc;
4230	}
4231
4232
4233	/**
4234	* Gets the address of a symbol in an open image.
4235	*
4236	* @returns IPRT status code.
4237	* @param pDevExt Device globals.
4238	* @param pSession Session data.
4239	* @param pReq The request buffer.
4240	*/
4241	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
4242	{
4243	PSUPDRVLDRIMAGE pImage;
4244	PSUPDRVLDRUSAGE pUsage;
4245	uint32_t i;
4246	PSUPLDRSYM paSyms;
4247	const char *pchStrings;
4248	const size_t cbSymbol = strlen(pReq->u.In.szSymbol) + 1;
4249	void *pvSymbol = NULL;
4250	int rc = VERR_GENERAL_FAILURE;
4251	Log3(("supdrvIOCtl_LdrGetSymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
4252
4253	/*
4254	* Find the ldr image.
4255	*/
4256	supdrvLdrLock(pDevExt);
4257	pUsage = pSession->pLdrUsage;
4258	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4259	pUsage = pUsage->pNext;
4260	if (!pUsage)
4261	{
4262	supdrvLdrUnlock(pDevExt);
4263	Log(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
4264	return VERR_INVALID_HANDLE;
4265	}
4266	pImage = pUsage->pImage;
4267	if (pImage->uState != SUP_IOCTL_LDR_LOAD)
4268	{
4269	unsigned uState = pImage->uState;
4270	supdrvLdrUnlock(pDevExt);
4271	Log(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", uState, uState)); NOREF(uState);
4272	return VERR_ALREADY_LOADED;
4273	}
4274
4275	/*
4276	* Search the symbol strings.
4277	*/
4278	pchStrings = pImage->pachStrTab;
4279	paSyms = pImage->paSymbols;
4280	for (i = 0; i < pImage->cSymbols; i++)
4281	{
4282	if ( paSyms[i].offSymbol < pImage->cbImageBits /* paranoia */
4283	&& paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4284	&& !memcmp(pchStrings + paSyms[i].offName, pReq->u.In.szSymbol, cbSymbol))
4285	{
4286	pvSymbol = (uint8_t *)pImage->pvImage + paSyms[i].offSymbol;
4287	rc = VINF_SUCCESS;
4288	break;
4289	}
4290	}
4291	supdrvLdrUnlock(pDevExt);
4292	pReq->u.Out.pvSymbol = pvSymbol;
4293	return rc;
4294	}
4295
4296
4297	/**
4298	* Gets the address of a symbol in an open image or the support driver.
4299	*
4300	* @returns VINF_SUCCESS on success.
4301	* @returns
4302	* @param pDevExt Device globals.
4303	* @param pSession Session data.
4304	* @param pReq The request buffer.
4305	*/
4306	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq)
4307	{
4308	int rc = VINF_SUCCESS;
4309	const char *pszSymbol = pReq->u.In.pszSymbol;
4310	const char *pszModule = pReq->u.In.pszModule;
4311	size_t cbSymbol;
4312	char const *pszEnd;
4313	uint32_t i;
4314
4315	/*
4316	* Input validation.
4317	*/
4318	AssertPtrReturn(pszSymbol, VERR_INVALID_POINTER);
4319	pszEnd = RTStrEnd(pszSymbol, 512);
4320	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4321	cbSymbol = pszEnd - pszSymbol + 1;
4322
4323	if (pszModule)
4324	{
4325	AssertPtrReturn(pszModule, VERR_INVALID_POINTER);
4326	pszEnd = RTStrEnd(pszModule, 64);
4327	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4328	}
4329	Log3(("supdrvIDC_LdrGetSymbol: pszModule=%p:{%s} pszSymbol=%p:{%s}\n", pszModule, pszModule, pszSymbol, pszSymbol));
4330
4331
4332	if ( !pszModule
4333	\|\| !strcmp(pszModule, "SupDrv"))
4334	{
4335	/*
4336	* Search the support driver export table.
4337	*/
4338	for (i = 0; i < RT_ELEMENTS(g_aFunctions); i++)
4339	if (!strcmp(g_aFunctions[i].szName, pszSymbol))
4340	{
4341	pReq->u.Out.pfnSymbol = g_aFunctions[i].pfn;
4342	break;
4343	}
4344	}
4345	else
4346	{
4347	/*
4348	* Find the loader image.
4349	*/
4350	PSUPDRVLDRIMAGE pImage;
4351
4352	supdrvLdrLock(pDevExt);
4353
4354	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
4355	if (!strcmp(pImage->szName, pszModule))
4356	break;
4357	if (pImage && pImage->uState == SUP_IOCTL_LDR_LOAD)
4358	{
4359	/*
4360	* Search the symbol strings.
4361	*/
4362	const char *pchStrings = pImage->pachStrTab;
4363	PCSUPLDRSYM paSyms = pImage->paSymbols;
4364	for (i = 0; i < pImage->cSymbols; i++)
4365	{
4366	if ( paSyms[i].offSymbol < pImage->cbImageBits /* paranoia */
4367	&& paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4368	&& !memcmp(pchStrings + paSyms[i].offName, pszSymbol, cbSymbol))
4369	{
4370	/*
4371	* Found it! Calc the symbol address and add a reference to the module.
4372	*/
4373	pReq->u.Out.pfnSymbol = (PFNRT)((uint8_t *)pImage->pvImage + paSyms[i].offSymbol);
4374	rc = supdrvLdrAddUsage(pSession, pImage);
4375	break;
4376	}
4377	}
4378	}
4379	else
4380	rc = pImage ? VERR_WRONG_ORDER : VERR_MODULE_NOT_FOUND;
4381
4382	supdrvLdrUnlock(pDevExt);
4383	}
4384	return rc;
4385	}
4386
4387
4388	/**
4389	* Updates the VMMR0 entry point pointers.
4390	*
4391	* @returns IPRT status code.
4392	* @param pDevExt Device globals.
4393	* @param pSession Session data.
4394	* @param pVMMR0 VMMR0 image handle.
4395	* @param pvVMMR0EntryInt VMMR0EntryInt address.
4396	* @param pvVMMR0EntryFast VMMR0EntryFast address.
4397	* @param pvVMMR0EntryEx VMMR0EntryEx address.
4398	* @remark Caller must own the loader mutex.
4399	*/
4400	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx)
4401	{
4402	int rc = VINF_SUCCESS;
4403	LogFlow(("supdrvLdrSetR0EP pvVMMR0=%p pvVMMR0EntryInt=%p\n", pvVMMR0, pvVMMR0EntryInt));
4404
4405
4406	/*
4407	* Check if not yet set.
4408	*/
4409	if (!pDevExt->pvVMMR0)
4410	{
4411	pDevExt->pvVMMR0 = pvVMMR0;
4412	pDevExt->pfnVMMR0EntryInt = pvVMMR0EntryInt;
4413	pDevExt->pfnVMMR0EntryFast = pvVMMR0EntryFast;
4414	pDevExt->pfnVMMR0EntryEx = pvVMMR0EntryEx;
4415	}
4416	else
4417	{
4418	/*
4419	* Return failure or success depending on whether the values match or not.
4420	*/
4421	if ( pDevExt->pvVMMR0 != pvVMMR0
4422	\|\| (void *)pDevExt->pfnVMMR0EntryInt != pvVMMR0EntryInt
4423	\|\| (void *)pDevExt->pfnVMMR0EntryFast != pvVMMR0EntryFast
4424	\|\| (void *)pDevExt->pfnVMMR0EntryEx != pvVMMR0EntryEx)
4425	{
4426	AssertMsgFailed(("SUP_IOCTL_LDR_SETR0EP: Already set pointing to a different module!\n"));
4427	rc = VERR_INVALID_PARAMETER;
4428	}
4429	}
4430	return rc;
4431	}
4432
4433
4434	/**
4435	* Unsets the VMMR0 entry point installed by supdrvLdrSetR0EP.
4436	*
4437	* @param pDevExt Device globals.
4438	*/
4439	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt)
4440	{
4441	pDevExt->pvVMMR0 = NULL;
4442	pDevExt->pfnVMMR0EntryInt = NULL;
4443	pDevExt->pfnVMMR0EntryFast = NULL;
4444	pDevExt->pfnVMMR0EntryEx = NULL;
4445	}
4446
4447
4448	/**
4449	* Adds a usage reference in the specified session of an image.
4450	*
4451	* Called while owning the loader semaphore.
4452	*
4453	* @returns VINF_SUCCESS on success and VERR_NO_MEMORY on failure.
4454	* @param pSession Session in question.
4455	* @param pImage Image which the session is using.
4456	*/
4457	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage)
4458	{
4459	PSUPDRVLDRUSAGE pUsage;
4460	LogFlow(("supdrvLdrAddUsage: pImage=%p\n", pImage));
4461
4462	/*
4463	* Referenced it already?
4464	*/
4465	pUsage = pSession->pLdrUsage;
4466	while (pUsage)
4467	{
4468	if (pUsage->pImage == pImage)
4469	{
4470	pUsage->cUsage++;
4471	return VINF_SUCCESS;
4472	}
4473	pUsage = pUsage->pNext;
4474	}
4475
4476	/*
4477	* Allocate new usage record.
4478	*/
4479	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
4480	AssertReturn(pUsage, /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_5);
4481	pUsage->cUsage = 1;
4482	pUsage->pImage = pImage;
4483	pUsage->pNext = pSession->pLdrUsage;
4484	pSession->pLdrUsage = pUsage;
4485	return VINF_SUCCESS;
4486	}
4487
4488
4489	/**
4490	* Frees a load image.
4491	*
4492	* @param pDevExt Pointer to device extension.
4493	* @param pImage Pointer to the image we're gonna free.
4494	* This image must exit!
4495	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
4496	*/
4497	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
4498	{
4499	PSUPDRVLDRIMAGE pImagePrev;
4500	LogFlow(("supdrvLdrFree: pImage=%p\n", pImage));
4501
4502	/* find it - arg. should've used doubly linked list. */
4503	Assert(pDevExt->pLdrImages);
4504	pImagePrev = NULL;
4505	if (pDevExt->pLdrImages != pImage)
4506	{
4507	pImagePrev = pDevExt->pLdrImages;
4508	while (pImagePrev->pNext != pImage)
4509	pImagePrev = pImagePrev->pNext;
4510	Assert(pImagePrev->pNext == pImage);
4511	}
4512
4513	/* unlink */
4514	if (pImagePrev)
4515	pImagePrev->pNext = pImage->pNext;
4516	else
4517	pDevExt->pLdrImages = pImage->pNext;
4518
4519	/* check if this is VMMR0.r0 unset its entry point pointers. */
4520	if (pDevExt->pvVMMR0 == pImage->pvImage)
4521	supdrvLdrUnsetVMMR0EPs(pDevExt);
4522
4523	/* check for objects with destructors in this image. (Shouldn't happen.) */
4524	if (pDevExt->pObjs)
4525	{
4526	unsigned cObjs = 0;
4527	PSUPDRVOBJ pObj;
4528	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
4529	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
4530	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4531	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4532	{
4533	pObj->pfnDestructor = NULL;
4534	cObjs++;
4535	}
4536	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
4537	if (cObjs)
4538	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
4539	}
4540
4541	/* call termination function if fully loaded. */
4542	if ( pImage->pfnModuleTerm
4543	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
4544	{
4545	LogFlow(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
4546	pImage->pfnModuleTerm();
4547	}
4548
4549	/* do native unload if appropriate. */
4550	if (pImage->fNative)
4551	supdrvOSLdrUnload(pDevExt, pImage);
4552
4553	/* free the image */
4554	pImage->cUsage = 0;
4555	pImage->pNext = 0;
4556	pImage->uState = SUP_IOCTL_LDR_FREE;
4557	RTMemExecFree(pImage->pvImageAlloc, pImage->cbImageBits + 31);
4558	pImage->pvImageAlloc = NULL;
4559	RTMemFree(pImage->pachStrTab);
4560	pImage->pachStrTab = NULL;
4561	RTMemFree(pImage->paSymbols);
4562	pImage->paSymbols = NULL;
4563	RTMemFree(pImage);
4564	}
4565
4566
4567	/**
4568	* Acquires the loader lock.
4569	*
4570	* @returns IPRT status code.
4571	* @param pDevExt The device extension.
4572	*/
4573	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt)
4574	{
4575	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4576	int rc = RTSemMutexRequest(pDevExt->mtxLdr, RT_INDEFINITE_WAIT);
4577	#else
4578	int rc = RTSemFastMutexRequest(pDevExt->mtxLdr);
4579	#endif
4580	AssertRC(rc);
4581	return rc;
4582	}
4583
4584
4585	/**
4586	* Releases the loader lock.
4587	*
4588	* @returns IPRT status code.
4589	* @param pDevExt The device extension.
4590	*/
4591	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt)
4592	{
4593	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4594	return RTSemMutexRelease(pDevExt->mtxLdr);
4595	#else
4596	return RTSemFastMutexRelease(pDevExt->mtxLdr);
4597	#endif
4598	}
4599
4600
4601	/**
4602	* Implements the service call request.
4603	*
4604	* @returns VBox status code.
4605	* @param pDevExt The device extension.
4606	* @param pSession The calling session.
4607	* @param pReq The request packet, valid.
4608	*/
4609	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq)
4610	{
4611	#if !defined(RT_OS_WINDOWS) \|\| defined(DEBUG)
4612	int rc;
4613
4614	/*
4615	* Find the module first in the module referenced by the calling session.
4616	*/
4617	rc = supdrvLdrLock(pDevExt);
4618	if (RT_SUCCESS(rc))
4619	{
4620	PFNSUPR0SERVICEREQHANDLER pfnServiceReqHandler = NULL;
4621	PSUPDRVLDRUSAGE pUsage;
4622
4623	for (pUsage = pSession->pLdrUsage; pUsage; pUsage = pUsage->pNext)
4624	if ( pUsage->pImage->pfnServiceReqHandler
4625	&& !strcmp(pUsage->pImage->szName, pReq->u.In.szName))
4626	{
4627	pfnServiceReqHandler = pUsage->pImage->pfnServiceReqHandler;
4628	break;
4629	}
4630	supdrvLdrUnlock(pDevExt);
4631
4632	if (pfnServiceReqHandler)
4633	{
4634	/*
4635	* Call it.
4636	*/
4637	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
4638	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, NULL);
4639	else
4640	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0]);
4641	}
4642	else
4643	rc = VERR_SUPDRV_SERVICE_NOT_FOUND;
4644	}
4645
4646	/* log it */
4647	if ( RT_FAILURE(rc)
4648	&& rc != VERR_INTERRUPTED
4649	&& rc != VERR_TIMEOUT)
4650	Log(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4651	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4652	else
4653	Log4(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4654	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4655	return rc;
4656	#else /* RT_OS_WINDOWS && !DEBUG */
4657	return VERR_NOT_IMPLEMENTED;
4658	#endif /* RT_OS_WINDOWS && !DEBUG */
4659	}
4660
4661
4662	/**
4663	* Implements the logger settings request.
4664	*
4665	* @returns VBox status code.
4666	* @param pDevExt The device extension.
4667	* @param pSession The caller's session.
4668	* @param pReq The request.
4669	*/
4670	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq)
4671	{
4672	const char *pszGroup = &pReq->u.In.szStrings[pReq->u.In.offGroups];
4673	const char *pszFlags = &pReq->u.In.szStrings[pReq->u.In.offFlags];
4674	const char *pszDest = &pReq->u.In.szStrings[pReq->u.In.offDestination];
4675	PRTLOGGER pLogger = NULL;
4676	int rc;
4677
4678	/*
4679	* Some further validation.
4680	*/
4681	switch (pReq->u.In.fWhat)
4682	{
4683	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4684	case SUPLOGGERSETTINGS_WHAT_CREATE:
4685	break;
4686
4687	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4688	if (pszGroup \|\| pszFlags \|\| *pszDest)
4689	return VERR_INVALID_PARAMETER;
4690	if (pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_RELEASE)
4691	return VERR_ACCESS_DENIED;
4692	break;
4693
4694	default:
4695	return VERR_INTERNAL_ERROR;
4696	}
4697
4698	/*
4699	* Get the logger.
4700	*/
4701	switch (pReq->u.In.fWhich)
4702	{
4703	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4704	pLogger = RTLogGetDefaultInstance();
4705	break;
4706
4707	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4708	pLogger = RTLogRelDefaultInstance();
4709	break;
4710
4711	default:
4712	return VERR_INTERNAL_ERROR;
4713	}
4714
4715	/*
4716	* Do the job.
4717	*/
4718	switch (pReq->u.In.fWhat)
4719	{
4720	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4721	if (pLogger)
4722	{
4723	rc = RTLogFlags(pLogger, pszFlags);
4724	if (RT_SUCCESS(rc))
4725	rc = RTLogGroupSettings(pLogger, pszGroup);
4726	NOREF(pszDest);
4727	}
4728	else
4729	rc = VERR_NOT_FOUND;
4730	break;
4731
4732	case SUPLOGGERSETTINGS_WHAT_CREATE:
4733	{
4734	if (pLogger)
4735	rc = VERR_ALREADY_EXISTS;
4736	else
4737	{
4738	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
4739
4740	rc = RTLogCreate(&pLogger,
4741	0 /* fFlags */,
4742	pszGroup,
4743	pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_DEBUG
4744	? "VBOX_LOG"
4745	: "VBOX_RELEASE_LOG",
4746	RT_ELEMENTS(s_apszGroups),
4747	s_apszGroups,
4748	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER,
4749	NULL);
4750	if (RT_SUCCESS(rc))
4751	{
4752	rc = RTLogFlags(pLogger, pszFlags);
4753	NOREF(pszDest);
4754	if (RT_SUCCESS(rc))
4755	{
4756	switch (pReq->u.In.fWhich)
4757	{
4758	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4759	pLogger = RTLogSetDefaultInstance(pLogger);
4760	break;
4761	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4762	pLogger = RTLogRelSetDefaultInstance(pLogger);
4763	break;
4764	}
4765	}
4766	RTLogDestroy(pLogger);
4767	}
4768	}
4769	break;
4770	}
4771
4772	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4773	switch (pReq->u.In.fWhich)
4774	{
4775	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4776	pLogger = RTLogSetDefaultInstance(NULL);
4777	break;
4778	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4779	pLogger = RTLogRelSetDefaultInstance(NULL);
4780	break;
4781	}
4782	rc = RTLogDestroy(pLogger);
4783	break;
4784
4785	default:
4786	{
4787	rc = VERR_INTERNAL_ERROR;
4788	break;
4789	}
4790	}
4791
4792	return rc;
4793	}
4794
4795
4796	/**
4797	* Creates the GIP.
4798	*
4799	* @returns VBox status code.
4800	* @param pDevExt Instance data. GIP stuff may be updated.
4801	*/
4802	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
4803	{
4804	PSUPGLOBALINFOPAGE pGip;
4805	RTHCPHYS HCPhysGip;
4806	uint32_t u32SystemResolution;
4807	uint32_t u32Interval;
4808	unsigned cCpus;
4809	int rc;
4810
4811
4812	LogFlow(("supdrvGipCreate:\n"));
4813
4814	/* assert order */
4815	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
4816	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
4817	Assert(!pDevExt->pGipTimer);
4818
4819	/*
4820	* Check the CPU count.
4821	*/
4822	cCpus = RTMpGetArraySize();
4823	if ( cCpus > RTCPUSET_MAX_CPUS
4824	\|\| cCpus > 256 /ApicId is used for the mappings/)
4825	{
4826	SUPR0Printf("VBoxDrv: Too many CPUs (%u) for the GIP (max %u)\n", cCpus, RT_MIN(RTCPUSET_MAX_CPUS, 256));
4827	return VERR_TOO_MANY_CPUS;
4828	}
4829
4830	/*
4831	* Allocate a contiguous set of pages with a default kernel mapping.
4832	*/
4833	rc = RTR0MemObjAllocCont(&pDevExt->GipMemObj, RT_UOFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), false /fExecutable/);
4834	if (RT_FAILURE(rc))
4835	{
4836	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
4837	return rc;
4838	}
4839	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
4840	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
4841
4842	/*
4843	* Find a reasonable update interval and initialize the structure.
4844	*/
4845	u32Interval = u32SystemResolution = RTTimerGetSystemGranularity();
4846	while (u32Interval < 10000000 /* 10 ms */)
4847	u32Interval += u32SystemResolution;
4848
4849	supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), 1000000000 / u32Interval /=Hz/, cCpus);
4850
4851	/*
4852	* Create the timer.
4853	* If CPU_ALL isn't supported we'll have to fall back to synchronous mode.
4854	*/
4855	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
4856	{
4857	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, RTTIMER_FLAGS_CPU_ALL, supdrvGipAsyncTimer, pDevExt);
4858	if (rc == VERR_NOT_SUPPORTED)
4859	{
4860	OSDBGPRINT(("supdrvGipCreate: omni timer not supported, falling back to synchronous mode\n"));
4861	pGip->u32Mode = SUPGIPMODE_SYNC_TSC;
4862	}
4863	}
4864	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
4865	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0, supdrvGipSyncTimer, pDevExt);
4866	if (RT_SUCCESS(rc))
4867	{
4868	rc = RTMpNotificationRegister(supdrvGipMpEvent, pDevExt);
4869	if (RT_SUCCESS(rc))
4870	{
4871	rc = RTMpOnAll(supdrvGipInitOnCpu, pDevExt, pGip);
4872	if (RT_SUCCESS(rc))
4873	{
4874	/*
4875	* We're good.
4876	*/
4877	Log(("supdrvGipCreate: %u ns interval.\n", u32Interval));
4878	g_pSUPGlobalInfoPage = pGip;
4879	return VINF_SUCCESS;
4880	}
4881
4882	OSDBGPRINT(("supdrvGipCreate: RTMpOnAll failed with rc=%Rrc\n", rc));
4883	RTMpNotificationDeregister(supdrvGipMpEvent, pDevExt);
4884
4885	}
4886	else
4887	OSDBGPRINT(("supdrvGipCreate: failed to register MP event notfication. rc=%Rrc\n", rc));
4888	}
4889	else
4890	{
4891	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %u ns interval. rc=%Rrc\n", u32Interval, rc));
4892	Assert(!pDevExt->pGipTimer);
4893	}
4894	supdrvGipDestroy(pDevExt);
4895	return rc;
4896	}
4897
4898
4899	/**
4900	* Terminates the GIP.
4901	*
4902	* @param pDevExt Instance data. GIP stuff may be updated.
4903	*/
4904	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
4905	{
4906	int rc;
4907	#ifdef DEBUG_DARWIN_GIP
4908	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
4909	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
4910	pDevExt->pGipTimer, pDevExt->GipMemObj));
4911	#endif
4912
4913	/*
4914	* Invalid the GIP data.
4915	*/
4916	if (pDevExt->pGip)
4917	{
4918	supdrvGipTerm(pDevExt->pGip);
4919	pDevExt->pGip = NULL;
4920	}
4921	g_pSUPGlobalInfoPage = NULL;
4922
4923	/*
4924	* Destroy the timer and free the GIP memory object.
4925	*/
4926	if (pDevExt->pGipTimer)
4927	{
4928	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
4929	pDevExt->pGipTimer = NULL;
4930	}
4931
4932	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
4933	{
4934	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
4935	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
4936	}
4937
4938	/*
4939	* Finally, make sure we've release the system timer resolution request
4940	* if one actually succeeded and is still pending.
4941	*/
4942	if (pDevExt->u32SystemTimerGranularityGrant)
4943	{
4944	rc = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
4945	pDevExt->u32SystemTimerGranularityGrant = 0;
4946	}
4947	}
4948
4949
4950	/**
4951	* Timer callback function sync GIP mode.
4952	* @param pTimer The timer.
4953	* @param pvUser The device extension.
4954	*/
4955	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
4956	{
4957	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
4958	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4959	uint64_t u64TSC = ASMReadTSC();
4960	uint64_t NanoTS = RTTimeSystemNanoTS();
4961
4962	supdrvGipUpdate(pDevExt->pGip, NanoTS, u64TSC, NIL_RTCPUID, iTick);
4963
4964	ASMSetFlags(fOldFlags);
4965	}
4966
4967
4968	/**
4969	* Timer callback function for async GIP mode.
4970	* @param pTimer The timer.
4971	* @param pvUser The device extension.
4972	*/
4973	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
4974	{
4975	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
4976	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4977	RTCPUID idCpu = RTMpCpuId();
4978	uint64_t u64TSC = ASMReadTSC();
4979	uint64_t NanoTS = RTTimeSystemNanoTS();
4980
4981	/** @todo reset the transaction number and whatnot when iTick == 1. */
4982	if (pDevExt->idGipMaster == idCpu)
4983	supdrvGipUpdate(pDevExt->pGip, NanoTS, u64TSC, idCpu, iTick);
4984	else
4985	supdrvGipUpdatePerCpu(pDevExt->pGip, NanoTS, u64TSC, idCpu, ASMGetApicId(), iTick);
4986
4987	ASMSetFlags(fOldFlags);
4988	}
4989
4990
4991	/**
4992	* The calling CPU should be accounted as online, update GIP accordingly.
4993	*
4994	* This is used by supdrvGipMpEvent as well as the supdrvGipCreate.
4995	*
4996	* @param pGip The GIP.
4997	* @param idCpu The CPU ID.
4998	*/
4999	static void supdrvGipMpEventOnline(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
5000	{
5001	int iCpuSet;
5002	uint8_t idApic;
5003	uint32_t i;
5004
5005	Assert(idCpu == RTMpCpuId());
5006	Assert(pGip->cPossibleCpus == RTMpGetCount());
5007
5008	/*
5009	* Update the globals.
5010	*/
5011	ASMAtomicWriteU16(&pGip->cPresentCpus, RTMpGetPresentCount());
5012	ASMAtomicWriteU16(&pGip->cOnlineCpus, RTMpGetOnlineCount());
5013	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5014	if (iCpuSet >= 0)
5015	{
5016	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5017	RTCpuSetAddByIndex(&pGip->OnlineCpuSet, iCpuSet);
5018	RTCpuSetAddByIndex(&pGip->PresentCpuSet, iCpuSet);
5019	}
5020
5021	/*
5022	* Find our entry, or allocate one if not found.
5023	* ASSUMES that CPU IDs are constant.
5024	*/
5025	for (i = 0; i < pGip->cCpus; i++)
5026	if (pGip->aCPUs[i].idCpu == idCpu)
5027	break;
5028
5029	if (i >= pGip->cCpus)
5030	for (i = 0; i < pGip->cCpus; i++)
5031	{
5032	bool fRc;
5033	ASMAtomicCmpXchgSize(&pGip->aCPUs[i].idCpu, idCpu, NIL_RTCPUID, fRc);
5034	if (fRc)
5035	break;
5036	}
5037
5038	AssertReturnVoid(i < pGip->cCpus);
5039
5040	/*
5041	* Update the entry.
5042	*/
5043	idApic = ASMGetApicId();
5044	ASMAtomicUoWriteU16(&pGip->aCPUs[i].idApic, idApic);
5045	ASMAtomicUoWriteS16(&pGip->aCPUs[i].iCpuSet, (int16_t)iCpuSet);
5046	ASMAtomicUoWriteSize(&pGip->aCPUs[i].idCpu, idCpu);
5047	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_ONLINE);
5048
5049	/*
5050	* Update the APIC ID and CPU set index mappings.
5051	*/
5052	ASMAtomicWriteU16(&pGip->aiCpuFromApicId[idApic], i);
5053	ASMAtomicWriteU16(&pGip->aiCpuFromCpuSetIdx[iCpuSet], i);
5054	}
5055
5056
5057	/**
5058	* The CPU should be accounted as offline, update the GIP accordingly.
5059	*
5060	* This is used by supdrvGipMpEvent.
5061	*
5062	* @param pGip The GIP.
5063	* @param idCpu The CPU ID.
5064	*/
5065	static void supdrvGipMpEventOffline(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
5066	{
5067	int iCpuSet;
5068	unsigned i;
5069
5070	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5071	AssertReturnVoid(iCpuSet >= 0);
5072
5073	i = pGip->aiCpuFromCpuSetIdx[iCpuSet];
5074	AssertReturnVoid(i < pGip->cCpus);
5075	AssertReturnVoid(pGip->aCPUs[i].idCpu == idCpu);
5076
5077	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5078	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_OFFLINE);
5079	RTCpuSetDelByIndex(&pGip->OnlineCpuSet, iCpuSet);
5080	}
5081
5082
5083	/**
5084	* Multiprocessor event notification callback.
5085	*
5086	* This is used to make sure that the GIP master gets passed on to
5087	* another CPU. It also updates the associated CPU data.
5088	*
5089	* @param enmEvent The event.
5090	* @param idCpu The cpu it applies to.
5091	* @param pvUser Pointer to the device extension.
5092	*/
5093	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser)
5094	{
5095	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5096	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5097
5098	/*
5099	* Update the GIP CPU data.
5100	*/
5101	if (pGip)
5102	{
5103	switch (enmEvent)
5104	{
5105	case RTMPEVENT_ONLINE:
5106	supdrvGipMpEventOnline(pGip, idCpu);
5107	break;
5108	case RTMPEVENT_OFFLINE:
5109	supdrvGipMpEventOffline(pGip, idCpu);
5110	break;
5111
5112	}
5113	}
5114
5115	/*
5116	* Make sure there is a master GIP.
5117	*/
5118	if (enmEvent == RTMPEVENT_OFFLINE)
5119	{
5120	RTCPUID idGipMaster;
5121	ASMAtomicReadSize(&pDevExt->idGipMaster, &idGipMaster);
5122	if (idGipMaster == idCpu)
5123	{
5124	/*
5125	* Find a new GIP master.
5126	*/
5127	bool fIgnored;
5128	unsigned i;
5129	RTCPUID idNewGipMaster = NIL_RTCPUID;
5130	RTCPUSET OnlineCpus;
5131	RTMpGetOnlineSet(&OnlineCpus);
5132
5133	for (i = 0; i < RTCPUSET_MAX_CPUS; i++)
5134	{
5135	RTCPUID idCurCpu = RTMpCpuIdFromSetIndex(i);
5136	if ( RTCpuSetIsMember(&OnlineCpus, idCurCpu)
5137	&& idCurCpu != idGipMaster)
5138	{
5139	idNewGipMaster = idCurCpu;
5140	break;
5141	}
5142	}
5143
5144	Log(("supdrvGipMpEvent: Gip master %#lx -> %#lx\n", (long)idGipMaster, (long)idNewGipMaster));
5145	ASMAtomicCmpXchgSize(&pDevExt->idGipMaster, idNewGipMaster, idGipMaster, fIgnored);
5146	NOREF(fIgnored);
5147	}
5148	}
5149	}
5150
5151
5152	/**
5153	* Callback used by supdrvDetermineAsyncTSC to read the TSC on a CPU.
5154	*
5155	* @param idCpu Ignored.
5156	* @param pvUser1 Where to put the TSC.
5157	* @param pvUser2 Ignored.
5158	*/
5159	static DECLCALLBACK(void) supdrvDetermineAsyncTscWorker(RTCPUID idCpu, void pvUser1, void pvUser2)
5160	{
5161	#if 1
5162	ASMAtomicWriteU64((uint64_t volatile *)pvUser1, ASMReadTSC());
5163	#else
5164	(uint64_t )pvUser1 = ASMReadTSC();
5165	#endif
5166	}
5167
5168
5169	/**
5170	* Determine if Async GIP mode is required because of TSC drift.
5171	*
5172	* When using the default/normal timer code it is essential that the time stamp counter
5173	* (TSC) runs never backwards, that is, a read operation to the counter should return
5174	* a bigger value than any previous read operation. This is guaranteed by the latest
5175	* AMD CPUs and by newer Intel CPUs which never enter the C2 state (P4). In any other
5176	* case we have to choose the asynchronous timer mode.
5177	*
5178	* @param poffMin Pointer to the determined difference between different cores.
5179	* @return false if the time stamp counters appear to be synchronized, true otherwise.
5180	*/
5181	static bool supdrvDetermineAsyncTsc(uint64_t *poffMin)
5182	{
5183	/*
5184	* Just iterate all the cpus 8 times and make sure that the TSC is
5185	* ever increasing. We don't bother taking TSC rollover into account.
5186	*/
5187	int iEndCpu = RTMpGetArraySize();
5188	int iCpu;
5189	int cLoops = 8;
5190	bool fAsync = false;
5191	int rc = VINF_SUCCESS;
5192	uint64_t offMax = 0;
5193	uint64_t offMin = ~(uint64_t)0;
5194	uint64_t PrevTsc = ASMReadTSC();
5195
5196	while (cLoops-- > 0)
5197	{
5198	for (iCpu = 0; iCpu < iEndCpu; iCpu++)
5199	{
5200	uint64_t CurTsc;
5201	rc = RTMpOnSpecific(RTMpCpuIdFromSetIndex(iCpu), supdrvDetermineAsyncTscWorker, &CurTsc, NULL);
5202	if (RT_SUCCESS(rc))
5203	{
5204	if (CurTsc <= PrevTsc)
5205	{
5206	fAsync = true;
5207	offMin = offMax = PrevTsc - CurTsc;
5208	Log(("supdrvDetermineAsyncTsc: iCpu=%d cLoops=%d CurTsc=%llx PrevTsc=%llx\n",
5209	iCpu, cLoops, CurTsc, PrevTsc));
5210	break;
5211	}
5212
5213	/* Gather statistics (except the first time). */
5214	if (iCpu != 0 \|\| cLoops != 7)
5215	{
5216	uint64_t off = CurTsc - PrevTsc;
5217	if (off < offMin)
5218	offMin = off;
5219	if (off > offMax)
5220	offMax = off;
5221	Log2(("%d/%d: off=%llx\n", cLoops, iCpu, off));
5222	}
5223
5224	/* Next */
5225	PrevTsc = CurTsc;
5226	}
5227	else if (rc == VERR_NOT_SUPPORTED)
5228	break;
5229	else
5230	AssertMsg(rc == VERR_CPU_NOT_FOUND \|\| rc == VERR_CPU_OFFLINE, ("%d\n", rc));
5231	}
5232
5233	/* broke out of the loop. */
5234	if (iCpu < iEndCpu)
5235	break;
5236	}
5237
5238	poffMin = offMin; / Almost RTMpOnSpecific profiling. */
5239	Log(("supdrvDetermineAsyncTsc: returns %d; iEndCpu=%d rc=%d offMin=%llx offMax=%llx\n",
5240	fAsync, iEndCpu, rc, offMin, offMax));
5241	#if !defined(RT_OS_SOLARIS) && !defined(RT_OS_OS2) && !defined(RT_OS_WINDOWS)
5242	OSDBGPRINT(("vboxdrv: fAsync=%d offMin=%#lx offMax=%#lx\n", fAsync, (long)offMin, (long)offMax));
5243	#endif
5244	return fAsync;
5245	}
5246
5247
5248	/**
5249	* Determine the GIP TSC mode.
5250	*
5251	* @returns The most suitable TSC mode.
5252	* @param pDevExt Pointer to the device instance data.
5253	*/
5254	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt)
5255	{
5256	/*
5257	* On SMP we're faced with two problems:
5258	* (1) There might be a skew between the CPU, so that cpu0
5259	* returns a TSC that is slightly different from cpu1.
5260	* (2) Power management (and other things) may cause the TSC
5261	* to run at a non-constant speed, and cause the speed
5262	* to be different on the cpus. This will result in (1).
5263	*
5264	* So, on SMP systems we'll have to select the ASYNC update method
5265	* if there are symptoms of these problems.
5266	*/
5267	if (RTMpGetCount() > 1)
5268	{
5269	uint32_t uEAX, uEBX, uECX, uEDX;
5270	uint64_t u64DiffCoresIgnored;
5271
5272	/* Permit the user and/or the OS specific bits to force async mode. */
5273	if (supdrvOSGetForcedAsyncTscMode(pDevExt))
5274	return SUPGIPMODE_ASYNC_TSC;
5275
5276	/* Try check for current differences between the cpus. */
5277	if (supdrvDetermineAsyncTsc(&u64DiffCoresIgnored))
5278	return SUPGIPMODE_ASYNC_TSC;
5279
5280	/*
5281	* If the CPU supports power management and is an AMD one we
5282	* won't trust it unless it has the TscInvariant bit is set.
5283	*/
5284	/* Check for "AuthenticAMD" */
5285	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
5286	if ( uEAX >= 1
5287	&& uEBX == X86_CPUID_VENDOR_AMD_EBX
5288	&& uECX == X86_CPUID_VENDOR_AMD_ECX
5289	&& uEDX == X86_CPUID_VENDOR_AMD_EDX)
5290	{
5291	/* Check for APM support and that TscInvariant is cleared. */
5292	ASMCpuId(0x80000000, &uEAX, &uEBX, &uECX, &uEDX);
5293	if (uEAX >= 0x80000007)
5294	{
5295	ASMCpuId(0x80000007, &uEAX, &uEBX, &uECX, &uEDX);
5296	if ( !(uEDX & RT_BIT(8))/* TscInvariant */
5297	&& (uEDX & 0x3e)) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
5298	return SUPGIPMODE_ASYNC_TSC;
5299	}
5300	}
5301	}
5302	return SUPGIPMODE_SYNC_TSC;
5303	}
5304
5305
5306
5307	/**
5308	* Initializes the GIP data.
5309	*
5310	* @param pDevExt Pointer to the device instance data.
5311	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5312	* @param HCPhys The physical address of the GIP.
5313	* @param u64NanoTS The current nanosecond timestamp.
5314	* @param uUpdateHz The update frequency.
5315	* @param cCpus The CPU count.
5316	*/
5317	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
5318	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned cCpus)
5319	{
5320	size_t const cbGip = RT_ALIGN_Z(RT_OFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), PAGE_SIZE);
5321	unsigned i;
5322	#ifdef DEBUG_DARWIN_GIP
5323	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5324	#else
5325	LogFlow(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5326	#endif
5327
5328	/*
5329	* Initialize the structure.
5330	*/
5331	memset(pGip, 0, cbGip);
5332	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
5333	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
5334	pGip->u32Mode = supdrvGipDeterminTscMode(pDevExt);
5335	pGip->cCpus = (uint16_t)cCpus;
5336	pGip->cPages = (uint16_t)(cbGip / PAGE_SIZE);
5337	pGip->u32UpdateHz = uUpdateHz;
5338	pGip->u32UpdateIntervalNS = 1000000000 / uUpdateHz;
5339	pGip->u64NanoTSLastUpdateHz = u64NanoTS;
5340	RTCpuSetEmpty(&pGip->OnlineCpuSet);
5341	RTCpuSetEmpty(&pGip->PresentCpuSet);
5342	RTMpGetSet(&pGip->PossibleCpuSet);
5343	pGip->cOnlineCpus = RTMpGetOnlineCount();
5344	pGip->cPresentCpus = RTMpGetPresentCount();
5345	pGip->cPossibleCpus = RTMpGetCount();
5346	pGip->idCpuMax = RTMpGetMaxCpuId();
5347	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromApicId); i++)
5348	pGip->aiCpuFromApicId[i] = 0;
5349	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx); i++)
5350	pGip->aiCpuFromCpuSetIdx[i] = UINT16_MAX;
5351
5352	for (i = 0; i < cCpus; i++)
5353	{
5354	pGip->aCPUs[i].u32TransactionId = 2;
5355	pGip->aCPUs[i].u64NanoTS = u64NanoTS;
5356	pGip->aCPUs[i].u64TSC = ASMReadTSC();
5357
5358	pGip->aCPUs[i].enmState = SUPGIPCPUSTATE_INVALID;
5359	pGip->aCPUs[i].idCpu = NIL_RTCPUID;
5360	pGip->aCPUs[i].iCpuSet = -1;
5361	pGip->aCPUs[i].idApic = UINT8_MAX;
5362
5363	/*
5364	* We don't know the following values until we've executed updates.
5365	* So, we'll just pretend it's a 4 GHz CPU and adjust the history it on
5366	* the 2nd timer callout.
5367	*/
5368	pGip->aCPUs[i].u64CpuHz = _4G + 1; /* tstGIP-2 depends on this. */
5369	pGip->aCPUs[i].u32UpdateIntervalTSC
5370	= pGip->aCPUs[i].au32TSCHistory[0]
5371	= pGip->aCPUs[i].au32TSCHistory[1]
5372	= pGip->aCPUs[i].au32TSCHistory[2]
5373	= pGip->aCPUs[i].au32TSCHistory[3]
5374	= pGip->aCPUs[i].au32TSCHistory[4]
5375	= pGip->aCPUs[i].au32TSCHistory[5]
5376	= pGip->aCPUs[i].au32TSCHistory[6]
5377	= pGip->aCPUs[i].au32TSCHistory[7]
5378	= /pGip->aCPUs[i].u64CpuHz/ (uint32_t)(_4G / uUpdateHz);
5379	}
5380
5381	/*
5382	* Link it to the device extension.
5383	*/
5384	pDevExt->pGip = pGip;
5385	pDevExt->HCPhysGip = HCPhys;
5386	pDevExt->cGipUsers = 0;
5387	}
5388
5389
5390	/**
5391	* On CPU initialization callback for RTMpOnAll.
5392	*
5393	* @param idCpu The CPU ID.
5394	* @param pvUser1 The device extension.
5395	* @param pvUser2 The GIP.
5396	*/
5397	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2)
5398	{
5399	/* This is good enough, even though it will update some of the globals a
5400	bit to much. */
5401	supdrvGipMpEventOnline((PSUPGLOBALINFOPAGE)pvUser2, idCpu);
5402	}
5403
5404
5405	/**
5406	* Invalidates the GIP data upon termination.
5407	*
5408	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5409	*/
5410	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
5411	{
5412	unsigned i;
5413	pGip->u32Magic = 0;
5414	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
5415	{
5416	pGip->aCPUs[i].u64NanoTS = 0;
5417	pGip->aCPUs[i].u64TSC = 0;
5418	pGip->aCPUs[i].iTSCHistoryHead = 0;
5419	}
5420	}
5421
5422
5423	/**
5424	* Worker routine for supdrvGipUpdate and supdrvGipUpdatePerCpu that
5425	* updates all the per cpu data except the transaction id.
5426	*
5427	* @param pGip The GIP.
5428	* @param pGipCpu Pointer to the per cpu data.
5429	* @param u64NanoTS The current time stamp.
5430	* @param u64TSC The current TSC.
5431	* @param iTick The current timer tick.
5432	*/
5433	static void supdrvGipDoUpdateCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick)
5434	{
5435	uint64_t u64TSCDelta;
5436	uint32_t u32UpdateIntervalTSC;
5437	uint32_t u32UpdateIntervalTSCSlack;
5438	unsigned iTSCHistoryHead;
5439	uint64_t u64CpuHz;
5440	uint32_t u32TransactionId;
5441
5442	/* Delta between this and the previous update. */
5443	ASMAtomicUoWriteU32(&pGipCpu->u32PrevUpdateIntervalNS, (uint32_t)(u64NanoTS - pGipCpu->u64NanoTS));
5444
5445	/*
5446	* Update the NanoTS.
5447	*/
5448	ASMAtomicWriteU64(&pGipCpu->u64NanoTS, u64NanoTS);
5449
5450	/*
5451	* Calc TSC delta.
5452	*/
5453	/** @todo validate the NanoTS delta, don't trust the OS to call us when it should... */
5454	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
5455	ASMAtomicWriteU64(&pGipCpu->u64TSC, u64TSC);
5456
5457	if (u64TSCDelta >> 32)
5458	{
5459	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
5460	pGipCpu->cErrors++;
5461	}
5462
5463	/*
5464	* On the 2nd and 3rd callout, reset the history with the current TSC
5465	* interval since the values entered by supdrvGipInit are totally off.
5466	* The interval on the 1st callout completely unreliable, the 2nd is a bit
5467	* better, while the 3rd should be most reliable.
5468	*/
5469	u32TransactionId = pGipCpu->u32TransactionId;
5470	if (RT_UNLIKELY( ( u32TransactionId == 5
5471	\|\| u32TransactionId == 7)
5472	&& ( iTick == 2
5473	\|\| iTick == 3) ))
5474	{
5475	unsigned i;
5476	for (i = 0; i < RT_ELEMENTS(pGipCpu->au32TSCHistory); i++)
5477	ASMAtomicUoWriteU32(&pGipCpu->au32TSCHistory[i], (uint32_t)u64TSCDelta);
5478	}
5479
5480	/*
5481	* TSC History.
5482	*/
5483	Assert(RT_ELEMENTS(pGipCpu->au32TSCHistory) == 8);
5484	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
5485	ASMAtomicWriteU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
5486	ASMAtomicWriteU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
5487
5488	/*
5489	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
5490	*/
5491	if (pGip->u32UpdateHz >= 1000)
5492	{
5493	uint32_t u32;
5494	u32 = pGipCpu->au32TSCHistory[0];
5495	u32 += pGipCpu->au32TSCHistory[1];
5496	u32 += pGipCpu->au32TSCHistory[2];
5497	u32 += pGipCpu->au32TSCHistory[3];
5498	u32 >>= 2;
5499	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
5500	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
5501	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
5502	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
5503	u32UpdateIntervalTSC >>= 2;
5504	u32UpdateIntervalTSC += u32;
5505	u32UpdateIntervalTSC >>= 1;
5506
5507	/* Value chosen for a 2GHz Athlon64 running linux 2.6.10/11, . */
5508	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
5509	}
5510	else if (pGip->u32UpdateHz >= 90)
5511	{
5512	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5513	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
5514	u32UpdateIntervalTSC >>= 1;
5515
5516	/* value chosen on a 2GHz thinkpad running windows */
5517	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
5518	}
5519	else
5520	{
5521	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5522
5523	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
5524	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
5525	}
5526	ASMAtomicWriteU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
5527
5528	/*
5529	* CpuHz.
5530	*/
5531	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, pGip->u32UpdateHz);
5532	ASMAtomicWriteU64(&pGipCpu->u64CpuHz, u64CpuHz);
5533	}
5534
5535
5536	/**
5537	* Updates the GIP.
5538	*
5539	* @param pGip Pointer to the GIP.
5540	* @param u64NanoTS The current nanosecond timesamp.
5541	* @param u64TSC The current TSC timesamp.
5542	* @param idCpu The CPU ID.
5543	* @param iTick The current timer tick.
5544	*/
5545	static void supdrvGipUpdate(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick)
5546	{
5547	/*
5548	* Determine the relevant CPU data.
5549	*/
5550	PSUPGIPCPU pGipCpu;
5551	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
5552	pGipCpu = &pGip->aCPUs[0];
5553	else
5554	{
5555	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
5556	if (RT_UNLIKELY(iCpu >= pGip->cCpus))
5557	return;
5558	pGipCpu = &pGip->aCPUs[iCpu];
5559	if (RT_UNLIKELY(pGipCpu->idCpu != idCpu))
5560	return;
5561	}
5562
5563	/*
5564	* Start update transaction.
5565	*/
5566	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5567	{
5568	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
5569	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5570	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5571	pGipCpu->cErrors++;
5572	return;
5573	}
5574
5575	/*
5576	* Recalc the update frequency every 0x800th time.
5577	*/
5578	if (!(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
5579	{
5580	if (pGip->u64NanoTSLastUpdateHz)
5581	{
5582	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
5583	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
5584	uint32_t u32UpdateHz = (uint32_t)((UINT64_C(1000000000) * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
5585	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
5586	{
5587	ASMAtomicWriteU32(&pGip->u32UpdateHz, u32UpdateHz);
5588	ASMAtomicWriteU32(&pGip->u32UpdateIntervalNS, 1000000000 / u32UpdateHz);
5589	}
5590	#endif
5591	}
5592	ASMAtomicWriteU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS);
5593	}
5594
5595	/*
5596	* Update the data.
5597	*/
5598	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS, u64TSC, iTick);
5599
5600	/*
5601	* Complete transaction.
5602	*/
5603	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5604	}
5605
5606
5607	/**
5608	* Updates the per cpu GIP data for the calling cpu.
5609	*
5610	* @param pGip Pointer to the GIP.
5611	* @param u64NanoTS The current nanosecond timesamp.
5612	* @param u64TSC The current TSC timesamp.
5613	* @param idCpu The CPU ID.
5614	* @param idApic The APIC id for the CPU index.
5615	* @param iTick The current timer tick.
5616	*/
5617	static void supdrvGipUpdatePerCpu(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC,
5618	RTCPUID idCpu, uint8_t idApic, uint64_t iTick)
5619	{
5620	unsigned iCpu = pGip->aiCpuFromApicId[idApic];
5621
5622	if (RT_LIKELY(iCpu < pGip->cCpus))
5623	{
5624	PSUPGIPCPU pGipCpu = &pGip->aCPUs[iCpu];
5625	if (pGipCpu->idCpu == idCpu)
5626	{
5627
5628	/*
5629	* Start update transaction.
5630	*/
5631	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5632	{
5633	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5634	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5635	pGipCpu->cErrors++;
5636	return;
5637	}
5638
5639	/*
5640	* Update the data.
5641	*/
5642	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS, u64TSC, iTick);
5643
5644	/*
5645	* Complete transaction.
5646	*/
5647	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5648	}
5649	}
5650	}

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

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDrv.c@ 37049

Download in other formats: