SUPDrv.c@ 37344

Last change on this file since 37344 was 37293, checked in by vboxsync, 14 years ago
Support/SupDrv.c: release assertion.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 213.9 KB

Line
1	/* $Revision: 37293 $ */
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	{ "RTMpNotificationRegister", (void *)RTMpNotificationRegister },
325	{ "RTMpNotificationDeregister", (void *)RTMpNotificationDeregister },
326	{ "RTMpOnAll", (void *)RTMpOnAll },
327	{ "RTMpOnOthers", (void *)RTMpOnOthers },
328	{ "RTMpOnSpecific", (void *)RTMpOnSpecific },
329	{ "RTMpPokeCpu", (void *)RTMpPokeCpu },
330	{ "RTPowerNotificationRegister", (void *)RTPowerNotificationRegister },
331	{ "RTPowerNotificationDeregister", (void *)RTPowerNotificationDeregister },
332	{ "RTLogRelDefaultInstance", (void *)RTLogRelDefaultInstance },
333	{ "RTLogSetDefaultInstanceThread", (void *)RTLogSetDefaultInstanceThread },
334	{ "RTLogLoggerExV", (void *)RTLogLoggerExV },
335	{ "RTLogPrintfV", (void *)RTLogPrintfV },
336	{ "RTR0AssertPanicSystem", (void *)RTR0AssertPanicSystem },
337	{ "RTAssertMsg1", (void *)RTAssertMsg1 },
338	{ "RTAssertMsg2V", (void *)RTAssertMsg2V },
339	{ "RTAssertSetQuiet", (void *)RTAssertSetQuiet },
340	{ "RTAssertMayPanic", (void *)RTAssertMayPanic },
341	{ "RTAssertSetMayPanic", (void *)RTAssertSetMayPanic },
342	{ "RTAssertAreQuiet", (void *)RTAssertAreQuiet },
343	{ "RTStrFormat", (void *)RTStrFormat },
344	{ "RTStrFormatNumber", (void *)RTStrFormatNumber },
345	{ "RTStrFormatTypeDeregister", (void *)RTStrFormatTypeDeregister },
346	{ "RTStrFormatTypeRegister", (void *)RTStrFormatTypeRegister },
347	{ "RTStrFormatTypeSetUser", (void *)RTStrFormatTypeSetUser },
348	{ "RTStrFormatV", (void *)RTStrFormatV },
349	{ "RTStrPrintf", (void *)RTStrPrintf },
350	{ "RTStrPrintfEx", (void *)RTStrPrintfEx },
351	{ "RTStrPrintfExV", (void *)RTStrPrintfExV },
352	{ "RTStrPrintfV", (void *)RTStrPrintfV },
353	{ "RTHandleTableAllocWithCtx", (void *)RTHandleTableAllocWithCtx },
354	{ "RTHandleTableCreate", (void *)RTHandleTableCreate },
355	{ "RTHandleTableCreateEx", (void *)RTHandleTableCreateEx },
356	{ "RTHandleTableDestroy", (void *)RTHandleTableDestroy },
357	{ "RTHandleTableFreeWithCtx", (void *)RTHandleTableFreeWithCtx },
358	{ "RTHandleTableLookupWithCtx", (void *)RTHandleTableLookupWithCtx },
359	{ "RTNetIPv4AddDataChecksum", (void *)RTNetIPv4AddDataChecksum },
360	{ "RTNetIPv4AddTCPChecksum", (void *)RTNetIPv4AddTCPChecksum },
361	{ "RTNetIPv4AddUDPChecksum", (void *)RTNetIPv4AddUDPChecksum },
362	{ "RTNetIPv4FinalizeChecksum", (void *)RTNetIPv4FinalizeChecksum },
363	{ "RTNetIPv4HdrChecksum", (void *)RTNetIPv4HdrChecksum },
364	{ "RTNetIPv4IsDHCPValid", (void *)RTNetIPv4IsDHCPValid },
365	{ "RTNetIPv4IsHdrValid", (void *)RTNetIPv4IsHdrValid },
366	{ "RTNetIPv4IsTCPSizeValid", (void *)RTNetIPv4IsTCPSizeValid },
367	{ "RTNetIPv4IsTCPValid", (void *)RTNetIPv4IsTCPValid },
368	{ "RTNetIPv4IsUDPSizeValid", (void *)RTNetIPv4IsUDPSizeValid },
369	{ "RTNetIPv4IsUDPValid", (void *)RTNetIPv4IsUDPValid },
370	{ "RTNetIPv4PseudoChecksum", (void *)RTNetIPv4PseudoChecksum },
371	{ "RTNetIPv4PseudoChecksumBits", (void *)RTNetIPv4PseudoChecksumBits },
372	{ "RTNetIPv4TCPChecksum", (void *)RTNetIPv4TCPChecksum },
373	{ "RTNetIPv4UDPChecksum", (void *)RTNetIPv4UDPChecksum },
374	{ "RTNetIPv6PseudoChecksum", (void *)RTNetIPv6PseudoChecksum },
375	{ "RTNetIPv6PseudoChecksumBits", (void *)RTNetIPv6PseudoChecksumBits },
376	{ "RTNetIPv6PseudoChecksumEx", (void *)RTNetIPv6PseudoChecksumEx },
377	{ "RTNetTCPChecksum", (void *)RTNetTCPChecksum },
378	{ "RTNetUDPChecksum", (void *)RTNetUDPChecksum },
379	{ "RTCrc32", (void *)RTCrc32 },
380	{ "RTCrc32Finish", (void *)RTCrc32Finish },
381	{ "RTCrc32Process", (void *)RTCrc32Process },
382	{ "RTCrc32Start", (void *)RTCrc32Start },
383	};
384
385	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
386	/**
387	* Drag in the rest of IRPT since we share it with the
388	* rest of the kernel modules on darwin.
389	*/
390	PFNRT g_apfnVBoxDrvIPRTDeps[] =
391	{
392	/* VBoxNetFlt */
393	(PFNRT)RTErrConvertFromErrno,
394	(PFNRT)RTUuidCompare,
395	(PFNRT)RTUuidCompareStr,
396	(PFNRT)RTUuidFromStr,
397	(PFNRT)RTStrDupTag,
398	(PFNRT)RTStrFree,
399	(PFNRT)RTStrCopy,
400	(PFNRT)RTStrNCmp,
401	/* VBoxNetAdp */
402	(PFNRT)RTRandBytes,
403	/* VBoxUSB */
404	(PFNRT)RTPathStripFilename,
405	NULL
406	};
407	#endif /* RT_OS_DARWIN \|\| RT_OS_SOLARIS \|\| RT_OS_SOLARIS */
408
409
410	/**
411	* Initializes the device extentsion structure.
412	*
413	* @returns IPRT status code.
414	* @param pDevExt The device extension to initialize.
415	* @param cbSession The size of the session structure. The size of
416	* SUPDRVSESSION may be smaller when SUPDRV_AGNOSTIC is
417	* defined because we're skipping the OS specific members
418	* then.
419	*/
420	int VBOXCALL supdrvInitDevExt(PSUPDRVDEVEXT pDevExt, size_t cbSession)
421	{
422	int rc;
423
424	#ifdef SUPDRV_WITH_RELEASE_LOGGER
425	/*
426	* Create the release log.
427	*/
428	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
429	PRTLOGGER pRelLogger;
430	rc = RTLogCreate(&pRelLogger, 0 /* fFlags */, "all",
431	"VBOX_RELEASE_LOG", RT_ELEMENTS(s_apszGroups), s_apszGroups, RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER, NULL);
432	if (RT_SUCCESS(rc))
433	RTLogRelSetDefaultInstance(pRelLogger);
434	/** @todo Add native hook for getting logger config parameters and setting
435	* them. On linux we should use the module parameter stuff... */
436	#endif
437
438	/*
439	* Initialize it.
440	*/
441	memset(pDevExt, 0, sizeof(*pDevExt));
442	rc = RTSpinlockCreate(&pDevExt->Spinlock);
443	if (RT_SUCCESS(rc))
444	{
445	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
446	rc = RTSemMutexCreate(&pDevExt->mtxLdr);
447	#else
448	rc = RTSemFastMutexCreate(&pDevExt->mtxLdr);
449	#endif
450	if (RT_SUCCESS(rc))
451	{
452	rc = RTSemFastMutexCreate(&pDevExt->mtxComponentFactory);
453	if (RT_SUCCESS(rc))
454	{
455	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
456	rc = RTSemMutexCreate(&pDevExt->mtxGip);
457	#else
458	rc = RTSemFastMutexCreate(&pDevExt->mtxGip);
459	#endif
460	if (RT_SUCCESS(rc))
461	{
462	rc = supdrvGipCreate(pDevExt);
463	if (RT_SUCCESS(rc))
464	{
465	pDevExt->u32Cookie = BIRD; /** @todo make this random? */
466	pDevExt->cbSession = (uint32_t)cbSession;
467
468	/*
469	* Fixup the absolute symbols.
470	*
471	* Because of the table indexing assumptions we'll have a little #ifdef orgy
472	* here rather than distributing this to OS specific files. At least for now.
473	*/
474	#ifdef RT_OS_DARWIN
475	# if ARCH_BITS == 32
476	if (SUPR0GetPagingMode() >= SUPPAGINGMODE_AMD64)
477	{
478	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
479	g_aFunctions[1].pfn = (void )0x80; / SUPR0Abs64bitKernelCS - KERNEL64_CS, seg.h */
480	g_aFunctions[2].pfn = (void )0x88; / SUPR0Abs64bitKernelSS - KERNEL64_SS, seg.h */
481	g_aFunctions[3].pfn = (void )0x88; / SUPR0Abs64bitKernelDS - KERNEL64_SS, seg.h */
482	}
483	else
484	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
485	g_aFunctions[4].pfn = (void )0x08; / SUPR0AbsKernelCS - KERNEL_CS, seg.h */
486	g_aFunctions[5].pfn = (void )0x10; / SUPR0AbsKernelSS - KERNEL_DS, seg.h */
487	g_aFunctions[6].pfn = (void )0x10; / SUPR0AbsKernelDS - KERNEL_DS, seg.h */
488	g_aFunctions[7].pfn = (void )0x10; / SUPR0AbsKernelES - KERNEL_DS, seg.h */
489	g_aFunctions[8].pfn = (void )0x10; / SUPR0AbsKernelFS - KERNEL_DS, seg.h */
490	g_aFunctions[9].pfn = (void )0x48; / SUPR0AbsKernelGS - CPU_DATA_GS, seg.h */
491	# else /* 64-bit darwin: */
492	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
493	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
494	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
495	g_aFunctions[3].pfn = (void )0; / SUPR0Abs64bitKernelDS */
496	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
497	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
498	g_aFunctions[6].pfn = (void )0; / SUPR0AbsKernelDS */
499	g_aFunctions[7].pfn = (void )0; / SUPR0AbsKernelES */
500	g_aFunctions[8].pfn = (void )0; / SUPR0AbsKernelFS */
501	g_aFunctions[9].pfn = (void )0; / SUPR0AbsKernelGS */
502
503	# endif
504	#else /* !RT_OS_DARWIN */
505	# if ARCH_BITS == 64
506	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
507	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
508	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
509	g_aFunctions[3].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0Abs64bitKernelDS */
510	# else
511	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
512	# endif
513	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
514	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
515	g_aFunctions[6].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0AbsKernelDS */
516	g_aFunctions[7].pfn = (void )(uintptr_t)ASMGetES(); / SUPR0AbsKernelES */
517	g_aFunctions[8].pfn = (void )(uintptr_t)ASMGetFS(); / SUPR0AbsKernelFS */
518	g_aFunctions[9].pfn = (void )(uintptr_t)ASMGetGS(); / SUPR0AbsKernelGS */
519	#endif /* !RT_OS_DARWIN */
520	return VINF_SUCCESS;
521	}
522
523	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
524	RTSemMutexDestroy(pDevExt->mtxGip);
525	pDevExt->mtxGip = NIL_RTSEMMUTEX;
526	#else
527	RTSemFastMutexDestroy(pDevExt->mtxGip);
528	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
529	#endif
530	}
531	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
532	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
533	}
534	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
535	RTSemMutexDestroy(pDevExt->mtxLdr);
536	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
537	#else
538	RTSemFastMutexDestroy(pDevExt->mtxLdr);
539	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
540	#endif
541	}
542	RTSpinlockDestroy(pDevExt->Spinlock);
543	pDevExt->Spinlock = NIL_RTSPINLOCK;
544	}
545	#ifdef SUPDRV_WITH_RELEASE_LOGGER
546	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
547	RTLogDestroy(RTLogSetDefaultInstance(NULL));
548	#endif
549
550	return rc;
551	}
552
553
554	/**
555	* Delete the device extension (e.g. cleanup members).
556	*
557	* @param pDevExt The device extension to delete.
558	*/
559	void VBOXCALL supdrvDeleteDevExt(PSUPDRVDEVEXT pDevExt)
560	{
561	PSUPDRVOBJ pObj;
562	PSUPDRVUSAGE pUsage;
563
564	/*
565	* Kill mutexes and spinlocks.
566	*/
567	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
568	RTSemMutexDestroy(pDevExt->mtxGip);
569	pDevExt->mtxGip = NIL_RTSEMMUTEX;
570	#else
571	RTSemFastMutexDestroy(pDevExt->mtxGip);
572	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
573	#endif
574	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
575	RTSemMutexDestroy(pDevExt->mtxLdr);
576	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
577	#else
578	RTSemFastMutexDestroy(pDevExt->mtxLdr);
579	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
580	#endif
581	RTSpinlockDestroy(pDevExt->Spinlock);
582	pDevExt->Spinlock = NIL_RTSPINLOCK;
583	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
584	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
585
586	/*
587	* Free lists.
588	*/
589	/* objects. */
590	pObj = pDevExt->pObjs;
591	Assert(!pObj); /* (can trigger on forced unloads) */
592	pDevExt->pObjs = NULL;
593	while (pObj)
594	{
595	void *pvFree = pObj;
596	pObj = pObj->pNext;
597	RTMemFree(pvFree);
598	}
599
600	/* usage records. */
601	pUsage = pDevExt->pUsageFree;
602	pDevExt->pUsageFree = NULL;
603	while (pUsage)
604	{
605	void *pvFree = pUsage;
606	pUsage = pUsage->pNext;
607	RTMemFree(pvFree);
608	}
609
610	/* kill the GIP. */
611	supdrvGipDestroy(pDevExt);
612
613	#ifdef SUPDRV_WITH_RELEASE_LOGGER
614	/* destroy the loggers. */
615	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
616	RTLogDestroy(RTLogSetDefaultInstance(NULL));
617	#endif
618	}
619
620
621	/**
622	* Create session.
623	*
624	* @returns IPRT status code.
625	* @param pDevExt Device extension.
626	* @param fUser Flag indicating whether this is a user or kernel session.
627	* @param ppSession Where to store the pointer to the session data.
628	*/
629	int VBOXCALL supdrvCreateSession(PSUPDRVDEVEXT pDevExt, bool fUser, PSUPDRVSESSION *ppSession)
630	{
631	/*
632	* Allocate memory for the session data.
633	*/
634	int rc;
635	PSUPDRVSESSION pSession = *ppSession = (PSUPDRVSESSION)RTMemAllocZ(pDevExt->cbSession);
636	if (pSession)
637	{
638	/* Initialize session data. */
639	rc = RTSpinlockCreate(&pSession->Spinlock);
640	if (!rc)
641	{
642	rc = RTHandleTableCreateEx(&pSession->hHandleTable,
643	RTHANDLETABLE_FLAGS_LOCKED \| RTHANDLETABLE_FLAGS_CONTEXT,
644	1 /uBase/, 32768 /cMax/, supdrvSessionObjHandleRetain, pSession);
645	if (RT_SUCCESS(rc))
646	{
647	Assert(pSession->Spinlock != NIL_RTSPINLOCK);
648	pSession->pDevExt = pDevExt;
649	pSession->u32Cookie = BIRD_INV;
650	/*pSession->pLdrUsage = NULL;
651	pSession->pVM = NULL;
652	pSession->pUsage = NULL;
653	pSession->pGip = NULL;
654	pSession->fGipReferenced = false;
655	pSession->Bundle.cUsed = 0; */
656	pSession->Uid = NIL_RTUID;
657	pSession->Gid = NIL_RTGID;
658	if (fUser)
659	{
660	pSession->Process = RTProcSelf();
661	pSession->R0Process = RTR0ProcHandleSelf();
662	}
663	else
664	{
665	pSession->Process = NIL_RTPROCESS;
666	pSession->R0Process = NIL_RTR0PROCESS;
667	}
668
669	VBOXDRV_SUPDRV_SESSION_CREATE(pSession, fUser);
670	LogFlow(("Created session %p initial cookie=%#x\n", pSession, pSession->u32Cookie));
671	return VINF_SUCCESS;
672	}
673
674	RTSpinlockDestroy(pSession->Spinlock);
675	}
676	RTMemFree(pSession);
677	*ppSession = NULL;
678	Log(("Failed to create spinlock, rc=%d!\n", rc));
679	}
680	else
681	rc = VERR_NO_MEMORY;
682
683	return rc;
684	}
685
686
687	/**
688	* Shared code for cleaning up a session.
689	*
690	* @param pDevExt Device extension.
691	* @param pSession Session data.
692	* This data will be freed by this routine.
693	*/
694	void VBOXCALL supdrvCloseSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
695	{
696	VBOXDRV_SUPDRV_SESSION_CLOSE(pSession);
697
698	/*
699	* Cleanup the session first.
700	*/
701	supdrvCleanupSession(pDevExt, pSession);
702
703	/*
704	* Free the rest of the session stuff.
705	*/
706	RTSpinlockDestroy(pSession->Spinlock);
707	pSession->Spinlock = NIL_RTSPINLOCK;
708	pSession->pDevExt = NULL;
709	RTMemFree(pSession);
710	LogFlow(("supdrvCloseSession: returns\n"));
711	}
712
713
714	/**
715	* Shared code for cleaning up a session (but not quite freeing it).
716	*
717	* This is primarily intended for MAC OS X where we have to clean up the memory
718	* stuff before the file handle is closed.
719	*
720	* @param pDevExt Device extension.
721	* @param pSession Session data.
722	* This data will be freed by this routine.
723	*/
724	void VBOXCALL supdrvCleanupSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
725	{
726	int rc;
727	PSUPDRVBUNDLE pBundle;
728	LogFlow(("supdrvCleanupSession: pSession=%p\n", pSession));
729
730	/*
731	* Remove logger instances related to this session.
732	*/
733	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pSession);
734
735	/*
736	* Destroy the handle table.
737	*/
738	rc = RTHandleTableDestroy(pSession->hHandleTable, supdrvSessionObjHandleDelete, pSession);
739	AssertRC(rc);
740	pSession->hHandleTable = NIL_RTHANDLETABLE;
741
742	/*
743	* Release object references made in this session.
744	* In theory there should be noone racing us in this session.
745	*/
746	Log2(("release objects - start\n"));
747	if (pSession->pUsage)
748	{
749	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
750	PSUPDRVUSAGE pUsage;
751	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
752
753	while ((pUsage = pSession->pUsage) != NULL)
754	{
755	PSUPDRVOBJ pObj = pUsage->pObj;
756	pSession->pUsage = pUsage->pNext;
757
758	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
759	if (pUsage->cUsage < pObj->cUsage)
760	{
761	pObj->cUsage -= pUsage->cUsage;
762	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
763	}
764	else
765	{
766	/* Destroy the object and free the record. */
767	if (pDevExt->pObjs == pObj)
768	pDevExt->pObjs = pObj->pNext;
769	else
770	{
771	PSUPDRVOBJ pObjPrev;
772	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
773	if (pObjPrev->pNext == pObj)
774	{
775	pObjPrev->pNext = pObj->pNext;
776	break;
777	}
778	Assert(pObjPrev);
779	}
780	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
781
782	Log(("supdrvCleanupSession: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
783	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
784	if (pObj->pfnDestructor)
785	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
786	RTMemFree(pObj);
787	}
788
789	/* free it and continue. */
790	RTMemFree(pUsage);
791
792	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
793	}
794
795	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
796	AssertMsg(!pSession->pUsage, ("Some buster reregistered an object during desturction!\n"));
797	}
798	Log2(("release objects - done\n"));
799
800	/*
801	* Release memory allocated in the session.
802	*
803	* We do not serialize this as we assume that the application will
804	* not allocated memory while closing the file handle object.
805	*/
806	Log2(("freeing memory:\n"));
807	pBundle = &pSession->Bundle;
808	while (pBundle)
809	{
810	PSUPDRVBUNDLE pToFree;
811	unsigned i;
812
813	/*
814	* Check and unlock all entries in the bundle.
815	*/
816	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
817	{
818	if (pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ)
819	{
820	Log2(("eType=%d pvR0=%p pvR3=%p cb=%ld\n", pBundle->aMem[i].eType, RTR0MemObjAddress(pBundle->aMem[i].MemObj),
821	(void *)RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3), (long)RTR0MemObjSize(pBundle->aMem[i].MemObj)));
822	if (pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ)
823	{
824	rc = RTR0MemObjFree(pBundle->aMem[i].MapObjR3, false);
825	AssertRC(rc); /** @todo figure out how to handle this. */
826	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
827	}
828	rc = RTR0MemObjFree(pBundle->aMem[i].MemObj, true /* fFreeMappings */);
829	AssertRC(rc); /** @todo figure out how to handle this. */
830	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
831	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
832	}
833	}
834
835	/*
836	* Advance and free previous bundle.
837	*/
838	pToFree = pBundle;
839	pBundle = pBundle->pNext;
840
841	pToFree->pNext = NULL;
842	pToFree->cUsed = 0;
843	if (pToFree != &pSession->Bundle)
844	RTMemFree(pToFree);
845	}
846	Log2(("freeing memory - done\n"));
847
848	/*
849	* Deregister component factories.
850	*/
851	RTSemFastMutexRequest(pDevExt->mtxComponentFactory);
852	Log2(("deregistering component factories:\n"));
853	if (pDevExt->pComponentFactoryHead)
854	{
855	PSUPDRVFACTORYREG pPrev = NULL;
856	PSUPDRVFACTORYREG pCur = pDevExt->pComponentFactoryHead;
857	while (pCur)
858	{
859	if (pCur->pSession == pSession)
860	{
861	/* unlink it */
862	PSUPDRVFACTORYREG pNext = pCur->pNext;
863	if (pPrev)
864	pPrev->pNext = pNext;
865	else
866	pDevExt->pComponentFactoryHead = pNext;
867
868	/* free it */
869	pCur->pNext = NULL;
870	pCur->pSession = NULL;
871	pCur->pFactory = NULL;
872	RTMemFree(pCur);
873
874	/* next */
875	pCur = pNext;
876	}
877	else
878	{
879	/* next */
880	pPrev = pCur;
881	pCur = pCur->pNext;
882	}
883	}
884	}
885	RTSemFastMutexRelease(pDevExt->mtxComponentFactory);
886	Log2(("deregistering component factories - done\n"));
887
888	/*
889	* Loaded images needs to be dereferenced and possibly freed up.
890	*/
891	supdrvLdrLock(pDevExt);
892	Log2(("freeing images:\n"));
893	if (pSession->pLdrUsage)
894	{
895	PSUPDRVLDRUSAGE pUsage = pSession->pLdrUsage;
896	pSession->pLdrUsage = NULL;
897	while (pUsage)
898	{
899	void *pvFree = pUsage;
900	PSUPDRVLDRIMAGE pImage = pUsage->pImage;
901	if (pImage->cUsage > pUsage->cUsage)
902	pImage->cUsage -= pUsage->cUsage;
903	else
904	supdrvLdrFree(pDevExt, pImage);
905	pUsage->pImage = NULL;
906	pUsage = pUsage->pNext;
907	RTMemFree(pvFree);
908	}
909	}
910	supdrvLdrUnlock(pDevExt);
911	Log2(("freeing images - done\n"));
912
913	/*
914	* Unmap the GIP.
915	*/
916	Log2(("umapping GIP:\n"));
917	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
918	{
919	SUPR0GipUnmap(pSession);
920	pSession->fGipReferenced = 0;
921	}
922	Log2(("umapping GIP - done\n"));
923	}
924
925
926	/**
927	* RTHandleTableDestroy callback used by supdrvCleanupSession.
928	*
929	* @returns IPRT status code, see SUPR0ObjAddRef.
930	* @param hHandleTable The handle table handle. Ignored.
931	* @param pvObj The object pointer.
932	* @param pvCtx Context, the handle type. Ignored.
933	* @param pvUser Session pointer.
934	*/
935	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser)
936	{
937	NOREF(pvCtx);
938	NOREF(hHandleTable);
939	return SUPR0ObjAddRefEx(pvObj, (PSUPDRVSESSION)pvUser, true /fNoBlocking/);
940	}
941
942
943	/**
944	* RTHandleTableDestroy callback used by supdrvCleanupSession.
945	*
946	* @param hHandleTable The handle table handle. Ignored.
947	* @param h The handle value. Ignored.
948	* @param pvObj The object pointer.
949	* @param pvCtx Context, the handle type. Ignored.
950	* @param pvUser Session pointer.
951	*/
952	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser)
953	{
954	NOREF(pvCtx);
955	NOREF(h);
956	NOREF(hHandleTable);
957	SUPR0ObjRelease(pvObj, (PSUPDRVSESSION)pvUser);
958	}
959
960
961	/**
962	* Fast path I/O Control worker.
963	*
964	* @returns VBox status code that should be passed down to ring-3 unchanged.
965	* @param uIOCtl Function number.
966	* @param idCpu VMCPU id.
967	* @param pDevExt Device extention.
968	* @param pSession Session data.
969	*/
970	int VBOXCALL supdrvIOCtlFast(uintptr_t uIOCtl, VMCPUID idCpu, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
971	{
972	/*
973	* We check the two prereqs after doing this only to allow the compiler to optimize things better.
974	*/
975	if (RT_LIKELY(pSession->pVM && pDevExt->pfnVMMR0EntryFast))
976	{
977	switch (uIOCtl)
978	{
979	case SUP_IOCTL_FAST_DO_RAW_RUN:
980	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_RAW_RUN);
981	break;
982	case SUP_IOCTL_FAST_DO_HWACC_RUN:
983	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_HWACC_RUN);
984	break;
985	case SUP_IOCTL_FAST_DO_NOP:
986	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_NOP);
987	break;
988	default:
989	return VERR_INTERNAL_ERROR;
990	}
991	return VINF_SUCCESS;
992	}
993	return VERR_INTERNAL_ERROR;
994	}
995
996
997	/**
998	* Helper for supdrvIOCtl. Check if pszStr contains any character of pszChars.
999	* We would use strpbrk here if this function would be contained in the RedHat kABI white
1000	* list, see http://www.kerneldrivers.org/RHEL5.
1001	*
1002	* @returns 1 if pszStr does contain any character of pszChars, 0 otherwise.
1003	* @param pszStr String to check
1004	* @param pszChars Character set
1005	*/
1006	static int supdrvCheckInvalidChar(const char pszStr, const char pszChars)
1007	{
1008	int chCur;
1009	while ((chCur = *pszStr++) != '\0')
1010	{
1011	int ch;
1012	const char *psz = pszChars;
1013	while ((ch = *psz++) != '\0')
1014	if (ch == chCur)
1015	return 1;
1016
1017	}
1018	return 0;
1019	}
1020
1021
1022	/**
1023	* I/O Control worker.
1024	*
1025	* @returns IPRT status code.
1026	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1027	*
1028	* @param uIOCtl Function number.
1029	* @param pDevExt Device extention.
1030	* @param pSession Session data.
1031	* @param pReqHdr The request header.
1032	*/
1033	int VBOXCALL supdrvIOCtl(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1034	{
1035	VBOXDRV_SUPDRV_IOCTL_ENTRY(pSession, uIOCtl, pReqHdr);
1036
1037	/*
1038	* Validate the request.
1039	*/
1040	/* this first check could probably be omitted as its also done by the OS specific code... */
1041	if (RT_UNLIKELY( (pReqHdr->fFlags & SUPREQHDR_FLAGS_MAGIC_MASK) != SUPREQHDR_FLAGS_MAGIC
1042	\|\| pReqHdr->cbIn < sizeof(*pReqHdr)
1043	\|\| pReqHdr->cbOut < sizeof(*pReqHdr)))
1044	{
1045	OSDBGPRINT(("vboxdrv: Bad ioctl request header; cbIn=%#lx cbOut=%#lx fFlags=%#lx\n",
1046	(long)pReqHdr->cbIn, (long)pReqHdr->cbOut, (long)pReqHdr->fFlags));
1047	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1048	return VERR_INVALID_PARAMETER;
1049	}
1050	if (RT_UNLIKELY(uIOCtl == SUP_IOCTL_COOKIE))
1051	{
1052	if (pReqHdr->u32Cookie != SUPCOOKIE_INITIAL_COOKIE)
1053	{
1054	OSDBGPRINT(("SUP_IOCTL_COOKIE: bad cookie %#lx\n", (long)pReqHdr->u32Cookie));
1055	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1056	return VERR_INVALID_PARAMETER;
1057	}
1058	}
1059	else if (RT_UNLIKELY( pReqHdr->u32Cookie != pDevExt->u32Cookie
1060	\|\| pReqHdr->u32SessionCookie != pSession->u32Cookie))
1061	{
1062	OSDBGPRINT(("vboxdrv: bad cookie %#lx / %#lx.\n", (long)pReqHdr->u32Cookie, (long)pReqHdr->u32SessionCookie));
1063	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1064	return VERR_INVALID_PARAMETER;
1065	}
1066
1067	/*
1068	* Validation macros
1069	*/
1070	#define REQ_CHECK_SIZES_EX(Name, cbInExpect, cbOutExpect) \
1071	do { \
1072	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect) \|\| pReqHdr->cbOut != (cbOutExpect))) \
1073	{ \
1074	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld. cbOut=%ld expected %ld.\n", \
1075	(long)pReq->Hdr.cbIn, (long)(cbInExpect), (long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
1076	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1077	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1078	} \
1079	} while (0)
1080
1081	#define REQ_CHECK_SIZES(Name) REQ_CHECK_SIZES_EX(Name, Name ## _SIZE_IN, Name ## _SIZE_OUT)
1082
1083	#define REQ_CHECK_SIZE_IN(Name, cbInExpect) \
1084	do { \
1085	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect))) \
1086	{ \
1087	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld.\n", \
1088	(long)pReq->Hdr.cbIn, (long)(cbInExpect))); \
1089	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1090	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1091	} \
1092	} while (0)
1093
1094	#define REQ_CHECK_SIZE_OUT(Name, cbOutExpect) \
1095	do { \
1096	if (RT_UNLIKELY(pReqHdr->cbOut != (cbOutExpect))) \
1097	{ \
1098	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbOut=%ld expected %ld.\n", \
1099	(long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
1100	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1101	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1102	} \
1103	} while (0)
1104
1105	#define REQ_CHECK_EXPR(Name, expr) \
1106	do { \
1107	if (RT_UNLIKELY(!(expr))) \
1108	{ \
1109	OSDBGPRINT(( #Name ": %s\n", #expr)); \
1110	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1111	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1112	} \
1113	} while (0)
1114
1115	#define REQ_CHECK_EXPR_FMT(expr, fmt) \
1116	do { \
1117	if (RT_UNLIKELY(!(expr))) \
1118	{ \
1119	OSDBGPRINT( fmt ); \
1120	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1121	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1122	} \
1123	} while (0)
1124
1125
1126	/*
1127	* The switch.
1128	*/
1129	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
1130	{
1131	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
1132	{
1133	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
1134	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
1135	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
1136	{
1137	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
1138	pReq->Hdr.rc = VERR_INVALID_MAGIC;
1139	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, VERR_INVALID_MAGIC);
1140	return 0;
1141	}
1142
1143	#if 0
1144	/*
1145	* Call out to the OS specific code and let it do permission checks on the
1146	* client process.
1147	*/
1148	if (!supdrvOSValidateClientProcess(pDevExt, pSession))
1149	{
1150	pReq->u.Out.u32Cookie = 0xffffffff;
1151	pReq->u.Out.u32SessionCookie = 0xffffffff;
1152	pReq->u.Out.u32SessionVersion = 0xffffffff;
1153	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1154	pReq->u.Out.pSession = NULL;
1155	pReq->u.Out.cFunctions = 0;
1156	pReq->Hdr.rc = VERR_PERMISSION_DENIED;
1157	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, VERR_PERMISSION_DENIED);
1158	return 0;
1159	}
1160	#endif
1161
1162	/*
1163	* Match the version.
1164	* The current logic is very simple, match the major interface version.
1165	*/
1166	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
1167	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
1168	{
1169	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1170	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
1171	pReq->u.Out.u32Cookie = 0xffffffff;
1172	pReq->u.Out.u32SessionCookie = 0xffffffff;
1173	pReq->u.Out.u32SessionVersion = 0xffffffff;
1174	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1175	pReq->u.Out.pSession = NULL;
1176	pReq->u.Out.cFunctions = 0;
1177	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1178	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1179	return 0;
1180	}
1181
1182	/*
1183	* Fill in return data and be gone.
1184	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
1185	* u32SessionVersion <= u32ReqVersion!
1186	*/
1187	/** @todo Somehow validate the client and negotiate a secure cookie... */
1188	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
1189	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
1190	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
1191	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1192	pReq->u.Out.pSession = pSession;
1193	pReq->u.Out.cFunctions = sizeof(g_aFunctions) / sizeof(g_aFunctions[0]);
1194	pReq->Hdr.rc = VINF_SUCCESS;
1195	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1196	return 0;
1197	}
1198
1199	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_QUERY_FUNCS(0)):
1200	{
1201	/* validate */
1202	PSUPQUERYFUNCS pReq = (PSUPQUERYFUNCS)pReqHdr;
1203	REQ_CHECK_SIZES_EX(SUP_IOCTL_QUERY_FUNCS, SUP_IOCTL_QUERY_FUNCS_SIZE_IN, SUP_IOCTL_QUERY_FUNCS_SIZE_OUT(RT_ELEMENTS(g_aFunctions)));
1204
1205	/* execute */
1206	pReq->u.Out.cFunctions = RT_ELEMENTS(g_aFunctions);
1207	memcpy(&pReq->u.Out.aFunctions[0], g_aFunctions, sizeof(g_aFunctions));
1208	pReq->Hdr.rc = VINF_SUCCESS;
1209	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1210	return 0;
1211	}
1212
1213	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_LOCK):
1214	{
1215	/* validate */
1216	PSUPPAGELOCK pReq = (PSUPPAGELOCK)pReqHdr;
1217	REQ_CHECK_SIZE_IN(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_IN);
1218	REQ_CHECK_SIZE_OUT(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_OUT(pReq->u.In.cPages));
1219	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.cPages > 0);
1220	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.pvR3 >= PAGE_SIZE);
1221
1222	/* execute */
1223	pReq->Hdr.rc = SUPR0LockMem(pSession, pReq->u.In.pvR3, pReq->u.In.cPages, &pReq->u.Out.aPages[0]);
1224	if (RT_FAILURE(pReq->Hdr.rc))
1225	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1226	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1227	return 0;
1228	}
1229
1230	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_UNLOCK):
1231	{
1232	/* validate */
1233	PSUPPAGEUNLOCK pReq = (PSUPPAGEUNLOCK)pReqHdr;
1234	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_UNLOCK);
1235
1236	/* execute */
1237	pReq->Hdr.rc = SUPR0UnlockMem(pSession, pReq->u.In.pvR3);
1238	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1239	return 0;
1240	}
1241
1242	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_ALLOC):
1243	{
1244	/* validate */
1245	PSUPCONTALLOC pReq = (PSUPCONTALLOC)pReqHdr;
1246	REQ_CHECK_SIZES(SUP_IOCTL_CONT_ALLOC);
1247
1248	/* execute */
1249	pReq->Hdr.rc = SUPR0ContAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.HCPhys);
1250	if (RT_FAILURE(pReq->Hdr.rc))
1251	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1252	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1253	return 0;
1254	}
1255
1256	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_FREE):
1257	{
1258	/* validate */
1259	PSUPCONTFREE pReq = (PSUPCONTFREE)pReqHdr;
1260	REQ_CHECK_SIZES(SUP_IOCTL_CONT_FREE);
1261
1262	/* execute */
1263	pReq->Hdr.rc = SUPR0ContFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1264	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1265	return 0;
1266	}
1267
1268	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_OPEN):
1269	{
1270	/* validate */
1271	PSUPLDROPEN pReq = (PSUPLDROPEN)pReqHdr;
1272	REQ_CHECK_SIZES(SUP_IOCTL_LDR_OPEN);
1273	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs > 0);
1274	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs < 16*_1M);
1275	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1276	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1277	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits < pReq->u.In.cbImageWithTabs);
1278	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.szName[0]);
1279	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1280	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, !supdrvCheckInvalidChar(pReq->u.In.szName, ";:()[]{}/\\\|&*%#@!~`\"'"));
1281	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szFilename, sizeof(pReq->u.In.szFilename)));
1282
1283	/* execute */
1284	pReq->Hdr.rc = supdrvIOCtl_LdrOpen(pDevExt, pSession, pReq);
1285	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1286	return 0;
1287	}
1288
1289	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOAD):
1290	{
1291	/* validate */
1292	PSUPLDRLOAD pReq = (PSUPLDRLOAD)pReqHdr;
1293	REQ_CHECK_EXPR(Name, pReq->Hdr.cbIn >= sizeof(*pReq));
1294	REQ_CHECK_SIZES_EX(SUP_IOCTL_LDR_LOAD, SUP_IOCTL_LDR_LOAD_SIZE_IN(pReq->u.In.cbImageWithTabs), SUP_IOCTL_LDR_LOAD_SIZE_OUT);
1295	REQ_CHECK_EXPR(SUP_IOCTL_LDR_LOAD, pReq->u.In.cSymbols <= 16384);
1296	REQ_CHECK_EXPR_FMT( !pReq->u.In.cSymbols
1297	\|\| ( pReq->u.In.offSymbols < pReq->u.In.cbImageWithTabs
1298	&& pReq->u.In.offSymbols + pReq->u.In.cSymbols * sizeof(SUPLDRSYM) <= pReq->u.In.cbImageWithTabs),
1299	("SUP_IOCTL_LDR_LOAD: offSymbols=%#lx cSymbols=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offSymbols,
1300	(long)pReq->u.In.cSymbols, (long)pReq->u.In.cbImageWithTabs));
1301	REQ_CHECK_EXPR_FMT( !pReq->u.In.cbStrTab
1302	\|\| ( pReq->u.In.offStrTab < pReq->u.In.cbImageWithTabs
1303	&& pReq->u.In.offStrTab + pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs
1304	&& pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs),
1305	("SUP_IOCTL_LDR_LOAD: offStrTab=%#lx cbStrTab=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offStrTab,
1306	(long)pReq->u.In.cbStrTab, (long)pReq->u.In.cbImageWithTabs));
1307
1308	if (pReq->u.In.cSymbols)
1309	{
1310	uint32_t i;
1311	PSUPLDRSYM paSyms = (PSUPLDRSYM)&pReq->u.In.abImage[pReq->u.In.offSymbols];
1312	for (i = 0; i < pReq->u.In.cSymbols; i++)
1313	{
1314	REQ_CHECK_EXPR_FMT(paSyms[i].offSymbol < pReq->u.In.cbImageWithTabs,
1315	("SUP_IOCTL_LDR_LOAD: sym #%ld: symb off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offSymbol, (long)pReq->u.In.cbImageWithTabs));
1316	REQ_CHECK_EXPR_FMT(paSyms[i].offName < pReq->u.In.cbStrTab,
1317	("SUP_IOCTL_LDR_LOAD: sym #%ld: name off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1318	REQ_CHECK_EXPR_FMT(RTStrEnd((char const *)&pReq->u.In.abImage[pReq->u.In.offStrTab + paSyms[i].offName],
1319	pReq->u.In.cbStrTab - paSyms[i].offName),
1320	("SUP_IOCTL_LDR_LOAD: sym #%ld: unterminated name! (%#lx / %#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1321	}
1322	}
1323
1324	/* execute */
1325	pReq->Hdr.rc = supdrvIOCtl_LdrLoad(pDevExt, pSession, pReq);
1326	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1327	return 0;
1328	}
1329
1330	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_FREE):
1331	{
1332	/* validate */
1333	PSUPLDRFREE pReq = (PSUPLDRFREE)pReqHdr;
1334	REQ_CHECK_SIZES(SUP_IOCTL_LDR_FREE);
1335
1336	/* execute */
1337	pReq->Hdr.rc = supdrvIOCtl_LdrFree(pDevExt, pSession, pReq);
1338	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1339	return 0;
1340	}
1341
1342	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_GET_SYMBOL):
1343	{
1344	/* validate */
1345	PSUPLDRGETSYMBOL pReq = (PSUPLDRGETSYMBOL)pReqHdr;
1346	REQ_CHECK_SIZES(SUP_IOCTL_LDR_GET_SYMBOL);
1347	REQ_CHECK_EXPR(SUP_IOCTL_LDR_GET_SYMBOL, RTStrEnd(pReq->u.In.szSymbol, sizeof(pReq->u.In.szSymbol)));
1348
1349	/* execute */
1350	pReq->Hdr.rc = supdrvIOCtl_LdrGetSymbol(pDevExt, pSession, pReq);
1351	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1352	return 0;
1353	}
1354
1355	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0(0)):
1356	{
1357	/* validate */
1358	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1359	Log4(("SUP_IOCTL_CALL_VMMR0: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1360	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1361
1362	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_VMMR0_SIZE(0))
1363	{
1364	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(0), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(0));
1365
1366	/* execute */
1367	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1368	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg, pSession);
1369	else
1370	pReq->Hdr.rc = VERR_WRONG_ORDER;
1371	}
1372	else
1373	{
1374	PSUPVMMR0REQHDR pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1375	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR)),
1376	("SUP_IOCTL_CALL_VMMR0: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR))));
1377	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1378	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(pVMMReq->cbReq));
1379
1380	/* execute */
1381	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1382	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1383	else
1384	pReq->Hdr.rc = VERR_WRONG_ORDER;
1385	}
1386
1387	if ( RT_FAILURE(pReq->Hdr.rc)
1388	&& pReq->Hdr.rc != VERR_INTERRUPTED
1389	&& pReq->Hdr.rc != VERR_TIMEOUT)
1390	Log(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1391	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1392	else
1393	Log4(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1394	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1395	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1396	return 0;
1397	}
1398
1399	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0_BIG):
1400	{
1401	/* validate */
1402	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1403	PSUPVMMR0REQHDR pVMMReq;
1404	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1405	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1406
1407	pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1408	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR)),
1409	("SUP_IOCTL_CALL_VMMR0_BIG: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR))));
1410	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0_BIG, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1411	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));
1412
1413	/* execute */
1414	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1415	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1416	else
1417	pReq->Hdr.rc = VERR_WRONG_ORDER;
1418
1419	if ( RT_FAILURE(pReq->Hdr.rc)
1420	&& pReq->Hdr.rc != VERR_INTERRUPTED
1421	&& pReq->Hdr.rc != VERR_TIMEOUT)
1422	Log(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1423	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1424	else
1425	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1426	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1427	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1428	return 0;
1429	}
1430
1431	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_PAGING_MODE):
1432	{
1433	/* validate */
1434	PSUPGETPAGINGMODE pReq = (PSUPGETPAGINGMODE)pReqHdr;
1435	REQ_CHECK_SIZES(SUP_IOCTL_GET_PAGING_MODE);
1436
1437	/* execute */
1438	pReq->Hdr.rc = VINF_SUCCESS;
1439	pReq->u.Out.enmMode = SUPR0GetPagingMode();
1440	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1441	return 0;
1442	}
1443
1444	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_ALLOC):
1445	{
1446	/* validate */
1447	PSUPLOWALLOC pReq = (PSUPLOWALLOC)pReqHdr;
1448	REQ_CHECK_EXPR(SUP_IOCTL_LOW_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_LOW_ALLOC_SIZE_IN);
1449	REQ_CHECK_SIZES_EX(SUP_IOCTL_LOW_ALLOC, SUP_IOCTL_LOW_ALLOC_SIZE_IN, SUP_IOCTL_LOW_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1450
1451	/* execute */
1452	pReq->Hdr.rc = SUPR0LowAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1453	if (RT_FAILURE(pReq->Hdr.rc))
1454	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1455	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1456	return 0;
1457	}
1458
1459	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_FREE):
1460	{
1461	/* validate */
1462	PSUPLOWFREE pReq = (PSUPLOWFREE)pReqHdr;
1463	REQ_CHECK_SIZES(SUP_IOCTL_LOW_FREE);
1464
1465	/* execute */
1466	pReq->Hdr.rc = SUPR0LowFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1467	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1468	return 0;
1469	}
1470
1471	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_MAP):
1472	{
1473	/* validate */
1474	PSUPGIPMAP pReq = (PSUPGIPMAP)pReqHdr;
1475	REQ_CHECK_SIZES(SUP_IOCTL_GIP_MAP);
1476
1477	/* execute */
1478	pReq->Hdr.rc = SUPR0GipMap(pSession, &pReq->u.Out.pGipR3, &pReq->u.Out.HCPhysGip);
1479	if (RT_SUCCESS(pReq->Hdr.rc))
1480	pReq->u.Out.pGipR0 = pDevExt->pGip;
1481	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1482	return 0;
1483	}
1484
1485	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_UNMAP):
1486	{
1487	/* validate */
1488	PSUPGIPUNMAP pReq = (PSUPGIPUNMAP)pReqHdr;
1489	REQ_CHECK_SIZES(SUP_IOCTL_GIP_UNMAP);
1490
1491	/* execute */
1492	pReq->Hdr.rc = SUPR0GipUnmap(pSession);
1493	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1494	return 0;
1495	}
1496
1497	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SET_VM_FOR_FAST):
1498	{
1499	/* validate */
1500	PSUPSETVMFORFAST pReq = (PSUPSETVMFORFAST)pReqHdr;
1501	REQ_CHECK_SIZES(SUP_IOCTL_SET_VM_FOR_FAST);
1502	REQ_CHECK_EXPR_FMT( !pReq->u.In.pVMR0
1503	\|\| ( VALID_PTR(pReq->u.In.pVMR0)
1504	&& !((uintptr_t)pReq->u.In.pVMR0 & (PAGE_SIZE - 1))),
1505	("SUP_IOCTL_SET_VM_FOR_FAST: pVMR0=%p!\n", pReq->u.In.pVMR0));
1506	/* execute */
1507	pSession->pVM = pReq->u.In.pVMR0;
1508	pReq->Hdr.rc = VINF_SUCCESS;
1509	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1510	return 0;
1511	}
1512
1513	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_ALLOC_EX):
1514	{
1515	/* validate */
1516	PSUPPAGEALLOCEX pReq = (PSUPPAGEALLOCEX)pReqHdr;
1517	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC_EX, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_EX_SIZE_IN);
1518	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));
1519	REQ_CHECK_EXPR_FMT(pReq->u.In.fKernelMapping \|\| pReq->u.In.fUserMapping,
1520	("SUP_IOCTL_PAGE_ALLOC_EX: No mapping requested!\n"));
1521	REQ_CHECK_EXPR_FMT(pReq->u.In.fUserMapping,
1522	("SUP_IOCTL_PAGE_ALLOC_EX: Must have user mapping!\n"));
1523	REQ_CHECK_EXPR_FMT(!pReq->u.In.fReserved0 && !pReq->u.In.fReserved1,
1524	("SUP_IOCTL_PAGE_ALLOC_EX: fReserved0=%d fReserved1=%d\n", pReq->u.In.fReserved0, pReq->u.In.fReserved1));
1525
1526	/* execute */
1527	pReq->Hdr.rc = SUPR0PageAllocEx(pSession, pReq->u.In.cPages, 0 /* fFlags */,
1528	pReq->u.In.fUserMapping ? &pReq->u.Out.pvR3 : NULL,
1529	pReq->u.In.fKernelMapping ? &pReq->u.Out.pvR0 : NULL,
1530	&pReq->u.Out.aPages[0]);
1531	if (RT_FAILURE(pReq->Hdr.rc))
1532	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1533	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1534	return 0;
1535	}
1536
1537	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_MAP_KERNEL):
1538	{
1539	/* validate */
1540	PSUPPAGEMAPKERNEL pReq = (PSUPPAGEMAPKERNEL)pReqHdr;
1541	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_MAP_KERNEL);
1542	REQ_CHECK_EXPR_FMT(!pReq->u.In.fFlags, ("SUP_IOCTL_PAGE_MAP_KERNEL: fFlags=%#x! MBZ\n", pReq->u.In.fFlags));
1543	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_MAP_KERNEL: offSub=%#x\n", pReq->u.In.offSub));
1544	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1545	("SUP_IOCTL_PAGE_MAP_KERNEL: cbSub=%#x\n", pReq->u.In.cbSub));
1546
1547	/* execute */
1548	pReq->Hdr.rc = SUPR0PageMapKernel(pSession, pReq->u.In.pvR3, pReq->u.In.offSub, pReq->u.In.cbSub,
1549	pReq->u.In.fFlags, &pReq->u.Out.pvR0);
1550	if (RT_FAILURE(pReq->Hdr.rc))
1551	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1552	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1553	return 0;
1554	}
1555
1556	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_PROTECT):
1557	{
1558	/* validate */
1559	PSUPPAGEPROTECT pReq = (PSUPPAGEPROTECT)pReqHdr;
1560	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_PROTECT);
1561	REQ_CHECK_EXPR_FMT(!(pReq->u.In.fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)),
1562	("SUP_IOCTL_PAGE_PROTECT: fProt=%#x!\n", pReq->u.In.fProt));
1563	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_PROTECT: offSub=%#x\n", pReq->u.In.offSub));
1564	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1565	("SUP_IOCTL_PAGE_PROTECT: cbSub=%#x\n", pReq->u.In.cbSub));
1566
1567	/* execute */
1568	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);
1569	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1570	return 0;
1571	}
1572
1573	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_FREE):
1574	{
1575	/* validate */
1576	PSUPPAGEFREE pReq = (PSUPPAGEFREE)pReqHdr;
1577	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_FREE);
1578
1579	/* execute */
1580	pReq->Hdr.rc = SUPR0PageFree(pSession, pReq->u.In.pvR3);
1581	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1582	return 0;
1583	}
1584
1585	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_SERVICE(0)):
1586	{
1587	/* validate */
1588	PSUPCALLSERVICE pReq = (PSUPCALLSERVICE)pReqHdr;
1589	Log4(("SUP_IOCTL_CALL_SERVICE: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1590	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1591
1592	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
1593	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(0), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(0));
1594	else
1595	{
1596	PSUPR0SERVICEREQHDR pSrvReq = (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0];
1597	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR)),
1598	("SUP_IOCTL_CALL_SERVICE: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR))));
1599	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, pSrvReq->u32Magic == SUPR0SERVICEREQHDR_MAGIC);
1600	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(pSrvReq->cbReq), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(pSrvReq->cbReq));
1601	}
1602	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1603
1604	/* execute */
1605	pReq->Hdr.rc = supdrvIOCtl_CallServiceModule(pDevExt, pSession, pReq);
1606	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1607	return 0;
1608	}
1609
1610	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOGGER_SETTINGS(0)):
1611	{
1612	/* validate */
1613	PSUPLOGGERSETTINGS pReq = (PSUPLOGGERSETTINGS)pReqHdr;
1614	size_t cbStrTab;
1615	REQ_CHECK_SIZE_OUT(SUP_IOCTL_LOGGER_SETTINGS, SUP_IOCTL_LOGGER_SETTINGS_SIZE_OUT);
1616	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->Hdr.cbIn >= SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(1));
1617	cbStrTab = pReq->Hdr.cbIn - SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(0);
1618	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offGroups < cbStrTab);
1619	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offFlags < cbStrTab);
1620	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offDestination < cbStrTab);
1621	REQ_CHECK_EXPR_FMT(pReq->u.In.szStrings[cbStrTab - 1] == '\0',
1622	("SUP_IOCTL_LOGGER_SETTINGS: cbIn=%#x cbStrTab=%#zx LastChar=%d\n",
1623	pReq->Hdr.cbIn, cbStrTab, pReq->u.In.szStrings[cbStrTab - 1]));
1624	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhich <= SUPLOGGERSETTINGS_WHICH_RELEASE);
1625	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhat <= SUPLOGGERSETTINGS_WHAT_DESTROY);
1626
1627	/* execute */
1628	pReq->Hdr.rc = supdrvIOCtl_LoggerSettings(pDevExt, pSession, pReq);
1629	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1630	return 0;
1631	}
1632
1633	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP2):
1634	{
1635	/* validate */
1636	PSUPSEMOP2 pReq = (PSUPSEMOP2)pReqHdr;
1637	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP2, SUP_IOCTL_SEM_OP2_SIZE_IN, SUP_IOCTL_SEM_OP2_SIZE_OUT);
1638	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP2, pReq->u.In.uReserved == 0);
1639
1640	/* execute */
1641	switch (pReq->u.In.uType)
1642	{
1643	case SUP_SEM_TYPE_EVENT:
1644	{
1645	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1646	switch (pReq->u.In.uOp)
1647	{
1648	case SUPSEMOP2_WAIT_MS_REL:
1649	pReq->Hdr.rc = SUPSemEventWaitNoResume(pSession, hEvent, pReq->u.In.uArg.cRelMsTimeout);
1650	break;
1651	case SUPSEMOP2_WAIT_NS_ABS:
1652	pReq->Hdr.rc = SUPSemEventWaitNsAbsIntr(pSession, hEvent, pReq->u.In.uArg.uAbsNsTimeout);
1653	break;
1654	case SUPSEMOP2_WAIT_NS_REL:
1655	pReq->Hdr.rc = SUPSemEventWaitNsRelIntr(pSession, hEvent, pReq->u.In.uArg.cRelNsTimeout);
1656	break;
1657	case SUPSEMOP2_SIGNAL:
1658	pReq->Hdr.rc = SUPSemEventSignal(pSession, hEvent);
1659	break;
1660	case SUPSEMOP2_CLOSE:
1661	pReq->Hdr.rc = SUPSemEventClose(pSession, hEvent);
1662	break;
1663	case SUPSEMOP2_RESET:
1664	default:
1665	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1666	break;
1667	}
1668	break;
1669	}
1670
1671	case SUP_SEM_TYPE_EVENT_MULTI:
1672	{
1673	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1674	switch (pReq->u.In.uOp)
1675	{
1676	case SUPSEMOP2_WAIT_MS_REL:
1677	pReq->Hdr.rc = SUPSemEventMultiWaitNoResume(pSession, hEventMulti, pReq->u.In.uArg.cRelMsTimeout);
1678	break;
1679	case SUPSEMOP2_WAIT_NS_ABS:
1680	pReq->Hdr.rc = SUPSemEventMultiWaitNsAbsIntr(pSession, hEventMulti, pReq->u.In.uArg.uAbsNsTimeout);
1681	break;
1682	case SUPSEMOP2_WAIT_NS_REL:
1683	pReq->Hdr.rc = SUPSemEventMultiWaitNsRelIntr(pSession, hEventMulti, pReq->u.In.uArg.cRelNsTimeout);
1684	break;
1685	case SUPSEMOP2_SIGNAL:
1686	pReq->Hdr.rc = SUPSemEventMultiSignal(pSession, hEventMulti);
1687	break;
1688	case SUPSEMOP2_CLOSE:
1689	pReq->Hdr.rc = SUPSemEventMultiClose(pSession, hEventMulti);
1690	break;
1691	case SUPSEMOP2_RESET:
1692	pReq->Hdr.rc = SUPSemEventMultiReset(pSession, hEventMulti);
1693	break;
1694	default:
1695	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1696	break;
1697	}
1698	break;
1699	}
1700
1701	default:
1702	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1703	break;
1704	}
1705	return 0;
1706	}
1707
1708	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP3):
1709	{
1710	/* validate */
1711	PSUPSEMOP3 pReq = (PSUPSEMOP3)pReqHdr;
1712	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP3, SUP_IOCTL_SEM_OP3_SIZE_IN, SUP_IOCTL_SEM_OP3_SIZE_OUT);
1713	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, pReq->u.In.u32Reserved == 0 && pReq->u.In.u64Reserved == 0);
1714
1715	/* execute */
1716	switch (pReq->u.In.uType)
1717	{
1718	case SUP_SEM_TYPE_EVENT:
1719	{
1720	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1721	switch (pReq->u.In.uOp)
1722	{
1723	case SUPSEMOP3_CREATE:
1724	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1725	pReq->Hdr.rc = SUPSemEventCreate(pSession, &hEvent);
1726	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEvent;
1727	break;
1728	case SUPSEMOP3_GET_RESOLUTION:
1729	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1730	pReq->Hdr.rc = VINF_SUCCESS;
1731	pReq->Hdr.cbOut = sizeof(*pReq);
1732	pReq->u.Out.cNsResolution = SUPSemEventGetResolution(pSession);
1733	break;
1734	default:
1735	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1736	break;
1737	}
1738	break;
1739	}
1740
1741	case SUP_SEM_TYPE_EVENT_MULTI:
1742	{
1743	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1744	switch (pReq->u.In.uOp)
1745	{
1746	case SUPSEMOP3_CREATE:
1747	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1748	pReq->Hdr.rc = SUPSemEventMultiCreate(pSession, &hEventMulti);
1749	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEventMulti;
1750	break;
1751	case SUPSEMOP3_GET_RESOLUTION:
1752	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1753	pReq->Hdr.rc = VINF_SUCCESS;
1754	pReq->u.Out.cNsResolution = SUPSemEventMultiGetResolution(pSession);
1755	break;
1756	default:
1757	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1758	break;
1759	}
1760	break;
1761	}
1762
1763	default:
1764	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1765	break;
1766	}
1767	return 0;
1768	}
1769
1770	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
1771	{
1772	/* validate */
1773	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
1774	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
1775	REQ_CHECK_EXPR(SUP_IOCTL_VT_CAPS, pReq->Hdr.cbIn <= SUP_IOCTL_VT_CAPS_SIZE_IN);
1776
1777	/* execute */
1778	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.Caps);
1779	if (RT_FAILURE(pReq->Hdr.rc))
1780	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1781	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1782	return 0;
1783	}
1784
1785	default:
1786	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1787	break;
1788	}
1789	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_GENERAL_FAILURE, VERR_GENERAL_FAILURE);
1790	return VERR_GENERAL_FAILURE;
1791	}
1792
1793
1794	/**
1795	* Inter-Driver Communication (IDC) worker.
1796	*
1797	* @returns VBox status code.
1798	* @retval VINF_SUCCESS on success.
1799	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1800	* @retval VERR_NOT_SUPPORTED if the request isn't supported.
1801	*
1802	* @param uReq The request (function) code.
1803	* @param pDevExt Device extention.
1804	* @param pSession Session data.
1805	* @param pReqHdr The request header.
1806	*/
1807	int VBOXCALL supdrvIDC(uintptr_t uReq, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQHDR pReqHdr)
1808	{
1809	/*
1810	* The OS specific code has already validated the pSession
1811	* pointer, and the request size being greater or equal to
1812	* size of the header.
1813	*
1814	* So, just check that pSession is a kernel context session.
1815	*/
1816	if (RT_UNLIKELY( pSession
1817	&& pSession->R0Process != NIL_RTR0PROCESS))
1818	return VERR_INVALID_PARAMETER;
1819
1820	/*
1821	* Validation macro.
1822	*/
1823	#define REQ_CHECK_IDC_SIZE(Name, cbExpect) \
1824	do { \
1825	if (RT_UNLIKELY(pReqHdr->cb != (cbExpect))) \
1826	{ \
1827	OSDBGPRINT(( #Name ": Invalid input/output sizes. cb=%ld expected %ld.\n", \
1828	(long)pReqHdr->cb, (long)(cbExpect))); \
1829	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1830	} \
1831	} while (0)
1832
1833	switch (uReq)
1834	{
1835	case SUPDRV_IDC_REQ_CONNECT:
1836	{
1837	PSUPDRVIDCREQCONNECT pReq = (PSUPDRVIDCREQCONNECT)pReqHdr;
1838	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_CONNECT, sizeof(*pReq));
1839
1840	/*
1841	* Validate the cookie and other input.
1842	*/
1843	if (pReq->Hdr.pSession != NULL)
1844	{
1845	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: pSession=%p expected NULL!\n", pReq->Hdr.pSession));
1846	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1847	}
1848	if (pReq->u.In.u32MagicCookie != SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE)
1849	{
1850	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: u32MagicCookie=%#x expected %#x!\n",
1851	(unsigned)pReq->u.In.u32MagicCookie, (unsigned)SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE));
1852	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1853	}
1854	if ( pReq->u.In.uMinVersion > pReq->u.In.uReqVersion
1855	\|\| (pReq->u.In.uMinVersion & UINT32_C(0xffff0000)) != (pReq->u.In.uReqVersion & UINT32_C(0xffff0000)))
1856	{
1857	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: uMinVersion=%#x uMaxVersion=%#x doesn't match!\n",
1858	pReq->u.In.uMinVersion, pReq->u.In.uReqVersion));
1859	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1860	}
1861
1862	/*
1863	* Match the version.
1864	* The current logic is very simple, match the major interface version.
1865	*/
1866	if ( pReq->u.In.uMinVersion > SUPDRV_IDC_VERSION
1867	\|\| (pReq->u.In.uMinVersion & 0xffff0000) != (SUPDRV_IDC_VERSION & 0xffff0000))
1868	{
1869	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1870	pReq->u.In.uReqVersion, pReq->u.In.uMinVersion, (unsigned)SUPDRV_IDC_VERSION));
1871	pReq->u.Out.pSession = NULL;
1872	pReq->u.Out.uSessionVersion = 0xffffffff;
1873	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1874	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1875	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1876	return VINF_SUCCESS;
1877	}
1878
1879	pReq->u.Out.pSession = NULL;
1880	pReq->u.Out.uSessionVersion = SUPDRV_IDC_VERSION;
1881	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1882	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1883
1884	/*
1885	* On NT we will already have a session associated with the
1886	* client, just like with the SUP_IOCTL_COOKIE request, while
1887	* the other doesn't.
1888	*/
1889	#ifdef RT_OS_WINDOWS
1890	pReq->Hdr.rc = VINF_SUCCESS;
1891	#else
1892	AssertReturn(!pSession, VERR_INTERNAL_ERROR);
1893	pReq->Hdr.rc = supdrvCreateSession(pDevExt, false /* fUser */, &pSession);
1894	if (RT_FAILURE(pReq->Hdr.rc))
1895	{
1896	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: failed to create session, rc=%d\n", pReq->Hdr.rc));
1897	return VINF_SUCCESS;
1898	}
1899	#endif
1900
1901	pReq->u.Out.pSession = pSession;
1902	pReq->Hdr.pSession = pSession;
1903
1904	return VINF_SUCCESS;
1905	}
1906
1907	case SUPDRV_IDC_REQ_DISCONNECT:
1908	{
1909	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_DISCONNECT, sizeof(*pReqHdr));
1910
1911	#ifdef RT_OS_WINDOWS
1912	/* Windows will destroy the session when the file object is destroyed. */
1913	#else
1914	supdrvCloseSession(pDevExt, pSession);
1915	#endif
1916	return pReqHdr->rc = VINF_SUCCESS;
1917	}
1918
1919	case SUPDRV_IDC_REQ_GET_SYMBOL:
1920	{
1921	PSUPDRVIDCREQGETSYM pReq = (PSUPDRVIDCREQGETSYM)pReqHdr;
1922	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_GET_SYMBOL, sizeof(*pReq));
1923
1924	pReq->Hdr.rc = supdrvIDC_LdrGetSymbol(pDevExt, pSession, pReq);
1925	return VINF_SUCCESS;
1926	}
1927
1928	case SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY:
1929	{
1930	PSUPDRVIDCREQCOMPREGFACTORY pReq = (PSUPDRVIDCREQCOMPREGFACTORY)pReqHdr;
1931	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY, sizeof(*pReq));
1932
1933	pReq->Hdr.rc = SUPR0ComponentRegisterFactory(pSession, pReq->u.In.pFactory);
1934	return VINF_SUCCESS;
1935	}
1936
1937	case SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY:
1938	{
1939	PSUPDRVIDCREQCOMPDEREGFACTORY pReq = (PSUPDRVIDCREQCOMPDEREGFACTORY)pReqHdr;
1940	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY, sizeof(*pReq));
1941
1942	pReq->Hdr.rc = SUPR0ComponentDeregisterFactory(pSession, pReq->u.In.pFactory);
1943	return VINF_SUCCESS;
1944	}
1945
1946	default:
1947	Log(("Unknown IDC %#lx\n", (long)uReq));
1948	break;
1949	}
1950
1951	#undef REQ_CHECK_IDC_SIZE
1952	return VERR_NOT_SUPPORTED;
1953	}
1954
1955
1956	/**
1957	* Register a object for reference counting.
1958	* The object is registered with one reference in the specified session.
1959	*
1960	* @returns Unique identifier on success (pointer).
1961	* All future reference must use this identifier.
1962	* @returns NULL on failure.
1963	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
1964	* @param pvUser1 The first user argument.
1965	* @param pvUser2 The second user argument.
1966	*/
1967	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
1968	{
1969	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1970	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1971	PSUPDRVOBJ pObj;
1972	PSUPDRVUSAGE pUsage;
1973
1974	/*
1975	* Validate the input.
1976	*/
1977	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
1978	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
1979	AssertPtrReturn(pfnDestructor, NULL);
1980
1981	/*
1982	* Allocate and initialize the object.
1983	*/
1984	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
1985	if (!pObj)
1986	return NULL;
1987	pObj->u32Magic = SUPDRVOBJ_MAGIC;
1988	pObj->enmType = enmType;
1989	pObj->pNext = NULL;
1990	pObj->cUsage = 1;
1991	pObj->pfnDestructor = pfnDestructor;
1992	pObj->pvUser1 = pvUser1;
1993	pObj->pvUser2 = pvUser2;
1994	pObj->CreatorUid = pSession->Uid;
1995	pObj->CreatorGid = pSession->Gid;
1996	pObj->CreatorProcess= pSession->Process;
1997	supdrvOSObjInitCreator(pObj, pSession);
1998
1999	/*
2000	* Allocate the usage record.
2001	* (We keep freed usage records around to simplify SUPR0ObjAddRefEx().)
2002	*/
2003	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2004
2005	pUsage = pDevExt->pUsageFree;
2006	if (pUsage)
2007	pDevExt->pUsageFree = pUsage->pNext;
2008	else
2009	{
2010	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2011	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
2012	if (!pUsage)
2013	{
2014	RTMemFree(pObj);
2015	return NULL;
2016	}
2017	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2018	}
2019
2020	/*
2021	* Insert the object and create the session usage record.
2022	*/
2023	/* The object. */
2024	pObj->pNext = pDevExt->pObjs;
2025	pDevExt->pObjs = pObj;
2026
2027	/* The session record. */
2028	pUsage->cUsage = 1;
2029	pUsage->pObj = pObj;
2030	pUsage->pNext = pSession->pUsage;
2031	/* Log2(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext)); */
2032	pSession->pUsage = pUsage;
2033
2034	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2035
2036	Log(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
2037	return pObj;
2038	}
2039
2040
2041	/**
2042	* Increment the reference counter for the object associating the reference
2043	* with the specified session.
2044	*
2045	* @returns IPRT status code.
2046	* @param pvObj The identifier returned by SUPR0ObjRegister().
2047	* @param pSession The session which is referencing the object.
2048	*
2049	* @remarks The caller should not own any spinlocks and must carefully protect
2050	* itself against potential race with the destructor so freed memory
2051	* isn't accessed here.
2052	*/
2053	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
2054	{
2055	return SUPR0ObjAddRefEx(pvObj, pSession, false /* fNoBlocking */);
2056	}
2057
2058
2059	/**
2060	* Increment the reference counter for the object associating the reference
2061	* with the specified session.
2062	*
2063	* @returns IPRT status code.
2064	* @retval VERR_TRY_AGAIN if fNoBlocking was set and a new usage record
2065	* couldn't be allocated. (If you see this you're not doing the right
2066	* thing and it won't ever work reliably.)
2067	*
2068	* @param pvObj The identifier returned by SUPR0ObjRegister().
2069	* @param pSession The session which is referencing the object.
2070	* @param fNoBlocking Set if it's not OK to block. Never try to make the
2071	* first reference to an object in a session with this
2072	* argument set.
2073	*
2074	* @remarks The caller should not own any spinlocks and must carefully protect
2075	* itself against potential race with the destructor so freed memory
2076	* isn't accessed here.
2077	*/
2078	SUPR0DECL(int) SUPR0ObjAddRefEx(void *pvObj, PSUPDRVSESSION pSession, bool fNoBlocking)
2079	{
2080	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2081	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2082	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2083	int rc = VINF_SUCCESS;
2084	PSUPDRVUSAGE pUsagePre;
2085	PSUPDRVUSAGE pUsage;
2086
2087	/*
2088	* Validate the input.
2089	* Be ready for the destruction race (someone might be stuck in the
2090	* destructor waiting a lock we own).
2091	*/
2092	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2093	AssertPtrReturn(pObj, VERR_INVALID_POINTER);
2094	AssertMsgReturn(pObj->u32Magic == SUPDRVOBJ_MAGIC \|\| pObj->u32Magic == SUPDRVOBJ_MAGIC_DEAD,
2095	("Invalid pvObj=%p magic=%#x (expected %#x or %#x)\n", pvObj, pObj->u32Magic, SUPDRVOBJ_MAGIC, SUPDRVOBJ_MAGIC_DEAD),
2096	VERR_INVALID_PARAMETER);
2097
2098	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2099
2100	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2101	{
2102	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2103
2104	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2105	return VERR_WRONG_ORDER;
2106	}
2107
2108	/*
2109	* Preallocate the usage record if we can.
2110	*/
2111	pUsagePre = pDevExt->pUsageFree;
2112	if (pUsagePre)
2113	pDevExt->pUsageFree = pUsagePre->pNext;
2114	else if (!fNoBlocking)
2115	{
2116	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2117	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
2118	if (!pUsagePre)
2119	return VERR_NO_MEMORY;
2120
2121	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2122	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2123	{
2124	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2125
2126	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2127	return VERR_WRONG_ORDER;
2128	}
2129	}
2130
2131	/*
2132	* Reference the object.
2133	*/
2134	pObj->cUsage++;
2135
2136	/*
2137	* Look for the session record.
2138	*/
2139	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
2140	{
2141	/Log(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2142	if (pUsage->pObj == pObj)
2143	break;
2144	}
2145	if (pUsage)
2146	pUsage->cUsage++;
2147	else if (pUsagePre)
2148	{
2149	/* create a new session record. */
2150	pUsagePre->cUsage = 1;
2151	pUsagePre->pObj = pObj;
2152	pUsagePre->pNext = pSession->pUsage;
2153	pSession->pUsage = pUsagePre;
2154	/Log(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));/
2155
2156	pUsagePre = NULL;
2157	}
2158	else
2159	{
2160	pObj->cUsage--;
2161	rc = VERR_TRY_AGAIN;
2162	}
2163
2164	/*
2165	* Put any unused usage record into the free list..
2166	*/
2167	if (pUsagePre)
2168	{
2169	pUsagePre->pNext = pDevExt->pUsageFree;
2170	pDevExt->pUsageFree = pUsagePre;
2171	}
2172
2173	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2174
2175	return rc;
2176	}
2177
2178
2179	/**
2180	* Decrement / destroy a reference counter record for an object.
2181	*
2182	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
2183	*
2184	* @returns IPRT status code.
2185	* @retval VINF_SUCCESS if not destroyed.
2186	* @retval VINF_OBJECT_DESTROYED if it's destroyed by this release call.
2187	* @retval VERR_INVALID_PARAMETER if the object isn't valid. Will assert in
2188	* string builds.
2189	*
2190	* @param pvObj The identifier returned by SUPR0ObjRegister().
2191	* @param pSession The session which is referencing the object.
2192	*/
2193	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
2194	{
2195	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2196	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2197	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2198	int rc = VERR_INVALID_PARAMETER;
2199	PSUPDRVUSAGE pUsage;
2200	PSUPDRVUSAGE pUsagePrev;
2201
2202	/*
2203	* Validate the input.
2204	*/
2205	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2206	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2207	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2208	VERR_INVALID_PARAMETER);
2209
2210	/*
2211	* Acquire the spinlock and look for the usage record.
2212	*/
2213	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2214
2215	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
2216	pUsage;
2217	pUsagePrev = pUsage, pUsage = pUsage->pNext)
2218	{
2219	/Log2(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2220	if (pUsage->pObj == pObj)
2221	{
2222	rc = VINF_SUCCESS;
2223	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
2224	if (pUsage->cUsage > 1)
2225	{
2226	pObj->cUsage--;
2227	pUsage->cUsage--;
2228	}
2229	else
2230	{
2231	/*
2232	* Free the session record.
2233	*/
2234	if (pUsagePrev)
2235	pUsagePrev->pNext = pUsage->pNext;
2236	else
2237	pSession->pUsage = pUsage->pNext;
2238	pUsage->pNext = pDevExt->pUsageFree;
2239	pDevExt->pUsageFree = pUsage;
2240
2241	/* What about the object? */
2242	if (pObj->cUsage > 1)
2243	pObj->cUsage--;
2244	else
2245	{
2246	/*
2247	* Object is to be destroyed, unlink it.
2248	*/
2249	pObj->u32Magic = SUPDRVOBJ_MAGIC_DEAD;
2250	rc = VINF_OBJECT_DESTROYED;
2251	if (pDevExt->pObjs == pObj)
2252	pDevExt->pObjs = pObj->pNext;
2253	else
2254	{
2255	PSUPDRVOBJ pObjPrev;
2256	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
2257	if (pObjPrev->pNext == pObj)
2258	{
2259	pObjPrev->pNext = pObj->pNext;
2260	break;
2261	}
2262	Assert(pObjPrev);
2263	}
2264	}
2265	}
2266	break;
2267	}
2268	}
2269
2270	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2271
2272	/*
2273	* Call the destructor and free the object if required.
2274	*/
2275	if (rc == VINF_OBJECT_DESTROYED)
2276	{
2277	Log(("SUPR0ObjRelease: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
2278	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
2279	if (pObj->pfnDestructor)
2280	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
2281	RTMemFree(pObj);
2282	}
2283
2284	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
2285	return rc;
2286	}
2287
2288
2289	/**
2290	* Verifies that the current process can access the specified object.
2291	*
2292	* @returns The following IPRT status code:
2293	* @retval VINF_SUCCESS if access was granted.
2294	* @retval VERR_PERMISSION_DENIED if denied access.
2295	* @retval VERR_INVALID_PARAMETER if invalid parameter.
2296	*
2297	* @param pvObj The identifier returned by SUPR0ObjRegister().
2298	* @param pSession The session which wishes to access the object.
2299	* @param pszObjName Object string name. This is optional and depends on the object type.
2300	*
2301	* @remark The caller is responsible for making sure the object isn't removed while
2302	* we're inside this function. If uncertain about this, just call AddRef before calling us.
2303	*/
2304	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
2305	{
2306	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2307	int rc;
2308
2309	/*
2310	* Validate the input.
2311	*/
2312	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2313	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2314	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2315	VERR_INVALID_PARAMETER);
2316
2317	/*
2318	* Check access. (returns true if a decision has been made.)
2319	*/
2320	rc = VERR_INTERNAL_ERROR;
2321	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
2322	return rc;
2323
2324	/*
2325	* Default policy is to allow the user to access his own
2326	* stuff but nothing else.
2327	*/
2328	if (pObj->CreatorUid == pSession->Uid)
2329	return VINF_SUCCESS;
2330	return VERR_PERMISSION_DENIED;
2331	}
2332
2333
2334	/**
2335	* Lock pages.
2336	*
2337	* @returns IPRT status code.
2338	* @param pSession Session to which the locked memory should be associated.
2339	* @param pvR3 Start of the memory range to lock.
2340	* This must be page aligned.
2341	* @param cPages Number of pages to lock.
2342	* @param paPages Where to put the physical addresses of locked memory.
2343	*/
2344	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
2345	{
2346	int rc;
2347	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2348	const size_t cb = (size_t)cPages << PAGE_SHIFT;
2349	LogFlow(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
2350
2351	/*
2352	* Verify input.
2353	*/
2354	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2355	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
2356	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
2357	\|\| !pvR3)
2358	{
2359	Log(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
2360	return VERR_INVALID_PARAMETER;
2361	}
2362
2363	/*
2364	* Let IPRT do the job.
2365	*/
2366	Mem.eType = MEMREF_TYPE_LOCKED;
2367	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTMEM_PROT_READ \| RTMEM_PROT_WRITE, RTR0ProcHandleSelf());
2368	if (RT_SUCCESS(rc))
2369	{
2370	uint32_t iPage = cPages;
2371	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
2372	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
2373
2374	while (iPage-- > 0)
2375	{
2376	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2377	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
2378	{
2379	AssertMsgFailed(("iPage=%d\n", iPage));
2380	rc = VERR_INTERNAL_ERROR;
2381	break;
2382	}
2383	}
2384	if (RT_SUCCESS(rc))
2385	rc = supdrvMemAdd(&Mem, pSession);
2386	if (RT_FAILURE(rc))
2387	{
2388	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
2389	AssertRC(rc2);
2390	}
2391	}
2392
2393	return rc;
2394	}
2395
2396
2397	/**
2398	* Unlocks the memory pointed to by pv.
2399	*
2400	* @returns IPRT status code.
2401	* @param pSession Session to which the memory was locked.
2402	* @param pvR3 Memory to unlock.
2403	*/
2404	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2405	{
2406	LogFlow(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2407	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2408	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
2409	}
2410
2411
2412	/**
2413	* Allocates a chunk of page aligned memory with contiguous and fixed physical
2414	* backing.
2415	*
2416	* @returns IPRT status code.
2417	* @param pSession Session data.
2418	* @param cPages Number of pages to allocate.
2419	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
2420	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
2421	* @param pHCPhys Where to put the physical address of allocated memory.
2422	*/
2423	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
2424	{
2425	int rc;
2426	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2427	LogFlow(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
2428
2429	/*
2430	* Validate input.
2431	*/
2432	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2433	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
2434	{
2435	Log(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
2436	pSession, ppvR0, ppvR3, pHCPhys));
2437	return VERR_INVALID_PARAMETER;
2438
2439	}
2440	if (cPages < 1 \|\| cPages >= 256)
2441	{
2442	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2443	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2444	}
2445
2446	/*
2447	* Let IPRT do the job.
2448	*/
2449	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
2450	if (RT_SUCCESS(rc))
2451	{
2452	int rc2;
2453	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2454	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2455	if (RT_SUCCESS(rc))
2456	{
2457	Mem.eType = MEMREF_TYPE_CONT;
2458	rc = supdrvMemAdd(&Mem, pSession);
2459	if (!rc)
2460	{
2461	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2462	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2463	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
2464	return 0;
2465	}
2466
2467	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2468	AssertRC(rc2);
2469	}
2470	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2471	AssertRC(rc2);
2472	}
2473
2474	return rc;
2475	}
2476
2477
2478	/**
2479	* Frees memory allocated using SUPR0ContAlloc().
2480	*
2481	* @returns IPRT status code.
2482	* @param pSession The session to which the memory was allocated.
2483	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2484	*/
2485	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2486	{
2487	LogFlow(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2488	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2489	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
2490	}
2491
2492
2493	/**
2494	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
2495	*
2496	* The memory isn't zeroed.
2497	*
2498	* @returns IPRT status code.
2499	* @param pSession Session data.
2500	* @param cPages Number of pages to allocate.
2501	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
2502	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
2503	* @param paPages Where to put the physical addresses of allocated memory.
2504	*/
2505	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
2506	{
2507	unsigned iPage;
2508	int rc;
2509	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2510	LogFlow(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
2511
2512	/*
2513	* Validate input.
2514	*/
2515	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2516	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
2517	{
2518	Log(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
2519	pSession, ppvR3, ppvR0, paPages));
2520	return VERR_INVALID_PARAMETER;
2521
2522	}
2523	if (cPages < 1 \|\| cPages >= 256)
2524	{
2525	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2526	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2527	}
2528
2529	/*
2530	* Let IPRT do the work.
2531	*/
2532	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
2533	if (RT_SUCCESS(rc))
2534	{
2535	int rc2;
2536	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2537	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2538	if (RT_SUCCESS(rc))
2539	{
2540	Mem.eType = MEMREF_TYPE_LOW;
2541	rc = supdrvMemAdd(&Mem, pSession);
2542	if (!rc)
2543	{
2544	for (iPage = 0; iPage < cPages; iPage++)
2545	{
2546	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2547	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%RHp\n", paPages[iPage]));
2548	}
2549	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2550	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2551	return 0;
2552	}
2553
2554	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2555	AssertRC(rc2);
2556	}
2557
2558	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2559	AssertRC(rc2);
2560	}
2561
2562	return rc;
2563	}
2564
2565
2566	/**
2567	* Frees memory allocated using SUPR0LowAlloc().
2568	*
2569	* @returns IPRT status code.
2570	* @param pSession The session to which the memory was allocated.
2571	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2572	*/
2573	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2574	{
2575	LogFlow(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2576	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2577	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
2578	}
2579
2580
2581
2582	/**
2583	* Allocates a chunk of memory with both R0 and R3 mappings.
2584	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
2585	*
2586	* @returns IPRT status code.
2587	* @param pSession The session to associated the allocation with.
2588	* @param cb Number of bytes to allocate.
2589	* @param ppvR0 Where to store the address of the Ring-0 mapping.
2590	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2591	*/
2592	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
2593	{
2594	int rc;
2595	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2596	LogFlow(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
2597
2598	/*
2599	* Validate input.
2600	*/
2601	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2602	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
2603	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
2604	if (cb < 1 \|\| cb >= _4M)
2605	{
2606	Log(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
2607	return VERR_INVALID_PARAMETER;
2608	}
2609
2610	/*
2611	* Let IPRT do the work.
2612	*/
2613	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
2614	if (RT_SUCCESS(rc))
2615	{
2616	int rc2;
2617	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2618	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2619	if (RT_SUCCESS(rc))
2620	{
2621	Mem.eType = MEMREF_TYPE_MEM;
2622	rc = supdrvMemAdd(&Mem, pSession);
2623	if (!rc)
2624	{
2625	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2626	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2627	return VINF_SUCCESS;
2628	}
2629
2630	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2631	AssertRC(rc2);
2632	}
2633
2634	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2635	AssertRC(rc2);
2636	}
2637
2638	return rc;
2639	}
2640
2641
2642	/**
2643	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
2644	*
2645	* @returns IPRT status code.
2646	* @param pSession The session to which the memory was allocated.
2647	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2648	* @param paPages Where to store the physical addresses.
2649	*/
2650	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
2651	{
2652	PSUPDRVBUNDLE pBundle;
2653	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2654	LogFlow(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
2655
2656	/*
2657	* Validate input.
2658	*/
2659	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2660	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
2661	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
2662
2663	/*
2664	* Search for the address.
2665	*/
2666	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2667	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2668	{
2669	if (pBundle->cUsed > 0)
2670	{
2671	unsigned i;
2672	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2673	{
2674	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
2675	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2676	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2677	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2678	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
2679	)
2680	)
2681	{
2682	const size_t cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2683	size_t iPage;
2684	for (iPage = 0; iPage < cPages; iPage++)
2685	{
2686	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2687	paPages[iPage].uReserved = 0;
2688	}
2689	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2690	return VINF_SUCCESS;
2691	}
2692	}
2693	}
2694	}
2695	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2696	Log(("Failed to find %p!!!\n", (void *)uPtr));
2697	return VERR_INVALID_PARAMETER;
2698	}
2699
2700
2701	/**
2702	* Free memory allocated by SUPR0MemAlloc().
2703	*
2704	* @returns IPRT status code.
2705	* @param pSession The session owning the allocation.
2706	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2707	*/
2708	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2709	{
2710	LogFlow(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2711	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2712	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
2713	}
2714
2715
2716	/**
2717	* Allocates a chunk of memory with a kernel or/and a user mode mapping.
2718	*
2719	* The memory is fixed and it's possible to query the physical addresses using
2720	* SUPR0MemGetPhys().
2721	*
2722	* @returns IPRT status code.
2723	* @param pSession The session to associated the allocation with.
2724	* @param cPages The number of pages to allocate.
2725	* @param fFlags Flags, reserved for the future. Must be zero.
2726	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2727	* NULL if no ring-3 mapping.
2728	* @param ppvR3 Where to store the address of the Ring-0 mapping.
2729	* NULL if no ring-0 mapping.
2730	* @param paPages Where to store the addresses of the pages. Optional.
2731	*/
2732	SUPR0DECL(int) SUPR0PageAllocEx(PSUPDRVSESSION pSession, uint32_t cPages, uint32_t fFlags, PRTR3PTR ppvR3, PRTR0PTR ppvR0, PRTHCPHYS paPages)
2733	{
2734	int rc;
2735	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2736	LogFlow(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
2737
2738	/*
2739	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2740	*/
2741	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2742	AssertPtrNullReturn(ppvR3, VERR_INVALID_POINTER);
2743	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2744	AssertReturn(ppvR3 \|\| ppvR0, VERR_INVALID_PARAMETER);
2745	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2746	if (cPages < 1 \|\| cPages > VBOX_MAX_ALLOC_PAGE_COUNT)
2747	{
2748	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)));
2749	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2750	}
2751
2752	/*
2753	* Let IPRT do the work.
2754	*/
2755	if (ppvR0)
2756	rc = RTR0MemObjAllocPage(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, true /* fExecutable */);
2757	else
2758	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
2759	if (RT_SUCCESS(rc))
2760	{
2761	int rc2;
2762	if (ppvR3)
2763	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2764	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2765	else
2766	Mem.MapObjR3 = NIL_RTR0MEMOBJ;
2767	if (RT_SUCCESS(rc))
2768	{
2769	Mem.eType = MEMREF_TYPE_PAGE;
2770	rc = supdrvMemAdd(&Mem, pSession);
2771	if (!rc)
2772	{
2773	if (ppvR3)
2774	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2775	if (ppvR0)
2776	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2777	if (paPages)
2778	{
2779	uint32_t iPage = cPages;
2780	while (iPage-- > 0)
2781	{
2782	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
2783	Assert(paPages[iPage] != NIL_RTHCPHYS);
2784	}
2785	}
2786	return VINF_SUCCESS;
2787	}
2788
2789	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2790	AssertRC(rc2);
2791	}
2792
2793	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2794	AssertRC(rc2);
2795	}
2796	return rc;
2797	}
2798
2799
2800	/**
2801	* Maps a chunk of memory previously allocated by SUPR0PageAllocEx into kernel
2802	* space.
2803	*
2804	* @returns IPRT status code.
2805	* @param pSession The session to associated the allocation with.
2806	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
2807	* @param offSub Where to start mapping. Must be page aligned.
2808	* @param cbSub How much to map. Must be page aligned.
2809	* @param fFlags Flags, MBZ.
2810	* @param ppvR0 Where to return the address of the ring-0 mapping on
2811	* success.
2812	*/
2813	SUPR0DECL(int) SUPR0PageMapKernel(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t offSub, uint32_t cbSub,
2814	uint32_t fFlags, PRTR0PTR ppvR0)
2815	{
2816	int rc;
2817	PSUPDRVBUNDLE pBundle;
2818	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2819	RTR0MEMOBJ hMemObj = NIL_RTR0MEMOBJ;
2820	LogFlow(("SUPR0PageMapKernel: pSession=%p pvR3=%p offSub=%#x cbSub=%#x\n", pSession, pvR3, offSub, cbSub));
2821
2822	/*
2823	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2824	*/
2825	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2826	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2827	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2828	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2829	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2830	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
2831
2832	/*
2833	* Find the memory object.
2834	*/
2835	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2836	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2837	{
2838	if (pBundle->cUsed > 0)
2839	{
2840	unsigned i;
2841	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2842	{
2843	if ( ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
2844	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2845	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2846	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
2847	\|\| ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED
2848	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2849	&& pBundle->aMem[i].MapObjR3 == NIL_RTR0MEMOBJ
2850	&& RTR0MemObjAddressR3(pBundle->aMem[i].MemObj) == pvR3))
2851	{
2852	hMemObj = pBundle->aMem[i].MemObj;
2853	break;
2854	}
2855	}
2856	}
2857	}
2858	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2859
2860	rc = VERR_INVALID_PARAMETER;
2861	if (hMemObj != NIL_RTR0MEMOBJ)
2862	{
2863	/*
2864	* Do some further input validations before calling IPRT.
2865	* (Cleanup is done indirectly by telling RTR0MemObjFree to include mappings.)
2866	*/
2867	size_t cbMemObj = RTR0MemObjSize(hMemObj);
2868	if ( offSub < cbMemObj
2869	&& cbSub <= cbMemObj
2870	&& offSub + cbSub <= cbMemObj)
2871	{
2872	RTR0MEMOBJ hMapObj;
2873	rc = RTR0MemObjMapKernelEx(&hMapObj, hMemObj, (void *)-1, 0,
2874	RTMEM_PROT_READ \| RTMEM_PROT_WRITE, offSub, cbSub);
2875	if (RT_SUCCESS(rc))
2876	*ppvR0 = RTR0MemObjAddress(hMapObj);
2877	}
2878	else
2879	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
2880
2881	}
2882	return rc;
2883	}
2884
2885
2886	/**
2887	* Changes the page level protection of one or more pages previously allocated
2888	* by SUPR0PageAllocEx.
2889	*
2890	* @returns IPRT status code.
2891	* @param pSession The session to associated the allocation with.
2892	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
2893	* NIL_RTR3PTR if the ring-3 mapping should be unaffected.
2894	* @param pvR0 The ring-0 address returned by SUPR0PageAllocEx.
2895	* NIL_RTR0PTR if the ring-0 mapping should be unaffected.
2896	* @param offSub Where to start changing. Must be page aligned.
2897	* @param cbSub How much to change. Must be page aligned.
2898	* @param fProt The new page level protection, see RTMEM_PROT_*.
2899	*/
2900	SUPR0DECL(int) SUPR0PageProtect(PSUPDRVSESSION pSession, RTR3PTR pvR3, RTR0PTR pvR0, uint32_t offSub, uint32_t cbSub, uint32_t fProt)
2901	{
2902	int rc;
2903	PSUPDRVBUNDLE pBundle;
2904	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2905	RTR0MEMOBJ hMemObjR0 = NIL_RTR0MEMOBJ;
2906	RTR0MEMOBJ hMemObjR3 = NIL_RTR0MEMOBJ;
2907	LogFlow(("SUPR0PageProtect: pSession=%p pvR3=%p pvR0=%p offSub=%#x cbSub=%#x fProt-%#x\n", pSession, pvR3, pvR0, offSub, cbSub, fProt));
2908
2909	/*
2910	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2911	*/
2912	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2913	AssertReturn(!(fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)), VERR_INVALID_PARAMETER);
2914	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2915	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2916	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
2917
2918	/*
2919	* Find the memory object.
2920	*/
2921	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2922	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2923	{
2924	if (pBundle->cUsed > 0)
2925	{
2926	unsigned i;
2927	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2928	{
2929	if ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
2930	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2931	&& ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2932	\|\| pvR3 == NIL_RTR3PTR)
2933	&& ( pvR0 == NIL_RTR0PTR
2934	\|\| RTR0MemObjAddress(pBundle->aMem[i].MemObj) == pvR0)
2935	&& ( pvR3 == NIL_RTR3PTR
2936	\|\| RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3))
2937	{
2938	if (pvR0 != NIL_RTR0PTR)
2939	hMemObjR0 = pBundle->aMem[i].MemObj;
2940	if (pvR3 != NIL_RTR3PTR)
2941	hMemObjR3 = pBundle->aMem[i].MapObjR3;
2942	break;
2943	}
2944	}
2945	}
2946	}
2947	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2948
2949	rc = VERR_INVALID_PARAMETER;
2950	if ( hMemObjR0 != NIL_RTR0MEMOBJ
2951	\|\| hMemObjR3 != NIL_RTR0MEMOBJ)
2952	{
2953	/*
2954	* Do some further input validations before calling IPRT.
2955	*/
2956	size_t cbMemObj = hMemObjR0 != NIL_RTR0PTR ? RTR0MemObjSize(hMemObjR0) : RTR0MemObjSize(hMemObjR3);
2957	if ( offSub < cbMemObj
2958	&& cbSub <= cbMemObj
2959	&& offSub + cbSub <= cbMemObj)
2960	{
2961	rc = VINF_SUCCESS;
2962	if (hMemObjR3 != NIL_RTR0PTR)
2963	rc = RTR0MemObjProtect(hMemObjR3, offSub, cbSub, fProt);
2964	if (hMemObjR0 != NIL_RTR0PTR && RT_SUCCESS(rc))
2965	rc = RTR0MemObjProtect(hMemObjR0, offSub, cbSub, fProt);
2966	}
2967	else
2968	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
2969
2970	}
2971	return rc;
2972
2973	}
2974
2975
2976	/**
2977	* Free memory allocated by SUPR0PageAlloc() and SUPR0PageAllocEx().
2978	*
2979	* @returns IPRT status code.
2980	* @param pSession The session owning the allocation.
2981	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc() or
2982	* SUPR0PageAllocEx().
2983	*/
2984	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2985	{
2986	LogFlow(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2987	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2988	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_PAGE);
2989	}
2990
2991
2992	/**
2993	* Gets the paging mode of the current CPU.
2994	*
2995	* @returns Paging mode, SUPPAGEINGMODE_INVALID on error.
2996	*/
2997	SUPR0DECL(SUPPAGINGMODE) SUPR0GetPagingMode(void)
2998	{
2999	SUPPAGINGMODE enmMode;
3000
3001	RTR0UINTREG cr0 = ASMGetCR0();
3002	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
3003	enmMode = SUPPAGINGMODE_INVALID;
3004	else
3005	{
3006	RTR0UINTREG cr4 = ASMGetCR4();
3007	uint32_t fNXEPlusLMA = 0;
3008	if (cr4 & X86_CR4_PAE)
3009	{
3010	uint32_t fAmdFeatures = ASMCpuId_EDX(0x80000001);
3011	if (fAmdFeatures & (X86_CPUID_AMD_FEATURE_EDX_NX \| X86_CPUID_AMD_FEATURE_EDX_LONG_MODE))
3012	{
3013	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
3014	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
3015	fNXEPlusLMA \|= RT_BIT(0);
3016	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
3017	fNXEPlusLMA \|= RT_BIT(1);
3018	}
3019	}
3020
3021	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
3022	{
3023	case 0:
3024	enmMode = SUPPAGINGMODE_32_BIT;
3025	break;
3026
3027	case X86_CR4_PGE:
3028	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
3029	break;
3030
3031	case X86_CR4_PAE:
3032	enmMode = SUPPAGINGMODE_PAE;
3033	break;
3034
3035	case X86_CR4_PAE \| RT_BIT(0):
3036	enmMode = SUPPAGINGMODE_PAE_NX;
3037	break;
3038
3039	case X86_CR4_PAE \| X86_CR4_PGE:
3040	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3041	break;
3042
3043	case X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3044	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3045	break;
3046
3047	case RT_BIT(1) \| X86_CR4_PAE:
3048	enmMode = SUPPAGINGMODE_AMD64;
3049	break;
3050
3051	case RT_BIT(1) \| X86_CR4_PAE \| RT_BIT(0):
3052	enmMode = SUPPAGINGMODE_AMD64_NX;
3053	break;
3054
3055	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
3056	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
3057	break;
3058
3059	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3060	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
3061	break;
3062
3063	default:
3064	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
3065	enmMode = SUPPAGINGMODE_INVALID;
3066	break;
3067	}
3068	}
3069	return enmMode;
3070	}
3071
3072
3073	/**
3074	* Enables or disabled hardware virtualization extensions using native OS APIs.
3075	*
3076	* @returns VBox status code.
3077	* @retval VINF_SUCCESS on success.
3078	* @retval VERR_NOT_SUPPORTED if not supported by the native OS.
3079	*
3080	* @param fEnable Whether to enable or disable.
3081	*/
3082	SUPR0DECL(int) SUPR0EnableVTx(bool fEnable)
3083	{
3084	#ifdef RT_OS_DARWIN
3085	return supdrvOSEnableVTx(fEnable);
3086	#else
3087	return VERR_NOT_SUPPORTED;
3088	#endif
3089	}
3090
3091
3092	/** @todo document me */
3093	SUPR0DECL(int) SUPR0QueryVTCaps(PSUPDRVSESSION pSession, uint32_t *pfCaps)
3094	{
3095	/*
3096	* Input validation.
3097	*/
3098	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3099	AssertPtrReturn(pfCaps, VERR_INVALID_POINTER);
3100
3101	*pfCaps = 0;
3102
3103	if (ASMHasCpuId())
3104	{
3105	uint32_t u32FeaturesECX;
3106	uint32_t u32Dummy;
3107	uint32_t u32FeaturesEDX;
3108	uint32_t u32VendorEBX, u32VendorECX, u32VendorEDX, u32AMDFeatureEDX, u32AMDFeatureECX;
3109	uint64_t val;
3110
3111	ASMCpuId(0, &u32Dummy, &u32VendorEBX, &u32VendorECX, &u32VendorEDX);
3112	ASMCpuId(1, &u32Dummy, &u32Dummy, &u32FeaturesECX, &u32FeaturesEDX);
3113	/* Query AMD features. */
3114	ASMCpuId(0x80000001, &u32Dummy, &u32Dummy, &u32AMDFeatureECX, &u32AMDFeatureEDX);
3115
3116	if ( u32VendorEBX == X86_CPUID_VENDOR_INTEL_EBX
3117	&& u32VendorECX == X86_CPUID_VENDOR_INTEL_ECX
3118	&& u32VendorEDX == X86_CPUID_VENDOR_INTEL_EDX
3119	)
3120	{
3121	if ( (u32FeaturesECX & X86_CPUID_FEATURE_ECX_VMX)
3122	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3123	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3124	)
3125	{
3126	val = ASMRdMsr(MSR_IA32_FEATURE_CONTROL);
3127	/*
3128	* Both the LOCK and VMXON bit must be set; otherwise VMXON will generate a #GP.
3129	* Once the lock bit is set, this MSR can no longer be modified.
3130	*/
3131	if ( (val & (MSR_IA32_FEATURE_CONTROL_VMXON\|MSR_IA32_FEATURE_CONTROL_LOCK))
3132	== (MSR_IA32_FEATURE_CONTROL_VMXON\|MSR_IA32_FEATURE_CONTROL_LOCK) /* enabled and locked */
3133	\|\| !(val & MSR_IA32_FEATURE_CONTROL_LOCK) /* not enabled, but not locked either */
3134	)
3135	{
3136	VMX_CAPABILITY vtCaps;
3137
3138	*pfCaps \|= SUPVTCAPS_VT_X;
3139
3140	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS);
3141	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC_USE_SECONDARY_EXEC_CTRL)
3142	{
3143	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS2);
3144	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC2_EPT)
3145	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3146	}
3147	return VINF_SUCCESS;
3148	}
3149	return VERR_VMX_MSR_LOCKED_OR_DISABLED;
3150	}
3151	return VERR_VMX_NO_VMX;
3152	}
3153
3154	if ( u32VendorEBX == X86_CPUID_VENDOR_AMD_EBX
3155	&& u32VendorECX == X86_CPUID_VENDOR_AMD_ECX
3156	&& u32VendorEDX == X86_CPUID_VENDOR_AMD_EDX
3157	)
3158	{
3159	if ( (u32AMDFeatureECX & X86_CPUID_AMD_FEATURE_ECX_SVM)
3160	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3161	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3162	)
3163	{
3164	/* Check if SVM is disabled */
3165	val = ASMRdMsr(MSR_K8_VM_CR);
3166	if (!(val & MSR_K8_VM_CR_SVM_DISABLE))
3167	{
3168	*pfCaps \|= SUPVTCAPS_AMD_V;
3169
3170	/* Query AMD features. */
3171	ASMCpuId(0x8000000A, &u32Dummy, &u32Dummy, &u32Dummy, &u32FeaturesEDX);
3172
3173	if (u32FeaturesEDX & AMD_CPUID_SVM_FEATURE_EDX_NESTED_PAGING)
3174	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3175
3176	return VINF_SUCCESS;
3177	}
3178	return VERR_SVM_DISABLED;
3179	}
3180	return VERR_SVM_NO_SVM;
3181	}
3182	}
3183
3184	return VERR_UNSUPPORTED_CPU;
3185	}
3186
3187
3188	/**
3189	* (Re-)initializes the per-cpu structure prior to starting or resuming the GIP
3190	* updating.
3191	*
3192	* @param pGipCpu The per CPU structure for this CPU.
3193	* @param u64NanoTS The current time.
3194	*/
3195	static void supdrvGipReInitCpu(PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
3196	{
3197	pGipCpu->u64TSC = ASMReadTSC() - pGipCpu->u32UpdateIntervalTSC;
3198	pGipCpu->u64NanoTS = u64NanoTS;
3199	}
3200
3201
3202	/**
3203	* Set the current TSC and NanoTS value for the CPU.
3204	*
3205	* @param idCpu The CPU ID. Unused - we have to use the APIC ID.
3206	* @param pvUser1 Pointer to the ring-0 GIP mapping.
3207	* @param pvUser2 Pointer to the variable holding the current time.
3208	*/
3209	static DECLCALLBACK(void) supdrvGipReInitCpuCallback(RTCPUID idCpu, void pvUser1, void pvUser2)
3210	{
3211	PSUPGLOBALINFOPAGE pGip = (PSUPGLOBALINFOPAGE)pvUser1;
3212	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
3213
3214	if (RT_LIKELY(iCpu < pGip->cCpus && pGip->aCPUs[iCpu].idCpu == idCpu))
3215	supdrvGipReInitCpu(&pGip->aCPUs[iCpu], (uint64_t )pvUser2);
3216
3217	NOREF(pvUser2);
3218	NOREF(idCpu);
3219	}
3220
3221
3222	/**
3223	* Maps the GIP into userspace and/or get the physical address of the GIP.
3224	*
3225	* @returns IPRT status code.
3226	* @param pSession Session to which the GIP mapping should belong.
3227	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
3228	* @param pHCPhysGip Where to store the physical address. (optional)
3229	*
3230	* @remark There is no reference counting on the mapping, so one call to this function
3231	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
3232	* and remove the session as a GIP user.
3233	*/
3234	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
3235	{
3236	int rc;
3237	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3238	RTR3PTR pGipR3 = NIL_RTR3PTR;
3239	RTHCPHYS HCPhys = NIL_RTHCPHYS;
3240	LogFlow(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
3241
3242	/*
3243	* Validate
3244	*/
3245	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3246	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
3247	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
3248
3249	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3250	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3251	#else
3252	RTSemFastMutexRequest(pDevExt->mtxGip);
3253	#endif
3254	if (pDevExt->pGip)
3255	{
3256	/*
3257	* Map it?
3258	*/
3259	rc = VINF_SUCCESS;
3260	if (ppGipR3)
3261	{
3262	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
3263	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
3264	RTMEM_PROT_READ, RTR0ProcHandleSelf());
3265	if (RT_SUCCESS(rc))
3266	pGipR3 = RTR0MemObjAddressR3(pSession->GipMapObjR3);
3267	}
3268
3269	/*
3270	* Get physical address.
3271	*/
3272	if (pHCPhysGip && RT_SUCCESS(rc))
3273	HCPhys = pDevExt->HCPhysGip;
3274
3275	/*
3276	* Reference globally.
3277	*/
3278	if (!pSession->fGipReferenced && RT_SUCCESS(rc))
3279	{
3280	pSession->fGipReferenced = 1;
3281	pDevExt->cGipUsers++;
3282	if (pDevExt->cGipUsers == 1)
3283	{
3284	PSUPGLOBALINFOPAGE pGipR0 = pDevExt->pGip;
3285	uint64_t u64NanoTS;
3286	uint32_t u32SystemResolution;
3287	unsigned i;
3288
3289	LogFlow(("SUPR0GipMap: Resumes GIP updating\n"));
3290
3291	/*
3292	* Try bump up the system timer resolution.
3293	* The more interrupts the better...
3294	*/
3295	if ( RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
3296	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
3297	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1953125 /* 512 HZ */, &u32SystemResolution))
3298	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 2000000 /* 500 HZ */, &u32SystemResolution))
3299	)
3300	{
3301	Assert(RTTimerGetSystemGranularity() <= u32SystemResolution);
3302	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
3303	}
3304
3305	if (pGipR0->aCPUs[0].u32TransactionId != 2 /* not the first time */)
3306	{
3307	for (i = 0; i < RT_ELEMENTS(pGipR0->aCPUs); i++)
3308	ASMAtomicUoWriteU32(&pGipR0->aCPUs[i].u32TransactionId,
3309	(pGipR0->aCPUs[i].u32TransactionId + GIP_UPDATEHZ_RECALC_FREQ * 2)
3310	& ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
3311	ASMAtomicWriteU64(&pGipR0->u64NanoTSLastUpdateHz, 0);
3312	}
3313
3314	u64NanoTS = RTTimeSystemNanoTS() - pGipR0->u32UpdateIntervalNS;
3315	if ( pGipR0->u32Mode == SUPGIPMODE_SYNC_TSC
3316	\|\| RTMpGetOnlineCount() == 1)
3317	supdrvGipReInitCpu(&pGipR0->aCPUs[0], u64NanoTS);
3318	else
3319	RTMpOnAll(supdrvGipReInitCpuCallback, pGipR0, &u64NanoTS);
3320
3321	#ifndef DO_NOT_START_GIP
3322	rc = RTTimerStart(pDevExt->pGipTimer, 0); AssertRC(rc);
3323	#endif
3324	rc = VINF_SUCCESS;
3325	}
3326	}
3327	}
3328	else
3329	{
3330	rc = VERR_GENERAL_FAILURE;
3331	Log(("SUPR0GipMap: GIP is not available!\n"));
3332	}
3333	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3334	RTSemMutexRelease(pDevExt->mtxGip);
3335	#else
3336	RTSemFastMutexRelease(pDevExt->mtxGip);
3337	#endif
3338
3339	/*
3340	* Write returns.
3341	*/
3342	if (pHCPhysGip)
3343	*pHCPhysGip = HCPhys;
3344	if (ppGipR3)
3345	*ppGipR3 = pGipR3;
3346
3347	#ifdef DEBUG_DARWIN_GIP
3348	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3349	#else
3350	LogFlow(( "SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3351	#endif
3352	return rc;
3353	}
3354
3355
3356	/**
3357	* Unmaps any user mapping of the GIP and terminates all GIP access
3358	* from this session.
3359	*
3360	* @returns IPRT status code.
3361	* @param pSession Session to which the GIP mapping should belong.
3362	*/
3363	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
3364	{
3365	int rc = VINF_SUCCESS;
3366	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3367	#ifdef DEBUG_DARWIN_GIP
3368	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
3369	pSession,
3370	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
3371	pSession->GipMapObjR3));
3372	#else
3373	LogFlow(("SUPR0GipUnmap: pSession=%p\n", pSession));
3374	#endif
3375	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3376
3377	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3378	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3379	#else
3380	RTSemFastMutexRequest(pDevExt->mtxGip);
3381	#endif
3382
3383	/*
3384	* Unmap anything?
3385	*/
3386	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
3387	{
3388	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
3389	AssertRC(rc);
3390	if (RT_SUCCESS(rc))
3391	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
3392	}
3393
3394	/*
3395	* Dereference global GIP.
3396	*/
3397	if (pSession->fGipReferenced && !rc)
3398	{
3399	pSession->fGipReferenced = 0;
3400	if ( pDevExt->cGipUsers > 0
3401	&& !--pDevExt->cGipUsers)
3402	{
3403	LogFlow(("SUPR0GipUnmap: Suspends GIP updating\n"));
3404	#ifndef DO_NOT_START_GIP
3405	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = VINF_SUCCESS;
3406	#endif
3407
3408	if (pDevExt->u32SystemTimerGranularityGrant)
3409	{
3410	int rc2 = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant);
3411	AssertRC(rc2);
3412	pDevExt->u32SystemTimerGranularityGrant = 0;
3413	}
3414	}
3415	}
3416
3417	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3418	RTSemMutexRelease(pDevExt->mtxGip);
3419	#else
3420	RTSemFastMutexRelease(pDevExt->mtxGip);
3421	#endif
3422
3423	return rc;
3424	}
3425
3426
3427	/**
3428	* Gets the GIP pointer.
3429	*
3430	* @returns Pointer to the GIP or NULL.
3431	*/
3432	SUPDECL(PSUPGLOBALINFOPAGE) SUPGetGIP(void)
3433	{
3434	return g_pSUPGlobalInfoPage;
3435	}
3436
3437
3438	/**
3439	* Register a component factory with the support driver.
3440	*
3441	* This is currently restricted to kernel sessions only.
3442	*
3443	* @returns VBox status code.
3444	* @retval VINF_SUCCESS on success.
3445	* @retval VERR_NO_MEMORY if we're out of memory.
3446	* @retval VERR_ALREADY_EXISTS if the factory has already been registered.
3447	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3448	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3449	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3450	*
3451	* @param pSession The SUPDRV session (must be a ring-0 session).
3452	* @param pFactory Pointer to the component factory registration structure.
3453	*
3454	* @remarks This interface is also available via SUPR0IdcComponentRegisterFactory.
3455	*/
3456	SUPR0DECL(int) SUPR0ComponentRegisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3457	{
3458	PSUPDRVFACTORYREG pNewReg;
3459	const char *psz;
3460	int rc;
3461
3462	/*
3463	* Validate parameters.
3464	*/
3465	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3466	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3467	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3468	AssertPtrReturn(pFactory->pfnQueryFactoryInterface, VERR_INVALID_POINTER);
3469	psz = RTStrEnd(pFactory->szName, sizeof(pFactory->szName));
3470	AssertReturn(psz, VERR_INVALID_PARAMETER);
3471
3472	/*
3473	* Allocate and initialize a new registration structure.
3474	*/
3475	pNewReg = (PSUPDRVFACTORYREG)RTMemAlloc(sizeof(SUPDRVFACTORYREG));
3476	if (pNewReg)
3477	{
3478	pNewReg->pNext = NULL;
3479	pNewReg->pFactory = pFactory;
3480	pNewReg->pSession = pSession;
3481	pNewReg->cchName = psz - &pFactory->szName[0];
3482
3483	/*
3484	* Add it to the tail of the list after checking for prior registration.
3485	*/
3486	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3487	if (RT_SUCCESS(rc))
3488	{
3489	PSUPDRVFACTORYREG pPrev = NULL;
3490	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3491	while (pCur && pCur->pFactory != pFactory)
3492	{
3493	pPrev = pCur;
3494	pCur = pCur->pNext;
3495	}
3496	if (!pCur)
3497	{
3498	if (pPrev)
3499	pPrev->pNext = pNewReg;
3500	else
3501	pSession->pDevExt->pComponentFactoryHead = pNewReg;
3502	rc = VINF_SUCCESS;
3503	}
3504	else
3505	rc = VERR_ALREADY_EXISTS;
3506
3507	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3508	}
3509
3510	if (RT_FAILURE(rc))
3511	RTMemFree(pNewReg);
3512	}
3513	else
3514	rc = VERR_NO_MEMORY;
3515	return rc;
3516	}
3517
3518
3519	/**
3520	* Deregister a component factory.
3521	*
3522	* @returns VBox status code.
3523	* @retval VINF_SUCCESS on success.
3524	* @retval VERR_NOT_FOUND if the factory wasn't registered.
3525	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3526	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3527	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3528	*
3529	* @param pSession The SUPDRV session (must be a ring-0 session).
3530	* @param pFactory Pointer to the component factory registration structure
3531	* previously passed SUPR0ComponentRegisterFactory().
3532	*
3533	* @remarks This interface is also available via SUPR0IdcComponentDeregisterFactory.
3534	*/
3535	SUPR0DECL(int) SUPR0ComponentDeregisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3536	{
3537	int rc;
3538
3539	/*
3540	* Validate parameters.
3541	*/
3542	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3543	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3544	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3545
3546	/*
3547	* Take the lock and look for the registration record.
3548	*/
3549	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3550	if (RT_SUCCESS(rc))
3551	{
3552	PSUPDRVFACTORYREG pPrev = NULL;
3553	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3554	while (pCur && pCur->pFactory != pFactory)
3555	{
3556	pPrev = pCur;
3557	pCur = pCur->pNext;
3558	}
3559	if (pCur)
3560	{
3561	if (!pPrev)
3562	pSession->pDevExt->pComponentFactoryHead = pCur->pNext;
3563	else
3564	pPrev->pNext = pCur->pNext;
3565
3566	pCur->pNext = NULL;
3567	pCur->pFactory = NULL;
3568	pCur->pSession = NULL;
3569	rc = VINF_SUCCESS;
3570	}
3571	else
3572	rc = VERR_NOT_FOUND;
3573
3574	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3575
3576	RTMemFree(pCur);
3577	}
3578	return rc;
3579	}
3580
3581
3582	/**
3583	* Queries a component factory.
3584	*
3585	* @returns VBox status code.
3586	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3587	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3588	* @retval VERR_SUPDRV_COMPONENT_NOT_FOUND if the component factory wasn't found.
3589	* @retval VERR_SUPDRV_INTERFACE_NOT_SUPPORTED if the interface wasn't supported.
3590	*
3591	* @param pSession The SUPDRV session.
3592	* @param pszName The name of the component factory.
3593	* @param pszInterfaceUuid The UUID of the factory interface (stringified).
3594	* @param ppvFactoryIf Where to store the factory interface.
3595	*/
3596	SUPR0DECL(int) SUPR0ComponentQueryFactory(PSUPDRVSESSION pSession, const char pszName, const char pszInterfaceUuid, void **ppvFactoryIf)
3597	{
3598	const char *pszEnd;
3599	size_t cchName;
3600	int rc;
3601
3602	/*
3603	* Validate parameters.
3604	*/
3605	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3606
3607	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
3608	pszEnd = RTStrEnd(pszName, RT_SIZEOFMEMB(SUPDRVFACTORY, szName));
3609	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3610	cchName = pszEnd - pszName;
3611
3612	AssertPtrReturn(pszInterfaceUuid, VERR_INVALID_POINTER);
3613	pszEnd = RTStrEnd(pszInterfaceUuid, RTUUID_STR_LENGTH);
3614	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3615
3616	AssertPtrReturn(ppvFactoryIf, VERR_INVALID_POINTER);
3617	*ppvFactoryIf = NULL;
3618
3619	/*
3620	* Take the lock and try all factories by this name.
3621	*/
3622	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3623	if (RT_SUCCESS(rc))
3624	{
3625	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3626	rc = VERR_SUPDRV_COMPONENT_NOT_FOUND;
3627	while (pCur)
3628	{
3629	if ( pCur->cchName == cchName
3630	&& !memcmp(pCur->pFactory->szName, pszName, cchName))
3631	{
3632	void *pvFactory = pCur->pFactory->pfnQueryFactoryInterface(pCur->pFactory, pSession, pszInterfaceUuid);
3633	if (pvFactory)
3634	{
3635	*ppvFactoryIf = pvFactory;
3636	rc = VINF_SUCCESS;
3637	break;
3638	}
3639	rc = VERR_SUPDRV_INTERFACE_NOT_SUPPORTED;
3640	}
3641
3642	/* next */
3643	pCur = pCur->pNext;
3644	}
3645
3646	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3647	}
3648	return rc;
3649	}
3650
3651
3652	/**
3653	* Adds a memory object to the session.
3654	*
3655	* @returns IPRT status code.
3656	* @param pMem Memory tracking structure containing the
3657	* information to track.
3658	* @param pSession The session.
3659	*/
3660	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
3661	{
3662	PSUPDRVBUNDLE pBundle;
3663	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3664
3665	/*
3666	* Find free entry and record the allocation.
3667	*/
3668	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3669	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3670	{
3671	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
3672	{
3673	unsigned i;
3674	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3675	{
3676	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
3677	{
3678	pBundle->cUsed++;
3679	pBundle->aMem[i] = *pMem;
3680	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3681	return VINF_SUCCESS;
3682	}
3683	}
3684	AssertFailed(); /* !!this can't be happening!!! */
3685	}
3686	}
3687	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3688
3689	/*
3690	* Need to allocate a new bundle.
3691	* Insert into the last entry in the bundle.
3692	*/
3693	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
3694	if (!pBundle)
3695	return VERR_NO_MEMORY;
3696
3697	/* take last entry. */
3698	pBundle->cUsed++;
3699	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
3700
3701	/* insert into list. */
3702	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3703	pBundle->pNext = pSession->Bundle.pNext;
3704	pSession->Bundle.pNext = pBundle;
3705	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3706
3707	return VINF_SUCCESS;
3708	}
3709
3710
3711	/**
3712	* Releases a memory object referenced by pointer and type.
3713	*
3714	* @returns IPRT status code.
3715	* @param pSession Session data.
3716	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
3717	* @param eType Memory type.
3718	*/
3719	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
3720	{
3721	PSUPDRVBUNDLE pBundle;
3722	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3723
3724	/*
3725	* Validate input.
3726	*/
3727	if (!uPtr)
3728	{
3729	Log(("Illegal address %p\n", (void *)uPtr));
3730	return VERR_INVALID_PARAMETER;
3731	}
3732
3733	/*
3734	* Search for the address.
3735	*/
3736	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3737	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3738	{
3739	if (pBundle->cUsed > 0)
3740	{
3741	unsigned i;
3742	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3743	{
3744	if ( pBundle->aMem[i].eType == eType
3745	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3746	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
3747	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3748	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
3749	)
3750	{
3751	/* Make a copy of it and release it outside the spinlock. */
3752	SUPDRVMEMREF Mem = pBundle->aMem[i];
3753	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
3754	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
3755	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
3756	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3757
3758	if (Mem.MapObjR3 != NIL_RTR0MEMOBJ)
3759	{
3760	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
3761	AssertRC(rc); /** @todo figure out how to handle this. */
3762	}
3763	if (Mem.MemObj != NIL_RTR0MEMOBJ)
3764	{
3765	int rc = RTR0MemObjFree(Mem.MemObj, true /* fFreeMappings */);
3766	AssertRC(rc); /** @todo figure out how to handle this. */
3767	}
3768	return VINF_SUCCESS;
3769	}
3770	}
3771	}
3772	}
3773	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3774	Log(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
3775	return VERR_INVALID_PARAMETER;
3776	}
3777
3778
3779	/**
3780	* Opens an image. If it's the first time it's opened the call must upload
3781	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
3782	*
3783	* This is the 1st step of the loading.
3784	*
3785	* @returns IPRT status code.
3786	* @param pDevExt Device globals.
3787	* @param pSession Session data.
3788	* @param pReq The open request.
3789	*/
3790	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
3791	{
3792	int rc;
3793	PSUPDRVLDRIMAGE pImage;
3794	void *pv;
3795	size_t cchName = strlen(pReq->u.In.szName); /* (caller checked < 32). */
3796	LogFlow(("supdrvIOCtl_LdrOpen: szName=%s cbImageWithTabs=%d\n", pReq->u.In.szName, pReq->u.In.cbImageWithTabs));
3797
3798	/*
3799	* Check if we got an instance of the image already.
3800	*/
3801	supdrvLdrLock(pDevExt);
3802	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
3803	{
3804	if ( pImage->szName[cchName] == '\0'
3805	&& !memcmp(pImage->szName, pReq->u.In.szName, cchName))
3806	{
3807	/** @todo check cbImageBits and cbImageWithTabs here, if they differs that indicates that the images are different. */
3808	pImage->cUsage++;
3809	pReq->u.Out.pvImageBase = pImage->pvImage;
3810	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
3811	pReq->u.Out.fNativeLoader = pImage->fNative;
3812	supdrvLdrAddUsage(pSession, pImage);
3813	supdrvLdrUnlock(pDevExt);
3814	return VINF_SUCCESS;
3815	}
3816	}
3817	/* (not found - add it!) */
3818
3819	/*
3820	* Allocate memory.
3821	*/
3822	pv = RTMemAlloc(RT_OFFSETOF(SUPDRVLDRIMAGE, szName[cchName + 1]));
3823	if (!pv)
3824	{
3825	supdrvLdrUnlock(pDevExt);
3826	Log(("supdrvIOCtl_LdrOpen: RTMemAlloc() failed\n"));
3827	return /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_2;
3828	}
3829
3830	/*
3831	* Setup and link in the LDR stuff.
3832	*/
3833	pImage = (PSUPDRVLDRIMAGE)pv;
3834	pImage->pvImage = NULL;
3835	pImage->pvImageAlloc = NULL;
3836	pImage->cbImageWithTabs = pReq->u.In.cbImageWithTabs;
3837	pImage->cbImageBits = pReq->u.In.cbImageBits;
3838	pImage->cSymbols = 0;
3839	pImage->paSymbols = NULL;
3840	pImage->pachStrTab = NULL;
3841	pImage->cbStrTab = 0;
3842	pImage->pfnModuleInit = NULL;
3843	pImage->pfnModuleTerm = NULL;
3844	pImage->pfnServiceReqHandler = NULL;
3845	pImage->uState = SUP_IOCTL_LDR_OPEN;
3846	pImage->cUsage = 1;
3847	memcpy(pImage->szName, pReq->u.In.szName, cchName + 1);
3848
3849	/*
3850	* Try load it using the native loader, if that isn't supported, fall back
3851	* on the older method.
3852	*/
3853	pImage->fNative = true;
3854	rc = supdrvOSLdrOpen(pDevExt, pImage, pReq->u.In.szFilename);
3855	if (rc == VERR_NOT_SUPPORTED)
3856	{
3857	pImage->pvImageAlloc = RTMemExecAlloc(pImage->cbImageBits + 31);
3858	pImage->pvImage = RT_ALIGN_P(pImage->pvImageAlloc, 32);
3859	pImage->fNative = false;
3860	rc = pImage->pvImageAlloc ? VINF_SUCCESS : VERR_NO_EXEC_MEMORY;
3861	}
3862	if (RT_FAILURE(rc))
3863	{
3864	supdrvLdrUnlock(pDevExt);
3865	RTMemFree(pImage);
3866	Log(("supdrvIOCtl_LdrOpen(%s): failed - %Rrc\n", pReq->u.In.szName, rc));
3867	return rc;
3868	}
3869	Assert(VALID_PTR(pImage->pvImage) \|\| RT_FAILURE(rc));
3870
3871	/*
3872	* Link it.
3873	*/
3874	pImage->pNext = pDevExt->pLdrImages;
3875	pDevExt->pLdrImages = pImage;
3876
3877	supdrvLdrAddUsage(pSession, pImage);
3878
3879	pReq->u.Out.pvImageBase = pImage->pvImage;
3880	pReq->u.Out.fNeedsLoading = true;
3881	pReq->u.Out.fNativeLoader = pImage->fNative;
3882	supdrvLdrUnlock(pDevExt);
3883
3884	#if defined(RT_OS_WINDOWS) && defined(DEBUG)
3885	SUPR0Printf("VBoxDrv: windbg> .reload /f %s=%#p\n", pImage->szName, pImage->pvImage);
3886	#endif
3887	return VINF_SUCCESS;
3888	}
3889
3890
3891	/**
3892	* Worker that validates a pointer to an image entrypoint.
3893	*
3894	* @returns IPRT status code.
3895	* @param pDevExt The device globals.
3896	* @param pImage The loader image.
3897	* @param pv The pointer into the image.
3898	* @param fMayBeNull Whether it may be NULL.
3899	* @param pszWhat What is this entrypoint? (for logging)
3900	* @param pbImageBits The image bits prepared by ring-3.
3901	*
3902	* @remarks Will leave the lock on failure.
3903	*/
3904	static int supdrvLdrValidatePointer(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage, void *pv,
3905	bool fMayBeNull, const uint8_t pbImageBits, const char pszWhat)
3906	{
3907	if (!fMayBeNull \|\| pv)
3908	{
3909	if ((uintptr_t)pv - (uintptr_t)pImage->pvImage >= pImage->cbImageBits)
3910	{
3911	supdrvLdrUnlock(pDevExt);
3912	Log(("Out of range (%p LB %#x): %s=%p\n", pImage->pvImage, pImage->cbImageBits, pszWhat, pv));
3913	return VERR_INVALID_PARAMETER;
3914	}
3915
3916	if (pImage->fNative)
3917	{
3918	int rc = supdrvOSLdrValidatePointer(pDevExt, pImage, pv, pbImageBits);
3919	if (RT_FAILURE(rc))
3920	{
3921	supdrvLdrUnlock(pDevExt);
3922	Log(("Bad entry point address: %s=%p (rc=%Rrc)\n", pszWhat, pv, rc));
3923	return rc;
3924	}
3925	}
3926	}
3927	return VINF_SUCCESS;
3928	}
3929
3930
3931	/**
3932	* Loads the image bits.
3933	*
3934	* This is the 2nd step of the loading.
3935	*
3936	* @returns IPRT status code.
3937	* @param pDevExt Device globals.
3938	* @param pSession Session data.
3939	* @param pReq The request.
3940	*/
3941	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
3942	{
3943	PSUPDRVLDRUSAGE pUsage;
3944	PSUPDRVLDRIMAGE pImage;
3945	int rc;
3946	LogFlow(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImageWithBits=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImageWithTabs));
3947
3948	/*
3949	* Find the ldr image.
3950	*/
3951	supdrvLdrLock(pDevExt);
3952	pUsage = pSession->pLdrUsage;
3953	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3954	pUsage = pUsage->pNext;
3955	if (!pUsage)
3956	{
3957	supdrvLdrUnlock(pDevExt);
3958	Log(("SUP_IOCTL_LDR_LOAD: couldn't find image!\n"));
3959	return VERR_INVALID_HANDLE;
3960	}
3961	pImage = pUsage->pImage;
3962
3963	/*
3964	* Validate input.
3965	*/
3966	if ( pImage->cbImageWithTabs != pReq->u.In.cbImageWithTabs
3967	\|\| pImage->cbImageBits != pReq->u.In.cbImageBits)
3968	{
3969	supdrvLdrUnlock(pDevExt);
3970	Log(("SUP_IOCTL_LDR_LOAD: image size mismatch!! %d(prep) != %d(load) or %d != %d\n",
3971	pImage->cbImageWithTabs, pReq->u.In.cbImageWithTabs, pImage->cbImageBits, pReq->u.In.cbImageBits));
3972	return VERR_INVALID_HANDLE;
3973	}
3974
3975	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
3976	{
3977	unsigned uState = pImage->uState;
3978	supdrvLdrUnlock(pDevExt);
3979	if (uState != SUP_IOCTL_LDR_LOAD)
3980	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
3981	return VERR_ALREADY_LOADED;
3982	}
3983
3984	switch (pReq->u.In.eEPType)
3985	{
3986	case SUPLDRLOADEP_NOTHING:
3987	break;
3988
3989	case SUPLDRLOADEP_VMMR0:
3990	rc = supdrvLdrValidatePointer( pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0, false, pReq->u.In.abImage, "pvVMMR0");
3991	if (RT_SUCCESS(rc))
3992	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt, false, pReq->u.In.abImage, "pvVMMR0EntryInt");
3993	if (RT_SUCCESS(rc))
3994	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, false, pReq->u.In.abImage, "pvVMMR0EntryFast");
3995	if (RT_SUCCESS(rc))
3996	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx, false, pReq->u.In.abImage, "pvVMMR0EntryEx");
3997	if (RT_FAILURE(rc))
3998	return rc;
3999	break;
4000
4001	case SUPLDRLOADEP_SERVICE:
4002	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.Service.pfnServiceReq, false, pReq->u.In.abImage, "pfnServiceReq");
4003	if (RT_FAILURE(rc))
4004	return rc;
4005	if ( pReq->u.In.EP.Service.apvReserved[0] != NIL_RTR0PTR
4006	\|\| pReq->u.In.EP.Service.apvReserved[1] != NIL_RTR0PTR
4007	\|\| pReq->u.In.EP.Service.apvReserved[2] != NIL_RTR0PTR)
4008	{
4009	supdrvLdrUnlock(pDevExt);
4010	Log(("Out of range (%p LB %#x): apvReserved={%p,%p,%p} MBZ!\n",
4011	pImage->pvImage, pReq->u.In.cbImageWithTabs,
4012	pReq->u.In.EP.Service.apvReserved[0],
4013	pReq->u.In.EP.Service.apvReserved[1],
4014	pReq->u.In.EP.Service.apvReserved[2]));
4015	return VERR_INVALID_PARAMETER;
4016	}
4017	break;
4018
4019	default:
4020	supdrvLdrUnlock(pDevExt);
4021	Log(("Invalid eEPType=%d\n", pReq->u.In.eEPType));
4022	return VERR_INVALID_PARAMETER;
4023	}
4024
4025	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleInit, true, pReq->u.In.abImage, "pfnModuleInit");
4026	if (RT_FAILURE(rc))
4027	return rc;
4028	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleTerm, true, pReq->u.In.abImage, "pfnModuleTerm");
4029	if (RT_FAILURE(rc))
4030	return rc;
4031
4032	/*
4033	* Allocate and copy the tables.
4034	* (No need to do try/except as this is a buffered request.)
4035	*/
4036	pImage->cbStrTab = pReq->u.In.cbStrTab;
4037	if (pImage->cbStrTab)
4038	{
4039	pImage->pachStrTab = (char *)RTMemAlloc(pImage->cbStrTab);
4040	if (pImage->pachStrTab)
4041	memcpy(pImage->pachStrTab, &pReq->u.In.abImage[pReq->u.In.offStrTab], pImage->cbStrTab);
4042	else
4043	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_3;
4044	}
4045
4046	pImage->cSymbols = pReq->u.In.cSymbols;
4047	if (RT_SUCCESS(rc) && pImage->cSymbols)
4048	{
4049	size_t cbSymbols = pImage->cSymbols * sizeof(SUPLDRSYM);
4050	pImage->paSymbols = (PSUPLDRSYM)RTMemAlloc(cbSymbols);
4051	if (pImage->paSymbols)
4052	memcpy(pImage->paSymbols, &pReq->u.In.abImage[pReq->u.In.offSymbols], cbSymbols);
4053	else
4054	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_4;
4055	}
4056
4057	/*
4058	* Copy the bits / complete native loading.
4059	*/
4060	if (RT_SUCCESS(rc))
4061	{
4062	pImage->uState = SUP_IOCTL_LDR_LOAD;
4063	pImage->pfnModuleInit = pReq->u.In.pfnModuleInit;
4064	pImage->pfnModuleTerm = pReq->u.In.pfnModuleTerm;
4065
4066	if (pImage->fNative)
4067	rc = supdrvOSLdrLoad(pDevExt, pImage, pReq->u.In.abImage, pReq);
4068	else
4069	memcpy(pImage->pvImage, &pReq->u.In.abImage[0], pImage->cbImageBits);
4070	}
4071
4072	/*
4073	* Update any entry points.
4074	*/
4075	if (RT_SUCCESS(rc))
4076	{
4077	switch (pReq->u.In.eEPType)
4078	{
4079	default:
4080	case SUPLDRLOADEP_NOTHING:
4081	rc = VINF_SUCCESS;
4082	break;
4083	case SUPLDRLOADEP_VMMR0:
4084	rc = supdrvLdrSetVMMR0EPs(pDevExt, pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
4085	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
4086	break;
4087	case SUPLDRLOADEP_SERVICE:
4088	pImage->pfnServiceReqHandler = pReq->u.In.EP.Service.pfnServiceReq;
4089	rc = VINF_SUCCESS;
4090	break;
4091	}
4092	}
4093
4094	/*
4095	* On success call the module initialization.
4096	*/
4097	LogFlow(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
4098	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
4099	{
4100	Log(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
4101	rc = pImage->pfnModuleInit();
4102	if (rc && pDevExt->pvVMMR0 == pImage->pvImage)
4103	supdrvLdrUnsetVMMR0EPs(pDevExt);
4104	}
4105
4106	if (RT_FAILURE(rc))
4107	{
4108	pImage->uState = SUP_IOCTL_LDR_OPEN;
4109	pImage->pfnModuleInit = NULL;
4110	pImage->pfnModuleTerm = NULL;
4111	pImage->pfnServiceReqHandler= NULL;
4112	pImage->cbStrTab = 0;
4113	RTMemFree(pImage->pachStrTab);
4114	pImage->pachStrTab = NULL;
4115	RTMemFree(pImage->paSymbols);
4116	pImage->paSymbols = NULL;
4117	pImage->cSymbols = 0;
4118	}
4119
4120	supdrvLdrUnlock(pDevExt);
4121	return rc;
4122	}
4123
4124
4125	/**
4126	* Frees a previously loaded (prep'ed) image.
4127	*
4128	* @returns IPRT status code.
4129	* @param pDevExt Device globals.
4130	* @param pSession Session data.
4131	* @param pReq The request.
4132	*/
4133	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
4134	{
4135	int rc;
4136	PSUPDRVLDRUSAGE pUsagePrev;
4137	PSUPDRVLDRUSAGE pUsage;
4138	PSUPDRVLDRIMAGE pImage;
4139	LogFlow(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
4140
4141	/*
4142	* Find the ldr image.
4143	*/
4144	supdrvLdrLock(pDevExt);
4145	pUsagePrev = NULL;
4146	pUsage = pSession->pLdrUsage;
4147	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4148	{
4149	pUsagePrev = pUsage;
4150	pUsage = pUsage->pNext;
4151	}
4152	if (!pUsage)
4153	{
4154	supdrvLdrUnlock(pDevExt);
4155	Log(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
4156	return VERR_INVALID_HANDLE;
4157	}
4158
4159	/*
4160	* Check if we can remove anything.
4161	*/
4162	rc = VINF_SUCCESS;
4163	pImage = pUsage->pImage;
4164	if (pImage->cUsage <= 1 \|\| pUsage->cUsage <= 1)
4165	{
4166	/*
4167	* Check if there are any objects with destructors in the image, if
4168	* so leave it for the session cleanup routine so we get a chance to
4169	* clean things up in the right order and not leave them all dangling.
4170	*/
4171	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
4172	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
4173	if (pImage->cUsage <= 1)
4174	{
4175	PSUPDRVOBJ pObj;
4176	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4177	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4178	{
4179	rc = VERR_DANGLING_OBJECTS;
4180	break;
4181	}
4182	}
4183	else
4184	{
4185	PSUPDRVUSAGE pGenUsage;
4186	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
4187	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4188	{
4189	rc = VERR_DANGLING_OBJECTS;
4190	break;
4191	}
4192	}
4193	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
4194	if (rc == VINF_SUCCESS)
4195	{
4196	/* unlink it */
4197	if (pUsagePrev)
4198	pUsagePrev->pNext = pUsage->pNext;
4199	else
4200	pSession->pLdrUsage = pUsage->pNext;
4201
4202	/* free it */
4203	pUsage->pImage = NULL;
4204	pUsage->pNext = NULL;
4205	RTMemFree(pUsage);
4206
4207	/*
4208	* Dereference the image.
4209	*/
4210	if (pImage->cUsage <= 1)
4211	supdrvLdrFree(pDevExt, pImage);
4212	else
4213	pImage->cUsage--;
4214	}
4215	else
4216	{
4217	Log(("supdrvIOCtl_LdrFree: Dangling objects in %p/%s!\n", pImage->pvImage, pImage->szName));
4218	rc = VINF_SUCCESS; /** @todo BRANCH-2.1: remove this after branching. */
4219	}
4220	}
4221	else
4222	{
4223	/*
4224	* Dereference both image and usage.
4225	*/
4226	pImage->cUsage--;
4227	pUsage->cUsage--;
4228	}
4229
4230	supdrvLdrUnlock(pDevExt);
4231	return rc;
4232	}
4233
4234
4235	/**
4236	* Gets the address of a symbol in an open image.
4237	*
4238	* @returns IPRT status code.
4239	* @param pDevExt Device globals.
4240	* @param pSession Session data.
4241	* @param pReq The request buffer.
4242	*/
4243	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
4244	{
4245	PSUPDRVLDRIMAGE pImage;
4246	PSUPDRVLDRUSAGE pUsage;
4247	uint32_t i;
4248	PSUPLDRSYM paSyms;
4249	const char *pchStrings;
4250	const size_t cbSymbol = strlen(pReq->u.In.szSymbol) + 1;
4251	void *pvSymbol = NULL;
4252	int rc = VERR_GENERAL_FAILURE;
4253	Log3(("supdrvIOCtl_LdrGetSymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
4254
4255	/*
4256	* Find the ldr image.
4257	*/
4258	supdrvLdrLock(pDevExt);
4259	pUsage = pSession->pLdrUsage;
4260	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4261	pUsage = pUsage->pNext;
4262	if (!pUsage)
4263	{
4264	supdrvLdrUnlock(pDevExt);
4265	Log(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
4266	return VERR_INVALID_HANDLE;
4267	}
4268	pImage = pUsage->pImage;
4269	if (pImage->uState != SUP_IOCTL_LDR_LOAD)
4270	{
4271	unsigned uState = pImage->uState;
4272	supdrvLdrUnlock(pDevExt);
4273	Log(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", uState, uState)); NOREF(uState);
4274	return VERR_ALREADY_LOADED;
4275	}
4276
4277	/*
4278	* Search the symbol strings.
4279	*
4280	* Note! The int32_t is for native loading on solaris where the data
4281	* and text segments are in very different places.
4282	*/
4283	pchStrings = pImage->pachStrTab;
4284	paSyms = pImage->paSymbols;
4285	for (i = 0; i < pImage->cSymbols; i++)
4286	{
4287	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4288	&& !memcmp(pchStrings + paSyms[i].offName, pReq->u.In.szSymbol, cbSymbol))
4289	{
4290	pvSymbol = (uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol;
4291	rc = VINF_SUCCESS;
4292	break;
4293	}
4294	}
4295	supdrvLdrUnlock(pDevExt);
4296	pReq->u.Out.pvSymbol = pvSymbol;
4297	return rc;
4298	}
4299
4300
4301	/**
4302	* Gets the address of a symbol in an open image or the support driver.
4303	*
4304	* @returns VINF_SUCCESS on success.
4305	* @returns
4306	* @param pDevExt Device globals.
4307	* @param pSession Session data.
4308	* @param pReq The request buffer.
4309	*/
4310	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq)
4311	{
4312	int rc = VINF_SUCCESS;
4313	const char *pszSymbol = pReq->u.In.pszSymbol;
4314	const char *pszModule = pReq->u.In.pszModule;
4315	size_t cbSymbol;
4316	char const *pszEnd;
4317	uint32_t i;
4318
4319	/*
4320	* Input validation.
4321	*/
4322	AssertPtrReturn(pszSymbol, VERR_INVALID_POINTER);
4323	pszEnd = RTStrEnd(pszSymbol, 512);
4324	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4325	cbSymbol = pszEnd - pszSymbol + 1;
4326
4327	if (pszModule)
4328	{
4329	AssertPtrReturn(pszModule, VERR_INVALID_POINTER);
4330	pszEnd = RTStrEnd(pszModule, 64);
4331	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4332	}
4333	Log3(("supdrvIDC_LdrGetSymbol: pszModule=%p:{%s} pszSymbol=%p:{%s}\n", pszModule, pszModule, pszSymbol, pszSymbol));
4334
4335
4336	if ( !pszModule
4337	\|\| !strcmp(pszModule, "SupDrv"))
4338	{
4339	/*
4340	* Search the support driver export table.
4341	*/
4342	for (i = 0; i < RT_ELEMENTS(g_aFunctions); i++)
4343	if (!strcmp(g_aFunctions[i].szName, pszSymbol))
4344	{
4345	pReq->u.Out.pfnSymbol = g_aFunctions[i].pfn;
4346	break;
4347	}
4348	}
4349	else
4350	{
4351	/*
4352	* Find the loader image.
4353	*/
4354	PSUPDRVLDRIMAGE pImage;
4355
4356	supdrvLdrLock(pDevExt);
4357
4358	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
4359	if (!strcmp(pImage->szName, pszModule))
4360	break;
4361	if (pImage && pImage->uState == SUP_IOCTL_LDR_LOAD)
4362	{
4363	/*
4364	* Search the symbol strings.
4365	*/
4366	const char *pchStrings = pImage->pachStrTab;
4367	PCSUPLDRSYM paSyms = pImage->paSymbols;
4368	for (i = 0; i < pImage->cSymbols; i++)
4369	{
4370	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4371	&& !memcmp(pchStrings + paSyms[i].offName, pszSymbol, cbSymbol))
4372	{
4373	/*
4374	* Found it! Calc the symbol address and add a reference to the module.
4375	*/
4376	pReq->u.Out.pfnSymbol = (PFNRT)((uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol);
4377	rc = supdrvLdrAddUsage(pSession, pImage);
4378	break;
4379	}
4380	}
4381	}
4382	else
4383	rc = pImage ? VERR_WRONG_ORDER : VERR_MODULE_NOT_FOUND;
4384
4385	supdrvLdrUnlock(pDevExt);
4386	}
4387	return rc;
4388	}
4389
4390
4391	/**
4392	* Updates the VMMR0 entry point pointers.
4393	*
4394	* @returns IPRT status code.
4395	* @param pDevExt Device globals.
4396	* @param pSession Session data.
4397	* @param pVMMR0 VMMR0 image handle.
4398	* @param pvVMMR0EntryInt VMMR0EntryInt address.
4399	* @param pvVMMR0EntryFast VMMR0EntryFast address.
4400	* @param pvVMMR0EntryEx VMMR0EntryEx address.
4401	* @remark Caller must own the loader mutex.
4402	*/
4403	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx)
4404	{
4405	int rc = VINF_SUCCESS;
4406	LogFlow(("supdrvLdrSetR0EP pvVMMR0=%p pvVMMR0EntryInt=%p\n", pvVMMR0, pvVMMR0EntryInt));
4407
4408
4409	/*
4410	* Check if not yet set.
4411	*/
4412	if (!pDevExt->pvVMMR0)
4413	{
4414	pDevExt->pvVMMR0 = pvVMMR0;
4415	pDevExt->pfnVMMR0EntryInt = pvVMMR0EntryInt;
4416	pDevExt->pfnVMMR0EntryFast = pvVMMR0EntryFast;
4417	pDevExt->pfnVMMR0EntryEx = pvVMMR0EntryEx;
4418	}
4419	else
4420	{
4421	/*
4422	* Return failure or success depending on whether the values match or not.
4423	*/
4424	if ( pDevExt->pvVMMR0 != pvVMMR0
4425	\|\| (void *)pDevExt->pfnVMMR0EntryInt != pvVMMR0EntryInt
4426	\|\| (void *)pDevExt->pfnVMMR0EntryFast != pvVMMR0EntryFast
4427	\|\| (void *)pDevExt->pfnVMMR0EntryEx != pvVMMR0EntryEx)
4428	{
4429	AssertMsgFailed(("SUP_IOCTL_LDR_SETR0EP: Already set pointing to a different module!\n"));
4430	rc = VERR_INVALID_PARAMETER;
4431	}
4432	}
4433	return rc;
4434	}
4435
4436
4437	/**
4438	* Unsets the VMMR0 entry point installed by supdrvLdrSetR0EP.
4439	*
4440	* @param pDevExt Device globals.
4441	*/
4442	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt)
4443	{
4444	pDevExt->pvVMMR0 = NULL;
4445	pDevExt->pfnVMMR0EntryInt = NULL;
4446	pDevExt->pfnVMMR0EntryFast = NULL;
4447	pDevExt->pfnVMMR0EntryEx = NULL;
4448	}
4449
4450
4451	/**
4452	* Adds a usage reference in the specified session of an image.
4453	*
4454	* Called while owning the loader semaphore.
4455	*
4456	* @returns VINF_SUCCESS on success and VERR_NO_MEMORY on failure.
4457	* @param pSession Session in question.
4458	* @param pImage Image which the session is using.
4459	*/
4460	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage)
4461	{
4462	PSUPDRVLDRUSAGE pUsage;
4463	LogFlow(("supdrvLdrAddUsage: pImage=%p\n", pImage));
4464
4465	/*
4466	* Referenced it already?
4467	*/
4468	pUsage = pSession->pLdrUsage;
4469	while (pUsage)
4470	{
4471	if (pUsage->pImage == pImage)
4472	{
4473	pUsage->cUsage++;
4474	return VINF_SUCCESS;
4475	}
4476	pUsage = pUsage->pNext;
4477	}
4478
4479	/*
4480	* Allocate new usage record.
4481	*/
4482	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
4483	AssertReturn(pUsage, /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_5);
4484	pUsage->cUsage = 1;
4485	pUsage->pImage = pImage;
4486	pUsage->pNext = pSession->pLdrUsage;
4487	pSession->pLdrUsage = pUsage;
4488	return VINF_SUCCESS;
4489	}
4490
4491
4492	/**
4493	* Frees a load image.
4494	*
4495	* @param pDevExt Pointer to device extension.
4496	* @param pImage Pointer to the image we're gonna free.
4497	* This image must exit!
4498	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
4499	*/
4500	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
4501	{
4502	PSUPDRVLDRIMAGE pImagePrev;
4503	LogFlow(("supdrvLdrFree: pImage=%p\n", pImage));
4504
4505	/* find it - arg. should've used doubly linked list. */
4506	Assert(pDevExt->pLdrImages);
4507	pImagePrev = NULL;
4508	if (pDevExt->pLdrImages != pImage)
4509	{
4510	pImagePrev = pDevExt->pLdrImages;
4511	while (pImagePrev->pNext != pImage)
4512	pImagePrev = pImagePrev->pNext;
4513	Assert(pImagePrev->pNext == pImage);
4514	}
4515
4516	/* unlink */
4517	if (pImagePrev)
4518	pImagePrev->pNext = pImage->pNext;
4519	else
4520	pDevExt->pLdrImages = pImage->pNext;
4521
4522	/* check if this is VMMR0.r0 unset its entry point pointers. */
4523	if (pDevExt->pvVMMR0 == pImage->pvImage)
4524	supdrvLdrUnsetVMMR0EPs(pDevExt);
4525
4526	/* check for objects with destructors in this image. (Shouldn't happen.) */
4527	if (pDevExt->pObjs)
4528	{
4529	unsigned cObjs = 0;
4530	PSUPDRVOBJ pObj;
4531	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
4532	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
4533	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4534	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4535	{
4536	pObj->pfnDestructor = NULL;
4537	cObjs++;
4538	}
4539	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
4540	if (cObjs)
4541	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
4542	}
4543
4544	/* call termination function if fully loaded. */
4545	if ( pImage->pfnModuleTerm
4546	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
4547	{
4548	LogFlow(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
4549	pImage->pfnModuleTerm();
4550	}
4551
4552	/* do native unload if appropriate. */
4553	if (pImage->fNative)
4554	supdrvOSLdrUnload(pDevExt, pImage);
4555
4556	/* free the image */
4557	pImage->cUsage = 0;
4558	pImage->pNext = 0;
4559	pImage->uState = SUP_IOCTL_LDR_FREE;
4560	RTMemExecFree(pImage->pvImageAlloc, pImage->cbImageBits + 31);
4561	pImage->pvImageAlloc = NULL;
4562	RTMemFree(pImage->pachStrTab);
4563	pImage->pachStrTab = NULL;
4564	RTMemFree(pImage->paSymbols);
4565	pImage->paSymbols = NULL;
4566	RTMemFree(pImage);
4567	}
4568
4569
4570	/**
4571	* Acquires the loader lock.
4572	*
4573	* @returns IPRT status code.
4574	* @param pDevExt The device extension.
4575	*/
4576	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt)
4577	{
4578	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4579	int rc = RTSemMutexRequest(pDevExt->mtxLdr, RT_INDEFINITE_WAIT);
4580	#else
4581	int rc = RTSemFastMutexRequest(pDevExt->mtxLdr);
4582	#endif
4583	AssertRC(rc);
4584	return rc;
4585	}
4586
4587
4588	/**
4589	* Releases the loader lock.
4590	*
4591	* @returns IPRT status code.
4592	* @param pDevExt The device extension.
4593	*/
4594	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt)
4595	{
4596	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4597	return RTSemMutexRelease(pDevExt->mtxLdr);
4598	#else
4599	return RTSemFastMutexRelease(pDevExt->mtxLdr);
4600	#endif
4601	}
4602
4603
4604	/**
4605	* Implements the service call request.
4606	*
4607	* @returns VBox status code.
4608	* @param pDevExt The device extension.
4609	* @param pSession The calling session.
4610	* @param pReq The request packet, valid.
4611	*/
4612	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq)
4613	{
4614	#if !defined(RT_OS_WINDOWS) \|\| defined(DEBUG)
4615	int rc;
4616
4617	/*
4618	* Find the module first in the module referenced by the calling session.
4619	*/
4620	rc = supdrvLdrLock(pDevExt);
4621	if (RT_SUCCESS(rc))
4622	{
4623	PFNSUPR0SERVICEREQHANDLER pfnServiceReqHandler = NULL;
4624	PSUPDRVLDRUSAGE pUsage;
4625
4626	for (pUsage = pSession->pLdrUsage; pUsage; pUsage = pUsage->pNext)
4627	if ( pUsage->pImage->pfnServiceReqHandler
4628	&& !strcmp(pUsage->pImage->szName, pReq->u.In.szName))
4629	{
4630	pfnServiceReqHandler = pUsage->pImage->pfnServiceReqHandler;
4631	break;
4632	}
4633	supdrvLdrUnlock(pDevExt);
4634
4635	if (pfnServiceReqHandler)
4636	{
4637	/*
4638	* Call it.
4639	*/
4640	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
4641	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, NULL);
4642	else
4643	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0]);
4644	}
4645	else
4646	rc = VERR_SUPDRV_SERVICE_NOT_FOUND;
4647	}
4648
4649	/* log it */
4650	if ( RT_FAILURE(rc)
4651	&& rc != VERR_INTERRUPTED
4652	&& rc != VERR_TIMEOUT)
4653	Log(("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	else
4656	Log4(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4657	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4658	return rc;
4659	#else /* RT_OS_WINDOWS && !DEBUG */
4660	return VERR_NOT_IMPLEMENTED;
4661	#endif /* RT_OS_WINDOWS && !DEBUG */
4662	}
4663
4664
4665	/**
4666	* Implements the logger settings request.
4667	*
4668	* @returns VBox status code.
4669	* @param pDevExt The device extension.
4670	* @param pSession The caller's session.
4671	* @param pReq The request.
4672	*/
4673	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq)
4674	{
4675	const char *pszGroup = &pReq->u.In.szStrings[pReq->u.In.offGroups];
4676	const char *pszFlags = &pReq->u.In.szStrings[pReq->u.In.offFlags];
4677	const char *pszDest = &pReq->u.In.szStrings[pReq->u.In.offDestination];
4678	PRTLOGGER pLogger = NULL;
4679	int rc;
4680
4681	/*
4682	* Some further validation.
4683	*/
4684	switch (pReq->u.In.fWhat)
4685	{
4686	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4687	case SUPLOGGERSETTINGS_WHAT_CREATE:
4688	break;
4689
4690	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4691	if (pszGroup \|\| pszFlags \|\| *pszDest)
4692	return VERR_INVALID_PARAMETER;
4693	if (pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_RELEASE)
4694	return VERR_ACCESS_DENIED;
4695	break;
4696
4697	default:
4698	return VERR_INTERNAL_ERROR;
4699	}
4700
4701	/*
4702	* Get the logger.
4703	*/
4704	switch (pReq->u.In.fWhich)
4705	{
4706	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4707	pLogger = RTLogGetDefaultInstance();
4708	break;
4709
4710	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4711	pLogger = RTLogRelDefaultInstance();
4712	break;
4713
4714	default:
4715	return VERR_INTERNAL_ERROR;
4716	}
4717
4718	/*
4719	* Do the job.
4720	*/
4721	switch (pReq->u.In.fWhat)
4722	{
4723	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4724	if (pLogger)
4725	{
4726	rc = RTLogFlags(pLogger, pszFlags);
4727	if (RT_SUCCESS(rc))
4728	rc = RTLogGroupSettings(pLogger, pszGroup);
4729	NOREF(pszDest);
4730	}
4731	else
4732	rc = VERR_NOT_FOUND;
4733	break;
4734
4735	case SUPLOGGERSETTINGS_WHAT_CREATE:
4736	{
4737	if (pLogger)
4738	rc = VERR_ALREADY_EXISTS;
4739	else
4740	{
4741	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
4742
4743	rc = RTLogCreate(&pLogger,
4744	0 /* fFlags */,
4745	pszGroup,
4746	pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_DEBUG
4747	? "VBOX_LOG"
4748	: "VBOX_RELEASE_LOG",
4749	RT_ELEMENTS(s_apszGroups),
4750	s_apszGroups,
4751	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER,
4752	NULL);
4753	if (RT_SUCCESS(rc))
4754	{
4755	rc = RTLogFlags(pLogger, pszFlags);
4756	NOREF(pszDest);
4757	if (RT_SUCCESS(rc))
4758	{
4759	switch (pReq->u.In.fWhich)
4760	{
4761	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4762	pLogger = RTLogSetDefaultInstance(pLogger);
4763	break;
4764	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4765	pLogger = RTLogRelSetDefaultInstance(pLogger);
4766	break;
4767	}
4768	}
4769	RTLogDestroy(pLogger);
4770	}
4771	}
4772	break;
4773	}
4774
4775	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4776	switch (pReq->u.In.fWhich)
4777	{
4778	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4779	pLogger = RTLogSetDefaultInstance(NULL);
4780	break;
4781	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4782	pLogger = RTLogRelSetDefaultInstance(NULL);
4783	break;
4784	}
4785	rc = RTLogDestroy(pLogger);
4786	break;
4787
4788	default:
4789	{
4790	rc = VERR_INTERNAL_ERROR;
4791	break;
4792	}
4793	}
4794
4795	return rc;
4796	}
4797
4798
4799	/**
4800	* Creates the GIP.
4801	*
4802	* @returns VBox status code.
4803	* @param pDevExt Instance data. GIP stuff may be updated.
4804	*/
4805	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
4806	{
4807	PSUPGLOBALINFOPAGE pGip;
4808	RTHCPHYS HCPhysGip;
4809	uint32_t u32SystemResolution;
4810	uint32_t u32Interval;
4811	unsigned cCpus;
4812	int rc;
4813
4814
4815	LogFlow(("supdrvGipCreate:\n"));
4816
4817	/* assert order */
4818	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
4819	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
4820	Assert(!pDevExt->pGipTimer);
4821
4822	/*
4823	* Check the CPU count.
4824	*/
4825	cCpus = RTMpGetArraySize();
4826	if ( cCpus > RTCPUSET_MAX_CPUS
4827	\|\| cCpus > 256 /ApicId is used for the mappings/)
4828	{
4829	SUPR0Printf("VBoxDrv: Too many CPUs (%u) for the GIP (max %u)\n", cCpus, RT_MIN(RTCPUSET_MAX_CPUS, 256));
4830	return VERR_TOO_MANY_CPUS;
4831	}
4832
4833	/*
4834	* Allocate a contiguous set of pages with a default kernel mapping.
4835	*/
4836	rc = RTR0MemObjAllocCont(&pDevExt->GipMemObj, RT_UOFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), false /fExecutable/);
4837	if (RT_FAILURE(rc))
4838	{
4839	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
4840	return rc;
4841	}
4842	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
4843	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
4844
4845	/*
4846	* Find a reasonable update interval and initialize the structure.
4847	*/
4848	u32Interval = u32SystemResolution = RTTimerGetSystemGranularity();
4849	while (u32Interval < 10000000 /* 10 ms */)
4850	u32Interval += u32SystemResolution;
4851
4852	supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), 1000000000 / u32Interval /=Hz/, cCpus);
4853
4854	/*
4855	* Create the timer.
4856	* If CPU_ALL isn't supported we'll have to fall back to synchronous mode.
4857	*/
4858	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
4859	{
4860	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, RTTIMER_FLAGS_CPU_ALL, supdrvGipAsyncTimer, pDevExt);
4861	if (rc == VERR_NOT_SUPPORTED)
4862	{
4863	OSDBGPRINT(("supdrvGipCreate: omni timer not supported, falling back to synchronous mode\n"));
4864	pGip->u32Mode = SUPGIPMODE_SYNC_TSC;
4865	}
4866	}
4867	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
4868	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0, supdrvGipSyncTimer, pDevExt);
4869	if (RT_SUCCESS(rc))
4870	{
4871	rc = RTMpNotificationRegister(supdrvGipMpEvent, pDevExt);
4872	if (RT_SUCCESS(rc))
4873	{
4874	rc = RTMpOnAll(supdrvGipInitOnCpu, pDevExt, pGip);
4875	if (RT_SUCCESS(rc))
4876	{
4877	/*
4878	* We're good.
4879	*/
4880	Log(("supdrvGipCreate: %u ns interval.\n", u32Interval));
4881	g_pSUPGlobalInfoPage = pGip;
4882	return VINF_SUCCESS;
4883	}
4884
4885	OSDBGPRINT(("supdrvGipCreate: RTMpOnAll failed with rc=%Rrc\n", rc));
4886	RTMpNotificationDeregister(supdrvGipMpEvent, pDevExt);
4887
4888	}
4889	else
4890	OSDBGPRINT(("supdrvGipCreate: failed to register MP event notfication. rc=%Rrc\n", rc));
4891	}
4892	else
4893	{
4894	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %u ns interval. rc=%Rrc\n", u32Interval, rc));
4895	Assert(!pDevExt->pGipTimer);
4896	}
4897	supdrvGipDestroy(pDevExt);
4898	return rc;
4899	}
4900
4901
4902	/**
4903	* Terminates the GIP.
4904	*
4905	* @param pDevExt Instance data. GIP stuff may be updated.
4906	*/
4907	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
4908	{
4909	int rc;
4910	#ifdef DEBUG_DARWIN_GIP
4911	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
4912	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
4913	pDevExt->pGipTimer, pDevExt->GipMemObj));
4914	#endif
4915
4916	/*
4917	* Invalid the GIP data.
4918	*/
4919	if (pDevExt->pGip)
4920	{
4921	supdrvGipTerm(pDevExt->pGip);
4922	pDevExt->pGip = NULL;
4923	}
4924	g_pSUPGlobalInfoPage = NULL;
4925
4926	/*
4927	* Destroy the timer and free the GIP memory object.
4928	*/
4929	if (pDevExt->pGipTimer)
4930	{
4931	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
4932	pDevExt->pGipTimer = NULL;
4933	}
4934
4935	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
4936	{
4937	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
4938	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
4939	}
4940
4941	/*
4942	* Finally, make sure we've release the system timer resolution request
4943	* if one actually succeeded and is still pending.
4944	*/
4945	if (pDevExt->u32SystemTimerGranularityGrant)
4946	{
4947	rc = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
4948	pDevExt->u32SystemTimerGranularityGrant = 0;
4949	}
4950	}
4951
4952
4953	/**
4954	* Timer callback function sync GIP mode.
4955	* @param pTimer The timer.
4956	* @param pvUser The device extension.
4957	*/
4958	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
4959	{
4960	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
4961	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4962	uint64_t u64TSC = ASMReadTSC();
4963	uint64_t NanoTS = RTTimeSystemNanoTS();
4964
4965	supdrvGipUpdate(pDevExt->pGip, NanoTS, u64TSC, NIL_RTCPUID, iTick);
4966
4967	ASMSetFlags(fOldFlags);
4968	}
4969
4970
4971	/**
4972	* Timer callback function for async GIP mode.
4973	* @param pTimer The timer.
4974	* @param pvUser The device extension.
4975	*/
4976	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
4977	{
4978	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
4979	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4980	RTCPUID idCpu = RTMpCpuId();
4981	uint64_t u64TSC = ASMReadTSC();
4982	uint64_t NanoTS = RTTimeSystemNanoTS();
4983
4984	/** @todo reset the transaction number and whatnot when iTick == 1. */
4985	if (pDevExt->idGipMaster == idCpu)
4986	supdrvGipUpdate(pDevExt->pGip, NanoTS, u64TSC, idCpu, iTick);
4987	else
4988	supdrvGipUpdatePerCpu(pDevExt->pGip, NanoTS, u64TSC, idCpu, ASMGetApicId(), iTick);
4989
4990	ASMSetFlags(fOldFlags);
4991	}
4992
4993
4994	/**
4995	* Finds our (@a idCpu) entry, or allocates a new one if not found.
4996	*
4997	* @returns Index of the CPU in the cache set.
4998	* @param pGip The GIP.
4999	* @param idCpu The CPU ID.
5000	*/
5001	static uint32_t supdrvGipCpuIndexFromCpuId(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
5002	{
5003	uint32_t i, cTries;
5004
5005	/*
5006	* ASSUMES that CPU IDs are constant.
5007	*/
5008	for (i = 0; i < pGip->cCpus; i++)
5009	if (pGip->aCPUs[i].idCpu == idCpu)
5010	return i;
5011
5012	cTries = 0;
5013	do
5014	{
5015	for (i = 0; i < pGip->cCpus; i++)
5016	{
5017	bool fRc;
5018	ASMAtomicCmpXchgSize(&pGip->aCPUs[i].idCpu, idCpu, NIL_RTCPUID, fRc);
5019	if (fRc)
5020	return i;
5021	}
5022	} while (cTries++ < 32);
5023	AssertReleaseFailed();
5024	return i - 1;
5025	}
5026
5027
5028	/**
5029	* The calling CPU should be accounted as online, update GIP accordingly.
5030	*
5031	* This is used by supdrvGipMpEvent as well as the supdrvGipCreate.
5032	*
5033	* @param pGip The GIP.
5034	* @param idCpu The CPU ID.
5035	*/
5036	static void supdrvGipMpEventOnline(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
5037	{
5038	int iCpuSet = 0;
5039	uint16_t idApic = UINT16_MAX;
5040	uint32_t i = 0;
5041
5042	AssertRelease(idCpu == RTMpCpuId());
5043	Assert(pGip->cPossibleCpus == RTMpGetCount());
5044
5045	/*
5046	* Update the globals.
5047	*/
5048	ASMAtomicWriteU16(&pGip->cPresentCpus, RTMpGetPresentCount());
5049	ASMAtomicWriteU16(&pGip->cOnlineCpus, RTMpGetOnlineCount());
5050	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5051	if (iCpuSet >= 0)
5052	{
5053	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5054	RTCpuSetAddByIndex(&pGip->OnlineCpuSet, iCpuSet);
5055	RTCpuSetAddByIndex(&pGip->PresentCpuSet, iCpuSet);
5056	}
5057
5058	/*
5059	* Update the entry.
5060	*/
5061	i = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
5062	idApic = ASMGetApicId();
5063	ASMAtomicUoWriteU16(&pGip->aCPUs[i].idApic, idApic);
5064	ASMAtomicUoWriteS16(&pGip->aCPUs[i].iCpuSet, (int16_t)iCpuSet);
5065	ASMAtomicUoWriteSize(&pGip->aCPUs[i].idCpu, idCpu);
5066
5067	/*
5068	* Update the APIC ID and CPU set index mappings.
5069	*/
5070	ASMAtomicWriteU16(&pGip->aiCpuFromApicId[idApic], i);
5071	ASMAtomicWriteU16(&pGip->aiCpuFromCpuSetIdx[iCpuSet], i);
5072
5073	/* commit it */
5074	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_ONLINE);
5075	}
5076
5077
5078	/**
5079	* The CPU should be accounted as offline, update the GIP accordingly.
5080	*
5081	* This is used by supdrvGipMpEvent.
5082	*
5083	* @param pGip The GIP.
5084	* @param idCpu The CPU ID.
5085	*/
5086	static void supdrvGipMpEventOffline(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
5087	{
5088	int iCpuSet;
5089	unsigned i;
5090
5091	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5092	AssertReturnVoid(iCpuSet >= 0);
5093
5094	i = pGip->aiCpuFromCpuSetIdx[iCpuSet];
5095	AssertReturnVoid(i < pGip->cCpus);
5096	AssertReturnVoid(pGip->aCPUs[i].idCpu == idCpu);
5097
5098	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5099	RTCpuSetDelByIndex(&pGip->OnlineCpuSet, iCpuSet);
5100
5101	/* commit it */
5102	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_OFFLINE);
5103	}
5104
5105
5106	/**
5107	* Multiprocessor event notification callback.
5108	*
5109	* This is used to make sure that the GIP master gets passed on to
5110	* another CPU. It also updates the associated CPU data.
5111	*
5112	* @param enmEvent The event.
5113	* @param idCpu The cpu it applies to.
5114	* @param pvUser Pointer to the device extension.
5115	*/
5116	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser)
5117	{
5118	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5119	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5120
5121	AssertRelease(idCpu == RTMpCpuId());
5122	AssertRelease(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
5123
5124	/*
5125	* Update the GIP CPU data.
5126	*/
5127	if (pGip)
5128	{
5129	switch (enmEvent)
5130	{
5131	case RTMPEVENT_ONLINE:
5132	supdrvGipMpEventOnline(pGip, idCpu);
5133	break;
5134	case RTMPEVENT_OFFLINE:
5135	supdrvGipMpEventOffline(pGip, idCpu);
5136	break;
5137
5138	}
5139	}
5140
5141	/*
5142	* Make sure there is a master GIP.
5143	*/
5144	if (enmEvent == RTMPEVENT_OFFLINE)
5145	{
5146	RTCPUID idGipMaster = ASMAtomicReadU32(&pDevExt->idGipMaster);
5147	if (idGipMaster == idCpu)
5148	{
5149	/*
5150	* Find a new GIP master.
5151	*/
5152	bool fIgnored;
5153	unsigned i;
5154	RTCPUID idNewGipMaster = NIL_RTCPUID;
5155	RTCPUSET OnlineCpus;
5156	RTMpGetOnlineSet(&OnlineCpus);
5157
5158	for (i = 0; i < RTCPUSET_MAX_CPUS; i++)
5159	{
5160	RTCPUID idCurCpu = RTMpCpuIdFromSetIndex(i);
5161	if ( RTCpuSetIsMember(&OnlineCpus, idCurCpu)
5162	&& idCurCpu != idGipMaster)
5163	{
5164	idNewGipMaster = idCurCpu;
5165	break;
5166	}
5167	}
5168
5169	Log(("supdrvGipMpEvent: Gip master %#lx -> %#lx\n", (long)idGipMaster, (long)idNewGipMaster));
5170	ASMAtomicCmpXchgSize(&pDevExt->idGipMaster, idNewGipMaster, idGipMaster, fIgnored);
5171	NOREF(fIgnored);
5172	}
5173	}
5174	}
5175
5176
5177	/**
5178	* Callback used by supdrvDetermineAsyncTSC to read the TSC on a CPU.
5179	*
5180	* @param idCpu Ignored.
5181	* @param pvUser1 Where to put the TSC.
5182	* @param pvUser2 Ignored.
5183	*/
5184	static DECLCALLBACK(void) supdrvDetermineAsyncTscWorker(RTCPUID idCpu, void pvUser1, void pvUser2)
5185	{
5186	#if 1
5187	ASMAtomicWriteU64((uint64_t volatile *)pvUser1, ASMReadTSC());
5188	#else
5189	(uint64_t )pvUser1 = ASMReadTSC();
5190	#endif
5191	}
5192
5193
5194	/**
5195	* Determine if Async GIP mode is required because of TSC drift.
5196	*
5197	* When using the default/normal timer code it is essential that the time stamp counter
5198	* (TSC) runs never backwards, that is, a read operation to the counter should return
5199	* a bigger value than any previous read operation. This is guaranteed by the latest
5200	* AMD CPUs and by newer Intel CPUs which never enter the C2 state (P4). In any other
5201	* case we have to choose the asynchronous timer mode.
5202	*
5203	* @param poffMin Pointer to the determined difference between different cores.
5204	* @return false if the time stamp counters appear to be synchronized, true otherwise.
5205	*/
5206	static bool supdrvDetermineAsyncTsc(uint64_t *poffMin)
5207	{
5208	/*
5209	* Just iterate all the cpus 8 times and make sure that the TSC is
5210	* ever increasing. We don't bother taking TSC rollover into account.
5211	*/
5212	int iEndCpu = RTMpGetArraySize();
5213	int iCpu;
5214	int cLoops = 8;
5215	bool fAsync = false;
5216	int rc = VINF_SUCCESS;
5217	uint64_t offMax = 0;
5218	uint64_t offMin = ~(uint64_t)0;
5219	uint64_t PrevTsc = ASMReadTSC();
5220
5221	while (cLoops-- > 0)
5222	{
5223	for (iCpu = 0; iCpu < iEndCpu; iCpu++)
5224	{
5225	uint64_t CurTsc;
5226	rc = RTMpOnSpecific(RTMpCpuIdFromSetIndex(iCpu), supdrvDetermineAsyncTscWorker, &CurTsc, NULL);
5227	if (RT_SUCCESS(rc))
5228	{
5229	if (CurTsc <= PrevTsc)
5230	{
5231	fAsync = true;
5232	offMin = offMax = PrevTsc - CurTsc;
5233	Log(("supdrvDetermineAsyncTsc: iCpu=%d cLoops=%d CurTsc=%llx PrevTsc=%llx\n",
5234	iCpu, cLoops, CurTsc, PrevTsc));
5235	break;
5236	}
5237
5238	/* Gather statistics (except the first time). */
5239	if (iCpu != 0 \|\| cLoops != 7)
5240	{
5241	uint64_t off = CurTsc - PrevTsc;
5242	if (off < offMin)
5243	offMin = off;
5244	if (off > offMax)
5245	offMax = off;
5246	Log2(("%d/%d: off=%llx\n", cLoops, iCpu, off));
5247	}
5248
5249	/* Next */
5250	PrevTsc = CurTsc;
5251	}
5252	else if (rc == VERR_NOT_SUPPORTED)
5253	break;
5254	else
5255	AssertMsg(rc == VERR_CPU_NOT_FOUND \|\| rc == VERR_CPU_OFFLINE, ("%d\n", rc));
5256	}
5257
5258	/* broke out of the loop. */
5259	if (iCpu < iEndCpu)
5260	break;
5261	}
5262
5263	poffMin = offMin; / Almost RTMpOnSpecific profiling. */
5264	Log(("supdrvDetermineAsyncTsc: returns %d; iEndCpu=%d rc=%d offMin=%llx offMax=%llx\n",
5265	fAsync, iEndCpu, rc, offMin, offMax));
5266	#if !defined(RT_OS_SOLARIS) && !defined(RT_OS_OS2) && !defined(RT_OS_WINDOWS)
5267	OSDBGPRINT(("vboxdrv: fAsync=%d offMin=%#lx offMax=%#lx\n", fAsync, (long)offMin, (long)offMax));
5268	#endif
5269	return fAsync;
5270	}
5271
5272
5273	/**
5274	* Determine the GIP TSC mode.
5275	*
5276	* @returns The most suitable TSC mode.
5277	* @param pDevExt Pointer to the device instance data.
5278	*/
5279	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt)
5280	{
5281	/*
5282	* On SMP we're faced with two problems:
5283	* (1) There might be a skew between the CPU, so that cpu0
5284	* returns a TSC that is slightly different from cpu1.
5285	* (2) Power management (and other things) may cause the TSC
5286	* to run at a non-constant speed, and cause the speed
5287	* to be different on the cpus. This will result in (1).
5288	*
5289	* So, on SMP systems we'll have to select the ASYNC update method
5290	* if there are symptoms of these problems.
5291	*/
5292	if (RTMpGetCount() > 1)
5293	{
5294	uint32_t uEAX, uEBX, uECX, uEDX;
5295	uint64_t u64DiffCoresIgnored;
5296
5297	/* Permit the user and/or the OS specific bits to force async mode. */
5298	if (supdrvOSGetForcedAsyncTscMode(pDevExt))
5299	return SUPGIPMODE_ASYNC_TSC;
5300
5301	/* Try check for current differences between the cpus. */
5302	if (supdrvDetermineAsyncTsc(&u64DiffCoresIgnored))
5303	return SUPGIPMODE_ASYNC_TSC;
5304
5305	/*
5306	* If the CPU supports power management and is an AMD one we
5307	* won't trust it unless it has the TscInvariant bit is set.
5308	*/
5309	/* Check for "AuthenticAMD" */
5310	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
5311	if ( uEAX >= 1
5312	&& uEBX == X86_CPUID_VENDOR_AMD_EBX
5313	&& uECX == X86_CPUID_VENDOR_AMD_ECX
5314	&& uEDX == X86_CPUID_VENDOR_AMD_EDX)
5315	{
5316	/* Check for APM support and that TscInvariant is cleared. */
5317	ASMCpuId(0x80000000, &uEAX, &uEBX, &uECX, &uEDX);
5318	if (uEAX >= 0x80000007)
5319	{
5320	ASMCpuId(0x80000007, &uEAX, &uEBX, &uECX, &uEDX);
5321	if ( !(uEDX & RT_BIT(8))/* TscInvariant */
5322	&& (uEDX & 0x3e)) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
5323	return SUPGIPMODE_ASYNC_TSC;
5324	}
5325	}
5326	}
5327	return SUPGIPMODE_SYNC_TSC;
5328	}
5329
5330
5331
5332	/**
5333	* Initializes the GIP data.
5334	*
5335	* @param pDevExt Pointer to the device instance data.
5336	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5337	* @param HCPhys The physical address of the GIP.
5338	* @param u64NanoTS The current nanosecond timestamp.
5339	* @param uUpdateHz The update frequency.
5340	* @param cCpus The CPU count.
5341	*/
5342	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
5343	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned cCpus)
5344	{
5345	size_t const cbGip = RT_ALIGN_Z(RT_OFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), PAGE_SIZE);
5346	unsigned i;
5347	#ifdef DEBUG_DARWIN_GIP
5348	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5349	#else
5350	LogFlow(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5351	#endif
5352
5353	/*
5354	* Initialize the structure.
5355	*/
5356	memset(pGip, 0, cbGip);
5357	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
5358	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
5359	pGip->u32Mode = supdrvGipDeterminTscMode(pDevExt);
5360	pGip->cCpus = (uint16_t)cCpus;
5361	pGip->cPages = (uint16_t)(cbGip / PAGE_SIZE);
5362	pGip->u32UpdateHz = uUpdateHz;
5363	pGip->u32UpdateIntervalNS = 1000000000 / uUpdateHz;
5364	pGip->u64NanoTSLastUpdateHz = u64NanoTS;
5365	RTCpuSetEmpty(&pGip->OnlineCpuSet);
5366	RTCpuSetEmpty(&pGip->PresentCpuSet);
5367	RTMpGetSet(&pGip->PossibleCpuSet);
5368	pGip->cOnlineCpus = RTMpGetOnlineCount();
5369	pGip->cPresentCpus = RTMpGetPresentCount();
5370	pGip->cPossibleCpus = RTMpGetCount();
5371	pGip->idCpuMax = RTMpGetMaxCpuId();
5372	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromApicId); i++)
5373	pGip->aiCpuFromApicId[i] = 0;
5374	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx); i++)
5375	pGip->aiCpuFromCpuSetIdx[i] = UINT16_MAX;
5376
5377	for (i = 0; i < cCpus; i++)
5378	{
5379	pGip->aCPUs[i].u32TransactionId = 2;
5380	pGip->aCPUs[i].u64NanoTS = u64NanoTS;
5381	pGip->aCPUs[i].u64TSC = ASMReadTSC();
5382
5383	pGip->aCPUs[i].enmState = SUPGIPCPUSTATE_INVALID;
5384	pGip->aCPUs[i].idCpu = NIL_RTCPUID;
5385	pGip->aCPUs[i].iCpuSet = -1;
5386	pGip->aCPUs[i].idApic = UINT16_MAX;
5387
5388	/*
5389	* We don't know the following values until we've executed updates.
5390	* So, we'll just pretend it's a 4 GHz CPU and adjust the history it on
5391	* the 2nd timer callout.
5392	*/
5393	pGip->aCPUs[i].u64CpuHz = _4G + 1; /* tstGIP-2 depends on this. */
5394	pGip->aCPUs[i].u32UpdateIntervalTSC
5395	= pGip->aCPUs[i].au32TSCHistory[0]
5396	= pGip->aCPUs[i].au32TSCHistory[1]
5397	= pGip->aCPUs[i].au32TSCHistory[2]
5398	= pGip->aCPUs[i].au32TSCHistory[3]
5399	= pGip->aCPUs[i].au32TSCHistory[4]
5400	= pGip->aCPUs[i].au32TSCHistory[5]
5401	= pGip->aCPUs[i].au32TSCHistory[6]
5402	= pGip->aCPUs[i].au32TSCHistory[7]
5403	= /pGip->aCPUs[i].u64CpuHz/ (uint32_t)(_4G / uUpdateHz);
5404	}
5405
5406	/*
5407	* Link it to the device extension.
5408	*/
5409	pDevExt->pGip = pGip;
5410	pDevExt->HCPhysGip = HCPhys;
5411	pDevExt->cGipUsers = 0;
5412	}
5413
5414
5415	/**
5416	* On CPU initialization callback for RTMpOnAll.
5417	*
5418	* @param idCpu The CPU ID.
5419	* @param pvUser1 The device extension.
5420	* @param pvUser2 The GIP.
5421	*/
5422	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2)
5423	{
5424	/* This is good enough, even though it will update some of the globals a
5425	bit to much. */
5426	supdrvGipMpEventOnline((PSUPGLOBALINFOPAGE)pvUser2, idCpu);
5427	}
5428
5429
5430	/**
5431	* Invalidates the GIP data upon termination.
5432	*
5433	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5434	*/
5435	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
5436	{
5437	unsigned i;
5438	pGip->u32Magic = 0;
5439	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
5440	{
5441	pGip->aCPUs[i].u64NanoTS = 0;
5442	pGip->aCPUs[i].u64TSC = 0;
5443	pGip->aCPUs[i].iTSCHistoryHead = 0;
5444	}
5445	}
5446
5447
5448	/**
5449	* Worker routine for supdrvGipUpdate and supdrvGipUpdatePerCpu that
5450	* updates all the per cpu data except the transaction id.
5451	*
5452	* @param pGip The GIP.
5453	* @param pGipCpu Pointer to the per cpu data.
5454	* @param u64NanoTS The current time stamp.
5455	* @param u64TSC The current TSC.
5456	* @param iTick The current timer tick.
5457	*/
5458	static void supdrvGipDoUpdateCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick)
5459	{
5460	uint64_t u64TSCDelta;
5461	uint32_t u32UpdateIntervalTSC;
5462	uint32_t u32UpdateIntervalTSCSlack;
5463	unsigned iTSCHistoryHead;
5464	uint64_t u64CpuHz;
5465	uint32_t u32TransactionId;
5466
5467	/* Delta between this and the previous update. */
5468	ASMAtomicUoWriteU32(&pGipCpu->u32PrevUpdateIntervalNS, (uint32_t)(u64NanoTS - pGipCpu->u64NanoTS));
5469
5470	/*
5471	* Update the NanoTS.
5472	*/
5473	ASMAtomicWriteU64(&pGipCpu->u64NanoTS, u64NanoTS);
5474
5475	/*
5476	* Calc TSC delta.
5477	*/
5478	/** @todo validate the NanoTS delta, don't trust the OS to call us when it should... */
5479	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
5480	ASMAtomicWriteU64(&pGipCpu->u64TSC, u64TSC);
5481
5482	if (u64TSCDelta >> 32)
5483	{
5484	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
5485	pGipCpu->cErrors++;
5486	}
5487
5488	/*
5489	* On the 2nd and 3rd callout, reset the history with the current TSC
5490	* interval since the values entered by supdrvGipInit are totally off.
5491	* The interval on the 1st callout completely unreliable, the 2nd is a bit
5492	* better, while the 3rd should be most reliable.
5493	*/
5494	u32TransactionId = pGipCpu->u32TransactionId;
5495	if (RT_UNLIKELY( ( u32TransactionId == 5
5496	\|\| u32TransactionId == 7)
5497	&& ( iTick == 2
5498	\|\| iTick == 3) ))
5499	{
5500	unsigned i;
5501	for (i = 0; i < RT_ELEMENTS(pGipCpu->au32TSCHistory); i++)
5502	ASMAtomicUoWriteU32(&pGipCpu->au32TSCHistory[i], (uint32_t)u64TSCDelta);
5503	}
5504
5505	/*
5506	* TSC History.
5507	*/
5508	Assert(RT_ELEMENTS(pGipCpu->au32TSCHistory) == 8);
5509	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
5510	ASMAtomicWriteU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
5511	ASMAtomicWriteU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
5512
5513	/*
5514	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
5515	*/
5516	if (pGip->u32UpdateHz >= 1000)
5517	{
5518	uint32_t u32;
5519	u32 = pGipCpu->au32TSCHistory[0];
5520	u32 += pGipCpu->au32TSCHistory[1];
5521	u32 += pGipCpu->au32TSCHistory[2];
5522	u32 += pGipCpu->au32TSCHistory[3];
5523	u32 >>= 2;
5524	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
5525	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
5526	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
5527	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
5528	u32UpdateIntervalTSC >>= 2;
5529	u32UpdateIntervalTSC += u32;
5530	u32UpdateIntervalTSC >>= 1;
5531
5532	/* Value chosen for a 2GHz Athlon64 running linux 2.6.10/11, . */
5533	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
5534	}
5535	else if (pGip->u32UpdateHz >= 90)
5536	{
5537	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5538	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
5539	u32UpdateIntervalTSC >>= 1;
5540
5541	/* value chosen on a 2GHz thinkpad running windows */
5542	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
5543	}
5544	else
5545	{
5546	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5547
5548	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
5549	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
5550	}
5551	ASMAtomicWriteU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
5552
5553	/*
5554	* CpuHz.
5555	*/
5556	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, pGip->u32UpdateHz);
5557	ASMAtomicWriteU64(&pGipCpu->u64CpuHz, u64CpuHz);
5558	}
5559
5560
5561	/**
5562	* Updates the GIP.
5563	*
5564	* @param pGip Pointer to the GIP.
5565	* @param u64NanoTS The current nanosecond timesamp.
5566	* @param u64TSC The current TSC timesamp.
5567	* @param idCpu The CPU ID.
5568	* @param iTick The current timer tick.
5569	*/
5570	static void supdrvGipUpdate(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick)
5571	{
5572	/*
5573	* Determine the relevant CPU data.
5574	*/
5575	PSUPGIPCPU pGipCpu;
5576	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
5577	pGipCpu = &pGip->aCPUs[0];
5578	else
5579	{
5580	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
5581	if (RT_UNLIKELY(iCpu >= pGip->cCpus))
5582	return;
5583	pGipCpu = &pGip->aCPUs[iCpu];
5584	if (RT_UNLIKELY(pGipCpu->idCpu != idCpu))
5585	return;
5586	}
5587
5588	/*
5589	* Start update transaction.
5590	*/
5591	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5592	{
5593	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
5594	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5595	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5596	pGipCpu->cErrors++;
5597	return;
5598	}
5599
5600	/*
5601	* Recalc the update frequency every 0x800th time.
5602	*/
5603	if (!(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
5604	{
5605	if (pGip->u64NanoTSLastUpdateHz)
5606	{
5607	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
5608	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
5609	uint32_t u32UpdateHz = (uint32_t)((UINT64_C(1000000000) * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
5610	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
5611	{
5612	ASMAtomicWriteU32(&pGip->u32UpdateHz, u32UpdateHz);
5613	ASMAtomicWriteU32(&pGip->u32UpdateIntervalNS, 1000000000 / u32UpdateHz);
5614	}
5615	#endif
5616	}
5617	ASMAtomicWriteU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS);
5618	}
5619
5620	/*
5621	* Update the data.
5622	*/
5623	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS, u64TSC, iTick);
5624
5625	/*
5626	* Complete transaction.
5627	*/
5628	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5629	}
5630
5631
5632	/**
5633	* Updates the per cpu GIP data for the calling cpu.
5634	*
5635	* @param pGip Pointer to the GIP.
5636	* @param u64NanoTS The current nanosecond timesamp.
5637	* @param u64TSC The current TSC timesamp.
5638	* @param idCpu The CPU ID.
5639	* @param idApic The APIC id for the CPU index.
5640	* @param iTick The current timer tick.
5641	*/
5642	static void supdrvGipUpdatePerCpu(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC,
5643	RTCPUID idCpu, uint8_t idApic, uint64_t iTick)
5644	{
5645	uint32_t iCpu;
5646
5647	/*
5648	* Avoid a potential race when a CPU online notification doesn't fire on
5649	* the onlined CPU but the tick creeps in before the event notification is
5650	* run.
5651	*/
5652	if (RT_UNLIKELY(iTick == 1))
5653	{
5654	iCpu = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
5655	if (pGip->aCPUs[iCpu].enmState == SUPGIPCPUSTATE_OFFLINE)
5656	supdrvGipMpEventOnline(pGip, idCpu);
5657	}
5658
5659	iCpu = pGip->aiCpuFromApicId[idApic];
5660	if (RT_LIKELY(iCpu < pGip->cCpus))
5661	{
5662	PSUPGIPCPU pGipCpu = &pGip->aCPUs[iCpu];
5663	if (pGipCpu->idCpu == idCpu)
5664	{
5665	/*
5666	* Start update transaction.
5667	*/
5668	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5669	{
5670	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5671	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5672	pGipCpu->cErrors++;
5673	return;
5674	}
5675
5676	/*
5677	* Update the data.
5678	*/
5679	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS, u64TSC, iTick);
5680
5681	/*
5682	* Complete transaction.
5683	*/
5684	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5685	}
5686	}
5687	}
5688

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

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDrv.c@ 37344

Download in other formats: