SUPDrv.c@ 40611

Last change on this file since 40611 was 40611, checked in by vboxsync, 13 years ago
Init the dtrace bits...
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 214.9 KB

Line
1	/* $Revision: 40611 $ */
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	#include <iprt/x86.h>
58
59	#include <VBox/param.h>
60	#include <VBox/log.h>
61	#include <VBox/err.h>
62	#include <VBox/vmm/hwacc_svm.h>
63	#include <VBox/vmm/hwacc_vmx.h>
64
65	#if defined(RT_OS_SOLARIS) \|\| defined(RT_OS_DARWIN)
66	# include "dtrace/SUPDrv.h"
67	#else
68	# define VBOXDRV_SESSION_CREATE(pvSession, fUser) do { } while (0)
69	# define VBOXDRV_SESSION_CLOSE(pvSession) do { } while (0)
70	# define VBOXDRV_IOCTL_ENTRY(pvSession, uIOCtl, pvReqHdr) do { } while (0)
71	# define VBOXDRV_IOCTL_RETURN(pvSession, uIOCtl, pvReqHdr, rcRet, rcReq) do { } while (0)
72	#endif
73
74	/*
75	* Logging assignments:
76	* Log - useful stuff, like failures.
77	* LogFlow - program flow, except the really noisy bits.
78	* Log2 - Cleanup.
79	* Log3 - Loader flow noise.
80	* Log4 - Call VMMR0 flow noise.
81	* Log5 - Native yet-to-be-defined noise.
82	* Log6 - Native ioctl flow noise.
83	*
84	* Logging requires BUILD_TYPE=debug and possibly changes to the logger
85	* instantiation in log-vbox.c(pp).
86	*/
87
88
89	/*******************************************************************************
90	* Defined Constants And Macros *
91	*******************************************************************************/
92	/** The frequency by which we recalculate the u32UpdateHz and
93	* u32UpdateIntervalNS GIP members. The value must be a power of 2. */
94	#define GIP_UPDATEHZ_RECALC_FREQ 0x800
95
96	/** @def VBOX_SVN_REV
97	* The makefile should define this if it can. */
98	#ifndef VBOX_SVN_REV
99	# define VBOX_SVN_REV 0
100	#endif
101
102	#if 0 /* Don't start the GIP timers. Useful when debugging the IPRT timer code. */
103	# define DO_NOT_START_GIP
104	#endif
105
106
107	/*******************************************************************************
108	* Internal Functions *
109	*******************************************************************************/
110	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser);
111	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser);
112	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession);
113	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType);
114	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq);
115	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq);
116	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq);
117	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq);
118	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq);
119	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx);
120	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt);
121	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage);
122	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage);
123	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt);
124	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt);
125	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq);
126	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq);
127	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt);
128	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt);
129	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
130	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
131	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser);
132	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
133	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned cCpus);
134	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2);
135	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip);
136	static void supdrvGipUpdate(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick);
137	static void supdrvGipUpdatePerCpu(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC,
138	RTCPUID idCpu, uint8_t idApic, uint64_t iTick);
139	static void supdrvGipInitCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pCpu, uint64_t u64NanoTS);
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	{ "RTAssertMsg2AddV", (void *)RTAssertMsg2AddV },
340	{ "RTAssertSetQuiet", (void *)RTAssertSetQuiet },
341	{ "RTAssertMayPanic", (void *)RTAssertMayPanic },
342	{ "RTAssertSetMayPanic", (void *)RTAssertSetMayPanic },
343	{ "RTAssertAreQuiet", (void *)RTAssertAreQuiet },
344	{ "RTStrFormat", (void *)RTStrFormat },
345	{ "RTStrFormatNumber", (void *)RTStrFormatNumber },
346	{ "RTStrFormatTypeDeregister", (void *)RTStrFormatTypeDeregister },
347	{ "RTStrFormatTypeRegister", (void *)RTStrFormatTypeRegister },
348	{ "RTStrFormatTypeSetUser", (void *)RTStrFormatTypeSetUser },
349	{ "RTStrFormatV", (void *)RTStrFormatV },
350	{ "RTStrPrintf", (void *)RTStrPrintf },
351	{ "RTStrPrintfEx", (void *)RTStrPrintfEx },
352	{ "RTStrPrintfExV", (void *)RTStrPrintfExV },
353	{ "RTStrPrintfV", (void *)RTStrPrintfV },
354	{ "RTHandleTableAllocWithCtx", (void *)RTHandleTableAllocWithCtx },
355	{ "RTHandleTableCreate", (void *)RTHandleTableCreate },
356	{ "RTHandleTableCreateEx", (void *)RTHandleTableCreateEx },
357	{ "RTHandleTableDestroy", (void *)RTHandleTableDestroy },
358	{ "RTHandleTableFreeWithCtx", (void *)RTHandleTableFreeWithCtx },
359	{ "RTHandleTableLookupWithCtx", (void *)RTHandleTableLookupWithCtx },
360	{ "RTNetIPv4AddDataChecksum", (void *)RTNetIPv4AddDataChecksum },
361	{ "RTNetIPv4AddTCPChecksum", (void *)RTNetIPv4AddTCPChecksum },
362	{ "RTNetIPv4AddUDPChecksum", (void *)RTNetIPv4AddUDPChecksum },
363	{ "RTNetIPv4FinalizeChecksum", (void *)RTNetIPv4FinalizeChecksum },
364	{ "RTNetIPv4HdrChecksum", (void *)RTNetIPv4HdrChecksum },
365	{ "RTNetIPv4IsDHCPValid", (void *)RTNetIPv4IsDHCPValid },
366	{ "RTNetIPv4IsHdrValid", (void *)RTNetIPv4IsHdrValid },
367	{ "RTNetIPv4IsTCPSizeValid", (void *)RTNetIPv4IsTCPSizeValid },
368	{ "RTNetIPv4IsTCPValid", (void *)RTNetIPv4IsTCPValid },
369	{ "RTNetIPv4IsUDPSizeValid", (void *)RTNetIPv4IsUDPSizeValid },
370	{ "RTNetIPv4IsUDPValid", (void *)RTNetIPv4IsUDPValid },
371	{ "RTNetIPv4PseudoChecksum", (void *)RTNetIPv4PseudoChecksum },
372	{ "RTNetIPv4PseudoChecksumBits", (void *)RTNetIPv4PseudoChecksumBits },
373	{ "RTNetIPv4TCPChecksum", (void *)RTNetIPv4TCPChecksum },
374	{ "RTNetIPv4UDPChecksum", (void *)RTNetIPv4UDPChecksum },
375	{ "RTNetIPv6PseudoChecksum", (void *)RTNetIPv6PseudoChecksum },
376	{ "RTNetIPv6PseudoChecksumBits", (void *)RTNetIPv6PseudoChecksumBits },
377	{ "RTNetIPv6PseudoChecksumEx", (void *)RTNetIPv6PseudoChecksumEx },
378	{ "RTNetTCPChecksum", (void *)RTNetTCPChecksum },
379	{ "RTNetUDPChecksum", (void *)RTNetUDPChecksum },
380	{ "RTCrc32", (void *)RTCrc32 },
381	{ "RTCrc32Finish", (void *)RTCrc32Finish },
382	{ "RTCrc32Process", (void *)RTCrc32Process },
383	{ "RTCrc32Start", (void *)RTCrc32Start },
384	};
385
386	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
387	/**
388	* Drag in the rest of IRPT since we share it with the
389	* rest of the kernel modules on darwin.
390	*/
391	PFNRT g_apfnVBoxDrvIPRTDeps[] =
392	{
393	/* VBoxNetFlt */
394	(PFNRT)RTErrConvertFromErrno,
395	(PFNRT)RTUuidCompare,
396	(PFNRT)RTUuidCompareStr,
397	(PFNRT)RTUuidFromStr,
398	(PFNRT)RTStrDupTag,
399	(PFNRT)RTStrFree,
400	(PFNRT)RTStrCopy,
401	(PFNRT)RTStrNCmp,
402	/* VBoxNetAdp */
403	(PFNRT)RTRandBytes,
404	/* VBoxUSB */
405	(PFNRT)RTPathStripFilename,
406	NULL
407	};
408	#endif /* RT_OS_DARWIN \|\| RT_OS_SOLARIS \|\| RT_OS_SOLARIS */
409
410
411	/**
412	* Initializes the device extentsion structure.
413	*
414	* @returns IPRT status code.
415	* @param pDevExt The device extension to initialize.
416	* @param cbSession The size of the session structure. The size of
417	* SUPDRVSESSION may be smaller when SUPDRV_AGNOSTIC is
418	* defined because we're skipping the OS specific members
419	* then.
420	*/
421	int VBOXCALL supdrvInitDevExt(PSUPDRVDEVEXT pDevExt, size_t cbSession)
422	{
423	int rc;
424
425	#ifdef SUPDRV_WITH_RELEASE_LOGGER
426	/*
427	* Create the release log.
428	*/
429	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
430	PRTLOGGER pRelLogger;
431	rc = RTLogCreate(&pRelLogger, 0 /* fFlags */, "all",
432	"VBOX_RELEASE_LOG", RT_ELEMENTS(s_apszGroups), s_apszGroups, RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER, NULL);
433	if (RT_SUCCESS(rc))
434	RTLogRelSetDefaultInstance(pRelLogger);
435	/** @todo Add native hook for getting logger config parameters and setting
436	* them. On linux we should use the module parameter stuff... */
437	#endif
438
439	/*
440	* Initialize it.
441	*/
442	memset(pDevExt, 0, sizeof(*pDevExt));
443	rc = RTSpinlockCreate(&pDevExt->Spinlock);
444	if (RT_SUCCESS(rc))
445	{
446	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
447	rc = RTSemMutexCreate(&pDevExt->mtxLdr);
448	#else
449	rc = RTSemFastMutexCreate(&pDevExt->mtxLdr);
450	#endif
451	if (RT_SUCCESS(rc))
452	{
453	rc = RTSemFastMutexCreate(&pDevExt->mtxComponentFactory);
454	if (RT_SUCCESS(rc))
455	{
456	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
457	rc = RTSemMutexCreate(&pDevExt->mtxGip);
458	#else
459	rc = RTSemFastMutexCreate(&pDevExt->mtxGip);
460	#endif
461	if (RT_SUCCESS(rc))
462	{
463	rc = supdrvGipCreate(pDevExt);
464	if (RT_SUCCESS(rc))
465	{
466	#ifdef VBOX_WITH_DTRACE_R0DRV
467	rc = supdrvDTraceInit(pDevExt);
468	if (RT_SUCCESS(rc))
469	#endif
470	{
471
472	pDevExt->u32Cookie = BIRD; /** @todo make this random? */
473	pDevExt->cbSession = (uint32_t)cbSession;
474
475	/*
476	* Fixup the absolute symbols.
477	*
478	* Because of the table indexing assumptions we'll have a little #ifdef orgy
479	* here rather than distributing this to OS specific files. At least for now.
480	*/
481	#ifdef RT_OS_DARWIN
482	# if ARCH_BITS == 32
483	if (SUPR0GetPagingMode() >= SUPPAGINGMODE_AMD64)
484	{
485	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
486	g_aFunctions[1].pfn = (void )0x80; / SUPR0Abs64bitKernelCS - KERNEL64_CS, seg.h */
487	g_aFunctions[2].pfn = (void )0x88; / SUPR0Abs64bitKernelSS - KERNEL64_SS, seg.h */
488	g_aFunctions[3].pfn = (void )0x88; / SUPR0Abs64bitKernelDS - KERNEL64_SS, seg.h */
489	}
490	else
491	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
492	g_aFunctions[4].pfn = (void )0x08; / SUPR0AbsKernelCS - KERNEL_CS, seg.h */
493	g_aFunctions[5].pfn = (void )0x10; / SUPR0AbsKernelSS - KERNEL_DS, seg.h */
494	g_aFunctions[6].pfn = (void )0x10; / SUPR0AbsKernelDS - KERNEL_DS, seg.h */
495	g_aFunctions[7].pfn = (void )0x10; / SUPR0AbsKernelES - KERNEL_DS, seg.h */
496	g_aFunctions[8].pfn = (void )0x10; / SUPR0AbsKernelFS - KERNEL_DS, seg.h */
497	g_aFunctions[9].pfn = (void )0x48; / SUPR0AbsKernelGS - CPU_DATA_GS, seg.h */
498	# else /* 64-bit darwin: */
499	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
500	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
501	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
502	g_aFunctions[3].pfn = (void )0; / SUPR0Abs64bitKernelDS */
503	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
504	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
505	g_aFunctions[6].pfn = (void )0; / SUPR0AbsKernelDS */
506	g_aFunctions[7].pfn = (void )0; / SUPR0AbsKernelES */
507	g_aFunctions[8].pfn = (void )0; / SUPR0AbsKernelFS */
508	g_aFunctions[9].pfn = (void )0; / SUPR0AbsKernelGS */
509
510	# endif
511	#else /* !RT_OS_DARWIN */
512	# if ARCH_BITS == 64
513	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
514	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
515	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
516	g_aFunctions[3].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0Abs64bitKernelDS */
517	# else
518	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
519	# endif
520	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
521	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
522	g_aFunctions[6].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0AbsKernelDS */
523	g_aFunctions[7].pfn = (void )(uintptr_t)ASMGetES(); / SUPR0AbsKernelES */
524	g_aFunctions[8].pfn = (void )(uintptr_t)ASMGetFS(); / SUPR0AbsKernelFS */
525	g_aFunctions[9].pfn = (void )(uintptr_t)ASMGetGS(); / SUPR0AbsKernelGS */
526	#endif /* !RT_OS_DARWIN */
527	return VINF_SUCCESS;
528	}
529
530	#ifdef VBOX_WITH_DTRACE_R0DRV
531	supdrvGipDestroy(pDevExt);
532	#endif
533	}
534
535	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
536	RTSemMutexDestroy(pDevExt->mtxGip);
537	pDevExt->mtxGip = NIL_RTSEMMUTEX;
538	#else
539	RTSemFastMutexDestroy(pDevExt->mtxGip);
540	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
541	#endif
542	}
543	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
544	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
545	}
546	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
547	RTSemMutexDestroy(pDevExt->mtxLdr);
548	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
549	#else
550	RTSemFastMutexDestroy(pDevExt->mtxLdr);
551	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
552	#endif
553	}
554	RTSpinlockDestroy(pDevExt->Spinlock);
555	pDevExt->Spinlock = NIL_RTSPINLOCK;
556	}
557	#ifdef SUPDRV_WITH_RELEASE_LOGGER
558	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
559	RTLogDestroy(RTLogSetDefaultInstance(NULL));
560	#endif
561
562	return rc;
563	}
564
565
566	/**
567	* Delete the device extension (e.g. cleanup members).
568	*
569	* @param pDevExt The device extension to delete.
570	*/
571	void VBOXCALL supdrvDeleteDevExt(PSUPDRVDEVEXT pDevExt)
572	{
573	PSUPDRVOBJ pObj;
574	PSUPDRVUSAGE pUsage;
575
576	/*
577	* Kill mutexes and spinlocks.
578	*/
579	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
580	RTSemMutexDestroy(pDevExt->mtxGip);
581	pDevExt->mtxGip = NIL_RTSEMMUTEX;
582	#else
583	RTSemFastMutexDestroy(pDevExt->mtxGip);
584	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
585	#endif
586	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
587	RTSemMutexDestroy(pDevExt->mtxLdr);
588	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
589	#else
590	RTSemFastMutexDestroy(pDevExt->mtxLdr);
591	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
592	#endif
593	RTSpinlockDestroy(pDevExt->Spinlock);
594	pDevExt->Spinlock = NIL_RTSPINLOCK;
595	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
596	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
597
598	/*
599	* Free lists.
600	*/
601	/* objects. */
602	pObj = pDevExt->pObjs;
603	Assert(!pObj); /* (can trigger on forced unloads) */
604	pDevExt->pObjs = NULL;
605	while (pObj)
606	{
607	void *pvFree = pObj;
608	pObj = pObj->pNext;
609	RTMemFree(pvFree);
610	}
611
612	/* usage records. */
613	pUsage = pDevExt->pUsageFree;
614	pDevExt->pUsageFree = NULL;
615	while (pUsage)
616	{
617	void *pvFree = pUsage;
618	pUsage = pUsage->pNext;
619	RTMemFree(pvFree);
620	}
621
622	/* kill the GIP. */
623	supdrvGipDestroy(pDevExt);
624
625	#ifdef VBOX_WITH_DTRACE_R0DRV
626	supdrvDTraceTerm(pDevExt);
627	#endif
628
629	#ifdef SUPDRV_WITH_RELEASE_LOGGER
630	/* destroy the loggers. */
631	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
632	RTLogDestroy(RTLogSetDefaultInstance(NULL));
633	#endif
634	}
635
636
637	/**
638	* Create session.
639	*
640	* @returns IPRT status code.
641	* @param pDevExt Device extension.
642	* @param fUser Flag indicating whether this is a user or kernel session.
643	* @param ppSession Where to store the pointer to the session data.
644	*/
645	int VBOXCALL supdrvCreateSession(PSUPDRVDEVEXT pDevExt, bool fUser, PSUPDRVSESSION *ppSession)
646	{
647	/*
648	* Allocate memory for the session data.
649	*/
650	int rc;
651	PSUPDRVSESSION pSession = *ppSession = (PSUPDRVSESSION)RTMemAllocZ(pDevExt->cbSession);
652	if (pSession)
653	{
654	/* Initialize session data. */
655	rc = RTSpinlockCreate(&pSession->Spinlock);
656	if (!rc)
657	{
658	rc = RTHandleTableCreateEx(&pSession->hHandleTable,
659	RTHANDLETABLE_FLAGS_LOCKED \| RTHANDLETABLE_FLAGS_CONTEXT,
660	1 /uBase/, 32768 /cMax/, supdrvSessionObjHandleRetain, pSession);
661	if (RT_SUCCESS(rc))
662	{
663	Assert(pSession->Spinlock != NIL_RTSPINLOCK);
664	pSession->pDevExt = pDevExt;
665	pSession->u32Cookie = BIRD_INV;
666	/*pSession->pLdrUsage = NULL;
667	pSession->pVM = NULL;
668	pSession->pUsage = NULL;
669	pSession->pGip = NULL;
670	pSession->fGipReferenced = false;
671	pSession->Bundle.cUsed = 0; */
672	pSession->Uid = NIL_RTUID;
673	pSession->Gid = NIL_RTGID;
674	if (fUser)
675	{
676	pSession->Process = RTProcSelf();
677	pSession->R0Process = RTR0ProcHandleSelf();
678	}
679	else
680	{
681	pSession->Process = NIL_RTPROCESS;
682	pSession->R0Process = NIL_RTR0PROCESS;
683	}
684
685	VBOXDRV_SESSION_CREATE(pSession, fUser);
686	LogFlow(("Created session %p initial cookie=%#x\n", pSession, pSession->u32Cookie));
687	return VINF_SUCCESS;
688	}
689
690	RTSpinlockDestroy(pSession->Spinlock);
691	}
692	RTMemFree(pSession);
693	*ppSession = NULL;
694	Log(("Failed to create spinlock, rc=%d!\n", rc));
695	}
696	else
697	rc = VERR_NO_MEMORY;
698
699	return rc;
700	}
701
702
703	/**
704	* Shared code for cleaning up a session.
705	*
706	* @param pDevExt Device extension.
707	* @param pSession Session data.
708	* This data will be freed by this routine.
709	*/
710	void VBOXCALL supdrvCloseSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
711	{
712	VBOXDRV_SESSION_CLOSE(pSession);
713
714	/*
715	* Cleanup the session first.
716	*/
717	supdrvCleanupSession(pDevExt, pSession);
718
719	/*
720	* Free the rest of the session stuff.
721	*/
722	RTSpinlockDestroy(pSession->Spinlock);
723	pSession->Spinlock = NIL_RTSPINLOCK;
724	pSession->pDevExt = NULL;
725	RTMemFree(pSession);
726	LogFlow(("supdrvCloseSession: returns\n"));
727	}
728
729
730	/**
731	* Shared code for cleaning up a session (but not quite freeing it).
732	*
733	* This is primarily intended for MAC OS X where we have to clean up the memory
734	* stuff before the file handle is closed.
735	*
736	* @param pDevExt Device extension.
737	* @param pSession Session data.
738	* This data will be freed by this routine.
739	*/
740	void VBOXCALL supdrvCleanupSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
741	{
742	int rc;
743	PSUPDRVBUNDLE pBundle;
744	LogFlow(("supdrvCleanupSession: pSession=%p\n", pSession));
745
746	/*
747	* Remove logger instances related to this session.
748	*/
749	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pSession);
750
751	/*
752	* Destroy the handle table.
753	*/
754	rc = RTHandleTableDestroy(pSession->hHandleTable, supdrvSessionObjHandleDelete, pSession);
755	AssertRC(rc);
756	pSession->hHandleTable = NIL_RTHANDLETABLE;
757
758	/*
759	* Release object references made in this session.
760	* In theory there should be noone racing us in this session.
761	*/
762	Log2(("release objects - start\n"));
763	if (pSession->pUsage)
764	{
765	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
766	PSUPDRVUSAGE pUsage;
767	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
768
769	while ((pUsage = pSession->pUsage) != NULL)
770	{
771	PSUPDRVOBJ pObj = pUsage->pObj;
772	pSession->pUsage = pUsage->pNext;
773
774	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
775	if (pUsage->cUsage < pObj->cUsage)
776	{
777	pObj->cUsage -= pUsage->cUsage;
778	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
779	}
780	else
781	{
782	/* Destroy the object and free the record. */
783	if (pDevExt->pObjs == pObj)
784	pDevExt->pObjs = pObj->pNext;
785	else
786	{
787	PSUPDRVOBJ pObjPrev;
788	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
789	if (pObjPrev->pNext == pObj)
790	{
791	pObjPrev->pNext = pObj->pNext;
792	break;
793	}
794	Assert(pObjPrev);
795	}
796	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
797
798	Log(("supdrvCleanupSession: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
799	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
800	if (pObj->pfnDestructor)
801	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
802	RTMemFree(pObj);
803	}
804
805	/* free it and continue. */
806	RTMemFree(pUsage);
807
808	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
809	}
810
811	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
812	AssertMsg(!pSession->pUsage, ("Some buster reregistered an object during desturction!\n"));
813	}
814	Log2(("release objects - done\n"));
815
816	/*
817	* Release memory allocated in the session.
818	*
819	* We do not serialize this as we assume that the application will
820	* not allocated memory while closing the file handle object.
821	*/
822	Log2(("freeing memory:\n"));
823	pBundle = &pSession->Bundle;
824	while (pBundle)
825	{
826	PSUPDRVBUNDLE pToFree;
827	unsigned i;
828
829	/*
830	* Check and unlock all entries in the bundle.
831	*/
832	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
833	{
834	if (pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ)
835	{
836	Log2(("eType=%d pvR0=%p pvR3=%p cb=%ld\n", pBundle->aMem[i].eType, RTR0MemObjAddress(pBundle->aMem[i].MemObj),
837	(void *)RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3), (long)RTR0MemObjSize(pBundle->aMem[i].MemObj)));
838	if (pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ)
839	{
840	rc = RTR0MemObjFree(pBundle->aMem[i].MapObjR3, false);
841	AssertRC(rc); /** @todo figure out how to handle this. */
842	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
843	}
844	rc = RTR0MemObjFree(pBundle->aMem[i].MemObj, true /* fFreeMappings */);
845	AssertRC(rc); /** @todo figure out how to handle this. */
846	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
847	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
848	}
849	}
850
851	/*
852	* Advance and free previous bundle.
853	*/
854	pToFree = pBundle;
855	pBundle = pBundle->pNext;
856
857	pToFree->pNext = NULL;
858	pToFree->cUsed = 0;
859	if (pToFree != &pSession->Bundle)
860	RTMemFree(pToFree);
861	}
862	Log2(("freeing memory - done\n"));
863
864	/*
865	* Deregister component factories.
866	*/
867	RTSemFastMutexRequest(pDevExt->mtxComponentFactory);
868	Log2(("deregistering component factories:\n"));
869	if (pDevExt->pComponentFactoryHead)
870	{
871	PSUPDRVFACTORYREG pPrev = NULL;
872	PSUPDRVFACTORYREG pCur = pDevExt->pComponentFactoryHead;
873	while (pCur)
874	{
875	if (pCur->pSession == pSession)
876	{
877	/* unlink it */
878	PSUPDRVFACTORYREG pNext = pCur->pNext;
879	if (pPrev)
880	pPrev->pNext = pNext;
881	else
882	pDevExt->pComponentFactoryHead = pNext;
883
884	/* free it */
885	pCur->pNext = NULL;
886	pCur->pSession = NULL;
887	pCur->pFactory = NULL;
888	RTMemFree(pCur);
889
890	/* next */
891	pCur = pNext;
892	}
893	else
894	{
895	/* next */
896	pPrev = pCur;
897	pCur = pCur->pNext;
898	}
899	}
900	}
901	RTSemFastMutexRelease(pDevExt->mtxComponentFactory);
902	Log2(("deregistering component factories - done\n"));
903
904	/*
905	* Loaded images needs to be dereferenced and possibly freed up.
906	*/
907	supdrvLdrLock(pDevExt);
908	Log2(("freeing images:\n"));
909	if (pSession->pLdrUsage)
910	{
911	PSUPDRVLDRUSAGE pUsage = pSession->pLdrUsage;
912	pSession->pLdrUsage = NULL;
913	while (pUsage)
914	{
915	void *pvFree = pUsage;
916	PSUPDRVLDRIMAGE pImage = pUsage->pImage;
917	if (pImage->cUsage > pUsage->cUsage)
918	pImage->cUsage -= pUsage->cUsage;
919	else
920	supdrvLdrFree(pDevExt, pImage);
921	pUsage->pImage = NULL;
922	pUsage = pUsage->pNext;
923	RTMemFree(pvFree);
924	}
925	}
926	supdrvLdrUnlock(pDevExt);
927	Log2(("freeing images - done\n"));
928
929	/*
930	* Unmap the GIP.
931	*/
932	Log2(("umapping GIP:\n"));
933	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
934	{
935	SUPR0GipUnmap(pSession);
936	pSession->fGipReferenced = 0;
937	}
938	Log2(("umapping GIP - done\n"));
939	}
940
941
942	/**
943	* RTHandleTableDestroy callback used by supdrvCleanupSession.
944	*
945	* @returns IPRT status code, see SUPR0ObjAddRef.
946	* @param hHandleTable The handle table handle. Ignored.
947	* @param pvObj The object pointer.
948	* @param pvCtx Context, the handle type. Ignored.
949	* @param pvUser Session pointer.
950	*/
951	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser)
952	{
953	NOREF(pvCtx);
954	NOREF(hHandleTable);
955	return SUPR0ObjAddRefEx(pvObj, (PSUPDRVSESSION)pvUser, true /fNoBlocking/);
956	}
957
958
959	/**
960	* RTHandleTableDestroy callback used by supdrvCleanupSession.
961	*
962	* @param hHandleTable The handle table handle. Ignored.
963	* @param h The handle value. Ignored.
964	* @param pvObj The object pointer.
965	* @param pvCtx Context, the handle type. Ignored.
966	* @param pvUser Session pointer.
967	*/
968	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser)
969	{
970	NOREF(pvCtx);
971	NOREF(h);
972	NOREF(hHandleTable);
973	SUPR0ObjRelease(pvObj, (PSUPDRVSESSION)pvUser);
974	}
975
976
977	/**
978	* Fast path I/O Control worker.
979	*
980	* @returns VBox status code that should be passed down to ring-3 unchanged.
981	* @param uIOCtl Function number.
982	* @param idCpu VMCPU id.
983	* @param pDevExt Device extention.
984	* @param pSession Session data.
985	*/
986	int VBOXCALL supdrvIOCtlFast(uintptr_t uIOCtl, VMCPUID idCpu, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
987	{
988	/*
989	* We check the two prereqs after doing this only to allow the compiler to optimize things better.
990	*/
991	if (RT_LIKELY( RT_VALID_PTR(pSession)
992	&& pSession->pVM
993	&& pDevExt->pfnVMMR0EntryFast))
994	{
995	switch (uIOCtl)
996	{
997	case SUP_IOCTL_FAST_DO_RAW_RUN:
998	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_RAW_RUN);
999	break;
1000	case SUP_IOCTL_FAST_DO_HWACC_RUN:
1001	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_HWACC_RUN);
1002	break;
1003	case SUP_IOCTL_FAST_DO_NOP:
1004	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_NOP);
1005	break;
1006	default:
1007	return VERR_INTERNAL_ERROR;
1008	}
1009	return VINF_SUCCESS;
1010	}
1011	return VERR_INTERNAL_ERROR;
1012	}
1013
1014
1015	/**
1016	* Helper for supdrvIOCtl. Check if pszStr contains any character of pszChars.
1017	* We would use strpbrk here if this function would be contained in the RedHat kABI white
1018	* list, see http://www.kerneldrivers.org/RHEL5.
1019	*
1020	* @returns 1 if pszStr does contain any character of pszChars, 0 otherwise.
1021	* @param pszStr String to check
1022	* @param pszChars Character set
1023	*/
1024	static int supdrvCheckInvalidChar(const char pszStr, const char pszChars)
1025	{
1026	int chCur;
1027	while ((chCur = *pszStr++) != '\0')
1028	{
1029	int ch;
1030	const char *psz = pszChars;
1031	while ((ch = *psz++) != '\0')
1032	if (ch == chCur)
1033	return 1;
1034
1035	}
1036	return 0;
1037	}
1038
1039
1040	/**
1041	* I/O Control worker.
1042	*
1043	* @returns IPRT status code.
1044	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1045	*
1046	* @param uIOCtl Function number.
1047	* @param pDevExt Device extention.
1048	* @param pSession Session data.
1049	* @param pReqHdr The request header.
1050	*/
1051	int VBOXCALL supdrvIOCtl(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1052	{
1053	VBOXDRV_IOCTL_ENTRY(pSession, uIOCtl, pReqHdr);
1054
1055	/*
1056	* Validate the request.
1057	*/
1058	/* this first check could probably be omitted as its also done by the OS specific code... */
1059	if (RT_UNLIKELY( (pReqHdr->fFlags & SUPREQHDR_FLAGS_MAGIC_MASK) != SUPREQHDR_FLAGS_MAGIC
1060	\|\| pReqHdr->cbIn < sizeof(*pReqHdr)
1061	\|\| pReqHdr->cbOut < sizeof(*pReqHdr)))
1062	{
1063	OSDBGPRINT(("vboxdrv: Bad ioctl request header; cbIn=%#lx cbOut=%#lx fFlags=%#lx\n",
1064	(long)pReqHdr->cbIn, (long)pReqHdr->cbOut, (long)pReqHdr->fFlags));
1065	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1066	return VERR_INVALID_PARAMETER;
1067	}
1068	if (RT_UNLIKELY(!RT_VALID_PTR(pSession)))
1069	{
1070	OSDBGPRINT(("vboxdrv: Invalid pSession valud %p (ioctl=%p)\n", pSession, (void *)uIOCtl));
1071	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1072	return VERR_INVALID_PARAMETER;
1073	}
1074	if (RT_UNLIKELY(uIOCtl == SUP_IOCTL_COOKIE))
1075	{
1076	if (pReqHdr->u32Cookie != SUPCOOKIE_INITIAL_COOKIE)
1077	{
1078	OSDBGPRINT(("SUP_IOCTL_COOKIE: bad cookie %#lx\n", (long)pReqHdr->u32Cookie));
1079	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1080	return VERR_INVALID_PARAMETER;
1081	}
1082	}
1083	else if (RT_UNLIKELY( pReqHdr->u32Cookie != pDevExt->u32Cookie
1084	\|\| pReqHdr->u32SessionCookie != pSession->u32Cookie))
1085	{
1086	OSDBGPRINT(("vboxdrv: bad cookie %#lx / %#lx.\n", (long)pReqHdr->u32Cookie, (long)pReqHdr->u32SessionCookie));
1087	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1088	return VERR_INVALID_PARAMETER;
1089	}
1090
1091	/*
1092	* Validation macros
1093	*/
1094	#define REQ_CHECK_SIZES_EX(Name, cbInExpect, cbOutExpect) \
1095	do { \
1096	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect) \|\| pReqHdr->cbOut != (cbOutExpect))) \
1097	{ \
1098	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld. cbOut=%ld expected %ld.\n", \
1099	(long)pReq->Hdr.cbIn, (long)(cbInExpect), (long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
1100	VBOXDRV_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_SIZES(Name) REQ_CHECK_SIZES_EX(Name, Name ## _SIZE_IN, Name ## _SIZE_OUT)
1106
1107	#define REQ_CHECK_SIZE_IN(Name, cbInExpect) \
1108	do { \
1109	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect))) \
1110	{ \
1111	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld.\n", \
1112	(long)pReq->Hdr.cbIn, (long)(cbInExpect))); \
1113	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1114	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1115	} \
1116	} while (0)
1117
1118	#define REQ_CHECK_SIZE_OUT(Name, cbOutExpect) \
1119	do { \
1120	if (RT_UNLIKELY(pReqHdr->cbOut != (cbOutExpect))) \
1121	{ \
1122	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbOut=%ld expected %ld.\n", \
1123	(long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
1124	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1125	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1126	} \
1127	} while (0)
1128
1129	#define REQ_CHECK_EXPR(Name, expr) \
1130	do { \
1131	if (RT_UNLIKELY(!(expr))) \
1132	{ \
1133	OSDBGPRINT(( #Name ": %s\n", #expr)); \
1134	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1135	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1136	} \
1137	} while (0)
1138
1139	#define REQ_CHECK_EXPR_FMT(expr, fmt) \
1140	do { \
1141	if (RT_UNLIKELY(!(expr))) \
1142	{ \
1143	OSDBGPRINT( fmt ); \
1144	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1145	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1146	} \
1147	} while (0)
1148
1149	/*
1150	* The switch.
1151	*/
1152	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
1153	{
1154	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
1155	{
1156	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
1157	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
1158	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
1159	{
1160	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
1161	pReq->Hdr.rc = VERR_INVALID_MAGIC;
1162	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, VERR_INVALID_MAGIC);
1163	return 0;
1164	}
1165
1166	#if 0
1167	/*
1168	* Call out to the OS specific code and let it do permission checks on the
1169	* client process.
1170	*/
1171	if (!supdrvOSValidateClientProcess(pDevExt, pSession))
1172	{
1173	pReq->u.Out.u32Cookie = 0xffffffff;
1174	pReq->u.Out.u32SessionCookie = 0xffffffff;
1175	pReq->u.Out.u32SessionVersion = 0xffffffff;
1176	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1177	pReq->u.Out.pSession = NULL;
1178	pReq->u.Out.cFunctions = 0;
1179	pReq->Hdr.rc = VERR_PERMISSION_DENIED;
1180	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, VERR_PERMISSION_DENIED);
1181	return 0;
1182	}
1183	#endif
1184
1185	/*
1186	* Match the version.
1187	* The current logic is very simple, match the major interface version.
1188	*/
1189	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
1190	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
1191	{
1192	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1193	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
1194	pReq->u.Out.u32Cookie = 0xffffffff;
1195	pReq->u.Out.u32SessionCookie = 0xffffffff;
1196	pReq->u.Out.u32SessionVersion = 0xffffffff;
1197	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1198	pReq->u.Out.pSession = NULL;
1199	pReq->u.Out.cFunctions = 0;
1200	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1201	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1202	return 0;
1203	}
1204
1205	/*
1206	* Fill in return data and be gone.
1207	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
1208	* u32SessionVersion <= u32ReqVersion!
1209	*/
1210	/** @todo Somehow validate the client and negotiate a secure cookie... */
1211	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
1212	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
1213	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
1214	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1215	pReq->u.Out.pSession = pSession;
1216	pReq->u.Out.cFunctions = sizeof(g_aFunctions) / sizeof(g_aFunctions[0]);
1217	pReq->Hdr.rc = VINF_SUCCESS;
1218	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1219	return 0;
1220	}
1221
1222	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_QUERY_FUNCS(0)):
1223	{
1224	/* validate */
1225	PSUPQUERYFUNCS pReq = (PSUPQUERYFUNCS)pReqHdr;
1226	REQ_CHECK_SIZES_EX(SUP_IOCTL_QUERY_FUNCS, SUP_IOCTL_QUERY_FUNCS_SIZE_IN, SUP_IOCTL_QUERY_FUNCS_SIZE_OUT(RT_ELEMENTS(g_aFunctions)));
1227
1228	/* execute */
1229	pReq->u.Out.cFunctions = RT_ELEMENTS(g_aFunctions);
1230	memcpy(&pReq->u.Out.aFunctions[0], g_aFunctions, sizeof(g_aFunctions));
1231	pReq->Hdr.rc = VINF_SUCCESS;
1232	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1233	return 0;
1234	}
1235
1236	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_LOCK):
1237	{
1238	/* validate */
1239	PSUPPAGELOCK pReq = (PSUPPAGELOCK)pReqHdr;
1240	REQ_CHECK_SIZE_IN(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_IN);
1241	REQ_CHECK_SIZE_OUT(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_OUT(pReq->u.In.cPages));
1242	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.cPages > 0);
1243	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.pvR3 >= PAGE_SIZE);
1244
1245	/* execute */
1246	pReq->Hdr.rc = SUPR0LockMem(pSession, pReq->u.In.pvR3, pReq->u.In.cPages, &pReq->u.Out.aPages[0]);
1247	if (RT_FAILURE(pReq->Hdr.rc))
1248	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1249	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1250	return 0;
1251	}
1252
1253	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_UNLOCK):
1254	{
1255	/* validate */
1256	PSUPPAGEUNLOCK pReq = (PSUPPAGEUNLOCK)pReqHdr;
1257	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_UNLOCK);
1258
1259	/* execute */
1260	pReq->Hdr.rc = SUPR0UnlockMem(pSession, pReq->u.In.pvR3);
1261	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1262	return 0;
1263	}
1264
1265	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_ALLOC):
1266	{
1267	/* validate */
1268	PSUPCONTALLOC pReq = (PSUPCONTALLOC)pReqHdr;
1269	REQ_CHECK_SIZES(SUP_IOCTL_CONT_ALLOC);
1270
1271	/* execute */
1272	pReq->Hdr.rc = SUPR0ContAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.HCPhys);
1273	if (RT_FAILURE(pReq->Hdr.rc))
1274	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1275	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1276	return 0;
1277	}
1278
1279	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_FREE):
1280	{
1281	/* validate */
1282	PSUPCONTFREE pReq = (PSUPCONTFREE)pReqHdr;
1283	REQ_CHECK_SIZES(SUP_IOCTL_CONT_FREE);
1284
1285	/* execute */
1286	pReq->Hdr.rc = SUPR0ContFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1287	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1288	return 0;
1289	}
1290
1291	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_OPEN):
1292	{
1293	/* validate */
1294	PSUPLDROPEN pReq = (PSUPLDROPEN)pReqHdr;
1295	REQ_CHECK_SIZES(SUP_IOCTL_LDR_OPEN);
1296	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs > 0);
1297	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs < 16*_1M);
1298	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1299	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1300	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits < pReq->u.In.cbImageWithTabs);
1301	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.szName[0]);
1302	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1303	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, !supdrvCheckInvalidChar(pReq->u.In.szName, ";:()[]{}/\\\|&*%#@!~`\"'"));
1304	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szFilename, sizeof(pReq->u.In.szFilename)));
1305
1306	/* execute */
1307	pReq->Hdr.rc = supdrvIOCtl_LdrOpen(pDevExt, pSession, pReq);
1308	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1309	return 0;
1310	}
1311
1312	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOAD):
1313	{
1314	/* validate */
1315	PSUPLDRLOAD pReq = (PSUPLDRLOAD)pReqHdr;
1316	REQ_CHECK_EXPR(Name, pReq->Hdr.cbIn >= sizeof(*pReq));
1317	REQ_CHECK_SIZES_EX(SUP_IOCTL_LDR_LOAD, SUP_IOCTL_LDR_LOAD_SIZE_IN(pReq->u.In.cbImageWithTabs), SUP_IOCTL_LDR_LOAD_SIZE_OUT);
1318	REQ_CHECK_EXPR(SUP_IOCTL_LDR_LOAD, pReq->u.In.cSymbols <= 16384);
1319	REQ_CHECK_EXPR_FMT( !pReq->u.In.cSymbols
1320	\|\| ( pReq->u.In.offSymbols < pReq->u.In.cbImageWithTabs
1321	&& pReq->u.In.offSymbols + pReq->u.In.cSymbols * sizeof(SUPLDRSYM) <= pReq->u.In.cbImageWithTabs),
1322	("SUP_IOCTL_LDR_LOAD: offSymbols=%#lx cSymbols=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offSymbols,
1323	(long)pReq->u.In.cSymbols, (long)pReq->u.In.cbImageWithTabs));
1324	REQ_CHECK_EXPR_FMT( !pReq->u.In.cbStrTab
1325	\|\| ( pReq->u.In.offStrTab < pReq->u.In.cbImageWithTabs
1326	&& pReq->u.In.offStrTab + pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs
1327	&& pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs),
1328	("SUP_IOCTL_LDR_LOAD: offStrTab=%#lx cbStrTab=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offStrTab,
1329	(long)pReq->u.In.cbStrTab, (long)pReq->u.In.cbImageWithTabs));
1330
1331	if (pReq->u.In.cSymbols)
1332	{
1333	uint32_t i;
1334	PSUPLDRSYM paSyms = (PSUPLDRSYM)&pReq->u.In.abImage[pReq->u.In.offSymbols];
1335	for (i = 0; i < pReq->u.In.cSymbols; i++)
1336	{
1337	REQ_CHECK_EXPR_FMT(paSyms[i].offSymbol < pReq->u.In.cbImageWithTabs,
1338	("SUP_IOCTL_LDR_LOAD: sym #%ld: symb off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offSymbol, (long)pReq->u.In.cbImageWithTabs));
1339	REQ_CHECK_EXPR_FMT(paSyms[i].offName < pReq->u.In.cbStrTab,
1340	("SUP_IOCTL_LDR_LOAD: sym #%ld: name off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1341	REQ_CHECK_EXPR_FMT(RTStrEnd((char const *)&pReq->u.In.abImage[pReq->u.In.offStrTab + paSyms[i].offName],
1342	pReq->u.In.cbStrTab - paSyms[i].offName),
1343	("SUP_IOCTL_LDR_LOAD: sym #%ld: unterminated name! (%#lx / %#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1344	}
1345	}
1346
1347	/* execute */
1348	pReq->Hdr.rc = supdrvIOCtl_LdrLoad(pDevExt, pSession, pReq);
1349	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1350	return 0;
1351	}
1352
1353	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_FREE):
1354	{
1355	/* validate */
1356	PSUPLDRFREE pReq = (PSUPLDRFREE)pReqHdr;
1357	REQ_CHECK_SIZES(SUP_IOCTL_LDR_FREE);
1358
1359	/* execute */
1360	pReq->Hdr.rc = supdrvIOCtl_LdrFree(pDevExt, pSession, pReq);
1361	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1362	return 0;
1363	}
1364
1365	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_GET_SYMBOL):
1366	{
1367	/* validate */
1368	PSUPLDRGETSYMBOL pReq = (PSUPLDRGETSYMBOL)pReqHdr;
1369	REQ_CHECK_SIZES(SUP_IOCTL_LDR_GET_SYMBOL);
1370	REQ_CHECK_EXPR(SUP_IOCTL_LDR_GET_SYMBOL, RTStrEnd(pReq->u.In.szSymbol, sizeof(pReq->u.In.szSymbol)));
1371
1372	/* execute */
1373	pReq->Hdr.rc = supdrvIOCtl_LdrGetSymbol(pDevExt, pSession, pReq);
1374	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1375	return 0;
1376	}
1377
1378	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0(0)):
1379	{
1380	/* validate */
1381	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1382	Log4(("SUP_IOCTL_CALL_VMMR0: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1383	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1384
1385	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_VMMR0_SIZE(0))
1386	{
1387	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(0), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(0));
1388
1389	/* execute */
1390	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1391	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg, pSession);
1392	else
1393	pReq->Hdr.rc = VERR_WRONG_ORDER;
1394	}
1395	else
1396	{
1397	PSUPVMMR0REQHDR pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1398	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR)),
1399	("SUP_IOCTL_CALL_VMMR0: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR))));
1400	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1401	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(pVMMReq->cbReq));
1402
1403	/* execute */
1404	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1405	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1406	else
1407	pReq->Hdr.rc = VERR_WRONG_ORDER;
1408	}
1409
1410	if ( RT_FAILURE(pReq->Hdr.rc)
1411	&& pReq->Hdr.rc != VERR_INTERRUPTED
1412	&& pReq->Hdr.rc != VERR_TIMEOUT)
1413	Log(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1414	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1415	else
1416	Log4(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1417	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1418	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1419	return 0;
1420	}
1421
1422	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0_BIG):
1423	{
1424	/* validate */
1425	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1426	PSUPVMMR0REQHDR pVMMReq;
1427	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1428	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1429
1430	pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1431	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR)),
1432	("SUP_IOCTL_CALL_VMMR0_BIG: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR))));
1433	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0_BIG, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1434	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));
1435
1436	/* execute */
1437	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1438	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1439	else
1440	pReq->Hdr.rc = VERR_WRONG_ORDER;
1441
1442	if ( RT_FAILURE(pReq->Hdr.rc)
1443	&& pReq->Hdr.rc != VERR_INTERRUPTED
1444	&& pReq->Hdr.rc != VERR_TIMEOUT)
1445	Log(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1446	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1447	else
1448	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1449	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1450	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1451	return 0;
1452	}
1453
1454	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_PAGING_MODE):
1455	{
1456	/* validate */
1457	PSUPGETPAGINGMODE pReq = (PSUPGETPAGINGMODE)pReqHdr;
1458	REQ_CHECK_SIZES(SUP_IOCTL_GET_PAGING_MODE);
1459
1460	/* execute */
1461	pReq->Hdr.rc = VINF_SUCCESS;
1462	pReq->u.Out.enmMode = SUPR0GetPagingMode();
1463	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1464	return 0;
1465	}
1466
1467	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_ALLOC):
1468	{
1469	/* validate */
1470	PSUPLOWALLOC pReq = (PSUPLOWALLOC)pReqHdr;
1471	REQ_CHECK_EXPR(SUP_IOCTL_LOW_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_LOW_ALLOC_SIZE_IN);
1472	REQ_CHECK_SIZES_EX(SUP_IOCTL_LOW_ALLOC, SUP_IOCTL_LOW_ALLOC_SIZE_IN, SUP_IOCTL_LOW_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1473
1474	/* execute */
1475	pReq->Hdr.rc = SUPR0LowAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1476	if (RT_FAILURE(pReq->Hdr.rc))
1477	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1478	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1479	return 0;
1480	}
1481
1482	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_FREE):
1483	{
1484	/* validate */
1485	PSUPLOWFREE pReq = (PSUPLOWFREE)pReqHdr;
1486	REQ_CHECK_SIZES(SUP_IOCTL_LOW_FREE);
1487
1488	/* execute */
1489	pReq->Hdr.rc = SUPR0LowFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1490	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1491	return 0;
1492	}
1493
1494	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_MAP):
1495	{
1496	/* validate */
1497	PSUPGIPMAP pReq = (PSUPGIPMAP)pReqHdr;
1498	REQ_CHECK_SIZES(SUP_IOCTL_GIP_MAP);
1499
1500	/* execute */
1501	pReq->Hdr.rc = SUPR0GipMap(pSession, &pReq->u.Out.pGipR3, &pReq->u.Out.HCPhysGip);
1502	if (RT_SUCCESS(pReq->Hdr.rc))
1503	pReq->u.Out.pGipR0 = pDevExt->pGip;
1504	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1505	return 0;
1506	}
1507
1508	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_UNMAP):
1509	{
1510	/* validate */
1511	PSUPGIPUNMAP pReq = (PSUPGIPUNMAP)pReqHdr;
1512	REQ_CHECK_SIZES(SUP_IOCTL_GIP_UNMAP);
1513
1514	/* execute */
1515	pReq->Hdr.rc = SUPR0GipUnmap(pSession);
1516	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1517	return 0;
1518	}
1519
1520	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SET_VM_FOR_FAST):
1521	{
1522	/* validate */
1523	PSUPSETVMFORFAST pReq = (PSUPSETVMFORFAST)pReqHdr;
1524	REQ_CHECK_SIZES(SUP_IOCTL_SET_VM_FOR_FAST);
1525	REQ_CHECK_EXPR_FMT( !pReq->u.In.pVMR0
1526	\|\| ( VALID_PTR(pReq->u.In.pVMR0)
1527	&& !((uintptr_t)pReq->u.In.pVMR0 & (PAGE_SIZE - 1))),
1528	("SUP_IOCTL_SET_VM_FOR_FAST: pVMR0=%p!\n", pReq->u.In.pVMR0));
1529	/* execute */
1530	pSession->pVM = pReq->u.In.pVMR0;
1531	pReq->Hdr.rc = VINF_SUCCESS;
1532	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1533	return 0;
1534	}
1535
1536	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_ALLOC_EX):
1537	{
1538	/* validate */
1539	PSUPPAGEALLOCEX pReq = (PSUPPAGEALLOCEX)pReqHdr;
1540	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC_EX, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_EX_SIZE_IN);
1541	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));
1542	REQ_CHECK_EXPR_FMT(pReq->u.In.fKernelMapping \|\| pReq->u.In.fUserMapping,
1543	("SUP_IOCTL_PAGE_ALLOC_EX: No mapping requested!\n"));
1544	REQ_CHECK_EXPR_FMT(pReq->u.In.fUserMapping,
1545	("SUP_IOCTL_PAGE_ALLOC_EX: Must have user mapping!\n"));
1546	REQ_CHECK_EXPR_FMT(!pReq->u.In.fReserved0 && !pReq->u.In.fReserved1,
1547	("SUP_IOCTL_PAGE_ALLOC_EX: fReserved0=%d fReserved1=%d\n", pReq->u.In.fReserved0, pReq->u.In.fReserved1));
1548
1549	/* execute */
1550	pReq->Hdr.rc = SUPR0PageAllocEx(pSession, pReq->u.In.cPages, 0 /* fFlags */,
1551	pReq->u.In.fUserMapping ? &pReq->u.Out.pvR3 : NULL,
1552	pReq->u.In.fKernelMapping ? &pReq->u.Out.pvR0 : NULL,
1553	&pReq->u.Out.aPages[0]);
1554	if (RT_FAILURE(pReq->Hdr.rc))
1555	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1556	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1557	return 0;
1558	}
1559
1560	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_MAP_KERNEL):
1561	{
1562	/* validate */
1563	PSUPPAGEMAPKERNEL pReq = (PSUPPAGEMAPKERNEL)pReqHdr;
1564	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_MAP_KERNEL);
1565	REQ_CHECK_EXPR_FMT(!pReq->u.In.fFlags, ("SUP_IOCTL_PAGE_MAP_KERNEL: fFlags=%#x! MBZ\n", pReq->u.In.fFlags));
1566	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_MAP_KERNEL: offSub=%#x\n", pReq->u.In.offSub));
1567	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1568	("SUP_IOCTL_PAGE_MAP_KERNEL: cbSub=%#x\n", pReq->u.In.cbSub));
1569
1570	/* execute */
1571	pReq->Hdr.rc = SUPR0PageMapKernel(pSession, pReq->u.In.pvR3, pReq->u.In.offSub, pReq->u.In.cbSub,
1572	pReq->u.In.fFlags, &pReq->u.Out.pvR0);
1573	if (RT_FAILURE(pReq->Hdr.rc))
1574	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1575	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1576	return 0;
1577	}
1578
1579	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_PROTECT):
1580	{
1581	/* validate */
1582	PSUPPAGEPROTECT pReq = (PSUPPAGEPROTECT)pReqHdr;
1583	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_PROTECT);
1584	REQ_CHECK_EXPR_FMT(!(pReq->u.In.fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)),
1585	("SUP_IOCTL_PAGE_PROTECT: fProt=%#x!\n", pReq->u.In.fProt));
1586	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_PROTECT: offSub=%#x\n", pReq->u.In.offSub));
1587	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1588	("SUP_IOCTL_PAGE_PROTECT: cbSub=%#x\n", pReq->u.In.cbSub));
1589
1590	/* execute */
1591	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);
1592	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1593	return 0;
1594	}
1595
1596	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_FREE):
1597	{
1598	/* validate */
1599	PSUPPAGEFREE pReq = (PSUPPAGEFREE)pReqHdr;
1600	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_FREE);
1601
1602	/* execute */
1603	pReq->Hdr.rc = SUPR0PageFree(pSession, pReq->u.In.pvR3);
1604	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1605	return 0;
1606	}
1607
1608	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_SERVICE(0)):
1609	{
1610	/* validate */
1611	PSUPCALLSERVICE pReq = (PSUPCALLSERVICE)pReqHdr;
1612	Log4(("SUP_IOCTL_CALL_SERVICE: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1613	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1614
1615	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
1616	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(0), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(0));
1617	else
1618	{
1619	PSUPR0SERVICEREQHDR pSrvReq = (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0];
1620	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR)),
1621	("SUP_IOCTL_CALL_SERVICE: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR))));
1622	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, pSrvReq->u32Magic == SUPR0SERVICEREQHDR_MAGIC);
1623	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(pSrvReq->cbReq), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(pSrvReq->cbReq));
1624	}
1625	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1626
1627	/* execute */
1628	pReq->Hdr.rc = supdrvIOCtl_CallServiceModule(pDevExt, pSession, pReq);
1629	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1630	return 0;
1631	}
1632
1633	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOGGER_SETTINGS(0)):
1634	{
1635	/* validate */
1636	PSUPLOGGERSETTINGS pReq = (PSUPLOGGERSETTINGS)pReqHdr;
1637	size_t cbStrTab;
1638	REQ_CHECK_SIZE_OUT(SUP_IOCTL_LOGGER_SETTINGS, SUP_IOCTL_LOGGER_SETTINGS_SIZE_OUT);
1639	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->Hdr.cbIn >= SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(1));
1640	cbStrTab = pReq->Hdr.cbIn - SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(0);
1641	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offGroups < cbStrTab);
1642	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offFlags < cbStrTab);
1643	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offDestination < cbStrTab);
1644	REQ_CHECK_EXPR_FMT(pReq->u.In.szStrings[cbStrTab - 1] == '\0',
1645	("SUP_IOCTL_LOGGER_SETTINGS: cbIn=%#x cbStrTab=%#zx LastChar=%d\n",
1646	pReq->Hdr.cbIn, cbStrTab, pReq->u.In.szStrings[cbStrTab - 1]));
1647	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhich <= SUPLOGGERSETTINGS_WHICH_RELEASE);
1648	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhat <= SUPLOGGERSETTINGS_WHAT_DESTROY);
1649
1650	/* execute */
1651	pReq->Hdr.rc = supdrvIOCtl_LoggerSettings(pDevExt, pSession, pReq);
1652	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1653	return 0;
1654	}
1655
1656	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP2):
1657	{
1658	/* validate */
1659	PSUPSEMOP2 pReq = (PSUPSEMOP2)pReqHdr;
1660	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP2, SUP_IOCTL_SEM_OP2_SIZE_IN, SUP_IOCTL_SEM_OP2_SIZE_OUT);
1661	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP2, pReq->u.In.uReserved == 0);
1662
1663	/* execute */
1664	switch (pReq->u.In.uType)
1665	{
1666	case SUP_SEM_TYPE_EVENT:
1667	{
1668	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1669	switch (pReq->u.In.uOp)
1670	{
1671	case SUPSEMOP2_WAIT_MS_REL:
1672	pReq->Hdr.rc = SUPSemEventWaitNoResume(pSession, hEvent, pReq->u.In.uArg.cRelMsTimeout);
1673	break;
1674	case SUPSEMOP2_WAIT_NS_ABS:
1675	pReq->Hdr.rc = SUPSemEventWaitNsAbsIntr(pSession, hEvent, pReq->u.In.uArg.uAbsNsTimeout);
1676	break;
1677	case SUPSEMOP2_WAIT_NS_REL:
1678	pReq->Hdr.rc = SUPSemEventWaitNsRelIntr(pSession, hEvent, pReq->u.In.uArg.cRelNsTimeout);
1679	break;
1680	case SUPSEMOP2_SIGNAL:
1681	pReq->Hdr.rc = SUPSemEventSignal(pSession, hEvent);
1682	break;
1683	case SUPSEMOP2_CLOSE:
1684	pReq->Hdr.rc = SUPSemEventClose(pSession, hEvent);
1685	break;
1686	case SUPSEMOP2_RESET:
1687	default:
1688	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1689	break;
1690	}
1691	break;
1692	}
1693
1694	case SUP_SEM_TYPE_EVENT_MULTI:
1695	{
1696	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1697	switch (pReq->u.In.uOp)
1698	{
1699	case SUPSEMOP2_WAIT_MS_REL:
1700	pReq->Hdr.rc = SUPSemEventMultiWaitNoResume(pSession, hEventMulti, pReq->u.In.uArg.cRelMsTimeout);
1701	break;
1702	case SUPSEMOP2_WAIT_NS_ABS:
1703	pReq->Hdr.rc = SUPSemEventMultiWaitNsAbsIntr(pSession, hEventMulti, pReq->u.In.uArg.uAbsNsTimeout);
1704	break;
1705	case SUPSEMOP2_WAIT_NS_REL:
1706	pReq->Hdr.rc = SUPSemEventMultiWaitNsRelIntr(pSession, hEventMulti, pReq->u.In.uArg.cRelNsTimeout);
1707	break;
1708	case SUPSEMOP2_SIGNAL:
1709	pReq->Hdr.rc = SUPSemEventMultiSignal(pSession, hEventMulti);
1710	break;
1711	case SUPSEMOP2_CLOSE:
1712	pReq->Hdr.rc = SUPSemEventMultiClose(pSession, hEventMulti);
1713	break;
1714	case SUPSEMOP2_RESET:
1715	pReq->Hdr.rc = SUPSemEventMultiReset(pSession, hEventMulti);
1716	break;
1717	default:
1718	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1719	break;
1720	}
1721	break;
1722	}
1723
1724	default:
1725	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1726	break;
1727	}
1728	return 0;
1729	}
1730
1731	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP3):
1732	{
1733	/* validate */
1734	PSUPSEMOP3 pReq = (PSUPSEMOP3)pReqHdr;
1735	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP3, SUP_IOCTL_SEM_OP3_SIZE_IN, SUP_IOCTL_SEM_OP3_SIZE_OUT);
1736	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, pReq->u.In.u32Reserved == 0 && pReq->u.In.u64Reserved == 0);
1737
1738	/* execute */
1739	switch (pReq->u.In.uType)
1740	{
1741	case SUP_SEM_TYPE_EVENT:
1742	{
1743	SUPSEMEVENT hEvent = (SUPSEMEVENT)(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, hEvent == NIL_SUPSEMEVENT);
1748	pReq->Hdr.rc = SUPSemEventCreate(pSession, &hEvent);
1749	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEvent;
1750	break;
1751	case SUPSEMOP3_GET_RESOLUTION:
1752	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1753	pReq->Hdr.rc = VINF_SUCCESS;
1754	pReq->Hdr.cbOut = sizeof(*pReq);
1755	pReq->u.Out.cNsResolution = SUPSemEventGetResolution(pSession);
1756	break;
1757	default:
1758	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1759	break;
1760	}
1761	break;
1762	}
1763
1764	case SUP_SEM_TYPE_EVENT_MULTI:
1765	{
1766	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1767	switch (pReq->u.In.uOp)
1768	{
1769	case SUPSEMOP3_CREATE:
1770	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1771	pReq->Hdr.rc = SUPSemEventMultiCreate(pSession, &hEventMulti);
1772	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEventMulti;
1773	break;
1774	case SUPSEMOP3_GET_RESOLUTION:
1775	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1776	pReq->Hdr.rc = VINF_SUCCESS;
1777	pReq->u.Out.cNsResolution = SUPSemEventMultiGetResolution(pSession);
1778	break;
1779	default:
1780	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1781	break;
1782	}
1783	break;
1784	}
1785
1786	default:
1787	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1788	break;
1789	}
1790	return 0;
1791	}
1792
1793	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
1794	{
1795	/* validate */
1796	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
1797	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
1798	REQ_CHECK_EXPR(SUP_IOCTL_VT_CAPS, pReq->Hdr.cbIn <= SUP_IOCTL_VT_CAPS_SIZE_IN);
1799
1800	/* execute */
1801	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.Caps);
1802	if (RT_FAILURE(pReq->Hdr.rc))
1803	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1804	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1805	return 0;
1806	}
1807
1808	default:
1809	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1810	break;
1811	}
1812	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_GENERAL_FAILURE, VERR_GENERAL_FAILURE);
1813	return VERR_GENERAL_FAILURE;
1814	}
1815
1816
1817	/**
1818	* Inter-Driver Communication (IDC) worker.
1819	*
1820	* @returns VBox status code.
1821	* @retval VINF_SUCCESS on success.
1822	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1823	* @retval VERR_NOT_SUPPORTED if the request isn't supported.
1824	*
1825	* @param uReq The request (function) code.
1826	* @param pDevExt Device extention.
1827	* @param pSession Session data.
1828	* @param pReqHdr The request header.
1829	*/
1830	int VBOXCALL supdrvIDC(uintptr_t uReq, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQHDR pReqHdr)
1831	{
1832	/*
1833	* The OS specific code has already validated the pSession
1834	* pointer, and the request size being greater or equal to
1835	* size of the header.
1836	*
1837	* So, just check that pSession is a kernel context session.
1838	*/
1839	if (RT_UNLIKELY( pSession
1840	&& pSession->R0Process != NIL_RTR0PROCESS))
1841	return VERR_INVALID_PARAMETER;
1842
1843	/*
1844	* Validation macro.
1845	*/
1846	#define REQ_CHECK_IDC_SIZE(Name, cbExpect) \
1847	do { \
1848	if (RT_UNLIKELY(pReqHdr->cb != (cbExpect))) \
1849	{ \
1850	OSDBGPRINT(( #Name ": Invalid input/output sizes. cb=%ld expected %ld.\n", \
1851	(long)pReqHdr->cb, (long)(cbExpect))); \
1852	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1853	} \
1854	} while (0)
1855
1856	switch (uReq)
1857	{
1858	case SUPDRV_IDC_REQ_CONNECT:
1859	{
1860	PSUPDRVIDCREQCONNECT pReq = (PSUPDRVIDCREQCONNECT)pReqHdr;
1861	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_CONNECT, sizeof(*pReq));
1862
1863	/*
1864	* Validate the cookie and other input.
1865	*/
1866	if (pReq->Hdr.pSession != NULL)
1867	{
1868	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Hdr.pSession=%p expected NULL!\n", pReq->Hdr.pSession));
1869	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1870	}
1871	if (pReq->u.In.u32MagicCookie != SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE)
1872	{
1873	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: u32MagicCookie=%#x expected %#x!\n",
1874	(unsigned)pReq->u.In.u32MagicCookie, (unsigned)SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE));
1875	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1876	}
1877	if ( pReq->u.In.uMinVersion > pReq->u.In.uReqVersion
1878	\|\| (pReq->u.In.uMinVersion & UINT32_C(0xffff0000)) != (pReq->u.In.uReqVersion & UINT32_C(0xffff0000)))
1879	{
1880	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: uMinVersion=%#x uMaxVersion=%#x doesn't match!\n",
1881	pReq->u.In.uMinVersion, pReq->u.In.uReqVersion));
1882	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1883	}
1884	if (pSession != NULL)
1885	{
1886	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: pSession=%p expected NULL!\n", pSession));
1887	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1888	}
1889
1890	/*
1891	* Match the version.
1892	* The current logic is very simple, match the major interface version.
1893	*/
1894	if ( pReq->u.In.uMinVersion > SUPDRV_IDC_VERSION
1895	\|\| (pReq->u.In.uMinVersion & 0xffff0000) != (SUPDRV_IDC_VERSION & 0xffff0000))
1896	{
1897	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1898	pReq->u.In.uReqVersion, pReq->u.In.uMinVersion, (unsigned)SUPDRV_IDC_VERSION));
1899	pReq->u.Out.pSession = NULL;
1900	pReq->u.Out.uSessionVersion = 0xffffffff;
1901	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1902	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1903	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1904	return VINF_SUCCESS;
1905	}
1906
1907	pReq->u.Out.pSession = NULL;
1908	pReq->u.Out.uSessionVersion = SUPDRV_IDC_VERSION;
1909	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1910	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1911
1912	pReq->Hdr.rc = supdrvCreateSession(pDevExt, false /* fUser */, &pSession);
1913	if (RT_FAILURE(pReq->Hdr.rc))
1914	{
1915	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: failed to create session, rc=%d\n", pReq->Hdr.rc));
1916	return VINF_SUCCESS;
1917	}
1918
1919	pReq->u.Out.pSession = pSession;
1920	pReq->Hdr.pSession = pSession;
1921
1922	return VINF_SUCCESS;
1923	}
1924
1925	case SUPDRV_IDC_REQ_DISCONNECT:
1926	{
1927	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_DISCONNECT, sizeof(*pReqHdr));
1928
1929	supdrvCloseSession(pDevExt, pSession);
1930	return pReqHdr->rc = VINF_SUCCESS;
1931	}
1932
1933	case SUPDRV_IDC_REQ_GET_SYMBOL:
1934	{
1935	PSUPDRVIDCREQGETSYM pReq = (PSUPDRVIDCREQGETSYM)pReqHdr;
1936	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_GET_SYMBOL, sizeof(*pReq));
1937
1938	pReq->Hdr.rc = supdrvIDC_LdrGetSymbol(pDevExt, pSession, pReq);
1939	return VINF_SUCCESS;
1940	}
1941
1942	case SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY:
1943	{
1944	PSUPDRVIDCREQCOMPREGFACTORY pReq = (PSUPDRVIDCREQCOMPREGFACTORY)pReqHdr;
1945	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY, sizeof(*pReq));
1946
1947	pReq->Hdr.rc = SUPR0ComponentRegisterFactory(pSession, pReq->u.In.pFactory);
1948	return VINF_SUCCESS;
1949	}
1950
1951	case SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY:
1952	{
1953	PSUPDRVIDCREQCOMPDEREGFACTORY pReq = (PSUPDRVIDCREQCOMPDEREGFACTORY)pReqHdr;
1954	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY, sizeof(*pReq));
1955
1956	pReq->Hdr.rc = SUPR0ComponentDeregisterFactory(pSession, pReq->u.In.pFactory);
1957	return VINF_SUCCESS;
1958	}
1959
1960	default:
1961	Log(("Unknown IDC %#lx\n", (long)uReq));
1962	break;
1963	}
1964
1965	#undef REQ_CHECK_IDC_SIZE
1966	return VERR_NOT_SUPPORTED;
1967	}
1968
1969
1970	/**
1971	* Register a object for reference counting.
1972	* The object is registered with one reference in the specified session.
1973	*
1974	* @returns Unique identifier on success (pointer).
1975	* All future reference must use this identifier.
1976	* @returns NULL on failure.
1977	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
1978	* @param pvUser1 The first user argument.
1979	* @param pvUser2 The second user argument.
1980	*/
1981	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
1982	{
1983	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1984	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1985	PSUPDRVOBJ pObj;
1986	PSUPDRVUSAGE pUsage;
1987
1988	/*
1989	* Validate the input.
1990	*/
1991	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
1992	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
1993	AssertPtrReturn(pfnDestructor, NULL);
1994
1995	/*
1996	* Allocate and initialize the object.
1997	*/
1998	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
1999	if (!pObj)
2000	return NULL;
2001	pObj->u32Magic = SUPDRVOBJ_MAGIC;
2002	pObj->enmType = enmType;
2003	pObj->pNext = NULL;
2004	pObj->cUsage = 1;
2005	pObj->pfnDestructor = pfnDestructor;
2006	pObj->pvUser1 = pvUser1;
2007	pObj->pvUser2 = pvUser2;
2008	pObj->CreatorUid = pSession->Uid;
2009	pObj->CreatorGid = pSession->Gid;
2010	pObj->CreatorProcess= pSession->Process;
2011	supdrvOSObjInitCreator(pObj, pSession);
2012
2013	/*
2014	* Allocate the usage record.
2015	* (We keep freed usage records around to simplify SUPR0ObjAddRefEx().)
2016	*/
2017	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2018
2019	pUsage = pDevExt->pUsageFree;
2020	if (pUsage)
2021	pDevExt->pUsageFree = pUsage->pNext;
2022	else
2023	{
2024	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2025	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
2026	if (!pUsage)
2027	{
2028	RTMemFree(pObj);
2029	return NULL;
2030	}
2031	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2032	}
2033
2034	/*
2035	* Insert the object and create the session usage record.
2036	*/
2037	/* The object. */
2038	pObj->pNext = pDevExt->pObjs;
2039	pDevExt->pObjs = pObj;
2040
2041	/* The session record. */
2042	pUsage->cUsage = 1;
2043	pUsage->pObj = pObj;
2044	pUsage->pNext = pSession->pUsage;
2045	/* Log2(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext)); */
2046	pSession->pUsage = pUsage;
2047
2048	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2049
2050	Log(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
2051	return pObj;
2052	}
2053
2054
2055	/**
2056	* Increment the reference counter for the object associating the reference
2057	* with the specified session.
2058	*
2059	* @returns IPRT status code.
2060	* @param pvObj The identifier returned by SUPR0ObjRegister().
2061	* @param pSession The session which is referencing the object.
2062	*
2063	* @remarks The caller should not own any spinlocks and must carefully protect
2064	* itself against potential race with the destructor so freed memory
2065	* isn't accessed here.
2066	*/
2067	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
2068	{
2069	return SUPR0ObjAddRefEx(pvObj, pSession, false /* fNoBlocking */);
2070	}
2071
2072
2073	/**
2074	* Increment the reference counter for the object associating the reference
2075	* with the specified session.
2076	*
2077	* @returns IPRT status code.
2078	* @retval VERR_TRY_AGAIN if fNoBlocking was set and a new usage record
2079	* couldn't be allocated. (If you see this you're not doing the right
2080	* thing and it won't ever work reliably.)
2081	*
2082	* @param pvObj The identifier returned by SUPR0ObjRegister().
2083	* @param pSession The session which is referencing the object.
2084	* @param fNoBlocking Set if it's not OK to block. Never try to make the
2085	* first reference to an object in a session with this
2086	* argument set.
2087	*
2088	* @remarks The caller should not own any spinlocks and must carefully protect
2089	* itself against potential race with the destructor so freed memory
2090	* isn't accessed here.
2091	*/
2092	SUPR0DECL(int) SUPR0ObjAddRefEx(void *pvObj, PSUPDRVSESSION pSession, bool fNoBlocking)
2093	{
2094	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2095	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2096	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2097	int rc = VINF_SUCCESS;
2098	PSUPDRVUSAGE pUsagePre;
2099	PSUPDRVUSAGE pUsage;
2100
2101	/*
2102	* Validate the input.
2103	* Be ready for the destruction race (someone might be stuck in the
2104	* destructor waiting a lock we own).
2105	*/
2106	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2107	AssertPtrReturn(pObj, VERR_INVALID_POINTER);
2108	AssertMsgReturn(pObj->u32Magic == SUPDRVOBJ_MAGIC \|\| pObj->u32Magic == SUPDRVOBJ_MAGIC_DEAD,
2109	("Invalid pvObj=%p magic=%#x (expected %#x or %#x)\n", pvObj, pObj->u32Magic, SUPDRVOBJ_MAGIC, SUPDRVOBJ_MAGIC_DEAD),
2110	VERR_INVALID_PARAMETER);
2111
2112	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2113
2114	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2115	{
2116	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2117
2118	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2119	return VERR_WRONG_ORDER;
2120	}
2121
2122	/*
2123	* Preallocate the usage record if we can.
2124	*/
2125	pUsagePre = pDevExt->pUsageFree;
2126	if (pUsagePre)
2127	pDevExt->pUsageFree = pUsagePre->pNext;
2128	else if (!fNoBlocking)
2129	{
2130	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2131	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
2132	if (!pUsagePre)
2133	return VERR_NO_MEMORY;
2134
2135	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2136	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2137	{
2138	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2139
2140	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2141	return VERR_WRONG_ORDER;
2142	}
2143	}
2144
2145	/*
2146	* Reference the object.
2147	*/
2148	pObj->cUsage++;
2149
2150	/*
2151	* Look for the session record.
2152	*/
2153	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
2154	{
2155	/Log(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2156	if (pUsage->pObj == pObj)
2157	break;
2158	}
2159	if (pUsage)
2160	pUsage->cUsage++;
2161	else if (pUsagePre)
2162	{
2163	/* create a new session record. */
2164	pUsagePre->cUsage = 1;
2165	pUsagePre->pObj = pObj;
2166	pUsagePre->pNext = pSession->pUsage;
2167	pSession->pUsage = pUsagePre;
2168	/Log(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));/
2169
2170	pUsagePre = NULL;
2171	}
2172	else
2173	{
2174	pObj->cUsage--;
2175	rc = VERR_TRY_AGAIN;
2176	}
2177
2178	/*
2179	* Put any unused usage record into the free list..
2180	*/
2181	if (pUsagePre)
2182	{
2183	pUsagePre->pNext = pDevExt->pUsageFree;
2184	pDevExt->pUsageFree = pUsagePre;
2185	}
2186
2187	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2188
2189	return rc;
2190	}
2191
2192
2193	/**
2194	* Decrement / destroy a reference counter record for an object.
2195	*
2196	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
2197	*
2198	* @returns IPRT status code.
2199	* @retval VINF_SUCCESS if not destroyed.
2200	* @retval VINF_OBJECT_DESTROYED if it's destroyed by this release call.
2201	* @retval VERR_INVALID_PARAMETER if the object isn't valid. Will assert in
2202	* string builds.
2203	*
2204	* @param pvObj The identifier returned by SUPR0ObjRegister().
2205	* @param pSession The session which is referencing the object.
2206	*/
2207	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
2208	{
2209	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2210	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2211	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2212	int rc = VERR_INVALID_PARAMETER;
2213	PSUPDRVUSAGE pUsage;
2214	PSUPDRVUSAGE pUsagePrev;
2215
2216	/*
2217	* Validate the input.
2218	*/
2219	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2220	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2221	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2222	VERR_INVALID_PARAMETER);
2223
2224	/*
2225	* Acquire the spinlock and look for the usage record.
2226	*/
2227	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2228
2229	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
2230	pUsage;
2231	pUsagePrev = pUsage, pUsage = pUsage->pNext)
2232	{
2233	/Log2(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2234	if (pUsage->pObj == pObj)
2235	{
2236	rc = VINF_SUCCESS;
2237	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
2238	if (pUsage->cUsage > 1)
2239	{
2240	pObj->cUsage--;
2241	pUsage->cUsage--;
2242	}
2243	else
2244	{
2245	/*
2246	* Free the session record.
2247	*/
2248	if (pUsagePrev)
2249	pUsagePrev->pNext = pUsage->pNext;
2250	else
2251	pSession->pUsage = pUsage->pNext;
2252	pUsage->pNext = pDevExt->pUsageFree;
2253	pDevExt->pUsageFree = pUsage;
2254
2255	/* What about the object? */
2256	if (pObj->cUsage > 1)
2257	pObj->cUsage--;
2258	else
2259	{
2260	/*
2261	* Object is to be destroyed, unlink it.
2262	*/
2263	pObj->u32Magic = SUPDRVOBJ_MAGIC_DEAD;
2264	rc = VINF_OBJECT_DESTROYED;
2265	if (pDevExt->pObjs == pObj)
2266	pDevExt->pObjs = pObj->pNext;
2267	else
2268	{
2269	PSUPDRVOBJ pObjPrev;
2270	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
2271	if (pObjPrev->pNext == pObj)
2272	{
2273	pObjPrev->pNext = pObj->pNext;
2274	break;
2275	}
2276	Assert(pObjPrev);
2277	}
2278	}
2279	}
2280	break;
2281	}
2282	}
2283
2284	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2285
2286	/*
2287	* Call the destructor and free the object if required.
2288	*/
2289	if (rc == VINF_OBJECT_DESTROYED)
2290	{
2291	Log(("SUPR0ObjRelease: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
2292	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
2293	if (pObj->pfnDestructor)
2294	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
2295	RTMemFree(pObj);
2296	}
2297
2298	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
2299	return rc;
2300	}
2301
2302
2303	/**
2304	* Verifies that the current process can access the specified object.
2305	*
2306	* @returns The following IPRT status code:
2307	* @retval VINF_SUCCESS if access was granted.
2308	* @retval VERR_PERMISSION_DENIED if denied access.
2309	* @retval VERR_INVALID_PARAMETER if invalid parameter.
2310	*
2311	* @param pvObj The identifier returned by SUPR0ObjRegister().
2312	* @param pSession The session which wishes to access the object.
2313	* @param pszObjName Object string name. This is optional and depends on the object type.
2314	*
2315	* @remark The caller is responsible for making sure the object isn't removed while
2316	* we're inside this function. If uncertain about this, just call AddRef before calling us.
2317	*/
2318	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
2319	{
2320	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2321	int rc;
2322
2323	/*
2324	* Validate the input.
2325	*/
2326	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2327	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2328	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2329	VERR_INVALID_PARAMETER);
2330
2331	/*
2332	* Check access. (returns true if a decision has been made.)
2333	*/
2334	rc = VERR_INTERNAL_ERROR;
2335	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
2336	return rc;
2337
2338	/*
2339	* Default policy is to allow the user to access his own
2340	* stuff but nothing else.
2341	*/
2342	if (pObj->CreatorUid == pSession->Uid)
2343	return VINF_SUCCESS;
2344	return VERR_PERMISSION_DENIED;
2345	}
2346
2347
2348	/**
2349	* Lock pages.
2350	*
2351	* @returns IPRT status code.
2352	* @param pSession Session to which the locked memory should be associated.
2353	* @param pvR3 Start of the memory range to lock.
2354	* This must be page aligned.
2355	* @param cPages Number of pages to lock.
2356	* @param paPages Where to put the physical addresses of locked memory.
2357	*/
2358	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
2359	{
2360	int rc;
2361	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2362	const size_t cb = (size_t)cPages << PAGE_SHIFT;
2363	LogFlow(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
2364
2365	/*
2366	* Verify input.
2367	*/
2368	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2369	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
2370	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
2371	\|\| !pvR3)
2372	{
2373	Log(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
2374	return VERR_INVALID_PARAMETER;
2375	}
2376
2377	/*
2378	* Let IPRT do the job.
2379	*/
2380	Mem.eType = MEMREF_TYPE_LOCKED;
2381	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTMEM_PROT_READ \| RTMEM_PROT_WRITE, RTR0ProcHandleSelf());
2382	if (RT_SUCCESS(rc))
2383	{
2384	uint32_t iPage = cPages;
2385	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
2386	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
2387
2388	while (iPage-- > 0)
2389	{
2390	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2391	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
2392	{
2393	AssertMsgFailed(("iPage=%d\n", iPage));
2394	rc = VERR_INTERNAL_ERROR;
2395	break;
2396	}
2397	}
2398	if (RT_SUCCESS(rc))
2399	rc = supdrvMemAdd(&Mem, pSession);
2400	if (RT_FAILURE(rc))
2401	{
2402	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
2403	AssertRC(rc2);
2404	}
2405	}
2406
2407	return rc;
2408	}
2409
2410
2411	/**
2412	* Unlocks the memory pointed to by pv.
2413	*
2414	* @returns IPRT status code.
2415	* @param pSession Session to which the memory was locked.
2416	* @param pvR3 Memory to unlock.
2417	*/
2418	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2419	{
2420	LogFlow(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2421	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2422	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
2423	}
2424
2425
2426	/**
2427	* Allocates a chunk of page aligned memory with contiguous and fixed physical
2428	* backing.
2429	*
2430	* @returns IPRT status code.
2431	* @param pSession Session data.
2432	* @param cPages Number of pages to allocate.
2433	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
2434	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
2435	* @param pHCPhys Where to put the physical address of allocated memory.
2436	*/
2437	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
2438	{
2439	int rc;
2440	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2441	LogFlow(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
2442
2443	/*
2444	* Validate input.
2445	*/
2446	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2447	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
2448	{
2449	Log(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
2450	pSession, ppvR0, ppvR3, pHCPhys));
2451	return VERR_INVALID_PARAMETER;
2452
2453	}
2454	if (cPages < 1 \|\| cPages >= 256)
2455	{
2456	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2457	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2458	}
2459
2460	/*
2461	* Let IPRT do the job.
2462	*/
2463	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
2464	if (RT_SUCCESS(rc))
2465	{
2466	int rc2;
2467	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2468	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2469	if (RT_SUCCESS(rc))
2470	{
2471	Mem.eType = MEMREF_TYPE_CONT;
2472	rc = supdrvMemAdd(&Mem, pSession);
2473	if (!rc)
2474	{
2475	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2476	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2477	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
2478	return 0;
2479	}
2480
2481	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2482	AssertRC(rc2);
2483	}
2484	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2485	AssertRC(rc2);
2486	}
2487
2488	return rc;
2489	}
2490
2491
2492	/**
2493	* Frees memory allocated using SUPR0ContAlloc().
2494	*
2495	* @returns IPRT status code.
2496	* @param pSession The session to which the memory was allocated.
2497	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2498	*/
2499	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2500	{
2501	LogFlow(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2502	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2503	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
2504	}
2505
2506
2507	/**
2508	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
2509	*
2510	* The memory isn't zeroed.
2511	*
2512	* @returns IPRT status code.
2513	* @param pSession Session data.
2514	* @param cPages Number of pages to allocate.
2515	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
2516	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
2517	* @param paPages Where to put the physical addresses of allocated memory.
2518	*/
2519	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
2520	{
2521	unsigned iPage;
2522	int rc;
2523	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2524	LogFlow(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
2525
2526	/*
2527	* Validate input.
2528	*/
2529	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2530	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
2531	{
2532	Log(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
2533	pSession, ppvR3, ppvR0, paPages));
2534	return VERR_INVALID_PARAMETER;
2535
2536	}
2537	if (cPages < 1 \|\| cPages >= 256)
2538	{
2539	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2540	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2541	}
2542
2543	/*
2544	* Let IPRT do the work.
2545	*/
2546	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
2547	if (RT_SUCCESS(rc))
2548	{
2549	int rc2;
2550	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2551	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2552	if (RT_SUCCESS(rc))
2553	{
2554	Mem.eType = MEMREF_TYPE_LOW;
2555	rc = supdrvMemAdd(&Mem, pSession);
2556	if (!rc)
2557	{
2558	for (iPage = 0; iPage < cPages; iPage++)
2559	{
2560	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2561	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%RHp\n", paPages[iPage]));
2562	}
2563	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2564	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2565	return 0;
2566	}
2567
2568	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2569	AssertRC(rc2);
2570	}
2571
2572	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2573	AssertRC(rc2);
2574	}
2575
2576	return rc;
2577	}
2578
2579
2580	/**
2581	* Frees memory allocated using SUPR0LowAlloc().
2582	*
2583	* @returns IPRT status code.
2584	* @param pSession The session to which the memory was allocated.
2585	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2586	*/
2587	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2588	{
2589	LogFlow(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2590	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2591	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
2592	}
2593
2594
2595
2596	/**
2597	* Allocates a chunk of memory with both R0 and R3 mappings.
2598	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
2599	*
2600	* @returns IPRT status code.
2601	* @param pSession The session to associated the allocation with.
2602	* @param cb Number of bytes to allocate.
2603	* @param ppvR0 Where to store the address of the Ring-0 mapping.
2604	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2605	*/
2606	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
2607	{
2608	int rc;
2609	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2610	LogFlow(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
2611
2612	/*
2613	* Validate input.
2614	*/
2615	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2616	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
2617	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
2618	if (cb < 1 \|\| cb >= _4M)
2619	{
2620	Log(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
2621	return VERR_INVALID_PARAMETER;
2622	}
2623
2624	/*
2625	* Let IPRT do the work.
2626	*/
2627	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
2628	if (RT_SUCCESS(rc))
2629	{
2630	int rc2;
2631	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2632	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2633	if (RT_SUCCESS(rc))
2634	{
2635	Mem.eType = MEMREF_TYPE_MEM;
2636	rc = supdrvMemAdd(&Mem, pSession);
2637	if (!rc)
2638	{
2639	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2640	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2641	return VINF_SUCCESS;
2642	}
2643
2644	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2645	AssertRC(rc2);
2646	}
2647
2648	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2649	AssertRC(rc2);
2650	}
2651
2652	return rc;
2653	}
2654
2655
2656	/**
2657	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
2658	*
2659	* @returns IPRT status code.
2660	* @param pSession The session to which the memory was allocated.
2661	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2662	* @param paPages Where to store the physical addresses.
2663	*/
2664	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
2665	{
2666	PSUPDRVBUNDLE pBundle;
2667	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2668	LogFlow(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
2669
2670	/*
2671	* Validate input.
2672	*/
2673	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2674	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
2675	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
2676
2677	/*
2678	* Search for the address.
2679	*/
2680	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2681	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2682	{
2683	if (pBundle->cUsed > 0)
2684	{
2685	unsigned i;
2686	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2687	{
2688	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
2689	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2690	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2691	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2692	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
2693	)
2694	)
2695	{
2696	const size_t cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2697	size_t iPage;
2698	for (iPage = 0; iPage < cPages; iPage++)
2699	{
2700	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2701	paPages[iPage].uReserved = 0;
2702	}
2703	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2704	return VINF_SUCCESS;
2705	}
2706	}
2707	}
2708	}
2709	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2710	Log(("Failed to find %p!!!\n", (void *)uPtr));
2711	return VERR_INVALID_PARAMETER;
2712	}
2713
2714
2715	/**
2716	* Free memory allocated by SUPR0MemAlloc().
2717	*
2718	* @returns IPRT status code.
2719	* @param pSession The session owning the allocation.
2720	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2721	*/
2722	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2723	{
2724	LogFlow(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2725	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2726	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
2727	}
2728
2729
2730	/**
2731	* Allocates a chunk of memory with a kernel or/and a user mode mapping.
2732	*
2733	* The memory is fixed and it's possible to query the physical addresses using
2734	* SUPR0MemGetPhys().
2735	*
2736	* @returns IPRT status code.
2737	* @param pSession The session to associated the allocation with.
2738	* @param cPages The number of pages to allocate.
2739	* @param fFlags Flags, reserved for the future. Must be zero.
2740	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2741	* NULL if no ring-3 mapping.
2742	* @param ppvR3 Where to store the address of the Ring-0 mapping.
2743	* NULL if no ring-0 mapping.
2744	* @param paPages Where to store the addresses of the pages. Optional.
2745	*/
2746	SUPR0DECL(int) SUPR0PageAllocEx(PSUPDRVSESSION pSession, uint32_t cPages, uint32_t fFlags, PRTR3PTR ppvR3, PRTR0PTR ppvR0, PRTHCPHYS paPages)
2747	{
2748	int rc;
2749	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2750	LogFlow(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
2751
2752	/*
2753	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2754	*/
2755	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2756	AssertPtrNullReturn(ppvR3, VERR_INVALID_POINTER);
2757	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2758	AssertReturn(ppvR3 \|\| ppvR0, VERR_INVALID_PARAMETER);
2759	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2760	if (cPages < 1 \|\| cPages > VBOX_MAX_ALLOC_PAGE_COUNT)
2761	{
2762	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)));
2763	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2764	}
2765
2766	/*
2767	* Let IPRT do the work.
2768	*/
2769	if (ppvR0)
2770	rc = RTR0MemObjAllocPage(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, true /* fExecutable */);
2771	else
2772	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
2773	if (RT_SUCCESS(rc))
2774	{
2775	int rc2;
2776	if (ppvR3)
2777	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2778	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2779	else
2780	Mem.MapObjR3 = NIL_RTR0MEMOBJ;
2781	if (RT_SUCCESS(rc))
2782	{
2783	Mem.eType = MEMREF_TYPE_PAGE;
2784	rc = supdrvMemAdd(&Mem, pSession);
2785	if (!rc)
2786	{
2787	if (ppvR3)
2788	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2789	if (ppvR0)
2790	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2791	if (paPages)
2792	{
2793	uint32_t iPage = cPages;
2794	while (iPage-- > 0)
2795	{
2796	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
2797	Assert(paPages[iPage] != NIL_RTHCPHYS);
2798	}
2799	}
2800	return VINF_SUCCESS;
2801	}
2802
2803	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2804	AssertRC(rc2);
2805	}
2806
2807	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2808	AssertRC(rc2);
2809	}
2810	return rc;
2811	}
2812
2813
2814	/**
2815	* Maps a chunk of memory previously allocated by SUPR0PageAllocEx into kernel
2816	* space.
2817	*
2818	* @returns IPRT status code.
2819	* @param pSession The session to associated the allocation with.
2820	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
2821	* @param offSub Where to start mapping. Must be page aligned.
2822	* @param cbSub How much to map. Must be page aligned.
2823	* @param fFlags Flags, MBZ.
2824	* @param ppvR0 Where to return the address of the ring-0 mapping on
2825	* success.
2826	*/
2827	SUPR0DECL(int) SUPR0PageMapKernel(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t offSub, uint32_t cbSub,
2828	uint32_t fFlags, PRTR0PTR ppvR0)
2829	{
2830	int rc;
2831	PSUPDRVBUNDLE pBundle;
2832	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2833	RTR0MEMOBJ hMemObj = NIL_RTR0MEMOBJ;
2834	LogFlow(("SUPR0PageMapKernel: pSession=%p pvR3=%p offSub=%#x cbSub=%#x\n", pSession, pvR3, offSub, cbSub));
2835
2836	/*
2837	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2838	*/
2839	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2840	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2841	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2842	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2843	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2844	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
2845
2846	/*
2847	* Find the memory object.
2848	*/
2849	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2850	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2851	{
2852	if (pBundle->cUsed > 0)
2853	{
2854	unsigned i;
2855	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2856	{
2857	if ( ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
2858	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2859	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2860	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
2861	\|\| ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED
2862	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2863	&& pBundle->aMem[i].MapObjR3 == NIL_RTR0MEMOBJ
2864	&& RTR0MemObjAddressR3(pBundle->aMem[i].MemObj) == pvR3))
2865	{
2866	hMemObj = pBundle->aMem[i].MemObj;
2867	break;
2868	}
2869	}
2870	}
2871	}
2872	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2873
2874	rc = VERR_INVALID_PARAMETER;
2875	if (hMemObj != NIL_RTR0MEMOBJ)
2876	{
2877	/*
2878	* Do some further input validations before calling IPRT.
2879	* (Cleanup is done indirectly by telling RTR0MemObjFree to include mappings.)
2880	*/
2881	size_t cbMemObj = RTR0MemObjSize(hMemObj);
2882	if ( offSub < cbMemObj
2883	&& cbSub <= cbMemObj
2884	&& offSub + cbSub <= cbMemObj)
2885	{
2886	RTR0MEMOBJ hMapObj;
2887	rc = RTR0MemObjMapKernelEx(&hMapObj, hMemObj, (void *)-1, 0,
2888	RTMEM_PROT_READ \| RTMEM_PROT_WRITE, offSub, cbSub);
2889	if (RT_SUCCESS(rc))
2890	*ppvR0 = RTR0MemObjAddress(hMapObj);
2891	}
2892	else
2893	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
2894
2895	}
2896	return rc;
2897	}
2898
2899
2900	/**
2901	* Changes the page level protection of one or more pages previously allocated
2902	* by SUPR0PageAllocEx.
2903	*
2904	* @returns IPRT status code.
2905	* @param pSession The session to associated the allocation with.
2906	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
2907	* NIL_RTR3PTR if the ring-3 mapping should be unaffected.
2908	* @param pvR0 The ring-0 address returned by SUPR0PageAllocEx.
2909	* NIL_RTR0PTR if the ring-0 mapping should be unaffected.
2910	* @param offSub Where to start changing. Must be page aligned.
2911	* @param cbSub How much to change. Must be page aligned.
2912	* @param fProt The new page level protection, see RTMEM_PROT_*.
2913	*/
2914	SUPR0DECL(int) SUPR0PageProtect(PSUPDRVSESSION pSession, RTR3PTR pvR3, RTR0PTR pvR0, uint32_t offSub, uint32_t cbSub, uint32_t fProt)
2915	{
2916	int rc;
2917	PSUPDRVBUNDLE pBundle;
2918	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2919	RTR0MEMOBJ hMemObjR0 = NIL_RTR0MEMOBJ;
2920	RTR0MEMOBJ hMemObjR3 = NIL_RTR0MEMOBJ;
2921	LogFlow(("SUPR0PageProtect: pSession=%p pvR3=%p pvR0=%p offSub=%#x cbSub=%#x fProt-%#x\n", pSession, pvR3, pvR0, offSub, cbSub, fProt));
2922
2923	/*
2924	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2925	*/
2926	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2927	AssertReturn(!(fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)), VERR_INVALID_PARAMETER);
2928	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2929	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2930	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
2931
2932	/*
2933	* Find the memory object.
2934	*/
2935	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2936	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2937	{
2938	if (pBundle->cUsed > 0)
2939	{
2940	unsigned i;
2941	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2942	{
2943	if ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
2944	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2945	&& ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2946	\|\| pvR3 == NIL_RTR3PTR)
2947	&& ( pvR0 == NIL_RTR0PTR
2948	\|\| RTR0MemObjAddress(pBundle->aMem[i].MemObj) == pvR0)
2949	&& ( pvR3 == NIL_RTR3PTR
2950	\|\| RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3))
2951	{
2952	if (pvR0 != NIL_RTR0PTR)
2953	hMemObjR0 = pBundle->aMem[i].MemObj;
2954	if (pvR3 != NIL_RTR3PTR)
2955	hMemObjR3 = pBundle->aMem[i].MapObjR3;
2956	break;
2957	}
2958	}
2959	}
2960	}
2961	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2962
2963	rc = VERR_INVALID_PARAMETER;
2964	if ( hMemObjR0 != NIL_RTR0MEMOBJ
2965	\|\| hMemObjR3 != NIL_RTR0MEMOBJ)
2966	{
2967	/*
2968	* Do some further input validations before calling IPRT.
2969	*/
2970	size_t cbMemObj = hMemObjR0 != NIL_RTR0PTR ? RTR0MemObjSize(hMemObjR0) : RTR0MemObjSize(hMemObjR3);
2971	if ( offSub < cbMemObj
2972	&& cbSub <= cbMemObj
2973	&& offSub + cbSub <= cbMemObj)
2974	{
2975	rc = VINF_SUCCESS;
2976	if (hMemObjR3 != NIL_RTR0PTR)
2977	rc = RTR0MemObjProtect(hMemObjR3, offSub, cbSub, fProt);
2978	if (hMemObjR0 != NIL_RTR0PTR && RT_SUCCESS(rc))
2979	rc = RTR0MemObjProtect(hMemObjR0, offSub, cbSub, fProt);
2980	}
2981	else
2982	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
2983
2984	}
2985	return rc;
2986
2987	}
2988
2989
2990	/**
2991	* Free memory allocated by SUPR0PageAlloc() and SUPR0PageAllocEx().
2992	*
2993	* @returns IPRT status code.
2994	* @param pSession The session owning the allocation.
2995	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc() or
2996	* SUPR0PageAllocEx().
2997	*/
2998	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2999	{
3000	LogFlow(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
3001	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3002	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_PAGE);
3003	}
3004
3005
3006	/**
3007	* Gets the paging mode of the current CPU.
3008	*
3009	* @returns Paging mode, SUPPAGEINGMODE_INVALID on error.
3010	*/
3011	SUPR0DECL(SUPPAGINGMODE) SUPR0GetPagingMode(void)
3012	{
3013	SUPPAGINGMODE enmMode;
3014
3015	RTR0UINTREG cr0 = ASMGetCR0();
3016	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
3017	enmMode = SUPPAGINGMODE_INVALID;
3018	else
3019	{
3020	RTR0UINTREG cr4 = ASMGetCR4();
3021	uint32_t fNXEPlusLMA = 0;
3022	if (cr4 & X86_CR4_PAE)
3023	{
3024	uint32_t fAmdFeatures = ASMCpuId_EDX(0x80000001);
3025	if (fAmdFeatures & (X86_CPUID_AMD_FEATURE_EDX_NX \| X86_CPUID_AMD_FEATURE_EDX_LONG_MODE))
3026	{
3027	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
3028	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
3029	fNXEPlusLMA \|= RT_BIT(0);
3030	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
3031	fNXEPlusLMA \|= RT_BIT(1);
3032	}
3033	}
3034
3035	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
3036	{
3037	case 0:
3038	enmMode = SUPPAGINGMODE_32_BIT;
3039	break;
3040
3041	case X86_CR4_PGE:
3042	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
3043	break;
3044
3045	case X86_CR4_PAE:
3046	enmMode = SUPPAGINGMODE_PAE;
3047	break;
3048
3049	case X86_CR4_PAE \| RT_BIT(0):
3050	enmMode = SUPPAGINGMODE_PAE_NX;
3051	break;
3052
3053	case X86_CR4_PAE \| X86_CR4_PGE:
3054	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3055	break;
3056
3057	case X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3058	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3059	break;
3060
3061	case RT_BIT(1) \| X86_CR4_PAE:
3062	enmMode = SUPPAGINGMODE_AMD64;
3063	break;
3064
3065	case RT_BIT(1) \| X86_CR4_PAE \| RT_BIT(0):
3066	enmMode = SUPPAGINGMODE_AMD64_NX;
3067	break;
3068
3069	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
3070	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
3071	break;
3072
3073	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3074	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
3075	break;
3076
3077	default:
3078	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
3079	enmMode = SUPPAGINGMODE_INVALID;
3080	break;
3081	}
3082	}
3083	return enmMode;
3084	}
3085
3086
3087	/**
3088	* Enables or disabled hardware virtualization extensions using native OS APIs.
3089	*
3090	* @returns VBox status code.
3091	* @retval VINF_SUCCESS on success.
3092	* @retval VERR_NOT_SUPPORTED if not supported by the native OS.
3093	*
3094	* @param fEnable Whether to enable or disable.
3095	*/
3096	SUPR0DECL(int) SUPR0EnableVTx(bool fEnable)
3097	{
3098	#ifdef RT_OS_DARWIN
3099	return supdrvOSEnableVTx(fEnable);
3100	#else
3101	return VERR_NOT_SUPPORTED;
3102	#endif
3103	}
3104
3105
3106	/** @todo document me */
3107	SUPR0DECL(int) SUPR0QueryVTCaps(PSUPDRVSESSION pSession, uint32_t *pfCaps)
3108	{
3109	/*
3110	* Input validation.
3111	*/
3112	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3113	AssertPtrReturn(pfCaps, VERR_INVALID_POINTER);
3114
3115	*pfCaps = 0;
3116
3117	if (ASMHasCpuId())
3118	{
3119	uint32_t u32FeaturesECX;
3120	uint32_t u32Dummy;
3121	uint32_t u32FeaturesEDX;
3122	uint32_t u32VendorEBX, u32VendorECX, u32VendorEDX, u32AMDFeatureEDX, u32AMDFeatureECX;
3123	uint64_t val;
3124
3125	ASMCpuId(0, &u32Dummy, &u32VendorEBX, &u32VendorECX, &u32VendorEDX);
3126	ASMCpuId(1, &u32Dummy, &u32Dummy, &u32FeaturesECX, &u32FeaturesEDX);
3127	/* Query AMD features. */
3128	ASMCpuId(0x80000001, &u32Dummy, &u32Dummy, &u32AMDFeatureECX, &u32AMDFeatureEDX);
3129
3130	if ( u32VendorEBX == X86_CPUID_VENDOR_INTEL_EBX
3131	&& u32VendorECX == X86_CPUID_VENDOR_INTEL_ECX
3132	&& u32VendorEDX == X86_CPUID_VENDOR_INTEL_EDX
3133	)
3134	{
3135	if ( (u32FeaturesECX & X86_CPUID_FEATURE_ECX_VMX)
3136	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3137	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3138	)
3139	{
3140	val = ASMRdMsr(MSR_IA32_FEATURE_CONTROL);
3141	/*
3142	* Both the LOCK and VMXON bit must be set; otherwise VMXON will generate a #GP.
3143	* Once the lock bit is set, this MSR can no longer be modified.
3144	*/
3145	if ( (val & (MSR_IA32_FEATURE_CONTROL_VMXON\|MSR_IA32_FEATURE_CONTROL_LOCK))
3146	== (MSR_IA32_FEATURE_CONTROL_VMXON\|MSR_IA32_FEATURE_CONTROL_LOCK) /* enabled and locked */
3147	\|\| !(val & MSR_IA32_FEATURE_CONTROL_LOCK) /* not enabled, but not locked either */
3148	)
3149	{
3150	VMX_CAPABILITY vtCaps;
3151
3152	*pfCaps \|= SUPVTCAPS_VT_X;
3153
3154	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS);
3155	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC_USE_SECONDARY_EXEC_CTRL)
3156	{
3157	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS2);
3158	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC2_EPT)
3159	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3160	}
3161	return VINF_SUCCESS;
3162	}
3163	return VERR_VMX_MSR_LOCKED_OR_DISABLED;
3164	}
3165	return VERR_VMX_NO_VMX;
3166	}
3167
3168	if ( u32VendorEBX == X86_CPUID_VENDOR_AMD_EBX
3169	&& u32VendorECX == X86_CPUID_VENDOR_AMD_ECX
3170	&& u32VendorEDX == X86_CPUID_VENDOR_AMD_EDX
3171	)
3172	{
3173	if ( (u32AMDFeatureECX & X86_CPUID_AMD_FEATURE_ECX_SVM)
3174	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3175	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3176	)
3177	{
3178	/* Check if SVM is disabled */
3179	val = ASMRdMsr(MSR_K8_VM_CR);
3180	if (!(val & MSR_K8_VM_CR_SVM_DISABLE))
3181	{
3182	*pfCaps \|= SUPVTCAPS_AMD_V;
3183
3184	/* Query AMD features. */
3185	ASMCpuId(0x8000000A, &u32Dummy, &u32Dummy, &u32Dummy, &u32FeaturesEDX);
3186
3187	if (u32FeaturesEDX & AMD_CPUID_SVM_FEATURE_EDX_NESTED_PAGING)
3188	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3189
3190	return VINF_SUCCESS;
3191	}
3192	return VERR_SVM_DISABLED;
3193	}
3194	return VERR_SVM_NO_SVM;
3195	}
3196	}
3197
3198	return VERR_UNSUPPORTED_CPU;
3199	}
3200
3201
3202	/**
3203	* (Re-)initializes the per-cpu structure prior to starting or resuming the GIP
3204	* updating.
3205	*
3206	* @param pGipCpu The per CPU structure for this CPU.
3207	* @param u64NanoTS The current time.
3208	*/
3209	static void supdrvGipReInitCpu(PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
3210	{
3211	pGipCpu->u64TSC = ASMReadTSC() - pGipCpu->u32UpdateIntervalTSC;
3212	pGipCpu->u64NanoTS = u64NanoTS;
3213	}
3214
3215
3216	/**
3217	* Set the current TSC and NanoTS value for the CPU.
3218	*
3219	* @param idCpu The CPU ID. Unused - we have to use the APIC ID.
3220	* @param pvUser1 Pointer to the ring-0 GIP mapping.
3221	* @param pvUser2 Pointer to the variable holding the current time.
3222	*/
3223	static DECLCALLBACK(void) supdrvGipReInitCpuCallback(RTCPUID idCpu, void pvUser1, void pvUser2)
3224	{
3225	PSUPGLOBALINFOPAGE pGip = (PSUPGLOBALINFOPAGE)pvUser1;
3226	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
3227
3228	if (RT_LIKELY(iCpu < pGip->cCpus && pGip->aCPUs[iCpu].idCpu == idCpu))
3229	supdrvGipReInitCpu(&pGip->aCPUs[iCpu], (uint64_t )pvUser2);
3230
3231	NOREF(pvUser2);
3232	NOREF(idCpu);
3233	}
3234
3235
3236	/**
3237	* Maps the GIP into userspace and/or get the physical address of the GIP.
3238	*
3239	* @returns IPRT status code.
3240	* @param pSession Session to which the GIP mapping should belong.
3241	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
3242	* @param pHCPhysGip Where to store the physical address. (optional)
3243	*
3244	* @remark There is no reference counting on the mapping, so one call to this function
3245	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
3246	* and remove the session as a GIP user.
3247	*/
3248	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
3249	{
3250	int rc;
3251	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3252	RTR3PTR pGipR3 = NIL_RTR3PTR;
3253	RTHCPHYS HCPhys = NIL_RTHCPHYS;
3254	LogFlow(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
3255
3256	/*
3257	* Validate
3258	*/
3259	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3260	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
3261	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
3262
3263	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3264	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3265	#else
3266	RTSemFastMutexRequest(pDevExt->mtxGip);
3267	#endif
3268	if (pDevExt->pGip)
3269	{
3270	/*
3271	* Map it?
3272	*/
3273	rc = VINF_SUCCESS;
3274	if (ppGipR3)
3275	{
3276	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
3277	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
3278	RTMEM_PROT_READ, RTR0ProcHandleSelf());
3279	if (RT_SUCCESS(rc))
3280	pGipR3 = RTR0MemObjAddressR3(pSession->GipMapObjR3);
3281	}
3282
3283	/*
3284	* Get physical address.
3285	*/
3286	if (pHCPhysGip && RT_SUCCESS(rc))
3287	HCPhys = pDevExt->HCPhysGip;
3288
3289	/*
3290	* Reference globally.
3291	*/
3292	if (!pSession->fGipReferenced && RT_SUCCESS(rc))
3293	{
3294	pSession->fGipReferenced = 1;
3295	pDevExt->cGipUsers++;
3296	if (pDevExt->cGipUsers == 1)
3297	{
3298	PSUPGLOBALINFOPAGE pGipR0 = pDevExt->pGip;
3299	uint64_t u64NanoTS;
3300	uint32_t u32SystemResolution;
3301	unsigned i;
3302
3303	LogFlow(("SUPR0GipMap: Resumes GIP updating\n"));
3304
3305	/*
3306	* Try bump up the system timer resolution.
3307	* The more interrupts the better...
3308	*/
3309	if ( RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
3310	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
3311	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1953125 /* 512 HZ */, &u32SystemResolution))
3312	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 2000000 /* 500 HZ */, &u32SystemResolution))
3313	)
3314	{
3315	Assert(RTTimerGetSystemGranularity() <= u32SystemResolution);
3316	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
3317	}
3318
3319	if (pGipR0->aCPUs[0].u32TransactionId != 2 /* not the first time */)
3320	{
3321	for (i = 0; i < RT_ELEMENTS(pGipR0->aCPUs); i++)
3322	ASMAtomicUoWriteU32(&pGipR0->aCPUs[i].u32TransactionId,
3323	(pGipR0->aCPUs[i].u32TransactionId + GIP_UPDATEHZ_RECALC_FREQ * 2)
3324	& ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
3325	ASMAtomicWriteU64(&pGipR0->u64NanoTSLastUpdateHz, 0);
3326	}
3327
3328	u64NanoTS = RTTimeSystemNanoTS() - pGipR0->u32UpdateIntervalNS;
3329	if ( pGipR0->u32Mode == SUPGIPMODE_SYNC_TSC
3330	\|\| RTMpGetOnlineCount() == 1)
3331	supdrvGipReInitCpu(&pGipR0->aCPUs[0], u64NanoTS);
3332	else
3333	RTMpOnAll(supdrvGipReInitCpuCallback, pGipR0, &u64NanoTS);
3334
3335	#ifndef DO_NOT_START_GIP
3336	rc = RTTimerStart(pDevExt->pGipTimer, 0); AssertRC(rc);
3337	#endif
3338	rc = VINF_SUCCESS;
3339	}
3340	}
3341	}
3342	else
3343	{
3344	rc = VERR_GENERAL_FAILURE;
3345	Log(("SUPR0GipMap: GIP is not available!\n"));
3346	}
3347	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3348	RTSemMutexRelease(pDevExt->mtxGip);
3349	#else
3350	RTSemFastMutexRelease(pDevExt->mtxGip);
3351	#endif
3352
3353	/*
3354	* Write returns.
3355	*/
3356	if (pHCPhysGip)
3357	*pHCPhysGip = HCPhys;
3358	if (ppGipR3)
3359	*ppGipR3 = pGipR3;
3360
3361	#ifdef DEBUG_DARWIN_GIP
3362	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3363	#else
3364	LogFlow(( "SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3365	#endif
3366	return rc;
3367	}
3368
3369
3370	/**
3371	* Unmaps any user mapping of the GIP and terminates all GIP access
3372	* from this session.
3373	*
3374	* @returns IPRT status code.
3375	* @param pSession Session to which the GIP mapping should belong.
3376	*/
3377	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
3378	{
3379	int rc = VINF_SUCCESS;
3380	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3381	#ifdef DEBUG_DARWIN_GIP
3382	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
3383	pSession,
3384	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
3385	pSession->GipMapObjR3));
3386	#else
3387	LogFlow(("SUPR0GipUnmap: pSession=%p\n", pSession));
3388	#endif
3389	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3390
3391	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3392	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3393	#else
3394	RTSemFastMutexRequest(pDevExt->mtxGip);
3395	#endif
3396
3397	/*
3398	* Unmap anything?
3399	*/
3400	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
3401	{
3402	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
3403	AssertRC(rc);
3404	if (RT_SUCCESS(rc))
3405	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
3406	}
3407
3408	/*
3409	* Dereference global GIP.
3410	*/
3411	if (pSession->fGipReferenced && !rc)
3412	{
3413	pSession->fGipReferenced = 0;
3414	if ( pDevExt->cGipUsers > 0
3415	&& !--pDevExt->cGipUsers)
3416	{
3417	LogFlow(("SUPR0GipUnmap: Suspends GIP updating\n"));
3418	#ifndef DO_NOT_START_GIP
3419	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = VINF_SUCCESS;
3420	#endif
3421
3422	if (pDevExt->u32SystemTimerGranularityGrant)
3423	{
3424	int rc2 = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant);
3425	AssertRC(rc2);
3426	pDevExt->u32SystemTimerGranularityGrant = 0;
3427	}
3428	}
3429	}
3430
3431	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3432	RTSemMutexRelease(pDevExt->mtxGip);
3433	#else
3434	RTSemFastMutexRelease(pDevExt->mtxGip);
3435	#endif
3436
3437	return rc;
3438	}
3439
3440
3441	/**
3442	* Gets the GIP pointer.
3443	*
3444	* @returns Pointer to the GIP or NULL.
3445	*/
3446	SUPDECL(PSUPGLOBALINFOPAGE) SUPGetGIP(void)
3447	{
3448	return g_pSUPGlobalInfoPage;
3449	}
3450
3451
3452	/**
3453	* Register a component factory with the support driver.
3454	*
3455	* This is currently restricted to kernel sessions only.
3456	*
3457	* @returns VBox status code.
3458	* @retval VINF_SUCCESS on success.
3459	* @retval VERR_NO_MEMORY if we're out of memory.
3460	* @retval VERR_ALREADY_EXISTS if the factory has already been registered.
3461	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3462	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3463	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3464	*
3465	* @param pSession The SUPDRV session (must be a ring-0 session).
3466	* @param pFactory Pointer to the component factory registration structure.
3467	*
3468	* @remarks This interface is also available via SUPR0IdcComponentRegisterFactory.
3469	*/
3470	SUPR0DECL(int) SUPR0ComponentRegisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3471	{
3472	PSUPDRVFACTORYREG pNewReg;
3473	const char *psz;
3474	int rc;
3475
3476	/*
3477	* Validate parameters.
3478	*/
3479	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3480	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3481	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3482	AssertPtrReturn(pFactory->pfnQueryFactoryInterface, VERR_INVALID_POINTER);
3483	psz = RTStrEnd(pFactory->szName, sizeof(pFactory->szName));
3484	AssertReturn(psz, VERR_INVALID_PARAMETER);
3485
3486	/*
3487	* Allocate and initialize a new registration structure.
3488	*/
3489	pNewReg = (PSUPDRVFACTORYREG)RTMemAlloc(sizeof(SUPDRVFACTORYREG));
3490	if (pNewReg)
3491	{
3492	pNewReg->pNext = NULL;
3493	pNewReg->pFactory = pFactory;
3494	pNewReg->pSession = pSession;
3495	pNewReg->cchName = psz - &pFactory->szName[0];
3496
3497	/*
3498	* Add it to the tail of the list after checking for prior registration.
3499	*/
3500	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3501	if (RT_SUCCESS(rc))
3502	{
3503	PSUPDRVFACTORYREG pPrev = NULL;
3504	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3505	while (pCur && pCur->pFactory != pFactory)
3506	{
3507	pPrev = pCur;
3508	pCur = pCur->pNext;
3509	}
3510	if (!pCur)
3511	{
3512	if (pPrev)
3513	pPrev->pNext = pNewReg;
3514	else
3515	pSession->pDevExt->pComponentFactoryHead = pNewReg;
3516	rc = VINF_SUCCESS;
3517	}
3518	else
3519	rc = VERR_ALREADY_EXISTS;
3520
3521	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3522	}
3523
3524	if (RT_FAILURE(rc))
3525	RTMemFree(pNewReg);
3526	}
3527	else
3528	rc = VERR_NO_MEMORY;
3529	return rc;
3530	}
3531
3532
3533	/**
3534	* Deregister a component factory.
3535	*
3536	* @returns VBox status code.
3537	* @retval VINF_SUCCESS on success.
3538	* @retval VERR_NOT_FOUND if the factory wasn't registered.
3539	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3540	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3541	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3542	*
3543	* @param pSession The SUPDRV session (must be a ring-0 session).
3544	* @param pFactory Pointer to the component factory registration structure
3545	* previously passed SUPR0ComponentRegisterFactory().
3546	*
3547	* @remarks This interface is also available via SUPR0IdcComponentDeregisterFactory.
3548	*/
3549	SUPR0DECL(int) SUPR0ComponentDeregisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3550	{
3551	int rc;
3552
3553	/*
3554	* Validate parameters.
3555	*/
3556	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3557	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3558	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3559
3560	/*
3561	* Take the lock and look for the registration record.
3562	*/
3563	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3564	if (RT_SUCCESS(rc))
3565	{
3566	PSUPDRVFACTORYREG pPrev = NULL;
3567	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3568	while (pCur && pCur->pFactory != pFactory)
3569	{
3570	pPrev = pCur;
3571	pCur = pCur->pNext;
3572	}
3573	if (pCur)
3574	{
3575	if (!pPrev)
3576	pSession->pDevExt->pComponentFactoryHead = pCur->pNext;
3577	else
3578	pPrev->pNext = pCur->pNext;
3579
3580	pCur->pNext = NULL;
3581	pCur->pFactory = NULL;
3582	pCur->pSession = NULL;
3583	rc = VINF_SUCCESS;
3584	}
3585	else
3586	rc = VERR_NOT_FOUND;
3587
3588	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3589
3590	RTMemFree(pCur);
3591	}
3592	return rc;
3593	}
3594
3595
3596	/**
3597	* Queries a component factory.
3598	*
3599	* @returns VBox status code.
3600	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3601	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3602	* @retval VERR_SUPDRV_COMPONENT_NOT_FOUND if the component factory wasn't found.
3603	* @retval VERR_SUPDRV_INTERFACE_NOT_SUPPORTED if the interface wasn't supported.
3604	*
3605	* @param pSession The SUPDRV session.
3606	* @param pszName The name of the component factory.
3607	* @param pszInterfaceUuid The UUID of the factory interface (stringified).
3608	* @param ppvFactoryIf Where to store the factory interface.
3609	*/
3610	SUPR0DECL(int) SUPR0ComponentQueryFactory(PSUPDRVSESSION pSession, const char pszName, const char pszInterfaceUuid, void **ppvFactoryIf)
3611	{
3612	const char *pszEnd;
3613	size_t cchName;
3614	int rc;
3615
3616	/*
3617	* Validate parameters.
3618	*/
3619	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3620
3621	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
3622	pszEnd = RTStrEnd(pszName, RT_SIZEOFMEMB(SUPDRVFACTORY, szName));
3623	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3624	cchName = pszEnd - pszName;
3625
3626	AssertPtrReturn(pszInterfaceUuid, VERR_INVALID_POINTER);
3627	pszEnd = RTStrEnd(pszInterfaceUuid, RTUUID_STR_LENGTH);
3628	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3629
3630	AssertPtrReturn(ppvFactoryIf, VERR_INVALID_POINTER);
3631	*ppvFactoryIf = NULL;
3632
3633	/*
3634	* Take the lock and try all factories by this name.
3635	*/
3636	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3637	if (RT_SUCCESS(rc))
3638	{
3639	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3640	rc = VERR_SUPDRV_COMPONENT_NOT_FOUND;
3641	while (pCur)
3642	{
3643	if ( pCur->cchName == cchName
3644	&& !memcmp(pCur->pFactory->szName, pszName, cchName))
3645	{
3646	void *pvFactory = pCur->pFactory->pfnQueryFactoryInterface(pCur->pFactory, pSession, pszInterfaceUuid);
3647	if (pvFactory)
3648	{
3649	*ppvFactoryIf = pvFactory;
3650	rc = VINF_SUCCESS;
3651	break;
3652	}
3653	rc = VERR_SUPDRV_INTERFACE_NOT_SUPPORTED;
3654	}
3655
3656	/* next */
3657	pCur = pCur->pNext;
3658	}
3659
3660	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3661	}
3662	return rc;
3663	}
3664
3665
3666	/**
3667	* Adds a memory object to the session.
3668	*
3669	* @returns IPRT status code.
3670	* @param pMem Memory tracking structure containing the
3671	* information to track.
3672	* @param pSession The session.
3673	*/
3674	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
3675	{
3676	PSUPDRVBUNDLE pBundle;
3677	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3678
3679	/*
3680	* Find free entry and record the allocation.
3681	*/
3682	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3683	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3684	{
3685	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
3686	{
3687	unsigned i;
3688	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3689	{
3690	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
3691	{
3692	pBundle->cUsed++;
3693	pBundle->aMem[i] = *pMem;
3694	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3695	return VINF_SUCCESS;
3696	}
3697	}
3698	AssertFailed(); /* !!this can't be happening!!! */
3699	}
3700	}
3701	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3702
3703	/*
3704	* Need to allocate a new bundle.
3705	* Insert into the last entry in the bundle.
3706	*/
3707	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
3708	if (!pBundle)
3709	return VERR_NO_MEMORY;
3710
3711	/* take last entry. */
3712	pBundle->cUsed++;
3713	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
3714
3715	/* insert into list. */
3716	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3717	pBundle->pNext = pSession->Bundle.pNext;
3718	pSession->Bundle.pNext = pBundle;
3719	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3720
3721	return VINF_SUCCESS;
3722	}
3723
3724
3725	/**
3726	* Releases a memory object referenced by pointer and type.
3727	*
3728	* @returns IPRT status code.
3729	* @param pSession Session data.
3730	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
3731	* @param eType Memory type.
3732	*/
3733	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
3734	{
3735	PSUPDRVBUNDLE pBundle;
3736	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3737
3738	/*
3739	* Validate input.
3740	*/
3741	if (!uPtr)
3742	{
3743	Log(("Illegal address %p\n", (void *)uPtr));
3744	return VERR_INVALID_PARAMETER;
3745	}
3746
3747	/*
3748	* Search for the address.
3749	*/
3750	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3751	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3752	{
3753	if (pBundle->cUsed > 0)
3754	{
3755	unsigned i;
3756	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3757	{
3758	if ( pBundle->aMem[i].eType == eType
3759	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3760	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
3761	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3762	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
3763	)
3764	{
3765	/* Make a copy of it and release it outside the spinlock. */
3766	SUPDRVMEMREF Mem = pBundle->aMem[i];
3767	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
3768	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
3769	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
3770	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3771
3772	if (Mem.MapObjR3 != NIL_RTR0MEMOBJ)
3773	{
3774	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
3775	AssertRC(rc); /** @todo figure out how to handle this. */
3776	}
3777	if (Mem.MemObj != NIL_RTR0MEMOBJ)
3778	{
3779	int rc = RTR0MemObjFree(Mem.MemObj, true /* fFreeMappings */);
3780	AssertRC(rc); /** @todo figure out how to handle this. */
3781	}
3782	return VINF_SUCCESS;
3783	}
3784	}
3785	}
3786	}
3787	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3788	Log(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
3789	return VERR_INVALID_PARAMETER;
3790	}
3791
3792
3793	/**
3794	* Opens an image. If it's the first time it's opened the call must upload
3795	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
3796	*
3797	* This is the 1st step of the loading.
3798	*
3799	* @returns IPRT status code.
3800	* @param pDevExt Device globals.
3801	* @param pSession Session data.
3802	* @param pReq The open request.
3803	*/
3804	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
3805	{
3806	int rc;
3807	PSUPDRVLDRIMAGE pImage;
3808	void *pv;
3809	size_t cchName = strlen(pReq->u.In.szName); /* (caller checked < 32). */
3810	LogFlow(("supdrvIOCtl_LdrOpen: szName=%s cbImageWithTabs=%d\n", pReq->u.In.szName, pReq->u.In.cbImageWithTabs));
3811
3812	/*
3813	* Check if we got an instance of the image already.
3814	*/
3815	supdrvLdrLock(pDevExt);
3816	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
3817	{
3818	if ( pImage->szName[cchName] == '\0'
3819	&& !memcmp(pImage->szName, pReq->u.In.szName, cchName))
3820	{
3821	/** @todo check cbImageBits and cbImageWithTabs here, if they differs that indicates that the images are different. */
3822	pImage->cUsage++;
3823	pReq->u.Out.pvImageBase = pImage->pvImage;
3824	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
3825	pReq->u.Out.fNativeLoader = pImage->fNative;
3826	supdrvLdrAddUsage(pSession, pImage);
3827	supdrvLdrUnlock(pDevExt);
3828	return VINF_SUCCESS;
3829	}
3830	}
3831	/* (not found - add it!) */
3832
3833	/*
3834	* Allocate memory.
3835	*/
3836	pv = RTMemAlloc(RT_OFFSETOF(SUPDRVLDRIMAGE, szName[cchName + 1]));
3837	if (!pv)
3838	{
3839	supdrvLdrUnlock(pDevExt);
3840	Log(("supdrvIOCtl_LdrOpen: RTMemAlloc() failed\n"));
3841	return /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_2;
3842	}
3843
3844	/*
3845	* Setup and link in the LDR stuff.
3846	*/
3847	pImage = (PSUPDRVLDRIMAGE)pv;
3848	pImage->pvImage = NULL;
3849	pImage->pvImageAlloc = NULL;
3850	pImage->cbImageWithTabs = pReq->u.In.cbImageWithTabs;
3851	pImage->cbImageBits = pReq->u.In.cbImageBits;
3852	pImage->cSymbols = 0;
3853	pImage->paSymbols = NULL;
3854	pImage->pachStrTab = NULL;
3855	pImage->cbStrTab = 0;
3856	pImage->pfnModuleInit = NULL;
3857	pImage->pfnModuleTerm = NULL;
3858	pImage->pfnServiceReqHandler = NULL;
3859	pImage->uState = SUP_IOCTL_LDR_OPEN;
3860	pImage->cUsage = 1;
3861	memcpy(pImage->szName, pReq->u.In.szName, cchName + 1);
3862
3863	/*
3864	* Try load it using the native loader, if that isn't supported, fall back
3865	* on the older method.
3866	*/
3867	pImage->fNative = true;
3868	rc = supdrvOSLdrOpen(pDevExt, pImage, pReq->u.In.szFilename);
3869	if (rc == VERR_NOT_SUPPORTED)
3870	{
3871	pImage->pvImageAlloc = RTMemExecAlloc(pImage->cbImageBits + 31);
3872	pImage->pvImage = RT_ALIGN_P(pImage->pvImageAlloc, 32);
3873	pImage->fNative = false;
3874	rc = pImage->pvImageAlloc ? VINF_SUCCESS : VERR_NO_EXEC_MEMORY;
3875	}
3876	if (RT_FAILURE(rc))
3877	{
3878	supdrvLdrUnlock(pDevExt);
3879	RTMemFree(pImage);
3880	Log(("supdrvIOCtl_LdrOpen(%s): failed - %Rrc\n", pReq->u.In.szName, rc));
3881	return rc;
3882	}
3883	Assert(VALID_PTR(pImage->pvImage) \|\| RT_FAILURE(rc));
3884
3885	/*
3886	* Link it.
3887	*/
3888	pImage->pNext = pDevExt->pLdrImages;
3889	pDevExt->pLdrImages = pImage;
3890
3891	supdrvLdrAddUsage(pSession, pImage);
3892
3893	pReq->u.Out.pvImageBase = pImage->pvImage;
3894	pReq->u.Out.fNeedsLoading = true;
3895	pReq->u.Out.fNativeLoader = pImage->fNative;
3896	supdrvLdrUnlock(pDevExt);
3897
3898	#if defined(RT_OS_WINDOWS) && defined(DEBUG)
3899	SUPR0Printf("VBoxDrv: windbg> .reload /f %s=%#p\n", pImage->szName, pImage->pvImage);
3900	#endif
3901	return VINF_SUCCESS;
3902	}
3903
3904
3905	/**
3906	* Worker that validates a pointer to an image entrypoint.
3907	*
3908	* @returns IPRT status code.
3909	* @param pDevExt The device globals.
3910	* @param pImage The loader image.
3911	* @param pv The pointer into the image.
3912	* @param fMayBeNull Whether it may be NULL.
3913	* @param pszWhat What is this entrypoint? (for logging)
3914	* @param pbImageBits The image bits prepared by ring-3.
3915	*
3916	* @remarks Will leave the lock on failure.
3917	*/
3918	static int supdrvLdrValidatePointer(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage, void *pv,
3919	bool fMayBeNull, const uint8_t pbImageBits, const char pszWhat)
3920	{
3921	if (!fMayBeNull \|\| pv)
3922	{
3923	if ((uintptr_t)pv - (uintptr_t)pImage->pvImage >= pImage->cbImageBits)
3924	{
3925	supdrvLdrUnlock(pDevExt);
3926	Log(("Out of range (%p LB %#x): %s=%p\n", pImage->pvImage, pImage->cbImageBits, pszWhat, pv));
3927	return VERR_INVALID_PARAMETER;
3928	}
3929
3930	if (pImage->fNative)
3931	{
3932	int rc = supdrvOSLdrValidatePointer(pDevExt, pImage, pv, pbImageBits);
3933	if (RT_FAILURE(rc))
3934	{
3935	supdrvLdrUnlock(pDevExt);
3936	Log(("Bad entry point address: %s=%p (rc=%Rrc)\n", pszWhat, pv, rc));
3937	return rc;
3938	}
3939	}
3940	}
3941	return VINF_SUCCESS;
3942	}
3943
3944
3945	/**
3946	* Loads the image bits.
3947	*
3948	* This is the 2nd step of the loading.
3949	*
3950	* @returns IPRT status code.
3951	* @param pDevExt Device globals.
3952	* @param pSession Session data.
3953	* @param pReq The request.
3954	*/
3955	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
3956	{
3957	PSUPDRVLDRUSAGE pUsage;
3958	PSUPDRVLDRIMAGE pImage;
3959	int rc;
3960	LogFlow(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImageWithBits=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImageWithTabs));
3961
3962	/*
3963	* Find the ldr image.
3964	*/
3965	supdrvLdrLock(pDevExt);
3966	pUsage = pSession->pLdrUsage;
3967	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3968	pUsage = pUsage->pNext;
3969	if (!pUsage)
3970	{
3971	supdrvLdrUnlock(pDevExt);
3972	Log(("SUP_IOCTL_LDR_LOAD: couldn't find image!\n"));
3973	return VERR_INVALID_HANDLE;
3974	}
3975	pImage = pUsage->pImage;
3976
3977	/*
3978	* Validate input.
3979	*/
3980	if ( pImage->cbImageWithTabs != pReq->u.In.cbImageWithTabs
3981	\|\| pImage->cbImageBits != pReq->u.In.cbImageBits)
3982	{
3983	supdrvLdrUnlock(pDevExt);
3984	Log(("SUP_IOCTL_LDR_LOAD: image size mismatch!! %d(prep) != %d(load) or %d != %d\n",
3985	pImage->cbImageWithTabs, pReq->u.In.cbImageWithTabs, pImage->cbImageBits, pReq->u.In.cbImageBits));
3986	return VERR_INVALID_HANDLE;
3987	}
3988
3989	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
3990	{
3991	unsigned uState = pImage->uState;
3992	supdrvLdrUnlock(pDevExt);
3993	if (uState != SUP_IOCTL_LDR_LOAD)
3994	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
3995	return VERR_ALREADY_LOADED;
3996	}
3997
3998	switch (pReq->u.In.eEPType)
3999	{
4000	case SUPLDRLOADEP_NOTHING:
4001	break;
4002
4003	case SUPLDRLOADEP_VMMR0:
4004	rc = supdrvLdrValidatePointer( pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0, false, pReq->u.In.abImage, "pvVMMR0");
4005	if (RT_SUCCESS(rc))
4006	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt, false, pReq->u.In.abImage, "pvVMMR0EntryInt");
4007	if (RT_SUCCESS(rc))
4008	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, false, pReq->u.In.abImage, "pvVMMR0EntryFast");
4009	if (RT_SUCCESS(rc))
4010	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx, false, pReq->u.In.abImage, "pvVMMR0EntryEx");
4011	if (RT_FAILURE(rc))
4012	return rc;
4013	break;
4014
4015	case SUPLDRLOADEP_SERVICE:
4016	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.Service.pfnServiceReq, false, pReq->u.In.abImage, "pfnServiceReq");
4017	if (RT_FAILURE(rc))
4018	return rc;
4019	if ( pReq->u.In.EP.Service.apvReserved[0] != NIL_RTR0PTR
4020	\|\| pReq->u.In.EP.Service.apvReserved[1] != NIL_RTR0PTR
4021	\|\| pReq->u.In.EP.Service.apvReserved[2] != NIL_RTR0PTR)
4022	{
4023	supdrvLdrUnlock(pDevExt);
4024	Log(("Out of range (%p LB %#x): apvReserved={%p,%p,%p} MBZ!\n",
4025	pImage->pvImage, pReq->u.In.cbImageWithTabs,
4026	pReq->u.In.EP.Service.apvReserved[0],
4027	pReq->u.In.EP.Service.apvReserved[1],
4028	pReq->u.In.EP.Service.apvReserved[2]));
4029	return VERR_INVALID_PARAMETER;
4030	}
4031	break;
4032
4033	default:
4034	supdrvLdrUnlock(pDevExt);
4035	Log(("Invalid eEPType=%d\n", pReq->u.In.eEPType));
4036	return VERR_INVALID_PARAMETER;
4037	}
4038
4039	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleInit, true, pReq->u.In.abImage, "pfnModuleInit");
4040	if (RT_FAILURE(rc))
4041	return rc;
4042	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleTerm, true, pReq->u.In.abImage, "pfnModuleTerm");
4043	if (RT_FAILURE(rc))
4044	return rc;
4045
4046	/*
4047	* Allocate and copy the tables.
4048	* (No need to do try/except as this is a buffered request.)
4049	*/
4050	pImage->cbStrTab = pReq->u.In.cbStrTab;
4051	if (pImage->cbStrTab)
4052	{
4053	pImage->pachStrTab = (char *)RTMemAlloc(pImage->cbStrTab);
4054	if (pImage->pachStrTab)
4055	memcpy(pImage->pachStrTab, &pReq->u.In.abImage[pReq->u.In.offStrTab], pImage->cbStrTab);
4056	else
4057	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_3;
4058	}
4059
4060	pImage->cSymbols = pReq->u.In.cSymbols;
4061	if (RT_SUCCESS(rc) && pImage->cSymbols)
4062	{
4063	size_t cbSymbols = pImage->cSymbols * sizeof(SUPLDRSYM);
4064	pImage->paSymbols = (PSUPLDRSYM)RTMemAlloc(cbSymbols);
4065	if (pImage->paSymbols)
4066	memcpy(pImage->paSymbols, &pReq->u.In.abImage[pReq->u.In.offSymbols], cbSymbols);
4067	else
4068	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_4;
4069	}
4070
4071	/*
4072	* Copy the bits / complete native loading.
4073	*/
4074	if (RT_SUCCESS(rc))
4075	{
4076	pImage->uState = SUP_IOCTL_LDR_LOAD;
4077	pImage->pfnModuleInit = pReq->u.In.pfnModuleInit;
4078	pImage->pfnModuleTerm = pReq->u.In.pfnModuleTerm;
4079
4080	if (pImage->fNative)
4081	rc = supdrvOSLdrLoad(pDevExt, pImage, pReq->u.In.abImage, pReq);
4082	else
4083	memcpy(pImage->pvImage, &pReq->u.In.abImage[0], pImage->cbImageBits);
4084	}
4085
4086	/*
4087	* Update any entry points.
4088	*/
4089	if (RT_SUCCESS(rc))
4090	{
4091	switch (pReq->u.In.eEPType)
4092	{
4093	default:
4094	case SUPLDRLOADEP_NOTHING:
4095	rc = VINF_SUCCESS;
4096	break;
4097	case SUPLDRLOADEP_VMMR0:
4098	rc = supdrvLdrSetVMMR0EPs(pDevExt, pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
4099	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
4100	break;
4101	case SUPLDRLOADEP_SERVICE:
4102	pImage->pfnServiceReqHandler = pReq->u.In.EP.Service.pfnServiceReq;
4103	rc = VINF_SUCCESS;
4104	break;
4105	}
4106	}
4107
4108	/*
4109	* On success call the module initialization.
4110	*/
4111	LogFlow(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
4112	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
4113	{
4114	Log(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
4115	rc = pImage->pfnModuleInit();
4116	if (rc && pDevExt->pvVMMR0 == pImage->pvImage)
4117	supdrvLdrUnsetVMMR0EPs(pDevExt);
4118	}
4119
4120	if (RT_FAILURE(rc))
4121	{
4122	pImage->uState = SUP_IOCTL_LDR_OPEN;
4123	pImage->pfnModuleInit = NULL;
4124	pImage->pfnModuleTerm = NULL;
4125	pImage->pfnServiceReqHandler= NULL;
4126	pImage->cbStrTab = 0;
4127	RTMemFree(pImage->pachStrTab);
4128	pImage->pachStrTab = NULL;
4129	RTMemFree(pImage->paSymbols);
4130	pImage->paSymbols = NULL;
4131	pImage->cSymbols = 0;
4132	}
4133
4134	supdrvLdrUnlock(pDevExt);
4135	return rc;
4136	}
4137
4138
4139	/**
4140	* Frees a previously loaded (prep'ed) image.
4141	*
4142	* @returns IPRT status code.
4143	* @param pDevExt Device globals.
4144	* @param pSession Session data.
4145	* @param pReq The request.
4146	*/
4147	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
4148	{
4149	int rc;
4150	PSUPDRVLDRUSAGE pUsagePrev;
4151	PSUPDRVLDRUSAGE pUsage;
4152	PSUPDRVLDRIMAGE pImage;
4153	LogFlow(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
4154
4155	/*
4156	* Find the ldr image.
4157	*/
4158	supdrvLdrLock(pDevExt);
4159	pUsagePrev = NULL;
4160	pUsage = pSession->pLdrUsage;
4161	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4162	{
4163	pUsagePrev = pUsage;
4164	pUsage = pUsage->pNext;
4165	}
4166	if (!pUsage)
4167	{
4168	supdrvLdrUnlock(pDevExt);
4169	Log(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
4170	return VERR_INVALID_HANDLE;
4171	}
4172
4173	/*
4174	* Check if we can remove anything.
4175	*/
4176	rc = VINF_SUCCESS;
4177	pImage = pUsage->pImage;
4178	if (pImage->cUsage <= 1 \|\| pUsage->cUsage <= 1)
4179	{
4180	/*
4181	* Check if there are any objects with destructors in the image, if
4182	* so leave it for the session cleanup routine so we get a chance to
4183	* clean things up in the right order and not leave them all dangling.
4184	*/
4185	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
4186	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
4187	if (pImage->cUsage <= 1)
4188	{
4189	PSUPDRVOBJ pObj;
4190	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4191	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4192	{
4193	rc = VERR_DANGLING_OBJECTS;
4194	break;
4195	}
4196	}
4197	else
4198	{
4199	PSUPDRVUSAGE pGenUsage;
4200	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
4201	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4202	{
4203	rc = VERR_DANGLING_OBJECTS;
4204	break;
4205	}
4206	}
4207	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
4208	if (rc == VINF_SUCCESS)
4209	{
4210	/* unlink it */
4211	if (pUsagePrev)
4212	pUsagePrev->pNext = pUsage->pNext;
4213	else
4214	pSession->pLdrUsage = pUsage->pNext;
4215
4216	/* free it */
4217	pUsage->pImage = NULL;
4218	pUsage->pNext = NULL;
4219	RTMemFree(pUsage);
4220
4221	/*
4222	* Dereference the image.
4223	*/
4224	if (pImage->cUsage <= 1)
4225	supdrvLdrFree(pDevExt, pImage);
4226	else
4227	pImage->cUsage--;
4228	}
4229	else
4230	{
4231	Log(("supdrvIOCtl_LdrFree: Dangling objects in %p/%s!\n", pImage->pvImage, pImage->szName));
4232	rc = VINF_SUCCESS; /** @todo BRANCH-2.1: remove this after branching. */
4233	}
4234	}
4235	else
4236	{
4237	/*
4238	* Dereference both image and usage.
4239	*/
4240	pImage->cUsage--;
4241	pUsage->cUsage--;
4242	}
4243
4244	supdrvLdrUnlock(pDevExt);
4245	return rc;
4246	}
4247
4248
4249	/**
4250	* Gets the address of a symbol in an open image.
4251	*
4252	* @returns IPRT status code.
4253	* @param pDevExt Device globals.
4254	* @param pSession Session data.
4255	* @param pReq The request buffer.
4256	*/
4257	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
4258	{
4259	PSUPDRVLDRIMAGE pImage;
4260	PSUPDRVLDRUSAGE pUsage;
4261	uint32_t i;
4262	PSUPLDRSYM paSyms;
4263	const char *pchStrings;
4264	const size_t cbSymbol = strlen(pReq->u.In.szSymbol) + 1;
4265	void *pvSymbol = NULL;
4266	int rc = VERR_GENERAL_FAILURE;
4267	Log3(("supdrvIOCtl_LdrGetSymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
4268
4269	/*
4270	* Find the ldr image.
4271	*/
4272	supdrvLdrLock(pDevExt);
4273	pUsage = pSession->pLdrUsage;
4274	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4275	pUsage = pUsage->pNext;
4276	if (!pUsage)
4277	{
4278	supdrvLdrUnlock(pDevExt);
4279	Log(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
4280	return VERR_INVALID_HANDLE;
4281	}
4282	pImage = pUsage->pImage;
4283	if (pImage->uState != SUP_IOCTL_LDR_LOAD)
4284	{
4285	unsigned uState = pImage->uState;
4286	supdrvLdrUnlock(pDevExt);
4287	Log(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", uState, uState)); NOREF(uState);
4288	return VERR_ALREADY_LOADED;
4289	}
4290
4291	/*
4292	* Search the symbol strings.
4293	*
4294	* Note! The int32_t is for native loading on solaris where the data
4295	* and text segments are in very different places.
4296	*/
4297	pchStrings = pImage->pachStrTab;
4298	paSyms = pImage->paSymbols;
4299	for (i = 0; i < pImage->cSymbols; i++)
4300	{
4301	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4302	&& !memcmp(pchStrings + paSyms[i].offName, pReq->u.In.szSymbol, cbSymbol))
4303	{
4304	pvSymbol = (uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol;
4305	rc = VINF_SUCCESS;
4306	break;
4307	}
4308	}
4309	supdrvLdrUnlock(pDevExt);
4310	pReq->u.Out.pvSymbol = pvSymbol;
4311	return rc;
4312	}
4313
4314
4315	/**
4316	* Gets the address of a symbol in an open image or the support driver.
4317	*
4318	* @returns VINF_SUCCESS on success.
4319	* @returns
4320	* @param pDevExt Device globals.
4321	* @param pSession Session data.
4322	* @param pReq The request buffer.
4323	*/
4324	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq)
4325	{
4326	int rc = VINF_SUCCESS;
4327	const char *pszSymbol = pReq->u.In.pszSymbol;
4328	const char *pszModule = pReq->u.In.pszModule;
4329	size_t cbSymbol;
4330	char const *pszEnd;
4331	uint32_t i;
4332
4333	/*
4334	* Input validation.
4335	*/
4336	AssertPtrReturn(pszSymbol, VERR_INVALID_POINTER);
4337	pszEnd = RTStrEnd(pszSymbol, 512);
4338	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4339	cbSymbol = pszEnd - pszSymbol + 1;
4340
4341	if (pszModule)
4342	{
4343	AssertPtrReturn(pszModule, VERR_INVALID_POINTER);
4344	pszEnd = RTStrEnd(pszModule, 64);
4345	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4346	}
4347	Log3(("supdrvIDC_LdrGetSymbol: pszModule=%p:{%s} pszSymbol=%p:{%s}\n", pszModule, pszModule, pszSymbol, pszSymbol));
4348
4349
4350	if ( !pszModule
4351	\|\| !strcmp(pszModule, "SupDrv"))
4352	{
4353	/*
4354	* Search the support driver export table.
4355	*/
4356	for (i = 0; i < RT_ELEMENTS(g_aFunctions); i++)
4357	if (!strcmp(g_aFunctions[i].szName, pszSymbol))
4358	{
4359	pReq->u.Out.pfnSymbol = g_aFunctions[i].pfn;
4360	break;
4361	}
4362	}
4363	else
4364	{
4365	/*
4366	* Find the loader image.
4367	*/
4368	PSUPDRVLDRIMAGE pImage;
4369
4370	supdrvLdrLock(pDevExt);
4371
4372	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
4373	if (!strcmp(pImage->szName, pszModule))
4374	break;
4375	if (pImage && pImage->uState == SUP_IOCTL_LDR_LOAD)
4376	{
4377	/*
4378	* Search the symbol strings.
4379	*/
4380	const char *pchStrings = pImage->pachStrTab;
4381	PCSUPLDRSYM paSyms = pImage->paSymbols;
4382	for (i = 0; i < pImage->cSymbols; i++)
4383	{
4384	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4385	&& !memcmp(pchStrings + paSyms[i].offName, pszSymbol, cbSymbol))
4386	{
4387	/*
4388	* Found it! Calc the symbol address and add a reference to the module.
4389	*/
4390	pReq->u.Out.pfnSymbol = (PFNRT)((uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol);
4391	rc = supdrvLdrAddUsage(pSession, pImage);
4392	break;
4393	}
4394	}
4395	}
4396	else
4397	rc = pImage ? VERR_WRONG_ORDER : VERR_MODULE_NOT_FOUND;
4398
4399	supdrvLdrUnlock(pDevExt);
4400	}
4401	return rc;
4402	}
4403
4404
4405	/**
4406	* Updates the VMMR0 entry point pointers.
4407	*
4408	* @returns IPRT status code.
4409	* @param pDevExt Device globals.
4410	* @param pSession Session data.
4411	* @param pVMMR0 VMMR0 image handle.
4412	* @param pvVMMR0EntryInt VMMR0EntryInt address.
4413	* @param pvVMMR0EntryFast VMMR0EntryFast address.
4414	* @param pvVMMR0EntryEx VMMR0EntryEx address.
4415	* @remark Caller must own the loader mutex.
4416	*/
4417	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx)
4418	{
4419	int rc = VINF_SUCCESS;
4420	LogFlow(("supdrvLdrSetR0EP pvVMMR0=%p pvVMMR0EntryInt=%p\n", pvVMMR0, pvVMMR0EntryInt));
4421
4422
4423	/*
4424	* Check if not yet set.
4425	*/
4426	if (!pDevExt->pvVMMR0)
4427	{
4428	pDevExt->pvVMMR0 = pvVMMR0;
4429	pDevExt->pfnVMMR0EntryInt = pvVMMR0EntryInt;
4430	pDevExt->pfnVMMR0EntryFast = pvVMMR0EntryFast;
4431	pDevExt->pfnVMMR0EntryEx = pvVMMR0EntryEx;
4432	}
4433	else
4434	{
4435	/*
4436	* Return failure or success depending on whether the values match or not.
4437	*/
4438	if ( pDevExt->pvVMMR0 != pvVMMR0
4439	\|\| (void *)pDevExt->pfnVMMR0EntryInt != pvVMMR0EntryInt
4440	\|\| (void *)pDevExt->pfnVMMR0EntryFast != pvVMMR0EntryFast
4441	\|\| (void *)pDevExt->pfnVMMR0EntryEx != pvVMMR0EntryEx)
4442	{
4443	AssertMsgFailed(("SUP_IOCTL_LDR_SETR0EP: Already set pointing to a different module!\n"));
4444	rc = VERR_INVALID_PARAMETER;
4445	}
4446	}
4447	return rc;
4448	}
4449
4450
4451	/**
4452	* Unsets the VMMR0 entry point installed by supdrvLdrSetR0EP.
4453	*
4454	* @param pDevExt Device globals.
4455	*/
4456	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt)
4457	{
4458	pDevExt->pvVMMR0 = NULL;
4459	pDevExt->pfnVMMR0EntryInt = NULL;
4460	pDevExt->pfnVMMR0EntryFast = NULL;
4461	pDevExt->pfnVMMR0EntryEx = NULL;
4462	}
4463
4464
4465	/**
4466	* Adds a usage reference in the specified session of an image.
4467	*
4468	* Called while owning the loader semaphore.
4469	*
4470	* @returns VINF_SUCCESS on success and VERR_NO_MEMORY on failure.
4471	* @param pSession Session in question.
4472	* @param pImage Image which the session is using.
4473	*/
4474	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage)
4475	{
4476	PSUPDRVLDRUSAGE pUsage;
4477	LogFlow(("supdrvLdrAddUsage: pImage=%p\n", pImage));
4478
4479	/*
4480	* Referenced it already?
4481	*/
4482	pUsage = pSession->pLdrUsage;
4483	while (pUsage)
4484	{
4485	if (pUsage->pImage == pImage)
4486	{
4487	pUsage->cUsage++;
4488	return VINF_SUCCESS;
4489	}
4490	pUsage = pUsage->pNext;
4491	}
4492
4493	/*
4494	* Allocate new usage record.
4495	*/
4496	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
4497	AssertReturn(pUsage, /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_5);
4498	pUsage->cUsage = 1;
4499	pUsage->pImage = pImage;
4500	pUsage->pNext = pSession->pLdrUsage;
4501	pSession->pLdrUsage = pUsage;
4502	return VINF_SUCCESS;
4503	}
4504
4505
4506	/**
4507	* Frees a load image.
4508	*
4509	* @param pDevExt Pointer to device extension.
4510	* @param pImage Pointer to the image we're gonna free.
4511	* This image must exit!
4512	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
4513	*/
4514	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
4515	{
4516	PSUPDRVLDRIMAGE pImagePrev;
4517	LogFlow(("supdrvLdrFree: pImage=%p\n", pImage));
4518
4519	/* find it - arg. should've used doubly linked list. */
4520	Assert(pDevExt->pLdrImages);
4521	pImagePrev = NULL;
4522	if (pDevExt->pLdrImages != pImage)
4523	{
4524	pImagePrev = pDevExt->pLdrImages;
4525	while (pImagePrev->pNext != pImage)
4526	pImagePrev = pImagePrev->pNext;
4527	Assert(pImagePrev->pNext == pImage);
4528	}
4529
4530	/* unlink */
4531	if (pImagePrev)
4532	pImagePrev->pNext = pImage->pNext;
4533	else
4534	pDevExt->pLdrImages = pImage->pNext;
4535
4536	/* check if this is VMMR0.r0 unset its entry point pointers. */
4537	if (pDevExt->pvVMMR0 == pImage->pvImage)
4538	supdrvLdrUnsetVMMR0EPs(pDevExt);
4539
4540	/* check for objects with destructors in this image. (Shouldn't happen.) */
4541	if (pDevExt->pObjs)
4542	{
4543	unsigned cObjs = 0;
4544	PSUPDRVOBJ pObj;
4545	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
4546	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
4547	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4548	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4549	{
4550	pObj->pfnDestructor = NULL;
4551	cObjs++;
4552	}
4553	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
4554	if (cObjs)
4555	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
4556	}
4557
4558	/* call termination function if fully loaded. */
4559	if ( pImage->pfnModuleTerm
4560	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
4561	{
4562	LogFlow(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
4563	pImage->pfnModuleTerm();
4564	}
4565
4566	/* do native unload if appropriate. */
4567	if (pImage->fNative)
4568	supdrvOSLdrUnload(pDevExt, pImage);
4569
4570	/* free the image */
4571	pImage->cUsage = 0;
4572	pImage->pNext = 0;
4573	pImage->uState = SUP_IOCTL_LDR_FREE;
4574	RTMemExecFree(pImage->pvImageAlloc, pImage->cbImageBits + 31);
4575	pImage->pvImageAlloc = NULL;
4576	RTMemFree(pImage->pachStrTab);
4577	pImage->pachStrTab = NULL;
4578	RTMemFree(pImage->paSymbols);
4579	pImage->paSymbols = NULL;
4580	RTMemFree(pImage);
4581	}
4582
4583
4584	/**
4585	* Acquires the loader lock.
4586	*
4587	* @returns IPRT status code.
4588	* @param pDevExt The device extension.
4589	*/
4590	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt)
4591	{
4592	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4593	int rc = RTSemMutexRequest(pDevExt->mtxLdr, RT_INDEFINITE_WAIT);
4594	#else
4595	int rc = RTSemFastMutexRequest(pDevExt->mtxLdr);
4596	#endif
4597	AssertRC(rc);
4598	return rc;
4599	}
4600
4601
4602	/**
4603	* Releases the loader lock.
4604	*
4605	* @returns IPRT status code.
4606	* @param pDevExt The device extension.
4607	*/
4608	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt)
4609	{
4610	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4611	return RTSemMutexRelease(pDevExt->mtxLdr);
4612	#else
4613	return RTSemFastMutexRelease(pDevExt->mtxLdr);
4614	#endif
4615	}
4616
4617
4618	/**
4619	* Implements the service call request.
4620	*
4621	* @returns VBox status code.
4622	* @param pDevExt The device extension.
4623	* @param pSession The calling session.
4624	* @param pReq The request packet, valid.
4625	*/
4626	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq)
4627	{
4628	#if !defined(RT_OS_WINDOWS) \|\| defined(DEBUG)
4629	int rc;
4630
4631	/*
4632	* Find the module first in the module referenced by the calling session.
4633	*/
4634	rc = supdrvLdrLock(pDevExt);
4635	if (RT_SUCCESS(rc))
4636	{
4637	PFNSUPR0SERVICEREQHANDLER pfnServiceReqHandler = NULL;
4638	PSUPDRVLDRUSAGE pUsage;
4639
4640	for (pUsage = pSession->pLdrUsage; pUsage; pUsage = pUsage->pNext)
4641	if ( pUsage->pImage->pfnServiceReqHandler
4642	&& !strcmp(pUsage->pImage->szName, pReq->u.In.szName))
4643	{
4644	pfnServiceReqHandler = pUsage->pImage->pfnServiceReqHandler;
4645	break;
4646	}
4647	supdrvLdrUnlock(pDevExt);
4648
4649	if (pfnServiceReqHandler)
4650	{
4651	/*
4652	* Call it.
4653	*/
4654	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
4655	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, NULL);
4656	else
4657	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0]);
4658	}
4659	else
4660	rc = VERR_SUPDRV_SERVICE_NOT_FOUND;
4661	}
4662
4663	/* log it */
4664	if ( RT_FAILURE(rc)
4665	&& rc != VERR_INTERRUPTED
4666	&& rc != VERR_TIMEOUT)
4667	Log(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4668	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4669	else
4670	Log4(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4671	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4672	return rc;
4673	#else /* RT_OS_WINDOWS && !DEBUG */
4674	return VERR_NOT_IMPLEMENTED;
4675	#endif /* RT_OS_WINDOWS && !DEBUG */
4676	}
4677
4678
4679	/**
4680	* Implements the logger settings request.
4681	*
4682	* @returns VBox status code.
4683	* @param pDevExt The device extension.
4684	* @param pSession The caller's session.
4685	* @param pReq The request.
4686	*/
4687	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq)
4688	{
4689	const char *pszGroup = &pReq->u.In.szStrings[pReq->u.In.offGroups];
4690	const char *pszFlags = &pReq->u.In.szStrings[pReq->u.In.offFlags];
4691	const char *pszDest = &pReq->u.In.szStrings[pReq->u.In.offDestination];
4692	PRTLOGGER pLogger = NULL;
4693	int rc;
4694
4695	/*
4696	* Some further validation.
4697	*/
4698	switch (pReq->u.In.fWhat)
4699	{
4700	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4701	case SUPLOGGERSETTINGS_WHAT_CREATE:
4702	break;
4703
4704	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4705	if (pszGroup \|\| pszFlags \|\| *pszDest)
4706	return VERR_INVALID_PARAMETER;
4707	if (pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_RELEASE)
4708	return VERR_ACCESS_DENIED;
4709	break;
4710
4711	default:
4712	return VERR_INTERNAL_ERROR;
4713	}
4714
4715	/*
4716	* Get the logger.
4717	*/
4718	switch (pReq->u.In.fWhich)
4719	{
4720	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4721	pLogger = RTLogGetDefaultInstance();
4722	break;
4723
4724	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4725	pLogger = RTLogRelDefaultInstance();
4726	break;
4727
4728	default:
4729	return VERR_INTERNAL_ERROR;
4730	}
4731
4732	/*
4733	* Do the job.
4734	*/
4735	switch (pReq->u.In.fWhat)
4736	{
4737	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4738	if (pLogger)
4739	{
4740	rc = RTLogFlags(pLogger, pszFlags);
4741	if (RT_SUCCESS(rc))
4742	rc = RTLogGroupSettings(pLogger, pszGroup);
4743	NOREF(pszDest);
4744	}
4745	else
4746	rc = VERR_NOT_FOUND;
4747	break;
4748
4749	case SUPLOGGERSETTINGS_WHAT_CREATE:
4750	{
4751	if (pLogger)
4752	rc = VERR_ALREADY_EXISTS;
4753	else
4754	{
4755	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
4756
4757	rc = RTLogCreate(&pLogger,
4758	0 /* fFlags */,
4759	pszGroup,
4760	pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_DEBUG
4761	? "VBOX_LOG"
4762	: "VBOX_RELEASE_LOG",
4763	RT_ELEMENTS(s_apszGroups),
4764	s_apszGroups,
4765	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER,
4766	NULL);
4767	if (RT_SUCCESS(rc))
4768	{
4769	rc = RTLogFlags(pLogger, pszFlags);
4770	NOREF(pszDest);
4771	if (RT_SUCCESS(rc))
4772	{
4773	switch (pReq->u.In.fWhich)
4774	{
4775	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4776	pLogger = RTLogSetDefaultInstance(pLogger);
4777	break;
4778	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4779	pLogger = RTLogRelSetDefaultInstance(pLogger);
4780	break;
4781	}
4782	}
4783	RTLogDestroy(pLogger);
4784	}
4785	}
4786	break;
4787	}
4788
4789	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4790	switch (pReq->u.In.fWhich)
4791	{
4792	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4793	pLogger = RTLogSetDefaultInstance(NULL);
4794	break;
4795	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4796	pLogger = RTLogRelSetDefaultInstance(NULL);
4797	break;
4798	}
4799	rc = RTLogDestroy(pLogger);
4800	break;
4801
4802	default:
4803	{
4804	rc = VERR_INTERNAL_ERROR;
4805	break;
4806	}
4807	}
4808
4809	return rc;
4810	}
4811
4812
4813	/**
4814	* Creates the GIP.
4815	*
4816	* @returns VBox status code.
4817	* @param pDevExt Instance data. GIP stuff may be updated.
4818	*/
4819	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
4820	{
4821	PSUPGLOBALINFOPAGE pGip;
4822	RTHCPHYS HCPhysGip;
4823	uint32_t u32SystemResolution;
4824	uint32_t u32Interval;
4825	unsigned cCpus;
4826	int rc;
4827
4828
4829	LogFlow(("supdrvGipCreate:\n"));
4830
4831	/* assert order */
4832	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
4833	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
4834	Assert(!pDevExt->pGipTimer);
4835
4836	/*
4837	* Check the CPU count.
4838	*/
4839	cCpus = RTMpGetArraySize();
4840	if ( cCpus > RTCPUSET_MAX_CPUS
4841	\|\| cCpus > 256 /ApicId is used for the mappings/)
4842	{
4843	SUPR0Printf("VBoxDrv: Too many CPUs (%u) for the GIP (max %u)\n", cCpus, RT_MIN(RTCPUSET_MAX_CPUS, 256));
4844	return VERR_TOO_MANY_CPUS;
4845	}
4846
4847	/*
4848	* Allocate a contiguous set of pages with a default kernel mapping.
4849	*/
4850	rc = RTR0MemObjAllocCont(&pDevExt->GipMemObj, RT_UOFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), false /fExecutable/);
4851	if (RT_FAILURE(rc))
4852	{
4853	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
4854	return rc;
4855	}
4856	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
4857	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
4858
4859	/*
4860	* Find a reasonable update interval and initialize the structure.
4861	*/
4862	u32Interval = u32SystemResolution = RTTimerGetSystemGranularity();
4863	while (u32Interval < 10000000 /* 10 ms */)
4864	u32Interval += u32SystemResolution;
4865
4866	supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), 1000000000 / u32Interval /=Hz/, cCpus);
4867
4868	/*
4869	* Create the timer.
4870	* If CPU_ALL isn't supported we'll have to fall back to synchronous mode.
4871	*/
4872	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
4873	{
4874	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, RTTIMER_FLAGS_CPU_ALL, supdrvGipAsyncTimer, pDevExt);
4875	if (rc == VERR_NOT_SUPPORTED)
4876	{
4877	OSDBGPRINT(("supdrvGipCreate: omni timer not supported, falling back to synchronous mode\n"));
4878	pGip->u32Mode = SUPGIPMODE_SYNC_TSC;
4879	}
4880	}
4881	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
4882	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0, supdrvGipSyncTimer, pDevExt);
4883	if (RT_SUCCESS(rc))
4884	{
4885	rc = RTMpNotificationRegister(supdrvGipMpEvent, pDevExt);
4886	if (RT_SUCCESS(rc))
4887	{
4888	rc = RTMpOnAll(supdrvGipInitOnCpu, pDevExt, pGip);
4889	if (RT_SUCCESS(rc))
4890	{
4891	/*
4892	* We're good.
4893	*/
4894	Log(("supdrvGipCreate: %u ns interval.\n", u32Interval));
4895	g_pSUPGlobalInfoPage = pGip;
4896	return VINF_SUCCESS;
4897	}
4898
4899	OSDBGPRINT(("supdrvGipCreate: RTMpOnAll failed with rc=%Rrc\n", rc));
4900	RTMpNotificationDeregister(supdrvGipMpEvent, pDevExt);
4901
4902	}
4903	else
4904	OSDBGPRINT(("supdrvGipCreate: failed to register MP event notfication. rc=%Rrc\n", rc));
4905	}
4906	else
4907	{
4908	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %u ns interval. rc=%Rrc\n", u32Interval, rc));
4909	Assert(!pDevExt->pGipTimer);
4910	}
4911	supdrvGipDestroy(pDevExt);
4912	return rc;
4913	}
4914
4915
4916	/**
4917	* Terminates the GIP.
4918	*
4919	* @param pDevExt Instance data. GIP stuff may be updated.
4920	*/
4921	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
4922	{
4923	int rc;
4924	#ifdef DEBUG_DARWIN_GIP
4925	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
4926	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
4927	pDevExt->pGipTimer, pDevExt->GipMemObj));
4928	#endif
4929
4930	/*
4931	* Invalid the GIP data.
4932	*/
4933	if (pDevExt->pGip)
4934	{
4935	supdrvGipTerm(pDevExt->pGip);
4936	pDevExt->pGip = NULL;
4937	}
4938	g_pSUPGlobalInfoPage = NULL;
4939
4940	/*
4941	* Destroy the timer and free the GIP memory object.
4942	*/
4943	if (pDevExt->pGipTimer)
4944	{
4945	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
4946	pDevExt->pGipTimer = NULL;
4947	}
4948
4949	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
4950	{
4951	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
4952	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
4953	}
4954
4955	/*
4956	* Finally, make sure we've release the system timer resolution request
4957	* if one actually succeeded and is still pending.
4958	*/
4959	if (pDevExt->u32SystemTimerGranularityGrant)
4960	{
4961	rc = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
4962	pDevExt->u32SystemTimerGranularityGrant = 0;
4963	}
4964	}
4965
4966
4967	/**
4968	* Timer callback function sync GIP mode.
4969	* @param pTimer The timer.
4970	* @param pvUser The device extension.
4971	*/
4972	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
4973	{
4974	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
4975	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4976	uint64_t u64TSC = ASMReadTSC();
4977	uint64_t NanoTS = RTTimeSystemNanoTS();
4978
4979	supdrvGipUpdate(pDevExt->pGip, NanoTS, u64TSC, NIL_RTCPUID, iTick);
4980
4981	ASMSetFlags(fOldFlags);
4982	}
4983
4984
4985	/**
4986	* Timer callback function for async GIP mode.
4987	* @param pTimer The timer.
4988	* @param pvUser The device extension.
4989	*/
4990	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
4991	{
4992	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
4993	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4994	RTCPUID idCpu = RTMpCpuId();
4995	uint64_t u64TSC = ASMReadTSC();
4996	uint64_t NanoTS = RTTimeSystemNanoTS();
4997
4998	/** @todo reset the transaction number and whatnot when iTick == 1. */
4999	if (pDevExt->idGipMaster == idCpu)
5000	supdrvGipUpdate(pDevExt->pGip, NanoTS, u64TSC, idCpu, iTick);
5001	else
5002	supdrvGipUpdatePerCpu(pDevExt->pGip, NanoTS, u64TSC, idCpu, ASMGetApicId(), iTick);
5003
5004	ASMSetFlags(fOldFlags);
5005	}
5006
5007
5008	/**
5009	* Finds our (@a idCpu) entry, or allocates a new one if not found.
5010	*
5011	* @returns Index of the CPU in the cache set.
5012	* @param pGip The GIP.
5013	* @param idCpu The CPU ID.
5014	*/
5015	static uint32_t supdrvGipCpuIndexFromCpuId(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
5016	{
5017	uint32_t i, cTries;
5018
5019	/*
5020	* ASSUMES that CPU IDs are constant.
5021	*/
5022	for (i = 0; i < pGip->cCpus; i++)
5023	if (pGip->aCPUs[i].idCpu == idCpu)
5024	return i;
5025
5026	cTries = 0;
5027	do
5028	{
5029	for (i = 0; i < pGip->cCpus; i++)
5030	{
5031	bool fRc;
5032	ASMAtomicCmpXchgSize(&pGip->aCPUs[i].idCpu, idCpu, NIL_RTCPUID, fRc);
5033	if (fRc)
5034	return i;
5035	}
5036	} while (cTries++ < 32);
5037	AssertReleaseFailed();
5038	return i - 1;
5039	}
5040
5041
5042	/**
5043	* The calling CPU should be accounted as online, update GIP accordingly.
5044	*
5045	* This is used by supdrvGipMpEvent as well as the supdrvGipCreate.
5046	*
5047	* @param pGip The GIP.
5048	* @param idCpu The CPU ID.
5049	*/
5050	static void supdrvGipMpEventOnline(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
5051	{
5052	int iCpuSet = 0;
5053	uint16_t idApic = UINT16_MAX;
5054	uint32_t i = 0;
5055	uint64_t u64NanoTS = 0;
5056
5057	AssertRelease(idCpu == RTMpCpuId());
5058	Assert(pGip->cPossibleCpus == RTMpGetCount());
5059
5060	/*
5061	* Update the globals.
5062	*/
5063	ASMAtomicWriteU16(&pGip->cPresentCpus, RTMpGetPresentCount());
5064	ASMAtomicWriteU16(&pGip->cOnlineCpus, RTMpGetOnlineCount());
5065	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5066	if (iCpuSet >= 0)
5067	{
5068	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5069	RTCpuSetAddByIndex(&pGip->OnlineCpuSet, iCpuSet);
5070	RTCpuSetAddByIndex(&pGip->PresentCpuSet, iCpuSet);
5071	}
5072
5073	/*
5074	* Update the entry.
5075	*/
5076	u64NanoTS = RTTimeSystemNanoTS() - pGip->u32UpdateIntervalNS;
5077	i = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
5078	supdrvGipInitCpu(pGip, &pGip->aCPUs[i], u64NanoTS);
5079	idApic = ASMGetApicId();
5080	ASMAtomicWriteU16(&pGip->aCPUs[i].idApic, idApic);
5081	ASMAtomicWriteS16(&pGip->aCPUs[i].iCpuSet, (int16_t)iCpuSet);
5082	ASMAtomicWriteSize(&pGip->aCPUs[i].idCpu, idCpu);
5083
5084	/*
5085	* Update the APIC ID and CPU set index mappings.
5086	*/
5087	ASMAtomicWriteU16(&pGip->aiCpuFromApicId[idApic], i);
5088	ASMAtomicWriteU16(&pGip->aiCpuFromCpuSetIdx[iCpuSet], i);
5089
5090	/* commit it */
5091	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_ONLINE);
5092	}
5093
5094
5095	/**
5096	* The CPU should be accounted as offline, update the GIP accordingly.
5097	*
5098	* This is used by supdrvGipMpEvent.
5099	*
5100	* @param pGip The GIP.
5101	* @param idCpu The CPU ID.
5102	*/
5103	static void supdrvGipMpEventOffline(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
5104	{
5105	int iCpuSet;
5106	unsigned i;
5107
5108	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5109	AssertReturnVoid(iCpuSet >= 0);
5110
5111	i = pGip->aiCpuFromCpuSetIdx[iCpuSet];
5112	AssertReturnVoid(i < pGip->cCpus);
5113	AssertReturnVoid(pGip->aCPUs[i].idCpu == idCpu);
5114
5115	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5116	RTCpuSetDelByIndex(&pGip->OnlineCpuSet, iCpuSet);
5117
5118	/* commit it */
5119	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_OFFLINE);
5120	}
5121
5122
5123	/**
5124	* Multiprocessor event notification callback.
5125	*
5126	* This is used to make sure that the GIP master gets passed on to
5127	* another CPU. It also updates the associated CPU data.
5128	*
5129	* @param enmEvent The event.
5130	* @param idCpu The cpu it applies to.
5131	* @param pvUser Pointer to the device extension.
5132	*
5133	* @remarks This function -must- fire on the newly online'd CPU for the
5134	* RTMPEVENT_ONLINE case and can fire on any CPU for the
5135	* RTMPEVENT_OFFLINE case.
5136	*/
5137	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser)
5138	{
5139	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5140	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5141
5142	AssertRelease(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
5143
5144	/*
5145	* Update the GIP CPU data.
5146	*/
5147	if (pGip)
5148	{
5149	switch (enmEvent)
5150	{
5151	case RTMPEVENT_ONLINE:
5152	AssertRelease(idCpu == RTMpCpuId());
5153	supdrvGipMpEventOnline(pGip, idCpu);
5154	break;
5155	case RTMPEVENT_OFFLINE:
5156	supdrvGipMpEventOffline(pGip, idCpu);
5157	break;
5158
5159	}
5160	}
5161
5162	/*
5163	* Make sure there is a master GIP.
5164	*/
5165	if (enmEvent == RTMPEVENT_OFFLINE)
5166	{
5167	RTCPUID idGipMaster = ASMAtomicReadU32(&pDevExt->idGipMaster);
5168	if (idGipMaster == idCpu)
5169	{
5170	/*
5171	* Find a new GIP master.
5172	*/
5173	bool fIgnored;
5174	unsigned i;
5175	RTCPUID idNewGipMaster = NIL_RTCPUID;
5176	RTCPUSET OnlineCpus;
5177	RTMpGetOnlineSet(&OnlineCpus);
5178
5179	for (i = 0; i < RTCPUSET_MAX_CPUS; i++)
5180	{
5181	RTCPUID idCurCpu = RTMpCpuIdFromSetIndex(i);
5182	if ( RTCpuSetIsMember(&OnlineCpus, idCurCpu)
5183	&& idCurCpu != idGipMaster)
5184	{
5185	idNewGipMaster = idCurCpu;
5186	break;
5187	}
5188	}
5189
5190	Log(("supdrvGipMpEvent: Gip master %#lx -> %#lx\n", (long)idGipMaster, (long)idNewGipMaster));
5191	ASMAtomicCmpXchgSize(&pDevExt->idGipMaster, idNewGipMaster, idGipMaster, fIgnored);
5192	NOREF(fIgnored);
5193	}
5194	}
5195	}
5196
5197
5198	/**
5199	* Callback used by supdrvDetermineAsyncTSC to read the TSC on a CPU.
5200	*
5201	* @param idCpu Ignored.
5202	* @param pvUser1 Where to put the TSC.
5203	* @param pvUser2 Ignored.
5204	*/
5205	static DECLCALLBACK(void) supdrvDetermineAsyncTscWorker(RTCPUID idCpu, void pvUser1, void pvUser2)
5206	{
5207	#if 1
5208	ASMAtomicWriteU64((uint64_t volatile *)pvUser1, ASMReadTSC());
5209	#else
5210	(uint64_t )pvUser1 = ASMReadTSC();
5211	#endif
5212	}
5213
5214
5215	/**
5216	* Determine if Async GIP mode is required because of TSC drift.
5217	*
5218	* When using the default/normal timer code it is essential that the time stamp counter
5219	* (TSC) runs never backwards, that is, a read operation to the counter should return
5220	* a bigger value than any previous read operation. This is guaranteed by the latest
5221	* AMD CPUs and by newer Intel CPUs which never enter the C2 state (P4). In any other
5222	* case we have to choose the asynchronous timer mode.
5223	*
5224	* @param poffMin Pointer to the determined difference between different cores.
5225	* @return false if the time stamp counters appear to be synchronized, true otherwise.
5226	*/
5227	static bool supdrvDetermineAsyncTsc(uint64_t *poffMin)
5228	{
5229	/*
5230	* Just iterate all the cpus 8 times and make sure that the TSC is
5231	* ever increasing. We don't bother taking TSC rollover into account.
5232	*/
5233	int iEndCpu = RTMpGetArraySize();
5234	int iCpu;
5235	int cLoops = 8;
5236	bool fAsync = false;
5237	int rc = VINF_SUCCESS;
5238	uint64_t offMax = 0;
5239	uint64_t offMin = ~(uint64_t)0;
5240	uint64_t PrevTsc = ASMReadTSC();
5241
5242	while (cLoops-- > 0)
5243	{
5244	for (iCpu = 0; iCpu < iEndCpu; iCpu++)
5245	{
5246	uint64_t CurTsc;
5247	rc = RTMpOnSpecific(RTMpCpuIdFromSetIndex(iCpu), supdrvDetermineAsyncTscWorker, &CurTsc, NULL);
5248	if (RT_SUCCESS(rc))
5249	{
5250	if (CurTsc <= PrevTsc)
5251	{
5252	fAsync = true;
5253	offMin = offMax = PrevTsc - CurTsc;
5254	Log(("supdrvDetermineAsyncTsc: iCpu=%d cLoops=%d CurTsc=%llx PrevTsc=%llx\n",
5255	iCpu, cLoops, CurTsc, PrevTsc));
5256	break;
5257	}
5258
5259	/* Gather statistics (except the first time). */
5260	if (iCpu != 0 \|\| cLoops != 7)
5261	{
5262	uint64_t off = CurTsc - PrevTsc;
5263	if (off < offMin)
5264	offMin = off;
5265	if (off > offMax)
5266	offMax = off;
5267	Log2(("%d/%d: off=%llx\n", cLoops, iCpu, off));
5268	}
5269
5270	/* Next */
5271	PrevTsc = CurTsc;
5272	}
5273	else if (rc == VERR_NOT_SUPPORTED)
5274	break;
5275	else
5276	AssertMsg(rc == VERR_CPU_NOT_FOUND \|\| rc == VERR_CPU_OFFLINE, ("%d\n", rc));
5277	}
5278
5279	/* broke out of the loop. */
5280	if (iCpu < iEndCpu)
5281	break;
5282	}
5283
5284	poffMin = offMin; / Almost RTMpOnSpecific profiling. */
5285	Log(("supdrvDetermineAsyncTsc: returns %d; iEndCpu=%d rc=%d offMin=%llx offMax=%llx\n",
5286	fAsync, iEndCpu, rc, offMin, offMax));
5287	#if !defined(RT_OS_SOLARIS) && !defined(RT_OS_OS2) && !defined(RT_OS_WINDOWS)
5288	OSDBGPRINT(("vboxdrv: fAsync=%d offMin=%#lx offMax=%#lx\n", fAsync, (long)offMin, (long)offMax));
5289	#endif
5290	return fAsync;
5291	}
5292
5293
5294	/**
5295	* Determine the GIP TSC mode.
5296	*
5297	* @returns The most suitable TSC mode.
5298	* @param pDevExt Pointer to the device instance data.
5299	*/
5300	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt)
5301	{
5302	/*
5303	* On SMP we're faced with two problems:
5304	* (1) There might be a skew between the CPU, so that cpu0
5305	* returns a TSC that is slightly different from cpu1.
5306	* (2) Power management (and other things) may cause the TSC
5307	* to run at a non-constant speed, and cause the speed
5308	* to be different on the cpus. This will result in (1).
5309	*
5310	* So, on SMP systems we'll have to select the ASYNC update method
5311	* if there are symptoms of these problems.
5312	*/
5313	if (RTMpGetCount() > 1)
5314	{
5315	uint32_t uEAX, uEBX, uECX, uEDX;
5316	uint64_t u64DiffCoresIgnored;
5317
5318	/* Permit the user and/or the OS specific bits to force async mode. */
5319	if (supdrvOSGetForcedAsyncTscMode(pDevExt))
5320	return SUPGIPMODE_ASYNC_TSC;
5321
5322	/* Try check for current differences between the cpus. */
5323	if (supdrvDetermineAsyncTsc(&u64DiffCoresIgnored))
5324	return SUPGIPMODE_ASYNC_TSC;
5325
5326	/*
5327	* If the CPU supports power management and is an AMD one we
5328	* won't trust it unless it has the TscInvariant bit is set.
5329	*/
5330	/* Check for "AuthenticAMD" */
5331	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
5332	if ( uEAX >= 1
5333	&& uEBX == X86_CPUID_VENDOR_AMD_EBX
5334	&& uECX == X86_CPUID_VENDOR_AMD_ECX
5335	&& uEDX == X86_CPUID_VENDOR_AMD_EDX)
5336	{
5337	/* Check for APM support and that TscInvariant is cleared. */
5338	ASMCpuId(0x80000000, &uEAX, &uEBX, &uECX, &uEDX);
5339	if (uEAX >= 0x80000007)
5340	{
5341	ASMCpuId(0x80000007, &uEAX, &uEBX, &uECX, &uEDX);
5342	if ( !(uEDX & RT_BIT(8))/* TscInvariant */
5343	&& (uEDX & 0x3e)) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
5344	return SUPGIPMODE_ASYNC_TSC;
5345	}
5346	}
5347	}
5348	return SUPGIPMODE_SYNC_TSC;
5349	}
5350
5351
5352	/**
5353	* Initializes per-CPU GIP information.
5354	*
5355	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5356	* @param pCpu Pointer to which GIP CPU to initalize.
5357	* @param u64NanoTS The current nanosecond timestamp.
5358	*/
5359	static void supdrvGipInitCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pCpu, uint64_t u64NanoTS)
5360	{
5361	pCpu->u32TransactionId = 2;
5362	pCpu->u64NanoTS = u64NanoTS;
5363	pCpu->u64TSC = ASMReadTSC();
5364
5365	ASMAtomicWriteSize(&pCpu->enmState, SUPGIPCPUSTATE_INVALID);
5366	ASMAtomicWriteSize(&pCpu->idCpu, NIL_RTCPUID);
5367	ASMAtomicWriteS16(&pCpu->iCpuSet, -1);
5368	ASMAtomicWriteU16(&pCpu->idApic, UINT16_MAX);
5369
5370	/*
5371	* We don't know the following values until we've executed updates.
5372	* So, we'll just pretend it's a 4 GHz CPU and adjust the history it on
5373	* the 2nd timer callout.
5374	*/
5375	pCpu->u64CpuHz = _4G + 1; /* tstGIP-2 depends on this. */
5376	pCpu->u32UpdateIntervalTSC
5377	= pCpu->au32TSCHistory[0]
5378	= pCpu->au32TSCHistory[1]
5379	= pCpu->au32TSCHistory[2]
5380	= pCpu->au32TSCHistory[3]
5381	= pCpu->au32TSCHistory[4]
5382	= pCpu->au32TSCHistory[5]
5383	= pCpu->au32TSCHistory[6]
5384	= pCpu->au32TSCHistory[7]
5385	= (uint32_t)(_4G / pGip->u32UpdateHz);
5386	}
5387
5388
5389	/**
5390	* Initializes the GIP data.
5391	*
5392	* @param pDevExt Pointer to the device instance data.
5393	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5394	* @param HCPhys The physical address of the GIP.
5395	* @param u64NanoTS The current nanosecond timestamp.
5396	* @param uUpdateHz The update frequency.
5397	* @param cCpus The CPU count.
5398	*/
5399	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
5400	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned cCpus)
5401	{
5402	size_t const cbGip = RT_ALIGN_Z(RT_OFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), PAGE_SIZE);
5403	unsigned i;
5404	#ifdef DEBUG_DARWIN_GIP
5405	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5406	#else
5407	LogFlow(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5408	#endif
5409
5410	/*
5411	* Initialize the structure.
5412	*/
5413	memset(pGip, 0, cbGip);
5414	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
5415	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
5416	pGip->u32Mode = supdrvGipDeterminTscMode(pDevExt);
5417	pGip->cCpus = (uint16_t)cCpus;
5418	pGip->cPages = (uint16_t)(cbGip / PAGE_SIZE);
5419	pGip->u32UpdateHz = uUpdateHz;
5420	pGip->u32UpdateIntervalNS = 1000000000 / uUpdateHz;
5421	pGip->u64NanoTSLastUpdateHz = u64NanoTS;
5422	RTCpuSetEmpty(&pGip->OnlineCpuSet);
5423	RTCpuSetEmpty(&pGip->PresentCpuSet);
5424	RTMpGetSet(&pGip->PossibleCpuSet);
5425	pGip->cOnlineCpus = RTMpGetOnlineCount();
5426	pGip->cPresentCpus = RTMpGetPresentCount();
5427	pGip->cPossibleCpus = RTMpGetCount();
5428	pGip->idCpuMax = RTMpGetMaxCpuId();
5429	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromApicId); i++)
5430	pGip->aiCpuFromApicId[i] = 0;
5431	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx); i++)
5432	pGip->aiCpuFromCpuSetIdx[i] = UINT16_MAX;
5433
5434	for (i = 0; i < cCpus; i++)
5435	supdrvGipInitCpu(pGip, &pGip->aCPUs[i], u64NanoTS);
5436
5437	/*
5438	* Link it to the device extension.
5439	*/
5440	pDevExt->pGip = pGip;
5441	pDevExt->HCPhysGip = HCPhys;
5442	pDevExt->cGipUsers = 0;
5443	}
5444
5445
5446	/**
5447	* On CPU initialization callback for RTMpOnAll.
5448	*
5449	* @param idCpu The CPU ID.
5450	* @param pvUser1 The device extension.
5451	* @param pvUser2 The GIP.
5452	*/
5453	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2)
5454	{
5455	/* This is good enough, even though it will update some of the globals a
5456	bit to much. */
5457	supdrvGipMpEventOnline((PSUPGLOBALINFOPAGE)pvUser2, idCpu);
5458	}
5459
5460
5461	/**
5462	* Invalidates the GIP data upon termination.
5463	*
5464	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5465	*/
5466	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
5467	{
5468	unsigned i;
5469	pGip->u32Magic = 0;
5470	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
5471	{
5472	pGip->aCPUs[i].u64NanoTS = 0;
5473	pGip->aCPUs[i].u64TSC = 0;
5474	pGip->aCPUs[i].iTSCHistoryHead = 0;
5475	}
5476	}
5477
5478
5479	/**
5480	* Worker routine for supdrvGipUpdate and supdrvGipUpdatePerCpu that
5481	* updates all the per cpu data except the transaction id.
5482	*
5483	* @param pGip The GIP.
5484	* @param pGipCpu Pointer to the per cpu data.
5485	* @param u64NanoTS The current time stamp.
5486	* @param u64TSC The current TSC.
5487	* @param iTick The current timer tick.
5488	*/
5489	static void supdrvGipDoUpdateCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick)
5490	{
5491	uint64_t u64TSCDelta;
5492	uint32_t u32UpdateIntervalTSC;
5493	uint32_t u32UpdateIntervalTSCSlack;
5494	unsigned iTSCHistoryHead;
5495	uint64_t u64CpuHz;
5496	uint32_t u32TransactionId;
5497
5498	/* Delta between this and the previous update. */
5499	ASMAtomicUoWriteU32(&pGipCpu->u32PrevUpdateIntervalNS, (uint32_t)(u64NanoTS - pGipCpu->u64NanoTS));
5500
5501	/*
5502	* Update the NanoTS.
5503	*/
5504	ASMAtomicWriteU64(&pGipCpu->u64NanoTS, u64NanoTS);
5505
5506	/*
5507	* Calc TSC delta.
5508	*/
5509	/** @todo validate the NanoTS delta, don't trust the OS to call us when it should... */
5510	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
5511	ASMAtomicWriteU64(&pGipCpu->u64TSC, u64TSC);
5512
5513	if (u64TSCDelta >> 32)
5514	{
5515	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
5516	pGipCpu->cErrors++;
5517	}
5518
5519	/*
5520	* On the 2nd and 3rd callout, reset the history with the current TSC
5521	* interval since the values entered by supdrvGipInit are totally off.
5522	* The interval on the 1st callout completely unreliable, the 2nd is a bit
5523	* better, while the 3rd should be most reliable.
5524	*/
5525	u32TransactionId = pGipCpu->u32TransactionId;
5526	if (RT_UNLIKELY( ( u32TransactionId == 5
5527	\|\| u32TransactionId == 7)
5528	&& ( iTick == 2
5529	\|\| iTick == 3) ))
5530	{
5531	unsigned i;
5532	for (i = 0; i < RT_ELEMENTS(pGipCpu->au32TSCHistory); i++)
5533	ASMAtomicUoWriteU32(&pGipCpu->au32TSCHistory[i], (uint32_t)u64TSCDelta);
5534	}
5535
5536	/*
5537	* TSC History.
5538	*/
5539	Assert(RT_ELEMENTS(pGipCpu->au32TSCHistory) == 8);
5540	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
5541	ASMAtomicWriteU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
5542	ASMAtomicWriteU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
5543
5544	/*
5545	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
5546	*/
5547	if (pGip->u32UpdateHz >= 1000)
5548	{
5549	uint32_t u32;
5550	u32 = pGipCpu->au32TSCHistory[0];
5551	u32 += pGipCpu->au32TSCHistory[1];
5552	u32 += pGipCpu->au32TSCHistory[2];
5553	u32 += pGipCpu->au32TSCHistory[3];
5554	u32 >>= 2;
5555	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
5556	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
5557	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
5558	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
5559	u32UpdateIntervalTSC >>= 2;
5560	u32UpdateIntervalTSC += u32;
5561	u32UpdateIntervalTSC >>= 1;
5562
5563	/* Value chosen for a 2GHz Athlon64 running linux 2.6.10/11, . */
5564	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
5565	}
5566	else if (pGip->u32UpdateHz >= 90)
5567	{
5568	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5569	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
5570	u32UpdateIntervalTSC >>= 1;
5571
5572	/* value chosen on a 2GHz thinkpad running windows */
5573	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
5574	}
5575	else
5576	{
5577	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5578
5579	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
5580	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
5581	}
5582	ASMAtomicWriteU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
5583
5584	/*
5585	* CpuHz.
5586	*/
5587	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, pGip->u32UpdateHz);
5588	ASMAtomicWriteU64(&pGipCpu->u64CpuHz, u64CpuHz);
5589	}
5590
5591
5592	/**
5593	* Updates the GIP.
5594	*
5595	* @param pGip Pointer to the GIP.
5596	* @param u64NanoTS The current nanosecond timesamp.
5597	* @param u64TSC The current TSC timesamp.
5598	* @param idCpu The CPU ID.
5599	* @param iTick The current timer tick.
5600	*/
5601	static void supdrvGipUpdate(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick)
5602	{
5603	/*
5604	* Determine the relevant CPU data.
5605	*/
5606	PSUPGIPCPU pGipCpu;
5607	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
5608	pGipCpu = &pGip->aCPUs[0];
5609	else
5610	{
5611	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
5612	if (RT_UNLIKELY(iCpu >= pGip->cCpus))
5613	return;
5614	pGipCpu = &pGip->aCPUs[iCpu];
5615	if (RT_UNLIKELY(pGipCpu->idCpu != idCpu))
5616	return;
5617	}
5618
5619	/*
5620	* Start update transaction.
5621	*/
5622	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5623	{
5624	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
5625	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5626	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5627	pGipCpu->cErrors++;
5628	return;
5629	}
5630
5631	/*
5632	* Recalc the update frequency every 0x800th time.
5633	*/
5634	if (!(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
5635	{
5636	if (pGip->u64NanoTSLastUpdateHz)
5637	{
5638	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
5639	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
5640	uint32_t u32UpdateHz = (uint32_t)((UINT64_C(1000000000) * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
5641	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
5642	{
5643	ASMAtomicWriteU32(&pGip->u32UpdateHz, u32UpdateHz);
5644	ASMAtomicWriteU32(&pGip->u32UpdateIntervalNS, 1000000000 / u32UpdateHz);
5645	}
5646	#endif
5647	}
5648	ASMAtomicWriteU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS);
5649	}
5650
5651	/*
5652	* Update the data.
5653	*/
5654	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS, u64TSC, iTick);
5655
5656	/*
5657	* Complete transaction.
5658	*/
5659	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5660	}
5661
5662
5663	/**
5664	* Updates the per cpu GIP data for the calling cpu.
5665	*
5666	* @param pGip Pointer to the GIP.
5667	* @param u64NanoTS The current nanosecond timesamp.
5668	* @param u64TSC The current TSC timesamp.
5669	* @param idCpu The CPU ID.
5670	* @param idApic The APIC id for the CPU index.
5671	* @param iTick The current timer tick.
5672	*/
5673	static void supdrvGipUpdatePerCpu(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC,
5674	RTCPUID idCpu, uint8_t idApic, uint64_t iTick)
5675	{
5676	uint32_t iCpu;
5677
5678	/*
5679	* Avoid a potential race when a CPU online notification doesn't fire on
5680	* the onlined CPU but the tick creeps in before the event notification is
5681	* run.
5682	*/
5683	if (RT_UNLIKELY(iTick == 1))
5684	{
5685	iCpu = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
5686	if (pGip->aCPUs[iCpu].enmState == SUPGIPCPUSTATE_OFFLINE)
5687	supdrvGipMpEventOnline(pGip, idCpu);
5688	}
5689
5690	iCpu = pGip->aiCpuFromApicId[idApic];
5691	if (RT_LIKELY(iCpu < pGip->cCpus))
5692	{
5693	PSUPGIPCPU pGipCpu = &pGip->aCPUs[iCpu];
5694	if (pGipCpu->idCpu == idCpu)
5695	{
5696	/*
5697	* Start update transaction.
5698	*/
5699	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5700	{
5701	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5702	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5703	pGipCpu->cErrors++;
5704	return;
5705	}
5706
5707	/*
5708	* Update the data.
5709	*/
5710	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS, u64TSC, iTick);
5711
5712	/*
5713	* Complete transaction.
5714	*/
5715	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5716	}
5717	}
5718	}
5719

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

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDrv.c@ 40611

Download in other formats: