SUPDrv.c@ 36540

Last change on this file since 36540 was 36408, checked in by vboxsync, 14 years ago
log rotation review and adjustments: Don't delete any excess files if log roation is disabled - we don't know what these files might be. Moved RTLOGGERFILE into log.c. Keep RTLogCreate simple, anyone needing rotation can use RTLogCreateEx[V]. Made RTLogGetDestinations produce the log rotation bits.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 212.8 KB

Line
1	/* $Revision: 36408 $ */
2	/** @file
3	* VBoxDrv - The VirtualBox Support Driver - Common code.
4	*/
5
6	/*
7	* Copyright (C) 2006-2011 Oracle Corporation
8	*
9	* This file is part of VirtualBox Open Source Edition (OSE), as
10	* available from http://www.virtualbox.org. This file is free software;
11	* you can redistribute it and/or modify it under the terms of the GNU
12	* General Public License (GPL) as published by the Free Software
13	* Foundation, in version 2 as it comes in the "COPYING" file of the
14	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16	*
17	* The contents of this file may alternatively be used under the terms
18	* of the Common Development and Distribution License Version 1.0
19	* (CDDL) only, as it comes in the "COPYING.CDDL" file of the
20	* VirtualBox OSE distribution, in which case the provisions of the
21	* CDDL are applicable instead of those of the GPL.
22	*
23	* You may elect to license modified versions of this file under the
24	* terms and conditions of either the GPL or the CDDL or both.
25	*/
26
27	/*******************************************************************************
28	* Header Files *
29	*******************************************************************************/
30	#define LOG_GROUP LOG_GROUP_SUP_DRV
31	#define SUPDRV_AGNOSTIC
32	#include "SUPDrvInternal.h"
33	#ifndef PAGE_SHIFT
34	# include <iprt/param.h>
35	#endif
36	#include <iprt/asm.h>
37	#include <iprt/asm-amd64-x86.h>
38	#include <iprt/asm-math.h>
39	#include <iprt/cpuset.h>
40	#include <iprt/handletable.h>
41	#include <iprt/mem.h>
42	#include <iprt/mp.h>
43	#include <iprt/power.h>
44	#include <iprt/process.h>
45	#include <iprt/semaphore.h>
46	#include <iprt/spinlock.h>
47	#include <iprt/thread.h>
48	#include <iprt/uuid.h>
49	#include <iprt/net.h>
50	#include <iprt/crc.h>
51	#include <iprt/string.h>
52	#include <iprt/timer.h>
53	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
54	# include <iprt/rand.h>
55	# include <iprt/path.h>
56	#endif
57
58	#include <VBox/param.h>
59	#include <VBox/log.h>
60	#include <VBox/err.h>
61	#include <VBox/vmm/hwacc_svm.h>
62	#include <VBox/vmm/hwacc_vmx.h>
63	#include <VBox/x86.h>
64
65	#ifdef VBOX_WITH_DTRACE
66	# include "SUPDrv-dtrace.h"
67	#else
68	# define VBOXDRV_SUPDRV_SESSION_CREATE(pvSession, fUser) do { } while (0)
69	# define VBOXDRV_SUPDRV_SESSION_CLOSE(pvSession) do { } while (0)
70	# define VBOXDRV_SUPDRV_IOCCLOSE(pvSession) do { } while (0)
71	# define VBOXDRV_SUPDRV_IOCTL_ENTRY(pvSession, uIOCtl, pvReqHdr) do { } while (0)
72	# define VBOXDRV_SUPDRV_IOCTL_RETURN(pvSession, uIOCtl, pvReqHdr, rcRet, rcReq) do { } while (0)
73	#endif
74
75	/*
76	* Logging assignments:
77	* Log - useful stuff, like failures.
78	* LogFlow - program flow, except the really noisy bits.
79	* Log2 - Cleanup.
80	* Log3 - Loader flow noise.
81	* Log4 - Call VMMR0 flow noise.
82	* Log5 - Native yet-to-be-defined noise.
83	* Log6 - Native ioctl flow noise.
84	*
85	* Logging requires BUILD_TYPE=debug and possibly changes to the logger
86	* instantiation in log-vbox.c(pp).
87	*/
88
89
90	/*******************************************************************************
91	* Defined Constants And Macros *
92	*******************************************************************************/
93	/** The frequency by which we recalculate the u32UpdateHz and
94	* u32UpdateIntervalNS GIP members. The value must be a power of 2. */
95	#define GIP_UPDATEHZ_RECALC_FREQ 0x800
96
97	/** @def VBOX_SVN_REV
98	* The makefile should define this if it can. */
99	#ifndef VBOX_SVN_REV
100	# define VBOX_SVN_REV 0
101	#endif
102
103	#if 0 /* Don't start the GIP timers. Useful when debugging the IPRT timer code. */
104	# define DO_NOT_START_GIP
105	#endif
106
107
108	/*******************************************************************************
109	* Internal Functions *
110	*******************************************************************************/
111	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser);
112	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser);
113	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession);
114	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType);
115	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq);
116	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq);
117	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq);
118	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq);
119	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq);
120	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx);
121	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt);
122	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage);
123	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage);
124	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt);
125	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt);
126	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq);
127	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq);
128	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt);
129	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt);
130	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
131	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
132	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser);
133	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
134	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned cCpus);
135	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2);
136	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip);
137	static void supdrvGipUpdate(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick);
138	static void supdrvGipUpdatePerCpu(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC,
139	RTCPUID idCpu, uint8_t idApic, uint64_t iTick);
140
141
142	/*******************************************************************************
143	* Global Variables *
144	*******************************************************************************/
145	DECLEXPORT(PSUPGLOBALINFOPAGE) g_pSUPGlobalInfoPage = NULL;
146
147	/**
148	* Array of the R0 SUP API.
149	*/
150	static SUPFUNC g_aFunctions[] =
151	{
152	/* name function */
153	/* Entries with absolute addresses determined at runtime, fixup
154	code makes ugly ASSUMPTIONS about the order here: */
155	{ "SUPR0AbsIs64bit", (void *)0 },
156	{ "SUPR0Abs64bitKernelCS", (void *)0 },
157	{ "SUPR0Abs64bitKernelSS", (void *)0 },
158	{ "SUPR0Abs64bitKernelDS", (void *)0 },
159	{ "SUPR0AbsKernelCS", (void *)0 },
160	{ "SUPR0AbsKernelSS", (void *)0 },
161	{ "SUPR0AbsKernelDS", (void *)0 },
162	{ "SUPR0AbsKernelES", (void *)0 },
163	{ "SUPR0AbsKernelFS", (void *)0 },
164	{ "SUPR0AbsKernelGS", (void *)0 },
165	/* Normal function pointers: */
166	{ "SUPR0ComponentRegisterFactory", (void *)SUPR0ComponentRegisterFactory },
167	{ "SUPR0ComponentDeregisterFactory", (void *)SUPR0ComponentDeregisterFactory },
168	{ "SUPR0ComponentQueryFactory", (void *)SUPR0ComponentQueryFactory },
169	{ "SUPR0ObjRegister", (void *)SUPR0ObjRegister },
170	{ "SUPR0ObjAddRef", (void *)SUPR0ObjAddRef },
171	{ "SUPR0ObjAddRefEx", (void *)SUPR0ObjAddRefEx },
172	{ "SUPR0ObjRelease", (void *)SUPR0ObjRelease },
173	{ "SUPR0ObjVerifyAccess", (void *)SUPR0ObjVerifyAccess },
174	{ "SUPR0LockMem", (void *)SUPR0LockMem },
175	{ "SUPR0UnlockMem", (void *)SUPR0UnlockMem },
176	{ "SUPR0ContAlloc", (void *)SUPR0ContAlloc },
177	{ "SUPR0ContFree", (void *)SUPR0ContFree },
178	{ "SUPR0LowAlloc", (void *)SUPR0LowAlloc },
179	{ "SUPR0LowFree", (void *)SUPR0LowFree },
180	{ "SUPR0MemAlloc", (void *)SUPR0MemAlloc },
181	{ "SUPR0MemGetPhys", (void *)SUPR0MemGetPhys },
182	{ "SUPR0MemFree", (void *)SUPR0MemFree },
183	{ "SUPR0PageAllocEx", (void *)SUPR0PageAllocEx },
184	{ "SUPR0PageFree", (void *)SUPR0PageFree },
185	{ "SUPR0Printf", (void )SUPR0Printf }, /* @todo needs wrapping? */
186	{ "SUPSemEventCreate", (void *)SUPSemEventCreate },
187	{ "SUPSemEventClose", (void *)SUPSemEventClose },
188	{ "SUPSemEventSignal", (void *)SUPSemEventSignal },
189	{ "SUPSemEventWait", (void *)SUPSemEventWait },
190	{ "SUPSemEventWaitNoResume", (void *)SUPSemEventWaitNoResume },
191	{ "SUPSemEventWaitNsAbsIntr", (void *)SUPSemEventWaitNsAbsIntr },
192	{ "SUPSemEventWaitNsRelIntr", (void *)SUPSemEventWaitNsRelIntr },
193	{ "SUPSemEventGetResolution", (void *)SUPSemEventGetResolution },
194	{ "SUPSemEventMultiCreate", (void *)SUPSemEventMultiCreate },
195	{ "SUPSemEventMultiClose", (void *)SUPSemEventMultiClose },
196	{ "SUPSemEventMultiSignal", (void *)SUPSemEventMultiSignal },
197	{ "SUPSemEventMultiReset", (void *)SUPSemEventMultiReset },
198	{ "SUPSemEventMultiWait", (void *)SUPSemEventMultiWait },
199	{ "SUPSemEventMultiWaitNoResume", (void *)SUPSemEventMultiWaitNoResume },
200	{ "SUPSemEventMultiWaitNsAbsIntr", (void *)SUPSemEventMultiWaitNsAbsIntr },
201	{ "SUPSemEventMultiWaitNsRelIntr", (void *)SUPSemEventMultiWaitNsRelIntr },
202	{ "SUPSemEventMultiGetResolution", (void *)SUPSemEventMultiGetResolution },
203	{ "SUPR0GetPagingMode", (void *)SUPR0GetPagingMode },
204	{ "SUPR0EnableVTx", (void *)SUPR0EnableVTx },
205	{ "SUPGetGIP", (void *)SUPGetGIP },
206	{ "g_pSUPGlobalInfoPage", (void *)&g_pSUPGlobalInfoPage },
207	{ "RTMemAllocTag", (void *)RTMemAllocTag },
208	{ "RTMemAllocZTag", (void *)RTMemAllocZTag },
209	{ "RTMemAllocVarTag", (void *)RTMemAllocVarTag },
210	{ "RTMemAllocZVarTag", (void *)RTMemAllocZVarTag },
211	{ "RTMemFree", (void *)RTMemFree },
212	{ "RTMemDupTag", (void *)RTMemDupTag },
213	{ "RTMemDupExTag", (void *)RTMemDupExTag },
214	{ "RTMemReallocTag", (void *)RTMemReallocTag },
215	{ "RTR0MemObjAllocLowTag", (void *)RTR0MemObjAllocLowTag },
216	{ "RTR0MemObjAllocPageTag", (void *)RTR0MemObjAllocPageTag },
217	{ "RTR0MemObjAllocPhysTag", (void *)RTR0MemObjAllocPhysTag },
218	{ "RTR0MemObjAllocPhysExTag", (void *)RTR0MemObjAllocPhysExTag },
219	{ "RTR0MemObjAllocPhysNCTag", (void *)RTR0MemObjAllocPhysNCTag },
220	{ "RTR0MemObjAllocContTag", (void *)RTR0MemObjAllocContTag },
221	{ "RTR0MemObjEnterPhysTag", (void *)RTR0MemObjEnterPhysTag },
222	{ "RTR0MemObjLockUserTag", (void *)RTR0MemObjLockUserTag },
223	{ "RTR0MemObjMapKernelTag", (void *)RTR0MemObjMapKernelTag },
224	{ "RTR0MemObjMapKernelExTag", (void *)RTR0MemObjMapKernelExTag },
225	{ "RTR0MemObjMapUserTag", (void *)RTR0MemObjMapUserTag },
226	{ "RTR0MemObjProtect", (void *)RTR0MemObjProtect },
227	{ "RTR0MemObjAddress", (void *)RTR0MemObjAddress },
228	{ "RTR0MemObjAddressR3", (void *)RTR0MemObjAddressR3 },
229	{ "RTR0MemObjSize", (void *)RTR0MemObjSize },
230	{ "RTR0MemObjIsMapping", (void *)RTR0MemObjIsMapping },
231	{ "RTR0MemObjGetPagePhysAddr", (void *)RTR0MemObjGetPagePhysAddr },
232	{ "RTR0MemObjFree", (void *)RTR0MemObjFree },
233	{ "RTR0MemUserCopyFrom", (void *)RTR0MemUserCopyFrom },
234	{ "RTR0MemUserCopyTo", (void *)RTR0MemUserCopyTo },
235	{ "RTR0MemUserIsValidAddr", (void *)RTR0MemUserIsValidAddr },
236	{ "RTR0MemKernelIsValidAddr", (void *)RTR0MemKernelIsValidAddr },
237	{ "RTR0MemAreKrnlAndUsrDifferent", (void *)RTR0MemAreKrnlAndUsrDifferent },
238	{ "RTSemMutexCreate", (void *)RTSemMutexCreate },
239	{ "RTSemMutexRequest", (void *)RTSemMutexRequest },
240	{ "RTSemMutexRequestDebug", (void *)RTSemMutexRequestDebug },
241	{ "RTSemMutexRequestNoResume", (void *)RTSemMutexRequestNoResume },
242	{ "RTSemMutexRequestNoResumeDebug", (void *)RTSemMutexRequestNoResumeDebug },
243	{ "RTSemMutexRelease", (void *)RTSemMutexRelease },
244	{ "RTSemMutexDestroy", (void *)RTSemMutexDestroy },
245	{ "RTProcSelf", (void *)RTProcSelf },
246	{ "RTR0ProcHandleSelf", (void *)RTR0ProcHandleSelf },
247	{ "RTSemFastMutexCreate", (void *)RTSemFastMutexCreate },
248	{ "RTSemFastMutexDestroy", (void *)RTSemFastMutexDestroy },
249	{ "RTSemFastMutexRequest", (void *)RTSemFastMutexRequest },
250	{ "RTSemFastMutexRelease", (void *)RTSemFastMutexRelease },
251	{ "RTSemEventCreate", (void *)RTSemEventCreate },
252	{ "RTSemEventSignal", (void *)RTSemEventSignal },
253	{ "RTSemEventWait", (void *)RTSemEventWait },
254	{ "RTSemEventWaitNoResume", (void *)RTSemEventWaitNoResume },
255	{ "RTSemEventWaitEx", (void *)RTSemEventWaitEx },
256	{ "RTSemEventWaitExDebug", (void *)RTSemEventWaitExDebug },
257	{ "RTSemEventGetResolution", (void *)RTSemEventGetResolution },
258	{ "RTSemEventDestroy", (void *)RTSemEventDestroy },
259	{ "RTSemEventMultiCreate", (void *)RTSemEventMultiCreate },
260	{ "RTSemEventMultiSignal", (void *)RTSemEventMultiSignal },
261	{ "RTSemEventMultiReset", (void *)RTSemEventMultiReset },
262	{ "RTSemEventMultiWait", (void *)RTSemEventMultiWait },
263	{ "RTSemEventMultiWaitNoResume", (void *)RTSemEventMultiWaitNoResume },
264	{ "RTSemEventMultiWaitEx", (void *)RTSemEventMultiWaitEx },
265	{ "RTSemEventMultiWaitExDebug", (void *)RTSemEventMultiWaitExDebug },
266	{ "RTSemEventMultiGetResolution", (void *)RTSemEventMultiGetResolution },
267	{ "RTSemEventMultiDestroy", (void *)RTSemEventMultiDestroy },
268	{ "RTSpinlockCreate", (void *)RTSpinlockCreate },
269	{ "RTSpinlockDestroy", (void *)RTSpinlockDestroy },
270	{ "RTSpinlockAcquire", (void *)RTSpinlockAcquire },
271	{ "RTSpinlockRelease", (void *)RTSpinlockRelease },
272	{ "RTSpinlockAcquireNoInts", (void *)RTSpinlockAcquireNoInts },
273	{ "RTSpinlockReleaseNoInts", (void *)RTSpinlockReleaseNoInts },
274	{ "RTTimeNanoTS", (void *)RTTimeNanoTS },
275	{ "RTTimeMilliTS", (void *)RTTimeMilliTS },
276	{ "RTTimeSystemNanoTS", (void *)RTTimeSystemNanoTS },
277	{ "RTTimeSystemMilliTS", (void *)RTTimeSystemMilliTS },
278	{ "RTThreadNativeSelf", (void *)RTThreadNativeSelf },
279	{ "RTThreadSleep", (void *)RTThreadSleep },
280	{ "RTThreadYield", (void *)RTThreadYield },
281	{ "RTThreadSelf", (void *)RTThreadSelf },
282	{ "RTThreadCreate", (void *)RTThreadCreate },
283	{ "RTThreadGetNative", (void *)RTThreadGetNative },
284	{ "RTThreadWait", (void *)RTThreadWait },
285	{ "RTThreadWaitNoResume", (void *)RTThreadWaitNoResume },
286	{ "RTThreadGetName", (void *)RTThreadGetName },
287	{ "RTThreadSelfName", (void *)RTThreadSelfName },
288	{ "RTThreadGetType", (void *)RTThreadGetType },
289	{ "RTThreadUserSignal", (void *)RTThreadUserSignal },
290	{ "RTThreadUserReset", (void *)RTThreadUserReset },
291	{ "RTThreadUserWait", (void *)RTThreadUserWait },
292	{ "RTThreadUserWaitNoResume", (void *)RTThreadUserWaitNoResume },
293	{ "RTThreadPreemptIsEnabled", (void *)RTThreadPreemptIsEnabled },
294	{ "RTThreadPreemptIsPending", (void *)RTThreadPreemptIsPending },
295	{ "RTThreadPreemptIsPendingTrusty", (void *)RTThreadPreemptIsPendingTrusty },
296	{ "RTThreadPreemptIsPossible", (void *)RTThreadPreemptIsPossible },
297	{ "RTThreadPreemptDisable", (void *)RTThreadPreemptDisable },
298	{ "RTThreadPreemptRestore", (void *)RTThreadPreemptRestore },
299	{ "RTThreadIsInInterrupt", (void *)RTThreadIsInInterrupt },
300	{ "RTTimerCreate", (void *)RTTimerCreate },
301	{ "RTTimerCreateEx", (void *)RTTimerCreateEx },
302	{ "RTTimerDestroy", (void *)RTTimerDestroy },
303	{ "RTTimerStart", (void *)RTTimerStart },
304	{ "RTTimerStop", (void *)RTTimerStop },
305	{ "RTTimerChangeInterval", (void *)RTTimerChangeInterval },
306	{ "RTTimerGetSystemGranularity", (void *)RTTimerGetSystemGranularity },
307	{ "RTTimerRequestSystemGranularity", (void *)RTTimerRequestSystemGranularity },
308	{ "RTTimerReleaseSystemGranularity", (void *)RTTimerReleaseSystemGranularity },
309	{ "RTTimerCanDoHighResolution", (void *)RTTimerCanDoHighResolution },
310
311	{ "RTLogDefaultInstance", (void *)RTLogDefaultInstance },
312	{ "RTMpCpuId", (void *)RTMpCpuId },
313	{ "RTMpCpuIdFromSetIndex", (void *)RTMpCpuIdFromSetIndex },
314	{ "RTMpCpuIdToSetIndex", (void *)RTMpCpuIdToSetIndex },
315	{ "RTMpGetArraySize", (void *)RTMpGetArraySize },
316	{ "RTMpIsCpuPossible", (void *)RTMpIsCpuPossible },
317	{ "RTMpGetCount", (void *)RTMpGetCount },
318	{ "RTMpGetMaxCpuId", (void *)RTMpGetMaxCpuId },
319	{ "RTMpGetOnlineCount", (void *)RTMpGetOnlineCount },
320	{ "RTMpGetOnlineSet", (void *)RTMpGetOnlineSet },
321	{ "RTMpGetSet", (void *)RTMpGetSet },
322	{ "RTMpIsCpuOnline", (void *)RTMpIsCpuOnline },
323	{ "RTMpIsCpuWorkPending", (void *)RTMpIsCpuWorkPending },
324	{ "RTMpOnAll", (void *)RTMpOnAll },
325	{ "RTMpOnOthers", (void *)RTMpOnOthers },
326	{ "RTMpOnSpecific", (void *)RTMpOnSpecific },
327	{ "RTMpPokeCpu", (void *)RTMpPokeCpu },
328	{ "RTPowerNotificationRegister", (void *)RTPowerNotificationRegister },
329	{ "RTPowerNotificationDeregister", (void *)RTPowerNotificationDeregister },
330	{ "RTLogRelDefaultInstance", (void *)RTLogRelDefaultInstance },
331	{ "RTLogSetDefaultInstanceThread", (void *)RTLogSetDefaultInstanceThread },
332	{ "RTLogLoggerExV", (void *)RTLogLoggerExV },
333	{ "RTLogPrintfV", (void *)RTLogPrintfV },
334	{ "RTR0AssertPanicSystem", (void *)RTR0AssertPanicSystem },
335	{ "RTAssertMsg1", (void *)RTAssertMsg1 },
336	{ "RTAssertMsg2V", (void *)RTAssertMsg2V },
337	{ "RTAssertSetQuiet", (void *)RTAssertSetQuiet },
338	{ "RTAssertMayPanic", (void *)RTAssertMayPanic },
339	{ "RTAssertSetMayPanic", (void *)RTAssertSetMayPanic },
340	{ "RTAssertAreQuiet", (void *)RTAssertAreQuiet },
341	{ "RTStrFormat", (void *)RTStrFormat },
342	{ "RTStrFormatNumber", (void *)RTStrFormatNumber },
343	{ "RTStrFormatTypeDeregister", (void *)RTStrFormatTypeDeregister },
344	{ "RTStrFormatTypeRegister", (void *)RTStrFormatTypeRegister },
345	{ "RTStrFormatTypeSetUser", (void *)RTStrFormatTypeSetUser },
346	{ "RTStrFormatV", (void *)RTStrFormatV },
347	{ "RTStrPrintf", (void *)RTStrPrintf },
348	{ "RTStrPrintfEx", (void *)RTStrPrintfEx },
349	{ "RTStrPrintfExV", (void *)RTStrPrintfExV },
350	{ "RTStrPrintfV", (void *)RTStrPrintfV },
351	{ "RTHandleTableAllocWithCtx", (void *)RTHandleTableAllocWithCtx },
352	{ "RTHandleTableCreate", (void *)RTHandleTableCreate },
353	{ "RTHandleTableCreateEx", (void *)RTHandleTableCreateEx },
354	{ "RTHandleTableDestroy", (void *)RTHandleTableDestroy },
355	{ "RTHandleTableFreeWithCtx", (void *)RTHandleTableFreeWithCtx },
356	{ "RTHandleTableLookupWithCtx", (void *)RTHandleTableLookupWithCtx },
357	{ "RTNetIPv4AddDataChecksum", (void *)RTNetIPv4AddDataChecksum },
358	{ "RTNetIPv4AddTCPChecksum", (void *)RTNetIPv4AddTCPChecksum },
359	{ "RTNetIPv4AddUDPChecksum", (void *)RTNetIPv4AddUDPChecksum },
360	{ "RTNetIPv4FinalizeChecksum", (void *)RTNetIPv4FinalizeChecksum },
361	{ "RTNetIPv4HdrChecksum", (void *)RTNetIPv4HdrChecksum },
362	{ "RTNetIPv4IsDHCPValid", (void *)RTNetIPv4IsDHCPValid },
363	{ "RTNetIPv4IsHdrValid", (void *)RTNetIPv4IsHdrValid },
364	{ "RTNetIPv4IsTCPSizeValid", (void *)RTNetIPv4IsTCPSizeValid },
365	{ "RTNetIPv4IsTCPValid", (void *)RTNetIPv4IsTCPValid },
366	{ "RTNetIPv4IsUDPSizeValid", (void *)RTNetIPv4IsUDPSizeValid },
367	{ "RTNetIPv4IsUDPValid", (void *)RTNetIPv4IsUDPValid },
368	{ "RTNetIPv4PseudoChecksum", (void *)RTNetIPv4PseudoChecksum },
369	{ "RTNetIPv4PseudoChecksumBits", (void *)RTNetIPv4PseudoChecksumBits },
370	{ "RTNetIPv4TCPChecksum", (void *)RTNetIPv4TCPChecksum },
371	{ "RTNetIPv4UDPChecksum", (void *)RTNetIPv4UDPChecksum },
372	{ "RTNetIPv6PseudoChecksum", (void *)RTNetIPv6PseudoChecksum },
373	{ "RTNetIPv6PseudoChecksumBits", (void *)RTNetIPv6PseudoChecksumBits },
374	{ "RTNetIPv6PseudoChecksumEx", (void *)RTNetIPv6PseudoChecksumEx },
375	{ "RTNetTCPChecksum", (void *)RTNetTCPChecksum },
376	{ "RTNetUDPChecksum", (void *)RTNetUDPChecksum },
377	{ "RTCrc32", (void *)RTCrc32 },
378	{ "RTCrc32Finish", (void *)RTCrc32Finish },
379	{ "RTCrc32Process", (void *)RTCrc32Process },
380	{ "RTCrc32Start", (void *)RTCrc32Start },
381	};
382
383	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
384	/**
385	* Drag in the rest of IRPT since we share it with the
386	* rest of the kernel modules on darwin.
387	*/
388	PFNRT g_apfnVBoxDrvIPRTDeps[] =
389	{
390	/* VBoxNetFlt */
391	(PFNRT)RTErrConvertFromErrno,
392	(PFNRT)RTUuidCompare,
393	(PFNRT)RTUuidCompareStr,
394	(PFNRT)RTUuidFromStr,
395	(PFNRT)RTStrDupTag,
396	(PFNRT)RTStrFree,
397	/* VBoxNetAdp */
398	(PFNRT)RTRandBytes,
399	/* VBoxUSB */
400	(PFNRT)RTPathStripFilename,
401	NULL
402	};
403	#endif /* RT_OS_DARWIN \|\| RT_OS_SOLARIS \|\| RT_OS_SOLARIS */
404
405
406	/**
407	* Initializes the device extentsion structure.
408	*
409	* @returns IPRT status code.
410	* @param pDevExt The device extension to initialize.
411	* @param cbSession The size of the session structure. The size of
412	* SUPDRVSESSION may be smaller when SUPDRV_AGNOSTIC is
413	* defined because we're skipping the OS specific members
414	* then.
415	*/
416	int VBOXCALL supdrvInitDevExt(PSUPDRVDEVEXT pDevExt, size_t cbSession)
417	{
418	int rc;
419
420	#ifdef SUPDRV_WITH_RELEASE_LOGGER
421	/*
422	* Create the release log.
423	*/
424	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
425	PRTLOGGER pRelLogger;
426	rc = RTLogCreate(&pRelLogger, 0 /* fFlags */, "all",
427	"VBOX_RELEASE_LOG", RT_ELEMENTS(s_apszGroups), s_apszGroups, RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER, NULL);
428	if (RT_SUCCESS(rc))
429	RTLogRelSetDefaultInstance(pRelLogger);
430	/** @todo Add native hook for getting logger config parameters and setting
431	* them. On linux we should use the module parameter stuff... */
432	#endif
433
434	/*
435	* Initialize it.
436	*/
437	memset(pDevExt, 0, sizeof(*pDevExt));
438	rc = RTSpinlockCreate(&pDevExt->Spinlock);
439	if (RT_SUCCESS(rc))
440	{
441	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
442	rc = RTSemMutexCreate(&pDevExt->mtxLdr);
443	#else
444	rc = RTSemFastMutexCreate(&pDevExt->mtxLdr);
445	#endif
446	if (RT_SUCCESS(rc))
447	{
448	rc = RTSemFastMutexCreate(&pDevExt->mtxComponentFactory);
449	if (RT_SUCCESS(rc))
450	{
451	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
452	rc = RTSemMutexCreate(&pDevExt->mtxGip);
453	#else
454	rc = RTSemFastMutexCreate(&pDevExt->mtxGip);
455	#endif
456	if (RT_SUCCESS(rc))
457	{
458	rc = supdrvGipCreate(pDevExt);
459	if (RT_SUCCESS(rc))
460	{
461	pDevExt->u32Cookie = BIRD; /** @todo make this random? */
462	pDevExt->cbSession = (uint32_t)cbSession;
463
464	/*
465	* Fixup the absolute symbols.
466	*
467	* Because of the table indexing assumptions we'll have a little #ifdef orgy
468	* here rather than distributing this to OS specific files. At least for now.
469	*/
470	#ifdef RT_OS_DARWIN
471	# if ARCH_BITS == 32
472	if (SUPR0GetPagingMode() >= SUPPAGINGMODE_AMD64)
473	{
474	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
475	g_aFunctions[1].pfn = (void )0x80; / SUPR0Abs64bitKernelCS - KERNEL64_CS, seg.h */
476	g_aFunctions[2].pfn = (void )0x88; / SUPR0Abs64bitKernelSS - KERNEL64_SS, seg.h */
477	g_aFunctions[3].pfn = (void )0x88; / SUPR0Abs64bitKernelDS - KERNEL64_SS, seg.h */
478	}
479	else
480	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
481	g_aFunctions[4].pfn = (void )0x08; / SUPR0AbsKernelCS - KERNEL_CS, seg.h */
482	g_aFunctions[5].pfn = (void )0x10; / SUPR0AbsKernelSS - KERNEL_DS, seg.h */
483	g_aFunctions[6].pfn = (void )0x10; / SUPR0AbsKernelDS - KERNEL_DS, seg.h */
484	g_aFunctions[7].pfn = (void )0x10; / SUPR0AbsKernelES - KERNEL_DS, seg.h */
485	g_aFunctions[8].pfn = (void )0x10; / SUPR0AbsKernelFS - KERNEL_DS, seg.h */
486	g_aFunctions[9].pfn = (void )0x48; / SUPR0AbsKernelGS - CPU_DATA_GS, seg.h */
487	# else /* 64-bit darwin: */
488	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
489	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
490	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
491	g_aFunctions[3].pfn = (void )0; / SUPR0Abs64bitKernelDS */
492	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
493	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
494	g_aFunctions[6].pfn = (void )0; / SUPR0AbsKernelDS */
495	g_aFunctions[7].pfn = (void )0; / SUPR0AbsKernelES */
496	g_aFunctions[8].pfn = (void )0; / SUPR0AbsKernelFS */
497	g_aFunctions[9].pfn = (void )0; / SUPR0AbsKernelGS */
498
499	# endif
500	#else /* !RT_OS_DARWIN */
501	# if ARCH_BITS == 64
502	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
503	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
504	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
505	g_aFunctions[3].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0Abs64bitKernelDS */
506	# else
507	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
508	# endif
509	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
510	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
511	g_aFunctions[6].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0AbsKernelDS */
512	g_aFunctions[7].pfn = (void )(uintptr_t)ASMGetES(); / SUPR0AbsKernelES */
513	g_aFunctions[8].pfn = (void )(uintptr_t)ASMGetFS(); / SUPR0AbsKernelFS */
514	g_aFunctions[9].pfn = (void )(uintptr_t)ASMGetGS(); / SUPR0AbsKernelGS */
515	#endif /* !RT_OS_DARWIN */
516	return VINF_SUCCESS;
517	}
518
519	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
520	RTSemMutexDestroy(pDevExt->mtxGip);
521	pDevExt->mtxGip = NIL_RTSEMMUTEX;
522	#else
523	RTSemFastMutexDestroy(pDevExt->mtxGip);
524	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
525	#endif
526	}
527	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
528	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
529	}
530	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
531	RTSemMutexDestroy(pDevExt->mtxLdr);
532	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
533	#else
534	RTSemFastMutexDestroy(pDevExt->mtxLdr);
535	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
536	#endif
537	}
538	RTSpinlockDestroy(pDevExt->Spinlock);
539	pDevExt->Spinlock = NIL_RTSPINLOCK;
540	}
541	#ifdef SUPDRV_WITH_RELEASE_LOGGER
542	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
543	RTLogDestroy(RTLogSetDefaultInstance(NULL));
544	#endif
545
546	return rc;
547	}
548
549
550	/**
551	* Delete the device extension (e.g. cleanup members).
552	*
553	* @param pDevExt The device extension to delete.
554	*/
555	void VBOXCALL supdrvDeleteDevExt(PSUPDRVDEVEXT pDevExt)
556	{
557	PSUPDRVOBJ pObj;
558	PSUPDRVUSAGE pUsage;
559
560	/*
561	* Kill mutexes and spinlocks.
562	*/
563	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
564	RTSemMutexDestroy(pDevExt->mtxGip);
565	pDevExt->mtxGip = NIL_RTSEMMUTEX;
566	#else
567	RTSemFastMutexDestroy(pDevExt->mtxGip);
568	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
569	#endif
570	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
571	RTSemMutexDestroy(pDevExt->mtxLdr);
572	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
573	#else
574	RTSemFastMutexDestroy(pDevExt->mtxLdr);
575	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
576	#endif
577	RTSpinlockDestroy(pDevExt->Spinlock);
578	pDevExt->Spinlock = NIL_RTSPINLOCK;
579	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
580	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
581
582	/*
583	* Free lists.
584	*/
585	/* objects. */
586	pObj = pDevExt->pObjs;
587	Assert(!pObj); /* (can trigger on forced unloads) */
588	pDevExt->pObjs = NULL;
589	while (pObj)
590	{
591	void *pvFree = pObj;
592	pObj = pObj->pNext;
593	RTMemFree(pvFree);
594	}
595
596	/* usage records. */
597	pUsage = pDevExt->pUsageFree;
598	pDevExt->pUsageFree = NULL;
599	while (pUsage)
600	{
601	void *pvFree = pUsage;
602	pUsage = pUsage->pNext;
603	RTMemFree(pvFree);
604	}
605
606	/* kill the GIP. */
607	supdrvGipDestroy(pDevExt);
608
609	#ifdef SUPDRV_WITH_RELEASE_LOGGER
610	/* destroy the loggers. */
611	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
612	RTLogDestroy(RTLogSetDefaultInstance(NULL));
613	#endif
614	}
615
616
617	/**
618	* Create session.
619	*
620	* @returns IPRT status code.
621	* @param pDevExt Device extension.
622	* @param fUser Flag indicating whether this is a user or kernel session.
623	* @param ppSession Where to store the pointer to the session data.
624	*/
625	int VBOXCALL supdrvCreateSession(PSUPDRVDEVEXT pDevExt, bool fUser, PSUPDRVSESSION *ppSession)
626	{
627	/*
628	* Allocate memory for the session data.
629	*/
630	int rc;
631	PSUPDRVSESSION pSession = *ppSession = (PSUPDRVSESSION)RTMemAllocZ(pDevExt->cbSession);
632	if (pSession)
633	{
634	/* Initialize session data. */
635	rc = RTSpinlockCreate(&pSession->Spinlock);
636	if (!rc)
637	{
638	rc = RTHandleTableCreateEx(&pSession->hHandleTable,
639	RTHANDLETABLE_FLAGS_LOCKED \| RTHANDLETABLE_FLAGS_CONTEXT,
640	1 /uBase/, 32768 /cMax/, supdrvSessionObjHandleRetain, pSession);
641	if (RT_SUCCESS(rc))
642	{
643	Assert(pSession->Spinlock != NIL_RTSPINLOCK);
644	pSession->pDevExt = pDevExt;
645	pSession->u32Cookie = BIRD_INV;
646	/*pSession->pLdrUsage = NULL;
647	pSession->pVM = NULL;
648	pSession->pUsage = NULL;
649	pSession->pGip = NULL;
650	pSession->fGipReferenced = false;
651	pSession->Bundle.cUsed = 0; */
652	pSession->Uid = NIL_RTUID;
653	pSession->Gid = NIL_RTGID;
654	if (fUser)
655	{
656	pSession->Process = RTProcSelf();
657	pSession->R0Process = RTR0ProcHandleSelf();
658	}
659	else
660	{
661	pSession->Process = NIL_RTPROCESS;
662	pSession->R0Process = NIL_RTR0PROCESS;
663	}
664
665	VBOXDRV_SUPDRV_SESSION_CREATE(pSession, fUser);
666	LogFlow(("Created session %p initial cookie=%#x\n", pSession, pSession->u32Cookie));
667	return VINF_SUCCESS;
668	}
669
670	RTSpinlockDestroy(pSession->Spinlock);
671	}
672	RTMemFree(pSession);
673	*ppSession = NULL;
674	Log(("Failed to create spinlock, rc=%d!\n", rc));
675	}
676	else
677	rc = VERR_NO_MEMORY;
678
679	return rc;
680	}
681
682
683	/**
684	* Shared code for cleaning up a session.
685	*
686	* @param pDevExt Device extension.
687	* @param pSession Session data.
688	* This data will be freed by this routine.
689	*/
690	void VBOXCALL supdrvCloseSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
691	{
692	VBOXDRV_SUPDRV_SESSION_CLOSE(pSession);
693
694	/*
695	* Cleanup the session first.
696	*/
697	supdrvCleanupSession(pDevExt, pSession);
698
699	/*
700	* Free the rest of the session stuff.
701	*/
702	RTSpinlockDestroy(pSession->Spinlock);
703	pSession->Spinlock = NIL_RTSPINLOCK;
704	pSession->pDevExt = NULL;
705	RTMemFree(pSession);
706	LogFlow(("supdrvCloseSession: returns\n"));
707	}
708
709
710	/**
711	* Shared code for cleaning up a session (but not quite freeing it).
712	*
713	* This is primarily intended for MAC OS X where we have to clean up the memory
714	* stuff before the file handle is closed.
715	*
716	* @param pDevExt Device extension.
717	* @param pSession Session data.
718	* This data will be freed by this routine.
719	*/
720	void VBOXCALL supdrvCleanupSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
721	{
722	int rc;
723	PSUPDRVBUNDLE pBundle;
724	LogFlow(("supdrvCleanupSession: pSession=%p\n", pSession));
725
726	/*
727	* Remove logger instances related to this session.
728	*/
729	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pSession);
730
731	/*
732	* Destroy the handle table.
733	*/
734	rc = RTHandleTableDestroy(pSession->hHandleTable, supdrvSessionObjHandleDelete, pSession);
735	AssertRC(rc);
736	pSession->hHandleTable = NIL_RTHANDLETABLE;
737
738	/*
739	* Release object references made in this session.
740	* In theory there should be noone racing us in this session.
741	*/
742	Log2(("release objects - start\n"));
743	if (pSession->pUsage)
744	{
745	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
746	PSUPDRVUSAGE pUsage;
747	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
748
749	while ((pUsage = pSession->pUsage) != NULL)
750	{
751	PSUPDRVOBJ pObj = pUsage->pObj;
752	pSession->pUsage = pUsage->pNext;
753
754	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
755	if (pUsage->cUsage < pObj->cUsage)
756	{
757	pObj->cUsage -= pUsage->cUsage;
758	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
759	}
760	else
761	{
762	/* Destroy the object and free the record. */
763	if (pDevExt->pObjs == pObj)
764	pDevExt->pObjs = pObj->pNext;
765	else
766	{
767	PSUPDRVOBJ pObjPrev;
768	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
769	if (pObjPrev->pNext == pObj)
770	{
771	pObjPrev->pNext = pObj->pNext;
772	break;
773	}
774	Assert(pObjPrev);
775	}
776	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
777
778	Log(("supdrvCleanupSession: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
779	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
780	if (pObj->pfnDestructor)
781	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
782	RTMemFree(pObj);
783	}
784
785	/* free it and continue. */
786	RTMemFree(pUsage);
787
788	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
789	}
790
791	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
792	AssertMsg(!pSession->pUsage, ("Some buster reregistered an object during desturction!\n"));
793	}
794	Log2(("release objects - done\n"));
795
796	/*
797	* Release memory allocated in the session.
798	*
799	* We do not serialize this as we assume that the application will
800	* not allocated memory while closing the file handle object.
801	*/
802	Log2(("freeing memory:\n"));
803	pBundle = &pSession->Bundle;
804	while (pBundle)
805	{
806	PSUPDRVBUNDLE pToFree;
807	unsigned i;
808
809	/*
810	* Check and unlock all entries in the bundle.
811	*/
812	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
813	{
814	if (pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ)
815	{
816	Log2(("eType=%d pvR0=%p pvR3=%p cb=%ld\n", pBundle->aMem[i].eType, RTR0MemObjAddress(pBundle->aMem[i].MemObj),
817	(void *)RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3), (long)RTR0MemObjSize(pBundle->aMem[i].MemObj)));
818	if (pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ)
819	{
820	rc = RTR0MemObjFree(pBundle->aMem[i].MapObjR3, false);
821	AssertRC(rc); /** @todo figure out how to handle this. */
822	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
823	}
824	rc = RTR0MemObjFree(pBundle->aMem[i].MemObj, true /* fFreeMappings */);
825	AssertRC(rc); /** @todo figure out how to handle this. */
826	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
827	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
828	}
829	}
830
831	/*
832	* Advance and free previous bundle.
833	*/
834	pToFree = pBundle;
835	pBundle = pBundle->pNext;
836
837	pToFree->pNext = NULL;
838	pToFree->cUsed = 0;
839	if (pToFree != &pSession->Bundle)
840	RTMemFree(pToFree);
841	}
842	Log2(("freeing memory - done\n"));
843
844	/*
845	* Deregister component factories.
846	*/
847	RTSemFastMutexRequest(pDevExt->mtxComponentFactory);
848	Log2(("deregistering component factories:\n"));
849	if (pDevExt->pComponentFactoryHead)
850	{
851	PSUPDRVFACTORYREG pPrev = NULL;
852	PSUPDRVFACTORYREG pCur = pDevExt->pComponentFactoryHead;
853	while (pCur)
854	{
855	if (pCur->pSession == pSession)
856	{
857	/* unlink it */
858	PSUPDRVFACTORYREG pNext = pCur->pNext;
859	if (pPrev)
860	pPrev->pNext = pNext;
861	else
862	pDevExt->pComponentFactoryHead = pNext;
863
864	/* free it */
865	pCur->pNext = NULL;
866	pCur->pSession = NULL;
867	pCur->pFactory = NULL;
868	RTMemFree(pCur);
869
870	/* next */
871	pCur = pNext;
872	}
873	else
874	{
875	/* next */
876	pPrev = pCur;
877	pCur = pCur->pNext;
878	}
879	}
880	}
881	RTSemFastMutexRelease(pDevExt->mtxComponentFactory);
882	Log2(("deregistering component factories - done\n"));
883
884	/*
885	* Loaded images needs to be dereferenced and possibly freed up.
886	*/
887	supdrvLdrLock(pDevExt);
888	Log2(("freeing images:\n"));
889	if (pSession->pLdrUsage)
890	{
891	PSUPDRVLDRUSAGE pUsage = pSession->pLdrUsage;
892	pSession->pLdrUsage = NULL;
893	while (pUsage)
894	{
895	void *pvFree = pUsage;
896	PSUPDRVLDRIMAGE pImage = pUsage->pImage;
897	if (pImage->cUsage > pUsage->cUsage)
898	pImage->cUsage -= pUsage->cUsage;
899	else
900	supdrvLdrFree(pDevExt, pImage);
901	pUsage->pImage = NULL;
902	pUsage = pUsage->pNext;
903	RTMemFree(pvFree);
904	}
905	}
906	supdrvLdrUnlock(pDevExt);
907	Log2(("freeing images - done\n"));
908
909	/*
910	* Unmap the GIP.
911	*/
912	Log2(("umapping GIP:\n"));
913	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
914	{
915	SUPR0GipUnmap(pSession);
916	pSession->fGipReferenced = 0;
917	}
918	Log2(("umapping GIP - done\n"));
919	}
920
921
922	/**
923	* RTHandleTableDestroy callback used by supdrvCleanupSession.
924	*
925	* @returns IPRT status code, see SUPR0ObjAddRef.
926	* @param hHandleTable The handle table handle. Ignored.
927	* @param pvObj The object pointer.
928	* @param pvCtx Context, the handle type. Ignored.
929	* @param pvUser Session pointer.
930	*/
931	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser)
932	{
933	NOREF(pvCtx);
934	NOREF(hHandleTable);
935	return SUPR0ObjAddRefEx(pvObj, (PSUPDRVSESSION)pvUser, true /fNoBlocking/);
936	}
937
938
939	/**
940	* RTHandleTableDestroy callback used by supdrvCleanupSession.
941	*
942	* @param hHandleTable The handle table handle. Ignored.
943	* @param h The handle value. Ignored.
944	* @param pvObj The object pointer.
945	* @param pvCtx Context, the handle type. Ignored.
946	* @param pvUser Session pointer.
947	*/
948	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser)
949	{
950	NOREF(pvCtx);
951	NOREF(h);
952	NOREF(hHandleTable);
953	SUPR0ObjRelease(pvObj, (PSUPDRVSESSION)pvUser);
954	}
955
956
957	/**
958	* Fast path I/O Control worker.
959	*
960	* @returns VBox status code that should be passed down to ring-3 unchanged.
961	* @param uIOCtl Function number.
962	* @param idCpu VMCPU id.
963	* @param pDevExt Device extention.
964	* @param pSession Session data.
965	*/
966	int VBOXCALL supdrvIOCtlFast(uintptr_t uIOCtl, VMCPUID idCpu, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
967	{
968	/*
969	* We check the two prereqs after doing this only to allow the compiler to optimize things better.
970	*/
971	if (RT_LIKELY(pSession->pVM && pDevExt->pfnVMMR0EntryFast))
972	{
973	switch (uIOCtl)
974	{
975	case SUP_IOCTL_FAST_DO_RAW_RUN:
976	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_RAW_RUN);
977	break;
978	case SUP_IOCTL_FAST_DO_HWACC_RUN:
979	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_HWACC_RUN);
980	break;
981	case SUP_IOCTL_FAST_DO_NOP:
982	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_NOP);
983	break;
984	default:
985	return VERR_INTERNAL_ERROR;
986	}
987	return VINF_SUCCESS;
988	}
989	return VERR_INTERNAL_ERROR;
990	}
991
992
993	/**
994	* Helper for supdrvIOCtl. Check if pszStr contains any character of pszChars.
995	* We would use strpbrk here if this function would be contained in the RedHat kABI white
996	* list, see http://www.kerneldrivers.org/RHEL5.
997	*
998	* @returns 1 if pszStr does contain any character of pszChars, 0 otherwise.
999	* @param pszStr String to check
1000	* @param pszChars Character set
1001	*/
1002	static int supdrvCheckInvalidChar(const char pszStr, const char pszChars)
1003	{
1004	int chCur;
1005	while ((chCur = *pszStr++) != '\0')
1006	{
1007	int ch;
1008	const char *psz = pszChars;
1009	while ((ch = *psz++) != '\0')
1010	if (ch == chCur)
1011	return 1;
1012
1013	}
1014	return 0;
1015	}
1016
1017
1018	/**
1019	* I/O Control worker.
1020	*
1021	* @returns IPRT status code.
1022	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1023	*
1024	* @param uIOCtl Function number.
1025	* @param pDevExt Device extention.
1026	* @param pSession Session data.
1027	* @param pReqHdr The request header.
1028	*/
1029	int VBOXCALL supdrvIOCtl(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1030	{
1031	VBOXDRV_SUPDRV_IOCTL_ENTRY(pSession, uIOCtl, pReqHdr);
1032
1033	/*
1034	* Validate the request.
1035	*/
1036	/* this first check could probably be omitted as its also done by the OS specific code... */
1037	if (RT_UNLIKELY( (pReqHdr->fFlags & SUPREQHDR_FLAGS_MAGIC_MASK) != SUPREQHDR_FLAGS_MAGIC
1038	\|\| pReqHdr->cbIn < sizeof(*pReqHdr)
1039	\|\| pReqHdr->cbOut < sizeof(*pReqHdr)))
1040	{
1041	OSDBGPRINT(("vboxdrv: Bad ioctl request header; cbIn=%#lx cbOut=%#lx fFlags=%#lx\n",
1042	(long)pReqHdr->cbIn, (long)pReqHdr->cbOut, (long)pReqHdr->fFlags));
1043	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1044	return VERR_INVALID_PARAMETER;
1045	}
1046	if (RT_UNLIKELY(uIOCtl == SUP_IOCTL_COOKIE))
1047	{
1048	if (pReqHdr->u32Cookie != SUPCOOKIE_INITIAL_COOKIE)
1049	{
1050	OSDBGPRINT(("SUP_IOCTL_COOKIE: bad cookie %#lx\n", (long)pReqHdr->u32Cookie));
1051	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1052	return VERR_INVALID_PARAMETER;
1053	}
1054	}
1055	else if (RT_UNLIKELY( pReqHdr->u32Cookie != pDevExt->u32Cookie
1056	\|\| pReqHdr->u32SessionCookie != pSession->u32Cookie))
1057	{
1058	OSDBGPRINT(("vboxdrv: bad cookie %#lx / %#lx.\n", (long)pReqHdr->u32Cookie, (long)pReqHdr->u32SessionCookie));
1059	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1060	return VERR_INVALID_PARAMETER;
1061	}
1062
1063	/*
1064	* Validation macros
1065	*/
1066	#define REQ_CHECK_SIZES_EX(Name, cbInExpect, cbOutExpect) \
1067	do { \
1068	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect) \|\| pReqHdr->cbOut != (cbOutExpect))) \
1069	{ \
1070	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld. cbOut=%ld expected %ld.\n", \
1071	(long)pReq->Hdr.cbIn, (long)(cbInExpect), (long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
1072	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1073	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1074	} \
1075	} while (0)
1076
1077	#define REQ_CHECK_SIZES(Name) REQ_CHECK_SIZES_EX(Name, Name ## _SIZE_IN, Name ## _SIZE_OUT)
1078
1079	#define REQ_CHECK_SIZE_IN(Name, cbInExpect) \
1080	do { \
1081	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect))) \
1082	{ \
1083	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld.\n", \
1084	(long)pReq->Hdr.cbIn, (long)(cbInExpect))); \
1085	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1086	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1087	} \
1088	} while (0)
1089
1090	#define REQ_CHECK_SIZE_OUT(Name, cbOutExpect) \
1091	do { \
1092	if (RT_UNLIKELY(pReqHdr->cbOut != (cbOutExpect))) \
1093	{ \
1094	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbOut=%ld expected %ld.\n", \
1095	(long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
1096	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1097	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1098	} \
1099	} while (0)
1100
1101	#define REQ_CHECK_EXPR(Name, expr) \
1102	do { \
1103	if (RT_UNLIKELY(!(expr))) \
1104	{ \
1105	OSDBGPRINT(( #Name ": %s\n", #expr)); \
1106	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1107	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1108	} \
1109	} while (0)
1110
1111	#define REQ_CHECK_EXPR_FMT(expr, fmt) \
1112	do { \
1113	if (RT_UNLIKELY(!(expr))) \
1114	{ \
1115	OSDBGPRINT( fmt ); \
1116	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1117	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1118	} \
1119	} while (0)
1120
1121
1122	/*
1123	* The switch.
1124	*/
1125	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
1126	{
1127	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
1128	{
1129	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
1130	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
1131	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
1132	{
1133	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
1134	pReq->Hdr.rc = VERR_INVALID_MAGIC;
1135	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, VERR_INVALID_MAGIC);
1136	return 0;
1137	}
1138
1139	#if 0
1140	/*
1141	* Call out to the OS specific code and let it do permission checks on the
1142	* client process.
1143	*/
1144	if (!supdrvOSValidateClientProcess(pDevExt, pSession))
1145	{
1146	pReq->u.Out.u32Cookie = 0xffffffff;
1147	pReq->u.Out.u32SessionCookie = 0xffffffff;
1148	pReq->u.Out.u32SessionVersion = 0xffffffff;
1149	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1150	pReq->u.Out.pSession = NULL;
1151	pReq->u.Out.cFunctions = 0;
1152	pReq->Hdr.rc = VERR_PERMISSION_DENIED;
1153	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, VERR_PERMISSION_DENIED);
1154	return 0;
1155	}
1156	#endif
1157
1158	/*
1159	* Match the version.
1160	* The current logic is very simple, match the major interface version.
1161	*/
1162	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
1163	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
1164	{
1165	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1166	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
1167	pReq->u.Out.u32Cookie = 0xffffffff;
1168	pReq->u.Out.u32SessionCookie = 0xffffffff;
1169	pReq->u.Out.u32SessionVersion = 0xffffffff;
1170	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1171	pReq->u.Out.pSession = NULL;
1172	pReq->u.Out.cFunctions = 0;
1173	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1174	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1175	return 0;
1176	}
1177
1178	/*
1179	* Fill in return data and be gone.
1180	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
1181	* u32SessionVersion <= u32ReqVersion!
1182	*/
1183	/** @todo Somehow validate the client and negotiate a secure cookie... */
1184	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
1185	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
1186	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
1187	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1188	pReq->u.Out.pSession = pSession;
1189	pReq->u.Out.cFunctions = sizeof(g_aFunctions) / sizeof(g_aFunctions[0]);
1190	pReq->Hdr.rc = VINF_SUCCESS;
1191	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1192	return 0;
1193	}
1194
1195	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_QUERY_FUNCS(0)):
1196	{
1197	/* validate */
1198	PSUPQUERYFUNCS pReq = (PSUPQUERYFUNCS)pReqHdr;
1199	REQ_CHECK_SIZES_EX(SUP_IOCTL_QUERY_FUNCS, SUP_IOCTL_QUERY_FUNCS_SIZE_IN, SUP_IOCTL_QUERY_FUNCS_SIZE_OUT(RT_ELEMENTS(g_aFunctions)));
1200
1201	/* execute */
1202	pReq->u.Out.cFunctions = RT_ELEMENTS(g_aFunctions);
1203	memcpy(&pReq->u.Out.aFunctions[0], g_aFunctions, sizeof(g_aFunctions));
1204	pReq->Hdr.rc = VINF_SUCCESS;
1205	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1206	return 0;
1207	}
1208
1209	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_LOCK):
1210	{
1211	/* validate */
1212	PSUPPAGELOCK pReq = (PSUPPAGELOCK)pReqHdr;
1213	REQ_CHECK_SIZE_IN(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_IN);
1214	REQ_CHECK_SIZE_OUT(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_OUT(pReq->u.In.cPages));
1215	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.cPages > 0);
1216	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.pvR3 >= PAGE_SIZE);
1217
1218	/* execute */
1219	pReq->Hdr.rc = SUPR0LockMem(pSession, pReq->u.In.pvR3, pReq->u.In.cPages, &pReq->u.Out.aPages[0]);
1220	if (RT_FAILURE(pReq->Hdr.rc))
1221	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1222	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1223	return 0;
1224	}
1225
1226	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_UNLOCK):
1227	{
1228	/* validate */
1229	PSUPPAGEUNLOCK pReq = (PSUPPAGEUNLOCK)pReqHdr;
1230	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_UNLOCK);
1231
1232	/* execute */
1233	pReq->Hdr.rc = SUPR0UnlockMem(pSession, pReq->u.In.pvR3);
1234	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1235	return 0;
1236	}
1237
1238	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_ALLOC):
1239	{
1240	/* validate */
1241	PSUPCONTALLOC pReq = (PSUPCONTALLOC)pReqHdr;
1242	REQ_CHECK_SIZES(SUP_IOCTL_CONT_ALLOC);
1243
1244	/* execute */
1245	pReq->Hdr.rc = SUPR0ContAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.HCPhys);
1246	if (RT_FAILURE(pReq->Hdr.rc))
1247	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1248	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1249	return 0;
1250	}
1251
1252	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_FREE):
1253	{
1254	/* validate */
1255	PSUPCONTFREE pReq = (PSUPCONTFREE)pReqHdr;
1256	REQ_CHECK_SIZES(SUP_IOCTL_CONT_FREE);
1257
1258	/* execute */
1259	pReq->Hdr.rc = SUPR0ContFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1260	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1261	return 0;
1262	}
1263
1264	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_OPEN):
1265	{
1266	/* validate */
1267	PSUPLDROPEN pReq = (PSUPLDROPEN)pReqHdr;
1268	REQ_CHECK_SIZES(SUP_IOCTL_LDR_OPEN);
1269	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs > 0);
1270	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs < 16*_1M);
1271	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1272	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1273	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits < pReq->u.In.cbImageWithTabs);
1274	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.szName[0]);
1275	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1276	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, !supdrvCheckInvalidChar(pReq->u.In.szName, ";:()[]{}/\\\|&*%#@!~`\"'"));
1277	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szFilename, sizeof(pReq->u.In.szFilename)));
1278
1279	/* execute */
1280	pReq->Hdr.rc = supdrvIOCtl_LdrOpen(pDevExt, pSession, pReq);
1281	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1282	return 0;
1283	}
1284
1285	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOAD):
1286	{
1287	/* validate */
1288	PSUPLDRLOAD pReq = (PSUPLDRLOAD)pReqHdr;
1289	REQ_CHECK_EXPR(Name, pReq->Hdr.cbIn >= sizeof(*pReq));
1290	REQ_CHECK_SIZES_EX(SUP_IOCTL_LDR_LOAD, SUP_IOCTL_LDR_LOAD_SIZE_IN(pReq->u.In.cbImageWithTabs), SUP_IOCTL_LDR_LOAD_SIZE_OUT);
1291	REQ_CHECK_EXPR(SUP_IOCTL_LDR_LOAD, pReq->u.In.cSymbols <= 16384);
1292	REQ_CHECK_EXPR_FMT( !pReq->u.In.cSymbols
1293	\|\| ( pReq->u.In.offSymbols < pReq->u.In.cbImageWithTabs
1294	&& pReq->u.In.offSymbols + pReq->u.In.cSymbols * sizeof(SUPLDRSYM) <= pReq->u.In.cbImageWithTabs),
1295	("SUP_IOCTL_LDR_LOAD: offSymbols=%#lx cSymbols=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offSymbols,
1296	(long)pReq->u.In.cSymbols, (long)pReq->u.In.cbImageWithTabs));
1297	REQ_CHECK_EXPR_FMT( !pReq->u.In.cbStrTab
1298	\|\| ( pReq->u.In.offStrTab < pReq->u.In.cbImageWithTabs
1299	&& pReq->u.In.offStrTab + pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs
1300	&& pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs),
1301	("SUP_IOCTL_LDR_LOAD: offStrTab=%#lx cbStrTab=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offStrTab,
1302	(long)pReq->u.In.cbStrTab, (long)pReq->u.In.cbImageWithTabs));
1303
1304	if (pReq->u.In.cSymbols)
1305	{
1306	uint32_t i;
1307	PSUPLDRSYM paSyms = (PSUPLDRSYM)&pReq->u.In.abImage[pReq->u.In.offSymbols];
1308	for (i = 0; i < pReq->u.In.cSymbols; i++)
1309	{
1310	REQ_CHECK_EXPR_FMT(paSyms[i].offSymbol < pReq->u.In.cbImageWithTabs,
1311	("SUP_IOCTL_LDR_LOAD: sym #%ld: symb off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offSymbol, (long)pReq->u.In.cbImageWithTabs));
1312	REQ_CHECK_EXPR_FMT(paSyms[i].offName < pReq->u.In.cbStrTab,
1313	("SUP_IOCTL_LDR_LOAD: sym #%ld: name off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1314	REQ_CHECK_EXPR_FMT(RTStrEnd((char const *)&pReq->u.In.abImage[pReq->u.In.offStrTab + paSyms[i].offName],
1315	pReq->u.In.cbStrTab - paSyms[i].offName),
1316	("SUP_IOCTL_LDR_LOAD: sym #%ld: unterminated name! (%#lx / %#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1317	}
1318	}
1319
1320	/* execute */
1321	pReq->Hdr.rc = supdrvIOCtl_LdrLoad(pDevExt, pSession, pReq);
1322	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1323	return 0;
1324	}
1325
1326	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_FREE):
1327	{
1328	/* validate */
1329	PSUPLDRFREE pReq = (PSUPLDRFREE)pReqHdr;
1330	REQ_CHECK_SIZES(SUP_IOCTL_LDR_FREE);
1331
1332	/* execute */
1333	pReq->Hdr.rc = supdrvIOCtl_LdrFree(pDevExt, pSession, pReq);
1334	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1335	return 0;
1336	}
1337
1338	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_GET_SYMBOL):
1339	{
1340	/* validate */
1341	PSUPLDRGETSYMBOL pReq = (PSUPLDRGETSYMBOL)pReqHdr;
1342	REQ_CHECK_SIZES(SUP_IOCTL_LDR_GET_SYMBOL);
1343	REQ_CHECK_EXPR(SUP_IOCTL_LDR_GET_SYMBOL, RTStrEnd(pReq->u.In.szSymbol, sizeof(pReq->u.In.szSymbol)));
1344
1345	/* execute */
1346	pReq->Hdr.rc = supdrvIOCtl_LdrGetSymbol(pDevExt, pSession, pReq);
1347	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1348	return 0;
1349	}
1350
1351	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0(0)):
1352	{
1353	/* validate */
1354	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1355	Log4(("SUP_IOCTL_CALL_VMMR0: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1356	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1357
1358	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_VMMR0_SIZE(0))
1359	{
1360	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(0), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(0));
1361
1362	/* execute */
1363	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1364	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg, pSession);
1365	else
1366	pReq->Hdr.rc = VERR_WRONG_ORDER;
1367	}
1368	else
1369	{
1370	PSUPVMMR0REQHDR pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1371	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR)),
1372	("SUP_IOCTL_CALL_VMMR0: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR))));
1373	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1374	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(pVMMReq->cbReq));
1375
1376	/* execute */
1377	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1378	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1379	else
1380	pReq->Hdr.rc = VERR_WRONG_ORDER;
1381	}
1382
1383	if ( RT_FAILURE(pReq->Hdr.rc)
1384	&& pReq->Hdr.rc != VERR_INTERRUPTED
1385	&& pReq->Hdr.rc != VERR_TIMEOUT)
1386	Log(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1387	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1388	else
1389	Log4(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1390	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1391	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1392	return 0;
1393	}
1394
1395	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0_BIG):
1396	{
1397	/* validate */
1398	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1399	PSUPVMMR0REQHDR pVMMReq;
1400	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1401	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1402
1403	pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1404	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR)),
1405	("SUP_IOCTL_CALL_VMMR0_BIG: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR))));
1406	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0_BIG, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1407	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));
1408
1409	/* execute */
1410	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1411	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1412	else
1413	pReq->Hdr.rc = VERR_WRONG_ORDER;
1414
1415	if ( RT_FAILURE(pReq->Hdr.rc)
1416	&& pReq->Hdr.rc != VERR_INTERRUPTED
1417	&& pReq->Hdr.rc != VERR_TIMEOUT)
1418	Log(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1419	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1420	else
1421	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1422	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1423	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1424	return 0;
1425	}
1426
1427	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_PAGING_MODE):
1428	{
1429	/* validate */
1430	PSUPGETPAGINGMODE pReq = (PSUPGETPAGINGMODE)pReqHdr;
1431	REQ_CHECK_SIZES(SUP_IOCTL_GET_PAGING_MODE);
1432
1433	/* execute */
1434	pReq->Hdr.rc = VINF_SUCCESS;
1435	pReq->u.Out.enmMode = SUPR0GetPagingMode();
1436	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1437	return 0;
1438	}
1439
1440	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_ALLOC):
1441	{
1442	/* validate */
1443	PSUPLOWALLOC pReq = (PSUPLOWALLOC)pReqHdr;
1444	REQ_CHECK_EXPR(SUP_IOCTL_LOW_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_LOW_ALLOC_SIZE_IN);
1445	REQ_CHECK_SIZES_EX(SUP_IOCTL_LOW_ALLOC, SUP_IOCTL_LOW_ALLOC_SIZE_IN, SUP_IOCTL_LOW_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1446
1447	/* execute */
1448	pReq->Hdr.rc = SUPR0LowAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1449	if (RT_FAILURE(pReq->Hdr.rc))
1450	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1451	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1452	return 0;
1453	}
1454
1455	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_FREE):
1456	{
1457	/* validate */
1458	PSUPLOWFREE pReq = (PSUPLOWFREE)pReqHdr;
1459	REQ_CHECK_SIZES(SUP_IOCTL_LOW_FREE);
1460
1461	/* execute */
1462	pReq->Hdr.rc = SUPR0LowFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1463	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1464	return 0;
1465	}
1466
1467	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_MAP):
1468	{
1469	/* validate */
1470	PSUPGIPMAP pReq = (PSUPGIPMAP)pReqHdr;
1471	REQ_CHECK_SIZES(SUP_IOCTL_GIP_MAP);
1472
1473	/* execute */
1474	pReq->Hdr.rc = SUPR0GipMap(pSession, &pReq->u.Out.pGipR3, &pReq->u.Out.HCPhysGip);
1475	if (RT_SUCCESS(pReq->Hdr.rc))
1476	pReq->u.Out.pGipR0 = pDevExt->pGip;
1477	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1478	return 0;
1479	}
1480
1481	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_UNMAP):
1482	{
1483	/* validate */
1484	PSUPGIPUNMAP pReq = (PSUPGIPUNMAP)pReqHdr;
1485	REQ_CHECK_SIZES(SUP_IOCTL_GIP_UNMAP);
1486
1487	/* execute */
1488	pReq->Hdr.rc = SUPR0GipUnmap(pSession);
1489	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1490	return 0;
1491	}
1492
1493	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SET_VM_FOR_FAST):
1494	{
1495	/* validate */
1496	PSUPSETVMFORFAST pReq = (PSUPSETVMFORFAST)pReqHdr;
1497	REQ_CHECK_SIZES(SUP_IOCTL_SET_VM_FOR_FAST);
1498	REQ_CHECK_EXPR_FMT( !pReq->u.In.pVMR0
1499	\|\| ( VALID_PTR(pReq->u.In.pVMR0)
1500	&& !((uintptr_t)pReq->u.In.pVMR0 & (PAGE_SIZE - 1))),
1501	("SUP_IOCTL_SET_VM_FOR_FAST: pVMR0=%p!\n", pReq->u.In.pVMR0));
1502	/* execute */
1503	pSession->pVM = pReq->u.In.pVMR0;
1504	pReq->Hdr.rc = VINF_SUCCESS;
1505	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1506	return 0;
1507	}
1508
1509	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_ALLOC_EX):
1510	{
1511	/* validate */
1512	PSUPPAGEALLOCEX pReq = (PSUPPAGEALLOCEX)pReqHdr;
1513	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC_EX, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_EX_SIZE_IN);
1514	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));
1515	REQ_CHECK_EXPR_FMT(pReq->u.In.fKernelMapping \|\| pReq->u.In.fUserMapping,
1516	("SUP_IOCTL_PAGE_ALLOC_EX: No mapping requested!\n"));
1517	REQ_CHECK_EXPR_FMT(pReq->u.In.fUserMapping,
1518	("SUP_IOCTL_PAGE_ALLOC_EX: Must have user mapping!\n"));
1519	REQ_CHECK_EXPR_FMT(!pReq->u.In.fReserved0 && !pReq->u.In.fReserved1,
1520	("SUP_IOCTL_PAGE_ALLOC_EX: fReserved0=%d fReserved1=%d\n", pReq->u.In.fReserved0, pReq->u.In.fReserved1));
1521
1522	/* execute */
1523	pReq->Hdr.rc = SUPR0PageAllocEx(pSession, pReq->u.In.cPages, 0 /* fFlags */,
1524	pReq->u.In.fUserMapping ? &pReq->u.Out.pvR3 : NULL,
1525	pReq->u.In.fKernelMapping ? &pReq->u.Out.pvR0 : NULL,
1526	&pReq->u.Out.aPages[0]);
1527	if (RT_FAILURE(pReq->Hdr.rc))
1528	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1529	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1530	return 0;
1531	}
1532
1533	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_MAP_KERNEL):
1534	{
1535	/* validate */
1536	PSUPPAGEMAPKERNEL pReq = (PSUPPAGEMAPKERNEL)pReqHdr;
1537	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_MAP_KERNEL);
1538	REQ_CHECK_EXPR_FMT(!pReq->u.In.fFlags, ("SUP_IOCTL_PAGE_MAP_KERNEL: fFlags=%#x! MBZ\n", pReq->u.In.fFlags));
1539	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_MAP_KERNEL: offSub=%#x\n", pReq->u.In.offSub));
1540	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1541	("SUP_IOCTL_PAGE_MAP_KERNEL: cbSub=%#x\n", pReq->u.In.cbSub));
1542
1543	/* execute */
1544	pReq->Hdr.rc = SUPR0PageMapKernel(pSession, pReq->u.In.pvR3, pReq->u.In.offSub, pReq->u.In.cbSub,
1545	pReq->u.In.fFlags, &pReq->u.Out.pvR0);
1546	if (RT_FAILURE(pReq->Hdr.rc))
1547	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1548	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1549	return 0;
1550	}
1551
1552	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_PROTECT):
1553	{
1554	/* validate */
1555	PSUPPAGEPROTECT pReq = (PSUPPAGEPROTECT)pReqHdr;
1556	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_PROTECT);
1557	REQ_CHECK_EXPR_FMT(!(pReq->u.In.fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)),
1558	("SUP_IOCTL_PAGE_PROTECT: fProt=%#x!\n", pReq->u.In.fProt));
1559	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_PROTECT: offSub=%#x\n", pReq->u.In.offSub));
1560	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1561	("SUP_IOCTL_PAGE_PROTECT: cbSub=%#x\n", pReq->u.In.cbSub));
1562
1563	/* execute */
1564	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);
1565	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1566	return 0;
1567	}
1568
1569	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_FREE):
1570	{
1571	/* validate */
1572	PSUPPAGEFREE pReq = (PSUPPAGEFREE)pReqHdr;
1573	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_FREE);
1574
1575	/* execute */
1576	pReq->Hdr.rc = SUPR0PageFree(pSession, pReq->u.In.pvR3);
1577	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1578	return 0;
1579	}
1580
1581	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_SERVICE(0)):
1582	{
1583	/* validate */
1584	PSUPCALLSERVICE pReq = (PSUPCALLSERVICE)pReqHdr;
1585	Log4(("SUP_IOCTL_CALL_SERVICE: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1586	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1587
1588	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
1589	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(0), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(0));
1590	else
1591	{
1592	PSUPR0SERVICEREQHDR pSrvReq = (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0];
1593	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR)),
1594	("SUP_IOCTL_CALL_SERVICE: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR))));
1595	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, pSrvReq->u32Magic == SUPR0SERVICEREQHDR_MAGIC);
1596	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(pSrvReq->cbReq), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(pSrvReq->cbReq));
1597	}
1598	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1599
1600	/* execute */
1601	pReq->Hdr.rc = supdrvIOCtl_CallServiceModule(pDevExt, pSession, pReq);
1602	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1603	return 0;
1604	}
1605
1606	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOGGER_SETTINGS(0)):
1607	{
1608	/* validate */
1609	PSUPLOGGERSETTINGS pReq = (PSUPLOGGERSETTINGS)pReqHdr;
1610	size_t cbStrTab;
1611	REQ_CHECK_SIZE_OUT(SUP_IOCTL_LOGGER_SETTINGS, SUP_IOCTL_LOGGER_SETTINGS_SIZE_OUT);
1612	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->Hdr.cbIn >= SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(1));
1613	cbStrTab = pReq->Hdr.cbIn - SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(0);
1614	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offGroups < cbStrTab);
1615	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offFlags < cbStrTab);
1616	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offDestination < cbStrTab);
1617	REQ_CHECK_EXPR_FMT(pReq->u.In.szStrings[cbStrTab - 1] == '\0',
1618	("SUP_IOCTL_LOGGER_SETTINGS: cbIn=%#x cbStrTab=%#zx LastChar=%d\n",
1619	pReq->Hdr.cbIn, cbStrTab, pReq->u.In.szStrings[cbStrTab - 1]));
1620	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhich <= SUPLOGGERSETTINGS_WHICH_RELEASE);
1621	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhat <= SUPLOGGERSETTINGS_WHAT_DESTROY);
1622
1623	/* execute */
1624	pReq->Hdr.rc = supdrvIOCtl_LoggerSettings(pDevExt, pSession, pReq);
1625	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1626	return 0;
1627	}
1628
1629	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP2):
1630	{
1631	/* validate */
1632	PSUPSEMOP2 pReq = (PSUPSEMOP2)pReqHdr;
1633	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP2, SUP_IOCTL_SEM_OP2_SIZE_IN, SUP_IOCTL_SEM_OP2_SIZE_OUT);
1634	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP2, pReq->u.In.uReserved == 0);
1635
1636	/* execute */
1637	switch (pReq->u.In.uType)
1638	{
1639	case SUP_SEM_TYPE_EVENT:
1640	{
1641	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1642	switch (pReq->u.In.uOp)
1643	{
1644	case SUPSEMOP2_WAIT_MS_REL:
1645	pReq->Hdr.rc = SUPSemEventWaitNoResume(pSession, hEvent, pReq->u.In.uArg.cRelMsTimeout);
1646	break;
1647	case SUPSEMOP2_WAIT_NS_ABS:
1648	pReq->Hdr.rc = SUPSemEventWaitNsAbsIntr(pSession, hEvent, pReq->u.In.uArg.uAbsNsTimeout);
1649	break;
1650	case SUPSEMOP2_WAIT_NS_REL:
1651	pReq->Hdr.rc = SUPSemEventWaitNsRelIntr(pSession, hEvent, pReq->u.In.uArg.cRelNsTimeout);
1652	break;
1653	case SUPSEMOP2_SIGNAL:
1654	pReq->Hdr.rc = SUPSemEventSignal(pSession, hEvent);
1655	break;
1656	case SUPSEMOP2_CLOSE:
1657	pReq->Hdr.rc = SUPSemEventClose(pSession, hEvent);
1658	break;
1659	case SUPSEMOP2_RESET:
1660	default:
1661	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1662	break;
1663	}
1664	break;
1665	}
1666
1667	case SUP_SEM_TYPE_EVENT_MULTI:
1668	{
1669	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1670	switch (pReq->u.In.uOp)
1671	{
1672	case SUPSEMOP2_WAIT_MS_REL:
1673	pReq->Hdr.rc = SUPSemEventMultiWaitNoResume(pSession, hEventMulti, pReq->u.In.uArg.cRelMsTimeout);
1674	break;
1675	case SUPSEMOP2_WAIT_NS_ABS:
1676	pReq->Hdr.rc = SUPSemEventMultiWaitNsAbsIntr(pSession, hEventMulti, pReq->u.In.uArg.uAbsNsTimeout);
1677	break;
1678	case SUPSEMOP2_WAIT_NS_REL:
1679	pReq->Hdr.rc = SUPSemEventMultiWaitNsRelIntr(pSession, hEventMulti, pReq->u.In.uArg.cRelNsTimeout);
1680	break;
1681	case SUPSEMOP2_SIGNAL:
1682	pReq->Hdr.rc = SUPSemEventMultiSignal(pSession, hEventMulti);
1683	break;
1684	case SUPSEMOP2_CLOSE:
1685	pReq->Hdr.rc = SUPSemEventMultiClose(pSession, hEventMulti);
1686	break;
1687	case SUPSEMOP2_RESET:
1688	pReq->Hdr.rc = SUPSemEventMultiReset(pSession, hEventMulti);
1689	break;
1690	default:
1691	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1692	break;
1693	}
1694	break;
1695	}
1696
1697	default:
1698	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1699	break;
1700	}
1701	return 0;
1702	}
1703
1704	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP3):
1705	{
1706	/* validate */
1707	PSUPSEMOP3 pReq = (PSUPSEMOP3)pReqHdr;
1708	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP3, SUP_IOCTL_SEM_OP3_SIZE_IN, SUP_IOCTL_SEM_OP3_SIZE_OUT);
1709	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, pReq->u.In.u32Reserved == 0 && pReq->u.In.u64Reserved == 0);
1710
1711	/* execute */
1712	switch (pReq->u.In.uType)
1713	{
1714	case SUP_SEM_TYPE_EVENT:
1715	{
1716	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1717	switch (pReq->u.In.uOp)
1718	{
1719	case SUPSEMOP3_CREATE:
1720	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1721	pReq->Hdr.rc = SUPSemEventCreate(pSession, &hEvent);
1722	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEvent;
1723	break;
1724	case SUPSEMOP3_GET_RESOLUTION:
1725	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1726	pReq->Hdr.rc = VINF_SUCCESS;
1727	pReq->Hdr.cbOut = sizeof(*pReq);
1728	pReq->u.Out.cNsResolution = SUPSemEventGetResolution(pSession);
1729	break;
1730	default:
1731	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1732	break;
1733	}
1734	break;
1735	}
1736
1737	case SUP_SEM_TYPE_EVENT_MULTI:
1738	{
1739	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1740	switch (pReq->u.In.uOp)
1741	{
1742	case SUPSEMOP3_CREATE:
1743	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1744	pReq->Hdr.rc = SUPSemEventMultiCreate(pSession, &hEventMulti);
1745	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEventMulti;
1746	break;
1747	case SUPSEMOP3_GET_RESOLUTION:
1748	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1749	pReq->Hdr.rc = VINF_SUCCESS;
1750	pReq->u.Out.cNsResolution = SUPSemEventMultiGetResolution(pSession);
1751	break;
1752	default:
1753	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1754	break;
1755	}
1756	break;
1757	}
1758
1759	default:
1760	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1761	break;
1762	}
1763	return 0;
1764	}
1765
1766	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
1767	{
1768	/* validate */
1769	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
1770	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
1771	REQ_CHECK_EXPR(SUP_IOCTL_VT_CAPS, pReq->Hdr.cbIn <= SUP_IOCTL_VT_CAPS_SIZE_IN);
1772
1773	/* execute */
1774	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.Caps);
1775	if (RT_FAILURE(pReq->Hdr.rc))
1776	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1777	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1778	return 0;
1779	}
1780
1781	default:
1782	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1783	break;
1784	}
1785	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_GENERAL_FAILURE, VERR_GENERAL_FAILURE);
1786	return VERR_GENERAL_FAILURE;
1787	}
1788
1789
1790	/**
1791	* Inter-Driver Communication (IDC) worker.
1792	*
1793	* @returns VBox status code.
1794	* @retval VINF_SUCCESS on success.
1795	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1796	* @retval VERR_NOT_SUPPORTED if the request isn't supported.
1797	*
1798	* @param uReq The request (function) code.
1799	* @param pDevExt Device extention.
1800	* @param pSession Session data.
1801	* @param pReqHdr The request header.
1802	*/
1803	int VBOXCALL supdrvIDC(uintptr_t uReq, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQHDR pReqHdr)
1804	{
1805	/*
1806	* The OS specific code has already validated the pSession
1807	* pointer, and the request size being greater or equal to
1808	* size of the header.
1809	*
1810	* So, just check that pSession is a kernel context session.
1811	*/
1812	if (RT_UNLIKELY( pSession
1813	&& pSession->R0Process != NIL_RTR0PROCESS))
1814	return VERR_INVALID_PARAMETER;
1815
1816	/*
1817	* Validation macro.
1818	*/
1819	#define REQ_CHECK_IDC_SIZE(Name, cbExpect) \
1820	do { \
1821	if (RT_UNLIKELY(pReqHdr->cb != (cbExpect))) \
1822	{ \
1823	OSDBGPRINT(( #Name ": Invalid input/output sizes. cb=%ld expected %ld.\n", \
1824	(long)pReqHdr->cb, (long)(cbExpect))); \
1825	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1826	} \
1827	} while (0)
1828
1829	switch (uReq)
1830	{
1831	case SUPDRV_IDC_REQ_CONNECT:
1832	{
1833	PSUPDRVIDCREQCONNECT pReq = (PSUPDRVIDCREQCONNECT)pReqHdr;
1834	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_CONNECT, sizeof(*pReq));
1835
1836	/*
1837	* Validate the cookie and other input.
1838	*/
1839	if (pReq->Hdr.pSession != NULL)
1840	{
1841	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: pSession=%p expected NULL!\n", pReq->Hdr.pSession));
1842	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1843	}
1844	if (pReq->u.In.u32MagicCookie != SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE)
1845	{
1846	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: u32MagicCookie=%#x expected %#x!\n",
1847	(unsigned)pReq->u.In.u32MagicCookie, (unsigned)SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE));
1848	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1849	}
1850	if ( pReq->u.In.uMinVersion > pReq->u.In.uReqVersion
1851	\|\| (pReq->u.In.uMinVersion & UINT32_C(0xffff0000)) != (pReq->u.In.uReqVersion & UINT32_C(0xffff0000)))
1852	{
1853	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: uMinVersion=%#x uMaxVersion=%#x doesn't match!\n",
1854	pReq->u.In.uMinVersion, pReq->u.In.uReqVersion));
1855	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1856	}
1857
1858	/*
1859	* Match the version.
1860	* The current logic is very simple, match the major interface version.
1861	*/
1862	if ( pReq->u.In.uMinVersion > SUPDRV_IDC_VERSION
1863	\|\| (pReq->u.In.uMinVersion & 0xffff0000) != (SUPDRV_IDC_VERSION & 0xffff0000))
1864	{
1865	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1866	pReq->u.In.uReqVersion, pReq->u.In.uMinVersion, (unsigned)SUPDRV_IDC_VERSION));
1867	pReq->u.Out.pSession = NULL;
1868	pReq->u.Out.uSessionVersion = 0xffffffff;
1869	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1870	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1871	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1872	return VINF_SUCCESS;
1873	}
1874
1875	pReq->u.Out.pSession = NULL;
1876	pReq->u.Out.uSessionVersion = SUPDRV_IDC_VERSION;
1877	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1878	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1879
1880	/*
1881	* On NT we will already have a session associated with the
1882	* client, just like with the SUP_IOCTL_COOKIE request, while
1883	* the other doesn't.
1884	*/
1885	#ifdef RT_OS_WINDOWS
1886	pReq->Hdr.rc = VINF_SUCCESS;
1887	#else
1888	AssertReturn(!pSession, VERR_INTERNAL_ERROR);
1889	pReq->Hdr.rc = supdrvCreateSession(pDevExt, false /* fUser */, &pSession);
1890	if (RT_FAILURE(pReq->Hdr.rc))
1891	{
1892	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: failed to create session, rc=%d\n", pReq->Hdr.rc));
1893	return VINF_SUCCESS;
1894	}
1895	#endif
1896
1897	pReq->u.Out.pSession = pSession;
1898	pReq->Hdr.pSession = pSession;
1899
1900	return VINF_SUCCESS;
1901	}
1902
1903	case SUPDRV_IDC_REQ_DISCONNECT:
1904	{
1905	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_DISCONNECT, sizeof(*pReqHdr));
1906
1907	#ifdef RT_OS_WINDOWS
1908	/* Windows will destroy the session when the file object is destroyed. */
1909	#else
1910	supdrvCloseSession(pDevExt, pSession);
1911	#endif
1912	return pReqHdr->rc = VINF_SUCCESS;
1913	}
1914
1915	case SUPDRV_IDC_REQ_GET_SYMBOL:
1916	{
1917	PSUPDRVIDCREQGETSYM pReq = (PSUPDRVIDCREQGETSYM)pReqHdr;
1918	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_GET_SYMBOL, sizeof(*pReq));
1919
1920	pReq->Hdr.rc = supdrvIDC_LdrGetSymbol(pDevExt, pSession, pReq);
1921	return VINF_SUCCESS;
1922	}
1923
1924	case SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY:
1925	{
1926	PSUPDRVIDCREQCOMPREGFACTORY pReq = (PSUPDRVIDCREQCOMPREGFACTORY)pReqHdr;
1927	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY, sizeof(*pReq));
1928
1929	pReq->Hdr.rc = SUPR0ComponentRegisterFactory(pSession, pReq->u.In.pFactory);
1930	return VINF_SUCCESS;
1931	}
1932
1933	case SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY:
1934	{
1935	PSUPDRVIDCREQCOMPDEREGFACTORY pReq = (PSUPDRVIDCREQCOMPDEREGFACTORY)pReqHdr;
1936	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY, sizeof(*pReq));
1937
1938	pReq->Hdr.rc = SUPR0ComponentDeregisterFactory(pSession, pReq->u.In.pFactory);
1939	return VINF_SUCCESS;
1940	}
1941
1942	default:
1943	Log(("Unknown IDC %#lx\n", (long)uReq));
1944	break;
1945	}
1946
1947	#undef REQ_CHECK_IDC_SIZE
1948	return VERR_NOT_SUPPORTED;
1949	}
1950
1951
1952	/**
1953	* Register a object for reference counting.
1954	* The object is registered with one reference in the specified session.
1955	*
1956	* @returns Unique identifier on success (pointer).
1957	* All future reference must use this identifier.
1958	* @returns NULL on failure.
1959	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
1960	* @param pvUser1 The first user argument.
1961	* @param pvUser2 The second user argument.
1962	*/
1963	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
1964	{
1965	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1966	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1967	PSUPDRVOBJ pObj;
1968	PSUPDRVUSAGE pUsage;
1969
1970	/*
1971	* Validate the input.
1972	*/
1973	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
1974	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
1975	AssertPtrReturn(pfnDestructor, NULL);
1976
1977	/*
1978	* Allocate and initialize the object.
1979	*/
1980	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
1981	if (!pObj)
1982	return NULL;
1983	pObj->u32Magic = SUPDRVOBJ_MAGIC;
1984	pObj->enmType = enmType;
1985	pObj->pNext = NULL;
1986	pObj->cUsage = 1;
1987	pObj->pfnDestructor = pfnDestructor;
1988	pObj->pvUser1 = pvUser1;
1989	pObj->pvUser2 = pvUser2;
1990	pObj->CreatorUid = pSession->Uid;
1991	pObj->CreatorGid = pSession->Gid;
1992	pObj->CreatorProcess= pSession->Process;
1993	supdrvOSObjInitCreator(pObj, pSession);
1994
1995	/*
1996	* Allocate the usage record.
1997	* (We keep freed usage records around to simplify SUPR0ObjAddRefEx().)
1998	*/
1999	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2000
2001	pUsage = pDevExt->pUsageFree;
2002	if (pUsage)
2003	pDevExt->pUsageFree = pUsage->pNext;
2004	else
2005	{
2006	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2007	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
2008	if (!pUsage)
2009	{
2010	RTMemFree(pObj);
2011	return NULL;
2012	}
2013	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2014	}
2015
2016	/*
2017	* Insert the object and create the session usage record.
2018	*/
2019	/* The object. */
2020	pObj->pNext = pDevExt->pObjs;
2021	pDevExt->pObjs = pObj;
2022
2023	/* The session record. */
2024	pUsage->cUsage = 1;
2025	pUsage->pObj = pObj;
2026	pUsage->pNext = pSession->pUsage;
2027	/* Log2(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext)); */
2028	pSession->pUsage = pUsage;
2029
2030	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2031
2032	Log(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
2033	return pObj;
2034	}
2035
2036
2037	/**
2038	* Increment the reference counter for the object associating the reference
2039	* with the specified session.
2040	*
2041	* @returns IPRT status code.
2042	* @param pvObj The identifier returned by SUPR0ObjRegister().
2043	* @param pSession The session which is referencing the object.
2044	*
2045	* @remarks The caller should not own any spinlocks and must carefully protect
2046	* itself against potential race with the destructor so freed memory
2047	* isn't accessed here.
2048	*/
2049	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
2050	{
2051	return SUPR0ObjAddRefEx(pvObj, pSession, false /* fNoBlocking */);
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	* @retval VERR_TRY_AGAIN if fNoBlocking was set and a new usage record
2061	* couldn't be allocated. (If you see this you're not doing the right
2062	* thing and it won't ever work reliably.)
2063	*
2064	* @param pvObj The identifier returned by SUPR0ObjRegister().
2065	* @param pSession The session which is referencing the object.
2066	* @param fNoBlocking Set if it's not OK to block. Never try to make the
2067	* first reference to an object in a session with this
2068	* argument set.
2069	*
2070	* @remarks The caller should not own any spinlocks and must carefully protect
2071	* itself against potential race with the destructor so freed memory
2072	* isn't accessed here.
2073	*/
2074	SUPR0DECL(int) SUPR0ObjAddRefEx(void *pvObj, PSUPDRVSESSION pSession, bool fNoBlocking)
2075	{
2076	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2077	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2078	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2079	int rc = VINF_SUCCESS;
2080	PSUPDRVUSAGE pUsagePre;
2081	PSUPDRVUSAGE pUsage;
2082
2083	/*
2084	* Validate the input.
2085	* Be ready for the destruction race (someone might be stuck in the
2086	* destructor waiting a lock we own).
2087	*/
2088	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2089	AssertPtrReturn(pObj, VERR_INVALID_POINTER);
2090	AssertMsgReturn(pObj->u32Magic == SUPDRVOBJ_MAGIC \|\| pObj->u32Magic == SUPDRVOBJ_MAGIC_DEAD,
2091	("Invalid pvObj=%p magic=%#x (expected %#x or %#x)\n", pvObj, pObj->u32Magic, SUPDRVOBJ_MAGIC, SUPDRVOBJ_MAGIC_DEAD),
2092	VERR_INVALID_PARAMETER);
2093
2094	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2095
2096	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2097	{
2098	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2099
2100	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2101	return VERR_WRONG_ORDER;
2102	}
2103
2104	/*
2105	* Preallocate the usage record if we can.
2106	*/
2107	pUsagePre = pDevExt->pUsageFree;
2108	if (pUsagePre)
2109	pDevExt->pUsageFree = pUsagePre->pNext;
2110	else if (!fNoBlocking)
2111	{
2112	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2113	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
2114	if (!pUsagePre)
2115	return VERR_NO_MEMORY;
2116
2117	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2118	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2119	{
2120	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2121
2122	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2123	return VERR_WRONG_ORDER;
2124	}
2125	}
2126
2127	/*
2128	* Reference the object.
2129	*/
2130	pObj->cUsage++;
2131
2132	/*
2133	* Look for the session record.
2134	*/
2135	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
2136	{
2137	/Log(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2138	if (pUsage->pObj == pObj)
2139	break;
2140	}
2141	if (pUsage)
2142	pUsage->cUsage++;
2143	else if (pUsagePre)
2144	{
2145	/* create a new session record. */
2146	pUsagePre->cUsage = 1;
2147	pUsagePre->pObj = pObj;
2148	pUsagePre->pNext = pSession->pUsage;
2149	pSession->pUsage = pUsagePre;
2150	/Log(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));/
2151
2152	pUsagePre = NULL;
2153	}
2154	else
2155	{
2156	pObj->cUsage--;
2157	rc = VERR_TRY_AGAIN;
2158	}
2159
2160	/*
2161	* Put any unused usage record into the free list..
2162	*/
2163	if (pUsagePre)
2164	{
2165	pUsagePre->pNext = pDevExt->pUsageFree;
2166	pDevExt->pUsageFree = pUsagePre;
2167	}
2168
2169	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2170
2171	return rc;
2172	}
2173
2174
2175	/**
2176	* Decrement / destroy a reference counter record for an object.
2177	*
2178	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
2179	*
2180	* @returns IPRT status code.
2181	* @retval VINF_SUCCESS if not destroyed.
2182	* @retval VINF_OBJECT_DESTROYED if it's destroyed by this release call.
2183	* @retval VERR_INVALID_PARAMETER if the object isn't valid. Will assert in
2184	* string builds.
2185	*
2186	* @param pvObj The identifier returned by SUPR0ObjRegister().
2187	* @param pSession The session which is referencing the object.
2188	*/
2189	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
2190	{
2191	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2192	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2193	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2194	int rc = VERR_INVALID_PARAMETER;
2195	PSUPDRVUSAGE pUsage;
2196	PSUPDRVUSAGE pUsagePrev;
2197
2198	/*
2199	* Validate the input.
2200	*/
2201	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2202	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2203	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2204	VERR_INVALID_PARAMETER);
2205
2206	/*
2207	* Acquire the spinlock and look for the usage record.
2208	*/
2209	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2210
2211	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
2212	pUsage;
2213	pUsagePrev = pUsage, pUsage = pUsage->pNext)
2214	{
2215	/Log2(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2216	if (pUsage->pObj == pObj)
2217	{
2218	rc = VINF_SUCCESS;
2219	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
2220	if (pUsage->cUsage > 1)
2221	{
2222	pObj->cUsage--;
2223	pUsage->cUsage--;
2224	}
2225	else
2226	{
2227	/*
2228	* Free the session record.
2229	*/
2230	if (pUsagePrev)
2231	pUsagePrev->pNext = pUsage->pNext;
2232	else
2233	pSession->pUsage = pUsage->pNext;
2234	pUsage->pNext = pDevExt->pUsageFree;
2235	pDevExt->pUsageFree = pUsage;
2236
2237	/* What about the object? */
2238	if (pObj->cUsage > 1)
2239	pObj->cUsage--;
2240	else
2241	{
2242	/*
2243	* Object is to be destroyed, unlink it.
2244	*/
2245	pObj->u32Magic = SUPDRVOBJ_MAGIC_DEAD;
2246	rc = VINF_OBJECT_DESTROYED;
2247	if (pDevExt->pObjs == pObj)
2248	pDevExt->pObjs = pObj->pNext;
2249	else
2250	{
2251	PSUPDRVOBJ pObjPrev;
2252	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
2253	if (pObjPrev->pNext == pObj)
2254	{
2255	pObjPrev->pNext = pObj->pNext;
2256	break;
2257	}
2258	Assert(pObjPrev);
2259	}
2260	}
2261	}
2262	break;
2263	}
2264	}
2265
2266	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2267
2268	/*
2269	* Call the destructor and free the object if required.
2270	*/
2271	if (rc == VINF_OBJECT_DESTROYED)
2272	{
2273	Log(("SUPR0ObjRelease: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
2274	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
2275	if (pObj->pfnDestructor)
2276	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
2277	RTMemFree(pObj);
2278	}
2279
2280	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
2281	return rc;
2282	}
2283
2284
2285	/**
2286	* Verifies that the current process can access the specified object.
2287	*
2288	* @returns The following IPRT status code:
2289	* @retval VINF_SUCCESS if access was granted.
2290	* @retval VERR_PERMISSION_DENIED if denied access.
2291	* @retval VERR_INVALID_PARAMETER if invalid parameter.
2292	*
2293	* @param pvObj The identifier returned by SUPR0ObjRegister().
2294	* @param pSession The session which wishes to access the object.
2295	* @param pszObjName Object string name. This is optional and depends on the object type.
2296	*
2297	* @remark The caller is responsible for making sure the object isn't removed while
2298	* we're inside this function. If uncertain about this, just call AddRef before calling us.
2299	*/
2300	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
2301	{
2302	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2303	int rc;
2304
2305	/*
2306	* Validate the input.
2307	*/
2308	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2309	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2310	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2311	VERR_INVALID_PARAMETER);
2312
2313	/*
2314	* Check access. (returns true if a decision has been made.)
2315	*/
2316	rc = VERR_INTERNAL_ERROR;
2317	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
2318	return rc;
2319
2320	/*
2321	* Default policy is to allow the user to access his own
2322	* stuff but nothing else.
2323	*/
2324	if (pObj->CreatorUid == pSession->Uid)
2325	return VINF_SUCCESS;
2326	return VERR_PERMISSION_DENIED;
2327	}
2328
2329
2330	/**
2331	* Lock pages.
2332	*
2333	* @returns IPRT status code.
2334	* @param pSession Session to which the locked memory should be associated.
2335	* @param pvR3 Start of the memory range to lock.
2336	* This must be page aligned.
2337	* @param cPages Number of pages to lock.
2338	* @param paPages Where to put the physical addresses of locked memory.
2339	*/
2340	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
2341	{
2342	int rc;
2343	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2344	const size_t cb = (size_t)cPages << PAGE_SHIFT;
2345	LogFlow(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
2346
2347	/*
2348	* Verify input.
2349	*/
2350	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2351	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
2352	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
2353	\|\| !pvR3)
2354	{
2355	Log(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
2356	return VERR_INVALID_PARAMETER;
2357	}
2358
2359	/*
2360	* Let IPRT do the job.
2361	*/
2362	Mem.eType = MEMREF_TYPE_LOCKED;
2363	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTMEM_PROT_READ \| RTMEM_PROT_WRITE, RTR0ProcHandleSelf());
2364	if (RT_SUCCESS(rc))
2365	{
2366	uint32_t iPage = cPages;
2367	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
2368	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
2369
2370	while (iPage-- > 0)
2371	{
2372	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2373	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
2374	{
2375	AssertMsgFailed(("iPage=%d\n", iPage));
2376	rc = VERR_INTERNAL_ERROR;
2377	break;
2378	}
2379	}
2380	if (RT_SUCCESS(rc))
2381	rc = supdrvMemAdd(&Mem, pSession);
2382	if (RT_FAILURE(rc))
2383	{
2384	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
2385	AssertRC(rc2);
2386	}
2387	}
2388
2389	return rc;
2390	}
2391
2392
2393	/**
2394	* Unlocks the memory pointed to by pv.
2395	*
2396	* @returns IPRT status code.
2397	* @param pSession Session to which the memory was locked.
2398	* @param pvR3 Memory to unlock.
2399	*/
2400	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2401	{
2402	LogFlow(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2403	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2404	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
2405	}
2406
2407
2408	/**
2409	* Allocates a chunk of page aligned memory with contiguous and fixed physical
2410	* backing.
2411	*
2412	* @returns IPRT status code.
2413	* @param pSession Session data.
2414	* @param cPages Number of pages to allocate.
2415	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
2416	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
2417	* @param pHCPhys Where to put the physical address of allocated memory.
2418	*/
2419	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
2420	{
2421	int rc;
2422	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2423	LogFlow(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
2424
2425	/*
2426	* Validate input.
2427	*/
2428	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2429	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
2430	{
2431	Log(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
2432	pSession, ppvR0, ppvR3, pHCPhys));
2433	return VERR_INVALID_PARAMETER;
2434
2435	}
2436	if (cPages < 1 \|\| cPages >= 256)
2437	{
2438	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2439	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2440	}
2441
2442	/*
2443	* Let IPRT do the job.
2444	*/
2445	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
2446	if (RT_SUCCESS(rc))
2447	{
2448	int rc2;
2449	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2450	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2451	if (RT_SUCCESS(rc))
2452	{
2453	Mem.eType = MEMREF_TYPE_CONT;
2454	rc = supdrvMemAdd(&Mem, pSession);
2455	if (!rc)
2456	{
2457	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2458	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2459	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
2460	return 0;
2461	}
2462
2463	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2464	AssertRC(rc2);
2465	}
2466	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2467	AssertRC(rc2);
2468	}
2469
2470	return rc;
2471	}
2472
2473
2474	/**
2475	* Frees memory allocated using SUPR0ContAlloc().
2476	*
2477	* @returns IPRT status code.
2478	* @param pSession The session to which the memory was allocated.
2479	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2480	*/
2481	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2482	{
2483	LogFlow(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2484	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2485	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
2486	}
2487
2488
2489	/**
2490	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
2491	*
2492	* The memory isn't zeroed.
2493	*
2494	* @returns IPRT status code.
2495	* @param pSession Session data.
2496	* @param cPages Number of pages to allocate.
2497	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
2498	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
2499	* @param paPages Where to put the physical addresses of allocated memory.
2500	*/
2501	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
2502	{
2503	unsigned iPage;
2504	int rc;
2505	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2506	LogFlow(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
2507
2508	/*
2509	* Validate input.
2510	*/
2511	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2512	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
2513	{
2514	Log(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
2515	pSession, ppvR3, ppvR0, paPages));
2516	return VERR_INVALID_PARAMETER;
2517
2518	}
2519	if (cPages < 1 \|\| cPages >= 256)
2520	{
2521	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2522	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2523	}
2524
2525	/*
2526	* Let IPRT do the work.
2527	*/
2528	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
2529	if (RT_SUCCESS(rc))
2530	{
2531	int rc2;
2532	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2533	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2534	if (RT_SUCCESS(rc))
2535	{
2536	Mem.eType = MEMREF_TYPE_LOW;
2537	rc = supdrvMemAdd(&Mem, pSession);
2538	if (!rc)
2539	{
2540	for (iPage = 0; iPage < cPages; iPage++)
2541	{
2542	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2543	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%RHp\n", paPages[iPage]));
2544	}
2545	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2546	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2547	return 0;
2548	}
2549
2550	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2551	AssertRC(rc2);
2552	}
2553
2554	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2555	AssertRC(rc2);
2556	}
2557
2558	return rc;
2559	}
2560
2561
2562	/**
2563	* Frees memory allocated using SUPR0LowAlloc().
2564	*
2565	* @returns IPRT status code.
2566	* @param pSession The session to which the memory was allocated.
2567	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2568	*/
2569	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2570	{
2571	LogFlow(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2572	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2573	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
2574	}
2575
2576
2577
2578	/**
2579	* Allocates a chunk of memory with both R0 and R3 mappings.
2580	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
2581	*
2582	* @returns IPRT status code.
2583	* @param pSession The session to associated the allocation with.
2584	* @param cb Number of bytes to allocate.
2585	* @param ppvR0 Where to store the address of the Ring-0 mapping.
2586	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2587	*/
2588	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
2589	{
2590	int rc;
2591	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2592	LogFlow(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
2593
2594	/*
2595	* Validate input.
2596	*/
2597	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2598	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
2599	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
2600	if (cb < 1 \|\| cb >= _4M)
2601	{
2602	Log(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
2603	return VERR_INVALID_PARAMETER;
2604	}
2605
2606	/*
2607	* Let IPRT do the work.
2608	*/
2609	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
2610	if (RT_SUCCESS(rc))
2611	{
2612	int rc2;
2613	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2614	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2615	if (RT_SUCCESS(rc))
2616	{
2617	Mem.eType = MEMREF_TYPE_MEM;
2618	rc = supdrvMemAdd(&Mem, pSession);
2619	if (!rc)
2620	{
2621	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2622	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2623	return VINF_SUCCESS;
2624	}
2625
2626	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2627	AssertRC(rc2);
2628	}
2629
2630	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2631	AssertRC(rc2);
2632	}
2633
2634	return rc;
2635	}
2636
2637
2638	/**
2639	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
2640	*
2641	* @returns IPRT status code.
2642	* @param pSession The session to which the memory was allocated.
2643	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2644	* @param paPages Where to store the physical addresses.
2645	*/
2646	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
2647	{
2648	PSUPDRVBUNDLE pBundle;
2649	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2650	LogFlow(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
2651
2652	/*
2653	* Validate input.
2654	*/
2655	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2656	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
2657	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
2658
2659	/*
2660	* Search for the address.
2661	*/
2662	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2663	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2664	{
2665	if (pBundle->cUsed > 0)
2666	{
2667	unsigned i;
2668	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2669	{
2670	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
2671	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2672	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2673	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2674	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
2675	)
2676	)
2677	{
2678	const size_t cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2679	size_t iPage;
2680	for (iPage = 0; iPage < cPages; iPage++)
2681	{
2682	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2683	paPages[iPage].uReserved = 0;
2684	}
2685	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2686	return VINF_SUCCESS;
2687	}
2688	}
2689	}
2690	}
2691	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2692	Log(("Failed to find %p!!!\n", (void *)uPtr));
2693	return VERR_INVALID_PARAMETER;
2694	}
2695
2696
2697	/**
2698	* Free memory allocated by SUPR0MemAlloc().
2699	*
2700	* @returns IPRT status code.
2701	* @param pSession The session owning the allocation.
2702	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2703	*/
2704	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2705	{
2706	LogFlow(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2707	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2708	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
2709	}
2710
2711
2712	/**
2713	* Allocates a chunk of memory with a kernel or/and a user mode mapping.
2714	*
2715	* The memory is fixed and it's possible to query the physical addresses using
2716	* SUPR0MemGetPhys().
2717	*
2718	* @returns IPRT status code.
2719	* @param pSession The session to associated the allocation with.
2720	* @param cPages The number of pages to allocate.
2721	* @param fFlags Flags, reserved for the future. Must be zero.
2722	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2723	* NULL if no ring-3 mapping.
2724	* @param ppvR3 Where to store the address of the Ring-0 mapping.
2725	* NULL if no ring-0 mapping.
2726	* @param paPages Where to store the addresses of the pages. Optional.
2727	*/
2728	SUPR0DECL(int) SUPR0PageAllocEx(PSUPDRVSESSION pSession, uint32_t cPages, uint32_t fFlags, PRTR3PTR ppvR3, PRTR0PTR ppvR0, PRTHCPHYS paPages)
2729	{
2730	int rc;
2731	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2732	LogFlow(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
2733
2734	/*
2735	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2736	*/
2737	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2738	AssertPtrNullReturn(ppvR3, VERR_INVALID_POINTER);
2739	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2740	AssertReturn(ppvR3 \|\| ppvR0, VERR_INVALID_PARAMETER);
2741	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2742	if (cPages < 1 \|\| cPages > VBOX_MAX_ALLOC_PAGE_COUNT)
2743	{
2744	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)));
2745	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2746	}
2747
2748	/*
2749	* Let IPRT do the work.
2750	*/
2751	if (ppvR0)
2752	rc = RTR0MemObjAllocPage(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, true /* fExecutable */);
2753	else
2754	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
2755	if (RT_SUCCESS(rc))
2756	{
2757	int rc2;
2758	if (ppvR3)
2759	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2760	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2761	else
2762	Mem.MapObjR3 = NIL_RTR0MEMOBJ;
2763	if (RT_SUCCESS(rc))
2764	{
2765	Mem.eType = MEMREF_TYPE_PAGE;
2766	rc = supdrvMemAdd(&Mem, pSession);
2767	if (!rc)
2768	{
2769	if (ppvR3)
2770	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2771	if (ppvR0)
2772	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2773	if (paPages)
2774	{
2775	uint32_t iPage = cPages;
2776	while (iPage-- > 0)
2777	{
2778	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
2779	Assert(paPages[iPage] != NIL_RTHCPHYS);
2780	}
2781	}
2782	return VINF_SUCCESS;
2783	}
2784
2785	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2786	AssertRC(rc2);
2787	}
2788
2789	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2790	AssertRC(rc2);
2791	}
2792	return rc;
2793	}
2794
2795
2796	/**
2797	* Maps a chunk of memory previously allocated by SUPR0PageAllocEx into kernel
2798	* space.
2799	*
2800	* @returns IPRT status code.
2801	* @param pSession The session to associated the allocation with.
2802	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
2803	* @param offSub Where to start mapping. Must be page aligned.
2804	* @param cbSub How much to map. Must be page aligned.
2805	* @param fFlags Flags, MBZ.
2806	* @param ppvR0 Where to return the address of the ring-0 mapping on
2807	* success.
2808	*/
2809	SUPR0DECL(int) SUPR0PageMapKernel(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t offSub, uint32_t cbSub,
2810	uint32_t fFlags, PRTR0PTR ppvR0)
2811	{
2812	int rc;
2813	PSUPDRVBUNDLE pBundle;
2814	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2815	RTR0MEMOBJ hMemObj = NIL_RTR0MEMOBJ;
2816	LogFlow(("SUPR0PageMapKernel: pSession=%p pvR3=%p offSub=%#x cbSub=%#x\n", pSession, pvR3, offSub, cbSub));
2817
2818	/*
2819	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2820	*/
2821	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2822	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2823	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2824	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2825	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2826	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
2827
2828	/*
2829	* Find the memory object.
2830	*/
2831	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2832	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2833	{
2834	if (pBundle->cUsed > 0)
2835	{
2836	unsigned i;
2837	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2838	{
2839	if ( ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
2840	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2841	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2842	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
2843	\|\| ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED
2844	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2845	&& pBundle->aMem[i].MapObjR3 == NIL_RTR0MEMOBJ
2846	&& RTR0MemObjAddressR3(pBundle->aMem[i].MemObj) == pvR3))
2847	{
2848	hMemObj = pBundle->aMem[i].MemObj;
2849	break;
2850	}
2851	}
2852	}
2853	}
2854	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2855
2856	rc = VERR_INVALID_PARAMETER;
2857	if (hMemObj != NIL_RTR0MEMOBJ)
2858	{
2859	/*
2860	* Do some further input validations before calling IPRT.
2861	* (Cleanup is done indirectly by telling RTR0MemObjFree to include mappings.)
2862	*/
2863	size_t cbMemObj = RTR0MemObjSize(hMemObj);
2864	if ( offSub < cbMemObj
2865	&& cbSub <= cbMemObj
2866	&& offSub + cbSub <= cbMemObj)
2867	{
2868	RTR0MEMOBJ hMapObj;
2869	rc = RTR0MemObjMapKernelEx(&hMapObj, hMemObj, (void *)-1, 0,
2870	RTMEM_PROT_READ \| RTMEM_PROT_WRITE, offSub, cbSub);
2871	if (RT_SUCCESS(rc))
2872	*ppvR0 = RTR0MemObjAddress(hMapObj);
2873	}
2874	else
2875	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
2876
2877	}
2878	return rc;
2879	}
2880
2881
2882	/**
2883	* Changes the page level protection of one or more pages previously allocated
2884	* by SUPR0PageAllocEx.
2885	*
2886	* @returns IPRT status code.
2887	* @param pSession The session to associated the allocation with.
2888	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
2889	* NIL_RTR3PTR if the ring-3 mapping should be unaffected.
2890	* @param pvR0 The ring-0 address returned by SUPR0PageAllocEx.
2891	* NIL_RTR0PTR if the ring-0 mapping should be unaffected.
2892	* @param offSub Where to start changing. Must be page aligned.
2893	* @param cbSub How much to change. Must be page aligned.
2894	* @param fProt The new page level protection, see RTMEM_PROT_*.
2895	*/
2896	SUPR0DECL(int) SUPR0PageProtect(PSUPDRVSESSION pSession, RTR3PTR pvR3, RTR0PTR pvR0, uint32_t offSub, uint32_t cbSub, uint32_t fProt)
2897	{
2898	int rc;
2899	PSUPDRVBUNDLE pBundle;
2900	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2901	RTR0MEMOBJ hMemObjR0 = NIL_RTR0MEMOBJ;
2902	RTR0MEMOBJ hMemObjR3 = NIL_RTR0MEMOBJ;
2903	LogFlow(("SUPR0PageProtect: pSession=%p pvR3=%p pvR0=%p offSub=%#x cbSub=%#x fProt-%#x\n", pSession, pvR3, pvR0, offSub, cbSub, fProt));
2904
2905	/*
2906	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2907	*/
2908	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2909	AssertReturn(!(fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)), VERR_INVALID_PARAMETER);
2910	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2911	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2912	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
2913
2914	/*
2915	* Find the memory object.
2916	*/
2917	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2918	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2919	{
2920	if (pBundle->cUsed > 0)
2921	{
2922	unsigned i;
2923	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2924	{
2925	if ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
2926	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2927	&& ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2928	\|\| pvR3 == NIL_RTR3PTR)
2929	&& ( pvR0 == NIL_RTR0PTR
2930	\|\| RTR0MemObjAddress(pBundle->aMem[i].MemObj) == pvR0)
2931	&& ( pvR3 == NIL_RTR3PTR
2932	\|\| RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3))
2933	{
2934	if (pvR0 != NIL_RTR0PTR)
2935	hMemObjR0 = pBundle->aMem[i].MemObj;
2936	if (pvR3 != NIL_RTR3PTR)
2937	hMemObjR3 = pBundle->aMem[i].MapObjR3;
2938	break;
2939	}
2940	}
2941	}
2942	}
2943	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2944
2945	rc = VERR_INVALID_PARAMETER;
2946	if ( hMemObjR0 != NIL_RTR0MEMOBJ
2947	\|\| hMemObjR3 != NIL_RTR0MEMOBJ)
2948	{
2949	/*
2950	* Do some further input validations before calling IPRT.
2951	*/
2952	size_t cbMemObj = hMemObjR0 != NIL_RTR0PTR ? RTR0MemObjSize(hMemObjR0) : RTR0MemObjSize(hMemObjR3);
2953	if ( offSub < cbMemObj
2954	&& cbSub <= cbMemObj
2955	&& offSub + cbSub <= cbMemObj)
2956	{
2957	rc = VINF_SUCCESS;
2958	if (hMemObjR3 != NIL_RTR0PTR)
2959	rc = RTR0MemObjProtect(hMemObjR3, offSub, cbSub, fProt);
2960	if (hMemObjR0 != NIL_RTR0PTR && RT_SUCCESS(rc))
2961	rc = RTR0MemObjProtect(hMemObjR0, offSub, cbSub, fProt);
2962	}
2963	else
2964	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
2965
2966	}
2967	return rc;
2968
2969	}
2970
2971
2972	/**
2973	* Free memory allocated by SUPR0PageAlloc() and SUPR0PageAllocEx().
2974	*
2975	* @returns IPRT status code.
2976	* @param pSession The session owning the allocation.
2977	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc() or
2978	* SUPR0PageAllocEx().
2979	*/
2980	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2981	{
2982	LogFlow(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2983	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2984	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_PAGE);
2985	}
2986
2987
2988	/**
2989	* Gets the paging mode of the current CPU.
2990	*
2991	* @returns Paging mode, SUPPAGEINGMODE_INVALID on error.
2992	*/
2993	SUPR0DECL(SUPPAGINGMODE) SUPR0GetPagingMode(void)
2994	{
2995	SUPPAGINGMODE enmMode;
2996
2997	RTR0UINTREG cr0 = ASMGetCR0();
2998	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
2999	enmMode = SUPPAGINGMODE_INVALID;
3000	else
3001	{
3002	RTR0UINTREG cr4 = ASMGetCR4();
3003	uint32_t fNXEPlusLMA = 0;
3004	if (cr4 & X86_CR4_PAE)
3005	{
3006	uint32_t fAmdFeatures = ASMCpuId_EDX(0x80000001);
3007	if (fAmdFeatures & (X86_CPUID_AMD_FEATURE_EDX_NX \| X86_CPUID_AMD_FEATURE_EDX_LONG_MODE))
3008	{
3009	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
3010	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
3011	fNXEPlusLMA \|= RT_BIT(0);
3012	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
3013	fNXEPlusLMA \|= RT_BIT(1);
3014	}
3015	}
3016
3017	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
3018	{
3019	case 0:
3020	enmMode = SUPPAGINGMODE_32_BIT;
3021	break;
3022
3023	case X86_CR4_PGE:
3024	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
3025	break;
3026
3027	case X86_CR4_PAE:
3028	enmMode = SUPPAGINGMODE_PAE;
3029	break;
3030
3031	case X86_CR4_PAE \| RT_BIT(0):
3032	enmMode = SUPPAGINGMODE_PAE_NX;
3033	break;
3034
3035	case X86_CR4_PAE \| X86_CR4_PGE:
3036	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3037	break;
3038
3039	case X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3040	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3041	break;
3042
3043	case RT_BIT(1) \| X86_CR4_PAE:
3044	enmMode = SUPPAGINGMODE_AMD64;
3045	break;
3046
3047	case RT_BIT(1) \| X86_CR4_PAE \| RT_BIT(0):
3048	enmMode = SUPPAGINGMODE_AMD64_NX;
3049	break;
3050
3051	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
3052	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
3053	break;
3054
3055	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3056	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
3057	break;
3058
3059	default:
3060	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
3061	enmMode = SUPPAGINGMODE_INVALID;
3062	break;
3063	}
3064	}
3065	return enmMode;
3066	}
3067
3068
3069	/**
3070	* Enables or disabled hardware virtualization extensions using native OS APIs.
3071	*
3072	* @returns VBox status code.
3073	* @retval VINF_SUCCESS on success.
3074	* @retval VERR_NOT_SUPPORTED if not supported by the native OS.
3075	*
3076	* @param fEnable Whether to enable or disable.
3077	*/
3078	SUPR0DECL(int) SUPR0EnableVTx(bool fEnable)
3079	{
3080	#ifdef RT_OS_DARWIN
3081	return supdrvOSEnableVTx(fEnable);
3082	#else
3083	return VERR_NOT_SUPPORTED;
3084	#endif
3085	}
3086
3087
3088	/** @todo document me */
3089	SUPR0DECL(int) SUPR0QueryVTCaps(PSUPDRVSESSION pSession, uint32_t *pfCaps)
3090	{
3091	/*
3092	* Input validation.
3093	*/
3094	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3095	AssertPtrReturn(pfCaps, VERR_INVALID_POINTER);
3096
3097	*pfCaps = 0;
3098
3099	if (ASMHasCpuId())
3100	{
3101	uint32_t u32FeaturesECX;
3102	uint32_t u32Dummy;
3103	uint32_t u32FeaturesEDX;
3104	uint32_t u32VendorEBX, u32VendorECX, u32VendorEDX, u32AMDFeatureEDX, u32AMDFeatureECX;
3105	uint64_t val;
3106
3107	ASMCpuId(0, &u32Dummy, &u32VendorEBX, &u32VendorECX, &u32VendorEDX);
3108	ASMCpuId(1, &u32Dummy, &u32Dummy, &u32FeaturesECX, &u32FeaturesEDX);
3109	/* Query AMD features. */
3110	ASMCpuId(0x80000001, &u32Dummy, &u32Dummy, &u32AMDFeatureECX, &u32AMDFeatureEDX);
3111
3112	if ( u32VendorEBX == X86_CPUID_VENDOR_INTEL_EBX
3113	&& u32VendorECX == X86_CPUID_VENDOR_INTEL_ECX
3114	&& u32VendorEDX == X86_CPUID_VENDOR_INTEL_EDX
3115	)
3116	{
3117	if ( (u32FeaturesECX & X86_CPUID_FEATURE_ECX_VMX)
3118	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3119	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3120	)
3121	{
3122	val = ASMRdMsr(MSR_IA32_FEATURE_CONTROL);
3123	/*
3124	* Both the LOCK and VMXON bit must be set; otherwise VMXON will generate a #GP.
3125	* Once the lock bit is set, this MSR can no longer be modified.
3126	*/
3127	if ( (val & (MSR_IA32_FEATURE_CONTROL_VMXON\|MSR_IA32_FEATURE_CONTROL_LOCK))
3128	== (MSR_IA32_FEATURE_CONTROL_VMXON\|MSR_IA32_FEATURE_CONTROL_LOCK) /* enabled and locked */
3129	\|\| !(val & MSR_IA32_FEATURE_CONTROL_LOCK) /* not enabled, but not locked either */
3130	)
3131	{
3132	VMX_CAPABILITY vtCaps;
3133
3134	*pfCaps \|= SUPVTCAPS_VT_X;
3135
3136	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS);
3137	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC_USE_SECONDARY_EXEC_CTRL)
3138	{
3139	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS2);
3140	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC2_EPT)
3141	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3142	}
3143	return VINF_SUCCESS;
3144	}
3145	return VERR_VMX_MSR_LOCKED_OR_DISABLED;
3146	}
3147	return VERR_VMX_NO_VMX;
3148	}
3149
3150	if ( u32VendorEBX == X86_CPUID_VENDOR_AMD_EBX
3151	&& u32VendorECX == X86_CPUID_VENDOR_AMD_ECX
3152	&& u32VendorEDX == X86_CPUID_VENDOR_AMD_EDX
3153	)
3154	{
3155	if ( (u32AMDFeatureECX & X86_CPUID_AMD_FEATURE_ECX_SVM)
3156	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3157	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3158	)
3159	{
3160	/* Check if SVM is disabled */
3161	val = ASMRdMsr(MSR_K8_VM_CR);
3162	if (!(val & MSR_K8_VM_CR_SVM_DISABLE))
3163	{
3164	*pfCaps \|= SUPVTCAPS_AMD_V;
3165
3166	/* Query AMD features. */
3167	ASMCpuId(0x8000000A, &u32Dummy, &u32Dummy, &u32Dummy, &u32FeaturesEDX);
3168
3169	if (u32FeaturesEDX & AMD_CPUID_SVM_FEATURE_EDX_NESTED_PAGING)
3170	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3171
3172	return VINF_SUCCESS;
3173	}
3174	return VERR_SVM_DISABLED;
3175	}
3176	return VERR_SVM_NO_SVM;
3177	}
3178	}
3179
3180	return VERR_UNSUPPORTED_CPU;
3181	}
3182
3183
3184	/**
3185	* (Re-)initializes the per-cpu structure prior to starting or resuming the GIP
3186	* updating.
3187	*
3188	* @param pGipCpu The per CPU structure for this CPU.
3189	* @param u64NanoTS The current time.
3190	*/
3191	static void supdrvGipReInitCpu(PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
3192	{
3193	pGipCpu->u64TSC = ASMReadTSC() - pGipCpu->u32UpdateIntervalTSC;
3194	pGipCpu->u64NanoTS = u64NanoTS;
3195	}
3196
3197
3198	/**
3199	* Set the current TSC and NanoTS value for the CPU.
3200	*
3201	* @param idCpu The CPU ID. Unused - we have to use the APIC ID.
3202	* @param pvUser1 Pointer to the ring-0 GIP mapping.
3203	* @param pvUser2 Pointer to the variable holding the current time.
3204	*/
3205	static DECLCALLBACK(void) supdrvGipReInitCpuCallback(RTCPUID idCpu, void pvUser1, void pvUser2)
3206	{
3207	PSUPGLOBALINFOPAGE pGip = (PSUPGLOBALINFOPAGE)pvUser1;
3208	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
3209
3210	if (RT_LIKELY(iCpu < pGip->cCpus && pGip->aCPUs[iCpu].idCpu == idCpu))
3211	supdrvGipReInitCpu(&pGip->aCPUs[iCpu], (uint64_t )pvUser2);
3212
3213	NOREF(pvUser2);
3214	NOREF(idCpu);
3215	}
3216
3217
3218	/**
3219	* Maps the GIP into userspace and/or get the physical address of the GIP.
3220	*
3221	* @returns IPRT status code.
3222	* @param pSession Session to which the GIP mapping should belong.
3223	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
3224	* @param pHCPhysGip Where to store the physical address. (optional)
3225	*
3226	* @remark There is no reference counting on the mapping, so one call to this function
3227	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
3228	* and remove the session as a GIP user.
3229	*/
3230	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
3231	{
3232	int rc;
3233	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3234	RTR3PTR pGipR3 = NIL_RTR3PTR;
3235	RTHCPHYS HCPhys = NIL_RTHCPHYS;
3236	LogFlow(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
3237
3238	/*
3239	* Validate
3240	*/
3241	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3242	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
3243	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
3244
3245	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3246	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3247	#else
3248	RTSemFastMutexRequest(pDevExt->mtxGip);
3249	#endif
3250	if (pDevExt->pGip)
3251	{
3252	/*
3253	* Map it?
3254	*/
3255	rc = VINF_SUCCESS;
3256	if (ppGipR3)
3257	{
3258	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
3259	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
3260	RTMEM_PROT_READ, RTR0ProcHandleSelf());
3261	if (RT_SUCCESS(rc))
3262	pGipR3 = RTR0MemObjAddressR3(pSession->GipMapObjR3);
3263	}
3264
3265	/*
3266	* Get physical address.
3267	*/
3268	if (pHCPhysGip && RT_SUCCESS(rc))
3269	HCPhys = pDevExt->HCPhysGip;
3270
3271	/*
3272	* Reference globally.
3273	*/
3274	if (!pSession->fGipReferenced && RT_SUCCESS(rc))
3275	{
3276	pSession->fGipReferenced = 1;
3277	pDevExt->cGipUsers++;
3278	if (pDevExt->cGipUsers == 1)
3279	{
3280	PSUPGLOBALINFOPAGE pGipR0 = pDevExt->pGip;
3281	uint64_t u64NanoTS;
3282	uint32_t u32SystemResolution;
3283	unsigned i;
3284
3285	LogFlow(("SUPR0GipMap: Resumes GIP updating\n"));
3286
3287	/*
3288	* Try bump up the system timer resolution.
3289	* The more interrupts the better...
3290	*/
3291	if ( RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
3292	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
3293	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1953125 /* 512 HZ */, &u32SystemResolution))
3294	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 2000000 /* 500 HZ */, &u32SystemResolution))
3295	)
3296	{
3297	Assert(RTTimerGetSystemGranularity() <= u32SystemResolution);
3298	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
3299	}
3300
3301	if (pGipR0->aCPUs[0].u32TransactionId != 2 /* not the first time */)
3302	{
3303	for (i = 0; i < RT_ELEMENTS(pGipR0->aCPUs); i++)
3304	ASMAtomicUoWriteU32(&pGipR0->aCPUs[i].u32TransactionId,
3305	(pGipR0->aCPUs[i].u32TransactionId + GIP_UPDATEHZ_RECALC_FREQ * 2)
3306	& ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
3307	ASMAtomicWriteU64(&pGipR0->u64NanoTSLastUpdateHz, 0);
3308	}
3309
3310	u64NanoTS = RTTimeSystemNanoTS() - pGipR0->u32UpdateIntervalNS;
3311	if ( pGipR0->u32Mode == SUPGIPMODE_SYNC_TSC
3312	\|\| RTMpGetOnlineCount() == 1)
3313	supdrvGipReInitCpu(&pGipR0->aCPUs[0], u64NanoTS);
3314	else
3315	RTMpOnAll(supdrvGipReInitCpuCallback, pGipR0, &u64NanoTS);
3316
3317	#ifndef DO_NOT_START_GIP
3318	rc = RTTimerStart(pDevExt->pGipTimer, 0); AssertRC(rc);
3319	#endif
3320	rc = VINF_SUCCESS;
3321	}
3322	}
3323	}
3324	else
3325	{
3326	rc = VERR_GENERAL_FAILURE;
3327	Log(("SUPR0GipMap: GIP is not available!\n"));
3328	}
3329	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3330	RTSemMutexRelease(pDevExt->mtxGip);
3331	#else
3332	RTSemFastMutexRelease(pDevExt->mtxGip);
3333	#endif
3334
3335	/*
3336	* Write returns.
3337	*/
3338	if (pHCPhysGip)
3339	*pHCPhysGip = HCPhys;
3340	if (ppGipR3)
3341	*ppGipR3 = pGipR3;
3342
3343	#ifdef DEBUG_DARWIN_GIP
3344	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3345	#else
3346	LogFlow(( "SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3347	#endif
3348	return rc;
3349	}
3350
3351
3352	/**
3353	* Unmaps any user mapping of the GIP and terminates all GIP access
3354	* from this session.
3355	*
3356	* @returns IPRT status code.
3357	* @param pSession Session to which the GIP mapping should belong.
3358	*/
3359	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
3360	{
3361	int rc = VINF_SUCCESS;
3362	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3363	#ifdef DEBUG_DARWIN_GIP
3364	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
3365	pSession,
3366	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
3367	pSession->GipMapObjR3));
3368	#else
3369	LogFlow(("SUPR0GipUnmap: pSession=%p\n", pSession));
3370	#endif
3371	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3372
3373	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3374	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3375	#else
3376	RTSemFastMutexRequest(pDevExt->mtxGip);
3377	#endif
3378
3379	/*
3380	* Unmap anything?
3381	*/
3382	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
3383	{
3384	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
3385	AssertRC(rc);
3386	if (RT_SUCCESS(rc))
3387	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
3388	}
3389
3390	/*
3391	* Dereference global GIP.
3392	*/
3393	if (pSession->fGipReferenced && !rc)
3394	{
3395	pSession->fGipReferenced = 0;
3396	if ( pDevExt->cGipUsers > 0
3397	&& !--pDevExt->cGipUsers)
3398	{
3399	LogFlow(("SUPR0GipUnmap: Suspends GIP updating\n"));
3400	#ifndef DO_NOT_START_GIP
3401	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = VINF_SUCCESS;
3402	#endif
3403
3404	if (pDevExt->u32SystemTimerGranularityGrant)
3405	{
3406	int rc2 = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant);
3407	AssertRC(rc2);
3408	pDevExt->u32SystemTimerGranularityGrant = 0;
3409	}
3410	}
3411	}
3412
3413	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3414	RTSemMutexRelease(pDevExt->mtxGip);
3415	#else
3416	RTSemFastMutexRelease(pDevExt->mtxGip);
3417	#endif
3418
3419	return rc;
3420	}
3421
3422
3423	/**
3424	* Gets the GIP pointer.
3425	*
3426	* @returns Pointer to the GIP or NULL.
3427	*/
3428	SUPDECL(PSUPGLOBALINFOPAGE) SUPGetGIP(void)
3429	{
3430	return g_pSUPGlobalInfoPage;
3431	}
3432
3433
3434	/**
3435	* Register a component factory with the support driver.
3436	*
3437	* This is currently restricted to kernel sessions only.
3438	*
3439	* @returns VBox status code.
3440	* @retval VINF_SUCCESS on success.
3441	* @retval VERR_NO_MEMORY if we're out of memory.
3442	* @retval VERR_ALREADY_EXISTS if the factory has already been registered.
3443	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3444	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3445	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3446	*
3447	* @param pSession The SUPDRV session (must be a ring-0 session).
3448	* @param pFactory Pointer to the component factory registration structure.
3449	*
3450	* @remarks This interface is also available via SUPR0IdcComponentRegisterFactory.
3451	*/
3452	SUPR0DECL(int) SUPR0ComponentRegisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3453	{
3454	PSUPDRVFACTORYREG pNewReg;
3455	const char *psz;
3456	int rc;
3457
3458	/*
3459	* Validate parameters.
3460	*/
3461	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3462	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3463	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3464	AssertPtrReturn(pFactory->pfnQueryFactoryInterface, VERR_INVALID_POINTER);
3465	psz = RTStrEnd(pFactory->szName, sizeof(pFactory->szName));
3466	AssertReturn(psz, VERR_INVALID_PARAMETER);
3467
3468	/*
3469	* Allocate and initialize a new registration structure.
3470	*/
3471	pNewReg = (PSUPDRVFACTORYREG)RTMemAlloc(sizeof(SUPDRVFACTORYREG));
3472	if (pNewReg)
3473	{
3474	pNewReg->pNext = NULL;
3475	pNewReg->pFactory = pFactory;
3476	pNewReg->pSession = pSession;
3477	pNewReg->cchName = psz - &pFactory->szName[0];
3478
3479	/*
3480	* Add it to the tail of the list after checking for prior registration.
3481	*/
3482	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3483	if (RT_SUCCESS(rc))
3484	{
3485	PSUPDRVFACTORYREG pPrev = NULL;
3486	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3487	while (pCur && pCur->pFactory != pFactory)
3488	{
3489	pPrev = pCur;
3490	pCur = pCur->pNext;
3491	}
3492	if (!pCur)
3493	{
3494	if (pPrev)
3495	pPrev->pNext = pNewReg;
3496	else
3497	pSession->pDevExt->pComponentFactoryHead = pNewReg;
3498	rc = VINF_SUCCESS;
3499	}
3500	else
3501	rc = VERR_ALREADY_EXISTS;
3502
3503	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3504	}
3505
3506	if (RT_FAILURE(rc))
3507	RTMemFree(pNewReg);
3508	}
3509	else
3510	rc = VERR_NO_MEMORY;
3511	return rc;
3512	}
3513
3514
3515	/**
3516	* Deregister a component factory.
3517	*
3518	* @returns VBox status code.
3519	* @retval VINF_SUCCESS on success.
3520	* @retval VERR_NOT_FOUND if the factory wasn't registered.
3521	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3522	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3523	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3524	*
3525	* @param pSession The SUPDRV session (must be a ring-0 session).
3526	* @param pFactory Pointer to the component factory registration structure
3527	* previously passed SUPR0ComponentRegisterFactory().
3528	*
3529	* @remarks This interface is also available via SUPR0IdcComponentDeregisterFactory.
3530	*/
3531	SUPR0DECL(int) SUPR0ComponentDeregisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3532	{
3533	int rc;
3534
3535	/*
3536	* Validate parameters.
3537	*/
3538	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3539	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3540	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3541
3542	/*
3543	* Take the lock and look for the registration record.
3544	*/
3545	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3546	if (RT_SUCCESS(rc))
3547	{
3548	PSUPDRVFACTORYREG pPrev = NULL;
3549	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3550	while (pCur && pCur->pFactory != pFactory)
3551	{
3552	pPrev = pCur;
3553	pCur = pCur->pNext;
3554	}
3555	if (pCur)
3556	{
3557	if (!pPrev)
3558	pSession->pDevExt->pComponentFactoryHead = pCur->pNext;
3559	else
3560	pPrev->pNext = pCur->pNext;
3561
3562	pCur->pNext = NULL;
3563	pCur->pFactory = NULL;
3564	pCur->pSession = NULL;
3565	rc = VINF_SUCCESS;
3566	}
3567	else
3568	rc = VERR_NOT_FOUND;
3569
3570	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3571
3572	RTMemFree(pCur);
3573	}
3574	return rc;
3575	}
3576
3577
3578	/**
3579	* Queries a component factory.
3580	*
3581	* @returns VBox status code.
3582	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3583	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3584	* @retval VERR_SUPDRV_COMPONENT_NOT_FOUND if the component factory wasn't found.
3585	* @retval VERR_SUPDRV_INTERFACE_NOT_SUPPORTED if the interface wasn't supported.
3586	*
3587	* @param pSession The SUPDRV session.
3588	* @param pszName The name of the component factory.
3589	* @param pszInterfaceUuid The UUID of the factory interface (stringified).
3590	* @param ppvFactoryIf Where to store the factory interface.
3591	*/
3592	SUPR0DECL(int) SUPR0ComponentQueryFactory(PSUPDRVSESSION pSession, const char pszName, const char pszInterfaceUuid, void **ppvFactoryIf)
3593	{
3594	const char *pszEnd;
3595	size_t cchName;
3596	int rc;
3597
3598	/*
3599	* Validate parameters.
3600	*/
3601	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3602
3603	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
3604	pszEnd = RTStrEnd(pszName, RT_SIZEOFMEMB(SUPDRVFACTORY, szName));
3605	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3606	cchName = pszEnd - pszName;
3607
3608	AssertPtrReturn(pszInterfaceUuid, VERR_INVALID_POINTER);
3609	pszEnd = RTStrEnd(pszInterfaceUuid, RTUUID_STR_LENGTH);
3610	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3611
3612	AssertPtrReturn(ppvFactoryIf, VERR_INVALID_POINTER);
3613	*ppvFactoryIf = NULL;
3614
3615	/*
3616	* Take the lock and try all factories by this name.
3617	*/
3618	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3619	if (RT_SUCCESS(rc))
3620	{
3621	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3622	rc = VERR_SUPDRV_COMPONENT_NOT_FOUND;
3623	while (pCur)
3624	{
3625	if ( pCur->cchName == cchName
3626	&& !memcmp(pCur->pFactory->szName, pszName, cchName))
3627	{
3628	void *pvFactory = pCur->pFactory->pfnQueryFactoryInterface(pCur->pFactory, pSession, pszInterfaceUuid);
3629	if (pvFactory)
3630	{
3631	*ppvFactoryIf = pvFactory;
3632	rc = VINF_SUCCESS;
3633	break;
3634	}
3635	rc = VERR_SUPDRV_INTERFACE_NOT_SUPPORTED;
3636	}
3637
3638	/* next */
3639	pCur = pCur->pNext;
3640	}
3641
3642	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3643	}
3644	return rc;
3645	}
3646
3647
3648	/**
3649	* Adds a memory object to the session.
3650	*
3651	* @returns IPRT status code.
3652	* @param pMem Memory tracking structure containing the
3653	* information to track.
3654	* @param pSession The session.
3655	*/
3656	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
3657	{
3658	PSUPDRVBUNDLE pBundle;
3659	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3660
3661	/*
3662	* Find free entry and record the allocation.
3663	*/
3664	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3665	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3666	{
3667	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
3668	{
3669	unsigned i;
3670	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3671	{
3672	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
3673	{
3674	pBundle->cUsed++;
3675	pBundle->aMem[i] = *pMem;
3676	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3677	return VINF_SUCCESS;
3678	}
3679	}
3680	AssertFailed(); /* !!this can't be happening!!! */
3681	}
3682	}
3683	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3684
3685	/*
3686	* Need to allocate a new bundle.
3687	* Insert into the last entry in the bundle.
3688	*/
3689	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
3690	if (!pBundle)
3691	return VERR_NO_MEMORY;
3692
3693	/* take last entry. */
3694	pBundle->cUsed++;
3695	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
3696
3697	/* insert into list. */
3698	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3699	pBundle->pNext = pSession->Bundle.pNext;
3700	pSession->Bundle.pNext = pBundle;
3701	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3702
3703	return VINF_SUCCESS;
3704	}
3705
3706
3707	/**
3708	* Releases a memory object referenced by pointer and type.
3709	*
3710	* @returns IPRT status code.
3711	* @param pSession Session data.
3712	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
3713	* @param eType Memory type.
3714	*/
3715	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
3716	{
3717	PSUPDRVBUNDLE pBundle;
3718	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3719
3720	/*
3721	* Validate input.
3722	*/
3723	if (!uPtr)
3724	{
3725	Log(("Illegal address %p\n", (void *)uPtr));
3726	return VERR_INVALID_PARAMETER;
3727	}
3728
3729	/*
3730	* Search for the address.
3731	*/
3732	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3733	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3734	{
3735	if (pBundle->cUsed > 0)
3736	{
3737	unsigned i;
3738	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3739	{
3740	if ( pBundle->aMem[i].eType == eType
3741	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3742	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
3743	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3744	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
3745	)
3746	{
3747	/* Make a copy of it and release it outside the spinlock. */
3748	SUPDRVMEMREF Mem = pBundle->aMem[i];
3749	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
3750	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
3751	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
3752	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3753
3754	if (Mem.MapObjR3 != NIL_RTR0MEMOBJ)
3755	{
3756	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
3757	AssertRC(rc); /** @todo figure out how to handle this. */
3758	}
3759	if (Mem.MemObj != NIL_RTR0MEMOBJ)
3760	{
3761	int rc = RTR0MemObjFree(Mem.MemObj, true /* fFreeMappings */);
3762	AssertRC(rc); /** @todo figure out how to handle this. */
3763	}
3764	return VINF_SUCCESS;
3765	}
3766	}
3767	}
3768	}
3769	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3770	Log(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
3771	return VERR_INVALID_PARAMETER;
3772	}
3773
3774
3775	/**
3776	* Opens an image. If it's the first time it's opened the call must upload
3777	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
3778	*
3779	* This is the 1st step of the loading.
3780	*
3781	* @returns IPRT status code.
3782	* @param pDevExt Device globals.
3783	* @param pSession Session data.
3784	* @param pReq The open request.
3785	*/
3786	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
3787	{
3788	int rc;
3789	PSUPDRVLDRIMAGE pImage;
3790	void *pv;
3791	size_t cchName = strlen(pReq->u.In.szName); /* (caller checked < 32). */
3792	LogFlow(("supdrvIOCtl_LdrOpen: szName=%s cbImageWithTabs=%d\n", pReq->u.In.szName, pReq->u.In.cbImageWithTabs));
3793
3794	/*
3795	* Check if we got an instance of the image already.
3796	*/
3797	supdrvLdrLock(pDevExt);
3798	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
3799	{
3800	if ( pImage->szName[cchName] == '\0'
3801	&& !memcmp(pImage->szName, pReq->u.In.szName, cchName))
3802	{
3803	/** @todo check cbImageBits and cbImageWithTabs here, if they differs that indicates that the images are different. */
3804	pImage->cUsage++;
3805	pReq->u.Out.pvImageBase = pImage->pvImage;
3806	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
3807	pReq->u.Out.fNativeLoader = pImage->fNative;
3808	supdrvLdrAddUsage(pSession, pImage);
3809	supdrvLdrUnlock(pDevExt);
3810	return VINF_SUCCESS;
3811	}
3812	}
3813	/* (not found - add it!) */
3814
3815	/*
3816	* Allocate memory.
3817	*/
3818	pv = RTMemAlloc(RT_OFFSETOF(SUPDRVLDRIMAGE, szName[cchName + 1]));
3819	if (!pv)
3820	{
3821	supdrvLdrUnlock(pDevExt);
3822	Log(("supdrvIOCtl_LdrOpen: RTMemAlloc() failed\n"));
3823	return /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_2;
3824	}
3825
3826	/*
3827	* Setup and link in the LDR stuff.
3828	*/
3829	pImage = (PSUPDRVLDRIMAGE)pv;
3830	pImage->pvImage = NULL;
3831	pImage->pvImageAlloc = NULL;
3832	pImage->cbImageWithTabs = pReq->u.In.cbImageWithTabs;
3833	pImage->cbImageBits = pReq->u.In.cbImageBits;
3834	pImage->cSymbols = 0;
3835	pImage->paSymbols = NULL;
3836	pImage->pachStrTab = NULL;
3837	pImage->cbStrTab = 0;
3838	pImage->pfnModuleInit = NULL;
3839	pImage->pfnModuleTerm = NULL;
3840	pImage->pfnServiceReqHandler = NULL;
3841	pImage->uState = SUP_IOCTL_LDR_OPEN;
3842	pImage->cUsage = 1;
3843	memcpy(pImage->szName, pReq->u.In.szName, cchName + 1);
3844
3845	/*
3846	* Try load it using the native loader, if that isn't supported, fall back
3847	* on the older method.
3848	*/
3849	pImage->fNative = true;
3850	rc = supdrvOSLdrOpen(pDevExt, pImage, pReq->u.In.szFilename);
3851	if (rc == VERR_NOT_SUPPORTED)
3852	{
3853	pImage->pvImageAlloc = RTMemExecAlloc(pImage->cbImageBits + 31);
3854	pImage->pvImage = RT_ALIGN_P(pImage->pvImageAlloc, 32);
3855	pImage->fNative = false;
3856	rc = pImage->pvImageAlloc ? VINF_SUCCESS : VERR_NO_EXEC_MEMORY;
3857	}
3858	if (RT_FAILURE(rc))
3859	{
3860	supdrvLdrUnlock(pDevExt);
3861	RTMemFree(pImage);
3862	Log(("supdrvIOCtl_LdrOpen(%s): failed - %Rrc\n", pReq->u.In.szName, rc));
3863	return rc;
3864	}
3865	Assert(VALID_PTR(pImage->pvImage) \|\| RT_FAILURE(rc));
3866
3867	/*
3868	* Link it.
3869	*/
3870	pImage->pNext = pDevExt->pLdrImages;
3871	pDevExt->pLdrImages = pImage;
3872
3873	supdrvLdrAddUsage(pSession, pImage);
3874
3875	pReq->u.Out.pvImageBase = pImage->pvImage;
3876	pReq->u.Out.fNeedsLoading = true;
3877	pReq->u.Out.fNativeLoader = pImage->fNative;
3878	supdrvLdrUnlock(pDevExt);
3879
3880	#if defined(RT_OS_WINDOWS) && defined(DEBUG)
3881	SUPR0Printf("VBoxDrv: windbg> .reload /f %s=%#p\n", pImage->szName, pImage->pvImage);
3882	#endif
3883	return VINF_SUCCESS;
3884	}
3885
3886
3887	/**
3888	* Worker that validates a pointer to an image entrypoint.
3889	*
3890	* @returns IPRT status code.
3891	* @param pDevExt The device globals.
3892	* @param pImage The loader image.
3893	* @param pv The pointer into the image.
3894	* @param fMayBeNull Whether it may be NULL.
3895	* @param pszWhat What is this entrypoint? (for logging)
3896	* @param pbImageBits The image bits prepared by ring-3.
3897	*
3898	* @remarks Will leave the lock on failure.
3899	*/
3900	static int supdrvLdrValidatePointer(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage, void *pv,
3901	bool fMayBeNull, const uint8_t pbImageBits, const char pszWhat)
3902	{
3903	if (!fMayBeNull \|\| pv)
3904	{
3905	if ((uintptr_t)pv - (uintptr_t)pImage->pvImage >= pImage->cbImageBits)
3906	{
3907	supdrvLdrUnlock(pDevExt);
3908	Log(("Out of range (%p LB %#x): %s=%p\n", pImage->pvImage, pImage->cbImageBits, pszWhat, pv));
3909	return VERR_INVALID_PARAMETER;
3910	}
3911
3912	if (pImage->fNative)
3913	{
3914	int rc = supdrvOSLdrValidatePointer(pDevExt, pImage, pv, pbImageBits);
3915	if (RT_FAILURE(rc))
3916	{
3917	supdrvLdrUnlock(pDevExt);
3918	Log(("Bad entry point address: %s=%p (rc=%Rrc)\n", pszWhat, pv, rc));
3919	return rc;
3920	}
3921	}
3922	}
3923	return VINF_SUCCESS;
3924	}
3925
3926
3927	/**
3928	* Loads the image bits.
3929	*
3930	* This is the 2nd step of the loading.
3931	*
3932	* @returns IPRT status code.
3933	* @param pDevExt Device globals.
3934	* @param pSession Session data.
3935	* @param pReq The request.
3936	*/
3937	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
3938	{
3939	PSUPDRVLDRUSAGE pUsage;
3940	PSUPDRVLDRIMAGE pImage;
3941	int rc;
3942	LogFlow(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImageWithBits=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImageWithTabs));
3943
3944	/*
3945	* Find the ldr image.
3946	*/
3947	supdrvLdrLock(pDevExt);
3948	pUsage = pSession->pLdrUsage;
3949	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3950	pUsage = pUsage->pNext;
3951	if (!pUsage)
3952	{
3953	supdrvLdrUnlock(pDevExt);
3954	Log(("SUP_IOCTL_LDR_LOAD: couldn't find image!\n"));
3955	return VERR_INVALID_HANDLE;
3956	}
3957	pImage = pUsage->pImage;
3958
3959	/*
3960	* Validate input.
3961	*/
3962	if ( pImage->cbImageWithTabs != pReq->u.In.cbImageWithTabs
3963	\|\| pImage->cbImageBits != pReq->u.In.cbImageBits)
3964	{
3965	supdrvLdrUnlock(pDevExt);
3966	Log(("SUP_IOCTL_LDR_LOAD: image size mismatch!! %d(prep) != %d(load) or %d != %d\n",
3967	pImage->cbImageWithTabs, pReq->u.In.cbImageWithTabs, pImage->cbImageBits, pReq->u.In.cbImageBits));
3968	return VERR_INVALID_HANDLE;
3969	}
3970
3971	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
3972	{
3973	unsigned uState = pImage->uState;
3974	supdrvLdrUnlock(pDevExt);
3975	if (uState != SUP_IOCTL_LDR_LOAD)
3976	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
3977	return VERR_ALREADY_LOADED;
3978	}
3979
3980	switch (pReq->u.In.eEPType)
3981	{
3982	case SUPLDRLOADEP_NOTHING:
3983	break;
3984
3985	case SUPLDRLOADEP_VMMR0:
3986	rc = supdrvLdrValidatePointer( pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0, false, pReq->u.In.abImage, "pvVMMR0");
3987	if (RT_SUCCESS(rc))
3988	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt, false, pReq->u.In.abImage, "pvVMMR0EntryInt");
3989	if (RT_SUCCESS(rc))
3990	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, false, pReq->u.In.abImage, "pvVMMR0EntryFast");
3991	if (RT_SUCCESS(rc))
3992	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx, false, pReq->u.In.abImage, "pvVMMR0EntryEx");
3993	if (RT_FAILURE(rc))
3994	return rc;
3995	break;
3996
3997	case SUPLDRLOADEP_SERVICE:
3998	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.Service.pfnServiceReq, false, pReq->u.In.abImage, "pfnServiceReq");
3999	if (RT_FAILURE(rc))
4000	return rc;
4001	if ( pReq->u.In.EP.Service.apvReserved[0] != NIL_RTR0PTR
4002	\|\| pReq->u.In.EP.Service.apvReserved[1] != NIL_RTR0PTR
4003	\|\| pReq->u.In.EP.Service.apvReserved[2] != NIL_RTR0PTR)
4004	{
4005	supdrvLdrUnlock(pDevExt);
4006	Log(("Out of range (%p LB %#x): apvReserved={%p,%p,%p} MBZ!\n",
4007	pImage->pvImage, pReq->u.In.cbImageWithTabs,
4008	pReq->u.In.EP.Service.apvReserved[0],
4009	pReq->u.In.EP.Service.apvReserved[1],
4010	pReq->u.In.EP.Service.apvReserved[2]));
4011	return VERR_INVALID_PARAMETER;
4012	}
4013	break;
4014
4015	default:
4016	supdrvLdrUnlock(pDevExt);
4017	Log(("Invalid eEPType=%d\n", pReq->u.In.eEPType));
4018	return VERR_INVALID_PARAMETER;
4019	}
4020
4021	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleInit, true, pReq->u.In.abImage, "pfnModuleInit");
4022	if (RT_FAILURE(rc))
4023	return rc;
4024	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleTerm, true, pReq->u.In.abImage, "pfnModuleTerm");
4025	if (RT_FAILURE(rc))
4026	return rc;
4027
4028	/*
4029	* Allocate and copy the tables.
4030	* (No need to do try/except as this is a buffered request.)
4031	*/
4032	pImage->cbStrTab = pReq->u.In.cbStrTab;
4033	if (pImage->cbStrTab)
4034	{
4035	pImage->pachStrTab = (char *)RTMemAlloc(pImage->cbStrTab);
4036	if (pImage->pachStrTab)
4037	memcpy(pImage->pachStrTab, &pReq->u.In.abImage[pReq->u.In.offStrTab], pImage->cbStrTab);
4038	else
4039	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_3;
4040	}
4041
4042	pImage->cSymbols = pReq->u.In.cSymbols;
4043	if (RT_SUCCESS(rc) && pImage->cSymbols)
4044	{
4045	size_t cbSymbols = pImage->cSymbols * sizeof(SUPLDRSYM);
4046	pImage->paSymbols = (PSUPLDRSYM)RTMemAlloc(cbSymbols);
4047	if (pImage->paSymbols)
4048	memcpy(pImage->paSymbols, &pReq->u.In.abImage[pReq->u.In.offSymbols], cbSymbols);
4049	else
4050	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_4;
4051	}
4052
4053	/*
4054	* Copy the bits / complete native loading.
4055	*/
4056	if (RT_SUCCESS(rc))
4057	{
4058	pImage->uState = SUP_IOCTL_LDR_LOAD;
4059	pImage->pfnModuleInit = pReq->u.In.pfnModuleInit;
4060	pImage->pfnModuleTerm = pReq->u.In.pfnModuleTerm;
4061
4062	if (pImage->fNative)
4063	rc = supdrvOSLdrLoad(pDevExt, pImage, pReq->u.In.abImage);
4064	else
4065	memcpy(pImage->pvImage, &pReq->u.In.abImage[0], pImage->cbImageBits);
4066	}
4067
4068	/*
4069	* Update any entry points.
4070	*/
4071	if (RT_SUCCESS(rc))
4072	{
4073	switch (pReq->u.In.eEPType)
4074	{
4075	default:
4076	case SUPLDRLOADEP_NOTHING:
4077	rc = VINF_SUCCESS;
4078	break;
4079	case SUPLDRLOADEP_VMMR0:
4080	rc = supdrvLdrSetVMMR0EPs(pDevExt, pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
4081	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
4082	break;
4083	case SUPLDRLOADEP_SERVICE:
4084	pImage->pfnServiceReqHandler = pReq->u.In.EP.Service.pfnServiceReq;
4085	rc = VINF_SUCCESS;
4086	break;
4087	}
4088	}
4089
4090	/*
4091	* On success call the module initialization.
4092	*/
4093	LogFlow(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
4094	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
4095	{
4096	Log(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
4097	rc = pImage->pfnModuleInit();
4098	if (rc && pDevExt->pvVMMR0 == pImage->pvImage)
4099	supdrvLdrUnsetVMMR0EPs(pDevExt);
4100	}
4101
4102	if (RT_FAILURE(rc))
4103	{
4104	pImage->uState = SUP_IOCTL_LDR_OPEN;
4105	pImage->pfnModuleInit = NULL;
4106	pImage->pfnModuleTerm = NULL;
4107	pImage->pfnServiceReqHandler= NULL;
4108	pImage->cbStrTab = 0;
4109	RTMemFree(pImage->pachStrTab);
4110	pImage->pachStrTab = NULL;
4111	RTMemFree(pImage->paSymbols);
4112	pImage->paSymbols = NULL;
4113	pImage->cSymbols = 0;
4114	}
4115
4116	supdrvLdrUnlock(pDevExt);
4117	return rc;
4118	}
4119
4120
4121	/**
4122	* Frees a previously loaded (prep'ed) image.
4123	*
4124	* @returns IPRT status code.
4125	* @param pDevExt Device globals.
4126	* @param pSession Session data.
4127	* @param pReq The request.
4128	*/
4129	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
4130	{
4131	int rc;
4132	PSUPDRVLDRUSAGE pUsagePrev;
4133	PSUPDRVLDRUSAGE pUsage;
4134	PSUPDRVLDRIMAGE pImage;
4135	LogFlow(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
4136
4137	/*
4138	* Find the ldr image.
4139	*/
4140	supdrvLdrLock(pDevExt);
4141	pUsagePrev = NULL;
4142	pUsage = pSession->pLdrUsage;
4143	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4144	{
4145	pUsagePrev = pUsage;
4146	pUsage = pUsage->pNext;
4147	}
4148	if (!pUsage)
4149	{
4150	supdrvLdrUnlock(pDevExt);
4151	Log(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
4152	return VERR_INVALID_HANDLE;
4153	}
4154
4155	/*
4156	* Check if we can remove anything.
4157	*/
4158	rc = VINF_SUCCESS;
4159	pImage = pUsage->pImage;
4160	if (pImage->cUsage <= 1 \|\| pUsage->cUsage <= 1)
4161	{
4162	/*
4163	* Check if there are any objects with destructors in the image, if
4164	* so leave it for the session cleanup routine so we get a chance to
4165	* clean things up in the right order and not leave them all dangling.
4166	*/
4167	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
4168	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
4169	if (pImage->cUsage <= 1)
4170	{
4171	PSUPDRVOBJ pObj;
4172	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4173	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4174	{
4175	rc = VERR_DANGLING_OBJECTS;
4176	break;
4177	}
4178	}
4179	else
4180	{
4181	PSUPDRVUSAGE pGenUsage;
4182	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
4183	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4184	{
4185	rc = VERR_DANGLING_OBJECTS;
4186	break;
4187	}
4188	}
4189	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
4190	if (rc == VINF_SUCCESS)
4191	{
4192	/* unlink it */
4193	if (pUsagePrev)
4194	pUsagePrev->pNext = pUsage->pNext;
4195	else
4196	pSession->pLdrUsage = pUsage->pNext;
4197
4198	/* free it */
4199	pUsage->pImage = NULL;
4200	pUsage->pNext = NULL;
4201	RTMemFree(pUsage);
4202
4203	/*
4204	* Dereference the image.
4205	*/
4206	if (pImage->cUsage <= 1)
4207	supdrvLdrFree(pDevExt, pImage);
4208	else
4209	pImage->cUsage--;
4210	}
4211	else
4212	{
4213	Log(("supdrvIOCtl_LdrFree: Dangling objects in %p/%s!\n", pImage->pvImage, pImage->szName));
4214	rc = VINF_SUCCESS; /** @todo BRANCH-2.1: remove this after branching. */
4215	}
4216	}
4217	else
4218	{
4219	/*
4220	* Dereference both image and usage.
4221	*/
4222	pImage->cUsage--;
4223	pUsage->cUsage--;
4224	}
4225
4226	supdrvLdrUnlock(pDevExt);
4227	return rc;
4228	}
4229
4230
4231	/**
4232	* Gets the address of a symbol in an open image.
4233	*
4234	* @returns IPRT status code.
4235	* @param pDevExt Device globals.
4236	* @param pSession Session data.
4237	* @param pReq The request buffer.
4238	*/
4239	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
4240	{
4241	PSUPDRVLDRIMAGE pImage;
4242	PSUPDRVLDRUSAGE pUsage;
4243	uint32_t i;
4244	PSUPLDRSYM paSyms;
4245	const char *pchStrings;
4246	const size_t cbSymbol = strlen(pReq->u.In.szSymbol) + 1;
4247	void *pvSymbol = NULL;
4248	int rc = VERR_GENERAL_FAILURE;
4249	Log3(("supdrvIOCtl_LdrGetSymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
4250
4251	/*
4252	* Find the ldr image.
4253	*/
4254	supdrvLdrLock(pDevExt);
4255	pUsage = pSession->pLdrUsage;
4256	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4257	pUsage = pUsage->pNext;
4258	if (!pUsage)
4259	{
4260	supdrvLdrUnlock(pDevExt);
4261	Log(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
4262	return VERR_INVALID_HANDLE;
4263	}
4264	pImage = pUsage->pImage;
4265	if (pImage->uState != SUP_IOCTL_LDR_LOAD)
4266	{
4267	unsigned uState = pImage->uState;
4268	supdrvLdrUnlock(pDevExt);
4269	Log(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", uState, uState)); NOREF(uState);
4270	return VERR_ALREADY_LOADED;
4271	}
4272
4273	/*
4274	* Search the symbol strings.
4275	*/
4276	pchStrings = pImage->pachStrTab;
4277	paSyms = pImage->paSymbols;
4278	for (i = 0; i < pImage->cSymbols; i++)
4279	{
4280	if ( paSyms[i].offSymbol < pImage->cbImageBits /* paranoia */
4281	&& paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4282	&& !memcmp(pchStrings + paSyms[i].offName, pReq->u.In.szSymbol, cbSymbol))
4283	{
4284	pvSymbol = (uint8_t *)pImage->pvImage + paSyms[i].offSymbol;
4285	rc = VINF_SUCCESS;
4286	break;
4287	}
4288	}
4289	supdrvLdrUnlock(pDevExt);
4290	pReq->u.Out.pvSymbol = pvSymbol;
4291	return rc;
4292	}
4293
4294
4295	/**
4296	* Gets the address of a symbol in an open image or the support driver.
4297	*
4298	* @returns VINF_SUCCESS on success.
4299	* @returns
4300	* @param pDevExt Device globals.
4301	* @param pSession Session data.
4302	* @param pReq The request buffer.
4303	*/
4304	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq)
4305	{
4306	int rc = VINF_SUCCESS;
4307	const char *pszSymbol = pReq->u.In.pszSymbol;
4308	const char *pszModule = pReq->u.In.pszModule;
4309	size_t cbSymbol;
4310	char const *pszEnd;
4311	uint32_t i;
4312
4313	/*
4314	* Input validation.
4315	*/
4316	AssertPtrReturn(pszSymbol, VERR_INVALID_POINTER);
4317	pszEnd = RTStrEnd(pszSymbol, 512);
4318	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4319	cbSymbol = pszEnd - pszSymbol + 1;
4320
4321	if (pszModule)
4322	{
4323	AssertPtrReturn(pszModule, VERR_INVALID_POINTER);
4324	pszEnd = RTStrEnd(pszModule, 64);
4325	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4326	}
4327	Log3(("supdrvIDC_LdrGetSymbol: pszModule=%p:{%s} pszSymbol=%p:{%s}\n", pszModule, pszModule, pszSymbol, pszSymbol));
4328
4329
4330	if ( !pszModule
4331	\|\| !strcmp(pszModule, "SupDrv"))
4332	{
4333	/*
4334	* Search the support driver export table.
4335	*/
4336	for (i = 0; i < RT_ELEMENTS(g_aFunctions); i++)
4337	if (!strcmp(g_aFunctions[i].szName, pszSymbol))
4338	{
4339	pReq->u.Out.pfnSymbol = g_aFunctions[i].pfn;
4340	break;
4341	}
4342	}
4343	else
4344	{
4345	/*
4346	* Find the loader image.
4347	*/
4348	PSUPDRVLDRIMAGE pImage;
4349
4350	supdrvLdrLock(pDevExt);
4351
4352	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
4353	if (!strcmp(pImage->szName, pszModule))
4354	break;
4355	if (pImage && pImage->uState == SUP_IOCTL_LDR_LOAD)
4356	{
4357	/*
4358	* Search the symbol strings.
4359	*/
4360	const char *pchStrings = pImage->pachStrTab;
4361	PCSUPLDRSYM paSyms = pImage->paSymbols;
4362	for (i = 0; i < pImage->cSymbols; i++)
4363	{
4364	if ( paSyms[i].offSymbol < pImage->cbImageBits /* paranoia */
4365	&& paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4366	&& !memcmp(pchStrings + paSyms[i].offName, pszSymbol, cbSymbol))
4367	{
4368	/*
4369	* Found it! Calc the symbol address and add a reference to the module.
4370	*/
4371	pReq->u.Out.pfnSymbol = (PFNRT)((uint8_t *)pImage->pvImage + paSyms[i].offSymbol);
4372	rc = supdrvLdrAddUsage(pSession, pImage);
4373	break;
4374	}
4375	}
4376	}
4377	else
4378	rc = pImage ? VERR_WRONG_ORDER : VERR_MODULE_NOT_FOUND;
4379
4380	supdrvLdrUnlock(pDevExt);
4381	}
4382	return rc;
4383	}
4384
4385
4386	/**
4387	* Updates the VMMR0 entry point pointers.
4388	*
4389	* @returns IPRT status code.
4390	* @param pDevExt Device globals.
4391	* @param pSession Session data.
4392	* @param pVMMR0 VMMR0 image handle.
4393	* @param pvVMMR0EntryInt VMMR0EntryInt address.
4394	* @param pvVMMR0EntryFast VMMR0EntryFast address.
4395	* @param pvVMMR0EntryEx VMMR0EntryEx address.
4396	* @remark Caller must own the loader mutex.
4397	*/
4398	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx)
4399	{
4400	int rc = VINF_SUCCESS;
4401	LogFlow(("supdrvLdrSetR0EP pvVMMR0=%p pvVMMR0EntryInt=%p\n", pvVMMR0, pvVMMR0EntryInt));
4402
4403
4404	/*
4405	* Check if not yet set.
4406	*/
4407	if (!pDevExt->pvVMMR0)
4408	{
4409	pDevExt->pvVMMR0 = pvVMMR0;
4410	pDevExt->pfnVMMR0EntryInt = pvVMMR0EntryInt;
4411	pDevExt->pfnVMMR0EntryFast = pvVMMR0EntryFast;
4412	pDevExt->pfnVMMR0EntryEx = pvVMMR0EntryEx;
4413	}
4414	else
4415	{
4416	/*
4417	* Return failure or success depending on whether the values match or not.
4418	*/
4419	if ( pDevExt->pvVMMR0 != pvVMMR0
4420	\|\| (void *)pDevExt->pfnVMMR0EntryInt != pvVMMR0EntryInt
4421	\|\| (void *)pDevExt->pfnVMMR0EntryFast != pvVMMR0EntryFast
4422	\|\| (void *)pDevExt->pfnVMMR0EntryEx != pvVMMR0EntryEx)
4423	{
4424	AssertMsgFailed(("SUP_IOCTL_LDR_SETR0EP: Already set pointing to a different module!\n"));
4425	rc = VERR_INVALID_PARAMETER;
4426	}
4427	}
4428	return rc;
4429	}
4430
4431
4432	/**
4433	* Unsets the VMMR0 entry point installed by supdrvLdrSetR0EP.
4434	*
4435	* @param pDevExt Device globals.
4436	*/
4437	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt)
4438	{
4439	pDevExt->pvVMMR0 = NULL;
4440	pDevExt->pfnVMMR0EntryInt = NULL;
4441	pDevExt->pfnVMMR0EntryFast = NULL;
4442	pDevExt->pfnVMMR0EntryEx = NULL;
4443	}
4444
4445
4446	/**
4447	* Adds a usage reference in the specified session of an image.
4448	*
4449	* Called while owning the loader semaphore.
4450	*
4451	* @returns VINF_SUCCESS on success and VERR_NO_MEMORY on failure.
4452	* @param pSession Session in question.
4453	* @param pImage Image which the session is using.
4454	*/
4455	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage)
4456	{
4457	PSUPDRVLDRUSAGE pUsage;
4458	LogFlow(("supdrvLdrAddUsage: pImage=%p\n", pImage));
4459
4460	/*
4461	* Referenced it already?
4462	*/
4463	pUsage = pSession->pLdrUsage;
4464	while (pUsage)
4465	{
4466	if (pUsage->pImage == pImage)
4467	{
4468	pUsage->cUsage++;
4469	return VINF_SUCCESS;
4470	}
4471	pUsage = pUsage->pNext;
4472	}
4473
4474	/*
4475	* Allocate new usage record.
4476	*/
4477	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
4478	AssertReturn(pUsage, /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_5);
4479	pUsage->cUsage = 1;
4480	pUsage->pImage = pImage;
4481	pUsage->pNext = pSession->pLdrUsage;
4482	pSession->pLdrUsage = pUsage;
4483	return VINF_SUCCESS;
4484	}
4485
4486
4487	/**
4488	* Frees a load image.
4489	*
4490	* @param pDevExt Pointer to device extension.
4491	* @param pImage Pointer to the image we're gonna free.
4492	* This image must exit!
4493	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
4494	*/
4495	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
4496	{
4497	PSUPDRVLDRIMAGE pImagePrev;
4498	LogFlow(("supdrvLdrFree: pImage=%p\n", pImage));
4499
4500	/* find it - arg. should've used doubly linked list. */
4501	Assert(pDevExt->pLdrImages);
4502	pImagePrev = NULL;
4503	if (pDevExt->pLdrImages != pImage)
4504	{
4505	pImagePrev = pDevExt->pLdrImages;
4506	while (pImagePrev->pNext != pImage)
4507	pImagePrev = pImagePrev->pNext;
4508	Assert(pImagePrev->pNext == pImage);
4509	}
4510
4511	/* unlink */
4512	if (pImagePrev)
4513	pImagePrev->pNext = pImage->pNext;
4514	else
4515	pDevExt->pLdrImages = pImage->pNext;
4516
4517	/* check if this is VMMR0.r0 unset its entry point pointers. */
4518	if (pDevExt->pvVMMR0 == pImage->pvImage)
4519	supdrvLdrUnsetVMMR0EPs(pDevExt);
4520
4521	/* check for objects with destructors in this image. (Shouldn't happen.) */
4522	if (pDevExt->pObjs)
4523	{
4524	unsigned cObjs = 0;
4525	PSUPDRVOBJ pObj;
4526	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
4527	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
4528	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4529	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4530	{
4531	pObj->pfnDestructor = NULL;
4532	cObjs++;
4533	}
4534	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
4535	if (cObjs)
4536	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
4537	}
4538
4539	/* call termination function if fully loaded. */
4540	if ( pImage->pfnModuleTerm
4541	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
4542	{
4543	LogFlow(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
4544	pImage->pfnModuleTerm();
4545	}
4546
4547	/* do native unload if appropriate. */
4548	if (pImage->fNative)
4549	supdrvOSLdrUnload(pDevExt, pImage);
4550
4551	/* free the image */
4552	pImage->cUsage = 0;
4553	pImage->pNext = 0;
4554	pImage->uState = SUP_IOCTL_LDR_FREE;
4555	RTMemExecFree(pImage->pvImageAlloc, pImage->cbImageBits + 31);
4556	pImage->pvImageAlloc = NULL;
4557	RTMemFree(pImage->pachStrTab);
4558	pImage->pachStrTab = NULL;
4559	RTMemFree(pImage->paSymbols);
4560	pImage->paSymbols = NULL;
4561	RTMemFree(pImage);
4562	}
4563
4564
4565	/**
4566	* Acquires the loader lock.
4567	*
4568	* @returns IPRT status code.
4569	* @param pDevExt The device extension.
4570	*/
4571	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt)
4572	{
4573	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4574	int rc = RTSemMutexRequest(pDevExt->mtxLdr, RT_INDEFINITE_WAIT);
4575	#else
4576	int rc = RTSemFastMutexRequest(pDevExt->mtxLdr);
4577	#endif
4578	AssertRC(rc);
4579	return rc;
4580	}
4581
4582
4583	/**
4584	* Releases the loader lock.
4585	*
4586	* @returns IPRT status code.
4587	* @param pDevExt The device extension.
4588	*/
4589	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt)
4590	{
4591	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4592	return RTSemMutexRelease(pDevExt->mtxLdr);
4593	#else
4594	return RTSemFastMutexRelease(pDevExt->mtxLdr);
4595	#endif
4596	}
4597
4598
4599	/**
4600	* Implements the service call request.
4601	*
4602	* @returns VBox status code.
4603	* @param pDevExt The device extension.
4604	* @param pSession The calling session.
4605	* @param pReq The request packet, valid.
4606	*/
4607	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq)
4608	{
4609	#if !defined(RT_OS_WINDOWS) \|\| defined(DEBUG)
4610	int rc;
4611
4612	/*
4613	* Find the module first in the module referenced by the calling session.
4614	*/
4615	rc = supdrvLdrLock(pDevExt);
4616	if (RT_SUCCESS(rc))
4617	{
4618	PFNSUPR0SERVICEREQHANDLER pfnServiceReqHandler = NULL;
4619	PSUPDRVLDRUSAGE pUsage;
4620
4621	for (pUsage = pSession->pLdrUsage; pUsage; pUsage = pUsage->pNext)
4622	if ( pUsage->pImage->pfnServiceReqHandler
4623	&& !strcmp(pUsage->pImage->szName, pReq->u.In.szName))
4624	{
4625	pfnServiceReqHandler = pUsage->pImage->pfnServiceReqHandler;
4626	break;
4627	}
4628	supdrvLdrUnlock(pDevExt);
4629
4630	if (pfnServiceReqHandler)
4631	{
4632	/*
4633	* Call it.
4634	*/
4635	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
4636	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, NULL);
4637	else
4638	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0]);
4639	}
4640	else
4641	rc = VERR_SUPDRV_SERVICE_NOT_FOUND;
4642	}
4643
4644	/* log it */
4645	if ( RT_FAILURE(rc)
4646	&& rc != VERR_INTERRUPTED
4647	&& rc != VERR_TIMEOUT)
4648	Log(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4649	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4650	else
4651	Log4(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4652	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4653	return rc;
4654	#else /* RT_OS_WINDOWS && !DEBUG */
4655	return VERR_NOT_IMPLEMENTED;
4656	#endif /* RT_OS_WINDOWS && !DEBUG */
4657	}
4658
4659
4660	/**
4661	* Implements the logger settings request.
4662	*
4663	* @returns VBox status code.
4664	* @param pDevExt The device extension.
4665	* @param pSession The caller's session.
4666	* @param pReq The request.
4667	*/
4668	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq)
4669	{
4670	const char *pszGroup = &pReq->u.In.szStrings[pReq->u.In.offGroups];
4671	const char *pszFlags = &pReq->u.In.szStrings[pReq->u.In.offFlags];
4672	const char *pszDest = &pReq->u.In.szStrings[pReq->u.In.offDestination];
4673	PRTLOGGER pLogger = NULL;
4674	int rc;
4675
4676	/*
4677	* Some further validation.
4678	*/
4679	switch (pReq->u.In.fWhat)
4680	{
4681	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4682	case SUPLOGGERSETTINGS_WHAT_CREATE:
4683	break;
4684
4685	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4686	if (pszGroup \|\| pszFlags \|\| *pszDest)
4687	return VERR_INVALID_PARAMETER;
4688	if (pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_RELEASE)
4689	return VERR_ACCESS_DENIED;
4690	break;
4691
4692	default:
4693	return VERR_INTERNAL_ERROR;
4694	}
4695
4696	/*
4697	* Get the logger.
4698	*/
4699	switch (pReq->u.In.fWhich)
4700	{
4701	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4702	pLogger = RTLogGetDefaultInstance();
4703	break;
4704
4705	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4706	pLogger = RTLogRelDefaultInstance();
4707	break;
4708
4709	default:
4710	return VERR_INTERNAL_ERROR;
4711	}
4712
4713	/*
4714	* Do the job.
4715	*/
4716	switch (pReq->u.In.fWhat)
4717	{
4718	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4719	if (pLogger)
4720	{
4721	rc = RTLogFlags(pLogger, pszFlags);
4722	if (RT_SUCCESS(rc))
4723	rc = RTLogGroupSettings(pLogger, pszGroup);
4724	NOREF(pszDest);
4725	}
4726	else
4727	rc = VERR_NOT_FOUND;
4728	break;
4729
4730	case SUPLOGGERSETTINGS_WHAT_CREATE:
4731	{
4732	if (pLogger)
4733	rc = VERR_ALREADY_EXISTS;
4734	else
4735	{
4736	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
4737
4738	rc = RTLogCreate(&pLogger,
4739	0 /* fFlags */,
4740	pszGroup,
4741	pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_DEBUG
4742	? "VBOX_LOG"
4743	: "VBOX_RELEASE_LOG",
4744	RT_ELEMENTS(s_apszGroups),
4745	s_apszGroups,
4746	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER,
4747	NULL);
4748	if (RT_SUCCESS(rc))
4749	{
4750	rc = RTLogFlags(pLogger, pszFlags);
4751	NOREF(pszDest);
4752	if (RT_SUCCESS(rc))
4753	{
4754	switch (pReq->u.In.fWhich)
4755	{
4756	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4757	pLogger = RTLogSetDefaultInstance(pLogger);
4758	break;
4759	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4760	pLogger = RTLogRelSetDefaultInstance(pLogger);
4761	break;
4762	}
4763	}
4764	RTLogDestroy(pLogger);
4765	}
4766	}
4767	break;
4768	}
4769
4770	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4771	switch (pReq->u.In.fWhich)
4772	{
4773	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4774	pLogger = RTLogSetDefaultInstance(NULL);
4775	break;
4776	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4777	pLogger = RTLogRelSetDefaultInstance(NULL);
4778	break;
4779	}
4780	rc = RTLogDestroy(pLogger);
4781	break;
4782
4783	default:
4784	{
4785	rc = VERR_INTERNAL_ERROR;
4786	break;
4787	}
4788	}
4789
4790	return rc;
4791	}
4792
4793
4794	/**
4795	* Creates the GIP.
4796	*
4797	* @returns VBox status code.
4798	* @param pDevExt Instance data. GIP stuff may be updated.
4799	*/
4800	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
4801	{
4802	PSUPGLOBALINFOPAGE pGip;
4803	RTHCPHYS HCPhysGip;
4804	uint32_t u32SystemResolution;
4805	uint32_t u32Interval;
4806	unsigned cCpus;
4807	int rc;
4808
4809
4810	LogFlow(("supdrvGipCreate:\n"));
4811
4812	/* assert order */
4813	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
4814	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
4815	Assert(!pDevExt->pGipTimer);
4816
4817	/*
4818	* Check the CPU count.
4819	*/
4820	cCpus = RTMpGetArraySize();
4821	if ( cCpus > RTCPUSET_MAX_CPUS
4822	\|\| cCpus > 256 /ApicId is used for the mappings/)
4823	{
4824	SUPR0Printf("VBoxDrv: Too many CPUs (%u) for the GIP (max %u)\n", cCpus, RT_MIN(RTCPUSET_MAX_CPUS, 256));
4825	return VERR_TOO_MANY_CPUS;
4826	}
4827
4828	/*
4829	* Allocate a contiguous set of pages with a default kernel mapping.
4830	*/
4831	rc = RTR0MemObjAllocCont(&pDevExt->GipMemObj, RT_UOFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), false /fExecutable/);
4832	if (RT_FAILURE(rc))
4833	{
4834	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
4835	return rc;
4836	}
4837	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
4838	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
4839
4840	/*
4841	* Find a reasonable update interval and initialize the structure.
4842	*/
4843	u32Interval = u32SystemResolution = RTTimerGetSystemGranularity();
4844	while (u32Interval < 10000000 /* 10 ms */)
4845	u32Interval += u32SystemResolution;
4846
4847	supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), 1000000000 / u32Interval /=Hz/, cCpus);
4848
4849	/*
4850	* Create the timer.
4851	* If CPU_ALL isn't supported we'll have to fall back to synchronous mode.
4852	*/
4853	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
4854	{
4855	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, RTTIMER_FLAGS_CPU_ALL, supdrvGipAsyncTimer, pDevExt);
4856	if (rc == VERR_NOT_SUPPORTED)
4857	{
4858	OSDBGPRINT(("supdrvGipCreate: omni timer not supported, falling back to synchronous mode\n"));
4859	pGip->u32Mode = SUPGIPMODE_SYNC_TSC;
4860	}
4861	}
4862	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
4863	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0, supdrvGipSyncTimer, pDevExt);
4864	if (RT_SUCCESS(rc))
4865	{
4866	rc = RTMpNotificationRegister(supdrvGipMpEvent, pDevExt);
4867	if (RT_SUCCESS(rc))
4868	{
4869	rc = RTMpOnAll(supdrvGipInitOnCpu, pDevExt, pGip);
4870	if (RT_SUCCESS(rc))
4871	{
4872	/*
4873	* We're good.
4874	*/
4875	Log(("supdrvGipCreate: %u ns interval.\n", u32Interval));
4876	g_pSUPGlobalInfoPage = pGip;
4877	return VINF_SUCCESS;
4878	}
4879
4880	OSDBGPRINT(("supdrvGipCreate: RTMpOnAll failed with rc=%Rrc\n", rc));
4881	RTMpNotificationDeregister(supdrvGipMpEvent, pDevExt);
4882
4883	}
4884	else
4885	OSDBGPRINT(("supdrvGipCreate: failed to register MP event notfication. rc=%Rrc\n", rc));
4886	}
4887	else
4888	{
4889	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %u ns interval. rc=%Rrc\n", u32Interval, rc));
4890	Assert(!pDevExt->pGipTimer);
4891	}
4892	supdrvGipDestroy(pDevExt);
4893	return rc;
4894	}
4895
4896
4897	/**
4898	* Terminates the GIP.
4899	*
4900	* @param pDevExt Instance data. GIP stuff may be updated.
4901	*/
4902	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
4903	{
4904	int rc;
4905	#ifdef DEBUG_DARWIN_GIP
4906	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
4907	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
4908	pDevExt->pGipTimer, pDevExt->GipMemObj));
4909	#endif
4910
4911	/*
4912	* Invalid the GIP data.
4913	*/
4914	if (pDevExt->pGip)
4915	{
4916	supdrvGipTerm(pDevExt->pGip);
4917	pDevExt->pGip = NULL;
4918	}
4919	g_pSUPGlobalInfoPage = NULL;
4920
4921	/*
4922	* Destroy the timer and free the GIP memory object.
4923	*/
4924	if (pDevExt->pGipTimer)
4925	{
4926	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
4927	pDevExt->pGipTimer = NULL;
4928	}
4929
4930	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
4931	{
4932	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
4933	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
4934	}
4935
4936	/*
4937	* Finally, make sure we've release the system timer resolution request
4938	* if one actually succeeded and is still pending.
4939	*/
4940	if (pDevExt->u32SystemTimerGranularityGrant)
4941	{
4942	rc = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
4943	pDevExt->u32SystemTimerGranularityGrant = 0;
4944	}
4945	}
4946
4947
4948	/**
4949	* Timer callback function sync GIP mode.
4950	* @param pTimer The timer.
4951	* @param pvUser The device extension.
4952	*/
4953	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
4954	{
4955	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
4956	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4957	uint64_t u64TSC = ASMReadTSC();
4958	uint64_t NanoTS = RTTimeSystemNanoTS();
4959
4960	supdrvGipUpdate(pDevExt->pGip, NanoTS, u64TSC, NIL_RTCPUID, iTick);
4961
4962	ASMSetFlags(fOldFlags);
4963	}
4964
4965
4966	/**
4967	* Timer callback function for async GIP mode.
4968	* @param pTimer The timer.
4969	* @param pvUser The device extension.
4970	*/
4971	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
4972	{
4973	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
4974	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4975	RTCPUID idCpu = RTMpCpuId();
4976	uint64_t u64TSC = ASMReadTSC();
4977	uint64_t NanoTS = RTTimeSystemNanoTS();
4978
4979	/** @todo reset the transaction number and whatnot when iTick == 1. */
4980	if (pDevExt->idGipMaster == idCpu)
4981	supdrvGipUpdate(pDevExt->pGip, NanoTS, u64TSC, idCpu, iTick);
4982	else
4983	supdrvGipUpdatePerCpu(pDevExt->pGip, NanoTS, u64TSC, idCpu, ASMGetApicId(), iTick);
4984
4985	ASMSetFlags(fOldFlags);
4986	}
4987
4988
4989	/**
4990	* The calling CPU should be accounted as online, update GIP accordingly.
4991	*
4992	* This is used by supdrvGipMpEvent as well as the supdrvGipCreate.
4993	*
4994	* @param pGip The GIP.
4995	* @param idCpu The CPU ID.
4996	*/
4997	static void supdrvGipMpEventOnline(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
4998	{
4999	int iCpuSet;
5000	uint8_t idApic;
5001	uint32_t i;
5002
5003	Assert(idCpu == RTMpCpuId());
5004	Assert(pGip->cPossibleCpus == RTMpGetCount());
5005
5006	/*
5007	* Update the globals.
5008	*/
5009	ASMAtomicWriteU16(&pGip->cPresentCpus, RTMpGetPresentCount());
5010	ASMAtomicWriteU16(&pGip->cOnlineCpus, RTMpGetOnlineCount());
5011	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5012	if (iCpuSet >= 0)
5013	{
5014	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5015	RTCpuSetAddByIndex(&pGip->OnlineCpuSet, iCpuSet);
5016	RTCpuSetAddByIndex(&pGip->PresentCpuSet, iCpuSet);
5017	}
5018
5019	/*
5020	* Find our entry, or allocate one if not found.
5021	* ASSUMES that CPU IDs are constant.
5022	*/
5023	for (i = 0; i < pGip->cCpus; i++)
5024	if (pGip->aCPUs[i].idCpu == idCpu)
5025	break;
5026
5027	if (i >= pGip->cCpus)
5028	for (i = 0; i < pGip->cCpus; i++)
5029	{
5030	bool fRc;
5031	ASMAtomicCmpXchgSize(&pGip->aCPUs[i].idCpu, idCpu, NIL_RTCPUID, fRc);
5032	if (fRc)
5033	break;
5034	}
5035
5036	AssertReturnVoid(i < pGip->cCpus);
5037
5038	/*
5039	* Update the entry.
5040	*/
5041	idApic = ASMGetApicId();
5042	ASMAtomicUoWriteU16(&pGip->aCPUs[i].idApic, idApic);
5043	ASMAtomicUoWriteS16(&pGip->aCPUs[i].iCpuSet, (int16_t)iCpuSet);
5044	ASMAtomicUoWriteSize(&pGip->aCPUs[i].idCpu, idCpu);
5045	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_ONLINE);
5046
5047	/*
5048	* Update the APIC ID and CPU set index mappings.
5049	*/
5050	ASMAtomicWriteU16(&pGip->aiCpuFromApicId[idApic], i);
5051	ASMAtomicWriteU16(&pGip->aiCpuFromCpuSetIdx[iCpuSet], i);
5052	}
5053
5054
5055	/**
5056	* The CPU should be accounted as offline, update the GIP accordingly.
5057	*
5058	* This is used by supdrvGipMpEvent.
5059	*
5060	* @param pGip The GIP.
5061	* @param idCpu The CPU ID.
5062	*/
5063	static void supdrvGipMpEventOffline(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
5064	{
5065	int iCpuSet;
5066	unsigned i;
5067
5068	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5069	AssertReturnVoid(iCpuSet >= 0);
5070
5071	i = pGip->aiCpuFromCpuSetIdx[iCpuSet];
5072	AssertReturnVoid(i < pGip->cCpus);
5073	AssertReturnVoid(pGip->aCPUs[i].idCpu == idCpu);
5074
5075	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5076	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_OFFLINE);
5077	RTCpuSetDelByIndex(&pGip->OnlineCpuSet, iCpuSet);
5078	}
5079
5080
5081	/**
5082	* Multiprocessor event notification callback.
5083	*
5084	* This is used to make sure that the GIP master gets passed on to
5085	* another CPU. It also updates the associated CPU data.
5086	*
5087	* @param enmEvent The event.
5088	* @param idCpu The cpu it applies to.
5089	* @param pvUser Pointer to the device extension.
5090	*/
5091	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser)
5092	{
5093	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5094	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5095
5096	/*
5097	* Update the GIP CPU data.
5098	*/
5099	if (pGip)
5100	{
5101	switch (enmEvent)
5102	{
5103	case RTMPEVENT_ONLINE:
5104	supdrvGipMpEventOnline(pGip, idCpu);
5105	break;
5106	case RTMPEVENT_OFFLINE:
5107	supdrvGipMpEventOffline(pGip, idCpu);
5108	break;
5109
5110	}
5111	}
5112
5113	/*
5114	* Make sure there is a master GIP.
5115	*/
5116	if (enmEvent == RTMPEVENT_OFFLINE)
5117	{
5118	RTCPUID idGipMaster;
5119	ASMAtomicReadSize(&pDevExt->idGipMaster, &idGipMaster);
5120	if (idGipMaster == idCpu)
5121	{
5122	/*
5123	* Find a new GIP master.
5124	*/
5125	bool fIgnored;
5126	unsigned i;
5127	RTCPUID idNewGipMaster = NIL_RTCPUID;
5128	RTCPUSET OnlineCpus;
5129	RTMpGetOnlineSet(&OnlineCpus);
5130
5131	for (i = 0; i < RTCPUSET_MAX_CPUS; i++)
5132	{
5133	RTCPUID idCurCpu = RTMpCpuIdFromSetIndex(i);
5134	if ( RTCpuSetIsMember(&OnlineCpus, idCurCpu)
5135	&& idCurCpu != idGipMaster)
5136	{
5137	idNewGipMaster = idCurCpu;
5138	break;
5139	}
5140	}
5141
5142	Log(("supdrvGipMpEvent: Gip master %#lx -> %#lx\n", (long)idGipMaster, (long)idNewGipMaster));
5143	ASMAtomicCmpXchgSize(&pDevExt->idGipMaster, idNewGipMaster, idGipMaster, fIgnored);
5144	NOREF(fIgnored);
5145	}
5146	}
5147	}
5148
5149
5150	/**
5151	* Callback used by supdrvDetermineAsyncTSC to read the TSC on a CPU.
5152	*
5153	* @param idCpu Ignored.
5154	* @param pvUser1 Where to put the TSC.
5155	* @param pvUser2 Ignored.
5156	*/
5157	static DECLCALLBACK(void) supdrvDetermineAsyncTscWorker(RTCPUID idCpu, void pvUser1, void pvUser2)
5158	{
5159	#if 1
5160	ASMAtomicWriteU64((uint64_t volatile *)pvUser1, ASMReadTSC());
5161	#else
5162	(uint64_t )pvUser1 = ASMReadTSC();
5163	#endif
5164	}
5165
5166
5167	/**
5168	* Determine if Async GIP mode is required because of TSC drift.
5169	*
5170	* When using the default/normal timer code it is essential that the time stamp counter
5171	* (TSC) runs never backwards, that is, a read operation to the counter should return
5172	* a bigger value than any previous read operation. This is guaranteed by the latest
5173	* AMD CPUs and by newer Intel CPUs which never enter the C2 state (P4). In any other
5174	* case we have to choose the asynchronous timer mode.
5175	*
5176	* @param poffMin Pointer to the determined difference between different cores.
5177	* @return false if the time stamp counters appear to be synchronized, true otherwise.
5178	*/
5179	static bool supdrvDetermineAsyncTsc(uint64_t *poffMin)
5180	{
5181	/*
5182	* Just iterate all the cpus 8 times and make sure that the TSC is
5183	* ever increasing. We don't bother taking TSC rollover into account.
5184	*/
5185	int iEndCpu = RTMpGetArraySize();
5186	int iCpu;
5187	int cLoops = 8;
5188	bool fAsync = false;
5189	int rc = VINF_SUCCESS;
5190	uint64_t offMax = 0;
5191	uint64_t offMin = ~(uint64_t)0;
5192	uint64_t PrevTsc = ASMReadTSC();
5193
5194	while (cLoops-- > 0)
5195	{
5196	for (iCpu = 0; iCpu < iEndCpu; iCpu++)
5197	{
5198	uint64_t CurTsc;
5199	rc = RTMpOnSpecific(RTMpCpuIdFromSetIndex(iCpu), supdrvDetermineAsyncTscWorker, &CurTsc, NULL);
5200	if (RT_SUCCESS(rc))
5201	{
5202	if (CurTsc <= PrevTsc)
5203	{
5204	fAsync = true;
5205	offMin = offMax = PrevTsc - CurTsc;
5206	Log(("supdrvDetermineAsyncTsc: iCpu=%d cLoops=%d CurTsc=%llx PrevTsc=%llx\n",
5207	iCpu, cLoops, CurTsc, PrevTsc));
5208	break;
5209	}
5210
5211	/* Gather statistics (except the first time). */
5212	if (iCpu != 0 \|\| cLoops != 7)
5213	{
5214	uint64_t off = CurTsc - PrevTsc;
5215	if (off < offMin)
5216	offMin = off;
5217	if (off > offMax)
5218	offMax = off;
5219	Log2(("%d/%d: off=%llx\n", cLoops, iCpu, off));
5220	}
5221
5222	/* Next */
5223	PrevTsc = CurTsc;
5224	}
5225	else if (rc == VERR_NOT_SUPPORTED)
5226	break;
5227	else
5228	AssertMsg(rc == VERR_CPU_NOT_FOUND \|\| rc == VERR_CPU_OFFLINE, ("%d\n", rc));
5229	}
5230
5231	/* broke out of the loop. */
5232	if (iCpu < iEndCpu)
5233	break;
5234	}
5235
5236	poffMin = offMin; / Almost RTMpOnSpecific profiling. */
5237	Log(("supdrvDetermineAsyncTsc: returns %d; iEndCpu=%d rc=%d offMin=%llx offMax=%llx\n",
5238	fAsync, iEndCpu, rc, offMin, offMax));
5239	#if !defined(RT_OS_SOLARIS) && !defined(RT_OS_OS2) && !defined(RT_OS_WINDOWS)
5240	OSDBGPRINT(("vboxdrv: fAsync=%d offMin=%#lx offMax=%#lx\n", fAsync, (long)offMin, (long)offMax));
5241	#endif
5242	return fAsync;
5243	}
5244
5245
5246	/**
5247	* Determine the GIP TSC mode.
5248	*
5249	* @returns The most suitable TSC mode.
5250	* @param pDevExt Pointer to the device instance data.
5251	*/
5252	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt)
5253	{
5254	/*
5255	* On SMP we're faced with two problems:
5256	* (1) There might be a skew between the CPU, so that cpu0
5257	* returns a TSC that is slightly different from cpu1.
5258	* (2) Power management (and other things) may cause the TSC
5259	* to run at a non-constant speed, and cause the speed
5260	* to be different on the cpus. This will result in (1).
5261	*
5262	* So, on SMP systems we'll have to select the ASYNC update method
5263	* if there are symptoms of these problems.
5264	*/
5265	if (RTMpGetCount() > 1)
5266	{
5267	uint32_t uEAX, uEBX, uECX, uEDX;
5268	uint64_t u64DiffCoresIgnored;
5269
5270	/* Permit the user and/or the OS specific bits to force async mode. */
5271	if (supdrvOSGetForcedAsyncTscMode(pDevExt))
5272	return SUPGIPMODE_ASYNC_TSC;
5273
5274	/* Try check for current differences between the cpus. */
5275	if (supdrvDetermineAsyncTsc(&u64DiffCoresIgnored))
5276	return SUPGIPMODE_ASYNC_TSC;
5277
5278	/*
5279	* If the CPU supports power management and is an AMD one we
5280	* won't trust it unless it has the TscInvariant bit is set.
5281	*/
5282	/* Check for "AuthenticAMD" */
5283	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
5284	if ( uEAX >= 1
5285	&& uEBX == X86_CPUID_VENDOR_AMD_EBX
5286	&& uECX == X86_CPUID_VENDOR_AMD_ECX
5287	&& uEDX == X86_CPUID_VENDOR_AMD_EDX)
5288	{
5289	/* Check for APM support and that TscInvariant is cleared. */
5290	ASMCpuId(0x80000000, &uEAX, &uEBX, &uECX, &uEDX);
5291	if (uEAX >= 0x80000007)
5292	{
5293	ASMCpuId(0x80000007, &uEAX, &uEBX, &uECX, &uEDX);
5294	if ( !(uEDX & RT_BIT(8))/* TscInvariant */
5295	&& (uEDX & 0x3e)) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
5296	return SUPGIPMODE_ASYNC_TSC;
5297	}
5298	}
5299	}
5300	return SUPGIPMODE_SYNC_TSC;
5301	}
5302
5303
5304
5305	/**
5306	* Initializes the GIP data.
5307	*
5308	* @param pDevExt Pointer to the device instance data.
5309	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5310	* @param HCPhys The physical address of the GIP.
5311	* @param u64NanoTS The current nanosecond timestamp.
5312	* @param uUpdateHz The update frequency.
5313	* @param cCpus The CPU count.
5314	*/
5315	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
5316	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned cCpus)
5317	{
5318	size_t const cbGip = RT_ALIGN_Z(RT_OFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), PAGE_SIZE);
5319	unsigned i;
5320	#ifdef DEBUG_DARWIN_GIP
5321	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5322	#else
5323	LogFlow(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5324	#endif
5325
5326	/*
5327	* Initialize the structure.
5328	*/
5329	memset(pGip, 0, cbGip);
5330	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
5331	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
5332	pGip->u32Mode = supdrvGipDeterminTscMode(pDevExt);
5333	pGip->cCpus = (uint16_t)cCpus;
5334	pGip->cPages = (uint16_t)(cbGip / PAGE_SIZE);
5335	pGip->u32UpdateHz = uUpdateHz;
5336	pGip->u32UpdateIntervalNS = 1000000000 / uUpdateHz;
5337	pGip->u64NanoTSLastUpdateHz = u64NanoTS;
5338	RTCpuSetEmpty(&pGip->OnlineCpuSet);
5339	RTCpuSetEmpty(&pGip->PresentCpuSet);
5340	RTMpGetSet(&pGip->PossibleCpuSet);
5341	pGip->cOnlineCpus = RTMpGetOnlineCount();
5342	pGip->cPresentCpus = RTMpGetPresentCount();
5343	pGip->cPossibleCpus = RTMpGetCount();
5344	pGip->idCpuMax = RTMpGetMaxCpuId();
5345	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromApicId); i++)
5346	pGip->aiCpuFromApicId[i] = 0;
5347	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx); i++)
5348	pGip->aiCpuFromCpuSetIdx[i] = UINT16_MAX;
5349
5350	for (i = 0; i < cCpus; i++)
5351	{
5352	pGip->aCPUs[i].u32TransactionId = 2;
5353	pGip->aCPUs[i].u64NanoTS = u64NanoTS;
5354	pGip->aCPUs[i].u64TSC = ASMReadTSC();
5355
5356	pGip->aCPUs[i].enmState = SUPGIPCPUSTATE_INVALID;
5357	pGip->aCPUs[i].idCpu = NIL_RTCPUID;
5358	pGip->aCPUs[i].iCpuSet = -1;
5359	pGip->aCPUs[i].idApic = UINT8_MAX;
5360
5361	/*
5362	* We don't know the following values until we've executed updates.
5363	* So, we'll just pretend it's a 4 GHz CPU and adjust the history it on
5364	* the 2nd timer callout.
5365	*/
5366	pGip->aCPUs[i].u64CpuHz = _4G + 1; /* tstGIP-2 depends on this. */
5367	pGip->aCPUs[i].u32UpdateIntervalTSC
5368	= pGip->aCPUs[i].au32TSCHistory[0]
5369	= pGip->aCPUs[i].au32TSCHistory[1]
5370	= pGip->aCPUs[i].au32TSCHistory[2]
5371	= pGip->aCPUs[i].au32TSCHistory[3]
5372	= pGip->aCPUs[i].au32TSCHistory[4]
5373	= pGip->aCPUs[i].au32TSCHistory[5]
5374	= pGip->aCPUs[i].au32TSCHistory[6]
5375	= pGip->aCPUs[i].au32TSCHistory[7]
5376	= /pGip->aCPUs[i].u64CpuHz/ (uint32_t)(_4G / uUpdateHz);
5377	}
5378
5379	/*
5380	* Link it to the device extension.
5381	*/
5382	pDevExt->pGip = pGip;
5383	pDevExt->HCPhysGip = HCPhys;
5384	pDevExt->cGipUsers = 0;
5385	}
5386
5387
5388	/**
5389	* On CPU initialization callback for RTMpOnAll.
5390	*
5391	* @param idCpu The CPU ID.
5392	* @param pvUser1 The device extension.
5393	* @param pvUser2 The GIP.
5394	*/
5395	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2)
5396	{
5397	/* This is good enough, even though it will update some of the globals a
5398	bit to much. */
5399	supdrvGipMpEventOnline((PSUPGLOBALINFOPAGE)pvUser2, idCpu);
5400	}
5401
5402
5403	/**
5404	* Invalidates the GIP data upon termination.
5405	*
5406	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5407	*/
5408	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
5409	{
5410	unsigned i;
5411	pGip->u32Magic = 0;
5412	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
5413	{
5414	pGip->aCPUs[i].u64NanoTS = 0;
5415	pGip->aCPUs[i].u64TSC = 0;
5416	pGip->aCPUs[i].iTSCHistoryHead = 0;
5417	}
5418	}
5419
5420
5421	/**
5422	* Worker routine for supdrvGipUpdate and supdrvGipUpdatePerCpu that
5423	* updates all the per cpu data except the transaction id.
5424	*
5425	* @param pGip The GIP.
5426	* @param pGipCpu Pointer to the per cpu data.
5427	* @param u64NanoTS The current time stamp.
5428	* @param u64TSC The current TSC.
5429	* @param iTick The current timer tick.
5430	*/
5431	static void supdrvGipDoUpdateCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick)
5432	{
5433	uint64_t u64TSCDelta;
5434	uint32_t u32UpdateIntervalTSC;
5435	uint32_t u32UpdateIntervalTSCSlack;
5436	unsigned iTSCHistoryHead;
5437	uint64_t u64CpuHz;
5438	uint32_t u32TransactionId;
5439
5440	/* Delta between this and the previous update. */
5441	ASMAtomicUoWriteU32(&pGipCpu->u32PrevUpdateIntervalNS, (uint32_t)(u64NanoTS - pGipCpu->u64NanoTS));
5442
5443	/*
5444	* Update the NanoTS.
5445	*/
5446	ASMAtomicWriteU64(&pGipCpu->u64NanoTS, u64NanoTS);
5447
5448	/*
5449	* Calc TSC delta.
5450	*/
5451	/** @todo validate the NanoTS delta, don't trust the OS to call us when it should... */
5452	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
5453	ASMAtomicWriteU64(&pGipCpu->u64TSC, u64TSC);
5454
5455	if (u64TSCDelta >> 32)
5456	{
5457	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
5458	pGipCpu->cErrors++;
5459	}
5460
5461	/*
5462	* On the 2nd and 3rd callout, reset the history with the current TSC
5463	* interval since the values entered by supdrvGipInit are totally off.
5464	* The interval on the 1st callout completely unreliable, the 2nd is a bit
5465	* better, while the 3rd should be most reliable.
5466	*/
5467	u32TransactionId = pGipCpu->u32TransactionId;
5468	if (RT_UNLIKELY( ( u32TransactionId == 5
5469	\|\| u32TransactionId == 7)
5470	&& ( iTick == 2
5471	\|\| iTick == 3) ))
5472	{
5473	unsigned i;
5474	for (i = 0; i < RT_ELEMENTS(pGipCpu->au32TSCHistory); i++)
5475	ASMAtomicUoWriteU32(&pGipCpu->au32TSCHistory[i], (uint32_t)u64TSCDelta);
5476	}
5477
5478	/*
5479	* TSC History.
5480	*/
5481	Assert(RT_ELEMENTS(pGipCpu->au32TSCHistory) == 8);
5482	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
5483	ASMAtomicWriteU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
5484	ASMAtomicWriteU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
5485
5486	/*
5487	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
5488	*/
5489	if (pGip->u32UpdateHz >= 1000)
5490	{
5491	uint32_t u32;
5492	u32 = pGipCpu->au32TSCHistory[0];
5493	u32 += pGipCpu->au32TSCHistory[1];
5494	u32 += pGipCpu->au32TSCHistory[2];
5495	u32 += pGipCpu->au32TSCHistory[3];
5496	u32 >>= 2;
5497	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
5498	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
5499	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
5500	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
5501	u32UpdateIntervalTSC >>= 2;
5502	u32UpdateIntervalTSC += u32;
5503	u32UpdateIntervalTSC >>= 1;
5504
5505	/* Value chosen for a 2GHz Athlon64 running linux 2.6.10/11, . */
5506	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
5507	}
5508	else if (pGip->u32UpdateHz >= 90)
5509	{
5510	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5511	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
5512	u32UpdateIntervalTSC >>= 1;
5513
5514	/* value chosen on a 2GHz thinkpad running windows */
5515	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
5516	}
5517	else
5518	{
5519	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5520
5521	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
5522	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
5523	}
5524	ASMAtomicWriteU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
5525
5526	/*
5527	* CpuHz.
5528	*/
5529	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, pGip->u32UpdateHz);
5530	ASMAtomicWriteU64(&pGipCpu->u64CpuHz, u64CpuHz);
5531	}
5532
5533
5534	/**
5535	* Updates the GIP.
5536	*
5537	* @param pGip Pointer to the GIP.
5538	* @param u64NanoTS The current nanosecond timesamp.
5539	* @param u64TSC The current TSC timesamp.
5540	* @param idCpu The CPU ID.
5541	* @param iTick The current timer tick.
5542	*/
5543	static void supdrvGipUpdate(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick)
5544	{
5545	/*
5546	* Determine the relevant CPU data.
5547	*/
5548	PSUPGIPCPU pGipCpu;
5549	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
5550	pGipCpu = &pGip->aCPUs[0];
5551	else
5552	{
5553	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
5554	if (RT_UNLIKELY(iCpu >= pGip->cCpus))
5555	return;
5556	pGipCpu = &pGip->aCPUs[iCpu];
5557	if (RT_UNLIKELY(pGipCpu->idCpu != idCpu))
5558	return;
5559	}
5560
5561	/*
5562	* Start update transaction.
5563	*/
5564	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5565	{
5566	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
5567	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5568	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5569	pGipCpu->cErrors++;
5570	return;
5571	}
5572
5573	/*
5574	* Recalc the update frequency every 0x800th time.
5575	*/
5576	if (!(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
5577	{
5578	if (pGip->u64NanoTSLastUpdateHz)
5579	{
5580	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
5581	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
5582	uint32_t u32UpdateHz = (uint32_t)((UINT64_C(1000000000) * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
5583	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
5584	{
5585	ASMAtomicWriteU32(&pGip->u32UpdateHz, u32UpdateHz);
5586	ASMAtomicWriteU32(&pGip->u32UpdateIntervalNS, 1000000000 / u32UpdateHz);
5587	}
5588	#endif
5589	}
5590	ASMAtomicWriteU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS);
5591	}
5592
5593	/*
5594	* Update the data.
5595	*/
5596	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS, u64TSC, iTick);
5597
5598	/*
5599	* Complete transaction.
5600	*/
5601	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5602	}
5603
5604
5605	/**
5606	* Updates the per cpu GIP data for the calling cpu.
5607	*
5608	* @param pGip Pointer to the GIP.
5609	* @param u64NanoTS The current nanosecond timesamp.
5610	* @param u64TSC The current TSC timesamp.
5611	* @param idCpu The CPU ID.
5612	* @param idApic The APIC id for the CPU index.
5613	* @param iTick The current timer tick.
5614	*/
5615	static void supdrvGipUpdatePerCpu(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC,
5616	RTCPUID idCpu, uint8_t idApic, uint64_t iTick)
5617	{
5618	unsigned iCpu = pGip->aiCpuFromApicId[idApic];
5619
5620	if (RT_LIKELY(iCpu < pGip->cCpus))
5621	{
5622	PSUPGIPCPU pGipCpu = &pGip->aCPUs[iCpu];
5623	if (pGipCpu->idCpu == idCpu)
5624	{
5625
5626	/*
5627	* Start update transaction.
5628	*/
5629	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5630	{
5631	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5632	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5633	pGipCpu->cErrors++;
5634	return;
5635	}
5636
5637	/*
5638	* Update the data.
5639	*/
5640	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS, u64TSC, iTick);
5641
5642	/*
5643	* Complete transaction.
5644	*/
5645	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5646	}
5647	}
5648	}

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

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDrv.c@ 36540

Download in other formats: