SUPDrv.c@ 40282

Last change on this file since 40282 was 40227, checked in by vboxsync, 13 years ago
VMM/VMMR0 SupDrv Solaris/MpNotification: Fix guru meditation/panics when host CPUs are dynamically offline'd/online'd. More flexibility if offline notifications don't fire on the desired CPU.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 214.7 KB

Line
1	/* $Revision: 40227 $ */
2	/** @file
3	* VBoxDrv - The VirtualBox Support Driver - Common code.
4	*/
5
6	/*
7	* Copyright (C) 2006-2011 Oracle Corporation
8	*
9	* This file is part of VirtualBox Open Source Edition (OSE), as
10	* available from http://www.virtualbox.org. This file is free software;
11	* you can redistribute it and/or modify it under the terms of the GNU
12	* General Public License (GPL) as published by the Free Software
13	* Foundation, in version 2 as it comes in the "COPYING" file of the
14	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16	*
17	* The contents of this file may alternatively be used under the terms
18	* of the Common Development and Distribution License Version 1.0
19	* (CDDL) only, as it comes in the "COPYING.CDDL" file of the
20	* VirtualBox OSE distribution, in which case the provisions of the
21	* CDDL are applicable instead of those of the GPL.
22	*
23	* You may elect to license modified versions of this file under the
24	* terms and conditions of either the GPL or the CDDL or both.
25	*/
26
27	/*******************************************************************************
28	* Header Files *
29	*******************************************************************************/
30	#define LOG_GROUP LOG_GROUP_SUP_DRV
31	#define SUPDRV_AGNOSTIC
32	#include "SUPDrvInternal.h"
33	#ifndef PAGE_SHIFT
34	# include <iprt/param.h>
35	#endif
36	#include <iprt/asm.h>
37	#include <iprt/asm-amd64-x86.h>
38	#include <iprt/asm-math.h>
39	#include <iprt/cpuset.h>
40	#include <iprt/handletable.h>
41	#include <iprt/mem.h>
42	#include <iprt/mp.h>
43	#include <iprt/power.h>
44	#include <iprt/process.h>
45	#include <iprt/semaphore.h>
46	#include <iprt/spinlock.h>
47	#include <iprt/thread.h>
48	#include <iprt/uuid.h>
49	#include <iprt/net.h>
50	#include <iprt/crc.h>
51	#include <iprt/string.h>
52	#include <iprt/timer.h>
53	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
54	# include <iprt/rand.h>
55	# include <iprt/path.h>
56	#endif
57	#include <iprt/x86.h>
58
59	#include <VBox/param.h>
60	#include <VBox/log.h>
61	#include <VBox/err.h>
62	#include <VBox/vmm/hwacc_svm.h>
63	#include <VBox/vmm/hwacc_vmx.h>
64
65	#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	static void supdrvGipInitCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pCpu, uint64_t u64NanoTS);
141
142
143	/*******************************************************************************
144	* Global Variables *
145	*******************************************************************************/
146	DECLEXPORT(PSUPGLOBALINFOPAGE) g_pSUPGlobalInfoPage = NULL;
147
148	/**
149	* Array of the R0 SUP API.
150	*/
151	static SUPFUNC g_aFunctions[] =
152	{
153	/* name function */
154	/* Entries with absolute addresses determined at runtime, fixup
155	code makes ugly ASSUMPTIONS about the order here: */
156	{ "SUPR0AbsIs64bit", (void *)0 },
157	{ "SUPR0Abs64bitKernelCS", (void *)0 },
158	{ "SUPR0Abs64bitKernelSS", (void *)0 },
159	{ "SUPR0Abs64bitKernelDS", (void *)0 },
160	{ "SUPR0AbsKernelCS", (void *)0 },
161	{ "SUPR0AbsKernelSS", (void *)0 },
162	{ "SUPR0AbsKernelDS", (void *)0 },
163	{ "SUPR0AbsKernelES", (void *)0 },
164	{ "SUPR0AbsKernelFS", (void *)0 },
165	{ "SUPR0AbsKernelGS", (void *)0 },
166	/* Normal function pointers: */
167	{ "SUPR0ComponentRegisterFactory", (void *)SUPR0ComponentRegisterFactory },
168	{ "SUPR0ComponentDeregisterFactory", (void *)SUPR0ComponentDeregisterFactory },
169	{ "SUPR0ComponentQueryFactory", (void *)SUPR0ComponentQueryFactory },
170	{ "SUPR0ObjRegister", (void *)SUPR0ObjRegister },
171	{ "SUPR0ObjAddRef", (void *)SUPR0ObjAddRef },
172	{ "SUPR0ObjAddRefEx", (void *)SUPR0ObjAddRefEx },
173	{ "SUPR0ObjRelease", (void *)SUPR0ObjRelease },
174	{ "SUPR0ObjVerifyAccess", (void *)SUPR0ObjVerifyAccess },
175	{ "SUPR0LockMem", (void *)SUPR0LockMem },
176	{ "SUPR0UnlockMem", (void *)SUPR0UnlockMem },
177	{ "SUPR0ContAlloc", (void *)SUPR0ContAlloc },
178	{ "SUPR0ContFree", (void *)SUPR0ContFree },
179	{ "SUPR0LowAlloc", (void *)SUPR0LowAlloc },
180	{ "SUPR0LowFree", (void *)SUPR0LowFree },
181	{ "SUPR0MemAlloc", (void *)SUPR0MemAlloc },
182	{ "SUPR0MemGetPhys", (void *)SUPR0MemGetPhys },
183	{ "SUPR0MemFree", (void *)SUPR0MemFree },
184	{ "SUPR0PageAllocEx", (void *)SUPR0PageAllocEx },
185	{ "SUPR0PageFree", (void *)SUPR0PageFree },
186	{ "SUPR0Printf", (void )SUPR0Printf }, /* @todo needs wrapping? */
187	{ "SUPSemEventCreate", (void *)SUPSemEventCreate },
188	{ "SUPSemEventClose", (void *)SUPSemEventClose },
189	{ "SUPSemEventSignal", (void *)SUPSemEventSignal },
190	{ "SUPSemEventWait", (void *)SUPSemEventWait },
191	{ "SUPSemEventWaitNoResume", (void *)SUPSemEventWaitNoResume },
192	{ "SUPSemEventWaitNsAbsIntr", (void *)SUPSemEventWaitNsAbsIntr },
193	{ "SUPSemEventWaitNsRelIntr", (void *)SUPSemEventWaitNsRelIntr },
194	{ "SUPSemEventGetResolution", (void *)SUPSemEventGetResolution },
195	{ "SUPSemEventMultiCreate", (void *)SUPSemEventMultiCreate },
196	{ "SUPSemEventMultiClose", (void *)SUPSemEventMultiClose },
197	{ "SUPSemEventMultiSignal", (void *)SUPSemEventMultiSignal },
198	{ "SUPSemEventMultiReset", (void *)SUPSemEventMultiReset },
199	{ "SUPSemEventMultiWait", (void *)SUPSemEventMultiWait },
200	{ "SUPSemEventMultiWaitNoResume", (void *)SUPSemEventMultiWaitNoResume },
201	{ "SUPSemEventMultiWaitNsAbsIntr", (void *)SUPSemEventMultiWaitNsAbsIntr },
202	{ "SUPSemEventMultiWaitNsRelIntr", (void *)SUPSemEventMultiWaitNsRelIntr },
203	{ "SUPSemEventMultiGetResolution", (void *)SUPSemEventMultiGetResolution },
204	{ "SUPR0GetPagingMode", (void *)SUPR0GetPagingMode },
205	{ "SUPR0EnableVTx", (void *)SUPR0EnableVTx },
206	{ "SUPGetGIP", (void *)SUPGetGIP },
207	{ "g_pSUPGlobalInfoPage", (void *)&g_pSUPGlobalInfoPage },
208	{ "RTMemAllocTag", (void *)RTMemAllocTag },
209	{ "RTMemAllocZTag", (void *)RTMemAllocZTag },
210	{ "RTMemAllocVarTag", (void *)RTMemAllocVarTag },
211	{ "RTMemAllocZVarTag", (void *)RTMemAllocZVarTag },
212	{ "RTMemFree", (void *)RTMemFree },
213	{ "RTMemDupTag", (void *)RTMemDupTag },
214	{ "RTMemDupExTag", (void *)RTMemDupExTag },
215	{ "RTMemReallocTag", (void *)RTMemReallocTag },
216	{ "RTR0MemObjAllocLowTag", (void *)RTR0MemObjAllocLowTag },
217	{ "RTR0MemObjAllocPageTag", (void *)RTR0MemObjAllocPageTag },
218	{ "RTR0MemObjAllocPhysTag", (void *)RTR0MemObjAllocPhysTag },
219	{ "RTR0MemObjAllocPhysExTag", (void *)RTR0MemObjAllocPhysExTag },
220	{ "RTR0MemObjAllocPhysNCTag", (void *)RTR0MemObjAllocPhysNCTag },
221	{ "RTR0MemObjAllocContTag", (void *)RTR0MemObjAllocContTag },
222	{ "RTR0MemObjEnterPhysTag", (void *)RTR0MemObjEnterPhysTag },
223	{ "RTR0MemObjLockUserTag", (void *)RTR0MemObjLockUserTag },
224	{ "RTR0MemObjMapKernelTag", (void *)RTR0MemObjMapKernelTag },
225	{ "RTR0MemObjMapKernelExTag", (void *)RTR0MemObjMapKernelExTag },
226	{ "RTR0MemObjMapUserTag", (void *)RTR0MemObjMapUserTag },
227	{ "RTR0MemObjProtect", (void *)RTR0MemObjProtect },
228	{ "RTR0MemObjAddress", (void *)RTR0MemObjAddress },
229	{ "RTR0MemObjAddressR3", (void *)RTR0MemObjAddressR3 },
230	{ "RTR0MemObjSize", (void *)RTR0MemObjSize },
231	{ "RTR0MemObjIsMapping", (void *)RTR0MemObjIsMapping },
232	{ "RTR0MemObjGetPagePhysAddr", (void *)RTR0MemObjGetPagePhysAddr },
233	{ "RTR0MemObjFree", (void *)RTR0MemObjFree },
234	{ "RTR0MemUserCopyFrom", (void *)RTR0MemUserCopyFrom },
235	{ "RTR0MemUserCopyTo", (void *)RTR0MemUserCopyTo },
236	{ "RTR0MemUserIsValidAddr", (void *)RTR0MemUserIsValidAddr },
237	{ "RTR0MemKernelIsValidAddr", (void *)RTR0MemKernelIsValidAddr },
238	{ "RTR0MemAreKrnlAndUsrDifferent", (void *)RTR0MemAreKrnlAndUsrDifferent },
239	{ "RTSemMutexCreate", (void *)RTSemMutexCreate },
240	{ "RTSemMutexRequest", (void *)RTSemMutexRequest },
241	{ "RTSemMutexRequestDebug", (void *)RTSemMutexRequestDebug },
242	{ "RTSemMutexRequestNoResume", (void *)RTSemMutexRequestNoResume },
243	{ "RTSemMutexRequestNoResumeDebug", (void *)RTSemMutexRequestNoResumeDebug },
244	{ "RTSemMutexRelease", (void *)RTSemMutexRelease },
245	{ "RTSemMutexDestroy", (void *)RTSemMutexDestroy },
246	{ "RTProcSelf", (void *)RTProcSelf },
247	{ "RTR0ProcHandleSelf", (void *)RTR0ProcHandleSelf },
248	{ "RTSemFastMutexCreate", (void *)RTSemFastMutexCreate },
249	{ "RTSemFastMutexDestroy", (void *)RTSemFastMutexDestroy },
250	{ "RTSemFastMutexRequest", (void *)RTSemFastMutexRequest },
251	{ "RTSemFastMutexRelease", (void *)RTSemFastMutexRelease },
252	{ "RTSemEventCreate", (void *)RTSemEventCreate },
253	{ "RTSemEventSignal", (void *)RTSemEventSignal },
254	{ "RTSemEventWait", (void *)RTSemEventWait },
255	{ "RTSemEventWaitNoResume", (void *)RTSemEventWaitNoResume },
256	{ "RTSemEventWaitEx", (void *)RTSemEventWaitEx },
257	{ "RTSemEventWaitExDebug", (void *)RTSemEventWaitExDebug },
258	{ "RTSemEventGetResolution", (void *)RTSemEventGetResolution },
259	{ "RTSemEventDestroy", (void *)RTSemEventDestroy },
260	{ "RTSemEventMultiCreate", (void *)RTSemEventMultiCreate },
261	{ "RTSemEventMultiSignal", (void *)RTSemEventMultiSignal },
262	{ "RTSemEventMultiReset", (void *)RTSemEventMultiReset },
263	{ "RTSemEventMultiWait", (void *)RTSemEventMultiWait },
264	{ "RTSemEventMultiWaitNoResume", (void *)RTSemEventMultiWaitNoResume },
265	{ "RTSemEventMultiWaitEx", (void *)RTSemEventMultiWaitEx },
266	{ "RTSemEventMultiWaitExDebug", (void *)RTSemEventMultiWaitExDebug },
267	{ "RTSemEventMultiGetResolution", (void *)RTSemEventMultiGetResolution },
268	{ "RTSemEventMultiDestroy", (void *)RTSemEventMultiDestroy },
269	{ "RTSpinlockCreate", (void *)RTSpinlockCreate },
270	{ "RTSpinlockDestroy", (void *)RTSpinlockDestroy },
271	{ "RTSpinlockAcquire", (void *)RTSpinlockAcquire },
272	{ "RTSpinlockRelease", (void *)RTSpinlockRelease },
273	{ "RTSpinlockAcquireNoInts", (void *)RTSpinlockAcquireNoInts },
274	{ "RTSpinlockReleaseNoInts", (void *)RTSpinlockReleaseNoInts },
275	{ "RTTimeNanoTS", (void *)RTTimeNanoTS },
276	{ "RTTimeMilliTS", (void *)RTTimeMilliTS },
277	{ "RTTimeSystemNanoTS", (void *)RTTimeSystemNanoTS },
278	{ "RTTimeSystemMilliTS", (void *)RTTimeSystemMilliTS },
279	{ "RTThreadNativeSelf", (void *)RTThreadNativeSelf },
280	{ "RTThreadSleep", (void *)RTThreadSleep },
281	{ "RTThreadYield", (void *)RTThreadYield },
282	{ "RTThreadSelf", (void *)RTThreadSelf },
283	{ "RTThreadCreate", (void *)RTThreadCreate },
284	{ "RTThreadGetNative", (void *)RTThreadGetNative },
285	{ "RTThreadWait", (void *)RTThreadWait },
286	{ "RTThreadWaitNoResume", (void *)RTThreadWaitNoResume },
287	{ "RTThreadGetName", (void *)RTThreadGetName },
288	{ "RTThreadSelfName", (void *)RTThreadSelfName },
289	{ "RTThreadGetType", (void *)RTThreadGetType },
290	{ "RTThreadUserSignal", (void *)RTThreadUserSignal },
291	{ "RTThreadUserReset", (void *)RTThreadUserReset },
292	{ "RTThreadUserWait", (void *)RTThreadUserWait },
293	{ "RTThreadUserWaitNoResume", (void *)RTThreadUserWaitNoResume },
294	{ "RTThreadPreemptIsEnabled", (void *)RTThreadPreemptIsEnabled },
295	{ "RTThreadPreemptIsPending", (void *)RTThreadPreemptIsPending },
296	{ "RTThreadPreemptIsPendingTrusty", (void *)RTThreadPreemptIsPendingTrusty },
297	{ "RTThreadPreemptIsPossible", (void *)RTThreadPreemptIsPossible },
298	{ "RTThreadPreemptDisable", (void *)RTThreadPreemptDisable },
299	{ "RTThreadPreemptRestore", (void *)RTThreadPreemptRestore },
300	{ "RTThreadIsInInterrupt", (void *)RTThreadIsInInterrupt },
301	{ "RTTimerCreate", (void *)RTTimerCreate },
302	{ "RTTimerCreateEx", (void *)RTTimerCreateEx },
303	{ "RTTimerDestroy", (void *)RTTimerDestroy },
304	{ "RTTimerStart", (void *)RTTimerStart },
305	{ "RTTimerStop", (void *)RTTimerStop },
306	{ "RTTimerChangeInterval", (void *)RTTimerChangeInterval },
307	{ "RTTimerGetSystemGranularity", (void *)RTTimerGetSystemGranularity },
308	{ "RTTimerRequestSystemGranularity", (void *)RTTimerRequestSystemGranularity },
309	{ "RTTimerReleaseSystemGranularity", (void *)RTTimerReleaseSystemGranularity },
310	{ "RTTimerCanDoHighResolution", (void *)RTTimerCanDoHighResolution },
311
312	{ "RTLogDefaultInstance", (void *)RTLogDefaultInstance },
313	{ "RTMpCpuId", (void *)RTMpCpuId },
314	{ "RTMpCpuIdFromSetIndex", (void *)RTMpCpuIdFromSetIndex },
315	{ "RTMpCpuIdToSetIndex", (void *)RTMpCpuIdToSetIndex },
316	{ "RTMpGetArraySize", (void *)RTMpGetArraySize },
317	{ "RTMpIsCpuPossible", (void *)RTMpIsCpuPossible },
318	{ "RTMpGetCount", (void *)RTMpGetCount },
319	{ "RTMpGetMaxCpuId", (void *)RTMpGetMaxCpuId },
320	{ "RTMpGetOnlineCount", (void *)RTMpGetOnlineCount },
321	{ "RTMpGetOnlineSet", (void *)RTMpGetOnlineSet },
322	{ "RTMpGetSet", (void *)RTMpGetSet },
323	{ "RTMpIsCpuOnline", (void *)RTMpIsCpuOnline },
324	{ "RTMpIsCpuWorkPending", (void *)RTMpIsCpuWorkPending },
325	{ "RTMpNotificationRegister", (void *)RTMpNotificationRegister },
326	{ "RTMpNotificationDeregister", (void *)RTMpNotificationDeregister },
327	{ "RTMpOnAll", (void *)RTMpOnAll },
328	{ "RTMpOnOthers", (void *)RTMpOnOthers },
329	{ "RTMpOnSpecific", (void *)RTMpOnSpecific },
330	{ "RTMpPokeCpu", (void *)RTMpPokeCpu },
331	{ "RTPowerNotificationRegister", (void *)RTPowerNotificationRegister },
332	{ "RTPowerNotificationDeregister", (void *)RTPowerNotificationDeregister },
333	{ "RTLogRelDefaultInstance", (void *)RTLogRelDefaultInstance },
334	{ "RTLogSetDefaultInstanceThread", (void *)RTLogSetDefaultInstanceThread },
335	{ "RTLogLoggerExV", (void *)RTLogLoggerExV },
336	{ "RTLogPrintfV", (void *)RTLogPrintfV },
337	{ "RTR0AssertPanicSystem", (void *)RTR0AssertPanicSystem },
338	{ "RTAssertMsg1", (void *)RTAssertMsg1 },
339	{ "RTAssertMsg2V", (void *)RTAssertMsg2V },
340	{ "RTAssertMsg2AddV", (void *)RTAssertMsg2AddV },
341	{ "RTAssertSetQuiet", (void *)RTAssertSetQuiet },
342	{ "RTAssertMayPanic", (void *)RTAssertMayPanic },
343	{ "RTAssertSetMayPanic", (void *)RTAssertSetMayPanic },
344	{ "RTAssertAreQuiet", (void *)RTAssertAreQuiet },
345	{ "RTStrFormat", (void *)RTStrFormat },
346	{ "RTStrFormatNumber", (void *)RTStrFormatNumber },
347	{ "RTStrFormatTypeDeregister", (void *)RTStrFormatTypeDeregister },
348	{ "RTStrFormatTypeRegister", (void *)RTStrFormatTypeRegister },
349	{ "RTStrFormatTypeSetUser", (void *)RTStrFormatTypeSetUser },
350	{ "RTStrFormatV", (void *)RTStrFormatV },
351	{ "RTStrPrintf", (void *)RTStrPrintf },
352	{ "RTStrPrintfEx", (void *)RTStrPrintfEx },
353	{ "RTStrPrintfExV", (void *)RTStrPrintfExV },
354	{ "RTStrPrintfV", (void *)RTStrPrintfV },
355	{ "RTHandleTableAllocWithCtx", (void *)RTHandleTableAllocWithCtx },
356	{ "RTHandleTableCreate", (void *)RTHandleTableCreate },
357	{ "RTHandleTableCreateEx", (void *)RTHandleTableCreateEx },
358	{ "RTHandleTableDestroy", (void *)RTHandleTableDestroy },
359	{ "RTHandleTableFreeWithCtx", (void *)RTHandleTableFreeWithCtx },
360	{ "RTHandleTableLookupWithCtx", (void *)RTHandleTableLookupWithCtx },
361	{ "RTNetIPv4AddDataChecksum", (void *)RTNetIPv4AddDataChecksum },
362	{ "RTNetIPv4AddTCPChecksum", (void *)RTNetIPv4AddTCPChecksum },
363	{ "RTNetIPv4AddUDPChecksum", (void *)RTNetIPv4AddUDPChecksum },
364	{ "RTNetIPv4FinalizeChecksum", (void *)RTNetIPv4FinalizeChecksum },
365	{ "RTNetIPv4HdrChecksum", (void *)RTNetIPv4HdrChecksum },
366	{ "RTNetIPv4IsDHCPValid", (void *)RTNetIPv4IsDHCPValid },
367	{ "RTNetIPv4IsHdrValid", (void *)RTNetIPv4IsHdrValid },
368	{ "RTNetIPv4IsTCPSizeValid", (void *)RTNetIPv4IsTCPSizeValid },
369	{ "RTNetIPv4IsTCPValid", (void *)RTNetIPv4IsTCPValid },
370	{ "RTNetIPv4IsUDPSizeValid", (void *)RTNetIPv4IsUDPSizeValid },
371	{ "RTNetIPv4IsUDPValid", (void *)RTNetIPv4IsUDPValid },
372	{ "RTNetIPv4PseudoChecksum", (void *)RTNetIPv4PseudoChecksum },
373	{ "RTNetIPv4PseudoChecksumBits", (void *)RTNetIPv4PseudoChecksumBits },
374	{ "RTNetIPv4TCPChecksum", (void *)RTNetIPv4TCPChecksum },
375	{ "RTNetIPv4UDPChecksum", (void *)RTNetIPv4UDPChecksum },
376	{ "RTNetIPv6PseudoChecksum", (void *)RTNetIPv6PseudoChecksum },
377	{ "RTNetIPv6PseudoChecksumBits", (void *)RTNetIPv6PseudoChecksumBits },
378	{ "RTNetIPv6PseudoChecksumEx", (void *)RTNetIPv6PseudoChecksumEx },
379	{ "RTNetTCPChecksum", (void *)RTNetTCPChecksum },
380	{ "RTNetUDPChecksum", (void *)RTNetUDPChecksum },
381	{ "RTCrc32", (void *)RTCrc32 },
382	{ "RTCrc32Finish", (void *)RTCrc32Finish },
383	{ "RTCrc32Process", (void *)RTCrc32Process },
384	{ "RTCrc32Start", (void *)RTCrc32Start },
385	};
386
387	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
388	/**
389	* Drag in the rest of IRPT since we share it with the
390	* rest of the kernel modules on darwin.
391	*/
392	PFNRT g_apfnVBoxDrvIPRTDeps[] =
393	{
394	/* VBoxNetFlt */
395	(PFNRT)RTErrConvertFromErrno,
396	(PFNRT)RTUuidCompare,
397	(PFNRT)RTUuidCompareStr,
398	(PFNRT)RTUuidFromStr,
399	(PFNRT)RTStrDupTag,
400	(PFNRT)RTStrFree,
401	(PFNRT)RTStrCopy,
402	(PFNRT)RTStrNCmp,
403	/* VBoxNetAdp */
404	(PFNRT)RTRandBytes,
405	/* VBoxUSB */
406	(PFNRT)RTPathStripFilename,
407	NULL
408	};
409	#endif /* RT_OS_DARWIN \|\| RT_OS_SOLARIS \|\| RT_OS_SOLARIS */
410
411
412	/**
413	* Initializes the device extentsion structure.
414	*
415	* @returns IPRT status code.
416	* @param pDevExt The device extension to initialize.
417	* @param cbSession The size of the session structure. The size of
418	* SUPDRVSESSION may be smaller when SUPDRV_AGNOSTIC is
419	* defined because we're skipping the OS specific members
420	* then.
421	*/
422	int VBOXCALL supdrvInitDevExt(PSUPDRVDEVEXT pDevExt, size_t cbSession)
423	{
424	int rc;
425
426	#ifdef SUPDRV_WITH_RELEASE_LOGGER
427	/*
428	* Create the release log.
429	*/
430	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
431	PRTLOGGER pRelLogger;
432	rc = RTLogCreate(&pRelLogger, 0 /* fFlags */, "all",
433	"VBOX_RELEASE_LOG", RT_ELEMENTS(s_apszGroups), s_apszGroups, RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER, NULL);
434	if (RT_SUCCESS(rc))
435	RTLogRelSetDefaultInstance(pRelLogger);
436	/** @todo Add native hook for getting logger config parameters and setting
437	* them. On linux we should use the module parameter stuff... */
438	#endif
439
440	/*
441	* Initialize it.
442	*/
443	memset(pDevExt, 0, sizeof(*pDevExt));
444	rc = RTSpinlockCreate(&pDevExt->Spinlock);
445	if (RT_SUCCESS(rc))
446	{
447	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
448	rc = RTSemMutexCreate(&pDevExt->mtxLdr);
449	#else
450	rc = RTSemFastMutexCreate(&pDevExt->mtxLdr);
451	#endif
452	if (RT_SUCCESS(rc))
453	{
454	rc = RTSemFastMutexCreate(&pDevExt->mtxComponentFactory);
455	if (RT_SUCCESS(rc))
456	{
457	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
458	rc = RTSemMutexCreate(&pDevExt->mtxGip);
459	#else
460	rc = RTSemFastMutexCreate(&pDevExt->mtxGip);
461	#endif
462	if (RT_SUCCESS(rc))
463	{
464	rc = supdrvGipCreate(pDevExt);
465	if (RT_SUCCESS(rc))
466	{
467	pDevExt->u32Cookie = BIRD; /** @todo make this random? */
468	pDevExt->cbSession = (uint32_t)cbSession;
469
470	/*
471	* Fixup the absolute symbols.
472	*
473	* Because of the table indexing assumptions we'll have a little #ifdef orgy
474	* here rather than distributing this to OS specific files. At least for now.
475	*/
476	#ifdef RT_OS_DARWIN
477	# if ARCH_BITS == 32
478	if (SUPR0GetPagingMode() >= SUPPAGINGMODE_AMD64)
479	{
480	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
481	g_aFunctions[1].pfn = (void )0x80; / SUPR0Abs64bitKernelCS - KERNEL64_CS, seg.h */
482	g_aFunctions[2].pfn = (void )0x88; / SUPR0Abs64bitKernelSS - KERNEL64_SS, seg.h */
483	g_aFunctions[3].pfn = (void )0x88; / SUPR0Abs64bitKernelDS - KERNEL64_SS, seg.h */
484	}
485	else
486	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
487	g_aFunctions[4].pfn = (void )0x08; / SUPR0AbsKernelCS - KERNEL_CS, seg.h */
488	g_aFunctions[5].pfn = (void )0x10; / SUPR0AbsKernelSS - KERNEL_DS, seg.h */
489	g_aFunctions[6].pfn = (void )0x10; / SUPR0AbsKernelDS - KERNEL_DS, seg.h */
490	g_aFunctions[7].pfn = (void )0x10; / SUPR0AbsKernelES - KERNEL_DS, seg.h */
491	g_aFunctions[8].pfn = (void )0x10; / SUPR0AbsKernelFS - KERNEL_DS, seg.h */
492	g_aFunctions[9].pfn = (void )0x48; / SUPR0AbsKernelGS - CPU_DATA_GS, seg.h */
493	# else /* 64-bit darwin: */
494	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
495	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
496	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
497	g_aFunctions[3].pfn = (void )0; / SUPR0Abs64bitKernelDS */
498	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
499	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
500	g_aFunctions[6].pfn = (void )0; / SUPR0AbsKernelDS */
501	g_aFunctions[7].pfn = (void )0; / SUPR0AbsKernelES */
502	g_aFunctions[8].pfn = (void )0; / SUPR0AbsKernelFS */
503	g_aFunctions[9].pfn = (void )0; / SUPR0AbsKernelGS */
504
505	# endif
506	#else /* !RT_OS_DARWIN */
507	# if ARCH_BITS == 64
508	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
509	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
510	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
511	g_aFunctions[3].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0Abs64bitKernelDS */
512	# else
513	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
514	# endif
515	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
516	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
517	g_aFunctions[6].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0AbsKernelDS */
518	g_aFunctions[7].pfn = (void )(uintptr_t)ASMGetES(); / SUPR0AbsKernelES */
519	g_aFunctions[8].pfn = (void )(uintptr_t)ASMGetFS(); / SUPR0AbsKernelFS */
520	g_aFunctions[9].pfn = (void )(uintptr_t)ASMGetGS(); / SUPR0AbsKernelGS */
521	#endif /* !RT_OS_DARWIN */
522	return VINF_SUCCESS;
523	}
524
525	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
526	RTSemMutexDestroy(pDevExt->mtxGip);
527	pDevExt->mtxGip = NIL_RTSEMMUTEX;
528	#else
529	RTSemFastMutexDestroy(pDevExt->mtxGip);
530	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
531	#endif
532	}
533	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
534	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
535	}
536	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
537	RTSemMutexDestroy(pDevExt->mtxLdr);
538	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
539	#else
540	RTSemFastMutexDestroy(pDevExt->mtxLdr);
541	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
542	#endif
543	}
544	RTSpinlockDestroy(pDevExt->Spinlock);
545	pDevExt->Spinlock = NIL_RTSPINLOCK;
546	}
547	#ifdef SUPDRV_WITH_RELEASE_LOGGER
548	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
549	RTLogDestroy(RTLogSetDefaultInstance(NULL));
550	#endif
551
552	return rc;
553	}
554
555
556	/**
557	* Delete the device extension (e.g. cleanup members).
558	*
559	* @param pDevExt The device extension to delete.
560	*/
561	void VBOXCALL supdrvDeleteDevExt(PSUPDRVDEVEXT pDevExt)
562	{
563	PSUPDRVOBJ pObj;
564	PSUPDRVUSAGE pUsage;
565
566	/*
567	* Kill mutexes and spinlocks.
568	*/
569	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
570	RTSemMutexDestroy(pDevExt->mtxGip);
571	pDevExt->mtxGip = NIL_RTSEMMUTEX;
572	#else
573	RTSemFastMutexDestroy(pDevExt->mtxGip);
574	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
575	#endif
576	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
577	RTSemMutexDestroy(pDevExt->mtxLdr);
578	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
579	#else
580	RTSemFastMutexDestroy(pDevExt->mtxLdr);
581	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
582	#endif
583	RTSpinlockDestroy(pDevExt->Spinlock);
584	pDevExt->Spinlock = NIL_RTSPINLOCK;
585	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
586	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
587
588	/*
589	* Free lists.
590	*/
591	/* objects. */
592	pObj = pDevExt->pObjs;
593	Assert(!pObj); /* (can trigger on forced unloads) */
594	pDevExt->pObjs = NULL;
595	while (pObj)
596	{
597	void *pvFree = pObj;
598	pObj = pObj->pNext;
599	RTMemFree(pvFree);
600	}
601
602	/* usage records. */
603	pUsage = pDevExt->pUsageFree;
604	pDevExt->pUsageFree = NULL;
605	while (pUsage)
606	{
607	void *pvFree = pUsage;
608	pUsage = pUsage->pNext;
609	RTMemFree(pvFree);
610	}
611
612	/* kill the GIP. */
613	supdrvGipDestroy(pDevExt);
614
615	#ifdef SUPDRV_WITH_RELEASE_LOGGER
616	/* destroy the loggers. */
617	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
618	RTLogDestroy(RTLogSetDefaultInstance(NULL));
619	#endif
620	}
621
622
623	/**
624	* Create session.
625	*
626	* @returns IPRT status code.
627	* @param pDevExt Device extension.
628	* @param fUser Flag indicating whether this is a user or kernel session.
629	* @param ppSession Where to store the pointer to the session data.
630	*/
631	int VBOXCALL supdrvCreateSession(PSUPDRVDEVEXT pDevExt, bool fUser, PSUPDRVSESSION *ppSession)
632	{
633	/*
634	* Allocate memory for the session data.
635	*/
636	int rc;
637	PSUPDRVSESSION pSession = *ppSession = (PSUPDRVSESSION)RTMemAllocZ(pDevExt->cbSession);
638	if (pSession)
639	{
640	/* Initialize session data. */
641	rc = RTSpinlockCreate(&pSession->Spinlock);
642	if (!rc)
643	{
644	rc = RTHandleTableCreateEx(&pSession->hHandleTable,
645	RTHANDLETABLE_FLAGS_LOCKED \| RTHANDLETABLE_FLAGS_CONTEXT,
646	1 /uBase/, 32768 /cMax/, supdrvSessionObjHandleRetain, pSession);
647	if (RT_SUCCESS(rc))
648	{
649	Assert(pSession->Spinlock != NIL_RTSPINLOCK);
650	pSession->pDevExt = pDevExt;
651	pSession->u32Cookie = BIRD_INV;
652	/*pSession->pLdrUsage = NULL;
653	pSession->pVM = NULL;
654	pSession->pUsage = NULL;
655	pSession->pGip = NULL;
656	pSession->fGipReferenced = false;
657	pSession->Bundle.cUsed = 0; */
658	pSession->Uid = NIL_RTUID;
659	pSession->Gid = NIL_RTGID;
660	if (fUser)
661	{
662	pSession->Process = RTProcSelf();
663	pSession->R0Process = RTR0ProcHandleSelf();
664	}
665	else
666	{
667	pSession->Process = NIL_RTPROCESS;
668	pSession->R0Process = NIL_RTR0PROCESS;
669	}
670
671	VBOXDRV_SUPDRV_SESSION_CREATE(pSession, fUser);
672	LogFlow(("Created session %p initial cookie=%#x\n", pSession, pSession->u32Cookie));
673	return VINF_SUCCESS;
674	}
675
676	RTSpinlockDestroy(pSession->Spinlock);
677	}
678	RTMemFree(pSession);
679	*ppSession = NULL;
680	Log(("Failed to create spinlock, rc=%d!\n", rc));
681	}
682	else
683	rc = VERR_NO_MEMORY;
684
685	return rc;
686	}
687
688
689	/**
690	* Shared code for cleaning up a session.
691	*
692	* @param pDevExt Device extension.
693	* @param pSession Session data.
694	* This data will be freed by this routine.
695	*/
696	void VBOXCALL supdrvCloseSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
697	{
698	VBOXDRV_SUPDRV_SESSION_CLOSE(pSession);
699
700	/*
701	* Cleanup the session first.
702	*/
703	supdrvCleanupSession(pDevExt, pSession);
704
705	/*
706	* Free the rest of the session stuff.
707	*/
708	RTSpinlockDestroy(pSession->Spinlock);
709	pSession->Spinlock = NIL_RTSPINLOCK;
710	pSession->pDevExt = NULL;
711	RTMemFree(pSession);
712	LogFlow(("supdrvCloseSession: returns\n"));
713	}
714
715
716	/**
717	* Shared code for cleaning up a session (but not quite freeing it).
718	*
719	* This is primarily intended for MAC OS X where we have to clean up the memory
720	* stuff before the file handle is closed.
721	*
722	* @param pDevExt Device extension.
723	* @param pSession Session data.
724	* This data will be freed by this routine.
725	*/
726	void VBOXCALL supdrvCleanupSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
727	{
728	int rc;
729	PSUPDRVBUNDLE pBundle;
730	LogFlow(("supdrvCleanupSession: pSession=%p\n", pSession));
731
732	/*
733	* Remove logger instances related to this session.
734	*/
735	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pSession);
736
737	/*
738	* Destroy the handle table.
739	*/
740	rc = RTHandleTableDestroy(pSession->hHandleTable, supdrvSessionObjHandleDelete, pSession);
741	AssertRC(rc);
742	pSession->hHandleTable = NIL_RTHANDLETABLE;
743
744	/*
745	* Release object references made in this session.
746	* In theory there should be noone racing us in this session.
747	*/
748	Log2(("release objects - start\n"));
749	if (pSession->pUsage)
750	{
751	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
752	PSUPDRVUSAGE pUsage;
753	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
754
755	while ((pUsage = pSession->pUsage) != NULL)
756	{
757	PSUPDRVOBJ pObj = pUsage->pObj;
758	pSession->pUsage = pUsage->pNext;
759
760	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
761	if (pUsage->cUsage < pObj->cUsage)
762	{
763	pObj->cUsage -= pUsage->cUsage;
764	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
765	}
766	else
767	{
768	/* Destroy the object and free the record. */
769	if (pDevExt->pObjs == pObj)
770	pDevExt->pObjs = pObj->pNext;
771	else
772	{
773	PSUPDRVOBJ pObjPrev;
774	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
775	if (pObjPrev->pNext == pObj)
776	{
777	pObjPrev->pNext = pObj->pNext;
778	break;
779	}
780	Assert(pObjPrev);
781	}
782	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
783
784	Log(("supdrvCleanupSession: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
785	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
786	if (pObj->pfnDestructor)
787	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
788	RTMemFree(pObj);
789	}
790
791	/* free it and continue. */
792	RTMemFree(pUsage);
793
794	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
795	}
796
797	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
798	AssertMsg(!pSession->pUsage, ("Some buster reregistered an object during desturction!\n"));
799	}
800	Log2(("release objects - done\n"));
801
802	/*
803	* Release memory allocated in the session.
804	*
805	* We do not serialize this as we assume that the application will
806	* not allocated memory while closing the file handle object.
807	*/
808	Log2(("freeing memory:\n"));
809	pBundle = &pSession->Bundle;
810	while (pBundle)
811	{
812	PSUPDRVBUNDLE pToFree;
813	unsigned i;
814
815	/*
816	* Check and unlock all entries in the bundle.
817	*/
818	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
819	{
820	if (pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ)
821	{
822	Log2(("eType=%d pvR0=%p pvR3=%p cb=%ld\n", pBundle->aMem[i].eType, RTR0MemObjAddress(pBundle->aMem[i].MemObj),
823	(void *)RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3), (long)RTR0MemObjSize(pBundle->aMem[i].MemObj)));
824	if (pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ)
825	{
826	rc = RTR0MemObjFree(pBundle->aMem[i].MapObjR3, false);
827	AssertRC(rc); /** @todo figure out how to handle this. */
828	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
829	}
830	rc = RTR0MemObjFree(pBundle->aMem[i].MemObj, true /* fFreeMappings */);
831	AssertRC(rc); /** @todo figure out how to handle this. */
832	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
833	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
834	}
835	}
836
837	/*
838	* Advance and free previous bundle.
839	*/
840	pToFree = pBundle;
841	pBundle = pBundle->pNext;
842
843	pToFree->pNext = NULL;
844	pToFree->cUsed = 0;
845	if (pToFree != &pSession->Bundle)
846	RTMemFree(pToFree);
847	}
848	Log2(("freeing memory - done\n"));
849
850	/*
851	* Deregister component factories.
852	*/
853	RTSemFastMutexRequest(pDevExt->mtxComponentFactory);
854	Log2(("deregistering component factories:\n"));
855	if (pDevExt->pComponentFactoryHead)
856	{
857	PSUPDRVFACTORYREG pPrev = NULL;
858	PSUPDRVFACTORYREG pCur = pDevExt->pComponentFactoryHead;
859	while (pCur)
860	{
861	if (pCur->pSession == pSession)
862	{
863	/* unlink it */
864	PSUPDRVFACTORYREG pNext = pCur->pNext;
865	if (pPrev)
866	pPrev->pNext = pNext;
867	else
868	pDevExt->pComponentFactoryHead = pNext;
869
870	/* free it */
871	pCur->pNext = NULL;
872	pCur->pSession = NULL;
873	pCur->pFactory = NULL;
874	RTMemFree(pCur);
875
876	/* next */
877	pCur = pNext;
878	}
879	else
880	{
881	/* next */
882	pPrev = pCur;
883	pCur = pCur->pNext;
884	}
885	}
886	}
887	RTSemFastMutexRelease(pDevExt->mtxComponentFactory);
888	Log2(("deregistering component factories - done\n"));
889
890	/*
891	* Loaded images needs to be dereferenced and possibly freed up.
892	*/
893	supdrvLdrLock(pDevExt);
894	Log2(("freeing images:\n"));
895	if (pSession->pLdrUsage)
896	{
897	PSUPDRVLDRUSAGE pUsage = pSession->pLdrUsage;
898	pSession->pLdrUsage = NULL;
899	while (pUsage)
900	{
901	void *pvFree = pUsage;
902	PSUPDRVLDRIMAGE pImage = pUsage->pImage;
903	if (pImage->cUsage > pUsage->cUsage)
904	pImage->cUsage -= pUsage->cUsage;
905	else
906	supdrvLdrFree(pDevExt, pImage);
907	pUsage->pImage = NULL;
908	pUsage = pUsage->pNext;
909	RTMemFree(pvFree);
910	}
911	}
912	supdrvLdrUnlock(pDevExt);
913	Log2(("freeing images - done\n"));
914
915	/*
916	* Unmap the GIP.
917	*/
918	Log2(("umapping GIP:\n"));
919	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
920	{
921	SUPR0GipUnmap(pSession);
922	pSession->fGipReferenced = 0;
923	}
924	Log2(("umapping GIP - done\n"));
925	}
926
927
928	/**
929	* RTHandleTableDestroy callback used by supdrvCleanupSession.
930	*
931	* @returns IPRT status code, see SUPR0ObjAddRef.
932	* @param hHandleTable The handle table handle. Ignored.
933	* @param pvObj The object pointer.
934	* @param pvCtx Context, the handle type. Ignored.
935	* @param pvUser Session pointer.
936	*/
937	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser)
938	{
939	NOREF(pvCtx);
940	NOREF(hHandleTable);
941	return SUPR0ObjAddRefEx(pvObj, (PSUPDRVSESSION)pvUser, true /fNoBlocking/);
942	}
943
944
945	/**
946	* RTHandleTableDestroy callback used by supdrvCleanupSession.
947	*
948	* @param hHandleTable The handle table handle. Ignored.
949	* @param h The handle value. Ignored.
950	* @param pvObj The object pointer.
951	* @param pvCtx Context, the handle type. Ignored.
952	* @param pvUser Session pointer.
953	*/
954	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser)
955	{
956	NOREF(pvCtx);
957	NOREF(h);
958	NOREF(hHandleTable);
959	SUPR0ObjRelease(pvObj, (PSUPDRVSESSION)pvUser);
960	}
961
962
963	/**
964	* Fast path I/O Control worker.
965	*
966	* @returns VBox status code that should be passed down to ring-3 unchanged.
967	* @param uIOCtl Function number.
968	* @param idCpu VMCPU id.
969	* @param pDevExt Device extention.
970	* @param pSession Session data.
971	*/
972	int VBOXCALL supdrvIOCtlFast(uintptr_t uIOCtl, VMCPUID idCpu, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
973	{
974	/*
975	* We check the two prereqs after doing this only to allow the compiler to optimize things better.
976	*/
977	if (RT_LIKELY( RT_VALID_PTR(pSession)
978	&& pSession->pVM
979	&& pDevExt->pfnVMMR0EntryFast))
980	{
981	switch (uIOCtl)
982	{
983	case SUP_IOCTL_FAST_DO_RAW_RUN:
984	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_RAW_RUN);
985	break;
986	case SUP_IOCTL_FAST_DO_HWACC_RUN:
987	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_HWACC_RUN);
988	break;
989	case SUP_IOCTL_FAST_DO_NOP:
990	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_NOP);
991	break;
992	default:
993	return VERR_INTERNAL_ERROR;
994	}
995	return VINF_SUCCESS;
996	}
997	return VERR_INTERNAL_ERROR;
998	}
999
1000
1001	/**
1002	* Helper for supdrvIOCtl. Check if pszStr contains any character of pszChars.
1003	* We would use strpbrk here if this function would be contained in the RedHat kABI white
1004	* list, see http://www.kerneldrivers.org/RHEL5.
1005	*
1006	* @returns 1 if pszStr does contain any character of pszChars, 0 otherwise.
1007	* @param pszStr String to check
1008	* @param pszChars Character set
1009	*/
1010	static int supdrvCheckInvalidChar(const char pszStr, const char pszChars)
1011	{
1012	int chCur;
1013	while ((chCur = *pszStr++) != '\0')
1014	{
1015	int ch;
1016	const char *psz = pszChars;
1017	while ((ch = *psz++) != '\0')
1018	if (ch == chCur)
1019	return 1;
1020
1021	}
1022	return 0;
1023	}
1024
1025
1026	/**
1027	* I/O Control worker.
1028	*
1029	* @returns IPRT status code.
1030	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1031	*
1032	* @param uIOCtl Function number.
1033	* @param pDevExt Device extention.
1034	* @param pSession Session data.
1035	* @param pReqHdr The request header.
1036	*/
1037	int VBOXCALL supdrvIOCtl(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1038	{
1039	VBOXDRV_SUPDRV_IOCTL_ENTRY(pSession, uIOCtl, pReqHdr);
1040
1041	/*
1042	* Validate the request.
1043	*/
1044	/* this first check could probably be omitted as its also done by the OS specific code... */
1045	if (RT_UNLIKELY( (pReqHdr->fFlags & SUPREQHDR_FLAGS_MAGIC_MASK) != SUPREQHDR_FLAGS_MAGIC
1046	\|\| pReqHdr->cbIn < sizeof(*pReqHdr)
1047	\|\| pReqHdr->cbOut < sizeof(*pReqHdr)))
1048	{
1049	OSDBGPRINT(("vboxdrv: Bad ioctl request header; cbIn=%#lx cbOut=%#lx fFlags=%#lx\n",
1050	(long)pReqHdr->cbIn, (long)pReqHdr->cbOut, (long)pReqHdr->fFlags));
1051	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1052	return VERR_INVALID_PARAMETER;
1053	}
1054	if (RT_UNLIKELY(!RT_VALID_PTR(pSession)))
1055	{
1056	OSDBGPRINT(("vboxdrv: Invalid pSession valud %p (ioctl=%p)\n", pSession, (void *)uIOCtl));
1057	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1058	return VERR_INVALID_PARAMETER;
1059	}
1060	if (RT_UNLIKELY(uIOCtl == SUP_IOCTL_COOKIE))
1061	{
1062	if (pReqHdr->u32Cookie != SUPCOOKIE_INITIAL_COOKIE)
1063	{
1064	OSDBGPRINT(("SUP_IOCTL_COOKIE: bad cookie %#lx\n", (long)pReqHdr->u32Cookie));
1065	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1066	return VERR_INVALID_PARAMETER;
1067	}
1068	}
1069	else if (RT_UNLIKELY( pReqHdr->u32Cookie != pDevExt->u32Cookie
1070	\|\| pReqHdr->u32SessionCookie != pSession->u32Cookie))
1071	{
1072	OSDBGPRINT(("vboxdrv: bad cookie %#lx / %#lx.\n", (long)pReqHdr->u32Cookie, (long)pReqHdr->u32SessionCookie));
1073	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1074	return VERR_INVALID_PARAMETER;
1075	}
1076
1077	/*
1078	* Validation macros
1079	*/
1080	#define REQ_CHECK_SIZES_EX(Name, cbInExpect, cbOutExpect) \
1081	do { \
1082	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect) \|\| pReqHdr->cbOut != (cbOutExpect))) \
1083	{ \
1084	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld. cbOut=%ld expected %ld.\n", \
1085	(long)pReq->Hdr.cbIn, (long)(cbInExpect), (long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
1086	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1087	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1088	} \
1089	} while (0)
1090
1091	#define REQ_CHECK_SIZES(Name) REQ_CHECK_SIZES_EX(Name, Name ## _SIZE_IN, Name ## _SIZE_OUT)
1092
1093	#define REQ_CHECK_SIZE_IN(Name, cbInExpect) \
1094	do { \
1095	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect))) \
1096	{ \
1097	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld.\n", \
1098	(long)pReq->Hdr.cbIn, (long)(cbInExpect))); \
1099	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1100	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1101	} \
1102	} while (0)
1103
1104	#define REQ_CHECK_SIZE_OUT(Name, cbOutExpect) \
1105	do { \
1106	if (RT_UNLIKELY(pReqHdr->cbOut != (cbOutExpect))) \
1107	{ \
1108	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbOut=%ld expected %ld.\n", \
1109	(long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
1110	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1111	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1112	} \
1113	} while (0)
1114
1115	#define REQ_CHECK_EXPR(Name, expr) \
1116	do { \
1117	if (RT_UNLIKELY(!(expr))) \
1118	{ \
1119	OSDBGPRINT(( #Name ": %s\n", #expr)); \
1120	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1121	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1122	} \
1123	} while (0)
1124
1125	#define REQ_CHECK_EXPR_FMT(expr, fmt) \
1126	do { \
1127	if (RT_UNLIKELY(!(expr))) \
1128	{ \
1129	OSDBGPRINT( fmt ); \
1130	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1131	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1132	} \
1133	} while (0)
1134
1135
1136	/*
1137	* The switch.
1138	*/
1139	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
1140	{
1141	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
1142	{
1143	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
1144	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
1145	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
1146	{
1147	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
1148	pReq->Hdr.rc = VERR_INVALID_MAGIC;
1149	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, VERR_INVALID_MAGIC);
1150	return 0;
1151	}
1152
1153	#if 0
1154	/*
1155	* Call out to the OS specific code and let it do permission checks on the
1156	* client process.
1157	*/
1158	if (!supdrvOSValidateClientProcess(pDevExt, pSession))
1159	{
1160	pReq->u.Out.u32Cookie = 0xffffffff;
1161	pReq->u.Out.u32SessionCookie = 0xffffffff;
1162	pReq->u.Out.u32SessionVersion = 0xffffffff;
1163	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1164	pReq->u.Out.pSession = NULL;
1165	pReq->u.Out.cFunctions = 0;
1166	pReq->Hdr.rc = VERR_PERMISSION_DENIED;
1167	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, VERR_PERMISSION_DENIED);
1168	return 0;
1169	}
1170	#endif
1171
1172	/*
1173	* Match the version.
1174	* The current logic is very simple, match the major interface version.
1175	*/
1176	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
1177	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
1178	{
1179	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1180	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
1181	pReq->u.Out.u32Cookie = 0xffffffff;
1182	pReq->u.Out.u32SessionCookie = 0xffffffff;
1183	pReq->u.Out.u32SessionVersion = 0xffffffff;
1184	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1185	pReq->u.Out.pSession = NULL;
1186	pReq->u.Out.cFunctions = 0;
1187	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1188	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1189	return 0;
1190	}
1191
1192	/*
1193	* Fill in return data and be gone.
1194	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
1195	* u32SessionVersion <= u32ReqVersion!
1196	*/
1197	/** @todo Somehow validate the client and negotiate a secure cookie... */
1198	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
1199	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
1200	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
1201	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1202	pReq->u.Out.pSession = pSession;
1203	pReq->u.Out.cFunctions = sizeof(g_aFunctions) / sizeof(g_aFunctions[0]);
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_QUERY_FUNCS(0)):
1210	{
1211	/* validate */
1212	PSUPQUERYFUNCS pReq = (PSUPQUERYFUNCS)pReqHdr;
1213	REQ_CHECK_SIZES_EX(SUP_IOCTL_QUERY_FUNCS, SUP_IOCTL_QUERY_FUNCS_SIZE_IN, SUP_IOCTL_QUERY_FUNCS_SIZE_OUT(RT_ELEMENTS(g_aFunctions)));
1214
1215	/* execute */
1216	pReq->u.Out.cFunctions = RT_ELEMENTS(g_aFunctions);
1217	memcpy(&pReq->u.Out.aFunctions[0], g_aFunctions, sizeof(g_aFunctions));
1218	pReq->Hdr.rc = VINF_SUCCESS;
1219	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1220	return 0;
1221	}
1222
1223	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_LOCK):
1224	{
1225	/* validate */
1226	PSUPPAGELOCK pReq = (PSUPPAGELOCK)pReqHdr;
1227	REQ_CHECK_SIZE_IN(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_IN);
1228	REQ_CHECK_SIZE_OUT(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_OUT(pReq->u.In.cPages));
1229	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.cPages > 0);
1230	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.pvR3 >= PAGE_SIZE);
1231
1232	/* execute */
1233	pReq->Hdr.rc = SUPR0LockMem(pSession, pReq->u.In.pvR3, pReq->u.In.cPages, &pReq->u.Out.aPages[0]);
1234	if (RT_FAILURE(pReq->Hdr.rc))
1235	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1236	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1237	return 0;
1238	}
1239
1240	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_UNLOCK):
1241	{
1242	/* validate */
1243	PSUPPAGEUNLOCK pReq = (PSUPPAGEUNLOCK)pReqHdr;
1244	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_UNLOCK);
1245
1246	/* execute */
1247	pReq->Hdr.rc = SUPR0UnlockMem(pSession, pReq->u.In.pvR3);
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_ALLOC):
1253	{
1254	/* validate */
1255	PSUPCONTALLOC pReq = (PSUPCONTALLOC)pReqHdr;
1256	REQ_CHECK_SIZES(SUP_IOCTL_CONT_ALLOC);
1257
1258	/* execute */
1259	pReq->Hdr.rc = SUPR0ContAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.HCPhys);
1260	if (RT_FAILURE(pReq->Hdr.rc))
1261	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1262	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1263	return 0;
1264	}
1265
1266	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_FREE):
1267	{
1268	/* validate */
1269	PSUPCONTFREE pReq = (PSUPCONTFREE)pReqHdr;
1270	REQ_CHECK_SIZES(SUP_IOCTL_CONT_FREE);
1271
1272	/* execute */
1273	pReq->Hdr.rc = SUPR0ContFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1274	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1275	return 0;
1276	}
1277
1278	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_OPEN):
1279	{
1280	/* validate */
1281	PSUPLDROPEN pReq = (PSUPLDROPEN)pReqHdr;
1282	REQ_CHECK_SIZES(SUP_IOCTL_LDR_OPEN);
1283	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs > 0);
1284	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs < 16*_1M);
1285	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1286	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1287	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits < pReq->u.In.cbImageWithTabs);
1288	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.szName[0]);
1289	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1290	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, !supdrvCheckInvalidChar(pReq->u.In.szName, ";:()[]{}/\\\|&*%#@!~`\"'"));
1291	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szFilename, sizeof(pReq->u.In.szFilename)));
1292
1293	/* execute */
1294	pReq->Hdr.rc = supdrvIOCtl_LdrOpen(pDevExt, pSession, pReq);
1295	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1296	return 0;
1297	}
1298
1299	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOAD):
1300	{
1301	/* validate */
1302	PSUPLDRLOAD pReq = (PSUPLDRLOAD)pReqHdr;
1303	REQ_CHECK_EXPR(Name, pReq->Hdr.cbIn >= sizeof(*pReq));
1304	REQ_CHECK_SIZES_EX(SUP_IOCTL_LDR_LOAD, SUP_IOCTL_LDR_LOAD_SIZE_IN(pReq->u.In.cbImageWithTabs), SUP_IOCTL_LDR_LOAD_SIZE_OUT);
1305	REQ_CHECK_EXPR(SUP_IOCTL_LDR_LOAD, pReq->u.In.cSymbols <= 16384);
1306	REQ_CHECK_EXPR_FMT( !pReq->u.In.cSymbols
1307	\|\| ( pReq->u.In.offSymbols < pReq->u.In.cbImageWithTabs
1308	&& pReq->u.In.offSymbols + pReq->u.In.cSymbols * sizeof(SUPLDRSYM) <= pReq->u.In.cbImageWithTabs),
1309	("SUP_IOCTL_LDR_LOAD: offSymbols=%#lx cSymbols=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offSymbols,
1310	(long)pReq->u.In.cSymbols, (long)pReq->u.In.cbImageWithTabs));
1311	REQ_CHECK_EXPR_FMT( !pReq->u.In.cbStrTab
1312	\|\| ( pReq->u.In.offStrTab < pReq->u.In.cbImageWithTabs
1313	&& pReq->u.In.offStrTab + pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs
1314	&& pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs),
1315	("SUP_IOCTL_LDR_LOAD: offStrTab=%#lx cbStrTab=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offStrTab,
1316	(long)pReq->u.In.cbStrTab, (long)pReq->u.In.cbImageWithTabs));
1317
1318	if (pReq->u.In.cSymbols)
1319	{
1320	uint32_t i;
1321	PSUPLDRSYM paSyms = (PSUPLDRSYM)&pReq->u.In.abImage[pReq->u.In.offSymbols];
1322	for (i = 0; i < pReq->u.In.cSymbols; i++)
1323	{
1324	REQ_CHECK_EXPR_FMT(paSyms[i].offSymbol < pReq->u.In.cbImageWithTabs,
1325	("SUP_IOCTL_LDR_LOAD: sym #%ld: symb off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offSymbol, (long)pReq->u.In.cbImageWithTabs));
1326	REQ_CHECK_EXPR_FMT(paSyms[i].offName < pReq->u.In.cbStrTab,
1327	("SUP_IOCTL_LDR_LOAD: sym #%ld: name off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1328	REQ_CHECK_EXPR_FMT(RTStrEnd((char const *)&pReq->u.In.abImage[pReq->u.In.offStrTab + paSyms[i].offName],
1329	pReq->u.In.cbStrTab - paSyms[i].offName),
1330	("SUP_IOCTL_LDR_LOAD: sym #%ld: unterminated name! (%#lx / %#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1331	}
1332	}
1333
1334	/* execute */
1335	pReq->Hdr.rc = supdrvIOCtl_LdrLoad(pDevExt, pSession, pReq);
1336	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1337	return 0;
1338	}
1339
1340	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_FREE):
1341	{
1342	/* validate */
1343	PSUPLDRFREE pReq = (PSUPLDRFREE)pReqHdr;
1344	REQ_CHECK_SIZES(SUP_IOCTL_LDR_FREE);
1345
1346	/* execute */
1347	pReq->Hdr.rc = supdrvIOCtl_LdrFree(pDevExt, pSession, pReq);
1348	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1349	return 0;
1350	}
1351
1352	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_GET_SYMBOL):
1353	{
1354	/* validate */
1355	PSUPLDRGETSYMBOL pReq = (PSUPLDRGETSYMBOL)pReqHdr;
1356	REQ_CHECK_SIZES(SUP_IOCTL_LDR_GET_SYMBOL);
1357	REQ_CHECK_EXPR(SUP_IOCTL_LDR_GET_SYMBOL, RTStrEnd(pReq->u.In.szSymbol, sizeof(pReq->u.In.szSymbol)));
1358
1359	/* execute */
1360	pReq->Hdr.rc = supdrvIOCtl_LdrGetSymbol(pDevExt, pSession, pReq);
1361	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1362	return 0;
1363	}
1364
1365	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0(0)):
1366	{
1367	/* validate */
1368	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1369	Log4(("SUP_IOCTL_CALL_VMMR0: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1370	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1371
1372	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_VMMR0_SIZE(0))
1373	{
1374	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(0), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(0));
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, NULL, pReq->u.In.u64Arg, pSession);
1379	else
1380	pReq->Hdr.rc = VERR_WRONG_ORDER;
1381	}
1382	else
1383	{
1384	PSUPVMMR0REQHDR pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1385	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR)),
1386	("SUP_IOCTL_CALL_VMMR0: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR))));
1387	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1388	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(pVMMReq->cbReq));
1389
1390	/* execute */
1391	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1392	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1393	else
1394	pReq->Hdr.rc = VERR_WRONG_ORDER;
1395	}
1396
1397	if ( RT_FAILURE(pReq->Hdr.rc)
1398	&& pReq->Hdr.rc != VERR_INTERRUPTED
1399	&& pReq->Hdr.rc != VERR_TIMEOUT)
1400	Log(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1401	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1402	else
1403	Log4(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1404	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1405	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1406	return 0;
1407	}
1408
1409	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0_BIG):
1410	{
1411	/* validate */
1412	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1413	PSUPVMMR0REQHDR pVMMReq;
1414	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1415	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1416
1417	pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1418	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR)),
1419	("SUP_IOCTL_CALL_VMMR0_BIG: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR))));
1420	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0_BIG, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1421	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));
1422
1423	/* execute */
1424	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1425	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1426	else
1427	pReq->Hdr.rc = VERR_WRONG_ORDER;
1428
1429	if ( RT_FAILURE(pReq->Hdr.rc)
1430	&& pReq->Hdr.rc != VERR_INTERRUPTED
1431	&& pReq->Hdr.rc != VERR_TIMEOUT)
1432	Log(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1433	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1434	else
1435	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1436	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1437	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1438	return 0;
1439	}
1440
1441	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_PAGING_MODE):
1442	{
1443	/* validate */
1444	PSUPGETPAGINGMODE pReq = (PSUPGETPAGINGMODE)pReqHdr;
1445	REQ_CHECK_SIZES(SUP_IOCTL_GET_PAGING_MODE);
1446
1447	/* execute */
1448	pReq->Hdr.rc = VINF_SUCCESS;
1449	pReq->u.Out.enmMode = SUPR0GetPagingMode();
1450	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1451	return 0;
1452	}
1453
1454	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_ALLOC):
1455	{
1456	/* validate */
1457	PSUPLOWALLOC pReq = (PSUPLOWALLOC)pReqHdr;
1458	REQ_CHECK_EXPR(SUP_IOCTL_LOW_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_LOW_ALLOC_SIZE_IN);
1459	REQ_CHECK_SIZES_EX(SUP_IOCTL_LOW_ALLOC, SUP_IOCTL_LOW_ALLOC_SIZE_IN, SUP_IOCTL_LOW_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1460
1461	/* execute */
1462	pReq->Hdr.rc = SUPR0LowAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1463	if (RT_FAILURE(pReq->Hdr.rc))
1464	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1465	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1466	return 0;
1467	}
1468
1469	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_FREE):
1470	{
1471	/* validate */
1472	PSUPLOWFREE pReq = (PSUPLOWFREE)pReqHdr;
1473	REQ_CHECK_SIZES(SUP_IOCTL_LOW_FREE);
1474
1475	/* execute */
1476	pReq->Hdr.rc = SUPR0LowFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
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_MAP):
1482	{
1483	/* validate */
1484	PSUPGIPMAP pReq = (PSUPGIPMAP)pReqHdr;
1485	REQ_CHECK_SIZES(SUP_IOCTL_GIP_MAP);
1486
1487	/* execute */
1488	pReq->Hdr.rc = SUPR0GipMap(pSession, &pReq->u.Out.pGipR3, &pReq->u.Out.HCPhysGip);
1489	if (RT_SUCCESS(pReq->Hdr.rc))
1490	pReq->u.Out.pGipR0 = pDevExt->pGip;
1491	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1492	return 0;
1493	}
1494
1495	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_UNMAP):
1496	{
1497	/* validate */
1498	PSUPGIPUNMAP pReq = (PSUPGIPUNMAP)pReqHdr;
1499	REQ_CHECK_SIZES(SUP_IOCTL_GIP_UNMAP);
1500
1501	/* execute */
1502	pReq->Hdr.rc = SUPR0GipUnmap(pSession);
1503	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1504	return 0;
1505	}
1506
1507	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SET_VM_FOR_FAST):
1508	{
1509	/* validate */
1510	PSUPSETVMFORFAST pReq = (PSUPSETVMFORFAST)pReqHdr;
1511	REQ_CHECK_SIZES(SUP_IOCTL_SET_VM_FOR_FAST);
1512	REQ_CHECK_EXPR_FMT( !pReq->u.In.pVMR0
1513	\|\| ( VALID_PTR(pReq->u.In.pVMR0)
1514	&& !((uintptr_t)pReq->u.In.pVMR0 & (PAGE_SIZE - 1))),
1515	("SUP_IOCTL_SET_VM_FOR_FAST: pVMR0=%p!\n", pReq->u.In.pVMR0));
1516	/* execute */
1517	pSession->pVM = pReq->u.In.pVMR0;
1518	pReq->Hdr.rc = VINF_SUCCESS;
1519	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1520	return 0;
1521	}
1522
1523	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_ALLOC_EX):
1524	{
1525	/* validate */
1526	PSUPPAGEALLOCEX pReq = (PSUPPAGEALLOCEX)pReqHdr;
1527	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC_EX, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_EX_SIZE_IN);
1528	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));
1529	REQ_CHECK_EXPR_FMT(pReq->u.In.fKernelMapping \|\| pReq->u.In.fUserMapping,
1530	("SUP_IOCTL_PAGE_ALLOC_EX: No mapping requested!\n"));
1531	REQ_CHECK_EXPR_FMT(pReq->u.In.fUserMapping,
1532	("SUP_IOCTL_PAGE_ALLOC_EX: Must have user mapping!\n"));
1533	REQ_CHECK_EXPR_FMT(!pReq->u.In.fReserved0 && !pReq->u.In.fReserved1,
1534	("SUP_IOCTL_PAGE_ALLOC_EX: fReserved0=%d fReserved1=%d\n", pReq->u.In.fReserved0, pReq->u.In.fReserved1));
1535
1536	/* execute */
1537	pReq->Hdr.rc = SUPR0PageAllocEx(pSession, pReq->u.In.cPages, 0 /* fFlags */,
1538	pReq->u.In.fUserMapping ? &pReq->u.Out.pvR3 : NULL,
1539	pReq->u.In.fKernelMapping ? &pReq->u.Out.pvR0 : NULL,
1540	&pReq->u.Out.aPages[0]);
1541	if (RT_FAILURE(pReq->Hdr.rc))
1542	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1543	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1544	return 0;
1545	}
1546
1547	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_MAP_KERNEL):
1548	{
1549	/* validate */
1550	PSUPPAGEMAPKERNEL pReq = (PSUPPAGEMAPKERNEL)pReqHdr;
1551	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_MAP_KERNEL);
1552	REQ_CHECK_EXPR_FMT(!pReq->u.In.fFlags, ("SUP_IOCTL_PAGE_MAP_KERNEL: fFlags=%#x! MBZ\n", pReq->u.In.fFlags));
1553	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_MAP_KERNEL: offSub=%#x\n", pReq->u.In.offSub));
1554	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1555	("SUP_IOCTL_PAGE_MAP_KERNEL: cbSub=%#x\n", pReq->u.In.cbSub));
1556
1557	/* execute */
1558	pReq->Hdr.rc = SUPR0PageMapKernel(pSession, pReq->u.In.pvR3, pReq->u.In.offSub, pReq->u.In.cbSub,
1559	pReq->u.In.fFlags, &pReq->u.Out.pvR0);
1560	if (RT_FAILURE(pReq->Hdr.rc))
1561	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1562	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1563	return 0;
1564	}
1565
1566	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_PROTECT):
1567	{
1568	/* validate */
1569	PSUPPAGEPROTECT pReq = (PSUPPAGEPROTECT)pReqHdr;
1570	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_PROTECT);
1571	REQ_CHECK_EXPR_FMT(!(pReq->u.In.fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)),
1572	("SUP_IOCTL_PAGE_PROTECT: fProt=%#x!\n", pReq->u.In.fProt));
1573	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_PROTECT: offSub=%#x\n", pReq->u.In.offSub));
1574	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1575	("SUP_IOCTL_PAGE_PROTECT: cbSub=%#x\n", pReq->u.In.cbSub));
1576
1577	/* execute */
1578	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);
1579	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1580	return 0;
1581	}
1582
1583	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_FREE):
1584	{
1585	/* validate */
1586	PSUPPAGEFREE pReq = (PSUPPAGEFREE)pReqHdr;
1587	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_FREE);
1588
1589	/* execute */
1590	pReq->Hdr.rc = SUPR0PageFree(pSession, pReq->u.In.pvR3);
1591	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1592	return 0;
1593	}
1594
1595	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_SERVICE(0)):
1596	{
1597	/* validate */
1598	PSUPCALLSERVICE pReq = (PSUPCALLSERVICE)pReqHdr;
1599	Log4(("SUP_IOCTL_CALL_SERVICE: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1600	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1601
1602	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
1603	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(0), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(0));
1604	else
1605	{
1606	PSUPR0SERVICEREQHDR pSrvReq = (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0];
1607	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR)),
1608	("SUP_IOCTL_CALL_SERVICE: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR))));
1609	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, pSrvReq->u32Magic == SUPR0SERVICEREQHDR_MAGIC);
1610	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(pSrvReq->cbReq), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(pSrvReq->cbReq));
1611	}
1612	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1613
1614	/* execute */
1615	pReq->Hdr.rc = supdrvIOCtl_CallServiceModule(pDevExt, pSession, pReq);
1616	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1617	return 0;
1618	}
1619
1620	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOGGER_SETTINGS(0)):
1621	{
1622	/* validate */
1623	PSUPLOGGERSETTINGS pReq = (PSUPLOGGERSETTINGS)pReqHdr;
1624	size_t cbStrTab;
1625	REQ_CHECK_SIZE_OUT(SUP_IOCTL_LOGGER_SETTINGS, SUP_IOCTL_LOGGER_SETTINGS_SIZE_OUT);
1626	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->Hdr.cbIn >= SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(1));
1627	cbStrTab = pReq->Hdr.cbIn - SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(0);
1628	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offGroups < cbStrTab);
1629	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offFlags < cbStrTab);
1630	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offDestination < cbStrTab);
1631	REQ_CHECK_EXPR_FMT(pReq->u.In.szStrings[cbStrTab - 1] == '\0',
1632	("SUP_IOCTL_LOGGER_SETTINGS: cbIn=%#x cbStrTab=%#zx LastChar=%d\n",
1633	pReq->Hdr.cbIn, cbStrTab, pReq->u.In.szStrings[cbStrTab - 1]));
1634	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhich <= SUPLOGGERSETTINGS_WHICH_RELEASE);
1635	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhat <= SUPLOGGERSETTINGS_WHAT_DESTROY);
1636
1637	/* execute */
1638	pReq->Hdr.rc = supdrvIOCtl_LoggerSettings(pDevExt, pSession, pReq);
1639	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1640	return 0;
1641	}
1642
1643	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP2):
1644	{
1645	/* validate */
1646	PSUPSEMOP2 pReq = (PSUPSEMOP2)pReqHdr;
1647	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP2, SUP_IOCTL_SEM_OP2_SIZE_IN, SUP_IOCTL_SEM_OP2_SIZE_OUT);
1648	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP2, pReq->u.In.uReserved == 0);
1649
1650	/* execute */
1651	switch (pReq->u.In.uType)
1652	{
1653	case SUP_SEM_TYPE_EVENT:
1654	{
1655	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1656	switch (pReq->u.In.uOp)
1657	{
1658	case SUPSEMOP2_WAIT_MS_REL:
1659	pReq->Hdr.rc = SUPSemEventWaitNoResume(pSession, hEvent, pReq->u.In.uArg.cRelMsTimeout);
1660	break;
1661	case SUPSEMOP2_WAIT_NS_ABS:
1662	pReq->Hdr.rc = SUPSemEventWaitNsAbsIntr(pSession, hEvent, pReq->u.In.uArg.uAbsNsTimeout);
1663	break;
1664	case SUPSEMOP2_WAIT_NS_REL:
1665	pReq->Hdr.rc = SUPSemEventWaitNsRelIntr(pSession, hEvent, pReq->u.In.uArg.cRelNsTimeout);
1666	break;
1667	case SUPSEMOP2_SIGNAL:
1668	pReq->Hdr.rc = SUPSemEventSignal(pSession, hEvent);
1669	break;
1670	case SUPSEMOP2_CLOSE:
1671	pReq->Hdr.rc = SUPSemEventClose(pSession, hEvent);
1672	break;
1673	case SUPSEMOP2_RESET:
1674	default:
1675	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1676	break;
1677	}
1678	break;
1679	}
1680
1681	case SUP_SEM_TYPE_EVENT_MULTI:
1682	{
1683	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1684	switch (pReq->u.In.uOp)
1685	{
1686	case SUPSEMOP2_WAIT_MS_REL:
1687	pReq->Hdr.rc = SUPSemEventMultiWaitNoResume(pSession, hEventMulti, pReq->u.In.uArg.cRelMsTimeout);
1688	break;
1689	case SUPSEMOP2_WAIT_NS_ABS:
1690	pReq->Hdr.rc = SUPSemEventMultiWaitNsAbsIntr(pSession, hEventMulti, pReq->u.In.uArg.uAbsNsTimeout);
1691	break;
1692	case SUPSEMOP2_WAIT_NS_REL:
1693	pReq->Hdr.rc = SUPSemEventMultiWaitNsRelIntr(pSession, hEventMulti, pReq->u.In.uArg.cRelNsTimeout);
1694	break;
1695	case SUPSEMOP2_SIGNAL:
1696	pReq->Hdr.rc = SUPSemEventMultiSignal(pSession, hEventMulti);
1697	break;
1698	case SUPSEMOP2_CLOSE:
1699	pReq->Hdr.rc = SUPSemEventMultiClose(pSession, hEventMulti);
1700	break;
1701	case SUPSEMOP2_RESET:
1702	pReq->Hdr.rc = SUPSemEventMultiReset(pSession, hEventMulti);
1703	break;
1704	default:
1705	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1706	break;
1707	}
1708	break;
1709	}
1710
1711	default:
1712	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1713	break;
1714	}
1715	return 0;
1716	}
1717
1718	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP3):
1719	{
1720	/* validate */
1721	PSUPSEMOP3 pReq = (PSUPSEMOP3)pReqHdr;
1722	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP3, SUP_IOCTL_SEM_OP3_SIZE_IN, SUP_IOCTL_SEM_OP3_SIZE_OUT);
1723	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, pReq->u.In.u32Reserved == 0 && pReq->u.In.u64Reserved == 0);
1724
1725	/* execute */
1726	switch (pReq->u.In.uType)
1727	{
1728	case SUP_SEM_TYPE_EVENT:
1729	{
1730	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1731	switch (pReq->u.In.uOp)
1732	{
1733	case SUPSEMOP3_CREATE:
1734	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1735	pReq->Hdr.rc = SUPSemEventCreate(pSession, &hEvent);
1736	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEvent;
1737	break;
1738	case SUPSEMOP3_GET_RESOLUTION:
1739	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1740	pReq->Hdr.rc = VINF_SUCCESS;
1741	pReq->Hdr.cbOut = sizeof(*pReq);
1742	pReq->u.Out.cNsResolution = SUPSemEventGetResolution(pSession);
1743	break;
1744	default:
1745	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1746	break;
1747	}
1748	break;
1749	}
1750
1751	case SUP_SEM_TYPE_EVENT_MULTI:
1752	{
1753	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1754	switch (pReq->u.In.uOp)
1755	{
1756	case SUPSEMOP3_CREATE:
1757	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1758	pReq->Hdr.rc = SUPSemEventMultiCreate(pSession, &hEventMulti);
1759	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEventMulti;
1760	break;
1761	case SUPSEMOP3_GET_RESOLUTION:
1762	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1763	pReq->Hdr.rc = VINF_SUCCESS;
1764	pReq->u.Out.cNsResolution = SUPSemEventMultiGetResolution(pSession);
1765	break;
1766	default:
1767	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1768	break;
1769	}
1770	break;
1771	}
1772
1773	default:
1774	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1775	break;
1776	}
1777	return 0;
1778	}
1779
1780	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
1781	{
1782	/* validate */
1783	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
1784	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
1785	REQ_CHECK_EXPR(SUP_IOCTL_VT_CAPS, pReq->Hdr.cbIn <= SUP_IOCTL_VT_CAPS_SIZE_IN);
1786
1787	/* execute */
1788	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.Caps);
1789	if (RT_FAILURE(pReq->Hdr.rc))
1790	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1791	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1792	return 0;
1793	}
1794
1795	default:
1796	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1797	break;
1798	}
1799	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_GENERAL_FAILURE, VERR_GENERAL_FAILURE);
1800	return VERR_GENERAL_FAILURE;
1801	}
1802
1803
1804	/**
1805	* Inter-Driver Communication (IDC) worker.
1806	*
1807	* @returns VBox status code.
1808	* @retval VINF_SUCCESS on success.
1809	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1810	* @retval VERR_NOT_SUPPORTED if the request isn't supported.
1811	*
1812	* @param uReq The request (function) code.
1813	* @param pDevExt Device extention.
1814	* @param pSession Session data.
1815	* @param pReqHdr The request header.
1816	*/
1817	int VBOXCALL supdrvIDC(uintptr_t uReq, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQHDR pReqHdr)
1818	{
1819	/*
1820	* The OS specific code has already validated the pSession
1821	* pointer, and the request size being greater or equal to
1822	* size of the header.
1823	*
1824	* So, just check that pSession is a kernel context session.
1825	*/
1826	if (RT_UNLIKELY( pSession
1827	&& pSession->R0Process != NIL_RTR0PROCESS))
1828	return VERR_INVALID_PARAMETER;
1829
1830	/*
1831	* Validation macro.
1832	*/
1833	#define REQ_CHECK_IDC_SIZE(Name, cbExpect) \
1834	do { \
1835	if (RT_UNLIKELY(pReqHdr->cb != (cbExpect))) \
1836	{ \
1837	OSDBGPRINT(( #Name ": Invalid input/output sizes. cb=%ld expected %ld.\n", \
1838	(long)pReqHdr->cb, (long)(cbExpect))); \
1839	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1840	} \
1841	} while (0)
1842
1843	switch (uReq)
1844	{
1845	case SUPDRV_IDC_REQ_CONNECT:
1846	{
1847	PSUPDRVIDCREQCONNECT pReq = (PSUPDRVIDCREQCONNECT)pReqHdr;
1848	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_CONNECT, sizeof(*pReq));
1849
1850	/*
1851	* Validate the cookie and other input.
1852	*/
1853	if (pReq->Hdr.pSession != NULL)
1854	{
1855	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Hdr.pSession=%p expected NULL!\n", pReq->Hdr.pSession));
1856	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1857	}
1858	if (pReq->u.In.u32MagicCookie != SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE)
1859	{
1860	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: u32MagicCookie=%#x expected %#x!\n",
1861	(unsigned)pReq->u.In.u32MagicCookie, (unsigned)SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE));
1862	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1863	}
1864	if ( pReq->u.In.uMinVersion > pReq->u.In.uReqVersion
1865	\|\| (pReq->u.In.uMinVersion & UINT32_C(0xffff0000)) != (pReq->u.In.uReqVersion & UINT32_C(0xffff0000)))
1866	{
1867	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: uMinVersion=%#x uMaxVersion=%#x doesn't match!\n",
1868	pReq->u.In.uMinVersion, pReq->u.In.uReqVersion));
1869	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1870	}
1871	if (pSession != NULL)
1872	{
1873	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: pSession=%p expected NULL!\n", pSession));
1874	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1875	}
1876
1877	/*
1878	* Match the version.
1879	* The current logic is very simple, match the major interface version.
1880	*/
1881	if ( pReq->u.In.uMinVersion > SUPDRV_IDC_VERSION
1882	\|\| (pReq->u.In.uMinVersion & 0xffff0000) != (SUPDRV_IDC_VERSION & 0xffff0000))
1883	{
1884	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1885	pReq->u.In.uReqVersion, pReq->u.In.uMinVersion, (unsigned)SUPDRV_IDC_VERSION));
1886	pReq->u.Out.pSession = NULL;
1887	pReq->u.Out.uSessionVersion = 0xffffffff;
1888	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1889	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1890	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1891	return VINF_SUCCESS;
1892	}
1893
1894	pReq->u.Out.pSession = NULL;
1895	pReq->u.Out.uSessionVersion = SUPDRV_IDC_VERSION;
1896	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1897	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1898
1899	pReq->Hdr.rc = supdrvCreateSession(pDevExt, false /* fUser */, &pSession);
1900	if (RT_FAILURE(pReq->Hdr.rc))
1901	{
1902	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: failed to create session, rc=%d\n", pReq->Hdr.rc));
1903	return VINF_SUCCESS;
1904	}
1905
1906	pReq->u.Out.pSession = pSession;
1907	pReq->Hdr.pSession = pSession;
1908
1909	return VINF_SUCCESS;
1910	}
1911
1912	case SUPDRV_IDC_REQ_DISCONNECT:
1913	{
1914	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_DISCONNECT, sizeof(*pReqHdr));
1915
1916	supdrvCloseSession(pDevExt, pSession);
1917	return pReqHdr->rc = VINF_SUCCESS;
1918	}
1919
1920	case SUPDRV_IDC_REQ_GET_SYMBOL:
1921	{
1922	PSUPDRVIDCREQGETSYM pReq = (PSUPDRVIDCREQGETSYM)pReqHdr;
1923	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_GET_SYMBOL, sizeof(*pReq));
1924
1925	pReq->Hdr.rc = supdrvIDC_LdrGetSymbol(pDevExt, pSession, pReq);
1926	return VINF_SUCCESS;
1927	}
1928
1929	case SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY:
1930	{
1931	PSUPDRVIDCREQCOMPREGFACTORY pReq = (PSUPDRVIDCREQCOMPREGFACTORY)pReqHdr;
1932	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY, sizeof(*pReq));
1933
1934	pReq->Hdr.rc = SUPR0ComponentRegisterFactory(pSession, pReq->u.In.pFactory);
1935	return VINF_SUCCESS;
1936	}
1937
1938	case SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY:
1939	{
1940	PSUPDRVIDCREQCOMPDEREGFACTORY pReq = (PSUPDRVIDCREQCOMPDEREGFACTORY)pReqHdr;
1941	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY, sizeof(*pReq));
1942
1943	pReq->Hdr.rc = SUPR0ComponentDeregisterFactory(pSession, pReq->u.In.pFactory);
1944	return VINF_SUCCESS;
1945	}
1946
1947	default:
1948	Log(("Unknown IDC %#lx\n", (long)uReq));
1949	break;
1950	}
1951
1952	#undef REQ_CHECK_IDC_SIZE
1953	return VERR_NOT_SUPPORTED;
1954	}
1955
1956
1957	/**
1958	* Register a object for reference counting.
1959	* The object is registered with one reference in the specified session.
1960	*
1961	* @returns Unique identifier on success (pointer).
1962	* All future reference must use this identifier.
1963	* @returns NULL on failure.
1964	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
1965	* @param pvUser1 The first user argument.
1966	* @param pvUser2 The second user argument.
1967	*/
1968	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
1969	{
1970	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1971	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1972	PSUPDRVOBJ pObj;
1973	PSUPDRVUSAGE pUsage;
1974
1975	/*
1976	* Validate the input.
1977	*/
1978	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
1979	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
1980	AssertPtrReturn(pfnDestructor, NULL);
1981
1982	/*
1983	* Allocate and initialize the object.
1984	*/
1985	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
1986	if (!pObj)
1987	return NULL;
1988	pObj->u32Magic = SUPDRVOBJ_MAGIC;
1989	pObj->enmType = enmType;
1990	pObj->pNext = NULL;
1991	pObj->cUsage = 1;
1992	pObj->pfnDestructor = pfnDestructor;
1993	pObj->pvUser1 = pvUser1;
1994	pObj->pvUser2 = pvUser2;
1995	pObj->CreatorUid = pSession->Uid;
1996	pObj->CreatorGid = pSession->Gid;
1997	pObj->CreatorProcess= pSession->Process;
1998	supdrvOSObjInitCreator(pObj, pSession);
1999
2000	/*
2001	* Allocate the usage record.
2002	* (We keep freed usage records around to simplify SUPR0ObjAddRefEx().)
2003	*/
2004	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2005
2006	pUsage = pDevExt->pUsageFree;
2007	if (pUsage)
2008	pDevExt->pUsageFree = pUsage->pNext;
2009	else
2010	{
2011	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2012	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
2013	if (!pUsage)
2014	{
2015	RTMemFree(pObj);
2016	return NULL;
2017	}
2018	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2019	}
2020
2021	/*
2022	* Insert the object and create the session usage record.
2023	*/
2024	/* The object. */
2025	pObj->pNext = pDevExt->pObjs;
2026	pDevExt->pObjs = pObj;
2027
2028	/* The session record. */
2029	pUsage->cUsage = 1;
2030	pUsage->pObj = pObj;
2031	pUsage->pNext = pSession->pUsage;
2032	/* Log2(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext)); */
2033	pSession->pUsage = pUsage;
2034
2035	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2036
2037	Log(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
2038	return pObj;
2039	}
2040
2041
2042	/**
2043	* Increment the reference counter for the object associating the reference
2044	* with the specified session.
2045	*
2046	* @returns IPRT status code.
2047	* @param pvObj The identifier returned by SUPR0ObjRegister().
2048	* @param pSession The session which is referencing the object.
2049	*
2050	* @remarks The caller should not own any spinlocks and must carefully protect
2051	* itself against potential race with the destructor so freed memory
2052	* isn't accessed here.
2053	*/
2054	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
2055	{
2056	return SUPR0ObjAddRefEx(pvObj, pSession, false /* fNoBlocking */);
2057	}
2058
2059
2060	/**
2061	* Increment the reference counter for the object associating the reference
2062	* with the specified session.
2063	*
2064	* @returns IPRT status code.
2065	* @retval VERR_TRY_AGAIN if fNoBlocking was set and a new usage record
2066	* couldn't be allocated. (If you see this you're not doing the right
2067	* thing and it won't ever work reliably.)
2068	*
2069	* @param pvObj The identifier returned by SUPR0ObjRegister().
2070	* @param pSession The session which is referencing the object.
2071	* @param fNoBlocking Set if it's not OK to block. Never try to make the
2072	* first reference to an object in a session with this
2073	* argument set.
2074	*
2075	* @remarks The caller should not own any spinlocks and must carefully protect
2076	* itself against potential race with the destructor so freed memory
2077	* isn't accessed here.
2078	*/
2079	SUPR0DECL(int) SUPR0ObjAddRefEx(void *pvObj, PSUPDRVSESSION pSession, bool fNoBlocking)
2080	{
2081	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2082	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2083	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2084	int rc = VINF_SUCCESS;
2085	PSUPDRVUSAGE pUsagePre;
2086	PSUPDRVUSAGE pUsage;
2087
2088	/*
2089	* Validate the input.
2090	* Be ready for the destruction race (someone might be stuck in the
2091	* destructor waiting a lock we own).
2092	*/
2093	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2094	AssertPtrReturn(pObj, VERR_INVALID_POINTER);
2095	AssertMsgReturn(pObj->u32Magic == SUPDRVOBJ_MAGIC \|\| pObj->u32Magic == SUPDRVOBJ_MAGIC_DEAD,
2096	("Invalid pvObj=%p magic=%#x (expected %#x or %#x)\n", pvObj, pObj->u32Magic, SUPDRVOBJ_MAGIC, SUPDRVOBJ_MAGIC_DEAD),
2097	VERR_INVALID_PARAMETER);
2098
2099	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2100
2101	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2102	{
2103	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2104
2105	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2106	return VERR_WRONG_ORDER;
2107	}
2108
2109	/*
2110	* Preallocate the usage record if we can.
2111	*/
2112	pUsagePre = pDevExt->pUsageFree;
2113	if (pUsagePre)
2114	pDevExt->pUsageFree = pUsagePre->pNext;
2115	else if (!fNoBlocking)
2116	{
2117	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2118	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
2119	if (!pUsagePre)
2120	return VERR_NO_MEMORY;
2121
2122	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2123	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2124	{
2125	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2126
2127	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2128	return VERR_WRONG_ORDER;
2129	}
2130	}
2131
2132	/*
2133	* Reference the object.
2134	*/
2135	pObj->cUsage++;
2136
2137	/*
2138	* Look for the session record.
2139	*/
2140	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
2141	{
2142	/Log(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2143	if (pUsage->pObj == pObj)
2144	break;
2145	}
2146	if (pUsage)
2147	pUsage->cUsage++;
2148	else if (pUsagePre)
2149	{
2150	/* create a new session record. */
2151	pUsagePre->cUsage = 1;
2152	pUsagePre->pObj = pObj;
2153	pUsagePre->pNext = pSession->pUsage;
2154	pSession->pUsage = pUsagePre;
2155	/Log(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));/
2156
2157	pUsagePre = NULL;
2158	}
2159	else
2160	{
2161	pObj->cUsage--;
2162	rc = VERR_TRY_AGAIN;
2163	}
2164
2165	/*
2166	* Put any unused usage record into the free list..
2167	*/
2168	if (pUsagePre)
2169	{
2170	pUsagePre->pNext = pDevExt->pUsageFree;
2171	pDevExt->pUsageFree = pUsagePre;
2172	}
2173
2174	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2175
2176	return rc;
2177	}
2178
2179
2180	/**
2181	* Decrement / destroy a reference counter record for an object.
2182	*
2183	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
2184	*
2185	* @returns IPRT status code.
2186	* @retval VINF_SUCCESS if not destroyed.
2187	* @retval VINF_OBJECT_DESTROYED if it's destroyed by this release call.
2188	* @retval VERR_INVALID_PARAMETER if the object isn't valid. Will assert in
2189	* string builds.
2190	*
2191	* @param pvObj The identifier returned by SUPR0ObjRegister().
2192	* @param pSession The session which is referencing the object.
2193	*/
2194	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
2195	{
2196	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2197	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2198	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2199	int rc = VERR_INVALID_PARAMETER;
2200	PSUPDRVUSAGE pUsage;
2201	PSUPDRVUSAGE pUsagePrev;
2202
2203	/*
2204	* Validate the input.
2205	*/
2206	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2207	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2208	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2209	VERR_INVALID_PARAMETER);
2210
2211	/*
2212	* Acquire the spinlock and look for the usage record.
2213	*/
2214	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2215
2216	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
2217	pUsage;
2218	pUsagePrev = pUsage, pUsage = pUsage->pNext)
2219	{
2220	/Log2(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2221	if (pUsage->pObj == pObj)
2222	{
2223	rc = VINF_SUCCESS;
2224	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
2225	if (pUsage->cUsage > 1)
2226	{
2227	pObj->cUsage--;
2228	pUsage->cUsage--;
2229	}
2230	else
2231	{
2232	/*
2233	* Free the session record.
2234	*/
2235	if (pUsagePrev)
2236	pUsagePrev->pNext = pUsage->pNext;
2237	else
2238	pSession->pUsage = pUsage->pNext;
2239	pUsage->pNext = pDevExt->pUsageFree;
2240	pDevExt->pUsageFree = pUsage;
2241
2242	/* What about the object? */
2243	if (pObj->cUsage > 1)
2244	pObj->cUsage--;
2245	else
2246	{
2247	/*
2248	* Object is to be destroyed, unlink it.
2249	*/
2250	pObj->u32Magic = SUPDRVOBJ_MAGIC_DEAD;
2251	rc = VINF_OBJECT_DESTROYED;
2252	if (pDevExt->pObjs == pObj)
2253	pDevExt->pObjs = pObj->pNext;
2254	else
2255	{
2256	PSUPDRVOBJ pObjPrev;
2257	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
2258	if (pObjPrev->pNext == pObj)
2259	{
2260	pObjPrev->pNext = pObj->pNext;
2261	break;
2262	}
2263	Assert(pObjPrev);
2264	}
2265	}
2266	}
2267	break;
2268	}
2269	}
2270
2271	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2272
2273	/*
2274	* Call the destructor and free the object if required.
2275	*/
2276	if (rc == VINF_OBJECT_DESTROYED)
2277	{
2278	Log(("SUPR0ObjRelease: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
2279	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
2280	if (pObj->pfnDestructor)
2281	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
2282	RTMemFree(pObj);
2283	}
2284
2285	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
2286	return rc;
2287	}
2288
2289
2290	/**
2291	* Verifies that the current process can access the specified object.
2292	*
2293	* @returns The following IPRT status code:
2294	* @retval VINF_SUCCESS if access was granted.
2295	* @retval VERR_PERMISSION_DENIED if denied access.
2296	* @retval VERR_INVALID_PARAMETER if invalid parameter.
2297	*
2298	* @param pvObj The identifier returned by SUPR0ObjRegister().
2299	* @param pSession The session which wishes to access the object.
2300	* @param pszObjName Object string name. This is optional and depends on the object type.
2301	*
2302	* @remark The caller is responsible for making sure the object isn't removed while
2303	* we're inside this function. If uncertain about this, just call AddRef before calling us.
2304	*/
2305	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
2306	{
2307	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2308	int rc;
2309
2310	/*
2311	* Validate the input.
2312	*/
2313	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2314	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2315	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2316	VERR_INVALID_PARAMETER);
2317
2318	/*
2319	* Check access. (returns true if a decision has been made.)
2320	*/
2321	rc = VERR_INTERNAL_ERROR;
2322	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
2323	return rc;
2324
2325	/*
2326	* Default policy is to allow the user to access his own
2327	* stuff but nothing else.
2328	*/
2329	if (pObj->CreatorUid == pSession->Uid)
2330	return VINF_SUCCESS;
2331	return VERR_PERMISSION_DENIED;
2332	}
2333
2334
2335	/**
2336	* Lock pages.
2337	*
2338	* @returns IPRT status code.
2339	* @param pSession Session to which the locked memory should be associated.
2340	* @param pvR3 Start of the memory range to lock.
2341	* This must be page aligned.
2342	* @param cPages Number of pages to lock.
2343	* @param paPages Where to put the physical addresses of locked memory.
2344	*/
2345	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
2346	{
2347	int rc;
2348	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2349	const size_t cb = (size_t)cPages << PAGE_SHIFT;
2350	LogFlow(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
2351
2352	/*
2353	* Verify input.
2354	*/
2355	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2356	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
2357	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
2358	\|\| !pvR3)
2359	{
2360	Log(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
2361	return VERR_INVALID_PARAMETER;
2362	}
2363
2364	/*
2365	* Let IPRT do the job.
2366	*/
2367	Mem.eType = MEMREF_TYPE_LOCKED;
2368	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTMEM_PROT_READ \| RTMEM_PROT_WRITE, RTR0ProcHandleSelf());
2369	if (RT_SUCCESS(rc))
2370	{
2371	uint32_t iPage = cPages;
2372	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
2373	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
2374
2375	while (iPage-- > 0)
2376	{
2377	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2378	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
2379	{
2380	AssertMsgFailed(("iPage=%d\n", iPage));
2381	rc = VERR_INTERNAL_ERROR;
2382	break;
2383	}
2384	}
2385	if (RT_SUCCESS(rc))
2386	rc = supdrvMemAdd(&Mem, pSession);
2387	if (RT_FAILURE(rc))
2388	{
2389	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
2390	AssertRC(rc2);
2391	}
2392	}
2393
2394	return rc;
2395	}
2396
2397
2398	/**
2399	* Unlocks the memory pointed to by pv.
2400	*
2401	* @returns IPRT status code.
2402	* @param pSession Session to which the memory was locked.
2403	* @param pvR3 Memory to unlock.
2404	*/
2405	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2406	{
2407	LogFlow(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2408	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2409	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
2410	}
2411
2412
2413	/**
2414	* Allocates a chunk of page aligned memory with contiguous and fixed physical
2415	* backing.
2416	*
2417	* @returns IPRT status code.
2418	* @param pSession Session data.
2419	* @param cPages Number of pages to allocate.
2420	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
2421	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
2422	* @param pHCPhys Where to put the physical address of allocated memory.
2423	*/
2424	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
2425	{
2426	int rc;
2427	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2428	LogFlow(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
2429
2430	/*
2431	* Validate input.
2432	*/
2433	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2434	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
2435	{
2436	Log(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
2437	pSession, ppvR0, ppvR3, pHCPhys));
2438	return VERR_INVALID_PARAMETER;
2439
2440	}
2441	if (cPages < 1 \|\| cPages >= 256)
2442	{
2443	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2444	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2445	}
2446
2447	/*
2448	* Let IPRT do the job.
2449	*/
2450	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
2451	if (RT_SUCCESS(rc))
2452	{
2453	int rc2;
2454	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2455	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2456	if (RT_SUCCESS(rc))
2457	{
2458	Mem.eType = MEMREF_TYPE_CONT;
2459	rc = supdrvMemAdd(&Mem, pSession);
2460	if (!rc)
2461	{
2462	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2463	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2464	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
2465	return 0;
2466	}
2467
2468	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2469	AssertRC(rc2);
2470	}
2471	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2472	AssertRC(rc2);
2473	}
2474
2475	return rc;
2476	}
2477
2478
2479	/**
2480	* Frees memory allocated using SUPR0ContAlloc().
2481	*
2482	* @returns IPRT status code.
2483	* @param pSession The session to which the memory was allocated.
2484	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2485	*/
2486	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2487	{
2488	LogFlow(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2489	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2490	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
2491	}
2492
2493
2494	/**
2495	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
2496	*
2497	* The memory isn't zeroed.
2498	*
2499	* @returns IPRT status code.
2500	* @param pSession Session data.
2501	* @param cPages Number of pages to allocate.
2502	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
2503	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
2504	* @param paPages Where to put the physical addresses of allocated memory.
2505	*/
2506	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
2507	{
2508	unsigned iPage;
2509	int rc;
2510	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2511	LogFlow(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
2512
2513	/*
2514	* Validate input.
2515	*/
2516	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2517	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
2518	{
2519	Log(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
2520	pSession, ppvR3, ppvR0, paPages));
2521	return VERR_INVALID_PARAMETER;
2522
2523	}
2524	if (cPages < 1 \|\| cPages >= 256)
2525	{
2526	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2527	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2528	}
2529
2530	/*
2531	* Let IPRT do the work.
2532	*/
2533	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
2534	if (RT_SUCCESS(rc))
2535	{
2536	int rc2;
2537	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2538	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2539	if (RT_SUCCESS(rc))
2540	{
2541	Mem.eType = MEMREF_TYPE_LOW;
2542	rc = supdrvMemAdd(&Mem, pSession);
2543	if (!rc)
2544	{
2545	for (iPage = 0; iPage < cPages; iPage++)
2546	{
2547	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2548	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%RHp\n", paPages[iPage]));
2549	}
2550	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2551	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2552	return 0;
2553	}
2554
2555	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2556	AssertRC(rc2);
2557	}
2558
2559	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2560	AssertRC(rc2);
2561	}
2562
2563	return rc;
2564	}
2565
2566
2567	/**
2568	* Frees memory allocated using SUPR0LowAlloc().
2569	*
2570	* @returns IPRT status code.
2571	* @param pSession The session to which the memory was allocated.
2572	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2573	*/
2574	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2575	{
2576	LogFlow(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2577	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2578	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
2579	}
2580
2581
2582
2583	/**
2584	* Allocates a chunk of memory with both R0 and R3 mappings.
2585	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
2586	*
2587	* @returns IPRT status code.
2588	* @param pSession The session to associated the allocation with.
2589	* @param cb Number of bytes to allocate.
2590	* @param ppvR0 Where to store the address of the Ring-0 mapping.
2591	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2592	*/
2593	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
2594	{
2595	int rc;
2596	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2597	LogFlow(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
2598
2599	/*
2600	* Validate input.
2601	*/
2602	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2603	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
2604	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
2605	if (cb < 1 \|\| cb >= _4M)
2606	{
2607	Log(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
2608	return VERR_INVALID_PARAMETER;
2609	}
2610
2611	/*
2612	* Let IPRT do the work.
2613	*/
2614	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
2615	if (RT_SUCCESS(rc))
2616	{
2617	int rc2;
2618	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2619	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2620	if (RT_SUCCESS(rc))
2621	{
2622	Mem.eType = MEMREF_TYPE_MEM;
2623	rc = supdrvMemAdd(&Mem, pSession);
2624	if (!rc)
2625	{
2626	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2627	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2628	return VINF_SUCCESS;
2629	}
2630
2631	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2632	AssertRC(rc2);
2633	}
2634
2635	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2636	AssertRC(rc2);
2637	}
2638
2639	return rc;
2640	}
2641
2642
2643	/**
2644	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
2645	*
2646	* @returns IPRT status code.
2647	* @param pSession The session to which the memory was allocated.
2648	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2649	* @param paPages Where to store the physical addresses.
2650	*/
2651	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
2652	{
2653	PSUPDRVBUNDLE pBundle;
2654	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2655	LogFlow(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
2656
2657	/*
2658	* Validate input.
2659	*/
2660	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2661	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
2662	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
2663
2664	/*
2665	* Search for the address.
2666	*/
2667	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2668	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2669	{
2670	if (pBundle->cUsed > 0)
2671	{
2672	unsigned i;
2673	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2674	{
2675	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
2676	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2677	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2678	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2679	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
2680	)
2681	)
2682	{
2683	const size_t cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2684	size_t iPage;
2685	for (iPage = 0; iPage < cPages; iPage++)
2686	{
2687	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2688	paPages[iPage].uReserved = 0;
2689	}
2690	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2691	return VINF_SUCCESS;
2692	}
2693	}
2694	}
2695	}
2696	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2697	Log(("Failed to find %p!!!\n", (void *)uPtr));
2698	return VERR_INVALID_PARAMETER;
2699	}
2700
2701
2702	/**
2703	* Free memory allocated by SUPR0MemAlloc().
2704	*
2705	* @returns IPRT status code.
2706	* @param pSession The session owning the allocation.
2707	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2708	*/
2709	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2710	{
2711	LogFlow(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2712	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2713	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
2714	}
2715
2716
2717	/**
2718	* Allocates a chunk of memory with a kernel or/and a user mode mapping.
2719	*
2720	* The memory is fixed and it's possible to query the physical addresses using
2721	* SUPR0MemGetPhys().
2722	*
2723	* @returns IPRT status code.
2724	* @param pSession The session to associated the allocation with.
2725	* @param cPages The number of pages to allocate.
2726	* @param fFlags Flags, reserved for the future. Must be zero.
2727	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2728	* NULL if no ring-3 mapping.
2729	* @param ppvR3 Where to store the address of the Ring-0 mapping.
2730	* NULL if no ring-0 mapping.
2731	* @param paPages Where to store the addresses of the pages. Optional.
2732	*/
2733	SUPR0DECL(int) SUPR0PageAllocEx(PSUPDRVSESSION pSession, uint32_t cPages, uint32_t fFlags, PRTR3PTR ppvR3, PRTR0PTR ppvR0, PRTHCPHYS paPages)
2734	{
2735	int rc;
2736	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2737	LogFlow(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
2738
2739	/*
2740	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2741	*/
2742	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2743	AssertPtrNullReturn(ppvR3, VERR_INVALID_POINTER);
2744	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2745	AssertReturn(ppvR3 \|\| ppvR0, VERR_INVALID_PARAMETER);
2746	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2747	if (cPages < 1 \|\| cPages > VBOX_MAX_ALLOC_PAGE_COUNT)
2748	{
2749	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)));
2750	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2751	}
2752
2753	/*
2754	* Let IPRT do the work.
2755	*/
2756	if (ppvR0)
2757	rc = RTR0MemObjAllocPage(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, true /* fExecutable */);
2758	else
2759	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
2760	if (RT_SUCCESS(rc))
2761	{
2762	int rc2;
2763	if (ppvR3)
2764	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2765	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2766	else
2767	Mem.MapObjR3 = NIL_RTR0MEMOBJ;
2768	if (RT_SUCCESS(rc))
2769	{
2770	Mem.eType = MEMREF_TYPE_PAGE;
2771	rc = supdrvMemAdd(&Mem, pSession);
2772	if (!rc)
2773	{
2774	if (ppvR3)
2775	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2776	if (ppvR0)
2777	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2778	if (paPages)
2779	{
2780	uint32_t iPage = cPages;
2781	while (iPage-- > 0)
2782	{
2783	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
2784	Assert(paPages[iPage] != NIL_RTHCPHYS);
2785	}
2786	}
2787	return VINF_SUCCESS;
2788	}
2789
2790	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2791	AssertRC(rc2);
2792	}
2793
2794	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2795	AssertRC(rc2);
2796	}
2797	return rc;
2798	}
2799
2800
2801	/**
2802	* Maps a chunk of memory previously allocated by SUPR0PageAllocEx into kernel
2803	* space.
2804	*
2805	* @returns IPRT status code.
2806	* @param pSession The session to associated the allocation with.
2807	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
2808	* @param offSub Where to start mapping. Must be page aligned.
2809	* @param cbSub How much to map. Must be page aligned.
2810	* @param fFlags Flags, MBZ.
2811	* @param ppvR0 Where to return the address of the ring-0 mapping on
2812	* success.
2813	*/
2814	SUPR0DECL(int) SUPR0PageMapKernel(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t offSub, uint32_t cbSub,
2815	uint32_t fFlags, PRTR0PTR ppvR0)
2816	{
2817	int rc;
2818	PSUPDRVBUNDLE pBundle;
2819	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2820	RTR0MEMOBJ hMemObj = NIL_RTR0MEMOBJ;
2821	LogFlow(("SUPR0PageMapKernel: pSession=%p pvR3=%p offSub=%#x cbSub=%#x\n", pSession, pvR3, offSub, cbSub));
2822
2823	/*
2824	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2825	*/
2826	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2827	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2828	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2829	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2830	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2831	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
2832
2833	/*
2834	* Find the memory object.
2835	*/
2836	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2837	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2838	{
2839	if (pBundle->cUsed > 0)
2840	{
2841	unsigned i;
2842	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2843	{
2844	if ( ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
2845	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2846	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2847	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
2848	\|\| ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED
2849	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2850	&& pBundle->aMem[i].MapObjR3 == NIL_RTR0MEMOBJ
2851	&& RTR0MemObjAddressR3(pBundle->aMem[i].MemObj) == pvR3))
2852	{
2853	hMemObj = pBundle->aMem[i].MemObj;
2854	break;
2855	}
2856	}
2857	}
2858	}
2859	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2860
2861	rc = VERR_INVALID_PARAMETER;
2862	if (hMemObj != NIL_RTR0MEMOBJ)
2863	{
2864	/*
2865	* Do some further input validations before calling IPRT.
2866	* (Cleanup is done indirectly by telling RTR0MemObjFree to include mappings.)
2867	*/
2868	size_t cbMemObj = RTR0MemObjSize(hMemObj);
2869	if ( offSub < cbMemObj
2870	&& cbSub <= cbMemObj
2871	&& offSub + cbSub <= cbMemObj)
2872	{
2873	RTR0MEMOBJ hMapObj;
2874	rc = RTR0MemObjMapKernelEx(&hMapObj, hMemObj, (void *)-1, 0,
2875	RTMEM_PROT_READ \| RTMEM_PROT_WRITE, offSub, cbSub);
2876	if (RT_SUCCESS(rc))
2877	*ppvR0 = RTR0MemObjAddress(hMapObj);
2878	}
2879	else
2880	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
2881
2882	}
2883	return rc;
2884	}
2885
2886
2887	/**
2888	* Changes the page level protection of one or more pages previously allocated
2889	* by SUPR0PageAllocEx.
2890	*
2891	* @returns IPRT status code.
2892	* @param pSession The session to associated the allocation with.
2893	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
2894	* NIL_RTR3PTR if the ring-3 mapping should be unaffected.
2895	* @param pvR0 The ring-0 address returned by SUPR0PageAllocEx.
2896	* NIL_RTR0PTR if the ring-0 mapping should be unaffected.
2897	* @param offSub Where to start changing. Must be page aligned.
2898	* @param cbSub How much to change. Must be page aligned.
2899	* @param fProt The new page level protection, see RTMEM_PROT_*.
2900	*/
2901	SUPR0DECL(int) SUPR0PageProtect(PSUPDRVSESSION pSession, RTR3PTR pvR3, RTR0PTR pvR0, uint32_t offSub, uint32_t cbSub, uint32_t fProt)
2902	{
2903	int rc;
2904	PSUPDRVBUNDLE pBundle;
2905	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2906	RTR0MEMOBJ hMemObjR0 = NIL_RTR0MEMOBJ;
2907	RTR0MEMOBJ hMemObjR3 = NIL_RTR0MEMOBJ;
2908	LogFlow(("SUPR0PageProtect: pSession=%p pvR3=%p pvR0=%p offSub=%#x cbSub=%#x fProt-%#x\n", pSession, pvR3, pvR0, offSub, cbSub, fProt));
2909
2910	/*
2911	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2912	*/
2913	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2914	AssertReturn(!(fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)), VERR_INVALID_PARAMETER);
2915	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2916	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2917	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
2918
2919	/*
2920	* Find the memory object.
2921	*/
2922	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2923	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2924	{
2925	if (pBundle->cUsed > 0)
2926	{
2927	unsigned i;
2928	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2929	{
2930	if ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
2931	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2932	&& ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2933	\|\| pvR3 == NIL_RTR3PTR)
2934	&& ( pvR0 == NIL_RTR0PTR
2935	\|\| RTR0MemObjAddress(pBundle->aMem[i].MemObj) == pvR0)
2936	&& ( pvR3 == NIL_RTR3PTR
2937	\|\| RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3))
2938	{
2939	if (pvR0 != NIL_RTR0PTR)
2940	hMemObjR0 = pBundle->aMem[i].MemObj;
2941	if (pvR3 != NIL_RTR3PTR)
2942	hMemObjR3 = pBundle->aMem[i].MapObjR3;
2943	break;
2944	}
2945	}
2946	}
2947	}
2948	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2949
2950	rc = VERR_INVALID_PARAMETER;
2951	if ( hMemObjR0 != NIL_RTR0MEMOBJ
2952	\|\| hMemObjR3 != NIL_RTR0MEMOBJ)
2953	{
2954	/*
2955	* Do some further input validations before calling IPRT.
2956	*/
2957	size_t cbMemObj = hMemObjR0 != NIL_RTR0PTR ? RTR0MemObjSize(hMemObjR0) : RTR0MemObjSize(hMemObjR3);
2958	if ( offSub < cbMemObj
2959	&& cbSub <= cbMemObj
2960	&& offSub + cbSub <= cbMemObj)
2961	{
2962	rc = VINF_SUCCESS;
2963	if (hMemObjR3 != NIL_RTR0PTR)
2964	rc = RTR0MemObjProtect(hMemObjR3, offSub, cbSub, fProt);
2965	if (hMemObjR0 != NIL_RTR0PTR && RT_SUCCESS(rc))
2966	rc = RTR0MemObjProtect(hMemObjR0, offSub, cbSub, fProt);
2967	}
2968	else
2969	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
2970
2971	}
2972	return rc;
2973
2974	}
2975
2976
2977	/**
2978	* Free memory allocated by SUPR0PageAlloc() and SUPR0PageAllocEx().
2979	*
2980	* @returns IPRT status code.
2981	* @param pSession The session owning the allocation.
2982	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc() or
2983	* SUPR0PageAllocEx().
2984	*/
2985	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2986	{
2987	LogFlow(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2988	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2989	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_PAGE);
2990	}
2991
2992
2993	/**
2994	* Gets the paging mode of the current CPU.
2995	*
2996	* @returns Paging mode, SUPPAGEINGMODE_INVALID on error.
2997	*/
2998	SUPR0DECL(SUPPAGINGMODE) SUPR0GetPagingMode(void)
2999	{
3000	SUPPAGINGMODE enmMode;
3001
3002	RTR0UINTREG cr0 = ASMGetCR0();
3003	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
3004	enmMode = SUPPAGINGMODE_INVALID;
3005	else
3006	{
3007	RTR0UINTREG cr4 = ASMGetCR4();
3008	uint32_t fNXEPlusLMA = 0;
3009	if (cr4 & X86_CR4_PAE)
3010	{
3011	uint32_t fAmdFeatures = ASMCpuId_EDX(0x80000001);
3012	if (fAmdFeatures & (X86_CPUID_AMD_FEATURE_EDX_NX \| X86_CPUID_AMD_FEATURE_EDX_LONG_MODE))
3013	{
3014	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
3015	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
3016	fNXEPlusLMA \|= RT_BIT(0);
3017	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
3018	fNXEPlusLMA \|= RT_BIT(1);
3019	}
3020	}
3021
3022	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
3023	{
3024	case 0:
3025	enmMode = SUPPAGINGMODE_32_BIT;
3026	break;
3027
3028	case X86_CR4_PGE:
3029	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
3030	break;
3031
3032	case X86_CR4_PAE:
3033	enmMode = SUPPAGINGMODE_PAE;
3034	break;
3035
3036	case X86_CR4_PAE \| RT_BIT(0):
3037	enmMode = SUPPAGINGMODE_PAE_NX;
3038	break;
3039
3040	case X86_CR4_PAE \| X86_CR4_PGE:
3041	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3042	break;
3043
3044	case X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3045	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3046	break;
3047
3048	case RT_BIT(1) \| X86_CR4_PAE:
3049	enmMode = SUPPAGINGMODE_AMD64;
3050	break;
3051
3052	case RT_BIT(1) \| X86_CR4_PAE \| RT_BIT(0):
3053	enmMode = SUPPAGINGMODE_AMD64_NX;
3054	break;
3055
3056	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
3057	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
3058	break;
3059
3060	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3061	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
3062	break;
3063
3064	default:
3065	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
3066	enmMode = SUPPAGINGMODE_INVALID;
3067	break;
3068	}
3069	}
3070	return enmMode;
3071	}
3072
3073
3074	/**
3075	* Enables or disabled hardware virtualization extensions using native OS APIs.
3076	*
3077	* @returns VBox status code.
3078	* @retval VINF_SUCCESS on success.
3079	* @retval VERR_NOT_SUPPORTED if not supported by the native OS.
3080	*
3081	* @param fEnable Whether to enable or disable.
3082	*/
3083	SUPR0DECL(int) SUPR0EnableVTx(bool fEnable)
3084	{
3085	#ifdef RT_OS_DARWIN
3086	return supdrvOSEnableVTx(fEnable);
3087	#else
3088	return VERR_NOT_SUPPORTED;
3089	#endif
3090	}
3091
3092
3093	/** @todo document me */
3094	SUPR0DECL(int) SUPR0QueryVTCaps(PSUPDRVSESSION pSession, uint32_t *pfCaps)
3095	{
3096	/*
3097	* Input validation.
3098	*/
3099	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3100	AssertPtrReturn(pfCaps, VERR_INVALID_POINTER);
3101
3102	*pfCaps = 0;
3103
3104	if (ASMHasCpuId())
3105	{
3106	uint32_t u32FeaturesECX;
3107	uint32_t u32Dummy;
3108	uint32_t u32FeaturesEDX;
3109	uint32_t u32VendorEBX, u32VendorECX, u32VendorEDX, u32AMDFeatureEDX, u32AMDFeatureECX;
3110	uint64_t val;
3111
3112	ASMCpuId(0, &u32Dummy, &u32VendorEBX, &u32VendorECX, &u32VendorEDX);
3113	ASMCpuId(1, &u32Dummy, &u32Dummy, &u32FeaturesECX, &u32FeaturesEDX);
3114	/* Query AMD features. */
3115	ASMCpuId(0x80000001, &u32Dummy, &u32Dummy, &u32AMDFeatureECX, &u32AMDFeatureEDX);
3116
3117	if ( u32VendorEBX == X86_CPUID_VENDOR_INTEL_EBX
3118	&& u32VendorECX == X86_CPUID_VENDOR_INTEL_ECX
3119	&& u32VendorEDX == X86_CPUID_VENDOR_INTEL_EDX
3120	)
3121	{
3122	if ( (u32FeaturesECX & X86_CPUID_FEATURE_ECX_VMX)
3123	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3124	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3125	)
3126	{
3127	val = ASMRdMsr(MSR_IA32_FEATURE_CONTROL);
3128	/*
3129	* Both the LOCK and VMXON bit must be set; otherwise VMXON will generate a #GP.
3130	* Once the lock bit is set, this MSR can no longer be modified.
3131	*/
3132	if ( (val & (MSR_IA32_FEATURE_CONTROL_VMXON\|MSR_IA32_FEATURE_CONTROL_LOCK))
3133	== (MSR_IA32_FEATURE_CONTROL_VMXON\|MSR_IA32_FEATURE_CONTROL_LOCK) /* enabled and locked */
3134	\|\| !(val & MSR_IA32_FEATURE_CONTROL_LOCK) /* not enabled, but not locked either */
3135	)
3136	{
3137	VMX_CAPABILITY vtCaps;
3138
3139	*pfCaps \|= SUPVTCAPS_VT_X;
3140
3141	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS);
3142	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC_USE_SECONDARY_EXEC_CTRL)
3143	{
3144	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS2);
3145	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC2_EPT)
3146	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3147	}
3148	return VINF_SUCCESS;
3149	}
3150	return VERR_VMX_MSR_LOCKED_OR_DISABLED;
3151	}
3152	return VERR_VMX_NO_VMX;
3153	}
3154
3155	if ( u32VendorEBX == X86_CPUID_VENDOR_AMD_EBX
3156	&& u32VendorECX == X86_CPUID_VENDOR_AMD_ECX
3157	&& u32VendorEDX == X86_CPUID_VENDOR_AMD_EDX
3158	)
3159	{
3160	if ( (u32AMDFeatureECX & X86_CPUID_AMD_FEATURE_ECX_SVM)
3161	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3162	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3163	)
3164	{
3165	/* Check if SVM is disabled */
3166	val = ASMRdMsr(MSR_K8_VM_CR);
3167	if (!(val & MSR_K8_VM_CR_SVM_DISABLE))
3168	{
3169	*pfCaps \|= SUPVTCAPS_AMD_V;
3170
3171	/* Query AMD features. */
3172	ASMCpuId(0x8000000A, &u32Dummy, &u32Dummy, &u32Dummy, &u32FeaturesEDX);
3173
3174	if (u32FeaturesEDX & AMD_CPUID_SVM_FEATURE_EDX_NESTED_PAGING)
3175	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3176
3177	return VINF_SUCCESS;
3178	}
3179	return VERR_SVM_DISABLED;
3180	}
3181	return VERR_SVM_NO_SVM;
3182	}
3183	}
3184
3185	return VERR_UNSUPPORTED_CPU;
3186	}
3187
3188
3189	/**
3190	* (Re-)initializes the per-cpu structure prior to starting or resuming the GIP
3191	* updating.
3192	*
3193	* @param pGipCpu The per CPU structure for this CPU.
3194	* @param u64NanoTS The current time.
3195	*/
3196	static void supdrvGipReInitCpu(PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
3197	{
3198	pGipCpu->u64TSC = ASMReadTSC() - pGipCpu->u32UpdateIntervalTSC;
3199	pGipCpu->u64NanoTS = u64NanoTS;
3200	}
3201
3202
3203	/**
3204	* Set the current TSC and NanoTS value for the CPU.
3205	*
3206	* @param idCpu The CPU ID. Unused - we have to use the APIC ID.
3207	* @param pvUser1 Pointer to the ring-0 GIP mapping.
3208	* @param pvUser2 Pointer to the variable holding the current time.
3209	*/
3210	static DECLCALLBACK(void) supdrvGipReInitCpuCallback(RTCPUID idCpu, void pvUser1, void pvUser2)
3211	{
3212	PSUPGLOBALINFOPAGE pGip = (PSUPGLOBALINFOPAGE)pvUser1;
3213	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
3214
3215	if (RT_LIKELY(iCpu < pGip->cCpus && pGip->aCPUs[iCpu].idCpu == idCpu))
3216	supdrvGipReInitCpu(&pGip->aCPUs[iCpu], (uint64_t )pvUser2);
3217
3218	NOREF(pvUser2);
3219	NOREF(idCpu);
3220	}
3221
3222
3223	/**
3224	* Maps the GIP into userspace and/or get the physical address of the GIP.
3225	*
3226	* @returns IPRT status code.
3227	* @param pSession Session to which the GIP mapping should belong.
3228	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
3229	* @param pHCPhysGip Where to store the physical address. (optional)
3230	*
3231	* @remark There is no reference counting on the mapping, so one call to this function
3232	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
3233	* and remove the session as a GIP user.
3234	*/
3235	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
3236	{
3237	int rc;
3238	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3239	RTR3PTR pGipR3 = NIL_RTR3PTR;
3240	RTHCPHYS HCPhys = NIL_RTHCPHYS;
3241	LogFlow(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
3242
3243	/*
3244	* Validate
3245	*/
3246	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3247	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
3248	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
3249
3250	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3251	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3252	#else
3253	RTSemFastMutexRequest(pDevExt->mtxGip);
3254	#endif
3255	if (pDevExt->pGip)
3256	{
3257	/*
3258	* Map it?
3259	*/
3260	rc = VINF_SUCCESS;
3261	if (ppGipR3)
3262	{
3263	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
3264	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
3265	RTMEM_PROT_READ, RTR0ProcHandleSelf());
3266	if (RT_SUCCESS(rc))
3267	pGipR3 = RTR0MemObjAddressR3(pSession->GipMapObjR3);
3268	}
3269
3270	/*
3271	* Get physical address.
3272	*/
3273	if (pHCPhysGip && RT_SUCCESS(rc))
3274	HCPhys = pDevExt->HCPhysGip;
3275
3276	/*
3277	* Reference globally.
3278	*/
3279	if (!pSession->fGipReferenced && RT_SUCCESS(rc))
3280	{
3281	pSession->fGipReferenced = 1;
3282	pDevExt->cGipUsers++;
3283	if (pDevExt->cGipUsers == 1)
3284	{
3285	PSUPGLOBALINFOPAGE pGipR0 = pDevExt->pGip;
3286	uint64_t u64NanoTS;
3287	uint32_t u32SystemResolution;
3288	unsigned i;
3289
3290	LogFlow(("SUPR0GipMap: Resumes GIP updating\n"));
3291
3292	/*
3293	* Try bump up the system timer resolution.
3294	* The more interrupts the better...
3295	*/
3296	if ( RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
3297	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
3298	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1953125 /* 512 HZ */, &u32SystemResolution))
3299	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 2000000 /* 500 HZ */, &u32SystemResolution))
3300	)
3301	{
3302	Assert(RTTimerGetSystemGranularity() <= u32SystemResolution);
3303	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
3304	}
3305
3306	if (pGipR0->aCPUs[0].u32TransactionId != 2 /* not the first time */)
3307	{
3308	for (i = 0; i < RT_ELEMENTS(pGipR0->aCPUs); i++)
3309	ASMAtomicUoWriteU32(&pGipR0->aCPUs[i].u32TransactionId,
3310	(pGipR0->aCPUs[i].u32TransactionId + GIP_UPDATEHZ_RECALC_FREQ * 2)
3311	& ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
3312	ASMAtomicWriteU64(&pGipR0->u64NanoTSLastUpdateHz, 0);
3313	}
3314
3315	u64NanoTS = RTTimeSystemNanoTS() - pGipR0->u32UpdateIntervalNS;
3316	if ( pGipR0->u32Mode == SUPGIPMODE_SYNC_TSC
3317	\|\| RTMpGetOnlineCount() == 1)
3318	supdrvGipReInitCpu(&pGipR0->aCPUs[0], u64NanoTS);
3319	else
3320	RTMpOnAll(supdrvGipReInitCpuCallback, pGipR0, &u64NanoTS);
3321
3322	#ifndef DO_NOT_START_GIP
3323	rc = RTTimerStart(pDevExt->pGipTimer, 0); AssertRC(rc);
3324	#endif
3325	rc = VINF_SUCCESS;
3326	}
3327	}
3328	}
3329	else
3330	{
3331	rc = VERR_GENERAL_FAILURE;
3332	Log(("SUPR0GipMap: GIP is not available!\n"));
3333	}
3334	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3335	RTSemMutexRelease(pDevExt->mtxGip);
3336	#else
3337	RTSemFastMutexRelease(pDevExt->mtxGip);
3338	#endif
3339
3340	/*
3341	* Write returns.
3342	*/
3343	if (pHCPhysGip)
3344	*pHCPhysGip = HCPhys;
3345	if (ppGipR3)
3346	*ppGipR3 = pGipR3;
3347
3348	#ifdef DEBUG_DARWIN_GIP
3349	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3350	#else
3351	LogFlow(( "SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3352	#endif
3353	return rc;
3354	}
3355
3356
3357	/**
3358	* Unmaps any user mapping of the GIP and terminates all GIP access
3359	* from this session.
3360	*
3361	* @returns IPRT status code.
3362	* @param pSession Session to which the GIP mapping should belong.
3363	*/
3364	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
3365	{
3366	int rc = VINF_SUCCESS;
3367	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3368	#ifdef DEBUG_DARWIN_GIP
3369	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
3370	pSession,
3371	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
3372	pSession->GipMapObjR3));
3373	#else
3374	LogFlow(("SUPR0GipUnmap: pSession=%p\n", pSession));
3375	#endif
3376	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3377
3378	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3379	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3380	#else
3381	RTSemFastMutexRequest(pDevExt->mtxGip);
3382	#endif
3383
3384	/*
3385	* Unmap anything?
3386	*/
3387	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
3388	{
3389	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
3390	AssertRC(rc);
3391	if (RT_SUCCESS(rc))
3392	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
3393	}
3394
3395	/*
3396	* Dereference global GIP.
3397	*/
3398	if (pSession->fGipReferenced && !rc)
3399	{
3400	pSession->fGipReferenced = 0;
3401	if ( pDevExt->cGipUsers > 0
3402	&& !--pDevExt->cGipUsers)
3403	{
3404	LogFlow(("SUPR0GipUnmap: Suspends GIP updating\n"));
3405	#ifndef DO_NOT_START_GIP
3406	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = VINF_SUCCESS;
3407	#endif
3408
3409	if (pDevExt->u32SystemTimerGranularityGrant)
3410	{
3411	int rc2 = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant);
3412	AssertRC(rc2);
3413	pDevExt->u32SystemTimerGranularityGrant = 0;
3414	}
3415	}
3416	}
3417
3418	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3419	RTSemMutexRelease(pDevExt->mtxGip);
3420	#else
3421	RTSemFastMutexRelease(pDevExt->mtxGip);
3422	#endif
3423
3424	return rc;
3425	}
3426
3427
3428	/**
3429	* Gets the GIP pointer.
3430	*
3431	* @returns Pointer to the GIP or NULL.
3432	*/
3433	SUPDECL(PSUPGLOBALINFOPAGE) SUPGetGIP(void)
3434	{
3435	return g_pSUPGlobalInfoPage;
3436	}
3437
3438
3439	/**
3440	* Register a component factory with the support driver.
3441	*
3442	* This is currently restricted to kernel sessions only.
3443	*
3444	* @returns VBox status code.
3445	* @retval VINF_SUCCESS on success.
3446	* @retval VERR_NO_MEMORY if we're out of memory.
3447	* @retval VERR_ALREADY_EXISTS if the factory has already been registered.
3448	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3449	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3450	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3451	*
3452	* @param pSession The SUPDRV session (must be a ring-0 session).
3453	* @param pFactory Pointer to the component factory registration structure.
3454	*
3455	* @remarks This interface is also available via SUPR0IdcComponentRegisterFactory.
3456	*/
3457	SUPR0DECL(int) SUPR0ComponentRegisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3458	{
3459	PSUPDRVFACTORYREG pNewReg;
3460	const char *psz;
3461	int rc;
3462
3463	/*
3464	* Validate parameters.
3465	*/
3466	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3467	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3468	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3469	AssertPtrReturn(pFactory->pfnQueryFactoryInterface, VERR_INVALID_POINTER);
3470	psz = RTStrEnd(pFactory->szName, sizeof(pFactory->szName));
3471	AssertReturn(psz, VERR_INVALID_PARAMETER);
3472
3473	/*
3474	* Allocate and initialize a new registration structure.
3475	*/
3476	pNewReg = (PSUPDRVFACTORYREG)RTMemAlloc(sizeof(SUPDRVFACTORYREG));
3477	if (pNewReg)
3478	{
3479	pNewReg->pNext = NULL;
3480	pNewReg->pFactory = pFactory;
3481	pNewReg->pSession = pSession;
3482	pNewReg->cchName = psz - &pFactory->szName[0];
3483
3484	/*
3485	* Add it to the tail of the list after checking for prior registration.
3486	*/
3487	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3488	if (RT_SUCCESS(rc))
3489	{
3490	PSUPDRVFACTORYREG pPrev = NULL;
3491	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3492	while (pCur && pCur->pFactory != pFactory)
3493	{
3494	pPrev = pCur;
3495	pCur = pCur->pNext;
3496	}
3497	if (!pCur)
3498	{
3499	if (pPrev)
3500	pPrev->pNext = pNewReg;
3501	else
3502	pSession->pDevExt->pComponentFactoryHead = pNewReg;
3503	rc = VINF_SUCCESS;
3504	}
3505	else
3506	rc = VERR_ALREADY_EXISTS;
3507
3508	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3509	}
3510
3511	if (RT_FAILURE(rc))
3512	RTMemFree(pNewReg);
3513	}
3514	else
3515	rc = VERR_NO_MEMORY;
3516	return rc;
3517	}
3518
3519
3520	/**
3521	* Deregister a component factory.
3522	*
3523	* @returns VBox status code.
3524	* @retval VINF_SUCCESS on success.
3525	* @retval VERR_NOT_FOUND if the factory wasn't registered.
3526	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3527	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3528	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3529	*
3530	* @param pSession The SUPDRV session (must be a ring-0 session).
3531	* @param pFactory Pointer to the component factory registration structure
3532	* previously passed SUPR0ComponentRegisterFactory().
3533	*
3534	* @remarks This interface is also available via SUPR0IdcComponentDeregisterFactory.
3535	*/
3536	SUPR0DECL(int) SUPR0ComponentDeregisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3537	{
3538	int rc;
3539
3540	/*
3541	* Validate parameters.
3542	*/
3543	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3544	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3545	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3546
3547	/*
3548	* Take the lock and look for the registration record.
3549	*/
3550	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3551	if (RT_SUCCESS(rc))
3552	{
3553	PSUPDRVFACTORYREG pPrev = NULL;
3554	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3555	while (pCur && pCur->pFactory != pFactory)
3556	{
3557	pPrev = pCur;
3558	pCur = pCur->pNext;
3559	}
3560	if (pCur)
3561	{
3562	if (!pPrev)
3563	pSession->pDevExt->pComponentFactoryHead = pCur->pNext;
3564	else
3565	pPrev->pNext = pCur->pNext;
3566
3567	pCur->pNext = NULL;
3568	pCur->pFactory = NULL;
3569	pCur->pSession = NULL;
3570	rc = VINF_SUCCESS;
3571	}
3572	else
3573	rc = VERR_NOT_FOUND;
3574
3575	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3576
3577	RTMemFree(pCur);
3578	}
3579	return rc;
3580	}
3581
3582
3583	/**
3584	* Queries a component factory.
3585	*
3586	* @returns VBox status code.
3587	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3588	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3589	* @retval VERR_SUPDRV_COMPONENT_NOT_FOUND if the component factory wasn't found.
3590	* @retval VERR_SUPDRV_INTERFACE_NOT_SUPPORTED if the interface wasn't supported.
3591	*
3592	* @param pSession The SUPDRV session.
3593	* @param pszName The name of the component factory.
3594	* @param pszInterfaceUuid The UUID of the factory interface (stringified).
3595	* @param ppvFactoryIf Where to store the factory interface.
3596	*/
3597	SUPR0DECL(int) SUPR0ComponentQueryFactory(PSUPDRVSESSION pSession, const char pszName, const char pszInterfaceUuid, void **ppvFactoryIf)
3598	{
3599	const char *pszEnd;
3600	size_t cchName;
3601	int rc;
3602
3603	/*
3604	* Validate parameters.
3605	*/
3606	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3607
3608	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
3609	pszEnd = RTStrEnd(pszName, RT_SIZEOFMEMB(SUPDRVFACTORY, szName));
3610	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3611	cchName = pszEnd - pszName;
3612
3613	AssertPtrReturn(pszInterfaceUuid, VERR_INVALID_POINTER);
3614	pszEnd = RTStrEnd(pszInterfaceUuid, RTUUID_STR_LENGTH);
3615	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3616
3617	AssertPtrReturn(ppvFactoryIf, VERR_INVALID_POINTER);
3618	*ppvFactoryIf = NULL;
3619
3620	/*
3621	* Take the lock and try all factories by this name.
3622	*/
3623	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3624	if (RT_SUCCESS(rc))
3625	{
3626	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3627	rc = VERR_SUPDRV_COMPONENT_NOT_FOUND;
3628	while (pCur)
3629	{
3630	if ( pCur->cchName == cchName
3631	&& !memcmp(pCur->pFactory->szName, pszName, cchName))
3632	{
3633	void *pvFactory = pCur->pFactory->pfnQueryFactoryInterface(pCur->pFactory, pSession, pszInterfaceUuid);
3634	if (pvFactory)
3635	{
3636	*ppvFactoryIf = pvFactory;
3637	rc = VINF_SUCCESS;
3638	break;
3639	}
3640	rc = VERR_SUPDRV_INTERFACE_NOT_SUPPORTED;
3641	}
3642
3643	/* next */
3644	pCur = pCur->pNext;
3645	}
3646
3647	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3648	}
3649	return rc;
3650	}
3651
3652
3653	/**
3654	* Adds a memory object to the session.
3655	*
3656	* @returns IPRT status code.
3657	* @param pMem Memory tracking structure containing the
3658	* information to track.
3659	* @param pSession The session.
3660	*/
3661	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
3662	{
3663	PSUPDRVBUNDLE pBundle;
3664	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3665
3666	/*
3667	* Find free entry and record the allocation.
3668	*/
3669	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3670	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3671	{
3672	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
3673	{
3674	unsigned i;
3675	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3676	{
3677	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
3678	{
3679	pBundle->cUsed++;
3680	pBundle->aMem[i] = *pMem;
3681	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3682	return VINF_SUCCESS;
3683	}
3684	}
3685	AssertFailed(); /* !!this can't be happening!!! */
3686	}
3687	}
3688	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3689
3690	/*
3691	* Need to allocate a new bundle.
3692	* Insert into the last entry in the bundle.
3693	*/
3694	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
3695	if (!pBundle)
3696	return VERR_NO_MEMORY;
3697
3698	/* take last entry. */
3699	pBundle->cUsed++;
3700	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
3701
3702	/* insert into list. */
3703	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3704	pBundle->pNext = pSession->Bundle.pNext;
3705	pSession->Bundle.pNext = pBundle;
3706	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3707
3708	return VINF_SUCCESS;
3709	}
3710
3711
3712	/**
3713	* Releases a memory object referenced by pointer and type.
3714	*
3715	* @returns IPRT status code.
3716	* @param pSession Session data.
3717	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
3718	* @param eType Memory type.
3719	*/
3720	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
3721	{
3722	PSUPDRVBUNDLE pBundle;
3723	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3724
3725	/*
3726	* Validate input.
3727	*/
3728	if (!uPtr)
3729	{
3730	Log(("Illegal address %p\n", (void *)uPtr));
3731	return VERR_INVALID_PARAMETER;
3732	}
3733
3734	/*
3735	* Search for the address.
3736	*/
3737	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3738	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3739	{
3740	if (pBundle->cUsed > 0)
3741	{
3742	unsigned i;
3743	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3744	{
3745	if ( pBundle->aMem[i].eType == eType
3746	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3747	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
3748	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3749	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
3750	)
3751	{
3752	/* Make a copy of it and release it outside the spinlock. */
3753	SUPDRVMEMREF Mem = pBundle->aMem[i];
3754	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
3755	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
3756	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
3757	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3758
3759	if (Mem.MapObjR3 != NIL_RTR0MEMOBJ)
3760	{
3761	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
3762	AssertRC(rc); /** @todo figure out how to handle this. */
3763	}
3764	if (Mem.MemObj != NIL_RTR0MEMOBJ)
3765	{
3766	int rc = RTR0MemObjFree(Mem.MemObj, true /* fFreeMappings */);
3767	AssertRC(rc); /** @todo figure out how to handle this. */
3768	}
3769	return VINF_SUCCESS;
3770	}
3771	}
3772	}
3773	}
3774	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3775	Log(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
3776	return VERR_INVALID_PARAMETER;
3777	}
3778
3779
3780	/**
3781	* Opens an image. If it's the first time it's opened the call must upload
3782	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
3783	*
3784	* This is the 1st step of the loading.
3785	*
3786	* @returns IPRT status code.
3787	* @param pDevExt Device globals.
3788	* @param pSession Session data.
3789	* @param pReq The open request.
3790	*/
3791	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
3792	{
3793	int rc;
3794	PSUPDRVLDRIMAGE pImage;
3795	void *pv;
3796	size_t cchName = strlen(pReq->u.In.szName); /* (caller checked < 32). */
3797	LogFlow(("supdrvIOCtl_LdrOpen: szName=%s cbImageWithTabs=%d\n", pReq->u.In.szName, pReq->u.In.cbImageWithTabs));
3798
3799	/*
3800	* Check if we got an instance of the image already.
3801	*/
3802	supdrvLdrLock(pDevExt);
3803	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
3804	{
3805	if ( pImage->szName[cchName] == '\0'
3806	&& !memcmp(pImage->szName, pReq->u.In.szName, cchName))
3807	{
3808	/** @todo check cbImageBits and cbImageWithTabs here, if they differs that indicates that the images are different. */
3809	pImage->cUsage++;
3810	pReq->u.Out.pvImageBase = pImage->pvImage;
3811	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
3812	pReq->u.Out.fNativeLoader = pImage->fNative;
3813	supdrvLdrAddUsage(pSession, pImage);
3814	supdrvLdrUnlock(pDevExt);
3815	return VINF_SUCCESS;
3816	}
3817	}
3818	/* (not found - add it!) */
3819
3820	/*
3821	* Allocate memory.
3822	*/
3823	pv = RTMemAlloc(RT_OFFSETOF(SUPDRVLDRIMAGE, szName[cchName + 1]));
3824	if (!pv)
3825	{
3826	supdrvLdrUnlock(pDevExt);
3827	Log(("supdrvIOCtl_LdrOpen: RTMemAlloc() failed\n"));
3828	return /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_2;
3829	}
3830
3831	/*
3832	* Setup and link in the LDR stuff.
3833	*/
3834	pImage = (PSUPDRVLDRIMAGE)pv;
3835	pImage->pvImage = NULL;
3836	pImage->pvImageAlloc = NULL;
3837	pImage->cbImageWithTabs = pReq->u.In.cbImageWithTabs;
3838	pImage->cbImageBits = pReq->u.In.cbImageBits;
3839	pImage->cSymbols = 0;
3840	pImage->paSymbols = NULL;
3841	pImage->pachStrTab = NULL;
3842	pImage->cbStrTab = 0;
3843	pImage->pfnModuleInit = NULL;
3844	pImage->pfnModuleTerm = NULL;
3845	pImage->pfnServiceReqHandler = NULL;
3846	pImage->uState = SUP_IOCTL_LDR_OPEN;
3847	pImage->cUsage = 1;
3848	memcpy(pImage->szName, pReq->u.In.szName, cchName + 1);
3849
3850	/*
3851	* Try load it using the native loader, if that isn't supported, fall back
3852	* on the older method.
3853	*/
3854	pImage->fNative = true;
3855	rc = supdrvOSLdrOpen(pDevExt, pImage, pReq->u.In.szFilename);
3856	if (rc == VERR_NOT_SUPPORTED)
3857	{
3858	pImage->pvImageAlloc = RTMemExecAlloc(pImage->cbImageBits + 31);
3859	pImage->pvImage = RT_ALIGN_P(pImage->pvImageAlloc, 32);
3860	pImage->fNative = false;
3861	rc = pImage->pvImageAlloc ? VINF_SUCCESS : VERR_NO_EXEC_MEMORY;
3862	}
3863	if (RT_FAILURE(rc))
3864	{
3865	supdrvLdrUnlock(pDevExt);
3866	RTMemFree(pImage);
3867	Log(("supdrvIOCtl_LdrOpen(%s): failed - %Rrc\n", pReq->u.In.szName, rc));
3868	return rc;
3869	}
3870	Assert(VALID_PTR(pImage->pvImage) \|\| RT_FAILURE(rc));
3871
3872	/*
3873	* Link it.
3874	*/
3875	pImage->pNext = pDevExt->pLdrImages;
3876	pDevExt->pLdrImages = pImage;
3877
3878	supdrvLdrAddUsage(pSession, pImage);
3879
3880	pReq->u.Out.pvImageBase = pImage->pvImage;
3881	pReq->u.Out.fNeedsLoading = true;
3882	pReq->u.Out.fNativeLoader = pImage->fNative;
3883	supdrvLdrUnlock(pDevExt);
3884
3885	#if defined(RT_OS_WINDOWS) && defined(DEBUG)
3886	SUPR0Printf("VBoxDrv: windbg> .reload /f %s=%#p\n", pImage->szName, pImage->pvImage);
3887	#endif
3888	return VINF_SUCCESS;
3889	}
3890
3891
3892	/**
3893	* Worker that validates a pointer to an image entrypoint.
3894	*
3895	* @returns IPRT status code.
3896	* @param pDevExt The device globals.
3897	* @param pImage The loader image.
3898	* @param pv The pointer into the image.
3899	* @param fMayBeNull Whether it may be NULL.
3900	* @param pszWhat What is this entrypoint? (for logging)
3901	* @param pbImageBits The image bits prepared by ring-3.
3902	*
3903	* @remarks Will leave the lock on failure.
3904	*/
3905	static int supdrvLdrValidatePointer(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage, void *pv,
3906	bool fMayBeNull, const uint8_t pbImageBits, const char pszWhat)
3907	{
3908	if (!fMayBeNull \|\| pv)
3909	{
3910	if ((uintptr_t)pv - (uintptr_t)pImage->pvImage >= pImage->cbImageBits)
3911	{
3912	supdrvLdrUnlock(pDevExt);
3913	Log(("Out of range (%p LB %#x): %s=%p\n", pImage->pvImage, pImage->cbImageBits, pszWhat, pv));
3914	return VERR_INVALID_PARAMETER;
3915	}
3916
3917	if (pImage->fNative)
3918	{
3919	int rc = supdrvOSLdrValidatePointer(pDevExt, pImage, pv, pbImageBits);
3920	if (RT_FAILURE(rc))
3921	{
3922	supdrvLdrUnlock(pDevExt);
3923	Log(("Bad entry point address: %s=%p (rc=%Rrc)\n", pszWhat, pv, rc));
3924	return rc;
3925	}
3926	}
3927	}
3928	return VINF_SUCCESS;
3929	}
3930
3931
3932	/**
3933	* Loads the image bits.
3934	*
3935	* This is the 2nd step of the loading.
3936	*
3937	* @returns IPRT status code.
3938	* @param pDevExt Device globals.
3939	* @param pSession Session data.
3940	* @param pReq The request.
3941	*/
3942	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
3943	{
3944	PSUPDRVLDRUSAGE pUsage;
3945	PSUPDRVLDRIMAGE pImage;
3946	int rc;
3947	LogFlow(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImageWithBits=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImageWithTabs));
3948
3949	/*
3950	* Find the ldr image.
3951	*/
3952	supdrvLdrLock(pDevExt);
3953	pUsage = pSession->pLdrUsage;
3954	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3955	pUsage = pUsage->pNext;
3956	if (!pUsage)
3957	{
3958	supdrvLdrUnlock(pDevExt);
3959	Log(("SUP_IOCTL_LDR_LOAD: couldn't find image!\n"));
3960	return VERR_INVALID_HANDLE;
3961	}
3962	pImage = pUsage->pImage;
3963
3964	/*
3965	* Validate input.
3966	*/
3967	if ( pImage->cbImageWithTabs != pReq->u.In.cbImageWithTabs
3968	\|\| pImage->cbImageBits != pReq->u.In.cbImageBits)
3969	{
3970	supdrvLdrUnlock(pDevExt);
3971	Log(("SUP_IOCTL_LDR_LOAD: image size mismatch!! %d(prep) != %d(load) or %d != %d\n",
3972	pImage->cbImageWithTabs, pReq->u.In.cbImageWithTabs, pImage->cbImageBits, pReq->u.In.cbImageBits));
3973	return VERR_INVALID_HANDLE;
3974	}
3975
3976	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
3977	{
3978	unsigned uState = pImage->uState;
3979	supdrvLdrUnlock(pDevExt);
3980	if (uState != SUP_IOCTL_LDR_LOAD)
3981	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
3982	return VERR_ALREADY_LOADED;
3983	}
3984
3985	switch (pReq->u.In.eEPType)
3986	{
3987	case SUPLDRLOADEP_NOTHING:
3988	break;
3989
3990	case SUPLDRLOADEP_VMMR0:
3991	rc = supdrvLdrValidatePointer( pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0, false, pReq->u.In.abImage, "pvVMMR0");
3992	if (RT_SUCCESS(rc))
3993	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt, false, pReq->u.In.abImage, "pvVMMR0EntryInt");
3994	if (RT_SUCCESS(rc))
3995	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, false, pReq->u.In.abImage, "pvVMMR0EntryFast");
3996	if (RT_SUCCESS(rc))
3997	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx, false, pReq->u.In.abImage, "pvVMMR0EntryEx");
3998	if (RT_FAILURE(rc))
3999	return rc;
4000	break;
4001
4002	case SUPLDRLOADEP_SERVICE:
4003	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.Service.pfnServiceReq, false, pReq->u.In.abImage, "pfnServiceReq");
4004	if (RT_FAILURE(rc))
4005	return rc;
4006	if ( pReq->u.In.EP.Service.apvReserved[0] != NIL_RTR0PTR
4007	\|\| pReq->u.In.EP.Service.apvReserved[1] != NIL_RTR0PTR
4008	\|\| pReq->u.In.EP.Service.apvReserved[2] != NIL_RTR0PTR)
4009	{
4010	supdrvLdrUnlock(pDevExt);
4011	Log(("Out of range (%p LB %#x): apvReserved={%p,%p,%p} MBZ!\n",
4012	pImage->pvImage, pReq->u.In.cbImageWithTabs,
4013	pReq->u.In.EP.Service.apvReserved[0],
4014	pReq->u.In.EP.Service.apvReserved[1],
4015	pReq->u.In.EP.Service.apvReserved[2]));
4016	return VERR_INVALID_PARAMETER;
4017	}
4018	break;
4019
4020	default:
4021	supdrvLdrUnlock(pDevExt);
4022	Log(("Invalid eEPType=%d\n", pReq->u.In.eEPType));
4023	return VERR_INVALID_PARAMETER;
4024	}
4025
4026	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleInit, true, pReq->u.In.abImage, "pfnModuleInit");
4027	if (RT_FAILURE(rc))
4028	return rc;
4029	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleTerm, true, pReq->u.In.abImage, "pfnModuleTerm");
4030	if (RT_FAILURE(rc))
4031	return rc;
4032
4033	/*
4034	* Allocate and copy the tables.
4035	* (No need to do try/except as this is a buffered request.)
4036	*/
4037	pImage->cbStrTab = pReq->u.In.cbStrTab;
4038	if (pImage->cbStrTab)
4039	{
4040	pImage->pachStrTab = (char *)RTMemAlloc(pImage->cbStrTab);
4041	if (pImage->pachStrTab)
4042	memcpy(pImage->pachStrTab, &pReq->u.In.abImage[pReq->u.In.offStrTab], pImage->cbStrTab);
4043	else
4044	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_3;
4045	}
4046
4047	pImage->cSymbols = pReq->u.In.cSymbols;
4048	if (RT_SUCCESS(rc) && pImage->cSymbols)
4049	{
4050	size_t cbSymbols = pImage->cSymbols * sizeof(SUPLDRSYM);
4051	pImage->paSymbols = (PSUPLDRSYM)RTMemAlloc(cbSymbols);
4052	if (pImage->paSymbols)
4053	memcpy(pImage->paSymbols, &pReq->u.In.abImage[pReq->u.In.offSymbols], cbSymbols);
4054	else
4055	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_4;
4056	}
4057
4058	/*
4059	* Copy the bits / complete native loading.
4060	*/
4061	if (RT_SUCCESS(rc))
4062	{
4063	pImage->uState = SUP_IOCTL_LDR_LOAD;
4064	pImage->pfnModuleInit = pReq->u.In.pfnModuleInit;
4065	pImage->pfnModuleTerm = pReq->u.In.pfnModuleTerm;
4066
4067	if (pImage->fNative)
4068	rc = supdrvOSLdrLoad(pDevExt, pImage, pReq->u.In.abImage, pReq);
4069	else
4070	memcpy(pImage->pvImage, &pReq->u.In.abImage[0], pImage->cbImageBits);
4071	}
4072
4073	/*
4074	* Update any entry points.
4075	*/
4076	if (RT_SUCCESS(rc))
4077	{
4078	switch (pReq->u.In.eEPType)
4079	{
4080	default:
4081	case SUPLDRLOADEP_NOTHING:
4082	rc = VINF_SUCCESS;
4083	break;
4084	case SUPLDRLOADEP_VMMR0:
4085	rc = supdrvLdrSetVMMR0EPs(pDevExt, pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
4086	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
4087	break;
4088	case SUPLDRLOADEP_SERVICE:
4089	pImage->pfnServiceReqHandler = pReq->u.In.EP.Service.pfnServiceReq;
4090	rc = VINF_SUCCESS;
4091	break;
4092	}
4093	}
4094
4095	/*
4096	* On success call the module initialization.
4097	*/
4098	LogFlow(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
4099	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
4100	{
4101	Log(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
4102	rc = pImage->pfnModuleInit();
4103	if (rc && pDevExt->pvVMMR0 == pImage->pvImage)
4104	supdrvLdrUnsetVMMR0EPs(pDevExt);
4105	}
4106
4107	if (RT_FAILURE(rc))
4108	{
4109	pImage->uState = SUP_IOCTL_LDR_OPEN;
4110	pImage->pfnModuleInit = NULL;
4111	pImage->pfnModuleTerm = NULL;
4112	pImage->pfnServiceReqHandler= NULL;
4113	pImage->cbStrTab = 0;
4114	RTMemFree(pImage->pachStrTab);
4115	pImage->pachStrTab = NULL;
4116	RTMemFree(pImage->paSymbols);
4117	pImage->paSymbols = NULL;
4118	pImage->cSymbols = 0;
4119	}
4120
4121	supdrvLdrUnlock(pDevExt);
4122	return rc;
4123	}
4124
4125
4126	/**
4127	* Frees a previously loaded (prep'ed) image.
4128	*
4129	* @returns IPRT status code.
4130	* @param pDevExt Device globals.
4131	* @param pSession Session data.
4132	* @param pReq The request.
4133	*/
4134	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
4135	{
4136	int rc;
4137	PSUPDRVLDRUSAGE pUsagePrev;
4138	PSUPDRVLDRUSAGE pUsage;
4139	PSUPDRVLDRIMAGE pImage;
4140	LogFlow(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
4141
4142	/*
4143	* Find the ldr image.
4144	*/
4145	supdrvLdrLock(pDevExt);
4146	pUsagePrev = NULL;
4147	pUsage = pSession->pLdrUsage;
4148	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4149	{
4150	pUsagePrev = pUsage;
4151	pUsage = pUsage->pNext;
4152	}
4153	if (!pUsage)
4154	{
4155	supdrvLdrUnlock(pDevExt);
4156	Log(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
4157	return VERR_INVALID_HANDLE;
4158	}
4159
4160	/*
4161	* Check if we can remove anything.
4162	*/
4163	rc = VINF_SUCCESS;
4164	pImage = pUsage->pImage;
4165	if (pImage->cUsage <= 1 \|\| pUsage->cUsage <= 1)
4166	{
4167	/*
4168	* Check if there are any objects with destructors in the image, if
4169	* so leave it for the session cleanup routine so we get a chance to
4170	* clean things up in the right order and not leave them all dangling.
4171	*/
4172	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
4173	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
4174	if (pImage->cUsage <= 1)
4175	{
4176	PSUPDRVOBJ pObj;
4177	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4178	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4179	{
4180	rc = VERR_DANGLING_OBJECTS;
4181	break;
4182	}
4183	}
4184	else
4185	{
4186	PSUPDRVUSAGE pGenUsage;
4187	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
4188	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4189	{
4190	rc = VERR_DANGLING_OBJECTS;
4191	break;
4192	}
4193	}
4194	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
4195	if (rc == VINF_SUCCESS)
4196	{
4197	/* unlink it */
4198	if (pUsagePrev)
4199	pUsagePrev->pNext = pUsage->pNext;
4200	else
4201	pSession->pLdrUsage = pUsage->pNext;
4202
4203	/* free it */
4204	pUsage->pImage = NULL;
4205	pUsage->pNext = NULL;
4206	RTMemFree(pUsage);
4207
4208	/*
4209	* Dereference the image.
4210	*/
4211	if (pImage->cUsage <= 1)
4212	supdrvLdrFree(pDevExt, pImage);
4213	else
4214	pImage->cUsage--;
4215	}
4216	else
4217	{
4218	Log(("supdrvIOCtl_LdrFree: Dangling objects in %p/%s!\n", pImage->pvImage, pImage->szName));
4219	rc = VINF_SUCCESS; /** @todo BRANCH-2.1: remove this after branching. */
4220	}
4221	}
4222	else
4223	{
4224	/*
4225	* Dereference both image and usage.
4226	*/
4227	pImage->cUsage--;
4228	pUsage->cUsage--;
4229	}
4230
4231	supdrvLdrUnlock(pDevExt);
4232	return rc;
4233	}
4234
4235
4236	/**
4237	* Gets the address of a symbol in an open image.
4238	*
4239	* @returns IPRT status code.
4240	* @param pDevExt Device globals.
4241	* @param pSession Session data.
4242	* @param pReq The request buffer.
4243	*/
4244	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
4245	{
4246	PSUPDRVLDRIMAGE pImage;
4247	PSUPDRVLDRUSAGE pUsage;
4248	uint32_t i;
4249	PSUPLDRSYM paSyms;
4250	const char *pchStrings;
4251	const size_t cbSymbol = strlen(pReq->u.In.szSymbol) + 1;
4252	void *pvSymbol = NULL;
4253	int rc = VERR_GENERAL_FAILURE;
4254	Log3(("supdrvIOCtl_LdrGetSymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
4255
4256	/*
4257	* Find the ldr image.
4258	*/
4259	supdrvLdrLock(pDevExt);
4260	pUsage = pSession->pLdrUsage;
4261	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4262	pUsage = pUsage->pNext;
4263	if (!pUsage)
4264	{
4265	supdrvLdrUnlock(pDevExt);
4266	Log(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
4267	return VERR_INVALID_HANDLE;
4268	}
4269	pImage = pUsage->pImage;
4270	if (pImage->uState != SUP_IOCTL_LDR_LOAD)
4271	{
4272	unsigned uState = pImage->uState;
4273	supdrvLdrUnlock(pDevExt);
4274	Log(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", uState, uState)); NOREF(uState);
4275	return VERR_ALREADY_LOADED;
4276	}
4277
4278	/*
4279	* Search the symbol strings.
4280	*
4281	* Note! The int32_t is for native loading on solaris where the data
4282	* and text segments are in very different places.
4283	*/
4284	pchStrings = pImage->pachStrTab;
4285	paSyms = pImage->paSymbols;
4286	for (i = 0; i < pImage->cSymbols; i++)
4287	{
4288	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4289	&& !memcmp(pchStrings + paSyms[i].offName, pReq->u.In.szSymbol, cbSymbol))
4290	{
4291	pvSymbol = (uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol;
4292	rc = VINF_SUCCESS;
4293	break;
4294	}
4295	}
4296	supdrvLdrUnlock(pDevExt);
4297	pReq->u.Out.pvSymbol = pvSymbol;
4298	return rc;
4299	}
4300
4301
4302	/**
4303	* Gets the address of a symbol in an open image or the support driver.
4304	*
4305	* @returns VINF_SUCCESS on success.
4306	* @returns
4307	* @param pDevExt Device globals.
4308	* @param pSession Session data.
4309	* @param pReq The request buffer.
4310	*/
4311	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq)
4312	{
4313	int rc = VINF_SUCCESS;
4314	const char *pszSymbol = pReq->u.In.pszSymbol;
4315	const char *pszModule = pReq->u.In.pszModule;
4316	size_t cbSymbol;
4317	char const *pszEnd;
4318	uint32_t i;
4319
4320	/*
4321	* Input validation.
4322	*/
4323	AssertPtrReturn(pszSymbol, VERR_INVALID_POINTER);
4324	pszEnd = RTStrEnd(pszSymbol, 512);
4325	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4326	cbSymbol = pszEnd - pszSymbol + 1;
4327
4328	if (pszModule)
4329	{
4330	AssertPtrReturn(pszModule, VERR_INVALID_POINTER);
4331	pszEnd = RTStrEnd(pszModule, 64);
4332	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4333	}
4334	Log3(("supdrvIDC_LdrGetSymbol: pszModule=%p:{%s} pszSymbol=%p:{%s}\n", pszModule, pszModule, pszSymbol, pszSymbol));
4335
4336
4337	if ( !pszModule
4338	\|\| !strcmp(pszModule, "SupDrv"))
4339	{
4340	/*
4341	* Search the support driver export table.
4342	*/
4343	for (i = 0; i < RT_ELEMENTS(g_aFunctions); i++)
4344	if (!strcmp(g_aFunctions[i].szName, pszSymbol))
4345	{
4346	pReq->u.Out.pfnSymbol = g_aFunctions[i].pfn;
4347	break;
4348	}
4349	}
4350	else
4351	{
4352	/*
4353	* Find the loader image.
4354	*/
4355	PSUPDRVLDRIMAGE pImage;
4356
4357	supdrvLdrLock(pDevExt);
4358
4359	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
4360	if (!strcmp(pImage->szName, pszModule))
4361	break;
4362	if (pImage && pImage->uState == SUP_IOCTL_LDR_LOAD)
4363	{
4364	/*
4365	* Search the symbol strings.
4366	*/
4367	const char *pchStrings = pImage->pachStrTab;
4368	PCSUPLDRSYM paSyms = pImage->paSymbols;
4369	for (i = 0; i < pImage->cSymbols; i++)
4370	{
4371	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4372	&& !memcmp(pchStrings + paSyms[i].offName, pszSymbol, cbSymbol))
4373	{
4374	/*
4375	* Found it! Calc the symbol address and add a reference to the module.
4376	*/
4377	pReq->u.Out.pfnSymbol = (PFNRT)((uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol);
4378	rc = supdrvLdrAddUsage(pSession, pImage);
4379	break;
4380	}
4381	}
4382	}
4383	else
4384	rc = pImage ? VERR_WRONG_ORDER : VERR_MODULE_NOT_FOUND;
4385
4386	supdrvLdrUnlock(pDevExt);
4387	}
4388	return rc;
4389	}
4390
4391
4392	/**
4393	* Updates the VMMR0 entry point pointers.
4394	*
4395	* @returns IPRT status code.
4396	* @param pDevExt Device globals.
4397	* @param pSession Session data.
4398	* @param pVMMR0 VMMR0 image handle.
4399	* @param pvVMMR0EntryInt VMMR0EntryInt address.
4400	* @param pvVMMR0EntryFast VMMR0EntryFast address.
4401	* @param pvVMMR0EntryEx VMMR0EntryEx address.
4402	* @remark Caller must own the loader mutex.
4403	*/
4404	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx)
4405	{
4406	int rc = VINF_SUCCESS;
4407	LogFlow(("supdrvLdrSetR0EP pvVMMR0=%p pvVMMR0EntryInt=%p\n", pvVMMR0, pvVMMR0EntryInt));
4408
4409
4410	/*
4411	* Check if not yet set.
4412	*/
4413	if (!pDevExt->pvVMMR0)
4414	{
4415	pDevExt->pvVMMR0 = pvVMMR0;
4416	pDevExt->pfnVMMR0EntryInt = pvVMMR0EntryInt;
4417	pDevExt->pfnVMMR0EntryFast = pvVMMR0EntryFast;
4418	pDevExt->pfnVMMR0EntryEx = pvVMMR0EntryEx;
4419	}
4420	else
4421	{
4422	/*
4423	* Return failure or success depending on whether the values match or not.
4424	*/
4425	if ( pDevExt->pvVMMR0 != pvVMMR0
4426	\|\| (void *)pDevExt->pfnVMMR0EntryInt != pvVMMR0EntryInt
4427	\|\| (void *)pDevExt->pfnVMMR0EntryFast != pvVMMR0EntryFast
4428	\|\| (void *)pDevExt->pfnVMMR0EntryEx != pvVMMR0EntryEx)
4429	{
4430	AssertMsgFailed(("SUP_IOCTL_LDR_SETR0EP: Already set pointing to a different module!\n"));
4431	rc = VERR_INVALID_PARAMETER;
4432	}
4433	}
4434	return rc;
4435	}
4436
4437
4438	/**
4439	* Unsets the VMMR0 entry point installed by supdrvLdrSetR0EP.
4440	*
4441	* @param pDevExt Device globals.
4442	*/
4443	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt)
4444	{
4445	pDevExt->pvVMMR0 = NULL;
4446	pDevExt->pfnVMMR0EntryInt = NULL;
4447	pDevExt->pfnVMMR0EntryFast = NULL;
4448	pDevExt->pfnVMMR0EntryEx = NULL;
4449	}
4450
4451
4452	/**
4453	* Adds a usage reference in the specified session of an image.
4454	*
4455	* Called while owning the loader semaphore.
4456	*
4457	* @returns VINF_SUCCESS on success and VERR_NO_MEMORY on failure.
4458	* @param pSession Session in question.
4459	* @param pImage Image which the session is using.
4460	*/
4461	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage)
4462	{
4463	PSUPDRVLDRUSAGE pUsage;
4464	LogFlow(("supdrvLdrAddUsage: pImage=%p\n", pImage));
4465
4466	/*
4467	* Referenced it already?
4468	*/
4469	pUsage = pSession->pLdrUsage;
4470	while (pUsage)
4471	{
4472	if (pUsage->pImage == pImage)
4473	{
4474	pUsage->cUsage++;
4475	return VINF_SUCCESS;
4476	}
4477	pUsage = pUsage->pNext;
4478	}
4479
4480	/*
4481	* Allocate new usage record.
4482	*/
4483	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
4484	AssertReturn(pUsage, /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_5);
4485	pUsage->cUsage = 1;
4486	pUsage->pImage = pImage;
4487	pUsage->pNext = pSession->pLdrUsage;
4488	pSession->pLdrUsage = pUsage;
4489	return VINF_SUCCESS;
4490	}
4491
4492
4493	/**
4494	* Frees a load image.
4495	*
4496	* @param pDevExt Pointer to device extension.
4497	* @param pImage Pointer to the image we're gonna free.
4498	* This image must exit!
4499	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
4500	*/
4501	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
4502	{
4503	PSUPDRVLDRIMAGE pImagePrev;
4504	LogFlow(("supdrvLdrFree: pImage=%p\n", pImage));
4505
4506	/* find it - arg. should've used doubly linked list. */
4507	Assert(pDevExt->pLdrImages);
4508	pImagePrev = NULL;
4509	if (pDevExt->pLdrImages != pImage)
4510	{
4511	pImagePrev = pDevExt->pLdrImages;
4512	while (pImagePrev->pNext != pImage)
4513	pImagePrev = pImagePrev->pNext;
4514	Assert(pImagePrev->pNext == pImage);
4515	}
4516
4517	/* unlink */
4518	if (pImagePrev)
4519	pImagePrev->pNext = pImage->pNext;
4520	else
4521	pDevExt->pLdrImages = pImage->pNext;
4522
4523	/* check if this is VMMR0.r0 unset its entry point pointers. */
4524	if (pDevExt->pvVMMR0 == pImage->pvImage)
4525	supdrvLdrUnsetVMMR0EPs(pDevExt);
4526
4527	/* check for objects with destructors in this image. (Shouldn't happen.) */
4528	if (pDevExt->pObjs)
4529	{
4530	unsigned cObjs = 0;
4531	PSUPDRVOBJ pObj;
4532	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
4533	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
4534	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4535	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4536	{
4537	pObj->pfnDestructor = NULL;
4538	cObjs++;
4539	}
4540	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
4541	if (cObjs)
4542	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
4543	}
4544
4545	/* call termination function if fully loaded. */
4546	if ( pImage->pfnModuleTerm
4547	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
4548	{
4549	LogFlow(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
4550	pImage->pfnModuleTerm();
4551	}
4552
4553	/* do native unload if appropriate. */
4554	if (pImage->fNative)
4555	supdrvOSLdrUnload(pDevExt, pImage);
4556
4557	/* free the image */
4558	pImage->cUsage = 0;
4559	pImage->pNext = 0;
4560	pImage->uState = SUP_IOCTL_LDR_FREE;
4561	RTMemExecFree(pImage->pvImageAlloc, pImage->cbImageBits + 31);
4562	pImage->pvImageAlloc = NULL;
4563	RTMemFree(pImage->pachStrTab);
4564	pImage->pachStrTab = NULL;
4565	RTMemFree(pImage->paSymbols);
4566	pImage->paSymbols = NULL;
4567	RTMemFree(pImage);
4568	}
4569
4570
4571	/**
4572	* Acquires the loader lock.
4573	*
4574	* @returns IPRT status code.
4575	* @param pDevExt The device extension.
4576	*/
4577	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt)
4578	{
4579	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4580	int rc = RTSemMutexRequest(pDevExt->mtxLdr, RT_INDEFINITE_WAIT);
4581	#else
4582	int rc = RTSemFastMutexRequest(pDevExt->mtxLdr);
4583	#endif
4584	AssertRC(rc);
4585	return rc;
4586	}
4587
4588
4589	/**
4590	* Releases the loader lock.
4591	*
4592	* @returns IPRT status code.
4593	* @param pDevExt The device extension.
4594	*/
4595	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt)
4596	{
4597	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4598	return RTSemMutexRelease(pDevExt->mtxLdr);
4599	#else
4600	return RTSemFastMutexRelease(pDevExt->mtxLdr);
4601	#endif
4602	}
4603
4604
4605	/**
4606	* Implements the service call request.
4607	*
4608	* @returns VBox status code.
4609	* @param pDevExt The device extension.
4610	* @param pSession The calling session.
4611	* @param pReq The request packet, valid.
4612	*/
4613	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq)
4614	{
4615	#if !defined(RT_OS_WINDOWS) \|\| defined(DEBUG)
4616	int rc;
4617
4618	/*
4619	* Find the module first in the module referenced by the calling session.
4620	*/
4621	rc = supdrvLdrLock(pDevExt);
4622	if (RT_SUCCESS(rc))
4623	{
4624	PFNSUPR0SERVICEREQHANDLER pfnServiceReqHandler = NULL;
4625	PSUPDRVLDRUSAGE pUsage;
4626
4627	for (pUsage = pSession->pLdrUsage; pUsage; pUsage = pUsage->pNext)
4628	if ( pUsage->pImage->pfnServiceReqHandler
4629	&& !strcmp(pUsage->pImage->szName, pReq->u.In.szName))
4630	{
4631	pfnServiceReqHandler = pUsage->pImage->pfnServiceReqHandler;
4632	break;
4633	}
4634	supdrvLdrUnlock(pDevExt);
4635
4636	if (pfnServiceReqHandler)
4637	{
4638	/*
4639	* Call it.
4640	*/
4641	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
4642	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, NULL);
4643	else
4644	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0]);
4645	}
4646	else
4647	rc = VERR_SUPDRV_SERVICE_NOT_FOUND;
4648	}
4649
4650	/* log it */
4651	if ( RT_FAILURE(rc)
4652	&& rc != VERR_INTERRUPTED
4653	&& rc != VERR_TIMEOUT)
4654	Log(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4655	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4656	else
4657	Log4(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4658	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4659	return rc;
4660	#else /* RT_OS_WINDOWS && !DEBUG */
4661	return VERR_NOT_IMPLEMENTED;
4662	#endif /* RT_OS_WINDOWS && !DEBUG */
4663	}
4664
4665
4666	/**
4667	* Implements the logger settings request.
4668	*
4669	* @returns VBox status code.
4670	* @param pDevExt The device extension.
4671	* @param pSession The caller's session.
4672	* @param pReq The request.
4673	*/
4674	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq)
4675	{
4676	const char *pszGroup = &pReq->u.In.szStrings[pReq->u.In.offGroups];
4677	const char *pszFlags = &pReq->u.In.szStrings[pReq->u.In.offFlags];
4678	const char *pszDest = &pReq->u.In.szStrings[pReq->u.In.offDestination];
4679	PRTLOGGER pLogger = NULL;
4680	int rc;
4681
4682	/*
4683	* Some further validation.
4684	*/
4685	switch (pReq->u.In.fWhat)
4686	{
4687	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4688	case SUPLOGGERSETTINGS_WHAT_CREATE:
4689	break;
4690
4691	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4692	if (pszGroup \|\| pszFlags \|\| *pszDest)
4693	return VERR_INVALID_PARAMETER;
4694	if (pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_RELEASE)
4695	return VERR_ACCESS_DENIED;
4696	break;
4697
4698	default:
4699	return VERR_INTERNAL_ERROR;
4700	}
4701
4702	/*
4703	* Get the logger.
4704	*/
4705	switch (pReq->u.In.fWhich)
4706	{
4707	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4708	pLogger = RTLogGetDefaultInstance();
4709	break;
4710
4711	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4712	pLogger = RTLogRelDefaultInstance();
4713	break;
4714
4715	default:
4716	return VERR_INTERNAL_ERROR;
4717	}
4718
4719	/*
4720	* Do the job.
4721	*/
4722	switch (pReq->u.In.fWhat)
4723	{
4724	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4725	if (pLogger)
4726	{
4727	rc = RTLogFlags(pLogger, pszFlags);
4728	if (RT_SUCCESS(rc))
4729	rc = RTLogGroupSettings(pLogger, pszGroup);
4730	NOREF(pszDest);
4731	}
4732	else
4733	rc = VERR_NOT_FOUND;
4734	break;
4735
4736	case SUPLOGGERSETTINGS_WHAT_CREATE:
4737	{
4738	if (pLogger)
4739	rc = VERR_ALREADY_EXISTS;
4740	else
4741	{
4742	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
4743
4744	rc = RTLogCreate(&pLogger,
4745	0 /* fFlags */,
4746	pszGroup,
4747	pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_DEBUG
4748	? "VBOX_LOG"
4749	: "VBOX_RELEASE_LOG",
4750	RT_ELEMENTS(s_apszGroups),
4751	s_apszGroups,
4752	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER,
4753	NULL);
4754	if (RT_SUCCESS(rc))
4755	{
4756	rc = RTLogFlags(pLogger, pszFlags);
4757	NOREF(pszDest);
4758	if (RT_SUCCESS(rc))
4759	{
4760	switch (pReq->u.In.fWhich)
4761	{
4762	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4763	pLogger = RTLogSetDefaultInstance(pLogger);
4764	break;
4765	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4766	pLogger = RTLogRelSetDefaultInstance(pLogger);
4767	break;
4768	}
4769	}
4770	RTLogDestroy(pLogger);
4771	}
4772	}
4773	break;
4774	}
4775
4776	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4777	switch (pReq->u.In.fWhich)
4778	{
4779	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4780	pLogger = RTLogSetDefaultInstance(NULL);
4781	break;
4782	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4783	pLogger = RTLogRelSetDefaultInstance(NULL);
4784	break;
4785	}
4786	rc = RTLogDestroy(pLogger);
4787	break;
4788
4789	default:
4790	{
4791	rc = VERR_INTERNAL_ERROR;
4792	break;
4793	}
4794	}
4795
4796	return rc;
4797	}
4798
4799
4800	/**
4801	* Creates the GIP.
4802	*
4803	* @returns VBox status code.
4804	* @param pDevExt Instance data. GIP stuff may be updated.
4805	*/
4806	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
4807	{
4808	PSUPGLOBALINFOPAGE pGip;
4809	RTHCPHYS HCPhysGip;
4810	uint32_t u32SystemResolution;
4811	uint32_t u32Interval;
4812	unsigned cCpus;
4813	int rc;
4814
4815
4816	LogFlow(("supdrvGipCreate:\n"));
4817
4818	/* assert order */
4819	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
4820	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
4821	Assert(!pDevExt->pGipTimer);
4822
4823	/*
4824	* Check the CPU count.
4825	*/
4826	cCpus = RTMpGetArraySize();
4827	if ( cCpus > RTCPUSET_MAX_CPUS
4828	\|\| cCpus > 256 /ApicId is used for the mappings/)
4829	{
4830	SUPR0Printf("VBoxDrv: Too many CPUs (%u) for the GIP (max %u)\n", cCpus, RT_MIN(RTCPUSET_MAX_CPUS, 256));
4831	return VERR_TOO_MANY_CPUS;
4832	}
4833
4834	/*
4835	* Allocate a contiguous set of pages with a default kernel mapping.
4836	*/
4837	rc = RTR0MemObjAllocCont(&pDevExt->GipMemObj, RT_UOFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), false /fExecutable/);
4838	if (RT_FAILURE(rc))
4839	{
4840	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
4841	return rc;
4842	}
4843	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
4844	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
4845
4846	/*
4847	* Find a reasonable update interval and initialize the structure.
4848	*/
4849	u32Interval = u32SystemResolution = RTTimerGetSystemGranularity();
4850	while (u32Interval < 10000000 /* 10 ms */)
4851	u32Interval += u32SystemResolution;
4852
4853	supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), 1000000000 / u32Interval /=Hz/, cCpus);
4854
4855	/*
4856	* Create the timer.
4857	* If CPU_ALL isn't supported we'll have to fall back to synchronous mode.
4858	*/
4859	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
4860	{
4861	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, RTTIMER_FLAGS_CPU_ALL, supdrvGipAsyncTimer, pDevExt);
4862	if (rc == VERR_NOT_SUPPORTED)
4863	{
4864	OSDBGPRINT(("supdrvGipCreate: omni timer not supported, falling back to synchronous mode\n"));
4865	pGip->u32Mode = SUPGIPMODE_SYNC_TSC;
4866	}
4867	}
4868	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
4869	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0, supdrvGipSyncTimer, pDevExt);
4870	if (RT_SUCCESS(rc))
4871	{
4872	rc = RTMpNotificationRegister(supdrvGipMpEvent, pDevExt);
4873	if (RT_SUCCESS(rc))
4874	{
4875	rc = RTMpOnAll(supdrvGipInitOnCpu, pDevExt, pGip);
4876	if (RT_SUCCESS(rc))
4877	{
4878	/*
4879	* We're good.
4880	*/
4881	Log(("supdrvGipCreate: %u ns interval.\n", u32Interval));
4882	g_pSUPGlobalInfoPage = pGip;
4883	return VINF_SUCCESS;
4884	}
4885
4886	OSDBGPRINT(("supdrvGipCreate: RTMpOnAll failed with rc=%Rrc\n", rc));
4887	RTMpNotificationDeregister(supdrvGipMpEvent, pDevExt);
4888
4889	}
4890	else
4891	OSDBGPRINT(("supdrvGipCreate: failed to register MP event notfication. rc=%Rrc\n", rc));
4892	}
4893	else
4894	{
4895	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %u ns interval. rc=%Rrc\n", u32Interval, rc));
4896	Assert(!pDevExt->pGipTimer);
4897	}
4898	supdrvGipDestroy(pDevExt);
4899	return rc;
4900	}
4901
4902
4903	/**
4904	* Terminates the GIP.
4905	*
4906	* @param pDevExt Instance data. GIP stuff may be updated.
4907	*/
4908	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
4909	{
4910	int rc;
4911	#ifdef DEBUG_DARWIN_GIP
4912	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
4913	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
4914	pDevExt->pGipTimer, pDevExt->GipMemObj));
4915	#endif
4916
4917	/*
4918	* Invalid the GIP data.
4919	*/
4920	if (pDevExt->pGip)
4921	{
4922	supdrvGipTerm(pDevExt->pGip);
4923	pDevExt->pGip = NULL;
4924	}
4925	g_pSUPGlobalInfoPage = NULL;
4926
4927	/*
4928	* Destroy the timer and free the GIP memory object.
4929	*/
4930	if (pDevExt->pGipTimer)
4931	{
4932	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
4933	pDevExt->pGipTimer = NULL;
4934	}
4935
4936	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
4937	{
4938	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
4939	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
4940	}
4941
4942	/*
4943	* Finally, make sure we've release the system timer resolution request
4944	* if one actually succeeded and is still pending.
4945	*/
4946	if (pDevExt->u32SystemTimerGranularityGrant)
4947	{
4948	rc = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
4949	pDevExt->u32SystemTimerGranularityGrant = 0;
4950	}
4951	}
4952
4953
4954	/**
4955	* Timer callback function sync GIP mode.
4956	* @param pTimer The timer.
4957	* @param pvUser The device extension.
4958	*/
4959	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
4960	{
4961	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
4962	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4963	uint64_t u64TSC = ASMReadTSC();
4964	uint64_t NanoTS = RTTimeSystemNanoTS();
4965
4966	supdrvGipUpdate(pDevExt->pGip, NanoTS, u64TSC, NIL_RTCPUID, iTick);
4967
4968	ASMSetFlags(fOldFlags);
4969	}
4970
4971
4972	/**
4973	* Timer callback function for async GIP mode.
4974	* @param pTimer The timer.
4975	* @param pvUser The device extension.
4976	*/
4977	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
4978	{
4979	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
4980	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4981	RTCPUID idCpu = RTMpCpuId();
4982	uint64_t u64TSC = ASMReadTSC();
4983	uint64_t NanoTS = RTTimeSystemNanoTS();
4984
4985	/** @todo reset the transaction number and whatnot when iTick == 1. */
4986	if (pDevExt->idGipMaster == idCpu)
4987	supdrvGipUpdate(pDevExt->pGip, NanoTS, u64TSC, idCpu, iTick);
4988	else
4989	supdrvGipUpdatePerCpu(pDevExt->pGip, NanoTS, u64TSC, idCpu, ASMGetApicId(), iTick);
4990
4991	ASMSetFlags(fOldFlags);
4992	}
4993
4994
4995	/**
4996	* Finds our (@a idCpu) entry, or allocates a new one if not found.
4997	*
4998	* @returns Index of the CPU in the cache set.
4999	* @param pGip The GIP.
5000	* @param idCpu The CPU ID.
5001	*/
5002	static uint32_t supdrvGipCpuIndexFromCpuId(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
5003	{
5004	uint32_t i, cTries;
5005
5006	/*
5007	* ASSUMES that CPU IDs are constant.
5008	*/
5009	for (i = 0; i < pGip->cCpus; i++)
5010	if (pGip->aCPUs[i].idCpu == idCpu)
5011	return i;
5012
5013	cTries = 0;
5014	do
5015	{
5016	for (i = 0; i < pGip->cCpus; i++)
5017	{
5018	bool fRc;
5019	ASMAtomicCmpXchgSize(&pGip->aCPUs[i].idCpu, idCpu, NIL_RTCPUID, fRc);
5020	if (fRc)
5021	return i;
5022	}
5023	} while (cTries++ < 32);
5024	AssertReleaseFailed();
5025	return i - 1;
5026	}
5027
5028
5029	/**
5030	* The calling CPU should be accounted as online, update GIP accordingly.
5031	*
5032	* This is used by supdrvGipMpEvent as well as the supdrvGipCreate.
5033	*
5034	* @param pGip The GIP.
5035	* @param idCpu The CPU ID.
5036	*/
5037	static void supdrvGipMpEventOnline(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
5038	{
5039	int iCpuSet = 0;
5040	uint16_t idApic = UINT16_MAX;
5041	uint32_t i = 0;
5042	uint64_t u64NanoTS = 0;
5043
5044	AssertRelease(idCpu == RTMpCpuId());
5045	Assert(pGip->cPossibleCpus == RTMpGetCount());
5046
5047	/*
5048	* Update the globals.
5049	*/
5050	ASMAtomicWriteU16(&pGip->cPresentCpus, RTMpGetPresentCount());
5051	ASMAtomicWriteU16(&pGip->cOnlineCpus, RTMpGetOnlineCount());
5052	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5053	if (iCpuSet >= 0)
5054	{
5055	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5056	RTCpuSetAddByIndex(&pGip->OnlineCpuSet, iCpuSet);
5057	RTCpuSetAddByIndex(&pGip->PresentCpuSet, iCpuSet);
5058	}
5059
5060	/*
5061	* Update the entry.
5062	*/
5063	u64NanoTS = RTTimeSystemNanoTS() - pGip->u32UpdateIntervalNS;
5064	i = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
5065	supdrvGipInitCpu(pGip, &pGip->aCPUs[i], u64NanoTS);
5066	idApic = ASMGetApicId();
5067	ASMAtomicUoWriteU16(&pGip->aCPUs[i].idApic, idApic);
5068	ASMAtomicUoWriteS16(&pGip->aCPUs[i].iCpuSet, (int16_t)iCpuSet);
5069	ASMAtomicUoWriteSize(&pGip->aCPUs[i].idCpu, idCpu);
5070
5071	/*
5072	* Update the APIC ID and CPU set index mappings.
5073	*/
5074	ASMAtomicWriteU16(&pGip->aiCpuFromApicId[idApic], i);
5075	ASMAtomicWriteU16(&pGip->aiCpuFromCpuSetIdx[iCpuSet], i);
5076
5077	/* commit it */
5078	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_ONLINE);
5079	}
5080
5081
5082	/**
5083	* The CPU should be accounted as offline, update the GIP accordingly.
5084	*
5085	* This is used by supdrvGipMpEvent.
5086	*
5087	* @param pGip The GIP.
5088	* @param idCpu The CPU ID.
5089	*/
5090	static void supdrvGipMpEventOffline(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
5091	{
5092	int iCpuSet;
5093	unsigned i;
5094
5095	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5096	AssertReturnVoid(iCpuSet >= 0);
5097
5098	i = pGip->aiCpuFromCpuSetIdx[iCpuSet];
5099	AssertReturnVoid(i < pGip->cCpus);
5100	AssertReturnVoid(pGip->aCPUs[i].idCpu == idCpu);
5101
5102	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5103	RTCpuSetDelByIndex(&pGip->OnlineCpuSet, iCpuSet);
5104
5105	/* commit it */
5106	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_OFFLINE);
5107	}
5108
5109
5110	/**
5111	* Multiprocessor event notification callback.
5112	*
5113	* This is used to make sure that the GIP master gets passed on to
5114	* another CPU. It also updates the associated CPU data.
5115	*
5116	* @param enmEvent The event.
5117	* @param idCpu The cpu it applies to.
5118	* @param pvUser Pointer to the device extension.
5119	*
5120	* @remarks This function -must- fire on the newly online'd CPU for the
5121	* RTMPEVENT_ONLINE case and can fire on any CPU for the
5122	* RTMPEVENT_OFFLINE case.
5123	*/
5124	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser)
5125	{
5126	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5127	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5128
5129	AssertRelease(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
5130
5131	/*
5132	* Update the GIP CPU data.
5133	*/
5134	if (pGip)
5135	{
5136	switch (enmEvent)
5137	{
5138	case RTMPEVENT_ONLINE:
5139	AssertRelease(idCpu == RTMpCpuId());
5140	supdrvGipMpEventOnline(pGip, idCpu);
5141	break;
5142	case RTMPEVENT_OFFLINE:
5143	supdrvGipMpEventOffline(pGip, idCpu);
5144	break;
5145
5146	}
5147	}
5148
5149	/*
5150	* Make sure there is a master GIP.
5151	*/
5152	if (enmEvent == RTMPEVENT_OFFLINE)
5153	{
5154	RTCPUID idGipMaster = ASMAtomicReadU32(&pDevExt->idGipMaster);
5155	if (idGipMaster == idCpu)
5156	{
5157	/*
5158	* Find a new GIP master.
5159	*/
5160	bool fIgnored;
5161	unsigned i;
5162	RTCPUID idNewGipMaster = NIL_RTCPUID;
5163	RTCPUSET OnlineCpus;
5164	RTMpGetOnlineSet(&OnlineCpus);
5165
5166	for (i = 0; i < RTCPUSET_MAX_CPUS; i++)
5167	{
5168	RTCPUID idCurCpu = RTMpCpuIdFromSetIndex(i);
5169	if ( RTCpuSetIsMember(&OnlineCpus, idCurCpu)
5170	&& idCurCpu != idGipMaster)
5171	{
5172	idNewGipMaster = idCurCpu;
5173	break;
5174	}
5175	}
5176
5177	Log(("supdrvGipMpEvent: Gip master %#lx -> %#lx\n", (long)idGipMaster, (long)idNewGipMaster));
5178	ASMAtomicCmpXchgSize(&pDevExt->idGipMaster, idNewGipMaster, idGipMaster, fIgnored);
5179	NOREF(fIgnored);
5180	}
5181	}
5182	}
5183
5184
5185	/**
5186	* Callback used by supdrvDetermineAsyncTSC to read the TSC on a CPU.
5187	*
5188	* @param idCpu Ignored.
5189	* @param pvUser1 Where to put the TSC.
5190	* @param pvUser2 Ignored.
5191	*/
5192	static DECLCALLBACK(void) supdrvDetermineAsyncTscWorker(RTCPUID idCpu, void pvUser1, void pvUser2)
5193	{
5194	#if 1
5195	ASMAtomicWriteU64((uint64_t volatile *)pvUser1, ASMReadTSC());
5196	#else
5197	(uint64_t )pvUser1 = ASMReadTSC();
5198	#endif
5199	}
5200
5201
5202	/**
5203	* Determine if Async GIP mode is required because of TSC drift.
5204	*
5205	* When using the default/normal timer code it is essential that the time stamp counter
5206	* (TSC) runs never backwards, that is, a read operation to the counter should return
5207	* a bigger value than any previous read operation. This is guaranteed by the latest
5208	* AMD CPUs and by newer Intel CPUs which never enter the C2 state (P4). In any other
5209	* case we have to choose the asynchronous timer mode.
5210	*
5211	* @param poffMin Pointer to the determined difference between different cores.
5212	* @return false if the time stamp counters appear to be synchronized, true otherwise.
5213	*/
5214	static bool supdrvDetermineAsyncTsc(uint64_t *poffMin)
5215	{
5216	/*
5217	* Just iterate all the cpus 8 times and make sure that the TSC is
5218	* ever increasing. We don't bother taking TSC rollover into account.
5219	*/
5220	int iEndCpu = RTMpGetArraySize();
5221	int iCpu;
5222	int cLoops = 8;
5223	bool fAsync = false;
5224	int rc = VINF_SUCCESS;
5225	uint64_t offMax = 0;
5226	uint64_t offMin = ~(uint64_t)0;
5227	uint64_t PrevTsc = ASMReadTSC();
5228
5229	while (cLoops-- > 0)
5230	{
5231	for (iCpu = 0; iCpu < iEndCpu; iCpu++)
5232	{
5233	uint64_t CurTsc;
5234	rc = RTMpOnSpecific(RTMpCpuIdFromSetIndex(iCpu), supdrvDetermineAsyncTscWorker, &CurTsc, NULL);
5235	if (RT_SUCCESS(rc))
5236	{
5237	if (CurTsc <= PrevTsc)
5238	{
5239	fAsync = true;
5240	offMin = offMax = PrevTsc - CurTsc;
5241	Log(("supdrvDetermineAsyncTsc: iCpu=%d cLoops=%d CurTsc=%llx PrevTsc=%llx\n",
5242	iCpu, cLoops, CurTsc, PrevTsc));
5243	break;
5244	}
5245
5246	/* Gather statistics (except the first time). */
5247	if (iCpu != 0 \|\| cLoops != 7)
5248	{
5249	uint64_t off = CurTsc - PrevTsc;
5250	if (off < offMin)
5251	offMin = off;
5252	if (off > offMax)
5253	offMax = off;
5254	Log2(("%d/%d: off=%llx\n", cLoops, iCpu, off));
5255	}
5256
5257	/* Next */
5258	PrevTsc = CurTsc;
5259	}
5260	else if (rc == VERR_NOT_SUPPORTED)
5261	break;
5262	else
5263	AssertMsg(rc == VERR_CPU_NOT_FOUND \|\| rc == VERR_CPU_OFFLINE, ("%d\n", rc));
5264	}
5265
5266	/* broke out of the loop. */
5267	if (iCpu < iEndCpu)
5268	break;
5269	}
5270
5271	poffMin = offMin; / Almost RTMpOnSpecific profiling. */
5272	Log(("supdrvDetermineAsyncTsc: returns %d; iEndCpu=%d rc=%d offMin=%llx offMax=%llx\n",
5273	fAsync, iEndCpu, rc, offMin, offMax));
5274	#if !defined(RT_OS_SOLARIS) && !defined(RT_OS_OS2) && !defined(RT_OS_WINDOWS)
5275	OSDBGPRINT(("vboxdrv: fAsync=%d offMin=%#lx offMax=%#lx\n", fAsync, (long)offMin, (long)offMax));
5276	#endif
5277	return fAsync;
5278	}
5279
5280
5281	/**
5282	* Determine the GIP TSC mode.
5283	*
5284	* @returns The most suitable TSC mode.
5285	* @param pDevExt Pointer to the device instance data.
5286	*/
5287	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt)
5288	{
5289	/*
5290	* On SMP we're faced with two problems:
5291	* (1) There might be a skew between the CPU, so that cpu0
5292	* returns a TSC that is slightly different from cpu1.
5293	* (2) Power management (and other things) may cause the TSC
5294	* to run at a non-constant speed, and cause the speed
5295	* to be different on the cpus. This will result in (1).
5296	*
5297	* So, on SMP systems we'll have to select the ASYNC update method
5298	* if there are symptoms of these problems.
5299	*/
5300	if (RTMpGetCount() > 1)
5301	{
5302	uint32_t uEAX, uEBX, uECX, uEDX;
5303	uint64_t u64DiffCoresIgnored;
5304
5305	/* Permit the user and/or the OS specific bits to force async mode. */
5306	if (supdrvOSGetForcedAsyncTscMode(pDevExt))
5307	return SUPGIPMODE_ASYNC_TSC;
5308
5309	/* Try check for current differences between the cpus. */
5310	if (supdrvDetermineAsyncTsc(&u64DiffCoresIgnored))
5311	return SUPGIPMODE_ASYNC_TSC;
5312
5313	/*
5314	* If the CPU supports power management and is an AMD one we
5315	* won't trust it unless it has the TscInvariant bit is set.
5316	*/
5317	/* Check for "AuthenticAMD" */
5318	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
5319	if ( uEAX >= 1
5320	&& uEBX == X86_CPUID_VENDOR_AMD_EBX
5321	&& uECX == X86_CPUID_VENDOR_AMD_ECX
5322	&& uEDX == X86_CPUID_VENDOR_AMD_EDX)
5323	{
5324	/* Check for APM support and that TscInvariant is cleared. */
5325	ASMCpuId(0x80000000, &uEAX, &uEBX, &uECX, &uEDX);
5326	if (uEAX >= 0x80000007)
5327	{
5328	ASMCpuId(0x80000007, &uEAX, &uEBX, &uECX, &uEDX);
5329	if ( !(uEDX & RT_BIT(8))/* TscInvariant */
5330	&& (uEDX & 0x3e)) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
5331	return SUPGIPMODE_ASYNC_TSC;
5332	}
5333	}
5334	}
5335	return SUPGIPMODE_SYNC_TSC;
5336	}
5337
5338
5339	/**
5340	* Initializes per-CPU GIP information.
5341	*
5342	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5343	* @param pCpu Pointer to which GIP CPU to initalize.
5344	* @param u64NanoTS The current nanosecond timestamp.
5345	*/
5346	static void supdrvGipInitCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pCpu, uint64_t u64NanoTS)
5347	{
5348	pCpu->u32TransactionId = 2;
5349	pCpu->u64NanoTS = u64NanoTS;
5350	pCpu->u64TSC = ASMReadTSC();
5351
5352	pCpu->enmState = SUPGIPCPUSTATE_INVALID;
5353	pCpu->idCpu = NIL_RTCPUID;
5354	pCpu->iCpuSet = -1;
5355	pCpu->idApic = UINT16_MAX;
5356
5357	/*
5358	* We don't know the following values until we've executed updates.
5359	* So, we'll just pretend it's a 4 GHz CPU and adjust the history it on
5360	* the 2nd timer callout.
5361	*/
5362	pCpu->u64CpuHz = _4G + 1; /* tstGIP-2 depends on this. */
5363	pCpu->u32UpdateIntervalTSC
5364	= pCpu->au32TSCHistory[0]
5365	= pCpu->au32TSCHistory[1]
5366	= pCpu->au32TSCHistory[2]
5367	= pCpu->au32TSCHistory[3]
5368	= pCpu->au32TSCHistory[4]
5369	= pCpu->au32TSCHistory[5]
5370	= pCpu->au32TSCHistory[6]
5371	= pCpu->au32TSCHistory[7]
5372	= (uint32_t)(_4G / pGip->u32UpdateHz);
5373	}
5374
5375
5376	/**
5377	* Initializes the GIP data.
5378	*
5379	* @param pDevExt Pointer to the device instance data.
5380	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5381	* @param HCPhys The physical address of the GIP.
5382	* @param u64NanoTS The current nanosecond timestamp.
5383	* @param uUpdateHz The update frequency.
5384	* @param cCpus The CPU count.
5385	*/
5386	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
5387	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned cCpus)
5388	{
5389	size_t const cbGip = RT_ALIGN_Z(RT_OFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), PAGE_SIZE);
5390	unsigned i;
5391	#ifdef DEBUG_DARWIN_GIP
5392	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5393	#else
5394	LogFlow(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5395	#endif
5396
5397	/*
5398	* Initialize the structure.
5399	*/
5400	memset(pGip, 0, cbGip);
5401	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
5402	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
5403	pGip->u32Mode = supdrvGipDeterminTscMode(pDevExt);
5404	pGip->cCpus = (uint16_t)cCpus;
5405	pGip->cPages = (uint16_t)(cbGip / PAGE_SIZE);
5406	pGip->u32UpdateHz = uUpdateHz;
5407	pGip->u32UpdateIntervalNS = 1000000000 / uUpdateHz;
5408	pGip->u64NanoTSLastUpdateHz = u64NanoTS;
5409	RTCpuSetEmpty(&pGip->OnlineCpuSet);
5410	RTCpuSetEmpty(&pGip->PresentCpuSet);
5411	RTMpGetSet(&pGip->PossibleCpuSet);
5412	pGip->cOnlineCpus = RTMpGetOnlineCount();
5413	pGip->cPresentCpus = RTMpGetPresentCount();
5414	pGip->cPossibleCpus = RTMpGetCount();
5415	pGip->idCpuMax = RTMpGetMaxCpuId();
5416	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromApicId); i++)
5417	pGip->aiCpuFromApicId[i] = 0;
5418	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx); i++)
5419	pGip->aiCpuFromCpuSetIdx[i] = UINT16_MAX;
5420
5421	for (i = 0; i < cCpus; i++)
5422	supdrvGipInitCpu(pGip, &pGip->aCPUs[i], u64NanoTS);
5423
5424	/*
5425	* Link it to the device extension.
5426	*/
5427	pDevExt->pGip = pGip;
5428	pDevExt->HCPhysGip = HCPhys;
5429	pDevExt->cGipUsers = 0;
5430	}
5431
5432
5433	/**
5434	* On CPU initialization callback for RTMpOnAll.
5435	*
5436	* @param idCpu The CPU ID.
5437	* @param pvUser1 The device extension.
5438	* @param pvUser2 The GIP.
5439	*/
5440	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2)
5441	{
5442	/* This is good enough, even though it will update some of the globals a
5443	bit to much. */
5444	supdrvGipMpEventOnline((PSUPGLOBALINFOPAGE)pvUser2, idCpu);
5445	}
5446
5447
5448	/**
5449	* Invalidates the GIP data upon termination.
5450	*
5451	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5452	*/
5453	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
5454	{
5455	unsigned i;
5456	pGip->u32Magic = 0;
5457	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
5458	{
5459	pGip->aCPUs[i].u64NanoTS = 0;
5460	pGip->aCPUs[i].u64TSC = 0;
5461	pGip->aCPUs[i].iTSCHistoryHead = 0;
5462	}
5463	}
5464
5465
5466	/**
5467	* Worker routine for supdrvGipUpdate and supdrvGipUpdatePerCpu that
5468	* updates all the per cpu data except the transaction id.
5469	*
5470	* @param pGip The GIP.
5471	* @param pGipCpu Pointer to the per cpu data.
5472	* @param u64NanoTS The current time stamp.
5473	* @param u64TSC The current TSC.
5474	* @param iTick The current timer tick.
5475	*/
5476	static void supdrvGipDoUpdateCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick)
5477	{
5478	uint64_t u64TSCDelta;
5479	uint32_t u32UpdateIntervalTSC;
5480	uint32_t u32UpdateIntervalTSCSlack;
5481	unsigned iTSCHistoryHead;
5482	uint64_t u64CpuHz;
5483	uint32_t u32TransactionId;
5484
5485	/* Delta between this and the previous update. */
5486	ASMAtomicUoWriteU32(&pGipCpu->u32PrevUpdateIntervalNS, (uint32_t)(u64NanoTS - pGipCpu->u64NanoTS));
5487
5488	/*
5489	* Update the NanoTS.
5490	*/
5491	ASMAtomicWriteU64(&pGipCpu->u64NanoTS, u64NanoTS);
5492
5493	/*
5494	* Calc TSC delta.
5495	*/
5496	/** @todo validate the NanoTS delta, don't trust the OS to call us when it should... */
5497	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
5498	ASMAtomicWriteU64(&pGipCpu->u64TSC, u64TSC);
5499
5500	if (u64TSCDelta >> 32)
5501	{
5502	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
5503	pGipCpu->cErrors++;
5504	}
5505
5506	/*
5507	* On the 2nd and 3rd callout, reset the history with the current TSC
5508	* interval since the values entered by supdrvGipInit are totally off.
5509	* The interval on the 1st callout completely unreliable, the 2nd is a bit
5510	* better, while the 3rd should be most reliable.
5511	*/
5512	u32TransactionId = pGipCpu->u32TransactionId;
5513	if (RT_UNLIKELY( ( u32TransactionId == 5
5514	\|\| u32TransactionId == 7)
5515	&& ( iTick == 2
5516	\|\| iTick == 3) ))
5517	{
5518	unsigned i;
5519	for (i = 0; i < RT_ELEMENTS(pGipCpu->au32TSCHistory); i++)
5520	ASMAtomicUoWriteU32(&pGipCpu->au32TSCHistory[i], (uint32_t)u64TSCDelta);
5521	}
5522
5523	/*
5524	* TSC History.
5525	*/
5526	Assert(RT_ELEMENTS(pGipCpu->au32TSCHistory) == 8);
5527	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
5528	ASMAtomicWriteU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
5529	ASMAtomicWriteU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
5530
5531	/*
5532	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
5533	*/
5534	if (pGip->u32UpdateHz >= 1000)
5535	{
5536	uint32_t u32;
5537	u32 = pGipCpu->au32TSCHistory[0];
5538	u32 += pGipCpu->au32TSCHistory[1];
5539	u32 += pGipCpu->au32TSCHistory[2];
5540	u32 += pGipCpu->au32TSCHistory[3];
5541	u32 >>= 2;
5542	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
5543	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
5544	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
5545	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
5546	u32UpdateIntervalTSC >>= 2;
5547	u32UpdateIntervalTSC += u32;
5548	u32UpdateIntervalTSC >>= 1;
5549
5550	/* Value chosen for a 2GHz Athlon64 running linux 2.6.10/11, . */
5551	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
5552	}
5553	else if (pGip->u32UpdateHz >= 90)
5554	{
5555	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5556	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
5557	u32UpdateIntervalTSC >>= 1;
5558
5559	/* value chosen on a 2GHz thinkpad running windows */
5560	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
5561	}
5562	else
5563	{
5564	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5565
5566	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
5567	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
5568	}
5569	ASMAtomicWriteU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
5570
5571	/*
5572	* CpuHz.
5573	*/
5574	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, pGip->u32UpdateHz);
5575	ASMAtomicWriteU64(&pGipCpu->u64CpuHz, u64CpuHz);
5576	}
5577
5578
5579	/**
5580	* Updates the GIP.
5581	*
5582	* @param pGip Pointer to the GIP.
5583	* @param u64NanoTS The current nanosecond timesamp.
5584	* @param u64TSC The current TSC timesamp.
5585	* @param idCpu The CPU ID.
5586	* @param iTick The current timer tick.
5587	*/
5588	static void supdrvGipUpdate(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick)
5589	{
5590	/*
5591	* Determine the relevant CPU data.
5592	*/
5593	PSUPGIPCPU pGipCpu;
5594	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
5595	pGipCpu = &pGip->aCPUs[0];
5596	else
5597	{
5598	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
5599	if (RT_UNLIKELY(iCpu >= pGip->cCpus))
5600	return;
5601	pGipCpu = &pGip->aCPUs[iCpu];
5602	if (RT_UNLIKELY(pGipCpu->idCpu != idCpu))
5603	return;
5604	}
5605
5606	/*
5607	* Start update transaction.
5608	*/
5609	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5610	{
5611	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
5612	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5613	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5614	pGipCpu->cErrors++;
5615	return;
5616	}
5617
5618	/*
5619	* Recalc the update frequency every 0x800th time.
5620	*/
5621	if (!(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
5622	{
5623	if (pGip->u64NanoTSLastUpdateHz)
5624	{
5625	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
5626	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
5627	uint32_t u32UpdateHz = (uint32_t)((UINT64_C(1000000000) * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
5628	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
5629	{
5630	ASMAtomicWriteU32(&pGip->u32UpdateHz, u32UpdateHz);
5631	ASMAtomicWriteU32(&pGip->u32UpdateIntervalNS, 1000000000 / u32UpdateHz);
5632	}
5633	#endif
5634	}
5635	ASMAtomicWriteU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS);
5636	}
5637
5638	/*
5639	* Update the data.
5640	*/
5641	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS, u64TSC, iTick);
5642
5643	/*
5644	* Complete transaction.
5645	*/
5646	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5647	}
5648
5649
5650	/**
5651	* Updates the per cpu GIP data for the calling cpu.
5652	*
5653	* @param pGip Pointer to the GIP.
5654	* @param u64NanoTS The current nanosecond timesamp.
5655	* @param u64TSC The current TSC timesamp.
5656	* @param idCpu The CPU ID.
5657	* @param idApic The APIC id for the CPU index.
5658	* @param iTick The current timer tick.
5659	*/
5660	static void supdrvGipUpdatePerCpu(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC,
5661	RTCPUID idCpu, uint8_t idApic, uint64_t iTick)
5662	{
5663	uint32_t iCpu;
5664
5665	/*
5666	* Avoid a potential race when a CPU online notification doesn't fire on
5667	* the onlined CPU but the tick creeps in before the event notification is
5668	* run.
5669	*/
5670	if (RT_UNLIKELY(iTick == 1))
5671	{
5672	iCpu = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
5673	if (pGip->aCPUs[iCpu].enmState == SUPGIPCPUSTATE_OFFLINE)
5674	supdrvGipMpEventOnline(pGip, idCpu);
5675	}
5676
5677	iCpu = pGip->aiCpuFromApicId[idApic];
5678	if (RT_LIKELY(iCpu < pGip->cCpus))
5679	{
5680	PSUPGIPCPU pGipCpu = &pGip->aCPUs[iCpu];
5681	if (pGipCpu->idCpu == idCpu)
5682	{
5683	/*
5684	* Start update transaction.
5685	*/
5686	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5687	{
5688	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5689	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5690	pGipCpu->cErrors++;
5691	return;
5692	}
5693
5694	/*
5695	* Update the data.
5696	*/
5697	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS, u64TSC, iTick);
5698
5699	/*
5700	* Complete transaction.
5701	*/
5702	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5703	}
5704	}
5705	}
5706

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

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDrv.c@ 40282

Download in other formats: