SUPDrv.c@ 38183

Last change on this file since 38183 was 38075, checked in by vboxsync, 13 years ago
SUPDrv: Check pSession in the IOCtl routine and clean up a little earlier on windows.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 214.3 KB

Line
1	/* $Revision: 38075 $ */
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
141
142	/*******************************************************************************
143	* Global Variables *
144	*******************************************************************************/
145	DECLEXPORT(PSUPGLOBALINFOPAGE) g_pSUPGlobalInfoPage = NULL;
146
147	/**
148	* Array of the R0 SUP API.
149	*/
150	static SUPFUNC g_aFunctions[] =
151	{
152	/* name function */
153	/* Entries with absolute addresses determined at runtime, fixup
154	code makes ugly ASSUMPTIONS about the order here: */
155	{ "SUPR0AbsIs64bit", (void *)0 },
156	{ "SUPR0Abs64bitKernelCS", (void *)0 },
157	{ "SUPR0Abs64bitKernelSS", (void *)0 },
158	{ "SUPR0Abs64bitKernelDS", (void *)0 },
159	{ "SUPR0AbsKernelCS", (void *)0 },
160	{ "SUPR0AbsKernelSS", (void *)0 },
161	{ "SUPR0AbsKernelDS", (void *)0 },
162	{ "SUPR0AbsKernelES", (void *)0 },
163	{ "SUPR0AbsKernelFS", (void *)0 },
164	{ "SUPR0AbsKernelGS", (void *)0 },
165	/* Normal function pointers: */
166	{ "SUPR0ComponentRegisterFactory", (void *)SUPR0ComponentRegisterFactory },
167	{ "SUPR0ComponentDeregisterFactory", (void *)SUPR0ComponentDeregisterFactory },
168	{ "SUPR0ComponentQueryFactory", (void *)SUPR0ComponentQueryFactory },
169	{ "SUPR0ObjRegister", (void *)SUPR0ObjRegister },
170	{ "SUPR0ObjAddRef", (void *)SUPR0ObjAddRef },
171	{ "SUPR0ObjAddRefEx", (void *)SUPR0ObjAddRefEx },
172	{ "SUPR0ObjRelease", (void *)SUPR0ObjRelease },
173	{ "SUPR0ObjVerifyAccess", (void *)SUPR0ObjVerifyAccess },
174	{ "SUPR0LockMem", (void *)SUPR0LockMem },
175	{ "SUPR0UnlockMem", (void *)SUPR0UnlockMem },
176	{ "SUPR0ContAlloc", (void *)SUPR0ContAlloc },
177	{ "SUPR0ContFree", (void *)SUPR0ContFree },
178	{ "SUPR0LowAlloc", (void *)SUPR0LowAlloc },
179	{ "SUPR0LowFree", (void *)SUPR0LowFree },
180	{ "SUPR0MemAlloc", (void *)SUPR0MemAlloc },
181	{ "SUPR0MemGetPhys", (void *)SUPR0MemGetPhys },
182	{ "SUPR0MemFree", (void *)SUPR0MemFree },
183	{ "SUPR0PageAllocEx", (void *)SUPR0PageAllocEx },
184	{ "SUPR0PageFree", (void *)SUPR0PageFree },
185	{ "SUPR0Printf", (void )SUPR0Printf }, /* @todo needs wrapping? */
186	{ "SUPSemEventCreate", (void *)SUPSemEventCreate },
187	{ "SUPSemEventClose", (void *)SUPSemEventClose },
188	{ "SUPSemEventSignal", (void *)SUPSemEventSignal },
189	{ "SUPSemEventWait", (void *)SUPSemEventWait },
190	{ "SUPSemEventWaitNoResume", (void *)SUPSemEventWaitNoResume },
191	{ "SUPSemEventWaitNsAbsIntr", (void *)SUPSemEventWaitNsAbsIntr },
192	{ "SUPSemEventWaitNsRelIntr", (void *)SUPSemEventWaitNsRelIntr },
193	{ "SUPSemEventGetResolution", (void *)SUPSemEventGetResolution },
194	{ "SUPSemEventMultiCreate", (void *)SUPSemEventMultiCreate },
195	{ "SUPSemEventMultiClose", (void *)SUPSemEventMultiClose },
196	{ "SUPSemEventMultiSignal", (void *)SUPSemEventMultiSignal },
197	{ "SUPSemEventMultiReset", (void *)SUPSemEventMultiReset },
198	{ "SUPSemEventMultiWait", (void *)SUPSemEventMultiWait },
199	{ "SUPSemEventMultiWaitNoResume", (void *)SUPSemEventMultiWaitNoResume },
200	{ "SUPSemEventMultiWaitNsAbsIntr", (void *)SUPSemEventMultiWaitNsAbsIntr },
201	{ "SUPSemEventMultiWaitNsRelIntr", (void *)SUPSemEventMultiWaitNsRelIntr },
202	{ "SUPSemEventMultiGetResolution", (void *)SUPSemEventMultiGetResolution },
203	{ "SUPR0GetPagingMode", (void *)SUPR0GetPagingMode },
204	{ "SUPR0EnableVTx", (void *)SUPR0EnableVTx },
205	{ "SUPGetGIP", (void *)SUPGetGIP },
206	{ "g_pSUPGlobalInfoPage", (void *)&g_pSUPGlobalInfoPage },
207	{ "RTMemAllocTag", (void *)RTMemAllocTag },
208	{ "RTMemAllocZTag", (void *)RTMemAllocZTag },
209	{ "RTMemAllocVarTag", (void *)RTMemAllocVarTag },
210	{ "RTMemAllocZVarTag", (void *)RTMemAllocZVarTag },
211	{ "RTMemFree", (void *)RTMemFree },
212	{ "RTMemDupTag", (void *)RTMemDupTag },
213	{ "RTMemDupExTag", (void *)RTMemDupExTag },
214	{ "RTMemReallocTag", (void *)RTMemReallocTag },
215	{ "RTR0MemObjAllocLowTag", (void *)RTR0MemObjAllocLowTag },
216	{ "RTR0MemObjAllocPageTag", (void *)RTR0MemObjAllocPageTag },
217	{ "RTR0MemObjAllocPhysTag", (void *)RTR0MemObjAllocPhysTag },
218	{ "RTR0MemObjAllocPhysExTag", (void *)RTR0MemObjAllocPhysExTag },
219	{ "RTR0MemObjAllocPhysNCTag", (void *)RTR0MemObjAllocPhysNCTag },
220	{ "RTR0MemObjAllocContTag", (void *)RTR0MemObjAllocContTag },
221	{ "RTR0MemObjEnterPhysTag", (void *)RTR0MemObjEnterPhysTag },
222	{ "RTR0MemObjLockUserTag", (void *)RTR0MemObjLockUserTag },
223	{ "RTR0MemObjMapKernelTag", (void *)RTR0MemObjMapKernelTag },
224	{ "RTR0MemObjMapKernelExTag", (void *)RTR0MemObjMapKernelExTag },
225	{ "RTR0MemObjMapUserTag", (void *)RTR0MemObjMapUserTag },
226	{ "RTR0MemObjProtect", (void *)RTR0MemObjProtect },
227	{ "RTR0MemObjAddress", (void *)RTR0MemObjAddress },
228	{ "RTR0MemObjAddressR3", (void *)RTR0MemObjAddressR3 },
229	{ "RTR0MemObjSize", (void *)RTR0MemObjSize },
230	{ "RTR0MemObjIsMapping", (void *)RTR0MemObjIsMapping },
231	{ "RTR0MemObjGetPagePhysAddr", (void *)RTR0MemObjGetPagePhysAddr },
232	{ "RTR0MemObjFree", (void *)RTR0MemObjFree },
233	{ "RTR0MemUserCopyFrom", (void *)RTR0MemUserCopyFrom },
234	{ "RTR0MemUserCopyTo", (void *)RTR0MemUserCopyTo },
235	{ "RTR0MemUserIsValidAddr", (void *)RTR0MemUserIsValidAddr },
236	{ "RTR0MemKernelIsValidAddr", (void *)RTR0MemKernelIsValidAddr },
237	{ "RTR0MemAreKrnlAndUsrDifferent", (void *)RTR0MemAreKrnlAndUsrDifferent },
238	{ "RTSemMutexCreate", (void *)RTSemMutexCreate },
239	{ "RTSemMutexRequest", (void *)RTSemMutexRequest },
240	{ "RTSemMutexRequestDebug", (void *)RTSemMutexRequestDebug },
241	{ "RTSemMutexRequestNoResume", (void *)RTSemMutexRequestNoResume },
242	{ "RTSemMutexRequestNoResumeDebug", (void *)RTSemMutexRequestNoResumeDebug },
243	{ "RTSemMutexRelease", (void *)RTSemMutexRelease },
244	{ "RTSemMutexDestroy", (void *)RTSemMutexDestroy },
245	{ "RTProcSelf", (void *)RTProcSelf },
246	{ "RTR0ProcHandleSelf", (void *)RTR0ProcHandleSelf },
247	{ "RTSemFastMutexCreate", (void *)RTSemFastMutexCreate },
248	{ "RTSemFastMutexDestroy", (void *)RTSemFastMutexDestroy },
249	{ "RTSemFastMutexRequest", (void *)RTSemFastMutexRequest },
250	{ "RTSemFastMutexRelease", (void *)RTSemFastMutexRelease },
251	{ "RTSemEventCreate", (void *)RTSemEventCreate },
252	{ "RTSemEventSignal", (void *)RTSemEventSignal },
253	{ "RTSemEventWait", (void *)RTSemEventWait },
254	{ "RTSemEventWaitNoResume", (void *)RTSemEventWaitNoResume },
255	{ "RTSemEventWaitEx", (void *)RTSemEventWaitEx },
256	{ "RTSemEventWaitExDebug", (void *)RTSemEventWaitExDebug },
257	{ "RTSemEventGetResolution", (void *)RTSemEventGetResolution },
258	{ "RTSemEventDestroy", (void *)RTSemEventDestroy },
259	{ "RTSemEventMultiCreate", (void *)RTSemEventMultiCreate },
260	{ "RTSemEventMultiSignal", (void *)RTSemEventMultiSignal },
261	{ "RTSemEventMultiReset", (void *)RTSemEventMultiReset },
262	{ "RTSemEventMultiWait", (void *)RTSemEventMultiWait },
263	{ "RTSemEventMultiWaitNoResume", (void *)RTSemEventMultiWaitNoResume },
264	{ "RTSemEventMultiWaitEx", (void *)RTSemEventMultiWaitEx },
265	{ "RTSemEventMultiWaitExDebug", (void *)RTSemEventMultiWaitExDebug },
266	{ "RTSemEventMultiGetResolution", (void *)RTSemEventMultiGetResolution },
267	{ "RTSemEventMultiDestroy", (void *)RTSemEventMultiDestroy },
268	{ "RTSpinlockCreate", (void *)RTSpinlockCreate },
269	{ "RTSpinlockDestroy", (void *)RTSpinlockDestroy },
270	{ "RTSpinlockAcquire", (void *)RTSpinlockAcquire },
271	{ "RTSpinlockRelease", (void *)RTSpinlockRelease },
272	{ "RTSpinlockAcquireNoInts", (void *)RTSpinlockAcquireNoInts },
273	{ "RTSpinlockReleaseNoInts", (void *)RTSpinlockReleaseNoInts },
274	{ "RTTimeNanoTS", (void *)RTTimeNanoTS },
275	{ "RTTimeMilliTS", (void *)RTTimeMilliTS },
276	{ "RTTimeSystemNanoTS", (void *)RTTimeSystemNanoTS },
277	{ "RTTimeSystemMilliTS", (void *)RTTimeSystemMilliTS },
278	{ "RTThreadNativeSelf", (void *)RTThreadNativeSelf },
279	{ "RTThreadSleep", (void *)RTThreadSleep },
280	{ "RTThreadYield", (void *)RTThreadYield },
281	{ "RTThreadSelf", (void *)RTThreadSelf },
282	{ "RTThreadCreate", (void *)RTThreadCreate },
283	{ "RTThreadGetNative", (void *)RTThreadGetNative },
284	{ "RTThreadWait", (void *)RTThreadWait },
285	{ "RTThreadWaitNoResume", (void *)RTThreadWaitNoResume },
286	{ "RTThreadGetName", (void *)RTThreadGetName },
287	{ "RTThreadSelfName", (void *)RTThreadSelfName },
288	{ "RTThreadGetType", (void *)RTThreadGetType },
289	{ "RTThreadUserSignal", (void *)RTThreadUserSignal },
290	{ "RTThreadUserReset", (void *)RTThreadUserReset },
291	{ "RTThreadUserWait", (void *)RTThreadUserWait },
292	{ "RTThreadUserWaitNoResume", (void *)RTThreadUserWaitNoResume },
293	{ "RTThreadPreemptIsEnabled", (void *)RTThreadPreemptIsEnabled },
294	{ "RTThreadPreemptIsPending", (void *)RTThreadPreemptIsPending },
295	{ "RTThreadPreemptIsPendingTrusty", (void *)RTThreadPreemptIsPendingTrusty },
296	{ "RTThreadPreemptIsPossible", (void *)RTThreadPreemptIsPossible },
297	{ "RTThreadPreemptDisable", (void *)RTThreadPreemptDisable },
298	{ "RTThreadPreemptRestore", (void *)RTThreadPreemptRestore },
299	{ "RTThreadIsInInterrupt", (void *)RTThreadIsInInterrupt },
300	{ "RTTimerCreate", (void *)RTTimerCreate },
301	{ "RTTimerCreateEx", (void *)RTTimerCreateEx },
302	{ "RTTimerDestroy", (void *)RTTimerDestroy },
303	{ "RTTimerStart", (void *)RTTimerStart },
304	{ "RTTimerStop", (void *)RTTimerStop },
305	{ "RTTimerChangeInterval", (void *)RTTimerChangeInterval },
306	{ "RTTimerGetSystemGranularity", (void *)RTTimerGetSystemGranularity },
307	{ "RTTimerRequestSystemGranularity", (void *)RTTimerRequestSystemGranularity },
308	{ "RTTimerReleaseSystemGranularity", (void *)RTTimerReleaseSystemGranularity },
309	{ "RTTimerCanDoHighResolution", (void *)RTTimerCanDoHighResolution },
310
311	{ "RTLogDefaultInstance", (void *)RTLogDefaultInstance },
312	{ "RTMpCpuId", (void *)RTMpCpuId },
313	{ "RTMpCpuIdFromSetIndex", (void *)RTMpCpuIdFromSetIndex },
314	{ "RTMpCpuIdToSetIndex", (void *)RTMpCpuIdToSetIndex },
315	{ "RTMpGetArraySize", (void *)RTMpGetArraySize },
316	{ "RTMpIsCpuPossible", (void *)RTMpIsCpuPossible },
317	{ "RTMpGetCount", (void *)RTMpGetCount },
318	{ "RTMpGetMaxCpuId", (void *)RTMpGetMaxCpuId },
319	{ "RTMpGetOnlineCount", (void *)RTMpGetOnlineCount },
320	{ "RTMpGetOnlineSet", (void *)RTMpGetOnlineSet },
321	{ "RTMpGetSet", (void *)RTMpGetSet },
322	{ "RTMpIsCpuOnline", (void *)RTMpIsCpuOnline },
323	{ "RTMpIsCpuWorkPending", (void *)RTMpIsCpuWorkPending },
324	{ "RTMpNotificationRegister", (void *)RTMpNotificationRegister },
325	{ "RTMpNotificationDeregister", (void *)RTMpNotificationDeregister },
326	{ "RTMpOnAll", (void *)RTMpOnAll },
327	{ "RTMpOnOthers", (void *)RTMpOnOthers },
328	{ "RTMpOnSpecific", (void *)RTMpOnSpecific },
329	{ "RTMpPokeCpu", (void *)RTMpPokeCpu },
330	{ "RTPowerNotificationRegister", (void *)RTPowerNotificationRegister },
331	{ "RTPowerNotificationDeregister", (void *)RTPowerNotificationDeregister },
332	{ "RTLogRelDefaultInstance", (void *)RTLogRelDefaultInstance },
333	{ "RTLogSetDefaultInstanceThread", (void *)RTLogSetDefaultInstanceThread },
334	{ "RTLogLoggerExV", (void *)RTLogLoggerExV },
335	{ "RTLogPrintfV", (void *)RTLogPrintfV },
336	{ "RTR0AssertPanicSystem", (void *)RTR0AssertPanicSystem },
337	{ "RTAssertMsg1", (void *)RTAssertMsg1 },
338	{ "RTAssertMsg2V", (void *)RTAssertMsg2V },
339	{ "RTAssertMsg2AddV", (void *)RTAssertMsg2AddV },
340	{ "RTAssertSetQuiet", (void *)RTAssertSetQuiet },
341	{ "RTAssertMayPanic", (void *)RTAssertMayPanic },
342	{ "RTAssertSetMayPanic", (void *)RTAssertSetMayPanic },
343	{ "RTAssertAreQuiet", (void *)RTAssertAreQuiet },
344	{ "RTStrFormat", (void *)RTStrFormat },
345	{ "RTStrFormatNumber", (void *)RTStrFormatNumber },
346	{ "RTStrFormatTypeDeregister", (void *)RTStrFormatTypeDeregister },
347	{ "RTStrFormatTypeRegister", (void *)RTStrFormatTypeRegister },
348	{ "RTStrFormatTypeSetUser", (void *)RTStrFormatTypeSetUser },
349	{ "RTStrFormatV", (void *)RTStrFormatV },
350	{ "RTStrPrintf", (void *)RTStrPrintf },
351	{ "RTStrPrintfEx", (void *)RTStrPrintfEx },
352	{ "RTStrPrintfExV", (void *)RTStrPrintfExV },
353	{ "RTStrPrintfV", (void *)RTStrPrintfV },
354	{ "RTHandleTableAllocWithCtx", (void *)RTHandleTableAllocWithCtx },
355	{ "RTHandleTableCreate", (void *)RTHandleTableCreate },
356	{ "RTHandleTableCreateEx", (void *)RTHandleTableCreateEx },
357	{ "RTHandleTableDestroy", (void *)RTHandleTableDestroy },
358	{ "RTHandleTableFreeWithCtx", (void *)RTHandleTableFreeWithCtx },
359	{ "RTHandleTableLookupWithCtx", (void *)RTHandleTableLookupWithCtx },
360	{ "RTNetIPv4AddDataChecksum", (void *)RTNetIPv4AddDataChecksum },
361	{ "RTNetIPv4AddTCPChecksum", (void *)RTNetIPv4AddTCPChecksum },
362	{ "RTNetIPv4AddUDPChecksum", (void *)RTNetIPv4AddUDPChecksum },
363	{ "RTNetIPv4FinalizeChecksum", (void *)RTNetIPv4FinalizeChecksum },
364	{ "RTNetIPv4HdrChecksum", (void *)RTNetIPv4HdrChecksum },
365	{ "RTNetIPv4IsDHCPValid", (void *)RTNetIPv4IsDHCPValid },
366	{ "RTNetIPv4IsHdrValid", (void *)RTNetIPv4IsHdrValid },
367	{ "RTNetIPv4IsTCPSizeValid", (void *)RTNetIPv4IsTCPSizeValid },
368	{ "RTNetIPv4IsTCPValid", (void *)RTNetIPv4IsTCPValid },
369	{ "RTNetIPv4IsUDPSizeValid", (void *)RTNetIPv4IsUDPSizeValid },
370	{ "RTNetIPv4IsUDPValid", (void *)RTNetIPv4IsUDPValid },
371	{ "RTNetIPv4PseudoChecksum", (void *)RTNetIPv4PseudoChecksum },
372	{ "RTNetIPv4PseudoChecksumBits", (void *)RTNetIPv4PseudoChecksumBits },
373	{ "RTNetIPv4TCPChecksum", (void *)RTNetIPv4TCPChecksum },
374	{ "RTNetIPv4UDPChecksum", (void *)RTNetIPv4UDPChecksum },
375	{ "RTNetIPv6PseudoChecksum", (void *)RTNetIPv6PseudoChecksum },
376	{ "RTNetIPv6PseudoChecksumBits", (void *)RTNetIPv6PseudoChecksumBits },
377	{ "RTNetIPv6PseudoChecksumEx", (void *)RTNetIPv6PseudoChecksumEx },
378	{ "RTNetTCPChecksum", (void *)RTNetTCPChecksum },
379	{ "RTNetUDPChecksum", (void *)RTNetUDPChecksum },
380	{ "RTCrc32", (void *)RTCrc32 },
381	{ "RTCrc32Finish", (void *)RTCrc32Finish },
382	{ "RTCrc32Process", (void *)RTCrc32Process },
383	{ "RTCrc32Start", (void *)RTCrc32Start },
384	};
385
386	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
387	/**
388	* Drag in the rest of IRPT since we share it with the
389	* rest of the kernel modules on darwin.
390	*/
391	PFNRT g_apfnVBoxDrvIPRTDeps[] =
392	{
393	/* VBoxNetFlt */
394	(PFNRT)RTErrConvertFromErrno,
395	(PFNRT)RTUuidCompare,
396	(PFNRT)RTUuidCompareStr,
397	(PFNRT)RTUuidFromStr,
398	(PFNRT)RTStrDupTag,
399	(PFNRT)RTStrFree,
400	(PFNRT)RTStrCopy,
401	(PFNRT)RTStrNCmp,
402	/* VBoxNetAdp */
403	(PFNRT)RTRandBytes,
404	/* VBoxUSB */
405	(PFNRT)RTPathStripFilename,
406	NULL
407	};
408	#endif /* RT_OS_DARWIN \|\| RT_OS_SOLARIS \|\| RT_OS_SOLARIS */
409
410
411	/**
412	* Initializes the device extentsion structure.
413	*
414	* @returns IPRT status code.
415	* @param pDevExt The device extension to initialize.
416	* @param cbSession The size of the session structure. The size of
417	* SUPDRVSESSION may be smaller when SUPDRV_AGNOSTIC is
418	* defined because we're skipping the OS specific members
419	* then.
420	*/
421	int VBOXCALL supdrvInitDevExt(PSUPDRVDEVEXT pDevExt, size_t cbSession)
422	{
423	int rc;
424
425	#ifdef SUPDRV_WITH_RELEASE_LOGGER
426	/*
427	* Create the release log.
428	*/
429	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
430	PRTLOGGER pRelLogger;
431	rc = RTLogCreate(&pRelLogger, 0 /* fFlags */, "all",
432	"VBOX_RELEASE_LOG", RT_ELEMENTS(s_apszGroups), s_apszGroups, RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER, NULL);
433	if (RT_SUCCESS(rc))
434	RTLogRelSetDefaultInstance(pRelLogger);
435	/** @todo Add native hook for getting logger config parameters and setting
436	* them. On linux we should use the module parameter stuff... */
437	#endif
438
439	/*
440	* Initialize it.
441	*/
442	memset(pDevExt, 0, sizeof(*pDevExt));
443	rc = RTSpinlockCreate(&pDevExt->Spinlock);
444	if (RT_SUCCESS(rc))
445	{
446	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
447	rc = RTSemMutexCreate(&pDevExt->mtxLdr);
448	#else
449	rc = RTSemFastMutexCreate(&pDevExt->mtxLdr);
450	#endif
451	if (RT_SUCCESS(rc))
452	{
453	rc = RTSemFastMutexCreate(&pDevExt->mtxComponentFactory);
454	if (RT_SUCCESS(rc))
455	{
456	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
457	rc = RTSemMutexCreate(&pDevExt->mtxGip);
458	#else
459	rc = RTSemFastMutexCreate(&pDevExt->mtxGip);
460	#endif
461	if (RT_SUCCESS(rc))
462	{
463	rc = supdrvGipCreate(pDevExt);
464	if (RT_SUCCESS(rc))
465	{
466	pDevExt->u32Cookie = BIRD; /** @todo make this random? */
467	pDevExt->cbSession = (uint32_t)cbSession;
468
469	/*
470	* Fixup the absolute symbols.
471	*
472	* Because of the table indexing assumptions we'll have a little #ifdef orgy
473	* here rather than distributing this to OS specific files. At least for now.
474	*/
475	#ifdef RT_OS_DARWIN
476	# if ARCH_BITS == 32
477	if (SUPR0GetPagingMode() >= SUPPAGINGMODE_AMD64)
478	{
479	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
480	g_aFunctions[1].pfn = (void )0x80; / SUPR0Abs64bitKernelCS - KERNEL64_CS, seg.h */
481	g_aFunctions[2].pfn = (void )0x88; / SUPR0Abs64bitKernelSS - KERNEL64_SS, seg.h */
482	g_aFunctions[3].pfn = (void )0x88; / SUPR0Abs64bitKernelDS - KERNEL64_SS, seg.h */
483	}
484	else
485	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
486	g_aFunctions[4].pfn = (void )0x08; / SUPR0AbsKernelCS - KERNEL_CS, seg.h */
487	g_aFunctions[5].pfn = (void )0x10; / SUPR0AbsKernelSS - KERNEL_DS, seg.h */
488	g_aFunctions[6].pfn = (void )0x10; / SUPR0AbsKernelDS - KERNEL_DS, seg.h */
489	g_aFunctions[7].pfn = (void )0x10; / SUPR0AbsKernelES - KERNEL_DS, seg.h */
490	g_aFunctions[8].pfn = (void )0x10; / SUPR0AbsKernelFS - KERNEL_DS, seg.h */
491	g_aFunctions[9].pfn = (void )0x48; / SUPR0AbsKernelGS - CPU_DATA_GS, seg.h */
492	# else /* 64-bit darwin: */
493	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
494	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
495	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
496	g_aFunctions[3].pfn = (void )0; / SUPR0Abs64bitKernelDS */
497	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
498	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
499	g_aFunctions[6].pfn = (void )0; / SUPR0AbsKernelDS */
500	g_aFunctions[7].pfn = (void )0; / SUPR0AbsKernelES */
501	g_aFunctions[8].pfn = (void )0; / SUPR0AbsKernelFS */
502	g_aFunctions[9].pfn = (void )0; / SUPR0AbsKernelGS */
503
504	# endif
505	#else /* !RT_OS_DARWIN */
506	# if ARCH_BITS == 64
507	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
508	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
509	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
510	g_aFunctions[3].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0Abs64bitKernelDS */
511	# else
512	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
513	# endif
514	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
515	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
516	g_aFunctions[6].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0AbsKernelDS */
517	g_aFunctions[7].pfn = (void )(uintptr_t)ASMGetES(); / SUPR0AbsKernelES */
518	g_aFunctions[8].pfn = (void )(uintptr_t)ASMGetFS(); / SUPR0AbsKernelFS */
519	g_aFunctions[9].pfn = (void )(uintptr_t)ASMGetGS(); / SUPR0AbsKernelGS */
520	#endif /* !RT_OS_DARWIN */
521	return VINF_SUCCESS;
522	}
523
524	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
525	RTSemMutexDestroy(pDevExt->mtxGip);
526	pDevExt->mtxGip = NIL_RTSEMMUTEX;
527	#else
528	RTSemFastMutexDestroy(pDevExt->mtxGip);
529	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
530	#endif
531	}
532	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
533	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
534	}
535	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
536	RTSemMutexDestroy(pDevExt->mtxLdr);
537	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
538	#else
539	RTSemFastMutexDestroy(pDevExt->mtxLdr);
540	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
541	#endif
542	}
543	RTSpinlockDestroy(pDevExt->Spinlock);
544	pDevExt->Spinlock = NIL_RTSPINLOCK;
545	}
546	#ifdef SUPDRV_WITH_RELEASE_LOGGER
547	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
548	RTLogDestroy(RTLogSetDefaultInstance(NULL));
549	#endif
550
551	return rc;
552	}
553
554
555	/**
556	* Delete the device extension (e.g. cleanup members).
557	*
558	* @param pDevExt The device extension to delete.
559	*/
560	void VBOXCALL supdrvDeleteDevExt(PSUPDRVDEVEXT pDevExt)
561	{
562	PSUPDRVOBJ pObj;
563	PSUPDRVUSAGE pUsage;
564
565	/*
566	* Kill mutexes and spinlocks.
567	*/
568	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
569	RTSemMutexDestroy(pDevExt->mtxGip);
570	pDevExt->mtxGip = NIL_RTSEMMUTEX;
571	#else
572	RTSemFastMutexDestroy(pDevExt->mtxGip);
573	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
574	#endif
575	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
576	RTSemMutexDestroy(pDevExt->mtxLdr);
577	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
578	#else
579	RTSemFastMutexDestroy(pDevExt->mtxLdr);
580	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
581	#endif
582	RTSpinlockDestroy(pDevExt->Spinlock);
583	pDevExt->Spinlock = NIL_RTSPINLOCK;
584	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
585	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
586
587	/*
588	* Free lists.
589	*/
590	/* objects. */
591	pObj = pDevExt->pObjs;
592	Assert(!pObj); /* (can trigger on forced unloads) */
593	pDevExt->pObjs = NULL;
594	while (pObj)
595	{
596	void *pvFree = pObj;
597	pObj = pObj->pNext;
598	RTMemFree(pvFree);
599	}
600
601	/* usage records. */
602	pUsage = pDevExt->pUsageFree;
603	pDevExt->pUsageFree = NULL;
604	while (pUsage)
605	{
606	void *pvFree = pUsage;
607	pUsage = pUsage->pNext;
608	RTMemFree(pvFree);
609	}
610
611	/* kill the GIP. */
612	supdrvGipDestroy(pDevExt);
613
614	#ifdef SUPDRV_WITH_RELEASE_LOGGER
615	/* destroy the loggers. */
616	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
617	RTLogDestroy(RTLogSetDefaultInstance(NULL));
618	#endif
619	}
620
621
622	/**
623	* Create session.
624	*
625	* @returns IPRT status code.
626	* @param pDevExt Device extension.
627	* @param fUser Flag indicating whether this is a user or kernel session.
628	* @param ppSession Where to store the pointer to the session data.
629	*/
630	int VBOXCALL supdrvCreateSession(PSUPDRVDEVEXT pDevExt, bool fUser, PSUPDRVSESSION *ppSession)
631	{
632	/*
633	* Allocate memory for the session data.
634	*/
635	int rc;
636	PSUPDRVSESSION pSession = *ppSession = (PSUPDRVSESSION)RTMemAllocZ(pDevExt->cbSession);
637	if (pSession)
638	{
639	/* Initialize session data. */
640	rc = RTSpinlockCreate(&pSession->Spinlock);
641	if (!rc)
642	{
643	rc = RTHandleTableCreateEx(&pSession->hHandleTable,
644	RTHANDLETABLE_FLAGS_LOCKED \| RTHANDLETABLE_FLAGS_CONTEXT,
645	1 /uBase/, 32768 /cMax/, supdrvSessionObjHandleRetain, pSession);
646	if (RT_SUCCESS(rc))
647	{
648	Assert(pSession->Spinlock != NIL_RTSPINLOCK);
649	pSession->pDevExt = pDevExt;
650	pSession->u32Cookie = BIRD_INV;
651	/*pSession->pLdrUsage = NULL;
652	pSession->pVM = NULL;
653	pSession->pUsage = NULL;
654	pSession->pGip = NULL;
655	pSession->fGipReferenced = false;
656	pSession->Bundle.cUsed = 0; */
657	pSession->Uid = NIL_RTUID;
658	pSession->Gid = NIL_RTGID;
659	if (fUser)
660	{
661	pSession->Process = RTProcSelf();
662	pSession->R0Process = RTR0ProcHandleSelf();
663	}
664	else
665	{
666	pSession->Process = NIL_RTPROCESS;
667	pSession->R0Process = NIL_RTR0PROCESS;
668	}
669
670	VBOXDRV_SUPDRV_SESSION_CREATE(pSession, fUser);
671	LogFlow(("Created session %p initial cookie=%#x\n", pSession, pSession->u32Cookie));
672	return VINF_SUCCESS;
673	}
674
675	RTSpinlockDestroy(pSession->Spinlock);
676	}
677	RTMemFree(pSession);
678	*ppSession = NULL;
679	Log(("Failed to create spinlock, rc=%d!\n", rc));
680	}
681	else
682	rc = VERR_NO_MEMORY;
683
684	return rc;
685	}
686
687
688	/**
689	* Shared code for cleaning up a session.
690	*
691	* @param pDevExt Device extension.
692	* @param pSession Session data.
693	* This data will be freed by this routine.
694	*/
695	void VBOXCALL supdrvCloseSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
696	{
697	VBOXDRV_SUPDRV_SESSION_CLOSE(pSession);
698
699	/*
700	* Cleanup the session first.
701	*/
702	supdrvCleanupSession(pDevExt, pSession);
703
704	/*
705	* Free the rest of the session stuff.
706	*/
707	RTSpinlockDestroy(pSession->Spinlock);
708	pSession->Spinlock = NIL_RTSPINLOCK;
709	pSession->pDevExt = NULL;
710	RTMemFree(pSession);
711	LogFlow(("supdrvCloseSession: returns\n"));
712	}
713
714
715	/**
716	* Shared code for cleaning up a session (but not quite freeing it).
717	*
718	* This is primarily intended for MAC OS X where we have to clean up the memory
719	* stuff before the file handle is closed.
720	*
721	* @param pDevExt Device extension.
722	* @param pSession Session data.
723	* This data will be freed by this routine.
724	*/
725	void VBOXCALL supdrvCleanupSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
726	{
727	int rc;
728	PSUPDRVBUNDLE pBundle;
729	LogFlow(("supdrvCleanupSession: pSession=%p\n", pSession));
730
731	/*
732	* Remove logger instances related to this session.
733	*/
734	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pSession);
735
736	/*
737	* Destroy the handle table.
738	*/
739	rc = RTHandleTableDestroy(pSession->hHandleTable, supdrvSessionObjHandleDelete, pSession);
740	AssertRC(rc);
741	pSession->hHandleTable = NIL_RTHANDLETABLE;
742
743	/*
744	* Release object references made in this session.
745	* In theory there should be noone racing us in this session.
746	*/
747	Log2(("release objects - start\n"));
748	if (pSession->pUsage)
749	{
750	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
751	PSUPDRVUSAGE pUsage;
752	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
753
754	while ((pUsage = pSession->pUsage) != NULL)
755	{
756	PSUPDRVOBJ pObj = pUsage->pObj;
757	pSession->pUsage = pUsage->pNext;
758
759	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
760	if (pUsage->cUsage < pObj->cUsage)
761	{
762	pObj->cUsage -= pUsage->cUsage;
763	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
764	}
765	else
766	{
767	/* Destroy the object and free the record. */
768	if (pDevExt->pObjs == pObj)
769	pDevExt->pObjs = pObj->pNext;
770	else
771	{
772	PSUPDRVOBJ pObjPrev;
773	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
774	if (pObjPrev->pNext == pObj)
775	{
776	pObjPrev->pNext = pObj->pNext;
777	break;
778	}
779	Assert(pObjPrev);
780	}
781	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
782
783	Log(("supdrvCleanupSession: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
784	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
785	if (pObj->pfnDestructor)
786	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
787	RTMemFree(pObj);
788	}
789
790	/* free it and continue. */
791	RTMemFree(pUsage);
792
793	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
794	}
795
796	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
797	AssertMsg(!pSession->pUsage, ("Some buster reregistered an object during desturction!\n"));
798	}
799	Log2(("release objects - done\n"));
800
801	/*
802	* Release memory allocated in the session.
803	*
804	* We do not serialize this as we assume that the application will
805	* not allocated memory while closing the file handle object.
806	*/
807	Log2(("freeing memory:\n"));
808	pBundle = &pSession->Bundle;
809	while (pBundle)
810	{
811	PSUPDRVBUNDLE pToFree;
812	unsigned i;
813
814	/*
815	* Check and unlock all entries in the bundle.
816	*/
817	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
818	{
819	if (pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ)
820	{
821	Log2(("eType=%d pvR0=%p pvR3=%p cb=%ld\n", pBundle->aMem[i].eType, RTR0MemObjAddress(pBundle->aMem[i].MemObj),
822	(void *)RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3), (long)RTR0MemObjSize(pBundle->aMem[i].MemObj)));
823	if (pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ)
824	{
825	rc = RTR0MemObjFree(pBundle->aMem[i].MapObjR3, false);
826	AssertRC(rc); /** @todo figure out how to handle this. */
827	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
828	}
829	rc = RTR0MemObjFree(pBundle->aMem[i].MemObj, true /* fFreeMappings */);
830	AssertRC(rc); /** @todo figure out how to handle this. */
831	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
832	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
833	}
834	}
835
836	/*
837	* Advance and free previous bundle.
838	*/
839	pToFree = pBundle;
840	pBundle = pBundle->pNext;
841
842	pToFree->pNext = NULL;
843	pToFree->cUsed = 0;
844	if (pToFree != &pSession->Bundle)
845	RTMemFree(pToFree);
846	}
847	Log2(("freeing memory - done\n"));
848
849	/*
850	* Deregister component factories.
851	*/
852	RTSemFastMutexRequest(pDevExt->mtxComponentFactory);
853	Log2(("deregistering component factories:\n"));
854	if (pDevExt->pComponentFactoryHead)
855	{
856	PSUPDRVFACTORYREG pPrev = NULL;
857	PSUPDRVFACTORYREG pCur = pDevExt->pComponentFactoryHead;
858	while (pCur)
859	{
860	if (pCur->pSession == pSession)
861	{
862	/* unlink it */
863	PSUPDRVFACTORYREG pNext = pCur->pNext;
864	if (pPrev)
865	pPrev->pNext = pNext;
866	else
867	pDevExt->pComponentFactoryHead = pNext;
868
869	/* free it */
870	pCur->pNext = NULL;
871	pCur->pSession = NULL;
872	pCur->pFactory = NULL;
873	RTMemFree(pCur);
874
875	/* next */
876	pCur = pNext;
877	}
878	else
879	{
880	/* next */
881	pPrev = pCur;
882	pCur = pCur->pNext;
883	}
884	}
885	}
886	RTSemFastMutexRelease(pDevExt->mtxComponentFactory);
887	Log2(("deregistering component factories - done\n"));
888
889	/*
890	* Loaded images needs to be dereferenced and possibly freed up.
891	*/
892	supdrvLdrLock(pDevExt);
893	Log2(("freeing images:\n"));
894	if (pSession->pLdrUsage)
895	{
896	PSUPDRVLDRUSAGE pUsage = pSession->pLdrUsage;
897	pSession->pLdrUsage = NULL;
898	while (pUsage)
899	{
900	void *pvFree = pUsage;
901	PSUPDRVLDRIMAGE pImage = pUsage->pImage;
902	if (pImage->cUsage > pUsage->cUsage)
903	pImage->cUsage -= pUsage->cUsage;
904	else
905	supdrvLdrFree(pDevExt, pImage);
906	pUsage->pImage = NULL;
907	pUsage = pUsage->pNext;
908	RTMemFree(pvFree);
909	}
910	}
911	supdrvLdrUnlock(pDevExt);
912	Log2(("freeing images - done\n"));
913
914	/*
915	* Unmap the GIP.
916	*/
917	Log2(("umapping GIP:\n"));
918	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
919	{
920	SUPR0GipUnmap(pSession);
921	pSession->fGipReferenced = 0;
922	}
923	Log2(("umapping GIP - done\n"));
924	}
925
926
927	/**
928	* RTHandleTableDestroy callback used by supdrvCleanupSession.
929	*
930	* @returns IPRT status code, see SUPR0ObjAddRef.
931	* @param hHandleTable The handle table handle. Ignored.
932	* @param pvObj The object pointer.
933	* @param pvCtx Context, the handle type. Ignored.
934	* @param pvUser Session pointer.
935	*/
936	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser)
937	{
938	NOREF(pvCtx);
939	NOREF(hHandleTable);
940	return SUPR0ObjAddRefEx(pvObj, (PSUPDRVSESSION)pvUser, true /fNoBlocking/);
941	}
942
943
944	/**
945	* RTHandleTableDestroy callback used by supdrvCleanupSession.
946	*
947	* @param hHandleTable The handle table handle. Ignored.
948	* @param h The handle value. Ignored.
949	* @param pvObj The object pointer.
950	* @param pvCtx Context, the handle type. Ignored.
951	* @param pvUser Session pointer.
952	*/
953	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser)
954	{
955	NOREF(pvCtx);
956	NOREF(h);
957	NOREF(hHandleTable);
958	SUPR0ObjRelease(pvObj, (PSUPDRVSESSION)pvUser);
959	}
960
961
962	/**
963	* Fast path I/O Control worker.
964	*
965	* @returns VBox status code that should be passed down to ring-3 unchanged.
966	* @param uIOCtl Function number.
967	* @param idCpu VMCPU id.
968	* @param pDevExt Device extention.
969	* @param pSession Session data.
970	*/
971	int VBOXCALL supdrvIOCtlFast(uintptr_t uIOCtl, VMCPUID idCpu, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
972	{
973	/*
974	* We check the two prereqs after doing this only to allow the compiler to optimize things better.
975	*/
976	if (RT_LIKELY( RT_VALID_PTR(pSession)
977	&& pSession->pVM
978	&& pDevExt->pfnVMMR0EntryFast))
979	{
980	switch (uIOCtl)
981	{
982	case SUP_IOCTL_FAST_DO_RAW_RUN:
983	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_RAW_RUN);
984	break;
985	case SUP_IOCTL_FAST_DO_HWACC_RUN:
986	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_HWACC_RUN);
987	break;
988	case SUP_IOCTL_FAST_DO_NOP:
989	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_NOP);
990	break;
991	default:
992	return VERR_INTERNAL_ERROR;
993	}
994	return VINF_SUCCESS;
995	}
996	return VERR_INTERNAL_ERROR;
997	}
998
999
1000	/**
1001	* Helper for supdrvIOCtl. Check if pszStr contains any character of pszChars.
1002	* We would use strpbrk here if this function would be contained in the RedHat kABI white
1003	* list, see http://www.kerneldrivers.org/RHEL5.
1004	*
1005	* @returns 1 if pszStr does contain any character of pszChars, 0 otherwise.
1006	* @param pszStr String to check
1007	* @param pszChars Character set
1008	*/
1009	static int supdrvCheckInvalidChar(const char pszStr, const char pszChars)
1010	{
1011	int chCur;
1012	while ((chCur = *pszStr++) != '\0')
1013	{
1014	int ch;
1015	const char *psz = pszChars;
1016	while ((ch = *psz++) != '\0')
1017	if (ch == chCur)
1018	return 1;
1019
1020	}
1021	return 0;
1022	}
1023
1024
1025	/**
1026	* I/O Control worker.
1027	*
1028	* @returns IPRT status code.
1029	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1030	*
1031	* @param uIOCtl Function number.
1032	* @param pDevExt Device extention.
1033	* @param pSession Session data.
1034	* @param pReqHdr The request header.
1035	*/
1036	int VBOXCALL supdrvIOCtl(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1037	{
1038	VBOXDRV_SUPDRV_IOCTL_ENTRY(pSession, uIOCtl, pReqHdr);
1039
1040	/*
1041	* Validate the request.
1042	*/
1043	/* this first check could probably be omitted as its also done by the OS specific code... */
1044	if (RT_UNLIKELY( (pReqHdr->fFlags & SUPREQHDR_FLAGS_MAGIC_MASK) != SUPREQHDR_FLAGS_MAGIC
1045	\|\| pReqHdr->cbIn < sizeof(*pReqHdr)
1046	\|\| pReqHdr->cbOut < sizeof(*pReqHdr)))
1047	{
1048	OSDBGPRINT(("vboxdrv: Bad ioctl request header; cbIn=%#lx cbOut=%#lx fFlags=%#lx\n",
1049	(long)pReqHdr->cbIn, (long)pReqHdr->cbOut, (long)pReqHdr->fFlags));
1050	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1051	return VERR_INVALID_PARAMETER;
1052	}
1053	if (RT_UNLIKELY(!RT_VALID_PTR(pSession)))
1054	{
1055	OSDBGPRINT(("vboxdrv: Invalid pSession valud %p (ioctl=%p)\n", pSession, (void *)uIOCtl));
1056	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1057	return VERR_INVALID_PARAMETER;
1058	}
1059	if (RT_UNLIKELY(uIOCtl == SUP_IOCTL_COOKIE))
1060	{
1061	if (pReqHdr->u32Cookie != SUPCOOKIE_INITIAL_COOKIE)
1062	{
1063	OSDBGPRINT(("SUP_IOCTL_COOKIE: bad cookie %#lx\n", (long)pReqHdr->u32Cookie));
1064	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1065	return VERR_INVALID_PARAMETER;
1066	}
1067	}
1068	else if (RT_UNLIKELY( pReqHdr->u32Cookie != pDevExt->u32Cookie
1069	\|\| pReqHdr->u32SessionCookie != pSession->u32Cookie))
1070	{
1071	OSDBGPRINT(("vboxdrv: bad cookie %#lx / %#lx.\n", (long)pReqHdr->u32Cookie, (long)pReqHdr->u32SessionCookie));
1072	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1073	return VERR_INVALID_PARAMETER;
1074	}
1075
1076	/*
1077	* Validation macros
1078	*/
1079	#define REQ_CHECK_SIZES_EX(Name, cbInExpect, cbOutExpect) \
1080	do { \
1081	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect) \|\| pReqHdr->cbOut != (cbOutExpect))) \
1082	{ \
1083	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld. cbOut=%ld expected %ld.\n", \
1084	(long)pReq->Hdr.cbIn, (long)(cbInExpect), (long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
1085	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1086	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1087	} \
1088	} while (0)
1089
1090	#define REQ_CHECK_SIZES(Name) REQ_CHECK_SIZES_EX(Name, Name ## _SIZE_IN, Name ## _SIZE_OUT)
1091
1092	#define REQ_CHECK_SIZE_IN(Name, cbInExpect) \
1093	do { \
1094	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect))) \
1095	{ \
1096	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld.\n", \
1097	(long)pReq->Hdr.cbIn, (long)(cbInExpect))); \
1098	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1099	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1100	} \
1101	} while (0)
1102
1103	#define REQ_CHECK_SIZE_OUT(Name, cbOutExpect) \
1104	do { \
1105	if (RT_UNLIKELY(pReqHdr->cbOut != (cbOutExpect))) \
1106	{ \
1107	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbOut=%ld expected %ld.\n", \
1108	(long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
1109	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1110	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1111	} \
1112	} while (0)
1113
1114	#define REQ_CHECK_EXPR(Name, expr) \
1115	do { \
1116	if (RT_UNLIKELY(!(expr))) \
1117	{ \
1118	OSDBGPRINT(( #Name ": %s\n", #expr)); \
1119	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1120	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1121	} \
1122	} while (0)
1123
1124	#define REQ_CHECK_EXPR_FMT(expr, fmt) \
1125	do { \
1126	if (RT_UNLIKELY(!(expr))) \
1127	{ \
1128	OSDBGPRINT( fmt ); \
1129	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1130	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1131	} \
1132	} while (0)
1133
1134
1135	/*
1136	* The switch.
1137	*/
1138	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
1139	{
1140	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
1141	{
1142	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
1143	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
1144	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
1145	{
1146	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
1147	pReq->Hdr.rc = VERR_INVALID_MAGIC;
1148	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, VERR_INVALID_MAGIC);
1149	return 0;
1150	}
1151
1152	#if 0
1153	/*
1154	* Call out to the OS specific code and let it do permission checks on the
1155	* client process.
1156	*/
1157	if (!supdrvOSValidateClientProcess(pDevExt, pSession))
1158	{
1159	pReq->u.Out.u32Cookie = 0xffffffff;
1160	pReq->u.Out.u32SessionCookie = 0xffffffff;
1161	pReq->u.Out.u32SessionVersion = 0xffffffff;
1162	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1163	pReq->u.Out.pSession = NULL;
1164	pReq->u.Out.cFunctions = 0;
1165	pReq->Hdr.rc = VERR_PERMISSION_DENIED;
1166	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, VERR_PERMISSION_DENIED);
1167	return 0;
1168	}
1169	#endif
1170
1171	/*
1172	* Match the version.
1173	* The current logic is very simple, match the major interface version.
1174	*/
1175	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
1176	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
1177	{
1178	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1179	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
1180	pReq->u.Out.u32Cookie = 0xffffffff;
1181	pReq->u.Out.u32SessionCookie = 0xffffffff;
1182	pReq->u.Out.u32SessionVersion = 0xffffffff;
1183	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1184	pReq->u.Out.pSession = NULL;
1185	pReq->u.Out.cFunctions = 0;
1186	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1187	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1188	return 0;
1189	}
1190
1191	/*
1192	* Fill in return data and be gone.
1193	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
1194	* u32SessionVersion <= u32ReqVersion!
1195	*/
1196	/** @todo Somehow validate the client and negotiate a secure cookie... */
1197	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
1198	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
1199	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
1200	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1201	pReq->u.Out.pSession = pSession;
1202	pReq->u.Out.cFunctions = sizeof(g_aFunctions) / sizeof(g_aFunctions[0]);
1203	pReq->Hdr.rc = VINF_SUCCESS;
1204	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1205	return 0;
1206	}
1207
1208	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_QUERY_FUNCS(0)):
1209	{
1210	/* validate */
1211	PSUPQUERYFUNCS pReq = (PSUPQUERYFUNCS)pReqHdr;
1212	REQ_CHECK_SIZES_EX(SUP_IOCTL_QUERY_FUNCS, SUP_IOCTL_QUERY_FUNCS_SIZE_IN, SUP_IOCTL_QUERY_FUNCS_SIZE_OUT(RT_ELEMENTS(g_aFunctions)));
1213
1214	/* execute */
1215	pReq->u.Out.cFunctions = RT_ELEMENTS(g_aFunctions);
1216	memcpy(&pReq->u.Out.aFunctions[0], g_aFunctions, sizeof(g_aFunctions));
1217	pReq->Hdr.rc = VINF_SUCCESS;
1218	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1219	return 0;
1220	}
1221
1222	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_LOCK):
1223	{
1224	/* validate */
1225	PSUPPAGELOCK pReq = (PSUPPAGELOCK)pReqHdr;
1226	REQ_CHECK_SIZE_IN(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_IN);
1227	REQ_CHECK_SIZE_OUT(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_OUT(pReq->u.In.cPages));
1228	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.cPages > 0);
1229	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.pvR3 >= PAGE_SIZE);
1230
1231	/* execute */
1232	pReq->Hdr.rc = SUPR0LockMem(pSession, pReq->u.In.pvR3, pReq->u.In.cPages, &pReq->u.Out.aPages[0]);
1233	if (RT_FAILURE(pReq->Hdr.rc))
1234	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1235	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1236	return 0;
1237	}
1238
1239	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_UNLOCK):
1240	{
1241	/* validate */
1242	PSUPPAGEUNLOCK pReq = (PSUPPAGEUNLOCK)pReqHdr;
1243	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_UNLOCK);
1244
1245	/* execute */
1246	pReq->Hdr.rc = SUPR0UnlockMem(pSession, pReq->u.In.pvR3);
1247	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1248	return 0;
1249	}
1250
1251	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_ALLOC):
1252	{
1253	/* validate */
1254	PSUPCONTALLOC pReq = (PSUPCONTALLOC)pReqHdr;
1255	REQ_CHECK_SIZES(SUP_IOCTL_CONT_ALLOC);
1256
1257	/* execute */
1258	pReq->Hdr.rc = SUPR0ContAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.HCPhys);
1259	if (RT_FAILURE(pReq->Hdr.rc))
1260	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1261	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1262	return 0;
1263	}
1264
1265	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_FREE):
1266	{
1267	/* validate */
1268	PSUPCONTFREE pReq = (PSUPCONTFREE)pReqHdr;
1269	REQ_CHECK_SIZES(SUP_IOCTL_CONT_FREE);
1270
1271	/* execute */
1272	pReq->Hdr.rc = SUPR0ContFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1273	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1274	return 0;
1275	}
1276
1277	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_OPEN):
1278	{
1279	/* validate */
1280	PSUPLDROPEN pReq = (PSUPLDROPEN)pReqHdr;
1281	REQ_CHECK_SIZES(SUP_IOCTL_LDR_OPEN);
1282	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs > 0);
1283	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs < 16*_1M);
1284	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
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 < pReq->u.In.cbImageWithTabs);
1287	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.szName[0]);
1288	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1289	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, !supdrvCheckInvalidChar(pReq->u.In.szName, ";:()[]{}/\\\|&*%#@!~`\"'"));
1290	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szFilename, sizeof(pReq->u.In.szFilename)));
1291
1292	/* execute */
1293	pReq->Hdr.rc = supdrvIOCtl_LdrOpen(pDevExt, pSession, pReq);
1294	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1295	return 0;
1296	}
1297
1298	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOAD):
1299	{
1300	/* validate */
1301	PSUPLDRLOAD pReq = (PSUPLDRLOAD)pReqHdr;
1302	REQ_CHECK_EXPR(Name, pReq->Hdr.cbIn >= sizeof(*pReq));
1303	REQ_CHECK_SIZES_EX(SUP_IOCTL_LDR_LOAD, SUP_IOCTL_LDR_LOAD_SIZE_IN(pReq->u.In.cbImageWithTabs), SUP_IOCTL_LDR_LOAD_SIZE_OUT);
1304	REQ_CHECK_EXPR(SUP_IOCTL_LDR_LOAD, pReq->u.In.cSymbols <= 16384);
1305	REQ_CHECK_EXPR_FMT( !pReq->u.In.cSymbols
1306	\|\| ( pReq->u.In.offSymbols < pReq->u.In.cbImageWithTabs
1307	&& pReq->u.In.offSymbols + pReq->u.In.cSymbols * sizeof(SUPLDRSYM) <= pReq->u.In.cbImageWithTabs),
1308	("SUP_IOCTL_LDR_LOAD: offSymbols=%#lx cSymbols=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offSymbols,
1309	(long)pReq->u.In.cSymbols, (long)pReq->u.In.cbImageWithTabs));
1310	REQ_CHECK_EXPR_FMT( !pReq->u.In.cbStrTab
1311	\|\| ( pReq->u.In.offStrTab < pReq->u.In.cbImageWithTabs
1312	&& pReq->u.In.offStrTab + pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs
1313	&& pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs),
1314	("SUP_IOCTL_LDR_LOAD: offStrTab=%#lx cbStrTab=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offStrTab,
1315	(long)pReq->u.In.cbStrTab, (long)pReq->u.In.cbImageWithTabs));
1316
1317	if (pReq->u.In.cSymbols)
1318	{
1319	uint32_t i;
1320	PSUPLDRSYM paSyms = (PSUPLDRSYM)&pReq->u.In.abImage[pReq->u.In.offSymbols];
1321	for (i = 0; i < pReq->u.In.cSymbols; i++)
1322	{
1323	REQ_CHECK_EXPR_FMT(paSyms[i].offSymbol < pReq->u.In.cbImageWithTabs,
1324	("SUP_IOCTL_LDR_LOAD: sym #%ld: symb off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offSymbol, (long)pReq->u.In.cbImageWithTabs));
1325	REQ_CHECK_EXPR_FMT(paSyms[i].offName < pReq->u.In.cbStrTab,
1326	("SUP_IOCTL_LDR_LOAD: sym #%ld: name off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1327	REQ_CHECK_EXPR_FMT(RTStrEnd((char const *)&pReq->u.In.abImage[pReq->u.In.offStrTab + paSyms[i].offName],
1328	pReq->u.In.cbStrTab - paSyms[i].offName),
1329	("SUP_IOCTL_LDR_LOAD: sym #%ld: unterminated name! (%#lx / %#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1330	}
1331	}
1332
1333	/* execute */
1334	pReq->Hdr.rc = supdrvIOCtl_LdrLoad(pDevExt, pSession, pReq);
1335	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1336	return 0;
1337	}
1338
1339	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_FREE):
1340	{
1341	/* validate */
1342	PSUPLDRFREE pReq = (PSUPLDRFREE)pReqHdr;
1343	REQ_CHECK_SIZES(SUP_IOCTL_LDR_FREE);
1344
1345	/* execute */
1346	pReq->Hdr.rc = supdrvIOCtl_LdrFree(pDevExt, pSession, pReq);
1347	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1348	return 0;
1349	}
1350
1351	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_GET_SYMBOL):
1352	{
1353	/* validate */
1354	PSUPLDRGETSYMBOL pReq = (PSUPLDRGETSYMBOL)pReqHdr;
1355	REQ_CHECK_SIZES(SUP_IOCTL_LDR_GET_SYMBOL);
1356	REQ_CHECK_EXPR(SUP_IOCTL_LDR_GET_SYMBOL, RTStrEnd(pReq->u.In.szSymbol, sizeof(pReq->u.In.szSymbol)));
1357
1358	/* execute */
1359	pReq->Hdr.rc = supdrvIOCtl_LdrGetSymbol(pDevExt, pSession, pReq);
1360	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1361	return 0;
1362	}
1363
1364	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0(0)):
1365	{
1366	/* validate */
1367	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1368	Log4(("SUP_IOCTL_CALL_VMMR0: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1369	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1370
1371	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_VMMR0_SIZE(0))
1372	{
1373	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(0), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(0));
1374
1375	/* execute */
1376	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1377	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg, pSession);
1378	else
1379	pReq->Hdr.rc = VERR_WRONG_ORDER;
1380	}
1381	else
1382	{
1383	PSUPVMMR0REQHDR pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1384	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR)),
1385	("SUP_IOCTL_CALL_VMMR0: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR))));
1386	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1387	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(pVMMReq->cbReq));
1388
1389	/* execute */
1390	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1391	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1392	else
1393	pReq->Hdr.rc = VERR_WRONG_ORDER;
1394	}
1395
1396	if ( RT_FAILURE(pReq->Hdr.rc)
1397	&& pReq->Hdr.rc != VERR_INTERRUPTED
1398	&& pReq->Hdr.rc != VERR_TIMEOUT)
1399	Log(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1400	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1401	else
1402	Log4(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1403	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1404	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1405	return 0;
1406	}
1407
1408	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0_BIG):
1409	{
1410	/* validate */
1411	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1412	PSUPVMMR0REQHDR pVMMReq;
1413	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1414	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1415
1416	pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1417	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR)),
1418	("SUP_IOCTL_CALL_VMMR0_BIG: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR))));
1419	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0_BIG, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1420	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));
1421
1422	/* execute */
1423	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1424	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1425	else
1426	pReq->Hdr.rc = VERR_WRONG_ORDER;
1427
1428	if ( RT_FAILURE(pReq->Hdr.rc)
1429	&& pReq->Hdr.rc != VERR_INTERRUPTED
1430	&& pReq->Hdr.rc != VERR_TIMEOUT)
1431	Log(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1432	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1433	else
1434	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1435	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1436	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1437	return 0;
1438	}
1439
1440	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_PAGING_MODE):
1441	{
1442	/* validate */
1443	PSUPGETPAGINGMODE pReq = (PSUPGETPAGINGMODE)pReqHdr;
1444	REQ_CHECK_SIZES(SUP_IOCTL_GET_PAGING_MODE);
1445
1446	/* execute */
1447	pReq->Hdr.rc = VINF_SUCCESS;
1448	pReq->u.Out.enmMode = SUPR0GetPagingMode();
1449	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1450	return 0;
1451	}
1452
1453	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_ALLOC):
1454	{
1455	/* validate */
1456	PSUPLOWALLOC pReq = (PSUPLOWALLOC)pReqHdr;
1457	REQ_CHECK_EXPR(SUP_IOCTL_LOW_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_LOW_ALLOC_SIZE_IN);
1458	REQ_CHECK_SIZES_EX(SUP_IOCTL_LOW_ALLOC, SUP_IOCTL_LOW_ALLOC_SIZE_IN, SUP_IOCTL_LOW_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1459
1460	/* execute */
1461	pReq->Hdr.rc = SUPR0LowAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1462	if (RT_FAILURE(pReq->Hdr.rc))
1463	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1464	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1465	return 0;
1466	}
1467
1468	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_FREE):
1469	{
1470	/* validate */
1471	PSUPLOWFREE pReq = (PSUPLOWFREE)pReqHdr;
1472	REQ_CHECK_SIZES(SUP_IOCTL_LOW_FREE);
1473
1474	/* execute */
1475	pReq->Hdr.rc = SUPR0LowFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1476	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1477	return 0;
1478	}
1479
1480	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_MAP):
1481	{
1482	/* validate */
1483	PSUPGIPMAP pReq = (PSUPGIPMAP)pReqHdr;
1484	REQ_CHECK_SIZES(SUP_IOCTL_GIP_MAP);
1485
1486	/* execute */
1487	pReq->Hdr.rc = SUPR0GipMap(pSession, &pReq->u.Out.pGipR3, &pReq->u.Out.HCPhysGip);
1488	if (RT_SUCCESS(pReq->Hdr.rc))
1489	pReq->u.Out.pGipR0 = pDevExt->pGip;
1490	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1491	return 0;
1492	}
1493
1494	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_UNMAP):
1495	{
1496	/* validate */
1497	PSUPGIPUNMAP pReq = (PSUPGIPUNMAP)pReqHdr;
1498	REQ_CHECK_SIZES(SUP_IOCTL_GIP_UNMAP);
1499
1500	/* execute */
1501	pReq->Hdr.rc = SUPR0GipUnmap(pSession);
1502	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1503	return 0;
1504	}
1505
1506	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SET_VM_FOR_FAST):
1507	{
1508	/* validate */
1509	PSUPSETVMFORFAST pReq = (PSUPSETVMFORFAST)pReqHdr;
1510	REQ_CHECK_SIZES(SUP_IOCTL_SET_VM_FOR_FAST);
1511	REQ_CHECK_EXPR_FMT( !pReq->u.In.pVMR0
1512	\|\| ( VALID_PTR(pReq->u.In.pVMR0)
1513	&& !((uintptr_t)pReq->u.In.pVMR0 & (PAGE_SIZE - 1))),
1514	("SUP_IOCTL_SET_VM_FOR_FAST: pVMR0=%p!\n", pReq->u.In.pVMR0));
1515	/* execute */
1516	pSession->pVM = pReq->u.In.pVMR0;
1517	pReq->Hdr.rc = VINF_SUCCESS;
1518	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1519	return 0;
1520	}
1521
1522	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_ALLOC_EX):
1523	{
1524	/* validate */
1525	PSUPPAGEALLOCEX pReq = (PSUPPAGEALLOCEX)pReqHdr;
1526	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC_EX, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_EX_SIZE_IN);
1527	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));
1528	REQ_CHECK_EXPR_FMT(pReq->u.In.fKernelMapping \|\| pReq->u.In.fUserMapping,
1529	("SUP_IOCTL_PAGE_ALLOC_EX: No mapping requested!\n"));
1530	REQ_CHECK_EXPR_FMT(pReq->u.In.fUserMapping,
1531	("SUP_IOCTL_PAGE_ALLOC_EX: Must have user mapping!\n"));
1532	REQ_CHECK_EXPR_FMT(!pReq->u.In.fReserved0 && !pReq->u.In.fReserved1,
1533	("SUP_IOCTL_PAGE_ALLOC_EX: fReserved0=%d fReserved1=%d\n", pReq->u.In.fReserved0, pReq->u.In.fReserved1));
1534
1535	/* execute */
1536	pReq->Hdr.rc = SUPR0PageAllocEx(pSession, pReq->u.In.cPages, 0 /* fFlags */,
1537	pReq->u.In.fUserMapping ? &pReq->u.Out.pvR3 : NULL,
1538	pReq->u.In.fKernelMapping ? &pReq->u.Out.pvR0 : NULL,
1539	&pReq->u.Out.aPages[0]);
1540	if (RT_FAILURE(pReq->Hdr.rc))
1541	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1542	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1543	return 0;
1544	}
1545
1546	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_MAP_KERNEL):
1547	{
1548	/* validate */
1549	PSUPPAGEMAPKERNEL pReq = (PSUPPAGEMAPKERNEL)pReqHdr;
1550	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_MAP_KERNEL);
1551	REQ_CHECK_EXPR_FMT(!pReq->u.In.fFlags, ("SUP_IOCTL_PAGE_MAP_KERNEL: fFlags=%#x! MBZ\n", pReq->u.In.fFlags));
1552	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_MAP_KERNEL: offSub=%#x\n", pReq->u.In.offSub));
1553	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1554	("SUP_IOCTL_PAGE_MAP_KERNEL: cbSub=%#x\n", pReq->u.In.cbSub));
1555
1556	/* execute */
1557	pReq->Hdr.rc = SUPR0PageMapKernel(pSession, pReq->u.In.pvR3, pReq->u.In.offSub, pReq->u.In.cbSub,
1558	pReq->u.In.fFlags, &pReq->u.Out.pvR0);
1559	if (RT_FAILURE(pReq->Hdr.rc))
1560	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1561	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1562	return 0;
1563	}
1564
1565	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_PROTECT):
1566	{
1567	/* validate */
1568	PSUPPAGEPROTECT pReq = (PSUPPAGEPROTECT)pReqHdr;
1569	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_PROTECT);
1570	REQ_CHECK_EXPR_FMT(!(pReq->u.In.fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)),
1571	("SUP_IOCTL_PAGE_PROTECT: fProt=%#x!\n", pReq->u.In.fProt));
1572	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_PROTECT: offSub=%#x\n", pReq->u.In.offSub));
1573	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1574	("SUP_IOCTL_PAGE_PROTECT: cbSub=%#x\n", pReq->u.In.cbSub));
1575
1576	/* execute */
1577	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);
1578	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1579	return 0;
1580	}
1581
1582	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_FREE):
1583	{
1584	/* validate */
1585	PSUPPAGEFREE pReq = (PSUPPAGEFREE)pReqHdr;
1586	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_FREE);
1587
1588	/* execute */
1589	pReq->Hdr.rc = SUPR0PageFree(pSession, pReq->u.In.pvR3);
1590	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1591	return 0;
1592	}
1593
1594	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_SERVICE(0)):
1595	{
1596	/* validate */
1597	PSUPCALLSERVICE pReq = (PSUPCALLSERVICE)pReqHdr;
1598	Log4(("SUP_IOCTL_CALL_SERVICE: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1599	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1600
1601	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
1602	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(0), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(0));
1603	else
1604	{
1605	PSUPR0SERVICEREQHDR pSrvReq = (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0];
1606	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR)),
1607	("SUP_IOCTL_CALL_SERVICE: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR))));
1608	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, pSrvReq->u32Magic == SUPR0SERVICEREQHDR_MAGIC);
1609	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(pSrvReq->cbReq), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(pSrvReq->cbReq));
1610	}
1611	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1612
1613	/* execute */
1614	pReq->Hdr.rc = supdrvIOCtl_CallServiceModule(pDevExt, pSession, pReq);
1615	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1616	return 0;
1617	}
1618
1619	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOGGER_SETTINGS(0)):
1620	{
1621	/* validate */
1622	PSUPLOGGERSETTINGS pReq = (PSUPLOGGERSETTINGS)pReqHdr;
1623	size_t cbStrTab;
1624	REQ_CHECK_SIZE_OUT(SUP_IOCTL_LOGGER_SETTINGS, SUP_IOCTL_LOGGER_SETTINGS_SIZE_OUT);
1625	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->Hdr.cbIn >= SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(1));
1626	cbStrTab = pReq->Hdr.cbIn - SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(0);
1627	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offGroups < cbStrTab);
1628	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offFlags < cbStrTab);
1629	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offDestination < cbStrTab);
1630	REQ_CHECK_EXPR_FMT(pReq->u.In.szStrings[cbStrTab - 1] == '\0',
1631	("SUP_IOCTL_LOGGER_SETTINGS: cbIn=%#x cbStrTab=%#zx LastChar=%d\n",
1632	pReq->Hdr.cbIn, cbStrTab, pReq->u.In.szStrings[cbStrTab - 1]));
1633	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhich <= SUPLOGGERSETTINGS_WHICH_RELEASE);
1634	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhat <= SUPLOGGERSETTINGS_WHAT_DESTROY);
1635
1636	/* execute */
1637	pReq->Hdr.rc = supdrvIOCtl_LoggerSettings(pDevExt, pSession, pReq);
1638	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1639	return 0;
1640	}
1641
1642	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP2):
1643	{
1644	/* validate */
1645	PSUPSEMOP2 pReq = (PSUPSEMOP2)pReqHdr;
1646	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP2, SUP_IOCTL_SEM_OP2_SIZE_IN, SUP_IOCTL_SEM_OP2_SIZE_OUT);
1647	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP2, pReq->u.In.uReserved == 0);
1648
1649	/* execute */
1650	switch (pReq->u.In.uType)
1651	{
1652	case SUP_SEM_TYPE_EVENT:
1653	{
1654	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1655	switch (pReq->u.In.uOp)
1656	{
1657	case SUPSEMOP2_WAIT_MS_REL:
1658	pReq->Hdr.rc = SUPSemEventWaitNoResume(pSession, hEvent, pReq->u.In.uArg.cRelMsTimeout);
1659	break;
1660	case SUPSEMOP2_WAIT_NS_ABS:
1661	pReq->Hdr.rc = SUPSemEventWaitNsAbsIntr(pSession, hEvent, pReq->u.In.uArg.uAbsNsTimeout);
1662	break;
1663	case SUPSEMOP2_WAIT_NS_REL:
1664	pReq->Hdr.rc = SUPSemEventWaitNsRelIntr(pSession, hEvent, pReq->u.In.uArg.cRelNsTimeout);
1665	break;
1666	case SUPSEMOP2_SIGNAL:
1667	pReq->Hdr.rc = SUPSemEventSignal(pSession, hEvent);
1668	break;
1669	case SUPSEMOP2_CLOSE:
1670	pReq->Hdr.rc = SUPSemEventClose(pSession, hEvent);
1671	break;
1672	case SUPSEMOP2_RESET:
1673	default:
1674	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1675	break;
1676	}
1677	break;
1678	}
1679
1680	case SUP_SEM_TYPE_EVENT_MULTI:
1681	{
1682	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1683	switch (pReq->u.In.uOp)
1684	{
1685	case SUPSEMOP2_WAIT_MS_REL:
1686	pReq->Hdr.rc = SUPSemEventMultiWaitNoResume(pSession, hEventMulti, pReq->u.In.uArg.cRelMsTimeout);
1687	break;
1688	case SUPSEMOP2_WAIT_NS_ABS:
1689	pReq->Hdr.rc = SUPSemEventMultiWaitNsAbsIntr(pSession, hEventMulti, pReq->u.In.uArg.uAbsNsTimeout);
1690	break;
1691	case SUPSEMOP2_WAIT_NS_REL:
1692	pReq->Hdr.rc = SUPSemEventMultiWaitNsRelIntr(pSession, hEventMulti, pReq->u.In.uArg.cRelNsTimeout);
1693	break;
1694	case SUPSEMOP2_SIGNAL:
1695	pReq->Hdr.rc = SUPSemEventMultiSignal(pSession, hEventMulti);
1696	break;
1697	case SUPSEMOP2_CLOSE:
1698	pReq->Hdr.rc = SUPSemEventMultiClose(pSession, hEventMulti);
1699	break;
1700	case SUPSEMOP2_RESET:
1701	pReq->Hdr.rc = SUPSemEventMultiReset(pSession, hEventMulti);
1702	break;
1703	default:
1704	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1705	break;
1706	}
1707	break;
1708	}
1709
1710	default:
1711	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1712	break;
1713	}
1714	return 0;
1715	}
1716
1717	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP3):
1718	{
1719	/* validate */
1720	PSUPSEMOP3 pReq = (PSUPSEMOP3)pReqHdr;
1721	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP3, SUP_IOCTL_SEM_OP3_SIZE_IN, SUP_IOCTL_SEM_OP3_SIZE_OUT);
1722	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, pReq->u.In.u32Reserved == 0 && pReq->u.In.u64Reserved == 0);
1723
1724	/* execute */
1725	switch (pReq->u.In.uType)
1726	{
1727	case SUP_SEM_TYPE_EVENT:
1728	{
1729	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1730	switch (pReq->u.In.uOp)
1731	{
1732	case SUPSEMOP3_CREATE:
1733	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1734	pReq->Hdr.rc = SUPSemEventCreate(pSession, &hEvent);
1735	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEvent;
1736	break;
1737	case SUPSEMOP3_GET_RESOLUTION:
1738	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1739	pReq->Hdr.rc = VINF_SUCCESS;
1740	pReq->Hdr.cbOut = sizeof(*pReq);
1741	pReq->u.Out.cNsResolution = SUPSemEventGetResolution(pSession);
1742	break;
1743	default:
1744	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1745	break;
1746	}
1747	break;
1748	}
1749
1750	case SUP_SEM_TYPE_EVENT_MULTI:
1751	{
1752	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1753	switch (pReq->u.In.uOp)
1754	{
1755	case SUPSEMOP3_CREATE:
1756	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1757	pReq->Hdr.rc = SUPSemEventMultiCreate(pSession, &hEventMulti);
1758	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEventMulti;
1759	break;
1760	case SUPSEMOP3_GET_RESOLUTION:
1761	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1762	pReq->Hdr.rc = VINF_SUCCESS;
1763	pReq->u.Out.cNsResolution = SUPSemEventMultiGetResolution(pSession);
1764	break;
1765	default:
1766	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1767	break;
1768	}
1769	break;
1770	}
1771
1772	default:
1773	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1774	break;
1775	}
1776	return 0;
1777	}
1778
1779	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
1780	{
1781	/* validate */
1782	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
1783	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
1784	REQ_CHECK_EXPR(SUP_IOCTL_VT_CAPS, pReq->Hdr.cbIn <= SUP_IOCTL_VT_CAPS_SIZE_IN);
1785
1786	/* execute */
1787	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.Caps);
1788	if (RT_FAILURE(pReq->Hdr.rc))
1789	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1790	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1791	return 0;
1792	}
1793
1794	default:
1795	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1796	break;
1797	}
1798	VBOXDRV_SUPDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_GENERAL_FAILURE, VERR_GENERAL_FAILURE);
1799	return VERR_GENERAL_FAILURE;
1800	}
1801
1802
1803	/**
1804	* Inter-Driver Communication (IDC) worker.
1805	*
1806	* @returns VBox status code.
1807	* @retval VINF_SUCCESS on success.
1808	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1809	* @retval VERR_NOT_SUPPORTED if the request isn't supported.
1810	*
1811	* @param uReq The request (function) code.
1812	* @param pDevExt Device extention.
1813	* @param pSession Session data.
1814	* @param pReqHdr The request header.
1815	*/
1816	int VBOXCALL supdrvIDC(uintptr_t uReq, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQHDR pReqHdr)
1817	{
1818	/*
1819	* The OS specific code has already validated the pSession
1820	* pointer, and the request size being greater or equal to
1821	* size of the header.
1822	*
1823	* So, just check that pSession is a kernel context session.
1824	*/
1825	if (RT_UNLIKELY( pSession
1826	&& pSession->R0Process != NIL_RTR0PROCESS))
1827	return VERR_INVALID_PARAMETER;
1828
1829	/*
1830	* Validation macro.
1831	*/
1832	#define REQ_CHECK_IDC_SIZE(Name, cbExpect) \
1833	do { \
1834	if (RT_UNLIKELY(pReqHdr->cb != (cbExpect))) \
1835	{ \
1836	OSDBGPRINT(( #Name ": Invalid input/output sizes. cb=%ld expected %ld.\n", \
1837	(long)pReqHdr->cb, (long)(cbExpect))); \
1838	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1839	} \
1840	} while (0)
1841
1842	switch (uReq)
1843	{
1844	case SUPDRV_IDC_REQ_CONNECT:
1845	{
1846	PSUPDRVIDCREQCONNECT pReq = (PSUPDRVIDCREQCONNECT)pReqHdr;
1847	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_CONNECT, sizeof(*pReq));
1848
1849	/*
1850	* Validate the cookie and other input.
1851	*/
1852	if (pReq->Hdr.pSession != NULL)
1853	{
1854	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: pSession=%p expected NULL!\n", pReq->Hdr.pSession));
1855	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1856	}
1857	if (pReq->u.In.u32MagicCookie != SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE)
1858	{
1859	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: u32MagicCookie=%#x expected %#x!\n",
1860	(unsigned)pReq->u.In.u32MagicCookie, (unsigned)SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE));
1861	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1862	}
1863	if ( pReq->u.In.uMinVersion > pReq->u.In.uReqVersion
1864	\|\| (pReq->u.In.uMinVersion & UINT32_C(0xffff0000)) != (pReq->u.In.uReqVersion & UINT32_C(0xffff0000)))
1865	{
1866	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: uMinVersion=%#x uMaxVersion=%#x doesn't match!\n",
1867	pReq->u.In.uMinVersion, pReq->u.In.uReqVersion));
1868	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1869	}
1870
1871	/*
1872	* Match the version.
1873	* The current logic is very simple, match the major interface version.
1874	*/
1875	if ( pReq->u.In.uMinVersion > SUPDRV_IDC_VERSION
1876	\|\| (pReq->u.In.uMinVersion & 0xffff0000) != (SUPDRV_IDC_VERSION & 0xffff0000))
1877	{
1878	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1879	pReq->u.In.uReqVersion, pReq->u.In.uMinVersion, (unsigned)SUPDRV_IDC_VERSION));
1880	pReq->u.Out.pSession = NULL;
1881	pReq->u.Out.uSessionVersion = 0xffffffff;
1882	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1883	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1884	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1885	return VINF_SUCCESS;
1886	}
1887
1888	pReq->u.Out.pSession = NULL;
1889	pReq->u.Out.uSessionVersion = SUPDRV_IDC_VERSION;
1890	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1891	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1892
1893	/*
1894	* On NT we will already have a session associated with the
1895	* client, just like with the SUP_IOCTL_COOKIE request, while
1896	* the other doesn't.
1897	*/
1898	#ifdef RT_OS_WINDOWS
1899	pReq->Hdr.rc = VINF_SUCCESS;
1900	#else
1901	AssertReturn(!pSession, VERR_INTERNAL_ERROR);
1902	pReq->Hdr.rc = supdrvCreateSession(pDevExt, false /* fUser */, &pSession);
1903	if (RT_FAILURE(pReq->Hdr.rc))
1904	{
1905	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: failed to create session, rc=%d\n", pReq->Hdr.rc));
1906	return VINF_SUCCESS;
1907	}
1908	#endif
1909
1910	pReq->u.Out.pSession = pSession;
1911	pReq->Hdr.pSession = pSession;
1912
1913	return VINF_SUCCESS;
1914	}
1915
1916	case SUPDRV_IDC_REQ_DISCONNECT:
1917	{
1918	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_DISCONNECT, sizeof(*pReqHdr));
1919
1920	#ifdef RT_OS_WINDOWS
1921	/* Windows will destroy the session when the file object is destroyed. */
1922	#else
1923	supdrvCloseSession(pDevExt, pSession);
1924	#endif
1925	return pReqHdr->rc = VINF_SUCCESS;
1926	}
1927
1928	case SUPDRV_IDC_REQ_GET_SYMBOL:
1929	{
1930	PSUPDRVIDCREQGETSYM pReq = (PSUPDRVIDCREQGETSYM)pReqHdr;
1931	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_GET_SYMBOL, sizeof(*pReq));
1932
1933	pReq->Hdr.rc = supdrvIDC_LdrGetSymbol(pDevExt, pSession, pReq);
1934	return VINF_SUCCESS;
1935	}
1936
1937	case SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY:
1938	{
1939	PSUPDRVIDCREQCOMPREGFACTORY pReq = (PSUPDRVIDCREQCOMPREGFACTORY)pReqHdr;
1940	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY, sizeof(*pReq));
1941
1942	pReq->Hdr.rc = SUPR0ComponentRegisterFactory(pSession, pReq->u.In.pFactory);
1943	return VINF_SUCCESS;
1944	}
1945
1946	case SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY:
1947	{
1948	PSUPDRVIDCREQCOMPDEREGFACTORY pReq = (PSUPDRVIDCREQCOMPDEREGFACTORY)pReqHdr;
1949	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY, sizeof(*pReq));
1950
1951	pReq->Hdr.rc = SUPR0ComponentDeregisterFactory(pSession, pReq->u.In.pFactory);
1952	return VINF_SUCCESS;
1953	}
1954
1955	default:
1956	Log(("Unknown IDC %#lx\n", (long)uReq));
1957	break;
1958	}
1959
1960	#undef REQ_CHECK_IDC_SIZE
1961	return VERR_NOT_SUPPORTED;
1962	}
1963
1964
1965	/**
1966	* Register a object for reference counting.
1967	* The object is registered with one reference in the specified session.
1968	*
1969	* @returns Unique identifier on success (pointer).
1970	* All future reference must use this identifier.
1971	* @returns NULL on failure.
1972	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
1973	* @param pvUser1 The first user argument.
1974	* @param pvUser2 The second user argument.
1975	*/
1976	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
1977	{
1978	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1979	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1980	PSUPDRVOBJ pObj;
1981	PSUPDRVUSAGE pUsage;
1982
1983	/*
1984	* Validate the input.
1985	*/
1986	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
1987	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
1988	AssertPtrReturn(pfnDestructor, NULL);
1989
1990	/*
1991	* Allocate and initialize the object.
1992	*/
1993	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
1994	if (!pObj)
1995	return NULL;
1996	pObj->u32Magic = SUPDRVOBJ_MAGIC;
1997	pObj->enmType = enmType;
1998	pObj->pNext = NULL;
1999	pObj->cUsage = 1;
2000	pObj->pfnDestructor = pfnDestructor;
2001	pObj->pvUser1 = pvUser1;
2002	pObj->pvUser2 = pvUser2;
2003	pObj->CreatorUid = pSession->Uid;
2004	pObj->CreatorGid = pSession->Gid;
2005	pObj->CreatorProcess= pSession->Process;
2006	supdrvOSObjInitCreator(pObj, pSession);
2007
2008	/*
2009	* Allocate the usage record.
2010	* (We keep freed usage records around to simplify SUPR0ObjAddRefEx().)
2011	*/
2012	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2013
2014	pUsage = pDevExt->pUsageFree;
2015	if (pUsage)
2016	pDevExt->pUsageFree = pUsage->pNext;
2017	else
2018	{
2019	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2020	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
2021	if (!pUsage)
2022	{
2023	RTMemFree(pObj);
2024	return NULL;
2025	}
2026	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2027	}
2028
2029	/*
2030	* Insert the object and create the session usage record.
2031	*/
2032	/* The object. */
2033	pObj->pNext = pDevExt->pObjs;
2034	pDevExt->pObjs = pObj;
2035
2036	/* The session record. */
2037	pUsage->cUsage = 1;
2038	pUsage->pObj = pObj;
2039	pUsage->pNext = pSession->pUsage;
2040	/* Log2(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext)); */
2041	pSession->pUsage = pUsage;
2042
2043	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2044
2045	Log(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
2046	return pObj;
2047	}
2048
2049
2050	/**
2051	* Increment the reference counter for the object associating the reference
2052	* with the specified session.
2053	*
2054	* @returns IPRT status code.
2055	* @param pvObj The identifier returned by SUPR0ObjRegister().
2056	* @param pSession The session which is referencing the object.
2057	*
2058	* @remarks The caller should not own any spinlocks and must carefully protect
2059	* itself against potential race with the destructor so freed memory
2060	* isn't accessed here.
2061	*/
2062	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
2063	{
2064	return SUPR0ObjAddRefEx(pvObj, pSession, false /* fNoBlocking */);
2065	}
2066
2067
2068	/**
2069	* Increment the reference counter for the object associating the reference
2070	* with the specified session.
2071	*
2072	* @returns IPRT status code.
2073	* @retval VERR_TRY_AGAIN if fNoBlocking was set and a new usage record
2074	* couldn't be allocated. (If you see this you're not doing the right
2075	* thing and it won't ever work reliably.)
2076	*
2077	* @param pvObj The identifier returned by SUPR0ObjRegister().
2078	* @param pSession The session which is referencing the object.
2079	* @param fNoBlocking Set if it's not OK to block. Never try to make the
2080	* first reference to an object in a session with this
2081	* argument set.
2082	*
2083	* @remarks The caller should not own any spinlocks and must carefully protect
2084	* itself against potential race with the destructor so freed memory
2085	* isn't accessed here.
2086	*/
2087	SUPR0DECL(int) SUPR0ObjAddRefEx(void *pvObj, PSUPDRVSESSION pSession, bool fNoBlocking)
2088	{
2089	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2090	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2091	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2092	int rc = VINF_SUCCESS;
2093	PSUPDRVUSAGE pUsagePre;
2094	PSUPDRVUSAGE pUsage;
2095
2096	/*
2097	* Validate the input.
2098	* Be ready for the destruction race (someone might be stuck in the
2099	* destructor waiting a lock we own).
2100	*/
2101	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2102	AssertPtrReturn(pObj, VERR_INVALID_POINTER);
2103	AssertMsgReturn(pObj->u32Magic == SUPDRVOBJ_MAGIC \|\| pObj->u32Magic == SUPDRVOBJ_MAGIC_DEAD,
2104	("Invalid pvObj=%p magic=%#x (expected %#x or %#x)\n", pvObj, pObj->u32Magic, SUPDRVOBJ_MAGIC, SUPDRVOBJ_MAGIC_DEAD),
2105	VERR_INVALID_PARAMETER);
2106
2107	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2108
2109	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2110	{
2111	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2112
2113	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2114	return VERR_WRONG_ORDER;
2115	}
2116
2117	/*
2118	* Preallocate the usage record if we can.
2119	*/
2120	pUsagePre = pDevExt->pUsageFree;
2121	if (pUsagePre)
2122	pDevExt->pUsageFree = pUsagePre->pNext;
2123	else if (!fNoBlocking)
2124	{
2125	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2126	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
2127	if (!pUsagePre)
2128	return VERR_NO_MEMORY;
2129
2130	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2131	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2132	{
2133	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2134
2135	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2136	return VERR_WRONG_ORDER;
2137	}
2138	}
2139
2140	/*
2141	* Reference the object.
2142	*/
2143	pObj->cUsage++;
2144
2145	/*
2146	* Look for the session record.
2147	*/
2148	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
2149	{
2150	/Log(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2151	if (pUsage->pObj == pObj)
2152	break;
2153	}
2154	if (pUsage)
2155	pUsage->cUsage++;
2156	else if (pUsagePre)
2157	{
2158	/* create a new session record. */
2159	pUsagePre->cUsage = 1;
2160	pUsagePre->pObj = pObj;
2161	pUsagePre->pNext = pSession->pUsage;
2162	pSession->pUsage = pUsagePre;
2163	/Log(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));/
2164
2165	pUsagePre = NULL;
2166	}
2167	else
2168	{
2169	pObj->cUsage--;
2170	rc = VERR_TRY_AGAIN;
2171	}
2172
2173	/*
2174	* Put any unused usage record into the free list..
2175	*/
2176	if (pUsagePre)
2177	{
2178	pUsagePre->pNext = pDevExt->pUsageFree;
2179	pDevExt->pUsageFree = pUsagePre;
2180	}
2181
2182	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2183
2184	return rc;
2185	}
2186
2187
2188	/**
2189	* Decrement / destroy a reference counter record for an object.
2190	*
2191	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
2192	*
2193	* @returns IPRT status code.
2194	* @retval VINF_SUCCESS if not destroyed.
2195	* @retval VINF_OBJECT_DESTROYED if it's destroyed by this release call.
2196	* @retval VERR_INVALID_PARAMETER if the object isn't valid. Will assert in
2197	* string builds.
2198	*
2199	* @param pvObj The identifier returned by SUPR0ObjRegister().
2200	* @param pSession The session which is referencing the object.
2201	*/
2202	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
2203	{
2204	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2205	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2206	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2207	int rc = VERR_INVALID_PARAMETER;
2208	PSUPDRVUSAGE pUsage;
2209	PSUPDRVUSAGE pUsagePrev;
2210
2211	/*
2212	* Validate the input.
2213	*/
2214	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2215	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2216	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2217	VERR_INVALID_PARAMETER);
2218
2219	/*
2220	* Acquire the spinlock and look for the usage record.
2221	*/
2222	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2223
2224	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
2225	pUsage;
2226	pUsagePrev = pUsage, pUsage = pUsage->pNext)
2227	{
2228	/Log2(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2229	if (pUsage->pObj == pObj)
2230	{
2231	rc = VINF_SUCCESS;
2232	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
2233	if (pUsage->cUsage > 1)
2234	{
2235	pObj->cUsage--;
2236	pUsage->cUsage--;
2237	}
2238	else
2239	{
2240	/*
2241	* Free the session record.
2242	*/
2243	if (pUsagePrev)
2244	pUsagePrev->pNext = pUsage->pNext;
2245	else
2246	pSession->pUsage = pUsage->pNext;
2247	pUsage->pNext = pDevExt->pUsageFree;
2248	pDevExt->pUsageFree = pUsage;
2249
2250	/* What about the object? */
2251	if (pObj->cUsage > 1)
2252	pObj->cUsage--;
2253	else
2254	{
2255	/*
2256	* Object is to be destroyed, unlink it.
2257	*/
2258	pObj->u32Magic = SUPDRVOBJ_MAGIC_DEAD;
2259	rc = VINF_OBJECT_DESTROYED;
2260	if (pDevExt->pObjs == pObj)
2261	pDevExt->pObjs = pObj->pNext;
2262	else
2263	{
2264	PSUPDRVOBJ pObjPrev;
2265	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
2266	if (pObjPrev->pNext == pObj)
2267	{
2268	pObjPrev->pNext = pObj->pNext;
2269	break;
2270	}
2271	Assert(pObjPrev);
2272	}
2273	}
2274	}
2275	break;
2276	}
2277	}
2278
2279	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2280
2281	/*
2282	* Call the destructor and free the object if required.
2283	*/
2284	if (rc == VINF_OBJECT_DESTROYED)
2285	{
2286	Log(("SUPR0ObjRelease: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
2287	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
2288	if (pObj->pfnDestructor)
2289	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
2290	RTMemFree(pObj);
2291	}
2292
2293	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
2294	return rc;
2295	}
2296
2297
2298	/**
2299	* Verifies that the current process can access the specified object.
2300	*
2301	* @returns The following IPRT status code:
2302	* @retval VINF_SUCCESS if access was granted.
2303	* @retval VERR_PERMISSION_DENIED if denied access.
2304	* @retval VERR_INVALID_PARAMETER if invalid parameter.
2305	*
2306	* @param pvObj The identifier returned by SUPR0ObjRegister().
2307	* @param pSession The session which wishes to access the object.
2308	* @param pszObjName Object string name. This is optional and depends on the object type.
2309	*
2310	* @remark The caller is responsible for making sure the object isn't removed while
2311	* we're inside this function. If uncertain about this, just call AddRef before calling us.
2312	*/
2313	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
2314	{
2315	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2316	int rc;
2317
2318	/*
2319	* Validate the input.
2320	*/
2321	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2322	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2323	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2324	VERR_INVALID_PARAMETER);
2325
2326	/*
2327	* Check access. (returns true if a decision has been made.)
2328	*/
2329	rc = VERR_INTERNAL_ERROR;
2330	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
2331	return rc;
2332
2333	/*
2334	* Default policy is to allow the user to access his own
2335	* stuff but nothing else.
2336	*/
2337	if (pObj->CreatorUid == pSession->Uid)
2338	return VINF_SUCCESS;
2339	return VERR_PERMISSION_DENIED;
2340	}
2341
2342
2343	/**
2344	* Lock pages.
2345	*
2346	* @returns IPRT status code.
2347	* @param pSession Session to which the locked memory should be associated.
2348	* @param pvR3 Start of the memory range to lock.
2349	* This must be page aligned.
2350	* @param cPages Number of pages to lock.
2351	* @param paPages Where to put the physical addresses of locked memory.
2352	*/
2353	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
2354	{
2355	int rc;
2356	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2357	const size_t cb = (size_t)cPages << PAGE_SHIFT;
2358	LogFlow(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
2359
2360	/*
2361	* Verify input.
2362	*/
2363	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2364	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
2365	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
2366	\|\| !pvR3)
2367	{
2368	Log(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
2369	return VERR_INVALID_PARAMETER;
2370	}
2371
2372	/*
2373	* Let IPRT do the job.
2374	*/
2375	Mem.eType = MEMREF_TYPE_LOCKED;
2376	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTMEM_PROT_READ \| RTMEM_PROT_WRITE, RTR0ProcHandleSelf());
2377	if (RT_SUCCESS(rc))
2378	{
2379	uint32_t iPage = cPages;
2380	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
2381	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
2382
2383	while (iPage-- > 0)
2384	{
2385	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2386	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
2387	{
2388	AssertMsgFailed(("iPage=%d\n", iPage));
2389	rc = VERR_INTERNAL_ERROR;
2390	break;
2391	}
2392	}
2393	if (RT_SUCCESS(rc))
2394	rc = supdrvMemAdd(&Mem, pSession);
2395	if (RT_FAILURE(rc))
2396	{
2397	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
2398	AssertRC(rc2);
2399	}
2400	}
2401
2402	return rc;
2403	}
2404
2405
2406	/**
2407	* Unlocks the memory pointed to by pv.
2408	*
2409	* @returns IPRT status code.
2410	* @param pSession Session to which the memory was locked.
2411	* @param pvR3 Memory to unlock.
2412	*/
2413	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2414	{
2415	LogFlow(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2416	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2417	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
2418	}
2419
2420
2421	/**
2422	* Allocates a chunk of page aligned memory with contiguous and fixed physical
2423	* backing.
2424	*
2425	* @returns IPRT status code.
2426	* @param pSession Session data.
2427	* @param cPages Number of pages to allocate.
2428	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
2429	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
2430	* @param pHCPhys Where to put the physical address of allocated memory.
2431	*/
2432	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
2433	{
2434	int rc;
2435	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2436	LogFlow(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
2437
2438	/*
2439	* Validate input.
2440	*/
2441	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2442	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
2443	{
2444	Log(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
2445	pSession, ppvR0, ppvR3, pHCPhys));
2446	return VERR_INVALID_PARAMETER;
2447
2448	}
2449	if (cPages < 1 \|\| cPages >= 256)
2450	{
2451	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2452	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2453	}
2454
2455	/*
2456	* Let IPRT do the job.
2457	*/
2458	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
2459	if (RT_SUCCESS(rc))
2460	{
2461	int rc2;
2462	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2463	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2464	if (RT_SUCCESS(rc))
2465	{
2466	Mem.eType = MEMREF_TYPE_CONT;
2467	rc = supdrvMemAdd(&Mem, pSession);
2468	if (!rc)
2469	{
2470	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2471	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2472	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
2473	return 0;
2474	}
2475
2476	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2477	AssertRC(rc2);
2478	}
2479	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2480	AssertRC(rc2);
2481	}
2482
2483	return rc;
2484	}
2485
2486
2487	/**
2488	* Frees memory allocated using SUPR0ContAlloc().
2489	*
2490	* @returns IPRT status code.
2491	* @param pSession The session to which the memory was allocated.
2492	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2493	*/
2494	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2495	{
2496	LogFlow(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2497	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2498	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
2499	}
2500
2501
2502	/**
2503	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
2504	*
2505	* The memory isn't zeroed.
2506	*
2507	* @returns IPRT status code.
2508	* @param pSession Session data.
2509	* @param cPages Number of pages to allocate.
2510	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
2511	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
2512	* @param paPages Where to put the physical addresses of allocated memory.
2513	*/
2514	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
2515	{
2516	unsigned iPage;
2517	int rc;
2518	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2519	LogFlow(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
2520
2521	/*
2522	* Validate input.
2523	*/
2524	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2525	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
2526	{
2527	Log(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
2528	pSession, ppvR3, ppvR0, paPages));
2529	return VERR_INVALID_PARAMETER;
2530
2531	}
2532	if (cPages < 1 \|\| cPages >= 256)
2533	{
2534	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2535	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2536	}
2537
2538	/*
2539	* Let IPRT do the work.
2540	*/
2541	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
2542	if (RT_SUCCESS(rc))
2543	{
2544	int rc2;
2545	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2546	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2547	if (RT_SUCCESS(rc))
2548	{
2549	Mem.eType = MEMREF_TYPE_LOW;
2550	rc = supdrvMemAdd(&Mem, pSession);
2551	if (!rc)
2552	{
2553	for (iPage = 0; iPage < cPages; iPage++)
2554	{
2555	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2556	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%RHp\n", paPages[iPage]));
2557	}
2558	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2559	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2560	return 0;
2561	}
2562
2563	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2564	AssertRC(rc2);
2565	}
2566
2567	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2568	AssertRC(rc2);
2569	}
2570
2571	return rc;
2572	}
2573
2574
2575	/**
2576	* Frees memory allocated using SUPR0LowAlloc().
2577	*
2578	* @returns IPRT status code.
2579	* @param pSession The session to which the memory was allocated.
2580	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2581	*/
2582	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2583	{
2584	LogFlow(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2585	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2586	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
2587	}
2588
2589
2590
2591	/**
2592	* Allocates a chunk of memory with both R0 and R3 mappings.
2593	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
2594	*
2595	* @returns IPRT status code.
2596	* @param pSession The session to associated the allocation with.
2597	* @param cb Number of bytes to allocate.
2598	* @param ppvR0 Where to store the address of the Ring-0 mapping.
2599	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2600	*/
2601	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
2602	{
2603	int rc;
2604	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2605	LogFlow(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
2606
2607	/*
2608	* Validate input.
2609	*/
2610	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2611	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
2612	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
2613	if (cb < 1 \|\| cb >= _4M)
2614	{
2615	Log(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
2616	return VERR_INVALID_PARAMETER;
2617	}
2618
2619	/*
2620	* Let IPRT do the work.
2621	*/
2622	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
2623	if (RT_SUCCESS(rc))
2624	{
2625	int rc2;
2626	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2627	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2628	if (RT_SUCCESS(rc))
2629	{
2630	Mem.eType = MEMREF_TYPE_MEM;
2631	rc = supdrvMemAdd(&Mem, pSession);
2632	if (!rc)
2633	{
2634	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2635	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2636	return VINF_SUCCESS;
2637	}
2638
2639	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2640	AssertRC(rc2);
2641	}
2642
2643	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2644	AssertRC(rc2);
2645	}
2646
2647	return rc;
2648	}
2649
2650
2651	/**
2652	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
2653	*
2654	* @returns IPRT status code.
2655	* @param pSession The session to which the memory was allocated.
2656	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2657	* @param paPages Where to store the physical addresses.
2658	*/
2659	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
2660	{
2661	PSUPDRVBUNDLE pBundle;
2662	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2663	LogFlow(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
2664
2665	/*
2666	* Validate input.
2667	*/
2668	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2669	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
2670	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
2671
2672	/*
2673	* Search for the address.
2674	*/
2675	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2676	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2677	{
2678	if (pBundle->cUsed > 0)
2679	{
2680	unsigned i;
2681	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2682	{
2683	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
2684	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2685	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2686	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2687	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
2688	)
2689	)
2690	{
2691	const size_t cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2692	size_t iPage;
2693	for (iPage = 0; iPage < cPages; iPage++)
2694	{
2695	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2696	paPages[iPage].uReserved = 0;
2697	}
2698	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2699	return VINF_SUCCESS;
2700	}
2701	}
2702	}
2703	}
2704	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2705	Log(("Failed to find %p!!!\n", (void *)uPtr));
2706	return VERR_INVALID_PARAMETER;
2707	}
2708
2709
2710	/**
2711	* Free memory allocated by SUPR0MemAlloc().
2712	*
2713	* @returns IPRT status code.
2714	* @param pSession The session owning the allocation.
2715	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2716	*/
2717	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2718	{
2719	LogFlow(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2720	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2721	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
2722	}
2723
2724
2725	/**
2726	* Allocates a chunk of memory with a kernel or/and a user mode mapping.
2727	*
2728	* The memory is fixed and it's possible to query the physical addresses using
2729	* SUPR0MemGetPhys().
2730	*
2731	* @returns IPRT status code.
2732	* @param pSession The session to associated the allocation with.
2733	* @param cPages The number of pages to allocate.
2734	* @param fFlags Flags, reserved for the future. Must be zero.
2735	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2736	* NULL if no ring-3 mapping.
2737	* @param ppvR3 Where to store the address of the Ring-0 mapping.
2738	* NULL if no ring-0 mapping.
2739	* @param paPages Where to store the addresses of the pages. Optional.
2740	*/
2741	SUPR0DECL(int) SUPR0PageAllocEx(PSUPDRVSESSION pSession, uint32_t cPages, uint32_t fFlags, PRTR3PTR ppvR3, PRTR0PTR ppvR0, PRTHCPHYS paPages)
2742	{
2743	int rc;
2744	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2745	LogFlow(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
2746
2747	/*
2748	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2749	*/
2750	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2751	AssertPtrNullReturn(ppvR3, VERR_INVALID_POINTER);
2752	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2753	AssertReturn(ppvR3 \|\| ppvR0, VERR_INVALID_PARAMETER);
2754	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2755	if (cPages < 1 \|\| cPages > VBOX_MAX_ALLOC_PAGE_COUNT)
2756	{
2757	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)));
2758	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2759	}
2760
2761	/*
2762	* Let IPRT do the work.
2763	*/
2764	if (ppvR0)
2765	rc = RTR0MemObjAllocPage(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, true /* fExecutable */);
2766	else
2767	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
2768	if (RT_SUCCESS(rc))
2769	{
2770	int rc2;
2771	if (ppvR3)
2772	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2773	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2774	else
2775	Mem.MapObjR3 = NIL_RTR0MEMOBJ;
2776	if (RT_SUCCESS(rc))
2777	{
2778	Mem.eType = MEMREF_TYPE_PAGE;
2779	rc = supdrvMemAdd(&Mem, pSession);
2780	if (!rc)
2781	{
2782	if (ppvR3)
2783	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2784	if (ppvR0)
2785	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2786	if (paPages)
2787	{
2788	uint32_t iPage = cPages;
2789	while (iPage-- > 0)
2790	{
2791	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
2792	Assert(paPages[iPage] != NIL_RTHCPHYS);
2793	}
2794	}
2795	return VINF_SUCCESS;
2796	}
2797
2798	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2799	AssertRC(rc2);
2800	}
2801
2802	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2803	AssertRC(rc2);
2804	}
2805	return rc;
2806	}
2807
2808
2809	/**
2810	* Maps a chunk of memory previously allocated by SUPR0PageAllocEx into kernel
2811	* space.
2812	*
2813	* @returns IPRT status code.
2814	* @param pSession The session to associated the allocation with.
2815	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
2816	* @param offSub Where to start mapping. Must be page aligned.
2817	* @param cbSub How much to map. Must be page aligned.
2818	* @param fFlags Flags, MBZ.
2819	* @param ppvR0 Where to return the address of the ring-0 mapping on
2820	* success.
2821	*/
2822	SUPR0DECL(int) SUPR0PageMapKernel(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t offSub, uint32_t cbSub,
2823	uint32_t fFlags, PRTR0PTR ppvR0)
2824	{
2825	int rc;
2826	PSUPDRVBUNDLE pBundle;
2827	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2828	RTR0MEMOBJ hMemObj = NIL_RTR0MEMOBJ;
2829	LogFlow(("SUPR0PageMapKernel: pSession=%p pvR3=%p offSub=%#x cbSub=%#x\n", pSession, pvR3, offSub, cbSub));
2830
2831	/*
2832	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2833	*/
2834	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2835	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2836	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2837	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2838	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2839	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
2840
2841	/*
2842	* Find the memory object.
2843	*/
2844	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2845	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2846	{
2847	if (pBundle->cUsed > 0)
2848	{
2849	unsigned i;
2850	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2851	{
2852	if ( ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
2853	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2854	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2855	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
2856	\|\| ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED
2857	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2858	&& pBundle->aMem[i].MapObjR3 == NIL_RTR0MEMOBJ
2859	&& RTR0MemObjAddressR3(pBundle->aMem[i].MemObj) == pvR3))
2860	{
2861	hMemObj = pBundle->aMem[i].MemObj;
2862	break;
2863	}
2864	}
2865	}
2866	}
2867	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2868
2869	rc = VERR_INVALID_PARAMETER;
2870	if (hMemObj != NIL_RTR0MEMOBJ)
2871	{
2872	/*
2873	* Do some further input validations before calling IPRT.
2874	* (Cleanup is done indirectly by telling RTR0MemObjFree to include mappings.)
2875	*/
2876	size_t cbMemObj = RTR0MemObjSize(hMemObj);
2877	if ( offSub < cbMemObj
2878	&& cbSub <= cbMemObj
2879	&& offSub + cbSub <= cbMemObj)
2880	{
2881	RTR0MEMOBJ hMapObj;
2882	rc = RTR0MemObjMapKernelEx(&hMapObj, hMemObj, (void *)-1, 0,
2883	RTMEM_PROT_READ \| RTMEM_PROT_WRITE, offSub, cbSub);
2884	if (RT_SUCCESS(rc))
2885	*ppvR0 = RTR0MemObjAddress(hMapObj);
2886	}
2887	else
2888	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
2889
2890	}
2891	return rc;
2892	}
2893
2894
2895	/**
2896	* Changes the page level protection of one or more pages previously allocated
2897	* by SUPR0PageAllocEx.
2898	*
2899	* @returns IPRT status code.
2900	* @param pSession The session to associated the allocation with.
2901	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
2902	* NIL_RTR3PTR if the ring-3 mapping should be unaffected.
2903	* @param pvR0 The ring-0 address returned by SUPR0PageAllocEx.
2904	* NIL_RTR0PTR if the ring-0 mapping should be unaffected.
2905	* @param offSub Where to start changing. Must be page aligned.
2906	* @param cbSub How much to change. Must be page aligned.
2907	* @param fProt The new page level protection, see RTMEM_PROT_*.
2908	*/
2909	SUPR0DECL(int) SUPR0PageProtect(PSUPDRVSESSION pSession, RTR3PTR pvR3, RTR0PTR pvR0, uint32_t offSub, uint32_t cbSub, uint32_t fProt)
2910	{
2911	int rc;
2912	PSUPDRVBUNDLE pBundle;
2913	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2914	RTR0MEMOBJ hMemObjR0 = NIL_RTR0MEMOBJ;
2915	RTR0MEMOBJ hMemObjR3 = NIL_RTR0MEMOBJ;
2916	LogFlow(("SUPR0PageProtect: pSession=%p pvR3=%p pvR0=%p offSub=%#x cbSub=%#x fProt-%#x\n", pSession, pvR3, pvR0, offSub, cbSub, fProt));
2917
2918	/*
2919	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2920	*/
2921	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2922	AssertReturn(!(fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)), VERR_INVALID_PARAMETER);
2923	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2924	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2925	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
2926
2927	/*
2928	* Find the memory object.
2929	*/
2930	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2931	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2932	{
2933	if (pBundle->cUsed > 0)
2934	{
2935	unsigned i;
2936	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2937	{
2938	if ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
2939	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2940	&& ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2941	\|\| pvR3 == NIL_RTR3PTR)
2942	&& ( pvR0 == NIL_RTR0PTR
2943	\|\| RTR0MemObjAddress(pBundle->aMem[i].MemObj) == pvR0)
2944	&& ( pvR3 == NIL_RTR3PTR
2945	\|\| RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3))
2946	{
2947	if (pvR0 != NIL_RTR0PTR)
2948	hMemObjR0 = pBundle->aMem[i].MemObj;
2949	if (pvR3 != NIL_RTR3PTR)
2950	hMemObjR3 = pBundle->aMem[i].MapObjR3;
2951	break;
2952	}
2953	}
2954	}
2955	}
2956	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2957
2958	rc = VERR_INVALID_PARAMETER;
2959	if ( hMemObjR0 != NIL_RTR0MEMOBJ
2960	\|\| hMemObjR3 != NIL_RTR0MEMOBJ)
2961	{
2962	/*
2963	* Do some further input validations before calling IPRT.
2964	*/
2965	size_t cbMemObj = hMemObjR0 != NIL_RTR0PTR ? RTR0MemObjSize(hMemObjR0) : RTR0MemObjSize(hMemObjR3);
2966	if ( offSub < cbMemObj
2967	&& cbSub <= cbMemObj
2968	&& offSub + cbSub <= cbMemObj)
2969	{
2970	rc = VINF_SUCCESS;
2971	if (hMemObjR3 != NIL_RTR0PTR)
2972	rc = RTR0MemObjProtect(hMemObjR3, offSub, cbSub, fProt);
2973	if (hMemObjR0 != NIL_RTR0PTR && RT_SUCCESS(rc))
2974	rc = RTR0MemObjProtect(hMemObjR0, offSub, cbSub, fProt);
2975	}
2976	else
2977	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
2978
2979	}
2980	return rc;
2981
2982	}
2983
2984
2985	/**
2986	* Free memory allocated by SUPR0PageAlloc() and SUPR0PageAllocEx().
2987	*
2988	* @returns IPRT status code.
2989	* @param pSession The session owning the allocation.
2990	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc() or
2991	* SUPR0PageAllocEx().
2992	*/
2993	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2994	{
2995	LogFlow(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2996	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2997	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_PAGE);
2998	}
2999
3000
3001	/**
3002	* Gets the paging mode of the current CPU.
3003	*
3004	* @returns Paging mode, SUPPAGEINGMODE_INVALID on error.
3005	*/
3006	SUPR0DECL(SUPPAGINGMODE) SUPR0GetPagingMode(void)
3007	{
3008	SUPPAGINGMODE enmMode;
3009
3010	RTR0UINTREG cr0 = ASMGetCR0();
3011	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
3012	enmMode = SUPPAGINGMODE_INVALID;
3013	else
3014	{
3015	RTR0UINTREG cr4 = ASMGetCR4();
3016	uint32_t fNXEPlusLMA = 0;
3017	if (cr4 & X86_CR4_PAE)
3018	{
3019	uint32_t fAmdFeatures = ASMCpuId_EDX(0x80000001);
3020	if (fAmdFeatures & (X86_CPUID_AMD_FEATURE_EDX_NX \| X86_CPUID_AMD_FEATURE_EDX_LONG_MODE))
3021	{
3022	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
3023	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
3024	fNXEPlusLMA \|= RT_BIT(0);
3025	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
3026	fNXEPlusLMA \|= RT_BIT(1);
3027	}
3028	}
3029
3030	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
3031	{
3032	case 0:
3033	enmMode = SUPPAGINGMODE_32_BIT;
3034	break;
3035
3036	case X86_CR4_PGE:
3037	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
3038	break;
3039
3040	case X86_CR4_PAE:
3041	enmMode = SUPPAGINGMODE_PAE;
3042	break;
3043
3044	case X86_CR4_PAE \| RT_BIT(0):
3045	enmMode = SUPPAGINGMODE_PAE_NX;
3046	break;
3047
3048	case X86_CR4_PAE \| X86_CR4_PGE:
3049	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3050	break;
3051
3052	case X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3053	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3054	break;
3055
3056	case RT_BIT(1) \| X86_CR4_PAE:
3057	enmMode = SUPPAGINGMODE_AMD64;
3058	break;
3059
3060	case RT_BIT(1) \| X86_CR4_PAE \| RT_BIT(0):
3061	enmMode = SUPPAGINGMODE_AMD64_NX;
3062	break;
3063
3064	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
3065	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
3066	break;
3067
3068	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3069	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
3070	break;
3071
3072	default:
3073	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
3074	enmMode = SUPPAGINGMODE_INVALID;
3075	break;
3076	}
3077	}
3078	return enmMode;
3079	}
3080
3081
3082	/**
3083	* Enables or disabled hardware virtualization extensions using native OS APIs.
3084	*
3085	* @returns VBox status code.
3086	* @retval VINF_SUCCESS on success.
3087	* @retval VERR_NOT_SUPPORTED if not supported by the native OS.
3088	*
3089	* @param fEnable Whether to enable or disable.
3090	*/
3091	SUPR0DECL(int) SUPR0EnableVTx(bool fEnable)
3092	{
3093	#ifdef RT_OS_DARWIN
3094	return supdrvOSEnableVTx(fEnable);
3095	#else
3096	return VERR_NOT_SUPPORTED;
3097	#endif
3098	}
3099
3100
3101	/** @todo document me */
3102	SUPR0DECL(int) SUPR0QueryVTCaps(PSUPDRVSESSION pSession, uint32_t *pfCaps)
3103	{
3104	/*
3105	* Input validation.
3106	*/
3107	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3108	AssertPtrReturn(pfCaps, VERR_INVALID_POINTER);
3109
3110	*pfCaps = 0;
3111
3112	if (ASMHasCpuId())
3113	{
3114	uint32_t u32FeaturesECX;
3115	uint32_t u32Dummy;
3116	uint32_t u32FeaturesEDX;
3117	uint32_t u32VendorEBX, u32VendorECX, u32VendorEDX, u32AMDFeatureEDX, u32AMDFeatureECX;
3118	uint64_t val;
3119
3120	ASMCpuId(0, &u32Dummy, &u32VendorEBX, &u32VendorECX, &u32VendorEDX);
3121	ASMCpuId(1, &u32Dummy, &u32Dummy, &u32FeaturesECX, &u32FeaturesEDX);
3122	/* Query AMD features. */
3123	ASMCpuId(0x80000001, &u32Dummy, &u32Dummy, &u32AMDFeatureECX, &u32AMDFeatureEDX);
3124
3125	if ( u32VendorEBX == X86_CPUID_VENDOR_INTEL_EBX
3126	&& u32VendorECX == X86_CPUID_VENDOR_INTEL_ECX
3127	&& u32VendorEDX == X86_CPUID_VENDOR_INTEL_EDX
3128	)
3129	{
3130	if ( (u32FeaturesECX & X86_CPUID_FEATURE_ECX_VMX)
3131	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3132	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3133	)
3134	{
3135	val = ASMRdMsr(MSR_IA32_FEATURE_CONTROL);
3136	/*
3137	* Both the LOCK and VMXON bit must be set; otherwise VMXON will generate a #GP.
3138	* Once the lock bit is set, this MSR can no longer be modified.
3139	*/
3140	if ( (val & (MSR_IA32_FEATURE_CONTROL_VMXON\|MSR_IA32_FEATURE_CONTROL_LOCK))
3141	== (MSR_IA32_FEATURE_CONTROL_VMXON\|MSR_IA32_FEATURE_CONTROL_LOCK) /* enabled and locked */
3142	\|\| !(val & MSR_IA32_FEATURE_CONTROL_LOCK) /* not enabled, but not locked either */
3143	)
3144	{
3145	VMX_CAPABILITY vtCaps;
3146
3147	*pfCaps \|= SUPVTCAPS_VT_X;
3148
3149	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS);
3150	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC_USE_SECONDARY_EXEC_CTRL)
3151	{
3152	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS2);
3153	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC2_EPT)
3154	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3155	}
3156	return VINF_SUCCESS;
3157	}
3158	return VERR_VMX_MSR_LOCKED_OR_DISABLED;
3159	}
3160	return VERR_VMX_NO_VMX;
3161	}
3162
3163	if ( u32VendorEBX == X86_CPUID_VENDOR_AMD_EBX
3164	&& u32VendorECX == X86_CPUID_VENDOR_AMD_ECX
3165	&& u32VendorEDX == X86_CPUID_VENDOR_AMD_EDX
3166	)
3167	{
3168	if ( (u32AMDFeatureECX & X86_CPUID_AMD_FEATURE_ECX_SVM)
3169	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3170	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3171	)
3172	{
3173	/* Check if SVM is disabled */
3174	val = ASMRdMsr(MSR_K8_VM_CR);
3175	if (!(val & MSR_K8_VM_CR_SVM_DISABLE))
3176	{
3177	*pfCaps \|= SUPVTCAPS_AMD_V;
3178
3179	/* Query AMD features. */
3180	ASMCpuId(0x8000000A, &u32Dummy, &u32Dummy, &u32Dummy, &u32FeaturesEDX);
3181
3182	if (u32FeaturesEDX & AMD_CPUID_SVM_FEATURE_EDX_NESTED_PAGING)
3183	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3184
3185	return VINF_SUCCESS;
3186	}
3187	return VERR_SVM_DISABLED;
3188	}
3189	return VERR_SVM_NO_SVM;
3190	}
3191	}
3192
3193	return VERR_UNSUPPORTED_CPU;
3194	}
3195
3196
3197	/**
3198	* (Re-)initializes the per-cpu structure prior to starting or resuming the GIP
3199	* updating.
3200	*
3201	* @param pGipCpu The per CPU structure for this CPU.
3202	* @param u64NanoTS The current time.
3203	*/
3204	static void supdrvGipReInitCpu(PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
3205	{
3206	pGipCpu->u64TSC = ASMReadTSC() - pGipCpu->u32UpdateIntervalTSC;
3207	pGipCpu->u64NanoTS = u64NanoTS;
3208	}
3209
3210
3211	/**
3212	* Set the current TSC and NanoTS value for the CPU.
3213	*
3214	* @param idCpu The CPU ID. Unused - we have to use the APIC ID.
3215	* @param pvUser1 Pointer to the ring-0 GIP mapping.
3216	* @param pvUser2 Pointer to the variable holding the current time.
3217	*/
3218	static DECLCALLBACK(void) supdrvGipReInitCpuCallback(RTCPUID idCpu, void pvUser1, void pvUser2)
3219	{
3220	PSUPGLOBALINFOPAGE pGip = (PSUPGLOBALINFOPAGE)pvUser1;
3221	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
3222
3223	if (RT_LIKELY(iCpu < pGip->cCpus && pGip->aCPUs[iCpu].idCpu == idCpu))
3224	supdrvGipReInitCpu(&pGip->aCPUs[iCpu], (uint64_t )pvUser2);
3225
3226	NOREF(pvUser2);
3227	NOREF(idCpu);
3228	}
3229
3230
3231	/**
3232	* Maps the GIP into userspace and/or get the physical address of the GIP.
3233	*
3234	* @returns IPRT status code.
3235	* @param pSession Session to which the GIP mapping should belong.
3236	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
3237	* @param pHCPhysGip Where to store the physical address. (optional)
3238	*
3239	* @remark There is no reference counting on the mapping, so one call to this function
3240	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
3241	* and remove the session as a GIP user.
3242	*/
3243	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
3244	{
3245	int rc;
3246	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3247	RTR3PTR pGipR3 = NIL_RTR3PTR;
3248	RTHCPHYS HCPhys = NIL_RTHCPHYS;
3249	LogFlow(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
3250
3251	/*
3252	* Validate
3253	*/
3254	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3255	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
3256	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
3257
3258	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3259	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3260	#else
3261	RTSemFastMutexRequest(pDevExt->mtxGip);
3262	#endif
3263	if (pDevExt->pGip)
3264	{
3265	/*
3266	* Map it?
3267	*/
3268	rc = VINF_SUCCESS;
3269	if (ppGipR3)
3270	{
3271	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
3272	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
3273	RTMEM_PROT_READ, RTR0ProcHandleSelf());
3274	if (RT_SUCCESS(rc))
3275	pGipR3 = RTR0MemObjAddressR3(pSession->GipMapObjR3);
3276	}
3277
3278	/*
3279	* Get physical address.
3280	*/
3281	if (pHCPhysGip && RT_SUCCESS(rc))
3282	HCPhys = pDevExt->HCPhysGip;
3283
3284	/*
3285	* Reference globally.
3286	*/
3287	if (!pSession->fGipReferenced && RT_SUCCESS(rc))
3288	{
3289	pSession->fGipReferenced = 1;
3290	pDevExt->cGipUsers++;
3291	if (pDevExt->cGipUsers == 1)
3292	{
3293	PSUPGLOBALINFOPAGE pGipR0 = pDevExt->pGip;
3294	uint64_t u64NanoTS;
3295	uint32_t u32SystemResolution;
3296	unsigned i;
3297
3298	LogFlow(("SUPR0GipMap: Resumes GIP updating\n"));
3299
3300	/*
3301	* Try bump up the system timer resolution.
3302	* The more interrupts the better...
3303	*/
3304	if ( RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
3305	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
3306	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1953125 /* 512 HZ */, &u32SystemResolution))
3307	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 2000000 /* 500 HZ */, &u32SystemResolution))
3308	)
3309	{
3310	Assert(RTTimerGetSystemGranularity() <= u32SystemResolution);
3311	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
3312	}
3313
3314	if (pGipR0->aCPUs[0].u32TransactionId != 2 /* not the first time */)
3315	{
3316	for (i = 0; i < RT_ELEMENTS(pGipR0->aCPUs); i++)
3317	ASMAtomicUoWriteU32(&pGipR0->aCPUs[i].u32TransactionId,
3318	(pGipR0->aCPUs[i].u32TransactionId + GIP_UPDATEHZ_RECALC_FREQ * 2)
3319	& ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
3320	ASMAtomicWriteU64(&pGipR0->u64NanoTSLastUpdateHz, 0);
3321	}
3322
3323	u64NanoTS = RTTimeSystemNanoTS() - pGipR0->u32UpdateIntervalNS;
3324	if ( pGipR0->u32Mode == SUPGIPMODE_SYNC_TSC
3325	\|\| RTMpGetOnlineCount() == 1)
3326	supdrvGipReInitCpu(&pGipR0->aCPUs[0], u64NanoTS);
3327	else
3328	RTMpOnAll(supdrvGipReInitCpuCallback, pGipR0, &u64NanoTS);
3329
3330	#ifndef DO_NOT_START_GIP
3331	rc = RTTimerStart(pDevExt->pGipTimer, 0); AssertRC(rc);
3332	#endif
3333	rc = VINF_SUCCESS;
3334	}
3335	}
3336	}
3337	else
3338	{
3339	rc = VERR_GENERAL_FAILURE;
3340	Log(("SUPR0GipMap: GIP is not available!\n"));
3341	}
3342	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3343	RTSemMutexRelease(pDevExt->mtxGip);
3344	#else
3345	RTSemFastMutexRelease(pDevExt->mtxGip);
3346	#endif
3347
3348	/*
3349	* Write returns.
3350	*/
3351	if (pHCPhysGip)
3352	*pHCPhysGip = HCPhys;
3353	if (ppGipR3)
3354	*ppGipR3 = pGipR3;
3355
3356	#ifdef DEBUG_DARWIN_GIP
3357	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3358	#else
3359	LogFlow(( "SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3360	#endif
3361	return rc;
3362	}
3363
3364
3365	/**
3366	* Unmaps any user mapping of the GIP and terminates all GIP access
3367	* from this session.
3368	*
3369	* @returns IPRT status code.
3370	* @param pSession Session to which the GIP mapping should belong.
3371	*/
3372	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
3373	{
3374	int rc = VINF_SUCCESS;
3375	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3376	#ifdef DEBUG_DARWIN_GIP
3377	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
3378	pSession,
3379	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
3380	pSession->GipMapObjR3));
3381	#else
3382	LogFlow(("SUPR0GipUnmap: pSession=%p\n", pSession));
3383	#endif
3384	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3385
3386	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3387	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3388	#else
3389	RTSemFastMutexRequest(pDevExt->mtxGip);
3390	#endif
3391
3392	/*
3393	* Unmap anything?
3394	*/
3395	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
3396	{
3397	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
3398	AssertRC(rc);
3399	if (RT_SUCCESS(rc))
3400	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
3401	}
3402
3403	/*
3404	* Dereference global GIP.
3405	*/
3406	if (pSession->fGipReferenced && !rc)
3407	{
3408	pSession->fGipReferenced = 0;
3409	if ( pDevExt->cGipUsers > 0
3410	&& !--pDevExt->cGipUsers)
3411	{
3412	LogFlow(("SUPR0GipUnmap: Suspends GIP updating\n"));
3413	#ifndef DO_NOT_START_GIP
3414	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = VINF_SUCCESS;
3415	#endif
3416
3417	if (pDevExt->u32SystemTimerGranularityGrant)
3418	{
3419	int rc2 = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant);
3420	AssertRC(rc2);
3421	pDevExt->u32SystemTimerGranularityGrant = 0;
3422	}
3423	}
3424	}
3425
3426	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3427	RTSemMutexRelease(pDevExt->mtxGip);
3428	#else
3429	RTSemFastMutexRelease(pDevExt->mtxGip);
3430	#endif
3431
3432	return rc;
3433	}
3434
3435
3436	/**
3437	* Gets the GIP pointer.
3438	*
3439	* @returns Pointer to the GIP or NULL.
3440	*/
3441	SUPDECL(PSUPGLOBALINFOPAGE) SUPGetGIP(void)
3442	{
3443	return g_pSUPGlobalInfoPage;
3444	}
3445
3446
3447	/**
3448	* Register a component factory with the support driver.
3449	*
3450	* This is currently restricted to kernel sessions only.
3451	*
3452	* @returns VBox status code.
3453	* @retval VINF_SUCCESS on success.
3454	* @retval VERR_NO_MEMORY if we're out of memory.
3455	* @retval VERR_ALREADY_EXISTS if the factory has already been registered.
3456	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3457	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3458	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3459	*
3460	* @param pSession The SUPDRV session (must be a ring-0 session).
3461	* @param pFactory Pointer to the component factory registration structure.
3462	*
3463	* @remarks This interface is also available via SUPR0IdcComponentRegisterFactory.
3464	*/
3465	SUPR0DECL(int) SUPR0ComponentRegisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3466	{
3467	PSUPDRVFACTORYREG pNewReg;
3468	const char *psz;
3469	int rc;
3470
3471	/*
3472	* Validate parameters.
3473	*/
3474	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3475	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3476	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3477	AssertPtrReturn(pFactory->pfnQueryFactoryInterface, VERR_INVALID_POINTER);
3478	psz = RTStrEnd(pFactory->szName, sizeof(pFactory->szName));
3479	AssertReturn(psz, VERR_INVALID_PARAMETER);
3480
3481	/*
3482	* Allocate and initialize a new registration structure.
3483	*/
3484	pNewReg = (PSUPDRVFACTORYREG)RTMemAlloc(sizeof(SUPDRVFACTORYREG));
3485	if (pNewReg)
3486	{
3487	pNewReg->pNext = NULL;
3488	pNewReg->pFactory = pFactory;
3489	pNewReg->pSession = pSession;
3490	pNewReg->cchName = psz - &pFactory->szName[0];
3491
3492	/*
3493	* Add it to the tail of the list after checking for prior registration.
3494	*/
3495	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3496	if (RT_SUCCESS(rc))
3497	{
3498	PSUPDRVFACTORYREG pPrev = NULL;
3499	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3500	while (pCur && pCur->pFactory != pFactory)
3501	{
3502	pPrev = pCur;
3503	pCur = pCur->pNext;
3504	}
3505	if (!pCur)
3506	{
3507	if (pPrev)
3508	pPrev->pNext = pNewReg;
3509	else
3510	pSession->pDevExt->pComponentFactoryHead = pNewReg;
3511	rc = VINF_SUCCESS;
3512	}
3513	else
3514	rc = VERR_ALREADY_EXISTS;
3515
3516	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3517	}
3518
3519	if (RT_FAILURE(rc))
3520	RTMemFree(pNewReg);
3521	}
3522	else
3523	rc = VERR_NO_MEMORY;
3524	return rc;
3525	}
3526
3527
3528	/**
3529	* Deregister a component factory.
3530	*
3531	* @returns VBox status code.
3532	* @retval VINF_SUCCESS on success.
3533	* @retval VERR_NOT_FOUND if the factory wasn't registered.
3534	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3535	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3536	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3537	*
3538	* @param pSession The SUPDRV session (must be a ring-0 session).
3539	* @param pFactory Pointer to the component factory registration structure
3540	* previously passed SUPR0ComponentRegisterFactory().
3541	*
3542	* @remarks This interface is also available via SUPR0IdcComponentDeregisterFactory.
3543	*/
3544	SUPR0DECL(int) SUPR0ComponentDeregisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3545	{
3546	int rc;
3547
3548	/*
3549	* Validate parameters.
3550	*/
3551	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3552	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3553	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3554
3555	/*
3556	* Take the lock and look for the registration record.
3557	*/
3558	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3559	if (RT_SUCCESS(rc))
3560	{
3561	PSUPDRVFACTORYREG pPrev = NULL;
3562	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3563	while (pCur && pCur->pFactory != pFactory)
3564	{
3565	pPrev = pCur;
3566	pCur = pCur->pNext;
3567	}
3568	if (pCur)
3569	{
3570	if (!pPrev)
3571	pSession->pDevExt->pComponentFactoryHead = pCur->pNext;
3572	else
3573	pPrev->pNext = pCur->pNext;
3574
3575	pCur->pNext = NULL;
3576	pCur->pFactory = NULL;
3577	pCur->pSession = NULL;
3578	rc = VINF_SUCCESS;
3579	}
3580	else
3581	rc = VERR_NOT_FOUND;
3582
3583	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3584
3585	RTMemFree(pCur);
3586	}
3587	return rc;
3588	}
3589
3590
3591	/**
3592	* Queries a component factory.
3593	*
3594	* @returns VBox status code.
3595	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3596	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3597	* @retval VERR_SUPDRV_COMPONENT_NOT_FOUND if the component factory wasn't found.
3598	* @retval VERR_SUPDRV_INTERFACE_NOT_SUPPORTED if the interface wasn't supported.
3599	*
3600	* @param pSession The SUPDRV session.
3601	* @param pszName The name of the component factory.
3602	* @param pszInterfaceUuid The UUID of the factory interface (stringified).
3603	* @param ppvFactoryIf Where to store the factory interface.
3604	*/
3605	SUPR0DECL(int) SUPR0ComponentQueryFactory(PSUPDRVSESSION pSession, const char pszName, const char pszInterfaceUuid, void **ppvFactoryIf)
3606	{
3607	const char *pszEnd;
3608	size_t cchName;
3609	int rc;
3610
3611	/*
3612	* Validate parameters.
3613	*/
3614	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3615
3616	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
3617	pszEnd = RTStrEnd(pszName, RT_SIZEOFMEMB(SUPDRVFACTORY, szName));
3618	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3619	cchName = pszEnd - pszName;
3620
3621	AssertPtrReturn(pszInterfaceUuid, VERR_INVALID_POINTER);
3622	pszEnd = RTStrEnd(pszInterfaceUuid, RTUUID_STR_LENGTH);
3623	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3624
3625	AssertPtrReturn(ppvFactoryIf, VERR_INVALID_POINTER);
3626	*ppvFactoryIf = NULL;
3627
3628	/*
3629	* Take the lock and try all factories by this name.
3630	*/
3631	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3632	if (RT_SUCCESS(rc))
3633	{
3634	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3635	rc = VERR_SUPDRV_COMPONENT_NOT_FOUND;
3636	while (pCur)
3637	{
3638	if ( pCur->cchName == cchName
3639	&& !memcmp(pCur->pFactory->szName, pszName, cchName))
3640	{
3641	void *pvFactory = pCur->pFactory->pfnQueryFactoryInterface(pCur->pFactory, pSession, pszInterfaceUuid);
3642	if (pvFactory)
3643	{
3644	*ppvFactoryIf = pvFactory;
3645	rc = VINF_SUCCESS;
3646	break;
3647	}
3648	rc = VERR_SUPDRV_INTERFACE_NOT_SUPPORTED;
3649	}
3650
3651	/* next */
3652	pCur = pCur->pNext;
3653	}
3654
3655	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3656	}
3657	return rc;
3658	}
3659
3660
3661	/**
3662	* Adds a memory object to the session.
3663	*
3664	* @returns IPRT status code.
3665	* @param pMem Memory tracking structure containing the
3666	* information to track.
3667	* @param pSession The session.
3668	*/
3669	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
3670	{
3671	PSUPDRVBUNDLE pBundle;
3672	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3673
3674	/*
3675	* Find free entry and record the allocation.
3676	*/
3677	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3678	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3679	{
3680	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
3681	{
3682	unsigned i;
3683	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3684	{
3685	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
3686	{
3687	pBundle->cUsed++;
3688	pBundle->aMem[i] = *pMem;
3689	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3690	return VINF_SUCCESS;
3691	}
3692	}
3693	AssertFailed(); /* !!this can't be happening!!! */
3694	}
3695	}
3696	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3697
3698	/*
3699	* Need to allocate a new bundle.
3700	* Insert into the last entry in the bundle.
3701	*/
3702	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
3703	if (!pBundle)
3704	return VERR_NO_MEMORY;
3705
3706	/* take last entry. */
3707	pBundle->cUsed++;
3708	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
3709
3710	/* insert into list. */
3711	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3712	pBundle->pNext = pSession->Bundle.pNext;
3713	pSession->Bundle.pNext = pBundle;
3714	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3715
3716	return VINF_SUCCESS;
3717	}
3718
3719
3720	/**
3721	* Releases a memory object referenced by pointer and type.
3722	*
3723	* @returns IPRT status code.
3724	* @param pSession Session data.
3725	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
3726	* @param eType Memory type.
3727	*/
3728	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
3729	{
3730	PSUPDRVBUNDLE pBundle;
3731	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3732
3733	/*
3734	* Validate input.
3735	*/
3736	if (!uPtr)
3737	{
3738	Log(("Illegal address %p\n", (void *)uPtr));
3739	return VERR_INVALID_PARAMETER;
3740	}
3741
3742	/*
3743	* Search for the address.
3744	*/
3745	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3746	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3747	{
3748	if (pBundle->cUsed > 0)
3749	{
3750	unsigned i;
3751	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3752	{
3753	if ( pBundle->aMem[i].eType == eType
3754	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3755	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
3756	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3757	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
3758	)
3759	{
3760	/* Make a copy of it and release it outside the spinlock. */
3761	SUPDRVMEMREF Mem = pBundle->aMem[i];
3762	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
3763	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
3764	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
3765	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3766
3767	if (Mem.MapObjR3 != NIL_RTR0MEMOBJ)
3768	{
3769	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
3770	AssertRC(rc); /** @todo figure out how to handle this. */
3771	}
3772	if (Mem.MemObj != NIL_RTR0MEMOBJ)
3773	{
3774	int rc = RTR0MemObjFree(Mem.MemObj, true /* fFreeMappings */);
3775	AssertRC(rc); /** @todo figure out how to handle this. */
3776	}
3777	return VINF_SUCCESS;
3778	}
3779	}
3780	}
3781	}
3782	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3783	Log(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
3784	return VERR_INVALID_PARAMETER;
3785	}
3786
3787
3788	/**
3789	* Opens an image. If it's the first time it's opened the call must upload
3790	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
3791	*
3792	* This is the 1st step of the loading.
3793	*
3794	* @returns IPRT status code.
3795	* @param pDevExt Device globals.
3796	* @param pSession Session data.
3797	* @param pReq The open request.
3798	*/
3799	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
3800	{
3801	int rc;
3802	PSUPDRVLDRIMAGE pImage;
3803	void *pv;
3804	size_t cchName = strlen(pReq->u.In.szName); /* (caller checked < 32). */
3805	LogFlow(("supdrvIOCtl_LdrOpen: szName=%s cbImageWithTabs=%d\n", pReq->u.In.szName, pReq->u.In.cbImageWithTabs));
3806
3807	/*
3808	* Check if we got an instance of the image already.
3809	*/
3810	supdrvLdrLock(pDevExt);
3811	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
3812	{
3813	if ( pImage->szName[cchName] == '\0'
3814	&& !memcmp(pImage->szName, pReq->u.In.szName, cchName))
3815	{
3816	/** @todo check cbImageBits and cbImageWithTabs here, if they differs that indicates that the images are different. */
3817	pImage->cUsage++;
3818	pReq->u.Out.pvImageBase = pImage->pvImage;
3819	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
3820	pReq->u.Out.fNativeLoader = pImage->fNative;
3821	supdrvLdrAddUsage(pSession, pImage);
3822	supdrvLdrUnlock(pDevExt);
3823	return VINF_SUCCESS;
3824	}
3825	}
3826	/* (not found - add it!) */
3827
3828	/*
3829	* Allocate memory.
3830	*/
3831	pv = RTMemAlloc(RT_OFFSETOF(SUPDRVLDRIMAGE, szName[cchName + 1]));
3832	if (!pv)
3833	{
3834	supdrvLdrUnlock(pDevExt);
3835	Log(("supdrvIOCtl_LdrOpen: RTMemAlloc() failed\n"));
3836	return /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_2;
3837	}
3838
3839	/*
3840	* Setup and link in the LDR stuff.
3841	*/
3842	pImage = (PSUPDRVLDRIMAGE)pv;
3843	pImage->pvImage = NULL;
3844	pImage->pvImageAlloc = NULL;
3845	pImage->cbImageWithTabs = pReq->u.In.cbImageWithTabs;
3846	pImage->cbImageBits = pReq->u.In.cbImageBits;
3847	pImage->cSymbols = 0;
3848	pImage->paSymbols = NULL;
3849	pImage->pachStrTab = NULL;
3850	pImage->cbStrTab = 0;
3851	pImage->pfnModuleInit = NULL;
3852	pImage->pfnModuleTerm = NULL;
3853	pImage->pfnServiceReqHandler = NULL;
3854	pImage->uState = SUP_IOCTL_LDR_OPEN;
3855	pImage->cUsage = 1;
3856	memcpy(pImage->szName, pReq->u.In.szName, cchName + 1);
3857
3858	/*
3859	* Try load it using the native loader, if that isn't supported, fall back
3860	* on the older method.
3861	*/
3862	pImage->fNative = true;
3863	rc = supdrvOSLdrOpen(pDevExt, pImage, pReq->u.In.szFilename);
3864	if (rc == VERR_NOT_SUPPORTED)
3865	{
3866	pImage->pvImageAlloc = RTMemExecAlloc(pImage->cbImageBits + 31);
3867	pImage->pvImage = RT_ALIGN_P(pImage->pvImageAlloc, 32);
3868	pImage->fNative = false;
3869	rc = pImage->pvImageAlloc ? VINF_SUCCESS : VERR_NO_EXEC_MEMORY;
3870	}
3871	if (RT_FAILURE(rc))
3872	{
3873	supdrvLdrUnlock(pDevExt);
3874	RTMemFree(pImage);
3875	Log(("supdrvIOCtl_LdrOpen(%s): failed - %Rrc\n", pReq->u.In.szName, rc));
3876	return rc;
3877	}
3878	Assert(VALID_PTR(pImage->pvImage) \|\| RT_FAILURE(rc));
3879
3880	/*
3881	* Link it.
3882	*/
3883	pImage->pNext = pDevExt->pLdrImages;
3884	pDevExt->pLdrImages = pImage;
3885
3886	supdrvLdrAddUsage(pSession, pImage);
3887
3888	pReq->u.Out.pvImageBase = pImage->pvImage;
3889	pReq->u.Out.fNeedsLoading = true;
3890	pReq->u.Out.fNativeLoader = pImage->fNative;
3891	supdrvLdrUnlock(pDevExt);
3892
3893	#if defined(RT_OS_WINDOWS) && defined(DEBUG)
3894	SUPR0Printf("VBoxDrv: windbg> .reload /f %s=%#p\n", pImage->szName, pImage->pvImage);
3895	#endif
3896	return VINF_SUCCESS;
3897	}
3898
3899
3900	/**
3901	* Worker that validates a pointer to an image entrypoint.
3902	*
3903	* @returns IPRT status code.
3904	* @param pDevExt The device globals.
3905	* @param pImage The loader image.
3906	* @param pv The pointer into the image.
3907	* @param fMayBeNull Whether it may be NULL.
3908	* @param pszWhat What is this entrypoint? (for logging)
3909	* @param pbImageBits The image bits prepared by ring-3.
3910	*
3911	* @remarks Will leave the lock on failure.
3912	*/
3913	static int supdrvLdrValidatePointer(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage, void *pv,
3914	bool fMayBeNull, const uint8_t pbImageBits, const char pszWhat)
3915	{
3916	if (!fMayBeNull \|\| pv)
3917	{
3918	if ((uintptr_t)pv - (uintptr_t)pImage->pvImage >= pImage->cbImageBits)
3919	{
3920	supdrvLdrUnlock(pDevExt);
3921	Log(("Out of range (%p LB %#x): %s=%p\n", pImage->pvImage, pImage->cbImageBits, pszWhat, pv));
3922	return VERR_INVALID_PARAMETER;
3923	}
3924
3925	if (pImage->fNative)
3926	{
3927	int rc = supdrvOSLdrValidatePointer(pDevExt, pImage, pv, pbImageBits);
3928	if (RT_FAILURE(rc))
3929	{
3930	supdrvLdrUnlock(pDevExt);
3931	Log(("Bad entry point address: %s=%p (rc=%Rrc)\n", pszWhat, pv, rc));
3932	return rc;
3933	}
3934	}
3935	}
3936	return VINF_SUCCESS;
3937	}
3938
3939
3940	/**
3941	* Loads the image bits.
3942	*
3943	* This is the 2nd step of the loading.
3944	*
3945	* @returns IPRT status code.
3946	* @param pDevExt Device globals.
3947	* @param pSession Session data.
3948	* @param pReq The request.
3949	*/
3950	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
3951	{
3952	PSUPDRVLDRUSAGE pUsage;
3953	PSUPDRVLDRIMAGE pImage;
3954	int rc;
3955	LogFlow(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImageWithBits=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImageWithTabs));
3956
3957	/*
3958	* Find the ldr image.
3959	*/
3960	supdrvLdrLock(pDevExt);
3961	pUsage = pSession->pLdrUsage;
3962	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3963	pUsage = pUsage->pNext;
3964	if (!pUsage)
3965	{
3966	supdrvLdrUnlock(pDevExt);
3967	Log(("SUP_IOCTL_LDR_LOAD: couldn't find image!\n"));
3968	return VERR_INVALID_HANDLE;
3969	}
3970	pImage = pUsage->pImage;
3971
3972	/*
3973	* Validate input.
3974	*/
3975	if ( pImage->cbImageWithTabs != pReq->u.In.cbImageWithTabs
3976	\|\| pImage->cbImageBits != pReq->u.In.cbImageBits)
3977	{
3978	supdrvLdrUnlock(pDevExt);
3979	Log(("SUP_IOCTL_LDR_LOAD: image size mismatch!! %d(prep) != %d(load) or %d != %d\n",
3980	pImage->cbImageWithTabs, pReq->u.In.cbImageWithTabs, pImage->cbImageBits, pReq->u.In.cbImageBits));
3981	return VERR_INVALID_HANDLE;
3982	}
3983
3984	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
3985	{
3986	unsigned uState = pImage->uState;
3987	supdrvLdrUnlock(pDevExt);
3988	if (uState != SUP_IOCTL_LDR_LOAD)
3989	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
3990	return VERR_ALREADY_LOADED;
3991	}
3992
3993	switch (pReq->u.In.eEPType)
3994	{
3995	case SUPLDRLOADEP_NOTHING:
3996	break;
3997
3998	case SUPLDRLOADEP_VMMR0:
3999	rc = supdrvLdrValidatePointer( pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0, false, pReq->u.In.abImage, "pvVMMR0");
4000	if (RT_SUCCESS(rc))
4001	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt, false, pReq->u.In.abImage, "pvVMMR0EntryInt");
4002	if (RT_SUCCESS(rc))
4003	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, false, pReq->u.In.abImage, "pvVMMR0EntryFast");
4004	if (RT_SUCCESS(rc))
4005	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx, false, pReq->u.In.abImage, "pvVMMR0EntryEx");
4006	if (RT_FAILURE(rc))
4007	return rc;
4008	break;
4009
4010	case SUPLDRLOADEP_SERVICE:
4011	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.Service.pfnServiceReq, false, pReq->u.In.abImage, "pfnServiceReq");
4012	if (RT_FAILURE(rc))
4013	return rc;
4014	if ( pReq->u.In.EP.Service.apvReserved[0] != NIL_RTR0PTR
4015	\|\| pReq->u.In.EP.Service.apvReserved[1] != NIL_RTR0PTR
4016	\|\| pReq->u.In.EP.Service.apvReserved[2] != NIL_RTR0PTR)
4017	{
4018	supdrvLdrUnlock(pDevExt);
4019	Log(("Out of range (%p LB %#x): apvReserved={%p,%p,%p} MBZ!\n",
4020	pImage->pvImage, pReq->u.In.cbImageWithTabs,
4021	pReq->u.In.EP.Service.apvReserved[0],
4022	pReq->u.In.EP.Service.apvReserved[1],
4023	pReq->u.In.EP.Service.apvReserved[2]));
4024	return VERR_INVALID_PARAMETER;
4025	}
4026	break;
4027
4028	default:
4029	supdrvLdrUnlock(pDevExt);
4030	Log(("Invalid eEPType=%d\n", pReq->u.In.eEPType));
4031	return VERR_INVALID_PARAMETER;
4032	}
4033
4034	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleInit, true, pReq->u.In.abImage, "pfnModuleInit");
4035	if (RT_FAILURE(rc))
4036	return rc;
4037	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleTerm, true, pReq->u.In.abImage, "pfnModuleTerm");
4038	if (RT_FAILURE(rc))
4039	return rc;
4040
4041	/*
4042	* Allocate and copy the tables.
4043	* (No need to do try/except as this is a buffered request.)
4044	*/
4045	pImage->cbStrTab = pReq->u.In.cbStrTab;
4046	if (pImage->cbStrTab)
4047	{
4048	pImage->pachStrTab = (char *)RTMemAlloc(pImage->cbStrTab);
4049	if (pImage->pachStrTab)
4050	memcpy(pImage->pachStrTab, &pReq->u.In.abImage[pReq->u.In.offStrTab], pImage->cbStrTab);
4051	else
4052	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_3;
4053	}
4054
4055	pImage->cSymbols = pReq->u.In.cSymbols;
4056	if (RT_SUCCESS(rc) && pImage->cSymbols)
4057	{
4058	size_t cbSymbols = pImage->cSymbols * sizeof(SUPLDRSYM);
4059	pImage->paSymbols = (PSUPLDRSYM)RTMemAlloc(cbSymbols);
4060	if (pImage->paSymbols)
4061	memcpy(pImage->paSymbols, &pReq->u.In.abImage[pReq->u.In.offSymbols], cbSymbols);
4062	else
4063	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_4;
4064	}
4065
4066	/*
4067	* Copy the bits / complete native loading.
4068	*/
4069	if (RT_SUCCESS(rc))
4070	{
4071	pImage->uState = SUP_IOCTL_LDR_LOAD;
4072	pImage->pfnModuleInit = pReq->u.In.pfnModuleInit;
4073	pImage->pfnModuleTerm = pReq->u.In.pfnModuleTerm;
4074
4075	if (pImage->fNative)
4076	rc = supdrvOSLdrLoad(pDevExt, pImage, pReq->u.In.abImage, pReq);
4077	else
4078	memcpy(pImage->pvImage, &pReq->u.In.abImage[0], pImage->cbImageBits);
4079	}
4080
4081	/*
4082	* Update any entry points.
4083	*/
4084	if (RT_SUCCESS(rc))
4085	{
4086	switch (pReq->u.In.eEPType)
4087	{
4088	default:
4089	case SUPLDRLOADEP_NOTHING:
4090	rc = VINF_SUCCESS;
4091	break;
4092	case SUPLDRLOADEP_VMMR0:
4093	rc = supdrvLdrSetVMMR0EPs(pDevExt, pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
4094	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
4095	break;
4096	case SUPLDRLOADEP_SERVICE:
4097	pImage->pfnServiceReqHandler = pReq->u.In.EP.Service.pfnServiceReq;
4098	rc = VINF_SUCCESS;
4099	break;
4100	}
4101	}
4102
4103	/*
4104	* On success call the module initialization.
4105	*/
4106	LogFlow(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
4107	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
4108	{
4109	Log(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
4110	rc = pImage->pfnModuleInit();
4111	if (rc && pDevExt->pvVMMR0 == pImage->pvImage)
4112	supdrvLdrUnsetVMMR0EPs(pDevExt);
4113	}
4114
4115	if (RT_FAILURE(rc))
4116	{
4117	pImage->uState = SUP_IOCTL_LDR_OPEN;
4118	pImage->pfnModuleInit = NULL;
4119	pImage->pfnModuleTerm = NULL;
4120	pImage->pfnServiceReqHandler= NULL;
4121	pImage->cbStrTab = 0;
4122	RTMemFree(pImage->pachStrTab);
4123	pImage->pachStrTab = NULL;
4124	RTMemFree(pImage->paSymbols);
4125	pImage->paSymbols = NULL;
4126	pImage->cSymbols = 0;
4127	}
4128
4129	supdrvLdrUnlock(pDevExt);
4130	return rc;
4131	}
4132
4133
4134	/**
4135	* Frees a previously loaded (prep'ed) image.
4136	*
4137	* @returns IPRT status code.
4138	* @param pDevExt Device globals.
4139	* @param pSession Session data.
4140	* @param pReq The request.
4141	*/
4142	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
4143	{
4144	int rc;
4145	PSUPDRVLDRUSAGE pUsagePrev;
4146	PSUPDRVLDRUSAGE pUsage;
4147	PSUPDRVLDRIMAGE pImage;
4148	LogFlow(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
4149
4150	/*
4151	* Find the ldr image.
4152	*/
4153	supdrvLdrLock(pDevExt);
4154	pUsagePrev = NULL;
4155	pUsage = pSession->pLdrUsage;
4156	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4157	{
4158	pUsagePrev = pUsage;
4159	pUsage = pUsage->pNext;
4160	}
4161	if (!pUsage)
4162	{
4163	supdrvLdrUnlock(pDevExt);
4164	Log(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
4165	return VERR_INVALID_HANDLE;
4166	}
4167
4168	/*
4169	* Check if we can remove anything.
4170	*/
4171	rc = VINF_SUCCESS;
4172	pImage = pUsage->pImage;
4173	if (pImage->cUsage <= 1 \|\| pUsage->cUsage <= 1)
4174	{
4175	/*
4176	* Check if there are any objects with destructors in the image, if
4177	* so leave it for the session cleanup routine so we get a chance to
4178	* clean things up in the right order and not leave them all dangling.
4179	*/
4180	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
4181	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
4182	if (pImage->cUsage <= 1)
4183	{
4184	PSUPDRVOBJ pObj;
4185	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4186	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4187	{
4188	rc = VERR_DANGLING_OBJECTS;
4189	break;
4190	}
4191	}
4192	else
4193	{
4194	PSUPDRVUSAGE pGenUsage;
4195	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
4196	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4197	{
4198	rc = VERR_DANGLING_OBJECTS;
4199	break;
4200	}
4201	}
4202	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
4203	if (rc == VINF_SUCCESS)
4204	{
4205	/* unlink it */
4206	if (pUsagePrev)
4207	pUsagePrev->pNext = pUsage->pNext;
4208	else
4209	pSession->pLdrUsage = pUsage->pNext;
4210
4211	/* free it */
4212	pUsage->pImage = NULL;
4213	pUsage->pNext = NULL;
4214	RTMemFree(pUsage);
4215
4216	/*
4217	* Dereference the image.
4218	*/
4219	if (pImage->cUsage <= 1)
4220	supdrvLdrFree(pDevExt, pImage);
4221	else
4222	pImage->cUsage--;
4223	}
4224	else
4225	{
4226	Log(("supdrvIOCtl_LdrFree: Dangling objects in %p/%s!\n", pImage->pvImage, pImage->szName));
4227	rc = VINF_SUCCESS; /** @todo BRANCH-2.1: remove this after branching. */
4228	}
4229	}
4230	else
4231	{
4232	/*
4233	* Dereference both image and usage.
4234	*/
4235	pImage->cUsage--;
4236	pUsage->cUsage--;
4237	}
4238
4239	supdrvLdrUnlock(pDevExt);
4240	return rc;
4241	}
4242
4243
4244	/**
4245	* Gets the address of a symbol in an open image.
4246	*
4247	* @returns IPRT status code.
4248	* @param pDevExt Device globals.
4249	* @param pSession Session data.
4250	* @param pReq The request buffer.
4251	*/
4252	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
4253	{
4254	PSUPDRVLDRIMAGE pImage;
4255	PSUPDRVLDRUSAGE pUsage;
4256	uint32_t i;
4257	PSUPLDRSYM paSyms;
4258	const char *pchStrings;
4259	const size_t cbSymbol = strlen(pReq->u.In.szSymbol) + 1;
4260	void *pvSymbol = NULL;
4261	int rc = VERR_GENERAL_FAILURE;
4262	Log3(("supdrvIOCtl_LdrGetSymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
4263
4264	/*
4265	* Find the ldr image.
4266	*/
4267	supdrvLdrLock(pDevExt);
4268	pUsage = pSession->pLdrUsage;
4269	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4270	pUsage = pUsage->pNext;
4271	if (!pUsage)
4272	{
4273	supdrvLdrUnlock(pDevExt);
4274	Log(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
4275	return VERR_INVALID_HANDLE;
4276	}
4277	pImage = pUsage->pImage;
4278	if (pImage->uState != SUP_IOCTL_LDR_LOAD)
4279	{
4280	unsigned uState = pImage->uState;
4281	supdrvLdrUnlock(pDevExt);
4282	Log(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", uState, uState)); NOREF(uState);
4283	return VERR_ALREADY_LOADED;
4284	}
4285
4286	/*
4287	* Search the symbol strings.
4288	*
4289	* Note! The int32_t is for native loading on solaris where the data
4290	* and text segments are in very different places.
4291	*/
4292	pchStrings = pImage->pachStrTab;
4293	paSyms = pImage->paSymbols;
4294	for (i = 0; i < pImage->cSymbols; i++)
4295	{
4296	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4297	&& !memcmp(pchStrings + paSyms[i].offName, pReq->u.In.szSymbol, cbSymbol))
4298	{
4299	pvSymbol = (uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol;
4300	rc = VINF_SUCCESS;
4301	break;
4302	}
4303	}
4304	supdrvLdrUnlock(pDevExt);
4305	pReq->u.Out.pvSymbol = pvSymbol;
4306	return rc;
4307	}
4308
4309
4310	/**
4311	* Gets the address of a symbol in an open image or the support driver.
4312	*
4313	* @returns VINF_SUCCESS on success.
4314	* @returns
4315	* @param pDevExt Device globals.
4316	* @param pSession Session data.
4317	* @param pReq The request buffer.
4318	*/
4319	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq)
4320	{
4321	int rc = VINF_SUCCESS;
4322	const char *pszSymbol = pReq->u.In.pszSymbol;
4323	const char *pszModule = pReq->u.In.pszModule;
4324	size_t cbSymbol;
4325	char const *pszEnd;
4326	uint32_t i;
4327
4328	/*
4329	* Input validation.
4330	*/
4331	AssertPtrReturn(pszSymbol, VERR_INVALID_POINTER);
4332	pszEnd = RTStrEnd(pszSymbol, 512);
4333	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4334	cbSymbol = pszEnd - pszSymbol + 1;
4335
4336	if (pszModule)
4337	{
4338	AssertPtrReturn(pszModule, VERR_INVALID_POINTER);
4339	pszEnd = RTStrEnd(pszModule, 64);
4340	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4341	}
4342	Log3(("supdrvIDC_LdrGetSymbol: pszModule=%p:{%s} pszSymbol=%p:{%s}\n", pszModule, pszModule, pszSymbol, pszSymbol));
4343
4344
4345	if ( !pszModule
4346	\|\| !strcmp(pszModule, "SupDrv"))
4347	{
4348	/*
4349	* Search the support driver export table.
4350	*/
4351	for (i = 0; i < RT_ELEMENTS(g_aFunctions); i++)
4352	if (!strcmp(g_aFunctions[i].szName, pszSymbol))
4353	{
4354	pReq->u.Out.pfnSymbol = g_aFunctions[i].pfn;
4355	break;
4356	}
4357	}
4358	else
4359	{
4360	/*
4361	* Find the loader image.
4362	*/
4363	PSUPDRVLDRIMAGE pImage;
4364
4365	supdrvLdrLock(pDevExt);
4366
4367	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
4368	if (!strcmp(pImage->szName, pszModule))
4369	break;
4370	if (pImage && pImage->uState == SUP_IOCTL_LDR_LOAD)
4371	{
4372	/*
4373	* Search the symbol strings.
4374	*/
4375	const char *pchStrings = pImage->pachStrTab;
4376	PCSUPLDRSYM paSyms = pImage->paSymbols;
4377	for (i = 0; i < pImage->cSymbols; i++)
4378	{
4379	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4380	&& !memcmp(pchStrings + paSyms[i].offName, pszSymbol, cbSymbol))
4381	{
4382	/*
4383	* Found it! Calc the symbol address and add a reference to the module.
4384	*/
4385	pReq->u.Out.pfnSymbol = (PFNRT)((uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol);
4386	rc = supdrvLdrAddUsage(pSession, pImage);
4387	break;
4388	}
4389	}
4390	}
4391	else
4392	rc = pImage ? VERR_WRONG_ORDER : VERR_MODULE_NOT_FOUND;
4393
4394	supdrvLdrUnlock(pDevExt);
4395	}
4396	return rc;
4397	}
4398
4399
4400	/**
4401	* Updates the VMMR0 entry point pointers.
4402	*
4403	* @returns IPRT status code.
4404	* @param pDevExt Device globals.
4405	* @param pSession Session data.
4406	* @param pVMMR0 VMMR0 image handle.
4407	* @param pvVMMR0EntryInt VMMR0EntryInt address.
4408	* @param pvVMMR0EntryFast VMMR0EntryFast address.
4409	* @param pvVMMR0EntryEx VMMR0EntryEx address.
4410	* @remark Caller must own the loader mutex.
4411	*/
4412	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx)
4413	{
4414	int rc = VINF_SUCCESS;
4415	LogFlow(("supdrvLdrSetR0EP pvVMMR0=%p pvVMMR0EntryInt=%p\n", pvVMMR0, pvVMMR0EntryInt));
4416
4417
4418	/*
4419	* Check if not yet set.
4420	*/
4421	if (!pDevExt->pvVMMR0)
4422	{
4423	pDevExt->pvVMMR0 = pvVMMR0;
4424	pDevExt->pfnVMMR0EntryInt = pvVMMR0EntryInt;
4425	pDevExt->pfnVMMR0EntryFast = pvVMMR0EntryFast;
4426	pDevExt->pfnVMMR0EntryEx = pvVMMR0EntryEx;
4427	}
4428	else
4429	{
4430	/*
4431	* Return failure or success depending on whether the values match or not.
4432	*/
4433	if ( pDevExt->pvVMMR0 != pvVMMR0
4434	\|\| (void *)pDevExt->pfnVMMR0EntryInt != pvVMMR0EntryInt
4435	\|\| (void *)pDevExt->pfnVMMR0EntryFast != pvVMMR0EntryFast
4436	\|\| (void *)pDevExt->pfnVMMR0EntryEx != pvVMMR0EntryEx)
4437	{
4438	AssertMsgFailed(("SUP_IOCTL_LDR_SETR0EP: Already set pointing to a different module!\n"));
4439	rc = VERR_INVALID_PARAMETER;
4440	}
4441	}
4442	return rc;
4443	}
4444
4445
4446	/**
4447	* Unsets the VMMR0 entry point installed by supdrvLdrSetR0EP.
4448	*
4449	* @param pDevExt Device globals.
4450	*/
4451	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt)
4452	{
4453	pDevExt->pvVMMR0 = NULL;
4454	pDevExt->pfnVMMR0EntryInt = NULL;
4455	pDevExt->pfnVMMR0EntryFast = NULL;
4456	pDevExt->pfnVMMR0EntryEx = NULL;
4457	}
4458
4459
4460	/**
4461	* Adds a usage reference in the specified session of an image.
4462	*
4463	* Called while owning the loader semaphore.
4464	*
4465	* @returns VINF_SUCCESS on success and VERR_NO_MEMORY on failure.
4466	* @param pSession Session in question.
4467	* @param pImage Image which the session is using.
4468	*/
4469	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage)
4470	{
4471	PSUPDRVLDRUSAGE pUsage;
4472	LogFlow(("supdrvLdrAddUsage: pImage=%p\n", pImage));
4473
4474	/*
4475	* Referenced it already?
4476	*/
4477	pUsage = pSession->pLdrUsage;
4478	while (pUsage)
4479	{
4480	if (pUsage->pImage == pImage)
4481	{
4482	pUsage->cUsage++;
4483	return VINF_SUCCESS;
4484	}
4485	pUsage = pUsage->pNext;
4486	}
4487
4488	/*
4489	* Allocate new usage record.
4490	*/
4491	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
4492	AssertReturn(pUsage, /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_5);
4493	pUsage->cUsage = 1;
4494	pUsage->pImage = pImage;
4495	pUsage->pNext = pSession->pLdrUsage;
4496	pSession->pLdrUsage = pUsage;
4497	return VINF_SUCCESS;
4498	}
4499
4500
4501	/**
4502	* Frees a load image.
4503	*
4504	* @param pDevExt Pointer to device extension.
4505	* @param pImage Pointer to the image we're gonna free.
4506	* This image must exit!
4507	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
4508	*/
4509	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
4510	{
4511	PSUPDRVLDRIMAGE pImagePrev;
4512	LogFlow(("supdrvLdrFree: pImage=%p\n", pImage));
4513
4514	/* find it - arg. should've used doubly linked list. */
4515	Assert(pDevExt->pLdrImages);
4516	pImagePrev = NULL;
4517	if (pDevExt->pLdrImages != pImage)
4518	{
4519	pImagePrev = pDevExt->pLdrImages;
4520	while (pImagePrev->pNext != pImage)
4521	pImagePrev = pImagePrev->pNext;
4522	Assert(pImagePrev->pNext == pImage);
4523	}
4524
4525	/* unlink */
4526	if (pImagePrev)
4527	pImagePrev->pNext = pImage->pNext;
4528	else
4529	pDevExt->pLdrImages = pImage->pNext;
4530
4531	/* check if this is VMMR0.r0 unset its entry point pointers. */
4532	if (pDevExt->pvVMMR0 == pImage->pvImage)
4533	supdrvLdrUnsetVMMR0EPs(pDevExt);
4534
4535	/* check for objects with destructors in this image. (Shouldn't happen.) */
4536	if (pDevExt->pObjs)
4537	{
4538	unsigned cObjs = 0;
4539	PSUPDRVOBJ pObj;
4540	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
4541	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
4542	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4543	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4544	{
4545	pObj->pfnDestructor = NULL;
4546	cObjs++;
4547	}
4548	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
4549	if (cObjs)
4550	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
4551	}
4552
4553	/* call termination function if fully loaded. */
4554	if ( pImage->pfnModuleTerm
4555	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
4556	{
4557	LogFlow(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
4558	pImage->pfnModuleTerm();
4559	}
4560
4561	/* do native unload if appropriate. */
4562	if (pImage->fNative)
4563	supdrvOSLdrUnload(pDevExt, pImage);
4564
4565	/* free the image */
4566	pImage->cUsage = 0;
4567	pImage->pNext = 0;
4568	pImage->uState = SUP_IOCTL_LDR_FREE;
4569	RTMemExecFree(pImage->pvImageAlloc, pImage->cbImageBits + 31);
4570	pImage->pvImageAlloc = NULL;
4571	RTMemFree(pImage->pachStrTab);
4572	pImage->pachStrTab = NULL;
4573	RTMemFree(pImage->paSymbols);
4574	pImage->paSymbols = NULL;
4575	RTMemFree(pImage);
4576	}
4577
4578
4579	/**
4580	* Acquires the loader lock.
4581	*
4582	* @returns IPRT status code.
4583	* @param pDevExt The device extension.
4584	*/
4585	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt)
4586	{
4587	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4588	int rc = RTSemMutexRequest(pDevExt->mtxLdr, RT_INDEFINITE_WAIT);
4589	#else
4590	int rc = RTSemFastMutexRequest(pDevExt->mtxLdr);
4591	#endif
4592	AssertRC(rc);
4593	return rc;
4594	}
4595
4596
4597	/**
4598	* Releases the loader lock.
4599	*
4600	* @returns IPRT status code.
4601	* @param pDevExt The device extension.
4602	*/
4603	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt)
4604	{
4605	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4606	return RTSemMutexRelease(pDevExt->mtxLdr);
4607	#else
4608	return RTSemFastMutexRelease(pDevExt->mtxLdr);
4609	#endif
4610	}
4611
4612
4613	/**
4614	* Implements the service call request.
4615	*
4616	* @returns VBox status code.
4617	* @param pDevExt The device extension.
4618	* @param pSession The calling session.
4619	* @param pReq The request packet, valid.
4620	*/
4621	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq)
4622	{
4623	#if !defined(RT_OS_WINDOWS) \|\| defined(DEBUG)
4624	int rc;
4625
4626	/*
4627	* Find the module first in the module referenced by the calling session.
4628	*/
4629	rc = supdrvLdrLock(pDevExt);
4630	if (RT_SUCCESS(rc))
4631	{
4632	PFNSUPR0SERVICEREQHANDLER pfnServiceReqHandler = NULL;
4633	PSUPDRVLDRUSAGE pUsage;
4634
4635	for (pUsage = pSession->pLdrUsage; pUsage; pUsage = pUsage->pNext)
4636	if ( pUsage->pImage->pfnServiceReqHandler
4637	&& !strcmp(pUsage->pImage->szName, pReq->u.In.szName))
4638	{
4639	pfnServiceReqHandler = pUsage->pImage->pfnServiceReqHandler;
4640	break;
4641	}
4642	supdrvLdrUnlock(pDevExt);
4643
4644	if (pfnServiceReqHandler)
4645	{
4646	/*
4647	* Call it.
4648	*/
4649	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
4650	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, NULL);
4651	else
4652	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0]);
4653	}
4654	else
4655	rc = VERR_SUPDRV_SERVICE_NOT_FOUND;
4656	}
4657
4658	/* log it */
4659	if ( RT_FAILURE(rc)
4660	&& rc != VERR_INTERRUPTED
4661	&& rc != VERR_TIMEOUT)
4662	Log(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4663	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4664	else
4665	Log4(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4666	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4667	return rc;
4668	#else /* RT_OS_WINDOWS && !DEBUG */
4669	return VERR_NOT_IMPLEMENTED;
4670	#endif /* RT_OS_WINDOWS && !DEBUG */
4671	}
4672
4673
4674	/**
4675	* Implements the logger settings request.
4676	*
4677	* @returns VBox status code.
4678	* @param pDevExt The device extension.
4679	* @param pSession The caller's session.
4680	* @param pReq The request.
4681	*/
4682	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq)
4683	{
4684	const char *pszGroup = &pReq->u.In.szStrings[pReq->u.In.offGroups];
4685	const char *pszFlags = &pReq->u.In.szStrings[pReq->u.In.offFlags];
4686	const char *pszDest = &pReq->u.In.szStrings[pReq->u.In.offDestination];
4687	PRTLOGGER pLogger = NULL;
4688	int rc;
4689
4690	/*
4691	* Some further validation.
4692	*/
4693	switch (pReq->u.In.fWhat)
4694	{
4695	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4696	case SUPLOGGERSETTINGS_WHAT_CREATE:
4697	break;
4698
4699	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4700	if (pszGroup \|\| pszFlags \|\| *pszDest)
4701	return VERR_INVALID_PARAMETER;
4702	if (pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_RELEASE)
4703	return VERR_ACCESS_DENIED;
4704	break;
4705
4706	default:
4707	return VERR_INTERNAL_ERROR;
4708	}
4709
4710	/*
4711	* Get the logger.
4712	*/
4713	switch (pReq->u.In.fWhich)
4714	{
4715	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4716	pLogger = RTLogGetDefaultInstance();
4717	break;
4718
4719	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4720	pLogger = RTLogRelDefaultInstance();
4721	break;
4722
4723	default:
4724	return VERR_INTERNAL_ERROR;
4725	}
4726
4727	/*
4728	* Do the job.
4729	*/
4730	switch (pReq->u.In.fWhat)
4731	{
4732	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4733	if (pLogger)
4734	{
4735	rc = RTLogFlags(pLogger, pszFlags);
4736	if (RT_SUCCESS(rc))
4737	rc = RTLogGroupSettings(pLogger, pszGroup);
4738	NOREF(pszDest);
4739	}
4740	else
4741	rc = VERR_NOT_FOUND;
4742	break;
4743
4744	case SUPLOGGERSETTINGS_WHAT_CREATE:
4745	{
4746	if (pLogger)
4747	rc = VERR_ALREADY_EXISTS;
4748	else
4749	{
4750	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
4751
4752	rc = RTLogCreate(&pLogger,
4753	0 /* fFlags */,
4754	pszGroup,
4755	pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_DEBUG
4756	? "VBOX_LOG"
4757	: "VBOX_RELEASE_LOG",
4758	RT_ELEMENTS(s_apszGroups),
4759	s_apszGroups,
4760	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER,
4761	NULL);
4762	if (RT_SUCCESS(rc))
4763	{
4764	rc = RTLogFlags(pLogger, pszFlags);
4765	NOREF(pszDest);
4766	if (RT_SUCCESS(rc))
4767	{
4768	switch (pReq->u.In.fWhich)
4769	{
4770	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4771	pLogger = RTLogSetDefaultInstance(pLogger);
4772	break;
4773	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4774	pLogger = RTLogRelSetDefaultInstance(pLogger);
4775	break;
4776	}
4777	}
4778	RTLogDestroy(pLogger);
4779	}
4780	}
4781	break;
4782	}
4783
4784	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4785	switch (pReq->u.In.fWhich)
4786	{
4787	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4788	pLogger = RTLogSetDefaultInstance(NULL);
4789	break;
4790	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4791	pLogger = RTLogRelSetDefaultInstance(NULL);
4792	break;
4793	}
4794	rc = RTLogDestroy(pLogger);
4795	break;
4796
4797	default:
4798	{
4799	rc = VERR_INTERNAL_ERROR;
4800	break;
4801	}
4802	}
4803
4804	return rc;
4805	}
4806
4807
4808	/**
4809	* Creates the GIP.
4810	*
4811	* @returns VBox status code.
4812	* @param pDevExt Instance data. GIP stuff may be updated.
4813	*/
4814	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
4815	{
4816	PSUPGLOBALINFOPAGE pGip;
4817	RTHCPHYS HCPhysGip;
4818	uint32_t u32SystemResolution;
4819	uint32_t u32Interval;
4820	unsigned cCpus;
4821	int rc;
4822
4823
4824	LogFlow(("supdrvGipCreate:\n"));
4825
4826	/* assert order */
4827	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
4828	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
4829	Assert(!pDevExt->pGipTimer);
4830
4831	/*
4832	* Check the CPU count.
4833	*/
4834	cCpus = RTMpGetArraySize();
4835	if ( cCpus > RTCPUSET_MAX_CPUS
4836	\|\| cCpus > 256 /ApicId is used for the mappings/)
4837	{
4838	SUPR0Printf("VBoxDrv: Too many CPUs (%u) for the GIP (max %u)\n", cCpus, RT_MIN(RTCPUSET_MAX_CPUS, 256));
4839	return VERR_TOO_MANY_CPUS;
4840	}
4841
4842	/*
4843	* Allocate a contiguous set of pages with a default kernel mapping.
4844	*/
4845	rc = RTR0MemObjAllocCont(&pDevExt->GipMemObj, RT_UOFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), false /fExecutable/);
4846	if (RT_FAILURE(rc))
4847	{
4848	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
4849	return rc;
4850	}
4851	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
4852	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
4853
4854	/*
4855	* Find a reasonable update interval and initialize the structure.
4856	*/
4857	u32Interval = u32SystemResolution = RTTimerGetSystemGranularity();
4858	while (u32Interval < 10000000 /* 10 ms */)
4859	u32Interval += u32SystemResolution;
4860
4861	supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), 1000000000 / u32Interval /=Hz/, cCpus);
4862
4863	/*
4864	* Create the timer.
4865	* If CPU_ALL isn't supported we'll have to fall back to synchronous mode.
4866	*/
4867	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
4868	{
4869	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, RTTIMER_FLAGS_CPU_ALL, supdrvGipAsyncTimer, pDevExt);
4870	if (rc == VERR_NOT_SUPPORTED)
4871	{
4872	OSDBGPRINT(("supdrvGipCreate: omni timer not supported, falling back to synchronous mode\n"));
4873	pGip->u32Mode = SUPGIPMODE_SYNC_TSC;
4874	}
4875	}
4876	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
4877	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0, supdrvGipSyncTimer, pDevExt);
4878	if (RT_SUCCESS(rc))
4879	{
4880	rc = RTMpNotificationRegister(supdrvGipMpEvent, pDevExt);
4881	if (RT_SUCCESS(rc))
4882	{
4883	rc = RTMpOnAll(supdrvGipInitOnCpu, pDevExt, pGip);
4884	if (RT_SUCCESS(rc))
4885	{
4886	/*
4887	* We're good.
4888	*/
4889	Log(("supdrvGipCreate: %u ns interval.\n", u32Interval));
4890	g_pSUPGlobalInfoPage = pGip;
4891	return VINF_SUCCESS;
4892	}
4893
4894	OSDBGPRINT(("supdrvGipCreate: RTMpOnAll failed with rc=%Rrc\n", rc));
4895	RTMpNotificationDeregister(supdrvGipMpEvent, pDevExt);
4896
4897	}
4898	else
4899	OSDBGPRINT(("supdrvGipCreate: failed to register MP event notfication. rc=%Rrc\n", rc));
4900	}
4901	else
4902	{
4903	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %u ns interval. rc=%Rrc\n", u32Interval, rc));
4904	Assert(!pDevExt->pGipTimer);
4905	}
4906	supdrvGipDestroy(pDevExt);
4907	return rc;
4908	}
4909
4910
4911	/**
4912	* Terminates the GIP.
4913	*
4914	* @param pDevExt Instance data. GIP stuff may be updated.
4915	*/
4916	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
4917	{
4918	int rc;
4919	#ifdef DEBUG_DARWIN_GIP
4920	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
4921	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
4922	pDevExt->pGipTimer, pDevExt->GipMemObj));
4923	#endif
4924
4925	/*
4926	* Invalid the GIP data.
4927	*/
4928	if (pDevExt->pGip)
4929	{
4930	supdrvGipTerm(pDevExt->pGip);
4931	pDevExt->pGip = NULL;
4932	}
4933	g_pSUPGlobalInfoPage = NULL;
4934
4935	/*
4936	* Destroy the timer and free the GIP memory object.
4937	*/
4938	if (pDevExt->pGipTimer)
4939	{
4940	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
4941	pDevExt->pGipTimer = NULL;
4942	}
4943
4944	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
4945	{
4946	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
4947	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
4948	}
4949
4950	/*
4951	* Finally, make sure we've release the system timer resolution request
4952	* if one actually succeeded and is still pending.
4953	*/
4954	if (pDevExt->u32SystemTimerGranularityGrant)
4955	{
4956	rc = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
4957	pDevExt->u32SystemTimerGranularityGrant = 0;
4958	}
4959	}
4960
4961
4962	/**
4963	* Timer callback function sync GIP mode.
4964	* @param pTimer The timer.
4965	* @param pvUser The device extension.
4966	*/
4967	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
4968	{
4969	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
4970	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4971	uint64_t u64TSC = ASMReadTSC();
4972	uint64_t NanoTS = RTTimeSystemNanoTS();
4973
4974	supdrvGipUpdate(pDevExt->pGip, NanoTS, u64TSC, NIL_RTCPUID, iTick);
4975
4976	ASMSetFlags(fOldFlags);
4977	}
4978
4979
4980	/**
4981	* Timer callback function for async GIP mode.
4982	* @param pTimer The timer.
4983	* @param pvUser The device extension.
4984	*/
4985	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
4986	{
4987	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
4988	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4989	RTCPUID idCpu = RTMpCpuId();
4990	uint64_t u64TSC = ASMReadTSC();
4991	uint64_t NanoTS = RTTimeSystemNanoTS();
4992
4993	/** @todo reset the transaction number and whatnot when iTick == 1. */
4994	if (pDevExt->idGipMaster == idCpu)
4995	supdrvGipUpdate(pDevExt->pGip, NanoTS, u64TSC, idCpu, iTick);
4996	else
4997	supdrvGipUpdatePerCpu(pDevExt->pGip, NanoTS, u64TSC, idCpu, ASMGetApicId(), iTick);
4998
4999	ASMSetFlags(fOldFlags);
5000	}
5001
5002
5003	/**
5004	* Finds our (@a idCpu) entry, or allocates a new one if not found.
5005	*
5006	* @returns Index of the CPU in the cache set.
5007	* @param pGip The GIP.
5008	* @param idCpu The CPU ID.
5009	*/
5010	static uint32_t supdrvGipCpuIndexFromCpuId(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
5011	{
5012	uint32_t i, cTries;
5013
5014	/*
5015	* ASSUMES that CPU IDs are constant.
5016	*/
5017	for (i = 0; i < pGip->cCpus; i++)
5018	if (pGip->aCPUs[i].idCpu == idCpu)
5019	return i;
5020
5021	cTries = 0;
5022	do
5023	{
5024	for (i = 0; i < pGip->cCpus; i++)
5025	{
5026	bool fRc;
5027	ASMAtomicCmpXchgSize(&pGip->aCPUs[i].idCpu, idCpu, NIL_RTCPUID, fRc);
5028	if (fRc)
5029	return i;
5030	}
5031	} while (cTries++ < 32);
5032	AssertReleaseFailed();
5033	return i - 1;
5034	}
5035
5036
5037	/**
5038	* The calling CPU should be accounted as online, update GIP accordingly.
5039	*
5040	* This is used by supdrvGipMpEvent as well as the supdrvGipCreate.
5041	*
5042	* @param pGip The GIP.
5043	* @param idCpu The CPU ID.
5044	*/
5045	static void supdrvGipMpEventOnline(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
5046	{
5047	int iCpuSet = 0;
5048	uint16_t idApic = UINT16_MAX;
5049	uint32_t i = 0;
5050
5051	AssertRelease(idCpu == RTMpCpuId());
5052	Assert(pGip->cPossibleCpus == RTMpGetCount());
5053
5054	/*
5055	* Update the globals.
5056	*/
5057	ASMAtomicWriteU16(&pGip->cPresentCpus, RTMpGetPresentCount());
5058	ASMAtomicWriteU16(&pGip->cOnlineCpus, RTMpGetOnlineCount());
5059	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5060	if (iCpuSet >= 0)
5061	{
5062	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5063	RTCpuSetAddByIndex(&pGip->OnlineCpuSet, iCpuSet);
5064	RTCpuSetAddByIndex(&pGip->PresentCpuSet, iCpuSet);
5065	}
5066
5067	/*
5068	* Update the entry.
5069	*/
5070	i = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
5071	idApic = ASMGetApicId();
5072	ASMAtomicUoWriteU16(&pGip->aCPUs[i].idApic, idApic);
5073	ASMAtomicUoWriteS16(&pGip->aCPUs[i].iCpuSet, (int16_t)iCpuSet);
5074	ASMAtomicUoWriteSize(&pGip->aCPUs[i].idCpu, idCpu);
5075
5076	/*
5077	* Update the APIC ID and CPU set index mappings.
5078	*/
5079	ASMAtomicWriteU16(&pGip->aiCpuFromApicId[idApic], i);
5080	ASMAtomicWriteU16(&pGip->aiCpuFromCpuSetIdx[iCpuSet], i);
5081
5082	/* commit it */
5083	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_ONLINE);
5084	}
5085
5086
5087	/**
5088	* The CPU should be accounted as offline, update the GIP accordingly.
5089	*
5090	* This is used by supdrvGipMpEvent.
5091	*
5092	* @param pGip The GIP.
5093	* @param idCpu The CPU ID.
5094	*/
5095	static void supdrvGipMpEventOffline(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
5096	{
5097	int iCpuSet;
5098	unsigned i;
5099
5100	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5101	AssertReturnVoid(iCpuSet >= 0);
5102
5103	i = pGip->aiCpuFromCpuSetIdx[iCpuSet];
5104	AssertReturnVoid(i < pGip->cCpus);
5105	AssertReturnVoid(pGip->aCPUs[i].idCpu == idCpu);
5106
5107	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5108	RTCpuSetDelByIndex(&pGip->OnlineCpuSet, iCpuSet);
5109
5110	/* commit it */
5111	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_OFFLINE);
5112	}
5113
5114
5115	/**
5116	* Multiprocessor event notification callback.
5117	*
5118	* This is used to make sure that the GIP master gets passed on to
5119	* another CPU. It also updates the associated CPU data.
5120	*
5121	* @param enmEvent The event.
5122	* @param idCpu The cpu it applies to.
5123	* @param pvUser Pointer to the device extension.
5124	*/
5125	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser)
5126	{
5127	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5128	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5129
5130	AssertRelease(idCpu == RTMpCpuId());
5131	AssertRelease(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
5132
5133	/*
5134	* Update the GIP CPU data.
5135	*/
5136	if (pGip)
5137	{
5138	switch (enmEvent)
5139	{
5140	case RTMPEVENT_ONLINE:
5141	supdrvGipMpEventOnline(pGip, idCpu);
5142	break;
5143	case RTMPEVENT_OFFLINE:
5144	supdrvGipMpEventOffline(pGip, idCpu);
5145	break;
5146
5147	}
5148	}
5149
5150	/*
5151	* Make sure there is a master GIP.
5152	*/
5153	if (enmEvent == RTMPEVENT_OFFLINE)
5154	{
5155	RTCPUID idGipMaster = ASMAtomicReadU32(&pDevExt->idGipMaster);
5156	if (idGipMaster == idCpu)
5157	{
5158	/*
5159	* Find a new GIP master.
5160	*/
5161	bool fIgnored;
5162	unsigned i;
5163	RTCPUID idNewGipMaster = NIL_RTCPUID;
5164	RTCPUSET OnlineCpus;
5165	RTMpGetOnlineSet(&OnlineCpus);
5166
5167	for (i = 0; i < RTCPUSET_MAX_CPUS; i++)
5168	{
5169	RTCPUID idCurCpu = RTMpCpuIdFromSetIndex(i);
5170	if ( RTCpuSetIsMember(&OnlineCpus, idCurCpu)
5171	&& idCurCpu != idGipMaster)
5172	{
5173	idNewGipMaster = idCurCpu;
5174	break;
5175	}
5176	}
5177
5178	Log(("supdrvGipMpEvent: Gip master %#lx -> %#lx\n", (long)idGipMaster, (long)idNewGipMaster));
5179	ASMAtomicCmpXchgSize(&pDevExt->idGipMaster, idNewGipMaster, idGipMaster, fIgnored);
5180	NOREF(fIgnored);
5181	}
5182	}
5183	}
5184
5185
5186	/**
5187	* Callback used by supdrvDetermineAsyncTSC to read the TSC on a CPU.
5188	*
5189	* @param idCpu Ignored.
5190	* @param pvUser1 Where to put the TSC.
5191	* @param pvUser2 Ignored.
5192	*/
5193	static DECLCALLBACK(void) supdrvDetermineAsyncTscWorker(RTCPUID idCpu, void pvUser1, void pvUser2)
5194	{
5195	#if 1
5196	ASMAtomicWriteU64((uint64_t volatile *)pvUser1, ASMReadTSC());
5197	#else
5198	(uint64_t )pvUser1 = ASMReadTSC();
5199	#endif
5200	}
5201
5202
5203	/**
5204	* Determine if Async GIP mode is required because of TSC drift.
5205	*
5206	* When using the default/normal timer code it is essential that the time stamp counter
5207	* (TSC) runs never backwards, that is, a read operation to the counter should return
5208	* a bigger value than any previous read operation. This is guaranteed by the latest
5209	* AMD CPUs and by newer Intel CPUs which never enter the C2 state (P4). In any other
5210	* case we have to choose the asynchronous timer mode.
5211	*
5212	* @param poffMin Pointer to the determined difference between different cores.
5213	* @return false if the time stamp counters appear to be synchronized, true otherwise.
5214	*/
5215	static bool supdrvDetermineAsyncTsc(uint64_t *poffMin)
5216	{
5217	/*
5218	* Just iterate all the cpus 8 times and make sure that the TSC is
5219	* ever increasing. We don't bother taking TSC rollover into account.
5220	*/
5221	int iEndCpu = RTMpGetArraySize();
5222	int iCpu;
5223	int cLoops = 8;
5224	bool fAsync = false;
5225	int rc = VINF_SUCCESS;
5226	uint64_t offMax = 0;
5227	uint64_t offMin = ~(uint64_t)0;
5228	uint64_t PrevTsc = ASMReadTSC();
5229
5230	while (cLoops-- > 0)
5231	{
5232	for (iCpu = 0; iCpu < iEndCpu; iCpu++)
5233	{
5234	uint64_t CurTsc;
5235	rc = RTMpOnSpecific(RTMpCpuIdFromSetIndex(iCpu), supdrvDetermineAsyncTscWorker, &CurTsc, NULL);
5236	if (RT_SUCCESS(rc))
5237	{
5238	if (CurTsc <= PrevTsc)
5239	{
5240	fAsync = true;
5241	offMin = offMax = PrevTsc - CurTsc;
5242	Log(("supdrvDetermineAsyncTsc: iCpu=%d cLoops=%d CurTsc=%llx PrevTsc=%llx\n",
5243	iCpu, cLoops, CurTsc, PrevTsc));
5244	break;
5245	}
5246
5247	/* Gather statistics (except the first time). */
5248	if (iCpu != 0 \|\| cLoops != 7)
5249	{
5250	uint64_t off = CurTsc - PrevTsc;
5251	if (off < offMin)
5252	offMin = off;
5253	if (off > offMax)
5254	offMax = off;
5255	Log2(("%d/%d: off=%llx\n", cLoops, iCpu, off));
5256	}
5257
5258	/* Next */
5259	PrevTsc = CurTsc;
5260	}
5261	else if (rc == VERR_NOT_SUPPORTED)
5262	break;
5263	else
5264	AssertMsg(rc == VERR_CPU_NOT_FOUND \|\| rc == VERR_CPU_OFFLINE, ("%d\n", rc));
5265	}
5266
5267	/* broke out of the loop. */
5268	if (iCpu < iEndCpu)
5269	break;
5270	}
5271
5272	poffMin = offMin; / Almost RTMpOnSpecific profiling. */
5273	Log(("supdrvDetermineAsyncTsc: returns %d; iEndCpu=%d rc=%d offMin=%llx offMax=%llx\n",
5274	fAsync, iEndCpu, rc, offMin, offMax));
5275	#if !defined(RT_OS_SOLARIS) && !defined(RT_OS_OS2) && !defined(RT_OS_WINDOWS)
5276	OSDBGPRINT(("vboxdrv: fAsync=%d offMin=%#lx offMax=%#lx\n", fAsync, (long)offMin, (long)offMax));
5277	#endif
5278	return fAsync;
5279	}
5280
5281
5282	/**
5283	* Determine the GIP TSC mode.
5284	*
5285	* @returns The most suitable TSC mode.
5286	* @param pDevExt Pointer to the device instance data.
5287	*/
5288	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt)
5289	{
5290	/*
5291	* On SMP we're faced with two problems:
5292	* (1) There might be a skew between the CPU, so that cpu0
5293	* returns a TSC that is slightly different from cpu1.
5294	* (2) Power management (and other things) may cause the TSC
5295	* to run at a non-constant speed, and cause the speed
5296	* to be different on the cpus. This will result in (1).
5297	*
5298	* So, on SMP systems we'll have to select the ASYNC update method
5299	* if there are symptoms of these problems.
5300	*/
5301	if (RTMpGetCount() > 1)
5302	{
5303	uint32_t uEAX, uEBX, uECX, uEDX;
5304	uint64_t u64DiffCoresIgnored;
5305
5306	/* Permit the user and/or the OS specific bits to force async mode. */
5307	if (supdrvOSGetForcedAsyncTscMode(pDevExt))
5308	return SUPGIPMODE_ASYNC_TSC;
5309
5310	/* Try check for current differences between the cpus. */
5311	if (supdrvDetermineAsyncTsc(&u64DiffCoresIgnored))
5312	return SUPGIPMODE_ASYNC_TSC;
5313
5314	/*
5315	* If the CPU supports power management and is an AMD one we
5316	* won't trust it unless it has the TscInvariant bit is set.
5317	*/
5318	/* Check for "AuthenticAMD" */
5319	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
5320	if ( uEAX >= 1
5321	&& uEBX == X86_CPUID_VENDOR_AMD_EBX
5322	&& uECX == X86_CPUID_VENDOR_AMD_ECX
5323	&& uEDX == X86_CPUID_VENDOR_AMD_EDX)
5324	{
5325	/* Check for APM support and that TscInvariant is cleared. */
5326	ASMCpuId(0x80000000, &uEAX, &uEBX, &uECX, &uEDX);
5327	if (uEAX >= 0x80000007)
5328	{
5329	ASMCpuId(0x80000007, &uEAX, &uEBX, &uECX, &uEDX);
5330	if ( !(uEDX & RT_BIT(8))/* TscInvariant */
5331	&& (uEDX & 0x3e)) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
5332	return SUPGIPMODE_ASYNC_TSC;
5333	}
5334	}
5335	}
5336	return SUPGIPMODE_SYNC_TSC;
5337	}
5338
5339
5340
5341	/**
5342	* Initializes the GIP data.
5343	*
5344	* @param pDevExt Pointer to the device instance data.
5345	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5346	* @param HCPhys The physical address of the GIP.
5347	* @param u64NanoTS The current nanosecond timestamp.
5348	* @param uUpdateHz The update frequency.
5349	* @param cCpus The CPU count.
5350	*/
5351	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
5352	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned cCpus)
5353	{
5354	size_t const cbGip = RT_ALIGN_Z(RT_OFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), PAGE_SIZE);
5355	unsigned i;
5356	#ifdef DEBUG_DARWIN_GIP
5357	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5358	#else
5359	LogFlow(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5360	#endif
5361
5362	/*
5363	* Initialize the structure.
5364	*/
5365	memset(pGip, 0, cbGip);
5366	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
5367	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
5368	pGip->u32Mode = supdrvGipDeterminTscMode(pDevExt);
5369	pGip->cCpus = (uint16_t)cCpus;
5370	pGip->cPages = (uint16_t)(cbGip / PAGE_SIZE);
5371	pGip->u32UpdateHz = uUpdateHz;
5372	pGip->u32UpdateIntervalNS = 1000000000 / uUpdateHz;
5373	pGip->u64NanoTSLastUpdateHz = u64NanoTS;
5374	RTCpuSetEmpty(&pGip->OnlineCpuSet);
5375	RTCpuSetEmpty(&pGip->PresentCpuSet);
5376	RTMpGetSet(&pGip->PossibleCpuSet);
5377	pGip->cOnlineCpus = RTMpGetOnlineCount();
5378	pGip->cPresentCpus = RTMpGetPresentCount();
5379	pGip->cPossibleCpus = RTMpGetCount();
5380	pGip->idCpuMax = RTMpGetMaxCpuId();
5381	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromApicId); i++)
5382	pGip->aiCpuFromApicId[i] = 0;
5383	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx); i++)
5384	pGip->aiCpuFromCpuSetIdx[i] = UINT16_MAX;
5385
5386	for (i = 0; i < cCpus; i++)
5387	{
5388	pGip->aCPUs[i].u32TransactionId = 2;
5389	pGip->aCPUs[i].u64NanoTS = u64NanoTS;
5390	pGip->aCPUs[i].u64TSC = ASMReadTSC();
5391
5392	pGip->aCPUs[i].enmState = SUPGIPCPUSTATE_INVALID;
5393	pGip->aCPUs[i].idCpu = NIL_RTCPUID;
5394	pGip->aCPUs[i].iCpuSet = -1;
5395	pGip->aCPUs[i].idApic = UINT16_MAX;
5396
5397	/*
5398	* We don't know the following values until we've executed updates.
5399	* So, we'll just pretend it's a 4 GHz CPU and adjust the history it on
5400	* the 2nd timer callout.
5401	*/
5402	pGip->aCPUs[i].u64CpuHz = _4G + 1; /* tstGIP-2 depends on this. */
5403	pGip->aCPUs[i].u32UpdateIntervalTSC
5404	= pGip->aCPUs[i].au32TSCHistory[0]
5405	= pGip->aCPUs[i].au32TSCHistory[1]
5406	= pGip->aCPUs[i].au32TSCHistory[2]
5407	= pGip->aCPUs[i].au32TSCHistory[3]
5408	= pGip->aCPUs[i].au32TSCHistory[4]
5409	= pGip->aCPUs[i].au32TSCHistory[5]
5410	= pGip->aCPUs[i].au32TSCHistory[6]
5411	= pGip->aCPUs[i].au32TSCHistory[7]
5412	= /pGip->aCPUs[i].u64CpuHz/ (uint32_t)(_4G / uUpdateHz);
5413	}
5414
5415	/*
5416	* Link it to the device extension.
5417	*/
5418	pDevExt->pGip = pGip;
5419	pDevExt->HCPhysGip = HCPhys;
5420	pDevExt->cGipUsers = 0;
5421	}
5422
5423
5424	/**
5425	* On CPU initialization callback for RTMpOnAll.
5426	*
5427	* @param idCpu The CPU ID.
5428	* @param pvUser1 The device extension.
5429	* @param pvUser2 The GIP.
5430	*/
5431	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2)
5432	{
5433	/* This is good enough, even though it will update some of the globals a
5434	bit to much. */
5435	supdrvGipMpEventOnline((PSUPGLOBALINFOPAGE)pvUser2, idCpu);
5436	}
5437
5438
5439	/**
5440	* Invalidates the GIP data upon termination.
5441	*
5442	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5443	*/
5444	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
5445	{
5446	unsigned i;
5447	pGip->u32Magic = 0;
5448	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
5449	{
5450	pGip->aCPUs[i].u64NanoTS = 0;
5451	pGip->aCPUs[i].u64TSC = 0;
5452	pGip->aCPUs[i].iTSCHistoryHead = 0;
5453	}
5454	}
5455
5456
5457	/**
5458	* Worker routine for supdrvGipUpdate and supdrvGipUpdatePerCpu that
5459	* updates all the per cpu data except the transaction id.
5460	*
5461	* @param pGip The GIP.
5462	* @param pGipCpu Pointer to the per cpu data.
5463	* @param u64NanoTS The current time stamp.
5464	* @param u64TSC The current TSC.
5465	* @param iTick The current timer tick.
5466	*/
5467	static void supdrvGipDoUpdateCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick)
5468	{
5469	uint64_t u64TSCDelta;
5470	uint32_t u32UpdateIntervalTSC;
5471	uint32_t u32UpdateIntervalTSCSlack;
5472	unsigned iTSCHistoryHead;
5473	uint64_t u64CpuHz;
5474	uint32_t u32TransactionId;
5475
5476	/* Delta between this and the previous update. */
5477	ASMAtomicUoWriteU32(&pGipCpu->u32PrevUpdateIntervalNS, (uint32_t)(u64NanoTS - pGipCpu->u64NanoTS));
5478
5479	/*
5480	* Update the NanoTS.
5481	*/
5482	ASMAtomicWriteU64(&pGipCpu->u64NanoTS, u64NanoTS);
5483
5484	/*
5485	* Calc TSC delta.
5486	*/
5487	/** @todo validate the NanoTS delta, don't trust the OS to call us when it should... */
5488	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
5489	ASMAtomicWriteU64(&pGipCpu->u64TSC, u64TSC);
5490
5491	if (u64TSCDelta >> 32)
5492	{
5493	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
5494	pGipCpu->cErrors++;
5495	}
5496
5497	/*
5498	* On the 2nd and 3rd callout, reset the history with the current TSC
5499	* interval since the values entered by supdrvGipInit are totally off.
5500	* The interval on the 1st callout completely unreliable, the 2nd is a bit
5501	* better, while the 3rd should be most reliable.
5502	*/
5503	u32TransactionId = pGipCpu->u32TransactionId;
5504	if (RT_UNLIKELY( ( u32TransactionId == 5
5505	\|\| u32TransactionId == 7)
5506	&& ( iTick == 2
5507	\|\| iTick == 3) ))
5508	{
5509	unsigned i;
5510	for (i = 0; i < RT_ELEMENTS(pGipCpu->au32TSCHistory); i++)
5511	ASMAtomicUoWriteU32(&pGipCpu->au32TSCHistory[i], (uint32_t)u64TSCDelta);
5512	}
5513
5514	/*
5515	* TSC History.
5516	*/
5517	Assert(RT_ELEMENTS(pGipCpu->au32TSCHistory) == 8);
5518	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
5519	ASMAtomicWriteU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
5520	ASMAtomicWriteU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
5521
5522	/*
5523	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
5524	*/
5525	if (pGip->u32UpdateHz >= 1000)
5526	{
5527	uint32_t u32;
5528	u32 = pGipCpu->au32TSCHistory[0];
5529	u32 += pGipCpu->au32TSCHistory[1];
5530	u32 += pGipCpu->au32TSCHistory[2];
5531	u32 += pGipCpu->au32TSCHistory[3];
5532	u32 >>= 2;
5533	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
5534	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
5535	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
5536	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
5537	u32UpdateIntervalTSC >>= 2;
5538	u32UpdateIntervalTSC += u32;
5539	u32UpdateIntervalTSC >>= 1;
5540
5541	/* Value chosen for a 2GHz Athlon64 running linux 2.6.10/11, . */
5542	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
5543	}
5544	else if (pGip->u32UpdateHz >= 90)
5545	{
5546	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5547	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
5548	u32UpdateIntervalTSC >>= 1;
5549
5550	/* value chosen on a 2GHz thinkpad running windows */
5551	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
5552	}
5553	else
5554	{
5555	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5556
5557	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
5558	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
5559	}
5560	ASMAtomicWriteU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
5561
5562	/*
5563	* CpuHz.
5564	*/
5565	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, pGip->u32UpdateHz);
5566	ASMAtomicWriteU64(&pGipCpu->u64CpuHz, u64CpuHz);
5567	}
5568
5569
5570	/**
5571	* Updates the GIP.
5572	*
5573	* @param pGip Pointer to the GIP.
5574	* @param u64NanoTS The current nanosecond timesamp.
5575	* @param u64TSC The current TSC timesamp.
5576	* @param idCpu The CPU ID.
5577	* @param iTick The current timer tick.
5578	*/
5579	static void supdrvGipUpdate(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick)
5580	{
5581	/*
5582	* Determine the relevant CPU data.
5583	*/
5584	PSUPGIPCPU pGipCpu;
5585	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
5586	pGipCpu = &pGip->aCPUs[0];
5587	else
5588	{
5589	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
5590	if (RT_UNLIKELY(iCpu >= pGip->cCpus))
5591	return;
5592	pGipCpu = &pGip->aCPUs[iCpu];
5593	if (RT_UNLIKELY(pGipCpu->idCpu != idCpu))
5594	return;
5595	}
5596
5597	/*
5598	* Start update transaction.
5599	*/
5600	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5601	{
5602	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
5603	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5604	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5605	pGipCpu->cErrors++;
5606	return;
5607	}
5608
5609	/*
5610	* Recalc the update frequency every 0x800th time.
5611	*/
5612	if (!(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
5613	{
5614	if (pGip->u64NanoTSLastUpdateHz)
5615	{
5616	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
5617	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
5618	uint32_t u32UpdateHz = (uint32_t)((UINT64_C(1000000000) * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
5619	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
5620	{
5621	ASMAtomicWriteU32(&pGip->u32UpdateHz, u32UpdateHz);
5622	ASMAtomicWriteU32(&pGip->u32UpdateIntervalNS, 1000000000 / u32UpdateHz);
5623	}
5624	#endif
5625	}
5626	ASMAtomicWriteU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS);
5627	}
5628
5629	/*
5630	* Update the data.
5631	*/
5632	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS, u64TSC, iTick);
5633
5634	/*
5635	* Complete transaction.
5636	*/
5637	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5638	}
5639
5640
5641	/**
5642	* Updates the per cpu GIP data for the calling cpu.
5643	*
5644	* @param pGip Pointer to the GIP.
5645	* @param u64NanoTS The current nanosecond timesamp.
5646	* @param u64TSC The current TSC timesamp.
5647	* @param idCpu The CPU ID.
5648	* @param idApic The APIC id for the CPU index.
5649	* @param iTick The current timer tick.
5650	*/
5651	static void supdrvGipUpdatePerCpu(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, uint64_t u64TSC,
5652	RTCPUID idCpu, uint8_t idApic, uint64_t iTick)
5653	{
5654	uint32_t iCpu;
5655
5656	/*
5657	* Avoid a potential race when a CPU online notification doesn't fire on
5658	* the onlined CPU but the tick creeps in before the event notification is
5659	* run.
5660	*/
5661	if (RT_UNLIKELY(iTick == 1))
5662	{
5663	iCpu = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
5664	if (pGip->aCPUs[iCpu].enmState == SUPGIPCPUSTATE_OFFLINE)
5665	supdrvGipMpEventOnline(pGip, idCpu);
5666	}
5667
5668	iCpu = pGip->aiCpuFromApicId[idApic];
5669	if (RT_LIKELY(iCpu < pGip->cCpus))
5670	{
5671	PSUPGIPCPU pGipCpu = &pGip->aCPUs[iCpu];
5672	if (pGipCpu->idCpu == idCpu)
5673	{
5674	/*
5675	* Start update transaction.
5676	*/
5677	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5678	{
5679	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5680	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5681	pGipCpu->cErrors++;
5682	return;
5683	}
5684
5685	/*
5686	* Update the data.
5687	*/
5688	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS, u64TSC, iTick);
5689
5690	/*
5691	* Complete transaction.
5692	*/
5693	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5694	}
5695	}
5696	}
5697

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

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDrv.c@ 38183

Download in other formats: