SUPDrv.c@ 40769

Last change on this file since 40769 was 40769, checked in by vboxsync, 13 years ago
HostDrivers/Support: Moved spinlock to DevExt for Mp event race updating the GIP.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 218.5 KB

Line
1	/* $Revision: 40769 $ */
2	/** @file
3	* VBoxDrv - The VirtualBox Support Driver - Common code.
4	*/
5
6	/*
7	* Copyright (C) 2006-2012 Oracle Corporation
8	*
9	* This file is part of VirtualBox Open Source Edition (OSE), as
10	* available from http://www.virtualbox.org. This file is free software;
11	* you can redistribute it and/or modify it under the terms of the GNU
12	* General Public License (GPL) as published by the Free Software
13	* Foundation, in version 2 as it comes in the "COPYING" file of the
14	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16	*
17	* The contents of this file may alternatively be used under the terms
18	* of the Common Development and Distribution License Version 1.0
19	* (CDDL) only, as it comes in the "COPYING.CDDL" file of the
20	* VirtualBox OSE distribution, in which case the provisions of the
21	* CDDL are applicable instead of those of the GPL.
22	*
23	* You may elect to license modified versions of this file under the
24	* terms and conditions of either the GPL or the CDDL or both.
25	*/
26
27	/*******************************************************************************
28	* Header Files *
29	*******************************************************************************/
30	#define LOG_GROUP LOG_GROUP_SUP_DRV
31	#define SUPDRV_AGNOSTIC
32	#include "SUPDrvInternal.h"
33	#ifndef PAGE_SHIFT
34	# include <iprt/param.h>
35	#endif
36	#include <iprt/asm.h>
37	#include <iprt/asm-amd64-x86.h>
38	#include <iprt/asm-math.h>
39	#include <iprt/cpuset.h>
40	#include <iprt/handletable.h>
41	#include <iprt/mem.h>
42	#include <iprt/mp.h>
43	#include <iprt/power.h>
44	#include <iprt/process.h>
45	#include <iprt/semaphore.h>
46	#include <iprt/spinlock.h>
47	#include <iprt/thread.h>
48	#include <iprt/uuid.h>
49	#include <iprt/net.h>
50	#include <iprt/crc.h>
51	#include <iprt/string.h>
52	#include <iprt/timer.h>
53	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
54	# include <iprt/rand.h>
55	# include <iprt/path.h>
56	#endif
57	#include <iprt/x86.h>
58
59	#include <VBox/param.h>
60	#include <VBox/log.h>
61	#include <VBox/err.h>
62	#include <VBox/vmm/hwacc_svm.h>
63	#include <VBox/vmm/hwacc_vmx.h>
64
65	#if defined(RT_OS_SOLARIS) \|\| defined(RT_OS_DARWIN)
66	# include "dtrace/SUPDrv.h"
67	#else
68	# define VBOXDRV_SESSION_CREATE(pvSession, fUser) do { } while (0)
69	# define VBOXDRV_SESSION_CLOSE(pvSession) do { } while (0)
70	# define VBOXDRV_IOCTL_ENTRY(pvSession, uIOCtl, pvReqHdr) do { } while (0)
71	# define VBOXDRV_IOCTL_RETURN(pvSession, uIOCtl, pvReqHdr, rcRet, rcReq) do { } while (0)
72	#endif
73
74	/*
75	* Logging assignments:
76	* Log - useful stuff, like failures.
77	* LogFlow - program flow, except the really noisy bits.
78	* Log2 - Cleanup.
79	* Log3 - Loader flow noise.
80	* Log4 - Call VMMR0 flow noise.
81	* Log5 - Native yet-to-be-defined noise.
82	* Log6 - Native ioctl flow noise.
83	*
84	* Logging requires BUILD_TYPE=debug and possibly changes to the logger
85	* instantiation in log-vbox.c(pp).
86	*/
87
88
89	/*******************************************************************************
90	* Defined Constants And Macros *
91	*******************************************************************************/
92	/** The frequency by which we recalculate the u32UpdateHz and
93	* u32UpdateIntervalNS GIP members. The value must be a power of 2. */
94	#define GIP_UPDATEHZ_RECALC_FREQ 0x800
95
96	/** @def VBOX_SVN_REV
97	* The makefile should define this if it can. */
98	#ifndef VBOX_SVN_REV
99	# define VBOX_SVN_REV 0
100	#endif
101
102	#if 0 /* Don't start the GIP timers. Useful when debugging the IPRT timer code. */
103	# define DO_NOT_START_GIP
104	#endif
105
106
107	/*******************************************************************************
108	* Internal Functions *
109	*******************************************************************************/
110	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser);
111	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser);
112	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession);
113	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType);
114	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq);
115	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq);
116	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq);
117	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq);
118	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq);
119	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx);
120	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt);
121	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage);
122	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage);
123	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt);
124	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt);
125	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq);
126	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq);
127	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt);
128	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt);
129	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
130	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
131	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser);
132	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
133	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned cCpus);
134	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2);
135	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip);
136	static void supdrvGipUpdate(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick);
137	static void supdrvGipUpdatePerCpu(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC,
138	RTCPUID idCpu, uint8_t idApic, uint64_t iTick);
139	static void supdrvGipInitCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pCpu, uint64_t u64NanoTS);
140
141
142	/*******************************************************************************
143	* Global Variables *
144	*******************************************************************************/
145	DECLEXPORT(PSUPGLOBALINFOPAGE) g_pSUPGlobalInfoPage = NULL;
146
147	/**
148	* Array of the R0 SUP API.
149	*/
150	static SUPFUNC g_aFunctions[] =
151	{
152	/* name function */
153	/* Entries with absolute addresses determined at runtime, fixup
154	code makes ugly ASSUMPTIONS about the order here: */
155	{ "SUPR0AbsIs64bit", (void *)0 },
156	{ "SUPR0Abs64bitKernelCS", (void *)0 },
157	{ "SUPR0Abs64bitKernelSS", (void *)0 },
158	{ "SUPR0Abs64bitKernelDS", (void *)0 },
159	{ "SUPR0AbsKernelCS", (void *)0 },
160	{ "SUPR0AbsKernelSS", (void *)0 },
161	{ "SUPR0AbsKernelDS", (void *)0 },
162	{ "SUPR0AbsKernelES", (void *)0 },
163	{ "SUPR0AbsKernelFS", (void *)0 },
164	{ "SUPR0AbsKernelGS", (void *)0 },
165	{ "SUPR0TracerFireProbe", (void *)SUPR0TracerFireProbe },
166	/* Normal function pointers: */
167	{ "SUPR0ComponentRegisterFactory", (void *)SUPR0ComponentRegisterFactory },
168	{ "SUPR0ComponentDeregisterFactory", (void *)SUPR0ComponentDeregisterFactory },
169	{ "SUPR0ComponentQueryFactory", (void *)SUPR0ComponentQueryFactory },
170	{ "SUPR0ObjRegister", (void *)SUPR0ObjRegister },
171	{ "SUPR0ObjAddRef", (void *)SUPR0ObjAddRef },
172	{ "SUPR0ObjAddRefEx", (void *)SUPR0ObjAddRefEx },
173	{ "SUPR0ObjRelease", (void *)SUPR0ObjRelease },
174	{ "SUPR0ObjVerifyAccess", (void *)SUPR0ObjVerifyAccess },
175	{ "SUPR0LockMem", (void *)SUPR0LockMem },
176	{ "SUPR0UnlockMem", (void *)SUPR0UnlockMem },
177	{ "SUPR0ContAlloc", (void *)SUPR0ContAlloc },
178	{ "SUPR0ContFree", (void *)SUPR0ContFree },
179	{ "SUPR0LowAlloc", (void *)SUPR0LowAlloc },
180	{ "SUPR0LowFree", (void *)SUPR0LowFree },
181	{ "SUPR0MemAlloc", (void *)SUPR0MemAlloc },
182	{ "SUPR0MemGetPhys", (void *)SUPR0MemGetPhys },
183	{ "SUPR0MemFree", (void *)SUPR0MemFree },
184	{ "SUPR0PageAllocEx", (void *)SUPR0PageAllocEx },
185	{ "SUPR0PageFree", (void *)SUPR0PageFree },
186	{ "SUPR0Printf", (void )SUPR0Printf }, /* @todo needs wrapping? */
187	{ "SUPSemEventCreate", (void *)SUPSemEventCreate },
188	{ "SUPSemEventClose", (void *)SUPSemEventClose },
189	{ "SUPSemEventSignal", (void *)SUPSemEventSignal },
190	{ "SUPSemEventWait", (void *)SUPSemEventWait },
191	{ "SUPSemEventWaitNoResume", (void *)SUPSemEventWaitNoResume },
192	{ "SUPSemEventWaitNsAbsIntr", (void *)SUPSemEventWaitNsAbsIntr },
193	{ "SUPSemEventWaitNsRelIntr", (void *)SUPSemEventWaitNsRelIntr },
194	{ "SUPSemEventGetResolution", (void *)SUPSemEventGetResolution },
195	{ "SUPSemEventMultiCreate", (void *)SUPSemEventMultiCreate },
196	{ "SUPSemEventMultiClose", (void *)SUPSemEventMultiClose },
197	{ "SUPSemEventMultiSignal", (void *)SUPSemEventMultiSignal },
198	{ "SUPSemEventMultiReset", (void *)SUPSemEventMultiReset },
199	{ "SUPSemEventMultiWait", (void *)SUPSemEventMultiWait },
200	{ "SUPSemEventMultiWaitNoResume", (void *)SUPSemEventMultiWaitNoResume },
201	{ "SUPSemEventMultiWaitNsAbsIntr", (void *)SUPSemEventMultiWaitNsAbsIntr },
202	{ "SUPSemEventMultiWaitNsRelIntr", (void *)SUPSemEventMultiWaitNsRelIntr },
203	{ "SUPSemEventMultiGetResolution", (void *)SUPSemEventMultiGetResolution },
204	{ "SUPR0GetPagingMode", (void *)SUPR0GetPagingMode },
205	{ "SUPR0EnableVTx", (void *)SUPR0EnableVTx },
206	{ "SUPR0TracerRegisterModule", (void *)SUPR0TracerRegisterModule },
207	{ "SUPGetGIP", (void *)SUPGetGIP },
208	{ "g_pSUPGlobalInfoPage", (void *)&g_pSUPGlobalInfoPage },
209	{ "RTMemAllocTag", (void *)RTMemAllocTag },
210	{ "RTMemAllocZTag", (void *)RTMemAllocZTag },
211	{ "RTMemAllocVarTag", (void *)RTMemAllocVarTag },
212	{ "RTMemAllocZVarTag", (void *)RTMemAllocZVarTag },
213	{ "RTMemFree", (void *)RTMemFree },
214	{ "RTMemDupTag", (void *)RTMemDupTag },
215	{ "RTMemDupExTag", (void *)RTMemDupExTag },
216	{ "RTMemReallocTag", (void *)RTMemReallocTag },
217	{ "RTR0MemObjAllocLowTag", (void *)RTR0MemObjAllocLowTag },
218	{ "RTR0MemObjAllocPageTag", (void *)RTR0MemObjAllocPageTag },
219	{ "RTR0MemObjAllocPhysTag", (void *)RTR0MemObjAllocPhysTag },
220	{ "RTR0MemObjAllocPhysExTag", (void *)RTR0MemObjAllocPhysExTag },
221	{ "RTR0MemObjAllocPhysNCTag", (void *)RTR0MemObjAllocPhysNCTag },
222	{ "RTR0MemObjAllocContTag", (void *)RTR0MemObjAllocContTag },
223	{ "RTR0MemObjEnterPhysTag", (void *)RTR0MemObjEnterPhysTag },
224	{ "RTR0MemObjLockUserTag", (void *)RTR0MemObjLockUserTag },
225	{ "RTR0MemObjMapKernelTag", (void *)RTR0MemObjMapKernelTag },
226	{ "RTR0MemObjMapKernelExTag", (void *)RTR0MemObjMapKernelExTag },
227	{ "RTR0MemObjMapUserTag", (void *)RTR0MemObjMapUserTag },
228	{ "RTR0MemObjProtect", (void *)RTR0MemObjProtect },
229	{ "RTR0MemObjAddress", (void *)RTR0MemObjAddress },
230	{ "RTR0MemObjAddressR3", (void *)RTR0MemObjAddressR3 },
231	{ "RTR0MemObjSize", (void *)RTR0MemObjSize },
232	{ "RTR0MemObjIsMapping", (void *)RTR0MemObjIsMapping },
233	{ "RTR0MemObjGetPagePhysAddr", (void *)RTR0MemObjGetPagePhysAddr },
234	{ "RTR0MemObjFree", (void *)RTR0MemObjFree },
235	{ "RTR0MemUserCopyFrom", (void *)RTR0MemUserCopyFrom },
236	{ "RTR0MemUserCopyTo", (void *)RTR0MemUserCopyTo },
237	{ "RTR0MemUserIsValidAddr", (void *)RTR0MemUserIsValidAddr },
238	{ "RTR0MemKernelIsValidAddr", (void *)RTR0MemKernelIsValidAddr },
239	{ "RTR0MemAreKrnlAndUsrDifferent", (void *)RTR0MemAreKrnlAndUsrDifferent },
240	{ "RTSemMutexCreate", (void *)RTSemMutexCreate },
241	{ "RTSemMutexRequest", (void *)RTSemMutexRequest },
242	{ "RTSemMutexRequestDebug", (void *)RTSemMutexRequestDebug },
243	{ "RTSemMutexRequestNoResume", (void *)RTSemMutexRequestNoResume },
244	{ "RTSemMutexRequestNoResumeDebug", (void *)RTSemMutexRequestNoResumeDebug },
245	{ "RTSemMutexRelease", (void *)RTSemMutexRelease },
246	{ "RTSemMutexDestroy", (void *)RTSemMutexDestroy },
247	{ "RTProcSelf", (void *)RTProcSelf },
248	{ "RTR0ProcHandleSelf", (void *)RTR0ProcHandleSelf },
249	{ "RTSemFastMutexCreate", (void *)RTSemFastMutexCreate },
250	{ "RTSemFastMutexDestroy", (void *)RTSemFastMutexDestroy },
251	{ "RTSemFastMutexRequest", (void *)RTSemFastMutexRequest },
252	{ "RTSemFastMutexRelease", (void *)RTSemFastMutexRelease },
253	{ "RTSemEventCreate", (void *)RTSemEventCreate },
254	{ "RTSemEventSignal", (void *)RTSemEventSignal },
255	{ "RTSemEventWait", (void *)RTSemEventWait },
256	{ "RTSemEventWaitNoResume", (void *)RTSemEventWaitNoResume },
257	{ "RTSemEventWaitEx", (void *)RTSemEventWaitEx },
258	{ "RTSemEventWaitExDebug", (void *)RTSemEventWaitExDebug },
259	{ "RTSemEventGetResolution", (void *)RTSemEventGetResolution },
260	{ "RTSemEventDestroy", (void *)RTSemEventDestroy },
261	{ "RTSemEventMultiCreate", (void *)RTSemEventMultiCreate },
262	{ "RTSemEventMultiSignal", (void *)RTSemEventMultiSignal },
263	{ "RTSemEventMultiReset", (void *)RTSemEventMultiReset },
264	{ "RTSemEventMultiWait", (void *)RTSemEventMultiWait },
265	{ "RTSemEventMultiWaitNoResume", (void *)RTSemEventMultiWaitNoResume },
266	{ "RTSemEventMultiWaitEx", (void *)RTSemEventMultiWaitEx },
267	{ "RTSemEventMultiWaitExDebug", (void *)RTSemEventMultiWaitExDebug },
268	{ "RTSemEventMultiGetResolution", (void *)RTSemEventMultiGetResolution },
269	{ "RTSemEventMultiDestroy", (void *)RTSemEventMultiDestroy },
270	{ "RTSpinlockCreate", (void *)RTSpinlockCreate },
271	{ "RTSpinlockDestroy", (void *)RTSpinlockDestroy },
272	{ "RTSpinlockAcquire", (void *)RTSpinlockAcquire },
273	{ "RTSpinlockRelease", (void *)RTSpinlockRelease },
274	{ "RTSpinlockAcquireNoInts", (void *)RTSpinlockAcquireNoInts },
275	{ "RTSpinlockReleaseNoInts", (void *)RTSpinlockReleaseNoInts },
276	{ "RTTimeNanoTS", (void *)RTTimeNanoTS },
277	{ "RTTimeMilliTS", (void *)RTTimeMilliTS },
278	{ "RTTimeSystemNanoTS", (void *)RTTimeSystemNanoTS },
279	{ "RTTimeSystemMilliTS", (void *)RTTimeSystemMilliTS },
280	{ "RTThreadNativeSelf", (void *)RTThreadNativeSelf },
281	{ "RTThreadSleep", (void *)RTThreadSleep },
282	{ "RTThreadYield", (void *)RTThreadYield },
283	{ "RTThreadSelf", (void *)RTThreadSelf },
284	{ "RTThreadCreate", (void *)RTThreadCreate },
285	{ "RTThreadGetNative", (void *)RTThreadGetNative },
286	{ "RTThreadWait", (void *)RTThreadWait },
287	{ "RTThreadWaitNoResume", (void *)RTThreadWaitNoResume },
288	{ "RTThreadGetName", (void *)RTThreadGetName },
289	{ "RTThreadSelfName", (void *)RTThreadSelfName },
290	{ "RTThreadGetType", (void *)RTThreadGetType },
291	{ "RTThreadUserSignal", (void *)RTThreadUserSignal },
292	{ "RTThreadUserReset", (void *)RTThreadUserReset },
293	{ "RTThreadUserWait", (void *)RTThreadUserWait },
294	{ "RTThreadUserWaitNoResume", (void *)RTThreadUserWaitNoResume },
295	{ "RTThreadPreemptIsEnabled", (void *)RTThreadPreemptIsEnabled },
296	{ "RTThreadPreemptIsPending", (void *)RTThreadPreemptIsPending },
297	{ "RTThreadPreemptIsPendingTrusty", (void *)RTThreadPreemptIsPendingTrusty },
298	{ "RTThreadPreemptIsPossible", (void *)RTThreadPreemptIsPossible },
299	{ "RTThreadPreemptDisable", (void *)RTThreadPreemptDisable },
300	{ "RTThreadPreemptRestore", (void *)RTThreadPreemptRestore },
301	{ "RTThreadIsInInterrupt", (void *)RTThreadIsInInterrupt },
302	{ "RTTimerCreate", (void *)RTTimerCreate },
303	{ "RTTimerCreateEx", (void *)RTTimerCreateEx },
304	{ "RTTimerDestroy", (void *)RTTimerDestroy },
305	{ "RTTimerStart", (void *)RTTimerStart },
306	{ "RTTimerStop", (void *)RTTimerStop },
307	{ "RTTimerChangeInterval", (void *)RTTimerChangeInterval },
308	{ "RTTimerGetSystemGranularity", (void *)RTTimerGetSystemGranularity },
309	{ "RTTimerRequestSystemGranularity", (void *)RTTimerRequestSystemGranularity },
310	{ "RTTimerReleaseSystemGranularity", (void *)RTTimerReleaseSystemGranularity },
311	{ "RTTimerCanDoHighResolution", (void *)RTTimerCanDoHighResolution },
312
313	{ "RTLogDefaultInstance", (void *)RTLogDefaultInstance },
314	{ "RTMpCpuId", (void *)RTMpCpuId },
315	{ "RTMpCpuIdFromSetIndex", (void *)RTMpCpuIdFromSetIndex },
316	{ "RTMpCpuIdToSetIndex", (void *)RTMpCpuIdToSetIndex },
317	{ "RTMpGetArraySize", (void *)RTMpGetArraySize },
318	{ "RTMpIsCpuPossible", (void *)RTMpIsCpuPossible },
319	{ "RTMpGetCount", (void *)RTMpGetCount },
320	{ "RTMpGetMaxCpuId", (void *)RTMpGetMaxCpuId },
321	{ "RTMpGetOnlineCount", (void *)RTMpGetOnlineCount },
322	{ "RTMpGetOnlineSet", (void *)RTMpGetOnlineSet },
323	{ "RTMpGetSet", (void *)RTMpGetSet },
324	{ "RTMpIsCpuOnline", (void *)RTMpIsCpuOnline },
325	{ "RTMpIsCpuWorkPending", (void *)RTMpIsCpuWorkPending },
326	{ "RTMpNotificationRegister", (void *)RTMpNotificationRegister },
327	{ "RTMpNotificationDeregister", (void *)RTMpNotificationDeregister },
328	{ "RTMpOnAll", (void *)RTMpOnAll },
329	{ "RTMpOnOthers", (void *)RTMpOnOthers },
330	{ "RTMpOnSpecific", (void *)RTMpOnSpecific },
331	{ "RTMpPokeCpu", (void *)RTMpPokeCpu },
332	{ "RTPowerNotificationRegister", (void *)RTPowerNotificationRegister },
333	{ "RTPowerNotificationDeregister", (void *)RTPowerNotificationDeregister },
334	{ "RTLogRelDefaultInstance", (void *)RTLogRelDefaultInstance },
335	{ "RTLogSetDefaultInstanceThread", (void *)RTLogSetDefaultInstanceThread },
336	{ "RTLogLoggerExV", (void *)RTLogLoggerExV },
337	{ "RTLogPrintfV", (void *)RTLogPrintfV },
338	{ "RTR0AssertPanicSystem", (void *)RTR0AssertPanicSystem },
339	{ "RTAssertMsg1", (void *)RTAssertMsg1 },
340	{ "RTAssertMsg2V", (void *)RTAssertMsg2V },
341	{ "RTAssertMsg2AddV", (void *)RTAssertMsg2AddV },
342	{ "RTAssertSetQuiet", (void *)RTAssertSetQuiet },
343	{ "RTAssertMayPanic", (void *)RTAssertMayPanic },
344	{ "RTAssertSetMayPanic", (void *)RTAssertSetMayPanic },
345	{ "RTAssertAreQuiet", (void *)RTAssertAreQuiet },
346	{ "RTStrFormat", (void *)RTStrFormat },
347	{ "RTStrFormatNumber", (void *)RTStrFormatNumber },
348	{ "RTStrFormatTypeDeregister", (void *)RTStrFormatTypeDeregister },
349	{ "RTStrFormatTypeRegister", (void *)RTStrFormatTypeRegister },
350	{ "RTStrFormatTypeSetUser", (void *)RTStrFormatTypeSetUser },
351	{ "RTStrFormatV", (void *)RTStrFormatV },
352	{ "RTStrPrintf", (void *)RTStrPrintf },
353	{ "RTStrPrintfEx", (void *)RTStrPrintfEx },
354	{ "RTStrPrintfExV", (void *)RTStrPrintfExV },
355	{ "RTStrPrintfV", (void *)RTStrPrintfV },
356	{ "RTHandleTableAllocWithCtx", (void *)RTHandleTableAllocWithCtx },
357	{ "RTHandleTableCreate", (void *)RTHandleTableCreate },
358	{ "RTHandleTableCreateEx", (void *)RTHandleTableCreateEx },
359	{ "RTHandleTableDestroy", (void *)RTHandleTableDestroy },
360	{ "RTHandleTableFreeWithCtx", (void *)RTHandleTableFreeWithCtx },
361	{ "RTHandleTableLookupWithCtx", (void *)RTHandleTableLookupWithCtx },
362	{ "RTNetIPv4AddDataChecksum", (void *)RTNetIPv4AddDataChecksum },
363	{ "RTNetIPv4AddTCPChecksum", (void *)RTNetIPv4AddTCPChecksum },
364	{ "RTNetIPv4AddUDPChecksum", (void *)RTNetIPv4AddUDPChecksum },
365	{ "RTNetIPv4FinalizeChecksum", (void *)RTNetIPv4FinalizeChecksum },
366	{ "RTNetIPv4HdrChecksum", (void *)RTNetIPv4HdrChecksum },
367	{ "RTNetIPv4IsDHCPValid", (void *)RTNetIPv4IsDHCPValid },
368	{ "RTNetIPv4IsHdrValid", (void *)RTNetIPv4IsHdrValid },
369	{ "RTNetIPv4IsTCPSizeValid", (void *)RTNetIPv4IsTCPSizeValid },
370	{ "RTNetIPv4IsTCPValid", (void *)RTNetIPv4IsTCPValid },
371	{ "RTNetIPv4IsUDPSizeValid", (void *)RTNetIPv4IsUDPSizeValid },
372	{ "RTNetIPv4IsUDPValid", (void *)RTNetIPv4IsUDPValid },
373	{ "RTNetIPv4PseudoChecksum", (void *)RTNetIPv4PseudoChecksum },
374	{ "RTNetIPv4PseudoChecksumBits", (void *)RTNetIPv4PseudoChecksumBits },
375	{ "RTNetIPv4TCPChecksum", (void *)RTNetIPv4TCPChecksum },
376	{ "RTNetIPv4UDPChecksum", (void *)RTNetIPv4UDPChecksum },
377	{ "RTNetIPv6PseudoChecksum", (void *)RTNetIPv6PseudoChecksum },
378	{ "RTNetIPv6PseudoChecksumBits", (void *)RTNetIPv6PseudoChecksumBits },
379	{ "RTNetIPv6PseudoChecksumEx", (void *)RTNetIPv6PseudoChecksumEx },
380	{ "RTNetTCPChecksum", (void *)RTNetTCPChecksum },
381	{ "RTNetUDPChecksum", (void *)RTNetUDPChecksum },
382	{ "RTCrc32", (void *)RTCrc32 },
383	{ "RTCrc32Finish", (void *)RTCrc32Finish },
384	{ "RTCrc32Process", (void *)RTCrc32Process },
385	{ "RTCrc32Start", (void *)RTCrc32Start },
386	};
387
388	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
389	/**
390	* Drag in the rest of IRPT since we share it with the
391	* rest of the kernel modules on darwin.
392	*/
393	PFNRT g_apfnVBoxDrvIPRTDeps[] =
394	{
395	/* VBoxNetFlt */
396	(PFNRT)RTErrConvertFromErrno,
397	(PFNRT)RTUuidCompare,
398	(PFNRT)RTUuidCompareStr,
399	(PFNRT)RTUuidFromStr,
400	(PFNRT)RTStrDupTag,
401	(PFNRT)RTStrFree,
402	(PFNRT)RTStrCopy,
403	(PFNRT)RTStrNCmp,
404	/* VBoxNetAdp */
405	(PFNRT)RTRandBytes,
406	/* VBoxUSB */
407	(PFNRT)RTPathStripFilename,
408	NULL
409	};
410	#endif /* RT_OS_DARWIN \|\| RT_OS_SOLARIS \|\| RT_OS_SOLARIS */
411
412
413	/**
414	* Initializes the device extentsion structure.
415	*
416	* @returns IPRT status code.
417	* @param pDevExt The device extension to initialize.
418	* @param cbSession The size of the session structure. The size of
419	* SUPDRVSESSION may be smaller when SUPDRV_AGNOSTIC is
420	* defined because we're skipping the OS specific members
421	* then.
422	*/
423	int VBOXCALL supdrvInitDevExt(PSUPDRVDEVEXT pDevExt, size_t cbSession)
424	{
425	int rc;
426
427	#ifdef SUPDRV_WITH_RELEASE_LOGGER
428	/*
429	* Create the release log.
430	*/
431	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
432	PRTLOGGER pRelLogger;
433	rc = RTLogCreate(&pRelLogger, 0 /* fFlags */, "all",
434	"VBOX_RELEASE_LOG", RT_ELEMENTS(s_apszGroups), s_apszGroups, RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER, NULL);
435	if (RT_SUCCESS(rc))
436	RTLogRelSetDefaultInstance(pRelLogger);
437	/** @todo Add native hook for getting logger config parameters and setting
438	* them. On linux we should use the module parameter stuff... */
439	#endif
440
441	/*
442	* Initialize it.
443	*/
444	memset(pDevExt, 0, sizeof(*pDevExt));
445	rc = RTSpinlockCreate(&pDevExt->Spinlock);
446	if (RT_SUCCESS(rc))
447	{
448	rc = RTSpinlockCreate(&pDevExt->spinGip);
449	if (RT_SUCCESS(rc))
450	{
451	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
452	rc = RTSemMutexCreate(&pDevExt->mtxLdr);
453	#else
454	rc = RTSemFastMutexCreate(&pDevExt->mtxLdr);
455	#endif
456	if (RT_SUCCESS(rc))
457	{
458	rc = RTSemFastMutexCreate(&pDevExt->mtxComponentFactory);
459	if (RT_SUCCESS(rc))
460	{
461	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
462	rc = RTSemMutexCreate(&pDevExt->mtxGip);
463	#else
464	rc = RTSemFastMutexCreate(&pDevExt->mtxGip);
465	#endif
466	if (RT_SUCCESS(rc))
467	{
468	rc = supdrvGipCreate(pDevExt);
469	if (RT_SUCCESS(rc))
470	{
471	#ifdef VBOX_WITH_SUPDRV_GENERIC_TRACER
472	rc = supdrvTracerInit(pDevExt);
473	#elif defined(VBOX_WITH_DTRACE_R0DRV)
474	rc = supdrvVtgInit(pDevExt, &g_aFunctions[10]);
475	if (RT_SUCCESS(rc))
476	#endif
477	{
478	pDevExt->pLdrInitImage = NULL;
479	pDevExt->hLdrInitThread = NIL_RTNATIVETHREAD;
480	pDevExt->u32Cookie = BIRD; /** @todo make this random? */
481	pDevExt->cbSession = (uint32_t)cbSession;
482
483	/*
484	* Fixup the absolute symbols.
485	*
486	* Because of the table indexing assumptions we'll have a little #ifdef orgy
487	* here rather than distributing this to OS specific files. At least for now.
488	*/
489	#ifdef RT_OS_DARWIN
490	# if ARCH_BITS == 32
491	if (SUPR0GetPagingMode() >= SUPPAGINGMODE_AMD64)
492	{
493	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
494	g_aFunctions[1].pfn = (void )0x80; / SUPR0Abs64bitKernelCS - KERNEL64_CS, seg.h */
495	g_aFunctions[2].pfn = (void )0x88; / SUPR0Abs64bitKernelSS - KERNEL64_SS, seg.h */
496	g_aFunctions[3].pfn = (void )0x88; / SUPR0Abs64bitKernelDS - KERNEL64_SS, seg.h */
497	}
498	else
499	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
500	g_aFunctions[4].pfn = (void )0x08; / SUPR0AbsKernelCS - KERNEL_CS, seg.h */
501	g_aFunctions[5].pfn = (void )0x10; / SUPR0AbsKernelSS - KERNEL_DS, seg.h */
502	g_aFunctions[6].pfn = (void )0x10; / SUPR0AbsKernelDS - KERNEL_DS, seg.h */
503	g_aFunctions[7].pfn = (void )0x10; / SUPR0AbsKernelES - KERNEL_DS, seg.h */
504	g_aFunctions[8].pfn = (void )0x10; / SUPR0AbsKernelFS - KERNEL_DS, seg.h */
505	g_aFunctions[9].pfn = (void )0x48; / SUPR0AbsKernelGS - CPU_DATA_GS, seg.h */
506	# else /* 64-bit darwin: */
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 )0; / SUPR0Abs64bitKernelDS */
511	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
512	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
513	g_aFunctions[6].pfn = (void )0; / SUPR0AbsKernelDS */
514	g_aFunctions[7].pfn = (void )0; / SUPR0AbsKernelES */
515	g_aFunctions[8].pfn = (void )0; / SUPR0AbsKernelFS */
516	g_aFunctions[9].pfn = (void )0; / SUPR0AbsKernelGS */
517
518	# endif
519	#else /* !RT_OS_DARWIN */
520	# if ARCH_BITS == 64
521	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
522	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
523	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
524	g_aFunctions[3].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0Abs64bitKernelDS */
525	# else
526	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
527	# endif
528	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
529	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
530	g_aFunctions[6].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0AbsKernelDS */
531	g_aFunctions[7].pfn = (void )(uintptr_t)ASMGetES(); / SUPR0AbsKernelES */
532	g_aFunctions[8].pfn = (void )(uintptr_t)ASMGetFS(); / SUPR0AbsKernelFS */
533	g_aFunctions[9].pfn = (void )(uintptr_t)ASMGetGS(); / SUPR0AbsKernelGS */
534	#endif /* !RT_OS_DARWIN */
535	return VINF_SUCCESS;
536	}
537
538	#if defined(VBOX_WITH_SUPDRV_GENERIC_TRACER) \|\| defined(VBOX_WITH_DTRACE_R0DRV)
539	supdrvGipDestroy(pDevExt);
540	#endif
541	}
542
543	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
544	RTSemMutexDestroy(pDevExt->mtxGip);
545	pDevExt->mtxGip = NIL_RTSEMMUTEX;
546	#else
547	RTSemFastMutexDestroy(pDevExt->mtxGip);
548	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
549	#endif
550	}
551	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
552	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
553	}
554	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
555	RTSemMutexDestroy(pDevExt->mtxLdr);
556	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
557	#else
558	RTSemFastMutexDestroy(pDevExt->mtxLdr);
559	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
560	#endif
561	}
562	RTSpinlockDestroy(pDevExt->spinGip);
563	pDevExt->spinGip = NIL_RTSPINLOCK;
564	}
565	RTSpinlockDestroy(pDevExt->Spinlock);
566	pDevExt->Spinlock = NIL_RTSPINLOCK;
567	}
568	#ifdef SUPDRV_WITH_RELEASE_LOGGER
569	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
570	RTLogDestroy(RTLogSetDefaultInstance(NULL));
571	#endif
572
573	return rc;
574	}
575
576
577	/**
578	* Delete the device extension (e.g. cleanup members).
579	*
580	* @param pDevExt The device extension to delete.
581	*/
582	void VBOXCALL supdrvDeleteDevExt(PSUPDRVDEVEXT pDevExt)
583	{
584	PSUPDRVOBJ pObj;
585	PSUPDRVUSAGE pUsage;
586
587	/*
588	* Kill mutexes and spinlocks.
589	*/
590	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
591	RTSemMutexDestroy(pDevExt->mtxGip);
592	pDevExt->mtxGip = NIL_RTSEMMUTEX;
593	#else
594	RTSemFastMutexDestroy(pDevExt->mtxGip);
595	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
596	#endif
597	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
598	RTSemMutexDestroy(pDevExt->mtxLdr);
599	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
600	#else
601	RTSemFastMutexDestroy(pDevExt->mtxLdr);
602	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
603	#endif
604	RTSpinlockDestroy(pDevExt->Spinlock);
605	pDevExt->Spinlock = NIL_RTSPINLOCK;
606	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
607	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
608
609	/*
610	* Free lists.
611	*/
612	/* objects. */
613	pObj = pDevExt->pObjs;
614	Assert(!pObj); /* (can trigger on forced unloads) */
615	pDevExt->pObjs = NULL;
616	while (pObj)
617	{
618	void *pvFree = pObj;
619	pObj = pObj->pNext;
620	RTMemFree(pvFree);
621	}
622
623	/* usage records. */
624	pUsage = pDevExt->pUsageFree;
625	pDevExt->pUsageFree = NULL;
626	while (pUsage)
627	{
628	void *pvFree = pUsage;
629	pUsage = pUsage->pNext;
630	RTMemFree(pvFree);
631	}
632
633	/* kill the GIP. */
634	supdrvGipDestroy(pDevExt);
635	RTSpinlockDestroy(pDevExt->spinGip);
636	pDevExt->spinGip = NIL_RTSPINLOCK;
637
638	#ifdef VBOX_WITH_SUPDRV_GENERIC_TRACER
639	supdrvTracerTerm(pDevExt);
640	#elif defined(VBOX_WITH_DTRACE_R0DRV)
641	supdrvVtgTerm(pDevExt);
642	#endif
643
644	#ifdef SUPDRV_WITH_RELEASE_LOGGER
645	/* destroy the loggers. */
646	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
647	RTLogDestroy(RTLogSetDefaultInstance(NULL));
648	#endif
649	}
650
651
652	/**
653	* Create session.
654	*
655	* @returns IPRT status code.
656	* @param pDevExt Device extension.
657	* @param fUser Flag indicating whether this is a user or kernel session.
658	* @param ppSession Where to store the pointer to the session data.
659	*/
660	int VBOXCALL supdrvCreateSession(PSUPDRVDEVEXT pDevExt, bool fUser, PSUPDRVSESSION *ppSession)
661	{
662	/*
663	* Allocate memory for the session data.
664	*/
665	int rc;
666	PSUPDRVSESSION pSession = *ppSession = (PSUPDRVSESSION)RTMemAllocZ(pDevExt->cbSession);
667	if (pSession)
668	{
669	/* Initialize session data. */
670	rc = RTSpinlockCreate(&pSession->Spinlock);
671	if (!rc)
672	{
673	rc = RTHandleTableCreateEx(&pSession->hHandleTable,
674	RTHANDLETABLE_FLAGS_LOCKED \| RTHANDLETABLE_FLAGS_CONTEXT,
675	1 /uBase/, 32768 /cMax/, supdrvSessionObjHandleRetain, pSession);
676	if (RT_SUCCESS(rc))
677	{
678	Assert(pSession->Spinlock != NIL_RTSPINLOCK);
679	pSession->pDevExt = pDevExt;
680	pSession->u32Cookie = BIRD_INV;
681	/*pSession->pLdrUsage = NULL;
682	pSession->pVM = NULL;
683	pSession->pUsage = NULL;
684	pSession->pGip = NULL;
685	pSession->fGipReferenced = false;
686	pSession->Bundle.cUsed = 0; */
687	pSession->Uid = NIL_RTUID;
688	pSession->Gid = NIL_RTGID;
689	if (fUser)
690	{
691	pSession->Process = RTProcSelf();
692	pSession->R0Process = RTR0ProcHandleSelf();
693	}
694	else
695	{
696	pSession->Process = NIL_RTPROCESS;
697	pSession->R0Process = NIL_RTR0PROCESS;
698	}
699
700	VBOXDRV_SESSION_CREATE(pSession, fUser);
701	LogFlow(("Created session %p initial cookie=%#x\n", pSession, pSession->u32Cookie));
702	return VINF_SUCCESS;
703	}
704
705	RTSpinlockDestroy(pSession->Spinlock);
706	}
707	RTMemFree(pSession);
708	*ppSession = NULL;
709	Log(("Failed to create spinlock, rc=%d!\n", rc));
710	}
711	else
712	rc = VERR_NO_MEMORY;
713
714	return rc;
715	}
716
717
718	/**
719	* Shared code for cleaning up a session.
720	*
721	* @param pDevExt Device extension.
722	* @param pSession Session data.
723	* This data will be freed by this routine.
724	*/
725	void VBOXCALL supdrvCloseSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
726	{
727	VBOXDRV_SESSION_CLOSE(pSession);
728
729	/*
730	* Cleanup the session first.
731	*/
732	supdrvCleanupSession(pDevExt, pSession);
733
734	/*
735	* Free the rest of the session stuff.
736	*/
737	RTSpinlockDestroy(pSession->Spinlock);
738	pSession->Spinlock = NIL_RTSPINLOCK;
739	pSession->pDevExt = NULL;
740	RTMemFree(pSession);
741	LogFlow(("supdrvCloseSession: returns\n"));
742	}
743
744
745	/**
746	* Shared code for cleaning up a session (but not quite freeing it).
747	*
748	* This is primarily intended for MAC OS X where we have to clean up the memory
749	* stuff before the file handle is closed.
750	*
751	* @param pDevExt Device extension.
752	* @param pSession Session data.
753	* This data will be freed by this routine.
754	*/
755	void VBOXCALL supdrvCleanupSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
756	{
757	int rc;
758	PSUPDRVBUNDLE pBundle;
759	LogFlow(("supdrvCleanupSession: pSession=%p\n", pSession));
760
761	/*
762	* Remove logger instances related to this session.
763	*/
764	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pSession);
765
766	/*
767	* Destroy the handle table.
768	*/
769	rc = RTHandleTableDestroy(pSession->hHandleTable, supdrvSessionObjHandleDelete, pSession);
770	AssertRC(rc);
771	pSession->hHandleTable = NIL_RTHANDLETABLE;
772
773	/*
774	* Release object references made in this session.
775	* In theory there should be noone racing us in this session.
776	*/
777	Log2(("release objects - start\n"));
778	if (pSession->pUsage)
779	{
780	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
781	PSUPDRVUSAGE pUsage;
782	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
783
784	while ((pUsage = pSession->pUsage) != NULL)
785	{
786	PSUPDRVOBJ pObj = pUsage->pObj;
787	pSession->pUsage = pUsage->pNext;
788
789	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
790	if (pUsage->cUsage < pObj->cUsage)
791	{
792	pObj->cUsage -= pUsage->cUsage;
793	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
794	}
795	else
796	{
797	/* Destroy the object and free the record. */
798	if (pDevExt->pObjs == pObj)
799	pDevExt->pObjs = pObj->pNext;
800	else
801	{
802	PSUPDRVOBJ pObjPrev;
803	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
804	if (pObjPrev->pNext == pObj)
805	{
806	pObjPrev->pNext = pObj->pNext;
807	break;
808	}
809	Assert(pObjPrev);
810	}
811	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
812
813	Log(("supdrvCleanupSession: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
814	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
815	if (pObj->pfnDestructor)
816	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
817	RTMemFree(pObj);
818	}
819
820	/* free it and continue. */
821	RTMemFree(pUsage);
822
823	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
824	}
825
826	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
827	AssertMsg(!pSession->pUsage, ("Some buster reregistered an object during desturction!\n"));
828	}
829	Log2(("release objects - done\n"));
830
831	#ifdef VBOX_WITH_SUPDRV_GENERIC_TRACER
832	/*
833	* Do tracer cleanups related to this session.
834	*/
835	Log2(("release tracer stuff - start\n"));
836	supdrvTracerCleanupSession(pDevExt, pSession);
837	Log2(("release tracer stuff - end\n"));
838	#endif
839
840	/*
841	* Release memory allocated in the session.
842	*
843	* We do not serialize this as we assume that the application will
844	* not allocated memory while closing the file handle object.
845	*/
846	Log2(("freeing memory:\n"));
847	pBundle = &pSession->Bundle;
848	while (pBundle)
849	{
850	PSUPDRVBUNDLE pToFree;
851	unsigned i;
852
853	/*
854	* Check and unlock all entries in the bundle.
855	*/
856	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
857	{
858	if (pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ)
859	{
860	Log2(("eType=%d pvR0=%p pvR3=%p cb=%ld\n", pBundle->aMem[i].eType, RTR0MemObjAddress(pBundle->aMem[i].MemObj),
861	(void *)RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3), (long)RTR0MemObjSize(pBundle->aMem[i].MemObj)));
862	if (pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ)
863	{
864	rc = RTR0MemObjFree(pBundle->aMem[i].MapObjR3, false);
865	AssertRC(rc); /** @todo figure out how to handle this. */
866	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
867	}
868	rc = RTR0MemObjFree(pBundle->aMem[i].MemObj, true /* fFreeMappings */);
869	AssertRC(rc); /** @todo figure out how to handle this. */
870	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
871	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
872	}
873	}
874
875	/*
876	* Advance and free previous bundle.
877	*/
878	pToFree = pBundle;
879	pBundle = pBundle->pNext;
880
881	pToFree->pNext = NULL;
882	pToFree->cUsed = 0;
883	if (pToFree != &pSession->Bundle)
884	RTMemFree(pToFree);
885	}
886	Log2(("freeing memory - done\n"));
887
888	/*
889	* Deregister component factories.
890	*/
891	RTSemFastMutexRequest(pDevExt->mtxComponentFactory);
892	Log2(("deregistering component factories:\n"));
893	if (pDevExt->pComponentFactoryHead)
894	{
895	PSUPDRVFACTORYREG pPrev = NULL;
896	PSUPDRVFACTORYREG pCur = pDevExt->pComponentFactoryHead;
897	while (pCur)
898	{
899	if (pCur->pSession == pSession)
900	{
901	/* unlink it */
902	PSUPDRVFACTORYREG pNext = pCur->pNext;
903	if (pPrev)
904	pPrev->pNext = pNext;
905	else
906	pDevExt->pComponentFactoryHead = pNext;
907
908	/* free it */
909	pCur->pNext = NULL;
910	pCur->pSession = NULL;
911	pCur->pFactory = NULL;
912	RTMemFree(pCur);
913
914	/* next */
915	pCur = pNext;
916	}
917	else
918	{
919	/* next */
920	pPrev = pCur;
921	pCur = pCur->pNext;
922	}
923	}
924	}
925	RTSemFastMutexRelease(pDevExt->mtxComponentFactory);
926	Log2(("deregistering component factories - done\n"));
927
928	/*
929	* Loaded images needs to be dereferenced and possibly freed up.
930	*/
931	supdrvLdrLock(pDevExt);
932	Log2(("freeing images:\n"));
933	if (pSession->pLdrUsage)
934	{
935	PSUPDRVLDRUSAGE pUsage = pSession->pLdrUsage;
936	pSession->pLdrUsage = NULL;
937	while (pUsage)
938	{
939	void *pvFree = pUsage;
940	PSUPDRVLDRIMAGE pImage = pUsage->pImage;
941	if (pImage->cUsage > pUsage->cUsage)
942	pImage->cUsage -= pUsage->cUsage;
943	else
944	supdrvLdrFree(pDevExt, pImage);
945	pUsage->pImage = NULL;
946	pUsage = pUsage->pNext;
947	RTMemFree(pvFree);
948	}
949	}
950	supdrvLdrUnlock(pDevExt);
951	Log2(("freeing images - done\n"));
952
953	/*
954	* Unmap the GIP.
955	*/
956	Log2(("umapping GIP:\n"));
957	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
958	{
959	SUPR0GipUnmap(pSession);
960	pSession->fGipReferenced = 0;
961	}
962	Log2(("umapping GIP - done\n"));
963	}
964
965
966	/**
967	* RTHandleTableDestroy callback used by supdrvCleanupSession.
968	*
969	* @returns IPRT status code, see SUPR0ObjAddRef.
970	* @param hHandleTable The handle table handle. Ignored.
971	* @param pvObj The object pointer.
972	* @param pvCtx Context, the handle type. Ignored.
973	* @param pvUser Session pointer.
974	*/
975	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser)
976	{
977	NOREF(pvCtx);
978	NOREF(hHandleTable);
979	return SUPR0ObjAddRefEx(pvObj, (PSUPDRVSESSION)pvUser, true /fNoBlocking/);
980	}
981
982
983	/**
984	* RTHandleTableDestroy callback used by supdrvCleanupSession.
985	*
986	* @param hHandleTable The handle table handle. Ignored.
987	* @param h The handle value. Ignored.
988	* @param pvObj The object pointer.
989	* @param pvCtx Context, the handle type. Ignored.
990	* @param pvUser Session pointer.
991	*/
992	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser)
993	{
994	NOREF(pvCtx);
995	NOREF(h);
996	NOREF(hHandleTable);
997	SUPR0ObjRelease(pvObj, (PSUPDRVSESSION)pvUser);
998	}
999
1000
1001	/**
1002	* Fast path I/O Control worker.
1003	*
1004	* @returns VBox status code that should be passed down to ring-3 unchanged.
1005	* @param uIOCtl Function number.
1006	* @param idCpu VMCPU id.
1007	* @param pDevExt Device extention.
1008	* @param pSession Session data.
1009	*/
1010	int VBOXCALL supdrvIOCtlFast(uintptr_t uIOCtl, VMCPUID idCpu, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
1011	{
1012	/*
1013	* We check the two prereqs after doing this only to allow the compiler to optimize things better.
1014	*/
1015	if (RT_LIKELY( RT_VALID_PTR(pSession)
1016	&& pSession->pVM
1017	&& pDevExt->pfnVMMR0EntryFast))
1018	{
1019	switch (uIOCtl)
1020	{
1021	case SUP_IOCTL_FAST_DO_RAW_RUN:
1022	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_RAW_RUN);
1023	break;
1024	case SUP_IOCTL_FAST_DO_HWACC_RUN:
1025	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_HWACC_RUN);
1026	break;
1027	case SUP_IOCTL_FAST_DO_NOP:
1028	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_NOP);
1029	break;
1030	default:
1031	return VERR_INTERNAL_ERROR;
1032	}
1033	return VINF_SUCCESS;
1034	}
1035	return VERR_INTERNAL_ERROR;
1036	}
1037
1038
1039	/**
1040	* Helper for supdrvIOCtl. Check if pszStr contains any character of pszChars.
1041	* We would use strpbrk here if this function would be contained in the RedHat kABI white
1042	* list, see http://www.kerneldrivers.org/RHEL5.
1043	*
1044	* @returns 1 if pszStr does contain any character of pszChars, 0 otherwise.
1045	* @param pszStr String to check
1046	* @param pszChars Character set
1047	*/
1048	static int supdrvCheckInvalidChar(const char pszStr, const char pszChars)
1049	{
1050	int chCur;
1051	while ((chCur = *pszStr++) != '\0')
1052	{
1053	int ch;
1054	const char *psz = pszChars;
1055	while ((ch = *psz++) != '\0')
1056	if (ch == chCur)
1057	return 1;
1058
1059	}
1060	return 0;
1061	}
1062
1063
1064	/**
1065	* I/O Control worker.
1066	*
1067	* @returns IPRT status code.
1068	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1069	*
1070	* @param uIOCtl Function number.
1071	* @param pDevExt Device extention.
1072	* @param pSession Session data.
1073	* @param pReqHdr The request header.
1074	*/
1075	int VBOXCALL supdrvIOCtl(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1076	{
1077	VBOXDRV_IOCTL_ENTRY(pSession, uIOCtl, pReqHdr);
1078
1079	/*
1080	* Validate the request.
1081	*/
1082	/* this first check could probably be omitted as its also done by the OS specific code... */
1083	if (RT_UNLIKELY( (pReqHdr->fFlags & SUPREQHDR_FLAGS_MAGIC_MASK) != SUPREQHDR_FLAGS_MAGIC
1084	\|\| pReqHdr->cbIn < sizeof(*pReqHdr)
1085	\|\| pReqHdr->cbOut < sizeof(*pReqHdr)))
1086	{
1087	OSDBGPRINT(("vboxdrv: Bad ioctl request header; cbIn=%#lx cbOut=%#lx fFlags=%#lx\n",
1088	(long)pReqHdr->cbIn, (long)pReqHdr->cbOut, (long)pReqHdr->fFlags));
1089	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1090	return VERR_INVALID_PARAMETER;
1091	}
1092	if (RT_UNLIKELY(!RT_VALID_PTR(pSession)))
1093	{
1094	OSDBGPRINT(("vboxdrv: Invalid pSession valud %p (ioctl=%p)\n", pSession, (void *)uIOCtl));
1095	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1096	return VERR_INVALID_PARAMETER;
1097	}
1098	if (RT_UNLIKELY(uIOCtl == SUP_IOCTL_COOKIE))
1099	{
1100	if (pReqHdr->u32Cookie != SUPCOOKIE_INITIAL_COOKIE)
1101	{
1102	OSDBGPRINT(("SUP_IOCTL_COOKIE: bad cookie %#lx\n", (long)pReqHdr->u32Cookie));
1103	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1104	return VERR_INVALID_PARAMETER;
1105	}
1106	}
1107	else if (RT_UNLIKELY( pReqHdr->u32Cookie != pDevExt->u32Cookie
1108	\|\| pReqHdr->u32SessionCookie != pSession->u32Cookie))
1109	{
1110	OSDBGPRINT(("vboxdrv: bad cookie %#lx / %#lx.\n", (long)pReqHdr->u32Cookie, (long)pReqHdr->u32SessionCookie));
1111	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1112	return VERR_INVALID_PARAMETER;
1113	}
1114
1115	/*
1116	* Validation macros
1117	*/
1118	#define REQ_CHECK_SIZES_EX(Name, cbInExpect, cbOutExpect) \
1119	do { \
1120	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect) \|\| pReqHdr->cbOut != (cbOutExpect))) \
1121	{ \
1122	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld. cbOut=%ld expected %ld.\n", \
1123	(long)pReq->Hdr.cbIn, (long)(cbInExpect), (long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
1124	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1125	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1126	} \
1127	} while (0)
1128
1129	#define REQ_CHECK_SIZES(Name) REQ_CHECK_SIZES_EX(Name, Name ## _SIZE_IN, Name ## _SIZE_OUT)
1130
1131	#define REQ_CHECK_SIZE_IN(Name, cbInExpect) \
1132	do { \
1133	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect))) \
1134	{ \
1135	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld.\n", \
1136	(long)pReq->Hdr.cbIn, (long)(cbInExpect))); \
1137	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1138	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1139	} \
1140	} while (0)
1141
1142	#define REQ_CHECK_SIZE_OUT(Name, cbOutExpect) \
1143	do { \
1144	if (RT_UNLIKELY(pReqHdr->cbOut != (cbOutExpect))) \
1145	{ \
1146	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbOut=%ld expected %ld.\n", \
1147	(long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
1148	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1149	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1150	} \
1151	} while (0)
1152
1153	#define REQ_CHECK_EXPR(Name, expr) \
1154	do { \
1155	if (RT_UNLIKELY(!(expr))) \
1156	{ \
1157	OSDBGPRINT(( #Name ": %s\n", #expr)); \
1158	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1159	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1160	} \
1161	} while (0)
1162
1163	#define REQ_CHECK_EXPR_FMT(expr, fmt) \
1164	do { \
1165	if (RT_UNLIKELY(!(expr))) \
1166	{ \
1167	OSDBGPRINT( fmt ); \
1168	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VERR_INVALID_PARAMETER); \
1169	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
1170	} \
1171	} while (0)
1172
1173	/*
1174	* The switch.
1175	*/
1176	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
1177	{
1178	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
1179	{
1180	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
1181	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
1182	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
1183	{
1184	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
1185	pReq->Hdr.rc = VERR_INVALID_MAGIC;
1186	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, VERR_INVALID_MAGIC);
1187	return 0;
1188	}
1189
1190	#if 0
1191	/*
1192	* Call out to the OS specific code and let it do permission checks on the
1193	* client process.
1194	*/
1195	if (!supdrvOSValidateClientProcess(pDevExt, pSession))
1196	{
1197	pReq->u.Out.u32Cookie = 0xffffffff;
1198	pReq->u.Out.u32SessionCookie = 0xffffffff;
1199	pReq->u.Out.u32SessionVersion = 0xffffffff;
1200	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1201	pReq->u.Out.pSession = NULL;
1202	pReq->u.Out.cFunctions = 0;
1203	pReq->Hdr.rc = VERR_PERMISSION_DENIED;
1204	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, VERR_PERMISSION_DENIED);
1205	return 0;
1206	}
1207	#endif
1208
1209	/*
1210	* Match the version.
1211	* The current logic is very simple, match the major interface version.
1212	*/
1213	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
1214	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
1215	{
1216	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1217	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
1218	pReq->u.Out.u32Cookie = 0xffffffff;
1219	pReq->u.Out.u32SessionCookie = 0xffffffff;
1220	pReq->u.Out.u32SessionVersion = 0xffffffff;
1221	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1222	pReq->u.Out.pSession = NULL;
1223	pReq->u.Out.cFunctions = 0;
1224	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1225	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1226	return 0;
1227	}
1228
1229	/*
1230	* Fill in return data and be gone.
1231	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
1232	* u32SessionVersion <= u32ReqVersion!
1233	*/
1234	/** @todo Somehow validate the client and negotiate a secure cookie... */
1235	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
1236	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
1237	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
1238	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1239	pReq->u.Out.pSession = pSession;
1240	pReq->u.Out.cFunctions = sizeof(g_aFunctions) / sizeof(g_aFunctions[0]);
1241	pReq->Hdr.rc = VINF_SUCCESS;
1242	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1243	return 0;
1244	}
1245
1246	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_QUERY_FUNCS(0)):
1247	{
1248	/* validate */
1249	PSUPQUERYFUNCS pReq = (PSUPQUERYFUNCS)pReqHdr;
1250	REQ_CHECK_SIZES_EX(SUP_IOCTL_QUERY_FUNCS, SUP_IOCTL_QUERY_FUNCS_SIZE_IN, SUP_IOCTL_QUERY_FUNCS_SIZE_OUT(RT_ELEMENTS(g_aFunctions)));
1251
1252	/* execute */
1253	pReq->u.Out.cFunctions = RT_ELEMENTS(g_aFunctions);
1254	memcpy(&pReq->u.Out.aFunctions[0], g_aFunctions, sizeof(g_aFunctions));
1255	pReq->Hdr.rc = VINF_SUCCESS;
1256	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1257	return 0;
1258	}
1259
1260	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_LOCK):
1261	{
1262	/* validate */
1263	PSUPPAGELOCK pReq = (PSUPPAGELOCK)pReqHdr;
1264	REQ_CHECK_SIZE_IN(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_IN);
1265	REQ_CHECK_SIZE_OUT(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_OUT(pReq->u.In.cPages));
1266	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.cPages > 0);
1267	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.pvR3 >= PAGE_SIZE);
1268
1269	/* execute */
1270	pReq->Hdr.rc = SUPR0LockMem(pSession, pReq->u.In.pvR3, pReq->u.In.cPages, &pReq->u.Out.aPages[0]);
1271	if (RT_FAILURE(pReq->Hdr.rc))
1272	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1273	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1274	return 0;
1275	}
1276
1277	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_UNLOCK):
1278	{
1279	/* validate */
1280	PSUPPAGEUNLOCK pReq = (PSUPPAGEUNLOCK)pReqHdr;
1281	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_UNLOCK);
1282
1283	/* execute */
1284	pReq->Hdr.rc = SUPR0UnlockMem(pSession, pReq->u.In.pvR3);
1285	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1286	return 0;
1287	}
1288
1289	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_ALLOC):
1290	{
1291	/* validate */
1292	PSUPCONTALLOC pReq = (PSUPCONTALLOC)pReqHdr;
1293	REQ_CHECK_SIZES(SUP_IOCTL_CONT_ALLOC);
1294
1295	/* execute */
1296	pReq->Hdr.rc = SUPR0ContAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.HCPhys);
1297	if (RT_FAILURE(pReq->Hdr.rc))
1298	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1299	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1300	return 0;
1301	}
1302
1303	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_FREE):
1304	{
1305	/* validate */
1306	PSUPCONTFREE pReq = (PSUPCONTFREE)pReqHdr;
1307	REQ_CHECK_SIZES(SUP_IOCTL_CONT_FREE);
1308
1309	/* execute */
1310	pReq->Hdr.rc = SUPR0ContFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1311	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1312	return 0;
1313	}
1314
1315	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_OPEN):
1316	{
1317	/* validate */
1318	PSUPLDROPEN pReq = (PSUPLDROPEN)pReqHdr;
1319	REQ_CHECK_SIZES(SUP_IOCTL_LDR_OPEN);
1320	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs > 0);
1321	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs < 16*_1M);
1322	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1323	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1324	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits < pReq->u.In.cbImageWithTabs);
1325	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.szName[0]);
1326	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1327	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, !supdrvCheckInvalidChar(pReq->u.In.szName, ";:()[]{}/\\\|&*%#@!~`\"'"));
1328	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szFilename, sizeof(pReq->u.In.szFilename)));
1329
1330	/* execute */
1331	pReq->Hdr.rc = supdrvIOCtl_LdrOpen(pDevExt, pSession, pReq);
1332	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1333	return 0;
1334	}
1335
1336	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOAD):
1337	{
1338	/* validate */
1339	PSUPLDRLOAD pReq = (PSUPLDRLOAD)pReqHdr;
1340	REQ_CHECK_EXPR(Name, pReq->Hdr.cbIn >= sizeof(*pReq));
1341	REQ_CHECK_SIZES_EX(SUP_IOCTL_LDR_LOAD, SUP_IOCTL_LDR_LOAD_SIZE_IN(pReq->u.In.cbImageWithTabs), SUP_IOCTL_LDR_LOAD_SIZE_OUT);
1342	REQ_CHECK_EXPR(SUP_IOCTL_LDR_LOAD, pReq->u.In.cSymbols <= 16384);
1343	REQ_CHECK_EXPR_FMT( !pReq->u.In.cSymbols
1344	\|\| ( pReq->u.In.offSymbols < pReq->u.In.cbImageWithTabs
1345	&& pReq->u.In.offSymbols + pReq->u.In.cSymbols * sizeof(SUPLDRSYM) <= pReq->u.In.cbImageWithTabs),
1346	("SUP_IOCTL_LDR_LOAD: offSymbols=%#lx cSymbols=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offSymbols,
1347	(long)pReq->u.In.cSymbols, (long)pReq->u.In.cbImageWithTabs));
1348	REQ_CHECK_EXPR_FMT( !pReq->u.In.cbStrTab
1349	\|\| ( pReq->u.In.offStrTab < pReq->u.In.cbImageWithTabs
1350	&& pReq->u.In.offStrTab + pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs
1351	&& pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs),
1352	("SUP_IOCTL_LDR_LOAD: offStrTab=%#lx cbStrTab=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offStrTab,
1353	(long)pReq->u.In.cbStrTab, (long)pReq->u.In.cbImageWithTabs));
1354
1355	if (pReq->u.In.cSymbols)
1356	{
1357	uint32_t i;
1358	PSUPLDRSYM paSyms = (PSUPLDRSYM)&pReq->u.In.abImage[pReq->u.In.offSymbols];
1359	for (i = 0; i < pReq->u.In.cSymbols; i++)
1360	{
1361	REQ_CHECK_EXPR_FMT(paSyms[i].offSymbol < pReq->u.In.cbImageWithTabs,
1362	("SUP_IOCTL_LDR_LOAD: sym #%ld: symb off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offSymbol, (long)pReq->u.In.cbImageWithTabs));
1363	REQ_CHECK_EXPR_FMT(paSyms[i].offName < pReq->u.In.cbStrTab,
1364	("SUP_IOCTL_LDR_LOAD: sym #%ld: name off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1365	REQ_CHECK_EXPR_FMT(RTStrEnd((char const *)&pReq->u.In.abImage[pReq->u.In.offStrTab + paSyms[i].offName],
1366	pReq->u.In.cbStrTab - paSyms[i].offName),
1367	("SUP_IOCTL_LDR_LOAD: sym #%ld: unterminated name! (%#lx / %#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1368	}
1369	}
1370
1371	/* execute */
1372	pReq->Hdr.rc = supdrvIOCtl_LdrLoad(pDevExt, pSession, pReq);
1373	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1374	return 0;
1375	}
1376
1377	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_FREE):
1378	{
1379	/* validate */
1380	PSUPLDRFREE pReq = (PSUPLDRFREE)pReqHdr;
1381	REQ_CHECK_SIZES(SUP_IOCTL_LDR_FREE);
1382
1383	/* execute */
1384	pReq->Hdr.rc = supdrvIOCtl_LdrFree(pDevExt, pSession, pReq);
1385	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1386	return 0;
1387	}
1388
1389	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_GET_SYMBOL):
1390	{
1391	/* validate */
1392	PSUPLDRGETSYMBOL pReq = (PSUPLDRGETSYMBOL)pReqHdr;
1393	REQ_CHECK_SIZES(SUP_IOCTL_LDR_GET_SYMBOL);
1394	REQ_CHECK_EXPR(SUP_IOCTL_LDR_GET_SYMBOL, RTStrEnd(pReq->u.In.szSymbol, sizeof(pReq->u.In.szSymbol)));
1395
1396	/* execute */
1397	pReq->Hdr.rc = supdrvIOCtl_LdrGetSymbol(pDevExt, pSession, pReq);
1398	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1399	return 0;
1400	}
1401
1402	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0(0)):
1403	{
1404	/* validate */
1405	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1406	Log4(("SUP_IOCTL_CALL_VMMR0: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1407	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1408
1409	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_VMMR0_SIZE(0))
1410	{
1411	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(0), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(0));
1412
1413	/* execute */
1414	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1415	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg, pSession);
1416	else
1417	pReq->Hdr.rc = VERR_WRONG_ORDER;
1418	}
1419	else
1420	{
1421	PSUPVMMR0REQHDR pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1422	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR)),
1423	("SUP_IOCTL_CALL_VMMR0: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR))));
1424	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1425	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(pVMMReq->cbReq));
1426
1427	/* execute */
1428	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1429	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1430	else
1431	pReq->Hdr.rc = VERR_WRONG_ORDER;
1432	}
1433
1434	if ( RT_FAILURE(pReq->Hdr.rc)
1435	&& pReq->Hdr.rc != VERR_INTERRUPTED
1436	&& pReq->Hdr.rc != VERR_TIMEOUT)
1437	Log(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1438	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1439	else
1440	Log4(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1441	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1442	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1443	return 0;
1444	}
1445
1446	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0_BIG):
1447	{
1448	/* validate */
1449	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1450	PSUPVMMR0REQHDR pVMMReq;
1451	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1452	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1453
1454	pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1455	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR)),
1456	("SUP_IOCTL_CALL_VMMR0_BIG: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR))));
1457	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0_BIG, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1458	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));
1459
1460	/* execute */
1461	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1462	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1463	else
1464	pReq->Hdr.rc = VERR_WRONG_ORDER;
1465
1466	if ( RT_FAILURE(pReq->Hdr.rc)
1467	&& pReq->Hdr.rc != VERR_INTERRUPTED
1468	&& pReq->Hdr.rc != VERR_TIMEOUT)
1469	Log(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1470	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1471	else
1472	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1473	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1474	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1475	return 0;
1476	}
1477
1478	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_PAGING_MODE):
1479	{
1480	/* validate */
1481	PSUPGETPAGINGMODE pReq = (PSUPGETPAGINGMODE)pReqHdr;
1482	REQ_CHECK_SIZES(SUP_IOCTL_GET_PAGING_MODE);
1483
1484	/* execute */
1485	pReq->Hdr.rc = VINF_SUCCESS;
1486	pReq->u.Out.enmMode = SUPR0GetPagingMode();
1487	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1488	return 0;
1489	}
1490
1491	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_ALLOC):
1492	{
1493	/* validate */
1494	PSUPLOWALLOC pReq = (PSUPLOWALLOC)pReqHdr;
1495	REQ_CHECK_EXPR(SUP_IOCTL_LOW_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_LOW_ALLOC_SIZE_IN);
1496	REQ_CHECK_SIZES_EX(SUP_IOCTL_LOW_ALLOC, SUP_IOCTL_LOW_ALLOC_SIZE_IN, SUP_IOCTL_LOW_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1497
1498	/* execute */
1499	pReq->Hdr.rc = SUPR0LowAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1500	if (RT_FAILURE(pReq->Hdr.rc))
1501	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1502	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1503	return 0;
1504	}
1505
1506	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_FREE):
1507	{
1508	/* validate */
1509	PSUPLOWFREE pReq = (PSUPLOWFREE)pReqHdr;
1510	REQ_CHECK_SIZES(SUP_IOCTL_LOW_FREE);
1511
1512	/* execute */
1513	pReq->Hdr.rc = SUPR0LowFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1514	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1515	return 0;
1516	}
1517
1518	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_MAP):
1519	{
1520	/* validate */
1521	PSUPGIPMAP pReq = (PSUPGIPMAP)pReqHdr;
1522	REQ_CHECK_SIZES(SUP_IOCTL_GIP_MAP);
1523
1524	/* execute */
1525	pReq->Hdr.rc = SUPR0GipMap(pSession, &pReq->u.Out.pGipR3, &pReq->u.Out.HCPhysGip);
1526	if (RT_SUCCESS(pReq->Hdr.rc))
1527	pReq->u.Out.pGipR0 = pDevExt->pGip;
1528	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1529	return 0;
1530	}
1531
1532	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_UNMAP):
1533	{
1534	/* validate */
1535	PSUPGIPUNMAP pReq = (PSUPGIPUNMAP)pReqHdr;
1536	REQ_CHECK_SIZES(SUP_IOCTL_GIP_UNMAP);
1537
1538	/* execute */
1539	pReq->Hdr.rc = SUPR0GipUnmap(pSession);
1540	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1541	return 0;
1542	}
1543
1544	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SET_VM_FOR_FAST):
1545	{
1546	/* validate */
1547	PSUPSETVMFORFAST pReq = (PSUPSETVMFORFAST)pReqHdr;
1548	REQ_CHECK_SIZES(SUP_IOCTL_SET_VM_FOR_FAST);
1549	REQ_CHECK_EXPR_FMT( !pReq->u.In.pVMR0
1550	\|\| ( VALID_PTR(pReq->u.In.pVMR0)
1551	&& !((uintptr_t)pReq->u.In.pVMR0 & (PAGE_SIZE - 1))),
1552	("SUP_IOCTL_SET_VM_FOR_FAST: pVMR0=%p!\n", pReq->u.In.pVMR0));
1553	/* execute */
1554	pSession->pVM = pReq->u.In.pVMR0;
1555	pReq->Hdr.rc = VINF_SUCCESS;
1556	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1557	return 0;
1558	}
1559
1560	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_ALLOC_EX):
1561	{
1562	/* validate */
1563	PSUPPAGEALLOCEX pReq = (PSUPPAGEALLOCEX)pReqHdr;
1564	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC_EX, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_EX_SIZE_IN);
1565	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));
1566	REQ_CHECK_EXPR_FMT(pReq->u.In.fKernelMapping \|\| pReq->u.In.fUserMapping,
1567	("SUP_IOCTL_PAGE_ALLOC_EX: No mapping requested!\n"));
1568	REQ_CHECK_EXPR_FMT(pReq->u.In.fUserMapping,
1569	("SUP_IOCTL_PAGE_ALLOC_EX: Must have user mapping!\n"));
1570	REQ_CHECK_EXPR_FMT(!pReq->u.In.fReserved0 && !pReq->u.In.fReserved1,
1571	("SUP_IOCTL_PAGE_ALLOC_EX: fReserved0=%d fReserved1=%d\n", pReq->u.In.fReserved0, pReq->u.In.fReserved1));
1572
1573	/* execute */
1574	pReq->Hdr.rc = SUPR0PageAllocEx(pSession, pReq->u.In.cPages, 0 /* fFlags */,
1575	pReq->u.In.fUserMapping ? &pReq->u.Out.pvR3 : NULL,
1576	pReq->u.In.fKernelMapping ? &pReq->u.Out.pvR0 : NULL,
1577	&pReq->u.Out.aPages[0]);
1578	if (RT_FAILURE(pReq->Hdr.rc))
1579	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1580	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1581	return 0;
1582	}
1583
1584	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_MAP_KERNEL):
1585	{
1586	/* validate */
1587	PSUPPAGEMAPKERNEL pReq = (PSUPPAGEMAPKERNEL)pReqHdr;
1588	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_MAP_KERNEL);
1589	REQ_CHECK_EXPR_FMT(!pReq->u.In.fFlags, ("SUP_IOCTL_PAGE_MAP_KERNEL: fFlags=%#x! MBZ\n", pReq->u.In.fFlags));
1590	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_MAP_KERNEL: offSub=%#x\n", pReq->u.In.offSub));
1591	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1592	("SUP_IOCTL_PAGE_MAP_KERNEL: cbSub=%#x\n", pReq->u.In.cbSub));
1593
1594	/* execute */
1595	pReq->Hdr.rc = SUPR0PageMapKernel(pSession, pReq->u.In.pvR3, pReq->u.In.offSub, pReq->u.In.cbSub,
1596	pReq->u.In.fFlags, &pReq->u.Out.pvR0);
1597	if (RT_FAILURE(pReq->Hdr.rc))
1598	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1599	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1600	return 0;
1601	}
1602
1603	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_PROTECT):
1604	{
1605	/* validate */
1606	PSUPPAGEPROTECT pReq = (PSUPPAGEPROTECT)pReqHdr;
1607	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_PROTECT);
1608	REQ_CHECK_EXPR_FMT(!(pReq->u.In.fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)),
1609	("SUP_IOCTL_PAGE_PROTECT: fProt=%#x!\n", pReq->u.In.fProt));
1610	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_PROTECT: offSub=%#x\n", pReq->u.In.offSub));
1611	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1612	("SUP_IOCTL_PAGE_PROTECT: cbSub=%#x\n", pReq->u.In.cbSub));
1613
1614	/* execute */
1615	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);
1616	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1617	return 0;
1618	}
1619
1620	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_FREE):
1621	{
1622	/* validate */
1623	PSUPPAGEFREE pReq = (PSUPPAGEFREE)pReqHdr;
1624	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_FREE);
1625
1626	/* execute */
1627	pReq->Hdr.rc = SUPR0PageFree(pSession, pReq->u.In.pvR3);
1628	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1629	return 0;
1630	}
1631
1632	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_SERVICE(0)):
1633	{
1634	/* validate */
1635	PSUPCALLSERVICE pReq = (PSUPCALLSERVICE)pReqHdr;
1636	Log4(("SUP_IOCTL_CALL_SERVICE: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1637	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1638
1639	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
1640	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(0), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(0));
1641	else
1642	{
1643	PSUPR0SERVICEREQHDR pSrvReq = (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0];
1644	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR)),
1645	("SUP_IOCTL_CALL_SERVICE: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR))));
1646	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, pSrvReq->u32Magic == SUPR0SERVICEREQHDR_MAGIC);
1647	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(pSrvReq->cbReq), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(pSrvReq->cbReq));
1648	}
1649	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1650
1651	/* execute */
1652	pReq->Hdr.rc = supdrvIOCtl_CallServiceModule(pDevExt, pSession, pReq);
1653	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1654	return 0;
1655	}
1656
1657	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOGGER_SETTINGS(0)):
1658	{
1659	/* validate */
1660	PSUPLOGGERSETTINGS pReq = (PSUPLOGGERSETTINGS)pReqHdr;
1661	size_t cbStrTab;
1662	REQ_CHECK_SIZE_OUT(SUP_IOCTL_LOGGER_SETTINGS, SUP_IOCTL_LOGGER_SETTINGS_SIZE_OUT);
1663	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->Hdr.cbIn >= SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(1));
1664	cbStrTab = pReq->Hdr.cbIn - SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(0);
1665	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offGroups < cbStrTab);
1666	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offFlags < cbStrTab);
1667	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offDestination < cbStrTab);
1668	REQ_CHECK_EXPR_FMT(pReq->u.In.szStrings[cbStrTab - 1] == '\0',
1669	("SUP_IOCTL_LOGGER_SETTINGS: cbIn=%#x cbStrTab=%#zx LastChar=%d\n",
1670	pReq->Hdr.cbIn, cbStrTab, pReq->u.In.szStrings[cbStrTab - 1]));
1671	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhich <= SUPLOGGERSETTINGS_WHICH_RELEASE);
1672	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhat <= SUPLOGGERSETTINGS_WHAT_DESTROY);
1673
1674	/* execute */
1675	pReq->Hdr.rc = supdrvIOCtl_LoggerSettings(pDevExt, pSession, pReq);
1676	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1677	return 0;
1678	}
1679
1680	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP2):
1681	{
1682	/* validate */
1683	PSUPSEMOP2 pReq = (PSUPSEMOP2)pReqHdr;
1684	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP2, SUP_IOCTL_SEM_OP2_SIZE_IN, SUP_IOCTL_SEM_OP2_SIZE_OUT);
1685	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP2, pReq->u.In.uReserved == 0);
1686
1687	/* execute */
1688	switch (pReq->u.In.uType)
1689	{
1690	case SUP_SEM_TYPE_EVENT:
1691	{
1692	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1693	switch (pReq->u.In.uOp)
1694	{
1695	case SUPSEMOP2_WAIT_MS_REL:
1696	pReq->Hdr.rc = SUPSemEventWaitNoResume(pSession, hEvent, pReq->u.In.uArg.cRelMsTimeout);
1697	break;
1698	case SUPSEMOP2_WAIT_NS_ABS:
1699	pReq->Hdr.rc = SUPSemEventWaitNsAbsIntr(pSession, hEvent, pReq->u.In.uArg.uAbsNsTimeout);
1700	break;
1701	case SUPSEMOP2_WAIT_NS_REL:
1702	pReq->Hdr.rc = SUPSemEventWaitNsRelIntr(pSession, hEvent, pReq->u.In.uArg.cRelNsTimeout);
1703	break;
1704	case SUPSEMOP2_SIGNAL:
1705	pReq->Hdr.rc = SUPSemEventSignal(pSession, hEvent);
1706	break;
1707	case SUPSEMOP2_CLOSE:
1708	pReq->Hdr.rc = SUPSemEventClose(pSession, hEvent);
1709	break;
1710	case SUPSEMOP2_RESET:
1711	default:
1712	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1713	break;
1714	}
1715	break;
1716	}
1717
1718	case SUP_SEM_TYPE_EVENT_MULTI:
1719	{
1720	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1721	switch (pReq->u.In.uOp)
1722	{
1723	case SUPSEMOP2_WAIT_MS_REL:
1724	pReq->Hdr.rc = SUPSemEventMultiWaitNoResume(pSession, hEventMulti, pReq->u.In.uArg.cRelMsTimeout);
1725	break;
1726	case SUPSEMOP2_WAIT_NS_ABS:
1727	pReq->Hdr.rc = SUPSemEventMultiWaitNsAbsIntr(pSession, hEventMulti, pReq->u.In.uArg.uAbsNsTimeout);
1728	break;
1729	case SUPSEMOP2_WAIT_NS_REL:
1730	pReq->Hdr.rc = SUPSemEventMultiWaitNsRelIntr(pSession, hEventMulti, pReq->u.In.uArg.cRelNsTimeout);
1731	break;
1732	case SUPSEMOP2_SIGNAL:
1733	pReq->Hdr.rc = SUPSemEventMultiSignal(pSession, hEventMulti);
1734	break;
1735	case SUPSEMOP2_CLOSE:
1736	pReq->Hdr.rc = SUPSemEventMultiClose(pSession, hEventMulti);
1737	break;
1738	case SUPSEMOP2_RESET:
1739	pReq->Hdr.rc = SUPSemEventMultiReset(pSession, hEventMulti);
1740	break;
1741	default:
1742	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1743	break;
1744	}
1745	break;
1746	}
1747
1748	default:
1749	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1750	break;
1751	}
1752	return 0;
1753	}
1754
1755	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP3):
1756	{
1757	/* validate */
1758	PSUPSEMOP3 pReq = (PSUPSEMOP3)pReqHdr;
1759	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP3, SUP_IOCTL_SEM_OP3_SIZE_IN, SUP_IOCTL_SEM_OP3_SIZE_OUT);
1760	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, pReq->u.In.u32Reserved == 0 && pReq->u.In.u64Reserved == 0);
1761
1762	/* execute */
1763	switch (pReq->u.In.uType)
1764	{
1765	case SUP_SEM_TYPE_EVENT:
1766	{
1767	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1768	switch (pReq->u.In.uOp)
1769	{
1770	case SUPSEMOP3_CREATE:
1771	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1772	pReq->Hdr.rc = SUPSemEventCreate(pSession, &hEvent);
1773	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEvent;
1774	break;
1775	case SUPSEMOP3_GET_RESOLUTION:
1776	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1777	pReq->Hdr.rc = VINF_SUCCESS;
1778	pReq->Hdr.cbOut = sizeof(*pReq);
1779	pReq->u.Out.cNsResolution = SUPSemEventGetResolution(pSession);
1780	break;
1781	default:
1782	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1783	break;
1784	}
1785	break;
1786	}
1787
1788	case SUP_SEM_TYPE_EVENT_MULTI:
1789	{
1790	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1791	switch (pReq->u.In.uOp)
1792	{
1793	case SUPSEMOP3_CREATE:
1794	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1795	pReq->Hdr.rc = SUPSemEventMultiCreate(pSession, &hEventMulti);
1796	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEventMulti;
1797	break;
1798	case SUPSEMOP3_GET_RESOLUTION:
1799	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1800	pReq->Hdr.rc = VINF_SUCCESS;
1801	pReq->u.Out.cNsResolution = SUPSemEventMultiGetResolution(pSession);
1802	break;
1803	default:
1804	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1805	break;
1806	}
1807	break;
1808	}
1809
1810	default:
1811	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1812	break;
1813	}
1814	return 0;
1815	}
1816
1817	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
1818	{
1819	/* validate */
1820	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
1821	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
1822	REQ_CHECK_EXPR(SUP_IOCTL_VT_CAPS, pReq->Hdr.cbIn <= SUP_IOCTL_VT_CAPS_SIZE_IN);
1823
1824	/* execute */
1825	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.Caps);
1826	if (RT_FAILURE(pReq->Hdr.rc))
1827	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1828	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VINF_SUCCESS, pReq->Hdr.rc);
1829	return 0;
1830	}
1831
1832	default:
1833	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1834	break;
1835	}
1836	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_GENERAL_FAILURE, VERR_GENERAL_FAILURE);
1837	return VERR_GENERAL_FAILURE;
1838	}
1839
1840
1841	/**
1842	* Inter-Driver Communication (IDC) worker.
1843	*
1844	* @returns VBox status code.
1845	* @retval VINF_SUCCESS on success.
1846	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1847	* @retval VERR_NOT_SUPPORTED if the request isn't supported.
1848	*
1849	* @param uReq The request (function) code.
1850	* @param pDevExt Device extention.
1851	* @param pSession Session data.
1852	* @param pReqHdr The request header.
1853	*/
1854	int VBOXCALL supdrvIDC(uintptr_t uReq, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQHDR pReqHdr)
1855	{
1856	/*
1857	* The OS specific code has already validated the pSession
1858	* pointer, and the request size being greater or equal to
1859	* size of the header.
1860	*
1861	* So, just check that pSession is a kernel context session.
1862	*/
1863	if (RT_UNLIKELY( pSession
1864	&& pSession->R0Process != NIL_RTR0PROCESS))
1865	return VERR_INVALID_PARAMETER;
1866
1867	/*
1868	* Validation macro.
1869	*/
1870	#define REQ_CHECK_IDC_SIZE(Name, cbExpect) \
1871	do { \
1872	if (RT_UNLIKELY(pReqHdr->cb != (cbExpect))) \
1873	{ \
1874	OSDBGPRINT(( #Name ": Invalid input/output sizes. cb=%ld expected %ld.\n", \
1875	(long)pReqHdr->cb, (long)(cbExpect))); \
1876	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1877	} \
1878	} while (0)
1879
1880	switch (uReq)
1881	{
1882	case SUPDRV_IDC_REQ_CONNECT:
1883	{
1884	PSUPDRVIDCREQCONNECT pReq = (PSUPDRVIDCREQCONNECT)pReqHdr;
1885	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_CONNECT, sizeof(*pReq));
1886
1887	/*
1888	* Validate the cookie and other input.
1889	*/
1890	if (pReq->Hdr.pSession != NULL)
1891	{
1892	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Hdr.pSession=%p expected NULL!\n", pReq->Hdr.pSession));
1893	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1894	}
1895	if (pReq->u.In.u32MagicCookie != SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE)
1896	{
1897	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: u32MagicCookie=%#x expected %#x!\n",
1898	(unsigned)pReq->u.In.u32MagicCookie, (unsigned)SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE));
1899	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1900	}
1901	if ( pReq->u.In.uMinVersion > pReq->u.In.uReqVersion
1902	\|\| (pReq->u.In.uMinVersion & UINT32_C(0xffff0000)) != (pReq->u.In.uReqVersion & UINT32_C(0xffff0000)))
1903	{
1904	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: uMinVersion=%#x uMaxVersion=%#x doesn't match!\n",
1905	pReq->u.In.uMinVersion, pReq->u.In.uReqVersion));
1906	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1907	}
1908	if (pSession != NULL)
1909	{
1910	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: pSession=%p expected NULL!\n", pSession));
1911	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1912	}
1913
1914	/*
1915	* Match the version.
1916	* The current logic is very simple, match the major interface version.
1917	*/
1918	if ( pReq->u.In.uMinVersion > SUPDRV_IDC_VERSION
1919	\|\| (pReq->u.In.uMinVersion & 0xffff0000) != (SUPDRV_IDC_VERSION & 0xffff0000))
1920	{
1921	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1922	pReq->u.In.uReqVersion, pReq->u.In.uMinVersion, (unsigned)SUPDRV_IDC_VERSION));
1923	pReq->u.Out.pSession = NULL;
1924	pReq->u.Out.uSessionVersion = 0xffffffff;
1925	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1926	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1927	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1928	return VINF_SUCCESS;
1929	}
1930
1931	pReq->u.Out.pSession = NULL;
1932	pReq->u.Out.uSessionVersion = SUPDRV_IDC_VERSION;
1933	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1934	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1935
1936	pReq->Hdr.rc = supdrvCreateSession(pDevExt, false /* fUser */, &pSession);
1937	if (RT_FAILURE(pReq->Hdr.rc))
1938	{
1939	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: failed to create session, rc=%d\n", pReq->Hdr.rc));
1940	return VINF_SUCCESS;
1941	}
1942
1943	pReq->u.Out.pSession = pSession;
1944	pReq->Hdr.pSession = pSession;
1945
1946	return VINF_SUCCESS;
1947	}
1948
1949	case SUPDRV_IDC_REQ_DISCONNECT:
1950	{
1951	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_DISCONNECT, sizeof(*pReqHdr));
1952
1953	supdrvCloseSession(pDevExt, pSession);
1954	return pReqHdr->rc = VINF_SUCCESS;
1955	}
1956
1957	case SUPDRV_IDC_REQ_GET_SYMBOL:
1958	{
1959	PSUPDRVIDCREQGETSYM pReq = (PSUPDRVIDCREQGETSYM)pReqHdr;
1960	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_GET_SYMBOL, sizeof(*pReq));
1961
1962	pReq->Hdr.rc = supdrvIDC_LdrGetSymbol(pDevExt, pSession, pReq);
1963	return VINF_SUCCESS;
1964	}
1965
1966	case SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY:
1967	{
1968	PSUPDRVIDCREQCOMPREGFACTORY pReq = (PSUPDRVIDCREQCOMPREGFACTORY)pReqHdr;
1969	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY, sizeof(*pReq));
1970
1971	pReq->Hdr.rc = SUPR0ComponentRegisterFactory(pSession, pReq->u.In.pFactory);
1972	return VINF_SUCCESS;
1973	}
1974
1975	case SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY:
1976	{
1977	PSUPDRVIDCREQCOMPDEREGFACTORY pReq = (PSUPDRVIDCREQCOMPDEREGFACTORY)pReqHdr;
1978	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY, sizeof(*pReq));
1979
1980	pReq->Hdr.rc = SUPR0ComponentDeregisterFactory(pSession, pReq->u.In.pFactory);
1981	return VINF_SUCCESS;
1982	}
1983
1984	default:
1985	Log(("Unknown IDC %#lx\n", (long)uReq));
1986	break;
1987	}
1988
1989	#undef REQ_CHECK_IDC_SIZE
1990	return VERR_NOT_SUPPORTED;
1991	}
1992
1993
1994	/**
1995	* Register a object for reference counting.
1996	* The object is registered with one reference in the specified session.
1997	*
1998	* @returns Unique identifier on success (pointer).
1999	* All future reference must use this identifier.
2000	* @returns NULL on failure.
2001	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
2002	* @param pvUser1 The first user argument.
2003	* @param pvUser2 The second user argument.
2004	*/
2005	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
2006	{
2007	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2008	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2009	PSUPDRVOBJ pObj;
2010	PSUPDRVUSAGE pUsage;
2011
2012	/*
2013	* Validate the input.
2014	*/
2015	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
2016	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
2017	AssertPtrReturn(pfnDestructor, NULL);
2018
2019	/*
2020	* Allocate and initialize the object.
2021	*/
2022	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
2023	if (!pObj)
2024	return NULL;
2025	pObj->u32Magic = SUPDRVOBJ_MAGIC;
2026	pObj->enmType = enmType;
2027	pObj->pNext = NULL;
2028	pObj->cUsage = 1;
2029	pObj->pfnDestructor = pfnDestructor;
2030	pObj->pvUser1 = pvUser1;
2031	pObj->pvUser2 = pvUser2;
2032	pObj->CreatorUid = pSession->Uid;
2033	pObj->CreatorGid = pSession->Gid;
2034	pObj->CreatorProcess= pSession->Process;
2035	supdrvOSObjInitCreator(pObj, pSession);
2036
2037	/*
2038	* Allocate the usage record.
2039	* (We keep freed usage records around to simplify SUPR0ObjAddRefEx().)
2040	*/
2041	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2042
2043	pUsage = pDevExt->pUsageFree;
2044	if (pUsage)
2045	pDevExt->pUsageFree = pUsage->pNext;
2046	else
2047	{
2048	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2049	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
2050	if (!pUsage)
2051	{
2052	RTMemFree(pObj);
2053	return NULL;
2054	}
2055	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2056	}
2057
2058	/*
2059	* Insert the object and create the session usage record.
2060	*/
2061	/* The object. */
2062	pObj->pNext = pDevExt->pObjs;
2063	pDevExt->pObjs = pObj;
2064
2065	/* The session record. */
2066	pUsage->cUsage = 1;
2067	pUsage->pObj = pObj;
2068	pUsage->pNext = pSession->pUsage;
2069	/* Log2(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext)); */
2070	pSession->pUsage = pUsage;
2071
2072	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2073
2074	Log(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
2075	return pObj;
2076	}
2077
2078
2079	/**
2080	* Increment the reference counter for the object associating the reference
2081	* with the specified session.
2082	*
2083	* @returns IPRT status code.
2084	* @param pvObj The identifier returned by SUPR0ObjRegister().
2085	* @param pSession The session which is referencing the object.
2086	*
2087	* @remarks The caller should not own any spinlocks and must carefully protect
2088	* itself against potential race with the destructor so freed memory
2089	* isn't accessed here.
2090	*/
2091	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
2092	{
2093	return SUPR0ObjAddRefEx(pvObj, pSession, false /* fNoBlocking */);
2094	}
2095
2096
2097	/**
2098	* Increment the reference counter for the object associating the reference
2099	* with the specified session.
2100	*
2101	* @returns IPRT status code.
2102	* @retval VERR_TRY_AGAIN if fNoBlocking was set and a new usage record
2103	* couldn't be allocated. (If you see this you're not doing the right
2104	* thing and it won't ever work reliably.)
2105	*
2106	* @param pvObj The identifier returned by SUPR0ObjRegister().
2107	* @param pSession The session which is referencing the object.
2108	* @param fNoBlocking Set if it's not OK to block. Never try to make the
2109	* first reference to an object in a session with this
2110	* argument set.
2111	*
2112	* @remarks The caller should not own any spinlocks and must carefully protect
2113	* itself against potential race with the destructor so freed memory
2114	* isn't accessed here.
2115	*/
2116	SUPR0DECL(int) SUPR0ObjAddRefEx(void *pvObj, PSUPDRVSESSION pSession, bool fNoBlocking)
2117	{
2118	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2119	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2120	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2121	int rc = VINF_SUCCESS;
2122	PSUPDRVUSAGE pUsagePre;
2123	PSUPDRVUSAGE pUsage;
2124
2125	/*
2126	* Validate the input.
2127	* Be ready for the destruction race (someone might be stuck in the
2128	* destructor waiting a lock we own).
2129	*/
2130	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2131	AssertPtrReturn(pObj, VERR_INVALID_POINTER);
2132	AssertMsgReturn(pObj->u32Magic == SUPDRVOBJ_MAGIC \|\| pObj->u32Magic == SUPDRVOBJ_MAGIC_DEAD,
2133	("Invalid pvObj=%p magic=%#x (expected %#x or %#x)\n", pvObj, pObj->u32Magic, SUPDRVOBJ_MAGIC, SUPDRVOBJ_MAGIC_DEAD),
2134	VERR_INVALID_PARAMETER);
2135
2136	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2137
2138	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2139	{
2140	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2141
2142	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2143	return VERR_WRONG_ORDER;
2144	}
2145
2146	/*
2147	* Preallocate the usage record if we can.
2148	*/
2149	pUsagePre = pDevExt->pUsageFree;
2150	if (pUsagePre)
2151	pDevExt->pUsageFree = pUsagePre->pNext;
2152	else if (!fNoBlocking)
2153	{
2154	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2155	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
2156	if (!pUsagePre)
2157	return VERR_NO_MEMORY;
2158
2159	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2160	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2161	{
2162	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2163
2164	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2165	return VERR_WRONG_ORDER;
2166	}
2167	}
2168
2169	/*
2170	* Reference the object.
2171	*/
2172	pObj->cUsage++;
2173
2174	/*
2175	* Look for the session record.
2176	*/
2177	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
2178	{
2179	/Log(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2180	if (pUsage->pObj == pObj)
2181	break;
2182	}
2183	if (pUsage)
2184	pUsage->cUsage++;
2185	else if (pUsagePre)
2186	{
2187	/* create a new session record. */
2188	pUsagePre->cUsage = 1;
2189	pUsagePre->pObj = pObj;
2190	pUsagePre->pNext = pSession->pUsage;
2191	pSession->pUsage = pUsagePre;
2192	/Log(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));/
2193
2194	pUsagePre = NULL;
2195	}
2196	else
2197	{
2198	pObj->cUsage--;
2199	rc = VERR_TRY_AGAIN;
2200	}
2201
2202	/*
2203	* Put any unused usage record into the free list..
2204	*/
2205	if (pUsagePre)
2206	{
2207	pUsagePre->pNext = pDevExt->pUsageFree;
2208	pDevExt->pUsageFree = pUsagePre;
2209	}
2210
2211	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2212
2213	return rc;
2214	}
2215
2216
2217	/**
2218	* Decrement / destroy a reference counter record for an object.
2219	*
2220	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
2221	*
2222	* @returns IPRT status code.
2223	* @retval VINF_SUCCESS if not destroyed.
2224	* @retval VINF_OBJECT_DESTROYED if it's destroyed by this release call.
2225	* @retval VERR_INVALID_PARAMETER if the object isn't valid. Will assert in
2226	* string builds.
2227	*
2228	* @param pvObj The identifier returned by SUPR0ObjRegister().
2229	* @param pSession The session which is referencing the object.
2230	*/
2231	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
2232	{
2233	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2234	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2235	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2236	int rc = VERR_INVALID_PARAMETER;
2237	PSUPDRVUSAGE pUsage;
2238	PSUPDRVUSAGE pUsagePrev;
2239
2240	/*
2241	* Validate the input.
2242	*/
2243	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2244	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2245	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2246	VERR_INVALID_PARAMETER);
2247
2248	/*
2249	* Acquire the spinlock and look for the usage record.
2250	*/
2251	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
2252
2253	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
2254	pUsage;
2255	pUsagePrev = pUsage, pUsage = pUsage->pNext)
2256	{
2257	/Log2(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2258	if (pUsage->pObj == pObj)
2259	{
2260	rc = VINF_SUCCESS;
2261	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
2262	if (pUsage->cUsage > 1)
2263	{
2264	pObj->cUsage--;
2265	pUsage->cUsage--;
2266	}
2267	else
2268	{
2269	/*
2270	* Free the session record.
2271	*/
2272	if (pUsagePrev)
2273	pUsagePrev->pNext = pUsage->pNext;
2274	else
2275	pSession->pUsage = pUsage->pNext;
2276	pUsage->pNext = pDevExt->pUsageFree;
2277	pDevExt->pUsageFree = pUsage;
2278
2279	/* What about the object? */
2280	if (pObj->cUsage > 1)
2281	pObj->cUsage--;
2282	else
2283	{
2284	/*
2285	* Object is to be destroyed, unlink it.
2286	*/
2287	pObj->u32Magic = SUPDRVOBJ_MAGIC_DEAD;
2288	rc = VINF_OBJECT_DESTROYED;
2289	if (pDevExt->pObjs == pObj)
2290	pDevExt->pObjs = pObj->pNext;
2291	else
2292	{
2293	PSUPDRVOBJ pObjPrev;
2294	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
2295	if (pObjPrev->pNext == pObj)
2296	{
2297	pObjPrev->pNext = pObj->pNext;
2298	break;
2299	}
2300	Assert(pObjPrev);
2301	}
2302	}
2303	}
2304	break;
2305	}
2306	}
2307
2308	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
2309
2310	/*
2311	* Call the destructor and free the object if required.
2312	*/
2313	if (rc == VINF_OBJECT_DESTROYED)
2314	{
2315	Log(("SUPR0ObjRelease: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
2316	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
2317	if (pObj->pfnDestructor)
2318	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
2319	RTMemFree(pObj);
2320	}
2321
2322	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
2323	return rc;
2324	}
2325
2326
2327	/**
2328	* Verifies that the current process can access the specified object.
2329	*
2330	* @returns The following IPRT status code:
2331	* @retval VINF_SUCCESS if access was granted.
2332	* @retval VERR_PERMISSION_DENIED if denied access.
2333	* @retval VERR_INVALID_PARAMETER if invalid parameter.
2334	*
2335	* @param pvObj The identifier returned by SUPR0ObjRegister().
2336	* @param pSession The session which wishes to access the object.
2337	* @param pszObjName Object string name. This is optional and depends on the object type.
2338	*
2339	* @remark The caller is responsible for making sure the object isn't removed while
2340	* we're inside this function. If uncertain about this, just call AddRef before calling us.
2341	*/
2342	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
2343	{
2344	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2345	int rc;
2346
2347	/*
2348	* Validate the input.
2349	*/
2350	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2351	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2352	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2353	VERR_INVALID_PARAMETER);
2354
2355	/*
2356	* Check access. (returns true if a decision has been made.)
2357	*/
2358	rc = VERR_INTERNAL_ERROR;
2359	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
2360	return rc;
2361
2362	/*
2363	* Default policy is to allow the user to access his own
2364	* stuff but nothing else.
2365	*/
2366	if (pObj->CreatorUid == pSession->Uid)
2367	return VINF_SUCCESS;
2368	return VERR_PERMISSION_DENIED;
2369	}
2370
2371
2372	/**
2373	* Lock pages.
2374	*
2375	* @returns IPRT status code.
2376	* @param pSession Session to which the locked memory should be associated.
2377	* @param pvR3 Start of the memory range to lock.
2378	* This must be page aligned.
2379	* @param cPages Number of pages to lock.
2380	* @param paPages Where to put the physical addresses of locked memory.
2381	*/
2382	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
2383	{
2384	int rc;
2385	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2386	const size_t cb = (size_t)cPages << PAGE_SHIFT;
2387	LogFlow(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
2388
2389	/*
2390	* Verify input.
2391	*/
2392	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2393	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
2394	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
2395	\|\| !pvR3)
2396	{
2397	Log(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
2398	return VERR_INVALID_PARAMETER;
2399	}
2400
2401	/*
2402	* Let IPRT do the job.
2403	*/
2404	Mem.eType = MEMREF_TYPE_LOCKED;
2405	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTMEM_PROT_READ \| RTMEM_PROT_WRITE, RTR0ProcHandleSelf());
2406	if (RT_SUCCESS(rc))
2407	{
2408	uint32_t iPage = cPages;
2409	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
2410	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
2411
2412	while (iPage-- > 0)
2413	{
2414	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2415	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
2416	{
2417	AssertMsgFailed(("iPage=%d\n", iPage));
2418	rc = VERR_INTERNAL_ERROR;
2419	break;
2420	}
2421	}
2422	if (RT_SUCCESS(rc))
2423	rc = supdrvMemAdd(&Mem, pSession);
2424	if (RT_FAILURE(rc))
2425	{
2426	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
2427	AssertRC(rc2);
2428	}
2429	}
2430
2431	return rc;
2432	}
2433
2434
2435	/**
2436	* Unlocks the memory pointed to by pv.
2437	*
2438	* @returns IPRT status code.
2439	* @param pSession Session to which the memory was locked.
2440	* @param pvR3 Memory to unlock.
2441	*/
2442	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2443	{
2444	LogFlow(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2445	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2446	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
2447	}
2448
2449
2450	/**
2451	* Allocates a chunk of page aligned memory with contiguous and fixed physical
2452	* backing.
2453	*
2454	* @returns IPRT status code.
2455	* @param pSession Session data.
2456	* @param cPages Number of pages to allocate.
2457	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
2458	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
2459	* @param pHCPhys Where to put the physical address of allocated memory.
2460	*/
2461	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
2462	{
2463	int rc;
2464	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2465	LogFlow(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
2466
2467	/*
2468	* Validate input.
2469	*/
2470	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2471	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
2472	{
2473	Log(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
2474	pSession, ppvR0, ppvR3, pHCPhys));
2475	return VERR_INVALID_PARAMETER;
2476
2477	}
2478	if (cPages < 1 \|\| cPages >= 256)
2479	{
2480	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2481	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2482	}
2483
2484	/*
2485	* Let IPRT do the job.
2486	*/
2487	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
2488	if (RT_SUCCESS(rc))
2489	{
2490	int rc2;
2491	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2492	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2493	if (RT_SUCCESS(rc))
2494	{
2495	Mem.eType = MEMREF_TYPE_CONT;
2496	rc = supdrvMemAdd(&Mem, pSession);
2497	if (!rc)
2498	{
2499	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2500	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2501	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
2502	return 0;
2503	}
2504
2505	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2506	AssertRC(rc2);
2507	}
2508	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2509	AssertRC(rc2);
2510	}
2511
2512	return rc;
2513	}
2514
2515
2516	/**
2517	* Frees memory allocated using SUPR0ContAlloc().
2518	*
2519	* @returns IPRT status code.
2520	* @param pSession The session to which the memory was allocated.
2521	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2522	*/
2523	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2524	{
2525	LogFlow(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2526	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2527	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
2528	}
2529
2530
2531	/**
2532	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
2533	*
2534	* The memory isn't zeroed.
2535	*
2536	* @returns IPRT status code.
2537	* @param pSession Session data.
2538	* @param cPages Number of pages to allocate.
2539	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
2540	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
2541	* @param paPages Where to put the physical addresses of allocated memory.
2542	*/
2543	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
2544	{
2545	unsigned iPage;
2546	int rc;
2547	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2548	LogFlow(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
2549
2550	/*
2551	* Validate input.
2552	*/
2553	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2554	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
2555	{
2556	Log(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
2557	pSession, ppvR3, ppvR0, paPages));
2558	return VERR_INVALID_PARAMETER;
2559
2560	}
2561	if (cPages < 1 \|\| cPages >= 256)
2562	{
2563	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2564	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2565	}
2566
2567	/*
2568	* Let IPRT do the work.
2569	*/
2570	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
2571	if (RT_SUCCESS(rc))
2572	{
2573	int rc2;
2574	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2575	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2576	if (RT_SUCCESS(rc))
2577	{
2578	Mem.eType = MEMREF_TYPE_LOW;
2579	rc = supdrvMemAdd(&Mem, pSession);
2580	if (!rc)
2581	{
2582	for (iPage = 0; iPage < cPages; iPage++)
2583	{
2584	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2585	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%RHp\n", paPages[iPage]));
2586	}
2587	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2588	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2589	return 0;
2590	}
2591
2592	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2593	AssertRC(rc2);
2594	}
2595
2596	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2597	AssertRC(rc2);
2598	}
2599
2600	return rc;
2601	}
2602
2603
2604	/**
2605	* Frees memory allocated using SUPR0LowAlloc().
2606	*
2607	* @returns IPRT status code.
2608	* @param pSession The session to which the memory was allocated.
2609	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2610	*/
2611	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2612	{
2613	LogFlow(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2614	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2615	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
2616	}
2617
2618
2619
2620	/**
2621	* Allocates a chunk of memory with both R0 and R3 mappings.
2622	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
2623	*
2624	* @returns IPRT status code.
2625	* @param pSession The session to associated the allocation with.
2626	* @param cb Number of bytes to allocate.
2627	* @param ppvR0 Where to store the address of the Ring-0 mapping.
2628	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2629	*/
2630	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
2631	{
2632	int rc;
2633	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2634	LogFlow(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
2635
2636	/*
2637	* Validate input.
2638	*/
2639	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2640	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
2641	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
2642	if (cb < 1 \|\| cb >= _4M)
2643	{
2644	Log(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
2645	return VERR_INVALID_PARAMETER;
2646	}
2647
2648	/*
2649	* Let IPRT do the work.
2650	*/
2651	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
2652	if (RT_SUCCESS(rc))
2653	{
2654	int rc2;
2655	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2656	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2657	if (RT_SUCCESS(rc))
2658	{
2659	Mem.eType = MEMREF_TYPE_MEM;
2660	rc = supdrvMemAdd(&Mem, pSession);
2661	if (!rc)
2662	{
2663	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2664	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2665	return VINF_SUCCESS;
2666	}
2667
2668	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2669	AssertRC(rc2);
2670	}
2671
2672	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2673	AssertRC(rc2);
2674	}
2675
2676	return rc;
2677	}
2678
2679
2680	/**
2681	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
2682	*
2683	* @returns IPRT status code.
2684	* @param pSession The session to which the memory was allocated.
2685	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2686	* @param paPages Where to store the physical addresses.
2687	*/
2688	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
2689	{
2690	PSUPDRVBUNDLE pBundle;
2691	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2692	LogFlow(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
2693
2694	/*
2695	* Validate input.
2696	*/
2697	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2698	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
2699	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
2700
2701	/*
2702	* Search for the address.
2703	*/
2704	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2705	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2706	{
2707	if (pBundle->cUsed > 0)
2708	{
2709	unsigned i;
2710	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2711	{
2712	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
2713	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2714	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2715	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2716	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
2717	)
2718	)
2719	{
2720	const size_t cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2721	size_t iPage;
2722	for (iPage = 0; iPage < cPages; iPage++)
2723	{
2724	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2725	paPages[iPage].uReserved = 0;
2726	}
2727	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2728	return VINF_SUCCESS;
2729	}
2730	}
2731	}
2732	}
2733	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2734	Log(("Failed to find %p!!!\n", (void *)uPtr));
2735	return VERR_INVALID_PARAMETER;
2736	}
2737
2738
2739	/**
2740	* Free memory allocated by SUPR0MemAlloc().
2741	*
2742	* @returns IPRT status code.
2743	* @param pSession The session owning the allocation.
2744	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2745	*/
2746	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2747	{
2748	LogFlow(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2749	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2750	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
2751	}
2752
2753
2754	/**
2755	* Allocates a chunk of memory with a kernel or/and a user mode mapping.
2756	*
2757	* The memory is fixed and it's possible to query the physical addresses using
2758	* SUPR0MemGetPhys().
2759	*
2760	* @returns IPRT status code.
2761	* @param pSession The session to associated the allocation with.
2762	* @param cPages The number of pages to allocate.
2763	* @param fFlags Flags, reserved for the future. Must be zero.
2764	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2765	* NULL if no ring-3 mapping.
2766	* @param ppvR3 Where to store the address of the Ring-0 mapping.
2767	* NULL if no ring-0 mapping.
2768	* @param paPages Where to store the addresses of the pages. Optional.
2769	*/
2770	SUPR0DECL(int) SUPR0PageAllocEx(PSUPDRVSESSION pSession, uint32_t cPages, uint32_t fFlags, PRTR3PTR ppvR3, PRTR0PTR ppvR0, PRTHCPHYS paPages)
2771	{
2772	int rc;
2773	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2774	LogFlow(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
2775
2776	/*
2777	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2778	*/
2779	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2780	AssertPtrNullReturn(ppvR3, VERR_INVALID_POINTER);
2781	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2782	AssertReturn(ppvR3 \|\| ppvR0, VERR_INVALID_PARAMETER);
2783	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2784	if (cPages < 1 \|\| cPages > VBOX_MAX_ALLOC_PAGE_COUNT)
2785	{
2786	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)));
2787	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2788	}
2789
2790	/*
2791	* Let IPRT do the work.
2792	*/
2793	if (ppvR0)
2794	rc = RTR0MemObjAllocPage(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, true /* fExecutable */);
2795	else
2796	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
2797	if (RT_SUCCESS(rc))
2798	{
2799	int rc2;
2800	if (ppvR3)
2801	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2802	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2803	else
2804	Mem.MapObjR3 = NIL_RTR0MEMOBJ;
2805	if (RT_SUCCESS(rc))
2806	{
2807	Mem.eType = MEMREF_TYPE_PAGE;
2808	rc = supdrvMemAdd(&Mem, pSession);
2809	if (!rc)
2810	{
2811	if (ppvR3)
2812	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2813	if (ppvR0)
2814	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2815	if (paPages)
2816	{
2817	uint32_t iPage = cPages;
2818	while (iPage-- > 0)
2819	{
2820	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
2821	Assert(paPages[iPage] != NIL_RTHCPHYS);
2822	}
2823	}
2824	return VINF_SUCCESS;
2825	}
2826
2827	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2828	AssertRC(rc2);
2829	}
2830
2831	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2832	AssertRC(rc2);
2833	}
2834	return rc;
2835	}
2836
2837
2838	/**
2839	* Maps a chunk of memory previously allocated by SUPR0PageAllocEx into kernel
2840	* space.
2841	*
2842	* @returns IPRT status code.
2843	* @param pSession The session to associated the allocation with.
2844	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
2845	* @param offSub Where to start mapping. Must be page aligned.
2846	* @param cbSub How much to map. Must be page aligned.
2847	* @param fFlags Flags, MBZ.
2848	* @param ppvR0 Where to return the address of the ring-0 mapping on
2849	* success.
2850	*/
2851	SUPR0DECL(int) SUPR0PageMapKernel(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t offSub, uint32_t cbSub,
2852	uint32_t fFlags, PRTR0PTR ppvR0)
2853	{
2854	int rc;
2855	PSUPDRVBUNDLE pBundle;
2856	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2857	RTR0MEMOBJ hMemObj = NIL_RTR0MEMOBJ;
2858	LogFlow(("SUPR0PageMapKernel: pSession=%p pvR3=%p offSub=%#x cbSub=%#x\n", pSession, pvR3, offSub, cbSub));
2859
2860	/*
2861	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2862	*/
2863	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2864	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2865	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2866	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2867	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2868	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
2869
2870	/*
2871	* Find the memory object.
2872	*/
2873	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2874	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2875	{
2876	if (pBundle->cUsed > 0)
2877	{
2878	unsigned i;
2879	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2880	{
2881	if ( ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
2882	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2883	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2884	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
2885	\|\| ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED
2886	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2887	&& pBundle->aMem[i].MapObjR3 == NIL_RTR0MEMOBJ
2888	&& RTR0MemObjAddressR3(pBundle->aMem[i].MemObj) == pvR3))
2889	{
2890	hMemObj = pBundle->aMem[i].MemObj;
2891	break;
2892	}
2893	}
2894	}
2895	}
2896	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2897
2898	rc = VERR_INVALID_PARAMETER;
2899	if (hMemObj != NIL_RTR0MEMOBJ)
2900	{
2901	/*
2902	* Do some further input validations before calling IPRT.
2903	* (Cleanup is done indirectly by telling RTR0MemObjFree to include mappings.)
2904	*/
2905	size_t cbMemObj = RTR0MemObjSize(hMemObj);
2906	if ( offSub < cbMemObj
2907	&& cbSub <= cbMemObj
2908	&& offSub + cbSub <= cbMemObj)
2909	{
2910	RTR0MEMOBJ hMapObj;
2911	rc = RTR0MemObjMapKernelEx(&hMapObj, hMemObj, (void *)-1, 0,
2912	RTMEM_PROT_READ \| RTMEM_PROT_WRITE, offSub, cbSub);
2913	if (RT_SUCCESS(rc))
2914	*ppvR0 = RTR0MemObjAddress(hMapObj);
2915	}
2916	else
2917	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
2918
2919	}
2920	return rc;
2921	}
2922
2923
2924	/**
2925	* Changes the page level protection of one or more pages previously allocated
2926	* by SUPR0PageAllocEx.
2927	*
2928	* @returns IPRT status code.
2929	* @param pSession The session to associated the allocation with.
2930	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
2931	* NIL_RTR3PTR if the ring-3 mapping should be unaffected.
2932	* @param pvR0 The ring-0 address returned by SUPR0PageAllocEx.
2933	* NIL_RTR0PTR if the ring-0 mapping should be unaffected.
2934	* @param offSub Where to start changing. Must be page aligned.
2935	* @param cbSub How much to change. Must be page aligned.
2936	* @param fProt The new page level protection, see RTMEM_PROT_*.
2937	*/
2938	SUPR0DECL(int) SUPR0PageProtect(PSUPDRVSESSION pSession, RTR3PTR pvR3, RTR0PTR pvR0, uint32_t offSub, uint32_t cbSub, uint32_t fProt)
2939	{
2940	int rc;
2941	PSUPDRVBUNDLE pBundle;
2942	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2943	RTR0MEMOBJ hMemObjR0 = NIL_RTR0MEMOBJ;
2944	RTR0MEMOBJ hMemObjR3 = NIL_RTR0MEMOBJ;
2945	LogFlow(("SUPR0PageProtect: pSession=%p pvR3=%p pvR0=%p offSub=%#x cbSub=%#x fProt-%#x\n", pSession, pvR3, pvR0, offSub, cbSub, fProt));
2946
2947	/*
2948	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2949	*/
2950	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2951	AssertReturn(!(fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)), VERR_INVALID_PARAMETER);
2952	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2953	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2954	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
2955
2956	/*
2957	* Find the memory object.
2958	*/
2959	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2960	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2961	{
2962	if (pBundle->cUsed > 0)
2963	{
2964	unsigned i;
2965	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2966	{
2967	if ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
2968	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2969	&& ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2970	\|\| pvR3 == NIL_RTR3PTR)
2971	&& ( pvR0 == NIL_RTR0PTR
2972	\|\| RTR0MemObjAddress(pBundle->aMem[i].MemObj) == pvR0)
2973	&& ( pvR3 == NIL_RTR3PTR
2974	\|\| RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3))
2975	{
2976	if (pvR0 != NIL_RTR0PTR)
2977	hMemObjR0 = pBundle->aMem[i].MemObj;
2978	if (pvR3 != NIL_RTR3PTR)
2979	hMemObjR3 = pBundle->aMem[i].MapObjR3;
2980	break;
2981	}
2982	}
2983	}
2984	}
2985	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2986
2987	rc = VERR_INVALID_PARAMETER;
2988	if ( hMemObjR0 != NIL_RTR0MEMOBJ
2989	\|\| hMemObjR3 != NIL_RTR0MEMOBJ)
2990	{
2991	/*
2992	* Do some further input validations before calling IPRT.
2993	*/
2994	size_t cbMemObj = hMemObjR0 != NIL_RTR0PTR ? RTR0MemObjSize(hMemObjR0) : RTR0MemObjSize(hMemObjR3);
2995	if ( offSub < cbMemObj
2996	&& cbSub <= cbMemObj
2997	&& offSub + cbSub <= cbMemObj)
2998	{
2999	rc = VINF_SUCCESS;
3000	if (hMemObjR3 != NIL_RTR0PTR)
3001	rc = RTR0MemObjProtect(hMemObjR3, offSub, cbSub, fProt);
3002	if (hMemObjR0 != NIL_RTR0PTR && RT_SUCCESS(rc))
3003	rc = RTR0MemObjProtect(hMemObjR0, offSub, cbSub, fProt);
3004	}
3005	else
3006	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
3007
3008	}
3009	return rc;
3010
3011	}
3012
3013
3014	/**
3015	* Free memory allocated by SUPR0PageAlloc() and SUPR0PageAllocEx().
3016	*
3017	* @returns IPRT status code.
3018	* @param pSession The session owning the allocation.
3019	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc() or
3020	* SUPR0PageAllocEx().
3021	*/
3022	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
3023	{
3024	LogFlow(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
3025	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3026	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_PAGE);
3027	}
3028
3029
3030	/**
3031	* Gets the paging mode of the current CPU.
3032	*
3033	* @returns Paging mode, SUPPAGEINGMODE_INVALID on error.
3034	*/
3035	SUPR0DECL(SUPPAGINGMODE) SUPR0GetPagingMode(void)
3036	{
3037	SUPPAGINGMODE enmMode;
3038
3039	RTR0UINTREG cr0 = ASMGetCR0();
3040	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
3041	enmMode = SUPPAGINGMODE_INVALID;
3042	else
3043	{
3044	RTR0UINTREG cr4 = ASMGetCR4();
3045	uint32_t fNXEPlusLMA = 0;
3046	if (cr4 & X86_CR4_PAE)
3047	{
3048	uint32_t fAmdFeatures = ASMCpuId_EDX(0x80000001);
3049	if (fAmdFeatures & (X86_CPUID_AMD_FEATURE_EDX_NX \| X86_CPUID_AMD_FEATURE_EDX_LONG_MODE))
3050	{
3051	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
3052	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
3053	fNXEPlusLMA \|= RT_BIT(0);
3054	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
3055	fNXEPlusLMA \|= RT_BIT(1);
3056	}
3057	}
3058
3059	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
3060	{
3061	case 0:
3062	enmMode = SUPPAGINGMODE_32_BIT;
3063	break;
3064
3065	case X86_CR4_PGE:
3066	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
3067	break;
3068
3069	case X86_CR4_PAE:
3070	enmMode = SUPPAGINGMODE_PAE;
3071	break;
3072
3073	case X86_CR4_PAE \| RT_BIT(0):
3074	enmMode = SUPPAGINGMODE_PAE_NX;
3075	break;
3076
3077	case X86_CR4_PAE \| X86_CR4_PGE:
3078	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3079	break;
3080
3081	case X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3082	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3083	break;
3084
3085	case RT_BIT(1) \| X86_CR4_PAE:
3086	enmMode = SUPPAGINGMODE_AMD64;
3087	break;
3088
3089	case RT_BIT(1) \| X86_CR4_PAE \| RT_BIT(0):
3090	enmMode = SUPPAGINGMODE_AMD64_NX;
3091	break;
3092
3093	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
3094	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
3095	break;
3096
3097	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3098	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
3099	break;
3100
3101	default:
3102	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
3103	enmMode = SUPPAGINGMODE_INVALID;
3104	break;
3105	}
3106	}
3107	return enmMode;
3108	}
3109
3110
3111	/**
3112	* Enables or disabled hardware virtualization extensions using native OS APIs.
3113	*
3114	* @returns VBox status code.
3115	* @retval VINF_SUCCESS on success.
3116	* @retval VERR_NOT_SUPPORTED if not supported by the native OS.
3117	*
3118	* @param fEnable Whether to enable or disable.
3119	*/
3120	SUPR0DECL(int) SUPR0EnableVTx(bool fEnable)
3121	{
3122	#ifdef RT_OS_DARWIN
3123	return supdrvOSEnableVTx(fEnable);
3124	#else
3125	return VERR_NOT_SUPPORTED;
3126	#endif
3127	}
3128
3129
3130	/** @todo document me */
3131	SUPR0DECL(int) SUPR0QueryVTCaps(PSUPDRVSESSION pSession, uint32_t *pfCaps)
3132	{
3133	/*
3134	* Input validation.
3135	*/
3136	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3137	AssertPtrReturn(pfCaps, VERR_INVALID_POINTER);
3138
3139	*pfCaps = 0;
3140
3141	if (ASMHasCpuId())
3142	{
3143	uint32_t u32FeaturesECX;
3144	uint32_t u32Dummy;
3145	uint32_t u32FeaturesEDX;
3146	uint32_t u32VendorEBX, u32VendorECX, u32VendorEDX, u32AMDFeatureEDX, u32AMDFeatureECX;
3147	uint64_t val;
3148
3149	ASMCpuId(0, &u32Dummy, &u32VendorEBX, &u32VendorECX, &u32VendorEDX);
3150	ASMCpuId(1, &u32Dummy, &u32Dummy, &u32FeaturesECX, &u32FeaturesEDX);
3151	/* Query AMD features. */
3152	ASMCpuId(0x80000001, &u32Dummy, &u32Dummy, &u32AMDFeatureECX, &u32AMDFeatureEDX);
3153
3154	if ( u32VendorEBX == X86_CPUID_VENDOR_INTEL_EBX
3155	&& u32VendorECX == X86_CPUID_VENDOR_INTEL_ECX
3156	&& u32VendorEDX == X86_CPUID_VENDOR_INTEL_EDX
3157	)
3158	{
3159	if ( (u32FeaturesECX & X86_CPUID_FEATURE_ECX_VMX)
3160	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3161	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3162	)
3163	{
3164	val = ASMRdMsr(MSR_IA32_FEATURE_CONTROL);
3165	/*
3166	* Both the LOCK and VMXON bit must be set; otherwise VMXON will generate a #GP.
3167	* Once the lock bit is set, this MSR can no longer be modified.
3168	*/
3169	if ( (val & (MSR_IA32_FEATURE_CONTROL_VMXON\|MSR_IA32_FEATURE_CONTROL_LOCK))
3170	== (MSR_IA32_FEATURE_CONTROL_VMXON\|MSR_IA32_FEATURE_CONTROL_LOCK) /* enabled and locked */
3171	\|\| !(val & MSR_IA32_FEATURE_CONTROL_LOCK) /* not enabled, but not locked either */
3172	)
3173	{
3174	VMX_CAPABILITY vtCaps;
3175
3176	*pfCaps \|= SUPVTCAPS_VT_X;
3177
3178	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS);
3179	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC_USE_SECONDARY_EXEC_CTRL)
3180	{
3181	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS2);
3182	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC2_EPT)
3183	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3184	}
3185	return VINF_SUCCESS;
3186	}
3187	return VERR_VMX_MSR_LOCKED_OR_DISABLED;
3188	}
3189	return VERR_VMX_NO_VMX;
3190	}
3191
3192	if ( u32VendorEBX == X86_CPUID_VENDOR_AMD_EBX
3193	&& u32VendorECX == X86_CPUID_VENDOR_AMD_ECX
3194	&& u32VendorEDX == X86_CPUID_VENDOR_AMD_EDX
3195	)
3196	{
3197	if ( (u32AMDFeatureECX & X86_CPUID_AMD_FEATURE_ECX_SVM)
3198	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3199	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3200	)
3201	{
3202	/* Check if SVM is disabled */
3203	val = ASMRdMsr(MSR_K8_VM_CR);
3204	if (!(val & MSR_K8_VM_CR_SVM_DISABLE))
3205	{
3206	*pfCaps \|= SUPVTCAPS_AMD_V;
3207
3208	/* Query AMD features. */
3209	ASMCpuId(0x8000000A, &u32Dummy, &u32Dummy, &u32Dummy, &u32FeaturesEDX);
3210
3211	if (u32FeaturesEDX & AMD_CPUID_SVM_FEATURE_EDX_NESTED_PAGING)
3212	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3213
3214	return VINF_SUCCESS;
3215	}
3216	return VERR_SVM_DISABLED;
3217	}
3218	return VERR_SVM_NO_SVM;
3219	}
3220	}
3221
3222	return VERR_UNSUPPORTED_CPU;
3223	}
3224
3225
3226	/**
3227	* (Re-)initializes the per-cpu structure prior to starting or resuming the GIP
3228	* updating.
3229	*
3230	* @param pGipCpu The per CPU structure for this CPU.
3231	* @param u64NanoTS The current time.
3232	*/
3233	static void supdrvGipReInitCpu(PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
3234	{
3235	pGipCpu->u64TSC = ASMReadTSC() - pGipCpu->u32UpdateIntervalTSC;
3236	pGipCpu->u64NanoTS = u64NanoTS;
3237	}
3238
3239
3240	/**
3241	* Set the current TSC and NanoTS value for the CPU.
3242	*
3243	* @param idCpu The CPU ID. Unused - we have to use the APIC ID.
3244	* @param pvUser1 Pointer to the ring-0 GIP mapping.
3245	* @param pvUser2 Pointer to the variable holding the current time.
3246	*/
3247	static DECLCALLBACK(void) supdrvGipReInitCpuCallback(RTCPUID idCpu, void pvUser1, void pvUser2)
3248	{
3249	PSUPGLOBALINFOPAGE pGip = (PSUPGLOBALINFOPAGE)pvUser1;
3250	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
3251
3252	if (RT_LIKELY(iCpu < pGip->cCpus && pGip->aCPUs[iCpu].idCpu == idCpu))
3253	supdrvGipReInitCpu(&pGip->aCPUs[iCpu], (uint64_t )pvUser2);
3254
3255	NOREF(pvUser2);
3256	NOREF(idCpu);
3257	}
3258
3259
3260	/**
3261	* Maps the GIP into userspace and/or get the physical address of the GIP.
3262	*
3263	* @returns IPRT status code.
3264	* @param pSession Session to which the GIP mapping should belong.
3265	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
3266	* @param pHCPhysGip Where to store the physical address. (optional)
3267	*
3268	* @remark There is no reference counting on the mapping, so one call to this function
3269	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
3270	* and remove the session as a GIP user.
3271	*/
3272	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
3273	{
3274	int rc;
3275	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3276	RTR3PTR pGipR3 = NIL_RTR3PTR;
3277	RTHCPHYS HCPhys = NIL_RTHCPHYS;
3278	LogFlow(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
3279
3280	/*
3281	* Validate
3282	*/
3283	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3284	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
3285	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
3286
3287	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3288	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3289	#else
3290	RTSemFastMutexRequest(pDevExt->mtxGip);
3291	#endif
3292	if (pDevExt->pGip)
3293	{
3294	/*
3295	* Map it?
3296	*/
3297	rc = VINF_SUCCESS;
3298	if (ppGipR3)
3299	{
3300	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
3301	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
3302	RTMEM_PROT_READ, RTR0ProcHandleSelf());
3303	if (RT_SUCCESS(rc))
3304	pGipR3 = RTR0MemObjAddressR3(pSession->GipMapObjR3);
3305	}
3306
3307	/*
3308	* Get physical address.
3309	*/
3310	if (pHCPhysGip && RT_SUCCESS(rc))
3311	HCPhys = pDevExt->HCPhysGip;
3312
3313	/*
3314	* Reference globally.
3315	*/
3316	if (!pSession->fGipReferenced && RT_SUCCESS(rc))
3317	{
3318	pSession->fGipReferenced = 1;
3319	pDevExt->cGipUsers++;
3320	if (pDevExt->cGipUsers == 1)
3321	{
3322	PSUPGLOBALINFOPAGE pGipR0 = pDevExt->pGip;
3323	uint64_t u64NanoTS;
3324	uint32_t u32SystemResolution;
3325	unsigned i;
3326
3327	LogFlow(("SUPR0GipMap: Resumes GIP updating\n"));
3328
3329	/*
3330	* Try bump up the system timer resolution.
3331	* The more interrupts the better...
3332	*/
3333	if ( RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
3334	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
3335	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1953125 /* 512 HZ */, &u32SystemResolution))
3336	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 2000000 /* 500 HZ */, &u32SystemResolution))
3337	)
3338	{
3339	Assert(RTTimerGetSystemGranularity() <= u32SystemResolution);
3340	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
3341	}
3342
3343	if (pGipR0->aCPUs[0].u32TransactionId != 2 /* not the first time */)
3344	{
3345	for (i = 0; i < RT_ELEMENTS(pGipR0->aCPUs); i++)
3346	ASMAtomicUoWriteU32(&pGipR0->aCPUs[i].u32TransactionId,
3347	(pGipR0->aCPUs[i].u32TransactionId + GIP_UPDATEHZ_RECALC_FREQ * 2)
3348	& ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
3349	ASMAtomicWriteU64(&pGipR0->u64NanoTSLastUpdateHz, 0);
3350	}
3351
3352	u64NanoTS = RTTimeSystemNanoTS() - pGipR0->u32UpdateIntervalNS;
3353	if ( pGipR0->u32Mode == SUPGIPMODE_SYNC_TSC
3354	\|\| RTMpGetOnlineCount() == 1)
3355	supdrvGipReInitCpu(&pGipR0->aCPUs[0], u64NanoTS);
3356	else
3357	RTMpOnAll(supdrvGipReInitCpuCallback, pGipR0, &u64NanoTS);
3358
3359	#ifndef DO_NOT_START_GIP
3360	rc = RTTimerStart(pDevExt->pGipTimer, 0); AssertRC(rc);
3361	#endif
3362	rc = VINF_SUCCESS;
3363	}
3364	}
3365	}
3366	else
3367	{
3368	rc = VERR_GENERAL_FAILURE;
3369	Log(("SUPR0GipMap: GIP is not available!\n"));
3370	}
3371	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3372	RTSemMutexRelease(pDevExt->mtxGip);
3373	#else
3374	RTSemFastMutexRelease(pDevExt->mtxGip);
3375	#endif
3376
3377	/*
3378	* Write returns.
3379	*/
3380	if (pHCPhysGip)
3381	*pHCPhysGip = HCPhys;
3382	if (ppGipR3)
3383	*ppGipR3 = pGipR3;
3384
3385	#ifdef DEBUG_DARWIN_GIP
3386	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3387	#else
3388	LogFlow(( "SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3389	#endif
3390	return rc;
3391	}
3392
3393
3394	/**
3395	* Unmaps any user mapping of the GIP and terminates all GIP access
3396	* from this session.
3397	*
3398	* @returns IPRT status code.
3399	* @param pSession Session to which the GIP mapping should belong.
3400	*/
3401	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
3402	{
3403	int rc = VINF_SUCCESS;
3404	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3405	#ifdef DEBUG_DARWIN_GIP
3406	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
3407	pSession,
3408	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
3409	pSession->GipMapObjR3));
3410	#else
3411	LogFlow(("SUPR0GipUnmap: pSession=%p\n", pSession));
3412	#endif
3413	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3414
3415	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3416	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3417	#else
3418	RTSemFastMutexRequest(pDevExt->mtxGip);
3419	#endif
3420
3421	/*
3422	* Unmap anything?
3423	*/
3424	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
3425	{
3426	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
3427	AssertRC(rc);
3428	if (RT_SUCCESS(rc))
3429	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
3430	}
3431
3432	/*
3433	* Dereference global GIP.
3434	*/
3435	if (pSession->fGipReferenced && !rc)
3436	{
3437	pSession->fGipReferenced = 0;
3438	if ( pDevExt->cGipUsers > 0
3439	&& !--pDevExt->cGipUsers)
3440	{
3441	LogFlow(("SUPR0GipUnmap: Suspends GIP updating\n"));
3442	#ifndef DO_NOT_START_GIP
3443	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = VINF_SUCCESS;
3444	#endif
3445
3446	if (pDevExt->u32SystemTimerGranularityGrant)
3447	{
3448	int rc2 = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant);
3449	AssertRC(rc2);
3450	pDevExt->u32SystemTimerGranularityGrant = 0;
3451	}
3452	}
3453	}
3454
3455	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3456	RTSemMutexRelease(pDevExt->mtxGip);
3457	#else
3458	RTSemFastMutexRelease(pDevExt->mtxGip);
3459	#endif
3460
3461	return rc;
3462	}
3463
3464
3465	/**
3466	* Gets the GIP pointer.
3467	*
3468	* @returns Pointer to the GIP or NULL.
3469	*/
3470	SUPDECL(PSUPGLOBALINFOPAGE) SUPGetGIP(void)
3471	{
3472	return g_pSUPGlobalInfoPage;
3473	}
3474
3475
3476	/**
3477	* Register a component factory with the support driver.
3478	*
3479	* This is currently restricted to kernel sessions only.
3480	*
3481	* @returns VBox status code.
3482	* @retval VINF_SUCCESS on success.
3483	* @retval VERR_NO_MEMORY if we're out of memory.
3484	* @retval VERR_ALREADY_EXISTS if the factory has already been registered.
3485	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3486	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3487	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3488	*
3489	* @param pSession The SUPDRV session (must be a ring-0 session).
3490	* @param pFactory Pointer to the component factory registration structure.
3491	*
3492	* @remarks This interface is also available via SUPR0IdcComponentRegisterFactory.
3493	*/
3494	SUPR0DECL(int) SUPR0ComponentRegisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3495	{
3496	PSUPDRVFACTORYREG pNewReg;
3497	const char *psz;
3498	int rc;
3499
3500	/*
3501	* Validate parameters.
3502	*/
3503	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3504	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3505	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3506	AssertPtrReturn(pFactory->pfnQueryFactoryInterface, VERR_INVALID_POINTER);
3507	psz = RTStrEnd(pFactory->szName, sizeof(pFactory->szName));
3508	AssertReturn(psz, VERR_INVALID_PARAMETER);
3509
3510	/*
3511	* Allocate and initialize a new registration structure.
3512	*/
3513	pNewReg = (PSUPDRVFACTORYREG)RTMemAlloc(sizeof(SUPDRVFACTORYREG));
3514	if (pNewReg)
3515	{
3516	pNewReg->pNext = NULL;
3517	pNewReg->pFactory = pFactory;
3518	pNewReg->pSession = pSession;
3519	pNewReg->cchName = psz - &pFactory->szName[0];
3520
3521	/*
3522	* Add it to the tail of the list after checking for prior registration.
3523	*/
3524	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3525	if (RT_SUCCESS(rc))
3526	{
3527	PSUPDRVFACTORYREG pPrev = NULL;
3528	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3529	while (pCur && pCur->pFactory != pFactory)
3530	{
3531	pPrev = pCur;
3532	pCur = pCur->pNext;
3533	}
3534	if (!pCur)
3535	{
3536	if (pPrev)
3537	pPrev->pNext = pNewReg;
3538	else
3539	pSession->pDevExt->pComponentFactoryHead = pNewReg;
3540	rc = VINF_SUCCESS;
3541	}
3542	else
3543	rc = VERR_ALREADY_EXISTS;
3544
3545	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3546	}
3547
3548	if (RT_FAILURE(rc))
3549	RTMemFree(pNewReg);
3550	}
3551	else
3552	rc = VERR_NO_MEMORY;
3553	return rc;
3554	}
3555
3556
3557	/**
3558	* Deregister a component factory.
3559	*
3560	* @returns VBox status code.
3561	* @retval VINF_SUCCESS on success.
3562	* @retval VERR_NOT_FOUND if the factory wasn't registered.
3563	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3564	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3565	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3566	*
3567	* @param pSession The SUPDRV session (must be a ring-0 session).
3568	* @param pFactory Pointer to the component factory registration structure
3569	* previously passed SUPR0ComponentRegisterFactory().
3570	*
3571	* @remarks This interface is also available via SUPR0IdcComponentDeregisterFactory.
3572	*/
3573	SUPR0DECL(int) SUPR0ComponentDeregisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3574	{
3575	int rc;
3576
3577	/*
3578	* Validate parameters.
3579	*/
3580	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3581	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3582	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3583
3584	/*
3585	* Take the lock and look for the registration record.
3586	*/
3587	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3588	if (RT_SUCCESS(rc))
3589	{
3590	PSUPDRVFACTORYREG pPrev = NULL;
3591	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3592	while (pCur && pCur->pFactory != pFactory)
3593	{
3594	pPrev = pCur;
3595	pCur = pCur->pNext;
3596	}
3597	if (pCur)
3598	{
3599	if (!pPrev)
3600	pSession->pDevExt->pComponentFactoryHead = pCur->pNext;
3601	else
3602	pPrev->pNext = pCur->pNext;
3603
3604	pCur->pNext = NULL;
3605	pCur->pFactory = NULL;
3606	pCur->pSession = NULL;
3607	rc = VINF_SUCCESS;
3608	}
3609	else
3610	rc = VERR_NOT_FOUND;
3611
3612	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3613
3614	RTMemFree(pCur);
3615	}
3616	return rc;
3617	}
3618
3619
3620	/**
3621	* Queries a component factory.
3622	*
3623	* @returns VBox status code.
3624	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3625	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3626	* @retval VERR_SUPDRV_COMPONENT_NOT_FOUND if the component factory wasn't found.
3627	* @retval VERR_SUPDRV_INTERFACE_NOT_SUPPORTED if the interface wasn't supported.
3628	*
3629	* @param pSession The SUPDRV session.
3630	* @param pszName The name of the component factory.
3631	* @param pszInterfaceUuid The UUID of the factory interface (stringified).
3632	* @param ppvFactoryIf Where to store the factory interface.
3633	*/
3634	SUPR0DECL(int) SUPR0ComponentQueryFactory(PSUPDRVSESSION pSession, const char pszName, const char pszInterfaceUuid, void **ppvFactoryIf)
3635	{
3636	const char *pszEnd;
3637	size_t cchName;
3638	int rc;
3639
3640	/*
3641	* Validate parameters.
3642	*/
3643	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3644
3645	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
3646	pszEnd = RTStrEnd(pszName, RT_SIZEOFMEMB(SUPDRVFACTORY, szName));
3647	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3648	cchName = pszEnd - pszName;
3649
3650	AssertPtrReturn(pszInterfaceUuid, VERR_INVALID_POINTER);
3651	pszEnd = RTStrEnd(pszInterfaceUuid, RTUUID_STR_LENGTH);
3652	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3653
3654	AssertPtrReturn(ppvFactoryIf, VERR_INVALID_POINTER);
3655	*ppvFactoryIf = NULL;
3656
3657	/*
3658	* Take the lock and try all factories by this name.
3659	*/
3660	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3661	if (RT_SUCCESS(rc))
3662	{
3663	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3664	rc = VERR_SUPDRV_COMPONENT_NOT_FOUND;
3665	while (pCur)
3666	{
3667	if ( pCur->cchName == cchName
3668	&& !memcmp(pCur->pFactory->szName, pszName, cchName))
3669	{
3670	void *pvFactory = pCur->pFactory->pfnQueryFactoryInterface(pCur->pFactory, pSession, pszInterfaceUuid);
3671	if (pvFactory)
3672	{
3673	*ppvFactoryIf = pvFactory;
3674	rc = VINF_SUCCESS;
3675	break;
3676	}
3677	rc = VERR_SUPDRV_INTERFACE_NOT_SUPPORTED;
3678	}
3679
3680	/* next */
3681	pCur = pCur->pNext;
3682	}
3683
3684	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3685	}
3686	return rc;
3687	}
3688
3689
3690	#if (!defined(VBOX_WITH_SUPDRV_GENERIC_TRACER) && !defined(VBOX_WITH_DTRACE_R0DRV)) \
3691	\|\| defined(RT_OS_SOLARIS)
3692	/**
3693	* Stub function.
3694	*/
3695	SUPR0DECL(void) SUPR0TracerFireProbe(uint32_t idProbe, uintptr_t uArg0, uintptr_t uArg1, uintptr_t uArg2,
3696	uintptr_t uArg3, uintptr_t uArg4)
3697	{
3698	NOREF(idProbe); NOREF(uArg0); NOREF(uArg1); NOREF(uArg2); NOREF(uArg3); NOREF(uArg4);
3699	}
3700	#endif
3701
3702	#if !defined(VBOX_WITH_SUPDRV_GENERIC_TRACER) && !defined(VBOX_WITH_DTRACE_R0DRV)
3703	/**
3704	* Stub function.
3705	*/
3706	SUPR0DECL(int) SUPR0TracerRegisterModule(void hMod, struct VTGOBJHDR pVtgHdr)
3707	{
3708	NOREF(hMod); NOREF(pVtgHdr);
3709	return VINF_SUCCESS;
3710	}
3711	#endif
3712
3713
3714	/**
3715	* Adds a memory object to the session.
3716	*
3717	* @returns IPRT status code.
3718	* @param pMem Memory tracking structure containing the
3719	* information to track.
3720	* @param pSession The session.
3721	*/
3722	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
3723	{
3724	PSUPDRVBUNDLE pBundle;
3725	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3726
3727	/*
3728	* Find free entry and record the allocation.
3729	*/
3730	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3731	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3732	{
3733	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
3734	{
3735	unsigned i;
3736	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3737	{
3738	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
3739	{
3740	pBundle->cUsed++;
3741	pBundle->aMem[i] = *pMem;
3742	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3743	return VINF_SUCCESS;
3744	}
3745	}
3746	AssertFailed(); /* !!this can't be happening!!! */
3747	}
3748	}
3749	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3750
3751	/*
3752	* Need to allocate a new bundle.
3753	* Insert into the last entry in the bundle.
3754	*/
3755	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
3756	if (!pBundle)
3757	return VERR_NO_MEMORY;
3758
3759	/* take last entry. */
3760	pBundle->cUsed++;
3761	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
3762
3763	/* insert into list. */
3764	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3765	pBundle->pNext = pSession->Bundle.pNext;
3766	pSession->Bundle.pNext = pBundle;
3767	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3768
3769	return VINF_SUCCESS;
3770	}
3771
3772
3773	/**
3774	* Releases a memory object referenced by pointer and type.
3775	*
3776	* @returns IPRT status code.
3777	* @param pSession Session data.
3778	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
3779	* @param eType Memory type.
3780	*/
3781	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
3782	{
3783	PSUPDRVBUNDLE pBundle;
3784	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3785
3786	/*
3787	* Validate input.
3788	*/
3789	if (!uPtr)
3790	{
3791	Log(("Illegal address %p\n", (void *)uPtr));
3792	return VERR_INVALID_PARAMETER;
3793	}
3794
3795	/*
3796	* Search for the address.
3797	*/
3798	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
3799	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3800	{
3801	if (pBundle->cUsed > 0)
3802	{
3803	unsigned i;
3804	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3805	{
3806	if ( pBundle->aMem[i].eType == eType
3807	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3808	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
3809	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3810	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
3811	)
3812	{
3813	/* Make a copy of it and release it outside the spinlock. */
3814	SUPDRVMEMREF Mem = pBundle->aMem[i];
3815	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
3816	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
3817	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
3818	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3819
3820	if (Mem.MapObjR3 != NIL_RTR0MEMOBJ)
3821	{
3822	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
3823	AssertRC(rc); /** @todo figure out how to handle this. */
3824	}
3825	if (Mem.MemObj != NIL_RTR0MEMOBJ)
3826	{
3827	int rc = RTR0MemObjFree(Mem.MemObj, true /* fFreeMappings */);
3828	AssertRC(rc); /** @todo figure out how to handle this. */
3829	}
3830	return VINF_SUCCESS;
3831	}
3832	}
3833	}
3834	}
3835	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
3836	Log(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
3837	return VERR_INVALID_PARAMETER;
3838	}
3839
3840
3841	/**
3842	* Opens an image. If it's the first time it's opened the call must upload
3843	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
3844	*
3845	* This is the 1st step of the loading.
3846	*
3847	* @returns IPRT status code.
3848	* @param pDevExt Device globals.
3849	* @param pSession Session data.
3850	* @param pReq The open request.
3851	*/
3852	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
3853	{
3854	int rc;
3855	PSUPDRVLDRIMAGE pImage;
3856	void *pv;
3857	size_t cchName = strlen(pReq->u.In.szName); /* (caller checked < 32). */
3858	LogFlow(("supdrvIOCtl_LdrOpen: szName=%s cbImageWithTabs=%d\n", pReq->u.In.szName, pReq->u.In.cbImageWithTabs));
3859
3860	/*
3861	* Check if we got an instance of the image already.
3862	*/
3863	supdrvLdrLock(pDevExt);
3864	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
3865	{
3866	if ( pImage->szName[cchName] == '\0'
3867	&& !memcmp(pImage->szName, pReq->u.In.szName, cchName))
3868	{
3869	/** @todo check cbImageBits and cbImageWithTabs here, if they differs that indicates that the images are different. */
3870	pImage->cUsage++;
3871	pReq->u.Out.pvImageBase = pImage->pvImage;
3872	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
3873	pReq->u.Out.fNativeLoader = pImage->fNative;
3874	supdrvLdrAddUsage(pSession, pImage);
3875	supdrvLdrUnlock(pDevExt);
3876	return VINF_SUCCESS;
3877	}
3878	}
3879	/* (not found - add it!) */
3880
3881	/*
3882	* Allocate memory.
3883	*/
3884	pv = RTMemAlloc(RT_OFFSETOF(SUPDRVLDRIMAGE, szName[cchName + 1]));
3885	if (!pv)
3886	{
3887	supdrvLdrUnlock(pDevExt);
3888	Log(("supdrvIOCtl_LdrOpen: RTMemAlloc() failed\n"));
3889	return /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_2;
3890	}
3891
3892	/*
3893	* Setup and link in the LDR stuff.
3894	*/
3895	pImage = (PSUPDRVLDRIMAGE)pv;
3896	pImage->pvImage = NULL;
3897	pImage->pvImageAlloc = NULL;
3898	pImage->cbImageWithTabs = pReq->u.In.cbImageWithTabs;
3899	pImage->cbImageBits = pReq->u.In.cbImageBits;
3900	pImage->cSymbols = 0;
3901	pImage->paSymbols = NULL;
3902	pImage->pachStrTab = NULL;
3903	pImage->cbStrTab = 0;
3904	pImage->pfnModuleInit = NULL;
3905	pImage->pfnModuleTerm = NULL;
3906	pImage->pfnServiceReqHandler = NULL;
3907	pImage->uState = SUP_IOCTL_LDR_OPEN;
3908	pImage->cUsage = 1;
3909	pImage->pDevExt = pDevExt;
3910	memcpy(pImage->szName, pReq->u.In.szName, cchName + 1);
3911
3912	/*
3913	* Try load it using the native loader, if that isn't supported, fall back
3914	* on the older method.
3915	*/
3916	pImage->fNative = true;
3917	rc = supdrvOSLdrOpen(pDevExt, pImage, pReq->u.In.szFilename);
3918	if (rc == VERR_NOT_SUPPORTED)
3919	{
3920	pImage->pvImageAlloc = RTMemExecAlloc(pImage->cbImageBits + 31);
3921	pImage->pvImage = RT_ALIGN_P(pImage->pvImageAlloc, 32);
3922	pImage->fNative = false;
3923	rc = pImage->pvImageAlloc ? VINF_SUCCESS : VERR_NO_EXEC_MEMORY;
3924	}
3925	if (RT_FAILURE(rc))
3926	{
3927	supdrvLdrUnlock(pDevExt);
3928	RTMemFree(pImage);
3929	Log(("supdrvIOCtl_LdrOpen(%s): failed - %Rrc\n", pReq->u.In.szName, rc));
3930	return rc;
3931	}
3932	Assert(VALID_PTR(pImage->pvImage) \|\| RT_FAILURE(rc));
3933
3934	/*
3935	* Link it.
3936	*/
3937	pImage->pNext = pDevExt->pLdrImages;
3938	pDevExt->pLdrImages = pImage;
3939
3940	supdrvLdrAddUsage(pSession, pImage);
3941
3942	pReq->u.Out.pvImageBase = pImage->pvImage;
3943	pReq->u.Out.fNeedsLoading = true;
3944	pReq->u.Out.fNativeLoader = pImage->fNative;
3945	supdrvLdrUnlock(pDevExt);
3946
3947	#if defined(RT_OS_WINDOWS) && defined(DEBUG)
3948	SUPR0Printf("VBoxDrv: windbg> .reload /f %s=%#p\n", pImage->szName, pImage->pvImage);
3949	#endif
3950	return VINF_SUCCESS;
3951	}
3952
3953
3954	/**
3955	* Worker that validates a pointer to an image entrypoint.
3956	*
3957	* @returns IPRT status code.
3958	* @param pDevExt The device globals.
3959	* @param pImage The loader image.
3960	* @param pv The pointer into the image.
3961	* @param fMayBeNull Whether it may be NULL.
3962	* @param pszWhat What is this entrypoint? (for logging)
3963	* @param pbImageBits The image bits prepared by ring-3.
3964	*
3965	* @remarks Will leave the lock on failure.
3966	*/
3967	static int supdrvLdrValidatePointer(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage, void *pv,
3968	bool fMayBeNull, const uint8_t pbImageBits, const char pszWhat)
3969	{
3970	if (!fMayBeNull \|\| pv)
3971	{
3972	if ((uintptr_t)pv - (uintptr_t)pImage->pvImage >= pImage->cbImageBits)
3973	{
3974	supdrvLdrUnlock(pDevExt);
3975	Log(("Out of range (%p LB %#x): %s=%p\n", pImage->pvImage, pImage->cbImageBits, pszWhat, pv));
3976	return VERR_INVALID_PARAMETER;
3977	}
3978
3979	if (pImage->fNative)
3980	{
3981	int rc = supdrvOSLdrValidatePointer(pDevExt, pImage, pv, pbImageBits);
3982	if (RT_FAILURE(rc))
3983	{
3984	supdrvLdrUnlock(pDevExt);
3985	Log(("Bad entry point address: %s=%p (rc=%Rrc)\n", pszWhat, pv, rc));
3986	return rc;
3987	}
3988	}
3989	}
3990	return VINF_SUCCESS;
3991	}
3992
3993
3994	/**
3995	* Loads the image bits.
3996	*
3997	* This is the 2nd step of the loading.
3998	*
3999	* @returns IPRT status code.
4000	* @param pDevExt Device globals.
4001	* @param pSession Session data.
4002	* @param pReq The request.
4003	*/
4004	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
4005	{
4006	PSUPDRVLDRUSAGE pUsage;
4007	PSUPDRVLDRIMAGE pImage;
4008	int rc;
4009	LogFlow(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImageWithBits=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImageWithTabs));
4010
4011	/*
4012	* Find the ldr image.
4013	*/
4014	supdrvLdrLock(pDevExt);
4015	pUsage = pSession->pLdrUsage;
4016	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4017	pUsage = pUsage->pNext;
4018	if (!pUsage)
4019	{
4020	supdrvLdrUnlock(pDevExt);
4021	Log(("SUP_IOCTL_LDR_LOAD: couldn't find image!\n"));
4022	return VERR_INVALID_HANDLE;
4023	}
4024	pImage = pUsage->pImage;
4025
4026	/*
4027	* Validate input.
4028	*/
4029	if ( pImage->cbImageWithTabs != pReq->u.In.cbImageWithTabs
4030	\|\| pImage->cbImageBits != pReq->u.In.cbImageBits)
4031	{
4032	supdrvLdrUnlock(pDevExt);
4033	Log(("SUP_IOCTL_LDR_LOAD: image size mismatch!! %d(prep) != %d(load) or %d != %d\n",
4034	pImage->cbImageWithTabs, pReq->u.In.cbImageWithTabs, pImage->cbImageBits, pReq->u.In.cbImageBits));
4035	return VERR_INVALID_HANDLE;
4036	}
4037
4038	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
4039	{
4040	unsigned uState = pImage->uState;
4041	supdrvLdrUnlock(pDevExt);
4042	if (uState != SUP_IOCTL_LDR_LOAD)
4043	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
4044	return VERR_ALREADY_LOADED;
4045	}
4046
4047	switch (pReq->u.In.eEPType)
4048	{
4049	case SUPLDRLOADEP_NOTHING:
4050	break;
4051
4052	case SUPLDRLOADEP_VMMR0:
4053	rc = supdrvLdrValidatePointer( pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0, false, pReq->u.In.abImage, "pvVMMR0");
4054	if (RT_SUCCESS(rc))
4055	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt, false, pReq->u.In.abImage, "pvVMMR0EntryInt");
4056	if (RT_SUCCESS(rc))
4057	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, false, pReq->u.In.abImage, "pvVMMR0EntryFast");
4058	if (RT_SUCCESS(rc))
4059	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx, false, pReq->u.In.abImage, "pvVMMR0EntryEx");
4060	if (RT_FAILURE(rc))
4061	return rc;
4062	break;
4063
4064	case SUPLDRLOADEP_SERVICE:
4065	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.Service.pfnServiceReq, false, pReq->u.In.abImage, "pfnServiceReq");
4066	if (RT_FAILURE(rc))
4067	return rc;
4068	if ( pReq->u.In.EP.Service.apvReserved[0] != NIL_RTR0PTR
4069	\|\| pReq->u.In.EP.Service.apvReserved[1] != NIL_RTR0PTR
4070	\|\| pReq->u.In.EP.Service.apvReserved[2] != NIL_RTR0PTR)
4071	{
4072	supdrvLdrUnlock(pDevExt);
4073	Log(("Out of range (%p LB %#x): apvReserved={%p,%p,%p} MBZ!\n",
4074	pImage->pvImage, pReq->u.In.cbImageWithTabs,
4075	pReq->u.In.EP.Service.apvReserved[0],
4076	pReq->u.In.EP.Service.apvReserved[1],
4077	pReq->u.In.EP.Service.apvReserved[2]));
4078	return VERR_INVALID_PARAMETER;
4079	}
4080	break;
4081
4082	default:
4083	supdrvLdrUnlock(pDevExt);
4084	Log(("Invalid eEPType=%d\n", pReq->u.In.eEPType));
4085	return VERR_INVALID_PARAMETER;
4086	}
4087
4088	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleInit, true, pReq->u.In.abImage, "pfnModuleInit");
4089	if (RT_FAILURE(rc))
4090	return rc;
4091	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleTerm, true, pReq->u.In.abImage, "pfnModuleTerm");
4092	if (RT_FAILURE(rc))
4093	return rc;
4094
4095	/*
4096	* Allocate and copy the tables.
4097	* (No need to do try/except as this is a buffered request.)
4098	*/
4099	pImage->cbStrTab = pReq->u.In.cbStrTab;
4100	if (pImage->cbStrTab)
4101	{
4102	pImage->pachStrTab = (char *)RTMemAlloc(pImage->cbStrTab);
4103	if (pImage->pachStrTab)
4104	memcpy(pImage->pachStrTab, &pReq->u.In.abImage[pReq->u.In.offStrTab], pImage->cbStrTab);
4105	else
4106	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_3;
4107	}
4108
4109	pImage->cSymbols = pReq->u.In.cSymbols;
4110	if (RT_SUCCESS(rc) && pImage->cSymbols)
4111	{
4112	size_t cbSymbols = pImage->cSymbols * sizeof(SUPLDRSYM);
4113	pImage->paSymbols = (PSUPLDRSYM)RTMemAlloc(cbSymbols);
4114	if (pImage->paSymbols)
4115	memcpy(pImage->paSymbols, &pReq->u.In.abImage[pReq->u.In.offSymbols], cbSymbols);
4116	else
4117	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_4;
4118	}
4119
4120	/*
4121	* Copy the bits / complete native loading.
4122	*/
4123	if (RT_SUCCESS(rc))
4124	{
4125	pImage->uState = SUP_IOCTL_LDR_LOAD;
4126	pImage->pfnModuleInit = pReq->u.In.pfnModuleInit;
4127	pImage->pfnModuleTerm = pReq->u.In.pfnModuleTerm;
4128
4129	if (pImage->fNative)
4130	rc = supdrvOSLdrLoad(pDevExt, pImage, pReq->u.In.abImage, pReq);
4131	else
4132	memcpy(pImage->pvImage, &pReq->u.In.abImage[0], pImage->cbImageBits);
4133	}
4134
4135	/*
4136	* Update any entry points.
4137	*/
4138	if (RT_SUCCESS(rc))
4139	{
4140	switch (pReq->u.In.eEPType)
4141	{
4142	default:
4143	case SUPLDRLOADEP_NOTHING:
4144	rc = VINF_SUCCESS;
4145	break;
4146	case SUPLDRLOADEP_VMMR0:
4147	rc = supdrvLdrSetVMMR0EPs(pDevExt, pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
4148	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
4149	break;
4150	case SUPLDRLOADEP_SERVICE:
4151	pImage->pfnServiceReqHandler = pReq->u.In.EP.Service.pfnServiceReq;
4152	rc = VINF_SUCCESS;
4153	break;
4154	}
4155	}
4156
4157	/*
4158	* On success call the module initialization.
4159	*/
4160	LogFlow(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
4161	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
4162	{
4163	Log(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
4164	pDevExt->pLdrInitImage = pImage;
4165	pDevExt->hLdrInitThread = RTThreadNativeSelf();
4166	rc = pImage->pfnModuleInit(pImage);
4167	pDevExt->pLdrInitImage = NULL;
4168	pDevExt->hLdrInitThread = NIL_RTNATIVETHREAD;
4169	if (RT_FAILURE(rc) && pDevExt->pvVMMR0 == pImage->pvImage)
4170	supdrvLdrUnsetVMMR0EPs(pDevExt);
4171	}
4172
4173	if (RT_FAILURE(rc))
4174	{
4175	/* Inform the tracing component in case ModuleInit registered TPs. */
4176	#ifdef VBOX_WITH_SUPDRV_GENERIC_TRACER
4177	supdrvTracerModuleUnloading(pDevExt, pImage);
4178	#elif defined(VBOX_WITH_DTRACE_R0DRV)
4179	supdrvVtgModuleUnloading(pDevExt, pImage);
4180	#endif
4181
4182	pImage->uState = SUP_IOCTL_LDR_OPEN;
4183	pImage->pfnModuleInit = NULL;
4184	pImage->pfnModuleTerm = NULL;
4185	pImage->pfnServiceReqHandler= NULL;
4186	pImage->cbStrTab = 0;
4187	RTMemFree(pImage->pachStrTab);
4188	pImage->pachStrTab = NULL;
4189	RTMemFree(pImage->paSymbols);
4190	pImage->paSymbols = NULL;
4191	pImage->cSymbols = 0;
4192	}
4193
4194	supdrvLdrUnlock(pDevExt);
4195	return rc;
4196	}
4197
4198
4199	/**
4200	* Frees a previously loaded (prep'ed) image.
4201	*
4202	* @returns IPRT status code.
4203	* @param pDevExt Device globals.
4204	* @param pSession Session data.
4205	* @param pReq The request.
4206	*/
4207	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
4208	{
4209	int rc;
4210	PSUPDRVLDRUSAGE pUsagePrev;
4211	PSUPDRVLDRUSAGE pUsage;
4212	PSUPDRVLDRIMAGE pImage;
4213	LogFlow(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
4214
4215	/*
4216	* Find the ldr image.
4217	*/
4218	supdrvLdrLock(pDevExt);
4219	pUsagePrev = NULL;
4220	pUsage = pSession->pLdrUsage;
4221	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4222	{
4223	pUsagePrev = pUsage;
4224	pUsage = pUsage->pNext;
4225	}
4226	if (!pUsage)
4227	{
4228	supdrvLdrUnlock(pDevExt);
4229	Log(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
4230	return VERR_INVALID_HANDLE;
4231	}
4232
4233	/*
4234	* Check if we can remove anything.
4235	*/
4236	rc = VINF_SUCCESS;
4237	pImage = pUsage->pImage;
4238	if (pImage->cUsage <= 1 \|\| pUsage->cUsage <= 1)
4239	{
4240	/*
4241	* Check if there are any objects with destructors in the image, if
4242	* so leave it for the session cleanup routine so we get a chance to
4243	* clean things up in the right order and not leave them all dangling.
4244	*/
4245	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
4246	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
4247	if (pImage->cUsage <= 1)
4248	{
4249	PSUPDRVOBJ pObj;
4250	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4251	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4252	{
4253	rc = VERR_DANGLING_OBJECTS;
4254	break;
4255	}
4256	}
4257	else
4258	{
4259	PSUPDRVUSAGE pGenUsage;
4260	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
4261	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4262	{
4263	rc = VERR_DANGLING_OBJECTS;
4264	break;
4265	}
4266	}
4267	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
4268	if (rc == VINF_SUCCESS)
4269	{
4270	/* unlink it */
4271	if (pUsagePrev)
4272	pUsagePrev->pNext = pUsage->pNext;
4273	else
4274	pSession->pLdrUsage = pUsage->pNext;
4275
4276	/* free it */
4277	pUsage->pImage = NULL;
4278	pUsage->pNext = NULL;
4279	RTMemFree(pUsage);
4280
4281	/*
4282	* Dereference the image.
4283	*/
4284	if (pImage->cUsage <= 1)
4285	supdrvLdrFree(pDevExt, pImage);
4286	else
4287	pImage->cUsage--;
4288	}
4289	else
4290	{
4291	Log(("supdrvIOCtl_LdrFree: Dangling objects in %p/%s!\n", pImage->pvImage, pImage->szName));
4292	rc = VINF_SUCCESS; /** @todo BRANCH-2.1: remove this after branching. */
4293	}
4294	}
4295	else
4296	{
4297	/*
4298	* Dereference both image and usage.
4299	*/
4300	pImage->cUsage--;
4301	pUsage->cUsage--;
4302	}
4303
4304	supdrvLdrUnlock(pDevExt);
4305	return rc;
4306	}
4307
4308
4309	/**
4310	* Gets the address of a symbol in an open image.
4311	*
4312	* @returns IPRT status code.
4313	* @param pDevExt Device globals.
4314	* @param pSession Session data.
4315	* @param pReq The request buffer.
4316	*/
4317	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
4318	{
4319	PSUPDRVLDRIMAGE pImage;
4320	PSUPDRVLDRUSAGE pUsage;
4321	uint32_t i;
4322	PSUPLDRSYM paSyms;
4323	const char *pchStrings;
4324	const size_t cbSymbol = strlen(pReq->u.In.szSymbol) + 1;
4325	void *pvSymbol = NULL;
4326	int rc = VERR_GENERAL_FAILURE;
4327	Log3(("supdrvIOCtl_LdrGetSymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
4328
4329	/*
4330	* Find the ldr image.
4331	*/
4332	supdrvLdrLock(pDevExt);
4333	pUsage = pSession->pLdrUsage;
4334	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4335	pUsage = pUsage->pNext;
4336	if (!pUsage)
4337	{
4338	supdrvLdrUnlock(pDevExt);
4339	Log(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
4340	return VERR_INVALID_HANDLE;
4341	}
4342	pImage = pUsage->pImage;
4343	if (pImage->uState != SUP_IOCTL_LDR_LOAD)
4344	{
4345	unsigned uState = pImage->uState;
4346	supdrvLdrUnlock(pDevExt);
4347	Log(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", uState, uState)); NOREF(uState);
4348	return VERR_ALREADY_LOADED;
4349	}
4350
4351	/*
4352	* Search the symbol strings.
4353	*
4354	* Note! The int32_t is for native loading on solaris where the data
4355	* and text segments are in very different places.
4356	*/
4357	pchStrings = pImage->pachStrTab;
4358	paSyms = pImage->paSymbols;
4359	for (i = 0; i < pImage->cSymbols; i++)
4360	{
4361	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4362	&& !memcmp(pchStrings + paSyms[i].offName, pReq->u.In.szSymbol, cbSymbol))
4363	{
4364	pvSymbol = (uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol;
4365	rc = VINF_SUCCESS;
4366	break;
4367	}
4368	}
4369	supdrvLdrUnlock(pDevExt);
4370	pReq->u.Out.pvSymbol = pvSymbol;
4371	return rc;
4372	}
4373
4374
4375	/**
4376	* Gets the address of a symbol in an open image or the support driver.
4377	*
4378	* @returns VINF_SUCCESS on success.
4379	* @returns
4380	* @param pDevExt Device globals.
4381	* @param pSession Session data.
4382	* @param pReq The request buffer.
4383	*/
4384	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq)
4385	{
4386	int rc = VINF_SUCCESS;
4387	const char *pszSymbol = pReq->u.In.pszSymbol;
4388	const char *pszModule = pReq->u.In.pszModule;
4389	size_t cbSymbol;
4390	char const *pszEnd;
4391	uint32_t i;
4392
4393	/*
4394	* Input validation.
4395	*/
4396	AssertPtrReturn(pszSymbol, VERR_INVALID_POINTER);
4397	pszEnd = RTStrEnd(pszSymbol, 512);
4398	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4399	cbSymbol = pszEnd - pszSymbol + 1;
4400
4401	if (pszModule)
4402	{
4403	AssertPtrReturn(pszModule, VERR_INVALID_POINTER);
4404	pszEnd = RTStrEnd(pszModule, 64);
4405	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4406	}
4407	Log3(("supdrvIDC_LdrGetSymbol: pszModule=%p:{%s} pszSymbol=%p:{%s}\n", pszModule, pszModule, pszSymbol, pszSymbol));
4408
4409
4410	if ( !pszModule
4411	\|\| !strcmp(pszModule, "SupDrv"))
4412	{
4413	/*
4414	* Search the support driver export table.
4415	*/
4416	for (i = 0; i < RT_ELEMENTS(g_aFunctions); i++)
4417	if (!strcmp(g_aFunctions[i].szName, pszSymbol))
4418	{
4419	pReq->u.Out.pfnSymbol = g_aFunctions[i].pfn;
4420	break;
4421	}
4422	}
4423	else
4424	{
4425	/*
4426	* Find the loader image.
4427	*/
4428	PSUPDRVLDRIMAGE pImage;
4429
4430	supdrvLdrLock(pDevExt);
4431
4432	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
4433	if (!strcmp(pImage->szName, pszModule))
4434	break;
4435	if (pImage && pImage->uState == SUP_IOCTL_LDR_LOAD)
4436	{
4437	/*
4438	* Search the symbol strings.
4439	*/
4440	const char *pchStrings = pImage->pachStrTab;
4441	PCSUPLDRSYM paSyms = pImage->paSymbols;
4442	for (i = 0; i < pImage->cSymbols; i++)
4443	{
4444	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4445	&& !memcmp(pchStrings + paSyms[i].offName, pszSymbol, cbSymbol))
4446	{
4447	/*
4448	* Found it! Calc the symbol address and add a reference to the module.
4449	*/
4450	pReq->u.Out.pfnSymbol = (PFNRT)((uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol);
4451	rc = supdrvLdrAddUsage(pSession, pImage);
4452	break;
4453	}
4454	}
4455	}
4456	else
4457	rc = pImage ? VERR_WRONG_ORDER : VERR_MODULE_NOT_FOUND;
4458
4459	supdrvLdrUnlock(pDevExt);
4460	}
4461	return rc;
4462	}
4463
4464
4465	/**
4466	* Updates the VMMR0 entry point pointers.
4467	*
4468	* @returns IPRT status code.
4469	* @param pDevExt Device globals.
4470	* @param pSession Session data.
4471	* @param pVMMR0 VMMR0 image handle.
4472	* @param pvVMMR0EntryInt VMMR0EntryInt address.
4473	* @param pvVMMR0EntryFast VMMR0EntryFast address.
4474	* @param pvVMMR0EntryEx VMMR0EntryEx address.
4475	* @remark Caller must own the loader mutex.
4476	*/
4477	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx)
4478	{
4479	int rc = VINF_SUCCESS;
4480	LogFlow(("supdrvLdrSetR0EP pvVMMR0=%p pvVMMR0EntryInt=%p\n", pvVMMR0, pvVMMR0EntryInt));
4481
4482
4483	/*
4484	* Check if not yet set.
4485	*/
4486	if (!pDevExt->pvVMMR0)
4487	{
4488	pDevExt->pvVMMR0 = pvVMMR0;
4489	pDevExt->pfnVMMR0EntryInt = pvVMMR0EntryInt;
4490	pDevExt->pfnVMMR0EntryFast = pvVMMR0EntryFast;
4491	pDevExt->pfnVMMR0EntryEx = pvVMMR0EntryEx;
4492	}
4493	else
4494	{
4495	/*
4496	* Return failure or success depending on whether the values match or not.
4497	*/
4498	if ( pDevExt->pvVMMR0 != pvVMMR0
4499	\|\| (void *)pDevExt->pfnVMMR0EntryInt != pvVMMR0EntryInt
4500	\|\| (void *)pDevExt->pfnVMMR0EntryFast != pvVMMR0EntryFast
4501	\|\| (void *)pDevExt->pfnVMMR0EntryEx != pvVMMR0EntryEx)
4502	{
4503	AssertMsgFailed(("SUP_IOCTL_LDR_SETR0EP: Already set pointing to a different module!\n"));
4504	rc = VERR_INVALID_PARAMETER;
4505	}
4506	}
4507	return rc;
4508	}
4509
4510
4511	/**
4512	* Unsets the VMMR0 entry point installed by supdrvLdrSetR0EP.
4513	*
4514	* @param pDevExt Device globals.
4515	*/
4516	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt)
4517	{
4518	pDevExt->pvVMMR0 = NULL;
4519	pDevExt->pfnVMMR0EntryInt = NULL;
4520	pDevExt->pfnVMMR0EntryFast = NULL;
4521	pDevExt->pfnVMMR0EntryEx = NULL;
4522	}
4523
4524
4525	/**
4526	* Adds a usage reference in the specified session of an image.
4527	*
4528	* Called while owning the loader semaphore.
4529	*
4530	* @returns VINF_SUCCESS on success and VERR_NO_MEMORY on failure.
4531	* @param pSession Session in question.
4532	* @param pImage Image which the session is using.
4533	*/
4534	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage)
4535	{
4536	PSUPDRVLDRUSAGE pUsage;
4537	LogFlow(("supdrvLdrAddUsage: pImage=%p\n", pImage));
4538
4539	/*
4540	* Referenced it already?
4541	*/
4542	pUsage = pSession->pLdrUsage;
4543	while (pUsage)
4544	{
4545	if (pUsage->pImage == pImage)
4546	{
4547	pUsage->cUsage++;
4548	return VINF_SUCCESS;
4549	}
4550	pUsage = pUsage->pNext;
4551	}
4552
4553	/*
4554	* Allocate new usage record.
4555	*/
4556	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
4557	AssertReturn(pUsage, /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_5);
4558	pUsage->cUsage = 1;
4559	pUsage->pImage = pImage;
4560	pUsage->pNext = pSession->pLdrUsage;
4561	pSession->pLdrUsage = pUsage;
4562	return VINF_SUCCESS;
4563	}
4564
4565
4566	/**
4567	* Frees a load image.
4568	*
4569	* @param pDevExt Pointer to device extension.
4570	* @param pImage Pointer to the image we're gonna free.
4571	* This image must exit!
4572	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
4573	*/
4574	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
4575	{
4576	PSUPDRVLDRIMAGE pImagePrev;
4577	LogFlow(("supdrvLdrFree: pImage=%p\n", pImage));
4578
4579	/* find it - arg. should've used doubly linked list. */
4580	Assert(pDevExt->pLdrImages);
4581	pImagePrev = NULL;
4582	if (pDevExt->pLdrImages != pImage)
4583	{
4584	pImagePrev = pDevExt->pLdrImages;
4585	while (pImagePrev->pNext != pImage)
4586	pImagePrev = pImagePrev->pNext;
4587	Assert(pImagePrev->pNext == pImage);
4588	}
4589
4590	/* unlink */
4591	if (pImagePrev)
4592	pImagePrev->pNext = pImage->pNext;
4593	else
4594	pDevExt->pLdrImages = pImage->pNext;
4595
4596	/* check if this is VMMR0.r0 unset its entry point pointers. */
4597	if (pDevExt->pvVMMR0 == pImage->pvImage)
4598	supdrvLdrUnsetVMMR0EPs(pDevExt);
4599
4600	/* check for objects with destructors in this image. (Shouldn't happen.) */
4601	if (pDevExt->pObjs)
4602	{
4603	unsigned cObjs = 0;
4604	PSUPDRVOBJ pObj;
4605	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
4606	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
4607	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4608	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4609	{
4610	pObj->pfnDestructor = NULL;
4611	cObjs++;
4612	}
4613	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
4614	if (cObjs)
4615	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
4616	}
4617
4618	/* call termination function if fully loaded. */
4619	if ( pImage->pfnModuleTerm
4620	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
4621	{
4622	LogFlow(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
4623	pImage->pfnModuleTerm(pImage);
4624	}
4625
4626	/* Inform the tracing component. */
4627	#ifdef VBOX_WITH_SUPDRV_GENERIC_TRACER
4628	supdrvTracerModuleUnloading(pDevExt, pImage);
4629	#elif defined(VBOX_WITH_DTRACE_R0DRV)
4630	supdrvVtgModuleUnloading(pDevExt, pImage);
4631	#endif
4632
4633	/* do native unload if appropriate. */
4634	if (pImage->fNative)
4635	supdrvOSLdrUnload(pDevExt, pImage);
4636
4637	/* free the image */
4638	pImage->cUsage = 0;
4639	pImage->pDevExt = NULL;
4640	pImage->pNext = NULL;
4641	pImage->uState = SUP_IOCTL_LDR_FREE;
4642	RTMemExecFree(pImage->pvImageAlloc, pImage->cbImageBits + 31);
4643	pImage->pvImageAlloc = NULL;
4644	RTMemFree(pImage->pachStrTab);
4645	pImage->pachStrTab = NULL;
4646	RTMemFree(pImage->paSymbols);
4647	pImage->paSymbols = NULL;
4648	RTMemFree(pImage);
4649	}
4650
4651
4652	/**
4653	* Acquires the loader lock.
4654	*
4655	* @returns IPRT status code.
4656	* @param pDevExt The device extension.
4657	*/
4658	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt)
4659	{
4660	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4661	int rc = RTSemMutexRequest(pDevExt->mtxLdr, RT_INDEFINITE_WAIT);
4662	#else
4663	int rc = RTSemFastMutexRequest(pDevExt->mtxLdr);
4664	#endif
4665	AssertRC(rc);
4666	return rc;
4667	}
4668
4669
4670	/**
4671	* Releases the loader lock.
4672	*
4673	* @returns IPRT status code.
4674	* @param pDevExt The device extension.
4675	*/
4676	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt)
4677	{
4678	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4679	return RTSemMutexRelease(pDevExt->mtxLdr);
4680	#else
4681	return RTSemFastMutexRelease(pDevExt->mtxLdr);
4682	#endif
4683	}
4684
4685
4686	/**
4687	* Implements the service call request.
4688	*
4689	* @returns VBox status code.
4690	* @param pDevExt The device extension.
4691	* @param pSession The calling session.
4692	* @param pReq The request packet, valid.
4693	*/
4694	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq)
4695	{
4696	#if !defined(RT_OS_WINDOWS) \|\| defined(DEBUG)
4697	int rc;
4698
4699	/*
4700	* Find the module first in the module referenced by the calling session.
4701	*/
4702	rc = supdrvLdrLock(pDevExt);
4703	if (RT_SUCCESS(rc))
4704	{
4705	PFNSUPR0SERVICEREQHANDLER pfnServiceReqHandler = NULL;
4706	PSUPDRVLDRUSAGE pUsage;
4707
4708	for (pUsage = pSession->pLdrUsage; pUsage; pUsage = pUsage->pNext)
4709	if ( pUsage->pImage->pfnServiceReqHandler
4710	&& !strcmp(pUsage->pImage->szName, pReq->u.In.szName))
4711	{
4712	pfnServiceReqHandler = pUsage->pImage->pfnServiceReqHandler;
4713	break;
4714	}
4715	supdrvLdrUnlock(pDevExt);
4716
4717	if (pfnServiceReqHandler)
4718	{
4719	/*
4720	* Call it.
4721	*/
4722	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
4723	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, NULL);
4724	else
4725	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0]);
4726	}
4727	else
4728	rc = VERR_SUPDRV_SERVICE_NOT_FOUND;
4729	}
4730
4731	/* log it */
4732	if ( RT_FAILURE(rc)
4733	&& rc != VERR_INTERRUPTED
4734	&& rc != VERR_TIMEOUT)
4735	Log(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4736	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4737	else
4738	Log4(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4739	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4740	return rc;
4741	#else /* RT_OS_WINDOWS && !DEBUG */
4742	return VERR_NOT_IMPLEMENTED;
4743	#endif /* RT_OS_WINDOWS && !DEBUG */
4744	}
4745
4746
4747	/**
4748	* Implements the logger settings request.
4749	*
4750	* @returns VBox status code.
4751	* @param pDevExt The device extension.
4752	* @param pSession The caller's session.
4753	* @param pReq The request.
4754	*/
4755	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq)
4756	{
4757	const char *pszGroup = &pReq->u.In.szStrings[pReq->u.In.offGroups];
4758	const char *pszFlags = &pReq->u.In.szStrings[pReq->u.In.offFlags];
4759	const char *pszDest = &pReq->u.In.szStrings[pReq->u.In.offDestination];
4760	PRTLOGGER pLogger = NULL;
4761	int rc;
4762
4763	/*
4764	* Some further validation.
4765	*/
4766	switch (pReq->u.In.fWhat)
4767	{
4768	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4769	case SUPLOGGERSETTINGS_WHAT_CREATE:
4770	break;
4771
4772	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4773	if (pszGroup \|\| pszFlags \|\| *pszDest)
4774	return VERR_INVALID_PARAMETER;
4775	if (pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_RELEASE)
4776	return VERR_ACCESS_DENIED;
4777	break;
4778
4779	default:
4780	return VERR_INTERNAL_ERROR;
4781	}
4782
4783	/*
4784	* Get the logger.
4785	*/
4786	switch (pReq->u.In.fWhich)
4787	{
4788	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4789	pLogger = RTLogGetDefaultInstance();
4790	break;
4791
4792	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4793	pLogger = RTLogRelDefaultInstance();
4794	break;
4795
4796	default:
4797	return VERR_INTERNAL_ERROR;
4798	}
4799
4800	/*
4801	* Do the job.
4802	*/
4803	switch (pReq->u.In.fWhat)
4804	{
4805	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4806	if (pLogger)
4807	{
4808	rc = RTLogFlags(pLogger, pszFlags);
4809	if (RT_SUCCESS(rc))
4810	rc = RTLogGroupSettings(pLogger, pszGroup);
4811	NOREF(pszDest);
4812	}
4813	else
4814	rc = VERR_NOT_FOUND;
4815	break;
4816
4817	case SUPLOGGERSETTINGS_WHAT_CREATE:
4818	{
4819	if (pLogger)
4820	rc = VERR_ALREADY_EXISTS;
4821	else
4822	{
4823	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
4824
4825	rc = RTLogCreate(&pLogger,
4826	0 /* fFlags */,
4827	pszGroup,
4828	pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_DEBUG
4829	? "VBOX_LOG"
4830	: "VBOX_RELEASE_LOG",
4831	RT_ELEMENTS(s_apszGroups),
4832	s_apszGroups,
4833	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER,
4834	NULL);
4835	if (RT_SUCCESS(rc))
4836	{
4837	rc = RTLogFlags(pLogger, pszFlags);
4838	NOREF(pszDest);
4839	if (RT_SUCCESS(rc))
4840	{
4841	switch (pReq->u.In.fWhich)
4842	{
4843	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4844	pLogger = RTLogSetDefaultInstance(pLogger);
4845	break;
4846	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4847	pLogger = RTLogRelSetDefaultInstance(pLogger);
4848	break;
4849	}
4850	}
4851	RTLogDestroy(pLogger);
4852	}
4853	}
4854	break;
4855	}
4856
4857	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4858	switch (pReq->u.In.fWhich)
4859	{
4860	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4861	pLogger = RTLogSetDefaultInstance(NULL);
4862	break;
4863	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4864	pLogger = RTLogRelSetDefaultInstance(NULL);
4865	break;
4866	}
4867	rc = RTLogDestroy(pLogger);
4868	break;
4869
4870	default:
4871	{
4872	rc = VERR_INTERNAL_ERROR;
4873	break;
4874	}
4875	}
4876
4877	return rc;
4878	}
4879
4880
4881	/**
4882	* Creates the GIP.
4883	*
4884	* @returns VBox status code.
4885	* @param pDevExt Instance data. GIP stuff may be updated.
4886	*/
4887	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
4888	{
4889	PSUPGLOBALINFOPAGE pGip;
4890	RTHCPHYS HCPhysGip;
4891	uint32_t u32SystemResolution;
4892	uint32_t u32Interval;
4893	unsigned cCpus;
4894	int rc;
4895
4896
4897	LogFlow(("supdrvGipCreate:\n"));
4898
4899	/* assert order */
4900	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
4901	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
4902	Assert(!pDevExt->pGipTimer);
4903
4904	/*
4905	* Check the CPU count.
4906	*/
4907	cCpus = RTMpGetArraySize();
4908	if ( cCpus > RTCPUSET_MAX_CPUS
4909	\|\| cCpus > 256 /ApicId is used for the mappings/)
4910	{
4911	SUPR0Printf("VBoxDrv: Too many CPUs (%u) for the GIP (max %u)\n", cCpus, RT_MIN(RTCPUSET_MAX_CPUS, 256));
4912	return VERR_TOO_MANY_CPUS;
4913	}
4914
4915	/*
4916	* Allocate a contiguous set of pages with a default kernel mapping.
4917	*/
4918	rc = RTR0MemObjAllocCont(&pDevExt->GipMemObj, RT_UOFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), false /fExecutable/);
4919	if (RT_FAILURE(rc))
4920	{
4921	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
4922	return rc;
4923	}
4924	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
4925	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
4926
4927	/*
4928	* Find a reasonable update interval and initialize the structure.
4929	*/
4930	u32Interval = u32SystemResolution = RTTimerGetSystemGranularity();
4931	while (u32Interval < 10000000 /* 10 ms */)
4932	u32Interval += u32SystemResolution;
4933
4934	supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), 1000000000 / u32Interval /=Hz/, cCpus);
4935
4936	/*
4937	* Create the timer.
4938	* If CPU_ALL isn't supported we'll have to fall back to synchronous mode.
4939	*/
4940	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
4941	{
4942	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, RTTIMER_FLAGS_CPU_ALL, supdrvGipAsyncTimer, pDevExt);
4943	if (rc == VERR_NOT_SUPPORTED)
4944	{
4945	OSDBGPRINT(("supdrvGipCreate: omni timer not supported, falling back to synchronous mode\n"));
4946	pGip->u32Mode = SUPGIPMODE_SYNC_TSC;
4947	}
4948	}
4949	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
4950	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0, supdrvGipSyncTimer, pDevExt);
4951	if (RT_SUCCESS(rc))
4952	{
4953	rc = RTMpNotificationRegister(supdrvGipMpEvent, pDevExt);
4954	if (RT_SUCCESS(rc))
4955	{
4956	rc = RTMpOnAll(supdrvGipInitOnCpu, pDevExt, pGip);
4957	if (RT_SUCCESS(rc))
4958	{
4959	/*
4960	* We're good.
4961	*/
4962	Log(("supdrvGipCreate: %u ns interval.\n", u32Interval));
4963	g_pSUPGlobalInfoPage = pGip;
4964	return VINF_SUCCESS;
4965	}
4966
4967	OSDBGPRINT(("supdrvGipCreate: RTMpOnAll failed with rc=%Rrc\n", rc));
4968	RTMpNotificationDeregister(supdrvGipMpEvent, pDevExt);
4969
4970	}
4971	else
4972	OSDBGPRINT(("supdrvGipCreate: failed to register MP event notfication. rc=%Rrc\n", rc));
4973	}
4974	else
4975	{
4976	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %u ns interval. rc=%Rrc\n", u32Interval, rc));
4977	Assert(!pDevExt->pGipTimer);
4978	}
4979	supdrvGipDestroy(pDevExt);
4980	return rc;
4981	}
4982
4983
4984	/**
4985	* Terminates the GIP.
4986	*
4987	* @param pDevExt Instance data. GIP stuff may be updated.
4988	*/
4989	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
4990	{
4991	int rc;
4992	#ifdef DEBUG_DARWIN_GIP
4993	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
4994	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
4995	pDevExt->pGipTimer, pDevExt->GipMemObj));
4996	#endif
4997
4998	/*
4999	* Invalid the GIP data.
5000	*/
5001	if (pDevExt->pGip)
5002	{
5003	supdrvGipTerm(pDevExt->pGip);
5004	pDevExt->pGip = NULL;
5005	}
5006	g_pSUPGlobalInfoPage = NULL;
5007
5008	/*
5009	* Destroy the timer and free the GIP memory object.
5010	*/
5011	if (pDevExt->pGipTimer)
5012	{
5013	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
5014	pDevExt->pGipTimer = NULL;
5015	}
5016
5017	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
5018	{
5019	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
5020	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
5021	}
5022
5023	/*
5024	* Finally, make sure we've release the system timer resolution request
5025	* if one actually succeeded and is still pending.
5026	*/
5027	if (pDevExt->u32SystemTimerGranularityGrant)
5028	{
5029	rc = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
5030	pDevExt->u32SystemTimerGranularityGrant = 0;
5031	}
5032	}
5033
5034
5035	/**
5036	* Timer callback function sync GIP mode.
5037	* @param pTimer The timer.
5038	* @param pvUser The device extension.
5039	*/
5040	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
5041	{
5042	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
5043	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5044	uint64_t u64TSC = ASMReadTSC();
5045	uint64_t NanoTS = RTTimeSystemNanoTS();
5046
5047	supdrvGipUpdate(pDevExt, NanoTS, u64TSC, NIL_RTCPUID, iTick);
5048
5049	ASMSetFlags(fOldFlags);
5050	}
5051
5052
5053	/**
5054	* Timer callback function for async GIP mode.
5055	* @param pTimer The timer.
5056	* @param pvUser The device extension.
5057	*/
5058	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
5059	{
5060	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
5061	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5062	RTCPUID idCpu = RTMpCpuId();
5063	uint64_t u64TSC = ASMReadTSC();
5064	uint64_t NanoTS = RTTimeSystemNanoTS();
5065
5066	/** @todo reset the transaction number and whatnot when iTick == 1. */
5067	if (pDevExt->idGipMaster == idCpu)
5068	supdrvGipUpdate(pDevExt, NanoTS, u64TSC, idCpu, iTick);
5069	else
5070	supdrvGipUpdatePerCpu(pDevExt, NanoTS, u64TSC, idCpu, ASMGetApicId(), iTick);
5071
5072	ASMSetFlags(fOldFlags);
5073	}
5074
5075
5076	/**
5077	* Finds our (@a idCpu) entry, or allocates a new one if not found.
5078	*
5079	* @returns Index of the CPU in the cache set.
5080	* @param pGip The GIP.
5081	* @param idCpu The CPU ID.
5082	*/
5083	static uint32_t supdrvGipCpuIndexFromCpuId(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
5084	{
5085	uint32_t i, cTries;
5086
5087	/*
5088	* ASSUMES that CPU IDs are constant.
5089	*/
5090	for (i = 0; i < pGip->cCpus; i++)
5091	if (pGip->aCPUs[i].idCpu == idCpu)
5092	return i;
5093
5094	cTries = 0;
5095	do
5096	{
5097	for (i = 0; i < pGip->cCpus; i++)
5098	{
5099	bool fRc;
5100	ASMAtomicCmpXchgSize(&pGip->aCPUs[i].idCpu, idCpu, NIL_RTCPUID, fRc);
5101	if (fRc)
5102	return i;
5103	}
5104	} while (cTries++ < 32);
5105	AssertReleaseFailed();
5106	return i - 1;
5107	}
5108
5109
5110	/**
5111	* The calling CPU should be accounted as online, update GIP accordingly.
5112	*
5113	* This is used by supdrvGipMpEvent as well as the supdrvGipCreate.
5114	*
5115	* @param pDevExt The device extension,
5116	* @param idCpu The CPU ID.
5117	*/
5118	static void supdrvGipMpEventOnline(PSUPDRVDEVEXT pDevExt, RTCPUID idCpu)
5119	{
5120	int iCpuSet = 0;
5121	uint16_t idApic = UINT16_MAX;
5122	uint32_t i = 0;
5123	uint64_t u64NanoTS = 0;
5124	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
5125	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5126
5127	AssertPtrReturnVoid(pGip);
5128	AssertRelease(idCpu == RTMpCpuId());
5129	Assert(pGip->cPossibleCpus == RTMpGetCount());
5130
5131	/*
5132	* Do this behind a spinlock with interrupts disabled as this can fire
5133	* on all CPUs simultaneously, see #6110.
5134	*/
5135	RTSpinlockAcquireNoInts(pDevExt->spinGip, &SpinlockTmp);
5136
5137	/*
5138	* Update the globals.
5139	*/
5140	ASMAtomicWriteU16(&pGip->cPresentCpus, RTMpGetPresentCount());
5141	ASMAtomicWriteU16(&pGip->cOnlineCpus, RTMpGetOnlineCount());
5142	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5143	if (iCpuSet >= 0)
5144	{
5145	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5146	RTCpuSetAddByIndex(&pGip->OnlineCpuSet, iCpuSet);
5147	RTCpuSetAddByIndex(&pGip->PresentCpuSet, iCpuSet);
5148	}
5149
5150	/*
5151	* Update the entry.
5152	*/
5153	u64NanoTS = RTTimeSystemNanoTS() - pGip->u32UpdateIntervalNS;
5154	i = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
5155	supdrvGipInitCpu(pGip, &pGip->aCPUs[i], u64NanoTS);
5156	idApic = ASMGetApicId();
5157	ASMAtomicWriteU16(&pGip->aCPUs[i].idApic, idApic);
5158	ASMAtomicWriteS16(&pGip->aCPUs[i].iCpuSet, (int16_t)iCpuSet);
5159	ASMAtomicWriteSize(&pGip->aCPUs[i].idCpu, idCpu);
5160
5161	/*
5162	* Update the APIC ID and CPU set index mappings.
5163	*/
5164	ASMAtomicWriteU16(&pGip->aiCpuFromApicId[idApic], i);
5165	ASMAtomicWriteU16(&pGip->aiCpuFromCpuSetIdx[iCpuSet], i);
5166
5167	/* commit it */
5168	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_ONLINE);
5169
5170	RTSpinlockReleaseNoInts(pDevExt->spinGip, &SpinlockTmp);
5171	}
5172
5173
5174	/**
5175	* The CPU should be accounted as offline, update the GIP accordingly.
5176	*
5177	* This is used by supdrvGipMpEvent.
5178	*
5179	* @param pDevExt The device extension.
5180	* @param idCpu The CPU ID.
5181	*/
5182	static void supdrvGipMpEventOffline(PSUPDRVDEVEXT pDevExt, RTCPUID idCpu)
5183	{
5184	int iCpuSet;
5185	unsigned i;
5186
5187	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5188	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
5189
5190	AssertPtrReturnVoid(pGip);
5191	RTSpinlockAcquireNoInts(pDevExt->spinGip, &SpinlockTmp);
5192
5193	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5194	AssertReturnVoid(iCpuSet >= 0);
5195
5196	i = pGip->aiCpuFromCpuSetIdx[iCpuSet];
5197	AssertReturnVoid(i < pGip->cCpus);
5198	AssertReturnVoid(pGip->aCPUs[i].idCpu == idCpu);
5199
5200	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5201	RTCpuSetDelByIndex(&pGip->OnlineCpuSet, iCpuSet);
5202
5203	/* commit it */
5204	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_OFFLINE);
5205
5206	RTSpinlockReleaseNoInts(pDevExt->spinGip, &SpinlockTmp);
5207	}
5208
5209
5210	/**
5211	* Multiprocessor event notification callback.
5212	*
5213	* This is used to make sure that the GIP master gets passed on to
5214	* another CPU. It also updates the associated CPU data.
5215	*
5216	* @param enmEvent The event.
5217	* @param idCpu The cpu it applies to.
5218	* @param pvUser Pointer to the device extension.
5219	*
5220	* @remarks This function -must- fire on the newly online'd CPU for the
5221	* RTMPEVENT_ONLINE case and can fire on any CPU for the
5222	* RTMPEVENT_OFFLINE case.
5223	*/
5224	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser)
5225	{
5226	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5227	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5228
5229	AssertRelease(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
5230
5231	/*
5232	* Update the GIP CPU data.
5233	*/
5234	if (pGip)
5235	{
5236	switch (enmEvent)
5237	{
5238	case RTMPEVENT_ONLINE:
5239	AssertRelease(idCpu == RTMpCpuId());
5240	supdrvGipMpEventOnline(pDevExt, idCpu);
5241	break;
5242	case RTMPEVENT_OFFLINE:
5243	supdrvGipMpEventOffline(pDevExt, idCpu);
5244	break;
5245
5246	}
5247	}
5248
5249	/*
5250	* Make sure there is a master GIP.
5251	*/
5252	if (enmEvent == RTMPEVENT_OFFLINE)
5253	{
5254	RTCPUID idGipMaster = ASMAtomicReadU32(&pDevExt->idGipMaster);
5255	if (idGipMaster == idCpu)
5256	{
5257	/*
5258	* Find a new GIP master.
5259	*/
5260	bool fIgnored;
5261	unsigned i;
5262	RTCPUID idNewGipMaster = NIL_RTCPUID;
5263	RTCPUSET OnlineCpus;
5264	RTMpGetOnlineSet(&OnlineCpus);
5265
5266	for (i = 0; i < RTCPUSET_MAX_CPUS; i++)
5267	{
5268	RTCPUID idCurCpu = RTMpCpuIdFromSetIndex(i);
5269	if ( RTCpuSetIsMember(&OnlineCpus, idCurCpu)
5270	&& idCurCpu != idGipMaster)
5271	{
5272	idNewGipMaster = idCurCpu;
5273	break;
5274	}
5275	}
5276
5277	Log(("supdrvGipMpEvent: Gip master %#lx -> %#lx\n", (long)idGipMaster, (long)idNewGipMaster));
5278	ASMAtomicCmpXchgSize(&pDevExt->idGipMaster, idNewGipMaster, idGipMaster, fIgnored);
5279	NOREF(fIgnored);
5280	}
5281	}
5282	}
5283
5284
5285	/**
5286	* Callback used by supdrvDetermineAsyncTSC to read the TSC on a CPU.
5287	*
5288	* @param idCpu Ignored.
5289	* @param pvUser1 Where to put the TSC.
5290	* @param pvUser2 Ignored.
5291	*/
5292	static DECLCALLBACK(void) supdrvDetermineAsyncTscWorker(RTCPUID idCpu, void pvUser1, void pvUser2)
5293	{
5294	#if 1
5295	ASMAtomicWriteU64((uint64_t volatile *)pvUser1, ASMReadTSC());
5296	#else
5297	(uint64_t )pvUser1 = ASMReadTSC();
5298	#endif
5299	}
5300
5301
5302	/**
5303	* Determine if Async GIP mode is required because of TSC drift.
5304	*
5305	* When using the default/normal timer code it is essential that the time stamp counter
5306	* (TSC) runs never backwards, that is, a read operation to the counter should return
5307	* a bigger value than any previous read operation. This is guaranteed by the latest
5308	* AMD CPUs and by newer Intel CPUs which never enter the C2 state (P4). In any other
5309	* case we have to choose the asynchronous timer mode.
5310	*
5311	* @param poffMin Pointer to the determined difference between different cores.
5312	* @return false if the time stamp counters appear to be synchronized, true otherwise.
5313	*/
5314	static bool supdrvDetermineAsyncTsc(uint64_t *poffMin)
5315	{
5316	/*
5317	* Just iterate all the cpus 8 times and make sure that the TSC is
5318	* ever increasing. We don't bother taking TSC rollover into account.
5319	*/
5320	int iEndCpu = RTMpGetArraySize();
5321	int iCpu;
5322	int cLoops = 8;
5323	bool fAsync = false;
5324	int rc = VINF_SUCCESS;
5325	uint64_t offMax = 0;
5326	uint64_t offMin = ~(uint64_t)0;
5327	uint64_t PrevTsc = ASMReadTSC();
5328
5329	while (cLoops-- > 0)
5330	{
5331	for (iCpu = 0; iCpu < iEndCpu; iCpu++)
5332	{
5333	uint64_t CurTsc;
5334	rc = RTMpOnSpecific(RTMpCpuIdFromSetIndex(iCpu), supdrvDetermineAsyncTscWorker, &CurTsc, NULL);
5335	if (RT_SUCCESS(rc))
5336	{
5337	if (CurTsc <= PrevTsc)
5338	{
5339	fAsync = true;
5340	offMin = offMax = PrevTsc - CurTsc;
5341	Log(("supdrvDetermineAsyncTsc: iCpu=%d cLoops=%d CurTsc=%llx PrevTsc=%llx\n",
5342	iCpu, cLoops, CurTsc, PrevTsc));
5343	break;
5344	}
5345
5346	/* Gather statistics (except the first time). */
5347	if (iCpu != 0 \|\| cLoops != 7)
5348	{
5349	uint64_t off = CurTsc - PrevTsc;
5350	if (off < offMin)
5351	offMin = off;
5352	if (off > offMax)
5353	offMax = off;
5354	Log2(("%d/%d: off=%llx\n", cLoops, iCpu, off));
5355	}
5356
5357	/* Next */
5358	PrevTsc = CurTsc;
5359	}
5360	else if (rc == VERR_NOT_SUPPORTED)
5361	break;
5362	else
5363	AssertMsg(rc == VERR_CPU_NOT_FOUND \|\| rc == VERR_CPU_OFFLINE, ("%d\n", rc));
5364	}
5365
5366	/* broke out of the loop. */
5367	if (iCpu < iEndCpu)
5368	break;
5369	}
5370
5371	poffMin = offMin; / Almost RTMpOnSpecific profiling. */
5372	Log(("supdrvDetermineAsyncTsc: returns %d; iEndCpu=%d rc=%d offMin=%llx offMax=%llx\n",
5373	fAsync, iEndCpu, rc, offMin, offMax));
5374	#if !defined(RT_OS_SOLARIS) && !defined(RT_OS_OS2) && !defined(RT_OS_WINDOWS)
5375	OSDBGPRINT(("vboxdrv: fAsync=%d offMin=%#lx offMax=%#lx\n", fAsync, (long)offMin, (long)offMax));
5376	#endif
5377	return fAsync;
5378	}
5379
5380
5381	/**
5382	* Determine the GIP TSC mode.
5383	*
5384	* @returns The most suitable TSC mode.
5385	* @param pDevExt Pointer to the device instance data.
5386	*/
5387	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt)
5388	{
5389	/*
5390	* On SMP we're faced with two problems:
5391	* (1) There might be a skew between the CPU, so that cpu0
5392	* returns a TSC that is slightly different from cpu1.
5393	* (2) Power management (and other things) may cause the TSC
5394	* to run at a non-constant speed, and cause the speed
5395	* to be different on the cpus. This will result in (1).
5396	*
5397	* So, on SMP systems we'll have to select the ASYNC update method
5398	* if there are symptoms of these problems.
5399	*/
5400	if (RTMpGetCount() > 1)
5401	{
5402	uint32_t uEAX, uEBX, uECX, uEDX;
5403	uint64_t u64DiffCoresIgnored;
5404
5405	/* Permit the user and/or the OS specific bits to force async mode. */
5406	if (supdrvOSGetForcedAsyncTscMode(pDevExt))
5407	return SUPGIPMODE_ASYNC_TSC;
5408
5409	/* Try check for current differences between the cpus. */
5410	if (supdrvDetermineAsyncTsc(&u64DiffCoresIgnored))
5411	return SUPGIPMODE_ASYNC_TSC;
5412
5413	/*
5414	* If the CPU supports power management and is an AMD one we
5415	* won't trust it unless it has the TscInvariant bit is set.
5416	*/
5417	/* Check for "AuthenticAMD" */
5418	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
5419	if ( uEAX >= 1
5420	&& uEBX == X86_CPUID_VENDOR_AMD_EBX
5421	&& uECX == X86_CPUID_VENDOR_AMD_ECX
5422	&& uEDX == X86_CPUID_VENDOR_AMD_EDX)
5423	{
5424	/* Check for APM support and that TscInvariant is cleared. */
5425	ASMCpuId(0x80000000, &uEAX, &uEBX, &uECX, &uEDX);
5426	if (uEAX >= 0x80000007)
5427	{
5428	ASMCpuId(0x80000007, &uEAX, &uEBX, &uECX, &uEDX);
5429	if ( !(uEDX & RT_BIT(8))/* TscInvariant */
5430	&& (uEDX & 0x3e)) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
5431	return SUPGIPMODE_ASYNC_TSC;
5432	}
5433	}
5434	}
5435	return SUPGIPMODE_SYNC_TSC;
5436	}
5437
5438
5439	/**
5440	* Initializes per-CPU GIP information.
5441	*
5442	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5443	* @param pCpu Pointer to which GIP CPU to initalize.
5444	* @param u64NanoTS The current nanosecond timestamp.
5445	*/
5446	static void supdrvGipInitCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pCpu, uint64_t u64NanoTS)
5447	{
5448	pCpu->u32TransactionId = 2;
5449	pCpu->u64NanoTS = u64NanoTS;
5450	pCpu->u64TSC = ASMReadTSC();
5451
5452	ASMAtomicWriteSize(&pCpu->enmState, SUPGIPCPUSTATE_INVALID);
5453	ASMAtomicWriteSize(&pCpu->idCpu, NIL_RTCPUID);
5454	ASMAtomicWriteS16(&pCpu->iCpuSet, -1);
5455	ASMAtomicWriteU16(&pCpu->idApic, UINT16_MAX);
5456
5457	/*
5458	* We don't know the following values until we've executed updates.
5459	* So, we'll just pretend it's a 4 GHz CPU and adjust the history it on
5460	* the 2nd timer callout.
5461	*/
5462	pCpu->u64CpuHz = _4G + 1; /* tstGIP-2 depends on this. */
5463	pCpu->u32UpdateIntervalTSC
5464	= pCpu->au32TSCHistory[0]
5465	= pCpu->au32TSCHistory[1]
5466	= pCpu->au32TSCHistory[2]
5467	= pCpu->au32TSCHistory[3]
5468	= pCpu->au32TSCHistory[4]
5469	= pCpu->au32TSCHistory[5]
5470	= pCpu->au32TSCHistory[6]
5471	= pCpu->au32TSCHistory[7]
5472	= (uint32_t)(_4G / pGip->u32UpdateHz);
5473	}
5474
5475
5476	/**
5477	* Initializes the GIP data.
5478	*
5479	* @param pDevExt Pointer to the device instance data.
5480	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5481	* @param HCPhys The physical address of the GIP.
5482	* @param u64NanoTS The current nanosecond timestamp.
5483	* @param uUpdateHz The update frequency.
5484	* @param cCpus The CPU count.
5485	*/
5486	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
5487	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned cCpus)
5488	{
5489	size_t const cbGip = RT_ALIGN_Z(RT_OFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), PAGE_SIZE);
5490	unsigned i;
5491	#ifdef DEBUG_DARWIN_GIP
5492	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5493	#else
5494	LogFlow(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5495	#endif
5496
5497	/*
5498	* Initialize the structure.
5499	*/
5500	memset(pGip, 0, cbGip);
5501	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
5502	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
5503	pGip->u32Mode = supdrvGipDeterminTscMode(pDevExt);
5504	pGip->cCpus = (uint16_t)cCpus;
5505	pGip->cPages = (uint16_t)(cbGip / PAGE_SIZE);
5506	pGip->u32UpdateHz = uUpdateHz;
5507	pGip->u32UpdateIntervalNS = 1000000000 / uUpdateHz;
5508	pGip->u64NanoTSLastUpdateHz = u64NanoTS;
5509	RTCpuSetEmpty(&pGip->OnlineCpuSet);
5510	RTCpuSetEmpty(&pGip->PresentCpuSet);
5511	RTMpGetSet(&pGip->PossibleCpuSet);
5512	pGip->cOnlineCpus = RTMpGetOnlineCount();
5513	pGip->cPresentCpus = RTMpGetPresentCount();
5514	pGip->cPossibleCpus = RTMpGetCount();
5515	pGip->idCpuMax = RTMpGetMaxCpuId();
5516	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromApicId); i++)
5517	pGip->aiCpuFromApicId[i] = 0;
5518	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx); i++)
5519	pGip->aiCpuFromCpuSetIdx[i] = UINT16_MAX;
5520
5521	for (i = 0; i < cCpus; i++)
5522	supdrvGipInitCpu(pGip, &pGip->aCPUs[i], u64NanoTS);
5523
5524	/*
5525	* Link it to the device extension.
5526	*/
5527	pDevExt->pGip = pGip;
5528	pDevExt->HCPhysGip = HCPhys;
5529	pDevExt->cGipUsers = 0;
5530	}
5531
5532
5533	/**
5534	* On CPU initialization callback for RTMpOnAll.
5535	*
5536	* @param idCpu The CPU ID.
5537	* @param pvUser1 The device extension.
5538	* @param pvUser2 The GIP.
5539	*/
5540	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2)
5541	{
5542	/* This is good enough, even though it will update some of the globals a
5543	bit to much. */
5544	supdrvGipMpEventOnline((PSUPDRVDEVEXT)pvUser1, idCpu);
5545	}
5546
5547
5548	/**
5549	* Invalidates the GIP data upon termination.
5550	*
5551	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5552	*/
5553	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
5554	{
5555	unsigned i;
5556	pGip->u32Magic = 0;
5557	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
5558	{
5559	pGip->aCPUs[i].u64NanoTS = 0;
5560	pGip->aCPUs[i].u64TSC = 0;
5561	pGip->aCPUs[i].iTSCHistoryHead = 0;
5562	}
5563	}
5564
5565
5566	/**
5567	* Worker routine for supdrvGipUpdate and supdrvGipUpdatePerCpu that
5568	* updates all the per cpu data except the transaction id.
5569	*
5570	* @param pDevExt The device extension.
5571	* @param pGipCpu Pointer to the per cpu data.
5572	* @param u64NanoTS The current time stamp.
5573	* @param u64TSC The current TSC.
5574	* @param iTick The current timer tick.
5575	*/
5576	static void supdrvGipDoUpdateCpu(PSUPDRVDEVEXT pDevExt, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick)
5577	{
5578	uint64_t u64TSCDelta;
5579	uint32_t u32UpdateIntervalTSC;
5580	uint32_t u32UpdateIntervalTSCSlack;
5581	unsigned iTSCHistoryHead;
5582	uint64_t u64CpuHz;
5583	uint32_t u32TransactionId;
5584
5585	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5586	AssertPtrReturnVoid(pGip);
5587
5588	/* Delta between this and the previous update. */
5589	ASMAtomicUoWriteU32(&pGipCpu->u32PrevUpdateIntervalNS, (uint32_t)(u64NanoTS - pGipCpu->u64NanoTS));
5590
5591	/*
5592	* Update the NanoTS.
5593	*/
5594	ASMAtomicWriteU64(&pGipCpu->u64NanoTS, u64NanoTS);
5595
5596	/*
5597	* Calc TSC delta.
5598	*/
5599	/** @todo validate the NanoTS delta, don't trust the OS to call us when it should... */
5600	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
5601	ASMAtomicWriteU64(&pGipCpu->u64TSC, u64TSC);
5602
5603	if (u64TSCDelta >> 32)
5604	{
5605	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
5606	pGipCpu->cErrors++;
5607	}
5608
5609	/*
5610	* On the 2nd and 3rd callout, reset the history with the current TSC
5611	* interval since the values entered by supdrvGipInit are totally off.
5612	* The interval on the 1st callout completely unreliable, the 2nd is a bit
5613	* better, while the 3rd should be most reliable.
5614	*/
5615	u32TransactionId = pGipCpu->u32TransactionId;
5616	if (RT_UNLIKELY( ( u32TransactionId == 5
5617	\|\| u32TransactionId == 7)
5618	&& ( iTick == 2
5619	\|\| iTick == 3) ))
5620	{
5621	unsigned i;
5622	for (i = 0; i < RT_ELEMENTS(pGipCpu->au32TSCHistory); i++)
5623	ASMAtomicUoWriteU32(&pGipCpu->au32TSCHistory[i], (uint32_t)u64TSCDelta);
5624	}
5625
5626	/*
5627	* TSC History.
5628	*/
5629	Assert(RT_ELEMENTS(pGipCpu->au32TSCHistory) == 8);
5630	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
5631	ASMAtomicWriteU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
5632	ASMAtomicWriteU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
5633
5634	/*
5635	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
5636	*/
5637	if (pGip->u32UpdateHz >= 1000)
5638	{
5639	uint32_t u32;
5640	u32 = pGipCpu->au32TSCHistory[0];
5641	u32 += pGipCpu->au32TSCHistory[1];
5642	u32 += pGipCpu->au32TSCHistory[2];
5643	u32 += pGipCpu->au32TSCHistory[3];
5644	u32 >>= 2;
5645	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
5646	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
5647	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
5648	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
5649	u32UpdateIntervalTSC >>= 2;
5650	u32UpdateIntervalTSC += u32;
5651	u32UpdateIntervalTSC >>= 1;
5652
5653	/* Value chosen for a 2GHz Athlon64 running linux 2.6.10/11, . */
5654	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
5655	}
5656	else if (pGip->u32UpdateHz >= 90)
5657	{
5658	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5659	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
5660	u32UpdateIntervalTSC >>= 1;
5661
5662	/* value chosen on a 2GHz thinkpad running windows */
5663	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
5664	}
5665	else
5666	{
5667	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5668
5669	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
5670	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
5671	}
5672	ASMAtomicWriteU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
5673
5674	/*
5675	* CpuHz.
5676	*/
5677	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, pGip->u32UpdateHz);
5678	ASMAtomicWriteU64(&pGipCpu->u64CpuHz, u64CpuHz);
5679	}
5680
5681
5682	/**
5683	* Updates the GIP.
5684	*
5685	* @param pDevExt The device extension.
5686	* @param u64NanoTS The current nanosecond timesamp.
5687	* @param u64TSC The current TSC timesamp.
5688	* @param idCpu The CPU ID.
5689	* @param iTick The current timer tick.
5690	*/
5691	static void supdrvGipUpdate(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick)
5692	{
5693	/*
5694	* Determine the relevant CPU data.
5695	*/
5696	PSUPGIPCPU pGipCpu;
5697	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5698	AssertPtrReturnVoid(pGip);
5699
5700	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
5701	pGipCpu = &pGip->aCPUs[0];
5702	else
5703	{
5704	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
5705	if (RT_UNLIKELY(iCpu >= pGip->cCpus))
5706	return;
5707	pGipCpu = &pGip->aCPUs[iCpu];
5708	if (RT_UNLIKELY(pGipCpu->idCpu != idCpu))
5709	return;
5710	}
5711
5712	/*
5713	* Start update transaction.
5714	*/
5715	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5716	{
5717	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
5718	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5719	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5720	pGipCpu->cErrors++;
5721	return;
5722	}
5723
5724	/*
5725	* Recalc the update frequency every 0x800th time.
5726	*/
5727	if (!(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
5728	{
5729	if (pGip->u64NanoTSLastUpdateHz)
5730	{
5731	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
5732	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
5733	uint32_t u32UpdateHz = (uint32_t)((UINT64_C(1000000000) * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
5734	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
5735	{
5736	ASMAtomicWriteU32(&pGip->u32UpdateHz, u32UpdateHz);
5737	ASMAtomicWriteU32(&pGip->u32UpdateIntervalNS, 1000000000 / u32UpdateHz);
5738	}
5739	#endif
5740	}
5741	ASMAtomicWriteU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS);
5742	}
5743
5744	/*
5745	* Update the data.
5746	*/
5747	supdrvGipDoUpdateCpu(pDevExt, pGipCpu, u64NanoTS, u64TSC, iTick);
5748
5749	/*
5750	* Complete transaction.
5751	*/
5752	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5753	}
5754
5755
5756	/**
5757	* Updates the per cpu GIP data for the calling cpu.
5758	*
5759	* @param pDevExt The device extension.
5760	* @param u64NanoTS The current nanosecond timesamp.
5761	* @param u64TSC The current TSC timesamp.
5762	* @param idCpu The CPU ID.
5763	* @param idApic The APIC id for the CPU index.
5764	* @param iTick The current timer tick.
5765	*/
5766	static void supdrvGipUpdatePerCpu(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC,
5767	RTCPUID idCpu, uint8_t idApic, uint64_t iTick)
5768	{
5769	uint32_t iCpu;
5770	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5771
5772	/*
5773	* Avoid a potential race when a CPU online notification doesn't fire on
5774	* the onlined CPU but the tick creeps in before the event notification is
5775	* run.
5776	*/
5777	if (RT_UNLIKELY(iTick == 1))
5778	{
5779	iCpu = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
5780	if (pGip->aCPUs[iCpu].enmState == SUPGIPCPUSTATE_OFFLINE)
5781	supdrvGipMpEventOnline(pDevExt, idCpu);
5782	}
5783
5784	iCpu = pGip->aiCpuFromApicId[idApic];
5785	if (RT_LIKELY(iCpu < pGip->cCpus))
5786	{
5787	PSUPGIPCPU pGipCpu = &pGip->aCPUs[iCpu];
5788	if (pGipCpu->idCpu == idCpu)
5789	{
5790	/*
5791	* Start update transaction.
5792	*/
5793	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5794	{
5795	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5796	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5797	pGipCpu->cErrors++;
5798	return;
5799	}
5800
5801	/*
5802	* Update the data.
5803	*/
5804	supdrvGipDoUpdateCpu(pDevExt, pGipCpu, u64NanoTS, u64TSC, iTick);
5805
5806	/*
5807	* Complete transaction.
5808	*/
5809	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5810	}
5811	}
5812	}
5813

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

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDrv.c@ 40769

Download in other formats: