SUPDrv.c@ 52478

Last change on this file since 52478 was 52192, checked in by vboxsync, 11 years ago
HostDrivers/Support, VMM: support CONFIG_PAX_KERNEXEC Linux kernels
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 238.7 KB

Line
1	/* $Id: SUPDrv.c 52192 2014-07-25 15:04:01Z vboxsync $ */
2	/** @file
3	* VBoxDrv - The VirtualBox Support Driver - Common code.
4	*/
5
6	/*
7	* Copyright (C) 2006-2013 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/hm_svm.h>
63	#include <VBox/vmm/hm_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 supdrvIOCtl_MsrProber(PSUPDRVDEVEXT pDevExt, PSUPMSRPROBER pReq);
128	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt);
129	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt);
130	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
131	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
132	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser);
133	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
134	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned cCpus);
135	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2);
136	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip);
137	static void supdrvGipUpdate(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick);
138	static void supdrvGipUpdatePerCpu(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC,
139	RTCPUID idCpu, uint8_t idApic, uint64_t iTick);
140	static void supdrvGipInitCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pCpu, uint64_t u64NanoTS);
141	static int supdrvIOCtl_ResumeSuspendedKbds(void);
142
143
144	/*******************************************************************************
145	* Global Variables *
146	*******************************************************************************/
147	DECLEXPORT(PSUPGLOBALINFOPAGE) g_pSUPGlobalInfoPage = NULL;
148
149	/**
150	* Array of the R0 SUP API.
151	*/
152	static SUPFUNC g_aFunctions[] =
153	{
154	/* SED: START */
155	/* name function */
156	/* Entries with absolute addresses determined at runtime, fixup
157	code makes ugly ASSUMPTIONS about the order here: */
158	{ "SUPR0AbsIs64bit", (void *)0 },
159	{ "SUPR0Abs64bitKernelCS", (void *)0 },
160	{ "SUPR0Abs64bitKernelSS", (void *)0 },
161	{ "SUPR0Abs64bitKernelDS", (void *)0 },
162	{ "SUPR0AbsKernelCS", (void *)0 },
163	{ "SUPR0AbsKernelSS", (void *)0 },
164	{ "SUPR0AbsKernelDS", (void *)0 },
165	{ "SUPR0AbsKernelES", (void *)0 },
166	{ "SUPR0AbsKernelFS", (void *)0 },
167	{ "SUPR0AbsKernelGS", (void *)0 },
168	/* Normal function pointers: */
169	{ "g_pSUPGlobalInfoPage", (void )&g_pSUPGlobalInfoPage }, / SED: DATA */
170	{ "SUPGetGIP", (void *)SUPGetGIP },
171	{ "SUPR0ComponentDeregisterFactory", (void *)SUPR0ComponentDeregisterFactory },
172	{ "SUPR0ComponentQueryFactory", (void *)SUPR0ComponentQueryFactory },
173	{ "SUPR0ComponentRegisterFactory", (void *)SUPR0ComponentRegisterFactory },
174	{ "SUPR0ContAlloc", (void *)SUPR0ContAlloc },
175	{ "SUPR0ContFree", (void *)SUPR0ContFree },
176	{ "SUPR0EnableVTx", (void *)SUPR0EnableVTx },
177	{ "SUPR0SuspendVTxOnCpu", (void *)SUPR0SuspendVTxOnCpu },
178	{ "SUPR0ResumeVTxOnCpu", (void *)SUPR0ResumeVTxOnCpu },
179	{ "SUPR0GetPagingMode", (void *)SUPR0GetPagingMode },
180	{ "SUPR0GetKernelFeatures", (void *)SUPR0GetKernelFeatures },
181	{ "SUPR0LockMem", (void *)SUPR0LockMem },
182	{ "SUPR0LowAlloc", (void *)SUPR0LowAlloc },
183	{ "SUPR0LowFree", (void *)SUPR0LowFree },
184	{ "SUPR0MemAlloc", (void *)SUPR0MemAlloc },
185	{ "SUPR0MemFree", (void *)SUPR0MemFree },
186	{ "SUPR0MemGetPhys", (void *)SUPR0MemGetPhys },
187	{ "SUPR0ObjAddRef", (void *)SUPR0ObjAddRef },
188	{ "SUPR0ObjAddRefEx", (void *)SUPR0ObjAddRefEx },
189	{ "SUPR0ObjRegister", (void *)SUPR0ObjRegister },
190	{ "SUPR0ObjRelease", (void *)SUPR0ObjRelease },
191	{ "SUPR0ObjVerifyAccess", (void *)SUPR0ObjVerifyAccess },
192	{ "SUPR0PageAllocEx", (void *)SUPR0PageAllocEx },
193	{ "SUPR0PageFree", (void *)SUPR0PageFree },
194	{ "SUPR0Printf", (void *)SUPR0Printf },
195	{ "SUPR0TracerDeregisterDrv", (void *)SUPR0TracerDeregisterDrv },
196	{ "SUPR0TracerDeregisterImpl", (void *)SUPR0TracerDeregisterImpl },
197	{ "SUPR0TracerFireProbe", (void *)SUPR0TracerFireProbe },
198	{ "SUPR0TracerRegisterDrv", (void *)SUPR0TracerRegisterDrv },
199	{ "SUPR0TracerRegisterImpl", (void *)SUPR0TracerRegisterImpl },
200	{ "SUPR0TracerRegisterModule", (void *)SUPR0TracerRegisterModule },
201	{ "SUPR0TracerUmodProbeFire", (void *)SUPR0TracerUmodProbeFire },
202	{ "SUPR0UnlockMem", (void *)SUPR0UnlockMem },
203	{ "SUPSemEventClose", (void *)SUPSemEventClose },
204	{ "SUPSemEventCreate", (void *)SUPSemEventCreate },
205	{ "SUPSemEventGetResolution", (void *)SUPSemEventGetResolution },
206	{ "SUPSemEventMultiClose", (void *)SUPSemEventMultiClose },
207	{ "SUPSemEventMultiCreate", (void *)SUPSemEventMultiCreate },
208	{ "SUPSemEventMultiGetResolution", (void *)SUPSemEventMultiGetResolution },
209	{ "SUPSemEventMultiReset", (void *)SUPSemEventMultiReset },
210	{ "SUPSemEventMultiSignal", (void *)SUPSemEventMultiSignal },
211	{ "SUPSemEventMultiWait", (void *)SUPSemEventMultiWait },
212	{ "SUPSemEventMultiWaitNoResume", (void *)SUPSemEventMultiWaitNoResume },
213	{ "SUPSemEventMultiWaitNsAbsIntr", (void *)SUPSemEventMultiWaitNsAbsIntr },
214	{ "SUPSemEventMultiWaitNsRelIntr", (void *)SUPSemEventMultiWaitNsRelIntr },
215	{ "SUPSemEventSignal", (void *)SUPSemEventSignal },
216	{ "SUPSemEventWait", (void *)SUPSemEventWait },
217	{ "SUPSemEventWaitNoResume", (void *)SUPSemEventWaitNoResume },
218	{ "SUPSemEventWaitNsAbsIntr", (void *)SUPSemEventWaitNsAbsIntr },
219	{ "SUPSemEventWaitNsRelIntr", (void *)SUPSemEventWaitNsRelIntr },
220
221	{ "RTAssertAreQuiet", (void *)RTAssertAreQuiet },
222	{ "RTAssertMayPanic", (void *)RTAssertMayPanic },
223	{ "RTAssertMsg1", (void *)RTAssertMsg1 },
224	{ "RTAssertMsg2AddV", (void *)RTAssertMsg2AddV },
225	{ "RTAssertMsg2V", (void *)RTAssertMsg2V },
226	{ "RTAssertSetMayPanic", (void *)RTAssertSetMayPanic },
227	{ "RTAssertSetQuiet", (void *)RTAssertSetQuiet },
228	{ "RTCrc32", (void *)RTCrc32 },
229	{ "RTCrc32Finish", (void *)RTCrc32Finish },
230	{ "RTCrc32Process", (void *)RTCrc32Process },
231	{ "RTCrc32Start", (void *)RTCrc32Start },
232	{ "RTErrConvertFromErrno", (void *)RTErrConvertFromErrno },
233	{ "RTErrConvertToErrno", (void *)RTErrConvertToErrno },
234	{ "RTHandleTableAllocWithCtx", (void *)RTHandleTableAllocWithCtx },
235	{ "RTHandleTableCreate", (void *)RTHandleTableCreate },
236	{ "RTHandleTableCreateEx", (void *)RTHandleTableCreateEx },
237	{ "RTHandleTableDestroy", (void *)RTHandleTableDestroy },
238	{ "RTHandleTableFreeWithCtx", (void *)RTHandleTableFreeWithCtx },
239	{ "RTHandleTableLookupWithCtx", (void *)RTHandleTableLookupWithCtx },
240	{ "RTLogDefaultInstance", (void *)RTLogDefaultInstance },
241	{ "RTLogGetDefaultInstance", (void *)RTLogGetDefaultInstance },
242	{ "RTLogLoggerExV", (void *)RTLogLoggerExV },
243	{ "RTLogPrintfV", (void *)RTLogPrintfV },
244	{ "RTLogRelDefaultInstance", (void *)RTLogRelDefaultInstance },
245	{ "RTLogSetDefaultInstanceThread", (void *)RTLogSetDefaultInstanceThread },
246	{ "RTMemAllocExTag", (void *)RTMemAllocExTag },
247	{ "RTMemAllocTag", (void *)RTMemAllocTag },
248	{ "RTMemAllocVarTag", (void *)RTMemAllocVarTag },
249	{ "RTMemAllocZTag", (void *)RTMemAllocZTag },
250	{ "RTMemAllocZVarTag", (void *)RTMemAllocZVarTag },
251	{ "RTMemDupExTag", (void *)RTMemDupExTag },
252	{ "RTMemDupTag", (void *)RTMemDupTag },
253	{ "RTMemFree", (void *)RTMemFree },
254	{ "RTMemFreeEx", (void *)RTMemFreeEx },
255	{ "RTMemReallocTag", (void *)RTMemReallocTag },
256	{ "RTMpCpuId", (void *)RTMpCpuId },
257	{ "RTMpCpuIdFromSetIndex", (void *)RTMpCpuIdFromSetIndex },
258	{ "RTMpCpuIdToSetIndex", (void *)RTMpCpuIdToSetIndex },
259	{ "RTMpGetArraySize", (void *)RTMpGetArraySize },
260	{ "RTMpGetCount", (void *)RTMpGetCount },
261	{ "RTMpGetMaxCpuId", (void *)RTMpGetMaxCpuId },
262	{ "RTMpGetOnlineCount", (void *)RTMpGetOnlineCount },
263	{ "RTMpGetOnlineSet", (void *)RTMpGetOnlineSet },
264	{ "RTMpGetSet", (void *)RTMpGetSet },
265	{ "RTMpIsCpuOnline", (void *)RTMpIsCpuOnline },
266	{ "RTMpIsCpuPossible", (void *)RTMpIsCpuPossible },
267	{ "RTMpIsCpuWorkPending", (void *)RTMpIsCpuWorkPending },
268	{ "RTMpNotificationDeregister", (void *)RTMpNotificationDeregister },
269	{ "RTMpNotificationRegister", (void *)RTMpNotificationRegister },
270	{ "RTMpOnAll", (void *)RTMpOnAll },
271	{ "RTMpOnOthers", (void *)RTMpOnOthers },
272	{ "RTMpOnSpecific", (void *)RTMpOnSpecific },
273	{ "RTMpPokeCpu", (void *)RTMpPokeCpu },
274	{ "RTNetIPv4AddDataChecksum", (void *)RTNetIPv4AddDataChecksum },
275	{ "RTNetIPv4AddTCPChecksum", (void *)RTNetIPv4AddTCPChecksum },
276	{ "RTNetIPv4AddUDPChecksum", (void *)RTNetIPv4AddUDPChecksum },
277	{ "RTNetIPv4FinalizeChecksum", (void *)RTNetIPv4FinalizeChecksum },
278	{ "RTNetIPv4HdrChecksum", (void *)RTNetIPv4HdrChecksum },
279	{ "RTNetIPv4IsDHCPValid", (void *)RTNetIPv4IsDHCPValid },
280	{ "RTNetIPv4IsHdrValid", (void *)RTNetIPv4IsHdrValid },
281	{ "RTNetIPv4IsTCPSizeValid", (void *)RTNetIPv4IsTCPSizeValid },
282	{ "RTNetIPv4IsTCPValid", (void *)RTNetIPv4IsTCPValid },
283	{ "RTNetIPv4IsUDPSizeValid", (void *)RTNetIPv4IsUDPSizeValid },
284	{ "RTNetIPv4IsUDPValid", (void *)RTNetIPv4IsUDPValid },
285	{ "RTNetIPv4PseudoChecksum", (void *)RTNetIPv4PseudoChecksum },
286	{ "RTNetIPv4PseudoChecksumBits", (void *)RTNetIPv4PseudoChecksumBits },
287	{ "RTNetIPv4TCPChecksum", (void *)RTNetIPv4TCPChecksum },
288	{ "RTNetIPv4UDPChecksum", (void *)RTNetIPv4UDPChecksum },
289	{ "RTNetIPv6PseudoChecksum", (void *)RTNetIPv6PseudoChecksum },
290	{ "RTNetIPv6PseudoChecksumBits", (void *)RTNetIPv6PseudoChecksumBits },
291	{ "RTNetIPv6PseudoChecksumEx", (void *)RTNetIPv6PseudoChecksumEx },
292	{ "RTNetTCPChecksum", (void *)RTNetTCPChecksum },
293	{ "RTNetUDPChecksum", (void *)RTNetUDPChecksum },
294	{ "RTPowerNotificationDeregister", (void *)RTPowerNotificationDeregister },
295	{ "RTPowerNotificationRegister", (void *)RTPowerNotificationRegister },
296	{ "RTProcSelf", (void *)RTProcSelf },
297	{ "RTR0AssertPanicSystem", (void *)RTR0AssertPanicSystem },
298	{ "RTR0MemAreKrnlAndUsrDifferent", (void *)RTR0MemAreKrnlAndUsrDifferent },
299	{ "RTR0MemKernelIsValidAddr", (void *)RTR0MemKernelIsValidAddr },
300	{ "RTR0MemKernelCopyFrom", (void *)RTR0MemKernelCopyFrom },
301	{ "RTR0MemKernelCopyTo", (void *)RTR0MemKernelCopyTo },
302	{ "RTR0MemObjAddress", (void *)RTR0MemObjAddress },
303	{ "RTR0MemObjAddressR3", (void *)RTR0MemObjAddressR3 },
304	{ "RTR0MemObjAllocContTag", (void *)RTR0MemObjAllocContTag },
305	{ "RTR0MemObjAllocLowTag", (void *)RTR0MemObjAllocLowTag },
306	{ "RTR0MemObjAllocPageTag", (void *)RTR0MemObjAllocPageTag },
307	{ "RTR0MemObjAllocPhysExTag", (void *)RTR0MemObjAllocPhysExTag },
308	{ "RTR0MemObjAllocPhysNCTag", (void *)RTR0MemObjAllocPhysNCTag },
309	{ "RTR0MemObjAllocPhysTag", (void *)RTR0MemObjAllocPhysTag },
310	{ "RTR0MemObjEnterPhysTag", (void *)RTR0MemObjEnterPhysTag },
311	{ "RTR0MemObjFree", (void *)RTR0MemObjFree },
312	{ "RTR0MemObjGetPagePhysAddr", (void *)RTR0MemObjGetPagePhysAddr },
313	{ "RTR0MemObjIsMapping", (void *)RTR0MemObjIsMapping },
314	{ "RTR0MemObjLockUserTag", (void *)RTR0MemObjLockUserTag },
315	{ "RTR0MemObjMapKernelExTag", (void *)RTR0MemObjMapKernelExTag },
316	{ "RTR0MemObjMapKernelTag", (void *)RTR0MemObjMapKernelTag },
317	{ "RTR0MemObjMapUserTag", (void *)RTR0MemObjMapUserTag },
318	{ "RTR0MemObjProtect", (void *)RTR0MemObjProtect },
319	{ "RTR0MemObjSize", (void *)RTR0MemObjSize },
320	{ "RTR0MemUserCopyFrom", (void *)RTR0MemUserCopyFrom },
321	{ "RTR0MemUserCopyTo", (void *)RTR0MemUserCopyTo },
322	{ "RTR0MemUserIsValidAddr", (void *)RTR0MemUserIsValidAddr },
323	{ "RTR0ProcHandleSelf", (void *)RTR0ProcHandleSelf },
324	{ "RTSemEventCreate", (void *)RTSemEventCreate },
325	{ "RTSemEventDestroy", (void *)RTSemEventDestroy },
326	{ "RTSemEventGetResolution", (void *)RTSemEventGetResolution },
327	{ "RTSemEventMultiCreate", (void *)RTSemEventMultiCreate },
328	{ "RTSemEventMultiDestroy", (void *)RTSemEventMultiDestroy },
329	{ "RTSemEventMultiGetResolution", (void *)RTSemEventMultiGetResolution },
330	{ "RTSemEventMultiReset", (void *)RTSemEventMultiReset },
331	{ "RTSemEventMultiSignal", (void *)RTSemEventMultiSignal },
332	{ "RTSemEventMultiWait", (void *)RTSemEventMultiWait },
333	{ "RTSemEventMultiWaitEx", (void *)RTSemEventMultiWaitEx },
334	{ "RTSemEventMultiWaitExDebug", (void *)RTSemEventMultiWaitExDebug },
335	{ "RTSemEventMultiWaitNoResume", (void *)RTSemEventMultiWaitNoResume },
336	{ "RTSemEventSignal", (void *)RTSemEventSignal },
337	{ "RTSemEventWait", (void *)RTSemEventWait },
338	{ "RTSemEventWaitEx", (void *)RTSemEventWaitEx },
339	{ "RTSemEventWaitExDebug", (void *)RTSemEventWaitExDebug },
340	{ "RTSemEventWaitNoResume", (void *)RTSemEventWaitNoResume },
341	{ "RTSemFastMutexCreate", (void *)RTSemFastMutexCreate },
342	{ "RTSemFastMutexDestroy", (void *)RTSemFastMutexDestroy },
343	{ "RTSemFastMutexRelease", (void *)RTSemFastMutexRelease },
344	{ "RTSemFastMutexRequest", (void *)RTSemFastMutexRequest },
345	{ "RTSemMutexCreate", (void *)RTSemMutexCreate },
346	{ "RTSemMutexDestroy", (void *)RTSemMutexDestroy },
347	{ "RTSemMutexRelease", (void *)RTSemMutexRelease },
348	{ "RTSemMutexRequest", (void *)RTSemMutexRequest },
349	{ "RTSemMutexRequestDebug", (void *)RTSemMutexRequestDebug },
350	{ "RTSemMutexRequestNoResume", (void *)RTSemMutexRequestNoResume },
351	{ "RTSemMutexRequestNoResumeDebug", (void *)RTSemMutexRequestNoResumeDebug },
352	{ "RTSpinlockAcquire", (void *)RTSpinlockAcquire },
353	{ "RTSpinlockCreate", (void *)RTSpinlockCreate },
354	{ "RTSpinlockDestroy", (void *)RTSpinlockDestroy },
355	{ "RTSpinlockRelease", (void *)RTSpinlockRelease },
356	{ "RTSpinlockReleaseNoInts", (void *)RTSpinlockReleaseNoInts },
357	{ "RTStrCopy", (void *)RTStrCopy },
358	{ "RTStrDupTag", (void *)RTStrDupTag },
359	{ "RTStrFormat", (void *)RTStrFormat },
360	{ "RTStrFormatNumber", (void *)RTStrFormatNumber },
361	{ "RTStrFormatTypeDeregister", (void *)RTStrFormatTypeDeregister },
362	{ "RTStrFormatTypeRegister", (void *)RTStrFormatTypeRegister },
363	{ "RTStrFormatTypeSetUser", (void *)RTStrFormatTypeSetUser },
364	{ "RTStrFormatV", (void *)RTStrFormatV },
365	{ "RTStrFree", (void *)RTStrFree },
366	{ "RTStrNCmp", (void *)RTStrNCmp },
367	{ "RTStrPrintf", (void *)RTStrPrintf },
368	{ "RTStrPrintfEx", (void *)RTStrPrintfEx },
369	{ "RTStrPrintfExV", (void *)RTStrPrintfExV },
370	{ "RTStrPrintfV", (void *)RTStrPrintfV },
371	{ "RTThreadCreate", (void *)RTThreadCreate },
372	{ "RTThreadCtxHooksAreRegistered", (void *)RTThreadCtxHooksAreRegistered },
373	{ "RTThreadCtxHooksCreate", (void *)RTThreadCtxHooksCreate },
374	{ "RTThreadCtxHooksDeregister", (void *)RTThreadCtxHooksDeregister },
375	{ "RTThreadCtxHooksRegister", (void *)RTThreadCtxHooksRegister },
376	{ "RTThreadCtxHooksRelease", (void *)RTThreadCtxHooksRelease },
377	{ "RTThreadCtxHooksRetain", (void *)RTThreadCtxHooksRetain },
378	{ "RTThreadGetName", (void *)RTThreadGetName },
379	{ "RTThreadGetNative", (void *)RTThreadGetNative },
380	{ "RTThreadGetType", (void *)RTThreadGetType },
381	{ "RTThreadIsInInterrupt", (void *)RTThreadIsInInterrupt },
382	{ "RTThreadNativeSelf", (void *)RTThreadNativeSelf },
383	{ "RTThreadPreemptDisable", (void *)RTThreadPreemptDisable },
384	{ "RTThreadPreemptIsEnabled", (void *)RTThreadPreemptIsEnabled },
385	{ "RTThreadPreemptIsPending", (void *)RTThreadPreemptIsPending },
386	{ "RTThreadPreemptIsPendingTrusty", (void *)RTThreadPreemptIsPendingTrusty },
387	{ "RTThreadPreemptIsPossible", (void *)RTThreadPreemptIsPossible },
388	{ "RTThreadPreemptRestore", (void *)RTThreadPreemptRestore },
389	{ "RTThreadSelf", (void *)RTThreadSelf },
390	{ "RTThreadSelfName", (void *)RTThreadSelfName },
391	{ "RTThreadSleep", (void *)RTThreadSleep },
392	{ "RTThreadUserReset", (void *)RTThreadUserReset },
393	{ "RTThreadUserSignal", (void *)RTThreadUserSignal },
394	{ "RTThreadUserWait", (void *)RTThreadUserWait },
395	{ "RTThreadUserWaitNoResume", (void *)RTThreadUserWaitNoResume },
396	{ "RTThreadWait", (void *)RTThreadWait },
397	{ "RTThreadWaitNoResume", (void *)RTThreadWaitNoResume },
398	{ "RTThreadYield", (void *)RTThreadYield },
399	{ "RTTimeMilliTS", (void *)RTTimeMilliTS },
400	{ "RTTimeNanoTS", (void *)RTTimeNanoTS },
401	{ "RTTimeNow", (void *)RTTimeNow },
402	{ "RTTimerCanDoHighResolution", (void *)RTTimerCanDoHighResolution },
403	{ "RTTimerChangeInterval", (void *)RTTimerChangeInterval },
404	{ "RTTimerCreate", (void *)RTTimerCreate },
405	{ "RTTimerCreateEx", (void *)RTTimerCreateEx },
406	{ "RTTimerDestroy", (void *)RTTimerDestroy },
407	{ "RTTimerGetSystemGranularity", (void *)RTTimerGetSystemGranularity },
408	{ "RTTimerReleaseSystemGranularity", (void *)RTTimerReleaseSystemGranularity },
409	{ "RTTimerRequestSystemGranularity", (void *)RTTimerRequestSystemGranularity },
410	{ "RTTimerStart", (void *)RTTimerStart },
411	{ "RTTimerStop", (void *)RTTimerStop },
412	{ "RTTimeSystemMilliTS", (void *)RTTimeSystemMilliTS },
413	{ "RTTimeSystemNanoTS", (void *)RTTimeSystemNanoTS },
414	{ "RTUuidCompare", (void *)RTUuidCompare },
415	{ "RTUuidCompareStr", (void *)RTUuidCompareStr },
416	{ "RTUuidFromStr", (void *)RTUuidFromStr },
417	/* SED: END */
418	};
419
420	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
421	/**
422	* Drag in the rest of IRPT since we share it with the
423	* rest of the kernel modules on darwin.
424	*/
425	PFNRT g_apfnVBoxDrvIPRTDeps[] =
426	{
427	/* VBoxNetAdp */
428	(PFNRT)RTRandBytes,
429	/* VBoxUSB */
430	(PFNRT)RTPathStripFilename,
431	NULL
432	};
433	#endif /* RT_OS_DARWIN \|\| RT_OS_SOLARIS \|\| RT_OS_SOLARIS */
434
435
436	/**
437	* Initializes the device extentsion structure.
438	*
439	* @returns IPRT status code.
440	* @param pDevExt The device extension to initialize.
441	* @param cbSession The size of the session structure. The size of
442	* SUPDRVSESSION may be smaller when SUPDRV_AGNOSTIC is
443	* defined because we're skipping the OS specific members
444	* then.
445	*/
446	int VBOXCALL supdrvInitDevExt(PSUPDRVDEVEXT pDevExt, size_t cbSession)
447	{
448	int rc;
449
450	#ifdef SUPDRV_WITH_RELEASE_LOGGER
451	/*
452	* Create the release log.
453	*/
454	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
455	PRTLOGGER pRelLogger;
456	rc = RTLogCreate(&pRelLogger, 0 /* fFlags */, "all",
457	"VBOX_RELEASE_LOG", RT_ELEMENTS(s_apszGroups), s_apszGroups, RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER, NULL);
458	if (RT_SUCCESS(rc))
459	RTLogRelSetDefaultInstance(pRelLogger);
460	/** @todo Add native hook for getting logger config parameters and setting
461	* them. On linux we should use the module parameter stuff... */
462	#endif
463
464	/*
465	* Initialize it.
466	*/
467	memset(pDevExt, 0, sizeof(pDevExt)); / Does not wipe OS specific tail section of the structure. */
468	pDevExt->Spinlock = NIL_RTSPINLOCK;
469	pDevExt->hGipSpinlock = NIL_RTSPINLOCK;
470	pDevExt->hSessionHashTabSpinlock = NIL_RTSPINLOCK;
471	rc = RTSpinlockCreate(&pDevExt->Spinlock, RTSPINLOCK_FLAGS_INTERRUPT_SAFE, "SUPDrvDevExt");
472	if (RT_SUCCESS(rc))
473	rc = RTSpinlockCreate(&pDevExt->hGipSpinlock, RTSPINLOCK_FLAGS_INTERRUPT_SAFE, "SUPDrvGip");
474	if (RT_SUCCESS(rc))
475	rc = RTSpinlockCreate(&pDevExt->hSessionHashTabSpinlock, RTSPINLOCK_FLAGS_INTERRUPT_SAFE, "SUPDrvSession");
476
477	if (RT_SUCCESS(rc))
478	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
479	rc = RTSemMutexCreate(&pDevExt->mtxLdr);
480	#else
481	rc = RTSemFastMutexCreate(&pDevExt->mtxLdr);
482	#endif
483	if (RT_SUCCESS(rc))
484	{
485	rc = RTSemFastMutexCreate(&pDevExt->mtxComponentFactory);
486	if (RT_SUCCESS(rc))
487	{
488	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
489	rc = RTSemMutexCreate(&pDevExt->mtxGip);
490	#else
491	rc = RTSemFastMutexCreate(&pDevExt->mtxGip);
492	#endif
493	if (RT_SUCCESS(rc))
494	{
495	rc = supdrvGipCreate(pDevExt);
496	if (RT_SUCCESS(rc))
497	{
498	rc = supdrvTracerInit(pDevExt);
499	if (RT_SUCCESS(rc))
500	{
501	pDevExt->pLdrInitImage = NULL;
502	pDevExt->hLdrInitThread = NIL_RTNATIVETHREAD;
503	pDevExt->u32Cookie = BIRD; /** @todo make this random? */
504	pDevExt->cbSession = (uint32_t)cbSession;
505
506	/*
507	* Fixup the absolute symbols.
508	*
509	* Because of the table indexing assumptions we'll have a little #ifdef orgy
510	* here rather than distributing this to OS specific files. At least for now.
511	*/
512	#ifdef RT_OS_DARWIN
513	# if ARCH_BITS == 32
514	if (SUPR0GetPagingMode() >= SUPPAGINGMODE_AMD64)
515	{
516	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
517	g_aFunctions[1].pfn = (void )0x80; / SUPR0Abs64bitKernelCS - KERNEL64_CS, seg.h */
518	g_aFunctions[2].pfn = (void )0x88; / SUPR0Abs64bitKernelSS - KERNEL64_SS, seg.h */
519	g_aFunctions[3].pfn = (void )0x88; / SUPR0Abs64bitKernelDS - KERNEL64_SS, seg.h */
520	}
521	else
522	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
523	g_aFunctions[4].pfn = (void )0x08; / SUPR0AbsKernelCS - KERNEL_CS, seg.h */
524	g_aFunctions[5].pfn = (void )0x10; / SUPR0AbsKernelSS - KERNEL_DS, seg.h */
525	g_aFunctions[6].pfn = (void )0x10; / SUPR0AbsKernelDS - KERNEL_DS, seg.h */
526	g_aFunctions[7].pfn = (void )0x10; / SUPR0AbsKernelES - KERNEL_DS, seg.h */
527	g_aFunctions[8].pfn = (void )0x10; / SUPR0AbsKernelFS - KERNEL_DS, seg.h */
528	g_aFunctions[9].pfn = (void )0x48; / SUPR0AbsKernelGS - CPU_DATA_GS, seg.h */
529	# else /* 64-bit darwin: */
530	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
531	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
532	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
533	g_aFunctions[3].pfn = (void )0; / SUPR0Abs64bitKernelDS */
534	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
535	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
536	g_aFunctions[6].pfn = (void )0; / SUPR0AbsKernelDS */
537	g_aFunctions[7].pfn = (void )0; / SUPR0AbsKernelES */
538	g_aFunctions[8].pfn = (void )0; / SUPR0AbsKernelFS */
539	g_aFunctions[9].pfn = (void )0; / SUPR0AbsKernelGS */
540
541	# endif
542	#else /* !RT_OS_DARWIN */
543	# if ARCH_BITS == 64
544	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
545	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
546	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
547	g_aFunctions[3].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0Abs64bitKernelDS */
548	# else
549	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
550	# endif
551	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
552	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
553	g_aFunctions[6].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0AbsKernelDS */
554	g_aFunctions[7].pfn = (void )(uintptr_t)ASMGetES(); / SUPR0AbsKernelES */
555	g_aFunctions[8].pfn = (void )(uintptr_t)ASMGetFS(); / SUPR0AbsKernelFS */
556	g_aFunctions[9].pfn = (void )(uintptr_t)ASMGetGS(); / SUPR0AbsKernelGS */
557	#endif /* !RT_OS_DARWIN */
558	return VINF_SUCCESS;
559	}
560
561	supdrvGipDestroy(pDevExt);
562	}
563
564	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
565	RTSemMutexDestroy(pDevExt->mtxGip);
566	pDevExt->mtxGip = NIL_RTSEMMUTEX;
567	#else
568	RTSemFastMutexDestroy(pDevExt->mtxGip);
569	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
570	#endif
571	}
572	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
573	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
574	}
575	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
576	RTSemMutexDestroy(pDevExt->mtxLdr);
577	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
578	#else
579	RTSemFastMutexDestroy(pDevExt->mtxLdr);
580	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
581	#endif
582	}
583
584	RTSpinlockDestroy(pDevExt->Spinlock);
585	pDevExt->Spinlock = NIL_RTSPINLOCK;
586	RTSpinlockDestroy(pDevExt->hGipSpinlock);
587	pDevExt->hGipSpinlock = NIL_RTSPINLOCK;
588	RTSpinlockDestroy(pDevExt->hSessionHashTabSpinlock);
589	pDevExt->hSessionHashTabSpinlock = NIL_RTSPINLOCK;
590
591	#ifdef SUPDRV_WITH_RELEASE_LOGGER
592	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
593	RTLogDestroy(RTLogSetDefaultInstance(NULL));
594	#endif
595
596	return rc;
597	}
598
599
600	/**
601	* Delete the device extension (e.g. cleanup members).
602	*
603	* @param pDevExt The device extension to delete.
604	*/
605	void VBOXCALL supdrvDeleteDevExt(PSUPDRVDEVEXT pDevExt)
606	{
607	PSUPDRVOBJ pObj;
608	PSUPDRVUSAGE pUsage;
609
610	/*
611	* Kill mutexes and spinlocks.
612	*/
613	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
614	RTSemMutexDestroy(pDevExt->mtxGip);
615	pDevExt->mtxGip = NIL_RTSEMMUTEX;
616	#else
617	RTSemFastMutexDestroy(pDevExt->mtxGip);
618	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
619	#endif
620	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
621	RTSemMutexDestroy(pDevExt->mtxLdr);
622	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
623	#else
624	RTSemFastMutexDestroy(pDevExt->mtxLdr);
625	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
626	#endif
627	RTSpinlockDestroy(pDevExt->Spinlock);
628	pDevExt->Spinlock = NIL_RTSPINLOCK;
629	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
630	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
631	RTSpinlockDestroy(pDevExt->hSessionHashTabSpinlock);
632	pDevExt->hSessionHashTabSpinlock = NIL_RTSPINLOCK;
633
634	/*
635	* Free lists.
636	*/
637	/* objects. */
638	pObj = pDevExt->pObjs;
639	Assert(!pObj); /* (can trigger on forced unloads) */
640	pDevExt->pObjs = NULL;
641	while (pObj)
642	{
643	void *pvFree = pObj;
644	pObj = pObj->pNext;
645	RTMemFree(pvFree);
646	}
647
648	/* usage records. */
649	pUsage = pDevExt->pUsageFree;
650	pDevExt->pUsageFree = NULL;
651	while (pUsage)
652	{
653	void *pvFree = pUsage;
654	pUsage = pUsage->pNext;
655	RTMemFree(pvFree);
656	}
657
658	/* kill the GIP. */
659	supdrvGipDestroy(pDevExt);
660	RTSpinlockDestroy(pDevExt->hGipSpinlock);
661	pDevExt->hGipSpinlock = NIL_RTSPINLOCK;
662
663	supdrvTracerTerm(pDevExt);
664
665	#ifdef SUPDRV_WITH_RELEASE_LOGGER
666	/* destroy the loggers. */
667	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
668	RTLogDestroy(RTLogSetDefaultInstance(NULL));
669	#endif
670	}
671
672
673	/**
674	* Create session.
675	*
676	* @returns IPRT status code.
677	* @param pDevExt Device extension.
678	* @param fUser Flag indicating whether this is a user or kernel
679	* session.
680	* @param fUnrestricted Unrestricted access (system) or restricted access
681	* (user)?
682	* @param ppSession Where to store the pointer to the session data.
683	*/
684	int VBOXCALL supdrvCreateSession(PSUPDRVDEVEXT pDevExt, bool fUser, bool fUnrestricted, PSUPDRVSESSION *ppSession)
685	{
686	int rc;
687	PSUPDRVSESSION pSession;
688
689	if (!SUP_IS_DEVEXT_VALID(pDevExt))
690	return VERR_INVALID_PARAMETER;
691
692	/*
693	* Allocate memory for the session data.
694	*/
695	pSession = *ppSession = (PSUPDRVSESSION)RTMemAllocZ(pDevExt->cbSession);
696	if (pSession)
697	{
698	/* Initialize session data. */
699	rc = RTSpinlockCreate(&pSession->Spinlock, RTSPINLOCK_FLAGS_INTERRUPT_UNSAFE, "SUPDrvSession");
700	if (!rc)
701	{
702	rc = RTHandleTableCreateEx(&pSession->hHandleTable,
703	RTHANDLETABLE_FLAGS_LOCKED_IRQ_SAFE \| RTHANDLETABLE_FLAGS_CONTEXT,
704	1 /uBase/, 32768 /cMax/, supdrvSessionObjHandleRetain, pSession);
705	if (RT_SUCCESS(rc))
706	{
707	Assert(pSession->Spinlock != NIL_RTSPINLOCK);
708	pSession->pDevExt = pDevExt;
709	pSession->u32Cookie = BIRD_INV;
710	pSession->fUnrestricted = fUnrestricted;
711	/pSession->fInHashTable = false; /
712	pSession->cRefs = 1;
713	/*pSession->pCommonNextHash = NULL;
714	pSession->ppOsSessionPtr = NULL; */
715	if (fUser)
716	{
717	pSession->Process = RTProcSelf();
718	pSession->R0Process = RTR0ProcHandleSelf();
719	}
720	else
721	{
722	pSession->Process = NIL_RTPROCESS;
723	pSession->R0Process = NIL_RTR0PROCESS;
724	}
725	/*pSession->pLdrUsage = NULL;
726	pSession->pVM = NULL;
727	pSession->pUsage = NULL;
728	pSession->pGip = NULL;
729	pSession->fGipReferenced = false;
730	pSession->Bundle.cUsed = 0; */
731	pSession->Uid = NIL_RTUID;
732	pSession->Gid = NIL_RTGID;
733	/pSession->uTracerData = 0;/
734	pSession->hTracerCaller = NIL_RTNATIVETHREAD;
735	RTListInit(&pSession->TpProviders);
736	/pSession->cTpProviders = 0;/
737	/pSession->cTpProbesFiring = 0;/
738	RTListInit(&pSession->TpUmods);
739	/RT_ZERO(pSession->apTpLookupTable);/
740
741	VBOXDRV_SESSION_CREATE(pSession, fUser);
742	LogFlow(("Created session %p initial cookie=%#x\n", pSession, pSession->u32Cookie));
743	return VINF_SUCCESS;
744	}
745
746	RTSpinlockDestroy(pSession->Spinlock);
747	}
748	RTMemFree(pSession);
749	*ppSession = NULL;
750	Log(("Failed to create spinlock, rc=%d!\n", rc));
751	}
752	else
753	rc = VERR_NO_MEMORY;
754
755	return rc;
756	}
757
758
759	/**
760	* Cleans up the session in the context of the process to which it belongs, the
761	* caller will free the session and the session spinlock.
762	*
763	* This should normally occur when the session is closed or as the process
764	* exits. Careful reference counting in the OS specfic code makes sure that
765	* there cannot be any races between process/handle cleanup callbacks and
766	* threads doing I/O control calls.
767	*
768	* @param pDevExt The device extension.
769	* @param pSession Session data.
770	*/
771	static void supdrvCleanupSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
772	{
773	int rc;
774	PSUPDRVBUNDLE pBundle;
775	LogFlow(("supdrvCleanupSession: pSession=%p\n", pSession));
776
777	Assert(!pSession->fInHashTable);
778	Assert(!pSession->ppOsSessionPtr);
779	AssertReleaseMsg(pSession->R0Process == RTR0ProcHandleSelf() \|\| pSession->R0Process == NIL_RTR0PROCESS,
780	("R0Process=%p cur=%p; Process=%u curpid=%u\n", RTR0ProcHandleSelf(), RTProcSelf()));
781
782	/*
783	* Remove logger instances related to this session.
784	*/
785	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pSession);
786
787	/*
788	* Destroy the handle table.
789	*/
790	rc = RTHandleTableDestroy(pSession->hHandleTable, supdrvSessionObjHandleDelete, pSession);
791	AssertRC(rc);
792	pSession->hHandleTable = NIL_RTHANDLETABLE;
793
794	/*
795	* Release object references made in this session.
796	* In theory there should be noone racing us in this session.
797	*/
798	Log2(("release objects - start\n"));
799	if (pSession->pUsage)
800	{
801	PSUPDRVUSAGE pUsage;
802	RTSpinlockAcquire(pDevExt->Spinlock);
803
804	while ((pUsage = pSession->pUsage) != NULL)
805	{
806	PSUPDRVOBJ pObj = pUsage->pObj;
807	pSession->pUsage = pUsage->pNext;
808
809	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
810	if (pUsage->cUsage < pObj->cUsage)
811	{
812	pObj->cUsage -= pUsage->cUsage;
813	RTSpinlockRelease(pDevExt->Spinlock);
814	}
815	else
816	{
817	/* Destroy the object and free the record. */
818	if (pDevExt->pObjs == pObj)
819	pDevExt->pObjs = pObj->pNext;
820	else
821	{
822	PSUPDRVOBJ pObjPrev;
823	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
824	if (pObjPrev->pNext == pObj)
825	{
826	pObjPrev->pNext = pObj->pNext;
827	break;
828	}
829	Assert(pObjPrev);
830	}
831	RTSpinlockRelease(pDevExt->Spinlock);
832
833	Log(("supdrvCleanupSession: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
834	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
835	if (pObj->pfnDestructor)
836	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
837	RTMemFree(pObj);
838	}
839
840	/* free it and continue. */
841	RTMemFree(pUsage);
842
843	RTSpinlockAcquire(pDevExt->Spinlock);
844	}
845
846	RTSpinlockRelease(pDevExt->Spinlock);
847	AssertMsg(!pSession->pUsage, ("Some buster reregistered an object during desturction!\n"));
848	}
849	Log2(("release objects - done\n"));
850
851	/*
852	* Do tracer cleanups related to this session.
853	*/
854	Log2(("release tracer stuff - start\n"));
855	supdrvTracerCleanupSession(pDevExt, pSession);
856	Log2(("release tracer stuff - end\n"));
857
858	/*
859	* Release memory allocated in the session.
860	*
861	* We do not serialize this as we assume that the application will
862	* not allocated memory while closing the file handle object.
863	*/
864	Log2(("freeing memory:\n"));
865	pBundle = &pSession->Bundle;
866	while (pBundle)
867	{
868	PSUPDRVBUNDLE pToFree;
869	unsigned i;
870
871	/*
872	* Check and unlock all entries in the bundle.
873	*/
874	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
875	{
876	if (pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ)
877	{
878	Log2(("eType=%d pvR0=%p pvR3=%p cb=%ld\n", pBundle->aMem[i].eType, RTR0MemObjAddress(pBundle->aMem[i].MemObj),
879	(void *)RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3), (long)RTR0MemObjSize(pBundle->aMem[i].MemObj)));
880	if (pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ)
881	{
882	rc = RTR0MemObjFree(pBundle->aMem[i].MapObjR3, false);
883	AssertRC(rc); /** @todo figure out how to handle this. */
884	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
885	}
886	rc = RTR0MemObjFree(pBundle->aMem[i].MemObj, true /* fFreeMappings */);
887	AssertRC(rc); /** @todo figure out how to handle this. */
888	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
889	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
890	}
891	}
892
893	/*
894	* Advance and free previous bundle.
895	*/
896	pToFree = pBundle;
897	pBundle = pBundle->pNext;
898
899	pToFree->pNext = NULL;
900	pToFree->cUsed = 0;
901	if (pToFree != &pSession->Bundle)
902	RTMemFree(pToFree);
903	}
904	Log2(("freeing memory - done\n"));
905
906	/*
907	* Deregister component factories.
908	*/
909	RTSemFastMutexRequest(pDevExt->mtxComponentFactory);
910	Log2(("deregistering component factories:\n"));
911	if (pDevExt->pComponentFactoryHead)
912	{
913	PSUPDRVFACTORYREG pPrev = NULL;
914	PSUPDRVFACTORYREG pCur = pDevExt->pComponentFactoryHead;
915	while (pCur)
916	{
917	if (pCur->pSession == pSession)
918	{
919	/* unlink it */
920	PSUPDRVFACTORYREG pNext = pCur->pNext;
921	if (pPrev)
922	pPrev->pNext = pNext;
923	else
924	pDevExt->pComponentFactoryHead = pNext;
925
926	/* free it */
927	pCur->pNext = NULL;
928	pCur->pSession = NULL;
929	pCur->pFactory = NULL;
930	RTMemFree(pCur);
931
932	/* next */
933	pCur = pNext;
934	}
935	else
936	{
937	/* next */
938	pPrev = pCur;
939	pCur = pCur->pNext;
940	}
941	}
942	}
943	RTSemFastMutexRelease(pDevExt->mtxComponentFactory);
944	Log2(("deregistering component factories - done\n"));
945
946	/*
947	* Loaded images needs to be dereferenced and possibly freed up.
948	*/
949	supdrvLdrLock(pDevExt);
950	Log2(("freeing images:\n"));
951	if (pSession->pLdrUsage)
952	{
953	PSUPDRVLDRUSAGE pUsage = pSession->pLdrUsage;
954	pSession->pLdrUsage = NULL;
955	while (pUsage)
956	{
957	void *pvFree = pUsage;
958	PSUPDRVLDRIMAGE pImage = pUsage->pImage;
959	if (pImage->cUsage > pUsage->cUsage)
960	pImage->cUsage -= pUsage->cUsage;
961	else
962	supdrvLdrFree(pDevExt, pImage);
963	pUsage->pImage = NULL;
964	pUsage = pUsage->pNext;
965	RTMemFree(pvFree);
966	}
967	}
968	supdrvLdrUnlock(pDevExt);
969	Log2(("freeing images - done\n"));
970
971	/*
972	* Unmap the GIP.
973	*/
974	Log2(("umapping GIP:\n"));
975	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
976	{
977	SUPR0GipUnmap(pSession);
978	pSession->fGipReferenced = 0;
979	}
980	Log2(("umapping GIP - done\n"));
981	}
982
983
984	/**
985	* Common code for freeing a session when the reference count reaches zero.
986	*
987	* @param pDevExt Device extension.
988	* @param pSession Session data.
989	* This data will be freed by this routine.
990	*/
991	static void supdrvDestroySession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
992	{
993	VBOXDRV_SESSION_CLOSE(pSession);
994
995	/*
996	* Cleanup the session first.
997	*/
998	supdrvCleanupSession(pDevExt, pSession);
999	supdrvOSCleanupSession(pDevExt, pSession);
1000
1001	/*
1002	* Free the rest of the session stuff.
1003	*/
1004	RTSpinlockDestroy(pSession->Spinlock);
1005	pSession->Spinlock = NIL_RTSPINLOCK;
1006	pSession->pDevExt = NULL;
1007	RTMemFree(pSession);
1008	LogFlow(("supdrvDestroySession: returns\n"));
1009	}
1010
1011
1012	/**
1013	* Inserts the session into the global hash table.
1014	*
1015	* @retval VINF_SUCCESS on success.
1016	* @retval VERR_WRONG_ORDER if the session was already inserted (asserted).
1017	* @retval VERR_INVALID_PARAMETER if the session handle is invalid or a ring-0
1018	* session (asserted).
1019	* @retval VERR_DUPLICATE if there is already a session for that pid.
1020	*
1021	* @param pDevExt The device extension.
1022	* @param pSession The session.
1023	* @param ppOsSessionPtr Pointer to the OS session pointer, if any is
1024	* available and used. This will set to point to the
1025	* session while under the protection of the session
1026	* hash table spinlock. It will also be kept in
1027	* PSUPDRVSESSION::ppOsSessionPtr for lookup and
1028	* cleanup use.
1029	* @param pvUser Argument for supdrvOSSessionHashTabInserted.
1030	*/
1031	int VBOXCALL supdrvSessionHashTabInsert(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVSESSION *ppOsSessionPtr,
1032	void *pvUser)
1033	{
1034	PSUPDRVSESSION pCur;
1035	unsigned iHash;
1036
1037	/*
1038	* Validate input.
1039	*/
1040	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1041	AssertReturn(pSession->R0Process != NIL_RTR0PROCESS, VERR_INVALID_PARAMETER);
1042
1043	/*
1044	* Calculate the hash table index and acquire the spinlock.
1045	*/
1046	iHash = SUPDRV_SESSION_HASH(pSession->Process);
1047
1048	RTSpinlockAcquire(pDevExt->hSessionHashTabSpinlock);
1049
1050	/*
1051	* If there are a collisions, we need to carefully check if we got a
1052	* duplicate. There can only be one open session per process.
1053	*/
1054	pCur = pDevExt->apSessionHashTab[iHash];
1055	if (pCur)
1056	{
1057	while (pCur && pCur->Process != pSession->Process)
1058	pCur = pCur->pCommonNextHash;
1059
1060	if (pCur)
1061	{
1062	RTSpinlockRelease(pDevExt->hSessionHashTabSpinlock);
1063	if (pCur == pSession)
1064	{
1065	Assert(pSession->fInHashTable);
1066	AssertFailed();
1067	return VERR_WRONG_ORDER;
1068	}
1069	Assert(!pSession->fInHashTable);
1070	if (pCur->R0Process == pSession->R0Process)
1071	return VERR_RESOURCE_IN_USE;
1072	return VERR_DUPLICATE;
1073	}
1074	}
1075	Assert(!pSession->fInHashTable);
1076	Assert(!pSession->ppOsSessionPtr);
1077
1078	/*
1079	* Insert it, doing a callout to the OS specific code in case it has
1080	* anything it wishes to do while we're holding the spinlock.
1081	*/
1082	pSession->pCommonNextHash = pDevExt->apSessionHashTab[iHash];
1083	pDevExt->apSessionHashTab[iHash] = pSession;
1084	pSession->fInHashTable = true;
1085	ASMAtomicIncS32(&pDevExt->cSessions);
1086
1087	pSession->ppOsSessionPtr = ppOsSessionPtr;
1088	if (ppOsSessionPtr)
1089	ASMAtomicWritePtr(ppOsSessionPtr, pSession);
1090
1091	supdrvOSSessionHashTabInserted(pDevExt, pSession, pvUser);
1092
1093	/*
1094	* Retain a reference for the pointer in the session table.
1095	*/
1096	ASMAtomicIncU32(&pSession->cRefs);
1097
1098	RTSpinlockRelease(pDevExt->hSessionHashTabSpinlock);
1099	return VINF_SUCCESS;
1100	}
1101
1102
1103	/**
1104	* Removes the session from the global hash table.
1105	*
1106	* @retval VINF_SUCCESS on success.
1107	* @retval VERR_NOT_FOUND if the session was already removed (asserted).
1108	* @retval VERR_INVALID_PARAMETER if the session handle is invalid or a ring-0
1109	* session (asserted).
1110	*
1111	* @param pDevExt The device extension.
1112	* @param pSession The session. The caller is expected to have a reference
1113	* to this so it won't croak on us when we release the hash
1114	* table reference.
1115	* @param pvUser OS specific context value for the
1116	* supdrvOSSessionHashTabInserted callback.
1117	*/
1118	int VBOXCALL supdrvSessionHashTabRemove(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, void *pvUser)
1119	{
1120	PSUPDRVSESSION pCur;
1121	unsigned iHash;
1122	int32_t cRefs;
1123
1124	/*
1125	* Validate input.
1126	*/
1127	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1128	AssertReturn(pSession->R0Process != NIL_RTR0PROCESS, VERR_INVALID_PARAMETER);
1129
1130	/*
1131	* Calculate the hash table index and acquire the spinlock.
1132	*/
1133	iHash = SUPDRV_SESSION_HASH(pSession->Process);
1134
1135	RTSpinlockAcquire(pDevExt->hSessionHashTabSpinlock);
1136
1137	/*
1138	* Unlink it.
1139	*/
1140	pCur = pDevExt->apSessionHashTab[iHash];
1141	if (pCur == pSession)
1142	pDevExt->apSessionHashTab[iHash] = pSession->pCommonNextHash;
1143	else
1144	{
1145	PSUPDRVSESSION pPrev = pCur;
1146	while (pCur && pCur != pSession)
1147	{
1148	pPrev = pCur;
1149	pCur = pCur->pCommonNextHash;
1150	}
1151	if (pCur)
1152	pPrev->pCommonNextHash = pCur->pCommonNextHash;
1153	else
1154	{
1155	Assert(!pSession->fInHashTable);
1156	RTSpinlockRelease(pDevExt->hSessionHashTabSpinlock);
1157	return VERR_NOT_FOUND;
1158	}
1159	}
1160
1161	pSession->pCommonNextHash = NULL;
1162	pSession->fInHashTable = false;
1163
1164	ASMAtomicDecU32(&pDevExt->cSessions);
1165
1166	/*
1167	* Clear OS specific session pointer if available and do the OS callback.
1168	*/
1169	if (pSession->ppOsSessionPtr)
1170	{
1171	ASMAtomicCmpXchgPtr(pSession->ppOsSessionPtr, NULL, pSession);
1172	pSession->ppOsSessionPtr = NULL;
1173	}
1174
1175	supdrvOSSessionHashTabRemoved(pDevExt, pSession, pvUser);
1176
1177	RTSpinlockRelease(pDevExt->hSessionHashTabSpinlock);
1178
1179	/*
1180	* Drop the reference the hash table had to the session. This shouldn't
1181	* be the last reference!
1182	*/
1183	cRefs = ASMAtomicDecU32(&pSession->cRefs);
1184	Assert(cRefs > 0 && cRefs < _1M);
1185	if (cRefs == 0)
1186	supdrvDestroySession(pDevExt, pSession);
1187
1188	return VINF_SUCCESS;
1189	}
1190
1191
1192	/**
1193	* Looks up the session for the current process in the global hash table or in
1194	* OS specific pointer.
1195	*
1196	* @returns Pointer to the session with a reference that the caller must
1197	* release. If no valid session was found, NULL is returned.
1198	*
1199	* @param pDevExt The device extension.
1200	* @param Process The process ID.
1201	* @param R0Process The ring-0 process handle.
1202	* @param ppOsSessionPtr The OS session pointer if available. If not NULL,
1203	* this is used instead of the hash table. For
1204	* additional safety it must then be equal to the
1205	* SUPDRVSESSION::ppOsSessionPtr member.
1206	* This can be NULL even if the OS has a session
1207	* pointer.
1208	*/
1209	PSUPDRVSESSION VBOXCALL supdrvSessionHashTabLookup(PSUPDRVDEVEXT pDevExt, RTPROCESS Process, RTR0PROCESS R0Process,
1210	PSUPDRVSESSION *ppOsSessionPtr)
1211	{
1212	PSUPDRVSESSION pCur;
1213	unsigned iHash;
1214
1215	/*
1216	* Validate input.
1217	*/
1218	AssertReturn(R0Process != NIL_RTR0PROCESS, NULL);
1219
1220	/*
1221	* Calculate the hash table index and acquire the spinlock.
1222	*/
1223	iHash = SUPDRV_SESSION_HASH(Process);
1224
1225	RTSpinlockAcquire(pDevExt->hSessionHashTabSpinlock);
1226
1227	/*
1228	* If an OS session pointer is provided, always use it.
1229	*/
1230	if (ppOsSessionPtr)
1231	{
1232	pCur = *ppOsSessionPtr;
1233	if ( pCur
1234	&& ( pCur->ppOsSessionPtr != ppOsSessionPtr
1235	\|\| pCur->Process != Process
1236	\|\| pCur->R0Process != R0Process) )
1237	pCur = NULL;
1238	}
1239	else
1240	{
1241	/*
1242	* Otherwise, do the hash table lookup.
1243	*/
1244	pCur = pDevExt->apSessionHashTab[iHash];
1245	while ( pCur
1246	&& ( pCur->Process != Process
1247	\|\| pCur->R0Process != R0Process) )
1248	pCur = pCur->pCommonNextHash;
1249	}
1250
1251	/*
1252	* Retain the session.
1253	*/
1254	if (pCur)
1255	{
1256	uint32_t cRefs = ASMAtomicIncU32(&pCur->cRefs);
1257	NOREF(cRefs);
1258	Assert(cRefs > 1 && cRefs < _1M);
1259	}
1260
1261	RTSpinlockRelease(pDevExt->hSessionHashTabSpinlock);
1262
1263	return pCur;
1264	}
1265
1266
1267	/**
1268	* Retain a session to make sure it doesn't go away while it is in use.
1269	*
1270	* @returns New reference count on success, UINT32_MAX on failure.
1271	* @param pSession Session data.
1272	*/
1273	uint32_t VBOXCALL supdrvSessionRetain(PSUPDRVSESSION pSession)
1274	{
1275	uint32_t cRefs;
1276	AssertPtrReturn(pSession, UINT32_MAX);
1277	AssertReturn(SUP_IS_SESSION_VALID(pSession), UINT32_MAX);
1278
1279	cRefs = ASMAtomicIncU32(&pSession->cRefs);
1280	AssertMsg(cRefs > 1 && cRefs < _1M, ("%#x %p\n", cRefs, pSession));
1281	return cRefs;
1282	}
1283
1284
1285	/**
1286	* Releases a given session.
1287	*
1288	* @returns New reference count on success (0 if closed), UINT32_MAX on failure.
1289	* @param pSession Session data.
1290	*/
1291	uint32_t VBOXCALL supdrvSessionRelease(PSUPDRVSESSION pSession)
1292	{
1293	uint32_t cRefs;
1294	AssertPtrReturn(pSession, UINT32_MAX);
1295	AssertReturn(SUP_IS_SESSION_VALID(pSession), UINT32_MAX);
1296
1297	cRefs = ASMAtomicDecU32(&pSession->cRefs);
1298	AssertMsg(cRefs < _1M, ("%#x %p\n", cRefs, pSession));
1299	if (cRefs == 0)
1300	supdrvDestroySession(pSession->pDevExt, pSession);
1301	return cRefs;
1302	}
1303
1304
1305	/**
1306	* RTHandleTableDestroy callback used by supdrvCleanupSession.
1307	*
1308	* @returns IPRT status code, see SUPR0ObjAddRef.
1309	* @param hHandleTable The handle table handle. Ignored.
1310	* @param pvObj The object pointer.
1311	* @param pvCtx Context, the handle type. Ignored.
1312	* @param pvUser Session pointer.
1313	*/
1314	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser)
1315	{
1316	NOREF(pvCtx);
1317	NOREF(hHandleTable);
1318	return SUPR0ObjAddRefEx(pvObj, (PSUPDRVSESSION)pvUser, true /fNoBlocking/);
1319	}
1320
1321
1322	/**
1323	* RTHandleTableDestroy callback used by supdrvCleanupSession.
1324	*
1325	* @param hHandleTable The handle table handle. Ignored.
1326	* @param h The handle value. Ignored.
1327	* @param pvObj The object pointer.
1328	* @param pvCtx Context, the handle type. Ignored.
1329	* @param pvUser Session pointer.
1330	*/
1331	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser)
1332	{
1333	NOREF(pvCtx);
1334	NOREF(h);
1335	NOREF(hHandleTable);
1336	SUPR0ObjRelease(pvObj, (PSUPDRVSESSION)pvUser);
1337	}
1338
1339
1340	/**
1341	* Fast path I/O Control worker.
1342	*
1343	* @returns VBox status code that should be passed down to ring-3 unchanged.
1344	* @param uIOCtl Function number.
1345	* @param idCpu VMCPU id.
1346	* @param pDevExt Device extention.
1347	* @param pSession Session data.
1348	*/
1349	int VBOXCALL supdrvIOCtlFast(uintptr_t uIOCtl, VMCPUID idCpu, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
1350	{
1351	/*
1352	* We check the two prereqs after doing this only to allow the compiler to optimize things better.
1353	*/
1354	if (RT_LIKELY( RT_VALID_PTR(pSession)
1355	&& pSession->pVM
1356	&& pDevExt->pfnVMMR0EntryFast))
1357	{
1358	switch (uIOCtl)
1359	{
1360	case SUP_IOCTL_FAST_DO_RAW_RUN:
1361	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_RAW_RUN);
1362	break;
1363	case SUP_IOCTL_FAST_DO_HM_RUN:
1364	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_HM_RUN);
1365	break;
1366	case SUP_IOCTL_FAST_DO_NOP:
1367	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_NOP);
1368	break;
1369	default:
1370	return VERR_INTERNAL_ERROR;
1371	}
1372	return VINF_SUCCESS;
1373	}
1374	return VERR_INTERNAL_ERROR;
1375	}
1376
1377
1378	/**
1379	* Helper for supdrvIOCtl. Check if pszStr contains any character of pszChars.
1380	* We would use strpbrk here if this function would be contained in the RedHat kABI white
1381	* list, see http://www.kerneldrivers.org/RHEL5.
1382	*
1383	* @returns 1 if pszStr does contain any character of pszChars, 0 otherwise.
1384	* @param pszStr String to check
1385	* @param pszChars Character set
1386	*/
1387	static int supdrvCheckInvalidChar(const char pszStr, const char pszChars)
1388	{
1389	int chCur;
1390	while ((chCur = *pszStr++) != '\0')
1391	{
1392	int ch;
1393	const char *psz = pszChars;
1394	while ((ch = *psz++) != '\0')
1395	if (ch == chCur)
1396	return 1;
1397
1398	}
1399	return 0;
1400	}
1401
1402
1403
1404	/**
1405	* I/O Control inner worker (tracing reasons).
1406	*
1407	* @returns IPRT status code.
1408	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1409	*
1410	* @param uIOCtl Function number.
1411	* @param pDevExt Device extention.
1412	* @param pSession Session data.
1413	* @param pReqHdr The request header.
1414	*/
1415	static int supdrvIOCtlInnerUnrestricted(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1416	{
1417	/*
1418	* Validation macros
1419	*/
1420	#define REQ_CHECK_SIZES_EX(Name, cbInExpect, cbOutExpect) \
1421	do { \
1422	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect) \|\| pReqHdr->cbOut != (cbOutExpect))) \
1423	{ \
1424	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld. cbOut=%ld expected %ld.\n", \
1425	(long)pReqHdr->cbIn, (long)(cbInExpect), (long)pReqHdr->cbOut, (long)(cbOutExpect))); \
1426	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1427	} \
1428	} while (0)
1429
1430	#define REQ_CHECK_SIZES(Name) REQ_CHECK_SIZES_EX(Name, Name ## _SIZE_IN, Name ## _SIZE_OUT)
1431
1432	#define REQ_CHECK_SIZE_IN(Name, cbInExpect) \
1433	do { \
1434	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect))) \
1435	{ \
1436	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld.\n", \
1437	(long)pReqHdr->cbIn, (long)(cbInExpect))); \
1438	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1439	} \
1440	} while (0)
1441
1442	#define REQ_CHECK_SIZE_OUT(Name, cbOutExpect) \
1443	do { \
1444	if (RT_UNLIKELY(pReqHdr->cbOut != (cbOutExpect))) \
1445	{ \
1446	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbOut=%ld expected %ld.\n", \
1447	(long)pReqHdr->cbOut, (long)(cbOutExpect))); \
1448	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1449	} \
1450	} while (0)
1451
1452	#define REQ_CHECK_EXPR(Name, expr) \
1453	do { \
1454	if (RT_UNLIKELY(!(expr))) \
1455	{ \
1456	OSDBGPRINT(( #Name ": %s\n", #expr)); \
1457	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1458	} \
1459	} while (0)
1460
1461	#define REQ_CHECK_EXPR_FMT(expr, fmt) \
1462	do { \
1463	if (RT_UNLIKELY(!(expr))) \
1464	{ \
1465	OSDBGPRINT( fmt ); \
1466	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1467	} \
1468	} while (0)
1469
1470	/*
1471	* The switch.
1472	*/
1473	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
1474	{
1475	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
1476	{
1477	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
1478	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
1479	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
1480	{
1481	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
1482	pReq->Hdr.rc = VERR_INVALID_MAGIC;
1483	return 0;
1484	}
1485
1486	#if 0
1487	/*
1488	* Call out to the OS specific code and let it do permission checks on the
1489	* client process.
1490	*/
1491	if (!supdrvOSValidateClientProcess(pDevExt, pSession))
1492	{
1493	pReq->u.Out.u32Cookie = 0xffffffff;
1494	pReq->u.Out.u32SessionCookie = 0xffffffff;
1495	pReq->u.Out.u32SessionVersion = 0xffffffff;
1496	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1497	pReq->u.Out.pSession = NULL;
1498	pReq->u.Out.cFunctions = 0;
1499	pReq->Hdr.rc = VERR_PERMISSION_DENIED;
1500	return 0;
1501	}
1502	#endif
1503
1504	/*
1505	* Match the version.
1506	* The current logic is very simple, match the major interface version.
1507	*/
1508	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
1509	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
1510	{
1511	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1512	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
1513	pReq->u.Out.u32Cookie = 0xffffffff;
1514	pReq->u.Out.u32SessionCookie = 0xffffffff;
1515	pReq->u.Out.u32SessionVersion = 0xffffffff;
1516	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1517	pReq->u.Out.pSession = NULL;
1518	pReq->u.Out.cFunctions = 0;
1519	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1520	return 0;
1521	}
1522
1523	/*
1524	* Fill in return data and be gone.
1525	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
1526	* u32SessionVersion <= u32ReqVersion!
1527	*/
1528	/** @todo Somehow validate the client and negotiate a secure cookie... */
1529	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
1530	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
1531	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
1532	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1533	pReq->u.Out.pSession = pSession;
1534	pReq->u.Out.cFunctions = sizeof(g_aFunctions) / sizeof(g_aFunctions[0]);
1535	pReq->Hdr.rc = VINF_SUCCESS;
1536	return 0;
1537	}
1538
1539	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_QUERY_FUNCS(0)):
1540	{
1541	/* validate */
1542	PSUPQUERYFUNCS pReq = (PSUPQUERYFUNCS)pReqHdr;
1543	REQ_CHECK_SIZES_EX(SUP_IOCTL_QUERY_FUNCS, SUP_IOCTL_QUERY_FUNCS_SIZE_IN, SUP_IOCTL_QUERY_FUNCS_SIZE_OUT(RT_ELEMENTS(g_aFunctions)));
1544
1545	/* execute */
1546	pReq->u.Out.cFunctions = RT_ELEMENTS(g_aFunctions);
1547	memcpy(&pReq->u.Out.aFunctions[0], g_aFunctions, sizeof(g_aFunctions));
1548	pReq->Hdr.rc = VINF_SUCCESS;
1549	return 0;
1550	}
1551
1552	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_LOCK):
1553	{
1554	/* validate */
1555	PSUPPAGELOCK pReq = (PSUPPAGELOCK)pReqHdr;
1556	REQ_CHECK_SIZE_IN(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_IN);
1557	REQ_CHECK_SIZE_OUT(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_OUT(pReq->u.In.cPages));
1558	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.cPages > 0);
1559	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.pvR3 >= PAGE_SIZE);
1560
1561	/* execute */
1562	pReq->Hdr.rc = SUPR0LockMem(pSession, pReq->u.In.pvR3, pReq->u.In.cPages, &pReq->u.Out.aPages[0]);
1563	if (RT_FAILURE(pReq->Hdr.rc))
1564	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1565	return 0;
1566	}
1567
1568	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_UNLOCK):
1569	{
1570	/* validate */
1571	PSUPPAGEUNLOCK pReq = (PSUPPAGEUNLOCK)pReqHdr;
1572	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_UNLOCK);
1573
1574	/* execute */
1575	pReq->Hdr.rc = SUPR0UnlockMem(pSession, pReq->u.In.pvR3);
1576	return 0;
1577	}
1578
1579	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_ALLOC):
1580	{
1581	/* validate */
1582	PSUPCONTALLOC pReq = (PSUPCONTALLOC)pReqHdr;
1583	REQ_CHECK_SIZES(SUP_IOCTL_CONT_ALLOC);
1584
1585	/* execute */
1586	pReq->Hdr.rc = SUPR0ContAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.HCPhys);
1587	if (RT_FAILURE(pReq->Hdr.rc))
1588	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1589	return 0;
1590	}
1591
1592	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_FREE):
1593	{
1594	/* validate */
1595	PSUPCONTFREE pReq = (PSUPCONTFREE)pReqHdr;
1596	REQ_CHECK_SIZES(SUP_IOCTL_CONT_FREE);
1597
1598	/* execute */
1599	pReq->Hdr.rc = SUPR0ContFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1600	return 0;
1601	}
1602
1603	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_OPEN):
1604	{
1605	/* validate */
1606	PSUPLDROPEN pReq = (PSUPLDROPEN)pReqHdr;
1607	REQ_CHECK_SIZES(SUP_IOCTL_LDR_OPEN);
1608	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs > 0);
1609	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs < 16*_1M);
1610	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1611	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1612	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits < pReq->u.In.cbImageWithTabs);
1613	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.szName[0]);
1614	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1615	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, !supdrvCheckInvalidChar(pReq->u.In.szName, ";:()[]{}/\\\|&*%#@!~`\"'"));
1616	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szFilename, sizeof(pReq->u.In.szFilename)));
1617
1618	/* execute */
1619	pReq->Hdr.rc = supdrvIOCtl_LdrOpen(pDevExt, pSession, pReq);
1620	return 0;
1621	}
1622
1623	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOAD):
1624	{
1625	/* validate */
1626	PSUPLDRLOAD pReq = (PSUPLDRLOAD)pReqHdr;
1627	REQ_CHECK_EXPR(Name, pReq->Hdr.cbIn >= sizeof(*pReq));
1628	REQ_CHECK_SIZES_EX(SUP_IOCTL_LDR_LOAD, SUP_IOCTL_LDR_LOAD_SIZE_IN(pReq->u.In.cbImageWithTabs), SUP_IOCTL_LDR_LOAD_SIZE_OUT);
1629	REQ_CHECK_EXPR(SUP_IOCTL_LDR_LOAD, pReq->u.In.cSymbols <= 16384);
1630	REQ_CHECK_EXPR_FMT( !pReq->u.In.cSymbols
1631	\|\| ( pReq->u.In.offSymbols < pReq->u.In.cbImageWithTabs
1632	&& pReq->u.In.offSymbols + pReq->u.In.cSymbols * sizeof(SUPLDRSYM) <= pReq->u.In.cbImageWithTabs),
1633	("SUP_IOCTL_LDR_LOAD: offSymbols=%#lx cSymbols=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offSymbols,
1634	(long)pReq->u.In.cSymbols, (long)pReq->u.In.cbImageWithTabs));
1635	REQ_CHECK_EXPR_FMT( !pReq->u.In.cbStrTab
1636	\|\| ( pReq->u.In.offStrTab < pReq->u.In.cbImageWithTabs
1637	&& pReq->u.In.offStrTab + pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs
1638	&& pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs),
1639	("SUP_IOCTL_LDR_LOAD: offStrTab=%#lx cbStrTab=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offStrTab,
1640	(long)pReq->u.In.cbStrTab, (long)pReq->u.In.cbImageWithTabs));
1641
1642	if (pReq->u.In.cSymbols)
1643	{
1644	uint32_t i;
1645	PSUPLDRSYM paSyms = (PSUPLDRSYM)&pReq->u.In.abImage[pReq->u.In.offSymbols];
1646	for (i = 0; i < pReq->u.In.cSymbols; i++)
1647	{
1648	REQ_CHECK_EXPR_FMT(paSyms[i].offSymbol < pReq->u.In.cbImageWithTabs,
1649	("SUP_IOCTL_LDR_LOAD: sym #%ld: symb off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offSymbol, (long)pReq->u.In.cbImageWithTabs));
1650	REQ_CHECK_EXPR_FMT(paSyms[i].offName < pReq->u.In.cbStrTab,
1651	("SUP_IOCTL_LDR_LOAD: sym #%ld: name off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1652	REQ_CHECK_EXPR_FMT(RTStrEnd((char const *)&pReq->u.In.abImage[pReq->u.In.offStrTab + paSyms[i].offName],
1653	pReq->u.In.cbStrTab - paSyms[i].offName),
1654	("SUP_IOCTL_LDR_LOAD: sym #%ld: unterminated name! (%#lx / %#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1655	}
1656	}
1657
1658	/* execute */
1659	pReq->Hdr.rc = supdrvIOCtl_LdrLoad(pDevExt, pSession, pReq);
1660	return 0;
1661	}
1662
1663	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_FREE):
1664	{
1665	/* validate */
1666	PSUPLDRFREE pReq = (PSUPLDRFREE)pReqHdr;
1667	REQ_CHECK_SIZES(SUP_IOCTL_LDR_FREE);
1668
1669	/* execute */
1670	pReq->Hdr.rc = supdrvIOCtl_LdrFree(pDevExt, pSession, pReq);
1671	return 0;
1672	}
1673
1674	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_GET_SYMBOL):
1675	{
1676	/* validate */
1677	PSUPLDRGETSYMBOL pReq = (PSUPLDRGETSYMBOL)pReqHdr;
1678	REQ_CHECK_SIZES(SUP_IOCTL_LDR_GET_SYMBOL);
1679	REQ_CHECK_EXPR(SUP_IOCTL_LDR_GET_SYMBOL, RTStrEnd(pReq->u.In.szSymbol, sizeof(pReq->u.In.szSymbol)));
1680
1681	/* execute */
1682	pReq->Hdr.rc = supdrvIOCtl_LdrGetSymbol(pDevExt, pSession, pReq);
1683	return 0;
1684	}
1685
1686	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0(0)):
1687	{
1688	/* validate */
1689	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1690	Log4(("SUP_IOCTL_CALL_VMMR0: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1691	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1692
1693	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_VMMR0_SIZE(0))
1694	{
1695	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(0), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(0));
1696
1697	/* execute */
1698	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1699	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg, pSession);
1700	else
1701	pReq->Hdr.rc = VERR_WRONG_ORDER;
1702	}
1703	else
1704	{
1705	PSUPVMMR0REQHDR pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1706	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR)),
1707	("SUP_IOCTL_CALL_VMMR0: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR))));
1708	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1709	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(pVMMReq->cbReq));
1710
1711	/* execute */
1712	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1713	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1714	else
1715	pReq->Hdr.rc = VERR_WRONG_ORDER;
1716	}
1717
1718	if ( RT_FAILURE(pReq->Hdr.rc)
1719	&& pReq->Hdr.rc != VERR_INTERRUPTED
1720	&& pReq->Hdr.rc != VERR_TIMEOUT)
1721	Log(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1722	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1723	else
1724	Log4(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1725	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1726	return 0;
1727	}
1728
1729	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0_BIG):
1730	{
1731	/* validate */
1732	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1733	PSUPVMMR0REQHDR pVMMReq;
1734	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1735	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1736
1737	pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1738	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR)),
1739	("SUP_IOCTL_CALL_VMMR0_BIG: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR))));
1740	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0_BIG, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1741	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));
1742
1743	/* execute */
1744	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1745	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1746	else
1747	pReq->Hdr.rc = VERR_WRONG_ORDER;
1748
1749	if ( RT_FAILURE(pReq->Hdr.rc)
1750	&& pReq->Hdr.rc != VERR_INTERRUPTED
1751	&& pReq->Hdr.rc != VERR_TIMEOUT)
1752	Log(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1753	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1754	else
1755	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1756	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1757	return 0;
1758	}
1759
1760	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_PAGING_MODE):
1761	{
1762	/* validate */
1763	PSUPGETPAGINGMODE pReq = (PSUPGETPAGINGMODE)pReqHdr;
1764	REQ_CHECK_SIZES(SUP_IOCTL_GET_PAGING_MODE);
1765
1766	/* execute */
1767	pReq->Hdr.rc = VINF_SUCCESS;
1768	pReq->u.Out.enmMode = SUPR0GetPagingMode();
1769	return 0;
1770	}
1771
1772	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_ALLOC):
1773	{
1774	/* validate */
1775	PSUPLOWALLOC pReq = (PSUPLOWALLOC)pReqHdr;
1776	REQ_CHECK_EXPR(SUP_IOCTL_LOW_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_LOW_ALLOC_SIZE_IN);
1777	REQ_CHECK_SIZES_EX(SUP_IOCTL_LOW_ALLOC, SUP_IOCTL_LOW_ALLOC_SIZE_IN, SUP_IOCTL_LOW_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1778
1779	/* execute */
1780	pReq->Hdr.rc = SUPR0LowAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1781	if (RT_FAILURE(pReq->Hdr.rc))
1782	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1783	return 0;
1784	}
1785
1786	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_FREE):
1787	{
1788	/* validate */
1789	PSUPLOWFREE pReq = (PSUPLOWFREE)pReqHdr;
1790	REQ_CHECK_SIZES(SUP_IOCTL_LOW_FREE);
1791
1792	/* execute */
1793	pReq->Hdr.rc = SUPR0LowFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1794	return 0;
1795	}
1796
1797	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_MAP):
1798	{
1799	/* validate */
1800	PSUPGIPMAP pReq = (PSUPGIPMAP)pReqHdr;
1801	REQ_CHECK_SIZES(SUP_IOCTL_GIP_MAP);
1802
1803	/* execute */
1804	pReq->Hdr.rc = SUPR0GipMap(pSession, &pReq->u.Out.pGipR3, &pReq->u.Out.HCPhysGip);
1805	if (RT_SUCCESS(pReq->Hdr.rc))
1806	pReq->u.Out.pGipR0 = pDevExt->pGip;
1807	return 0;
1808	}
1809
1810	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_UNMAP):
1811	{
1812	/* validate */
1813	PSUPGIPUNMAP pReq = (PSUPGIPUNMAP)pReqHdr;
1814	REQ_CHECK_SIZES(SUP_IOCTL_GIP_UNMAP);
1815
1816	/* execute */
1817	pReq->Hdr.rc = SUPR0GipUnmap(pSession);
1818	return 0;
1819	}
1820
1821	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SET_VM_FOR_FAST):
1822	{
1823	/* validate */
1824	PSUPSETVMFORFAST pReq = (PSUPSETVMFORFAST)pReqHdr;
1825	REQ_CHECK_SIZES(SUP_IOCTL_SET_VM_FOR_FAST);
1826	REQ_CHECK_EXPR_FMT( !pReq->u.In.pVMR0
1827	\|\| ( VALID_PTR(pReq->u.In.pVMR0)
1828	&& !((uintptr_t)pReq->u.In.pVMR0 & (PAGE_SIZE - 1))),
1829	("SUP_IOCTL_SET_VM_FOR_FAST: pVMR0=%p!\n", pReq->u.In.pVMR0));
1830	/* execute */
1831	pSession->pVM = pReq->u.In.pVMR0;
1832	pReq->Hdr.rc = VINF_SUCCESS;
1833	return 0;
1834	}
1835
1836	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_ALLOC_EX):
1837	{
1838	/* validate */
1839	PSUPPAGEALLOCEX pReq = (PSUPPAGEALLOCEX)pReqHdr;
1840	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC_EX, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_EX_SIZE_IN);
1841	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));
1842	REQ_CHECK_EXPR_FMT(pReq->u.In.fKernelMapping \|\| pReq->u.In.fUserMapping,
1843	("SUP_IOCTL_PAGE_ALLOC_EX: No mapping requested!\n"));
1844	REQ_CHECK_EXPR_FMT(pReq->u.In.fUserMapping,
1845	("SUP_IOCTL_PAGE_ALLOC_EX: Must have user mapping!\n"));
1846	REQ_CHECK_EXPR_FMT(!pReq->u.In.fReserved0 && !pReq->u.In.fReserved1,
1847	("SUP_IOCTL_PAGE_ALLOC_EX: fReserved0=%d fReserved1=%d\n", pReq->u.In.fReserved0, pReq->u.In.fReserved1));
1848
1849	/* execute */
1850	pReq->Hdr.rc = SUPR0PageAllocEx(pSession, pReq->u.In.cPages, 0 /* fFlags */,
1851	pReq->u.In.fUserMapping ? &pReq->u.Out.pvR3 : NULL,
1852	pReq->u.In.fKernelMapping ? &pReq->u.Out.pvR0 : NULL,
1853	&pReq->u.Out.aPages[0]);
1854	if (RT_FAILURE(pReq->Hdr.rc))
1855	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1856	return 0;
1857	}
1858
1859	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_MAP_KERNEL):
1860	{
1861	/* validate */
1862	PSUPPAGEMAPKERNEL pReq = (PSUPPAGEMAPKERNEL)pReqHdr;
1863	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_MAP_KERNEL);
1864	REQ_CHECK_EXPR_FMT(!pReq->u.In.fFlags, ("SUP_IOCTL_PAGE_MAP_KERNEL: fFlags=%#x! MBZ\n", pReq->u.In.fFlags));
1865	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_MAP_KERNEL: offSub=%#x\n", pReq->u.In.offSub));
1866	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1867	("SUP_IOCTL_PAGE_MAP_KERNEL: cbSub=%#x\n", pReq->u.In.cbSub));
1868
1869	/* execute */
1870	pReq->Hdr.rc = SUPR0PageMapKernel(pSession, pReq->u.In.pvR3, pReq->u.In.offSub, pReq->u.In.cbSub,
1871	pReq->u.In.fFlags, &pReq->u.Out.pvR0);
1872	if (RT_FAILURE(pReq->Hdr.rc))
1873	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1874	return 0;
1875	}
1876
1877	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_PROTECT):
1878	{
1879	/* validate */
1880	PSUPPAGEPROTECT pReq = (PSUPPAGEPROTECT)pReqHdr;
1881	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_PROTECT);
1882	REQ_CHECK_EXPR_FMT(!(pReq->u.In.fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)),
1883	("SUP_IOCTL_PAGE_PROTECT: fProt=%#x!\n", pReq->u.In.fProt));
1884	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_PROTECT: offSub=%#x\n", pReq->u.In.offSub));
1885	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1886	("SUP_IOCTL_PAGE_PROTECT: cbSub=%#x\n", pReq->u.In.cbSub));
1887
1888	/* execute */
1889	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);
1890	return 0;
1891	}
1892
1893	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_FREE):
1894	{
1895	/* validate */
1896	PSUPPAGEFREE pReq = (PSUPPAGEFREE)pReqHdr;
1897	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_FREE);
1898
1899	/* execute */
1900	pReq->Hdr.rc = SUPR0PageFree(pSession, pReq->u.In.pvR3);
1901	return 0;
1902	}
1903
1904	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_SERVICE(0)):
1905	{
1906	/* validate */
1907	PSUPCALLSERVICE pReq = (PSUPCALLSERVICE)pReqHdr;
1908	Log4(("SUP_IOCTL_CALL_SERVICE: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1909	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1910
1911	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
1912	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(0), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(0));
1913	else
1914	{
1915	PSUPR0SERVICEREQHDR pSrvReq = (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0];
1916	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR)),
1917	("SUP_IOCTL_CALL_SERVICE: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR))));
1918	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, pSrvReq->u32Magic == SUPR0SERVICEREQHDR_MAGIC);
1919	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(pSrvReq->cbReq), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(pSrvReq->cbReq));
1920	}
1921	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1922
1923	/* execute */
1924	pReq->Hdr.rc = supdrvIOCtl_CallServiceModule(pDevExt, pSession, pReq);
1925	return 0;
1926	}
1927
1928	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOGGER_SETTINGS(0)):
1929	{
1930	/* validate */
1931	PSUPLOGGERSETTINGS pReq = (PSUPLOGGERSETTINGS)pReqHdr;
1932	size_t cbStrTab;
1933	REQ_CHECK_SIZE_OUT(SUP_IOCTL_LOGGER_SETTINGS, SUP_IOCTL_LOGGER_SETTINGS_SIZE_OUT);
1934	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->Hdr.cbIn >= SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(1));
1935	cbStrTab = pReq->Hdr.cbIn - SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(0);
1936	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offGroups < cbStrTab);
1937	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offFlags < cbStrTab);
1938	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offDestination < cbStrTab);
1939	REQ_CHECK_EXPR_FMT(pReq->u.In.szStrings[cbStrTab - 1] == '\0',
1940	("SUP_IOCTL_LOGGER_SETTINGS: cbIn=%#x cbStrTab=%#zx LastChar=%d\n",
1941	pReq->Hdr.cbIn, cbStrTab, pReq->u.In.szStrings[cbStrTab - 1]));
1942	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhich <= SUPLOGGERSETTINGS_WHICH_RELEASE);
1943	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhat <= SUPLOGGERSETTINGS_WHAT_DESTROY);
1944
1945	/* execute */
1946	pReq->Hdr.rc = supdrvIOCtl_LoggerSettings(pDevExt, pSession, pReq);
1947	return 0;
1948	}
1949
1950	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP2):
1951	{
1952	/* validate */
1953	PSUPSEMOP2 pReq = (PSUPSEMOP2)pReqHdr;
1954	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP2, SUP_IOCTL_SEM_OP2_SIZE_IN, SUP_IOCTL_SEM_OP2_SIZE_OUT);
1955	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP2, pReq->u.In.uReserved == 0);
1956
1957	/* execute */
1958	switch (pReq->u.In.uType)
1959	{
1960	case SUP_SEM_TYPE_EVENT:
1961	{
1962	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1963	switch (pReq->u.In.uOp)
1964	{
1965	case SUPSEMOP2_WAIT_MS_REL:
1966	pReq->Hdr.rc = SUPSemEventWaitNoResume(pSession, hEvent, pReq->u.In.uArg.cRelMsTimeout);
1967	break;
1968	case SUPSEMOP2_WAIT_NS_ABS:
1969	pReq->Hdr.rc = SUPSemEventWaitNsAbsIntr(pSession, hEvent, pReq->u.In.uArg.uAbsNsTimeout);
1970	break;
1971	case SUPSEMOP2_WAIT_NS_REL:
1972	pReq->Hdr.rc = SUPSemEventWaitNsRelIntr(pSession, hEvent, pReq->u.In.uArg.cRelNsTimeout);
1973	break;
1974	case SUPSEMOP2_SIGNAL:
1975	pReq->Hdr.rc = SUPSemEventSignal(pSession, hEvent);
1976	break;
1977	case SUPSEMOP2_CLOSE:
1978	pReq->Hdr.rc = SUPSemEventClose(pSession, hEvent);
1979	break;
1980	case SUPSEMOP2_RESET:
1981	default:
1982	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1983	break;
1984	}
1985	break;
1986	}
1987
1988	case SUP_SEM_TYPE_EVENT_MULTI:
1989	{
1990	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1991	switch (pReq->u.In.uOp)
1992	{
1993	case SUPSEMOP2_WAIT_MS_REL:
1994	pReq->Hdr.rc = SUPSemEventMultiWaitNoResume(pSession, hEventMulti, pReq->u.In.uArg.cRelMsTimeout);
1995	break;
1996	case SUPSEMOP2_WAIT_NS_ABS:
1997	pReq->Hdr.rc = SUPSemEventMultiWaitNsAbsIntr(pSession, hEventMulti, pReq->u.In.uArg.uAbsNsTimeout);
1998	break;
1999	case SUPSEMOP2_WAIT_NS_REL:
2000	pReq->Hdr.rc = SUPSemEventMultiWaitNsRelIntr(pSession, hEventMulti, pReq->u.In.uArg.cRelNsTimeout);
2001	break;
2002	case SUPSEMOP2_SIGNAL:
2003	pReq->Hdr.rc = SUPSemEventMultiSignal(pSession, hEventMulti);
2004	break;
2005	case SUPSEMOP2_CLOSE:
2006	pReq->Hdr.rc = SUPSemEventMultiClose(pSession, hEventMulti);
2007	break;
2008	case SUPSEMOP2_RESET:
2009	pReq->Hdr.rc = SUPSemEventMultiReset(pSession, hEventMulti);
2010	break;
2011	default:
2012	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
2013	break;
2014	}
2015	break;
2016	}
2017
2018	default:
2019	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
2020	break;
2021	}
2022	return 0;
2023	}
2024
2025	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP3):
2026	{
2027	/* validate */
2028	PSUPSEMOP3 pReq = (PSUPSEMOP3)pReqHdr;
2029	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP3, SUP_IOCTL_SEM_OP3_SIZE_IN, SUP_IOCTL_SEM_OP3_SIZE_OUT);
2030	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, pReq->u.In.u32Reserved == 0 && pReq->u.In.u64Reserved == 0);
2031
2032	/* execute */
2033	switch (pReq->u.In.uType)
2034	{
2035	case SUP_SEM_TYPE_EVENT:
2036	{
2037	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
2038	switch (pReq->u.In.uOp)
2039	{
2040	case SUPSEMOP3_CREATE:
2041	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
2042	pReq->Hdr.rc = SUPSemEventCreate(pSession, &hEvent);
2043	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEvent;
2044	break;
2045	case SUPSEMOP3_GET_RESOLUTION:
2046	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
2047	pReq->Hdr.rc = VINF_SUCCESS;
2048	pReq->Hdr.cbOut = sizeof(*pReq);
2049	pReq->u.Out.cNsResolution = SUPSemEventGetResolution(pSession);
2050	break;
2051	default:
2052	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
2053	break;
2054	}
2055	break;
2056	}
2057
2058	case SUP_SEM_TYPE_EVENT_MULTI:
2059	{
2060	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
2061	switch (pReq->u.In.uOp)
2062	{
2063	case SUPSEMOP3_CREATE:
2064	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
2065	pReq->Hdr.rc = SUPSemEventMultiCreate(pSession, &hEventMulti);
2066	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEventMulti;
2067	break;
2068	case SUPSEMOP3_GET_RESOLUTION:
2069	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
2070	pReq->Hdr.rc = VINF_SUCCESS;
2071	pReq->u.Out.cNsResolution = SUPSemEventMultiGetResolution(pSession);
2072	break;
2073	default:
2074	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
2075	break;
2076	}
2077	break;
2078	}
2079
2080	default:
2081	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
2082	break;
2083	}
2084	return 0;
2085	}
2086
2087	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
2088	{
2089	/* validate */
2090	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
2091	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
2092
2093	/* execute */
2094	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.Caps);
2095	if (RT_FAILURE(pReq->Hdr.rc))
2096	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
2097	return 0;
2098	}
2099
2100	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_OPEN):
2101	{
2102	/* validate */
2103	PSUPTRACEROPEN pReq = (PSUPTRACEROPEN)pReqHdr;
2104	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_OPEN);
2105
2106	/* execute */
2107	pReq->Hdr.rc = supdrvIOCtl_TracerOpen(pDevExt, pSession, pReq->u.In.uCookie, pReq->u.In.uArg);
2108	return 0;
2109	}
2110
2111	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_CLOSE):
2112	{
2113	/* validate */
2114	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_CLOSE);
2115
2116	/* execute */
2117	pReqHdr->rc = supdrvIOCtl_TracerClose(pDevExt, pSession);
2118	return 0;
2119	}
2120
2121	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_IOCTL):
2122	{
2123	/* validate */
2124	PSUPTRACERIOCTL pReq = (PSUPTRACERIOCTL)pReqHdr;
2125	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_IOCTL);
2126
2127	/* execute */
2128	pReqHdr->rc = supdrvIOCtl_TracerIOCtl(pDevExt, pSession, pReq->u.In.uCmd, pReq->u.In.uArg, &pReq->u.Out.iRetVal);
2129	return 0;
2130	}
2131
2132	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_REG):
2133	{
2134	/* validate */
2135	PSUPTRACERUMODREG pReq = (PSUPTRACERUMODREG)pReqHdr;
2136	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_REG);
2137	if (!RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)))
2138	return VERR_INVALID_PARAMETER;
2139
2140	/* execute */
2141	pReqHdr->rc = supdrvIOCtl_TracerUmodRegister(pDevExt, pSession,
2142	pReq->u.In.R3PtrVtgHdr, pReq->u.In.uVtgHdrAddr,
2143	pReq->u.In.R3PtrStrTab, pReq->u.In.cbStrTab,
2144	pReq->u.In.szName, pReq->u.In.fFlags);
2145	return 0;
2146	}
2147
2148	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_DEREG):
2149	{
2150	/* validate */
2151	PSUPTRACERUMODDEREG pReq = (PSUPTRACERUMODDEREG)pReqHdr;
2152	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_DEREG);
2153
2154	/* execute */
2155	pReqHdr->rc = supdrvIOCtl_TracerUmodDeregister(pDevExt, pSession, pReq->u.In.pVtgHdr);
2156	return 0;
2157	}
2158
2159	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_FIRE_PROBE):
2160	{
2161	/* validate */
2162	PSUPTRACERUMODFIREPROBE pReq = (PSUPTRACERUMODFIREPROBE)pReqHdr;
2163	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_FIRE_PROBE);
2164
2165	supdrvIOCtl_TracerUmodProbeFire(pDevExt, pSession, &pReq->u.In);
2166	pReqHdr->rc = VINF_SUCCESS;
2167	return 0;
2168	}
2169
2170	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_MSR_PROBER):
2171	{
2172	/* validate */
2173	PSUPMSRPROBER pReq = (PSUPMSRPROBER)pReqHdr;
2174	REQ_CHECK_SIZES(SUP_IOCTL_MSR_PROBER);
2175	REQ_CHECK_EXPR(SUP_IOCTL_MSR_PROBER,
2176	pReq->u.In.enmOp > SUPMSRPROBEROP_INVALID && pReq->u.In.enmOp < SUPMSRPROBEROP_END);
2177
2178	pReqHdr->rc = supdrvIOCtl_MsrProber(pDevExt, pReq);
2179	return 0;
2180	}
2181
2182	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_RESUME_SUSPENDED_KBDS):
2183	{
2184	/* validate */
2185	REQ_CHECK_SIZES(SUP_IOCTL_RESUME_SUSPENDED_KBDS);
2186
2187	pReqHdr->rc = supdrvIOCtl_ResumeSuspendedKbds();
2188	return 0;
2189	}
2190
2191	default:
2192	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
2193	break;
2194	}
2195	return VERR_GENERAL_FAILURE;
2196	}
2197
2198
2199	/**
2200	* I/O Control inner worker for the restricted operations.
2201	*
2202	* @returns IPRT status code.
2203	* @retval VERR_INVALID_PARAMETER if the request is invalid.
2204	*
2205	* @param uIOCtl Function number.
2206	* @param pDevExt Device extention.
2207	* @param pSession Session data.
2208	* @param pReqHdr The request header.
2209	*/
2210	static int supdrvIOCtlInnerRestricted(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
2211	{
2212	/*
2213	* The switch.
2214	*/
2215	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
2216	{
2217	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
2218	{
2219	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
2220	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
2221	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
2222	{
2223	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
2224	pReq->Hdr.rc = VERR_INVALID_MAGIC;
2225	return 0;
2226	}
2227
2228	/*
2229	* Match the version.
2230	* The current logic is very simple, match the major interface version.
2231	*/
2232	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
2233	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
2234	{
2235	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
2236	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
2237	pReq->u.Out.u32Cookie = 0xffffffff;
2238	pReq->u.Out.u32SessionCookie = 0xffffffff;
2239	pReq->u.Out.u32SessionVersion = 0xffffffff;
2240	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
2241	pReq->u.Out.pSession = NULL;
2242	pReq->u.Out.cFunctions = 0;
2243	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
2244	return 0;
2245	}
2246
2247	/*
2248	* Fill in return data and be gone.
2249	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
2250	* u32SessionVersion <= u32ReqVersion!
2251	*/
2252	/** @todo Somehow validate the client and negotiate a secure cookie... */
2253	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
2254	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
2255	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
2256	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
2257	pReq->u.Out.pSession = pSession;
2258	pReq->u.Out.cFunctions = 0;
2259	pReq->Hdr.rc = VINF_SUCCESS;
2260	return 0;
2261	}
2262
2263	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
2264	{
2265	/* validate */
2266	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
2267	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
2268
2269	/* execute */
2270	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.Caps);
2271	if (RT_FAILURE(pReq->Hdr.rc))
2272	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
2273	return 0;
2274	}
2275
2276	default:
2277	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
2278	break;
2279	}
2280	return VERR_GENERAL_FAILURE;
2281	}
2282
2283
2284	/**
2285	* I/O Control worker.
2286	*
2287	* @returns IPRT status code.
2288	* @retval VERR_INVALID_PARAMETER if the request is invalid.
2289	*
2290	* @param uIOCtl Function number.
2291	* @param pDevExt Device extention.
2292	* @param pSession Session data.
2293	* @param pReqHdr The request header.
2294	*/
2295	int VBOXCALL supdrvIOCtl(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
2296	{
2297	int rc;
2298	VBOXDRV_IOCTL_ENTRY(pSession, uIOCtl, pReqHdr);
2299
2300	/*
2301	* Validate the request.
2302	*/
2303	/* this first check could probably be omitted as its also done by the OS specific code... */
2304	if (RT_UNLIKELY( (pReqHdr->fFlags & SUPREQHDR_FLAGS_MAGIC_MASK) != SUPREQHDR_FLAGS_MAGIC
2305	\|\| pReqHdr->cbIn < sizeof(*pReqHdr)
2306	\|\| pReqHdr->cbOut < sizeof(*pReqHdr)))
2307	{
2308	OSDBGPRINT(("vboxdrv: Bad ioctl request header; cbIn=%#lx cbOut=%#lx fFlags=%#lx\n",
2309	(long)pReqHdr->cbIn, (long)pReqHdr->cbOut, (long)pReqHdr->fFlags));
2310	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
2311	return VERR_INVALID_PARAMETER;
2312	}
2313	if (RT_UNLIKELY(!RT_VALID_PTR(pSession)))
2314	{
2315	OSDBGPRINT(("vboxdrv: Invalid pSession value %p (ioctl=%p)\n", pSession, (void *)uIOCtl));
2316	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
2317	return VERR_INVALID_PARAMETER;
2318	}
2319	if (RT_UNLIKELY(uIOCtl == SUP_IOCTL_COOKIE))
2320	{
2321	if (pReqHdr->u32Cookie != SUPCOOKIE_INITIAL_COOKIE)
2322	{
2323	OSDBGPRINT(("SUP_IOCTL_COOKIE: bad cookie %#lx\n", (long)pReqHdr->u32Cookie));
2324	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
2325	return VERR_INVALID_PARAMETER;
2326	}
2327	}
2328	else if (RT_UNLIKELY( pReqHdr->u32Cookie != pDevExt->u32Cookie
2329	\|\| pReqHdr->u32SessionCookie != pSession->u32Cookie))
2330	{
2331	OSDBGPRINT(("vboxdrv: bad cookie %#lx / %#lx.\n", (long)pReqHdr->u32Cookie, (long)pReqHdr->u32SessionCookie));
2332	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
2333	return VERR_INVALID_PARAMETER;
2334	}
2335
2336	/*
2337	* Hand it to an inner function to avoid lots of unnecessary return tracepoints.
2338	*/
2339	if (pSession->fUnrestricted)
2340	rc = supdrvIOCtlInnerUnrestricted(uIOCtl, pDevExt, pSession, pReqHdr);
2341	else
2342	rc = supdrvIOCtlInnerRestricted(uIOCtl, pDevExt, pSession, pReqHdr);
2343
2344	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, pReqHdr->rc, rc);
2345	return rc;
2346	}
2347
2348
2349	/**
2350	* Inter-Driver Communication (IDC) worker.
2351	*
2352	* @returns VBox status code.
2353	* @retval VINF_SUCCESS on success.
2354	* @retval VERR_INVALID_PARAMETER if the request is invalid.
2355	* @retval VERR_NOT_SUPPORTED if the request isn't supported.
2356	*
2357	* @param uReq The request (function) code.
2358	* @param pDevExt Device extention.
2359	* @param pSession Session data.
2360	* @param pReqHdr The request header.
2361	*/
2362	int VBOXCALL supdrvIDC(uintptr_t uReq, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQHDR pReqHdr)
2363	{
2364	/*
2365	* The OS specific code has already validated the pSession
2366	* pointer, and the request size being greater or equal to
2367	* size of the header.
2368	*
2369	* So, just check that pSession is a kernel context session.
2370	*/
2371	if (RT_UNLIKELY( pSession
2372	&& pSession->R0Process != NIL_RTR0PROCESS))
2373	return VERR_INVALID_PARAMETER;
2374
2375	/*
2376	* Validation macro.
2377	*/
2378	#define REQ_CHECK_IDC_SIZE(Name, cbExpect) \
2379	do { \
2380	if (RT_UNLIKELY(pReqHdr->cb != (cbExpect))) \
2381	{ \
2382	OSDBGPRINT(( #Name ": Invalid input/output sizes. cb=%ld expected %ld.\n", \
2383	(long)pReqHdr->cb, (long)(cbExpect))); \
2384	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
2385	} \
2386	} while (0)
2387
2388	switch (uReq)
2389	{
2390	case SUPDRV_IDC_REQ_CONNECT:
2391	{
2392	PSUPDRVIDCREQCONNECT pReq = (PSUPDRVIDCREQCONNECT)pReqHdr;
2393	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_CONNECT, sizeof(*pReq));
2394
2395	/*
2396	* Validate the cookie and other input.
2397	*/
2398	if (pReq->Hdr.pSession != NULL)
2399	{
2400	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Hdr.pSession=%p expected NULL!\n", pReq->Hdr.pSession));
2401	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2402	}
2403	if (pReq->u.In.u32MagicCookie != SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE)
2404	{
2405	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: u32MagicCookie=%#x expected %#x!\n",
2406	(unsigned)pReq->u.In.u32MagicCookie, (unsigned)SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE));
2407	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2408	}
2409	if ( pReq->u.In.uMinVersion > pReq->u.In.uReqVersion
2410	\|\| (pReq->u.In.uMinVersion & UINT32_C(0xffff0000)) != (pReq->u.In.uReqVersion & UINT32_C(0xffff0000)))
2411	{
2412	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: uMinVersion=%#x uMaxVersion=%#x doesn't match!\n",
2413	pReq->u.In.uMinVersion, pReq->u.In.uReqVersion));
2414	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2415	}
2416	if (pSession != NULL)
2417	{
2418	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: pSession=%p expected NULL!\n", pSession));
2419	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2420	}
2421
2422	/*
2423	* Match the version.
2424	* The current logic is very simple, match the major interface version.
2425	*/
2426	if ( pReq->u.In.uMinVersion > SUPDRV_IDC_VERSION
2427	\|\| (pReq->u.In.uMinVersion & 0xffff0000) != (SUPDRV_IDC_VERSION & 0xffff0000))
2428	{
2429	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
2430	pReq->u.In.uReqVersion, pReq->u.In.uMinVersion, (unsigned)SUPDRV_IDC_VERSION));
2431	pReq->u.Out.pSession = NULL;
2432	pReq->u.Out.uSessionVersion = 0xffffffff;
2433	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
2434	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
2435	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
2436	return VINF_SUCCESS;
2437	}
2438
2439	pReq->u.Out.pSession = NULL;
2440	pReq->u.Out.uSessionVersion = SUPDRV_IDC_VERSION;
2441	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
2442	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
2443
2444	pReq->Hdr.rc = supdrvCreateSession(pDevExt, false /* fUser /, true /fUnrestricted*/, &pSession);
2445	if (RT_FAILURE(pReq->Hdr.rc))
2446	{
2447	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: failed to create session, rc=%d\n", pReq->Hdr.rc));
2448	return VINF_SUCCESS;
2449	}
2450
2451	pReq->u.Out.pSession = pSession;
2452	pReq->Hdr.pSession = pSession;
2453
2454	return VINF_SUCCESS;
2455	}
2456
2457	case SUPDRV_IDC_REQ_DISCONNECT:
2458	{
2459	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_DISCONNECT, sizeof(*pReqHdr));
2460
2461	supdrvSessionRelease(pSession);
2462	return pReqHdr->rc = VINF_SUCCESS;
2463	}
2464
2465	case SUPDRV_IDC_REQ_GET_SYMBOL:
2466	{
2467	PSUPDRVIDCREQGETSYM pReq = (PSUPDRVIDCREQGETSYM)pReqHdr;
2468	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_GET_SYMBOL, sizeof(*pReq));
2469
2470	pReq->Hdr.rc = supdrvIDC_LdrGetSymbol(pDevExt, pSession, pReq);
2471	return VINF_SUCCESS;
2472	}
2473
2474	case SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY:
2475	{
2476	PSUPDRVIDCREQCOMPREGFACTORY pReq = (PSUPDRVIDCREQCOMPREGFACTORY)pReqHdr;
2477	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY, sizeof(*pReq));
2478
2479	pReq->Hdr.rc = SUPR0ComponentRegisterFactory(pSession, pReq->u.In.pFactory);
2480	return VINF_SUCCESS;
2481	}
2482
2483	case SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY:
2484	{
2485	PSUPDRVIDCREQCOMPDEREGFACTORY pReq = (PSUPDRVIDCREQCOMPDEREGFACTORY)pReqHdr;
2486	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY, sizeof(*pReq));
2487
2488	pReq->Hdr.rc = SUPR0ComponentDeregisterFactory(pSession, pReq->u.In.pFactory);
2489	return VINF_SUCCESS;
2490	}
2491
2492	default:
2493	Log(("Unknown IDC %#lx\n", (long)uReq));
2494	break;
2495	}
2496
2497	#undef REQ_CHECK_IDC_SIZE
2498	return VERR_NOT_SUPPORTED;
2499	}
2500
2501
2502	/**
2503	* Register a object for reference counting.
2504	* The object is registered with one reference in the specified session.
2505	*
2506	* @returns Unique identifier on success (pointer).
2507	* All future reference must use this identifier.
2508	* @returns NULL on failure.
2509	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
2510	* @param pvUser1 The first user argument.
2511	* @param pvUser2 The second user argument.
2512	*/
2513	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
2514	{
2515	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2516	PSUPDRVOBJ pObj;
2517	PSUPDRVUSAGE pUsage;
2518
2519	/*
2520	* Validate the input.
2521	*/
2522	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
2523	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
2524	AssertPtrReturn(pfnDestructor, NULL);
2525
2526	/*
2527	* Allocate and initialize the object.
2528	*/
2529	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
2530	if (!pObj)
2531	return NULL;
2532	pObj->u32Magic = SUPDRVOBJ_MAGIC;
2533	pObj->enmType = enmType;
2534	pObj->pNext = NULL;
2535	pObj->cUsage = 1;
2536	pObj->pfnDestructor = pfnDestructor;
2537	pObj->pvUser1 = pvUser1;
2538	pObj->pvUser2 = pvUser2;
2539	pObj->CreatorUid = pSession->Uid;
2540	pObj->CreatorGid = pSession->Gid;
2541	pObj->CreatorProcess= pSession->Process;
2542	supdrvOSObjInitCreator(pObj, pSession);
2543
2544	/*
2545	* Allocate the usage record.
2546	* (We keep freed usage records around to simplify SUPR0ObjAddRefEx().)
2547	*/
2548	RTSpinlockAcquire(pDevExt->Spinlock);
2549
2550	pUsage = pDevExt->pUsageFree;
2551	if (pUsage)
2552	pDevExt->pUsageFree = pUsage->pNext;
2553	else
2554	{
2555	RTSpinlockRelease(pDevExt->Spinlock);
2556	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
2557	if (!pUsage)
2558	{
2559	RTMemFree(pObj);
2560	return NULL;
2561	}
2562	RTSpinlockAcquire(pDevExt->Spinlock);
2563	}
2564
2565	/*
2566	* Insert the object and create the session usage record.
2567	*/
2568	/* The object. */
2569	pObj->pNext = pDevExt->pObjs;
2570	pDevExt->pObjs = pObj;
2571
2572	/* The session record. */
2573	pUsage->cUsage = 1;
2574	pUsage->pObj = pObj;
2575	pUsage->pNext = pSession->pUsage;
2576	/* Log2(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext)); */
2577	pSession->pUsage = pUsage;
2578
2579	RTSpinlockRelease(pDevExt->Spinlock);
2580
2581	Log(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
2582	return pObj;
2583	}
2584
2585
2586	/**
2587	* Increment the reference counter for the object associating the reference
2588	* with the specified session.
2589	*
2590	* @returns IPRT status code.
2591	* @param pvObj The identifier returned by SUPR0ObjRegister().
2592	* @param pSession The session which is referencing the object.
2593	*
2594	* @remarks The caller should not own any spinlocks and must carefully protect
2595	* itself against potential race with the destructor so freed memory
2596	* isn't accessed here.
2597	*/
2598	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
2599	{
2600	return SUPR0ObjAddRefEx(pvObj, pSession, false /* fNoBlocking */);
2601	}
2602
2603
2604	/**
2605	* Increment the reference counter for the object associating the reference
2606	* with the specified session.
2607	*
2608	* @returns IPRT status code.
2609	* @retval VERR_TRY_AGAIN if fNoBlocking was set and a new usage record
2610	* couldn't be allocated. (If you see this you're not doing the right
2611	* thing and it won't ever work reliably.)
2612	*
2613	* @param pvObj The identifier returned by SUPR0ObjRegister().
2614	* @param pSession The session which is referencing the object.
2615	* @param fNoBlocking Set if it's not OK to block. Never try to make the
2616	* first reference to an object in a session with this
2617	* argument set.
2618	*
2619	* @remarks The caller should not own any spinlocks and must carefully protect
2620	* itself against potential race with the destructor so freed memory
2621	* isn't accessed here.
2622	*/
2623	SUPR0DECL(int) SUPR0ObjAddRefEx(void *pvObj, PSUPDRVSESSION pSession, bool fNoBlocking)
2624	{
2625	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2626	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2627	int rc = VINF_SUCCESS;
2628	PSUPDRVUSAGE pUsagePre;
2629	PSUPDRVUSAGE pUsage;
2630
2631	/*
2632	* Validate the input.
2633	* Be ready for the destruction race (someone might be stuck in the
2634	* destructor waiting a lock we own).
2635	*/
2636	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2637	AssertPtrReturn(pObj, VERR_INVALID_POINTER);
2638	AssertMsgReturn(pObj->u32Magic == SUPDRVOBJ_MAGIC \|\| pObj->u32Magic == SUPDRVOBJ_MAGIC_DEAD,
2639	("Invalid pvObj=%p magic=%#x (expected %#x or %#x)\n", pvObj, pObj->u32Magic, SUPDRVOBJ_MAGIC, SUPDRVOBJ_MAGIC_DEAD),
2640	VERR_INVALID_PARAMETER);
2641
2642	RTSpinlockAcquire(pDevExt->Spinlock);
2643
2644	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2645	{
2646	RTSpinlockRelease(pDevExt->Spinlock);
2647
2648	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2649	return VERR_WRONG_ORDER;
2650	}
2651
2652	/*
2653	* Preallocate the usage record if we can.
2654	*/
2655	pUsagePre = pDevExt->pUsageFree;
2656	if (pUsagePre)
2657	pDevExt->pUsageFree = pUsagePre->pNext;
2658	else if (!fNoBlocking)
2659	{
2660	RTSpinlockRelease(pDevExt->Spinlock);
2661	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
2662	if (!pUsagePre)
2663	return VERR_NO_MEMORY;
2664
2665	RTSpinlockAcquire(pDevExt->Spinlock);
2666	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2667	{
2668	RTSpinlockRelease(pDevExt->Spinlock);
2669
2670	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2671	return VERR_WRONG_ORDER;
2672	}
2673	}
2674
2675	/*
2676	* Reference the object.
2677	*/
2678	pObj->cUsage++;
2679
2680	/*
2681	* Look for the session record.
2682	*/
2683	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
2684	{
2685	/Log(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2686	if (pUsage->pObj == pObj)
2687	break;
2688	}
2689	if (pUsage)
2690	pUsage->cUsage++;
2691	else if (pUsagePre)
2692	{
2693	/* create a new session record. */
2694	pUsagePre->cUsage = 1;
2695	pUsagePre->pObj = pObj;
2696	pUsagePre->pNext = pSession->pUsage;
2697	pSession->pUsage = pUsagePre;
2698	/Log(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));/
2699
2700	pUsagePre = NULL;
2701	}
2702	else
2703	{
2704	pObj->cUsage--;
2705	rc = VERR_TRY_AGAIN;
2706	}
2707
2708	/*
2709	* Put any unused usage record into the free list..
2710	*/
2711	if (pUsagePre)
2712	{
2713	pUsagePre->pNext = pDevExt->pUsageFree;
2714	pDevExt->pUsageFree = pUsagePre;
2715	}
2716
2717	RTSpinlockRelease(pDevExt->Spinlock);
2718
2719	return rc;
2720	}
2721
2722
2723	/**
2724	* Decrement / destroy a reference counter record for an object.
2725	*
2726	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
2727	*
2728	* @returns IPRT status code.
2729	* @retval VINF_SUCCESS if not destroyed.
2730	* @retval VINF_OBJECT_DESTROYED if it's destroyed by this release call.
2731	* @retval VERR_INVALID_PARAMETER if the object isn't valid. Will assert in
2732	* string builds.
2733	*
2734	* @param pvObj The identifier returned by SUPR0ObjRegister().
2735	* @param pSession The session which is referencing the object.
2736	*/
2737	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
2738	{
2739	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2740	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2741	int rc = VERR_INVALID_PARAMETER;
2742	PSUPDRVUSAGE pUsage;
2743	PSUPDRVUSAGE pUsagePrev;
2744
2745	/*
2746	* Validate the input.
2747	*/
2748	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2749	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2750	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2751	VERR_INVALID_PARAMETER);
2752
2753	/*
2754	* Acquire the spinlock and look for the usage record.
2755	*/
2756	RTSpinlockAcquire(pDevExt->Spinlock);
2757
2758	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
2759	pUsage;
2760	pUsagePrev = pUsage, pUsage = pUsage->pNext)
2761	{
2762	/Log2(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2763	if (pUsage->pObj == pObj)
2764	{
2765	rc = VINF_SUCCESS;
2766	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
2767	if (pUsage->cUsage > 1)
2768	{
2769	pObj->cUsage--;
2770	pUsage->cUsage--;
2771	}
2772	else
2773	{
2774	/*
2775	* Free the session record.
2776	*/
2777	if (pUsagePrev)
2778	pUsagePrev->pNext = pUsage->pNext;
2779	else
2780	pSession->pUsage = pUsage->pNext;
2781	pUsage->pNext = pDevExt->pUsageFree;
2782	pDevExt->pUsageFree = pUsage;
2783
2784	/* What about the object? */
2785	if (pObj->cUsage > 1)
2786	pObj->cUsage--;
2787	else
2788	{
2789	/*
2790	* Object is to be destroyed, unlink it.
2791	*/
2792	pObj->u32Magic = SUPDRVOBJ_MAGIC_DEAD;
2793	rc = VINF_OBJECT_DESTROYED;
2794	if (pDevExt->pObjs == pObj)
2795	pDevExt->pObjs = pObj->pNext;
2796	else
2797	{
2798	PSUPDRVOBJ pObjPrev;
2799	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
2800	if (pObjPrev->pNext == pObj)
2801	{
2802	pObjPrev->pNext = pObj->pNext;
2803	break;
2804	}
2805	Assert(pObjPrev);
2806	}
2807	}
2808	}
2809	break;
2810	}
2811	}
2812
2813	RTSpinlockRelease(pDevExt->Spinlock);
2814
2815	/*
2816	* Call the destructor and free the object if required.
2817	*/
2818	if (rc == VINF_OBJECT_DESTROYED)
2819	{
2820	Log(("SUPR0ObjRelease: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
2821	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
2822	if (pObj->pfnDestructor)
2823	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
2824	RTMemFree(pObj);
2825	}
2826
2827	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
2828	return rc;
2829	}
2830
2831
2832	/**
2833	* Verifies that the current process can access the specified object.
2834	*
2835	* @returns The following IPRT status code:
2836	* @retval VINF_SUCCESS if access was granted.
2837	* @retval VERR_PERMISSION_DENIED if denied access.
2838	* @retval VERR_INVALID_PARAMETER if invalid parameter.
2839	*
2840	* @param pvObj The identifier returned by SUPR0ObjRegister().
2841	* @param pSession The session which wishes to access the object.
2842	* @param pszObjName Object string name. This is optional and depends on the object type.
2843	*
2844	* @remark The caller is responsible for making sure the object isn't removed while
2845	* we're inside this function. If uncertain about this, just call AddRef before calling us.
2846	*/
2847	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
2848	{
2849	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2850	int rc;
2851
2852	/*
2853	* Validate the input.
2854	*/
2855	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2856	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2857	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2858	VERR_INVALID_PARAMETER);
2859
2860	/*
2861	* Check access. (returns true if a decision has been made.)
2862	*/
2863	rc = VERR_INTERNAL_ERROR;
2864	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
2865	return rc;
2866
2867	/*
2868	* Default policy is to allow the user to access his own
2869	* stuff but nothing else.
2870	*/
2871	if (pObj->CreatorUid == pSession->Uid)
2872	return VINF_SUCCESS;
2873	return VERR_PERMISSION_DENIED;
2874	}
2875
2876
2877	/**
2878	* Lock pages.
2879	*
2880	* @returns IPRT status code.
2881	* @param pSession Session to which the locked memory should be associated.
2882	* @param pvR3 Start of the memory range to lock.
2883	* This must be page aligned.
2884	* @param cPages Number of pages to lock.
2885	* @param paPages Where to put the physical addresses of locked memory.
2886	*/
2887	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
2888	{
2889	int rc;
2890	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2891	const size_t cb = (size_t)cPages << PAGE_SHIFT;
2892	LogFlow(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
2893
2894	/*
2895	* Verify input.
2896	*/
2897	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2898	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
2899	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
2900	\|\| !pvR3)
2901	{
2902	Log(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
2903	return VERR_INVALID_PARAMETER;
2904	}
2905
2906	/*
2907	* Let IPRT do the job.
2908	*/
2909	Mem.eType = MEMREF_TYPE_LOCKED;
2910	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTMEM_PROT_READ \| RTMEM_PROT_WRITE, RTR0ProcHandleSelf());
2911	if (RT_SUCCESS(rc))
2912	{
2913	uint32_t iPage = cPages;
2914	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
2915	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
2916
2917	while (iPage-- > 0)
2918	{
2919	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2920	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
2921	{
2922	AssertMsgFailed(("iPage=%d\n", iPage));
2923	rc = VERR_INTERNAL_ERROR;
2924	break;
2925	}
2926	}
2927	if (RT_SUCCESS(rc))
2928	rc = supdrvMemAdd(&Mem, pSession);
2929	if (RT_FAILURE(rc))
2930	{
2931	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
2932	AssertRC(rc2);
2933	}
2934	}
2935
2936	return rc;
2937	}
2938
2939
2940	/**
2941	* Unlocks the memory pointed to by pv.
2942	*
2943	* @returns IPRT status code.
2944	* @param pSession Session to which the memory was locked.
2945	* @param pvR3 Memory to unlock.
2946	*/
2947	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2948	{
2949	LogFlow(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2950	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2951	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
2952	}
2953
2954
2955	/**
2956	* Allocates a chunk of page aligned memory with contiguous and fixed physical
2957	* backing.
2958	*
2959	* @returns IPRT status code.
2960	* @param pSession Session data.
2961	* @param cPages Number of pages to allocate.
2962	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
2963	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
2964	* @param pHCPhys Where to put the physical address of allocated memory.
2965	*/
2966	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
2967	{
2968	int rc;
2969	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2970	LogFlow(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
2971
2972	/*
2973	* Validate input.
2974	*/
2975	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2976	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
2977	{
2978	Log(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
2979	pSession, ppvR0, ppvR3, pHCPhys));
2980	return VERR_INVALID_PARAMETER;
2981
2982	}
2983	if (cPages < 1 \|\| cPages >= 256)
2984	{
2985	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2986	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2987	}
2988
2989	/*
2990	* Let IPRT do the job.
2991	*/
2992	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
2993	if (RT_SUCCESS(rc))
2994	{
2995	int rc2;
2996	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2997	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2998	if (RT_SUCCESS(rc))
2999	{
3000	Mem.eType = MEMREF_TYPE_CONT;
3001	rc = supdrvMemAdd(&Mem, pSession);
3002	if (!rc)
3003	{
3004	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
3005	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
3006	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
3007	return 0;
3008	}
3009
3010	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
3011	AssertRC(rc2);
3012	}
3013	rc2 = RTR0MemObjFree(Mem.MemObj, false);
3014	AssertRC(rc2);
3015	}
3016
3017	return rc;
3018	}
3019
3020
3021	/**
3022	* Frees memory allocated using SUPR0ContAlloc().
3023	*
3024	* @returns IPRT status code.
3025	* @param pSession The session to which the memory was allocated.
3026	* @param uPtr Pointer to the memory (ring-3 or ring-0).
3027	*/
3028	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
3029	{
3030	LogFlow(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
3031	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3032	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
3033	}
3034
3035
3036	/**
3037	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
3038	*
3039	* The memory isn't zeroed.
3040	*
3041	* @returns IPRT status code.
3042	* @param pSession Session data.
3043	* @param cPages Number of pages to allocate.
3044	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
3045	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
3046	* @param paPages Where to put the physical addresses of allocated memory.
3047	*/
3048	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
3049	{
3050	unsigned iPage;
3051	int rc;
3052	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
3053	LogFlow(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
3054
3055	/*
3056	* Validate input.
3057	*/
3058	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3059	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
3060	{
3061	Log(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
3062	pSession, ppvR3, ppvR0, paPages));
3063	return VERR_INVALID_PARAMETER;
3064
3065	}
3066	if (cPages < 1 \|\| cPages >= 256)
3067	{
3068	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
3069	return VERR_PAGE_COUNT_OUT_OF_RANGE;
3070	}
3071
3072	/*
3073	* Let IPRT do the work.
3074	*/
3075	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
3076	if (RT_SUCCESS(rc))
3077	{
3078	int rc2;
3079	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
3080	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
3081	if (RT_SUCCESS(rc))
3082	{
3083	Mem.eType = MEMREF_TYPE_LOW;
3084	rc = supdrvMemAdd(&Mem, pSession);
3085	if (!rc)
3086	{
3087	for (iPage = 0; iPage < cPages; iPage++)
3088	{
3089	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
3090	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%RHp\n", paPages[iPage]));
3091	}
3092	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
3093	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
3094	return 0;
3095	}
3096
3097	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
3098	AssertRC(rc2);
3099	}
3100
3101	rc2 = RTR0MemObjFree(Mem.MemObj, false);
3102	AssertRC(rc2);
3103	}
3104
3105	return rc;
3106	}
3107
3108
3109	/**
3110	* Frees memory allocated using SUPR0LowAlloc().
3111	*
3112	* @returns IPRT status code.
3113	* @param pSession The session to which the memory was allocated.
3114	* @param uPtr Pointer to the memory (ring-3 or ring-0).
3115	*/
3116	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
3117	{
3118	LogFlow(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
3119	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3120	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
3121	}
3122
3123
3124
3125	/**
3126	* Allocates a chunk of memory with both R0 and R3 mappings.
3127	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
3128	*
3129	* @returns IPRT status code.
3130	* @param pSession The session to associated the allocation with.
3131	* @param cb Number of bytes to allocate.
3132	* @param ppvR0 Where to store the address of the Ring-0 mapping.
3133	* @param ppvR3 Where to store the address of the Ring-3 mapping.
3134	*/
3135	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
3136	{
3137	int rc;
3138	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
3139	LogFlow(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
3140
3141	/*
3142	* Validate input.
3143	*/
3144	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3145	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
3146	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
3147	if (cb < 1 \|\| cb >= _4M)
3148	{
3149	Log(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
3150	return VERR_INVALID_PARAMETER;
3151	}
3152
3153	/*
3154	* Let IPRT do the work.
3155	*/
3156	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
3157	if (RT_SUCCESS(rc))
3158	{
3159	int rc2;
3160	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
3161	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
3162	if (RT_SUCCESS(rc))
3163	{
3164	Mem.eType = MEMREF_TYPE_MEM;
3165	rc = supdrvMemAdd(&Mem, pSession);
3166	if (!rc)
3167	{
3168	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
3169	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
3170	return VINF_SUCCESS;
3171	}
3172
3173	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
3174	AssertRC(rc2);
3175	}
3176
3177	rc2 = RTR0MemObjFree(Mem.MemObj, false);
3178	AssertRC(rc2);
3179	}
3180
3181	return rc;
3182	}
3183
3184
3185	/**
3186	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
3187	*
3188	* @returns IPRT status code.
3189	* @param pSession The session to which the memory was allocated.
3190	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
3191	* @param paPages Where to store the physical addresses.
3192	*/
3193	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
3194	{
3195	PSUPDRVBUNDLE pBundle;
3196	LogFlow(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
3197
3198	/*
3199	* Validate input.
3200	*/
3201	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3202	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
3203	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
3204
3205	/*
3206	* Search for the address.
3207	*/
3208	RTSpinlockAcquire(pSession->Spinlock);
3209	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3210	{
3211	if (pBundle->cUsed > 0)
3212	{
3213	unsigned i;
3214	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3215	{
3216	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
3217	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3218	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
3219	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3220	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
3221	)
3222	)
3223	{
3224	const size_t cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
3225	size_t iPage;
3226	for (iPage = 0; iPage < cPages; iPage++)
3227	{
3228	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
3229	paPages[iPage].uReserved = 0;
3230	}
3231	RTSpinlockRelease(pSession->Spinlock);
3232	return VINF_SUCCESS;
3233	}
3234	}
3235	}
3236	}
3237	RTSpinlockRelease(pSession->Spinlock);
3238	Log(("Failed to find %p!!!\n", (void *)uPtr));
3239	return VERR_INVALID_PARAMETER;
3240	}
3241
3242
3243	/**
3244	* Free memory allocated by SUPR0MemAlloc().
3245	*
3246	* @returns IPRT status code.
3247	* @param pSession The session owning the allocation.
3248	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
3249	*/
3250	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
3251	{
3252	LogFlow(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
3253	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3254	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
3255	}
3256
3257
3258	/**
3259	* Allocates a chunk of memory with a kernel or/and a user mode mapping.
3260	*
3261	* The memory is fixed and it's possible to query the physical addresses using
3262	* SUPR0MemGetPhys().
3263	*
3264	* @returns IPRT status code.
3265	* @param pSession The session to associated the allocation with.
3266	* @param cPages The number of pages to allocate.
3267	* @param fFlags Flags, reserved for the future. Must be zero.
3268	* @param ppvR3 Where to store the address of the Ring-3 mapping.
3269	* NULL if no ring-3 mapping.
3270	* @param ppvR3 Where to store the address of the Ring-0 mapping.
3271	* NULL if no ring-0 mapping.
3272	* @param paPages Where to store the addresses of the pages. Optional.
3273	*/
3274	SUPR0DECL(int) SUPR0PageAllocEx(PSUPDRVSESSION pSession, uint32_t cPages, uint32_t fFlags, PRTR3PTR ppvR3, PRTR0PTR ppvR0, PRTHCPHYS paPages)
3275	{
3276	int rc;
3277	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
3278	LogFlow(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
3279
3280	/*
3281	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
3282	*/
3283	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3284	AssertPtrNullReturn(ppvR3, VERR_INVALID_POINTER);
3285	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
3286	AssertReturn(ppvR3 \|\| ppvR0, VERR_INVALID_PARAMETER);
3287	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
3288	if (cPages < 1 \|\| cPages > VBOX_MAX_ALLOC_PAGE_COUNT)
3289	{
3290	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)));
3291	return VERR_PAGE_COUNT_OUT_OF_RANGE;
3292	}
3293
3294	/*
3295	* Let IPRT do the work.
3296	*/
3297	if (ppvR0)
3298	rc = RTR0MemObjAllocPage(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, true /* fExecutable */);
3299	else
3300	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
3301	if (RT_SUCCESS(rc))
3302	{
3303	int rc2;
3304	if (ppvR3)
3305	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
3306	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
3307	else
3308	Mem.MapObjR3 = NIL_RTR0MEMOBJ;
3309	if (RT_SUCCESS(rc))
3310	{
3311	Mem.eType = MEMREF_TYPE_PAGE;
3312	rc = supdrvMemAdd(&Mem, pSession);
3313	if (!rc)
3314	{
3315	if (ppvR3)
3316	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
3317	if (ppvR0)
3318	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
3319	if (paPages)
3320	{
3321	uint32_t iPage = cPages;
3322	while (iPage-- > 0)
3323	{
3324	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
3325	Assert(paPages[iPage] != NIL_RTHCPHYS);
3326	}
3327	}
3328	return VINF_SUCCESS;
3329	}
3330
3331	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
3332	AssertRC(rc2);
3333	}
3334
3335	rc2 = RTR0MemObjFree(Mem.MemObj, false);
3336	AssertRC(rc2);
3337	}
3338	return rc;
3339	}
3340
3341
3342	/**
3343	* Maps a chunk of memory previously allocated by SUPR0PageAllocEx into kernel
3344	* space.
3345	*
3346	* @returns IPRT status code.
3347	* @param pSession The session to associated the allocation with.
3348	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
3349	* @param offSub Where to start mapping. Must be page aligned.
3350	* @param cbSub How much to map. Must be page aligned.
3351	* @param fFlags Flags, MBZ.
3352	* @param ppvR0 Where to return the address of the ring-0 mapping on
3353	* success.
3354	*/
3355	SUPR0DECL(int) SUPR0PageMapKernel(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t offSub, uint32_t cbSub,
3356	uint32_t fFlags, PRTR0PTR ppvR0)
3357	{
3358	int rc;
3359	PSUPDRVBUNDLE pBundle;
3360	RTR0MEMOBJ hMemObj = NIL_RTR0MEMOBJ;
3361	LogFlow(("SUPR0PageMapKernel: pSession=%p pvR3=%p offSub=%#x cbSub=%#x\n", pSession, pvR3, offSub, cbSub));
3362
3363	/*
3364	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
3365	*/
3366	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3367	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
3368	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
3369	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3370	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3371	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
3372
3373	/*
3374	* Find the memory object.
3375	*/
3376	RTSpinlockAcquire(pSession->Spinlock);
3377	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3378	{
3379	if (pBundle->cUsed > 0)
3380	{
3381	unsigned i;
3382	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3383	{
3384	if ( ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
3385	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3386	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3387	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
3388	\|\| ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED
3389	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3390	&& pBundle->aMem[i].MapObjR3 == NIL_RTR0MEMOBJ
3391	&& RTR0MemObjAddressR3(pBundle->aMem[i].MemObj) == pvR3))
3392	{
3393	hMemObj = pBundle->aMem[i].MemObj;
3394	break;
3395	}
3396	}
3397	}
3398	}
3399	RTSpinlockRelease(pSession->Spinlock);
3400
3401	rc = VERR_INVALID_PARAMETER;
3402	if (hMemObj != NIL_RTR0MEMOBJ)
3403	{
3404	/*
3405	* Do some further input validations before calling IPRT.
3406	* (Cleanup is done indirectly by telling RTR0MemObjFree to include mappings.)
3407	*/
3408	size_t cbMemObj = RTR0MemObjSize(hMemObj);
3409	if ( offSub < cbMemObj
3410	&& cbSub <= cbMemObj
3411	&& offSub + cbSub <= cbMemObj)
3412	{
3413	RTR0MEMOBJ hMapObj;
3414	rc = RTR0MemObjMapKernelEx(&hMapObj, hMemObj, (void *)-1, 0,
3415	RTMEM_PROT_READ \| RTMEM_PROT_WRITE, offSub, cbSub);
3416	if (RT_SUCCESS(rc))
3417	*ppvR0 = RTR0MemObjAddress(hMapObj);
3418	}
3419	else
3420	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
3421
3422	}
3423	return rc;
3424	}
3425
3426
3427	/**
3428	* Changes the page level protection of one or more pages previously allocated
3429	* by SUPR0PageAllocEx.
3430	*
3431	* @returns IPRT status code.
3432	* @param pSession The session to associated the allocation with.
3433	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
3434	* NIL_RTR3PTR if the ring-3 mapping should be unaffected.
3435	* @param pvR0 The ring-0 address returned by SUPR0PageAllocEx.
3436	* NIL_RTR0PTR if the ring-0 mapping should be unaffected.
3437	* @param offSub Where to start changing. Must be page aligned.
3438	* @param cbSub How much to change. Must be page aligned.
3439	* @param fProt The new page level protection, see RTMEM_PROT_*.
3440	*/
3441	SUPR0DECL(int) SUPR0PageProtect(PSUPDRVSESSION pSession, RTR3PTR pvR3, RTR0PTR pvR0, uint32_t offSub, uint32_t cbSub, uint32_t fProt)
3442	{
3443	int rc;
3444	PSUPDRVBUNDLE pBundle;
3445	RTR0MEMOBJ hMemObjR0 = NIL_RTR0MEMOBJ;
3446	RTR0MEMOBJ hMemObjR3 = NIL_RTR0MEMOBJ;
3447	LogFlow(("SUPR0PageProtect: pSession=%p pvR3=%p pvR0=%p offSub=%#x cbSub=%#x fProt-%#x\n", pSession, pvR3, pvR0, offSub, cbSub, fProt));
3448
3449	/*
3450	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
3451	*/
3452	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3453	AssertReturn(!(fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)), VERR_INVALID_PARAMETER);
3454	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3455	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3456	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
3457
3458	/*
3459	* Find the memory object.
3460	*/
3461	RTSpinlockAcquire(pSession->Spinlock);
3462	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3463	{
3464	if (pBundle->cUsed > 0)
3465	{
3466	unsigned i;
3467	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3468	{
3469	if ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
3470	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3471	&& ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3472	\|\| pvR3 == NIL_RTR3PTR)
3473	&& ( pvR0 == NIL_RTR0PTR
3474	\|\| RTR0MemObjAddress(pBundle->aMem[i].MemObj) == pvR0)
3475	&& ( pvR3 == NIL_RTR3PTR
3476	\|\| RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3))
3477	{
3478	if (pvR0 != NIL_RTR0PTR)
3479	hMemObjR0 = pBundle->aMem[i].MemObj;
3480	if (pvR3 != NIL_RTR3PTR)
3481	hMemObjR3 = pBundle->aMem[i].MapObjR3;
3482	break;
3483	}
3484	}
3485	}
3486	}
3487	RTSpinlockRelease(pSession->Spinlock);
3488
3489	rc = VERR_INVALID_PARAMETER;
3490	if ( hMemObjR0 != NIL_RTR0MEMOBJ
3491	\|\| hMemObjR3 != NIL_RTR0MEMOBJ)
3492	{
3493	/*
3494	* Do some further input validations before calling IPRT.
3495	*/
3496	size_t cbMemObj = hMemObjR0 != NIL_RTR0PTR ? RTR0MemObjSize(hMemObjR0) : RTR0MemObjSize(hMemObjR3);
3497	if ( offSub < cbMemObj
3498	&& cbSub <= cbMemObj
3499	&& offSub + cbSub <= cbMemObj)
3500	{
3501	rc = VINF_SUCCESS;
3502	if (hMemObjR3 != NIL_RTR0PTR)
3503	rc = RTR0MemObjProtect(hMemObjR3, offSub, cbSub, fProt);
3504	if (hMemObjR0 != NIL_RTR0PTR && RT_SUCCESS(rc))
3505	rc = RTR0MemObjProtect(hMemObjR0, offSub, cbSub, fProt);
3506	}
3507	else
3508	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
3509
3510	}
3511	return rc;
3512
3513	}
3514
3515
3516	/**
3517	* Free memory allocated by SUPR0PageAlloc() and SUPR0PageAllocEx().
3518	*
3519	* @returns IPRT status code.
3520	* @param pSession The session owning the allocation.
3521	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc() or
3522	* SUPR0PageAllocEx().
3523	*/
3524	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
3525	{
3526	LogFlow(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
3527	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3528	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_PAGE);
3529	}
3530
3531
3532	/**
3533	* Gets the paging mode of the current CPU.
3534	*
3535	* @returns Paging mode, SUPPAGEINGMODE_INVALID on error.
3536	*/
3537	SUPR0DECL(SUPPAGINGMODE) SUPR0GetPagingMode(void)
3538	{
3539	SUPPAGINGMODE enmMode;
3540
3541	RTR0UINTREG cr0 = ASMGetCR0();
3542	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
3543	enmMode = SUPPAGINGMODE_INVALID;
3544	else
3545	{
3546	RTR0UINTREG cr4 = ASMGetCR4();
3547	uint32_t fNXEPlusLMA = 0;
3548	if (cr4 & X86_CR4_PAE)
3549	{
3550	uint32_t fExtFeatures = ASMCpuId_EDX(0x80000001);
3551	if (fExtFeatures & (X86_CPUID_EXT_FEATURE_EDX_NX \| X86_CPUID_EXT_FEATURE_EDX_LONG_MODE))
3552	{
3553	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
3554	if ((fExtFeatures & X86_CPUID_EXT_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
3555	fNXEPlusLMA \|= RT_BIT(0);
3556	if ((fExtFeatures & X86_CPUID_EXT_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
3557	fNXEPlusLMA \|= RT_BIT(1);
3558	}
3559	}
3560
3561	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
3562	{
3563	case 0:
3564	enmMode = SUPPAGINGMODE_32_BIT;
3565	break;
3566
3567	case X86_CR4_PGE:
3568	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
3569	break;
3570
3571	case X86_CR4_PAE:
3572	enmMode = SUPPAGINGMODE_PAE;
3573	break;
3574
3575	case X86_CR4_PAE \| RT_BIT(0):
3576	enmMode = SUPPAGINGMODE_PAE_NX;
3577	break;
3578
3579	case X86_CR4_PAE \| X86_CR4_PGE:
3580	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3581	break;
3582
3583	case X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3584	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3585	break;
3586
3587	case RT_BIT(1) \| X86_CR4_PAE:
3588	enmMode = SUPPAGINGMODE_AMD64;
3589	break;
3590
3591	case RT_BIT(1) \| X86_CR4_PAE \| RT_BIT(0):
3592	enmMode = SUPPAGINGMODE_AMD64_NX;
3593	break;
3594
3595	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
3596	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
3597	break;
3598
3599	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3600	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
3601	break;
3602
3603	default:
3604	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
3605	enmMode = SUPPAGINGMODE_INVALID;
3606	break;
3607	}
3608	}
3609	return enmMode;
3610	}
3611
3612
3613	/**
3614	* Enables or disabled hardware virtualization extensions using native OS APIs.
3615	*
3616	* @returns VBox status code.
3617	* @retval VINF_SUCCESS on success.
3618	* @retval VERR_NOT_SUPPORTED if not supported by the native OS.
3619	*
3620	* @param fEnable Whether to enable or disable.
3621	*/
3622	SUPR0DECL(int) SUPR0EnableVTx(bool fEnable)
3623	{
3624	#ifdef RT_OS_DARWIN
3625	return supdrvOSEnableVTx(fEnable);
3626	#else
3627	return VERR_NOT_SUPPORTED;
3628	#endif
3629	}
3630
3631
3632	/**
3633	* Suspends hardware virtualization extensions using the native OS API.
3634	*
3635	* This is called prior to entering raw-mode context.
3636	*
3637	* @returns @c true if suspended, @c false if not.
3638	*/
3639	SUPR0DECL(bool) SUPR0SuspendVTxOnCpu(void)
3640	{
3641	#ifdef RT_OS_DARWIN
3642	return supdrvOSSuspendVTxOnCpu();
3643	#else
3644	return false;
3645	#endif
3646	}
3647
3648
3649	/**
3650	* Resumes hardware virtualization extensions using the native OS API.
3651	*
3652	* This is called after to entering raw-mode context.
3653	*
3654	* @param fSuspended The return value of SUPR0SuspendVTxOnCpu.
3655	*/
3656	SUPR0DECL(void) SUPR0ResumeVTxOnCpu(bool fSuspended)
3657	{
3658	#ifdef RT_OS_DARWIN
3659	supdrvOSResumeVTxOnCpu(fSuspended);
3660	#else
3661	Assert(!fSuspended);
3662	#endif
3663	}
3664
3665
3666	/**
3667	* Queries the AMD-V and VT-x capabilities of the calling CPU.
3668	*
3669	* @returns VBox status code.
3670	* @retval VERR_VMX_NO_VMX
3671	* @retval VERR_VMX_MSR_ALL_VMXON_DISABLED
3672	* @retval VERR_VMX_MSR_VMXON_DISABLED
3673	* @retval VERR_VMX_MSR_LOCKING_FAILED
3674	* @retval VERR_SVM_NO_SVM
3675	* @retval VERR_SVM_DISABLED
3676	* @retval VERR_UNSUPPORTED_CPU if not identifiable as an AMD, Intel or VIA
3677	* (centaur) CPU.
3678	*
3679	* @param pSession The session handle.
3680	* @param pfCaps Where to store the capabilities.
3681	*/
3682	SUPR0DECL(int) SUPR0QueryVTCaps(PSUPDRVSESSION pSession, uint32_t *pfCaps)
3683	{
3684	int rc = VERR_UNSUPPORTED_CPU;
3685	bool fIsSmxModeAmbiguous = false;
3686	RTTHREADPREEMPTSTATE PreemptState = RTTHREADPREEMPTSTATE_INITIALIZER;
3687
3688	/*
3689	* Input validation.
3690	*/
3691	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3692	AssertPtrReturn(pfCaps, VERR_INVALID_POINTER);
3693
3694	*pfCaps = 0;
3695	/* We may modify MSRs and re-read them, disable preemption so we make sure we don't migrate CPUs. */
3696	RTThreadPreemptDisable(&PreemptState);
3697	if (ASMHasCpuId())
3698	{
3699	uint32_t fFeaturesECX, fFeaturesEDX, uDummy;
3700	uint32_t uMaxId, uVendorEBX, uVendorECX, uVendorEDX;
3701
3702	ASMCpuId(0, &uMaxId, &uVendorEBX, &uVendorECX, &uVendorEDX);
3703	ASMCpuId(1, &uDummy, &uDummy, &fFeaturesECX, &fFeaturesEDX);
3704
3705	if ( ASMIsValidStdRange(uMaxId)
3706	&& ( ASMIsIntelCpuEx( uVendorEBX, uVendorECX, uVendorEDX)
3707	\|\| ASMIsViaCentaurCpuEx(uVendorEBX, uVendorECX, uVendorEDX) )
3708	)
3709	{
3710	if ( (fFeaturesECX & X86_CPUID_FEATURE_ECX_VMX)
3711	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3712	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3713	)
3714	{
3715	/** @todo Unify code with hmR0InitIntelCpu(). */
3716	uint64_t u64FeatMsr = ASMRdMsr(MSR_IA32_FEATURE_CONTROL);
3717	bool const fMaybeSmxMode = RT_BOOL(ASMGetCR4() & X86_CR4_SMXE);
3718	bool fMsrLocked = RT_BOOL(u64FeatMsr & MSR_IA32_FEATURE_CONTROL_LOCK);
3719	bool fSmxVmxAllowed = RT_BOOL(u64FeatMsr & MSR_IA32_FEATURE_CONTROL_SMX_VMXON);
3720	bool fVmxAllowed = RT_BOOL(u64FeatMsr & MSR_IA32_FEATURE_CONTROL_VMXON);
3721
3722	/* Check if the LOCK bit is set but excludes the required VMXON bit. */
3723	if (fMsrLocked)
3724	{
3725	if (fVmxAllowed && fSmxVmxAllowed)
3726	rc = VINF_SUCCESS;
3727	else if (!fVmxAllowed && !fSmxVmxAllowed)
3728	rc = VERR_VMX_MSR_ALL_VMXON_DISABLED;
3729	else if (!fMaybeSmxMode)
3730	{
3731	if (fVmxAllowed)
3732	rc = VINF_SUCCESS;
3733	else
3734	rc = VERR_VMX_MSR_VMXON_DISABLED;
3735	}
3736	else
3737	{
3738	/*
3739	* CR4.SMXE is set but this doesn't mean the CPU is necessarily in SMX mode. We shall assume
3740	* that it is -not- and that it is a stupid BIOS/OS setting CR4.SMXE for no good reason.
3741	* See @bugref{6873}.
3742	*/
3743	Assert(fMaybeSmxMode == true);
3744	fIsSmxModeAmbiguous = true;
3745	rc = VINF_SUCCESS;
3746	}
3747	}
3748	else
3749	{
3750	/*
3751	* MSR is not yet locked; we can change it ourselves here.
3752	* Once the lock bit is set, this MSR can no longer be modified.
3753	*
3754	* Set both the VMXON and SMX_VMXON bits as we can't determine SMX mode
3755	* accurately. See @bugref{6873}.
3756	*/
3757	u64FeatMsr \|= MSR_IA32_FEATURE_CONTROL_LOCK
3758	\| MSR_IA32_FEATURE_CONTROL_SMX_VMXON
3759	\| MSR_IA32_FEATURE_CONTROL_VMXON;
3760	ASMWrMsr(MSR_IA32_FEATURE_CONTROL, u64FeatMsr);
3761
3762	/* Verify. */
3763	u64FeatMsr = ASMRdMsr(MSR_IA32_FEATURE_CONTROL);
3764	fMsrLocked = RT_BOOL(u64FeatMsr & MSR_IA32_FEATURE_CONTROL_LOCK);
3765	fSmxVmxAllowed = fMsrLocked && RT_BOOL(u64FeatMsr & MSR_IA32_FEATURE_CONTROL_SMX_VMXON);
3766	fVmxAllowed = fMsrLocked && RT_BOOL(u64FeatMsr & MSR_IA32_FEATURE_CONTROL_VMXON);
3767	if (fSmxVmxAllowed && fVmxAllowed)
3768	rc = VINF_SUCCESS;
3769	else
3770	rc = VERR_VMX_MSR_LOCKING_FAILED;
3771	}
3772
3773	if (rc == VINF_SUCCESS)
3774	{
3775	VMXCAPABILITY vtCaps;
3776
3777	*pfCaps \|= SUPVTCAPS_VT_X;
3778
3779	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS);
3780	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC_USE_SECONDARY_EXEC_CTRL)
3781	{
3782	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS2);
3783	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC2_EPT)
3784	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3785	}
3786	}
3787	}
3788	else
3789	rc = VERR_VMX_NO_VMX;
3790	}
3791	else if ( ASMIsAmdCpuEx(uVendorEBX, uVendorECX, uVendorEDX)
3792	&& ASMIsValidStdRange(uMaxId))
3793	{
3794	uint32_t fExtFeaturesEcx, uExtMaxId;
3795	ASMCpuId(0x80000000, &uExtMaxId, &uDummy, &uDummy, &uDummy);
3796	ASMCpuId(0x80000001, &uDummy, &uDummy, &fExtFeaturesEcx, &uDummy);
3797	if ( ASMIsValidExtRange(uExtMaxId)
3798	&& uExtMaxId >= 0x8000000a
3799	&& (fExtFeaturesEcx & X86_CPUID_AMD_FEATURE_ECX_SVM)
3800	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3801	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3802	)
3803	{
3804	/* Check if SVM is disabled */
3805	uint64_t u64FeatMsr = ASMRdMsr(MSR_K8_VM_CR);
3806	if (!(u64FeatMsr & MSR_K8_VM_CR_SVM_DISABLE))
3807	{
3808	uint32_t fSvmFeatures;
3809	*pfCaps \|= SUPVTCAPS_AMD_V;
3810
3811	/* Query AMD-V features. */
3812	ASMCpuId(0x8000000a, &uDummy, &uDummy, &uDummy, &fSvmFeatures);
3813	if (fSvmFeatures & AMD_CPUID_SVM_FEATURE_EDX_NESTED_PAGING)
3814	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3815
3816	rc = VINF_SUCCESS;
3817	}
3818	else
3819	rc = VERR_SVM_DISABLED;
3820	}
3821	else
3822	rc = VERR_SVM_NO_SVM;
3823	}
3824	}
3825
3826	RTThreadPreemptRestore(&PreemptState);
3827	if (fIsSmxModeAmbiguous)
3828	SUPR0Printf(("WARNING! CR4 hints SMX mode but your CPU is too secretive. Proceeding anyway... We wish you good luck!\n"));
3829	return rc;
3830	}
3831
3832
3833	/**
3834	* (Re-)initializes the per-cpu structure prior to starting or resuming the GIP
3835	* updating.
3836	*
3837	* @param pGipCpu The per CPU structure for this CPU.
3838	* @param u64NanoTS The current time.
3839	*/
3840	static void supdrvGipReInitCpu(PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
3841	{
3842	pGipCpu->u64TSC = ASMReadTSC() - pGipCpu->u32UpdateIntervalTSC;
3843	pGipCpu->u64NanoTS = u64NanoTS;
3844	}
3845
3846
3847	/**
3848	* Set the current TSC and NanoTS value for the CPU.
3849	*
3850	* @param idCpu The CPU ID. Unused - we have to use the APIC ID.
3851	* @param pvUser1 Pointer to the ring-0 GIP mapping.
3852	* @param pvUser2 Pointer to the variable holding the current time.
3853	*/
3854	static DECLCALLBACK(void) supdrvGipReInitCpuCallback(RTCPUID idCpu, void pvUser1, void pvUser2)
3855	{
3856	PSUPGLOBALINFOPAGE pGip = (PSUPGLOBALINFOPAGE)pvUser1;
3857	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
3858
3859	if (RT_LIKELY(iCpu < pGip->cCpus && pGip->aCPUs[iCpu].idCpu == idCpu))
3860	supdrvGipReInitCpu(&pGip->aCPUs[iCpu], (uint64_t )pvUser2);
3861
3862	NOREF(pvUser2);
3863	NOREF(idCpu);
3864	}
3865
3866
3867	/**
3868	* Maps the GIP into userspace and/or get the physical address of the GIP.
3869	*
3870	* @returns IPRT status code.
3871	* @param pSession Session to which the GIP mapping should belong.
3872	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
3873	* @param pHCPhysGip Where to store the physical address. (optional)
3874	*
3875	* @remark There is no reference counting on the mapping, so one call to this function
3876	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
3877	* and remove the session as a GIP user.
3878	*/
3879	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
3880	{
3881	int rc;
3882	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3883	RTR3PTR pGipR3 = NIL_RTR3PTR;
3884	RTHCPHYS HCPhys = NIL_RTHCPHYS;
3885	LogFlow(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
3886
3887	/*
3888	* Validate
3889	*/
3890	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3891	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
3892	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
3893
3894	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3895	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3896	#else
3897	RTSemFastMutexRequest(pDevExt->mtxGip);
3898	#endif
3899	if (pDevExt->pGip)
3900	{
3901	/*
3902	* Map it?
3903	*/
3904	rc = VINF_SUCCESS;
3905	if (ppGipR3)
3906	{
3907	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
3908	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
3909	RTMEM_PROT_READ, RTR0ProcHandleSelf());
3910	if (RT_SUCCESS(rc))
3911	pGipR3 = RTR0MemObjAddressR3(pSession->GipMapObjR3);
3912	}
3913
3914	/*
3915	* Get physical address.
3916	*/
3917	if (pHCPhysGip && RT_SUCCESS(rc))
3918	HCPhys = pDevExt->HCPhysGip;
3919
3920	/*
3921	* Reference globally.
3922	*/
3923	if (!pSession->fGipReferenced && RT_SUCCESS(rc))
3924	{
3925	pSession->fGipReferenced = 1;
3926	pDevExt->cGipUsers++;
3927	if (pDevExt->cGipUsers == 1)
3928	{
3929	PSUPGLOBALINFOPAGE pGipR0 = pDevExt->pGip;
3930	uint64_t u64NanoTS;
3931	uint32_t u32SystemResolution;
3932	unsigned i;
3933
3934	LogFlow(("SUPR0GipMap: Resumes GIP updating\n"));
3935
3936	/*
3937	* Try bump up the system timer resolution.
3938	* The more interrupts the better...
3939	*/
3940	if ( RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
3941	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
3942	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1953125 /* 512 HZ */, &u32SystemResolution))
3943	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 2000000 /* 500 HZ */, &u32SystemResolution))
3944	)
3945	{
3946	Assert(RTTimerGetSystemGranularity() <= u32SystemResolution);
3947	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
3948	}
3949
3950	if (pGipR0->aCPUs[0].u32TransactionId != 2 /* not the first time */)
3951	{
3952	for (i = 0; i < RT_ELEMENTS(pGipR0->aCPUs); i++)
3953	ASMAtomicUoWriteU32(&pGipR0->aCPUs[i].u32TransactionId,
3954	(pGipR0->aCPUs[i].u32TransactionId + GIP_UPDATEHZ_RECALC_FREQ * 2)
3955	& ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
3956	ASMAtomicWriteU64(&pGipR0->u64NanoTSLastUpdateHz, 0);
3957	}
3958
3959	u64NanoTS = RTTimeSystemNanoTS() - pGipR0->u32UpdateIntervalNS;
3960	if ( pGipR0->u32Mode == SUPGIPMODE_SYNC_TSC
3961	\|\| RTMpGetOnlineCount() == 1)
3962	supdrvGipReInitCpu(&pGipR0->aCPUs[0], u64NanoTS);
3963	else
3964	RTMpOnAll(supdrvGipReInitCpuCallback, pGipR0, &u64NanoTS);
3965
3966	#ifndef DO_NOT_START_GIP
3967	rc = RTTimerStart(pDevExt->pGipTimer, 0); AssertRC(rc);
3968	#endif
3969	rc = VINF_SUCCESS;
3970	}
3971	}
3972	}
3973	else
3974	{
3975	rc = VERR_GENERAL_FAILURE;
3976	Log(("SUPR0GipMap: GIP is not available!\n"));
3977	}
3978	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3979	RTSemMutexRelease(pDevExt->mtxGip);
3980	#else
3981	RTSemFastMutexRelease(pDevExt->mtxGip);
3982	#endif
3983
3984	/*
3985	* Write returns.
3986	*/
3987	if (pHCPhysGip)
3988	*pHCPhysGip = HCPhys;
3989	if (ppGipR3)
3990	*ppGipR3 = pGipR3;
3991
3992	#ifdef DEBUG_DARWIN_GIP
3993	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3994	#else
3995	LogFlow(( "SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3996	#endif
3997	return rc;
3998	}
3999
4000
4001	/**
4002	* Unmaps any user mapping of the GIP and terminates all GIP access
4003	* from this session.
4004	*
4005	* @returns IPRT status code.
4006	* @param pSession Session to which the GIP mapping should belong.
4007	*/
4008	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
4009	{
4010	int rc = VINF_SUCCESS;
4011	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
4012	#ifdef DEBUG_DARWIN_GIP
4013	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
4014	pSession,
4015	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
4016	pSession->GipMapObjR3));
4017	#else
4018	LogFlow(("SUPR0GipUnmap: pSession=%p\n", pSession));
4019	#endif
4020	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
4021
4022	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
4023	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
4024	#else
4025	RTSemFastMutexRequest(pDevExt->mtxGip);
4026	#endif
4027
4028	/*
4029	* Unmap anything?
4030	*/
4031	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
4032	{
4033	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
4034	AssertRC(rc);
4035	if (RT_SUCCESS(rc))
4036	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
4037	}
4038
4039	/*
4040	* Dereference global GIP.
4041	*/
4042	if (pSession->fGipReferenced && !rc)
4043	{
4044	pSession->fGipReferenced = 0;
4045	if ( pDevExt->cGipUsers > 0
4046	&& !--pDevExt->cGipUsers)
4047	{
4048	LogFlow(("SUPR0GipUnmap: Suspends GIP updating\n"));
4049	#ifndef DO_NOT_START_GIP
4050	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = VINF_SUCCESS;
4051	#endif
4052
4053	if (pDevExt->u32SystemTimerGranularityGrant)
4054	{
4055	int rc2 = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant);
4056	AssertRC(rc2);
4057	pDevExt->u32SystemTimerGranularityGrant = 0;
4058	}
4059	}
4060	}
4061
4062	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
4063	RTSemMutexRelease(pDevExt->mtxGip);
4064	#else
4065	RTSemFastMutexRelease(pDevExt->mtxGip);
4066	#endif
4067
4068	return rc;
4069	}
4070
4071
4072	/**
4073	* Gets the GIP pointer.
4074	*
4075	* @returns Pointer to the GIP or NULL.
4076	*/
4077	SUPDECL(PSUPGLOBALINFOPAGE) SUPGetGIP(void)
4078	{
4079	return g_pSUPGlobalInfoPage;
4080	}
4081
4082
4083	/**
4084	* Register a component factory with the support driver.
4085	*
4086	* This is currently restricted to kernel sessions only.
4087	*
4088	* @returns VBox status code.
4089	* @retval VINF_SUCCESS on success.
4090	* @retval VERR_NO_MEMORY if we're out of memory.
4091	* @retval VERR_ALREADY_EXISTS if the factory has already been registered.
4092	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
4093	* @retval VERR_INVALID_PARAMETER on invalid parameter.
4094	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
4095	*
4096	* @param pSession The SUPDRV session (must be a ring-0 session).
4097	* @param pFactory Pointer to the component factory registration structure.
4098	*
4099	* @remarks This interface is also available via SUPR0IdcComponentRegisterFactory.
4100	*/
4101	SUPR0DECL(int) SUPR0ComponentRegisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
4102	{
4103	PSUPDRVFACTORYREG pNewReg;
4104	const char *psz;
4105	int rc;
4106
4107	/*
4108	* Validate parameters.
4109	*/
4110	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
4111	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
4112	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
4113	AssertPtrReturn(pFactory->pfnQueryFactoryInterface, VERR_INVALID_POINTER);
4114	psz = RTStrEnd(pFactory->szName, sizeof(pFactory->szName));
4115	AssertReturn(psz, VERR_INVALID_PARAMETER);
4116
4117	/*
4118	* Allocate and initialize a new registration structure.
4119	*/
4120	pNewReg = (PSUPDRVFACTORYREG)RTMemAlloc(sizeof(SUPDRVFACTORYREG));
4121	if (pNewReg)
4122	{
4123	pNewReg->pNext = NULL;
4124	pNewReg->pFactory = pFactory;
4125	pNewReg->pSession = pSession;
4126	pNewReg->cchName = psz - &pFactory->szName[0];
4127
4128	/*
4129	* Add it to the tail of the list after checking for prior registration.
4130	*/
4131	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
4132	if (RT_SUCCESS(rc))
4133	{
4134	PSUPDRVFACTORYREG pPrev = NULL;
4135	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
4136	while (pCur && pCur->pFactory != pFactory)
4137	{
4138	pPrev = pCur;
4139	pCur = pCur->pNext;
4140	}
4141	if (!pCur)
4142	{
4143	if (pPrev)
4144	pPrev->pNext = pNewReg;
4145	else
4146	pSession->pDevExt->pComponentFactoryHead = pNewReg;
4147	rc = VINF_SUCCESS;
4148	}
4149	else
4150	rc = VERR_ALREADY_EXISTS;
4151
4152	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
4153	}
4154
4155	if (RT_FAILURE(rc))
4156	RTMemFree(pNewReg);
4157	}
4158	else
4159	rc = VERR_NO_MEMORY;
4160	return rc;
4161	}
4162
4163
4164	/**
4165	* Deregister a component factory.
4166	*
4167	* @returns VBox status code.
4168	* @retval VINF_SUCCESS on success.
4169	* @retval VERR_NOT_FOUND if the factory wasn't registered.
4170	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
4171	* @retval VERR_INVALID_PARAMETER on invalid parameter.
4172	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
4173	*
4174	* @param pSession The SUPDRV session (must be a ring-0 session).
4175	* @param pFactory Pointer to the component factory registration structure
4176	* previously passed SUPR0ComponentRegisterFactory().
4177	*
4178	* @remarks This interface is also available via SUPR0IdcComponentDeregisterFactory.
4179	*/
4180	SUPR0DECL(int) SUPR0ComponentDeregisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
4181	{
4182	int rc;
4183
4184	/*
4185	* Validate parameters.
4186	*/
4187	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
4188	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
4189	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
4190
4191	/*
4192	* Take the lock and look for the registration record.
4193	*/
4194	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
4195	if (RT_SUCCESS(rc))
4196	{
4197	PSUPDRVFACTORYREG pPrev = NULL;
4198	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
4199	while (pCur && pCur->pFactory != pFactory)
4200	{
4201	pPrev = pCur;
4202	pCur = pCur->pNext;
4203	}
4204	if (pCur)
4205	{
4206	if (!pPrev)
4207	pSession->pDevExt->pComponentFactoryHead = pCur->pNext;
4208	else
4209	pPrev->pNext = pCur->pNext;
4210
4211	pCur->pNext = NULL;
4212	pCur->pFactory = NULL;
4213	pCur->pSession = NULL;
4214	rc = VINF_SUCCESS;
4215	}
4216	else
4217	rc = VERR_NOT_FOUND;
4218
4219	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
4220
4221	RTMemFree(pCur);
4222	}
4223	return rc;
4224	}
4225
4226
4227	/**
4228	* Queries a component factory.
4229	*
4230	* @returns VBox status code.
4231	* @retval VERR_INVALID_PARAMETER on invalid parameter.
4232	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
4233	* @retval VERR_SUPDRV_COMPONENT_NOT_FOUND if the component factory wasn't found.
4234	* @retval VERR_SUPDRV_INTERFACE_NOT_SUPPORTED if the interface wasn't supported.
4235	*
4236	* @param pSession The SUPDRV session.
4237	* @param pszName The name of the component factory.
4238	* @param pszInterfaceUuid The UUID of the factory interface (stringified).
4239	* @param ppvFactoryIf Where to store the factory interface.
4240	*/
4241	SUPR0DECL(int) SUPR0ComponentQueryFactory(PSUPDRVSESSION pSession, const char pszName, const char pszInterfaceUuid, void **ppvFactoryIf)
4242	{
4243	const char *pszEnd;
4244	size_t cchName;
4245	int rc;
4246
4247	/*
4248	* Validate parameters.
4249	*/
4250	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
4251
4252	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
4253	pszEnd = RTStrEnd(pszName, RT_SIZEOFMEMB(SUPDRVFACTORY, szName));
4254	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4255	cchName = pszEnd - pszName;
4256
4257	AssertPtrReturn(pszInterfaceUuid, VERR_INVALID_POINTER);
4258	pszEnd = RTStrEnd(pszInterfaceUuid, RTUUID_STR_LENGTH);
4259	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4260
4261	AssertPtrReturn(ppvFactoryIf, VERR_INVALID_POINTER);
4262	*ppvFactoryIf = NULL;
4263
4264	/*
4265	* Take the lock and try all factories by this name.
4266	*/
4267	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
4268	if (RT_SUCCESS(rc))
4269	{
4270	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
4271	rc = VERR_SUPDRV_COMPONENT_NOT_FOUND;
4272	while (pCur)
4273	{
4274	if ( pCur->cchName == cchName
4275	&& !memcmp(pCur->pFactory->szName, pszName, cchName))
4276	{
4277	void *pvFactory = pCur->pFactory->pfnQueryFactoryInterface(pCur->pFactory, pSession, pszInterfaceUuid);
4278	if (pvFactory)
4279	{
4280	*ppvFactoryIf = pvFactory;
4281	rc = VINF_SUCCESS;
4282	break;
4283	}
4284	rc = VERR_SUPDRV_INTERFACE_NOT_SUPPORTED;
4285	}
4286
4287	/* next */
4288	pCur = pCur->pNext;
4289	}
4290
4291	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
4292	}
4293	return rc;
4294	}
4295
4296
4297	/**
4298	* Adds a memory object to the session.
4299	*
4300	* @returns IPRT status code.
4301	* @param pMem Memory tracking structure containing the
4302	* information to track.
4303	* @param pSession The session.
4304	*/
4305	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
4306	{
4307	PSUPDRVBUNDLE pBundle;
4308
4309	/*
4310	* Find free entry and record the allocation.
4311	*/
4312	RTSpinlockAcquire(pSession->Spinlock);
4313	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
4314	{
4315	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
4316	{
4317	unsigned i;
4318	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
4319	{
4320	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
4321	{
4322	pBundle->cUsed++;
4323	pBundle->aMem[i] = *pMem;
4324	RTSpinlockRelease(pSession->Spinlock);
4325	return VINF_SUCCESS;
4326	}
4327	}
4328	AssertFailed(); /* !!this can't be happening!!! */
4329	}
4330	}
4331	RTSpinlockRelease(pSession->Spinlock);
4332
4333	/*
4334	* Need to allocate a new bundle.
4335	* Insert into the last entry in the bundle.
4336	*/
4337	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
4338	if (!pBundle)
4339	return VERR_NO_MEMORY;
4340
4341	/* take last entry. */
4342	pBundle->cUsed++;
4343	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
4344
4345	/* insert into list. */
4346	RTSpinlockAcquire(pSession->Spinlock);
4347	pBundle->pNext = pSession->Bundle.pNext;
4348	pSession->Bundle.pNext = pBundle;
4349	RTSpinlockRelease(pSession->Spinlock);
4350
4351	return VINF_SUCCESS;
4352	}
4353
4354
4355	/**
4356	* Releases a memory object referenced by pointer and type.
4357	*
4358	* @returns IPRT status code.
4359	* @param pSession Session data.
4360	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
4361	* @param eType Memory type.
4362	*/
4363	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
4364	{
4365	PSUPDRVBUNDLE pBundle;
4366
4367	/*
4368	* Validate input.
4369	*/
4370	if (!uPtr)
4371	{
4372	Log(("Illegal address %p\n", (void *)uPtr));
4373	return VERR_INVALID_PARAMETER;
4374	}
4375
4376	/*
4377	* Search for the address.
4378	*/
4379	RTSpinlockAcquire(pSession->Spinlock);
4380	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
4381	{
4382	if (pBundle->cUsed > 0)
4383	{
4384	unsigned i;
4385	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
4386	{
4387	if ( pBundle->aMem[i].eType == eType
4388	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
4389	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
4390	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
4391	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
4392	)
4393	{
4394	/* Make a copy of it and release it outside the spinlock. */
4395	SUPDRVMEMREF Mem = pBundle->aMem[i];
4396	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
4397	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
4398	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
4399	RTSpinlockRelease(pSession->Spinlock);
4400
4401	if (Mem.MapObjR3 != NIL_RTR0MEMOBJ)
4402	{
4403	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
4404	AssertRC(rc); /** @todo figure out how to handle this. */
4405	}
4406	if (Mem.MemObj != NIL_RTR0MEMOBJ)
4407	{
4408	int rc = RTR0MemObjFree(Mem.MemObj, true /* fFreeMappings */);
4409	AssertRC(rc); /** @todo figure out how to handle this. */
4410	}
4411	return VINF_SUCCESS;
4412	}
4413	}
4414	}
4415	}
4416	RTSpinlockRelease(pSession->Spinlock);
4417	Log(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
4418	return VERR_INVALID_PARAMETER;
4419	}
4420
4421
4422	/**
4423	* Opens an image. If it's the first time it's opened the call must upload
4424	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
4425	*
4426	* This is the 1st step of the loading.
4427	*
4428	* @returns IPRT status code.
4429	* @param pDevExt Device globals.
4430	* @param pSession Session data.
4431	* @param pReq The open request.
4432	*/
4433	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
4434	{
4435	int rc;
4436	PSUPDRVLDRIMAGE pImage;
4437	void *pv;
4438	size_t cchName = strlen(pReq->u.In.szName); /* (caller checked < 32). */
4439	LogFlow(("supdrvIOCtl_LdrOpen: szName=%s cbImageWithTabs=%d\n", pReq->u.In.szName, pReq->u.In.cbImageWithTabs));
4440
4441	/*
4442	* Check if we got an instance of the image already.
4443	*/
4444	supdrvLdrLock(pDevExt);
4445	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
4446	{
4447	if ( pImage->szName[cchName] == '\0'
4448	&& !memcmp(pImage->szName, pReq->u.In.szName, cchName))
4449	{
4450	/** @todo check cbImageBits and cbImageWithTabs here, if they differs that indicates that the images are different. */
4451	pImage->cUsage++;
4452	pReq->u.Out.pvImageBase = pImage->pvImage;
4453	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
4454	pReq->u.Out.fNativeLoader = pImage->fNative;
4455	supdrvLdrAddUsage(pSession, pImage);
4456	supdrvLdrUnlock(pDevExt);
4457	return VINF_SUCCESS;
4458	}
4459	}
4460	/* (not found - add it!) */
4461
4462	/*
4463	* Allocate memory.
4464	*/
4465	Assert(cchName < sizeof(pImage->szName));
4466	pv = RTMemAlloc(sizeof(SUPDRVLDRIMAGE));
4467	if (!pv)
4468	{
4469	supdrvLdrUnlock(pDevExt);
4470	Log(("supdrvIOCtl_LdrOpen: RTMemAlloc() failed\n"));
4471	return /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_2;
4472	}
4473
4474	/*
4475	* Setup and link in the LDR stuff.
4476	*/
4477	pImage = (PSUPDRVLDRIMAGE)pv;
4478	pImage->pvImage = NULL;
4479	pImage->pvImageAlloc = NULL;
4480	pImage->cbImageWithTabs = pReq->u.In.cbImageWithTabs;
4481	pImage->cbImageBits = pReq->u.In.cbImageBits;
4482	pImage->cSymbols = 0;
4483	pImage->paSymbols = NULL;
4484	pImage->pachStrTab = NULL;
4485	pImage->cbStrTab = 0;
4486	pImage->pfnModuleInit = NULL;
4487	pImage->pfnModuleTerm = NULL;
4488	pImage->pfnServiceReqHandler = NULL;
4489	pImage->uState = SUP_IOCTL_LDR_OPEN;
4490	pImage->cUsage = 1;
4491	pImage->pDevExt = pDevExt;
4492	memcpy(pImage->szName, pReq->u.In.szName, cchName + 1);
4493
4494	/*
4495	* Try load it using the native loader, if that isn't supported, fall back
4496	* on the older method.
4497	*/
4498	pImage->fNative = true;
4499	rc = supdrvOSLdrOpen(pDevExt, pImage, pReq->u.In.szFilename);
4500	if (rc == VERR_NOT_SUPPORTED)
4501	{
4502	pImage->pvImageAlloc = RTMemExecAlloc(pImage->cbImageBits + 31);
4503	pImage->pvImage = RT_ALIGN_P(pImage->pvImageAlloc, 32);
4504	pImage->fNative = false;
4505	rc = pImage->pvImageAlloc ? VINF_SUCCESS : VERR_NO_EXEC_MEMORY;
4506	}
4507	if (RT_FAILURE(rc))
4508	{
4509	supdrvLdrUnlock(pDevExt);
4510	RTMemFree(pImage);
4511	Log(("supdrvIOCtl_LdrOpen(%s): failed - %Rrc\n", pReq->u.In.szName, rc));
4512	return rc;
4513	}
4514	Assert(VALID_PTR(pImage->pvImage) \|\| RT_FAILURE(rc));
4515
4516	/*
4517	* Link it.
4518	*/
4519	pImage->pNext = pDevExt->pLdrImages;
4520	pDevExt->pLdrImages = pImage;
4521
4522	supdrvLdrAddUsage(pSession, pImage);
4523
4524	pReq->u.Out.pvImageBase = pImage->pvImage;
4525	pReq->u.Out.fNeedsLoading = true;
4526	pReq->u.Out.fNativeLoader = pImage->fNative;
4527	supdrvOSLdrNotifyOpened(pDevExt, pImage);
4528
4529	supdrvLdrUnlock(pDevExt);
4530	return VINF_SUCCESS;
4531	}
4532
4533
4534	/**
4535	* Worker that validates a pointer to an image entrypoint.
4536	*
4537	* @returns IPRT status code.
4538	* @param pDevExt The device globals.
4539	* @param pImage The loader image.
4540	* @param pv The pointer into the image.
4541	* @param fMayBeNull Whether it may be NULL.
4542	* @param pszWhat What is this entrypoint? (for logging)
4543	* @param pbImageBits The image bits prepared by ring-3.
4544	*
4545	* @remarks Will leave the lock on failure.
4546	*/
4547	static int supdrvLdrValidatePointer(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage, void *pv,
4548	bool fMayBeNull, const uint8_t pbImageBits, const char pszWhat)
4549	{
4550	if (!fMayBeNull \|\| pv)
4551	{
4552	if ((uintptr_t)pv - (uintptr_t)pImage->pvImage >= pImage->cbImageBits)
4553	{
4554	supdrvLdrUnlock(pDevExt);
4555	Log(("Out of range (%p LB %#x): %s=%p\n", pImage->pvImage, pImage->cbImageBits, pszWhat, pv));
4556	return VERR_INVALID_PARAMETER;
4557	}
4558
4559	if (pImage->fNative)
4560	{
4561	int rc = supdrvOSLdrValidatePointer(pDevExt, pImage, pv, pbImageBits);
4562	if (RT_FAILURE(rc))
4563	{
4564	supdrvLdrUnlock(pDevExt);
4565	Log(("Bad entry point address: %s=%p (rc=%Rrc)\n", pszWhat, pv, rc));
4566	return rc;
4567	}
4568	}
4569	}
4570	return VINF_SUCCESS;
4571	}
4572
4573
4574	/**
4575	* Loads the image bits.
4576	*
4577	* This is the 2nd step of the loading.
4578	*
4579	* @returns IPRT status code.
4580	* @param pDevExt Device globals.
4581	* @param pSession Session data.
4582	* @param pReq The request.
4583	*/
4584	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
4585	{
4586	PSUPDRVLDRUSAGE pUsage;
4587	PSUPDRVLDRIMAGE pImage;
4588	int rc;
4589	LogFlow(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImageWithBits=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImageWithTabs));
4590
4591	/*
4592	* Find the ldr image.
4593	*/
4594	supdrvLdrLock(pDevExt);
4595	pUsage = pSession->pLdrUsage;
4596	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4597	pUsage = pUsage->pNext;
4598	if (!pUsage)
4599	{
4600	supdrvLdrUnlock(pDevExt);
4601	Log(("SUP_IOCTL_LDR_LOAD: couldn't find image!\n"));
4602	return VERR_INVALID_HANDLE;
4603	}
4604	pImage = pUsage->pImage;
4605
4606	/*
4607	* Validate input.
4608	*/
4609	if ( pImage->cbImageWithTabs != pReq->u.In.cbImageWithTabs
4610	\|\| pImage->cbImageBits != pReq->u.In.cbImageBits)
4611	{
4612	supdrvLdrUnlock(pDevExt);
4613	Log(("SUP_IOCTL_LDR_LOAD: image size mismatch!! %d(prep) != %d(load) or %d != %d\n",
4614	pImage->cbImageWithTabs, pReq->u.In.cbImageWithTabs, pImage->cbImageBits, pReq->u.In.cbImageBits));
4615	return VERR_INVALID_HANDLE;
4616	}
4617
4618	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
4619	{
4620	unsigned uState = pImage->uState;
4621	supdrvLdrUnlock(pDevExt);
4622	if (uState != SUP_IOCTL_LDR_LOAD)
4623	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
4624	return VERR_ALREADY_LOADED;
4625	}
4626
4627	switch (pReq->u.In.eEPType)
4628	{
4629	case SUPLDRLOADEP_NOTHING:
4630	break;
4631
4632	case SUPLDRLOADEP_VMMR0:
4633	rc = supdrvLdrValidatePointer( pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0, false, pReq->u.In.abImage, "pvVMMR0");
4634	if (RT_SUCCESS(rc))
4635	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt, false, pReq->u.In.abImage, "pvVMMR0EntryInt");
4636	if (RT_SUCCESS(rc))
4637	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, false, pReq->u.In.abImage, "pvVMMR0EntryFast");
4638	if (RT_SUCCESS(rc))
4639	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx, false, pReq->u.In.abImage, "pvVMMR0EntryEx");
4640	if (RT_FAILURE(rc))
4641	return rc;
4642	break;
4643
4644	case SUPLDRLOADEP_SERVICE:
4645	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.Service.pfnServiceReq, false, pReq->u.In.abImage, "pfnServiceReq");
4646	if (RT_FAILURE(rc))
4647	return rc;
4648	if ( pReq->u.In.EP.Service.apvReserved[0] != NIL_RTR0PTR
4649	\|\| pReq->u.In.EP.Service.apvReserved[1] != NIL_RTR0PTR
4650	\|\| pReq->u.In.EP.Service.apvReserved[2] != NIL_RTR0PTR)
4651	{
4652	supdrvLdrUnlock(pDevExt);
4653	Log(("Out of range (%p LB %#x): apvReserved={%p,%p,%p} MBZ!\n",
4654	pImage->pvImage, pReq->u.In.cbImageWithTabs,
4655	pReq->u.In.EP.Service.apvReserved[0],
4656	pReq->u.In.EP.Service.apvReserved[1],
4657	pReq->u.In.EP.Service.apvReserved[2]));
4658	return VERR_INVALID_PARAMETER;
4659	}
4660	break;
4661
4662	default:
4663	supdrvLdrUnlock(pDevExt);
4664	Log(("Invalid eEPType=%d\n", pReq->u.In.eEPType));
4665	return VERR_INVALID_PARAMETER;
4666	}
4667
4668	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleInit, true, pReq->u.In.abImage, "pfnModuleInit");
4669	if (RT_FAILURE(rc))
4670	return rc;
4671	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleTerm, true, pReq->u.In.abImage, "pfnModuleTerm");
4672	if (RT_FAILURE(rc))
4673	return rc;
4674
4675	/*
4676	* Allocate and copy the tables.
4677	* (No need to do try/except as this is a buffered request.)
4678	*/
4679	pImage->cbStrTab = pReq->u.In.cbStrTab;
4680	if (pImage->cbStrTab)
4681	{
4682	pImage->pachStrTab = (char *)RTMemAlloc(pImage->cbStrTab);
4683	if (pImage->pachStrTab)
4684	memcpy(pImage->pachStrTab, &pReq->u.In.abImage[pReq->u.In.offStrTab], pImage->cbStrTab);
4685	else
4686	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_3;
4687	}
4688
4689	pImage->cSymbols = pReq->u.In.cSymbols;
4690	if (RT_SUCCESS(rc) && pImage->cSymbols)
4691	{
4692	size_t cbSymbols = pImage->cSymbols * sizeof(SUPLDRSYM);
4693	pImage->paSymbols = (PSUPLDRSYM)RTMemAlloc(cbSymbols);
4694	if (pImage->paSymbols)
4695	memcpy(pImage->paSymbols, &pReq->u.In.abImage[pReq->u.In.offSymbols], cbSymbols);
4696	else
4697	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_4;
4698	}
4699
4700	/*
4701	* Copy the bits / complete native loading.
4702	*/
4703	if (RT_SUCCESS(rc))
4704	{
4705	pImage->uState = SUP_IOCTL_LDR_LOAD;
4706	pImage->pfnModuleInit = pReq->u.In.pfnModuleInit;
4707	pImage->pfnModuleTerm = pReq->u.In.pfnModuleTerm;
4708
4709	if (pImage->fNative)
4710	rc = supdrvOSLdrLoad(pDevExt, pImage, pReq->u.In.abImage, pReq);
4711	else
4712	{
4713	memcpy(pImage->pvImage, &pReq->u.In.abImage[0], pImage->cbImageBits);
4714	Log(("vboxdrv: Loaded '%s' at %p\n", pImage->szName, pImage->pvImage));
4715	}
4716	}
4717
4718	/*
4719	* Update any entry points.
4720	*/
4721	if (RT_SUCCESS(rc))
4722	{
4723	switch (pReq->u.In.eEPType)
4724	{
4725	default:
4726	case SUPLDRLOADEP_NOTHING:
4727	rc = VINF_SUCCESS;
4728	break;
4729	case SUPLDRLOADEP_VMMR0:
4730	rc = supdrvLdrSetVMMR0EPs(pDevExt, pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
4731	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
4732	break;
4733	case SUPLDRLOADEP_SERVICE:
4734	pImage->pfnServiceReqHandler = pReq->u.In.EP.Service.pfnServiceReq;
4735	rc = VINF_SUCCESS;
4736	break;
4737	}
4738	}
4739
4740	/*
4741	* On success call the module initialization.
4742	*/
4743	LogFlow(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
4744	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
4745	{
4746	Log(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
4747	pDevExt->pLdrInitImage = pImage;
4748	pDevExt->hLdrInitThread = RTThreadNativeSelf();
4749	rc = pImage->pfnModuleInit(pImage);
4750	pDevExt->pLdrInitImage = NULL;
4751	pDevExt->hLdrInitThread = NIL_RTNATIVETHREAD;
4752	if (RT_FAILURE(rc) && pDevExt->pvVMMR0 == pImage->pvImage)
4753	supdrvLdrUnsetVMMR0EPs(pDevExt);
4754	}
4755	SUPR0Printf("vboxdrv: %p %s\n", pImage->pvImage, pImage->szName);
4756
4757	if (RT_FAILURE(rc))
4758	{
4759	/* Inform the tracing component in case ModuleInit registered TPs. */
4760	supdrvTracerModuleUnloading(pDevExt, pImage);
4761
4762	pImage->uState = SUP_IOCTL_LDR_OPEN;
4763	pImage->pfnModuleInit = NULL;
4764	pImage->pfnModuleTerm = NULL;
4765	pImage->pfnServiceReqHandler= NULL;
4766	pImage->cbStrTab = 0;
4767	RTMemFree(pImage->pachStrTab);
4768	pImage->pachStrTab = NULL;
4769	RTMemFree(pImage->paSymbols);
4770	pImage->paSymbols = NULL;
4771	pImage->cSymbols = 0;
4772	}
4773
4774	supdrvLdrUnlock(pDevExt);
4775	return rc;
4776	}
4777
4778
4779	/**
4780	* Frees a previously loaded (prep'ed) image.
4781	*
4782	* @returns IPRT status code.
4783	* @param pDevExt Device globals.
4784	* @param pSession Session data.
4785	* @param pReq The request.
4786	*/
4787	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
4788	{
4789	int rc;
4790	PSUPDRVLDRUSAGE pUsagePrev;
4791	PSUPDRVLDRUSAGE pUsage;
4792	PSUPDRVLDRIMAGE pImage;
4793	LogFlow(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
4794
4795	/*
4796	* Find the ldr image.
4797	*/
4798	supdrvLdrLock(pDevExt);
4799	pUsagePrev = NULL;
4800	pUsage = pSession->pLdrUsage;
4801	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4802	{
4803	pUsagePrev = pUsage;
4804	pUsage = pUsage->pNext;
4805	}
4806	if (!pUsage)
4807	{
4808	supdrvLdrUnlock(pDevExt);
4809	Log(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
4810	return VERR_INVALID_HANDLE;
4811	}
4812
4813	/*
4814	* Check if we can remove anything.
4815	*/
4816	rc = VINF_SUCCESS;
4817	pImage = pUsage->pImage;
4818	if (pImage->cUsage <= 1 \|\| pUsage->cUsage <= 1)
4819	{
4820	/*
4821	* Check if there are any objects with destructors in the image, if
4822	* so leave it for the session cleanup routine so we get a chance to
4823	* clean things up in the right order and not leave them all dangling.
4824	*/
4825	RTSpinlockAcquire(pDevExt->Spinlock);
4826	if (pImage->cUsage <= 1)
4827	{
4828	PSUPDRVOBJ pObj;
4829	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4830	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4831	{
4832	rc = VERR_DANGLING_OBJECTS;
4833	break;
4834	}
4835	}
4836	else
4837	{
4838	PSUPDRVUSAGE pGenUsage;
4839	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
4840	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4841	{
4842	rc = VERR_DANGLING_OBJECTS;
4843	break;
4844	}
4845	}
4846	RTSpinlockRelease(pDevExt->Spinlock);
4847	if (rc == VINF_SUCCESS)
4848	{
4849	/* unlink it */
4850	if (pUsagePrev)
4851	pUsagePrev->pNext = pUsage->pNext;
4852	else
4853	pSession->pLdrUsage = pUsage->pNext;
4854
4855	/* free it */
4856	pUsage->pImage = NULL;
4857	pUsage->pNext = NULL;
4858	RTMemFree(pUsage);
4859
4860	/*
4861	* Dereference the image.
4862	*/
4863	if (pImage->cUsage <= 1)
4864	supdrvLdrFree(pDevExt, pImage);
4865	else
4866	pImage->cUsage--;
4867	}
4868	else
4869	{
4870	Log(("supdrvIOCtl_LdrFree: Dangling objects in %p/%s!\n", pImage->pvImage, pImage->szName));
4871	rc = VINF_SUCCESS; /** @todo BRANCH-2.1: remove this after branching. */
4872	}
4873	}
4874	else
4875	{
4876	/*
4877	* Dereference both image and usage.
4878	*/
4879	pImage->cUsage--;
4880	pUsage->cUsage--;
4881	}
4882
4883	supdrvLdrUnlock(pDevExt);
4884	return rc;
4885	}
4886
4887
4888	/**
4889	* Gets the address of a symbol in an open image.
4890	*
4891	* @returns IPRT status code.
4892	* @param pDevExt Device globals.
4893	* @param pSession Session data.
4894	* @param pReq The request buffer.
4895	*/
4896	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
4897	{
4898	PSUPDRVLDRIMAGE pImage;
4899	PSUPDRVLDRUSAGE pUsage;
4900	uint32_t i;
4901	PSUPLDRSYM paSyms;
4902	const char *pchStrings;
4903	const size_t cbSymbol = strlen(pReq->u.In.szSymbol) + 1;
4904	void *pvSymbol = NULL;
4905	int rc = VERR_GENERAL_FAILURE;
4906	Log3(("supdrvIOCtl_LdrGetSymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
4907
4908	/*
4909	* Find the ldr image.
4910	*/
4911	supdrvLdrLock(pDevExt);
4912	pUsage = pSession->pLdrUsage;
4913	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4914	pUsage = pUsage->pNext;
4915	if (!pUsage)
4916	{
4917	supdrvLdrUnlock(pDevExt);
4918	Log(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
4919	return VERR_INVALID_HANDLE;
4920	}
4921	pImage = pUsage->pImage;
4922	if (pImage->uState != SUP_IOCTL_LDR_LOAD)
4923	{
4924	unsigned uState = pImage->uState;
4925	supdrvLdrUnlock(pDevExt);
4926	Log(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", uState, uState)); NOREF(uState);
4927	return VERR_ALREADY_LOADED;
4928	}
4929
4930	/*
4931	* Search the symbol strings.
4932	*
4933	* Note! The int32_t is for native loading on solaris where the data
4934	* and text segments are in very different places.
4935	*/
4936	pchStrings = pImage->pachStrTab;
4937	paSyms = pImage->paSymbols;
4938	for (i = 0; i < pImage->cSymbols; i++)
4939	{
4940	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4941	&& !memcmp(pchStrings + paSyms[i].offName, pReq->u.In.szSymbol, cbSymbol))
4942	{
4943	pvSymbol = (uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol;
4944	rc = VINF_SUCCESS;
4945	break;
4946	}
4947	}
4948	supdrvLdrUnlock(pDevExt);
4949	pReq->u.Out.pvSymbol = pvSymbol;
4950	return rc;
4951	}
4952
4953
4954	/**
4955	* Gets the address of a symbol in an open image or the support driver.
4956	*
4957	* @returns VINF_SUCCESS on success.
4958	* @returns
4959	* @param pDevExt Device globals.
4960	* @param pSession Session data.
4961	* @param pReq The request buffer.
4962	*/
4963	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq)
4964	{
4965	int rc = VINF_SUCCESS;
4966	const char *pszSymbol = pReq->u.In.pszSymbol;
4967	const char *pszModule = pReq->u.In.pszModule;
4968	size_t cbSymbol;
4969	char const *pszEnd;
4970	uint32_t i;
4971
4972	/*
4973	* Input validation.
4974	*/
4975	AssertPtrReturn(pszSymbol, VERR_INVALID_POINTER);
4976	pszEnd = RTStrEnd(pszSymbol, 512);
4977	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4978	cbSymbol = pszEnd - pszSymbol + 1;
4979
4980	if (pszModule)
4981	{
4982	AssertPtrReturn(pszModule, VERR_INVALID_POINTER);
4983	pszEnd = RTStrEnd(pszModule, 64);
4984	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4985	}
4986	Log3(("supdrvIDC_LdrGetSymbol: pszModule=%p:{%s} pszSymbol=%p:{%s}\n", pszModule, pszModule, pszSymbol, pszSymbol));
4987
4988
4989	if ( !pszModule
4990	\|\| !strcmp(pszModule, "SupDrv"))
4991	{
4992	/*
4993	* Search the support driver export table.
4994	*/
4995	for (i = 0; i < RT_ELEMENTS(g_aFunctions); i++)
4996	if (!strcmp(g_aFunctions[i].szName, pszSymbol))
4997	{
4998	pReq->u.Out.pfnSymbol = g_aFunctions[i].pfn;
4999	break;
5000	}
5001	}
5002	else
5003	{
5004	/*
5005	* Find the loader image.
5006	*/
5007	PSUPDRVLDRIMAGE pImage;
5008
5009	supdrvLdrLock(pDevExt);
5010
5011	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
5012	if (!strcmp(pImage->szName, pszModule))
5013	break;
5014	if (pImage && pImage->uState == SUP_IOCTL_LDR_LOAD)
5015	{
5016	/*
5017	* Search the symbol strings.
5018	*/
5019	const char *pchStrings = pImage->pachStrTab;
5020	PCSUPLDRSYM paSyms = pImage->paSymbols;
5021	for (i = 0; i < pImage->cSymbols; i++)
5022	{
5023	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
5024	&& !memcmp(pchStrings + paSyms[i].offName, pszSymbol, cbSymbol))
5025	{
5026	/*
5027	* Found it! Calc the symbol address and add a reference to the module.
5028	*/
5029	pReq->u.Out.pfnSymbol = (PFNRT)((uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol);
5030	rc = supdrvLdrAddUsage(pSession, pImage);
5031	break;
5032	}
5033	}
5034	}
5035	else
5036	rc = pImage ? VERR_WRONG_ORDER : VERR_MODULE_NOT_FOUND;
5037
5038	supdrvLdrUnlock(pDevExt);
5039	}
5040	return rc;
5041	}
5042
5043
5044	/**
5045	* Updates the VMMR0 entry point pointers.
5046	*
5047	* @returns IPRT status code.
5048	* @param pDevExt Device globals.
5049	* @param pSession Session data.
5050	* @param pVMMR0 VMMR0 image handle.
5051	* @param pvVMMR0EntryInt VMMR0EntryInt address.
5052	* @param pvVMMR0EntryFast VMMR0EntryFast address.
5053	* @param pvVMMR0EntryEx VMMR0EntryEx address.
5054	* @remark Caller must own the loader mutex.
5055	*/
5056	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx)
5057	{
5058	int rc = VINF_SUCCESS;
5059	LogFlow(("supdrvLdrSetR0EP pvVMMR0=%p pvVMMR0EntryInt=%p\n", pvVMMR0, pvVMMR0EntryInt));
5060
5061
5062	/*
5063	* Check if not yet set.
5064	*/
5065	if (!pDevExt->pvVMMR0)
5066	{
5067	pDevExt->pvVMMR0 = pvVMMR0;
5068	pDevExt->pfnVMMR0EntryInt = pvVMMR0EntryInt;
5069	pDevExt->pfnVMMR0EntryFast = pvVMMR0EntryFast;
5070	pDevExt->pfnVMMR0EntryEx = pvVMMR0EntryEx;
5071	}
5072	else
5073	{
5074	/*
5075	* Return failure or success depending on whether the values match or not.
5076	*/
5077	if ( pDevExt->pvVMMR0 != pvVMMR0
5078	\|\| (void *)pDevExt->pfnVMMR0EntryInt != pvVMMR0EntryInt
5079	\|\| (void *)pDevExt->pfnVMMR0EntryFast != pvVMMR0EntryFast
5080	\|\| (void *)pDevExt->pfnVMMR0EntryEx != pvVMMR0EntryEx)
5081	{
5082	AssertMsgFailed(("SUP_IOCTL_LDR_SETR0EP: Already set pointing to a different module!\n"));
5083	rc = VERR_INVALID_PARAMETER;
5084	}
5085	}
5086	return rc;
5087	}
5088
5089
5090	/**
5091	* Unsets the VMMR0 entry point installed by supdrvLdrSetR0EP.
5092	*
5093	* @param pDevExt Device globals.
5094	*/
5095	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt)
5096	{
5097	pDevExt->pvVMMR0 = NULL;
5098	pDevExt->pfnVMMR0EntryInt = NULL;
5099	pDevExt->pfnVMMR0EntryFast = NULL;
5100	pDevExt->pfnVMMR0EntryEx = NULL;
5101	}
5102
5103
5104	/**
5105	* Adds a usage reference in the specified session of an image.
5106	*
5107	* Called while owning the loader semaphore.
5108	*
5109	* @returns VINF_SUCCESS on success and VERR_NO_MEMORY on failure.
5110	* @param pSession Session in question.
5111	* @param pImage Image which the session is using.
5112	*/
5113	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage)
5114	{
5115	PSUPDRVLDRUSAGE pUsage;
5116	LogFlow(("supdrvLdrAddUsage: pImage=%p\n", pImage));
5117
5118	/*
5119	* Referenced it already?
5120	*/
5121	pUsage = pSession->pLdrUsage;
5122	while (pUsage)
5123	{
5124	if (pUsage->pImage == pImage)
5125	{
5126	pUsage->cUsage++;
5127	return VINF_SUCCESS;
5128	}
5129	pUsage = pUsage->pNext;
5130	}
5131
5132	/*
5133	* Allocate new usage record.
5134	*/
5135	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
5136	AssertReturn(pUsage, /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_5);
5137	pUsage->cUsage = 1;
5138	pUsage->pImage = pImage;
5139	pUsage->pNext = pSession->pLdrUsage;
5140	pSession->pLdrUsage = pUsage;
5141	return VINF_SUCCESS;
5142	}
5143
5144
5145	/**
5146	* Frees a load image.
5147	*
5148	* @param pDevExt Pointer to device extension.
5149	* @param pImage Pointer to the image we're gonna free.
5150	* This image must exit!
5151	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
5152	*/
5153	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
5154	{
5155	PSUPDRVLDRIMAGE pImagePrev;
5156	LogFlow(("supdrvLdrFree: pImage=%p\n", pImage));
5157
5158	/* find it - arg. should've used doubly linked list. */
5159	Assert(pDevExt->pLdrImages);
5160	pImagePrev = NULL;
5161	if (pDevExt->pLdrImages != pImage)
5162	{
5163	pImagePrev = pDevExt->pLdrImages;
5164	while (pImagePrev->pNext != pImage)
5165	pImagePrev = pImagePrev->pNext;
5166	Assert(pImagePrev->pNext == pImage);
5167	}
5168
5169	/* unlink */
5170	if (pImagePrev)
5171	pImagePrev->pNext = pImage->pNext;
5172	else
5173	pDevExt->pLdrImages = pImage->pNext;
5174
5175	/* check if this is VMMR0.r0 unset its entry point pointers. */
5176	if (pDevExt->pvVMMR0 == pImage->pvImage)
5177	supdrvLdrUnsetVMMR0EPs(pDevExt);
5178
5179	/* check for objects with destructors in this image. (Shouldn't happen.) */
5180	if (pDevExt->pObjs)
5181	{
5182	unsigned cObjs = 0;
5183	PSUPDRVOBJ pObj;
5184	RTSpinlockAcquire(pDevExt->Spinlock);
5185	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
5186	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
5187	{
5188	pObj->pfnDestructor = NULL;
5189	cObjs++;
5190	}
5191	RTSpinlockRelease(pDevExt->Spinlock);
5192	if (cObjs)
5193	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
5194	}
5195
5196	/* call termination function if fully loaded. */
5197	if ( pImage->pfnModuleTerm
5198	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
5199	{
5200	LogFlow(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
5201	pImage->pfnModuleTerm(pImage);
5202	}
5203
5204	/* Inform the tracing component. */
5205	supdrvTracerModuleUnloading(pDevExt, pImage);
5206
5207	/* do native unload if appropriate. */
5208	if (pImage->fNative)
5209	supdrvOSLdrUnload(pDevExt, pImage);
5210
5211	/* free the image */
5212	pImage->cUsage = 0;
5213	pImage->pDevExt = NULL;
5214	pImage->pNext = NULL;
5215	pImage->uState = SUP_IOCTL_LDR_FREE;
5216	RTMemExecFree(pImage->pvImageAlloc, pImage->cbImageBits + 31);
5217	pImage->pvImageAlloc = NULL;
5218	RTMemFree(pImage->pachStrTab);
5219	pImage->pachStrTab = NULL;
5220	RTMemFree(pImage->paSymbols);
5221	pImage->paSymbols = NULL;
5222	RTMemFree(pImage);
5223	}
5224
5225
5226	/**
5227	* Acquires the loader lock.
5228	*
5229	* @returns IPRT status code.
5230	* @param pDevExt The device extension.
5231	*/
5232	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt)
5233	{
5234	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
5235	int rc = RTSemMutexRequest(pDevExt->mtxLdr, RT_INDEFINITE_WAIT);
5236	#else
5237	int rc = RTSemFastMutexRequest(pDevExt->mtxLdr);
5238	#endif
5239	AssertRC(rc);
5240	return rc;
5241	}
5242
5243
5244	/**
5245	* Releases the loader lock.
5246	*
5247	* @returns IPRT status code.
5248	* @param pDevExt The device extension.
5249	*/
5250	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt)
5251	{
5252	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
5253	return RTSemMutexRelease(pDevExt->mtxLdr);
5254	#else
5255	return RTSemFastMutexRelease(pDevExt->mtxLdr);
5256	#endif
5257	}
5258
5259
5260	/**
5261	* Implements the service call request.
5262	*
5263	* @returns VBox status code.
5264	* @param pDevExt The device extension.
5265	* @param pSession The calling session.
5266	* @param pReq The request packet, valid.
5267	*/
5268	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq)
5269	{
5270	#if !defined(RT_OS_WINDOWS) \|\| defined(RT_ARCH_AMD64) \|\| defined(DEBUG)
5271	int rc;
5272
5273	/*
5274	* Find the module first in the module referenced by the calling session.
5275	*/
5276	rc = supdrvLdrLock(pDevExt);
5277	if (RT_SUCCESS(rc))
5278	{
5279	PFNSUPR0SERVICEREQHANDLER pfnServiceReqHandler = NULL;
5280	PSUPDRVLDRUSAGE pUsage;
5281
5282	for (pUsage = pSession->pLdrUsage; pUsage; pUsage = pUsage->pNext)
5283	if ( pUsage->pImage->pfnServiceReqHandler
5284	&& !strcmp(pUsage->pImage->szName, pReq->u.In.szName))
5285	{
5286	pfnServiceReqHandler = pUsage->pImage->pfnServiceReqHandler;
5287	break;
5288	}
5289	supdrvLdrUnlock(pDevExt);
5290
5291	if (pfnServiceReqHandler)
5292	{
5293	/*
5294	* Call it.
5295	*/
5296	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
5297	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, NULL);
5298	else
5299	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0]);
5300	}
5301	else
5302	rc = VERR_SUPDRV_SERVICE_NOT_FOUND;
5303	}
5304
5305	/* log it */
5306	if ( RT_FAILURE(rc)
5307	&& rc != VERR_INTERRUPTED
5308	&& rc != VERR_TIMEOUT)
5309	Log(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
5310	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
5311	else
5312	Log4(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
5313	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
5314	return rc;
5315	#else /* RT_OS_WINDOWS && !RT_ARCH_AMD64 && !DEBUG */
5316	return VERR_NOT_IMPLEMENTED;
5317	#endif /* RT_OS_WINDOWS && !RT_ARCH_AMD64 && !DEBUG */
5318	}
5319
5320
5321	/**
5322	* Implements the logger settings request.
5323	*
5324	* @returns VBox status code.
5325	* @param pDevExt The device extension.
5326	* @param pSession The caller's session.
5327	* @param pReq The request.
5328	*/
5329	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq)
5330	{
5331	const char *pszGroup = &pReq->u.In.szStrings[pReq->u.In.offGroups];
5332	const char *pszFlags = &pReq->u.In.szStrings[pReq->u.In.offFlags];
5333	const char *pszDest = &pReq->u.In.szStrings[pReq->u.In.offDestination];
5334	PRTLOGGER pLogger = NULL;
5335	int rc;
5336
5337	/*
5338	* Some further validation.
5339	*/
5340	switch (pReq->u.In.fWhat)
5341	{
5342	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
5343	case SUPLOGGERSETTINGS_WHAT_CREATE:
5344	break;
5345
5346	case SUPLOGGERSETTINGS_WHAT_DESTROY:
5347	if (pszGroup \|\| pszFlags \|\| *pszDest)
5348	return VERR_INVALID_PARAMETER;
5349	if (pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_RELEASE)
5350	return VERR_ACCESS_DENIED;
5351	break;
5352
5353	default:
5354	return VERR_INTERNAL_ERROR;
5355	}
5356
5357	/*
5358	* Get the logger.
5359	*/
5360	switch (pReq->u.In.fWhich)
5361	{
5362	case SUPLOGGERSETTINGS_WHICH_DEBUG:
5363	pLogger = RTLogGetDefaultInstance();
5364	break;
5365
5366	case SUPLOGGERSETTINGS_WHICH_RELEASE:
5367	pLogger = RTLogRelDefaultInstance();
5368	break;
5369
5370	default:
5371	return VERR_INTERNAL_ERROR;
5372	}
5373
5374	/*
5375	* Do the job.
5376	*/
5377	switch (pReq->u.In.fWhat)
5378	{
5379	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
5380	if (pLogger)
5381	{
5382	rc = RTLogFlags(pLogger, pszFlags);
5383	if (RT_SUCCESS(rc))
5384	rc = RTLogGroupSettings(pLogger, pszGroup);
5385	NOREF(pszDest);
5386	}
5387	else
5388	rc = VERR_NOT_FOUND;
5389	break;
5390
5391	case SUPLOGGERSETTINGS_WHAT_CREATE:
5392	{
5393	if (pLogger)
5394	rc = VERR_ALREADY_EXISTS;
5395	else
5396	{
5397	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
5398
5399	rc = RTLogCreate(&pLogger,
5400	0 /* fFlags */,
5401	pszGroup,
5402	pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_DEBUG
5403	? "VBOX_LOG"
5404	: "VBOX_RELEASE_LOG",
5405	RT_ELEMENTS(s_apszGroups),
5406	s_apszGroups,
5407	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER,
5408	NULL);
5409	if (RT_SUCCESS(rc))
5410	{
5411	rc = RTLogFlags(pLogger, pszFlags);
5412	NOREF(pszDest);
5413	if (RT_SUCCESS(rc))
5414	{
5415	switch (pReq->u.In.fWhich)
5416	{
5417	case SUPLOGGERSETTINGS_WHICH_DEBUG:
5418	pLogger = RTLogSetDefaultInstance(pLogger);
5419	break;
5420	case SUPLOGGERSETTINGS_WHICH_RELEASE:
5421	pLogger = RTLogRelSetDefaultInstance(pLogger);
5422	break;
5423	}
5424	}
5425	RTLogDestroy(pLogger);
5426	}
5427	}
5428	break;
5429	}
5430
5431	case SUPLOGGERSETTINGS_WHAT_DESTROY:
5432	switch (pReq->u.In.fWhich)
5433	{
5434	case SUPLOGGERSETTINGS_WHICH_DEBUG:
5435	pLogger = RTLogSetDefaultInstance(NULL);
5436	break;
5437	case SUPLOGGERSETTINGS_WHICH_RELEASE:
5438	pLogger = RTLogRelSetDefaultInstance(NULL);
5439	break;
5440	}
5441	rc = RTLogDestroy(pLogger);
5442	break;
5443
5444	default:
5445	{
5446	rc = VERR_INTERNAL_ERROR;
5447	break;
5448	}
5449	}
5450
5451	return rc;
5452	}
5453
5454
5455	/**
5456	* Implements the MSR prober operations.
5457	*
5458	* @returns VBox status code.
5459	* @param pDevExt The device extension.
5460	* @param pReq The request.
5461	*/
5462	static int supdrvIOCtl_MsrProber(PSUPDRVDEVEXT pDevExt, PSUPMSRPROBER pReq)
5463	{
5464	#ifdef SUPDRV_WITH_MSR_PROBER
5465	RTCPUID const idCpu = pReq->u.In.idCpu == UINT32_MAX ? NIL_RTCPUID : pReq->u.In.idCpu;
5466	int rc;
5467
5468	switch (pReq->u.In.enmOp)
5469	{
5470	case SUPMSRPROBEROP_READ:
5471	{
5472	uint64_t uValue;
5473	rc = supdrvOSMsrProberRead(pReq->u.In.uMsr, idCpu, &uValue);
5474	if (RT_SUCCESS(rc))
5475	{
5476	pReq->u.Out.uResults.Read.uValue = uValue;
5477	pReq->u.Out.uResults.Read.fGp = false;
5478	}
5479	else if (rc == VERR_ACCESS_DENIED)
5480	{
5481	pReq->u.Out.uResults.Read.uValue = 0;
5482	pReq->u.Out.uResults.Read.fGp = true;
5483	rc = VINF_SUCCESS;
5484	}
5485	break;
5486	}
5487
5488	case SUPMSRPROBEROP_WRITE:
5489	rc = supdrvOSMsrProberWrite(pReq->u.In.uMsr, idCpu, pReq->u.In.uArgs.Write.uToWrite);
5490	if (RT_SUCCESS(rc))
5491	pReq->u.Out.uResults.Write.fGp = false;
5492	else if (rc == VERR_ACCESS_DENIED)
5493	{
5494	pReq->u.Out.uResults.Write.fGp = true;
5495	rc = VINF_SUCCESS;
5496	}
5497	break;
5498
5499	case SUPMSRPROBEROP_MODIFY:
5500	case SUPMSRPROBEROP_MODIFY_FASTER:
5501	rc = supdrvOSMsrProberModify(idCpu, pReq);
5502	break;
5503
5504	default:
5505	return VERR_INVALID_FUNCTION;
5506	}
5507	return rc;
5508	#else
5509	return VERR_NOT_IMPLEMENTED;
5510	#endif
5511	}
5512
5513
5514
5515	/**
5516	* Creates the GIP.
5517	*
5518	* @returns VBox status code.
5519	* @param pDevExt Instance data. GIP stuff may be updated.
5520	*/
5521	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
5522	{
5523	PSUPGLOBALINFOPAGE pGip;
5524	RTHCPHYS HCPhysGip;
5525	uint32_t u32SystemResolution;
5526	uint32_t u32Interval;
5527	unsigned cCpus;
5528	int rc;
5529
5530
5531	LogFlow(("supdrvGipCreate:\n"));
5532
5533	/* assert order */
5534	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
5535	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
5536	Assert(!pDevExt->pGipTimer);
5537
5538	/*
5539	* Check the CPU count.
5540	*/
5541	cCpus = RTMpGetArraySize();
5542	if ( cCpus > RTCPUSET_MAX_CPUS
5543	\|\| cCpus > 256 /ApicId is used for the mappings/)
5544	{
5545	SUPR0Printf("VBoxDrv: Too many CPUs (%u) for the GIP (max %u)\n", cCpus, RT_MIN(RTCPUSET_MAX_CPUS, 256));
5546	return VERR_TOO_MANY_CPUS;
5547	}
5548
5549	/*
5550	* Allocate a contiguous set of pages with a default kernel mapping.
5551	*/
5552	rc = RTR0MemObjAllocCont(&pDevExt->GipMemObj, RT_UOFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), false /fExecutable/);
5553	if (RT_FAILURE(rc))
5554	{
5555	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
5556	return rc;
5557	}
5558	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
5559	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
5560
5561	/*
5562	* Find a reasonable update interval and initialize the structure.
5563	*/
5564	u32Interval = u32SystemResolution = RTTimerGetSystemGranularity();
5565	while (u32Interval < 10000000 /* 10 ms */)
5566	u32Interval += u32SystemResolution;
5567
5568	supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), 1000000000 / u32Interval /=Hz/, cCpus);
5569
5570	/*
5571	* Create the timer.
5572	* If CPU_ALL isn't supported we'll have to fall back to synchronous mode.
5573	*/
5574	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
5575	{
5576	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, RTTIMER_FLAGS_CPU_ALL, supdrvGipAsyncTimer, pDevExt);
5577	if (rc == VERR_NOT_SUPPORTED)
5578	{
5579	OSDBGPRINT(("supdrvGipCreate: omni timer not supported, falling back to synchronous mode\n"));
5580	pGip->u32Mode = SUPGIPMODE_SYNC_TSC;
5581	}
5582	}
5583	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
5584	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0 /* fFlags */, supdrvGipSyncTimer, pDevExt);
5585	if (RT_SUCCESS(rc))
5586	{
5587	rc = RTMpNotificationRegister(supdrvGipMpEvent, pDevExt);
5588	if (RT_SUCCESS(rc))
5589	{
5590	rc = RTMpOnAll(supdrvGipInitOnCpu, pDevExt, pGip);
5591	if (RT_SUCCESS(rc))
5592	{
5593	/*
5594	* We're good.
5595	*/
5596	Log(("supdrvGipCreate: %u ns interval.\n", u32Interval));
5597	g_pSUPGlobalInfoPage = pGip;
5598	return VINF_SUCCESS;
5599	}
5600
5601	OSDBGPRINT(("supdrvGipCreate: RTMpOnAll failed with rc=%Rrc\n", rc));
5602	RTMpNotificationDeregister(supdrvGipMpEvent, pDevExt);
5603
5604	}
5605	else
5606	OSDBGPRINT(("supdrvGipCreate: failed to register MP event notfication. rc=%Rrc\n", rc));
5607	}
5608	else
5609	{
5610	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %u ns interval. rc=%Rrc\n", u32Interval, rc));
5611	Assert(!pDevExt->pGipTimer);
5612	}
5613	supdrvGipDestroy(pDevExt);
5614	return rc;
5615	}
5616
5617
5618	/**
5619	* Terminates the GIP.
5620	*
5621	* @param pDevExt Instance data. GIP stuff may be updated.
5622	*/
5623	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
5624	{
5625	int rc;
5626	#ifdef DEBUG_DARWIN_GIP
5627	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
5628	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
5629	pDevExt->pGipTimer, pDevExt->GipMemObj));
5630	#endif
5631
5632	/*
5633	* Invalid the GIP data.
5634	*/
5635	if (pDevExt->pGip)
5636	{
5637	supdrvGipTerm(pDevExt->pGip);
5638	pDevExt->pGip = NULL;
5639	}
5640	g_pSUPGlobalInfoPage = NULL;
5641
5642	/*
5643	* Destroy the timer and free the GIP memory object.
5644	*/
5645	if (pDevExt->pGipTimer)
5646	{
5647	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
5648	pDevExt->pGipTimer = NULL;
5649	}
5650
5651	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
5652	{
5653	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
5654	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
5655	}
5656
5657	/*
5658	* Finally, make sure we've release the system timer resolution request
5659	* if one actually succeeded and is still pending.
5660	*/
5661	if (pDevExt->u32SystemTimerGranularityGrant)
5662	{
5663	rc = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
5664	pDevExt->u32SystemTimerGranularityGrant = 0;
5665	}
5666	}
5667
5668
5669	/**
5670	* Timer callback function sync GIP mode.
5671	* @param pTimer The timer.
5672	* @param pvUser The device extension.
5673	*/
5674	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
5675	{
5676	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
5677	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5678	uint64_t u64TSC = ASMReadTSC();
5679	uint64_t NanoTS = RTTimeSystemNanoTS();
5680
5681	supdrvGipUpdate(pDevExt, NanoTS, u64TSC, NIL_RTCPUID, iTick);
5682
5683	ASMSetFlags(fOldFlags);
5684	}
5685
5686
5687	/**
5688	* Timer callback function for async GIP mode.
5689	* @param pTimer The timer.
5690	* @param pvUser The device extension.
5691	*/
5692	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
5693	{
5694	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
5695	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5696	RTCPUID idCpu = RTMpCpuId();
5697	uint64_t u64TSC = ASMReadTSC();
5698	uint64_t NanoTS = RTTimeSystemNanoTS();
5699
5700	/** @todo reset the transaction number and whatnot when iTick == 1. */
5701	if (pDevExt->idGipMaster == idCpu)
5702	supdrvGipUpdate(pDevExt, NanoTS, u64TSC, idCpu, iTick);
5703	else
5704	supdrvGipUpdatePerCpu(pDevExt, NanoTS, u64TSC, idCpu, ASMGetApicId(), iTick);
5705
5706	ASMSetFlags(fOldFlags);
5707	}
5708
5709
5710	/**
5711	* Finds our (@a idCpu) entry, or allocates a new one if not found.
5712	*
5713	* @returns Index of the CPU in the cache set.
5714	* @param pGip The GIP.
5715	* @param idCpu The CPU ID.
5716	*/
5717	static uint32_t supdrvGipCpuIndexFromCpuId(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
5718	{
5719	uint32_t i, cTries;
5720
5721	/*
5722	* ASSUMES that CPU IDs are constant.
5723	*/
5724	for (i = 0; i < pGip->cCpus; i++)
5725	if (pGip->aCPUs[i].idCpu == idCpu)
5726	return i;
5727
5728	cTries = 0;
5729	do
5730	{
5731	for (i = 0; i < pGip->cCpus; i++)
5732	{
5733	bool fRc;
5734	ASMAtomicCmpXchgSize(&pGip->aCPUs[i].idCpu, idCpu, NIL_RTCPUID, fRc);
5735	if (fRc)
5736	return i;
5737	}
5738	} while (cTries++ < 32);
5739	AssertReleaseFailed();
5740	return i - 1;
5741	}
5742
5743
5744	/**
5745	* The calling CPU should be accounted as online, update GIP accordingly.
5746	*
5747	* This is used by supdrvGipMpEvent as well as the supdrvGipCreate.
5748	*
5749	* @param pDevExt The device extension.
5750	* @param idCpu The CPU ID.
5751	*/
5752	static void supdrvGipMpEventOnline(PSUPDRVDEVEXT pDevExt, RTCPUID idCpu)
5753	{
5754	int iCpuSet = 0;
5755	uint16_t idApic = UINT16_MAX;
5756	uint32_t i = 0;
5757	uint64_t u64NanoTS = 0;
5758	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5759
5760	AssertPtrReturnVoid(pGip);
5761	AssertRelease(idCpu == RTMpCpuId());
5762	Assert(pGip->cPossibleCpus == RTMpGetCount());
5763
5764	/*
5765	* Do this behind a spinlock with interrupts disabled as this can fire
5766	* on all CPUs simultaneously, see @bugref{6110}.
5767	*/
5768	RTSpinlockAcquire(pDevExt->hGipSpinlock);
5769
5770	/*
5771	* Update the globals.
5772	*/
5773	ASMAtomicWriteU16(&pGip->cPresentCpus, RTMpGetPresentCount());
5774	ASMAtomicWriteU16(&pGip->cOnlineCpus, RTMpGetOnlineCount());
5775	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5776	if (iCpuSet >= 0)
5777	{
5778	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5779	RTCpuSetAddByIndex(&pGip->OnlineCpuSet, iCpuSet);
5780	RTCpuSetAddByIndex(&pGip->PresentCpuSet, iCpuSet);
5781	}
5782
5783	/*
5784	* Update the entry.
5785	*/
5786	u64NanoTS = RTTimeSystemNanoTS() - pGip->u32UpdateIntervalNS;
5787	i = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
5788	supdrvGipInitCpu(pGip, &pGip->aCPUs[i], u64NanoTS);
5789	idApic = ASMGetApicId();
5790	ASMAtomicWriteU16(&pGip->aCPUs[i].idApic, idApic);
5791	ASMAtomicWriteS16(&pGip->aCPUs[i].iCpuSet, (int16_t)iCpuSet);
5792	ASMAtomicWriteSize(&pGip->aCPUs[i].idCpu, idCpu);
5793
5794	/*
5795	* Update the APIC ID and CPU set index mappings.
5796	*/
5797	ASMAtomicWriteU16(&pGip->aiCpuFromApicId[idApic], i);
5798	ASMAtomicWriteU16(&pGip->aiCpuFromCpuSetIdx[iCpuSet], i);
5799
5800	/* commit it */
5801	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_ONLINE);
5802
5803	RTSpinlockReleaseNoInts(pDevExt->hGipSpinlock);
5804	}
5805
5806
5807	/**
5808	* The CPU should be accounted as offline, update the GIP accordingly.
5809	*
5810	* This is used by supdrvGipMpEvent.
5811	*
5812	* @param pDevExt The device extension.
5813	* @param idCpu The CPU ID.
5814	*/
5815	static void supdrvGipMpEventOffline(PSUPDRVDEVEXT pDevExt, RTCPUID idCpu)
5816	{
5817	int iCpuSet;
5818	unsigned i;
5819
5820	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5821
5822	AssertPtrReturnVoid(pGip);
5823	RTSpinlockAcquire(pDevExt->hGipSpinlock);
5824
5825	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5826	AssertReturnVoid(iCpuSet >= 0);
5827
5828	i = pGip->aiCpuFromCpuSetIdx[iCpuSet];
5829	AssertReturnVoid(i < pGip->cCpus);
5830	AssertReturnVoid(pGip->aCPUs[i].idCpu == idCpu);
5831
5832	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5833	RTCpuSetDelByIndex(&pGip->OnlineCpuSet, iCpuSet);
5834
5835	/* commit it */
5836	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_OFFLINE);
5837
5838	RTSpinlockReleaseNoInts(pDevExt->hGipSpinlock);
5839	}
5840
5841
5842	/**
5843	* Multiprocessor event notification callback.
5844	*
5845	* This is used to make sure that the GIP master gets passed on to
5846	* another CPU. It also updates the associated CPU data.
5847	*
5848	* @param enmEvent The event.
5849	* @param idCpu The cpu it applies to.
5850	* @param pvUser Pointer to the device extension.
5851	*
5852	* @remarks This function -must- fire on the newly online'd CPU for the
5853	* RTMPEVENT_ONLINE case and can fire on any CPU for the
5854	* RTMPEVENT_OFFLINE case.
5855	*/
5856	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser)
5857	{
5858	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5859	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5860
5861	AssertRelease(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
5862
5863	/*
5864	* Update the GIP CPU data.
5865	*/
5866	if (pGip)
5867	{
5868	switch (enmEvent)
5869	{
5870	case RTMPEVENT_ONLINE:
5871	AssertRelease(idCpu == RTMpCpuId());
5872	supdrvGipMpEventOnline(pDevExt, idCpu);
5873	break;
5874	case RTMPEVENT_OFFLINE:
5875	supdrvGipMpEventOffline(pDevExt, idCpu);
5876	break;
5877
5878	}
5879	}
5880
5881	/*
5882	* Make sure there is a master GIP.
5883	*/
5884	if (enmEvent == RTMPEVENT_OFFLINE)
5885	{
5886	RTCPUID idGipMaster = ASMAtomicReadU32(&pDevExt->idGipMaster);
5887	if (idGipMaster == idCpu)
5888	{
5889	/*
5890	* Find a new GIP master.
5891	*/
5892	bool fIgnored;
5893	unsigned i;
5894	RTCPUID idNewGipMaster = NIL_RTCPUID;
5895	RTCPUSET OnlineCpus;
5896	RTMpGetOnlineSet(&OnlineCpus);
5897
5898	for (i = 0; i < RTCPUSET_MAX_CPUS; i++)
5899	{
5900	RTCPUID idCurCpu = RTMpCpuIdFromSetIndex(i);
5901	if ( RTCpuSetIsMember(&OnlineCpus, idCurCpu)
5902	&& idCurCpu != idGipMaster)
5903	{
5904	idNewGipMaster = idCurCpu;
5905	break;
5906	}
5907	}
5908
5909	Log(("supdrvGipMpEvent: Gip master %#lx -> %#lx\n", (long)idGipMaster, (long)idNewGipMaster));
5910	ASMAtomicCmpXchgSize(&pDevExt->idGipMaster, idNewGipMaster, idGipMaster, fIgnored);
5911	NOREF(fIgnored);
5912	}
5913	}
5914	}
5915
5916
5917	/**
5918	* Callback used by supdrvDetermineAsyncTSC to read the TSC on a CPU.
5919	*
5920	* @param idCpu Ignored.
5921	* @param pvUser1 Where to put the TSC.
5922	* @param pvUser2 Ignored.
5923	*/
5924	static DECLCALLBACK(void) supdrvDetermineAsyncTscWorker(RTCPUID idCpu, void pvUser1, void pvUser2)
5925	{
5926	#if 1
5927	ASMAtomicWriteU64((uint64_t volatile *)pvUser1, ASMReadTSC());
5928	#else
5929	(uint64_t )pvUser1 = ASMReadTSC();
5930	#endif
5931	}
5932
5933
5934	/**
5935	* Determine if Async GIP mode is required because of TSC drift.
5936	*
5937	* When using the default/normal timer code it is essential that the time stamp counter
5938	* (TSC) runs never backwards, that is, a read operation to the counter should return
5939	* a bigger value than any previous read operation. This is guaranteed by the latest
5940	* AMD CPUs and by newer Intel CPUs which never enter the C2 state (P4). In any other
5941	* case we have to choose the asynchronous timer mode.
5942	*
5943	* @param poffMin Pointer to the determined difference between different cores.
5944	* @return false if the time stamp counters appear to be synchronized, true otherwise.
5945	*/
5946	static bool supdrvDetermineAsyncTsc(uint64_t *poffMin)
5947	{
5948	/*
5949	* Just iterate all the cpus 8 times and make sure that the TSC is
5950	* ever increasing. We don't bother taking TSC rollover into account.
5951	*/
5952	int iEndCpu = RTMpGetArraySize();
5953	int iCpu;
5954	int cLoops = 8;
5955	bool fAsync = false;
5956	int rc = VINF_SUCCESS;
5957	uint64_t offMax = 0;
5958	uint64_t offMin = ~(uint64_t)0;
5959	uint64_t PrevTsc = ASMReadTSC();
5960
5961	while (cLoops-- > 0)
5962	{
5963	for (iCpu = 0; iCpu < iEndCpu; iCpu++)
5964	{
5965	uint64_t CurTsc;
5966	rc = RTMpOnSpecific(RTMpCpuIdFromSetIndex(iCpu), supdrvDetermineAsyncTscWorker, &CurTsc, NULL);
5967	if (RT_SUCCESS(rc))
5968	{
5969	if (CurTsc <= PrevTsc)
5970	{
5971	fAsync = true;
5972	offMin = offMax = PrevTsc - CurTsc;
5973	Log(("supdrvDetermineAsyncTsc: iCpu=%d cLoops=%d CurTsc=%llx PrevTsc=%llx\n",
5974	iCpu, cLoops, CurTsc, PrevTsc));
5975	break;
5976	}
5977
5978	/* Gather statistics (except the first time). */
5979	if (iCpu != 0 \|\| cLoops != 7)
5980	{
5981	uint64_t off = CurTsc - PrevTsc;
5982	if (off < offMin)
5983	offMin = off;
5984	if (off > offMax)
5985	offMax = off;
5986	Log2(("%d/%d: off=%llx\n", cLoops, iCpu, off));
5987	}
5988
5989	/* Next */
5990	PrevTsc = CurTsc;
5991	}
5992	else if (rc == VERR_NOT_SUPPORTED)
5993	break;
5994	else
5995	AssertMsg(rc == VERR_CPU_NOT_FOUND \|\| rc == VERR_CPU_OFFLINE, ("%d\n", rc));
5996	}
5997
5998	/* broke out of the loop. */
5999	if (iCpu < iEndCpu)
6000	break;
6001	}
6002
6003	poffMin = offMin; / Almost RTMpOnSpecific profiling. */
6004	Log(("supdrvDetermineAsyncTsc: returns %d; iEndCpu=%d rc=%d offMin=%llx offMax=%llx\n",
6005	fAsync, iEndCpu, rc, offMin, offMax));
6006	#if !defined(RT_OS_SOLARIS) && !defined(RT_OS_OS2) && !defined(RT_OS_WINDOWS)
6007	OSDBGPRINT(("vboxdrv: fAsync=%d offMin=%#lx offMax=%#lx\n", fAsync, (long)offMin, (long)offMax));
6008	#endif
6009	return fAsync;
6010	}
6011
6012
6013	/**
6014	* Determine the GIP TSC mode.
6015	*
6016	* @returns The most suitable TSC mode.
6017	* @param pDevExt Pointer to the device instance data.
6018	*/
6019	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt)
6020	{
6021	/*
6022	* On SMP we're faced with two problems:
6023	* (1) There might be a skew between the CPU, so that cpu0
6024	* returns a TSC that is slightly different from cpu1.
6025	* (2) Power management (and other things) may cause the TSC
6026	* to run at a non-constant speed, and cause the speed
6027	* to be different on the cpus. This will result in (1).
6028	*
6029	* So, on SMP systems we'll have to select the ASYNC update method
6030	* if there are symptoms of these problems.
6031	*/
6032	if (RTMpGetCount() > 1)
6033	{
6034	uint32_t uEAX, uEBX, uECX, uEDX;
6035	uint64_t u64DiffCoresIgnored;
6036
6037	/* Permit the user and/or the OS specific bits to force async mode. */
6038	if (supdrvOSGetForcedAsyncTscMode(pDevExt))
6039	return SUPGIPMODE_ASYNC_TSC;
6040
6041	/* Try check for current differences between the cpus. */
6042	if (supdrvDetermineAsyncTsc(&u64DiffCoresIgnored))
6043	return SUPGIPMODE_ASYNC_TSC;
6044
6045	/*
6046	* If the CPU supports power management and is an AMD one we
6047	* won't trust it unless it has the TscInvariant bit is set.
6048	*/
6049	/* Check for "AuthenticAMD" */
6050	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
6051	if ( uEAX >= 1
6052	&& ASMIsAmdCpuEx(uEBX, uECX, uEDX))
6053	{
6054	/* Check for APM support and that TscInvariant is cleared. */
6055	ASMCpuId(0x80000000, &uEAX, &uEBX, &uECX, &uEDX);
6056	if (uEAX >= 0x80000007)
6057	{
6058	ASMCpuId(0x80000007, &uEAX, &uEBX, &uECX, &uEDX);
6059	if ( !(uEDX & RT_BIT(8))/* TscInvariant */
6060	&& (uEDX & 0x3e)) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
6061	return SUPGIPMODE_ASYNC_TSC;
6062	}
6063	}
6064	}
6065	return SUPGIPMODE_SYNC_TSC;
6066	}
6067
6068
6069	/**
6070	* Initializes per-CPU GIP information.
6071	*
6072	* @param pGip Pointer to the read-write kernel mapping of the GIP.
6073	* @param pCpu Pointer to which GIP CPU to initalize.
6074	* @param u64NanoTS The current nanosecond timestamp.
6075	*/
6076	static void supdrvGipInitCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pCpu, uint64_t u64NanoTS)
6077	{
6078	pCpu->u32TransactionId = 2;
6079	pCpu->u64NanoTS = u64NanoTS;
6080	pCpu->u64TSC = ASMReadTSC();
6081
6082	ASMAtomicWriteSize(&pCpu->enmState, SUPGIPCPUSTATE_INVALID);
6083	ASMAtomicWriteSize(&pCpu->idCpu, NIL_RTCPUID);
6084	ASMAtomicWriteS16(&pCpu->iCpuSet, -1);
6085	ASMAtomicWriteU16(&pCpu->idApic, UINT16_MAX);
6086
6087	/*
6088	* We don't know the following values until we've executed updates.
6089	* So, we'll just pretend it's a 4 GHz CPU and adjust the history it on
6090	* the 2nd timer callout.
6091	*/
6092	pCpu->u64CpuHz = _4G + 1; /* tstGIP-2 depends on this. */
6093	pCpu->u32UpdateIntervalTSC
6094	= pCpu->au32TSCHistory[0]
6095	= pCpu->au32TSCHistory[1]
6096	= pCpu->au32TSCHistory[2]
6097	= pCpu->au32TSCHistory[3]
6098	= pCpu->au32TSCHistory[4]
6099	= pCpu->au32TSCHistory[5]
6100	= pCpu->au32TSCHistory[6]
6101	= pCpu->au32TSCHistory[7]
6102	= (uint32_t)(_4G / pGip->u32UpdateHz);
6103	}
6104
6105
6106	/**
6107	* Initializes the GIP data.
6108	*
6109	* @param pDevExt Pointer to the device instance data.
6110	* @param pGip Pointer to the read-write kernel mapping of the GIP.
6111	* @param HCPhys The physical address of the GIP.
6112	* @param u64NanoTS The current nanosecond timestamp.
6113	* @param uUpdateHz The update frequency.
6114	* @param cCpus The CPU count.
6115	*/
6116	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
6117	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned cCpus)
6118	{
6119	size_t const cbGip = RT_ALIGN_Z(RT_OFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), PAGE_SIZE);
6120	unsigned i;
6121	#ifdef DEBUG_DARWIN_GIP
6122	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
6123	#else
6124	LogFlow(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
6125	#endif
6126
6127	/*
6128	* Initialize the structure.
6129	*/
6130	memset(pGip, 0, cbGip);
6131	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
6132	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
6133	pGip->u32Mode = supdrvGipDeterminTscMode(pDevExt);
6134	pGip->cCpus = (uint16_t)cCpus;
6135	pGip->cPages = (uint16_t)(cbGip / PAGE_SIZE);
6136	pGip->u32UpdateHz = uUpdateHz;
6137	pGip->u32UpdateIntervalNS = 1000000000 / uUpdateHz;
6138	pGip->u64NanoTSLastUpdateHz = u64NanoTS;
6139	RTCpuSetEmpty(&pGip->OnlineCpuSet);
6140	RTCpuSetEmpty(&pGip->PresentCpuSet);
6141	RTMpGetSet(&pGip->PossibleCpuSet);
6142	pGip->cOnlineCpus = RTMpGetOnlineCount();
6143	pGip->cPresentCpus = RTMpGetPresentCount();
6144	pGip->cPossibleCpus = RTMpGetCount();
6145	pGip->idCpuMax = RTMpGetMaxCpuId();
6146	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromApicId); i++)
6147	pGip->aiCpuFromApicId[i] = 0;
6148	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx); i++)
6149	pGip->aiCpuFromCpuSetIdx[i] = UINT16_MAX;
6150
6151	for (i = 0; i < cCpus; i++)
6152	supdrvGipInitCpu(pGip, &pGip->aCPUs[i], u64NanoTS);
6153
6154	/*
6155	* Link it to the device extension.
6156	*/
6157	pDevExt->pGip = pGip;
6158	pDevExt->HCPhysGip = HCPhys;
6159	pDevExt->cGipUsers = 0;
6160	}
6161
6162
6163	/**
6164	* On CPU initialization callback for RTMpOnAll.
6165	*
6166	* @param idCpu The CPU ID.
6167	* @param pvUser1 The device extension.
6168	* @param pvUser2 The GIP.
6169	*/
6170	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2)
6171	{
6172	/* This is good enough, even though it will update some of the globals a
6173	bit to much. */
6174	supdrvGipMpEventOnline((PSUPDRVDEVEXT)pvUser1, idCpu);
6175	}
6176
6177
6178	/**
6179	* Invalidates the GIP data upon termination.
6180	*
6181	* @param pGip Pointer to the read-write kernel mapping of the GIP.
6182	*/
6183	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
6184	{
6185	unsigned i;
6186	pGip->u32Magic = 0;
6187	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
6188	{
6189	pGip->aCPUs[i].u64NanoTS = 0;
6190	pGip->aCPUs[i].u64TSC = 0;
6191	pGip->aCPUs[i].iTSCHistoryHead = 0;
6192	}
6193	}
6194
6195
6196	/**
6197	* Worker routine for supdrvGipUpdate and supdrvGipUpdatePerCpu that
6198	* updates all the per cpu data except the transaction id.
6199	*
6200	* @param pDevExt The device extension.
6201	* @param pGipCpu Pointer to the per cpu data.
6202	* @param u64NanoTS The current time stamp.
6203	* @param u64TSC The current TSC.
6204	* @param iTick The current timer tick.
6205	*/
6206	static void supdrvGipDoUpdateCpu(PSUPDRVDEVEXT pDevExt, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick)
6207	{
6208	uint64_t u64TSCDelta;
6209	uint32_t u32UpdateIntervalTSC;
6210	uint32_t u32UpdateIntervalTSCSlack;
6211	unsigned iTSCHistoryHead;
6212	uint64_t u64CpuHz;
6213	uint32_t u32TransactionId;
6214
6215	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
6216	AssertPtrReturnVoid(pGip);
6217
6218	/* Delta between this and the previous update. */
6219	ASMAtomicUoWriteU32(&pGipCpu->u32PrevUpdateIntervalNS, (uint32_t)(u64NanoTS - pGipCpu->u64NanoTS));
6220
6221	/*
6222	* Update the NanoTS.
6223	*/
6224	ASMAtomicWriteU64(&pGipCpu->u64NanoTS, u64NanoTS);
6225
6226	/*
6227	* Calc TSC delta.
6228	*/
6229	/** @todo validate the NanoTS delta, don't trust the OS to call us when it should... */
6230	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
6231	ASMAtomicWriteU64(&pGipCpu->u64TSC, u64TSC);
6232
6233	if (u64TSCDelta >> 32)
6234	{
6235	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
6236	pGipCpu->cErrors++;
6237	}
6238
6239	/*
6240	* On the 2nd and 3rd callout, reset the history with the current TSC
6241	* interval since the values entered by supdrvGipInit are totally off.
6242	* The interval on the 1st callout completely unreliable, the 2nd is a bit
6243	* better, while the 3rd should be most reliable.
6244	*/
6245	u32TransactionId = pGipCpu->u32TransactionId;
6246	if (RT_UNLIKELY( ( u32TransactionId == 5
6247	\|\| u32TransactionId == 7)
6248	&& ( iTick == 2
6249	\|\| iTick == 3) ))
6250	{
6251	unsigned i;
6252	for (i = 0; i < RT_ELEMENTS(pGipCpu->au32TSCHistory); i++)
6253	ASMAtomicUoWriteU32(&pGipCpu->au32TSCHistory[i], (uint32_t)u64TSCDelta);
6254	}
6255
6256	/*
6257	* TSC History.
6258	*/
6259	Assert(RT_ELEMENTS(pGipCpu->au32TSCHistory) == 8);
6260	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
6261	ASMAtomicWriteU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
6262	ASMAtomicWriteU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
6263
6264	/*
6265	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
6266	*/
6267	if (pGip->u32UpdateHz >= 1000)
6268	{
6269	uint32_t u32;
6270	u32 = pGipCpu->au32TSCHistory[0];
6271	u32 += pGipCpu->au32TSCHistory[1];
6272	u32 += pGipCpu->au32TSCHistory[2];
6273	u32 += pGipCpu->au32TSCHistory[3];
6274	u32 >>= 2;
6275	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
6276	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
6277	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
6278	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
6279	u32UpdateIntervalTSC >>= 2;
6280	u32UpdateIntervalTSC += u32;
6281	u32UpdateIntervalTSC >>= 1;
6282
6283	/* Value chosen for a 2GHz Athlon64 running linux 2.6.10/11, . */
6284	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
6285	}
6286	else if (pGip->u32UpdateHz >= 90)
6287	{
6288	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
6289	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
6290	u32UpdateIntervalTSC >>= 1;
6291
6292	/* value chosen on a 2GHz thinkpad running windows */
6293	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
6294	}
6295	else
6296	{
6297	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
6298
6299	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
6300	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
6301	}
6302	ASMAtomicWriteU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
6303
6304	/*
6305	* CpuHz.
6306	*/
6307	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, pGip->u32UpdateHz);
6308	ASMAtomicWriteU64(&pGipCpu->u64CpuHz, u64CpuHz);
6309	}
6310
6311
6312	/**
6313	* Updates the GIP.
6314	*
6315	* @param pDevExt The device extension.
6316	* @param u64NanoTS The current nanosecond timesamp.
6317	* @param u64TSC The current TSC timesamp.
6318	* @param idCpu The CPU ID.
6319	* @param iTick The current timer tick.
6320	*/
6321	static void supdrvGipUpdate(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick)
6322	{
6323	/*
6324	* Determine the relevant CPU data.
6325	*/
6326	PSUPGIPCPU pGipCpu;
6327	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
6328	AssertPtrReturnVoid(pGip);
6329
6330	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
6331	pGipCpu = &pGip->aCPUs[0];
6332	else
6333	{
6334	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
6335	if (RT_UNLIKELY(iCpu >= pGip->cCpus))
6336	return;
6337	pGipCpu = &pGip->aCPUs[iCpu];
6338	if (RT_UNLIKELY(pGipCpu->idCpu != idCpu))
6339	return;
6340	}
6341
6342	/*
6343	* Start update transaction.
6344	*/
6345	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
6346	{
6347	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
6348	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
6349	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
6350	pGipCpu->cErrors++;
6351	return;
6352	}
6353
6354	/*
6355	* Recalc the update frequency every 0x800th time.
6356	*/
6357	if (!(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
6358	{
6359	if (pGip->u64NanoTSLastUpdateHz)
6360	{
6361	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
6362	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
6363	uint32_t u32UpdateHz = (uint32_t)((UINT64_C(1000000000) * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
6364	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
6365	{
6366	ASMAtomicWriteU32(&pGip->u32UpdateHz, u32UpdateHz);
6367	ASMAtomicWriteU32(&pGip->u32UpdateIntervalNS, 1000000000 / u32UpdateHz);
6368	}
6369	#endif
6370	}
6371	ASMAtomicWriteU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS);
6372	}
6373
6374	/*
6375	* Update the data.
6376	*/
6377	supdrvGipDoUpdateCpu(pDevExt, pGipCpu, u64NanoTS, u64TSC, iTick);
6378
6379	/*
6380	* Complete transaction.
6381	*/
6382	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
6383	}
6384
6385
6386	/**
6387	* Updates the per cpu GIP data for the calling cpu.
6388	*
6389	* @param pDevExt The device extension.
6390	* @param u64NanoTS The current nanosecond timesamp.
6391	* @param u64TSC The current TSC timesamp.
6392	* @param idCpu The CPU ID.
6393	* @param idApic The APIC id for the CPU index.
6394	* @param iTick The current timer tick.
6395	*/
6396	static void supdrvGipUpdatePerCpu(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC,
6397	RTCPUID idCpu, uint8_t idApic, uint64_t iTick)
6398	{
6399	uint32_t iCpu;
6400	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
6401
6402	/*
6403	* Avoid a potential race when a CPU online notification doesn't fire on
6404	* the onlined CPU but the tick creeps in before the event notification is
6405	* run.
6406	*/
6407	if (RT_UNLIKELY(iTick == 1))
6408	{
6409	iCpu = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
6410	if (pGip->aCPUs[iCpu].enmState == SUPGIPCPUSTATE_OFFLINE)
6411	supdrvGipMpEventOnline(pDevExt, idCpu);
6412	}
6413
6414	iCpu = pGip->aiCpuFromApicId[idApic];
6415	if (RT_LIKELY(iCpu < pGip->cCpus))
6416	{
6417	PSUPGIPCPU pGipCpu = &pGip->aCPUs[iCpu];
6418	if (pGipCpu->idCpu == idCpu)
6419	{
6420	/*
6421	* Start update transaction.
6422	*/
6423	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
6424	{
6425	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
6426	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
6427	pGipCpu->cErrors++;
6428	return;
6429	}
6430
6431	/*
6432	* Update the data.
6433	*/
6434	supdrvGipDoUpdateCpu(pDevExt, pGipCpu, u64NanoTS, u64TSC, iTick);
6435
6436	/*
6437	* Complete transaction.
6438	*/
6439	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
6440	}
6441	}
6442	}
6443
6444	/**
6445	* Resume built-in keyboard on MacBook Air and Pro hosts.
6446	* If there is no built-in keyboard device, return success anyway.
6447	*
6448	* @returns 0 on Mac OS X platform, VERR_NOT_IMPLEMENTED on the other ones.
6449	*/
6450	static int supdrvIOCtl_ResumeSuspendedKbds(void)
6451	{
6452	#if defined(RT_OS_DARWIN)
6453	return supdrvDarwinResumeSuspendedKbds();
6454	#else
6455	return VERR_NOT_IMPLEMENTED;
6456	#endif
6457	}
6458

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

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDrv.c@ 52478

Download in other formats: