SUPDrv.c@ 51850

Last change on this file since 51850 was 51850, checked in by vboxsync, 11 years ago
hrmpf
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 238.6 KB

Line
1	/* $Id: SUPDrv.c 51850 2014-07-03 13:39:52Z 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	{ "SUPR0LockMem", (void *)SUPR0LockMem },
181	{ "SUPR0LowAlloc", (void *)SUPR0LowAlloc },
182	{ "SUPR0LowFree", (void *)SUPR0LowFree },
183	{ "SUPR0MemAlloc", (void *)SUPR0MemAlloc },
184	{ "SUPR0MemFree", (void *)SUPR0MemFree },
185	{ "SUPR0MemGetPhys", (void *)SUPR0MemGetPhys },
186	{ "SUPR0ObjAddRef", (void *)SUPR0ObjAddRef },
187	{ "SUPR0ObjAddRefEx", (void *)SUPR0ObjAddRefEx },
188	{ "SUPR0ObjRegister", (void *)SUPR0ObjRegister },
189	{ "SUPR0ObjRelease", (void *)SUPR0ObjRelease },
190	{ "SUPR0ObjVerifyAccess", (void *)SUPR0ObjVerifyAccess },
191	{ "SUPR0PageAllocEx", (void *)SUPR0PageAllocEx },
192	{ "SUPR0PageFree", (void *)SUPR0PageFree },
193	{ "SUPR0Printf", (void *)SUPR0Printf },
194	{ "SUPR0TracerDeregisterDrv", (void *)SUPR0TracerDeregisterDrv },
195	{ "SUPR0TracerDeregisterImpl", (void *)SUPR0TracerDeregisterImpl },
196	{ "SUPR0TracerFireProbe", (void *)SUPR0TracerFireProbe },
197	{ "SUPR0TracerRegisterDrv", (void *)SUPR0TracerRegisterDrv },
198	{ "SUPR0TracerRegisterImpl", (void *)SUPR0TracerRegisterImpl },
199	{ "SUPR0TracerRegisterModule", (void *)SUPR0TracerRegisterModule },
200	{ "SUPR0TracerUmodProbeFire", (void *)SUPR0TracerUmodProbeFire },
201	{ "SUPR0UnlockMem", (void *)SUPR0UnlockMem },
202	{ "SUPSemEventClose", (void *)SUPSemEventClose },
203	{ "SUPSemEventCreate", (void *)SUPSemEventCreate },
204	{ "SUPSemEventGetResolution", (void *)SUPSemEventGetResolution },
205	{ "SUPSemEventMultiClose", (void *)SUPSemEventMultiClose },
206	{ "SUPSemEventMultiCreate", (void *)SUPSemEventMultiCreate },
207	{ "SUPSemEventMultiGetResolution", (void *)SUPSemEventMultiGetResolution },
208	{ "SUPSemEventMultiReset", (void *)SUPSemEventMultiReset },
209	{ "SUPSemEventMultiSignal", (void *)SUPSemEventMultiSignal },
210	{ "SUPSemEventMultiWait", (void *)SUPSemEventMultiWait },
211	{ "SUPSemEventMultiWaitNoResume", (void *)SUPSemEventMultiWaitNoResume },
212	{ "SUPSemEventMultiWaitNsAbsIntr", (void *)SUPSemEventMultiWaitNsAbsIntr },
213	{ "SUPSemEventMultiWaitNsRelIntr", (void *)SUPSemEventMultiWaitNsRelIntr },
214	{ "SUPSemEventSignal", (void *)SUPSemEventSignal },
215	{ "SUPSemEventWait", (void *)SUPSemEventWait },
216	{ "SUPSemEventWaitNoResume", (void *)SUPSemEventWaitNoResume },
217	{ "SUPSemEventWaitNsAbsIntr", (void *)SUPSemEventWaitNsAbsIntr },
218	{ "SUPSemEventWaitNsRelIntr", (void *)SUPSemEventWaitNsRelIntr },
219
220	{ "RTAssertAreQuiet", (void *)RTAssertAreQuiet },
221	{ "RTAssertMayPanic", (void *)RTAssertMayPanic },
222	{ "RTAssertMsg1", (void *)RTAssertMsg1 },
223	{ "RTAssertMsg2AddV", (void *)RTAssertMsg2AddV },
224	{ "RTAssertMsg2V", (void *)RTAssertMsg2V },
225	{ "RTAssertSetMayPanic", (void *)RTAssertSetMayPanic },
226	{ "RTAssertSetQuiet", (void *)RTAssertSetQuiet },
227	{ "RTCrc32", (void *)RTCrc32 },
228	{ "RTCrc32Finish", (void *)RTCrc32Finish },
229	{ "RTCrc32Process", (void *)RTCrc32Process },
230	{ "RTCrc32Start", (void *)RTCrc32Start },
231	{ "RTErrConvertFromErrno", (void *)RTErrConvertFromErrno },
232	{ "RTErrConvertToErrno", (void *)RTErrConvertToErrno },
233	{ "RTHandleTableAllocWithCtx", (void *)RTHandleTableAllocWithCtx },
234	{ "RTHandleTableCreate", (void *)RTHandleTableCreate },
235	{ "RTHandleTableCreateEx", (void *)RTHandleTableCreateEx },
236	{ "RTHandleTableDestroy", (void *)RTHandleTableDestroy },
237	{ "RTHandleTableFreeWithCtx", (void *)RTHandleTableFreeWithCtx },
238	{ "RTHandleTableLookupWithCtx", (void *)RTHandleTableLookupWithCtx },
239	{ "RTLogDefaultInstance", (void *)RTLogDefaultInstance },
240	{ "RTLogGetDefaultInstance", (void *)RTLogGetDefaultInstance },
241	{ "RTLogLoggerExV", (void *)RTLogLoggerExV },
242	{ "RTLogPrintfV", (void *)RTLogPrintfV },
243	{ "RTLogRelDefaultInstance", (void *)RTLogRelDefaultInstance },
244	{ "RTLogSetDefaultInstanceThread", (void *)RTLogSetDefaultInstanceThread },
245	{ "RTMemAllocExTag", (void *)RTMemAllocExTag },
246	{ "RTMemAllocTag", (void *)RTMemAllocTag },
247	{ "RTMemAllocVarTag", (void *)RTMemAllocVarTag },
248	{ "RTMemAllocZTag", (void *)RTMemAllocZTag },
249	{ "RTMemAllocZVarTag", (void *)RTMemAllocZVarTag },
250	{ "RTMemDupExTag", (void *)RTMemDupExTag },
251	{ "RTMemDupTag", (void *)RTMemDupTag },
252	{ "RTMemFree", (void *)RTMemFree },
253	{ "RTMemFreeEx", (void *)RTMemFreeEx },
254	{ "RTMemReallocTag", (void *)RTMemReallocTag },
255	{ "RTMpCpuId", (void *)RTMpCpuId },
256	{ "RTMpCpuIdFromSetIndex", (void *)RTMpCpuIdFromSetIndex },
257	{ "RTMpCpuIdToSetIndex", (void *)RTMpCpuIdToSetIndex },
258	{ "RTMpGetArraySize", (void *)RTMpGetArraySize },
259	{ "RTMpGetCount", (void *)RTMpGetCount },
260	{ "RTMpGetMaxCpuId", (void *)RTMpGetMaxCpuId },
261	{ "RTMpGetOnlineCount", (void *)RTMpGetOnlineCount },
262	{ "RTMpGetOnlineSet", (void *)RTMpGetOnlineSet },
263	{ "RTMpGetSet", (void *)RTMpGetSet },
264	{ "RTMpIsCpuOnline", (void *)RTMpIsCpuOnline },
265	{ "RTMpIsCpuPossible", (void *)RTMpIsCpuPossible },
266	{ "RTMpIsCpuWorkPending", (void *)RTMpIsCpuWorkPending },
267	{ "RTMpNotificationDeregister", (void *)RTMpNotificationDeregister },
268	{ "RTMpNotificationRegister", (void *)RTMpNotificationRegister },
269	{ "RTMpOnAll", (void *)RTMpOnAll },
270	{ "RTMpOnOthers", (void *)RTMpOnOthers },
271	{ "RTMpOnSpecific", (void *)RTMpOnSpecific },
272	{ "RTMpPokeCpu", (void *)RTMpPokeCpu },
273	{ "RTNetIPv4AddDataChecksum", (void *)RTNetIPv4AddDataChecksum },
274	{ "RTNetIPv4AddTCPChecksum", (void *)RTNetIPv4AddTCPChecksum },
275	{ "RTNetIPv4AddUDPChecksum", (void *)RTNetIPv4AddUDPChecksum },
276	{ "RTNetIPv4FinalizeChecksum", (void *)RTNetIPv4FinalizeChecksum },
277	{ "RTNetIPv4HdrChecksum", (void *)RTNetIPv4HdrChecksum },
278	{ "RTNetIPv4IsDHCPValid", (void *)RTNetIPv4IsDHCPValid },
279	{ "RTNetIPv4IsHdrValid", (void *)RTNetIPv4IsHdrValid },
280	{ "RTNetIPv4IsTCPSizeValid", (void *)RTNetIPv4IsTCPSizeValid },
281	{ "RTNetIPv4IsTCPValid", (void *)RTNetIPv4IsTCPValid },
282	{ "RTNetIPv4IsUDPSizeValid", (void *)RTNetIPv4IsUDPSizeValid },
283	{ "RTNetIPv4IsUDPValid", (void *)RTNetIPv4IsUDPValid },
284	{ "RTNetIPv4PseudoChecksum", (void *)RTNetIPv4PseudoChecksum },
285	{ "RTNetIPv4PseudoChecksumBits", (void *)RTNetIPv4PseudoChecksumBits },
286	{ "RTNetIPv4TCPChecksum", (void *)RTNetIPv4TCPChecksum },
287	{ "RTNetIPv4UDPChecksum", (void *)RTNetIPv4UDPChecksum },
288	{ "RTNetIPv6PseudoChecksum", (void *)RTNetIPv6PseudoChecksum },
289	{ "RTNetIPv6PseudoChecksumBits", (void *)RTNetIPv6PseudoChecksumBits },
290	{ "RTNetIPv6PseudoChecksumEx", (void *)RTNetIPv6PseudoChecksumEx },
291	{ "RTNetTCPChecksum", (void *)RTNetTCPChecksum },
292	{ "RTNetUDPChecksum", (void *)RTNetUDPChecksum },
293	{ "RTPowerNotificationDeregister", (void *)RTPowerNotificationDeregister },
294	{ "RTPowerNotificationRegister", (void *)RTPowerNotificationRegister },
295	{ "RTProcSelf", (void *)RTProcSelf },
296	{ "RTR0AssertPanicSystem", (void *)RTR0AssertPanicSystem },
297	{ "RTR0MemAreKrnlAndUsrDifferent", (void *)RTR0MemAreKrnlAndUsrDifferent },
298	{ "RTR0MemKernelIsValidAddr", (void *)RTR0MemKernelIsValidAddr },
299	{ "RTR0MemKernelCopyFrom", (void *)RTR0MemKernelCopyFrom },
300	{ "RTR0MemKernelCopyTo", (void *)RTR0MemKernelCopyTo },
301	{ "RTR0MemObjAddress", (void *)RTR0MemObjAddress },
302	{ "RTR0MemObjAddressR3", (void *)RTR0MemObjAddressR3 },
303	{ "RTR0MemObjAllocContTag", (void *)RTR0MemObjAllocContTag },
304	{ "RTR0MemObjAllocLowTag", (void *)RTR0MemObjAllocLowTag },
305	{ "RTR0MemObjAllocPageTag", (void *)RTR0MemObjAllocPageTag },
306	{ "RTR0MemObjAllocPhysExTag", (void *)RTR0MemObjAllocPhysExTag },
307	{ "RTR0MemObjAllocPhysNCTag", (void *)RTR0MemObjAllocPhysNCTag },
308	{ "RTR0MemObjAllocPhysTag", (void *)RTR0MemObjAllocPhysTag },
309	{ "RTR0MemObjEnterPhysTag", (void *)RTR0MemObjEnterPhysTag },
310	{ "RTR0MemObjFree", (void *)RTR0MemObjFree },
311	{ "RTR0MemObjGetPagePhysAddr", (void *)RTR0MemObjGetPagePhysAddr },
312	{ "RTR0MemObjIsMapping", (void *)RTR0MemObjIsMapping },
313	{ "RTR0MemObjLockUserTag", (void *)RTR0MemObjLockUserTag },
314	{ "RTR0MemObjMapKernelExTag", (void *)RTR0MemObjMapKernelExTag },
315	{ "RTR0MemObjMapKernelTag", (void *)RTR0MemObjMapKernelTag },
316	{ "RTR0MemObjMapUserTag", (void *)RTR0MemObjMapUserTag },
317	{ "RTR0MemObjProtect", (void *)RTR0MemObjProtect },
318	{ "RTR0MemObjSize", (void *)RTR0MemObjSize },
319	{ "RTR0MemUserCopyFrom", (void *)RTR0MemUserCopyFrom },
320	{ "RTR0MemUserCopyTo", (void *)RTR0MemUserCopyTo },
321	{ "RTR0MemUserIsValidAddr", (void *)RTR0MemUserIsValidAddr },
322	{ "RTR0ProcHandleSelf", (void *)RTR0ProcHandleSelf },
323	{ "RTSemEventCreate", (void *)RTSemEventCreate },
324	{ "RTSemEventDestroy", (void *)RTSemEventDestroy },
325	{ "RTSemEventGetResolution", (void *)RTSemEventGetResolution },
326	{ "RTSemEventMultiCreate", (void *)RTSemEventMultiCreate },
327	{ "RTSemEventMultiDestroy", (void *)RTSemEventMultiDestroy },
328	{ "RTSemEventMultiGetResolution", (void *)RTSemEventMultiGetResolution },
329	{ "RTSemEventMultiReset", (void *)RTSemEventMultiReset },
330	{ "RTSemEventMultiSignal", (void *)RTSemEventMultiSignal },
331	{ "RTSemEventMultiWait", (void *)RTSemEventMultiWait },
332	{ "RTSemEventMultiWaitEx", (void *)RTSemEventMultiWaitEx },
333	{ "RTSemEventMultiWaitExDebug", (void *)RTSemEventMultiWaitExDebug },
334	{ "RTSemEventMultiWaitNoResume", (void *)RTSemEventMultiWaitNoResume },
335	{ "RTSemEventSignal", (void *)RTSemEventSignal },
336	{ "RTSemEventWait", (void *)RTSemEventWait },
337	{ "RTSemEventWaitEx", (void *)RTSemEventWaitEx },
338	{ "RTSemEventWaitExDebug", (void *)RTSemEventWaitExDebug },
339	{ "RTSemEventWaitNoResume", (void *)RTSemEventWaitNoResume },
340	{ "RTSemFastMutexCreate", (void *)RTSemFastMutexCreate },
341	{ "RTSemFastMutexDestroy", (void *)RTSemFastMutexDestroy },
342	{ "RTSemFastMutexRelease", (void *)RTSemFastMutexRelease },
343	{ "RTSemFastMutexRequest", (void *)RTSemFastMutexRequest },
344	{ "RTSemMutexCreate", (void *)RTSemMutexCreate },
345	{ "RTSemMutexDestroy", (void *)RTSemMutexDestroy },
346	{ "RTSemMutexRelease", (void *)RTSemMutexRelease },
347	{ "RTSemMutexRequest", (void *)RTSemMutexRequest },
348	{ "RTSemMutexRequestDebug", (void *)RTSemMutexRequestDebug },
349	{ "RTSemMutexRequestNoResume", (void *)RTSemMutexRequestNoResume },
350	{ "RTSemMutexRequestNoResumeDebug", (void *)RTSemMutexRequestNoResumeDebug },
351	{ "RTSpinlockAcquire", (void *)RTSpinlockAcquire },
352	{ "RTSpinlockCreate", (void *)RTSpinlockCreate },
353	{ "RTSpinlockDestroy", (void *)RTSpinlockDestroy },
354	{ "RTSpinlockRelease", (void *)RTSpinlockRelease },
355	{ "RTSpinlockReleaseNoInts", (void *)RTSpinlockReleaseNoInts },
356	{ "RTStrCopy", (void *)RTStrCopy },
357	{ "RTStrDupTag", (void *)RTStrDupTag },
358	{ "RTStrFormat", (void *)RTStrFormat },
359	{ "RTStrFormatNumber", (void *)RTStrFormatNumber },
360	{ "RTStrFormatTypeDeregister", (void *)RTStrFormatTypeDeregister },
361	{ "RTStrFormatTypeRegister", (void *)RTStrFormatTypeRegister },
362	{ "RTStrFormatTypeSetUser", (void *)RTStrFormatTypeSetUser },
363	{ "RTStrFormatV", (void *)RTStrFormatV },
364	{ "RTStrFree", (void *)RTStrFree },
365	{ "RTStrNCmp", (void *)RTStrNCmp },
366	{ "RTStrPrintf", (void *)RTStrPrintf },
367	{ "RTStrPrintfEx", (void *)RTStrPrintfEx },
368	{ "RTStrPrintfExV", (void *)RTStrPrintfExV },
369	{ "RTStrPrintfV", (void *)RTStrPrintfV },
370	{ "RTThreadCreate", (void *)RTThreadCreate },
371	{ "RTThreadCtxHooksAreRegistered", (void *)RTThreadCtxHooksAreRegistered },
372	{ "RTThreadCtxHooksCreate", (void *)RTThreadCtxHooksCreate },
373	{ "RTThreadCtxHooksDeregister", (void *)RTThreadCtxHooksDeregister },
374	{ "RTThreadCtxHooksRegister", (void *)RTThreadCtxHooksRegister },
375	{ "RTThreadCtxHooksRelease", (void *)RTThreadCtxHooksRelease },
376	{ "RTThreadCtxHooksRetain", (void *)RTThreadCtxHooksRetain },
377	{ "RTThreadGetName", (void *)RTThreadGetName },
378	{ "RTThreadGetNative", (void *)RTThreadGetNative },
379	{ "RTThreadGetType", (void *)RTThreadGetType },
380	{ "RTThreadIsInInterrupt", (void *)RTThreadIsInInterrupt },
381	{ "RTThreadNativeSelf", (void *)RTThreadNativeSelf },
382	{ "RTThreadPreemptDisable", (void *)RTThreadPreemptDisable },
383	{ "RTThreadPreemptIsEnabled", (void *)RTThreadPreemptIsEnabled },
384	{ "RTThreadPreemptIsPending", (void *)RTThreadPreemptIsPending },
385	{ "RTThreadPreemptIsPendingTrusty", (void *)RTThreadPreemptIsPendingTrusty },
386	{ "RTThreadPreemptIsPossible", (void *)RTThreadPreemptIsPossible },
387	{ "RTThreadPreemptRestore", (void *)RTThreadPreemptRestore },
388	{ "RTThreadSelf", (void *)RTThreadSelf },
389	{ "RTThreadSelfName", (void *)RTThreadSelfName },
390	{ "RTThreadSleep", (void *)RTThreadSleep },
391	{ "RTThreadUserReset", (void *)RTThreadUserReset },
392	{ "RTThreadUserSignal", (void *)RTThreadUserSignal },
393	{ "RTThreadUserWait", (void *)RTThreadUserWait },
394	{ "RTThreadUserWaitNoResume", (void *)RTThreadUserWaitNoResume },
395	{ "RTThreadWait", (void *)RTThreadWait },
396	{ "RTThreadWaitNoResume", (void *)RTThreadWaitNoResume },
397	{ "RTThreadYield", (void *)RTThreadYield },
398	{ "RTTimeMilliTS", (void *)RTTimeMilliTS },
399	{ "RTTimeNanoTS", (void *)RTTimeNanoTS },
400	{ "RTTimeNow", (void *)RTTimeNow },
401	{ "RTTimerCanDoHighResolution", (void *)RTTimerCanDoHighResolution },
402	{ "RTTimerChangeInterval", (void *)RTTimerChangeInterval },
403	{ "RTTimerCreate", (void *)RTTimerCreate },
404	{ "RTTimerCreateEx", (void *)RTTimerCreateEx },
405	{ "RTTimerDestroy", (void *)RTTimerDestroy },
406	{ "RTTimerGetSystemGranularity", (void *)RTTimerGetSystemGranularity },
407	{ "RTTimerReleaseSystemGranularity", (void *)RTTimerReleaseSystemGranularity },
408	{ "RTTimerRequestSystemGranularity", (void *)RTTimerRequestSystemGranularity },
409	{ "RTTimerStart", (void *)RTTimerStart },
410	{ "RTTimerStop", (void *)RTTimerStop },
411	{ "RTTimeSystemMilliTS", (void *)RTTimeSystemMilliTS },
412	{ "RTTimeSystemNanoTS", (void *)RTTimeSystemNanoTS },
413	{ "RTUuidCompare", (void *)RTUuidCompare },
414	{ "RTUuidCompareStr", (void *)RTUuidCompareStr },
415	{ "RTUuidFromStr", (void *)RTUuidFromStr },
416	/* SED: END */
417	};
418
419	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
420	/**
421	* Drag in the rest of IRPT since we share it with the
422	* rest of the kernel modules on darwin.
423	*/
424	PFNRT g_apfnVBoxDrvIPRTDeps[] =
425	{
426	/* VBoxNetAdp */
427	(PFNRT)RTRandBytes,
428	/* VBoxUSB */
429	(PFNRT)RTPathStripFilename,
430	NULL
431	};
432	#endif /* RT_OS_DARWIN \|\| RT_OS_SOLARIS \|\| RT_OS_SOLARIS */
433
434
435	/**
436	* Initializes the device extentsion structure.
437	*
438	* @returns IPRT status code.
439	* @param pDevExt The device extension to initialize.
440	* @param cbSession The size of the session structure. The size of
441	* SUPDRVSESSION may be smaller when SUPDRV_AGNOSTIC is
442	* defined because we're skipping the OS specific members
443	* then.
444	*/
445	int VBOXCALL supdrvInitDevExt(PSUPDRVDEVEXT pDevExt, size_t cbSession)
446	{
447	int rc;
448
449	#ifdef SUPDRV_WITH_RELEASE_LOGGER
450	/*
451	* Create the release log.
452	*/
453	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
454	PRTLOGGER pRelLogger;
455	rc = RTLogCreate(&pRelLogger, 0 /* fFlags */, "all",
456	"VBOX_RELEASE_LOG", RT_ELEMENTS(s_apszGroups), s_apszGroups, RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER, NULL);
457	if (RT_SUCCESS(rc))
458	RTLogRelSetDefaultInstance(pRelLogger);
459	/** @todo Add native hook for getting logger config parameters and setting
460	* them. On linux we should use the module parameter stuff... */
461	#endif
462
463	/*
464	* Initialize it.
465	*/
466	memset(pDevExt, 0, sizeof(pDevExt)); / Does not wipe OS specific tail section of the structure. */
467	pDevExt->Spinlock = NIL_RTSPINLOCK;
468	pDevExt->hGipSpinlock = NIL_RTSPINLOCK;
469	pDevExt->hSessionHashTabSpinlock = NIL_RTSPINLOCK;
470	rc = RTSpinlockCreate(&pDevExt->Spinlock, RTSPINLOCK_FLAGS_INTERRUPT_SAFE, "SUPDrvDevExt");
471	if (RT_SUCCESS(rc))
472	rc = RTSpinlockCreate(&pDevExt->hGipSpinlock, RTSPINLOCK_FLAGS_INTERRUPT_SAFE, "SUPDrvGip");
473	if (RT_SUCCESS(rc))
474	rc = RTSpinlockCreate(&pDevExt->hSessionHashTabSpinlock, RTSPINLOCK_FLAGS_INTERRUPT_SAFE, "SUPDrvSession");
475
476	if (RT_SUCCESS(rc))
477	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
478	rc = RTSemMutexCreate(&pDevExt->mtxLdr);
479	#else
480	rc = RTSemFastMutexCreate(&pDevExt->mtxLdr);
481	#endif
482	if (RT_SUCCESS(rc))
483	{
484	rc = RTSemFastMutexCreate(&pDevExt->mtxComponentFactory);
485	if (RT_SUCCESS(rc))
486	{
487	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
488	rc = RTSemMutexCreate(&pDevExt->mtxGip);
489	#else
490	rc = RTSemFastMutexCreate(&pDevExt->mtxGip);
491	#endif
492	if (RT_SUCCESS(rc))
493	{
494	rc = supdrvGipCreate(pDevExt);
495	if (RT_SUCCESS(rc))
496	{
497	rc = supdrvTracerInit(pDevExt);
498	if (RT_SUCCESS(rc))
499	{
500	pDevExt->pLdrInitImage = NULL;
501	pDevExt->hLdrInitThread = NIL_RTNATIVETHREAD;
502	pDevExt->u32Cookie = BIRD; /** @todo make this random? */
503	pDevExt->cbSession = (uint32_t)cbSession;
504
505	/*
506	* Fixup the absolute symbols.
507	*
508	* Because of the table indexing assumptions we'll have a little #ifdef orgy
509	* here rather than distributing this to OS specific files. At least for now.
510	*/
511	#ifdef RT_OS_DARWIN
512	# if ARCH_BITS == 32
513	if (SUPR0GetPagingMode() >= SUPPAGINGMODE_AMD64)
514	{
515	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
516	g_aFunctions[1].pfn = (void )0x80; / SUPR0Abs64bitKernelCS - KERNEL64_CS, seg.h */
517	g_aFunctions[2].pfn = (void )0x88; / SUPR0Abs64bitKernelSS - KERNEL64_SS, seg.h */
518	g_aFunctions[3].pfn = (void )0x88; / SUPR0Abs64bitKernelDS - KERNEL64_SS, seg.h */
519	}
520	else
521	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
522	g_aFunctions[4].pfn = (void )0x08; / SUPR0AbsKernelCS - KERNEL_CS, seg.h */
523	g_aFunctions[5].pfn = (void )0x10; / SUPR0AbsKernelSS - KERNEL_DS, seg.h */
524	g_aFunctions[6].pfn = (void )0x10; / SUPR0AbsKernelDS - KERNEL_DS, seg.h */
525	g_aFunctions[7].pfn = (void )0x10; / SUPR0AbsKernelES - KERNEL_DS, seg.h */
526	g_aFunctions[8].pfn = (void )0x10; / SUPR0AbsKernelFS - KERNEL_DS, seg.h */
527	g_aFunctions[9].pfn = (void )0x48; / SUPR0AbsKernelGS - CPU_DATA_GS, seg.h */
528	# else /* 64-bit darwin: */
529	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
530	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
531	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
532	g_aFunctions[3].pfn = (void )0; / SUPR0Abs64bitKernelDS */
533	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
534	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
535	g_aFunctions[6].pfn = (void )0; / SUPR0AbsKernelDS */
536	g_aFunctions[7].pfn = (void )0; / SUPR0AbsKernelES */
537	g_aFunctions[8].pfn = (void )0; / SUPR0AbsKernelFS */
538	g_aFunctions[9].pfn = (void )0; / SUPR0AbsKernelGS */
539
540	# endif
541	#else /* !RT_OS_DARWIN */
542	# if ARCH_BITS == 64
543	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
544	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
545	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
546	g_aFunctions[3].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0Abs64bitKernelDS */
547	# else
548	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
549	# endif
550	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
551	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
552	g_aFunctions[6].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0AbsKernelDS */
553	g_aFunctions[7].pfn = (void )(uintptr_t)ASMGetES(); / SUPR0AbsKernelES */
554	g_aFunctions[8].pfn = (void )(uintptr_t)ASMGetFS(); / SUPR0AbsKernelFS */
555	g_aFunctions[9].pfn = (void )(uintptr_t)ASMGetGS(); / SUPR0AbsKernelGS */
556	#endif /* !RT_OS_DARWIN */
557	return VINF_SUCCESS;
558	}
559
560	supdrvGipDestroy(pDevExt);
561	}
562
563	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
564	RTSemMutexDestroy(pDevExt->mtxGip);
565	pDevExt->mtxGip = NIL_RTSEMMUTEX;
566	#else
567	RTSemFastMutexDestroy(pDevExt->mtxGip);
568	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
569	#endif
570	}
571	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
572	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
573	}
574	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
575	RTSemMutexDestroy(pDevExt->mtxLdr);
576	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
577	#else
578	RTSemFastMutexDestroy(pDevExt->mtxLdr);
579	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
580	#endif
581	}
582
583	RTSpinlockDestroy(pDevExt->Spinlock);
584	pDevExt->Spinlock = NIL_RTSPINLOCK;
585	RTSpinlockDestroy(pDevExt->hGipSpinlock);
586	pDevExt->hGipSpinlock = NIL_RTSPINLOCK;
587	RTSpinlockDestroy(pDevExt->hSessionHashTabSpinlock);
588	pDevExt->hSessionHashTabSpinlock = NIL_RTSPINLOCK;
589
590	#ifdef SUPDRV_WITH_RELEASE_LOGGER
591	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
592	RTLogDestroy(RTLogSetDefaultInstance(NULL));
593	#endif
594
595	return rc;
596	}
597
598
599	/**
600	* Delete the device extension (e.g. cleanup members).
601	*
602	* @param pDevExt The device extension to delete.
603	*/
604	void VBOXCALL supdrvDeleteDevExt(PSUPDRVDEVEXT pDevExt)
605	{
606	PSUPDRVOBJ pObj;
607	PSUPDRVUSAGE pUsage;
608
609	/*
610	* Kill mutexes and spinlocks.
611	*/
612	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
613	RTSemMutexDestroy(pDevExt->mtxGip);
614	pDevExt->mtxGip = NIL_RTSEMMUTEX;
615	#else
616	RTSemFastMutexDestroy(pDevExt->mtxGip);
617	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
618	#endif
619	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
620	RTSemMutexDestroy(pDevExt->mtxLdr);
621	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
622	#else
623	RTSemFastMutexDestroy(pDevExt->mtxLdr);
624	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
625	#endif
626	RTSpinlockDestroy(pDevExt->Spinlock);
627	pDevExt->Spinlock = NIL_RTSPINLOCK;
628	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
629	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
630	RTSpinlockDestroy(pDevExt->hSessionHashTabSpinlock);
631	pDevExt->hSessionHashTabSpinlock = NIL_RTSPINLOCK;
632
633	/*
634	* Free lists.
635	*/
636	/* objects. */
637	pObj = pDevExt->pObjs;
638	Assert(!pObj); /* (can trigger on forced unloads) */
639	pDevExt->pObjs = NULL;
640	while (pObj)
641	{
642	void *pvFree = pObj;
643	pObj = pObj->pNext;
644	RTMemFree(pvFree);
645	}
646
647	/* usage records. */
648	pUsage = pDevExt->pUsageFree;
649	pDevExt->pUsageFree = NULL;
650	while (pUsage)
651	{
652	void *pvFree = pUsage;
653	pUsage = pUsage->pNext;
654	RTMemFree(pvFree);
655	}
656
657	/* kill the GIP. */
658	supdrvGipDestroy(pDevExt);
659	RTSpinlockDestroy(pDevExt->hGipSpinlock);
660	pDevExt->hGipSpinlock = NIL_RTSPINLOCK;
661
662	supdrvTracerTerm(pDevExt);
663
664	#ifdef SUPDRV_WITH_RELEASE_LOGGER
665	/* destroy the loggers. */
666	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
667	RTLogDestroy(RTLogSetDefaultInstance(NULL));
668	#endif
669	}
670
671
672	/**
673	* Create session.
674	*
675	* @returns IPRT status code.
676	* @param pDevExt Device extension.
677	* @param fUser Flag indicating whether this is a user or kernel
678	* session.
679	* @param fUnrestricted Unrestricted access (system) or restricted access
680	* (user)?
681	* @param ppSession Where to store the pointer to the session data.
682	*/
683	int VBOXCALL supdrvCreateSession(PSUPDRVDEVEXT pDevExt, bool fUser, bool fUnrestricted, PSUPDRVSESSION *ppSession)
684	{
685	int rc;
686	PSUPDRVSESSION pSession;
687
688	if (!SUP_IS_DEVEXT_VALID(pDevExt))
689	return VERR_INVALID_PARAMETER;
690
691	/*
692	* Allocate memory for the session data.
693	*/
694	pSession = *ppSession = (PSUPDRVSESSION)RTMemAllocZ(pDevExt->cbSession);
695	if (pSession)
696	{
697	/* Initialize session data. */
698	rc = RTSpinlockCreate(&pSession->Spinlock, RTSPINLOCK_FLAGS_INTERRUPT_UNSAFE, "SUPDrvSession");
699	if (!rc)
700	{
701	rc = RTHandleTableCreateEx(&pSession->hHandleTable,
702	RTHANDLETABLE_FLAGS_LOCKED_IRQ_SAFE \| RTHANDLETABLE_FLAGS_CONTEXT,
703	1 /uBase/, 32768 /cMax/, supdrvSessionObjHandleRetain, pSession);
704	if (RT_SUCCESS(rc))
705	{
706	Assert(pSession->Spinlock != NIL_RTSPINLOCK);
707	pSession->pDevExt = pDevExt;
708	pSession->u32Cookie = BIRD_INV;
709	pSession->fUnrestricted = fUnrestricted;
710	/pSession->fInHashTable = false; /
711	pSession->cRefs = 1;
712	/*pSession->pCommonNextHash = NULL;
713	pSession->ppOsSessionPtr = NULL; */
714	if (fUser)
715	{
716	pSession->Process = RTProcSelf();
717	pSession->R0Process = RTR0ProcHandleSelf();
718	}
719	else
720	{
721	pSession->Process = NIL_RTPROCESS;
722	pSession->R0Process = NIL_RTR0PROCESS;
723	}
724	/*pSession->pLdrUsage = NULL;
725	pSession->pVM = NULL;
726	pSession->pUsage = NULL;
727	pSession->pGip = NULL;
728	pSession->fGipReferenced = false;
729	pSession->Bundle.cUsed = 0; */
730	pSession->Uid = NIL_RTUID;
731	pSession->Gid = NIL_RTGID;
732	/pSession->uTracerData = 0;/
733	pSession->hTracerCaller = NIL_RTNATIVETHREAD;
734	RTListInit(&pSession->TpProviders);
735	/pSession->cTpProviders = 0;/
736	/pSession->cTpProbesFiring = 0;/
737	RTListInit(&pSession->TpUmods);
738	/RT_ZERO(pSession->apTpLookupTable);/
739
740	VBOXDRV_SESSION_CREATE(pSession, fUser);
741	LogFlow(("Created session %p initial cookie=%#x\n", pSession, pSession->u32Cookie));
742	return VINF_SUCCESS;
743	}
744
745	RTSpinlockDestroy(pSession->Spinlock);
746	}
747	RTMemFree(pSession);
748	*ppSession = NULL;
749	Log(("Failed to create spinlock, rc=%d!\n", rc));
750	}
751	else
752	rc = VERR_NO_MEMORY;
753
754	return rc;
755	}
756
757
758	/**
759	* Cleans up the session in the context of the process to which it belongs, the
760	* caller will free the session and the session spinlock.
761	*
762	* This should normally occur when the session is closed or as the process
763	* exits. Careful reference counting in the OS specfic code makes sure that
764	* there cannot be any races between process/handle cleanup callbacks and
765	* threads doing I/O control calls.
766	*
767	* @param pDevExt The device extension.
768	* @param pSession Session data.
769	*/
770	static void supdrvCleanupSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
771	{
772	int rc;
773	PSUPDRVBUNDLE pBundle;
774	LogFlow(("supdrvCleanupSession: pSession=%p\n", pSession));
775
776	Assert(!pSession->fInHashTable);
777	Assert(!pSession->ppOsSessionPtr);
778	AssertReleaseMsg(pSession->R0Process == RTR0ProcHandleSelf() \|\| pSession->R0Process == NIL_RTR0PROCESS,
779	("R0Process=%p cur=%p; Process=%u curpid=%u\n", RTR0ProcHandleSelf(), RTProcSelf()));
780
781	/*
782	* Remove logger instances related to this session.
783	*/
784	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pSession);
785
786	/*
787	* Destroy the handle table.
788	*/
789	rc = RTHandleTableDestroy(pSession->hHandleTable, supdrvSessionObjHandleDelete, pSession);
790	AssertRC(rc);
791	pSession->hHandleTable = NIL_RTHANDLETABLE;
792
793	/*
794	* Release object references made in this session.
795	* In theory there should be noone racing us in this session.
796	*/
797	Log2(("release objects - start\n"));
798	if (pSession->pUsage)
799	{
800	PSUPDRVUSAGE pUsage;
801	RTSpinlockAcquire(pDevExt->Spinlock);
802
803	while ((pUsage = pSession->pUsage) != NULL)
804	{
805	PSUPDRVOBJ pObj = pUsage->pObj;
806	pSession->pUsage = pUsage->pNext;
807
808	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
809	if (pUsage->cUsage < pObj->cUsage)
810	{
811	pObj->cUsage -= pUsage->cUsage;
812	RTSpinlockRelease(pDevExt->Spinlock);
813	}
814	else
815	{
816	/* Destroy the object and free the record. */
817	if (pDevExt->pObjs == pObj)
818	pDevExt->pObjs = pObj->pNext;
819	else
820	{
821	PSUPDRVOBJ pObjPrev;
822	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
823	if (pObjPrev->pNext == pObj)
824	{
825	pObjPrev->pNext = pObj->pNext;
826	break;
827	}
828	Assert(pObjPrev);
829	}
830	RTSpinlockRelease(pDevExt->Spinlock);
831
832	Log(("supdrvCleanupSession: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
833	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
834	if (pObj->pfnDestructor)
835	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
836	RTMemFree(pObj);
837	}
838
839	/* free it and continue. */
840	RTMemFree(pUsage);
841
842	RTSpinlockAcquire(pDevExt->Spinlock);
843	}
844
845	RTSpinlockRelease(pDevExt->Spinlock);
846	AssertMsg(!pSession->pUsage, ("Some buster reregistered an object during desturction!\n"));
847	}
848	Log2(("release objects - done\n"));
849
850	/*
851	* Do tracer cleanups related to this session.
852	*/
853	Log2(("release tracer stuff - start\n"));
854	supdrvTracerCleanupSession(pDevExt, pSession);
855	Log2(("release tracer stuff - end\n"));
856
857	/*
858	* Release memory allocated in the session.
859	*
860	* We do not serialize this as we assume that the application will
861	* not allocated memory while closing the file handle object.
862	*/
863	Log2(("freeing memory:\n"));
864	pBundle = &pSession->Bundle;
865	while (pBundle)
866	{
867	PSUPDRVBUNDLE pToFree;
868	unsigned i;
869
870	/*
871	* Check and unlock all entries in the bundle.
872	*/
873	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
874	{
875	if (pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ)
876	{
877	Log2(("eType=%d pvR0=%p pvR3=%p cb=%ld\n", pBundle->aMem[i].eType, RTR0MemObjAddress(pBundle->aMem[i].MemObj),
878	(void *)RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3), (long)RTR0MemObjSize(pBundle->aMem[i].MemObj)));
879	if (pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ)
880	{
881	rc = RTR0MemObjFree(pBundle->aMem[i].MapObjR3, false);
882	AssertRC(rc); /** @todo figure out how to handle this. */
883	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
884	}
885	rc = RTR0MemObjFree(pBundle->aMem[i].MemObj, true /* fFreeMappings */);
886	AssertRC(rc); /** @todo figure out how to handle this. */
887	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
888	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
889	}
890	}
891
892	/*
893	* Advance and free previous bundle.
894	*/
895	pToFree = pBundle;
896	pBundle = pBundle->pNext;
897
898	pToFree->pNext = NULL;
899	pToFree->cUsed = 0;
900	if (pToFree != &pSession->Bundle)
901	RTMemFree(pToFree);
902	}
903	Log2(("freeing memory - done\n"));
904
905	/*
906	* Deregister component factories.
907	*/
908	RTSemFastMutexRequest(pDevExt->mtxComponentFactory);
909	Log2(("deregistering component factories:\n"));
910	if (pDevExt->pComponentFactoryHead)
911	{
912	PSUPDRVFACTORYREG pPrev = NULL;
913	PSUPDRVFACTORYREG pCur = pDevExt->pComponentFactoryHead;
914	while (pCur)
915	{
916	if (pCur->pSession == pSession)
917	{
918	/* unlink it */
919	PSUPDRVFACTORYREG pNext = pCur->pNext;
920	if (pPrev)
921	pPrev->pNext = pNext;
922	else
923	pDevExt->pComponentFactoryHead = pNext;
924
925	/* free it */
926	pCur->pNext = NULL;
927	pCur->pSession = NULL;
928	pCur->pFactory = NULL;
929	RTMemFree(pCur);
930
931	/* next */
932	pCur = pNext;
933	}
934	else
935	{
936	/* next */
937	pPrev = pCur;
938	pCur = pCur->pNext;
939	}
940	}
941	}
942	RTSemFastMutexRelease(pDevExt->mtxComponentFactory);
943	Log2(("deregistering component factories - done\n"));
944
945	/*
946	* Loaded images needs to be dereferenced and possibly freed up.
947	*/
948	supdrvLdrLock(pDevExt);
949	Log2(("freeing images:\n"));
950	if (pSession->pLdrUsage)
951	{
952	PSUPDRVLDRUSAGE pUsage = pSession->pLdrUsage;
953	pSession->pLdrUsage = NULL;
954	while (pUsage)
955	{
956	void *pvFree = pUsage;
957	PSUPDRVLDRIMAGE pImage = pUsage->pImage;
958	if (pImage->cUsage > pUsage->cUsage)
959	pImage->cUsage -= pUsage->cUsage;
960	else
961	supdrvLdrFree(pDevExt, pImage);
962	pUsage->pImage = NULL;
963	pUsage = pUsage->pNext;
964	RTMemFree(pvFree);
965	}
966	}
967	supdrvLdrUnlock(pDevExt);
968	Log2(("freeing images - done\n"));
969
970	/*
971	* Unmap the GIP.
972	*/
973	Log2(("umapping GIP:\n"));
974	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
975	{
976	SUPR0GipUnmap(pSession);
977	pSession->fGipReferenced = 0;
978	}
979	Log2(("umapping GIP - done\n"));
980	}
981
982
983	/**
984	* Common code for freeing a session when the reference count reaches zero.
985	*
986	* @param pDevExt Device extension.
987	* @param pSession Session data.
988	* This data will be freed by this routine.
989	*/
990	static void supdrvDestroySession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
991	{
992	VBOXDRV_SESSION_CLOSE(pSession);
993
994	/*
995	* Cleanup the session first.
996	*/
997	supdrvCleanupSession(pDevExt, pSession);
998	supdrvOSCleanupSession(pDevExt, pSession);
999
1000	/*
1001	* Free the rest of the session stuff.
1002	*/
1003	RTSpinlockDestroy(pSession->Spinlock);
1004	pSession->Spinlock = NIL_RTSPINLOCK;
1005	pSession->pDevExt = NULL;
1006	RTMemFree(pSession);
1007	LogFlow(("supdrvDestroySession: returns\n"));
1008	}
1009
1010
1011	/**
1012	* Inserts the session into the global hash table.
1013	*
1014	* @retval VINF_SUCCESS on success.
1015	* @retval VERR_WRONG_ORDER if the session was already inserted (asserted).
1016	* @retval VERR_INVALID_PARAMETER if the session handle is invalid or a ring-0
1017	* session (asserted).
1018	* @retval VERR_DUPLICATE if there is already a session for that pid.
1019	*
1020	* @param pDevExt The device extension.
1021	* @param pSession The session.
1022	* @param ppOsSessionPtr Pointer to the OS session pointer, if any is
1023	* available and used. This will set to point to the
1024	* session while under the protection of the session
1025	* hash table spinlock. It will also be kept in
1026	* PSUPDRVSESSION::ppOsSessionPtr for lookup and
1027	* cleanup use.
1028	* @param pvUser Argument for supdrvOSSessionHashTabInserted.
1029	*/
1030	int VBOXCALL supdrvSessionHashTabInsert(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVSESSION *ppOsSessionPtr,
1031	void *pvUser)
1032	{
1033	PSUPDRVSESSION pCur;
1034	unsigned iHash;
1035
1036	/*
1037	* Validate input.
1038	*/
1039	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1040	AssertReturn(pSession->R0Process != NIL_RTR0PROCESS, VERR_INVALID_PARAMETER);
1041
1042	/*
1043	* Calculate the hash table index and acquire the spinlock.
1044	*/
1045	iHash = SUPDRV_SESSION_HASH(pSession->Process);
1046
1047	RTSpinlockAcquire(pDevExt->hSessionHashTabSpinlock);
1048
1049	/*
1050	* If there are a collisions, we need to carefully check if we got a
1051	* duplicate. There can only be one open session per process.
1052	*/
1053	pCur = pDevExt->apSessionHashTab[iHash];
1054	if (pCur)
1055	{
1056	while (pCur && pCur->Process != pSession->Process)
1057	pCur = pCur->pCommonNextHash;
1058
1059	if (pCur)
1060	{
1061	RTSpinlockRelease(pDevExt->hSessionHashTabSpinlock);
1062	if (pCur == pSession)
1063	{
1064	Assert(pSession->fInHashTable);
1065	AssertFailed();
1066	return VERR_WRONG_ORDER;
1067	}
1068	Assert(!pSession->fInHashTable);
1069	if (pCur->R0Process == pSession->R0Process)
1070	return VERR_RESOURCE_IN_USE;
1071	return VERR_DUPLICATE;
1072	}
1073	}
1074	Assert(!pSession->fInHashTable);
1075	Assert(!pSession->ppOsSessionPtr);
1076
1077	/*
1078	* Insert it, doing a callout to the OS specific code in case it has
1079	* anything it wishes to do while we're holding the spinlock.
1080	*/
1081	pSession->pCommonNextHash = pDevExt->apSessionHashTab[iHash];
1082	pDevExt->apSessionHashTab[iHash] = pSession;
1083	pSession->fInHashTable = true;
1084	ASMAtomicIncS32(&pDevExt->cSessions);
1085
1086	pSession->ppOsSessionPtr = ppOsSessionPtr;
1087	if (ppOsSessionPtr)
1088	ASMAtomicWritePtr(ppOsSessionPtr, pSession);
1089
1090	supdrvOSSessionHashTabInserted(pDevExt, pSession, pvUser);
1091
1092	/*
1093	* Retain a reference for the pointer in the session table.
1094	*/
1095	ASMAtomicIncU32(&pSession->cRefs);
1096
1097	RTSpinlockRelease(pDevExt->hSessionHashTabSpinlock);
1098	return VINF_SUCCESS;
1099	}
1100
1101
1102	/**
1103	* Removes the session from the global hash table.
1104	*
1105	* @retval VINF_SUCCESS on success.
1106	* @retval VERR_NOT_FOUND if the session was already removed (asserted).
1107	* @retval VERR_INVALID_PARAMETER if the session handle is invalid or a ring-0
1108	* session (asserted).
1109	*
1110	* @param pDevExt The device extension.
1111	* @param pSession The session. The caller is expected to have a reference
1112	* to this so it won't croak on us when we release the hash
1113	* table reference.
1114	* @param pvUser OS specific context value for the
1115	* supdrvOSSessionHashTabInserted callback.
1116	*/
1117	int VBOXCALL supdrvSessionHashTabRemove(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, void *pvUser)
1118	{
1119	PSUPDRVSESSION pCur;
1120	unsigned iHash;
1121	int32_t cRefs;
1122
1123	/*
1124	* Validate input.
1125	*/
1126	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1127	AssertReturn(pSession->R0Process != NIL_RTR0PROCESS, VERR_INVALID_PARAMETER);
1128
1129	/*
1130	* Calculate the hash table index and acquire the spinlock.
1131	*/
1132	iHash = SUPDRV_SESSION_HASH(pSession->Process);
1133
1134	RTSpinlockAcquire(pDevExt->hSessionHashTabSpinlock);
1135
1136	/*
1137	* Unlink it.
1138	*/
1139	pCur = pDevExt->apSessionHashTab[iHash];
1140	if (pCur == pSession)
1141	pDevExt->apSessionHashTab[iHash] = pSession->pCommonNextHash;
1142	else
1143	{
1144	PSUPDRVSESSION pPrev = pCur;
1145	while (pCur && pCur != pSession)
1146	{
1147	pPrev = pCur;
1148	pCur = pCur->pCommonNextHash;
1149	}
1150	if (pCur)
1151	pPrev->pCommonNextHash = pCur->pCommonNextHash;
1152	else
1153	{
1154	Assert(!pSession->fInHashTable);
1155	RTSpinlockRelease(pDevExt->hSessionHashTabSpinlock);
1156	return VERR_NOT_FOUND;
1157	}
1158	}
1159
1160	pSession->pCommonNextHash = NULL;
1161	pSession->fInHashTable = false;
1162
1163	ASMAtomicDecU32(&pDevExt->cSessions);
1164
1165	/*
1166	* Clear OS specific session pointer if available and do the OS callback.
1167	*/
1168	if (pSession->ppOsSessionPtr)
1169	{
1170	ASMAtomicCmpXchgPtr(pSession->ppOsSessionPtr, NULL, pSession);
1171	pSession->ppOsSessionPtr = NULL;
1172	}
1173
1174	supdrvOSSessionHashTabRemoved(pDevExt, pSession, pvUser);
1175
1176	RTSpinlockRelease(pDevExt->hSessionHashTabSpinlock);
1177
1178	/*
1179	* Drop the reference the hash table had to the session. This shouldn't
1180	* be the last reference!
1181	*/
1182	cRefs = ASMAtomicDecU32(&pSession->cRefs);
1183	Assert(cRefs > 0 && cRefs < _1M);
1184	if (cRefs == 0)
1185	supdrvDestroySession(pDevExt, pSession);
1186
1187	return VINF_SUCCESS;
1188	}
1189
1190
1191	/**
1192	* Looks up the session for the current process in the global hash table or in
1193	* OS specific pointer.
1194	*
1195	* @returns Pointer to the session with a reference that the caller must
1196	* release. If no valid session was found, NULL is returned.
1197	*
1198	* @param pDevExt The device extension.
1199	* @param Process The process ID.
1200	* @param R0Process The ring-0 process handle.
1201	* @param ppOsSessionPtr The OS session pointer if available. If not NULL,
1202	* this is used instead of the hash table. For
1203	* additional safety it must then be equal to the
1204	* SUPDRVSESSION::ppOsSessionPtr member.
1205	* This can be NULL even if the OS has a session
1206	* pointer.
1207	*/
1208	PSUPDRVSESSION VBOXCALL supdrvSessionHashTabLookup(PSUPDRVDEVEXT pDevExt, RTPROCESS Process, RTR0PROCESS R0Process,
1209	PSUPDRVSESSION *ppOsSessionPtr)
1210	{
1211	PSUPDRVSESSION pCur;
1212	unsigned iHash;
1213
1214	/*
1215	* Validate input.
1216	*/
1217	AssertReturn(R0Process != NIL_RTR0PROCESS, NULL);
1218
1219	/*
1220	* Calculate the hash table index and acquire the spinlock.
1221	*/
1222	iHash = SUPDRV_SESSION_HASH(Process);
1223
1224	RTSpinlockAcquire(pDevExt->hSessionHashTabSpinlock);
1225
1226	/*
1227	* If an OS session pointer is provided, always use it.
1228	*/
1229	if (ppOsSessionPtr)
1230	{
1231	pCur = *ppOsSessionPtr;
1232	if ( pCur
1233	&& ( pCur->ppOsSessionPtr != ppOsSessionPtr
1234	\|\| pCur->Process != Process
1235	\|\| pCur->R0Process != R0Process) )
1236	pCur = NULL;
1237	}
1238	else
1239	{
1240	/*
1241	* Otherwise, do the hash table lookup.
1242	*/
1243	pCur = pDevExt->apSessionHashTab[iHash];
1244	while ( pCur
1245	&& ( pCur->Process != Process
1246	\|\| pCur->R0Process != R0Process) )
1247	pCur = pCur->pCommonNextHash;
1248	}
1249
1250	/*
1251	* Retain the session.
1252	*/
1253	if (pCur)
1254	{
1255	uint32_t cRefs = ASMAtomicIncU32(&pCur->cRefs);
1256	NOREF(cRefs);
1257	Assert(cRefs > 1 && cRefs < _1M);
1258	}
1259
1260	RTSpinlockRelease(pDevExt->hSessionHashTabSpinlock);
1261
1262	return pCur;
1263	}
1264
1265
1266	/**
1267	* Retain a session to make sure it doesn't go away while it is in use.
1268	*
1269	* @returns New reference count on success, UINT32_MAX on failure.
1270	* @param pSession Session data.
1271	*/
1272	uint32_t VBOXCALL supdrvSessionRetain(PSUPDRVSESSION pSession)
1273	{
1274	uint32_t cRefs;
1275	AssertPtrReturn(pSession, UINT32_MAX);
1276	AssertReturn(SUP_IS_SESSION_VALID(pSession), UINT32_MAX);
1277
1278	cRefs = ASMAtomicIncU32(&pSession->cRefs);
1279	AssertMsg(cRefs > 1 && cRefs < _1M, ("%#x %p\n", cRefs, pSession));
1280	return cRefs;
1281	}
1282
1283
1284	/**
1285	* Releases a given session.
1286	*
1287	* @returns New reference count on success (0 if closed), UINT32_MAX on failure.
1288	* @param pSession Session data.
1289	*/
1290	uint32_t VBOXCALL supdrvSessionRelease(PSUPDRVSESSION pSession)
1291	{
1292	uint32_t cRefs;
1293	AssertPtrReturn(pSession, UINT32_MAX);
1294	AssertReturn(SUP_IS_SESSION_VALID(pSession), UINT32_MAX);
1295
1296	cRefs = ASMAtomicDecU32(&pSession->cRefs);
1297	AssertMsg(cRefs < _1M, ("%#x %p\n", cRefs, pSession));
1298	if (cRefs == 0)
1299	supdrvDestroySession(pSession->pDevExt, pSession);
1300	return cRefs;
1301	}
1302
1303
1304	/**
1305	* RTHandleTableDestroy callback used by supdrvCleanupSession.
1306	*
1307	* @returns IPRT status code, see SUPR0ObjAddRef.
1308	* @param hHandleTable The handle table handle. Ignored.
1309	* @param pvObj The object pointer.
1310	* @param pvCtx Context, the handle type. Ignored.
1311	* @param pvUser Session pointer.
1312	*/
1313	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser)
1314	{
1315	NOREF(pvCtx);
1316	NOREF(hHandleTable);
1317	return SUPR0ObjAddRefEx(pvObj, (PSUPDRVSESSION)pvUser, true /fNoBlocking/);
1318	}
1319
1320
1321	/**
1322	* RTHandleTableDestroy callback used by supdrvCleanupSession.
1323	*
1324	* @param hHandleTable The handle table handle. Ignored.
1325	* @param h The handle value. Ignored.
1326	* @param pvObj The object pointer.
1327	* @param pvCtx Context, the handle type. Ignored.
1328	* @param pvUser Session pointer.
1329	*/
1330	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser)
1331	{
1332	NOREF(pvCtx);
1333	NOREF(h);
1334	NOREF(hHandleTable);
1335	SUPR0ObjRelease(pvObj, (PSUPDRVSESSION)pvUser);
1336	}
1337
1338
1339	/**
1340	* Fast path I/O Control worker.
1341	*
1342	* @returns VBox status code that should be passed down to ring-3 unchanged.
1343	* @param uIOCtl Function number.
1344	* @param idCpu VMCPU id.
1345	* @param pDevExt Device extention.
1346	* @param pSession Session data.
1347	*/
1348	int VBOXCALL supdrvIOCtlFast(uintptr_t uIOCtl, VMCPUID idCpu, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
1349	{
1350	/*
1351	* We check the two prereqs after doing this only to allow the compiler to optimize things better.
1352	*/
1353	if (RT_LIKELY( RT_VALID_PTR(pSession)
1354	&& pSession->pVM
1355	&& pDevExt->pfnVMMR0EntryFast))
1356	{
1357	switch (uIOCtl)
1358	{
1359	case SUP_IOCTL_FAST_DO_RAW_RUN:
1360	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_RAW_RUN);
1361	break;
1362	case SUP_IOCTL_FAST_DO_HM_RUN:
1363	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_HM_RUN);
1364	break;
1365	case SUP_IOCTL_FAST_DO_NOP:
1366	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_NOP);
1367	break;
1368	default:
1369	return VERR_INTERNAL_ERROR;
1370	}
1371	return VINF_SUCCESS;
1372	}
1373	return VERR_INTERNAL_ERROR;
1374	}
1375
1376
1377	/**
1378	* Helper for supdrvIOCtl. Check if pszStr contains any character of pszChars.
1379	* We would use strpbrk here if this function would be contained in the RedHat kABI white
1380	* list, see http://www.kerneldrivers.org/RHEL5.
1381	*
1382	* @returns 1 if pszStr does contain any character of pszChars, 0 otherwise.
1383	* @param pszStr String to check
1384	* @param pszChars Character set
1385	*/
1386	static int supdrvCheckInvalidChar(const char pszStr, const char pszChars)
1387	{
1388	int chCur;
1389	while ((chCur = *pszStr++) != '\0')
1390	{
1391	int ch;
1392	const char *psz = pszChars;
1393	while ((ch = *psz++) != '\0')
1394	if (ch == chCur)
1395	return 1;
1396
1397	}
1398	return 0;
1399	}
1400
1401
1402
1403	/**
1404	* I/O Control inner worker (tracing reasons).
1405	*
1406	* @returns IPRT status code.
1407	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1408	*
1409	* @param uIOCtl Function number.
1410	* @param pDevExt Device extention.
1411	* @param pSession Session data.
1412	* @param pReqHdr The request header.
1413	*/
1414	static int supdrvIOCtlInnerUnrestricted(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1415	{
1416	/*
1417	* Validation macros
1418	*/
1419	#define REQ_CHECK_SIZES_EX(Name, cbInExpect, cbOutExpect) \
1420	do { \
1421	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect) \|\| pReqHdr->cbOut != (cbOutExpect))) \
1422	{ \
1423	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld. cbOut=%ld expected %ld.\n", \
1424	(long)pReqHdr->cbIn, (long)(cbInExpect), (long)pReqHdr->cbOut, (long)(cbOutExpect))); \
1425	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1426	} \
1427	} while (0)
1428
1429	#define REQ_CHECK_SIZES(Name) REQ_CHECK_SIZES_EX(Name, Name ## _SIZE_IN, Name ## _SIZE_OUT)
1430
1431	#define REQ_CHECK_SIZE_IN(Name, cbInExpect) \
1432	do { \
1433	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect))) \
1434	{ \
1435	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld.\n", \
1436	(long)pReqHdr->cbIn, (long)(cbInExpect))); \
1437	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1438	} \
1439	} while (0)
1440
1441	#define REQ_CHECK_SIZE_OUT(Name, cbOutExpect) \
1442	do { \
1443	if (RT_UNLIKELY(pReqHdr->cbOut != (cbOutExpect))) \
1444	{ \
1445	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbOut=%ld expected %ld.\n", \
1446	(long)pReqHdr->cbOut, (long)(cbOutExpect))); \
1447	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1448	} \
1449	} while (0)
1450
1451	#define REQ_CHECK_EXPR(Name, expr) \
1452	do { \
1453	if (RT_UNLIKELY(!(expr))) \
1454	{ \
1455	OSDBGPRINT(( #Name ": %s\n", #expr)); \
1456	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1457	} \
1458	} while (0)
1459
1460	#define REQ_CHECK_EXPR_FMT(expr, fmt) \
1461	do { \
1462	if (RT_UNLIKELY(!(expr))) \
1463	{ \
1464	OSDBGPRINT( fmt ); \
1465	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1466	} \
1467	} while (0)
1468
1469	/*
1470	* The switch.
1471	*/
1472	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
1473	{
1474	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
1475	{
1476	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
1477	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
1478	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
1479	{
1480	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
1481	pReq->Hdr.rc = VERR_INVALID_MAGIC;
1482	return 0;
1483	}
1484
1485	#if 0
1486	/*
1487	* Call out to the OS specific code and let it do permission checks on the
1488	* client process.
1489	*/
1490	if (!supdrvOSValidateClientProcess(pDevExt, pSession))
1491	{
1492	pReq->u.Out.u32Cookie = 0xffffffff;
1493	pReq->u.Out.u32SessionCookie = 0xffffffff;
1494	pReq->u.Out.u32SessionVersion = 0xffffffff;
1495	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1496	pReq->u.Out.pSession = NULL;
1497	pReq->u.Out.cFunctions = 0;
1498	pReq->Hdr.rc = VERR_PERMISSION_DENIED;
1499	return 0;
1500	}
1501	#endif
1502
1503	/*
1504	* Match the version.
1505	* The current logic is very simple, match the major interface version.
1506	*/
1507	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
1508	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
1509	{
1510	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1511	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
1512	pReq->u.Out.u32Cookie = 0xffffffff;
1513	pReq->u.Out.u32SessionCookie = 0xffffffff;
1514	pReq->u.Out.u32SessionVersion = 0xffffffff;
1515	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1516	pReq->u.Out.pSession = NULL;
1517	pReq->u.Out.cFunctions = 0;
1518	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1519	return 0;
1520	}
1521
1522	/*
1523	* Fill in return data and be gone.
1524	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
1525	* u32SessionVersion <= u32ReqVersion!
1526	*/
1527	/** @todo Somehow validate the client and negotiate a secure cookie... */
1528	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
1529	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
1530	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
1531	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1532	pReq->u.Out.pSession = pSession;
1533	pReq->u.Out.cFunctions = sizeof(g_aFunctions) / sizeof(g_aFunctions[0]);
1534	pReq->Hdr.rc = VINF_SUCCESS;
1535	return 0;
1536	}
1537
1538	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_QUERY_FUNCS(0)):
1539	{
1540	/* validate */
1541	PSUPQUERYFUNCS pReq = (PSUPQUERYFUNCS)pReqHdr;
1542	REQ_CHECK_SIZES_EX(SUP_IOCTL_QUERY_FUNCS, SUP_IOCTL_QUERY_FUNCS_SIZE_IN, SUP_IOCTL_QUERY_FUNCS_SIZE_OUT(RT_ELEMENTS(g_aFunctions)));
1543
1544	/* execute */
1545	pReq->u.Out.cFunctions = RT_ELEMENTS(g_aFunctions);
1546	memcpy(&pReq->u.Out.aFunctions[0], g_aFunctions, sizeof(g_aFunctions));
1547	pReq->Hdr.rc = VINF_SUCCESS;
1548	return 0;
1549	}
1550
1551	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_LOCK):
1552	{
1553	/* validate */
1554	PSUPPAGELOCK pReq = (PSUPPAGELOCK)pReqHdr;
1555	REQ_CHECK_SIZE_IN(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_IN);
1556	REQ_CHECK_SIZE_OUT(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_OUT(pReq->u.In.cPages));
1557	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.cPages > 0);
1558	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.pvR3 >= PAGE_SIZE);
1559
1560	/* execute */
1561	pReq->Hdr.rc = SUPR0LockMem(pSession, pReq->u.In.pvR3, pReq->u.In.cPages, &pReq->u.Out.aPages[0]);
1562	if (RT_FAILURE(pReq->Hdr.rc))
1563	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1564	return 0;
1565	}
1566
1567	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_UNLOCK):
1568	{
1569	/* validate */
1570	PSUPPAGEUNLOCK pReq = (PSUPPAGEUNLOCK)pReqHdr;
1571	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_UNLOCK);
1572
1573	/* execute */
1574	pReq->Hdr.rc = SUPR0UnlockMem(pSession, pReq->u.In.pvR3);
1575	return 0;
1576	}
1577
1578	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_ALLOC):
1579	{
1580	/* validate */
1581	PSUPCONTALLOC pReq = (PSUPCONTALLOC)pReqHdr;
1582	REQ_CHECK_SIZES(SUP_IOCTL_CONT_ALLOC);
1583
1584	/* execute */
1585	pReq->Hdr.rc = SUPR0ContAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.HCPhys);
1586	if (RT_FAILURE(pReq->Hdr.rc))
1587	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1588	return 0;
1589	}
1590
1591	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_FREE):
1592	{
1593	/* validate */
1594	PSUPCONTFREE pReq = (PSUPCONTFREE)pReqHdr;
1595	REQ_CHECK_SIZES(SUP_IOCTL_CONT_FREE);
1596
1597	/* execute */
1598	pReq->Hdr.rc = SUPR0ContFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1599	return 0;
1600	}
1601
1602	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_OPEN):
1603	{
1604	/* validate */
1605	PSUPLDROPEN pReq = (PSUPLDROPEN)pReqHdr;
1606	REQ_CHECK_SIZES(SUP_IOCTL_LDR_OPEN);
1607	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs > 0);
1608	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs < 16*_1M);
1609	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
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 < pReq->u.In.cbImageWithTabs);
1612	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.szName[0]);
1613	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1614	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, !supdrvCheckInvalidChar(pReq->u.In.szName, ";:()[]{}/\\\|&*%#@!~`\"'"));
1615	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szFilename, sizeof(pReq->u.In.szFilename)));
1616
1617	/* execute */
1618	pReq->Hdr.rc = supdrvIOCtl_LdrOpen(pDevExt, pSession, pReq);
1619	return 0;
1620	}
1621
1622	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOAD):
1623	{
1624	/* validate */
1625	PSUPLDRLOAD pReq = (PSUPLDRLOAD)pReqHdr;
1626	REQ_CHECK_EXPR(Name, pReq->Hdr.cbIn >= sizeof(*pReq));
1627	REQ_CHECK_SIZES_EX(SUP_IOCTL_LDR_LOAD, SUP_IOCTL_LDR_LOAD_SIZE_IN(pReq->u.In.cbImageWithTabs), SUP_IOCTL_LDR_LOAD_SIZE_OUT);
1628	REQ_CHECK_EXPR(SUP_IOCTL_LDR_LOAD, pReq->u.In.cSymbols <= 16384);
1629	REQ_CHECK_EXPR_FMT( !pReq->u.In.cSymbols
1630	\|\| ( pReq->u.In.offSymbols < pReq->u.In.cbImageWithTabs
1631	&& pReq->u.In.offSymbols + pReq->u.In.cSymbols * sizeof(SUPLDRSYM) <= pReq->u.In.cbImageWithTabs),
1632	("SUP_IOCTL_LDR_LOAD: offSymbols=%#lx cSymbols=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offSymbols,
1633	(long)pReq->u.In.cSymbols, (long)pReq->u.In.cbImageWithTabs));
1634	REQ_CHECK_EXPR_FMT( !pReq->u.In.cbStrTab
1635	\|\| ( pReq->u.In.offStrTab < pReq->u.In.cbImageWithTabs
1636	&& pReq->u.In.offStrTab + pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs
1637	&& pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs),
1638	("SUP_IOCTL_LDR_LOAD: offStrTab=%#lx cbStrTab=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offStrTab,
1639	(long)pReq->u.In.cbStrTab, (long)pReq->u.In.cbImageWithTabs));
1640
1641	if (pReq->u.In.cSymbols)
1642	{
1643	uint32_t i;
1644	PSUPLDRSYM paSyms = (PSUPLDRSYM)&pReq->u.In.abImage[pReq->u.In.offSymbols];
1645	for (i = 0; i < pReq->u.In.cSymbols; i++)
1646	{
1647	REQ_CHECK_EXPR_FMT(paSyms[i].offSymbol < pReq->u.In.cbImageWithTabs,
1648	("SUP_IOCTL_LDR_LOAD: sym #%ld: symb off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offSymbol, (long)pReq->u.In.cbImageWithTabs));
1649	REQ_CHECK_EXPR_FMT(paSyms[i].offName < pReq->u.In.cbStrTab,
1650	("SUP_IOCTL_LDR_LOAD: sym #%ld: name off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1651	REQ_CHECK_EXPR_FMT(RTStrEnd((char const *)&pReq->u.In.abImage[pReq->u.In.offStrTab + paSyms[i].offName],
1652	pReq->u.In.cbStrTab - paSyms[i].offName),
1653	("SUP_IOCTL_LDR_LOAD: sym #%ld: unterminated name! (%#lx / %#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1654	}
1655	}
1656
1657	/* execute */
1658	pReq->Hdr.rc = supdrvIOCtl_LdrLoad(pDevExt, pSession, pReq);
1659	return 0;
1660	}
1661
1662	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_FREE):
1663	{
1664	/* validate */
1665	PSUPLDRFREE pReq = (PSUPLDRFREE)pReqHdr;
1666	REQ_CHECK_SIZES(SUP_IOCTL_LDR_FREE);
1667
1668	/* execute */
1669	pReq->Hdr.rc = supdrvIOCtl_LdrFree(pDevExt, pSession, pReq);
1670	return 0;
1671	}
1672
1673	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_GET_SYMBOL):
1674	{
1675	/* validate */
1676	PSUPLDRGETSYMBOL pReq = (PSUPLDRGETSYMBOL)pReqHdr;
1677	REQ_CHECK_SIZES(SUP_IOCTL_LDR_GET_SYMBOL);
1678	REQ_CHECK_EXPR(SUP_IOCTL_LDR_GET_SYMBOL, RTStrEnd(pReq->u.In.szSymbol, sizeof(pReq->u.In.szSymbol)));
1679
1680	/* execute */
1681	pReq->Hdr.rc = supdrvIOCtl_LdrGetSymbol(pDevExt, pSession, pReq);
1682	return 0;
1683	}
1684
1685	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0(0)):
1686	{
1687	/* validate */
1688	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1689	Log4(("SUP_IOCTL_CALL_VMMR0: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1690	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1691
1692	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_VMMR0_SIZE(0))
1693	{
1694	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(0), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(0));
1695
1696	/* execute */
1697	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1698	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg, pSession);
1699	else
1700	pReq->Hdr.rc = VERR_WRONG_ORDER;
1701	}
1702	else
1703	{
1704	PSUPVMMR0REQHDR pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1705	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR)),
1706	("SUP_IOCTL_CALL_VMMR0: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR))));
1707	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1708	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(pVMMReq->cbReq));
1709
1710	/* execute */
1711	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1712	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1713	else
1714	pReq->Hdr.rc = VERR_WRONG_ORDER;
1715	}
1716
1717	if ( RT_FAILURE(pReq->Hdr.rc)
1718	&& pReq->Hdr.rc != VERR_INTERRUPTED
1719	&& pReq->Hdr.rc != VERR_TIMEOUT)
1720	Log(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1721	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1722	else
1723	Log4(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1724	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1725	return 0;
1726	}
1727
1728	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0_BIG):
1729	{
1730	/* validate */
1731	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1732	PSUPVMMR0REQHDR pVMMReq;
1733	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1734	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1735
1736	pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1737	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR)),
1738	("SUP_IOCTL_CALL_VMMR0_BIG: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR))));
1739	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0_BIG, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1740	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));
1741
1742	/* execute */
1743	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1744	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1745	else
1746	pReq->Hdr.rc = VERR_WRONG_ORDER;
1747
1748	if ( RT_FAILURE(pReq->Hdr.rc)
1749	&& pReq->Hdr.rc != VERR_INTERRUPTED
1750	&& pReq->Hdr.rc != VERR_TIMEOUT)
1751	Log(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1752	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1753	else
1754	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1755	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1756	return 0;
1757	}
1758
1759	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_PAGING_MODE):
1760	{
1761	/* validate */
1762	PSUPGETPAGINGMODE pReq = (PSUPGETPAGINGMODE)pReqHdr;
1763	REQ_CHECK_SIZES(SUP_IOCTL_GET_PAGING_MODE);
1764
1765	/* execute */
1766	pReq->Hdr.rc = VINF_SUCCESS;
1767	pReq->u.Out.enmMode = SUPR0GetPagingMode();
1768	return 0;
1769	}
1770
1771	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_ALLOC):
1772	{
1773	/* validate */
1774	PSUPLOWALLOC pReq = (PSUPLOWALLOC)pReqHdr;
1775	REQ_CHECK_EXPR(SUP_IOCTL_LOW_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_LOW_ALLOC_SIZE_IN);
1776	REQ_CHECK_SIZES_EX(SUP_IOCTL_LOW_ALLOC, SUP_IOCTL_LOW_ALLOC_SIZE_IN, SUP_IOCTL_LOW_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1777
1778	/* execute */
1779	pReq->Hdr.rc = SUPR0LowAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1780	if (RT_FAILURE(pReq->Hdr.rc))
1781	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1782	return 0;
1783	}
1784
1785	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_FREE):
1786	{
1787	/* validate */
1788	PSUPLOWFREE pReq = (PSUPLOWFREE)pReqHdr;
1789	REQ_CHECK_SIZES(SUP_IOCTL_LOW_FREE);
1790
1791	/* execute */
1792	pReq->Hdr.rc = SUPR0LowFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1793	return 0;
1794	}
1795
1796	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_MAP):
1797	{
1798	/* validate */
1799	PSUPGIPMAP pReq = (PSUPGIPMAP)pReqHdr;
1800	REQ_CHECK_SIZES(SUP_IOCTL_GIP_MAP);
1801
1802	/* execute */
1803	pReq->Hdr.rc = SUPR0GipMap(pSession, &pReq->u.Out.pGipR3, &pReq->u.Out.HCPhysGip);
1804	if (RT_SUCCESS(pReq->Hdr.rc))
1805	pReq->u.Out.pGipR0 = pDevExt->pGip;
1806	return 0;
1807	}
1808
1809	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_UNMAP):
1810	{
1811	/* validate */
1812	PSUPGIPUNMAP pReq = (PSUPGIPUNMAP)pReqHdr;
1813	REQ_CHECK_SIZES(SUP_IOCTL_GIP_UNMAP);
1814
1815	/* execute */
1816	pReq->Hdr.rc = SUPR0GipUnmap(pSession);
1817	return 0;
1818	}
1819
1820	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SET_VM_FOR_FAST):
1821	{
1822	/* validate */
1823	PSUPSETVMFORFAST pReq = (PSUPSETVMFORFAST)pReqHdr;
1824	REQ_CHECK_SIZES(SUP_IOCTL_SET_VM_FOR_FAST);
1825	REQ_CHECK_EXPR_FMT( !pReq->u.In.pVMR0
1826	\|\| ( VALID_PTR(pReq->u.In.pVMR0)
1827	&& !((uintptr_t)pReq->u.In.pVMR0 & (PAGE_SIZE - 1))),
1828	("SUP_IOCTL_SET_VM_FOR_FAST: pVMR0=%p!\n", pReq->u.In.pVMR0));
1829	/* execute */
1830	pSession->pVM = pReq->u.In.pVMR0;
1831	pReq->Hdr.rc = VINF_SUCCESS;
1832	return 0;
1833	}
1834
1835	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_ALLOC_EX):
1836	{
1837	/* validate */
1838	PSUPPAGEALLOCEX pReq = (PSUPPAGEALLOCEX)pReqHdr;
1839	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC_EX, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_EX_SIZE_IN);
1840	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));
1841	REQ_CHECK_EXPR_FMT(pReq->u.In.fKernelMapping \|\| pReq->u.In.fUserMapping,
1842	("SUP_IOCTL_PAGE_ALLOC_EX: No mapping requested!\n"));
1843	REQ_CHECK_EXPR_FMT(pReq->u.In.fUserMapping,
1844	("SUP_IOCTL_PAGE_ALLOC_EX: Must have user mapping!\n"));
1845	REQ_CHECK_EXPR_FMT(!pReq->u.In.fReserved0 && !pReq->u.In.fReserved1,
1846	("SUP_IOCTL_PAGE_ALLOC_EX: fReserved0=%d fReserved1=%d\n", pReq->u.In.fReserved0, pReq->u.In.fReserved1));
1847
1848	/* execute */
1849	pReq->Hdr.rc = SUPR0PageAllocEx(pSession, pReq->u.In.cPages, 0 /* fFlags */,
1850	pReq->u.In.fUserMapping ? &pReq->u.Out.pvR3 : NULL,
1851	pReq->u.In.fKernelMapping ? &pReq->u.Out.pvR0 : NULL,
1852	&pReq->u.Out.aPages[0]);
1853	if (RT_FAILURE(pReq->Hdr.rc))
1854	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1855	return 0;
1856	}
1857
1858	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_MAP_KERNEL):
1859	{
1860	/* validate */
1861	PSUPPAGEMAPKERNEL pReq = (PSUPPAGEMAPKERNEL)pReqHdr;
1862	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_MAP_KERNEL);
1863	REQ_CHECK_EXPR_FMT(!pReq->u.In.fFlags, ("SUP_IOCTL_PAGE_MAP_KERNEL: fFlags=%#x! MBZ\n", pReq->u.In.fFlags));
1864	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_MAP_KERNEL: offSub=%#x\n", pReq->u.In.offSub));
1865	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1866	("SUP_IOCTL_PAGE_MAP_KERNEL: cbSub=%#x\n", pReq->u.In.cbSub));
1867
1868	/* execute */
1869	pReq->Hdr.rc = SUPR0PageMapKernel(pSession, pReq->u.In.pvR3, pReq->u.In.offSub, pReq->u.In.cbSub,
1870	pReq->u.In.fFlags, &pReq->u.Out.pvR0);
1871	if (RT_FAILURE(pReq->Hdr.rc))
1872	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1873	return 0;
1874	}
1875
1876	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_PROTECT):
1877	{
1878	/* validate */
1879	PSUPPAGEPROTECT pReq = (PSUPPAGEPROTECT)pReqHdr;
1880	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_PROTECT);
1881	REQ_CHECK_EXPR_FMT(!(pReq->u.In.fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)),
1882	("SUP_IOCTL_PAGE_PROTECT: fProt=%#x!\n", pReq->u.In.fProt));
1883	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_PROTECT: offSub=%#x\n", pReq->u.In.offSub));
1884	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1885	("SUP_IOCTL_PAGE_PROTECT: cbSub=%#x\n", pReq->u.In.cbSub));
1886
1887	/* execute */
1888	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);
1889	return 0;
1890	}
1891
1892	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_FREE):
1893	{
1894	/* validate */
1895	PSUPPAGEFREE pReq = (PSUPPAGEFREE)pReqHdr;
1896	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_FREE);
1897
1898	/* execute */
1899	pReq->Hdr.rc = SUPR0PageFree(pSession, pReq->u.In.pvR3);
1900	return 0;
1901	}
1902
1903	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_SERVICE(0)):
1904	{
1905	/* validate */
1906	PSUPCALLSERVICE pReq = (PSUPCALLSERVICE)pReqHdr;
1907	Log4(("SUP_IOCTL_CALL_SERVICE: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1908	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1909
1910	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
1911	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(0), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(0));
1912	else
1913	{
1914	PSUPR0SERVICEREQHDR pSrvReq = (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0];
1915	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR)),
1916	("SUP_IOCTL_CALL_SERVICE: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR))));
1917	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, pSrvReq->u32Magic == SUPR0SERVICEREQHDR_MAGIC);
1918	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(pSrvReq->cbReq), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(pSrvReq->cbReq));
1919	}
1920	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1921
1922	/* execute */
1923	pReq->Hdr.rc = supdrvIOCtl_CallServiceModule(pDevExt, pSession, pReq);
1924	return 0;
1925	}
1926
1927	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOGGER_SETTINGS(0)):
1928	{
1929	/* validate */
1930	PSUPLOGGERSETTINGS pReq = (PSUPLOGGERSETTINGS)pReqHdr;
1931	size_t cbStrTab;
1932	REQ_CHECK_SIZE_OUT(SUP_IOCTL_LOGGER_SETTINGS, SUP_IOCTL_LOGGER_SETTINGS_SIZE_OUT);
1933	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->Hdr.cbIn >= SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(1));
1934	cbStrTab = pReq->Hdr.cbIn - SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(0);
1935	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offGroups < cbStrTab);
1936	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offFlags < cbStrTab);
1937	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offDestination < cbStrTab);
1938	REQ_CHECK_EXPR_FMT(pReq->u.In.szStrings[cbStrTab - 1] == '\0',
1939	("SUP_IOCTL_LOGGER_SETTINGS: cbIn=%#x cbStrTab=%#zx LastChar=%d\n",
1940	pReq->Hdr.cbIn, cbStrTab, pReq->u.In.szStrings[cbStrTab - 1]));
1941	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhich <= SUPLOGGERSETTINGS_WHICH_RELEASE);
1942	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhat <= SUPLOGGERSETTINGS_WHAT_DESTROY);
1943
1944	/* execute */
1945	pReq->Hdr.rc = supdrvIOCtl_LoggerSettings(pDevExt, pSession, pReq);
1946	return 0;
1947	}
1948
1949	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP2):
1950	{
1951	/* validate */
1952	PSUPSEMOP2 pReq = (PSUPSEMOP2)pReqHdr;
1953	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP2, SUP_IOCTL_SEM_OP2_SIZE_IN, SUP_IOCTL_SEM_OP2_SIZE_OUT);
1954	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP2, pReq->u.In.uReserved == 0);
1955
1956	/* execute */
1957	switch (pReq->u.In.uType)
1958	{
1959	case SUP_SEM_TYPE_EVENT:
1960	{
1961	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1962	switch (pReq->u.In.uOp)
1963	{
1964	case SUPSEMOP2_WAIT_MS_REL:
1965	pReq->Hdr.rc = SUPSemEventWaitNoResume(pSession, hEvent, pReq->u.In.uArg.cRelMsTimeout);
1966	break;
1967	case SUPSEMOP2_WAIT_NS_ABS:
1968	pReq->Hdr.rc = SUPSemEventWaitNsAbsIntr(pSession, hEvent, pReq->u.In.uArg.uAbsNsTimeout);
1969	break;
1970	case SUPSEMOP2_WAIT_NS_REL:
1971	pReq->Hdr.rc = SUPSemEventWaitNsRelIntr(pSession, hEvent, pReq->u.In.uArg.cRelNsTimeout);
1972	break;
1973	case SUPSEMOP2_SIGNAL:
1974	pReq->Hdr.rc = SUPSemEventSignal(pSession, hEvent);
1975	break;
1976	case SUPSEMOP2_CLOSE:
1977	pReq->Hdr.rc = SUPSemEventClose(pSession, hEvent);
1978	break;
1979	case SUPSEMOP2_RESET:
1980	default:
1981	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1982	break;
1983	}
1984	break;
1985	}
1986
1987	case SUP_SEM_TYPE_EVENT_MULTI:
1988	{
1989	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1990	switch (pReq->u.In.uOp)
1991	{
1992	case SUPSEMOP2_WAIT_MS_REL:
1993	pReq->Hdr.rc = SUPSemEventMultiWaitNoResume(pSession, hEventMulti, pReq->u.In.uArg.cRelMsTimeout);
1994	break;
1995	case SUPSEMOP2_WAIT_NS_ABS:
1996	pReq->Hdr.rc = SUPSemEventMultiWaitNsAbsIntr(pSession, hEventMulti, pReq->u.In.uArg.uAbsNsTimeout);
1997	break;
1998	case SUPSEMOP2_WAIT_NS_REL:
1999	pReq->Hdr.rc = SUPSemEventMultiWaitNsRelIntr(pSession, hEventMulti, pReq->u.In.uArg.cRelNsTimeout);
2000	break;
2001	case SUPSEMOP2_SIGNAL:
2002	pReq->Hdr.rc = SUPSemEventMultiSignal(pSession, hEventMulti);
2003	break;
2004	case SUPSEMOP2_CLOSE:
2005	pReq->Hdr.rc = SUPSemEventMultiClose(pSession, hEventMulti);
2006	break;
2007	case SUPSEMOP2_RESET:
2008	pReq->Hdr.rc = SUPSemEventMultiReset(pSession, hEventMulti);
2009	break;
2010	default:
2011	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
2012	break;
2013	}
2014	break;
2015	}
2016
2017	default:
2018	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
2019	break;
2020	}
2021	return 0;
2022	}
2023
2024	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP3):
2025	{
2026	/* validate */
2027	PSUPSEMOP3 pReq = (PSUPSEMOP3)pReqHdr;
2028	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP3, SUP_IOCTL_SEM_OP3_SIZE_IN, SUP_IOCTL_SEM_OP3_SIZE_OUT);
2029	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, pReq->u.In.u32Reserved == 0 && pReq->u.In.u64Reserved == 0);
2030
2031	/* execute */
2032	switch (pReq->u.In.uType)
2033	{
2034	case SUP_SEM_TYPE_EVENT:
2035	{
2036	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
2037	switch (pReq->u.In.uOp)
2038	{
2039	case SUPSEMOP3_CREATE:
2040	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
2041	pReq->Hdr.rc = SUPSemEventCreate(pSession, &hEvent);
2042	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEvent;
2043	break;
2044	case SUPSEMOP3_GET_RESOLUTION:
2045	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
2046	pReq->Hdr.rc = VINF_SUCCESS;
2047	pReq->Hdr.cbOut = sizeof(*pReq);
2048	pReq->u.Out.cNsResolution = SUPSemEventGetResolution(pSession);
2049	break;
2050	default:
2051	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
2052	break;
2053	}
2054	break;
2055	}
2056
2057	case SUP_SEM_TYPE_EVENT_MULTI:
2058	{
2059	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
2060	switch (pReq->u.In.uOp)
2061	{
2062	case SUPSEMOP3_CREATE:
2063	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
2064	pReq->Hdr.rc = SUPSemEventMultiCreate(pSession, &hEventMulti);
2065	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEventMulti;
2066	break;
2067	case SUPSEMOP3_GET_RESOLUTION:
2068	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
2069	pReq->Hdr.rc = VINF_SUCCESS;
2070	pReq->u.Out.cNsResolution = SUPSemEventMultiGetResolution(pSession);
2071	break;
2072	default:
2073	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
2074	break;
2075	}
2076	break;
2077	}
2078
2079	default:
2080	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
2081	break;
2082	}
2083	return 0;
2084	}
2085
2086	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
2087	{
2088	/* validate */
2089	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
2090	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
2091
2092	/* execute */
2093	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.Caps);
2094	if (RT_FAILURE(pReq->Hdr.rc))
2095	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
2096	return 0;
2097	}
2098
2099	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_OPEN):
2100	{
2101	/* validate */
2102	PSUPTRACEROPEN pReq = (PSUPTRACEROPEN)pReqHdr;
2103	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_OPEN);
2104
2105	/* execute */
2106	pReq->Hdr.rc = supdrvIOCtl_TracerOpen(pDevExt, pSession, pReq->u.In.uCookie, pReq->u.In.uArg);
2107	return 0;
2108	}
2109
2110	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_CLOSE):
2111	{
2112	/* validate */
2113	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_CLOSE);
2114
2115	/* execute */
2116	pReqHdr->rc = supdrvIOCtl_TracerClose(pDevExt, pSession);
2117	return 0;
2118	}
2119
2120	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_IOCTL):
2121	{
2122	/* validate */
2123	PSUPTRACERIOCTL pReq = (PSUPTRACERIOCTL)pReqHdr;
2124	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_IOCTL);
2125
2126	/* execute */
2127	pReqHdr->rc = supdrvIOCtl_TracerIOCtl(pDevExt, pSession, pReq->u.In.uCmd, pReq->u.In.uArg, &pReq->u.Out.iRetVal);
2128	return 0;
2129	}
2130
2131	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_REG):
2132	{
2133	/* validate */
2134	PSUPTRACERUMODREG pReq = (PSUPTRACERUMODREG)pReqHdr;
2135	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_REG);
2136	if (!RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)))
2137	return VERR_INVALID_PARAMETER;
2138
2139	/* execute */
2140	pReqHdr->rc = supdrvIOCtl_TracerUmodRegister(pDevExt, pSession,
2141	pReq->u.In.R3PtrVtgHdr, pReq->u.In.uVtgHdrAddr,
2142	pReq->u.In.R3PtrStrTab, pReq->u.In.cbStrTab,
2143	pReq->u.In.szName, pReq->u.In.fFlags);
2144	return 0;
2145	}
2146
2147	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_DEREG):
2148	{
2149	/* validate */
2150	PSUPTRACERUMODDEREG pReq = (PSUPTRACERUMODDEREG)pReqHdr;
2151	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_DEREG);
2152
2153	/* execute */
2154	pReqHdr->rc = supdrvIOCtl_TracerUmodDeregister(pDevExt, pSession, pReq->u.In.pVtgHdr);
2155	return 0;
2156	}
2157
2158	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_FIRE_PROBE):
2159	{
2160	/* validate */
2161	PSUPTRACERUMODFIREPROBE pReq = (PSUPTRACERUMODFIREPROBE)pReqHdr;
2162	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_FIRE_PROBE);
2163
2164	supdrvIOCtl_TracerUmodProbeFire(pDevExt, pSession, &pReq->u.In);
2165	pReqHdr->rc = VINF_SUCCESS;
2166	return 0;
2167	}
2168
2169	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_MSR_PROBER):
2170	{
2171	/* validate */
2172	PSUPMSRPROBER pReq = (PSUPMSRPROBER)pReqHdr;
2173	REQ_CHECK_SIZES(SUP_IOCTL_MSR_PROBER);
2174	REQ_CHECK_EXPR(SUP_IOCTL_MSR_PROBER,
2175	pReq->u.In.enmOp > SUPMSRPROBEROP_INVALID && pReq->u.In.enmOp < SUPMSRPROBEROP_END);
2176
2177	pReqHdr->rc = supdrvIOCtl_MsrProber(pDevExt, pReq);
2178	return 0;
2179	}
2180
2181	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_RESUME_SUSPENDED_KBDS):
2182	{
2183	/* validate */
2184	REQ_CHECK_SIZES(SUP_IOCTL_RESUME_SUSPENDED_KBDS);
2185
2186	pReqHdr->rc = supdrvIOCtl_ResumeSuspendedKbds();
2187	return 0;
2188	}
2189
2190	default:
2191	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
2192	break;
2193	}
2194	return VERR_GENERAL_FAILURE;
2195	}
2196
2197
2198	/**
2199	* I/O Control inner worker for the restricted operations.
2200	*
2201	* @returns IPRT status code.
2202	* @retval VERR_INVALID_PARAMETER if the request is invalid.
2203	*
2204	* @param uIOCtl Function number.
2205	* @param pDevExt Device extention.
2206	* @param pSession Session data.
2207	* @param pReqHdr The request header.
2208	*/
2209	static int supdrvIOCtlInnerRestricted(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
2210	{
2211	/*
2212	* The switch.
2213	*/
2214	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
2215	{
2216	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
2217	{
2218	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
2219	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
2220	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
2221	{
2222	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
2223	pReq->Hdr.rc = VERR_INVALID_MAGIC;
2224	return 0;
2225	}
2226
2227	/*
2228	* Match the version.
2229	* The current logic is very simple, match the major interface version.
2230	*/
2231	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
2232	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
2233	{
2234	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
2235	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
2236	pReq->u.Out.u32Cookie = 0xffffffff;
2237	pReq->u.Out.u32SessionCookie = 0xffffffff;
2238	pReq->u.Out.u32SessionVersion = 0xffffffff;
2239	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
2240	pReq->u.Out.pSession = NULL;
2241	pReq->u.Out.cFunctions = 0;
2242	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
2243	return 0;
2244	}
2245
2246	/*
2247	* Fill in return data and be gone.
2248	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
2249	* u32SessionVersion <= u32ReqVersion!
2250	*/
2251	/** @todo Somehow validate the client and negotiate a secure cookie... */
2252	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
2253	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
2254	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
2255	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
2256	pReq->u.Out.pSession = pSession;
2257	pReq->u.Out.cFunctions = 0;
2258	pReq->Hdr.rc = VINF_SUCCESS;
2259	return 0;
2260	}
2261
2262	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
2263	{
2264	/* validate */
2265	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
2266	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
2267
2268	/* execute */
2269	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.Caps);
2270	if (RT_FAILURE(pReq->Hdr.rc))
2271	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
2272	return 0;
2273	}
2274
2275	default:
2276	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
2277	break;
2278	}
2279	return VERR_GENERAL_FAILURE;
2280	}
2281
2282
2283	/**
2284	* I/O Control worker.
2285	*
2286	* @returns IPRT status code.
2287	* @retval VERR_INVALID_PARAMETER if the request is invalid.
2288	*
2289	* @param uIOCtl Function number.
2290	* @param pDevExt Device extention.
2291	* @param pSession Session data.
2292	* @param pReqHdr The request header.
2293	*/
2294	int VBOXCALL supdrvIOCtl(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
2295	{
2296	int rc;
2297	VBOXDRV_IOCTL_ENTRY(pSession, uIOCtl, pReqHdr);
2298
2299	/*
2300	* Validate the request.
2301	*/
2302	/* this first check could probably be omitted as its also done by the OS specific code... */
2303	if (RT_UNLIKELY( (pReqHdr->fFlags & SUPREQHDR_FLAGS_MAGIC_MASK) != SUPREQHDR_FLAGS_MAGIC
2304	\|\| pReqHdr->cbIn < sizeof(*pReqHdr)
2305	\|\| pReqHdr->cbOut < sizeof(*pReqHdr)))
2306	{
2307	OSDBGPRINT(("vboxdrv: Bad ioctl request header; cbIn=%#lx cbOut=%#lx fFlags=%#lx\n",
2308	(long)pReqHdr->cbIn, (long)pReqHdr->cbOut, (long)pReqHdr->fFlags));
2309	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
2310	return VERR_INVALID_PARAMETER;
2311	}
2312	if (RT_UNLIKELY(!RT_VALID_PTR(pSession)))
2313	{
2314	OSDBGPRINT(("vboxdrv: Invalid pSession value %p (ioctl=%p)\n", pSession, (void *)uIOCtl));
2315	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
2316	return VERR_INVALID_PARAMETER;
2317	}
2318	if (RT_UNLIKELY(uIOCtl == SUP_IOCTL_COOKIE))
2319	{
2320	if (pReqHdr->u32Cookie != SUPCOOKIE_INITIAL_COOKIE)
2321	{
2322	OSDBGPRINT(("SUP_IOCTL_COOKIE: bad cookie %#lx\n", (long)pReqHdr->u32Cookie));
2323	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
2324	return VERR_INVALID_PARAMETER;
2325	}
2326	}
2327	else if (RT_UNLIKELY( pReqHdr->u32Cookie != pDevExt->u32Cookie
2328	\|\| pReqHdr->u32SessionCookie != pSession->u32Cookie))
2329	{
2330	OSDBGPRINT(("vboxdrv: bad cookie %#lx / %#lx.\n", (long)pReqHdr->u32Cookie, (long)pReqHdr->u32SessionCookie));
2331	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
2332	return VERR_INVALID_PARAMETER;
2333	}
2334
2335	/*
2336	* Hand it to an inner function to avoid lots of unnecessary return tracepoints.
2337	*/
2338	if (pSession->fUnrestricted)
2339	rc = supdrvIOCtlInnerUnrestricted(uIOCtl, pDevExt, pSession, pReqHdr);
2340	else
2341	rc = supdrvIOCtlInnerRestricted(uIOCtl, pDevExt, pSession, pReqHdr);
2342
2343	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, pReqHdr->rc, rc);
2344	return rc;
2345	}
2346
2347
2348	/**
2349	* Inter-Driver Communication (IDC) worker.
2350	*
2351	* @returns VBox status code.
2352	* @retval VINF_SUCCESS on success.
2353	* @retval VERR_INVALID_PARAMETER if the request is invalid.
2354	* @retval VERR_NOT_SUPPORTED if the request isn't supported.
2355	*
2356	* @param uReq The request (function) code.
2357	* @param pDevExt Device extention.
2358	* @param pSession Session data.
2359	* @param pReqHdr The request header.
2360	*/
2361	int VBOXCALL supdrvIDC(uintptr_t uReq, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQHDR pReqHdr)
2362	{
2363	/*
2364	* The OS specific code has already validated the pSession
2365	* pointer, and the request size being greater or equal to
2366	* size of the header.
2367	*
2368	* So, just check that pSession is a kernel context session.
2369	*/
2370	if (RT_UNLIKELY( pSession
2371	&& pSession->R0Process != NIL_RTR0PROCESS))
2372	return VERR_INVALID_PARAMETER;
2373
2374	/*
2375	* Validation macro.
2376	*/
2377	#define REQ_CHECK_IDC_SIZE(Name, cbExpect) \
2378	do { \
2379	if (RT_UNLIKELY(pReqHdr->cb != (cbExpect))) \
2380	{ \
2381	OSDBGPRINT(( #Name ": Invalid input/output sizes. cb=%ld expected %ld.\n", \
2382	(long)pReqHdr->cb, (long)(cbExpect))); \
2383	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
2384	} \
2385	} while (0)
2386
2387	switch (uReq)
2388	{
2389	case SUPDRV_IDC_REQ_CONNECT:
2390	{
2391	PSUPDRVIDCREQCONNECT pReq = (PSUPDRVIDCREQCONNECT)pReqHdr;
2392	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_CONNECT, sizeof(*pReq));
2393
2394	/*
2395	* Validate the cookie and other input.
2396	*/
2397	if (pReq->Hdr.pSession != NULL)
2398	{
2399	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Hdr.pSession=%p expected NULL!\n", pReq->Hdr.pSession));
2400	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2401	}
2402	if (pReq->u.In.u32MagicCookie != SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE)
2403	{
2404	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: u32MagicCookie=%#x expected %#x!\n",
2405	(unsigned)pReq->u.In.u32MagicCookie, (unsigned)SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE));
2406	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2407	}
2408	if ( pReq->u.In.uMinVersion > pReq->u.In.uReqVersion
2409	\|\| (pReq->u.In.uMinVersion & UINT32_C(0xffff0000)) != (pReq->u.In.uReqVersion & UINT32_C(0xffff0000)))
2410	{
2411	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: uMinVersion=%#x uMaxVersion=%#x doesn't match!\n",
2412	pReq->u.In.uMinVersion, pReq->u.In.uReqVersion));
2413	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2414	}
2415	if (pSession != NULL)
2416	{
2417	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: pSession=%p expected NULL!\n", pSession));
2418	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2419	}
2420
2421	/*
2422	* Match the version.
2423	* The current logic is very simple, match the major interface version.
2424	*/
2425	if ( pReq->u.In.uMinVersion > SUPDRV_IDC_VERSION
2426	\|\| (pReq->u.In.uMinVersion & 0xffff0000) != (SUPDRV_IDC_VERSION & 0xffff0000))
2427	{
2428	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
2429	pReq->u.In.uReqVersion, pReq->u.In.uMinVersion, (unsigned)SUPDRV_IDC_VERSION));
2430	pReq->u.Out.pSession = NULL;
2431	pReq->u.Out.uSessionVersion = 0xffffffff;
2432	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
2433	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
2434	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
2435	return VINF_SUCCESS;
2436	}
2437
2438	pReq->u.Out.pSession = NULL;
2439	pReq->u.Out.uSessionVersion = SUPDRV_IDC_VERSION;
2440	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
2441	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
2442
2443	pReq->Hdr.rc = supdrvCreateSession(pDevExt, false /* fUser /, true /fUnrestricted*/, &pSession);
2444	if (RT_FAILURE(pReq->Hdr.rc))
2445	{
2446	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: failed to create session, rc=%d\n", pReq->Hdr.rc));
2447	return VINF_SUCCESS;
2448	}
2449
2450	pReq->u.Out.pSession = pSession;
2451	pReq->Hdr.pSession = pSession;
2452
2453	return VINF_SUCCESS;
2454	}
2455
2456	case SUPDRV_IDC_REQ_DISCONNECT:
2457	{
2458	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_DISCONNECT, sizeof(*pReqHdr));
2459
2460	supdrvSessionRelease(pSession);
2461	return pReqHdr->rc = VINF_SUCCESS;
2462	}
2463
2464	case SUPDRV_IDC_REQ_GET_SYMBOL:
2465	{
2466	PSUPDRVIDCREQGETSYM pReq = (PSUPDRVIDCREQGETSYM)pReqHdr;
2467	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_GET_SYMBOL, sizeof(*pReq));
2468
2469	pReq->Hdr.rc = supdrvIDC_LdrGetSymbol(pDevExt, pSession, pReq);
2470	return VINF_SUCCESS;
2471	}
2472
2473	case SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY:
2474	{
2475	PSUPDRVIDCREQCOMPREGFACTORY pReq = (PSUPDRVIDCREQCOMPREGFACTORY)pReqHdr;
2476	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY, sizeof(*pReq));
2477
2478	pReq->Hdr.rc = SUPR0ComponentRegisterFactory(pSession, pReq->u.In.pFactory);
2479	return VINF_SUCCESS;
2480	}
2481
2482	case SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY:
2483	{
2484	PSUPDRVIDCREQCOMPDEREGFACTORY pReq = (PSUPDRVIDCREQCOMPDEREGFACTORY)pReqHdr;
2485	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY, sizeof(*pReq));
2486
2487	pReq->Hdr.rc = SUPR0ComponentDeregisterFactory(pSession, pReq->u.In.pFactory);
2488	return VINF_SUCCESS;
2489	}
2490
2491	default:
2492	Log(("Unknown IDC %#lx\n", (long)uReq));
2493	break;
2494	}
2495
2496	#undef REQ_CHECK_IDC_SIZE
2497	return VERR_NOT_SUPPORTED;
2498	}
2499
2500
2501	/**
2502	* Register a object for reference counting.
2503	* The object is registered with one reference in the specified session.
2504	*
2505	* @returns Unique identifier on success (pointer).
2506	* All future reference must use this identifier.
2507	* @returns NULL on failure.
2508	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
2509	* @param pvUser1 The first user argument.
2510	* @param pvUser2 The second user argument.
2511	*/
2512	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
2513	{
2514	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2515	PSUPDRVOBJ pObj;
2516	PSUPDRVUSAGE pUsage;
2517
2518	/*
2519	* Validate the input.
2520	*/
2521	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
2522	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
2523	AssertPtrReturn(pfnDestructor, NULL);
2524
2525	/*
2526	* Allocate and initialize the object.
2527	*/
2528	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
2529	if (!pObj)
2530	return NULL;
2531	pObj->u32Magic = SUPDRVOBJ_MAGIC;
2532	pObj->enmType = enmType;
2533	pObj->pNext = NULL;
2534	pObj->cUsage = 1;
2535	pObj->pfnDestructor = pfnDestructor;
2536	pObj->pvUser1 = pvUser1;
2537	pObj->pvUser2 = pvUser2;
2538	pObj->CreatorUid = pSession->Uid;
2539	pObj->CreatorGid = pSession->Gid;
2540	pObj->CreatorProcess= pSession->Process;
2541	supdrvOSObjInitCreator(pObj, pSession);
2542
2543	/*
2544	* Allocate the usage record.
2545	* (We keep freed usage records around to simplify SUPR0ObjAddRefEx().)
2546	*/
2547	RTSpinlockAcquire(pDevExt->Spinlock);
2548
2549	pUsage = pDevExt->pUsageFree;
2550	if (pUsage)
2551	pDevExt->pUsageFree = pUsage->pNext;
2552	else
2553	{
2554	RTSpinlockRelease(pDevExt->Spinlock);
2555	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
2556	if (!pUsage)
2557	{
2558	RTMemFree(pObj);
2559	return NULL;
2560	}
2561	RTSpinlockAcquire(pDevExt->Spinlock);
2562	}
2563
2564	/*
2565	* Insert the object and create the session usage record.
2566	*/
2567	/* The object. */
2568	pObj->pNext = pDevExt->pObjs;
2569	pDevExt->pObjs = pObj;
2570
2571	/* The session record. */
2572	pUsage->cUsage = 1;
2573	pUsage->pObj = pObj;
2574	pUsage->pNext = pSession->pUsage;
2575	/* Log2(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext)); */
2576	pSession->pUsage = pUsage;
2577
2578	RTSpinlockRelease(pDevExt->Spinlock);
2579
2580	Log(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
2581	return pObj;
2582	}
2583
2584
2585	/**
2586	* Increment the reference counter for the object associating the reference
2587	* with the specified session.
2588	*
2589	* @returns IPRT status code.
2590	* @param pvObj The identifier returned by SUPR0ObjRegister().
2591	* @param pSession The session which is referencing the object.
2592	*
2593	* @remarks The caller should not own any spinlocks and must carefully protect
2594	* itself against potential race with the destructor so freed memory
2595	* isn't accessed here.
2596	*/
2597	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
2598	{
2599	return SUPR0ObjAddRefEx(pvObj, pSession, false /* fNoBlocking */);
2600	}
2601
2602
2603	/**
2604	* Increment the reference counter for the object associating the reference
2605	* with the specified session.
2606	*
2607	* @returns IPRT status code.
2608	* @retval VERR_TRY_AGAIN if fNoBlocking was set and a new usage record
2609	* couldn't be allocated. (If you see this you're not doing the right
2610	* thing and it won't ever work reliably.)
2611	*
2612	* @param pvObj The identifier returned by SUPR0ObjRegister().
2613	* @param pSession The session which is referencing the object.
2614	* @param fNoBlocking Set if it's not OK to block. Never try to make the
2615	* first reference to an object in a session with this
2616	* argument set.
2617	*
2618	* @remarks The caller should not own any spinlocks and must carefully protect
2619	* itself against potential race with the destructor so freed memory
2620	* isn't accessed here.
2621	*/
2622	SUPR0DECL(int) SUPR0ObjAddRefEx(void *pvObj, PSUPDRVSESSION pSession, bool fNoBlocking)
2623	{
2624	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2625	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2626	int rc = VINF_SUCCESS;
2627	PSUPDRVUSAGE pUsagePre;
2628	PSUPDRVUSAGE pUsage;
2629
2630	/*
2631	* Validate the input.
2632	* Be ready for the destruction race (someone might be stuck in the
2633	* destructor waiting a lock we own).
2634	*/
2635	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2636	AssertPtrReturn(pObj, VERR_INVALID_POINTER);
2637	AssertMsgReturn(pObj->u32Magic == SUPDRVOBJ_MAGIC \|\| pObj->u32Magic == SUPDRVOBJ_MAGIC_DEAD,
2638	("Invalid pvObj=%p magic=%#x (expected %#x or %#x)\n", pvObj, pObj->u32Magic, SUPDRVOBJ_MAGIC, SUPDRVOBJ_MAGIC_DEAD),
2639	VERR_INVALID_PARAMETER);
2640
2641	RTSpinlockAcquire(pDevExt->Spinlock);
2642
2643	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2644	{
2645	RTSpinlockRelease(pDevExt->Spinlock);
2646
2647	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2648	return VERR_WRONG_ORDER;
2649	}
2650
2651	/*
2652	* Preallocate the usage record if we can.
2653	*/
2654	pUsagePre = pDevExt->pUsageFree;
2655	if (pUsagePre)
2656	pDevExt->pUsageFree = pUsagePre->pNext;
2657	else if (!fNoBlocking)
2658	{
2659	RTSpinlockRelease(pDevExt->Spinlock);
2660	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
2661	if (!pUsagePre)
2662	return VERR_NO_MEMORY;
2663
2664	RTSpinlockAcquire(pDevExt->Spinlock);
2665	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2666	{
2667	RTSpinlockRelease(pDevExt->Spinlock);
2668
2669	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2670	return VERR_WRONG_ORDER;
2671	}
2672	}
2673
2674	/*
2675	* Reference the object.
2676	*/
2677	pObj->cUsage++;
2678
2679	/*
2680	* Look for the session record.
2681	*/
2682	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
2683	{
2684	/Log(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2685	if (pUsage->pObj == pObj)
2686	break;
2687	}
2688	if (pUsage)
2689	pUsage->cUsage++;
2690	else if (pUsagePre)
2691	{
2692	/* create a new session record. */
2693	pUsagePre->cUsage = 1;
2694	pUsagePre->pObj = pObj;
2695	pUsagePre->pNext = pSession->pUsage;
2696	pSession->pUsage = pUsagePre;
2697	/Log(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));/
2698
2699	pUsagePre = NULL;
2700	}
2701	else
2702	{
2703	pObj->cUsage--;
2704	rc = VERR_TRY_AGAIN;
2705	}
2706
2707	/*
2708	* Put any unused usage record into the free list..
2709	*/
2710	if (pUsagePre)
2711	{
2712	pUsagePre->pNext = pDevExt->pUsageFree;
2713	pDevExt->pUsageFree = pUsagePre;
2714	}
2715
2716	RTSpinlockRelease(pDevExt->Spinlock);
2717
2718	return rc;
2719	}
2720
2721
2722	/**
2723	* Decrement / destroy a reference counter record for an object.
2724	*
2725	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
2726	*
2727	* @returns IPRT status code.
2728	* @retval VINF_SUCCESS if not destroyed.
2729	* @retval VINF_OBJECT_DESTROYED if it's destroyed by this release call.
2730	* @retval VERR_INVALID_PARAMETER if the object isn't valid. Will assert in
2731	* string builds.
2732	*
2733	* @param pvObj The identifier returned by SUPR0ObjRegister().
2734	* @param pSession The session which is referencing the object.
2735	*/
2736	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
2737	{
2738	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2739	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2740	int rc = VERR_INVALID_PARAMETER;
2741	PSUPDRVUSAGE pUsage;
2742	PSUPDRVUSAGE pUsagePrev;
2743
2744	/*
2745	* Validate the input.
2746	*/
2747	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2748	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2749	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2750	VERR_INVALID_PARAMETER);
2751
2752	/*
2753	* Acquire the spinlock and look for the usage record.
2754	*/
2755	RTSpinlockAcquire(pDevExt->Spinlock);
2756
2757	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
2758	pUsage;
2759	pUsagePrev = pUsage, pUsage = pUsage->pNext)
2760	{
2761	/Log2(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2762	if (pUsage->pObj == pObj)
2763	{
2764	rc = VINF_SUCCESS;
2765	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
2766	if (pUsage->cUsage > 1)
2767	{
2768	pObj->cUsage--;
2769	pUsage->cUsage--;
2770	}
2771	else
2772	{
2773	/*
2774	* Free the session record.
2775	*/
2776	if (pUsagePrev)
2777	pUsagePrev->pNext = pUsage->pNext;
2778	else
2779	pSession->pUsage = pUsage->pNext;
2780	pUsage->pNext = pDevExt->pUsageFree;
2781	pDevExt->pUsageFree = pUsage;
2782
2783	/* What about the object? */
2784	if (pObj->cUsage > 1)
2785	pObj->cUsage--;
2786	else
2787	{
2788	/*
2789	* Object is to be destroyed, unlink it.
2790	*/
2791	pObj->u32Magic = SUPDRVOBJ_MAGIC_DEAD;
2792	rc = VINF_OBJECT_DESTROYED;
2793	if (pDevExt->pObjs == pObj)
2794	pDevExt->pObjs = pObj->pNext;
2795	else
2796	{
2797	PSUPDRVOBJ pObjPrev;
2798	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
2799	if (pObjPrev->pNext == pObj)
2800	{
2801	pObjPrev->pNext = pObj->pNext;
2802	break;
2803	}
2804	Assert(pObjPrev);
2805	}
2806	}
2807	}
2808	break;
2809	}
2810	}
2811
2812	RTSpinlockRelease(pDevExt->Spinlock);
2813
2814	/*
2815	* Call the destructor and free the object if required.
2816	*/
2817	if (rc == VINF_OBJECT_DESTROYED)
2818	{
2819	Log(("SUPR0ObjRelease: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
2820	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
2821	if (pObj->pfnDestructor)
2822	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
2823	RTMemFree(pObj);
2824	}
2825
2826	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
2827	return rc;
2828	}
2829
2830
2831	/**
2832	* Verifies that the current process can access the specified object.
2833	*
2834	* @returns The following IPRT status code:
2835	* @retval VINF_SUCCESS if access was granted.
2836	* @retval VERR_PERMISSION_DENIED if denied access.
2837	* @retval VERR_INVALID_PARAMETER if invalid parameter.
2838	*
2839	* @param pvObj The identifier returned by SUPR0ObjRegister().
2840	* @param pSession The session which wishes to access the object.
2841	* @param pszObjName Object string name. This is optional and depends on the object type.
2842	*
2843	* @remark The caller is responsible for making sure the object isn't removed while
2844	* we're inside this function. If uncertain about this, just call AddRef before calling us.
2845	*/
2846	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
2847	{
2848	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2849	int rc;
2850
2851	/*
2852	* Validate the input.
2853	*/
2854	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2855	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2856	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2857	VERR_INVALID_PARAMETER);
2858
2859	/*
2860	* Check access. (returns true if a decision has been made.)
2861	*/
2862	rc = VERR_INTERNAL_ERROR;
2863	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
2864	return rc;
2865
2866	/*
2867	* Default policy is to allow the user to access his own
2868	* stuff but nothing else.
2869	*/
2870	if (pObj->CreatorUid == pSession->Uid)
2871	return VINF_SUCCESS;
2872	return VERR_PERMISSION_DENIED;
2873	}
2874
2875
2876	/**
2877	* Lock pages.
2878	*
2879	* @returns IPRT status code.
2880	* @param pSession Session to which the locked memory should be associated.
2881	* @param pvR3 Start of the memory range to lock.
2882	* This must be page aligned.
2883	* @param cPages Number of pages to lock.
2884	* @param paPages Where to put the physical addresses of locked memory.
2885	*/
2886	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
2887	{
2888	int rc;
2889	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2890	const size_t cb = (size_t)cPages << PAGE_SHIFT;
2891	LogFlow(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
2892
2893	/*
2894	* Verify input.
2895	*/
2896	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2897	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
2898	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
2899	\|\| !pvR3)
2900	{
2901	Log(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
2902	return VERR_INVALID_PARAMETER;
2903	}
2904
2905	/*
2906	* Let IPRT do the job.
2907	*/
2908	Mem.eType = MEMREF_TYPE_LOCKED;
2909	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTMEM_PROT_READ \| RTMEM_PROT_WRITE, RTR0ProcHandleSelf());
2910	if (RT_SUCCESS(rc))
2911	{
2912	uint32_t iPage = cPages;
2913	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
2914	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
2915
2916	while (iPage-- > 0)
2917	{
2918	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2919	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
2920	{
2921	AssertMsgFailed(("iPage=%d\n", iPage));
2922	rc = VERR_INTERNAL_ERROR;
2923	break;
2924	}
2925	}
2926	if (RT_SUCCESS(rc))
2927	rc = supdrvMemAdd(&Mem, pSession);
2928	if (RT_FAILURE(rc))
2929	{
2930	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
2931	AssertRC(rc2);
2932	}
2933	}
2934
2935	return rc;
2936	}
2937
2938
2939	/**
2940	* Unlocks the memory pointed to by pv.
2941	*
2942	* @returns IPRT status code.
2943	* @param pSession Session to which the memory was locked.
2944	* @param pvR3 Memory to unlock.
2945	*/
2946	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2947	{
2948	LogFlow(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2949	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2950	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
2951	}
2952
2953
2954	/**
2955	* Allocates a chunk of page aligned memory with contiguous and fixed physical
2956	* backing.
2957	*
2958	* @returns IPRT status code.
2959	* @param pSession Session data.
2960	* @param cPages Number of pages to allocate.
2961	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
2962	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
2963	* @param pHCPhys Where to put the physical address of allocated memory.
2964	*/
2965	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
2966	{
2967	int rc;
2968	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2969	LogFlow(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
2970
2971	/*
2972	* Validate input.
2973	*/
2974	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2975	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
2976	{
2977	Log(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
2978	pSession, ppvR0, ppvR3, pHCPhys));
2979	return VERR_INVALID_PARAMETER;
2980
2981	}
2982	if (cPages < 1 \|\| cPages >= 256)
2983	{
2984	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2985	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2986	}
2987
2988	/*
2989	* Let IPRT do the job.
2990	*/
2991	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
2992	if (RT_SUCCESS(rc))
2993	{
2994	int rc2;
2995	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2996	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2997	if (RT_SUCCESS(rc))
2998	{
2999	Mem.eType = MEMREF_TYPE_CONT;
3000	rc = supdrvMemAdd(&Mem, pSession);
3001	if (!rc)
3002	{
3003	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
3004	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
3005	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
3006	return 0;
3007	}
3008
3009	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
3010	AssertRC(rc2);
3011	}
3012	rc2 = RTR0MemObjFree(Mem.MemObj, false);
3013	AssertRC(rc2);
3014	}
3015
3016	return rc;
3017	}
3018
3019
3020	/**
3021	* Frees memory allocated using SUPR0ContAlloc().
3022	*
3023	* @returns IPRT status code.
3024	* @param pSession The session to which the memory was allocated.
3025	* @param uPtr Pointer to the memory (ring-3 or ring-0).
3026	*/
3027	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
3028	{
3029	LogFlow(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
3030	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3031	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
3032	}
3033
3034
3035	/**
3036	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
3037	*
3038	* The memory isn't zeroed.
3039	*
3040	* @returns IPRT status code.
3041	* @param pSession Session data.
3042	* @param cPages Number of pages to allocate.
3043	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
3044	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
3045	* @param paPages Where to put the physical addresses of allocated memory.
3046	*/
3047	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
3048	{
3049	unsigned iPage;
3050	int rc;
3051	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
3052	LogFlow(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
3053
3054	/*
3055	* Validate input.
3056	*/
3057	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3058	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
3059	{
3060	Log(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
3061	pSession, ppvR3, ppvR0, paPages));
3062	return VERR_INVALID_PARAMETER;
3063
3064	}
3065	if (cPages < 1 \|\| cPages >= 256)
3066	{
3067	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
3068	return VERR_PAGE_COUNT_OUT_OF_RANGE;
3069	}
3070
3071	/*
3072	* Let IPRT do the work.
3073	*/
3074	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
3075	if (RT_SUCCESS(rc))
3076	{
3077	int rc2;
3078	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
3079	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
3080	if (RT_SUCCESS(rc))
3081	{
3082	Mem.eType = MEMREF_TYPE_LOW;
3083	rc = supdrvMemAdd(&Mem, pSession);
3084	if (!rc)
3085	{
3086	for (iPage = 0; iPage < cPages; iPage++)
3087	{
3088	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
3089	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%RHp\n", paPages[iPage]));
3090	}
3091	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
3092	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
3093	return 0;
3094	}
3095
3096	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
3097	AssertRC(rc2);
3098	}
3099
3100	rc2 = RTR0MemObjFree(Mem.MemObj, false);
3101	AssertRC(rc2);
3102	}
3103
3104	return rc;
3105	}
3106
3107
3108	/**
3109	* Frees memory allocated using SUPR0LowAlloc().
3110	*
3111	* @returns IPRT status code.
3112	* @param pSession The session to which the memory was allocated.
3113	* @param uPtr Pointer to the memory (ring-3 or ring-0).
3114	*/
3115	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
3116	{
3117	LogFlow(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
3118	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3119	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
3120	}
3121
3122
3123
3124	/**
3125	* Allocates a chunk of memory with both R0 and R3 mappings.
3126	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
3127	*
3128	* @returns IPRT status code.
3129	* @param pSession The session to associated the allocation with.
3130	* @param cb Number of bytes to allocate.
3131	* @param ppvR0 Where to store the address of the Ring-0 mapping.
3132	* @param ppvR3 Where to store the address of the Ring-3 mapping.
3133	*/
3134	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
3135	{
3136	int rc;
3137	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
3138	LogFlow(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
3139
3140	/*
3141	* Validate input.
3142	*/
3143	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3144	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
3145	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
3146	if (cb < 1 \|\| cb >= _4M)
3147	{
3148	Log(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
3149	return VERR_INVALID_PARAMETER;
3150	}
3151
3152	/*
3153	* Let IPRT do the work.
3154	*/
3155	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
3156	if (RT_SUCCESS(rc))
3157	{
3158	int rc2;
3159	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
3160	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
3161	if (RT_SUCCESS(rc))
3162	{
3163	Mem.eType = MEMREF_TYPE_MEM;
3164	rc = supdrvMemAdd(&Mem, pSession);
3165	if (!rc)
3166	{
3167	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
3168	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
3169	return VINF_SUCCESS;
3170	}
3171
3172	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
3173	AssertRC(rc2);
3174	}
3175
3176	rc2 = RTR0MemObjFree(Mem.MemObj, false);
3177	AssertRC(rc2);
3178	}
3179
3180	return rc;
3181	}
3182
3183
3184	/**
3185	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
3186	*
3187	* @returns IPRT status code.
3188	* @param pSession The session to which the memory was allocated.
3189	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
3190	* @param paPages Where to store the physical addresses.
3191	*/
3192	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
3193	{
3194	PSUPDRVBUNDLE pBundle;
3195	LogFlow(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
3196
3197	/*
3198	* Validate input.
3199	*/
3200	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3201	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
3202	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
3203
3204	/*
3205	* Search for the address.
3206	*/
3207	RTSpinlockAcquire(pSession->Spinlock);
3208	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3209	{
3210	if (pBundle->cUsed > 0)
3211	{
3212	unsigned i;
3213	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3214	{
3215	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
3216	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3217	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
3218	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3219	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
3220	)
3221	)
3222	{
3223	const size_t cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
3224	size_t iPage;
3225	for (iPage = 0; iPage < cPages; iPage++)
3226	{
3227	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
3228	paPages[iPage].uReserved = 0;
3229	}
3230	RTSpinlockRelease(pSession->Spinlock);
3231	return VINF_SUCCESS;
3232	}
3233	}
3234	}
3235	}
3236	RTSpinlockRelease(pSession->Spinlock);
3237	Log(("Failed to find %p!!!\n", (void *)uPtr));
3238	return VERR_INVALID_PARAMETER;
3239	}
3240
3241
3242	/**
3243	* Free memory allocated by SUPR0MemAlloc().
3244	*
3245	* @returns IPRT status code.
3246	* @param pSession The session owning the allocation.
3247	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
3248	*/
3249	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
3250	{
3251	LogFlow(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
3252	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3253	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
3254	}
3255
3256
3257	/**
3258	* Allocates a chunk of memory with a kernel or/and a user mode mapping.
3259	*
3260	* The memory is fixed and it's possible to query the physical addresses using
3261	* SUPR0MemGetPhys().
3262	*
3263	* @returns IPRT status code.
3264	* @param pSession The session to associated the allocation with.
3265	* @param cPages The number of pages to allocate.
3266	* @param fFlags Flags, reserved for the future. Must be zero.
3267	* @param ppvR3 Where to store the address of the Ring-3 mapping.
3268	* NULL if no ring-3 mapping.
3269	* @param ppvR3 Where to store the address of the Ring-0 mapping.
3270	* NULL if no ring-0 mapping.
3271	* @param paPages Where to store the addresses of the pages. Optional.
3272	*/
3273	SUPR0DECL(int) SUPR0PageAllocEx(PSUPDRVSESSION pSession, uint32_t cPages, uint32_t fFlags, PRTR3PTR ppvR3, PRTR0PTR ppvR0, PRTHCPHYS paPages)
3274	{
3275	int rc;
3276	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
3277	LogFlow(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
3278
3279	/*
3280	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
3281	*/
3282	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3283	AssertPtrNullReturn(ppvR3, VERR_INVALID_POINTER);
3284	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
3285	AssertReturn(ppvR3 \|\| ppvR0, VERR_INVALID_PARAMETER);
3286	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
3287	if (cPages < 1 \|\| cPages > VBOX_MAX_ALLOC_PAGE_COUNT)
3288	{
3289	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)));
3290	return VERR_PAGE_COUNT_OUT_OF_RANGE;
3291	}
3292
3293	/*
3294	* Let IPRT do the work.
3295	*/
3296	if (ppvR0)
3297	rc = RTR0MemObjAllocPage(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, true /* fExecutable */);
3298	else
3299	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
3300	if (RT_SUCCESS(rc))
3301	{
3302	int rc2;
3303	if (ppvR3)
3304	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
3305	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
3306	else
3307	Mem.MapObjR3 = NIL_RTR0MEMOBJ;
3308	if (RT_SUCCESS(rc))
3309	{
3310	Mem.eType = MEMREF_TYPE_PAGE;
3311	rc = supdrvMemAdd(&Mem, pSession);
3312	if (!rc)
3313	{
3314	if (ppvR3)
3315	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
3316	if (ppvR0)
3317	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
3318	if (paPages)
3319	{
3320	uint32_t iPage = cPages;
3321	while (iPage-- > 0)
3322	{
3323	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
3324	Assert(paPages[iPage] != NIL_RTHCPHYS);
3325	}
3326	}
3327	return VINF_SUCCESS;
3328	}
3329
3330	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
3331	AssertRC(rc2);
3332	}
3333
3334	rc2 = RTR0MemObjFree(Mem.MemObj, false);
3335	AssertRC(rc2);
3336	}
3337	return rc;
3338	}
3339
3340
3341	/**
3342	* Maps a chunk of memory previously allocated by SUPR0PageAllocEx into kernel
3343	* space.
3344	*
3345	* @returns IPRT status code.
3346	* @param pSession The session to associated the allocation with.
3347	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
3348	* @param offSub Where to start mapping. Must be page aligned.
3349	* @param cbSub How much to map. Must be page aligned.
3350	* @param fFlags Flags, MBZ.
3351	* @param ppvR0 Where to return the address of the ring-0 mapping on
3352	* success.
3353	*/
3354	SUPR0DECL(int) SUPR0PageMapKernel(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t offSub, uint32_t cbSub,
3355	uint32_t fFlags, PRTR0PTR ppvR0)
3356	{
3357	int rc;
3358	PSUPDRVBUNDLE pBundle;
3359	RTR0MEMOBJ hMemObj = NIL_RTR0MEMOBJ;
3360	LogFlow(("SUPR0PageMapKernel: pSession=%p pvR3=%p offSub=%#x cbSub=%#x\n", pSession, pvR3, offSub, cbSub));
3361
3362	/*
3363	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
3364	*/
3365	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3366	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
3367	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
3368	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3369	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3370	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
3371
3372	/*
3373	* Find the memory object.
3374	*/
3375	RTSpinlockAcquire(pSession->Spinlock);
3376	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3377	{
3378	if (pBundle->cUsed > 0)
3379	{
3380	unsigned i;
3381	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3382	{
3383	if ( ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
3384	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3385	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3386	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
3387	\|\| ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED
3388	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3389	&& pBundle->aMem[i].MapObjR3 == NIL_RTR0MEMOBJ
3390	&& RTR0MemObjAddressR3(pBundle->aMem[i].MemObj) == pvR3))
3391	{
3392	hMemObj = pBundle->aMem[i].MemObj;
3393	break;
3394	}
3395	}
3396	}
3397	}
3398	RTSpinlockRelease(pSession->Spinlock);
3399
3400	rc = VERR_INVALID_PARAMETER;
3401	if (hMemObj != NIL_RTR0MEMOBJ)
3402	{
3403	/*
3404	* Do some further input validations before calling IPRT.
3405	* (Cleanup is done indirectly by telling RTR0MemObjFree to include mappings.)
3406	*/
3407	size_t cbMemObj = RTR0MemObjSize(hMemObj);
3408	if ( offSub < cbMemObj
3409	&& cbSub <= cbMemObj
3410	&& offSub + cbSub <= cbMemObj)
3411	{
3412	RTR0MEMOBJ hMapObj;
3413	rc = RTR0MemObjMapKernelEx(&hMapObj, hMemObj, (void *)-1, 0,
3414	RTMEM_PROT_READ \| RTMEM_PROT_WRITE, offSub, cbSub);
3415	if (RT_SUCCESS(rc))
3416	*ppvR0 = RTR0MemObjAddress(hMapObj);
3417	}
3418	else
3419	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
3420
3421	}
3422	return rc;
3423	}
3424
3425
3426	/**
3427	* Changes the page level protection of one or more pages previously allocated
3428	* by SUPR0PageAllocEx.
3429	*
3430	* @returns IPRT status code.
3431	* @param pSession The session to associated the allocation with.
3432	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
3433	* NIL_RTR3PTR if the ring-3 mapping should be unaffected.
3434	* @param pvR0 The ring-0 address returned by SUPR0PageAllocEx.
3435	* NIL_RTR0PTR if the ring-0 mapping should be unaffected.
3436	* @param offSub Where to start changing. Must be page aligned.
3437	* @param cbSub How much to change. Must be page aligned.
3438	* @param fProt The new page level protection, see RTMEM_PROT_*.
3439	*/
3440	SUPR0DECL(int) SUPR0PageProtect(PSUPDRVSESSION pSession, RTR3PTR pvR3, RTR0PTR pvR0, uint32_t offSub, uint32_t cbSub, uint32_t fProt)
3441	{
3442	int rc;
3443	PSUPDRVBUNDLE pBundle;
3444	RTR0MEMOBJ hMemObjR0 = NIL_RTR0MEMOBJ;
3445	RTR0MEMOBJ hMemObjR3 = NIL_RTR0MEMOBJ;
3446	LogFlow(("SUPR0PageProtect: pSession=%p pvR3=%p pvR0=%p offSub=%#x cbSub=%#x fProt-%#x\n", pSession, pvR3, pvR0, offSub, cbSub, fProt));
3447
3448	/*
3449	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
3450	*/
3451	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3452	AssertReturn(!(fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)), VERR_INVALID_PARAMETER);
3453	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3454	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3455	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
3456
3457	/*
3458	* Find the memory object.
3459	*/
3460	RTSpinlockAcquire(pSession->Spinlock);
3461	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3462	{
3463	if (pBundle->cUsed > 0)
3464	{
3465	unsigned i;
3466	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3467	{
3468	if ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
3469	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3470	&& ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3471	\|\| pvR3 == NIL_RTR3PTR)
3472	&& ( pvR0 == NIL_RTR0PTR
3473	\|\| RTR0MemObjAddress(pBundle->aMem[i].MemObj) == pvR0)
3474	&& ( pvR3 == NIL_RTR3PTR
3475	\|\| RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3))
3476	{
3477	if (pvR0 != NIL_RTR0PTR)
3478	hMemObjR0 = pBundle->aMem[i].MemObj;
3479	if (pvR3 != NIL_RTR3PTR)
3480	hMemObjR3 = pBundle->aMem[i].MapObjR3;
3481	break;
3482	}
3483	}
3484	}
3485	}
3486	RTSpinlockRelease(pSession->Spinlock);
3487
3488	rc = VERR_INVALID_PARAMETER;
3489	if ( hMemObjR0 != NIL_RTR0MEMOBJ
3490	\|\| hMemObjR3 != NIL_RTR0MEMOBJ)
3491	{
3492	/*
3493	* Do some further input validations before calling IPRT.
3494	*/
3495	size_t cbMemObj = hMemObjR0 != NIL_RTR0PTR ? RTR0MemObjSize(hMemObjR0) : RTR0MemObjSize(hMemObjR3);
3496	if ( offSub < cbMemObj
3497	&& cbSub <= cbMemObj
3498	&& offSub + cbSub <= cbMemObj)
3499	{
3500	rc = VINF_SUCCESS;
3501	if (hMemObjR3 != NIL_RTR0PTR)
3502	rc = RTR0MemObjProtect(hMemObjR3, offSub, cbSub, fProt);
3503	if (hMemObjR0 != NIL_RTR0PTR && RT_SUCCESS(rc))
3504	rc = RTR0MemObjProtect(hMemObjR0, offSub, cbSub, fProt);
3505	}
3506	else
3507	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
3508
3509	}
3510	return rc;
3511
3512	}
3513
3514
3515	/**
3516	* Free memory allocated by SUPR0PageAlloc() and SUPR0PageAllocEx().
3517	*
3518	* @returns IPRT status code.
3519	* @param pSession The session owning the allocation.
3520	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc() or
3521	* SUPR0PageAllocEx().
3522	*/
3523	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
3524	{
3525	LogFlow(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
3526	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3527	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_PAGE);
3528	}
3529
3530
3531	/**
3532	* Gets the paging mode of the current CPU.
3533	*
3534	* @returns Paging mode, SUPPAGEINGMODE_INVALID on error.
3535	*/
3536	SUPR0DECL(SUPPAGINGMODE) SUPR0GetPagingMode(void)
3537	{
3538	SUPPAGINGMODE enmMode;
3539
3540	RTR0UINTREG cr0 = ASMGetCR0();
3541	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
3542	enmMode = SUPPAGINGMODE_INVALID;
3543	else
3544	{
3545	RTR0UINTREG cr4 = ASMGetCR4();
3546	uint32_t fNXEPlusLMA = 0;
3547	if (cr4 & X86_CR4_PAE)
3548	{
3549	uint32_t fExtFeatures = ASMCpuId_EDX(0x80000001);
3550	if (fExtFeatures & (X86_CPUID_EXT_FEATURE_EDX_NX \| X86_CPUID_EXT_FEATURE_EDX_LONG_MODE))
3551	{
3552	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
3553	if ((fExtFeatures & X86_CPUID_EXT_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
3554	fNXEPlusLMA \|= RT_BIT(0);
3555	if ((fExtFeatures & X86_CPUID_EXT_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
3556	fNXEPlusLMA \|= RT_BIT(1);
3557	}
3558	}
3559
3560	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
3561	{
3562	case 0:
3563	enmMode = SUPPAGINGMODE_32_BIT;
3564	break;
3565
3566	case X86_CR4_PGE:
3567	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
3568	break;
3569
3570	case X86_CR4_PAE:
3571	enmMode = SUPPAGINGMODE_PAE;
3572	break;
3573
3574	case X86_CR4_PAE \| RT_BIT(0):
3575	enmMode = SUPPAGINGMODE_PAE_NX;
3576	break;
3577
3578	case X86_CR4_PAE \| X86_CR4_PGE:
3579	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3580	break;
3581
3582	case X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3583	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3584	break;
3585
3586	case RT_BIT(1) \| X86_CR4_PAE:
3587	enmMode = SUPPAGINGMODE_AMD64;
3588	break;
3589
3590	case RT_BIT(1) \| X86_CR4_PAE \| RT_BIT(0):
3591	enmMode = SUPPAGINGMODE_AMD64_NX;
3592	break;
3593
3594	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
3595	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
3596	break;
3597
3598	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3599	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
3600	break;
3601
3602	default:
3603	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
3604	enmMode = SUPPAGINGMODE_INVALID;
3605	break;
3606	}
3607	}
3608	return enmMode;
3609	}
3610
3611
3612	/**
3613	* Enables or disabled hardware virtualization extensions using native OS APIs.
3614	*
3615	* @returns VBox status code.
3616	* @retval VINF_SUCCESS on success.
3617	* @retval VERR_NOT_SUPPORTED if not supported by the native OS.
3618	*
3619	* @param fEnable Whether to enable or disable.
3620	*/
3621	SUPR0DECL(int) SUPR0EnableVTx(bool fEnable)
3622	{
3623	#ifdef RT_OS_DARWIN
3624	return supdrvOSEnableVTx(fEnable);
3625	#else
3626	return VERR_NOT_SUPPORTED;
3627	#endif
3628	}
3629
3630
3631	/**
3632	* Suspends hardware virtualization extensions using the native OS API.
3633	*
3634	* This is called prior to entering raw-mode context.
3635	*
3636	* @returns @c true if suspended, @c false if not.
3637	*/
3638	SUPR0DECL(bool) SUPR0SuspendVTxOnCpu(void)
3639	{
3640	#ifdef RT_OS_DARWIN
3641	return supdrvOSSuspendVTxOnCpu();
3642	#else
3643	return false;
3644	#endif
3645	}
3646
3647
3648	/**
3649	* Resumes hardware virtualization extensions using the native OS API.
3650	*
3651	* This is called after to entering raw-mode context.
3652	*
3653	* @param fSuspended The return value of SUPR0SuspendVTxOnCpu.
3654	*/
3655	SUPR0DECL(void) SUPR0ResumeVTxOnCpu(bool fSuspended)
3656	{
3657	#ifdef RT_OS_DARWIN
3658	supdrvOSResumeVTxOnCpu(fSuspended);
3659	#else
3660	Assert(!fSuspended);
3661	#endif
3662	}
3663
3664
3665	/**
3666	* Queries the AMD-V and VT-x capabilities of the calling CPU.
3667	*
3668	* @returns VBox status code.
3669	* @retval VERR_VMX_NO_VMX
3670	* @retval VERR_VMX_MSR_ALL_VMXON_DISABLED
3671	* @retval VERR_VMX_MSR_VMXON_DISABLED
3672	* @retval VERR_VMX_MSR_LOCKING_FAILED
3673	* @retval VERR_SVM_NO_SVM
3674	* @retval VERR_SVM_DISABLED
3675	* @retval VERR_UNSUPPORTED_CPU if not identifiable as an AMD, Intel or VIA
3676	* (centaur) CPU.
3677	*
3678	* @param pSession The session handle.
3679	* @param pfCaps Where to store the capabilities.
3680	*/
3681	SUPR0DECL(int) SUPR0QueryVTCaps(PSUPDRVSESSION pSession, uint32_t *pfCaps)
3682	{
3683	int rc = VERR_UNSUPPORTED_CPU;
3684	bool fIsSmxModeAmbiguous = false;
3685	RTTHREADPREEMPTSTATE PreemptState = RTTHREADPREEMPTSTATE_INITIALIZER;
3686
3687	/*
3688	* Input validation.
3689	*/
3690	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3691	AssertPtrReturn(pfCaps, VERR_INVALID_POINTER);
3692
3693	*pfCaps = 0;
3694	/* We may modify MSRs and re-read them, disable preemption so we make sure we don't migrate CPUs. */
3695	RTThreadPreemptDisable(&PreemptState);
3696	if (ASMHasCpuId())
3697	{
3698	uint32_t fFeaturesECX, fFeaturesEDX, uDummy;
3699	uint32_t uMaxId, uVendorEBX, uVendorECX, uVendorEDX;
3700
3701	ASMCpuId(0, &uMaxId, &uVendorEBX, &uVendorECX, &uVendorEDX);
3702	ASMCpuId(1, &uDummy, &uDummy, &fFeaturesECX, &fFeaturesEDX);
3703
3704	if ( ASMIsValidStdRange(uMaxId)
3705	&& ( ASMIsIntelCpuEx( uVendorEBX, uVendorECX, uVendorEDX)
3706	\|\| ASMIsViaCentaurCpuEx(uVendorEBX, uVendorECX, uVendorEDX) )
3707	)
3708	{
3709	if ( (fFeaturesECX & X86_CPUID_FEATURE_ECX_VMX)
3710	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3711	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3712	)
3713	{
3714	/** @todo Unify code with hmR0InitIntelCpu(). */
3715	uint64_t u64FeatMsr = ASMRdMsr(MSR_IA32_FEATURE_CONTROL);
3716	bool const fMaybeSmxMode = RT_BOOL(ASMGetCR4() & X86_CR4_SMXE);
3717	bool fMsrLocked = RT_BOOL(u64FeatMsr & MSR_IA32_FEATURE_CONTROL_LOCK);
3718	bool fSmxVmxAllowed = RT_BOOL(u64FeatMsr & MSR_IA32_FEATURE_CONTROL_SMX_VMXON);
3719	bool fVmxAllowed = RT_BOOL(u64FeatMsr & MSR_IA32_FEATURE_CONTROL_VMXON);
3720
3721	/* Check if the LOCK bit is set but excludes the required VMXON bit. */
3722	if (fMsrLocked)
3723	{
3724	if (fVmxAllowed && fSmxVmxAllowed)
3725	rc = VINF_SUCCESS;
3726	else if (!fVmxAllowed && !fSmxVmxAllowed)
3727	rc = VERR_VMX_MSR_ALL_VMXON_DISABLED;
3728	else if (!fMaybeSmxMode)
3729	{
3730	if (fVmxAllowed)
3731	rc = VINF_SUCCESS;
3732	else
3733	rc = VERR_VMX_MSR_VMXON_DISABLED;
3734	}
3735	else
3736	{
3737	/*
3738	* CR4.SMXE is set but this doesn't mean the CPU is necessarily in SMX mode. We shall assume
3739	* that it is -not- and that it is a stupid BIOS/OS setting CR4.SMXE for no good reason.
3740	* See @bugref{6873}.
3741	*/
3742	Assert(fMaybeSmxMode == true);
3743	fIsSmxModeAmbiguous = true;
3744	rc = VINF_SUCCESS;
3745	}
3746	}
3747	else
3748	{
3749	/*
3750	* MSR is not yet locked; we can change it ourselves here.
3751	* Once the lock bit is set, this MSR can no longer be modified.
3752	*
3753	* Set both the VMXON and SMX_VMXON bits as we can't determine SMX mode
3754	* accurately. See @bugref{6873}.
3755	*/
3756	u64FeatMsr \|= MSR_IA32_FEATURE_CONTROL_LOCK
3757	\| MSR_IA32_FEATURE_CONTROL_SMX_VMXON
3758	\| MSR_IA32_FEATURE_CONTROL_VMXON;
3759	ASMWrMsr(MSR_IA32_FEATURE_CONTROL, u64FeatMsr);
3760
3761	/* Verify. */
3762	u64FeatMsr = ASMRdMsr(MSR_IA32_FEATURE_CONTROL);
3763	fMsrLocked = RT_BOOL(u64FeatMsr & MSR_IA32_FEATURE_CONTROL_LOCK);
3764	fSmxVmxAllowed = fMsrLocked && RT_BOOL(u64FeatMsr & MSR_IA32_FEATURE_CONTROL_SMX_VMXON);
3765	fVmxAllowed = fMsrLocked && RT_BOOL(u64FeatMsr & MSR_IA32_FEATURE_CONTROL_VMXON);
3766	if (fSmxVmxAllowed && fVmxAllowed)
3767	rc = VINF_SUCCESS;
3768	else
3769	rc = VERR_VMX_MSR_LOCKING_FAILED;
3770	}
3771
3772	if (rc == VINF_SUCCESS)
3773	{
3774	VMXCAPABILITY vtCaps;
3775
3776	*pfCaps \|= SUPVTCAPS_VT_X;
3777
3778	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS);
3779	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC_USE_SECONDARY_EXEC_CTRL)
3780	{
3781	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS2);
3782	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC2_EPT)
3783	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3784	}
3785	}
3786	}
3787	else
3788	rc = VERR_VMX_NO_VMX;
3789	}
3790	else if ( ASMIsAmdCpuEx(uVendorEBX, uVendorECX, uVendorEDX)
3791	&& ASMIsValidStdRange(uMaxId))
3792	{
3793	uint32_t fExtFeaturesEcx, uExtMaxId;
3794	ASMCpuId(0x80000000, &uExtMaxId, &uDummy, &uDummy, &uDummy);
3795	ASMCpuId(0x80000001, &uDummy, &uDummy, &fExtFeaturesEcx, &uDummy);
3796	if ( ASMIsValidExtRange(uExtMaxId)
3797	&& uExtMaxId >= 0x8000000a
3798	&& (fExtFeaturesEcx & X86_CPUID_AMD_FEATURE_ECX_SVM)
3799	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3800	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3801	)
3802	{
3803	/* Check if SVM is disabled */
3804	uint64_t u64FeatMsr = ASMRdMsr(MSR_K8_VM_CR);
3805	if (!(u64FeatMsr & MSR_K8_VM_CR_SVM_DISABLE))
3806	{
3807	uint32_t fSvmFeatures;
3808	*pfCaps \|= SUPVTCAPS_AMD_V;
3809
3810	/* Query AMD-V features. */
3811	ASMCpuId(0x8000000a, &uDummy, &uDummy, &uDummy, &fSvmFeatures);
3812	if (fSvmFeatures & AMD_CPUID_SVM_FEATURE_EDX_NESTED_PAGING)
3813	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3814
3815	rc = VINF_SUCCESS;
3816	}
3817	else
3818	rc = VERR_SVM_DISABLED;
3819	}
3820	else
3821	rc = VERR_SVM_NO_SVM;
3822	}
3823	}
3824
3825	RTThreadPreemptRestore(&PreemptState);
3826	if (fIsSmxModeAmbiguous)
3827	SUPR0Printf(("WARNING! CR4 hints SMX mode but your CPU is too secretive. Proceeding anyway... We wish you good luck!\n"));
3828	return rc;
3829	}
3830
3831
3832	/**
3833	* (Re-)initializes the per-cpu structure prior to starting or resuming the GIP
3834	* updating.
3835	*
3836	* @param pGipCpu The per CPU structure for this CPU.
3837	* @param u64NanoTS The current time.
3838	*/
3839	static void supdrvGipReInitCpu(PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
3840	{
3841	pGipCpu->u64TSC = ASMReadTSC() - pGipCpu->u32UpdateIntervalTSC;
3842	pGipCpu->u64NanoTS = u64NanoTS;
3843	}
3844
3845
3846	/**
3847	* Set the current TSC and NanoTS value for the CPU.
3848	*
3849	* @param idCpu The CPU ID. Unused - we have to use the APIC ID.
3850	* @param pvUser1 Pointer to the ring-0 GIP mapping.
3851	* @param pvUser2 Pointer to the variable holding the current time.
3852	*/
3853	static DECLCALLBACK(void) supdrvGipReInitCpuCallback(RTCPUID idCpu, void pvUser1, void pvUser2)
3854	{
3855	PSUPGLOBALINFOPAGE pGip = (PSUPGLOBALINFOPAGE)pvUser1;
3856	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
3857
3858	if (RT_LIKELY(iCpu < pGip->cCpus && pGip->aCPUs[iCpu].idCpu == idCpu))
3859	supdrvGipReInitCpu(&pGip->aCPUs[iCpu], (uint64_t )pvUser2);
3860
3861	NOREF(pvUser2);
3862	NOREF(idCpu);
3863	}
3864
3865
3866	/**
3867	* Maps the GIP into userspace and/or get the physical address of the GIP.
3868	*
3869	* @returns IPRT status code.
3870	* @param pSession Session to which the GIP mapping should belong.
3871	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
3872	* @param pHCPhysGip Where to store the physical address. (optional)
3873	*
3874	* @remark There is no reference counting on the mapping, so one call to this function
3875	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
3876	* and remove the session as a GIP user.
3877	*/
3878	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
3879	{
3880	int rc;
3881	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3882	RTR3PTR pGipR3 = NIL_RTR3PTR;
3883	RTHCPHYS HCPhys = NIL_RTHCPHYS;
3884	LogFlow(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
3885
3886	/*
3887	* Validate
3888	*/
3889	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3890	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
3891	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
3892
3893	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3894	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3895	#else
3896	RTSemFastMutexRequest(pDevExt->mtxGip);
3897	#endif
3898	if (pDevExt->pGip)
3899	{
3900	/*
3901	* Map it?
3902	*/
3903	rc = VINF_SUCCESS;
3904	if (ppGipR3)
3905	{
3906	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
3907	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
3908	RTMEM_PROT_READ, RTR0ProcHandleSelf());
3909	if (RT_SUCCESS(rc))
3910	pGipR3 = RTR0MemObjAddressR3(pSession->GipMapObjR3);
3911	}
3912
3913	/*
3914	* Get physical address.
3915	*/
3916	if (pHCPhysGip && RT_SUCCESS(rc))
3917	HCPhys = pDevExt->HCPhysGip;
3918
3919	/*
3920	* Reference globally.
3921	*/
3922	if (!pSession->fGipReferenced && RT_SUCCESS(rc))
3923	{
3924	pSession->fGipReferenced = 1;
3925	pDevExt->cGipUsers++;
3926	if (pDevExt->cGipUsers == 1)
3927	{
3928	PSUPGLOBALINFOPAGE pGipR0 = pDevExt->pGip;
3929	uint64_t u64NanoTS;
3930	uint32_t u32SystemResolution;
3931	unsigned i;
3932
3933	LogFlow(("SUPR0GipMap: Resumes GIP updating\n"));
3934
3935	/*
3936	* Try bump up the system timer resolution.
3937	* The more interrupts the better...
3938	*/
3939	if ( RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
3940	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
3941	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1953125 /* 512 HZ */, &u32SystemResolution))
3942	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 2000000 /* 500 HZ */, &u32SystemResolution))
3943	)
3944	{
3945	Assert(RTTimerGetSystemGranularity() <= u32SystemResolution);
3946	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
3947	}
3948
3949	if (pGipR0->aCPUs[0].u32TransactionId != 2 /* not the first time */)
3950	{
3951	for (i = 0; i < RT_ELEMENTS(pGipR0->aCPUs); i++)
3952	ASMAtomicUoWriteU32(&pGipR0->aCPUs[i].u32TransactionId,
3953	(pGipR0->aCPUs[i].u32TransactionId + GIP_UPDATEHZ_RECALC_FREQ * 2)
3954	& ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
3955	ASMAtomicWriteU64(&pGipR0->u64NanoTSLastUpdateHz, 0);
3956	}
3957
3958	u64NanoTS = RTTimeSystemNanoTS() - pGipR0->u32UpdateIntervalNS;
3959	if ( pGipR0->u32Mode == SUPGIPMODE_SYNC_TSC
3960	\|\| RTMpGetOnlineCount() == 1)
3961	supdrvGipReInitCpu(&pGipR0->aCPUs[0], u64NanoTS);
3962	else
3963	RTMpOnAll(supdrvGipReInitCpuCallback, pGipR0, &u64NanoTS);
3964
3965	#ifndef DO_NOT_START_GIP
3966	rc = RTTimerStart(pDevExt->pGipTimer, 0); AssertRC(rc);
3967	#endif
3968	rc = VINF_SUCCESS;
3969	}
3970	}
3971	}
3972	else
3973	{
3974	rc = VERR_GENERAL_FAILURE;
3975	Log(("SUPR0GipMap: GIP is not available!\n"));
3976	}
3977	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3978	RTSemMutexRelease(pDevExt->mtxGip);
3979	#else
3980	RTSemFastMutexRelease(pDevExt->mtxGip);
3981	#endif
3982
3983	/*
3984	* Write returns.
3985	*/
3986	if (pHCPhysGip)
3987	*pHCPhysGip = HCPhys;
3988	if (ppGipR3)
3989	*ppGipR3 = pGipR3;
3990
3991	#ifdef DEBUG_DARWIN_GIP
3992	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3993	#else
3994	LogFlow(( "SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3995	#endif
3996	return rc;
3997	}
3998
3999
4000	/**
4001	* Unmaps any user mapping of the GIP and terminates all GIP access
4002	* from this session.
4003	*
4004	* @returns IPRT status code.
4005	* @param pSession Session to which the GIP mapping should belong.
4006	*/
4007	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
4008	{
4009	int rc = VINF_SUCCESS;
4010	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
4011	#ifdef DEBUG_DARWIN_GIP
4012	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
4013	pSession,
4014	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
4015	pSession->GipMapObjR3));
4016	#else
4017	LogFlow(("SUPR0GipUnmap: pSession=%p\n", pSession));
4018	#endif
4019	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
4020
4021	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
4022	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
4023	#else
4024	RTSemFastMutexRequest(pDevExt->mtxGip);
4025	#endif
4026
4027	/*
4028	* Unmap anything?
4029	*/
4030	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
4031	{
4032	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
4033	AssertRC(rc);
4034	if (RT_SUCCESS(rc))
4035	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
4036	}
4037
4038	/*
4039	* Dereference global GIP.
4040	*/
4041	if (pSession->fGipReferenced && !rc)
4042	{
4043	pSession->fGipReferenced = 0;
4044	if ( pDevExt->cGipUsers > 0
4045	&& !--pDevExt->cGipUsers)
4046	{
4047	LogFlow(("SUPR0GipUnmap: Suspends GIP updating\n"));
4048	#ifndef DO_NOT_START_GIP
4049	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = VINF_SUCCESS;
4050	#endif
4051
4052	if (pDevExt->u32SystemTimerGranularityGrant)
4053	{
4054	int rc2 = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant);
4055	AssertRC(rc2);
4056	pDevExt->u32SystemTimerGranularityGrant = 0;
4057	}
4058	}
4059	}
4060
4061	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
4062	RTSemMutexRelease(pDevExt->mtxGip);
4063	#else
4064	RTSemFastMutexRelease(pDevExt->mtxGip);
4065	#endif
4066
4067	return rc;
4068	}
4069
4070
4071	/**
4072	* Gets the GIP pointer.
4073	*
4074	* @returns Pointer to the GIP or NULL.
4075	*/
4076	SUPDECL(PSUPGLOBALINFOPAGE) SUPGetGIP(void)
4077	{
4078	return g_pSUPGlobalInfoPage;
4079	}
4080
4081
4082	/**
4083	* Register a component factory with the support driver.
4084	*
4085	* This is currently restricted to kernel sessions only.
4086	*
4087	* @returns VBox status code.
4088	* @retval VINF_SUCCESS on success.
4089	* @retval VERR_NO_MEMORY if we're out of memory.
4090	* @retval VERR_ALREADY_EXISTS if the factory has already been registered.
4091	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
4092	* @retval VERR_INVALID_PARAMETER on invalid parameter.
4093	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
4094	*
4095	* @param pSession The SUPDRV session (must be a ring-0 session).
4096	* @param pFactory Pointer to the component factory registration structure.
4097	*
4098	* @remarks This interface is also available via SUPR0IdcComponentRegisterFactory.
4099	*/
4100	SUPR0DECL(int) SUPR0ComponentRegisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
4101	{
4102	PSUPDRVFACTORYREG pNewReg;
4103	const char *psz;
4104	int rc;
4105
4106	/*
4107	* Validate parameters.
4108	*/
4109	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
4110	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
4111	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
4112	AssertPtrReturn(pFactory->pfnQueryFactoryInterface, VERR_INVALID_POINTER);
4113	psz = RTStrEnd(pFactory->szName, sizeof(pFactory->szName));
4114	AssertReturn(psz, VERR_INVALID_PARAMETER);
4115
4116	/*
4117	* Allocate and initialize a new registration structure.
4118	*/
4119	pNewReg = (PSUPDRVFACTORYREG)RTMemAlloc(sizeof(SUPDRVFACTORYREG));
4120	if (pNewReg)
4121	{
4122	pNewReg->pNext = NULL;
4123	pNewReg->pFactory = pFactory;
4124	pNewReg->pSession = pSession;
4125	pNewReg->cchName = psz - &pFactory->szName[0];
4126
4127	/*
4128	* Add it to the tail of the list after checking for prior registration.
4129	*/
4130	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
4131	if (RT_SUCCESS(rc))
4132	{
4133	PSUPDRVFACTORYREG pPrev = NULL;
4134	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
4135	while (pCur && pCur->pFactory != pFactory)
4136	{
4137	pPrev = pCur;
4138	pCur = pCur->pNext;
4139	}
4140	if (!pCur)
4141	{
4142	if (pPrev)
4143	pPrev->pNext = pNewReg;
4144	else
4145	pSession->pDevExt->pComponentFactoryHead = pNewReg;
4146	rc = VINF_SUCCESS;
4147	}
4148	else
4149	rc = VERR_ALREADY_EXISTS;
4150
4151	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
4152	}
4153
4154	if (RT_FAILURE(rc))
4155	RTMemFree(pNewReg);
4156	}
4157	else
4158	rc = VERR_NO_MEMORY;
4159	return rc;
4160	}
4161
4162
4163	/**
4164	* Deregister a component factory.
4165	*
4166	* @returns VBox status code.
4167	* @retval VINF_SUCCESS on success.
4168	* @retval VERR_NOT_FOUND if the factory wasn't registered.
4169	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
4170	* @retval VERR_INVALID_PARAMETER on invalid parameter.
4171	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
4172	*
4173	* @param pSession The SUPDRV session (must be a ring-0 session).
4174	* @param pFactory Pointer to the component factory registration structure
4175	* previously passed SUPR0ComponentRegisterFactory().
4176	*
4177	* @remarks This interface is also available via SUPR0IdcComponentDeregisterFactory.
4178	*/
4179	SUPR0DECL(int) SUPR0ComponentDeregisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
4180	{
4181	int rc;
4182
4183	/*
4184	* Validate parameters.
4185	*/
4186	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
4187	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
4188	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
4189
4190	/*
4191	* Take the lock and look for the registration record.
4192	*/
4193	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
4194	if (RT_SUCCESS(rc))
4195	{
4196	PSUPDRVFACTORYREG pPrev = NULL;
4197	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
4198	while (pCur && pCur->pFactory != pFactory)
4199	{
4200	pPrev = pCur;
4201	pCur = pCur->pNext;
4202	}
4203	if (pCur)
4204	{
4205	if (!pPrev)
4206	pSession->pDevExt->pComponentFactoryHead = pCur->pNext;
4207	else
4208	pPrev->pNext = pCur->pNext;
4209
4210	pCur->pNext = NULL;
4211	pCur->pFactory = NULL;
4212	pCur->pSession = NULL;
4213	rc = VINF_SUCCESS;
4214	}
4215	else
4216	rc = VERR_NOT_FOUND;
4217
4218	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
4219
4220	RTMemFree(pCur);
4221	}
4222	return rc;
4223	}
4224
4225
4226	/**
4227	* Queries a component factory.
4228	*
4229	* @returns VBox status code.
4230	* @retval VERR_INVALID_PARAMETER on invalid parameter.
4231	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
4232	* @retval VERR_SUPDRV_COMPONENT_NOT_FOUND if the component factory wasn't found.
4233	* @retval VERR_SUPDRV_INTERFACE_NOT_SUPPORTED if the interface wasn't supported.
4234	*
4235	* @param pSession The SUPDRV session.
4236	* @param pszName The name of the component factory.
4237	* @param pszInterfaceUuid The UUID of the factory interface (stringified).
4238	* @param ppvFactoryIf Where to store the factory interface.
4239	*/
4240	SUPR0DECL(int) SUPR0ComponentQueryFactory(PSUPDRVSESSION pSession, const char pszName, const char pszInterfaceUuid, void **ppvFactoryIf)
4241	{
4242	const char *pszEnd;
4243	size_t cchName;
4244	int rc;
4245
4246	/*
4247	* Validate parameters.
4248	*/
4249	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
4250
4251	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
4252	pszEnd = RTStrEnd(pszName, RT_SIZEOFMEMB(SUPDRVFACTORY, szName));
4253	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4254	cchName = pszEnd - pszName;
4255
4256	AssertPtrReturn(pszInterfaceUuid, VERR_INVALID_POINTER);
4257	pszEnd = RTStrEnd(pszInterfaceUuid, RTUUID_STR_LENGTH);
4258	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4259
4260	AssertPtrReturn(ppvFactoryIf, VERR_INVALID_POINTER);
4261	*ppvFactoryIf = NULL;
4262
4263	/*
4264	* Take the lock and try all factories by this name.
4265	*/
4266	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
4267	if (RT_SUCCESS(rc))
4268	{
4269	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
4270	rc = VERR_SUPDRV_COMPONENT_NOT_FOUND;
4271	while (pCur)
4272	{
4273	if ( pCur->cchName == cchName
4274	&& !memcmp(pCur->pFactory->szName, pszName, cchName))
4275	{
4276	void *pvFactory = pCur->pFactory->pfnQueryFactoryInterface(pCur->pFactory, pSession, pszInterfaceUuid);
4277	if (pvFactory)
4278	{
4279	*ppvFactoryIf = pvFactory;
4280	rc = VINF_SUCCESS;
4281	break;
4282	}
4283	rc = VERR_SUPDRV_INTERFACE_NOT_SUPPORTED;
4284	}
4285
4286	/* next */
4287	pCur = pCur->pNext;
4288	}
4289
4290	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
4291	}
4292	return rc;
4293	}
4294
4295
4296	/**
4297	* Adds a memory object to the session.
4298	*
4299	* @returns IPRT status code.
4300	* @param pMem Memory tracking structure containing the
4301	* information to track.
4302	* @param pSession The session.
4303	*/
4304	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
4305	{
4306	PSUPDRVBUNDLE pBundle;
4307
4308	/*
4309	* Find free entry and record the allocation.
4310	*/
4311	RTSpinlockAcquire(pSession->Spinlock);
4312	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
4313	{
4314	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
4315	{
4316	unsigned i;
4317	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
4318	{
4319	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
4320	{
4321	pBundle->cUsed++;
4322	pBundle->aMem[i] = *pMem;
4323	RTSpinlockRelease(pSession->Spinlock);
4324	return VINF_SUCCESS;
4325	}
4326	}
4327	AssertFailed(); /* !!this can't be happening!!! */
4328	}
4329	}
4330	RTSpinlockRelease(pSession->Spinlock);
4331
4332	/*
4333	* Need to allocate a new bundle.
4334	* Insert into the last entry in the bundle.
4335	*/
4336	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
4337	if (!pBundle)
4338	return VERR_NO_MEMORY;
4339
4340	/* take last entry. */
4341	pBundle->cUsed++;
4342	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
4343
4344	/* insert into list. */
4345	RTSpinlockAcquire(pSession->Spinlock);
4346	pBundle->pNext = pSession->Bundle.pNext;
4347	pSession->Bundle.pNext = pBundle;
4348	RTSpinlockRelease(pSession->Spinlock);
4349
4350	return VINF_SUCCESS;
4351	}
4352
4353
4354	/**
4355	* Releases a memory object referenced by pointer and type.
4356	*
4357	* @returns IPRT status code.
4358	* @param pSession Session data.
4359	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
4360	* @param eType Memory type.
4361	*/
4362	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
4363	{
4364	PSUPDRVBUNDLE pBundle;
4365
4366	/*
4367	* Validate input.
4368	*/
4369	if (!uPtr)
4370	{
4371	Log(("Illegal address %p\n", (void *)uPtr));
4372	return VERR_INVALID_PARAMETER;
4373	}
4374
4375	/*
4376	* Search for the address.
4377	*/
4378	RTSpinlockAcquire(pSession->Spinlock);
4379	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
4380	{
4381	if (pBundle->cUsed > 0)
4382	{
4383	unsigned i;
4384	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
4385	{
4386	if ( pBundle->aMem[i].eType == eType
4387	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
4388	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
4389	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
4390	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
4391	)
4392	{
4393	/* Make a copy of it and release it outside the spinlock. */
4394	SUPDRVMEMREF Mem = pBundle->aMem[i];
4395	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
4396	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
4397	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
4398	RTSpinlockRelease(pSession->Spinlock);
4399
4400	if (Mem.MapObjR3 != NIL_RTR0MEMOBJ)
4401	{
4402	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
4403	AssertRC(rc); /** @todo figure out how to handle this. */
4404	}
4405	if (Mem.MemObj != NIL_RTR0MEMOBJ)
4406	{
4407	int rc = RTR0MemObjFree(Mem.MemObj, true /* fFreeMappings */);
4408	AssertRC(rc); /** @todo figure out how to handle this. */
4409	}
4410	return VINF_SUCCESS;
4411	}
4412	}
4413	}
4414	}
4415	RTSpinlockRelease(pSession->Spinlock);
4416	Log(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
4417	return VERR_INVALID_PARAMETER;
4418	}
4419
4420
4421	/**
4422	* Opens an image. If it's the first time it's opened the call must upload
4423	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
4424	*
4425	* This is the 1st step of the loading.
4426	*
4427	* @returns IPRT status code.
4428	* @param pDevExt Device globals.
4429	* @param pSession Session data.
4430	* @param pReq The open request.
4431	*/
4432	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
4433	{
4434	int rc;
4435	PSUPDRVLDRIMAGE pImage;
4436	void *pv;
4437	size_t cchName = strlen(pReq->u.In.szName); /* (caller checked < 32). */
4438	LogFlow(("supdrvIOCtl_LdrOpen: szName=%s cbImageWithTabs=%d\n", pReq->u.In.szName, pReq->u.In.cbImageWithTabs));
4439
4440	/*
4441	* Check if we got an instance of the image already.
4442	*/
4443	supdrvLdrLock(pDevExt);
4444	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
4445	{
4446	if ( pImage->szName[cchName] == '\0'
4447	&& !memcmp(pImage->szName, pReq->u.In.szName, cchName))
4448	{
4449	/** @todo check cbImageBits and cbImageWithTabs here, if they differs that indicates that the images are different. */
4450	pImage->cUsage++;
4451	pReq->u.Out.pvImageBase = pImage->pvImage;
4452	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
4453	pReq->u.Out.fNativeLoader = pImage->fNative;
4454	supdrvLdrAddUsage(pSession, pImage);
4455	supdrvLdrUnlock(pDevExt);
4456	return VINF_SUCCESS;
4457	}
4458	}
4459	/* (not found - add it!) */
4460
4461	/*
4462	* Allocate memory.
4463	*/
4464	Assert(cchName < sizeof(pImage->szName));
4465	pv = RTMemAlloc(sizeof(SUPDRVLDRIMAGE));
4466	if (!pv)
4467	{
4468	supdrvLdrUnlock(pDevExt);
4469	Log(("supdrvIOCtl_LdrOpen: RTMemAlloc() failed\n"));
4470	return /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_2;
4471	}
4472
4473	/*
4474	* Setup and link in the LDR stuff.
4475	*/
4476	pImage = (PSUPDRVLDRIMAGE)pv;
4477	pImage->pvImage = NULL;
4478	pImage->pvImageAlloc = NULL;
4479	pImage->cbImageWithTabs = pReq->u.In.cbImageWithTabs;
4480	pImage->cbImageBits = pReq->u.In.cbImageBits;
4481	pImage->cSymbols = 0;
4482	pImage->paSymbols = NULL;
4483	pImage->pachStrTab = NULL;
4484	pImage->cbStrTab = 0;
4485	pImage->pfnModuleInit = NULL;
4486	pImage->pfnModuleTerm = NULL;
4487	pImage->pfnServiceReqHandler = NULL;
4488	pImage->uState = SUP_IOCTL_LDR_OPEN;
4489	pImage->cUsage = 1;
4490	pImage->pDevExt = pDevExt;
4491	memcpy(pImage->szName, pReq->u.In.szName, cchName + 1);
4492
4493	/*
4494	* Try load it using the native loader, if that isn't supported, fall back
4495	* on the older method.
4496	*/
4497	pImage->fNative = true;
4498	rc = supdrvOSLdrOpen(pDevExt, pImage, pReq->u.In.szFilename);
4499	if (rc == VERR_NOT_SUPPORTED)
4500	{
4501	pImage->pvImageAlloc = RTMemExecAlloc(pImage->cbImageBits + 31);
4502	pImage->pvImage = RT_ALIGN_P(pImage->pvImageAlloc, 32);
4503	pImage->fNative = false;
4504	rc = pImage->pvImageAlloc ? VINF_SUCCESS : VERR_NO_EXEC_MEMORY;
4505	}
4506	if (RT_FAILURE(rc))
4507	{
4508	supdrvLdrUnlock(pDevExt);
4509	RTMemFree(pImage);
4510	Log(("supdrvIOCtl_LdrOpen(%s): failed - %Rrc\n", pReq->u.In.szName, rc));
4511	return rc;
4512	}
4513	Assert(VALID_PTR(pImage->pvImage) \|\| RT_FAILURE(rc));
4514
4515	/*
4516	* Link it.
4517	*/
4518	pImage->pNext = pDevExt->pLdrImages;
4519	pDevExt->pLdrImages = pImage;
4520
4521	supdrvLdrAddUsage(pSession, pImage);
4522
4523	pReq->u.Out.pvImageBase = pImage->pvImage;
4524	pReq->u.Out.fNeedsLoading = true;
4525	pReq->u.Out.fNativeLoader = pImage->fNative;
4526	supdrvOSLdrNotifyOpened(pDevExt, pImage);
4527
4528	supdrvLdrUnlock(pDevExt);
4529	return VINF_SUCCESS;
4530	}
4531
4532
4533	/**
4534	* Worker that validates a pointer to an image entrypoint.
4535	*
4536	* @returns IPRT status code.
4537	* @param pDevExt The device globals.
4538	* @param pImage The loader image.
4539	* @param pv The pointer into the image.
4540	* @param fMayBeNull Whether it may be NULL.
4541	* @param pszWhat What is this entrypoint? (for logging)
4542	* @param pbImageBits The image bits prepared by ring-3.
4543	*
4544	* @remarks Will leave the lock on failure.
4545	*/
4546	static int supdrvLdrValidatePointer(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage, void *pv,
4547	bool fMayBeNull, const uint8_t pbImageBits, const char pszWhat)
4548	{
4549	if (!fMayBeNull \|\| pv)
4550	{
4551	if ((uintptr_t)pv - (uintptr_t)pImage->pvImage >= pImage->cbImageBits)
4552	{
4553	supdrvLdrUnlock(pDevExt);
4554	Log(("Out of range (%p LB %#x): %s=%p\n", pImage->pvImage, pImage->cbImageBits, pszWhat, pv));
4555	return VERR_INVALID_PARAMETER;
4556	}
4557
4558	if (pImage->fNative)
4559	{
4560	int rc = supdrvOSLdrValidatePointer(pDevExt, pImage, pv, pbImageBits);
4561	if (RT_FAILURE(rc))
4562	{
4563	supdrvLdrUnlock(pDevExt);
4564	Log(("Bad entry point address: %s=%p (rc=%Rrc)\n", pszWhat, pv, rc));
4565	return rc;
4566	}
4567	}
4568	}
4569	return VINF_SUCCESS;
4570	}
4571
4572
4573	/**
4574	* Loads the image bits.
4575	*
4576	* This is the 2nd step of the loading.
4577	*
4578	* @returns IPRT status code.
4579	* @param pDevExt Device globals.
4580	* @param pSession Session data.
4581	* @param pReq The request.
4582	*/
4583	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
4584	{
4585	PSUPDRVLDRUSAGE pUsage;
4586	PSUPDRVLDRIMAGE pImage;
4587	int rc;
4588	LogFlow(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImageWithBits=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImageWithTabs));
4589
4590	/*
4591	* Find the ldr image.
4592	*/
4593	supdrvLdrLock(pDevExt);
4594	pUsage = pSession->pLdrUsage;
4595	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4596	pUsage = pUsage->pNext;
4597	if (!pUsage)
4598	{
4599	supdrvLdrUnlock(pDevExt);
4600	Log(("SUP_IOCTL_LDR_LOAD: couldn't find image!\n"));
4601	return VERR_INVALID_HANDLE;
4602	}
4603	pImage = pUsage->pImage;
4604
4605	/*
4606	* Validate input.
4607	*/
4608	if ( pImage->cbImageWithTabs != pReq->u.In.cbImageWithTabs
4609	\|\| pImage->cbImageBits != pReq->u.In.cbImageBits)
4610	{
4611	supdrvLdrUnlock(pDevExt);
4612	Log(("SUP_IOCTL_LDR_LOAD: image size mismatch!! %d(prep) != %d(load) or %d != %d\n",
4613	pImage->cbImageWithTabs, pReq->u.In.cbImageWithTabs, pImage->cbImageBits, pReq->u.In.cbImageBits));
4614	return VERR_INVALID_HANDLE;
4615	}
4616
4617	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
4618	{
4619	unsigned uState = pImage->uState;
4620	supdrvLdrUnlock(pDevExt);
4621	if (uState != SUP_IOCTL_LDR_LOAD)
4622	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
4623	return VERR_ALREADY_LOADED;
4624	}
4625
4626	switch (pReq->u.In.eEPType)
4627	{
4628	case SUPLDRLOADEP_NOTHING:
4629	break;
4630
4631	case SUPLDRLOADEP_VMMR0:
4632	rc = supdrvLdrValidatePointer( pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0, false, pReq->u.In.abImage, "pvVMMR0");
4633	if (RT_SUCCESS(rc))
4634	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt, false, pReq->u.In.abImage, "pvVMMR0EntryInt");
4635	if (RT_SUCCESS(rc))
4636	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, false, pReq->u.In.abImage, "pvVMMR0EntryFast");
4637	if (RT_SUCCESS(rc))
4638	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx, false, pReq->u.In.abImage, "pvVMMR0EntryEx");
4639	if (RT_FAILURE(rc))
4640	return rc;
4641	break;
4642
4643	case SUPLDRLOADEP_SERVICE:
4644	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.Service.pfnServiceReq, false, pReq->u.In.abImage, "pfnServiceReq");
4645	if (RT_FAILURE(rc))
4646	return rc;
4647	if ( pReq->u.In.EP.Service.apvReserved[0] != NIL_RTR0PTR
4648	\|\| pReq->u.In.EP.Service.apvReserved[1] != NIL_RTR0PTR
4649	\|\| pReq->u.In.EP.Service.apvReserved[2] != NIL_RTR0PTR)
4650	{
4651	supdrvLdrUnlock(pDevExt);
4652	Log(("Out of range (%p LB %#x): apvReserved={%p,%p,%p} MBZ!\n",
4653	pImage->pvImage, pReq->u.In.cbImageWithTabs,
4654	pReq->u.In.EP.Service.apvReserved[0],
4655	pReq->u.In.EP.Service.apvReserved[1],
4656	pReq->u.In.EP.Service.apvReserved[2]));
4657	return VERR_INVALID_PARAMETER;
4658	}
4659	break;
4660
4661	default:
4662	supdrvLdrUnlock(pDevExt);
4663	Log(("Invalid eEPType=%d\n", pReq->u.In.eEPType));
4664	return VERR_INVALID_PARAMETER;
4665	}
4666
4667	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleInit, true, pReq->u.In.abImage, "pfnModuleInit");
4668	if (RT_FAILURE(rc))
4669	return rc;
4670	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleTerm, true, pReq->u.In.abImage, "pfnModuleTerm");
4671	if (RT_FAILURE(rc))
4672	return rc;
4673
4674	/*
4675	* Allocate and copy the tables.
4676	* (No need to do try/except as this is a buffered request.)
4677	*/
4678	pImage->cbStrTab = pReq->u.In.cbStrTab;
4679	if (pImage->cbStrTab)
4680	{
4681	pImage->pachStrTab = (char *)RTMemAlloc(pImage->cbStrTab);
4682	if (pImage->pachStrTab)
4683	memcpy(pImage->pachStrTab, &pReq->u.In.abImage[pReq->u.In.offStrTab], pImage->cbStrTab);
4684	else
4685	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_3;
4686	}
4687
4688	pImage->cSymbols = pReq->u.In.cSymbols;
4689	if (RT_SUCCESS(rc) && pImage->cSymbols)
4690	{
4691	size_t cbSymbols = pImage->cSymbols * sizeof(SUPLDRSYM);
4692	pImage->paSymbols = (PSUPLDRSYM)RTMemAlloc(cbSymbols);
4693	if (pImage->paSymbols)
4694	memcpy(pImage->paSymbols, &pReq->u.In.abImage[pReq->u.In.offSymbols], cbSymbols);
4695	else
4696	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_4;
4697	}
4698
4699	/*
4700	* Copy the bits / complete native loading.
4701	*/
4702	if (RT_SUCCESS(rc))
4703	{
4704	pImage->uState = SUP_IOCTL_LDR_LOAD;
4705	pImage->pfnModuleInit = pReq->u.In.pfnModuleInit;
4706	pImage->pfnModuleTerm = pReq->u.In.pfnModuleTerm;
4707
4708	if (pImage->fNative)
4709	rc = supdrvOSLdrLoad(pDevExt, pImage, pReq->u.In.abImage, pReq);
4710	else
4711	{
4712	memcpy(pImage->pvImage, &pReq->u.In.abImage[0], pImage->cbImageBits);
4713	Log(("vboxdrv: Loaded '%s' at %p\n", pImage->szName, pImage->pvImage));
4714	}
4715	}
4716
4717	/*
4718	* Update any entry points.
4719	*/
4720	if (RT_SUCCESS(rc))
4721	{
4722	switch (pReq->u.In.eEPType)
4723	{
4724	default:
4725	case SUPLDRLOADEP_NOTHING:
4726	rc = VINF_SUCCESS;
4727	break;
4728	case SUPLDRLOADEP_VMMR0:
4729	rc = supdrvLdrSetVMMR0EPs(pDevExt, pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
4730	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
4731	break;
4732	case SUPLDRLOADEP_SERVICE:
4733	pImage->pfnServiceReqHandler = pReq->u.In.EP.Service.pfnServiceReq;
4734	rc = VINF_SUCCESS;
4735	break;
4736	}
4737	}
4738
4739	/*
4740	* On success call the module initialization.
4741	*/
4742	LogFlow(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
4743	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
4744	{
4745	Log(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
4746	pDevExt->pLdrInitImage = pImage;
4747	pDevExt->hLdrInitThread = RTThreadNativeSelf();
4748	rc = pImage->pfnModuleInit(pImage);
4749	pDevExt->pLdrInitImage = NULL;
4750	pDevExt->hLdrInitThread = NIL_RTNATIVETHREAD;
4751	if (RT_FAILURE(rc) && pDevExt->pvVMMR0 == pImage->pvImage)
4752	supdrvLdrUnsetVMMR0EPs(pDevExt);
4753	}
4754	SUPR0Printf("vboxdrv: %p %s\n", pImage->pvImage, pImage->szName);
4755
4756	if (RT_FAILURE(rc))
4757	{
4758	/* Inform the tracing component in case ModuleInit registered TPs. */
4759	supdrvTracerModuleUnloading(pDevExt, pImage);
4760
4761	pImage->uState = SUP_IOCTL_LDR_OPEN;
4762	pImage->pfnModuleInit = NULL;
4763	pImage->pfnModuleTerm = NULL;
4764	pImage->pfnServiceReqHandler= NULL;
4765	pImage->cbStrTab = 0;
4766	RTMemFree(pImage->pachStrTab);
4767	pImage->pachStrTab = NULL;
4768	RTMemFree(pImage->paSymbols);
4769	pImage->paSymbols = NULL;
4770	pImage->cSymbols = 0;
4771	}
4772
4773	supdrvLdrUnlock(pDevExt);
4774	return rc;
4775	}
4776
4777
4778	/**
4779	* Frees a previously loaded (prep'ed) image.
4780	*
4781	* @returns IPRT status code.
4782	* @param pDevExt Device globals.
4783	* @param pSession Session data.
4784	* @param pReq The request.
4785	*/
4786	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
4787	{
4788	int rc;
4789	PSUPDRVLDRUSAGE pUsagePrev;
4790	PSUPDRVLDRUSAGE pUsage;
4791	PSUPDRVLDRIMAGE pImage;
4792	LogFlow(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
4793
4794	/*
4795	* Find the ldr image.
4796	*/
4797	supdrvLdrLock(pDevExt);
4798	pUsagePrev = NULL;
4799	pUsage = pSession->pLdrUsage;
4800	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4801	{
4802	pUsagePrev = pUsage;
4803	pUsage = pUsage->pNext;
4804	}
4805	if (!pUsage)
4806	{
4807	supdrvLdrUnlock(pDevExt);
4808	Log(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
4809	return VERR_INVALID_HANDLE;
4810	}
4811
4812	/*
4813	* Check if we can remove anything.
4814	*/
4815	rc = VINF_SUCCESS;
4816	pImage = pUsage->pImage;
4817	if (pImage->cUsage <= 1 \|\| pUsage->cUsage <= 1)
4818	{
4819	/*
4820	* Check if there are any objects with destructors in the image, if
4821	* so leave it for the session cleanup routine so we get a chance to
4822	* clean things up in the right order and not leave them all dangling.
4823	*/
4824	RTSpinlockAcquire(pDevExt->Spinlock);
4825	if (pImage->cUsage <= 1)
4826	{
4827	PSUPDRVOBJ pObj;
4828	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4829	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4830	{
4831	rc = VERR_DANGLING_OBJECTS;
4832	break;
4833	}
4834	}
4835	else
4836	{
4837	PSUPDRVUSAGE pGenUsage;
4838	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
4839	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4840	{
4841	rc = VERR_DANGLING_OBJECTS;
4842	break;
4843	}
4844	}
4845	RTSpinlockRelease(pDevExt->Spinlock);
4846	if (rc == VINF_SUCCESS)
4847	{
4848	/* unlink it */
4849	if (pUsagePrev)
4850	pUsagePrev->pNext = pUsage->pNext;
4851	else
4852	pSession->pLdrUsage = pUsage->pNext;
4853
4854	/* free it */
4855	pUsage->pImage = NULL;
4856	pUsage->pNext = NULL;
4857	RTMemFree(pUsage);
4858
4859	/*
4860	* Dereference the image.
4861	*/
4862	if (pImage->cUsage <= 1)
4863	supdrvLdrFree(pDevExt, pImage);
4864	else
4865	pImage->cUsage--;
4866	}
4867	else
4868	{
4869	Log(("supdrvIOCtl_LdrFree: Dangling objects in %p/%s!\n", pImage->pvImage, pImage->szName));
4870	rc = VINF_SUCCESS; /** @todo BRANCH-2.1: remove this after branching. */
4871	}
4872	}
4873	else
4874	{
4875	/*
4876	* Dereference both image and usage.
4877	*/
4878	pImage->cUsage--;
4879	pUsage->cUsage--;
4880	}
4881
4882	supdrvLdrUnlock(pDevExt);
4883	return rc;
4884	}
4885
4886
4887	/**
4888	* Gets the address of a symbol in an open image.
4889	*
4890	* @returns IPRT status code.
4891	* @param pDevExt Device globals.
4892	* @param pSession Session data.
4893	* @param pReq The request buffer.
4894	*/
4895	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
4896	{
4897	PSUPDRVLDRIMAGE pImage;
4898	PSUPDRVLDRUSAGE pUsage;
4899	uint32_t i;
4900	PSUPLDRSYM paSyms;
4901	const char *pchStrings;
4902	const size_t cbSymbol = strlen(pReq->u.In.szSymbol) + 1;
4903	void *pvSymbol = NULL;
4904	int rc = VERR_GENERAL_FAILURE;
4905	Log3(("supdrvIOCtl_LdrGetSymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
4906
4907	/*
4908	* Find the ldr image.
4909	*/
4910	supdrvLdrLock(pDevExt);
4911	pUsage = pSession->pLdrUsage;
4912	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4913	pUsage = pUsage->pNext;
4914	if (!pUsage)
4915	{
4916	supdrvLdrUnlock(pDevExt);
4917	Log(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
4918	return VERR_INVALID_HANDLE;
4919	}
4920	pImage = pUsage->pImage;
4921	if (pImage->uState != SUP_IOCTL_LDR_LOAD)
4922	{
4923	unsigned uState = pImage->uState;
4924	supdrvLdrUnlock(pDevExt);
4925	Log(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", uState, uState)); NOREF(uState);
4926	return VERR_ALREADY_LOADED;
4927	}
4928
4929	/*
4930	* Search the symbol strings.
4931	*
4932	* Note! The int32_t is for native loading on solaris where the data
4933	* and text segments are in very different places.
4934	*/
4935	pchStrings = pImage->pachStrTab;
4936	paSyms = pImage->paSymbols;
4937	for (i = 0; i < pImage->cSymbols; i++)
4938	{
4939	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4940	&& !memcmp(pchStrings + paSyms[i].offName, pReq->u.In.szSymbol, cbSymbol))
4941	{
4942	pvSymbol = (uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol;
4943	rc = VINF_SUCCESS;
4944	break;
4945	}
4946	}
4947	supdrvLdrUnlock(pDevExt);
4948	pReq->u.Out.pvSymbol = pvSymbol;
4949	return rc;
4950	}
4951
4952
4953	/**
4954	* Gets the address of a symbol in an open image or the support driver.
4955	*
4956	* @returns VINF_SUCCESS on success.
4957	* @returns
4958	* @param pDevExt Device globals.
4959	* @param pSession Session data.
4960	* @param pReq The request buffer.
4961	*/
4962	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq)
4963	{
4964	int rc = VINF_SUCCESS;
4965	const char *pszSymbol = pReq->u.In.pszSymbol;
4966	const char *pszModule = pReq->u.In.pszModule;
4967	size_t cbSymbol;
4968	char const *pszEnd;
4969	uint32_t i;
4970
4971	/*
4972	* Input validation.
4973	*/
4974	AssertPtrReturn(pszSymbol, VERR_INVALID_POINTER);
4975	pszEnd = RTStrEnd(pszSymbol, 512);
4976	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4977	cbSymbol = pszEnd - pszSymbol + 1;
4978
4979	if (pszModule)
4980	{
4981	AssertPtrReturn(pszModule, VERR_INVALID_POINTER);
4982	pszEnd = RTStrEnd(pszModule, 64);
4983	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4984	}
4985	Log3(("supdrvIDC_LdrGetSymbol: pszModule=%p:{%s} pszSymbol=%p:{%s}\n", pszModule, pszModule, pszSymbol, pszSymbol));
4986
4987
4988	if ( !pszModule
4989	\|\| !strcmp(pszModule, "SupDrv"))
4990	{
4991	/*
4992	* Search the support driver export table.
4993	*/
4994	for (i = 0; i < RT_ELEMENTS(g_aFunctions); i++)
4995	if (!strcmp(g_aFunctions[i].szName, pszSymbol))
4996	{
4997	pReq->u.Out.pfnSymbol = g_aFunctions[i].pfn;
4998	break;
4999	}
5000	}
5001	else
5002	{
5003	/*
5004	* Find the loader image.
5005	*/
5006	PSUPDRVLDRIMAGE pImage;
5007
5008	supdrvLdrLock(pDevExt);
5009
5010	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
5011	if (!strcmp(pImage->szName, pszModule))
5012	break;
5013	if (pImage && pImage->uState == SUP_IOCTL_LDR_LOAD)
5014	{
5015	/*
5016	* Search the symbol strings.
5017	*/
5018	const char *pchStrings = pImage->pachStrTab;
5019	PCSUPLDRSYM paSyms = pImage->paSymbols;
5020	for (i = 0; i < pImage->cSymbols; i++)
5021	{
5022	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
5023	&& !memcmp(pchStrings + paSyms[i].offName, pszSymbol, cbSymbol))
5024	{
5025	/*
5026	* Found it! Calc the symbol address and add a reference to the module.
5027	*/
5028	pReq->u.Out.pfnSymbol = (PFNRT)((uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol);
5029	rc = supdrvLdrAddUsage(pSession, pImage);
5030	break;
5031	}
5032	}
5033	}
5034	else
5035	rc = pImage ? VERR_WRONG_ORDER : VERR_MODULE_NOT_FOUND;
5036
5037	supdrvLdrUnlock(pDevExt);
5038	}
5039	return rc;
5040	}
5041
5042
5043	/**
5044	* Updates the VMMR0 entry point pointers.
5045	*
5046	* @returns IPRT status code.
5047	* @param pDevExt Device globals.
5048	* @param pSession Session data.
5049	* @param pVMMR0 VMMR0 image handle.
5050	* @param pvVMMR0EntryInt VMMR0EntryInt address.
5051	* @param pvVMMR0EntryFast VMMR0EntryFast address.
5052	* @param pvVMMR0EntryEx VMMR0EntryEx address.
5053	* @remark Caller must own the loader mutex.
5054	*/
5055	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx)
5056	{
5057	int rc = VINF_SUCCESS;
5058	LogFlow(("supdrvLdrSetR0EP pvVMMR0=%p pvVMMR0EntryInt=%p\n", pvVMMR0, pvVMMR0EntryInt));
5059
5060
5061	/*
5062	* Check if not yet set.
5063	*/
5064	if (!pDevExt->pvVMMR0)
5065	{
5066	pDevExt->pvVMMR0 = pvVMMR0;
5067	pDevExt->pfnVMMR0EntryInt = pvVMMR0EntryInt;
5068	pDevExt->pfnVMMR0EntryFast = pvVMMR0EntryFast;
5069	pDevExt->pfnVMMR0EntryEx = pvVMMR0EntryEx;
5070	}
5071	else
5072	{
5073	/*
5074	* Return failure or success depending on whether the values match or not.
5075	*/
5076	if ( pDevExt->pvVMMR0 != pvVMMR0
5077	\|\| (void *)pDevExt->pfnVMMR0EntryInt != pvVMMR0EntryInt
5078	\|\| (void *)pDevExt->pfnVMMR0EntryFast != pvVMMR0EntryFast
5079	\|\| (void *)pDevExt->pfnVMMR0EntryEx != pvVMMR0EntryEx)
5080	{
5081	AssertMsgFailed(("SUP_IOCTL_LDR_SETR0EP: Already set pointing to a different module!\n"));
5082	rc = VERR_INVALID_PARAMETER;
5083	}
5084	}
5085	return rc;
5086	}
5087
5088
5089	/**
5090	* Unsets the VMMR0 entry point installed by supdrvLdrSetR0EP.
5091	*
5092	* @param pDevExt Device globals.
5093	*/
5094	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt)
5095	{
5096	pDevExt->pvVMMR0 = NULL;
5097	pDevExt->pfnVMMR0EntryInt = NULL;
5098	pDevExt->pfnVMMR0EntryFast = NULL;
5099	pDevExt->pfnVMMR0EntryEx = NULL;
5100	}
5101
5102
5103	/**
5104	* Adds a usage reference in the specified session of an image.
5105	*
5106	* Called while owning the loader semaphore.
5107	*
5108	* @returns VINF_SUCCESS on success and VERR_NO_MEMORY on failure.
5109	* @param pSession Session in question.
5110	* @param pImage Image which the session is using.
5111	*/
5112	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage)
5113	{
5114	PSUPDRVLDRUSAGE pUsage;
5115	LogFlow(("supdrvLdrAddUsage: pImage=%p\n", pImage));
5116
5117	/*
5118	* Referenced it already?
5119	*/
5120	pUsage = pSession->pLdrUsage;
5121	while (pUsage)
5122	{
5123	if (pUsage->pImage == pImage)
5124	{
5125	pUsage->cUsage++;
5126	return VINF_SUCCESS;
5127	}
5128	pUsage = pUsage->pNext;
5129	}
5130
5131	/*
5132	* Allocate new usage record.
5133	*/
5134	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
5135	AssertReturn(pUsage, /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_5);
5136	pUsage->cUsage = 1;
5137	pUsage->pImage = pImage;
5138	pUsage->pNext = pSession->pLdrUsage;
5139	pSession->pLdrUsage = pUsage;
5140	return VINF_SUCCESS;
5141	}
5142
5143
5144	/**
5145	* Frees a load image.
5146	*
5147	* @param pDevExt Pointer to device extension.
5148	* @param pImage Pointer to the image we're gonna free.
5149	* This image must exit!
5150	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
5151	*/
5152	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
5153	{
5154	PSUPDRVLDRIMAGE pImagePrev;
5155	LogFlow(("supdrvLdrFree: pImage=%p\n", pImage));
5156
5157	/* find it - arg. should've used doubly linked list. */
5158	Assert(pDevExt->pLdrImages);
5159	pImagePrev = NULL;
5160	if (pDevExt->pLdrImages != pImage)
5161	{
5162	pImagePrev = pDevExt->pLdrImages;
5163	while (pImagePrev->pNext != pImage)
5164	pImagePrev = pImagePrev->pNext;
5165	Assert(pImagePrev->pNext == pImage);
5166	}
5167
5168	/* unlink */
5169	if (pImagePrev)
5170	pImagePrev->pNext = pImage->pNext;
5171	else
5172	pDevExt->pLdrImages = pImage->pNext;
5173
5174	/* check if this is VMMR0.r0 unset its entry point pointers. */
5175	if (pDevExt->pvVMMR0 == pImage->pvImage)
5176	supdrvLdrUnsetVMMR0EPs(pDevExt);
5177
5178	/* check for objects with destructors in this image. (Shouldn't happen.) */
5179	if (pDevExt->pObjs)
5180	{
5181	unsigned cObjs = 0;
5182	PSUPDRVOBJ pObj;
5183	RTSpinlockAcquire(pDevExt->Spinlock);
5184	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
5185	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
5186	{
5187	pObj->pfnDestructor = NULL;
5188	cObjs++;
5189	}
5190	RTSpinlockRelease(pDevExt->Spinlock);
5191	if (cObjs)
5192	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
5193	}
5194
5195	/* call termination function if fully loaded. */
5196	if ( pImage->pfnModuleTerm
5197	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
5198	{
5199	LogFlow(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
5200	pImage->pfnModuleTerm(pImage);
5201	}
5202
5203	/* Inform the tracing component. */
5204	supdrvTracerModuleUnloading(pDevExt, pImage);
5205
5206	/* do native unload if appropriate. */
5207	if (pImage->fNative)
5208	supdrvOSLdrUnload(pDevExt, pImage);
5209
5210	/* free the image */
5211	pImage->cUsage = 0;
5212	pImage->pDevExt = NULL;
5213	pImage->pNext = NULL;
5214	pImage->uState = SUP_IOCTL_LDR_FREE;
5215	RTMemExecFree(pImage->pvImageAlloc, pImage->cbImageBits + 31);
5216	pImage->pvImageAlloc = NULL;
5217	RTMemFree(pImage->pachStrTab);
5218	pImage->pachStrTab = NULL;
5219	RTMemFree(pImage->paSymbols);
5220	pImage->paSymbols = NULL;
5221	RTMemFree(pImage);
5222	}
5223
5224
5225	/**
5226	* Acquires the loader lock.
5227	*
5228	* @returns IPRT status code.
5229	* @param pDevExt The device extension.
5230	*/
5231	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt)
5232	{
5233	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
5234	int rc = RTSemMutexRequest(pDevExt->mtxLdr, RT_INDEFINITE_WAIT);
5235	#else
5236	int rc = RTSemFastMutexRequest(pDevExt->mtxLdr);
5237	#endif
5238	AssertRC(rc);
5239	return rc;
5240	}
5241
5242
5243	/**
5244	* Releases the loader lock.
5245	*
5246	* @returns IPRT status code.
5247	* @param pDevExt The device extension.
5248	*/
5249	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt)
5250	{
5251	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
5252	return RTSemMutexRelease(pDevExt->mtxLdr);
5253	#else
5254	return RTSemFastMutexRelease(pDevExt->mtxLdr);
5255	#endif
5256	}
5257
5258
5259	/**
5260	* Implements the service call request.
5261	*
5262	* @returns VBox status code.
5263	* @param pDevExt The device extension.
5264	* @param pSession The calling session.
5265	* @param pReq The request packet, valid.
5266	*/
5267	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq)
5268	{
5269	#if !defined(RT_OS_WINDOWS) \|\| defined(RT_ARCH_AMD64) \|\| defined(DEBUG)
5270	int rc;
5271
5272	/*
5273	* Find the module first in the module referenced by the calling session.
5274	*/
5275	rc = supdrvLdrLock(pDevExt);
5276	if (RT_SUCCESS(rc))
5277	{
5278	PFNSUPR0SERVICEREQHANDLER pfnServiceReqHandler = NULL;
5279	PSUPDRVLDRUSAGE pUsage;
5280
5281	for (pUsage = pSession->pLdrUsage; pUsage; pUsage = pUsage->pNext)
5282	if ( pUsage->pImage->pfnServiceReqHandler
5283	&& !strcmp(pUsage->pImage->szName, pReq->u.In.szName))
5284	{
5285	pfnServiceReqHandler = pUsage->pImage->pfnServiceReqHandler;
5286	break;
5287	}
5288	supdrvLdrUnlock(pDevExt);
5289
5290	if (pfnServiceReqHandler)
5291	{
5292	/*
5293	* Call it.
5294	*/
5295	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
5296	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, NULL);
5297	else
5298	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0]);
5299	}
5300	else
5301	rc = VERR_SUPDRV_SERVICE_NOT_FOUND;
5302	}
5303
5304	/* log it */
5305	if ( RT_FAILURE(rc)
5306	&& rc != VERR_INTERRUPTED
5307	&& rc != VERR_TIMEOUT)
5308	Log(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
5309	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
5310	else
5311	Log4(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
5312	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
5313	return rc;
5314	#else /* RT_OS_WINDOWS && !RT_ARCH_AMD64 && !DEBUG */
5315	return VERR_NOT_IMPLEMENTED;
5316	#endif /* RT_OS_WINDOWS && !RT_ARCH_AMD64 && !DEBUG */
5317	}
5318
5319
5320	/**
5321	* Implements the logger settings request.
5322	*
5323	* @returns VBox status code.
5324	* @param pDevExt The device extension.
5325	* @param pSession The caller's session.
5326	* @param pReq The request.
5327	*/
5328	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq)
5329	{
5330	const char *pszGroup = &pReq->u.In.szStrings[pReq->u.In.offGroups];
5331	const char *pszFlags = &pReq->u.In.szStrings[pReq->u.In.offFlags];
5332	const char *pszDest = &pReq->u.In.szStrings[pReq->u.In.offDestination];
5333	PRTLOGGER pLogger = NULL;
5334	int rc;
5335
5336	/*
5337	* Some further validation.
5338	*/
5339	switch (pReq->u.In.fWhat)
5340	{
5341	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
5342	case SUPLOGGERSETTINGS_WHAT_CREATE:
5343	break;
5344
5345	case SUPLOGGERSETTINGS_WHAT_DESTROY:
5346	if (pszGroup \|\| pszFlags \|\| *pszDest)
5347	return VERR_INVALID_PARAMETER;
5348	if (pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_RELEASE)
5349	return VERR_ACCESS_DENIED;
5350	break;
5351
5352	default:
5353	return VERR_INTERNAL_ERROR;
5354	}
5355
5356	/*
5357	* Get the logger.
5358	*/
5359	switch (pReq->u.In.fWhich)
5360	{
5361	case SUPLOGGERSETTINGS_WHICH_DEBUG:
5362	pLogger = RTLogGetDefaultInstance();
5363	break;
5364
5365	case SUPLOGGERSETTINGS_WHICH_RELEASE:
5366	pLogger = RTLogRelDefaultInstance();
5367	break;
5368
5369	default:
5370	return VERR_INTERNAL_ERROR;
5371	}
5372
5373	/*
5374	* Do the job.
5375	*/
5376	switch (pReq->u.In.fWhat)
5377	{
5378	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
5379	if (pLogger)
5380	{
5381	rc = RTLogFlags(pLogger, pszFlags);
5382	if (RT_SUCCESS(rc))
5383	rc = RTLogGroupSettings(pLogger, pszGroup);
5384	NOREF(pszDest);
5385	}
5386	else
5387	rc = VERR_NOT_FOUND;
5388	break;
5389
5390	case SUPLOGGERSETTINGS_WHAT_CREATE:
5391	{
5392	if (pLogger)
5393	rc = VERR_ALREADY_EXISTS;
5394	else
5395	{
5396	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
5397
5398	rc = RTLogCreate(&pLogger,
5399	0 /* fFlags */,
5400	pszGroup,
5401	pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_DEBUG
5402	? "VBOX_LOG"
5403	: "VBOX_RELEASE_LOG",
5404	RT_ELEMENTS(s_apszGroups),
5405	s_apszGroups,
5406	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER,
5407	NULL);
5408	if (RT_SUCCESS(rc))
5409	{
5410	rc = RTLogFlags(pLogger, pszFlags);
5411	NOREF(pszDest);
5412	if (RT_SUCCESS(rc))
5413	{
5414	switch (pReq->u.In.fWhich)
5415	{
5416	case SUPLOGGERSETTINGS_WHICH_DEBUG:
5417	pLogger = RTLogSetDefaultInstance(pLogger);
5418	break;
5419	case SUPLOGGERSETTINGS_WHICH_RELEASE:
5420	pLogger = RTLogRelSetDefaultInstance(pLogger);
5421	break;
5422	}
5423	}
5424	RTLogDestroy(pLogger);
5425	}
5426	}
5427	break;
5428	}
5429
5430	case SUPLOGGERSETTINGS_WHAT_DESTROY:
5431	switch (pReq->u.In.fWhich)
5432	{
5433	case SUPLOGGERSETTINGS_WHICH_DEBUG:
5434	pLogger = RTLogSetDefaultInstance(NULL);
5435	break;
5436	case SUPLOGGERSETTINGS_WHICH_RELEASE:
5437	pLogger = RTLogRelSetDefaultInstance(NULL);
5438	break;
5439	}
5440	rc = RTLogDestroy(pLogger);
5441	break;
5442
5443	default:
5444	{
5445	rc = VERR_INTERNAL_ERROR;
5446	break;
5447	}
5448	}
5449
5450	return rc;
5451	}
5452
5453
5454	/**
5455	* Implements the MSR prober operations.
5456	*
5457	* @returns VBox status code.
5458	* @param pDevExt The device extension.
5459	* @param pReq The request.
5460	*/
5461	static int supdrvIOCtl_MsrProber(PSUPDRVDEVEXT pDevExt, PSUPMSRPROBER pReq)
5462	{
5463	#ifdef SUPDRV_WITH_MSR_PROBER
5464	RTCPUID const idCpu = pReq->u.In.idCpu == UINT32_MAX ? NIL_RTCPUID : pReq->u.In.idCpu;
5465	int rc;
5466
5467	switch (pReq->u.In.enmOp)
5468	{
5469	case SUPMSRPROBEROP_READ:
5470	{
5471	uint64_t uValue;
5472	rc = supdrvOSMsrProberRead(pReq->u.In.uMsr, idCpu, &uValue);
5473	if (RT_SUCCESS(rc))
5474	{
5475	pReq->u.Out.uResults.Read.uValue = uValue;
5476	pReq->u.Out.uResults.Read.fGp = false;
5477	}
5478	else if (rc == VERR_ACCESS_DENIED)
5479	{
5480	pReq->u.Out.uResults.Read.uValue = 0;
5481	pReq->u.Out.uResults.Read.fGp = true;
5482	rc = VINF_SUCCESS;
5483	}
5484	break;
5485	}
5486
5487	case SUPMSRPROBEROP_WRITE:
5488	rc = supdrvOSMsrProberWrite(pReq->u.In.uMsr, idCpu, pReq->u.In.uArgs.Write.uToWrite);
5489	if (RT_SUCCESS(rc))
5490	pReq->u.Out.uResults.Write.fGp = false;
5491	else if (rc == VERR_ACCESS_DENIED)
5492	{
5493	pReq->u.Out.uResults.Write.fGp = true;
5494	rc = VINF_SUCCESS;
5495	}
5496	break;
5497
5498	case SUPMSRPROBEROP_MODIFY:
5499	case SUPMSRPROBEROP_MODIFY_FASTER:
5500	rc = supdrvOSMsrProberModify(idCpu, pReq);
5501	break;
5502
5503	default:
5504	return VERR_INVALID_FUNCTION;
5505	}
5506	return rc;
5507	#else
5508	return VERR_NOT_IMPLEMENTED;
5509	#endif
5510	}
5511
5512
5513
5514	/**
5515	* Creates the GIP.
5516	*
5517	* @returns VBox status code.
5518	* @param pDevExt Instance data. GIP stuff may be updated.
5519	*/
5520	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
5521	{
5522	PSUPGLOBALINFOPAGE pGip;
5523	RTHCPHYS HCPhysGip;
5524	uint32_t u32SystemResolution;
5525	uint32_t u32Interval;
5526	unsigned cCpus;
5527	int rc;
5528
5529
5530	LogFlow(("supdrvGipCreate:\n"));
5531
5532	/* assert order */
5533	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
5534	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
5535	Assert(!pDevExt->pGipTimer);
5536
5537	/*
5538	* Check the CPU count.
5539	*/
5540	cCpus = RTMpGetArraySize();
5541	if ( cCpus > RTCPUSET_MAX_CPUS
5542	\|\| cCpus > 256 /ApicId is used for the mappings/)
5543	{
5544	SUPR0Printf("VBoxDrv: Too many CPUs (%u) for the GIP (max %u)\n", cCpus, RT_MIN(RTCPUSET_MAX_CPUS, 256));
5545	return VERR_TOO_MANY_CPUS;
5546	}
5547
5548	/*
5549	* Allocate a contiguous set of pages with a default kernel mapping.
5550	*/
5551	rc = RTR0MemObjAllocCont(&pDevExt->GipMemObj, RT_UOFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), false /fExecutable/);
5552	if (RT_FAILURE(rc))
5553	{
5554	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
5555	return rc;
5556	}
5557	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
5558	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
5559
5560	/*
5561	* Find a reasonable update interval and initialize the structure.
5562	*/
5563	u32Interval = u32SystemResolution = RTTimerGetSystemGranularity();
5564	while (u32Interval < 10000000 /* 10 ms */)
5565	u32Interval += u32SystemResolution;
5566
5567	supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), 1000000000 / u32Interval /=Hz/, cCpus);
5568
5569	/*
5570	* Create the timer.
5571	* If CPU_ALL isn't supported we'll have to fall back to synchronous mode.
5572	*/
5573	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
5574	{
5575	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, RTTIMER_FLAGS_CPU_ALL, supdrvGipAsyncTimer, pDevExt);
5576	if (rc == VERR_NOT_SUPPORTED)
5577	{
5578	OSDBGPRINT(("supdrvGipCreate: omni timer not supported, falling back to synchronous mode\n"));
5579	pGip->u32Mode = SUPGIPMODE_SYNC_TSC;
5580	}
5581	}
5582	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
5583	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0 /* fFlags */, supdrvGipSyncTimer, pDevExt);
5584	if (RT_SUCCESS(rc))
5585	{
5586	rc = RTMpNotificationRegister(supdrvGipMpEvent, pDevExt);
5587	if (RT_SUCCESS(rc))
5588	{
5589	rc = RTMpOnAll(supdrvGipInitOnCpu, pDevExt, pGip);
5590	if (RT_SUCCESS(rc))
5591	{
5592	/*
5593	* We're good.
5594	*/
5595	Log(("supdrvGipCreate: %u ns interval.\n", u32Interval));
5596	g_pSUPGlobalInfoPage = pGip;
5597	return VINF_SUCCESS;
5598	}
5599
5600	OSDBGPRINT(("supdrvGipCreate: RTMpOnAll failed with rc=%Rrc\n", rc));
5601	RTMpNotificationDeregister(supdrvGipMpEvent, pDevExt);
5602
5603	}
5604	else
5605	OSDBGPRINT(("supdrvGipCreate: failed to register MP event notfication. rc=%Rrc\n", rc));
5606	}
5607	else
5608	{
5609	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %u ns interval. rc=%Rrc\n", u32Interval, rc));
5610	Assert(!pDevExt->pGipTimer);
5611	}
5612	supdrvGipDestroy(pDevExt);
5613	return rc;
5614	}
5615
5616
5617	/**
5618	* Terminates the GIP.
5619	*
5620	* @param pDevExt Instance data. GIP stuff may be updated.
5621	*/
5622	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
5623	{
5624	int rc;
5625	#ifdef DEBUG_DARWIN_GIP
5626	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
5627	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
5628	pDevExt->pGipTimer, pDevExt->GipMemObj));
5629	#endif
5630
5631	/*
5632	* Invalid the GIP data.
5633	*/
5634	if (pDevExt->pGip)
5635	{
5636	supdrvGipTerm(pDevExt->pGip);
5637	pDevExt->pGip = NULL;
5638	}
5639	g_pSUPGlobalInfoPage = NULL;
5640
5641	/*
5642	* Destroy the timer and free the GIP memory object.
5643	*/
5644	if (pDevExt->pGipTimer)
5645	{
5646	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
5647	pDevExt->pGipTimer = NULL;
5648	}
5649
5650	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
5651	{
5652	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
5653	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
5654	}
5655
5656	/*
5657	* Finally, make sure we've release the system timer resolution request
5658	* if one actually succeeded and is still pending.
5659	*/
5660	if (pDevExt->u32SystemTimerGranularityGrant)
5661	{
5662	rc = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
5663	pDevExt->u32SystemTimerGranularityGrant = 0;
5664	}
5665	}
5666
5667
5668	/**
5669	* Timer callback function sync GIP mode.
5670	* @param pTimer The timer.
5671	* @param pvUser The device extension.
5672	*/
5673	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
5674	{
5675	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
5676	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5677	uint64_t u64TSC = ASMReadTSC();
5678	uint64_t NanoTS = RTTimeSystemNanoTS();
5679
5680	supdrvGipUpdate(pDevExt, NanoTS, u64TSC, NIL_RTCPUID, iTick);
5681
5682	ASMSetFlags(fOldFlags);
5683	}
5684
5685
5686	/**
5687	* Timer callback function for async GIP mode.
5688	* @param pTimer The timer.
5689	* @param pvUser The device extension.
5690	*/
5691	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
5692	{
5693	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
5694	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5695	RTCPUID idCpu = RTMpCpuId();
5696	uint64_t u64TSC = ASMReadTSC();
5697	uint64_t NanoTS = RTTimeSystemNanoTS();
5698
5699	/** @todo reset the transaction number and whatnot when iTick == 1. */
5700	if (pDevExt->idGipMaster == idCpu)
5701	supdrvGipUpdate(pDevExt, NanoTS, u64TSC, idCpu, iTick);
5702	else
5703	supdrvGipUpdatePerCpu(pDevExt, NanoTS, u64TSC, idCpu, ASMGetApicId(), iTick);
5704
5705	ASMSetFlags(fOldFlags);
5706	}
5707
5708
5709	/**
5710	* Finds our (@a idCpu) entry, or allocates a new one if not found.
5711	*
5712	* @returns Index of the CPU in the cache set.
5713	* @param pGip The GIP.
5714	* @param idCpu The CPU ID.
5715	*/
5716	static uint32_t supdrvGipCpuIndexFromCpuId(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
5717	{
5718	uint32_t i, cTries;
5719
5720	/*
5721	* ASSUMES that CPU IDs are constant.
5722	*/
5723	for (i = 0; i < pGip->cCpus; i++)
5724	if (pGip->aCPUs[i].idCpu == idCpu)
5725	return i;
5726
5727	cTries = 0;
5728	do
5729	{
5730	for (i = 0; i < pGip->cCpus; i++)
5731	{
5732	bool fRc;
5733	ASMAtomicCmpXchgSize(&pGip->aCPUs[i].idCpu, idCpu, NIL_RTCPUID, fRc);
5734	if (fRc)
5735	return i;
5736	}
5737	} while (cTries++ < 32);
5738	AssertReleaseFailed();
5739	return i - 1;
5740	}
5741
5742
5743	/**
5744	* The calling CPU should be accounted as online, update GIP accordingly.
5745	*
5746	* This is used by supdrvGipMpEvent as well as the supdrvGipCreate.
5747	*
5748	* @param pDevExt The device extension.
5749	* @param idCpu The CPU ID.
5750	*/
5751	static void supdrvGipMpEventOnline(PSUPDRVDEVEXT pDevExt, RTCPUID idCpu)
5752	{
5753	int iCpuSet = 0;
5754	uint16_t idApic = UINT16_MAX;
5755	uint32_t i = 0;
5756	uint64_t u64NanoTS = 0;
5757	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5758
5759	AssertPtrReturnVoid(pGip);
5760	AssertRelease(idCpu == RTMpCpuId());
5761	Assert(pGip->cPossibleCpus == RTMpGetCount());
5762
5763	/*
5764	* Do this behind a spinlock with interrupts disabled as this can fire
5765	* on all CPUs simultaneously, see @bugref{6110}.
5766	*/
5767	RTSpinlockAcquire(pDevExt->hGipSpinlock);
5768
5769	/*
5770	* Update the globals.
5771	*/
5772	ASMAtomicWriteU16(&pGip->cPresentCpus, RTMpGetPresentCount());
5773	ASMAtomicWriteU16(&pGip->cOnlineCpus, RTMpGetOnlineCount());
5774	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5775	if (iCpuSet >= 0)
5776	{
5777	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5778	RTCpuSetAddByIndex(&pGip->OnlineCpuSet, iCpuSet);
5779	RTCpuSetAddByIndex(&pGip->PresentCpuSet, iCpuSet);
5780	}
5781
5782	/*
5783	* Update the entry.
5784	*/
5785	u64NanoTS = RTTimeSystemNanoTS() - pGip->u32UpdateIntervalNS;
5786	i = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
5787	supdrvGipInitCpu(pGip, &pGip->aCPUs[i], u64NanoTS);
5788	idApic = ASMGetApicId();
5789	ASMAtomicWriteU16(&pGip->aCPUs[i].idApic, idApic);
5790	ASMAtomicWriteS16(&pGip->aCPUs[i].iCpuSet, (int16_t)iCpuSet);
5791	ASMAtomicWriteSize(&pGip->aCPUs[i].idCpu, idCpu);
5792
5793	/*
5794	* Update the APIC ID and CPU set index mappings.
5795	*/
5796	ASMAtomicWriteU16(&pGip->aiCpuFromApicId[idApic], i);
5797	ASMAtomicWriteU16(&pGip->aiCpuFromCpuSetIdx[iCpuSet], i);
5798
5799	/* commit it */
5800	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_ONLINE);
5801
5802	RTSpinlockReleaseNoInts(pDevExt->hGipSpinlock);
5803	}
5804
5805
5806	/**
5807	* The CPU should be accounted as offline, update the GIP accordingly.
5808	*
5809	* This is used by supdrvGipMpEvent.
5810	*
5811	* @param pDevExt The device extension.
5812	* @param idCpu The CPU ID.
5813	*/
5814	static void supdrvGipMpEventOffline(PSUPDRVDEVEXT pDevExt, RTCPUID idCpu)
5815	{
5816	int iCpuSet;
5817	unsigned i;
5818
5819	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5820
5821	AssertPtrReturnVoid(pGip);
5822	RTSpinlockAcquire(pDevExt->hGipSpinlock);
5823
5824	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5825	AssertReturnVoid(iCpuSet >= 0);
5826
5827	i = pGip->aiCpuFromCpuSetIdx[iCpuSet];
5828	AssertReturnVoid(i < pGip->cCpus);
5829	AssertReturnVoid(pGip->aCPUs[i].idCpu == idCpu);
5830
5831	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5832	RTCpuSetDelByIndex(&pGip->OnlineCpuSet, iCpuSet);
5833
5834	/* commit it */
5835	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_OFFLINE);
5836
5837	RTSpinlockReleaseNoInts(pDevExt->hGipSpinlock);
5838	}
5839
5840
5841	/**
5842	* Multiprocessor event notification callback.
5843	*
5844	* This is used to make sure that the GIP master gets passed on to
5845	* another CPU. It also updates the associated CPU data.
5846	*
5847	* @param enmEvent The event.
5848	* @param idCpu The cpu it applies to.
5849	* @param pvUser Pointer to the device extension.
5850	*
5851	* @remarks This function -must- fire on the newly online'd CPU for the
5852	* RTMPEVENT_ONLINE case and can fire on any CPU for the
5853	* RTMPEVENT_OFFLINE case.
5854	*/
5855	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser)
5856	{
5857	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5858	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5859
5860	AssertRelease(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
5861
5862	/*
5863	* Update the GIP CPU data.
5864	*/
5865	if (pGip)
5866	{
5867	switch (enmEvent)
5868	{
5869	case RTMPEVENT_ONLINE:
5870	AssertRelease(idCpu == RTMpCpuId());
5871	supdrvGipMpEventOnline(pDevExt, idCpu);
5872	break;
5873	case RTMPEVENT_OFFLINE:
5874	supdrvGipMpEventOffline(pDevExt, idCpu);
5875	break;
5876
5877	}
5878	}
5879
5880	/*
5881	* Make sure there is a master GIP.
5882	*/
5883	if (enmEvent == RTMPEVENT_OFFLINE)
5884	{
5885	RTCPUID idGipMaster = ASMAtomicReadU32(&pDevExt->idGipMaster);
5886	if (idGipMaster == idCpu)
5887	{
5888	/*
5889	* Find a new GIP master.
5890	*/
5891	bool fIgnored;
5892	unsigned i;
5893	RTCPUID idNewGipMaster = NIL_RTCPUID;
5894	RTCPUSET OnlineCpus;
5895	RTMpGetOnlineSet(&OnlineCpus);
5896
5897	for (i = 0; i < RTCPUSET_MAX_CPUS; i++)
5898	{
5899	RTCPUID idCurCpu = RTMpCpuIdFromSetIndex(i);
5900	if ( RTCpuSetIsMember(&OnlineCpus, idCurCpu)
5901	&& idCurCpu != idGipMaster)
5902	{
5903	idNewGipMaster = idCurCpu;
5904	break;
5905	}
5906	}
5907
5908	Log(("supdrvGipMpEvent: Gip master %#lx -> %#lx\n", (long)idGipMaster, (long)idNewGipMaster));
5909	ASMAtomicCmpXchgSize(&pDevExt->idGipMaster, idNewGipMaster, idGipMaster, fIgnored);
5910	NOREF(fIgnored);
5911	}
5912	}
5913	}
5914
5915
5916	/**
5917	* Callback used by supdrvDetermineAsyncTSC to read the TSC on a CPU.
5918	*
5919	* @param idCpu Ignored.
5920	* @param pvUser1 Where to put the TSC.
5921	* @param pvUser2 Ignored.
5922	*/
5923	static DECLCALLBACK(void) supdrvDetermineAsyncTscWorker(RTCPUID idCpu, void pvUser1, void pvUser2)
5924	{
5925	#if 1
5926	ASMAtomicWriteU64((uint64_t volatile *)pvUser1, ASMReadTSC());
5927	#else
5928	(uint64_t )pvUser1 = ASMReadTSC();
5929	#endif
5930	}
5931
5932
5933	/**
5934	* Determine if Async GIP mode is required because of TSC drift.
5935	*
5936	* When using the default/normal timer code it is essential that the time stamp counter
5937	* (TSC) runs never backwards, that is, a read operation to the counter should return
5938	* a bigger value than any previous read operation. This is guaranteed by the latest
5939	* AMD CPUs and by newer Intel CPUs which never enter the C2 state (P4). In any other
5940	* case we have to choose the asynchronous timer mode.
5941	*
5942	* @param poffMin Pointer to the determined difference between different cores.
5943	* @return false if the time stamp counters appear to be synchronized, true otherwise.
5944	*/
5945	static bool supdrvDetermineAsyncTsc(uint64_t *poffMin)
5946	{
5947	/*
5948	* Just iterate all the cpus 8 times and make sure that the TSC is
5949	* ever increasing. We don't bother taking TSC rollover into account.
5950	*/
5951	int iEndCpu = RTMpGetArraySize();
5952	int iCpu;
5953	int cLoops = 8;
5954	bool fAsync = false;
5955	int rc = VINF_SUCCESS;
5956	uint64_t offMax = 0;
5957	uint64_t offMin = ~(uint64_t)0;
5958	uint64_t PrevTsc = ASMReadTSC();
5959
5960	while (cLoops-- > 0)
5961	{
5962	for (iCpu = 0; iCpu < iEndCpu; iCpu++)
5963	{
5964	uint64_t CurTsc;
5965	rc = RTMpOnSpecific(RTMpCpuIdFromSetIndex(iCpu), supdrvDetermineAsyncTscWorker, &CurTsc, NULL);
5966	if (RT_SUCCESS(rc))
5967	{
5968	if (CurTsc <= PrevTsc)
5969	{
5970	fAsync = true;
5971	offMin = offMax = PrevTsc - CurTsc;
5972	Log(("supdrvDetermineAsyncTsc: iCpu=%d cLoops=%d CurTsc=%llx PrevTsc=%llx\n",
5973	iCpu, cLoops, CurTsc, PrevTsc));
5974	break;
5975	}
5976
5977	/* Gather statistics (except the first time). */
5978	if (iCpu != 0 \|\| cLoops != 7)
5979	{
5980	uint64_t off = CurTsc - PrevTsc;
5981	if (off < offMin)
5982	offMin = off;
5983	if (off > offMax)
5984	offMax = off;
5985	Log2(("%d/%d: off=%llx\n", cLoops, iCpu, off));
5986	}
5987
5988	/* Next */
5989	PrevTsc = CurTsc;
5990	}
5991	else if (rc == VERR_NOT_SUPPORTED)
5992	break;
5993	else
5994	AssertMsg(rc == VERR_CPU_NOT_FOUND \|\| rc == VERR_CPU_OFFLINE, ("%d\n", rc));
5995	}
5996
5997	/* broke out of the loop. */
5998	if (iCpu < iEndCpu)
5999	break;
6000	}
6001
6002	poffMin = offMin; / Almost RTMpOnSpecific profiling. */
6003	Log(("supdrvDetermineAsyncTsc: returns %d; iEndCpu=%d rc=%d offMin=%llx offMax=%llx\n",
6004	fAsync, iEndCpu, rc, offMin, offMax));
6005	#if !defined(RT_OS_SOLARIS) && !defined(RT_OS_OS2) && !defined(RT_OS_WINDOWS)
6006	OSDBGPRINT(("vboxdrv: fAsync=%d offMin=%#lx offMax=%#lx\n", fAsync, (long)offMin, (long)offMax));
6007	#endif
6008	return fAsync;
6009	}
6010
6011
6012	/**
6013	* Determine the GIP TSC mode.
6014	*
6015	* @returns The most suitable TSC mode.
6016	* @param pDevExt Pointer to the device instance data.
6017	*/
6018	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt)
6019	{
6020	/*
6021	* On SMP we're faced with two problems:
6022	* (1) There might be a skew between the CPU, so that cpu0
6023	* returns a TSC that is slightly different from cpu1.
6024	* (2) Power management (and other things) may cause the TSC
6025	* to run at a non-constant speed, and cause the speed
6026	* to be different on the cpus. This will result in (1).
6027	*
6028	* So, on SMP systems we'll have to select the ASYNC update method
6029	* if there are symptoms of these problems.
6030	*/
6031	if (RTMpGetCount() > 1)
6032	{
6033	uint32_t uEAX, uEBX, uECX, uEDX;
6034	uint64_t u64DiffCoresIgnored;
6035
6036	/* Permit the user and/or the OS specific bits to force async mode. */
6037	if (supdrvOSGetForcedAsyncTscMode(pDevExt))
6038	return SUPGIPMODE_ASYNC_TSC;
6039
6040	/* Try check for current differences between the cpus. */
6041	if (supdrvDetermineAsyncTsc(&u64DiffCoresIgnored))
6042	return SUPGIPMODE_ASYNC_TSC;
6043
6044	/*
6045	* If the CPU supports power management and is an AMD one we
6046	* won't trust it unless it has the TscInvariant bit is set.
6047	*/
6048	/* Check for "AuthenticAMD" */
6049	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
6050	if ( uEAX >= 1
6051	&& ASMIsAmdCpuEx(uEBX, uECX, uEDX))
6052	{
6053	/* Check for APM support and that TscInvariant is cleared. */
6054	ASMCpuId(0x80000000, &uEAX, &uEBX, &uECX, &uEDX);
6055	if (uEAX >= 0x80000007)
6056	{
6057	ASMCpuId(0x80000007, &uEAX, &uEBX, &uECX, &uEDX);
6058	if ( !(uEDX & RT_BIT(8))/* TscInvariant */
6059	&& (uEDX & 0x3e)) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
6060	return SUPGIPMODE_ASYNC_TSC;
6061	}
6062	}
6063	}
6064	return SUPGIPMODE_SYNC_TSC;
6065	}
6066
6067
6068	/**
6069	* Initializes per-CPU GIP information.
6070	*
6071	* @param pGip Pointer to the read-write kernel mapping of the GIP.
6072	* @param pCpu Pointer to which GIP CPU to initalize.
6073	* @param u64NanoTS The current nanosecond timestamp.
6074	*/
6075	static void supdrvGipInitCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pCpu, uint64_t u64NanoTS)
6076	{
6077	pCpu->u32TransactionId = 2;
6078	pCpu->u64NanoTS = u64NanoTS;
6079	pCpu->u64TSC = ASMReadTSC();
6080
6081	ASMAtomicWriteSize(&pCpu->enmState, SUPGIPCPUSTATE_INVALID);
6082	ASMAtomicWriteSize(&pCpu->idCpu, NIL_RTCPUID);
6083	ASMAtomicWriteS16(&pCpu->iCpuSet, -1);
6084	ASMAtomicWriteU16(&pCpu->idApic, UINT16_MAX);
6085
6086	/*
6087	* We don't know the following values until we've executed updates.
6088	* So, we'll just pretend it's a 4 GHz CPU and adjust the history it on
6089	* the 2nd timer callout.
6090	*/
6091	pCpu->u64CpuHz = _4G + 1; /* tstGIP-2 depends on this. */
6092	pCpu->u32UpdateIntervalTSC
6093	= pCpu->au32TSCHistory[0]
6094	= pCpu->au32TSCHistory[1]
6095	= pCpu->au32TSCHistory[2]
6096	= pCpu->au32TSCHistory[3]
6097	= pCpu->au32TSCHistory[4]
6098	= pCpu->au32TSCHistory[5]
6099	= pCpu->au32TSCHistory[6]
6100	= pCpu->au32TSCHistory[7]
6101	= (uint32_t)(_4G / pGip->u32UpdateHz);
6102	}
6103
6104
6105	/**
6106	* Initializes the GIP data.
6107	*
6108	* @param pDevExt Pointer to the device instance data.
6109	* @param pGip Pointer to the read-write kernel mapping of the GIP.
6110	* @param HCPhys The physical address of the GIP.
6111	* @param u64NanoTS The current nanosecond timestamp.
6112	* @param uUpdateHz The update frequency.
6113	* @param cCpus The CPU count.
6114	*/
6115	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
6116	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned cCpus)
6117	{
6118	size_t const cbGip = RT_ALIGN_Z(RT_OFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), PAGE_SIZE);
6119	unsigned i;
6120	#ifdef DEBUG_DARWIN_GIP
6121	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
6122	#else
6123	LogFlow(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
6124	#endif
6125
6126	/*
6127	* Initialize the structure.
6128	*/
6129	memset(pGip, 0, cbGip);
6130	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
6131	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
6132	pGip->u32Mode = supdrvGipDeterminTscMode(pDevExt);
6133	pGip->cCpus = (uint16_t)cCpus;
6134	pGip->cPages = (uint16_t)(cbGip / PAGE_SIZE);
6135	pGip->u32UpdateHz = uUpdateHz;
6136	pGip->u32UpdateIntervalNS = 1000000000 / uUpdateHz;
6137	pGip->u64NanoTSLastUpdateHz = u64NanoTS;
6138	RTCpuSetEmpty(&pGip->OnlineCpuSet);
6139	RTCpuSetEmpty(&pGip->PresentCpuSet);
6140	RTMpGetSet(&pGip->PossibleCpuSet);
6141	pGip->cOnlineCpus = RTMpGetOnlineCount();
6142	pGip->cPresentCpus = RTMpGetPresentCount();
6143	pGip->cPossibleCpus = RTMpGetCount();
6144	pGip->idCpuMax = RTMpGetMaxCpuId();
6145	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromApicId); i++)
6146	pGip->aiCpuFromApicId[i] = 0;
6147	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx); i++)
6148	pGip->aiCpuFromCpuSetIdx[i] = UINT16_MAX;
6149
6150	for (i = 0; i < cCpus; i++)
6151	supdrvGipInitCpu(pGip, &pGip->aCPUs[i], u64NanoTS);
6152
6153	/*
6154	* Link it to the device extension.
6155	*/
6156	pDevExt->pGip = pGip;
6157	pDevExt->HCPhysGip = HCPhys;
6158	pDevExt->cGipUsers = 0;
6159	}
6160
6161
6162	/**
6163	* On CPU initialization callback for RTMpOnAll.
6164	*
6165	* @param idCpu The CPU ID.
6166	* @param pvUser1 The device extension.
6167	* @param pvUser2 The GIP.
6168	*/
6169	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2)
6170	{
6171	/* This is good enough, even though it will update some of the globals a
6172	bit to much. */
6173	supdrvGipMpEventOnline((PSUPDRVDEVEXT)pvUser1, idCpu);
6174	}
6175
6176
6177	/**
6178	* Invalidates the GIP data upon termination.
6179	*
6180	* @param pGip Pointer to the read-write kernel mapping of the GIP.
6181	*/
6182	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
6183	{
6184	unsigned i;
6185	pGip->u32Magic = 0;
6186	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
6187	{
6188	pGip->aCPUs[i].u64NanoTS = 0;
6189	pGip->aCPUs[i].u64TSC = 0;
6190	pGip->aCPUs[i].iTSCHistoryHead = 0;
6191	}
6192	}
6193
6194
6195	/**
6196	* Worker routine for supdrvGipUpdate and supdrvGipUpdatePerCpu that
6197	* updates all the per cpu data except the transaction id.
6198	*
6199	* @param pDevExt The device extension.
6200	* @param pGipCpu Pointer to the per cpu data.
6201	* @param u64NanoTS The current time stamp.
6202	* @param u64TSC The current TSC.
6203	* @param iTick The current timer tick.
6204	*/
6205	static void supdrvGipDoUpdateCpu(PSUPDRVDEVEXT pDevExt, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick)
6206	{
6207	uint64_t u64TSCDelta;
6208	uint32_t u32UpdateIntervalTSC;
6209	uint32_t u32UpdateIntervalTSCSlack;
6210	unsigned iTSCHistoryHead;
6211	uint64_t u64CpuHz;
6212	uint32_t u32TransactionId;
6213
6214	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
6215	AssertPtrReturnVoid(pGip);
6216
6217	/* Delta between this and the previous update. */
6218	ASMAtomicUoWriteU32(&pGipCpu->u32PrevUpdateIntervalNS, (uint32_t)(u64NanoTS - pGipCpu->u64NanoTS));
6219
6220	/*
6221	* Update the NanoTS.
6222	*/
6223	ASMAtomicWriteU64(&pGipCpu->u64NanoTS, u64NanoTS);
6224
6225	/*
6226	* Calc TSC delta.
6227	*/
6228	/** @todo validate the NanoTS delta, don't trust the OS to call us when it should... */
6229	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
6230	ASMAtomicWriteU64(&pGipCpu->u64TSC, u64TSC);
6231
6232	if (u64TSCDelta >> 32)
6233	{
6234	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
6235	pGipCpu->cErrors++;
6236	}
6237
6238	/*
6239	* On the 2nd and 3rd callout, reset the history with the current TSC
6240	* interval since the values entered by supdrvGipInit are totally off.
6241	* The interval on the 1st callout completely unreliable, the 2nd is a bit
6242	* better, while the 3rd should be most reliable.
6243	*/
6244	u32TransactionId = pGipCpu->u32TransactionId;
6245	if (RT_UNLIKELY( ( u32TransactionId == 5
6246	\|\| u32TransactionId == 7)
6247	&& ( iTick == 2
6248	\|\| iTick == 3) ))
6249	{
6250	unsigned i;
6251	for (i = 0; i < RT_ELEMENTS(pGipCpu->au32TSCHistory); i++)
6252	ASMAtomicUoWriteU32(&pGipCpu->au32TSCHistory[i], (uint32_t)u64TSCDelta);
6253	}
6254
6255	/*
6256	* TSC History.
6257	*/
6258	Assert(RT_ELEMENTS(pGipCpu->au32TSCHistory) == 8);
6259	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
6260	ASMAtomicWriteU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
6261	ASMAtomicWriteU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
6262
6263	/*
6264	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
6265	*/
6266	if (pGip->u32UpdateHz >= 1000)
6267	{
6268	uint32_t u32;
6269	u32 = pGipCpu->au32TSCHistory[0];
6270	u32 += pGipCpu->au32TSCHistory[1];
6271	u32 += pGipCpu->au32TSCHistory[2];
6272	u32 += pGipCpu->au32TSCHistory[3];
6273	u32 >>= 2;
6274	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
6275	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
6276	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
6277	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
6278	u32UpdateIntervalTSC >>= 2;
6279	u32UpdateIntervalTSC += u32;
6280	u32UpdateIntervalTSC >>= 1;
6281
6282	/* Value chosen for a 2GHz Athlon64 running linux 2.6.10/11, . */
6283	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
6284	}
6285	else if (pGip->u32UpdateHz >= 90)
6286	{
6287	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
6288	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
6289	u32UpdateIntervalTSC >>= 1;
6290
6291	/* value chosen on a 2GHz thinkpad running windows */
6292	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
6293	}
6294	else
6295	{
6296	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
6297
6298	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
6299	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
6300	}
6301	ASMAtomicWriteU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
6302
6303	/*
6304	* CpuHz.
6305	*/
6306	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, pGip->u32UpdateHz);
6307	ASMAtomicWriteU64(&pGipCpu->u64CpuHz, u64CpuHz);
6308	}
6309
6310
6311	/**
6312	* Updates the GIP.
6313	*
6314	* @param pDevExt The device extension.
6315	* @param u64NanoTS The current nanosecond timesamp.
6316	* @param u64TSC The current TSC timesamp.
6317	* @param idCpu The CPU ID.
6318	* @param iTick The current timer tick.
6319	*/
6320	static void supdrvGipUpdate(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick)
6321	{
6322	/*
6323	* Determine the relevant CPU data.
6324	*/
6325	PSUPGIPCPU pGipCpu;
6326	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
6327	AssertPtrReturnVoid(pGip);
6328
6329	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
6330	pGipCpu = &pGip->aCPUs[0];
6331	else
6332	{
6333	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
6334	if (RT_UNLIKELY(iCpu >= pGip->cCpus))
6335	return;
6336	pGipCpu = &pGip->aCPUs[iCpu];
6337	if (RT_UNLIKELY(pGipCpu->idCpu != idCpu))
6338	return;
6339	}
6340
6341	/*
6342	* Start update transaction.
6343	*/
6344	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
6345	{
6346	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
6347	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
6348	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
6349	pGipCpu->cErrors++;
6350	return;
6351	}
6352
6353	/*
6354	* Recalc the update frequency every 0x800th time.
6355	*/
6356	if (!(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
6357	{
6358	if (pGip->u64NanoTSLastUpdateHz)
6359	{
6360	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
6361	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
6362	uint32_t u32UpdateHz = (uint32_t)((UINT64_C(1000000000) * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
6363	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
6364	{
6365	ASMAtomicWriteU32(&pGip->u32UpdateHz, u32UpdateHz);
6366	ASMAtomicWriteU32(&pGip->u32UpdateIntervalNS, 1000000000 / u32UpdateHz);
6367	}
6368	#endif
6369	}
6370	ASMAtomicWriteU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS);
6371	}
6372
6373	/*
6374	* Update the data.
6375	*/
6376	supdrvGipDoUpdateCpu(pDevExt, pGipCpu, u64NanoTS, u64TSC, iTick);
6377
6378	/*
6379	* Complete transaction.
6380	*/
6381	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
6382	}
6383
6384
6385	/**
6386	* Updates the per cpu GIP data for the calling cpu.
6387	*
6388	* @param pDevExt The device extension.
6389	* @param u64NanoTS The current nanosecond timesamp.
6390	* @param u64TSC The current TSC timesamp.
6391	* @param idCpu The CPU ID.
6392	* @param idApic The APIC id for the CPU index.
6393	* @param iTick The current timer tick.
6394	*/
6395	static void supdrvGipUpdatePerCpu(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC,
6396	RTCPUID idCpu, uint8_t idApic, uint64_t iTick)
6397	{
6398	uint32_t iCpu;
6399	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
6400
6401	/*
6402	* Avoid a potential race when a CPU online notification doesn't fire on
6403	* the onlined CPU but the tick creeps in before the event notification is
6404	* run.
6405	*/
6406	if (RT_UNLIKELY(iTick == 1))
6407	{
6408	iCpu = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
6409	if (pGip->aCPUs[iCpu].enmState == SUPGIPCPUSTATE_OFFLINE)
6410	supdrvGipMpEventOnline(pDevExt, idCpu);
6411	}
6412
6413	iCpu = pGip->aiCpuFromApicId[idApic];
6414	if (RT_LIKELY(iCpu < pGip->cCpus))
6415	{
6416	PSUPGIPCPU pGipCpu = &pGip->aCPUs[iCpu];
6417	if (pGipCpu->idCpu == idCpu)
6418	{
6419	/*
6420	* Start update transaction.
6421	*/
6422	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
6423	{
6424	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
6425	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
6426	pGipCpu->cErrors++;
6427	return;
6428	}
6429
6430	/*
6431	* Update the data.
6432	*/
6433	supdrvGipDoUpdateCpu(pDevExt, pGipCpu, u64NanoTS, u64TSC, iTick);
6434
6435	/*
6436	* Complete transaction.
6437	*/
6438	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
6439	}
6440	}
6441	}
6442
6443	/**
6444	* Resume built-in keyboard on MacBook Air and Pro hosts.
6445	* If there is no built-in keyboard device, return success anyway.
6446	*
6447	* @returns 0 on Mac OS X platform, VERR_NOT_IMPLEMENTED on the other ones.
6448	*/
6449	static int supdrvIOCtl_ResumeSuspendedKbds(void)
6450	{
6451	#if defined(RT_OS_DARWIN)
6452	return supdrvDarwinResumeSuspendedKbds();
6453	#else
6454	return VERR_NOT_IMPLEMENTED;
6455	#endif
6456	}
6457

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

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDrv.c@ 51850

Download in other formats: