SUPDrv.c@ 48328

Last change on this file since 48328 was 48328, checked in by vboxsync, 11 years ago
VMM: Assert, doxygen, todo question.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 225.0 KB

Line
1	/* $Id: SUPDrv.c 48328 2013-09-05 22:20:18Z 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 supdrvGipCreate(PSUPDRVDEVEXT pDevExt);
128	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt);
129	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
130	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
131	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser);
132	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
133	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned cCpus);
134	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2);
135	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip);
136	static void supdrvGipUpdate(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick);
137	static void supdrvGipUpdatePerCpu(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC,
138	RTCPUID idCpu, uint8_t idApic, uint64_t iTick);
139	static void supdrvGipInitCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pCpu, uint64_t u64NanoTS);
140
141
142	/*******************************************************************************
143	* Global Variables *
144	*******************************************************************************/
145	DECLEXPORT(PSUPGLOBALINFOPAGE) g_pSUPGlobalInfoPage = NULL;
146
147	/**
148	* Array of the R0 SUP API.
149	*/
150	static SUPFUNC g_aFunctions[] =
151	{
152	/* SED: START */
153	/* name function */
154	/* Entries with absolute addresses determined at runtime, fixup
155	code makes ugly ASSUMPTIONS about the order here: */
156	{ "SUPR0AbsIs64bit", (void *)0 },
157	{ "SUPR0Abs64bitKernelCS", (void *)0 },
158	{ "SUPR0Abs64bitKernelSS", (void *)0 },
159	{ "SUPR0Abs64bitKernelDS", (void *)0 },
160	{ "SUPR0AbsKernelCS", (void *)0 },
161	{ "SUPR0AbsKernelSS", (void *)0 },
162	{ "SUPR0AbsKernelDS", (void *)0 },
163	{ "SUPR0AbsKernelES", (void *)0 },
164	{ "SUPR0AbsKernelFS", (void *)0 },
165	{ "SUPR0AbsKernelGS", (void *)0 },
166	/* Normal function pointers: */
167	{ "g_pSUPGlobalInfoPage", (void )&g_pSUPGlobalInfoPage }, / SED: DATA */
168	{ "SUPGetGIP", (void *)SUPGetGIP },
169	{ "SUPR0ComponentDeregisterFactory", (void *)SUPR0ComponentDeregisterFactory },
170	{ "SUPR0ComponentQueryFactory", (void *)SUPR0ComponentQueryFactory },
171	{ "SUPR0ComponentRegisterFactory", (void *)SUPR0ComponentRegisterFactory },
172	{ "SUPR0ContAlloc", (void *)SUPR0ContAlloc },
173	{ "SUPR0ContFree", (void *)SUPR0ContFree },
174	{ "SUPR0EnableVTx", (void *)SUPR0EnableVTx },
175	{ "SUPR0SuspendVTxOnCpu", (void *)SUPR0SuspendVTxOnCpu },
176	{ "SUPR0ResumeVTxOnCpu", (void *)SUPR0ResumeVTxOnCpu },
177	{ "SUPR0GetPagingMode", (void *)SUPR0GetPagingMode },
178	{ "SUPR0LockMem", (void *)SUPR0LockMem },
179	{ "SUPR0LowAlloc", (void *)SUPR0LowAlloc },
180	{ "SUPR0LowFree", (void *)SUPR0LowFree },
181	{ "SUPR0MemAlloc", (void *)SUPR0MemAlloc },
182	{ "SUPR0MemFree", (void *)SUPR0MemFree },
183	{ "SUPR0MemGetPhys", (void *)SUPR0MemGetPhys },
184	{ "SUPR0ObjAddRef", (void *)SUPR0ObjAddRef },
185	{ "SUPR0ObjAddRefEx", (void *)SUPR0ObjAddRefEx },
186	{ "SUPR0ObjRegister", (void *)SUPR0ObjRegister },
187	{ "SUPR0ObjRelease", (void *)SUPR0ObjRelease },
188	{ "SUPR0ObjVerifyAccess", (void *)SUPR0ObjVerifyAccess },
189	{ "SUPR0PageAllocEx", (void *)SUPR0PageAllocEx },
190	{ "SUPR0PageFree", (void *)SUPR0PageFree },
191	{ "SUPR0Printf", (void *)SUPR0Printf },
192	{ "SUPR0TracerDeregisterDrv", (void *)SUPR0TracerDeregisterDrv },
193	{ "SUPR0TracerDeregisterImpl", (void *)SUPR0TracerDeregisterImpl },
194	{ "SUPR0TracerFireProbe", (void *)SUPR0TracerFireProbe },
195	{ "SUPR0TracerRegisterDrv", (void *)SUPR0TracerRegisterDrv },
196	{ "SUPR0TracerRegisterImpl", (void *)SUPR0TracerRegisterImpl },
197	{ "SUPR0TracerRegisterModule", (void *)SUPR0TracerRegisterModule },
198	{ "SUPR0TracerUmodProbeFire", (void *)SUPR0TracerUmodProbeFire },
199	{ "SUPR0UnlockMem", (void *)SUPR0UnlockMem },
200	{ "SUPSemEventClose", (void *)SUPSemEventClose },
201	{ "SUPSemEventCreate", (void *)SUPSemEventCreate },
202	{ "SUPSemEventGetResolution", (void *)SUPSemEventGetResolution },
203	{ "SUPSemEventMultiClose", (void *)SUPSemEventMultiClose },
204	{ "SUPSemEventMultiCreate", (void *)SUPSemEventMultiCreate },
205	{ "SUPSemEventMultiGetResolution", (void *)SUPSemEventMultiGetResolution },
206	{ "SUPSemEventMultiReset", (void *)SUPSemEventMultiReset },
207	{ "SUPSemEventMultiSignal", (void *)SUPSemEventMultiSignal },
208	{ "SUPSemEventMultiWait", (void *)SUPSemEventMultiWait },
209	{ "SUPSemEventMultiWaitNoResume", (void *)SUPSemEventMultiWaitNoResume },
210	{ "SUPSemEventMultiWaitNsAbsIntr", (void *)SUPSemEventMultiWaitNsAbsIntr },
211	{ "SUPSemEventMultiWaitNsRelIntr", (void *)SUPSemEventMultiWaitNsRelIntr },
212	{ "SUPSemEventSignal", (void *)SUPSemEventSignal },
213	{ "SUPSemEventWait", (void *)SUPSemEventWait },
214	{ "SUPSemEventWaitNoResume", (void *)SUPSemEventWaitNoResume },
215	{ "SUPSemEventWaitNsAbsIntr", (void *)SUPSemEventWaitNsAbsIntr },
216	{ "SUPSemEventWaitNsRelIntr", (void *)SUPSemEventWaitNsRelIntr },
217
218	{ "RTAssertAreQuiet", (void *)RTAssertAreQuiet },
219	{ "RTAssertMayPanic", (void *)RTAssertMayPanic },
220	{ "RTAssertMsg1", (void *)RTAssertMsg1 },
221	{ "RTAssertMsg2AddV", (void *)RTAssertMsg2AddV },
222	{ "RTAssertMsg2V", (void *)RTAssertMsg2V },
223	{ "RTAssertSetMayPanic", (void *)RTAssertSetMayPanic },
224	{ "RTAssertSetQuiet", (void *)RTAssertSetQuiet },
225	{ "RTCrc32", (void *)RTCrc32 },
226	{ "RTCrc32Finish", (void *)RTCrc32Finish },
227	{ "RTCrc32Process", (void *)RTCrc32Process },
228	{ "RTCrc32Start", (void *)RTCrc32Start },
229	{ "RTErrConvertFromErrno", (void *)RTErrConvertFromErrno },
230	{ "RTErrConvertToErrno", (void *)RTErrConvertToErrno },
231	{ "RTHandleTableAllocWithCtx", (void *)RTHandleTableAllocWithCtx },
232	{ "RTHandleTableCreate", (void *)RTHandleTableCreate },
233	{ "RTHandleTableCreateEx", (void *)RTHandleTableCreateEx },
234	{ "RTHandleTableDestroy", (void *)RTHandleTableDestroy },
235	{ "RTHandleTableFreeWithCtx", (void *)RTHandleTableFreeWithCtx },
236	{ "RTHandleTableLookupWithCtx", (void *)RTHandleTableLookupWithCtx },
237	{ "RTLogDefaultInstance", (void *)RTLogDefaultInstance },
238	{ "RTLogGetDefaultInstance", (void *)RTLogGetDefaultInstance },
239	{ "RTLogLoggerExV", (void *)RTLogLoggerExV },
240	{ "RTLogPrintfV", (void *)RTLogPrintfV },
241	{ "RTLogRelDefaultInstance", (void *)RTLogRelDefaultInstance },
242	{ "RTLogSetDefaultInstanceThread", (void *)RTLogSetDefaultInstanceThread },
243	{ "RTMemAllocExTag", (void *)RTMemAllocExTag },
244	{ "RTMemAllocTag", (void *)RTMemAllocTag },
245	{ "RTMemAllocVarTag", (void *)RTMemAllocVarTag },
246	{ "RTMemAllocZTag", (void *)RTMemAllocZTag },
247	{ "RTMemAllocZVarTag", (void *)RTMemAllocZVarTag },
248	{ "RTMemDupExTag", (void *)RTMemDupExTag },
249	{ "RTMemDupTag", (void *)RTMemDupTag },
250	{ "RTMemFree", (void *)RTMemFree },
251	{ "RTMemFreeEx", (void *)RTMemFreeEx },
252	{ "RTMemReallocTag", (void *)RTMemReallocTag },
253	{ "RTMpCpuId", (void *)RTMpCpuId },
254	{ "RTMpCpuIdFromSetIndex", (void *)RTMpCpuIdFromSetIndex },
255	{ "RTMpCpuIdToSetIndex", (void *)RTMpCpuIdToSetIndex },
256	{ "RTMpGetArraySize", (void *)RTMpGetArraySize },
257	{ "RTMpGetCount", (void *)RTMpGetCount },
258	{ "RTMpGetMaxCpuId", (void *)RTMpGetMaxCpuId },
259	{ "RTMpGetOnlineCount", (void *)RTMpGetOnlineCount },
260	{ "RTMpGetOnlineSet", (void *)RTMpGetOnlineSet },
261	{ "RTMpGetSet", (void *)RTMpGetSet },
262	{ "RTMpIsCpuOnline", (void *)RTMpIsCpuOnline },
263	{ "RTMpIsCpuPossible", (void *)RTMpIsCpuPossible },
264	{ "RTMpIsCpuWorkPending", (void *)RTMpIsCpuWorkPending },
265	{ "RTMpNotificationDeregister", (void *)RTMpNotificationDeregister },
266	{ "RTMpNotificationRegister", (void *)RTMpNotificationRegister },
267	{ "RTMpOnAll", (void *)RTMpOnAll },
268	{ "RTMpOnOthers", (void *)RTMpOnOthers },
269	{ "RTMpOnSpecific", (void *)RTMpOnSpecific },
270	{ "RTMpPokeCpu", (void *)RTMpPokeCpu },
271	{ "RTNetIPv4AddDataChecksum", (void *)RTNetIPv4AddDataChecksum },
272	{ "RTNetIPv4AddTCPChecksum", (void *)RTNetIPv4AddTCPChecksum },
273	{ "RTNetIPv4AddUDPChecksum", (void *)RTNetIPv4AddUDPChecksum },
274	{ "RTNetIPv4FinalizeChecksum", (void *)RTNetIPv4FinalizeChecksum },
275	{ "RTNetIPv4HdrChecksum", (void *)RTNetIPv4HdrChecksum },
276	{ "RTNetIPv4IsDHCPValid", (void *)RTNetIPv4IsDHCPValid },
277	{ "RTNetIPv4IsHdrValid", (void *)RTNetIPv4IsHdrValid },
278	{ "RTNetIPv4IsTCPSizeValid", (void *)RTNetIPv4IsTCPSizeValid },
279	{ "RTNetIPv4IsTCPValid", (void *)RTNetIPv4IsTCPValid },
280	{ "RTNetIPv4IsUDPSizeValid", (void *)RTNetIPv4IsUDPSizeValid },
281	{ "RTNetIPv4IsUDPValid", (void *)RTNetIPv4IsUDPValid },
282	{ "RTNetIPv4PseudoChecksum", (void *)RTNetIPv4PseudoChecksum },
283	{ "RTNetIPv4PseudoChecksumBits", (void *)RTNetIPv4PseudoChecksumBits },
284	{ "RTNetIPv4TCPChecksum", (void *)RTNetIPv4TCPChecksum },
285	{ "RTNetIPv4UDPChecksum", (void *)RTNetIPv4UDPChecksum },
286	{ "RTNetIPv6PseudoChecksum", (void *)RTNetIPv6PseudoChecksum },
287	{ "RTNetIPv6PseudoChecksumBits", (void *)RTNetIPv6PseudoChecksumBits },
288	{ "RTNetIPv6PseudoChecksumEx", (void *)RTNetIPv6PseudoChecksumEx },
289	{ "RTNetTCPChecksum", (void *)RTNetTCPChecksum },
290	{ "RTNetUDPChecksum", (void *)RTNetUDPChecksum },
291	{ "RTPowerNotificationDeregister", (void *)RTPowerNotificationDeregister },
292	{ "RTPowerNotificationRegister", (void *)RTPowerNotificationRegister },
293	{ "RTProcSelf", (void *)RTProcSelf },
294	{ "RTR0AssertPanicSystem", (void *)RTR0AssertPanicSystem },
295	{ "RTR0MemAreKrnlAndUsrDifferent", (void *)RTR0MemAreKrnlAndUsrDifferent },
296	{ "RTR0MemKernelIsValidAddr", (void *)RTR0MemKernelIsValidAddr },
297	{ "RTR0MemKernelCopyFrom", (void *)RTR0MemKernelCopyFrom },
298	{ "RTR0MemKernelCopyTo", (void *)RTR0MemKernelCopyTo },
299	{ "RTR0MemObjAddress", (void *)RTR0MemObjAddress },
300	{ "RTR0MemObjAddressR3", (void *)RTR0MemObjAddressR3 },
301	{ "RTR0MemObjAllocContTag", (void *)RTR0MemObjAllocContTag },
302	{ "RTR0MemObjAllocLowTag", (void *)RTR0MemObjAllocLowTag },
303	{ "RTR0MemObjAllocPageTag", (void *)RTR0MemObjAllocPageTag },
304	{ "RTR0MemObjAllocPhysExTag", (void *)RTR0MemObjAllocPhysExTag },
305	{ "RTR0MemObjAllocPhysNCTag", (void *)RTR0MemObjAllocPhysNCTag },
306	{ "RTR0MemObjAllocPhysTag", (void *)RTR0MemObjAllocPhysTag },
307	{ "RTR0MemObjEnterPhysTag", (void *)RTR0MemObjEnterPhysTag },
308	{ "RTR0MemObjFree", (void *)RTR0MemObjFree },
309	{ "RTR0MemObjGetPagePhysAddr", (void *)RTR0MemObjGetPagePhysAddr },
310	{ "RTR0MemObjIsMapping", (void *)RTR0MemObjIsMapping },
311	{ "RTR0MemObjLockUserTag", (void *)RTR0MemObjLockUserTag },
312	{ "RTR0MemObjMapKernelExTag", (void *)RTR0MemObjMapKernelExTag },
313	{ "RTR0MemObjMapKernelTag", (void *)RTR0MemObjMapKernelTag },
314	{ "RTR0MemObjMapUserTag", (void *)RTR0MemObjMapUserTag },
315	{ "RTR0MemObjProtect", (void *)RTR0MemObjProtect },
316	{ "RTR0MemObjSize", (void *)RTR0MemObjSize },
317	{ "RTR0MemUserCopyFrom", (void *)RTR0MemUserCopyFrom },
318	{ "RTR0MemUserCopyTo", (void *)RTR0MemUserCopyTo },
319	{ "RTR0MemUserIsValidAddr", (void *)RTR0MemUserIsValidAddr },
320	{ "RTR0ProcHandleSelf", (void *)RTR0ProcHandleSelf },
321	{ "RTSemEventCreate", (void *)RTSemEventCreate },
322	{ "RTSemEventDestroy", (void *)RTSemEventDestroy },
323	{ "RTSemEventGetResolution", (void *)RTSemEventGetResolution },
324	{ "RTSemEventMultiCreate", (void *)RTSemEventMultiCreate },
325	{ "RTSemEventMultiDestroy", (void *)RTSemEventMultiDestroy },
326	{ "RTSemEventMultiGetResolution", (void *)RTSemEventMultiGetResolution },
327	{ "RTSemEventMultiReset", (void *)RTSemEventMultiReset },
328	{ "RTSemEventMultiSignal", (void *)RTSemEventMultiSignal },
329	{ "RTSemEventMultiWait", (void *)RTSemEventMultiWait },
330	{ "RTSemEventMultiWaitEx", (void *)RTSemEventMultiWaitEx },
331	{ "RTSemEventMultiWaitExDebug", (void *)RTSemEventMultiWaitExDebug },
332	{ "RTSemEventMultiWaitNoResume", (void *)RTSemEventMultiWaitNoResume },
333	{ "RTSemEventSignal", (void *)RTSemEventSignal },
334	{ "RTSemEventWait", (void *)RTSemEventWait },
335	{ "RTSemEventWaitEx", (void *)RTSemEventWaitEx },
336	{ "RTSemEventWaitExDebug", (void *)RTSemEventWaitExDebug },
337	{ "RTSemEventWaitNoResume", (void *)RTSemEventWaitNoResume },
338	{ "RTSemFastMutexCreate", (void *)RTSemFastMutexCreate },
339	{ "RTSemFastMutexDestroy", (void *)RTSemFastMutexDestroy },
340	{ "RTSemFastMutexRelease", (void *)RTSemFastMutexRelease },
341	{ "RTSemFastMutexRequest", (void *)RTSemFastMutexRequest },
342	{ "RTSemMutexCreate", (void *)RTSemMutexCreate },
343	{ "RTSemMutexDestroy", (void *)RTSemMutexDestroy },
344	{ "RTSemMutexRelease", (void *)RTSemMutexRelease },
345	{ "RTSemMutexRequest", (void *)RTSemMutexRequest },
346	{ "RTSemMutexRequestDebug", (void *)RTSemMutexRequestDebug },
347	{ "RTSemMutexRequestNoResume", (void *)RTSemMutexRequestNoResume },
348	{ "RTSemMutexRequestNoResumeDebug", (void *)RTSemMutexRequestNoResumeDebug },
349	{ "RTSpinlockAcquire", (void *)RTSpinlockAcquire },
350	{ "RTSpinlockCreate", (void *)RTSpinlockCreate },
351	{ "RTSpinlockDestroy", (void *)RTSpinlockDestroy },
352	{ "RTSpinlockRelease", (void *)RTSpinlockRelease },
353	{ "RTSpinlockReleaseNoInts", (void *)RTSpinlockReleaseNoInts },
354	{ "RTStrCopy", (void *)RTStrCopy },
355	{ "RTStrDupTag", (void *)RTStrDupTag },
356	{ "RTStrFormat", (void *)RTStrFormat },
357	{ "RTStrFormatNumber", (void *)RTStrFormatNumber },
358	{ "RTStrFormatTypeDeregister", (void *)RTStrFormatTypeDeregister },
359	{ "RTStrFormatTypeRegister", (void *)RTStrFormatTypeRegister },
360	{ "RTStrFormatTypeSetUser", (void *)RTStrFormatTypeSetUser },
361	{ "RTStrFormatV", (void *)RTStrFormatV },
362	{ "RTStrFree", (void *)RTStrFree },
363	{ "RTStrNCmp", (void *)RTStrNCmp },
364	{ "RTStrPrintf", (void *)RTStrPrintf },
365	{ "RTStrPrintfEx", (void *)RTStrPrintfEx },
366	{ "RTStrPrintfExV", (void *)RTStrPrintfExV },
367	{ "RTStrPrintfV", (void *)RTStrPrintfV },
368	{ "RTThreadCreate", (void *)RTThreadCreate },
369	{ "RTThreadCtxHooksAreRegistered", (void *)RTThreadCtxHooksAreRegistered },
370	{ "RTThreadCtxHooksCreate", (void *)RTThreadCtxHooksCreate },
371	{ "RTThreadCtxHooksDeregister", (void *)RTThreadCtxHooksDeregister },
372	{ "RTThreadCtxHooksRegister", (void *)RTThreadCtxHooksRegister },
373	{ "RTThreadCtxHooksRelease", (void *)RTThreadCtxHooksRelease },
374	{ "RTThreadCtxHooksRetain", (void *)RTThreadCtxHooksRetain },
375	{ "RTThreadGetName", (void *)RTThreadGetName },
376	{ "RTThreadGetNative", (void *)RTThreadGetNative },
377	{ "RTThreadGetType", (void *)RTThreadGetType },
378	{ "RTThreadIsInInterrupt", (void *)RTThreadIsInInterrupt },
379	{ "RTThreadNativeSelf", (void *)RTThreadNativeSelf },
380	{ "RTThreadPreemptDisable", (void *)RTThreadPreemptDisable },
381	{ "RTThreadPreemptIsEnabled", (void *)RTThreadPreemptIsEnabled },
382	{ "RTThreadPreemptIsPending", (void *)RTThreadPreemptIsPending },
383	{ "RTThreadPreemptIsPendingTrusty", (void *)RTThreadPreemptIsPendingTrusty },
384	{ "RTThreadPreemptIsPossible", (void *)RTThreadPreemptIsPossible },
385	{ "RTThreadPreemptRestore", (void *)RTThreadPreemptRestore },
386	{ "RTThreadSelf", (void *)RTThreadSelf },
387	{ "RTThreadSelfName", (void *)RTThreadSelfName },
388	{ "RTThreadSleep", (void *)RTThreadSleep },
389	{ "RTThreadUserReset", (void *)RTThreadUserReset },
390	{ "RTThreadUserSignal", (void *)RTThreadUserSignal },
391	{ "RTThreadUserWait", (void *)RTThreadUserWait },
392	{ "RTThreadUserWaitNoResume", (void *)RTThreadUserWaitNoResume },
393	{ "RTThreadWait", (void *)RTThreadWait },
394	{ "RTThreadWaitNoResume", (void *)RTThreadWaitNoResume },
395	{ "RTThreadYield", (void *)RTThreadYield },
396	{ "RTTimeMilliTS", (void *)RTTimeMilliTS },
397	{ "RTTimeNanoTS", (void *)RTTimeNanoTS },
398	{ "RTTimeNow", (void *)RTTimeNow },
399	{ "RTTimerCanDoHighResolution", (void *)RTTimerCanDoHighResolution },
400	{ "RTTimerChangeInterval", (void *)RTTimerChangeInterval },
401	{ "RTTimerCreate", (void *)RTTimerCreate },
402	{ "RTTimerCreateEx", (void *)RTTimerCreateEx },
403	{ "RTTimerDestroy", (void *)RTTimerDestroy },
404	{ "RTTimerGetSystemGranularity", (void *)RTTimerGetSystemGranularity },
405	{ "RTTimerReleaseSystemGranularity", (void *)RTTimerReleaseSystemGranularity },
406	{ "RTTimerRequestSystemGranularity", (void *)RTTimerRequestSystemGranularity },
407	{ "RTTimerStart", (void *)RTTimerStart },
408	{ "RTTimerStop", (void *)RTTimerStop },
409	{ "RTTimeSystemMilliTS", (void *)RTTimeSystemMilliTS },
410	{ "RTTimeSystemNanoTS", (void *)RTTimeSystemNanoTS },
411	{ "RTUuidCompare", (void *)RTUuidCompare },
412	{ "RTUuidCompareStr", (void *)RTUuidCompareStr },
413	{ "RTUuidFromStr", (void *)RTUuidFromStr },
414	/* SED: END */
415	};
416
417	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
418	/**
419	* Drag in the rest of IRPT since we share it with the
420	* rest of the kernel modules on darwin.
421	*/
422	PFNRT g_apfnVBoxDrvIPRTDeps[] =
423	{
424	/* VBoxNetAdp */
425	(PFNRT)RTRandBytes,
426	/* VBoxUSB */
427	(PFNRT)RTPathStripFilename,
428	NULL
429	};
430	#endif /* RT_OS_DARWIN \|\| RT_OS_SOLARIS \|\| RT_OS_SOLARIS */
431
432
433	/**
434	* Initializes the device extentsion structure.
435	*
436	* @returns IPRT status code.
437	* @param pDevExt The device extension to initialize.
438	* @param cbSession The size of the session structure. The size of
439	* SUPDRVSESSION may be smaller when SUPDRV_AGNOSTIC is
440	* defined because we're skipping the OS specific members
441	* then.
442	*/
443	int VBOXCALL supdrvInitDevExt(PSUPDRVDEVEXT pDevExt, size_t cbSession)
444	{
445	int rc;
446
447	#ifdef SUPDRV_WITH_RELEASE_LOGGER
448	/*
449	* Create the release log.
450	*/
451	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
452	PRTLOGGER pRelLogger;
453	rc = RTLogCreate(&pRelLogger, 0 /* fFlags */, "all",
454	"VBOX_RELEASE_LOG", RT_ELEMENTS(s_apszGroups), s_apszGroups, RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER, NULL);
455	if (RT_SUCCESS(rc))
456	RTLogRelSetDefaultInstance(pRelLogger);
457	/** @todo Add native hook for getting logger config parameters and setting
458	* them. On linux we should use the module parameter stuff... */
459	#endif
460
461	/*
462	* Initialize it.
463	*/
464	memset(pDevExt, 0, sizeof(*pDevExt));
465	rc = RTSpinlockCreate(&pDevExt->Spinlock, RTSPINLOCK_FLAGS_INTERRUPT_SAFE, "SUPDrvDevExt");
466	if (RT_SUCCESS(rc))
467	{
468	rc = RTSpinlockCreate(&pDevExt->hGipSpinlock, RTSPINLOCK_FLAGS_INTERRUPT_SAFE, "SUPDrvGip");
469	if (RT_SUCCESS(rc))
470	{
471	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
472	rc = RTSemMutexCreate(&pDevExt->mtxLdr);
473	#else
474	rc = RTSemFastMutexCreate(&pDevExt->mtxLdr);
475	#endif
476	if (RT_SUCCESS(rc))
477	{
478	rc = RTSemFastMutexCreate(&pDevExt->mtxComponentFactory);
479	if (RT_SUCCESS(rc))
480	{
481	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
482	rc = RTSemMutexCreate(&pDevExt->mtxGip);
483	#else
484	rc = RTSemFastMutexCreate(&pDevExt->mtxGip);
485	#endif
486	if (RT_SUCCESS(rc))
487	{
488	rc = supdrvGipCreate(pDevExt);
489	if (RT_SUCCESS(rc))
490	{
491	rc = supdrvTracerInit(pDevExt);
492	if (RT_SUCCESS(rc))
493	{
494	pDevExt->pLdrInitImage = NULL;
495	pDevExt->hLdrInitThread = NIL_RTNATIVETHREAD;
496	pDevExt->u32Cookie = BIRD; /** @todo make this random? */
497	pDevExt->cbSession = (uint32_t)cbSession;
498
499	/*
500	* Fixup the absolute symbols.
501	*
502	* Because of the table indexing assumptions we'll have a little #ifdef orgy
503	* here rather than distributing this to OS specific files. At least for now.
504	*/
505	#ifdef RT_OS_DARWIN
506	# if ARCH_BITS == 32
507	if (SUPR0GetPagingMode() >= SUPPAGINGMODE_AMD64)
508	{
509	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
510	g_aFunctions[1].pfn = (void )0x80; / SUPR0Abs64bitKernelCS - KERNEL64_CS, seg.h */
511	g_aFunctions[2].pfn = (void )0x88; / SUPR0Abs64bitKernelSS - KERNEL64_SS, seg.h */
512	g_aFunctions[3].pfn = (void )0x88; / SUPR0Abs64bitKernelDS - KERNEL64_SS, seg.h */
513	}
514	else
515	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
516	g_aFunctions[4].pfn = (void )0x08; / SUPR0AbsKernelCS - KERNEL_CS, seg.h */
517	g_aFunctions[5].pfn = (void )0x10; / SUPR0AbsKernelSS - KERNEL_DS, seg.h */
518	g_aFunctions[6].pfn = (void )0x10; / SUPR0AbsKernelDS - KERNEL_DS, seg.h */
519	g_aFunctions[7].pfn = (void )0x10; / SUPR0AbsKernelES - KERNEL_DS, seg.h */
520	g_aFunctions[8].pfn = (void )0x10; / SUPR0AbsKernelFS - KERNEL_DS, seg.h */
521	g_aFunctions[9].pfn = (void )0x48; / SUPR0AbsKernelGS - CPU_DATA_GS, seg.h */
522	# else /* 64-bit darwin: */
523	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
524	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
525	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
526	g_aFunctions[3].pfn = (void )0; / SUPR0Abs64bitKernelDS */
527	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
528	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
529	g_aFunctions[6].pfn = (void )0; / SUPR0AbsKernelDS */
530	g_aFunctions[7].pfn = (void )0; / SUPR0AbsKernelES */
531	g_aFunctions[8].pfn = (void )0; / SUPR0AbsKernelFS */
532	g_aFunctions[9].pfn = (void )0; / SUPR0AbsKernelGS */
533
534	# endif
535	#else /* !RT_OS_DARWIN */
536	# if ARCH_BITS == 64
537	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
538	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
539	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
540	g_aFunctions[3].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0Abs64bitKernelDS */
541	# else
542	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
543	# endif
544	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
545	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
546	g_aFunctions[6].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0AbsKernelDS */
547	g_aFunctions[7].pfn = (void )(uintptr_t)ASMGetES(); / SUPR0AbsKernelES */
548	g_aFunctions[8].pfn = (void )(uintptr_t)ASMGetFS(); / SUPR0AbsKernelFS */
549	g_aFunctions[9].pfn = (void )(uintptr_t)ASMGetGS(); / SUPR0AbsKernelGS */
550	#endif /* !RT_OS_DARWIN */
551	return VINF_SUCCESS;
552	}
553
554	supdrvGipDestroy(pDevExt);
555	}
556
557	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
558	RTSemMutexDestroy(pDevExt->mtxGip);
559	pDevExt->mtxGip = NIL_RTSEMMUTEX;
560	#else
561	RTSemFastMutexDestroy(pDevExt->mtxGip);
562	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
563	#endif
564	}
565	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
566	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
567	}
568	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
569	RTSemMutexDestroy(pDevExt->mtxLdr);
570	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
571	#else
572	RTSemFastMutexDestroy(pDevExt->mtxLdr);
573	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
574	#endif
575	}
576	RTSpinlockDestroy(pDevExt->hGipSpinlock);
577	pDevExt->hGipSpinlock = NIL_RTSPINLOCK;
578	}
579	RTSpinlockDestroy(pDevExt->Spinlock);
580	pDevExt->Spinlock = NIL_RTSPINLOCK;
581	}
582	#ifdef SUPDRV_WITH_RELEASE_LOGGER
583	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
584	RTLogDestroy(RTLogSetDefaultInstance(NULL));
585	#endif
586
587	return rc;
588	}
589
590
591	/**
592	* Delete the device extension (e.g. cleanup members).
593	*
594	* @param pDevExt The device extension to delete.
595	*/
596	void VBOXCALL supdrvDeleteDevExt(PSUPDRVDEVEXT pDevExt)
597	{
598	PSUPDRVOBJ pObj;
599	PSUPDRVUSAGE pUsage;
600
601	/*
602	* Kill mutexes and spinlocks.
603	*/
604	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
605	RTSemMutexDestroy(pDevExt->mtxGip);
606	pDevExt->mtxGip = NIL_RTSEMMUTEX;
607	#else
608	RTSemFastMutexDestroy(pDevExt->mtxGip);
609	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
610	#endif
611	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
612	RTSemMutexDestroy(pDevExt->mtxLdr);
613	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
614	#else
615	RTSemFastMutexDestroy(pDevExt->mtxLdr);
616	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
617	#endif
618	RTSpinlockDestroy(pDevExt->Spinlock);
619	pDevExt->Spinlock = NIL_RTSPINLOCK;
620	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
621	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
622
623	/*
624	* Free lists.
625	*/
626	/* objects. */
627	pObj = pDevExt->pObjs;
628	Assert(!pObj); /* (can trigger on forced unloads) */
629	pDevExt->pObjs = NULL;
630	while (pObj)
631	{
632	void *pvFree = pObj;
633	pObj = pObj->pNext;
634	RTMemFree(pvFree);
635	}
636
637	/* usage records. */
638	pUsage = pDevExt->pUsageFree;
639	pDevExt->pUsageFree = NULL;
640	while (pUsage)
641	{
642	void *pvFree = pUsage;
643	pUsage = pUsage->pNext;
644	RTMemFree(pvFree);
645	}
646
647	/* kill the GIP. */
648	supdrvGipDestroy(pDevExt);
649	RTSpinlockDestroy(pDevExt->hGipSpinlock);
650	pDevExt->hGipSpinlock = NIL_RTSPINLOCK;
651
652	supdrvTracerTerm(pDevExt);
653
654	#ifdef SUPDRV_WITH_RELEASE_LOGGER
655	/* destroy the loggers. */
656	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
657	RTLogDestroy(RTLogSetDefaultInstance(NULL));
658	#endif
659	}
660
661
662	/**
663	* Create session.
664	*
665	* @returns IPRT status code.
666	* @param pDevExt Device extension.
667	* @param fUser Flag indicating whether this is a user or kernel
668	* session.
669	* @param fUnrestricted Unrestricted access (system) or restricted access
670	* (user)?
671	* @param ppSession Where to store the pointer to the session data.
672	*/
673	int VBOXCALL supdrvCreateSession(PSUPDRVDEVEXT pDevExt, bool fUser, bool fUnrestricted, PSUPDRVSESSION *ppSession)
674	{
675	int rc;
676	PSUPDRVSESSION pSession;
677
678	if (!SUP_IS_DEVEXT_VALID(pDevExt))
679	return VERR_INVALID_PARAMETER;
680
681	/*
682	* Allocate memory for the session data.
683	*/
684	pSession = *ppSession = (PSUPDRVSESSION)RTMemAllocZ(pDevExt->cbSession);
685	if (pSession)
686	{
687	/* Initialize session data. */
688	rc = RTSpinlockCreate(&pSession->Spinlock, RTSPINLOCK_FLAGS_INTERRUPT_UNSAFE, "SUPDrvSession");
689	if (!rc)
690	{
691	rc = RTHandleTableCreateEx(&pSession->hHandleTable,
692	RTHANDLETABLE_FLAGS_LOCKED_IRQ_SAFE \| RTHANDLETABLE_FLAGS_CONTEXT,
693	1 /uBase/, 32768 /cMax/, supdrvSessionObjHandleRetain, pSession);
694	if (RT_SUCCESS(rc))
695	{
696	Assert(pSession->Spinlock != NIL_RTSPINLOCK);
697	pSession->pDevExt = pDevExt;
698	pSession->u32Cookie = BIRD_INV;
699	pSession->fUnrestricted = fUnrestricted;
700	pSession->cRefs = 1;
701	/*pSession->pLdrUsage = NULL;
702	pSession->pVM = NULL;
703	pSession->pUsage = NULL;
704	pSession->pGip = NULL;
705	pSession->fGipReferenced = false;
706	pSession->Bundle.cUsed = 0; */
707	pSession->Uid = NIL_RTUID;
708	pSession->Gid = NIL_RTGID;
709	if (fUser)
710	{
711	pSession->Process = RTProcSelf();
712	pSession->R0Process = RTR0ProcHandleSelf();
713	}
714	else
715	{
716	pSession->Process = NIL_RTPROCESS;
717	pSession->R0Process = NIL_RTR0PROCESS;
718	}
719	/pSession->uTracerData = 0;/
720	pSession->hTracerCaller = NIL_RTNATIVETHREAD;
721	RTListInit(&pSession->TpProviders);
722	/pSession->cTpProviders = 0;/
723	/pSession->cTpProbesFiring = 0;/
724	RTListInit(&pSession->TpUmods);
725	/RT_ZERO(pSession->apTpLookupTable);/
726
727	VBOXDRV_SESSION_CREATE(pSession, fUser);
728	LogFlow(("Created session %p initial cookie=%#x\n", pSession, pSession->u32Cookie));
729	return VINF_SUCCESS;
730	}
731
732	RTSpinlockDestroy(pSession->Spinlock);
733	}
734	RTMemFree(pSession);
735	*ppSession = NULL;
736	Log(("Failed to create spinlock, rc=%d!\n", rc));
737	}
738	else
739	rc = VERR_NO_MEMORY;
740
741	return rc;
742	}
743
744
745	/**
746	* Shared code for cleaning up a session (but not quite freeing it).
747	*
748	* This is primarily intended for MAC OS X where we have to clean up the memory
749	* stuff before the file handle is closed.
750	*
751	* @param pDevExt Device extension.
752	* @param pSession Session data.
753	* This data will be freed by this routine.
754	*/
755	static void supdrvCleanupSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
756	{
757	int rc;
758	PSUPDRVBUNDLE pBundle;
759	LogFlow(("supdrvCleanupSession: pSession=%p\n", pSession));
760
761	/*
762	* Remove logger instances related to this session.
763	*/
764	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pSession);
765
766	/*
767	* Destroy the handle table.
768	*/
769	rc = RTHandleTableDestroy(pSession->hHandleTable, supdrvSessionObjHandleDelete, pSession);
770	AssertRC(rc);
771	pSession->hHandleTable = NIL_RTHANDLETABLE;
772
773	/*
774	* Release object references made in this session.
775	* In theory there should be noone racing us in this session.
776	*/
777	Log2(("release objects - start\n"));
778	if (pSession->pUsage)
779	{
780	PSUPDRVUSAGE pUsage;
781	RTSpinlockAcquire(pDevExt->Spinlock);
782
783	while ((pUsage = pSession->pUsage) != NULL)
784	{
785	PSUPDRVOBJ pObj = pUsage->pObj;
786	pSession->pUsage = pUsage->pNext;
787
788	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
789	if (pUsage->cUsage < pObj->cUsage)
790	{
791	pObj->cUsage -= pUsage->cUsage;
792	RTSpinlockRelease(pDevExt->Spinlock);
793	}
794	else
795	{
796	/* Destroy the object and free the record. */
797	if (pDevExt->pObjs == pObj)
798	pDevExt->pObjs = pObj->pNext;
799	else
800	{
801	PSUPDRVOBJ pObjPrev;
802	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
803	if (pObjPrev->pNext == pObj)
804	{
805	pObjPrev->pNext = pObj->pNext;
806	break;
807	}
808	Assert(pObjPrev);
809	}
810	RTSpinlockRelease(pDevExt->Spinlock);
811
812	Log(("supdrvCleanupSession: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
813	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
814	if (pObj->pfnDestructor)
815	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
816	RTMemFree(pObj);
817	}
818
819	/* free it and continue. */
820	RTMemFree(pUsage);
821
822	RTSpinlockAcquire(pDevExt->Spinlock);
823	}
824
825	RTSpinlockRelease(pDevExt->Spinlock);
826	AssertMsg(!pSession->pUsage, ("Some buster reregistered an object during desturction!\n"));
827	}
828	Log2(("release objects - done\n"));
829
830	/*
831	* Do tracer cleanups related to this session.
832	*/
833	Log2(("release tracer stuff - start\n"));
834	supdrvTracerCleanupSession(pDevExt, pSession);
835	Log2(("release tracer stuff - end\n"));
836
837	/*
838	* Release memory allocated in the session.
839	*
840	* We do not serialize this as we assume that the application will
841	* not allocated memory while closing the file handle object.
842	*/
843	Log2(("freeing memory:\n"));
844	pBundle = &pSession->Bundle;
845	while (pBundle)
846	{
847	PSUPDRVBUNDLE pToFree;
848	unsigned i;
849
850	/*
851	* Check and unlock all entries in the bundle.
852	*/
853	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
854	{
855	if (pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ)
856	{
857	Log2(("eType=%d pvR0=%p pvR3=%p cb=%ld\n", pBundle->aMem[i].eType, RTR0MemObjAddress(pBundle->aMem[i].MemObj),
858	(void *)RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3), (long)RTR0MemObjSize(pBundle->aMem[i].MemObj)));
859	if (pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ)
860	{
861	rc = RTR0MemObjFree(pBundle->aMem[i].MapObjR3, false);
862	AssertRC(rc); /** @todo figure out how to handle this. */
863	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
864	}
865	rc = RTR0MemObjFree(pBundle->aMem[i].MemObj, true /* fFreeMappings */);
866	AssertRC(rc); /** @todo figure out how to handle this. */
867	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
868	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
869	}
870	}
871
872	/*
873	* Advance and free previous bundle.
874	*/
875	pToFree = pBundle;
876	pBundle = pBundle->pNext;
877
878	pToFree->pNext = NULL;
879	pToFree->cUsed = 0;
880	if (pToFree != &pSession->Bundle)
881	RTMemFree(pToFree);
882	}
883	Log2(("freeing memory - done\n"));
884
885	/*
886	* Deregister component factories.
887	*/
888	RTSemFastMutexRequest(pDevExt->mtxComponentFactory);
889	Log2(("deregistering component factories:\n"));
890	if (pDevExt->pComponentFactoryHead)
891	{
892	PSUPDRVFACTORYREG pPrev = NULL;
893	PSUPDRVFACTORYREG pCur = pDevExt->pComponentFactoryHead;
894	while (pCur)
895	{
896	if (pCur->pSession == pSession)
897	{
898	/* unlink it */
899	PSUPDRVFACTORYREG pNext = pCur->pNext;
900	if (pPrev)
901	pPrev->pNext = pNext;
902	else
903	pDevExt->pComponentFactoryHead = pNext;
904
905	/* free it */
906	pCur->pNext = NULL;
907	pCur->pSession = NULL;
908	pCur->pFactory = NULL;
909	RTMemFree(pCur);
910
911	/* next */
912	pCur = pNext;
913	}
914	else
915	{
916	/* next */
917	pPrev = pCur;
918	pCur = pCur->pNext;
919	}
920	}
921	}
922	RTSemFastMutexRelease(pDevExt->mtxComponentFactory);
923	Log2(("deregistering component factories - done\n"));
924
925	/*
926	* Loaded images needs to be dereferenced and possibly freed up.
927	*/
928	supdrvLdrLock(pDevExt);
929	Log2(("freeing images:\n"));
930	if (pSession->pLdrUsage)
931	{
932	PSUPDRVLDRUSAGE pUsage = pSession->pLdrUsage;
933	pSession->pLdrUsage = NULL;
934	while (pUsage)
935	{
936	void *pvFree = pUsage;
937	PSUPDRVLDRIMAGE pImage = pUsage->pImage;
938	if (pImage->cUsage > pUsage->cUsage)
939	pImage->cUsage -= pUsage->cUsage;
940	else
941	supdrvLdrFree(pDevExt, pImage);
942	pUsage->pImage = NULL;
943	pUsage = pUsage->pNext;
944	RTMemFree(pvFree);
945	}
946	}
947	supdrvLdrUnlock(pDevExt);
948	Log2(("freeing images - done\n"));
949
950	/*
951	* Unmap the GIP.
952	*/
953	Log2(("umapping GIP:\n"));
954	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
955	{
956	SUPR0GipUnmap(pSession);
957	pSession->fGipReferenced = 0;
958	}
959	Log2(("umapping GIP - done\n"));
960	}
961
962
963	/**
964	* Shared code for cleaning up a session.
965	*
966	* @param pDevExt Device extension.
967	* @param pSession Session data.
968	* This data will be freed by this routine.
969	*/
970	static void supdrvCloseSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
971	{
972	VBOXDRV_SESSION_CLOSE(pSession);
973
974	/*
975	* Cleanup the session first.
976	*/
977	supdrvCleanupSession(pDevExt, pSession);
978
979	/*
980	* Free the rest of the session stuff.
981	*/
982	RTSpinlockDestroy(pSession->Spinlock);
983	pSession->Spinlock = NIL_RTSPINLOCK;
984	pSession->pDevExt = NULL;
985	RTMemFree(pSession);
986	LogFlow(("supdrvCloseSession: returns\n"));
987	}
988
989
990	/**
991	* Retain a session to make sure it doesn't go away while it is in use.
992	*
993	* @returns New reference count on success, UINT32_MAX on failure.
994	* @param pSession Session data.
995	*/
996	uint32_t VBOXCALL supdrvSessionRetain(PSUPDRVSESSION pSession)
997	{
998	uint32_t cRefs;
999	AssertPtrReturn(pSession, UINT32_MAX);
1000	AssertReturn(SUP_IS_SESSION_VALID(pSession), UINT32_MAX);
1001
1002	cRefs = ASMAtomicIncU32(&pSession->cRefs);
1003	AssertMsg(cRefs > 1 && cRefs < _1M, ("%#x %p\n", cRefs, pSession));
1004	return cRefs;
1005	}
1006
1007
1008	/**
1009	* Releases a given session.
1010	*
1011	* @returns New reference count on success (0 if closed), UINT32_MAX on failure.
1012	* @param pSession Session data.
1013	*/
1014	uint32_t VBOXCALL supdrvSessionRelease(PSUPDRVSESSION pSession)
1015	{
1016	uint32_t cRefs;
1017	AssertPtrReturn(pSession, UINT32_MAX);
1018	AssertReturn(SUP_IS_SESSION_VALID(pSession), UINT32_MAX);
1019
1020	cRefs = ASMAtomicDecU32(&pSession->cRefs);
1021	AssertMsg(cRefs < _1M, ("%#x %p\n", cRefs, pSession));
1022	if (cRefs == 0)
1023	supdrvCloseSession(pSession->pDevExt, pSession);
1024	return cRefs;
1025	}
1026
1027
1028	/**
1029	* RTHandleTableDestroy callback used by supdrvCleanupSession.
1030	*
1031	* @returns IPRT status code, see SUPR0ObjAddRef.
1032	* @param hHandleTable The handle table handle. Ignored.
1033	* @param pvObj The object pointer.
1034	* @param pvCtx Context, the handle type. Ignored.
1035	* @param pvUser Session pointer.
1036	*/
1037	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser)
1038	{
1039	NOREF(pvCtx);
1040	NOREF(hHandleTable);
1041	return SUPR0ObjAddRefEx(pvObj, (PSUPDRVSESSION)pvUser, true /fNoBlocking/);
1042	}
1043
1044
1045	/**
1046	* RTHandleTableDestroy callback used by supdrvCleanupSession.
1047	*
1048	* @param hHandleTable The handle table handle. Ignored.
1049	* @param h The handle value. Ignored.
1050	* @param pvObj The object pointer.
1051	* @param pvCtx Context, the handle type. Ignored.
1052	* @param pvUser Session pointer.
1053	*/
1054	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser)
1055	{
1056	NOREF(pvCtx);
1057	NOREF(h);
1058	NOREF(hHandleTable);
1059	SUPR0ObjRelease(pvObj, (PSUPDRVSESSION)pvUser);
1060	}
1061
1062
1063	/**
1064	* Fast path I/O Control worker.
1065	*
1066	* @returns VBox status code that should be passed down to ring-3 unchanged.
1067	* @param uIOCtl Function number.
1068	* @param idCpu VMCPU id.
1069	* @param pDevExt Device extention.
1070	* @param pSession Session data.
1071	*/
1072	int VBOXCALL supdrvIOCtlFast(uintptr_t uIOCtl, VMCPUID idCpu, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
1073	{
1074	/*
1075	* We check the two prereqs after doing this only to allow the compiler to optimize things better.
1076	*/
1077	if (RT_LIKELY( RT_VALID_PTR(pSession)
1078	&& pSession->pVM
1079	&& pDevExt->pfnVMMR0EntryFast))
1080	{
1081	switch (uIOCtl)
1082	{
1083	case SUP_IOCTL_FAST_DO_RAW_RUN:
1084	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_RAW_RUN);
1085	break;
1086	case SUP_IOCTL_FAST_DO_HM_RUN:
1087	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_HM_RUN);
1088	break;
1089	case SUP_IOCTL_FAST_DO_NOP:
1090	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_NOP);
1091	break;
1092	default:
1093	return VERR_INTERNAL_ERROR;
1094	}
1095	return VINF_SUCCESS;
1096	}
1097	return VERR_INTERNAL_ERROR;
1098	}
1099
1100
1101	/**
1102	* Helper for supdrvIOCtl. Check if pszStr contains any character of pszChars.
1103	* We would use strpbrk here if this function would be contained in the RedHat kABI white
1104	* list, see http://www.kerneldrivers.org/RHEL5.
1105	*
1106	* @returns 1 if pszStr does contain any character of pszChars, 0 otherwise.
1107	* @param pszStr String to check
1108	* @param pszChars Character set
1109	*/
1110	static int supdrvCheckInvalidChar(const char pszStr, const char pszChars)
1111	{
1112	int chCur;
1113	while ((chCur = *pszStr++) != '\0')
1114	{
1115	int ch;
1116	const char *psz = pszChars;
1117	while ((ch = *psz++) != '\0')
1118	if (ch == chCur)
1119	return 1;
1120
1121	}
1122	return 0;
1123	}
1124
1125
1126
1127	/**
1128	* I/O Control inner worker (tracing reasons).
1129	*
1130	* @returns IPRT status code.
1131	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1132	*
1133	* @param uIOCtl Function number.
1134	* @param pDevExt Device extention.
1135	* @param pSession Session data.
1136	* @param pReqHdr The request header.
1137	*/
1138	static int supdrvIOCtlInnerUnrestricted(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1139	{
1140	/*
1141	* Validation macros
1142	*/
1143	#define REQ_CHECK_SIZES_EX(Name, cbInExpect, cbOutExpect) \
1144	do { \
1145	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect) \|\| pReqHdr->cbOut != (cbOutExpect))) \
1146	{ \
1147	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld. cbOut=%ld expected %ld.\n", \
1148	(long)pReqHdr->cbIn, (long)(cbInExpect), (long)pReqHdr->cbOut, (long)(cbOutExpect))); \
1149	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1150	} \
1151	} while (0)
1152
1153	#define REQ_CHECK_SIZES(Name) REQ_CHECK_SIZES_EX(Name, Name ## _SIZE_IN, Name ## _SIZE_OUT)
1154
1155	#define REQ_CHECK_SIZE_IN(Name, cbInExpect) \
1156	do { \
1157	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect))) \
1158	{ \
1159	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld.\n", \
1160	(long)pReqHdr->cbIn, (long)(cbInExpect))); \
1161	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1162	} \
1163	} while (0)
1164
1165	#define REQ_CHECK_SIZE_OUT(Name, cbOutExpect) \
1166	do { \
1167	if (RT_UNLIKELY(pReqHdr->cbOut != (cbOutExpect))) \
1168	{ \
1169	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbOut=%ld expected %ld.\n", \
1170	(long)pReqHdr->cbOut, (long)(cbOutExpect))); \
1171	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1172	} \
1173	} while (0)
1174
1175	#define REQ_CHECK_EXPR(Name, expr) \
1176	do { \
1177	if (RT_UNLIKELY(!(expr))) \
1178	{ \
1179	OSDBGPRINT(( #Name ": %s\n", #expr)); \
1180	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1181	} \
1182	} while (0)
1183
1184	#define REQ_CHECK_EXPR_FMT(expr, fmt) \
1185	do { \
1186	if (RT_UNLIKELY(!(expr))) \
1187	{ \
1188	OSDBGPRINT( fmt ); \
1189	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1190	} \
1191	} while (0)
1192
1193	/*
1194	* The switch.
1195	*/
1196	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
1197	{
1198	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
1199	{
1200	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
1201	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
1202	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
1203	{
1204	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
1205	pReq->Hdr.rc = VERR_INVALID_MAGIC;
1206	return 0;
1207	}
1208
1209	#if 0
1210	/*
1211	* Call out to the OS specific code and let it do permission checks on the
1212	* client process.
1213	*/
1214	if (!supdrvOSValidateClientProcess(pDevExt, pSession))
1215	{
1216	pReq->u.Out.u32Cookie = 0xffffffff;
1217	pReq->u.Out.u32SessionCookie = 0xffffffff;
1218	pReq->u.Out.u32SessionVersion = 0xffffffff;
1219	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1220	pReq->u.Out.pSession = NULL;
1221	pReq->u.Out.cFunctions = 0;
1222	pReq->Hdr.rc = VERR_PERMISSION_DENIED;
1223	return 0;
1224	}
1225	#endif
1226
1227	/*
1228	* Match the version.
1229	* The current logic is very simple, match the major interface version.
1230	*/
1231	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
1232	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
1233	{
1234	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1235	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
1236	pReq->u.Out.u32Cookie = 0xffffffff;
1237	pReq->u.Out.u32SessionCookie = 0xffffffff;
1238	pReq->u.Out.u32SessionVersion = 0xffffffff;
1239	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1240	pReq->u.Out.pSession = NULL;
1241	pReq->u.Out.cFunctions = 0;
1242	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1243	return 0;
1244	}
1245
1246	/*
1247	* Fill in return data and be gone.
1248	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
1249	* u32SessionVersion <= u32ReqVersion!
1250	*/
1251	/** @todo Somehow validate the client and negotiate a secure cookie... */
1252	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
1253	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
1254	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
1255	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1256	pReq->u.Out.pSession = pSession;
1257	pReq->u.Out.cFunctions = sizeof(g_aFunctions) / sizeof(g_aFunctions[0]);
1258	pReq->Hdr.rc = VINF_SUCCESS;
1259	return 0;
1260	}
1261
1262	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_QUERY_FUNCS(0)):
1263	{
1264	/* validate */
1265	PSUPQUERYFUNCS pReq = (PSUPQUERYFUNCS)pReqHdr;
1266	REQ_CHECK_SIZES_EX(SUP_IOCTL_QUERY_FUNCS, SUP_IOCTL_QUERY_FUNCS_SIZE_IN, SUP_IOCTL_QUERY_FUNCS_SIZE_OUT(RT_ELEMENTS(g_aFunctions)));
1267
1268	/* execute */
1269	pReq->u.Out.cFunctions = RT_ELEMENTS(g_aFunctions);
1270	memcpy(&pReq->u.Out.aFunctions[0], g_aFunctions, sizeof(g_aFunctions));
1271	pReq->Hdr.rc = VINF_SUCCESS;
1272	return 0;
1273	}
1274
1275	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_LOCK):
1276	{
1277	/* validate */
1278	PSUPPAGELOCK pReq = (PSUPPAGELOCK)pReqHdr;
1279	REQ_CHECK_SIZE_IN(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_IN);
1280	REQ_CHECK_SIZE_OUT(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_OUT(pReq->u.In.cPages));
1281	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.cPages > 0);
1282	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.pvR3 >= PAGE_SIZE);
1283
1284	/* execute */
1285	pReq->Hdr.rc = SUPR0LockMem(pSession, pReq->u.In.pvR3, pReq->u.In.cPages, &pReq->u.Out.aPages[0]);
1286	if (RT_FAILURE(pReq->Hdr.rc))
1287	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1288	return 0;
1289	}
1290
1291	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_UNLOCK):
1292	{
1293	/* validate */
1294	PSUPPAGEUNLOCK pReq = (PSUPPAGEUNLOCK)pReqHdr;
1295	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_UNLOCK);
1296
1297	/* execute */
1298	pReq->Hdr.rc = SUPR0UnlockMem(pSession, pReq->u.In.pvR3);
1299	return 0;
1300	}
1301
1302	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_ALLOC):
1303	{
1304	/* validate */
1305	PSUPCONTALLOC pReq = (PSUPCONTALLOC)pReqHdr;
1306	REQ_CHECK_SIZES(SUP_IOCTL_CONT_ALLOC);
1307
1308	/* execute */
1309	pReq->Hdr.rc = SUPR0ContAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.HCPhys);
1310	if (RT_FAILURE(pReq->Hdr.rc))
1311	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1312	return 0;
1313	}
1314
1315	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_FREE):
1316	{
1317	/* validate */
1318	PSUPCONTFREE pReq = (PSUPCONTFREE)pReqHdr;
1319	REQ_CHECK_SIZES(SUP_IOCTL_CONT_FREE);
1320
1321	/* execute */
1322	pReq->Hdr.rc = SUPR0ContFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1323	return 0;
1324	}
1325
1326	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_OPEN):
1327	{
1328	/* validate */
1329	PSUPLDROPEN pReq = (PSUPLDROPEN)pReqHdr;
1330	REQ_CHECK_SIZES(SUP_IOCTL_LDR_OPEN);
1331	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs > 0);
1332	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs < 16*_1M);
1333	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1334	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1335	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits < pReq->u.In.cbImageWithTabs);
1336	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.szName[0]);
1337	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1338	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, !supdrvCheckInvalidChar(pReq->u.In.szName, ";:()[]{}/\\\|&*%#@!~`\"'"));
1339	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szFilename, sizeof(pReq->u.In.szFilename)));
1340
1341	/* execute */
1342	pReq->Hdr.rc = supdrvIOCtl_LdrOpen(pDevExt, pSession, pReq);
1343	return 0;
1344	}
1345
1346	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOAD):
1347	{
1348	/* validate */
1349	PSUPLDRLOAD pReq = (PSUPLDRLOAD)pReqHdr;
1350	REQ_CHECK_EXPR(Name, pReq->Hdr.cbIn >= sizeof(*pReq));
1351	REQ_CHECK_SIZES_EX(SUP_IOCTL_LDR_LOAD, SUP_IOCTL_LDR_LOAD_SIZE_IN(pReq->u.In.cbImageWithTabs), SUP_IOCTL_LDR_LOAD_SIZE_OUT);
1352	REQ_CHECK_EXPR(SUP_IOCTL_LDR_LOAD, pReq->u.In.cSymbols <= 16384);
1353	REQ_CHECK_EXPR_FMT( !pReq->u.In.cSymbols
1354	\|\| ( pReq->u.In.offSymbols < pReq->u.In.cbImageWithTabs
1355	&& pReq->u.In.offSymbols + pReq->u.In.cSymbols * sizeof(SUPLDRSYM) <= pReq->u.In.cbImageWithTabs),
1356	("SUP_IOCTL_LDR_LOAD: offSymbols=%#lx cSymbols=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offSymbols,
1357	(long)pReq->u.In.cSymbols, (long)pReq->u.In.cbImageWithTabs));
1358	REQ_CHECK_EXPR_FMT( !pReq->u.In.cbStrTab
1359	\|\| ( pReq->u.In.offStrTab < pReq->u.In.cbImageWithTabs
1360	&& pReq->u.In.offStrTab + pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs
1361	&& pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs),
1362	("SUP_IOCTL_LDR_LOAD: offStrTab=%#lx cbStrTab=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offStrTab,
1363	(long)pReq->u.In.cbStrTab, (long)pReq->u.In.cbImageWithTabs));
1364
1365	if (pReq->u.In.cSymbols)
1366	{
1367	uint32_t i;
1368	PSUPLDRSYM paSyms = (PSUPLDRSYM)&pReq->u.In.abImage[pReq->u.In.offSymbols];
1369	for (i = 0; i < pReq->u.In.cSymbols; i++)
1370	{
1371	REQ_CHECK_EXPR_FMT(paSyms[i].offSymbol < pReq->u.In.cbImageWithTabs,
1372	("SUP_IOCTL_LDR_LOAD: sym #%ld: symb off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offSymbol, (long)pReq->u.In.cbImageWithTabs));
1373	REQ_CHECK_EXPR_FMT(paSyms[i].offName < pReq->u.In.cbStrTab,
1374	("SUP_IOCTL_LDR_LOAD: sym #%ld: name off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1375	REQ_CHECK_EXPR_FMT(RTStrEnd((char const *)&pReq->u.In.abImage[pReq->u.In.offStrTab + paSyms[i].offName],
1376	pReq->u.In.cbStrTab - paSyms[i].offName),
1377	("SUP_IOCTL_LDR_LOAD: sym #%ld: unterminated name! (%#lx / %#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1378	}
1379	}
1380
1381	/* execute */
1382	pReq->Hdr.rc = supdrvIOCtl_LdrLoad(pDevExt, pSession, pReq);
1383	return 0;
1384	}
1385
1386	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_FREE):
1387	{
1388	/* validate */
1389	PSUPLDRFREE pReq = (PSUPLDRFREE)pReqHdr;
1390	REQ_CHECK_SIZES(SUP_IOCTL_LDR_FREE);
1391
1392	/* execute */
1393	pReq->Hdr.rc = supdrvIOCtl_LdrFree(pDevExt, pSession, pReq);
1394	return 0;
1395	}
1396
1397	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_GET_SYMBOL):
1398	{
1399	/* validate */
1400	PSUPLDRGETSYMBOL pReq = (PSUPLDRGETSYMBOL)pReqHdr;
1401	REQ_CHECK_SIZES(SUP_IOCTL_LDR_GET_SYMBOL);
1402	REQ_CHECK_EXPR(SUP_IOCTL_LDR_GET_SYMBOL, RTStrEnd(pReq->u.In.szSymbol, sizeof(pReq->u.In.szSymbol)));
1403
1404	/* execute */
1405	pReq->Hdr.rc = supdrvIOCtl_LdrGetSymbol(pDevExt, pSession, pReq);
1406	return 0;
1407	}
1408
1409	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0(0)):
1410	{
1411	/* validate */
1412	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1413	Log4(("SUP_IOCTL_CALL_VMMR0: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1414	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1415
1416	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_VMMR0_SIZE(0))
1417	{
1418	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(0), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(0));
1419
1420	/* execute */
1421	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1422	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg, pSession);
1423	else
1424	pReq->Hdr.rc = VERR_WRONG_ORDER;
1425	}
1426	else
1427	{
1428	PSUPVMMR0REQHDR pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1429	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR)),
1430	("SUP_IOCTL_CALL_VMMR0: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR))));
1431	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1432	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(pVMMReq->cbReq));
1433
1434	/* execute */
1435	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1436	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1437	else
1438	pReq->Hdr.rc = VERR_WRONG_ORDER;
1439	}
1440
1441	if ( RT_FAILURE(pReq->Hdr.rc)
1442	&& pReq->Hdr.rc != VERR_INTERRUPTED
1443	&& pReq->Hdr.rc != VERR_TIMEOUT)
1444	Log(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1445	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1446	else
1447	Log4(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1448	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1449	return 0;
1450	}
1451
1452	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0_BIG):
1453	{
1454	/* validate */
1455	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1456	PSUPVMMR0REQHDR pVMMReq;
1457	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1458	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1459
1460	pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1461	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR)),
1462	("SUP_IOCTL_CALL_VMMR0_BIG: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR))));
1463	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0_BIG, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1464	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));
1465
1466	/* execute */
1467	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1468	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1469	else
1470	pReq->Hdr.rc = VERR_WRONG_ORDER;
1471
1472	if ( RT_FAILURE(pReq->Hdr.rc)
1473	&& pReq->Hdr.rc != VERR_INTERRUPTED
1474	&& pReq->Hdr.rc != VERR_TIMEOUT)
1475	Log(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1476	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1477	else
1478	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1479	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1480	return 0;
1481	}
1482
1483	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_PAGING_MODE):
1484	{
1485	/* validate */
1486	PSUPGETPAGINGMODE pReq = (PSUPGETPAGINGMODE)pReqHdr;
1487	REQ_CHECK_SIZES(SUP_IOCTL_GET_PAGING_MODE);
1488
1489	/* execute */
1490	pReq->Hdr.rc = VINF_SUCCESS;
1491	pReq->u.Out.enmMode = SUPR0GetPagingMode();
1492	return 0;
1493	}
1494
1495	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_ALLOC):
1496	{
1497	/* validate */
1498	PSUPLOWALLOC pReq = (PSUPLOWALLOC)pReqHdr;
1499	REQ_CHECK_EXPR(SUP_IOCTL_LOW_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_LOW_ALLOC_SIZE_IN);
1500	REQ_CHECK_SIZES_EX(SUP_IOCTL_LOW_ALLOC, SUP_IOCTL_LOW_ALLOC_SIZE_IN, SUP_IOCTL_LOW_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1501
1502	/* execute */
1503	pReq->Hdr.rc = SUPR0LowAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1504	if (RT_FAILURE(pReq->Hdr.rc))
1505	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1506	return 0;
1507	}
1508
1509	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_FREE):
1510	{
1511	/* validate */
1512	PSUPLOWFREE pReq = (PSUPLOWFREE)pReqHdr;
1513	REQ_CHECK_SIZES(SUP_IOCTL_LOW_FREE);
1514
1515	/* execute */
1516	pReq->Hdr.rc = SUPR0LowFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1517	return 0;
1518	}
1519
1520	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_MAP):
1521	{
1522	/* validate */
1523	PSUPGIPMAP pReq = (PSUPGIPMAP)pReqHdr;
1524	REQ_CHECK_SIZES(SUP_IOCTL_GIP_MAP);
1525
1526	/* execute */
1527	pReq->Hdr.rc = SUPR0GipMap(pSession, &pReq->u.Out.pGipR3, &pReq->u.Out.HCPhysGip);
1528	if (RT_SUCCESS(pReq->Hdr.rc))
1529	pReq->u.Out.pGipR0 = pDevExt->pGip;
1530	return 0;
1531	}
1532
1533	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_UNMAP):
1534	{
1535	/* validate */
1536	PSUPGIPUNMAP pReq = (PSUPGIPUNMAP)pReqHdr;
1537	REQ_CHECK_SIZES(SUP_IOCTL_GIP_UNMAP);
1538
1539	/* execute */
1540	pReq->Hdr.rc = SUPR0GipUnmap(pSession);
1541	return 0;
1542	}
1543
1544	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SET_VM_FOR_FAST):
1545	{
1546	/* validate */
1547	PSUPSETVMFORFAST pReq = (PSUPSETVMFORFAST)pReqHdr;
1548	REQ_CHECK_SIZES(SUP_IOCTL_SET_VM_FOR_FAST);
1549	REQ_CHECK_EXPR_FMT( !pReq->u.In.pVMR0
1550	\|\| ( VALID_PTR(pReq->u.In.pVMR0)
1551	&& !((uintptr_t)pReq->u.In.pVMR0 & (PAGE_SIZE - 1))),
1552	("SUP_IOCTL_SET_VM_FOR_FAST: pVMR0=%p!\n", pReq->u.In.pVMR0));
1553	/* execute */
1554	pSession->pVM = pReq->u.In.pVMR0;
1555	pReq->Hdr.rc = VINF_SUCCESS;
1556	return 0;
1557	}
1558
1559	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_ALLOC_EX):
1560	{
1561	/* validate */
1562	PSUPPAGEALLOCEX pReq = (PSUPPAGEALLOCEX)pReqHdr;
1563	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC_EX, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_EX_SIZE_IN);
1564	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));
1565	REQ_CHECK_EXPR_FMT(pReq->u.In.fKernelMapping \|\| pReq->u.In.fUserMapping,
1566	("SUP_IOCTL_PAGE_ALLOC_EX: No mapping requested!\n"));
1567	REQ_CHECK_EXPR_FMT(pReq->u.In.fUserMapping,
1568	("SUP_IOCTL_PAGE_ALLOC_EX: Must have user mapping!\n"));
1569	REQ_CHECK_EXPR_FMT(!pReq->u.In.fReserved0 && !pReq->u.In.fReserved1,
1570	("SUP_IOCTL_PAGE_ALLOC_EX: fReserved0=%d fReserved1=%d\n", pReq->u.In.fReserved0, pReq->u.In.fReserved1));
1571
1572	/* execute */
1573	pReq->Hdr.rc = SUPR0PageAllocEx(pSession, pReq->u.In.cPages, 0 /* fFlags */,
1574	pReq->u.In.fUserMapping ? &pReq->u.Out.pvR3 : NULL,
1575	pReq->u.In.fKernelMapping ? &pReq->u.Out.pvR0 : NULL,
1576	&pReq->u.Out.aPages[0]);
1577	if (RT_FAILURE(pReq->Hdr.rc))
1578	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1579	return 0;
1580	}
1581
1582	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_MAP_KERNEL):
1583	{
1584	/* validate */
1585	PSUPPAGEMAPKERNEL pReq = (PSUPPAGEMAPKERNEL)pReqHdr;
1586	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_MAP_KERNEL);
1587	REQ_CHECK_EXPR_FMT(!pReq->u.In.fFlags, ("SUP_IOCTL_PAGE_MAP_KERNEL: fFlags=%#x! MBZ\n", pReq->u.In.fFlags));
1588	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_MAP_KERNEL: offSub=%#x\n", pReq->u.In.offSub));
1589	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1590	("SUP_IOCTL_PAGE_MAP_KERNEL: cbSub=%#x\n", pReq->u.In.cbSub));
1591
1592	/* execute */
1593	pReq->Hdr.rc = SUPR0PageMapKernel(pSession, pReq->u.In.pvR3, pReq->u.In.offSub, pReq->u.In.cbSub,
1594	pReq->u.In.fFlags, &pReq->u.Out.pvR0);
1595	if (RT_FAILURE(pReq->Hdr.rc))
1596	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1597	return 0;
1598	}
1599
1600	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_PROTECT):
1601	{
1602	/* validate */
1603	PSUPPAGEPROTECT pReq = (PSUPPAGEPROTECT)pReqHdr;
1604	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_PROTECT);
1605	REQ_CHECK_EXPR_FMT(!(pReq->u.In.fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)),
1606	("SUP_IOCTL_PAGE_PROTECT: fProt=%#x!\n", pReq->u.In.fProt));
1607	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_PROTECT: offSub=%#x\n", pReq->u.In.offSub));
1608	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1609	("SUP_IOCTL_PAGE_PROTECT: cbSub=%#x\n", pReq->u.In.cbSub));
1610
1611	/* execute */
1612	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);
1613	return 0;
1614	}
1615
1616	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_FREE):
1617	{
1618	/* validate */
1619	PSUPPAGEFREE pReq = (PSUPPAGEFREE)pReqHdr;
1620	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_FREE);
1621
1622	/* execute */
1623	pReq->Hdr.rc = SUPR0PageFree(pSession, pReq->u.In.pvR3);
1624	return 0;
1625	}
1626
1627	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_SERVICE(0)):
1628	{
1629	/* validate */
1630	PSUPCALLSERVICE pReq = (PSUPCALLSERVICE)pReqHdr;
1631	Log4(("SUP_IOCTL_CALL_SERVICE: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1632	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1633
1634	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
1635	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(0), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(0));
1636	else
1637	{
1638	PSUPR0SERVICEREQHDR pSrvReq = (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0];
1639	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR)),
1640	("SUP_IOCTL_CALL_SERVICE: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR))));
1641	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, pSrvReq->u32Magic == SUPR0SERVICEREQHDR_MAGIC);
1642	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(pSrvReq->cbReq), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(pSrvReq->cbReq));
1643	}
1644	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1645
1646	/* execute */
1647	pReq->Hdr.rc = supdrvIOCtl_CallServiceModule(pDevExt, pSession, pReq);
1648	return 0;
1649	}
1650
1651	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOGGER_SETTINGS(0)):
1652	{
1653	/* validate */
1654	PSUPLOGGERSETTINGS pReq = (PSUPLOGGERSETTINGS)pReqHdr;
1655	size_t cbStrTab;
1656	REQ_CHECK_SIZE_OUT(SUP_IOCTL_LOGGER_SETTINGS, SUP_IOCTL_LOGGER_SETTINGS_SIZE_OUT);
1657	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->Hdr.cbIn >= SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(1));
1658	cbStrTab = pReq->Hdr.cbIn - SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(0);
1659	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offGroups < cbStrTab);
1660	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offFlags < cbStrTab);
1661	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offDestination < cbStrTab);
1662	REQ_CHECK_EXPR_FMT(pReq->u.In.szStrings[cbStrTab - 1] == '\0',
1663	("SUP_IOCTL_LOGGER_SETTINGS: cbIn=%#x cbStrTab=%#zx LastChar=%d\n",
1664	pReq->Hdr.cbIn, cbStrTab, pReq->u.In.szStrings[cbStrTab - 1]));
1665	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhich <= SUPLOGGERSETTINGS_WHICH_RELEASE);
1666	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhat <= SUPLOGGERSETTINGS_WHAT_DESTROY);
1667
1668	/* execute */
1669	pReq->Hdr.rc = supdrvIOCtl_LoggerSettings(pDevExt, pSession, pReq);
1670	return 0;
1671	}
1672
1673	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP2):
1674	{
1675	/* validate */
1676	PSUPSEMOP2 pReq = (PSUPSEMOP2)pReqHdr;
1677	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP2, SUP_IOCTL_SEM_OP2_SIZE_IN, SUP_IOCTL_SEM_OP2_SIZE_OUT);
1678	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP2, pReq->u.In.uReserved == 0);
1679
1680	/* execute */
1681	switch (pReq->u.In.uType)
1682	{
1683	case SUP_SEM_TYPE_EVENT:
1684	{
1685	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1686	switch (pReq->u.In.uOp)
1687	{
1688	case SUPSEMOP2_WAIT_MS_REL:
1689	pReq->Hdr.rc = SUPSemEventWaitNoResume(pSession, hEvent, pReq->u.In.uArg.cRelMsTimeout);
1690	break;
1691	case SUPSEMOP2_WAIT_NS_ABS:
1692	pReq->Hdr.rc = SUPSemEventWaitNsAbsIntr(pSession, hEvent, pReq->u.In.uArg.uAbsNsTimeout);
1693	break;
1694	case SUPSEMOP2_WAIT_NS_REL:
1695	pReq->Hdr.rc = SUPSemEventWaitNsRelIntr(pSession, hEvent, pReq->u.In.uArg.cRelNsTimeout);
1696	break;
1697	case SUPSEMOP2_SIGNAL:
1698	pReq->Hdr.rc = SUPSemEventSignal(pSession, hEvent);
1699	break;
1700	case SUPSEMOP2_CLOSE:
1701	pReq->Hdr.rc = SUPSemEventClose(pSession, hEvent);
1702	break;
1703	case SUPSEMOP2_RESET:
1704	default:
1705	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1706	break;
1707	}
1708	break;
1709	}
1710
1711	case SUP_SEM_TYPE_EVENT_MULTI:
1712	{
1713	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1714	switch (pReq->u.In.uOp)
1715	{
1716	case SUPSEMOP2_WAIT_MS_REL:
1717	pReq->Hdr.rc = SUPSemEventMultiWaitNoResume(pSession, hEventMulti, pReq->u.In.uArg.cRelMsTimeout);
1718	break;
1719	case SUPSEMOP2_WAIT_NS_ABS:
1720	pReq->Hdr.rc = SUPSemEventMultiWaitNsAbsIntr(pSession, hEventMulti, pReq->u.In.uArg.uAbsNsTimeout);
1721	break;
1722	case SUPSEMOP2_WAIT_NS_REL:
1723	pReq->Hdr.rc = SUPSemEventMultiWaitNsRelIntr(pSession, hEventMulti, pReq->u.In.uArg.cRelNsTimeout);
1724	break;
1725	case SUPSEMOP2_SIGNAL:
1726	pReq->Hdr.rc = SUPSemEventMultiSignal(pSession, hEventMulti);
1727	break;
1728	case SUPSEMOP2_CLOSE:
1729	pReq->Hdr.rc = SUPSemEventMultiClose(pSession, hEventMulti);
1730	break;
1731	case SUPSEMOP2_RESET:
1732	pReq->Hdr.rc = SUPSemEventMultiReset(pSession, hEventMulti);
1733	break;
1734	default:
1735	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1736	break;
1737	}
1738	break;
1739	}
1740
1741	default:
1742	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1743	break;
1744	}
1745	return 0;
1746	}
1747
1748	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP3):
1749	{
1750	/* validate */
1751	PSUPSEMOP3 pReq = (PSUPSEMOP3)pReqHdr;
1752	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP3, SUP_IOCTL_SEM_OP3_SIZE_IN, SUP_IOCTL_SEM_OP3_SIZE_OUT);
1753	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, pReq->u.In.u32Reserved == 0 && pReq->u.In.u64Reserved == 0);
1754
1755	/* execute */
1756	switch (pReq->u.In.uType)
1757	{
1758	case SUP_SEM_TYPE_EVENT:
1759	{
1760	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1761	switch (pReq->u.In.uOp)
1762	{
1763	case SUPSEMOP3_CREATE:
1764	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1765	pReq->Hdr.rc = SUPSemEventCreate(pSession, &hEvent);
1766	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEvent;
1767	break;
1768	case SUPSEMOP3_GET_RESOLUTION:
1769	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1770	pReq->Hdr.rc = VINF_SUCCESS;
1771	pReq->Hdr.cbOut = sizeof(*pReq);
1772	pReq->u.Out.cNsResolution = SUPSemEventGetResolution(pSession);
1773	break;
1774	default:
1775	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1776	break;
1777	}
1778	break;
1779	}
1780
1781	case SUP_SEM_TYPE_EVENT_MULTI:
1782	{
1783	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1784	switch (pReq->u.In.uOp)
1785	{
1786	case SUPSEMOP3_CREATE:
1787	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1788	pReq->Hdr.rc = SUPSemEventMultiCreate(pSession, &hEventMulti);
1789	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEventMulti;
1790	break;
1791	case SUPSEMOP3_GET_RESOLUTION:
1792	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1793	pReq->Hdr.rc = VINF_SUCCESS;
1794	pReq->u.Out.cNsResolution = SUPSemEventMultiGetResolution(pSession);
1795	break;
1796	default:
1797	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1798	break;
1799	}
1800	break;
1801	}
1802
1803	default:
1804	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1805	break;
1806	}
1807	return 0;
1808	}
1809
1810	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
1811	{
1812	/* validate */
1813	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
1814	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
1815
1816	/* execute */
1817	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.Caps);
1818	if (RT_FAILURE(pReq->Hdr.rc))
1819	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1820	return 0;
1821	}
1822
1823	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_OPEN):
1824	{
1825	/* validate */
1826	PSUPTRACEROPEN pReq = (PSUPTRACEROPEN)pReqHdr;
1827	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_OPEN);
1828
1829	/* execute */
1830	pReq->Hdr.rc = supdrvIOCtl_TracerOpen(pDevExt, pSession, pReq->u.In.uCookie, pReq->u.In.uArg);
1831	return 0;
1832	}
1833
1834	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_CLOSE):
1835	{
1836	/* validate */
1837	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_CLOSE);
1838
1839	/* execute */
1840	pReqHdr->rc = supdrvIOCtl_TracerClose(pDevExt, pSession);
1841	return 0;
1842	}
1843
1844	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_IOCTL):
1845	{
1846	/* validate */
1847	PSUPTRACERIOCTL pReq = (PSUPTRACERIOCTL)pReqHdr;
1848	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_IOCTL);
1849
1850	/* execute */
1851	pReqHdr->rc = supdrvIOCtl_TracerIOCtl(pDevExt, pSession, pReq->u.In.uCmd, pReq->u.In.uArg, &pReq->u.Out.iRetVal);
1852	return 0;
1853	}
1854
1855	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_REG):
1856	{
1857	/* validate */
1858	PSUPTRACERUMODREG pReq = (PSUPTRACERUMODREG)pReqHdr;
1859	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_REG);
1860	if (!RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)))
1861	return VERR_INVALID_PARAMETER;
1862
1863	/* execute */
1864	pReqHdr->rc = supdrvIOCtl_TracerUmodRegister(pDevExt, pSession,
1865	pReq->u.In.R3PtrVtgHdr, pReq->u.In.uVtgHdrAddr,
1866	pReq->u.In.R3PtrStrTab, pReq->u.In.cbStrTab,
1867	pReq->u.In.szName, pReq->u.In.fFlags);
1868	return 0;
1869	}
1870
1871	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_DEREG):
1872	{
1873	/* validate */
1874	PSUPTRACERUMODDEREG pReq = (PSUPTRACERUMODDEREG)pReqHdr;
1875	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_DEREG);
1876
1877	/* execute */
1878	pReqHdr->rc = supdrvIOCtl_TracerUmodDeregister(pDevExt, pSession, pReq->u.In.pVtgHdr);
1879	return 0;
1880	}
1881
1882	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_FIRE_PROBE):
1883	{
1884	/* validate */
1885	PSUPTRACERUMODFIREPROBE pReq = (PSUPTRACERUMODFIREPROBE)pReqHdr;
1886	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_FIRE_PROBE);
1887
1888	supdrvIOCtl_TracerUmodProbeFire(pDevExt, pSession, &pReq->u.In);
1889	pReqHdr->rc = VINF_SUCCESS;
1890	return 0;
1891	}
1892
1893	default:
1894	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1895	break;
1896	}
1897	return VERR_GENERAL_FAILURE;
1898	}
1899
1900
1901	/**
1902	* I/O Control inner worker for the restricted operations.
1903	*
1904	* @returns IPRT status code.
1905	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1906	*
1907	* @param uIOCtl Function number.
1908	* @param pDevExt Device extention.
1909	* @param pSession Session data.
1910	* @param pReqHdr The request header.
1911	*/
1912	static int supdrvIOCtlInnerRestricted(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1913	{
1914	/*
1915	* The switch.
1916	*/
1917	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
1918	{
1919	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
1920	{
1921	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
1922	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
1923	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
1924	{
1925	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
1926	pReq->Hdr.rc = VERR_INVALID_MAGIC;
1927	return 0;
1928	}
1929
1930	/*
1931	* Match the version.
1932	* The current logic is very simple, match the major interface version.
1933	*/
1934	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
1935	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
1936	{
1937	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1938	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
1939	pReq->u.Out.u32Cookie = 0xffffffff;
1940	pReq->u.Out.u32SessionCookie = 0xffffffff;
1941	pReq->u.Out.u32SessionVersion = 0xffffffff;
1942	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1943	pReq->u.Out.pSession = NULL;
1944	pReq->u.Out.cFunctions = 0;
1945	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1946	return 0;
1947	}
1948
1949	/*
1950	* Fill in return data and be gone.
1951	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
1952	* u32SessionVersion <= u32ReqVersion!
1953	*/
1954	/** @todo Somehow validate the client and negotiate a secure cookie... */
1955	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
1956	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
1957	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
1958	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1959	pReq->u.Out.pSession = pSession;
1960	pReq->u.Out.cFunctions = 0;
1961	pReq->Hdr.rc = VINF_SUCCESS;
1962	return 0;
1963	}
1964
1965	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
1966	{
1967	/* validate */
1968	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
1969	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
1970
1971	/* execute */
1972	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.Caps);
1973	if (RT_FAILURE(pReq->Hdr.rc))
1974	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1975	return 0;
1976	}
1977
1978	default:
1979	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1980	break;
1981	}
1982	return VERR_GENERAL_FAILURE;
1983	}
1984
1985
1986	/**
1987	* I/O Control worker.
1988	*
1989	* @returns IPRT status code.
1990	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1991	*
1992	* @param uIOCtl Function number.
1993	* @param pDevExt Device extention.
1994	* @param pSession Session data.
1995	* @param pReqHdr The request header.
1996	*/
1997	int VBOXCALL supdrvIOCtl(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1998	{
1999	int rc;
2000	VBOXDRV_IOCTL_ENTRY(pSession, uIOCtl, pReqHdr);
2001
2002	/*
2003	* Validate the request.
2004	*/
2005	/* this first check could probably be omitted as its also done by the OS specific code... */
2006	if (RT_UNLIKELY( (pReqHdr->fFlags & SUPREQHDR_FLAGS_MAGIC_MASK) != SUPREQHDR_FLAGS_MAGIC
2007	\|\| pReqHdr->cbIn < sizeof(*pReqHdr)
2008	\|\| pReqHdr->cbOut < sizeof(*pReqHdr)))
2009	{
2010	OSDBGPRINT(("vboxdrv: Bad ioctl request header; cbIn=%#lx cbOut=%#lx fFlags=%#lx\n",
2011	(long)pReqHdr->cbIn, (long)pReqHdr->cbOut, (long)pReqHdr->fFlags));
2012	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
2013	return VERR_INVALID_PARAMETER;
2014	}
2015	if (RT_UNLIKELY(!RT_VALID_PTR(pSession)))
2016	{
2017	OSDBGPRINT(("vboxdrv: Invalid pSession valud %p (ioctl=%p)\n", pSession, (void *)uIOCtl));
2018	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
2019	return VERR_INVALID_PARAMETER;
2020	}
2021	if (RT_UNLIKELY(uIOCtl == SUP_IOCTL_COOKIE))
2022	{
2023	if (pReqHdr->u32Cookie != SUPCOOKIE_INITIAL_COOKIE)
2024	{
2025	OSDBGPRINT(("SUP_IOCTL_COOKIE: bad cookie %#lx\n", (long)pReqHdr->u32Cookie));
2026	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
2027	return VERR_INVALID_PARAMETER;
2028	}
2029	}
2030	else if (RT_UNLIKELY( pReqHdr->u32Cookie != pDevExt->u32Cookie
2031	\|\| pReqHdr->u32SessionCookie != pSession->u32Cookie))
2032	{
2033	OSDBGPRINT(("vboxdrv: bad cookie %#lx / %#lx.\n", (long)pReqHdr->u32Cookie, (long)pReqHdr->u32SessionCookie));
2034	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
2035	return VERR_INVALID_PARAMETER;
2036	}
2037
2038	/*
2039	* Hand it to an inner function to avoid lots of unnecessary return tracepoints.
2040	*/
2041	if (pSession->fUnrestricted)
2042	rc = supdrvIOCtlInnerUnrestricted(uIOCtl, pDevExt, pSession, pReqHdr);
2043	else
2044	rc = supdrvIOCtlInnerRestricted(uIOCtl, pDevExt, pSession, pReqHdr);
2045
2046	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, pReqHdr->rc, rc);
2047	return rc;
2048	}
2049
2050
2051	/**
2052	* Inter-Driver Communication (IDC) worker.
2053	*
2054	* @returns VBox status code.
2055	* @retval VINF_SUCCESS on success.
2056	* @retval VERR_INVALID_PARAMETER if the request is invalid.
2057	* @retval VERR_NOT_SUPPORTED if the request isn't supported.
2058	*
2059	* @param uReq The request (function) code.
2060	* @param pDevExt Device extention.
2061	* @param pSession Session data.
2062	* @param pReqHdr The request header.
2063	*/
2064	int VBOXCALL supdrvIDC(uintptr_t uReq, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQHDR pReqHdr)
2065	{
2066	/*
2067	* The OS specific code has already validated the pSession
2068	* pointer, and the request size being greater or equal to
2069	* size of the header.
2070	*
2071	* So, just check that pSession is a kernel context session.
2072	*/
2073	if (RT_UNLIKELY( pSession
2074	&& pSession->R0Process != NIL_RTR0PROCESS))
2075	return VERR_INVALID_PARAMETER;
2076
2077	/*
2078	* Validation macro.
2079	*/
2080	#define REQ_CHECK_IDC_SIZE(Name, cbExpect) \
2081	do { \
2082	if (RT_UNLIKELY(pReqHdr->cb != (cbExpect))) \
2083	{ \
2084	OSDBGPRINT(( #Name ": Invalid input/output sizes. cb=%ld expected %ld.\n", \
2085	(long)pReqHdr->cb, (long)(cbExpect))); \
2086	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
2087	} \
2088	} while (0)
2089
2090	switch (uReq)
2091	{
2092	case SUPDRV_IDC_REQ_CONNECT:
2093	{
2094	PSUPDRVIDCREQCONNECT pReq = (PSUPDRVIDCREQCONNECT)pReqHdr;
2095	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_CONNECT, sizeof(*pReq));
2096
2097	/*
2098	* Validate the cookie and other input.
2099	*/
2100	if (pReq->Hdr.pSession != NULL)
2101	{
2102	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Hdr.pSession=%p expected NULL!\n", pReq->Hdr.pSession));
2103	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2104	}
2105	if (pReq->u.In.u32MagicCookie != SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE)
2106	{
2107	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: u32MagicCookie=%#x expected %#x!\n",
2108	(unsigned)pReq->u.In.u32MagicCookie, (unsigned)SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE));
2109	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2110	}
2111	if ( pReq->u.In.uMinVersion > pReq->u.In.uReqVersion
2112	\|\| (pReq->u.In.uMinVersion & UINT32_C(0xffff0000)) != (pReq->u.In.uReqVersion & UINT32_C(0xffff0000)))
2113	{
2114	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: uMinVersion=%#x uMaxVersion=%#x doesn't match!\n",
2115	pReq->u.In.uMinVersion, pReq->u.In.uReqVersion));
2116	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2117	}
2118	if (pSession != NULL)
2119	{
2120	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: pSession=%p expected NULL!\n", pSession));
2121	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2122	}
2123
2124	/*
2125	* Match the version.
2126	* The current logic is very simple, match the major interface version.
2127	*/
2128	if ( pReq->u.In.uMinVersion > SUPDRV_IDC_VERSION
2129	\|\| (pReq->u.In.uMinVersion & 0xffff0000) != (SUPDRV_IDC_VERSION & 0xffff0000))
2130	{
2131	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
2132	pReq->u.In.uReqVersion, pReq->u.In.uMinVersion, (unsigned)SUPDRV_IDC_VERSION));
2133	pReq->u.Out.pSession = NULL;
2134	pReq->u.Out.uSessionVersion = 0xffffffff;
2135	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
2136	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
2137	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
2138	return VINF_SUCCESS;
2139	}
2140
2141	pReq->u.Out.pSession = NULL;
2142	pReq->u.Out.uSessionVersion = SUPDRV_IDC_VERSION;
2143	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
2144	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
2145
2146	pReq->Hdr.rc = supdrvCreateSession(pDevExt, false /* fUser /, true /fUnrestricted*/, &pSession);
2147	if (RT_FAILURE(pReq->Hdr.rc))
2148	{
2149	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: failed to create session, rc=%d\n", pReq->Hdr.rc));
2150	return VINF_SUCCESS;
2151	}
2152
2153	pReq->u.Out.pSession = pSession;
2154	pReq->Hdr.pSession = pSession;
2155
2156	return VINF_SUCCESS;
2157	}
2158
2159	case SUPDRV_IDC_REQ_DISCONNECT:
2160	{
2161	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_DISCONNECT, sizeof(*pReqHdr));
2162
2163	supdrvSessionRelease(pSession);
2164	return pReqHdr->rc = VINF_SUCCESS;
2165	}
2166
2167	case SUPDRV_IDC_REQ_GET_SYMBOL:
2168	{
2169	PSUPDRVIDCREQGETSYM pReq = (PSUPDRVIDCREQGETSYM)pReqHdr;
2170	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_GET_SYMBOL, sizeof(*pReq));
2171
2172	pReq->Hdr.rc = supdrvIDC_LdrGetSymbol(pDevExt, pSession, pReq);
2173	return VINF_SUCCESS;
2174	}
2175
2176	case SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY:
2177	{
2178	PSUPDRVIDCREQCOMPREGFACTORY pReq = (PSUPDRVIDCREQCOMPREGFACTORY)pReqHdr;
2179	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY, sizeof(*pReq));
2180
2181	pReq->Hdr.rc = SUPR0ComponentRegisterFactory(pSession, pReq->u.In.pFactory);
2182	return VINF_SUCCESS;
2183	}
2184
2185	case SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY:
2186	{
2187	PSUPDRVIDCREQCOMPDEREGFACTORY pReq = (PSUPDRVIDCREQCOMPDEREGFACTORY)pReqHdr;
2188	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY, sizeof(*pReq));
2189
2190	pReq->Hdr.rc = SUPR0ComponentDeregisterFactory(pSession, pReq->u.In.pFactory);
2191	return VINF_SUCCESS;
2192	}
2193
2194	default:
2195	Log(("Unknown IDC %#lx\n", (long)uReq));
2196	break;
2197	}
2198
2199	#undef REQ_CHECK_IDC_SIZE
2200	return VERR_NOT_SUPPORTED;
2201	}
2202
2203
2204	/**
2205	* Register a object for reference counting.
2206	* The object is registered with one reference in the specified session.
2207	*
2208	* @returns Unique identifier on success (pointer).
2209	* All future reference must use this identifier.
2210	* @returns NULL on failure.
2211	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
2212	* @param pvUser1 The first user argument.
2213	* @param pvUser2 The second user argument.
2214	*/
2215	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
2216	{
2217	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2218	PSUPDRVOBJ pObj;
2219	PSUPDRVUSAGE pUsage;
2220
2221	/*
2222	* Validate the input.
2223	*/
2224	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
2225	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
2226	AssertPtrReturn(pfnDestructor, NULL);
2227
2228	/*
2229	* Allocate and initialize the object.
2230	*/
2231	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
2232	if (!pObj)
2233	return NULL;
2234	pObj->u32Magic = SUPDRVOBJ_MAGIC;
2235	pObj->enmType = enmType;
2236	pObj->pNext = NULL;
2237	pObj->cUsage = 1;
2238	pObj->pfnDestructor = pfnDestructor;
2239	pObj->pvUser1 = pvUser1;
2240	pObj->pvUser2 = pvUser2;
2241	pObj->CreatorUid = pSession->Uid;
2242	pObj->CreatorGid = pSession->Gid;
2243	pObj->CreatorProcess= pSession->Process;
2244	supdrvOSObjInitCreator(pObj, pSession);
2245
2246	/*
2247	* Allocate the usage record.
2248	* (We keep freed usage records around to simplify SUPR0ObjAddRefEx().)
2249	*/
2250	RTSpinlockAcquire(pDevExt->Spinlock);
2251
2252	pUsage = pDevExt->pUsageFree;
2253	if (pUsage)
2254	pDevExt->pUsageFree = pUsage->pNext;
2255	else
2256	{
2257	RTSpinlockRelease(pDevExt->Spinlock);
2258	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
2259	if (!pUsage)
2260	{
2261	RTMemFree(pObj);
2262	return NULL;
2263	}
2264	RTSpinlockAcquire(pDevExt->Spinlock);
2265	}
2266
2267	/*
2268	* Insert the object and create the session usage record.
2269	*/
2270	/* The object. */
2271	pObj->pNext = pDevExt->pObjs;
2272	pDevExt->pObjs = pObj;
2273
2274	/* The session record. */
2275	pUsage->cUsage = 1;
2276	pUsage->pObj = pObj;
2277	pUsage->pNext = pSession->pUsage;
2278	/* Log2(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext)); */
2279	pSession->pUsage = pUsage;
2280
2281	RTSpinlockRelease(pDevExt->Spinlock);
2282
2283	Log(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
2284	return pObj;
2285	}
2286
2287
2288	/**
2289	* Increment the reference counter for the object associating the reference
2290	* with the specified session.
2291	*
2292	* @returns IPRT status code.
2293	* @param pvObj The identifier returned by SUPR0ObjRegister().
2294	* @param pSession The session which is referencing the object.
2295	*
2296	* @remarks The caller should not own any spinlocks and must carefully protect
2297	* itself against potential race with the destructor so freed memory
2298	* isn't accessed here.
2299	*/
2300	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
2301	{
2302	return SUPR0ObjAddRefEx(pvObj, pSession, false /* fNoBlocking */);
2303	}
2304
2305
2306	/**
2307	* Increment the reference counter for the object associating the reference
2308	* with the specified session.
2309	*
2310	* @returns IPRT status code.
2311	* @retval VERR_TRY_AGAIN if fNoBlocking was set and a new usage record
2312	* couldn't be allocated. (If you see this you're not doing the right
2313	* thing and it won't ever work reliably.)
2314	*
2315	* @param pvObj The identifier returned by SUPR0ObjRegister().
2316	* @param pSession The session which is referencing the object.
2317	* @param fNoBlocking Set if it's not OK to block. Never try to make the
2318	* first reference to an object in a session with this
2319	* argument set.
2320	*
2321	* @remarks The caller should not own any spinlocks and must carefully protect
2322	* itself against potential race with the destructor so freed memory
2323	* isn't accessed here.
2324	*/
2325	SUPR0DECL(int) SUPR0ObjAddRefEx(void *pvObj, PSUPDRVSESSION pSession, bool fNoBlocking)
2326	{
2327	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2328	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2329	int rc = VINF_SUCCESS;
2330	PSUPDRVUSAGE pUsagePre;
2331	PSUPDRVUSAGE pUsage;
2332
2333	/*
2334	* Validate the input.
2335	* Be ready for the destruction race (someone might be stuck in the
2336	* destructor waiting a lock we own).
2337	*/
2338	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2339	AssertPtrReturn(pObj, VERR_INVALID_POINTER);
2340	AssertMsgReturn(pObj->u32Magic == SUPDRVOBJ_MAGIC \|\| pObj->u32Magic == SUPDRVOBJ_MAGIC_DEAD,
2341	("Invalid pvObj=%p magic=%#x (expected %#x or %#x)\n", pvObj, pObj->u32Magic, SUPDRVOBJ_MAGIC, SUPDRVOBJ_MAGIC_DEAD),
2342	VERR_INVALID_PARAMETER);
2343
2344	RTSpinlockAcquire(pDevExt->Spinlock);
2345
2346	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2347	{
2348	RTSpinlockRelease(pDevExt->Spinlock);
2349
2350	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2351	return VERR_WRONG_ORDER;
2352	}
2353
2354	/*
2355	* Preallocate the usage record if we can.
2356	*/
2357	pUsagePre = pDevExt->pUsageFree;
2358	if (pUsagePre)
2359	pDevExt->pUsageFree = pUsagePre->pNext;
2360	else if (!fNoBlocking)
2361	{
2362	RTSpinlockRelease(pDevExt->Spinlock);
2363	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
2364	if (!pUsagePre)
2365	return VERR_NO_MEMORY;
2366
2367	RTSpinlockAcquire(pDevExt->Spinlock);
2368	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2369	{
2370	RTSpinlockRelease(pDevExt->Spinlock);
2371
2372	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2373	return VERR_WRONG_ORDER;
2374	}
2375	}
2376
2377	/*
2378	* Reference the object.
2379	*/
2380	pObj->cUsage++;
2381
2382	/*
2383	* Look for the session record.
2384	*/
2385	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
2386	{
2387	/Log(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2388	if (pUsage->pObj == pObj)
2389	break;
2390	}
2391	if (pUsage)
2392	pUsage->cUsage++;
2393	else if (pUsagePre)
2394	{
2395	/* create a new session record. */
2396	pUsagePre->cUsage = 1;
2397	pUsagePre->pObj = pObj;
2398	pUsagePre->pNext = pSession->pUsage;
2399	pSession->pUsage = pUsagePre;
2400	/Log(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));/
2401
2402	pUsagePre = NULL;
2403	}
2404	else
2405	{
2406	pObj->cUsage--;
2407	rc = VERR_TRY_AGAIN;
2408	}
2409
2410	/*
2411	* Put any unused usage record into the free list..
2412	*/
2413	if (pUsagePre)
2414	{
2415	pUsagePre->pNext = pDevExt->pUsageFree;
2416	pDevExt->pUsageFree = pUsagePre;
2417	}
2418
2419	RTSpinlockRelease(pDevExt->Spinlock);
2420
2421	return rc;
2422	}
2423
2424
2425	/**
2426	* Decrement / destroy a reference counter record for an object.
2427	*
2428	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
2429	*
2430	* @returns IPRT status code.
2431	* @retval VINF_SUCCESS if not destroyed.
2432	* @retval VINF_OBJECT_DESTROYED if it's destroyed by this release call.
2433	* @retval VERR_INVALID_PARAMETER if the object isn't valid. Will assert in
2434	* string builds.
2435	*
2436	* @param pvObj The identifier returned by SUPR0ObjRegister().
2437	* @param pSession The session which is referencing the object.
2438	*/
2439	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
2440	{
2441	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2442	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2443	int rc = VERR_INVALID_PARAMETER;
2444	PSUPDRVUSAGE pUsage;
2445	PSUPDRVUSAGE pUsagePrev;
2446
2447	/*
2448	* Validate the input.
2449	*/
2450	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2451	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2452	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2453	VERR_INVALID_PARAMETER);
2454
2455	/*
2456	* Acquire the spinlock and look for the usage record.
2457	*/
2458	RTSpinlockAcquire(pDevExt->Spinlock);
2459
2460	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
2461	pUsage;
2462	pUsagePrev = pUsage, pUsage = pUsage->pNext)
2463	{
2464	/Log2(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2465	if (pUsage->pObj == pObj)
2466	{
2467	rc = VINF_SUCCESS;
2468	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
2469	if (pUsage->cUsage > 1)
2470	{
2471	pObj->cUsage--;
2472	pUsage->cUsage--;
2473	}
2474	else
2475	{
2476	/*
2477	* Free the session record.
2478	*/
2479	if (pUsagePrev)
2480	pUsagePrev->pNext = pUsage->pNext;
2481	else
2482	pSession->pUsage = pUsage->pNext;
2483	pUsage->pNext = pDevExt->pUsageFree;
2484	pDevExt->pUsageFree = pUsage;
2485
2486	/* What about the object? */
2487	if (pObj->cUsage > 1)
2488	pObj->cUsage--;
2489	else
2490	{
2491	/*
2492	* Object is to be destroyed, unlink it.
2493	*/
2494	pObj->u32Magic = SUPDRVOBJ_MAGIC_DEAD;
2495	rc = VINF_OBJECT_DESTROYED;
2496	if (pDevExt->pObjs == pObj)
2497	pDevExt->pObjs = pObj->pNext;
2498	else
2499	{
2500	PSUPDRVOBJ pObjPrev;
2501	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
2502	if (pObjPrev->pNext == pObj)
2503	{
2504	pObjPrev->pNext = pObj->pNext;
2505	break;
2506	}
2507	Assert(pObjPrev);
2508	}
2509	}
2510	}
2511	break;
2512	}
2513	}
2514
2515	RTSpinlockRelease(pDevExt->Spinlock);
2516
2517	/*
2518	* Call the destructor and free the object if required.
2519	*/
2520	if (rc == VINF_OBJECT_DESTROYED)
2521	{
2522	Log(("SUPR0ObjRelease: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
2523	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
2524	if (pObj->pfnDestructor)
2525	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
2526	RTMemFree(pObj);
2527	}
2528
2529	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
2530	return rc;
2531	}
2532
2533
2534	/**
2535	* Verifies that the current process can access the specified object.
2536	*
2537	* @returns The following IPRT status code:
2538	* @retval VINF_SUCCESS if access was granted.
2539	* @retval VERR_PERMISSION_DENIED if denied access.
2540	* @retval VERR_INVALID_PARAMETER if invalid parameter.
2541	*
2542	* @param pvObj The identifier returned by SUPR0ObjRegister().
2543	* @param pSession The session which wishes to access the object.
2544	* @param pszObjName Object string name. This is optional and depends on the object type.
2545	*
2546	* @remark The caller is responsible for making sure the object isn't removed while
2547	* we're inside this function. If uncertain about this, just call AddRef before calling us.
2548	*/
2549	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
2550	{
2551	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2552	int rc;
2553
2554	/*
2555	* Validate the input.
2556	*/
2557	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2558	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2559	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2560	VERR_INVALID_PARAMETER);
2561
2562	/*
2563	* Check access. (returns true if a decision has been made.)
2564	*/
2565	rc = VERR_INTERNAL_ERROR;
2566	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
2567	return rc;
2568
2569	/*
2570	* Default policy is to allow the user to access his own
2571	* stuff but nothing else.
2572	*/
2573	if (pObj->CreatorUid == pSession->Uid)
2574	return VINF_SUCCESS;
2575	return VERR_PERMISSION_DENIED;
2576	}
2577
2578
2579	/**
2580	* Lock pages.
2581	*
2582	* @returns IPRT status code.
2583	* @param pSession Session to which the locked memory should be associated.
2584	* @param pvR3 Start of the memory range to lock.
2585	* This must be page aligned.
2586	* @param cPages Number of pages to lock.
2587	* @param paPages Where to put the physical addresses of locked memory.
2588	*/
2589	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
2590	{
2591	int rc;
2592	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2593	const size_t cb = (size_t)cPages << PAGE_SHIFT;
2594	LogFlow(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
2595
2596	/*
2597	* Verify input.
2598	*/
2599	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2600	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
2601	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
2602	\|\| !pvR3)
2603	{
2604	Log(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
2605	return VERR_INVALID_PARAMETER;
2606	}
2607
2608	/*
2609	* Let IPRT do the job.
2610	*/
2611	Mem.eType = MEMREF_TYPE_LOCKED;
2612	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTMEM_PROT_READ \| RTMEM_PROT_WRITE, RTR0ProcHandleSelf());
2613	if (RT_SUCCESS(rc))
2614	{
2615	uint32_t iPage = cPages;
2616	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
2617	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
2618
2619	while (iPage-- > 0)
2620	{
2621	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2622	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
2623	{
2624	AssertMsgFailed(("iPage=%d\n", iPage));
2625	rc = VERR_INTERNAL_ERROR;
2626	break;
2627	}
2628	}
2629	if (RT_SUCCESS(rc))
2630	rc = supdrvMemAdd(&Mem, pSession);
2631	if (RT_FAILURE(rc))
2632	{
2633	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
2634	AssertRC(rc2);
2635	}
2636	}
2637
2638	return rc;
2639	}
2640
2641
2642	/**
2643	* Unlocks the memory pointed to by pv.
2644	*
2645	* @returns IPRT status code.
2646	* @param pSession Session to which the memory was locked.
2647	* @param pvR3 Memory to unlock.
2648	*/
2649	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2650	{
2651	LogFlow(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2652	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2653	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
2654	}
2655
2656
2657	/**
2658	* Allocates a chunk of page aligned memory with contiguous and fixed physical
2659	* backing.
2660	*
2661	* @returns IPRT status code.
2662	* @param pSession Session data.
2663	* @param cPages Number of pages to allocate.
2664	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
2665	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
2666	* @param pHCPhys Where to put the physical address of allocated memory.
2667	*/
2668	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
2669	{
2670	int rc;
2671	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2672	LogFlow(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
2673
2674	/*
2675	* Validate input.
2676	*/
2677	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2678	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
2679	{
2680	Log(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
2681	pSession, ppvR0, ppvR3, pHCPhys));
2682	return VERR_INVALID_PARAMETER;
2683
2684	}
2685	if (cPages < 1 \|\| cPages >= 256)
2686	{
2687	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2688	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2689	}
2690
2691	/*
2692	* Let IPRT do the job.
2693	*/
2694	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
2695	if (RT_SUCCESS(rc))
2696	{
2697	int rc2;
2698	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2699	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2700	if (RT_SUCCESS(rc))
2701	{
2702	Mem.eType = MEMREF_TYPE_CONT;
2703	rc = supdrvMemAdd(&Mem, pSession);
2704	if (!rc)
2705	{
2706	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2707	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2708	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
2709	return 0;
2710	}
2711
2712	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2713	AssertRC(rc2);
2714	}
2715	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2716	AssertRC(rc2);
2717	}
2718
2719	return rc;
2720	}
2721
2722
2723	/**
2724	* Frees memory allocated using SUPR0ContAlloc().
2725	*
2726	* @returns IPRT status code.
2727	* @param pSession The session to which the memory was allocated.
2728	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2729	*/
2730	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2731	{
2732	LogFlow(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2733	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2734	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
2735	}
2736
2737
2738	/**
2739	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
2740	*
2741	* The memory isn't zeroed.
2742	*
2743	* @returns IPRT status code.
2744	* @param pSession Session data.
2745	* @param cPages Number of pages to allocate.
2746	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
2747	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
2748	* @param paPages Where to put the physical addresses of allocated memory.
2749	*/
2750	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
2751	{
2752	unsigned iPage;
2753	int rc;
2754	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2755	LogFlow(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
2756
2757	/*
2758	* Validate input.
2759	*/
2760	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2761	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
2762	{
2763	Log(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
2764	pSession, ppvR3, ppvR0, paPages));
2765	return VERR_INVALID_PARAMETER;
2766
2767	}
2768	if (cPages < 1 \|\| cPages >= 256)
2769	{
2770	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2771	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2772	}
2773
2774	/*
2775	* Let IPRT do the work.
2776	*/
2777	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
2778	if (RT_SUCCESS(rc))
2779	{
2780	int rc2;
2781	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2782	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2783	if (RT_SUCCESS(rc))
2784	{
2785	Mem.eType = MEMREF_TYPE_LOW;
2786	rc = supdrvMemAdd(&Mem, pSession);
2787	if (!rc)
2788	{
2789	for (iPage = 0; iPage < cPages; iPage++)
2790	{
2791	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2792	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%RHp\n", paPages[iPage]));
2793	}
2794	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2795	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2796	return 0;
2797	}
2798
2799	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2800	AssertRC(rc2);
2801	}
2802
2803	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2804	AssertRC(rc2);
2805	}
2806
2807	return rc;
2808	}
2809
2810
2811	/**
2812	* Frees memory allocated using SUPR0LowAlloc().
2813	*
2814	* @returns IPRT status code.
2815	* @param pSession The session to which the memory was allocated.
2816	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2817	*/
2818	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2819	{
2820	LogFlow(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2821	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2822	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
2823	}
2824
2825
2826
2827	/**
2828	* Allocates a chunk of memory with both R0 and R3 mappings.
2829	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
2830	*
2831	* @returns IPRT status code.
2832	* @param pSession The session to associated the allocation with.
2833	* @param cb Number of bytes to allocate.
2834	* @param ppvR0 Where to store the address of the Ring-0 mapping.
2835	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2836	*/
2837	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
2838	{
2839	int rc;
2840	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2841	LogFlow(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
2842
2843	/*
2844	* Validate input.
2845	*/
2846	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2847	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
2848	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
2849	if (cb < 1 \|\| cb >= _4M)
2850	{
2851	Log(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
2852	return VERR_INVALID_PARAMETER;
2853	}
2854
2855	/*
2856	* Let IPRT do the work.
2857	*/
2858	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
2859	if (RT_SUCCESS(rc))
2860	{
2861	int rc2;
2862	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2863	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2864	if (RT_SUCCESS(rc))
2865	{
2866	Mem.eType = MEMREF_TYPE_MEM;
2867	rc = supdrvMemAdd(&Mem, pSession);
2868	if (!rc)
2869	{
2870	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2871	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2872	return VINF_SUCCESS;
2873	}
2874
2875	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2876	AssertRC(rc2);
2877	}
2878
2879	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2880	AssertRC(rc2);
2881	}
2882
2883	return rc;
2884	}
2885
2886
2887	/**
2888	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
2889	*
2890	* @returns IPRT status code.
2891	* @param pSession The session to which the memory was allocated.
2892	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2893	* @param paPages Where to store the physical addresses.
2894	*/
2895	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
2896	{
2897	PSUPDRVBUNDLE pBundle;
2898	LogFlow(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
2899
2900	/*
2901	* Validate input.
2902	*/
2903	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2904	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
2905	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
2906
2907	/*
2908	* Search for the address.
2909	*/
2910	RTSpinlockAcquire(pSession->Spinlock);
2911	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2912	{
2913	if (pBundle->cUsed > 0)
2914	{
2915	unsigned i;
2916	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2917	{
2918	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
2919	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2920	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2921	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2922	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
2923	)
2924	)
2925	{
2926	const size_t cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2927	size_t iPage;
2928	for (iPage = 0; iPage < cPages; iPage++)
2929	{
2930	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2931	paPages[iPage].uReserved = 0;
2932	}
2933	RTSpinlockRelease(pSession->Spinlock);
2934	return VINF_SUCCESS;
2935	}
2936	}
2937	}
2938	}
2939	RTSpinlockRelease(pSession->Spinlock);
2940	Log(("Failed to find %p!!!\n", (void *)uPtr));
2941	return VERR_INVALID_PARAMETER;
2942	}
2943
2944
2945	/**
2946	* Free memory allocated by SUPR0MemAlloc().
2947	*
2948	* @returns IPRT status code.
2949	* @param pSession The session owning the allocation.
2950	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2951	*/
2952	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2953	{
2954	LogFlow(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2955	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2956	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
2957	}
2958
2959
2960	/**
2961	* Allocates a chunk of memory with a kernel or/and a user mode mapping.
2962	*
2963	* The memory is fixed and it's possible to query the physical addresses using
2964	* SUPR0MemGetPhys().
2965	*
2966	* @returns IPRT status code.
2967	* @param pSession The session to associated the allocation with.
2968	* @param cPages The number of pages to allocate.
2969	* @param fFlags Flags, reserved for the future. Must be zero.
2970	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2971	* NULL if no ring-3 mapping.
2972	* @param ppvR3 Where to store the address of the Ring-0 mapping.
2973	* NULL if no ring-0 mapping.
2974	* @param paPages Where to store the addresses of the pages. Optional.
2975	*/
2976	SUPR0DECL(int) SUPR0PageAllocEx(PSUPDRVSESSION pSession, uint32_t cPages, uint32_t fFlags, PRTR3PTR ppvR3, PRTR0PTR ppvR0, PRTHCPHYS paPages)
2977	{
2978	int rc;
2979	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2980	LogFlow(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
2981
2982	/*
2983	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2984	*/
2985	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2986	AssertPtrNullReturn(ppvR3, VERR_INVALID_POINTER);
2987	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2988	AssertReturn(ppvR3 \|\| ppvR0, VERR_INVALID_PARAMETER);
2989	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2990	if (cPages < 1 \|\| cPages > VBOX_MAX_ALLOC_PAGE_COUNT)
2991	{
2992	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)));
2993	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2994	}
2995
2996	/*
2997	* Let IPRT do the work.
2998	*/
2999	if (ppvR0)
3000	rc = RTR0MemObjAllocPage(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, true /* fExecutable */);
3001	else
3002	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
3003	if (RT_SUCCESS(rc))
3004	{
3005	int rc2;
3006	if (ppvR3)
3007	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
3008	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
3009	else
3010	Mem.MapObjR3 = NIL_RTR0MEMOBJ;
3011	if (RT_SUCCESS(rc))
3012	{
3013	Mem.eType = MEMREF_TYPE_PAGE;
3014	rc = supdrvMemAdd(&Mem, pSession);
3015	if (!rc)
3016	{
3017	if (ppvR3)
3018	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
3019	if (ppvR0)
3020	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
3021	if (paPages)
3022	{
3023	uint32_t iPage = cPages;
3024	while (iPage-- > 0)
3025	{
3026	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
3027	Assert(paPages[iPage] != NIL_RTHCPHYS);
3028	}
3029	}
3030	return VINF_SUCCESS;
3031	}
3032
3033	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
3034	AssertRC(rc2);
3035	}
3036
3037	rc2 = RTR0MemObjFree(Mem.MemObj, false);
3038	AssertRC(rc2);
3039	}
3040	return rc;
3041	}
3042
3043
3044	/**
3045	* Maps a chunk of memory previously allocated by SUPR0PageAllocEx into kernel
3046	* space.
3047	*
3048	* @returns IPRT status code.
3049	* @param pSession The session to associated the allocation with.
3050	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
3051	* @param offSub Where to start mapping. Must be page aligned.
3052	* @param cbSub How much to map. Must be page aligned.
3053	* @param fFlags Flags, MBZ.
3054	* @param ppvR0 Where to return the address of the ring-0 mapping on
3055	* success.
3056	*/
3057	SUPR0DECL(int) SUPR0PageMapKernel(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t offSub, uint32_t cbSub,
3058	uint32_t fFlags, PRTR0PTR ppvR0)
3059	{
3060	int rc;
3061	PSUPDRVBUNDLE pBundle;
3062	RTR0MEMOBJ hMemObj = NIL_RTR0MEMOBJ;
3063	LogFlow(("SUPR0PageMapKernel: pSession=%p pvR3=%p offSub=%#x cbSub=%#x\n", pSession, pvR3, offSub, cbSub));
3064
3065	/*
3066	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
3067	*/
3068	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3069	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
3070	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
3071	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3072	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3073	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
3074
3075	/*
3076	* Find the memory object.
3077	*/
3078	RTSpinlockAcquire(pSession->Spinlock);
3079	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3080	{
3081	if (pBundle->cUsed > 0)
3082	{
3083	unsigned i;
3084	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3085	{
3086	if ( ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
3087	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3088	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3089	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
3090	\|\| ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED
3091	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3092	&& pBundle->aMem[i].MapObjR3 == NIL_RTR0MEMOBJ
3093	&& RTR0MemObjAddressR3(pBundle->aMem[i].MemObj) == pvR3))
3094	{
3095	hMemObj = pBundle->aMem[i].MemObj;
3096	break;
3097	}
3098	}
3099	}
3100	}
3101	RTSpinlockRelease(pSession->Spinlock);
3102
3103	rc = VERR_INVALID_PARAMETER;
3104	if (hMemObj != NIL_RTR0MEMOBJ)
3105	{
3106	/*
3107	* Do some further input validations before calling IPRT.
3108	* (Cleanup is done indirectly by telling RTR0MemObjFree to include mappings.)
3109	*/
3110	size_t cbMemObj = RTR0MemObjSize(hMemObj);
3111	if ( offSub < cbMemObj
3112	&& cbSub <= cbMemObj
3113	&& offSub + cbSub <= cbMemObj)
3114	{
3115	RTR0MEMOBJ hMapObj;
3116	rc = RTR0MemObjMapKernelEx(&hMapObj, hMemObj, (void *)-1, 0,
3117	RTMEM_PROT_READ \| RTMEM_PROT_WRITE, offSub, cbSub);
3118	if (RT_SUCCESS(rc))
3119	*ppvR0 = RTR0MemObjAddress(hMapObj);
3120	}
3121	else
3122	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
3123
3124	}
3125	return rc;
3126	}
3127
3128
3129	/**
3130	* Changes the page level protection of one or more pages previously allocated
3131	* by SUPR0PageAllocEx.
3132	*
3133	* @returns IPRT status code.
3134	* @param pSession The session to associated the allocation with.
3135	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
3136	* NIL_RTR3PTR if the ring-3 mapping should be unaffected.
3137	* @param pvR0 The ring-0 address returned by SUPR0PageAllocEx.
3138	* NIL_RTR0PTR if the ring-0 mapping should be unaffected.
3139	* @param offSub Where to start changing. Must be page aligned.
3140	* @param cbSub How much to change. Must be page aligned.
3141	* @param fProt The new page level protection, see RTMEM_PROT_*.
3142	*/
3143	SUPR0DECL(int) SUPR0PageProtect(PSUPDRVSESSION pSession, RTR3PTR pvR3, RTR0PTR pvR0, uint32_t offSub, uint32_t cbSub, uint32_t fProt)
3144	{
3145	int rc;
3146	PSUPDRVBUNDLE pBundle;
3147	RTR0MEMOBJ hMemObjR0 = NIL_RTR0MEMOBJ;
3148	RTR0MEMOBJ hMemObjR3 = NIL_RTR0MEMOBJ;
3149	LogFlow(("SUPR0PageProtect: pSession=%p pvR3=%p pvR0=%p offSub=%#x cbSub=%#x fProt-%#x\n", pSession, pvR3, pvR0, offSub, cbSub, fProt));
3150
3151	/*
3152	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
3153	*/
3154	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3155	AssertReturn(!(fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)), VERR_INVALID_PARAMETER);
3156	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3157	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3158	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
3159
3160	/*
3161	* Find the memory object.
3162	*/
3163	RTSpinlockAcquire(pSession->Spinlock);
3164	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3165	{
3166	if (pBundle->cUsed > 0)
3167	{
3168	unsigned i;
3169	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3170	{
3171	if ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
3172	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3173	&& ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3174	\|\| pvR3 == NIL_RTR3PTR)
3175	&& ( pvR0 == NIL_RTR0PTR
3176	\|\| RTR0MemObjAddress(pBundle->aMem[i].MemObj) == pvR0)
3177	&& ( pvR3 == NIL_RTR3PTR
3178	\|\| RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3))
3179	{
3180	if (pvR0 != NIL_RTR0PTR)
3181	hMemObjR0 = pBundle->aMem[i].MemObj;
3182	if (pvR3 != NIL_RTR3PTR)
3183	hMemObjR3 = pBundle->aMem[i].MapObjR3;
3184	break;
3185	}
3186	}
3187	}
3188	}
3189	RTSpinlockRelease(pSession->Spinlock);
3190
3191	rc = VERR_INVALID_PARAMETER;
3192	if ( hMemObjR0 != NIL_RTR0MEMOBJ
3193	\|\| hMemObjR3 != NIL_RTR0MEMOBJ)
3194	{
3195	/*
3196	* Do some further input validations before calling IPRT.
3197	*/
3198	size_t cbMemObj = hMemObjR0 != NIL_RTR0PTR ? RTR0MemObjSize(hMemObjR0) : RTR0MemObjSize(hMemObjR3);
3199	if ( offSub < cbMemObj
3200	&& cbSub <= cbMemObj
3201	&& offSub + cbSub <= cbMemObj)
3202	{
3203	rc = VINF_SUCCESS;
3204	if (hMemObjR3 != NIL_RTR0PTR)
3205	rc = RTR0MemObjProtect(hMemObjR3, offSub, cbSub, fProt);
3206	if (hMemObjR0 != NIL_RTR0PTR && RT_SUCCESS(rc))
3207	rc = RTR0MemObjProtect(hMemObjR0, offSub, cbSub, fProt);
3208	}
3209	else
3210	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
3211
3212	}
3213	return rc;
3214
3215	}
3216
3217
3218	/**
3219	* Free memory allocated by SUPR0PageAlloc() and SUPR0PageAllocEx().
3220	*
3221	* @returns IPRT status code.
3222	* @param pSession The session owning the allocation.
3223	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc() or
3224	* SUPR0PageAllocEx().
3225	*/
3226	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
3227	{
3228	LogFlow(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
3229	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3230	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_PAGE);
3231	}
3232
3233
3234	/**
3235	* Gets the paging mode of the current CPU.
3236	*
3237	* @returns Paging mode, SUPPAGEINGMODE_INVALID on error.
3238	*/
3239	SUPR0DECL(SUPPAGINGMODE) SUPR0GetPagingMode(void)
3240	{
3241	SUPPAGINGMODE enmMode;
3242
3243	RTR0UINTREG cr0 = ASMGetCR0();
3244	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
3245	enmMode = SUPPAGINGMODE_INVALID;
3246	else
3247	{
3248	RTR0UINTREG cr4 = ASMGetCR4();
3249	uint32_t fNXEPlusLMA = 0;
3250	if (cr4 & X86_CR4_PAE)
3251	{
3252	uint32_t fExtFeatures = ASMCpuId_EDX(0x80000001);
3253	if (fExtFeatures & (X86_CPUID_EXT_FEATURE_EDX_NX \| X86_CPUID_EXT_FEATURE_EDX_LONG_MODE))
3254	{
3255	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
3256	if ((fExtFeatures & X86_CPUID_EXT_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
3257	fNXEPlusLMA \|= RT_BIT(0);
3258	if ((fExtFeatures & X86_CPUID_EXT_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
3259	fNXEPlusLMA \|= RT_BIT(1);
3260	}
3261	}
3262
3263	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
3264	{
3265	case 0:
3266	enmMode = SUPPAGINGMODE_32_BIT;
3267	break;
3268
3269	case X86_CR4_PGE:
3270	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
3271	break;
3272
3273	case X86_CR4_PAE:
3274	enmMode = SUPPAGINGMODE_PAE;
3275	break;
3276
3277	case X86_CR4_PAE \| RT_BIT(0):
3278	enmMode = SUPPAGINGMODE_PAE_NX;
3279	break;
3280
3281	case X86_CR4_PAE \| X86_CR4_PGE:
3282	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3283	break;
3284
3285	case X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3286	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3287	break;
3288
3289	case RT_BIT(1) \| X86_CR4_PAE:
3290	enmMode = SUPPAGINGMODE_AMD64;
3291	break;
3292
3293	case RT_BIT(1) \| X86_CR4_PAE \| RT_BIT(0):
3294	enmMode = SUPPAGINGMODE_AMD64_NX;
3295	break;
3296
3297	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
3298	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
3299	break;
3300
3301	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3302	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
3303	break;
3304
3305	default:
3306	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
3307	enmMode = SUPPAGINGMODE_INVALID;
3308	break;
3309	}
3310	}
3311	return enmMode;
3312	}
3313
3314
3315	/**
3316	* Enables or disabled hardware virtualization extensions using native OS APIs.
3317	*
3318	* @returns VBox status code.
3319	* @retval VINF_SUCCESS on success.
3320	* @retval VERR_NOT_SUPPORTED if not supported by the native OS.
3321	*
3322	* @param fEnable Whether to enable or disable.
3323	*/
3324	SUPR0DECL(int) SUPR0EnableVTx(bool fEnable)
3325	{
3326	#ifdef RT_OS_DARWIN
3327	return supdrvOSEnableVTx(fEnable);
3328	#else
3329	return VERR_NOT_SUPPORTED;
3330	#endif
3331	}
3332
3333
3334	/**
3335	* Suspends hardware virtualization extensions using the native OS API.
3336	*
3337	* This is called prior to entering raw-mode context.
3338	*
3339	* @returns @c true if suspended, @c false if not.
3340	*/
3341	SUPR0DECL(bool) SUPR0SuspendVTxOnCpu(void)
3342	{
3343	#ifdef RT_OS_DARWIN
3344	return supdrvOSSuspendVTxOnCpu();
3345	#else
3346	return false;
3347	#endif
3348	}
3349
3350
3351	/**
3352	* Resumes hardware virtualization extensions using the native OS API.
3353	*
3354	* This is called after to entering raw-mode context.
3355	*
3356	* @param fSuspended The return value of SUPR0SuspendVTxOnCpu.
3357	*/
3358	SUPR0DECL(void) SUPR0ResumeVTxOnCpu(bool fSuspended)
3359	{
3360	#ifdef RT_OS_DARWIN
3361	supdrvOSResumeVTxOnCpu(fSuspended);
3362	#else
3363	Assert(!fSuspended);
3364	#endif
3365	}
3366
3367
3368	/**
3369	* Queries the AMD-V and VT-x capabilities of the calling CPU.
3370	*
3371	* @returns VBox status code.
3372	* @retval VERR_VMX_NO_VMX
3373	* @retval VERR_VMX_MSR_SMX_VMXON_DISABLED
3374	* @retval VERR_VMX_MSR_VMXON_DISABLED
3375	* @retval VERR_SVM_NO_SVM
3376	* @retval VERR_SVM_DISABLED
3377	* @retval VERR_UNSUPPORTED_CPU if not identifiable as an AMD, Intel or VIA
3378	* (centaur) CPU.
3379	*
3380	* @param pSession The session handle.
3381	* @param pfCaps Where to store the capabilities.
3382	*/
3383	SUPR0DECL(int) SUPR0QueryVTCaps(PSUPDRVSESSION pSession, uint32_t *pfCaps)
3384	{
3385	/*
3386	* Input validation.
3387	*/
3388	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3389	AssertPtrReturn(pfCaps, VERR_INVALID_POINTER);
3390
3391	*pfCaps = 0;
3392
3393	/** @todo r=ramshankar: Although we're only reading CPUIDs/MSRs here which
3394	* should be identical on all CPUs on the system, it's probably
3395	* cleaner to prevent migration nonetheless? */
3396	if (ASMHasCpuId())
3397	{
3398	uint32_t fFeaturesECX, fFeaturesEDX, uDummy;
3399	uint32_t uMaxId, uVendorEBX, uVendorECX, uVendorEDX;
3400	uint64_t u64FeatMsr;
3401
3402	ASMCpuId(0, &uMaxId, &uVendorEBX, &uVendorECX, &uVendorEDX);
3403	ASMCpuId(1, &uDummy, &uDummy, &fFeaturesECX, &fFeaturesEDX);
3404
3405	if ( ASMIsValidStdRange(uMaxId)
3406	&& ( ASMIsIntelCpuEx( uVendorEBX, uVendorECX, uVendorEDX)
3407	\|\| ASMIsViaCentaurCpuEx(uVendorEBX, uVendorECX, uVendorEDX) )
3408	)
3409	{
3410	if ( (fFeaturesECX & X86_CPUID_FEATURE_ECX_VMX)
3411	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3412	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3413	)
3414	{
3415	bool fInSmxMode;
3416	bool fMsrLocked;
3417	bool fSmxVmxAllowed;
3418	bool fVmxAllowed;
3419	bool fAllowed;
3420
3421	/*
3422	* We require the lock bit and the appropriate VMXON bit to be set otherwise VMXON will generate a #GP
3423	* This is a simplified check (assumes BIOS does it job and properly locks the control bit). For the more
3424	* extensive procedure see hmR0InitIntelCpu().
3425	*/
3426	u64FeatMsr = ASMRdMsr(MSR_IA32_FEATURE_CONTROL);
3427	fInSmxMode = !!(ASMGetCR4() & X86_CR4_SMXE);
3428	fMsrLocked = !!(u64FeatMsr & MSR_IA32_FEATURE_CONTROL_LOCK);
3429	fSmxVmxAllowed = fMsrLocked && !!(u64FeatMsr & MSR_IA32_FEATURE_CONTROL_SMX_VMXON);
3430	fVmxAllowed = fMsrLocked && !!(u64FeatMsr & MSR_IA32_FEATURE_CONTROL_VMXON);
3431	fAllowed = fInSmxMode ? fSmxVmxAllowed : fVmxAllowed;
3432	if (fAllowed)
3433	{
3434	VMX_CAPABILITY vtCaps;
3435
3436	*pfCaps \|= SUPVTCAPS_VT_X;
3437
3438	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS);
3439	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC_USE_SECONDARY_EXEC_CTRL)
3440	{
3441	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS2);
3442	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC2_EPT)
3443	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3444	}
3445	return VINF_SUCCESS;
3446	}
3447	return fInSmxMode ? VERR_VMX_MSR_SMX_VMXON_DISABLED : VERR_VMX_MSR_VMXON_DISABLED;
3448	}
3449	return VERR_VMX_NO_VMX;
3450	}
3451
3452	if ( ASMIsAmdCpuEx(uVendorEBX, uVendorECX, uVendorEDX)
3453	&& ASMIsValidStdRange(uMaxId))
3454	{
3455	uint32_t fExtFeaturesEcx, uExtMaxId;
3456	ASMCpuId(0x80000000, &uExtMaxId, &uDummy, &uDummy, &uDummy);
3457	ASMCpuId(0x80000001, &uDummy, &uDummy, &fExtFeaturesEcx, &uDummy);
3458	if ( ASMIsValidExtRange(uExtMaxId)
3459	&& uExtMaxId >= 0x8000000a
3460	&& (fExtFeaturesEcx & X86_CPUID_AMD_FEATURE_ECX_SVM)
3461	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3462	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3463	)
3464	{
3465	/* Check if SVM is disabled */
3466	u64FeatMsr = ASMRdMsr(MSR_K8_VM_CR);
3467	if (!(u64FeatMsr & MSR_K8_VM_CR_SVM_DISABLE))
3468	{
3469	uint32_t fSvmFeatures;
3470	*pfCaps \|= SUPVTCAPS_AMD_V;
3471
3472	/* Query AMD-V features. */
3473	ASMCpuId(0x8000000a, &uDummy, &uDummy, &uDummy, &fSvmFeatures);
3474	if (fSvmFeatures & AMD_CPUID_SVM_FEATURE_EDX_NESTED_PAGING)
3475	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3476
3477	return VINF_SUCCESS;
3478	}
3479	return VERR_SVM_DISABLED;
3480	}
3481	return VERR_SVM_NO_SVM;
3482	}
3483	}
3484
3485	return VERR_UNSUPPORTED_CPU;
3486	}
3487
3488
3489	/**
3490	* (Re-)initializes the per-cpu structure prior to starting or resuming the GIP
3491	* updating.
3492	*
3493	* @param pGipCpu The per CPU structure for this CPU.
3494	* @param u64NanoTS The current time.
3495	*/
3496	static void supdrvGipReInitCpu(PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
3497	{
3498	pGipCpu->u64TSC = ASMReadTSC() - pGipCpu->u32UpdateIntervalTSC;
3499	pGipCpu->u64NanoTS = u64NanoTS;
3500	}
3501
3502
3503	/**
3504	* Set the current TSC and NanoTS value for the CPU.
3505	*
3506	* @param idCpu The CPU ID. Unused - we have to use the APIC ID.
3507	* @param pvUser1 Pointer to the ring-0 GIP mapping.
3508	* @param pvUser2 Pointer to the variable holding the current time.
3509	*/
3510	static DECLCALLBACK(void) supdrvGipReInitCpuCallback(RTCPUID idCpu, void pvUser1, void pvUser2)
3511	{
3512	PSUPGLOBALINFOPAGE pGip = (PSUPGLOBALINFOPAGE)pvUser1;
3513	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
3514
3515	if (RT_LIKELY(iCpu < pGip->cCpus && pGip->aCPUs[iCpu].idCpu == idCpu))
3516	supdrvGipReInitCpu(&pGip->aCPUs[iCpu], (uint64_t )pvUser2);
3517
3518	NOREF(pvUser2);
3519	NOREF(idCpu);
3520	}
3521
3522
3523	/**
3524	* Maps the GIP into userspace and/or get the physical address of the GIP.
3525	*
3526	* @returns IPRT status code.
3527	* @param pSession Session to which the GIP mapping should belong.
3528	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
3529	* @param pHCPhysGip Where to store the physical address. (optional)
3530	*
3531	* @remark There is no reference counting on the mapping, so one call to this function
3532	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
3533	* and remove the session as a GIP user.
3534	*/
3535	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
3536	{
3537	int rc;
3538	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3539	RTR3PTR pGipR3 = NIL_RTR3PTR;
3540	RTHCPHYS HCPhys = NIL_RTHCPHYS;
3541	LogFlow(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
3542
3543	/*
3544	* Validate
3545	*/
3546	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3547	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
3548	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
3549
3550	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3551	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3552	#else
3553	RTSemFastMutexRequest(pDevExt->mtxGip);
3554	#endif
3555	if (pDevExt->pGip)
3556	{
3557	/*
3558	* Map it?
3559	*/
3560	rc = VINF_SUCCESS;
3561	if (ppGipR3)
3562	{
3563	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
3564	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
3565	RTMEM_PROT_READ, RTR0ProcHandleSelf());
3566	if (RT_SUCCESS(rc))
3567	pGipR3 = RTR0MemObjAddressR3(pSession->GipMapObjR3);
3568	}
3569
3570	/*
3571	* Get physical address.
3572	*/
3573	if (pHCPhysGip && RT_SUCCESS(rc))
3574	HCPhys = pDevExt->HCPhysGip;
3575
3576	/*
3577	* Reference globally.
3578	*/
3579	if (!pSession->fGipReferenced && RT_SUCCESS(rc))
3580	{
3581	pSession->fGipReferenced = 1;
3582	pDevExt->cGipUsers++;
3583	if (pDevExt->cGipUsers == 1)
3584	{
3585	PSUPGLOBALINFOPAGE pGipR0 = pDevExt->pGip;
3586	uint64_t u64NanoTS;
3587	uint32_t u32SystemResolution;
3588	unsigned i;
3589
3590	LogFlow(("SUPR0GipMap: Resumes GIP updating\n"));
3591
3592	/*
3593	* Try bump up the system timer resolution.
3594	* The more interrupts the better...
3595	*/
3596	if ( RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
3597	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
3598	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1953125 /* 512 HZ */, &u32SystemResolution))
3599	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 2000000 /* 500 HZ */, &u32SystemResolution))
3600	)
3601	{
3602	Assert(RTTimerGetSystemGranularity() <= u32SystemResolution);
3603	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
3604	}
3605
3606	if (pGipR0->aCPUs[0].u32TransactionId != 2 /* not the first time */)
3607	{
3608	for (i = 0; i < RT_ELEMENTS(pGipR0->aCPUs); i++)
3609	ASMAtomicUoWriteU32(&pGipR0->aCPUs[i].u32TransactionId,
3610	(pGipR0->aCPUs[i].u32TransactionId + GIP_UPDATEHZ_RECALC_FREQ * 2)
3611	& ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
3612	ASMAtomicWriteU64(&pGipR0->u64NanoTSLastUpdateHz, 0);
3613	}
3614
3615	u64NanoTS = RTTimeSystemNanoTS() - pGipR0->u32UpdateIntervalNS;
3616	if ( pGipR0->u32Mode == SUPGIPMODE_SYNC_TSC
3617	\|\| RTMpGetOnlineCount() == 1)
3618	supdrvGipReInitCpu(&pGipR0->aCPUs[0], u64NanoTS);
3619	else
3620	RTMpOnAll(supdrvGipReInitCpuCallback, pGipR0, &u64NanoTS);
3621
3622	#ifndef DO_NOT_START_GIP
3623	rc = RTTimerStart(pDevExt->pGipTimer, 0); AssertRC(rc);
3624	#endif
3625	rc = VINF_SUCCESS;
3626	}
3627	}
3628	}
3629	else
3630	{
3631	rc = VERR_GENERAL_FAILURE;
3632	Log(("SUPR0GipMap: GIP is not available!\n"));
3633	}
3634	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3635	RTSemMutexRelease(pDevExt->mtxGip);
3636	#else
3637	RTSemFastMutexRelease(pDevExt->mtxGip);
3638	#endif
3639
3640	/*
3641	* Write returns.
3642	*/
3643	if (pHCPhysGip)
3644	*pHCPhysGip = HCPhys;
3645	if (ppGipR3)
3646	*ppGipR3 = pGipR3;
3647
3648	#ifdef DEBUG_DARWIN_GIP
3649	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3650	#else
3651	LogFlow(( "SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3652	#endif
3653	return rc;
3654	}
3655
3656
3657	/**
3658	* Unmaps any user mapping of the GIP and terminates all GIP access
3659	* from this session.
3660	*
3661	* @returns IPRT status code.
3662	* @param pSession Session to which the GIP mapping should belong.
3663	*/
3664	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
3665	{
3666	int rc = VINF_SUCCESS;
3667	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3668	#ifdef DEBUG_DARWIN_GIP
3669	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
3670	pSession,
3671	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
3672	pSession->GipMapObjR3));
3673	#else
3674	LogFlow(("SUPR0GipUnmap: pSession=%p\n", pSession));
3675	#endif
3676	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3677
3678	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3679	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3680	#else
3681	RTSemFastMutexRequest(pDevExt->mtxGip);
3682	#endif
3683
3684	/*
3685	* Unmap anything?
3686	*/
3687	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
3688	{
3689	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
3690	AssertRC(rc);
3691	if (RT_SUCCESS(rc))
3692	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
3693	}
3694
3695	/*
3696	* Dereference global GIP.
3697	*/
3698	if (pSession->fGipReferenced && !rc)
3699	{
3700	pSession->fGipReferenced = 0;
3701	if ( pDevExt->cGipUsers > 0
3702	&& !--pDevExt->cGipUsers)
3703	{
3704	LogFlow(("SUPR0GipUnmap: Suspends GIP updating\n"));
3705	#ifndef DO_NOT_START_GIP
3706	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = VINF_SUCCESS;
3707	#endif
3708
3709	if (pDevExt->u32SystemTimerGranularityGrant)
3710	{
3711	int rc2 = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant);
3712	AssertRC(rc2);
3713	pDevExt->u32SystemTimerGranularityGrant = 0;
3714	}
3715	}
3716	}
3717
3718	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3719	RTSemMutexRelease(pDevExt->mtxGip);
3720	#else
3721	RTSemFastMutexRelease(pDevExt->mtxGip);
3722	#endif
3723
3724	return rc;
3725	}
3726
3727
3728	/**
3729	* Gets the GIP pointer.
3730	*
3731	* @returns Pointer to the GIP or NULL.
3732	*/
3733	SUPDECL(PSUPGLOBALINFOPAGE) SUPGetGIP(void)
3734	{
3735	return g_pSUPGlobalInfoPage;
3736	}
3737
3738
3739	/**
3740	* Register a component factory with the support driver.
3741	*
3742	* This is currently restricted to kernel sessions only.
3743	*
3744	* @returns VBox status code.
3745	* @retval VINF_SUCCESS on success.
3746	* @retval VERR_NO_MEMORY if we're out of memory.
3747	* @retval VERR_ALREADY_EXISTS if the factory has already been registered.
3748	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3749	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3750	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3751	*
3752	* @param pSession The SUPDRV session (must be a ring-0 session).
3753	* @param pFactory Pointer to the component factory registration structure.
3754	*
3755	* @remarks This interface is also available via SUPR0IdcComponentRegisterFactory.
3756	*/
3757	SUPR0DECL(int) SUPR0ComponentRegisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3758	{
3759	PSUPDRVFACTORYREG pNewReg;
3760	const char *psz;
3761	int rc;
3762
3763	/*
3764	* Validate parameters.
3765	*/
3766	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3767	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3768	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3769	AssertPtrReturn(pFactory->pfnQueryFactoryInterface, VERR_INVALID_POINTER);
3770	psz = RTStrEnd(pFactory->szName, sizeof(pFactory->szName));
3771	AssertReturn(psz, VERR_INVALID_PARAMETER);
3772
3773	/*
3774	* Allocate and initialize a new registration structure.
3775	*/
3776	pNewReg = (PSUPDRVFACTORYREG)RTMemAlloc(sizeof(SUPDRVFACTORYREG));
3777	if (pNewReg)
3778	{
3779	pNewReg->pNext = NULL;
3780	pNewReg->pFactory = pFactory;
3781	pNewReg->pSession = pSession;
3782	pNewReg->cchName = psz - &pFactory->szName[0];
3783
3784	/*
3785	* Add it to the tail of the list after checking for prior registration.
3786	*/
3787	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3788	if (RT_SUCCESS(rc))
3789	{
3790	PSUPDRVFACTORYREG pPrev = NULL;
3791	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3792	while (pCur && pCur->pFactory != pFactory)
3793	{
3794	pPrev = pCur;
3795	pCur = pCur->pNext;
3796	}
3797	if (!pCur)
3798	{
3799	if (pPrev)
3800	pPrev->pNext = pNewReg;
3801	else
3802	pSession->pDevExt->pComponentFactoryHead = pNewReg;
3803	rc = VINF_SUCCESS;
3804	}
3805	else
3806	rc = VERR_ALREADY_EXISTS;
3807
3808	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3809	}
3810
3811	if (RT_FAILURE(rc))
3812	RTMemFree(pNewReg);
3813	}
3814	else
3815	rc = VERR_NO_MEMORY;
3816	return rc;
3817	}
3818
3819
3820	/**
3821	* Deregister a component factory.
3822	*
3823	* @returns VBox status code.
3824	* @retval VINF_SUCCESS on success.
3825	* @retval VERR_NOT_FOUND if the factory wasn't registered.
3826	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3827	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3828	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3829	*
3830	* @param pSession The SUPDRV session (must be a ring-0 session).
3831	* @param pFactory Pointer to the component factory registration structure
3832	* previously passed SUPR0ComponentRegisterFactory().
3833	*
3834	* @remarks This interface is also available via SUPR0IdcComponentDeregisterFactory.
3835	*/
3836	SUPR0DECL(int) SUPR0ComponentDeregisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3837	{
3838	int rc;
3839
3840	/*
3841	* Validate parameters.
3842	*/
3843	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3844	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3845	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3846
3847	/*
3848	* Take the lock and look for the registration record.
3849	*/
3850	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3851	if (RT_SUCCESS(rc))
3852	{
3853	PSUPDRVFACTORYREG pPrev = NULL;
3854	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3855	while (pCur && pCur->pFactory != pFactory)
3856	{
3857	pPrev = pCur;
3858	pCur = pCur->pNext;
3859	}
3860	if (pCur)
3861	{
3862	if (!pPrev)
3863	pSession->pDevExt->pComponentFactoryHead = pCur->pNext;
3864	else
3865	pPrev->pNext = pCur->pNext;
3866
3867	pCur->pNext = NULL;
3868	pCur->pFactory = NULL;
3869	pCur->pSession = NULL;
3870	rc = VINF_SUCCESS;
3871	}
3872	else
3873	rc = VERR_NOT_FOUND;
3874
3875	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3876
3877	RTMemFree(pCur);
3878	}
3879	return rc;
3880	}
3881
3882
3883	/**
3884	* Queries a component factory.
3885	*
3886	* @returns VBox status code.
3887	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3888	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3889	* @retval VERR_SUPDRV_COMPONENT_NOT_FOUND if the component factory wasn't found.
3890	* @retval VERR_SUPDRV_INTERFACE_NOT_SUPPORTED if the interface wasn't supported.
3891	*
3892	* @param pSession The SUPDRV session.
3893	* @param pszName The name of the component factory.
3894	* @param pszInterfaceUuid The UUID of the factory interface (stringified).
3895	* @param ppvFactoryIf Where to store the factory interface.
3896	*/
3897	SUPR0DECL(int) SUPR0ComponentQueryFactory(PSUPDRVSESSION pSession, const char pszName, const char pszInterfaceUuid, void **ppvFactoryIf)
3898	{
3899	const char *pszEnd;
3900	size_t cchName;
3901	int rc;
3902
3903	/*
3904	* Validate parameters.
3905	*/
3906	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3907
3908	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
3909	pszEnd = RTStrEnd(pszName, RT_SIZEOFMEMB(SUPDRVFACTORY, szName));
3910	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3911	cchName = pszEnd - pszName;
3912
3913	AssertPtrReturn(pszInterfaceUuid, VERR_INVALID_POINTER);
3914	pszEnd = RTStrEnd(pszInterfaceUuid, RTUUID_STR_LENGTH);
3915	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3916
3917	AssertPtrReturn(ppvFactoryIf, VERR_INVALID_POINTER);
3918	*ppvFactoryIf = NULL;
3919
3920	/*
3921	* Take the lock and try all factories by this name.
3922	*/
3923	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3924	if (RT_SUCCESS(rc))
3925	{
3926	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3927	rc = VERR_SUPDRV_COMPONENT_NOT_FOUND;
3928	while (pCur)
3929	{
3930	if ( pCur->cchName == cchName
3931	&& !memcmp(pCur->pFactory->szName, pszName, cchName))
3932	{
3933	void *pvFactory = pCur->pFactory->pfnQueryFactoryInterface(pCur->pFactory, pSession, pszInterfaceUuid);
3934	if (pvFactory)
3935	{
3936	*ppvFactoryIf = pvFactory;
3937	rc = VINF_SUCCESS;
3938	break;
3939	}
3940	rc = VERR_SUPDRV_INTERFACE_NOT_SUPPORTED;
3941	}
3942
3943	/* next */
3944	pCur = pCur->pNext;
3945	}
3946
3947	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3948	}
3949	return rc;
3950	}
3951
3952
3953	/**
3954	* Adds a memory object to the session.
3955	*
3956	* @returns IPRT status code.
3957	* @param pMem Memory tracking structure containing the
3958	* information to track.
3959	* @param pSession The session.
3960	*/
3961	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
3962	{
3963	PSUPDRVBUNDLE pBundle;
3964
3965	/*
3966	* Find free entry and record the allocation.
3967	*/
3968	RTSpinlockAcquire(pSession->Spinlock);
3969	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3970	{
3971	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
3972	{
3973	unsigned i;
3974	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3975	{
3976	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
3977	{
3978	pBundle->cUsed++;
3979	pBundle->aMem[i] = *pMem;
3980	RTSpinlockRelease(pSession->Spinlock);
3981	return VINF_SUCCESS;
3982	}
3983	}
3984	AssertFailed(); /* !!this can't be happening!!! */
3985	}
3986	}
3987	RTSpinlockRelease(pSession->Spinlock);
3988
3989	/*
3990	* Need to allocate a new bundle.
3991	* Insert into the last entry in the bundle.
3992	*/
3993	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
3994	if (!pBundle)
3995	return VERR_NO_MEMORY;
3996
3997	/* take last entry. */
3998	pBundle->cUsed++;
3999	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
4000
4001	/* insert into list. */
4002	RTSpinlockAcquire(pSession->Spinlock);
4003	pBundle->pNext = pSession->Bundle.pNext;
4004	pSession->Bundle.pNext = pBundle;
4005	RTSpinlockRelease(pSession->Spinlock);
4006
4007	return VINF_SUCCESS;
4008	}
4009
4010
4011	/**
4012	* Releases a memory object referenced by pointer and type.
4013	*
4014	* @returns IPRT status code.
4015	* @param pSession Session data.
4016	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
4017	* @param eType Memory type.
4018	*/
4019	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
4020	{
4021	PSUPDRVBUNDLE pBundle;
4022
4023	/*
4024	* Validate input.
4025	*/
4026	if (!uPtr)
4027	{
4028	Log(("Illegal address %p\n", (void *)uPtr));
4029	return VERR_INVALID_PARAMETER;
4030	}
4031
4032	/*
4033	* Search for the address.
4034	*/
4035	RTSpinlockAcquire(pSession->Spinlock);
4036	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
4037	{
4038	if (pBundle->cUsed > 0)
4039	{
4040	unsigned i;
4041	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
4042	{
4043	if ( pBundle->aMem[i].eType == eType
4044	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
4045	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
4046	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
4047	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
4048	)
4049	{
4050	/* Make a copy of it and release it outside the spinlock. */
4051	SUPDRVMEMREF Mem = pBundle->aMem[i];
4052	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
4053	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
4054	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
4055	RTSpinlockRelease(pSession->Spinlock);
4056
4057	if (Mem.MapObjR3 != NIL_RTR0MEMOBJ)
4058	{
4059	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
4060	AssertRC(rc); /** @todo figure out how to handle this. */
4061	}
4062	if (Mem.MemObj != NIL_RTR0MEMOBJ)
4063	{
4064	int rc = RTR0MemObjFree(Mem.MemObj, true /* fFreeMappings */);
4065	AssertRC(rc); /** @todo figure out how to handle this. */
4066	}
4067	return VINF_SUCCESS;
4068	}
4069	}
4070	}
4071	}
4072	RTSpinlockRelease(pSession->Spinlock);
4073	Log(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
4074	return VERR_INVALID_PARAMETER;
4075	}
4076
4077
4078	/**
4079	* Opens an image. If it's the first time it's opened the call must upload
4080	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
4081	*
4082	* This is the 1st step of the loading.
4083	*
4084	* @returns IPRT status code.
4085	* @param pDevExt Device globals.
4086	* @param pSession Session data.
4087	* @param pReq The open request.
4088	*/
4089	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
4090	{
4091	int rc;
4092	PSUPDRVLDRIMAGE pImage;
4093	void *pv;
4094	size_t cchName = strlen(pReq->u.In.szName); /* (caller checked < 32). */
4095	LogFlow(("supdrvIOCtl_LdrOpen: szName=%s cbImageWithTabs=%d\n", pReq->u.In.szName, pReq->u.In.cbImageWithTabs));
4096
4097	/*
4098	* Check if we got an instance of the image already.
4099	*/
4100	supdrvLdrLock(pDevExt);
4101	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
4102	{
4103	if ( pImage->szName[cchName] == '\0'
4104	&& !memcmp(pImage->szName, pReq->u.In.szName, cchName))
4105	{
4106	/** @todo check cbImageBits and cbImageWithTabs here, if they differs that indicates that the images are different. */
4107	pImage->cUsage++;
4108	pReq->u.Out.pvImageBase = pImage->pvImage;
4109	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
4110	pReq->u.Out.fNativeLoader = pImage->fNative;
4111	supdrvLdrAddUsage(pSession, pImage);
4112	supdrvLdrUnlock(pDevExt);
4113	return VINF_SUCCESS;
4114	}
4115	}
4116	/* (not found - add it!) */
4117
4118	/*
4119	* Allocate memory.
4120	*/
4121	Assert(cchName < sizeof(pImage->szName));
4122	pv = RTMemAlloc(sizeof(SUPDRVLDRIMAGE));
4123	if (!pv)
4124	{
4125	supdrvLdrUnlock(pDevExt);
4126	Log(("supdrvIOCtl_LdrOpen: RTMemAlloc() failed\n"));
4127	return /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_2;
4128	}
4129
4130	/*
4131	* Setup and link in the LDR stuff.
4132	*/
4133	pImage = (PSUPDRVLDRIMAGE)pv;
4134	pImage->pvImage = NULL;
4135	pImage->pvImageAlloc = NULL;
4136	pImage->cbImageWithTabs = pReq->u.In.cbImageWithTabs;
4137	pImage->cbImageBits = pReq->u.In.cbImageBits;
4138	pImage->cSymbols = 0;
4139	pImage->paSymbols = NULL;
4140	pImage->pachStrTab = NULL;
4141	pImage->cbStrTab = 0;
4142	pImage->pfnModuleInit = NULL;
4143	pImage->pfnModuleTerm = NULL;
4144	pImage->pfnServiceReqHandler = NULL;
4145	pImage->uState = SUP_IOCTL_LDR_OPEN;
4146	pImage->cUsage = 1;
4147	pImage->pDevExt = pDevExt;
4148	memcpy(pImage->szName, pReq->u.In.szName, cchName + 1);
4149
4150	/*
4151	* Try load it using the native loader, if that isn't supported, fall back
4152	* on the older method.
4153	*/
4154	pImage->fNative = true;
4155	rc = supdrvOSLdrOpen(pDevExt, pImage, pReq->u.In.szFilename);
4156	if (rc == VERR_NOT_SUPPORTED)
4157	{
4158	pImage->pvImageAlloc = RTMemExecAlloc(pImage->cbImageBits + 31);
4159	pImage->pvImage = RT_ALIGN_P(pImage->pvImageAlloc, 32);
4160	pImage->fNative = false;
4161	rc = pImage->pvImageAlloc ? VINF_SUCCESS : VERR_NO_EXEC_MEMORY;
4162	}
4163	if (RT_FAILURE(rc))
4164	{
4165	supdrvLdrUnlock(pDevExt);
4166	RTMemFree(pImage);
4167	Log(("supdrvIOCtl_LdrOpen(%s): failed - %Rrc\n", pReq->u.In.szName, rc));
4168	return rc;
4169	}
4170	Assert(VALID_PTR(pImage->pvImage) \|\| RT_FAILURE(rc));
4171
4172	/*
4173	* Link it.
4174	*/
4175	pImage->pNext = pDevExt->pLdrImages;
4176	pDevExt->pLdrImages = pImage;
4177
4178	supdrvLdrAddUsage(pSession, pImage);
4179
4180	pReq->u.Out.pvImageBase = pImage->pvImage;
4181	pReq->u.Out.fNeedsLoading = true;
4182	pReq->u.Out.fNativeLoader = pImage->fNative;
4183	supdrvOSLdrNotifyOpened(pDevExt, pImage);
4184
4185	supdrvLdrUnlock(pDevExt);
4186	return VINF_SUCCESS;
4187	}
4188
4189
4190	/**
4191	* Worker that validates a pointer to an image entrypoint.
4192	*
4193	* @returns IPRT status code.
4194	* @param pDevExt The device globals.
4195	* @param pImage The loader image.
4196	* @param pv The pointer into the image.
4197	* @param fMayBeNull Whether it may be NULL.
4198	* @param pszWhat What is this entrypoint? (for logging)
4199	* @param pbImageBits The image bits prepared by ring-3.
4200	*
4201	* @remarks Will leave the lock on failure.
4202	*/
4203	static int supdrvLdrValidatePointer(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage, void *pv,
4204	bool fMayBeNull, const uint8_t pbImageBits, const char pszWhat)
4205	{
4206	if (!fMayBeNull \|\| pv)
4207	{
4208	if ((uintptr_t)pv - (uintptr_t)pImage->pvImage >= pImage->cbImageBits)
4209	{
4210	supdrvLdrUnlock(pDevExt);
4211	Log(("Out of range (%p LB %#x): %s=%p\n", pImage->pvImage, pImage->cbImageBits, pszWhat, pv));
4212	return VERR_INVALID_PARAMETER;
4213	}
4214
4215	if (pImage->fNative)
4216	{
4217	int rc = supdrvOSLdrValidatePointer(pDevExt, pImage, pv, pbImageBits);
4218	if (RT_FAILURE(rc))
4219	{
4220	supdrvLdrUnlock(pDevExt);
4221	Log(("Bad entry point address: %s=%p (rc=%Rrc)\n", pszWhat, pv, rc));
4222	return rc;
4223	}
4224	}
4225	}
4226	return VINF_SUCCESS;
4227	}
4228
4229
4230	/**
4231	* Loads the image bits.
4232	*
4233	* This is the 2nd step of the loading.
4234	*
4235	* @returns IPRT status code.
4236	* @param pDevExt Device globals.
4237	* @param pSession Session data.
4238	* @param pReq The request.
4239	*/
4240	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
4241	{
4242	PSUPDRVLDRUSAGE pUsage;
4243	PSUPDRVLDRIMAGE pImage;
4244	int rc;
4245	LogFlow(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImageWithBits=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImageWithTabs));
4246
4247	/*
4248	* Find the ldr image.
4249	*/
4250	supdrvLdrLock(pDevExt);
4251	pUsage = pSession->pLdrUsage;
4252	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4253	pUsage = pUsage->pNext;
4254	if (!pUsage)
4255	{
4256	supdrvLdrUnlock(pDevExt);
4257	Log(("SUP_IOCTL_LDR_LOAD: couldn't find image!\n"));
4258	return VERR_INVALID_HANDLE;
4259	}
4260	pImage = pUsage->pImage;
4261
4262	/*
4263	* Validate input.
4264	*/
4265	if ( pImage->cbImageWithTabs != pReq->u.In.cbImageWithTabs
4266	\|\| pImage->cbImageBits != pReq->u.In.cbImageBits)
4267	{
4268	supdrvLdrUnlock(pDevExt);
4269	Log(("SUP_IOCTL_LDR_LOAD: image size mismatch!! %d(prep) != %d(load) or %d != %d\n",
4270	pImage->cbImageWithTabs, pReq->u.In.cbImageWithTabs, pImage->cbImageBits, pReq->u.In.cbImageBits));
4271	return VERR_INVALID_HANDLE;
4272	}
4273
4274	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
4275	{
4276	unsigned uState = pImage->uState;
4277	supdrvLdrUnlock(pDevExt);
4278	if (uState != SUP_IOCTL_LDR_LOAD)
4279	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
4280	return VERR_ALREADY_LOADED;
4281	}
4282
4283	switch (pReq->u.In.eEPType)
4284	{
4285	case SUPLDRLOADEP_NOTHING:
4286	break;
4287
4288	case SUPLDRLOADEP_VMMR0:
4289	rc = supdrvLdrValidatePointer( pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0, false, pReq->u.In.abImage, "pvVMMR0");
4290	if (RT_SUCCESS(rc))
4291	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt, false, pReq->u.In.abImage, "pvVMMR0EntryInt");
4292	if (RT_SUCCESS(rc))
4293	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, false, pReq->u.In.abImage, "pvVMMR0EntryFast");
4294	if (RT_SUCCESS(rc))
4295	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx, false, pReq->u.In.abImage, "pvVMMR0EntryEx");
4296	if (RT_FAILURE(rc))
4297	return rc;
4298	break;
4299
4300	case SUPLDRLOADEP_SERVICE:
4301	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.Service.pfnServiceReq, false, pReq->u.In.abImage, "pfnServiceReq");
4302	if (RT_FAILURE(rc))
4303	return rc;
4304	if ( pReq->u.In.EP.Service.apvReserved[0] != NIL_RTR0PTR
4305	\|\| pReq->u.In.EP.Service.apvReserved[1] != NIL_RTR0PTR
4306	\|\| pReq->u.In.EP.Service.apvReserved[2] != NIL_RTR0PTR)
4307	{
4308	supdrvLdrUnlock(pDevExt);
4309	Log(("Out of range (%p LB %#x): apvReserved={%p,%p,%p} MBZ!\n",
4310	pImage->pvImage, pReq->u.In.cbImageWithTabs,
4311	pReq->u.In.EP.Service.apvReserved[0],
4312	pReq->u.In.EP.Service.apvReserved[1],
4313	pReq->u.In.EP.Service.apvReserved[2]));
4314	return VERR_INVALID_PARAMETER;
4315	}
4316	break;
4317
4318	default:
4319	supdrvLdrUnlock(pDevExt);
4320	Log(("Invalid eEPType=%d\n", pReq->u.In.eEPType));
4321	return VERR_INVALID_PARAMETER;
4322	}
4323
4324	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleInit, true, pReq->u.In.abImage, "pfnModuleInit");
4325	if (RT_FAILURE(rc))
4326	return rc;
4327	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleTerm, true, pReq->u.In.abImage, "pfnModuleTerm");
4328	if (RT_FAILURE(rc))
4329	return rc;
4330
4331	/*
4332	* Allocate and copy the tables.
4333	* (No need to do try/except as this is a buffered request.)
4334	*/
4335	pImage->cbStrTab = pReq->u.In.cbStrTab;
4336	if (pImage->cbStrTab)
4337	{
4338	pImage->pachStrTab = (char *)RTMemAlloc(pImage->cbStrTab);
4339	if (pImage->pachStrTab)
4340	memcpy(pImage->pachStrTab, &pReq->u.In.abImage[pReq->u.In.offStrTab], pImage->cbStrTab);
4341	else
4342	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_3;
4343	}
4344
4345	pImage->cSymbols = pReq->u.In.cSymbols;
4346	if (RT_SUCCESS(rc) && pImage->cSymbols)
4347	{
4348	size_t cbSymbols = pImage->cSymbols * sizeof(SUPLDRSYM);
4349	pImage->paSymbols = (PSUPLDRSYM)RTMemAlloc(cbSymbols);
4350	if (pImage->paSymbols)
4351	memcpy(pImage->paSymbols, &pReq->u.In.abImage[pReq->u.In.offSymbols], cbSymbols);
4352	else
4353	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_4;
4354	}
4355
4356	/*
4357	* Copy the bits / complete native loading.
4358	*/
4359	if (RT_SUCCESS(rc))
4360	{
4361	pImage->uState = SUP_IOCTL_LDR_LOAD;
4362	pImage->pfnModuleInit = pReq->u.In.pfnModuleInit;
4363	pImage->pfnModuleTerm = pReq->u.In.pfnModuleTerm;
4364
4365	if (pImage->fNative)
4366	rc = supdrvOSLdrLoad(pDevExt, pImage, pReq->u.In.abImage, pReq);
4367	else
4368	{
4369	memcpy(pImage->pvImage, &pReq->u.In.abImage[0], pImage->cbImageBits);
4370	Log(("vboxdrv: Loaded '%s' at %p\n", pImage->szName, pImage->pvImage));
4371	}
4372	}
4373
4374	/*
4375	* Update any entry points.
4376	*/
4377	if (RT_SUCCESS(rc))
4378	{
4379	switch (pReq->u.In.eEPType)
4380	{
4381	default:
4382	case SUPLDRLOADEP_NOTHING:
4383	rc = VINF_SUCCESS;
4384	break;
4385	case SUPLDRLOADEP_VMMR0:
4386	rc = supdrvLdrSetVMMR0EPs(pDevExt, pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
4387	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
4388	break;
4389	case SUPLDRLOADEP_SERVICE:
4390	pImage->pfnServiceReqHandler = pReq->u.In.EP.Service.pfnServiceReq;
4391	rc = VINF_SUCCESS;
4392	break;
4393	}
4394	}
4395
4396	/*
4397	* On success call the module initialization.
4398	*/
4399	LogFlow(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
4400	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
4401	{
4402	Log(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
4403	pDevExt->pLdrInitImage = pImage;
4404	pDevExt->hLdrInitThread = RTThreadNativeSelf();
4405	rc = pImage->pfnModuleInit(pImage);
4406	pDevExt->pLdrInitImage = NULL;
4407	pDevExt->hLdrInitThread = NIL_RTNATIVETHREAD;
4408	if (RT_FAILURE(rc) && pDevExt->pvVMMR0 == pImage->pvImage)
4409	supdrvLdrUnsetVMMR0EPs(pDevExt);
4410	}
4411
4412	if (RT_FAILURE(rc))
4413	{
4414	/* Inform the tracing component in case ModuleInit registered TPs. */
4415	supdrvTracerModuleUnloading(pDevExt, pImage);
4416
4417	pImage->uState = SUP_IOCTL_LDR_OPEN;
4418	pImage->pfnModuleInit = NULL;
4419	pImage->pfnModuleTerm = NULL;
4420	pImage->pfnServiceReqHandler= NULL;
4421	pImage->cbStrTab = 0;
4422	RTMemFree(pImage->pachStrTab);
4423	pImage->pachStrTab = NULL;
4424	RTMemFree(pImage->paSymbols);
4425	pImage->paSymbols = NULL;
4426	pImage->cSymbols = 0;
4427	}
4428
4429	supdrvLdrUnlock(pDevExt);
4430	return rc;
4431	}
4432
4433
4434	/**
4435	* Frees a previously loaded (prep'ed) image.
4436	*
4437	* @returns IPRT status code.
4438	* @param pDevExt Device globals.
4439	* @param pSession Session data.
4440	* @param pReq The request.
4441	*/
4442	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
4443	{
4444	int rc;
4445	PSUPDRVLDRUSAGE pUsagePrev;
4446	PSUPDRVLDRUSAGE pUsage;
4447	PSUPDRVLDRIMAGE pImage;
4448	LogFlow(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
4449
4450	/*
4451	* Find the ldr image.
4452	*/
4453	supdrvLdrLock(pDevExt);
4454	pUsagePrev = NULL;
4455	pUsage = pSession->pLdrUsage;
4456	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4457	{
4458	pUsagePrev = pUsage;
4459	pUsage = pUsage->pNext;
4460	}
4461	if (!pUsage)
4462	{
4463	supdrvLdrUnlock(pDevExt);
4464	Log(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
4465	return VERR_INVALID_HANDLE;
4466	}
4467
4468	/*
4469	* Check if we can remove anything.
4470	*/
4471	rc = VINF_SUCCESS;
4472	pImage = pUsage->pImage;
4473	if (pImage->cUsage <= 1 \|\| pUsage->cUsage <= 1)
4474	{
4475	/*
4476	* Check if there are any objects with destructors in the image, if
4477	* so leave it for the session cleanup routine so we get a chance to
4478	* clean things up in the right order and not leave them all dangling.
4479	*/
4480	RTSpinlockAcquire(pDevExt->Spinlock);
4481	if (pImage->cUsage <= 1)
4482	{
4483	PSUPDRVOBJ pObj;
4484	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4485	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4486	{
4487	rc = VERR_DANGLING_OBJECTS;
4488	break;
4489	}
4490	}
4491	else
4492	{
4493	PSUPDRVUSAGE pGenUsage;
4494	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
4495	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4496	{
4497	rc = VERR_DANGLING_OBJECTS;
4498	break;
4499	}
4500	}
4501	RTSpinlockRelease(pDevExt->Spinlock);
4502	if (rc == VINF_SUCCESS)
4503	{
4504	/* unlink it */
4505	if (pUsagePrev)
4506	pUsagePrev->pNext = pUsage->pNext;
4507	else
4508	pSession->pLdrUsage = pUsage->pNext;
4509
4510	/* free it */
4511	pUsage->pImage = NULL;
4512	pUsage->pNext = NULL;
4513	RTMemFree(pUsage);
4514
4515	/*
4516	* Dereference the image.
4517	*/
4518	if (pImage->cUsage <= 1)
4519	supdrvLdrFree(pDevExt, pImage);
4520	else
4521	pImage->cUsage--;
4522	}
4523	else
4524	{
4525	Log(("supdrvIOCtl_LdrFree: Dangling objects in %p/%s!\n", pImage->pvImage, pImage->szName));
4526	rc = VINF_SUCCESS; /** @todo BRANCH-2.1: remove this after branching. */
4527	}
4528	}
4529	else
4530	{
4531	/*
4532	* Dereference both image and usage.
4533	*/
4534	pImage->cUsage--;
4535	pUsage->cUsage--;
4536	}
4537
4538	supdrvLdrUnlock(pDevExt);
4539	return rc;
4540	}
4541
4542
4543	/**
4544	* Gets the address of a symbol in an open image.
4545	*
4546	* @returns IPRT status code.
4547	* @param pDevExt Device globals.
4548	* @param pSession Session data.
4549	* @param pReq The request buffer.
4550	*/
4551	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
4552	{
4553	PSUPDRVLDRIMAGE pImage;
4554	PSUPDRVLDRUSAGE pUsage;
4555	uint32_t i;
4556	PSUPLDRSYM paSyms;
4557	const char *pchStrings;
4558	const size_t cbSymbol = strlen(pReq->u.In.szSymbol) + 1;
4559	void *pvSymbol = NULL;
4560	int rc = VERR_GENERAL_FAILURE;
4561	Log3(("supdrvIOCtl_LdrGetSymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
4562
4563	/*
4564	* Find the ldr image.
4565	*/
4566	supdrvLdrLock(pDevExt);
4567	pUsage = pSession->pLdrUsage;
4568	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4569	pUsage = pUsage->pNext;
4570	if (!pUsage)
4571	{
4572	supdrvLdrUnlock(pDevExt);
4573	Log(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
4574	return VERR_INVALID_HANDLE;
4575	}
4576	pImage = pUsage->pImage;
4577	if (pImage->uState != SUP_IOCTL_LDR_LOAD)
4578	{
4579	unsigned uState = pImage->uState;
4580	supdrvLdrUnlock(pDevExt);
4581	Log(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", uState, uState)); NOREF(uState);
4582	return VERR_ALREADY_LOADED;
4583	}
4584
4585	/*
4586	* Search the symbol strings.
4587	*
4588	* Note! The int32_t is for native loading on solaris where the data
4589	* and text segments are in very different places.
4590	*/
4591	pchStrings = pImage->pachStrTab;
4592	paSyms = pImage->paSymbols;
4593	for (i = 0; i < pImage->cSymbols; i++)
4594	{
4595	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4596	&& !memcmp(pchStrings + paSyms[i].offName, pReq->u.In.szSymbol, cbSymbol))
4597	{
4598	pvSymbol = (uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol;
4599	rc = VINF_SUCCESS;
4600	break;
4601	}
4602	}
4603	supdrvLdrUnlock(pDevExt);
4604	pReq->u.Out.pvSymbol = pvSymbol;
4605	return rc;
4606	}
4607
4608
4609	/**
4610	* Gets the address of a symbol in an open image or the support driver.
4611	*
4612	* @returns VINF_SUCCESS on success.
4613	* @returns
4614	* @param pDevExt Device globals.
4615	* @param pSession Session data.
4616	* @param pReq The request buffer.
4617	*/
4618	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq)
4619	{
4620	int rc = VINF_SUCCESS;
4621	const char *pszSymbol = pReq->u.In.pszSymbol;
4622	const char *pszModule = pReq->u.In.pszModule;
4623	size_t cbSymbol;
4624	char const *pszEnd;
4625	uint32_t i;
4626
4627	/*
4628	* Input validation.
4629	*/
4630	AssertPtrReturn(pszSymbol, VERR_INVALID_POINTER);
4631	pszEnd = RTStrEnd(pszSymbol, 512);
4632	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4633	cbSymbol = pszEnd - pszSymbol + 1;
4634
4635	if (pszModule)
4636	{
4637	AssertPtrReturn(pszModule, VERR_INVALID_POINTER);
4638	pszEnd = RTStrEnd(pszModule, 64);
4639	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4640	}
4641	Log3(("supdrvIDC_LdrGetSymbol: pszModule=%p:{%s} pszSymbol=%p:{%s}\n", pszModule, pszModule, pszSymbol, pszSymbol));
4642
4643
4644	if ( !pszModule
4645	\|\| !strcmp(pszModule, "SupDrv"))
4646	{
4647	/*
4648	* Search the support driver export table.
4649	*/
4650	for (i = 0; i < RT_ELEMENTS(g_aFunctions); i++)
4651	if (!strcmp(g_aFunctions[i].szName, pszSymbol))
4652	{
4653	pReq->u.Out.pfnSymbol = g_aFunctions[i].pfn;
4654	break;
4655	}
4656	}
4657	else
4658	{
4659	/*
4660	* Find the loader image.
4661	*/
4662	PSUPDRVLDRIMAGE pImage;
4663
4664	supdrvLdrLock(pDevExt);
4665
4666	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
4667	if (!strcmp(pImage->szName, pszModule))
4668	break;
4669	if (pImage && pImage->uState == SUP_IOCTL_LDR_LOAD)
4670	{
4671	/*
4672	* Search the symbol strings.
4673	*/
4674	const char *pchStrings = pImage->pachStrTab;
4675	PCSUPLDRSYM paSyms = pImage->paSymbols;
4676	for (i = 0; i < pImage->cSymbols; i++)
4677	{
4678	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4679	&& !memcmp(pchStrings + paSyms[i].offName, pszSymbol, cbSymbol))
4680	{
4681	/*
4682	* Found it! Calc the symbol address and add a reference to the module.
4683	*/
4684	pReq->u.Out.pfnSymbol = (PFNRT)((uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol);
4685	rc = supdrvLdrAddUsage(pSession, pImage);
4686	break;
4687	}
4688	}
4689	}
4690	else
4691	rc = pImage ? VERR_WRONG_ORDER : VERR_MODULE_NOT_FOUND;
4692
4693	supdrvLdrUnlock(pDevExt);
4694	}
4695	return rc;
4696	}
4697
4698
4699	/**
4700	* Updates the VMMR0 entry point pointers.
4701	*
4702	* @returns IPRT status code.
4703	* @param pDevExt Device globals.
4704	* @param pSession Session data.
4705	* @param pVMMR0 VMMR0 image handle.
4706	* @param pvVMMR0EntryInt VMMR0EntryInt address.
4707	* @param pvVMMR0EntryFast VMMR0EntryFast address.
4708	* @param pvVMMR0EntryEx VMMR0EntryEx address.
4709	* @remark Caller must own the loader mutex.
4710	*/
4711	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx)
4712	{
4713	int rc = VINF_SUCCESS;
4714	LogFlow(("supdrvLdrSetR0EP pvVMMR0=%p pvVMMR0EntryInt=%p\n", pvVMMR0, pvVMMR0EntryInt));
4715
4716
4717	/*
4718	* Check if not yet set.
4719	*/
4720	if (!pDevExt->pvVMMR0)
4721	{
4722	pDevExt->pvVMMR0 = pvVMMR0;
4723	pDevExt->pfnVMMR0EntryInt = pvVMMR0EntryInt;
4724	pDevExt->pfnVMMR0EntryFast = pvVMMR0EntryFast;
4725	pDevExt->pfnVMMR0EntryEx = pvVMMR0EntryEx;
4726	}
4727	else
4728	{
4729	/*
4730	* Return failure or success depending on whether the values match or not.
4731	*/
4732	if ( pDevExt->pvVMMR0 != pvVMMR0
4733	\|\| (void *)pDevExt->pfnVMMR0EntryInt != pvVMMR0EntryInt
4734	\|\| (void *)pDevExt->pfnVMMR0EntryFast != pvVMMR0EntryFast
4735	\|\| (void *)pDevExt->pfnVMMR0EntryEx != pvVMMR0EntryEx)
4736	{
4737	AssertMsgFailed(("SUP_IOCTL_LDR_SETR0EP: Already set pointing to a different module!\n"));
4738	rc = VERR_INVALID_PARAMETER;
4739	}
4740	}
4741	return rc;
4742	}
4743
4744
4745	/**
4746	* Unsets the VMMR0 entry point installed by supdrvLdrSetR0EP.
4747	*
4748	* @param pDevExt Device globals.
4749	*/
4750	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt)
4751	{
4752	pDevExt->pvVMMR0 = NULL;
4753	pDevExt->pfnVMMR0EntryInt = NULL;
4754	pDevExt->pfnVMMR0EntryFast = NULL;
4755	pDevExt->pfnVMMR0EntryEx = NULL;
4756	}
4757
4758
4759	/**
4760	* Adds a usage reference in the specified session of an image.
4761	*
4762	* Called while owning the loader semaphore.
4763	*
4764	* @returns VINF_SUCCESS on success and VERR_NO_MEMORY on failure.
4765	* @param pSession Session in question.
4766	* @param pImage Image which the session is using.
4767	*/
4768	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage)
4769	{
4770	PSUPDRVLDRUSAGE pUsage;
4771	LogFlow(("supdrvLdrAddUsage: pImage=%p\n", pImage));
4772
4773	/*
4774	* Referenced it already?
4775	*/
4776	pUsage = pSession->pLdrUsage;
4777	while (pUsage)
4778	{
4779	if (pUsage->pImage == pImage)
4780	{
4781	pUsage->cUsage++;
4782	return VINF_SUCCESS;
4783	}
4784	pUsage = pUsage->pNext;
4785	}
4786
4787	/*
4788	* Allocate new usage record.
4789	*/
4790	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
4791	AssertReturn(pUsage, /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_5);
4792	pUsage->cUsage = 1;
4793	pUsage->pImage = pImage;
4794	pUsage->pNext = pSession->pLdrUsage;
4795	pSession->pLdrUsage = pUsage;
4796	return VINF_SUCCESS;
4797	}
4798
4799
4800	/**
4801	* Frees a load image.
4802	*
4803	* @param pDevExt Pointer to device extension.
4804	* @param pImage Pointer to the image we're gonna free.
4805	* This image must exit!
4806	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
4807	*/
4808	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
4809	{
4810	PSUPDRVLDRIMAGE pImagePrev;
4811	LogFlow(("supdrvLdrFree: pImage=%p\n", pImage));
4812
4813	/* find it - arg. should've used doubly linked list. */
4814	Assert(pDevExt->pLdrImages);
4815	pImagePrev = NULL;
4816	if (pDevExt->pLdrImages != pImage)
4817	{
4818	pImagePrev = pDevExt->pLdrImages;
4819	while (pImagePrev->pNext != pImage)
4820	pImagePrev = pImagePrev->pNext;
4821	Assert(pImagePrev->pNext == pImage);
4822	}
4823
4824	/* unlink */
4825	if (pImagePrev)
4826	pImagePrev->pNext = pImage->pNext;
4827	else
4828	pDevExt->pLdrImages = pImage->pNext;
4829
4830	/* check if this is VMMR0.r0 unset its entry point pointers. */
4831	if (pDevExt->pvVMMR0 == pImage->pvImage)
4832	supdrvLdrUnsetVMMR0EPs(pDevExt);
4833
4834	/* check for objects with destructors in this image. (Shouldn't happen.) */
4835	if (pDevExt->pObjs)
4836	{
4837	unsigned cObjs = 0;
4838	PSUPDRVOBJ pObj;
4839	RTSpinlockAcquire(pDevExt->Spinlock);
4840	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4841	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4842	{
4843	pObj->pfnDestructor = NULL;
4844	cObjs++;
4845	}
4846	RTSpinlockRelease(pDevExt->Spinlock);
4847	if (cObjs)
4848	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
4849	}
4850
4851	/* call termination function if fully loaded. */
4852	if ( pImage->pfnModuleTerm
4853	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
4854	{
4855	LogFlow(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
4856	pImage->pfnModuleTerm(pImage);
4857	}
4858
4859	/* Inform the tracing component. */
4860	supdrvTracerModuleUnloading(pDevExt, pImage);
4861
4862	/* do native unload if appropriate. */
4863	if (pImage->fNative)
4864	supdrvOSLdrUnload(pDevExt, pImage);
4865
4866	/* free the image */
4867	pImage->cUsage = 0;
4868	pImage->pDevExt = NULL;
4869	pImage->pNext = NULL;
4870	pImage->uState = SUP_IOCTL_LDR_FREE;
4871	RTMemExecFree(pImage->pvImageAlloc, pImage->cbImageBits + 31);
4872	pImage->pvImageAlloc = NULL;
4873	RTMemFree(pImage->pachStrTab);
4874	pImage->pachStrTab = NULL;
4875	RTMemFree(pImage->paSymbols);
4876	pImage->paSymbols = NULL;
4877	RTMemFree(pImage);
4878	}
4879
4880
4881	/**
4882	* Acquires the loader lock.
4883	*
4884	* @returns IPRT status code.
4885	* @param pDevExt The device extension.
4886	*/
4887	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt)
4888	{
4889	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4890	int rc = RTSemMutexRequest(pDevExt->mtxLdr, RT_INDEFINITE_WAIT);
4891	#else
4892	int rc = RTSemFastMutexRequest(pDevExt->mtxLdr);
4893	#endif
4894	AssertRC(rc);
4895	return rc;
4896	}
4897
4898
4899	/**
4900	* Releases the loader lock.
4901	*
4902	* @returns IPRT status code.
4903	* @param pDevExt The device extension.
4904	*/
4905	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt)
4906	{
4907	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4908	return RTSemMutexRelease(pDevExt->mtxLdr);
4909	#else
4910	return RTSemFastMutexRelease(pDevExt->mtxLdr);
4911	#endif
4912	}
4913
4914
4915	/**
4916	* Implements the service call request.
4917	*
4918	* @returns VBox status code.
4919	* @param pDevExt The device extension.
4920	* @param pSession The calling session.
4921	* @param pReq The request packet, valid.
4922	*/
4923	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq)
4924	{
4925	#if !defined(RT_OS_WINDOWS) \|\| defined(RT_ARCH_AMD64) \|\| defined(DEBUG)
4926	int rc;
4927
4928	/*
4929	* Find the module first in the module referenced by the calling session.
4930	*/
4931	rc = supdrvLdrLock(pDevExt);
4932	if (RT_SUCCESS(rc))
4933	{
4934	PFNSUPR0SERVICEREQHANDLER pfnServiceReqHandler = NULL;
4935	PSUPDRVLDRUSAGE pUsage;
4936
4937	for (pUsage = pSession->pLdrUsage; pUsage; pUsage = pUsage->pNext)
4938	if ( pUsage->pImage->pfnServiceReqHandler
4939	&& !strcmp(pUsage->pImage->szName, pReq->u.In.szName))
4940	{
4941	pfnServiceReqHandler = pUsage->pImage->pfnServiceReqHandler;
4942	break;
4943	}
4944	supdrvLdrUnlock(pDevExt);
4945
4946	if (pfnServiceReqHandler)
4947	{
4948	/*
4949	* Call it.
4950	*/
4951	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
4952	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, NULL);
4953	else
4954	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0]);
4955	}
4956	else
4957	rc = VERR_SUPDRV_SERVICE_NOT_FOUND;
4958	}
4959
4960	/* log it */
4961	if ( RT_FAILURE(rc)
4962	&& rc != VERR_INTERRUPTED
4963	&& rc != VERR_TIMEOUT)
4964	Log(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4965	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4966	else
4967	Log4(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4968	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4969	return rc;
4970	#else /* RT_OS_WINDOWS && !RT_ARCH_AMD64 && !DEBUG */
4971	return VERR_NOT_IMPLEMENTED;
4972	#endif /* RT_OS_WINDOWS && !RT_ARCH_AMD64 && !DEBUG */
4973	}
4974
4975
4976	/**
4977	* Implements the logger settings request.
4978	*
4979	* @returns VBox status code.
4980	* @param pDevExt The device extension.
4981	* @param pSession The caller's session.
4982	* @param pReq The request.
4983	*/
4984	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq)
4985	{
4986	const char *pszGroup = &pReq->u.In.szStrings[pReq->u.In.offGroups];
4987	const char *pszFlags = &pReq->u.In.szStrings[pReq->u.In.offFlags];
4988	const char *pszDest = &pReq->u.In.szStrings[pReq->u.In.offDestination];
4989	PRTLOGGER pLogger = NULL;
4990	int rc;
4991
4992	/*
4993	* Some further validation.
4994	*/
4995	switch (pReq->u.In.fWhat)
4996	{
4997	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4998	case SUPLOGGERSETTINGS_WHAT_CREATE:
4999	break;
5000
5001	case SUPLOGGERSETTINGS_WHAT_DESTROY:
5002	if (pszGroup \|\| pszFlags \|\| *pszDest)
5003	return VERR_INVALID_PARAMETER;
5004	if (pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_RELEASE)
5005	return VERR_ACCESS_DENIED;
5006	break;
5007
5008	default:
5009	return VERR_INTERNAL_ERROR;
5010	}
5011
5012	/*
5013	* Get the logger.
5014	*/
5015	switch (pReq->u.In.fWhich)
5016	{
5017	case SUPLOGGERSETTINGS_WHICH_DEBUG:
5018	pLogger = RTLogGetDefaultInstance();
5019	break;
5020
5021	case SUPLOGGERSETTINGS_WHICH_RELEASE:
5022	pLogger = RTLogRelDefaultInstance();
5023	break;
5024
5025	default:
5026	return VERR_INTERNAL_ERROR;
5027	}
5028
5029	/*
5030	* Do the job.
5031	*/
5032	switch (pReq->u.In.fWhat)
5033	{
5034	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
5035	if (pLogger)
5036	{
5037	rc = RTLogFlags(pLogger, pszFlags);
5038	if (RT_SUCCESS(rc))
5039	rc = RTLogGroupSettings(pLogger, pszGroup);
5040	NOREF(pszDest);
5041	}
5042	else
5043	rc = VERR_NOT_FOUND;
5044	break;
5045
5046	case SUPLOGGERSETTINGS_WHAT_CREATE:
5047	{
5048	if (pLogger)
5049	rc = VERR_ALREADY_EXISTS;
5050	else
5051	{
5052	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
5053
5054	rc = RTLogCreate(&pLogger,
5055	0 /* fFlags */,
5056	pszGroup,
5057	pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_DEBUG
5058	? "VBOX_LOG"
5059	: "VBOX_RELEASE_LOG",
5060	RT_ELEMENTS(s_apszGroups),
5061	s_apszGroups,
5062	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER,
5063	NULL);
5064	if (RT_SUCCESS(rc))
5065	{
5066	rc = RTLogFlags(pLogger, pszFlags);
5067	NOREF(pszDest);
5068	if (RT_SUCCESS(rc))
5069	{
5070	switch (pReq->u.In.fWhich)
5071	{
5072	case SUPLOGGERSETTINGS_WHICH_DEBUG:
5073	pLogger = RTLogSetDefaultInstance(pLogger);
5074	break;
5075	case SUPLOGGERSETTINGS_WHICH_RELEASE:
5076	pLogger = RTLogRelSetDefaultInstance(pLogger);
5077	break;
5078	}
5079	}
5080	RTLogDestroy(pLogger);
5081	}
5082	}
5083	break;
5084	}
5085
5086	case SUPLOGGERSETTINGS_WHAT_DESTROY:
5087	switch (pReq->u.In.fWhich)
5088	{
5089	case SUPLOGGERSETTINGS_WHICH_DEBUG:
5090	pLogger = RTLogSetDefaultInstance(NULL);
5091	break;
5092	case SUPLOGGERSETTINGS_WHICH_RELEASE:
5093	pLogger = RTLogRelSetDefaultInstance(NULL);
5094	break;
5095	}
5096	rc = RTLogDestroy(pLogger);
5097	break;
5098
5099	default:
5100	{
5101	rc = VERR_INTERNAL_ERROR;
5102	break;
5103	}
5104	}
5105
5106	return rc;
5107	}
5108
5109
5110	/**
5111	* Creates the GIP.
5112	*
5113	* @returns VBox status code.
5114	* @param pDevExt Instance data. GIP stuff may be updated.
5115	*/
5116	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
5117	{
5118	PSUPGLOBALINFOPAGE pGip;
5119	RTHCPHYS HCPhysGip;
5120	uint32_t u32SystemResolution;
5121	uint32_t u32Interval;
5122	unsigned cCpus;
5123	int rc;
5124
5125
5126	LogFlow(("supdrvGipCreate:\n"));
5127
5128	/* assert order */
5129	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
5130	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
5131	Assert(!pDevExt->pGipTimer);
5132
5133	/*
5134	* Check the CPU count.
5135	*/
5136	cCpus = RTMpGetArraySize();
5137	if ( cCpus > RTCPUSET_MAX_CPUS
5138	\|\| cCpus > 256 /ApicId is used for the mappings/)
5139	{
5140	SUPR0Printf("VBoxDrv: Too many CPUs (%u) for the GIP (max %u)\n", cCpus, RT_MIN(RTCPUSET_MAX_CPUS, 256));
5141	return VERR_TOO_MANY_CPUS;
5142	}
5143
5144	/*
5145	* Allocate a contiguous set of pages with a default kernel mapping.
5146	*/
5147	rc = RTR0MemObjAllocCont(&pDevExt->GipMemObj, RT_UOFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), false /fExecutable/);
5148	if (RT_FAILURE(rc))
5149	{
5150	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
5151	return rc;
5152	}
5153	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
5154	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
5155
5156	/*
5157	* Find a reasonable update interval and initialize the structure.
5158	*/
5159	u32Interval = u32SystemResolution = RTTimerGetSystemGranularity();
5160	while (u32Interval < 10000000 /* 10 ms */)
5161	u32Interval += u32SystemResolution;
5162
5163	supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), 1000000000 / u32Interval /=Hz/, cCpus);
5164
5165	/*
5166	* Create the timer.
5167	* If CPU_ALL isn't supported we'll have to fall back to synchronous mode.
5168	*/
5169	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
5170	{
5171	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, RTTIMER_FLAGS_CPU_ALL, supdrvGipAsyncTimer, pDevExt);
5172	if (rc == VERR_NOT_SUPPORTED)
5173	{
5174	OSDBGPRINT(("supdrvGipCreate: omni timer not supported, falling back to synchronous mode\n"));
5175	pGip->u32Mode = SUPGIPMODE_SYNC_TSC;
5176	}
5177	}
5178	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
5179	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0 /* fFlags */, supdrvGipSyncTimer, pDevExt);
5180	if (RT_SUCCESS(rc))
5181	{
5182	rc = RTMpNotificationRegister(supdrvGipMpEvent, pDevExt);
5183	if (RT_SUCCESS(rc))
5184	{
5185	rc = RTMpOnAll(supdrvGipInitOnCpu, pDevExt, pGip);
5186	if (RT_SUCCESS(rc))
5187	{
5188	/*
5189	* We're good.
5190	*/
5191	Log(("supdrvGipCreate: %u ns interval.\n", u32Interval));
5192	g_pSUPGlobalInfoPage = pGip;
5193	return VINF_SUCCESS;
5194	}
5195
5196	OSDBGPRINT(("supdrvGipCreate: RTMpOnAll failed with rc=%Rrc\n", rc));
5197	RTMpNotificationDeregister(supdrvGipMpEvent, pDevExt);
5198
5199	}
5200	else
5201	OSDBGPRINT(("supdrvGipCreate: failed to register MP event notfication. rc=%Rrc\n", rc));
5202	}
5203	else
5204	{
5205	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %u ns interval. rc=%Rrc\n", u32Interval, rc));
5206	Assert(!pDevExt->pGipTimer);
5207	}
5208	supdrvGipDestroy(pDevExt);
5209	return rc;
5210	}
5211
5212
5213	/**
5214	* Terminates the GIP.
5215	*
5216	* @param pDevExt Instance data. GIP stuff may be updated.
5217	*/
5218	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
5219	{
5220	int rc;
5221	#ifdef DEBUG_DARWIN_GIP
5222	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
5223	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
5224	pDevExt->pGipTimer, pDevExt->GipMemObj));
5225	#endif
5226
5227	/*
5228	* Invalid the GIP data.
5229	*/
5230	if (pDevExt->pGip)
5231	{
5232	supdrvGipTerm(pDevExt->pGip);
5233	pDevExt->pGip = NULL;
5234	}
5235	g_pSUPGlobalInfoPage = NULL;
5236
5237	/*
5238	* Destroy the timer and free the GIP memory object.
5239	*/
5240	if (pDevExt->pGipTimer)
5241	{
5242	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
5243	pDevExt->pGipTimer = NULL;
5244	}
5245
5246	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
5247	{
5248	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
5249	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
5250	}
5251
5252	/*
5253	* Finally, make sure we've release the system timer resolution request
5254	* if one actually succeeded and is still pending.
5255	*/
5256	if (pDevExt->u32SystemTimerGranularityGrant)
5257	{
5258	rc = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
5259	pDevExt->u32SystemTimerGranularityGrant = 0;
5260	}
5261	}
5262
5263
5264	/**
5265	* Timer callback function sync GIP mode.
5266	* @param pTimer The timer.
5267	* @param pvUser The device extension.
5268	*/
5269	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
5270	{
5271	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
5272	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5273	uint64_t u64TSC = ASMReadTSC();
5274	uint64_t NanoTS = RTTimeSystemNanoTS();
5275
5276	supdrvGipUpdate(pDevExt, NanoTS, u64TSC, NIL_RTCPUID, iTick);
5277
5278	ASMSetFlags(fOldFlags);
5279	}
5280
5281
5282	/**
5283	* Timer callback function for async GIP mode.
5284	* @param pTimer The timer.
5285	* @param pvUser The device extension.
5286	*/
5287	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
5288	{
5289	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
5290	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5291	RTCPUID idCpu = RTMpCpuId();
5292	uint64_t u64TSC = ASMReadTSC();
5293	uint64_t NanoTS = RTTimeSystemNanoTS();
5294
5295	/** @todo reset the transaction number and whatnot when iTick == 1. */
5296	if (pDevExt->idGipMaster == idCpu)
5297	supdrvGipUpdate(pDevExt, NanoTS, u64TSC, idCpu, iTick);
5298	else
5299	supdrvGipUpdatePerCpu(pDevExt, NanoTS, u64TSC, idCpu, ASMGetApicId(), iTick);
5300
5301	ASMSetFlags(fOldFlags);
5302	}
5303
5304
5305	/**
5306	* Finds our (@a idCpu) entry, or allocates a new one if not found.
5307	*
5308	* @returns Index of the CPU in the cache set.
5309	* @param pGip The GIP.
5310	* @param idCpu The CPU ID.
5311	*/
5312	static uint32_t supdrvGipCpuIndexFromCpuId(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
5313	{
5314	uint32_t i, cTries;
5315
5316	/*
5317	* ASSUMES that CPU IDs are constant.
5318	*/
5319	for (i = 0; i < pGip->cCpus; i++)
5320	if (pGip->aCPUs[i].idCpu == idCpu)
5321	return i;
5322
5323	cTries = 0;
5324	do
5325	{
5326	for (i = 0; i < pGip->cCpus; i++)
5327	{
5328	bool fRc;
5329	ASMAtomicCmpXchgSize(&pGip->aCPUs[i].idCpu, idCpu, NIL_RTCPUID, fRc);
5330	if (fRc)
5331	return i;
5332	}
5333	} while (cTries++ < 32);
5334	AssertReleaseFailed();
5335	return i - 1;
5336	}
5337
5338
5339	/**
5340	* The calling CPU should be accounted as online, update GIP accordingly.
5341	*
5342	* This is used by supdrvGipMpEvent as well as the supdrvGipCreate.
5343	*
5344	* @param pDevExt The device extension.
5345	* @param idCpu The CPU ID.
5346	*/
5347	static void supdrvGipMpEventOnline(PSUPDRVDEVEXT pDevExt, RTCPUID idCpu)
5348	{
5349	int iCpuSet = 0;
5350	uint16_t idApic = UINT16_MAX;
5351	uint32_t i = 0;
5352	uint64_t u64NanoTS = 0;
5353	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5354
5355	AssertPtrReturnVoid(pGip);
5356	AssertRelease(idCpu == RTMpCpuId());
5357	Assert(pGip->cPossibleCpus == RTMpGetCount());
5358
5359	/*
5360	* Do this behind a spinlock with interrupts disabled as this can fire
5361	* on all CPUs simultaneously, see @bugref{6110}.
5362	*/
5363	RTSpinlockAcquire(pDevExt->hGipSpinlock);
5364
5365	/*
5366	* Update the globals.
5367	*/
5368	ASMAtomicWriteU16(&pGip->cPresentCpus, RTMpGetPresentCount());
5369	ASMAtomicWriteU16(&pGip->cOnlineCpus, RTMpGetOnlineCount());
5370	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5371	if (iCpuSet >= 0)
5372	{
5373	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5374	RTCpuSetAddByIndex(&pGip->OnlineCpuSet, iCpuSet);
5375	RTCpuSetAddByIndex(&pGip->PresentCpuSet, iCpuSet);
5376	}
5377
5378	/*
5379	* Update the entry.
5380	*/
5381	u64NanoTS = RTTimeSystemNanoTS() - pGip->u32UpdateIntervalNS;
5382	i = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
5383	supdrvGipInitCpu(pGip, &pGip->aCPUs[i], u64NanoTS);
5384	idApic = ASMGetApicId();
5385	ASMAtomicWriteU16(&pGip->aCPUs[i].idApic, idApic);
5386	ASMAtomicWriteS16(&pGip->aCPUs[i].iCpuSet, (int16_t)iCpuSet);
5387	ASMAtomicWriteSize(&pGip->aCPUs[i].idCpu, idCpu);
5388
5389	/*
5390	* Update the APIC ID and CPU set index mappings.
5391	*/
5392	ASMAtomicWriteU16(&pGip->aiCpuFromApicId[idApic], i);
5393	ASMAtomicWriteU16(&pGip->aiCpuFromCpuSetIdx[iCpuSet], i);
5394
5395	/* commit it */
5396	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_ONLINE);
5397
5398	RTSpinlockReleaseNoInts(pDevExt->hGipSpinlock);
5399	}
5400
5401
5402	/**
5403	* The CPU should be accounted as offline, update the GIP accordingly.
5404	*
5405	* This is used by supdrvGipMpEvent.
5406	*
5407	* @param pDevExt The device extension.
5408	* @param idCpu The CPU ID.
5409	*/
5410	static void supdrvGipMpEventOffline(PSUPDRVDEVEXT pDevExt, RTCPUID idCpu)
5411	{
5412	int iCpuSet;
5413	unsigned i;
5414
5415	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5416
5417	AssertPtrReturnVoid(pGip);
5418	RTSpinlockAcquire(pDevExt->hGipSpinlock);
5419
5420	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5421	AssertReturnVoid(iCpuSet >= 0);
5422
5423	i = pGip->aiCpuFromCpuSetIdx[iCpuSet];
5424	AssertReturnVoid(i < pGip->cCpus);
5425	AssertReturnVoid(pGip->aCPUs[i].idCpu == idCpu);
5426
5427	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5428	RTCpuSetDelByIndex(&pGip->OnlineCpuSet, iCpuSet);
5429
5430	/* commit it */
5431	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_OFFLINE);
5432
5433	RTSpinlockReleaseNoInts(pDevExt->hGipSpinlock);
5434	}
5435
5436
5437	/**
5438	* Multiprocessor event notification callback.
5439	*
5440	* This is used to make sure that the GIP master gets passed on to
5441	* another CPU. It also updates the associated CPU data.
5442	*
5443	* @param enmEvent The event.
5444	* @param idCpu The cpu it applies to.
5445	* @param pvUser Pointer to the device extension.
5446	*
5447	* @remarks This function -must- fire on the newly online'd CPU for the
5448	* RTMPEVENT_ONLINE case and can fire on any CPU for the
5449	* RTMPEVENT_OFFLINE case.
5450	*/
5451	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser)
5452	{
5453	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5454	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5455
5456	AssertRelease(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
5457
5458	/*
5459	* Update the GIP CPU data.
5460	*/
5461	if (pGip)
5462	{
5463	switch (enmEvent)
5464	{
5465	case RTMPEVENT_ONLINE:
5466	AssertRelease(idCpu == RTMpCpuId());
5467	supdrvGipMpEventOnline(pDevExt, idCpu);
5468	break;
5469	case RTMPEVENT_OFFLINE:
5470	supdrvGipMpEventOffline(pDevExt, idCpu);
5471	break;
5472
5473	}
5474	}
5475
5476	/*
5477	* Make sure there is a master GIP.
5478	*/
5479	if (enmEvent == RTMPEVENT_OFFLINE)
5480	{
5481	RTCPUID idGipMaster = ASMAtomicReadU32(&pDevExt->idGipMaster);
5482	if (idGipMaster == idCpu)
5483	{
5484	/*
5485	* Find a new GIP master.
5486	*/
5487	bool fIgnored;
5488	unsigned i;
5489	RTCPUID idNewGipMaster = NIL_RTCPUID;
5490	RTCPUSET OnlineCpus;
5491	RTMpGetOnlineSet(&OnlineCpus);
5492
5493	for (i = 0; i < RTCPUSET_MAX_CPUS; i++)
5494	{
5495	RTCPUID idCurCpu = RTMpCpuIdFromSetIndex(i);
5496	if ( RTCpuSetIsMember(&OnlineCpus, idCurCpu)
5497	&& idCurCpu != idGipMaster)
5498	{
5499	idNewGipMaster = idCurCpu;
5500	break;
5501	}
5502	}
5503
5504	Log(("supdrvGipMpEvent: Gip master %#lx -> %#lx\n", (long)idGipMaster, (long)idNewGipMaster));
5505	ASMAtomicCmpXchgSize(&pDevExt->idGipMaster, idNewGipMaster, idGipMaster, fIgnored);
5506	NOREF(fIgnored);
5507	}
5508	}
5509	}
5510
5511
5512	/**
5513	* Callback used by supdrvDetermineAsyncTSC to read the TSC on a CPU.
5514	*
5515	* @param idCpu Ignored.
5516	* @param pvUser1 Where to put the TSC.
5517	* @param pvUser2 Ignored.
5518	*/
5519	static DECLCALLBACK(void) supdrvDetermineAsyncTscWorker(RTCPUID idCpu, void pvUser1, void pvUser2)
5520	{
5521	#if 1
5522	ASMAtomicWriteU64((uint64_t volatile *)pvUser1, ASMReadTSC());
5523	#else
5524	(uint64_t )pvUser1 = ASMReadTSC();
5525	#endif
5526	}
5527
5528
5529	/**
5530	* Determine if Async GIP mode is required because of TSC drift.
5531	*
5532	* When using the default/normal timer code it is essential that the time stamp counter
5533	* (TSC) runs never backwards, that is, a read operation to the counter should return
5534	* a bigger value than any previous read operation. This is guaranteed by the latest
5535	* AMD CPUs and by newer Intel CPUs which never enter the C2 state (P4). In any other
5536	* case we have to choose the asynchronous timer mode.
5537	*
5538	* @param poffMin Pointer to the determined difference between different cores.
5539	* @return false if the time stamp counters appear to be synchronized, true otherwise.
5540	*/
5541	static bool supdrvDetermineAsyncTsc(uint64_t *poffMin)
5542	{
5543	/*
5544	* Just iterate all the cpus 8 times and make sure that the TSC is
5545	* ever increasing. We don't bother taking TSC rollover into account.
5546	*/
5547	int iEndCpu = RTMpGetArraySize();
5548	int iCpu;
5549	int cLoops = 8;
5550	bool fAsync = false;
5551	int rc = VINF_SUCCESS;
5552	uint64_t offMax = 0;
5553	uint64_t offMin = ~(uint64_t)0;
5554	uint64_t PrevTsc = ASMReadTSC();
5555
5556	while (cLoops-- > 0)
5557	{
5558	for (iCpu = 0; iCpu < iEndCpu; iCpu++)
5559	{
5560	uint64_t CurTsc;
5561	rc = RTMpOnSpecific(RTMpCpuIdFromSetIndex(iCpu), supdrvDetermineAsyncTscWorker, &CurTsc, NULL);
5562	if (RT_SUCCESS(rc))
5563	{
5564	if (CurTsc <= PrevTsc)
5565	{
5566	fAsync = true;
5567	offMin = offMax = PrevTsc - CurTsc;
5568	Log(("supdrvDetermineAsyncTsc: iCpu=%d cLoops=%d CurTsc=%llx PrevTsc=%llx\n",
5569	iCpu, cLoops, CurTsc, PrevTsc));
5570	break;
5571	}
5572
5573	/* Gather statistics (except the first time). */
5574	if (iCpu != 0 \|\| cLoops != 7)
5575	{
5576	uint64_t off = CurTsc - PrevTsc;
5577	if (off < offMin)
5578	offMin = off;
5579	if (off > offMax)
5580	offMax = off;
5581	Log2(("%d/%d: off=%llx\n", cLoops, iCpu, off));
5582	}
5583
5584	/* Next */
5585	PrevTsc = CurTsc;
5586	}
5587	else if (rc == VERR_NOT_SUPPORTED)
5588	break;
5589	else
5590	AssertMsg(rc == VERR_CPU_NOT_FOUND \|\| rc == VERR_CPU_OFFLINE, ("%d\n", rc));
5591	}
5592
5593	/* broke out of the loop. */
5594	if (iCpu < iEndCpu)
5595	break;
5596	}
5597
5598	poffMin = offMin; / Almost RTMpOnSpecific profiling. */
5599	Log(("supdrvDetermineAsyncTsc: returns %d; iEndCpu=%d rc=%d offMin=%llx offMax=%llx\n",
5600	fAsync, iEndCpu, rc, offMin, offMax));
5601	#if !defined(RT_OS_SOLARIS) && !defined(RT_OS_OS2) && !defined(RT_OS_WINDOWS)
5602	OSDBGPRINT(("vboxdrv: fAsync=%d offMin=%#lx offMax=%#lx\n", fAsync, (long)offMin, (long)offMax));
5603	#endif
5604	return fAsync;
5605	}
5606
5607
5608	/**
5609	* Determine the GIP TSC mode.
5610	*
5611	* @returns The most suitable TSC mode.
5612	* @param pDevExt Pointer to the device instance data.
5613	*/
5614	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt)
5615	{
5616	/*
5617	* On SMP we're faced with two problems:
5618	* (1) There might be a skew between the CPU, so that cpu0
5619	* returns a TSC that is slightly different from cpu1.
5620	* (2) Power management (and other things) may cause the TSC
5621	* to run at a non-constant speed, and cause the speed
5622	* to be different on the cpus. This will result in (1).
5623	*
5624	* So, on SMP systems we'll have to select the ASYNC update method
5625	* if there are symptoms of these problems.
5626	*/
5627	if (RTMpGetCount() > 1)
5628	{
5629	uint32_t uEAX, uEBX, uECX, uEDX;
5630	uint64_t u64DiffCoresIgnored;
5631
5632	/* Permit the user and/or the OS specific bits to force async mode. */
5633	if (supdrvOSGetForcedAsyncTscMode(pDevExt))
5634	return SUPGIPMODE_ASYNC_TSC;
5635
5636	/* Try check for current differences between the cpus. */
5637	if (supdrvDetermineAsyncTsc(&u64DiffCoresIgnored))
5638	return SUPGIPMODE_ASYNC_TSC;
5639
5640	/*
5641	* If the CPU supports power management and is an AMD one we
5642	* won't trust it unless it has the TscInvariant bit is set.
5643	*/
5644	/* Check for "AuthenticAMD" */
5645	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
5646	if ( uEAX >= 1
5647	&& ASMIsAmdCpuEx(uEBX, uECX, uEDX))
5648	{
5649	/* Check for APM support and that TscInvariant is cleared. */
5650	ASMCpuId(0x80000000, &uEAX, &uEBX, &uECX, &uEDX);
5651	if (uEAX >= 0x80000007)
5652	{
5653	ASMCpuId(0x80000007, &uEAX, &uEBX, &uECX, &uEDX);
5654	if ( !(uEDX & RT_BIT(8))/* TscInvariant */
5655	&& (uEDX & 0x3e)) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
5656	return SUPGIPMODE_ASYNC_TSC;
5657	}
5658	}
5659	}
5660	return SUPGIPMODE_SYNC_TSC;
5661	}
5662
5663
5664	/**
5665	* Initializes per-CPU GIP information.
5666	*
5667	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5668	* @param pCpu Pointer to which GIP CPU to initalize.
5669	* @param u64NanoTS The current nanosecond timestamp.
5670	*/
5671	static void supdrvGipInitCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pCpu, uint64_t u64NanoTS)
5672	{
5673	pCpu->u32TransactionId = 2;
5674	pCpu->u64NanoTS = u64NanoTS;
5675	pCpu->u64TSC = ASMReadTSC();
5676
5677	ASMAtomicWriteSize(&pCpu->enmState, SUPGIPCPUSTATE_INVALID);
5678	ASMAtomicWriteSize(&pCpu->idCpu, NIL_RTCPUID);
5679	ASMAtomicWriteS16(&pCpu->iCpuSet, -1);
5680	ASMAtomicWriteU16(&pCpu->idApic, UINT16_MAX);
5681
5682	/*
5683	* We don't know the following values until we've executed updates.
5684	* So, we'll just pretend it's a 4 GHz CPU and adjust the history it on
5685	* the 2nd timer callout.
5686	*/
5687	pCpu->u64CpuHz = _4G + 1; /* tstGIP-2 depends on this. */
5688	pCpu->u32UpdateIntervalTSC
5689	= pCpu->au32TSCHistory[0]
5690	= pCpu->au32TSCHistory[1]
5691	= pCpu->au32TSCHistory[2]
5692	= pCpu->au32TSCHistory[3]
5693	= pCpu->au32TSCHistory[4]
5694	= pCpu->au32TSCHistory[5]
5695	= pCpu->au32TSCHistory[6]
5696	= pCpu->au32TSCHistory[7]
5697	= (uint32_t)(_4G / pGip->u32UpdateHz);
5698	}
5699
5700
5701	/**
5702	* Initializes the GIP data.
5703	*
5704	* @param pDevExt Pointer to the device instance data.
5705	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5706	* @param HCPhys The physical address of the GIP.
5707	* @param u64NanoTS The current nanosecond timestamp.
5708	* @param uUpdateHz The update frequency.
5709	* @param cCpus The CPU count.
5710	*/
5711	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
5712	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned cCpus)
5713	{
5714	size_t const cbGip = RT_ALIGN_Z(RT_OFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), PAGE_SIZE);
5715	unsigned i;
5716	#ifdef DEBUG_DARWIN_GIP
5717	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5718	#else
5719	LogFlow(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5720	#endif
5721
5722	/*
5723	* Initialize the structure.
5724	*/
5725	memset(pGip, 0, cbGip);
5726	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
5727	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
5728	pGip->u32Mode = supdrvGipDeterminTscMode(pDevExt);
5729	pGip->cCpus = (uint16_t)cCpus;
5730	pGip->cPages = (uint16_t)(cbGip / PAGE_SIZE);
5731	pGip->u32UpdateHz = uUpdateHz;
5732	pGip->u32UpdateIntervalNS = 1000000000 / uUpdateHz;
5733	pGip->u64NanoTSLastUpdateHz = u64NanoTS;
5734	RTCpuSetEmpty(&pGip->OnlineCpuSet);
5735	RTCpuSetEmpty(&pGip->PresentCpuSet);
5736	RTMpGetSet(&pGip->PossibleCpuSet);
5737	pGip->cOnlineCpus = RTMpGetOnlineCount();
5738	pGip->cPresentCpus = RTMpGetPresentCount();
5739	pGip->cPossibleCpus = RTMpGetCount();
5740	pGip->idCpuMax = RTMpGetMaxCpuId();
5741	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromApicId); i++)
5742	pGip->aiCpuFromApicId[i] = 0;
5743	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx); i++)
5744	pGip->aiCpuFromCpuSetIdx[i] = UINT16_MAX;
5745
5746	for (i = 0; i < cCpus; i++)
5747	supdrvGipInitCpu(pGip, &pGip->aCPUs[i], u64NanoTS);
5748
5749	/*
5750	* Link it to the device extension.
5751	*/
5752	pDevExt->pGip = pGip;
5753	pDevExt->HCPhysGip = HCPhys;
5754	pDevExt->cGipUsers = 0;
5755	}
5756
5757
5758	/**
5759	* On CPU initialization callback for RTMpOnAll.
5760	*
5761	* @param idCpu The CPU ID.
5762	* @param pvUser1 The device extension.
5763	* @param pvUser2 The GIP.
5764	*/
5765	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2)
5766	{
5767	/* This is good enough, even though it will update some of the globals a
5768	bit to much. */
5769	supdrvGipMpEventOnline((PSUPDRVDEVEXT)pvUser1, idCpu);
5770	}
5771
5772
5773	/**
5774	* Invalidates the GIP data upon termination.
5775	*
5776	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5777	*/
5778	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
5779	{
5780	unsigned i;
5781	pGip->u32Magic = 0;
5782	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
5783	{
5784	pGip->aCPUs[i].u64NanoTS = 0;
5785	pGip->aCPUs[i].u64TSC = 0;
5786	pGip->aCPUs[i].iTSCHistoryHead = 0;
5787	}
5788	}
5789
5790
5791	/**
5792	* Worker routine for supdrvGipUpdate and supdrvGipUpdatePerCpu that
5793	* updates all the per cpu data except the transaction id.
5794	*
5795	* @param pDevExt The device extension.
5796	* @param pGipCpu Pointer to the per cpu data.
5797	* @param u64NanoTS The current time stamp.
5798	* @param u64TSC The current TSC.
5799	* @param iTick The current timer tick.
5800	*/
5801	static void supdrvGipDoUpdateCpu(PSUPDRVDEVEXT pDevExt, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick)
5802	{
5803	uint64_t u64TSCDelta;
5804	uint32_t u32UpdateIntervalTSC;
5805	uint32_t u32UpdateIntervalTSCSlack;
5806	unsigned iTSCHistoryHead;
5807	uint64_t u64CpuHz;
5808	uint32_t u32TransactionId;
5809
5810	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5811	AssertPtrReturnVoid(pGip);
5812
5813	/* Delta between this and the previous update. */
5814	ASMAtomicUoWriteU32(&pGipCpu->u32PrevUpdateIntervalNS, (uint32_t)(u64NanoTS - pGipCpu->u64NanoTS));
5815
5816	/*
5817	* Update the NanoTS.
5818	*/
5819	ASMAtomicWriteU64(&pGipCpu->u64NanoTS, u64NanoTS);
5820
5821	/*
5822	* Calc TSC delta.
5823	*/
5824	/** @todo validate the NanoTS delta, don't trust the OS to call us when it should... */
5825	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
5826	ASMAtomicWriteU64(&pGipCpu->u64TSC, u64TSC);
5827
5828	if (u64TSCDelta >> 32)
5829	{
5830	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
5831	pGipCpu->cErrors++;
5832	}
5833
5834	/*
5835	* On the 2nd and 3rd callout, reset the history with the current TSC
5836	* interval since the values entered by supdrvGipInit are totally off.
5837	* The interval on the 1st callout completely unreliable, the 2nd is a bit
5838	* better, while the 3rd should be most reliable.
5839	*/
5840	u32TransactionId = pGipCpu->u32TransactionId;
5841	if (RT_UNLIKELY( ( u32TransactionId == 5
5842	\|\| u32TransactionId == 7)
5843	&& ( iTick == 2
5844	\|\| iTick == 3) ))
5845	{
5846	unsigned i;
5847	for (i = 0; i < RT_ELEMENTS(pGipCpu->au32TSCHistory); i++)
5848	ASMAtomicUoWriteU32(&pGipCpu->au32TSCHistory[i], (uint32_t)u64TSCDelta);
5849	}
5850
5851	/*
5852	* TSC History.
5853	*/
5854	Assert(RT_ELEMENTS(pGipCpu->au32TSCHistory) == 8);
5855	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
5856	ASMAtomicWriteU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
5857	ASMAtomicWriteU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
5858
5859	/*
5860	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
5861	*/
5862	if (pGip->u32UpdateHz >= 1000)
5863	{
5864	uint32_t u32;
5865	u32 = pGipCpu->au32TSCHistory[0];
5866	u32 += pGipCpu->au32TSCHistory[1];
5867	u32 += pGipCpu->au32TSCHistory[2];
5868	u32 += pGipCpu->au32TSCHistory[3];
5869	u32 >>= 2;
5870	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
5871	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
5872	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
5873	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
5874	u32UpdateIntervalTSC >>= 2;
5875	u32UpdateIntervalTSC += u32;
5876	u32UpdateIntervalTSC >>= 1;
5877
5878	/* Value chosen for a 2GHz Athlon64 running linux 2.6.10/11, . */
5879	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
5880	}
5881	else if (pGip->u32UpdateHz >= 90)
5882	{
5883	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5884	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
5885	u32UpdateIntervalTSC >>= 1;
5886
5887	/* value chosen on a 2GHz thinkpad running windows */
5888	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
5889	}
5890	else
5891	{
5892	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5893
5894	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
5895	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
5896	}
5897	ASMAtomicWriteU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
5898
5899	/*
5900	* CpuHz.
5901	*/
5902	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, pGip->u32UpdateHz);
5903	ASMAtomicWriteU64(&pGipCpu->u64CpuHz, u64CpuHz);
5904	}
5905
5906
5907	/**
5908	* Updates the GIP.
5909	*
5910	* @param pDevExt The device extension.
5911	* @param u64NanoTS The current nanosecond timesamp.
5912	* @param u64TSC The current TSC timesamp.
5913	* @param idCpu The CPU ID.
5914	* @param iTick The current timer tick.
5915	*/
5916	static void supdrvGipUpdate(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick)
5917	{
5918	/*
5919	* Determine the relevant CPU data.
5920	*/
5921	PSUPGIPCPU pGipCpu;
5922	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5923	AssertPtrReturnVoid(pGip);
5924
5925	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
5926	pGipCpu = &pGip->aCPUs[0];
5927	else
5928	{
5929	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
5930	if (RT_UNLIKELY(iCpu >= pGip->cCpus))
5931	return;
5932	pGipCpu = &pGip->aCPUs[iCpu];
5933	if (RT_UNLIKELY(pGipCpu->idCpu != idCpu))
5934	return;
5935	}
5936
5937	/*
5938	* Start update transaction.
5939	*/
5940	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5941	{
5942	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
5943	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5944	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5945	pGipCpu->cErrors++;
5946	return;
5947	}
5948
5949	/*
5950	* Recalc the update frequency every 0x800th time.
5951	*/
5952	if (!(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
5953	{
5954	if (pGip->u64NanoTSLastUpdateHz)
5955	{
5956	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
5957	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
5958	uint32_t u32UpdateHz = (uint32_t)((UINT64_C(1000000000) * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
5959	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
5960	{
5961	ASMAtomicWriteU32(&pGip->u32UpdateHz, u32UpdateHz);
5962	ASMAtomicWriteU32(&pGip->u32UpdateIntervalNS, 1000000000 / u32UpdateHz);
5963	}
5964	#endif
5965	}
5966	ASMAtomicWriteU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS);
5967	}
5968
5969	/*
5970	* Update the data.
5971	*/
5972	supdrvGipDoUpdateCpu(pDevExt, pGipCpu, u64NanoTS, u64TSC, iTick);
5973
5974	/*
5975	* Complete transaction.
5976	*/
5977	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5978	}
5979
5980
5981	/**
5982	* Updates the per cpu GIP data for the calling cpu.
5983	*
5984	* @param pDevExt The device extension.
5985	* @param u64NanoTS The current nanosecond timesamp.
5986	* @param u64TSC The current TSC timesamp.
5987	* @param idCpu The CPU ID.
5988	* @param idApic The APIC id for the CPU index.
5989	* @param iTick The current timer tick.
5990	*/
5991	static void supdrvGipUpdatePerCpu(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC,
5992	RTCPUID idCpu, uint8_t idApic, uint64_t iTick)
5993	{
5994	uint32_t iCpu;
5995	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5996
5997	/*
5998	* Avoid a potential race when a CPU online notification doesn't fire on
5999	* the onlined CPU but the tick creeps in before the event notification is
6000	* run.
6001	*/
6002	if (RT_UNLIKELY(iTick == 1))
6003	{
6004	iCpu = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
6005	if (pGip->aCPUs[iCpu].enmState == SUPGIPCPUSTATE_OFFLINE)
6006	supdrvGipMpEventOnline(pDevExt, idCpu);
6007	}
6008
6009	iCpu = pGip->aiCpuFromApicId[idApic];
6010	if (RT_LIKELY(iCpu < pGip->cCpus))
6011	{
6012	PSUPGIPCPU pGipCpu = &pGip->aCPUs[iCpu];
6013	if (pGipCpu->idCpu == idCpu)
6014	{
6015	/*
6016	* Start update transaction.
6017	*/
6018	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
6019	{
6020	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
6021	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
6022	pGipCpu->cErrors++;
6023	return;
6024	}
6025
6026	/*
6027	* Update the data.
6028	*/
6029	supdrvGipDoUpdateCpu(pDevExt, pGipCpu, u64NanoTS, u64TSC, iTick);
6030
6031	/*
6032	* Complete transaction.
6033	*/
6034	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
6035	}
6036	}
6037	}
6038

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

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDrv.c@ 48328

Download in other formats: