SUPDrv.c@ 49786

Last change on this file since 49786 was 49786, checked in by vboxsync, 11 years ago
HostDrivers: Fix SUPR0QueryVTCaps() to be preemption safe.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 228.4 KB

Line
1	/* $Id: SUPDrv.c 49786 2013-12-05 10:56:33Z vboxsync $ */
2	/** @file
3	* VBoxDrv - The VirtualBox Support Driver - Common code.
4	*/
5
6	/*
7	* Copyright (C) 2006-2013 Oracle Corporation
8	*
9	* This file is part of VirtualBox Open Source Edition (OSE), as
10	* available from http://www.virtualbox.org. This file is free software;
11	* you can redistribute it and/or modify it under the terms of the GNU
12	* General Public License (GPL) as published by the Free Software
13	* Foundation, in version 2 as it comes in the "COPYING" file of the
14	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16	*
17	* The contents of this file may alternatively be used under the terms
18	* of the Common Development and Distribution License Version 1.0
19	* (CDDL) only, as it comes in the "COPYING.CDDL" file of the
20	* VirtualBox OSE distribution, in which case the provisions of the
21	* CDDL are applicable instead of those of the GPL.
22	*
23	* You may elect to license modified versions of this file under the
24	* terms and conditions of either the GPL or the CDDL or both.
25	*/
26
27	/*******************************************************************************
28	* Header Files *
29	*******************************************************************************/
30	#define LOG_GROUP LOG_GROUP_SUP_DRV
31	#define SUPDRV_AGNOSTIC
32	#include "SUPDrvInternal.h"
33	#ifndef PAGE_SHIFT
34	# include <iprt/param.h>
35	#endif
36	#include <iprt/asm.h>
37	#include <iprt/asm-amd64-x86.h>
38	#include <iprt/asm-math.h>
39	#include <iprt/cpuset.h>
40	#include <iprt/handletable.h>
41	#include <iprt/mem.h>
42	#include <iprt/mp.h>
43	#include <iprt/power.h>
44	#include <iprt/process.h>
45	#include <iprt/semaphore.h>
46	#include <iprt/spinlock.h>
47	#include <iprt/thread.h>
48	#include <iprt/uuid.h>
49	#include <iprt/net.h>
50	#include <iprt/crc.h>
51	#include <iprt/string.h>
52	#include <iprt/timer.h>
53	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
54	# include <iprt/rand.h>
55	# include <iprt/path.h>
56	#endif
57	#include <iprt/x86.h>
58
59	#include <VBox/param.h>
60	#include <VBox/log.h>
61	#include <VBox/err.h>
62	#include <VBox/vmm/hm_svm.h>
63	#include <VBox/vmm/hm_vmx.h>
64
65	#if defined(RT_OS_SOLARIS) \|\| defined(RT_OS_DARWIN)
66	# include "dtrace/SUPDrv.h"
67	#else
68	# define VBOXDRV_SESSION_CREATE(pvSession, fUser) do { } while (0)
69	# define VBOXDRV_SESSION_CLOSE(pvSession) do { } while (0)
70	# define VBOXDRV_IOCTL_ENTRY(pvSession, uIOCtl, pvReqHdr) do { } while (0)
71	# define VBOXDRV_IOCTL_RETURN(pvSession, uIOCtl, pvReqHdr, rcRet, rcReq) do { } while (0)
72	#endif
73
74	/*
75	* Logging assignments:
76	* Log - useful stuff, like failures.
77	* LogFlow - program flow, except the really noisy bits.
78	* Log2 - Cleanup.
79	* Log3 - Loader flow noise.
80	* Log4 - Call VMMR0 flow noise.
81	* Log5 - Native yet-to-be-defined noise.
82	* Log6 - Native ioctl flow noise.
83	*
84	* Logging requires BUILD_TYPE=debug and possibly changes to the logger
85	* instantiation in log-vbox.c(pp).
86	*/
87
88
89	/*******************************************************************************
90	* Defined Constants And Macros *
91	*******************************************************************************/
92	/** The frequency by which we recalculate the u32UpdateHz and
93	* u32UpdateIntervalNS GIP members. The value must be a power of 2. */
94	#define GIP_UPDATEHZ_RECALC_FREQ 0x800
95
96	/** @def VBOX_SVN_REV
97	* The makefile should define this if it can. */
98	#ifndef VBOX_SVN_REV
99	# define VBOX_SVN_REV 0
100	#endif
101
102	#if 0 /* Don't start the GIP timers. Useful when debugging the IPRT timer code. */
103	# define DO_NOT_START_GIP
104	#endif
105
106
107	/*******************************************************************************
108	* Internal Functions *
109	*******************************************************************************/
110	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser);
111	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser);
112	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession);
113	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType);
114	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq);
115	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq);
116	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq);
117	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq);
118	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq);
119	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx);
120	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt);
121	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage);
122	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage);
123	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt);
124	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt);
125	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq);
126	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq);
127	static int supdrvIOCtl_MsrProber(PSUPDRVDEVEXT pDevExt, PSUPMSRPROBER pReq);
128	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt);
129	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt);
130	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
131	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
132	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser);
133	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
134	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned cCpus);
135	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2);
136	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip);
137	static void supdrvGipUpdate(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick);
138	static void supdrvGipUpdatePerCpu(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC,
139	RTCPUID idCpu, uint8_t idApic, uint64_t iTick);
140	static void supdrvGipInitCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pCpu, uint64_t u64NanoTS);
141
142
143	/*******************************************************************************
144	* Global Variables *
145	*******************************************************************************/
146	DECLEXPORT(PSUPGLOBALINFOPAGE) g_pSUPGlobalInfoPage = NULL;
147
148	/**
149	* Array of the R0 SUP API.
150	*/
151	static SUPFUNC g_aFunctions[] =
152	{
153	/* SED: START */
154	/* name function */
155	/* Entries with absolute addresses determined at runtime, fixup
156	code makes ugly ASSUMPTIONS about the order here: */
157	{ "SUPR0AbsIs64bit", (void *)0 },
158	{ "SUPR0Abs64bitKernelCS", (void *)0 },
159	{ "SUPR0Abs64bitKernelSS", (void *)0 },
160	{ "SUPR0Abs64bitKernelDS", (void *)0 },
161	{ "SUPR0AbsKernelCS", (void *)0 },
162	{ "SUPR0AbsKernelSS", (void *)0 },
163	{ "SUPR0AbsKernelDS", (void *)0 },
164	{ "SUPR0AbsKernelES", (void *)0 },
165	{ "SUPR0AbsKernelFS", (void *)0 },
166	{ "SUPR0AbsKernelGS", (void *)0 },
167	/* Normal function pointers: */
168	{ "g_pSUPGlobalInfoPage", (void )&g_pSUPGlobalInfoPage }, / SED: DATA */
169	{ "SUPGetGIP", (void *)SUPGetGIP },
170	{ "SUPR0ComponentDeregisterFactory", (void *)SUPR0ComponentDeregisterFactory },
171	{ "SUPR0ComponentQueryFactory", (void *)SUPR0ComponentQueryFactory },
172	{ "SUPR0ComponentRegisterFactory", (void *)SUPR0ComponentRegisterFactory },
173	{ "SUPR0ContAlloc", (void *)SUPR0ContAlloc },
174	{ "SUPR0ContFree", (void *)SUPR0ContFree },
175	{ "SUPR0EnableVTx", (void *)SUPR0EnableVTx },
176	{ "SUPR0SuspendVTxOnCpu", (void *)SUPR0SuspendVTxOnCpu },
177	{ "SUPR0ResumeVTxOnCpu", (void *)SUPR0ResumeVTxOnCpu },
178	{ "SUPR0GetPagingMode", (void *)SUPR0GetPagingMode },
179	{ "SUPR0LockMem", (void *)SUPR0LockMem },
180	{ "SUPR0LowAlloc", (void *)SUPR0LowAlloc },
181	{ "SUPR0LowFree", (void *)SUPR0LowFree },
182	{ "SUPR0MemAlloc", (void *)SUPR0MemAlloc },
183	{ "SUPR0MemFree", (void *)SUPR0MemFree },
184	{ "SUPR0MemGetPhys", (void *)SUPR0MemGetPhys },
185	{ "SUPR0ObjAddRef", (void *)SUPR0ObjAddRef },
186	{ "SUPR0ObjAddRefEx", (void *)SUPR0ObjAddRefEx },
187	{ "SUPR0ObjRegister", (void *)SUPR0ObjRegister },
188	{ "SUPR0ObjRelease", (void *)SUPR0ObjRelease },
189	{ "SUPR0ObjVerifyAccess", (void *)SUPR0ObjVerifyAccess },
190	{ "SUPR0PageAllocEx", (void *)SUPR0PageAllocEx },
191	{ "SUPR0PageFree", (void *)SUPR0PageFree },
192	{ "SUPR0Printf", (void *)SUPR0Printf },
193	{ "SUPR0TracerDeregisterDrv", (void *)SUPR0TracerDeregisterDrv },
194	{ "SUPR0TracerDeregisterImpl", (void *)SUPR0TracerDeregisterImpl },
195	{ "SUPR0TracerFireProbe", (void *)SUPR0TracerFireProbe },
196	{ "SUPR0TracerRegisterDrv", (void *)SUPR0TracerRegisterDrv },
197	{ "SUPR0TracerRegisterImpl", (void *)SUPR0TracerRegisterImpl },
198	{ "SUPR0TracerRegisterModule", (void *)SUPR0TracerRegisterModule },
199	{ "SUPR0TracerUmodProbeFire", (void *)SUPR0TracerUmodProbeFire },
200	{ "SUPR0UnlockMem", (void *)SUPR0UnlockMem },
201	{ "SUPSemEventClose", (void *)SUPSemEventClose },
202	{ "SUPSemEventCreate", (void *)SUPSemEventCreate },
203	{ "SUPSemEventGetResolution", (void *)SUPSemEventGetResolution },
204	{ "SUPSemEventMultiClose", (void *)SUPSemEventMultiClose },
205	{ "SUPSemEventMultiCreate", (void *)SUPSemEventMultiCreate },
206	{ "SUPSemEventMultiGetResolution", (void *)SUPSemEventMultiGetResolution },
207	{ "SUPSemEventMultiReset", (void *)SUPSemEventMultiReset },
208	{ "SUPSemEventMultiSignal", (void *)SUPSemEventMultiSignal },
209	{ "SUPSemEventMultiWait", (void *)SUPSemEventMultiWait },
210	{ "SUPSemEventMultiWaitNoResume", (void *)SUPSemEventMultiWaitNoResume },
211	{ "SUPSemEventMultiWaitNsAbsIntr", (void *)SUPSemEventMultiWaitNsAbsIntr },
212	{ "SUPSemEventMultiWaitNsRelIntr", (void *)SUPSemEventMultiWaitNsRelIntr },
213	{ "SUPSemEventSignal", (void *)SUPSemEventSignal },
214	{ "SUPSemEventWait", (void *)SUPSemEventWait },
215	{ "SUPSemEventWaitNoResume", (void *)SUPSemEventWaitNoResume },
216	{ "SUPSemEventWaitNsAbsIntr", (void *)SUPSemEventWaitNsAbsIntr },
217	{ "SUPSemEventWaitNsRelIntr", (void *)SUPSemEventWaitNsRelIntr },
218
219	{ "RTAssertAreQuiet", (void *)RTAssertAreQuiet },
220	{ "RTAssertMayPanic", (void *)RTAssertMayPanic },
221	{ "RTAssertMsg1", (void *)RTAssertMsg1 },
222	{ "RTAssertMsg2AddV", (void *)RTAssertMsg2AddV },
223	{ "RTAssertMsg2V", (void *)RTAssertMsg2V },
224	{ "RTAssertSetMayPanic", (void *)RTAssertSetMayPanic },
225	{ "RTAssertSetQuiet", (void *)RTAssertSetQuiet },
226	{ "RTCrc32", (void *)RTCrc32 },
227	{ "RTCrc32Finish", (void *)RTCrc32Finish },
228	{ "RTCrc32Process", (void *)RTCrc32Process },
229	{ "RTCrc32Start", (void *)RTCrc32Start },
230	{ "RTErrConvertFromErrno", (void *)RTErrConvertFromErrno },
231	{ "RTErrConvertToErrno", (void *)RTErrConvertToErrno },
232	{ "RTHandleTableAllocWithCtx", (void *)RTHandleTableAllocWithCtx },
233	{ "RTHandleTableCreate", (void *)RTHandleTableCreate },
234	{ "RTHandleTableCreateEx", (void *)RTHandleTableCreateEx },
235	{ "RTHandleTableDestroy", (void *)RTHandleTableDestroy },
236	{ "RTHandleTableFreeWithCtx", (void *)RTHandleTableFreeWithCtx },
237	{ "RTHandleTableLookupWithCtx", (void *)RTHandleTableLookupWithCtx },
238	{ "RTLogDefaultInstance", (void *)RTLogDefaultInstance },
239	{ "RTLogGetDefaultInstance", (void *)RTLogGetDefaultInstance },
240	{ "RTLogLoggerExV", (void *)RTLogLoggerExV },
241	{ "RTLogPrintfV", (void *)RTLogPrintfV },
242	{ "RTLogRelDefaultInstance", (void *)RTLogRelDefaultInstance },
243	{ "RTLogSetDefaultInstanceThread", (void *)RTLogSetDefaultInstanceThread },
244	{ "RTMemAllocExTag", (void *)RTMemAllocExTag },
245	{ "RTMemAllocTag", (void *)RTMemAllocTag },
246	{ "RTMemAllocVarTag", (void *)RTMemAllocVarTag },
247	{ "RTMemAllocZTag", (void *)RTMemAllocZTag },
248	{ "RTMemAllocZVarTag", (void *)RTMemAllocZVarTag },
249	{ "RTMemDupExTag", (void *)RTMemDupExTag },
250	{ "RTMemDupTag", (void *)RTMemDupTag },
251	{ "RTMemFree", (void *)RTMemFree },
252	{ "RTMemFreeEx", (void *)RTMemFreeEx },
253	{ "RTMemReallocTag", (void *)RTMemReallocTag },
254	{ "RTMpCpuId", (void *)RTMpCpuId },
255	{ "RTMpCpuIdFromSetIndex", (void *)RTMpCpuIdFromSetIndex },
256	{ "RTMpCpuIdToSetIndex", (void *)RTMpCpuIdToSetIndex },
257	{ "RTMpGetArraySize", (void *)RTMpGetArraySize },
258	{ "RTMpGetCount", (void *)RTMpGetCount },
259	{ "RTMpGetMaxCpuId", (void *)RTMpGetMaxCpuId },
260	{ "RTMpGetOnlineCount", (void *)RTMpGetOnlineCount },
261	{ "RTMpGetOnlineSet", (void *)RTMpGetOnlineSet },
262	{ "RTMpGetSet", (void *)RTMpGetSet },
263	{ "RTMpIsCpuOnline", (void *)RTMpIsCpuOnline },
264	{ "RTMpIsCpuPossible", (void *)RTMpIsCpuPossible },
265	{ "RTMpIsCpuWorkPending", (void *)RTMpIsCpuWorkPending },
266	{ "RTMpNotificationDeregister", (void *)RTMpNotificationDeregister },
267	{ "RTMpNotificationRegister", (void *)RTMpNotificationRegister },
268	{ "RTMpOnAll", (void *)RTMpOnAll },
269	{ "RTMpOnOthers", (void *)RTMpOnOthers },
270	{ "RTMpOnSpecific", (void *)RTMpOnSpecific },
271	{ "RTMpPokeCpu", (void *)RTMpPokeCpu },
272	{ "RTNetIPv4AddDataChecksum", (void *)RTNetIPv4AddDataChecksum },
273	{ "RTNetIPv4AddTCPChecksum", (void *)RTNetIPv4AddTCPChecksum },
274	{ "RTNetIPv4AddUDPChecksum", (void *)RTNetIPv4AddUDPChecksum },
275	{ "RTNetIPv4FinalizeChecksum", (void *)RTNetIPv4FinalizeChecksum },
276	{ "RTNetIPv4HdrChecksum", (void *)RTNetIPv4HdrChecksum },
277	{ "RTNetIPv4IsDHCPValid", (void *)RTNetIPv4IsDHCPValid },
278	{ "RTNetIPv4IsHdrValid", (void *)RTNetIPv4IsHdrValid },
279	{ "RTNetIPv4IsTCPSizeValid", (void *)RTNetIPv4IsTCPSizeValid },
280	{ "RTNetIPv4IsTCPValid", (void *)RTNetIPv4IsTCPValid },
281	{ "RTNetIPv4IsUDPSizeValid", (void *)RTNetIPv4IsUDPSizeValid },
282	{ "RTNetIPv4IsUDPValid", (void *)RTNetIPv4IsUDPValid },
283	{ "RTNetIPv4PseudoChecksum", (void *)RTNetIPv4PseudoChecksum },
284	{ "RTNetIPv4PseudoChecksumBits", (void *)RTNetIPv4PseudoChecksumBits },
285	{ "RTNetIPv4TCPChecksum", (void *)RTNetIPv4TCPChecksum },
286	{ "RTNetIPv4UDPChecksum", (void *)RTNetIPv4UDPChecksum },
287	{ "RTNetIPv6PseudoChecksum", (void *)RTNetIPv6PseudoChecksum },
288	{ "RTNetIPv6PseudoChecksumBits", (void *)RTNetIPv6PseudoChecksumBits },
289	{ "RTNetIPv6PseudoChecksumEx", (void *)RTNetIPv6PseudoChecksumEx },
290	{ "RTNetTCPChecksum", (void *)RTNetTCPChecksum },
291	{ "RTNetUDPChecksum", (void *)RTNetUDPChecksum },
292	{ "RTPowerNotificationDeregister", (void *)RTPowerNotificationDeregister },
293	{ "RTPowerNotificationRegister", (void *)RTPowerNotificationRegister },
294	{ "RTProcSelf", (void *)RTProcSelf },
295	{ "RTR0AssertPanicSystem", (void *)RTR0AssertPanicSystem },
296	{ "RTR0MemAreKrnlAndUsrDifferent", (void *)RTR0MemAreKrnlAndUsrDifferent },
297	{ "RTR0MemKernelIsValidAddr", (void *)RTR0MemKernelIsValidAddr },
298	{ "RTR0MemKernelCopyFrom", (void *)RTR0MemKernelCopyFrom },
299	{ "RTR0MemKernelCopyTo", (void *)RTR0MemKernelCopyTo },
300	{ "RTR0MemObjAddress", (void *)RTR0MemObjAddress },
301	{ "RTR0MemObjAddressR3", (void *)RTR0MemObjAddressR3 },
302	{ "RTR0MemObjAllocContTag", (void *)RTR0MemObjAllocContTag },
303	{ "RTR0MemObjAllocLowTag", (void *)RTR0MemObjAllocLowTag },
304	{ "RTR0MemObjAllocPageTag", (void *)RTR0MemObjAllocPageTag },
305	{ "RTR0MemObjAllocPhysExTag", (void *)RTR0MemObjAllocPhysExTag },
306	{ "RTR0MemObjAllocPhysNCTag", (void *)RTR0MemObjAllocPhysNCTag },
307	{ "RTR0MemObjAllocPhysTag", (void *)RTR0MemObjAllocPhysTag },
308	{ "RTR0MemObjEnterPhysTag", (void *)RTR0MemObjEnterPhysTag },
309	{ "RTR0MemObjFree", (void *)RTR0MemObjFree },
310	{ "RTR0MemObjGetPagePhysAddr", (void *)RTR0MemObjGetPagePhysAddr },
311	{ "RTR0MemObjIsMapping", (void *)RTR0MemObjIsMapping },
312	{ "RTR0MemObjLockUserTag", (void *)RTR0MemObjLockUserTag },
313	{ "RTR0MemObjMapKernelExTag", (void *)RTR0MemObjMapKernelExTag },
314	{ "RTR0MemObjMapKernelTag", (void *)RTR0MemObjMapKernelTag },
315	{ "RTR0MemObjMapUserTag", (void *)RTR0MemObjMapUserTag },
316	{ "RTR0MemObjProtect", (void *)RTR0MemObjProtect },
317	{ "RTR0MemObjSize", (void *)RTR0MemObjSize },
318	{ "RTR0MemUserCopyFrom", (void *)RTR0MemUserCopyFrom },
319	{ "RTR0MemUserCopyTo", (void *)RTR0MemUserCopyTo },
320	{ "RTR0MemUserIsValidAddr", (void *)RTR0MemUserIsValidAddr },
321	{ "RTR0ProcHandleSelf", (void *)RTR0ProcHandleSelf },
322	{ "RTSemEventCreate", (void *)RTSemEventCreate },
323	{ "RTSemEventDestroy", (void *)RTSemEventDestroy },
324	{ "RTSemEventGetResolution", (void *)RTSemEventGetResolution },
325	{ "RTSemEventMultiCreate", (void *)RTSemEventMultiCreate },
326	{ "RTSemEventMultiDestroy", (void *)RTSemEventMultiDestroy },
327	{ "RTSemEventMultiGetResolution", (void *)RTSemEventMultiGetResolution },
328	{ "RTSemEventMultiReset", (void *)RTSemEventMultiReset },
329	{ "RTSemEventMultiSignal", (void *)RTSemEventMultiSignal },
330	{ "RTSemEventMultiWait", (void *)RTSemEventMultiWait },
331	{ "RTSemEventMultiWaitEx", (void *)RTSemEventMultiWaitEx },
332	{ "RTSemEventMultiWaitExDebug", (void *)RTSemEventMultiWaitExDebug },
333	{ "RTSemEventMultiWaitNoResume", (void *)RTSemEventMultiWaitNoResume },
334	{ "RTSemEventSignal", (void *)RTSemEventSignal },
335	{ "RTSemEventWait", (void *)RTSemEventWait },
336	{ "RTSemEventWaitEx", (void *)RTSemEventWaitEx },
337	{ "RTSemEventWaitExDebug", (void *)RTSemEventWaitExDebug },
338	{ "RTSemEventWaitNoResume", (void *)RTSemEventWaitNoResume },
339	{ "RTSemFastMutexCreate", (void *)RTSemFastMutexCreate },
340	{ "RTSemFastMutexDestroy", (void *)RTSemFastMutexDestroy },
341	{ "RTSemFastMutexRelease", (void *)RTSemFastMutexRelease },
342	{ "RTSemFastMutexRequest", (void *)RTSemFastMutexRequest },
343	{ "RTSemMutexCreate", (void *)RTSemMutexCreate },
344	{ "RTSemMutexDestroy", (void *)RTSemMutexDestroy },
345	{ "RTSemMutexRelease", (void *)RTSemMutexRelease },
346	{ "RTSemMutexRequest", (void *)RTSemMutexRequest },
347	{ "RTSemMutexRequestDebug", (void *)RTSemMutexRequestDebug },
348	{ "RTSemMutexRequestNoResume", (void *)RTSemMutexRequestNoResume },
349	{ "RTSemMutexRequestNoResumeDebug", (void *)RTSemMutexRequestNoResumeDebug },
350	{ "RTSpinlockAcquire", (void *)RTSpinlockAcquire },
351	{ "RTSpinlockCreate", (void *)RTSpinlockCreate },
352	{ "RTSpinlockDestroy", (void *)RTSpinlockDestroy },
353	{ "RTSpinlockRelease", (void *)RTSpinlockRelease },
354	{ "RTSpinlockReleaseNoInts", (void *)RTSpinlockReleaseNoInts },
355	{ "RTStrCopy", (void *)RTStrCopy },
356	{ "RTStrDupTag", (void *)RTStrDupTag },
357	{ "RTStrFormat", (void *)RTStrFormat },
358	{ "RTStrFormatNumber", (void *)RTStrFormatNumber },
359	{ "RTStrFormatTypeDeregister", (void *)RTStrFormatTypeDeregister },
360	{ "RTStrFormatTypeRegister", (void *)RTStrFormatTypeRegister },
361	{ "RTStrFormatTypeSetUser", (void *)RTStrFormatTypeSetUser },
362	{ "RTStrFormatV", (void *)RTStrFormatV },
363	{ "RTStrFree", (void *)RTStrFree },
364	{ "RTStrNCmp", (void *)RTStrNCmp },
365	{ "RTStrPrintf", (void *)RTStrPrintf },
366	{ "RTStrPrintfEx", (void *)RTStrPrintfEx },
367	{ "RTStrPrintfExV", (void *)RTStrPrintfExV },
368	{ "RTStrPrintfV", (void *)RTStrPrintfV },
369	{ "RTThreadCreate", (void *)RTThreadCreate },
370	{ "RTThreadCtxHooksAreRegistered", (void *)RTThreadCtxHooksAreRegistered },
371	{ "RTThreadCtxHooksCreate", (void *)RTThreadCtxHooksCreate },
372	{ "RTThreadCtxHooksDeregister", (void *)RTThreadCtxHooksDeregister },
373	{ "RTThreadCtxHooksRegister", (void *)RTThreadCtxHooksRegister },
374	{ "RTThreadCtxHooksRelease", (void *)RTThreadCtxHooksRelease },
375	{ "RTThreadCtxHooksRetain", (void *)RTThreadCtxHooksRetain },
376	{ "RTThreadGetName", (void *)RTThreadGetName },
377	{ "RTThreadGetNative", (void *)RTThreadGetNative },
378	{ "RTThreadGetType", (void *)RTThreadGetType },
379	{ "RTThreadIsInInterrupt", (void *)RTThreadIsInInterrupt },
380	{ "RTThreadNativeSelf", (void *)RTThreadNativeSelf },
381	{ "RTThreadPreemptDisable", (void *)RTThreadPreemptDisable },
382	{ "RTThreadPreemptIsEnabled", (void *)RTThreadPreemptIsEnabled },
383	{ "RTThreadPreemptIsPending", (void *)RTThreadPreemptIsPending },
384	{ "RTThreadPreemptIsPendingTrusty", (void *)RTThreadPreemptIsPendingTrusty },
385	{ "RTThreadPreemptIsPossible", (void *)RTThreadPreemptIsPossible },
386	{ "RTThreadPreemptRestore", (void *)RTThreadPreemptRestore },
387	{ "RTThreadSelf", (void *)RTThreadSelf },
388	{ "RTThreadSelfName", (void *)RTThreadSelfName },
389	{ "RTThreadSleep", (void *)RTThreadSleep },
390	{ "RTThreadUserReset", (void *)RTThreadUserReset },
391	{ "RTThreadUserSignal", (void *)RTThreadUserSignal },
392	{ "RTThreadUserWait", (void *)RTThreadUserWait },
393	{ "RTThreadUserWaitNoResume", (void *)RTThreadUserWaitNoResume },
394	{ "RTThreadWait", (void *)RTThreadWait },
395	{ "RTThreadWaitNoResume", (void *)RTThreadWaitNoResume },
396	{ "RTThreadYield", (void *)RTThreadYield },
397	{ "RTTimeMilliTS", (void *)RTTimeMilliTS },
398	{ "RTTimeNanoTS", (void *)RTTimeNanoTS },
399	{ "RTTimeNow", (void *)RTTimeNow },
400	{ "RTTimerCanDoHighResolution", (void *)RTTimerCanDoHighResolution },
401	{ "RTTimerChangeInterval", (void *)RTTimerChangeInterval },
402	{ "RTTimerCreate", (void *)RTTimerCreate },
403	{ "RTTimerCreateEx", (void *)RTTimerCreateEx },
404	{ "RTTimerDestroy", (void *)RTTimerDestroy },
405	{ "RTTimerGetSystemGranularity", (void *)RTTimerGetSystemGranularity },
406	{ "RTTimerReleaseSystemGranularity", (void *)RTTimerReleaseSystemGranularity },
407	{ "RTTimerRequestSystemGranularity", (void *)RTTimerRequestSystemGranularity },
408	{ "RTTimerStart", (void *)RTTimerStart },
409	{ "RTTimerStop", (void *)RTTimerStop },
410	{ "RTTimeSystemMilliTS", (void *)RTTimeSystemMilliTS },
411	{ "RTTimeSystemNanoTS", (void *)RTTimeSystemNanoTS },
412	{ "RTUuidCompare", (void *)RTUuidCompare },
413	{ "RTUuidCompareStr", (void *)RTUuidCompareStr },
414	{ "RTUuidFromStr", (void *)RTUuidFromStr },
415	/* SED: END */
416	};
417
418	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
419	/**
420	* Drag in the rest of IRPT since we share it with the
421	* rest of the kernel modules on darwin.
422	*/
423	PFNRT g_apfnVBoxDrvIPRTDeps[] =
424	{
425	/* VBoxNetAdp */
426	(PFNRT)RTRandBytes,
427	/* VBoxUSB */
428	(PFNRT)RTPathStripFilename,
429	NULL
430	};
431	#endif /* RT_OS_DARWIN \|\| RT_OS_SOLARIS \|\| RT_OS_SOLARIS */
432
433
434	/**
435	* Initializes the device extentsion structure.
436	*
437	* @returns IPRT status code.
438	* @param pDevExt The device extension to initialize.
439	* @param cbSession The size of the session structure. The size of
440	* SUPDRVSESSION may be smaller when SUPDRV_AGNOSTIC is
441	* defined because we're skipping the OS specific members
442	* then.
443	*/
444	int VBOXCALL supdrvInitDevExt(PSUPDRVDEVEXT pDevExt, size_t cbSession)
445	{
446	int rc;
447
448	#ifdef SUPDRV_WITH_RELEASE_LOGGER
449	/*
450	* Create the release log.
451	*/
452	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
453	PRTLOGGER pRelLogger;
454	rc = RTLogCreate(&pRelLogger, 0 /* fFlags */, "all",
455	"VBOX_RELEASE_LOG", RT_ELEMENTS(s_apszGroups), s_apszGroups, RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER, NULL);
456	if (RT_SUCCESS(rc))
457	RTLogRelSetDefaultInstance(pRelLogger);
458	/** @todo Add native hook for getting logger config parameters and setting
459	* them. On linux we should use the module parameter stuff... */
460	#endif
461
462	/*
463	* Initialize it.
464	*/
465	memset(pDevExt, 0, sizeof(*pDevExt));
466	rc = RTSpinlockCreate(&pDevExt->Spinlock, RTSPINLOCK_FLAGS_INTERRUPT_SAFE, "SUPDrvDevExt");
467	if (RT_SUCCESS(rc))
468	{
469	rc = RTSpinlockCreate(&pDevExt->hGipSpinlock, RTSPINLOCK_FLAGS_INTERRUPT_SAFE, "SUPDrvGip");
470	if (RT_SUCCESS(rc))
471	{
472	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
473	rc = RTSemMutexCreate(&pDevExt->mtxLdr);
474	#else
475	rc = RTSemFastMutexCreate(&pDevExt->mtxLdr);
476	#endif
477	if (RT_SUCCESS(rc))
478	{
479	rc = RTSemFastMutexCreate(&pDevExt->mtxComponentFactory);
480	if (RT_SUCCESS(rc))
481	{
482	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
483	rc = RTSemMutexCreate(&pDevExt->mtxGip);
484	#else
485	rc = RTSemFastMutexCreate(&pDevExt->mtxGip);
486	#endif
487	if (RT_SUCCESS(rc))
488	{
489	rc = supdrvGipCreate(pDevExt);
490	if (RT_SUCCESS(rc))
491	{
492	rc = supdrvTracerInit(pDevExt);
493	if (RT_SUCCESS(rc))
494	{
495	pDevExt->pLdrInitImage = NULL;
496	pDevExt->hLdrInitThread = NIL_RTNATIVETHREAD;
497	pDevExt->u32Cookie = BIRD; /** @todo make this random? */
498	pDevExt->cbSession = (uint32_t)cbSession;
499
500	/*
501	* Fixup the absolute symbols.
502	*
503	* Because of the table indexing assumptions we'll have a little #ifdef orgy
504	* here rather than distributing this to OS specific files. At least for now.
505	*/
506	#ifdef RT_OS_DARWIN
507	# if ARCH_BITS == 32
508	if (SUPR0GetPagingMode() >= SUPPAGINGMODE_AMD64)
509	{
510	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
511	g_aFunctions[1].pfn = (void )0x80; / SUPR0Abs64bitKernelCS - KERNEL64_CS, seg.h */
512	g_aFunctions[2].pfn = (void )0x88; / SUPR0Abs64bitKernelSS - KERNEL64_SS, seg.h */
513	g_aFunctions[3].pfn = (void )0x88; / SUPR0Abs64bitKernelDS - KERNEL64_SS, seg.h */
514	}
515	else
516	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
517	g_aFunctions[4].pfn = (void )0x08; / SUPR0AbsKernelCS - KERNEL_CS, seg.h */
518	g_aFunctions[5].pfn = (void )0x10; / SUPR0AbsKernelSS - KERNEL_DS, seg.h */
519	g_aFunctions[6].pfn = (void )0x10; / SUPR0AbsKernelDS - KERNEL_DS, seg.h */
520	g_aFunctions[7].pfn = (void )0x10; / SUPR0AbsKernelES - KERNEL_DS, seg.h */
521	g_aFunctions[8].pfn = (void )0x10; / SUPR0AbsKernelFS - KERNEL_DS, seg.h */
522	g_aFunctions[9].pfn = (void )0x48; / SUPR0AbsKernelGS - CPU_DATA_GS, seg.h */
523	# else /* 64-bit darwin: */
524	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
525	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
526	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
527	g_aFunctions[3].pfn = (void )0; / SUPR0Abs64bitKernelDS */
528	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
529	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
530	g_aFunctions[6].pfn = (void )0; / SUPR0AbsKernelDS */
531	g_aFunctions[7].pfn = (void )0; / SUPR0AbsKernelES */
532	g_aFunctions[8].pfn = (void )0; / SUPR0AbsKernelFS */
533	g_aFunctions[9].pfn = (void )0; / SUPR0AbsKernelGS */
534
535	# endif
536	#else /* !RT_OS_DARWIN */
537	# if ARCH_BITS == 64
538	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
539	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
540	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
541	g_aFunctions[3].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0Abs64bitKernelDS */
542	# else
543	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
544	# endif
545	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
546	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
547	g_aFunctions[6].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0AbsKernelDS */
548	g_aFunctions[7].pfn = (void )(uintptr_t)ASMGetES(); / SUPR0AbsKernelES */
549	g_aFunctions[8].pfn = (void )(uintptr_t)ASMGetFS(); / SUPR0AbsKernelFS */
550	g_aFunctions[9].pfn = (void )(uintptr_t)ASMGetGS(); / SUPR0AbsKernelGS */
551	#endif /* !RT_OS_DARWIN */
552	return VINF_SUCCESS;
553	}
554
555	supdrvGipDestroy(pDevExt);
556	}
557
558	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
559	RTSemMutexDestroy(pDevExt->mtxGip);
560	pDevExt->mtxGip = NIL_RTSEMMUTEX;
561	#else
562	RTSemFastMutexDestroy(pDevExt->mtxGip);
563	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
564	#endif
565	}
566	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
567	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
568	}
569	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
570	RTSemMutexDestroy(pDevExt->mtxLdr);
571	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
572	#else
573	RTSemFastMutexDestroy(pDevExt->mtxLdr);
574	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
575	#endif
576	}
577	RTSpinlockDestroy(pDevExt->hGipSpinlock);
578	pDevExt->hGipSpinlock = NIL_RTSPINLOCK;
579	}
580	RTSpinlockDestroy(pDevExt->Spinlock);
581	pDevExt->Spinlock = NIL_RTSPINLOCK;
582	}
583	#ifdef SUPDRV_WITH_RELEASE_LOGGER
584	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
585	RTLogDestroy(RTLogSetDefaultInstance(NULL));
586	#endif
587
588	return rc;
589	}
590
591
592	/**
593	* Delete the device extension (e.g. cleanup members).
594	*
595	* @param pDevExt The device extension to delete.
596	*/
597	void VBOXCALL supdrvDeleteDevExt(PSUPDRVDEVEXT pDevExt)
598	{
599	PSUPDRVOBJ pObj;
600	PSUPDRVUSAGE pUsage;
601
602	/*
603	* Kill mutexes and spinlocks.
604	*/
605	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
606	RTSemMutexDestroy(pDevExt->mtxGip);
607	pDevExt->mtxGip = NIL_RTSEMMUTEX;
608	#else
609	RTSemFastMutexDestroy(pDevExt->mtxGip);
610	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
611	#endif
612	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
613	RTSemMutexDestroy(pDevExt->mtxLdr);
614	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
615	#else
616	RTSemFastMutexDestroy(pDevExt->mtxLdr);
617	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
618	#endif
619	RTSpinlockDestroy(pDevExt->Spinlock);
620	pDevExt->Spinlock = NIL_RTSPINLOCK;
621	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
622	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
623
624	/*
625	* Free lists.
626	*/
627	/* objects. */
628	pObj = pDevExt->pObjs;
629	Assert(!pObj); /* (can trigger on forced unloads) */
630	pDevExt->pObjs = NULL;
631	while (pObj)
632	{
633	void *pvFree = pObj;
634	pObj = pObj->pNext;
635	RTMemFree(pvFree);
636	}
637
638	/* usage records. */
639	pUsage = pDevExt->pUsageFree;
640	pDevExt->pUsageFree = NULL;
641	while (pUsage)
642	{
643	void *pvFree = pUsage;
644	pUsage = pUsage->pNext;
645	RTMemFree(pvFree);
646	}
647
648	/* kill the GIP. */
649	supdrvGipDestroy(pDevExt);
650	RTSpinlockDestroy(pDevExt->hGipSpinlock);
651	pDevExt->hGipSpinlock = NIL_RTSPINLOCK;
652
653	supdrvTracerTerm(pDevExt);
654
655	#ifdef SUPDRV_WITH_RELEASE_LOGGER
656	/* destroy the loggers. */
657	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
658	RTLogDestroy(RTLogSetDefaultInstance(NULL));
659	#endif
660	}
661
662
663	/**
664	* Create session.
665	*
666	* @returns IPRT status code.
667	* @param pDevExt Device extension.
668	* @param fUser Flag indicating whether this is a user or kernel
669	* session.
670	* @param fUnrestricted Unrestricted access (system) or restricted access
671	* (user)?
672	* @param ppSession Where to store the pointer to the session data.
673	*/
674	int VBOXCALL supdrvCreateSession(PSUPDRVDEVEXT pDevExt, bool fUser, bool fUnrestricted, PSUPDRVSESSION *ppSession)
675	{
676	int rc;
677	PSUPDRVSESSION pSession;
678
679	if (!SUP_IS_DEVEXT_VALID(pDevExt))
680	return VERR_INVALID_PARAMETER;
681
682	/*
683	* Allocate memory for the session data.
684	*/
685	pSession = *ppSession = (PSUPDRVSESSION)RTMemAllocZ(pDevExt->cbSession);
686	if (pSession)
687	{
688	/* Initialize session data. */
689	rc = RTSpinlockCreate(&pSession->Spinlock, RTSPINLOCK_FLAGS_INTERRUPT_UNSAFE, "SUPDrvSession");
690	if (!rc)
691	{
692	rc = RTHandleTableCreateEx(&pSession->hHandleTable,
693	RTHANDLETABLE_FLAGS_LOCKED_IRQ_SAFE \| RTHANDLETABLE_FLAGS_CONTEXT,
694	1 /uBase/, 32768 /cMax/, supdrvSessionObjHandleRetain, pSession);
695	if (RT_SUCCESS(rc))
696	{
697	Assert(pSession->Spinlock != NIL_RTSPINLOCK);
698	pSession->pDevExt = pDevExt;
699	pSession->u32Cookie = BIRD_INV;
700	pSession->fUnrestricted = fUnrestricted;
701	pSession->cRefs = 1;
702	/*pSession->pLdrUsage = NULL;
703	pSession->pVM = NULL;
704	pSession->pUsage = NULL;
705	pSession->pGip = NULL;
706	pSession->fGipReferenced = false;
707	pSession->Bundle.cUsed = 0; */
708	pSession->Uid = NIL_RTUID;
709	pSession->Gid = NIL_RTGID;
710	if (fUser)
711	{
712	pSession->Process = RTProcSelf();
713	pSession->R0Process = RTR0ProcHandleSelf();
714	}
715	else
716	{
717	pSession->Process = NIL_RTPROCESS;
718	pSession->R0Process = NIL_RTR0PROCESS;
719	}
720	/pSession->uTracerData = 0;/
721	pSession->hTracerCaller = NIL_RTNATIVETHREAD;
722	RTListInit(&pSession->TpProviders);
723	/pSession->cTpProviders = 0;/
724	/pSession->cTpProbesFiring = 0;/
725	RTListInit(&pSession->TpUmods);
726	/RT_ZERO(pSession->apTpLookupTable);/
727
728	VBOXDRV_SESSION_CREATE(pSession, fUser);
729	LogFlow(("Created session %p initial cookie=%#x\n", pSession, pSession->u32Cookie));
730	return VINF_SUCCESS;
731	}
732
733	RTSpinlockDestroy(pSession->Spinlock);
734	}
735	RTMemFree(pSession);
736	*ppSession = NULL;
737	Log(("Failed to create spinlock, rc=%d!\n", rc));
738	}
739	else
740	rc = VERR_NO_MEMORY;
741
742	return rc;
743	}
744
745
746	/**
747	* Shared code for cleaning up a session (but not quite freeing it).
748	*
749	* This is primarily intended for MAC OS X where we have to clean up the memory
750	* stuff before the file handle is closed.
751	*
752	* @param pDevExt Device extension.
753	* @param pSession Session data.
754	* This data will be freed by this routine.
755	*/
756	static void supdrvCleanupSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
757	{
758	int rc;
759	PSUPDRVBUNDLE pBundle;
760	LogFlow(("supdrvCleanupSession: pSession=%p\n", pSession));
761
762	/*
763	* Remove logger instances related to this session.
764	*/
765	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pSession);
766
767	/*
768	* Destroy the handle table.
769	*/
770	rc = RTHandleTableDestroy(pSession->hHandleTable, supdrvSessionObjHandleDelete, pSession);
771	AssertRC(rc);
772	pSession->hHandleTable = NIL_RTHANDLETABLE;
773
774	/*
775	* Release object references made in this session.
776	* In theory there should be noone racing us in this session.
777	*/
778	Log2(("release objects - start\n"));
779	if (pSession->pUsage)
780	{
781	PSUPDRVUSAGE pUsage;
782	RTSpinlockAcquire(pDevExt->Spinlock);
783
784	while ((pUsage = pSession->pUsage) != NULL)
785	{
786	PSUPDRVOBJ pObj = pUsage->pObj;
787	pSession->pUsage = pUsage->pNext;
788
789	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
790	if (pUsage->cUsage < pObj->cUsage)
791	{
792	pObj->cUsage -= pUsage->cUsage;
793	RTSpinlockRelease(pDevExt->Spinlock);
794	}
795	else
796	{
797	/* Destroy the object and free the record. */
798	if (pDevExt->pObjs == pObj)
799	pDevExt->pObjs = pObj->pNext;
800	else
801	{
802	PSUPDRVOBJ pObjPrev;
803	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
804	if (pObjPrev->pNext == pObj)
805	{
806	pObjPrev->pNext = pObj->pNext;
807	break;
808	}
809	Assert(pObjPrev);
810	}
811	RTSpinlockRelease(pDevExt->Spinlock);
812
813	Log(("supdrvCleanupSession: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
814	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
815	if (pObj->pfnDestructor)
816	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
817	RTMemFree(pObj);
818	}
819
820	/* free it and continue. */
821	RTMemFree(pUsage);
822
823	RTSpinlockAcquire(pDevExt->Spinlock);
824	}
825
826	RTSpinlockRelease(pDevExt->Spinlock);
827	AssertMsg(!pSession->pUsage, ("Some buster reregistered an object during desturction!\n"));
828	}
829	Log2(("release objects - done\n"));
830
831	/*
832	* Do tracer cleanups related to this session.
833	*/
834	Log2(("release tracer stuff - start\n"));
835	supdrvTracerCleanupSession(pDevExt, pSession);
836	Log2(("release tracer stuff - end\n"));
837
838	/*
839	* Release memory allocated in the session.
840	*
841	* We do not serialize this as we assume that the application will
842	* not allocated memory while closing the file handle object.
843	*/
844	Log2(("freeing memory:\n"));
845	pBundle = &pSession->Bundle;
846	while (pBundle)
847	{
848	PSUPDRVBUNDLE pToFree;
849	unsigned i;
850
851	/*
852	* Check and unlock all entries in the bundle.
853	*/
854	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
855	{
856	if (pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ)
857	{
858	Log2(("eType=%d pvR0=%p pvR3=%p cb=%ld\n", pBundle->aMem[i].eType, RTR0MemObjAddress(pBundle->aMem[i].MemObj),
859	(void *)RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3), (long)RTR0MemObjSize(pBundle->aMem[i].MemObj)));
860	if (pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ)
861	{
862	rc = RTR0MemObjFree(pBundle->aMem[i].MapObjR3, false);
863	AssertRC(rc); /** @todo figure out how to handle this. */
864	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
865	}
866	rc = RTR0MemObjFree(pBundle->aMem[i].MemObj, true /* fFreeMappings */);
867	AssertRC(rc); /** @todo figure out how to handle this. */
868	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
869	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
870	}
871	}
872
873	/*
874	* Advance and free previous bundle.
875	*/
876	pToFree = pBundle;
877	pBundle = pBundle->pNext;
878
879	pToFree->pNext = NULL;
880	pToFree->cUsed = 0;
881	if (pToFree != &pSession->Bundle)
882	RTMemFree(pToFree);
883	}
884	Log2(("freeing memory - done\n"));
885
886	/*
887	* Deregister component factories.
888	*/
889	RTSemFastMutexRequest(pDevExt->mtxComponentFactory);
890	Log2(("deregistering component factories:\n"));
891	if (pDevExt->pComponentFactoryHead)
892	{
893	PSUPDRVFACTORYREG pPrev = NULL;
894	PSUPDRVFACTORYREG pCur = pDevExt->pComponentFactoryHead;
895	while (pCur)
896	{
897	if (pCur->pSession == pSession)
898	{
899	/* unlink it */
900	PSUPDRVFACTORYREG pNext = pCur->pNext;
901	if (pPrev)
902	pPrev->pNext = pNext;
903	else
904	pDevExt->pComponentFactoryHead = pNext;
905
906	/* free it */
907	pCur->pNext = NULL;
908	pCur->pSession = NULL;
909	pCur->pFactory = NULL;
910	RTMemFree(pCur);
911
912	/* next */
913	pCur = pNext;
914	}
915	else
916	{
917	/* next */
918	pPrev = pCur;
919	pCur = pCur->pNext;
920	}
921	}
922	}
923	RTSemFastMutexRelease(pDevExt->mtxComponentFactory);
924	Log2(("deregistering component factories - done\n"));
925
926	/*
927	* Loaded images needs to be dereferenced and possibly freed up.
928	*/
929	supdrvLdrLock(pDevExt);
930	Log2(("freeing images:\n"));
931	if (pSession->pLdrUsage)
932	{
933	PSUPDRVLDRUSAGE pUsage = pSession->pLdrUsage;
934	pSession->pLdrUsage = NULL;
935	while (pUsage)
936	{
937	void *pvFree = pUsage;
938	PSUPDRVLDRIMAGE pImage = pUsage->pImage;
939	if (pImage->cUsage > pUsage->cUsage)
940	pImage->cUsage -= pUsage->cUsage;
941	else
942	supdrvLdrFree(pDevExt, pImage);
943	pUsage->pImage = NULL;
944	pUsage = pUsage->pNext;
945	RTMemFree(pvFree);
946	}
947	}
948	supdrvLdrUnlock(pDevExt);
949	Log2(("freeing images - done\n"));
950
951	/*
952	* Unmap the GIP.
953	*/
954	Log2(("umapping GIP:\n"));
955	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
956	{
957	SUPR0GipUnmap(pSession);
958	pSession->fGipReferenced = 0;
959	}
960	Log2(("umapping GIP - done\n"));
961	}
962
963
964	/**
965	* Shared code for cleaning up a session.
966	*
967	* @param pDevExt Device extension.
968	* @param pSession Session data.
969	* This data will be freed by this routine.
970	*/
971	static void supdrvCloseSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
972	{
973	VBOXDRV_SESSION_CLOSE(pSession);
974
975	/*
976	* Cleanup the session first.
977	*/
978	supdrvCleanupSession(pDevExt, pSession);
979
980	/*
981	* Free the rest of the session stuff.
982	*/
983	RTSpinlockDestroy(pSession->Spinlock);
984	pSession->Spinlock = NIL_RTSPINLOCK;
985	pSession->pDevExt = NULL;
986	RTMemFree(pSession);
987	LogFlow(("supdrvCloseSession: returns\n"));
988	}
989
990
991	/**
992	* Retain a session to make sure it doesn't go away while it is in use.
993	*
994	* @returns New reference count on success, UINT32_MAX on failure.
995	* @param pSession Session data.
996	*/
997	uint32_t VBOXCALL supdrvSessionRetain(PSUPDRVSESSION pSession)
998	{
999	uint32_t cRefs;
1000	AssertPtrReturn(pSession, UINT32_MAX);
1001	AssertReturn(SUP_IS_SESSION_VALID(pSession), UINT32_MAX);
1002
1003	cRefs = ASMAtomicIncU32(&pSession->cRefs);
1004	AssertMsg(cRefs > 1 && cRefs < _1M, ("%#x %p\n", cRefs, pSession));
1005	return cRefs;
1006	}
1007
1008
1009	/**
1010	* Releases a given session.
1011	*
1012	* @returns New reference count on success (0 if closed), UINT32_MAX on failure.
1013	* @param pSession Session data.
1014	*/
1015	uint32_t VBOXCALL supdrvSessionRelease(PSUPDRVSESSION pSession)
1016	{
1017	uint32_t cRefs;
1018	AssertPtrReturn(pSession, UINT32_MAX);
1019	AssertReturn(SUP_IS_SESSION_VALID(pSession), UINT32_MAX);
1020
1021	cRefs = ASMAtomicDecU32(&pSession->cRefs);
1022	AssertMsg(cRefs < _1M, ("%#x %p\n", cRefs, pSession));
1023	if (cRefs == 0)
1024	supdrvCloseSession(pSession->pDevExt, pSession);
1025	return cRefs;
1026	}
1027
1028
1029	/**
1030	* RTHandleTableDestroy callback used by supdrvCleanupSession.
1031	*
1032	* @returns IPRT status code, see SUPR0ObjAddRef.
1033	* @param hHandleTable The handle table handle. Ignored.
1034	* @param pvObj The object pointer.
1035	* @param pvCtx Context, the handle type. Ignored.
1036	* @param pvUser Session pointer.
1037	*/
1038	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser)
1039	{
1040	NOREF(pvCtx);
1041	NOREF(hHandleTable);
1042	return SUPR0ObjAddRefEx(pvObj, (PSUPDRVSESSION)pvUser, true /fNoBlocking/);
1043	}
1044
1045
1046	/**
1047	* RTHandleTableDestroy callback used by supdrvCleanupSession.
1048	*
1049	* @param hHandleTable The handle table handle. Ignored.
1050	* @param h The handle value. Ignored.
1051	* @param pvObj The object pointer.
1052	* @param pvCtx Context, the handle type. Ignored.
1053	* @param pvUser Session pointer.
1054	*/
1055	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser)
1056	{
1057	NOREF(pvCtx);
1058	NOREF(h);
1059	NOREF(hHandleTable);
1060	SUPR0ObjRelease(pvObj, (PSUPDRVSESSION)pvUser);
1061	}
1062
1063
1064	/**
1065	* Fast path I/O Control worker.
1066	*
1067	* @returns VBox status code that should be passed down to ring-3 unchanged.
1068	* @param uIOCtl Function number.
1069	* @param idCpu VMCPU id.
1070	* @param pDevExt Device extention.
1071	* @param pSession Session data.
1072	*/
1073	int VBOXCALL supdrvIOCtlFast(uintptr_t uIOCtl, VMCPUID idCpu, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
1074	{
1075	/*
1076	* We check the two prereqs after doing this only to allow the compiler to optimize things better.
1077	*/
1078	if (RT_LIKELY( RT_VALID_PTR(pSession)
1079	&& pSession->pVM
1080	&& pDevExt->pfnVMMR0EntryFast))
1081	{
1082	switch (uIOCtl)
1083	{
1084	case SUP_IOCTL_FAST_DO_RAW_RUN:
1085	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_RAW_RUN);
1086	break;
1087	case SUP_IOCTL_FAST_DO_HM_RUN:
1088	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_HM_RUN);
1089	break;
1090	case SUP_IOCTL_FAST_DO_NOP:
1091	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_NOP);
1092	break;
1093	default:
1094	return VERR_INTERNAL_ERROR;
1095	}
1096	return VINF_SUCCESS;
1097	}
1098	return VERR_INTERNAL_ERROR;
1099	}
1100
1101
1102	/**
1103	* Helper for supdrvIOCtl. Check if pszStr contains any character of pszChars.
1104	* We would use strpbrk here if this function would be contained in the RedHat kABI white
1105	* list, see http://www.kerneldrivers.org/RHEL5.
1106	*
1107	* @returns 1 if pszStr does contain any character of pszChars, 0 otherwise.
1108	* @param pszStr String to check
1109	* @param pszChars Character set
1110	*/
1111	static int supdrvCheckInvalidChar(const char pszStr, const char pszChars)
1112	{
1113	int chCur;
1114	while ((chCur = *pszStr++) != '\0')
1115	{
1116	int ch;
1117	const char *psz = pszChars;
1118	while ((ch = *psz++) != '\0')
1119	if (ch == chCur)
1120	return 1;
1121
1122	}
1123	return 0;
1124	}
1125
1126
1127
1128	/**
1129	* I/O Control inner worker (tracing reasons).
1130	*
1131	* @returns IPRT status code.
1132	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1133	*
1134	* @param uIOCtl Function number.
1135	* @param pDevExt Device extention.
1136	* @param pSession Session data.
1137	* @param pReqHdr The request header.
1138	*/
1139	static int supdrvIOCtlInnerUnrestricted(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1140	{
1141	/*
1142	* Validation macros
1143	*/
1144	#define REQ_CHECK_SIZES_EX(Name, cbInExpect, cbOutExpect) \
1145	do { \
1146	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect) \|\| pReqHdr->cbOut != (cbOutExpect))) \
1147	{ \
1148	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld. cbOut=%ld expected %ld.\n", \
1149	(long)pReqHdr->cbIn, (long)(cbInExpect), (long)pReqHdr->cbOut, (long)(cbOutExpect))); \
1150	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1151	} \
1152	} while (0)
1153
1154	#define REQ_CHECK_SIZES(Name) REQ_CHECK_SIZES_EX(Name, Name ## _SIZE_IN, Name ## _SIZE_OUT)
1155
1156	#define REQ_CHECK_SIZE_IN(Name, cbInExpect) \
1157	do { \
1158	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect))) \
1159	{ \
1160	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld.\n", \
1161	(long)pReqHdr->cbIn, (long)(cbInExpect))); \
1162	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1163	} \
1164	} while (0)
1165
1166	#define REQ_CHECK_SIZE_OUT(Name, cbOutExpect) \
1167	do { \
1168	if (RT_UNLIKELY(pReqHdr->cbOut != (cbOutExpect))) \
1169	{ \
1170	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbOut=%ld expected %ld.\n", \
1171	(long)pReqHdr->cbOut, (long)(cbOutExpect))); \
1172	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1173	} \
1174	} while (0)
1175
1176	#define REQ_CHECK_EXPR(Name, expr) \
1177	do { \
1178	if (RT_UNLIKELY(!(expr))) \
1179	{ \
1180	OSDBGPRINT(( #Name ": %s\n", #expr)); \
1181	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1182	} \
1183	} while (0)
1184
1185	#define REQ_CHECK_EXPR_FMT(expr, fmt) \
1186	do { \
1187	if (RT_UNLIKELY(!(expr))) \
1188	{ \
1189	OSDBGPRINT( fmt ); \
1190	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1191	} \
1192	} while (0)
1193
1194	/*
1195	* The switch.
1196	*/
1197	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
1198	{
1199	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
1200	{
1201	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
1202	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
1203	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
1204	{
1205	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
1206	pReq->Hdr.rc = VERR_INVALID_MAGIC;
1207	return 0;
1208	}
1209
1210	#if 0
1211	/*
1212	* Call out to the OS specific code and let it do permission checks on the
1213	* client process.
1214	*/
1215	if (!supdrvOSValidateClientProcess(pDevExt, pSession))
1216	{
1217	pReq->u.Out.u32Cookie = 0xffffffff;
1218	pReq->u.Out.u32SessionCookie = 0xffffffff;
1219	pReq->u.Out.u32SessionVersion = 0xffffffff;
1220	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1221	pReq->u.Out.pSession = NULL;
1222	pReq->u.Out.cFunctions = 0;
1223	pReq->Hdr.rc = VERR_PERMISSION_DENIED;
1224	return 0;
1225	}
1226	#endif
1227
1228	/*
1229	* Match the version.
1230	* The current logic is very simple, match the major interface version.
1231	*/
1232	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
1233	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
1234	{
1235	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1236	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
1237	pReq->u.Out.u32Cookie = 0xffffffff;
1238	pReq->u.Out.u32SessionCookie = 0xffffffff;
1239	pReq->u.Out.u32SessionVersion = 0xffffffff;
1240	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1241	pReq->u.Out.pSession = NULL;
1242	pReq->u.Out.cFunctions = 0;
1243	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1244	return 0;
1245	}
1246
1247	/*
1248	* Fill in return data and be gone.
1249	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
1250	* u32SessionVersion <= u32ReqVersion!
1251	*/
1252	/** @todo Somehow validate the client and negotiate a secure cookie... */
1253	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
1254	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
1255	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
1256	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1257	pReq->u.Out.pSession = pSession;
1258	pReq->u.Out.cFunctions = sizeof(g_aFunctions) / sizeof(g_aFunctions[0]);
1259	pReq->Hdr.rc = VINF_SUCCESS;
1260	return 0;
1261	}
1262
1263	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_QUERY_FUNCS(0)):
1264	{
1265	/* validate */
1266	PSUPQUERYFUNCS pReq = (PSUPQUERYFUNCS)pReqHdr;
1267	REQ_CHECK_SIZES_EX(SUP_IOCTL_QUERY_FUNCS, SUP_IOCTL_QUERY_FUNCS_SIZE_IN, SUP_IOCTL_QUERY_FUNCS_SIZE_OUT(RT_ELEMENTS(g_aFunctions)));
1268
1269	/* execute */
1270	pReq->u.Out.cFunctions = RT_ELEMENTS(g_aFunctions);
1271	memcpy(&pReq->u.Out.aFunctions[0], g_aFunctions, sizeof(g_aFunctions));
1272	pReq->Hdr.rc = VINF_SUCCESS;
1273	return 0;
1274	}
1275
1276	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_LOCK):
1277	{
1278	/* validate */
1279	PSUPPAGELOCK pReq = (PSUPPAGELOCK)pReqHdr;
1280	REQ_CHECK_SIZE_IN(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_IN);
1281	REQ_CHECK_SIZE_OUT(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_OUT(pReq->u.In.cPages));
1282	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.cPages > 0);
1283	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.pvR3 >= PAGE_SIZE);
1284
1285	/* execute */
1286	pReq->Hdr.rc = SUPR0LockMem(pSession, pReq->u.In.pvR3, pReq->u.In.cPages, &pReq->u.Out.aPages[0]);
1287	if (RT_FAILURE(pReq->Hdr.rc))
1288	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1289	return 0;
1290	}
1291
1292	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_UNLOCK):
1293	{
1294	/* validate */
1295	PSUPPAGEUNLOCK pReq = (PSUPPAGEUNLOCK)pReqHdr;
1296	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_UNLOCK);
1297
1298	/* execute */
1299	pReq->Hdr.rc = SUPR0UnlockMem(pSession, pReq->u.In.pvR3);
1300	return 0;
1301	}
1302
1303	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_ALLOC):
1304	{
1305	/* validate */
1306	PSUPCONTALLOC pReq = (PSUPCONTALLOC)pReqHdr;
1307	REQ_CHECK_SIZES(SUP_IOCTL_CONT_ALLOC);
1308
1309	/* execute */
1310	pReq->Hdr.rc = SUPR0ContAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.HCPhys);
1311	if (RT_FAILURE(pReq->Hdr.rc))
1312	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1313	return 0;
1314	}
1315
1316	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_FREE):
1317	{
1318	/* validate */
1319	PSUPCONTFREE pReq = (PSUPCONTFREE)pReqHdr;
1320	REQ_CHECK_SIZES(SUP_IOCTL_CONT_FREE);
1321
1322	/* execute */
1323	pReq->Hdr.rc = SUPR0ContFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1324	return 0;
1325	}
1326
1327	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_OPEN):
1328	{
1329	/* validate */
1330	PSUPLDROPEN pReq = (PSUPLDROPEN)pReqHdr;
1331	REQ_CHECK_SIZES(SUP_IOCTL_LDR_OPEN);
1332	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs > 0);
1333	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs < 16*_1M);
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 > 0);
1336	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits < pReq->u.In.cbImageWithTabs);
1337	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.szName[0]);
1338	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1339	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, !supdrvCheckInvalidChar(pReq->u.In.szName, ";:()[]{}/\\\|&*%#@!~`\"'"));
1340	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szFilename, sizeof(pReq->u.In.szFilename)));
1341
1342	/* execute */
1343	pReq->Hdr.rc = supdrvIOCtl_LdrOpen(pDevExt, pSession, pReq);
1344	return 0;
1345	}
1346
1347	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOAD):
1348	{
1349	/* validate */
1350	PSUPLDRLOAD pReq = (PSUPLDRLOAD)pReqHdr;
1351	REQ_CHECK_EXPR(Name, pReq->Hdr.cbIn >= sizeof(*pReq));
1352	REQ_CHECK_SIZES_EX(SUP_IOCTL_LDR_LOAD, SUP_IOCTL_LDR_LOAD_SIZE_IN(pReq->u.In.cbImageWithTabs), SUP_IOCTL_LDR_LOAD_SIZE_OUT);
1353	REQ_CHECK_EXPR(SUP_IOCTL_LDR_LOAD, pReq->u.In.cSymbols <= 16384);
1354	REQ_CHECK_EXPR_FMT( !pReq->u.In.cSymbols
1355	\|\| ( pReq->u.In.offSymbols < pReq->u.In.cbImageWithTabs
1356	&& pReq->u.In.offSymbols + pReq->u.In.cSymbols * sizeof(SUPLDRSYM) <= pReq->u.In.cbImageWithTabs),
1357	("SUP_IOCTL_LDR_LOAD: offSymbols=%#lx cSymbols=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offSymbols,
1358	(long)pReq->u.In.cSymbols, (long)pReq->u.In.cbImageWithTabs));
1359	REQ_CHECK_EXPR_FMT( !pReq->u.In.cbStrTab
1360	\|\| ( pReq->u.In.offStrTab < pReq->u.In.cbImageWithTabs
1361	&& pReq->u.In.offStrTab + pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs
1362	&& pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs),
1363	("SUP_IOCTL_LDR_LOAD: offStrTab=%#lx cbStrTab=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offStrTab,
1364	(long)pReq->u.In.cbStrTab, (long)pReq->u.In.cbImageWithTabs));
1365
1366	if (pReq->u.In.cSymbols)
1367	{
1368	uint32_t i;
1369	PSUPLDRSYM paSyms = (PSUPLDRSYM)&pReq->u.In.abImage[pReq->u.In.offSymbols];
1370	for (i = 0; i < pReq->u.In.cSymbols; i++)
1371	{
1372	REQ_CHECK_EXPR_FMT(paSyms[i].offSymbol < pReq->u.In.cbImageWithTabs,
1373	("SUP_IOCTL_LDR_LOAD: sym #%ld: symb off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offSymbol, (long)pReq->u.In.cbImageWithTabs));
1374	REQ_CHECK_EXPR_FMT(paSyms[i].offName < pReq->u.In.cbStrTab,
1375	("SUP_IOCTL_LDR_LOAD: sym #%ld: name off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1376	REQ_CHECK_EXPR_FMT(RTStrEnd((char const *)&pReq->u.In.abImage[pReq->u.In.offStrTab + paSyms[i].offName],
1377	pReq->u.In.cbStrTab - paSyms[i].offName),
1378	("SUP_IOCTL_LDR_LOAD: sym #%ld: unterminated name! (%#lx / %#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1379	}
1380	}
1381
1382	/* execute */
1383	pReq->Hdr.rc = supdrvIOCtl_LdrLoad(pDevExt, pSession, pReq);
1384	return 0;
1385	}
1386
1387	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_FREE):
1388	{
1389	/* validate */
1390	PSUPLDRFREE pReq = (PSUPLDRFREE)pReqHdr;
1391	REQ_CHECK_SIZES(SUP_IOCTL_LDR_FREE);
1392
1393	/* execute */
1394	pReq->Hdr.rc = supdrvIOCtl_LdrFree(pDevExt, pSession, pReq);
1395	return 0;
1396	}
1397
1398	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_GET_SYMBOL):
1399	{
1400	/* validate */
1401	PSUPLDRGETSYMBOL pReq = (PSUPLDRGETSYMBOL)pReqHdr;
1402	REQ_CHECK_SIZES(SUP_IOCTL_LDR_GET_SYMBOL);
1403	REQ_CHECK_EXPR(SUP_IOCTL_LDR_GET_SYMBOL, RTStrEnd(pReq->u.In.szSymbol, sizeof(pReq->u.In.szSymbol)));
1404
1405	/* execute */
1406	pReq->Hdr.rc = supdrvIOCtl_LdrGetSymbol(pDevExt, pSession, pReq);
1407	return 0;
1408	}
1409
1410	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0(0)):
1411	{
1412	/* validate */
1413	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1414	Log4(("SUP_IOCTL_CALL_VMMR0: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1415	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1416
1417	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_VMMR0_SIZE(0))
1418	{
1419	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(0), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(0));
1420
1421	/* execute */
1422	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1423	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg, pSession);
1424	else
1425	pReq->Hdr.rc = VERR_WRONG_ORDER;
1426	}
1427	else
1428	{
1429	PSUPVMMR0REQHDR pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1430	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR)),
1431	("SUP_IOCTL_CALL_VMMR0: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR))));
1432	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1433	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(pVMMReq->cbReq));
1434
1435	/* execute */
1436	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1437	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1438	else
1439	pReq->Hdr.rc = VERR_WRONG_ORDER;
1440	}
1441
1442	if ( RT_FAILURE(pReq->Hdr.rc)
1443	&& pReq->Hdr.rc != VERR_INTERRUPTED
1444	&& pReq->Hdr.rc != VERR_TIMEOUT)
1445	Log(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1446	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1447	else
1448	Log4(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1449	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1450	return 0;
1451	}
1452
1453	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0_BIG):
1454	{
1455	/* validate */
1456	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1457	PSUPVMMR0REQHDR pVMMReq;
1458	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1459	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1460
1461	pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1462	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR)),
1463	("SUP_IOCTL_CALL_VMMR0_BIG: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR))));
1464	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0_BIG, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1465	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));
1466
1467	/* execute */
1468	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1469	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1470	else
1471	pReq->Hdr.rc = VERR_WRONG_ORDER;
1472
1473	if ( RT_FAILURE(pReq->Hdr.rc)
1474	&& pReq->Hdr.rc != VERR_INTERRUPTED
1475	&& pReq->Hdr.rc != VERR_TIMEOUT)
1476	Log(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1477	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1478	else
1479	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1480	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1481	return 0;
1482	}
1483
1484	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_PAGING_MODE):
1485	{
1486	/* validate */
1487	PSUPGETPAGINGMODE pReq = (PSUPGETPAGINGMODE)pReqHdr;
1488	REQ_CHECK_SIZES(SUP_IOCTL_GET_PAGING_MODE);
1489
1490	/* execute */
1491	pReq->Hdr.rc = VINF_SUCCESS;
1492	pReq->u.Out.enmMode = SUPR0GetPagingMode();
1493	return 0;
1494	}
1495
1496	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_ALLOC):
1497	{
1498	/* validate */
1499	PSUPLOWALLOC pReq = (PSUPLOWALLOC)pReqHdr;
1500	REQ_CHECK_EXPR(SUP_IOCTL_LOW_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_LOW_ALLOC_SIZE_IN);
1501	REQ_CHECK_SIZES_EX(SUP_IOCTL_LOW_ALLOC, SUP_IOCTL_LOW_ALLOC_SIZE_IN, SUP_IOCTL_LOW_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1502
1503	/* execute */
1504	pReq->Hdr.rc = SUPR0LowAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1505	if (RT_FAILURE(pReq->Hdr.rc))
1506	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1507	return 0;
1508	}
1509
1510	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_FREE):
1511	{
1512	/* validate */
1513	PSUPLOWFREE pReq = (PSUPLOWFREE)pReqHdr;
1514	REQ_CHECK_SIZES(SUP_IOCTL_LOW_FREE);
1515
1516	/* execute */
1517	pReq->Hdr.rc = SUPR0LowFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1518	return 0;
1519	}
1520
1521	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_MAP):
1522	{
1523	/* validate */
1524	PSUPGIPMAP pReq = (PSUPGIPMAP)pReqHdr;
1525	REQ_CHECK_SIZES(SUP_IOCTL_GIP_MAP);
1526
1527	/* execute */
1528	pReq->Hdr.rc = SUPR0GipMap(pSession, &pReq->u.Out.pGipR3, &pReq->u.Out.HCPhysGip);
1529	if (RT_SUCCESS(pReq->Hdr.rc))
1530	pReq->u.Out.pGipR0 = pDevExt->pGip;
1531	return 0;
1532	}
1533
1534	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_UNMAP):
1535	{
1536	/* validate */
1537	PSUPGIPUNMAP pReq = (PSUPGIPUNMAP)pReqHdr;
1538	REQ_CHECK_SIZES(SUP_IOCTL_GIP_UNMAP);
1539
1540	/* execute */
1541	pReq->Hdr.rc = SUPR0GipUnmap(pSession);
1542	return 0;
1543	}
1544
1545	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SET_VM_FOR_FAST):
1546	{
1547	/* validate */
1548	PSUPSETVMFORFAST pReq = (PSUPSETVMFORFAST)pReqHdr;
1549	REQ_CHECK_SIZES(SUP_IOCTL_SET_VM_FOR_FAST);
1550	REQ_CHECK_EXPR_FMT( !pReq->u.In.pVMR0
1551	\|\| ( VALID_PTR(pReq->u.In.pVMR0)
1552	&& !((uintptr_t)pReq->u.In.pVMR0 & (PAGE_SIZE - 1))),
1553	("SUP_IOCTL_SET_VM_FOR_FAST: pVMR0=%p!\n", pReq->u.In.pVMR0));
1554	/* execute */
1555	pSession->pVM = pReq->u.In.pVMR0;
1556	pReq->Hdr.rc = VINF_SUCCESS;
1557	return 0;
1558	}
1559
1560	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_ALLOC_EX):
1561	{
1562	/* validate */
1563	PSUPPAGEALLOCEX pReq = (PSUPPAGEALLOCEX)pReqHdr;
1564	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC_EX, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_EX_SIZE_IN);
1565	REQ_CHECK_SIZES_EX(SUP_IOCTL_PAGE_ALLOC_EX, SUP_IOCTL_PAGE_ALLOC_EX_SIZE_IN, SUP_IOCTL_PAGE_ALLOC_EX_SIZE_OUT(pReq->u.In.cPages));
1566	REQ_CHECK_EXPR_FMT(pReq->u.In.fKernelMapping \|\| pReq->u.In.fUserMapping,
1567	("SUP_IOCTL_PAGE_ALLOC_EX: No mapping requested!\n"));
1568	REQ_CHECK_EXPR_FMT(pReq->u.In.fUserMapping,
1569	("SUP_IOCTL_PAGE_ALLOC_EX: Must have user mapping!\n"));
1570	REQ_CHECK_EXPR_FMT(!pReq->u.In.fReserved0 && !pReq->u.In.fReserved1,
1571	("SUP_IOCTL_PAGE_ALLOC_EX: fReserved0=%d fReserved1=%d\n", pReq->u.In.fReserved0, pReq->u.In.fReserved1));
1572
1573	/* execute */
1574	pReq->Hdr.rc = SUPR0PageAllocEx(pSession, pReq->u.In.cPages, 0 /* fFlags */,
1575	pReq->u.In.fUserMapping ? &pReq->u.Out.pvR3 : NULL,
1576	pReq->u.In.fKernelMapping ? &pReq->u.Out.pvR0 : NULL,
1577	&pReq->u.Out.aPages[0]);
1578	if (RT_FAILURE(pReq->Hdr.rc))
1579	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1580	return 0;
1581	}
1582
1583	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_MAP_KERNEL):
1584	{
1585	/* validate */
1586	PSUPPAGEMAPKERNEL pReq = (PSUPPAGEMAPKERNEL)pReqHdr;
1587	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_MAP_KERNEL);
1588	REQ_CHECK_EXPR_FMT(!pReq->u.In.fFlags, ("SUP_IOCTL_PAGE_MAP_KERNEL: fFlags=%#x! MBZ\n", pReq->u.In.fFlags));
1589	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_MAP_KERNEL: offSub=%#x\n", pReq->u.In.offSub));
1590	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1591	("SUP_IOCTL_PAGE_MAP_KERNEL: cbSub=%#x\n", pReq->u.In.cbSub));
1592
1593	/* execute */
1594	pReq->Hdr.rc = SUPR0PageMapKernel(pSession, pReq->u.In.pvR3, pReq->u.In.offSub, pReq->u.In.cbSub,
1595	pReq->u.In.fFlags, &pReq->u.Out.pvR0);
1596	if (RT_FAILURE(pReq->Hdr.rc))
1597	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1598	return 0;
1599	}
1600
1601	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_PROTECT):
1602	{
1603	/* validate */
1604	PSUPPAGEPROTECT pReq = (PSUPPAGEPROTECT)pReqHdr;
1605	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_PROTECT);
1606	REQ_CHECK_EXPR_FMT(!(pReq->u.In.fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)),
1607	("SUP_IOCTL_PAGE_PROTECT: fProt=%#x!\n", pReq->u.In.fProt));
1608	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_PROTECT: offSub=%#x\n", pReq->u.In.offSub));
1609	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1610	("SUP_IOCTL_PAGE_PROTECT: cbSub=%#x\n", pReq->u.In.cbSub));
1611
1612	/* execute */
1613	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);
1614	return 0;
1615	}
1616
1617	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_FREE):
1618	{
1619	/* validate */
1620	PSUPPAGEFREE pReq = (PSUPPAGEFREE)pReqHdr;
1621	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_FREE);
1622
1623	/* execute */
1624	pReq->Hdr.rc = SUPR0PageFree(pSession, pReq->u.In.pvR3);
1625	return 0;
1626	}
1627
1628	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_SERVICE(0)):
1629	{
1630	/* validate */
1631	PSUPCALLSERVICE pReq = (PSUPCALLSERVICE)pReqHdr;
1632	Log4(("SUP_IOCTL_CALL_SERVICE: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1633	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1634
1635	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
1636	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(0), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(0));
1637	else
1638	{
1639	PSUPR0SERVICEREQHDR pSrvReq = (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0];
1640	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR)),
1641	("SUP_IOCTL_CALL_SERVICE: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR))));
1642	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, pSrvReq->u32Magic == SUPR0SERVICEREQHDR_MAGIC);
1643	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(pSrvReq->cbReq), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(pSrvReq->cbReq));
1644	}
1645	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1646
1647	/* execute */
1648	pReq->Hdr.rc = supdrvIOCtl_CallServiceModule(pDevExt, pSession, pReq);
1649	return 0;
1650	}
1651
1652	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOGGER_SETTINGS(0)):
1653	{
1654	/* validate */
1655	PSUPLOGGERSETTINGS pReq = (PSUPLOGGERSETTINGS)pReqHdr;
1656	size_t cbStrTab;
1657	REQ_CHECK_SIZE_OUT(SUP_IOCTL_LOGGER_SETTINGS, SUP_IOCTL_LOGGER_SETTINGS_SIZE_OUT);
1658	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->Hdr.cbIn >= SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(1));
1659	cbStrTab = pReq->Hdr.cbIn - SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(0);
1660	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offGroups < cbStrTab);
1661	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offFlags < cbStrTab);
1662	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offDestination < cbStrTab);
1663	REQ_CHECK_EXPR_FMT(pReq->u.In.szStrings[cbStrTab - 1] == '\0',
1664	("SUP_IOCTL_LOGGER_SETTINGS: cbIn=%#x cbStrTab=%#zx LastChar=%d\n",
1665	pReq->Hdr.cbIn, cbStrTab, pReq->u.In.szStrings[cbStrTab - 1]));
1666	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhich <= SUPLOGGERSETTINGS_WHICH_RELEASE);
1667	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhat <= SUPLOGGERSETTINGS_WHAT_DESTROY);
1668
1669	/* execute */
1670	pReq->Hdr.rc = supdrvIOCtl_LoggerSettings(pDevExt, pSession, pReq);
1671	return 0;
1672	}
1673
1674	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP2):
1675	{
1676	/* validate */
1677	PSUPSEMOP2 pReq = (PSUPSEMOP2)pReqHdr;
1678	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP2, SUP_IOCTL_SEM_OP2_SIZE_IN, SUP_IOCTL_SEM_OP2_SIZE_OUT);
1679	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP2, pReq->u.In.uReserved == 0);
1680
1681	/* execute */
1682	switch (pReq->u.In.uType)
1683	{
1684	case SUP_SEM_TYPE_EVENT:
1685	{
1686	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1687	switch (pReq->u.In.uOp)
1688	{
1689	case SUPSEMOP2_WAIT_MS_REL:
1690	pReq->Hdr.rc = SUPSemEventWaitNoResume(pSession, hEvent, pReq->u.In.uArg.cRelMsTimeout);
1691	break;
1692	case SUPSEMOP2_WAIT_NS_ABS:
1693	pReq->Hdr.rc = SUPSemEventWaitNsAbsIntr(pSession, hEvent, pReq->u.In.uArg.uAbsNsTimeout);
1694	break;
1695	case SUPSEMOP2_WAIT_NS_REL:
1696	pReq->Hdr.rc = SUPSemEventWaitNsRelIntr(pSession, hEvent, pReq->u.In.uArg.cRelNsTimeout);
1697	break;
1698	case SUPSEMOP2_SIGNAL:
1699	pReq->Hdr.rc = SUPSemEventSignal(pSession, hEvent);
1700	break;
1701	case SUPSEMOP2_CLOSE:
1702	pReq->Hdr.rc = SUPSemEventClose(pSession, hEvent);
1703	break;
1704	case SUPSEMOP2_RESET:
1705	default:
1706	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1707	break;
1708	}
1709	break;
1710	}
1711
1712	case SUP_SEM_TYPE_EVENT_MULTI:
1713	{
1714	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1715	switch (pReq->u.In.uOp)
1716	{
1717	case SUPSEMOP2_WAIT_MS_REL:
1718	pReq->Hdr.rc = SUPSemEventMultiWaitNoResume(pSession, hEventMulti, pReq->u.In.uArg.cRelMsTimeout);
1719	break;
1720	case SUPSEMOP2_WAIT_NS_ABS:
1721	pReq->Hdr.rc = SUPSemEventMultiWaitNsAbsIntr(pSession, hEventMulti, pReq->u.In.uArg.uAbsNsTimeout);
1722	break;
1723	case SUPSEMOP2_WAIT_NS_REL:
1724	pReq->Hdr.rc = SUPSemEventMultiWaitNsRelIntr(pSession, hEventMulti, pReq->u.In.uArg.cRelNsTimeout);
1725	break;
1726	case SUPSEMOP2_SIGNAL:
1727	pReq->Hdr.rc = SUPSemEventMultiSignal(pSession, hEventMulti);
1728	break;
1729	case SUPSEMOP2_CLOSE:
1730	pReq->Hdr.rc = SUPSemEventMultiClose(pSession, hEventMulti);
1731	break;
1732	case SUPSEMOP2_RESET:
1733	pReq->Hdr.rc = SUPSemEventMultiReset(pSession, hEventMulti);
1734	break;
1735	default:
1736	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1737	break;
1738	}
1739	break;
1740	}
1741
1742	default:
1743	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1744	break;
1745	}
1746	return 0;
1747	}
1748
1749	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP3):
1750	{
1751	/* validate */
1752	PSUPSEMOP3 pReq = (PSUPSEMOP3)pReqHdr;
1753	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP3, SUP_IOCTL_SEM_OP3_SIZE_IN, SUP_IOCTL_SEM_OP3_SIZE_OUT);
1754	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, pReq->u.In.u32Reserved == 0 && pReq->u.In.u64Reserved == 0);
1755
1756	/* execute */
1757	switch (pReq->u.In.uType)
1758	{
1759	case SUP_SEM_TYPE_EVENT:
1760	{
1761	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1762	switch (pReq->u.In.uOp)
1763	{
1764	case SUPSEMOP3_CREATE:
1765	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1766	pReq->Hdr.rc = SUPSemEventCreate(pSession, &hEvent);
1767	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEvent;
1768	break;
1769	case SUPSEMOP3_GET_RESOLUTION:
1770	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1771	pReq->Hdr.rc = VINF_SUCCESS;
1772	pReq->Hdr.cbOut = sizeof(*pReq);
1773	pReq->u.Out.cNsResolution = SUPSemEventGetResolution(pSession);
1774	break;
1775	default:
1776	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1777	break;
1778	}
1779	break;
1780	}
1781
1782	case SUP_SEM_TYPE_EVENT_MULTI:
1783	{
1784	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1785	switch (pReq->u.In.uOp)
1786	{
1787	case SUPSEMOP3_CREATE:
1788	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1789	pReq->Hdr.rc = SUPSemEventMultiCreate(pSession, &hEventMulti);
1790	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEventMulti;
1791	break;
1792	case SUPSEMOP3_GET_RESOLUTION:
1793	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1794	pReq->Hdr.rc = VINF_SUCCESS;
1795	pReq->u.Out.cNsResolution = SUPSemEventMultiGetResolution(pSession);
1796	break;
1797	default:
1798	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1799	break;
1800	}
1801	break;
1802	}
1803
1804	default:
1805	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1806	break;
1807	}
1808	return 0;
1809	}
1810
1811	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
1812	{
1813	/* validate */
1814	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
1815	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
1816
1817	/* execute */
1818	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.Caps);
1819	if (RT_FAILURE(pReq->Hdr.rc))
1820	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1821	return 0;
1822	}
1823
1824	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_OPEN):
1825	{
1826	/* validate */
1827	PSUPTRACEROPEN pReq = (PSUPTRACEROPEN)pReqHdr;
1828	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_OPEN);
1829
1830	/* execute */
1831	pReq->Hdr.rc = supdrvIOCtl_TracerOpen(pDevExt, pSession, pReq->u.In.uCookie, pReq->u.In.uArg);
1832	return 0;
1833	}
1834
1835	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_CLOSE):
1836	{
1837	/* validate */
1838	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_CLOSE);
1839
1840	/* execute */
1841	pReqHdr->rc = supdrvIOCtl_TracerClose(pDevExt, pSession);
1842	return 0;
1843	}
1844
1845	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_IOCTL):
1846	{
1847	/* validate */
1848	PSUPTRACERIOCTL pReq = (PSUPTRACERIOCTL)pReqHdr;
1849	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_IOCTL);
1850
1851	/* execute */
1852	pReqHdr->rc = supdrvIOCtl_TracerIOCtl(pDevExt, pSession, pReq->u.In.uCmd, pReq->u.In.uArg, &pReq->u.Out.iRetVal);
1853	return 0;
1854	}
1855
1856	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_REG):
1857	{
1858	/* validate */
1859	PSUPTRACERUMODREG pReq = (PSUPTRACERUMODREG)pReqHdr;
1860	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_REG);
1861	if (!RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)))
1862	return VERR_INVALID_PARAMETER;
1863
1864	/* execute */
1865	pReqHdr->rc = supdrvIOCtl_TracerUmodRegister(pDevExt, pSession,
1866	pReq->u.In.R3PtrVtgHdr, pReq->u.In.uVtgHdrAddr,
1867	pReq->u.In.R3PtrStrTab, pReq->u.In.cbStrTab,
1868	pReq->u.In.szName, pReq->u.In.fFlags);
1869	return 0;
1870	}
1871
1872	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_DEREG):
1873	{
1874	/* validate */
1875	PSUPTRACERUMODDEREG pReq = (PSUPTRACERUMODDEREG)pReqHdr;
1876	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_DEREG);
1877
1878	/* execute */
1879	pReqHdr->rc = supdrvIOCtl_TracerUmodDeregister(pDevExt, pSession, pReq->u.In.pVtgHdr);
1880	return 0;
1881	}
1882
1883	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_FIRE_PROBE):
1884	{
1885	/* validate */
1886	PSUPTRACERUMODFIREPROBE pReq = (PSUPTRACERUMODFIREPROBE)pReqHdr;
1887	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_FIRE_PROBE);
1888
1889	supdrvIOCtl_TracerUmodProbeFire(pDevExt, pSession, &pReq->u.In);
1890	pReqHdr->rc = VINF_SUCCESS;
1891	return 0;
1892	}
1893
1894	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_MSR_PROBER):
1895	{
1896	/* validate */
1897	PSUPMSRPROBER pReq = (PSUPMSRPROBER)pReqHdr;
1898	REQ_CHECK_SIZES(SUP_IOCTL_MSR_PROBER);
1899	REQ_CHECK_EXPR(SUP_IOCTL_MSR_PROBER,
1900	pReq->u.In.enmOp > SUPMSRPROBEROP_INVALID && pReq->u.In.enmOp < SUPMSRPROBEROP_END);
1901
1902	pReqHdr->rc = supdrvIOCtl_MsrProber(pDevExt, pReq);
1903	return 0;
1904	}
1905
1906
1907	default:
1908	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1909	break;
1910	}
1911	return VERR_GENERAL_FAILURE;
1912	}
1913
1914
1915	/**
1916	* I/O Control inner worker for the restricted operations.
1917	*
1918	* @returns IPRT status code.
1919	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1920	*
1921	* @param uIOCtl Function number.
1922	* @param pDevExt Device extention.
1923	* @param pSession Session data.
1924	* @param pReqHdr The request header.
1925	*/
1926	static int supdrvIOCtlInnerRestricted(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1927	{
1928	/*
1929	* The switch.
1930	*/
1931	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
1932	{
1933	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
1934	{
1935	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
1936	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
1937	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
1938	{
1939	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
1940	pReq->Hdr.rc = VERR_INVALID_MAGIC;
1941	return 0;
1942	}
1943
1944	/*
1945	* Match the version.
1946	* The current logic is very simple, match the major interface version.
1947	*/
1948	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
1949	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
1950	{
1951	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1952	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
1953	pReq->u.Out.u32Cookie = 0xffffffff;
1954	pReq->u.Out.u32SessionCookie = 0xffffffff;
1955	pReq->u.Out.u32SessionVersion = 0xffffffff;
1956	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1957	pReq->u.Out.pSession = NULL;
1958	pReq->u.Out.cFunctions = 0;
1959	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1960	return 0;
1961	}
1962
1963	/*
1964	* Fill in return data and be gone.
1965	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
1966	* u32SessionVersion <= u32ReqVersion!
1967	*/
1968	/** @todo Somehow validate the client and negotiate a secure cookie... */
1969	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
1970	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
1971	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
1972	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1973	pReq->u.Out.pSession = pSession;
1974	pReq->u.Out.cFunctions = 0;
1975	pReq->Hdr.rc = VINF_SUCCESS;
1976	return 0;
1977	}
1978
1979	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
1980	{
1981	/* validate */
1982	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
1983	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
1984
1985	/* execute */
1986	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.Caps);
1987	if (RT_FAILURE(pReq->Hdr.rc))
1988	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1989	return 0;
1990	}
1991
1992	default:
1993	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1994	break;
1995	}
1996	return VERR_GENERAL_FAILURE;
1997	}
1998
1999
2000	/**
2001	* I/O Control worker.
2002	*
2003	* @returns IPRT status code.
2004	* @retval VERR_INVALID_PARAMETER if the request is invalid.
2005	*
2006	* @param uIOCtl Function number.
2007	* @param pDevExt Device extention.
2008	* @param pSession Session data.
2009	* @param pReqHdr The request header.
2010	*/
2011	int VBOXCALL supdrvIOCtl(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
2012	{
2013	int rc;
2014	VBOXDRV_IOCTL_ENTRY(pSession, uIOCtl, pReqHdr);
2015
2016	/*
2017	* Validate the request.
2018	*/
2019	/* this first check could probably be omitted as its also done by the OS specific code... */
2020	if (RT_UNLIKELY( (pReqHdr->fFlags & SUPREQHDR_FLAGS_MAGIC_MASK) != SUPREQHDR_FLAGS_MAGIC
2021	\|\| pReqHdr->cbIn < sizeof(*pReqHdr)
2022	\|\| pReqHdr->cbOut < sizeof(*pReqHdr)))
2023	{
2024	OSDBGPRINT(("vboxdrv: Bad ioctl request header; cbIn=%#lx cbOut=%#lx fFlags=%#lx\n",
2025	(long)pReqHdr->cbIn, (long)pReqHdr->cbOut, (long)pReqHdr->fFlags));
2026	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
2027	return VERR_INVALID_PARAMETER;
2028	}
2029	if (RT_UNLIKELY(!RT_VALID_PTR(pSession)))
2030	{
2031	OSDBGPRINT(("vboxdrv: Invalid pSession valud %p (ioctl=%p)\n", pSession, (void *)uIOCtl));
2032	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
2033	return VERR_INVALID_PARAMETER;
2034	}
2035	if (RT_UNLIKELY(uIOCtl == SUP_IOCTL_COOKIE))
2036	{
2037	if (pReqHdr->u32Cookie != SUPCOOKIE_INITIAL_COOKIE)
2038	{
2039	OSDBGPRINT(("SUP_IOCTL_COOKIE: bad cookie %#lx\n", (long)pReqHdr->u32Cookie));
2040	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
2041	return VERR_INVALID_PARAMETER;
2042	}
2043	}
2044	else if (RT_UNLIKELY( pReqHdr->u32Cookie != pDevExt->u32Cookie
2045	\|\| pReqHdr->u32SessionCookie != pSession->u32Cookie))
2046	{
2047	OSDBGPRINT(("vboxdrv: bad cookie %#lx / %#lx.\n", (long)pReqHdr->u32Cookie, (long)pReqHdr->u32SessionCookie));
2048	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
2049	return VERR_INVALID_PARAMETER;
2050	}
2051
2052	/*
2053	* Hand it to an inner function to avoid lots of unnecessary return tracepoints.
2054	*/
2055	if (pSession->fUnrestricted)
2056	rc = supdrvIOCtlInnerUnrestricted(uIOCtl, pDevExt, pSession, pReqHdr);
2057	else
2058	rc = supdrvIOCtlInnerRestricted(uIOCtl, pDevExt, pSession, pReqHdr);
2059
2060	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, pReqHdr->rc, rc);
2061	return rc;
2062	}
2063
2064
2065	/**
2066	* Inter-Driver Communication (IDC) worker.
2067	*
2068	* @returns VBox status code.
2069	* @retval VINF_SUCCESS on success.
2070	* @retval VERR_INVALID_PARAMETER if the request is invalid.
2071	* @retval VERR_NOT_SUPPORTED if the request isn't supported.
2072	*
2073	* @param uReq The request (function) code.
2074	* @param pDevExt Device extention.
2075	* @param pSession Session data.
2076	* @param pReqHdr The request header.
2077	*/
2078	int VBOXCALL supdrvIDC(uintptr_t uReq, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQHDR pReqHdr)
2079	{
2080	/*
2081	* The OS specific code has already validated the pSession
2082	* pointer, and the request size being greater or equal to
2083	* size of the header.
2084	*
2085	* So, just check that pSession is a kernel context session.
2086	*/
2087	if (RT_UNLIKELY( pSession
2088	&& pSession->R0Process != NIL_RTR0PROCESS))
2089	return VERR_INVALID_PARAMETER;
2090
2091	/*
2092	* Validation macro.
2093	*/
2094	#define REQ_CHECK_IDC_SIZE(Name, cbExpect) \
2095	do { \
2096	if (RT_UNLIKELY(pReqHdr->cb != (cbExpect))) \
2097	{ \
2098	OSDBGPRINT(( #Name ": Invalid input/output sizes. cb=%ld expected %ld.\n", \
2099	(long)pReqHdr->cb, (long)(cbExpect))); \
2100	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
2101	} \
2102	} while (0)
2103
2104	switch (uReq)
2105	{
2106	case SUPDRV_IDC_REQ_CONNECT:
2107	{
2108	PSUPDRVIDCREQCONNECT pReq = (PSUPDRVIDCREQCONNECT)pReqHdr;
2109	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_CONNECT, sizeof(*pReq));
2110
2111	/*
2112	* Validate the cookie and other input.
2113	*/
2114	if (pReq->Hdr.pSession != NULL)
2115	{
2116	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Hdr.pSession=%p expected NULL!\n", pReq->Hdr.pSession));
2117	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2118	}
2119	if (pReq->u.In.u32MagicCookie != SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE)
2120	{
2121	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: u32MagicCookie=%#x expected %#x!\n",
2122	(unsigned)pReq->u.In.u32MagicCookie, (unsigned)SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE));
2123	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2124	}
2125	if ( pReq->u.In.uMinVersion > pReq->u.In.uReqVersion
2126	\|\| (pReq->u.In.uMinVersion & UINT32_C(0xffff0000)) != (pReq->u.In.uReqVersion & UINT32_C(0xffff0000)))
2127	{
2128	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: uMinVersion=%#x uMaxVersion=%#x doesn't match!\n",
2129	pReq->u.In.uMinVersion, pReq->u.In.uReqVersion));
2130	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2131	}
2132	if (pSession != NULL)
2133	{
2134	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: pSession=%p expected NULL!\n", pSession));
2135	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2136	}
2137
2138	/*
2139	* Match the version.
2140	* The current logic is very simple, match the major interface version.
2141	*/
2142	if ( pReq->u.In.uMinVersion > SUPDRV_IDC_VERSION
2143	\|\| (pReq->u.In.uMinVersion & 0xffff0000) != (SUPDRV_IDC_VERSION & 0xffff0000))
2144	{
2145	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
2146	pReq->u.In.uReqVersion, pReq->u.In.uMinVersion, (unsigned)SUPDRV_IDC_VERSION));
2147	pReq->u.Out.pSession = NULL;
2148	pReq->u.Out.uSessionVersion = 0xffffffff;
2149	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
2150	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
2151	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
2152	return VINF_SUCCESS;
2153	}
2154
2155	pReq->u.Out.pSession = NULL;
2156	pReq->u.Out.uSessionVersion = SUPDRV_IDC_VERSION;
2157	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
2158	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
2159
2160	pReq->Hdr.rc = supdrvCreateSession(pDevExt, false /* fUser /, true /fUnrestricted*/, &pSession);
2161	if (RT_FAILURE(pReq->Hdr.rc))
2162	{
2163	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: failed to create session, rc=%d\n", pReq->Hdr.rc));
2164	return VINF_SUCCESS;
2165	}
2166
2167	pReq->u.Out.pSession = pSession;
2168	pReq->Hdr.pSession = pSession;
2169
2170	return VINF_SUCCESS;
2171	}
2172
2173	case SUPDRV_IDC_REQ_DISCONNECT:
2174	{
2175	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_DISCONNECT, sizeof(*pReqHdr));
2176
2177	supdrvSessionRelease(pSession);
2178	return pReqHdr->rc = VINF_SUCCESS;
2179	}
2180
2181	case SUPDRV_IDC_REQ_GET_SYMBOL:
2182	{
2183	PSUPDRVIDCREQGETSYM pReq = (PSUPDRVIDCREQGETSYM)pReqHdr;
2184	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_GET_SYMBOL, sizeof(*pReq));
2185
2186	pReq->Hdr.rc = supdrvIDC_LdrGetSymbol(pDevExt, pSession, pReq);
2187	return VINF_SUCCESS;
2188	}
2189
2190	case SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY:
2191	{
2192	PSUPDRVIDCREQCOMPREGFACTORY pReq = (PSUPDRVIDCREQCOMPREGFACTORY)pReqHdr;
2193	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY, sizeof(*pReq));
2194
2195	pReq->Hdr.rc = SUPR0ComponentRegisterFactory(pSession, pReq->u.In.pFactory);
2196	return VINF_SUCCESS;
2197	}
2198
2199	case SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY:
2200	{
2201	PSUPDRVIDCREQCOMPDEREGFACTORY pReq = (PSUPDRVIDCREQCOMPDEREGFACTORY)pReqHdr;
2202	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY, sizeof(*pReq));
2203
2204	pReq->Hdr.rc = SUPR0ComponentDeregisterFactory(pSession, pReq->u.In.pFactory);
2205	return VINF_SUCCESS;
2206	}
2207
2208	default:
2209	Log(("Unknown IDC %#lx\n", (long)uReq));
2210	break;
2211	}
2212
2213	#undef REQ_CHECK_IDC_SIZE
2214	return VERR_NOT_SUPPORTED;
2215	}
2216
2217
2218	/**
2219	* Register a object for reference counting.
2220	* The object is registered with one reference in the specified session.
2221	*
2222	* @returns Unique identifier on success (pointer).
2223	* All future reference must use this identifier.
2224	* @returns NULL on failure.
2225	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
2226	* @param pvUser1 The first user argument.
2227	* @param pvUser2 The second user argument.
2228	*/
2229	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
2230	{
2231	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2232	PSUPDRVOBJ pObj;
2233	PSUPDRVUSAGE pUsage;
2234
2235	/*
2236	* Validate the input.
2237	*/
2238	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
2239	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
2240	AssertPtrReturn(pfnDestructor, NULL);
2241
2242	/*
2243	* Allocate and initialize the object.
2244	*/
2245	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
2246	if (!pObj)
2247	return NULL;
2248	pObj->u32Magic = SUPDRVOBJ_MAGIC;
2249	pObj->enmType = enmType;
2250	pObj->pNext = NULL;
2251	pObj->cUsage = 1;
2252	pObj->pfnDestructor = pfnDestructor;
2253	pObj->pvUser1 = pvUser1;
2254	pObj->pvUser2 = pvUser2;
2255	pObj->CreatorUid = pSession->Uid;
2256	pObj->CreatorGid = pSession->Gid;
2257	pObj->CreatorProcess= pSession->Process;
2258	supdrvOSObjInitCreator(pObj, pSession);
2259
2260	/*
2261	* Allocate the usage record.
2262	* (We keep freed usage records around to simplify SUPR0ObjAddRefEx().)
2263	*/
2264	RTSpinlockAcquire(pDevExt->Spinlock);
2265
2266	pUsage = pDevExt->pUsageFree;
2267	if (pUsage)
2268	pDevExt->pUsageFree = pUsage->pNext;
2269	else
2270	{
2271	RTSpinlockRelease(pDevExt->Spinlock);
2272	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
2273	if (!pUsage)
2274	{
2275	RTMemFree(pObj);
2276	return NULL;
2277	}
2278	RTSpinlockAcquire(pDevExt->Spinlock);
2279	}
2280
2281	/*
2282	* Insert the object and create the session usage record.
2283	*/
2284	/* The object. */
2285	pObj->pNext = pDevExt->pObjs;
2286	pDevExt->pObjs = pObj;
2287
2288	/* The session record. */
2289	pUsage->cUsage = 1;
2290	pUsage->pObj = pObj;
2291	pUsage->pNext = pSession->pUsage;
2292	/* Log2(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext)); */
2293	pSession->pUsage = pUsage;
2294
2295	RTSpinlockRelease(pDevExt->Spinlock);
2296
2297	Log(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
2298	return pObj;
2299	}
2300
2301
2302	/**
2303	* Increment the reference counter for the object associating the reference
2304	* with the specified session.
2305	*
2306	* @returns IPRT status code.
2307	* @param pvObj The identifier returned by SUPR0ObjRegister().
2308	* @param pSession The session which is referencing the object.
2309	*
2310	* @remarks The caller should not own any spinlocks and must carefully protect
2311	* itself against potential race with the destructor so freed memory
2312	* isn't accessed here.
2313	*/
2314	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
2315	{
2316	return SUPR0ObjAddRefEx(pvObj, pSession, false /* fNoBlocking */);
2317	}
2318
2319
2320	/**
2321	* Increment the reference counter for the object associating the reference
2322	* with the specified session.
2323	*
2324	* @returns IPRT status code.
2325	* @retval VERR_TRY_AGAIN if fNoBlocking was set and a new usage record
2326	* couldn't be allocated. (If you see this you're not doing the right
2327	* thing and it won't ever work reliably.)
2328	*
2329	* @param pvObj The identifier returned by SUPR0ObjRegister().
2330	* @param pSession The session which is referencing the object.
2331	* @param fNoBlocking Set if it's not OK to block. Never try to make the
2332	* first reference to an object in a session with this
2333	* argument set.
2334	*
2335	* @remarks The caller should not own any spinlocks and must carefully protect
2336	* itself against potential race with the destructor so freed memory
2337	* isn't accessed here.
2338	*/
2339	SUPR0DECL(int) SUPR0ObjAddRefEx(void *pvObj, PSUPDRVSESSION pSession, bool fNoBlocking)
2340	{
2341	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2342	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2343	int rc = VINF_SUCCESS;
2344	PSUPDRVUSAGE pUsagePre;
2345	PSUPDRVUSAGE pUsage;
2346
2347	/*
2348	* Validate the input.
2349	* Be ready for the destruction race (someone might be stuck in the
2350	* destructor waiting a lock we own).
2351	*/
2352	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2353	AssertPtrReturn(pObj, VERR_INVALID_POINTER);
2354	AssertMsgReturn(pObj->u32Magic == SUPDRVOBJ_MAGIC \|\| pObj->u32Magic == SUPDRVOBJ_MAGIC_DEAD,
2355	("Invalid pvObj=%p magic=%#x (expected %#x or %#x)\n", pvObj, pObj->u32Magic, SUPDRVOBJ_MAGIC, SUPDRVOBJ_MAGIC_DEAD),
2356	VERR_INVALID_PARAMETER);
2357
2358	RTSpinlockAcquire(pDevExt->Spinlock);
2359
2360	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2361	{
2362	RTSpinlockRelease(pDevExt->Spinlock);
2363
2364	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2365	return VERR_WRONG_ORDER;
2366	}
2367
2368	/*
2369	* Preallocate the usage record if we can.
2370	*/
2371	pUsagePre = pDevExt->pUsageFree;
2372	if (pUsagePre)
2373	pDevExt->pUsageFree = pUsagePre->pNext;
2374	else if (!fNoBlocking)
2375	{
2376	RTSpinlockRelease(pDevExt->Spinlock);
2377	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
2378	if (!pUsagePre)
2379	return VERR_NO_MEMORY;
2380
2381	RTSpinlockAcquire(pDevExt->Spinlock);
2382	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2383	{
2384	RTSpinlockRelease(pDevExt->Spinlock);
2385
2386	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2387	return VERR_WRONG_ORDER;
2388	}
2389	}
2390
2391	/*
2392	* Reference the object.
2393	*/
2394	pObj->cUsage++;
2395
2396	/*
2397	* Look for the session record.
2398	*/
2399	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
2400	{
2401	/Log(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2402	if (pUsage->pObj == pObj)
2403	break;
2404	}
2405	if (pUsage)
2406	pUsage->cUsage++;
2407	else if (pUsagePre)
2408	{
2409	/* create a new session record. */
2410	pUsagePre->cUsage = 1;
2411	pUsagePre->pObj = pObj;
2412	pUsagePre->pNext = pSession->pUsage;
2413	pSession->pUsage = pUsagePre;
2414	/Log(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));/
2415
2416	pUsagePre = NULL;
2417	}
2418	else
2419	{
2420	pObj->cUsage--;
2421	rc = VERR_TRY_AGAIN;
2422	}
2423
2424	/*
2425	* Put any unused usage record into the free list..
2426	*/
2427	if (pUsagePre)
2428	{
2429	pUsagePre->pNext = pDevExt->pUsageFree;
2430	pDevExt->pUsageFree = pUsagePre;
2431	}
2432
2433	RTSpinlockRelease(pDevExt->Spinlock);
2434
2435	return rc;
2436	}
2437
2438
2439	/**
2440	* Decrement / destroy a reference counter record for an object.
2441	*
2442	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
2443	*
2444	* @returns IPRT status code.
2445	* @retval VINF_SUCCESS if not destroyed.
2446	* @retval VINF_OBJECT_DESTROYED if it's destroyed by this release call.
2447	* @retval VERR_INVALID_PARAMETER if the object isn't valid. Will assert in
2448	* string builds.
2449	*
2450	* @param pvObj The identifier returned by SUPR0ObjRegister().
2451	* @param pSession The session which is referencing the object.
2452	*/
2453	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
2454	{
2455	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2456	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2457	int rc = VERR_INVALID_PARAMETER;
2458	PSUPDRVUSAGE pUsage;
2459	PSUPDRVUSAGE pUsagePrev;
2460
2461	/*
2462	* Validate the input.
2463	*/
2464	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2465	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2466	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2467	VERR_INVALID_PARAMETER);
2468
2469	/*
2470	* Acquire the spinlock and look for the usage record.
2471	*/
2472	RTSpinlockAcquire(pDevExt->Spinlock);
2473
2474	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
2475	pUsage;
2476	pUsagePrev = pUsage, pUsage = pUsage->pNext)
2477	{
2478	/Log2(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2479	if (pUsage->pObj == pObj)
2480	{
2481	rc = VINF_SUCCESS;
2482	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
2483	if (pUsage->cUsage > 1)
2484	{
2485	pObj->cUsage--;
2486	pUsage->cUsage--;
2487	}
2488	else
2489	{
2490	/*
2491	* Free the session record.
2492	*/
2493	if (pUsagePrev)
2494	pUsagePrev->pNext = pUsage->pNext;
2495	else
2496	pSession->pUsage = pUsage->pNext;
2497	pUsage->pNext = pDevExt->pUsageFree;
2498	pDevExt->pUsageFree = pUsage;
2499
2500	/* What about the object? */
2501	if (pObj->cUsage > 1)
2502	pObj->cUsage--;
2503	else
2504	{
2505	/*
2506	* Object is to be destroyed, unlink it.
2507	*/
2508	pObj->u32Magic = SUPDRVOBJ_MAGIC_DEAD;
2509	rc = VINF_OBJECT_DESTROYED;
2510	if (pDevExt->pObjs == pObj)
2511	pDevExt->pObjs = pObj->pNext;
2512	else
2513	{
2514	PSUPDRVOBJ pObjPrev;
2515	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
2516	if (pObjPrev->pNext == pObj)
2517	{
2518	pObjPrev->pNext = pObj->pNext;
2519	break;
2520	}
2521	Assert(pObjPrev);
2522	}
2523	}
2524	}
2525	break;
2526	}
2527	}
2528
2529	RTSpinlockRelease(pDevExt->Spinlock);
2530
2531	/*
2532	* Call the destructor and free the object if required.
2533	*/
2534	if (rc == VINF_OBJECT_DESTROYED)
2535	{
2536	Log(("SUPR0ObjRelease: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
2537	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
2538	if (pObj->pfnDestructor)
2539	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
2540	RTMemFree(pObj);
2541	}
2542
2543	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
2544	return rc;
2545	}
2546
2547
2548	/**
2549	* Verifies that the current process can access the specified object.
2550	*
2551	* @returns The following IPRT status code:
2552	* @retval VINF_SUCCESS if access was granted.
2553	* @retval VERR_PERMISSION_DENIED if denied access.
2554	* @retval VERR_INVALID_PARAMETER if invalid parameter.
2555	*
2556	* @param pvObj The identifier returned by SUPR0ObjRegister().
2557	* @param pSession The session which wishes to access the object.
2558	* @param pszObjName Object string name. This is optional and depends on the object type.
2559	*
2560	* @remark The caller is responsible for making sure the object isn't removed while
2561	* we're inside this function. If uncertain about this, just call AddRef before calling us.
2562	*/
2563	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
2564	{
2565	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2566	int rc;
2567
2568	/*
2569	* Validate the input.
2570	*/
2571	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2572	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2573	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2574	VERR_INVALID_PARAMETER);
2575
2576	/*
2577	* Check access. (returns true if a decision has been made.)
2578	*/
2579	rc = VERR_INTERNAL_ERROR;
2580	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
2581	return rc;
2582
2583	/*
2584	* Default policy is to allow the user to access his own
2585	* stuff but nothing else.
2586	*/
2587	if (pObj->CreatorUid == pSession->Uid)
2588	return VINF_SUCCESS;
2589	return VERR_PERMISSION_DENIED;
2590	}
2591
2592
2593	/**
2594	* Lock pages.
2595	*
2596	* @returns IPRT status code.
2597	* @param pSession Session to which the locked memory should be associated.
2598	* @param pvR3 Start of the memory range to lock.
2599	* This must be page aligned.
2600	* @param cPages Number of pages to lock.
2601	* @param paPages Where to put the physical addresses of locked memory.
2602	*/
2603	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
2604	{
2605	int rc;
2606	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2607	const size_t cb = (size_t)cPages << PAGE_SHIFT;
2608	LogFlow(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
2609
2610	/*
2611	* Verify input.
2612	*/
2613	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2614	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
2615	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
2616	\|\| !pvR3)
2617	{
2618	Log(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
2619	return VERR_INVALID_PARAMETER;
2620	}
2621
2622	/*
2623	* Let IPRT do the job.
2624	*/
2625	Mem.eType = MEMREF_TYPE_LOCKED;
2626	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTMEM_PROT_READ \| RTMEM_PROT_WRITE, RTR0ProcHandleSelf());
2627	if (RT_SUCCESS(rc))
2628	{
2629	uint32_t iPage = cPages;
2630	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
2631	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
2632
2633	while (iPage-- > 0)
2634	{
2635	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2636	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
2637	{
2638	AssertMsgFailed(("iPage=%d\n", iPage));
2639	rc = VERR_INTERNAL_ERROR;
2640	break;
2641	}
2642	}
2643	if (RT_SUCCESS(rc))
2644	rc = supdrvMemAdd(&Mem, pSession);
2645	if (RT_FAILURE(rc))
2646	{
2647	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
2648	AssertRC(rc2);
2649	}
2650	}
2651
2652	return rc;
2653	}
2654
2655
2656	/**
2657	* Unlocks the memory pointed to by pv.
2658	*
2659	* @returns IPRT status code.
2660	* @param pSession Session to which the memory was locked.
2661	* @param pvR3 Memory to unlock.
2662	*/
2663	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2664	{
2665	LogFlow(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2666	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2667	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
2668	}
2669
2670
2671	/**
2672	* Allocates a chunk of page aligned memory with contiguous and fixed physical
2673	* backing.
2674	*
2675	* @returns IPRT status code.
2676	* @param pSession Session data.
2677	* @param cPages Number of pages to allocate.
2678	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
2679	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
2680	* @param pHCPhys Where to put the physical address of allocated memory.
2681	*/
2682	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
2683	{
2684	int rc;
2685	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2686	LogFlow(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
2687
2688	/*
2689	* Validate input.
2690	*/
2691	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2692	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
2693	{
2694	Log(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
2695	pSession, ppvR0, ppvR3, pHCPhys));
2696	return VERR_INVALID_PARAMETER;
2697
2698	}
2699	if (cPages < 1 \|\| cPages >= 256)
2700	{
2701	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2702	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2703	}
2704
2705	/*
2706	* Let IPRT do the job.
2707	*/
2708	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
2709	if (RT_SUCCESS(rc))
2710	{
2711	int rc2;
2712	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2713	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2714	if (RT_SUCCESS(rc))
2715	{
2716	Mem.eType = MEMREF_TYPE_CONT;
2717	rc = supdrvMemAdd(&Mem, pSession);
2718	if (!rc)
2719	{
2720	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2721	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2722	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
2723	return 0;
2724	}
2725
2726	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2727	AssertRC(rc2);
2728	}
2729	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2730	AssertRC(rc2);
2731	}
2732
2733	return rc;
2734	}
2735
2736
2737	/**
2738	* Frees memory allocated using SUPR0ContAlloc().
2739	*
2740	* @returns IPRT status code.
2741	* @param pSession The session to which the memory was allocated.
2742	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2743	*/
2744	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2745	{
2746	LogFlow(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2747	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2748	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
2749	}
2750
2751
2752	/**
2753	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
2754	*
2755	* The memory isn't zeroed.
2756	*
2757	* @returns IPRT status code.
2758	* @param pSession Session data.
2759	* @param cPages Number of pages to allocate.
2760	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
2761	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
2762	* @param paPages Where to put the physical addresses of allocated memory.
2763	*/
2764	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
2765	{
2766	unsigned iPage;
2767	int rc;
2768	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2769	LogFlow(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
2770
2771	/*
2772	* Validate input.
2773	*/
2774	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2775	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
2776	{
2777	Log(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
2778	pSession, ppvR3, ppvR0, paPages));
2779	return VERR_INVALID_PARAMETER;
2780
2781	}
2782	if (cPages < 1 \|\| cPages >= 256)
2783	{
2784	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2785	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2786	}
2787
2788	/*
2789	* Let IPRT do the work.
2790	*/
2791	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
2792	if (RT_SUCCESS(rc))
2793	{
2794	int rc2;
2795	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2796	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2797	if (RT_SUCCESS(rc))
2798	{
2799	Mem.eType = MEMREF_TYPE_LOW;
2800	rc = supdrvMemAdd(&Mem, pSession);
2801	if (!rc)
2802	{
2803	for (iPage = 0; iPage < cPages; iPage++)
2804	{
2805	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2806	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%RHp\n", paPages[iPage]));
2807	}
2808	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2809	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2810	return 0;
2811	}
2812
2813	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2814	AssertRC(rc2);
2815	}
2816
2817	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2818	AssertRC(rc2);
2819	}
2820
2821	return rc;
2822	}
2823
2824
2825	/**
2826	* Frees memory allocated using SUPR0LowAlloc().
2827	*
2828	* @returns IPRT status code.
2829	* @param pSession The session to which the memory was allocated.
2830	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2831	*/
2832	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2833	{
2834	LogFlow(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2835	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2836	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
2837	}
2838
2839
2840
2841	/**
2842	* Allocates a chunk of memory with both R0 and R3 mappings.
2843	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
2844	*
2845	* @returns IPRT status code.
2846	* @param pSession The session to associated the allocation with.
2847	* @param cb Number of bytes to allocate.
2848	* @param ppvR0 Where to store the address of the Ring-0 mapping.
2849	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2850	*/
2851	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
2852	{
2853	int rc;
2854	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2855	LogFlow(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
2856
2857	/*
2858	* Validate input.
2859	*/
2860	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2861	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
2862	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
2863	if (cb < 1 \|\| cb >= _4M)
2864	{
2865	Log(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
2866	return VERR_INVALID_PARAMETER;
2867	}
2868
2869	/*
2870	* Let IPRT do the work.
2871	*/
2872	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
2873	if (RT_SUCCESS(rc))
2874	{
2875	int rc2;
2876	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2877	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2878	if (RT_SUCCESS(rc))
2879	{
2880	Mem.eType = MEMREF_TYPE_MEM;
2881	rc = supdrvMemAdd(&Mem, pSession);
2882	if (!rc)
2883	{
2884	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2885	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2886	return VINF_SUCCESS;
2887	}
2888
2889	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2890	AssertRC(rc2);
2891	}
2892
2893	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2894	AssertRC(rc2);
2895	}
2896
2897	return rc;
2898	}
2899
2900
2901	/**
2902	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
2903	*
2904	* @returns IPRT status code.
2905	* @param pSession The session to which the memory was allocated.
2906	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2907	* @param paPages Where to store the physical addresses.
2908	*/
2909	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
2910	{
2911	PSUPDRVBUNDLE pBundle;
2912	LogFlow(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
2913
2914	/*
2915	* Validate input.
2916	*/
2917	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2918	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
2919	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
2920
2921	/*
2922	* Search for the address.
2923	*/
2924	RTSpinlockAcquire(pSession->Spinlock);
2925	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2926	{
2927	if (pBundle->cUsed > 0)
2928	{
2929	unsigned i;
2930	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2931	{
2932	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
2933	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2934	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2935	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2936	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
2937	)
2938	)
2939	{
2940	const size_t cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2941	size_t iPage;
2942	for (iPage = 0; iPage < cPages; iPage++)
2943	{
2944	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2945	paPages[iPage].uReserved = 0;
2946	}
2947	RTSpinlockRelease(pSession->Spinlock);
2948	return VINF_SUCCESS;
2949	}
2950	}
2951	}
2952	}
2953	RTSpinlockRelease(pSession->Spinlock);
2954	Log(("Failed to find %p!!!\n", (void *)uPtr));
2955	return VERR_INVALID_PARAMETER;
2956	}
2957
2958
2959	/**
2960	* Free memory allocated by SUPR0MemAlloc().
2961	*
2962	* @returns IPRT status code.
2963	* @param pSession The session owning the allocation.
2964	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2965	*/
2966	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2967	{
2968	LogFlow(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2969	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2970	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
2971	}
2972
2973
2974	/**
2975	* Allocates a chunk of memory with a kernel or/and a user mode mapping.
2976	*
2977	* The memory is fixed and it's possible to query the physical addresses using
2978	* SUPR0MemGetPhys().
2979	*
2980	* @returns IPRT status code.
2981	* @param pSession The session to associated the allocation with.
2982	* @param cPages The number of pages to allocate.
2983	* @param fFlags Flags, reserved for the future. Must be zero.
2984	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2985	* NULL if no ring-3 mapping.
2986	* @param ppvR3 Where to store the address of the Ring-0 mapping.
2987	* NULL if no ring-0 mapping.
2988	* @param paPages Where to store the addresses of the pages. Optional.
2989	*/
2990	SUPR0DECL(int) SUPR0PageAllocEx(PSUPDRVSESSION pSession, uint32_t cPages, uint32_t fFlags, PRTR3PTR ppvR3, PRTR0PTR ppvR0, PRTHCPHYS paPages)
2991	{
2992	int rc;
2993	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2994	LogFlow(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
2995
2996	/*
2997	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2998	*/
2999	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3000	AssertPtrNullReturn(ppvR3, VERR_INVALID_POINTER);
3001	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
3002	AssertReturn(ppvR3 \|\| ppvR0, VERR_INVALID_PARAMETER);
3003	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
3004	if (cPages < 1 \|\| cPages > VBOX_MAX_ALLOC_PAGE_COUNT)
3005	{
3006	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)));
3007	return VERR_PAGE_COUNT_OUT_OF_RANGE;
3008	}
3009
3010	/*
3011	* Let IPRT do the work.
3012	*/
3013	if (ppvR0)
3014	rc = RTR0MemObjAllocPage(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, true /* fExecutable */);
3015	else
3016	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
3017	if (RT_SUCCESS(rc))
3018	{
3019	int rc2;
3020	if (ppvR3)
3021	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
3022	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
3023	else
3024	Mem.MapObjR3 = NIL_RTR0MEMOBJ;
3025	if (RT_SUCCESS(rc))
3026	{
3027	Mem.eType = MEMREF_TYPE_PAGE;
3028	rc = supdrvMemAdd(&Mem, pSession);
3029	if (!rc)
3030	{
3031	if (ppvR3)
3032	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
3033	if (ppvR0)
3034	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
3035	if (paPages)
3036	{
3037	uint32_t iPage = cPages;
3038	while (iPage-- > 0)
3039	{
3040	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
3041	Assert(paPages[iPage] != NIL_RTHCPHYS);
3042	}
3043	}
3044	return VINF_SUCCESS;
3045	}
3046
3047	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
3048	AssertRC(rc2);
3049	}
3050
3051	rc2 = RTR0MemObjFree(Mem.MemObj, false);
3052	AssertRC(rc2);
3053	}
3054	return rc;
3055	}
3056
3057
3058	/**
3059	* Maps a chunk of memory previously allocated by SUPR0PageAllocEx into kernel
3060	* space.
3061	*
3062	* @returns IPRT status code.
3063	* @param pSession The session to associated the allocation with.
3064	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
3065	* @param offSub Where to start mapping. Must be page aligned.
3066	* @param cbSub How much to map. Must be page aligned.
3067	* @param fFlags Flags, MBZ.
3068	* @param ppvR0 Where to return the address of the ring-0 mapping on
3069	* success.
3070	*/
3071	SUPR0DECL(int) SUPR0PageMapKernel(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t offSub, uint32_t cbSub,
3072	uint32_t fFlags, PRTR0PTR ppvR0)
3073	{
3074	int rc;
3075	PSUPDRVBUNDLE pBundle;
3076	RTR0MEMOBJ hMemObj = NIL_RTR0MEMOBJ;
3077	LogFlow(("SUPR0PageMapKernel: pSession=%p pvR3=%p offSub=%#x cbSub=%#x\n", pSession, pvR3, offSub, cbSub));
3078
3079	/*
3080	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
3081	*/
3082	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3083	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
3084	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
3085	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3086	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3087	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
3088
3089	/*
3090	* Find the memory object.
3091	*/
3092	RTSpinlockAcquire(pSession->Spinlock);
3093	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3094	{
3095	if (pBundle->cUsed > 0)
3096	{
3097	unsigned i;
3098	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3099	{
3100	if ( ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
3101	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3102	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3103	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
3104	\|\| ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED
3105	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3106	&& pBundle->aMem[i].MapObjR3 == NIL_RTR0MEMOBJ
3107	&& RTR0MemObjAddressR3(pBundle->aMem[i].MemObj) == pvR3))
3108	{
3109	hMemObj = pBundle->aMem[i].MemObj;
3110	break;
3111	}
3112	}
3113	}
3114	}
3115	RTSpinlockRelease(pSession->Spinlock);
3116
3117	rc = VERR_INVALID_PARAMETER;
3118	if (hMemObj != NIL_RTR0MEMOBJ)
3119	{
3120	/*
3121	* Do some further input validations before calling IPRT.
3122	* (Cleanup is done indirectly by telling RTR0MemObjFree to include mappings.)
3123	*/
3124	size_t cbMemObj = RTR0MemObjSize(hMemObj);
3125	if ( offSub < cbMemObj
3126	&& cbSub <= cbMemObj
3127	&& offSub + cbSub <= cbMemObj)
3128	{
3129	RTR0MEMOBJ hMapObj;
3130	rc = RTR0MemObjMapKernelEx(&hMapObj, hMemObj, (void *)-1, 0,
3131	RTMEM_PROT_READ \| RTMEM_PROT_WRITE, offSub, cbSub);
3132	if (RT_SUCCESS(rc))
3133	*ppvR0 = RTR0MemObjAddress(hMapObj);
3134	}
3135	else
3136	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
3137
3138	}
3139	return rc;
3140	}
3141
3142
3143	/**
3144	* Changes the page level protection of one or more pages previously allocated
3145	* by SUPR0PageAllocEx.
3146	*
3147	* @returns IPRT status code.
3148	* @param pSession The session to associated the allocation with.
3149	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
3150	* NIL_RTR3PTR if the ring-3 mapping should be unaffected.
3151	* @param pvR0 The ring-0 address returned by SUPR0PageAllocEx.
3152	* NIL_RTR0PTR if the ring-0 mapping should be unaffected.
3153	* @param offSub Where to start changing. Must be page aligned.
3154	* @param cbSub How much to change. Must be page aligned.
3155	* @param fProt The new page level protection, see RTMEM_PROT_*.
3156	*/
3157	SUPR0DECL(int) SUPR0PageProtect(PSUPDRVSESSION pSession, RTR3PTR pvR3, RTR0PTR pvR0, uint32_t offSub, uint32_t cbSub, uint32_t fProt)
3158	{
3159	int rc;
3160	PSUPDRVBUNDLE pBundle;
3161	RTR0MEMOBJ hMemObjR0 = NIL_RTR0MEMOBJ;
3162	RTR0MEMOBJ hMemObjR3 = NIL_RTR0MEMOBJ;
3163	LogFlow(("SUPR0PageProtect: pSession=%p pvR3=%p pvR0=%p offSub=%#x cbSub=%#x fProt-%#x\n", pSession, pvR3, pvR0, offSub, cbSub, fProt));
3164
3165	/*
3166	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
3167	*/
3168	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3169	AssertReturn(!(fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)), VERR_INVALID_PARAMETER);
3170	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3171	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3172	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
3173
3174	/*
3175	* Find the memory object.
3176	*/
3177	RTSpinlockAcquire(pSession->Spinlock);
3178	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3179	{
3180	if (pBundle->cUsed > 0)
3181	{
3182	unsigned i;
3183	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3184	{
3185	if ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
3186	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3187	&& ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3188	\|\| pvR3 == NIL_RTR3PTR)
3189	&& ( pvR0 == NIL_RTR0PTR
3190	\|\| RTR0MemObjAddress(pBundle->aMem[i].MemObj) == pvR0)
3191	&& ( pvR3 == NIL_RTR3PTR
3192	\|\| RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3))
3193	{
3194	if (pvR0 != NIL_RTR0PTR)
3195	hMemObjR0 = pBundle->aMem[i].MemObj;
3196	if (pvR3 != NIL_RTR3PTR)
3197	hMemObjR3 = pBundle->aMem[i].MapObjR3;
3198	break;
3199	}
3200	}
3201	}
3202	}
3203	RTSpinlockRelease(pSession->Spinlock);
3204
3205	rc = VERR_INVALID_PARAMETER;
3206	if ( hMemObjR0 != NIL_RTR0MEMOBJ
3207	\|\| hMemObjR3 != NIL_RTR0MEMOBJ)
3208	{
3209	/*
3210	* Do some further input validations before calling IPRT.
3211	*/
3212	size_t cbMemObj = hMemObjR0 != NIL_RTR0PTR ? RTR0MemObjSize(hMemObjR0) : RTR0MemObjSize(hMemObjR3);
3213	if ( offSub < cbMemObj
3214	&& cbSub <= cbMemObj
3215	&& offSub + cbSub <= cbMemObj)
3216	{
3217	rc = VINF_SUCCESS;
3218	if (hMemObjR3 != NIL_RTR0PTR)
3219	rc = RTR0MemObjProtect(hMemObjR3, offSub, cbSub, fProt);
3220	if (hMemObjR0 != NIL_RTR0PTR && RT_SUCCESS(rc))
3221	rc = RTR0MemObjProtect(hMemObjR0, offSub, cbSub, fProt);
3222	}
3223	else
3224	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
3225
3226	}
3227	return rc;
3228
3229	}
3230
3231
3232	/**
3233	* Free memory allocated by SUPR0PageAlloc() and SUPR0PageAllocEx().
3234	*
3235	* @returns IPRT status code.
3236	* @param pSession The session owning the allocation.
3237	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc() or
3238	* SUPR0PageAllocEx().
3239	*/
3240	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
3241	{
3242	LogFlow(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
3243	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3244	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_PAGE);
3245	}
3246
3247
3248	/**
3249	* Gets the paging mode of the current CPU.
3250	*
3251	* @returns Paging mode, SUPPAGEINGMODE_INVALID on error.
3252	*/
3253	SUPR0DECL(SUPPAGINGMODE) SUPR0GetPagingMode(void)
3254	{
3255	SUPPAGINGMODE enmMode;
3256
3257	RTR0UINTREG cr0 = ASMGetCR0();
3258	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
3259	enmMode = SUPPAGINGMODE_INVALID;
3260	else
3261	{
3262	RTR0UINTREG cr4 = ASMGetCR4();
3263	uint32_t fNXEPlusLMA = 0;
3264	if (cr4 & X86_CR4_PAE)
3265	{
3266	uint32_t fExtFeatures = ASMCpuId_EDX(0x80000001);
3267	if (fExtFeatures & (X86_CPUID_EXT_FEATURE_EDX_NX \| X86_CPUID_EXT_FEATURE_EDX_LONG_MODE))
3268	{
3269	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
3270	if ((fExtFeatures & X86_CPUID_EXT_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
3271	fNXEPlusLMA \|= RT_BIT(0);
3272	if ((fExtFeatures & X86_CPUID_EXT_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
3273	fNXEPlusLMA \|= RT_BIT(1);
3274	}
3275	}
3276
3277	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
3278	{
3279	case 0:
3280	enmMode = SUPPAGINGMODE_32_BIT;
3281	break;
3282
3283	case X86_CR4_PGE:
3284	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
3285	break;
3286
3287	case X86_CR4_PAE:
3288	enmMode = SUPPAGINGMODE_PAE;
3289	break;
3290
3291	case X86_CR4_PAE \| RT_BIT(0):
3292	enmMode = SUPPAGINGMODE_PAE_NX;
3293	break;
3294
3295	case X86_CR4_PAE \| X86_CR4_PGE:
3296	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3297	break;
3298
3299	case X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3300	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3301	break;
3302
3303	case RT_BIT(1) \| X86_CR4_PAE:
3304	enmMode = SUPPAGINGMODE_AMD64;
3305	break;
3306
3307	case RT_BIT(1) \| X86_CR4_PAE \| RT_BIT(0):
3308	enmMode = SUPPAGINGMODE_AMD64_NX;
3309	break;
3310
3311	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
3312	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
3313	break;
3314
3315	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3316	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
3317	break;
3318
3319	default:
3320	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
3321	enmMode = SUPPAGINGMODE_INVALID;
3322	break;
3323	}
3324	}
3325	return enmMode;
3326	}
3327
3328
3329	/**
3330	* Enables or disabled hardware virtualization extensions using native OS APIs.
3331	*
3332	* @returns VBox status code.
3333	* @retval VINF_SUCCESS on success.
3334	* @retval VERR_NOT_SUPPORTED if not supported by the native OS.
3335	*
3336	* @param fEnable Whether to enable or disable.
3337	*/
3338	SUPR0DECL(int) SUPR0EnableVTx(bool fEnable)
3339	{
3340	#ifdef RT_OS_DARWIN
3341	return supdrvOSEnableVTx(fEnable);
3342	#else
3343	return VERR_NOT_SUPPORTED;
3344	#endif
3345	}
3346
3347
3348	/**
3349	* Suspends hardware virtualization extensions using the native OS API.
3350	*
3351	* This is called prior to entering raw-mode context.
3352	*
3353	* @returns @c true if suspended, @c false if not.
3354	*/
3355	SUPR0DECL(bool) SUPR0SuspendVTxOnCpu(void)
3356	{
3357	#ifdef RT_OS_DARWIN
3358	return supdrvOSSuspendVTxOnCpu();
3359	#else
3360	return false;
3361	#endif
3362	}
3363
3364
3365	/**
3366	* Resumes hardware virtualization extensions using the native OS API.
3367	*
3368	* This is called after to entering raw-mode context.
3369	*
3370	* @param fSuspended The return value of SUPR0SuspendVTxOnCpu.
3371	*/
3372	SUPR0DECL(void) SUPR0ResumeVTxOnCpu(bool fSuspended)
3373	{
3374	#ifdef RT_OS_DARWIN
3375	supdrvOSResumeVTxOnCpu(fSuspended);
3376	#else
3377	Assert(!fSuspended);
3378	#endif
3379	}
3380
3381
3382	/**
3383	* Queries the AMD-V and VT-x capabilities of the calling CPU.
3384	*
3385	* @returns VBox status code.
3386	* @retval VERR_VMX_NO_VMX
3387	* @retval VERR_VMX_MSR_SMX_VMXON_DISABLED
3388	* @retval VERR_VMX_MSR_VMXON_DISABLED
3389	* @retval VERR_VMX_MSR_LOCKING_FAILED
3390	* @retval VERR_SVM_NO_SVM
3391	* @retval VERR_SVM_DISABLED
3392	* @retval VERR_UNSUPPORTED_CPU if not identifiable as an AMD, Intel or VIA
3393	* (centaur) CPU.
3394	*
3395	* @param pSession The session handle.
3396	* @param pfCaps Where to store the capabilities.
3397	*/
3398	SUPR0DECL(int) SUPR0QueryVTCaps(PSUPDRVSESSION pSession, uint32_t *pfCaps)
3399	{
3400	int rc = VERR_UNSUPPORTED_CPU;
3401	RTTHREADPREEMPTSTATE PreemptState = RTTHREADPREEMPTSTATE_INITIALIZER;
3402
3403	/*
3404	* Input validation.
3405	*/
3406	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3407	AssertPtrReturn(pfCaps, VERR_INVALID_POINTER);
3408
3409	*pfCaps = 0;
3410	/* We may modify MSRs and re-read them, disable preemption so we make sure we don't migrate CPUs. */
3411	RTThreadPreemptDisable(&PreemptState);
3412	if (ASMHasCpuId())
3413	{
3414	uint32_t fFeaturesECX, fFeaturesEDX, uDummy;
3415	uint32_t uMaxId, uVendorEBX, uVendorECX, uVendorEDX;
3416
3417	ASMCpuId(0, &uMaxId, &uVendorEBX, &uVendorECX, &uVendorEDX);
3418	ASMCpuId(1, &uDummy, &uDummy, &fFeaturesECX, &fFeaturesEDX);
3419
3420	if ( ASMIsValidStdRange(uMaxId)
3421	&& ( ASMIsIntelCpuEx( uVendorEBX, uVendorECX, uVendorEDX)
3422	\|\| ASMIsViaCentaurCpuEx(uVendorEBX, uVendorECX, uVendorEDX) )
3423	)
3424	{
3425	if ( (fFeaturesECX & X86_CPUID_FEATURE_ECX_VMX)
3426	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3427	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3428	)
3429	{
3430	/** @todo Unify code with hmR0InitIntelCpu(). */
3431	uint64_t u64FeatMsr = ASMRdMsr(MSR_IA32_FEATURE_CONTROL);
3432	bool const fInSmxMode = RT_BOOL(ASMGetCR4() & X86_CR4_SMXE);
3433	bool fMsrLocked = RT_BOOL(u64FeatMsr & MSR_IA32_FEATURE_CONTROL_LOCK);
3434	bool fSmxVmxAllowed = RT_BOOL(u64FeatMsr & MSR_IA32_FEATURE_CONTROL_SMX_VMXON);
3435	bool fVmxAllowed = RT_BOOL(u64FeatMsr & MSR_IA32_FEATURE_CONTROL_VMXON);
3436
3437	/* Check if the LOCK bit is set but excludes the required VMXON bit. */
3438	if (fMsrLocked)
3439	{
3440	if (fInSmxMode && !fSmxVmxAllowed)
3441	rc = VERR_VMX_MSR_SMX_VMXON_DISABLED;
3442	else if (!fVmxAllowed)
3443	rc = VERR_VMX_MSR_VMXON_DISABLED;
3444	else
3445	rc = VINF_SUCCESS;
3446	}
3447	else
3448	{
3449	/*
3450	* MSR is not yet locked; we can change it ourselves here.
3451	* Once the lock bit is set, this MSR can no longer be modified.
3452	*/
3453	bool fAllowed;
3454	u64FeatMsr \|= MSR_IA32_FEATURE_CONTROL_LOCK;
3455	if (fInSmxMode)
3456	u64FeatMsr \|= MSR_IA32_FEATURE_CONTROL_SMX_VMXON;
3457	else
3458	u64FeatMsr \|= MSR_IA32_FEATURE_CONTROL_VMXON;
3459
3460	ASMWrMsr(MSR_IA32_FEATURE_CONTROL, u64FeatMsr);
3461
3462	/* Verify. */
3463	u64FeatMsr = ASMRdMsr(MSR_IA32_FEATURE_CONTROL);
3464	fMsrLocked = RT_BOOL(u64FeatMsr & MSR_IA32_FEATURE_CONTROL_LOCK);
3465	fSmxVmxAllowed = fMsrLocked && RT_BOOL(u64FeatMsr & MSR_IA32_FEATURE_CONTROL_SMX_VMXON);
3466	fVmxAllowed = fMsrLocked && RT_BOOL(u64FeatMsr & MSR_IA32_FEATURE_CONTROL_VMXON);
3467	fAllowed = fInSmxMode ? fSmxVmxAllowed : fVmxAllowed;
3468	if (fAllowed)
3469	rc = VINF_SUCCESS;
3470	else
3471	rc = VERR_VMX_MSR_LOCKING_FAILED;
3472	}
3473
3474	if (rc == VINF_SUCCESS)
3475	{
3476	VMX_CAPABILITY vtCaps;
3477
3478	*pfCaps \|= SUPVTCAPS_VT_X;
3479
3480	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS);
3481	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC_USE_SECONDARY_EXEC_CTRL)
3482	{
3483	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS2);
3484	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC2_EPT)
3485	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3486	}
3487	}
3488	}
3489	else
3490	rc = VERR_VMX_NO_VMX;
3491	}
3492	else if ( ASMIsAmdCpuEx(uVendorEBX, uVendorECX, uVendorEDX)
3493	&& ASMIsValidStdRange(uMaxId))
3494	{
3495	uint32_t fExtFeaturesEcx, uExtMaxId;
3496	ASMCpuId(0x80000000, &uExtMaxId, &uDummy, &uDummy, &uDummy);
3497	ASMCpuId(0x80000001, &uDummy, &uDummy, &fExtFeaturesEcx, &uDummy);
3498	if ( ASMIsValidExtRange(uExtMaxId)
3499	&& uExtMaxId >= 0x8000000a
3500	&& (fExtFeaturesEcx & X86_CPUID_AMD_FEATURE_ECX_SVM)
3501	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3502	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3503	)
3504	{
3505	/* Check if SVM is disabled */
3506	uint64_t u64FeatMsr = ASMRdMsr(MSR_K8_VM_CR);
3507	if (!(u64FeatMsr & MSR_K8_VM_CR_SVM_DISABLE))
3508	{
3509	uint32_t fSvmFeatures;
3510	*pfCaps \|= SUPVTCAPS_AMD_V;
3511
3512	/* Query AMD-V features. */
3513	ASMCpuId(0x8000000a, &uDummy, &uDummy, &uDummy, &fSvmFeatures);
3514	if (fSvmFeatures & AMD_CPUID_SVM_FEATURE_EDX_NESTED_PAGING)
3515	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3516
3517	rc = VINF_SUCCESS;
3518	}
3519	else
3520	rc = VERR_SVM_DISABLED;
3521	}
3522	else
3523	rc = VERR_SVM_NO_SVM;
3524	}
3525	}
3526
3527	RTThreadPreemptRestore(&PreemptState);
3528	return rc;
3529	}
3530
3531
3532	/**
3533	* (Re-)initializes the per-cpu structure prior to starting or resuming the GIP
3534	* updating.
3535	*
3536	* @param pGipCpu The per CPU structure for this CPU.
3537	* @param u64NanoTS The current time.
3538	*/
3539	static void supdrvGipReInitCpu(PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
3540	{
3541	pGipCpu->u64TSC = ASMReadTSC() - pGipCpu->u32UpdateIntervalTSC;
3542	pGipCpu->u64NanoTS = u64NanoTS;
3543	}
3544
3545
3546	/**
3547	* Set the current TSC and NanoTS value for the CPU.
3548	*
3549	* @param idCpu The CPU ID. Unused - we have to use the APIC ID.
3550	* @param pvUser1 Pointer to the ring-0 GIP mapping.
3551	* @param pvUser2 Pointer to the variable holding the current time.
3552	*/
3553	static DECLCALLBACK(void) supdrvGipReInitCpuCallback(RTCPUID idCpu, void pvUser1, void pvUser2)
3554	{
3555	PSUPGLOBALINFOPAGE pGip = (PSUPGLOBALINFOPAGE)pvUser1;
3556	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
3557
3558	if (RT_LIKELY(iCpu < pGip->cCpus && pGip->aCPUs[iCpu].idCpu == idCpu))
3559	supdrvGipReInitCpu(&pGip->aCPUs[iCpu], (uint64_t )pvUser2);
3560
3561	NOREF(pvUser2);
3562	NOREF(idCpu);
3563	}
3564
3565
3566	/**
3567	* Maps the GIP into userspace and/or get the physical address of the GIP.
3568	*
3569	* @returns IPRT status code.
3570	* @param pSession Session to which the GIP mapping should belong.
3571	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
3572	* @param pHCPhysGip Where to store the physical address. (optional)
3573	*
3574	* @remark There is no reference counting on the mapping, so one call to this function
3575	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
3576	* and remove the session as a GIP user.
3577	*/
3578	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
3579	{
3580	int rc;
3581	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3582	RTR3PTR pGipR3 = NIL_RTR3PTR;
3583	RTHCPHYS HCPhys = NIL_RTHCPHYS;
3584	LogFlow(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
3585
3586	/*
3587	* Validate
3588	*/
3589	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3590	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
3591	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
3592
3593	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3594	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3595	#else
3596	RTSemFastMutexRequest(pDevExt->mtxGip);
3597	#endif
3598	if (pDevExt->pGip)
3599	{
3600	/*
3601	* Map it?
3602	*/
3603	rc = VINF_SUCCESS;
3604	if (ppGipR3)
3605	{
3606	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
3607	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
3608	RTMEM_PROT_READ, RTR0ProcHandleSelf());
3609	if (RT_SUCCESS(rc))
3610	pGipR3 = RTR0MemObjAddressR3(pSession->GipMapObjR3);
3611	}
3612
3613	/*
3614	* Get physical address.
3615	*/
3616	if (pHCPhysGip && RT_SUCCESS(rc))
3617	HCPhys = pDevExt->HCPhysGip;
3618
3619	/*
3620	* Reference globally.
3621	*/
3622	if (!pSession->fGipReferenced && RT_SUCCESS(rc))
3623	{
3624	pSession->fGipReferenced = 1;
3625	pDevExt->cGipUsers++;
3626	if (pDevExt->cGipUsers == 1)
3627	{
3628	PSUPGLOBALINFOPAGE pGipR0 = pDevExt->pGip;
3629	uint64_t u64NanoTS;
3630	uint32_t u32SystemResolution;
3631	unsigned i;
3632
3633	LogFlow(("SUPR0GipMap: Resumes GIP updating\n"));
3634
3635	/*
3636	* Try bump up the system timer resolution.
3637	* The more interrupts the better...
3638	*/
3639	if ( RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
3640	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
3641	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1953125 /* 512 HZ */, &u32SystemResolution))
3642	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 2000000 /* 500 HZ */, &u32SystemResolution))
3643	)
3644	{
3645	Assert(RTTimerGetSystemGranularity() <= u32SystemResolution);
3646	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
3647	}
3648
3649	if (pGipR0->aCPUs[0].u32TransactionId != 2 /* not the first time */)
3650	{
3651	for (i = 0; i < RT_ELEMENTS(pGipR0->aCPUs); i++)
3652	ASMAtomicUoWriteU32(&pGipR0->aCPUs[i].u32TransactionId,
3653	(pGipR0->aCPUs[i].u32TransactionId + GIP_UPDATEHZ_RECALC_FREQ * 2)
3654	& ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
3655	ASMAtomicWriteU64(&pGipR0->u64NanoTSLastUpdateHz, 0);
3656	}
3657
3658	u64NanoTS = RTTimeSystemNanoTS() - pGipR0->u32UpdateIntervalNS;
3659	if ( pGipR0->u32Mode == SUPGIPMODE_SYNC_TSC
3660	\|\| RTMpGetOnlineCount() == 1)
3661	supdrvGipReInitCpu(&pGipR0->aCPUs[0], u64NanoTS);
3662	else
3663	RTMpOnAll(supdrvGipReInitCpuCallback, pGipR0, &u64NanoTS);
3664
3665	#ifndef DO_NOT_START_GIP
3666	rc = RTTimerStart(pDevExt->pGipTimer, 0); AssertRC(rc);
3667	#endif
3668	rc = VINF_SUCCESS;
3669	}
3670	}
3671	}
3672	else
3673	{
3674	rc = VERR_GENERAL_FAILURE;
3675	Log(("SUPR0GipMap: GIP is not available!\n"));
3676	}
3677	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3678	RTSemMutexRelease(pDevExt->mtxGip);
3679	#else
3680	RTSemFastMutexRelease(pDevExt->mtxGip);
3681	#endif
3682
3683	/*
3684	* Write returns.
3685	*/
3686	if (pHCPhysGip)
3687	*pHCPhysGip = HCPhys;
3688	if (ppGipR3)
3689	*ppGipR3 = pGipR3;
3690
3691	#ifdef DEBUG_DARWIN_GIP
3692	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3693	#else
3694	LogFlow(( "SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3695	#endif
3696	return rc;
3697	}
3698
3699
3700	/**
3701	* Unmaps any user mapping of the GIP and terminates all GIP access
3702	* from this session.
3703	*
3704	* @returns IPRT status code.
3705	* @param pSession Session to which the GIP mapping should belong.
3706	*/
3707	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
3708	{
3709	int rc = VINF_SUCCESS;
3710	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3711	#ifdef DEBUG_DARWIN_GIP
3712	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
3713	pSession,
3714	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
3715	pSession->GipMapObjR3));
3716	#else
3717	LogFlow(("SUPR0GipUnmap: pSession=%p\n", pSession));
3718	#endif
3719	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3720
3721	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3722	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3723	#else
3724	RTSemFastMutexRequest(pDevExt->mtxGip);
3725	#endif
3726
3727	/*
3728	* Unmap anything?
3729	*/
3730	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
3731	{
3732	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
3733	AssertRC(rc);
3734	if (RT_SUCCESS(rc))
3735	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
3736	}
3737
3738	/*
3739	* Dereference global GIP.
3740	*/
3741	if (pSession->fGipReferenced && !rc)
3742	{
3743	pSession->fGipReferenced = 0;
3744	if ( pDevExt->cGipUsers > 0
3745	&& !--pDevExt->cGipUsers)
3746	{
3747	LogFlow(("SUPR0GipUnmap: Suspends GIP updating\n"));
3748	#ifndef DO_NOT_START_GIP
3749	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = VINF_SUCCESS;
3750	#endif
3751
3752	if (pDevExt->u32SystemTimerGranularityGrant)
3753	{
3754	int rc2 = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant);
3755	AssertRC(rc2);
3756	pDevExt->u32SystemTimerGranularityGrant = 0;
3757	}
3758	}
3759	}
3760
3761	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3762	RTSemMutexRelease(pDevExt->mtxGip);
3763	#else
3764	RTSemFastMutexRelease(pDevExt->mtxGip);
3765	#endif
3766
3767	return rc;
3768	}
3769
3770
3771	/**
3772	* Gets the GIP pointer.
3773	*
3774	* @returns Pointer to the GIP or NULL.
3775	*/
3776	SUPDECL(PSUPGLOBALINFOPAGE) SUPGetGIP(void)
3777	{
3778	return g_pSUPGlobalInfoPage;
3779	}
3780
3781
3782	/**
3783	* Register a component factory with the support driver.
3784	*
3785	* This is currently restricted to kernel sessions only.
3786	*
3787	* @returns VBox status code.
3788	* @retval VINF_SUCCESS on success.
3789	* @retval VERR_NO_MEMORY if we're out of memory.
3790	* @retval VERR_ALREADY_EXISTS if the factory has already been registered.
3791	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3792	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3793	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3794	*
3795	* @param pSession The SUPDRV session (must be a ring-0 session).
3796	* @param pFactory Pointer to the component factory registration structure.
3797	*
3798	* @remarks This interface is also available via SUPR0IdcComponentRegisterFactory.
3799	*/
3800	SUPR0DECL(int) SUPR0ComponentRegisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3801	{
3802	PSUPDRVFACTORYREG pNewReg;
3803	const char *psz;
3804	int rc;
3805
3806	/*
3807	* Validate parameters.
3808	*/
3809	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3810	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3811	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3812	AssertPtrReturn(pFactory->pfnQueryFactoryInterface, VERR_INVALID_POINTER);
3813	psz = RTStrEnd(pFactory->szName, sizeof(pFactory->szName));
3814	AssertReturn(psz, VERR_INVALID_PARAMETER);
3815
3816	/*
3817	* Allocate and initialize a new registration structure.
3818	*/
3819	pNewReg = (PSUPDRVFACTORYREG)RTMemAlloc(sizeof(SUPDRVFACTORYREG));
3820	if (pNewReg)
3821	{
3822	pNewReg->pNext = NULL;
3823	pNewReg->pFactory = pFactory;
3824	pNewReg->pSession = pSession;
3825	pNewReg->cchName = psz - &pFactory->szName[0];
3826
3827	/*
3828	* Add it to the tail of the list after checking for prior registration.
3829	*/
3830	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3831	if (RT_SUCCESS(rc))
3832	{
3833	PSUPDRVFACTORYREG pPrev = NULL;
3834	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3835	while (pCur && pCur->pFactory != pFactory)
3836	{
3837	pPrev = pCur;
3838	pCur = pCur->pNext;
3839	}
3840	if (!pCur)
3841	{
3842	if (pPrev)
3843	pPrev->pNext = pNewReg;
3844	else
3845	pSession->pDevExt->pComponentFactoryHead = pNewReg;
3846	rc = VINF_SUCCESS;
3847	}
3848	else
3849	rc = VERR_ALREADY_EXISTS;
3850
3851	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3852	}
3853
3854	if (RT_FAILURE(rc))
3855	RTMemFree(pNewReg);
3856	}
3857	else
3858	rc = VERR_NO_MEMORY;
3859	return rc;
3860	}
3861
3862
3863	/**
3864	* Deregister a component factory.
3865	*
3866	* @returns VBox status code.
3867	* @retval VINF_SUCCESS on success.
3868	* @retval VERR_NOT_FOUND if the factory wasn't registered.
3869	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3870	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3871	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3872	*
3873	* @param pSession The SUPDRV session (must be a ring-0 session).
3874	* @param pFactory Pointer to the component factory registration structure
3875	* previously passed SUPR0ComponentRegisterFactory().
3876	*
3877	* @remarks This interface is also available via SUPR0IdcComponentDeregisterFactory.
3878	*/
3879	SUPR0DECL(int) SUPR0ComponentDeregisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3880	{
3881	int rc;
3882
3883	/*
3884	* Validate parameters.
3885	*/
3886	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3887	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3888	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3889
3890	/*
3891	* Take the lock and look for the registration record.
3892	*/
3893	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3894	if (RT_SUCCESS(rc))
3895	{
3896	PSUPDRVFACTORYREG pPrev = NULL;
3897	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3898	while (pCur && pCur->pFactory != pFactory)
3899	{
3900	pPrev = pCur;
3901	pCur = pCur->pNext;
3902	}
3903	if (pCur)
3904	{
3905	if (!pPrev)
3906	pSession->pDevExt->pComponentFactoryHead = pCur->pNext;
3907	else
3908	pPrev->pNext = pCur->pNext;
3909
3910	pCur->pNext = NULL;
3911	pCur->pFactory = NULL;
3912	pCur->pSession = NULL;
3913	rc = VINF_SUCCESS;
3914	}
3915	else
3916	rc = VERR_NOT_FOUND;
3917
3918	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3919
3920	RTMemFree(pCur);
3921	}
3922	return rc;
3923	}
3924
3925
3926	/**
3927	* Queries a component factory.
3928	*
3929	* @returns VBox status code.
3930	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3931	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3932	* @retval VERR_SUPDRV_COMPONENT_NOT_FOUND if the component factory wasn't found.
3933	* @retval VERR_SUPDRV_INTERFACE_NOT_SUPPORTED if the interface wasn't supported.
3934	*
3935	* @param pSession The SUPDRV session.
3936	* @param pszName The name of the component factory.
3937	* @param pszInterfaceUuid The UUID of the factory interface (stringified).
3938	* @param ppvFactoryIf Where to store the factory interface.
3939	*/
3940	SUPR0DECL(int) SUPR0ComponentQueryFactory(PSUPDRVSESSION pSession, const char pszName, const char pszInterfaceUuid, void **ppvFactoryIf)
3941	{
3942	const char *pszEnd;
3943	size_t cchName;
3944	int rc;
3945
3946	/*
3947	* Validate parameters.
3948	*/
3949	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3950
3951	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
3952	pszEnd = RTStrEnd(pszName, RT_SIZEOFMEMB(SUPDRVFACTORY, szName));
3953	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3954	cchName = pszEnd - pszName;
3955
3956	AssertPtrReturn(pszInterfaceUuid, VERR_INVALID_POINTER);
3957	pszEnd = RTStrEnd(pszInterfaceUuid, RTUUID_STR_LENGTH);
3958	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3959
3960	AssertPtrReturn(ppvFactoryIf, VERR_INVALID_POINTER);
3961	*ppvFactoryIf = NULL;
3962
3963	/*
3964	* Take the lock and try all factories by this name.
3965	*/
3966	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3967	if (RT_SUCCESS(rc))
3968	{
3969	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3970	rc = VERR_SUPDRV_COMPONENT_NOT_FOUND;
3971	while (pCur)
3972	{
3973	if ( pCur->cchName == cchName
3974	&& !memcmp(pCur->pFactory->szName, pszName, cchName))
3975	{
3976	void *pvFactory = pCur->pFactory->pfnQueryFactoryInterface(pCur->pFactory, pSession, pszInterfaceUuid);
3977	if (pvFactory)
3978	{
3979	*ppvFactoryIf = pvFactory;
3980	rc = VINF_SUCCESS;
3981	break;
3982	}
3983	rc = VERR_SUPDRV_INTERFACE_NOT_SUPPORTED;
3984	}
3985
3986	/* next */
3987	pCur = pCur->pNext;
3988	}
3989
3990	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3991	}
3992	return rc;
3993	}
3994
3995
3996	/**
3997	* Adds a memory object to the session.
3998	*
3999	* @returns IPRT status code.
4000	* @param pMem Memory tracking structure containing the
4001	* information to track.
4002	* @param pSession The session.
4003	*/
4004	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
4005	{
4006	PSUPDRVBUNDLE pBundle;
4007
4008	/*
4009	* Find free entry and record the allocation.
4010	*/
4011	RTSpinlockAcquire(pSession->Spinlock);
4012	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
4013	{
4014	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
4015	{
4016	unsigned i;
4017	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
4018	{
4019	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
4020	{
4021	pBundle->cUsed++;
4022	pBundle->aMem[i] = *pMem;
4023	RTSpinlockRelease(pSession->Spinlock);
4024	return VINF_SUCCESS;
4025	}
4026	}
4027	AssertFailed(); /* !!this can't be happening!!! */
4028	}
4029	}
4030	RTSpinlockRelease(pSession->Spinlock);
4031
4032	/*
4033	* Need to allocate a new bundle.
4034	* Insert into the last entry in the bundle.
4035	*/
4036	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
4037	if (!pBundle)
4038	return VERR_NO_MEMORY;
4039
4040	/* take last entry. */
4041	pBundle->cUsed++;
4042	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
4043
4044	/* insert into list. */
4045	RTSpinlockAcquire(pSession->Spinlock);
4046	pBundle->pNext = pSession->Bundle.pNext;
4047	pSession->Bundle.pNext = pBundle;
4048	RTSpinlockRelease(pSession->Spinlock);
4049
4050	return VINF_SUCCESS;
4051	}
4052
4053
4054	/**
4055	* Releases a memory object referenced by pointer and type.
4056	*
4057	* @returns IPRT status code.
4058	* @param pSession Session data.
4059	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
4060	* @param eType Memory type.
4061	*/
4062	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
4063	{
4064	PSUPDRVBUNDLE pBundle;
4065
4066	/*
4067	* Validate input.
4068	*/
4069	if (!uPtr)
4070	{
4071	Log(("Illegal address %p\n", (void *)uPtr));
4072	return VERR_INVALID_PARAMETER;
4073	}
4074
4075	/*
4076	* Search for the address.
4077	*/
4078	RTSpinlockAcquire(pSession->Spinlock);
4079	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
4080	{
4081	if (pBundle->cUsed > 0)
4082	{
4083	unsigned i;
4084	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
4085	{
4086	if ( pBundle->aMem[i].eType == eType
4087	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
4088	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
4089	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
4090	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
4091	)
4092	{
4093	/* Make a copy of it and release it outside the spinlock. */
4094	SUPDRVMEMREF Mem = pBundle->aMem[i];
4095	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
4096	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
4097	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
4098	RTSpinlockRelease(pSession->Spinlock);
4099
4100	if (Mem.MapObjR3 != NIL_RTR0MEMOBJ)
4101	{
4102	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
4103	AssertRC(rc); /** @todo figure out how to handle this. */
4104	}
4105	if (Mem.MemObj != NIL_RTR0MEMOBJ)
4106	{
4107	int rc = RTR0MemObjFree(Mem.MemObj, true /* fFreeMappings */);
4108	AssertRC(rc); /** @todo figure out how to handle this. */
4109	}
4110	return VINF_SUCCESS;
4111	}
4112	}
4113	}
4114	}
4115	RTSpinlockRelease(pSession->Spinlock);
4116	Log(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
4117	return VERR_INVALID_PARAMETER;
4118	}
4119
4120
4121	/**
4122	* Opens an image. If it's the first time it's opened the call must upload
4123	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
4124	*
4125	* This is the 1st step of the loading.
4126	*
4127	* @returns IPRT status code.
4128	* @param pDevExt Device globals.
4129	* @param pSession Session data.
4130	* @param pReq The open request.
4131	*/
4132	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
4133	{
4134	int rc;
4135	PSUPDRVLDRIMAGE pImage;
4136	void *pv;
4137	size_t cchName = strlen(pReq->u.In.szName); /* (caller checked < 32). */
4138	LogFlow(("supdrvIOCtl_LdrOpen: szName=%s cbImageWithTabs=%d\n", pReq->u.In.szName, pReq->u.In.cbImageWithTabs));
4139
4140	/*
4141	* Check if we got an instance of the image already.
4142	*/
4143	supdrvLdrLock(pDevExt);
4144	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
4145	{
4146	if ( pImage->szName[cchName] == '\0'
4147	&& !memcmp(pImage->szName, pReq->u.In.szName, cchName))
4148	{
4149	/** @todo check cbImageBits and cbImageWithTabs here, if they differs that indicates that the images are different. */
4150	pImage->cUsage++;
4151	pReq->u.Out.pvImageBase = pImage->pvImage;
4152	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
4153	pReq->u.Out.fNativeLoader = pImage->fNative;
4154	supdrvLdrAddUsage(pSession, pImage);
4155	supdrvLdrUnlock(pDevExt);
4156	return VINF_SUCCESS;
4157	}
4158	}
4159	/* (not found - add it!) */
4160
4161	/*
4162	* Allocate memory.
4163	*/
4164	Assert(cchName < sizeof(pImage->szName));
4165	pv = RTMemAlloc(sizeof(SUPDRVLDRIMAGE));
4166	if (!pv)
4167	{
4168	supdrvLdrUnlock(pDevExt);
4169	Log(("supdrvIOCtl_LdrOpen: RTMemAlloc() failed\n"));
4170	return /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_2;
4171	}
4172
4173	/*
4174	* Setup and link in the LDR stuff.
4175	*/
4176	pImage = (PSUPDRVLDRIMAGE)pv;
4177	pImage->pvImage = NULL;
4178	pImage->pvImageAlloc = NULL;
4179	pImage->cbImageWithTabs = pReq->u.In.cbImageWithTabs;
4180	pImage->cbImageBits = pReq->u.In.cbImageBits;
4181	pImage->cSymbols = 0;
4182	pImage->paSymbols = NULL;
4183	pImage->pachStrTab = NULL;
4184	pImage->cbStrTab = 0;
4185	pImage->pfnModuleInit = NULL;
4186	pImage->pfnModuleTerm = NULL;
4187	pImage->pfnServiceReqHandler = NULL;
4188	pImage->uState = SUP_IOCTL_LDR_OPEN;
4189	pImage->cUsage = 1;
4190	pImage->pDevExt = pDevExt;
4191	memcpy(pImage->szName, pReq->u.In.szName, cchName + 1);
4192
4193	/*
4194	* Try load it using the native loader, if that isn't supported, fall back
4195	* on the older method.
4196	*/
4197	pImage->fNative = true;
4198	rc = supdrvOSLdrOpen(pDevExt, pImage, pReq->u.In.szFilename);
4199	if (rc == VERR_NOT_SUPPORTED)
4200	{
4201	pImage->pvImageAlloc = RTMemExecAlloc(pImage->cbImageBits + 31);
4202	pImage->pvImage = RT_ALIGN_P(pImage->pvImageAlloc, 32);
4203	pImage->fNative = false;
4204	rc = pImage->pvImageAlloc ? VINF_SUCCESS : VERR_NO_EXEC_MEMORY;
4205	}
4206	if (RT_FAILURE(rc))
4207	{
4208	supdrvLdrUnlock(pDevExt);
4209	RTMemFree(pImage);
4210	Log(("supdrvIOCtl_LdrOpen(%s): failed - %Rrc\n", pReq->u.In.szName, rc));
4211	return rc;
4212	}
4213	Assert(VALID_PTR(pImage->pvImage) \|\| RT_FAILURE(rc));
4214
4215	/*
4216	* Link it.
4217	*/
4218	pImage->pNext = pDevExt->pLdrImages;
4219	pDevExt->pLdrImages = pImage;
4220
4221	supdrvLdrAddUsage(pSession, pImage);
4222
4223	pReq->u.Out.pvImageBase = pImage->pvImage;
4224	pReq->u.Out.fNeedsLoading = true;
4225	pReq->u.Out.fNativeLoader = pImage->fNative;
4226	supdrvOSLdrNotifyOpened(pDevExt, pImage);
4227
4228	supdrvLdrUnlock(pDevExt);
4229	return VINF_SUCCESS;
4230	}
4231
4232
4233	/**
4234	* Worker that validates a pointer to an image entrypoint.
4235	*
4236	* @returns IPRT status code.
4237	* @param pDevExt The device globals.
4238	* @param pImage The loader image.
4239	* @param pv The pointer into the image.
4240	* @param fMayBeNull Whether it may be NULL.
4241	* @param pszWhat What is this entrypoint? (for logging)
4242	* @param pbImageBits The image bits prepared by ring-3.
4243	*
4244	* @remarks Will leave the lock on failure.
4245	*/
4246	static int supdrvLdrValidatePointer(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage, void *pv,
4247	bool fMayBeNull, const uint8_t pbImageBits, const char pszWhat)
4248	{
4249	if (!fMayBeNull \|\| pv)
4250	{
4251	if ((uintptr_t)pv - (uintptr_t)pImage->pvImage >= pImage->cbImageBits)
4252	{
4253	supdrvLdrUnlock(pDevExt);
4254	Log(("Out of range (%p LB %#x): %s=%p\n", pImage->pvImage, pImage->cbImageBits, pszWhat, pv));
4255	return VERR_INVALID_PARAMETER;
4256	}
4257
4258	if (pImage->fNative)
4259	{
4260	int rc = supdrvOSLdrValidatePointer(pDevExt, pImage, pv, pbImageBits);
4261	if (RT_FAILURE(rc))
4262	{
4263	supdrvLdrUnlock(pDevExt);
4264	Log(("Bad entry point address: %s=%p (rc=%Rrc)\n", pszWhat, pv, rc));
4265	return rc;
4266	}
4267	}
4268	}
4269	return VINF_SUCCESS;
4270	}
4271
4272
4273	/**
4274	* Loads the image bits.
4275	*
4276	* This is the 2nd step of the loading.
4277	*
4278	* @returns IPRT status code.
4279	* @param pDevExt Device globals.
4280	* @param pSession Session data.
4281	* @param pReq The request.
4282	*/
4283	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
4284	{
4285	PSUPDRVLDRUSAGE pUsage;
4286	PSUPDRVLDRIMAGE pImage;
4287	int rc;
4288	LogFlow(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImageWithBits=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImageWithTabs));
4289
4290	/*
4291	* Find the ldr image.
4292	*/
4293	supdrvLdrLock(pDevExt);
4294	pUsage = pSession->pLdrUsage;
4295	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4296	pUsage = pUsage->pNext;
4297	if (!pUsage)
4298	{
4299	supdrvLdrUnlock(pDevExt);
4300	Log(("SUP_IOCTL_LDR_LOAD: couldn't find image!\n"));
4301	return VERR_INVALID_HANDLE;
4302	}
4303	pImage = pUsage->pImage;
4304
4305	/*
4306	* Validate input.
4307	*/
4308	if ( pImage->cbImageWithTabs != pReq->u.In.cbImageWithTabs
4309	\|\| pImage->cbImageBits != pReq->u.In.cbImageBits)
4310	{
4311	supdrvLdrUnlock(pDevExt);
4312	Log(("SUP_IOCTL_LDR_LOAD: image size mismatch!! %d(prep) != %d(load) or %d != %d\n",
4313	pImage->cbImageWithTabs, pReq->u.In.cbImageWithTabs, pImage->cbImageBits, pReq->u.In.cbImageBits));
4314	return VERR_INVALID_HANDLE;
4315	}
4316
4317	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
4318	{
4319	unsigned uState = pImage->uState;
4320	supdrvLdrUnlock(pDevExt);
4321	if (uState != SUP_IOCTL_LDR_LOAD)
4322	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
4323	return VERR_ALREADY_LOADED;
4324	}
4325
4326	switch (pReq->u.In.eEPType)
4327	{
4328	case SUPLDRLOADEP_NOTHING:
4329	break;
4330
4331	case SUPLDRLOADEP_VMMR0:
4332	rc = supdrvLdrValidatePointer( pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0, false, pReq->u.In.abImage, "pvVMMR0");
4333	if (RT_SUCCESS(rc))
4334	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt, false, pReq->u.In.abImage, "pvVMMR0EntryInt");
4335	if (RT_SUCCESS(rc))
4336	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, false, pReq->u.In.abImage, "pvVMMR0EntryFast");
4337	if (RT_SUCCESS(rc))
4338	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx, false, pReq->u.In.abImage, "pvVMMR0EntryEx");
4339	if (RT_FAILURE(rc))
4340	return rc;
4341	break;
4342
4343	case SUPLDRLOADEP_SERVICE:
4344	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.Service.pfnServiceReq, false, pReq->u.In.abImage, "pfnServiceReq");
4345	if (RT_FAILURE(rc))
4346	return rc;
4347	if ( pReq->u.In.EP.Service.apvReserved[0] != NIL_RTR0PTR
4348	\|\| pReq->u.In.EP.Service.apvReserved[1] != NIL_RTR0PTR
4349	\|\| pReq->u.In.EP.Service.apvReserved[2] != NIL_RTR0PTR)
4350	{
4351	supdrvLdrUnlock(pDevExt);
4352	Log(("Out of range (%p LB %#x): apvReserved={%p,%p,%p} MBZ!\n",
4353	pImage->pvImage, pReq->u.In.cbImageWithTabs,
4354	pReq->u.In.EP.Service.apvReserved[0],
4355	pReq->u.In.EP.Service.apvReserved[1],
4356	pReq->u.In.EP.Service.apvReserved[2]));
4357	return VERR_INVALID_PARAMETER;
4358	}
4359	break;
4360
4361	default:
4362	supdrvLdrUnlock(pDevExt);
4363	Log(("Invalid eEPType=%d\n", pReq->u.In.eEPType));
4364	return VERR_INVALID_PARAMETER;
4365	}
4366
4367	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleInit, true, pReq->u.In.abImage, "pfnModuleInit");
4368	if (RT_FAILURE(rc))
4369	return rc;
4370	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleTerm, true, pReq->u.In.abImage, "pfnModuleTerm");
4371	if (RT_FAILURE(rc))
4372	return rc;
4373
4374	/*
4375	* Allocate and copy the tables.
4376	* (No need to do try/except as this is a buffered request.)
4377	*/
4378	pImage->cbStrTab = pReq->u.In.cbStrTab;
4379	if (pImage->cbStrTab)
4380	{
4381	pImage->pachStrTab = (char *)RTMemAlloc(pImage->cbStrTab);
4382	if (pImage->pachStrTab)
4383	memcpy(pImage->pachStrTab, &pReq->u.In.abImage[pReq->u.In.offStrTab], pImage->cbStrTab);
4384	else
4385	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_3;
4386	}
4387
4388	pImage->cSymbols = pReq->u.In.cSymbols;
4389	if (RT_SUCCESS(rc) && pImage->cSymbols)
4390	{
4391	size_t cbSymbols = pImage->cSymbols * sizeof(SUPLDRSYM);
4392	pImage->paSymbols = (PSUPLDRSYM)RTMemAlloc(cbSymbols);
4393	if (pImage->paSymbols)
4394	memcpy(pImage->paSymbols, &pReq->u.In.abImage[pReq->u.In.offSymbols], cbSymbols);
4395	else
4396	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_4;
4397	}
4398
4399	/*
4400	* Copy the bits / complete native loading.
4401	*/
4402	if (RT_SUCCESS(rc))
4403	{
4404	pImage->uState = SUP_IOCTL_LDR_LOAD;
4405	pImage->pfnModuleInit = pReq->u.In.pfnModuleInit;
4406	pImage->pfnModuleTerm = pReq->u.In.pfnModuleTerm;
4407
4408	if (pImage->fNative)
4409	rc = supdrvOSLdrLoad(pDevExt, pImage, pReq->u.In.abImage, pReq);
4410	else
4411	{
4412	memcpy(pImage->pvImage, &pReq->u.In.abImage[0], pImage->cbImageBits);
4413	Log(("vboxdrv: Loaded '%s' at %p\n", pImage->szName, pImage->pvImage));
4414	}
4415	}
4416
4417	/*
4418	* Update any entry points.
4419	*/
4420	if (RT_SUCCESS(rc))
4421	{
4422	switch (pReq->u.In.eEPType)
4423	{
4424	default:
4425	case SUPLDRLOADEP_NOTHING:
4426	rc = VINF_SUCCESS;
4427	break;
4428	case SUPLDRLOADEP_VMMR0:
4429	rc = supdrvLdrSetVMMR0EPs(pDevExt, pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
4430	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
4431	break;
4432	case SUPLDRLOADEP_SERVICE:
4433	pImage->pfnServiceReqHandler = pReq->u.In.EP.Service.pfnServiceReq;
4434	rc = VINF_SUCCESS;
4435	break;
4436	}
4437	}
4438
4439	/*
4440	* On success call the module initialization.
4441	*/
4442	LogFlow(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
4443	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
4444	{
4445	Log(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
4446	pDevExt->pLdrInitImage = pImage;
4447	pDevExt->hLdrInitThread = RTThreadNativeSelf();
4448	rc = pImage->pfnModuleInit(pImage);
4449	pDevExt->pLdrInitImage = NULL;
4450	pDevExt->hLdrInitThread = NIL_RTNATIVETHREAD;
4451	if (RT_FAILURE(rc) && pDevExt->pvVMMR0 == pImage->pvImage)
4452	supdrvLdrUnsetVMMR0EPs(pDevExt);
4453	}
4454
4455	if (RT_FAILURE(rc))
4456	{
4457	/* Inform the tracing component in case ModuleInit registered TPs. */
4458	supdrvTracerModuleUnloading(pDevExt, pImage);
4459
4460	pImage->uState = SUP_IOCTL_LDR_OPEN;
4461	pImage->pfnModuleInit = NULL;
4462	pImage->pfnModuleTerm = NULL;
4463	pImage->pfnServiceReqHandler= NULL;
4464	pImage->cbStrTab = 0;
4465	RTMemFree(pImage->pachStrTab);
4466	pImage->pachStrTab = NULL;
4467	RTMemFree(pImage->paSymbols);
4468	pImage->paSymbols = NULL;
4469	pImage->cSymbols = 0;
4470	}
4471
4472	supdrvLdrUnlock(pDevExt);
4473	return rc;
4474	}
4475
4476
4477	/**
4478	* Frees a previously loaded (prep'ed) image.
4479	*
4480	* @returns IPRT status code.
4481	* @param pDevExt Device globals.
4482	* @param pSession Session data.
4483	* @param pReq The request.
4484	*/
4485	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
4486	{
4487	int rc;
4488	PSUPDRVLDRUSAGE pUsagePrev;
4489	PSUPDRVLDRUSAGE pUsage;
4490	PSUPDRVLDRIMAGE pImage;
4491	LogFlow(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
4492
4493	/*
4494	* Find the ldr image.
4495	*/
4496	supdrvLdrLock(pDevExt);
4497	pUsagePrev = NULL;
4498	pUsage = pSession->pLdrUsage;
4499	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4500	{
4501	pUsagePrev = pUsage;
4502	pUsage = pUsage->pNext;
4503	}
4504	if (!pUsage)
4505	{
4506	supdrvLdrUnlock(pDevExt);
4507	Log(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
4508	return VERR_INVALID_HANDLE;
4509	}
4510
4511	/*
4512	* Check if we can remove anything.
4513	*/
4514	rc = VINF_SUCCESS;
4515	pImage = pUsage->pImage;
4516	if (pImage->cUsage <= 1 \|\| pUsage->cUsage <= 1)
4517	{
4518	/*
4519	* Check if there are any objects with destructors in the image, if
4520	* so leave it for the session cleanup routine so we get a chance to
4521	* clean things up in the right order and not leave them all dangling.
4522	*/
4523	RTSpinlockAcquire(pDevExt->Spinlock);
4524	if (pImage->cUsage <= 1)
4525	{
4526	PSUPDRVOBJ pObj;
4527	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4528	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4529	{
4530	rc = VERR_DANGLING_OBJECTS;
4531	break;
4532	}
4533	}
4534	else
4535	{
4536	PSUPDRVUSAGE pGenUsage;
4537	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
4538	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4539	{
4540	rc = VERR_DANGLING_OBJECTS;
4541	break;
4542	}
4543	}
4544	RTSpinlockRelease(pDevExt->Spinlock);
4545	if (rc == VINF_SUCCESS)
4546	{
4547	/* unlink it */
4548	if (pUsagePrev)
4549	pUsagePrev->pNext = pUsage->pNext;
4550	else
4551	pSession->pLdrUsage = pUsage->pNext;
4552
4553	/* free it */
4554	pUsage->pImage = NULL;
4555	pUsage->pNext = NULL;
4556	RTMemFree(pUsage);
4557
4558	/*
4559	* Dereference the image.
4560	*/
4561	if (pImage->cUsage <= 1)
4562	supdrvLdrFree(pDevExt, pImage);
4563	else
4564	pImage->cUsage--;
4565	}
4566	else
4567	{
4568	Log(("supdrvIOCtl_LdrFree: Dangling objects in %p/%s!\n", pImage->pvImage, pImage->szName));
4569	rc = VINF_SUCCESS; /** @todo BRANCH-2.1: remove this after branching. */
4570	}
4571	}
4572	else
4573	{
4574	/*
4575	* Dereference both image and usage.
4576	*/
4577	pImage->cUsage--;
4578	pUsage->cUsage--;
4579	}
4580
4581	supdrvLdrUnlock(pDevExt);
4582	return rc;
4583	}
4584
4585
4586	/**
4587	* Gets the address of a symbol in an open image.
4588	*
4589	* @returns IPRT status code.
4590	* @param pDevExt Device globals.
4591	* @param pSession Session data.
4592	* @param pReq The request buffer.
4593	*/
4594	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
4595	{
4596	PSUPDRVLDRIMAGE pImage;
4597	PSUPDRVLDRUSAGE pUsage;
4598	uint32_t i;
4599	PSUPLDRSYM paSyms;
4600	const char *pchStrings;
4601	const size_t cbSymbol = strlen(pReq->u.In.szSymbol) + 1;
4602	void *pvSymbol = NULL;
4603	int rc = VERR_GENERAL_FAILURE;
4604	Log3(("supdrvIOCtl_LdrGetSymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
4605
4606	/*
4607	* Find the ldr image.
4608	*/
4609	supdrvLdrLock(pDevExt);
4610	pUsage = pSession->pLdrUsage;
4611	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4612	pUsage = pUsage->pNext;
4613	if (!pUsage)
4614	{
4615	supdrvLdrUnlock(pDevExt);
4616	Log(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
4617	return VERR_INVALID_HANDLE;
4618	}
4619	pImage = pUsage->pImage;
4620	if (pImage->uState != SUP_IOCTL_LDR_LOAD)
4621	{
4622	unsigned uState = pImage->uState;
4623	supdrvLdrUnlock(pDevExt);
4624	Log(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", uState, uState)); NOREF(uState);
4625	return VERR_ALREADY_LOADED;
4626	}
4627
4628	/*
4629	* Search the symbol strings.
4630	*
4631	* Note! The int32_t is for native loading on solaris where the data
4632	* and text segments are in very different places.
4633	*/
4634	pchStrings = pImage->pachStrTab;
4635	paSyms = pImage->paSymbols;
4636	for (i = 0; i < pImage->cSymbols; i++)
4637	{
4638	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4639	&& !memcmp(pchStrings + paSyms[i].offName, pReq->u.In.szSymbol, cbSymbol))
4640	{
4641	pvSymbol = (uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol;
4642	rc = VINF_SUCCESS;
4643	break;
4644	}
4645	}
4646	supdrvLdrUnlock(pDevExt);
4647	pReq->u.Out.pvSymbol = pvSymbol;
4648	return rc;
4649	}
4650
4651
4652	/**
4653	* Gets the address of a symbol in an open image or the support driver.
4654	*
4655	* @returns VINF_SUCCESS on success.
4656	* @returns
4657	* @param pDevExt Device globals.
4658	* @param pSession Session data.
4659	* @param pReq The request buffer.
4660	*/
4661	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq)
4662	{
4663	int rc = VINF_SUCCESS;
4664	const char *pszSymbol = pReq->u.In.pszSymbol;
4665	const char *pszModule = pReq->u.In.pszModule;
4666	size_t cbSymbol;
4667	char const *pszEnd;
4668	uint32_t i;
4669
4670	/*
4671	* Input validation.
4672	*/
4673	AssertPtrReturn(pszSymbol, VERR_INVALID_POINTER);
4674	pszEnd = RTStrEnd(pszSymbol, 512);
4675	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4676	cbSymbol = pszEnd - pszSymbol + 1;
4677
4678	if (pszModule)
4679	{
4680	AssertPtrReturn(pszModule, VERR_INVALID_POINTER);
4681	pszEnd = RTStrEnd(pszModule, 64);
4682	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4683	}
4684	Log3(("supdrvIDC_LdrGetSymbol: pszModule=%p:{%s} pszSymbol=%p:{%s}\n", pszModule, pszModule, pszSymbol, pszSymbol));
4685
4686
4687	if ( !pszModule
4688	\|\| !strcmp(pszModule, "SupDrv"))
4689	{
4690	/*
4691	* Search the support driver export table.
4692	*/
4693	for (i = 0; i < RT_ELEMENTS(g_aFunctions); i++)
4694	if (!strcmp(g_aFunctions[i].szName, pszSymbol))
4695	{
4696	pReq->u.Out.pfnSymbol = g_aFunctions[i].pfn;
4697	break;
4698	}
4699	}
4700	else
4701	{
4702	/*
4703	* Find the loader image.
4704	*/
4705	PSUPDRVLDRIMAGE pImage;
4706
4707	supdrvLdrLock(pDevExt);
4708
4709	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
4710	if (!strcmp(pImage->szName, pszModule))
4711	break;
4712	if (pImage && pImage->uState == SUP_IOCTL_LDR_LOAD)
4713	{
4714	/*
4715	* Search the symbol strings.
4716	*/
4717	const char *pchStrings = pImage->pachStrTab;
4718	PCSUPLDRSYM paSyms = pImage->paSymbols;
4719	for (i = 0; i < pImage->cSymbols; i++)
4720	{
4721	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4722	&& !memcmp(pchStrings + paSyms[i].offName, pszSymbol, cbSymbol))
4723	{
4724	/*
4725	* Found it! Calc the symbol address and add a reference to the module.
4726	*/
4727	pReq->u.Out.pfnSymbol = (PFNRT)((uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol);
4728	rc = supdrvLdrAddUsage(pSession, pImage);
4729	break;
4730	}
4731	}
4732	}
4733	else
4734	rc = pImage ? VERR_WRONG_ORDER : VERR_MODULE_NOT_FOUND;
4735
4736	supdrvLdrUnlock(pDevExt);
4737	}
4738	return rc;
4739	}
4740
4741
4742	/**
4743	* Updates the VMMR0 entry point pointers.
4744	*
4745	* @returns IPRT status code.
4746	* @param pDevExt Device globals.
4747	* @param pSession Session data.
4748	* @param pVMMR0 VMMR0 image handle.
4749	* @param pvVMMR0EntryInt VMMR0EntryInt address.
4750	* @param pvVMMR0EntryFast VMMR0EntryFast address.
4751	* @param pvVMMR0EntryEx VMMR0EntryEx address.
4752	* @remark Caller must own the loader mutex.
4753	*/
4754	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx)
4755	{
4756	int rc = VINF_SUCCESS;
4757	LogFlow(("supdrvLdrSetR0EP pvVMMR0=%p pvVMMR0EntryInt=%p\n", pvVMMR0, pvVMMR0EntryInt));
4758
4759
4760	/*
4761	* Check if not yet set.
4762	*/
4763	if (!pDevExt->pvVMMR0)
4764	{
4765	pDevExt->pvVMMR0 = pvVMMR0;
4766	pDevExt->pfnVMMR0EntryInt = pvVMMR0EntryInt;
4767	pDevExt->pfnVMMR0EntryFast = pvVMMR0EntryFast;
4768	pDevExt->pfnVMMR0EntryEx = pvVMMR0EntryEx;
4769	}
4770	else
4771	{
4772	/*
4773	* Return failure or success depending on whether the values match or not.
4774	*/
4775	if ( pDevExt->pvVMMR0 != pvVMMR0
4776	\|\| (void *)pDevExt->pfnVMMR0EntryInt != pvVMMR0EntryInt
4777	\|\| (void *)pDevExt->pfnVMMR0EntryFast != pvVMMR0EntryFast
4778	\|\| (void *)pDevExt->pfnVMMR0EntryEx != pvVMMR0EntryEx)
4779	{
4780	AssertMsgFailed(("SUP_IOCTL_LDR_SETR0EP: Already set pointing to a different module!\n"));
4781	rc = VERR_INVALID_PARAMETER;
4782	}
4783	}
4784	return rc;
4785	}
4786
4787
4788	/**
4789	* Unsets the VMMR0 entry point installed by supdrvLdrSetR0EP.
4790	*
4791	* @param pDevExt Device globals.
4792	*/
4793	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt)
4794	{
4795	pDevExt->pvVMMR0 = NULL;
4796	pDevExt->pfnVMMR0EntryInt = NULL;
4797	pDevExt->pfnVMMR0EntryFast = NULL;
4798	pDevExt->pfnVMMR0EntryEx = NULL;
4799	}
4800
4801
4802	/**
4803	* Adds a usage reference in the specified session of an image.
4804	*
4805	* Called while owning the loader semaphore.
4806	*
4807	* @returns VINF_SUCCESS on success and VERR_NO_MEMORY on failure.
4808	* @param pSession Session in question.
4809	* @param pImage Image which the session is using.
4810	*/
4811	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage)
4812	{
4813	PSUPDRVLDRUSAGE pUsage;
4814	LogFlow(("supdrvLdrAddUsage: pImage=%p\n", pImage));
4815
4816	/*
4817	* Referenced it already?
4818	*/
4819	pUsage = pSession->pLdrUsage;
4820	while (pUsage)
4821	{
4822	if (pUsage->pImage == pImage)
4823	{
4824	pUsage->cUsage++;
4825	return VINF_SUCCESS;
4826	}
4827	pUsage = pUsage->pNext;
4828	}
4829
4830	/*
4831	* Allocate new usage record.
4832	*/
4833	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
4834	AssertReturn(pUsage, /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_5);
4835	pUsage->cUsage = 1;
4836	pUsage->pImage = pImage;
4837	pUsage->pNext = pSession->pLdrUsage;
4838	pSession->pLdrUsage = pUsage;
4839	return VINF_SUCCESS;
4840	}
4841
4842
4843	/**
4844	* Frees a load image.
4845	*
4846	* @param pDevExt Pointer to device extension.
4847	* @param pImage Pointer to the image we're gonna free.
4848	* This image must exit!
4849	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
4850	*/
4851	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
4852	{
4853	PSUPDRVLDRIMAGE pImagePrev;
4854	LogFlow(("supdrvLdrFree: pImage=%p\n", pImage));
4855
4856	/* find it - arg. should've used doubly linked list. */
4857	Assert(pDevExt->pLdrImages);
4858	pImagePrev = NULL;
4859	if (pDevExt->pLdrImages != pImage)
4860	{
4861	pImagePrev = pDevExt->pLdrImages;
4862	while (pImagePrev->pNext != pImage)
4863	pImagePrev = pImagePrev->pNext;
4864	Assert(pImagePrev->pNext == pImage);
4865	}
4866
4867	/* unlink */
4868	if (pImagePrev)
4869	pImagePrev->pNext = pImage->pNext;
4870	else
4871	pDevExt->pLdrImages = pImage->pNext;
4872
4873	/* check if this is VMMR0.r0 unset its entry point pointers. */
4874	if (pDevExt->pvVMMR0 == pImage->pvImage)
4875	supdrvLdrUnsetVMMR0EPs(pDevExt);
4876
4877	/* check for objects with destructors in this image. (Shouldn't happen.) */
4878	if (pDevExt->pObjs)
4879	{
4880	unsigned cObjs = 0;
4881	PSUPDRVOBJ pObj;
4882	RTSpinlockAcquire(pDevExt->Spinlock);
4883	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4884	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4885	{
4886	pObj->pfnDestructor = NULL;
4887	cObjs++;
4888	}
4889	RTSpinlockRelease(pDevExt->Spinlock);
4890	if (cObjs)
4891	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
4892	}
4893
4894	/* call termination function if fully loaded. */
4895	if ( pImage->pfnModuleTerm
4896	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
4897	{
4898	LogFlow(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
4899	pImage->pfnModuleTerm(pImage);
4900	}
4901
4902	/* Inform the tracing component. */
4903	supdrvTracerModuleUnloading(pDevExt, pImage);
4904
4905	/* do native unload if appropriate. */
4906	if (pImage->fNative)
4907	supdrvOSLdrUnload(pDevExt, pImage);
4908
4909	/* free the image */
4910	pImage->cUsage = 0;
4911	pImage->pDevExt = NULL;
4912	pImage->pNext = NULL;
4913	pImage->uState = SUP_IOCTL_LDR_FREE;
4914	RTMemExecFree(pImage->pvImageAlloc, pImage->cbImageBits + 31);
4915	pImage->pvImageAlloc = NULL;
4916	RTMemFree(pImage->pachStrTab);
4917	pImage->pachStrTab = NULL;
4918	RTMemFree(pImage->paSymbols);
4919	pImage->paSymbols = NULL;
4920	RTMemFree(pImage);
4921	}
4922
4923
4924	/**
4925	* Acquires the loader lock.
4926	*
4927	* @returns IPRT status code.
4928	* @param pDevExt The device extension.
4929	*/
4930	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt)
4931	{
4932	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4933	int rc = RTSemMutexRequest(pDevExt->mtxLdr, RT_INDEFINITE_WAIT);
4934	#else
4935	int rc = RTSemFastMutexRequest(pDevExt->mtxLdr);
4936	#endif
4937	AssertRC(rc);
4938	return rc;
4939	}
4940
4941
4942	/**
4943	* Releases the loader lock.
4944	*
4945	* @returns IPRT status code.
4946	* @param pDevExt The device extension.
4947	*/
4948	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt)
4949	{
4950	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4951	return RTSemMutexRelease(pDevExt->mtxLdr);
4952	#else
4953	return RTSemFastMutexRelease(pDevExt->mtxLdr);
4954	#endif
4955	}
4956
4957
4958	/**
4959	* Implements the service call request.
4960	*
4961	* @returns VBox status code.
4962	* @param pDevExt The device extension.
4963	* @param pSession The calling session.
4964	* @param pReq The request packet, valid.
4965	*/
4966	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq)
4967	{
4968	#if !defined(RT_OS_WINDOWS) \|\| defined(RT_ARCH_AMD64) \|\| defined(DEBUG)
4969	int rc;
4970
4971	/*
4972	* Find the module first in the module referenced by the calling session.
4973	*/
4974	rc = supdrvLdrLock(pDevExt);
4975	if (RT_SUCCESS(rc))
4976	{
4977	PFNSUPR0SERVICEREQHANDLER pfnServiceReqHandler = NULL;
4978	PSUPDRVLDRUSAGE pUsage;
4979
4980	for (pUsage = pSession->pLdrUsage; pUsage; pUsage = pUsage->pNext)
4981	if ( pUsage->pImage->pfnServiceReqHandler
4982	&& !strcmp(pUsage->pImage->szName, pReq->u.In.szName))
4983	{
4984	pfnServiceReqHandler = pUsage->pImage->pfnServiceReqHandler;
4985	break;
4986	}
4987	supdrvLdrUnlock(pDevExt);
4988
4989	if (pfnServiceReqHandler)
4990	{
4991	/*
4992	* Call it.
4993	*/
4994	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
4995	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, NULL);
4996	else
4997	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0]);
4998	}
4999	else
5000	rc = VERR_SUPDRV_SERVICE_NOT_FOUND;
5001	}
5002
5003	/* log it */
5004	if ( RT_FAILURE(rc)
5005	&& rc != VERR_INTERRUPTED
5006	&& rc != VERR_TIMEOUT)
5007	Log(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
5008	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
5009	else
5010	Log4(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
5011	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
5012	return rc;
5013	#else /* RT_OS_WINDOWS && !RT_ARCH_AMD64 && !DEBUG */
5014	return VERR_NOT_IMPLEMENTED;
5015	#endif /* RT_OS_WINDOWS && !RT_ARCH_AMD64 && !DEBUG */
5016	}
5017
5018
5019	/**
5020	* Implements the logger settings request.
5021	*
5022	* @returns VBox status code.
5023	* @param pDevExt The device extension.
5024	* @param pSession The caller's session.
5025	* @param pReq The request.
5026	*/
5027	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq)
5028	{
5029	const char *pszGroup = &pReq->u.In.szStrings[pReq->u.In.offGroups];
5030	const char *pszFlags = &pReq->u.In.szStrings[pReq->u.In.offFlags];
5031	const char *pszDest = &pReq->u.In.szStrings[pReq->u.In.offDestination];
5032	PRTLOGGER pLogger = NULL;
5033	int rc;
5034
5035	/*
5036	* Some further validation.
5037	*/
5038	switch (pReq->u.In.fWhat)
5039	{
5040	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
5041	case SUPLOGGERSETTINGS_WHAT_CREATE:
5042	break;
5043
5044	case SUPLOGGERSETTINGS_WHAT_DESTROY:
5045	if (pszGroup \|\| pszFlags \|\| *pszDest)
5046	return VERR_INVALID_PARAMETER;
5047	if (pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_RELEASE)
5048	return VERR_ACCESS_DENIED;
5049	break;
5050
5051	default:
5052	return VERR_INTERNAL_ERROR;
5053	}
5054
5055	/*
5056	* Get the logger.
5057	*/
5058	switch (pReq->u.In.fWhich)
5059	{
5060	case SUPLOGGERSETTINGS_WHICH_DEBUG:
5061	pLogger = RTLogGetDefaultInstance();
5062	break;
5063
5064	case SUPLOGGERSETTINGS_WHICH_RELEASE:
5065	pLogger = RTLogRelDefaultInstance();
5066	break;
5067
5068	default:
5069	return VERR_INTERNAL_ERROR;
5070	}
5071
5072	/*
5073	* Do the job.
5074	*/
5075	switch (pReq->u.In.fWhat)
5076	{
5077	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
5078	if (pLogger)
5079	{
5080	rc = RTLogFlags(pLogger, pszFlags);
5081	if (RT_SUCCESS(rc))
5082	rc = RTLogGroupSettings(pLogger, pszGroup);
5083	NOREF(pszDest);
5084	}
5085	else
5086	rc = VERR_NOT_FOUND;
5087	break;
5088
5089	case SUPLOGGERSETTINGS_WHAT_CREATE:
5090	{
5091	if (pLogger)
5092	rc = VERR_ALREADY_EXISTS;
5093	else
5094	{
5095	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
5096
5097	rc = RTLogCreate(&pLogger,
5098	0 /* fFlags */,
5099	pszGroup,
5100	pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_DEBUG
5101	? "VBOX_LOG"
5102	: "VBOX_RELEASE_LOG",
5103	RT_ELEMENTS(s_apszGroups),
5104	s_apszGroups,
5105	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER,
5106	NULL);
5107	if (RT_SUCCESS(rc))
5108	{
5109	rc = RTLogFlags(pLogger, pszFlags);
5110	NOREF(pszDest);
5111	if (RT_SUCCESS(rc))
5112	{
5113	switch (pReq->u.In.fWhich)
5114	{
5115	case SUPLOGGERSETTINGS_WHICH_DEBUG:
5116	pLogger = RTLogSetDefaultInstance(pLogger);
5117	break;
5118	case SUPLOGGERSETTINGS_WHICH_RELEASE:
5119	pLogger = RTLogRelSetDefaultInstance(pLogger);
5120	break;
5121	}
5122	}
5123	RTLogDestroy(pLogger);
5124	}
5125	}
5126	break;
5127	}
5128
5129	case SUPLOGGERSETTINGS_WHAT_DESTROY:
5130	switch (pReq->u.In.fWhich)
5131	{
5132	case SUPLOGGERSETTINGS_WHICH_DEBUG:
5133	pLogger = RTLogSetDefaultInstance(NULL);
5134	break;
5135	case SUPLOGGERSETTINGS_WHICH_RELEASE:
5136	pLogger = RTLogRelSetDefaultInstance(NULL);
5137	break;
5138	}
5139	rc = RTLogDestroy(pLogger);
5140	break;
5141
5142	default:
5143	{
5144	rc = VERR_INTERNAL_ERROR;
5145	break;
5146	}
5147	}
5148
5149	return rc;
5150	}
5151
5152
5153	/**
5154	* Implements the MSR prober operations.
5155	*
5156	* @returns VBox status code.
5157	* @param pDevExt The device extension.
5158	* @param pReq The request.
5159	*/
5160	static int supdrvIOCtl_MsrProber(PSUPDRVDEVEXT pDevExt, PSUPMSRPROBER pReq)
5161	{
5162	#ifdef SUPDRV_WITH_MSR_PROBER
5163	RTCPUID const idCpu = pReq->u.In.idCpu == UINT32_MAX ? NIL_RTCPUID : pReq->u.In.idCpu;
5164	int rc;
5165
5166	switch (pReq->u.In.enmOp)
5167	{
5168	case SUPMSRPROBEROP_READ:
5169	{
5170	uint64_t uValue;
5171	rc = supdrvOSMsrProberRead(pReq->u.In.uMsr, idCpu, &uValue);
5172	if (RT_SUCCESS(rc))
5173	{
5174	pReq->u.Out.uResults.Read.uValue = uValue;
5175	pReq->u.Out.uResults.Read.fGp = false;
5176	}
5177	else if (rc == VERR_ACCESS_DENIED)
5178	{
5179	pReq->u.Out.uResults.Read.uValue = 0;
5180	pReq->u.Out.uResults.Read.fGp = true;
5181	rc = VINF_SUCCESS;
5182	}
5183	break;
5184	}
5185
5186	case SUPMSRPROBEROP_WRITE:
5187	rc = supdrvOSMsrProberWrite(pReq->u.In.uMsr, idCpu, pReq->u.In.uArgs.Write.uToWrite);
5188	if (RT_SUCCESS(rc))
5189	pReq->u.Out.uResults.Write.fGp = false;
5190	else if (rc == VERR_ACCESS_DENIED)
5191	{
5192	pReq->u.Out.uResults.Write.fGp = true;
5193	rc = VINF_SUCCESS;
5194	}
5195	break;
5196
5197	case SUPMSRPROBEROP_MODIFY:
5198	case SUPMSRPROBEROP_MODIFY_FASTER:
5199	rc = supdrvOSMsrProberModify(idCpu, pReq);
5200	break;
5201
5202	default:
5203	return VERR_INVALID_FUNCTION;
5204	}
5205	return rc;
5206	#else
5207	return VERR_NOT_IMPLEMENTED;
5208	#endif
5209	}
5210
5211
5212
5213	/**
5214	* Creates the GIP.
5215	*
5216	* @returns VBox status code.
5217	* @param pDevExt Instance data. GIP stuff may be updated.
5218	*/
5219	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
5220	{
5221	PSUPGLOBALINFOPAGE pGip;
5222	RTHCPHYS HCPhysGip;
5223	uint32_t u32SystemResolution;
5224	uint32_t u32Interval;
5225	unsigned cCpus;
5226	int rc;
5227
5228
5229	LogFlow(("supdrvGipCreate:\n"));
5230
5231	/* assert order */
5232	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
5233	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
5234	Assert(!pDevExt->pGipTimer);
5235
5236	/*
5237	* Check the CPU count.
5238	*/
5239	cCpus = RTMpGetArraySize();
5240	if ( cCpus > RTCPUSET_MAX_CPUS
5241	\|\| cCpus > 256 /ApicId is used for the mappings/)
5242	{
5243	SUPR0Printf("VBoxDrv: Too many CPUs (%u) for the GIP (max %u)\n", cCpus, RT_MIN(RTCPUSET_MAX_CPUS, 256));
5244	return VERR_TOO_MANY_CPUS;
5245	}
5246
5247	/*
5248	* Allocate a contiguous set of pages with a default kernel mapping.
5249	*/
5250	rc = RTR0MemObjAllocCont(&pDevExt->GipMemObj, RT_UOFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), false /fExecutable/);
5251	if (RT_FAILURE(rc))
5252	{
5253	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
5254	return rc;
5255	}
5256	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
5257	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
5258
5259	/*
5260	* Find a reasonable update interval and initialize the structure.
5261	*/
5262	u32Interval = u32SystemResolution = RTTimerGetSystemGranularity();
5263	while (u32Interval < 10000000 /* 10 ms */)
5264	u32Interval += u32SystemResolution;
5265
5266	supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), 1000000000 / u32Interval /=Hz/, cCpus);
5267
5268	/*
5269	* Create the timer.
5270	* If CPU_ALL isn't supported we'll have to fall back to synchronous mode.
5271	*/
5272	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
5273	{
5274	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, RTTIMER_FLAGS_CPU_ALL, supdrvGipAsyncTimer, pDevExt);
5275	if (rc == VERR_NOT_SUPPORTED)
5276	{
5277	OSDBGPRINT(("supdrvGipCreate: omni timer not supported, falling back to synchronous mode\n"));
5278	pGip->u32Mode = SUPGIPMODE_SYNC_TSC;
5279	}
5280	}
5281	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
5282	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0 /* fFlags */, supdrvGipSyncTimer, pDevExt);
5283	if (RT_SUCCESS(rc))
5284	{
5285	rc = RTMpNotificationRegister(supdrvGipMpEvent, pDevExt);
5286	if (RT_SUCCESS(rc))
5287	{
5288	rc = RTMpOnAll(supdrvGipInitOnCpu, pDevExt, pGip);
5289	if (RT_SUCCESS(rc))
5290	{
5291	/*
5292	* We're good.
5293	*/
5294	Log(("supdrvGipCreate: %u ns interval.\n", u32Interval));
5295	g_pSUPGlobalInfoPage = pGip;
5296	return VINF_SUCCESS;
5297	}
5298
5299	OSDBGPRINT(("supdrvGipCreate: RTMpOnAll failed with rc=%Rrc\n", rc));
5300	RTMpNotificationDeregister(supdrvGipMpEvent, pDevExt);
5301
5302	}
5303	else
5304	OSDBGPRINT(("supdrvGipCreate: failed to register MP event notfication. rc=%Rrc\n", rc));
5305	}
5306	else
5307	{
5308	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %u ns interval. rc=%Rrc\n", u32Interval, rc));
5309	Assert(!pDevExt->pGipTimer);
5310	}
5311	supdrvGipDestroy(pDevExt);
5312	return rc;
5313	}
5314
5315
5316	/**
5317	* Terminates the GIP.
5318	*
5319	* @param pDevExt Instance data. GIP stuff may be updated.
5320	*/
5321	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
5322	{
5323	int rc;
5324	#ifdef DEBUG_DARWIN_GIP
5325	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
5326	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
5327	pDevExt->pGipTimer, pDevExt->GipMemObj));
5328	#endif
5329
5330	/*
5331	* Invalid the GIP data.
5332	*/
5333	if (pDevExt->pGip)
5334	{
5335	supdrvGipTerm(pDevExt->pGip);
5336	pDevExt->pGip = NULL;
5337	}
5338	g_pSUPGlobalInfoPage = NULL;
5339
5340	/*
5341	* Destroy the timer and free the GIP memory object.
5342	*/
5343	if (pDevExt->pGipTimer)
5344	{
5345	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
5346	pDevExt->pGipTimer = NULL;
5347	}
5348
5349	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
5350	{
5351	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
5352	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
5353	}
5354
5355	/*
5356	* Finally, make sure we've release the system timer resolution request
5357	* if one actually succeeded and is still pending.
5358	*/
5359	if (pDevExt->u32SystemTimerGranularityGrant)
5360	{
5361	rc = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
5362	pDevExt->u32SystemTimerGranularityGrant = 0;
5363	}
5364	}
5365
5366
5367	/**
5368	* Timer callback function sync GIP mode.
5369	* @param pTimer The timer.
5370	* @param pvUser The device extension.
5371	*/
5372	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
5373	{
5374	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
5375	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5376	uint64_t u64TSC = ASMReadTSC();
5377	uint64_t NanoTS = RTTimeSystemNanoTS();
5378
5379	supdrvGipUpdate(pDevExt, NanoTS, u64TSC, NIL_RTCPUID, iTick);
5380
5381	ASMSetFlags(fOldFlags);
5382	}
5383
5384
5385	/**
5386	* Timer callback function for async GIP mode.
5387	* @param pTimer The timer.
5388	* @param pvUser The device extension.
5389	*/
5390	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
5391	{
5392	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
5393	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5394	RTCPUID idCpu = RTMpCpuId();
5395	uint64_t u64TSC = ASMReadTSC();
5396	uint64_t NanoTS = RTTimeSystemNanoTS();
5397
5398	/** @todo reset the transaction number and whatnot when iTick == 1. */
5399	if (pDevExt->idGipMaster == idCpu)
5400	supdrvGipUpdate(pDevExt, NanoTS, u64TSC, idCpu, iTick);
5401	else
5402	supdrvGipUpdatePerCpu(pDevExt, NanoTS, u64TSC, idCpu, ASMGetApicId(), iTick);
5403
5404	ASMSetFlags(fOldFlags);
5405	}
5406
5407
5408	/**
5409	* Finds our (@a idCpu) entry, or allocates a new one if not found.
5410	*
5411	* @returns Index of the CPU in the cache set.
5412	* @param pGip The GIP.
5413	* @param idCpu The CPU ID.
5414	*/
5415	static uint32_t supdrvGipCpuIndexFromCpuId(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
5416	{
5417	uint32_t i, cTries;
5418
5419	/*
5420	* ASSUMES that CPU IDs are constant.
5421	*/
5422	for (i = 0; i < pGip->cCpus; i++)
5423	if (pGip->aCPUs[i].idCpu == idCpu)
5424	return i;
5425
5426	cTries = 0;
5427	do
5428	{
5429	for (i = 0; i < pGip->cCpus; i++)
5430	{
5431	bool fRc;
5432	ASMAtomicCmpXchgSize(&pGip->aCPUs[i].idCpu, idCpu, NIL_RTCPUID, fRc);
5433	if (fRc)
5434	return i;
5435	}
5436	} while (cTries++ < 32);
5437	AssertReleaseFailed();
5438	return i - 1;
5439	}
5440
5441
5442	/**
5443	* The calling CPU should be accounted as online, update GIP accordingly.
5444	*
5445	* This is used by supdrvGipMpEvent as well as the supdrvGipCreate.
5446	*
5447	* @param pDevExt The device extension.
5448	* @param idCpu The CPU ID.
5449	*/
5450	static void supdrvGipMpEventOnline(PSUPDRVDEVEXT pDevExt, RTCPUID idCpu)
5451	{
5452	int iCpuSet = 0;
5453	uint16_t idApic = UINT16_MAX;
5454	uint32_t i = 0;
5455	uint64_t u64NanoTS = 0;
5456	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5457
5458	AssertPtrReturnVoid(pGip);
5459	AssertRelease(idCpu == RTMpCpuId());
5460	Assert(pGip->cPossibleCpus == RTMpGetCount());
5461
5462	/*
5463	* Do this behind a spinlock with interrupts disabled as this can fire
5464	* on all CPUs simultaneously, see @bugref{6110}.
5465	*/
5466	RTSpinlockAcquire(pDevExt->hGipSpinlock);
5467
5468	/*
5469	* Update the globals.
5470	*/
5471	ASMAtomicWriteU16(&pGip->cPresentCpus, RTMpGetPresentCount());
5472	ASMAtomicWriteU16(&pGip->cOnlineCpus, RTMpGetOnlineCount());
5473	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5474	if (iCpuSet >= 0)
5475	{
5476	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5477	RTCpuSetAddByIndex(&pGip->OnlineCpuSet, iCpuSet);
5478	RTCpuSetAddByIndex(&pGip->PresentCpuSet, iCpuSet);
5479	}
5480
5481	/*
5482	* Update the entry.
5483	*/
5484	u64NanoTS = RTTimeSystemNanoTS() - pGip->u32UpdateIntervalNS;
5485	i = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
5486	supdrvGipInitCpu(pGip, &pGip->aCPUs[i], u64NanoTS);
5487	idApic = ASMGetApicId();
5488	ASMAtomicWriteU16(&pGip->aCPUs[i].idApic, idApic);
5489	ASMAtomicWriteS16(&pGip->aCPUs[i].iCpuSet, (int16_t)iCpuSet);
5490	ASMAtomicWriteSize(&pGip->aCPUs[i].idCpu, idCpu);
5491
5492	/*
5493	* Update the APIC ID and CPU set index mappings.
5494	*/
5495	ASMAtomicWriteU16(&pGip->aiCpuFromApicId[idApic], i);
5496	ASMAtomicWriteU16(&pGip->aiCpuFromCpuSetIdx[iCpuSet], i);
5497
5498	/* commit it */
5499	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_ONLINE);
5500
5501	RTSpinlockReleaseNoInts(pDevExt->hGipSpinlock);
5502	}
5503
5504
5505	/**
5506	* The CPU should be accounted as offline, update the GIP accordingly.
5507	*
5508	* This is used by supdrvGipMpEvent.
5509	*
5510	* @param pDevExt The device extension.
5511	* @param idCpu The CPU ID.
5512	*/
5513	static void supdrvGipMpEventOffline(PSUPDRVDEVEXT pDevExt, RTCPUID idCpu)
5514	{
5515	int iCpuSet;
5516	unsigned i;
5517
5518	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5519
5520	AssertPtrReturnVoid(pGip);
5521	RTSpinlockAcquire(pDevExt->hGipSpinlock);
5522
5523	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5524	AssertReturnVoid(iCpuSet >= 0);
5525
5526	i = pGip->aiCpuFromCpuSetIdx[iCpuSet];
5527	AssertReturnVoid(i < pGip->cCpus);
5528	AssertReturnVoid(pGip->aCPUs[i].idCpu == idCpu);
5529
5530	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5531	RTCpuSetDelByIndex(&pGip->OnlineCpuSet, iCpuSet);
5532
5533	/* commit it */
5534	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_OFFLINE);
5535
5536	RTSpinlockReleaseNoInts(pDevExt->hGipSpinlock);
5537	}
5538
5539
5540	/**
5541	* Multiprocessor event notification callback.
5542	*
5543	* This is used to make sure that the GIP master gets passed on to
5544	* another CPU. It also updates the associated CPU data.
5545	*
5546	* @param enmEvent The event.
5547	* @param idCpu The cpu it applies to.
5548	* @param pvUser Pointer to the device extension.
5549	*
5550	* @remarks This function -must- fire on the newly online'd CPU for the
5551	* RTMPEVENT_ONLINE case and can fire on any CPU for the
5552	* RTMPEVENT_OFFLINE case.
5553	*/
5554	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser)
5555	{
5556	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5557	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5558
5559	AssertRelease(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
5560
5561	/*
5562	* Update the GIP CPU data.
5563	*/
5564	if (pGip)
5565	{
5566	switch (enmEvent)
5567	{
5568	case RTMPEVENT_ONLINE:
5569	AssertRelease(idCpu == RTMpCpuId());
5570	supdrvGipMpEventOnline(pDevExt, idCpu);
5571	break;
5572	case RTMPEVENT_OFFLINE:
5573	supdrvGipMpEventOffline(pDevExt, idCpu);
5574	break;
5575
5576	}
5577	}
5578
5579	/*
5580	* Make sure there is a master GIP.
5581	*/
5582	if (enmEvent == RTMPEVENT_OFFLINE)
5583	{
5584	RTCPUID idGipMaster = ASMAtomicReadU32(&pDevExt->idGipMaster);
5585	if (idGipMaster == idCpu)
5586	{
5587	/*
5588	* Find a new GIP master.
5589	*/
5590	bool fIgnored;
5591	unsigned i;
5592	RTCPUID idNewGipMaster = NIL_RTCPUID;
5593	RTCPUSET OnlineCpus;
5594	RTMpGetOnlineSet(&OnlineCpus);
5595
5596	for (i = 0; i < RTCPUSET_MAX_CPUS; i++)
5597	{
5598	RTCPUID idCurCpu = RTMpCpuIdFromSetIndex(i);
5599	if ( RTCpuSetIsMember(&OnlineCpus, idCurCpu)
5600	&& idCurCpu != idGipMaster)
5601	{
5602	idNewGipMaster = idCurCpu;
5603	break;
5604	}
5605	}
5606
5607	Log(("supdrvGipMpEvent: Gip master %#lx -> %#lx\n", (long)idGipMaster, (long)idNewGipMaster));
5608	ASMAtomicCmpXchgSize(&pDevExt->idGipMaster, idNewGipMaster, idGipMaster, fIgnored);
5609	NOREF(fIgnored);
5610	}
5611	}
5612	}
5613
5614
5615	/**
5616	* Callback used by supdrvDetermineAsyncTSC to read the TSC on a CPU.
5617	*
5618	* @param idCpu Ignored.
5619	* @param pvUser1 Where to put the TSC.
5620	* @param pvUser2 Ignored.
5621	*/
5622	static DECLCALLBACK(void) supdrvDetermineAsyncTscWorker(RTCPUID idCpu, void pvUser1, void pvUser2)
5623	{
5624	#if 1
5625	ASMAtomicWriteU64((uint64_t volatile *)pvUser1, ASMReadTSC());
5626	#else
5627	(uint64_t )pvUser1 = ASMReadTSC();
5628	#endif
5629	}
5630
5631
5632	/**
5633	* Determine if Async GIP mode is required because of TSC drift.
5634	*
5635	* When using the default/normal timer code it is essential that the time stamp counter
5636	* (TSC) runs never backwards, that is, a read operation to the counter should return
5637	* a bigger value than any previous read operation. This is guaranteed by the latest
5638	* AMD CPUs and by newer Intel CPUs which never enter the C2 state (P4). In any other
5639	* case we have to choose the asynchronous timer mode.
5640	*
5641	* @param poffMin Pointer to the determined difference between different cores.
5642	* @return false if the time stamp counters appear to be synchronized, true otherwise.
5643	*/
5644	static bool supdrvDetermineAsyncTsc(uint64_t *poffMin)
5645	{
5646	/*
5647	* Just iterate all the cpus 8 times and make sure that the TSC is
5648	* ever increasing. We don't bother taking TSC rollover into account.
5649	*/
5650	int iEndCpu = RTMpGetArraySize();
5651	int iCpu;
5652	int cLoops = 8;
5653	bool fAsync = false;
5654	int rc = VINF_SUCCESS;
5655	uint64_t offMax = 0;
5656	uint64_t offMin = ~(uint64_t)0;
5657	uint64_t PrevTsc = ASMReadTSC();
5658
5659	while (cLoops-- > 0)
5660	{
5661	for (iCpu = 0; iCpu < iEndCpu; iCpu++)
5662	{
5663	uint64_t CurTsc;
5664	rc = RTMpOnSpecific(RTMpCpuIdFromSetIndex(iCpu), supdrvDetermineAsyncTscWorker, &CurTsc, NULL);
5665	if (RT_SUCCESS(rc))
5666	{
5667	if (CurTsc <= PrevTsc)
5668	{
5669	fAsync = true;
5670	offMin = offMax = PrevTsc - CurTsc;
5671	Log(("supdrvDetermineAsyncTsc: iCpu=%d cLoops=%d CurTsc=%llx PrevTsc=%llx\n",
5672	iCpu, cLoops, CurTsc, PrevTsc));
5673	break;
5674	}
5675
5676	/* Gather statistics (except the first time). */
5677	if (iCpu != 0 \|\| cLoops != 7)
5678	{
5679	uint64_t off = CurTsc - PrevTsc;
5680	if (off < offMin)
5681	offMin = off;
5682	if (off > offMax)
5683	offMax = off;
5684	Log2(("%d/%d: off=%llx\n", cLoops, iCpu, off));
5685	}
5686
5687	/* Next */
5688	PrevTsc = CurTsc;
5689	}
5690	else if (rc == VERR_NOT_SUPPORTED)
5691	break;
5692	else
5693	AssertMsg(rc == VERR_CPU_NOT_FOUND \|\| rc == VERR_CPU_OFFLINE, ("%d\n", rc));
5694	}
5695
5696	/* broke out of the loop. */
5697	if (iCpu < iEndCpu)
5698	break;
5699	}
5700
5701	poffMin = offMin; / Almost RTMpOnSpecific profiling. */
5702	Log(("supdrvDetermineAsyncTsc: returns %d; iEndCpu=%d rc=%d offMin=%llx offMax=%llx\n",
5703	fAsync, iEndCpu, rc, offMin, offMax));
5704	#if !defined(RT_OS_SOLARIS) && !defined(RT_OS_OS2) && !defined(RT_OS_WINDOWS)
5705	OSDBGPRINT(("vboxdrv: fAsync=%d offMin=%#lx offMax=%#lx\n", fAsync, (long)offMin, (long)offMax));
5706	#endif
5707	return fAsync;
5708	}
5709
5710
5711	/**
5712	* Determine the GIP TSC mode.
5713	*
5714	* @returns The most suitable TSC mode.
5715	* @param pDevExt Pointer to the device instance data.
5716	*/
5717	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt)
5718	{
5719	/*
5720	* On SMP we're faced with two problems:
5721	* (1) There might be a skew between the CPU, so that cpu0
5722	* returns a TSC that is slightly different from cpu1.
5723	* (2) Power management (and other things) may cause the TSC
5724	* to run at a non-constant speed, and cause the speed
5725	* to be different on the cpus. This will result in (1).
5726	*
5727	* So, on SMP systems we'll have to select the ASYNC update method
5728	* if there are symptoms of these problems.
5729	*/
5730	if (RTMpGetCount() > 1)
5731	{
5732	uint32_t uEAX, uEBX, uECX, uEDX;
5733	uint64_t u64DiffCoresIgnored;
5734
5735	/* Permit the user and/or the OS specific bits to force async mode. */
5736	if (supdrvOSGetForcedAsyncTscMode(pDevExt))
5737	return SUPGIPMODE_ASYNC_TSC;
5738
5739	/* Try check for current differences between the cpus. */
5740	if (supdrvDetermineAsyncTsc(&u64DiffCoresIgnored))
5741	return SUPGIPMODE_ASYNC_TSC;
5742
5743	/*
5744	* If the CPU supports power management and is an AMD one we
5745	* won't trust it unless it has the TscInvariant bit is set.
5746	*/
5747	/* Check for "AuthenticAMD" */
5748	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
5749	if ( uEAX >= 1
5750	&& ASMIsAmdCpuEx(uEBX, uECX, uEDX))
5751	{
5752	/* Check for APM support and that TscInvariant is cleared. */
5753	ASMCpuId(0x80000000, &uEAX, &uEBX, &uECX, &uEDX);
5754	if (uEAX >= 0x80000007)
5755	{
5756	ASMCpuId(0x80000007, &uEAX, &uEBX, &uECX, &uEDX);
5757	if ( !(uEDX & RT_BIT(8))/* TscInvariant */
5758	&& (uEDX & 0x3e)) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
5759	return SUPGIPMODE_ASYNC_TSC;
5760	}
5761	}
5762	}
5763	return SUPGIPMODE_SYNC_TSC;
5764	}
5765
5766
5767	/**
5768	* Initializes per-CPU GIP information.
5769	*
5770	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5771	* @param pCpu Pointer to which GIP CPU to initalize.
5772	* @param u64NanoTS The current nanosecond timestamp.
5773	*/
5774	static void supdrvGipInitCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pCpu, uint64_t u64NanoTS)
5775	{
5776	pCpu->u32TransactionId = 2;
5777	pCpu->u64NanoTS = u64NanoTS;
5778	pCpu->u64TSC = ASMReadTSC();
5779
5780	ASMAtomicWriteSize(&pCpu->enmState, SUPGIPCPUSTATE_INVALID);
5781	ASMAtomicWriteSize(&pCpu->idCpu, NIL_RTCPUID);
5782	ASMAtomicWriteS16(&pCpu->iCpuSet, -1);
5783	ASMAtomicWriteU16(&pCpu->idApic, UINT16_MAX);
5784
5785	/*
5786	* We don't know the following values until we've executed updates.
5787	* So, we'll just pretend it's a 4 GHz CPU and adjust the history it on
5788	* the 2nd timer callout.
5789	*/
5790	pCpu->u64CpuHz = _4G + 1; /* tstGIP-2 depends on this. */
5791	pCpu->u32UpdateIntervalTSC
5792	= pCpu->au32TSCHistory[0]
5793	= pCpu->au32TSCHistory[1]
5794	= pCpu->au32TSCHistory[2]
5795	= pCpu->au32TSCHistory[3]
5796	= pCpu->au32TSCHistory[4]
5797	= pCpu->au32TSCHistory[5]
5798	= pCpu->au32TSCHistory[6]
5799	= pCpu->au32TSCHistory[7]
5800	= (uint32_t)(_4G / pGip->u32UpdateHz);
5801	}
5802
5803
5804	/**
5805	* Initializes the GIP data.
5806	*
5807	* @param pDevExt Pointer to the device instance data.
5808	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5809	* @param HCPhys The physical address of the GIP.
5810	* @param u64NanoTS The current nanosecond timestamp.
5811	* @param uUpdateHz The update frequency.
5812	* @param cCpus The CPU count.
5813	*/
5814	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
5815	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned cCpus)
5816	{
5817	size_t const cbGip = RT_ALIGN_Z(RT_OFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), PAGE_SIZE);
5818	unsigned i;
5819	#ifdef DEBUG_DARWIN_GIP
5820	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5821	#else
5822	LogFlow(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5823	#endif
5824
5825	/*
5826	* Initialize the structure.
5827	*/
5828	memset(pGip, 0, cbGip);
5829	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
5830	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
5831	pGip->u32Mode = supdrvGipDeterminTscMode(pDevExt);
5832	pGip->cCpus = (uint16_t)cCpus;
5833	pGip->cPages = (uint16_t)(cbGip / PAGE_SIZE);
5834	pGip->u32UpdateHz = uUpdateHz;
5835	pGip->u32UpdateIntervalNS = 1000000000 / uUpdateHz;
5836	pGip->u64NanoTSLastUpdateHz = u64NanoTS;
5837	RTCpuSetEmpty(&pGip->OnlineCpuSet);
5838	RTCpuSetEmpty(&pGip->PresentCpuSet);
5839	RTMpGetSet(&pGip->PossibleCpuSet);
5840	pGip->cOnlineCpus = RTMpGetOnlineCount();
5841	pGip->cPresentCpus = RTMpGetPresentCount();
5842	pGip->cPossibleCpus = RTMpGetCount();
5843	pGip->idCpuMax = RTMpGetMaxCpuId();
5844	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromApicId); i++)
5845	pGip->aiCpuFromApicId[i] = 0;
5846	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx); i++)
5847	pGip->aiCpuFromCpuSetIdx[i] = UINT16_MAX;
5848
5849	for (i = 0; i < cCpus; i++)
5850	supdrvGipInitCpu(pGip, &pGip->aCPUs[i], u64NanoTS);
5851
5852	/*
5853	* Link it to the device extension.
5854	*/
5855	pDevExt->pGip = pGip;
5856	pDevExt->HCPhysGip = HCPhys;
5857	pDevExt->cGipUsers = 0;
5858	}
5859
5860
5861	/**
5862	* On CPU initialization callback for RTMpOnAll.
5863	*
5864	* @param idCpu The CPU ID.
5865	* @param pvUser1 The device extension.
5866	* @param pvUser2 The GIP.
5867	*/
5868	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2)
5869	{
5870	/* This is good enough, even though it will update some of the globals a
5871	bit to much. */
5872	supdrvGipMpEventOnline((PSUPDRVDEVEXT)pvUser1, idCpu);
5873	}
5874
5875
5876	/**
5877	* Invalidates the GIP data upon termination.
5878	*
5879	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5880	*/
5881	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
5882	{
5883	unsigned i;
5884	pGip->u32Magic = 0;
5885	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
5886	{
5887	pGip->aCPUs[i].u64NanoTS = 0;
5888	pGip->aCPUs[i].u64TSC = 0;
5889	pGip->aCPUs[i].iTSCHistoryHead = 0;
5890	}
5891	}
5892
5893
5894	/**
5895	* Worker routine for supdrvGipUpdate and supdrvGipUpdatePerCpu that
5896	* updates all the per cpu data except the transaction id.
5897	*
5898	* @param pDevExt The device extension.
5899	* @param pGipCpu Pointer to the per cpu data.
5900	* @param u64NanoTS The current time stamp.
5901	* @param u64TSC The current TSC.
5902	* @param iTick The current timer tick.
5903	*/
5904	static void supdrvGipDoUpdateCpu(PSUPDRVDEVEXT pDevExt, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick)
5905	{
5906	uint64_t u64TSCDelta;
5907	uint32_t u32UpdateIntervalTSC;
5908	uint32_t u32UpdateIntervalTSCSlack;
5909	unsigned iTSCHistoryHead;
5910	uint64_t u64CpuHz;
5911	uint32_t u32TransactionId;
5912
5913	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5914	AssertPtrReturnVoid(pGip);
5915
5916	/* Delta between this and the previous update. */
5917	ASMAtomicUoWriteU32(&pGipCpu->u32PrevUpdateIntervalNS, (uint32_t)(u64NanoTS - pGipCpu->u64NanoTS));
5918
5919	/*
5920	* Update the NanoTS.
5921	*/
5922	ASMAtomicWriteU64(&pGipCpu->u64NanoTS, u64NanoTS);
5923
5924	/*
5925	* Calc TSC delta.
5926	*/
5927	/** @todo validate the NanoTS delta, don't trust the OS to call us when it should... */
5928	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
5929	ASMAtomicWriteU64(&pGipCpu->u64TSC, u64TSC);
5930
5931	if (u64TSCDelta >> 32)
5932	{
5933	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
5934	pGipCpu->cErrors++;
5935	}
5936
5937	/*
5938	* On the 2nd and 3rd callout, reset the history with the current TSC
5939	* interval since the values entered by supdrvGipInit are totally off.
5940	* The interval on the 1st callout completely unreliable, the 2nd is a bit
5941	* better, while the 3rd should be most reliable.
5942	*/
5943	u32TransactionId = pGipCpu->u32TransactionId;
5944	if (RT_UNLIKELY( ( u32TransactionId == 5
5945	\|\| u32TransactionId == 7)
5946	&& ( iTick == 2
5947	\|\| iTick == 3) ))
5948	{
5949	unsigned i;
5950	for (i = 0; i < RT_ELEMENTS(pGipCpu->au32TSCHistory); i++)
5951	ASMAtomicUoWriteU32(&pGipCpu->au32TSCHistory[i], (uint32_t)u64TSCDelta);
5952	}
5953
5954	/*
5955	* TSC History.
5956	*/
5957	Assert(RT_ELEMENTS(pGipCpu->au32TSCHistory) == 8);
5958	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
5959	ASMAtomicWriteU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
5960	ASMAtomicWriteU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
5961
5962	/*
5963	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
5964	*/
5965	if (pGip->u32UpdateHz >= 1000)
5966	{
5967	uint32_t u32;
5968	u32 = pGipCpu->au32TSCHistory[0];
5969	u32 += pGipCpu->au32TSCHistory[1];
5970	u32 += pGipCpu->au32TSCHistory[2];
5971	u32 += pGipCpu->au32TSCHistory[3];
5972	u32 >>= 2;
5973	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
5974	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
5975	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
5976	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
5977	u32UpdateIntervalTSC >>= 2;
5978	u32UpdateIntervalTSC += u32;
5979	u32UpdateIntervalTSC >>= 1;
5980
5981	/* Value chosen for a 2GHz Athlon64 running linux 2.6.10/11, . */
5982	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
5983	}
5984	else if (pGip->u32UpdateHz >= 90)
5985	{
5986	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5987	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
5988	u32UpdateIntervalTSC >>= 1;
5989
5990	/* value chosen on a 2GHz thinkpad running windows */
5991	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
5992	}
5993	else
5994	{
5995	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5996
5997	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
5998	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
5999	}
6000	ASMAtomicWriteU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
6001
6002	/*
6003	* CpuHz.
6004	*/
6005	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, pGip->u32UpdateHz);
6006	ASMAtomicWriteU64(&pGipCpu->u64CpuHz, u64CpuHz);
6007	}
6008
6009
6010	/**
6011	* Updates the GIP.
6012	*
6013	* @param pDevExt The device extension.
6014	* @param u64NanoTS The current nanosecond timesamp.
6015	* @param u64TSC The current TSC timesamp.
6016	* @param idCpu The CPU ID.
6017	* @param iTick The current timer tick.
6018	*/
6019	static void supdrvGipUpdate(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick)
6020	{
6021	/*
6022	* Determine the relevant CPU data.
6023	*/
6024	PSUPGIPCPU pGipCpu;
6025	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
6026	AssertPtrReturnVoid(pGip);
6027
6028	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
6029	pGipCpu = &pGip->aCPUs[0];
6030	else
6031	{
6032	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
6033	if (RT_UNLIKELY(iCpu >= pGip->cCpus))
6034	return;
6035	pGipCpu = &pGip->aCPUs[iCpu];
6036	if (RT_UNLIKELY(pGipCpu->idCpu != idCpu))
6037	return;
6038	}
6039
6040	/*
6041	* Start update transaction.
6042	*/
6043	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
6044	{
6045	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
6046	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
6047	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
6048	pGipCpu->cErrors++;
6049	return;
6050	}
6051
6052	/*
6053	* Recalc the update frequency every 0x800th time.
6054	*/
6055	if (!(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
6056	{
6057	if (pGip->u64NanoTSLastUpdateHz)
6058	{
6059	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
6060	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
6061	uint32_t u32UpdateHz = (uint32_t)((UINT64_C(1000000000) * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
6062	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
6063	{
6064	ASMAtomicWriteU32(&pGip->u32UpdateHz, u32UpdateHz);
6065	ASMAtomicWriteU32(&pGip->u32UpdateIntervalNS, 1000000000 / u32UpdateHz);
6066	}
6067	#endif
6068	}
6069	ASMAtomicWriteU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS);
6070	}
6071
6072	/*
6073	* Update the data.
6074	*/
6075	supdrvGipDoUpdateCpu(pDevExt, pGipCpu, u64NanoTS, u64TSC, iTick);
6076
6077	/*
6078	* Complete transaction.
6079	*/
6080	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
6081	}
6082
6083
6084	/**
6085	* Updates the per cpu GIP data for the calling cpu.
6086	*
6087	* @param pDevExt The device extension.
6088	* @param u64NanoTS The current nanosecond timesamp.
6089	* @param u64TSC The current TSC timesamp.
6090	* @param idCpu The CPU ID.
6091	* @param idApic The APIC id for the CPU index.
6092	* @param iTick The current timer tick.
6093	*/
6094	static void supdrvGipUpdatePerCpu(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC,
6095	RTCPUID idCpu, uint8_t idApic, uint64_t iTick)
6096	{
6097	uint32_t iCpu;
6098	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
6099
6100	/*
6101	* Avoid a potential race when a CPU online notification doesn't fire on
6102	* the onlined CPU but the tick creeps in before the event notification is
6103	* run.
6104	*/
6105	if (RT_UNLIKELY(iTick == 1))
6106	{
6107	iCpu = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
6108	if (pGip->aCPUs[iCpu].enmState == SUPGIPCPUSTATE_OFFLINE)
6109	supdrvGipMpEventOnline(pDevExt, idCpu);
6110	}
6111
6112	iCpu = pGip->aiCpuFromApicId[idApic];
6113	if (RT_LIKELY(iCpu < pGip->cCpus))
6114	{
6115	PSUPGIPCPU pGipCpu = &pGip->aCPUs[iCpu];
6116	if (pGipCpu->idCpu == idCpu)
6117	{
6118	/*
6119	* Start update transaction.
6120	*/
6121	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
6122	{
6123	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
6124	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
6125	pGipCpu->cErrors++;
6126	return;
6127	}
6128
6129	/*
6130	* Update the data.
6131	*/
6132	supdrvGipDoUpdateCpu(pDevExt, pGipCpu, u64NanoTS, u64TSC, iTick);
6133
6134	/*
6135	* Complete transaction.
6136	*/
6137	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
6138	}
6139	}
6140	}
6141

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

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDrv.c@ 49786

Download in other formats: