SUPDrv.c@ 47363

Last change on this file since 47363 was 47352, checked in by vboxsync, 11 years ago
Runtime/r0drv: Reference counting for kernel thread-context hooks.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 223.1 KB

Line
1	/* $Id: SUPDrv.c 47352 2013-07-23 16:19:20Z vboxsync $ */
2	/** @file
3	* VBoxDrv - The VirtualBox Support Driver - Common code.
4	*/
5
6	/*
7	* Copyright (C) 2006-2013 Oracle Corporation
8	*
9	* This file is part of VirtualBox Open Source Edition (OSE), as
10	* available from http://www.virtualbox.org. This file is free software;
11	* you can redistribute it and/or modify it under the terms of the GNU
12	* General Public License (GPL) as published by the Free Software
13	* Foundation, in version 2 as it comes in the "COPYING" file of the
14	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16	*
17	* The contents of this file may alternatively be used under the terms
18	* of the Common Development and Distribution License Version 1.0
19	* (CDDL) only, as it comes in the "COPYING.CDDL" file of the
20	* VirtualBox OSE distribution, in which case the provisions of the
21	* CDDL are applicable instead of those of the GPL.
22	*
23	* You may elect to license modified versions of this file under the
24	* terms and conditions of either the GPL or the CDDL or both.
25	*/
26
27	/*******************************************************************************
28	* Header Files *
29	*******************************************************************************/
30	#define LOG_GROUP LOG_GROUP_SUP_DRV
31	#define SUPDRV_AGNOSTIC
32	#include "SUPDrvInternal.h"
33	#ifndef PAGE_SHIFT
34	# include <iprt/param.h>
35	#endif
36	#include <iprt/asm.h>
37	#include <iprt/asm-amd64-x86.h>
38	#include <iprt/asm-math.h>
39	#include <iprt/cpuset.h>
40	#include <iprt/handletable.h>
41	#include <iprt/mem.h>
42	#include <iprt/mp.h>
43	#include <iprt/power.h>
44	#include <iprt/process.h>
45	#include <iprt/semaphore.h>
46	#include <iprt/spinlock.h>
47	#include <iprt/thread.h>
48	#include <iprt/uuid.h>
49	#include <iprt/net.h>
50	#include <iprt/crc.h>
51	#include <iprt/string.h>
52	#include <iprt/timer.h>
53	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
54	# include <iprt/rand.h>
55	# include <iprt/path.h>
56	#endif
57	#include <iprt/x86.h>
58
59	#include <VBox/param.h>
60	#include <VBox/log.h>
61	#include <VBox/err.h>
62	#include <VBox/vmm/hm_svm.h>
63	#include <VBox/vmm/hm_vmx.h>
64
65	#if defined(RT_OS_SOLARIS) \|\| defined(RT_OS_DARWIN)
66	# include "dtrace/SUPDrv.h"
67	#else
68	# define VBOXDRV_SESSION_CREATE(pvSession, fUser) do { } while (0)
69	# define VBOXDRV_SESSION_CLOSE(pvSession) do { } while (0)
70	# define VBOXDRV_IOCTL_ENTRY(pvSession, uIOCtl, pvReqHdr) do { } while (0)
71	# define VBOXDRV_IOCTL_RETURN(pvSession, uIOCtl, pvReqHdr, rcRet, rcReq) do { } while (0)
72	#endif
73
74	/*
75	* Logging assignments:
76	* Log - useful stuff, like failures.
77	* LogFlow - program flow, except the really noisy bits.
78	* Log2 - Cleanup.
79	* Log3 - Loader flow noise.
80	* Log4 - Call VMMR0 flow noise.
81	* Log5 - Native yet-to-be-defined noise.
82	* Log6 - Native ioctl flow noise.
83	*
84	* Logging requires BUILD_TYPE=debug and possibly changes to the logger
85	* instantiation in log-vbox.c(pp).
86	*/
87
88
89	/*******************************************************************************
90	* Defined Constants And Macros *
91	*******************************************************************************/
92	/** The frequency by which we recalculate the u32UpdateHz and
93	* u32UpdateIntervalNS GIP members. The value must be a power of 2. */
94	#define GIP_UPDATEHZ_RECALC_FREQ 0x800
95
96	/** @def VBOX_SVN_REV
97	* The makefile should define this if it can. */
98	#ifndef VBOX_SVN_REV
99	# define VBOX_SVN_REV 0
100	#endif
101
102	#if 0 /* Don't start the GIP timers. Useful when debugging the IPRT timer code. */
103	# define DO_NOT_START_GIP
104	#endif
105
106
107	/*******************************************************************************
108	* Internal Functions *
109	*******************************************************************************/
110	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser);
111	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser);
112	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession);
113	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType);
114	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq);
115	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq);
116	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq);
117	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq);
118	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq);
119	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx);
120	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt);
121	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage);
122	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage);
123	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt);
124	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt);
125	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq);
126	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq);
127	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt);
128	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt);
129	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
130	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
131	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser);
132	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
133	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned cCpus);
134	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2);
135	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip);
136	static void supdrvGipUpdate(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick);
137	static void supdrvGipUpdatePerCpu(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC,
138	RTCPUID idCpu, uint8_t idApic, uint64_t iTick);
139	static void supdrvGipInitCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pCpu, uint64_t u64NanoTS);
140
141
142	/*******************************************************************************
143	* Global Variables *
144	*******************************************************************************/
145	DECLEXPORT(PSUPGLOBALINFOPAGE) g_pSUPGlobalInfoPage = NULL;
146
147	/**
148	* Array of the R0 SUP API.
149	*/
150	static SUPFUNC g_aFunctions[] =
151	{
152	/* SED: START */
153	/* name function */
154	/* Entries with absolute addresses determined at runtime, fixup
155	code makes ugly ASSUMPTIONS about the order here: */
156	{ "SUPR0AbsIs64bit", (void *)0 },
157	{ "SUPR0Abs64bitKernelCS", (void *)0 },
158	{ "SUPR0Abs64bitKernelSS", (void *)0 },
159	{ "SUPR0Abs64bitKernelDS", (void *)0 },
160	{ "SUPR0AbsKernelCS", (void *)0 },
161	{ "SUPR0AbsKernelSS", (void *)0 },
162	{ "SUPR0AbsKernelDS", (void *)0 },
163	{ "SUPR0AbsKernelES", (void *)0 },
164	{ "SUPR0AbsKernelFS", (void *)0 },
165	{ "SUPR0AbsKernelGS", (void *)0 },
166	/* Normal function pointers: */
167	{ "g_pSUPGlobalInfoPage", (void )&g_pSUPGlobalInfoPage }, / SED: DATA */
168	{ "SUPGetGIP", (void *)SUPGetGIP },
169	{ "SUPR0ComponentDeregisterFactory", (void *)SUPR0ComponentDeregisterFactory },
170	{ "SUPR0ComponentQueryFactory", (void *)SUPR0ComponentQueryFactory },
171	{ "SUPR0ComponentRegisterFactory", (void *)SUPR0ComponentRegisterFactory },
172	{ "SUPR0ContAlloc", (void *)SUPR0ContAlloc },
173	{ "SUPR0ContFree", (void *)SUPR0ContFree },
174	{ "SUPR0EnableVTx", (void *)SUPR0EnableVTx },
175	{ "SUPR0SuspendVTxOnCpu", (void *)SUPR0SuspendVTxOnCpu },
176	{ "SUPR0ResumeVTxOnCpu", (void *)SUPR0ResumeVTxOnCpu },
177	{ "SUPR0GetPagingMode", (void *)SUPR0GetPagingMode },
178	{ "SUPR0LockMem", (void *)SUPR0LockMem },
179	{ "SUPR0LowAlloc", (void *)SUPR0LowAlloc },
180	{ "SUPR0LowFree", (void *)SUPR0LowFree },
181	{ "SUPR0MemAlloc", (void *)SUPR0MemAlloc },
182	{ "SUPR0MemFree", (void *)SUPR0MemFree },
183	{ "SUPR0MemGetPhys", (void *)SUPR0MemGetPhys },
184	{ "SUPR0ObjAddRef", (void *)SUPR0ObjAddRef },
185	{ "SUPR0ObjAddRefEx", (void *)SUPR0ObjAddRefEx },
186	{ "SUPR0ObjRegister", (void *)SUPR0ObjRegister },
187	{ "SUPR0ObjRelease", (void *)SUPR0ObjRelease },
188	{ "SUPR0ObjVerifyAccess", (void *)SUPR0ObjVerifyAccess },
189	{ "SUPR0PageAllocEx", (void *)SUPR0PageAllocEx },
190	{ "SUPR0PageFree", (void *)SUPR0PageFree },
191	{ "SUPR0Printf", (void *)SUPR0Printf },
192	{ "SUPR0TracerDeregisterDrv", (void *)SUPR0TracerDeregisterDrv },
193	{ "SUPR0TracerDeregisterImpl", (void *)SUPR0TracerDeregisterImpl },
194	{ "SUPR0TracerFireProbe", (void *)SUPR0TracerFireProbe },
195	{ "SUPR0TracerRegisterDrv", (void *)SUPR0TracerRegisterDrv },
196	{ "SUPR0TracerRegisterImpl", (void *)SUPR0TracerRegisterImpl },
197	{ "SUPR0TracerRegisterModule", (void *)SUPR0TracerRegisterModule },
198	{ "SUPR0TracerUmodProbeFire", (void *)SUPR0TracerUmodProbeFire },
199	{ "SUPR0UnlockMem", (void *)SUPR0UnlockMem },
200	{ "SUPSemEventClose", (void *)SUPSemEventClose },
201	{ "SUPSemEventCreate", (void *)SUPSemEventCreate },
202	{ "SUPSemEventGetResolution", (void *)SUPSemEventGetResolution },
203	{ "SUPSemEventMultiClose", (void *)SUPSemEventMultiClose },
204	{ "SUPSemEventMultiCreate", (void *)SUPSemEventMultiCreate },
205	{ "SUPSemEventMultiGetResolution", (void *)SUPSemEventMultiGetResolution },
206	{ "SUPSemEventMultiReset", (void *)SUPSemEventMultiReset },
207	{ "SUPSemEventMultiSignal", (void *)SUPSemEventMultiSignal },
208	{ "SUPSemEventMultiWait", (void *)SUPSemEventMultiWait },
209	{ "SUPSemEventMultiWaitNoResume", (void *)SUPSemEventMultiWaitNoResume },
210	{ "SUPSemEventMultiWaitNsAbsIntr", (void *)SUPSemEventMultiWaitNsAbsIntr },
211	{ "SUPSemEventMultiWaitNsRelIntr", (void *)SUPSemEventMultiWaitNsRelIntr },
212	{ "SUPSemEventSignal", (void *)SUPSemEventSignal },
213	{ "SUPSemEventWait", (void *)SUPSemEventWait },
214	{ "SUPSemEventWaitNoResume", (void *)SUPSemEventWaitNoResume },
215	{ "SUPSemEventWaitNsAbsIntr", (void *)SUPSemEventWaitNsAbsIntr },
216	{ "SUPSemEventWaitNsRelIntr", (void *)SUPSemEventWaitNsRelIntr },
217
218	{ "RTAssertAreQuiet", (void *)RTAssertAreQuiet },
219	{ "RTAssertMayPanic", (void *)RTAssertMayPanic },
220	{ "RTAssertMsg1", (void *)RTAssertMsg1 },
221	{ "RTAssertMsg2AddV", (void *)RTAssertMsg2AddV },
222	{ "RTAssertMsg2V", (void *)RTAssertMsg2V },
223	{ "RTAssertSetMayPanic", (void *)RTAssertSetMayPanic },
224	{ "RTAssertSetQuiet", (void *)RTAssertSetQuiet },
225	{ "RTCrc32", (void *)RTCrc32 },
226	{ "RTCrc32Finish", (void *)RTCrc32Finish },
227	{ "RTCrc32Process", (void *)RTCrc32Process },
228	{ "RTCrc32Start", (void *)RTCrc32Start },
229	{ "RTErrConvertFromErrno", (void *)RTErrConvertFromErrno },
230	{ "RTErrConvertToErrno", (void *)RTErrConvertToErrno },
231	{ "RTHandleTableAllocWithCtx", (void *)RTHandleTableAllocWithCtx },
232	{ "RTHandleTableCreate", (void *)RTHandleTableCreate },
233	{ "RTHandleTableCreateEx", (void *)RTHandleTableCreateEx },
234	{ "RTHandleTableDestroy", (void *)RTHandleTableDestroy },
235	{ "RTHandleTableFreeWithCtx", (void *)RTHandleTableFreeWithCtx },
236	{ "RTHandleTableLookupWithCtx", (void *)RTHandleTableLookupWithCtx },
237	{ "RTLogDefaultInstance", (void *)RTLogDefaultInstance },
238	{ "RTLogGetDefaultInstance", (void *)RTLogGetDefaultInstance },
239	{ "RTLogLoggerExV", (void *)RTLogLoggerExV },
240	{ "RTLogPrintfV", (void *)RTLogPrintfV },
241	{ "RTLogRelDefaultInstance", (void *)RTLogRelDefaultInstance },
242	{ "RTLogSetDefaultInstanceThread", (void *)RTLogSetDefaultInstanceThread },
243	{ "RTMemAllocExTag", (void *)RTMemAllocExTag },
244	{ "RTMemAllocTag", (void *)RTMemAllocTag },
245	{ "RTMemAllocVarTag", (void *)RTMemAllocVarTag },
246	{ "RTMemAllocZTag", (void *)RTMemAllocZTag },
247	{ "RTMemAllocZVarTag", (void *)RTMemAllocZVarTag },
248	{ "RTMemDupExTag", (void *)RTMemDupExTag },
249	{ "RTMemDupTag", (void *)RTMemDupTag },
250	{ "RTMemFree", (void *)RTMemFree },
251	{ "RTMemFreeEx", (void *)RTMemFreeEx },
252	{ "RTMemReallocTag", (void *)RTMemReallocTag },
253	{ "RTMpCpuId", (void *)RTMpCpuId },
254	{ "RTMpCpuIdFromSetIndex", (void *)RTMpCpuIdFromSetIndex },
255	{ "RTMpCpuIdToSetIndex", (void *)RTMpCpuIdToSetIndex },
256	{ "RTMpGetArraySize", (void *)RTMpGetArraySize },
257	{ "RTMpGetCount", (void *)RTMpGetCount },
258	{ "RTMpGetMaxCpuId", (void *)RTMpGetMaxCpuId },
259	{ "RTMpGetOnlineCount", (void *)RTMpGetOnlineCount },
260	{ "RTMpGetOnlineSet", (void *)RTMpGetOnlineSet },
261	{ "RTMpGetSet", (void *)RTMpGetSet },
262	{ "RTMpIsCpuOnline", (void *)RTMpIsCpuOnline },
263	{ "RTMpIsCpuPossible", (void *)RTMpIsCpuPossible },
264	{ "RTMpIsCpuWorkPending", (void *)RTMpIsCpuWorkPending },
265	{ "RTMpNotificationDeregister", (void *)RTMpNotificationDeregister },
266	{ "RTMpNotificationRegister", (void *)RTMpNotificationRegister },
267	{ "RTMpOnAll", (void *)RTMpOnAll },
268	{ "RTMpOnOthers", (void *)RTMpOnOthers },
269	{ "RTMpOnSpecific", (void *)RTMpOnSpecific },
270	{ "RTMpPokeCpu", (void *)RTMpPokeCpu },
271	{ "RTNetIPv4AddDataChecksum", (void *)RTNetIPv4AddDataChecksum },
272	{ "RTNetIPv4AddTCPChecksum", (void *)RTNetIPv4AddTCPChecksum },
273	{ "RTNetIPv4AddUDPChecksum", (void *)RTNetIPv4AddUDPChecksum },
274	{ "RTNetIPv4FinalizeChecksum", (void *)RTNetIPv4FinalizeChecksum },
275	{ "RTNetIPv4HdrChecksum", (void *)RTNetIPv4HdrChecksum },
276	{ "RTNetIPv4IsDHCPValid", (void *)RTNetIPv4IsDHCPValid },
277	{ "RTNetIPv4IsHdrValid", (void *)RTNetIPv4IsHdrValid },
278	{ "RTNetIPv4IsTCPSizeValid", (void *)RTNetIPv4IsTCPSizeValid },
279	{ "RTNetIPv4IsTCPValid", (void *)RTNetIPv4IsTCPValid },
280	{ "RTNetIPv4IsUDPSizeValid", (void *)RTNetIPv4IsUDPSizeValid },
281	{ "RTNetIPv4IsUDPValid", (void *)RTNetIPv4IsUDPValid },
282	{ "RTNetIPv4PseudoChecksum", (void *)RTNetIPv4PseudoChecksum },
283	{ "RTNetIPv4PseudoChecksumBits", (void *)RTNetIPv4PseudoChecksumBits },
284	{ "RTNetIPv4TCPChecksum", (void *)RTNetIPv4TCPChecksum },
285	{ "RTNetIPv4UDPChecksum", (void *)RTNetIPv4UDPChecksum },
286	{ "RTNetIPv6PseudoChecksum", (void *)RTNetIPv6PseudoChecksum },
287	{ "RTNetIPv6PseudoChecksumBits", (void *)RTNetIPv6PseudoChecksumBits },
288	{ "RTNetIPv6PseudoChecksumEx", (void *)RTNetIPv6PseudoChecksumEx },
289	{ "RTNetTCPChecksum", (void *)RTNetTCPChecksum },
290	{ "RTNetUDPChecksum", (void *)RTNetUDPChecksum },
291	{ "RTPowerNotificationDeregister", (void *)RTPowerNotificationDeregister },
292	{ "RTPowerNotificationRegister", (void *)RTPowerNotificationRegister },
293	{ "RTProcSelf", (void *)RTProcSelf },
294	{ "RTR0AssertPanicSystem", (void *)RTR0AssertPanicSystem },
295	{ "RTR0MemAreKrnlAndUsrDifferent", (void *)RTR0MemAreKrnlAndUsrDifferent },
296	{ "RTR0MemKernelIsValidAddr", (void *)RTR0MemKernelIsValidAddr },
297	{ "RTR0MemKernelCopyFrom", (void *)RTR0MemKernelCopyFrom },
298	{ "RTR0MemKernelCopyTo", (void *)RTR0MemKernelCopyTo },
299	{ "RTR0MemObjAddress", (void *)RTR0MemObjAddress },
300	{ "RTR0MemObjAddressR3", (void *)RTR0MemObjAddressR3 },
301	{ "RTR0MemObjAllocContTag", (void *)RTR0MemObjAllocContTag },
302	{ "RTR0MemObjAllocLowTag", (void *)RTR0MemObjAllocLowTag },
303	{ "RTR0MemObjAllocPageTag", (void *)RTR0MemObjAllocPageTag },
304	{ "RTR0MemObjAllocPhysExTag", (void *)RTR0MemObjAllocPhysExTag },
305	{ "RTR0MemObjAllocPhysNCTag", (void *)RTR0MemObjAllocPhysNCTag },
306	{ "RTR0MemObjAllocPhysTag", (void *)RTR0MemObjAllocPhysTag },
307	{ "RTR0MemObjEnterPhysTag", (void *)RTR0MemObjEnterPhysTag },
308	{ "RTR0MemObjFree", (void *)RTR0MemObjFree },
309	{ "RTR0MemObjGetPagePhysAddr", (void *)RTR0MemObjGetPagePhysAddr },
310	{ "RTR0MemObjIsMapping", (void *)RTR0MemObjIsMapping },
311	{ "RTR0MemObjLockUserTag", (void *)RTR0MemObjLockUserTag },
312	{ "RTR0MemObjMapKernelExTag", (void *)RTR0MemObjMapKernelExTag },
313	{ "RTR0MemObjMapKernelTag", (void *)RTR0MemObjMapKernelTag },
314	{ "RTR0MemObjMapUserTag", (void *)RTR0MemObjMapUserTag },
315	{ "RTR0MemObjProtect", (void *)RTR0MemObjProtect },
316	{ "RTR0MemObjSize", (void *)RTR0MemObjSize },
317	{ "RTR0MemUserCopyFrom", (void *)RTR0MemUserCopyFrom },
318	{ "RTR0MemUserCopyTo", (void *)RTR0MemUserCopyTo },
319	{ "RTR0MemUserIsValidAddr", (void *)RTR0MemUserIsValidAddr },
320	{ "RTR0ProcHandleSelf", (void *)RTR0ProcHandleSelf },
321	{ "RTSemEventCreate", (void *)RTSemEventCreate },
322	{ "RTSemEventDestroy", (void *)RTSemEventDestroy },
323	{ "RTSemEventGetResolution", (void *)RTSemEventGetResolution },
324	{ "RTSemEventMultiCreate", (void *)RTSemEventMultiCreate },
325	{ "RTSemEventMultiDestroy", (void *)RTSemEventMultiDestroy },
326	{ "RTSemEventMultiGetResolution", (void *)RTSemEventMultiGetResolution },
327	{ "RTSemEventMultiReset", (void *)RTSemEventMultiReset },
328	{ "RTSemEventMultiSignal", (void *)RTSemEventMultiSignal },
329	{ "RTSemEventMultiWait", (void *)RTSemEventMultiWait },
330	{ "RTSemEventMultiWaitEx", (void *)RTSemEventMultiWaitEx },
331	{ "RTSemEventMultiWaitExDebug", (void *)RTSemEventMultiWaitExDebug },
332	{ "RTSemEventMultiWaitNoResume", (void *)RTSemEventMultiWaitNoResume },
333	{ "RTSemEventSignal", (void *)RTSemEventSignal },
334	{ "RTSemEventWait", (void *)RTSemEventWait },
335	{ "RTSemEventWaitEx", (void *)RTSemEventWaitEx },
336	{ "RTSemEventWaitExDebug", (void *)RTSemEventWaitExDebug },
337	{ "RTSemEventWaitNoResume", (void *)RTSemEventWaitNoResume },
338	{ "RTSemFastMutexCreate", (void *)RTSemFastMutexCreate },
339	{ "RTSemFastMutexDestroy", (void *)RTSemFastMutexDestroy },
340	{ "RTSemFastMutexRelease", (void *)RTSemFastMutexRelease },
341	{ "RTSemFastMutexRequest", (void *)RTSemFastMutexRequest },
342	{ "RTSemMutexCreate", (void *)RTSemMutexCreate },
343	{ "RTSemMutexDestroy", (void *)RTSemMutexDestroy },
344	{ "RTSemMutexRelease", (void *)RTSemMutexRelease },
345	{ "RTSemMutexRequest", (void *)RTSemMutexRequest },
346	{ "RTSemMutexRequestDebug", (void *)RTSemMutexRequestDebug },
347	{ "RTSemMutexRequestNoResume", (void *)RTSemMutexRequestNoResume },
348	{ "RTSemMutexRequestNoResumeDebug", (void *)RTSemMutexRequestNoResumeDebug },
349	{ "RTSpinlockAcquire", (void *)RTSpinlockAcquire },
350	{ "RTSpinlockCreate", (void *)RTSpinlockCreate },
351	{ "RTSpinlockDestroy", (void *)RTSpinlockDestroy },
352	{ "RTSpinlockRelease", (void *)RTSpinlockRelease },
353	{ "RTSpinlockReleaseNoInts", (void *)RTSpinlockReleaseNoInts },
354	{ "RTStrCopy", (void *)RTStrCopy },
355	{ "RTStrDupTag", (void *)RTStrDupTag },
356	{ "RTStrFormat", (void *)RTStrFormat },
357	{ "RTStrFormatNumber", (void *)RTStrFormatNumber },
358	{ "RTStrFormatTypeDeregister", (void *)RTStrFormatTypeDeregister },
359	{ "RTStrFormatTypeRegister", (void *)RTStrFormatTypeRegister },
360	{ "RTStrFormatTypeSetUser", (void *)RTStrFormatTypeSetUser },
361	{ "RTStrFormatV", (void *)RTStrFormatV },
362	{ "RTStrFree", (void *)RTStrFree },
363	{ "RTStrNCmp", (void *)RTStrNCmp },
364	{ "RTStrPrintf", (void *)RTStrPrintf },
365	{ "RTStrPrintfEx", (void *)RTStrPrintfEx },
366	{ "RTStrPrintfExV", (void *)RTStrPrintfExV },
367	{ "RTStrPrintfV", (void *)RTStrPrintfV },
368	{ "RTThreadCreate", (void *)RTThreadCreate },
369	{ "RTThreadCtxHooksCreate", (void *)RTThreadCtxHooksCreate },
370	{ "RTThreadCtxHooksDeregister", (void *)RTThreadCtxHooksDeregister },
371	{ "RTThreadCtxHooksRegister", (void *)RTThreadCtxHooksRegister },
372	{ "RTThreadCtxHooksRelease", (void *)RTThreadCtxHooksRelease },
373	{ "RTThreadCtxHooksRetain", (void *)RTThreadCtxHooksRetain },
374	{ "RTThreadGetName", (void *)RTThreadGetName },
375	{ "RTThreadGetNative", (void *)RTThreadGetNative },
376	{ "RTThreadGetType", (void *)RTThreadGetType },
377	{ "RTThreadIsInInterrupt", (void *)RTThreadIsInInterrupt },
378	{ "RTThreadNativeSelf", (void *)RTThreadNativeSelf },
379	{ "RTThreadPreemptDisable", (void *)RTThreadPreemptDisable },
380	{ "RTThreadPreemptIsEnabled", (void *)RTThreadPreemptIsEnabled },
381	{ "RTThreadPreemptIsPending", (void *)RTThreadPreemptIsPending },
382	{ "RTThreadPreemptIsPendingTrusty", (void *)RTThreadPreemptIsPendingTrusty },
383	{ "RTThreadPreemptIsPossible", (void *)RTThreadPreemptIsPossible },
384	{ "RTThreadPreemptRestore", (void *)RTThreadPreemptRestore },
385	{ "RTThreadSelf", (void *)RTThreadSelf },
386	{ "RTThreadSelfName", (void *)RTThreadSelfName },
387	{ "RTThreadSleep", (void *)RTThreadSleep },
388	{ "RTThreadUserReset", (void *)RTThreadUserReset },
389	{ "RTThreadUserSignal", (void *)RTThreadUserSignal },
390	{ "RTThreadUserWait", (void *)RTThreadUserWait },
391	{ "RTThreadUserWaitNoResume", (void *)RTThreadUserWaitNoResume },
392	{ "RTThreadWait", (void *)RTThreadWait },
393	{ "RTThreadWaitNoResume", (void *)RTThreadWaitNoResume },
394	{ "RTThreadYield", (void *)RTThreadYield },
395	{ "RTTimeMilliTS", (void *)RTTimeMilliTS },
396	{ "RTTimeNanoTS", (void *)RTTimeNanoTS },
397	{ "RTTimeNow", (void *)RTTimeNow },
398	{ "RTTimerCanDoHighResolution", (void *)RTTimerCanDoHighResolution },
399	{ "RTTimerChangeInterval", (void *)RTTimerChangeInterval },
400	{ "RTTimerCreate", (void *)RTTimerCreate },
401	{ "RTTimerCreateEx", (void *)RTTimerCreateEx },
402	{ "RTTimerDestroy", (void *)RTTimerDestroy },
403	{ "RTTimerGetSystemGranularity", (void *)RTTimerGetSystemGranularity },
404	{ "RTTimerReleaseSystemGranularity", (void *)RTTimerReleaseSystemGranularity },
405	{ "RTTimerRequestSystemGranularity", (void *)RTTimerRequestSystemGranularity },
406	{ "RTTimerStart", (void *)RTTimerStart },
407	{ "RTTimerStop", (void *)RTTimerStop },
408	{ "RTTimeSystemMilliTS", (void *)RTTimeSystemMilliTS },
409	{ "RTTimeSystemNanoTS", (void *)RTTimeSystemNanoTS },
410	{ "RTUuidCompare", (void *)RTUuidCompare },
411	{ "RTUuidCompareStr", (void *)RTUuidCompareStr },
412	{ "RTUuidFromStr", (void *)RTUuidFromStr },
413	/* SED: END */
414	};
415
416	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
417	/**
418	* Drag in the rest of IRPT since we share it with the
419	* rest of the kernel modules on darwin.
420	*/
421	PFNRT g_apfnVBoxDrvIPRTDeps[] =
422	{
423	/* VBoxNetAdp */
424	(PFNRT)RTRandBytes,
425	/* VBoxUSB */
426	(PFNRT)RTPathStripFilename,
427	NULL
428	};
429	#endif /* RT_OS_DARWIN \|\| RT_OS_SOLARIS \|\| RT_OS_SOLARIS */
430
431
432	/**
433	* Initializes the device extentsion structure.
434	*
435	* @returns IPRT status code.
436	* @param pDevExt The device extension to initialize.
437	* @param cbSession The size of the session structure. The size of
438	* SUPDRVSESSION may be smaller when SUPDRV_AGNOSTIC is
439	* defined because we're skipping the OS specific members
440	* then.
441	*/
442	int VBOXCALL supdrvInitDevExt(PSUPDRVDEVEXT pDevExt, size_t cbSession)
443	{
444	int rc;
445
446	#ifdef SUPDRV_WITH_RELEASE_LOGGER
447	/*
448	* Create the release log.
449	*/
450	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
451	PRTLOGGER pRelLogger;
452	rc = RTLogCreate(&pRelLogger, 0 /* fFlags */, "all",
453	"VBOX_RELEASE_LOG", RT_ELEMENTS(s_apszGroups), s_apszGroups, RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER, NULL);
454	if (RT_SUCCESS(rc))
455	RTLogRelSetDefaultInstance(pRelLogger);
456	/** @todo Add native hook for getting logger config parameters and setting
457	* them. On linux we should use the module parameter stuff... */
458	#endif
459
460	/*
461	* Initialize it.
462	*/
463	memset(pDevExt, 0, sizeof(*pDevExt));
464	rc = RTSpinlockCreate(&pDevExt->Spinlock, RTSPINLOCK_FLAGS_INTERRUPT_SAFE, "SUPDrvDevExt");
465	if (RT_SUCCESS(rc))
466	{
467	rc = RTSpinlockCreate(&pDevExt->hGipSpinlock, RTSPINLOCK_FLAGS_INTERRUPT_SAFE, "SUPDrvGip");
468	if (RT_SUCCESS(rc))
469	{
470	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
471	rc = RTSemMutexCreate(&pDevExt->mtxLdr);
472	#else
473	rc = RTSemFastMutexCreate(&pDevExt->mtxLdr);
474	#endif
475	if (RT_SUCCESS(rc))
476	{
477	rc = RTSemFastMutexCreate(&pDevExt->mtxComponentFactory);
478	if (RT_SUCCESS(rc))
479	{
480	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
481	rc = RTSemMutexCreate(&pDevExt->mtxGip);
482	#else
483	rc = RTSemFastMutexCreate(&pDevExt->mtxGip);
484	#endif
485	if (RT_SUCCESS(rc))
486	{
487	rc = supdrvGipCreate(pDevExt);
488	if (RT_SUCCESS(rc))
489	{
490	rc = supdrvTracerInit(pDevExt);
491	if (RT_SUCCESS(rc))
492	{
493	pDevExt->pLdrInitImage = NULL;
494	pDevExt->hLdrInitThread = NIL_RTNATIVETHREAD;
495	pDevExt->u32Cookie = BIRD; /** @todo make this random? */
496	pDevExt->cbSession = (uint32_t)cbSession;
497
498	/*
499	* Fixup the absolute symbols.
500	*
501	* Because of the table indexing assumptions we'll have a little #ifdef orgy
502	* here rather than distributing this to OS specific files. At least for now.
503	*/
504	#ifdef RT_OS_DARWIN
505	# if ARCH_BITS == 32
506	if (SUPR0GetPagingMode() >= SUPPAGINGMODE_AMD64)
507	{
508	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
509	g_aFunctions[1].pfn = (void )0x80; / SUPR0Abs64bitKernelCS - KERNEL64_CS, seg.h */
510	g_aFunctions[2].pfn = (void )0x88; / SUPR0Abs64bitKernelSS - KERNEL64_SS, seg.h */
511	g_aFunctions[3].pfn = (void )0x88; / SUPR0Abs64bitKernelDS - KERNEL64_SS, seg.h */
512	}
513	else
514	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
515	g_aFunctions[4].pfn = (void )0x08; / SUPR0AbsKernelCS - KERNEL_CS, seg.h */
516	g_aFunctions[5].pfn = (void )0x10; / SUPR0AbsKernelSS - KERNEL_DS, seg.h */
517	g_aFunctions[6].pfn = (void )0x10; / SUPR0AbsKernelDS - KERNEL_DS, seg.h */
518	g_aFunctions[7].pfn = (void )0x10; / SUPR0AbsKernelES - KERNEL_DS, seg.h */
519	g_aFunctions[8].pfn = (void )0x10; / SUPR0AbsKernelFS - KERNEL_DS, seg.h */
520	g_aFunctions[9].pfn = (void )0x48; / SUPR0AbsKernelGS - CPU_DATA_GS, seg.h */
521	# else /* 64-bit darwin: */
522	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
523	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
524	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
525	g_aFunctions[3].pfn = (void )0; / SUPR0Abs64bitKernelDS */
526	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
527	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
528	g_aFunctions[6].pfn = (void )0; / SUPR0AbsKernelDS */
529	g_aFunctions[7].pfn = (void )0; / SUPR0AbsKernelES */
530	g_aFunctions[8].pfn = (void )0; / SUPR0AbsKernelFS */
531	g_aFunctions[9].pfn = (void )0; / SUPR0AbsKernelGS */
532
533	# endif
534	#else /* !RT_OS_DARWIN */
535	# if ARCH_BITS == 64
536	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
537	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
538	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
539	g_aFunctions[3].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0Abs64bitKernelDS */
540	# else
541	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
542	# endif
543	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
544	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
545	g_aFunctions[6].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0AbsKernelDS */
546	g_aFunctions[7].pfn = (void )(uintptr_t)ASMGetES(); / SUPR0AbsKernelES */
547	g_aFunctions[8].pfn = (void )(uintptr_t)ASMGetFS(); / SUPR0AbsKernelFS */
548	g_aFunctions[9].pfn = (void )(uintptr_t)ASMGetGS(); / SUPR0AbsKernelGS */
549	#endif /* !RT_OS_DARWIN */
550	return VINF_SUCCESS;
551	}
552
553	supdrvGipDestroy(pDevExt);
554	}
555
556	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
557	RTSemMutexDestroy(pDevExt->mtxGip);
558	pDevExt->mtxGip = NIL_RTSEMMUTEX;
559	#else
560	RTSemFastMutexDestroy(pDevExt->mtxGip);
561	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
562	#endif
563	}
564	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
565	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
566	}
567	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
568	RTSemMutexDestroy(pDevExt->mtxLdr);
569	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
570	#else
571	RTSemFastMutexDestroy(pDevExt->mtxLdr);
572	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
573	#endif
574	}
575	RTSpinlockDestroy(pDevExt->hGipSpinlock);
576	pDevExt->hGipSpinlock = NIL_RTSPINLOCK;
577	}
578	RTSpinlockDestroy(pDevExt->Spinlock);
579	pDevExt->Spinlock = NIL_RTSPINLOCK;
580	}
581	#ifdef SUPDRV_WITH_RELEASE_LOGGER
582	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
583	RTLogDestroy(RTLogSetDefaultInstance(NULL));
584	#endif
585
586	return rc;
587	}
588
589
590	/**
591	* Delete the device extension (e.g. cleanup members).
592	*
593	* @param pDevExt The device extension to delete.
594	*/
595	void VBOXCALL supdrvDeleteDevExt(PSUPDRVDEVEXT pDevExt)
596	{
597	PSUPDRVOBJ pObj;
598	PSUPDRVUSAGE pUsage;
599
600	/*
601	* Kill mutexes and spinlocks.
602	*/
603	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
604	RTSemMutexDestroy(pDevExt->mtxGip);
605	pDevExt->mtxGip = NIL_RTSEMMUTEX;
606	#else
607	RTSemFastMutexDestroy(pDevExt->mtxGip);
608	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
609	#endif
610	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
611	RTSemMutexDestroy(pDevExt->mtxLdr);
612	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
613	#else
614	RTSemFastMutexDestroy(pDevExt->mtxLdr);
615	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
616	#endif
617	RTSpinlockDestroy(pDevExt->Spinlock);
618	pDevExt->Spinlock = NIL_RTSPINLOCK;
619	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
620	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
621
622	/*
623	* Free lists.
624	*/
625	/* objects. */
626	pObj = pDevExt->pObjs;
627	Assert(!pObj); /* (can trigger on forced unloads) */
628	pDevExt->pObjs = NULL;
629	while (pObj)
630	{
631	void *pvFree = pObj;
632	pObj = pObj->pNext;
633	RTMemFree(pvFree);
634	}
635
636	/* usage records. */
637	pUsage = pDevExt->pUsageFree;
638	pDevExt->pUsageFree = NULL;
639	while (pUsage)
640	{
641	void *pvFree = pUsage;
642	pUsage = pUsage->pNext;
643	RTMemFree(pvFree);
644	}
645
646	/* kill the GIP. */
647	supdrvGipDestroy(pDevExt);
648	RTSpinlockDestroy(pDevExt->hGipSpinlock);
649	pDevExt->hGipSpinlock = NIL_RTSPINLOCK;
650
651	supdrvTracerTerm(pDevExt);
652
653	#ifdef SUPDRV_WITH_RELEASE_LOGGER
654	/* destroy the loggers. */
655	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
656	RTLogDestroy(RTLogSetDefaultInstance(NULL));
657	#endif
658	}
659
660
661	/**
662	* Create session.
663	*
664	* @returns IPRT status code.
665	* @param pDevExt Device extension.
666	* @param fUser Flag indicating whether this is a user or kernel
667	* session.
668	* @param fUnrestricted Unrestricted access (system) or restricted access
669	* (user)?
670	* @param ppSession Where to store the pointer to the session data.
671	*/
672	int VBOXCALL supdrvCreateSession(PSUPDRVDEVEXT pDevExt, bool fUser, bool fUnrestricted, PSUPDRVSESSION *ppSession)
673	{
674	int rc;
675	PSUPDRVSESSION pSession;
676
677	if (!SUP_IS_DEVEXT_VALID(pDevExt))
678	return VERR_INVALID_PARAMETER;
679
680	/*
681	* Allocate memory for the session data.
682	*/
683	pSession = *ppSession = (PSUPDRVSESSION)RTMemAllocZ(pDevExt->cbSession);
684	if (pSession)
685	{
686	/* Initialize session data. */
687	rc = RTSpinlockCreate(&pSession->Spinlock, RTSPINLOCK_FLAGS_INTERRUPT_UNSAFE, "SUPDrvSession");
688	if (!rc)
689	{
690	rc = RTHandleTableCreateEx(&pSession->hHandleTable,
691	RTHANDLETABLE_FLAGS_LOCKED_IRQ_SAFE \| RTHANDLETABLE_FLAGS_CONTEXT,
692	1 /uBase/, 32768 /cMax/, supdrvSessionObjHandleRetain, pSession);
693	if (RT_SUCCESS(rc))
694	{
695	Assert(pSession->Spinlock != NIL_RTSPINLOCK);
696	pSession->pDevExt = pDevExt;
697	pSession->u32Cookie = BIRD_INV;
698	pSession->fUnrestricted = fUnrestricted;
699	/*pSession->pLdrUsage = NULL;
700	pSession->pVM = NULL;
701	pSession->pUsage = NULL;
702	pSession->pGip = NULL;
703	pSession->fGipReferenced = false;
704	pSession->Bundle.cUsed = 0; */
705	pSession->Uid = NIL_RTUID;
706	pSession->Gid = NIL_RTGID;
707	if (fUser)
708	{
709	pSession->Process = RTProcSelf();
710	pSession->R0Process = RTR0ProcHandleSelf();
711	}
712	else
713	{
714	pSession->Process = NIL_RTPROCESS;
715	pSession->R0Process = NIL_RTR0PROCESS;
716	}
717	/pSession->uTracerData = 0;/
718	pSession->hTracerCaller = NIL_RTNATIVETHREAD;
719	RTListInit(&pSession->TpProviders);
720	/pSession->cTpProviders = 0;/
721	/pSession->cTpProbesFiring = 0;/
722	RTListInit(&pSession->TpUmods);
723	/RT_ZERO(pSession->apTpLookupTable);/
724
725	VBOXDRV_SESSION_CREATE(pSession, fUser);
726	LogFlow(("Created session %p initial cookie=%#x\n", pSession, pSession->u32Cookie));
727	return VINF_SUCCESS;
728	}
729
730	RTSpinlockDestroy(pSession->Spinlock);
731	}
732	RTMemFree(pSession);
733	*ppSession = NULL;
734	Log(("Failed to create spinlock, rc=%d!\n", rc));
735	}
736	else
737	rc = VERR_NO_MEMORY;
738
739	return rc;
740	}
741
742
743	/**
744	* Shared code for cleaning up a session.
745	*
746	* @param pDevExt Device extension.
747	* @param pSession Session data.
748	* This data will be freed by this routine.
749	*/
750	void VBOXCALL supdrvCloseSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
751	{
752	VBOXDRV_SESSION_CLOSE(pSession);
753
754	/*
755	* Cleanup the session first.
756	*/
757	supdrvCleanupSession(pDevExt, pSession);
758
759	/*
760	* Free the rest of the session stuff.
761	*/
762	RTSpinlockDestroy(pSession->Spinlock);
763	pSession->Spinlock = NIL_RTSPINLOCK;
764	pSession->pDevExt = NULL;
765	RTMemFree(pSession);
766	LogFlow(("supdrvCloseSession: returns\n"));
767	}
768
769
770	/**
771	* Shared code for cleaning up a session (but not quite freeing it).
772	*
773	* This is primarily intended for MAC OS X where we have to clean up the memory
774	* stuff before the file handle is closed.
775	*
776	* @param pDevExt Device extension.
777	* @param pSession Session data.
778	* This data will be freed by this routine.
779	*/
780	void VBOXCALL supdrvCleanupSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
781	{
782	int rc;
783	PSUPDRVBUNDLE pBundle;
784	LogFlow(("supdrvCleanupSession: pSession=%p\n", pSession));
785
786	/*
787	* Remove logger instances related to this session.
788	*/
789	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pSession);
790
791	/*
792	* Destroy the handle table.
793	*/
794	rc = RTHandleTableDestroy(pSession->hHandleTable, supdrvSessionObjHandleDelete, pSession);
795	AssertRC(rc);
796	pSession->hHandleTable = NIL_RTHANDLETABLE;
797
798	/*
799	* Release object references made in this session.
800	* In theory there should be noone racing us in this session.
801	*/
802	Log2(("release objects - start\n"));
803	if (pSession->pUsage)
804	{
805	PSUPDRVUSAGE pUsage;
806	RTSpinlockAcquire(pDevExt->Spinlock);
807
808	while ((pUsage = pSession->pUsage) != NULL)
809	{
810	PSUPDRVOBJ pObj = pUsage->pObj;
811	pSession->pUsage = pUsage->pNext;
812
813	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
814	if (pUsage->cUsage < pObj->cUsage)
815	{
816	pObj->cUsage -= pUsage->cUsage;
817	RTSpinlockRelease(pDevExt->Spinlock);
818	}
819	else
820	{
821	/* Destroy the object and free the record. */
822	if (pDevExt->pObjs == pObj)
823	pDevExt->pObjs = pObj->pNext;
824	else
825	{
826	PSUPDRVOBJ pObjPrev;
827	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
828	if (pObjPrev->pNext == pObj)
829	{
830	pObjPrev->pNext = pObj->pNext;
831	break;
832	}
833	Assert(pObjPrev);
834	}
835	RTSpinlockRelease(pDevExt->Spinlock);
836
837	Log(("supdrvCleanupSession: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
838	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
839	if (pObj->pfnDestructor)
840	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
841	RTMemFree(pObj);
842	}
843
844	/* free it and continue. */
845	RTMemFree(pUsage);
846
847	RTSpinlockAcquire(pDevExt->Spinlock);
848	}
849
850	RTSpinlockRelease(pDevExt->Spinlock);
851	AssertMsg(!pSession->pUsage, ("Some buster reregistered an object during desturction!\n"));
852	}
853	Log2(("release objects - done\n"));
854
855	/*
856	* Do tracer cleanups related to this session.
857	*/
858	Log2(("release tracer stuff - start\n"));
859	supdrvTracerCleanupSession(pDevExt, pSession);
860	Log2(("release tracer stuff - end\n"));
861
862	/*
863	* Release memory allocated in the session.
864	*
865	* We do not serialize this as we assume that the application will
866	* not allocated memory while closing the file handle object.
867	*/
868	Log2(("freeing memory:\n"));
869	pBundle = &pSession->Bundle;
870	while (pBundle)
871	{
872	PSUPDRVBUNDLE pToFree;
873	unsigned i;
874
875	/*
876	* Check and unlock all entries in the bundle.
877	*/
878	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
879	{
880	if (pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ)
881	{
882	Log2(("eType=%d pvR0=%p pvR3=%p cb=%ld\n", pBundle->aMem[i].eType, RTR0MemObjAddress(pBundle->aMem[i].MemObj),
883	(void *)RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3), (long)RTR0MemObjSize(pBundle->aMem[i].MemObj)));
884	if (pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ)
885	{
886	rc = RTR0MemObjFree(pBundle->aMem[i].MapObjR3, false);
887	AssertRC(rc); /** @todo figure out how to handle this. */
888	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
889	}
890	rc = RTR0MemObjFree(pBundle->aMem[i].MemObj, true /* fFreeMappings */);
891	AssertRC(rc); /** @todo figure out how to handle this. */
892	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
893	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
894	}
895	}
896
897	/*
898	* Advance and free previous bundle.
899	*/
900	pToFree = pBundle;
901	pBundle = pBundle->pNext;
902
903	pToFree->pNext = NULL;
904	pToFree->cUsed = 0;
905	if (pToFree != &pSession->Bundle)
906	RTMemFree(pToFree);
907	}
908	Log2(("freeing memory - done\n"));
909
910	/*
911	* Deregister component factories.
912	*/
913	RTSemFastMutexRequest(pDevExt->mtxComponentFactory);
914	Log2(("deregistering component factories:\n"));
915	if (pDevExt->pComponentFactoryHead)
916	{
917	PSUPDRVFACTORYREG pPrev = NULL;
918	PSUPDRVFACTORYREG pCur = pDevExt->pComponentFactoryHead;
919	while (pCur)
920	{
921	if (pCur->pSession == pSession)
922	{
923	/* unlink it */
924	PSUPDRVFACTORYREG pNext = pCur->pNext;
925	if (pPrev)
926	pPrev->pNext = pNext;
927	else
928	pDevExt->pComponentFactoryHead = pNext;
929
930	/* free it */
931	pCur->pNext = NULL;
932	pCur->pSession = NULL;
933	pCur->pFactory = NULL;
934	RTMemFree(pCur);
935
936	/* next */
937	pCur = pNext;
938	}
939	else
940	{
941	/* next */
942	pPrev = pCur;
943	pCur = pCur->pNext;
944	}
945	}
946	}
947	RTSemFastMutexRelease(pDevExt->mtxComponentFactory);
948	Log2(("deregistering component factories - done\n"));
949
950	/*
951	* Loaded images needs to be dereferenced and possibly freed up.
952	*/
953	supdrvLdrLock(pDevExt);
954	Log2(("freeing images:\n"));
955	if (pSession->pLdrUsage)
956	{
957	PSUPDRVLDRUSAGE pUsage = pSession->pLdrUsage;
958	pSession->pLdrUsage = NULL;
959	while (pUsage)
960	{
961	void *pvFree = pUsage;
962	PSUPDRVLDRIMAGE pImage = pUsage->pImage;
963	if (pImage->cUsage > pUsage->cUsage)
964	pImage->cUsage -= pUsage->cUsage;
965	else
966	supdrvLdrFree(pDevExt, pImage);
967	pUsage->pImage = NULL;
968	pUsage = pUsage->pNext;
969	RTMemFree(pvFree);
970	}
971	}
972	supdrvLdrUnlock(pDevExt);
973	Log2(("freeing images - done\n"));
974
975	/*
976	* Unmap the GIP.
977	*/
978	Log2(("umapping GIP:\n"));
979	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
980	{
981	SUPR0GipUnmap(pSession);
982	pSession->fGipReferenced = 0;
983	}
984	Log2(("umapping GIP - done\n"));
985	}
986
987
988	/**
989	* RTHandleTableDestroy callback used by supdrvCleanupSession.
990	*
991	* @returns IPRT status code, see SUPR0ObjAddRef.
992	* @param hHandleTable The handle table handle. Ignored.
993	* @param pvObj The object pointer.
994	* @param pvCtx Context, the handle type. Ignored.
995	* @param pvUser Session pointer.
996	*/
997	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser)
998	{
999	NOREF(pvCtx);
1000	NOREF(hHandleTable);
1001	return SUPR0ObjAddRefEx(pvObj, (PSUPDRVSESSION)pvUser, true /fNoBlocking/);
1002	}
1003
1004
1005	/**
1006	* RTHandleTableDestroy callback used by supdrvCleanupSession.
1007	*
1008	* @param hHandleTable The handle table handle. Ignored.
1009	* @param h The handle value. Ignored.
1010	* @param pvObj The object pointer.
1011	* @param pvCtx Context, the handle type. Ignored.
1012	* @param pvUser Session pointer.
1013	*/
1014	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser)
1015	{
1016	NOREF(pvCtx);
1017	NOREF(h);
1018	NOREF(hHandleTable);
1019	SUPR0ObjRelease(pvObj, (PSUPDRVSESSION)pvUser);
1020	}
1021
1022
1023	/**
1024	* Fast path I/O Control worker.
1025	*
1026	* @returns VBox status code that should be passed down to ring-3 unchanged.
1027	* @param uIOCtl Function number.
1028	* @param idCpu VMCPU id.
1029	* @param pDevExt Device extention.
1030	* @param pSession Session data.
1031	*/
1032	int VBOXCALL supdrvIOCtlFast(uintptr_t uIOCtl, VMCPUID idCpu, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
1033	{
1034	/*
1035	* We check the two prereqs after doing this only to allow the compiler to optimize things better.
1036	*/
1037	if (RT_LIKELY( RT_VALID_PTR(pSession)
1038	&& pSession->pVM
1039	&& pDevExt->pfnVMMR0EntryFast))
1040	{
1041	switch (uIOCtl)
1042	{
1043	case SUP_IOCTL_FAST_DO_RAW_RUN:
1044	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_RAW_RUN);
1045	break;
1046	case SUP_IOCTL_FAST_DO_HM_RUN:
1047	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_HM_RUN);
1048	break;
1049	case SUP_IOCTL_FAST_DO_NOP:
1050	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_NOP);
1051	break;
1052	default:
1053	return VERR_INTERNAL_ERROR;
1054	}
1055	return VINF_SUCCESS;
1056	}
1057	return VERR_INTERNAL_ERROR;
1058	}
1059
1060
1061	/**
1062	* Helper for supdrvIOCtl. Check if pszStr contains any character of pszChars.
1063	* We would use strpbrk here if this function would be contained in the RedHat kABI white
1064	* list, see http://www.kerneldrivers.org/RHEL5.
1065	*
1066	* @returns 1 if pszStr does contain any character of pszChars, 0 otherwise.
1067	* @param pszStr String to check
1068	* @param pszChars Character set
1069	*/
1070	static int supdrvCheckInvalidChar(const char pszStr, const char pszChars)
1071	{
1072	int chCur;
1073	while ((chCur = *pszStr++) != '\0')
1074	{
1075	int ch;
1076	const char *psz = pszChars;
1077	while ((ch = *psz++) != '\0')
1078	if (ch == chCur)
1079	return 1;
1080
1081	}
1082	return 0;
1083	}
1084
1085
1086
1087	/**
1088	* I/O Control inner worker (tracing reasons).
1089	*
1090	* @returns IPRT status code.
1091	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1092	*
1093	* @param uIOCtl Function number.
1094	* @param pDevExt Device extention.
1095	* @param pSession Session data.
1096	* @param pReqHdr The request header.
1097	*/
1098	static int supdrvIOCtlInnerUnrestricted(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1099	{
1100	/*
1101	* Validation macros
1102	*/
1103	#define REQ_CHECK_SIZES_EX(Name, cbInExpect, cbOutExpect) \
1104	do { \
1105	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect) \|\| pReqHdr->cbOut != (cbOutExpect))) \
1106	{ \
1107	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld. cbOut=%ld expected %ld.\n", \
1108	(long)pReqHdr->cbIn, (long)(cbInExpect), (long)pReqHdr->cbOut, (long)(cbOutExpect))); \
1109	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1110	} \
1111	} while (0)
1112
1113	#define REQ_CHECK_SIZES(Name) REQ_CHECK_SIZES_EX(Name, Name ## _SIZE_IN, Name ## _SIZE_OUT)
1114
1115	#define REQ_CHECK_SIZE_IN(Name, cbInExpect) \
1116	do { \
1117	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect))) \
1118	{ \
1119	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld.\n", \
1120	(long)pReqHdr->cbIn, (long)(cbInExpect))); \
1121	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1122	} \
1123	} while (0)
1124
1125	#define REQ_CHECK_SIZE_OUT(Name, cbOutExpect) \
1126	do { \
1127	if (RT_UNLIKELY(pReqHdr->cbOut != (cbOutExpect))) \
1128	{ \
1129	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbOut=%ld expected %ld.\n", \
1130	(long)pReqHdr->cbOut, (long)(cbOutExpect))); \
1131	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1132	} \
1133	} while (0)
1134
1135	#define REQ_CHECK_EXPR(Name, expr) \
1136	do { \
1137	if (RT_UNLIKELY(!(expr))) \
1138	{ \
1139	OSDBGPRINT(( #Name ": %s\n", #expr)); \
1140	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1141	} \
1142	} while (0)
1143
1144	#define REQ_CHECK_EXPR_FMT(expr, fmt) \
1145	do { \
1146	if (RT_UNLIKELY(!(expr))) \
1147	{ \
1148	OSDBGPRINT( fmt ); \
1149	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1150	} \
1151	} while (0)
1152
1153	/*
1154	* The switch.
1155	*/
1156	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
1157	{
1158	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
1159	{
1160	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
1161	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
1162	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
1163	{
1164	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
1165	pReq->Hdr.rc = VERR_INVALID_MAGIC;
1166	return 0;
1167	}
1168
1169	#if 0
1170	/*
1171	* Call out to the OS specific code and let it do permission checks on the
1172	* client process.
1173	*/
1174	if (!supdrvOSValidateClientProcess(pDevExt, pSession))
1175	{
1176	pReq->u.Out.u32Cookie = 0xffffffff;
1177	pReq->u.Out.u32SessionCookie = 0xffffffff;
1178	pReq->u.Out.u32SessionVersion = 0xffffffff;
1179	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1180	pReq->u.Out.pSession = NULL;
1181	pReq->u.Out.cFunctions = 0;
1182	pReq->Hdr.rc = VERR_PERMISSION_DENIED;
1183	return 0;
1184	}
1185	#endif
1186
1187	/*
1188	* Match the version.
1189	* The current logic is very simple, match the major interface version.
1190	*/
1191	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
1192	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
1193	{
1194	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1195	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
1196	pReq->u.Out.u32Cookie = 0xffffffff;
1197	pReq->u.Out.u32SessionCookie = 0xffffffff;
1198	pReq->u.Out.u32SessionVersion = 0xffffffff;
1199	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1200	pReq->u.Out.pSession = NULL;
1201	pReq->u.Out.cFunctions = 0;
1202	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1203	return 0;
1204	}
1205
1206	/*
1207	* Fill in return data and be gone.
1208	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
1209	* u32SessionVersion <= u32ReqVersion!
1210	*/
1211	/** @todo Somehow validate the client and negotiate a secure cookie... */
1212	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
1213	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
1214	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
1215	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1216	pReq->u.Out.pSession = pSession;
1217	pReq->u.Out.cFunctions = sizeof(g_aFunctions) / sizeof(g_aFunctions[0]);
1218	pReq->Hdr.rc = VINF_SUCCESS;
1219	return 0;
1220	}
1221
1222	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_QUERY_FUNCS(0)):
1223	{
1224	/* validate */
1225	PSUPQUERYFUNCS pReq = (PSUPQUERYFUNCS)pReqHdr;
1226	REQ_CHECK_SIZES_EX(SUP_IOCTL_QUERY_FUNCS, SUP_IOCTL_QUERY_FUNCS_SIZE_IN, SUP_IOCTL_QUERY_FUNCS_SIZE_OUT(RT_ELEMENTS(g_aFunctions)));
1227
1228	/* execute */
1229	pReq->u.Out.cFunctions = RT_ELEMENTS(g_aFunctions);
1230	memcpy(&pReq->u.Out.aFunctions[0], g_aFunctions, sizeof(g_aFunctions));
1231	pReq->Hdr.rc = VINF_SUCCESS;
1232	return 0;
1233	}
1234
1235	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_LOCK):
1236	{
1237	/* validate */
1238	PSUPPAGELOCK pReq = (PSUPPAGELOCK)pReqHdr;
1239	REQ_CHECK_SIZE_IN(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_IN);
1240	REQ_CHECK_SIZE_OUT(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_OUT(pReq->u.In.cPages));
1241	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.cPages > 0);
1242	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.pvR3 >= PAGE_SIZE);
1243
1244	/* execute */
1245	pReq->Hdr.rc = SUPR0LockMem(pSession, pReq->u.In.pvR3, pReq->u.In.cPages, &pReq->u.Out.aPages[0]);
1246	if (RT_FAILURE(pReq->Hdr.rc))
1247	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1248	return 0;
1249	}
1250
1251	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_UNLOCK):
1252	{
1253	/* validate */
1254	PSUPPAGEUNLOCK pReq = (PSUPPAGEUNLOCK)pReqHdr;
1255	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_UNLOCK);
1256
1257	/* execute */
1258	pReq->Hdr.rc = SUPR0UnlockMem(pSession, pReq->u.In.pvR3);
1259	return 0;
1260	}
1261
1262	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_ALLOC):
1263	{
1264	/* validate */
1265	PSUPCONTALLOC pReq = (PSUPCONTALLOC)pReqHdr;
1266	REQ_CHECK_SIZES(SUP_IOCTL_CONT_ALLOC);
1267
1268	/* execute */
1269	pReq->Hdr.rc = SUPR0ContAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.HCPhys);
1270	if (RT_FAILURE(pReq->Hdr.rc))
1271	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1272	return 0;
1273	}
1274
1275	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_FREE):
1276	{
1277	/* validate */
1278	PSUPCONTFREE pReq = (PSUPCONTFREE)pReqHdr;
1279	REQ_CHECK_SIZES(SUP_IOCTL_CONT_FREE);
1280
1281	/* execute */
1282	pReq->Hdr.rc = SUPR0ContFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1283	return 0;
1284	}
1285
1286	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_OPEN):
1287	{
1288	/* validate */
1289	PSUPLDROPEN pReq = (PSUPLDROPEN)pReqHdr;
1290	REQ_CHECK_SIZES(SUP_IOCTL_LDR_OPEN);
1291	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs > 0);
1292	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs < 16*_1M);
1293	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1294	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1295	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits < pReq->u.In.cbImageWithTabs);
1296	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.szName[0]);
1297	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1298	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, !supdrvCheckInvalidChar(pReq->u.In.szName, ";:()[]{}/\\\|&*%#@!~`\"'"));
1299	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szFilename, sizeof(pReq->u.In.szFilename)));
1300
1301	/* execute */
1302	pReq->Hdr.rc = supdrvIOCtl_LdrOpen(pDevExt, pSession, pReq);
1303	return 0;
1304	}
1305
1306	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOAD):
1307	{
1308	/* validate */
1309	PSUPLDRLOAD pReq = (PSUPLDRLOAD)pReqHdr;
1310	REQ_CHECK_EXPR(Name, pReq->Hdr.cbIn >= sizeof(*pReq));
1311	REQ_CHECK_SIZES_EX(SUP_IOCTL_LDR_LOAD, SUP_IOCTL_LDR_LOAD_SIZE_IN(pReq->u.In.cbImageWithTabs), SUP_IOCTL_LDR_LOAD_SIZE_OUT);
1312	REQ_CHECK_EXPR(SUP_IOCTL_LDR_LOAD, pReq->u.In.cSymbols <= 16384);
1313	REQ_CHECK_EXPR_FMT( !pReq->u.In.cSymbols
1314	\|\| ( pReq->u.In.offSymbols < pReq->u.In.cbImageWithTabs
1315	&& pReq->u.In.offSymbols + pReq->u.In.cSymbols * sizeof(SUPLDRSYM) <= pReq->u.In.cbImageWithTabs),
1316	("SUP_IOCTL_LDR_LOAD: offSymbols=%#lx cSymbols=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offSymbols,
1317	(long)pReq->u.In.cSymbols, (long)pReq->u.In.cbImageWithTabs));
1318	REQ_CHECK_EXPR_FMT( !pReq->u.In.cbStrTab
1319	\|\| ( pReq->u.In.offStrTab < pReq->u.In.cbImageWithTabs
1320	&& pReq->u.In.offStrTab + pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs
1321	&& pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs),
1322	("SUP_IOCTL_LDR_LOAD: offStrTab=%#lx cbStrTab=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offStrTab,
1323	(long)pReq->u.In.cbStrTab, (long)pReq->u.In.cbImageWithTabs));
1324
1325	if (pReq->u.In.cSymbols)
1326	{
1327	uint32_t i;
1328	PSUPLDRSYM paSyms = (PSUPLDRSYM)&pReq->u.In.abImage[pReq->u.In.offSymbols];
1329	for (i = 0; i < pReq->u.In.cSymbols; i++)
1330	{
1331	REQ_CHECK_EXPR_FMT(paSyms[i].offSymbol < pReq->u.In.cbImageWithTabs,
1332	("SUP_IOCTL_LDR_LOAD: sym #%ld: symb off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offSymbol, (long)pReq->u.In.cbImageWithTabs));
1333	REQ_CHECK_EXPR_FMT(paSyms[i].offName < pReq->u.In.cbStrTab,
1334	("SUP_IOCTL_LDR_LOAD: sym #%ld: name off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1335	REQ_CHECK_EXPR_FMT(RTStrEnd((char const *)&pReq->u.In.abImage[pReq->u.In.offStrTab + paSyms[i].offName],
1336	pReq->u.In.cbStrTab - paSyms[i].offName),
1337	("SUP_IOCTL_LDR_LOAD: sym #%ld: unterminated name! (%#lx / %#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1338	}
1339	}
1340
1341	/* execute */
1342	pReq->Hdr.rc = supdrvIOCtl_LdrLoad(pDevExt, pSession, pReq);
1343	return 0;
1344	}
1345
1346	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_FREE):
1347	{
1348	/* validate */
1349	PSUPLDRFREE pReq = (PSUPLDRFREE)pReqHdr;
1350	REQ_CHECK_SIZES(SUP_IOCTL_LDR_FREE);
1351
1352	/* execute */
1353	pReq->Hdr.rc = supdrvIOCtl_LdrFree(pDevExt, pSession, pReq);
1354	return 0;
1355	}
1356
1357	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_GET_SYMBOL):
1358	{
1359	/* validate */
1360	PSUPLDRGETSYMBOL pReq = (PSUPLDRGETSYMBOL)pReqHdr;
1361	REQ_CHECK_SIZES(SUP_IOCTL_LDR_GET_SYMBOL);
1362	REQ_CHECK_EXPR(SUP_IOCTL_LDR_GET_SYMBOL, RTStrEnd(pReq->u.In.szSymbol, sizeof(pReq->u.In.szSymbol)));
1363
1364	/* execute */
1365	pReq->Hdr.rc = supdrvIOCtl_LdrGetSymbol(pDevExt, pSession, pReq);
1366	return 0;
1367	}
1368
1369	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0(0)):
1370	{
1371	/* validate */
1372	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1373	Log4(("SUP_IOCTL_CALL_VMMR0: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1374	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1375
1376	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_VMMR0_SIZE(0))
1377	{
1378	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(0), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(0));
1379
1380	/* execute */
1381	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1382	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg, pSession);
1383	else
1384	pReq->Hdr.rc = VERR_WRONG_ORDER;
1385	}
1386	else
1387	{
1388	PSUPVMMR0REQHDR pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1389	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR)),
1390	("SUP_IOCTL_CALL_VMMR0: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR))));
1391	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1392	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(pVMMReq->cbReq));
1393
1394	/* execute */
1395	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1396	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1397	else
1398	pReq->Hdr.rc = VERR_WRONG_ORDER;
1399	}
1400
1401	if ( RT_FAILURE(pReq->Hdr.rc)
1402	&& pReq->Hdr.rc != VERR_INTERRUPTED
1403	&& pReq->Hdr.rc != VERR_TIMEOUT)
1404	Log(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1405	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1406	else
1407	Log4(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1408	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1409	return 0;
1410	}
1411
1412	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0_BIG):
1413	{
1414	/* validate */
1415	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1416	PSUPVMMR0REQHDR pVMMReq;
1417	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1418	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1419
1420	pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1421	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR)),
1422	("SUP_IOCTL_CALL_VMMR0_BIG: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR))));
1423	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0_BIG, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1424	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));
1425
1426	/* execute */
1427	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1428	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1429	else
1430	pReq->Hdr.rc = VERR_WRONG_ORDER;
1431
1432	if ( RT_FAILURE(pReq->Hdr.rc)
1433	&& pReq->Hdr.rc != VERR_INTERRUPTED
1434	&& pReq->Hdr.rc != VERR_TIMEOUT)
1435	Log(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1436	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1437	else
1438	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1439	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1440	return 0;
1441	}
1442
1443	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_PAGING_MODE):
1444	{
1445	/* validate */
1446	PSUPGETPAGINGMODE pReq = (PSUPGETPAGINGMODE)pReqHdr;
1447	REQ_CHECK_SIZES(SUP_IOCTL_GET_PAGING_MODE);
1448
1449	/* execute */
1450	pReq->Hdr.rc = VINF_SUCCESS;
1451	pReq->u.Out.enmMode = SUPR0GetPagingMode();
1452	return 0;
1453	}
1454
1455	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_ALLOC):
1456	{
1457	/* validate */
1458	PSUPLOWALLOC pReq = (PSUPLOWALLOC)pReqHdr;
1459	REQ_CHECK_EXPR(SUP_IOCTL_LOW_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_LOW_ALLOC_SIZE_IN);
1460	REQ_CHECK_SIZES_EX(SUP_IOCTL_LOW_ALLOC, SUP_IOCTL_LOW_ALLOC_SIZE_IN, SUP_IOCTL_LOW_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1461
1462	/* execute */
1463	pReq->Hdr.rc = SUPR0LowAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1464	if (RT_FAILURE(pReq->Hdr.rc))
1465	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1466	return 0;
1467	}
1468
1469	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_FREE):
1470	{
1471	/* validate */
1472	PSUPLOWFREE pReq = (PSUPLOWFREE)pReqHdr;
1473	REQ_CHECK_SIZES(SUP_IOCTL_LOW_FREE);
1474
1475	/* execute */
1476	pReq->Hdr.rc = SUPR0LowFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1477	return 0;
1478	}
1479
1480	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_MAP):
1481	{
1482	/* validate */
1483	PSUPGIPMAP pReq = (PSUPGIPMAP)pReqHdr;
1484	REQ_CHECK_SIZES(SUP_IOCTL_GIP_MAP);
1485
1486	/* execute */
1487	pReq->Hdr.rc = SUPR0GipMap(pSession, &pReq->u.Out.pGipR3, &pReq->u.Out.HCPhysGip);
1488	if (RT_SUCCESS(pReq->Hdr.rc))
1489	pReq->u.Out.pGipR0 = pDevExt->pGip;
1490	return 0;
1491	}
1492
1493	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_UNMAP):
1494	{
1495	/* validate */
1496	PSUPGIPUNMAP pReq = (PSUPGIPUNMAP)pReqHdr;
1497	REQ_CHECK_SIZES(SUP_IOCTL_GIP_UNMAP);
1498
1499	/* execute */
1500	pReq->Hdr.rc = SUPR0GipUnmap(pSession);
1501	return 0;
1502	}
1503
1504	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SET_VM_FOR_FAST):
1505	{
1506	/* validate */
1507	PSUPSETVMFORFAST pReq = (PSUPSETVMFORFAST)pReqHdr;
1508	REQ_CHECK_SIZES(SUP_IOCTL_SET_VM_FOR_FAST);
1509	REQ_CHECK_EXPR_FMT( !pReq->u.In.pVMR0
1510	\|\| ( VALID_PTR(pReq->u.In.pVMR0)
1511	&& !((uintptr_t)pReq->u.In.pVMR0 & (PAGE_SIZE - 1))),
1512	("SUP_IOCTL_SET_VM_FOR_FAST: pVMR0=%p!\n", pReq->u.In.pVMR0));
1513	/* execute */
1514	pSession->pVM = pReq->u.In.pVMR0;
1515	pReq->Hdr.rc = VINF_SUCCESS;
1516	return 0;
1517	}
1518
1519	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_ALLOC_EX):
1520	{
1521	/* validate */
1522	PSUPPAGEALLOCEX pReq = (PSUPPAGEALLOCEX)pReqHdr;
1523	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC_EX, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_EX_SIZE_IN);
1524	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));
1525	REQ_CHECK_EXPR_FMT(pReq->u.In.fKernelMapping \|\| pReq->u.In.fUserMapping,
1526	("SUP_IOCTL_PAGE_ALLOC_EX: No mapping requested!\n"));
1527	REQ_CHECK_EXPR_FMT(pReq->u.In.fUserMapping,
1528	("SUP_IOCTL_PAGE_ALLOC_EX: Must have user mapping!\n"));
1529	REQ_CHECK_EXPR_FMT(!pReq->u.In.fReserved0 && !pReq->u.In.fReserved1,
1530	("SUP_IOCTL_PAGE_ALLOC_EX: fReserved0=%d fReserved1=%d\n", pReq->u.In.fReserved0, pReq->u.In.fReserved1));
1531
1532	/* execute */
1533	pReq->Hdr.rc = SUPR0PageAllocEx(pSession, pReq->u.In.cPages, 0 /* fFlags */,
1534	pReq->u.In.fUserMapping ? &pReq->u.Out.pvR3 : NULL,
1535	pReq->u.In.fKernelMapping ? &pReq->u.Out.pvR0 : NULL,
1536	&pReq->u.Out.aPages[0]);
1537	if (RT_FAILURE(pReq->Hdr.rc))
1538	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1539	return 0;
1540	}
1541
1542	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_MAP_KERNEL):
1543	{
1544	/* validate */
1545	PSUPPAGEMAPKERNEL pReq = (PSUPPAGEMAPKERNEL)pReqHdr;
1546	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_MAP_KERNEL);
1547	REQ_CHECK_EXPR_FMT(!pReq->u.In.fFlags, ("SUP_IOCTL_PAGE_MAP_KERNEL: fFlags=%#x! MBZ\n", pReq->u.In.fFlags));
1548	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_MAP_KERNEL: offSub=%#x\n", pReq->u.In.offSub));
1549	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1550	("SUP_IOCTL_PAGE_MAP_KERNEL: cbSub=%#x\n", pReq->u.In.cbSub));
1551
1552	/* execute */
1553	pReq->Hdr.rc = SUPR0PageMapKernel(pSession, pReq->u.In.pvR3, pReq->u.In.offSub, pReq->u.In.cbSub,
1554	pReq->u.In.fFlags, &pReq->u.Out.pvR0);
1555	if (RT_FAILURE(pReq->Hdr.rc))
1556	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1557	return 0;
1558	}
1559
1560	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_PROTECT):
1561	{
1562	/* validate */
1563	PSUPPAGEPROTECT pReq = (PSUPPAGEPROTECT)pReqHdr;
1564	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_PROTECT);
1565	REQ_CHECK_EXPR_FMT(!(pReq->u.In.fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)),
1566	("SUP_IOCTL_PAGE_PROTECT: fProt=%#x!\n", pReq->u.In.fProt));
1567	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_PROTECT: offSub=%#x\n", pReq->u.In.offSub));
1568	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1569	("SUP_IOCTL_PAGE_PROTECT: cbSub=%#x\n", pReq->u.In.cbSub));
1570
1571	/* execute */
1572	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);
1573	return 0;
1574	}
1575
1576	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_FREE):
1577	{
1578	/* validate */
1579	PSUPPAGEFREE pReq = (PSUPPAGEFREE)pReqHdr;
1580	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_FREE);
1581
1582	/* execute */
1583	pReq->Hdr.rc = SUPR0PageFree(pSession, pReq->u.In.pvR3);
1584	return 0;
1585	}
1586
1587	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_SERVICE(0)):
1588	{
1589	/* validate */
1590	PSUPCALLSERVICE pReq = (PSUPCALLSERVICE)pReqHdr;
1591	Log4(("SUP_IOCTL_CALL_SERVICE: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1592	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1593
1594	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
1595	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(0), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(0));
1596	else
1597	{
1598	PSUPR0SERVICEREQHDR pSrvReq = (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0];
1599	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR)),
1600	("SUP_IOCTL_CALL_SERVICE: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR))));
1601	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, pSrvReq->u32Magic == SUPR0SERVICEREQHDR_MAGIC);
1602	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(pSrvReq->cbReq), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(pSrvReq->cbReq));
1603	}
1604	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1605
1606	/* execute */
1607	pReq->Hdr.rc = supdrvIOCtl_CallServiceModule(pDevExt, pSession, pReq);
1608	return 0;
1609	}
1610
1611	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOGGER_SETTINGS(0)):
1612	{
1613	/* validate */
1614	PSUPLOGGERSETTINGS pReq = (PSUPLOGGERSETTINGS)pReqHdr;
1615	size_t cbStrTab;
1616	REQ_CHECK_SIZE_OUT(SUP_IOCTL_LOGGER_SETTINGS, SUP_IOCTL_LOGGER_SETTINGS_SIZE_OUT);
1617	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->Hdr.cbIn >= SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(1));
1618	cbStrTab = pReq->Hdr.cbIn - SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(0);
1619	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offGroups < cbStrTab);
1620	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offFlags < cbStrTab);
1621	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offDestination < cbStrTab);
1622	REQ_CHECK_EXPR_FMT(pReq->u.In.szStrings[cbStrTab - 1] == '\0',
1623	("SUP_IOCTL_LOGGER_SETTINGS: cbIn=%#x cbStrTab=%#zx LastChar=%d\n",
1624	pReq->Hdr.cbIn, cbStrTab, pReq->u.In.szStrings[cbStrTab - 1]));
1625	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhich <= SUPLOGGERSETTINGS_WHICH_RELEASE);
1626	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhat <= SUPLOGGERSETTINGS_WHAT_DESTROY);
1627
1628	/* execute */
1629	pReq->Hdr.rc = supdrvIOCtl_LoggerSettings(pDevExt, pSession, pReq);
1630	return 0;
1631	}
1632
1633	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP2):
1634	{
1635	/* validate */
1636	PSUPSEMOP2 pReq = (PSUPSEMOP2)pReqHdr;
1637	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP2, SUP_IOCTL_SEM_OP2_SIZE_IN, SUP_IOCTL_SEM_OP2_SIZE_OUT);
1638	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP2, pReq->u.In.uReserved == 0);
1639
1640	/* execute */
1641	switch (pReq->u.In.uType)
1642	{
1643	case SUP_SEM_TYPE_EVENT:
1644	{
1645	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1646	switch (pReq->u.In.uOp)
1647	{
1648	case SUPSEMOP2_WAIT_MS_REL:
1649	pReq->Hdr.rc = SUPSemEventWaitNoResume(pSession, hEvent, pReq->u.In.uArg.cRelMsTimeout);
1650	break;
1651	case SUPSEMOP2_WAIT_NS_ABS:
1652	pReq->Hdr.rc = SUPSemEventWaitNsAbsIntr(pSession, hEvent, pReq->u.In.uArg.uAbsNsTimeout);
1653	break;
1654	case SUPSEMOP2_WAIT_NS_REL:
1655	pReq->Hdr.rc = SUPSemEventWaitNsRelIntr(pSession, hEvent, pReq->u.In.uArg.cRelNsTimeout);
1656	break;
1657	case SUPSEMOP2_SIGNAL:
1658	pReq->Hdr.rc = SUPSemEventSignal(pSession, hEvent);
1659	break;
1660	case SUPSEMOP2_CLOSE:
1661	pReq->Hdr.rc = SUPSemEventClose(pSession, hEvent);
1662	break;
1663	case SUPSEMOP2_RESET:
1664	default:
1665	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1666	break;
1667	}
1668	break;
1669	}
1670
1671	case SUP_SEM_TYPE_EVENT_MULTI:
1672	{
1673	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1674	switch (pReq->u.In.uOp)
1675	{
1676	case SUPSEMOP2_WAIT_MS_REL:
1677	pReq->Hdr.rc = SUPSemEventMultiWaitNoResume(pSession, hEventMulti, pReq->u.In.uArg.cRelMsTimeout);
1678	break;
1679	case SUPSEMOP2_WAIT_NS_ABS:
1680	pReq->Hdr.rc = SUPSemEventMultiWaitNsAbsIntr(pSession, hEventMulti, pReq->u.In.uArg.uAbsNsTimeout);
1681	break;
1682	case SUPSEMOP2_WAIT_NS_REL:
1683	pReq->Hdr.rc = SUPSemEventMultiWaitNsRelIntr(pSession, hEventMulti, pReq->u.In.uArg.cRelNsTimeout);
1684	break;
1685	case SUPSEMOP2_SIGNAL:
1686	pReq->Hdr.rc = SUPSemEventMultiSignal(pSession, hEventMulti);
1687	break;
1688	case SUPSEMOP2_CLOSE:
1689	pReq->Hdr.rc = SUPSemEventMultiClose(pSession, hEventMulti);
1690	break;
1691	case SUPSEMOP2_RESET:
1692	pReq->Hdr.rc = SUPSemEventMultiReset(pSession, hEventMulti);
1693	break;
1694	default:
1695	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1696	break;
1697	}
1698	break;
1699	}
1700
1701	default:
1702	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1703	break;
1704	}
1705	return 0;
1706	}
1707
1708	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP3):
1709	{
1710	/* validate */
1711	PSUPSEMOP3 pReq = (PSUPSEMOP3)pReqHdr;
1712	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP3, SUP_IOCTL_SEM_OP3_SIZE_IN, SUP_IOCTL_SEM_OP3_SIZE_OUT);
1713	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, pReq->u.In.u32Reserved == 0 && pReq->u.In.u64Reserved == 0);
1714
1715	/* execute */
1716	switch (pReq->u.In.uType)
1717	{
1718	case SUP_SEM_TYPE_EVENT:
1719	{
1720	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1721	switch (pReq->u.In.uOp)
1722	{
1723	case SUPSEMOP3_CREATE:
1724	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1725	pReq->Hdr.rc = SUPSemEventCreate(pSession, &hEvent);
1726	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEvent;
1727	break;
1728	case SUPSEMOP3_GET_RESOLUTION:
1729	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1730	pReq->Hdr.rc = VINF_SUCCESS;
1731	pReq->Hdr.cbOut = sizeof(*pReq);
1732	pReq->u.Out.cNsResolution = SUPSemEventGetResolution(pSession);
1733	break;
1734	default:
1735	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1736	break;
1737	}
1738	break;
1739	}
1740
1741	case SUP_SEM_TYPE_EVENT_MULTI:
1742	{
1743	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1744	switch (pReq->u.In.uOp)
1745	{
1746	case SUPSEMOP3_CREATE:
1747	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1748	pReq->Hdr.rc = SUPSemEventMultiCreate(pSession, &hEventMulti);
1749	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEventMulti;
1750	break;
1751	case SUPSEMOP3_GET_RESOLUTION:
1752	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1753	pReq->Hdr.rc = VINF_SUCCESS;
1754	pReq->u.Out.cNsResolution = SUPSemEventMultiGetResolution(pSession);
1755	break;
1756	default:
1757	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1758	break;
1759	}
1760	break;
1761	}
1762
1763	default:
1764	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1765	break;
1766	}
1767	return 0;
1768	}
1769
1770	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
1771	{
1772	/* validate */
1773	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
1774	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
1775
1776	/* execute */
1777	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.Caps);
1778	if (RT_FAILURE(pReq->Hdr.rc))
1779	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1780	return 0;
1781	}
1782
1783	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_OPEN):
1784	{
1785	/* validate */
1786	PSUPTRACEROPEN pReq = (PSUPTRACEROPEN)pReqHdr;
1787	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_OPEN);
1788
1789	/* execute */
1790	pReq->Hdr.rc = supdrvIOCtl_TracerOpen(pDevExt, pSession, pReq->u.In.uCookie, pReq->u.In.uArg);
1791	return 0;
1792	}
1793
1794	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_CLOSE):
1795	{
1796	/* validate */
1797	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_CLOSE);
1798
1799	/* execute */
1800	pReqHdr->rc = supdrvIOCtl_TracerClose(pDevExt, pSession);
1801	return 0;
1802	}
1803
1804	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_IOCTL):
1805	{
1806	/* validate */
1807	PSUPTRACERIOCTL pReq = (PSUPTRACERIOCTL)pReqHdr;
1808	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_IOCTL);
1809
1810	/* execute */
1811	pReqHdr->rc = supdrvIOCtl_TracerIOCtl(pDevExt, pSession, pReq->u.In.uCmd, pReq->u.In.uArg, &pReq->u.Out.iRetVal);
1812	return 0;
1813	}
1814
1815	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_REG):
1816	{
1817	/* validate */
1818	PSUPTRACERUMODREG pReq = (PSUPTRACERUMODREG)pReqHdr;
1819	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_REG);
1820	if (!RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)))
1821	return VERR_INVALID_PARAMETER;
1822
1823	/* execute */
1824	pReqHdr->rc = supdrvIOCtl_TracerUmodRegister(pDevExt, pSession,
1825	pReq->u.In.R3PtrVtgHdr, pReq->u.In.uVtgHdrAddr,
1826	pReq->u.In.R3PtrStrTab, pReq->u.In.cbStrTab,
1827	pReq->u.In.szName, pReq->u.In.fFlags);
1828	return 0;
1829	}
1830
1831	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_DEREG):
1832	{
1833	/* validate */
1834	PSUPTRACERUMODDEREG pReq = (PSUPTRACERUMODDEREG)pReqHdr;
1835	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_DEREG);
1836
1837	/* execute */
1838	pReqHdr->rc = supdrvIOCtl_TracerUmodDeregister(pDevExt, pSession, pReq->u.In.pVtgHdr);
1839	return 0;
1840	}
1841
1842	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_FIRE_PROBE):
1843	{
1844	/* validate */
1845	PSUPTRACERUMODFIREPROBE pReq = (PSUPTRACERUMODFIREPROBE)pReqHdr;
1846	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_FIRE_PROBE);
1847
1848	supdrvIOCtl_TracerUmodProbeFire(pDevExt, pSession, &pReq->u.In);
1849	pReqHdr->rc = VINF_SUCCESS;
1850	return 0;
1851	}
1852
1853	default:
1854	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1855	break;
1856	}
1857	return VERR_GENERAL_FAILURE;
1858	}
1859
1860
1861	/**
1862	* I/O Control inner worker for the restricted operations.
1863	*
1864	* @returns IPRT status code.
1865	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1866	*
1867	* @param uIOCtl Function number.
1868	* @param pDevExt Device extention.
1869	* @param pSession Session data.
1870	* @param pReqHdr The request header.
1871	*/
1872	static int supdrvIOCtlInnerRestricted(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1873	{
1874	/*
1875	* The switch.
1876	*/
1877	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
1878	{
1879	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
1880	{
1881	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
1882	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
1883	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
1884	{
1885	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
1886	pReq->Hdr.rc = VERR_INVALID_MAGIC;
1887	return 0;
1888	}
1889
1890	/*
1891	* Match the version.
1892	* The current logic is very simple, match the major interface version.
1893	*/
1894	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
1895	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
1896	{
1897	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1898	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
1899	pReq->u.Out.u32Cookie = 0xffffffff;
1900	pReq->u.Out.u32SessionCookie = 0xffffffff;
1901	pReq->u.Out.u32SessionVersion = 0xffffffff;
1902	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1903	pReq->u.Out.pSession = NULL;
1904	pReq->u.Out.cFunctions = 0;
1905	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1906	return 0;
1907	}
1908
1909	/*
1910	* Fill in return data and be gone.
1911	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
1912	* u32SessionVersion <= u32ReqVersion!
1913	*/
1914	/** @todo Somehow validate the client and negotiate a secure cookie... */
1915	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
1916	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
1917	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
1918	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1919	pReq->u.Out.pSession = pSession;
1920	pReq->u.Out.cFunctions = 0;
1921	pReq->Hdr.rc = VINF_SUCCESS;
1922	return 0;
1923	}
1924
1925	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
1926	{
1927	/* validate */
1928	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
1929	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
1930
1931	/* execute */
1932	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.Caps);
1933	if (RT_FAILURE(pReq->Hdr.rc))
1934	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1935	return 0;
1936	}
1937
1938	default:
1939	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1940	break;
1941	}
1942	return VERR_GENERAL_FAILURE;
1943	}
1944
1945
1946	/**
1947	* I/O Control worker.
1948	*
1949	* @returns IPRT status code.
1950	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1951	*
1952	* @param uIOCtl Function number.
1953	* @param pDevExt Device extention.
1954	* @param pSession Session data.
1955	* @param pReqHdr The request header.
1956	*/
1957	int VBOXCALL supdrvIOCtl(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1958	{
1959	int rc;
1960	VBOXDRV_IOCTL_ENTRY(pSession, uIOCtl, pReqHdr);
1961
1962	/*
1963	* Validate the request.
1964	*/
1965	/* this first check could probably be omitted as its also done by the OS specific code... */
1966	if (RT_UNLIKELY( (pReqHdr->fFlags & SUPREQHDR_FLAGS_MAGIC_MASK) != SUPREQHDR_FLAGS_MAGIC
1967	\|\| pReqHdr->cbIn < sizeof(*pReqHdr)
1968	\|\| pReqHdr->cbOut < sizeof(*pReqHdr)))
1969	{
1970	OSDBGPRINT(("vboxdrv: Bad ioctl request header; cbIn=%#lx cbOut=%#lx fFlags=%#lx\n",
1971	(long)pReqHdr->cbIn, (long)pReqHdr->cbOut, (long)pReqHdr->fFlags));
1972	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1973	return VERR_INVALID_PARAMETER;
1974	}
1975	if (RT_UNLIKELY(!RT_VALID_PTR(pSession)))
1976	{
1977	OSDBGPRINT(("vboxdrv: Invalid pSession valud %p (ioctl=%p)\n", pSession, (void *)uIOCtl));
1978	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1979	return VERR_INVALID_PARAMETER;
1980	}
1981	if (RT_UNLIKELY(uIOCtl == SUP_IOCTL_COOKIE))
1982	{
1983	if (pReqHdr->u32Cookie != SUPCOOKIE_INITIAL_COOKIE)
1984	{
1985	OSDBGPRINT(("SUP_IOCTL_COOKIE: bad cookie %#lx\n", (long)pReqHdr->u32Cookie));
1986	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1987	return VERR_INVALID_PARAMETER;
1988	}
1989	}
1990	else if (RT_UNLIKELY( pReqHdr->u32Cookie != pDevExt->u32Cookie
1991	\|\| pReqHdr->u32SessionCookie != pSession->u32Cookie))
1992	{
1993	OSDBGPRINT(("vboxdrv: bad cookie %#lx / %#lx.\n", (long)pReqHdr->u32Cookie, (long)pReqHdr->u32SessionCookie));
1994	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1995	return VERR_INVALID_PARAMETER;
1996	}
1997
1998	/*
1999	* Hand it to an inner function to avoid lots of unnecessary return tracepoints.
2000	*/
2001	if (pSession->fUnrestricted)
2002	rc = supdrvIOCtlInnerUnrestricted(uIOCtl, pDevExt, pSession, pReqHdr);
2003	else
2004	rc = supdrvIOCtlInnerRestricted(uIOCtl, pDevExt, pSession, pReqHdr);
2005
2006	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, pReqHdr->rc, rc);
2007	return rc;
2008	}
2009
2010
2011	/**
2012	* Inter-Driver Communication (IDC) worker.
2013	*
2014	* @returns VBox status code.
2015	* @retval VINF_SUCCESS on success.
2016	* @retval VERR_INVALID_PARAMETER if the request is invalid.
2017	* @retval VERR_NOT_SUPPORTED if the request isn't supported.
2018	*
2019	* @param uReq The request (function) code.
2020	* @param pDevExt Device extention.
2021	* @param pSession Session data.
2022	* @param pReqHdr The request header.
2023	*/
2024	int VBOXCALL supdrvIDC(uintptr_t uReq, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQHDR pReqHdr)
2025	{
2026	/*
2027	* The OS specific code has already validated the pSession
2028	* pointer, and the request size being greater or equal to
2029	* size of the header.
2030	*
2031	* So, just check that pSession is a kernel context session.
2032	*/
2033	if (RT_UNLIKELY( pSession
2034	&& pSession->R0Process != NIL_RTR0PROCESS))
2035	return VERR_INVALID_PARAMETER;
2036
2037	/*
2038	* Validation macro.
2039	*/
2040	#define REQ_CHECK_IDC_SIZE(Name, cbExpect) \
2041	do { \
2042	if (RT_UNLIKELY(pReqHdr->cb != (cbExpect))) \
2043	{ \
2044	OSDBGPRINT(( #Name ": Invalid input/output sizes. cb=%ld expected %ld.\n", \
2045	(long)pReqHdr->cb, (long)(cbExpect))); \
2046	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
2047	} \
2048	} while (0)
2049
2050	switch (uReq)
2051	{
2052	case SUPDRV_IDC_REQ_CONNECT:
2053	{
2054	PSUPDRVIDCREQCONNECT pReq = (PSUPDRVIDCREQCONNECT)pReqHdr;
2055	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_CONNECT, sizeof(*pReq));
2056
2057	/*
2058	* Validate the cookie and other input.
2059	*/
2060	if (pReq->Hdr.pSession != NULL)
2061	{
2062	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Hdr.pSession=%p expected NULL!\n", pReq->Hdr.pSession));
2063	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2064	}
2065	if (pReq->u.In.u32MagicCookie != SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE)
2066	{
2067	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: u32MagicCookie=%#x expected %#x!\n",
2068	(unsigned)pReq->u.In.u32MagicCookie, (unsigned)SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE));
2069	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2070	}
2071	if ( pReq->u.In.uMinVersion > pReq->u.In.uReqVersion
2072	\|\| (pReq->u.In.uMinVersion & UINT32_C(0xffff0000)) != (pReq->u.In.uReqVersion & UINT32_C(0xffff0000)))
2073	{
2074	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: uMinVersion=%#x uMaxVersion=%#x doesn't match!\n",
2075	pReq->u.In.uMinVersion, pReq->u.In.uReqVersion));
2076	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2077	}
2078	if (pSession != NULL)
2079	{
2080	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: pSession=%p expected NULL!\n", pSession));
2081	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2082	}
2083
2084	/*
2085	* Match the version.
2086	* The current logic is very simple, match the major interface version.
2087	*/
2088	if ( pReq->u.In.uMinVersion > SUPDRV_IDC_VERSION
2089	\|\| (pReq->u.In.uMinVersion & 0xffff0000) != (SUPDRV_IDC_VERSION & 0xffff0000))
2090	{
2091	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
2092	pReq->u.In.uReqVersion, pReq->u.In.uMinVersion, (unsigned)SUPDRV_IDC_VERSION));
2093	pReq->u.Out.pSession = NULL;
2094	pReq->u.Out.uSessionVersion = 0xffffffff;
2095	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
2096	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
2097	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
2098	return VINF_SUCCESS;
2099	}
2100
2101	pReq->u.Out.pSession = NULL;
2102	pReq->u.Out.uSessionVersion = SUPDRV_IDC_VERSION;
2103	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
2104	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
2105
2106	pReq->Hdr.rc = supdrvCreateSession(pDevExt, false /* fUser /, true /fUnrestricted*/, &pSession);
2107	if (RT_FAILURE(pReq->Hdr.rc))
2108	{
2109	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: failed to create session, rc=%d\n", pReq->Hdr.rc));
2110	return VINF_SUCCESS;
2111	}
2112
2113	pReq->u.Out.pSession = pSession;
2114	pReq->Hdr.pSession = pSession;
2115
2116	return VINF_SUCCESS;
2117	}
2118
2119	case SUPDRV_IDC_REQ_DISCONNECT:
2120	{
2121	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_DISCONNECT, sizeof(*pReqHdr));
2122
2123	supdrvCloseSession(pDevExt, pSession);
2124	return pReqHdr->rc = VINF_SUCCESS;
2125	}
2126
2127	case SUPDRV_IDC_REQ_GET_SYMBOL:
2128	{
2129	PSUPDRVIDCREQGETSYM pReq = (PSUPDRVIDCREQGETSYM)pReqHdr;
2130	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_GET_SYMBOL, sizeof(*pReq));
2131
2132	pReq->Hdr.rc = supdrvIDC_LdrGetSymbol(pDevExt, pSession, pReq);
2133	return VINF_SUCCESS;
2134	}
2135
2136	case SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY:
2137	{
2138	PSUPDRVIDCREQCOMPREGFACTORY pReq = (PSUPDRVIDCREQCOMPREGFACTORY)pReqHdr;
2139	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY, sizeof(*pReq));
2140
2141	pReq->Hdr.rc = SUPR0ComponentRegisterFactory(pSession, pReq->u.In.pFactory);
2142	return VINF_SUCCESS;
2143	}
2144
2145	case SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY:
2146	{
2147	PSUPDRVIDCREQCOMPDEREGFACTORY pReq = (PSUPDRVIDCREQCOMPDEREGFACTORY)pReqHdr;
2148	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY, sizeof(*pReq));
2149
2150	pReq->Hdr.rc = SUPR0ComponentDeregisterFactory(pSession, pReq->u.In.pFactory);
2151	return VINF_SUCCESS;
2152	}
2153
2154	default:
2155	Log(("Unknown IDC %#lx\n", (long)uReq));
2156	break;
2157	}
2158
2159	#undef REQ_CHECK_IDC_SIZE
2160	return VERR_NOT_SUPPORTED;
2161	}
2162
2163
2164	/**
2165	* Register a object for reference counting.
2166	* The object is registered with one reference in the specified session.
2167	*
2168	* @returns Unique identifier on success (pointer).
2169	* All future reference must use this identifier.
2170	* @returns NULL on failure.
2171	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
2172	* @param pvUser1 The first user argument.
2173	* @param pvUser2 The second user argument.
2174	*/
2175	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
2176	{
2177	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2178	PSUPDRVOBJ pObj;
2179	PSUPDRVUSAGE pUsage;
2180
2181	/*
2182	* Validate the input.
2183	*/
2184	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
2185	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
2186	AssertPtrReturn(pfnDestructor, NULL);
2187
2188	/*
2189	* Allocate and initialize the object.
2190	*/
2191	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
2192	if (!pObj)
2193	return NULL;
2194	pObj->u32Magic = SUPDRVOBJ_MAGIC;
2195	pObj->enmType = enmType;
2196	pObj->pNext = NULL;
2197	pObj->cUsage = 1;
2198	pObj->pfnDestructor = pfnDestructor;
2199	pObj->pvUser1 = pvUser1;
2200	pObj->pvUser2 = pvUser2;
2201	pObj->CreatorUid = pSession->Uid;
2202	pObj->CreatorGid = pSession->Gid;
2203	pObj->CreatorProcess= pSession->Process;
2204	supdrvOSObjInitCreator(pObj, pSession);
2205
2206	/*
2207	* Allocate the usage record.
2208	* (We keep freed usage records around to simplify SUPR0ObjAddRefEx().)
2209	*/
2210	RTSpinlockAcquire(pDevExt->Spinlock);
2211
2212	pUsage = pDevExt->pUsageFree;
2213	if (pUsage)
2214	pDevExt->pUsageFree = pUsage->pNext;
2215	else
2216	{
2217	RTSpinlockRelease(pDevExt->Spinlock);
2218	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
2219	if (!pUsage)
2220	{
2221	RTMemFree(pObj);
2222	return NULL;
2223	}
2224	RTSpinlockAcquire(pDevExt->Spinlock);
2225	}
2226
2227	/*
2228	* Insert the object and create the session usage record.
2229	*/
2230	/* The object. */
2231	pObj->pNext = pDevExt->pObjs;
2232	pDevExt->pObjs = pObj;
2233
2234	/* The session record. */
2235	pUsage->cUsage = 1;
2236	pUsage->pObj = pObj;
2237	pUsage->pNext = pSession->pUsage;
2238	/* Log2(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext)); */
2239	pSession->pUsage = pUsage;
2240
2241	RTSpinlockRelease(pDevExt->Spinlock);
2242
2243	Log(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
2244	return pObj;
2245	}
2246
2247
2248	/**
2249	* Increment the reference counter for the object associating the reference
2250	* with the specified session.
2251	*
2252	* @returns IPRT status code.
2253	* @param pvObj The identifier returned by SUPR0ObjRegister().
2254	* @param pSession The session which is referencing the object.
2255	*
2256	* @remarks The caller should not own any spinlocks and must carefully protect
2257	* itself against potential race with the destructor so freed memory
2258	* isn't accessed here.
2259	*/
2260	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
2261	{
2262	return SUPR0ObjAddRefEx(pvObj, pSession, false /* fNoBlocking */);
2263	}
2264
2265
2266	/**
2267	* Increment the reference counter for the object associating the reference
2268	* with the specified session.
2269	*
2270	* @returns IPRT status code.
2271	* @retval VERR_TRY_AGAIN if fNoBlocking was set and a new usage record
2272	* couldn't be allocated. (If you see this you're not doing the right
2273	* thing and it won't ever work reliably.)
2274	*
2275	* @param pvObj The identifier returned by SUPR0ObjRegister().
2276	* @param pSession The session which is referencing the object.
2277	* @param fNoBlocking Set if it's not OK to block. Never try to make the
2278	* first reference to an object in a session with this
2279	* argument set.
2280	*
2281	* @remarks The caller should not own any spinlocks and must carefully protect
2282	* itself against potential race with the destructor so freed memory
2283	* isn't accessed here.
2284	*/
2285	SUPR0DECL(int) SUPR0ObjAddRefEx(void *pvObj, PSUPDRVSESSION pSession, bool fNoBlocking)
2286	{
2287	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2288	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2289	int rc = VINF_SUCCESS;
2290	PSUPDRVUSAGE pUsagePre;
2291	PSUPDRVUSAGE pUsage;
2292
2293	/*
2294	* Validate the input.
2295	* Be ready for the destruction race (someone might be stuck in the
2296	* destructor waiting a lock we own).
2297	*/
2298	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2299	AssertPtrReturn(pObj, VERR_INVALID_POINTER);
2300	AssertMsgReturn(pObj->u32Magic == SUPDRVOBJ_MAGIC \|\| pObj->u32Magic == SUPDRVOBJ_MAGIC_DEAD,
2301	("Invalid pvObj=%p magic=%#x (expected %#x or %#x)\n", pvObj, pObj->u32Magic, SUPDRVOBJ_MAGIC, SUPDRVOBJ_MAGIC_DEAD),
2302	VERR_INVALID_PARAMETER);
2303
2304	RTSpinlockAcquire(pDevExt->Spinlock);
2305
2306	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2307	{
2308	RTSpinlockRelease(pDevExt->Spinlock);
2309
2310	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2311	return VERR_WRONG_ORDER;
2312	}
2313
2314	/*
2315	* Preallocate the usage record if we can.
2316	*/
2317	pUsagePre = pDevExt->pUsageFree;
2318	if (pUsagePre)
2319	pDevExt->pUsageFree = pUsagePre->pNext;
2320	else if (!fNoBlocking)
2321	{
2322	RTSpinlockRelease(pDevExt->Spinlock);
2323	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
2324	if (!pUsagePre)
2325	return VERR_NO_MEMORY;
2326
2327	RTSpinlockAcquire(pDevExt->Spinlock);
2328	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2329	{
2330	RTSpinlockRelease(pDevExt->Spinlock);
2331
2332	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2333	return VERR_WRONG_ORDER;
2334	}
2335	}
2336
2337	/*
2338	* Reference the object.
2339	*/
2340	pObj->cUsage++;
2341
2342	/*
2343	* Look for the session record.
2344	*/
2345	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
2346	{
2347	/Log(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2348	if (pUsage->pObj == pObj)
2349	break;
2350	}
2351	if (pUsage)
2352	pUsage->cUsage++;
2353	else if (pUsagePre)
2354	{
2355	/* create a new session record. */
2356	pUsagePre->cUsage = 1;
2357	pUsagePre->pObj = pObj;
2358	pUsagePre->pNext = pSession->pUsage;
2359	pSession->pUsage = pUsagePre;
2360	/Log(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));/
2361
2362	pUsagePre = NULL;
2363	}
2364	else
2365	{
2366	pObj->cUsage--;
2367	rc = VERR_TRY_AGAIN;
2368	}
2369
2370	/*
2371	* Put any unused usage record into the free list..
2372	*/
2373	if (pUsagePre)
2374	{
2375	pUsagePre->pNext = pDevExt->pUsageFree;
2376	pDevExt->pUsageFree = pUsagePre;
2377	}
2378
2379	RTSpinlockRelease(pDevExt->Spinlock);
2380
2381	return rc;
2382	}
2383
2384
2385	/**
2386	* Decrement / destroy a reference counter record for an object.
2387	*
2388	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
2389	*
2390	* @returns IPRT status code.
2391	* @retval VINF_SUCCESS if not destroyed.
2392	* @retval VINF_OBJECT_DESTROYED if it's destroyed by this release call.
2393	* @retval VERR_INVALID_PARAMETER if the object isn't valid. Will assert in
2394	* string builds.
2395	*
2396	* @param pvObj The identifier returned by SUPR0ObjRegister().
2397	* @param pSession The session which is referencing the object.
2398	*/
2399	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
2400	{
2401	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2402	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2403	int rc = VERR_INVALID_PARAMETER;
2404	PSUPDRVUSAGE pUsage;
2405	PSUPDRVUSAGE pUsagePrev;
2406
2407	/*
2408	* Validate the input.
2409	*/
2410	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2411	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2412	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2413	VERR_INVALID_PARAMETER);
2414
2415	/*
2416	* Acquire the spinlock and look for the usage record.
2417	*/
2418	RTSpinlockAcquire(pDevExt->Spinlock);
2419
2420	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
2421	pUsage;
2422	pUsagePrev = pUsage, pUsage = pUsage->pNext)
2423	{
2424	/Log2(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2425	if (pUsage->pObj == pObj)
2426	{
2427	rc = VINF_SUCCESS;
2428	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
2429	if (pUsage->cUsage > 1)
2430	{
2431	pObj->cUsage--;
2432	pUsage->cUsage--;
2433	}
2434	else
2435	{
2436	/*
2437	* Free the session record.
2438	*/
2439	if (pUsagePrev)
2440	pUsagePrev->pNext = pUsage->pNext;
2441	else
2442	pSession->pUsage = pUsage->pNext;
2443	pUsage->pNext = pDevExt->pUsageFree;
2444	pDevExt->pUsageFree = pUsage;
2445
2446	/* What about the object? */
2447	if (pObj->cUsage > 1)
2448	pObj->cUsage--;
2449	else
2450	{
2451	/*
2452	* Object is to be destroyed, unlink it.
2453	*/
2454	pObj->u32Magic = SUPDRVOBJ_MAGIC_DEAD;
2455	rc = VINF_OBJECT_DESTROYED;
2456	if (pDevExt->pObjs == pObj)
2457	pDevExt->pObjs = pObj->pNext;
2458	else
2459	{
2460	PSUPDRVOBJ pObjPrev;
2461	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
2462	if (pObjPrev->pNext == pObj)
2463	{
2464	pObjPrev->pNext = pObj->pNext;
2465	break;
2466	}
2467	Assert(pObjPrev);
2468	}
2469	}
2470	}
2471	break;
2472	}
2473	}
2474
2475	RTSpinlockRelease(pDevExt->Spinlock);
2476
2477	/*
2478	* Call the destructor and free the object if required.
2479	*/
2480	if (rc == VINF_OBJECT_DESTROYED)
2481	{
2482	Log(("SUPR0ObjRelease: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
2483	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
2484	if (pObj->pfnDestructor)
2485	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
2486	RTMemFree(pObj);
2487	}
2488
2489	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
2490	return rc;
2491	}
2492
2493
2494	/**
2495	* Verifies that the current process can access the specified object.
2496	*
2497	* @returns The following IPRT status code:
2498	* @retval VINF_SUCCESS if access was granted.
2499	* @retval VERR_PERMISSION_DENIED if denied access.
2500	* @retval VERR_INVALID_PARAMETER if invalid parameter.
2501	*
2502	* @param pvObj The identifier returned by SUPR0ObjRegister().
2503	* @param pSession The session which wishes to access the object.
2504	* @param pszObjName Object string name. This is optional and depends on the object type.
2505	*
2506	* @remark The caller is responsible for making sure the object isn't removed while
2507	* we're inside this function. If uncertain about this, just call AddRef before calling us.
2508	*/
2509	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
2510	{
2511	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2512	int rc;
2513
2514	/*
2515	* Validate the input.
2516	*/
2517	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2518	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2519	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2520	VERR_INVALID_PARAMETER);
2521
2522	/*
2523	* Check access. (returns true if a decision has been made.)
2524	*/
2525	rc = VERR_INTERNAL_ERROR;
2526	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
2527	return rc;
2528
2529	/*
2530	* Default policy is to allow the user to access his own
2531	* stuff but nothing else.
2532	*/
2533	if (pObj->CreatorUid == pSession->Uid)
2534	return VINF_SUCCESS;
2535	return VERR_PERMISSION_DENIED;
2536	}
2537
2538
2539	/**
2540	* Lock pages.
2541	*
2542	* @returns IPRT status code.
2543	* @param pSession Session to which the locked memory should be associated.
2544	* @param pvR3 Start of the memory range to lock.
2545	* This must be page aligned.
2546	* @param cPages Number of pages to lock.
2547	* @param paPages Where to put the physical addresses of locked memory.
2548	*/
2549	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
2550	{
2551	int rc;
2552	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2553	const size_t cb = (size_t)cPages << PAGE_SHIFT;
2554	LogFlow(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
2555
2556	/*
2557	* Verify input.
2558	*/
2559	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2560	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
2561	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
2562	\|\| !pvR3)
2563	{
2564	Log(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
2565	return VERR_INVALID_PARAMETER;
2566	}
2567
2568	/*
2569	* Let IPRT do the job.
2570	*/
2571	Mem.eType = MEMREF_TYPE_LOCKED;
2572	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTMEM_PROT_READ \| RTMEM_PROT_WRITE, RTR0ProcHandleSelf());
2573	if (RT_SUCCESS(rc))
2574	{
2575	uint32_t iPage = cPages;
2576	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
2577	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
2578
2579	while (iPage-- > 0)
2580	{
2581	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2582	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
2583	{
2584	AssertMsgFailed(("iPage=%d\n", iPage));
2585	rc = VERR_INTERNAL_ERROR;
2586	break;
2587	}
2588	}
2589	if (RT_SUCCESS(rc))
2590	rc = supdrvMemAdd(&Mem, pSession);
2591	if (RT_FAILURE(rc))
2592	{
2593	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
2594	AssertRC(rc2);
2595	}
2596	}
2597
2598	return rc;
2599	}
2600
2601
2602	/**
2603	* Unlocks the memory pointed to by pv.
2604	*
2605	* @returns IPRT status code.
2606	* @param pSession Session to which the memory was locked.
2607	* @param pvR3 Memory to unlock.
2608	*/
2609	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2610	{
2611	LogFlow(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2612	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2613	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
2614	}
2615
2616
2617	/**
2618	* Allocates a chunk of page aligned memory with contiguous and fixed physical
2619	* backing.
2620	*
2621	* @returns IPRT status code.
2622	* @param pSession Session data.
2623	* @param cPages Number of pages to allocate.
2624	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
2625	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
2626	* @param pHCPhys Where to put the physical address of allocated memory.
2627	*/
2628	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
2629	{
2630	int rc;
2631	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2632	LogFlow(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
2633
2634	/*
2635	* Validate input.
2636	*/
2637	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2638	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
2639	{
2640	Log(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
2641	pSession, ppvR0, ppvR3, pHCPhys));
2642	return VERR_INVALID_PARAMETER;
2643
2644	}
2645	if (cPages < 1 \|\| cPages >= 256)
2646	{
2647	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2648	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2649	}
2650
2651	/*
2652	* Let IPRT do the job.
2653	*/
2654	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
2655	if (RT_SUCCESS(rc))
2656	{
2657	int rc2;
2658	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2659	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2660	if (RT_SUCCESS(rc))
2661	{
2662	Mem.eType = MEMREF_TYPE_CONT;
2663	rc = supdrvMemAdd(&Mem, pSession);
2664	if (!rc)
2665	{
2666	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2667	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2668	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
2669	return 0;
2670	}
2671
2672	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2673	AssertRC(rc2);
2674	}
2675	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2676	AssertRC(rc2);
2677	}
2678
2679	return rc;
2680	}
2681
2682
2683	/**
2684	* Frees memory allocated using SUPR0ContAlloc().
2685	*
2686	* @returns IPRT status code.
2687	* @param pSession The session to which the memory was allocated.
2688	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2689	*/
2690	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2691	{
2692	LogFlow(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2693	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2694	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
2695	}
2696
2697
2698	/**
2699	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
2700	*
2701	* The memory isn't zeroed.
2702	*
2703	* @returns IPRT status code.
2704	* @param pSession Session data.
2705	* @param cPages Number of pages to allocate.
2706	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
2707	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
2708	* @param paPages Where to put the physical addresses of allocated memory.
2709	*/
2710	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
2711	{
2712	unsigned iPage;
2713	int rc;
2714	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2715	LogFlow(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
2716
2717	/*
2718	* Validate input.
2719	*/
2720	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2721	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
2722	{
2723	Log(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
2724	pSession, ppvR3, ppvR0, paPages));
2725	return VERR_INVALID_PARAMETER;
2726
2727	}
2728	if (cPages < 1 \|\| cPages >= 256)
2729	{
2730	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2731	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2732	}
2733
2734	/*
2735	* Let IPRT do the work.
2736	*/
2737	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
2738	if (RT_SUCCESS(rc))
2739	{
2740	int rc2;
2741	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2742	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2743	if (RT_SUCCESS(rc))
2744	{
2745	Mem.eType = MEMREF_TYPE_LOW;
2746	rc = supdrvMemAdd(&Mem, pSession);
2747	if (!rc)
2748	{
2749	for (iPage = 0; iPage < cPages; iPage++)
2750	{
2751	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2752	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%RHp\n", paPages[iPage]));
2753	}
2754	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2755	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2756	return 0;
2757	}
2758
2759	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2760	AssertRC(rc2);
2761	}
2762
2763	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2764	AssertRC(rc2);
2765	}
2766
2767	return rc;
2768	}
2769
2770
2771	/**
2772	* Frees memory allocated using SUPR0LowAlloc().
2773	*
2774	* @returns IPRT status code.
2775	* @param pSession The session to which the memory was allocated.
2776	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2777	*/
2778	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2779	{
2780	LogFlow(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2781	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2782	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
2783	}
2784
2785
2786
2787	/**
2788	* Allocates a chunk of memory with both R0 and R3 mappings.
2789	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
2790	*
2791	* @returns IPRT status code.
2792	* @param pSession The session to associated the allocation with.
2793	* @param cb Number of bytes to allocate.
2794	* @param ppvR0 Where to store the address of the Ring-0 mapping.
2795	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2796	*/
2797	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
2798	{
2799	int rc;
2800	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2801	LogFlow(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
2802
2803	/*
2804	* Validate input.
2805	*/
2806	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2807	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
2808	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
2809	if (cb < 1 \|\| cb >= _4M)
2810	{
2811	Log(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
2812	return VERR_INVALID_PARAMETER;
2813	}
2814
2815	/*
2816	* Let IPRT do the work.
2817	*/
2818	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
2819	if (RT_SUCCESS(rc))
2820	{
2821	int rc2;
2822	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2823	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2824	if (RT_SUCCESS(rc))
2825	{
2826	Mem.eType = MEMREF_TYPE_MEM;
2827	rc = supdrvMemAdd(&Mem, pSession);
2828	if (!rc)
2829	{
2830	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2831	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2832	return VINF_SUCCESS;
2833	}
2834
2835	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2836	AssertRC(rc2);
2837	}
2838
2839	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2840	AssertRC(rc2);
2841	}
2842
2843	return rc;
2844	}
2845
2846
2847	/**
2848	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
2849	*
2850	* @returns IPRT status code.
2851	* @param pSession The session to which the memory was allocated.
2852	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2853	* @param paPages Where to store the physical addresses.
2854	*/
2855	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
2856	{
2857	PSUPDRVBUNDLE pBundle;
2858	LogFlow(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
2859
2860	/*
2861	* Validate input.
2862	*/
2863	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2864	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
2865	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
2866
2867	/*
2868	* Search for the address.
2869	*/
2870	RTSpinlockAcquire(pSession->Spinlock);
2871	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2872	{
2873	if (pBundle->cUsed > 0)
2874	{
2875	unsigned i;
2876	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2877	{
2878	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
2879	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2880	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2881	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2882	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
2883	)
2884	)
2885	{
2886	const size_t cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2887	size_t iPage;
2888	for (iPage = 0; iPage < cPages; iPage++)
2889	{
2890	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2891	paPages[iPage].uReserved = 0;
2892	}
2893	RTSpinlockRelease(pSession->Spinlock);
2894	return VINF_SUCCESS;
2895	}
2896	}
2897	}
2898	}
2899	RTSpinlockRelease(pSession->Spinlock);
2900	Log(("Failed to find %p!!!\n", (void *)uPtr));
2901	return VERR_INVALID_PARAMETER;
2902	}
2903
2904
2905	/**
2906	* Free memory allocated by SUPR0MemAlloc().
2907	*
2908	* @returns IPRT status code.
2909	* @param pSession The session owning the allocation.
2910	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2911	*/
2912	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2913	{
2914	LogFlow(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2915	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2916	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
2917	}
2918
2919
2920	/**
2921	* Allocates a chunk of memory with a kernel or/and a user mode mapping.
2922	*
2923	* The memory is fixed and it's possible to query the physical addresses using
2924	* SUPR0MemGetPhys().
2925	*
2926	* @returns IPRT status code.
2927	* @param pSession The session to associated the allocation with.
2928	* @param cPages The number of pages to allocate.
2929	* @param fFlags Flags, reserved for the future. Must be zero.
2930	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2931	* NULL if no ring-3 mapping.
2932	* @param ppvR3 Where to store the address of the Ring-0 mapping.
2933	* NULL if no ring-0 mapping.
2934	* @param paPages Where to store the addresses of the pages. Optional.
2935	*/
2936	SUPR0DECL(int) SUPR0PageAllocEx(PSUPDRVSESSION pSession, uint32_t cPages, uint32_t fFlags, PRTR3PTR ppvR3, PRTR0PTR ppvR0, PRTHCPHYS paPages)
2937	{
2938	int rc;
2939	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2940	LogFlow(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
2941
2942	/*
2943	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2944	*/
2945	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2946	AssertPtrNullReturn(ppvR3, VERR_INVALID_POINTER);
2947	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2948	AssertReturn(ppvR3 \|\| ppvR0, VERR_INVALID_PARAMETER);
2949	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2950	if (cPages < 1 \|\| cPages > VBOX_MAX_ALLOC_PAGE_COUNT)
2951	{
2952	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)));
2953	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2954	}
2955
2956	/*
2957	* Let IPRT do the work.
2958	*/
2959	if (ppvR0)
2960	rc = RTR0MemObjAllocPage(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, true /* fExecutable */);
2961	else
2962	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
2963	if (RT_SUCCESS(rc))
2964	{
2965	int rc2;
2966	if (ppvR3)
2967	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2968	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2969	else
2970	Mem.MapObjR3 = NIL_RTR0MEMOBJ;
2971	if (RT_SUCCESS(rc))
2972	{
2973	Mem.eType = MEMREF_TYPE_PAGE;
2974	rc = supdrvMemAdd(&Mem, pSession);
2975	if (!rc)
2976	{
2977	if (ppvR3)
2978	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2979	if (ppvR0)
2980	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2981	if (paPages)
2982	{
2983	uint32_t iPage = cPages;
2984	while (iPage-- > 0)
2985	{
2986	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
2987	Assert(paPages[iPage] != NIL_RTHCPHYS);
2988	}
2989	}
2990	return VINF_SUCCESS;
2991	}
2992
2993	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2994	AssertRC(rc2);
2995	}
2996
2997	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2998	AssertRC(rc2);
2999	}
3000	return rc;
3001	}
3002
3003
3004	/**
3005	* Maps a chunk of memory previously allocated by SUPR0PageAllocEx into kernel
3006	* space.
3007	*
3008	* @returns IPRT status code.
3009	* @param pSession The session to associated the allocation with.
3010	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
3011	* @param offSub Where to start mapping. Must be page aligned.
3012	* @param cbSub How much to map. Must be page aligned.
3013	* @param fFlags Flags, MBZ.
3014	* @param ppvR0 Where to return the address of the ring-0 mapping on
3015	* success.
3016	*/
3017	SUPR0DECL(int) SUPR0PageMapKernel(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t offSub, uint32_t cbSub,
3018	uint32_t fFlags, PRTR0PTR ppvR0)
3019	{
3020	int rc;
3021	PSUPDRVBUNDLE pBundle;
3022	RTR0MEMOBJ hMemObj = NIL_RTR0MEMOBJ;
3023	LogFlow(("SUPR0PageMapKernel: pSession=%p pvR3=%p offSub=%#x cbSub=%#x\n", pSession, pvR3, offSub, cbSub));
3024
3025	/*
3026	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
3027	*/
3028	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3029	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
3030	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
3031	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3032	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3033	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
3034
3035	/*
3036	* Find the memory object.
3037	*/
3038	RTSpinlockAcquire(pSession->Spinlock);
3039	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3040	{
3041	if (pBundle->cUsed > 0)
3042	{
3043	unsigned i;
3044	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3045	{
3046	if ( ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
3047	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3048	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3049	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
3050	\|\| ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED
3051	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3052	&& pBundle->aMem[i].MapObjR3 == NIL_RTR0MEMOBJ
3053	&& RTR0MemObjAddressR3(pBundle->aMem[i].MemObj) == pvR3))
3054	{
3055	hMemObj = pBundle->aMem[i].MemObj;
3056	break;
3057	}
3058	}
3059	}
3060	}
3061	RTSpinlockRelease(pSession->Spinlock);
3062
3063	rc = VERR_INVALID_PARAMETER;
3064	if (hMemObj != NIL_RTR0MEMOBJ)
3065	{
3066	/*
3067	* Do some further input validations before calling IPRT.
3068	* (Cleanup is done indirectly by telling RTR0MemObjFree to include mappings.)
3069	*/
3070	size_t cbMemObj = RTR0MemObjSize(hMemObj);
3071	if ( offSub < cbMemObj
3072	&& cbSub <= cbMemObj
3073	&& offSub + cbSub <= cbMemObj)
3074	{
3075	RTR0MEMOBJ hMapObj;
3076	rc = RTR0MemObjMapKernelEx(&hMapObj, hMemObj, (void *)-1, 0,
3077	RTMEM_PROT_READ \| RTMEM_PROT_WRITE, offSub, cbSub);
3078	if (RT_SUCCESS(rc))
3079	*ppvR0 = RTR0MemObjAddress(hMapObj);
3080	}
3081	else
3082	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
3083
3084	}
3085	return rc;
3086	}
3087
3088
3089	/**
3090	* Changes the page level protection of one or more pages previously allocated
3091	* by SUPR0PageAllocEx.
3092	*
3093	* @returns IPRT status code.
3094	* @param pSession The session to associated the allocation with.
3095	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
3096	* NIL_RTR3PTR if the ring-3 mapping should be unaffected.
3097	* @param pvR0 The ring-0 address returned by SUPR0PageAllocEx.
3098	* NIL_RTR0PTR if the ring-0 mapping should be unaffected.
3099	* @param offSub Where to start changing. Must be page aligned.
3100	* @param cbSub How much to change. Must be page aligned.
3101	* @param fProt The new page level protection, see RTMEM_PROT_*.
3102	*/
3103	SUPR0DECL(int) SUPR0PageProtect(PSUPDRVSESSION pSession, RTR3PTR pvR3, RTR0PTR pvR0, uint32_t offSub, uint32_t cbSub, uint32_t fProt)
3104	{
3105	int rc;
3106	PSUPDRVBUNDLE pBundle;
3107	RTR0MEMOBJ hMemObjR0 = NIL_RTR0MEMOBJ;
3108	RTR0MEMOBJ hMemObjR3 = NIL_RTR0MEMOBJ;
3109	LogFlow(("SUPR0PageProtect: pSession=%p pvR3=%p pvR0=%p offSub=%#x cbSub=%#x fProt-%#x\n", pSession, pvR3, pvR0, offSub, cbSub, fProt));
3110
3111	/*
3112	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
3113	*/
3114	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3115	AssertReturn(!(fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)), VERR_INVALID_PARAMETER);
3116	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3117	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3118	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
3119
3120	/*
3121	* Find the memory object.
3122	*/
3123	RTSpinlockAcquire(pSession->Spinlock);
3124	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3125	{
3126	if (pBundle->cUsed > 0)
3127	{
3128	unsigned i;
3129	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3130	{
3131	if ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
3132	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3133	&& ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3134	\|\| pvR3 == NIL_RTR3PTR)
3135	&& ( pvR0 == NIL_RTR0PTR
3136	\|\| RTR0MemObjAddress(pBundle->aMem[i].MemObj) == pvR0)
3137	&& ( pvR3 == NIL_RTR3PTR
3138	\|\| RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3))
3139	{
3140	if (pvR0 != NIL_RTR0PTR)
3141	hMemObjR0 = pBundle->aMem[i].MemObj;
3142	if (pvR3 != NIL_RTR3PTR)
3143	hMemObjR3 = pBundle->aMem[i].MapObjR3;
3144	break;
3145	}
3146	}
3147	}
3148	}
3149	RTSpinlockRelease(pSession->Spinlock);
3150
3151	rc = VERR_INVALID_PARAMETER;
3152	if ( hMemObjR0 != NIL_RTR0MEMOBJ
3153	\|\| hMemObjR3 != NIL_RTR0MEMOBJ)
3154	{
3155	/*
3156	* Do some further input validations before calling IPRT.
3157	*/
3158	size_t cbMemObj = hMemObjR0 != NIL_RTR0PTR ? RTR0MemObjSize(hMemObjR0) : RTR0MemObjSize(hMemObjR3);
3159	if ( offSub < cbMemObj
3160	&& cbSub <= cbMemObj
3161	&& offSub + cbSub <= cbMemObj)
3162	{
3163	rc = VINF_SUCCESS;
3164	if (hMemObjR3 != NIL_RTR0PTR)
3165	rc = RTR0MemObjProtect(hMemObjR3, offSub, cbSub, fProt);
3166	if (hMemObjR0 != NIL_RTR0PTR && RT_SUCCESS(rc))
3167	rc = RTR0MemObjProtect(hMemObjR0, offSub, cbSub, fProt);
3168	}
3169	else
3170	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
3171
3172	}
3173	return rc;
3174
3175	}
3176
3177
3178	/**
3179	* Free memory allocated by SUPR0PageAlloc() and SUPR0PageAllocEx().
3180	*
3181	* @returns IPRT status code.
3182	* @param pSession The session owning the allocation.
3183	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc() or
3184	* SUPR0PageAllocEx().
3185	*/
3186	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
3187	{
3188	LogFlow(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
3189	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3190	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_PAGE);
3191	}
3192
3193
3194	/**
3195	* Gets the paging mode of the current CPU.
3196	*
3197	* @returns Paging mode, SUPPAGEINGMODE_INVALID on error.
3198	*/
3199	SUPR0DECL(SUPPAGINGMODE) SUPR0GetPagingMode(void)
3200	{
3201	SUPPAGINGMODE enmMode;
3202
3203	RTR0UINTREG cr0 = ASMGetCR0();
3204	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
3205	enmMode = SUPPAGINGMODE_INVALID;
3206	else
3207	{
3208	RTR0UINTREG cr4 = ASMGetCR4();
3209	uint32_t fNXEPlusLMA = 0;
3210	if (cr4 & X86_CR4_PAE)
3211	{
3212	uint32_t fExtFeatures = ASMCpuId_EDX(0x80000001);
3213	if (fExtFeatures & (X86_CPUID_EXT_FEATURE_EDX_NX \| X86_CPUID_EXT_FEATURE_EDX_LONG_MODE))
3214	{
3215	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
3216	if ((fExtFeatures & X86_CPUID_EXT_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
3217	fNXEPlusLMA \|= RT_BIT(0);
3218	if ((fExtFeatures & X86_CPUID_EXT_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
3219	fNXEPlusLMA \|= RT_BIT(1);
3220	}
3221	}
3222
3223	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
3224	{
3225	case 0:
3226	enmMode = SUPPAGINGMODE_32_BIT;
3227	break;
3228
3229	case X86_CR4_PGE:
3230	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
3231	break;
3232
3233	case X86_CR4_PAE:
3234	enmMode = SUPPAGINGMODE_PAE;
3235	break;
3236
3237	case X86_CR4_PAE \| RT_BIT(0):
3238	enmMode = SUPPAGINGMODE_PAE_NX;
3239	break;
3240
3241	case X86_CR4_PAE \| X86_CR4_PGE:
3242	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3243	break;
3244
3245	case X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3246	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3247	break;
3248
3249	case RT_BIT(1) \| X86_CR4_PAE:
3250	enmMode = SUPPAGINGMODE_AMD64;
3251	break;
3252
3253	case RT_BIT(1) \| X86_CR4_PAE \| RT_BIT(0):
3254	enmMode = SUPPAGINGMODE_AMD64_NX;
3255	break;
3256
3257	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
3258	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
3259	break;
3260
3261	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3262	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
3263	break;
3264
3265	default:
3266	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
3267	enmMode = SUPPAGINGMODE_INVALID;
3268	break;
3269	}
3270	}
3271	return enmMode;
3272	}
3273
3274
3275	/**
3276	* Enables or disabled hardware virtualization extensions using native OS APIs.
3277	*
3278	* @returns VBox status code.
3279	* @retval VINF_SUCCESS on success.
3280	* @retval VERR_NOT_SUPPORTED if not supported by the native OS.
3281	*
3282	* @param fEnable Whether to enable or disable.
3283	*/
3284	SUPR0DECL(int) SUPR0EnableVTx(bool fEnable)
3285	{
3286	#ifdef RT_OS_DARWIN
3287	return supdrvOSEnableVTx(fEnable);
3288	#else
3289	return VERR_NOT_SUPPORTED;
3290	#endif
3291	}
3292
3293
3294	/**
3295	* Suspends hardware virtualization extensions using the native OS API.
3296	*
3297	* This is called prior to entering raw-mode context.
3298	*
3299	* @returns @c true if suspended, @c false if not.
3300	*/
3301	SUPR0DECL(bool) SUPR0SuspendVTxOnCpu(void)
3302	{
3303	#ifdef RT_OS_DARWIN
3304	return supdrvOSSuspendVTxOnCpu();
3305	#else
3306	return false;
3307	#endif
3308	}
3309
3310
3311	/**
3312	* Resumes hardware virtualization extensions using the native OS API.
3313	*
3314	* This is called after to entering raw-mode context.
3315	*
3316	* @param fSuspended The return value of SUPR0SuspendVTxOnCpu.
3317	*/
3318	SUPR0DECL(void) SUPR0ResumeVTxOnCpu(bool fSuspended)
3319	{
3320	#ifdef RT_OS_DARWIN
3321	supdrvOSResumeVTxOnCpu(fSuspended);
3322	#else
3323	Assert(!fSuspended);
3324	#endif
3325	}
3326
3327
3328	/**
3329	* Quries the AMD-V and VT-x capabilities of the calling CPU.
3330	*
3331	* @returns VBox status code.
3332	* @retval VERR_VMX_NO_VMX
3333	* @retval VERR_VMX_MSR_LOCKED_OR_DISABLED
3334	* @retval VERR_SVM_NO_SVM
3335	* @retval VERR_SVM_DISABLED
3336	* @retval VERR_UNSUPPORTED_CPU if not identifiable as an AMD, Intel or VIA
3337	* (centaur) CPU.
3338	*
3339	* @param pSession The session handle.
3340	* @param pfCaps Where to store the capabilities.
3341	*/
3342	SUPR0DECL(int) SUPR0QueryVTCaps(PSUPDRVSESSION pSession, uint32_t *pfCaps)
3343	{
3344	/*
3345	* Input validation.
3346	*/
3347	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3348	AssertPtrReturn(pfCaps, VERR_INVALID_POINTER);
3349
3350	*pfCaps = 0;
3351
3352	if (ASMHasCpuId())
3353	{
3354	uint32_t fFeaturesECX, fFeaturesEDX, uDummy;
3355	uint32_t uMaxId, uVendorEBX, uVendorECX, uVendorEDX;
3356	uint64_t u64Value;
3357
3358	ASMCpuId(0, &uMaxId, &uVendorEBX, &uVendorECX, &uVendorEDX);
3359	ASMCpuId(1, &uDummy, &uDummy, &fFeaturesECX, &fFeaturesEDX);
3360
3361	if ( ASMIsValidStdRange(uMaxId)
3362	&& ( ASMIsIntelCpuEx( uVendorEBX, uVendorECX, uVendorEDX)
3363	\|\| ASMIsViaCentaurCpuEx(uVendorEBX, uVendorECX, uVendorEDX) )
3364	)
3365	{
3366	if ( (fFeaturesECX & X86_CPUID_FEATURE_ECX_VMX)
3367	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3368	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3369	)
3370	{
3371	/*
3372	* Both the LOCK and VMXON bit must be set; otherwise VMXON will generate a #GP.
3373	* Once the lock bit is set, this MSR can no longer be modified.
3374	*/
3375	u64Value = ASMRdMsr(MSR_IA32_FEATURE_CONTROL);
3376	if ( (u64Value & (MSR_IA32_FEATURE_CONTROL_VMXON \| MSR_IA32_FEATURE_CONTROL_LOCK))
3377	== (MSR_IA32_FEATURE_CONTROL_VMXON \| MSR_IA32_FEATURE_CONTROL_LOCK) /* enabled and locked */
3378	\|\| !(u64Value & MSR_IA32_FEATURE_CONTROL_LOCK) /* not enabled, but not locked either */
3379	)
3380	{
3381	VMX_CAPABILITY vtCaps;
3382
3383	*pfCaps \|= SUPVTCAPS_VT_X;
3384
3385	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS);
3386	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC_USE_SECONDARY_EXEC_CTRL)
3387	{
3388	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS2);
3389	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC2_EPT)
3390	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3391	}
3392	return VINF_SUCCESS;
3393	}
3394	return VERR_VMX_MSR_LOCKED_OR_DISABLED;
3395	}
3396	return VERR_VMX_NO_VMX;
3397	}
3398
3399	if ( ASMIsAmdCpuEx(uVendorEBX, uVendorECX, uVendorEDX)
3400	&& ASMIsValidStdRange(uMaxId))
3401	{
3402	uint32_t fExtFeaturesEcx, uExtMaxId;
3403	ASMCpuId(0x80000000, &uExtMaxId, &uDummy, &uDummy, &uDummy);
3404	ASMCpuId(0x80000001, &uDummy, &uDummy, &fExtFeaturesEcx, &uDummy);
3405	if ( ASMIsValidExtRange(uExtMaxId)
3406	&& uExtMaxId >= 0x8000000a
3407	&& (fExtFeaturesEcx & X86_CPUID_AMD_FEATURE_ECX_SVM)
3408	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3409	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3410	)
3411	{
3412	/* Check if SVM is disabled */
3413	u64Value = ASMRdMsr(MSR_K8_VM_CR);
3414	if (!(u64Value & MSR_K8_VM_CR_SVM_DISABLE))
3415	{
3416	uint32_t fSvmFeatures;
3417	*pfCaps \|= SUPVTCAPS_AMD_V;
3418
3419	/* Query AMD-V features. */
3420	ASMCpuId(0x8000000a, &uDummy, &uDummy, &uDummy, &fSvmFeatures);
3421	if (fSvmFeatures & AMD_CPUID_SVM_FEATURE_EDX_NESTED_PAGING)
3422	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3423
3424	return VINF_SUCCESS;
3425	}
3426	return VERR_SVM_DISABLED;
3427	}
3428	return VERR_SVM_NO_SVM;
3429	}
3430	}
3431
3432	return VERR_UNSUPPORTED_CPU;
3433	}
3434
3435
3436	/**
3437	* (Re-)initializes the per-cpu structure prior to starting or resuming the GIP
3438	* updating.
3439	*
3440	* @param pGipCpu The per CPU structure for this CPU.
3441	* @param u64NanoTS The current time.
3442	*/
3443	static void supdrvGipReInitCpu(PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
3444	{
3445	pGipCpu->u64TSC = ASMReadTSC() - pGipCpu->u32UpdateIntervalTSC;
3446	pGipCpu->u64NanoTS = u64NanoTS;
3447	}
3448
3449
3450	/**
3451	* Set the current TSC and NanoTS value for the CPU.
3452	*
3453	* @param idCpu The CPU ID. Unused - we have to use the APIC ID.
3454	* @param pvUser1 Pointer to the ring-0 GIP mapping.
3455	* @param pvUser2 Pointer to the variable holding the current time.
3456	*/
3457	static DECLCALLBACK(void) supdrvGipReInitCpuCallback(RTCPUID idCpu, void pvUser1, void pvUser2)
3458	{
3459	PSUPGLOBALINFOPAGE pGip = (PSUPGLOBALINFOPAGE)pvUser1;
3460	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
3461
3462	if (RT_LIKELY(iCpu < pGip->cCpus && pGip->aCPUs[iCpu].idCpu == idCpu))
3463	supdrvGipReInitCpu(&pGip->aCPUs[iCpu], (uint64_t )pvUser2);
3464
3465	NOREF(pvUser2);
3466	NOREF(idCpu);
3467	}
3468
3469
3470	/**
3471	* Maps the GIP into userspace and/or get the physical address of the GIP.
3472	*
3473	* @returns IPRT status code.
3474	* @param pSession Session to which the GIP mapping should belong.
3475	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
3476	* @param pHCPhysGip Where to store the physical address. (optional)
3477	*
3478	* @remark There is no reference counting on the mapping, so one call to this function
3479	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
3480	* and remove the session as a GIP user.
3481	*/
3482	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
3483	{
3484	int rc;
3485	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3486	RTR3PTR pGipR3 = NIL_RTR3PTR;
3487	RTHCPHYS HCPhys = NIL_RTHCPHYS;
3488	LogFlow(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
3489
3490	/*
3491	* Validate
3492	*/
3493	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3494	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
3495	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
3496
3497	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3498	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3499	#else
3500	RTSemFastMutexRequest(pDevExt->mtxGip);
3501	#endif
3502	if (pDevExt->pGip)
3503	{
3504	/*
3505	* Map it?
3506	*/
3507	rc = VINF_SUCCESS;
3508	if (ppGipR3)
3509	{
3510	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
3511	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
3512	RTMEM_PROT_READ, RTR0ProcHandleSelf());
3513	if (RT_SUCCESS(rc))
3514	pGipR3 = RTR0MemObjAddressR3(pSession->GipMapObjR3);
3515	}
3516
3517	/*
3518	* Get physical address.
3519	*/
3520	if (pHCPhysGip && RT_SUCCESS(rc))
3521	HCPhys = pDevExt->HCPhysGip;
3522
3523	/*
3524	* Reference globally.
3525	*/
3526	if (!pSession->fGipReferenced && RT_SUCCESS(rc))
3527	{
3528	pSession->fGipReferenced = 1;
3529	pDevExt->cGipUsers++;
3530	if (pDevExt->cGipUsers == 1)
3531	{
3532	PSUPGLOBALINFOPAGE pGipR0 = pDevExt->pGip;
3533	uint64_t u64NanoTS;
3534	uint32_t u32SystemResolution;
3535	unsigned i;
3536
3537	LogFlow(("SUPR0GipMap: Resumes GIP updating\n"));
3538
3539	/*
3540	* Try bump up the system timer resolution.
3541	* The more interrupts the better...
3542	*/
3543	if ( RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
3544	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
3545	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1953125 /* 512 HZ */, &u32SystemResolution))
3546	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 2000000 /* 500 HZ */, &u32SystemResolution))
3547	)
3548	{
3549	Assert(RTTimerGetSystemGranularity() <= u32SystemResolution);
3550	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
3551	}
3552
3553	if (pGipR0->aCPUs[0].u32TransactionId != 2 /* not the first time */)
3554	{
3555	for (i = 0; i < RT_ELEMENTS(pGipR0->aCPUs); i++)
3556	ASMAtomicUoWriteU32(&pGipR0->aCPUs[i].u32TransactionId,
3557	(pGipR0->aCPUs[i].u32TransactionId + GIP_UPDATEHZ_RECALC_FREQ * 2)
3558	& ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
3559	ASMAtomicWriteU64(&pGipR0->u64NanoTSLastUpdateHz, 0);
3560	}
3561
3562	u64NanoTS = RTTimeSystemNanoTS() - pGipR0->u32UpdateIntervalNS;
3563	if ( pGipR0->u32Mode == SUPGIPMODE_SYNC_TSC
3564	\|\| RTMpGetOnlineCount() == 1)
3565	supdrvGipReInitCpu(&pGipR0->aCPUs[0], u64NanoTS);
3566	else
3567	RTMpOnAll(supdrvGipReInitCpuCallback, pGipR0, &u64NanoTS);
3568
3569	#ifndef DO_NOT_START_GIP
3570	rc = RTTimerStart(pDevExt->pGipTimer, 0); AssertRC(rc);
3571	#endif
3572	rc = VINF_SUCCESS;
3573	}
3574	}
3575	}
3576	else
3577	{
3578	rc = VERR_GENERAL_FAILURE;
3579	Log(("SUPR0GipMap: GIP is not available!\n"));
3580	}
3581	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3582	RTSemMutexRelease(pDevExt->mtxGip);
3583	#else
3584	RTSemFastMutexRelease(pDevExt->mtxGip);
3585	#endif
3586
3587	/*
3588	* Write returns.
3589	*/
3590	if (pHCPhysGip)
3591	*pHCPhysGip = HCPhys;
3592	if (ppGipR3)
3593	*ppGipR3 = pGipR3;
3594
3595	#ifdef DEBUG_DARWIN_GIP
3596	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3597	#else
3598	LogFlow(( "SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3599	#endif
3600	return rc;
3601	}
3602
3603
3604	/**
3605	* Unmaps any user mapping of the GIP and terminates all GIP access
3606	* from this session.
3607	*
3608	* @returns IPRT status code.
3609	* @param pSession Session to which the GIP mapping should belong.
3610	*/
3611	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
3612	{
3613	int rc = VINF_SUCCESS;
3614	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3615	#ifdef DEBUG_DARWIN_GIP
3616	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
3617	pSession,
3618	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
3619	pSession->GipMapObjR3));
3620	#else
3621	LogFlow(("SUPR0GipUnmap: pSession=%p\n", pSession));
3622	#endif
3623	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3624
3625	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3626	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3627	#else
3628	RTSemFastMutexRequest(pDevExt->mtxGip);
3629	#endif
3630
3631	/*
3632	* Unmap anything?
3633	*/
3634	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
3635	{
3636	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
3637	AssertRC(rc);
3638	if (RT_SUCCESS(rc))
3639	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
3640	}
3641
3642	/*
3643	* Dereference global GIP.
3644	*/
3645	if (pSession->fGipReferenced && !rc)
3646	{
3647	pSession->fGipReferenced = 0;
3648	if ( pDevExt->cGipUsers > 0
3649	&& !--pDevExt->cGipUsers)
3650	{
3651	LogFlow(("SUPR0GipUnmap: Suspends GIP updating\n"));
3652	#ifndef DO_NOT_START_GIP
3653	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = VINF_SUCCESS;
3654	#endif
3655
3656	if (pDevExt->u32SystemTimerGranularityGrant)
3657	{
3658	int rc2 = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant);
3659	AssertRC(rc2);
3660	pDevExt->u32SystemTimerGranularityGrant = 0;
3661	}
3662	}
3663	}
3664
3665	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3666	RTSemMutexRelease(pDevExt->mtxGip);
3667	#else
3668	RTSemFastMutexRelease(pDevExt->mtxGip);
3669	#endif
3670
3671	return rc;
3672	}
3673
3674
3675	/**
3676	* Gets the GIP pointer.
3677	*
3678	* @returns Pointer to the GIP or NULL.
3679	*/
3680	SUPDECL(PSUPGLOBALINFOPAGE) SUPGetGIP(void)
3681	{
3682	return g_pSUPGlobalInfoPage;
3683	}
3684
3685
3686	/**
3687	* Register a component factory with the support driver.
3688	*
3689	* This is currently restricted to kernel sessions only.
3690	*
3691	* @returns VBox status code.
3692	* @retval VINF_SUCCESS on success.
3693	* @retval VERR_NO_MEMORY if we're out of memory.
3694	* @retval VERR_ALREADY_EXISTS if the factory has already been registered.
3695	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3696	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3697	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3698	*
3699	* @param pSession The SUPDRV session (must be a ring-0 session).
3700	* @param pFactory Pointer to the component factory registration structure.
3701	*
3702	* @remarks This interface is also available via SUPR0IdcComponentRegisterFactory.
3703	*/
3704	SUPR0DECL(int) SUPR0ComponentRegisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3705	{
3706	PSUPDRVFACTORYREG pNewReg;
3707	const char *psz;
3708	int rc;
3709
3710	/*
3711	* Validate parameters.
3712	*/
3713	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3714	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3715	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3716	AssertPtrReturn(pFactory->pfnQueryFactoryInterface, VERR_INVALID_POINTER);
3717	psz = RTStrEnd(pFactory->szName, sizeof(pFactory->szName));
3718	AssertReturn(psz, VERR_INVALID_PARAMETER);
3719
3720	/*
3721	* Allocate and initialize a new registration structure.
3722	*/
3723	pNewReg = (PSUPDRVFACTORYREG)RTMemAlloc(sizeof(SUPDRVFACTORYREG));
3724	if (pNewReg)
3725	{
3726	pNewReg->pNext = NULL;
3727	pNewReg->pFactory = pFactory;
3728	pNewReg->pSession = pSession;
3729	pNewReg->cchName = psz - &pFactory->szName[0];
3730
3731	/*
3732	* Add it to the tail of the list after checking for prior registration.
3733	*/
3734	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3735	if (RT_SUCCESS(rc))
3736	{
3737	PSUPDRVFACTORYREG pPrev = NULL;
3738	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3739	while (pCur && pCur->pFactory != pFactory)
3740	{
3741	pPrev = pCur;
3742	pCur = pCur->pNext;
3743	}
3744	if (!pCur)
3745	{
3746	if (pPrev)
3747	pPrev->pNext = pNewReg;
3748	else
3749	pSession->pDevExt->pComponentFactoryHead = pNewReg;
3750	rc = VINF_SUCCESS;
3751	}
3752	else
3753	rc = VERR_ALREADY_EXISTS;
3754
3755	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3756	}
3757
3758	if (RT_FAILURE(rc))
3759	RTMemFree(pNewReg);
3760	}
3761	else
3762	rc = VERR_NO_MEMORY;
3763	return rc;
3764	}
3765
3766
3767	/**
3768	* Deregister a component factory.
3769	*
3770	* @returns VBox status code.
3771	* @retval VINF_SUCCESS on success.
3772	* @retval VERR_NOT_FOUND if the factory wasn't registered.
3773	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3774	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3775	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3776	*
3777	* @param pSession The SUPDRV session (must be a ring-0 session).
3778	* @param pFactory Pointer to the component factory registration structure
3779	* previously passed SUPR0ComponentRegisterFactory().
3780	*
3781	* @remarks This interface is also available via SUPR0IdcComponentDeregisterFactory.
3782	*/
3783	SUPR0DECL(int) SUPR0ComponentDeregisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3784	{
3785	int rc;
3786
3787	/*
3788	* Validate parameters.
3789	*/
3790	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3791	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3792	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3793
3794	/*
3795	* Take the lock and look for the registration record.
3796	*/
3797	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3798	if (RT_SUCCESS(rc))
3799	{
3800	PSUPDRVFACTORYREG pPrev = NULL;
3801	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3802	while (pCur && pCur->pFactory != pFactory)
3803	{
3804	pPrev = pCur;
3805	pCur = pCur->pNext;
3806	}
3807	if (pCur)
3808	{
3809	if (!pPrev)
3810	pSession->pDevExt->pComponentFactoryHead = pCur->pNext;
3811	else
3812	pPrev->pNext = pCur->pNext;
3813
3814	pCur->pNext = NULL;
3815	pCur->pFactory = NULL;
3816	pCur->pSession = NULL;
3817	rc = VINF_SUCCESS;
3818	}
3819	else
3820	rc = VERR_NOT_FOUND;
3821
3822	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3823
3824	RTMemFree(pCur);
3825	}
3826	return rc;
3827	}
3828
3829
3830	/**
3831	* Queries a component factory.
3832	*
3833	* @returns VBox status code.
3834	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3835	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3836	* @retval VERR_SUPDRV_COMPONENT_NOT_FOUND if the component factory wasn't found.
3837	* @retval VERR_SUPDRV_INTERFACE_NOT_SUPPORTED if the interface wasn't supported.
3838	*
3839	* @param pSession The SUPDRV session.
3840	* @param pszName The name of the component factory.
3841	* @param pszInterfaceUuid The UUID of the factory interface (stringified).
3842	* @param ppvFactoryIf Where to store the factory interface.
3843	*/
3844	SUPR0DECL(int) SUPR0ComponentQueryFactory(PSUPDRVSESSION pSession, const char pszName, const char pszInterfaceUuid, void **ppvFactoryIf)
3845	{
3846	const char *pszEnd;
3847	size_t cchName;
3848	int rc;
3849
3850	/*
3851	* Validate parameters.
3852	*/
3853	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3854
3855	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
3856	pszEnd = RTStrEnd(pszName, RT_SIZEOFMEMB(SUPDRVFACTORY, szName));
3857	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3858	cchName = pszEnd - pszName;
3859
3860	AssertPtrReturn(pszInterfaceUuid, VERR_INVALID_POINTER);
3861	pszEnd = RTStrEnd(pszInterfaceUuid, RTUUID_STR_LENGTH);
3862	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3863
3864	AssertPtrReturn(ppvFactoryIf, VERR_INVALID_POINTER);
3865	*ppvFactoryIf = NULL;
3866
3867	/*
3868	* Take the lock and try all factories by this name.
3869	*/
3870	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3871	if (RT_SUCCESS(rc))
3872	{
3873	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3874	rc = VERR_SUPDRV_COMPONENT_NOT_FOUND;
3875	while (pCur)
3876	{
3877	if ( pCur->cchName == cchName
3878	&& !memcmp(pCur->pFactory->szName, pszName, cchName))
3879	{
3880	void *pvFactory = pCur->pFactory->pfnQueryFactoryInterface(pCur->pFactory, pSession, pszInterfaceUuid);
3881	if (pvFactory)
3882	{
3883	*ppvFactoryIf = pvFactory;
3884	rc = VINF_SUCCESS;
3885	break;
3886	}
3887	rc = VERR_SUPDRV_INTERFACE_NOT_SUPPORTED;
3888	}
3889
3890	/* next */
3891	pCur = pCur->pNext;
3892	}
3893
3894	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3895	}
3896	return rc;
3897	}
3898
3899
3900	/**
3901	* Adds a memory object to the session.
3902	*
3903	* @returns IPRT status code.
3904	* @param pMem Memory tracking structure containing the
3905	* information to track.
3906	* @param pSession The session.
3907	*/
3908	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
3909	{
3910	PSUPDRVBUNDLE pBundle;
3911
3912	/*
3913	* Find free entry and record the allocation.
3914	*/
3915	RTSpinlockAcquire(pSession->Spinlock);
3916	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3917	{
3918	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
3919	{
3920	unsigned i;
3921	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3922	{
3923	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
3924	{
3925	pBundle->cUsed++;
3926	pBundle->aMem[i] = *pMem;
3927	RTSpinlockRelease(pSession->Spinlock);
3928	return VINF_SUCCESS;
3929	}
3930	}
3931	AssertFailed(); /* !!this can't be happening!!! */
3932	}
3933	}
3934	RTSpinlockRelease(pSession->Spinlock);
3935
3936	/*
3937	* Need to allocate a new bundle.
3938	* Insert into the last entry in the bundle.
3939	*/
3940	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
3941	if (!pBundle)
3942	return VERR_NO_MEMORY;
3943
3944	/* take last entry. */
3945	pBundle->cUsed++;
3946	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
3947
3948	/* insert into list. */
3949	RTSpinlockAcquire(pSession->Spinlock);
3950	pBundle->pNext = pSession->Bundle.pNext;
3951	pSession->Bundle.pNext = pBundle;
3952	RTSpinlockRelease(pSession->Spinlock);
3953
3954	return VINF_SUCCESS;
3955	}
3956
3957
3958	/**
3959	* Releases a memory object referenced by pointer and type.
3960	*
3961	* @returns IPRT status code.
3962	* @param pSession Session data.
3963	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
3964	* @param eType Memory type.
3965	*/
3966	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
3967	{
3968	PSUPDRVBUNDLE pBundle;
3969
3970	/*
3971	* Validate input.
3972	*/
3973	if (!uPtr)
3974	{
3975	Log(("Illegal address %p\n", (void *)uPtr));
3976	return VERR_INVALID_PARAMETER;
3977	}
3978
3979	/*
3980	* Search for the address.
3981	*/
3982	RTSpinlockAcquire(pSession->Spinlock);
3983	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3984	{
3985	if (pBundle->cUsed > 0)
3986	{
3987	unsigned i;
3988	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3989	{
3990	if ( pBundle->aMem[i].eType == eType
3991	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3992	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
3993	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3994	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
3995	)
3996	{
3997	/* Make a copy of it and release it outside the spinlock. */
3998	SUPDRVMEMREF Mem = pBundle->aMem[i];
3999	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
4000	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
4001	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
4002	RTSpinlockRelease(pSession->Spinlock);
4003
4004	if (Mem.MapObjR3 != NIL_RTR0MEMOBJ)
4005	{
4006	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
4007	AssertRC(rc); /** @todo figure out how to handle this. */
4008	}
4009	if (Mem.MemObj != NIL_RTR0MEMOBJ)
4010	{
4011	int rc = RTR0MemObjFree(Mem.MemObj, true /* fFreeMappings */);
4012	AssertRC(rc); /** @todo figure out how to handle this. */
4013	}
4014	return VINF_SUCCESS;
4015	}
4016	}
4017	}
4018	}
4019	RTSpinlockRelease(pSession->Spinlock);
4020	Log(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
4021	return VERR_INVALID_PARAMETER;
4022	}
4023
4024
4025	/**
4026	* Opens an image. If it's the first time it's opened the call must upload
4027	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
4028	*
4029	* This is the 1st step of the loading.
4030	*
4031	* @returns IPRT status code.
4032	* @param pDevExt Device globals.
4033	* @param pSession Session data.
4034	* @param pReq The open request.
4035	*/
4036	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
4037	{
4038	int rc;
4039	PSUPDRVLDRIMAGE pImage;
4040	void *pv;
4041	size_t cchName = strlen(pReq->u.In.szName); /* (caller checked < 32). */
4042	LogFlow(("supdrvIOCtl_LdrOpen: szName=%s cbImageWithTabs=%d\n", pReq->u.In.szName, pReq->u.In.cbImageWithTabs));
4043
4044	/*
4045	* Check if we got an instance of the image already.
4046	*/
4047	supdrvLdrLock(pDevExt);
4048	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
4049	{
4050	if ( pImage->szName[cchName] == '\0'
4051	&& !memcmp(pImage->szName, pReq->u.In.szName, cchName))
4052	{
4053	/** @todo check cbImageBits and cbImageWithTabs here, if they differs that indicates that the images are different. */
4054	pImage->cUsage++;
4055	pReq->u.Out.pvImageBase = pImage->pvImage;
4056	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
4057	pReq->u.Out.fNativeLoader = pImage->fNative;
4058	supdrvLdrAddUsage(pSession, pImage);
4059	supdrvLdrUnlock(pDevExt);
4060	return VINF_SUCCESS;
4061	}
4062	}
4063	/* (not found - add it!) */
4064
4065	/*
4066	* Allocate memory.
4067	*/
4068	Assert(cchName < sizeof(pImage->szName));
4069	pv = RTMemAlloc(sizeof(SUPDRVLDRIMAGE));
4070	if (!pv)
4071	{
4072	supdrvLdrUnlock(pDevExt);
4073	Log(("supdrvIOCtl_LdrOpen: RTMemAlloc() failed\n"));
4074	return /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_2;
4075	}
4076
4077	/*
4078	* Setup and link in the LDR stuff.
4079	*/
4080	pImage = (PSUPDRVLDRIMAGE)pv;
4081	pImage->pvImage = NULL;
4082	pImage->pvImageAlloc = NULL;
4083	pImage->cbImageWithTabs = pReq->u.In.cbImageWithTabs;
4084	pImage->cbImageBits = pReq->u.In.cbImageBits;
4085	pImage->cSymbols = 0;
4086	pImage->paSymbols = NULL;
4087	pImage->pachStrTab = NULL;
4088	pImage->cbStrTab = 0;
4089	pImage->pfnModuleInit = NULL;
4090	pImage->pfnModuleTerm = NULL;
4091	pImage->pfnServiceReqHandler = NULL;
4092	pImage->uState = SUP_IOCTL_LDR_OPEN;
4093	pImage->cUsage = 1;
4094	pImage->pDevExt = pDevExt;
4095	memcpy(pImage->szName, pReq->u.In.szName, cchName + 1);
4096
4097	/*
4098	* Try load it using the native loader, if that isn't supported, fall back
4099	* on the older method.
4100	*/
4101	pImage->fNative = true;
4102	rc = supdrvOSLdrOpen(pDevExt, pImage, pReq->u.In.szFilename);
4103	if (rc == VERR_NOT_SUPPORTED)
4104	{
4105	pImage->pvImageAlloc = RTMemExecAlloc(pImage->cbImageBits + 31);
4106	pImage->pvImage = RT_ALIGN_P(pImage->pvImageAlloc, 32);
4107	pImage->fNative = false;
4108	rc = pImage->pvImageAlloc ? VINF_SUCCESS : VERR_NO_EXEC_MEMORY;
4109	}
4110	if (RT_FAILURE(rc))
4111	{
4112	supdrvLdrUnlock(pDevExt);
4113	RTMemFree(pImage);
4114	Log(("supdrvIOCtl_LdrOpen(%s): failed - %Rrc\n", pReq->u.In.szName, rc));
4115	return rc;
4116	}
4117	Assert(VALID_PTR(pImage->pvImage) \|\| RT_FAILURE(rc));
4118
4119	/*
4120	* Link it.
4121	*/
4122	pImage->pNext = pDevExt->pLdrImages;
4123	pDevExt->pLdrImages = pImage;
4124
4125	supdrvLdrAddUsage(pSession, pImage);
4126
4127	pReq->u.Out.pvImageBase = pImage->pvImage;
4128	pReq->u.Out.fNeedsLoading = true;
4129	pReq->u.Out.fNativeLoader = pImage->fNative;
4130	supdrvOSLdrNotifyOpened(pDevExt, pImage);
4131
4132	supdrvLdrUnlock(pDevExt);
4133	return VINF_SUCCESS;
4134	}
4135
4136
4137	/**
4138	* Worker that validates a pointer to an image entrypoint.
4139	*
4140	* @returns IPRT status code.
4141	* @param pDevExt The device globals.
4142	* @param pImage The loader image.
4143	* @param pv The pointer into the image.
4144	* @param fMayBeNull Whether it may be NULL.
4145	* @param pszWhat What is this entrypoint? (for logging)
4146	* @param pbImageBits The image bits prepared by ring-3.
4147	*
4148	* @remarks Will leave the lock on failure.
4149	*/
4150	static int supdrvLdrValidatePointer(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage, void *pv,
4151	bool fMayBeNull, const uint8_t pbImageBits, const char pszWhat)
4152	{
4153	if (!fMayBeNull \|\| pv)
4154	{
4155	if ((uintptr_t)pv - (uintptr_t)pImage->pvImage >= pImage->cbImageBits)
4156	{
4157	supdrvLdrUnlock(pDevExt);
4158	Log(("Out of range (%p LB %#x): %s=%p\n", pImage->pvImage, pImage->cbImageBits, pszWhat, pv));
4159	return VERR_INVALID_PARAMETER;
4160	}
4161
4162	if (pImage->fNative)
4163	{
4164	int rc = supdrvOSLdrValidatePointer(pDevExt, pImage, pv, pbImageBits);
4165	if (RT_FAILURE(rc))
4166	{
4167	supdrvLdrUnlock(pDevExt);
4168	Log(("Bad entry point address: %s=%p (rc=%Rrc)\n", pszWhat, pv, rc));
4169	return rc;
4170	}
4171	}
4172	}
4173	return VINF_SUCCESS;
4174	}
4175
4176
4177	/**
4178	* Loads the image bits.
4179	*
4180	* This is the 2nd step of the loading.
4181	*
4182	* @returns IPRT status code.
4183	* @param pDevExt Device globals.
4184	* @param pSession Session data.
4185	* @param pReq The request.
4186	*/
4187	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
4188	{
4189	PSUPDRVLDRUSAGE pUsage;
4190	PSUPDRVLDRIMAGE pImage;
4191	int rc;
4192	LogFlow(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImageWithBits=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImageWithTabs));
4193
4194	/*
4195	* Find the ldr image.
4196	*/
4197	supdrvLdrLock(pDevExt);
4198	pUsage = pSession->pLdrUsage;
4199	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4200	pUsage = pUsage->pNext;
4201	if (!pUsage)
4202	{
4203	supdrvLdrUnlock(pDevExt);
4204	Log(("SUP_IOCTL_LDR_LOAD: couldn't find image!\n"));
4205	return VERR_INVALID_HANDLE;
4206	}
4207	pImage = pUsage->pImage;
4208
4209	/*
4210	* Validate input.
4211	*/
4212	if ( pImage->cbImageWithTabs != pReq->u.In.cbImageWithTabs
4213	\|\| pImage->cbImageBits != pReq->u.In.cbImageBits)
4214	{
4215	supdrvLdrUnlock(pDevExt);
4216	Log(("SUP_IOCTL_LDR_LOAD: image size mismatch!! %d(prep) != %d(load) or %d != %d\n",
4217	pImage->cbImageWithTabs, pReq->u.In.cbImageWithTabs, pImage->cbImageBits, pReq->u.In.cbImageBits));
4218	return VERR_INVALID_HANDLE;
4219	}
4220
4221	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
4222	{
4223	unsigned uState = pImage->uState;
4224	supdrvLdrUnlock(pDevExt);
4225	if (uState != SUP_IOCTL_LDR_LOAD)
4226	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
4227	return VERR_ALREADY_LOADED;
4228	}
4229
4230	switch (pReq->u.In.eEPType)
4231	{
4232	case SUPLDRLOADEP_NOTHING:
4233	break;
4234
4235	case SUPLDRLOADEP_VMMR0:
4236	rc = supdrvLdrValidatePointer( pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0, false, pReq->u.In.abImage, "pvVMMR0");
4237	if (RT_SUCCESS(rc))
4238	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt, false, pReq->u.In.abImage, "pvVMMR0EntryInt");
4239	if (RT_SUCCESS(rc))
4240	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, false, pReq->u.In.abImage, "pvVMMR0EntryFast");
4241	if (RT_SUCCESS(rc))
4242	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx, false, pReq->u.In.abImage, "pvVMMR0EntryEx");
4243	if (RT_FAILURE(rc))
4244	return rc;
4245	break;
4246
4247	case SUPLDRLOADEP_SERVICE:
4248	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.Service.pfnServiceReq, false, pReq->u.In.abImage, "pfnServiceReq");
4249	if (RT_FAILURE(rc))
4250	return rc;
4251	if ( pReq->u.In.EP.Service.apvReserved[0] != NIL_RTR0PTR
4252	\|\| pReq->u.In.EP.Service.apvReserved[1] != NIL_RTR0PTR
4253	\|\| pReq->u.In.EP.Service.apvReserved[2] != NIL_RTR0PTR)
4254	{
4255	supdrvLdrUnlock(pDevExt);
4256	Log(("Out of range (%p LB %#x): apvReserved={%p,%p,%p} MBZ!\n",
4257	pImage->pvImage, pReq->u.In.cbImageWithTabs,
4258	pReq->u.In.EP.Service.apvReserved[0],
4259	pReq->u.In.EP.Service.apvReserved[1],
4260	pReq->u.In.EP.Service.apvReserved[2]));
4261	return VERR_INVALID_PARAMETER;
4262	}
4263	break;
4264
4265	default:
4266	supdrvLdrUnlock(pDevExt);
4267	Log(("Invalid eEPType=%d\n", pReq->u.In.eEPType));
4268	return VERR_INVALID_PARAMETER;
4269	}
4270
4271	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleInit, true, pReq->u.In.abImage, "pfnModuleInit");
4272	if (RT_FAILURE(rc))
4273	return rc;
4274	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleTerm, true, pReq->u.In.abImage, "pfnModuleTerm");
4275	if (RT_FAILURE(rc))
4276	return rc;
4277
4278	/*
4279	* Allocate and copy the tables.
4280	* (No need to do try/except as this is a buffered request.)
4281	*/
4282	pImage->cbStrTab = pReq->u.In.cbStrTab;
4283	if (pImage->cbStrTab)
4284	{
4285	pImage->pachStrTab = (char *)RTMemAlloc(pImage->cbStrTab);
4286	if (pImage->pachStrTab)
4287	memcpy(pImage->pachStrTab, &pReq->u.In.abImage[pReq->u.In.offStrTab], pImage->cbStrTab);
4288	else
4289	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_3;
4290	}
4291
4292	pImage->cSymbols = pReq->u.In.cSymbols;
4293	if (RT_SUCCESS(rc) && pImage->cSymbols)
4294	{
4295	size_t cbSymbols = pImage->cSymbols * sizeof(SUPLDRSYM);
4296	pImage->paSymbols = (PSUPLDRSYM)RTMemAlloc(cbSymbols);
4297	if (pImage->paSymbols)
4298	memcpy(pImage->paSymbols, &pReq->u.In.abImage[pReq->u.In.offSymbols], cbSymbols);
4299	else
4300	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_4;
4301	}
4302
4303	/*
4304	* Copy the bits / complete native loading.
4305	*/
4306	if (RT_SUCCESS(rc))
4307	{
4308	pImage->uState = SUP_IOCTL_LDR_LOAD;
4309	pImage->pfnModuleInit = pReq->u.In.pfnModuleInit;
4310	pImage->pfnModuleTerm = pReq->u.In.pfnModuleTerm;
4311
4312	if (pImage->fNative)
4313	rc = supdrvOSLdrLoad(pDevExt, pImage, pReq->u.In.abImage, pReq);
4314	else
4315	{
4316	memcpy(pImage->pvImage, &pReq->u.In.abImage[0], pImage->cbImageBits);
4317	Log(("vboxdrv: Loaded '%s' at %p\n", pImage->szName, pImage->pvImage));
4318	}
4319	}
4320
4321	/*
4322	* Update any entry points.
4323	*/
4324	if (RT_SUCCESS(rc))
4325	{
4326	switch (pReq->u.In.eEPType)
4327	{
4328	default:
4329	case SUPLDRLOADEP_NOTHING:
4330	rc = VINF_SUCCESS;
4331	break;
4332	case SUPLDRLOADEP_VMMR0:
4333	rc = supdrvLdrSetVMMR0EPs(pDevExt, pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
4334	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
4335	break;
4336	case SUPLDRLOADEP_SERVICE:
4337	pImage->pfnServiceReqHandler = pReq->u.In.EP.Service.pfnServiceReq;
4338	rc = VINF_SUCCESS;
4339	break;
4340	}
4341	}
4342
4343	/*
4344	* On success call the module initialization.
4345	*/
4346	LogFlow(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
4347	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
4348	{
4349	Log(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
4350	pDevExt->pLdrInitImage = pImage;
4351	pDevExt->hLdrInitThread = RTThreadNativeSelf();
4352	rc = pImage->pfnModuleInit(pImage);
4353	pDevExt->pLdrInitImage = NULL;
4354	pDevExt->hLdrInitThread = NIL_RTNATIVETHREAD;
4355	if (RT_FAILURE(rc) && pDevExt->pvVMMR0 == pImage->pvImage)
4356	supdrvLdrUnsetVMMR0EPs(pDevExt);
4357	}
4358
4359	if (RT_FAILURE(rc))
4360	{
4361	/* Inform the tracing component in case ModuleInit registered TPs. */
4362	supdrvTracerModuleUnloading(pDevExt, pImage);
4363
4364	pImage->uState = SUP_IOCTL_LDR_OPEN;
4365	pImage->pfnModuleInit = NULL;
4366	pImage->pfnModuleTerm = NULL;
4367	pImage->pfnServiceReqHandler= NULL;
4368	pImage->cbStrTab = 0;
4369	RTMemFree(pImage->pachStrTab);
4370	pImage->pachStrTab = NULL;
4371	RTMemFree(pImage->paSymbols);
4372	pImage->paSymbols = NULL;
4373	pImage->cSymbols = 0;
4374	}
4375
4376	supdrvLdrUnlock(pDevExt);
4377	return rc;
4378	}
4379
4380
4381	/**
4382	* Frees a previously loaded (prep'ed) image.
4383	*
4384	* @returns IPRT status code.
4385	* @param pDevExt Device globals.
4386	* @param pSession Session data.
4387	* @param pReq The request.
4388	*/
4389	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
4390	{
4391	int rc;
4392	PSUPDRVLDRUSAGE pUsagePrev;
4393	PSUPDRVLDRUSAGE pUsage;
4394	PSUPDRVLDRIMAGE pImage;
4395	LogFlow(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
4396
4397	/*
4398	* Find the ldr image.
4399	*/
4400	supdrvLdrLock(pDevExt);
4401	pUsagePrev = NULL;
4402	pUsage = pSession->pLdrUsage;
4403	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4404	{
4405	pUsagePrev = pUsage;
4406	pUsage = pUsage->pNext;
4407	}
4408	if (!pUsage)
4409	{
4410	supdrvLdrUnlock(pDevExt);
4411	Log(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
4412	return VERR_INVALID_HANDLE;
4413	}
4414
4415	/*
4416	* Check if we can remove anything.
4417	*/
4418	rc = VINF_SUCCESS;
4419	pImage = pUsage->pImage;
4420	if (pImage->cUsage <= 1 \|\| pUsage->cUsage <= 1)
4421	{
4422	/*
4423	* Check if there are any objects with destructors in the image, if
4424	* so leave it for the session cleanup routine so we get a chance to
4425	* clean things up in the right order and not leave them all dangling.
4426	*/
4427	RTSpinlockAcquire(pDevExt->Spinlock);
4428	if (pImage->cUsage <= 1)
4429	{
4430	PSUPDRVOBJ pObj;
4431	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4432	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4433	{
4434	rc = VERR_DANGLING_OBJECTS;
4435	break;
4436	}
4437	}
4438	else
4439	{
4440	PSUPDRVUSAGE pGenUsage;
4441	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
4442	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4443	{
4444	rc = VERR_DANGLING_OBJECTS;
4445	break;
4446	}
4447	}
4448	RTSpinlockRelease(pDevExt->Spinlock);
4449	if (rc == VINF_SUCCESS)
4450	{
4451	/* unlink it */
4452	if (pUsagePrev)
4453	pUsagePrev->pNext = pUsage->pNext;
4454	else
4455	pSession->pLdrUsage = pUsage->pNext;
4456
4457	/* free it */
4458	pUsage->pImage = NULL;
4459	pUsage->pNext = NULL;
4460	RTMemFree(pUsage);
4461
4462	/*
4463	* Dereference the image.
4464	*/
4465	if (pImage->cUsage <= 1)
4466	supdrvLdrFree(pDevExt, pImage);
4467	else
4468	pImage->cUsage--;
4469	}
4470	else
4471	{
4472	Log(("supdrvIOCtl_LdrFree: Dangling objects in %p/%s!\n", pImage->pvImage, pImage->szName));
4473	rc = VINF_SUCCESS; /** @todo BRANCH-2.1: remove this after branching. */
4474	}
4475	}
4476	else
4477	{
4478	/*
4479	* Dereference both image and usage.
4480	*/
4481	pImage->cUsage--;
4482	pUsage->cUsage--;
4483	}
4484
4485	supdrvLdrUnlock(pDevExt);
4486	return rc;
4487	}
4488
4489
4490	/**
4491	* Gets the address of a symbol in an open image.
4492	*
4493	* @returns IPRT status code.
4494	* @param pDevExt Device globals.
4495	* @param pSession Session data.
4496	* @param pReq The request buffer.
4497	*/
4498	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
4499	{
4500	PSUPDRVLDRIMAGE pImage;
4501	PSUPDRVLDRUSAGE pUsage;
4502	uint32_t i;
4503	PSUPLDRSYM paSyms;
4504	const char *pchStrings;
4505	const size_t cbSymbol = strlen(pReq->u.In.szSymbol) + 1;
4506	void *pvSymbol = NULL;
4507	int rc = VERR_GENERAL_FAILURE;
4508	Log3(("supdrvIOCtl_LdrGetSymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
4509
4510	/*
4511	* Find the ldr image.
4512	*/
4513	supdrvLdrLock(pDevExt);
4514	pUsage = pSession->pLdrUsage;
4515	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4516	pUsage = pUsage->pNext;
4517	if (!pUsage)
4518	{
4519	supdrvLdrUnlock(pDevExt);
4520	Log(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
4521	return VERR_INVALID_HANDLE;
4522	}
4523	pImage = pUsage->pImage;
4524	if (pImage->uState != SUP_IOCTL_LDR_LOAD)
4525	{
4526	unsigned uState = pImage->uState;
4527	supdrvLdrUnlock(pDevExt);
4528	Log(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", uState, uState)); NOREF(uState);
4529	return VERR_ALREADY_LOADED;
4530	}
4531
4532	/*
4533	* Search the symbol strings.
4534	*
4535	* Note! The int32_t is for native loading on solaris where the data
4536	* and text segments are in very different places.
4537	*/
4538	pchStrings = pImage->pachStrTab;
4539	paSyms = pImage->paSymbols;
4540	for (i = 0; i < pImage->cSymbols; i++)
4541	{
4542	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4543	&& !memcmp(pchStrings + paSyms[i].offName, pReq->u.In.szSymbol, cbSymbol))
4544	{
4545	pvSymbol = (uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol;
4546	rc = VINF_SUCCESS;
4547	break;
4548	}
4549	}
4550	supdrvLdrUnlock(pDevExt);
4551	pReq->u.Out.pvSymbol = pvSymbol;
4552	return rc;
4553	}
4554
4555
4556	/**
4557	* Gets the address of a symbol in an open image or the support driver.
4558	*
4559	* @returns VINF_SUCCESS on success.
4560	* @returns
4561	* @param pDevExt Device globals.
4562	* @param pSession Session data.
4563	* @param pReq The request buffer.
4564	*/
4565	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq)
4566	{
4567	int rc = VINF_SUCCESS;
4568	const char *pszSymbol = pReq->u.In.pszSymbol;
4569	const char *pszModule = pReq->u.In.pszModule;
4570	size_t cbSymbol;
4571	char const *pszEnd;
4572	uint32_t i;
4573
4574	/*
4575	* Input validation.
4576	*/
4577	AssertPtrReturn(pszSymbol, VERR_INVALID_POINTER);
4578	pszEnd = RTStrEnd(pszSymbol, 512);
4579	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4580	cbSymbol = pszEnd - pszSymbol + 1;
4581
4582	if (pszModule)
4583	{
4584	AssertPtrReturn(pszModule, VERR_INVALID_POINTER);
4585	pszEnd = RTStrEnd(pszModule, 64);
4586	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4587	}
4588	Log3(("supdrvIDC_LdrGetSymbol: pszModule=%p:{%s} pszSymbol=%p:{%s}\n", pszModule, pszModule, pszSymbol, pszSymbol));
4589
4590
4591	if ( !pszModule
4592	\|\| !strcmp(pszModule, "SupDrv"))
4593	{
4594	/*
4595	* Search the support driver export table.
4596	*/
4597	for (i = 0; i < RT_ELEMENTS(g_aFunctions); i++)
4598	if (!strcmp(g_aFunctions[i].szName, pszSymbol))
4599	{
4600	pReq->u.Out.pfnSymbol = g_aFunctions[i].pfn;
4601	break;
4602	}
4603	}
4604	else
4605	{
4606	/*
4607	* Find the loader image.
4608	*/
4609	PSUPDRVLDRIMAGE pImage;
4610
4611	supdrvLdrLock(pDevExt);
4612
4613	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
4614	if (!strcmp(pImage->szName, pszModule))
4615	break;
4616	if (pImage && pImage->uState == SUP_IOCTL_LDR_LOAD)
4617	{
4618	/*
4619	* Search the symbol strings.
4620	*/
4621	const char *pchStrings = pImage->pachStrTab;
4622	PCSUPLDRSYM paSyms = pImage->paSymbols;
4623	for (i = 0; i < pImage->cSymbols; i++)
4624	{
4625	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4626	&& !memcmp(pchStrings + paSyms[i].offName, pszSymbol, cbSymbol))
4627	{
4628	/*
4629	* Found it! Calc the symbol address and add a reference to the module.
4630	*/
4631	pReq->u.Out.pfnSymbol = (PFNRT)((uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol);
4632	rc = supdrvLdrAddUsage(pSession, pImage);
4633	break;
4634	}
4635	}
4636	}
4637	else
4638	rc = pImage ? VERR_WRONG_ORDER : VERR_MODULE_NOT_FOUND;
4639
4640	supdrvLdrUnlock(pDevExt);
4641	}
4642	return rc;
4643	}
4644
4645
4646	/**
4647	* Updates the VMMR0 entry point pointers.
4648	*
4649	* @returns IPRT status code.
4650	* @param pDevExt Device globals.
4651	* @param pSession Session data.
4652	* @param pVMMR0 VMMR0 image handle.
4653	* @param pvVMMR0EntryInt VMMR0EntryInt address.
4654	* @param pvVMMR0EntryFast VMMR0EntryFast address.
4655	* @param pvVMMR0EntryEx VMMR0EntryEx address.
4656	* @remark Caller must own the loader mutex.
4657	*/
4658	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx)
4659	{
4660	int rc = VINF_SUCCESS;
4661	LogFlow(("supdrvLdrSetR0EP pvVMMR0=%p pvVMMR0EntryInt=%p\n", pvVMMR0, pvVMMR0EntryInt));
4662
4663
4664	/*
4665	* Check if not yet set.
4666	*/
4667	if (!pDevExt->pvVMMR0)
4668	{
4669	pDevExt->pvVMMR0 = pvVMMR0;
4670	pDevExt->pfnVMMR0EntryInt = pvVMMR0EntryInt;
4671	pDevExt->pfnVMMR0EntryFast = pvVMMR0EntryFast;
4672	pDevExt->pfnVMMR0EntryEx = pvVMMR0EntryEx;
4673	}
4674	else
4675	{
4676	/*
4677	* Return failure or success depending on whether the values match or not.
4678	*/
4679	if ( pDevExt->pvVMMR0 != pvVMMR0
4680	\|\| (void *)pDevExt->pfnVMMR0EntryInt != pvVMMR0EntryInt
4681	\|\| (void *)pDevExt->pfnVMMR0EntryFast != pvVMMR0EntryFast
4682	\|\| (void *)pDevExt->pfnVMMR0EntryEx != pvVMMR0EntryEx)
4683	{
4684	AssertMsgFailed(("SUP_IOCTL_LDR_SETR0EP: Already set pointing to a different module!\n"));
4685	rc = VERR_INVALID_PARAMETER;
4686	}
4687	}
4688	return rc;
4689	}
4690
4691
4692	/**
4693	* Unsets the VMMR0 entry point installed by supdrvLdrSetR0EP.
4694	*
4695	* @param pDevExt Device globals.
4696	*/
4697	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt)
4698	{
4699	pDevExt->pvVMMR0 = NULL;
4700	pDevExt->pfnVMMR0EntryInt = NULL;
4701	pDevExt->pfnVMMR0EntryFast = NULL;
4702	pDevExt->pfnVMMR0EntryEx = NULL;
4703	}
4704
4705
4706	/**
4707	* Adds a usage reference in the specified session of an image.
4708	*
4709	* Called while owning the loader semaphore.
4710	*
4711	* @returns VINF_SUCCESS on success and VERR_NO_MEMORY on failure.
4712	* @param pSession Session in question.
4713	* @param pImage Image which the session is using.
4714	*/
4715	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage)
4716	{
4717	PSUPDRVLDRUSAGE pUsage;
4718	LogFlow(("supdrvLdrAddUsage: pImage=%p\n", pImage));
4719
4720	/*
4721	* Referenced it already?
4722	*/
4723	pUsage = pSession->pLdrUsage;
4724	while (pUsage)
4725	{
4726	if (pUsage->pImage == pImage)
4727	{
4728	pUsage->cUsage++;
4729	return VINF_SUCCESS;
4730	}
4731	pUsage = pUsage->pNext;
4732	}
4733
4734	/*
4735	* Allocate new usage record.
4736	*/
4737	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
4738	AssertReturn(pUsage, /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_5);
4739	pUsage->cUsage = 1;
4740	pUsage->pImage = pImage;
4741	pUsage->pNext = pSession->pLdrUsage;
4742	pSession->pLdrUsage = pUsage;
4743	return VINF_SUCCESS;
4744	}
4745
4746
4747	/**
4748	* Frees a load image.
4749	*
4750	* @param pDevExt Pointer to device extension.
4751	* @param pImage Pointer to the image we're gonna free.
4752	* This image must exit!
4753	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
4754	*/
4755	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
4756	{
4757	PSUPDRVLDRIMAGE pImagePrev;
4758	LogFlow(("supdrvLdrFree: pImage=%p\n", pImage));
4759
4760	/* find it - arg. should've used doubly linked list. */
4761	Assert(pDevExt->pLdrImages);
4762	pImagePrev = NULL;
4763	if (pDevExt->pLdrImages != pImage)
4764	{
4765	pImagePrev = pDevExt->pLdrImages;
4766	while (pImagePrev->pNext != pImage)
4767	pImagePrev = pImagePrev->pNext;
4768	Assert(pImagePrev->pNext == pImage);
4769	}
4770
4771	/* unlink */
4772	if (pImagePrev)
4773	pImagePrev->pNext = pImage->pNext;
4774	else
4775	pDevExt->pLdrImages = pImage->pNext;
4776
4777	/* check if this is VMMR0.r0 unset its entry point pointers. */
4778	if (pDevExt->pvVMMR0 == pImage->pvImage)
4779	supdrvLdrUnsetVMMR0EPs(pDevExt);
4780
4781	/* check for objects with destructors in this image. (Shouldn't happen.) */
4782	if (pDevExt->pObjs)
4783	{
4784	unsigned cObjs = 0;
4785	PSUPDRVOBJ pObj;
4786	RTSpinlockAcquire(pDevExt->Spinlock);
4787	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4788	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4789	{
4790	pObj->pfnDestructor = NULL;
4791	cObjs++;
4792	}
4793	RTSpinlockRelease(pDevExt->Spinlock);
4794	if (cObjs)
4795	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
4796	}
4797
4798	/* call termination function if fully loaded. */
4799	if ( pImage->pfnModuleTerm
4800	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
4801	{
4802	LogFlow(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
4803	pImage->pfnModuleTerm(pImage);
4804	}
4805
4806	/* Inform the tracing component. */
4807	supdrvTracerModuleUnloading(pDevExt, pImage);
4808
4809	/* do native unload if appropriate. */
4810	if (pImage->fNative)
4811	supdrvOSLdrUnload(pDevExt, pImage);
4812
4813	/* free the image */
4814	pImage->cUsage = 0;
4815	pImage->pDevExt = NULL;
4816	pImage->pNext = NULL;
4817	pImage->uState = SUP_IOCTL_LDR_FREE;
4818	RTMemExecFree(pImage->pvImageAlloc, pImage->cbImageBits + 31);
4819	pImage->pvImageAlloc = NULL;
4820	RTMemFree(pImage->pachStrTab);
4821	pImage->pachStrTab = NULL;
4822	RTMemFree(pImage->paSymbols);
4823	pImage->paSymbols = NULL;
4824	RTMemFree(pImage);
4825	}
4826
4827
4828	/**
4829	* Acquires the loader lock.
4830	*
4831	* @returns IPRT status code.
4832	* @param pDevExt The device extension.
4833	*/
4834	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt)
4835	{
4836	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4837	int rc = RTSemMutexRequest(pDevExt->mtxLdr, RT_INDEFINITE_WAIT);
4838	#else
4839	int rc = RTSemFastMutexRequest(pDevExt->mtxLdr);
4840	#endif
4841	AssertRC(rc);
4842	return rc;
4843	}
4844
4845
4846	/**
4847	* Releases the loader lock.
4848	*
4849	* @returns IPRT status code.
4850	* @param pDevExt The device extension.
4851	*/
4852	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt)
4853	{
4854	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4855	return RTSemMutexRelease(pDevExt->mtxLdr);
4856	#else
4857	return RTSemFastMutexRelease(pDevExt->mtxLdr);
4858	#endif
4859	}
4860
4861
4862	/**
4863	* Implements the service call request.
4864	*
4865	* @returns VBox status code.
4866	* @param pDevExt The device extension.
4867	* @param pSession The calling session.
4868	* @param pReq The request packet, valid.
4869	*/
4870	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq)
4871	{
4872	#if !defined(RT_OS_WINDOWS) \|\| defined(RT_ARCH_AMD64) \|\| defined(DEBUG)
4873	int rc;
4874
4875	/*
4876	* Find the module first in the module referenced by the calling session.
4877	*/
4878	rc = supdrvLdrLock(pDevExt);
4879	if (RT_SUCCESS(rc))
4880	{
4881	PFNSUPR0SERVICEREQHANDLER pfnServiceReqHandler = NULL;
4882	PSUPDRVLDRUSAGE pUsage;
4883
4884	for (pUsage = pSession->pLdrUsage; pUsage; pUsage = pUsage->pNext)
4885	if ( pUsage->pImage->pfnServiceReqHandler
4886	&& !strcmp(pUsage->pImage->szName, pReq->u.In.szName))
4887	{
4888	pfnServiceReqHandler = pUsage->pImage->pfnServiceReqHandler;
4889	break;
4890	}
4891	supdrvLdrUnlock(pDevExt);
4892
4893	if (pfnServiceReqHandler)
4894	{
4895	/*
4896	* Call it.
4897	*/
4898	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
4899	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, NULL);
4900	else
4901	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0]);
4902	}
4903	else
4904	rc = VERR_SUPDRV_SERVICE_NOT_FOUND;
4905	}
4906
4907	/* log it */
4908	if ( RT_FAILURE(rc)
4909	&& rc != VERR_INTERRUPTED
4910	&& rc != VERR_TIMEOUT)
4911	Log(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4912	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4913	else
4914	Log4(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4915	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4916	return rc;
4917	#else /* RT_OS_WINDOWS && !RT_ARCH_AMD64 && !DEBUG */
4918	return VERR_NOT_IMPLEMENTED;
4919	#endif /* RT_OS_WINDOWS && !RT_ARCH_AMD64 && !DEBUG */
4920	}
4921
4922
4923	/**
4924	* Implements the logger settings request.
4925	*
4926	* @returns VBox status code.
4927	* @param pDevExt The device extension.
4928	* @param pSession The caller's session.
4929	* @param pReq The request.
4930	*/
4931	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq)
4932	{
4933	const char *pszGroup = &pReq->u.In.szStrings[pReq->u.In.offGroups];
4934	const char *pszFlags = &pReq->u.In.szStrings[pReq->u.In.offFlags];
4935	const char *pszDest = &pReq->u.In.szStrings[pReq->u.In.offDestination];
4936	PRTLOGGER pLogger = NULL;
4937	int rc;
4938
4939	/*
4940	* Some further validation.
4941	*/
4942	switch (pReq->u.In.fWhat)
4943	{
4944	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4945	case SUPLOGGERSETTINGS_WHAT_CREATE:
4946	break;
4947
4948	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4949	if (pszGroup \|\| pszFlags \|\| *pszDest)
4950	return VERR_INVALID_PARAMETER;
4951	if (pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_RELEASE)
4952	return VERR_ACCESS_DENIED;
4953	break;
4954
4955	default:
4956	return VERR_INTERNAL_ERROR;
4957	}
4958
4959	/*
4960	* Get the logger.
4961	*/
4962	switch (pReq->u.In.fWhich)
4963	{
4964	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4965	pLogger = RTLogGetDefaultInstance();
4966	break;
4967
4968	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4969	pLogger = RTLogRelDefaultInstance();
4970	break;
4971
4972	default:
4973	return VERR_INTERNAL_ERROR;
4974	}
4975
4976	/*
4977	* Do the job.
4978	*/
4979	switch (pReq->u.In.fWhat)
4980	{
4981	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4982	if (pLogger)
4983	{
4984	rc = RTLogFlags(pLogger, pszFlags);
4985	if (RT_SUCCESS(rc))
4986	rc = RTLogGroupSettings(pLogger, pszGroup);
4987	NOREF(pszDest);
4988	}
4989	else
4990	rc = VERR_NOT_FOUND;
4991	break;
4992
4993	case SUPLOGGERSETTINGS_WHAT_CREATE:
4994	{
4995	if (pLogger)
4996	rc = VERR_ALREADY_EXISTS;
4997	else
4998	{
4999	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
5000
5001	rc = RTLogCreate(&pLogger,
5002	0 /* fFlags */,
5003	pszGroup,
5004	pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_DEBUG
5005	? "VBOX_LOG"
5006	: "VBOX_RELEASE_LOG",
5007	RT_ELEMENTS(s_apszGroups),
5008	s_apszGroups,
5009	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER,
5010	NULL);
5011	if (RT_SUCCESS(rc))
5012	{
5013	rc = RTLogFlags(pLogger, pszFlags);
5014	NOREF(pszDest);
5015	if (RT_SUCCESS(rc))
5016	{
5017	switch (pReq->u.In.fWhich)
5018	{
5019	case SUPLOGGERSETTINGS_WHICH_DEBUG:
5020	pLogger = RTLogSetDefaultInstance(pLogger);
5021	break;
5022	case SUPLOGGERSETTINGS_WHICH_RELEASE:
5023	pLogger = RTLogRelSetDefaultInstance(pLogger);
5024	break;
5025	}
5026	}
5027	RTLogDestroy(pLogger);
5028	}
5029	}
5030	break;
5031	}
5032
5033	case SUPLOGGERSETTINGS_WHAT_DESTROY:
5034	switch (pReq->u.In.fWhich)
5035	{
5036	case SUPLOGGERSETTINGS_WHICH_DEBUG:
5037	pLogger = RTLogSetDefaultInstance(NULL);
5038	break;
5039	case SUPLOGGERSETTINGS_WHICH_RELEASE:
5040	pLogger = RTLogRelSetDefaultInstance(NULL);
5041	break;
5042	}
5043	rc = RTLogDestroy(pLogger);
5044	break;
5045
5046	default:
5047	{
5048	rc = VERR_INTERNAL_ERROR;
5049	break;
5050	}
5051	}
5052
5053	return rc;
5054	}
5055
5056
5057	/**
5058	* Creates the GIP.
5059	*
5060	* @returns VBox status code.
5061	* @param pDevExt Instance data. GIP stuff may be updated.
5062	*/
5063	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
5064	{
5065	PSUPGLOBALINFOPAGE pGip;
5066	RTHCPHYS HCPhysGip;
5067	uint32_t u32SystemResolution;
5068	uint32_t u32Interval;
5069	unsigned cCpus;
5070	int rc;
5071
5072
5073	LogFlow(("supdrvGipCreate:\n"));
5074
5075	/* assert order */
5076	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
5077	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
5078	Assert(!pDevExt->pGipTimer);
5079
5080	/*
5081	* Check the CPU count.
5082	*/
5083	cCpus = RTMpGetArraySize();
5084	if ( cCpus > RTCPUSET_MAX_CPUS
5085	\|\| cCpus > 256 /ApicId is used for the mappings/)
5086	{
5087	SUPR0Printf("VBoxDrv: Too many CPUs (%u) for the GIP (max %u)\n", cCpus, RT_MIN(RTCPUSET_MAX_CPUS, 256));
5088	return VERR_TOO_MANY_CPUS;
5089	}
5090
5091	/*
5092	* Allocate a contiguous set of pages with a default kernel mapping.
5093	*/
5094	rc = RTR0MemObjAllocCont(&pDevExt->GipMemObj, RT_UOFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), false /fExecutable/);
5095	if (RT_FAILURE(rc))
5096	{
5097	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
5098	return rc;
5099	}
5100	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
5101	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
5102
5103	/*
5104	* Find a reasonable update interval and initialize the structure.
5105	*/
5106	u32Interval = u32SystemResolution = RTTimerGetSystemGranularity();
5107	while (u32Interval < 10000000 /* 10 ms */)
5108	u32Interval += u32SystemResolution;
5109
5110	supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), 1000000000 / u32Interval /=Hz/, cCpus);
5111
5112	/*
5113	* Create the timer.
5114	* If CPU_ALL isn't supported we'll have to fall back to synchronous mode.
5115	*/
5116	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
5117	{
5118	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, RTTIMER_FLAGS_CPU_ALL, supdrvGipAsyncTimer, pDevExt);
5119	if (rc == VERR_NOT_SUPPORTED)
5120	{
5121	OSDBGPRINT(("supdrvGipCreate: omni timer not supported, falling back to synchronous mode\n"));
5122	pGip->u32Mode = SUPGIPMODE_SYNC_TSC;
5123	}
5124	}
5125	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
5126	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0, supdrvGipSyncTimer, pDevExt);
5127	if (RT_SUCCESS(rc))
5128	{
5129	rc = RTMpNotificationRegister(supdrvGipMpEvent, pDevExt);
5130	if (RT_SUCCESS(rc))
5131	{
5132	rc = RTMpOnAll(supdrvGipInitOnCpu, pDevExt, pGip);
5133	if (RT_SUCCESS(rc))
5134	{
5135	/*
5136	* We're good.
5137	*/
5138	Log(("supdrvGipCreate: %u ns interval.\n", u32Interval));
5139	g_pSUPGlobalInfoPage = pGip;
5140	return VINF_SUCCESS;
5141	}
5142
5143	OSDBGPRINT(("supdrvGipCreate: RTMpOnAll failed with rc=%Rrc\n", rc));
5144	RTMpNotificationDeregister(supdrvGipMpEvent, pDevExt);
5145
5146	}
5147	else
5148	OSDBGPRINT(("supdrvGipCreate: failed to register MP event notfication. rc=%Rrc\n", rc));
5149	}
5150	else
5151	{
5152	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %u ns interval. rc=%Rrc\n", u32Interval, rc));
5153	Assert(!pDevExt->pGipTimer);
5154	}
5155	supdrvGipDestroy(pDevExt);
5156	return rc;
5157	}
5158
5159
5160	/**
5161	* Terminates the GIP.
5162	*
5163	* @param pDevExt Instance data. GIP stuff may be updated.
5164	*/
5165	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
5166	{
5167	int rc;
5168	#ifdef DEBUG_DARWIN_GIP
5169	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
5170	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
5171	pDevExt->pGipTimer, pDevExt->GipMemObj));
5172	#endif
5173
5174	/*
5175	* Invalid the GIP data.
5176	*/
5177	if (pDevExt->pGip)
5178	{
5179	supdrvGipTerm(pDevExt->pGip);
5180	pDevExt->pGip = NULL;
5181	}
5182	g_pSUPGlobalInfoPage = NULL;
5183
5184	/*
5185	* Destroy the timer and free the GIP memory object.
5186	*/
5187	if (pDevExt->pGipTimer)
5188	{
5189	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
5190	pDevExt->pGipTimer = NULL;
5191	}
5192
5193	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
5194	{
5195	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
5196	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
5197	}
5198
5199	/*
5200	* Finally, make sure we've release the system timer resolution request
5201	* if one actually succeeded and is still pending.
5202	*/
5203	if (pDevExt->u32SystemTimerGranularityGrant)
5204	{
5205	rc = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
5206	pDevExt->u32SystemTimerGranularityGrant = 0;
5207	}
5208	}
5209
5210
5211	/**
5212	* Timer callback function sync GIP mode.
5213	* @param pTimer The timer.
5214	* @param pvUser The device extension.
5215	*/
5216	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
5217	{
5218	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
5219	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5220	uint64_t u64TSC = ASMReadTSC();
5221	uint64_t NanoTS = RTTimeSystemNanoTS();
5222
5223	supdrvGipUpdate(pDevExt, NanoTS, u64TSC, NIL_RTCPUID, iTick);
5224
5225	ASMSetFlags(fOldFlags);
5226	}
5227
5228
5229	/**
5230	* Timer callback function for async GIP mode.
5231	* @param pTimer The timer.
5232	* @param pvUser The device extension.
5233	*/
5234	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
5235	{
5236	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
5237	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5238	RTCPUID idCpu = RTMpCpuId();
5239	uint64_t u64TSC = ASMReadTSC();
5240	uint64_t NanoTS = RTTimeSystemNanoTS();
5241
5242	/** @todo reset the transaction number and whatnot when iTick == 1. */
5243	if (pDevExt->idGipMaster == idCpu)
5244	supdrvGipUpdate(pDevExt, NanoTS, u64TSC, idCpu, iTick);
5245	else
5246	supdrvGipUpdatePerCpu(pDevExt, NanoTS, u64TSC, idCpu, ASMGetApicId(), iTick);
5247
5248	ASMSetFlags(fOldFlags);
5249	}
5250
5251
5252	/**
5253	* Finds our (@a idCpu) entry, or allocates a new one if not found.
5254	*
5255	* @returns Index of the CPU in the cache set.
5256	* @param pGip The GIP.
5257	* @param idCpu The CPU ID.
5258	*/
5259	static uint32_t supdrvGipCpuIndexFromCpuId(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
5260	{
5261	uint32_t i, cTries;
5262
5263	/*
5264	* ASSUMES that CPU IDs are constant.
5265	*/
5266	for (i = 0; i < pGip->cCpus; i++)
5267	if (pGip->aCPUs[i].idCpu == idCpu)
5268	return i;
5269
5270	cTries = 0;
5271	do
5272	{
5273	for (i = 0; i < pGip->cCpus; i++)
5274	{
5275	bool fRc;
5276	ASMAtomicCmpXchgSize(&pGip->aCPUs[i].idCpu, idCpu, NIL_RTCPUID, fRc);
5277	if (fRc)
5278	return i;
5279	}
5280	} while (cTries++ < 32);
5281	AssertReleaseFailed();
5282	return i - 1;
5283	}
5284
5285
5286	/**
5287	* The calling CPU should be accounted as online, update GIP accordingly.
5288	*
5289	* This is used by supdrvGipMpEvent as well as the supdrvGipCreate.
5290	*
5291	* @param pDevExt The device extension.
5292	* @param idCpu The CPU ID.
5293	*/
5294	static void supdrvGipMpEventOnline(PSUPDRVDEVEXT pDevExt, RTCPUID idCpu)
5295	{
5296	int iCpuSet = 0;
5297	uint16_t idApic = UINT16_MAX;
5298	uint32_t i = 0;
5299	uint64_t u64NanoTS = 0;
5300	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5301
5302	AssertPtrReturnVoid(pGip);
5303	AssertRelease(idCpu == RTMpCpuId());
5304	Assert(pGip->cPossibleCpus == RTMpGetCount());
5305
5306	/*
5307	* Do this behind a spinlock with interrupts disabled as this can fire
5308	* on all CPUs simultaneously, see @bugref{6110}.
5309	*/
5310	RTSpinlockAcquire(pDevExt->hGipSpinlock);
5311
5312	/*
5313	* Update the globals.
5314	*/
5315	ASMAtomicWriteU16(&pGip->cPresentCpus, RTMpGetPresentCount());
5316	ASMAtomicWriteU16(&pGip->cOnlineCpus, RTMpGetOnlineCount());
5317	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5318	if (iCpuSet >= 0)
5319	{
5320	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5321	RTCpuSetAddByIndex(&pGip->OnlineCpuSet, iCpuSet);
5322	RTCpuSetAddByIndex(&pGip->PresentCpuSet, iCpuSet);
5323	}
5324
5325	/*
5326	* Update the entry.
5327	*/
5328	u64NanoTS = RTTimeSystemNanoTS() - pGip->u32UpdateIntervalNS;
5329	i = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
5330	supdrvGipInitCpu(pGip, &pGip->aCPUs[i], u64NanoTS);
5331	idApic = ASMGetApicId();
5332	ASMAtomicWriteU16(&pGip->aCPUs[i].idApic, idApic);
5333	ASMAtomicWriteS16(&pGip->aCPUs[i].iCpuSet, (int16_t)iCpuSet);
5334	ASMAtomicWriteSize(&pGip->aCPUs[i].idCpu, idCpu);
5335
5336	/*
5337	* Update the APIC ID and CPU set index mappings.
5338	*/
5339	ASMAtomicWriteU16(&pGip->aiCpuFromApicId[idApic], i);
5340	ASMAtomicWriteU16(&pGip->aiCpuFromCpuSetIdx[iCpuSet], i);
5341
5342	/* commit it */
5343	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_ONLINE);
5344
5345	RTSpinlockReleaseNoInts(pDevExt->hGipSpinlock);
5346	}
5347
5348
5349	/**
5350	* The CPU should be accounted as offline, update the GIP accordingly.
5351	*
5352	* This is used by supdrvGipMpEvent.
5353	*
5354	* @param pDevExt The device extension.
5355	* @param idCpu The CPU ID.
5356	*/
5357	static void supdrvGipMpEventOffline(PSUPDRVDEVEXT pDevExt, RTCPUID idCpu)
5358	{
5359	int iCpuSet;
5360	unsigned i;
5361
5362	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5363
5364	AssertPtrReturnVoid(pGip);
5365	RTSpinlockAcquire(pDevExt->hGipSpinlock);
5366
5367	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5368	AssertReturnVoid(iCpuSet >= 0);
5369
5370	i = pGip->aiCpuFromCpuSetIdx[iCpuSet];
5371	AssertReturnVoid(i < pGip->cCpus);
5372	AssertReturnVoid(pGip->aCPUs[i].idCpu == idCpu);
5373
5374	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5375	RTCpuSetDelByIndex(&pGip->OnlineCpuSet, iCpuSet);
5376
5377	/* commit it */
5378	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_OFFLINE);
5379
5380	RTSpinlockReleaseNoInts(pDevExt->hGipSpinlock);
5381	}
5382
5383
5384	/**
5385	* Multiprocessor event notification callback.
5386	*
5387	* This is used to make sure that the GIP master gets passed on to
5388	* another CPU. It also updates the associated CPU data.
5389	*
5390	* @param enmEvent The event.
5391	* @param idCpu The cpu it applies to.
5392	* @param pvUser Pointer to the device extension.
5393	*
5394	* @remarks This function -must- fire on the newly online'd CPU for the
5395	* RTMPEVENT_ONLINE case and can fire on any CPU for the
5396	* RTMPEVENT_OFFLINE case.
5397	*/
5398	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser)
5399	{
5400	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5401	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5402
5403	AssertRelease(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
5404
5405	/*
5406	* Update the GIP CPU data.
5407	*/
5408	if (pGip)
5409	{
5410	switch (enmEvent)
5411	{
5412	case RTMPEVENT_ONLINE:
5413	AssertRelease(idCpu == RTMpCpuId());
5414	supdrvGipMpEventOnline(pDevExt, idCpu);
5415	break;
5416	case RTMPEVENT_OFFLINE:
5417	supdrvGipMpEventOffline(pDevExt, idCpu);
5418	break;
5419
5420	}
5421	}
5422
5423	/*
5424	* Make sure there is a master GIP.
5425	*/
5426	if (enmEvent == RTMPEVENT_OFFLINE)
5427	{
5428	RTCPUID idGipMaster = ASMAtomicReadU32(&pDevExt->idGipMaster);
5429	if (idGipMaster == idCpu)
5430	{
5431	/*
5432	* Find a new GIP master.
5433	*/
5434	bool fIgnored;
5435	unsigned i;
5436	RTCPUID idNewGipMaster = NIL_RTCPUID;
5437	RTCPUSET OnlineCpus;
5438	RTMpGetOnlineSet(&OnlineCpus);
5439
5440	for (i = 0; i < RTCPUSET_MAX_CPUS; i++)
5441	{
5442	RTCPUID idCurCpu = RTMpCpuIdFromSetIndex(i);
5443	if ( RTCpuSetIsMember(&OnlineCpus, idCurCpu)
5444	&& idCurCpu != idGipMaster)
5445	{
5446	idNewGipMaster = idCurCpu;
5447	break;
5448	}
5449	}
5450
5451	Log(("supdrvGipMpEvent: Gip master %#lx -> %#lx\n", (long)idGipMaster, (long)idNewGipMaster));
5452	ASMAtomicCmpXchgSize(&pDevExt->idGipMaster, idNewGipMaster, idGipMaster, fIgnored);
5453	NOREF(fIgnored);
5454	}
5455	}
5456	}
5457
5458
5459	/**
5460	* Callback used by supdrvDetermineAsyncTSC to read the TSC on a CPU.
5461	*
5462	* @param idCpu Ignored.
5463	* @param pvUser1 Where to put the TSC.
5464	* @param pvUser2 Ignored.
5465	*/
5466	static DECLCALLBACK(void) supdrvDetermineAsyncTscWorker(RTCPUID idCpu, void pvUser1, void pvUser2)
5467	{
5468	#if 1
5469	ASMAtomicWriteU64((uint64_t volatile *)pvUser1, ASMReadTSC());
5470	#else
5471	(uint64_t )pvUser1 = ASMReadTSC();
5472	#endif
5473	}
5474
5475
5476	/**
5477	* Determine if Async GIP mode is required because of TSC drift.
5478	*
5479	* When using the default/normal timer code it is essential that the time stamp counter
5480	* (TSC) runs never backwards, that is, a read operation to the counter should return
5481	* a bigger value than any previous read operation. This is guaranteed by the latest
5482	* AMD CPUs and by newer Intel CPUs which never enter the C2 state (P4). In any other
5483	* case we have to choose the asynchronous timer mode.
5484	*
5485	* @param poffMin Pointer to the determined difference between different cores.
5486	* @return false if the time stamp counters appear to be synchronized, true otherwise.
5487	*/
5488	static bool supdrvDetermineAsyncTsc(uint64_t *poffMin)
5489	{
5490	/*
5491	* Just iterate all the cpus 8 times and make sure that the TSC is
5492	* ever increasing. We don't bother taking TSC rollover into account.
5493	*/
5494	int iEndCpu = RTMpGetArraySize();
5495	int iCpu;
5496	int cLoops = 8;
5497	bool fAsync = false;
5498	int rc = VINF_SUCCESS;
5499	uint64_t offMax = 0;
5500	uint64_t offMin = ~(uint64_t)0;
5501	uint64_t PrevTsc = ASMReadTSC();
5502
5503	while (cLoops-- > 0)
5504	{
5505	for (iCpu = 0; iCpu < iEndCpu; iCpu++)
5506	{
5507	uint64_t CurTsc;
5508	rc = RTMpOnSpecific(RTMpCpuIdFromSetIndex(iCpu), supdrvDetermineAsyncTscWorker, &CurTsc, NULL);
5509	if (RT_SUCCESS(rc))
5510	{
5511	if (CurTsc <= PrevTsc)
5512	{
5513	fAsync = true;
5514	offMin = offMax = PrevTsc - CurTsc;
5515	Log(("supdrvDetermineAsyncTsc: iCpu=%d cLoops=%d CurTsc=%llx PrevTsc=%llx\n",
5516	iCpu, cLoops, CurTsc, PrevTsc));
5517	break;
5518	}
5519
5520	/* Gather statistics (except the first time). */
5521	if (iCpu != 0 \|\| cLoops != 7)
5522	{
5523	uint64_t off = CurTsc - PrevTsc;
5524	if (off < offMin)
5525	offMin = off;
5526	if (off > offMax)
5527	offMax = off;
5528	Log2(("%d/%d: off=%llx\n", cLoops, iCpu, off));
5529	}
5530
5531	/* Next */
5532	PrevTsc = CurTsc;
5533	}
5534	else if (rc == VERR_NOT_SUPPORTED)
5535	break;
5536	else
5537	AssertMsg(rc == VERR_CPU_NOT_FOUND \|\| rc == VERR_CPU_OFFLINE, ("%d\n", rc));
5538	}
5539
5540	/* broke out of the loop. */
5541	if (iCpu < iEndCpu)
5542	break;
5543	}
5544
5545	poffMin = offMin; / Almost RTMpOnSpecific profiling. */
5546	Log(("supdrvDetermineAsyncTsc: returns %d; iEndCpu=%d rc=%d offMin=%llx offMax=%llx\n",
5547	fAsync, iEndCpu, rc, offMin, offMax));
5548	#if !defined(RT_OS_SOLARIS) && !defined(RT_OS_OS2) && !defined(RT_OS_WINDOWS)
5549	OSDBGPRINT(("vboxdrv: fAsync=%d offMin=%#lx offMax=%#lx\n", fAsync, (long)offMin, (long)offMax));
5550	#endif
5551	return fAsync;
5552	}
5553
5554
5555	/**
5556	* Determine the GIP TSC mode.
5557	*
5558	* @returns The most suitable TSC mode.
5559	* @param pDevExt Pointer to the device instance data.
5560	*/
5561	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt)
5562	{
5563	/*
5564	* On SMP we're faced with two problems:
5565	* (1) There might be a skew between the CPU, so that cpu0
5566	* returns a TSC that is slightly different from cpu1.
5567	* (2) Power management (and other things) may cause the TSC
5568	* to run at a non-constant speed, and cause the speed
5569	* to be different on the cpus. This will result in (1).
5570	*
5571	* So, on SMP systems we'll have to select the ASYNC update method
5572	* if there are symptoms of these problems.
5573	*/
5574	if (RTMpGetCount() > 1)
5575	{
5576	uint32_t uEAX, uEBX, uECX, uEDX;
5577	uint64_t u64DiffCoresIgnored;
5578
5579	/* Permit the user and/or the OS specific bits to force async mode. */
5580	if (supdrvOSGetForcedAsyncTscMode(pDevExt))
5581	return SUPGIPMODE_ASYNC_TSC;
5582
5583	/* Try check for current differences between the cpus. */
5584	if (supdrvDetermineAsyncTsc(&u64DiffCoresIgnored))
5585	return SUPGIPMODE_ASYNC_TSC;
5586
5587	/*
5588	* If the CPU supports power management and is an AMD one we
5589	* won't trust it unless it has the TscInvariant bit is set.
5590	*/
5591	/* Check for "AuthenticAMD" */
5592	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
5593	if ( uEAX >= 1
5594	&& ASMIsAmdCpuEx(uEBX, uECX, uEDX))
5595	{
5596	/* Check for APM support and that TscInvariant is cleared. */
5597	ASMCpuId(0x80000000, &uEAX, &uEBX, &uECX, &uEDX);
5598	if (uEAX >= 0x80000007)
5599	{
5600	ASMCpuId(0x80000007, &uEAX, &uEBX, &uECX, &uEDX);
5601	if ( !(uEDX & RT_BIT(8))/* TscInvariant */
5602	&& (uEDX & 0x3e)) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
5603	return SUPGIPMODE_ASYNC_TSC;
5604	}
5605	}
5606	}
5607	return SUPGIPMODE_SYNC_TSC;
5608	}
5609
5610
5611	/**
5612	* Initializes per-CPU GIP information.
5613	*
5614	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5615	* @param pCpu Pointer to which GIP CPU to initalize.
5616	* @param u64NanoTS The current nanosecond timestamp.
5617	*/
5618	static void supdrvGipInitCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pCpu, uint64_t u64NanoTS)
5619	{
5620	pCpu->u32TransactionId = 2;
5621	pCpu->u64NanoTS = u64NanoTS;
5622	pCpu->u64TSC = ASMReadTSC();
5623
5624	ASMAtomicWriteSize(&pCpu->enmState, SUPGIPCPUSTATE_INVALID);
5625	ASMAtomicWriteSize(&pCpu->idCpu, NIL_RTCPUID);
5626	ASMAtomicWriteS16(&pCpu->iCpuSet, -1);
5627	ASMAtomicWriteU16(&pCpu->idApic, UINT16_MAX);
5628
5629	/*
5630	* We don't know the following values until we've executed updates.
5631	* So, we'll just pretend it's a 4 GHz CPU and adjust the history it on
5632	* the 2nd timer callout.
5633	*/
5634	pCpu->u64CpuHz = _4G + 1; /* tstGIP-2 depends on this. */
5635	pCpu->u32UpdateIntervalTSC
5636	= pCpu->au32TSCHistory[0]
5637	= pCpu->au32TSCHistory[1]
5638	= pCpu->au32TSCHistory[2]
5639	= pCpu->au32TSCHistory[3]
5640	= pCpu->au32TSCHistory[4]
5641	= pCpu->au32TSCHistory[5]
5642	= pCpu->au32TSCHistory[6]
5643	= pCpu->au32TSCHistory[7]
5644	= (uint32_t)(_4G / pGip->u32UpdateHz);
5645	}
5646
5647
5648	/**
5649	* Initializes the GIP data.
5650	*
5651	* @param pDevExt Pointer to the device instance data.
5652	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5653	* @param HCPhys The physical address of the GIP.
5654	* @param u64NanoTS The current nanosecond timestamp.
5655	* @param uUpdateHz The update frequency.
5656	* @param cCpus The CPU count.
5657	*/
5658	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
5659	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned cCpus)
5660	{
5661	size_t const cbGip = RT_ALIGN_Z(RT_OFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), PAGE_SIZE);
5662	unsigned i;
5663	#ifdef DEBUG_DARWIN_GIP
5664	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5665	#else
5666	LogFlow(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5667	#endif
5668
5669	/*
5670	* Initialize the structure.
5671	*/
5672	memset(pGip, 0, cbGip);
5673	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
5674	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
5675	pGip->u32Mode = supdrvGipDeterminTscMode(pDevExt);
5676	pGip->cCpus = (uint16_t)cCpus;
5677	pGip->cPages = (uint16_t)(cbGip / PAGE_SIZE);
5678	pGip->u32UpdateHz = uUpdateHz;
5679	pGip->u32UpdateIntervalNS = 1000000000 / uUpdateHz;
5680	pGip->u64NanoTSLastUpdateHz = u64NanoTS;
5681	RTCpuSetEmpty(&pGip->OnlineCpuSet);
5682	RTCpuSetEmpty(&pGip->PresentCpuSet);
5683	RTMpGetSet(&pGip->PossibleCpuSet);
5684	pGip->cOnlineCpus = RTMpGetOnlineCount();
5685	pGip->cPresentCpus = RTMpGetPresentCount();
5686	pGip->cPossibleCpus = RTMpGetCount();
5687	pGip->idCpuMax = RTMpGetMaxCpuId();
5688	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromApicId); i++)
5689	pGip->aiCpuFromApicId[i] = 0;
5690	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx); i++)
5691	pGip->aiCpuFromCpuSetIdx[i] = UINT16_MAX;
5692
5693	for (i = 0; i < cCpus; i++)
5694	supdrvGipInitCpu(pGip, &pGip->aCPUs[i], u64NanoTS);
5695
5696	/*
5697	* Link it to the device extension.
5698	*/
5699	pDevExt->pGip = pGip;
5700	pDevExt->HCPhysGip = HCPhys;
5701	pDevExt->cGipUsers = 0;
5702	}
5703
5704
5705	/**
5706	* On CPU initialization callback for RTMpOnAll.
5707	*
5708	* @param idCpu The CPU ID.
5709	* @param pvUser1 The device extension.
5710	* @param pvUser2 The GIP.
5711	*/
5712	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2)
5713	{
5714	/* This is good enough, even though it will update some of the globals a
5715	bit to much. */
5716	supdrvGipMpEventOnline((PSUPDRVDEVEXT)pvUser1, idCpu);
5717	}
5718
5719
5720	/**
5721	* Invalidates the GIP data upon termination.
5722	*
5723	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5724	*/
5725	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
5726	{
5727	unsigned i;
5728	pGip->u32Magic = 0;
5729	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
5730	{
5731	pGip->aCPUs[i].u64NanoTS = 0;
5732	pGip->aCPUs[i].u64TSC = 0;
5733	pGip->aCPUs[i].iTSCHistoryHead = 0;
5734	}
5735	}
5736
5737
5738	/**
5739	* Worker routine for supdrvGipUpdate and supdrvGipUpdatePerCpu that
5740	* updates all the per cpu data except the transaction id.
5741	*
5742	* @param pDevExt The device extension.
5743	* @param pGipCpu Pointer to the per cpu data.
5744	* @param u64NanoTS The current time stamp.
5745	* @param u64TSC The current TSC.
5746	* @param iTick The current timer tick.
5747	*/
5748	static void supdrvGipDoUpdateCpu(PSUPDRVDEVEXT pDevExt, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick)
5749	{
5750	uint64_t u64TSCDelta;
5751	uint32_t u32UpdateIntervalTSC;
5752	uint32_t u32UpdateIntervalTSCSlack;
5753	unsigned iTSCHistoryHead;
5754	uint64_t u64CpuHz;
5755	uint32_t u32TransactionId;
5756
5757	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5758	AssertPtrReturnVoid(pGip);
5759
5760	/* Delta between this and the previous update. */
5761	ASMAtomicUoWriteU32(&pGipCpu->u32PrevUpdateIntervalNS, (uint32_t)(u64NanoTS - pGipCpu->u64NanoTS));
5762
5763	/*
5764	* Update the NanoTS.
5765	*/
5766	ASMAtomicWriteU64(&pGipCpu->u64NanoTS, u64NanoTS);
5767
5768	/*
5769	* Calc TSC delta.
5770	*/
5771	/** @todo validate the NanoTS delta, don't trust the OS to call us when it should... */
5772	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
5773	ASMAtomicWriteU64(&pGipCpu->u64TSC, u64TSC);
5774
5775	if (u64TSCDelta >> 32)
5776	{
5777	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
5778	pGipCpu->cErrors++;
5779	}
5780
5781	/*
5782	* On the 2nd and 3rd callout, reset the history with the current TSC
5783	* interval since the values entered by supdrvGipInit are totally off.
5784	* The interval on the 1st callout completely unreliable, the 2nd is a bit
5785	* better, while the 3rd should be most reliable.
5786	*/
5787	u32TransactionId = pGipCpu->u32TransactionId;
5788	if (RT_UNLIKELY( ( u32TransactionId == 5
5789	\|\| u32TransactionId == 7)
5790	&& ( iTick == 2
5791	\|\| iTick == 3) ))
5792	{
5793	unsigned i;
5794	for (i = 0; i < RT_ELEMENTS(pGipCpu->au32TSCHistory); i++)
5795	ASMAtomicUoWriteU32(&pGipCpu->au32TSCHistory[i], (uint32_t)u64TSCDelta);
5796	}
5797
5798	/*
5799	* TSC History.
5800	*/
5801	Assert(RT_ELEMENTS(pGipCpu->au32TSCHistory) == 8);
5802	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
5803	ASMAtomicWriteU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
5804	ASMAtomicWriteU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
5805
5806	/*
5807	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
5808	*/
5809	if (pGip->u32UpdateHz >= 1000)
5810	{
5811	uint32_t u32;
5812	u32 = pGipCpu->au32TSCHistory[0];
5813	u32 += pGipCpu->au32TSCHistory[1];
5814	u32 += pGipCpu->au32TSCHistory[2];
5815	u32 += pGipCpu->au32TSCHistory[3];
5816	u32 >>= 2;
5817	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
5818	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
5819	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
5820	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
5821	u32UpdateIntervalTSC >>= 2;
5822	u32UpdateIntervalTSC += u32;
5823	u32UpdateIntervalTSC >>= 1;
5824
5825	/* Value chosen for a 2GHz Athlon64 running linux 2.6.10/11, . */
5826	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
5827	}
5828	else if (pGip->u32UpdateHz >= 90)
5829	{
5830	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5831	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
5832	u32UpdateIntervalTSC >>= 1;
5833
5834	/* value chosen on a 2GHz thinkpad running windows */
5835	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
5836	}
5837	else
5838	{
5839	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5840
5841	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
5842	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
5843	}
5844	ASMAtomicWriteU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
5845
5846	/*
5847	* CpuHz.
5848	*/
5849	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, pGip->u32UpdateHz);
5850	ASMAtomicWriteU64(&pGipCpu->u64CpuHz, u64CpuHz);
5851	}
5852
5853
5854	/**
5855	* Updates the GIP.
5856	*
5857	* @param pDevExt The device extension.
5858	* @param u64NanoTS The current nanosecond timesamp.
5859	* @param u64TSC The current TSC timesamp.
5860	* @param idCpu The CPU ID.
5861	* @param iTick The current timer tick.
5862	*/
5863	static void supdrvGipUpdate(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick)
5864	{
5865	/*
5866	* Determine the relevant CPU data.
5867	*/
5868	PSUPGIPCPU pGipCpu;
5869	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5870	AssertPtrReturnVoid(pGip);
5871
5872	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
5873	pGipCpu = &pGip->aCPUs[0];
5874	else
5875	{
5876	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
5877	if (RT_UNLIKELY(iCpu >= pGip->cCpus))
5878	return;
5879	pGipCpu = &pGip->aCPUs[iCpu];
5880	if (RT_UNLIKELY(pGipCpu->idCpu != idCpu))
5881	return;
5882	}
5883
5884	/*
5885	* Start update transaction.
5886	*/
5887	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5888	{
5889	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
5890	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5891	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5892	pGipCpu->cErrors++;
5893	return;
5894	}
5895
5896	/*
5897	* Recalc the update frequency every 0x800th time.
5898	*/
5899	if (!(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
5900	{
5901	if (pGip->u64NanoTSLastUpdateHz)
5902	{
5903	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
5904	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
5905	uint32_t u32UpdateHz = (uint32_t)((UINT64_C(1000000000) * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
5906	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
5907	{
5908	ASMAtomicWriteU32(&pGip->u32UpdateHz, u32UpdateHz);
5909	ASMAtomicWriteU32(&pGip->u32UpdateIntervalNS, 1000000000 / u32UpdateHz);
5910	}
5911	#endif
5912	}
5913	ASMAtomicWriteU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS);
5914	}
5915
5916	/*
5917	* Update the data.
5918	*/
5919	supdrvGipDoUpdateCpu(pDevExt, pGipCpu, u64NanoTS, u64TSC, iTick);
5920
5921	/*
5922	* Complete transaction.
5923	*/
5924	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5925	}
5926
5927
5928	/**
5929	* Updates the per cpu GIP data for the calling cpu.
5930	*
5931	* @param pDevExt The device extension.
5932	* @param u64NanoTS The current nanosecond timesamp.
5933	* @param u64TSC The current TSC timesamp.
5934	* @param idCpu The CPU ID.
5935	* @param idApic The APIC id for the CPU index.
5936	* @param iTick The current timer tick.
5937	*/
5938	static void supdrvGipUpdatePerCpu(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC,
5939	RTCPUID idCpu, uint8_t idApic, uint64_t iTick)
5940	{
5941	uint32_t iCpu;
5942	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5943
5944	/*
5945	* Avoid a potential race when a CPU online notification doesn't fire on
5946	* the onlined CPU but the tick creeps in before the event notification is
5947	* run.
5948	*/
5949	if (RT_UNLIKELY(iTick == 1))
5950	{
5951	iCpu = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
5952	if (pGip->aCPUs[iCpu].enmState == SUPGIPCPUSTATE_OFFLINE)
5953	supdrvGipMpEventOnline(pDevExt, idCpu);
5954	}
5955
5956	iCpu = pGip->aiCpuFromApicId[idApic];
5957	if (RT_LIKELY(iCpu < pGip->cCpus))
5958	{
5959	PSUPGIPCPU pGipCpu = &pGip->aCPUs[iCpu];
5960	if (pGipCpu->idCpu == idCpu)
5961	{
5962	/*
5963	* Start update transaction.
5964	*/
5965	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5966	{
5967	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5968	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5969	pGipCpu->cErrors++;
5970	return;
5971	}
5972
5973	/*
5974	* Update the data.
5975	*/
5976	supdrvGipDoUpdateCpu(pDevExt, pGipCpu, u64NanoTS, u64TSC, iTick);
5977
5978	/*
5979	* Complete transaction.
5980	*/
5981	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5982	}
5983	}
5984	}
5985

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

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDrv.c@ 47363

Download in other formats: