SUPDrv.c@ 46889

Last change on this file since 46889 was 46889, checked in by vboxsync, 11 years ago
amd64 is fine
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 222.7 KB

Line
1	/* $Id: SUPDrv.c 46889 2013-07-01 15:49:59Z 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	{ "RTThreadGetName", (void *)RTThreadGetName },
370	{ "RTThreadGetNative", (void *)RTThreadGetNative },
371	{ "RTThreadGetType", (void *)RTThreadGetType },
372	{ "RTThreadIsInInterrupt", (void *)RTThreadIsInInterrupt },
373	{ "RTThreadNativeSelf", (void *)RTThreadNativeSelf },
374	{ "RTThreadPreemptDisable", (void *)RTThreadPreemptDisable },
375	{ "RTThreadPreemptIsEnabled", (void *)RTThreadPreemptIsEnabled },
376	{ "RTThreadPreemptIsPending", (void *)RTThreadPreemptIsPending },
377	{ "RTThreadPreemptIsPendingTrusty", (void *)RTThreadPreemptIsPendingTrusty },
378	{ "RTThreadPreemptIsPossible", (void *)RTThreadPreemptIsPossible },
379	{ "RTThreadPreemptRestore", (void *)RTThreadPreemptRestore },
380	{ "RTThreadSelf", (void *)RTThreadSelf },
381	{ "RTThreadSelfName", (void *)RTThreadSelfName },
382	{ "RTThreadSleep", (void *)RTThreadSleep },
383	{ "RTThreadUserReset", (void *)RTThreadUserReset },
384	{ "RTThreadUserSignal", (void *)RTThreadUserSignal },
385	{ "RTThreadUserWait", (void *)RTThreadUserWait },
386	{ "RTThreadUserWaitNoResume", (void *)RTThreadUserWaitNoResume },
387	{ "RTThreadWait", (void *)RTThreadWait },
388	{ "RTThreadWaitNoResume", (void *)RTThreadWaitNoResume },
389	{ "RTThreadYield", (void *)RTThreadYield },
390	{ "RTTimeMilliTS", (void *)RTTimeMilliTS },
391	{ "RTTimeNanoTS", (void *)RTTimeNanoTS },
392	{ "RTTimeNow", (void *)RTTimeNow },
393	{ "RTTimerCanDoHighResolution", (void *)RTTimerCanDoHighResolution },
394	{ "RTTimerChangeInterval", (void *)RTTimerChangeInterval },
395	{ "RTTimerCreate", (void *)RTTimerCreate },
396	{ "RTTimerCreateEx", (void *)RTTimerCreateEx },
397	{ "RTTimerDestroy", (void *)RTTimerDestroy },
398	{ "RTTimerGetSystemGranularity", (void *)RTTimerGetSystemGranularity },
399	{ "RTTimerReleaseSystemGranularity", (void *)RTTimerReleaseSystemGranularity },
400	{ "RTTimerRequestSystemGranularity", (void *)RTTimerRequestSystemGranularity },
401	{ "RTTimerStart", (void *)RTTimerStart },
402	{ "RTTimerStop", (void *)RTTimerStop },
403	{ "RTTimeSystemMilliTS", (void *)RTTimeSystemMilliTS },
404	{ "RTTimeSystemNanoTS", (void *)RTTimeSystemNanoTS },
405	{ "RTUuidCompare", (void *)RTUuidCompare },
406	{ "RTUuidCompareStr", (void *)RTUuidCompareStr },
407	{ "RTUuidFromStr", (void *)RTUuidFromStr },
408	/* SED: END */
409	};
410
411	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
412	/**
413	* Drag in the rest of IRPT since we share it with the
414	* rest of the kernel modules on darwin.
415	*/
416	PFNRT g_apfnVBoxDrvIPRTDeps[] =
417	{
418	/* VBoxNetAdp */
419	(PFNRT)RTRandBytes,
420	/* VBoxUSB */
421	(PFNRT)RTPathStripFilename,
422	NULL
423	};
424	#endif /* RT_OS_DARWIN \|\| RT_OS_SOLARIS \|\| RT_OS_SOLARIS */
425
426
427	/**
428	* Initializes the device extentsion structure.
429	*
430	* @returns IPRT status code.
431	* @param pDevExt The device extension to initialize.
432	* @param cbSession The size of the session structure. The size of
433	* SUPDRVSESSION may be smaller when SUPDRV_AGNOSTIC is
434	* defined because we're skipping the OS specific members
435	* then.
436	*/
437	int VBOXCALL supdrvInitDevExt(PSUPDRVDEVEXT pDevExt, size_t cbSession)
438	{
439	int rc;
440
441	#ifdef SUPDRV_WITH_RELEASE_LOGGER
442	/*
443	* Create the release log.
444	*/
445	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
446	PRTLOGGER pRelLogger;
447	rc = RTLogCreate(&pRelLogger, 0 /* fFlags */, "all",
448	"VBOX_RELEASE_LOG", RT_ELEMENTS(s_apszGroups), s_apszGroups, RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER, NULL);
449	if (RT_SUCCESS(rc))
450	RTLogRelSetDefaultInstance(pRelLogger);
451	/** @todo Add native hook for getting logger config parameters and setting
452	* them. On linux we should use the module parameter stuff... */
453	#endif
454
455	/*
456	* Initialize it.
457	*/
458	memset(pDevExt, 0, sizeof(*pDevExt));
459	rc = RTSpinlockCreate(&pDevExt->Spinlock, RTSPINLOCK_FLAGS_INTERRUPT_SAFE, "SUPDrvDevExt");
460	if (RT_SUCCESS(rc))
461	{
462	rc = RTSpinlockCreate(&pDevExt->hGipSpinlock, RTSPINLOCK_FLAGS_INTERRUPT_SAFE, "SUPDrvGip");
463	if (RT_SUCCESS(rc))
464	{
465	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
466	rc = RTSemMutexCreate(&pDevExt->mtxLdr);
467	#else
468	rc = RTSemFastMutexCreate(&pDevExt->mtxLdr);
469	#endif
470	if (RT_SUCCESS(rc))
471	{
472	rc = RTSemFastMutexCreate(&pDevExt->mtxComponentFactory);
473	if (RT_SUCCESS(rc))
474	{
475	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
476	rc = RTSemMutexCreate(&pDevExt->mtxGip);
477	#else
478	rc = RTSemFastMutexCreate(&pDevExt->mtxGip);
479	#endif
480	if (RT_SUCCESS(rc))
481	{
482	rc = supdrvGipCreate(pDevExt);
483	if (RT_SUCCESS(rc))
484	{
485	rc = supdrvTracerInit(pDevExt);
486	if (RT_SUCCESS(rc))
487	{
488	pDevExt->pLdrInitImage = NULL;
489	pDevExt->hLdrInitThread = NIL_RTNATIVETHREAD;
490	pDevExt->u32Cookie = BIRD; /** @todo make this random? */
491	pDevExt->cbSession = (uint32_t)cbSession;
492
493	/*
494	* Fixup the absolute symbols.
495	*
496	* Because of the table indexing assumptions we'll have a little #ifdef orgy
497	* here rather than distributing this to OS specific files. At least for now.
498	*/
499	#ifdef RT_OS_DARWIN
500	# if ARCH_BITS == 32
501	if (SUPR0GetPagingMode() >= SUPPAGINGMODE_AMD64)
502	{
503	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
504	g_aFunctions[1].pfn = (void )0x80; / SUPR0Abs64bitKernelCS - KERNEL64_CS, seg.h */
505	g_aFunctions[2].pfn = (void )0x88; / SUPR0Abs64bitKernelSS - KERNEL64_SS, seg.h */
506	g_aFunctions[3].pfn = (void )0x88; / SUPR0Abs64bitKernelDS - KERNEL64_SS, seg.h */
507	}
508	else
509	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
510	g_aFunctions[4].pfn = (void )0x08; / SUPR0AbsKernelCS - KERNEL_CS, seg.h */
511	g_aFunctions[5].pfn = (void )0x10; / SUPR0AbsKernelSS - KERNEL_DS, seg.h */
512	g_aFunctions[6].pfn = (void )0x10; / SUPR0AbsKernelDS - KERNEL_DS, seg.h */
513	g_aFunctions[7].pfn = (void )0x10; / SUPR0AbsKernelES - KERNEL_DS, seg.h */
514	g_aFunctions[8].pfn = (void )0x10; / SUPR0AbsKernelFS - KERNEL_DS, seg.h */
515	g_aFunctions[9].pfn = (void )0x48; / SUPR0AbsKernelGS - CPU_DATA_GS, seg.h */
516	# else /* 64-bit darwin: */
517	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
518	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
519	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
520	g_aFunctions[3].pfn = (void )0; / SUPR0Abs64bitKernelDS */
521	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
522	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
523	g_aFunctions[6].pfn = (void )0; / SUPR0AbsKernelDS */
524	g_aFunctions[7].pfn = (void )0; / SUPR0AbsKernelES */
525	g_aFunctions[8].pfn = (void )0; / SUPR0AbsKernelFS */
526	g_aFunctions[9].pfn = (void )0; / SUPR0AbsKernelGS */
527
528	# endif
529	#else /* !RT_OS_DARWIN */
530	# if ARCH_BITS == 64
531	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
532	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
533	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
534	g_aFunctions[3].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0Abs64bitKernelDS */
535	# else
536	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
537	# endif
538	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
539	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
540	g_aFunctions[6].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0AbsKernelDS */
541	g_aFunctions[7].pfn = (void )(uintptr_t)ASMGetES(); / SUPR0AbsKernelES */
542	g_aFunctions[8].pfn = (void )(uintptr_t)ASMGetFS(); / SUPR0AbsKernelFS */
543	g_aFunctions[9].pfn = (void )(uintptr_t)ASMGetGS(); / SUPR0AbsKernelGS */
544	#endif /* !RT_OS_DARWIN */
545	return VINF_SUCCESS;
546	}
547
548	supdrvGipDestroy(pDevExt);
549	}
550
551	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
552	RTSemMutexDestroy(pDevExt->mtxGip);
553	pDevExt->mtxGip = NIL_RTSEMMUTEX;
554	#else
555	RTSemFastMutexDestroy(pDevExt->mtxGip);
556	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
557	#endif
558	}
559	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
560	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
561	}
562	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
563	RTSemMutexDestroy(pDevExt->mtxLdr);
564	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
565	#else
566	RTSemFastMutexDestroy(pDevExt->mtxLdr);
567	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
568	#endif
569	}
570	RTSpinlockDestroy(pDevExt->hGipSpinlock);
571	pDevExt->hGipSpinlock = NIL_RTSPINLOCK;
572	}
573	RTSpinlockDestroy(pDevExt->Spinlock);
574	pDevExt->Spinlock = NIL_RTSPINLOCK;
575	}
576	#ifdef SUPDRV_WITH_RELEASE_LOGGER
577	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
578	RTLogDestroy(RTLogSetDefaultInstance(NULL));
579	#endif
580
581	return rc;
582	}
583
584
585	/**
586	* Delete the device extension (e.g. cleanup members).
587	*
588	* @param pDevExt The device extension to delete.
589	*/
590	void VBOXCALL supdrvDeleteDevExt(PSUPDRVDEVEXT pDevExt)
591	{
592	PSUPDRVOBJ pObj;
593	PSUPDRVUSAGE pUsage;
594
595	/*
596	* Kill mutexes and spinlocks.
597	*/
598	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
599	RTSemMutexDestroy(pDevExt->mtxGip);
600	pDevExt->mtxGip = NIL_RTSEMMUTEX;
601	#else
602	RTSemFastMutexDestroy(pDevExt->mtxGip);
603	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
604	#endif
605	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
606	RTSemMutexDestroy(pDevExt->mtxLdr);
607	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
608	#else
609	RTSemFastMutexDestroy(pDevExt->mtxLdr);
610	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
611	#endif
612	RTSpinlockDestroy(pDevExt->Spinlock);
613	pDevExt->Spinlock = NIL_RTSPINLOCK;
614	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
615	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
616
617	/*
618	* Free lists.
619	*/
620	/* objects. */
621	pObj = pDevExt->pObjs;
622	Assert(!pObj); /* (can trigger on forced unloads) */
623	pDevExt->pObjs = NULL;
624	while (pObj)
625	{
626	void *pvFree = pObj;
627	pObj = pObj->pNext;
628	RTMemFree(pvFree);
629	}
630
631	/* usage records. */
632	pUsage = pDevExt->pUsageFree;
633	pDevExt->pUsageFree = NULL;
634	while (pUsage)
635	{
636	void *pvFree = pUsage;
637	pUsage = pUsage->pNext;
638	RTMemFree(pvFree);
639	}
640
641	/* kill the GIP. */
642	supdrvGipDestroy(pDevExt);
643	RTSpinlockDestroy(pDevExt->hGipSpinlock);
644	pDevExt->hGipSpinlock = NIL_RTSPINLOCK;
645
646	supdrvTracerTerm(pDevExt);
647
648	#ifdef SUPDRV_WITH_RELEASE_LOGGER
649	/* destroy the loggers. */
650	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
651	RTLogDestroy(RTLogSetDefaultInstance(NULL));
652	#endif
653	}
654
655
656	/**
657	* Create session.
658	*
659	* @returns IPRT status code.
660	* @param pDevExt Device extension.
661	* @param fUser Flag indicating whether this is a user or kernel
662	* session.
663	* @param fUnrestricted Unrestricted access (system) or restricted access
664	* (user)?
665	* @param ppSession Where to store the pointer to the session data.
666	*/
667	int VBOXCALL supdrvCreateSession(PSUPDRVDEVEXT pDevExt, bool fUser, bool fUnrestricted, PSUPDRVSESSION *ppSession)
668	{
669	int rc;
670	PSUPDRVSESSION pSession;
671
672	if (!SUP_IS_DEVEXT_VALID(pDevExt))
673	return VERR_INVALID_PARAMETER;
674
675	/*
676	* Allocate memory for the session data.
677	*/
678	pSession = *ppSession = (PSUPDRVSESSION)RTMemAllocZ(pDevExt->cbSession);
679	if (pSession)
680	{
681	/* Initialize session data. */
682	rc = RTSpinlockCreate(&pSession->Spinlock, RTSPINLOCK_FLAGS_INTERRUPT_UNSAFE, "SUPDrvSession");
683	if (!rc)
684	{
685	rc = RTHandleTableCreateEx(&pSession->hHandleTable,
686	RTHANDLETABLE_FLAGS_LOCKED_IRQ_SAFE \| RTHANDLETABLE_FLAGS_CONTEXT,
687	1 /uBase/, 32768 /cMax/, supdrvSessionObjHandleRetain, pSession);
688	if (RT_SUCCESS(rc))
689	{
690	Assert(pSession->Spinlock != NIL_RTSPINLOCK);
691	pSession->pDevExt = pDevExt;
692	pSession->u32Cookie = BIRD_INV;
693	pSession->fUnrestricted = fUnrestricted;
694	/*pSession->pLdrUsage = NULL;
695	pSession->pVM = NULL;
696	pSession->pUsage = NULL;
697	pSession->pGip = NULL;
698	pSession->fGipReferenced = false;
699	pSession->Bundle.cUsed = 0; */
700	pSession->Uid = NIL_RTUID;
701	pSession->Gid = NIL_RTGID;
702	if (fUser)
703	{
704	pSession->Process = RTProcSelf();
705	pSession->R0Process = RTR0ProcHandleSelf();
706	}
707	else
708	{
709	pSession->Process = NIL_RTPROCESS;
710	pSession->R0Process = NIL_RTR0PROCESS;
711	}
712	/pSession->uTracerData = 0;/
713	pSession->hTracerCaller = NIL_RTNATIVETHREAD;
714	RTListInit(&pSession->TpProviders);
715	/pSession->cTpProviders = 0;/
716	/pSession->cTpProbesFiring = 0;/
717	RTListInit(&pSession->TpUmods);
718	/RT_ZERO(pSession->apTpLookupTable);/
719
720	VBOXDRV_SESSION_CREATE(pSession, fUser);
721	LogFlow(("Created session %p initial cookie=%#x\n", pSession, pSession->u32Cookie));
722	return VINF_SUCCESS;
723	}
724
725	RTSpinlockDestroy(pSession->Spinlock);
726	}
727	RTMemFree(pSession);
728	*ppSession = NULL;
729	Log(("Failed to create spinlock, rc=%d!\n", rc));
730	}
731	else
732	rc = VERR_NO_MEMORY;
733
734	return rc;
735	}
736
737
738	/**
739	* Shared code for cleaning up a session.
740	*
741	* @param pDevExt Device extension.
742	* @param pSession Session data.
743	* This data will be freed by this routine.
744	*/
745	void VBOXCALL supdrvCloseSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
746	{
747	VBOXDRV_SESSION_CLOSE(pSession);
748
749	/*
750	* Cleanup the session first.
751	*/
752	supdrvCleanupSession(pDevExt, pSession);
753
754	/*
755	* Free the rest of the session stuff.
756	*/
757	RTSpinlockDestroy(pSession->Spinlock);
758	pSession->Spinlock = NIL_RTSPINLOCK;
759	pSession->pDevExt = NULL;
760	RTMemFree(pSession);
761	LogFlow(("supdrvCloseSession: returns\n"));
762	}
763
764
765	/**
766	* Shared code for cleaning up a session (but not quite freeing it).
767	*
768	* This is primarily intended for MAC OS X where we have to clean up the memory
769	* stuff before the file handle is closed.
770	*
771	* @param pDevExt Device extension.
772	* @param pSession Session data.
773	* This data will be freed by this routine.
774	*/
775	void VBOXCALL supdrvCleanupSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
776	{
777	int rc;
778	PSUPDRVBUNDLE pBundle;
779	LogFlow(("supdrvCleanupSession: pSession=%p\n", pSession));
780
781	/*
782	* Remove logger instances related to this session.
783	*/
784	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pSession);
785
786	/*
787	* Destroy the handle table.
788	*/
789	rc = RTHandleTableDestroy(pSession->hHandleTable, supdrvSessionObjHandleDelete, pSession);
790	AssertRC(rc);
791	pSession->hHandleTable = NIL_RTHANDLETABLE;
792
793	/*
794	* Release object references made in this session.
795	* In theory there should be noone racing us in this session.
796	*/
797	Log2(("release objects - start\n"));
798	if (pSession->pUsage)
799	{
800	PSUPDRVUSAGE pUsage;
801	RTSpinlockAcquire(pDevExt->Spinlock);
802
803	while ((pUsage = pSession->pUsage) != NULL)
804	{
805	PSUPDRVOBJ pObj = pUsage->pObj;
806	pSession->pUsage = pUsage->pNext;
807
808	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
809	if (pUsage->cUsage < pObj->cUsage)
810	{
811	pObj->cUsage -= pUsage->cUsage;
812	RTSpinlockRelease(pDevExt->Spinlock);
813	}
814	else
815	{
816	/* Destroy the object and free the record. */
817	if (pDevExt->pObjs == pObj)
818	pDevExt->pObjs = pObj->pNext;
819	else
820	{
821	PSUPDRVOBJ pObjPrev;
822	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
823	if (pObjPrev->pNext == pObj)
824	{
825	pObjPrev->pNext = pObj->pNext;
826	break;
827	}
828	Assert(pObjPrev);
829	}
830	RTSpinlockRelease(pDevExt->Spinlock);
831
832	Log(("supdrvCleanupSession: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
833	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
834	if (pObj->pfnDestructor)
835	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
836	RTMemFree(pObj);
837	}
838
839	/* free it and continue. */
840	RTMemFree(pUsage);
841
842	RTSpinlockAcquire(pDevExt->Spinlock);
843	}
844
845	RTSpinlockRelease(pDevExt->Spinlock);
846	AssertMsg(!pSession->pUsage, ("Some buster reregistered an object during desturction!\n"));
847	}
848	Log2(("release objects - done\n"));
849
850	/*
851	* Do tracer cleanups related to this session.
852	*/
853	Log2(("release tracer stuff - start\n"));
854	supdrvTracerCleanupSession(pDevExt, pSession);
855	Log2(("release tracer stuff - end\n"));
856
857	/*
858	* Release memory allocated in the session.
859	*
860	* We do not serialize this as we assume that the application will
861	* not allocated memory while closing the file handle object.
862	*/
863	Log2(("freeing memory:\n"));
864	pBundle = &pSession->Bundle;
865	while (pBundle)
866	{
867	PSUPDRVBUNDLE pToFree;
868	unsigned i;
869
870	/*
871	* Check and unlock all entries in the bundle.
872	*/
873	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
874	{
875	if (pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ)
876	{
877	Log2(("eType=%d pvR0=%p pvR3=%p cb=%ld\n", pBundle->aMem[i].eType, RTR0MemObjAddress(pBundle->aMem[i].MemObj),
878	(void *)RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3), (long)RTR0MemObjSize(pBundle->aMem[i].MemObj)));
879	if (pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ)
880	{
881	rc = RTR0MemObjFree(pBundle->aMem[i].MapObjR3, false);
882	AssertRC(rc); /** @todo figure out how to handle this. */
883	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
884	}
885	rc = RTR0MemObjFree(pBundle->aMem[i].MemObj, true /* fFreeMappings */);
886	AssertRC(rc); /** @todo figure out how to handle this. */
887	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
888	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
889	}
890	}
891
892	/*
893	* Advance and free previous bundle.
894	*/
895	pToFree = pBundle;
896	pBundle = pBundle->pNext;
897
898	pToFree->pNext = NULL;
899	pToFree->cUsed = 0;
900	if (pToFree != &pSession->Bundle)
901	RTMemFree(pToFree);
902	}
903	Log2(("freeing memory - done\n"));
904
905	/*
906	* Deregister component factories.
907	*/
908	RTSemFastMutexRequest(pDevExt->mtxComponentFactory);
909	Log2(("deregistering component factories:\n"));
910	if (pDevExt->pComponentFactoryHead)
911	{
912	PSUPDRVFACTORYREG pPrev = NULL;
913	PSUPDRVFACTORYREG pCur = pDevExt->pComponentFactoryHead;
914	while (pCur)
915	{
916	if (pCur->pSession == pSession)
917	{
918	/* unlink it */
919	PSUPDRVFACTORYREG pNext = pCur->pNext;
920	if (pPrev)
921	pPrev->pNext = pNext;
922	else
923	pDevExt->pComponentFactoryHead = pNext;
924
925	/* free it */
926	pCur->pNext = NULL;
927	pCur->pSession = NULL;
928	pCur->pFactory = NULL;
929	RTMemFree(pCur);
930
931	/* next */
932	pCur = pNext;
933	}
934	else
935	{
936	/* next */
937	pPrev = pCur;
938	pCur = pCur->pNext;
939	}
940	}
941	}
942	RTSemFastMutexRelease(pDevExt->mtxComponentFactory);
943	Log2(("deregistering component factories - done\n"));
944
945	/*
946	* Loaded images needs to be dereferenced and possibly freed up.
947	*/
948	supdrvLdrLock(pDevExt);
949	Log2(("freeing images:\n"));
950	if (pSession->pLdrUsage)
951	{
952	PSUPDRVLDRUSAGE pUsage = pSession->pLdrUsage;
953	pSession->pLdrUsage = NULL;
954	while (pUsage)
955	{
956	void *pvFree = pUsage;
957	PSUPDRVLDRIMAGE pImage = pUsage->pImage;
958	if (pImage->cUsage > pUsage->cUsage)
959	pImage->cUsage -= pUsage->cUsage;
960	else
961	supdrvLdrFree(pDevExt, pImage);
962	pUsage->pImage = NULL;
963	pUsage = pUsage->pNext;
964	RTMemFree(pvFree);
965	}
966	}
967	supdrvLdrUnlock(pDevExt);
968	Log2(("freeing images - done\n"));
969
970	/*
971	* Unmap the GIP.
972	*/
973	Log2(("umapping GIP:\n"));
974	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
975	{
976	SUPR0GipUnmap(pSession);
977	pSession->fGipReferenced = 0;
978	}
979	Log2(("umapping GIP - done\n"));
980	}
981
982
983	/**
984	* RTHandleTableDestroy callback used by supdrvCleanupSession.
985	*
986	* @returns IPRT status code, see SUPR0ObjAddRef.
987	* @param hHandleTable The handle table handle. Ignored.
988	* @param pvObj The object pointer.
989	* @param pvCtx Context, the handle type. Ignored.
990	* @param pvUser Session pointer.
991	*/
992	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser)
993	{
994	NOREF(pvCtx);
995	NOREF(hHandleTable);
996	return SUPR0ObjAddRefEx(pvObj, (PSUPDRVSESSION)pvUser, true /fNoBlocking/);
997	}
998
999
1000	/**
1001	* RTHandleTableDestroy callback used by supdrvCleanupSession.
1002	*
1003	* @param hHandleTable The handle table handle. Ignored.
1004	* @param h The handle value. Ignored.
1005	* @param pvObj The object pointer.
1006	* @param pvCtx Context, the handle type. Ignored.
1007	* @param pvUser Session pointer.
1008	*/
1009	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser)
1010	{
1011	NOREF(pvCtx);
1012	NOREF(h);
1013	NOREF(hHandleTable);
1014	SUPR0ObjRelease(pvObj, (PSUPDRVSESSION)pvUser);
1015	}
1016
1017
1018	/**
1019	* Fast path I/O Control worker.
1020	*
1021	* @returns VBox status code that should be passed down to ring-3 unchanged.
1022	* @param uIOCtl Function number.
1023	* @param idCpu VMCPU id.
1024	* @param pDevExt Device extention.
1025	* @param pSession Session data.
1026	*/
1027	int VBOXCALL supdrvIOCtlFast(uintptr_t uIOCtl, VMCPUID idCpu, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
1028	{
1029	/*
1030	* We check the two prereqs after doing this only to allow the compiler to optimize things better.
1031	*/
1032	if (RT_LIKELY( RT_VALID_PTR(pSession)
1033	&& pSession->pVM
1034	&& pDevExt->pfnVMMR0EntryFast))
1035	{
1036	switch (uIOCtl)
1037	{
1038	case SUP_IOCTL_FAST_DO_RAW_RUN:
1039	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_RAW_RUN);
1040	break;
1041	case SUP_IOCTL_FAST_DO_HM_RUN:
1042	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_HM_RUN);
1043	break;
1044	case SUP_IOCTL_FAST_DO_NOP:
1045	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_NOP);
1046	break;
1047	default:
1048	return VERR_INTERNAL_ERROR;
1049	}
1050	return VINF_SUCCESS;
1051	}
1052	return VERR_INTERNAL_ERROR;
1053	}
1054
1055
1056	/**
1057	* Helper for supdrvIOCtl. Check if pszStr contains any character of pszChars.
1058	* We would use strpbrk here if this function would be contained in the RedHat kABI white
1059	* list, see http://www.kerneldrivers.org/RHEL5.
1060	*
1061	* @returns 1 if pszStr does contain any character of pszChars, 0 otherwise.
1062	* @param pszStr String to check
1063	* @param pszChars Character set
1064	*/
1065	static int supdrvCheckInvalidChar(const char pszStr, const char pszChars)
1066	{
1067	int chCur;
1068	while ((chCur = *pszStr++) != '\0')
1069	{
1070	int ch;
1071	const char *psz = pszChars;
1072	while ((ch = *psz++) != '\0')
1073	if (ch == chCur)
1074	return 1;
1075
1076	}
1077	return 0;
1078	}
1079
1080
1081
1082	/**
1083	* I/O Control inner worker (tracing reasons).
1084	*
1085	* @returns IPRT status code.
1086	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1087	*
1088	* @param uIOCtl Function number.
1089	* @param pDevExt Device extention.
1090	* @param pSession Session data.
1091	* @param pReqHdr The request header.
1092	*/
1093	static int supdrvIOCtlInnerUnrestricted(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1094	{
1095	/*
1096	* Validation macros
1097	*/
1098	#define REQ_CHECK_SIZES_EX(Name, cbInExpect, cbOutExpect) \
1099	do { \
1100	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect) \|\| pReqHdr->cbOut != (cbOutExpect))) \
1101	{ \
1102	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld. cbOut=%ld expected %ld.\n", \
1103	(long)pReqHdr->cbIn, (long)(cbInExpect), (long)pReqHdr->cbOut, (long)(cbOutExpect))); \
1104	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1105	} \
1106	} while (0)
1107
1108	#define REQ_CHECK_SIZES(Name) REQ_CHECK_SIZES_EX(Name, Name ## _SIZE_IN, Name ## _SIZE_OUT)
1109
1110	#define REQ_CHECK_SIZE_IN(Name, cbInExpect) \
1111	do { \
1112	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect))) \
1113	{ \
1114	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld.\n", \
1115	(long)pReqHdr->cbIn, (long)(cbInExpect))); \
1116	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1117	} \
1118	} while (0)
1119
1120	#define REQ_CHECK_SIZE_OUT(Name, cbOutExpect) \
1121	do { \
1122	if (RT_UNLIKELY(pReqHdr->cbOut != (cbOutExpect))) \
1123	{ \
1124	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbOut=%ld expected %ld.\n", \
1125	(long)pReqHdr->cbOut, (long)(cbOutExpect))); \
1126	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1127	} \
1128	} while (0)
1129
1130	#define REQ_CHECK_EXPR(Name, expr) \
1131	do { \
1132	if (RT_UNLIKELY(!(expr))) \
1133	{ \
1134	OSDBGPRINT(( #Name ": %s\n", #expr)); \
1135	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1136	} \
1137	} while (0)
1138
1139	#define REQ_CHECK_EXPR_FMT(expr, fmt) \
1140	do { \
1141	if (RT_UNLIKELY(!(expr))) \
1142	{ \
1143	OSDBGPRINT( fmt ); \
1144	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1145	} \
1146	} while (0)
1147
1148	/*
1149	* The switch.
1150	*/
1151	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
1152	{
1153	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
1154	{
1155	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
1156	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
1157	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
1158	{
1159	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
1160	pReq->Hdr.rc = VERR_INVALID_MAGIC;
1161	return 0;
1162	}
1163
1164	#if 0
1165	/*
1166	* Call out to the OS specific code and let it do permission checks on the
1167	* client process.
1168	*/
1169	if (!supdrvOSValidateClientProcess(pDevExt, pSession))
1170	{
1171	pReq->u.Out.u32Cookie = 0xffffffff;
1172	pReq->u.Out.u32SessionCookie = 0xffffffff;
1173	pReq->u.Out.u32SessionVersion = 0xffffffff;
1174	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1175	pReq->u.Out.pSession = NULL;
1176	pReq->u.Out.cFunctions = 0;
1177	pReq->Hdr.rc = VERR_PERMISSION_DENIED;
1178	return 0;
1179	}
1180	#endif
1181
1182	/*
1183	* Match the version.
1184	* The current logic is very simple, match the major interface version.
1185	*/
1186	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
1187	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
1188	{
1189	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1190	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
1191	pReq->u.Out.u32Cookie = 0xffffffff;
1192	pReq->u.Out.u32SessionCookie = 0xffffffff;
1193	pReq->u.Out.u32SessionVersion = 0xffffffff;
1194	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1195	pReq->u.Out.pSession = NULL;
1196	pReq->u.Out.cFunctions = 0;
1197	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1198	return 0;
1199	}
1200
1201	/*
1202	* Fill in return data and be gone.
1203	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
1204	* u32SessionVersion <= u32ReqVersion!
1205	*/
1206	/** @todo Somehow validate the client and negotiate a secure cookie... */
1207	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
1208	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
1209	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
1210	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1211	pReq->u.Out.pSession = pSession;
1212	pReq->u.Out.cFunctions = sizeof(g_aFunctions) / sizeof(g_aFunctions[0]);
1213	pReq->Hdr.rc = VINF_SUCCESS;
1214	return 0;
1215	}
1216
1217	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_QUERY_FUNCS(0)):
1218	{
1219	/* validate */
1220	PSUPQUERYFUNCS pReq = (PSUPQUERYFUNCS)pReqHdr;
1221	REQ_CHECK_SIZES_EX(SUP_IOCTL_QUERY_FUNCS, SUP_IOCTL_QUERY_FUNCS_SIZE_IN, SUP_IOCTL_QUERY_FUNCS_SIZE_OUT(RT_ELEMENTS(g_aFunctions)));
1222
1223	/* execute */
1224	pReq->u.Out.cFunctions = RT_ELEMENTS(g_aFunctions);
1225	memcpy(&pReq->u.Out.aFunctions[0], g_aFunctions, sizeof(g_aFunctions));
1226	pReq->Hdr.rc = VINF_SUCCESS;
1227	return 0;
1228	}
1229
1230	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_LOCK):
1231	{
1232	/* validate */
1233	PSUPPAGELOCK pReq = (PSUPPAGELOCK)pReqHdr;
1234	REQ_CHECK_SIZE_IN(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_IN);
1235	REQ_CHECK_SIZE_OUT(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_OUT(pReq->u.In.cPages));
1236	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.cPages > 0);
1237	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.pvR3 >= PAGE_SIZE);
1238
1239	/* execute */
1240	pReq->Hdr.rc = SUPR0LockMem(pSession, pReq->u.In.pvR3, pReq->u.In.cPages, &pReq->u.Out.aPages[0]);
1241	if (RT_FAILURE(pReq->Hdr.rc))
1242	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1243	return 0;
1244	}
1245
1246	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_UNLOCK):
1247	{
1248	/* validate */
1249	PSUPPAGEUNLOCK pReq = (PSUPPAGEUNLOCK)pReqHdr;
1250	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_UNLOCK);
1251
1252	/* execute */
1253	pReq->Hdr.rc = SUPR0UnlockMem(pSession, pReq->u.In.pvR3);
1254	return 0;
1255	}
1256
1257	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_ALLOC):
1258	{
1259	/* validate */
1260	PSUPCONTALLOC pReq = (PSUPCONTALLOC)pReqHdr;
1261	REQ_CHECK_SIZES(SUP_IOCTL_CONT_ALLOC);
1262
1263	/* execute */
1264	pReq->Hdr.rc = SUPR0ContAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.HCPhys);
1265	if (RT_FAILURE(pReq->Hdr.rc))
1266	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1267	return 0;
1268	}
1269
1270	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_FREE):
1271	{
1272	/* validate */
1273	PSUPCONTFREE pReq = (PSUPCONTFREE)pReqHdr;
1274	REQ_CHECK_SIZES(SUP_IOCTL_CONT_FREE);
1275
1276	/* execute */
1277	pReq->Hdr.rc = SUPR0ContFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1278	return 0;
1279	}
1280
1281	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_OPEN):
1282	{
1283	/* validate */
1284	PSUPLDROPEN pReq = (PSUPLDROPEN)pReqHdr;
1285	REQ_CHECK_SIZES(SUP_IOCTL_LDR_OPEN);
1286	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs > 0);
1287	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs < 16*_1M);
1288	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1289	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1290	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits < pReq->u.In.cbImageWithTabs);
1291	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.szName[0]);
1292	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1293	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, !supdrvCheckInvalidChar(pReq->u.In.szName, ";:()[]{}/\\\|&*%#@!~`\"'"));
1294	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szFilename, sizeof(pReq->u.In.szFilename)));
1295
1296	/* execute */
1297	pReq->Hdr.rc = supdrvIOCtl_LdrOpen(pDevExt, pSession, pReq);
1298	return 0;
1299	}
1300
1301	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOAD):
1302	{
1303	/* validate */
1304	PSUPLDRLOAD pReq = (PSUPLDRLOAD)pReqHdr;
1305	REQ_CHECK_EXPR(Name, pReq->Hdr.cbIn >= sizeof(*pReq));
1306	REQ_CHECK_SIZES_EX(SUP_IOCTL_LDR_LOAD, SUP_IOCTL_LDR_LOAD_SIZE_IN(pReq->u.In.cbImageWithTabs), SUP_IOCTL_LDR_LOAD_SIZE_OUT);
1307	REQ_CHECK_EXPR(SUP_IOCTL_LDR_LOAD, pReq->u.In.cSymbols <= 16384);
1308	REQ_CHECK_EXPR_FMT( !pReq->u.In.cSymbols
1309	\|\| ( pReq->u.In.offSymbols < pReq->u.In.cbImageWithTabs
1310	&& pReq->u.In.offSymbols + pReq->u.In.cSymbols * sizeof(SUPLDRSYM) <= pReq->u.In.cbImageWithTabs),
1311	("SUP_IOCTL_LDR_LOAD: offSymbols=%#lx cSymbols=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offSymbols,
1312	(long)pReq->u.In.cSymbols, (long)pReq->u.In.cbImageWithTabs));
1313	REQ_CHECK_EXPR_FMT( !pReq->u.In.cbStrTab
1314	\|\| ( pReq->u.In.offStrTab < pReq->u.In.cbImageWithTabs
1315	&& pReq->u.In.offStrTab + pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs
1316	&& pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs),
1317	("SUP_IOCTL_LDR_LOAD: offStrTab=%#lx cbStrTab=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offStrTab,
1318	(long)pReq->u.In.cbStrTab, (long)pReq->u.In.cbImageWithTabs));
1319
1320	if (pReq->u.In.cSymbols)
1321	{
1322	uint32_t i;
1323	PSUPLDRSYM paSyms = (PSUPLDRSYM)&pReq->u.In.abImage[pReq->u.In.offSymbols];
1324	for (i = 0; i < pReq->u.In.cSymbols; i++)
1325	{
1326	REQ_CHECK_EXPR_FMT(paSyms[i].offSymbol < pReq->u.In.cbImageWithTabs,
1327	("SUP_IOCTL_LDR_LOAD: sym #%ld: symb off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offSymbol, (long)pReq->u.In.cbImageWithTabs));
1328	REQ_CHECK_EXPR_FMT(paSyms[i].offName < pReq->u.In.cbStrTab,
1329	("SUP_IOCTL_LDR_LOAD: sym #%ld: name off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1330	REQ_CHECK_EXPR_FMT(RTStrEnd((char const *)&pReq->u.In.abImage[pReq->u.In.offStrTab + paSyms[i].offName],
1331	pReq->u.In.cbStrTab - paSyms[i].offName),
1332	("SUP_IOCTL_LDR_LOAD: sym #%ld: unterminated name! (%#lx / %#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1333	}
1334	}
1335
1336	/* execute */
1337	pReq->Hdr.rc = supdrvIOCtl_LdrLoad(pDevExt, pSession, pReq);
1338	return 0;
1339	}
1340
1341	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_FREE):
1342	{
1343	/* validate */
1344	PSUPLDRFREE pReq = (PSUPLDRFREE)pReqHdr;
1345	REQ_CHECK_SIZES(SUP_IOCTL_LDR_FREE);
1346
1347	/* execute */
1348	pReq->Hdr.rc = supdrvIOCtl_LdrFree(pDevExt, pSession, pReq);
1349	return 0;
1350	}
1351
1352	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_GET_SYMBOL):
1353	{
1354	/* validate */
1355	PSUPLDRGETSYMBOL pReq = (PSUPLDRGETSYMBOL)pReqHdr;
1356	REQ_CHECK_SIZES(SUP_IOCTL_LDR_GET_SYMBOL);
1357	REQ_CHECK_EXPR(SUP_IOCTL_LDR_GET_SYMBOL, RTStrEnd(pReq->u.In.szSymbol, sizeof(pReq->u.In.szSymbol)));
1358
1359	/* execute */
1360	pReq->Hdr.rc = supdrvIOCtl_LdrGetSymbol(pDevExt, pSession, pReq);
1361	return 0;
1362	}
1363
1364	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0(0)):
1365	{
1366	/* validate */
1367	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1368	Log4(("SUP_IOCTL_CALL_VMMR0: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1369	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1370
1371	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_VMMR0_SIZE(0))
1372	{
1373	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(0), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(0));
1374
1375	/* execute */
1376	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1377	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg, pSession);
1378	else
1379	pReq->Hdr.rc = VERR_WRONG_ORDER;
1380	}
1381	else
1382	{
1383	PSUPVMMR0REQHDR pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1384	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR)),
1385	("SUP_IOCTL_CALL_VMMR0: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR))));
1386	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1387	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(pVMMReq->cbReq));
1388
1389	/* execute */
1390	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1391	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1392	else
1393	pReq->Hdr.rc = VERR_WRONG_ORDER;
1394	}
1395
1396	if ( RT_FAILURE(pReq->Hdr.rc)
1397	&& pReq->Hdr.rc != VERR_INTERRUPTED
1398	&& pReq->Hdr.rc != VERR_TIMEOUT)
1399	Log(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1400	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1401	else
1402	Log4(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1403	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1404	return 0;
1405	}
1406
1407	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0_BIG):
1408	{
1409	/* validate */
1410	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1411	PSUPVMMR0REQHDR pVMMReq;
1412	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1413	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1414
1415	pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1416	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR)),
1417	("SUP_IOCTL_CALL_VMMR0_BIG: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR))));
1418	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0_BIG, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1419	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));
1420
1421	/* execute */
1422	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1423	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1424	else
1425	pReq->Hdr.rc = VERR_WRONG_ORDER;
1426
1427	if ( RT_FAILURE(pReq->Hdr.rc)
1428	&& pReq->Hdr.rc != VERR_INTERRUPTED
1429	&& pReq->Hdr.rc != VERR_TIMEOUT)
1430	Log(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1431	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1432	else
1433	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1434	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1435	return 0;
1436	}
1437
1438	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_PAGING_MODE):
1439	{
1440	/* validate */
1441	PSUPGETPAGINGMODE pReq = (PSUPGETPAGINGMODE)pReqHdr;
1442	REQ_CHECK_SIZES(SUP_IOCTL_GET_PAGING_MODE);
1443
1444	/* execute */
1445	pReq->Hdr.rc = VINF_SUCCESS;
1446	pReq->u.Out.enmMode = SUPR0GetPagingMode();
1447	return 0;
1448	}
1449
1450	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_ALLOC):
1451	{
1452	/* validate */
1453	PSUPLOWALLOC pReq = (PSUPLOWALLOC)pReqHdr;
1454	REQ_CHECK_EXPR(SUP_IOCTL_LOW_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_LOW_ALLOC_SIZE_IN);
1455	REQ_CHECK_SIZES_EX(SUP_IOCTL_LOW_ALLOC, SUP_IOCTL_LOW_ALLOC_SIZE_IN, SUP_IOCTL_LOW_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1456
1457	/* execute */
1458	pReq->Hdr.rc = SUPR0LowAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1459	if (RT_FAILURE(pReq->Hdr.rc))
1460	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1461	return 0;
1462	}
1463
1464	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_FREE):
1465	{
1466	/* validate */
1467	PSUPLOWFREE pReq = (PSUPLOWFREE)pReqHdr;
1468	REQ_CHECK_SIZES(SUP_IOCTL_LOW_FREE);
1469
1470	/* execute */
1471	pReq->Hdr.rc = SUPR0LowFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1472	return 0;
1473	}
1474
1475	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_MAP):
1476	{
1477	/* validate */
1478	PSUPGIPMAP pReq = (PSUPGIPMAP)pReqHdr;
1479	REQ_CHECK_SIZES(SUP_IOCTL_GIP_MAP);
1480
1481	/* execute */
1482	pReq->Hdr.rc = SUPR0GipMap(pSession, &pReq->u.Out.pGipR3, &pReq->u.Out.HCPhysGip);
1483	if (RT_SUCCESS(pReq->Hdr.rc))
1484	pReq->u.Out.pGipR0 = pDevExt->pGip;
1485	return 0;
1486	}
1487
1488	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_UNMAP):
1489	{
1490	/* validate */
1491	PSUPGIPUNMAP pReq = (PSUPGIPUNMAP)pReqHdr;
1492	REQ_CHECK_SIZES(SUP_IOCTL_GIP_UNMAP);
1493
1494	/* execute */
1495	pReq->Hdr.rc = SUPR0GipUnmap(pSession);
1496	return 0;
1497	}
1498
1499	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SET_VM_FOR_FAST):
1500	{
1501	/* validate */
1502	PSUPSETVMFORFAST pReq = (PSUPSETVMFORFAST)pReqHdr;
1503	REQ_CHECK_SIZES(SUP_IOCTL_SET_VM_FOR_FAST);
1504	REQ_CHECK_EXPR_FMT( !pReq->u.In.pVMR0
1505	\|\| ( VALID_PTR(pReq->u.In.pVMR0)
1506	&& !((uintptr_t)pReq->u.In.pVMR0 & (PAGE_SIZE - 1))),
1507	("SUP_IOCTL_SET_VM_FOR_FAST: pVMR0=%p!\n", pReq->u.In.pVMR0));
1508	/* execute */
1509	pSession->pVM = pReq->u.In.pVMR0;
1510	pReq->Hdr.rc = VINF_SUCCESS;
1511	return 0;
1512	}
1513
1514	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_ALLOC_EX):
1515	{
1516	/* validate */
1517	PSUPPAGEALLOCEX pReq = (PSUPPAGEALLOCEX)pReqHdr;
1518	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC_EX, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_EX_SIZE_IN);
1519	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));
1520	REQ_CHECK_EXPR_FMT(pReq->u.In.fKernelMapping \|\| pReq->u.In.fUserMapping,
1521	("SUP_IOCTL_PAGE_ALLOC_EX: No mapping requested!\n"));
1522	REQ_CHECK_EXPR_FMT(pReq->u.In.fUserMapping,
1523	("SUP_IOCTL_PAGE_ALLOC_EX: Must have user mapping!\n"));
1524	REQ_CHECK_EXPR_FMT(!pReq->u.In.fReserved0 && !pReq->u.In.fReserved1,
1525	("SUP_IOCTL_PAGE_ALLOC_EX: fReserved0=%d fReserved1=%d\n", pReq->u.In.fReserved0, pReq->u.In.fReserved1));
1526
1527	/* execute */
1528	pReq->Hdr.rc = SUPR0PageAllocEx(pSession, pReq->u.In.cPages, 0 /* fFlags */,
1529	pReq->u.In.fUserMapping ? &pReq->u.Out.pvR3 : NULL,
1530	pReq->u.In.fKernelMapping ? &pReq->u.Out.pvR0 : NULL,
1531	&pReq->u.Out.aPages[0]);
1532	if (RT_FAILURE(pReq->Hdr.rc))
1533	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1534	return 0;
1535	}
1536
1537	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_MAP_KERNEL):
1538	{
1539	/* validate */
1540	PSUPPAGEMAPKERNEL pReq = (PSUPPAGEMAPKERNEL)pReqHdr;
1541	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_MAP_KERNEL);
1542	REQ_CHECK_EXPR_FMT(!pReq->u.In.fFlags, ("SUP_IOCTL_PAGE_MAP_KERNEL: fFlags=%#x! MBZ\n", pReq->u.In.fFlags));
1543	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_MAP_KERNEL: offSub=%#x\n", pReq->u.In.offSub));
1544	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1545	("SUP_IOCTL_PAGE_MAP_KERNEL: cbSub=%#x\n", pReq->u.In.cbSub));
1546
1547	/* execute */
1548	pReq->Hdr.rc = SUPR0PageMapKernel(pSession, pReq->u.In.pvR3, pReq->u.In.offSub, pReq->u.In.cbSub,
1549	pReq->u.In.fFlags, &pReq->u.Out.pvR0);
1550	if (RT_FAILURE(pReq->Hdr.rc))
1551	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1552	return 0;
1553	}
1554
1555	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_PROTECT):
1556	{
1557	/* validate */
1558	PSUPPAGEPROTECT pReq = (PSUPPAGEPROTECT)pReqHdr;
1559	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_PROTECT);
1560	REQ_CHECK_EXPR_FMT(!(pReq->u.In.fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)),
1561	("SUP_IOCTL_PAGE_PROTECT: fProt=%#x!\n", pReq->u.In.fProt));
1562	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_PROTECT: offSub=%#x\n", pReq->u.In.offSub));
1563	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1564	("SUP_IOCTL_PAGE_PROTECT: cbSub=%#x\n", pReq->u.In.cbSub));
1565
1566	/* execute */
1567	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);
1568	return 0;
1569	}
1570
1571	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_FREE):
1572	{
1573	/* validate */
1574	PSUPPAGEFREE pReq = (PSUPPAGEFREE)pReqHdr;
1575	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_FREE);
1576
1577	/* execute */
1578	pReq->Hdr.rc = SUPR0PageFree(pSession, pReq->u.In.pvR3);
1579	return 0;
1580	}
1581
1582	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_SERVICE(0)):
1583	{
1584	/* validate */
1585	PSUPCALLSERVICE pReq = (PSUPCALLSERVICE)pReqHdr;
1586	Log4(("SUP_IOCTL_CALL_SERVICE: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1587	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1588
1589	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
1590	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(0), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(0));
1591	else
1592	{
1593	PSUPR0SERVICEREQHDR pSrvReq = (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0];
1594	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR)),
1595	("SUP_IOCTL_CALL_SERVICE: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR))));
1596	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, pSrvReq->u32Magic == SUPR0SERVICEREQHDR_MAGIC);
1597	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(pSrvReq->cbReq), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(pSrvReq->cbReq));
1598	}
1599	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1600
1601	/* execute */
1602	pReq->Hdr.rc = supdrvIOCtl_CallServiceModule(pDevExt, pSession, pReq);
1603	return 0;
1604	}
1605
1606	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOGGER_SETTINGS(0)):
1607	{
1608	/* validate */
1609	PSUPLOGGERSETTINGS pReq = (PSUPLOGGERSETTINGS)pReqHdr;
1610	size_t cbStrTab;
1611	REQ_CHECK_SIZE_OUT(SUP_IOCTL_LOGGER_SETTINGS, SUP_IOCTL_LOGGER_SETTINGS_SIZE_OUT);
1612	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->Hdr.cbIn >= SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(1));
1613	cbStrTab = pReq->Hdr.cbIn - SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(0);
1614	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offGroups < cbStrTab);
1615	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offFlags < cbStrTab);
1616	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offDestination < cbStrTab);
1617	REQ_CHECK_EXPR_FMT(pReq->u.In.szStrings[cbStrTab - 1] == '\0',
1618	("SUP_IOCTL_LOGGER_SETTINGS: cbIn=%#x cbStrTab=%#zx LastChar=%d\n",
1619	pReq->Hdr.cbIn, cbStrTab, pReq->u.In.szStrings[cbStrTab - 1]));
1620	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhich <= SUPLOGGERSETTINGS_WHICH_RELEASE);
1621	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhat <= SUPLOGGERSETTINGS_WHAT_DESTROY);
1622
1623	/* execute */
1624	pReq->Hdr.rc = supdrvIOCtl_LoggerSettings(pDevExt, pSession, pReq);
1625	return 0;
1626	}
1627
1628	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP2):
1629	{
1630	/* validate */
1631	PSUPSEMOP2 pReq = (PSUPSEMOP2)pReqHdr;
1632	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP2, SUP_IOCTL_SEM_OP2_SIZE_IN, SUP_IOCTL_SEM_OP2_SIZE_OUT);
1633	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP2, pReq->u.In.uReserved == 0);
1634
1635	/* execute */
1636	switch (pReq->u.In.uType)
1637	{
1638	case SUP_SEM_TYPE_EVENT:
1639	{
1640	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1641	switch (pReq->u.In.uOp)
1642	{
1643	case SUPSEMOP2_WAIT_MS_REL:
1644	pReq->Hdr.rc = SUPSemEventWaitNoResume(pSession, hEvent, pReq->u.In.uArg.cRelMsTimeout);
1645	break;
1646	case SUPSEMOP2_WAIT_NS_ABS:
1647	pReq->Hdr.rc = SUPSemEventWaitNsAbsIntr(pSession, hEvent, pReq->u.In.uArg.uAbsNsTimeout);
1648	break;
1649	case SUPSEMOP2_WAIT_NS_REL:
1650	pReq->Hdr.rc = SUPSemEventWaitNsRelIntr(pSession, hEvent, pReq->u.In.uArg.cRelNsTimeout);
1651	break;
1652	case SUPSEMOP2_SIGNAL:
1653	pReq->Hdr.rc = SUPSemEventSignal(pSession, hEvent);
1654	break;
1655	case SUPSEMOP2_CLOSE:
1656	pReq->Hdr.rc = SUPSemEventClose(pSession, hEvent);
1657	break;
1658	case SUPSEMOP2_RESET:
1659	default:
1660	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1661	break;
1662	}
1663	break;
1664	}
1665
1666	case SUP_SEM_TYPE_EVENT_MULTI:
1667	{
1668	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1669	switch (pReq->u.In.uOp)
1670	{
1671	case SUPSEMOP2_WAIT_MS_REL:
1672	pReq->Hdr.rc = SUPSemEventMultiWaitNoResume(pSession, hEventMulti, pReq->u.In.uArg.cRelMsTimeout);
1673	break;
1674	case SUPSEMOP2_WAIT_NS_ABS:
1675	pReq->Hdr.rc = SUPSemEventMultiWaitNsAbsIntr(pSession, hEventMulti, pReq->u.In.uArg.uAbsNsTimeout);
1676	break;
1677	case SUPSEMOP2_WAIT_NS_REL:
1678	pReq->Hdr.rc = SUPSemEventMultiWaitNsRelIntr(pSession, hEventMulti, pReq->u.In.uArg.cRelNsTimeout);
1679	break;
1680	case SUPSEMOP2_SIGNAL:
1681	pReq->Hdr.rc = SUPSemEventMultiSignal(pSession, hEventMulti);
1682	break;
1683	case SUPSEMOP2_CLOSE:
1684	pReq->Hdr.rc = SUPSemEventMultiClose(pSession, hEventMulti);
1685	break;
1686	case SUPSEMOP2_RESET:
1687	pReq->Hdr.rc = SUPSemEventMultiReset(pSession, hEventMulti);
1688	break;
1689	default:
1690	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1691	break;
1692	}
1693	break;
1694	}
1695
1696	default:
1697	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1698	break;
1699	}
1700	return 0;
1701	}
1702
1703	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP3):
1704	{
1705	/* validate */
1706	PSUPSEMOP3 pReq = (PSUPSEMOP3)pReqHdr;
1707	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP3, SUP_IOCTL_SEM_OP3_SIZE_IN, SUP_IOCTL_SEM_OP3_SIZE_OUT);
1708	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, pReq->u.In.u32Reserved == 0 && pReq->u.In.u64Reserved == 0);
1709
1710	/* execute */
1711	switch (pReq->u.In.uType)
1712	{
1713	case SUP_SEM_TYPE_EVENT:
1714	{
1715	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1716	switch (pReq->u.In.uOp)
1717	{
1718	case SUPSEMOP3_CREATE:
1719	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1720	pReq->Hdr.rc = SUPSemEventCreate(pSession, &hEvent);
1721	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEvent;
1722	break;
1723	case SUPSEMOP3_GET_RESOLUTION:
1724	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1725	pReq->Hdr.rc = VINF_SUCCESS;
1726	pReq->Hdr.cbOut = sizeof(*pReq);
1727	pReq->u.Out.cNsResolution = SUPSemEventGetResolution(pSession);
1728	break;
1729	default:
1730	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1731	break;
1732	}
1733	break;
1734	}
1735
1736	case SUP_SEM_TYPE_EVENT_MULTI:
1737	{
1738	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1739	switch (pReq->u.In.uOp)
1740	{
1741	case SUPSEMOP3_CREATE:
1742	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1743	pReq->Hdr.rc = SUPSemEventMultiCreate(pSession, &hEventMulti);
1744	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEventMulti;
1745	break;
1746	case SUPSEMOP3_GET_RESOLUTION:
1747	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1748	pReq->Hdr.rc = VINF_SUCCESS;
1749	pReq->u.Out.cNsResolution = SUPSemEventMultiGetResolution(pSession);
1750	break;
1751	default:
1752	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1753	break;
1754	}
1755	break;
1756	}
1757
1758	default:
1759	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1760	break;
1761	}
1762	return 0;
1763	}
1764
1765	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
1766	{
1767	/* validate */
1768	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
1769	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
1770
1771	/* execute */
1772	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.Caps);
1773	if (RT_FAILURE(pReq->Hdr.rc))
1774	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1775	return 0;
1776	}
1777
1778	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_OPEN):
1779	{
1780	/* validate */
1781	PSUPTRACEROPEN pReq = (PSUPTRACEROPEN)pReqHdr;
1782	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_OPEN);
1783
1784	/* execute */
1785	pReq->Hdr.rc = supdrvIOCtl_TracerOpen(pDevExt, pSession, pReq->u.In.uCookie, pReq->u.In.uArg);
1786	return 0;
1787	}
1788
1789	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_CLOSE):
1790	{
1791	/* validate */
1792	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_CLOSE);
1793
1794	/* execute */
1795	pReqHdr->rc = supdrvIOCtl_TracerClose(pDevExt, pSession);
1796	return 0;
1797	}
1798
1799	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_IOCTL):
1800	{
1801	/* validate */
1802	PSUPTRACERIOCTL pReq = (PSUPTRACERIOCTL)pReqHdr;
1803	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_IOCTL);
1804
1805	/* execute */
1806	pReqHdr->rc = supdrvIOCtl_TracerIOCtl(pDevExt, pSession, pReq->u.In.uCmd, pReq->u.In.uArg, &pReq->u.Out.iRetVal);
1807	return 0;
1808	}
1809
1810	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_REG):
1811	{
1812	/* validate */
1813	PSUPTRACERUMODREG pReq = (PSUPTRACERUMODREG)pReqHdr;
1814	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_REG);
1815	if (!RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)))
1816	return VERR_INVALID_PARAMETER;
1817
1818	/* execute */
1819	pReqHdr->rc = supdrvIOCtl_TracerUmodRegister(pDevExt, pSession,
1820	pReq->u.In.R3PtrVtgHdr, pReq->u.In.uVtgHdrAddr,
1821	pReq->u.In.R3PtrStrTab, pReq->u.In.cbStrTab,
1822	pReq->u.In.szName, pReq->u.In.fFlags);
1823	return 0;
1824	}
1825
1826	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_DEREG):
1827	{
1828	/* validate */
1829	PSUPTRACERUMODDEREG pReq = (PSUPTRACERUMODDEREG)pReqHdr;
1830	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_DEREG);
1831
1832	/* execute */
1833	pReqHdr->rc = supdrvIOCtl_TracerUmodDeregister(pDevExt, pSession, pReq->u.In.pVtgHdr);
1834	return 0;
1835	}
1836
1837	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_FIRE_PROBE):
1838	{
1839	/* validate */
1840	PSUPTRACERUMODFIREPROBE pReq = (PSUPTRACERUMODFIREPROBE)pReqHdr;
1841	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_FIRE_PROBE);
1842
1843	supdrvIOCtl_TracerUmodProbeFire(pDevExt, pSession, &pReq->u.In);
1844	pReqHdr->rc = VINF_SUCCESS;
1845	return 0;
1846	}
1847
1848	default:
1849	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1850	break;
1851	}
1852	return VERR_GENERAL_FAILURE;
1853	}
1854
1855
1856	/**
1857	* I/O Control inner worker for the restricted operations.
1858	*
1859	* @returns IPRT status code.
1860	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1861	*
1862	* @param uIOCtl Function number.
1863	* @param pDevExt Device extention.
1864	* @param pSession Session data.
1865	* @param pReqHdr The request header.
1866	*/
1867	static int supdrvIOCtlInnerRestricted(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1868	{
1869	/*
1870	* The switch.
1871	*/
1872	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
1873	{
1874	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
1875	{
1876	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
1877	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
1878	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
1879	{
1880	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
1881	pReq->Hdr.rc = VERR_INVALID_MAGIC;
1882	return 0;
1883	}
1884
1885	/*
1886	* Match the version.
1887	* The current logic is very simple, match the major interface version.
1888	*/
1889	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
1890	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
1891	{
1892	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1893	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
1894	pReq->u.Out.u32Cookie = 0xffffffff;
1895	pReq->u.Out.u32SessionCookie = 0xffffffff;
1896	pReq->u.Out.u32SessionVersion = 0xffffffff;
1897	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1898	pReq->u.Out.pSession = NULL;
1899	pReq->u.Out.cFunctions = 0;
1900	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1901	return 0;
1902	}
1903
1904	/*
1905	* Fill in return data and be gone.
1906	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
1907	* u32SessionVersion <= u32ReqVersion!
1908	*/
1909	/** @todo Somehow validate the client and negotiate a secure cookie... */
1910	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
1911	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
1912	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
1913	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1914	pReq->u.Out.pSession = pSession;
1915	pReq->u.Out.cFunctions = 0;
1916	pReq->Hdr.rc = VINF_SUCCESS;
1917	return 0;
1918	}
1919
1920	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
1921	{
1922	/* validate */
1923	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
1924	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
1925
1926	/* execute */
1927	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.Caps);
1928	if (RT_FAILURE(pReq->Hdr.rc))
1929	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1930	return 0;
1931	}
1932
1933	default:
1934	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1935	break;
1936	}
1937	return VERR_GENERAL_FAILURE;
1938	}
1939
1940
1941	/**
1942	* I/O Control worker.
1943	*
1944	* @returns IPRT status code.
1945	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1946	*
1947	* @param uIOCtl Function number.
1948	* @param pDevExt Device extention.
1949	* @param pSession Session data.
1950	* @param pReqHdr The request header.
1951	*/
1952	int VBOXCALL supdrvIOCtl(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1953	{
1954	int rc;
1955	VBOXDRV_IOCTL_ENTRY(pSession, uIOCtl, pReqHdr);
1956
1957	/*
1958	* Validate the request.
1959	*/
1960	/* this first check could probably be omitted as its also done by the OS specific code... */
1961	if (RT_UNLIKELY( (pReqHdr->fFlags & SUPREQHDR_FLAGS_MAGIC_MASK) != SUPREQHDR_FLAGS_MAGIC
1962	\|\| pReqHdr->cbIn < sizeof(*pReqHdr)
1963	\|\| pReqHdr->cbOut < sizeof(*pReqHdr)))
1964	{
1965	OSDBGPRINT(("vboxdrv: Bad ioctl request header; cbIn=%#lx cbOut=%#lx fFlags=%#lx\n",
1966	(long)pReqHdr->cbIn, (long)pReqHdr->cbOut, (long)pReqHdr->fFlags));
1967	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1968	return VERR_INVALID_PARAMETER;
1969	}
1970	if (RT_UNLIKELY(!RT_VALID_PTR(pSession)))
1971	{
1972	OSDBGPRINT(("vboxdrv: Invalid pSession valud %p (ioctl=%p)\n", pSession, (void *)uIOCtl));
1973	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1974	return VERR_INVALID_PARAMETER;
1975	}
1976	if (RT_UNLIKELY(uIOCtl == SUP_IOCTL_COOKIE))
1977	{
1978	if (pReqHdr->u32Cookie != SUPCOOKIE_INITIAL_COOKIE)
1979	{
1980	OSDBGPRINT(("SUP_IOCTL_COOKIE: bad cookie %#lx\n", (long)pReqHdr->u32Cookie));
1981	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1982	return VERR_INVALID_PARAMETER;
1983	}
1984	}
1985	else if (RT_UNLIKELY( pReqHdr->u32Cookie != pDevExt->u32Cookie
1986	\|\| pReqHdr->u32SessionCookie != pSession->u32Cookie))
1987	{
1988	OSDBGPRINT(("vboxdrv: bad cookie %#lx / %#lx.\n", (long)pReqHdr->u32Cookie, (long)pReqHdr->u32SessionCookie));
1989	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1990	return VERR_INVALID_PARAMETER;
1991	}
1992
1993	/*
1994	* Hand it to an inner function to avoid lots of unnecessary return tracepoints.
1995	*/
1996	if (pSession->fUnrestricted)
1997	rc = supdrvIOCtlInnerUnrestricted(uIOCtl, pDevExt, pSession, pReqHdr);
1998	else
1999	rc = supdrvIOCtlInnerRestricted(uIOCtl, pDevExt, pSession, pReqHdr);
2000
2001	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, pReqHdr->rc, rc);
2002	return rc;
2003	}
2004
2005
2006	/**
2007	* Inter-Driver Communication (IDC) worker.
2008	*
2009	* @returns VBox status code.
2010	* @retval VINF_SUCCESS on success.
2011	* @retval VERR_INVALID_PARAMETER if the request is invalid.
2012	* @retval VERR_NOT_SUPPORTED if the request isn't supported.
2013	*
2014	* @param uReq The request (function) code.
2015	* @param pDevExt Device extention.
2016	* @param pSession Session data.
2017	* @param pReqHdr The request header.
2018	*/
2019	int VBOXCALL supdrvIDC(uintptr_t uReq, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQHDR pReqHdr)
2020	{
2021	/*
2022	* The OS specific code has already validated the pSession
2023	* pointer, and the request size being greater or equal to
2024	* size of the header.
2025	*
2026	* So, just check that pSession is a kernel context session.
2027	*/
2028	if (RT_UNLIKELY( pSession
2029	&& pSession->R0Process != NIL_RTR0PROCESS))
2030	return VERR_INVALID_PARAMETER;
2031
2032	/*
2033	* Validation macro.
2034	*/
2035	#define REQ_CHECK_IDC_SIZE(Name, cbExpect) \
2036	do { \
2037	if (RT_UNLIKELY(pReqHdr->cb != (cbExpect))) \
2038	{ \
2039	OSDBGPRINT(( #Name ": Invalid input/output sizes. cb=%ld expected %ld.\n", \
2040	(long)pReqHdr->cb, (long)(cbExpect))); \
2041	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
2042	} \
2043	} while (0)
2044
2045	switch (uReq)
2046	{
2047	case SUPDRV_IDC_REQ_CONNECT:
2048	{
2049	PSUPDRVIDCREQCONNECT pReq = (PSUPDRVIDCREQCONNECT)pReqHdr;
2050	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_CONNECT, sizeof(*pReq));
2051
2052	/*
2053	* Validate the cookie and other input.
2054	*/
2055	if (pReq->Hdr.pSession != NULL)
2056	{
2057	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Hdr.pSession=%p expected NULL!\n", pReq->Hdr.pSession));
2058	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2059	}
2060	if (pReq->u.In.u32MagicCookie != SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE)
2061	{
2062	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: u32MagicCookie=%#x expected %#x!\n",
2063	(unsigned)pReq->u.In.u32MagicCookie, (unsigned)SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE));
2064	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2065	}
2066	if ( pReq->u.In.uMinVersion > pReq->u.In.uReqVersion
2067	\|\| (pReq->u.In.uMinVersion & UINT32_C(0xffff0000)) != (pReq->u.In.uReqVersion & UINT32_C(0xffff0000)))
2068	{
2069	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: uMinVersion=%#x uMaxVersion=%#x doesn't match!\n",
2070	pReq->u.In.uMinVersion, pReq->u.In.uReqVersion));
2071	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2072	}
2073	if (pSession != NULL)
2074	{
2075	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: pSession=%p expected NULL!\n", pSession));
2076	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2077	}
2078
2079	/*
2080	* Match the version.
2081	* The current logic is very simple, match the major interface version.
2082	*/
2083	if ( pReq->u.In.uMinVersion > SUPDRV_IDC_VERSION
2084	\|\| (pReq->u.In.uMinVersion & 0xffff0000) != (SUPDRV_IDC_VERSION & 0xffff0000))
2085	{
2086	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
2087	pReq->u.In.uReqVersion, pReq->u.In.uMinVersion, (unsigned)SUPDRV_IDC_VERSION));
2088	pReq->u.Out.pSession = NULL;
2089	pReq->u.Out.uSessionVersion = 0xffffffff;
2090	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
2091	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
2092	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
2093	return VINF_SUCCESS;
2094	}
2095
2096	pReq->u.Out.pSession = NULL;
2097	pReq->u.Out.uSessionVersion = SUPDRV_IDC_VERSION;
2098	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
2099	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
2100
2101	pReq->Hdr.rc = supdrvCreateSession(pDevExt, false /* fUser /, true /fUnrestricted*/, &pSession);
2102	if (RT_FAILURE(pReq->Hdr.rc))
2103	{
2104	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: failed to create session, rc=%d\n", pReq->Hdr.rc));
2105	return VINF_SUCCESS;
2106	}
2107
2108	pReq->u.Out.pSession = pSession;
2109	pReq->Hdr.pSession = pSession;
2110
2111	return VINF_SUCCESS;
2112	}
2113
2114	case SUPDRV_IDC_REQ_DISCONNECT:
2115	{
2116	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_DISCONNECT, sizeof(*pReqHdr));
2117
2118	supdrvCloseSession(pDevExt, pSession);
2119	return pReqHdr->rc = VINF_SUCCESS;
2120	}
2121
2122	case SUPDRV_IDC_REQ_GET_SYMBOL:
2123	{
2124	PSUPDRVIDCREQGETSYM pReq = (PSUPDRVIDCREQGETSYM)pReqHdr;
2125	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_GET_SYMBOL, sizeof(*pReq));
2126
2127	pReq->Hdr.rc = supdrvIDC_LdrGetSymbol(pDevExt, pSession, pReq);
2128	return VINF_SUCCESS;
2129	}
2130
2131	case SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY:
2132	{
2133	PSUPDRVIDCREQCOMPREGFACTORY pReq = (PSUPDRVIDCREQCOMPREGFACTORY)pReqHdr;
2134	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY, sizeof(*pReq));
2135
2136	pReq->Hdr.rc = SUPR0ComponentRegisterFactory(pSession, pReq->u.In.pFactory);
2137	return VINF_SUCCESS;
2138	}
2139
2140	case SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY:
2141	{
2142	PSUPDRVIDCREQCOMPDEREGFACTORY pReq = (PSUPDRVIDCREQCOMPDEREGFACTORY)pReqHdr;
2143	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY, sizeof(*pReq));
2144
2145	pReq->Hdr.rc = SUPR0ComponentDeregisterFactory(pSession, pReq->u.In.pFactory);
2146	return VINF_SUCCESS;
2147	}
2148
2149	default:
2150	Log(("Unknown IDC %#lx\n", (long)uReq));
2151	break;
2152	}
2153
2154	#undef REQ_CHECK_IDC_SIZE
2155	return VERR_NOT_SUPPORTED;
2156	}
2157
2158
2159	/**
2160	* Register a object for reference counting.
2161	* The object is registered with one reference in the specified session.
2162	*
2163	* @returns Unique identifier on success (pointer).
2164	* All future reference must use this identifier.
2165	* @returns NULL on failure.
2166	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
2167	* @param pvUser1 The first user argument.
2168	* @param pvUser2 The second user argument.
2169	*/
2170	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
2171	{
2172	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2173	PSUPDRVOBJ pObj;
2174	PSUPDRVUSAGE pUsage;
2175
2176	/*
2177	* Validate the input.
2178	*/
2179	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
2180	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
2181	AssertPtrReturn(pfnDestructor, NULL);
2182
2183	/*
2184	* Allocate and initialize the object.
2185	*/
2186	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
2187	if (!pObj)
2188	return NULL;
2189	pObj->u32Magic = SUPDRVOBJ_MAGIC;
2190	pObj->enmType = enmType;
2191	pObj->pNext = NULL;
2192	pObj->cUsage = 1;
2193	pObj->pfnDestructor = pfnDestructor;
2194	pObj->pvUser1 = pvUser1;
2195	pObj->pvUser2 = pvUser2;
2196	pObj->CreatorUid = pSession->Uid;
2197	pObj->CreatorGid = pSession->Gid;
2198	pObj->CreatorProcess= pSession->Process;
2199	supdrvOSObjInitCreator(pObj, pSession);
2200
2201	/*
2202	* Allocate the usage record.
2203	* (We keep freed usage records around to simplify SUPR0ObjAddRefEx().)
2204	*/
2205	RTSpinlockAcquire(pDevExt->Spinlock);
2206
2207	pUsage = pDevExt->pUsageFree;
2208	if (pUsage)
2209	pDevExt->pUsageFree = pUsage->pNext;
2210	else
2211	{
2212	RTSpinlockRelease(pDevExt->Spinlock);
2213	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
2214	if (!pUsage)
2215	{
2216	RTMemFree(pObj);
2217	return NULL;
2218	}
2219	RTSpinlockAcquire(pDevExt->Spinlock);
2220	}
2221
2222	/*
2223	* Insert the object and create the session usage record.
2224	*/
2225	/* The object. */
2226	pObj->pNext = pDevExt->pObjs;
2227	pDevExt->pObjs = pObj;
2228
2229	/* The session record. */
2230	pUsage->cUsage = 1;
2231	pUsage->pObj = pObj;
2232	pUsage->pNext = pSession->pUsage;
2233	/* Log2(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext)); */
2234	pSession->pUsage = pUsage;
2235
2236	RTSpinlockRelease(pDevExt->Spinlock);
2237
2238	Log(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
2239	return pObj;
2240	}
2241
2242
2243	/**
2244	* Increment the reference counter for the object associating the reference
2245	* with the specified session.
2246	*
2247	* @returns IPRT status code.
2248	* @param pvObj The identifier returned by SUPR0ObjRegister().
2249	* @param pSession The session which is referencing the object.
2250	*
2251	* @remarks The caller should not own any spinlocks and must carefully protect
2252	* itself against potential race with the destructor so freed memory
2253	* isn't accessed here.
2254	*/
2255	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
2256	{
2257	return SUPR0ObjAddRefEx(pvObj, pSession, false /* fNoBlocking */);
2258	}
2259
2260
2261	/**
2262	* Increment the reference counter for the object associating the reference
2263	* with the specified session.
2264	*
2265	* @returns IPRT status code.
2266	* @retval VERR_TRY_AGAIN if fNoBlocking was set and a new usage record
2267	* couldn't be allocated. (If you see this you're not doing the right
2268	* thing and it won't ever work reliably.)
2269	*
2270	* @param pvObj The identifier returned by SUPR0ObjRegister().
2271	* @param pSession The session which is referencing the object.
2272	* @param fNoBlocking Set if it's not OK to block. Never try to make the
2273	* first reference to an object in a session with this
2274	* argument set.
2275	*
2276	* @remarks The caller should not own any spinlocks and must carefully protect
2277	* itself against potential race with the destructor so freed memory
2278	* isn't accessed here.
2279	*/
2280	SUPR0DECL(int) SUPR0ObjAddRefEx(void *pvObj, PSUPDRVSESSION pSession, bool fNoBlocking)
2281	{
2282	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2283	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2284	int rc = VINF_SUCCESS;
2285	PSUPDRVUSAGE pUsagePre;
2286	PSUPDRVUSAGE pUsage;
2287
2288	/*
2289	* Validate the input.
2290	* Be ready for the destruction race (someone might be stuck in the
2291	* destructor waiting a lock we own).
2292	*/
2293	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2294	AssertPtrReturn(pObj, VERR_INVALID_POINTER);
2295	AssertMsgReturn(pObj->u32Magic == SUPDRVOBJ_MAGIC \|\| pObj->u32Magic == SUPDRVOBJ_MAGIC_DEAD,
2296	("Invalid pvObj=%p magic=%#x (expected %#x or %#x)\n", pvObj, pObj->u32Magic, SUPDRVOBJ_MAGIC, SUPDRVOBJ_MAGIC_DEAD),
2297	VERR_INVALID_PARAMETER);
2298
2299	RTSpinlockAcquire(pDevExt->Spinlock);
2300
2301	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2302	{
2303	RTSpinlockRelease(pDevExt->Spinlock);
2304
2305	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2306	return VERR_WRONG_ORDER;
2307	}
2308
2309	/*
2310	* Preallocate the usage record if we can.
2311	*/
2312	pUsagePre = pDevExt->pUsageFree;
2313	if (pUsagePre)
2314	pDevExt->pUsageFree = pUsagePre->pNext;
2315	else if (!fNoBlocking)
2316	{
2317	RTSpinlockRelease(pDevExt->Spinlock);
2318	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
2319	if (!pUsagePre)
2320	return VERR_NO_MEMORY;
2321
2322	RTSpinlockAcquire(pDevExt->Spinlock);
2323	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2324	{
2325	RTSpinlockRelease(pDevExt->Spinlock);
2326
2327	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2328	return VERR_WRONG_ORDER;
2329	}
2330	}
2331
2332	/*
2333	* Reference the object.
2334	*/
2335	pObj->cUsage++;
2336
2337	/*
2338	* Look for the session record.
2339	*/
2340	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
2341	{
2342	/Log(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2343	if (pUsage->pObj == pObj)
2344	break;
2345	}
2346	if (pUsage)
2347	pUsage->cUsage++;
2348	else if (pUsagePre)
2349	{
2350	/* create a new session record. */
2351	pUsagePre->cUsage = 1;
2352	pUsagePre->pObj = pObj;
2353	pUsagePre->pNext = pSession->pUsage;
2354	pSession->pUsage = pUsagePre;
2355	/Log(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));/
2356
2357	pUsagePre = NULL;
2358	}
2359	else
2360	{
2361	pObj->cUsage--;
2362	rc = VERR_TRY_AGAIN;
2363	}
2364
2365	/*
2366	* Put any unused usage record into the free list..
2367	*/
2368	if (pUsagePre)
2369	{
2370	pUsagePre->pNext = pDevExt->pUsageFree;
2371	pDevExt->pUsageFree = pUsagePre;
2372	}
2373
2374	RTSpinlockRelease(pDevExt->Spinlock);
2375
2376	return rc;
2377	}
2378
2379
2380	/**
2381	* Decrement / destroy a reference counter record for an object.
2382	*
2383	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
2384	*
2385	* @returns IPRT status code.
2386	* @retval VINF_SUCCESS if not destroyed.
2387	* @retval VINF_OBJECT_DESTROYED if it's destroyed by this release call.
2388	* @retval VERR_INVALID_PARAMETER if the object isn't valid. Will assert in
2389	* string builds.
2390	*
2391	* @param pvObj The identifier returned by SUPR0ObjRegister().
2392	* @param pSession The session which is referencing the object.
2393	*/
2394	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
2395	{
2396	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2397	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2398	int rc = VERR_INVALID_PARAMETER;
2399	PSUPDRVUSAGE pUsage;
2400	PSUPDRVUSAGE pUsagePrev;
2401
2402	/*
2403	* Validate the input.
2404	*/
2405	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2406	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2407	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2408	VERR_INVALID_PARAMETER);
2409
2410	/*
2411	* Acquire the spinlock and look for the usage record.
2412	*/
2413	RTSpinlockAcquire(pDevExt->Spinlock);
2414
2415	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
2416	pUsage;
2417	pUsagePrev = pUsage, pUsage = pUsage->pNext)
2418	{
2419	/Log2(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2420	if (pUsage->pObj == pObj)
2421	{
2422	rc = VINF_SUCCESS;
2423	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
2424	if (pUsage->cUsage > 1)
2425	{
2426	pObj->cUsage--;
2427	pUsage->cUsage--;
2428	}
2429	else
2430	{
2431	/*
2432	* Free the session record.
2433	*/
2434	if (pUsagePrev)
2435	pUsagePrev->pNext = pUsage->pNext;
2436	else
2437	pSession->pUsage = pUsage->pNext;
2438	pUsage->pNext = pDevExt->pUsageFree;
2439	pDevExt->pUsageFree = pUsage;
2440
2441	/* What about the object? */
2442	if (pObj->cUsage > 1)
2443	pObj->cUsage--;
2444	else
2445	{
2446	/*
2447	* Object is to be destroyed, unlink it.
2448	*/
2449	pObj->u32Magic = SUPDRVOBJ_MAGIC_DEAD;
2450	rc = VINF_OBJECT_DESTROYED;
2451	if (pDevExt->pObjs == pObj)
2452	pDevExt->pObjs = pObj->pNext;
2453	else
2454	{
2455	PSUPDRVOBJ pObjPrev;
2456	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
2457	if (pObjPrev->pNext == pObj)
2458	{
2459	pObjPrev->pNext = pObj->pNext;
2460	break;
2461	}
2462	Assert(pObjPrev);
2463	}
2464	}
2465	}
2466	break;
2467	}
2468	}
2469
2470	RTSpinlockRelease(pDevExt->Spinlock);
2471
2472	/*
2473	* Call the destructor and free the object if required.
2474	*/
2475	if (rc == VINF_OBJECT_DESTROYED)
2476	{
2477	Log(("SUPR0ObjRelease: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
2478	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
2479	if (pObj->pfnDestructor)
2480	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
2481	RTMemFree(pObj);
2482	}
2483
2484	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
2485	return rc;
2486	}
2487
2488
2489	/**
2490	* Verifies that the current process can access the specified object.
2491	*
2492	* @returns The following IPRT status code:
2493	* @retval VINF_SUCCESS if access was granted.
2494	* @retval VERR_PERMISSION_DENIED if denied access.
2495	* @retval VERR_INVALID_PARAMETER if invalid parameter.
2496	*
2497	* @param pvObj The identifier returned by SUPR0ObjRegister().
2498	* @param pSession The session which wishes to access the object.
2499	* @param pszObjName Object string name. This is optional and depends on the object type.
2500	*
2501	* @remark The caller is responsible for making sure the object isn't removed while
2502	* we're inside this function. If uncertain about this, just call AddRef before calling us.
2503	*/
2504	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
2505	{
2506	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2507	int rc;
2508
2509	/*
2510	* Validate the input.
2511	*/
2512	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2513	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2514	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2515	VERR_INVALID_PARAMETER);
2516
2517	/*
2518	* Check access. (returns true if a decision has been made.)
2519	*/
2520	rc = VERR_INTERNAL_ERROR;
2521	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
2522	return rc;
2523
2524	/*
2525	* Default policy is to allow the user to access his own
2526	* stuff but nothing else.
2527	*/
2528	if (pObj->CreatorUid == pSession->Uid)
2529	return VINF_SUCCESS;
2530	return VERR_PERMISSION_DENIED;
2531	}
2532
2533
2534	/**
2535	* Lock pages.
2536	*
2537	* @returns IPRT status code.
2538	* @param pSession Session to which the locked memory should be associated.
2539	* @param pvR3 Start of the memory range to lock.
2540	* This must be page aligned.
2541	* @param cPages Number of pages to lock.
2542	* @param paPages Where to put the physical addresses of locked memory.
2543	*/
2544	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
2545	{
2546	int rc;
2547	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2548	const size_t cb = (size_t)cPages << PAGE_SHIFT;
2549	LogFlow(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
2550
2551	/*
2552	* Verify input.
2553	*/
2554	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2555	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
2556	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
2557	\|\| !pvR3)
2558	{
2559	Log(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
2560	return VERR_INVALID_PARAMETER;
2561	}
2562
2563	/*
2564	* Let IPRT do the job.
2565	*/
2566	Mem.eType = MEMREF_TYPE_LOCKED;
2567	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTMEM_PROT_READ \| RTMEM_PROT_WRITE, RTR0ProcHandleSelf());
2568	if (RT_SUCCESS(rc))
2569	{
2570	uint32_t iPage = cPages;
2571	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
2572	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
2573
2574	while (iPage-- > 0)
2575	{
2576	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2577	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
2578	{
2579	AssertMsgFailed(("iPage=%d\n", iPage));
2580	rc = VERR_INTERNAL_ERROR;
2581	break;
2582	}
2583	}
2584	if (RT_SUCCESS(rc))
2585	rc = supdrvMemAdd(&Mem, pSession);
2586	if (RT_FAILURE(rc))
2587	{
2588	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
2589	AssertRC(rc2);
2590	}
2591	}
2592
2593	return rc;
2594	}
2595
2596
2597	/**
2598	* Unlocks the memory pointed to by pv.
2599	*
2600	* @returns IPRT status code.
2601	* @param pSession Session to which the memory was locked.
2602	* @param pvR3 Memory to unlock.
2603	*/
2604	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2605	{
2606	LogFlow(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2607	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2608	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
2609	}
2610
2611
2612	/**
2613	* Allocates a chunk of page aligned memory with contiguous and fixed physical
2614	* backing.
2615	*
2616	* @returns IPRT status code.
2617	* @param pSession Session data.
2618	* @param cPages Number of pages to allocate.
2619	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
2620	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
2621	* @param pHCPhys Where to put the physical address of allocated memory.
2622	*/
2623	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
2624	{
2625	int rc;
2626	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2627	LogFlow(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
2628
2629	/*
2630	* Validate input.
2631	*/
2632	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2633	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
2634	{
2635	Log(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
2636	pSession, ppvR0, ppvR3, pHCPhys));
2637	return VERR_INVALID_PARAMETER;
2638
2639	}
2640	if (cPages < 1 \|\| cPages >= 256)
2641	{
2642	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2643	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2644	}
2645
2646	/*
2647	* Let IPRT do the job.
2648	*/
2649	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
2650	if (RT_SUCCESS(rc))
2651	{
2652	int rc2;
2653	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2654	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2655	if (RT_SUCCESS(rc))
2656	{
2657	Mem.eType = MEMREF_TYPE_CONT;
2658	rc = supdrvMemAdd(&Mem, pSession);
2659	if (!rc)
2660	{
2661	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2662	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2663	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
2664	return 0;
2665	}
2666
2667	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2668	AssertRC(rc2);
2669	}
2670	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2671	AssertRC(rc2);
2672	}
2673
2674	return rc;
2675	}
2676
2677
2678	/**
2679	* Frees memory allocated using SUPR0ContAlloc().
2680	*
2681	* @returns IPRT status code.
2682	* @param pSession The session to which the memory was allocated.
2683	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2684	*/
2685	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2686	{
2687	LogFlow(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2688	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2689	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
2690	}
2691
2692
2693	/**
2694	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
2695	*
2696	* The memory isn't zeroed.
2697	*
2698	* @returns IPRT status code.
2699	* @param pSession Session data.
2700	* @param cPages Number of pages to allocate.
2701	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
2702	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
2703	* @param paPages Where to put the physical addresses of allocated memory.
2704	*/
2705	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
2706	{
2707	unsigned iPage;
2708	int rc;
2709	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2710	LogFlow(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
2711
2712	/*
2713	* Validate input.
2714	*/
2715	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2716	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
2717	{
2718	Log(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
2719	pSession, ppvR3, ppvR0, paPages));
2720	return VERR_INVALID_PARAMETER;
2721
2722	}
2723	if (cPages < 1 \|\| cPages >= 256)
2724	{
2725	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2726	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2727	}
2728
2729	/*
2730	* Let IPRT do the work.
2731	*/
2732	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
2733	if (RT_SUCCESS(rc))
2734	{
2735	int rc2;
2736	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2737	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2738	if (RT_SUCCESS(rc))
2739	{
2740	Mem.eType = MEMREF_TYPE_LOW;
2741	rc = supdrvMemAdd(&Mem, pSession);
2742	if (!rc)
2743	{
2744	for (iPage = 0; iPage < cPages; iPage++)
2745	{
2746	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2747	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%RHp\n", paPages[iPage]));
2748	}
2749	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2750	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2751	return 0;
2752	}
2753
2754	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2755	AssertRC(rc2);
2756	}
2757
2758	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2759	AssertRC(rc2);
2760	}
2761
2762	return rc;
2763	}
2764
2765
2766	/**
2767	* Frees memory allocated using SUPR0LowAlloc().
2768	*
2769	* @returns IPRT status code.
2770	* @param pSession The session to which the memory was allocated.
2771	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2772	*/
2773	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2774	{
2775	LogFlow(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2776	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2777	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
2778	}
2779
2780
2781
2782	/**
2783	* Allocates a chunk of memory with both R0 and R3 mappings.
2784	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
2785	*
2786	* @returns IPRT status code.
2787	* @param pSession The session to associated the allocation with.
2788	* @param cb Number of bytes to allocate.
2789	* @param ppvR0 Where to store the address of the Ring-0 mapping.
2790	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2791	*/
2792	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
2793	{
2794	int rc;
2795	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2796	LogFlow(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
2797
2798	/*
2799	* Validate input.
2800	*/
2801	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2802	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
2803	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
2804	if (cb < 1 \|\| cb >= _4M)
2805	{
2806	Log(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
2807	return VERR_INVALID_PARAMETER;
2808	}
2809
2810	/*
2811	* Let IPRT do the work.
2812	*/
2813	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
2814	if (RT_SUCCESS(rc))
2815	{
2816	int rc2;
2817	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2818	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2819	if (RT_SUCCESS(rc))
2820	{
2821	Mem.eType = MEMREF_TYPE_MEM;
2822	rc = supdrvMemAdd(&Mem, pSession);
2823	if (!rc)
2824	{
2825	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2826	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2827	return VINF_SUCCESS;
2828	}
2829
2830	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2831	AssertRC(rc2);
2832	}
2833
2834	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2835	AssertRC(rc2);
2836	}
2837
2838	return rc;
2839	}
2840
2841
2842	/**
2843	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
2844	*
2845	* @returns IPRT status code.
2846	* @param pSession The session to which the memory was allocated.
2847	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2848	* @param paPages Where to store the physical addresses.
2849	*/
2850	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
2851	{
2852	PSUPDRVBUNDLE pBundle;
2853	LogFlow(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
2854
2855	/*
2856	* Validate input.
2857	*/
2858	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2859	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
2860	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
2861
2862	/*
2863	* Search for the address.
2864	*/
2865	RTSpinlockAcquire(pSession->Spinlock);
2866	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2867	{
2868	if (pBundle->cUsed > 0)
2869	{
2870	unsigned i;
2871	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2872	{
2873	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
2874	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2875	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2876	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2877	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
2878	)
2879	)
2880	{
2881	const size_t cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2882	size_t iPage;
2883	for (iPage = 0; iPage < cPages; iPage++)
2884	{
2885	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2886	paPages[iPage].uReserved = 0;
2887	}
2888	RTSpinlockRelease(pSession->Spinlock);
2889	return VINF_SUCCESS;
2890	}
2891	}
2892	}
2893	}
2894	RTSpinlockRelease(pSession->Spinlock);
2895	Log(("Failed to find %p!!!\n", (void *)uPtr));
2896	return VERR_INVALID_PARAMETER;
2897	}
2898
2899
2900	/**
2901	* Free memory allocated by SUPR0MemAlloc().
2902	*
2903	* @returns IPRT status code.
2904	* @param pSession The session owning the allocation.
2905	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2906	*/
2907	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2908	{
2909	LogFlow(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2910	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2911	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
2912	}
2913
2914
2915	/**
2916	* Allocates a chunk of memory with a kernel or/and a user mode mapping.
2917	*
2918	* The memory is fixed and it's possible to query the physical addresses using
2919	* SUPR0MemGetPhys().
2920	*
2921	* @returns IPRT status code.
2922	* @param pSession The session to associated the allocation with.
2923	* @param cPages The number of pages to allocate.
2924	* @param fFlags Flags, reserved for the future. Must be zero.
2925	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2926	* NULL if no ring-3 mapping.
2927	* @param ppvR3 Where to store the address of the Ring-0 mapping.
2928	* NULL if no ring-0 mapping.
2929	* @param paPages Where to store the addresses of the pages. Optional.
2930	*/
2931	SUPR0DECL(int) SUPR0PageAllocEx(PSUPDRVSESSION pSession, uint32_t cPages, uint32_t fFlags, PRTR3PTR ppvR3, PRTR0PTR ppvR0, PRTHCPHYS paPages)
2932	{
2933	int rc;
2934	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2935	LogFlow(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
2936
2937	/*
2938	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2939	*/
2940	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2941	AssertPtrNullReturn(ppvR3, VERR_INVALID_POINTER);
2942	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2943	AssertReturn(ppvR3 \|\| ppvR0, VERR_INVALID_PARAMETER);
2944	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2945	if (cPages < 1 \|\| cPages > VBOX_MAX_ALLOC_PAGE_COUNT)
2946	{
2947	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)));
2948	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2949	}
2950
2951	/*
2952	* Let IPRT do the work.
2953	*/
2954	if (ppvR0)
2955	rc = RTR0MemObjAllocPage(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, true /* fExecutable */);
2956	else
2957	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
2958	if (RT_SUCCESS(rc))
2959	{
2960	int rc2;
2961	if (ppvR3)
2962	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2963	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2964	else
2965	Mem.MapObjR3 = NIL_RTR0MEMOBJ;
2966	if (RT_SUCCESS(rc))
2967	{
2968	Mem.eType = MEMREF_TYPE_PAGE;
2969	rc = supdrvMemAdd(&Mem, pSession);
2970	if (!rc)
2971	{
2972	if (ppvR3)
2973	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2974	if (ppvR0)
2975	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2976	if (paPages)
2977	{
2978	uint32_t iPage = cPages;
2979	while (iPage-- > 0)
2980	{
2981	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
2982	Assert(paPages[iPage] != NIL_RTHCPHYS);
2983	}
2984	}
2985	return VINF_SUCCESS;
2986	}
2987
2988	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2989	AssertRC(rc2);
2990	}
2991
2992	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2993	AssertRC(rc2);
2994	}
2995	return rc;
2996	}
2997
2998
2999	/**
3000	* Maps a chunk of memory previously allocated by SUPR0PageAllocEx into kernel
3001	* space.
3002	*
3003	* @returns IPRT status code.
3004	* @param pSession The session to associated the allocation with.
3005	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
3006	* @param offSub Where to start mapping. Must be page aligned.
3007	* @param cbSub How much to map. Must be page aligned.
3008	* @param fFlags Flags, MBZ.
3009	* @param ppvR0 Where to return the address of the ring-0 mapping on
3010	* success.
3011	*/
3012	SUPR0DECL(int) SUPR0PageMapKernel(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t offSub, uint32_t cbSub,
3013	uint32_t fFlags, PRTR0PTR ppvR0)
3014	{
3015	int rc;
3016	PSUPDRVBUNDLE pBundle;
3017	RTR0MEMOBJ hMemObj = NIL_RTR0MEMOBJ;
3018	LogFlow(("SUPR0PageMapKernel: pSession=%p pvR3=%p offSub=%#x cbSub=%#x\n", pSession, pvR3, offSub, cbSub));
3019
3020	/*
3021	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
3022	*/
3023	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3024	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
3025	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
3026	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3027	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3028	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
3029
3030	/*
3031	* Find the memory object.
3032	*/
3033	RTSpinlockAcquire(pSession->Spinlock);
3034	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3035	{
3036	if (pBundle->cUsed > 0)
3037	{
3038	unsigned i;
3039	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3040	{
3041	if ( ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
3042	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3043	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3044	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
3045	\|\| ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED
3046	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3047	&& pBundle->aMem[i].MapObjR3 == NIL_RTR0MEMOBJ
3048	&& RTR0MemObjAddressR3(pBundle->aMem[i].MemObj) == pvR3))
3049	{
3050	hMemObj = pBundle->aMem[i].MemObj;
3051	break;
3052	}
3053	}
3054	}
3055	}
3056	RTSpinlockRelease(pSession->Spinlock);
3057
3058	rc = VERR_INVALID_PARAMETER;
3059	if (hMemObj != NIL_RTR0MEMOBJ)
3060	{
3061	/*
3062	* Do some further input validations before calling IPRT.
3063	* (Cleanup is done indirectly by telling RTR0MemObjFree to include mappings.)
3064	*/
3065	size_t cbMemObj = RTR0MemObjSize(hMemObj);
3066	if ( offSub < cbMemObj
3067	&& cbSub <= cbMemObj
3068	&& offSub + cbSub <= cbMemObj)
3069	{
3070	RTR0MEMOBJ hMapObj;
3071	rc = RTR0MemObjMapKernelEx(&hMapObj, hMemObj, (void *)-1, 0,
3072	RTMEM_PROT_READ \| RTMEM_PROT_WRITE, offSub, cbSub);
3073	if (RT_SUCCESS(rc))
3074	*ppvR0 = RTR0MemObjAddress(hMapObj);
3075	}
3076	else
3077	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
3078
3079	}
3080	return rc;
3081	}
3082
3083
3084	/**
3085	* Changes the page level protection of one or more pages previously allocated
3086	* by SUPR0PageAllocEx.
3087	*
3088	* @returns IPRT status code.
3089	* @param pSession The session to associated the allocation with.
3090	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
3091	* NIL_RTR3PTR if the ring-3 mapping should be unaffected.
3092	* @param pvR0 The ring-0 address returned by SUPR0PageAllocEx.
3093	* NIL_RTR0PTR if the ring-0 mapping should be unaffected.
3094	* @param offSub Where to start changing. Must be page aligned.
3095	* @param cbSub How much to change. Must be page aligned.
3096	* @param fProt The new page level protection, see RTMEM_PROT_*.
3097	*/
3098	SUPR0DECL(int) SUPR0PageProtect(PSUPDRVSESSION pSession, RTR3PTR pvR3, RTR0PTR pvR0, uint32_t offSub, uint32_t cbSub, uint32_t fProt)
3099	{
3100	int rc;
3101	PSUPDRVBUNDLE pBundle;
3102	RTR0MEMOBJ hMemObjR0 = NIL_RTR0MEMOBJ;
3103	RTR0MEMOBJ hMemObjR3 = NIL_RTR0MEMOBJ;
3104	LogFlow(("SUPR0PageProtect: pSession=%p pvR3=%p pvR0=%p offSub=%#x cbSub=%#x fProt-%#x\n", pSession, pvR3, pvR0, offSub, cbSub, fProt));
3105
3106	/*
3107	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
3108	*/
3109	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3110	AssertReturn(!(fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)), VERR_INVALID_PARAMETER);
3111	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3112	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3113	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
3114
3115	/*
3116	* Find the memory object.
3117	*/
3118	RTSpinlockAcquire(pSession->Spinlock);
3119	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3120	{
3121	if (pBundle->cUsed > 0)
3122	{
3123	unsigned i;
3124	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3125	{
3126	if ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
3127	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3128	&& ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3129	\|\| pvR3 == NIL_RTR3PTR)
3130	&& ( pvR0 == NIL_RTR0PTR
3131	\|\| RTR0MemObjAddress(pBundle->aMem[i].MemObj) == pvR0)
3132	&& ( pvR3 == NIL_RTR3PTR
3133	\|\| RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3))
3134	{
3135	if (pvR0 != NIL_RTR0PTR)
3136	hMemObjR0 = pBundle->aMem[i].MemObj;
3137	if (pvR3 != NIL_RTR3PTR)
3138	hMemObjR3 = pBundle->aMem[i].MapObjR3;
3139	break;
3140	}
3141	}
3142	}
3143	}
3144	RTSpinlockRelease(pSession->Spinlock);
3145
3146	rc = VERR_INVALID_PARAMETER;
3147	if ( hMemObjR0 != NIL_RTR0MEMOBJ
3148	\|\| hMemObjR3 != NIL_RTR0MEMOBJ)
3149	{
3150	/*
3151	* Do some further input validations before calling IPRT.
3152	*/
3153	size_t cbMemObj = hMemObjR0 != NIL_RTR0PTR ? RTR0MemObjSize(hMemObjR0) : RTR0MemObjSize(hMemObjR3);
3154	if ( offSub < cbMemObj
3155	&& cbSub <= cbMemObj
3156	&& offSub + cbSub <= cbMemObj)
3157	{
3158	rc = VINF_SUCCESS;
3159	if (hMemObjR3 != NIL_RTR0PTR)
3160	rc = RTR0MemObjProtect(hMemObjR3, offSub, cbSub, fProt);
3161	if (hMemObjR0 != NIL_RTR0PTR && RT_SUCCESS(rc))
3162	rc = RTR0MemObjProtect(hMemObjR0, offSub, cbSub, fProt);
3163	}
3164	else
3165	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
3166
3167	}
3168	return rc;
3169
3170	}
3171
3172
3173	/**
3174	* Free memory allocated by SUPR0PageAlloc() and SUPR0PageAllocEx().
3175	*
3176	* @returns IPRT status code.
3177	* @param pSession The session owning the allocation.
3178	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc() or
3179	* SUPR0PageAllocEx().
3180	*/
3181	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
3182	{
3183	LogFlow(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
3184	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3185	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_PAGE);
3186	}
3187
3188
3189	/**
3190	* Gets the paging mode of the current CPU.
3191	*
3192	* @returns Paging mode, SUPPAGEINGMODE_INVALID on error.
3193	*/
3194	SUPR0DECL(SUPPAGINGMODE) SUPR0GetPagingMode(void)
3195	{
3196	SUPPAGINGMODE enmMode;
3197
3198	RTR0UINTREG cr0 = ASMGetCR0();
3199	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
3200	enmMode = SUPPAGINGMODE_INVALID;
3201	else
3202	{
3203	RTR0UINTREG cr4 = ASMGetCR4();
3204	uint32_t fNXEPlusLMA = 0;
3205	if (cr4 & X86_CR4_PAE)
3206	{
3207	uint32_t fExtFeatures = ASMCpuId_EDX(0x80000001);
3208	if (fExtFeatures & (X86_CPUID_EXT_FEATURE_EDX_NX \| X86_CPUID_EXT_FEATURE_EDX_LONG_MODE))
3209	{
3210	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
3211	if ((fExtFeatures & X86_CPUID_EXT_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
3212	fNXEPlusLMA \|= RT_BIT(0);
3213	if ((fExtFeatures & X86_CPUID_EXT_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
3214	fNXEPlusLMA \|= RT_BIT(1);
3215	}
3216	}
3217
3218	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
3219	{
3220	case 0:
3221	enmMode = SUPPAGINGMODE_32_BIT;
3222	break;
3223
3224	case X86_CR4_PGE:
3225	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
3226	break;
3227
3228	case X86_CR4_PAE:
3229	enmMode = SUPPAGINGMODE_PAE;
3230	break;
3231
3232	case X86_CR4_PAE \| RT_BIT(0):
3233	enmMode = SUPPAGINGMODE_PAE_NX;
3234	break;
3235
3236	case X86_CR4_PAE \| X86_CR4_PGE:
3237	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3238	break;
3239
3240	case X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3241	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3242	break;
3243
3244	case RT_BIT(1) \| X86_CR4_PAE:
3245	enmMode = SUPPAGINGMODE_AMD64;
3246	break;
3247
3248	case RT_BIT(1) \| X86_CR4_PAE \| RT_BIT(0):
3249	enmMode = SUPPAGINGMODE_AMD64_NX;
3250	break;
3251
3252	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
3253	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
3254	break;
3255
3256	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3257	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
3258	break;
3259
3260	default:
3261	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
3262	enmMode = SUPPAGINGMODE_INVALID;
3263	break;
3264	}
3265	}
3266	return enmMode;
3267	}
3268
3269
3270	/**
3271	* Enables or disabled hardware virtualization extensions using native OS APIs.
3272	*
3273	* @returns VBox status code.
3274	* @retval VINF_SUCCESS on success.
3275	* @retval VERR_NOT_SUPPORTED if not supported by the native OS.
3276	*
3277	* @param fEnable Whether to enable or disable.
3278	*/
3279	SUPR0DECL(int) SUPR0EnableVTx(bool fEnable)
3280	{
3281	#ifdef RT_OS_DARWIN
3282	return supdrvOSEnableVTx(fEnable);
3283	#else
3284	return VERR_NOT_SUPPORTED;
3285	#endif
3286	}
3287
3288
3289	/**
3290	* Suspends hardware virtualization extensions using the native OS API.
3291	*
3292	* This is called prior to entering raw-mode context.
3293	*
3294	* @returns @c true if suspended, @c false if not.
3295	*/
3296	SUPR0DECL(bool) SUPR0SuspendVTxOnCpu(void)
3297	{
3298	#ifdef RT_OS_DARWIN
3299	return supdrvOSSuspendVTxOnCpu();
3300	#else
3301	return false;
3302	#endif
3303	}
3304
3305
3306	/**
3307	* Resumes hardware virtualization extensions using the native OS API.
3308	*
3309	* This is called after to entering raw-mode context.
3310	*
3311	* @param fSuspended The return value of SUPR0SuspendVTxOnCpu.
3312	*/
3313	SUPR0DECL(void) SUPR0ResumeVTxOnCpu(bool fSuspended)
3314	{
3315	#ifdef RT_OS_DARWIN
3316	supdrvOSResumeVTxOnCpu(fSuspended);
3317	#else
3318	Assert(!fSuspended);
3319	#endif
3320	}
3321
3322
3323	/**
3324	* Quries the AMD-V and VT-x capabilities of the calling CPU.
3325	*
3326	* @returns VBox status code.
3327	* @retval VERR_VMX_NO_VMX
3328	* @retval VERR_VMX_MSR_LOCKED_OR_DISABLED
3329	* @retval VERR_SVM_NO_SVM
3330	* @retval VERR_SVM_DISABLED
3331	* @retval VERR_UNSUPPORTED_CPU if not identifiable as an AMD, Intel or VIA
3332	* (centaur) CPU.
3333	*
3334	* @param pSession The session handle.
3335	* @param pfCaps Where to store the capabilities.
3336	*/
3337	SUPR0DECL(int) SUPR0QueryVTCaps(PSUPDRVSESSION pSession, uint32_t *pfCaps)
3338	{
3339	/*
3340	* Input validation.
3341	*/
3342	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3343	AssertPtrReturn(pfCaps, VERR_INVALID_POINTER);
3344
3345	*pfCaps = 0;
3346
3347	if (ASMHasCpuId())
3348	{
3349	uint32_t fFeaturesECX, fFeaturesEDX, uDummy;
3350	uint32_t uMaxId, uVendorEBX, uVendorECX, uVendorEDX;
3351	uint64_t u64Value;
3352
3353	ASMCpuId(0, &uMaxId, &uVendorEBX, &uVendorECX, &uVendorEDX);
3354	ASMCpuId(1, &uDummy, &uDummy, &fFeaturesECX, &fFeaturesEDX);
3355
3356	if ( ASMIsValidStdRange(uMaxId)
3357	&& ( ASMIsIntelCpuEx( uVendorEBX, uVendorECX, uVendorEDX)
3358	\|\| ASMIsViaCentaurCpuEx(uVendorEBX, uVendorECX, uVendorEDX) )
3359	)
3360	{
3361	if ( (fFeaturesECX & X86_CPUID_FEATURE_ECX_VMX)
3362	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3363	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3364	)
3365	{
3366	/*
3367	* Both the LOCK and VMXON bit must be set; otherwise VMXON will generate a #GP.
3368	* Once the lock bit is set, this MSR can no longer be modified.
3369	*/
3370	u64Value = ASMRdMsr(MSR_IA32_FEATURE_CONTROL);
3371	if ( (u64Value & (MSR_IA32_FEATURE_CONTROL_VMXON \| MSR_IA32_FEATURE_CONTROL_LOCK))
3372	== (MSR_IA32_FEATURE_CONTROL_VMXON \| MSR_IA32_FEATURE_CONTROL_LOCK) /* enabled and locked */
3373	\|\| !(u64Value & MSR_IA32_FEATURE_CONTROL_LOCK) /* not enabled, but not locked either */
3374	)
3375	{
3376	VMX_CAPABILITY vtCaps;
3377
3378	*pfCaps \|= SUPVTCAPS_VT_X;
3379
3380	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS);
3381	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC_USE_SECONDARY_EXEC_CTRL)
3382	{
3383	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS2);
3384	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC2_EPT)
3385	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3386	}
3387	return VINF_SUCCESS;
3388	}
3389	return VERR_VMX_MSR_LOCKED_OR_DISABLED;
3390	}
3391	return VERR_VMX_NO_VMX;
3392	}
3393
3394	if ( ASMIsAmdCpuEx(uVendorEBX, uVendorECX, uVendorEDX)
3395	&& ASMIsValidStdRange(uMaxId))
3396	{
3397	uint32_t fExtFeaturesEcx, uExtMaxId;
3398	ASMCpuId(0x80000000, &uExtMaxId, &uDummy, &uDummy, &uDummy);
3399	ASMCpuId(0x80000001, &uDummy, &uDummy, &fExtFeaturesEcx, &uDummy);
3400	if ( ASMIsValidExtRange(uExtMaxId)
3401	&& uExtMaxId >= 0x8000000a
3402	&& (fExtFeaturesEcx & X86_CPUID_AMD_FEATURE_ECX_SVM)
3403	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3404	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3405	)
3406	{
3407	/* Check if SVM is disabled */
3408	u64Value = ASMRdMsr(MSR_K8_VM_CR);
3409	if (!(u64Value & MSR_K8_VM_CR_SVM_DISABLE))
3410	{
3411	uint32_t fSvmFeatures;
3412	*pfCaps \|= SUPVTCAPS_AMD_V;
3413
3414	/* Query AMD-V features. */
3415	ASMCpuId(0x8000000a, &uDummy, &uDummy, &uDummy, &fSvmFeatures);
3416	if (fSvmFeatures & AMD_CPUID_SVM_FEATURE_EDX_NESTED_PAGING)
3417	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3418
3419	return VINF_SUCCESS;
3420	}
3421	return VERR_SVM_DISABLED;
3422	}
3423	return VERR_SVM_NO_SVM;
3424	}
3425	}
3426
3427	return VERR_UNSUPPORTED_CPU;
3428	}
3429
3430
3431	/**
3432	* (Re-)initializes the per-cpu structure prior to starting or resuming the GIP
3433	* updating.
3434	*
3435	* @param pGipCpu The per CPU structure for this CPU.
3436	* @param u64NanoTS The current time.
3437	*/
3438	static void supdrvGipReInitCpu(PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
3439	{
3440	pGipCpu->u64TSC = ASMReadTSC() - pGipCpu->u32UpdateIntervalTSC;
3441	pGipCpu->u64NanoTS = u64NanoTS;
3442	}
3443
3444
3445	/**
3446	* Set the current TSC and NanoTS value for the CPU.
3447	*
3448	* @param idCpu The CPU ID. Unused - we have to use the APIC ID.
3449	* @param pvUser1 Pointer to the ring-0 GIP mapping.
3450	* @param pvUser2 Pointer to the variable holding the current time.
3451	*/
3452	static DECLCALLBACK(void) supdrvGipReInitCpuCallback(RTCPUID idCpu, void pvUser1, void pvUser2)
3453	{
3454	PSUPGLOBALINFOPAGE pGip = (PSUPGLOBALINFOPAGE)pvUser1;
3455	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
3456
3457	if (RT_LIKELY(iCpu < pGip->cCpus && pGip->aCPUs[iCpu].idCpu == idCpu))
3458	supdrvGipReInitCpu(&pGip->aCPUs[iCpu], (uint64_t )pvUser2);
3459
3460	NOREF(pvUser2);
3461	NOREF(idCpu);
3462	}
3463
3464
3465	/**
3466	* Maps the GIP into userspace and/or get the physical address of the GIP.
3467	*
3468	* @returns IPRT status code.
3469	* @param pSession Session to which the GIP mapping should belong.
3470	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
3471	* @param pHCPhysGip Where to store the physical address. (optional)
3472	*
3473	* @remark There is no reference counting on the mapping, so one call to this function
3474	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
3475	* and remove the session as a GIP user.
3476	*/
3477	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
3478	{
3479	int rc;
3480	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3481	RTR3PTR pGipR3 = NIL_RTR3PTR;
3482	RTHCPHYS HCPhys = NIL_RTHCPHYS;
3483	LogFlow(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
3484
3485	/*
3486	* Validate
3487	*/
3488	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3489	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
3490	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
3491
3492	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3493	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3494	#else
3495	RTSemFastMutexRequest(pDevExt->mtxGip);
3496	#endif
3497	if (pDevExt->pGip)
3498	{
3499	/*
3500	* Map it?
3501	*/
3502	rc = VINF_SUCCESS;
3503	if (ppGipR3)
3504	{
3505	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
3506	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
3507	RTMEM_PROT_READ, RTR0ProcHandleSelf());
3508	if (RT_SUCCESS(rc))
3509	pGipR3 = RTR0MemObjAddressR3(pSession->GipMapObjR3);
3510	}
3511
3512	/*
3513	* Get physical address.
3514	*/
3515	if (pHCPhysGip && RT_SUCCESS(rc))
3516	HCPhys = pDevExt->HCPhysGip;
3517
3518	/*
3519	* Reference globally.
3520	*/
3521	if (!pSession->fGipReferenced && RT_SUCCESS(rc))
3522	{
3523	pSession->fGipReferenced = 1;
3524	pDevExt->cGipUsers++;
3525	if (pDevExt->cGipUsers == 1)
3526	{
3527	PSUPGLOBALINFOPAGE pGipR0 = pDevExt->pGip;
3528	uint64_t u64NanoTS;
3529	uint32_t u32SystemResolution;
3530	unsigned i;
3531
3532	LogFlow(("SUPR0GipMap: Resumes GIP updating\n"));
3533
3534	/*
3535	* Try bump up the system timer resolution.
3536	* The more interrupts the better...
3537	*/
3538	if ( RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
3539	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
3540	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1953125 /* 512 HZ */, &u32SystemResolution))
3541	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 2000000 /* 500 HZ */, &u32SystemResolution))
3542	)
3543	{
3544	Assert(RTTimerGetSystemGranularity() <= u32SystemResolution);
3545	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
3546	}
3547
3548	if (pGipR0->aCPUs[0].u32TransactionId != 2 /* not the first time */)
3549	{
3550	for (i = 0; i < RT_ELEMENTS(pGipR0->aCPUs); i++)
3551	ASMAtomicUoWriteU32(&pGipR0->aCPUs[i].u32TransactionId,
3552	(pGipR0->aCPUs[i].u32TransactionId + GIP_UPDATEHZ_RECALC_FREQ * 2)
3553	& ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
3554	ASMAtomicWriteU64(&pGipR0->u64NanoTSLastUpdateHz, 0);
3555	}
3556
3557	u64NanoTS = RTTimeSystemNanoTS() - pGipR0->u32UpdateIntervalNS;
3558	if ( pGipR0->u32Mode == SUPGIPMODE_SYNC_TSC
3559	\|\| RTMpGetOnlineCount() == 1)
3560	supdrvGipReInitCpu(&pGipR0->aCPUs[0], u64NanoTS);
3561	else
3562	RTMpOnAll(supdrvGipReInitCpuCallback, pGipR0, &u64NanoTS);
3563
3564	#ifndef DO_NOT_START_GIP
3565	rc = RTTimerStart(pDevExt->pGipTimer, 0); AssertRC(rc);
3566	#endif
3567	rc = VINF_SUCCESS;
3568	}
3569	}
3570	}
3571	else
3572	{
3573	rc = VERR_GENERAL_FAILURE;
3574	Log(("SUPR0GipMap: GIP is not available!\n"));
3575	}
3576	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3577	RTSemMutexRelease(pDevExt->mtxGip);
3578	#else
3579	RTSemFastMutexRelease(pDevExt->mtxGip);
3580	#endif
3581
3582	/*
3583	* Write returns.
3584	*/
3585	if (pHCPhysGip)
3586	*pHCPhysGip = HCPhys;
3587	if (ppGipR3)
3588	*ppGipR3 = pGipR3;
3589
3590	#ifdef DEBUG_DARWIN_GIP
3591	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3592	#else
3593	LogFlow(( "SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3594	#endif
3595	return rc;
3596	}
3597
3598
3599	/**
3600	* Unmaps any user mapping of the GIP and terminates all GIP access
3601	* from this session.
3602	*
3603	* @returns IPRT status code.
3604	* @param pSession Session to which the GIP mapping should belong.
3605	*/
3606	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
3607	{
3608	int rc = VINF_SUCCESS;
3609	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3610	#ifdef DEBUG_DARWIN_GIP
3611	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
3612	pSession,
3613	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
3614	pSession->GipMapObjR3));
3615	#else
3616	LogFlow(("SUPR0GipUnmap: pSession=%p\n", pSession));
3617	#endif
3618	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3619
3620	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3621	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3622	#else
3623	RTSemFastMutexRequest(pDevExt->mtxGip);
3624	#endif
3625
3626	/*
3627	* Unmap anything?
3628	*/
3629	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
3630	{
3631	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
3632	AssertRC(rc);
3633	if (RT_SUCCESS(rc))
3634	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
3635	}
3636
3637	/*
3638	* Dereference global GIP.
3639	*/
3640	if (pSession->fGipReferenced && !rc)
3641	{
3642	pSession->fGipReferenced = 0;
3643	if ( pDevExt->cGipUsers > 0
3644	&& !--pDevExt->cGipUsers)
3645	{
3646	LogFlow(("SUPR0GipUnmap: Suspends GIP updating\n"));
3647	#ifndef DO_NOT_START_GIP
3648	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = VINF_SUCCESS;
3649	#endif
3650
3651	if (pDevExt->u32SystemTimerGranularityGrant)
3652	{
3653	int rc2 = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant);
3654	AssertRC(rc2);
3655	pDevExt->u32SystemTimerGranularityGrant = 0;
3656	}
3657	}
3658	}
3659
3660	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3661	RTSemMutexRelease(pDevExt->mtxGip);
3662	#else
3663	RTSemFastMutexRelease(pDevExt->mtxGip);
3664	#endif
3665
3666	return rc;
3667	}
3668
3669
3670	/**
3671	* Gets the GIP pointer.
3672	*
3673	* @returns Pointer to the GIP or NULL.
3674	*/
3675	SUPDECL(PSUPGLOBALINFOPAGE) SUPGetGIP(void)
3676	{
3677	return g_pSUPGlobalInfoPage;
3678	}
3679
3680
3681	/**
3682	* Register a component factory with the support driver.
3683	*
3684	* This is currently restricted to kernel sessions only.
3685	*
3686	* @returns VBox status code.
3687	* @retval VINF_SUCCESS on success.
3688	* @retval VERR_NO_MEMORY if we're out of memory.
3689	* @retval VERR_ALREADY_EXISTS if the factory has already been registered.
3690	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3691	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3692	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3693	*
3694	* @param pSession The SUPDRV session (must be a ring-0 session).
3695	* @param pFactory Pointer to the component factory registration structure.
3696	*
3697	* @remarks This interface is also available via SUPR0IdcComponentRegisterFactory.
3698	*/
3699	SUPR0DECL(int) SUPR0ComponentRegisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3700	{
3701	PSUPDRVFACTORYREG pNewReg;
3702	const char *psz;
3703	int rc;
3704
3705	/*
3706	* Validate parameters.
3707	*/
3708	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3709	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3710	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3711	AssertPtrReturn(pFactory->pfnQueryFactoryInterface, VERR_INVALID_POINTER);
3712	psz = RTStrEnd(pFactory->szName, sizeof(pFactory->szName));
3713	AssertReturn(psz, VERR_INVALID_PARAMETER);
3714
3715	/*
3716	* Allocate and initialize a new registration structure.
3717	*/
3718	pNewReg = (PSUPDRVFACTORYREG)RTMemAlloc(sizeof(SUPDRVFACTORYREG));
3719	if (pNewReg)
3720	{
3721	pNewReg->pNext = NULL;
3722	pNewReg->pFactory = pFactory;
3723	pNewReg->pSession = pSession;
3724	pNewReg->cchName = psz - &pFactory->szName[0];
3725
3726	/*
3727	* Add it to the tail of the list after checking for prior registration.
3728	*/
3729	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3730	if (RT_SUCCESS(rc))
3731	{
3732	PSUPDRVFACTORYREG pPrev = NULL;
3733	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3734	while (pCur && pCur->pFactory != pFactory)
3735	{
3736	pPrev = pCur;
3737	pCur = pCur->pNext;
3738	}
3739	if (!pCur)
3740	{
3741	if (pPrev)
3742	pPrev->pNext = pNewReg;
3743	else
3744	pSession->pDevExt->pComponentFactoryHead = pNewReg;
3745	rc = VINF_SUCCESS;
3746	}
3747	else
3748	rc = VERR_ALREADY_EXISTS;
3749
3750	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3751	}
3752
3753	if (RT_FAILURE(rc))
3754	RTMemFree(pNewReg);
3755	}
3756	else
3757	rc = VERR_NO_MEMORY;
3758	return rc;
3759	}
3760
3761
3762	/**
3763	* Deregister a component factory.
3764	*
3765	* @returns VBox status code.
3766	* @retval VINF_SUCCESS on success.
3767	* @retval VERR_NOT_FOUND if the factory wasn't registered.
3768	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3769	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3770	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3771	*
3772	* @param pSession The SUPDRV session (must be a ring-0 session).
3773	* @param pFactory Pointer to the component factory registration structure
3774	* previously passed SUPR0ComponentRegisterFactory().
3775	*
3776	* @remarks This interface is also available via SUPR0IdcComponentDeregisterFactory.
3777	*/
3778	SUPR0DECL(int) SUPR0ComponentDeregisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3779	{
3780	int rc;
3781
3782	/*
3783	* Validate parameters.
3784	*/
3785	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3786	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3787	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3788
3789	/*
3790	* Take the lock and look for the registration record.
3791	*/
3792	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3793	if (RT_SUCCESS(rc))
3794	{
3795	PSUPDRVFACTORYREG pPrev = NULL;
3796	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3797	while (pCur && pCur->pFactory != pFactory)
3798	{
3799	pPrev = pCur;
3800	pCur = pCur->pNext;
3801	}
3802	if (pCur)
3803	{
3804	if (!pPrev)
3805	pSession->pDevExt->pComponentFactoryHead = pCur->pNext;
3806	else
3807	pPrev->pNext = pCur->pNext;
3808
3809	pCur->pNext = NULL;
3810	pCur->pFactory = NULL;
3811	pCur->pSession = NULL;
3812	rc = VINF_SUCCESS;
3813	}
3814	else
3815	rc = VERR_NOT_FOUND;
3816
3817	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3818
3819	RTMemFree(pCur);
3820	}
3821	return rc;
3822	}
3823
3824
3825	/**
3826	* Queries a component factory.
3827	*
3828	* @returns VBox status code.
3829	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3830	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3831	* @retval VERR_SUPDRV_COMPONENT_NOT_FOUND if the component factory wasn't found.
3832	* @retval VERR_SUPDRV_INTERFACE_NOT_SUPPORTED if the interface wasn't supported.
3833	*
3834	* @param pSession The SUPDRV session.
3835	* @param pszName The name of the component factory.
3836	* @param pszInterfaceUuid The UUID of the factory interface (stringified).
3837	* @param ppvFactoryIf Where to store the factory interface.
3838	*/
3839	SUPR0DECL(int) SUPR0ComponentQueryFactory(PSUPDRVSESSION pSession, const char pszName, const char pszInterfaceUuid, void **ppvFactoryIf)
3840	{
3841	const char *pszEnd;
3842	size_t cchName;
3843	int rc;
3844
3845	/*
3846	* Validate parameters.
3847	*/
3848	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3849
3850	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
3851	pszEnd = RTStrEnd(pszName, RT_SIZEOFMEMB(SUPDRVFACTORY, szName));
3852	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3853	cchName = pszEnd - pszName;
3854
3855	AssertPtrReturn(pszInterfaceUuid, VERR_INVALID_POINTER);
3856	pszEnd = RTStrEnd(pszInterfaceUuid, RTUUID_STR_LENGTH);
3857	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3858
3859	AssertPtrReturn(ppvFactoryIf, VERR_INVALID_POINTER);
3860	*ppvFactoryIf = NULL;
3861
3862	/*
3863	* Take the lock and try all factories by this name.
3864	*/
3865	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3866	if (RT_SUCCESS(rc))
3867	{
3868	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3869	rc = VERR_SUPDRV_COMPONENT_NOT_FOUND;
3870	while (pCur)
3871	{
3872	if ( pCur->cchName == cchName
3873	&& !memcmp(pCur->pFactory->szName, pszName, cchName))
3874	{
3875	void *pvFactory = pCur->pFactory->pfnQueryFactoryInterface(pCur->pFactory, pSession, pszInterfaceUuid);
3876	if (pvFactory)
3877	{
3878	*ppvFactoryIf = pvFactory;
3879	rc = VINF_SUCCESS;
3880	break;
3881	}
3882	rc = VERR_SUPDRV_INTERFACE_NOT_SUPPORTED;
3883	}
3884
3885	/* next */
3886	pCur = pCur->pNext;
3887	}
3888
3889	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3890	}
3891	return rc;
3892	}
3893
3894
3895	/**
3896	* Adds a memory object to the session.
3897	*
3898	* @returns IPRT status code.
3899	* @param pMem Memory tracking structure containing the
3900	* information to track.
3901	* @param pSession The session.
3902	*/
3903	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
3904	{
3905	PSUPDRVBUNDLE pBundle;
3906
3907	/*
3908	* Find free entry and record the allocation.
3909	*/
3910	RTSpinlockAcquire(pSession->Spinlock);
3911	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3912	{
3913	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
3914	{
3915	unsigned i;
3916	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3917	{
3918	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
3919	{
3920	pBundle->cUsed++;
3921	pBundle->aMem[i] = *pMem;
3922	RTSpinlockRelease(pSession->Spinlock);
3923	return VINF_SUCCESS;
3924	}
3925	}
3926	AssertFailed(); /* !!this can't be happening!!! */
3927	}
3928	}
3929	RTSpinlockRelease(pSession->Spinlock);
3930
3931	/*
3932	* Need to allocate a new bundle.
3933	* Insert into the last entry in the bundle.
3934	*/
3935	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
3936	if (!pBundle)
3937	return VERR_NO_MEMORY;
3938
3939	/* take last entry. */
3940	pBundle->cUsed++;
3941	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
3942
3943	/* insert into list. */
3944	RTSpinlockAcquire(pSession->Spinlock);
3945	pBundle->pNext = pSession->Bundle.pNext;
3946	pSession->Bundle.pNext = pBundle;
3947	RTSpinlockRelease(pSession->Spinlock);
3948
3949	return VINF_SUCCESS;
3950	}
3951
3952
3953	/**
3954	* Releases a memory object referenced by pointer and type.
3955	*
3956	* @returns IPRT status code.
3957	* @param pSession Session data.
3958	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
3959	* @param eType Memory type.
3960	*/
3961	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
3962	{
3963	PSUPDRVBUNDLE pBundle;
3964
3965	/*
3966	* Validate input.
3967	*/
3968	if (!uPtr)
3969	{
3970	Log(("Illegal address %p\n", (void *)uPtr));
3971	return VERR_INVALID_PARAMETER;
3972	}
3973
3974	/*
3975	* Search for the address.
3976	*/
3977	RTSpinlockAcquire(pSession->Spinlock);
3978	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3979	{
3980	if (pBundle->cUsed > 0)
3981	{
3982	unsigned i;
3983	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3984	{
3985	if ( pBundle->aMem[i].eType == eType
3986	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3987	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
3988	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3989	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
3990	)
3991	{
3992	/* Make a copy of it and release it outside the spinlock. */
3993	SUPDRVMEMREF Mem = pBundle->aMem[i];
3994	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
3995	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
3996	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
3997	RTSpinlockRelease(pSession->Spinlock);
3998
3999	if (Mem.MapObjR3 != NIL_RTR0MEMOBJ)
4000	{
4001	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
4002	AssertRC(rc); /** @todo figure out how to handle this. */
4003	}
4004	if (Mem.MemObj != NIL_RTR0MEMOBJ)
4005	{
4006	int rc = RTR0MemObjFree(Mem.MemObj, true /* fFreeMappings */);
4007	AssertRC(rc); /** @todo figure out how to handle this. */
4008	}
4009	return VINF_SUCCESS;
4010	}
4011	}
4012	}
4013	}
4014	RTSpinlockRelease(pSession->Spinlock);
4015	Log(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
4016	return VERR_INVALID_PARAMETER;
4017	}
4018
4019
4020	/**
4021	* Opens an image. If it's the first time it's opened the call must upload
4022	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
4023	*
4024	* This is the 1st step of the loading.
4025	*
4026	* @returns IPRT status code.
4027	* @param pDevExt Device globals.
4028	* @param pSession Session data.
4029	* @param pReq The open request.
4030	*/
4031	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
4032	{
4033	int rc;
4034	PSUPDRVLDRIMAGE pImage;
4035	void *pv;
4036	size_t cchName = strlen(pReq->u.In.szName); /* (caller checked < 32). */
4037	LogFlow(("supdrvIOCtl_LdrOpen: szName=%s cbImageWithTabs=%d\n", pReq->u.In.szName, pReq->u.In.cbImageWithTabs));
4038
4039	/*
4040	* Check if we got an instance of the image already.
4041	*/
4042	supdrvLdrLock(pDevExt);
4043	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
4044	{
4045	if ( pImage->szName[cchName] == '\0'
4046	&& !memcmp(pImage->szName, pReq->u.In.szName, cchName))
4047	{
4048	/** @todo check cbImageBits and cbImageWithTabs here, if they differs that indicates that the images are different. */
4049	pImage->cUsage++;
4050	pReq->u.Out.pvImageBase = pImage->pvImage;
4051	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
4052	pReq->u.Out.fNativeLoader = pImage->fNative;
4053	supdrvLdrAddUsage(pSession, pImage);
4054	supdrvLdrUnlock(pDevExt);
4055	return VINF_SUCCESS;
4056	}
4057	}
4058	/* (not found - add it!) */
4059
4060	/*
4061	* Allocate memory.
4062	*/
4063	Assert(cchName < sizeof(pImage->szName));
4064	pv = RTMemAlloc(sizeof(SUPDRVLDRIMAGE));
4065	if (!pv)
4066	{
4067	supdrvLdrUnlock(pDevExt);
4068	Log(("supdrvIOCtl_LdrOpen: RTMemAlloc() failed\n"));
4069	return /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_2;
4070	}
4071
4072	/*
4073	* Setup and link in the LDR stuff.
4074	*/
4075	pImage = (PSUPDRVLDRIMAGE)pv;
4076	pImage->pvImage = NULL;
4077	pImage->pvImageAlloc = NULL;
4078	pImage->cbImageWithTabs = pReq->u.In.cbImageWithTabs;
4079	pImage->cbImageBits = pReq->u.In.cbImageBits;
4080	pImage->cSymbols = 0;
4081	pImage->paSymbols = NULL;
4082	pImage->pachStrTab = NULL;
4083	pImage->cbStrTab = 0;
4084	pImage->pfnModuleInit = NULL;
4085	pImage->pfnModuleTerm = NULL;
4086	pImage->pfnServiceReqHandler = NULL;
4087	pImage->uState = SUP_IOCTL_LDR_OPEN;
4088	pImage->cUsage = 1;
4089	pImage->pDevExt = pDevExt;
4090	memcpy(pImage->szName, pReq->u.In.szName, cchName + 1);
4091
4092	/*
4093	* Try load it using the native loader, if that isn't supported, fall back
4094	* on the older method.
4095	*/
4096	pImage->fNative = true;
4097	rc = supdrvOSLdrOpen(pDevExt, pImage, pReq->u.In.szFilename);
4098	if (rc == VERR_NOT_SUPPORTED)
4099	{
4100	pImage->pvImageAlloc = RTMemExecAlloc(pImage->cbImageBits + 31);
4101	pImage->pvImage = RT_ALIGN_P(pImage->pvImageAlloc, 32);
4102	pImage->fNative = false;
4103	rc = pImage->pvImageAlloc ? VINF_SUCCESS : VERR_NO_EXEC_MEMORY;
4104	}
4105	if (RT_FAILURE(rc))
4106	{
4107	supdrvLdrUnlock(pDevExt);
4108	RTMemFree(pImage);
4109	Log(("supdrvIOCtl_LdrOpen(%s): failed - %Rrc\n", pReq->u.In.szName, rc));
4110	return rc;
4111	}
4112	Assert(VALID_PTR(pImage->pvImage) \|\| RT_FAILURE(rc));
4113
4114	/*
4115	* Link it.
4116	*/
4117	pImage->pNext = pDevExt->pLdrImages;
4118	pDevExt->pLdrImages = pImage;
4119
4120	supdrvLdrAddUsage(pSession, pImage);
4121
4122	pReq->u.Out.pvImageBase = pImage->pvImage;
4123	pReq->u.Out.fNeedsLoading = true;
4124	pReq->u.Out.fNativeLoader = pImage->fNative;
4125	supdrvOSLdrNotifyOpened(pDevExt, pImage);
4126
4127	supdrvLdrUnlock(pDevExt);
4128	return VINF_SUCCESS;
4129	}
4130
4131
4132	/**
4133	* Worker that validates a pointer to an image entrypoint.
4134	*
4135	* @returns IPRT status code.
4136	* @param pDevExt The device globals.
4137	* @param pImage The loader image.
4138	* @param pv The pointer into the image.
4139	* @param fMayBeNull Whether it may be NULL.
4140	* @param pszWhat What is this entrypoint? (for logging)
4141	* @param pbImageBits The image bits prepared by ring-3.
4142	*
4143	* @remarks Will leave the lock on failure.
4144	*/
4145	static int supdrvLdrValidatePointer(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage, void *pv,
4146	bool fMayBeNull, const uint8_t pbImageBits, const char pszWhat)
4147	{
4148	if (!fMayBeNull \|\| pv)
4149	{
4150	if ((uintptr_t)pv - (uintptr_t)pImage->pvImage >= pImage->cbImageBits)
4151	{
4152	supdrvLdrUnlock(pDevExt);
4153	Log(("Out of range (%p LB %#x): %s=%p\n", pImage->pvImage, pImage->cbImageBits, pszWhat, pv));
4154	return VERR_INVALID_PARAMETER;
4155	}
4156
4157	if (pImage->fNative)
4158	{
4159	int rc = supdrvOSLdrValidatePointer(pDevExt, pImage, pv, pbImageBits);
4160	if (RT_FAILURE(rc))
4161	{
4162	supdrvLdrUnlock(pDevExt);
4163	Log(("Bad entry point address: %s=%p (rc=%Rrc)\n", pszWhat, pv, rc));
4164	return rc;
4165	}
4166	}
4167	}
4168	return VINF_SUCCESS;
4169	}
4170
4171
4172	/**
4173	* Loads the image bits.
4174	*
4175	* This is the 2nd step of the loading.
4176	*
4177	* @returns IPRT status code.
4178	* @param pDevExt Device globals.
4179	* @param pSession Session data.
4180	* @param pReq The request.
4181	*/
4182	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
4183	{
4184	PSUPDRVLDRUSAGE pUsage;
4185	PSUPDRVLDRIMAGE pImage;
4186	int rc;
4187	LogFlow(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImageWithBits=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImageWithTabs));
4188
4189	/*
4190	* Find the ldr image.
4191	*/
4192	supdrvLdrLock(pDevExt);
4193	pUsage = pSession->pLdrUsage;
4194	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4195	pUsage = pUsage->pNext;
4196	if (!pUsage)
4197	{
4198	supdrvLdrUnlock(pDevExt);
4199	Log(("SUP_IOCTL_LDR_LOAD: couldn't find image!\n"));
4200	return VERR_INVALID_HANDLE;
4201	}
4202	pImage = pUsage->pImage;
4203
4204	/*
4205	* Validate input.
4206	*/
4207	if ( pImage->cbImageWithTabs != pReq->u.In.cbImageWithTabs
4208	\|\| pImage->cbImageBits != pReq->u.In.cbImageBits)
4209	{
4210	supdrvLdrUnlock(pDevExt);
4211	Log(("SUP_IOCTL_LDR_LOAD: image size mismatch!! %d(prep) != %d(load) or %d != %d\n",
4212	pImage->cbImageWithTabs, pReq->u.In.cbImageWithTabs, pImage->cbImageBits, pReq->u.In.cbImageBits));
4213	return VERR_INVALID_HANDLE;
4214	}
4215
4216	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
4217	{
4218	unsigned uState = pImage->uState;
4219	supdrvLdrUnlock(pDevExt);
4220	if (uState != SUP_IOCTL_LDR_LOAD)
4221	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
4222	return VERR_ALREADY_LOADED;
4223	}
4224
4225	switch (pReq->u.In.eEPType)
4226	{
4227	case SUPLDRLOADEP_NOTHING:
4228	break;
4229
4230	case SUPLDRLOADEP_VMMR0:
4231	rc = supdrvLdrValidatePointer( pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0, false, pReq->u.In.abImage, "pvVMMR0");
4232	if (RT_SUCCESS(rc))
4233	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt, false, pReq->u.In.abImage, "pvVMMR0EntryInt");
4234	if (RT_SUCCESS(rc))
4235	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, false, pReq->u.In.abImage, "pvVMMR0EntryFast");
4236	if (RT_SUCCESS(rc))
4237	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx, false, pReq->u.In.abImage, "pvVMMR0EntryEx");
4238	if (RT_FAILURE(rc))
4239	return rc;
4240	break;
4241
4242	case SUPLDRLOADEP_SERVICE:
4243	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.Service.pfnServiceReq, false, pReq->u.In.abImage, "pfnServiceReq");
4244	if (RT_FAILURE(rc))
4245	return rc;
4246	if ( pReq->u.In.EP.Service.apvReserved[0] != NIL_RTR0PTR
4247	\|\| pReq->u.In.EP.Service.apvReserved[1] != NIL_RTR0PTR
4248	\|\| pReq->u.In.EP.Service.apvReserved[2] != NIL_RTR0PTR)
4249	{
4250	supdrvLdrUnlock(pDevExt);
4251	Log(("Out of range (%p LB %#x): apvReserved={%p,%p,%p} MBZ!\n",
4252	pImage->pvImage, pReq->u.In.cbImageWithTabs,
4253	pReq->u.In.EP.Service.apvReserved[0],
4254	pReq->u.In.EP.Service.apvReserved[1],
4255	pReq->u.In.EP.Service.apvReserved[2]));
4256	return VERR_INVALID_PARAMETER;
4257	}
4258	break;
4259
4260	default:
4261	supdrvLdrUnlock(pDevExt);
4262	Log(("Invalid eEPType=%d\n", pReq->u.In.eEPType));
4263	return VERR_INVALID_PARAMETER;
4264	}
4265
4266	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleInit, true, pReq->u.In.abImage, "pfnModuleInit");
4267	if (RT_FAILURE(rc))
4268	return rc;
4269	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleTerm, true, pReq->u.In.abImage, "pfnModuleTerm");
4270	if (RT_FAILURE(rc))
4271	return rc;
4272
4273	/*
4274	* Allocate and copy the tables.
4275	* (No need to do try/except as this is a buffered request.)
4276	*/
4277	pImage->cbStrTab = pReq->u.In.cbStrTab;
4278	if (pImage->cbStrTab)
4279	{
4280	pImage->pachStrTab = (char *)RTMemAlloc(pImage->cbStrTab);
4281	if (pImage->pachStrTab)
4282	memcpy(pImage->pachStrTab, &pReq->u.In.abImage[pReq->u.In.offStrTab], pImage->cbStrTab);
4283	else
4284	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_3;
4285	}
4286
4287	pImage->cSymbols = pReq->u.In.cSymbols;
4288	if (RT_SUCCESS(rc) && pImage->cSymbols)
4289	{
4290	size_t cbSymbols = pImage->cSymbols * sizeof(SUPLDRSYM);
4291	pImage->paSymbols = (PSUPLDRSYM)RTMemAlloc(cbSymbols);
4292	if (pImage->paSymbols)
4293	memcpy(pImage->paSymbols, &pReq->u.In.abImage[pReq->u.In.offSymbols], cbSymbols);
4294	else
4295	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_4;
4296	}
4297
4298	/*
4299	* Copy the bits / complete native loading.
4300	*/
4301	if (RT_SUCCESS(rc))
4302	{
4303	pImage->uState = SUP_IOCTL_LDR_LOAD;
4304	pImage->pfnModuleInit = pReq->u.In.pfnModuleInit;
4305	pImage->pfnModuleTerm = pReq->u.In.pfnModuleTerm;
4306
4307	if (pImage->fNative)
4308	rc = supdrvOSLdrLoad(pDevExt, pImage, pReq->u.In.abImage, pReq);
4309	else
4310	{
4311	memcpy(pImage->pvImage, &pReq->u.In.abImage[0], pImage->cbImageBits);
4312	Log(("vboxdrv: Loaded '%s' at %p\n", pImage->szName, pImage->pvImage));
4313	}
4314	}
4315
4316	/*
4317	* Update any entry points.
4318	*/
4319	if (RT_SUCCESS(rc))
4320	{
4321	switch (pReq->u.In.eEPType)
4322	{
4323	default:
4324	case SUPLDRLOADEP_NOTHING:
4325	rc = VINF_SUCCESS;
4326	break;
4327	case SUPLDRLOADEP_VMMR0:
4328	rc = supdrvLdrSetVMMR0EPs(pDevExt, pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
4329	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
4330	break;
4331	case SUPLDRLOADEP_SERVICE:
4332	pImage->pfnServiceReqHandler = pReq->u.In.EP.Service.pfnServiceReq;
4333	rc = VINF_SUCCESS;
4334	break;
4335	}
4336	}
4337
4338	/*
4339	* On success call the module initialization.
4340	*/
4341	LogFlow(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
4342	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
4343	{
4344	Log(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
4345	pDevExt->pLdrInitImage = pImage;
4346	pDevExt->hLdrInitThread = RTThreadNativeSelf();
4347	rc = pImage->pfnModuleInit(pImage);
4348	pDevExt->pLdrInitImage = NULL;
4349	pDevExt->hLdrInitThread = NIL_RTNATIVETHREAD;
4350	if (RT_FAILURE(rc) && pDevExt->pvVMMR0 == pImage->pvImage)
4351	supdrvLdrUnsetVMMR0EPs(pDevExt);
4352	}
4353
4354	if (RT_FAILURE(rc))
4355	{
4356	/* Inform the tracing component in case ModuleInit registered TPs. */
4357	supdrvTracerModuleUnloading(pDevExt, pImage);
4358
4359	pImage->uState = SUP_IOCTL_LDR_OPEN;
4360	pImage->pfnModuleInit = NULL;
4361	pImage->pfnModuleTerm = NULL;
4362	pImage->pfnServiceReqHandler= NULL;
4363	pImage->cbStrTab = 0;
4364	RTMemFree(pImage->pachStrTab);
4365	pImage->pachStrTab = NULL;
4366	RTMemFree(pImage->paSymbols);
4367	pImage->paSymbols = NULL;
4368	pImage->cSymbols = 0;
4369	}
4370
4371	supdrvLdrUnlock(pDevExt);
4372	return rc;
4373	}
4374
4375
4376	/**
4377	* Frees a previously loaded (prep'ed) image.
4378	*
4379	* @returns IPRT status code.
4380	* @param pDevExt Device globals.
4381	* @param pSession Session data.
4382	* @param pReq The request.
4383	*/
4384	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
4385	{
4386	int rc;
4387	PSUPDRVLDRUSAGE pUsagePrev;
4388	PSUPDRVLDRUSAGE pUsage;
4389	PSUPDRVLDRIMAGE pImage;
4390	LogFlow(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
4391
4392	/*
4393	* Find the ldr image.
4394	*/
4395	supdrvLdrLock(pDevExt);
4396	pUsagePrev = NULL;
4397	pUsage = pSession->pLdrUsage;
4398	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4399	{
4400	pUsagePrev = pUsage;
4401	pUsage = pUsage->pNext;
4402	}
4403	if (!pUsage)
4404	{
4405	supdrvLdrUnlock(pDevExt);
4406	Log(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
4407	return VERR_INVALID_HANDLE;
4408	}
4409
4410	/*
4411	* Check if we can remove anything.
4412	*/
4413	rc = VINF_SUCCESS;
4414	pImage = pUsage->pImage;
4415	if (pImage->cUsage <= 1 \|\| pUsage->cUsage <= 1)
4416	{
4417	/*
4418	* Check if there are any objects with destructors in the image, if
4419	* so leave it for the session cleanup routine so we get a chance to
4420	* clean things up in the right order and not leave them all dangling.
4421	*/
4422	RTSpinlockAcquire(pDevExt->Spinlock);
4423	if (pImage->cUsage <= 1)
4424	{
4425	PSUPDRVOBJ pObj;
4426	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4427	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4428	{
4429	rc = VERR_DANGLING_OBJECTS;
4430	break;
4431	}
4432	}
4433	else
4434	{
4435	PSUPDRVUSAGE pGenUsage;
4436	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
4437	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4438	{
4439	rc = VERR_DANGLING_OBJECTS;
4440	break;
4441	}
4442	}
4443	RTSpinlockRelease(pDevExt->Spinlock);
4444	if (rc == VINF_SUCCESS)
4445	{
4446	/* unlink it */
4447	if (pUsagePrev)
4448	pUsagePrev->pNext = pUsage->pNext;
4449	else
4450	pSession->pLdrUsage = pUsage->pNext;
4451
4452	/* free it */
4453	pUsage->pImage = NULL;
4454	pUsage->pNext = NULL;
4455	RTMemFree(pUsage);
4456
4457	/*
4458	* Dereference the image.
4459	*/
4460	if (pImage->cUsage <= 1)
4461	supdrvLdrFree(pDevExt, pImage);
4462	else
4463	pImage->cUsage--;
4464	}
4465	else
4466	{
4467	Log(("supdrvIOCtl_LdrFree: Dangling objects in %p/%s!\n", pImage->pvImage, pImage->szName));
4468	rc = VINF_SUCCESS; /** @todo BRANCH-2.1: remove this after branching. */
4469	}
4470	}
4471	else
4472	{
4473	/*
4474	* Dereference both image and usage.
4475	*/
4476	pImage->cUsage--;
4477	pUsage->cUsage--;
4478	}
4479
4480	supdrvLdrUnlock(pDevExt);
4481	return rc;
4482	}
4483
4484
4485	/**
4486	* Gets the address of a symbol in an open image.
4487	*
4488	* @returns IPRT status code.
4489	* @param pDevExt Device globals.
4490	* @param pSession Session data.
4491	* @param pReq The request buffer.
4492	*/
4493	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
4494	{
4495	PSUPDRVLDRIMAGE pImage;
4496	PSUPDRVLDRUSAGE pUsage;
4497	uint32_t i;
4498	PSUPLDRSYM paSyms;
4499	const char *pchStrings;
4500	const size_t cbSymbol = strlen(pReq->u.In.szSymbol) + 1;
4501	void *pvSymbol = NULL;
4502	int rc = VERR_GENERAL_FAILURE;
4503	Log3(("supdrvIOCtl_LdrGetSymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
4504
4505	/*
4506	* Find the ldr image.
4507	*/
4508	supdrvLdrLock(pDevExt);
4509	pUsage = pSession->pLdrUsage;
4510	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4511	pUsage = pUsage->pNext;
4512	if (!pUsage)
4513	{
4514	supdrvLdrUnlock(pDevExt);
4515	Log(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
4516	return VERR_INVALID_HANDLE;
4517	}
4518	pImage = pUsage->pImage;
4519	if (pImage->uState != SUP_IOCTL_LDR_LOAD)
4520	{
4521	unsigned uState = pImage->uState;
4522	supdrvLdrUnlock(pDevExt);
4523	Log(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", uState, uState)); NOREF(uState);
4524	return VERR_ALREADY_LOADED;
4525	}
4526
4527	/*
4528	* Search the symbol strings.
4529	*
4530	* Note! The int32_t is for native loading on solaris where the data
4531	* and text segments are in very different places.
4532	*/
4533	pchStrings = pImage->pachStrTab;
4534	paSyms = pImage->paSymbols;
4535	for (i = 0; i < pImage->cSymbols; i++)
4536	{
4537	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4538	&& !memcmp(pchStrings + paSyms[i].offName, pReq->u.In.szSymbol, cbSymbol))
4539	{
4540	pvSymbol = (uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol;
4541	rc = VINF_SUCCESS;
4542	break;
4543	}
4544	}
4545	supdrvLdrUnlock(pDevExt);
4546	pReq->u.Out.pvSymbol = pvSymbol;
4547	return rc;
4548	}
4549
4550
4551	/**
4552	* Gets the address of a symbol in an open image or the support driver.
4553	*
4554	* @returns VINF_SUCCESS on success.
4555	* @returns
4556	* @param pDevExt Device globals.
4557	* @param pSession Session data.
4558	* @param pReq The request buffer.
4559	*/
4560	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq)
4561	{
4562	int rc = VINF_SUCCESS;
4563	const char *pszSymbol = pReq->u.In.pszSymbol;
4564	const char *pszModule = pReq->u.In.pszModule;
4565	size_t cbSymbol;
4566	char const *pszEnd;
4567	uint32_t i;
4568
4569	/*
4570	* Input validation.
4571	*/
4572	AssertPtrReturn(pszSymbol, VERR_INVALID_POINTER);
4573	pszEnd = RTStrEnd(pszSymbol, 512);
4574	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4575	cbSymbol = pszEnd - pszSymbol + 1;
4576
4577	if (pszModule)
4578	{
4579	AssertPtrReturn(pszModule, VERR_INVALID_POINTER);
4580	pszEnd = RTStrEnd(pszModule, 64);
4581	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4582	}
4583	Log3(("supdrvIDC_LdrGetSymbol: pszModule=%p:{%s} pszSymbol=%p:{%s}\n", pszModule, pszModule, pszSymbol, pszSymbol));
4584
4585
4586	if ( !pszModule
4587	\|\| !strcmp(pszModule, "SupDrv"))
4588	{
4589	/*
4590	* Search the support driver export table.
4591	*/
4592	for (i = 0; i < RT_ELEMENTS(g_aFunctions); i++)
4593	if (!strcmp(g_aFunctions[i].szName, pszSymbol))
4594	{
4595	pReq->u.Out.pfnSymbol = g_aFunctions[i].pfn;
4596	break;
4597	}
4598	}
4599	else
4600	{
4601	/*
4602	* Find the loader image.
4603	*/
4604	PSUPDRVLDRIMAGE pImage;
4605
4606	supdrvLdrLock(pDevExt);
4607
4608	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
4609	if (!strcmp(pImage->szName, pszModule))
4610	break;
4611	if (pImage && pImage->uState == SUP_IOCTL_LDR_LOAD)
4612	{
4613	/*
4614	* Search the symbol strings.
4615	*/
4616	const char *pchStrings = pImage->pachStrTab;
4617	PCSUPLDRSYM paSyms = pImage->paSymbols;
4618	for (i = 0; i < pImage->cSymbols; i++)
4619	{
4620	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4621	&& !memcmp(pchStrings + paSyms[i].offName, pszSymbol, cbSymbol))
4622	{
4623	/*
4624	* Found it! Calc the symbol address and add a reference to the module.
4625	*/
4626	pReq->u.Out.pfnSymbol = (PFNRT)((uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol);
4627	rc = supdrvLdrAddUsage(pSession, pImage);
4628	break;
4629	}
4630	}
4631	}
4632	else
4633	rc = pImage ? VERR_WRONG_ORDER : VERR_MODULE_NOT_FOUND;
4634
4635	supdrvLdrUnlock(pDevExt);
4636	}
4637	return rc;
4638	}
4639
4640
4641	/**
4642	* Updates the VMMR0 entry point pointers.
4643	*
4644	* @returns IPRT status code.
4645	* @param pDevExt Device globals.
4646	* @param pSession Session data.
4647	* @param pVMMR0 VMMR0 image handle.
4648	* @param pvVMMR0EntryInt VMMR0EntryInt address.
4649	* @param pvVMMR0EntryFast VMMR0EntryFast address.
4650	* @param pvVMMR0EntryEx VMMR0EntryEx address.
4651	* @remark Caller must own the loader mutex.
4652	*/
4653	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx)
4654	{
4655	int rc = VINF_SUCCESS;
4656	LogFlow(("supdrvLdrSetR0EP pvVMMR0=%p pvVMMR0EntryInt=%p\n", pvVMMR0, pvVMMR0EntryInt));
4657
4658
4659	/*
4660	* Check if not yet set.
4661	*/
4662	if (!pDevExt->pvVMMR0)
4663	{
4664	pDevExt->pvVMMR0 = pvVMMR0;
4665	pDevExt->pfnVMMR0EntryInt = pvVMMR0EntryInt;
4666	pDevExt->pfnVMMR0EntryFast = pvVMMR0EntryFast;
4667	pDevExt->pfnVMMR0EntryEx = pvVMMR0EntryEx;
4668	}
4669	else
4670	{
4671	/*
4672	* Return failure or success depending on whether the values match or not.
4673	*/
4674	if ( pDevExt->pvVMMR0 != pvVMMR0
4675	\|\| (void *)pDevExt->pfnVMMR0EntryInt != pvVMMR0EntryInt
4676	\|\| (void *)pDevExt->pfnVMMR0EntryFast != pvVMMR0EntryFast
4677	\|\| (void *)pDevExt->pfnVMMR0EntryEx != pvVMMR0EntryEx)
4678	{
4679	AssertMsgFailed(("SUP_IOCTL_LDR_SETR0EP: Already set pointing to a different module!\n"));
4680	rc = VERR_INVALID_PARAMETER;
4681	}
4682	}
4683	return rc;
4684	}
4685
4686
4687	/**
4688	* Unsets the VMMR0 entry point installed by supdrvLdrSetR0EP.
4689	*
4690	* @param pDevExt Device globals.
4691	*/
4692	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt)
4693	{
4694	pDevExt->pvVMMR0 = NULL;
4695	pDevExt->pfnVMMR0EntryInt = NULL;
4696	pDevExt->pfnVMMR0EntryFast = NULL;
4697	pDevExt->pfnVMMR0EntryEx = NULL;
4698	}
4699
4700
4701	/**
4702	* Adds a usage reference in the specified session of an image.
4703	*
4704	* Called while owning the loader semaphore.
4705	*
4706	* @returns VINF_SUCCESS on success and VERR_NO_MEMORY on failure.
4707	* @param pSession Session in question.
4708	* @param pImage Image which the session is using.
4709	*/
4710	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage)
4711	{
4712	PSUPDRVLDRUSAGE pUsage;
4713	LogFlow(("supdrvLdrAddUsage: pImage=%p\n", pImage));
4714
4715	/*
4716	* Referenced it already?
4717	*/
4718	pUsage = pSession->pLdrUsage;
4719	while (pUsage)
4720	{
4721	if (pUsage->pImage == pImage)
4722	{
4723	pUsage->cUsage++;
4724	return VINF_SUCCESS;
4725	}
4726	pUsage = pUsage->pNext;
4727	}
4728
4729	/*
4730	* Allocate new usage record.
4731	*/
4732	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
4733	AssertReturn(pUsage, /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_5);
4734	pUsage->cUsage = 1;
4735	pUsage->pImage = pImage;
4736	pUsage->pNext = pSession->pLdrUsage;
4737	pSession->pLdrUsage = pUsage;
4738	return VINF_SUCCESS;
4739	}
4740
4741
4742	/**
4743	* Frees a load image.
4744	*
4745	* @param pDevExt Pointer to device extension.
4746	* @param pImage Pointer to the image we're gonna free.
4747	* This image must exit!
4748	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
4749	*/
4750	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
4751	{
4752	PSUPDRVLDRIMAGE pImagePrev;
4753	LogFlow(("supdrvLdrFree: pImage=%p\n", pImage));
4754
4755	/* find it - arg. should've used doubly linked list. */
4756	Assert(pDevExt->pLdrImages);
4757	pImagePrev = NULL;
4758	if (pDevExt->pLdrImages != pImage)
4759	{
4760	pImagePrev = pDevExt->pLdrImages;
4761	while (pImagePrev->pNext != pImage)
4762	pImagePrev = pImagePrev->pNext;
4763	Assert(pImagePrev->pNext == pImage);
4764	}
4765
4766	/* unlink */
4767	if (pImagePrev)
4768	pImagePrev->pNext = pImage->pNext;
4769	else
4770	pDevExt->pLdrImages = pImage->pNext;
4771
4772	/* check if this is VMMR0.r0 unset its entry point pointers. */
4773	if (pDevExt->pvVMMR0 == pImage->pvImage)
4774	supdrvLdrUnsetVMMR0EPs(pDevExt);
4775
4776	/* check for objects with destructors in this image. (Shouldn't happen.) */
4777	if (pDevExt->pObjs)
4778	{
4779	unsigned cObjs = 0;
4780	PSUPDRVOBJ pObj;
4781	RTSpinlockAcquire(pDevExt->Spinlock);
4782	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4783	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4784	{
4785	pObj->pfnDestructor = NULL;
4786	cObjs++;
4787	}
4788	RTSpinlockRelease(pDevExt->Spinlock);
4789	if (cObjs)
4790	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
4791	}
4792
4793	/* call termination function if fully loaded. */
4794	if ( pImage->pfnModuleTerm
4795	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
4796	{
4797	LogFlow(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
4798	pImage->pfnModuleTerm(pImage);
4799	}
4800
4801	/* Inform the tracing component. */
4802	supdrvTracerModuleUnloading(pDevExt, pImage);
4803
4804	/* do native unload if appropriate. */
4805	if (pImage->fNative)
4806	supdrvOSLdrUnload(pDevExt, pImage);
4807
4808	/* free the image */
4809	pImage->cUsage = 0;
4810	pImage->pDevExt = NULL;
4811	pImage->pNext = NULL;
4812	pImage->uState = SUP_IOCTL_LDR_FREE;
4813	RTMemExecFree(pImage->pvImageAlloc, pImage->cbImageBits + 31);
4814	pImage->pvImageAlloc = NULL;
4815	RTMemFree(pImage->pachStrTab);
4816	pImage->pachStrTab = NULL;
4817	RTMemFree(pImage->paSymbols);
4818	pImage->paSymbols = NULL;
4819	RTMemFree(pImage);
4820	}
4821
4822
4823	/**
4824	* Acquires the loader lock.
4825	*
4826	* @returns IPRT status code.
4827	* @param pDevExt The device extension.
4828	*/
4829	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt)
4830	{
4831	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4832	int rc = RTSemMutexRequest(pDevExt->mtxLdr, RT_INDEFINITE_WAIT);
4833	#else
4834	int rc = RTSemFastMutexRequest(pDevExt->mtxLdr);
4835	#endif
4836	AssertRC(rc);
4837	return rc;
4838	}
4839
4840
4841	/**
4842	* Releases the loader lock.
4843	*
4844	* @returns IPRT status code.
4845	* @param pDevExt The device extension.
4846	*/
4847	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt)
4848	{
4849	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4850	return RTSemMutexRelease(pDevExt->mtxLdr);
4851	#else
4852	return RTSemFastMutexRelease(pDevExt->mtxLdr);
4853	#endif
4854	}
4855
4856
4857	/**
4858	* Implements the service call request.
4859	*
4860	* @returns VBox status code.
4861	* @param pDevExt The device extension.
4862	* @param pSession The calling session.
4863	* @param pReq The request packet, valid.
4864	*/
4865	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq)
4866	{
4867	#if !defined(RT_OS_WINDOWS) \|\| defined(RT_ARCH_AMD64) \|\| defined(DEBUG)
4868	int rc;
4869
4870	/*
4871	* Find the module first in the module referenced by the calling session.
4872	*/
4873	rc = supdrvLdrLock(pDevExt);
4874	if (RT_SUCCESS(rc))
4875	{
4876	PFNSUPR0SERVICEREQHANDLER pfnServiceReqHandler = NULL;
4877	PSUPDRVLDRUSAGE pUsage;
4878
4879	for (pUsage = pSession->pLdrUsage; pUsage; pUsage = pUsage->pNext)
4880	if ( pUsage->pImage->pfnServiceReqHandler
4881	&& !strcmp(pUsage->pImage->szName, pReq->u.In.szName))
4882	{
4883	pfnServiceReqHandler = pUsage->pImage->pfnServiceReqHandler;
4884	break;
4885	}
4886	supdrvLdrUnlock(pDevExt);
4887
4888	if (pfnServiceReqHandler)
4889	{
4890	/*
4891	* Call it.
4892	*/
4893	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
4894	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, NULL);
4895	else
4896	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0]);
4897	}
4898	else
4899	rc = VERR_SUPDRV_SERVICE_NOT_FOUND;
4900	}
4901
4902	/* log it */
4903	if ( RT_FAILURE(rc)
4904	&& rc != VERR_INTERRUPTED
4905	&& rc != VERR_TIMEOUT)
4906	Log(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4907	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4908	else
4909	Log4(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4910	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4911	return rc;
4912	#else /* RT_OS_WINDOWS && !RT_ARCH_AMD64 && !DEBUG */
4913	return VERR_NOT_IMPLEMENTED;
4914	#endif /* RT_OS_WINDOWS && !RT_ARCH_AMD64 && !DEBUG */
4915	}
4916
4917
4918	/**
4919	* Implements the logger settings request.
4920	*
4921	* @returns VBox status code.
4922	* @param pDevExt The device extension.
4923	* @param pSession The caller's session.
4924	* @param pReq The request.
4925	*/
4926	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq)
4927	{
4928	const char *pszGroup = &pReq->u.In.szStrings[pReq->u.In.offGroups];
4929	const char *pszFlags = &pReq->u.In.szStrings[pReq->u.In.offFlags];
4930	const char *pszDest = &pReq->u.In.szStrings[pReq->u.In.offDestination];
4931	PRTLOGGER pLogger = NULL;
4932	int rc;
4933
4934	/*
4935	* Some further validation.
4936	*/
4937	switch (pReq->u.In.fWhat)
4938	{
4939	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4940	case SUPLOGGERSETTINGS_WHAT_CREATE:
4941	break;
4942
4943	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4944	if (pszGroup \|\| pszFlags \|\| *pszDest)
4945	return VERR_INVALID_PARAMETER;
4946	if (pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_RELEASE)
4947	return VERR_ACCESS_DENIED;
4948	break;
4949
4950	default:
4951	return VERR_INTERNAL_ERROR;
4952	}
4953
4954	/*
4955	* Get the logger.
4956	*/
4957	switch (pReq->u.In.fWhich)
4958	{
4959	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4960	pLogger = RTLogGetDefaultInstance();
4961	break;
4962
4963	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4964	pLogger = RTLogRelDefaultInstance();
4965	break;
4966
4967	default:
4968	return VERR_INTERNAL_ERROR;
4969	}
4970
4971	/*
4972	* Do the job.
4973	*/
4974	switch (pReq->u.In.fWhat)
4975	{
4976	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4977	if (pLogger)
4978	{
4979	rc = RTLogFlags(pLogger, pszFlags);
4980	if (RT_SUCCESS(rc))
4981	rc = RTLogGroupSettings(pLogger, pszGroup);
4982	NOREF(pszDest);
4983	}
4984	else
4985	rc = VERR_NOT_FOUND;
4986	break;
4987
4988	case SUPLOGGERSETTINGS_WHAT_CREATE:
4989	{
4990	if (pLogger)
4991	rc = VERR_ALREADY_EXISTS;
4992	else
4993	{
4994	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
4995
4996	rc = RTLogCreate(&pLogger,
4997	0 /* fFlags */,
4998	pszGroup,
4999	pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_DEBUG
5000	? "VBOX_LOG"
5001	: "VBOX_RELEASE_LOG",
5002	RT_ELEMENTS(s_apszGroups),
5003	s_apszGroups,
5004	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER,
5005	NULL);
5006	if (RT_SUCCESS(rc))
5007	{
5008	rc = RTLogFlags(pLogger, pszFlags);
5009	NOREF(pszDest);
5010	if (RT_SUCCESS(rc))
5011	{
5012	switch (pReq->u.In.fWhich)
5013	{
5014	case SUPLOGGERSETTINGS_WHICH_DEBUG:
5015	pLogger = RTLogSetDefaultInstance(pLogger);
5016	break;
5017	case SUPLOGGERSETTINGS_WHICH_RELEASE:
5018	pLogger = RTLogRelSetDefaultInstance(pLogger);
5019	break;
5020	}
5021	}
5022	RTLogDestroy(pLogger);
5023	}
5024	}
5025	break;
5026	}
5027
5028	case SUPLOGGERSETTINGS_WHAT_DESTROY:
5029	switch (pReq->u.In.fWhich)
5030	{
5031	case SUPLOGGERSETTINGS_WHICH_DEBUG:
5032	pLogger = RTLogSetDefaultInstance(NULL);
5033	break;
5034	case SUPLOGGERSETTINGS_WHICH_RELEASE:
5035	pLogger = RTLogRelSetDefaultInstance(NULL);
5036	break;
5037	}
5038	rc = RTLogDestroy(pLogger);
5039	break;
5040
5041	default:
5042	{
5043	rc = VERR_INTERNAL_ERROR;
5044	break;
5045	}
5046	}
5047
5048	return rc;
5049	}
5050
5051
5052	/**
5053	* Creates the GIP.
5054	*
5055	* @returns VBox status code.
5056	* @param pDevExt Instance data. GIP stuff may be updated.
5057	*/
5058	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
5059	{
5060	PSUPGLOBALINFOPAGE pGip;
5061	RTHCPHYS HCPhysGip;
5062	uint32_t u32SystemResolution;
5063	uint32_t u32Interval;
5064	unsigned cCpus;
5065	int rc;
5066
5067
5068	LogFlow(("supdrvGipCreate:\n"));
5069
5070	/* assert order */
5071	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
5072	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
5073	Assert(!pDevExt->pGipTimer);
5074
5075	/*
5076	* Check the CPU count.
5077	*/
5078	cCpus = RTMpGetArraySize();
5079	if ( cCpus > RTCPUSET_MAX_CPUS
5080	\|\| cCpus > 256 /ApicId is used for the mappings/)
5081	{
5082	SUPR0Printf("VBoxDrv: Too many CPUs (%u) for the GIP (max %u)\n", cCpus, RT_MIN(RTCPUSET_MAX_CPUS, 256));
5083	return VERR_TOO_MANY_CPUS;
5084	}
5085
5086	/*
5087	* Allocate a contiguous set of pages with a default kernel mapping.
5088	*/
5089	rc = RTR0MemObjAllocCont(&pDevExt->GipMemObj, RT_UOFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), false /fExecutable/);
5090	if (RT_FAILURE(rc))
5091	{
5092	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
5093	return rc;
5094	}
5095	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
5096	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
5097
5098	/*
5099	* Find a reasonable update interval and initialize the structure.
5100	*/
5101	u32Interval = u32SystemResolution = RTTimerGetSystemGranularity();
5102	while (u32Interval < 10000000 /* 10 ms */)
5103	u32Interval += u32SystemResolution;
5104
5105	supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), 1000000000 / u32Interval /=Hz/, cCpus);
5106
5107	/*
5108	* Create the timer.
5109	* If CPU_ALL isn't supported we'll have to fall back to synchronous mode.
5110	*/
5111	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
5112	{
5113	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, RTTIMER_FLAGS_CPU_ALL, supdrvGipAsyncTimer, pDevExt);
5114	if (rc == VERR_NOT_SUPPORTED)
5115	{
5116	OSDBGPRINT(("supdrvGipCreate: omni timer not supported, falling back to synchronous mode\n"));
5117	pGip->u32Mode = SUPGIPMODE_SYNC_TSC;
5118	}
5119	}
5120	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
5121	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0, supdrvGipSyncTimer, pDevExt);
5122	if (RT_SUCCESS(rc))
5123	{
5124	rc = RTMpNotificationRegister(supdrvGipMpEvent, pDevExt);
5125	if (RT_SUCCESS(rc))
5126	{
5127	rc = RTMpOnAll(supdrvGipInitOnCpu, pDevExt, pGip);
5128	if (RT_SUCCESS(rc))
5129	{
5130	/*
5131	* We're good.
5132	*/
5133	Log(("supdrvGipCreate: %u ns interval.\n", u32Interval));
5134	g_pSUPGlobalInfoPage = pGip;
5135	return VINF_SUCCESS;
5136	}
5137
5138	OSDBGPRINT(("supdrvGipCreate: RTMpOnAll failed with rc=%Rrc\n", rc));
5139	RTMpNotificationDeregister(supdrvGipMpEvent, pDevExt);
5140
5141	}
5142	else
5143	OSDBGPRINT(("supdrvGipCreate: failed to register MP event notfication. rc=%Rrc\n", rc));
5144	}
5145	else
5146	{
5147	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %u ns interval. rc=%Rrc\n", u32Interval, rc));
5148	Assert(!pDevExt->pGipTimer);
5149	}
5150	supdrvGipDestroy(pDevExt);
5151	return rc;
5152	}
5153
5154
5155	/**
5156	* Terminates the GIP.
5157	*
5158	* @param pDevExt Instance data. GIP stuff may be updated.
5159	*/
5160	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
5161	{
5162	int rc;
5163	#ifdef DEBUG_DARWIN_GIP
5164	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
5165	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
5166	pDevExt->pGipTimer, pDevExt->GipMemObj));
5167	#endif
5168
5169	/*
5170	* Invalid the GIP data.
5171	*/
5172	if (pDevExt->pGip)
5173	{
5174	supdrvGipTerm(pDevExt->pGip);
5175	pDevExt->pGip = NULL;
5176	}
5177	g_pSUPGlobalInfoPage = NULL;
5178
5179	/*
5180	* Destroy the timer and free the GIP memory object.
5181	*/
5182	if (pDevExt->pGipTimer)
5183	{
5184	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
5185	pDevExt->pGipTimer = NULL;
5186	}
5187
5188	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
5189	{
5190	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
5191	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
5192	}
5193
5194	/*
5195	* Finally, make sure we've release the system timer resolution request
5196	* if one actually succeeded and is still pending.
5197	*/
5198	if (pDevExt->u32SystemTimerGranularityGrant)
5199	{
5200	rc = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
5201	pDevExt->u32SystemTimerGranularityGrant = 0;
5202	}
5203	}
5204
5205
5206	/**
5207	* Timer callback function sync GIP mode.
5208	* @param pTimer The timer.
5209	* @param pvUser The device extension.
5210	*/
5211	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
5212	{
5213	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
5214	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5215	uint64_t u64TSC = ASMReadTSC();
5216	uint64_t NanoTS = RTTimeSystemNanoTS();
5217
5218	supdrvGipUpdate(pDevExt, NanoTS, u64TSC, NIL_RTCPUID, iTick);
5219
5220	ASMSetFlags(fOldFlags);
5221	}
5222
5223
5224	/**
5225	* Timer callback function for async GIP mode.
5226	* @param pTimer The timer.
5227	* @param pvUser The device extension.
5228	*/
5229	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
5230	{
5231	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
5232	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5233	RTCPUID idCpu = RTMpCpuId();
5234	uint64_t u64TSC = ASMReadTSC();
5235	uint64_t NanoTS = RTTimeSystemNanoTS();
5236
5237	/** @todo reset the transaction number and whatnot when iTick == 1. */
5238	if (pDevExt->idGipMaster == idCpu)
5239	supdrvGipUpdate(pDevExt, NanoTS, u64TSC, idCpu, iTick);
5240	else
5241	supdrvGipUpdatePerCpu(pDevExt, NanoTS, u64TSC, idCpu, ASMGetApicId(), iTick);
5242
5243	ASMSetFlags(fOldFlags);
5244	}
5245
5246
5247	/**
5248	* Finds our (@a idCpu) entry, or allocates a new one if not found.
5249	*
5250	* @returns Index of the CPU in the cache set.
5251	* @param pGip The GIP.
5252	* @param idCpu The CPU ID.
5253	*/
5254	static uint32_t supdrvGipCpuIndexFromCpuId(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
5255	{
5256	uint32_t i, cTries;
5257
5258	/*
5259	* ASSUMES that CPU IDs are constant.
5260	*/
5261	for (i = 0; i < pGip->cCpus; i++)
5262	if (pGip->aCPUs[i].idCpu == idCpu)
5263	return i;
5264
5265	cTries = 0;
5266	do
5267	{
5268	for (i = 0; i < pGip->cCpus; i++)
5269	{
5270	bool fRc;
5271	ASMAtomicCmpXchgSize(&pGip->aCPUs[i].idCpu, idCpu, NIL_RTCPUID, fRc);
5272	if (fRc)
5273	return i;
5274	}
5275	} while (cTries++ < 32);
5276	AssertReleaseFailed();
5277	return i - 1;
5278	}
5279
5280
5281	/**
5282	* The calling CPU should be accounted as online, update GIP accordingly.
5283	*
5284	* This is used by supdrvGipMpEvent as well as the supdrvGipCreate.
5285	*
5286	* @param pDevExt The device extension.
5287	* @param idCpu The CPU ID.
5288	*/
5289	static void supdrvGipMpEventOnline(PSUPDRVDEVEXT pDevExt, RTCPUID idCpu)
5290	{
5291	int iCpuSet = 0;
5292	uint16_t idApic = UINT16_MAX;
5293	uint32_t i = 0;
5294	uint64_t u64NanoTS = 0;
5295	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5296
5297	AssertPtrReturnVoid(pGip);
5298	AssertRelease(idCpu == RTMpCpuId());
5299	Assert(pGip->cPossibleCpus == RTMpGetCount());
5300
5301	/*
5302	* Do this behind a spinlock with interrupts disabled as this can fire
5303	* on all CPUs simultaneously, see @bugref{6110}.
5304	*/
5305	RTSpinlockAcquire(pDevExt->hGipSpinlock);
5306
5307	/*
5308	* Update the globals.
5309	*/
5310	ASMAtomicWriteU16(&pGip->cPresentCpus, RTMpGetPresentCount());
5311	ASMAtomicWriteU16(&pGip->cOnlineCpus, RTMpGetOnlineCount());
5312	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5313	if (iCpuSet >= 0)
5314	{
5315	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5316	RTCpuSetAddByIndex(&pGip->OnlineCpuSet, iCpuSet);
5317	RTCpuSetAddByIndex(&pGip->PresentCpuSet, iCpuSet);
5318	}
5319
5320	/*
5321	* Update the entry.
5322	*/
5323	u64NanoTS = RTTimeSystemNanoTS() - pGip->u32UpdateIntervalNS;
5324	i = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
5325	supdrvGipInitCpu(pGip, &pGip->aCPUs[i], u64NanoTS);
5326	idApic = ASMGetApicId();
5327	ASMAtomicWriteU16(&pGip->aCPUs[i].idApic, idApic);
5328	ASMAtomicWriteS16(&pGip->aCPUs[i].iCpuSet, (int16_t)iCpuSet);
5329	ASMAtomicWriteSize(&pGip->aCPUs[i].idCpu, idCpu);
5330
5331	/*
5332	* Update the APIC ID and CPU set index mappings.
5333	*/
5334	ASMAtomicWriteU16(&pGip->aiCpuFromApicId[idApic], i);
5335	ASMAtomicWriteU16(&pGip->aiCpuFromCpuSetIdx[iCpuSet], i);
5336
5337	/* commit it */
5338	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_ONLINE);
5339
5340	RTSpinlockReleaseNoInts(pDevExt->hGipSpinlock);
5341	}
5342
5343
5344	/**
5345	* The CPU should be accounted as offline, update the GIP accordingly.
5346	*
5347	* This is used by supdrvGipMpEvent.
5348	*
5349	* @param pDevExt The device extension.
5350	* @param idCpu The CPU ID.
5351	*/
5352	static void supdrvGipMpEventOffline(PSUPDRVDEVEXT pDevExt, RTCPUID idCpu)
5353	{
5354	int iCpuSet;
5355	unsigned i;
5356
5357	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5358
5359	AssertPtrReturnVoid(pGip);
5360	RTSpinlockAcquire(pDevExt->hGipSpinlock);
5361
5362	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5363	AssertReturnVoid(iCpuSet >= 0);
5364
5365	i = pGip->aiCpuFromCpuSetIdx[iCpuSet];
5366	AssertReturnVoid(i < pGip->cCpus);
5367	AssertReturnVoid(pGip->aCPUs[i].idCpu == idCpu);
5368
5369	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5370	RTCpuSetDelByIndex(&pGip->OnlineCpuSet, iCpuSet);
5371
5372	/* commit it */
5373	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_OFFLINE);
5374
5375	RTSpinlockReleaseNoInts(pDevExt->hGipSpinlock);
5376	}
5377
5378
5379	/**
5380	* Multiprocessor event notification callback.
5381	*
5382	* This is used to make sure that the GIP master gets passed on to
5383	* another CPU. It also updates the associated CPU data.
5384	*
5385	* @param enmEvent The event.
5386	* @param idCpu The cpu it applies to.
5387	* @param pvUser Pointer to the device extension.
5388	*
5389	* @remarks This function -must- fire on the newly online'd CPU for the
5390	* RTMPEVENT_ONLINE case and can fire on any CPU for the
5391	* RTMPEVENT_OFFLINE case.
5392	*/
5393	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser)
5394	{
5395	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5396	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5397
5398	AssertRelease(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
5399
5400	/*
5401	* Update the GIP CPU data.
5402	*/
5403	if (pGip)
5404	{
5405	switch (enmEvent)
5406	{
5407	case RTMPEVENT_ONLINE:
5408	AssertRelease(idCpu == RTMpCpuId());
5409	supdrvGipMpEventOnline(pDevExt, idCpu);
5410	break;
5411	case RTMPEVENT_OFFLINE:
5412	supdrvGipMpEventOffline(pDevExt, idCpu);
5413	break;
5414
5415	}
5416	}
5417
5418	/*
5419	* Make sure there is a master GIP.
5420	*/
5421	if (enmEvent == RTMPEVENT_OFFLINE)
5422	{
5423	RTCPUID idGipMaster = ASMAtomicReadU32(&pDevExt->idGipMaster);
5424	if (idGipMaster == idCpu)
5425	{
5426	/*
5427	* Find a new GIP master.
5428	*/
5429	bool fIgnored;
5430	unsigned i;
5431	RTCPUID idNewGipMaster = NIL_RTCPUID;
5432	RTCPUSET OnlineCpus;
5433	RTMpGetOnlineSet(&OnlineCpus);
5434
5435	for (i = 0; i < RTCPUSET_MAX_CPUS; i++)
5436	{
5437	RTCPUID idCurCpu = RTMpCpuIdFromSetIndex(i);
5438	if ( RTCpuSetIsMember(&OnlineCpus, idCurCpu)
5439	&& idCurCpu != idGipMaster)
5440	{
5441	idNewGipMaster = idCurCpu;
5442	break;
5443	}
5444	}
5445
5446	Log(("supdrvGipMpEvent: Gip master %#lx -> %#lx\n", (long)idGipMaster, (long)idNewGipMaster));
5447	ASMAtomicCmpXchgSize(&pDevExt->idGipMaster, idNewGipMaster, idGipMaster, fIgnored);
5448	NOREF(fIgnored);
5449	}
5450	}
5451	}
5452
5453
5454	/**
5455	* Callback used by supdrvDetermineAsyncTSC to read the TSC on a CPU.
5456	*
5457	* @param idCpu Ignored.
5458	* @param pvUser1 Where to put the TSC.
5459	* @param pvUser2 Ignored.
5460	*/
5461	static DECLCALLBACK(void) supdrvDetermineAsyncTscWorker(RTCPUID idCpu, void pvUser1, void pvUser2)
5462	{
5463	#if 1
5464	ASMAtomicWriteU64((uint64_t volatile *)pvUser1, ASMReadTSC());
5465	#else
5466	(uint64_t )pvUser1 = ASMReadTSC();
5467	#endif
5468	}
5469
5470
5471	/**
5472	* Determine if Async GIP mode is required because of TSC drift.
5473	*
5474	* When using the default/normal timer code it is essential that the time stamp counter
5475	* (TSC) runs never backwards, that is, a read operation to the counter should return
5476	* a bigger value than any previous read operation. This is guaranteed by the latest
5477	* AMD CPUs and by newer Intel CPUs which never enter the C2 state (P4). In any other
5478	* case we have to choose the asynchronous timer mode.
5479	*
5480	* @param poffMin Pointer to the determined difference between different cores.
5481	* @return false if the time stamp counters appear to be synchronized, true otherwise.
5482	*/
5483	static bool supdrvDetermineAsyncTsc(uint64_t *poffMin)
5484	{
5485	/*
5486	* Just iterate all the cpus 8 times and make sure that the TSC is
5487	* ever increasing. We don't bother taking TSC rollover into account.
5488	*/
5489	int iEndCpu = RTMpGetArraySize();
5490	int iCpu;
5491	int cLoops = 8;
5492	bool fAsync = false;
5493	int rc = VINF_SUCCESS;
5494	uint64_t offMax = 0;
5495	uint64_t offMin = ~(uint64_t)0;
5496	uint64_t PrevTsc = ASMReadTSC();
5497
5498	while (cLoops-- > 0)
5499	{
5500	for (iCpu = 0; iCpu < iEndCpu; iCpu++)
5501	{
5502	uint64_t CurTsc;
5503	rc = RTMpOnSpecific(RTMpCpuIdFromSetIndex(iCpu), supdrvDetermineAsyncTscWorker, &CurTsc, NULL);
5504	if (RT_SUCCESS(rc))
5505	{
5506	if (CurTsc <= PrevTsc)
5507	{
5508	fAsync = true;
5509	offMin = offMax = PrevTsc - CurTsc;
5510	Log(("supdrvDetermineAsyncTsc: iCpu=%d cLoops=%d CurTsc=%llx PrevTsc=%llx\n",
5511	iCpu, cLoops, CurTsc, PrevTsc));
5512	break;
5513	}
5514
5515	/* Gather statistics (except the first time). */
5516	if (iCpu != 0 \|\| cLoops != 7)
5517	{
5518	uint64_t off = CurTsc - PrevTsc;
5519	if (off < offMin)
5520	offMin = off;
5521	if (off > offMax)
5522	offMax = off;
5523	Log2(("%d/%d: off=%llx\n", cLoops, iCpu, off));
5524	}
5525
5526	/* Next */
5527	PrevTsc = CurTsc;
5528	}
5529	else if (rc == VERR_NOT_SUPPORTED)
5530	break;
5531	else
5532	AssertMsg(rc == VERR_CPU_NOT_FOUND \|\| rc == VERR_CPU_OFFLINE, ("%d\n", rc));
5533	}
5534
5535	/* broke out of the loop. */
5536	if (iCpu < iEndCpu)
5537	break;
5538	}
5539
5540	poffMin = offMin; / Almost RTMpOnSpecific profiling. */
5541	Log(("supdrvDetermineAsyncTsc: returns %d; iEndCpu=%d rc=%d offMin=%llx offMax=%llx\n",
5542	fAsync, iEndCpu, rc, offMin, offMax));
5543	#if !defined(RT_OS_SOLARIS) && !defined(RT_OS_OS2) && !defined(RT_OS_WINDOWS)
5544	OSDBGPRINT(("vboxdrv: fAsync=%d offMin=%#lx offMax=%#lx\n", fAsync, (long)offMin, (long)offMax));
5545	#endif
5546	return fAsync;
5547	}
5548
5549
5550	/**
5551	* Determine the GIP TSC mode.
5552	*
5553	* @returns The most suitable TSC mode.
5554	* @param pDevExt Pointer to the device instance data.
5555	*/
5556	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt)
5557	{
5558	/*
5559	* On SMP we're faced with two problems:
5560	* (1) There might be a skew between the CPU, so that cpu0
5561	* returns a TSC that is slightly different from cpu1.
5562	* (2) Power management (and other things) may cause the TSC
5563	* to run at a non-constant speed, and cause the speed
5564	* to be different on the cpus. This will result in (1).
5565	*
5566	* So, on SMP systems we'll have to select the ASYNC update method
5567	* if there are symptoms of these problems.
5568	*/
5569	if (RTMpGetCount() > 1)
5570	{
5571	uint32_t uEAX, uEBX, uECX, uEDX;
5572	uint64_t u64DiffCoresIgnored;
5573
5574	/* Permit the user and/or the OS specific bits to force async mode. */
5575	if (supdrvOSGetForcedAsyncTscMode(pDevExt))
5576	return SUPGIPMODE_ASYNC_TSC;
5577
5578	/* Try check for current differences between the cpus. */
5579	if (supdrvDetermineAsyncTsc(&u64DiffCoresIgnored))
5580	return SUPGIPMODE_ASYNC_TSC;
5581
5582	/*
5583	* If the CPU supports power management and is an AMD one we
5584	* won't trust it unless it has the TscInvariant bit is set.
5585	*/
5586	/* Check for "AuthenticAMD" */
5587	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
5588	if ( uEAX >= 1
5589	&& ASMIsAmdCpuEx(uEBX, uECX, uEDX))
5590	{
5591	/* Check for APM support and that TscInvariant is cleared. */
5592	ASMCpuId(0x80000000, &uEAX, &uEBX, &uECX, &uEDX);
5593	if (uEAX >= 0x80000007)
5594	{
5595	ASMCpuId(0x80000007, &uEAX, &uEBX, &uECX, &uEDX);
5596	if ( !(uEDX & RT_BIT(8))/* TscInvariant */
5597	&& (uEDX & 0x3e)) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
5598	return SUPGIPMODE_ASYNC_TSC;
5599	}
5600	}
5601	}
5602	return SUPGIPMODE_SYNC_TSC;
5603	}
5604
5605
5606	/**
5607	* Initializes per-CPU GIP information.
5608	*
5609	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5610	* @param pCpu Pointer to which GIP CPU to initalize.
5611	* @param u64NanoTS The current nanosecond timestamp.
5612	*/
5613	static void supdrvGipInitCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pCpu, uint64_t u64NanoTS)
5614	{
5615	pCpu->u32TransactionId = 2;
5616	pCpu->u64NanoTS = u64NanoTS;
5617	pCpu->u64TSC = ASMReadTSC();
5618
5619	ASMAtomicWriteSize(&pCpu->enmState, SUPGIPCPUSTATE_INVALID);
5620	ASMAtomicWriteSize(&pCpu->idCpu, NIL_RTCPUID);
5621	ASMAtomicWriteS16(&pCpu->iCpuSet, -1);
5622	ASMAtomicWriteU16(&pCpu->idApic, UINT16_MAX);
5623
5624	/*
5625	* We don't know the following values until we've executed updates.
5626	* So, we'll just pretend it's a 4 GHz CPU and adjust the history it on
5627	* the 2nd timer callout.
5628	*/
5629	pCpu->u64CpuHz = _4G + 1; /* tstGIP-2 depends on this. */
5630	pCpu->u32UpdateIntervalTSC
5631	= pCpu->au32TSCHistory[0]
5632	= pCpu->au32TSCHistory[1]
5633	= pCpu->au32TSCHistory[2]
5634	= pCpu->au32TSCHistory[3]
5635	= pCpu->au32TSCHistory[4]
5636	= pCpu->au32TSCHistory[5]
5637	= pCpu->au32TSCHistory[6]
5638	= pCpu->au32TSCHistory[7]
5639	= (uint32_t)(_4G / pGip->u32UpdateHz);
5640	}
5641
5642
5643	/**
5644	* Initializes the GIP data.
5645	*
5646	* @param pDevExt Pointer to the device instance data.
5647	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5648	* @param HCPhys The physical address of the GIP.
5649	* @param u64NanoTS The current nanosecond timestamp.
5650	* @param uUpdateHz The update frequency.
5651	* @param cCpus The CPU count.
5652	*/
5653	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
5654	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned cCpus)
5655	{
5656	size_t const cbGip = RT_ALIGN_Z(RT_OFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), PAGE_SIZE);
5657	unsigned i;
5658	#ifdef DEBUG_DARWIN_GIP
5659	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5660	#else
5661	LogFlow(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5662	#endif
5663
5664	/*
5665	* Initialize the structure.
5666	*/
5667	memset(pGip, 0, cbGip);
5668	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
5669	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
5670	pGip->u32Mode = supdrvGipDeterminTscMode(pDevExt);
5671	pGip->cCpus = (uint16_t)cCpus;
5672	pGip->cPages = (uint16_t)(cbGip / PAGE_SIZE);
5673	pGip->u32UpdateHz = uUpdateHz;
5674	pGip->u32UpdateIntervalNS = 1000000000 / uUpdateHz;
5675	pGip->u64NanoTSLastUpdateHz = u64NanoTS;
5676	RTCpuSetEmpty(&pGip->OnlineCpuSet);
5677	RTCpuSetEmpty(&pGip->PresentCpuSet);
5678	RTMpGetSet(&pGip->PossibleCpuSet);
5679	pGip->cOnlineCpus = RTMpGetOnlineCount();
5680	pGip->cPresentCpus = RTMpGetPresentCount();
5681	pGip->cPossibleCpus = RTMpGetCount();
5682	pGip->idCpuMax = RTMpGetMaxCpuId();
5683	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromApicId); i++)
5684	pGip->aiCpuFromApicId[i] = 0;
5685	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx); i++)
5686	pGip->aiCpuFromCpuSetIdx[i] = UINT16_MAX;
5687
5688	for (i = 0; i < cCpus; i++)
5689	supdrvGipInitCpu(pGip, &pGip->aCPUs[i], u64NanoTS);
5690
5691	/*
5692	* Link it to the device extension.
5693	*/
5694	pDevExt->pGip = pGip;
5695	pDevExt->HCPhysGip = HCPhys;
5696	pDevExt->cGipUsers = 0;
5697	}
5698
5699
5700	/**
5701	* On CPU initialization callback for RTMpOnAll.
5702	*
5703	* @param idCpu The CPU ID.
5704	* @param pvUser1 The device extension.
5705	* @param pvUser2 The GIP.
5706	*/
5707	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2)
5708	{
5709	/* This is good enough, even though it will update some of the globals a
5710	bit to much. */
5711	supdrvGipMpEventOnline((PSUPDRVDEVEXT)pvUser1, idCpu);
5712	}
5713
5714
5715	/**
5716	* Invalidates the GIP data upon termination.
5717	*
5718	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5719	*/
5720	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
5721	{
5722	unsigned i;
5723	pGip->u32Magic = 0;
5724	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
5725	{
5726	pGip->aCPUs[i].u64NanoTS = 0;
5727	pGip->aCPUs[i].u64TSC = 0;
5728	pGip->aCPUs[i].iTSCHistoryHead = 0;
5729	}
5730	}
5731
5732
5733	/**
5734	* Worker routine for supdrvGipUpdate and supdrvGipUpdatePerCpu that
5735	* updates all the per cpu data except the transaction id.
5736	*
5737	* @param pDevExt The device extension.
5738	* @param pGipCpu Pointer to the per cpu data.
5739	* @param u64NanoTS The current time stamp.
5740	* @param u64TSC The current TSC.
5741	* @param iTick The current timer tick.
5742	*/
5743	static void supdrvGipDoUpdateCpu(PSUPDRVDEVEXT pDevExt, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick)
5744	{
5745	uint64_t u64TSCDelta;
5746	uint32_t u32UpdateIntervalTSC;
5747	uint32_t u32UpdateIntervalTSCSlack;
5748	unsigned iTSCHistoryHead;
5749	uint64_t u64CpuHz;
5750	uint32_t u32TransactionId;
5751
5752	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5753	AssertPtrReturnVoid(pGip);
5754
5755	/* Delta between this and the previous update. */
5756	ASMAtomicUoWriteU32(&pGipCpu->u32PrevUpdateIntervalNS, (uint32_t)(u64NanoTS - pGipCpu->u64NanoTS));
5757
5758	/*
5759	* Update the NanoTS.
5760	*/
5761	ASMAtomicWriteU64(&pGipCpu->u64NanoTS, u64NanoTS);
5762
5763	/*
5764	* Calc TSC delta.
5765	*/
5766	/** @todo validate the NanoTS delta, don't trust the OS to call us when it should... */
5767	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
5768	ASMAtomicWriteU64(&pGipCpu->u64TSC, u64TSC);
5769
5770	if (u64TSCDelta >> 32)
5771	{
5772	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
5773	pGipCpu->cErrors++;
5774	}
5775
5776	/*
5777	* On the 2nd and 3rd callout, reset the history with the current TSC
5778	* interval since the values entered by supdrvGipInit are totally off.
5779	* The interval on the 1st callout completely unreliable, the 2nd is a bit
5780	* better, while the 3rd should be most reliable.
5781	*/
5782	u32TransactionId = pGipCpu->u32TransactionId;
5783	if (RT_UNLIKELY( ( u32TransactionId == 5
5784	\|\| u32TransactionId == 7)
5785	&& ( iTick == 2
5786	\|\| iTick == 3) ))
5787	{
5788	unsigned i;
5789	for (i = 0; i < RT_ELEMENTS(pGipCpu->au32TSCHistory); i++)
5790	ASMAtomicUoWriteU32(&pGipCpu->au32TSCHistory[i], (uint32_t)u64TSCDelta);
5791	}
5792
5793	/*
5794	* TSC History.
5795	*/
5796	Assert(RT_ELEMENTS(pGipCpu->au32TSCHistory) == 8);
5797	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
5798	ASMAtomicWriteU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
5799	ASMAtomicWriteU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
5800
5801	/*
5802	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
5803	*/
5804	if (pGip->u32UpdateHz >= 1000)
5805	{
5806	uint32_t u32;
5807	u32 = pGipCpu->au32TSCHistory[0];
5808	u32 += pGipCpu->au32TSCHistory[1];
5809	u32 += pGipCpu->au32TSCHistory[2];
5810	u32 += pGipCpu->au32TSCHistory[3];
5811	u32 >>= 2;
5812	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
5813	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
5814	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
5815	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
5816	u32UpdateIntervalTSC >>= 2;
5817	u32UpdateIntervalTSC += u32;
5818	u32UpdateIntervalTSC >>= 1;
5819
5820	/* Value chosen for a 2GHz Athlon64 running linux 2.6.10/11, . */
5821	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
5822	}
5823	else if (pGip->u32UpdateHz >= 90)
5824	{
5825	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5826	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
5827	u32UpdateIntervalTSC >>= 1;
5828
5829	/* value chosen on a 2GHz thinkpad running windows */
5830	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
5831	}
5832	else
5833	{
5834	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5835
5836	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
5837	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
5838	}
5839	ASMAtomicWriteU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
5840
5841	/*
5842	* CpuHz.
5843	*/
5844	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, pGip->u32UpdateHz);
5845	ASMAtomicWriteU64(&pGipCpu->u64CpuHz, u64CpuHz);
5846	}
5847
5848
5849	/**
5850	* Updates the GIP.
5851	*
5852	* @param pDevExt The device extension.
5853	* @param u64NanoTS The current nanosecond timesamp.
5854	* @param u64TSC The current TSC timesamp.
5855	* @param idCpu The CPU ID.
5856	* @param iTick The current timer tick.
5857	*/
5858	static void supdrvGipUpdate(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick)
5859	{
5860	/*
5861	* Determine the relevant CPU data.
5862	*/
5863	PSUPGIPCPU pGipCpu;
5864	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5865	AssertPtrReturnVoid(pGip);
5866
5867	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
5868	pGipCpu = &pGip->aCPUs[0];
5869	else
5870	{
5871	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
5872	if (RT_UNLIKELY(iCpu >= pGip->cCpus))
5873	return;
5874	pGipCpu = &pGip->aCPUs[iCpu];
5875	if (RT_UNLIKELY(pGipCpu->idCpu != idCpu))
5876	return;
5877	}
5878
5879	/*
5880	* Start update transaction.
5881	*/
5882	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5883	{
5884	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
5885	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5886	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5887	pGipCpu->cErrors++;
5888	return;
5889	}
5890
5891	/*
5892	* Recalc the update frequency every 0x800th time.
5893	*/
5894	if (!(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
5895	{
5896	if (pGip->u64NanoTSLastUpdateHz)
5897	{
5898	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
5899	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
5900	uint32_t u32UpdateHz = (uint32_t)((UINT64_C(1000000000) * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
5901	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
5902	{
5903	ASMAtomicWriteU32(&pGip->u32UpdateHz, u32UpdateHz);
5904	ASMAtomicWriteU32(&pGip->u32UpdateIntervalNS, 1000000000 / u32UpdateHz);
5905	}
5906	#endif
5907	}
5908	ASMAtomicWriteU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS);
5909	}
5910
5911	/*
5912	* Update the data.
5913	*/
5914	supdrvGipDoUpdateCpu(pDevExt, pGipCpu, u64NanoTS, u64TSC, iTick);
5915
5916	/*
5917	* Complete transaction.
5918	*/
5919	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5920	}
5921
5922
5923	/**
5924	* Updates the per cpu GIP data for the calling cpu.
5925	*
5926	* @param pDevExt The device extension.
5927	* @param u64NanoTS The current nanosecond timesamp.
5928	* @param u64TSC The current TSC timesamp.
5929	* @param idCpu The CPU ID.
5930	* @param idApic The APIC id for the CPU index.
5931	* @param iTick The current timer tick.
5932	*/
5933	static void supdrvGipUpdatePerCpu(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC,
5934	RTCPUID idCpu, uint8_t idApic, uint64_t iTick)
5935	{
5936	uint32_t iCpu;
5937	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5938
5939	/*
5940	* Avoid a potential race when a CPU online notification doesn't fire on
5941	* the onlined CPU but the tick creeps in before the event notification is
5942	* run.
5943	*/
5944	if (RT_UNLIKELY(iTick == 1))
5945	{
5946	iCpu = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
5947	if (pGip->aCPUs[iCpu].enmState == SUPGIPCPUSTATE_OFFLINE)
5948	supdrvGipMpEventOnline(pDevExt, idCpu);
5949	}
5950
5951	iCpu = pGip->aiCpuFromApicId[idApic];
5952	if (RT_LIKELY(iCpu < pGip->cCpus))
5953	{
5954	PSUPGIPCPU pGipCpu = &pGip->aCPUs[iCpu];
5955	if (pGipCpu->idCpu == idCpu)
5956	{
5957	/*
5958	* Start update transaction.
5959	*/
5960	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5961	{
5962	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5963	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5964	pGipCpu->cErrors++;
5965	return;
5966	}
5967
5968	/*
5969	* Update the data.
5970	*/
5971	supdrvGipDoUpdateCpu(pDevExt, pGipCpu, u64NanoTS, u64TSC, iTick);
5972
5973	/*
5974	* Complete transaction.
5975	*/
5976	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5977	}
5978	}
5979	}
5980

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

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDrv.c@ 46889

Download in other formats: