SUPDrv.c@ 47199

Last change on this file since 47199 was 47199, checked in by vboxsync, 11 years ago
IPRT: Kernel thread-context hooks, linux implementation. Extended tstR0ThreadPreemption testcase to test the thread-context hooks.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 223.0 KB

Line
1	/* $Id: SUPDrv.c 47199 2013-07-16 15:45:42Z vboxsync $ */
2	/** @file
3	* VBoxDrv - The VirtualBox Support Driver - Common code.
4	*/
5
6	/*
7	* Copyright (C) 2006-2013 Oracle Corporation
8	*
9	* This file is part of VirtualBox Open Source Edition (OSE), as
10	* available from http://www.virtualbox.org. This file is free software;
11	* you can redistribute it and/or modify it under the terms of the GNU
12	* General Public License (GPL) as published by the Free Software
13	* Foundation, in version 2 as it comes in the "COPYING" file of the
14	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16	*
17	* The contents of this file may alternatively be used under the terms
18	* of the Common Development and Distribution License Version 1.0
19	* (CDDL) only, as it comes in the "COPYING.CDDL" file of the
20	* VirtualBox OSE distribution, in which case the provisions of the
21	* CDDL are applicable instead of those of the GPL.
22	*
23	* You may elect to license modified versions of this file under the
24	* terms and conditions of either the GPL or the CDDL or both.
25	*/
26
27	/*******************************************************************************
28	* Header Files *
29	*******************************************************************************/
30	#define LOG_GROUP LOG_GROUP_SUP_DRV
31	#define SUPDRV_AGNOSTIC
32	#include "SUPDrvInternal.h"
33	#ifndef PAGE_SHIFT
34	# include <iprt/param.h>
35	#endif
36	#include <iprt/asm.h>
37	#include <iprt/asm-amd64-x86.h>
38	#include <iprt/asm-math.h>
39	#include <iprt/cpuset.h>
40	#include <iprt/handletable.h>
41	#include <iprt/mem.h>
42	#include <iprt/mp.h>
43	#include <iprt/power.h>
44	#include <iprt/process.h>
45	#include <iprt/semaphore.h>
46	#include <iprt/spinlock.h>
47	#include <iprt/thread.h>
48	#include <iprt/uuid.h>
49	#include <iprt/net.h>
50	#include <iprt/crc.h>
51	#include <iprt/string.h>
52	#include <iprt/timer.h>
53	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
54	# include <iprt/rand.h>
55	# include <iprt/path.h>
56	#endif
57	#include <iprt/x86.h>
58
59	#include <VBox/param.h>
60	#include <VBox/log.h>
61	#include <VBox/err.h>
62	#include <VBox/vmm/hm_svm.h>
63	#include <VBox/vmm/hm_vmx.h>
64
65	#if defined(RT_OS_SOLARIS) \|\| defined(RT_OS_DARWIN)
66	# include "dtrace/SUPDrv.h"
67	#else
68	# define VBOXDRV_SESSION_CREATE(pvSession, fUser) do { } while (0)
69	# define VBOXDRV_SESSION_CLOSE(pvSession) do { } while (0)
70	# define VBOXDRV_IOCTL_ENTRY(pvSession, uIOCtl, pvReqHdr) do { } while (0)
71	# define VBOXDRV_IOCTL_RETURN(pvSession, uIOCtl, pvReqHdr, rcRet, rcReq) do { } while (0)
72	#endif
73
74	/*
75	* Logging assignments:
76	* Log - useful stuff, like failures.
77	* LogFlow - program flow, except the really noisy bits.
78	* Log2 - Cleanup.
79	* Log3 - Loader flow noise.
80	* Log4 - Call VMMR0 flow noise.
81	* Log5 - Native yet-to-be-defined noise.
82	* Log6 - Native ioctl flow noise.
83	*
84	* Logging requires BUILD_TYPE=debug and possibly changes to the logger
85	* instantiation in log-vbox.c(pp).
86	*/
87
88
89	/*******************************************************************************
90	* Defined Constants And Macros *
91	*******************************************************************************/
92	/** The frequency by which we recalculate the u32UpdateHz and
93	* u32UpdateIntervalNS GIP members. The value must be a power of 2. */
94	#define GIP_UPDATEHZ_RECALC_FREQ 0x800
95
96	/** @def VBOX_SVN_REV
97	* The makefile should define this if it can. */
98	#ifndef VBOX_SVN_REV
99	# define VBOX_SVN_REV 0
100	#endif
101
102	#if 0 /* Don't start the GIP timers. Useful when debugging the IPRT timer code. */
103	# define DO_NOT_START_GIP
104	#endif
105
106
107	/*******************************************************************************
108	* Internal Functions *
109	*******************************************************************************/
110	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser);
111	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser);
112	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession);
113	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType);
114	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq);
115	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq);
116	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq);
117	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq);
118	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq);
119	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx);
120	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt);
121	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage);
122	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage);
123	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt);
124	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt);
125	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq);
126	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq);
127	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt);
128	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt);
129	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
130	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
131	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser);
132	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
133	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned cCpus);
134	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2);
135	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip);
136	static void supdrvGipUpdate(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick);
137	static void supdrvGipUpdatePerCpu(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC,
138	RTCPUID idCpu, uint8_t idApic, uint64_t iTick);
139	static void supdrvGipInitCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pCpu, uint64_t u64NanoTS);
140
141
142	/*******************************************************************************
143	* Global Variables *
144	*******************************************************************************/
145	DECLEXPORT(PSUPGLOBALINFOPAGE) g_pSUPGlobalInfoPage = NULL;
146
147	/**
148	* Array of the R0 SUP API.
149	*/
150	static SUPFUNC g_aFunctions[] =
151	{
152	/* SED: START */
153	/* name function */
154	/* Entries with absolute addresses determined at runtime, fixup
155	code makes ugly ASSUMPTIONS about the order here: */
156	{ "SUPR0AbsIs64bit", (void *)0 },
157	{ "SUPR0Abs64bitKernelCS", (void *)0 },
158	{ "SUPR0Abs64bitKernelSS", (void *)0 },
159	{ "SUPR0Abs64bitKernelDS", (void *)0 },
160	{ "SUPR0AbsKernelCS", (void *)0 },
161	{ "SUPR0AbsKernelSS", (void *)0 },
162	{ "SUPR0AbsKernelDS", (void *)0 },
163	{ "SUPR0AbsKernelES", (void *)0 },
164	{ "SUPR0AbsKernelFS", (void *)0 },
165	{ "SUPR0AbsKernelGS", (void *)0 },
166	/* Normal function pointers: */
167	{ "g_pSUPGlobalInfoPage", (void )&g_pSUPGlobalInfoPage }, / SED: DATA */
168	{ "SUPGetGIP", (void *)SUPGetGIP },
169	{ "SUPR0ComponentDeregisterFactory", (void *)SUPR0ComponentDeregisterFactory },
170	{ "SUPR0ComponentQueryFactory", (void *)SUPR0ComponentQueryFactory },
171	{ "SUPR0ComponentRegisterFactory", (void *)SUPR0ComponentRegisterFactory },
172	{ "SUPR0ContAlloc", (void *)SUPR0ContAlloc },
173	{ "SUPR0ContFree", (void *)SUPR0ContFree },
174	{ "SUPR0EnableVTx", (void *)SUPR0EnableVTx },
175	{ "SUPR0SuspendVTxOnCpu", (void *)SUPR0SuspendVTxOnCpu },
176	{ "SUPR0ResumeVTxOnCpu", (void *)SUPR0ResumeVTxOnCpu },
177	{ "SUPR0GetPagingMode", (void *)SUPR0GetPagingMode },
178	{ "SUPR0LockMem", (void *)SUPR0LockMem },
179	{ "SUPR0LowAlloc", (void *)SUPR0LowAlloc },
180	{ "SUPR0LowFree", (void *)SUPR0LowFree },
181	{ "SUPR0MemAlloc", (void *)SUPR0MemAlloc },
182	{ "SUPR0MemFree", (void *)SUPR0MemFree },
183	{ "SUPR0MemGetPhys", (void *)SUPR0MemGetPhys },
184	{ "SUPR0ObjAddRef", (void *)SUPR0ObjAddRef },
185	{ "SUPR0ObjAddRefEx", (void *)SUPR0ObjAddRefEx },
186	{ "SUPR0ObjRegister", (void *)SUPR0ObjRegister },
187	{ "SUPR0ObjRelease", (void *)SUPR0ObjRelease },
188	{ "SUPR0ObjVerifyAccess", (void *)SUPR0ObjVerifyAccess },
189	{ "SUPR0PageAllocEx", (void *)SUPR0PageAllocEx },
190	{ "SUPR0PageFree", (void *)SUPR0PageFree },
191	{ "SUPR0Printf", (void *)SUPR0Printf },
192	{ "SUPR0TracerDeregisterDrv", (void *)SUPR0TracerDeregisterDrv },
193	{ "SUPR0TracerDeregisterImpl", (void *)SUPR0TracerDeregisterImpl },
194	{ "SUPR0TracerFireProbe", (void *)SUPR0TracerFireProbe },
195	{ "SUPR0TracerRegisterDrv", (void *)SUPR0TracerRegisterDrv },
196	{ "SUPR0TracerRegisterImpl", (void *)SUPR0TracerRegisterImpl },
197	{ "SUPR0TracerRegisterModule", (void *)SUPR0TracerRegisterModule },
198	{ "SUPR0TracerUmodProbeFire", (void *)SUPR0TracerUmodProbeFire },
199	{ "SUPR0UnlockMem", (void *)SUPR0UnlockMem },
200	{ "SUPSemEventClose", (void *)SUPSemEventClose },
201	{ "SUPSemEventCreate", (void *)SUPSemEventCreate },
202	{ "SUPSemEventGetResolution", (void *)SUPSemEventGetResolution },
203	{ "SUPSemEventMultiClose", (void *)SUPSemEventMultiClose },
204	{ "SUPSemEventMultiCreate", (void *)SUPSemEventMultiCreate },
205	{ "SUPSemEventMultiGetResolution", (void *)SUPSemEventMultiGetResolution },
206	{ "SUPSemEventMultiReset", (void *)SUPSemEventMultiReset },
207	{ "SUPSemEventMultiSignal", (void *)SUPSemEventMultiSignal },
208	{ "SUPSemEventMultiWait", (void *)SUPSemEventMultiWait },
209	{ "SUPSemEventMultiWaitNoResume", (void *)SUPSemEventMultiWaitNoResume },
210	{ "SUPSemEventMultiWaitNsAbsIntr", (void *)SUPSemEventMultiWaitNsAbsIntr },
211	{ "SUPSemEventMultiWaitNsRelIntr", (void *)SUPSemEventMultiWaitNsRelIntr },
212	{ "SUPSemEventSignal", (void *)SUPSemEventSignal },
213	{ "SUPSemEventWait", (void *)SUPSemEventWait },
214	{ "SUPSemEventWaitNoResume", (void *)SUPSemEventWaitNoResume },
215	{ "SUPSemEventWaitNsAbsIntr", (void *)SUPSemEventWaitNsAbsIntr },
216	{ "SUPSemEventWaitNsRelIntr", (void *)SUPSemEventWaitNsRelIntr },
217
218	{ "RTAssertAreQuiet", (void *)RTAssertAreQuiet },
219	{ "RTAssertMayPanic", (void *)RTAssertMayPanic },
220	{ "RTAssertMsg1", (void *)RTAssertMsg1 },
221	{ "RTAssertMsg2AddV", (void *)RTAssertMsg2AddV },
222	{ "RTAssertMsg2V", (void *)RTAssertMsg2V },
223	{ "RTAssertSetMayPanic", (void *)RTAssertSetMayPanic },
224	{ "RTAssertSetQuiet", (void *)RTAssertSetQuiet },
225	{ "RTCrc32", (void *)RTCrc32 },
226	{ "RTCrc32Finish", (void *)RTCrc32Finish },
227	{ "RTCrc32Process", (void *)RTCrc32Process },
228	{ "RTCrc32Start", (void *)RTCrc32Start },
229	{ "RTErrConvertFromErrno", (void *)RTErrConvertFromErrno },
230	{ "RTErrConvertToErrno", (void *)RTErrConvertToErrno },
231	{ "RTHandleTableAllocWithCtx", (void *)RTHandleTableAllocWithCtx },
232	{ "RTHandleTableCreate", (void *)RTHandleTableCreate },
233	{ "RTHandleTableCreateEx", (void *)RTHandleTableCreateEx },
234	{ "RTHandleTableDestroy", (void *)RTHandleTableDestroy },
235	{ "RTHandleTableFreeWithCtx", (void *)RTHandleTableFreeWithCtx },
236	{ "RTHandleTableLookupWithCtx", (void *)RTHandleTableLookupWithCtx },
237	{ "RTLogDefaultInstance", (void *)RTLogDefaultInstance },
238	{ "RTLogGetDefaultInstance", (void *)RTLogGetDefaultInstance },
239	{ "RTLogLoggerExV", (void *)RTLogLoggerExV },
240	{ "RTLogPrintfV", (void *)RTLogPrintfV },
241	{ "RTLogRelDefaultInstance", (void *)RTLogRelDefaultInstance },
242	{ "RTLogSetDefaultInstanceThread", (void *)RTLogSetDefaultInstanceThread },
243	{ "RTMemAllocExTag", (void *)RTMemAllocExTag },
244	{ "RTMemAllocTag", (void *)RTMemAllocTag },
245	{ "RTMemAllocVarTag", (void *)RTMemAllocVarTag },
246	{ "RTMemAllocZTag", (void *)RTMemAllocZTag },
247	{ "RTMemAllocZVarTag", (void *)RTMemAllocZVarTag },
248	{ "RTMemDupExTag", (void *)RTMemDupExTag },
249	{ "RTMemDupTag", (void *)RTMemDupTag },
250	{ "RTMemFree", (void *)RTMemFree },
251	{ "RTMemFreeEx", (void *)RTMemFreeEx },
252	{ "RTMemReallocTag", (void *)RTMemReallocTag },
253	{ "RTMpCpuId", (void *)RTMpCpuId },
254	{ "RTMpCpuIdFromSetIndex", (void *)RTMpCpuIdFromSetIndex },
255	{ "RTMpCpuIdToSetIndex", (void *)RTMpCpuIdToSetIndex },
256	{ "RTMpGetArraySize", (void *)RTMpGetArraySize },
257	{ "RTMpGetCount", (void *)RTMpGetCount },
258	{ "RTMpGetMaxCpuId", (void *)RTMpGetMaxCpuId },
259	{ "RTMpGetOnlineCount", (void *)RTMpGetOnlineCount },
260	{ "RTMpGetOnlineSet", (void *)RTMpGetOnlineSet },
261	{ "RTMpGetSet", (void *)RTMpGetSet },
262	{ "RTMpIsCpuOnline", (void *)RTMpIsCpuOnline },
263	{ "RTMpIsCpuPossible", (void *)RTMpIsCpuPossible },
264	{ "RTMpIsCpuWorkPending", (void *)RTMpIsCpuWorkPending },
265	{ "RTMpNotificationDeregister", (void *)RTMpNotificationDeregister },
266	{ "RTMpNotificationRegister", (void *)RTMpNotificationRegister },
267	{ "RTMpOnAll", (void *)RTMpOnAll },
268	{ "RTMpOnOthers", (void *)RTMpOnOthers },
269	{ "RTMpOnSpecific", (void *)RTMpOnSpecific },
270	{ "RTMpPokeCpu", (void *)RTMpPokeCpu },
271	{ "RTNetIPv4AddDataChecksum", (void *)RTNetIPv4AddDataChecksum },
272	{ "RTNetIPv4AddTCPChecksum", (void *)RTNetIPv4AddTCPChecksum },
273	{ "RTNetIPv4AddUDPChecksum", (void *)RTNetIPv4AddUDPChecksum },
274	{ "RTNetIPv4FinalizeChecksum", (void *)RTNetIPv4FinalizeChecksum },
275	{ "RTNetIPv4HdrChecksum", (void *)RTNetIPv4HdrChecksum },
276	{ "RTNetIPv4IsDHCPValid", (void *)RTNetIPv4IsDHCPValid },
277	{ "RTNetIPv4IsHdrValid", (void *)RTNetIPv4IsHdrValid },
278	{ "RTNetIPv4IsTCPSizeValid", (void *)RTNetIPv4IsTCPSizeValid },
279	{ "RTNetIPv4IsTCPValid", (void *)RTNetIPv4IsTCPValid },
280	{ "RTNetIPv4IsUDPSizeValid", (void *)RTNetIPv4IsUDPSizeValid },
281	{ "RTNetIPv4IsUDPValid", (void *)RTNetIPv4IsUDPValid },
282	{ "RTNetIPv4PseudoChecksum", (void *)RTNetIPv4PseudoChecksum },
283	{ "RTNetIPv4PseudoChecksumBits", (void *)RTNetIPv4PseudoChecksumBits },
284	{ "RTNetIPv4TCPChecksum", (void *)RTNetIPv4TCPChecksum },
285	{ "RTNetIPv4UDPChecksum", (void *)RTNetIPv4UDPChecksum },
286	{ "RTNetIPv6PseudoChecksum", (void *)RTNetIPv6PseudoChecksum },
287	{ "RTNetIPv6PseudoChecksumBits", (void *)RTNetIPv6PseudoChecksumBits },
288	{ "RTNetIPv6PseudoChecksumEx", (void *)RTNetIPv6PseudoChecksumEx },
289	{ "RTNetTCPChecksum", (void *)RTNetTCPChecksum },
290	{ "RTNetUDPChecksum", (void *)RTNetUDPChecksum },
291	{ "RTPowerNotificationDeregister", (void *)RTPowerNotificationDeregister },
292	{ "RTPowerNotificationRegister", (void *)RTPowerNotificationRegister },
293	{ "RTProcSelf", (void *)RTProcSelf },
294	{ "RTR0AssertPanicSystem", (void *)RTR0AssertPanicSystem },
295	{ "RTR0MemAreKrnlAndUsrDifferent", (void *)RTR0MemAreKrnlAndUsrDifferent },
296	{ "RTR0MemKernelIsValidAddr", (void *)RTR0MemKernelIsValidAddr },
297	{ "RTR0MemKernelCopyFrom", (void *)RTR0MemKernelCopyFrom },
298	{ "RTR0MemKernelCopyTo", (void *)RTR0MemKernelCopyTo },
299	{ "RTR0MemObjAddress", (void *)RTR0MemObjAddress },
300	{ "RTR0MemObjAddressR3", (void *)RTR0MemObjAddressR3 },
301	{ "RTR0MemObjAllocContTag", (void *)RTR0MemObjAllocContTag },
302	{ "RTR0MemObjAllocLowTag", (void *)RTR0MemObjAllocLowTag },
303	{ "RTR0MemObjAllocPageTag", (void *)RTR0MemObjAllocPageTag },
304	{ "RTR0MemObjAllocPhysExTag", (void *)RTR0MemObjAllocPhysExTag },
305	{ "RTR0MemObjAllocPhysNCTag", (void *)RTR0MemObjAllocPhysNCTag },
306	{ "RTR0MemObjAllocPhysTag", (void *)RTR0MemObjAllocPhysTag },
307	{ "RTR0MemObjEnterPhysTag", (void *)RTR0MemObjEnterPhysTag },
308	{ "RTR0MemObjFree", (void *)RTR0MemObjFree },
309	{ "RTR0MemObjGetPagePhysAddr", (void *)RTR0MemObjGetPagePhysAddr },
310	{ "RTR0MemObjIsMapping", (void *)RTR0MemObjIsMapping },
311	{ "RTR0MemObjLockUserTag", (void *)RTR0MemObjLockUserTag },
312	{ "RTR0MemObjMapKernelExTag", (void *)RTR0MemObjMapKernelExTag },
313	{ "RTR0MemObjMapKernelTag", (void *)RTR0MemObjMapKernelTag },
314	{ "RTR0MemObjMapUserTag", (void *)RTR0MemObjMapUserTag },
315	{ "RTR0MemObjProtect", (void *)RTR0MemObjProtect },
316	{ "RTR0MemObjSize", (void *)RTR0MemObjSize },
317	{ "RTR0MemUserCopyFrom", (void *)RTR0MemUserCopyFrom },
318	{ "RTR0MemUserCopyTo", (void *)RTR0MemUserCopyTo },
319	{ "RTR0MemUserIsValidAddr", (void *)RTR0MemUserIsValidAddr },
320	{ "RTR0ProcHandleSelf", (void *)RTR0ProcHandleSelf },
321	{ "RTSemEventCreate", (void *)RTSemEventCreate },
322	{ "RTSemEventDestroy", (void *)RTSemEventDestroy },
323	{ "RTSemEventGetResolution", (void *)RTSemEventGetResolution },
324	{ "RTSemEventMultiCreate", (void *)RTSemEventMultiCreate },
325	{ "RTSemEventMultiDestroy", (void *)RTSemEventMultiDestroy },
326	{ "RTSemEventMultiGetResolution", (void *)RTSemEventMultiGetResolution },
327	{ "RTSemEventMultiReset", (void *)RTSemEventMultiReset },
328	{ "RTSemEventMultiSignal", (void *)RTSemEventMultiSignal },
329	{ "RTSemEventMultiWait", (void *)RTSemEventMultiWait },
330	{ "RTSemEventMultiWaitEx", (void *)RTSemEventMultiWaitEx },
331	{ "RTSemEventMultiWaitExDebug", (void *)RTSemEventMultiWaitExDebug },
332	{ "RTSemEventMultiWaitNoResume", (void *)RTSemEventMultiWaitNoResume },
333	{ "RTSemEventSignal", (void *)RTSemEventSignal },
334	{ "RTSemEventWait", (void *)RTSemEventWait },
335	{ "RTSemEventWaitEx", (void *)RTSemEventWaitEx },
336	{ "RTSemEventWaitExDebug", (void *)RTSemEventWaitExDebug },
337	{ "RTSemEventWaitNoResume", (void *)RTSemEventWaitNoResume },
338	{ "RTSemFastMutexCreate", (void *)RTSemFastMutexCreate },
339	{ "RTSemFastMutexDestroy", (void *)RTSemFastMutexDestroy },
340	{ "RTSemFastMutexRelease", (void *)RTSemFastMutexRelease },
341	{ "RTSemFastMutexRequest", (void *)RTSemFastMutexRequest },
342	{ "RTSemMutexCreate", (void *)RTSemMutexCreate },
343	{ "RTSemMutexDestroy", (void *)RTSemMutexDestroy },
344	{ "RTSemMutexRelease", (void *)RTSemMutexRelease },
345	{ "RTSemMutexRequest", (void *)RTSemMutexRequest },
346	{ "RTSemMutexRequestDebug", (void *)RTSemMutexRequestDebug },
347	{ "RTSemMutexRequestNoResume", (void *)RTSemMutexRequestNoResume },
348	{ "RTSemMutexRequestNoResumeDebug", (void *)RTSemMutexRequestNoResumeDebug },
349	{ "RTSpinlockAcquire", (void *)RTSpinlockAcquire },
350	{ "RTSpinlockCreate", (void *)RTSpinlockCreate },
351	{ "RTSpinlockDestroy", (void *)RTSpinlockDestroy },
352	{ "RTSpinlockRelease", (void *)RTSpinlockRelease },
353	{ "RTSpinlockReleaseNoInts", (void *)RTSpinlockReleaseNoInts },
354	{ "RTStrCopy", (void *)RTStrCopy },
355	{ "RTStrDupTag", (void *)RTStrDupTag },
356	{ "RTStrFormat", (void *)RTStrFormat },
357	{ "RTStrFormatNumber", (void *)RTStrFormatNumber },
358	{ "RTStrFormatTypeDeregister", (void *)RTStrFormatTypeDeregister },
359	{ "RTStrFormatTypeRegister", (void *)RTStrFormatTypeRegister },
360	{ "RTStrFormatTypeSetUser", (void *)RTStrFormatTypeSetUser },
361	{ "RTStrFormatV", (void *)RTStrFormatV },
362	{ "RTStrFree", (void *)RTStrFree },
363	{ "RTStrNCmp", (void *)RTStrNCmp },
364	{ "RTStrPrintf", (void *)RTStrPrintf },
365	{ "RTStrPrintfEx", (void *)RTStrPrintfEx },
366	{ "RTStrPrintfExV", (void *)RTStrPrintfExV },
367	{ "RTStrPrintfV", (void *)RTStrPrintfV },
368	{ "RTThreadCreate", (void *)RTThreadCreate },
369	{ "RTThreadCtxHooksCreate", (void *)RTThreadCtxHooksCreate },
370	{ "RTThreadCtxHooksDeregister", (void *)RTThreadCtxHooksDeregister },
371	{ "RTThreadCtxHooksDestroy", (void *)RTThreadCtxHooksDestroy },
372	{ "RTThreadCtxHooksRegister", (void *)RTThreadCtxHooksRegister },
373	{ "RTThreadGetName", (void *)RTThreadGetName },
374	{ "RTThreadGetNative", (void *)RTThreadGetNative },
375	{ "RTThreadGetType", (void *)RTThreadGetType },
376	{ "RTThreadIsInInterrupt", (void *)RTThreadIsInInterrupt },
377	{ "RTThreadNativeSelf", (void *)RTThreadNativeSelf },
378	{ "RTThreadPreemptDisable", (void *)RTThreadPreemptDisable },
379	{ "RTThreadPreemptIsEnabled", (void *)RTThreadPreemptIsEnabled },
380	{ "RTThreadPreemptIsPending", (void *)RTThreadPreemptIsPending },
381	{ "RTThreadPreemptIsPendingTrusty", (void *)RTThreadPreemptIsPendingTrusty },
382	{ "RTThreadPreemptIsPossible", (void *)RTThreadPreemptIsPossible },
383	{ "RTThreadPreemptRestore", (void *)RTThreadPreemptRestore },
384	{ "RTThreadSelf", (void *)RTThreadSelf },
385	{ "RTThreadSelfName", (void *)RTThreadSelfName },
386	{ "RTThreadSleep", (void *)RTThreadSleep },
387	{ "RTThreadUserReset", (void *)RTThreadUserReset },
388	{ "RTThreadUserSignal", (void *)RTThreadUserSignal },
389	{ "RTThreadUserWait", (void *)RTThreadUserWait },
390	{ "RTThreadUserWaitNoResume", (void *)RTThreadUserWaitNoResume },
391	{ "RTThreadWait", (void *)RTThreadWait },
392	{ "RTThreadWaitNoResume", (void *)RTThreadWaitNoResume },
393	{ "RTThreadYield", (void *)RTThreadYield },
394	{ "RTTimeMilliTS", (void *)RTTimeMilliTS },
395	{ "RTTimeNanoTS", (void *)RTTimeNanoTS },
396	{ "RTTimeNow", (void *)RTTimeNow },
397	{ "RTTimerCanDoHighResolution", (void *)RTTimerCanDoHighResolution },
398	{ "RTTimerChangeInterval", (void *)RTTimerChangeInterval },
399	{ "RTTimerCreate", (void *)RTTimerCreate },
400	{ "RTTimerCreateEx", (void *)RTTimerCreateEx },
401	{ "RTTimerDestroy", (void *)RTTimerDestroy },
402	{ "RTTimerGetSystemGranularity", (void *)RTTimerGetSystemGranularity },
403	{ "RTTimerReleaseSystemGranularity", (void *)RTTimerReleaseSystemGranularity },
404	{ "RTTimerRequestSystemGranularity", (void *)RTTimerRequestSystemGranularity },
405	{ "RTTimerStart", (void *)RTTimerStart },
406	{ "RTTimerStop", (void *)RTTimerStop },
407	{ "RTTimeSystemMilliTS", (void *)RTTimeSystemMilliTS },
408	{ "RTTimeSystemNanoTS", (void *)RTTimeSystemNanoTS },
409	{ "RTUuidCompare", (void *)RTUuidCompare },
410	{ "RTUuidCompareStr", (void *)RTUuidCompareStr },
411	{ "RTUuidFromStr", (void *)RTUuidFromStr },
412	/* SED: END */
413	};
414
415	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
416	/**
417	* Drag in the rest of IRPT since we share it with the
418	* rest of the kernel modules on darwin.
419	*/
420	PFNRT g_apfnVBoxDrvIPRTDeps[] =
421	{
422	/* VBoxNetAdp */
423	(PFNRT)RTRandBytes,
424	/* VBoxUSB */
425	(PFNRT)RTPathStripFilename,
426	NULL
427	};
428	#endif /* RT_OS_DARWIN \|\| RT_OS_SOLARIS \|\| RT_OS_SOLARIS */
429
430
431	/**
432	* Initializes the device extentsion structure.
433	*
434	* @returns IPRT status code.
435	* @param pDevExt The device extension to initialize.
436	* @param cbSession The size of the session structure. The size of
437	* SUPDRVSESSION may be smaller when SUPDRV_AGNOSTIC is
438	* defined because we're skipping the OS specific members
439	* then.
440	*/
441	int VBOXCALL supdrvInitDevExt(PSUPDRVDEVEXT pDevExt, size_t cbSession)
442	{
443	int rc;
444
445	#ifdef SUPDRV_WITH_RELEASE_LOGGER
446	/*
447	* Create the release log.
448	*/
449	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
450	PRTLOGGER pRelLogger;
451	rc = RTLogCreate(&pRelLogger, 0 /* fFlags */, "all",
452	"VBOX_RELEASE_LOG", RT_ELEMENTS(s_apszGroups), s_apszGroups, RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER, NULL);
453	if (RT_SUCCESS(rc))
454	RTLogRelSetDefaultInstance(pRelLogger);
455	/** @todo Add native hook for getting logger config parameters and setting
456	* them. On linux we should use the module parameter stuff... */
457	#endif
458
459	/*
460	* Initialize it.
461	*/
462	memset(pDevExt, 0, sizeof(*pDevExt));
463	rc = RTSpinlockCreate(&pDevExt->Spinlock, RTSPINLOCK_FLAGS_INTERRUPT_SAFE, "SUPDrvDevExt");
464	if (RT_SUCCESS(rc))
465	{
466	rc = RTSpinlockCreate(&pDevExt->hGipSpinlock, RTSPINLOCK_FLAGS_INTERRUPT_SAFE, "SUPDrvGip");
467	if (RT_SUCCESS(rc))
468	{
469	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
470	rc = RTSemMutexCreate(&pDevExt->mtxLdr);
471	#else
472	rc = RTSemFastMutexCreate(&pDevExt->mtxLdr);
473	#endif
474	if (RT_SUCCESS(rc))
475	{
476	rc = RTSemFastMutexCreate(&pDevExt->mtxComponentFactory);
477	if (RT_SUCCESS(rc))
478	{
479	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
480	rc = RTSemMutexCreate(&pDevExt->mtxGip);
481	#else
482	rc = RTSemFastMutexCreate(&pDevExt->mtxGip);
483	#endif
484	if (RT_SUCCESS(rc))
485	{
486	rc = supdrvGipCreate(pDevExt);
487	if (RT_SUCCESS(rc))
488	{
489	rc = supdrvTracerInit(pDevExt);
490	if (RT_SUCCESS(rc))
491	{
492	pDevExt->pLdrInitImage = NULL;
493	pDevExt->hLdrInitThread = NIL_RTNATIVETHREAD;
494	pDevExt->u32Cookie = BIRD; /** @todo make this random? */
495	pDevExt->cbSession = (uint32_t)cbSession;
496
497	/*
498	* Fixup the absolute symbols.
499	*
500	* Because of the table indexing assumptions we'll have a little #ifdef orgy
501	* here rather than distributing this to OS specific files. At least for now.
502	*/
503	#ifdef RT_OS_DARWIN
504	# if ARCH_BITS == 32
505	if (SUPR0GetPagingMode() >= SUPPAGINGMODE_AMD64)
506	{
507	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
508	g_aFunctions[1].pfn = (void )0x80; / SUPR0Abs64bitKernelCS - KERNEL64_CS, seg.h */
509	g_aFunctions[2].pfn = (void )0x88; / SUPR0Abs64bitKernelSS - KERNEL64_SS, seg.h */
510	g_aFunctions[3].pfn = (void )0x88; / SUPR0Abs64bitKernelDS - KERNEL64_SS, seg.h */
511	}
512	else
513	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
514	g_aFunctions[4].pfn = (void )0x08; / SUPR0AbsKernelCS - KERNEL_CS, seg.h */
515	g_aFunctions[5].pfn = (void )0x10; / SUPR0AbsKernelSS - KERNEL_DS, seg.h */
516	g_aFunctions[6].pfn = (void )0x10; / SUPR0AbsKernelDS - KERNEL_DS, seg.h */
517	g_aFunctions[7].pfn = (void )0x10; / SUPR0AbsKernelES - KERNEL_DS, seg.h */
518	g_aFunctions[8].pfn = (void )0x10; / SUPR0AbsKernelFS - KERNEL_DS, seg.h */
519	g_aFunctions[9].pfn = (void )0x48; / SUPR0AbsKernelGS - CPU_DATA_GS, seg.h */
520	# else /* 64-bit darwin: */
521	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
522	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
523	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
524	g_aFunctions[3].pfn = (void )0; / SUPR0Abs64bitKernelDS */
525	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
526	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
527	g_aFunctions[6].pfn = (void )0; / SUPR0AbsKernelDS */
528	g_aFunctions[7].pfn = (void )0; / SUPR0AbsKernelES */
529	g_aFunctions[8].pfn = (void )0; / SUPR0AbsKernelFS */
530	g_aFunctions[9].pfn = (void )0; / SUPR0AbsKernelGS */
531
532	# endif
533	#else /* !RT_OS_DARWIN */
534	# if ARCH_BITS == 64
535	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
536	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
537	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
538	g_aFunctions[3].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0Abs64bitKernelDS */
539	# else
540	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
541	# endif
542	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
543	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
544	g_aFunctions[6].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0AbsKernelDS */
545	g_aFunctions[7].pfn = (void )(uintptr_t)ASMGetES(); / SUPR0AbsKernelES */
546	g_aFunctions[8].pfn = (void )(uintptr_t)ASMGetFS(); / SUPR0AbsKernelFS */
547	g_aFunctions[9].pfn = (void )(uintptr_t)ASMGetGS(); / SUPR0AbsKernelGS */
548	#endif /* !RT_OS_DARWIN */
549	return VINF_SUCCESS;
550	}
551
552	supdrvGipDestroy(pDevExt);
553	}
554
555	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
556	RTSemMutexDestroy(pDevExt->mtxGip);
557	pDevExt->mtxGip = NIL_RTSEMMUTEX;
558	#else
559	RTSemFastMutexDestroy(pDevExt->mtxGip);
560	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
561	#endif
562	}
563	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
564	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
565	}
566	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
567	RTSemMutexDestroy(pDevExt->mtxLdr);
568	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
569	#else
570	RTSemFastMutexDestroy(pDevExt->mtxLdr);
571	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
572	#endif
573	}
574	RTSpinlockDestroy(pDevExt->hGipSpinlock);
575	pDevExt->hGipSpinlock = NIL_RTSPINLOCK;
576	}
577	RTSpinlockDestroy(pDevExt->Spinlock);
578	pDevExt->Spinlock = NIL_RTSPINLOCK;
579	}
580	#ifdef SUPDRV_WITH_RELEASE_LOGGER
581	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
582	RTLogDestroy(RTLogSetDefaultInstance(NULL));
583	#endif
584
585	return rc;
586	}
587
588
589	/**
590	* Delete the device extension (e.g. cleanup members).
591	*
592	* @param pDevExt The device extension to delete.
593	*/
594	void VBOXCALL supdrvDeleteDevExt(PSUPDRVDEVEXT pDevExt)
595	{
596	PSUPDRVOBJ pObj;
597	PSUPDRVUSAGE pUsage;
598
599	/*
600	* Kill mutexes and spinlocks.
601	*/
602	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
603	RTSemMutexDestroy(pDevExt->mtxGip);
604	pDevExt->mtxGip = NIL_RTSEMMUTEX;
605	#else
606	RTSemFastMutexDestroy(pDevExt->mtxGip);
607	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
608	#endif
609	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
610	RTSemMutexDestroy(pDevExt->mtxLdr);
611	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
612	#else
613	RTSemFastMutexDestroy(pDevExt->mtxLdr);
614	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
615	#endif
616	RTSpinlockDestroy(pDevExt->Spinlock);
617	pDevExt->Spinlock = NIL_RTSPINLOCK;
618	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
619	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
620
621	/*
622	* Free lists.
623	*/
624	/* objects. */
625	pObj = pDevExt->pObjs;
626	Assert(!pObj); /* (can trigger on forced unloads) */
627	pDevExt->pObjs = NULL;
628	while (pObj)
629	{
630	void *pvFree = pObj;
631	pObj = pObj->pNext;
632	RTMemFree(pvFree);
633	}
634
635	/* usage records. */
636	pUsage = pDevExt->pUsageFree;
637	pDevExt->pUsageFree = NULL;
638	while (pUsage)
639	{
640	void *pvFree = pUsage;
641	pUsage = pUsage->pNext;
642	RTMemFree(pvFree);
643	}
644
645	/* kill the GIP. */
646	supdrvGipDestroy(pDevExt);
647	RTSpinlockDestroy(pDevExt->hGipSpinlock);
648	pDevExt->hGipSpinlock = NIL_RTSPINLOCK;
649
650	supdrvTracerTerm(pDevExt);
651
652	#ifdef SUPDRV_WITH_RELEASE_LOGGER
653	/* destroy the loggers. */
654	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
655	RTLogDestroy(RTLogSetDefaultInstance(NULL));
656	#endif
657	}
658
659
660	/**
661	* Create session.
662	*
663	* @returns IPRT status code.
664	* @param pDevExt Device extension.
665	* @param fUser Flag indicating whether this is a user or kernel
666	* session.
667	* @param fUnrestricted Unrestricted access (system) or restricted access
668	* (user)?
669	* @param ppSession Where to store the pointer to the session data.
670	*/
671	int VBOXCALL supdrvCreateSession(PSUPDRVDEVEXT pDevExt, bool fUser, bool fUnrestricted, PSUPDRVSESSION *ppSession)
672	{
673	int rc;
674	PSUPDRVSESSION pSession;
675
676	if (!SUP_IS_DEVEXT_VALID(pDevExt))
677	return VERR_INVALID_PARAMETER;
678
679	/*
680	* Allocate memory for the session data.
681	*/
682	pSession = *ppSession = (PSUPDRVSESSION)RTMemAllocZ(pDevExt->cbSession);
683	if (pSession)
684	{
685	/* Initialize session data. */
686	rc = RTSpinlockCreate(&pSession->Spinlock, RTSPINLOCK_FLAGS_INTERRUPT_UNSAFE, "SUPDrvSession");
687	if (!rc)
688	{
689	rc = RTHandleTableCreateEx(&pSession->hHandleTable,
690	RTHANDLETABLE_FLAGS_LOCKED_IRQ_SAFE \| RTHANDLETABLE_FLAGS_CONTEXT,
691	1 /uBase/, 32768 /cMax/, supdrvSessionObjHandleRetain, pSession);
692	if (RT_SUCCESS(rc))
693	{
694	Assert(pSession->Spinlock != NIL_RTSPINLOCK);
695	pSession->pDevExt = pDevExt;
696	pSession->u32Cookie = BIRD_INV;
697	pSession->fUnrestricted = fUnrestricted;
698	/*pSession->pLdrUsage = NULL;
699	pSession->pVM = NULL;
700	pSession->pUsage = NULL;
701	pSession->pGip = NULL;
702	pSession->fGipReferenced = false;
703	pSession->Bundle.cUsed = 0; */
704	pSession->Uid = NIL_RTUID;
705	pSession->Gid = NIL_RTGID;
706	if (fUser)
707	{
708	pSession->Process = RTProcSelf();
709	pSession->R0Process = RTR0ProcHandleSelf();
710	}
711	else
712	{
713	pSession->Process = NIL_RTPROCESS;
714	pSession->R0Process = NIL_RTR0PROCESS;
715	}
716	/pSession->uTracerData = 0;/
717	pSession->hTracerCaller = NIL_RTNATIVETHREAD;
718	RTListInit(&pSession->TpProviders);
719	/pSession->cTpProviders = 0;/
720	/pSession->cTpProbesFiring = 0;/
721	RTListInit(&pSession->TpUmods);
722	/RT_ZERO(pSession->apTpLookupTable);/
723
724	VBOXDRV_SESSION_CREATE(pSession, fUser);
725	LogFlow(("Created session %p initial cookie=%#x\n", pSession, pSession->u32Cookie));
726	return VINF_SUCCESS;
727	}
728
729	RTSpinlockDestroy(pSession->Spinlock);
730	}
731	RTMemFree(pSession);
732	*ppSession = NULL;
733	Log(("Failed to create spinlock, rc=%d!\n", rc));
734	}
735	else
736	rc = VERR_NO_MEMORY;
737
738	return rc;
739	}
740
741
742	/**
743	* Shared code for cleaning up a session.
744	*
745	* @param pDevExt Device extension.
746	* @param pSession Session data.
747	* This data will be freed by this routine.
748	*/
749	void VBOXCALL supdrvCloseSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
750	{
751	VBOXDRV_SESSION_CLOSE(pSession);
752
753	/*
754	* Cleanup the session first.
755	*/
756	supdrvCleanupSession(pDevExt, pSession);
757
758	/*
759	* Free the rest of the session stuff.
760	*/
761	RTSpinlockDestroy(pSession->Spinlock);
762	pSession->Spinlock = NIL_RTSPINLOCK;
763	pSession->pDevExt = NULL;
764	RTMemFree(pSession);
765	LogFlow(("supdrvCloseSession: returns\n"));
766	}
767
768
769	/**
770	* Shared code for cleaning up a session (but not quite freeing it).
771	*
772	* This is primarily intended for MAC OS X where we have to clean up the memory
773	* stuff before the file handle is closed.
774	*
775	* @param pDevExt Device extension.
776	* @param pSession Session data.
777	* This data will be freed by this routine.
778	*/
779	void VBOXCALL supdrvCleanupSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
780	{
781	int rc;
782	PSUPDRVBUNDLE pBundle;
783	LogFlow(("supdrvCleanupSession: pSession=%p\n", pSession));
784
785	/*
786	* Remove logger instances related to this session.
787	*/
788	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pSession);
789
790	/*
791	* Destroy the handle table.
792	*/
793	rc = RTHandleTableDestroy(pSession->hHandleTable, supdrvSessionObjHandleDelete, pSession);
794	AssertRC(rc);
795	pSession->hHandleTable = NIL_RTHANDLETABLE;
796
797	/*
798	* Release object references made in this session.
799	* In theory there should be noone racing us in this session.
800	*/
801	Log2(("release objects - start\n"));
802	if (pSession->pUsage)
803	{
804	PSUPDRVUSAGE pUsage;
805	RTSpinlockAcquire(pDevExt->Spinlock);
806
807	while ((pUsage = pSession->pUsage) != NULL)
808	{
809	PSUPDRVOBJ pObj = pUsage->pObj;
810	pSession->pUsage = pUsage->pNext;
811
812	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
813	if (pUsage->cUsage < pObj->cUsage)
814	{
815	pObj->cUsage -= pUsage->cUsage;
816	RTSpinlockRelease(pDevExt->Spinlock);
817	}
818	else
819	{
820	/* Destroy the object and free the record. */
821	if (pDevExt->pObjs == pObj)
822	pDevExt->pObjs = pObj->pNext;
823	else
824	{
825	PSUPDRVOBJ pObjPrev;
826	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
827	if (pObjPrev->pNext == pObj)
828	{
829	pObjPrev->pNext = pObj->pNext;
830	break;
831	}
832	Assert(pObjPrev);
833	}
834	RTSpinlockRelease(pDevExt->Spinlock);
835
836	Log(("supdrvCleanupSession: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
837	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
838	if (pObj->pfnDestructor)
839	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
840	RTMemFree(pObj);
841	}
842
843	/* free it and continue. */
844	RTMemFree(pUsage);
845
846	RTSpinlockAcquire(pDevExt->Spinlock);
847	}
848
849	RTSpinlockRelease(pDevExt->Spinlock);
850	AssertMsg(!pSession->pUsage, ("Some buster reregistered an object during desturction!\n"));
851	}
852	Log2(("release objects - done\n"));
853
854	/*
855	* Do tracer cleanups related to this session.
856	*/
857	Log2(("release tracer stuff - start\n"));
858	supdrvTracerCleanupSession(pDevExt, pSession);
859	Log2(("release tracer stuff - end\n"));
860
861	/*
862	* Release memory allocated in the session.
863	*
864	* We do not serialize this as we assume that the application will
865	* not allocated memory while closing the file handle object.
866	*/
867	Log2(("freeing memory:\n"));
868	pBundle = &pSession->Bundle;
869	while (pBundle)
870	{
871	PSUPDRVBUNDLE pToFree;
872	unsigned i;
873
874	/*
875	* Check and unlock all entries in the bundle.
876	*/
877	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
878	{
879	if (pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ)
880	{
881	Log2(("eType=%d pvR0=%p pvR3=%p cb=%ld\n", pBundle->aMem[i].eType, RTR0MemObjAddress(pBundle->aMem[i].MemObj),
882	(void *)RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3), (long)RTR0MemObjSize(pBundle->aMem[i].MemObj)));
883	if (pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ)
884	{
885	rc = RTR0MemObjFree(pBundle->aMem[i].MapObjR3, false);
886	AssertRC(rc); /** @todo figure out how to handle this. */
887	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
888	}
889	rc = RTR0MemObjFree(pBundle->aMem[i].MemObj, true /* fFreeMappings */);
890	AssertRC(rc); /** @todo figure out how to handle this. */
891	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
892	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
893	}
894	}
895
896	/*
897	* Advance and free previous bundle.
898	*/
899	pToFree = pBundle;
900	pBundle = pBundle->pNext;
901
902	pToFree->pNext = NULL;
903	pToFree->cUsed = 0;
904	if (pToFree != &pSession->Bundle)
905	RTMemFree(pToFree);
906	}
907	Log2(("freeing memory - done\n"));
908
909	/*
910	* Deregister component factories.
911	*/
912	RTSemFastMutexRequest(pDevExt->mtxComponentFactory);
913	Log2(("deregistering component factories:\n"));
914	if (pDevExt->pComponentFactoryHead)
915	{
916	PSUPDRVFACTORYREG pPrev = NULL;
917	PSUPDRVFACTORYREG pCur = pDevExt->pComponentFactoryHead;
918	while (pCur)
919	{
920	if (pCur->pSession == pSession)
921	{
922	/* unlink it */
923	PSUPDRVFACTORYREG pNext = pCur->pNext;
924	if (pPrev)
925	pPrev->pNext = pNext;
926	else
927	pDevExt->pComponentFactoryHead = pNext;
928
929	/* free it */
930	pCur->pNext = NULL;
931	pCur->pSession = NULL;
932	pCur->pFactory = NULL;
933	RTMemFree(pCur);
934
935	/* next */
936	pCur = pNext;
937	}
938	else
939	{
940	/* next */
941	pPrev = pCur;
942	pCur = pCur->pNext;
943	}
944	}
945	}
946	RTSemFastMutexRelease(pDevExt->mtxComponentFactory);
947	Log2(("deregistering component factories - done\n"));
948
949	/*
950	* Loaded images needs to be dereferenced and possibly freed up.
951	*/
952	supdrvLdrLock(pDevExt);
953	Log2(("freeing images:\n"));
954	if (pSession->pLdrUsage)
955	{
956	PSUPDRVLDRUSAGE pUsage = pSession->pLdrUsage;
957	pSession->pLdrUsage = NULL;
958	while (pUsage)
959	{
960	void *pvFree = pUsage;
961	PSUPDRVLDRIMAGE pImage = pUsage->pImage;
962	if (pImage->cUsage > pUsage->cUsage)
963	pImage->cUsage -= pUsage->cUsage;
964	else
965	supdrvLdrFree(pDevExt, pImage);
966	pUsage->pImage = NULL;
967	pUsage = pUsage->pNext;
968	RTMemFree(pvFree);
969	}
970	}
971	supdrvLdrUnlock(pDevExt);
972	Log2(("freeing images - done\n"));
973
974	/*
975	* Unmap the GIP.
976	*/
977	Log2(("umapping GIP:\n"));
978	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
979	{
980	SUPR0GipUnmap(pSession);
981	pSession->fGipReferenced = 0;
982	}
983	Log2(("umapping GIP - done\n"));
984	}
985
986
987	/**
988	* RTHandleTableDestroy callback used by supdrvCleanupSession.
989	*
990	* @returns IPRT status code, see SUPR0ObjAddRef.
991	* @param hHandleTable The handle table handle. Ignored.
992	* @param pvObj The object pointer.
993	* @param pvCtx Context, the handle type. Ignored.
994	* @param pvUser Session pointer.
995	*/
996	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser)
997	{
998	NOREF(pvCtx);
999	NOREF(hHandleTable);
1000	return SUPR0ObjAddRefEx(pvObj, (PSUPDRVSESSION)pvUser, true /fNoBlocking/);
1001	}
1002
1003
1004	/**
1005	* RTHandleTableDestroy callback used by supdrvCleanupSession.
1006	*
1007	* @param hHandleTable The handle table handle. Ignored.
1008	* @param h The handle value. Ignored.
1009	* @param pvObj The object pointer.
1010	* @param pvCtx Context, the handle type. Ignored.
1011	* @param pvUser Session pointer.
1012	*/
1013	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser)
1014	{
1015	NOREF(pvCtx);
1016	NOREF(h);
1017	NOREF(hHandleTable);
1018	SUPR0ObjRelease(pvObj, (PSUPDRVSESSION)pvUser);
1019	}
1020
1021
1022	/**
1023	* Fast path I/O Control worker.
1024	*
1025	* @returns VBox status code that should be passed down to ring-3 unchanged.
1026	* @param uIOCtl Function number.
1027	* @param idCpu VMCPU id.
1028	* @param pDevExt Device extention.
1029	* @param pSession Session data.
1030	*/
1031	int VBOXCALL supdrvIOCtlFast(uintptr_t uIOCtl, VMCPUID idCpu, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
1032	{
1033	/*
1034	* We check the two prereqs after doing this only to allow the compiler to optimize things better.
1035	*/
1036	if (RT_LIKELY( RT_VALID_PTR(pSession)
1037	&& pSession->pVM
1038	&& pDevExt->pfnVMMR0EntryFast))
1039	{
1040	switch (uIOCtl)
1041	{
1042	case SUP_IOCTL_FAST_DO_RAW_RUN:
1043	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_RAW_RUN);
1044	break;
1045	case SUP_IOCTL_FAST_DO_HM_RUN:
1046	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_HM_RUN);
1047	break;
1048	case SUP_IOCTL_FAST_DO_NOP:
1049	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_NOP);
1050	break;
1051	default:
1052	return VERR_INTERNAL_ERROR;
1053	}
1054	return VINF_SUCCESS;
1055	}
1056	return VERR_INTERNAL_ERROR;
1057	}
1058
1059
1060	/**
1061	* Helper for supdrvIOCtl. Check if pszStr contains any character of pszChars.
1062	* We would use strpbrk here if this function would be contained in the RedHat kABI white
1063	* list, see http://www.kerneldrivers.org/RHEL5.
1064	*
1065	* @returns 1 if pszStr does contain any character of pszChars, 0 otherwise.
1066	* @param pszStr String to check
1067	* @param pszChars Character set
1068	*/
1069	static int supdrvCheckInvalidChar(const char pszStr, const char pszChars)
1070	{
1071	int chCur;
1072	while ((chCur = *pszStr++) != '\0')
1073	{
1074	int ch;
1075	const char *psz = pszChars;
1076	while ((ch = *psz++) != '\0')
1077	if (ch == chCur)
1078	return 1;
1079
1080	}
1081	return 0;
1082	}
1083
1084
1085
1086	/**
1087	* I/O Control inner worker (tracing reasons).
1088	*
1089	* @returns IPRT status code.
1090	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1091	*
1092	* @param uIOCtl Function number.
1093	* @param pDevExt Device extention.
1094	* @param pSession Session data.
1095	* @param pReqHdr The request header.
1096	*/
1097	static int supdrvIOCtlInnerUnrestricted(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1098	{
1099	/*
1100	* Validation macros
1101	*/
1102	#define REQ_CHECK_SIZES_EX(Name, cbInExpect, cbOutExpect) \
1103	do { \
1104	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect) \|\| pReqHdr->cbOut != (cbOutExpect))) \
1105	{ \
1106	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld. cbOut=%ld expected %ld.\n", \
1107	(long)pReqHdr->cbIn, (long)(cbInExpect), (long)pReqHdr->cbOut, (long)(cbOutExpect))); \
1108	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1109	} \
1110	} while (0)
1111
1112	#define REQ_CHECK_SIZES(Name) REQ_CHECK_SIZES_EX(Name, Name ## _SIZE_IN, Name ## _SIZE_OUT)
1113
1114	#define REQ_CHECK_SIZE_IN(Name, cbInExpect) \
1115	do { \
1116	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect))) \
1117	{ \
1118	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld.\n", \
1119	(long)pReqHdr->cbIn, (long)(cbInExpect))); \
1120	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1121	} \
1122	} while (0)
1123
1124	#define REQ_CHECK_SIZE_OUT(Name, cbOutExpect) \
1125	do { \
1126	if (RT_UNLIKELY(pReqHdr->cbOut != (cbOutExpect))) \
1127	{ \
1128	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbOut=%ld expected %ld.\n", \
1129	(long)pReqHdr->cbOut, (long)(cbOutExpect))); \
1130	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1131	} \
1132	} while (0)
1133
1134	#define REQ_CHECK_EXPR(Name, expr) \
1135	do { \
1136	if (RT_UNLIKELY(!(expr))) \
1137	{ \
1138	OSDBGPRINT(( #Name ": %s\n", #expr)); \
1139	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1140	} \
1141	} while (0)
1142
1143	#define REQ_CHECK_EXPR_FMT(expr, fmt) \
1144	do { \
1145	if (RT_UNLIKELY(!(expr))) \
1146	{ \
1147	OSDBGPRINT( fmt ); \
1148	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1149	} \
1150	} while (0)
1151
1152	/*
1153	* The switch.
1154	*/
1155	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
1156	{
1157	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
1158	{
1159	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
1160	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
1161	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
1162	{
1163	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
1164	pReq->Hdr.rc = VERR_INVALID_MAGIC;
1165	return 0;
1166	}
1167
1168	#if 0
1169	/*
1170	* Call out to the OS specific code and let it do permission checks on the
1171	* client process.
1172	*/
1173	if (!supdrvOSValidateClientProcess(pDevExt, pSession))
1174	{
1175	pReq->u.Out.u32Cookie = 0xffffffff;
1176	pReq->u.Out.u32SessionCookie = 0xffffffff;
1177	pReq->u.Out.u32SessionVersion = 0xffffffff;
1178	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1179	pReq->u.Out.pSession = NULL;
1180	pReq->u.Out.cFunctions = 0;
1181	pReq->Hdr.rc = VERR_PERMISSION_DENIED;
1182	return 0;
1183	}
1184	#endif
1185
1186	/*
1187	* Match the version.
1188	* The current logic is very simple, match the major interface version.
1189	*/
1190	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
1191	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
1192	{
1193	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1194	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
1195	pReq->u.Out.u32Cookie = 0xffffffff;
1196	pReq->u.Out.u32SessionCookie = 0xffffffff;
1197	pReq->u.Out.u32SessionVersion = 0xffffffff;
1198	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1199	pReq->u.Out.pSession = NULL;
1200	pReq->u.Out.cFunctions = 0;
1201	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1202	return 0;
1203	}
1204
1205	/*
1206	* Fill in return data and be gone.
1207	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
1208	* u32SessionVersion <= u32ReqVersion!
1209	*/
1210	/** @todo Somehow validate the client and negotiate a secure cookie... */
1211	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
1212	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
1213	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
1214	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1215	pReq->u.Out.pSession = pSession;
1216	pReq->u.Out.cFunctions = sizeof(g_aFunctions) / sizeof(g_aFunctions[0]);
1217	pReq->Hdr.rc = VINF_SUCCESS;
1218	return 0;
1219	}
1220
1221	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_QUERY_FUNCS(0)):
1222	{
1223	/* validate */
1224	PSUPQUERYFUNCS pReq = (PSUPQUERYFUNCS)pReqHdr;
1225	REQ_CHECK_SIZES_EX(SUP_IOCTL_QUERY_FUNCS, SUP_IOCTL_QUERY_FUNCS_SIZE_IN, SUP_IOCTL_QUERY_FUNCS_SIZE_OUT(RT_ELEMENTS(g_aFunctions)));
1226
1227	/* execute */
1228	pReq->u.Out.cFunctions = RT_ELEMENTS(g_aFunctions);
1229	memcpy(&pReq->u.Out.aFunctions[0], g_aFunctions, sizeof(g_aFunctions));
1230	pReq->Hdr.rc = VINF_SUCCESS;
1231	return 0;
1232	}
1233
1234	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_LOCK):
1235	{
1236	/* validate */
1237	PSUPPAGELOCK pReq = (PSUPPAGELOCK)pReqHdr;
1238	REQ_CHECK_SIZE_IN(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_IN);
1239	REQ_CHECK_SIZE_OUT(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_OUT(pReq->u.In.cPages));
1240	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.cPages > 0);
1241	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.pvR3 >= PAGE_SIZE);
1242
1243	/* execute */
1244	pReq->Hdr.rc = SUPR0LockMem(pSession, pReq->u.In.pvR3, pReq->u.In.cPages, &pReq->u.Out.aPages[0]);
1245	if (RT_FAILURE(pReq->Hdr.rc))
1246	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1247	return 0;
1248	}
1249
1250	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_UNLOCK):
1251	{
1252	/* validate */
1253	PSUPPAGEUNLOCK pReq = (PSUPPAGEUNLOCK)pReqHdr;
1254	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_UNLOCK);
1255
1256	/* execute */
1257	pReq->Hdr.rc = SUPR0UnlockMem(pSession, pReq->u.In.pvR3);
1258	return 0;
1259	}
1260
1261	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_ALLOC):
1262	{
1263	/* validate */
1264	PSUPCONTALLOC pReq = (PSUPCONTALLOC)pReqHdr;
1265	REQ_CHECK_SIZES(SUP_IOCTL_CONT_ALLOC);
1266
1267	/* execute */
1268	pReq->Hdr.rc = SUPR0ContAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.HCPhys);
1269	if (RT_FAILURE(pReq->Hdr.rc))
1270	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1271	return 0;
1272	}
1273
1274	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_FREE):
1275	{
1276	/* validate */
1277	PSUPCONTFREE pReq = (PSUPCONTFREE)pReqHdr;
1278	REQ_CHECK_SIZES(SUP_IOCTL_CONT_FREE);
1279
1280	/* execute */
1281	pReq->Hdr.rc = SUPR0ContFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1282	return 0;
1283	}
1284
1285	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_OPEN):
1286	{
1287	/* validate */
1288	PSUPLDROPEN pReq = (PSUPLDROPEN)pReqHdr;
1289	REQ_CHECK_SIZES(SUP_IOCTL_LDR_OPEN);
1290	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs > 0);
1291	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs < 16*_1M);
1292	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1293	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1294	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits < pReq->u.In.cbImageWithTabs);
1295	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.szName[0]);
1296	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1297	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, !supdrvCheckInvalidChar(pReq->u.In.szName, ";:()[]{}/\\\|&*%#@!~`\"'"));
1298	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szFilename, sizeof(pReq->u.In.szFilename)));
1299
1300	/* execute */
1301	pReq->Hdr.rc = supdrvIOCtl_LdrOpen(pDevExt, pSession, pReq);
1302	return 0;
1303	}
1304
1305	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOAD):
1306	{
1307	/* validate */
1308	PSUPLDRLOAD pReq = (PSUPLDRLOAD)pReqHdr;
1309	REQ_CHECK_EXPR(Name, pReq->Hdr.cbIn >= sizeof(*pReq));
1310	REQ_CHECK_SIZES_EX(SUP_IOCTL_LDR_LOAD, SUP_IOCTL_LDR_LOAD_SIZE_IN(pReq->u.In.cbImageWithTabs), SUP_IOCTL_LDR_LOAD_SIZE_OUT);
1311	REQ_CHECK_EXPR(SUP_IOCTL_LDR_LOAD, pReq->u.In.cSymbols <= 16384);
1312	REQ_CHECK_EXPR_FMT( !pReq->u.In.cSymbols
1313	\|\| ( pReq->u.In.offSymbols < pReq->u.In.cbImageWithTabs
1314	&& pReq->u.In.offSymbols + pReq->u.In.cSymbols * sizeof(SUPLDRSYM) <= pReq->u.In.cbImageWithTabs),
1315	("SUP_IOCTL_LDR_LOAD: offSymbols=%#lx cSymbols=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offSymbols,
1316	(long)pReq->u.In.cSymbols, (long)pReq->u.In.cbImageWithTabs));
1317	REQ_CHECK_EXPR_FMT( !pReq->u.In.cbStrTab
1318	\|\| ( pReq->u.In.offStrTab < pReq->u.In.cbImageWithTabs
1319	&& pReq->u.In.offStrTab + pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs
1320	&& pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs),
1321	("SUP_IOCTL_LDR_LOAD: offStrTab=%#lx cbStrTab=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offStrTab,
1322	(long)pReq->u.In.cbStrTab, (long)pReq->u.In.cbImageWithTabs));
1323
1324	if (pReq->u.In.cSymbols)
1325	{
1326	uint32_t i;
1327	PSUPLDRSYM paSyms = (PSUPLDRSYM)&pReq->u.In.abImage[pReq->u.In.offSymbols];
1328	for (i = 0; i < pReq->u.In.cSymbols; i++)
1329	{
1330	REQ_CHECK_EXPR_FMT(paSyms[i].offSymbol < pReq->u.In.cbImageWithTabs,
1331	("SUP_IOCTL_LDR_LOAD: sym #%ld: symb off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offSymbol, (long)pReq->u.In.cbImageWithTabs));
1332	REQ_CHECK_EXPR_FMT(paSyms[i].offName < pReq->u.In.cbStrTab,
1333	("SUP_IOCTL_LDR_LOAD: sym #%ld: name off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1334	REQ_CHECK_EXPR_FMT(RTStrEnd((char const *)&pReq->u.In.abImage[pReq->u.In.offStrTab + paSyms[i].offName],
1335	pReq->u.In.cbStrTab - paSyms[i].offName),
1336	("SUP_IOCTL_LDR_LOAD: sym #%ld: unterminated name! (%#lx / %#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1337	}
1338	}
1339
1340	/* execute */
1341	pReq->Hdr.rc = supdrvIOCtl_LdrLoad(pDevExt, pSession, pReq);
1342	return 0;
1343	}
1344
1345	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_FREE):
1346	{
1347	/* validate */
1348	PSUPLDRFREE pReq = (PSUPLDRFREE)pReqHdr;
1349	REQ_CHECK_SIZES(SUP_IOCTL_LDR_FREE);
1350
1351	/* execute */
1352	pReq->Hdr.rc = supdrvIOCtl_LdrFree(pDevExt, pSession, pReq);
1353	return 0;
1354	}
1355
1356	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_GET_SYMBOL):
1357	{
1358	/* validate */
1359	PSUPLDRGETSYMBOL pReq = (PSUPLDRGETSYMBOL)pReqHdr;
1360	REQ_CHECK_SIZES(SUP_IOCTL_LDR_GET_SYMBOL);
1361	REQ_CHECK_EXPR(SUP_IOCTL_LDR_GET_SYMBOL, RTStrEnd(pReq->u.In.szSymbol, sizeof(pReq->u.In.szSymbol)));
1362
1363	/* execute */
1364	pReq->Hdr.rc = supdrvIOCtl_LdrGetSymbol(pDevExt, pSession, pReq);
1365	return 0;
1366	}
1367
1368	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0(0)):
1369	{
1370	/* validate */
1371	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1372	Log4(("SUP_IOCTL_CALL_VMMR0: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1373	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1374
1375	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_VMMR0_SIZE(0))
1376	{
1377	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(0), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(0));
1378
1379	/* execute */
1380	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1381	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg, pSession);
1382	else
1383	pReq->Hdr.rc = VERR_WRONG_ORDER;
1384	}
1385	else
1386	{
1387	PSUPVMMR0REQHDR pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1388	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR)),
1389	("SUP_IOCTL_CALL_VMMR0: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR))));
1390	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1391	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(pVMMReq->cbReq));
1392
1393	/* execute */
1394	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1395	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1396	else
1397	pReq->Hdr.rc = VERR_WRONG_ORDER;
1398	}
1399
1400	if ( RT_FAILURE(pReq->Hdr.rc)
1401	&& pReq->Hdr.rc != VERR_INTERRUPTED
1402	&& pReq->Hdr.rc != VERR_TIMEOUT)
1403	Log(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1404	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1405	else
1406	Log4(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1407	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1408	return 0;
1409	}
1410
1411	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0_BIG):
1412	{
1413	/* validate */
1414	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1415	PSUPVMMR0REQHDR pVMMReq;
1416	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1417	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1418
1419	pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1420	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR)),
1421	("SUP_IOCTL_CALL_VMMR0_BIG: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR))));
1422	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0_BIG, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1423	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));
1424
1425	/* execute */
1426	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1427	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1428	else
1429	pReq->Hdr.rc = VERR_WRONG_ORDER;
1430
1431	if ( RT_FAILURE(pReq->Hdr.rc)
1432	&& pReq->Hdr.rc != VERR_INTERRUPTED
1433	&& pReq->Hdr.rc != VERR_TIMEOUT)
1434	Log(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1435	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1436	else
1437	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1438	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1439	return 0;
1440	}
1441
1442	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_PAGING_MODE):
1443	{
1444	/* validate */
1445	PSUPGETPAGINGMODE pReq = (PSUPGETPAGINGMODE)pReqHdr;
1446	REQ_CHECK_SIZES(SUP_IOCTL_GET_PAGING_MODE);
1447
1448	/* execute */
1449	pReq->Hdr.rc = VINF_SUCCESS;
1450	pReq->u.Out.enmMode = SUPR0GetPagingMode();
1451	return 0;
1452	}
1453
1454	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_ALLOC):
1455	{
1456	/* validate */
1457	PSUPLOWALLOC pReq = (PSUPLOWALLOC)pReqHdr;
1458	REQ_CHECK_EXPR(SUP_IOCTL_LOW_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_LOW_ALLOC_SIZE_IN);
1459	REQ_CHECK_SIZES_EX(SUP_IOCTL_LOW_ALLOC, SUP_IOCTL_LOW_ALLOC_SIZE_IN, SUP_IOCTL_LOW_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1460
1461	/* execute */
1462	pReq->Hdr.rc = SUPR0LowAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1463	if (RT_FAILURE(pReq->Hdr.rc))
1464	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1465	return 0;
1466	}
1467
1468	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_FREE):
1469	{
1470	/* validate */
1471	PSUPLOWFREE pReq = (PSUPLOWFREE)pReqHdr;
1472	REQ_CHECK_SIZES(SUP_IOCTL_LOW_FREE);
1473
1474	/* execute */
1475	pReq->Hdr.rc = SUPR0LowFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1476	return 0;
1477	}
1478
1479	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_MAP):
1480	{
1481	/* validate */
1482	PSUPGIPMAP pReq = (PSUPGIPMAP)pReqHdr;
1483	REQ_CHECK_SIZES(SUP_IOCTL_GIP_MAP);
1484
1485	/* execute */
1486	pReq->Hdr.rc = SUPR0GipMap(pSession, &pReq->u.Out.pGipR3, &pReq->u.Out.HCPhysGip);
1487	if (RT_SUCCESS(pReq->Hdr.rc))
1488	pReq->u.Out.pGipR0 = pDevExt->pGip;
1489	return 0;
1490	}
1491
1492	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_UNMAP):
1493	{
1494	/* validate */
1495	PSUPGIPUNMAP pReq = (PSUPGIPUNMAP)pReqHdr;
1496	REQ_CHECK_SIZES(SUP_IOCTL_GIP_UNMAP);
1497
1498	/* execute */
1499	pReq->Hdr.rc = SUPR0GipUnmap(pSession);
1500	return 0;
1501	}
1502
1503	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SET_VM_FOR_FAST):
1504	{
1505	/* validate */
1506	PSUPSETVMFORFAST pReq = (PSUPSETVMFORFAST)pReqHdr;
1507	REQ_CHECK_SIZES(SUP_IOCTL_SET_VM_FOR_FAST);
1508	REQ_CHECK_EXPR_FMT( !pReq->u.In.pVMR0
1509	\|\| ( VALID_PTR(pReq->u.In.pVMR0)
1510	&& !((uintptr_t)pReq->u.In.pVMR0 & (PAGE_SIZE - 1))),
1511	("SUP_IOCTL_SET_VM_FOR_FAST: pVMR0=%p!\n", pReq->u.In.pVMR0));
1512	/* execute */
1513	pSession->pVM = pReq->u.In.pVMR0;
1514	pReq->Hdr.rc = VINF_SUCCESS;
1515	return 0;
1516	}
1517
1518	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_ALLOC_EX):
1519	{
1520	/* validate */
1521	PSUPPAGEALLOCEX pReq = (PSUPPAGEALLOCEX)pReqHdr;
1522	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC_EX, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_EX_SIZE_IN);
1523	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));
1524	REQ_CHECK_EXPR_FMT(pReq->u.In.fKernelMapping \|\| pReq->u.In.fUserMapping,
1525	("SUP_IOCTL_PAGE_ALLOC_EX: No mapping requested!\n"));
1526	REQ_CHECK_EXPR_FMT(pReq->u.In.fUserMapping,
1527	("SUP_IOCTL_PAGE_ALLOC_EX: Must have user mapping!\n"));
1528	REQ_CHECK_EXPR_FMT(!pReq->u.In.fReserved0 && !pReq->u.In.fReserved1,
1529	("SUP_IOCTL_PAGE_ALLOC_EX: fReserved0=%d fReserved1=%d\n", pReq->u.In.fReserved0, pReq->u.In.fReserved1));
1530
1531	/* execute */
1532	pReq->Hdr.rc = SUPR0PageAllocEx(pSession, pReq->u.In.cPages, 0 /* fFlags */,
1533	pReq->u.In.fUserMapping ? &pReq->u.Out.pvR3 : NULL,
1534	pReq->u.In.fKernelMapping ? &pReq->u.Out.pvR0 : NULL,
1535	&pReq->u.Out.aPages[0]);
1536	if (RT_FAILURE(pReq->Hdr.rc))
1537	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1538	return 0;
1539	}
1540
1541	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_MAP_KERNEL):
1542	{
1543	/* validate */
1544	PSUPPAGEMAPKERNEL pReq = (PSUPPAGEMAPKERNEL)pReqHdr;
1545	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_MAP_KERNEL);
1546	REQ_CHECK_EXPR_FMT(!pReq->u.In.fFlags, ("SUP_IOCTL_PAGE_MAP_KERNEL: fFlags=%#x! MBZ\n", pReq->u.In.fFlags));
1547	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_MAP_KERNEL: offSub=%#x\n", pReq->u.In.offSub));
1548	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1549	("SUP_IOCTL_PAGE_MAP_KERNEL: cbSub=%#x\n", pReq->u.In.cbSub));
1550
1551	/* execute */
1552	pReq->Hdr.rc = SUPR0PageMapKernel(pSession, pReq->u.In.pvR3, pReq->u.In.offSub, pReq->u.In.cbSub,
1553	pReq->u.In.fFlags, &pReq->u.Out.pvR0);
1554	if (RT_FAILURE(pReq->Hdr.rc))
1555	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1556	return 0;
1557	}
1558
1559	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_PROTECT):
1560	{
1561	/* validate */
1562	PSUPPAGEPROTECT pReq = (PSUPPAGEPROTECT)pReqHdr;
1563	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_PROTECT);
1564	REQ_CHECK_EXPR_FMT(!(pReq->u.In.fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)),
1565	("SUP_IOCTL_PAGE_PROTECT: fProt=%#x!\n", pReq->u.In.fProt));
1566	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_PROTECT: offSub=%#x\n", pReq->u.In.offSub));
1567	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1568	("SUP_IOCTL_PAGE_PROTECT: cbSub=%#x\n", pReq->u.In.cbSub));
1569
1570	/* execute */
1571	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);
1572	return 0;
1573	}
1574
1575	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_FREE):
1576	{
1577	/* validate */
1578	PSUPPAGEFREE pReq = (PSUPPAGEFREE)pReqHdr;
1579	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_FREE);
1580
1581	/* execute */
1582	pReq->Hdr.rc = SUPR0PageFree(pSession, pReq->u.In.pvR3);
1583	return 0;
1584	}
1585
1586	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_SERVICE(0)):
1587	{
1588	/* validate */
1589	PSUPCALLSERVICE pReq = (PSUPCALLSERVICE)pReqHdr;
1590	Log4(("SUP_IOCTL_CALL_SERVICE: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1591	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1592
1593	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
1594	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(0), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(0));
1595	else
1596	{
1597	PSUPR0SERVICEREQHDR pSrvReq = (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0];
1598	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR)),
1599	("SUP_IOCTL_CALL_SERVICE: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR))));
1600	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, pSrvReq->u32Magic == SUPR0SERVICEREQHDR_MAGIC);
1601	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(pSrvReq->cbReq), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(pSrvReq->cbReq));
1602	}
1603	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1604
1605	/* execute */
1606	pReq->Hdr.rc = supdrvIOCtl_CallServiceModule(pDevExt, pSession, pReq);
1607	return 0;
1608	}
1609
1610	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOGGER_SETTINGS(0)):
1611	{
1612	/* validate */
1613	PSUPLOGGERSETTINGS pReq = (PSUPLOGGERSETTINGS)pReqHdr;
1614	size_t cbStrTab;
1615	REQ_CHECK_SIZE_OUT(SUP_IOCTL_LOGGER_SETTINGS, SUP_IOCTL_LOGGER_SETTINGS_SIZE_OUT);
1616	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->Hdr.cbIn >= SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(1));
1617	cbStrTab = pReq->Hdr.cbIn - SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(0);
1618	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offGroups < cbStrTab);
1619	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offFlags < cbStrTab);
1620	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offDestination < cbStrTab);
1621	REQ_CHECK_EXPR_FMT(pReq->u.In.szStrings[cbStrTab - 1] == '\0',
1622	("SUP_IOCTL_LOGGER_SETTINGS: cbIn=%#x cbStrTab=%#zx LastChar=%d\n",
1623	pReq->Hdr.cbIn, cbStrTab, pReq->u.In.szStrings[cbStrTab - 1]));
1624	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhich <= SUPLOGGERSETTINGS_WHICH_RELEASE);
1625	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhat <= SUPLOGGERSETTINGS_WHAT_DESTROY);
1626
1627	/* execute */
1628	pReq->Hdr.rc = supdrvIOCtl_LoggerSettings(pDevExt, pSession, pReq);
1629	return 0;
1630	}
1631
1632	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP2):
1633	{
1634	/* validate */
1635	PSUPSEMOP2 pReq = (PSUPSEMOP2)pReqHdr;
1636	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP2, SUP_IOCTL_SEM_OP2_SIZE_IN, SUP_IOCTL_SEM_OP2_SIZE_OUT);
1637	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP2, pReq->u.In.uReserved == 0);
1638
1639	/* execute */
1640	switch (pReq->u.In.uType)
1641	{
1642	case SUP_SEM_TYPE_EVENT:
1643	{
1644	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1645	switch (pReq->u.In.uOp)
1646	{
1647	case SUPSEMOP2_WAIT_MS_REL:
1648	pReq->Hdr.rc = SUPSemEventWaitNoResume(pSession, hEvent, pReq->u.In.uArg.cRelMsTimeout);
1649	break;
1650	case SUPSEMOP2_WAIT_NS_ABS:
1651	pReq->Hdr.rc = SUPSemEventWaitNsAbsIntr(pSession, hEvent, pReq->u.In.uArg.uAbsNsTimeout);
1652	break;
1653	case SUPSEMOP2_WAIT_NS_REL:
1654	pReq->Hdr.rc = SUPSemEventWaitNsRelIntr(pSession, hEvent, pReq->u.In.uArg.cRelNsTimeout);
1655	break;
1656	case SUPSEMOP2_SIGNAL:
1657	pReq->Hdr.rc = SUPSemEventSignal(pSession, hEvent);
1658	break;
1659	case SUPSEMOP2_CLOSE:
1660	pReq->Hdr.rc = SUPSemEventClose(pSession, hEvent);
1661	break;
1662	case SUPSEMOP2_RESET:
1663	default:
1664	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1665	break;
1666	}
1667	break;
1668	}
1669
1670	case SUP_SEM_TYPE_EVENT_MULTI:
1671	{
1672	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1673	switch (pReq->u.In.uOp)
1674	{
1675	case SUPSEMOP2_WAIT_MS_REL:
1676	pReq->Hdr.rc = SUPSemEventMultiWaitNoResume(pSession, hEventMulti, pReq->u.In.uArg.cRelMsTimeout);
1677	break;
1678	case SUPSEMOP2_WAIT_NS_ABS:
1679	pReq->Hdr.rc = SUPSemEventMultiWaitNsAbsIntr(pSession, hEventMulti, pReq->u.In.uArg.uAbsNsTimeout);
1680	break;
1681	case SUPSEMOP2_WAIT_NS_REL:
1682	pReq->Hdr.rc = SUPSemEventMultiWaitNsRelIntr(pSession, hEventMulti, pReq->u.In.uArg.cRelNsTimeout);
1683	break;
1684	case SUPSEMOP2_SIGNAL:
1685	pReq->Hdr.rc = SUPSemEventMultiSignal(pSession, hEventMulti);
1686	break;
1687	case SUPSEMOP2_CLOSE:
1688	pReq->Hdr.rc = SUPSemEventMultiClose(pSession, hEventMulti);
1689	break;
1690	case SUPSEMOP2_RESET:
1691	pReq->Hdr.rc = SUPSemEventMultiReset(pSession, hEventMulti);
1692	break;
1693	default:
1694	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1695	break;
1696	}
1697	break;
1698	}
1699
1700	default:
1701	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1702	break;
1703	}
1704	return 0;
1705	}
1706
1707	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP3):
1708	{
1709	/* validate */
1710	PSUPSEMOP3 pReq = (PSUPSEMOP3)pReqHdr;
1711	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP3, SUP_IOCTL_SEM_OP3_SIZE_IN, SUP_IOCTL_SEM_OP3_SIZE_OUT);
1712	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, pReq->u.In.u32Reserved == 0 && pReq->u.In.u64Reserved == 0);
1713
1714	/* execute */
1715	switch (pReq->u.In.uType)
1716	{
1717	case SUP_SEM_TYPE_EVENT:
1718	{
1719	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1720	switch (pReq->u.In.uOp)
1721	{
1722	case SUPSEMOP3_CREATE:
1723	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1724	pReq->Hdr.rc = SUPSemEventCreate(pSession, &hEvent);
1725	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEvent;
1726	break;
1727	case SUPSEMOP3_GET_RESOLUTION:
1728	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1729	pReq->Hdr.rc = VINF_SUCCESS;
1730	pReq->Hdr.cbOut = sizeof(*pReq);
1731	pReq->u.Out.cNsResolution = SUPSemEventGetResolution(pSession);
1732	break;
1733	default:
1734	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1735	break;
1736	}
1737	break;
1738	}
1739
1740	case SUP_SEM_TYPE_EVENT_MULTI:
1741	{
1742	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1743	switch (pReq->u.In.uOp)
1744	{
1745	case SUPSEMOP3_CREATE:
1746	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1747	pReq->Hdr.rc = SUPSemEventMultiCreate(pSession, &hEventMulti);
1748	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEventMulti;
1749	break;
1750	case SUPSEMOP3_GET_RESOLUTION:
1751	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1752	pReq->Hdr.rc = VINF_SUCCESS;
1753	pReq->u.Out.cNsResolution = SUPSemEventMultiGetResolution(pSession);
1754	break;
1755	default:
1756	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1757	break;
1758	}
1759	break;
1760	}
1761
1762	default:
1763	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1764	break;
1765	}
1766	return 0;
1767	}
1768
1769	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
1770	{
1771	/* validate */
1772	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
1773	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
1774
1775	/* execute */
1776	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.Caps);
1777	if (RT_FAILURE(pReq->Hdr.rc))
1778	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1779	return 0;
1780	}
1781
1782	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_OPEN):
1783	{
1784	/* validate */
1785	PSUPTRACEROPEN pReq = (PSUPTRACEROPEN)pReqHdr;
1786	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_OPEN);
1787
1788	/* execute */
1789	pReq->Hdr.rc = supdrvIOCtl_TracerOpen(pDevExt, pSession, pReq->u.In.uCookie, pReq->u.In.uArg);
1790	return 0;
1791	}
1792
1793	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_CLOSE):
1794	{
1795	/* validate */
1796	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_CLOSE);
1797
1798	/* execute */
1799	pReqHdr->rc = supdrvIOCtl_TracerClose(pDevExt, pSession);
1800	return 0;
1801	}
1802
1803	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_IOCTL):
1804	{
1805	/* validate */
1806	PSUPTRACERIOCTL pReq = (PSUPTRACERIOCTL)pReqHdr;
1807	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_IOCTL);
1808
1809	/* execute */
1810	pReqHdr->rc = supdrvIOCtl_TracerIOCtl(pDevExt, pSession, pReq->u.In.uCmd, pReq->u.In.uArg, &pReq->u.Out.iRetVal);
1811	return 0;
1812	}
1813
1814	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_REG):
1815	{
1816	/* validate */
1817	PSUPTRACERUMODREG pReq = (PSUPTRACERUMODREG)pReqHdr;
1818	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_REG);
1819	if (!RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)))
1820	return VERR_INVALID_PARAMETER;
1821
1822	/* execute */
1823	pReqHdr->rc = supdrvIOCtl_TracerUmodRegister(pDevExt, pSession,
1824	pReq->u.In.R3PtrVtgHdr, pReq->u.In.uVtgHdrAddr,
1825	pReq->u.In.R3PtrStrTab, pReq->u.In.cbStrTab,
1826	pReq->u.In.szName, pReq->u.In.fFlags);
1827	return 0;
1828	}
1829
1830	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_DEREG):
1831	{
1832	/* validate */
1833	PSUPTRACERUMODDEREG pReq = (PSUPTRACERUMODDEREG)pReqHdr;
1834	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_DEREG);
1835
1836	/* execute */
1837	pReqHdr->rc = supdrvIOCtl_TracerUmodDeregister(pDevExt, pSession, pReq->u.In.pVtgHdr);
1838	return 0;
1839	}
1840
1841	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_FIRE_PROBE):
1842	{
1843	/* validate */
1844	PSUPTRACERUMODFIREPROBE pReq = (PSUPTRACERUMODFIREPROBE)pReqHdr;
1845	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_FIRE_PROBE);
1846
1847	supdrvIOCtl_TracerUmodProbeFire(pDevExt, pSession, &pReq->u.In);
1848	pReqHdr->rc = VINF_SUCCESS;
1849	return 0;
1850	}
1851
1852	default:
1853	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1854	break;
1855	}
1856	return VERR_GENERAL_FAILURE;
1857	}
1858
1859
1860	/**
1861	* I/O Control inner worker for the restricted operations.
1862	*
1863	* @returns IPRT status code.
1864	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1865	*
1866	* @param uIOCtl Function number.
1867	* @param pDevExt Device extention.
1868	* @param pSession Session data.
1869	* @param pReqHdr The request header.
1870	*/
1871	static int supdrvIOCtlInnerRestricted(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1872	{
1873	/*
1874	* The switch.
1875	*/
1876	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
1877	{
1878	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
1879	{
1880	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
1881	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
1882	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
1883	{
1884	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
1885	pReq->Hdr.rc = VERR_INVALID_MAGIC;
1886	return 0;
1887	}
1888
1889	/*
1890	* Match the version.
1891	* The current logic is very simple, match the major interface version.
1892	*/
1893	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
1894	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
1895	{
1896	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1897	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
1898	pReq->u.Out.u32Cookie = 0xffffffff;
1899	pReq->u.Out.u32SessionCookie = 0xffffffff;
1900	pReq->u.Out.u32SessionVersion = 0xffffffff;
1901	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1902	pReq->u.Out.pSession = NULL;
1903	pReq->u.Out.cFunctions = 0;
1904	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1905	return 0;
1906	}
1907
1908	/*
1909	* Fill in return data and be gone.
1910	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
1911	* u32SessionVersion <= u32ReqVersion!
1912	*/
1913	/** @todo Somehow validate the client and negotiate a secure cookie... */
1914	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
1915	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
1916	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
1917	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1918	pReq->u.Out.pSession = pSession;
1919	pReq->u.Out.cFunctions = 0;
1920	pReq->Hdr.rc = VINF_SUCCESS;
1921	return 0;
1922	}
1923
1924	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
1925	{
1926	/* validate */
1927	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
1928	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
1929
1930	/* execute */
1931	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.Caps);
1932	if (RT_FAILURE(pReq->Hdr.rc))
1933	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1934	return 0;
1935	}
1936
1937	default:
1938	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1939	break;
1940	}
1941	return VERR_GENERAL_FAILURE;
1942	}
1943
1944
1945	/**
1946	* I/O Control worker.
1947	*
1948	* @returns IPRT status code.
1949	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1950	*
1951	* @param uIOCtl Function number.
1952	* @param pDevExt Device extention.
1953	* @param pSession Session data.
1954	* @param pReqHdr The request header.
1955	*/
1956	int VBOXCALL supdrvIOCtl(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1957	{
1958	int rc;
1959	VBOXDRV_IOCTL_ENTRY(pSession, uIOCtl, pReqHdr);
1960
1961	/*
1962	* Validate the request.
1963	*/
1964	/* this first check could probably be omitted as its also done by the OS specific code... */
1965	if (RT_UNLIKELY( (pReqHdr->fFlags & SUPREQHDR_FLAGS_MAGIC_MASK) != SUPREQHDR_FLAGS_MAGIC
1966	\|\| pReqHdr->cbIn < sizeof(*pReqHdr)
1967	\|\| pReqHdr->cbOut < sizeof(*pReqHdr)))
1968	{
1969	OSDBGPRINT(("vboxdrv: Bad ioctl request header; cbIn=%#lx cbOut=%#lx fFlags=%#lx\n",
1970	(long)pReqHdr->cbIn, (long)pReqHdr->cbOut, (long)pReqHdr->fFlags));
1971	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1972	return VERR_INVALID_PARAMETER;
1973	}
1974	if (RT_UNLIKELY(!RT_VALID_PTR(pSession)))
1975	{
1976	OSDBGPRINT(("vboxdrv: Invalid pSession valud %p (ioctl=%p)\n", pSession, (void *)uIOCtl));
1977	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1978	return VERR_INVALID_PARAMETER;
1979	}
1980	if (RT_UNLIKELY(uIOCtl == SUP_IOCTL_COOKIE))
1981	{
1982	if (pReqHdr->u32Cookie != SUPCOOKIE_INITIAL_COOKIE)
1983	{
1984	OSDBGPRINT(("SUP_IOCTL_COOKIE: bad cookie %#lx\n", (long)pReqHdr->u32Cookie));
1985	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1986	return VERR_INVALID_PARAMETER;
1987	}
1988	}
1989	else if (RT_UNLIKELY( pReqHdr->u32Cookie != pDevExt->u32Cookie
1990	\|\| pReqHdr->u32SessionCookie != pSession->u32Cookie))
1991	{
1992	OSDBGPRINT(("vboxdrv: bad cookie %#lx / %#lx.\n", (long)pReqHdr->u32Cookie, (long)pReqHdr->u32SessionCookie));
1993	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1994	return VERR_INVALID_PARAMETER;
1995	}
1996
1997	/*
1998	* Hand it to an inner function to avoid lots of unnecessary return tracepoints.
1999	*/
2000	if (pSession->fUnrestricted)
2001	rc = supdrvIOCtlInnerUnrestricted(uIOCtl, pDevExt, pSession, pReqHdr);
2002	else
2003	rc = supdrvIOCtlInnerRestricted(uIOCtl, pDevExt, pSession, pReqHdr);
2004
2005	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, pReqHdr->rc, rc);
2006	return rc;
2007	}
2008
2009
2010	/**
2011	* Inter-Driver Communication (IDC) worker.
2012	*
2013	* @returns VBox status code.
2014	* @retval VINF_SUCCESS on success.
2015	* @retval VERR_INVALID_PARAMETER if the request is invalid.
2016	* @retval VERR_NOT_SUPPORTED if the request isn't supported.
2017	*
2018	* @param uReq The request (function) code.
2019	* @param pDevExt Device extention.
2020	* @param pSession Session data.
2021	* @param pReqHdr The request header.
2022	*/
2023	int VBOXCALL supdrvIDC(uintptr_t uReq, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQHDR pReqHdr)
2024	{
2025	/*
2026	* The OS specific code has already validated the pSession
2027	* pointer, and the request size being greater or equal to
2028	* size of the header.
2029	*
2030	* So, just check that pSession is a kernel context session.
2031	*/
2032	if (RT_UNLIKELY( pSession
2033	&& pSession->R0Process != NIL_RTR0PROCESS))
2034	return VERR_INVALID_PARAMETER;
2035
2036	/*
2037	* Validation macro.
2038	*/
2039	#define REQ_CHECK_IDC_SIZE(Name, cbExpect) \
2040	do { \
2041	if (RT_UNLIKELY(pReqHdr->cb != (cbExpect))) \
2042	{ \
2043	OSDBGPRINT(( #Name ": Invalid input/output sizes. cb=%ld expected %ld.\n", \
2044	(long)pReqHdr->cb, (long)(cbExpect))); \
2045	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
2046	} \
2047	} while (0)
2048
2049	switch (uReq)
2050	{
2051	case SUPDRV_IDC_REQ_CONNECT:
2052	{
2053	PSUPDRVIDCREQCONNECT pReq = (PSUPDRVIDCREQCONNECT)pReqHdr;
2054	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_CONNECT, sizeof(*pReq));
2055
2056	/*
2057	* Validate the cookie and other input.
2058	*/
2059	if (pReq->Hdr.pSession != NULL)
2060	{
2061	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Hdr.pSession=%p expected NULL!\n", pReq->Hdr.pSession));
2062	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2063	}
2064	if (pReq->u.In.u32MagicCookie != SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE)
2065	{
2066	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: u32MagicCookie=%#x expected %#x!\n",
2067	(unsigned)pReq->u.In.u32MagicCookie, (unsigned)SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE));
2068	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2069	}
2070	if ( pReq->u.In.uMinVersion > pReq->u.In.uReqVersion
2071	\|\| (pReq->u.In.uMinVersion & UINT32_C(0xffff0000)) != (pReq->u.In.uReqVersion & UINT32_C(0xffff0000)))
2072	{
2073	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: uMinVersion=%#x uMaxVersion=%#x doesn't match!\n",
2074	pReq->u.In.uMinVersion, pReq->u.In.uReqVersion));
2075	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2076	}
2077	if (pSession != NULL)
2078	{
2079	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: pSession=%p expected NULL!\n", pSession));
2080	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2081	}
2082
2083	/*
2084	* Match the version.
2085	* The current logic is very simple, match the major interface version.
2086	*/
2087	if ( pReq->u.In.uMinVersion > SUPDRV_IDC_VERSION
2088	\|\| (pReq->u.In.uMinVersion & 0xffff0000) != (SUPDRV_IDC_VERSION & 0xffff0000))
2089	{
2090	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
2091	pReq->u.In.uReqVersion, pReq->u.In.uMinVersion, (unsigned)SUPDRV_IDC_VERSION));
2092	pReq->u.Out.pSession = NULL;
2093	pReq->u.Out.uSessionVersion = 0xffffffff;
2094	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
2095	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
2096	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
2097	return VINF_SUCCESS;
2098	}
2099
2100	pReq->u.Out.pSession = NULL;
2101	pReq->u.Out.uSessionVersion = SUPDRV_IDC_VERSION;
2102	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
2103	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
2104
2105	pReq->Hdr.rc = supdrvCreateSession(pDevExt, false /* fUser /, true /fUnrestricted*/, &pSession);
2106	if (RT_FAILURE(pReq->Hdr.rc))
2107	{
2108	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: failed to create session, rc=%d\n", pReq->Hdr.rc));
2109	return VINF_SUCCESS;
2110	}
2111
2112	pReq->u.Out.pSession = pSession;
2113	pReq->Hdr.pSession = pSession;
2114
2115	return VINF_SUCCESS;
2116	}
2117
2118	case SUPDRV_IDC_REQ_DISCONNECT:
2119	{
2120	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_DISCONNECT, sizeof(*pReqHdr));
2121
2122	supdrvCloseSession(pDevExt, pSession);
2123	return pReqHdr->rc = VINF_SUCCESS;
2124	}
2125
2126	case SUPDRV_IDC_REQ_GET_SYMBOL:
2127	{
2128	PSUPDRVIDCREQGETSYM pReq = (PSUPDRVIDCREQGETSYM)pReqHdr;
2129	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_GET_SYMBOL, sizeof(*pReq));
2130
2131	pReq->Hdr.rc = supdrvIDC_LdrGetSymbol(pDevExt, pSession, pReq);
2132	return VINF_SUCCESS;
2133	}
2134
2135	case SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY:
2136	{
2137	PSUPDRVIDCREQCOMPREGFACTORY pReq = (PSUPDRVIDCREQCOMPREGFACTORY)pReqHdr;
2138	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY, sizeof(*pReq));
2139
2140	pReq->Hdr.rc = SUPR0ComponentRegisterFactory(pSession, pReq->u.In.pFactory);
2141	return VINF_SUCCESS;
2142	}
2143
2144	case SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY:
2145	{
2146	PSUPDRVIDCREQCOMPDEREGFACTORY pReq = (PSUPDRVIDCREQCOMPDEREGFACTORY)pReqHdr;
2147	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY, sizeof(*pReq));
2148
2149	pReq->Hdr.rc = SUPR0ComponentDeregisterFactory(pSession, pReq->u.In.pFactory);
2150	return VINF_SUCCESS;
2151	}
2152
2153	default:
2154	Log(("Unknown IDC %#lx\n", (long)uReq));
2155	break;
2156	}
2157
2158	#undef REQ_CHECK_IDC_SIZE
2159	return VERR_NOT_SUPPORTED;
2160	}
2161
2162
2163	/**
2164	* Register a object for reference counting.
2165	* The object is registered with one reference in the specified session.
2166	*
2167	* @returns Unique identifier on success (pointer).
2168	* All future reference must use this identifier.
2169	* @returns NULL on failure.
2170	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
2171	* @param pvUser1 The first user argument.
2172	* @param pvUser2 The second user argument.
2173	*/
2174	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
2175	{
2176	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2177	PSUPDRVOBJ pObj;
2178	PSUPDRVUSAGE pUsage;
2179
2180	/*
2181	* Validate the input.
2182	*/
2183	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
2184	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
2185	AssertPtrReturn(pfnDestructor, NULL);
2186
2187	/*
2188	* Allocate and initialize the object.
2189	*/
2190	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
2191	if (!pObj)
2192	return NULL;
2193	pObj->u32Magic = SUPDRVOBJ_MAGIC;
2194	pObj->enmType = enmType;
2195	pObj->pNext = NULL;
2196	pObj->cUsage = 1;
2197	pObj->pfnDestructor = pfnDestructor;
2198	pObj->pvUser1 = pvUser1;
2199	pObj->pvUser2 = pvUser2;
2200	pObj->CreatorUid = pSession->Uid;
2201	pObj->CreatorGid = pSession->Gid;
2202	pObj->CreatorProcess= pSession->Process;
2203	supdrvOSObjInitCreator(pObj, pSession);
2204
2205	/*
2206	* Allocate the usage record.
2207	* (We keep freed usage records around to simplify SUPR0ObjAddRefEx().)
2208	*/
2209	RTSpinlockAcquire(pDevExt->Spinlock);
2210
2211	pUsage = pDevExt->pUsageFree;
2212	if (pUsage)
2213	pDevExt->pUsageFree = pUsage->pNext;
2214	else
2215	{
2216	RTSpinlockRelease(pDevExt->Spinlock);
2217	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
2218	if (!pUsage)
2219	{
2220	RTMemFree(pObj);
2221	return NULL;
2222	}
2223	RTSpinlockAcquire(pDevExt->Spinlock);
2224	}
2225
2226	/*
2227	* Insert the object and create the session usage record.
2228	*/
2229	/* The object. */
2230	pObj->pNext = pDevExt->pObjs;
2231	pDevExt->pObjs = pObj;
2232
2233	/* The session record. */
2234	pUsage->cUsage = 1;
2235	pUsage->pObj = pObj;
2236	pUsage->pNext = pSession->pUsage;
2237	/* Log2(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext)); */
2238	pSession->pUsage = pUsage;
2239
2240	RTSpinlockRelease(pDevExt->Spinlock);
2241
2242	Log(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
2243	return pObj;
2244	}
2245
2246
2247	/**
2248	* Increment the reference counter for the object associating the reference
2249	* with the specified session.
2250	*
2251	* @returns IPRT status code.
2252	* @param pvObj The identifier returned by SUPR0ObjRegister().
2253	* @param pSession The session which is referencing the object.
2254	*
2255	* @remarks The caller should not own any spinlocks and must carefully protect
2256	* itself against potential race with the destructor so freed memory
2257	* isn't accessed here.
2258	*/
2259	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
2260	{
2261	return SUPR0ObjAddRefEx(pvObj, pSession, false /* fNoBlocking */);
2262	}
2263
2264
2265	/**
2266	* Increment the reference counter for the object associating the reference
2267	* with the specified session.
2268	*
2269	* @returns IPRT status code.
2270	* @retval VERR_TRY_AGAIN if fNoBlocking was set and a new usage record
2271	* couldn't be allocated. (If you see this you're not doing the right
2272	* thing and it won't ever work reliably.)
2273	*
2274	* @param pvObj The identifier returned by SUPR0ObjRegister().
2275	* @param pSession The session which is referencing the object.
2276	* @param fNoBlocking Set if it's not OK to block. Never try to make the
2277	* first reference to an object in a session with this
2278	* argument set.
2279	*
2280	* @remarks The caller should not own any spinlocks and must carefully protect
2281	* itself against potential race with the destructor so freed memory
2282	* isn't accessed here.
2283	*/
2284	SUPR0DECL(int) SUPR0ObjAddRefEx(void *pvObj, PSUPDRVSESSION pSession, bool fNoBlocking)
2285	{
2286	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2287	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2288	int rc = VINF_SUCCESS;
2289	PSUPDRVUSAGE pUsagePre;
2290	PSUPDRVUSAGE pUsage;
2291
2292	/*
2293	* Validate the input.
2294	* Be ready for the destruction race (someone might be stuck in the
2295	* destructor waiting a lock we own).
2296	*/
2297	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2298	AssertPtrReturn(pObj, VERR_INVALID_POINTER);
2299	AssertMsgReturn(pObj->u32Magic == SUPDRVOBJ_MAGIC \|\| pObj->u32Magic == SUPDRVOBJ_MAGIC_DEAD,
2300	("Invalid pvObj=%p magic=%#x (expected %#x or %#x)\n", pvObj, pObj->u32Magic, SUPDRVOBJ_MAGIC, SUPDRVOBJ_MAGIC_DEAD),
2301	VERR_INVALID_PARAMETER);
2302
2303	RTSpinlockAcquire(pDevExt->Spinlock);
2304
2305	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2306	{
2307	RTSpinlockRelease(pDevExt->Spinlock);
2308
2309	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2310	return VERR_WRONG_ORDER;
2311	}
2312
2313	/*
2314	* Preallocate the usage record if we can.
2315	*/
2316	pUsagePre = pDevExt->pUsageFree;
2317	if (pUsagePre)
2318	pDevExt->pUsageFree = pUsagePre->pNext;
2319	else if (!fNoBlocking)
2320	{
2321	RTSpinlockRelease(pDevExt->Spinlock);
2322	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
2323	if (!pUsagePre)
2324	return VERR_NO_MEMORY;
2325
2326	RTSpinlockAcquire(pDevExt->Spinlock);
2327	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2328	{
2329	RTSpinlockRelease(pDevExt->Spinlock);
2330
2331	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2332	return VERR_WRONG_ORDER;
2333	}
2334	}
2335
2336	/*
2337	* Reference the object.
2338	*/
2339	pObj->cUsage++;
2340
2341	/*
2342	* Look for the session record.
2343	*/
2344	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
2345	{
2346	/Log(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2347	if (pUsage->pObj == pObj)
2348	break;
2349	}
2350	if (pUsage)
2351	pUsage->cUsage++;
2352	else if (pUsagePre)
2353	{
2354	/* create a new session record. */
2355	pUsagePre->cUsage = 1;
2356	pUsagePre->pObj = pObj;
2357	pUsagePre->pNext = pSession->pUsage;
2358	pSession->pUsage = pUsagePre;
2359	/Log(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));/
2360
2361	pUsagePre = NULL;
2362	}
2363	else
2364	{
2365	pObj->cUsage--;
2366	rc = VERR_TRY_AGAIN;
2367	}
2368
2369	/*
2370	* Put any unused usage record into the free list..
2371	*/
2372	if (pUsagePre)
2373	{
2374	pUsagePre->pNext = pDevExt->pUsageFree;
2375	pDevExt->pUsageFree = pUsagePre;
2376	}
2377
2378	RTSpinlockRelease(pDevExt->Spinlock);
2379
2380	return rc;
2381	}
2382
2383
2384	/**
2385	* Decrement / destroy a reference counter record for an object.
2386	*
2387	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
2388	*
2389	* @returns IPRT status code.
2390	* @retval VINF_SUCCESS if not destroyed.
2391	* @retval VINF_OBJECT_DESTROYED if it's destroyed by this release call.
2392	* @retval VERR_INVALID_PARAMETER if the object isn't valid. Will assert in
2393	* string builds.
2394	*
2395	* @param pvObj The identifier returned by SUPR0ObjRegister().
2396	* @param pSession The session which is referencing the object.
2397	*/
2398	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
2399	{
2400	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2401	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2402	int rc = VERR_INVALID_PARAMETER;
2403	PSUPDRVUSAGE pUsage;
2404	PSUPDRVUSAGE pUsagePrev;
2405
2406	/*
2407	* Validate the input.
2408	*/
2409	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2410	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2411	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2412	VERR_INVALID_PARAMETER);
2413
2414	/*
2415	* Acquire the spinlock and look for the usage record.
2416	*/
2417	RTSpinlockAcquire(pDevExt->Spinlock);
2418
2419	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
2420	pUsage;
2421	pUsagePrev = pUsage, pUsage = pUsage->pNext)
2422	{
2423	/Log2(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2424	if (pUsage->pObj == pObj)
2425	{
2426	rc = VINF_SUCCESS;
2427	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
2428	if (pUsage->cUsage > 1)
2429	{
2430	pObj->cUsage--;
2431	pUsage->cUsage--;
2432	}
2433	else
2434	{
2435	/*
2436	* Free the session record.
2437	*/
2438	if (pUsagePrev)
2439	pUsagePrev->pNext = pUsage->pNext;
2440	else
2441	pSession->pUsage = pUsage->pNext;
2442	pUsage->pNext = pDevExt->pUsageFree;
2443	pDevExt->pUsageFree = pUsage;
2444
2445	/* What about the object? */
2446	if (pObj->cUsage > 1)
2447	pObj->cUsage--;
2448	else
2449	{
2450	/*
2451	* Object is to be destroyed, unlink it.
2452	*/
2453	pObj->u32Magic = SUPDRVOBJ_MAGIC_DEAD;
2454	rc = VINF_OBJECT_DESTROYED;
2455	if (pDevExt->pObjs == pObj)
2456	pDevExt->pObjs = pObj->pNext;
2457	else
2458	{
2459	PSUPDRVOBJ pObjPrev;
2460	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
2461	if (pObjPrev->pNext == pObj)
2462	{
2463	pObjPrev->pNext = pObj->pNext;
2464	break;
2465	}
2466	Assert(pObjPrev);
2467	}
2468	}
2469	}
2470	break;
2471	}
2472	}
2473
2474	RTSpinlockRelease(pDevExt->Spinlock);
2475
2476	/*
2477	* Call the destructor and free the object if required.
2478	*/
2479	if (rc == VINF_OBJECT_DESTROYED)
2480	{
2481	Log(("SUPR0ObjRelease: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
2482	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
2483	if (pObj->pfnDestructor)
2484	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
2485	RTMemFree(pObj);
2486	}
2487
2488	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
2489	return rc;
2490	}
2491
2492
2493	/**
2494	* Verifies that the current process can access the specified object.
2495	*
2496	* @returns The following IPRT status code:
2497	* @retval VINF_SUCCESS if access was granted.
2498	* @retval VERR_PERMISSION_DENIED if denied access.
2499	* @retval VERR_INVALID_PARAMETER if invalid parameter.
2500	*
2501	* @param pvObj The identifier returned by SUPR0ObjRegister().
2502	* @param pSession The session which wishes to access the object.
2503	* @param pszObjName Object string name. This is optional and depends on the object type.
2504	*
2505	* @remark The caller is responsible for making sure the object isn't removed while
2506	* we're inside this function. If uncertain about this, just call AddRef before calling us.
2507	*/
2508	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
2509	{
2510	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2511	int rc;
2512
2513	/*
2514	* Validate the input.
2515	*/
2516	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2517	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2518	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2519	VERR_INVALID_PARAMETER);
2520
2521	/*
2522	* Check access. (returns true if a decision has been made.)
2523	*/
2524	rc = VERR_INTERNAL_ERROR;
2525	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
2526	return rc;
2527
2528	/*
2529	* Default policy is to allow the user to access his own
2530	* stuff but nothing else.
2531	*/
2532	if (pObj->CreatorUid == pSession->Uid)
2533	return VINF_SUCCESS;
2534	return VERR_PERMISSION_DENIED;
2535	}
2536
2537
2538	/**
2539	* Lock pages.
2540	*
2541	* @returns IPRT status code.
2542	* @param pSession Session to which the locked memory should be associated.
2543	* @param pvR3 Start of the memory range to lock.
2544	* This must be page aligned.
2545	* @param cPages Number of pages to lock.
2546	* @param paPages Where to put the physical addresses of locked memory.
2547	*/
2548	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
2549	{
2550	int rc;
2551	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2552	const size_t cb = (size_t)cPages << PAGE_SHIFT;
2553	LogFlow(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
2554
2555	/*
2556	* Verify input.
2557	*/
2558	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2559	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
2560	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
2561	\|\| !pvR3)
2562	{
2563	Log(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
2564	return VERR_INVALID_PARAMETER;
2565	}
2566
2567	/*
2568	* Let IPRT do the job.
2569	*/
2570	Mem.eType = MEMREF_TYPE_LOCKED;
2571	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTMEM_PROT_READ \| RTMEM_PROT_WRITE, RTR0ProcHandleSelf());
2572	if (RT_SUCCESS(rc))
2573	{
2574	uint32_t iPage = cPages;
2575	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
2576	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
2577
2578	while (iPage-- > 0)
2579	{
2580	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2581	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
2582	{
2583	AssertMsgFailed(("iPage=%d\n", iPage));
2584	rc = VERR_INTERNAL_ERROR;
2585	break;
2586	}
2587	}
2588	if (RT_SUCCESS(rc))
2589	rc = supdrvMemAdd(&Mem, pSession);
2590	if (RT_FAILURE(rc))
2591	{
2592	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
2593	AssertRC(rc2);
2594	}
2595	}
2596
2597	return rc;
2598	}
2599
2600
2601	/**
2602	* Unlocks the memory pointed to by pv.
2603	*
2604	* @returns IPRT status code.
2605	* @param pSession Session to which the memory was locked.
2606	* @param pvR3 Memory to unlock.
2607	*/
2608	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2609	{
2610	LogFlow(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2611	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2612	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
2613	}
2614
2615
2616	/**
2617	* Allocates a chunk of page aligned memory with contiguous and fixed physical
2618	* backing.
2619	*
2620	* @returns IPRT status code.
2621	* @param pSession Session data.
2622	* @param cPages Number of pages to allocate.
2623	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
2624	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
2625	* @param pHCPhys Where to put the physical address of allocated memory.
2626	*/
2627	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
2628	{
2629	int rc;
2630	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2631	LogFlow(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
2632
2633	/*
2634	* Validate input.
2635	*/
2636	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2637	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
2638	{
2639	Log(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
2640	pSession, ppvR0, ppvR3, pHCPhys));
2641	return VERR_INVALID_PARAMETER;
2642
2643	}
2644	if (cPages < 1 \|\| cPages >= 256)
2645	{
2646	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2647	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2648	}
2649
2650	/*
2651	* Let IPRT do the job.
2652	*/
2653	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
2654	if (RT_SUCCESS(rc))
2655	{
2656	int rc2;
2657	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2658	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2659	if (RT_SUCCESS(rc))
2660	{
2661	Mem.eType = MEMREF_TYPE_CONT;
2662	rc = supdrvMemAdd(&Mem, pSession);
2663	if (!rc)
2664	{
2665	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2666	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2667	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
2668	return 0;
2669	}
2670
2671	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2672	AssertRC(rc2);
2673	}
2674	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2675	AssertRC(rc2);
2676	}
2677
2678	return rc;
2679	}
2680
2681
2682	/**
2683	* Frees memory allocated using SUPR0ContAlloc().
2684	*
2685	* @returns IPRT status code.
2686	* @param pSession The session to which the memory was allocated.
2687	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2688	*/
2689	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2690	{
2691	LogFlow(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2692	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2693	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
2694	}
2695
2696
2697	/**
2698	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
2699	*
2700	* The memory isn't zeroed.
2701	*
2702	* @returns IPRT status code.
2703	* @param pSession Session data.
2704	* @param cPages Number of pages to allocate.
2705	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
2706	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
2707	* @param paPages Where to put the physical addresses of allocated memory.
2708	*/
2709	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
2710	{
2711	unsigned iPage;
2712	int rc;
2713	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2714	LogFlow(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
2715
2716	/*
2717	* Validate input.
2718	*/
2719	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2720	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
2721	{
2722	Log(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
2723	pSession, ppvR3, ppvR0, paPages));
2724	return VERR_INVALID_PARAMETER;
2725
2726	}
2727	if (cPages < 1 \|\| cPages >= 256)
2728	{
2729	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2730	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2731	}
2732
2733	/*
2734	* Let IPRT do the work.
2735	*/
2736	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
2737	if (RT_SUCCESS(rc))
2738	{
2739	int rc2;
2740	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2741	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2742	if (RT_SUCCESS(rc))
2743	{
2744	Mem.eType = MEMREF_TYPE_LOW;
2745	rc = supdrvMemAdd(&Mem, pSession);
2746	if (!rc)
2747	{
2748	for (iPage = 0; iPage < cPages; iPage++)
2749	{
2750	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2751	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%RHp\n", paPages[iPage]));
2752	}
2753	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2754	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2755	return 0;
2756	}
2757
2758	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2759	AssertRC(rc2);
2760	}
2761
2762	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2763	AssertRC(rc2);
2764	}
2765
2766	return rc;
2767	}
2768
2769
2770	/**
2771	* Frees memory allocated using SUPR0LowAlloc().
2772	*
2773	* @returns IPRT status code.
2774	* @param pSession The session to which the memory was allocated.
2775	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2776	*/
2777	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2778	{
2779	LogFlow(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2780	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2781	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
2782	}
2783
2784
2785
2786	/**
2787	* Allocates a chunk of memory with both R0 and R3 mappings.
2788	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
2789	*
2790	* @returns IPRT status code.
2791	* @param pSession The session to associated the allocation with.
2792	* @param cb Number of bytes to allocate.
2793	* @param ppvR0 Where to store the address of the Ring-0 mapping.
2794	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2795	*/
2796	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
2797	{
2798	int rc;
2799	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2800	LogFlow(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
2801
2802	/*
2803	* Validate input.
2804	*/
2805	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2806	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
2807	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
2808	if (cb < 1 \|\| cb >= _4M)
2809	{
2810	Log(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
2811	return VERR_INVALID_PARAMETER;
2812	}
2813
2814	/*
2815	* Let IPRT do the work.
2816	*/
2817	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
2818	if (RT_SUCCESS(rc))
2819	{
2820	int rc2;
2821	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2822	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2823	if (RT_SUCCESS(rc))
2824	{
2825	Mem.eType = MEMREF_TYPE_MEM;
2826	rc = supdrvMemAdd(&Mem, pSession);
2827	if (!rc)
2828	{
2829	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2830	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2831	return VINF_SUCCESS;
2832	}
2833
2834	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2835	AssertRC(rc2);
2836	}
2837
2838	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2839	AssertRC(rc2);
2840	}
2841
2842	return rc;
2843	}
2844
2845
2846	/**
2847	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
2848	*
2849	* @returns IPRT status code.
2850	* @param pSession The session to which the memory was allocated.
2851	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2852	* @param paPages Where to store the physical addresses.
2853	*/
2854	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
2855	{
2856	PSUPDRVBUNDLE pBundle;
2857	LogFlow(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
2858
2859	/*
2860	* Validate input.
2861	*/
2862	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2863	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
2864	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
2865
2866	/*
2867	* Search for the address.
2868	*/
2869	RTSpinlockAcquire(pSession->Spinlock);
2870	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2871	{
2872	if (pBundle->cUsed > 0)
2873	{
2874	unsigned i;
2875	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2876	{
2877	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
2878	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2879	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2880	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2881	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
2882	)
2883	)
2884	{
2885	const size_t cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2886	size_t iPage;
2887	for (iPage = 0; iPage < cPages; iPage++)
2888	{
2889	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2890	paPages[iPage].uReserved = 0;
2891	}
2892	RTSpinlockRelease(pSession->Spinlock);
2893	return VINF_SUCCESS;
2894	}
2895	}
2896	}
2897	}
2898	RTSpinlockRelease(pSession->Spinlock);
2899	Log(("Failed to find %p!!!\n", (void *)uPtr));
2900	return VERR_INVALID_PARAMETER;
2901	}
2902
2903
2904	/**
2905	* Free memory allocated by SUPR0MemAlloc().
2906	*
2907	* @returns IPRT status code.
2908	* @param pSession The session owning the allocation.
2909	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2910	*/
2911	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2912	{
2913	LogFlow(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2914	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2915	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
2916	}
2917
2918
2919	/**
2920	* Allocates a chunk of memory with a kernel or/and a user mode mapping.
2921	*
2922	* The memory is fixed and it's possible to query the physical addresses using
2923	* SUPR0MemGetPhys().
2924	*
2925	* @returns IPRT status code.
2926	* @param pSession The session to associated the allocation with.
2927	* @param cPages The number of pages to allocate.
2928	* @param fFlags Flags, reserved for the future. Must be zero.
2929	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2930	* NULL if no ring-3 mapping.
2931	* @param ppvR3 Where to store the address of the Ring-0 mapping.
2932	* NULL if no ring-0 mapping.
2933	* @param paPages Where to store the addresses of the pages. Optional.
2934	*/
2935	SUPR0DECL(int) SUPR0PageAllocEx(PSUPDRVSESSION pSession, uint32_t cPages, uint32_t fFlags, PRTR3PTR ppvR3, PRTR0PTR ppvR0, PRTHCPHYS paPages)
2936	{
2937	int rc;
2938	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2939	LogFlow(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
2940
2941	/*
2942	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2943	*/
2944	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2945	AssertPtrNullReturn(ppvR3, VERR_INVALID_POINTER);
2946	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2947	AssertReturn(ppvR3 \|\| ppvR0, VERR_INVALID_PARAMETER);
2948	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2949	if (cPages < 1 \|\| cPages > VBOX_MAX_ALLOC_PAGE_COUNT)
2950	{
2951	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)));
2952	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2953	}
2954
2955	/*
2956	* Let IPRT do the work.
2957	*/
2958	if (ppvR0)
2959	rc = RTR0MemObjAllocPage(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, true /* fExecutable */);
2960	else
2961	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
2962	if (RT_SUCCESS(rc))
2963	{
2964	int rc2;
2965	if (ppvR3)
2966	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2967	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2968	else
2969	Mem.MapObjR3 = NIL_RTR0MEMOBJ;
2970	if (RT_SUCCESS(rc))
2971	{
2972	Mem.eType = MEMREF_TYPE_PAGE;
2973	rc = supdrvMemAdd(&Mem, pSession);
2974	if (!rc)
2975	{
2976	if (ppvR3)
2977	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2978	if (ppvR0)
2979	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2980	if (paPages)
2981	{
2982	uint32_t iPage = cPages;
2983	while (iPage-- > 0)
2984	{
2985	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
2986	Assert(paPages[iPage] != NIL_RTHCPHYS);
2987	}
2988	}
2989	return VINF_SUCCESS;
2990	}
2991
2992	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2993	AssertRC(rc2);
2994	}
2995
2996	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2997	AssertRC(rc2);
2998	}
2999	return rc;
3000	}
3001
3002
3003	/**
3004	* Maps a chunk of memory previously allocated by SUPR0PageAllocEx into kernel
3005	* space.
3006	*
3007	* @returns IPRT status code.
3008	* @param pSession The session to associated the allocation with.
3009	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
3010	* @param offSub Where to start mapping. Must be page aligned.
3011	* @param cbSub How much to map. Must be page aligned.
3012	* @param fFlags Flags, MBZ.
3013	* @param ppvR0 Where to return the address of the ring-0 mapping on
3014	* success.
3015	*/
3016	SUPR0DECL(int) SUPR0PageMapKernel(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t offSub, uint32_t cbSub,
3017	uint32_t fFlags, PRTR0PTR ppvR0)
3018	{
3019	int rc;
3020	PSUPDRVBUNDLE pBundle;
3021	RTR0MEMOBJ hMemObj = NIL_RTR0MEMOBJ;
3022	LogFlow(("SUPR0PageMapKernel: pSession=%p pvR3=%p offSub=%#x cbSub=%#x\n", pSession, pvR3, offSub, cbSub));
3023
3024	/*
3025	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
3026	*/
3027	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3028	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
3029	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
3030	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3031	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3032	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
3033
3034	/*
3035	* Find the memory object.
3036	*/
3037	RTSpinlockAcquire(pSession->Spinlock);
3038	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3039	{
3040	if (pBundle->cUsed > 0)
3041	{
3042	unsigned i;
3043	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3044	{
3045	if ( ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
3046	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3047	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3048	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
3049	\|\| ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED
3050	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3051	&& pBundle->aMem[i].MapObjR3 == NIL_RTR0MEMOBJ
3052	&& RTR0MemObjAddressR3(pBundle->aMem[i].MemObj) == pvR3))
3053	{
3054	hMemObj = pBundle->aMem[i].MemObj;
3055	break;
3056	}
3057	}
3058	}
3059	}
3060	RTSpinlockRelease(pSession->Spinlock);
3061
3062	rc = VERR_INVALID_PARAMETER;
3063	if (hMemObj != NIL_RTR0MEMOBJ)
3064	{
3065	/*
3066	* Do some further input validations before calling IPRT.
3067	* (Cleanup is done indirectly by telling RTR0MemObjFree to include mappings.)
3068	*/
3069	size_t cbMemObj = RTR0MemObjSize(hMemObj);
3070	if ( offSub < cbMemObj
3071	&& cbSub <= cbMemObj
3072	&& offSub + cbSub <= cbMemObj)
3073	{
3074	RTR0MEMOBJ hMapObj;
3075	rc = RTR0MemObjMapKernelEx(&hMapObj, hMemObj, (void *)-1, 0,
3076	RTMEM_PROT_READ \| RTMEM_PROT_WRITE, offSub, cbSub);
3077	if (RT_SUCCESS(rc))
3078	*ppvR0 = RTR0MemObjAddress(hMapObj);
3079	}
3080	else
3081	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
3082
3083	}
3084	return rc;
3085	}
3086
3087
3088	/**
3089	* Changes the page level protection of one or more pages previously allocated
3090	* by SUPR0PageAllocEx.
3091	*
3092	* @returns IPRT status code.
3093	* @param pSession The session to associated the allocation with.
3094	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
3095	* NIL_RTR3PTR if the ring-3 mapping should be unaffected.
3096	* @param pvR0 The ring-0 address returned by SUPR0PageAllocEx.
3097	* NIL_RTR0PTR if the ring-0 mapping should be unaffected.
3098	* @param offSub Where to start changing. Must be page aligned.
3099	* @param cbSub How much to change. Must be page aligned.
3100	* @param fProt The new page level protection, see RTMEM_PROT_*.
3101	*/
3102	SUPR0DECL(int) SUPR0PageProtect(PSUPDRVSESSION pSession, RTR3PTR pvR3, RTR0PTR pvR0, uint32_t offSub, uint32_t cbSub, uint32_t fProt)
3103	{
3104	int rc;
3105	PSUPDRVBUNDLE pBundle;
3106	RTR0MEMOBJ hMemObjR0 = NIL_RTR0MEMOBJ;
3107	RTR0MEMOBJ hMemObjR3 = NIL_RTR0MEMOBJ;
3108	LogFlow(("SUPR0PageProtect: pSession=%p pvR3=%p pvR0=%p offSub=%#x cbSub=%#x fProt-%#x\n", pSession, pvR3, pvR0, offSub, cbSub, fProt));
3109
3110	/*
3111	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
3112	*/
3113	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3114	AssertReturn(!(fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)), VERR_INVALID_PARAMETER);
3115	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3116	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3117	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
3118
3119	/*
3120	* Find the memory object.
3121	*/
3122	RTSpinlockAcquire(pSession->Spinlock);
3123	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3124	{
3125	if (pBundle->cUsed > 0)
3126	{
3127	unsigned i;
3128	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3129	{
3130	if ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
3131	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3132	&& ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3133	\|\| pvR3 == NIL_RTR3PTR)
3134	&& ( pvR0 == NIL_RTR0PTR
3135	\|\| RTR0MemObjAddress(pBundle->aMem[i].MemObj) == pvR0)
3136	&& ( pvR3 == NIL_RTR3PTR
3137	\|\| RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3))
3138	{
3139	if (pvR0 != NIL_RTR0PTR)
3140	hMemObjR0 = pBundle->aMem[i].MemObj;
3141	if (pvR3 != NIL_RTR3PTR)
3142	hMemObjR3 = pBundle->aMem[i].MapObjR3;
3143	break;
3144	}
3145	}
3146	}
3147	}
3148	RTSpinlockRelease(pSession->Spinlock);
3149
3150	rc = VERR_INVALID_PARAMETER;
3151	if ( hMemObjR0 != NIL_RTR0MEMOBJ
3152	\|\| hMemObjR3 != NIL_RTR0MEMOBJ)
3153	{
3154	/*
3155	* Do some further input validations before calling IPRT.
3156	*/
3157	size_t cbMemObj = hMemObjR0 != NIL_RTR0PTR ? RTR0MemObjSize(hMemObjR0) : RTR0MemObjSize(hMemObjR3);
3158	if ( offSub < cbMemObj
3159	&& cbSub <= cbMemObj
3160	&& offSub + cbSub <= cbMemObj)
3161	{
3162	rc = VINF_SUCCESS;
3163	if (hMemObjR3 != NIL_RTR0PTR)
3164	rc = RTR0MemObjProtect(hMemObjR3, offSub, cbSub, fProt);
3165	if (hMemObjR0 != NIL_RTR0PTR && RT_SUCCESS(rc))
3166	rc = RTR0MemObjProtect(hMemObjR0, offSub, cbSub, fProt);
3167	}
3168	else
3169	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
3170
3171	}
3172	return rc;
3173
3174	}
3175
3176
3177	/**
3178	* Free memory allocated by SUPR0PageAlloc() and SUPR0PageAllocEx().
3179	*
3180	* @returns IPRT status code.
3181	* @param pSession The session owning the allocation.
3182	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc() or
3183	* SUPR0PageAllocEx().
3184	*/
3185	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
3186	{
3187	LogFlow(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
3188	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3189	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_PAGE);
3190	}
3191
3192
3193	/**
3194	* Gets the paging mode of the current CPU.
3195	*
3196	* @returns Paging mode, SUPPAGEINGMODE_INVALID on error.
3197	*/
3198	SUPR0DECL(SUPPAGINGMODE) SUPR0GetPagingMode(void)
3199	{
3200	SUPPAGINGMODE enmMode;
3201
3202	RTR0UINTREG cr0 = ASMGetCR0();
3203	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
3204	enmMode = SUPPAGINGMODE_INVALID;
3205	else
3206	{
3207	RTR0UINTREG cr4 = ASMGetCR4();
3208	uint32_t fNXEPlusLMA = 0;
3209	if (cr4 & X86_CR4_PAE)
3210	{
3211	uint32_t fExtFeatures = ASMCpuId_EDX(0x80000001);
3212	if (fExtFeatures & (X86_CPUID_EXT_FEATURE_EDX_NX \| X86_CPUID_EXT_FEATURE_EDX_LONG_MODE))
3213	{
3214	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
3215	if ((fExtFeatures & X86_CPUID_EXT_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
3216	fNXEPlusLMA \|= RT_BIT(0);
3217	if ((fExtFeatures & X86_CPUID_EXT_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
3218	fNXEPlusLMA \|= RT_BIT(1);
3219	}
3220	}
3221
3222	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
3223	{
3224	case 0:
3225	enmMode = SUPPAGINGMODE_32_BIT;
3226	break;
3227
3228	case X86_CR4_PGE:
3229	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
3230	break;
3231
3232	case X86_CR4_PAE:
3233	enmMode = SUPPAGINGMODE_PAE;
3234	break;
3235
3236	case X86_CR4_PAE \| RT_BIT(0):
3237	enmMode = SUPPAGINGMODE_PAE_NX;
3238	break;
3239
3240	case X86_CR4_PAE \| X86_CR4_PGE:
3241	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3242	break;
3243
3244	case X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3245	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3246	break;
3247
3248	case RT_BIT(1) \| X86_CR4_PAE:
3249	enmMode = SUPPAGINGMODE_AMD64;
3250	break;
3251
3252	case RT_BIT(1) \| X86_CR4_PAE \| RT_BIT(0):
3253	enmMode = SUPPAGINGMODE_AMD64_NX;
3254	break;
3255
3256	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
3257	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
3258	break;
3259
3260	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3261	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
3262	break;
3263
3264	default:
3265	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
3266	enmMode = SUPPAGINGMODE_INVALID;
3267	break;
3268	}
3269	}
3270	return enmMode;
3271	}
3272
3273
3274	/**
3275	* Enables or disabled hardware virtualization extensions using native OS APIs.
3276	*
3277	* @returns VBox status code.
3278	* @retval VINF_SUCCESS on success.
3279	* @retval VERR_NOT_SUPPORTED if not supported by the native OS.
3280	*
3281	* @param fEnable Whether to enable or disable.
3282	*/
3283	SUPR0DECL(int) SUPR0EnableVTx(bool fEnable)
3284	{
3285	#ifdef RT_OS_DARWIN
3286	return supdrvOSEnableVTx(fEnable);
3287	#else
3288	return VERR_NOT_SUPPORTED;
3289	#endif
3290	}
3291
3292
3293	/**
3294	* Suspends hardware virtualization extensions using the native OS API.
3295	*
3296	* This is called prior to entering raw-mode context.
3297	*
3298	* @returns @c true if suspended, @c false if not.
3299	*/
3300	SUPR0DECL(bool) SUPR0SuspendVTxOnCpu(void)
3301	{
3302	#ifdef RT_OS_DARWIN
3303	return supdrvOSSuspendVTxOnCpu();
3304	#else
3305	return false;
3306	#endif
3307	}
3308
3309
3310	/**
3311	* Resumes hardware virtualization extensions using the native OS API.
3312	*
3313	* This is called after to entering raw-mode context.
3314	*
3315	* @param fSuspended The return value of SUPR0SuspendVTxOnCpu.
3316	*/
3317	SUPR0DECL(void) SUPR0ResumeVTxOnCpu(bool fSuspended)
3318	{
3319	#ifdef RT_OS_DARWIN
3320	supdrvOSResumeVTxOnCpu(fSuspended);
3321	#else
3322	Assert(!fSuspended);
3323	#endif
3324	}
3325
3326
3327	/**
3328	* Quries the AMD-V and VT-x capabilities of the calling CPU.
3329	*
3330	* @returns VBox status code.
3331	* @retval VERR_VMX_NO_VMX
3332	* @retval VERR_VMX_MSR_LOCKED_OR_DISABLED
3333	* @retval VERR_SVM_NO_SVM
3334	* @retval VERR_SVM_DISABLED
3335	* @retval VERR_UNSUPPORTED_CPU if not identifiable as an AMD, Intel or VIA
3336	* (centaur) CPU.
3337	*
3338	* @param pSession The session handle.
3339	* @param pfCaps Where to store the capabilities.
3340	*/
3341	SUPR0DECL(int) SUPR0QueryVTCaps(PSUPDRVSESSION pSession, uint32_t *pfCaps)
3342	{
3343	/*
3344	* Input validation.
3345	*/
3346	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3347	AssertPtrReturn(pfCaps, VERR_INVALID_POINTER);
3348
3349	*pfCaps = 0;
3350
3351	if (ASMHasCpuId())
3352	{
3353	uint32_t fFeaturesECX, fFeaturesEDX, uDummy;
3354	uint32_t uMaxId, uVendorEBX, uVendorECX, uVendorEDX;
3355	uint64_t u64Value;
3356
3357	ASMCpuId(0, &uMaxId, &uVendorEBX, &uVendorECX, &uVendorEDX);
3358	ASMCpuId(1, &uDummy, &uDummy, &fFeaturesECX, &fFeaturesEDX);
3359
3360	if ( ASMIsValidStdRange(uMaxId)
3361	&& ( ASMIsIntelCpuEx( uVendorEBX, uVendorECX, uVendorEDX)
3362	\|\| ASMIsViaCentaurCpuEx(uVendorEBX, uVendorECX, uVendorEDX) )
3363	)
3364	{
3365	if ( (fFeaturesECX & X86_CPUID_FEATURE_ECX_VMX)
3366	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3367	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3368	)
3369	{
3370	/*
3371	* Both the LOCK and VMXON bit must be set; otherwise VMXON will generate a #GP.
3372	* Once the lock bit is set, this MSR can no longer be modified.
3373	*/
3374	u64Value = ASMRdMsr(MSR_IA32_FEATURE_CONTROL);
3375	if ( (u64Value & (MSR_IA32_FEATURE_CONTROL_VMXON \| MSR_IA32_FEATURE_CONTROL_LOCK))
3376	== (MSR_IA32_FEATURE_CONTROL_VMXON \| MSR_IA32_FEATURE_CONTROL_LOCK) /* enabled and locked */
3377	\|\| !(u64Value & MSR_IA32_FEATURE_CONTROL_LOCK) /* not enabled, but not locked either */
3378	)
3379	{
3380	VMX_CAPABILITY vtCaps;
3381
3382	*pfCaps \|= SUPVTCAPS_VT_X;
3383
3384	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS);
3385	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC_USE_SECONDARY_EXEC_CTRL)
3386	{
3387	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS2);
3388	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC2_EPT)
3389	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3390	}
3391	return VINF_SUCCESS;
3392	}
3393	return VERR_VMX_MSR_LOCKED_OR_DISABLED;
3394	}
3395	return VERR_VMX_NO_VMX;
3396	}
3397
3398	if ( ASMIsAmdCpuEx(uVendorEBX, uVendorECX, uVendorEDX)
3399	&& ASMIsValidStdRange(uMaxId))
3400	{
3401	uint32_t fExtFeaturesEcx, uExtMaxId;
3402	ASMCpuId(0x80000000, &uExtMaxId, &uDummy, &uDummy, &uDummy);
3403	ASMCpuId(0x80000001, &uDummy, &uDummy, &fExtFeaturesEcx, &uDummy);
3404	if ( ASMIsValidExtRange(uExtMaxId)
3405	&& uExtMaxId >= 0x8000000a
3406	&& (fExtFeaturesEcx & X86_CPUID_AMD_FEATURE_ECX_SVM)
3407	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3408	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3409	)
3410	{
3411	/* Check if SVM is disabled */
3412	u64Value = ASMRdMsr(MSR_K8_VM_CR);
3413	if (!(u64Value & MSR_K8_VM_CR_SVM_DISABLE))
3414	{
3415	uint32_t fSvmFeatures;
3416	*pfCaps \|= SUPVTCAPS_AMD_V;
3417
3418	/* Query AMD-V features. */
3419	ASMCpuId(0x8000000a, &uDummy, &uDummy, &uDummy, &fSvmFeatures);
3420	if (fSvmFeatures & AMD_CPUID_SVM_FEATURE_EDX_NESTED_PAGING)
3421	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3422
3423	return VINF_SUCCESS;
3424	}
3425	return VERR_SVM_DISABLED;
3426	}
3427	return VERR_SVM_NO_SVM;
3428	}
3429	}
3430
3431	return VERR_UNSUPPORTED_CPU;
3432	}
3433
3434
3435	/**
3436	* (Re-)initializes the per-cpu structure prior to starting or resuming the GIP
3437	* updating.
3438	*
3439	* @param pGipCpu The per CPU structure for this CPU.
3440	* @param u64NanoTS The current time.
3441	*/
3442	static void supdrvGipReInitCpu(PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
3443	{
3444	pGipCpu->u64TSC = ASMReadTSC() - pGipCpu->u32UpdateIntervalTSC;
3445	pGipCpu->u64NanoTS = u64NanoTS;
3446	}
3447
3448
3449	/**
3450	* Set the current TSC and NanoTS value for the CPU.
3451	*
3452	* @param idCpu The CPU ID. Unused - we have to use the APIC ID.
3453	* @param pvUser1 Pointer to the ring-0 GIP mapping.
3454	* @param pvUser2 Pointer to the variable holding the current time.
3455	*/
3456	static DECLCALLBACK(void) supdrvGipReInitCpuCallback(RTCPUID idCpu, void pvUser1, void pvUser2)
3457	{
3458	PSUPGLOBALINFOPAGE pGip = (PSUPGLOBALINFOPAGE)pvUser1;
3459	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
3460
3461	if (RT_LIKELY(iCpu < pGip->cCpus && pGip->aCPUs[iCpu].idCpu == idCpu))
3462	supdrvGipReInitCpu(&pGip->aCPUs[iCpu], (uint64_t )pvUser2);
3463
3464	NOREF(pvUser2);
3465	NOREF(idCpu);
3466	}
3467
3468
3469	/**
3470	* Maps the GIP into userspace and/or get the physical address of the GIP.
3471	*
3472	* @returns IPRT status code.
3473	* @param pSession Session to which the GIP mapping should belong.
3474	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
3475	* @param pHCPhysGip Where to store the physical address. (optional)
3476	*
3477	* @remark There is no reference counting on the mapping, so one call to this function
3478	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
3479	* and remove the session as a GIP user.
3480	*/
3481	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
3482	{
3483	int rc;
3484	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3485	RTR3PTR pGipR3 = NIL_RTR3PTR;
3486	RTHCPHYS HCPhys = NIL_RTHCPHYS;
3487	LogFlow(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
3488
3489	/*
3490	* Validate
3491	*/
3492	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3493	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
3494	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
3495
3496	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3497	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3498	#else
3499	RTSemFastMutexRequest(pDevExt->mtxGip);
3500	#endif
3501	if (pDevExt->pGip)
3502	{
3503	/*
3504	* Map it?
3505	*/
3506	rc = VINF_SUCCESS;
3507	if (ppGipR3)
3508	{
3509	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
3510	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
3511	RTMEM_PROT_READ, RTR0ProcHandleSelf());
3512	if (RT_SUCCESS(rc))
3513	pGipR3 = RTR0MemObjAddressR3(pSession->GipMapObjR3);
3514	}
3515
3516	/*
3517	* Get physical address.
3518	*/
3519	if (pHCPhysGip && RT_SUCCESS(rc))
3520	HCPhys = pDevExt->HCPhysGip;
3521
3522	/*
3523	* Reference globally.
3524	*/
3525	if (!pSession->fGipReferenced && RT_SUCCESS(rc))
3526	{
3527	pSession->fGipReferenced = 1;
3528	pDevExt->cGipUsers++;
3529	if (pDevExt->cGipUsers == 1)
3530	{
3531	PSUPGLOBALINFOPAGE pGipR0 = pDevExt->pGip;
3532	uint64_t u64NanoTS;
3533	uint32_t u32SystemResolution;
3534	unsigned i;
3535
3536	LogFlow(("SUPR0GipMap: Resumes GIP updating\n"));
3537
3538	/*
3539	* Try bump up the system timer resolution.
3540	* The more interrupts the better...
3541	*/
3542	if ( RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
3543	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
3544	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1953125 /* 512 HZ */, &u32SystemResolution))
3545	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 2000000 /* 500 HZ */, &u32SystemResolution))
3546	)
3547	{
3548	Assert(RTTimerGetSystemGranularity() <= u32SystemResolution);
3549	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
3550	}
3551
3552	if (pGipR0->aCPUs[0].u32TransactionId != 2 /* not the first time */)
3553	{
3554	for (i = 0; i < RT_ELEMENTS(pGipR0->aCPUs); i++)
3555	ASMAtomicUoWriteU32(&pGipR0->aCPUs[i].u32TransactionId,
3556	(pGipR0->aCPUs[i].u32TransactionId + GIP_UPDATEHZ_RECALC_FREQ * 2)
3557	& ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
3558	ASMAtomicWriteU64(&pGipR0->u64NanoTSLastUpdateHz, 0);
3559	}
3560
3561	u64NanoTS = RTTimeSystemNanoTS() - pGipR0->u32UpdateIntervalNS;
3562	if ( pGipR0->u32Mode == SUPGIPMODE_SYNC_TSC
3563	\|\| RTMpGetOnlineCount() == 1)
3564	supdrvGipReInitCpu(&pGipR0->aCPUs[0], u64NanoTS);
3565	else
3566	RTMpOnAll(supdrvGipReInitCpuCallback, pGipR0, &u64NanoTS);
3567
3568	#ifndef DO_NOT_START_GIP
3569	rc = RTTimerStart(pDevExt->pGipTimer, 0); AssertRC(rc);
3570	#endif
3571	rc = VINF_SUCCESS;
3572	}
3573	}
3574	}
3575	else
3576	{
3577	rc = VERR_GENERAL_FAILURE;
3578	Log(("SUPR0GipMap: GIP is not available!\n"));
3579	}
3580	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3581	RTSemMutexRelease(pDevExt->mtxGip);
3582	#else
3583	RTSemFastMutexRelease(pDevExt->mtxGip);
3584	#endif
3585
3586	/*
3587	* Write returns.
3588	*/
3589	if (pHCPhysGip)
3590	*pHCPhysGip = HCPhys;
3591	if (ppGipR3)
3592	*ppGipR3 = pGipR3;
3593
3594	#ifdef DEBUG_DARWIN_GIP
3595	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3596	#else
3597	LogFlow(( "SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3598	#endif
3599	return rc;
3600	}
3601
3602
3603	/**
3604	* Unmaps any user mapping of the GIP and terminates all GIP access
3605	* from this session.
3606	*
3607	* @returns IPRT status code.
3608	* @param pSession Session to which the GIP mapping should belong.
3609	*/
3610	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
3611	{
3612	int rc = VINF_SUCCESS;
3613	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3614	#ifdef DEBUG_DARWIN_GIP
3615	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
3616	pSession,
3617	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
3618	pSession->GipMapObjR3));
3619	#else
3620	LogFlow(("SUPR0GipUnmap: pSession=%p\n", pSession));
3621	#endif
3622	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3623
3624	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3625	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3626	#else
3627	RTSemFastMutexRequest(pDevExt->mtxGip);
3628	#endif
3629
3630	/*
3631	* Unmap anything?
3632	*/
3633	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
3634	{
3635	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
3636	AssertRC(rc);
3637	if (RT_SUCCESS(rc))
3638	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
3639	}
3640
3641	/*
3642	* Dereference global GIP.
3643	*/
3644	if (pSession->fGipReferenced && !rc)
3645	{
3646	pSession->fGipReferenced = 0;
3647	if ( pDevExt->cGipUsers > 0
3648	&& !--pDevExt->cGipUsers)
3649	{
3650	LogFlow(("SUPR0GipUnmap: Suspends GIP updating\n"));
3651	#ifndef DO_NOT_START_GIP
3652	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = VINF_SUCCESS;
3653	#endif
3654
3655	if (pDevExt->u32SystemTimerGranularityGrant)
3656	{
3657	int rc2 = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant);
3658	AssertRC(rc2);
3659	pDevExt->u32SystemTimerGranularityGrant = 0;
3660	}
3661	}
3662	}
3663
3664	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3665	RTSemMutexRelease(pDevExt->mtxGip);
3666	#else
3667	RTSemFastMutexRelease(pDevExt->mtxGip);
3668	#endif
3669
3670	return rc;
3671	}
3672
3673
3674	/**
3675	* Gets the GIP pointer.
3676	*
3677	* @returns Pointer to the GIP or NULL.
3678	*/
3679	SUPDECL(PSUPGLOBALINFOPAGE) SUPGetGIP(void)
3680	{
3681	return g_pSUPGlobalInfoPage;
3682	}
3683
3684
3685	/**
3686	* Register a component factory with the support driver.
3687	*
3688	* This is currently restricted to kernel sessions only.
3689	*
3690	* @returns VBox status code.
3691	* @retval VINF_SUCCESS on success.
3692	* @retval VERR_NO_MEMORY if we're out of memory.
3693	* @retval VERR_ALREADY_EXISTS if the factory has already been registered.
3694	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3695	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3696	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3697	*
3698	* @param pSession The SUPDRV session (must be a ring-0 session).
3699	* @param pFactory Pointer to the component factory registration structure.
3700	*
3701	* @remarks This interface is also available via SUPR0IdcComponentRegisterFactory.
3702	*/
3703	SUPR0DECL(int) SUPR0ComponentRegisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3704	{
3705	PSUPDRVFACTORYREG pNewReg;
3706	const char *psz;
3707	int rc;
3708
3709	/*
3710	* Validate parameters.
3711	*/
3712	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3713	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3714	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3715	AssertPtrReturn(pFactory->pfnQueryFactoryInterface, VERR_INVALID_POINTER);
3716	psz = RTStrEnd(pFactory->szName, sizeof(pFactory->szName));
3717	AssertReturn(psz, VERR_INVALID_PARAMETER);
3718
3719	/*
3720	* Allocate and initialize a new registration structure.
3721	*/
3722	pNewReg = (PSUPDRVFACTORYREG)RTMemAlloc(sizeof(SUPDRVFACTORYREG));
3723	if (pNewReg)
3724	{
3725	pNewReg->pNext = NULL;
3726	pNewReg->pFactory = pFactory;
3727	pNewReg->pSession = pSession;
3728	pNewReg->cchName = psz - &pFactory->szName[0];
3729
3730	/*
3731	* Add it to the tail of the list after checking for prior registration.
3732	*/
3733	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3734	if (RT_SUCCESS(rc))
3735	{
3736	PSUPDRVFACTORYREG pPrev = NULL;
3737	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3738	while (pCur && pCur->pFactory != pFactory)
3739	{
3740	pPrev = pCur;
3741	pCur = pCur->pNext;
3742	}
3743	if (!pCur)
3744	{
3745	if (pPrev)
3746	pPrev->pNext = pNewReg;
3747	else
3748	pSession->pDevExt->pComponentFactoryHead = pNewReg;
3749	rc = VINF_SUCCESS;
3750	}
3751	else
3752	rc = VERR_ALREADY_EXISTS;
3753
3754	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3755	}
3756
3757	if (RT_FAILURE(rc))
3758	RTMemFree(pNewReg);
3759	}
3760	else
3761	rc = VERR_NO_MEMORY;
3762	return rc;
3763	}
3764
3765
3766	/**
3767	* Deregister a component factory.
3768	*
3769	* @returns VBox status code.
3770	* @retval VINF_SUCCESS on success.
3771	* @retval VERR_NOT_FOUND if the factory wasn't registered.
3772	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3773	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3774	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3775	*
3776	* @param pSession The SUPDRV session (must be a ring-0 session).
3777	* @param pFactory Pointer to the component factory registration structure
3778	* previously passed SUPR0ComponentRegisterFactory().
3779	*
3780	* @remarks This interface is also available via SUPR0IdcComponentDeregisterFactory.
3781	*/
3782	SUPR0DECL(int) SUPR0ComponentDeregisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3783	{
3784	int rc;
3785
3786	/*
3787	* Validate parameters.
3788	*/
3789	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3790	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3791	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3792
3793	/*
3794	* Take the lock and look for the registration record.
3795	*/
3796	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3797	if (RT_SUCCESS(rc))
3798	{
3799	PSUPDRVFACTORYREG pPrev = NULL;
3800	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3801	while (pCur && pCur->pFactory != pFactory)
3802	{
3803	pPrev = pCur;
3804	pCur = pCur->pNext;
3805	}
3806	if (pCur)
3807	{
3808	if (!pPrev)
3809	pSession->pDevExt->pComponentFactoryHead = pCur->pNext;
3810	else
3811	pPrev->pNext = pCur->pNext;
3812
3813	pCur->pNext = NULL;
3814	pCur->pFactory = NULL;
3815	pCur->pSession = NULL;
3816	rc = VINF_SUCCESS;
3817	}
3818	else
3819	rc = VERR_NOT_FOUND;
3820
3821	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3822
3823	RTMemFree(pCur);
3824	}
3825	return rc;
3826	}
3827
3828
3829	/**
3830	* Queries a component factory.
3831	*
3832	* @returns VBox status code.
3833	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3834	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3835	* @retval VERR_SUPDRV_COMPONENT_NOT_FOUND if the component factory wasn't found.
3836	* @retval VERR_SUPDRV_INTERFACE_NOT_SUPPORTED if the interface wasn't supported.
3837	*
3838	* @param pSession The SUPDRV session.
3839	* @param pszName The name of the component factory.
3840	* @param pszInterfaceUuid The UUID of the factory interface (stringified).
3841	* @param ppvFactoryIf Where to store the factory interface.
3842	*/
3843	SUPR0DECL(int) SUPR0ComponentQueryFactory(PSUPDRVSESSION pSession, const char pszName, const char pszInterfaceUuid, void **ppvFactoryIf)
3844	{
3845	const char *pszEnd;
3846	size_t cchName;
3847	int rc;
3848
3849	/*
3850	* Validate parameters.
3851	*/
3852	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3853
3854	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
3855	pszEnd = RTStrEnd(pszName, RT_SIZEOFMEMB(SUPDRVFACTORY, szName));
3856	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3857	cchName = pszEnd - pszName;
3858
3859	AssertPtrReturn(pszInterfaceUuid, VERR_INVALID_POINTER);
3860	pszEnd = RTStrEnd(pszInterfaceUuid, RTUUID_STR_LENGTH);
3861	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3862
3863	AssertPtrReturn(ppvFactoryIf, VERR_INVALID_POINTER);
3864	*ppvFactoryIf = NULL;
3865
3866	/*
3867	* Take the lock and try all factories by this name.
3868	*/
3869	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3870	if (RT_SUCCESS(rc))
3871	{
3872	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3873	rc = VERR_SUPDRV_COMPONENT_NOT_FOUND;
3874	while (pCur)
3875	{
3876	if ( pCur->cchName == cchName
3877	&& !memcmp(pCur->pFactory->szName, pszName, cchName))
3878	{
3879	void *pvFactory = pCur->pFactory->pfnQueryFactoryInterface(pCur->pFactory, pSession, pszInterfaceUuid);
3880	if (pvFactory)
3881	{
3882	*ppvFactoryIf = pvFactory;
3883	rc = VINF_SUCCESS;
3884	break;
3885	}
3886	rc = VERR_SUPDRV_INTERFACE_NOT_SUPPORTED;
3887	}
3888
3889	/* next */
3890	pCur = pCur->pNext;
3891	}
3892
3893	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3894	}
3895	return rc;
3896	}
3897
3898
3899	/**
3900	* Adds a memory object to the session.
3901	*
3902	* @returns IPRT status code.
3903	* @param pMem Memory tracking structure containing the
3904	* information to track.
3905	* @param pSession The session.
3906	*/
3907	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
3908	{
3909	PSUPDRVBUNDLE pBundle;
3910
3911	/*
3912	* Find free entry and record the allocation.
3913	*/
3914	RTSpinlockAcquire(pSession->Spinlock);
3915	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3916	{
3917	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
3918	{
3919	unsigned i;
3920	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3921	{
3922	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
3923	{
3924	pBundle->cUsed++;
3925	pBundle->aMem[i] = *pMem;
3926	RTSpinlockRelease(pSession->Spinlock);
3927	return VINF_SUCCESS;
3928	}
3929	}
3930	AssertFailed(); /* !!this can't be happening!!! */
3931	}
3932	}
3933	RTSpinlockRelease(pSession->Spinlock);
3934
3935	/*
3936	* Need to allocate a new bundle.
3937	* Insert into the last entry in the bundle.
3938	*/
3939	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
3940	if (!pBundle)
3941	return VERR_NO_MEMORY;
3942
3943	/* take last entry. */
3944	pBundle->cUsed++;
3945	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
3946
3947	/* insert into list. */
3948	RTSpinlockAcquire(pSession->Spinlock);
3949	pBundle->pNext = pSession->Bundle.pNext;
3950	pSession->Bundle.pNext = pBundle;
3951	RTSpinlockRelease(pSession->Spinlock);
3952
3953	return VINF_SUCCESS;
3954	}
3955
3956
3957	/**
3958	* Releases a memory object referenced by pointer and type.
3959	*
3960	* @returns IPRT status code.
3961	* @param pSession Session data.
3962	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
3963	* @param eType Memory type.
3964	*/
3965	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
3966	{
3967	PSUPDRVBUNDLE pBundle;
3968
3969	/*
3970	* Validate input.
3971	*/
3972	if (!uPtr)
3973	{
3974	Log(("Illegal address %p\n", (void *)uPtr));
3975	return VERR_INVALID_PARAMETER;
3976	}
3977
3978	/*
3979	* Search for the address.
3980	*/
3981	RTSpinlockAcquire(pSession->Spinlock);
3982	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3983	{
3984	if (pBundle->cUsed > 0)
3985	{
3986	unsigned i;
3987	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3988	{
3989	if ( pBundle->aMem[i].eType == eType
3990	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3991	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
3992	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3993	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
3994	)
3995	{
3996	/* Make a copy of it and release it outside the spinlock. */
3997	SUPDRVMEMREF Mem = pBundle->aMem[i];
3998	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
3999	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
4000	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
4001	RTSpinlockRelease(pSession->Spinlock);
4002
4003	if (Mem.MapObjR3 != NIL_RTR0MEMOBJ)
4004	{
4005	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
4006	AssertRC(rc); /** @todo figure out how to handle this. */
4007	}
4008	if (Mem.MemObj != NIL_RTR0MEMOBJ)
4009	{
4010	int rc = RTR0MemObjFree(Mem.MemObj, true /* fFreeMappings */);
4011	AssertRC(rc); /** @todo figure out how to handle this. */
4012	}
4013	return VINF_SUCCESS;
4014	}
4015	}
4016	}
4017	}
4018	RTSpinlockRelease(pSession->Spinlock);
4019	Log(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
4020	return VERR_INVALID_PARAMETER;
4021	}
4022
4023
4024	/**
4025	* Opens an image. If it's the first time it's opened the call must upload
4026	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
4027	*
4028	* This is the 1st step of the loading.
4029	*
4030	* @returns IPRT status code.
4031	* @param pDevExt Device globals.
4032	* @param pSession Session data.
4033	* @param pReq The open request.
4034	*/
4035	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
4036	{
4037	int rc;
4038	PSUPDRVLDRIMAGE pImage;
4039	void *pv;
4040	size_t cchName = strlen(pReq->u.In.szName); /* (caller checked < 32). */
4041	LogFlow(("supdrvIOCtl_LdrOpen: szName=%s cbImageWithTabs=%d\n", pReq->u.In.szName, pReq->u.In.cbImageWithTabs));
4042
4043	/*
4044	* Check if we got an instance of the image already.
4045	*/
4046	supdrvLdrLock(pDevExt);
4047	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
4048	{
4049	if ( pImage->szName[cchName] == '\0'
4050	&& !memcmp(pImage->szName, pReq->u.In.szName, cchName))
4051	{
4052	/** @todo check cbImageBits and cbImageWithTabs here, if they differs that indicates that the images are different. */
4053	pImage->cUsage++;
4054	pReq->u.Out.pvImageBase = pImage->pvImage;
4055	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
4056	pReq->u.Out.fNativeLoader = pImage->fNative;
4057	supdrvLdrAddUsage(pSession, pImage);
4058	supdrvLdrUnlock(pDevExt);
4059	return VINF_SUCCESS;
4060	}
4061	}
4062	/* (not found - add it!) */
4063
4064	/*
4065	* Allocate memory.
4066	*/
4067	Assert(cchName < sizeof(pImage->szName));
4068	pv = RTMemAlloc(sizeof(SUPDRVLDRIMAGE));
4069	if (!pv)
4070	{
4071	supdrvLdrUnlock(pDevExt);
4072	Log(("supdrvIOCtl_LdrOpen: RTMemAlloc() failed\n"));
4073	return /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_2;
4074	}
4075
4076	/*
4077	* Setup and link in the LDR stuff.
4078	*/
4079	pImage = (PSUPDRVLDRIMAGE)pv;
4080	pImage->pvImage = NULL;
4081	pImage->pvImageAlloc = NULL;
4082	pImage->cbImageWithTabs = pReq->u.In.cbImageWithTabs;
4083	pImage->cbImageBits = pReq->u.In.cbImageBits;
4084	pImage->cSymbols = 0;
4085	pImage->paSymbols = NULL;
4086	pImage->pachStrTab = NULL;
4087	pImage->cbStrTab = 0;
4088	pImage->pfnModuleInit = NULL;
4089	pImage->pfnModuleTerm = NULL;
4090	pImage->pfnServiceReqHandler = NULL;
4091	pImage->uState = SUP_IOCTL_LDR_OPEN;
4092	pImage->cUsage = 1;
4093	pImage->pDevExt = pDevExt;
4094	memcpy(pImage->szName, pReq->u.In.szName, cchName + 1);
4095
4096	/*
4097	* Try load it using the native loader, if that isn't supported, fall back
4098	* on the older method.
4099	*/
4100	pImage->fNative = true;
4101	rc = supdrvOSLdrOpen(pDevExt, pImage, pReq->u.In.szFilename);
4102	if (rc == VERR_NOT_SUPPORTED)
4103	{
4104	pImage->pvImageAlloc = RTMemExecAlloc(pImage->cbImageBits + 31);
4105	pImage->pvImage = RT_ALIGN_P(pImage->pvImageAlloc, 32);
4106	pImage->fNative = false;
4107	rc = pImage->pvImageAlloc ? VINF_SUCCESS : VERR_NO_EXEC_MEMORY;
4108	}
4109	if (RT_FAILURE(rc))
4110	{
4111	supdrvLdrUnlock(pDevExt);
4112	RTMemFree(pImage);
4113	Log(("supdrvIOCtl_LdrOpen(%s): failed - %Rrc\n", pReq->u.In.szName, rc));
4114	return rc;
4115	}
4116	Assert(VALID_PTR(pImage->pvImage) \|\| RT_FAILURE(rc));
4117
4118	/*
4119	* Link it.
4120	*/
4121	pImage->pNext = pDevExt->pLdrImages;
4122	pDevExt->pLdrImages = pImage;
4123
4124	supdrvLdrAddUsage(pSession, pImage);
4125
4126	pReq->u.Out.pvImageBase = pImage->pvImage;
4127	pReq->u.Out.fNeedsLoading = true;
4128	pReq->u.Out.fNativeLoader = pImage->fNative;
4129	supdrvOSLdrNotifyOpened(pDevExt, pImage);
4130
4131	supdrvLdrUnlock(pDevExt);
4132	return VINF_SUCCESS;
4133	}
4134
4135
4136	/**
4137	* Worker that validates a pointer to an image entrypoint.
4138	*
4139	* @returns IPRT status code.
4140	* @param pDevExt The device globals.
4141	* @param pImage The loader image.
4142	* @param pv The pointer into the image.
4143	* @param fMayBeNull Whether it may be NULL.
4144	* @param pszWhat What is this entrypoint? (for logging)
4145	* @param pbImageBits The image bits prepared by ring-3.
4146	*
4147	* @remarks Will leave the lock on failure.
4148	*/
4149	static int supdrvLdrValidatePointer(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage, void *pv,
4150	bool fMayBeNull, const uint8_t pbImageBits, const char pszWhat)
4151	{
4152	if (!fMayBeNull \|\| pv)
4153	{
4154	if ((uintptr_t)pv - (uintptr_t)pImage->pvImage >= pImage->cbImageBits)
4155	{
4156	supdrvLdrUnlock(pDevExt);
4157	Log(("Out of range (%p LB %#x): %s=%p\n", pImage->pvImage, pImage->cbImageBits, pszWhat, pv));
4158	return VERR_INVALID_PARAMETER;
4159	}
4160
4161	if (pImage->fNative)
4162	{
4163	int rc = supdrvOSLdrValidatePointer(pDevExt, pImage, pv, pbImageBits);
4164	if (RT_FAILURE(rc))
4165	{
4166	supdrvLdrUnlock(pDevExt);
4167	Log(("Bad entry point address: %s=%p (rc=%Rrc)\n", pszWhat, pv, rc));
4168	return rc;
4169	}
4170	}
4171	}
4172	return VINF_SUCCESS;
4173	}
4174
4175
4176	/**
4177	* Loads the image bits.
4178	*
4179	* This is the 2nd step of the loading.
4180	*
4181	* @returns IPRT status code.
4182	* @param pDevExt Device globals.
4183	* @param pSession Session data.
4184	* @param pReq The request.
4185	*/
4186	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
4187	{
4188	PSUPDRVLDRUSAGE pUsage;
4189	PSUPDRVLDRIMAGE pImage;
4190	int rc;
4191	LogFlow(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImageWithBits=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImageWithTabs));
4192
4193	/*
4194	* Find the ldr image.
4195	*/
4196	supdrvLdrLock(pDevExt);
4197	pUsage = pSession->pLdrUsage;
4198	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4199	pUsage = pUsage->pNext;
4200	if (!pUsage)
4201	{
4202	supdrvLdrUnlock(pDevExt);
4203	Log(("SUP_IOCTL_LDR_LOAD: couldn't find image!\n"));
4204	return VERR_INVALID_HANDLE;
4205	}
4206	pImage = pUsage->pImage;
4207
4208	/*
4209	* Validate input.
4210	*/
4211	if ( pImage->cbImageWithTabs != pReq->u.In.cbImageWithTabs
4212	\|\| pImage->cbImageBits != pReq->u.In.cbImageBits)
4213	{
4214	supdrvLdrUnlock(pDevExt);
4215	Log(("SUP_IOCTL_LDR_LOAD: image size mismatch!! %d(prep) != %d(load) or %d != %d\n",
4216	pImage->cbImageWithTabs, pReq->u.In.cbImageWithTabs, pImage->cbImageBits, pReq->u.In.cbImageBits));
4217	return VERR_INVALID_HANDLE;
4218	}
4219
4220	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
4221	{
4222	unsigned uState = pImage->uState;
4223	supdrvLdrUnlock(pDevExt);
4224	if (uState != SUP_IOCTL_LDR_LOAD)
4225	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
4226	return VERR_ALREADY_LOADED;
4227	}
4228
4229	switch (pReq->u.In.eEPType)
4230	{
4231	case SUPLDRLOADEP_NOTHING:
4232	break;
4233
4234	case SUPLDRLOADEP_VMMR0:
4235	rc = supdrvLdrValidatePointer( pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0, false, pReq->u.In.abImage, "pvVMMR0");
4236	if (RT_SUCCESS(rc))
4237	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt, false, pReq->u.In.abImage, "pvVMMR0EntryInt");
4238	if (RT_SUCCESS(rc))
4239	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, false, pReq->u.In.abImage, "pvVMMR0EntryFast");
4240	if (RT_SUCCESS(rc))
4241	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx, false, pReq->u.In.abImage, "pvVMMR0EntryEx");
4242	if (RT_FAILURE(rc))
4243	return rc;
4244	break;
4245
4246	case SUPLDRLOADEP_SERVICE:
4247	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.Service.pfnServiceReq, false, pReq->u.In.abImage, "pfnServiceReq");
4248	if (RT_FAILURE(rc))
4249	return rc;
4250	if ( pReq->u.In.EP.Service.apvReserved[0] != NIL_RTR0PTR
4251	\|\| pReq->u.In.EP.Service.apvReserved[1] != NIL_RTR0PTR
4252	\|\| pReq->u.In.EP.Service.apvReserved[2] != NIL_RTR0PTR)
4253	{
4254	supdrvLdrUnlock(pDevExt);
4255	Log(("Out of range (%p LB %#x): apvReserved={%p,%p,%p} MBZ!\n",
4256	pImage->pvImage, pReq->u.In.cbImageWithTabs,
4257	pReq->u.In.EP.Service.apvReserved[0],
4258	pReq->u.In.EP.Service.apvReserved[1],
4259	pReq->u.In.EP.Service.apvReserved[2]));
4260	return VERR_INVALID_PARAMETER;
4261	}
4262	break;
4263
4264	default:
4265	supdrvLdrUnlock(pDevExt);
4266	Log(("Invalid eEPType=%d\n", pReq->u.In.eEPType));
4267	return VERR_INVALID_PARAMETER;
4268	}
4269
4270	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleInit, true, pReq->u.In.abImage, "pfnModuleInit");
4271	if (RT_FAILURE(rc))
4272	return rc;
4273	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleTerm, true, pReq->u.In.abImage, "pfnModuleTerm");
4274	if (RT_FAILURE(rc))
4275	return rc;
4276
4277	/*
4278	* Allocate and copy the tables.
4279	* (No need to do try/except as this is a buffered request.)
4280	*/
4281	pImage->cbStrTab = pReq->u.In.cbStrTab;
4282	if (pImage->cbStrTab)
4283	{
4284	pImage->pachStrTab = (char *)RTMemAlloc(pImage->cbStrTab);
4285	if (pImage->pachStrTab)
4286	memcpy(pImage->pachStrTab, &pReq->u.In.abImage[pReq->u.In.offStrTab], pImage->cbStrTab);
4287	else
4288	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_3;
4289	}
4290
4291	pImage->cSymbols = pReq->u.In.cSymbols;
4292	if (RT_SUCCESS(rc) && pImage->cSymbols)
4293	{
4294	size_t cbSymbols = pImage->cSymbols * sizeof(SUPLDRSYM);
4295	pImage->paSymbols = (PSUPLDRSYM)RTMemAlloc(cbSymbols);
4296	if (pImage->paSymbols)
4297	memcpy(pImage->paSymbols, &pReq->u.In.abImage[pReq->u.In.offSymbols], cbSymbols);
4298	else
4299	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_4;
4300	}
4301
4302	/*
4303	* Copy the bits / complete native loading.
4304	*/
4305	if (RT_SUCCESS(rc))
4306	{
4307	pImage->uState = SUP_IOCTL_LDR_LOAD;
4308	pImage->pfnModuleInit = pReq->u.In.pfnModuleInit;
4309	pImage->pfnModuleTerm = pReq->u.In.pfnModuleTerm;
4310
4311	if (pImage->fNative)
4312	rc = supdrvOSLdrLoad(pDevExt, pImage, pReq->u.In.abImage, pReq);
4313	else
4314	{
4315	memcpy(pImage->pvImage, &pReq->u.In.abImage[0], pImage->cbImageBits);
4316	Log(("vboxdrv: Loaded '%s' at %p\n", pImage->szName, pImage->pvImage));
4317	}
4318	}
4319
4320	/*
4321	* Update any entry points.
4322	*/
4323	if (RT_SUCCESS(rc))
4324	{
4325	switch (pReq->u.In.eEPType)
4326	{
4327	default:
4328	case SUPLDRLOADEP_NOTHING:
4329	rc = VINF_SUCCESS;
4330	break;
4331	case SUPLDRLOADEP_VMMR0:
4332	rc = supdrvLdrSetVMMR0EPs(pDevExt, pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
4333	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
4334	break;
4335	case SUPLDRLOADEP_SERVICE:
4336	pImage->pfnServiceReqHandler = pReq->u.In.EP.Service.pfnServiceReq;
4337	rc = VINF_SUCCESS;
4338	break;
4339	}
4340	}
4341
4342	/*
4343	* On success call the module initialization.
4344	*/
4345	LogFlow(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
4346	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
4347	{
4348	Log(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
4349	pDevExt->pLdrInitImage = pImage;
4350	pDevExt->hLdrInitThread = RTThreadNativeSelf();
4351	rc = pImage->pfnModuleInit(pImage);
4352	pDevExt->pLdrInitImage = NULL;
4353	pDevExt->hLdrInitThread = NIL_RTNATIVETHREAD;
4354	if (RT_FAILURE(rc) && pDevExt->pvVMMR0 == pImage->pvImage)
4355	supdrvLdrUnsetVMMR0EPs(pDevExt);
4356	}
4357
4358	if (RT_FAILURE(rc))
4359	{
4360	/* Inform the tracing component in case ModuleInit registered TPs. */
4361	supdrvTracerModuleUnloading(pDevExt, pImage);
4362
4363	pImage->uState = SUP_IOCTL_LDR_OPEN;
4364	pImage->pfnModuleInit = NULL;
4365	pImage->pfnModuleTerm = NULL;
4366	pImage->pfnServiceReqHandler= NULL;
4367	pImage->cbStrTab = 0;
4368	RTMemFree(pImage->pachStrTab);
4369	pImage->pachStrTab = NULL;
4370	RTMemFree(pImage->paSymbols);
4371	pImage->paSymbols = NULL;
4372	pImage->cSymbols = 0;
4373	}
4374
4375	supdrvLdrUnlock(pDevExt);
4376	return rc;
4377	}
4378
4379
4380	/**
4381	* Frees a previously loaded (prep'ed) image.
4382	*
4383	* @returns IPRT status code.
4384	* @param pDevExt Device globals.
4385	* @param pSession Session data.
4386	* @param pReq The request.
4387	*/
4388	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
4389	{
4390	int rc;
4391	PSUPDRVLDRUSAGE pUsagePrev;
4392	PSUPDRVLDRUSAGE pUsage;
4393	PSUPDRVLDRIMAGE pImage;
4394	LogFlow(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
4395
4396	/*
4397	* Find the ldr image.
4398	*/
4399	supdrvLdrLock(pDevExt);
4400	pUsagePrev = NULL;
4401	pUsage = pSession->pLdrUsage;
4402	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4403	{
4404	pUsagePrev = pUsage;
4405	pUsage = pUsage->pNext;
4406	}
4407	if (!pUsage)
4408	{
4409	supdrvLdrUnlock(pDevExt);
4410	Log(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
4411	return VERR_INVALID_HANDLE;
4412	}
4413
4414	/*
4415	* Check if we can remove anything.
4416	*/
4417	rc = VINF_SUCCESS;
4418	pImage = pUsage->pImage;
4419	if (pImage->cUsage <= 1 \|\| pUsage->cUsage <= 1)
4420	{
4421	/*
4422	* Check if there are any objects with destructors in the image, if
4423	* so leave it for the session cleanup routine so we get a chance to
4424	* clean things up in the right order and not leave them all dangling.
4425	*/
4426	RTSpinlockAcquire(pDevExt->Spinlock);
4427	if (pImage->cUsage <= 1)
4428	{
4429	PSUPDRVOBJ pObj;
4430	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4431	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4432	{
4433	rc = VERR_DANGLING_OBJECTS;
4434	break;
4435	}
4436	}
4437	else
4438	{
4439	PSUPDRVUSAGE pGenUsage;
4440	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
4441	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4442	{
4443	rc = VERR_DANGLING_OBJECTS;
4444	break;
4445	}
4446	}
4447	RTSpinlockRelease(pDevExt->Spinlock);
4448	if (rc == VINF_SUCCESS)
4449	{
4450	/* unlink it */
4451	if (pUsagePrev)
4452	pUsagePrev->pNext = pUsage->pNext;
4453	else
4454	pSession->pLdrUsage = pUsage->pNext;
4455
4456	/* free it */
4457	pUsage->pImage = NULL;
4458	pUsage->pNext = NULL;
4459	RTMemFree(pUsage);
4460
4461	/*
4462	* Dereference the image.
4463	*/
4464	if (pImage->cUsage <= 1)
4465	supdrvLdrFree(pDevExt, pImage);
4466	else
4467	pImage->cUsage--;
4468	}
4469	else
4470	{
4471	Log(("supdrvIOCtl_LdrFree: Dangling objects in %p/%s!\n", pImage->pvImage, pImage->szName));
4472	rc = VINF_SUCCESS; /** @todo BRANCH-2.1: remove this after branching. */
4473	}
4474	}
4475	else
4476	{
4477	/*
4478	* Dereference both image and usage.
4479	*/
4480	pImage->cUsage--;
4481	pUsage->cUsage--;
4482	}
4483
4484	supdrvLdrUnlock(pDevExt);
4485	return rc;
4486	}
4487
4488
4489	/**
4490	* Gets the address of a symbol in an open image.
4491	*
4492	* @returns IPRT status code.
4493	* @param pDevExt Device globals.
4494	* @param pSession Session data.
4495	* @param pReq The request buffer.
4496	*/
4497	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
4498	{
4499	PSUPDRVLDRIMAGE pImage;
4500	PSUPDRVLDRUSAGE pUsage;
4501	uint32_t i;
4502	PSUPLDRSYM paSyms;
4503	const char *pchStrings;
4504	const size_t cbSymbol = strlen(pReq->u.In.szSymbol) + 1;
4505	void *pvSymbol = NULL;
4506	int rc = VERR_GENERAL_FAILURE;
4507	Log3(("supdrvIOCtl_LdrGetSymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
4508
4509	/*
4510	* Find the ldr image.
4511	*/
4512	supdrvLdrLock(pDevExt);
4513	pUsage = pSession->pLdrUsage;
4514	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4515	pUsage = pUsage->pNext;
4516	if (!pUsage)
4517	{
4518	supdrvLdrUnlock(pDevExt);
4519	Log(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
4520	return VERR_INVALID_HANDLE;
4521	}
4522	pImage = pUsage->pImage;
4523	if (pImage->uState != SUP_IOCTL_LDR_LOAD)
4524	{
4525	unsigned uState = pImage->uState;
4526	supdrvLdrUnlock(pDevExt);
4527	Log(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", uState, uState)); NOREF(uState);
4528	return VERR_ALREADY_LOADED;
4529	}
4530
4531	/*
4532	* Search the symbol strings.
4533	*
4534	* Note! The int32_t is for native loading on solaris where the data
4535	* and text segments are in very different places.
4536	*/
4537	pchStrings = pImage->pachStrTab;
4538	paSyms = pImage->paSymbols;
4539	for (i = 0; i < pImage->cSymbols; i++)
4540	{
4541	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4542	&& !memcmp(pchStrings + paSyms[i].offName, pReq->u.In.szSymbol, cbSymbol))
4543	{
4544	pvSymbol = (uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol;
4545	rc = VINF_SUCCESS;
4546	break;
4547	}
4548	}
4549	supdrvLdrUnlock(pDevExt);
4550	pReq->u.Out.pvSymbol = pvSymbol;
4551	return rc;
4552	}
4553
4554
4555	/**
4556	* Gets the address of a symbol in an open image or the support driver.
4557	*
4558	* @returns VINF_SUCCESS on success.
4559	* @returns
4560	* @param pDevExt Device globals.
4561	* @param pSession Session data.
4562	* @param pReq The request buffer.
4563	*/
4564	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq)
4565	{
4566	int rc = VINF_SUCCESS;
4567	const char *pszSymbol = pReq->u.In.pszSymbol;
4568	const char *pszModule = pReq->u.In.pszModule;
4569	size_t cbSymbol;
4570	char const *pszEnd;
4571	uint32_t i;
4572
4573	/*
4574	* Input validation.
4575	*/
4576	AssertPtrReturn(pszSymbol, VERR_INVALID_POINTER);
4577	pszEnd = RTStrEnd(pszSymbol, 512);
4578	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4579	cbSymbol = pszEnd - pszSymbol + 1;
4580
4581	if (pszModule)
4582	{
4583	AssertPtrReturn(pszModule, VERR_INVALID_POINTER);
4584	pszEnd = RTStrEnd(pszModule, 64);
4585	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4586	}
4587	Log3(("supdrvIDC_LdrGetSymbol: pszModule=%p:{%s} pszSymbol=%p:{%s}\n", pszModule, pszModule, pszSymbol, pszSymbol));
4588
4589
4590	if ( !pszModule
4591	\|\| !strcmp(pszModule, "SupDrv"))
4592	{
4593	/*
4594	* Search the support driver export table.
4595	*/
4596	for (i = 0; i < RT_ELEMENTS(g_aFunctions); i++)
4597	if (!strcmp(g_aFunctions[i].szName, pszSymbol))
4598	{
4599	pReq->u.Out.pfnSymbol = g_aFunctions[i].pfn;
4600	break;
4601	}
4602	}
4603	else
4604	{
4605	/*
4606	* Find the loader image.
4607	*/
4608	PSUPDRVLDRIMAGE pImage;
4609
4610	supdrvLdrLock(pDevExt);
4611
4612	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
4613	if (!strcmp(pImage->szName, pszModule))
4614	break;
4615	if (pImage && pImage->uState == SUP_IOCTL_LDR_LOAD)
4616	{
4617	/*
4618	* Search the symbol strings.
4619	*/
4620	const char *pchStrings = pImage->pachStrTab;
4621	PCSUPLDRSYM paSyms = pImage->paSymbols;
4622	for (i = 0; i < pImage->cSymbols; i++)
4623	{
4624	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4625	&& !memcmp(pchStrings + paSyms[i].offName, pszSymbol, cbSymbol))
4626	{
4627	/*
4628	* Found it! Calc the symbol address and add a reference to the module.
4629	*/
4630	pReq->u.Out.pfnSymbol = (PFNRT)((uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol);
4631	rc = supdrvLdrAddUsage(pSession, pImage);
4632	break;
4633	}
4634	}
4635	}
4636	else
4637	rc = pImage ? VERR_WRONG_ORDER : VERR_MODULE_NOT_FOUND;
4638
4639	supdrvLdrUnlock(pDevExt);
4640	}
4641	return rc;
4642	}
4643
4644
4645	/**
4646	* Updates the VMMR0 entry point pointers.
4647	*
4648	* @returns IPRT status code.
4649	* @param pDevExt Device globals.
4650	* @param pSession Session data.
4651	* @param pVMMR0 VMMR0 image handle.
4652	* @param pvVMMR0EntryInt VMMR0EntryInt address.
4653	* @param pvVMMR0EntryFast VMMR0EntryFast address.
4654	* @param pvVMMR0EntryEx VMMR0EntryEx address.
4655	* @remark Caller must own the loader mutex.
4656	*/
4657	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx)
4658	{
4659	int rc = VINF_SUCCESS;
4660	LogFlow(("supdrvLdrSetR0EP pvVMMR0=%p pvVMMR0EntryInt=%p\n", pvVMMR0, pvVMMR0EntryInt));
4661
4662
4663	/*
4664	* Check if not yet set.
4665	*/
4666	if (!pDevExt->pvVMMR0)
4667	{
4668	pDevExt->pvVMMR0 = pvVMMR0;
4669	pDevExt->pfnVMMR0EntryInt = pvVMMR0EntryInt;
4670	pDevExt->pfnVMMR0EntryFast = pvVMMR0EntryFast;
4671	pDevExt->pfnVMMR0EntryEx = pvVMMR0EntryEx;
4672	}
4673	else
4674	{
4675	/*
4676	* Return failure or success depending on whether the values match or not.
4677	*/
4678	if ( pDevExt->pvVMMR0 != pvVMMR0
4679	\|\| (void *)pDevExt->pfnVMMR0EntryInt != pvVMMR0EntryInt
4680	\|\| (void *)pDevExt->pfnVMMR0EntryFast != pvVMMR0EntryFast
4681	\|\| (void *)pDevExt->pfnVMMR0EntryEx != pvVMMR0EntryEx)
4682	{
4683	AssertMsgFailed(("SUP_IOCTL_LDR_SETR0EP: Already set pointing to a different module!\n"));
4684	rc = VERR_INVALID_PARAMETER;
4685	}
4686	}
4687	return rc;
4688	}
4689
4690
4691	/**
4692	* Unsets the VMMR0 entry point installed by supdrvLdrSetR0EP.
4693	*
4694	* @param pDevExt Device globals.
4695	*/
4696	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt)
4697	{
4698	pDevExt->pvVMMR0 = NULL;
4699	pDevExt->pfnVMMR0EntryInt = NULL;
4700	pDevExt->pfnVMMR0EntryFast = NULL;
4701	pDevExt->pfnVMMR0EntryEx = NULL;
4702	}
4703
4704
4705	/**
4706	* Adds a usage reference in the specified session of an image.
4707	*
4708	* Called while owning the loader semaphore.
4709	*
4710	* @returns VINF_SUCCESS on success and VERR_NO_MEMORY on failure.
4711	* @param pSession Session in question.
4712	* @param pImage Image which the session is using.
4713	*/
4714	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage)
4715	{
4716	PSUPDRVLDRUSAGE pUsage;
4717	LogFlow(("supdrvLdrAddUsage: pImage=%p\n", pImage));
4718
4719	/*
4720	* Referenced it already?
4721	*/
4722	pUsage = pSession->pLdrUsage;
4723	while (pUsage)
4724	{
4725	if (pUsage->pImage == pImage)
4726	{
4727	pUsage->cUsage++;
4728	return VINF_SUCCESS;
4729	}
4730	pUsage = pUsage->pNext;
4731	}
4732
4733	/*
4734	* Allocate new usage record.
4735	*/
4736	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
4737	AssertReturn(pUsage, /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_5);
4738	pUsage->cUsage = 1;
4739	pUsage->pImage = pImage;
4740	pUsage->pNext = pSession->pLdrUsage;
4741	pSession->pLdrUsage = pUsage;
4742	return VINF_SUCCESS;
4743	}
4744
4745
4746	/**
4747	* Frees a load image.
4748	*
4749	* @param pDevExt Pointer to device extension.
4750	* @param pImage Pointer to the image we're gonna free.
4751	* This image must exit!
4752	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
4753	*/
4754	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
4755	{
4756	PSUPDRVLDRIMAGE pImagePrev;
4757	LogFlow(("supdrvLdrFree: pImage=%p\n", pImage));
4758
4759	/* find it - arg. should've used doubly linked list. */
4760	Assert(pDevExt->pLdrImages);
4761	pImagePrev = NULL;
4762	if (pDevExt->pLdrImages != pImage)
4763	{
4764	pImagePrev = pDevExt->pLdrImages;
4765	while (pImagePrev->pNext != pImage)
4766	pImagePrev = pImagePrev->pNext;
4767	Assert(pImagePrev->pNext == pImage);
4768	}
4769
4770	/* unlink */
4771	if (pImagePrev)
4772	pImagePrev->pNext = pImage->pNext;
4773	else
4774	pDevExt->pLdrImages = pImage->pNext;
4775
4776	/* check if this is VMMR0.r0 unset its entry point pointers. */
4777	if (pDevExt->pvVMMR0 == pImage->pvImage)
4778	supdrvLdrUnsetVMMR0EPs(pDevExt);
4779
4780	/* check for objects with destructors in this image. (Shouldn't happen.) */
4781	if (pDevExt->pObjs)
4782	{
4783	unsigned cObjs = 0;
4784	PSUPDRVOBJ pObj;
4785	RTSpinlockAcquire(pDevExt->Spinlock);
4786	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4787	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4788	{
4789	pObj->pfnDestructor = NULL;
4790	cObjs++;
4791	}
4792	RTSpinlockRelease(pDevExt->Spinlock);
4793	if (cObjs)
4794	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
4795	}
4796
4797	/* call termination function if fully loaded. */
4798	if ( pImage->pfnModuleTerm
4799	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
4800	{
4801	LogFlow(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
4802	pImage->pfnModuleTerm(pImage);
4803	}
4804
4805	/* Inform the tracing component. */
4806	supdrvTracerModuleUnloading(pDevExt, pImage);
4807
4808	/* do native unload if appropriate. */
4809	if (pImage->fNative)
4810	supdrvOSLdrUnload(pDevExt, pImage);
4811
4812	/* free the image */
4813	pImage->cUsage = 0;
4814	pImage->pDevExt = NULL;
4815	pImage->pNext = NULL;
4816	pImage->uState = SUP_IOCTL_LDR_FREE;
4817	RTMemExecFree(pImage->pvImageAlloc, pImage->cbImageBits + 31);
4818	pImage->pvImageAlloc = NULL;
4819	RTMemFree(pImage->pachStrTab);
4820	pImage->pachStrTab = NULL;
4821	RTMemFree(pImage->paSymbols);
4822	pImage->paSymbols = NULL;
4823	RTMemFree(pImage);
4824	}
4825
4826
4827	/**
4828	* Acquires the loader lock.
4829	*
4830	* @returns IPRT status code.
4831	* @param pDevExt The device extension.
4832	*/
4833	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt)
4834	{
4835	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4836	int rc = RTSemMutexRequest(pDevExt->mtxLdr, RT_INDEFINITE_WAIT);
4837	#else
4838	int rc = RTSemFastMutexRequest(pDevExt->mtxLdr);
4839	#endif
4840	AssertRC(rc);
4841	return rc;
4842	}
4843
4844
4845	/**
4846	* Releases the loader lock.
4847	*
4848	* @returns IPRT status code.
4849	* @param pDevExt The device extension.
4850	*/
4851	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt)
4852	{
4853	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4854	return RTSemMutexRelease(pDevExt->mtxLdr);
4855	#else
4856	return RTSemFastMutexRelease(pDevExt->mtxLdr);
4857	#endif
4858	}
4859
4860
4861	/**
4862	* Implements the service call request.
4863	*
4864	* @returns VBox status code.
4865	* @param pDevExt The device extension.
4866	* @param pSession The calling session.
4867	* @param pReq The request packet, valid.
4868	*/
4869	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq)
4870	{
4871	#if !defined(RT_OS_WINDOWS) \|\| defined(RT_ARCH_AMD64) \|\| defined(DEBUG)
4872	int rc;
4873
4874	/*
4875	* Find the module first in the module referenced by the calling session.
4876	*/
4877	rc = supdrvLdrLock(pDevExt);
4878	if (RT_SUCCESS(rc))
4879	{
4880	PFNSUPR0SERVICEREQHANDLER pfnServiceReqHandler = NULL;
4881	PSUPDRVLDRUSAGE pUsage;
4882
4883	for (pUsage = pSession->pLdrUsage; pUsage; pUsage = pUsage->pNext)
4884	if ( pUsage->pImage->pfnServiceReqHandler
4885	&& !strcmp(pUsage->pImage->szName, pReq->u.In.szName))
4886	{
4887	pfnServiceReqHandler = pUsage->pImage->pfnServiceReqHandler;
4888	break;
4889	}
4890	supdrvLdrUnlock(pDevExt);
4891
4892	if (pfnServiceReqHandler)
4893	{
4894	/*
4895	* Call it.
4896	*/
4897	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
4898	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, NULL);
4899	else
4900	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0]);
4901	}
4902	else
4903	rc = VERR_SUPDRV_SERVICE_NOT_FOUND;
4904	}
4905
4906	/* log it */
4907	if ( RT_FAILURE(rc)
4908	&& rc != VERR_INTERRUPTED
4909	&& rc != VERR_TIMEOUT)
4910	Log(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4911	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4912	else
4913	Log4(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4914	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4915	return rc;
4916	#else /* RT_OS_WINDOWS && !RT_ARCH_AMD64 && !DEBUG */
4917	return VERR_NOT_IMPLEMENTED;
4918	#endif /* RT_OS_WINDOWS && !RT_ARCH_AMD64 && !DEBUG */
4919	}
4920
4921
4922	/**
4923	* Implements the logger settings request.
4924	*
4925	* @returns VBox status code.
4926	* @param pDevExt The device extension.
4927	* @param pSession The caller's session.
4928	* @param pReq The request.
4929	*/
4930	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq)
4931	{
4932	const char *pszGroup = &pReq->u.In.szStrings[pReq->u.In.offGroups];
4933	const char *pszFlags = &pReq->u.In.szStrings[pReq->u.In.offFlags];
4934	const char *pszDest = &pReq->u.In.szStrings[pReq->u.In.offDestination];
4935	PRTLOGGER pLogger = NULL;
4936	int rc;
4937
4938	/*
4939	* Some further validation.
4940	*/
4941	switch (pReq->u.In.fWhat)
4942	{
4943	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4944	case SUPLOGGERSETTINGS_WHAT_CREATE:
4945	break;
4946
4947	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4948	if (pszGroup \|\| pszFlags \|\| *pszDest)
4949	return VERR_INVALID_PARAMETER;
4950	if (pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_RELEASE)
4951	return VERR_ACCESS_DENIED;
4952	break;
4953
4954	default:
4955	return VERR_INTERNAL_ERROR;
4956	}
4957
4958	/*
4959	* Get the logger.
4960	*/
4961	switch (pReq->u.In.fWhich)
4962	{
4963	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4964	pLogger = RTLogGetDefaultInstance();
4965	break;
4966
4967	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4968	pLogger = RTLogRelDefaultInstance();
4969	break;
4970
4971	default:
4972	return VERR_INTERNAL_ERROR;
4973	}
4974
4975	/*
4976	* Do the job.
4977	*/
4978	switch (pReq->u.In.fWhat)
4979	{
4980	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4981	if (pLogger)
4982	{
4983	rc = RTLogFlags(pLogger, pszFlags);
4984	if (RT_SUCCESS(rc))
4985	rc = RTLogGroupSettings(pLogger, pszGroup);
4986	NOREF(pszDest);
4987	}
4988	else
4989	rc = VERR_NOT_FOUND;
4990	break;
4991
4992	case SUPLOGGERSETTINGS_WHAT_CREATE:
4993	{
4994	if (pLogger)
4995	rc = VERR_ALREADY_EXISTS;
4996	else
4997	{
4998	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
4999
5000	rc = RTLogCreate(&pLogger,
5001	0 /* fFlags */,
5002	pszGroup,
5003	pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_DEBUG
5004	? "VBOX_LOG"
5005	: "VBOX_RELEASE_LOG",
5006	RT_ELEMENTS(s_apszGroups),
5007	s_apszGroups,
5008	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER,
5009	NULL);
5010	if (RT_SUCCESS(rc))
5011	{
5012	rc = RTLogFlags(pLogger, pszFlags);
5013	NOREF(pszDest);
5014	if (RT_SUCCESS(rc))
5015	{
5016	switch (pReq->u.In.fWhich)
5017	{
5018	case SUPLOGGERSETTINGS_WHICH_DEBUG:
5019	pLogger = RTLogSetDefaultInstance(pLogger);
5020	break;
5021	case SUPLOGGERSETTINGS_WHICH_RELEASE:
5022	pLogger = RTLogRelSetDefaultInstance(pLogger);
5023	break;
5024	}
5025	}
5026	RTLogDestroy(pLogger);
5027	}
5028	}
5029	break;
5030	}
5031
5032	case SUPLOGGERSETTINGS_WHAT_DESTROY:
5033	switch (pReq->u.In.fWhich)
5034	{
5035	case SUPLOGGERSETTINGS_WHICH_DEBUG:
5036	pLogger = RTLogSetDefaultInstance(NULL);
5037	break;
5038	case SUPLOGGERSETTINGS_WHICH_RELEASE:
5039	pLogger = RTLogRelSetDefaultInstance(NULL);
5040	break;
5041	}
5042	rc = RTLogDestroy(pLogger);
5043	break;
5044
5045	default:
5046	{
5047	rc = VERR_INTERNAL_ERROR;
5048	break;
5049	}
5050	}
5051
5052	return rc;
5053	}
5054
5055
5056	/**
5057	* Creates the GIP.
5058	*
5059	* @returns VBox status code.
5060	* @param pDevExt Instance data. GIP stuff may be updated.
5061	*/
5062	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
5063	{
5064	PSUPGLOBALINFOPAGE pGip;
5065	RTHCPHYS HCPhysGip;
5066	uint32_t u32SystemResolution;
5067	uint32_t u32Interval;
5068	unsigned cCpus;
5069	int rc;
5070
5071
5072	LogFlow(("supdrvGipCreate:\n"));
5073
5074	/* assert order */
5075	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
5076	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
5077	Assert(!pDevExt->pGipTimer);
5078
5079	/*
5080	* Check the CPU count.
5081	*/
5082	cCpus = RTMpGetArraySize();
5083	if ( cCpus > RTCPUSET_MAX_CPUS
5084	\|\| cCpus > 256 /ApicId is used for the mappings/)
5085	{
5086	SUPR0Printf("VBoxDrv: Too many CPUs (%u) for the GIP (max %u)\n", cCpus, RT_MIN(RTCPUSET_MAX_CPUS, 256));
5087	return VERR_TOO_MANY_CPUS;
5088	}
5089
5090	/*
5091	* Allocate a contiguous set of pages with a default kernel mapping.
5092	*/
5093	rc = RTR0MemObjAllocCont(&pDevExt->GipMemObj, RT_UOFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), false /fExecutable/);
5094	if (RT_FAILURE(rc))
5095	{
5096	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
5097	return rc;
5098	}
5099	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
5100	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
5101
5102	/*
5103	* Find a reasonable update interval and initialize the structure.
5104	*/
5105	u32Interval = u32SystemResolution = RTTimerGetSystemGranularity();
5106	while (u32Interval < 10000000 /* 10 ms */)
5107	u32Interval += u32SystemResolution;
5108
5109	supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), 1000000000 / u32Interval /=Hz/, cCpus);
5110
5111	/*
5112	* Create the timer.
5113	* If CPU_ALL isn't supported we'll have to fall back to synchronous mode.
5114	*/
5115	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
5116	{
5117	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, RTTIMER_FLAGS_CPU_ALL, supdrvGipAsyncTimer, pDevExt);
5118	if (rc == VERR_NOT_SUPPORTED)
5119	{
5120	OSDBGPRINT(("supdrvGipCreate: omni timer not supported, falling back to synchronous mode\n"));
5121	pGip->u32Mode = SUPGIPMODE_SYNC_TSC;
5122	}
5123	}
5124	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
5125	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0, supdrvGipSyncTimer, pDevExt);
5126	if (RT_SUCCESS(rc))
5127	{
5128	rc = RTMpNotificationRegister(supdrvGipMpEvent, pDevExt);
5129	if (RT_SUCCESS(rc))
5130	{
5131	rc = RTMpOnAll(supdrvGipInitOnCpu, pDevExt, pGip);
5132	if (RT_SUCCESS(rc))
5133	{
5134	/*
5135	* We're good.
5136	*/
5137	Log(("supdrvGipCreate: %u ns interval.\n", u32Interval));
5138	g_pSUPGlobalInfoPage = pGip;
5139	return VINF_SUCCESS;
5140	}
5141
5142	OSDBGPRINT(("supdrvGipCreate: RTMpOnAll failed with rc=%Rrc\n", rc));
5143	RTMpNotificationDeregister(supdrvGipMpEvent, pDevExt);
5144
5145	}
5146	else
5147	OSDBGPRINT(("supdrvGipCreate: failed to register MP event notfication. rc=%Rrc\n", rc));
5148	}
5149	else
5150	{
5151	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %u ns interval. rc=%Rrc\n", u32Interval, rc));
5152	Assert(!pDevExt->pGipTimer);
5153	}
5154	supdrvGipDestroy(pDevExt);
5155	return rc;
5156	}
5157
5158
5159	/**
5160	* Terminates the GIP.
5161	*
5162	* @param pDevExt Instance data. GIP stuff may be updated.
5163	*/
5164	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
5165	{
5166	int rc;
5167	#ifdef DEBUG_DARWIN_GIP
5168	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
5169	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
5170	pDevExt->pGipTimer, pDevExt->GipMemObj));
5171	#endif
5172
5173	/*
5174	* Invalid the GIP data.
5175	*/
5176	if (pDevExt->pGip)
5177	{
5178	supdrvGipTerm(pDevExt->pGip);
5179	pDevExt->pGip = NULL;
5180	}
5181	g_pSUPGlobalInfoPage = NULL;
5182
5183	/*
5184	* Destroy the timer and free the GIP memory object.
5185	*/
5186	if (pDevExt->pGipTimer)
5187	{
5188	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
5189	pDevExt->pGipTimer = NULL;
5190	}
5191
5192	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
5193	{
5194	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
5195	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
5196	}
5197
5198	/*
5199	* Finally, make sure we've release the system timer resolution request
5200	* if one actually succeeded and is still pending.
5201	*/
5202	if (pDevExt->u32SystemTimerGranularityGrant)
5203	{
5204	rc = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
5205	pDevExt->u32SystemTimerGranularityGrant = 0;
5206	}
5207	}
5208
5209
5210	/**
5211	* Timer callback function sync GIP mode.
5212	* @param pTimer The timer.
5213	* @param pvUser The device extension.
5214	*/
5215	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
5216	{
5217	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
5218	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5219	uint64_t u64TSC = ASMReadTSC();
5220	uint64_t NanoTS = RTTimeSystemNanoTS();
5221
5222	supdrvGipUpdate(pDevExt, NanoTS, u64TSC, NIL_RTCPUID, iTick);
5223
5224	ASMSetFlags(fOldFlags);
5225	}
5226
5227
5228	/**
5229	* Timer callback function for async GIP mode.
5230	* @param pTimer The timer.
5231	* @param pvUser The device extension.
5232	*/
5233	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
5234	{
5235	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
5236	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5237	RTCPUID idCpu = RTMpCpuId();
5238	uint64_t u64TSC = ASMReadTSC();
5239	uint64_t NanoTS = RTTimeSystemNanoTS();
5240
5241	/** @todo reset the transaction number and whatnot when iTick == 1. */
5242	if (pDevExt->idGipMaster == idCpu)
5243	supdrvGipUpdate(pDevExt, NanoTS, u64TSC, idCpu, iTick);
5244	else
5245	supdrvGipUpdatePerCpu(pDevExt, NanoTS, u64TSC, idCpu, ASMGetApicId(), iTick);
5246
5247	ASMSetFlags(fOldFlags);
5248	}
5249
5250
5251	/**
5252	* Finds our (@a idCpu) entry, or allocates a new one if not found.
5253	*
5254	* @returns Index of the CPU in the cache set.
5255	* @param pGip The GIP.
5256	* @param idCpu The CPU ID.
5257	*/
5258	static uint32_t supdrvGipCpuIndexFromCpuId(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
5259	{
5260	uint32_t i, cTries;
5261
5262	/*
5263	* ASSUMES that CPU IDs are constant.
5264	*/
5265	for (i = 0; i < pGip->cCpus; i++)
5266	if (pGip->aCPUs[i].idCpu == idCpu)
5267	return i;
5268
5269	cTries = 0;
5270	do
5271	{
5272	for (i = 0; i < pGip->cCpus; i++)
5273	{
5274	bool fRc;
5275	ASMAtomicCmpXchgSize(&pGip->aCPUs[i].idCpu, idCpu, NIL_RTCPUID, fRc);
5276	if (fRc)
5277	return i;
5278	}
5279	} while (cTries++ < 32);
5280	AssertReleaseFailed();
5281	return i - 1;
5282	}
5283
5284
5285	/**
5286	* The calling CPU should be accounted as online, update GIP accordingly.
5287	*
5288	* This is used by supdrvGipMpEvent as well as the supdrvGipCreate.
5289	*
5290	* @param pDevExt The device extension.
5291	* @param idCpu The CPU ID.
5292	*/
5293	static void supdrvGipMpEventOnline(PSUPDRVDEVEXT pDevExt, RTCPUID idCpu)
5294	{
5295	int iCpuSet = 0;
5296	uint16_t idApic = UINT16_MAX;
5297	uint32_t i = 0;
5298	uint64_t u64NanoTS = 0;
5299	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5300
5301	AssertPtrReturnVoid(pGip);
5302	AssertRelease(idCpu == RTMpCpuId());
5303	Assert(pGip->cPossibleCpus == RTMpGetCount());
5304
5305	/*
5306	* Do this behind a spinlock with interrupts disabled as this can fire
5307	* on all CPUs simultaneously, see @bugref{6110}.
5308	*/
5309	RTSpinlockAcquire(pDevExt->hGipSpinlock);
5310
5311	/*
5312	* Update the globals.
5313	*/
5314	ASMAtomicWriteU16(&pGip->cPresentCpus, RTMpGetPresentCount());
5315	ASMAtomicWriteU16(&pGip->cOnlineCpus, RTMpGetOnlineCount());
5316	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5317	if (iCpuSet >= 0)
5318	{
5319	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5320	RTCpuSetAddByIndex(&pGip->OnlineCpuSet, iCpuSet);
5321	RTCpuSetAddByIndex(&pGip->PresentCpuSet, iCpuSet);
5322	}
5323
5324	/*
5325	* Update the entry.
5326	*/
5327	u64NanoTS = RTTimeSystemNanoTS() - pGip->u32UpdateIntervalNS;
5328	i = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
5329	supdrvGipInitCpu(pGip, &pGip->aCPUs[i], u64NanoTS);
5330	idApic = ASMGetApicId();
5331	ASMAtomicWriteU16(&pGip->aCPUs[i].idApic, idApic);
5332	ASMAtomicWriteS16(&pGip->aCPUs[i].iCpuSet, (int16_t)iCpuSet);
5333	ASMAtomicWriteSize(&pGip->aCPUs[i].idCpu, idCpu);
5334
5335	/*
5336	* Update the APIC ID and CPU set index mappings.
5337	*/
5338	ASMAtomicWriteU16(&pGip->aiCpuFromApicId[idApic], i);
5339	ASMAtomicWriteU16(&pGip->aiCpuFromCpuSetIdx[iCpuSet], i);
5340
5341	/* commit it */
5342	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_ONLINE);
5343
5344	RTSpinlockReleaseNoInts(pDevExt->hGipSpinlock);
5345	}
5346
5347
5348	/**
5349	* The CPU should be accounted as offline, update the GIP accordingly.
5350	*
5351	* This is used by supdrvGipMpEvent.
5352	*
5353	* @param pDevExt The device extension.
5354	* @param idCpu The CPU ID.
5355	*/
5356	static void supdrvGipMpEventOffline(PSUPDRVDEVEXT pDevExt, RTCPUID idCpu)
5357	{
5358	int iCpuSet;
5359	unsigned i;
5360
5361	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5362
5363	AssertPtrReturnVoid(pGip);
5364	RTSpinlockAcquire(pDevExt->hGipSpinlock);
5365
5366	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5367	AssertReturnVoid(iCpuSet >= 0);
5368
5369	i = pGip->aiCpuFromCpuSetIdx[iCpuSet];
5370	AssertReturnVoid(i < pGip->cCpus);
5371	AssertReturnVoid(pGip->aCPUs[i].idCpu == idCpu);
5372
5373	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5374	RTCpuSetDelByIndex(&pGip->OnlineCpuSet, iCpuSet);
5375
5376	/* commit it */
5377	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_OFFLINE);
5378
5379	RTSpinlockReleaseNoInts(pDevExt->hGipSpinlock);
5380	}
5381
5382
5383	/**
5384	* Multiprocessor event notification callback.
5385	*
5386	* This is used to make sure that the GIP master gets passed on to
5387	* another CPU. It also updates the associated CPU data.
5388	*
5389	* @param enmEvent The event.
5390	* @param idCpu The cpu it applies to.
5391	* @param pvUser Pointer to the device extension.
5392	*
5393	* @remarks This function -must- fire on the newly online'd CPU for the
5394	* RTMPEVENT_ONLINE case and can fire on any CPU for the
5395	* RTMPEVENT_OFFLINE case.
5396	*/
5397	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser)
5398	{
5399	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5400	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5401
5402	AssertRelease(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
5403
5404	/*
5405	* Update the GIP CPU data.
5406	*/
5407	if (pGip)
5408	{
5409	switch (enmEvent)
5410	{
5411	case RTMPEVENT_ONLINE:
5412	AssertRelease(idCpu == RTMpCpuId());
5413	supdrvGipMpEventOnline(pDevExt, idCpu);
5414	break;
5415	case RTMPEVENT_OFFLINE:
5416	supdrvGipMpEventOffline(pDevExt, idCpu);
5417	break;
5418
5419	}
5420	}
5421
5422	/*
5423	* Make sure there is a master GIP.
5424	*/
5425	if (enmEvent == RTMPEVENT_OFFLINE)
5426	{
5427	RTCPUID idGipMaster = ASMAtomicReadU32(&pDevExt->idGipMaster);
5428	if (idGipMaster == idCpu)
5429	{
5430	/*
5431	* Find a new GIP master.
5432	*/
5433	bool fIgnored;
5434	unsigned i;
5435	RTCPUID idNewGipMaster = NIL_RTCPUID;
5436	RTCPUSET OnlineCpus;
5437	RTMpGetOnlineSet(&OnlineCpus);
5438
5439	for (i = 0; i < RTCPUSET_MAX_CPUS; i++)
5440	{
5441	RTCPUID idCurCpu = RTMpCpuIdFromSetIndex(i);
5442	if ( RTCpuSetIsMember(&OnlineCpus, idCurCpu)
5443	&& idCurCpu != idGipMaster)
5444	{
5445	idNewGipMaster = idCurCpu;
5446	break;
5447	}
5448	}
5449
5450	Log(("supdrvGipMpEvent: Gip master %#lx -> %#lx\n", (long)idGipMaster, (long)idNewGipMaster));
5451	ASMAtomicCmpXchgSize(&pDevExt->idGipMaster, idNewGipMaster, idGipMaster, fIgnored);
5452	NOREF(fIgnored);
5453	}
5454	}
5455	}
5456
5457
5458	/**
5459	* Callback used by supdrvDetermineAsyncTSC to read the TSC on a CPU.
5460	*
5461	* @param idCpu Ignored.
5462	* @param pvUser1 Where to put the TSC.
5463	* @param pvUser2 Ignored.
5464	*/
5465	static DECLCALLBACK(void) supdrvDetermineAsyncTscWorker(RTCPUID idCpu, void pvUser1, void pvUser2)
5466	{
5467	#if 1
5468	ASMAtomicWriteU64((uint64_t volatile *)pvUser1, ASMReadTSC());
5469	#else
5470	(uint64_t )pvUser1 = ASMReadTSC();
5471	#endif
5472	}
5473
5474
5475	/**
5476	* Determine if Async GIP mode is required because of TSC drift.
5477	*
5478	* When using the default/normal timer code it is essential that the time stamp counter
5479	* (TSC) runs never backwards, that is, a read operation to the counter should return
5480	* a bigger value than any previous read operation. This is guaranteed by the latest
5481	* AMD CPUs and by newer Intel CPUs which never enter the C2 state (P4). In any other
5482	* case we have to choose the asynchronous timer mode.
5483	*
5484	* @param poffMin Pointer to the determined difference between different cores.
5485	* @return false if the time stamp counters appear to be synchronized, true otherwise.
5486	*/
5487	static bool supdrvDetermineAsyncTsc(uint64_t *poffMin)
5488	{
5489	/*
5490	* Just iterate all the cpus 8 times and make sure that the TSC is
5491	* ever increasing. We don't bother taking TSC rollover into account.
5492	*/
5493	int iEndCpu = RTMpGetArraySize();
5494	int iCpu;
5495	int cLoops = 8;
5496	bool fAsync = false;
5497	int rc = VINF_SUCCESS;
5498	uint64_t offMax = 0;
5499	uint64_t offMin = ~(uint64_t)0;
5500	uint64_t PrevTsc = ASMReadTSC();
5501
5502	while (cLoops-- > 0)
5503	{
5504	for (iCpu = 0; iCpu < iEndCpu; iCpu++)
5505	{
5506	uint64_t CurTsc;
5507	rc = RTMpOnSpecific(RTMpCpuIdFromSetIndex(iCpu), supdrvDetermineAsyncTscWorker, &CurTsc, NULL);
5508	if (RT_SUCCESS(rc))
5509	{
5510	if (CurTsc <= PrevTsc)
5511	{
5512	fAsync = true;
5513	offMin = offMax = PrevTsc - CurTsc;
5514	Log(("supdrvDetermineAsyncTsc: iCpu=%d cLoops=%d CurTsc=%llx PrevTsc=%llx\n",
5515	iCpu, cLoops, CurTsc, PrevTsc));
5516	break;
5517	}
5518
5519	/* Gather statistics (except the first time). */
5520	if (iCpu != 0 \|\| cLoops != 7)
5521	{
5522	uint64_t off = CurTsc - PrevTsc;
5523	if (off < offMin)
5524	offMin = off;
5525	if (off > offMax)
5526	offMax = off;
5527	Log2(("%d/%d: off=%llx\n", cLoops, iCpu, off));
5528	}
5529
5530	/* Next */
5531	PrevTsc = CurTsc;
5532	}
5533	else if (rc == VERR_NOT_SUPPORTED)
5534	break;
5535	else
5536	AssertMsg(rc == VERR_CPU_NOT_FOUND \|\| rc == VERR_CPU_OFFLINE, ("%d\n", rc));
5537	}
5538
5539	/* broke out of the loop. */
5540	if (iCpu < iEndCpu)
5541	break;
5542	}
5543
5544	poffMin = offMin; / Almost RTMpOnSpecific profiling. */
5545	Log(("supdrvDetermineAsyncTsc: returns %d; iEndCpu=%d rc=%d offMin=%llx offMax=%llx\n",
5546	fAsync, iEndCpu, rc, offMin, offMax));
5547	#if !defined(RT_OS_SOLARIS) && !defined(RT_OS_OS2) && !defined(RT_OS_WINDOWS)
5548	OSDBGPRINT(("vboxdrv: fAsync=%d offMin=%#lx offMax=%#lx\n", fAsync, (long)offMin, (long)offMax));
5549	#endif
5550	return fAsync;
5551	}
5552
5553
5554	/**
5555	* Determine the GIP TSC mode.
5556	*
5557	* @returns The most suitable TSC mode.
5558	* @param pDevExt Pointer to the device instance data.
5559	*/
5560	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt)
5561	{
5562	/*
5563	* On SMP we're faced with two problems:
5564	* (1) There might be a skew between the CPU, so that cpu0
5565	* returns a TSC that is slightly different from cpu1.
5566	* (2) Power management (and other things) may cause the TSC
5567	* to run at a non-constant speed, and cause the speed
5568	* to be different on the cpus. This will result in (1).
5569	*
5570	* So, on SMP systems we'll have to select the ASYNC update method
5571	* if there are symptoms of these problems.
5572	*/
5573	if (RTMpGetCount() > 1)
5574	{
5575	uint32_t uEAX, uEBX, uECX, uEDX;
5576	uint64_t u64DiffCoresIgnored;
5577
5578	/* Permit the user and/or the OS specific bits to force async mode. */
5579	if (supdrvOSGetForcedAsyncTscMode(pDevExt))
5580	return SUPGIPMODE_ASYNC_TSC;
5581
5582	/* Try check for current differences between the cpus. */
5583	if (supdrvDetermineAsyncTsc(&u64DiffCoresIgnored))
5584	return SUPGIPMODE_ASYNC_TSC;
5585
5586	/*
5587	* If the CPU supports power management and is an AMD one we
5588	* won't trust it unless it has the TscInvariant bit is set.
5589	*/
5590	/* Check for "AuthenticAMD" */
5591	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
5592	if ( uEAX >= 1
5593	&& ASMIsAmdCpuEx(uEBX, uECX, uEDX))
5594	{
5595	/* Check for APM support and that TscInvariant is cleared. */
5596	ASMCpuId(0x80000000, &uEAX, &uEBX, &uECX, &uEDX);
5597	if (uEAX >= 0x80000007)
5598	{
5599	ASMCpuId(0x80000007, &uEAX, &uEBX, &uECX, &uEDX);
5600	if ( !(uEDX & RT_BIT(8))/* TscInvariant */
5601	&& (uEDX & 0x3e)) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
5602	return SUPGIPMODE_ASYNC_TSC;
5603	}
5604	}
5605	}
5606	return SUPGIPMODE_SYNC_TSC;
5607	}
5608
5609
5610	/**
5611	* Initializes per-CPU GIP information.
5612	*
5613	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5614	* @param pCpu Pointer to which GIP CPU to initalize.
5615	* @param u64NanoTS The current nanosecond timestamp.
5616	*/
5617	static void supdrvGipInitCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pCpu, uint64_t u64NanoTS)
5618	{
5619	pCpu->u32TransactionId = 2;
5620	pCpu->u64NanoTS = u64NanoTS;
5621	pCpu->u64TSC = ASMReadTSC();
5622
5623	ASMAtomicWriteSize(&pCpu->enmState, SUPGIPCPUSTATE_INVALID);
5624	ASMAtomicWriteSize(&pCpu->idCpu, NIL_RTCPUID);
5625	ASMAtomicWriteS16(&pCpu->iCpuSet, -1);
5626	ASMAtomicWriteU16(&pCpu->idApic, UINT16_MAX);
5627
5628	/*
5629	* We don't know the following values until we've executed updates.
5630	* So, we'll just pretend it's a 4 GHz CPU and adjust the history it on
5631	* the 2nd timer callout.
5632	*/
5633	pCpu->u64CpuHz = _4G + 1; /* tstGIP-2 depends on this. */
5634	pCpu->u32UpdateIntervalTSC
5635	= pCpu->au32TSCHistory[0]
5636	= pCpu->au32TSCHistory[1]
5637	= pCpu->au32TSCHistory[2]
5638	= pCpu->au32TSCHistory[3]
5639	= pCpu->au32TSCHistory[4]
5640	= pCpu->au32TSCHistory[5]
5641	= pCpu->au32TSCHistory[6]
5642	= pCpu->au32TSCHistory[7]
5643	= (uint32_t)(_4G / pGip->u32UpdateHz);
5644	}
5645
5646
5647	/**
5648	* Initializes the GIP data.
5649	*
5650	* @param pDevExt Pointer to the device instance data.
5651	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5652	* @param HCPhys The physical address of the GIP.
5653	* @param u64NanoTS The current nanosecond timestamp.
5654	* @param uUpdateHz The update frequency.
5655	* @param cCpus The CPU count.
5656	*/
5657	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
5658	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned cCpus)
5659	{
5660	size_t const cbGip = RT_ALIGN_Z(RT_OFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), PAGE_SIZE);
5661	unsigned i;
5662	#ifdef DEBUG_DARWIN_GIP
5663	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5664	#else
5665	LogFlow(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5666	#endif
5667
5668	/*
5669	* Initialize the structure.
5670	*/
5671	memset(pGip, 0, cbGip);
5672	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
5673	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
5674	pGip->u32Mode = supdrvGipDeterminTscMode(pDevExt);
5675	pGip->cCpus = (uint16_t)cCpus;
5676	pGip->cPages = (uint16_t)(cbGip / PAGE_SIZE);
5677	pGip->u32UpdateHz = uUpdateHz;
5678	pGip->u32UpdateIntervalNS = 1000000000 / uUpdateHz;
5679	pGip->u64NanoTSLastUpdateHz = u64NanoTS;
5680	RTCpuSetEmpty(&pGip->OnlineCpuSet);
5681	RTCpuSetEmpty(&pGip->PresentCpuSet);
5682	RTMpGetSet(&pGip->PossibleCpuSet);
5683	pGip->cOnlineCpus = RTMpGetOnlineCount();
5684	pGip->cPresentCpus = RTMpGetPresentCount();
5685	pGip->cPossibleCpus = RTMpGetCount();
5686	pGip->idCpuMax = RTMpGetMaxCpuId();
5687	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromApicId); i++)
5688	pGip->aiCpuFromApicId[i] = 0;
5689	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx); i++)
5690	pGip->aiCpuFromCpuSetIdx[i] = UINT16_MAX;
5691
5692	for (i = 0; i < cCpus; i++)
5693	supdrvGipInitCpu(pGip, &pGip->aCPUs[i], u64NanoTS);
5694
5695	/*
5696	* Link it to the device extension.
5697	*/
5698	pDevExt->pGip = pGip;
5699	pDevExt->HCPhysGip = HCPhys;
5700	pDevExt->cGipUsers = 0;
5701	}
5702
5703
5704	/**
5705	* On CPU initialization callback for RTMpOnAll.
5706	*
5707	* @param idCpu The CPU ID.
5708	* @param pvUser1 The device extension.
5709	* @param pvUser2 The GIP.
5710	*/
5711	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2)
5712	{
5713	/* This is good enough, even though it will update some of the globals a
5714	bit to much. */
5715	supdrvGipMpEventOnline((PSUPDRVDEVEXT)pvUser1, idCpu);
5716	}
5717
5718
5719	/**
5720	* Invalidates the GIP data upon termination.
5721	*
5722	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5723	*/
5724	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
5725	{
5726	unsigned i;
5727	pGip->u32Magic = 0;
5728	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
5729	{
5730	pGip->aCPUs[i].u64NanoTS = 0;
5731	pGip->aCPUs[i].u64TSC = 0;
5732	pGip->aCPUs[i].iTSCHistoryHead = 0;
5733	}
5734	}
5735
5736
5737	/**
5738	* Worker routine for supdrvGipUpdate and supdrvGipUpdatePerCpu that
5739	* updates all the per cpu data except the transaction id.
5740	*
5741	* @param pDevExt The device extension.
5742	* @param pGipCpu Pointer to the per cpu data.
5743	* @param u64NanoTS The current time stamp.
5744	* @param u64TSC The current TSC.
5745	* @param iTick The current timer tick.
5746	*/
5747	static void supdrvGipDoUpdateCpu(PSUPDRVDEVEXT pDevExt, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick)
5748	{
5749	uint64_t u64TSCDelta;
5750	uint32_t u32UpdateIntervalTSC;
5751	uint32_t u32UpdateIntervalTSCSlack;
5752	unsigned iTSCHistoryHead;
5753	uint64_t u64CpuHz;
5754	uint32_t u32TransactionId;
5755
5756	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5757	AssertPtrReturnVoid(pGip);
5758
5759	/* Delta between this and the previous update. */
5760	ASMAtomicUoWriteU32(&pGipCpu->u32PrevUpdateIntervalNS, (uint32_t)(u64NanoTS - pGipCpu->u64NanoTS));
5761
5762	/*
5763	* Update the NanoTS.
5764	*/
5765	ASMAtomicWriteU64(&pGipCpu->u64NanoTS, u64NanoTS);
5766
5767	/*
5768	* Calc TSC delta.
5769	*/
5770	/** @todo validate the NanoTS delta, don't trust the OS to call us when it should... */
5771	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
5772	ASMAtomicWriteU64(&pGipCpu->u64TSC, u64TSC);
5773
5774	if (u64TSCDelta >> 32)
5775	{
5776	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
5777	pGipCpu->cErrors++;
5778	}
5779
5780	/*
5781	* On the 2nd and 3rd callout, reset the history with the current TSC
5782	* interval since the values entered by supdrvGipInit are totally off.
5783	* The interval on the 1st callout completely unreliable, the 2nd is a bit
5784	* better, while the 3rd should be most reliable.
5785	*/
5786	u32TransactionId = pGipCpu->u32TransactionId;
5787	if (RT_UNLIKELY( ( u32TransactionId == 5
5788	\|\| u32TransactionId == 7)
5789	&& ( iTick == 2
5790	\|\| iTick == 3) ))
5791	{
5792	unsigned i;
5793	for (i = 0; i < RT_ELEMENTS(pGipCpu->au32TSCHistory); i++)
5794	ASMAtomicUoWriteU32(&pGipCpu->au32TSCHistory[i], (uint32_t)u64TSCDelta);
5795	}
5796
5797	/*
5798	* TSC History.
5799	*/
5800	Assert(RT_ELEMENTS(pGipCpu->au32TSCHistory) == 8);
5801	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
5802	ASMAtomicWriteU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
5803	ASMAtomicWriteU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
5804
5805	/*
5806	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
5807	*/
5808	if (pGip->u32UpdateHz >= 1000)
5809	{
5810	uint32_t u32;
5811	u32 = pGipCpu->au32TSCHistory[0];
5812	u32 += pGipCpu->au32TSCHistory[1];
5813	u32 += pGipCpu->au32TSCHistory[2];
5814	u32 += pGipCpu->au32TSCHistory[3];
5815	u32 >>= 2;
5816	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
5817	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
5818	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
5819	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
5820	u32UpdateIntervalTSC >>= 2;
5821	u32UpdateIntervalTSC += u32;
5822	u32UpdateIntervalTSC >>= 1;
5823
5824	/* Value chosen for a 2GHz Athlon64 running linux 2.6.10/11, . */
5825	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
5826	}
5827	else if (pGip->u32UpdateHz >= 90)
5828	{
5829	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5830	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
5831	u32UpdateIntervalTSC >>= 1;
5832
5833	/* value chosen on a 2GHz thinkpad running windows */
5834	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
5835	}
5836	else
5837	{
5838	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5839
5840	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
5841	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
5842	}
5843	ASMAtomicWriteU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
5844
5845	/*
5846	* CpuHz.
5847	*/
5848	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, pGip->u32UpdateHz);
5849	ASMAtomicWriteU64(&pGipCpu->u64CpuHz, u64CpuHz);
5850	}
5851
5852
5853	/**
5854	* Updates the GIP.
5855	*
5856	* @param pDevExt The device extension.
5857	* @param u64NanoTS The current nanosecond timesamp.
5858	* @param u64TSC The current TSC timesamp.
5859	* @param idCpu The CPU ID.
5860	* @param iTick The current timer tick.
5861	*/
5862	static void supdrvGipUpdate(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick)
5863	{
5864	/*
5865	* Determine the relevant CPU data.
5866	*/
5867	PSUPGIPCPU pGipCpu;
5868	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5869	AssertPtrReturnVoid(pGip);
5870
5871	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
5872	pGipCpu = &pGip->aCPUs[0];
5873	else
5874	{
5875	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
5876	if (RT_UNLIKELY(iCpu >= pGip->cCpus))
5877	return;
5878	pGipCpu = &pGip->aCPUs[iCpu];
5879	if (RT_UNLIKELY(pGipCpu->idCpu != idCpu))
5880	return;
5881	}
5882
5883	/*
5884	* Start update transaction.
5885	*/
5886	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5887	{
5888	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
5889	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5890	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5891	pGipCpu->cErrors++;
5892	return;
5893	}
5894
5895	/*
5896	* Recalc the update frequency every 0x800th time.
5897	*/
5898	if (!(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
5899	{
5900	if (pGip->u64NanoTSLastUpdateHz)
5901	{
5902	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
5903	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
5904	uint32_t u32UpdateHz = (uint32_t)((UINT64_C(1000000000) * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
5905	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
5906	{
5907	ASMAtomicWriteU32(&pGip->u32UpdateHz, u32UpdateHz);
5908	ASMAtomicWriteU32(&pGip->u32UpdateIntervalNS, 1000000000 / u32UpdateHz);
5909	}
5910	#endif
5911	}
5912	ASMAtomicWriteU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS);
5913	}
5914
5915	/*
5916	* Update the data.
5917	*/
5918	supdrvGipDoUpdateCpu(pDevExt, pGipCpu, u64NanoTS, u64TSC, iTick);
5919
5920	/*
5921	* Complete transaction.
5922	*/
5923	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5924	}
5925
5926
5927	/**
5928	* Updates the per cpu GIP data for the calling cpu.
5929	*
5930	* @param pDevExt The device extension.
5931	* @param u64NanoTS The current nanosecond timesamp.
5932	* @param u64TSC The current TSC timesamp.
5933	* @param idCpu The CPU ID.
5934	* @param idApic The APIC id for the CPU index.
5935	* @param iTick The current timer tick.
5936	*/
5937	static void supdrvGipUpdatePerCpu(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC,
5938	RTCPUID idCpu, uint8_t idApic, uint64_t iTick)
5939	{
5940	uint32_t iCpu;
5941	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5942
5943	/*
5944	* Avoid a potential race when a CPU online notification doesn't fire on
5945	* the onlined CPU but the tick creeps in before the event notification is
5946	* run.
5947	*/
5948	if (RT_UNLIKELY(iTick == 1))
5949	{
5950	iCpu = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
5951	if (pGip->aCPUs[iCpu].enmState == SUPGIPCPUSTATE_OFFLINE)
5952	supdrvGipMpEventOnline(pDevExt, idCpu);
5953	}
5954
5955	iCpu = pGip->aiCpuFromApicId[idApic];
5956	if (RT_LIKELY(iCpu < pGip->cCpus))
5957	{
5958	PSUPGIPCPU pGipCpu = &pGip->aCPUs[iCpu];
5959	if (pGipCpu->idCpu == idCpu)
5960	{
5961	/*
5962	* Start update transaction.
5963	*/
5964	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5965	{
5966	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5967	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5968	pGipCpu->cErrors++;
5969	return;
5970	}
5971
5972	/*
5973	* Update the data.
5974	*/
5975	supdrvGipDoUpdateCpu(pDevExt, pGipCpu, u64NanoTS, u64TSC, iTick);
5976
5977	/*
5978	* Complete transaction.
5979	*/
5980	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5981	}
5982	}
5983	}
5984

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

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDrv.c@ 47199

Download in other formats: