SUPDrv.c@ 47537

Last change on this file since 47537 was 47537, checked in by vboxsync, 11 years ago
HostDrivers/Support: Introduce reference counting for the session object to keep it alive while there is someone accessing it, fixes R0 assertion on OS X when killing the VM process
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 224.2 KB

Line
1	/* $Id: SUPDrv.c 47537 2013-08-05 10:00:02Z vboxsync $ */
2	/** @file
3	* VBoxDrv - The VirtualBox Support Driver - Common code.
4	*/
5
6	/*
7	* Copyright (C) 2006-2013 Oracle Corporation
8	*
9	* This file is part of VirtualBox Open Source Edition (OSE), as
10	* available from http://www.virtualbox.org. This file is free software;
11	* you can redistribute it and/or modify it under the terms of the GNU
12	* General Public License (GPL) as published by the Free Software
13	* Foundation, in version 2 as it comes in the "COPYING" file of the
14	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16	*
17	* The contents of this file may alternatively be used under the terms
18	* of the Common Development and Distribution License Version 1.0
19	* (CDDL) only, as it comes in the "COPYING.CDDL" file of the
20	* VirtualBox OSE distribution, in which case the provisions of the
21	* CDDL are applicable instead of those of the GPL.
22	*
23	* You may elect to license modified versions of this file under the
24	* terms and conditions of either the GPL or the CDDL or both.
25	*/
26
27	/*******************************************************************************
28	* Header Files *
29	*******************************************************************************/
30	#define LOG_GROUP LOG_GROUP_SUP_DRV
31	#define SUPDRV_AGNOSTIC
32	#include "SUPDrvInternal.h"
33	#ifndef PAGE_SHIFT
34	# include <iprt/param.h>
35	#endif
36	#include <iprt/asm.h>
37	#include <iprt/asm-amd64-x86.h>
38	#include <iprt/asm-math.h>
39	#include <iprt/cpuset.h>
40	#include <iprt/handletable.h>
41	#include <iprt/mem.h>
42	#include <iprt/mp.h>
43	#include <iprt/power.h>
44	#include <iprt/process.h>
45	#include <iprt/semaphore.h>
46	#include <iprt/spinlock.h>
47	#include <iprt/thread.h>
48	#include <iprt/uuid.h>
49	#include <iprt/net.h>
50	#include <iprt/crc.h>
51	#include <iprt/string.h>
52	#include <iprt/timer.h>
53	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
54	# include <iprt/rand.h>
55	# include <iprt/path.h>
56	#endif
57	#include <iprt/x86.h>
58
59	#include <VBox/param.h>
60	#include <VBox/log.h>
61	#include <VBox/err.h>
62	#include <VBox/vmm/hm_svm.h>
63	#include <VBox/vmm/hm_vmx.h>
64
65	#if defined(RT_OS_SOLARIS) \|\| defined(RT_OS_DARWIN)
66	# include "dtrace/SUPDrv.h"
67	#else
68	# define VBOXDRV_SESSION_CREATE(pvSession, fUser) do { } while (0)
69	# define VBOXDRV_SESSION_CLOSE(pvSession) do { } while (0)
70	# define VBOXDRV_IOCTL_ENTRY(pvSession, uIOCtl, pvReqHdr) do { } while (0)
71	# define VBOXDRV_IOCTL_RETURN(pvSession, uIOCtl, pvReqHdr, rcRet, rcReq) do { } while (0)
72	#endif
73
74	/*
75	* Logging assignments:
76	* Log - useful stuff, like failures.
77	* LogFlow - program flow, except the really noisy bits.
78	* Log2 - Cleanup.
79	* Log3 - Loader flow noise.
80	* Log4 - Call VMMR0 flow noise.
81	* Log5 - Native yet-to-be-defined noise.
82	* Log6 - Native ioctl flow noise.
83	*
84	* Logging requires BUILD_TYPE=debug and possibly changes to the logger
85	* instantiation in log-vbox.c(pp).
86	*/
87
88
89	/*******************************************************************************
90	* Defined Constants And Macros *
91	*******************************************************************************/
92	/** The frequency by which we recalculate the u32UpdateHz and
93	* u32UpdateIntervalNS GIP members. The value must be a power of 2. */
94	#define GIP_UPDATEHZ_RECALC_FREQ 0x800
95
96	/** @def VBOX_SVN_REV
97	* The makefile should define this if it can. */
98	#ifndef VBOX_SVN_REV
99	# define VBOX_SVN_REV 0
100	#endif
101
102	#if 0 /* Don't start the GIP timers. Useful when debugging the IPRT timer code. */
103	# define DO_NOT_START_GIP
104	#endif
105
106
107	/*******************************************************************************
108	* Internal Functions *
109	*******************************************************************************/
110	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser);
111	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser);
112	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession);
113	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType);
114	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq);
115	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq);
116	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq);
117	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq);
118	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq);
119	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx);
120	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt);
121	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage);
122	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage);
123	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt);
124	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt);
125	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq);
126	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq);
127	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt);
128	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt);
129	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
130	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
131	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser);
132	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
133	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned cCpus);
134	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2);
135	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip);
136	static void supdrvGipUpdate(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick);
137	static void supdrvGipUpdatePerCpu(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC,
138	RTCPUID idCpu, uint8_t idApic, uint64_t iTick);
139	static void supdrvGipInitCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pCpu, uint64_t u64NanoTS);
140
141
142	/*******************************************************************************
143	* Global Variables *
144	*******************************************************************************/
145	DECLEXPORT(PSUPGLOBALINFOPAGE) g_pSUPGlobalInfoPage = NULL;
146
147	/**
148	* Array of the R0 SUP API.
149	*/
150	static SUPFUNC g_aFunctions[] =
151	{
152	/* SED: START */
153	/* name function */
154	/* Entries with absolute addresses determined at runtime, fixup
155	code makes ugly ASSUMPTIONS about the order here: */
156	{ "SUPR0AbsIs64bit", (void *)0 },
157	{ "SUPR0Abs64bitKernelCS", (void *)0 },
158	{ "SUPR0Abs64bitKernelSS", (void *)0 },
159	{ "SUPR0Abs64bitKernelDS", (void *)0 },
160	{ "SUPR0AbsKernelCS", (void *)0 },
161	{ "SUPR0AbsKernelSS", (void *)0 },
162	{ "SUPR0AbsKernelDS", (void *)0 },
163	{ "SUPR0AbsKernelES", (void *)0 },
164	{ "SUPR0AbsKernelFS", (void *)0 },
165	{ "SUPR0AbsKernelGS", (void *)0 },
166	/* Normal function pointers: */
167	{ "g_pSUPGlobalInfoPage", (void )&g_pSUPGlobalInfoPage }, / SED: DATA */
168	{ "SUPGetGIP", (void *)SUPGetGIP },
169	{ "SUPR0ComponentDeregisterFactory", (void *)SUPR0ComponentDeregisterFactory },
170	{ "SUPR0ComponentQueryFactory", (void *)SUPR0ComponentQueryFactory },
171	{ "SUPR0ComponentRegisterFactory", (void *)SUPR0ComponentRegisterFactory },
172	{ "SUPR0ContAlloc", (void *)SUPR0ContAlloc },
173	{ "SUPR0ContFree", (void *)SUPR0ContFree },
174	{ "SUPR0EnableVTx", (void *)SUPR0EnableVTx },
175	{ "SUPR0SuspendVTxOnCpu", (void *)SUPR0SuspendVTxOnCpu },
176	{ "SUPR0ResumeVTxOnCpu", (void *)SUPR0ResumeVTxOnCpu },
177	{ "SUPR0GetPagingMode", (void *)SUPR0GetPagingMode },
178	{ "SUPR0LockMem", (void *)SUPR0LockMem },
179	{ "SUPR0LowAlloc", (void *)SUPR0LowAlloc },
180	{ "SUPR0LowFree", (void *)SUPR0LowFree },
181	{ "SUPR0MemAlloc", (void *)SUPR0MemAlloc },
182	{ "SUPR0MemFree", (void *)SUPR0MemFree },
183	{ "SUPR0MemGetPhys", (void *)SUPR0MemGetPhys },
184	{ "SUPR0ObjAddRef", (void *)SUPR0ObjAddRef },
185	{ "SUPR0ObjAddRefEx", (void *)SUPR0ObjAddRefEx },
186	{ "SUPR0ObjRegister", (void *)SUPR0ObjRegister },
187	{ "SUPR0ObjRelease", (void *)SUPR0ObjRelease },
188	{ "SUPR0ObjVerifyAccess", (void *)SUPR0ObjVerifyAccess },
189	{ "SUPR0PageAllocEx", (void *)SUPR0PageAllocEx },
190	{ "SUPR0PageFree", (void *)SUPR0PageFree },
191	{ "SUPR0Printf", (void *)SUPR0Printf },
192	{ "SUPR0TracerDeregisterDrv", (void *)SUPR0TracerDeregisterDrv },
193	{ "SUPR0TracerDeregisterImpl", (void *)SUPR0TracerDeregisterImpl },
194	{ "SUPR0TracerFireProbe", (void *)SUPR0TracerFireProbe },
195	{ "SUPR0TracerRegisterDrv", (void *)SUPR0TracerRegisterDrv },
196	{ "SUPR0TracerRegisterImpl", (void *)SUPR0TracerRegisterImpl },
197	{ "SUPR0TracerRegisterModule", (void *)SUPR0TracerRegisterModule },
198	{ "SUPR0TracerUmodProbeFire", (void *)SUPR0TracerUmodProbeFire },
199	{ "SUPR0UnlockMem", (void *)SUPR0UnlockMem },
200	{ "SUPSemEventClose", (void *)SUPSemEventClose },
201	{ "SUPSemEventCreate", (void *)SUPSemEventCreate },
202	{ "SUPSemEventGetResolution", (void *)SUPSemEventGetResolution },
203	{ "SUPSemEventMultiClose", (void *)SUPSemEventMultiClose },
204	{ "SUPSemEventMultiCreate", (void *)SUPSemEventMultiCreate },
205	{ "SUPSemEventMultiGetResolution", (void *)SUPSemEventMultiGetResolution },
206	{ "SUPSemEventMultiReset", (void *)SUPSemEventMultiReset },
207	{ "SUPSemEventMultiSignal", (void *)SUPSemEventMultiSignal },
208	{ "SUPSemEventMultiWait", (void *)SUPSemEventMultiWait },
209	{ "SUPSemEventMultiWaitNoResume", (void *)SUPSemEventMultiWaitNoResume },
210	{ "SUPSemEventMultiWaitNsAbsIntr", (void *)SUPSemEventMultiWaitNsAbsIntr },
211	{ "SUPSemEventMultiWaitNsRelIntr", (void *)SUPSemEventMultiWaitNsRelIntr },
212	{ "SUPSemEventSignal", (void *)SUPSemEventSignal },
213	{ "SUPSemEventWait", (void *)SUPSemEventWait },
214	{ "SUPSemEventWaitNoResume", (void *)SUPSemEventWaitNoResume },
215	{ "SUPSemEventWaitNsAbsIntr", (void *)SUPSemEventWaitNsAbsIntr },
216	{ "SUPSemEventWaitNsRelIntr", (void *)SUPSemEventWaitNsRelIntr },
217
218	{ "RTAssertAreQuiet", (void *)RTAssertAreQuiet },
219	{ "RTAssertMayPanic", (void *)RTAssertMayPanic },
220	{ "RTAssertMsg1", (void *)RTAssertMsg1 },
221	{ "RTAssertMsg2AddV", (void *)RTAssertMsg2AddV },
222	{ "RTAssertMsg2V", (void *)RTAssertMsg2V },
223	{ "RTAssertSetMayPanic", (void *)RTAssertSetMayPanic },
224	{ "RTAssertSetQuiet", (void *)RTAssertSetQuiet },
225	{ "RTCrc32", (void *)RTCrc32 },
226	{ "RTCrc32Finish", (void *)RTCrc32Finish },
227	{ "RTCrc32Process", (void *)RTCrc32Process },
228	{ "RTCrc32Start", (void *)RTCrc32Start },
229	{ "RTErrConvertFromErrno", (void *)RTErrConvertFromErrno },
230	{ "RTErrConvertToErrno", (void *)RTErrConvertToErrno },
231	{ "RTHandleTableAllocWithCtx", (void *)RTHandleTableAllocWithCtx },
232	{ "RTHandleTableCreate", (void *)RTHandleTableCreate },
233	{ "RTHandleTableCreateEx", (void *)RTHandleTableCreateEx },
234	{ "RTHandleTableDestroy", (void *)RTHandleTableDestroy },
235	{ "RTHandleTableFreeWithCtx", (void *)RTHandleTableFreeWithCtx },
236	{ "RTHandleTableLookupWithCtx", (void *)RTHandleTableLookupWithCtx },
237	{ "RTLogDefaultInstance", (void *)RTLogDefaultInstance },
238	{ "RTLogGetDefaultInstance", (void *)RTLogGetDefaultInstance },
239	{ "RTLogLoggerExV", (void *)RTLogLoggerExV },
240	{ "RTLogPrintfV", (void *)RTLogPrintfV },
241	{ "RTLogRelDefaultInstance", (void *)RTLogRelDefaultInstance },
242	{ "RTLogSetDefaultInstanceThread", (void *)RTLogSetDefaultInstanceThread },
243	{ "RTMemAllocExTag", (void *)RTMemAllocExTag },
244	{ "RTMemAllocTag", (void *)RTMemAllocTag },
245	{ "RTMemAllocVarTag", (void *)RTMemAllocVarTag },
246	{ "RTMemAllocZTag", (void *)RTMemAllocZTag },
247	{ "RTMemAllocZVarTag", (void *)RTMemAllocZVarTag },
248	{ "RTMemDupExTag", (void *)RTMemDupExTag },
249	{ "RTMemDupTag", (void *)RTMemDupTag },
250	{ "RTMemFree", (void *)RTMemFree },
251	{ "RTMemFreeEx", (void *)RTMemFreeEx },
252	{ "RTMemReallocTag", (void *)RTMemReallocTag },
253	{ "RTMpCpuId", (void *)RTMpCpuId },
254	{ "RTMpCpuIdFromSetIndex", (void *)RTMpCpuIdFromSetIndex },
255	{ "RTMpCpuIdToSetIndex", (void *)RTMpCpuIdToSetIndex },
256	{ "RTMpGetArraySize", (void *)RTMpGetArraySize },
257	{ "RTMpGetCount", (void *)RTMpGetCount },
258	{ "RTMpGetMaxCpuId", (void *)RTMpGetMaxCpuId },
259	{ "RTMpGetOnlineCount", (void *)RTMpGetOnlineCount },
260	{ "RTMpGetOnlineSet", (void *)RTMpGetOnlineSet },
261	{ "RTMpGetSet", (void *)RTMpGetSet },
262	{ "RTMpIsCpuOnline", (void *)RTMpIsCpuOnline },
263	{ "RTMpIsCpuPossible", (void *)RTMpIsCpuPossible },
264	{ "RTMpIsCpuWorkPending", (void *)RTMpIsCpuWorkPending },
265	{ "RTMpNotificationDeregister", (void *)RTMpNotificationDeregister },
266	{ "RTMpNotificationRegister", (void *)RTMpNotificationRegister },
267	{ "RTMpOnAll", (void *)RTMpOnAll },
268	{ "RTMpOnOthers", (void *)RTMpOnOthers },
269	{ "RTMpOnSpecific", (void *)RTMpOnSpecific },
270	{ "RTMpPokeCpu", (void *)RTMpPokeCpu },
271	{ "RTNetIPv4AddDataChecksum", (void *)RTNetIPv4AddDataChecksum },
272	{ "RTNetIPv4AddTCPChecksum", (void *)RTNetIPv4AddTCPChecksum },
273	{ "RTNetIPv4AddUDPChecksum", (void *)RTNetIPv4AddUDPChecksum },
274	{ "RTNetIPv4FinalizeChecksum", (void *)RTNetIPv4FinalizeChecksum },
275	{ "RTNetIPv4HdrChecksum", (void *)RTNetIPv4HdrChecksum },
276	{ "RTNetIPv4IsDHCPValid", (void *)RTNetIPv4IsDHCPValid },
277	{ "RTNetIPv4IsHdrValid", (void *)RTNetIPv4IsHdrValid },
278	{ "RTNetIPv4IsTCPSizeValid", (void *)RTNetIPv4IsTCPSizeValid },
279	{ "RTNetIPv4IsTCPValid", (void *)RTNetIPv4IsTCPValid },
280	{ "RTNetIPv4IsUDPSizeValid", (void *)RTNetIPv4IsUDPSizeValid },
281	{ "RTNetIPv4IsUDPValid", (void *)RTNetIPv4IsUDPValid },
282	{ "RTNetIPv4PseudoChecksum", (void *)RTNetIPv4PseudoChecksum },
283	{ "RTNetIPv4PseudoChecksumBits", (void *)RTNetIPv4PseudoChecksumBits },
284	{ "RTNetIPv4TCPChecksum", (void *)RTNetIPv4TCPChecksum },
285	{ "RTNetIPv4UDPChecksum", (void *)RTNetIPv4UDPChecksum },
286	{ "RTNetIPv6PseudoChecksum", (void *)RTNetIPv6PseudoChecksum },
287	{ "RTNetIPv6PseudoChecksumBits", (void *)RTNetIPv6PseudoChecksumBits },
288	{ "RTNetIPv6PseudoChecksumEx", (void *)RTNetIPv6PseudoChecksumEx },
289	{ "RTNetTCPChecksum", (void *)RTNetTCPChecksum },
290	{ "RTNetUDPChecksum", (void *)RTNetUDPChecksum },
291	{ "RTPowerNotificationDeregister", (void *)RTPowerNotificationDeregister },
292	{ "RTPowerNotificationRegister", (void *)RTPowerNotificationRegister },
293	{ "RTProcSelf", (void *)RTProcSelf },
294	{ "RTR0AssertPanicSystem", (void *)RTR0AssertPanicSystem },
295	{ "RTR0MemAreKrnlAndUsrDifferent", (void *)RTR0MemAreKrnlAndUsrDifferent },
296	{ "RTR0MemKernelIsValidAddr", (void *)RTR0MemKernelIsValidAddr },
297	{ "RTR0MemKernelCopyFrom", (void *)RTR0MemKernelCopyFrom },
298	{ "RTR0MemKernelCopyTo", (void *)RTR0MemKernelCopyTo },
299	{ "RTR0MemObjAddress", (void *)RTR0MemObjAddress },
300	{ "RTR0MemObjAddressR3", (void *)RTR0MemObjAddressR3 },
301	{ "RTR0MemObjAllocContTag", (void *)RTR0MemObjAllocContTag },
302	{ "RTR0MemObjAllocLowTag", (void *)RTR0MemObjAllocLowTag },
303	{ "RTR0MemObjAllocPageTag", (void *)RTR0MemObjAllocPageTag },
304	{ "RTR0MemObjAllocPhysExTag", (void *)RTR0MemObjAllocPhysExTag },
305	{ "RTR0MemObjAllocPhysNCTag", (void *)RTR0MemObjAllocPhysNCTag },
306	{ "RTR0MemObjAllocPhysTag", (void *)RTR0MemObjAllocPhysTag },
307	{ "RTR0MemObjEnterPhysTag", (void *)RTR0MemObjEnterPhysTag },
308	{ "RTR0MemObjFree", (void *)RTR0MemObjFree },
309	{ "RTR0MemObjGetPagePhysAddr", (void *)RTR0MemObjGetPagePhysAddr },
310	{ "RTR0MemObjIsMapping", (void *)RTR0MemObjIsMapping },
311	{ "RTR0MemObjLockUserTag", (void *)RTR0MemObjLockUserTag },
312	{ "RTR0MemObjMapKernelExTag", (void *)RTR0MemObjMapKernelExTag },
313	{ "RTR0MemObjMapKernelTag", (void *)RTR0MemObjMapKernelTag },
314	{ "RTR0MemObjMapUserTag", (void *)RTR0MemObjMapUserTag },
315	{ "RTR0MemObjProtect", (void *)RTR0MemObjProtect },
316	{ "RTR0MemObjSize", (void *)RTR0MemObjSize },
317	{ "RTR0MemUserCopyFrom", (void *)RTR0MemUserCopyFrom },
318	{ "RTR0MemUserCopyTo", (void *)RTR0MemUserCopyTo },
319	{ "RTR0MemUserIsValidAddr", (void *)RTR0MemUserIsValidAddr },
320	{ "RTR0ProcHandleSelf", (void *)RTR0ProcHandleSelf },
321	{ "RTSemEventCreate", (void *)RTSemEventCreate },
322	{ "RTSemEventDestroy", (void *)RTSemEventDestroy },
323	{ "RTSemEventGetResolution", (void *)RTSemEventGetResolution },
324	{ "RTSemEventMultiCreate", (void *)RTSemEventMultiCreate },
325	{ "RTSemEventMultiDestroy", (void *)RTSemEventMultiDestroy },
326	{ "RTSemEventMultiGetResolution", (void *)RTSemEventMultiGetResolution },
327	{ "RTSemEventMultiReset", (void *)RTSemEventMultiReset },
328	{ "RTSemEventMultiSignal", (void *)RTSemEventMultiSignal },
329	{ "RTSemEventMultiWait", (void *)RTSemEventMultiWait },
330	{ "RTSemEventMultiWaitEx", (void *)RTSemEventMultiWaitEx },
331	{ "RTSemEventMultiWaitExDebug", (void *)RTSemEventMultiWaitExDebug },
332	{ "RTSemEventMultiWaitNoResume", (void *)RTSemEventMultiWaitNoResume },
333	{ "RTSemEventSignal", (void *)RTSemEventSignal },
334	{ "RTSemEventWait", (void *)RTSemEventWait },
335	{ "RTSemEventWaitEx", (void *)RTSemEventWaitEx },
336	{ "RTSemEventWaitExDebug", (void *)RTSemEventWaitExDebug },
337	{ "RTSemEventWaitNoResume", (void *)RTSemEventWaitNoResume },
338	{ "RTSemFastMutexCreate", (void *)RTSemFastMutexCreate },
339	{ "RTSemFastMutexDestroy", (void *)RTSemFastMutexDestroy },
340	{ "RTSemFastMutexRelease", (void *)RTSemFastMutexRelease },
341	{ "RTSemFastMutexRequest", (void *)RTSemFastMutexRequest },
342	{ "RTSemMutexCreate", (void *)RTSemMutexCreate },
343	{ "RTSemMutexDestroy", (void *)RTSemMutexDestroy },
344	{ "RTSemMutexRelease", (void *)RTSemMutexRelease },
345	{ "RTSemMutexRequest", (void *)RTSemMutexRequest },
346	{ "RTSemMutexRequestDebug", (void *)RTSemMutexRequestDebug },
347	{ "RTSemMutexRequestNoResume", (void *)RTSemMutexRequestNoResume },
348	{ "RTSemMutexRequestNoResumeDebug", (void *)RTSemMutexRequestNoResumeDebug },
349	{ "RTSpinlockAcquire", (void *)RTSpinlockAcquire },
350	{ "RTSpinlockCreate", (void *)RTSpinlockCreate },
351	{ "RTSpinlockDestroy", (void *)RTSpinlockDestroy },
352	{ "RTSpinlockRelease", (void *)RTSpinlockRelease },
353	{ "RTSpinlockReleaseNoInts", (void *)RTSpinlockReleaseNoInts },
354	{ "RTStrCopy", (void *)RTStrCopy },
355	{ "RTStrDupTag", (void *)RTStrDupTag },
356	{ "RTStrFormat", (void *)RTStrFormat },
357	{ "RTStrFormatNumber", (void *)RTStrFormatNumber },
358	{ "RTStrFormatTypeDeregister", (void *)RTStrFormatTypeDeregister },
359	{ "RTStrFormatTypeRegister", (void *)RTStrFormatTypeRegister },
360	{ "RTStrFormatTypeSetUser", (void *)RTStrFormatTypeSetUser },
361	{ "RTStrFormatV", (void *)RTStrFormatV },
362	{ "RTStrFree", (void *)RTStrFree },
363	{ "RTStrNCmp", (void *)RTStrNCmp },
364	{ "RTStrPrintf", (void *)RTStrPrintf },
365	{ "RTStrPrintfEx", (void *)RTStrPrintfEx },
366	{ "RTStrPrintfExV", (void *)RTStrPrintfExV },
367	{ "RTStrPrintfV", (void *)RTStrPrintfV },
368	{ "RTThreadCreate", (void *)RTThreadCreate },
369	{ "RTThreadCtxHooksCreate", (void *)RTThreadCtxHooksCreate },
370	{ "RTThreadCtxHooksDeregister", (void *)RTThreadCtxHooksDeregister },
371	{ "RTThreadCtxHooksRegister", (void *)RTThreadCtxHooksRegister },
372	{ "RTThreadCtxHooksRelease", (void *)RTThreadCtxHooksRelease },
373	{ "RTThreadCtxHooksRetain", (void *)RTThreadCtxHooksRetain },
374	{ "RTThreadGetName", (void *)RTThreadGetName },
375	{ "RTThreadGetNative", (void *)RTThreadGetNative },
376	{ "RTThreadGetType", (void *)RTThreadGetType },
377	{ "RTThreadIsInInterrupt", (void *)RTThreadIsInInterrupt },
378	{ "RTThreadNativeSelf", (void *)RTThreadNativeSelf },
379	{ "RTThreadPreemptDisable", (void *)RTThreadPreemptDisable },
380	{ "RTThreadPreemptIsEnabled", (void *)RTThreadPreemptIsEnabled },
381	{ "RTThreadPreemptIsPending", (void *)RTThreadPreemptIsPending },
382	{ "RTThreadPreemptIsPendingTrusty", (void *)RTThreadPreemptIsPendingTrusty },
383	{ "RTThreadPreemptIsPossible", (void *)RTThreadPreemptIsPossible },
384	{ "RTThreadPreemptRestore", (void *)RTThreadPreemptRestore },
385	{ "RTThreadSelf", (void *)RTThreadSelf },
386	{ "RTThreadSelfName", (void *)RTThreadSelfName },
387	{ "RTThreadSleep", (void *)RTThreadSleep },
388	{ "RTThreadUserReset", (void *)RTThreadUserReset },
389	{ "RTThreadUserSignal", (void *)RTThreadUserSignal },
390	{ "RTThreadUserWait", (void *)RTThreadUserWait },
391	{ "RTThreadUserWaitNoResume", (void *)RTThreadUserWaitNoResume },
392	{ "RTThreadWait", (void *)RTThreadWait },
393	{ "RTThreadWaitNoResume", (void *)RTThreadWaitNoResume },
394	{ "RTThreadYield", (void *)RTThreadYield },
395	{ "RTTimeMilliTS", (void *)RTTimeMilliTS },
396	{ "RTTimeNanoTS", (void *)RTTimeNanoTS },
397	{ "RTTimeNow", (void *)RTTimeNow },
398	{ "RTTimerCanDoHighResolution", (void *)RTTimerCanDoHighResolution },
399	{ "RTTimerChangeInterval", (void *)RTTimerChangeInterval },
400	{ "RTTimerCreate", (void *)RTTimerCreate },
401	{ "RTTimerCreateEx", (void *)RTTimerCreateEx },
402	{ "RTTimerDestroy", (void *)RTTimerDestroy },
403	{ "RTTimerGetSystemGranularity", (void *)RTTimerGetSystemGranularity },
404	{ "RTTimerReleaseSystemGranularity", (void *)RTTimerReleaseSystemGranularity },
405	{ "RTTimerRequestSystemGranularity", (void *)RTTimerRequestSystemGranularity },
406	{ "RTTimerStart", (void *)RTTimerStart },
407	{ "RTTimerStop", (void *)RTTimerStop },
408	{ "RTTimeSystemMilliTS", (void *)RTTimeSystemMilliTS },
409	{ "RTTimeSystemNanoTS", (void *)RTTimeSystemNanoTS },
410	{ "RTUuidCompare", (void *)RTUuidCompare },
411	{ "RTUuidCompareStr", (void *)RTUuidCompareStr },
412	{ "RTUuidFromStr", (void *)RTUuidFromStr },
413	/* SED: END */
414	};
415
416	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
417	/**
418	* Drag in the rest of IRPT since we share it with the
419	* rest of the kernel modules on darwin.
420	*/
421	PFNRT g_apfnVBoxDrvIPRTDeps[] =
422	{
423	/* VBoxNetAdp */
424	(PFNRT)RTRandBytes,
425	/* VBoxUSB */
426	(PFNRT)RTPathStripFilename,
427	NULL
428	};
429	#endif /* RT_OS_DARWIN \|\| RT_OS_SOLARIS \|\| RT_OS_SOLARIS */
430
431
432	/**
433	* Initializes the device extentsion structure.
434	*
435	* @returns IPRT status code.
436	* @param pDevExt The device extension to initialize.
437	* @param cbSession The size of the session structure. The size of
438	* SUPDRVSESSION may be smaller when SUPDRV_AGNOSTIC is
439	* defined because we're skipping the OS specific members
440	* then.
441	*/
442	int VBOXCALL supdrvInitDevExt(PSUPDRVDEVEXT pDevExt, size_t cbSession)
443	{
444	int rc;
445
446	#ifdef SUPDRV_WITH_RELEASE_LOGGER
447	/*
448	* Create the release log.
449	*/
450	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
451	PRTLOGGER pRelLogger;
452	rc = RTLogCreate(&pRelLogger, 0 /* fFlags */, "all",
453	"VBOX_RELEASE_LOG", RT_ELEMENTS(s_apszGroups), s_apszGroups, RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER, NULL);
454	if (RT_SUCCESS(rc))
455	RTLogRelSetDefaultInstance(pRelLogger);
456	/** @todo Add native hook for getting logger config parameters and setting
457	* them. On linux we should use the module parameter stuff... */
458	#endif
459
460	/*
461	* Initialize it.
462	*/
463	memset(pDevExt, 0, sizeof(*pDevExt));
464	rc = RTSpinlockCreate(&pDevExt->Spinlock, RTSPINLOCK_FLAGS_INTERRUPT_SAFE, "SUPDrvDevExt");
465	if (RT_SUCCESS(rc))
466	{
467	rc = RTSpinlockCreate(&pDevExt->hGipSpinlock, RTSPINLOCK_FLAGS_INTERRUPT_SAFE, "SUPDrvGip");
468	if (RT_SUCCESS(rc))
469	{
470	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
471	rc = RTSemMutexCreate(&pDevExt->mtxLdr);
472	#else
473	rc = RTSemFastMutexCreate(&pDevExt->mtxLdr);
474	#endif
475	if (RT_SUCCESS(rc))
476	{
477	rc = RTSemFastMutexCreate(&pDevExt->mtxComponentFactory);
478	if (RT_SUCCESS(rc))
479	{
480	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
481	rc = RTSemMutexCreate(&pDevExt->mtxGip);
482	#else
483	rc = RTSemFastMutexCreate(&pDevExt->mtxGip);
484	#endif
485	if (RT_SUCCESS(rc))
486	{
487	rc = supdrvGipCreate(pDevExt);
488	if (RT_SUCCESS(rc))
489	{
490	rc = supdrvTracerInit(pDevExt);
491	if (RT_SUCCESS(rc))
492	{
493	pDevExt->pLdrInitImage = NULL;
494	pDevExt->hLdrInitThread = NIL_RTNATIVETHREAD;
495	pDevExt->u32Cookie = BIRD; /** @todo make this random? */
496	pDevExt->cbSession = (uint32_t)cbSession;
497
498	/*
499	* Fixup the absolute symbols.
500	*
501	* Because of the table indexing assumptions we'll have a little #ifdef orgy
502	* here rather than distributing this to OS specific files. At least for now.
503	*/
504	#ifdef RT_OS_DARWIN
505	# if ARCH_BITS == 32
506	if (SUPR0GetPagingMode() >= SUPPAGINGMODE_AMD64)
507	{
508	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
509	g_aFunctions[1].pfn = (void )0x80; / SUPR0Abs64bitKernelCS - KERNEL64_CS, seg.h */
510	g_aFunctions[2].pfn = (void )0x88; / SUPR0Abs64bitKernelSS - KERNEL64_SS, seg.h */
511	g_aFunctions[3].pfn = (void )0x88; / SUPR0Abs64bitKernelDS - KERNEL64_SS, seg.h */
512	}
513	else
514	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
515	g_aFunctions[4].pfn = (void )0x08; / SUPR0AbsKernelCS - KERNEL_CS, seg.h */
516	g_aFunctions[5].pfn = (void )0x10; / SUPR0AbsKernelSS - KERNEL_DS, seg.h */
517	g_aFunctions[6].pfn = (void )0x10; / SUPR0AbsKernelDS - KERNEL_DS, seg.h */
518	g_aFunctions[7].pfn = (void )0x10; / SUPR0AbsKernelES - KERNEL_DS, seg.h */
519	g_aFunctions[8].pfn = (void )0x10; / SUPR0AbsKernelFS - KERNEL_DS, seg.h */
520	g_aFunctions[9].pfn = (void )0x48; / SUPR0AbsKernelGS - CPU_DATA_GS, seg.h */
521	# else /* 64-bit darwin: */
522	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
523	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
524	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
525	g_aFunctions[3].pfn = (void )0; / SUPR0Abs64bitKernelDS */
526	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
527	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
528	g_aFunctions[6].pfn = (void )0; / SUPR0AbsKernelDS */
529	g_aFunctions[7].pfn = (void )0; / SUPR0AbsKernelES */
530	g_aFunctions[8].pfn = (void )0; / SUPR0AbsKernelFS */
531	g_aFunctions[9].pfn = (void )0; / SUPR0AbsKernelGS */
532
533	# endif
534	#else /* !RT_OS_DARWIN */
535	# if ARCH_BITS == 64
536	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
537	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
538	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
539	g_aFunctions[3].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0Abs64bitKernelDS */
540	# else
541	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
542	# endif
543	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
544	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
545	g_aFunctions[6].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0AbsKernelDS */
546	g_aFunctions[7].pfn = (void )(uintptr_t)ASMGetES(); / SUPR0AbsKernelES */
547	g_aFunctions[8].pfn = (void )(uintptr_t)ASMGetFS(); / SUPR0AbsKernelFS */
548	g_aFunctions[9].pfn = (void )(uintptr_t)ASMGetGS(); / SUPR0AbsKernelGS */
549	#endif /* !RT_OS_DARWIN */
550	return VINF_SUCCESS;
551	}
552
553	supdrvGipDestroy(pDevExt);
554	}
555
556	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
557	RTSemMutexDestroy(pDevExt->mtxGip);
558	pDevExt->mtxGip = NIL_RTSEMMUTEX;
559	#else
560	RTSemFastMutexDestroy(pDevExt->mtxGip);
561	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
562	#endif
563	}
564	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
565	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
566	}
567	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
568	RTSemMutexDestroy(pDevExt->mtxLdr);
569	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
570	#else
571	RTSemFastMutexDestroy(pDevExt->mtxLdr);
572	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
573	#endif
574	}
575	RTSpinlockDestroy(pDevExt->hGipSpinlock);
576	pDevExt->hGipSpinlock = NIL_RTSPINLOCK;
577	}
578	RTSpinlockDestroy(pDevExt->Spinlock);
579	pDevExt->Spinlock = NIL_RTSPINLOCK;
580	}
581	#ifdef SUPDRV_WITH_RELEASE_LOGGER
582	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
583	RTLogDestroy(RTLogSetDefaultInstance(NULL));
584	#endif
585
586	return rc;
587	}
588
589
590	/**
591	* Delete the device extension (e.g. cleanup members).
592	*
593	* @param pDevExt The device extension to delete.
594	*/
595	void VBOXCALL supdrvDeleteDevExt(PSUPDRVDEVEXT pDevExt)
596	{
597	PSUPDRVOBJ pObj;
598	PSUPDRVUSAGE pUsage;
599
600	/*
601	* Kill mutexes and spinlocks.
602	*/
603	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
604	RTSemMutexDestroy(pDevExt->mtxGip);
605	pDevExt->mtxGip = NIL_RTSEMMUTEX;
606	#else
607	RTSemFastMutexDestroy(pDevExt->mtxGip);
608	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
609	#endif
610	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
611	RTSemMutexDestroy(pDevExt->mtxLdr);
612	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
613	#else
614	RTSemFastMutexDestroy(pDevExt->mtxLdr);
615	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
616	#endif
617	RTSpinlockDestroy(pDevExt->Spinlock);
618	pDevExt->Spinlock = NIL_RTSPINLOCK;
619	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
620	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
621
622	/*
623	* Free lists.
624	*/
625	/* objects. */
626	pObj = pDevExt->pObjs;
627	Assert(!pObj); /* (can trigger on forced unloads) */
628	pDevExt->pObjs = NULL;
629	while (pObj)
630	{
631	void *pvFree = pObj;
632	pObj = pObj->pNext;
633	RTMemFree(pvFree);
634	}
635
636	/* usage records. */
637	pUsage = pDevExt->pUsageFree;
638	pDevExt->pUsageFree = NULL;
639	while (pUsage)
640	{
641	void *pvFree = pUsage;
642	pUsage = pUsage->pNext;
643	RTMemFree(pvFree);
644	}
645
646	/* kill the GIP. */
647	supdrvGipDestroy(pDevExt);
648	RTSpinlockDestroy(pDevExt->hGipSpinlock);
649	pDevExt->hGipSpinlock = NIL_RTSPINLOCK;
650
651	supdrvTracerTerm(pDevExt);
652
653	#ifdef SUPDRV_WITH_RELEASE_LOGGER
654	/* destroy the loggers. */
655	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
656	RTLogDestroy(RTLogSetDefaultInstance(NULL));
657	#endif
658	}
659
660
661	/**
662	* Create session.
663	*
664	* @returns IPRT status code.
665	* @param pDevExt Device extension.
666	* @param fUser Flag indicating whether this is a user or kernel
667	* session.
668	* @param fUnrestricted Unrestricted access (system) or restricted access
669	* (user)?
670	* @param ppSession Where to store the pointer to the session data.
671	*/
672	int VBOXCALL supdrvCreateSession(PSUPDRVDEVEXT pDevExt, bool fUser, bool fUnrestricted, PSUPDRVSESSION *ppSession)
673	{
674	int rc;
675	PSUPDRVSESSION pSession;
676
677	if (!SUP_IS_DEVEXT_VALID(pDevExt))
678	return VERR_INVALID_PARAMETER;
679
680	/*
681	* Allocate memory for the session data.
682	*/
683	pSession = *ppSession = (PSUPDRVSESSION)RTMemAllocZ(pDevExt->cbSession);
684	if (pSession)
685	{
686	/* Initialize session data. */
687	rc = RTSpinlockCreate(&pSession->Spinlock, RTSPINLOCK_FLAGS_INTERRUPT_UNSAFE, "SUPDrvSession");
688	if (!rc)
689	{
690	rc = RTHandleTableCreateEx(&pSession->hHandleTable,
691	RTHANDLETABLE_FLAGS_LOCKED_IRQ_SAFE \| RTHANDLETABLE_FLAGS_CONTEXT,
692	1 /uBase/, 32768 /cMax/, supdrvSessionObjHandleRetain, pSession);
693	if (RT_SUCCESS(rc))
694	{
695	Assert(pSession->Spinlock != NIL_RTSPINLOCK);
696	pSession->pDevExt = pDevExt;
697	pSession->u32Cookie = BIRD_INV;
698	pSession->fUnrestricted = fUnrestricted;
699	pSession->cRefs = 1;
700	/*pSession->pLdrUsage = NULL;
701	pSession->pVM = NULL;
702	pSession->pUsage = NULL;
703	pSession->pGip = NULL;
704	pSession->fGipReferenced = false;
705	pSession->Bundle.cUsed = 0; */
706	pSession->Uid = NIL_RTUID;
707	pSession->Gid = NIL_RTGID;
708	if (fUser)
709	{
710	pSession->Process = RTProcSelf();
711	pSession->R0Process = RTR0ProcHandleSelf();
712	}
713	else
714	{
715	pSession->Process = NIL_RTPROCESS;
716	pSession->R0Process = NIL_RTR0PROCESS;
717	}
718	/pSession->uTracerData = 0;/
719	pSession->hTracerCaller = NIL_RTNATIVETHREAD;
720	RTListInit(&pSession->TpProviders);
721	/pSession->cTpProviders = 0;/
722	/pSession->cTpProbesFiring = 0;/
723	RTListInit(&pSession->TpUmods);
724	/RT_ZERO(pSession->apTpLookupTable);/
725
726	VBOXDRV_SESSION_CREATE(pSession, fUser);
727	LogFlow(("Created session %p initial cookie=%#x\n", pSession, pSession->u32Cookie));
728	return VINF_SUCCESS;
729	}
730
731	RTSpinlockDestroy(pSession->Spinlock);
732	}
733	RTMemFree(pSession);
734	*ppSession = NULL;
735	Log(("Failed to create spinlock, rc=%d!\n", rc));
736	}
737	else
738	rc = VERR_NO_MEMORY;
739
740	return rc;
741	}
742
743
744	/**
745	* Shared code for cleaning up a session (but not quite freeing it).
746	*
747	* This is primarily intended for MAC OS X where we have to clean up the memory
748	* stuff before the file handle is closed.
749	*
750	* @param pDevExt Device extension.
751	* @param pSession Session data.
752	* This data will be freed by this routine.
753	*/
754	static void supdrvCleanupSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
755	{
756	int rc;
757	PSUPDRVBUNDLE pBundle;
758	LogFlow(("supdrvCleanupSession: pSession=%p\n", pSession));
759
760	/*
761	* Remove logger instances related to this session.
762	*/
763	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pSession);
764
765	/*
766	* Destroy the handle table.
767	*/
768	rc = RTHandleTableDestroy(pSession->hHandleTable, supdrvSessionObjHandleDelete, pSession);
769	AssertRC(rc);
770	pSession->hHandleTable = NIL_RTHANDLETABLE;
771
772	/*
773	* Release object references made in this session.
774	* In theory there should be noone racing us in this session.
775	*/
776	Log2(("release objects - start\n"));
777	if (pSession->pUsage)
778	{
779	PSUPDRVUSAGE pUsage;
780	RTSpinlockAcquire(pDevExt->Spinlock);
781
782	while ((pUsage = pSession->pUsage) != NULL)
783	{
784	PSUPDRVOBJ pObj = pUsage->pObj;
785	pSession->pUsage = pUsage->pNext;
786
787	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
788	if (pUsage->cUsage < pObj->cUsage)
789	{
790	pObj->cUsage -= pUsage->cUsage;
791	RTSpinlockRelease(pDevExt->Spinlock);
792	}
793	else
794	{
795	/* Destroy the object and free the record. */
796	if (pDevExt->pObjs == pObj)
797	pDevExt->pObjs = pObj->pNext;
798	else
799	{
800	PSUPDRVOBJ pObjPrev;
801	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
802	if (pObjPrev->pNext == pObj)
803	{
804	pObjPrev->pNext = pObj->pNext;
805	break;
806	}
807	Assert(pObjPrev);
808	}
809	RTSpinlockRelease(pDevExt->Spinlock);
810
811	Log(("supdrvCleanupSession: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
812	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
813	if (pObj->pfnDestructor)
814	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
815	RTMemFree(pObj);
816	}
817
818	/* free it and continue. */
819	RTMemFree(pUsage);
820
821	RTSpinlockAcquire(pDevExt->Spinlock);
822	}
823
824	RTSpinlockRelease(pDevExt->Spinlock);
825	AssertMsg(!pSession->pUsage, ("Some buster reregistered an object during desturction!\n"));
826	}
827	Log2(("release objects - done\n"));
828
829	/*
830	* Do tracer cleanups related to this session.
831	*/
832	Log2(("release tracer stuff - start\n"));
833	supdrvTracerCleanupSession(pDevExt, pSession);
834	Log2(("release tracer stuff - end\n"));
835
836	/*
837	* Release memory allocated in the session.
838	*
839	* We do not serialize this as we assume that the application will
840	* not allocated memory while closing the file handle object.
841	*/
842	Log2(("freeing memory:\n"));
843	pBundle = &pSession->Bundle;
844	while (pBundle)
845	{
846	PSUPDRVBUNDLE pToFree;
847	unsigned i;
848
849	/*
850	* Check and unlock all entries in the bundle.
851	*/
852	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
853	{
854	if (pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ)
855	{
856	Log2(("eType=%d pvR0=%p pvR3=%p cb=%ld\n", pBundle->aMem[i].eType, RTR0MemObjAddress(pBundle->aMem[i].MemObj),
857	(void *)RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3), (long)RTR0MemObjSize(pBundle->aMem[i].MemObj)));
858	if (pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ)
859	{
860	rc = RTR0MemObjFree(pBundle->aMem[i].MapObjR3, false);
861	AssertRC(rc); /** @todo figure out how to handle this. */
862	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
863	}
864	rc = RTR0MemObjFree(pBundle->aMem[i].MemObj, true /* fFreeMappings */);
865	AssertRC(rc); /** @todo figure out how to handle this. */
866	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
867	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
868	}
869	}
870
871	/*
872	* Advance and free previous bundle.
873	*/
874	pToFree = pBundle;
875	pBundle = pBundle->pNext;
876
877	pToFree->pNext = NULL;
878	pToFree->cUsed = 0;
879	if (pToFree != &pSession->Bundle)
880	RTMemFree(pToFree);
881	}
882	Log2(("freeing memory - done\n"));
883
884	/*
885	* Deregister component factories.
886	*/
887	RTSemFastMutexRequest(pDevExt->mtxComponentFactory);
888	Log2(("deregistering component factories:\n"));
889	if (pDevExt->pComponentFactoryHead)
890	{
891	PSUPDRVFACTORYREG pPrev = NULL;
892	PSUPDRVFACTORYREG pCur = pDevExt->pComponentFactoryHead;
893	while (pCur)
894	{
895	if (pCur->pSession == pSession)
896	{
897	/* unlink it */
898	PSUPDRVFACTORYREG pNext = pCur->pNext;
899	if (pPrev)
900	pPrev->pNext = pNext;
901	else
902	pDevExt->pComponentFactoryHead = pNext;
903
904	/* free it */
905	pCur->pNext = NULL;
906	pCur->pSession = NULL;
907	pCur->pFactory = NULL;
908	RTMemFree(pCur);
909
910	/* next */
911	pCur = pNext;
912	}
913	else
914	{
915	/* next */
916	pPrev = pCur;
917	pCur = pCur->pNext;
918	}
919	}
920	}
921	RTSemFastMutexRelease(pDevExt->mtxComponentFactory);
922	Log2(("deregistering component factories - done\n"));
923
924	/*
925	* Loaded images needs to be dereferenced and possibly freed up.
926	*/
927	supdrvLdrLock(pDevExt);
928	Log2(("freeing images:\n"));
929	if (pSession->pLdrUsage)
930	{
931	PSUPDRVLDRUSAGE pUsage = pSession->pLdrUsage;
932	pSession->pLdrUsage = NULL;
933	while (pUsage)
934	{
935	void *pvFree = pUsage;
936	PSUPDRVLDRIMAGE pImage = pUsage->pImage;
937	if (pImage->cUsage > pUsage->cUsage)
938	pImage->cUsage -= pUsage->cUsage;
939	else
940	supdrvLdrFree(pDevExt, pImage);
941	pUsage->pImage = NULL;
942	pUsage = pUsage->pNext;
943	RTMemFree(pvFree);
944	}
945	}
946	supdrvLdrUnlock(pDevExt);
947	Log2(("freeing images - done\n"));
948
949	/*
950	* Unmap the GIP.
951	*/
952	Log2(("umapping GIP:\n"));
953	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
954	{
955	SUPR0GipUnmap(pSession);
956	pSession->fGipReferenced = 0;
957	}
958	Log2(("umapping GIP - done\n"));
959	}
960
961
962	/**
963	* Shared code for cleaning up a session.
964	*
965	* @param pDevExt Device extension.
966	* @param pSession Session data.
967	* This data will be freed by this routine.
968	*/
969	static void supdrvCloseSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
970	{
971	VBOXDRV_SESSION_CLOSE(pSession);
972
973	/*
974	* Cleanup the session first.
975	*/
976	supdrvCleanupSession(pDevExt, pSession);
977
978	/*
979	* Free the rest of the session stuff.
980	*/
981	RTSpinlockDestroy(pSession->Spinlock);
982	pSession->Spinlock = NIL_RTSPINLOCK;
983	pSession->pDevExt = NULL;
984	RTMemFree(pSession);
985	LogFlow(("supdrvCloseSession: returns\n"));
986	}
987
988
989	/**
990	* Retain a session to make sure it doesn't go away while it is in use.
991	*
992	* @returns New reference count on success, UINT32_MAX on failure.
993	* @param pSession Session data.
994	*/
995	uint32_t VBOXCALL supdrvSessionRetain(PSUPDRVSESSION pSession)
996	{
997	AssertPtrReturn(pSession, UINT32_MAX);
998	AssertReturn(SUP_IS_SESSION_VALID(pSession), UINT32_MAX);
999
1000	uint32_t cRefs = ASMAtomicIncU32(&pSession->cRefs);
1001	AssertMsg(cRefs > 1 && cRefs < _1M, ("%#x %p\n", cRefs, pSession));
1002	return cRefs;
1003	}
1004
1005	/**
1006	* Releases a given session.
1007	*
1008	* @returns New reference count on success (0 if closed), UINT32_MAX on failure.
1009	* @param pSession Session data.
1010	*/
1011	uint32_t VBOXCALL supdrvSessionRelease(PSUPDRVSESSION pSession)
1012	{
1013	AssertPtrReturn(pSession, UINT32_MAX);
1014	AssertReturn(SUP_IS_SESSION_VALID(pSession), UINT32_MAX);
1015
1016	uint32_t cRefs = ASMAtomicDecU32(&pSession->cRefs);
1017	AssertMsg(cRefs < _1M, ("%#x %p\n", cRefs, pSession));
1018	if (cRefs == 0)
1019	supdrvCloseSession(pSession->pDevExt, pSession);
1020	return cRefs;
1021	}
1022
1023	/**
1024	* RTHandleTableDestroy callback used by supdrvCleanupSession.
1025	*
1026	* @returns IPRT status code, see SUPR0ObjAddRef.
1027	* @param hHandleTable The handle table handle. Ignored.
1028	* @param pvObj The object pointer.
1029	* @param pvCtx Context, the handle type. Ignored.
1030	* @param pvUser Session pointer.
1031	*/
1032	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser)
1033	{
1034	NOREF(pvCtx);
1035	NOREF(hHandleTable);
1036	return SUPR0ObjAddRefEx(pvObj, (PSUPDRVSESSION)pvUser, true /fNoBlocking/);
1037	}
1038
1039
1040	/**
1041	* RTHandleTableDestroy callback used by supdrvCleanupSession.
1042	*
1043	* @param hHandleTable The handle table handle. Ignored.
1044	* @param h The handle value. Ignored.
1045	* @param pvObj The object pointer.
1046	* @param pvCtx Context, the handle type. Ignored.
1047	* @param pvUser Session pointer.
1048	*/
1049	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser)
1050	{
1051	NOREF(pvCtx);
1052	NOREF(h);
1053	NOREF(hHandleTable);
1054	SUPR0ObjRelease(pvObj, (PSUPDRVSESSION)pvUser);
1055	}
1056
1057
1058	/**
1059	* Fast path I/O Control worker.
1060	*
1061	* @returns VBox status code that should be passed down to ring-3 unchanged.
1062	* @param uIOCtl Function number.
1063	* @param idCpu VMCPU id.
1064	* @param pDevExt Device extention.
1065	* @param pSession Session data.
1066	*/
1067	int VBOXCALL supdrvIOCtlFast(uintptr_t uIOCtl, VMCPUID idCpu, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
1068	{
1069	/*
1070	* We check the two prereqs after doing this only to allow the compiler to optimize things better.
1071	*/
1072	if (RT_LIKELY( RT_VALID_PTR(pSession)
1073	&& pSession->pVM
1074	&& pDevExt->pfnVMMR0EntryFast))
1075	{
1076	switch (uIOCtl)
1077	{
1078	case SUP_IOCTL_FAST_DO_RAW_RUN:
1079	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_RAW_RUN);
1080	break;
1081	case SUP_IOCTL_FAST_DO_HM_RUN:
1082	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_HM_RUN);
1083	break;
1084	case SUP_IOCTL_FAST_DO_NOP:
1085	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_NOP);
1086	break;
1087	default:
1088	return VERR_INTERNAL_ERROR;
1089	}
1090	return VINF_SUCCESS;
1091	}
1092	return VERR_INTERNAL_ERROR;
1093	}
1094
1095
1096	/**
1097	* Helper for supdrvIOCtl. Check if pszStr contains any character of pszChars.
1098	* We would use strpbrk here if this function would be contained in the RedHat kABI white
1099	* list, see http://www.kerneldrivers.org/RHEL5.
1100	*
1101	* @returns 1 if pszStr does contain any character of pszChars, 0 otherwise.
1102	* @param pszStr String to check
1103	* @param pszChars Character set
1104	*/
1105	static int supdrvCheckInvalidChar(const char pszStr, const char pszChars)
1106	{
1107	int chCur;
1108	while ((chCur = *pszStr++) != '\0')
1109	{
1110	int ch;
1111	const char *psz = pszChars;
1112	while ((ch = *psz++) != '\0')
1113	if (ch == chCur)
1114	return 1;
1115
1116	}
1117	return 0;
1118	}
1119
1120
1121
1122	/**
1123	* I/O Control inner worker (tracing reasons).
1124	*
1125	* @returns IPRT status code.
1126	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1127	*
1128	* @param uIOCtl Function number.
1129	* @param pDevExt Device extention.
1130	* @param pSession Session data.
1131	* @param pReqHdr The request header.
1132	*/
1133	static int supdrvIOCtlInnerUnrestricted(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1134	{
1135	/*
1136	* Validation macros
1137	*/
1138	#define REQ_CHECK_SIZES_EX(Name, cbInExpect, cbOutExpect) \
1139	do { \
1140	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect) \|\| pReqHdr->cbOut != (cbOutExpect))) \
1141	{ \
1142	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld. cbOut=%ld expected %ld.\n", \
1143	(long)pReqHdr->cbIn, (long)(cbInExpect), (long)pReqHdr->cbOut, (long)(cbOutExpect))); \
1144	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1145	} \
1146	} while (0)
1147
1148	#define REQ_CHECK_SIZES(Name) REQ_CHECK_SIZES_EX(Name, Name ## _SIZE_IN, Name ## _SIZE_OUT)
1149
1150	#define REQ_CHECK_SIZE_IN(Name, cbInExpect) \
1151	do { \
1152	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect))) \
1153	{ \
1154	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld.\n", \
1155	(long)pReqHdr->cbIn, (long)(cbInExpect))); \
1156	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1157	} \
1158	} while (0)
1159
1160	#define REQ_CHECK_SIZE_OUT(Name, cbOutExpect) \
1161	do { \
1162	if (RT_UNLIKELY(pReqHdr->cbOut != (cbOutExpect))) \
1163	{ \
1164	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbOut=%ld expected %ld.\n", \
1165	(long)pReqHdr->cbOut, (long)(cbOutExpect))); \
1166	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1167	} \
1168	} while (0)
1169
1170	#define REQ_CHECK_EXPR(Name, expr) \
1171	do { \
1172	if (RT_UNLIKELY(!(expr))) \
1173	{ \
1174	OSDBGPRINT(( #Name ": %s\n", #expr)); \
1175	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1176	} \
1177	} while (0)
1178
1179	#define REQ_CHECK_EXPR_FMT(expr, fmt) \
1180	do { \
1181	if (RT_UNLIKELY(!(expr))) \
1182	{ \
1183	OSDBGPRINT( fmt ); \
1184	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1185	} \
1186	} while (0)
1187
1188	/*
1189	* The switch.
1190	*/
1191	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
1192	{
1193	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
1194	{
1195	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
1196	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
1197	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
1198	{
1199	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
1200	pReq->Hdr.rc = VERR_INVALID_MAGIC;
1201	return 0;
1202	}
1203
1204	#if 0
1205	/*
1206	* Call out to the OS specific code and let it do permission checks on the
1207	* client process.
1208	*/
1209	if (!supdrvOSValidateClientProcess(pDevExt, pSession))
1210	{
1211	pReq->u.Out.u32Cookie = 0xffffffff;
1212	pReq->u.Out.u32SessionCookie = 0xffffffff;
1213	pReq->u.Out.u32SessionVersion = 0xffffffff;
1214	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1215	pReq->u.Out.pSession = NULL;
1216	pReq->u.Out.cFunctions = 0;
1217	pReq->Hdr.rc = VERR_PERMISSION_DENIED;
1218	return 0;
1219	}
1220	#endif
1221
1222	/*
1223	* Match the version.
1224	* The current logic is very simple, match the major interface version.
1225	*/
1226	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
1227	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
1228	{
1229	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1230	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
1231	pReq->u.Out.u32Cookie = 0xffffffff;
1232	pReq->u.Out.u32SessionCookie = 0xffffffff;
1233	pReq->u.Out.u32SessionVersion = 0xffffffff;
1234	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1235	pReq->u.Out.pSession = NULL;
1236	pReq->u.Out.cFunctions = 0;
1237	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1238	return 0;
1239	}
1240
1241	/*
1242	* Fill in return data and be gone.
1243	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
1244	* u32SessionVersion <= u32ReqVersion!
1245	*/
1246	/** @todo Somehow validate the client and negotiate a secure cookie... */
1247	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
1248	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
1249	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
1250	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1251	pReq->u.Out.pSession = pSession;
1252	pReq->u.Out.cFunctions = sizeof(g_aFunctions) / sizeof(g_aFunctions[0]);
1253	pReq->Hdr.rc = VINF_SUCCESS;
1254	return 0;
1255	}
1256
1257	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_QUERY_FUNCS(0)):
1258	{
1259	/* validate */
1260	PSUPQUERYFUNCS pReq = (PSUPQUERYFUNCS)pReqHdr;
1261	REQ_CHECK_SIZES_EX(SUP_IOCTL_QUERY_FUNCS, SUP_IOCTL_QUERY_FUNCS_SIZE_IN, SUP_IOCTL_QUERY_FUNCS_SIZE_OUT(RT_ELEMENTS(g_aFunctions)));
1262
1263	/* execute */
1264	pReq->u.Out.cFunctions = RT_ELEMENTS(g_aFunctions);
1265	memcpy(&pReq->u.Out.aFunctions[0], g_aFunctions, sizeof(g_aFunctions));
1266	pReq->Hdr.rc = VINF_SUCCESS;
1267	return 0;
1268	}
1269
1270	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_LOCK):
1271	{
1272	/* validate */
1273	PSUPPAGELOCK pReq = (PSUPPAGELOCK)pReqHdr;
1274	REQ_CHECK_SIZE_IN(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_IN);
1275	REQ_CHECK_SIZE_OUT(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_OUT(pReq->u.In.cPages));
1276	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.cPages > 0);
1277	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.pvR3 >= PAGE_SIZE);
1278
1279	/* execute */
1280	pReq->Hdr.rc = SUPR0LockMem(pSession, pReq->u.In.pvR3, pReq->u.In.cPages, &pReq->u.Out.aPages[0]);
1281	if (RT_FAILURE(pReq->Hdr.rc))
1282	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1283	return 0;
1284	}
1285
1286	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_UNLOCK):
1287	{
1288	/* validate */
1289	PSUPPAGEUNLOCK pReq = (PSUPPAGEUNLOCK)pReqHdr;
1290	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_UNLOCK);
1291
1292	/* execute */
1293	pReq->Hdr.rc = SUPR0UnlockMem(pSession, pReq->u.In.pvR3);
1294	return 0;
1295	}
1296
1297	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_ALLOC):
1298	{
1299	/* validate */
1300	PSUPCONTALLOC pReq = (PSUPCONTALLOC)pReqHdr;
1301	REQ_CHECK_SIZES(SUP_IOCTL_CONT_ALLOC);
1302
1303	/* execute */
1304	pReq->Hdr.rc = SUPR0ContAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.HCPhys);
1305	if (RT_FAILURE(pReq->Hdr.rc))
1306	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1307	return 0;
1308	}
1309
1310	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_FREE):
1311	{
1312	/* validate */
1313	PSUPCONTFREE pReq = (PSUPCONTFREE)pReqHdr;
1314	REQ_CHECK_SIZES(SUP_IOCTL_CONT_FREE);
1315
1316	/* execute */
1317	pReq->Hdr.rc = SUPR0ContFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1318	return 0;
1319	}
1320
1321	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_OPEN):
1322	{
1323	/* validate */
1324	PSUPLDROPEN pReq = (PSUPLDROPEN)pReqHdr;
1325	REQ_CHECK_SIZES(SUP_IOCTL_LDR_OPEN);
1326	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs > 0);
1327	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs < 16*_1M);
1328	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1329	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1330	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits < pReq->u.In.cbImageWithTabs);
1331	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.szName[0]);
1332	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1333	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, !supdrvCheckInvalidChar(pReq->u.In.szName, ";:()[]{}/\\\|&*%#@!~`\"'"));
1334	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szFilename, sizeof(pReq->u.In.szFilename)));
1335
1336	/* execute */
1337	pReq->Hdr.rc = supdrvIOCtl_LdrOpen(pDevExt, pSession, pReq);
1338	return 0;
1339	}
1340
1341	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOAD):
1342	{
1343	/* validate */
1344	PSUPLDRLOAD pReq = (PSUPLDRLOAD)pReqHdr;
1345	REQ_CHECK_EXPR(Name, pReq->Hdr.cbIn >= sizeof(*pReq));
1346	REQ_CHECK_SIZES_EX(SUP_IOCTL_LDR_LOAD, SUP_IOCTL_LDR_LOAD_SIZE_IN(pReq->u.In.cbImageWithTabs), SUP_IOCTL_LDR_LOAD_SIZE_OUT);
1347	REQ_CHECK_EXPR(SUP_IOCTL_LDR_LOAD, pReq->u.In.cSymbols <= 16384);
1348	REQ_CHECK_EXPR_FMT( !pReq->u.In.cSymbols
1349	\|\| ( pReq->u.In.offSymbols < pReq->u.In.cbImageWithTabs
1350	&& pReq->u.In.offSymbols + pReq->u.In.cSymbols * sizeof(SUPLDRSYM) <= pReq->u.In.cbImageWithTabs),
1351	("SUP_IOCTL_LDR_LOAD: offSymbols=%#lx cSymbols=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offSymbols,
1352	(long)pReq->u.In.cSymbols, (long)pReq->u.In.cbImageWithTabs));
1353	REQ_CHECK_EXPR_FMT( !pReq->u.In.cbStrTab
1354	\|\| ( pReq->u.In.offStrTab < pReq->u.In.cbImageWithTabs
1355	&& pReq->u.In.offStrTab + pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs
1356	&& pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs),
1357	("SUP_IOCTL_LDR_LOAD: offStrTab=%#lx cbStrTab=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offStrTab,
1358	(long)pReq->u.In.cbStrTab, (long)pReq->u.In.cbImageWithTabs));
1359
1360	if (pReq->u.In.cSymbols)
1361	{
1362	uint32_t i;
1363	PSUPLDRSYM paSyms = (PSUPLDRSYM)&pReq->u.In.abImage[pReq->u.In.offSymbols];
1364	for (i = 0; i < pReq->u.In.cSymbols; i++)
1365	{
1366	REQ_CHECK_EXPR_FMT(paSyms[i].offSymbol < pReq->u.In.cbImageWithTabs,
1367	("SUP_IOCTL_LDR_LOAD: sym #%ld: symb off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offSymbol, (long)pReq->u.In.cbImageWithTabs));
1368	REQ_CHECK_EXPR_FMT(paSyms[i].offName < pReq->u.In.cbStrTab,
1369	("SUP_IOCTL_LDR_LOAD: sym #%ld: name off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1370	REQ_CHECK_EXPR_FMT(RTStrEnd((char const *)&pReq->u.In.abImage[pReq->u.In.offStrTab + paSyms[i].offName],
1371	pReq->u.In.cbStrTab - paSyms[i].offName),
1372	("SUP_IOCTL_LDR_LOAD: sym #%ld: unterminated name! (%#lx / %#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1373	}
1374	}
1375
1376	/* execute */
1377	pReq->Hdr.rc = supdrvIOCtl_LdrLoad(pDevExt, pSession, pReq);
1378	return 0;
1379	}
1380
1381	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_FREE):
1382	{
1383	/* validate */
1384	PSUPLDRFREE pReq = (PSUPLDRFREE)pReqHdr;
1385	REQ_CHECK_SIZES(SUP_IOCTL_LDR_FREE);
1386
1387	/* execute */
1388	pReq->Hdr.rc = supdrvIOCtl_LdrFree(pDevExt, pSession, pReq);
1389	return 0;
1390	}
1391
1392	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_GET_SYMBOL):
1393	{
1394	/* validate */
1395	PSUPLDRGETSYMBOL pReq = (PSUPLDRGETSYMBOL)pReqHdr;
1396	REQ_CHECK_SIZES(SUP_IOCTL_LDR_GET_SYMBOL);
1397	REQ_CHECK_EXPR(SUP_IOCTL_LDR_GET_SYMBOL, RTStrEnd(pReq->u.In.szSymbol, sizeof(pReq->u.In.szSymbol)));
1398
1399	/* execute */
1400	pReq->Hdr.rc = supdrvIOCtl_LdrGetSymbol(pDevExt, pSession, pReq);
1401	return 0;
1402	}
1403
1404	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0(0)):
1405	{
1406	/* validate */
1407	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1408	Log4(("SUP_IOCTL_CALL_VMMR0: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1409	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1410
1411	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_VMMR0_SIZE(0))
1412	{
1413	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(0), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(0));
1414
1415	/* execute */
1416	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1417	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg, pSession);
1418	else
1419	pReq->Hdr.rc = VERR_WRONG_ORDER;
1420	}
1421	else
1422	{
1423	PSUPVMMR0REQHDR pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1424	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR)),
1425	("SUP_IOCTL_CALL_VMMR0: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR))));
1426	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1427	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(pVMMReq->cbReq));
1428
1429	/* execute */
1430	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1431	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1432	else
1433	pReq->Hdr.rc = VERR_WRONG_ORDER;
1434	}
1435
1436	if ( RT_FAILURE(pReq->Hdr.rc)
1437	&& pReq->Hdr.rc != VERR_INTERRUPTED
1438	&& pReq->Hdr.rc != VERR_TIMEOUT)
1439	Log(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1440	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1441	else
1442	Log4(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1443	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1444	return 0;
1445	}
1446
1447	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0_BIG):
1448	{
1449	/* validate */
1450	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1451	PSUPVMMR0REQHDR pVMMReq;
1452	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1453	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1454
1455	pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1456	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR)),
1457	("SUP_IOCTL_CALL_VMMR0_BIG: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR))));
1458	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0_BIG, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1459	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));
1460
1461	/* execute */
1462	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1463	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1464	else
1465	pReq->Hdr.rc = VERR_WRONG_ORDER;
1466
1467	if ( RT_FAILURE(pReq->Hdr.rc)
1468	&& pReq->Hdr.rc != VERR_INTERRUPTED
1469	&& pReq->Hdr.rc != VERR_TIMEOUT)
1470	Log(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1471	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1472	else
1473	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1474	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1475	return 0;
1476	}
1477
1478	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_PAGING_MODE):
1479	{
1480	/* validate */
1481	PSUPGETPAGINGMODE pReq = (PSUPGETPAGINGMODE)pReqHdr;
1482	REQ_CHECK_SIZES(SUP_IOCTL_GET_PAGING_MODE);
1483
1484	/* execute */
1485	pReq->Hdr.rc = VINF_SUCCESS;
1486	pReq->u.Out.enmMode = SUPR0GetPagingMode();
1487	return 0;
1488	}
1489
1490	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_ALLOC):
1491	{
1492	/* validate */
1493	PSUPLOWALLOC pReq = (PSUPLOWALLOC)pReqHdr;
1494	REQ_CHECK_EXPR(SUP_IOCTL_LOW_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_LOW_ALLOC_SIZE_IN);
1495	REQ_CHECK_SIZES_EX(SUP_IOCTL_LOW_ALLOC, SUP_IOCTL_LOW_ALLOC_SIZE_IN, SUP_IOCTL_LOW_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1496
1497	/* execute */
1498	pReq->Hdr.rc = SUPR0LowAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1499	if (RT_FAILURE(pReq->Hdr.rc))
1500	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1501	return 0;
1502	}
1503
1504	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_FREE):
1505	{
1506	/* validate */
1507	PSUPLOWFREE pReq = (PSUPLOWFREE)pReqHdr;
1508	REQ_CHECK_SIZES(SUP_IOCTL_LOW_FREE);
1509
1510	/* execute */
1511	pReq->Hdr.rc = SUPR0LowFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1512	return 0;
1513	}
1514
1515	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_MAP):
1516	{
1517	/* validate */
1518	PSUPGIPMAP pReq = (PSUPGIPMAP)pReqHdr;
1519	REQ_CHECK_SIZES(SUP_IOCTL_GIP_MAP);
1520
1521	/* execute */
1522	pReq->Hdr.rc = SUPR0GipMap(pSession, &pReq->u.Out.pGipR3, &pReq->u.Out.HCPhysGip);
1523	if (RT_SUCCESS(pReq->Hdr.rc))
1524	pReq->u.Out.pGipR0 = pDevExt->pGip;
1525	return 0;
1526	}
1527
1528	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_UNMAP):
1529	{
1530	/* validate */
1531	PSUPGIPUNMAP pReq = (PSUPGIPUNMAP)pReqHdr;
1532	REQ_CHECK_SIZES(SUP_IOCTL_GIP_UNMAP);
1533
1534	/* execute */
1535	pReq->Hdr.rc = SUPR0GipUnmap(pSession);
1536	return 0;
1537	}
1538
1539	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SET_VM_FOR_FAST):
1540	{
1541	/* validate */
1542	PSUPSETVMFORFAST pReq = (PSUPSETVMFORFAST)pReqHdr;
1543	REQ_CHECK_SIZES(SUP_IOCTL_SET_VM_FOR_FAST);
1544	REQ_CHECK_EXPR_FMT( !pReq->u.In.pVMR0
1545	\|\| ( VALID_PTR(pReq->u.In.pVMR0)
1546	&& !((uintptr_t)pReq->u.In.pVMR0 & (PAGE_SIZE - 1))),
1547	("SUP_IOCTL_SET_VM_FOR_FAST: pVMR0=%p!\n", pReq->u.In.pVMR0));
1548	/* execute */
1549	pSession->pVM = pReq->u.In.pVMR0;
1550	pReq->Hdr.rc = VINF_SUCCESS;
1551	return 0;
1552	}
1553
1554	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_ALLOC_EX):
1555	{
1556	/* validate */
1557	PSUPPAGEALLOCEX pReq = (PSUPPAGEALLOCEX)pReqHdr;
1558	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC_EX, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_EX_SIZE_IN);
1559	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));
1560	REQ_CHECK_EXPR_FMT(pReq->u.In.fKernelMapping \|\| pReq->u.In.fUserMapping,
1561	("SUP_IOCTL_PAGE_ALLOC_EX: No mapping requested!\n"));
1562	REQ_CHECK_EXPR_FMT(pReq->u.In.fUserMapping,
1563	("SUP_IOCTL_PAGE_ALLOC_EX: Must have user mapping!\n"));
1564	REQ_CHECK_EXPR_FMT(!pReq->u.In.fReserved0 && !pReq->u.In.fReserved1,
1565	("SUP_IOCTL_PAGE_ALLOC_EX: fReserved0=%d fReserved1=%d\n", pReq->u.In.fReserved0, pReq->u.In.fReserved1));
1566
1567	/* execute */
1568	pReq->Hdr.rc = SUPR0PageAllocEx(pSession, pReq->u.In.cPages, 0 /* fFlags */,
1569	pReq->u.In.fUserMapping ? &pReq->u.Out.pvR3 : NULL,
1570	pReq->u.In.fKernelMapping ? &pReq->u.Out.pvR0 : NULL,
1571	&pReq->u.Out.aPages[0]);
1572	if (RT_FAILURE(pReq->Hdr.rc))
1573	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1574	return 0;
1575	}
1576
1577	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_MAP_KERNEL):
1578	{
1579	/* validate */
1580	PSUPPAGEMAPKERNEL pReq = (PSUPPAGEMAPKERNEL)pReqHdr;
1581	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_MAP_KERNEL);
1582	REQ_CHECK_EXPR_FMT(!pReq->u.In.fFlags, ("SUP_IOCTL_PAGE_MAP_KERNEL: fFlags=%#x! MBZ\n", pReq->u.In.fFlags));
1583	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_MAP_KERNEL: offSub=%#x\n", pReq->u.In.offSub));
1584	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1585	("SUP_IOCTL_PAGE_MAP_KERNEL: cbSub=%#x\n", pReq->u.In.cbSub));
1586
1587	/* execute */
1588	pReq->Hdr.rc = SUPR0PageMapKernel(pSession, pReq->u.In.pvR3, pReq->u.In.offSub, pReq->u.In.cbSub,
1589	pReq->u.In.fFlags, &pReq->u.Out.pvR0);
1590	if (RT_FAILURE(pReq->Hdr.rc))
1591	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1592	return 0;
1593	}
1594
1595	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_PROTECT):
1596	{
1597	/* validate */
1598	PSUPPAGEPROTECT pReq = (PSUPPAGEPROTECT)pReqHdr;
1599	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_PROTECT);
1600	REQ_CHECK_EXPR_FMT(!(pReq->u.In.fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)),
1601	("SUP_IOCTL_PAGE_PROTECT: fProt=%#x!\n", pReq->u.In.fProt));
1602	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_PROTECT: offSub=%#x\n", pReq->u.In.offSub));
1603	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1604	("SUP_IOCTL_PAGE_PROTECT: cbSub=%#x\n", pReq->u.In.cbSub));
1605
1606	/* execute */
1607	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);
1608	return 0;
1609	}
1610
1611	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_FREE):
1612	{
1613	/* validate */
1614	PSUPPAGEFREE pReq = (PSUPPAGEFREE)pReqHdr;
1615	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_FREE);
1616
1617	/* execute */
1618	pReq->Hdr.rc = SUPR0PageFree(pSession, pReq->u.In.pvR3);
1619	return 0;
1620	}
1621
1622	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_SERVICE(0)):
1623	{
1624	/* validate */
1625	PSUPCALLSERVICE pReq = (PSUPCALLSERVICE)pReqHdr;
1626	Log4(("SUP_IOCTL_CALL_SERVICE: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1627	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1628
1629	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
1630	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(0), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(0));
1631	else
1632	{
1633	PSUPR0SERVICEREQHDR pSrvReq = (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0];
1634	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR)),
1635	("SUP_IOCTL_CALL_SERVICE: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR))));
1636	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, pSrvReq->u32Magic == SUPR0SERVICEREQHDR_MAGIC);
1637	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(pSrvReq->cbReq), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(pSrvReq->cbReq));
1638	}
1639	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1640
1641	/* execute */
1642	pReq->Hdr.rc = supdrvIOCtl_CallServiceModule(pDevExt, pSession, pReq);
1643	return 0;
1644	}
1645
1646	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOGGER_SETTINGS(0)):
1647	{
1648	/* validate */
1649	PSUPLOGGERSETTINGS pReq = (PSUPLOGGERSETTINGS)pReqHdr;
1650	size_t cbStrTab;
1651	REQ_CHECK_SIZE_OUT(SUP_IOCTL_LOGGER_SETTINGS, SUP_IOCTL_LOGGER_SETTINGS_SIZE_OUT);
1652	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->Hdr.cbIn >= SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(1));
1653	cbStrTab = pReq->Hdr.cbIn - SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(0);
1654	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offGroups < cbStrTab);
1655	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offFlags < cbStrTab);
1656	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offDestination < cbStrTab);
1657	REQ_CHECK_EXPR_FMT(pReq->u.In.szStrings[cbStrTab - 1] == '\0',
1658	("SUP_IOCTL_LOGGER_SETTINGS: cbIn=%#x cbStrTab=%#zx LastChar=%d\n",
1659	pReq->Hdr.cbIn, cbStrTab, pReq->u.In.szStrings[cbStrTab - 1]));
1660	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhich <= SUPLOGGERSETTINGS_WHICH_RELEASE);
1661	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhat <= SUPLOGGERSETTINGS_WHAT_DESTROY);
1662
1663	/* execute */
1664	pReq->Hdr.rc = supdrvIOCtl_LoggerSettings(pDevExt, pSession, pReq);
1665	return 0;
1666	}
1667
1668	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP2):
1669	{
1670	/* validate */
1671	PSUPSEMOP2 pReq = (PSUPSEMOP2)pReqHdr;
1672	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP2, SUP_IOCTL_SEM_OP2_SIZE_IN, SUP_IOCTL_SEM_OP2_SIZE_OUT);
1673	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP2, pReq->u.In.uReserved == 0);
1674
1675	/* execute */
1676	switch (pReq->u.In.uType)
1677	{
1678	case SUP_SEM_TYPE_EVENT:
1679	{
1680	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1681	switch (pReq->u.In.uOp)
1682	{
1683	case SUPSEMOP2_WAIT_MS_REL:
1684	pReq->Hdr.rc = SUPSemEventWaitNoResume(pSession, hEvent, pReq->u.In.uArg.cRelMsTimeout);
1685	break;
1686	case SUPSEMOP2_WAIT_NS_ABS:
1687	pReq->Hdr.rc = SUPSemEventWaitNsAbsIntr(pSession, hEvent, pReq->u.In.uArg.uAbsNsTimeout);
1688	break;
1689	case SUPSEMOP2_WAIT_NS_REL:
1690	pReq->Hdr.rc = SUPSemEventWaitNsRelIntr(pSession, hEvent, pReq->u.In.uArg.cRelNsTimeout);
1691	break;
1692	case SUPSEMOP2_SIGNAL:
1693	pReq->Hdr.rc = SUPSemEventSignal(pSession, hEvent);
1694	break;
1695	case SUPSEMOP2_CLOSE:
1696	pReq->Hdr.rc = SUPSemEventClose(pSession, hEvent);
1697	break;
1698	case SUPSEMOP2_RESET:
1699	default:
1700	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1701	break;
1702	}
1703	break;
1704	}
1705
1706	case SUP_SEM_TYPE_EVENT_MULTI:
1707	{
1708	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1709	switch (pReq->u.In.uOp)
1710	{
1711	case SUPSEMOP2_WAIT_MS_REL:
1712	pReq->Hdr.rc = SUPSemEventMultiWaitNoResume(pSession, hEventMulti, pReq->u.In.uArg.cRelMsTimeout);
1713	break;
1714	case SUPSEMOP2_WAIT_NS_ABS:
1715	pReq->Hdr.rc = SUPSemEventMultiWaitNsAbsIntr(pSession, hEventMulti, pReq->u.In.uArg.uAbsNsTimeout);
1716	break;
1717	case SUPSEMOP2_WAIT_NS_REL:
1718	pReq->Hdr.rc = SUPSemEventMultiWaitNsRelIntr(pSession, hEventMulti, pReq->u.In.uArg.cRelNsTimeout);
1719	break;
1720	case SUPSEMOP2_SIGNAL:
1721	pReq->Hdr.rc = SUPSemEventMultiSignal(pSession, hEventMulti);
1722	break;
1723	case SUPSEMOP2_CLOSE:
1724	pReq->Hdr.rc = SUPSemEventMultiClose(pSession, hEventMulti);
1725	break;
1726	case SUPSEMOP2_RESET:
1727	pReq->Hdr.rc = SUPSemEventMultiReset(pSession, hEventMulti);
1728	break;
1729	default:
1730	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1731	break;
1732	}
1733	break;
1734	}
1735
1736	default:
1737	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1738	break;
1739	}
1740	return 0;
1741	}
1742
1743	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP3):
1744	{
1745	/* validate */
1746	PSUPSEMOP3 pReq = (PSUPSEMOP3)pReqHdr;
1747	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP3, SUP_IOCTL_SEM_OP3_SIZE_IN, SUP_IOCTL_SEM_OP3_SIZE_OUT);
1748	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, pReq->u.In.u32Reserved == 0 && pReq->u.In.u64Reserved == 0);
1749
1750	/* execute */
1751	switch (pReq->u.In.uType)
1752	{
1753	case SUP_SEM_TYPE_EVENT:
1754	{
1755	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1756	switch (pReq->u.In.uOp)
1757	{
1758	case SUPSEMOP3_CREATE:
1759	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1760	pReq->Hdr.rc = SUPSemEventCreate(pSession, &hEvent);
1761	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEvent;
1762	break;
1763	case SUPSEMOP3_GET_RESOLUTION:
1764	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1765	pReq->Hdr.rc = VINF_SUCCESS;
1766	pReq->Hdr.cbOut = sizeof(*pReq);
1767	pReq->u.Out.cNsResolution = SUPSemEventGetResolution(pSession);
1768	break;
1769	default:
1770	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1771	break;
1772	}
1773	break;
1774	}
1775
1776	case SUP_SEM_TYPE_EVENT_MULTI:
1777	{
1778	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1779	switch (pReq->u.In.uOp)
1780	{
1781	case SUPSEMOP3_CREATE:
1782	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1783	pReq->Hdr.rc = SUPSemEventMultiCreate(pSession, &hEventMulti);
1784	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEventMulti;
1785	break;
1786	case SUPSEMOP3_GET_RESOLUTION:
1787	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1788	pReq->Hdr.rc = VINF_SUCCESS;
1789	pReq->u.Out.cNsResolution = SUPSemEventMultiGetResolution(pSession);
1790	break;
1791	default:
1792	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1793	break;
1794	}
1795	break;
1796	}
1797
1798	default:
1799	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1800	break;
1801	}
1802	return 0;
1803	}
1804
1805	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
1806	{
1807	/* validate */
1808	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
1809	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
1810
1811	/* execute */
1812	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.Caps);
1813	if (RT_FAILURE(pReq->Hdr.rc))
1814	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1815	return 0;
1816	}
1817
1818	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_OPEN):
1819	{
1820	/* validate */
1821	PSUPTRACEROPEN pReq = (PSUPTRACEROPEN)pReqHdr;
1822	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_OPEN);
1823
1824	/* execute */
1825	pReq->Hdr.rc = supdrvIOCtl_TracerOpen(pDevExt, pSession, pReq->u.In.uCookie, pReq->u.In.uArg);
1826	return 0;
1827	}
1828
1829	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_CLOSE):
1830	{
1831	/* validate */
1832	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_CLOSE);
1833
1834	/* execute */
1835	pReqHdr->rc = supdrvIOCtl_TracerClose(pDevExt, pSession);
1836	return 0;
1837	}
1838
1839	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_IOCTL):
1840	{
1841	/* validate */
1842	PSUPTRACERIOCTL pReq = (PSUPTRACERIOCTL)pReqHdr;
1843	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_IOCTL);
1844
1845	/* execute */
1846	pReqHdr->rc = supdrvIOCtl_TracerIOCtl(pDevExt, pSession, pReq->u.In.uCmd, pReq->u.In.uArg, &pReq->u.Out.iRetVal);
1847	return 0;
1848	}
1849
1850	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_REG):
1851	{
1852	/* validate */
1853	PSUPTRACERUMODREG pReq = (PSUPTRACERUMODREG)pReqHdr;
1854	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_REG);
1855	if (!RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)))
1856	return VERR_INVALID_PARAMETER;
1857
1858	/* execute */
1859	pReqHdr->rc = supdrvIOCtl_TracerUmodRegister(pDevExt, pSession,
1860	pReq->u.In.R3PtrVtgHdr, pReq->u.In.uVtgHdrAddr,
1861	pReq->u.In.R3PtrStrTab, pReq->u.In.cbStrTab,
1862	pReq->u.In.szName, pReq->u.In.fFlags);
1863	return 0;
1864	}
1865
1866	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_DEREG):
1867	{
1868	/* validate */
1869	PSUPTRACERUMODDEREG pReq = (PSUPTRACERUMODDEREG)pReqHdr;
1870	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_DEREG);
1871
1872	/* execute */
1873	pReqHdr->rc = supdrvIOCtl_TracerUmodDeregister(pDevExt, pSession, pReq->u.In.pVtgHdr);
1874	return 0;
1875	}
1876
1877	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_FIRE_PROBE):
1878	{
1879	/* validate */
1880	PSUPTRACERUMODFIREPROBE pReq = (PSUPTRACERUMODFIREPROBE)pReqHdr;
1881	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_FIRE_PROBE);
1882
1883	supdrvIOCtl_TracerUmodProbeFire(pDevExt, pSession, &pReq->u.In);
1884	pReqHdr->rc = VINF_SUCCESS;
1885	return 0;
1886	}
1887
1888	default:
1889	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1890	break;
1891	}
1892	return VERR_GENERAL_FAILURE;
1893	}
1894
1895
1896	/**
1897	* I/O Control inner worker for the restricted operations.
1898	*
1899	* @returns IPRT status code.
1900	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1901	*
1902	* @param uIOCtl Function number.
1903	* @param pDevExt Device extention.
1904	* @param pSession Session data.
1905	* @param pReqHdr The request header.
1906	*/
1907	static int supdrvIOCtlInnerRestricted(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1908	{
1909	/*
1910	* The switch.
1911	*/
1912	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
1913	{
1914	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
1915	{
1916	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
1917	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
1918	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
1919	{
1920	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
1921	pReq->Hdr.rc = VERR_INVALID_MAGIC;
1922	return 0;
1923	}
1924
1925	/*
1926	* Match the version.
1927	* The current logic is very simple, match the major interface version.
1928	*/
1929	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
1930	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
1931	{
1932	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1933	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
1934	pReq->u.Out.u32Cookie = 0xffffffff;
1935	pReq->u.Out.u32SessionCookie = 0xffffffff;
1936	pReq->u.Out.u32SessionVersion = 0xffffffff;
1937	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1938	pReq->u.Out.pSession = NULL;
1939	pReq->u.Out.cFunctions = 0;
1940	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1941	return 0;
1942	}
1943
1944	/*
1945	* Fill in return data and be gone.
1946	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
1947	* u32SessionVersion <= u32ReqVersion!
1948	*/
1949	/** @todo Somehow validate the client and negotiate a secure cookie... */
1950	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
1951	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
1952	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
1953	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1954	pReq->u.Out.pSession = pSession;
1955	pReq->u.Out.cFunctions = 0;
1956	pReq->Hdr.rc = VINF_SUCCESS;
1957	return 0;
1958	}
1959
1960	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
1961	{
1962	/* validate */
1963	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
1964	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
1965
1966	/* execute */
1967	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.Caps);
1968	if (RT_FAILURE(pReq->Hdr.rc))
1969	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1970	return 0;
1971	}
1972
1973	default:
1974	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1975	break;
1976	}
1977	return VERR_GENERAL_FAILURE;
1978	}
1979
1980
1981	/**
1982	* I/O Control worker.
1983	*
1984	* @returns IPRT status code.
1985	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1986	*
1987	* @param uIOCtl Function number.
1988	* @param pDevExt Device extention.
1989	* @param pSession Session data.
1990	* @param pReqHdr The request header.
1991	*/
1992	int VBOXCALL supdrvIOCtl(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1993	{
1994	int rc;
1995	VBOXDRV_IOCTL_ENTRY(pSession, uIOCtl, pReqHdr);
1996
1997	/*
1998	* Validate the request.
1999	*/
2000	/* this first check could probably be omitted as its also done by the OS specific code... */
2001	if (RT_UNLIKELY( (pReqHdr->fFlags & SUPREQHDR_FLAGS_MAGIC_MASK) != SUPREQHDR_FLAGS_MAGIC
2002	\|\| pReqHdr->cbIn < sizeof(*pReqHdr)
2003	\|\| pReqHdr->cbOut < sizeof(*pReqHdr)))
2004	{
2005	OSDBGPRINT(("vboxdrv: Bad ioctl request header; cbIn=%#lx cbOut=%#lx fFlags=%#lx\n",
2006	(long)pReqHdr->cbIn, (long)pReqHdr->cbOut, (long)pReqHdr->fFlags));
2007	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
2008	return VERR_INVALID_PARAMETER;
2009	}
2010	if (RT_UNLIKELY(!RT_VALID_PTR(pSession)))
2011	{
2012	OSDBGPRINT(("vboxdrv: Invalid pSession valud %p (ioctl=%p)\n", pSession, (void *)uIOCtl));
2013	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
2014	return VERR_INVALID_PARAMETER;
2015	}
2016	if (RT_UNLIKELY(uIOCtl == SUP_IOCTL_COOKIE))
2017	{
2018	if (pReqHdr->u32Cookie != SUPCOOKIE_INITIAL_COOKIE)
2019	{
2020	OSDBGPRINT(("SUP_IOCTL_COOKIE: bad cookie %#lx\n", (long)pReqHdr->u32Cookie));
2021	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
2022	return VERR_INVALID_PARAMETER;
2023	}
2024	}
2025	else if (RT_UNLIKELY( pReqHdr->u32Cookie != pDevExt->u32Cookie
2026	\|\| pReqHdr->u32SessionCookie != pSession->u32Cookie))
2027	{
2028	OSDBGPRINT(("vboxdrv: bad cookie %#lx / %#lx.\n", (long)pReqHdr->u32Cookie, (long)pReqHdr->u32SessionCookie));
2029	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
2030	return VERR_INVALID_PARAMETER;
2031	}
2032
2033	/*
2034	* Hand it to an inner function to avoid lots of unnecessary return tracepoints.
2035	*/
2036	if (pSession->fUnrestricted)
2037	rc = supdrvIOCtlInnerUnrestricted(uIOCtl, pDevExt, pSession, pReqHdr);
2038	else
2039	rc = supdrvIOCtlInnerRestricted(uIOCtl, pDevExt, pSession, pReqHdr);
2040
2041	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, pReqHdr->rc, rc);
2042	return rc;
2043	}
2044
2045
2046	/**
2047	* Inter-Driver Communication (IDC) worker.
2048	*
2049	* @returns VBox status code.
2050	* @retval VINF_SUCCESS on success.
2051	* @retval VERR_INVALID_PARAMETER if the request is invalid.
2052	* @retval VERR_NOT_SUPPORTED if the request isn't supported.
2053	*
2054	* @param uReq The request (function) code.
2055	* @param pDevExt Device extention.
2056	* @param pSession Session data.
2057	* @param pReqHdr The request header.
2058	*/
2059	int VBOXCALL supdrvIDC(uintptr_t uReq, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQHDR pReqHdr)
2060	{
2061	/*
2062	* The OS specific code has already validated the pSession
2063	* pointer, and the request size being greater or equal to
2064	* size of the header.
2065	*
2066	* So, just check that pSession is a kernel context session.
2067	*/
2068	if (RT_UNLIKELY( pSession
2069	&& pSession->R0Process != NIL_RTR0PROCESS))
2070	return VERR_INVALID_PARAMETER;
2071
2072	/*
2073	* Validation macro.
2074	*/
2075	#define REQ_CHECK_IDC_SIZE(Name, cbExpect) \
2076	do { \
2077	if (RT_UNLIKELY(pReqHdr->cb != (cbExpect))) \
2078	{ \
2079	OSDBGPRINT(( #Name ": Invalid input/output sizes. cb=%ld expected %ld.\n", \
2080	(long)pReqHdr->cb, (long)(cbExpect))); \
2081	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
2082	} \
2083	} while (0)
2084
2085	switch (uReq)
2086	{
2087	case SUPDRV_IDC_REQ_CONNECT:
2088	{
2089	PSUPDRVIDCREQCONNECT pReq = (PSUPDRVIDCREQCONNECT)pReqHdr;
2090	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_CONNECT, sizeof(*pReq));
2091
2092	/*
2093	* Validate the cookie and other input.
2094	*/
2095	if (pReq->Hdr.pSession != NULL)
2096	{
2097	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Hdr.pSession=%p expected NULL!\n", pReq->Hdr.pSession));
2098	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2099	}
2100	if (pReq->u.In.u32MagicCookie != SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE)
2101	{
2102	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: u32MagicCookie=%#x expected %#x!\n",
2103	(unsigned)pReq->u.In.u32MagicCookie, (unsigned)SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE));
2104	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2105	}
2106	if ( pReq->u.In.uMinVersion > pReq->u.In.uReqVersion
2107	\|\| (pReq->u.In.uMinVersion & UINT32_C(0xffff0000)) != (pReq->u.In.uReqVersion & UINT32_C(0xffff0000)))
2108	{
2109	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: uMinVersion=%#x uMaxVersion=%#x doesn't match!\n",
2110	pReq->u.In.uMinVersion, pReq->u.In.uReqVersion));
2111	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2112	}
2113	if (pSession != NULL)
2114	{
2115	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: pSession=%p expected NULL!\n", pSession));
2116	return pReqHdr->rc = VERR_INVALID_PARAMETER;
2117	}
2118
2119	/*
2120	* Match the version.
2121	* The current logic is very simple, match the major interface version.
2122	*/
2123	if ( pReq->u.In.uMinVersion > SUPDRV_IDC_VERSION
2124	\|\| (pReq->u.In.uMinVersion & 0xffff0000) != (SUPDRV_IDC_VERSION & 0xffff0000))
2125	{
2126	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
2127	pReq->u.In.uReqVersion, pReq->u.In.uMinVersion, (unsigned)SUPDRV_IDC_VERSION));
2128	pReq->u.Out.pSession = NULL;
2129	pReq->u.Out.uSessionVersion = 0xffffffff;
2130	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
2131	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
2132	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
2133	return VINF_SUCCESS;
2134	}
2135
2136	pReq->u.Out.pSession = NULL;
2137	pReq->u.Out.uSessionVersion = SUPDRV_IDC_VERSION;
2138	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
2139	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
2140
2141	pReq->Hdr.rc = supdrvCreateSession(pDevExt, false /* fUser /, true /fUnrestricted*/, &pSession);
2142	if (RT_FAILURE(pReq->Hdr.rc))
2143	{
2144	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: failed to create session, rc=%d\n", pReq->Hdr.rc));
2145	return VINF_SUCCESS;
2146	}
2147
2148	pReq->u.Out.pSession = pSession;
2149	pReq->Hdr.pSession = pSession;
2150
2151	return VINF_SUCCESS;
2152	}
2153
2154	case SUPDRV_IDC_REQ_DISCONNECT:
2155	{
2156	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_DISCONNECT, sizeof(*pReqHdr));
2157
2158	supdrvSessionRelease(pSession);
2159	return pReqHdr->rc = VINF_SUCCESS;
2160	}
2161
2162	case SUPDRV_IDC_REQ_GET_SYMBOL:
2163	{
2164	PSUPDRVIDCREQGETSYM pReq = (PSUPDRVIDCREQGETSYM)pReqHdr;
2165	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_GET_SYMBOL, sizeof(*pReq));
2166
2167	pReq->Hdr.rc = supdrvIDC_LdrGetSymbol(pDevExt, pSession, pReq);
2168	return VINF_SUCCESS;
2169	}
2170
2171	case SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY:
2172	{
2173	PSUPDRVIDCREQCOMPREGFACTORY pReq = (PSUPDRVIDCREQCOMPREGFACTORY)pReqHdr;
2174	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY, sizeof(*pReq));
2175
2176	pReq->Hdr.rc = SUPR0ComponentRegisterFactory(pSession, pReq->u.In.pFactory);
2177	return VINF_SUCCESS;
2178	}
2179
2180	case SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY:
2181	{
2182	PSUPDRVIDCREQCOMPDEREGFACTORY pReq = (PSUPDRVIDCREQCOMPDEREGFACTORY)pReqHdr;
2183	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY, sizeof(*pReq));
2184
2185	pReq->Hdr.rc = SUPR0ComponentDeregisterFactory(pSession, pReq->u.In.pFactory);
2186	return VINF_SUCCESS;
2187	}
2188
2189	default:
2190	Log(("Unknown IDC %#lx\n", (long)uReq));
2191	break;
2192	}
2193
2194	#undef REQ_CHECK_IDC_SIZE
2195	return VERR_NOT_SUPPORTED;
2196	}
2197
2198
2199	/**
2200	* Register a object for reference counting.
2201	* The object is registered with one reference in the specified session.
2202	*
2203	* @returns Unique identifier on success (pointer).
2204	* All future reference must use this identifier.
2205	* @returns NULL on failure.
2206	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
2207	* @param pvUser1 The first user argument.
2208	* @param pvUser2 The second user argument.
2209	*/
2210	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
2211	{
2212	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2213	PSUPDRVOBJ pObj;
2214	PSUPDRVUSAGE pUsage;
2215
2216	/*
2217	* Validate the input.
2218	*/
2219	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
2220	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
2221	AssertPtrReturn(pfnDestructor, NULL);
2222
2223	/*
2224	* Allocate and initialize the object.
2225	*/
2226	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
2227	if (!pObj)
2228	return NULL;
2229	pObj->u32Magic = SUPDRVOBJ_MAGIC;
2230	pObj->enmType = enmType;
2231	pObj->pNext = NULL;
2232	pObj->cUsage = 1;
2233	pObj->pfnDestructor = pfnDestructor;
2234	pObj->pvUser1 = pvUser1;
2235	pObj->pvUser2 = pvUser2;
2236	pObj->CreatorUid = pSession->Uid;
2237	pObj->CreatorGid = pSession->Gid;
2238	pObj->CreatorProcess= pSession->Process;
2239	supdrvOSObjInitCreator(pObj, pSession);
2240
2241	/*
2242	* Allocate the usage record.
2243	* (We keep freed usage records around to simplify SUPR0ObjAddRefEx().)
2244	*/
2245	RTSpinlockAcquire(pDevExt->Spinlock);
2246
2247	pUsage = pDevExt->pUsageFree;
2248	if (pUsage)
2249	pDevExt->pUsageFree = pUsage->pNext;
2250	else
2251	{
2252	RTSpinlockRelease(pDevExt->Spinlock);
2253	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
2254	if (!pUsage)
2255	{
2256	RTMemFree(pObj);
2257	return NULL;
2258	}
2259	RTSpinlockAcquire(pDevExt->Spinlock);
2260	}
2261
2262	/*
2263	* Insert the object and create the session usage record.
2264	*/
2265	/* The object. */
2266	pObj->pNext = pDevExt->pObjs;
2267	pDevExt->pObjs = pObj;
2268
2269	/* The session record. */
2270	pUsage->cUsage = 1;
2271	pUsage->pObj = pObj;
2272	pUsage->pNext = pSession->pUsage;
2273	/* Log2(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext)); */
2274	pSession->pUsage = pUsage;
2275
2276	RTSpinlockRelease(pDevExt->Spinlock);
2277
2278	Log(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
2279	return pObj;
2280	}
2281
2282
2283	/**
2284	* Increment the reference counter for the object associating the reference
2285	* with the specified session.
2286	*
2287	* @returns IPRT status code.
2288	* @param pvObj The identifier returned by SUPR0ObjRegister().
2289	* @param pSession The session which is referencing the object.
2290	*
2291	* @remarks The caller should not own any spinlocks and must carefully protect
2292	* itself against potential race with the destructor so freed memory
2293	* isn't accessed here.
2294	*/
2295	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
2296	{
2297	return SUPR0ObjAddRefEx(pvObj, pSession, false /* fNoBlocking */);
2298	}
2299
2300
2301	/**
2302	* Increment the reference counter for the object associating the reference
2303	* with the specified session.
2304	*
2305	* @returns IPRT status code.
2306	* @retval VERR_TRY_AGAIN if fNoBlocking was set and a new usage record
2307	* couldn't be allocated. (If you see this you're not doing the right
2308	* thing and it won't ever work reliably.)
2309	*
2310	* @param pvObj The identifier returned by SUPR0ObjRegister().
2311	* @param pSession The session which is referencing the object.
2312	* @param fNoBlocking Set if it's not OK to block. Never try to make the
2313	* first reference to an object in a session with this
2314	* argument set.
2315	*
2316	* @remarks The caller should not own any spinlocks and must carefully protect
2317	* itself against potential race with the destructor so freed memory
2318	* isn't accessed here.
2319	*/
2320	SUPR0DECL(int) SUPR0ObjAddRefEx(void *pvObj, PSUPDRVSESSION pSession, bool fNoBlocking)
2321	{
2322	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2323	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2324	int rc = VINF_SUCCESS;
2325	PSUPDRVUSAGE pUsagePre;
2326	PSUPDRVUSAGE pUsage;
2327
2328	/*
2329	* Validate the input.
2330	* Be ready for the destruction race (someone might be stuck in the
2331	* destructor waiting a lock we own).
2332	*/
2333	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2334	AssertPtrReturn(pObj, VERR_INVALID_POINTER);
2335	AssertMsgReturn(pObj->u32Magic == SUPDRVOBJ_MAGIC \|\| pObj->u32Magic == SUPDRVOBJ_MAGIC_DEAD,
2336	("Invalid pvObj=%p magic=%#x (expected %#x or %#x)\n", pvObj, pObj->u32Magic, SUPDRVOBJ_MAGIC, SUPDRVOBJ_MAGIC_DEAD),
2337	VERR_INVALID_PARAMETER);
2338
2339	RTSpinlockAcquire(pDevExt->Spinlock);
2340
2341	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2342	{
2343	RTSpinlockRelease(pDevExt->Spinlock);
2344
2345	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2346	return VERR_WRONG_ORDER;
2347	}
2348
2349	/*
2350	* Preallocate the usage record if we can.
2351	*/
2352	pUsagePre = pDevExt->pUsageFree;
2353	if (pUsagePre)
2354	pDevExt->pUsageFree = pUsagePre->pNext;
2355	else if (!fNoBlocking)
2356	{
2357	RTSpinlockRelease(pDevExt->Spinlock);
2358	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
2359	if (!pUsagePre)
2360	return VERR_NO_MEMORY;
2361
2362	RTSpinlockAcquire(pDevExt->Spinlock);
2363	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2364	{
2365	RTSpinlockRelease(pDevExt->Spinlock);
2366
2367	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2368	return VERR_WRONG_ORDER;
2369	}
2370	}
2371
2372	/*
2373	* Reference the object.
2374	*/
2375	pObj->cUsage++;
2376
2377	/*
2378	* Look for the session record.
2379	*/
2380	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
2381	{
2382	/Log(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2383	if (pUsage->pObj == pObj)
2384	break;
2385	}
2386	if (pUsage)
2387	pUsage->cUsage++;
2388	else if (pUsagePre)
2389	{
2390	/* create a new session record. */
2391	pUsagePre->cUsage = 1;
2392	pUsagePre->pObj = pObj;
2393	pUsagePre->pNext = pSession->pUsage;
2394	pSession->pUsage = pUsagePre;
2395	/Log(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));/
2396
2397	pUsagePre = NULL;
2398	}
2399	else
2400	{
2401	pObj->cUsage--;
2402	rc = VERR_TRY_AGAIN;
2403	}
2404
2405	/*
2406	* Put any unused usage record into the free list..
2407	*/
2408	if (pUsagePre)
2409	{
2410	pUsagePre->pNext = pDevExt->pUsageFree;
2411	pDevExt->pUsageFree = pUsagePre;
2412	}
2413
2414	RTSpinlockRelease(pDevExt->Spinlock);
2415
2416	return rc;
2417	}
2418
2419
2420	/**
2421	* Decrement / destroy a reference counter record for an object.
2422	*
2423	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
2424	*
2425	* @returns IPRT status code.
2426	* @retval VINF_SUCCESS if not destroyed.
2427	* @retval VINF_OBJECT_DESTROYED if it's destroyed by this release call.
2428	* @retval VERR_INVALID_PARAMETER if the object isn't valid. Will assert in
2429	* string builds.
2430	*
2431	* @param pvObj The identifier returned by SUPR0ObjRegister().
2432	* @param pSession The session which is referencing the object.
2433	*/
2434	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
2435	{
2436	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2437	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2438	int rc = VERR_INVALID_PARAMETER;
2439	PSUPDRVUSAGE pUsage;
2440	PSUPDRVUSAGE pUsagePrev;
2441
2442	/*
2443	* Validate the input.
2444	*/
2445	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2446	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2447	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2448	VERR_INVALID_PARAMETER);
2449
2450	/*
2451	* Acquire the spinlock and look for the usage record.
2452	*/
2453	RTSpinlockAcquire(pDevExt->Spinlock);
2454
2455	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
2456	pUsage;
2457	pUsagePrev = pUsage, pUsage = pUsage->pNext)
2458	{
2459	/Log2(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2460	if (pUsage->pObj == pObj)
2461	{
2462	rc = VINF_SUCCESS;
2463	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
2464	if (pUsage->cUsage > 1)
2465	{
2466	pObj->cUsage--;
2467	pUsage->cUsage--;
2468	}
2469	else
2470	{
2471	/*
2472	* Free the session record.
2473	*/
2474	if (pUsagePrev)
2475	pUsagePrev->pNext = pUsage->pNext;
2476	else
2477	pSession->pUsage = pUsage->pNext;
2478	pUsage->pNext = pDevExt->pUsageFree;
2479	pDevExt->pUsageFree = pUsage;
2480
2481	/* What about the object? */
2482	if (pObj->cUsage > 1)
2483	pObj->cUsage--;
2484	else
2485	{
2486	/*
2487	* Object is to be destroyed, unlink it.
2488	*/
2489	pObj->u32Magic = SUPDRVOBJ_MAGIC_DEAD;
2490	rc = VINF_OBJECT_DESTROYED;
2491	if (pDevExt->pObjs == pObj)
2492	pDevExt->pObjs = pObj->pNext;
2493	else
2494	{
2495	PSUPDRVOBJ pObjPrev;
2496	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
2497	if (pObjPrev->pNext == pObj)
2498	{
2499	pObjPrev->pNext = pObj->pNext;
2500	break;
2501	}
2502	Assert(pObjPrev);
2503	}
2504	}
2505	}
2506	break;
2507	}
2508	}
2509
2510	RTSpinlockRelease(pDevExt->Spinlock);
2511
2512	/*
2513	* Call the destructor and free the object if required.
2514	*/
2515	if (rc == VINF_OBJECT_DESTROYED)
2516	{
2517	Log(("SUPR0ObjRelease: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
2518	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
2519	if (pObj->pfnDestructor)
2520	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
2521	RTMemFree(pObj);
2522	}
2523
2524	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
2525	return rc;
2526	}
2527
2528
2529	/**
2530	* Verifies that the current process can access the specified object.
2531	*
2532	* @returns The following IPRT status code:
2533	* @retval VINF_SUCCESS if access was granted.
2534	* @retval VERR_PERMISSION_DENIED if denied access.
2535	* @retval VERR_INVALID_PARAMETER if invalid parameter.
2536	*
2537	* @param pvObj The identifier returned by SUPR0ObjRegister().
2538	* @param pSession The session which wishes to access the object.
2539	* @param pszObjName Object string name. This is optional and depends on the object type.
2540	*
2541	* @remark The caller is responsible for making sure the object isn't removed while
2542	* we're inside this function. If uncertain about this, just call AddRef before calling us.
2543	*/
2544	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
2545	{
2546	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2547	int rc;
2548
2549	/*
2550	* Validate the input.
2551	*/
2552	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2553	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2554	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2555	VERR_INVALID_PARAMETER);
2556
2557	/*
2558	* Check access. (returns true if a decision has been made.)
2559	*/
2560	rc = VERR_INTERNAL_ERROR;
2561	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
2562	return rc;
2563
2564	/*
2565	* Default policy is to allow the user to access his own
2566	* stuff but nothing else.
2567	*/
2568	if (pObj->CreatorUid == pSession->Uid)
2569	return VINF_SUCCESS;
2570	return VERR_PERMISSION_DENIED;
2571	}
2572
2573
2574	/**
2575	* Lock pages.
2576	*
2577	* @returns IPRT status code.
2578	* @param pSession Session to which the locked memory should be associated.
2579	* @param pvR3 Start of the memory range to lock.
2580	* This must be page aligned.
2581	* @param cPages Number of pages to lock.
2582	* @param paPages Where to put the physical addresses of locked memory.
2583	*/
2584	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
2585	{
2586	int rc;
2587	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2588	const size_t cb = (size_t)cPages << PAGE_SHIFT;
2589	LogFlow(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
2590
2591	/*
2592	* Verify input.
2593	*/
2594	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2595	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
2596	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
2597	\|\| !pvR3)
2598	{
2599	Log(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
2600	return VERR_INVALID_PARAMETER;
2601	}
2602
2603	/*
2604	* Let IPRT do the job.
2605	*/
2606	Mem.eType = MEMREF_TYPE_LOCKED;
2607	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTMEM_PROT_READ \| RTMEM_PROT_WRITE, RTR0ProcHandleSelf());
2608	if (RT_SUCCESS(rc))
2609	{
2610	uint32_t iPage = cPages;
2611	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
2612	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
2613
2614	while (iPage-- > 0)
2615	{
2616	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2617	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
2618	{
2619	AssertMsgFailed(("iPage=%d\n", iPage));
2620	rc = VERR_INTERNAL_ERROR;
2621	break;
2622	}
2623	}
2624	if (RT_SUCCESS(rc))
2625	rc = supdrvMemAdd(&Mem, pSession);
2626	if (RT_FAILURE(rc))
2627	{
2628	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
2629	AssertRC(rc2);
2630	}
2631	}
2632
2633	return rc;
2634	}
2635
2636
2637	/**
2638	* Unlocks the memory pointed to by pv.
2639	*
2640	* @returns IPRT status code.
2641	* @param pSession Session to which the memory was locked.
2642	* @param pvR3 Memory to unlock.
2643	*/
2644	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2645	{
2646	LogFlow(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2647	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2648	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
2649	}
2650
2651
2652	/**
2653	* Allocates a chunk of page aligned memory with contiguous and fixed physical
2654	* backing.
2655	*
2656	* @returns IPRT status code.
2657	* @param pSession Session data.
2658	* @param cPages Number of pages to allocate.
2659	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
2660	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
2661	* @param pHCPhys Where to put the physical address of allocated memory.
2662	*/
2663	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
2664	{
2665	int rc;
2666	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2667	LogFlow(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
2668
2669	/*
2670	* Validate input.
2671	*/
2672	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2673	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
2674	{
2675	Log(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
2676	pSession, ppvR0, ppvR3, pHCPhys));
2677	return VERR_INVALID_PARAMETER;
2678
2679	}
2680	if (cPages < 1 \|\| cPages >= 256)
2681	{
2682	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2683	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2684	}
2685
2686	/*
2687	* Let IPRT do the job.
2688	*/
2689	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
2690	if (RT_SUCCESS(rc))
2691	{
2692	int rc2;
2693	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2694	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2695	if (RT_SUCCESS(rc))
2696	{
2697	Mem.eType = MEMREF_TYPE_CONT;
2698	rc = supdrvMemAdd(&Mem, pSession);
2699	if (!rc)
2700	{
2701	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2702	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2703	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
2704	return 0;
2705	}
2706
2707	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2708	AssertRC(rc2);
2709	}
2710	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2711	AssertRC(rc2);
2712	}
2713
2714	return rc;
2715	}
2716
2717
2718	/**
2719	* Frees memory allocated using SUPR0ContAlloc().
2720	*
2721	* @returns IPRT status code.
2722	* @param pSession The session to which the memory was allocated.
2723	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2724	*/
2725	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2726	{
2727	LogFlow(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2728	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2729	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
2730	}
2731
2732
2733	/**
2734	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
2735	*
2736	* The memory isn't zeroed.
2737	*
2738	* @returns IPRT status code.
2739	* @param pSession Session data.
2740	* @param cPages Number of pages to allocate.
2741	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
2742	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
2743	* @param paPages Where to put the physical addresses of allocated memory.
2744	*/
2745	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
2746	{
2747	unsigned iPage;
2748	int rc;
2749	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2750	LogFlow(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
2751
2752	/*
2753	* Validate input.
2754	*/
2755	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2756	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
2757	{
2758	Log(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
2759	pSession, ppvR3, ppvR0, paPages));
2760	return VERR_INVALID_PARAMETER;
2761
2762	}
2763	if (cPages < 1 \|\| cPages >= 256)
2764	{
2765	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2766	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2767	}
2768
2769	/*
2770	* Let IPRT do the work.
2771	*/
2772	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
2773	if (RT_SUCCESS(rc))
2774	{
2775	int rc2;
2776	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2777	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2778	if (RT_SUCCESS(rc))
2779	{
2780	Mem.eType = MEMREF_TYPE_LOW;
2781	rc = supdrvMemAdd(&Mem, pSession);
2782	if (!rc)
2783	{
2784	for (iPage = 0; iPage < cPages; iPage++)
2785	{
2786	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2787	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%RHp\n", paPages[iPage]));
2788	}
2789	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2790	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2791	return 0;
2792	}
2793
2794	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2795	AssertRC(rc2);
2796	}
2797
2798	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2799	AssertRC(rc2);
2800	}
2801
2802	return rc;
2803	}
2804
2805
2806	/**
2807	* Frees memory allocated using SUPR0LowAlloc().
2808	*
2809	* @returns IPRT status code.
2810	* @param pSession The session to which the memory was allocated.
2811	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2812	*/
2813	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2814	{
2815	LogFlow(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2816	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2817	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
2818	}
2819
2820
2821
2822	/**
2823	* Allocates a chunk of memory with both R0 and R3 mappings.
2824	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
2825	*
2826	* @returns IPRT status code.
2827	* @param pSession The session to associated the allocation with.
2828	* @param cb Number of bytes to allocate.
2829	* @param ppvR0 Where to store the address of the Ring-0 mapping.
2830	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2831	*/
2832	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
2833	{
2834	int rc;
2835	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2836	LogFlow(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
2837
2838	/*
2839	* Validate input.
2840	*/
2841	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2842	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
2843	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
2844	if (cb < 1 \|\| cb >= _4M)
2845	{
2846	Log(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
2847	return VERR_INVALID_PARAMETER;
2848	}
2849
2850	/*
2851	* Let IPRT do the work.
2852	*/
2853	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
2854	if (RT_SUCCESS(rc))
2855	{
2856	int rc2;
2857	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2858	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2859	if (RT_SUCCESS(rc))
2860	{
2861	Mem.eType = MEMREF_TYPE_MEM;
2862	rc = supdrvMemAdd(&Mem, pSession);
2863	if (!rc)
2864	{
2865	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2866	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2867	return VINF_SUCCESS;
2868	}
2869
2870	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2871	AssertRC(rc2);
2872	}
2873
2874	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2875	AssertRC(rc2);
2876	}
2877
2878	return rc;
2879	}
2880
2881
2882	/**
2883	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
2884	*
2885	* @returns IPRT status code.
2886	* @param pSession The session to which the memory was allocated.
2887	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2888	* @param paPages Where to store the physical addresses.
2889	*/
2890	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
2891	{
2892	PSUPDRVBUNDLE pBundle;
2893	LogFlow(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
2894
2895	/*
2896	* Validate input.
2897	*/
2898	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2899	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
2900	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
2901
2902	/*
2903	* Search for the address.
2904	*/
2905	RTSpinlockAcquire(pSession->Spinlock);
2906	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2907	{
2908	if (pBundle->cUsed > 0)
2909	{
2910	unsigned i;
2911	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2912	{
2913	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
2914	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2915	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2916	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2917	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
2918	)
2919	)
2920	{
2921	const size_t cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2922	size_t iPage;
2923	for (iPage = 0; iPage < cPages; iPage++)
2924	{
2925	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2926	paPages[iPage].uReserved = 0;
2927	}
2928	RTSpinlockRelease(pSession->Spinlock);
2929	return VINF_SUCCESS;
2930	}
2931	}
2932	}
2933	}
2934	RTSpinlockRelease(pSession->Spinlock);
2935	Log(("Failed to find %p!!!\n", (void *)uPtr));
2936	return VERR_INVALID_PARAMETER;
2937	}
2938
2939
2940	/**
2941	* Free memory allocated by SUPR0MemAlloc().
2942	*
2943	* @returns IPRT status code.
2944	* @param pSession The session owning the allocation.
2945	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2946	*/
2947	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2948	{
2949	LogFlow(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2950	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2951	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
2952	}
2953
2954
2955	/**
2956	* Allocates a chunk of memory with a kernel or/and a user mode mapping.
2957	*
2958	* The memory is fixed and it's possible to query the physical addresses using
2959	* SUPR0MemGetPhys().
2960	*
2961	* @returns IPRT status code.
2962	* @param pSession The session to associated the allocation with.
2963	* @param cPages The number of pages to allocate.
2964	* @param fFlags Flags, reserved for the future. Must be zero.
2965	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2966	* NULL if no ring-3 mapping.
2967	* @param ppvR3 Where to store the address of the Ring-0 mapping.
2968	* NULL if no ring-0 mapping.
2969	* @param paPages Where to store the addresses of the pages. Optional.
2970	*/
2971	SUPR0DECL(int) SUPR0PageAllocEx(PSUPDRVSESSION pSession, uint32_t cPages, uint32_t fFlags, PRTR3PTR ppvR3, PRTR0PTR ppvR0, PRTHCPHYS paPages)
2972	{
2973	int rc;
2974	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2975	LogFlow(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
2976
2977	/*
2978	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2979	*/
2980	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2981	AssertPtrNullReturn(ppvR3, VERR_INVALID_POINTER);
2982	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2983	AssertReturn(ppvR3 \|\| ppvR0, VERR_INVALID_PARAMETER);
2984	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2985	if (cPages < 1 \|\| cPages > VBOX_MAX_ALLOC_PAGE_COUNT)
2986	{
2987	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)));
2988	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2989	}
2990
2991	/*
2992	* Let IPRT do the work.
2993	*/
2994	if (ppvR0)
2995	rc = RTR0MemObjAllocPage(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, true /* fExecutable */);
2996	else
2997	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
2998	if (RT_SUCCESS(rc))
2999	{
3000	int rc2;
3001	if (ppvR3)
3002	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
3003	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
3004	else
3005	Mem.MapObjR3 = NIL_RTR0MEMOBJ;
3006	if (RT_SUCCESS(rc))
3007	{
3008	Mem.eType = MEMREF_TYPE_PAGE;
3009	rc = supdrvMemAdd(&Mem, pSession);
3010	if (!rc)
3011	{
3012	if (ppvR3)
3013	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
3014	if (ppvR0)
3015	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
3016	if (paPages)
3017	{
3018	uint32_t iPage = cPages;
3019	while (iPage-- > 0)
3020	{
3021	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
3022	Assert(paPages[iPage] != NIL_RTHCPHYS);
3023	}
3024	}
3025	return VINF_SUCCESS;
3026	}
3027
3028	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
3029	AssertRC(rc2);
3030	}
3031
3032	rc2 = RTR0MemObjFree(Mem.MemObj, false);
3033	AssertRC(rc2);
3034	}
3035	return rc;
3036	}
3037
3038
3039	/**
3040	* Maps a chunk of memory previously allocated by SUPR0PageAllocEx into kernel
3041	* space.
3042	*
3043	* @returns IPRT status code.
3044	* @param pSession The session to associated the allocation with.
3045	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
3046	* @param offSub Where to start mapping. Must be page aligned.
3047	* @param cbSub How much to map. Must be page aligned.
3048	* @param fFlags Flags, MBZ.
3049	* @param ppvR0 Where to return the address of the ring-0 mapping on
3050	* success.
3051	*/
3052	SUPR0DECL(int) SUPR0PageMapKernel(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t offSub, uint32_t cbSub,
3053	uint32_t fFlags, PRTR0PTR ppvR0)
3054	{
3055	int rc;
3056	PSUPDRVBUNDLE pBundle;
3057	RTR0MEMOBJ hMemObj = NIL_RTR0MEMOBJ;
3058	LogFlow(("SUPR0PageMapKernel: pSession=%p pvR3=%p offSub=%#x cbSub=%#x\n", pSession, pvR3, offSub, cbSub));
3059
3060	/*
3061	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
3062	*/
3063	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3064	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
3065	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
3066	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3067	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3068	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
3069
3070	/*
3071	* Find the memory object.
3072	*/
3073	RTSpinlockAcquire(pSession->Spinlock);
3074	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3075	{
3076	if (pBundle->cUsed > 0)
3077	{
3078	unsigned i;
3079	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3080	{
3081	if ( ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
3082	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3083	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3084	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
3085	\|\| ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED
3086	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3087	&& pBundle->aMem[i].MapObjR3 == NIL_RTR0MEMOBJ
3088	&& RTR0MemObjAddressR3(pBundle->aMem[i].MemObj) == pvR3))
3089	{
3090	hMemObj = pBundle->aMem[i].MemObj;
3091	break;
3092	}
3093	}
3094	}
3095	}
3096	RTSpinlockRelease(pSession->Spinlock);
3097
3098	rc = VERR_INVALID_PARAMETER;
3099	if (hMemObj != NIL_RTR0MEMOBJ)
3100	{
3101	/*
3102	* Do some further input validations before calling IPRT.
3103	* (Cleanup is done indirectly by telling RTR0MemObjFree to include mappings.)
3104	*/
3105	size_t cbMemObj = RTR0MemObjSize(hMemObj);
3106	if ( offSub < cbMemObj
3107	&& cbSub <= cbMemObj
3108	&& offSub + cbSub <= cbMemObj)
3109	{
3110	RTR0MEMOBJ hMapObj;
3111	rc = RTR0MemObjMapKernelEx(&hMapObj, hMemObj, (void *)-1, 0,
3112	RTMEM_PROT_READ \| RTMEM_PROT_WRITE, offSub, cbSub);
3113	if (RT_SUCCESS(rc))
3114	*ppvR0 = RTR0MemObjAddress(hMapObj);
3115	}
3116	else
3117	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
3118
3119	}
3120	return rc;
3121	}
3122
3123
3124	/**
3125	* Changes the page level protection of one or more pages previously allocated
3126	* by SUPR0PageAllocEx.
3127	*
3128	* @returns IPRT status code.
3129	* @param pSession The session to associated the allocation with.
3130	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
3131	* NIL_RTR3PTR if the ring-3 mapping should be unaffected.
3132	* @param pvR0 The ring-0 address returned by SUPR0PageAllocEx.
3133	* NIL_RTR0PTR if the ring-0 mapping should be unaffected.
3134	* @param offSub Where to start changing. Must be page aligned.
3135	* @param cbSub How much to change. Must be page aligned.
3136	* @param fProt The new page level protection, see RTMEM_PROT_*.
3137	*/
3138	SUPR0DECL(int) SUPR0PageProtect(PSUPDRVSESSION pSession, RTR3PTR pvR3, RTR0PTR pvR0, uint32_t offSub, uint32_t cbSub, uint32_t fProt)
3139	{
3140	int rc;
3141	PSUPDRVBUNDLE pBundle;
3142	RTR0MEMOBJ hMemObjR0 = NIL_RTR0MEMOBJ;
3143	RTR0MEMOBJ hMemObjR3 = NIL_RTR0MEMOBJ;
3144	LogFlow(("SUPR0PageProtect: pSession=%p pvR3=%p pvR0=%p offSub=%#x cbSub=%#x fProt-%#x\n", pSession, pvR3, pvR0, offSub, cbSub, fProt));
3145
3146	/*
3147	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
3148	*/
3149	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3150	AssertReturn(!(fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)), VERR_INVALID_PARAMETER);
3151	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3152	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3153	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
3154
3155	/*
3156	* Find the memory object.
3157	*/
3158	RTSpinlockAcquire(pSession->Spinlock);
3159	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3160	{
3161	if (pBundle->cUsed > 0)
3162	{
3163	unsigned i;
3164	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3165	{
3166	if ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
3167	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3168	&& ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3169	\|\| pvR3 == NIL_RTR3PTR)
3170	&& ( pvR0 == NIL_RTR0PTR
3171	\|\| RTR0MemObjAddress(pBundle->aMem[i].MemObj) == pvR0)
3172	&& ( pvR3 == NIL_RTR3PTR
3173	\|\| RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3))
3174	{
3175	if (pvR0 != NIL_RTR0PTR)
3176	hMemObjR0 = pBundle->aMem[i].MemObj;
3177	if (pvR3 != NIL_RTR3PTR)
3178	hMemObjR3 = pBundle->aMem[i].MapObjR3;
3179	break;
3180	}
3181	}
3182	}
3183	}
3184	RTSpinlockRelease(pSession->Spinlock);
3185
3186	rc = VERR_INVALID_PARAMETER;
3187	if ( hMemObjR0 != NIL_RTR0MEMOBJ
3188	\|\| hMemObjR3 != NIL_RTR0MEMOBJ)
3189	{
3190	/*
3191	* Do some further input validations before calling IPRT.
3192	*/
3193	size_t cbMemObj = hMemObjR0 != NIL_RTR0PTR ? RTR0MemObjSize(hMemObjR0) : RTR0MemObjSize(hMemObjR3);
3194	if ( offSub < cbMemObj
3195	&& cbSub <= cbMemObj
3196	&& offSub + cbSub <= cbMemObj)
3197	{
3198	rc = VINF_SUCCESS;
3199	if (hMemObjR3 != NIL_RTR0PTR)
3200	rc = RTR0MemObjProtect(hMemObjR3, offSub, cbSub, fProt);
3201	if (hMemObjR0 != NIL_RTR0PTR && RT_SUCCESS(rc))
3202	rc = RTR0MemObjProtect(hMemObjR0, offSub, cbSub, fProt);
3203	}
3204	else
3205	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
3206
3207	}
3208	return rc;
3209
3210	}
3211
3212
3213	/**
3214	* Free memory allocated by SUPR0PageAlloc() and SUPR0PageAllocEx().
3215	*
3216	* @returns IPRT status code.
3217	* @param pSession The session owning the allocation.
3218	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc() or
3219	* SUPR0PageAllocEx().
3220	*/
3221	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
3222	{
3223	LogFlow(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
3224	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3225	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_PAGE);
3226	}
3227
3228
3229	/**
3230	* Gets the paging mode of the current CPU.
3231	*
3232	* @returns Paging mode, SUPPAGEINGMODE_INVALID on error.
3233	*/
3234	SUPR0DECL(SUPPAGINGMODE) SUPR0GetPagingMode(void)
3235	{
3236	SUPPAGINGMODE enmMode;
3237
3238	RTR0UINTREG cr0 = ASMGetCR0();
3239	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
3240	enmMode = SUPPAGINGMODE_INVALID;
3241	else
3242	{
3243	RTR0UINTREG cr4 = ASMGetCR4();
3244	uint32_t fNXEPlusLMA = 0;
3245	if (cr4 & X86_CR4_PAE)
3246	{
3247	uint32_t fExtFeatures = ASMCpuId_EDX(0x80000001);
3248	if (fExtFeatures & (X86_CPUID_EXT_FEATURE_EDX_NX \| X86_CPUID_EXT_FEATURE_EDX_LONG_MODE))
3249	{
3250	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
3251	if ((fExtFeatures & X86_CPUID_EXT_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
3252	fNXEPlusLMA \|= RT_BIT(0);
3253	if ((fExtFeatures & X86_CPUID_EXT_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
3254	fNXEPlusLMA \|= RT_BIT(1);
3255	}
3256	}
3257
3258	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
3259	{
3260	case 0:
3261	enmMode = SUPPAGINGMODE_32_BIT;
3262	break;
3263
3264	case X86_CR4_PGE:
3265	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
3266	break;
3267
3268	case X86_CR4_PAE:
3269	enmMode = SUPPAGINGMODE_PAE;
3270	break;
3271
3272	case X86_CR4_PAE \| RT_BIT(0):
3273	enmMode = SUPPAGINGMODE_PAE_NX;
3274	break;
3275
3276	case X86_CR4_PAE \| X86_CR4_PGE:
3277	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3278	break;
3279
3280	case X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3281	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3282	break;
3283
3284	case RT_BIT(1) \| X86_CR4_PAE:
3285	enmMode = SUPPAGINGMODE_AMD64;
3286	break;
3287
3288	case RT_BIT(1) \| X86_CR4_PAE \| RT_BIT(0):
3289	enmMode = SUPPAGINGMODE_AMD64_NX;
3290	break;
3291
3292	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
3293	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
3294	break;
3295
3296	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3297	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
3298	break;
3299
3300	default:
3301	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
3302	enmMode = SUPPAGINGMODE_INVALID;
3303	break;
3304	}
3305	}
3306	return enmMode;
3307	}
3308
3309
3310	/**
3311	* Enables or disabled hardware virtualization extensions using native OS APIs.
3312	*
3313	* @returns VBox status code.
3314	* @retval VINF_SUCCESS on success.
3315	* @retval VERR_NOT_SUPPORTED if not supported by the native OS.
3316	*
3317	* @param fEnable Whether to enable or disable.
3318	*/
3319	SUPR0DECL(int) SUPR0EnableVTx(bool fEnable)
3320	{
3321	#ifdef RT_OS_DARWIN
3322	return supdrvOSEnableVTx(fEnable);
3323	#else
3324	return VERR_NOT_SUPPORTED;
3325	#endif
3326	}
3327
3328
3329	/**
3330	* Suspends hardware virtualization extensions using the native OS API.
3331	*
3332	* This is called prior to entering raw-mode context.
3333	*
3334	* @returns @c true if suspended, @c false if not.
3335	*/
3336	SUPR0DECL(bool) SUPR0SuspendVTxOnCpu(void)
3337	{
3338	#ifdef RT_OS_DARWIN
3339	return supdrvOSSuspendVTxOnCpu();
3340	#else
3341	return false;
3342	#endif
3343	}
3344
3345
3346	/**
3347	* Resumes hardware virtualization extensions using the native OS API.
3348	*
3349	* This is called after to entering raw-mode context.
3350	*
3351	* @param fSuspended The return value of SUPR0SuspendVTxOnCpu.
3352	*/
3353	SUPR0DECL(void) SUPR0ResumeVTxOnCpu(bool fSuspended)
3354	{
3355	#ifdef RT_OS_DARWIN
3356	supdrvOSResumeVTxOnCpu(fSuspended);
3357	#else
3358	Assert(!fSuspended);
3359	#endif
3360	}
3361
3362
3363	/**
3364	* Quries the AMD-V and VT-x capabilities of the calling CPU.
3365	*
3366	* @returns VBox status code.
3367	* @retval VERR_VMX_NO_VMX
3368	* @retval VERR_VMX_MSR_LOCKED_OR_DISABLED
3369	* @retval VERR_SVM_NO_SVM
3370	* @retval VERR_SVM_DISABLED
3371	* @retval VERR_UNSUPPORTED_CPU if not identifiable as an AMD, Intel or VIA
3372	* (centaur) CPU.
3373	*
3374	* @param pSession The session handle.
3375	* @param pfCaps Where to store the capabilities.
3376	*/
3377	SUPR0DECL(int) SUPR0QueryVTCaps(PSUPDRVSESSION pSession, uint32_t *pfCaps)
3378	{
3379	/*
3380	* Input validation.
3381	*/
3382	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3383	AssertPtrReturn(pfCaps, VERR_INVALID_POINTER);
3384
3385	*pfCaps = 0;
3386
3387	if (ASMHasCpuId())
3388	{
3389	uint32_t fFeaturesECX, fFeaturesEDX, uDummy;
3390	uint32_t uMaxId, uVendorEBX, uVendorECX, uVendorEDX;
3391	uint64_t u64Value;
3392
3393	ASMCpuId(0, &uMaxId, &uVendorEBX, &uVendorECX, &uVendorEDX);
3394	ASMCpuId(1, &uDummy, &uDummy, &fFeaturesECX, &fFeaturesEDX);
3395
3396	if ( ASMIsValidStdRange(uMaxId)
3397	&& ( ASMIsIntelCpuEx( uVendorEBX, uVendorECX, uVendorEDX)
3398	\|\| ASMIsViaCentaurCpuEx(uVendorEBX, uVendorECX, uVendorEDX) )
3399	)
3400	{
3401	if ( (fFeaturesECX & X86_CPUID_FEATURE_ECX_VMX)
3402	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3403	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3404	)
3405	{
3406	/*
3407	* Both the LOCK and VMXON bit must be set; otherwise VMXON will generate a #GP.
3408	* Once the lock bit is set, this MSR can no longer be modified.
3409	*/
3410	u64Value = ASMRdMsr(MSR_IA32_FEATURE_CONTROL);
3411	if ( (u64Value & (MSR_IA32_FEATURE_CONTROL_VMXON \| MSR_IA32_FEATURE_CONTROL_LOCK))
3412	== (MSR_IA32_FEATURE_CONTROL_VMXON \| MSR_IA32_FEATURE_CONTROL_LOCK) /* enabled and locked */
3413	\|\| !(u64Value & MSR_IA32_FEATURE_CONTROL_LOCK) /* not enabled, but not locked either */
3414	)
3415	{
3416	VMX_CAPABILITY vtCaps;
3417
3418	*pfCaps \|= SUPVTCAPS_VT_X;
3419
3420	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS);
3421	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC_USE_SECONDARY_EXEC_CTRL)
3422	{
3423	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS2);
3424	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC2_EPT)
3425	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3426	}
3427	return VINF_SUCCESS;
3428	}
3429	return VERR_VMX_MSR_LOCKED_OR_DISABLED;
3430	}
3431	return VERR_VMX_NO_VMX;
3432	}
3433
3434	if ( ASMIsAmdCpuEx(uVendorEBX, uVendorECX, uVendorEDX)
3435	&& ASMIsValidStdRange(uMaxId))
3436	{
3437	uint32_t fExtFeaturesEcx, uExtMaxId;
3438	ASMCpuId(0x80000000, &uExtMaxId, &uDummy, &uDummy, &uDummy);
3439	ASMCpuId(0x80000001, &uDummy, &uDummy, &fExtFeaturesEcx, &uDummy);
3440	if ( ASMIsValidExtRange(uExtMaxId)
3441	&& uExtMaxId >= 0x8000000a
3442	&& (fExtFeaturesEcx & X86_CPUID_AMD_FEATURE_ECX_SVM)
3443	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3444	&& (fFeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3445	)
3446	{
3447	/* Check if SVM is disabled */
3448	u64Value = ASMRdMsr(MSR_K8_VM_CR);
3449	if (!(u64Value & MSR_K8_VM_CR_SVM_DISABLE))
3450	{
3451	uint32_t fSvmFeatures;
3452	*pfCaps \|= SUPVTCAPS_AMD_V;
3453
3454	/* Query AMD-V features. */
3455	ASMCpuId(0x8000000a, &uDummy, &uDummy, &uDummy, &fSvmFeatures);
3456	if (fSvmFeatures & AMD_CPUID_SVM_FEATURE_EDX_NESTED_PAGING)
3457	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3458
3459	return VINF_SUCCESS;
3460	}
3461	return VERR_SVM_DISABLED;
3462	}
3463	return VERR_SVM_NO_SVM;
3464	}
3465	}
3466
3467	return VERR_UNSUPPORTED_CPU;
3468	}
3469
3470
3471	/**
3472	* (Re-)initializes the per-cpu structure prior to starting or resuming the GIP
3473	* updating.
3474	*
3475	* @param pGipCpu The per CPU structure for this CPU.
3476	* @param u64NanoTS The current time.
3477	*/
3478	static void supdrvGipReInitCpu(PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
3479	{
3480	pGipCpu->u64TSC = ASMReadTSC() - pGipCpu->u32UpdateIntervalTSC;
3481	pGipCpu->u64NanoTS = u64NanoTS;
3482	}
3483
3484
3485	/**
3486	* Set the current TSC and NanoTS value for the CPU.
3487	*
3488	* @param idCpu The CPU ID. Unused - we have to use the APIC ID.
3489	* @param pvUser1 Pointer to the ring-0 GIP mapping.
3490	* @param pvUser2 Pointer to the variable holding the current time.
3491	*/
3492	static DECLCALLBACK(void) supdrvGipReInitCpuCallback(RTCPUID idCpu, void pvUser1, void pvUser2)
3493	{
3494	PSUPGLOBALINFOPAGE pGip = (PSUPGLOBALINFOPAGE)pvUser1;
3495	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
3496
3497	if (RT_LIKELY(iCpu < pGip->cCpus && pGip->aCPUs[iCpu].idCpu == idCpu))
3498	supdrvGipReInitCpu(&pGip->aCPUs[iCpu], (uint64_t )pvUser2);
3499
3500	NOREF(pvUser2);
3501	NOREF(idCpu);
3502	}
3503
3504
3505	/**
3506	* Maps the GIP into userspace and/or get the physical address of the GIP.
3507	*
3508	* @returns IPRT status code.
3509	* @param pSession Session to which the GIP mapping should belong.
3510	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
3511	* @param pHCPhysGip Where to store the physical address. (optional)
3512	*
3513	* @remark There is no reference counting on the mapping, so one call to this function
3514	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
3515	* and remove the session as a GIP user.
3516	*/
3517	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
3518	{
3519	int rc;
3520	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3521	RTR3PTR pGipR3 = NIL_RTR3PTR;
3522	RTHCPHYS HCPhys = NIL_RTHCPHYS;
3523	LogFlow(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
3524
3525	/*
3526	* Validate
3527	*/
3528	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3529	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
3530	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
3531
3532	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3533	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3534	#else
3535	RTSemFastMutexRequest(pDevExt->mtxGip);
3536	#endif
3537	if (pDevExt->pGip)
3538	{
3539	/*
3540	* Map it?
3541	*/
3542	rc = VINF_SUCCESS;
3543	if (ppGipR3)
3544	{
3545	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
3546	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
3547	RTMEM_PROT_READ, RTR0ProcHandleSelf());
3548	if (RT_SUCCESS(rc))
3549	pGipR3 = RTR0MemObjAddressR3(pSession->GipMapObjR3);
3550	}
3551
3552	/*
3553	* Get physical address.
3554	*/
3555	if (pHCPhysGip && RT_SUCCESS(rc))
3556	HCPhys = pDevExt->HCPhysGip;
3557
3558	/*
3559	* Reference globally.
3560	*/
3561	if (!pSession->fGipReferenced && RT_SUCCESS(rc))
3562	{
3563	pSession->fGipReferenced = 1;
3564	pDevExt->cGipUsers++;
3565	if (pDevExt->cGipUsers == 1)
3566	{
3567	PSUPGLOBALINFOPAGE pGipR0 = pDevExt->pGip;
3568	uint64_t u64NanoTS;
3569	uint32_t u32SystemResolution;
3570	unsigned i;
3571
3572	LogFlow(("SUPR0GipMap: Resumes GIP updating\n"));
3573
3574	/*
3575	* Try bump up the system timer resolution.
3576	* The more interrupts the better...
3577	*/
3578	if ( RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
3579	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
3580	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1953125 /* 512 HZ */, &u32SystemResolution))
3581	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 2000000 /* 500 HZ */, &u32SystemResolution))
3582	)
3583	{
3584	Assert(RTTimerGetSystemGranularity() <= u32SystemResolution);
3585	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
3586	}
3587
3588	if (pGipR0->aCPUs[0].u32TransactionId != 2 /* not the first time */)
3589	{
3590	for (i = 0; i < RT_ELEMENTS(pGipR0->aCPUs); i++)
3591	ASMAtomicUoWriteU32(&pGipR0->aCPUs[i].u32TransactionId,
3592	(pGipR0->aCPUs[i].u32TransactionId + GIP_UPDATEHZ_RECALC_FREQ * 2)
3593	& ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
3594	ASMAtomicWriteU64(&pGipR0->u64NanoTSLastUpdateHz, 0);
3595	}
3596
3597	u64NanoTS = RTTimeSystemNanoTS() - pGipR0->u32UpdateIntervalNS;
3598	if ( pGipR0->u32Mode == SUPGIPMODE_SYNC_TSC
3599	\|\| RTMpGetOnlineCount() == 1)
3600	supdrvGipReInitCpu(&pGipR0->aCPUs[0], u64NanoTS);
3601	else
3602	RTMpOnAll(supdrvGipReInitCpuCallback, pGipR0, &u64NanoTS);
3603
3604	#ifndef DO_NOT_START_GIP
3605	rc = RTTimerStart(pDevExt->pGipTimer, 0); AssertRC(rc);
3606	#endif
3607	rc = VINF_SUCCESS;
3608	}
3609	}
3610	}
3611	else
3612	{
3613	rc = VERR_GENERAL_FAILURE;
3614	Log(("SUPR0GipMap: GIP is not available!\n"));
3615	}
3616	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3617	RTSemMutexRelease(pDevExt->mtxGip);
3618	#else
3619	RTSemFastMutexRelease(pDevExt->mtxGip);
3620	#endif
3621
3622	/*
3623	* Write returns.
3624	*/
3625	if (pHCPhysGip)
3626	*pHCPhysGip = HCPhys;
3627	if (ppGipR3)
3628	*ppGipR3 = pGipR3;
3629
3630	#ifdef DEBUG_DARWIN_GIP
3631	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3632	#else
3633	LogFlow(( "SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3634	#endif
3635	return rc;
3636	}
3637
3638
3639	/**
3640	* Unmaps any user mapping of the GIP and terminates all GIP access
3641	* from this session.
3642	*
3643	* @returns IPRT status code.
3644	* @param pSession Session to which the GIP mapping should belong.
3645	*/
3646	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
3647	{
3648	int rc = VINF_SUCCESS;
3649	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3650	#ifdef DEBUG_DARWIN_GIP
3651	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
3652	pSession,
3653	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
3654	pSession->GipMapObjR3));
3655	#else
3656	LogFlow(("SUPR0GipUnmap: pSession=%p\n", pSession));
3657	#endif
3658	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3659
3660	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3661	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3662	#else
3663	RTSemFastMutexRequest(pDevExt->mtxGip);
3664	#endif
3665
3666	/*
3667	* Unmap anything?
3668	*/
3669	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
3670	{
3671	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
3672	AssertRC(rc);
3673	if (RT_SUCCESS(rc))
3674	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
3675	}
3676
3677	/*
3678	* Dereference global GIP.
3679	*/
3680	if (pSession->fGipReferenced && !rc)
3681	{
3682	pSession->fGipReferenced = 0;
3683	if ( pDevExt->cGipUsers > 0
3684	&& !--pDevExt->cGipUsers)
3685	{
3686	LogFlow(("SUPR0GipUnmap: Suspends GIP updating\n"));
3687	#ifndef DO_NOT_START_GIP
3688	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = VINF_SUCCESS;
3689	#endif
3690
3691	if (pDevExt->u32SystemTimerGranularityGrant)
3692	{
3693	int rc2 = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant);
3694	AssertRC(rc2);
3695	pDevExt->u32SystemTimerGranularityGrant = 0;
3696	}
3697	}
3698	}
3699
3700	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3701	RTSemMutexRelease(pDevExt->mtxGip);
3702	#else
3703	RTSemFastMutexRelease(pDevExt->mtxGip);
3704	#endif
3705
3706	return rc;
3707	}
3708
3709
3710	/**
3711	* Gets the GIP pointer.
3712	*
3713	* @returns Pointer to the GIP or NULL.
3714	*/
3715	SUPDECL(PSUPGLOBALINFOPAGE) SUPGetGIP(void)
3716	{
3717	return g_pSUPGlobalInfoPage;
3718	}
3719
3720
3721	/**
3722	* Register a component factory with the support driver.
3723	*
3724	* This is currently restricted to kernel sessions only.
3725	*
3726	* @returns VBox status code.
3727	* @retval VINF_SUCCESS on success.
3728	* @retval VERR_NO_MEMORY if we're out of memory.
3729	* @retval VERR_ALREADY_EXISTS if the factory has already been registered.
3730	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3731	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3732	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3733	*
3734	* @param pSession The SUPDRV session (must be a ring-0 session).
3735	* @param pFactory Pointer to the component factory registration structure.
3736	*
3737	* @remarks This interface is also available via SUPR0IdcComponentRegisterFactory.
3738	*/
3739	SUPR0DECL(int) SUPR0ComponentRegisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3740	{
3741	PSUPDRVFACTORYREG pNewReg;
3742	const char *psz;
3743	int rc;
3744
3745	/*
3746	* Validate parameters.
3747	*/
3748	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3749	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3750	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3751	AssertPtrReturn(pFactory->pfnQueryFactoryInterface, VERR_INVALID_POINTER);
3752	psz = RTStrEnd(pFactory->szName, sizeof(pFactory->szName));
3753	AssertReturn(psz, VERR_INVALID_PARAMETER);
3754
3755	/*
3756	* Allocate and initialize a new registration structure.
3757	*/
3758	pNewReg = (PSUPDRVFACTORYREG)RTMemAlloc(sizeof(SUPDRVFACTORYREG));
3759	if (pNewReg)
3760	{
3761	pNewReg->pNext = NULL;
3762	pNewReg->pFactory = pFactory;
3763	pNewReg->pSession = pSession;
3764	pNewReg->cchName = psz - &pFactory->szName[0];
3765
3766	/*
3767	* Add it to the tail of the list after checking for prior registration.
3768	*/
3769	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3770	if (RT_SUCCESS(rc))
3771	{
3772	PSUPDRVFACTORYREG pPrev = NULL;
3773	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3774	while (pCur && pCur->pFactory != pFactory)
3775	{
3776	pPrev = pCur;
3777	pCur = pCur->pNext;
3778	}
3779	if (!pCur)
3780	{
3781	if (pPrev)
3782	pPrev->pNext = pNewReg;
3783	else
3784	pSession->pDevExt->pComponentFactoryHead = pNewReg;
3785	rc = VINF_SUCCESS;
3786	}
3787	else
3788	rc = VERR_ALREADY_EXISTS;
3789
3790	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3791	}
3792
3793	if (RT_FAILURE(rc))
3794	RTMemFree(pNewReg);
3795	}
3796	else
3797	rc = VERR_NO_MEMORY;
3798	return rc;
3799	}
3800
3801
3802	/**
3803	* Deregister a component factory.
3804	*
3805	* @returns VBox status code.
3806	* @retval VINF_SUCCESS on success.
3807	* @retval VERR_NOT_FOUND if the factory wasn't registered.
3808	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3809	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3810	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3811	*
3812	* @param pSession The SUPDRV session (must be a ring-0 session).
3813	* @param pFactory Pointer to the component factory registration structure
3814	* previously passed SUPR0ComponentRegisterFactory().
3815	*
3816	* @remarks This interface is also available via SUPR0IdcComponentDeregisterFactory.
3817	*/
3818	SUPR0DECL(int) SUPR0ComponentDeregisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3819	{
3820	int rc;
3821
3822	/*
3823	* Validate parameters.
3824	*/
3825	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3826	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3827	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3828
3829	/*
3830	* Take the lock and look for the registration record.
3831	*/
3832	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3833	if (RT_SUCCESS(rc))
3834	{
3835	PSUPDRVFACTORYREG pPrev = NULL;
3836	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3837	while (pCur && pCur->pFactory != pFactory)
3838	{
3839	pPrev = pCur;
3840	pCur = pCur->pNext;
3841	}
3842	if (pCur)
3843	{
3844	if (!pPrev)
3845	pSession->pDevExt->pComponentFactoryHead = pCur->pNext;
3846	else
3847	pPrev->pNext = pCur->pNext;
3848
3849	pCur->pNext = NULL;
3850	pCur->pFactory = NULL;
3851	pCur->pSession = NULL;
3852	rc = VINF_SUCCESS;
3853	}
3854	else
3855	rc = VERR_NOT_FOUND;
3856
3857	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3858
3859	RTMemFree(pCur);
3860	}
3861	return rc;
3862	}
3863
3864
3865	/**
3866	* Queries a component factory.
3867	*
3868	* @returns VBox status code.
3869	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3870	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3871	* @retval VERR_SUPDRV_COMPONENT_NOT_FOUND if the component factory wasn't found.
3872	* @retval VERR_SUPDRV_INTERFACE_NOT_SUPPORTED if the interface wasn't supported.
3873	*
3874	* @param pSession The SUPDRV session.
3875	* @param pszName The name of the component factory.
3876	* @param pszInterfaceUuid The UUID of the factory interface (stringified).
3877	* @param ppvFactoryIf Where to store the factory interface.
3878	*/
3879	SUPR0DECL(int) SUPR0ComponentQueryFactory(PSUPDRVSESSION pSession, const char pszName, const char pszInterfaceUuid, void **ppvFactoryIf)
3880	{
3881	const char *pszEnd;
3882	size_t cchName;
3883	int rc;
3884
3885	/*
3886	* Validate parameters.
3887	*/
3888	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3889
3890	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
3891	pszEnd = RTStrEnd(pszName, RT_SIZEOFMEMB(SUPDRVFACTORY, szName));
3892	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3893	cchName = pszEnd - pszName;
3894
3895	AssertPtrReturn(pszInterfaceUuid, VERR_INVALID_POINTER);
3896	pszEnd = RTStrEnd(pszInterfaceUuid, RTUUID_STR_LENGTH);
3897	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3898
3899	AssertPtrReturn(ppvFactoryIf, VERR_INVALID_POINTER);
3900	*ppvFactoryIf = NULL;
3901
3902	/*
3903	* Take the lock and try all factories by this name.
3904	*/
3905	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3906	if (RT_SUCCESS(rc))
3907	{
3908	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3909	rc = VERR_SUPDRV_COMPONENT_NOT_FOUND;
3910	while (pCur)
3911	{
3912	if ( pCur->cchName == cchName
3913	&& !memcmp(pCur->pFactory->szName, pszName, cchName))
3914	{
3915	void *pvFactory = pCur->pFactory->pfnQueryFactoryInterface(pCur->pFactory, pSession, pszInterfaceUuid);
3916	if (pvFactory)
3917	{
3918	*ppvFactoryIf = pvFactory;
3919	rc = VINF_SUCCESS;
3920	break;
3921	}
3922	rc = VERR_SUPDRV_INTERFACE_NOT_SUPPORTED;
3923	}
3924
3925	/* next */
3926	pCur = pCur->pNext;
3927	}
3928
3929	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3930	}
3931	return rc;
3932	}
3933
3934
3935	/**
3936	* Adds a memory object to the session.
3937	*
3938	* @returns IPRT status code.
3939	* @param pMem Memory tracking structure containing the
3940	* information to track.
3941	* @param pSession The session.
3942	*/
3943	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
3944	{
3945	PSUPDRVBUNDLE pBundle;
3946
3947	/*
3948	* Find free entry and record the allocation.
3949	*/
3950	RTSpinlockAcquire(pSession->Spinlock);
3951	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3952	{
3953	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
3954	{
3955	unsigned i;
3956	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3957	{
3958	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
3959	{
3960	pBundle->cUsed++;
3961	pBundle->aMem[i] = *pMem;
3962	RTSpinlockRelease(pSession->Spinlock);
3963	return VINF_SUCCESS;
3964	}
3965	}
3966	AssertFailed(); /* !!this can't be happening!!! */
3967	}
3968	}
3969	RTSpinlockRelease(pSession->Spinlock);
3970
3971	/*
3972	* Need to allocate a new bundle.
3973	* Insert into the last entry in the bundle.
3974	*/
3975	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
3976	if (!pBundle)
3977	return VERR_NO_MEMORY;
3978
3979	/* take last entry. */
3980	pBundle->cUsed++;
3981	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
3982
3983	/* insert into list. */
3984	RTSpinlockAcquire(pSession->Spinlock);
3985	pBundle->pNext = pSession->Bundle.pNext;
3986	pSession->Bundle.pNext = pBundle;
3987	RTSpinlockRelease(pSession->Spinlock);
3988
3989	return VINF_SUCCESS;
3990	}
3991
3992
3993	/**
3994	* Releases a memory object referenced by pointer and type.
3995	*
3996	* @returns IPRT status code.
3997	* @param pSession Session data.
3998	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
3999	* @param eType Memory type.
4000	*/
4001	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
4002	{
4003	PSUPDRVBUNDLE pBundle;
4004
4005	/*
4006	* Validate input.
4007	*/
4008	if (!uPtr)
4009	{
4010	Log(("Illegal address %p\n", (void *)uPtr));
4011	return VERR_INVALID_PARAMETER;
4012	}
4013
4014	/*
4015	* Search for the address.
4016	*/
4017	RTSpinlockAcquire(pSession->Spinlock);
4018	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
4019	{
4020	if (pBundle->cUsed > 0)
4021	{
4022	unsigned i;
4023	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
4024	{
4025	if ( pBundle->aMem[i].eType == eType
4026	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
4027	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
4028	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
4029	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
4030	)
4031	{
4032	/* Make a copy of it and release it outside the spinlock. */
4033	SUPDRVMEMREF Mem = pBundle->aMem[i];
4034	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
4035	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
4036	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
4037	RTSpinlockRelease(pSession->Spinlock);
4038
4039	if (Mem.MapObjR3 != NIL_RTR0MEMOBJ)
4040	{
4041	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
4042	AssertRC(rc); /** @todo figure out how to handle this. */
4043	}
4044	if (Mem.MemObj != NIL_RTR0MEMOBJ)
4045	{
4046	int rc = RTR0MemObjFree(Mem.MemObj, true /* fFreeMappings */);
4047	AssertRC(rc); /** @todo figure out how to handle this. */
4048	}
4049	return VINF_SUCCESS;
4050	}
4051	}
4052	}
4053	}
4054	RTSpinlockRelease(pSession->Spinlock);
4055	Log(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
4056	return VERR_INVALID_PARAMETER;
4057	}
4058
4059
4060	/**
4061	* Opens an image. If it's the first time it's opened the call must upload
4062	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
4063	*
4064	* This is the 1st step of the loading.
4065	*
4066	* @returns IPRT status code.
4067	* @param pDevExt Device globals.
4068	* @param pSession Session data.
4069	* @param pReq The open request.
4070	*/
4071	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
4072	{
4073	int rc;
4074	PSUPDRVLDRIMAGE pImage;
4075	void *pv;
4076	size_t cchName = strlen(pReq->u.In.szName); /* (caller checked < 32). */
4077	LogFlow(("supdrvIOCtl_LdrOpen: szName=%s cbImageWithTabs=%d\n", pReq->u.In.szName, pReq->u.In.cbImageWithTabs));
4078
4079	/*
4080	* Check if we got an instance of the image already.
4081	*/
4082	supdrvLdrLock(pDevExt);
4083	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
4084	{
4085	if ( pImage->szName[cchName] == '\0'
4086	&& !memcmp(pImage->szName, pReq->u.In.szName, cchName))
4087	{
4088	/** @todo check cbImageBits and cbImageWithTabs here, if they differs that indicates that the images are different. */
4089	pImage->cUsage++;
4090	pReq->u.Out.pvImageBase = pImage->pvImage;
4091	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
4092	pReq->u.Out.fNativeLoader = pImage->fNative;
4093	supdrvLdrAddUsage(pSession, pImage);
4094	supdrvLdrUnlock(pDevExt);
4095	return VINF_SUCCESS;
4096	}
4097	}
4098	/* (not found - add it!) */
4099
4100	/*
4101	* Allocate memory.
4102	*/
4103	Assert(cchName < sizeof(pImage->szName));
4104	pv = RTMemAlloc(sizeof(SUPDRVLDRIMAGE));
4105	if (!pv)
4106	{
4107	supdrvLdrUnlock(pDevExt);
4108	Log(("supdrvIOCtl_LdrOpen: RTMemAlloc() failed\n"));
4109	return /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_2;
4110	}
4111
4112	/*
4113	* Setup and link in the LDR stuff.
4114	*/
4115	pImage = (PSUPDRVLDRIMAGE)pv;
4116	pImage->pvImage = NULL;
4117	pImage->pvImageAlloc = NULL;
4118	pImage->cbImageWithTabs = pReq->u.In.cbImageWithTabs;
4119	pImage->cbImageBits = pReq->u.In.cbImageBits;
4120	pImage->cSymbols = 0;
4121	pImage->paSymbols = NULL;
4122	pImage->pachStrTab = NULL;
4123	pImage->cbStrTab = 0;
4124	pImage->pfnModuleInit = NULL;
4125	pImage->pfnModuleTerm = NULL;
4126	pImage->pfnServiceReqHandler = NULL;
4127	pImage->uState = SUP_IOCTL_LDR_OPEN;
4128	pImage->cUsage = 1;
4129	pImage->pDevExt = pDevExt;
4130	memcpy(pImage->szName, pReq->u.In.szName, cchName + 1);
4131
4132	/*
4133	* Try load it using the native loader, if that isn't supported, fall back
4134	* on the older method.
4135	*/
4136	pImage->fNative = true;
4137	rc = supdrvOSLdrOpen(pDevExt, pImage, pReq->u.In.szFilename);
4138	if (rc == VERR_NOT_SUPPORTED)
4139	{
4140	pImage->pvImageAlloc = RTMemExecAlloc(pImage->cbImageBits + 31);
4141	pImage->pvImage = RT_ALIGN_P(pImage->pvImageAlloc, 32);
4142	pImage->fNative = false;
4143	rc = pImage->pvImageAlloc ? VINF_SUCCESS : VERR_NO_EXEC_MEMORY;
4144	}
4145	if (RT_FAILURE(rc))
4146	{
4147	supdrvLdrUnlock(pDevExt);
4148	RTMemFree(pImage);
4149	Log(("supdrvIOCtl_LdrOpen(%s): failed - %Rrc\n", pReq->u.In.szName, rc));
4150	return rc;
4151	}
4152	Assert(VALID_PTR(pImage->pvImage) \|\| RT_FAILURE(rc));
4153
4154	/*
4155	* Link it.
4156	*/
4157	pImage->pNext = pDevExt->pLdrImages;
4158	pDevExt->pLdrImages = pImage;
4159
4160	supdrvLdrAddUsage(pSession, pImage);
4161
4162	pReq->u.Out.pvImageBase = pImage->pvImage;
4163	pReq->u.Out.fNeedsLoading = true;
4164	pReq->u.Out.fNativeLoader = pImage->fNative;
4165	supdrvOSLdrNotifyOpened(pDevExt, pImage);
4166
4167	supdrvLdrUnlock(pDevExt);
4168	return VINF_SUCCESS;
4169	}
4170
4171
4172	/**
4173	* Worker that validates a pointer to an image entrypoint.
4174	*
4175	* @returns IPRT status code.
4176	* @param pDevExt The device globals.
4177	* @param pImage The loader image.
4178	* @param pv The pointer into the image.
4179	* @param fMayBeNull Whether it may be NULL.
4180	* @param pszWhat What is this entrypoint? (for logging)
4181	* @param pbImageBits The image bits prepared by ring-3.
4182	*
4183	* @remarks Will leave the lock on failure.
4184	*/
4185	static int supdrvLdrValidatePointer(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage, void *pv,
4186	bool fMayBeNull, const uint8_t pbImageBits, const char pszWhat)
4187	{
4188	if (!fMayBeNull \|\| pv)
4189	{
4190	if ((uintptr_t)pv - (uintptr_t)pImage->pvImage >= pImage->cbImageBits)
4191	{
4192	supdrvLdrUnlock(pDevExt);
4193	Log(("Out of range (%p LB %#x): %s=%p\n", pImage->pvImage, pImage->cbImageBits, pszWhat, pv));
4194	return VERR_INVALID_PARAMETER;
4195	}
4196
4197	if (pImage->fNative)
4198	{
4199	int rc = supdrvOSLdrValidatePointer(pDevExt, pImage, pv, pbImageBits);
4200	if (RT_FAILURE(rc))
4201	{
4202	supdrvLdrUnlock(pDevExt);
4203	Log(("Bad entry point address: %s=%p (rc=%Rrc)\n", pszWhat, pv, rc));
4204	return rc;
4205	}
4206	}
4207	}
4208	return VINF_SUCCESS;
4209	}
4210
4211
4212	/**
4213	* Loads the image bits.
4214	*
4215	* This is the 2nd step of the loading.
4216	*
4217	* @returns IPRT status code.
4218	* @param pDevExt Device globals.
4219	* @param pSession Session data.
4220	* @param pReq The request.
4221	*/
4222	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
4223	{
4224	PSUPDRVLDRUSAGE pUsage;
4225	PSUPDRVLDRIMAGE pImage;
4226	int rc;
4227	LogFlow(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImageWithBits=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImageWithTabs));
4228
4229	/*
4230	* Find the ldr image.
4231	*/
4232	supdrvLdrLock(pDevExt);
4233	pUsage = pSession->pLdrUsage;
4234	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4235	pUsage = pUsage->pNext;
4236	if (!pUsage)
4237	{
4238	supdrvLdrUnlock(pDevExt);
4239	Log(("SUP_IOCTL_LDR_LOAD: couldn't find image!\n"));
4240	return VERR_INVALID_HANDLE;
4241	}
4242	pImage = pUsage->pImage;
4243
4244	/*
4245	* Validate input.
4246	*/
4247	if ( pImage->cbImageWithTabs != pReq->u.In.cbImageWithTabs
4248	\|\| pImage->cbImageBits != pReq->u.In.cbImageBits)
4249	{
4250	supdrvLdrUnlock(pDevExt);
4251	Log(("SUP_IOCTL_LDR_LOAD: image size mismatch!! %d(prep) != %d(load) or %d != %d\n",
4252	pImage->cbImageWithTabs, pReq->u.In.cbImageWithTabs, pImage->cbImageBits, pReq->u.In.cbImageBits));
4253	return VERR_INVALID_HANDLE;
4254	}
4255
4256	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
4257	{
4258	unsigned uState = pImage->uState;
4259	supdrvLdrUnlock(pDevExt);
4260	if (uState != SUP_IOCTL_LDR_LOAD)
4261	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
4262	return VERR_ALREADY_LOADED;
4263	}
4264
4265	switch (pReq->u.In.eEPType)
4266	{
4267	case SUPLDRLOADEP_NOTHING:
4268	break;
4269
4270	case SUPLDRLOADEP_VMMR0:
4271	rc = supdrvLdrValidatePointer( pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0, false, pReq->u.In.abImage, "pvVMMR0");
4272	if (RT_SUCCESS(rc))
4273	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt, false, pReq->u.In.abImage, "pvVMMR0EntryInt");
4274	if (RT_SUCCESS(rc))
4275	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, false, pReq->u.In.abImage, "pvVMMR0EntryFast");
4276	if (RT_SUCCESS(rc))
4277	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx, false, pReq->u.In.abImage, "pvVMMR0EntryEx");
4278	if (RT_FAILURE(rc))
4279	return rc;
4280	break;
4281
4282	case SUPLDRLOADEP_SERVICE:
4283	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.Service.pfnServiceReq, false, pReq->u.In.abImage, "pfnServiceReq");
4284	if (RT_FAILURE(rc))
4285	return rc;
4286	if ( pReq->u.In.EP.Service.apvReserved[0] != NIL_RTR0PTR
4287	\|\| pReq->u.In.EP.Service.apvReserved[1] != NIL_RTR0PTR
4288	\|\| pReq->u.In.EP.Service.apvReserved[2] != NIL_RTR0PTR)
4289	{
4290	supdrvLdrUnlock(pDevExt);
4291	Log(("Out of range (%p LB %#x): apvReserved={%p,%p,%p} MBZ!\n",
4292	pImage->pvImage, pReq->u.In.cbImageWithTabs,
4293	pReq->u.In.EP.Service.apvReserved[0],
4294	pReq->u.In.EP.Service.apvReserved[1],
4295	pReq->u.In.EP.Service.apvReserved[2]));
4296	return VERR_INVALID_PARAMETER;
4297	}
4298	break;
4299
4300	default:
4301	supdrvLdrUnlock(pDevExt);
4302	Log(("Invalid eEPType=%d\n", pReq->u.In.eEPType));
4303	return VERR_INVALID_PARAMETER;
4304	}
4305
4306	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleInit, true, pReq->u.In.abImage, "pfnModuleInit");
4307	if (RT_FAILURE(rc))
4308	return rc;
4309	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleTerm, true, pReq->u.In.abImage, "pfnModuleTerm");
4310	if (RT_FAILURE(rc))
4311	return rc;
4312
4313	/*
4314	* Allocate and copy the tables.
4315	* (No need to do try/except as this is a buffered request.)
4316	*/
4317	pImage->cbStrTab = pReq->u.In.cbStrTab;
4318	if (pImage->cbStrTab)
4319	{
4320	pImage->pachStrTab = (char *)RTMemAlloc(pImage->cbStrTab);
4321	if (pImage->pachStrTab)
4322	memcpy(pImage->pachStrTab, &pReq->u.In.abImage[pReq->u.In.offStrTab], pImage->cbStrTab);
4323	else
4324	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_3;
4325	}
4326
4327	pImage->cSymbols = pReq->u.In.cSymbols;
4328	if (RT_SUCCESS(rc) && pImage->cSymbols)
4329	{
4330	size_t cbSymbols = pImage->cSymbols * sizeof(SUPLDRSYM);
4331	pImage->paSymbols = (PSUPLDRSYM)RTMemAlloc(cbSymbols);
4332	if (pImage->paSymbols)
4333	memcpy(pImage->paSymbols, &pReq->u.In.abImage[pReq->u.In.offSymbols], cbSymbols);
4334	else
4335	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_4;
4336	}
4337
4338	/*
4339	* Copy the bits / complete native loading.
4340	*/
4341	if (RT_SUCCESS(rc))
4342	{
4343	pImage->uState = SUP_IOCTL_LDR_LOAD;
4344	pImage->pfnModuleInit = pReq->u.In.pfnModuleInit;
4345	pImage->pfnModuleTerm = pReq->u.In.pfnModuleTerm;
4346
4347	if (pImage->fNative)
4348	rc = supdrvOSLdrLoad(pDevExt, pImage, pReq->u.In.abImage, pReq);
4349	else
4350	{
4351	memcpy(pImage->pvImage, &pReq->u.In.abImage[0], pImage->cbImageBits);
4352	Log(("vboxdrv: Loaded '%s' at %p\n", pImage->szName, pImage->pvImage));
4353	}
4354	}
4355
4356	/*
4357	* Update any entry points.
4358	*/
4359	if (RT_SUCCESS(rc))
4360	{
4361	switch (pReq->u.In.eEPType)
4362	{
4363	default:
4364	case SUPLDRLOADEP_NOTHING:
4365	rc = VINF_SUCCESS;
4366	break;
4367	case SUPLDRLOADEP_VMMR0:
4368	rc = supdrvLdrSetVMMR0EPs(pDevExt, pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
4369	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
4370	break;
4371	case SUPLDRLOADEP_SERVICE:
4372	pImage->pfnServiceReqHandler = pReq->u.In.EP.Service.pfnServiceReq;
4373	rc = VINF_SUCCESS;
4374	break;
4375	}
4376	}
4377
4378	/*
4379	* On success call the module initialization.
4380	*/
4381	LogFlow(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
4382	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
4383	{
4384	Log(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
4385	pDevExt->pLdrInitImage = pImage;
4386	pDevExt->hLdrInitThread = RTThreadNativeSelf();
4387	rc = pImage->pfnModuleInit(pImage);
4388	pDevExt->pLdrInitImage = NULL;
4389	pDevExt->hLdrInitThread = NIL_RTNATIVETHREAD;
4390	if (RT_FAILURE(rc) && pDevExt->pvVMMR0 == pImage->pvImage)
4391	supdrvLdrUnsetVMMR0EPs(pDevExt);
4392	}
4393
4394	if (RT_FAILURE(rc))
4395	{
4396	/* Inform the tracing component in case ModuleInit registered TPs. */
4397	supdrvTracerModuleUnloading(pDevExt, pImage);
4398
4399	pImage->uState = SUP_IOCTL_LDR_OPEN;
4400	pImage->pfnModuleInit = NULL;
4401	pImage->pfnModuleTerm = NULL;
4402	pImage->pfnServiceReqHandler= NULL;
4403	pImage->cbStrTab = 0;
4404	RTMemFree(pImage->pachStrTab);
4405	pImage->pachStrTab = NULL;
4406	RTMemFree(pImage->paSymbols);
4407	pImage->paSymbols = NULL;
4408	pImage->cSymbols = 0;
4409	}
4410
4411	supdrvLdrUnlock(pDevExt);
4412	return rc;
4413	}
4414
4415
4416	/**
4417	* Frees a previously loaded (prep'ed) image.
4418	*
4419	* @returns IPRT status code.
4420	* @param pDevExt Device globals.
4421	* @param pSession Session data.
4422	* @param pReq The request.
4423	*/
4424	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
4425	{
4426	int rc;
4427	PSUPDRVLDRUSAGE pUsagePrev;
4428	PSUPDRVLDRUSAGE pUsage;
4429	PSUPDRVLDRIMAGE pImage;
4430	LogFlow(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
4431
4432	/*
4433	* Find the ldr image.
4434	*/
4435	supdrvLdrLock(pDevExt);
4436	pUsagePrev = NULL;
4437	pUsage = pSession->pLdrUsage;
4438	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4439	{
4440	pUsagePrev = pUsage;
4441	pUsage = pUsage->pNext;
4442	}
4443	if (!pUsage)
4444	{
4445	supdrvLdrUnlock(pDevExt);
4446	Log(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
4447	return VERR_INVALID_HANDLE;
4448	}
4449
4450	/*
4451	* Check if we can remove anything.
4452	*/
4453	rc = VINF_SUCCESS;
4454	pImage = pUsage->pImage;
4455	if (pImage->cUsage <= 1 \|\| pUsage->cUsage <= 1)
4456	{
4457	/*
4458	* Check if there are any objects with destructors in the image, if
4459	* so leave it for the session cleanup routine so we get a chance to
4460	* clean things up in the right order and not leave them all dangling.
4461	*/
4462	RTSpinlockAcquire(pDevExt->Spinlock);
4463	if (pImage->cUsage <= 1)
4464	{
4465	PSUPDRVOBJ pObj;
4466	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4467	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4468	{
4469	rc = VERR_DANGLING_OBJECTS;
4470	break;
4471	}
4472	}
4473	else
4474	{
4475	PSUPDRVUSAGE pGenUsage;
4476	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
4477	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4478	{
4479	rc = VERR_DANGLING_OBJECTS;
4480	break;
4481	}
4482	}
4483	RTSpinlockRelease(pDevExt->Spinlock);
4484	if (rc == VINF_SUCCESS)
4485	{
4486	/* unlink it */
4487	if (pUsagePrev)
4488	pUsagePrev->pNext = pUsage->pNext;
4489	else
4490	pSession->pLdrUsage = pUsage->pNext;
4491
4492	/* free it */
4493	pUsage->pImage = NULL;
4494	pUsage->pNext = NULL;
4495	RTMemFree(pUsage);
4496
4497	/*
4498	* Dereference the image.
4499	*/
4500	if (pImage->cUsage <= 1)
4501	supdrvLdrFree(pDevExt, pImage);
4502	else
4503	pImage->cUsage--;
4504	}
4505	else
4506	{
4507	Log(("supdrvIOCtl_LdrFree: Dangling objects in %p/%s!\n", pImage->pvImage, pImage->szName));
4508	rc = VINF_SUCCESS; /** @todo BRANCH-2.1: remove this after branching. */
4509	}
4510	}
4511	else
4512	{
4513	/*
4514	* Dereference both image and usage.
4515	*/
4516	pImage->cUsage--;
4517	pUsage->cUsage--;
4518	}
4519
4520	supdrvLdrUnlock(pDevExt);
4521	return rc;
4522	}
4523
4524
4525	/**
4526	* Gets the address of a symbol in an open image.
4527	*
4528	* @returns IPRT status code.
4529	* @param pDevExt Device globals.
4530	* @param pSession Session data.
4531	* @param pReq The request buffer.
4532	*/
4533	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
4534	{
4535	PSUPDRVLDRIMAGE pImage;
4536	PSUPDRVLDRUSAGE pUsage;
4537	uint32_t i;
4538	PSUPLDRSYM paSyms;
4539	const char *pchStrings;
4540	const size_t cbSymbol = strlen(pReq->u.In.szSymbol) + 1;
4541	void *pvSymbol = NULL;
4542	int rc = VERR_GENERAL_FAILURE;
4543	Log3(("supdrvIOCtl_LdrGetSymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
4544
4545	/*
4546	* Find the ldr image.
4547	*/
4548	supdrvLdrLock(pDevExt);
4549	pUsage = pSession->pLdrUsage;
4550	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4551	pUsage = pUsage->pNext;
4552	if (!pUsage)
4553	{
4554	supdrvLdrUnlock(pDevExt);
4555	Log(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
4556	return VERR_INVALID_HANDLE;
4557	}
4558	pImage = pUsage->pImage;
4559	if (pImage->uState != SUP_IOCTL_LDR_LOAD)
4560	{
4561	unsigned uState = pImage->uState;
4562	supdrvLdrUnlock(pDevExt);
4563	Log(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", uState, uState)); NOREF(uState);
4564	return VERR_ALREADY_LOADED;
4565	}
4566
4567	/*
4568	* Search the symbol strings.
4569	*
4570	* Note! The int32_t is for native loading on solaris where the data
4571	* and text segments are in very different places.
4572	*/
4573	pchStrings = pImage->pachStrTab;
4574	paSyms = pImage->paSymbols;
4575	for (i = 0; i < pImage->cSymbols; i++)
4576	{
4577	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4578	&& !memcmp(pchStrings + paSyms[i].offName, pReq->u.In.szSymbol, cbSymbol))
4579	{
4580	pvSymbol = (uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol;
4581	rc = VINF_SUCCESS;
4582	break;
4583	}
4584	}
4585	supdrvLdrUnlock(pDevExt);
4586	pReq->u.Out.pvSymbol = pvSymbol;
4587	return rc;
4588	}
4589
4590
4591	/**
4592	* Gets the address of a symbol in an open image or the support driver.
4593	*
4594	* @returns VINF_SUCCESS on success.
4595	* @returns
4596	* @param pDevExt Device globals.
4597	* @param pSession Session data.
4598	* @param pReq The request buffer.
4599	*/
4600	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq)
4601	{
4602	int rc = VINF_SUCCESS;
4603	const char *pszSymbol = pReq->u.In.pszSymbol;
4604	const char *pszModule = pReq->u.In.pszModule;
4605	size_t cbSymbol;
4606	char const *pszEnd;
4607	uint32_t i;
4608
4609	/*
4610	* Input validation.
4611	*/
4612	AssertPtrReturn(pszSymbol, VERR_INVALID_POINTER);
4613	pszEnd = RTStrEnd(pszSymbol, 512);
4614	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4615	cbSymbol = pszEnd - pszSymbol + 1;
4616
4617	if (pszModule)
4618	{
4619	AssertPtrReturn(pszModule, VERR_INVALID_POINTER);
4620	pszEnd = RTStrEnd(pszModule, 64);
4621	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4622	}
4623	Log3(("supdrvIDC_LdrGetSymbol: pszModule=%p:{%s} pszSymbol=%p:{%s}\n", pszModule, pszModule, pszSymbol, pszSymbol));
4624
4625
4626	if ( !pszModule
4627	\|\| !strcmp(pszModule, "SupDrv"))
4628	{
4629	/*
4630	* Search the support driver export table.
4631	*/
4632	for (i = 0; i < RT_ELEMENTS(g_aFunctions); i++)
4633	if (!strcmp(g_aFunctions[i].szName, pszSymbol))
4634	{
4635	pReq->u.Out.pfnSymbol = g_aFunctions[i].pfn;
4636	break;
4637	}
4638	}
4639	else
4640	{
4641	/*
4642	* Find the loader image.
4643	*/
4644	PSUPDRVLDRIMAGE pImage;
4645
4646	supdrvLdrLock(pDevExt);
4647
4648	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
4649	if (!strcmp(pImage->szName, pszModule))
4650	break;
4651	if (pImage && pImage->uState == SUP_IOCTL_LDR_LOAD)
4652	{
4653	/*
4654	* Search the symbol strings.
4655	*/
4656	const char *pchStrings = pImage->pachStrTab;
4657	PCSUPLDRSYM paSyms = pImage->paSymbols;
4658	for (i = 0; i < pImage->cSymbols; i++)
4659	{
4660	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4661	&& !memcmp(pchStrings + paSyms[i].offName, pszSymbol, cbSymbol))
4662	{
4663	/*
4664	* Found it! Calc the symbol address and add a reference to the module.
4665	*/
4666	pReq->u.Out.pfnSymbol = (PFNRT)((uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol);
4667	rc = supdrvLdrAddUsage(pSession, pImage);
4668	break;
4669	}
4670	}
4671	}
4672	else
4673	rc = pImage ? VERR_WRONG_ORDER : VERR_MODULE_NOT_FOUND;
4674
4675	supdrvLdrUnlock(pDevExt);
4676	}
4677	return rc;
4678	}
4679
4680
4681	/**
4682	* Updates the VMMR0 entry point pointers.
4683	*
4684	* @returns IPRT status code.
4685	* @param pDevExt Device globals.
4686	* @param pSession Session data.
4687	* @param pVMMR0 VMMR0 image handle.
4688	* @param pvVMMR0EntryInt VMMR0EntryInt address.
4689	* @param pvVMMR0EntryFast VMMR0EntryFast address.
4690	* @param pvVMMR0EntryEx VMMR0EntryEx address.
4691	* @remark Caller must own the loader mutex.
4692	*/
4693	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx)
4694	{
4695	int rc = VINF_SUCCESS;
4696	LogFlow(("supdrvLdrSetR0EP pvVMMR0=%p pvVMMR0EntryInt=%p\n", pvVMMR0, pvVMMR0EntryInt));
4697
4698
4699	/*
4700	* Check if not yet set.
4701	*/
4702	if (!pDevExt->pvVMMR0)
4703	{
4704	pDevExt->pvVMMR0 = pvVMMR0;
4705	pDevExt->pfnVMMR0EntryInt = pvVMMR0EntryInt;
4706	pDevExt->pfnVMMR0EntryFast = pvVMMR0EntryFast;
4707	pDevExt->pfnVMMR0EntryEx = pvVMMR0EntryEx;
4708	}
4709	else
4710	{
4711	/*
4712	* Return failure or success depending on whether the values match or not.
4713	*/
4714	if ( pDevExt->pvVMMR0 != pvVMMR0
4715	\|\| (void *)pDevExt->pfnVMMR0EntryInt != pvVMMR0EntryInt
4716	\|\| (void *)pDevExt->pfnVMMR0EntryFast != pvVMMR0EntryFast
4717	\|\| (void *)pDevExt->pfnVMMR0EntryEx != pvVMMR0EntryEx)
4718	{
4719	AssertMsgFailed(("SUP_IOCTL_LDR_SETR0EP: Already set pointing to a different module!\n"));
4720	rc = VERR_INVALID_PARAMETER;
4721	}
4722	}
4723	return rc;
4724	}
4725
4726
4727	/**
4728	* Unsets the VMMR0 entry point installed by supdrvLdrSetR0EP.
4729	*
4730	* @param pDevExt Device globals.
4731	*/
4732	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt)
4733	{
4734	pDevExt->pvVMMR0 = NULL;
4735	pDevExt->pfnVMMR0EntryInt = NULL;
4736	pDevExt->pfnVMMR0EntryFast = NULL;
4737	pDevExt->pfnVMMR0EntryEx = NULL;
4738	}
4739
4740
4741	/**
4742	* Adds a usage reference in the specified session of an image.
4743	*
4744	* Called while owning the loader semaphore.
4745	*
4746	* @returns VINF_SUCCESS on success and VERR_NO_MEMORY on failure.
4747	* @param pSession Session in question.
4748	* @param pImage Image which the session is using.
4749	*/
4750	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage)
4751	{
4752	PSUPDRVLDRUSAGE pUsage;
4753	LogFlow(("supdrvLdrAddUsage: pImage=%p\n", pImage));
4754
4755	/*
4756	* Referenced it already?
4757	*/
4758	pUsage = pSession->pLdrUsage;
4759	while (pUsage)
4760	{
4761	if (pUsage->pImage == pImage)
4762	{
4763	pUsage->cUsage++;
4764	return VINF_SUCCESS;
4765	}
4766	pUsage = pUsage->pNext;
4767	}
4768
4769	/*
4770	* Allocate new usage record.
4771	*/
4772	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
4773	AssertReturn(pUsage, /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_5);
4774	pUsage->cUsage = 1;
4775	pUsage->pImage = pImage;
4776	pUsage->pNext = pSession->pLdrUsage;
4777	pSession->pLdrUsage = pUsage;
4778	return VINF_SUCCESS;
4779	}
4780
4781
4782	/**
4783	* Frees a load image.
4784	*
4785	* @param pDevExt Pointer to device extension.
4786	* @param pImage Pointer to the image we're gonna free.
4787	* This image must exit!
4788	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
4789	*/
4790	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
4791	{
4792	PSUPDRVLDRIMAGE pImagePrev;
4793	LogFlow(("supdrvLdrFree: pImage=%p\n", pImage));
4794
4795	/* find it - arg. should've used doubly linked list. */
4796	Assert(pDevExt->pLdrImages);
4797	pImagePrev = NULL;
4798	if (pDevExt->pLdrImages != pImage)
4799	{
4800	pImagePrev = pDevExt->pLdrImages;
4801	while (pImagePrev->pNext != pImage)
4802	pImagePrev = pImagePrev->pNext;
4803	Assert(pImagePrev->pNext == pImage);
4804	}
4805
4806	/* unlink */
4807	if (pImagePrev)
4808	pImagePrev->pNext = pImage->pNext;
4809	else
4810	pDevExt->pLdrImages = pImage->pNext;
4811
4812	/* check if this is VMMR0.r0 unset its entry point pointers. */
4813	if (pDevExt->pvVMMR0 == pImage->pvImage)
4814	supdrvLdrUnsetVMMR0EPs(pDevExt);
4815
4816	/* check for objects with destructors in this image. (Shouldn't happen.) */
4817	if (pDevExt->pObjs)
4818	{
4819	unsigned cObjs = 0;
4820	PSUPDRVOBJ pObj;
4821	RTSpinlockAcquire(pDevExt->Spinlock);
4822	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4823	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4824	{
4825	pObj->pfnDestructor = NULL;
4826	cObjs++;
4827	}
4828	RTSpinlockRelease(pDevExt->Spinlock);
4829	if (cObjs)
4830	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
4831	}
4832
4833	/* call termination function if fully loaded. */
4834	if ( pImage->pfnModuleTerm
4835	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
4836	{
4837	LogFlow(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
4838	pImage->pfnModuleTerm(pImage);
4839	}
4840
4841	/* Inform the tracing component. */
4842	supdrvTracerModuleUnloading(pDevExt, pImage);
4843
4844	/* do native unload if appropriate. */
4845	if (pImage->fNative)
4846	supdrvOSLdrUnload(pDevExt, pImage);
4847
4848	/* free the image */
4849	pImage->cUsage = 0;
4850	pImage->pDevExt = NULL;
4851	pImage->pNext = NULL;
4852	pImage->uState = SUP_IOCTL_LDR_FREE;
4853	RTMemExecFree(pImage->pvImageAlloc, pImage->cbImageBits + 31);
4854	pImage->pvImageAlloc = NULL;
4855	RTMemFree(pImage->pachStrTab);
4856	pImage->pachStrTab = NULL;
4857	RTMemFree(pImage->paSymbols);
4858	pImage->paSymbols = NULL;
4859	RTMemFree(pImage);
4860	}
4861
4862
4863	/**
4864	* Acquires the loader lock.
4865	*
4866	* @returns IPRT status code.
4867	* @param pDevExt The device extension.
4868	*/
4869	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt)
4870	{
4871	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4872	int rc = RTSemMutexRequest(pDevExt->mtxLdr, RT_INDEFINITE_WAIT);
4873	#else
4874	int rc = RTSemFastMutexRequest(pDevExt->mtxLdr);
4875	#endif
4876	AssertRC(rc);
4877	return rc;
4878	}
4879
4880
4881	/**
4882	* Releases the loader lock.
4883	*
4884	* @returns IPRT status code.
4885	* @param pDevExt The device extension.
4886	*/
4887	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt)
4888	{
4889	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4890	return RTSemMutexRelease(pDevExt->mtxLdr);
4891	#else
4892	return RTSemFastMutexRelease(pDevExt->mtxLdr);
4893	#endif
4894	}
4895
4896
4897	/**
4898	* Implements the service call request.
4899	*
4900	* @returns VBox status code.
4901	* @param pDevExt The device extension.
4902	* @param pSession The calling session.
4903	* @param pReq The request packet, valid.
4904	*/
4905	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq)
4906	{
4907	#if !defined(RT_OS_WINDOWS) \|\| defined(RT_ARCH_AMD64) \|\| defined(DEBUG)
4908	int rc;
4909
4910	/*
4911	* Find the module first in the module referenced by the calling session.
4912	*/
4913	rc = supdrvLdrLock(pDevExt);
4914	if (RT_SUCCESS(rc))
4915	{
4916	PFNSUPR0SERVICEREQHANDLER pfnServiceReqHandler = NULL;
4917	PSUPDRVLDRUSAGE pUsage;
4918
4919	for (pUsage = pSession->pLdrUsage; pUsage; pUsage = pUsage->pNext)
4920	if ( pUsage->pImage->pfnServiceReqHandler
4921	&& !strcmp(pUsage->pImage->szName, pReq->u.In.szName))
4922	{
4923	pfnServiceReqHandler = pUsage->pImage->pfnServiceReqHandler;
4924	break;
4925	}
4926	supdrvLdrUnlock(pDevExt);
4927
4928	if (pfnServiceReqHandler)
4929	{
4930	/*
4931	* Call it.
4932	*/
4933	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
4934	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, NULL);
4935	else
4936	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0]);
4937	}
4938	else
4939	rc = VERR_SUPDRV_SERVICE_NOT_FOUND;
4940	}
4941
4942	/* log it */
4943	if ( RT_FAILURE(rc)
4944	&& rc != VERR_INTERRUPTED
4945	&& rc != VERR_TIMEOUT)
4946	Log(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4947	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4948	else
4949	Log4(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4950	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4951	return rc;
4952	#else /* RT_OS_WINDOWS && !RT_ARCH_AMD64 && !DEBUG */
4953	return VERR_NOT_IMPLEMENTED;
4954	#endif /* RT_OS_WINDOWS && !RT_ARCH_AMD64 && !DEBUG */
4955	}
4956
4957
4958	/**
4959	* Implements the logger settings request.
4960	*
4961	* @returns VBox status code.
4962	* @param pDevExt The device extension.
4963	* @param pSession The caller's session.
4964	* @param pReq The request.
4965	*/
4966	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq)
4967	{
4968	const char *pszGroup = &pReq->u.In.szStrings[pReq->u.In.offGroups];
4969	const char *pszFlags = &pReq->u.In.szStrings[pReq->u.In.offFlags];
4970	const char *pszDest = &pReq->u.In.szStrings[pReq->u.In.offDestination];
4971	PRTLOGGER pLogger = NULL;
4972	int rc;
4973
4974	/*
4975	* Some further validation.
4976	*/
4977	switch (pReq->u.In.fWhat)
4978	{
4979	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4980	case SUPLOGGERSETTINGS_WHAT_CREATE:
4981	break;
4982
4983	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4984	if (pszGroup \|\| pszFlags \|\| *pszDest)
4985	return VERR_INVALID_PARAMETER;
4986	if (pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_RELEASE)
4987	return VERR_ACCESS_DENIED;
4988	break;
4989
4990	default:
4991	return VERR_INTERNAL_ERROR;
4992	}
4993
4994	/*
4995	* Get the logger.
4996	*/
4997	switch (pReq->u.In.fWhich)
4998	{
4999	case SUPLOGGERSETTINGS_WHICH_DEBUG:
5000	pLogger = RTLogGetDefaultInstance();
5001	break;
5002
5003	case SUPLOGGERSETTINGS_WHICH_RELEASE:
5004	pLogger = RTLogRelDefaultInstance();
5005	break;
5006
5007	default:
5008	return VERR_INTERNAL_ERROR;
5009	}
5010
5011	/*
5012	* Do the job.
5013	*/
5014	switch (pReq->u.In.fWhat)
5015	{
5016	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
5017	if (pLogger)
5018	{
5019	rc = RTLogFlags(pLogger, pszFlags);
5020	if (RT_SUCCESS(rc))
5021	rc = RTLogGroupSettings(pLogger, pszGroup);
5022	NOREF(pszDest);
5023	}
5024	else
5025	rc = VERR_NOT_FOUND;
5026	break;
5027
5028	case SUPLOGGERSETTINGS_WHAT_CREATE:
5029	{
5030	if (pLogger)
5031	rc = VERR_ALREADY_EXISTS;
5032	else
5033	{
5034	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
5035
5036	rc = RTLogCreate(&pLogger,
5037	0 /* fFlags */,
5038	pszGroup,
5039	pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_DEBUG
5040	? "VBOX_LOG"
5041	: "VBOX_RELEASE_LOG",
5042	RT_ELEMENTS(s_apszGroups),
5043	s_apszGroups,
5044	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER,
5045	NULL);
5046	if (RT_SUCCESS(rc))
5047	{
5048	rc = RTLogFlags(pLogger, pszFlags);
5049	NOREF(pszDest);
5050	if (RT_SUCCESS(rc))
5051	{
5052	switch (pReq->u.In.fWhich)
5053	{
5054	case SUPLOGGERSETTINGS_WHICH_DEBUG:
5055	pLogger = RTLogSetDefaultInstance(pLogger);
5056	break;
5057	case SUPLOGGERSETTINGS_WHICH_RELEASE:
5058	pLogger = RTLogRelSetDefaultInstance(pLogger);
5059	break;
5060	}
5061	}
5062	RTLogDestroy(pLogger);
5063	}
5064	}
5065	break;
5066	}
5067
5068	case SUPLOGGERSETTINGS_WHAT_DESTROY:
5069	switch (pReq->u.In.fWhich)
5070	{
5071	case SUPLOGGERSETTINGS_WHICH_DEBUG:
5072	pLogger = RTLogSetDefaultInstance(NULL);
5073	break;
5074	case SUPLOGGERSETTINGS_WHICH_RELEASE:
5075	pLogger = RTLogRelSetDefaultInstance(NULL);
5076	break;
5077	}
5078	rc = RTLogDestroy(pLogger);
5079	break;
5080
5081	default:
5082	{
5083	rc = VERR_INTERNAL_ERROR;
5084	break;
5085	}
5086	}
5087
5088	return rc;
5089	}
5090
5091
5092	/**
5093	* Creates the GIP.
5094	*
5095	* @returns VBox status code.
5096	* @param pDevExt Instance data. GIP stuff may be updated.
5097	*/
5098	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
5099	{
5100	PSUPGLOBALINFOPAGE pGip;
5101	RTHCPHYS HCPhysGip;
5102	uint32_t u32SystemResolution;
5103	uint32_t u32Interval;
5104	unsigned cCpus;
5105	int rc;
5106
5107
5108	LogFlow(("supdrvGipCreate:\n"));
5109
5110	/* assert order */
5111	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
5112	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
5113	Assert(!pDevExt->pGipTimer);
5114
5115	/*
5116	* Check the CPU count.
5117	*/
5118	cCpus = RTMpGetArraySize();
5119	if ( cCpus > RTCPUSET_MAX_CPUS
5120	\|\| cCpus > 256 /ApicId is used for the mappings/)
5121	{
5122	SUPR0Printf("VBoxDrv: Too many CPUs (%u) for the GIP (max %u)\n", cCpus, RT_MIN(RTCPUSET_MAX_CPUS, 256));
5123	return VERR_TOO_MANY_CPUS;
5124	}
5125
5126	/*
5127	* Allocate a contiguous set of pages with a default kernel mapping.
5128	*/
5129	rc = RTR0MemObjAllocCont(&pDevExt->GipMemObj, RT_UOFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), false /fExecutable/);
5130	if (RT_FAILURE(rc))
5131	{
5132	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
5133	return rc;
5134	}
5135	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
5136	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
5137
5138	/*
5139	* Find a reasonable update interval and initialize the structure.
5140	*/
5141	u32Interval = u32SystemResolution = RTTimerGetSystemGranularity();
5142	while (u32Interval < 10000000 /* 10 ms */)
5143	u32Interval += u32SystemResolution;
5144
5145	supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), 1000000000 / u32Interval /=Hz/, cCpus);
5146
5147	/*
5148	* Create the timer.
5149	* If CPU_ALL isn't supported we'll have to fall back to synchronous mode.
5150	*/
5151	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
5152	{
5153	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, RTTIMER_FLAGS_CPU_ALL, supdrvGipAsyncTimer, pDevExt);
5154	if (rc == VERR_NOT_SUPPORTED)
5155	{
5156	OSDBGPRINT(("supdrvGipCreate: omni timer not supported, falling back to synchronous mode\n"));
5157	pGip->u32Mode = SUPGIPMODE_SYNC_TSC;
5158	}
5159	}
5160	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
5161	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0, supdrvGipSyncTimer, pDevExt);
5162	if (RT_SUCCESS(rc))
5163	{
5164	rc = RTMpNotificationRegister(supdrvGipMpEvent, pDevExt);
5165	if (RT_SUCCESS(rc))
5166	{
5167	rc = RTMpOnAll(supdrvGipInitOnCpu, pDevExt, pGip);
5168	if (RT_SUCCESS(rc))
5169	{
5170	/*
5171	* We're good.
5172	*/
5173	Log(("supdrvGipCreate: %u ns interval.\n", u32Interval));
5174	g_pSUPGlobalInfoPage = pGip;
5175	return VINF_SUCCESS;
5176	}
5177
5178	OSDBGPRINT(("supdrvGipCreate: RTMpOnAll failed with rc=%Rrc\n", rc));
5179	RTMpNotificationDeregister(supdrvGipMpEvent, pDevExt);
5180
5181	}
5182	else
5183	OSDBGPRINT(("supdrvGipCreate: failed to register MP event notfication. rc=%Rrc\n", rc));
5184	}
5185	else
5186	{
5187	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %u ns interval. rc=%Rrc\n", u32Interval, rc));
5188	Assert(!pDevExt->pGipTimer);
5189	}
5190	supdrvGipDestroy(pDevExt);
5191	return rc;
5192	}
5193
5194
5195	/**
5196	* Terminates the GIP.
5197	*
5198	* @param pDevExt Instance data. GIP stuff may be updated.
5199	*/
5200	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
5201	{
5202	int rc;
5203	#ifdef DEBUG_DARWIN_GIP
5204	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
5205	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
5206	pDevExt->pGipTimer, pDevExt->GipMemObj));
5207	#endif
5208
5209	/*
5210	* Invalid the GIP data.
5211	*/
5212	if (pDevExt->pGip)
5213	{
5214	supdrvGipTerm(pDevExt->pGip);
5215	pDevExt->pGip = NULL;
5216	}
5217	g_pSUPGlobalInfoPage = NULL;
5218
5219	/*
5220	* Destroy the timer and free the GIP memory object.
5221	*/
5222	if (pDevExt->pGipTimer)
5223	{
5224	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
5225	pDevExt->pGipTimer = NULL;
5226	}
5227
5228	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
5229	{
5230	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
5231	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
5232	}
5233
5234	/*
5235	* Finally, make sure we've release the system timer resolution request
5236	* if one actually succeeded and is still pending.
5237	*/
5238	if (pDevExt->u32SystemTimerGranularityGrant)
5239	{
5240	rc = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
5241	pDevExt->u32SystemTimerGranularityGrant = 0;
5242	}
5243	}
5244
5245
5246	/**
5247	* Timer callback function sync GIP mode.
5248	* @param pTimer The timer.
5249	* @param pvUser The device extension.
5250	*/
5251	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
5252	{
5253	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
5254	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5255	uint64_t u64TSC = ASMReadTSC();
5256	uint64_t NanoTS = RTTimeSystemNanoTS();
5257
5258	supdrvGipUpdate(pDevExt, NanoTS, u64TSC, NIL_RTCPUID, iTick);
5259
5260	ASMSetFlags(fOldFlags);
5261	}
5262
5263
5264	/**
5265	* Timer callback function for async GIP mode.
5266	* @param pTimer The timer.
5267	* @param pvUser The device extension.
5268	*/
5269	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
5270	{
5271	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
5272	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5273	RTCPUID idCpu = RTMpCpuId();
5274	uint64_t u64TSC = ASMReadTSC();
5275	uint64_t NanoTS = RTTimeSystemNanoTS();
5276
5277	/** @todo reset the transaction number and whatnot when iTick == 1. */
5278	if (pDevExt->idGipMaster == idCpu)
5279	supdrvGipUpdate(pDevExt, NanoTS, u64TSC, idCpu, iTick);
5280	else
5281	supdrvGipUpdatePerCpu(pDevExt, NanoTS, u64TSC, idCpu, ASMGetApicId(), iTick);
5282
5283	ASMSetFlags(fOldFlags);
5284	}
5285
5286
5287	/**
5288	* Finds our (@a idCpu) entry, or allocates a new one if not found.
5289	*
5290	* @returns Index of the CPU in the cache set.
5291	* @param pGip The GIP.
5292	* @param idCpu The CPU ID.
5293	*/
5294	static uint32_t supdrvGipCpuIndexFromCpuId(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
5295	{
5296	uint32_t i, cTries;
5297
5298	/*
5299	* ASSUMES that CPU IDs are constant.
5300	*/
5301	for (i = 0; i < pGip->cCpus; i++)
5302	if (pGip->aCPUs[i].idCpu == idCpu)
5303	return i;
5304
5305	cTries = 0;
5306	do
5307	{
5308	for (i = 0; i < pGip->cCpus; i++)
5309	{
5310	bool fRc;
5311	ASMAtomicCmpXchgSize(&pGip->aCPUs[i].idCpu, idCpu, NIL_RTCPUID, fRc);
5312	if (fRc)
5313	return i;
5314	}
5315	} while (cTries++ < 32);
5316	AssertReleaseFailed();
5317	return i - 1;
5318	}
5319
5320
5321	/**
5322	* The calling CPU should be accounted as online, update GIP accordingly.
5323	*
5324	* This is used by supdrvGipMpEvent as well as the supdrvGipCreate.
5325	*
5326	* @param pDevExt The device extension.
5327	* @param idCpu The CPU ID.
5328	*/
5329	static void supdrvGipMpEventOnline(PSUPDRVDEVEXT pDevExt, RTCPUID idCpu)
5330	{
5331	int iCpuSet = 0;
5332	uint16_t idApic = UINT16_MAX;
5333	uint32_t i = 0;
5334	uint64_t u64NanoTS = 0;
5335	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5336
5337	AssertPtrReturnVoid(pGip);
5338	AssertRelease(idCpu == RTMpCpuId());
5339	Assert(pGip->cPossibleCpus == RTMpGetCount());
5340
5341	/*
5342	* Do this behind a spinlock with interrupts disabled as this can fire
5343	* on all CPUs simultaneously, see @bugref{6110}.
5344	*/
5345	RTSpinlockAcquire(pDevExt->hGipSpinlock);
5346
5347	/*
5348	* Update the globals.
5349	*/
5350	ASMAtomicWriteU16(&pGip->cPresentCpus, RTMpGetPresentCount());
5351	ASMAtomicWriteU16(&pGip->cOnlineCpus, RTMpGetOnlineCount());
5352	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5353	if (iCpuSet >= 0)
5354	{
5355	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5356	RTCpuSetAddByIndex(&pGip->OnlineCpuSet, iCpuSet);
5357	RTCpuSetAddByIndex(&pGip->PresentCpuSet, iCpuSet);
5358	}
5359
5360	/*
5361	* Update the entry.
5362	*/
5363	u64NanoTS = RTTimeSystemNanoTS() - pGip->u32UpdateIntervalNS;
5364	i = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
5365	supdrvGipInitCpu(pGip, &pGip->aCPUs[i], u64NanoTS);
5366	idApic = ASMGetApicId();
5367	ASMAtomicWriteU16(&pGip->aCPUs[i].idApic, idApic);
5368	ASMAtomicWriteS16(&pGip->aCPUs[i].iCpuSet, (int16_t)iCpuSet);
5369	ASMAtomicWriteSize(&pGip->aCPUs[i].idCpu, idCpu);
5370
5371	/*
5372	* Update the APIC ID and CPU set index mappings.
5373	*/
5374	ASMAtomicWriteU16(&pGip->aiCpuFromApicId[idApic], i);
5375	ASMAtomicWriteU16(&pGip->aiCpuFromCpuSetIdx[iCpuSet], i);
5376
5377	/* commit it */
5378	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_ONLINE);
5379
5380	RTSpinlockReleaseNoInts(pDevExt->hGipSpinlock);
5381	}
5382
5383
5384	/**
5385	* The CPU should be accounted as offline, update the GIP accordingly.
5386	*
5387	* This is used by supdrvGipMpEvent.
5388	*
5389	* @param pDevExt The device extension.
5390	* @param idCpu The CPU ID.
5391	*/
5392	static void supdrvGipMpEventOffline(PSUPDRVDEVEXT pDevExt, RTCPUID idCpu)
5393	{
5394	int iCpuSet;
5395	unsigned i;
5396
5397	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5398
5399	AssertPtrReturnVoid(pGip);
5400	RTSpinlockAcquire(pDevExt->hGipSpinlock);
5401
5402	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5403	AssertReturnVoid(iCpuSet >= 0);
5404
5405	i = pGip->aiCpuFromCpuSetIdx[iCpuSet];
5406	AssertReturnVoid(i < pGip->cCpus);
5407	AssertReturnVoid(pGip->aCPUs[i].idCpu == idCpu);
5408
5409	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5410	RTCpuSetDelByIndex(&pGip->OnlineCpuSet, iCpuSet);
5411
5412	/* commit it */
5413	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_OFFLINE);
5414
5415	RTSpinlockReleaseNoInts(pDevExt->hGipSpinlock);
5416	}
5417
5418
5419	/**
5420	* Multiprocessor event notification callback.
5421	*
5422	* This is used to make sure that the GIP master gets passed on to
5423	* another CPU. It also updates the associated CPU data.
5424	*
5425	* @param enmEvent The event.
5426	* @param idCpu The cpu it applies to.
5427	* @param pvUser Pointer to the device extension.
5428	*
5429	* @remarks This function -must- fire on the newly online'd CPU for the
5430	* RTMPEVENT_ONLINE case and can fire on any CPU for the
5431	* RTMPEVENT_OFFLINE case.
5432	*/
5433	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser)
5434	{
5435	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5436	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5437
5438	AssertRelease(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
5439
5440	/*
5441	* Update the GIP CPU data.
5442	*/
5443	if (pGip)
5444	{
5445	switch (enmEvent)
5446	{
5447	case RTMPEVENT_ONLINE:
5448	AssertRelease(idCpu == RTMpCpuId());
5449	supdrvGipMpEventOnline(pDevExt, idCpu);
5450	break;
5451	case RTMPEVENT_OFFLINE:
5452	supdrvGipMpEventOffline(pDevExt, idCpu);
5453	break;
5454
5455	}
5456	}
5457
5458	/*
5459	* Make sure there is a master GIP.
5460	*/
5461	if (enmEvent == RTMPEVENT_OFFLINE)
5462	{
5463	RTCPUID idGipMaster = ASMAtomicReadU32(&pDevExt->idGipMaster);
5464	if (idGipMaster == idCpu)
5465	{
5466	/*
5467	* Find a new GIP master.
5468	*/
5469	bool fIgnored;
5470	unsigned i;
5471	RTCPUID idNewGipMaster = NIL_RTCPUID;
5472	RTCPUSET OnlineCpus;
5473	RTMpGetOnlineSet(&OnlineCpus);
5474
5475	for (i = 0; i < RTCPUSET_MAX_CPUS; i++)
5476	{
5477	RTCPUID idCurCpu = RTMpCpuIdFromSetIndex(i);
5478	if ( RTCpuSetIsMember(&OnlineCpus, idCurCpu)
5479	&& idCurCpu != idGipMaster)
5480	{
5481	idNewGipMaster = idCurCpu;
5482	break;
5483	}
5484	}
5485
5486	Log(("supdrvGipMpEvent: Gip master %#lx -> %#lx\n", (long)idGipMaster, (long)idNewGipMaster));
5487	ASMAtomicCmpXchgSize(&pDevExt->idGipMaster, idNewGipMaster, idGipMaster, fIgnored);
5488	NOREF(fIgnored);
5489	}
5490	}
5491	}
5492
5493
5494	/**
5495	* Callback used by supdrvDetermineAsyncTSC to read the TSC on a CPU.
5496	*
5497	* @param idCpu Ignored.
5498	* @param pvUser1 Where to put the TSC.
5499	* @param pvUser2 Ignored.
5500	*/
5501	static DECLCALLBACK(void) supdrvDetermineAsyncTscWorker(RTCPUID idCpu, void pvUser1, void pvUser2)
5502	{
5503	#if 1
5504	ASMAtomicWriteU64((uint64_t volatile *)pvUser1, ASMReadTSC());
5505	#else
5506	(uint64_t )pvUser1 = ASMReadTSC();
5507	#endif
5508	}
5509
5510
5511	/**
5512	* Determine if Async GIP mode is required because of TSC drift.
5513	*
5514	* When using the default/normal timer code it is essential that the time stamp counter
5515	* (TSC) runs never backwards, that is, a read operation to the counter should return
5516	* a bigger value than any previous read operation. This is guaranteed by the latest
5517	* AMD CPUs and by newer Intel CPUs which never enter the C2 state (P4). In any other
5518	* case we have to choose the asynchronous timer mode.
5519	*
5520	* @param poffMin Pointer to the determined difference between different cores.
5521	* @return false if the time stamp counters appear to be synchronized, true otherwise.
5522	*/
5523	static bool supdrvDetermineAsyncTsc(uint64_t *poffMin)
5524	{
5525	/*
5526	* Just iterate all the cpus 8 times and make sure that the TSC is
5527	* ever increasing. We don't bother taking TSC rollover into account.
5528	*/
5529	int iEndCpu = RTMpGetArraySize();
5530	int iCpu;
5531	int cLoops = 8;
5532	bool fAsync = false;
5533	int rc = VINF_SUCCESS;
5534	uint64_t offMax = 0;
5535	uint64_t offMin = ~(uint64_t)0;
5536	uint64_t PrevTsc = ASMReadTSC();
5537
5538	while (cLoops-- > 0)
5539	{
5540	for (iCpu = 0; iCpu < iEndCpu; iCpu++)
5541	{
5542	uint64_t CurTsc;
5543	rc = RTMpOnSpecific(RTMpCpuIdFromSetIndex(iCpu), supdrvDetermineAsyncTscWorker, &CurTsc, NULL);
5544	if (RT_SUCCESS(rc))
5545	{
5546	if (CurTsc <= PrevTsc)
5547	{
5548	fAsync = true;
5549	offMin = offMax = PrevTsc - CurTsc;
5550	Log(("supdrvDetermineAsyncTsc: iCpu=%d cLoops=%d CurTsc=%llx PrevTsc=%llx\n",
5551	iCpu, cLoops, CurTsc, PrevTsc));
5552	break;
5553	}
5554
5555	/* Gather statistics (except the first time). */
5556	if (iCpu != 0 \|\| cLoops != 7)
5557	{
5558	uint64_t off = CurTsc - PrevTsc;
5559	if (off < offMin)
5560	offMin = off;
5561	if (off > offMax)
5562	offMax = off;
5563	Log2(("%d/%d: off=%llx\n", cLoops, iCpu, off));
5564	}
5565
5566	/* Next */
5567	PrevTsc = CurTsc;
5568	}
5569	else if (rc == VERR_NOT_SUPPORTED)
5570	break;
5571	else
5572	AssertMsg(rc == VERR_CPU_NOT_FOUND \|\| rc == VERR_CPU_OFFLINE, ("%d\n", rc));
5573	}
5574
5575	/* broke out of the loop. */
5576	if (iCpu < iEndCpu)
5577	break;
5578	}
5579
5580	poffMin = offMin; / Almost RTMpOnSpecific profiling. */
5581	Log(("supdrvDetermineAsyncTsc: returns %d; iEndCpu=%d rc=%d offMin=%llx offMax=%llx\n",
5582	fAsync, iEndCpu, rc, offMin, offMax));
5583	#if !defined(RT_OS_SOLARIS) && !defined(RT_OS_OS2) && !defined(RT_OS_WINDOWS)
5584	OSDBGPRINT(("vboxdrv: fAsync=%d offMin=%#lx offMax=%#lx\n", fAsync, (long)offMin, (long)offMax));
5585	#endif
5586	return fAsync;
5587	}
5588
5589
5590	/**
5591	* Determine the GIP TSC mode.
5592	*
5593	* @returns The most suitable TSC mode.
5594	* @param pDevExt Pointer to the device instance data.
5595	*/
5596	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt)
5597	{
5598	/*
5599	* On SMP we're faced with two problems:
5600	* (1) There might be a skew between the CPU, so that cpu0
5601	* returns a TSC that is slightly different from cpu1.
5602	* (2) Power management (and other things) may cause the TSC
5603	* to run at a non-constant speed, and cause the speed
5604	* to be different on the cpus. This will result in (1).
5605	*
5606	* So, on SMP systems we'll have to select the ASYNC update method
5607	* if there are symptoms of these problems.
5608	*/
5609	if (RTMpGetCount() > 1)
5610	{
5611	uint32_t uEAX, uEBX, uECX, uEDX;
5612	uint64_t u64DiffCoresIgnored;
5613
5614	/* Permit the user and/or the OS specific bits to force async mode. */
5615	if (supdrvOSGetForcedAsyncTscMode(pDevExt))
5616	return SUPGIPMODE_ASYNC_TSC;
5617
5618	/* Try check for current differences between the cpus. */
5619	if (supdrvDetermineAsyncTsc(&u64DiffCoresIgnored))
5620	return SUPGIPMODE_ASYNC_TSC;
5621
5622	/*
5623	* If the CPU supports power management and is an AMD one we
5624	* won't trust it unless it has the TscInvariant bit is set.
5625	*/
5626	/* Check for "AuthenticAMD" */
5627	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
5628	if ( uEAX >= 1
5629	&& ASMIsAmdCpuEx(uEBX, uECX, uEDX))
5630	{
5631	/* Check for APM support and that TscInvariant is cleared. */
5632	ASMCpuId(0x80000000, &uEAX, &uEBX, &uECX, &uEDX);
5633	if (uEAX >= 0x80000007)
5634	{
5635	ASMCpuId(0x80000007, &uEAX, &uEBX, &uECX, &uEDX);
5636	if ( !(uEDX & RT_BIT(8))/* TscInvariant */
5637	&& (uEDX & 0x3e)) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
5638	return SUPGIPMODE_ASYNC_TSC;
5639	}
5640	}
5641	}
5642	return SUPGIPMODE_SYNC_TSC;
5643	}
5644
5645
5646	/**
5647	* Initializes per-CPU GIP information.
5648	*
5649	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5650	* @param pCpu Pointer to which GIP CPU to initalize.
5651	* @param u64NanoTS The current nanosecond timestamp.
5652	*/
5653	static void supdrvGipInitCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pCpu, uint64_t u64NanoTS)
5654	{
5655	pCpu->u32TransactionId = 2;
5656	pCpu->u64NanoTS = u64NanoTS;
5657	pCpu->u64TSC = ASMReadTSC();
5658
5659	ASMAtomicWriteSize(&pCpu->enmState, SUPGIPCPUSTATE_INVALID);
5660	ASMAtomicWriteSize(&pCpu->idCpu, NIL_RTCPUID);
5661	ASMAtomicWriteS16(&pCpu->iCpuSet, -1);
5662	ASMAtomicWriteU16(&pCpu->idApic, UINT16_MAX);
5663
5664	/*
5665	* We don't know the following values until we've executed updates.
5666	* So, we'll just pretend it's a 4 GHz CPU and adjust the history it on
5667	* the 2nd timer callout.
5668	*/
5669	pCpu->u64CpuHz = _4G + 1; /* tstGIP-2 depends on this. */
5670	pCpu->u32UpdateIntervalTSC
5671	= pCpu->au32TSCHistory[0]
5672	= pCpu->au32TSCHistory[1]
5673	= pCpu->au32TSCHistory[2]
5674	= pCpu->au32TSCHistory[3]
5675	= pCpu->au32TSCHistory[4]
5676	= pCpu->au32TSCHistory[5]
5677	= pCpu->au32TSCHistory[6]
5678	= pCpu->au32TSCHistory[7]
5679	= (uint32_t)(_4G / pGip->u32UpdateHz);
5680	}
5681
5682
5683	/**
5684	* Initializes the GIP data.
5685	*
5686	* @param pDevExt Pointer to the device instance data.
5687	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5688	* @param HCPhys The physical address of the GIP.
5689	* @param u64NanoTS The current nanosecond timestamp.
5690	* @param uUpdateHz The update frequency.
5691	* @param cCpus The CPU count.
5692	*/
5693	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
5694	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned cCpus)
5695	{
5696	size_t const cbGip = RT_ALIGN_Z(RT_OFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), PAGE_SIZE);
5697	unsigned i;
5698	#ifdef DEBUG_DARWIN_GIP
5699	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5700	#else
5701	LogFlow(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5702	#endif
5703
5704	/*
5705	* Initialize the structure.
5706	*/
5707	memset(pGip, 0, cbGip);
5708	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
5709	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
5710	pGip->u32Mode = supdrvGipDeterminTscMode(pDevExt);
5711	pGip->cCpus = (uint16_t)cCpus;
5712	pGip->cPages = (uint16_t)(cbGip / PAGE_SIZE);
5713	pGip->u32UpdateHz = uUpdateHz;
5714	pGip->u32UpdateIntervalNS = 1000000000 / uUpdateHz;
5715	pGip->u64NanoTSLastUpdateHz = u64NanoTS;
5716	RTCpuSetEmpty(&pGip->OnlineCpuSet);
5717	RTCpuSetEmpty(&pGip->PresentCpuSet);
5718	RTMpGetSet(&pGip->PossibleCpuSet);
5719	pGip->cOnlineCpus = RTMpGetOnlineCount();
5720	pGip->cPresentCpus = RTMpGetPresentCount();
5721	pGip->cPossibleCpus = RTMpGetCount();
5722	pGip->idCpuMax = RTMpGetMaxCpuId();
5723	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromApicId); i++)
5724	pGip->aiCpuFromApicId[i] = 0;
5725	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx); i++)
5726	pGip->aiCpuFromCpuSetIdx[i] = UINT16_MAX;
5727
5728	for (i = 0; i < cCpus; i++)
5729	supdrvGipInitCpu(pGip, &pGip->aCPUs[i], u64NanoTS);
5730
5731	/*
5732	* Link it to the device extension.
5733	*/
5734	pDevExt->pGip = pGip;
5735	pDevExt->HCPhysGip = HCPhys;
5736	pDevExt->cGipUsers = 0;
5737	}
5738
5739
5740	/**
5741	* On CPU initialization callback for RTMpOnAll.
5742	*
5743	* @param idCpu The CPU ID.
5744	* @param pvUser1 The device extension.
5745	* @param pvUser2 The GIP.
5746	*/
5747	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2)
5748	{
5749	/* This is good enough, even though it will update some of the globals a
5750	bit to much. */
5751	supdrvGipMpEventOnline((PSUPDRVDEVEXT)pvUser1, idCpu);
5752	}
5753
5754
5755	/**
5756	* Invalidates the GIP data upon termination.
5757	*
5758	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5759	*/
5760	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
5761	{
5762	unsigned i;
5763	pGip->u32Magic = 0;
5764	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
5765	{
5766	pGip->aCPUs[i].u64NanoTS = 0;
5767	pGip->aCPUs[i].u64TSC = 0;
5768	pGip->aCPUs[i].iTSCHistoryHead = 0;
5769	}
5770	}
5771
5772
5773	/**
5774	* Worker routine for supdrvGipUpdate and supdrvGipUpdatePerCpu that
5775	* updates all the per cpu data except the transaction id.
5776	*
5777	* @param pDevExt The device extension.
5778	* @param pGipCpu Pointer to the per cpu data.
5779	* @param u64NanoTS The current time stamp.
5780	* @param u64TSC The current TSC.
5781	* @param iTick The current timer tick.
5782	*/
5783	static void supdrvGipDoUpdateCpu(PSUPDRVDEVEXT pDevExt, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick)
5784	{
5785	uint64_t u64TSCDelta;
5786	uint32_t u32UpdateIntervalTSC;
5787	uint32_t u32UpdateIntervalTSCSlack;
5788	unsigned iTSCHistoryHead;
5789	uint64_t u64CpuHz;
5790	uint32_t u32TransactionId;
5791
5792	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5793	AssertPtrReturnVoid(pGip);
5794
5795	/* Delta between this and the previous update. */
5796	ASMAtomicUoWriteU32(&pGipCpu->u32PrevUpdateIntervalNS, (uint32_t)(u64NanoTS - pGipCpu->u64NanoTS));
5797
5798	/*
5799	* Update the NanoTS.
5800	*/
5801	ASMAtomicWriteU64(&pGipCpu->u64NanoTS, u64NanoTS);
5802
5803	/*
5804	* Calc TSC delta.
5805	*/
5806	/** @todo validate the NanoTS delta, don't trust the OS to call us when it should... */
5807	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
5808	ASMAtomicWriteU64(&pGipCpu->u64TSC, u64TSC);
5809
5810	if (u64TSCDelta >> 32)
5811	{
5812	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
5813	pGipCpu->cErrors++;
5814	}
5815
5816	/*
5817	* On the 2nd and 3rd callout, reset the history with the current TSC
5818	* interval since the values entered by supdrvGipInit are totally off.
5819	* The interval on the 1st callout completely unreliable, the 2nd is a bit
5820	* better, while the 3rd should be most reliable.
5821	*/
5822	u32TransactionId = pGipCpu->u32TransactionId;
5823	if (RT_UNLIKELY( ( u32TransactionId == 5
5824	\|\| u32TransactionId == 7)
5825	&& ( iTick == 2
5826	\|\| iTick == 3) ))
5827	{
5828	unsigned i;
5829	for (i = 0; i < RT_ELEMENTS(pGipCpu->au32TSCHistory); i++)
5830	ASMAtomicUoWriteU32(&pGipCpu->au32TSCHistory[i], (uint32_t)u64TSCDelta);
5831	}
5832
5833	/*
5834	* TSC History.
5835	*/
5836	Assert(RT_ELEMENTS(pGipCpu->au32TSCHistory) == 8);
5837	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
5838	ASMAtomicWriteU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
5839	ASMAtomicWriteU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
5840
5841	/*
5842	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
5843	*/
5844	if (pGip->u32UpdateHz >= 1000)
5845	{
5846	uint32_t u32;
5847	u32 = pGipCpu->au32TSCHistory[0];
5848	u32 += pGipCpu->au32TSCHistory[1];
5849	u32 += pGipCpu->au32TSCHistory[2];
5850	u32 += pGipCpu->au32TSCHistory[3];
5851	u32 >>= 2;
5852	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
5853	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
5854	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
5855	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
5856	u32UpdateIntervalTSC >>= 2;
5857	u32UpdateIntervalTSC += u32;
5858	u32UpdateIntervalTSC >>= 1;
5859
5860	/* Value chosen for a 2GHz Athlon64 running linux 2.6.10/11, . */
5861	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
5862	}
5863	else if (pGip->u32UpdateHz >= 90)
5864	{
5865	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5866	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
5867	u32UpdateIntervalTSC >>= 1;
5868
5869	/* value chosen on a 2GHz thinkpad running windows */
5870	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
5871	}
5872	else
5873	{
5874	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5875
5876	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
5877	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
5878	}
5879	ASMAtomicWriteU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
5880
5881	/*
5882	* CpuHz.
5883	*/
5884	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, pGip->u32UpdateHz);
5885	ASMAtomicWriteU64(&pGipCpu->u64CpuHz, u64CpuHz);
5886	}
5887
5888
5889	/**
5890	* Updates the GIP.
5891	*
5892	* @param pDevExt The device extension.
5893	* @param u64NanoTS The current nanosecond timesamp.
5894	* @param u64TSC The current TSC timesamp.
5895	* @param idCpu The CPU ID.
5896	* @param iTick The current timer tick.
5897	*/
5898	static void supdrvGipUpdate(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick)
5899	{
5900	/*
5901	* Determine the relevant CPU data.
5902	*/
5903	PSUPGIPCPU pGipCpu;
5904	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5905	AssertPtrReturnVoid(pGip);
5906
5907	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
5908	pGipCpu = &pGip->aCPUs[0];
5909	else
5910	{
5911	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
5912	if (RT_UNLIKELY(iCpu >= pGip->cCpus))
5913	return;
5914	pGipCpu = &pGip->aCPUs[iCpu];
5915	if (RT_UNLIKELY(pGipCpu->idCpu != idCpu))
5916	return;
5917	}
5918
5919	/*
5920	* Start update transaction.
5921	*/
5922	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5923	{
5924	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
5925	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5926	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5927	pGipCpu->cErrors++;
5928	return;
5929	}
5930
5931	/*
5932	* Recalc the update frequency every 0x800th time.
5933	*/
5934	if (!(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
5935	{
5936	if (pGip->u64NanoTSLastUpdateHz)
5937	{
5938	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
5939	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
5940	uint32_t u32UpdateHz = (uint32_t)((UINT64_C(1000000000) * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
5941	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
5942	{
5943	ASMAtomicWriteU32(&pGip->u32UpdateHz, u32UpdateHz);
5944	ASMAtomicWriteU32(&pGip->u32UpdateIntervalNS, 1000000000 / u32UpdateHz);
5945	}
5946	#endif
5947	}
5948	ASMAtomicWriteU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS);
5949	}
5950
5951	/*
5952	* Update the data.
5953	*/
5954	supdrvGipDoUpdateCpu(pDevExt, pGipCpu, u64NanoTS, u64TSC, iTick);
5955
5956	/*
5957	* Complete transaction.
5958	*/
5959	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5960	}
5961
5962
5963	/**
5964	* Updates the per cpu GIP data for the calling cpu.
5965	*
5966	* @param pDevExt The device extension.
5967	* @param u64NanoTS The current nanosecond timesamp.
5968	* @param u64TSC The current TSC timesamp.
5969	* @param idCpu The CPU ID.
5970	* @param idApic The APIC id for the CPU index.
5971	* @param iTick The current timer tick.
5972	*/
5973	static void supdrvGipUpdatePerCpu(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC,
5974	RTCPUID idCpu, uint8_t idApic, uint64_t iTick)
5975	{
5976	uint32_t iCpu;
5977	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5978
5979	/*
5980	* Avoid a potential race when a CPU online notification doesn't fire on
5981	* the onlined CPU but the tick creeps in before the event notification is
5982	* run.
5983	*/
5984	if (RT_UNLIKELY(iTick == 1))
5985	{
5986	iCpu = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
5987	if (pGip->aCPUs[iCpu].enmState == SUPGIPCPUSTATE_OFFLINE)
5988	supdrvGipMpEventOnline(pDevExt, idCpu);
5989	}
5990
5991	iCpu = pGip->aiCpuFromApicId[idApic];
5992	if (RT_LIKELY(iCpu < pGip->cCpus))
5993	{
5994	PSUPGIPCPU pGipCpu = &pGip->aCPUs[iCpu];
5995	if (pGipCpu->idCpu == idCpu)
5996	{
5997	/*
5998	* Start update transaction.
5999	*/
6000	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
6001	{
6002	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
6003	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
6004	pGipCpu->cErrors++;
6005	return;
6006	}
6007
6008	/*
6009	* Update the data.
6010	*/
6011	supdrvGipDoUpdateCpu(pDevExt, pGipCpu, u64NanoTS, u64TSC, iTick);
6012
6013	/*
6014	* Complete transaction.
6015	*/
6016	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
6017	}
6018	}
6019	}
6020

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

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDrv.c@ 47537

Download in other formats: