SUPDrv.c@ 41147

Last change on this file since 41147 was 41147, checked in by vboxsync, 13 years ago
SUPDrv,VMM: Tracepoints in raw-mode.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 217.6 KB

Line
1	/* $Id: SUPDrv.c 41147 2012-05-03 20:15:27Z vboxsync $ */
2	/** @file
3	* VBoxDrv - The VirtualBox Support Driver - Common code.
4	*/
5
6	/*
7	* Copyright (C) 2006-2012 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/hwacc_svm.h>
63	#include <VBox/vmm/hwacc_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	{ "SUPR0GetPagingMode", (void *)SUPR0GetPagingMode },
176	{ "SUPR0LockMem", (void *)SUPR0LockMem },
177	{ "SUPR0LowAlloc", (void *)SUPR0LowAlloc },
178	{ "SUPR0LowFree", (void *)SUPR0LowFree },
179	{ "SUPR0MemAlloc", (void *)SUPR0MemAlloc },
180	{ "SUPR0MemFree", (void *)SUPR0MemFree },
181	{ "SUPR0MemGetPhys", (void *)SUPR0MemGetPhys },
182	{ "SUPR0ObjAddRef", (void *)SUPR0ObjAddRef },
183	{ "SUPR0ObjAddRefEx", (void *)SUPR0ObjAddRefEx },
184	{ "SUPR0ObjRegister", (void *)SUPR0ObjRegister },
185	{ "SUPR0ObjRelease", (void *)SUPR0ObjRelease },
186	{ "SUPR0ObjVerifyAccess", (void *)SUPR0ObjVerifyAccess },
187	{ "SUPR0PageAllocEx", (void *)SUPR0PageAllocEx },
188	{ "SUPR0PageFree", (void *)SUPR0PageFree },
189	{ "SUPR0Printf", (void *)SUPR0Printf },
190	{ "SUPR0TracerDeregisterDrv", (void *)SUPR0TracerDeregisterDrv },
191	{ "SUPR0TracerDeregisterImpl", (void *)SUPR0TracerDeregisterImpl },
192	{ "SUPR0TracerFireProbe", (void *)SUPR0TracerFireProbe },
193	{ "SUPR0TracerRegisterDrv", (void *)SUPR0TracerRegisterDrv },
194	{ "SUPR0TracerRegisterImpl", (void *)SUPR0TracerRegisterImpl },
195	{ "SUPR0TracerRegisterModule", (void *)SUPR0TracerRegisterModule },
196	{ "SUPR0TracerUmodProbeFire", (void *)SUPR0TracerUmodProbeFire },
197	{ "SUPR0UnlockMem", (void *)SUPR0UnlockMem },
198	{ "SUPSemEventClose", (void *)SUPSemEventClose },
199	{ "SUPSemEventCreate", (void *)SUPSemEventCreate },
200	{ "SUPSemEventGetResolution", (void *)SUPSemEventGetResolution },
201	{ "SUPSemEventMultiClose", (void *)SUPSemEventMultiClose },
202	{ "SUPSemEventMultiCreate", (void *)SUPSemEventMultiCreate },
203	{ "SUPSemEventMultiGetResolution", (void *)SUPSemEventMultiGetResolution },
204	{ "SUPSemEventMultiReset", (void *)SUPSemEventMultiReset },
205	{ "SUPSemEventMultiSignal", (void *)SUPSemEventMultiSignal },
206	{ "SUPSemEventMultiWait", (void *)SUPSemEventMultiWait },
207	{ "SUPSemEventMultiWaitNoResume", (void *)SUPSemEventMultiWaitNoResume },
208	{ "SUPSemEventMultiWaitNsAbsIntr", (void *)SUPSemEventMultiWaitNsAbsIntr },
209	{ "SUPSemEventMultiWaitNsRelIntr", (void *)SUPSemEventMultiWaitNsRelIntr },
210	{ "SUPSemEventSignal", (void *)SUPSemEventSignal },
211	{ "SUPSemEventWait", (void *)SUPSemEventWait },
212	{ "SUPSemEventWaitNoResume", (void *)SUPSemEventWaitNoResume },
213	{ "SUPSemEventWaitNsAbsIntr", (void *)SUPSemEventWaitNsAbsIntr },
214	{ "SUPSemEventWaitNsRelIntr", (void *)SUPSemEventWaitNsRelIntr },
215
216	{ "RTAssertAreQuiet", (void *)RTAssertAreQuiet },
217	{ "RTAssertMayPanic", (void *)RTAssertMayPanic },
218	{ "RTAssertMsg1", (void *)RTAssertMsg1 },
219	{ "RTAssertMsg2AddV", (void *)RTAssertMsg2AddV },
220	{ "RTAssertMsg2V", (void *)RTAssertMsg2V },
221	{ "RTAssertSetMayPanic", (void *)RTAssertSetMayPanic },
222	{ "RTAssertSetQuiet", (void *)RTAssertSetQuiet },
223	{ "RTCrc32", (void *)RTCrc32 },
224	{ "RTCrc32Finish", (void *)RTCrc32Finish },
225	{ "RTCrc32Process", (void *)RTCrc32Process },
226	{ "RTCrc32Start", (void *)RTCrc32Start },
227	{ "RTErrConvertFromErrno", (void *)RTErrConvertFromErrno },
228	{ "RTErrConvertToErrno", (void *)RTErrConvertToErrno },
229	{ "RTHandleTableAllocWithCtx", (void *)RTHandleTableAllocWithCtx },
230	{ "RTHandleTableCreate", (void *)RTHandleTableCreate },
231	{ "RTHandleTableCreateEx", (void *)RTHandleTableCreateEx },
232	{ "RTHandleTableDestroy", (void *)RTHandleTableDestroy },
233	{ "RTHandleTableFreeWithCtx", (void *)RTHandleTableFreeWithCtx },
234	{ "RTHandleTableLookupWithCtx", (void *)RTHandleTableLookupWithCtx },
235	{ "RTLogDefaultInstance", (void *)RTLogDefaultInstance },
236	{ "RTLogLoggerExV", (void *)RTLogLoggerExV },
237	{ "RTLogPrintfV", (void *)RTLogPrintfV },
238	{ "RTLogRelDefaultInstance", (void *)RTLogRelDefaultInstance },
239	{ "RTLogSetDefaultInstanceThread", (void *)RTLogSetDefaultInstanceThread },
240	{ "RTMemAllocExTag", (void *)RTMemAllocExTag },
241	{ "RTMemAllocTag", (void *)RTMemAllocTag },
242	{ "RTMemAllocVarTag", (void *)RTMemAllocVarTag },
243	{ "RTMemAllocZTag", (void *)RTMemAllocZTag },
244	{ "RTMemAllocZVarTag", (void *)RTMemAllocZVarTag },
245	{ "RTMemDupExTag", (void *)RTMemDupExTag },
246	{ "RTMemDupTag", (void *)RTMemDupTag },
247	{ "RTMemFree", (void *)RTMemFree },
248	{ "RTMemFreeEx", (void *)RTMemFreeEx },
249	{ "RTMemReallocTag", (void *)RTMemReallocTag },
250	{ "RTMpCpuId", (void *)RTMpCpuId },
251	{ "RTMpCpuIdFromSetIndex", (void *)RTMpCpuIdFromSetIndex },
252	{ "RTMpCpuIdToSetIndex", (void *)RTMpCpuIdToSetIndex },
253	{ "RTMpGetArraySize", (void *)RTMpGetArraySize },
254	{ "RTMpGetCount", (void *)RTMpGetCount },
255	{ "RTMpGetMaxCpuId", (void *)RTMpGetMaxCpuId },
256	{ "RTMpGetOnlineCount", (void *)RTMpGetOnlineCount },
257	{ "RTMpGetOnlineSet", (void *)RTMpGetOnlineSet },
258	{ "RTMpGetSet", (void *)RTMpGetSet },
259	{ "RTMpIsCpuOnline", (void *)RTMpIsCpuOnline },
260	{ "RTMpIsCpuPossible", (void *)RTMpIsCpuPossible },
261	{ "RTMpIsCpuWorkPending", (void *)RTMpIsCpuWorkPending },
262	{ "RTMpNotificationDeregister", (void *)RTMpNotificationDeregister },
263	{ "RTMpNotificationRegister", (void *)RTMpNotificationRegister },
264	{ "RTMpOnAll", (void *)RTMpOnAll },
265	{ "RTMpOnOthers", (void *)RTMpOnOthers },
266	{ "RTMpOnSpecific", (void *)RTMpOnSpecific },
267	{ "RTMpPokeCpu", (void *)RTMpPokeCpu },
268	{ "RTNetIPv4AddDataChecksum", (void *)RTNetIPv4AddDataChecksum },
269	{ "RTNetIPv4AddTCPChecksum", (void *)RTNetIPv4AddTCPChecksum },
270	{ "RTNetIPv4AddUDPChecksum", (void *)RTNetIPv4AddUDPChecksum },
271	{ "RTNetIPv4FinalizeChecksum", (void *)RTNetIPv4FinalizeChecksum },
272	{ "RTNetIPv4HdrChecksum", (void *)RTNetIPv4HdrChecksum },
273	{ "RTNetIPv4IsDHCPValid", (void *)RTNetIPv4IsDHCPValid },
274	{ "RTNetIPv4IsHdrValid", (void *)RTNetIPv4IsHdrValid },
275	{ "RTNetIPv4IsTCPSizeValid", (void *)RTNetIPv4IsTCPSizeValid },
276	{ "RTNetIPv4IsTCPValid", (void *)RTNetIPv4IsTCPValid },
277	{ "RTNetIPv4IsUDPSizeValid", (void *)RTNetIPv4IsUDPSizeValid },
278	{ "RTNetIPv4IsUDPValid", (void *)RTNetIPv4IsUDPValid },
279	{ "RTNetIPv4PseudoChecksum", (void *)RTNetIPv4PseudoChecksum },
280	{ "RTNetIPv4PseudoChecksumBits", (void *)RTNetIPv4PseudoChecksumBits },
281	{ "RTNetIPv4TCPChecksum", (void *)RTNetIPv4TCPChecksum },
282	{ "RTNetIPv4UDPChecksum", (void *)RTNetIPv4UDPChecksum },
283	{ "RTNetIPv6PseudoChecksum", (void *)RTNetIPv6PseudoChecksum },
284	{ "RTNetIPv6PseudoChecksumBits", (void *)RTNetIPv6PseudoChecksumBits },
285	{ "RTNetIPv6PseudoChecksumEx", (void *)RTNetIPv6PseudoChecksumEx },
286	{ "RTNetTCPChecksum", (void *)RTNetTCPChecksum },
287	{ "RTNetUDPChecksum", (void *)RTNetUDPChecksum },
288	{ "RTPowerNotificationDeregister", (void *)RTPowerNotificationDeregister },
289	{ "RTPowerNotificationRegister", (void *)RTPowerNotificationRegister },
290	{ "RTProcSelf", (void *)RTProcSelf },
291	{ "RTR0AssertPanicSystem", (void *)RTR0AssertPanicSystem },
292	{ "RTR0MemAreKrnlAndUsrDifferent", (void *)RTR0MemAreKrnlAndUsrDifferent },
293	{ "RTR0MemKernelIsValidAddr", (void *)RTR0MemKernelIsValidAddr },
294	{ "RTR0MemObjAddress", (void *)RTR0MemObjAddress },
295	{ "RTR0MemObjAddressR3", (void *)RTR0MemObjAddressR3 },
296	{ "RTR0MemObjAllocContTag", (void *)RTR0MemObjAllocContTag },
297	{ "RTR0MemObjAllocLowTag", (void *)RTR0MemObjAllocLowTag },
298	{ "RTR0MemObjAllocPageTag", (void *)RTR0MemObjAllocPageTag },
299	{ "RTR0MemObjAllocPhysExTag", (void *)RTR0MemObjAllocPhysExTag },
300	{ "RTR0MemObjAllocPhysNCTag", (void *)RTR0MemObjAllocPhysNCTag },
301	{ "RTR0MemObjAllocPhysTag", (void *)RTR0MemObjAllocPhysTag },
302	{ "RTR0MemObjEnterPhysTag", (void *)RTR0MemObjEnterPhysTag },
303	{ "RTR0MemObjFree", (void *)RTR0MemObjFree },
304	{ "RTR0MemObjGetPagePhysAddr", (void *)RTR0MemObjGetPagePhysAddr },
305	{ "RTR0MemObjIsMapping", (void *)RTR0MemObjIsMapping },
306	{ "RTR0MemObjLockUserTag", (void *)RTR0MemObjLockUserTag },
307	{ "RTR0MemObjMapKernelExTag", (void *)RTR0MemObjMapKernelExTag },
308	{ "RTR0MemObjMapKernelTag", (void *)RTR0MemObjMapKernelTag },
309	{ "RTR0MemObjMapUserTag", (void *)RTR0MemObjMapUserTag },
310	{ "RTR0MemObjProtect", (void *)RTR0MemObjProtect },
311	{ "RTR0MemObjSize", (void *)RTR0MemObjSize },
312	{ "RTR0MemUserCopyFrom", (void *)RTR0MemUserCopyFrom },
313	{ "RTR0MemUserCopyTo", (void *)RTR0MemUserCopyTo },
314	{ "RTR0MemUserIsValidAddr", (void *)RTR0MemUserIsValidAddr },
315	{ "RTR0ProcHandleSelf", (void *)RTR0ProcHandleSelf },
316	{ "RTSemEventCreate", (void *)RTSemEventCreate },
317	{ "RTSemEventDestroy", (void *)RTSemEventDestroy },
318	{ "RTSemEventGetResolution", (void *)RTSemEventGetResolution },
319	{ "RTSemEventMultiCreate", (void *)RTSemEventMultiCreate },
320	{ "RTSemEventMultiDestroy", (void *)RTSemEventMultiDestroy },
321	{ "RTSemEventMultiGetResolution", (void *)RTSemEventMultiGetResolution },
322	{ "RTSemEventMultiReset", (void *)RTSemEventMultiReset },
323	{ "RTSemEventMultiSignal", (void *)RTSemEventMultiSignal },
324	{ "RTSemEventMultiWait", (void *)RTSemEventMultiWait },
325	{ "RTSemEventMultiWaitEx", (void *)RTSemEventMultiWaitEx },
326	{ "RTSemEventMultiWaitExDebug", (void *)RTSemEventMultiWaitExDebug },
327	{ "RTSemEventMultiWaitNoResume", (void *)RTSemEventMultiWaitNoResume },
328	{ "RTSemEventSignal", (void *)RTSemEventSignal },
329	{ "RTSemEventWait", (void *)RTSemEventWait },
330	{ "RTSemEventWaitEx", (void *)RTSemEventWaitEx },
331	{ "RTSemEventWaitExDebug", (void *)RTSemEventWaitExDebug },
332	{ "RTSemEventWaitNoResume", (void *)RTSemEventWaitNoResume },
333	{ "RTSemFastMutexCreate", (void *)RTSemFastMutexCreate },
334	{ "RTSemFastMutexDestroy", (void *)RTSemFastMutexDestroy },
335	{ "RTSemFastMutexRelease", (void *)RTSemFastMutexRelease },
336	{ "RTSemFastMutexRequest", (void *)RTSemFastMutexRequest },
337	{ "RTSemMutexCreate", (void *)RTSemMutexCreate },
338	{ "RTSemMutexDestroy", (void *)RTSemMutexDestroy },
339	{ "RTSemMutexRelease", (void *)RTSemMutexRelease },
340	{ "RTSemMutexRequest", (void *)RTSemMutexRequest },
341	{ "RTSemMutexRequestDebug", (void *)RTSemMutexRequestDebug },
342	{ "RTSemMutexRequestNoResume", (void *)RTSemMutexRequestNoResume },
343	{ "RTSemMutexRequestNoResumeDebug", (void *)RTSemMutexRequestNoResumeDebug },
344	{ "RTSpinlockAcquire", (void *)RTSpinlockAcquire },
345	{ "RTSpinlockCreate", (void *)RTSpinlockCreate },
346	{ "RTSpinlockDestroy", (void *)RTSpinlockDestroy },
347	{ "RTSpinlockRelease", (void *)RTSpinlockRelease },
348	{ "RTSpinlockReleaseNoInts", (void *)RTSpinlockReleaseNoInts },
349	{ "RTStrCopy", (void *)RTStrCopy },
350	{ "RTStrDupTag", (void *)RTStrDupTag },
351	{ "RTStrFormat", (void *)RTStrFormat },
352	{ "RTStrFormatNumber", (void *)RTStrFormatNumber },
353	{ "RTStrFormatTypeDeregister", (void *)RTStrFormatTypeDeregister },
354	{ "RTStrFormatTypeRegister", (void *)RTStrFormatTypeRegister },
355	{ "RTStrFormatTypeSetUser", (void *)RTStrFormatTypeSetUser },
356	{ "RTStrFormatV", (void *)RTStrFormatV },
357	{ "RTStrFree", (void *)RTStrFree },
358	{ "RTStrNCmp", (void *)RTStrNCmp },
359	{ "RTStrPrintf", (void *)RTStrPrintf },
360	{ "RTStrPrintfEx", (void *)RTStrPrintfEx },
361	{ "RTStrPrintfExV", (void *)RTStrPrintfExV },
362	{ "RTStrPrintfV", (void *)RTStrPrintfV },
363	{ "RTThreadCreate", (void *)RTThreadCreate },
364	{ "RTThreadGetName", (void *)RTThreadGetName },
365	{ "RTThreadGetNative", (void *)RTThreadGetNative },
366	{ "RTThreadGetType", (void *)RTThreadGetType },
367	{ "RTThreadIsInInterrupt", (void *)RTThreadIsInInterrupt },
368	{ "RTThreadNativeSelf", (void *)RTThreadNativeSelf },
369	{ "RTThreadPreemptDisable", (void *)RTThreadPreemptDisable },
370	{ "RTThreadPreemptIsEnabled", (void *)RTThreadPreemptIsEnabled },
371	{ "RTThreadPreemptIsPending", (void *)RTThreadPreemptIsPending },
372	{ "RTThreadPreemptIsPendingTrusty", (void *)RTThreadPreemptIsPendingTrusty },
373	{ "RTThreadPreemptIsPossible", (void *)RTThreadPreemptIsPossible },
374	{ "RTThreadPreemptRestore", (void *)RTThreadPreemptRestore },
375	{ "RTThreadSelf", (void *)RTThreadSelf },
376	{ "RTThreadSelfName", (void *)RTThreadSelfName },
377	{ "RTThreadSleep", (void *)RTThreadSleep },
378	{ "RTThreadUserReset", (void *)RTThreadUserReset },
379	{ "RTThreadUserSignal", (void *)RTThreadUserSignal },
380	{ "RTThreadUserWait", (void *)RTThreadUserWait },
381	{ "RTThreadUserWaitNoResume", (void *)RTThreadUserWaitNoResume },
382	{ "RTThreadWait", (void *)RTThreadWait },
383	{ "RTThreadWaitNoResume", (void *)RTThreadWaitNoResume },
384	{ "RTThreadYield", (void *)RTThreadYield },
385	{ "RTTimeMilliTS", (void *)RTTimeMilliTS },
386	{ "RTTimeNanoTS", (void *)RTTimeNanoTS },
387	{ "RTTimeNow", (void *)RTTimeNow },
388	{ "RTTimerCanDoHighResolution", (void *)RTTimerCanDoHighResolution },
389	{ "RTTimerChangeInterval", (void *)RTTimerChangeInterval },
390	{ "RTTimerCreate", (void *)RTTimerCreate },
391	{ "RTTimerCreateEx", (void *)RTTimerCreateEx },
392	{ "RTTimerDestroy", (void *)RTTimerDestroy },
393	{ "RTTimerGetSystemGranularity", (void *)RTTimerGetSystemGranularity },
394	{ "RTTimerReleaseSystemGranularity", (void *)RTTimerReleaseSystemGranularity },
395	{ "RTTimerRequestSystemGranularity", (void *)RTTimerRequestSystemGranularity },
396	{ "RTTimerStart", (void *)RTTimerStart },
397	{ "RTTimerStop", (void *)RTTimerStop },
398	{ "RTTimeSystemMilliTS", (void *)RTTimeSystemMilliTS },
399	{ "RTTimeSystemNanoTS", (void *)RTTimeSystemNanoTS },
400	{ "RTUuidCompare", (void *)RTUuidCompare },
401	{ "RTUuidCompareStr", (void *)RTUuidCompareStr },
402	{ "RTUuidFromStr", (void *)RTUuidFromStr },
403	/* SED: END */
404	};
405
406	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
407	/**
408	* Drag in the rest of IRPT since we share it with the
409	* rest of the kernel modules on darwin.
410	*/
411	PFNRT g_apfnVBoxDrvIPRTDeps[] =
412	{
413	/* VBoxNetAdp */
414	(PFNRT)RTRandBytes,
415	/* VBoxUSB */
416	(PFNRT)RTPathStripFilename,
417	NULL
418	};
419	#endif /* RT_OS_DARWIN \|\| RT_OS_SOLARIS \|\| RT_OS_SOLARIS */
420
421
422	/**
423	* Initializes the device extentsion structure.
424	*
425	* @returns IPRT status code.
426	* @param pDevExt The device extension to initialize.
427	* @param cbSession The size of the session structure. The size of
428	* SUPDRVSESSION may be smaller when SUPDRV_AGNOSTIC is
429	* defined because we're skipping the OS specific members
430	* then.
431	*/
432	int VBOXCALL supdrvInitDevExt(PSUPDRVDEVEXT pDevExt, size_t cbSession)
433	{
434	int rc;
435
436	#ifdef SUPDRV_WITH_RELEASE_LOGGER
437	/*
438	* Create the release log.
439	*/
440	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
441	PRTLOGGER pRelLogger;
442	rc = RTLogCreate(&pRelLogger, 0 /* fFlags */, "all",
443	"VBOX_RELEASE_LOG", RT_ELEMENTS(s_apszGroups), s_apszGroups, RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER, NULL);
444	if (RT_SUCCESS(rc))
445	RTLogRelSetDefaultInstance(pRelLogger);
446	/** @todo Add native hook for getting logger config parameters and setting
447	* them. On linux we should use the module parameter stuff... */
448	#endif
449
450	/*
451	* Initialize it.
452	*/
453	memset(pDevExt, 0, sizeof(*pDevExt));
454	rc = RTSpinlockCreate(&pDevExt->Spinlock, RTSPINLOCK_FLAGS_INTERRUPT_UNSAFE, "SUPDrvDevExt");
455	if (RT_SUCCESS(rc))
456	{
457	rc = RTSpinlockCreate(&pDevExt->hGipSpinlock, RTSPINLOCK_FLAGS_INTERRUPT_SAFE, "SUPDrvGip");
458	if (RT_SUCCESS(rc))
459	{
460	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
461	rc = RTSemMutexCreate(&pDevExt->mtxLdr);
462	#else
463	rc = RTSemFastMutexCreate(&pDevExt->mtxLdr);
464	#endif
465	if (RT_SUCCESS(rc))
466	{
467	rc = RTSemFastMutexCreate(&pDevExt->mtxComponentFactory);
468	if (RT_SUCCESS(rc))
469	{
470	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
471	rc = RTSemMutexCreate(&pDevExt->mtxGip);
472	#else
473	rc = RTSemFastMutexCreate(&pDevExt->mtxGip);
474	#endif
475	if (RT_SUCCESS(rc))
476	{
477	rc = supdrvGipCreate(pDevExt);
478	if (RT_SUCCESS(rc))
479	{
480	rc = supdrvTracerInit(pDevExt);
481	if (RT_SUCCESS(rc))
482	{
483	pDevExt->pLdrInitImage = NULL;
484	pDevExt->hLdrInitThread = NIL_RTNATIVETHREAD;
485	pDevExt->u32Cookie = BIRD; /** @todo make this random? */
486	pDevExt->cbSession = (uint32_t)cbSession;
487
488	/*
489	* Fixup the absolute symbols.
490	*
491	* Because of the table indexing assumptions we'll have a little #ifdef orgy
492	* here rather than distributing this to OS specific files. At least for now.
493	*/
494	#ifdef RT_OS_DARWIN
495	# if ARCH_BITS == 32
496	if (SUPR0GetPagingMode() >= SUPPAGINGMODE_AMD64)
497	{
498	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
499	g_aFunctions[1].pfn = (void )0x80; / SUPR0Abs64bitKernelCS - KERNEL64_CS, seg.h */
500	g_aFunctions[2].pfn = (void )0x88; / SUPR0Abs64bitKernelSS - KERNEL64_SS, seg.h */
501	g_aFunctions[3].pfn = (void )0x88; / SUPR0Abs64bitKernelDS - KERNEL64_SS, seg.h */
502	}
503	else
504	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
505	g_aFunctions[4].pfn = (void )0x08; / SUPR0AbsKernelCS - KERNEL_CS, seg.h */
506	g_aFunctions[5].pfn = (void )0x10; / SUPR0AbsKernelSS - KERNEL_DS, seg.h */
507	g_aFunctions[6].pfn = (void )0x10; / SUPR0AbsKernelDS - KERNEL_DS, seg.h */
508	g_aFunctions[7].pfn = (void )0x10; / SUPR0AbsKernelES - KERNEL_DS, seg.h */
509	g_aFunctions[8].pfn = (void )0x10; / SUPR0AbsKernelFS - KERNEL_DS, seg.h */
510	g_aFunctions[9].pfn = (void )0x48; / SUPR0AbsKernelGS - CPU_DATA_GS, seg.h */
511	# else /* 64-bit darwin: */
512	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
513	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
514	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
515	g_aFunctions[3].pfn = (void )0; / SUPR0Abs64bitKernelDS */
516	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
517	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
518	g_aFunctions[6].pfn = (void )0; / SUPR0AbsKernelDS */
519	g_aFunctions[7].pfn = (void )0; / SUPR0AbsKernelES */
520	g_aFunctions[8].pfn = (void )0; / SUPR0AbsKernelFS */
521	g_aFunctions[9].pfn = (void )0; / SUPR0AbsKernelGS */
522
523	# endif
524	#else /* !RT_OS_DARWIN */
525	# if ARCH_BITS == 64
526	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
527	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
528	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
529	g_aFunctions[3].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0Abs64bitKernelDS */
530	# else
531	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
532	# endif
533	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
534	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
535	g_aFunctions[6].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0AbsKernelDS */
536	g_aFunctions[7].pfn = (void )(uintptr_t)ASMGetES(); / SUPR0AbsKernelES */
537	g_aFunctions[8].pfn = (void )(uintptr_t)ASMGetFS(); / SUPR0AbsKernelFS */
538	g_aFunctions[9].pfn = (void )(uintptr_t)ASMGetGS(); / SUPR0AbsKernelGS */
539	#endif /* !RT_OS_DARWIN */
540	return VINF_SUCCESS;
541	}
542
543	supdrvGipDestroy(pDevExt);
544	}
545
546	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
547	RTSemMutexDestroy(pDevExt->mtxGip);
548	pDevExt->mtxGip = NIL_RTSEMMUTEX;
549	#else
550	RTSemFastMutexDestroy(pDevExt->mtxGip);
551	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
552	#endif
553	}
554	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
555	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
556	}
557	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
558	RTSemMutexDestroy(pDevExt->mtxLdr);
559	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
560	#else
561	RTSemFastMutexDestroy(pDevExt->mtxLdr);
562	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
563	#endif
564	}
565	RTSpinlockDestroy(pDevExt->hGipSpinlock);
566	pDevExt->hGipSpinlock = NIL_RTSPINLOCK;
567	}
568	RTSpinlockDestroy(pDevExt->Spinlock);
569	pDevExt->Spinlock = NIL_RTSPINLOCK;
570	}
571	#ifdef SUPDRV_WITH_RELEASE_LOGGER
572	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
573	RTLogDestroy(RTLogSetDefaultInstance(NULL));
574	#endif
575
576	return rc;
577	}
578
579
580	/**
581	* Delete the device extension (e.g. cleanup members).
582	*
583	* @param pDevExt The device extension to delete.
584	*/
585	void VBOXCALL supdrvDeleteDevExt(PSUPDRVDEVEXT pDevExt)
586	{
587	PSUPDRVOBJ pObj;
588	PSUPDRVUSAGE pUsage;
589
590	/*
591	* Kill mutexes and spinlocks.
592	*/
593	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
594	RTSemMutexDestroy(pDevExt->mtxGip);
595	pDevExt->mtxGip = NIL_RTSEMMUTEX;
596	#else
597	RTSemFastMutexDestroy(pDevExt->mtxGip);
598	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
599	#endif
600	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
601	RTSemMutexDestroy(pDevExt->mtxLdr);
602	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
603	#else
604	RTSemFastMutexDestroy(pDevExt->mtxLdr);
605	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
606	#endif
607	RTSpinlockDestroy(pDevExt->Spinlock);
608	pDevExt->Spinlock = NIL_RTSPINLOCK;
609	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
610	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
611
612	/*
613	* Free lists.
614	*/
615	/* objects. */
616	pObj = pDevExt->pObjs;
617	Assert(!pObj); /* (can trigger on forced unloads) */
618	pDevExt->pObjs = NULL;
619	while (pObj)
620	{
621	void *pvFree = pObj;
622	pObj = pObj->pNext;
623	RTMemFree(pvFree);
624	}
625
626	/* usage records. */
627	pUsage = pDevExt->pUsageFree;
628	pDevExt->pUsageFree = NULL;
629	while (pUsage)
630	{
631	void *pvFree = pUsage;
632	pUsage = pUsage->pNext;
633	RTMemFree(pvFree);
634	}
635
636	/* kill the GIP. */
637	supdrvGipDestroy(pDevExt);
638	RTSpinlockDestroy(pDevExt->hGipSpinlock);
639	pDevExt->hGipSpinlock = NIL_RTSPINLOCK;
640
641	supdrvTracerTerm(pDevExt);
642
643	#ifdef SUPDRV_WITH_RELEASE_LOGGER
644	/* destroy the loggers. */
645	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
646	RTLogDestroy(RTLogSetDefaultInstance(NULL));
647	#endif
648	}
649
650
651	/**
652	* Create session.
653	*
654	* @returns IPRT status code.
655	* @param pDevExt Device extension.
656	* @param fUser Flag indicating whether this is a user or kernel session.
657	* @param ppSession Where to store the pointer to the session data.
658	*/
659	int VBOXCALL supdrvCreateSession(PSUPDRVDEVEXT pDevExt, bool fUser, PSUPDRVSESSION *ppSession)
660	{
661	/*
662	* Allocate memory for the session data.
663	*/
664	int rc;
665	PSUPDRVSESSION pSession = *ppSession = (PSUPDRVSESSION)RTMemAllocZ(pDevExt->cbSession);
666	if (pSession)
667	{
668	/* Initialize session data. */
669	rc = RTSpinlockCreate(&pSession->Spinlock, RTSPINLOCK_FLAGS_INTERRUPT_UNSAFE, "SUPDrvSession");
670	if (!rc)
671	{
672	rc = RTHandleTableCreateEx(&pSession->hHandleTable,
673	RTHANDLETABLE_FLAGS_LOCKED \| RTHANDLETABLE_FLAGS_CONTEXT,
674	1 /uBase/, 32768 /cMax/, supdrvSessionObjHandleRetain, pSession);
675	if (RT_SUCCESS(rc))
676	{
677	Assert(pSession->Spinlock != NIL_RTSPINLOCK);
678	pSession->pDevExt = pDevExt;
679	pSession->u32Cookie = BIRD_INV;
680	/*pSession->pLdrUsage = NULL;
681	pSession->pVM = NULL;
682	pSession->pUsage = NULL;
683	pSession->pGip = NULL;
684	pSession->fGipReferenced = false;
685	pSession->Bundle.cUsed = 0; */
686	pSession->Uid = NIL_RTUID;
687	pSession->Gid = NIL_RTGID;
688	if (fUser)
689	{
690	pSession->Process = RTProcSelf();
691	pSession->R0Process = RTR0ProcHandleSelf();
692	}
693	else
694	{
695	pSession->Process = NIL_RTPROCESS;
696	pSession->R0Process = NIL_RTR0PROCESS;
697	}
698	/pSession->uTracerData = 0;/
699	pSession->hTracerCaller = NIL_RTNATIVETHREAD;
700	RTListInit(&pSession->TpProviders);
701	/pSession->cTpProviders = 0;/
702	/pSession->cTpProbesFiring = 0;/
703	RTListInit(&pSession->TpUmods);
704	/RT_ZERO(pSession->apTpLookupTable);/
705
706	VBOXDRV_SESSION_CREATE(pSession, fUser);
707	LogFlow(("Created session %p initial cookie=%#x\n", pSession, pSession->u32Cookie));
708	return VINF_SUCCESS;
709	}
710
711	RTSpinlockDestroy(pSession->Spinlock);
712	}
713	RTMemFree(pSession);
714	*ppSession = NULL;
715	Log(("Failed to create spinlock, rc=%d!\n", rc));
716	}
717	else
718	rc = VERR_NO_MEMORY;
719
720	return rc;
721	}
722
723
724	/**
725	* Shared code for cleaning up a session.
726	*
727	* @param pDevExt Device extension.
728	* @param pSession Session data.
729	* This data will be freed by this routine.
730	*/
731	void VBOXCALL supdrvCloseSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
732	{
733	VBOXDRV_SESSION_CLOSE(pSession);
734
735	/*
736	* Cleanup the session first.
737	*/
738	supdrvCleanupSession(pDevExt, pSession);
739
740	/*
741	* Free the rest of the session stuff.
742	*/
743	RTSpinlockDestroy(pSession->Spinlock);
744	pSession->Spinlock = NIL_RTSPINLOCK;
745	pSession->pDevExt = NULL;
746	RTMemFree(pSession);
747	LogFlow(("supdrvCloseSession: returns\n"));
748	}
749
750
751	/**
752	* Shared code for cleaning up a session (but not quite freeing it).
753	*
754	* This is primarily intended for MAC OS X where we have to clean up the memory
755	* stuff before the file handle is closed.
756	*
757	* @param pDevExt Device extension.
758	* @param pSession Session data.
759	* This data will be freed by this routine.
760	*/
761	void VBOXCALL supdrvCleanupSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
762	{
763	int rc;
764	PSUPDRVBUNDLE pBundle;
765	LogFlow(("supdrvCleanupSession: pSession=%p\n", pSession));
766
767	/*
768	* Remove logger instances related to this session.
769	*/
770	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pSession);
771
772	/*
773	* Destroy the handle table.
774	*/
775	rc = RTHandleTableDestroy(pSession->hHandleTable, supdrvSessionObjHandleDelete, pSession);
776	AssertRC(rc);
777	pSession->hHandleTable = NIL_RTHANDLETABLE;
778
779	/*
780	* Release object references made in this session.
781	* In theory there should be noone racing us in this session.
782	*/
783	Log2(("release objects - start\n"));
784	if (pSession->pUsage)
785	{
786	PSUPDRVUSAGE pUsage;
787	RTSpinlockAcquire(pDevExt->Spinlock);
788
789	while ((pUsage = pSession->pUsage) != NULL)
790	{
791	PSUPDRVOBJ pObj = pUsage->pObj;
792	pSession->pUsage = pUsage->pNext;
793
794	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
795	if (pUsage->cUsage < pObj->cUsage)
796	{
797	pObj->cUsage -= pUsage->cUsage;
798	RTSpinlockRelease(pDevExt->Spinlock);
799	}
800	else
801	{
802	/* Destroy the object and free the record. */
803	if (pDevExt->pObjs == pObj)
804	pDevExt->pObjs = pObj->pNext;
805	else
806	{
807	PSUPDRVOBJ pObjPrev;
808	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
809	if (pObjPrev->pNext == pObj)
810	{
811	pObjPrev->pNext = pObj->pNext;
812	break;
813	}
814	Assert(pObjPrev);
815	}
816	RTSpinlockRelease(pDevExt->Spinlock);
817
818	Log(("supdrvCleanupSession: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
819	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
820	if (pObj->pfnDestructor)
821	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
822	RTMemFree(pObj);
823	}
824
825	/* free it and continue. */
826	RTMemFree(pUsage);
827
828	RTSpinlockAcquire(pDevExt->Spinlock);
829	}
830
831	RTSpinlockRelease(pDevExt->Spinlock);
832	AssertMsg(!pSession->pUsage, ("Some buster reregistered an object during desturction!\n"));
833	}
834	Log2(("release objects - done\n"));
835
836	/*
837	* Do tracer cleanups related to this session.
838	*/
839	Log2(("release tracer stuff - start\n"));
840	supdrvTracerCleanupSession(pDevExt, pSession);
841	Log2(("release tracer stuff - end\n"));
842
843	/*
844	* Release memory allocated in the session.
845	*
846	* We do not serialize this as we assume that the application will
847	* not allocated memory while closing the file handle object.
848	*/
849	Log2(("freeing memory:\n"));
850	pBundle = &pSession->Bundle;
851	while (pBundle)
852	{
853	PSUPDRVBUNDLE pToFree;
854	unsigned i;
855
856	/*
857	* Check and unlock all entries in the bundle.
858	*/
859	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
860	{
861	if (pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ)
862	{
863	Log2(("eType=%d pvR0=%p pvR3=%p cb=%ld\n", pBundle->aMem[i].eType, RTR0MemObjAddress(pBundle->aMem[i].MemObj),
864	(void *)RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3), (long)RTR0MemObjSize(pBundle->aMem[i].MemObj)));
865	if (pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ)
866	{
867	rc = RTR0MemObjFree(pBundle->aMem[i].MapObjR3, false);
868	AssertRC(rc); /** @todo figure out how to handle this. */
869	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
870	}
871	rc = RTR0MemObjFree(pBundle->aMem[i].MemObj, true /* fFreeMappings */);
872	AssertRC(rc); /** @todo figure out how to handle this. */
873	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
874	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
875	}
876	}
877
878	/*
879	* Advance and free previous bundle.
880	*/
881	pToFree = pBundle;
882	pBundle = pBundle->pNext;
883
884	pToFree->pNext = NULL;
885	pToFree->cUsed = 0;
886	if (pToFree != &pSession->Bundle)
887	RTMemFree(pToFree);
888	}
889	Log2(("freeing memory - done\n"));
890
891	/*
892	* Deregister component factories.
893	*/
894	RTSemFastMutexRequest(pDevExt->mtxComponentFactory);
895	Log2(("deregistering component factories:\n"));
896	if (pDevExt->pComponentFactoryHead)
897	{
898	PSUPDRVFACTORYREG pPrev = NULL;
899	PSUPDRVFACTORYREG pCur = pDevExt->pComponentFactoryHead;
900	while (pCur)
901	{
902	if (pCur->pSession == pSession)
903	{
904	/* unlink it */
905	PSUPDRVFACTORYREG pNext = pCur->pNext;
906	if (pPrev)
907	pPrev->pNext = pNext;
908	else
909	pDevExt->pComponentFactoryHead = pNext;
910
911	/* free it */
912	pCur->pNext = NULL;
913	pCur->pSession = NULL;
914	pCur->pFactory = NULL;
915	RTMemFree(pCur);
916
917	/* next */
918	pCur = pNext;
919	}
920	else
921	{
922	/* next */
923	pPrev = pCur;
924	pCur = pCur->pNext;
925	}
926	}
927	}
928	RTSemFastMutexRelease(pDevExt->mtxComponentFactory);
929	Log2(("deregistering component factories - done\n"));
930
931	/*
932	* Loaded images needs to be dereferenced and possibly freed up.
933	*/
934	supdrvLdrLock(pDevExt);
935	Log2(("freeing images:\n"));
936	if (pSession->pLdrUsage)
937	{
938	PSUPDRVLDRUSAGE pUsage = pSession->pLdrUsage;
939	pSession->pLdrUsage = NULL;
940	while (pUsage)
941	{
942	void *pvFree = pUsage;
943	PSUPDRVLDRIMAGE pImage = pUsage->pImage;
944	if (pImage->cUsage > pUsage->cUsage)
945	pImage->cUsage -= pUsage->cUsage;
946	else
947	supdrvLdrFree(pDevExt, pImage);
948	pUsage->pImage = NULL;
949	pUsage = pUsage->pNext;
950	RTMemFree(pvFree);
951	}
952	}
953	supdrvLdrUnlock(pDevExt);
954	Log2(("freeing images - done\n"));
955
956	/*
957	* Unmap the GIP.
958	*/
959	Log2(("umapping GIP:\n"));
960	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
961	{
962	SUPR0GipUnmap(pSession);
963	pSession->fGipReferenced = 0;
964	}
965	Log2(("umapping GIP - done\n"));
966	}
967
968
969	/**
970	* RTHandleTableDestroy callback used by supdrvCleanupSession.
971	*
972	* @returns IPRT status code, see SUPR0ObjAddRef.
973	* @param hHandleTable The handle table handle. Ignored.
974	* @param pvObj The object pointer.
975	* @param pvCtx Context, the handle type. Ignored.
976	* @param pvUser Session pointer.
977	*/
978	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser)
979	{
980	NOREF(pvCtx);
981	NOREF(hHandleTable);
982	return SUPR0ObjAddRefEx(pvObj, (PSUPDRVSESSION)pvUser, true /fNoBlocking/);
983	}
984
985
986	/**
987	* RTHandleTableDestroy callback used by supdrvCleanupSession.
988	*
989	* @param hHandleTable The handle table handle. Ignored.
990	* @param h The handle value. Ignored.
991	* @param pvObj The object pointer.
992	* @param pvCtx Context, the handle type. Ignored.
993	* @param pvUser Session pointer.
994	*/
995	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser)
996	{
997	NOREF(pvCtx);
998	NOREF(h);
999	NOREF(hHandleTable);
1000	SUPR0ObjRelease(pvObj, (PSUPDRVSESSION)pvUser);
1001	}
1002
1003
1004	/**
1005	* Fast path I/O Control worker.
1006	*
1007	* @returns VBox status code that should be passed down to ring-3 unchanged.
1008	* @param uIOCtl Function number.
1009	* @param idCpu VMCPU id.
1010	* @param pDevExt Device extention.
1011	* @param pSession Session data.
1012	*/
1013	int VBOXCALL supdrvIOCtlFast(uintptr_t uIOCtl, VMCPUID idCpu, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
1014	{
1015	/*
1016	* We check the two prereqs after doing this only to allow the compiler to optimize things better.
1017	*/
1018	if (RT_LIKELY( RT_VALID_PTR(pSession)
1019	&& pSession->pVM
1020	&& pDevExt->pfnVMMR0EntryFast))
1021	{
1022	switch (uIOCtl)
1023	{
1024	case SUP_IOCTL_FAST_DO_RAW_RUN:
1025	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_RAW_RUN);
1026	break;
1027	case SUP_IOCTL_FAST_DO_HWACC_RUN:
1028	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_HWACC_RUN);
1029	break;
1030	case SUP_IOCTL_FAST_DO_NOP:
1031	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_NOP);
1032	break;
1033	default:
1034	return VERR_INTERNAL_ERROR;
1035	}
1036	return VINF_SUCCESS;
1037	}
1038	return VERR_INTERNAL_ERROR;
1039	}
1040
1041
1042	/**
1043	* Helper for supdrvIOCtl. Check if pszStr contains any character of pszChars.
1044	* We would use strpbrk here if this function would be contained in the RedHat kABI white
1045	* list, see http://www.kerneldrivers.org/RHEL5.
1046	*
1047	* @returns 1 if pszStr does contain any character of pszChars, 0 otherwise.
1048	* @param pszStr String to check
1049	* @param pszChars Character set
1050	*/
1051	static int supdrvCheckInvalidChar(const char pszStr, const char pszChars)
1052	{
1053	int chCur;
1054	while ((chCur = *pszStr++) != '\0')
1055	{
1056	int ch;
1057	const char *psz = pszChars;
1058	while ((ch = *psz++) != '\0')
1059	if (ch == chCur)
1060	return 1;
1061
1062	}
1063	return 0;
1064	}
1065
1066
1067
1068	/**
1069	* I/O Control inner worker (tracing reasons).
1070	*
1071	* @returns IPRT status code.
1072	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1073	*
1074	* @param uIOCtl Function number.
1075	* @param pDevExt Device extention.
1076	* @param pSession Session data.
1077	* @param pReqHdr The request header.
1078	*/
1079	static int supdrvIOCtlInner(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1080	{
1081	/*
1082	* Validation macros
1083	*/
1084	#define REQ_CHECK_SIZES_EX(Name, cbInExpect, cbOutExpect) \
1085	do { \
1086	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect) \|\| pReqHdr->cbOut != (cbOutExpect))) \
1087	{ \
1088	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld. cbOut=%ld expected %ld.\n", \
1089	(long)pReqHdr->cbIn, (long)(cbInExpect), (long)pReqHdr->cbOut, (long)(cbOutExpect))); \
1090	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1091	} \
1092	} while (0)
1093
1094	#define REQ_CHECK_SIZES(Name) REQ_CHECK_SIZES_EX(Name, Name ## _SIZE_IN, Name ## _SIZE_OUT)
1095
1096	#define REQ_CHECK_SIZE_IN(Name, cbInExpect) \
1097	do { \
1098	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect))) \
1099	{ \
1100	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld.\n", \
1101	(long)pReqHdr->cbIn, (long)(cbInExpect))); \
1102	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1103	} \
1104	} while (0)
1105
1106	#define REQ_CHECK_SIZE_OUT(Name, cbOutExpect) \
1107	do { \
1108	if (RT_UNLIKELY(pReqHdr->cbOut != (cbOutExpect))) \
1109	{ \
1110	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbOut=%ld expected %ld.\n", \
1111	(long)pReqHdr->cbOut, (long)(cbOutExpect))); \
1112	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1113	} \
1114	} while (0)
1115
1116	#define REQ_CHECK_EXPR(Name, expr) \
1117	do { \
1118	if (RT_UNLIKELY(!(expr))) \
1119	{ \
1120	OSDBGPRINT(( #Name ": %s\n", #expr)); \
1121	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1122	} \
1123	} while (0)
1124
1125	#define REQ_CHECK_EXPR_FMT(expr, fmt) \
1126	do { \
1127	if (RT_UNLIKELY(!(expr))) \
1128	{ \
1129	OSDBGPRINT( fmt ); \
1130	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1131	} \
1132	} while (0)
1133
1134	/*
1135	* The switch.
1136	*/
1137	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
1138	{
1139	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
1140	{
1141	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
1142	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
1143	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
1144	{
1145	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
1146	pReq->Hdr.rc = VERR_INVALID_MAGIC;
1147	return 0;
1148	}
1149
1150	#if 0
1151	/*
1152	* Call out to the OS specific code and let it do permission checks on the
1153	* client process.
1154	*/
1155	if (!supdrvOSValidateClientProcess(pDevExt, pSession))
1156	{
1157	pReq->u.Out.u32Cookie = 0xffffffff;
1158	pReq->u.Out.u32SessionCookie = 0xffffffff;
1159	pReq->u.Out.u32SessionVersion = 0xffffffff;
1160	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1161	pReq->u.Out.pSession = NULL;
1162	pReq->u.Out.cFunctions = 0;
1163	pReq->Hdr.rc = VERR_PERMISSION_DENIED;
1164	return 0;
1165	}
1166	#endif
1167
1168	/*
1169	* Match the version.
1170	* The current logic is very simple, match the major interface version.
1171	*/
1172	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
1173	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
1174	{
1175	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1176	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
1177	pReq->u.Out.u32Cookie = 0xffffffff;
1178	pReq->u.Out.u32SessionCookie = 0xffffffff;
1179	pReq->u.Out.u32SessionVersion = 0xffffffff;
1180	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1181	pReq->u.Out.pSession = NULL;
1182	pReq->u.Out.cFunctions = 0;
1183	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1184	return 0;
1185	}
1186
1187	/*
1188	* Fill in return data and be gone.
1189	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
1190	* u32SessionVersion <= u32ReqVersion!
1191	*/
1192	/** @todo Somehow validate the client and negotiate a secure cookie... */
1193	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
1194	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
1195	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
1196	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1197	pReq->u.Out.pSession = pSession;
1198	pReq->u.Out.cFunctions = sizeof(g_aFunctions) / sizeof(g_aFunctions[0]);
1199	pReq->Hdr.rc = VINF_SUCCESS;
1200	return 0;
1201	}
1202
1203	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_QUERY_FUNCS(0)):
1204	{
1205	/* validate */
1206	PSUPQUERYFUNCS pReq = (PSUPQUERYFUNCS)pReqHdr;
1207	REQ_CHECK_SIZES_EX(SUP_IOCTL_QUERY_FUNCS, SUP_IOCTL_QUERY_FUNCS_SIZE_IN, SUP_IOCTL_QUERY_FUNCS_SIZE_OUT(RT_ELEMENTS(g_aFunctions)));
1208
1209	/* execute */
1210	pReq->u.Out.cFunctions = RT_ELEMENTS(g_aFunctions);
1211	memcpy(&pReq->u.Out.aFunctions[0], g_aFunctions, sizeof(g_aFunctions));
1212	pReq->Hdr.rc = VINF_SUCCESS;
1213	return 0;
1214	}
1215
1216	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_LOCK):
1217	{
1218	/* validate */
1219	PSUPPAGELOCK pReq = (PSUPPAGELOCK)pReqHdr;
1220	REQ_CHECK_SIZE_IN(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_IN);
1221	REQ_CHECK_SIZE_OUT(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_OUT(pReq->u.In.cPages));
1222	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.cPages > 0);
1223	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.pvR3 >= PAGE_SIZE);
1224
1225	/* execute */
1226	pReq->Hdr.rc = SUPR0LockMem(pSession, pReq->u.In.pvR3, pReq->u.In.cPages, &pReq->u.Out.aPages[0]);
1227	if (RT_FAILURE(pReq->Hdr.rc))
1228	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1229	return 0;
1230	}
1231
1232	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_UNLOCK):
1233	{
1234	/* validate */
1235	PSUPPAGEUNLOCK pReq = (PSUPPAGEUNLOCK)pReqHdr;
1236	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_UNLOCK);
1237
1238	/* execute */
1239	pReq->Hdr.rc = SUPR0UnlockMem(pSession, pReq->u.In.pvR3);
1240	return 0;
1241	}
1242
1243	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_ALLOC):
1244	{
1245	/* validate */
1246	PSUPCONTALLOC pReq = (PSUPCONTALLOC)pReqHdr;
1247	REQ_CHECK_SIZES(SUP_IOCTL_CONT_ALLOC);
1248
1249	/* execute */
1250	pReq->Hdr.rc = SUPR0ContAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.HCPhys);
1251	if (RT_FAILURE(pReq->Hdr.rc))
1252	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1253	return 0;
1254	}
1255
1256	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_FREE):
1257	{
1258	/* validate */
1259	PSUPCONTFREE pReq = (PSUPCONTFREE)pReqHdr;
1260	REQ_CHECK_SIZES(SUP_IOCTL_CONT_FREE);
1261
1262	/* execute */
1263	pReq->Hdr.rc = SUPR0ContFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1264	return 0;
1265	}
1266
1267	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_OPEN):
1268	{
1269	/* validate */
1270	PSUPLDROPEN pReq = (PSUPLDROPEN)pReqHdr;
1271	REQ_CHECK_SIZES(SUP_IOCTL_LDR_OPEN);
1272	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs > 0);
1273	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs < 16*_1M);
1274	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1275	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1276	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits < pReq->u.In.cbImageWithTabs);
1277	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.szName[0]);
1278	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1279	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, !supdrvCheckInvalidChar(pReq->u.In.szName, ";:()[]{}/\\\|&*%#@!~`\"'"));
1280	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szFilename, sizeof(pReq->u.In.szFilename)));
1281
1282	/* execute */
1283	pReq->Hdr.rc = supdrvIOCtl_LdrOpen(pDevExt, pSession, pReq);
1284	return 0;
1285	}
1286
1287	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOAD):
1288	{
1289	/* validate */
1290	PSUPLDRLOAD pReq = (PSUPLDRLOAD)pReqHdr;
1291	REQ_CHECK_EXPR(Name, pReq->Hdr.cbIn >= sizeof(*pReq));
1292	REQ_CHECK_SIZES_EX(SUP_IOCTL_LDR_LOAD, SUP_IOCTL_LDR_LOAD_SIZE_IN(pReq->u.In.cbImageWithTabs), SUP_IOCTL_LDR_LOAD_SIZE_OUT);
1293	REQ_CHECK_EXPR(SUP_IOCTL_LDR_LOAD, pReq->u.In.cSymbols <= 16384);
1294	REQ_CHECK_EXPR_FMT( !pReq->u.In.cSymbols
1295	\|\| ( pReq->u.In.offSymbols < pReq->u.In.cbImageWithTabs
1296	&& pReq->u.In.offSymbols + pReq->u.In.cSymbols * sizeof(SUPLDRSYM) <= pReq->u.In.cbImageWithTabs),
1297	("SUP_IOCTL_LDR_LOAD: offSymbols=%#lx cSymbols=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offSymbols,
1298	(long)pReq->u.In.cSymbols, (long)pReq->u.In.cbImageWithTabs));
1299	REQ_CHECK_EXPR_FMT( !pReq->u.In.cbStrTab
1300	\|\| ( pReq->u.In.offStrTab < pReq->u.In.cbImageWithTabs
1301	&& pReq->u.In.offStrTab + pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs
1302	&& pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs),
1303	("SUP_IOCTL_LDR_LOAD: offStrTab=%#lx cbStrTab=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offStrTab,
1304	(long)pReq->u.In.cbStrTab, (long)pReq->u.In.cbImageWithTabs));
1305
1306	if (pReq->u.In.cSymbols)
1307	{
1308	uint32_t i;
1309	PSUPLDRSYM paSyms = (PSUPLDRSYM)&pReq->u.In.abImage[pReq->u.In.offSymbols];
1310	for (i = 0; i < pReq->u.In.cSymbols; i++)
1311	{
1312	REQ_CHECK_EXPR_FMT(paSyms[i].offSymbol < pReq->u.In.cbImageWithTabs,
1313	("SUP_IOCTL_LDR_LOAD: sym #%ld: symb off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offSymbol, (long)pReq->u.In.cbImageWithTabs));
1314	REQ_CHECK_EXPR_FMT(paSyms[i].offName < pReq->u.In.cbStrTab,
1315	("SUP_IOCTL_LDR_LOAD: sym #%ld: name off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1316	REQ_CHECK_EXPR_FMT(RTStrEnd((char const *)&pReq->u.In.abImage[pReq->u.In.offStrTab + paSyms[i].offName],
1317	pReq->u.In.cbStrTab - paSyms[i].offName),
1318	("SUP_IOCTL_LDR_LOAD: sym #%ld: unterminated name! (%#lx / %#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1319	}
1320	}
1321
1322	/* execute */
1323	pReq->Hdr.rc = supdrvIOCtl_LdrLoad(pDevExt, pSession, pReq);
1324	return 0;
1325	}
1326
1327	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_FREE):
1328	{
1329	/* validate */
1330	PSUPLDRFREE pReq = (PSUPLDRFREE)pReqHdr;
1331	REQ_CHECK_SIZES(SUP_IOCTL_LDR_FREE);
1332
1333	/* execute */
1334	pReq->Hdr.rc = supdrvIOCtl_LdrFree(pDevExt, pSession, pReq);
1335	return 0;
1336	}
1337
1338	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_GET_SYMBOL):
1339	{
1340	/* validate */
1341	PSUPLDRGETSYMBOL pReq = (PSUPLDRGETSYMBOL)pReqHdr;
1342	REQ_CHECK_SIZES(SUP_IOCTL_LDR_GET_SYMBOL);
1343	REQ_CHECK_EXPR(SUP_IOCTL_LDR_GET_SYMBOL, RTStrEnd(pReq->u.In.szSymbol, sizeof(pReq->u.In.szSymbol)));
1344
1345	/* execute */
1346	pReq->Hdr.rc = supdrvIOCtl_LdrGetSymbol(pDevExt, pSession, pReq);
1347	return 0;
1348	}
1349
1350	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0(0)):
1351	{
1352	/* validate */
1353	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1354	Log4(("SUP_IOCTL_CALL_VMMR0: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1355	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1356
1357	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_VMMR0_SIZE(0))
1358	{
1359	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(0), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(0));
1360
1361	/* execute */
1362	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1363	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg, pSession);
1364	else
1365	pReq->Hdr.rc = VERR_WRONG_ORDER;
1366	}
1367	else
1368	{
1369	PSUPVMMR0REQHDR pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1370	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR)),
1371	("SUP_IOCTL_CALL_VMMR0: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR))));
1372	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1373	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(pVMMReq->cbReq));
1374
1375	/* execute */
1376	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1377	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1378	else
1379	pReq->Hdr.rc = VERR_WRONG_ORDER;
1380	}
1381
1382	if ( RT_FAILURE(pReq->Hdr.rc)
1383	&& pReq->Hdr.rc != VERR_INTERRUPTED
1384	&& pReq->Hdr.rc != VERR_TIMEOUT)
1385	Log(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1386	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1387	else
1388	Log4(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1389	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1390	return 0;
1391	}
1392
1393	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0_BIG):
1394	{
1395	/* validate */
1396	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1397	PSUPVMMR0REQHDR pVMMReq;
1398	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1399	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1400
1401	pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1402	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR)),
1403	("SUP_IOCTL_CALL_VMMR0_BIG: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR))));
1404	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0_BIG, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1405	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));
1406
1407	/* execute */
1408	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1409	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1410	else
1411	pReq->Hdr.rc = VERR_WRONG_ORDER;
1412
1413	if ( RT_FAILURE(pReq->Hdr.rc)
1414	&& pReq->Hdr.rc != VERR_INTERRUPTED
1415	&& pReq->Hdr.rc != VERR_TIMEOUT)
1416	Log(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1417	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1418	else
1419	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1420	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1421	return 0;
1422	}
1423
1424	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_PAGING_MODE):
1425	{
1426	/* validate */
1427	PSUPGETPAGINGMODE pReq = (PSUPGETPAGINGMODE)pReqHdr;
1428	REQ_CHECK_SIZES(SUP_IOCTL_GET_PAGING_MODE);
1429
1430	/* execute */
1431	pReq->Hdr.rc = VINF_SUCCESS;
1432	pReq->u.Out.enmMode = SUPR0GetPagingMode();
1433	return 0;
1434	}
1435
1436	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_ALLOC):
1437	{
1438	/* validate */
1439	PSUPLOWALLOC pReq = (PSUPLOWALLOC)pReqHdr;
1440	REQ_CHECK_EXPR(SUP_IOCTL_LOW_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_LOW_ALLOC_SIZE_IN);
1441	REQ_CHECK_SIZES_EX(SUP_IOCTL_LOW_ALLOC, SUP_IOCTL_LOW_ALLOC_SIZE_IN, SUP_IOCTL_LOW_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1442
1443	/* execute */
1444	pReq->Hdr.rc = SUPR0LowAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1445	if (RT_FAILURE(pReq->Hdr.rc))
1446	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1447	return 0;
1448	}
1449
1450	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_FREE):
1451	{
1452	/* validate */
1453	PSUPLOWFREE pReq = (PSUPLOWFREE)pReqHdr;
1454	REQ_CHECK_SIZES(SUP_IOCTL_LOW_FREE);
1455
1456	/* execute */
1457	pReq->Hdr.rc = SUPR0LowFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1458	return 0;
1459	}
1460
1461	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_MAP):
1462	{
1463	/* validate */
1464	PSUPGIPMAP pReq = (PSUPGIPMAP)pReqHdr;
1465	REQ_CHECK_SIZES(SUP_IOCTL_GIP_MAP);
1466
1467	/* execute */
1468	pReq->Hdr.rc = SUPR0GipMap(pSession, &pReq->u.Out.pGipR3, &pReq->u.Out.HCPhysGip);
1469	if (RT_SUCCESS(pReq->Hdr.rc))
1470	pReq->u.Out.pGipR0 = pDevExt->pGip;
1471	return 0;
1472	}
1473
1474	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_UNMAP):
1475	{
1476	/* validate */
1477	PSUPGIPUNMAP pReq = (PSUPGIPUNMAP)pReqHdr;
1478	REQ_CHECK_SIZES(SUP_IOCTL_GIP_UNMAP);
1479
1480	/* execute */
1481	pReq->Hdr.rc = SUPR0GipUnmap(pSession);
1482	return 0;
1483	}
1484
1485	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SET_VM_FOR_FAST):
1486	{
1487	/* validate */
1488	PSUPSETVMFORFAST pReq = (PSUPSETVMFORFAST)pReqHdr;
1489	REQ_CHECK_SIZES(SUP_IOCTL_SET_VM_FOR_FAST);
1490	REQ_CHECK_EXPR_FMT( !pReq->u.In.pVMR0
1491	\|\| ( VALID_PTR(pReq->u.In.pVMR0)
1492	&& !((uintptr_t)pReq->u.In.pVMR0 & (PAGE_SIZE - 1))),
1493	("SUP_IOCTL_SET_VM_FOR_FAST: pVMR0=%p!\n", pReq->u.In.pVMR0));
1494	/* execute */
1495	pSession->pVM = pReq->u.In.pVMR0;
1496	pReq->Hdr.rc = VINF_SUCCESS;
1497	return 0;
1498	}
1499
1500	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_ALLOC_EX):
1501	{
1502	/* validate */
1503	PSUPPAGEALLOCEX pReq = (PSUPPAGEALLOCEX)pReqHdr;
1504	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC_EX, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_EX_SIZE_IN);
1505	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));
1506	REQ_CHECK_EXPR_FMT(pReq->u.In.fKernelMapping \|\| pReq->u.In.fUserMapping,
1507	("SUP_IOCTL_PAGE_ALLOC_EX: No mapping requested!\n"));
1508	REQ_CHECK_EXPR_FMT(pReq->u.In.fUserMapping,
1509	("SUP_IOCTL_PAGE_ALLOC_EX: Must have user mapping!\n"));
1510	REQ_CHECK_EXPR_FMT(!pReq->u.In.fReserved0 && !pReq->u.In.fReserved1,
1511	("SUP_IOCTL_PAGE_ALLOC_EX: fReserved0=%d fReserved1=%d\n", pReq->u.In.fReserved0, pReq->u.In.fReserved1));
1512
1513	/* execute */
1514	pReq->Hdr.rc = SUPR0PageAllocEx(pSession, pReq->u.In.cPages, 0 /* fFlags */,
1515	pReq->u.In.fUserMapping ? &pReq->u.Out.pvR3 : NULL,
1516	pReq->u.In.fKernelMapping ? &pReq->u.Out.pvR0 : NULL,
1517	&pReq->u.Out.aPages[0]);
1518	if (RT_FAILURE(pReq->Hdr.rc))
1519	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1520	return 0;
1521	}
1522
1523	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_MAP_KERNEL):
1524	{
1525	/* validate */
1526	PSUPPAGEMAPKERNEL pReq = (PSUPPAGEMAPKERNEL)pReqHdr;
1527	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_MAP_KERNEL);
1528	REQ_CHECK_EXPR_FMT(!pReq->u.In.fFlags, ("SUP_IOCTL_PAGE_MAP_KERNEL: fFlags=%#x! MBZ\n", pReq->u.In.fFlags));
1529	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_MAP_KERNEL: offSub=%#x\n", pReq->u.In.offSub));
1530	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1531	("SUP_IOCTL_PAGE_MAP_KERNEL: cbSub=%#x\n", pReq->u.In.cbSub));
1532
1533	/* execute */
1534	pReq->Hdr.rc = SUPR0PageMapKernel(pSession, pReq->u.In.pvR3, pReq->u.In.offSub, pReq->u.In.cbSub,
1535	pReq->u.In.fFlags, &pReq->u.Out.pvR0);
1536	if (RT_FAILURE(pReq->Hdr.rc))
1537	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1538	return 0;
1539	}
1540
1541	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_PROTECT):
1542	{
1543	/* validate */
1544	PSUPPAGEPROTECT pReq = (PSUPPAGEPROTECT)pReqHdr;
1545	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_PROTECT);
1546	REQ_CHECK_EXPR_FMT(!(pReq->u.In.fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)),
1547	("SUP_IOCTL_PAGE_PROTECT: fProt=%#x!\n", pReq->u.In.fProt));
1548	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_PROTECT: offSub=%#x\n", pReq->u.In.offSub));
1549	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1550	("SUP_IOCTL_PAGE_PROTECT: cbSub=%#x\n", pReq->u.In.cbSub));
1551
1552	/* execute */
1553	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);
1554	return 0;
1555	}
1556
1557	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_FREE):
1558	{
1559	/* validate */
1560	PSUPPAGEFREE pReq = (PSUPPAGEFREE)pReqHdr;
1561	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_FREE);
1562
1563	/* execute */
1564	pReq->Hdr.rc = SUPR0PageFree(pSession, pReq->u.In.pvR3);
1565	return 0;
1566	}
1567
1568	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_SERVICE(0)):
1569	{
1570	/* validate */
1571	PSUPCALLSERVICE pReq = (PSUPCALLSERVICE)pReqHdr;
1572	Log4(("SUP_IOCTL_CALL_SERVICE: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1573	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1574
1575	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
1576	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(0), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(0));
1577	else
1578	{
1579	PSUPR0SERVICEREQHDR pSrvReq = (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0];
1580	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR)),
1581	("SUP_IOCTL_CALL_SERVICE: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR))));
1582	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, pSrvReq->u32Magic == SUPR0SERVICEREQHDR_MAGIC);
1583	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(pSrvReq->cbReq), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(pSrvReq->cbReq));
1584	}
1585	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1586
1587	/* execute */
1588	pReq->Hdr.rc = supdrvIOCtl_CallServiceModule(pDevExt, pSession, pReq);
1589	return 0;
1590	}
1591
1592	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOGGER_SETTINGS(0)):
1593	{
1594	/* validate */
1595	PSUPLOGGERSETTINGS pReq = (PSUPLOGGERSETTINGS)pReqHdr;
1596	size_t cbStrTab;
1597	REQ_CHECK_SIZE_OUT(SUP_IOCTL_LOGGER_SETTINGS, SUP_IOCTL_LOGGER_SETTINGS_SIZE_OUT);
1598	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->Hdr.cbIn >= SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(1));
1599	cbStrTab = pReq->Hdr.cbIn - SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(0);
1600	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offGroups < cbStrTab);
1601	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offFlags < cbStrTab);
1602	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offDestination < cbStrTab);
1603	REQ_CHECK_EXPR_FMT(pReq->u.In.szStrings[cbStrTab - 1] == '\0',
1604	("SUP_IOCTL_LOGGER_SETTINGS: cbIn=%#x cbStrTab=%#zx LastChar=%d\n",
1605	pReq->Hdr.cbIn, cbStrTab, pReq->u.In.szStrings[cbStrTab - 1]));
1606	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhich <= SUPLOGGERSETTINGS_WHICH_RELEASE);
1607	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhat <= SUPLOGGERSETTINGS_WHAT_DESTROY);
1608
1609	/* execute */
1610	pReq->Hdr.rc = supdrvIOCtl_LoggerSettings(pDevExt, pSession, pReq);
1611	return 0;
1612	}
1613
1614	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP2):
1615	{
1616	/* validate */
1617	PSUPSEMOP2 pReq = (PSUPSEMOP2)pReqHdr;
1618	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP2, SUP_IOCTL_SEM_OP2_SIZE_IN, SUP_IOCTL_SEM_OP2_SIZE_OUT);
1619	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP2, pReq->u.In.uReserved == 0);
1620
1621	/* execute */
1622	switch (pReq->u.In.uType)
1623	{
1624	case SUP_SEM_TYPE_EVENT:
1625	{
1626	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1627	switch (pReq->u.In.uOp)
1628	{
1629	case SUPSEMOP2_WAIT_MS_REL:
1630	pReq->Hdr.rc = SUPSemEventWaitNoResume(pSession, hEvent, pReq->u.In.uArg.cRelMsTimeout);
1631	break;
1632	case SUPSEMOP2_WAIT_NS_ABS:
1633	pReq->Hdr.rc = SUPSemEventWaitNsAbsIntr(pSession, hEvent, pReq->u.In.uArg.uAbsNsTimeout);
1634	break;
1635	case SUPSEMOP2_WAIT_NS_REL:
1636	pReq->Hdr.rc = SUPSemEventWaitNsRelIntr(pSession, hEvent, pReq->u.In.uArg.cRelNsTimeout);
1637	break;
1638	case SUPSEMOP2_SIGNAL:
1639	pReq->Hdr.rc = SUPSemEventSignal(pSession, hEvent);
1640	break;
1641	case SUPSEMOP2_CLOSE:
1642	pReq->Hdr.rc = SUPSemEventClose(pSession, hEvent);
1643	break;
1644	case SUPSEMOP2_RESET:
1645	default:
1646	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1647	break;
1648	}
1649	break;
1650	}
1651
1652	case SUP_SEM_TYPE_EVENT_MULTI:
1653	{
1654	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1655	switch (pReq->u.In.uOp)
1656	{
1657	case SUPSEMOP2_WAIT_MS_REL:
1658	pReq->Hdr.rc = SUPSemEventMultiWaitNoResume(pSession, hEventMulti, pReq->u.In.uArg.cRelMsTimeout);
1659	break;
1660	case SUPSEMOP2_WAIT_NS_ABS:
1661	pReq->Hdr.rc = SUPSemEventMultiWaitNsAbsIntr(pSession, hEventMulti, pReq->u.In.uArg.uAbsNsTimeout);
1662	break;
1663	case SUPSEMOP2_WAIT_NS_REL:
1664	pReq->Hdr.rc = SUPSemEventMultiWaitNsRelIntr(pSession, hEventMulti, pReq->u.In.uArg.cRelNsTimeout);
1665	break;
1666	case SUPSEMOP2_SIGNAL:
1667	pReq->Hdr.rc = SUPSemEventMultiSignal(pSession, hEventMulti);
1668	break;
1669	case SUPSEMOP2_CLOSE:
1670	pReq->Hdr.rc = SUPSemEventMultiClose(pSession, hEventMulti);
1671	break;
1672	case SUPSEMOP2_RESET:
1673	pReq->Hdr.rc = SUPSemEventMultiReset(pSession, hEventMulti);
1674	break;
1675	default:
1676	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1677	break;
1678	}
1679	break;
1680	}
1681
1682	default:
1683	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1684	break;
1685	}
1686	return 0;
1687	}
1688
1689	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP3):
1690	{
1691	/* validate */
1692	PSUPSEMOP3 pReq = (PSUPSEMOP3)pReqHdr;
1693	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP3, SUP_IOCTL_SEM_OP3_SIZE_IN, SUP_IOCTL_SEM_OP3_SIZE_OUT);
1694	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, pReq->u.In.u32Reserved == 0 && pReq->u.In.u64Reserved == 0);
1695
1696	/* execute */
1697	switch (pReq->u.In.uType)
1698	{
1699	case SUP_SEM_TYPE_EVENT:
1700	{
1701	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1702	switch (pReq->u.In.uOp)
1703	{
1704	case SUPSEMOP3_CREATE:
1705	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1706	pReq->Hdr.rc = SUPSemEventCreate(pSession, &hEvent);
1707	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEvent;
1708	break;
1709	case SUPSEMOP3_GET_RESOLUTION:
1710	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1711	pReq->Hdr.rc = VINF_SUCCESS;
1712	pReq->Hdr.cbOut = sizeof(*pReq);
1713	pReq->u.Out.cNsResolution = SUPSemEventGetResolution(pSession);
1714	break;
1715	default:
1716	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1717	break;
1718	}
1719	break;
1720	}
1721
1722	case SUP_SEM_TYPE_EVENT_MULTI:
1723	{
1724	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1725	switch (pReq->u.In.uOp)
1726	{
1727	case SUPSEMOP3_CREATE:
1728	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1729	pReq->Hdr.rc = SUPSemEventMultiCreate(pSession, &hEventMulti);
1730	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEventMulti;
1731	break;
1732	case SUPSEMOP3_GET_RESOLUTION:
1733	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1734	pReq->Hdr.rc = VINF_SUCCESS;
1735	pReq->u.Out.cNsResolution = SUPSemEventMultiGetResolution(pSession);
1736	break;
1737	default:
1738	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1739	break;
1740	}
1741	break;
1742	}
1743
1744	default:
1745	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1746	break;
1747	}
1748	return 0;
1749	}
1750
1751	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
1752	{
1753	/* validate */
1754	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
1755	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
1756	REQ_CHECK_EXPR(SUP_IOCTL_VT_CAPS, pReq->Hdr.cbIn <= SUP_IOCTL_VT_CAPS_SIZE_IN);
1757
1758	/* execute */
1759	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.Caps);
1760	if (RT_FAILURE(pReq->Hdr.rc))
1761	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1762	return 0;
1763	}
1764
1765	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_OPEN):
1766	{
1767	/* validate */
1768	PSUPTRACEROPEN pReq = (PSUPTRACEROPEN)pReqHdr;
1769	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_OPEN);
1770
1771	/* execute */
1772	pReq->Hdr.rc = supdrvIOCtl_TracerOpen(pDevExt, pSession, pReq->u.In.uCookie, pReq->u.In.uArg);
1773	return 0;
1774	}
1775
1776	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_CLOSE):
1777	{
1778	/* validate */
1779	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_CLOSE);
1780
1781	/* execute */
1782	pReqHdr->rc = supdrvIOCtl_TracerClose(pDevExt, pSession);
1783	return 0;
1784	}
1785
1786	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_IOCTL):
1787	{
1788	/* validate */
1789	PSUPTRACERIOCTL pReq = (PSUPTRACERIOCTL)pReqHdr;
1790	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_IOCTL);
1791
1792	/* execute */
1793	pReqHdr->rc = supdrvIOCtl_TracerIOCtl(pDevExt, pSession, pReq->u.In.uCmd, pReq->u.In.uArg, &pReq->u.Out.iRetVal);
1794	return 0;
1795	}
1796
1797	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_REG):
1798	{
1799	/* validate */
1800	PSUPTRACERUMODREG pReq = (PSUPTRACERUMODREG)pReqHdr;
1801	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_REG);
1802	if (!RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)))
1803	return VERR_INVALID_PARAMETER;
1804
1805	/* execute */
1806	pReqHdr->rc = supdrvIOCtl_TracerUmodRegister(pDevExt, pSession,
1807	pReq->u.In.R3PtrVtgHdr, pReq->u.In.uVtgHdrAddr,
1808	pReq->u.In.R3PtrStrTab, pReq->u.In.cbStrTab,
1809	pReq->u.In.szName, pReq->u.In.fFlags);
1810	return 0;
1811	}
1812
1813	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_DEREG):
1814	{
1815	/* validate */
1816	PSUPTRACERUMODDEREG pReq = (PSUPTRACERUMODDEREG)pReqHdr;
1817	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_DEREG);
1818
1819	/* execute */
1820	pReqHdr->rc = supdrvIOCtl_TracerUmodDeregister(pDevExt, pSession, pReq->u.In.pVtgHdr);
1821	return 0;
1822	}
1823
1824	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_FIRE_PROBE):
1825	{
1826	/* validate */
1827	PSUPTRACERUMODFIREPROBE pReq = (PSUPTRACERUMODFIREPROBE)pReqHdr;
1828	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_FIRE_PROBE);
1829
1830	supdrvIOCtl_TracerUmodProbeFire(pDevExt, pSession, &pReq->u.In);
1831	pReqHdr->rc = VINF_SUCCESS;
1832	return 0;
1833	}
1834
1835	default:
1836	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1837	break;
1838	}
1839	return VERR_GENERAL_FAILURE;
1840	}
1841
1842
1843	/**
1844	* I/O Control worker.
1845	*
1846	* @returns IPRT status code.
1847	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1848	*
1849	* @param uIOCtl Function number.
1850	* @param pDevExt Device extention.
1851	* @param pSession Session data.
1852	* @param pReqHdr The request header.
1853	*/
1854	int VBOXCALL supdrvIOCtl(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1855	{
1856	int rc;
1857	VBOXDRV_IOCTL_ENTRY(pSession, uIOCtl, pReqHdr);
1858
1859	/*
1860	* Validate the request.
1861	*/
1862	/* this first check could probably be omitted as its also done by the OS specific code... */
1863	if (RT_UNLIKELY( (pReqHdr->fFlags & SUPREQHDR_FLAGS_MAGIC_MASK) != SUPREQHDR_FLAGS_MAGIC
1864	\|\| pReqHdr->cbIn < sizeof(*pReqHdr)
1865	\|\| pReqHdr->cbOut < sizeof(*pReqHdr)))
1866	{
1867	OSDBGPRINT(("vboxdrv: Bad ioctl request header; cbIn=%#lx cbOut=%#lx fFlags=%#lx\n",
1868	(long)pReqHdr->cbIn, (long)pReqHdr->cbOut, (long)pReqHdr->fFlags));
1869	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1870	return VERR_INVALID_PARAMETER;
1871	}
1872	if (RT_UNLIKELY(!RT_VALID_PTR(pSession)))
1873	{
1874	OSDBGPRINT(("vboxdrv: Invalid pSession valud %p (ioctl=%p)\n", pSession, (void *)uIOCtl));
1875	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1876	return VERR_INVALID_PARAMETER;
1877	}
1878	if (RT_UNLIKELY(uIOCtl == SUP_IOCTL_COOKIE))
1879	{
1880	if (pReqHdr->u32Cookie != SUPCOOKIE_INITIAL_COOKIE)
1881	{
1882	OSDBGPRINT(("SUP_IOCTL_COOKIE: bad cookie %#lx\n", (long)pReqHdr->u32Cookie));
1883	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1884	return VERR_INVALID_PARAMETER;
1885	}
1886	}
1887	else if (RT_UNLIKELY( pReqHdr->u32Cookie != pDevExt->u32Cookie
1888	\|\| pReqHdr->u32SessionCookie != pSession->u32Cookie))
1889	{
1890	OSDBGPRINT(("vboxdrv: bad cookie %#lx / %#lx.\n", (long)pReqHdr->u32Cookie, (long)pReqHdr->u32SessionCookie));
1891	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1892	return VERR_INVALID_PARAMETER;
1893	}
1894
1895	/*
1896	* Hand it to an inner function to avoid lots of unnecessary return tracepoints
1897	*/
1898	rc = supdrvIOCtlInner(uIOCtl, pDevExt, pSession, pReqHdr);
1899
1900	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, pReqHdr->rc, rc);
1901	return rc;
1902	}
1903
1904
1905	/**
1906	* Inter-Driver Communication (IDC) worker.
1907	*
1908	* @returns VBox status code.
1909	* @retval VINF_SUCCESS on success.
1910	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1911	* @retval VERR_NOT_SUPPORTED if the request isn't supported.
1912	*
1913	* @param uReq The request (function) code.
1914	* @param pDevExt Device extention.
1915	* @param pSession Session data.
1916	* @param pReqHdr The request header.
1917	*/
1918	int VBOXCALL supdrvIDC(uintptr_t uReq, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQHDR pReqHdr)
1919	{
1920	/*
1921	* The OS specific code has already validated the pSession
1922	* pointer, and the request size being greater or equal to
1923	* size of the header.
1924	*
1925	* So, just check that pSession is a kernel context session.
1926	*/
1927	if (RT_UNLIKELY( pSession
1928	&& pSession->R0Process != NIL_RTR0PROCESS))
1929	return VERR_INVALID_PARAMETER;
1930
1931	/*
1932	* Validation macro.
1933	*/
1934	#define REQ_CHECK_IDC_SIZE(Name, cbExpect) \
1935	do { \
1936	if (RT_UNLIKELY(pReqHdr->cb != (cbExpect))) \
1937	{ \
1938	OSDBGPRINT(( #Name ": Invalid input/output sizes. cb=%ld expected %ld.\n", \
1939	(long)pReqHdr->cb, (long)(cbExpect))); \
1940	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1941	} \
1942	} while (0)
1943
1944	switch (uReq)
1945	{
1946	case SUPDRV_IDC_REQ_CONNECT:
1947	{
1948	PSUPDRVIDCREQCONNECT pReq = (PSUPDRVIDCREQCONNECT)pReqHdr;
1949	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_CONNECT, sizeof(*pReq));
1950
1951	/*
1952	* Validate the cookie and other input.
1953	*/
1954	if (pReq->Hdr.pSession != NULL)
1955	{
1956	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Hdr.pSession=%p expected NULL!\n", pReq->Hdr.pSession));
1957	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1958	}
1959	if (pReq->u.In.u32MagicCookie != SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE)
1960	{
1961	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: u32MagicCookie=%#x expected %#x!\n",
1962	(unsigned)pReq->u.In.u32MagicCookie, (unsigned)SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE));
1963	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1964	}
1965	if ( pReq->u.In.uMinVersion > pReq->u.In.uReqVersion
1966	\|\| (pReq->u.In.uMinVersion & UINT32_C(0xffff0000)) != (pReq->u.In.uReqVersion & UINT32_C(0xffff0000)))
1967	{
1968	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: uMinVersion=%#x uMaxVersion=%#x doesn't match!\n",
1969	pReq->u.In.uMinVersion, pReq->u.In.uReqVersion));
1970	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1971	}
1972	if (pSession != NULL)
1973	{
1974	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: pSession=%p expected NULL!\n", pSession));
1975	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1976	}
1977
1978	/*
1979	* Match the version.
1980	* The current logic is very simple, match the major interface version.
1981	*/
1982	if ( pReq->u.In.uMinVersion > SUPDRV_IDC_VERSION
1983	\|\| (pReq->u.In.uMinVersion & 0xffff0000) != (SUPDRV_IDC_VERSION & 0xffff0000))
1984	{
1985	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1986	pReq->u.In.uReqVersion, pReq->u.In.uMinVersion, (unsigned)SUPDRV_IDC_VERSION));
1987	pReq->u.Out.pSession = NULL;
1988	pReq->u.Out.uSessionVersion = 0xffffffff;
1989	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1990	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1991	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1992	return VINF_SUCCESS;
1993	}
1994
1995	pReq->u.Out.pSession = NULL;
1996	pReq->u.Out.uSessionVersion = SUPDRV_IDC_VERSION;
1997	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1998	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1999
2000	pReq->Hdr.rc = supdrvCreateSession(pDevExt, false /* fUser */, &pSession);
2001	if (RT_FAILURE(pReq->Hdr.rc))
2002	{
2003	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: failed to create session, rc=%d\n", pReq->Hdr.rc));
2004	return VINF_SUCCESS;
2005	}
2006
2007	pReq->u.Out.pSession = pSession;
2008	pReq->Hdr.pSession = pSession;
2009
2010	return VINF_SUCCESS;
2011	}
2012
2013	case SUPDRV_IDC_REQ_DISCONNECT:
2014	{
2015	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_DISCONNECT, sizeof(*pReqHdr));
2016
2017	supdrvCloseSession(pDevExt, pSession);
2018	return pReqHdr->rc = VINF_SUCCESS;
2019	}
2020
2021	case SUPDRV_IDC_REQ_GET_SYMBOL:
2022	{
2023	PSUPDRVIDCREQGETSYM pReq = (PSUPDRVIDCREQGETSYM)pReqHdr;
2024	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_GET_SYMBOL, sizeof(*pReq));
2025
2026	pReq->Hdr.rc = supdrvIDC_LdrGetSymbol(pDevExt, pSession, pReq);
2027	return VINF_SUCCESS;
2028	}
2029
2030	case SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY:
2031	{
2032	PSUPDRVIDCREQCOMPREGFACTORY pReq = (PSUPDRVIDCREQCOMPREGFACTORY)pReqHdr;
2033	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY, sizeof(*pReq));
2034
2035	pReq->Hdr.rc = SUPR0ComponentRegisterFactory(pSession, pReq->u.In.pFactory);
2036	return VINF_SUCCESS;
2037	}
2038
2039	case SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY:
2040	{
2041	PSUPDRVIDCREQCOMPDEREGFACTORY pReq = (PSUPDRVIDCREQCOMPDEREGFACTORY)pReqHdr;
2042	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY, sizeof(*pReq));
2043
2044	pReq->Hdr.rc = SUPR0ComponentDeregisterFactory(pSession, pReq->u.In.pFactory);
2045	return VINF_SUCCESS;
2046	}
2047
2048	default:
2049	Log(("Unknown IDC %#lx\n", (long)uReq));
2050	break;
2051	}
2052
2053	#undef REQ_CHECK_IDC_SIZE
2054	return VERR_NOT_SUPPORTED;
2055	}
2056
2057
2058	/**
2059	* Register a object for reference counting.
2060	* The object is registered with one reference in the specified session.
2061	*
2062	* @returns Unique identifier on success (pointer).
2063	* All future reference must use this identifier.
2064	* @returns NULL on failure.
2065	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
2066	* @param pvUser1 The first user argument.
2067	* @param pvUser2 The second user argument.
2068	*/
2069	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
2070	{
2071	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2072	PSUPDRVOBJ pObj;
2073	PSUPDRVUSAGE pUsage;
2074
2075	/*
2076	* Validate the input.
2077	*/
2078	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
2079	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
2080	AssertPtrReturn(pfnDestructor, NULL);
2081
2082	/*
2083	* Allocate and initialize the object.
2084	*/
2085	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
2086	if (!pObj)
2087	return NULL;
2088	pObj->u32Magic = SUPDRVOBJ_MAGIC;
2089	pObj->enmType = enmType;
2090	pObj->pNext = NULL;
2091	pObj->cUsage = 1;
2092	pObj->pfnDestructor = pfnDestructor;
2093	pObj->pvUser1 = pvUser1;
2094	pObj->pvUser2 = pvUser2;
2095	pObj->CreatorUid = pSession->Uid;
2096	pObj->CreatorGid = pSession->Gid;
2097	pObj->CreatorProcess= pSession->Process;
2098	supdrvOSObjInitCreator(pObj, pSession);
2099
2100	/*
2101	* Allocate the usage record.
2102	* (We keep freed usage records around to simplify SUPR0ObjAddRefEx().)
2103	*/
2104	RTSpinlockAcquire(pDevExt->Spinlock);
2105
2106	pUsage = pDevExt->pUsageFree;
2107	if (pUsage)
2108	pDevExt->pUsageFree = pUsage->pNext;
2109	else
2110	{
2111	RTSpinlockRelease(pDevExt->Spinlock);
2112	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
2113	if (!pUsage)
2114	{
2115	RTMemFree(pObj);
2116	return NULL;
2117	}
2118	RTSpinlockAcquire(pDevExt->Spinlock);
2119	}
2120
2121	/*
2122	* Insert the object and create the session usage record.
2123	*/
2124	/* The object. */
2125	pObj->pNext = pDevExt->pObjs;
2126	pDevExt->pObjs = pObj;
2127
2128	/* The session record. */
2129	pUsage->cUsage = 1;
2130	pUsage->pObj = pObj;
2131	pUsage->pNext = pSession->pUsage;
2132	/* Log2(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext)); */
2133	pSession->pUsage = pUsage;
2134
2135	RTSpinlockRelease(pDevExt->Spinlock);
2136
2137	Log(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
2138	return pObj;
2139	}
2140
2141
2142	/**
2143	* Increment the reference counter for the object associating the reference
2144	* with the specified session.
2145	*
2146	* @returns IPRT status code.
2147	* @param pvObj The identifier returned by SUPR0ObjRegister().
2148	* @param pSession The session which is referencing the object.
2149	*
2150	* @remarks The caller should not own any spinlocks and must carefully protect
2151	* itself against potential race with the destructor so freed memory
2152	* isn't accessed here.
2153	*/
2154	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
2155	{
2156	return SUPR0ObjAddRefEx(pvObj, pSession, false /* fNoBlocking */);
2157	}
2158
2159
2160	/**
2161	* Increment the reference counter for the object associating the reference
2162	* with the specified session.
2163	*
2164	* @returns IPRT status code.
2165	* @retval VERR_TRY_AGAIN if fNoBlocking was set and a new usage record
2166	* couldn't be allocated. (If you see this you're not doing the right
2167	* thing and it won't ever work reliably.)
2168	*
2169	* @param pvObj The identifier returned by SUPR0ObjRegister().
2170	* @param pSession The session which is referencing the object.
2171	* @param fNoBlocking Set if it's not OK to block. Never try to make the
2172	* first reference to an object in a session with this
2173	* argument set.
2174	*
2175	* @remarks The caller should not own any spinlocks and must carefully protect
2176	* itself against potential race with the destructor so freed memory
2177	* isn't accessed here.
2178	*/
2179	SUPR0DECL(int) SUPR0ObjAddRefEx(void *pvObj, PSUPDRVSESSION pSession, bool fNoBlocking)
2180	{
2181	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2182	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2183	int rc = VINF_SUCCESS;
2184	PSUPDRVUSAGE pUsagePre;
2185	PSUPDRVUSAGE pUsage;
2186
2187	/*
2188	* Validate the input.
2189	* Be ready for the destruction race (someone might be stuck in the
2190	* destructor waiting a lock we own).
2191	*/
2192	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2193	AssertPtrReturn(pObj, VERR_INVALID_POINTER);
2194	AssertMsgReturn(pObj->u32Magic == SUPDRVOBJ_MAGIC \|\| pObj->u32Magic == SUPDRVOBJ_MAGIC_DEAD,
2195	("Invalid pvObj=%p magic=%#x (expected %#x or %#x)\n", pvObj, pObj->u32Magic, SUPDRVOBJ_MAGIC, SUPDRVOBJ_MAGIC_DEAD),
2196	VERR_INVALID_PARAMETER);
2197
2198	RTSpinlockAcquire(pDevExt->Spinlock);
2199
2200	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2201	{
2202	RTSpinlockRelease(pDevExt->Spinlock);
2203
2204	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2205	return VERR_WRONG_ORDER;
2206	}
2207
2208	/*
2209	* Preallocate the usage record if we can.
2210	*/
2211	pUsagePre = pDevExt->pUsageFree;
2212	if (pUsagePre)
2213	pDevExt->pUsageFree = pUsagePre->pNext;
2214	else if (!fNoBlocking)
2215	{
2216	RTSpinlockRelease(pDevExt->Spinlock);
2217	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
2218	if (!pUsagePre)
2219	return VERR_NO_MEMORY;
2220
2221	RTSpinlockAcquire(pDevExt->Spinlock);
2222	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2223	{
2224	RTSpinlockRelease(pDevExt->Spinlock);
2225
2226	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2227	return VERR_WRONG_ORDER;
2228	}
2229	}
2230
2231	/*
2232	* Reference the object.
2233	*/
2234	pObj->cUsage++;
2235
2236	/*
2237	* Look for the session record.
2238	*/
2239	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
2240	{
2241	/Log(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2242	if (pUsage->pObj == pObj)
2243	break;
2244	}
2245	if (pUsage)
2246	pUsage->cUsage++;
2247	else if (pUsagePre)
2248	{
2249	/* create a new session record. */
2250	pUsagePre->cUsage = 1;
2251	pUsagePre->pObj = pObj;
2252	pUsagePre->pNext = pSession->pUsage;
2253	pSession->pUsage = pUsagePre;
2254	/Log(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));/
2255
2256	pUsagePre = NULL;
2257	}
2258	else
2259	{
2260	pObj->cUsage--;
2261	rc = VERR_TRY_AGAIN;
2262	}
2263
2264	/*
2265	* Put any unused usage record into the free list..
2266	*/
2267	if (pUsagePre)
2268	{
2269	pUsagePre->pNext = pDevExt->pUsageFree;
2270	pDevExt->pUsageFree = pUsagePre;
2271	}
2272
2273	RTSpinlockRelease(pDevExt->Spinlock);
2274
2275	return rc;
2276	}
2277
2278
2279	/**
2280	* Decrement / destroy a reference counter record for an object.
2281	*
2282	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
2283	*
2284	* @returns IPRT status code.
2285	* @retval VINF_SUCCESS if not destroyed.
2286	* @retval VINF_OBJECT_DESTROYED if it's destroyed by this release call.
2287	* @retval VERR_INVALID_PARAMETER if the object isn't valid. Will assert in
2288	* string builds.
2289	*
2290	* @param pvObj The identifier returned by SUPR0ObjRegister().
2291	* @param pSession The session which is referencing the object.
2292	*/
2293	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
2294	{
2295	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2296	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2297	int rc = VERR_INVALID_PARAMETER;
2298	PSUPDRVUSAGE pUsage;
2299	PSUPDRVUSAGE pUsagePrev;
2300
2301	/*
2302	* Validate the input.
2303	*/
2304	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2305	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2306	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2307	VERR_INVALID_PARAMETER);
2308
2309	/*
2310	* Acquire the spinlock and look for the usage record.
2311	*/
2312	RTSpinlockAcquire(pDevExt->Spinlock);
2313
2314	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
2315	pUsage;
2316	pUsagePrev = pUsage, pUsage = pUsage->pNext)
2317	{
2318	/Log2(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2319	if (pUsage->pObj == pObj)
2320	{
2321	rc = VINF_SUCCESS;
2322	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
2323	if (pUsage->cUsage > 1)
2324	{
2325	pObj->cUsage--;
2326	pUsage->cUsage--;
2327	}
2328	else
2329	{
2330	/*
2331	* Free the session record.
2332	*/
2333	if (pUsagePrev)
2334	pUsagePrev->pNext = pUsage->pNext;
2335	else
2336	pSession->pUsage = pUsage->pNext;
2337	pUsage->pNext = pDevExt->pUsageFree;
2338	pDevExt->pUsageFree = pUsage;
2339
2340	/* What about the object? */
2341	if (pObj->cUsage > 1)
2342	pObj->cUsage--;
2343	else
2344	{
2345	/*
2346	* Object is to be destroyed, unlink it.
2347	*/
2348	pObj->u32Magic = SUPDRVOBJ_MAGIC_DEAD;
2349	rc = VINF_OBJECT_DESTROYED;
2350	if (pDevExt->pObjs == pObj)
2351	pDevExt->pObjs = pObj->pNext;
2352	else
2353	{
2354	PSUPDRVOBJ pObjPrev;
2355	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
2356	if (pObjPrev->pNext == pObj)
2357	{
2358	pObjPrev->pNext = pObj->pNext;
2359	break;
2360	}
2361	Assert(pObjPrev);
2362	}
2363	}
2364	}
2365	break;
2366	}
2367	}
2368
2369	RTSpinlockRelease(pDevExt->Spinlock);
2370
2371	/*
2372	* Call the destructor and free the object if required.
2373	*/
2374	if (rc == VINF_OBJECT_DESTROYED)
2375	{
2376	Log(("SUPR0ObjRelease: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
2377	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
2378	if (pObj->pfnDestructor)
2379	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
2380	RTMemFree(pObj);
2381	}
2382
2383	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
2384	return rc;
2385	}
2386
2387
2388	/**
2389	* Verifies that the current process can access the specified object.
2390	*
2391	* @returns The following IPRT status code:
2392	* @retval VINF_SUCCESS if access was granted.
2393	* @retval VERR_PERMISSION_DENIED if denied access.
2394	* @retval VERR_INVALID_PARAMETER if invalid parameter.
2395	*
2396	* @param pvObj The identifier returned by SUPR0ObjRegister().
2397	* @param pSession The session which wishes to access the object.
2398	* @param pszObjName Object string name. This is optional and depends on the object type.
2399	*
2400	* @remark The caller is responsible for making sure the object isn't removed while
2401	* we're inside this function. If uncertain about this, just call AddRef before calling us.
2402	*/
2403	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
2404	{
2405	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2406	int rc;
2407
2408	/*
2409	* Validate the input.
2410	*/
2411	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2412	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2413	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2414	VERR_INVALID_PARAMETER);
2415
2416	/*
2417	* Check access. (returns true if a decision has been made.)
2418	*/
2419	rc = VERR_INTERNAL_ERROR;
2420	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
2421	return rc;
2422
2423	/*
2424	* Default policy is to allow the user to access his own
2425	* stuff but nothing else.
2426	*/
2427	if (pObj->CreatorUid == pSession->Uid)
2428	return VINF_SUCCESS;
2429	return VERR_PERMISSION_DENIED;
2430	}
2431
2432
2433	/**
2434	* Lock pages.
2435	*
2436	* @returns IPRT status code.
2437	* @param pSession Session to which the locked memory should be associated.
2438	* @param pvR3 Start of the memory range to lock.
2439	* This must be page aligned.
2440	* @param cPages Number of pages to lock.
2441	* @param paPages Where to put the physical addresses of locked memory.
2442	*/
2443	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
2444	{
2445	int rc;
2446	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2447	const size_t cb = (size_t)cPages << PAGE_SHIFT;
2448	LogFlow(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
2449
2450	/*
2451	* Verify input.
2452	*/
2453	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2454	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
2455	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
2456	\|\| !pvR3)
2457	{
2458	Log(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
2459	return VERR_INVALID_PARAMETER;
2460	}
2461
2462	/*
2463	* Let IPRT do the job.
2464	*/
2465	Mem.eType = MEMREF_TYPE_LOCKED;
2466	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTMEM_PROT_READ \| RTMEM_PROT_WRITE, RTR0ProcHandleSelf());
2467	if (RT_SUCCESS(rc))
2468	{
2469	uint32_t iPage = cPages;
2470	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
2471	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
2472
2473	while (iPage-- > 0)
2474	{
2475	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2476	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
2477	{
2478	AssertMsgFailed(("iPage=%d\n", iPage));
2479	rc = VERR_INTERNAL_ERROR;
2480	break;
2481	}
2482	}
2483	if (RT_SUCCESS(rc))
2484	rc = supdrvMemAdd(&Mem, pSession);
2485	if (RT_FAILURE(rc))
2486	{
2487	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
2488	AssertRC(rc2);
2489	}
2490	}
2491
2492	return rc;
2493	}
2494
2495
2496	/**
2497	* Unlocks the memory pointed to by pv.
2498	*
2499	* @returns IPRT status code.
2500	* @param pSession Session to which the memory was locked.
2501	* @param pvR3 Memory to unlock.
2502	*/
2503	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2504	{
2505	LogFlow(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2506	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2507	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
2508	}
2509
2510
2511	/**
2512	* Allocates a chunk of page aligned memory with contiguous and fixed physical
2513	* backing.
2514	*
2515	* @returns IPRT status code.
2516	* @param pSession Session data.
2517	* @param cPages Number of pages to allocate.
2518	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
2519	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
2520	* @param pHCPhys Where to put the physical address of allocated memory.
2521	*/
2522	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
2523	{
2524	int rc;
2525	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2526	LogFlow(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
2527
2528	/*
2529	* Validate input.
2530	*/
2531	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2532	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
2533	{
2534	Log(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
2535	pSession, ppvR0, ppvR3, pHCPhys));
2536	return VERR_INVALID_PARAMETER;
2537
2538	}
2539	if (cPages < 1 \|\| cPages >= 256)
2540	{
2541	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2542	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2543	}
2544
2545	/*
2546	* Let IPRT do the job.
2547	*/
2548	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
2549	if (RT_SUCCESS(rc))
2550	{
2551	int rc2;
2552	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2553	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2554	if (RT_SUCCESS(rc))
2555	{
2556	Mem.eType = MEMREF_TYPE_CONT;
2557	rc = supdrvMemAdd(&Mem, pSession);
2558	if (!rc)
2559	{
2560	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2561	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2562	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
2563	return 0;
2564	}
2565
2566	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2567	AssertRC(rc2);
2568	}
2569	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2570	AssertRC(rc2);
2571	}
2572
2573	return rc;
2574	}
2575
2576
2577	/**
2578	* Frees memory allocated using SUPR0ContAlloc().
2579	*
2580	* @returns IPRT status code.
2581	* @param pSession The session to which the memory was allocated.
2582	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2583	*/
2584	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2585	{
2586	LogFlow(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2587	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2588	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
2589	}
2590
2591
2592	/**
2593	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
2594	*
2595	* The memory isn't zeroed.
2596	*
2597	* @returns IPRT status code.
2598	* @param pSession Session data.
2599	* @param cPages Number of pages to allocate.
2600	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
2601	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
2602	* @param paPages Where to put the physical addresses of allocated memory.
2603	*/
2604	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
2605	{
2606	unsigned iPage;
2607	int rc;
2608	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2609	LogFlow(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
2610
2611	/*
2612	* Validate input.
2613	*/
2614	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2615	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
2616	{
2617	Log(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
2618	pSession, ppvR3, ppvR0, paPages));
2619	return VERR_INVALID_PARAMETER;
2620
2621	}
2622	if (cPages < 1 \|\| cPages >= 256)
2623	{
2624	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2625	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2626	}
2627
2628	/*
2629	* Let IPRT do the work.
2630	*/
2631	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
2632	if (RT_SUCCESS(rc))
2633	{
2634	int rc2;
2635	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2636	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2637	if (RT_SUCCESS(rc))
2638	{
2639	Mem.eType = MEMREF_TYPE_LOW;
2640	rc = supdrvMemAdd(&Mem, pSession);
2641	if (!rc)
2642	{
2643	for (iPage = 0; iPage < cPages; iPage++)
2644	{
2645	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2646	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%RHp\n", paPages[iPage]));
2647	}
2648	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2649	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2650	return 0;
2651	}
2652
2653	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2654	AssertRC(rc2);
2655	}
2656
2657	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2658	AssertRC(rc2);
2659	}
2660
2661	return rc;
2662	}
2663
2664
2665	/**
2666	* Frees memory allocated using SUPR0LowAlloc().
2667	*
2668	* @returns IPRT status code.
2669	* @param pSession The session to which the memory was allocated.
2670	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2671	*/
2672	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2673	{
2674	LogFlow(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2675	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2676	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
2677	}
2678
2679
2680
2681	/**
2682	* Allocates a chunk of memory with both R0 and R3 mappings.
2683	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
2684	*
2685	* @returns IPRT status code.
2686	* @param pSession The session to associated the allocation with.
2687	* @param cb Number of bytes to allocate.
2688	* @param ppvR0 Where to store the address of the Ring-0 mapping.
2689	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2690	*/
2691	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
2692	{
2693	int rc;
2694	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2695	LogFlow(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
2696
2697	/*
2698	* Validate input.
2699	*/
2700	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2701	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
2702	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
2703	if (cb < 1 \|\| cb >= _4M)
2704	{
2705	Log(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
2706	return VERR_INVALID_PARAMETER;
2707	}
2708
2709	/*
2710	* Let IPRT do the work.
2711	*/
2712	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
2713	if (RT_SUCCESS(rc))
2714	{
2715	int rc2;
2716	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2717	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2718	if (RT_SUCCESS(rc))
2719	{
2720	Mem.eType = MEMREF_TYPE_MEM;
2721	rc = supdrvMemAdd(&Mem, pSession);
2722	if (!rc)
2723	{
2724	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2725	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2726	return VINF_SUCCESS;
2727	}
2728
2729	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2730	AssertRC(rc2);
2731	}
2732
2733	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2734	AssertRC(rc2);
2735	}
2736
2737	return rc;
2738	}
2739
2740
2741	/**
2742	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
2743	*
2744	* @returns IPRT status code.
2745	* @param pSession The session to which the memory was allocated.
2746	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2747	* @param paPages Where to store the physical addresses.
2748	*/
2749	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
2750	{
2751	PSUPDRVBUNDLE pBundle;
2752	LogFlow(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
2753
2754	/*
2755	* Validate input.
2756	*/
2757	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2758	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
2759	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
2760
2761	/*
2762	* Search for the address.
2763	*/
2764	RTSpinlockAcquire(pSession->Spinlock);
2765	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2766	{
2767	if (pBundle->cUsed > 0)
2768	{
2769	unsigned i;
2770	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2771	{
2772	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
2773	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2774	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2775	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2776	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
2777	)
2778	)
2779	{
2780	const size_t cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2781	size_t iPage;
2782	for (iPage = 0; iPage < cPages; iPage++)
2783	{
2784	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2785	paPages[iPage].uReserved = 0;
2786	}
2787	RTSpinlockRelease(pSession->Spinlock);
2788	return VINF_SUCCESS;
2789	}
2790	}
2791	}
2792	}
2793	RTSpinlockRelease(pSession->Spinlock);
2794	Log(("Failed to find %p!!!\n", (void *)uPtr));
2795	return VERR_INVALID_PARAMETER;
2796	}
2797
2798
2799	/**
2800	* Free memory allocated by SUPR0MemAlloc().
2801	*
2802	* @returns IPRT status code.
2803	* @param pSession The session owning the allocation.
2804	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2805	*/
2806	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2807	{
2808	LogFlow(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2809	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2810	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
2811	}
2812
2813
2814	/**
2815	* Allocates a chunk of memory with a kernel or/and a user mode mapping.
2816	*
2817	* The memory is fixed and it's possible to query the physical addresses using
2818	* SUPR0MemGetPhys().
2819	*
2820	* @returns IPRT status code.
2821	* @param pSession The session to associated the allocation with.
2822	* @param cPages The number of pages to allocate.
2823	* @param fFlags Flags, reserved for the future. Must be zero.
2824	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2825	* NULL if no ring-3 mapping.
2826	* @param ppvR3 Where to store the address of the Ring-0 mapping.
2827	* NULL if no ring-0 mapping.
2828	* @param paPages Where to store the addresses of the pages. Optional.
2829	*/
2830	SUPR0DECL(int) SUPR0PageAllocEx(PSUPDRVSESSION pSession, uint32_t cPages, uint32_t fFlags, PRTR3PTR ppvR3, PRTR0PTR ppvR0, PRTHCPHYS paPages)
2831	{
2832	int rc;
2833	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2834	LogFlow(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
2835
2836	/*
2837	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2838	*/
2839	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2840	AssertPtrNullReturn(ppvR3, VERR_INVALID_POINTER);
2841	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2842	AssertReturn(ppvR3 \|\| ppvR0, VERR_INVALID_PARAMETER);
2843	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2844	if (cPages < 1 \|\| cPages > VBOX_MAX_ALLOC_PAGE_COUNT)
2845	{
2846	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)));
2847	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2848	}
2849
2850	/*
2851	* Let IPRT do the work.
2852	*/
2853	if (ppvR0)
2854	rc = RTR0MemObjAllocPage(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, true /* fExecutable */);
2855	else
2856	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
2857	if (RT_SUCCESS(rc))
2858	{
2859	int rc2;
2860	if (ppvR3)
2861	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2862	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2863	else
2864	Mem.MapObjR3 = NIL_RTR0MEMOBJ;
2865	if (RT_SUCCESS(rc))
2866	{
2867	Mem.eType = MEMREF_TYPE_PAGE;
2868	rc = supdrvMemAdd(&Mem, pSession);
2869	if (!rc)
2870	{
2871	if (ppvR3)
2872	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2873	if (ppvR0)
2874	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2875	if (paPages)
2876	{
2877	uint32_t iPage = cPages;
2878	while (iPage-- > 0)
2879	{
2880	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
2881	Assert(paPages[iPage] != NIL_RTHCPHYS);
2882	}
2883	}
2884	return VINF_SUCCESS;
2885	}
2886
2887	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2888	AssertRC(rc2);
2889	}
2890
2891	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2892	AssertRC(rc2);
2893	}
2894	return rc;
2895	}
2896
2897
2898	/**
2899	* Maps a chunk of memory previously allocated by SUPR0PageAllocEx into kernel
2900	* space.
2901	*
2902	* @returns IPRT status code.
2903	* @param pSession The session to associated the allocation with.
2904	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
2905	* @param offSub Where to start mapping. Must be page aligned.
2906	* @param cbSub How much to map. Must be page aligned.
2907	* @param fFlags Flags, MBZ.
2908	* @param ppvR0 Where to return the address of the ring-0 mapping on
2909	* success.
2910	*/
2911	SUPR0DECL(int) SUPR0PageMapKernel(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t offSub, uint32_t cbSub,
2912	uint32_t fFlags, PRTR0PTR ppvR0)
2913	{
2914	int rc;
2915	PSUPDRVBUNDLE pBundle;
2916	RTR0MEMOBJ hMemObj = NIL_RTR0MEMOBJ;
2917	LogFlow(("SUPR0PageMapKernel: pSession=%p pvR3=%p offSub=%#x cbSub=%#x\n", pSession, pvR3, offSub, cbSub));
2918
2919	/*
2920	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2921	*/
2922	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2923	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2924	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2925	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2926	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2927	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
2928
2929	/*
2930	* Find the memory object.
2931	*/
2932	RTSpinlockAcquire(pSession->Spinlock);
2933	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2934	{
2935	if (pBundle->cUsed > 0)
2936	{
2937	unsigned i;
2938	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2939	{
2940	if ( ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
2941	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2942	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2943	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
2944	\|\| ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED
2945	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2946	&& pBundle->aMem[i].MapObjR3 == NIL_RTR0MEMOBJ
2947	&& RTR0MemObjAddressR3(pBundle->aMem[i].MemObj) == pvR3))
2948	{
2949	hMemObj = pBundle->aMem[i].MemObj;
2950	break;
2951	}
2952	}
2953	}
2954	}
2955	RTSpinlockRelease(pSession->Spinlock);
2956
2957	rc = VERR_INVALID_PARAMETER;
2958	if (hMemObj != NIL_RTR0MEMOBJ)
2959	{
2960	/*
2961	* Do some further input validations before calling IPRT.
2962	* (Cleanup is done indirectly by telling RTR0MemObjFree to include mappings.)
2963	*/
2964	size_t cbMemObj = RTR0MemObjSize(hMemObj);
2965	if ( offSub < cbMemObj
2966	&& cbSub <= cbMemObj
2967	&& offSub + cbSub <= cbMemObj)
2968	{
2969	RTR0MEMOBJ hMapObj;
2970	rc = RTR0MemObjMapKernelEx(&hMapObj, hMemObj, (void *)-1, 0,
2971	RTMEM_PROT_READ \| RTMEM_PROT_WRITE, offSub, cbSub);
2972	if (RT_SUCCESS(rc))
2973	*ppvR0 = RTR0MemObjAddress(hMapObj);
2974	}
2975	else
2976	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
2977
2978	}
2979	return rc;
2980	}
2981
2982
2983	/**
2984	* Changes the page level protection of one or more pages previously allocated
2985	* by SUPR0PageAllocEx.
2986	*
2987	* @returns IPRT status code.
2988	* @param pSession The session to associated the allocation with.
2989	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
2990	* NIL_RTR3PTR if the ring-3 mapping should be unaffected.
2991	* @param pvR0 The ring-0 address returned by SUPR0PageAllocEx.
2992	* NIL_RTR0PTR if the ring-0 mapping should be unaffected.
2993	* @param offSub Where to start changing. Must be page aligned.
2994	* @param cbSub How much to change. Must be page aligned.
2995	* @param fProt The new page level protection, see RTMEM_PROT_*.
2996	*/
2997	SUPR0DECL(int) SUPR0PageProtect(PSUPDRVSESSION pSession, RTR3PTR pvR3, RTR0PTR pvR0, uint32_t offSub, uint32_t cbSub, uint32_t fProt)
2998	{
2999	int rc;
3000	PSUPDRVBUNDLE pBundle;
3001	RTR0MEMOBJ hMemObjR0 = NIL_RTR0MEMOBJ;
3002	RTR0MEMOBJ hMemObjR3 = NIL_RTR0MEMOBJ;
3003	LogFlow(("SUPR0PageProtect: pSession=%p pvR3=%p pvR0=%p offSub=%#x cbSub=%#x fProt-%#x\n", pSession, pvR3, pvR0, offSub, cbSub, fProt));
3004
3005	/*
3006	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
3007	*/
3008	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3009	AssertReturn(!(fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)), VERR_INVALID_PARAMETER);
3010	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3011	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3012	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
3013
3014	/*
3015	* Find the memory object.
3016	*/
3017	RTSpinlockAcquire(pSession->Spinlock);
3018	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3019	{
3020	if (pBundle->cUsed > 0)
3021	{
3022	unsigned i;
3023	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3024	{
3025	if ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
3026	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3027	&& ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3028	\|\| pvR3 == NIL_RTR3PTR)
3029	&& ( pvR0 == NIL_RTR0PTR
3030	\|\| RTR0MemObjAddress(pBundle->aMem[i].MemObj) == pvR0)
3031	&& ( pvR3 == NIL_RTR3PTR
3032	\|\| RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3))
3033	{
3034	if (pvR0 != NIL_RTR0PTR)
3035	hMemObjR0 = pBundle->aMem[i].MemObj;
3036	if (pvR3 != NIL_RTR3PTR)
3037	hMemObjR3 = pBundle->aMem[i].MapObjR3;
3038	break;
3039	}
3040	}
3041	}
3042	}
3043	RTSpinlockRelease(pSession->Spinlock);
3044
3045	rc = VERR_INVALID_PARAMETER;
3046	if ( hMemObjR0 != NIL_RTR0MEMOBJ
3047	\|\| hMemObjR3 != NIL_RTR0MEMOBJ)
3048	{
3049	/*
3050	* Do some further input validations before calling IPRT.
3051	*/
3052	size_t cbMemObj = hMemObjR0 != NIL_RTR0PTR ? RTR0MemObjSize(hMemObjR0) : RTR0MemObjSize(hMemObjR3);
3053	if ( offSub < cbMemObj
3054	&& cbSub <= cbMemObj
3055	&& offSub + cbSub <= cbMemObj)
3056	{
3057	rc = VINF_SUCCESS;
3058	if (hMemObjR3 != NIL_RTR0PTR)
3059	rc = RTR0MemObjProtect(hMemObjR3, offSub, cbSub, fProt);
3060	if (hMemObjR0 != NIL_RTR0PTR && RT_SUCCESS(rc))
3061	rc = RTR0MemObjProtect(hMemObjR0, offSub, cbSub, fProt);
3062	}
3063	else
3064	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
3065
3066	}
3067	return rc;
3068
3069	}
3070
3071
3072	/**
3073	* Free memory allocated by SUPR0PageAlloc() and SUPR0PageAllocEx().
3074	*
3075	* @returns IPRT status code.
3076	* @param pSession The session owning the allocation.
3077	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc() or
3078	* SUPR0PageAllocEx().
3079	*/
3080	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
3081	{
3082	LogFlow(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
3083	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3084	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_PAGE);
3085	}
3086
3087
3088	/**
3089	* Gets the paging mode of the current CPU.
3090	*
3091	* @returns Paging mode, SUPPAGEINGMODE_INVALID on error.
3092	*/
3093	SUPR0DECL(SUPPAGINGMODE) SUPR0GetPagingMode(void)
3094	{
3095	SUPPAGINGMODE enmMode;
3096
3097	RTR0UINTREG cr0 = ASMGetCR0();
3098	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
3099	enmMode = SUPPAGINGMODE_INVALID;
3100	else
3101	{
3102	RTR0UINTREG cr4 = ASMGetCR4();
3103	uint32_t fNXEPlusLMA = 0;
3104	if (cr4 & X86_CR4_PAE)
3105	{
3106	uint32_t fAmdFeatures = ASMCpuId_EDX(0x80000001);
3107	if (fAmdFeatures & (X86_CPUID_AMD_FEATURE_EDX_NX \| X86_CPUID_AMD_FEATURE_EDX_LONG_MODE))
3108	{
3109	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
3110	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
3111	fNXEPlusLMA \|= RT_BIT(0);
3112	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
3113	fNXEPlusLMA \|= RT_BIT(1);
3114	}
3115	}
3116
3117	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
3118	{
3119	case 0:
3120	enmMode = SUPPAGINGMODE_32_BIT;
3121	break;
3122
3123	case X86_CR4_PGE:
3124	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
3125	break;
3126
3127	case X86_CR4_PAE:
3128	enmMode = SUPPAGINGMODE_PAE;
3129	break;
3130
3131	case X86_CR4_PAE \| RT_BIT(0):
3132	enmMode = SUPPAGINGMODE_PAE_NX;
3133	break;
3134
3135	case X86_CR4_PAE \| X86_CR4_PGE:
3136	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3137	break;
3138
3139	case X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3140	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3141	break;
3142
3143	case RT_BIT(1) \| X86_CR4_PAE:
3144	enmMode = SUPPAGINGMODE_AMD64;
3145	break;
3146
3147	case RT_BIT(1) \| X86_CR4_PAE \| RT_BIT(0):
3148	enmMode = SUPPAGINGMODE_AMD64_NX;
3149	break;
3150
3151	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
3152	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
3153	break;
3154
3155	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3156	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
3157	break;
3158
3159	default:
3160	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
3161	enmMode = SUPPAGINGMODE_INVALID;
3162	break;
3163	}
3164	}
3165	return enmMode;
3166	}
3167
3168
3169	/**
3170	* Enables or disabled hardware virtualization extensions using native OS APIs.
3171	*
3172	* @returns VBox status code.
3173	* @retval VINF_SUCCESS on success.
3174	* @retval VERR_NOT_SUPPORTED if not supported by the native OS.
3175	*
3176	* @param fEnable Whether to enable or disable.
3177	*/
3178	SUPR0DECL(int) SUPR0EnableVTx(bool fEnable)
3179	{
3180	#ifdef RT_OS_DARWIN
3181	return supdrvOSEnableVTx(fEnable);
3182	#else
3183	return VERR_NOT_SUPPORTED;
3184	#endif
3185	}
3186
3187
3188	/** @todo document me */
3189	SUPR0DECL(int) SUPR0QueryVTCaps(PSUPDRVSESSION pSession, uint32_t *pfCaps)
3190	{
3191	/*
3192	* Input validation.
3193	*/
3194	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3195	AssertPtrReturn(pfCaps, VERR_INVALID_POINTER);
3196
3197	*pfCaps = 0;
3198
3199	if (ASMHasCpuId())
3200	{
3201	uint32_t u32FeaturesECX;
3202	uint32_t u32Dummy;
3203	uint32_t u32FeaturesEDX;
3204	uint32_t u32VendorEBX, u32VendorECX, u32VendorEDX, u32AMDFeatureEDX, u32AMDFeatureECX;
3205	uint64_t val;
3206
3207	ASMCpuId(0, &u32Dummy, &u32VendorEBX, &u32VendorECX, &u32VendorEDX);
3208	ASMCpuId(1, &u32Dummy, &u32Dummy, &u32FeaturesECX, &u32FeaturesEDX);
3209	/* Query AMD features. */
3210	ASMCpuId(0x80000001, &u32Dummy, &u32Dummy, &u32AMDFeatureECX, &u32AMDFeatureEDX);
3211
3212	if ( u32VendorEBX == X86_CPUID_VENDOR_INTEL_EBX
3213	&& u32VendorECX == X86_CPUID_VENDOR_INTEL_ECX
3214	&& u32VendorEDX == X86_CPUID_VENDOR_INTEL_EDX
3215	)
3216	{
3217	if ( (u32FeaturesECX & X86_CPUID_FEATURE_ECX_VMX)
3218	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3219	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3220	)
3221	{
3222	val = ASMRdMsr(MSR_IA32_FEATURE_CONTROL);
3223	/*
3224	* Both the LOCK and VMXON bit must be set; otherwise VMXON will generate a #GP.
3225	* Once the lock bit is set, this MSR can no longer be modified.
3226	*/
3227	if ( (val & (MSR_IA32_FEATURE_CONTROL_VMXON\|MSR_IA32_FEATURE_CONTROL_LOCK))
3228	== (MSR_IA32_FEATURE_CONTROL_VMXON\|MSR_IA32_FEATURE_CONTROL_LOCK) /* enabled and locked */
3229	\|\| !(val & MSR_IA32_FEATURE_CONTROL_LOCK) /* not enabled, but not locked either */
3230	)
3231	{
3232	VMX_CAPABILITY vtCaps;
3233
3234	*pfCaps \|= SUPVTCAPS_VT_X;
3235
3236	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS);
3237	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC_USE_SECONDARY_EXEC_CTRL)
3238	{
3239	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS2);
3240	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC2_EPT)
3241	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3242	}
3243	return VINF_SUCCESS;
3244	}
3245	return VERR_VMX_MSR_LOCKED_OR_DISABLED;
3246	}
3247	return VERR_VMX_NO_VMX;
3248	}
3249
3250	if ( u32VendorEBX == X86_CPUID_VENDOR_AMD_EBX
3251	&& u32VendorECX == X86_CPUID_VENDOR_AMD_ECX
3252	&& u32VendorEDX == X86_CPUID_VENDOR_AMD_EDX
3253	)
3254	{
3255	if ( (u32AMDFeatureECX & X86_CPUID_AMD_FEATURE_ECX_SVM)
3256	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3257	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3258	)
3259	{
3260	/* Check if SVM is disabled */
3261	val = ASMRdMsr(MSR_K8_VM_CR);
3262	if (!(val & MSR_K8_VM_CR_SVM_DISABLE))
3263	{
3264	*pfCaps \|= SUPVTCAPS_AMD_V;
3265
3266	/* Query AMD features. */
3267	ASMCpuId(0x8000000A, &u32Dummy, &u32Dummy, &u32Dummy, &u32FeaturesEDX);
3268
3269	if (u32FeaturesEDX & AMD_CPUID_SVM_FEATURE_EDX_NESTED_PAGING)
3270	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3271
3272	return VINF_SUCCESS;
3273	}
3274	return VERR_SVM_DISABLED;
3275	}
3276	return VERR_SVM_NO_SVM;
3277	}
3278	}
3279
3280	return VERR_UNSUPPORTED_CPU;
3281	}
3282
3283
3284	/**
3285	* (Re-)initializes the per-cpu structure prior to starting or resuming the GIP
3286	* updating.
3287	*
3288	* @param pGipCpu The per CPU structure for this CPU.
3289	* @param u64NanoTS The current time.
3290	*/
3291	static void supdrvGipReInitCpu(PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
3292	{
3293	pGipCpu->u64TSC = ASMReadTSC() - pGipCpu->u32UpdateIntervalTSC;
3294	pGipCpu->u64NanoTS = u64NanoTS;
3295	}
3296
3297
3298	/**
3299	* Set the current TSC and NanoTS value for the CPU.
3300	*
3301	* @param idCpu The CPU ID. Unused - we have to use the APIC ID.
3302	* @param pvUser1 Pointer to the ring-0 GIP mapping.
3303	* @param pvUser2 Pointer to the variable holding the current time.
3304	*/
3305	static DECLCALLBACK(void) supdrvGipReInitCpuCallback(RTCPUID idCpu, void pvUser1, void pvUser2)
3306	{
3307	PSUPGLOBALINFOPAGE pGip = (PSUPGLOBALINFOPAGE)pvUser1;
3308	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
3309
3310	if (RT_LIKELY(iCpu < pGip->cCpus && pGip->aCPUs[iCpu].idCpu == idCpu))
3311	supdrvGipReInitCpu(&pGip->aCPUs[iCpu], (uint64_t )pvUser2);
3312
3313	NOREF(pvUser2);
3314	NOREF(idCpu);
3315	}
3316
3317
3318	/**
3319	* Maps the GIP into userspace and/or get the physical address of the GIP.
3320	*
3321	* @returns IPRT status code.
3322	* @param pSession Session to which the GIP mapping should belong.
3323	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
3324	* @param pHCPhysGip Where to store the physical address. (optional)
3325	*
3326	* @remark There is no reference counting on the mapping, so one call to this function
3327	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
3328	* and remove the session as a GIP user.
3329	*/
3330	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
3331	{
3332	int rc;
3333	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3334	RTR3PTR pGipR3 = NIL_RTR3PTR;
3335	RTHCPHYS HCPhys = NIL_RTHCPHYS;
3336	LogFlow(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
3337
3338	/*
3339	* Validate
3340	*/
3341	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3342	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
3343	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
3344
3345	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3346	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3347	#else
3348	RTSemFastMutexRequest(pDevExt->mtxGip);
3349	#endif
3350	if (pDevExt->pGip)
3351	{
3352	/*
3353	* Map it?
3354	*/
3355	rc = VINF_SUCCESS;
3356	if (ppGipR3)
3357	{
3358	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
3359	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
3360	RTMEM_PROT_READ, RTR0ProcHandleSelf());
3361	if (RT_SUCCESS(rc))
3362	pGipR3 = RTR0MemObjAddressR3(pSession->GipMapObjR3);
3363	}
3364
3365	/*
3366	* Get physical address.
3367	*/
3368	if (pHCPhysGip && RT_SUCCESS(rc))
3369	HCPhys = pDevExt->HCPhysGip;
3370
3371	/*
3372	* Reference globally.
3373	*/
3374	if (!pSession->fGipReferenced && RT_SUCCESS(rc))
3375	{
3376	pSession->fGipReferenced = 1;
3377	pDevExt->cGipUsers++;
3378	if (pDevExt->cGipUsers == 1)
3379	{
3380	PSUPGLOBALINFOPAGE pGipR0 = pDevExt->pGip;
3381	uint64_t u64NanoTS;
3382	uint32_t u32SystemResolution;
3383	unsigned i;
3384
3385	LogFlow(("SUPR0GipMap: Resumes GIP updating\n"));
3386
3387	/*
3388	* Try bump up the system timer resolution.
3389	* The more interrupts the better...
3390	*/
3391	if ( RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
3392	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
3393	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1953125 /* 512 HZ */, &u32SystemResolution))
3394	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 2000000 /* 500 HZ */, &u32SystemResolution))
3395	)
3396	{
3397	Assert(RTTimerGetSystemGranularity() <= u32SystemResolution);
3398	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
3399	}
3400
3401	if (pGipR0->aCPUs[0].u32TransactionId != 2 /* not the first time */)
3402	{
3403	for (i = 0; i < RT_ELEMENTS(pGipR0->aCPUs); i++)
3404	ASMAtomicUoWriteU32(&pGipR0->aCPUs[i].u32TransactionId,
3405	(pGipR0->aCPUs[i].u32TransactionId + GIP_UPDATEHZ_RECALC_FREQ * 2)
3406	& ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
3407	ASMAtomicWriteU64(&pGipR0->u64NanoTSLastUpdateHz, 0);
3408	}
3409
3410	u64NanoTS = RTTimeSystemNanoTS() - pGipR0->u32UpdateIntervalNS;
3411	if ( pGipR0->u32Mode == SUPGIPMODE_SYNC_TSC
3412	\|\| RTMpGetOnlineCount() == 1)
3413	supdrvGipReInitCpu(&pGipR0->aCPUs[0], u64NanoTS);
3414	else
3415	RTMpOnAll(supdrvGipReInitCpuCallback, pGipR0, &u64NanoTS);
3416
3417	#ifndef DO_NOT_START_GIP
3418	rc = RTTimerStart(pDevExt->pGipTimer, 0); AssertRC(rc);
3419	#endif
3420	rc = VINF_SUCCESS;
3421	}
3422	}
3423	}
3424	else
3425	{
3426	rc = VERR_GENERAL_FAILURE;
3427	Log(("SUPR0GipMap: GIP is not available!\n"));
3428	}
3429	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3430	RTSemMutexRelease(pDevExt->mtxGip);
3431	#else
3432	RTSemFastMutexRelease(pDevExt->mtxGip);
3433	#endif
3434
3435	/*
3436	* Write returns.
3437	*/
3438	if (pHCPhysGip)
3439	*pHCPhysGip = HCPhys;
3440	if (ppGipR3)
3441	*ppGipR3 = pGipR3;
3442
3443	#ifdef DEBUG_DARWIN_GIP
3444	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3445	#else
3446	LogFlow(( "SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3447	#endif
3448	return rc;
3449	}
3450
3451
3452	/**
3453	* Unmaps any user mapping of the GIP and terminates all GIP access
3454	* from this session.
3455	*
3456	* @returns IPRT status code.
3457	* @param pSession Session to which the GIP mapping should belong.
3458	*/
3459	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
3460	{
3461	int rc = VINF_SUCCESS;
3462	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3463	#ifdef DEBUG_DARWIN_GIP
3464	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
3465	pSession,
3466	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
3467	pSession->GipMapObjR3));
3468	#else
3469	LogFlow(("SUPR0GipUnmap: pSession=%p\n", pSession));
3470	#endif
3471	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3472
3473	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3474	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3475	#else
3476	RTSemFastMutexRequest(pDevExt->mtxGip);
3477	#endif
3478
3479	/*
3480	* Unmap anything?
3481	*/
3482	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
3483	{
3484	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
3485	AssertRC(rc);
3486	if (RT_SUCCESS(rc))
3487	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
3488	}
3489
3490	/*
3491	* Dereference global GIP.
3492	*/
3493	if (pSession->fGipReferenced && !rc)
3494	{
3495	pSession->fGipReferenced = 0;
3496	if ( pDevExt->cGipUsers > 0
3497	&& !--pDevExt->cGipUsers)
3498	{
3499	LogFlow(("SUPR0GipUnmap: Suspends GIP updating\n"));
3500	#ifndef DO_NOT_START_GIP
3501	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = VINF_SUCCESS;
3502	#endif
3503
3504	if (pDevExt->u32SystemTimerGranularityGrant)
3505	{
3506	int rc2 = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant);
3507	AssertRC(rc2);
3508	pDevExt->u32SystemTimerGranularityGrant = 0;
3509	}
3510	}
3511	}
3512
3513	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3514	RTSemMutexRelease(pDevExt->mtxGip);
3515	#else
3516	RTSemFastMutexRelease(pDevExt->mtxGip);
3517	#endif
3518
3519	return rc;
3520	}
3521
3522
3523	/**
3524	* Gets the GIP pointer.
3525	*
3526	* @returns Pointer to the GIP or NULL.
3527	*/
3528	SUPDECL(PSUPGLOBALINFOPAGE) SUPGetGIP(void)
3529	{
3530	return g_pSUPGlobalInfoPage;
3531	}
3532
3533
3534	/**
3535	* Register a component factory with the support driver.
3536	*
3537	* This is currently restricted to kernel sessions only.
3538	*
3539	* @returns VBox status code.
3540	* @retval VINF_SUCCESS on success.
3541	* @retval VERR_NO_MEMORY if we're out of memory.
3542	* @retval VERR_ALREADY_EXISTS if the factory has already been registered.
3543	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3544	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3545	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3546	*
3547	* @param pSession The SUPDRV session (must be a ring-0 session).
3548	* @param pFactory Pointer to the component factory registration structure.
3549	*
3550	* @remarks This interface is also available via SUPR0IdcComponentRegisterFactory.
3551	*/
3552	SUPR0DECL(int) SUPR0ComponentRegisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3553	{
3554	PSUPDRVFACTORYREG pNewReg;
3555	const char *psz;
3556	int rc;
3557
3558	/*
3559	* Validate parameters.
3560	*/
3561	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3562	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3563	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3564	AssertPtrReturn(pFactory->pfnQueryFactoryInterface, VERR_INVALID_POINTER);
3565	psz = RTStrEnd(pFactory->szName, sizeof(pFactory->szName));
3566	AssertReturn(psz, VERR_INVALID_PARAMETER);
3567
3568	/*
3569	* Allocate and initialize a new registration structure.
3570	*/
3571	pNewReg = (PSUPDRVFACTORYREG)RTMemAlloc(sizeof(SUPDRVFACTORYREG));
3572	if (pNewReg)
3573	{
3574	pNewReg->pNext = NULL;
3575	pNewReg->pFactory = pFactory;
3576	pNewReg->pSession = pSession;
3577	pNewReg->cchName = psz - &pFactory->szName[0];
3578
3579	/*
3580	* Add it to the tail of the list after checking for prior registration.
3581	*/
3582	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3583	if (RT_SUCCESS(rc))
3584	{
3585	PSUPDRVFACTORYREG pPrev = NULL;
3586	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3587	while (pCur && pCur->pFactory != pFactory)
3588	{
3589	pPrev = pCur;
3590	pCur = pCur->pNext;
3591	}
3592	if (!pCur)
3593	{
3594	if (pPrev)
3595	pPrev->pNext = pNewReg;
3596	else
3597	pSession->pDevExt->pComponentFactoryHead = pNewReg;
3598	rc = VINF_SUCCESS;
3599	}
3600	else
3601	rc = VERR_ALREADY_EXISTS;
3602
3603	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3604	}
3605
3606	if (RT_FAILURE(rc))
3607	RTMemFree(pNewReg);
3608	}
3609	else
3610	rc = VERR_NO_MEMORY;
3611	return rc;
3612	}
3613
3614
3615	/**
3616	* Deregister a component factory.
3617	*
3618	* @returns VBox status code.
3619	* @retval VINF_SUCCESS on success.
3620	* @retval VERR_NOT_FOUND if the factory wasn't registered.
3621	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3622	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3623	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3624	*
3625	* @param pSession The SUPDRV session (must be a ring-0 session).
3626	* @param pFactory Pointer to the component factory registration structure
3627	* previously passed SUPR0ComponentRegisterFactory().
3628	*
3629	* @remarks This interface is also available via SUPR0IdcComponentDeregisterFactory.
3630	*/
3631	SUPR0DECL(int) SUPR0ComponentDeregisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3632	{
3633	int rc;
3634
3635	/*
3636	* Validate parameters.
3637	*/
3638	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3639	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3640	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3641
3642	/*
3643	* Take the lock and look for the registration record.
3644	*/
3645	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3646	if (RT_SUCCESS(rc))
3647	{
3648	PSUPDRVFACTORYREG pPrev = NULL;
3649	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3650	while (pCur && pCur->pFactory != pFactory)
3651	{
3652	pPrev = pCur;
3653	pCur = pCur->pNext;
3654	}
3655	if (pCur)
3656	{
3657	if (!pPrev)
3658	pSession->pDevExt->pComponentFactoryHead = pCur->pNext;
3659	else
3660	pPrev->pNext = pCur->pNext;
3661
3662	pCur->pNext = NULL;
3663	pCur->pFactory = NULL;
3664	pCur->pSession = NULL;
3665	rc = VINF_SUCCESS;
3666	}
3667	else
3668	rc = VERR_NOT_FOUND;
3669
3670	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3671
3672	RTMemFree(pCur);
3673	}
3674	return rc;
3675	}
3676
3677
3678	/**
3679	* Queries a component factory.
3680	*
3681	* @returns VBox status code.
3682	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3683	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3684	* @retval VERR_SUPDRV_COMPONENT_NOT_FOUND if the component factory wasn't found.
3685	* @retval VERR_SUPDRV_INTERFACE_NOT_SUPPORTED if the interface wasn't supported.
3686	*
3687	* @param pSession The SUPDRV session.
3688	* @param pszName The name of the component factory.
3689	* @param pszInterfaceUuid The UUID of the factory interface (stringified).
3690	* @param ppvFactoryIf Where to store the factory interface.
3691	*/
3692	SUPR0DECL(int) SUPR0ComponentQueryFactory(PSUPDRVSESSION pSession, const char pszName, const char pszInterfaceUuid, void **ppvFactoryIf)
3693	{
3694	const char *pszEnd;
3695	size_t cchName;
3696	int rc;
3697
3698	/*
3699	* Validate parameters.
3700	*/
3701	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3702
3703	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
3704	pszEnd = RTStrEnd(pszName, RT_SIZEOFMEMB(SUPDRVFACTORY, szName));
3705	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3706	cchName = pszEnd - pszName;
3707
3708	AssertPtrReturn(pszInterfaceUuid, VERR_INVALID_POINTER);
3709	pszEnd = RTStrEnd(pszInterfaceUuid, RTUUID_STR_LENGTH);
3710	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3711
3712	AssertPtrReturn(ppvFactoryIf, VERR_INVALID_POINTER);
3713	*ppvFactoryIf = NULL;
3714
3715	/*
3716	* Take the lock and try all factories by this name.
3717	*/
3718	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3719	if (RT_SUCCESS(rc))
3720	{
3721	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3722	rc = VERR_SUPDRV_COMPONENT_NOT_FOUND;
3723	while (pCur)
3724	{
3725	if ( pCur->cchName == cchName
3726	&& !memcmp(pCur->pFactory->szName, pszName, cchName))
3727	{
3728	void *pvFactory = pCur->pFactory->pfnQueryFactoryInterface(pCur->pFactory, pSession, pszInterfaceUuid);
3729	if (pvFactory)
3730	{
3731	*ppvFactoryIf = pvFactory;
3732	rc = VINF_SUCCESS;
3733	break;
3734	}
3735	rc = VERR_SUPDRV_INTERFACE_NOT_SUPPORTED;
3736	}
3737
3738	/* next */
3739	pCur = pCur->pNext;
3740	}
3741
3742	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3743	}
3744	return rc;
3745	}
3746
3747
3748	/**
3749	* Adds a memory object to the session.
3750	*
3751	* @returns IPRT status code.
3752	* @param pMem Memory tracking structure containing the
3753	* information to track.
3754	* @param pSession The session.
3755	*/
3756	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
3757	{
3758	PSUPDRVBUNDLE pBundle;
3759
3760	/*
3761	* Find free entry and record the allocation.
3762	*/
3763	RTSpinlockAcquire(pSession->Spinlock);
3764	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3765	{
3766	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
3767	{
3768	unsigned i;
3769	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3770	{
3771	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
3772	{
3773	pBundle->cUsed++;
3774	pBundle->aMem[i] = *pMem;
3775	RTSpinlockRelease(pSession->Spinlock);
3776	return VINF_SUCCESS;
3777	}
3778	}
3779	AssertFailed(); /* !!this can't be happening!!! */
3780	}
3781	}
3782	RTSpinlockRelease(pSession->Spinlock);
3783
3784	/*
3785	* Need to allocate a new bundle.
3786	* Insert into the last entry in the bundle.
3787	*/
3788	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
3789	if (!pBundle)
3790	return VERR_NO_MEMORY;
3791
3792	/* take last entry. */
3793	pBundle->cUsed++;
3794	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
3795
3796	/* insert into list. */
3797	RTSpinlockAcquire(pSession->Spinlock);
3798	pBundle->pNext = pSession->Bundle.pNext;
3799	pSession->Bundle.pNext = pBundle;
3800	RTSpinlockRelease(pSession->Spinlock);
3801
3802	return VINF_SUCCESS;
3803	}
3804
3805
3806	/**
3807	* Releases a memory object referenced by pointer and type.
3808	*
3809	* @returns IPRT status code.
3810	* @param pSession Session data.
3811	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
3812	* @param eType Memory type.
3813	*/
3814	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
3815	{
3816	PSUPDRVBUNDLE pBundle;
3817
3818	/*
3819	* Validate input.
3820	*/
3821	if (!uPtr)
3822	{
3823	Log(("Illegal address %p\n", (void *)uPtr));
3824	return VERR_INVALID_PARAMETER;
3825	}
3826
3827	/*
3828	* Search for the address.
3829	*/
3830	RTSpinlockAcquire(pSession->Spinlock);
3831	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3832	{
3833	if (pBundle->cUsed > 0)
3834	{
3835	unsigned i;
3836	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3837	{
3838	if ( pBundle->aMem[i].eType == eType
3839	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3840	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
3841	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3842	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
3843	)
3844	{
3845	/* Make a copy of it and release it outside the spinlock. */
3846	SUPDRVMEMREF Mem = pBundle->aMem[i];
3847	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
3848	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
3849	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
3850	RTSpinlockRelease(pSession->Spinlock);
3851
3852	if (Mem.MapObjR3 != NIL_RTR0MEMOBJ)
3853	{
3854	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
3855	AssertRC(rc); /** @todo figure out how to handle this. */
3856	}
3857	if (Mem.MemObj != NIL_RTR0MEMOBJ)
3858	{
3859	int rc = RTR0MemObjFree(Mem.MemObj, true /* fFreeMappings */);
3860	AssertRC(rc); /** @todo figure out how to handle this. */
3861	}
3862	return VINF_SUCCESS;
3863	}
3864	}
3865	}
3866	}
3867	RTSpinlockRelease(pSession->Spinlock);
3868	Log(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
3869	return VERR_INVALID_PARAMETER;
3870	}
3871
3872
3873	/**
3874	* Opens an image. If it's the first time it's opened the call must upload
3875	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
3876	*
3877	* This is the 1st step of the loading.
3878	*
3879	* @returns IPRT status code.
3880	* @param pDevExt Device globals.
3881	* @param pSession Session data.
3882	* @param pReq The open request.
3883	*/
3884	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
3885	{
3886	int rc;
3887	PSUPDRVLDRIMAGE pImage;
3888	void *pv;
3889	size_t cchName = strlen(pReq->u.In.szName); /* (caller checked < 32). */
3890	LogFlow(("supdrvIOCtl_LdrOpen: szName=%s cbImageWithTabs=%d\n", pReq->u.In.szName, pReq->u.In.cbImageWithTabs));
3891
3892	/*
3893	* Check if we got an instance of the image already.
3894	*/
3895	supdrvLdrLock(pDevExt);
3896	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
3897	{
3898	if ( pImage->szName[cchName] == '\0'
3899	&& !memcmp(pImage->szName, pReq->u.In.szName, cchName))
3900	{
3901	/** @todo check cbImageBits and cbImageWithTabs here, if they differs that indicates that the images are different. */
3902	pImage->cUsage++;
3903	pReq->u.Out.pvImageBase = pImage->pvImage;
3904	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
3905	pReq->u.Out.fNativeLoader = pImage->fNative;
3906	supdrvLdrAddUsage(pSession, pImage);
3907	supdrvLdrUnlock(pDevExt);
3908	return VINF_SUCCESS;
3909	}
3910	}
3911	/* (not found - add it!) */
3912
3913	/*
3914	* Allocate memory.
3915	*/
3916	Assert(cchName < sizeof(pImage->szName));
3917	pv = RTMemAlloc(sizeof(SUPDRVLDRIMAGE));
3918	if (!pv)
3919	{
3920	supdrvLdrUnlock(pDevExt);
3921	Log(("supdrvIOCtl_LdrOpen: RTMemAlloc() failed\n"));
3922	return /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_2;
3923	}
3924
3925	/*
3926	* Setup and link in the LDR stuff.
3927	*/
3928	pImage = (PSUPDRVLDRIMAGE)pv;
3929	pImage->pvImage = NULL;
3930	pImage->pvImageAlloc = NULL;
3931	pImage->cbImageWithTabs = pReq->u.In.cbImageWithTabs;
3932	pImage->cbImageBits = pReq->u.In.cbImageBits;
3933	pImage->cSymbols = 0;
3934	pImage->paSymbols = NULL;
3935	pImage->pachStrTab = NULL;
3936	pImage->cbStrTab = 0;
3937	pImage->pfnModuleInit = NULL;
3938	pImage->pfnModuleTerm = NULL;
3939	pImage->pfnServiceReqHandler = NULL;
3940	pImage->uState = SUP_IOCTL_LDR_OPEN;
3941	pImage->cUsage = 1;
3942	pImage->pDevExt = pDevExt;
3943	memcpy(pImage->szName, pReq->u.In.szName, cchName + 1);
3944
3945	/*
3946	* Try load it using the native loader, if that isn't supported, fall back
3947	* on the older method.
3948	*/
3949	pImage->fNative = true;
3950	rc = supdrvOSLdrOpen(pDevExt, pImage, pReq->u.In.szFilename);
3951	if (rc == VERR_NOT_SUPPORTED)
3952	{
3953	pImage->pvImageAlloc = RTMemExecAlloc(pImage->cbImageBits + 31);
3954	pImage->pvImage = RT_ALIGN_P(pImage->pvImageAlloc, 32);
3955	pImage->fNative = false;
3956	rc = pImage->pvImageAlloc ? VINF_SUCCESS : VERR_NO_EXEC_MEMORY;
3957	}
3958	if (RT_FAILURE(rc))
3959	{
3960	supdrvLdrUnlock(pDevExt);
3961	RTMemFree(pImage);
3962	Log(("supdrvIOCtl_LdrOpen(%s): failed - %Rrc\n", pReq->u.In.szName, rc));
3963	return rc;
3964	}
3965	Assert(VALID_PTR(pImage->pvImage) \|\| RT_FAILURE(rc));
3966
3967	/*
3968	* Link it.
3969	*/
3970	pImage->pNext = pDevExt->pLdrImages;
3971	pDevExt->pLdrImages = pImage;
3972
3973	supdrvLdrAddUsage(pSession, pImage);
3974
3975	pReq->u.Out.pvImageBase = pImage->pvImage;
3976	pReq->u.Out.fNeedsLoading = true;
3977	pReq->u.Out.fNativeLoader = pImage->fNative;
3978	supdrvOSLdrNotifyOpened(pDevExt, pImage);
3979
3980	supdrvLdrUnlock(pDevExt);
3981	return VINF_SUCCESS;
3982	}
3983
3984
3985	/**
3986	* Worker that validates a pointer to an image entrypoint.
3987	*
3988	* @returns IPRT status code.
3989	* @param pDevExt The device globals.
3990	* @param pImage The loader image.
3991	* @param pv The pointer into the image.
3992	* @param fMayBeNull Whether it may be NULL.
3993	* @param pszWhat What is this entrypoint? (for logging)
3994	* @param pbImageBits The image bits prepared by ring-3.
3995	*
3996	* @remarks Will leave the lock on failure.
3997	*/
3998	static int supdrvLdrValidatePointer(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage, void *pv,
3999	bool fMayBeNull, const uint8_t pbImageBits, const char pszWhat)
4000	{
4001	if (!fMayBeNull \|\| pv)
4002	{
4003	if ((uintptr_t)pv - (uintptr_t)pImage->pvImage >= pImage->cbImageBits)
4004	{
4005	supdrvLdrUnlock(pDevExt);
4006	Log(("Out of range (%p LB %#x): %s=%p\n", pImage->pvImage, pImage->cbImageBits, pszWhat, pv));
4007	return VERR_INVALID_PARAMETER;
4008	}
4009
4010	if (pImage->fNative)
4011	{
4012	int rc = supdrvOSLdrValidatePointer(pDevExt, pImage, pv, pbImageBits);
4013	if (RT_FAILURE(rc))
4014	{
4015	supdrvLdrUnlock(pDevExt);
4016	Log(("Bad entry point address: %s=%p (rc=%Rrc)\n", pszWhat, pv, rc));
4017	return rc;
4018	}
4019	}
4020	}
4021	return VINF_SUCCESS;
4022	}
4023
4024
4025	/**
4026	* Loads the image bits.
4027	*
4028	* This is the 2nd step of the loading.
4029	*
4030	* @returns IPRT status code.
4031	* @param pDevExt Device globals.
4032	* @param pSession Session data.
4033	* @param pReq The request.
4034	*/
4035	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
4036	{
4037	PSUPDRVLDRUSAGE pUsage;
4038	PSUPDRVLDRIMAGE pImage;
4039	int rc;
4040	LogFlow(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImageWithBits=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImageWithTabs));
4041
4042	/*
4043	* Find the ldr image.
4044	*/
4045	supdrvLdrLock(pDevExt);
4046	pUsage = pSession->pLdrUsage;
4047	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4048	pUsage = pUsage->pNext;
4049	if (!pUsage)
4050	{
4051	supdrvLdrUnlock(pDevExt);
4052	Log(("SUP_IOCTL_LDR_LOAD: couldn't find image!\n"));
4053	return VERR_INVALID_HANDLE;
4054	}
4055	pImage = pUsage->pImage;
4056
4057	/*
4058	* Validate input.
4059	*/
4060	if ( pImage->cbImageWithTabs != pReq->u.In.cbImageWithTabs
4061	\|\| pImage->cbImageBits != pReq->u.In.cbImageBits)
4062	{
4063	supdrvLdrUnlock(pDevExt);
4064	Log(("SUP_IOCTL_LDR_LOAD: image size mismatch!! %d(prep) != %d(load) or %d != %d\n",
4065	pImage->cbImageWithTabs, pReq->u.In.cbImageWithTabs, pImage->cbImageBits, pReq->u.In.cbImageBits));
4066	return VERR_INVALID_HANDLE;
4067	}
4068
4069	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
4070	{
4071	unsigned uState = pImage->uState;
4072	supdrvLdrUnlock(pDevExt);
4073	if (uState != SUP_IOCTL_LDR_LOAD)
4074	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
4075	return VERR_ALREADY_LOADED;
4076	}
4077
4078	switch (pReq->u.In.eEPType)
4079	{
4080	case SUPLDRLOADEP_NOTHING:
4081	break;
4082
4083	case SUPLDRLOADEP_VMMR0:
4084	rc = supdrvLdrValidatePointer( pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0, false, pReq->u.In.abImage, "pvVMMR0");
4085	if (RT_SUCCESS(rc))
4086	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt, false, pReq->u.In.abImage, "pvVMMR0EntryInt");
4087	if (RT_SUCCESS(rc))
4088	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, false, pReq->u.In.abImage, "pvVMMR0EntryFast");
4089	if (RT_SUCCESS(rc))
4090	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx, false, pReq->u.In.abImage, "pvVMMR0EntryEx");
4091	if (RT_FAILURE(rc))
4092	return rc;
4093	break;
4094
4095	case SUPLDRLOADEP_SERVICE:
4096	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.Service.pfnServiceReq, false, pReq->u.In.abImage, "pfnServiceReq");
4097	if (RT_FAILURE(rc))
4098	return rc;
4099	if ( pReq->u.In.EP.Service.apvReserved[0] != NIL_RTR0PTR
4100	\|\| pReq->u.In.EP.Service.apvReserved[1] != NIL_RTR0PTR
4101	\|\| pReq->u.In.EP.Service.apvReserved[2] != NIL_RTR0PTR)
4102	{
4103	supdrvLdrUnlock(pDevExt);
4104	Log(("Out of range (%p LB %#x): apvReserved={%p,%p,%p} MBZ!\n",
4105	pImage->pvImage, pReq->u.In.cbImageWithTabs,
4106	pReq->u.In.EP.Service.apvReserved[0],
4107	pReq->u.In.EP.Service.apvReserved[1],
4108	pReq->u.In.EP.Service.apvReserved[2]));
4109	return VERR_INVALID_PARAMETER;
4110	}
4111	break;
4112
4113	default:
4114	supdrvLdrUnlock(pDevExt);
4115	Log(("Invalid eEPType=%d\n", pReq->u.In.eEPType));
4116	return VERR_INVALID_PARAMETER;
4117	}
4118
4119	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleInit, true, pReq->u.In.abImage, "pfnModuleInit");
4120	if (RT_FAILURE(rc))
4121	return rc;
4122	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleTerm, true, pReq->u.In.abImage, "pfnModuleTerm");
4123	if (RT_FAILURE(rc))
4124	return rc;
4125
4126	/*
4127	* Allocate and copy the tables.
4128	* (No need to do try/except as this is a buffered request.)
4129	*/
4130	pImage->cbStrTab = pReq->u.In.cbStrTab;
4131	if (pImage->cbStrTab)
4132	{
4133	pImage->pachStrTab = (char *)RTMemAlloc(pImage->cbStrTab);
4134	if (pImage->pachStrTab)
4135	memcpy(pImage->pachStrTab, &pReq->u.In.abImage[pReq->u.In.offStrTab], pImage->cbStrTab);
4136	else
4137	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_3;
4138	}
4139
4140	pImage->cSymbols = pReq->u.In.cSymbols;
4141	if (RT_SUCCESS(rc) && pImage->cSymbols)
4142	{
4143	size_t cbSymbols = pImage->cSymbols * sizeof(SUPLDRSYM);
4144	pImage->paSymbols = (PSUPLDRSYM)RTMemAlloc(cbSymbols);
4145	if (pImage->paSymbols)
4146	memcpy(pImage->paSymbols, &pReq->u.In.abImage[pReq->u.In.offSymbols], cbSymbols);
4147	else
4148	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_4;
4149	}
4150
4151	/*
4152	* Copy the bits / complete native loading.
4153	*/
4154	if (RT_SUCCESS(rc))
4155	{
4156	pImage->uState = SUP_IOCTL_LDR_LOAD;
4157	pImage->pfnModuleInit = pReq->u.In.pfnModuleInit;
4158	pImage->pfnModuleTerm = pReq->u.In.pfnModuleTerm;
4159
4160	if (pImage->fNative)
4161	rc = supdrvOSLdrLoad(pDevExt, pImage, pReq->u.In.abImage, pReq);
4162	else
4163	{
4164	memcpy(pImage->pvImage, &pReq->u.In.abImage[0], pImage->cbImageBits);
4165	SUPR0Printf("vboxdrv: Loaded '%s' at %p\n", pImage->szName, pImage->pvImage);
4166	}
4167	}
4168
4169	/*
4170	* Update any entry points.
4171	*/
4172	if (RT_SUCCESS(rc))
4173	{
4174	switch (pReq->u.In.eEPType)
4175	{
4176	default:
4177	case SUPLDRLOADEP_NOTHING:
4178	rc = VINF_SUCCESS;
4179	break;
4180	case SUPLDRLOADEP_VMMR0:
4181	rc = supdrvLdrSetVMMR0EPs(pDevExt, pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
4182	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
4183	break;
4184	case SUPLDRLOADEP_SERVICE:
4185	pImage->pfnServiceReqHandler = pReq->u.In.EP.Service.pfnServiceReq;
4186	rc = VINF_SUCCESS;
4187	break;
4188	}
4189	}
4190
4191	/*
4192	* On success call the module initialization.
4193	*/
4194	LogFlow(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
4195	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
4196	{
4197	Log(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
4198	pDevExt->pLdrInitImage = pImage;
4199	pDevExt->hLdrInitThread = RTThreadNativeSelf();
4200	rc = pImage->pfnModuleInit(pImage);
4201	pDevExt->pLdrInitImage = NULL;
4202	pDevExt->hLdrInitThread = NIL_RTNATIVETHREAD;
4203	if (RT_FAILURE(rc) && pDevExt->pvVMMR0 == pImage->pvImage)
4204	supdrvLdrUnsetVMMR0EPs(pDevExt);
4205	}
4206
4207	if (RT_FAILURE(rc))
4208	{
4209	/* Inform the tracing component in case ModuleInit registered TPs. */
4210	supdrvTracerModuleUnloading(pDevExt, pImage);
4211
4212	pImage->uState = SUP_IOCTL_LDR_OPEN;
4213	pImage->pfnModuleInit = NULL;
4214	pImage->pfnModuleTerm = NULL;
4215	pImage->pfnServiceReqHandler= NULL;
4216	pImage->cbStrTab = 0;
4217	RTMemFree(pImage->pachStrTab);
4218	pImage->pachStrTab = NULL;
4219	RTMemFree(pImage->paSymbols);
4220	pImage->paSymbols = NULL;
4221	pImage->cSymbols = 0;
4222	}
4223
4224	supdrvLdrUnlock(pDevExt);
4225	return rc;
4226	}
4227
4228
4229	/**
4230	* Frees a previously loaded (prep'ed) image.
4231	*
4232	* @returns IPRT status code.
4233	* @param pDevExt Device globals.
4234	* @param pSession Session data.
4235	* @param pReq The request.
4236	*/
4237	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
4238	{
4239	int rc;
4240	PSUPDRVLDRUSAGE pUsagePrev;
4241	PSUPDRVLDRUSAGE pUsage;
4242	PSUPDRVLDRIMAGE pImage;
4243	LogFlow(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
4244
4245	/*
4246	* Find the ldr image.
4247	*/
4248	supdrvLdrLock(pDevExt);
4249	pUsagePrev = NULL;
4250	pUsage = pSession->pLdrUsage;
4251	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4252	{
4253	pUsagePrev = pUsage;
4254	pUsage = pUsage->pNext;
4255	}
4256	if (!pUsage)
4257	{
4258	supdrvLdrUnlock(pDevExt);
4259	Log(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
4260	return VERR_INVALID_HANDLE;
4261	}
4262
4263	/*
4264	* Check if we can remove anything.
4265	*/
4266	rc = VINF_SUCCESS;
4267	pImage = pUsage->pImage;
4268	if (pImage->cUsage <= 1 \|\| pUsage->cUsage <= 1)
4269	{
4270	/*
4271	* Check if there are any objects with destructors in the image, if
4272	* so leave it for the session cleanup routine so we get a chance to
4273	* clean things up in the right order and not leave them all dangling.
4274	*/
4275	RTSpinlockAcquire(pDevExt->Spinlock);
4276	if (pImage->cUsage <= 1)
4277	{
4278	PSUPDRVOBJ pObj;
4279	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4280	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4281	{
4282	rc = VERR_DANGLING_OBJECTS;
4283	break;
4284	}
4285	}
4286	else
4287	{
4288	PSUPDRVUSAGE pGenUsage;
4289	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
4290	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4291	{
4292	rc = VERR_DANGLING_OBJECTS;
4293	break;
4294	}
4295	}
4296	RTSpinlockRelease(pDevExt->Spinlock);
4297	if (rc == VINF_SUCCESS)
4298	{
4299	/* unlink it */
4300	if (pUsagePrev)
4301	pUsagePrev->pNext = pUsage->pNext;
4302	else
4303	pSession->pLdrUsage = pUsage->pNext;
4304
4305	/* free it */
4306	pUsage->pImage = NULL;
4307	pUsage->pNext = NULL;
4308	RTMemFree(pUsage);
4309
4310	/*
4311	* Dereference the image.
4312	*/
4313	if (pImage->cUsage <= 1)
4314	supdrvLdrFree(pDevExt, pImage);
4315	else
4316	pImage->cUsage--;
4317	}
4318	else
4319	{
4320	Log(("supdrvIOCtl_LdrFree: Dangling objects in %p/%s!\n", pImage->pvImage, pImage->szName));
4321	rc = VINF_SUCCESS; /** @todo BRANCH-2.1: remove this after branching. */
4322	}
4323	}
4324	else
4325	{
4326	/*
4327	* Dereference both image and usage.
4328	*/
4329	pImage->cUsage--;
4330	pUsage->cUsage--;
4331	}
4332
4333	supdrvLdrUnlock(pDevExt);
4334	return rc;
4335	}
4336
4337
4338	/**
4339	* Gets the address of a symbol in an open image.
4340	*
4341	* @returns IPRT status code.
4342	* @param pDevExt Device globals.
4343	* @param pSession Session data.
4344	* @param pReq The request buffer.
4345	*/
4346	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
4347	{
4348	PSUPDRVLDRIMAGE pImage;
4349	PSUPDRVLDRUSAGE pUsage;
4350	uint32_t i;
4351	PSUPLDRSYM paSyms;
4352	const char *pchStrings;
4353	const size_t cbSymbol = strlen(pReq->u.In.szSymbol) + 1;
4354	void *pvSymbol = NULL;
4355	int rc = VERR_GENERAL_FAILURE;
4356	Log3(("supdrvIOCtl_LdrGetSymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
4357
4358	/*
4359	* Find the ldr image.
4360	*/
4361	supdrvLdrLock(pDevExt);
4362	pUsage = pSession->pLdrUsage;
4363	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4364	pUsage = pUsage->pNext;
4365	if (!pUsage)
4366	{
4367	supdrvLdrUnlock(pDevExt);
4368	Log(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
4369	return VERR_INVALID_HANDLE;
4370	}
4371	pImage = pUsage->pImage;
4372	if (pImage->uState != SUP_IOCTL_LDR_LOAD)
4373	{
4374	unsigned uState = pImage->uState;
4375	supdrvLdrUnlock(pDevExt);
4376	Log(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", uState, uState)); NOREF(uState);
4377	return VERR_ALREADY_LOADED;
4378	}
4379
4380	/*
4381	* Search the symbol strings.
4382	*
4383	* Note! The int32_t is for native loading on solaris where the data
4384	* and text segments are in very different places.
4385	*/
4386	pchStrings = pImage->pachStrTab;
4387	paSyms = pImage->paSymbols;
4388	for (i = 0; i < pImage->cSymbols; i++)
4389	{
4390	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4391	&& !memcmp(pchStrings + paSyms[i].offName, pReq->u.In.szSymbol, cbSymbol))
4392	{
4393	pvSymbol = (uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol;
4394	rc = VINF_SUCCESS;
4395	break;
4396	}
4397	}
4398	supdrvLdrUnlock(pDevExt);
4399	pReq->u.Out.pvSymbol = pvSymbol;
4400	return rc;
4401	}
4402
4403
4404	/**
4405	* Gets the address of a symbol in an open image or the support driver.
4406	*
4407	* @returns VINF_SUCCESS on success.
4408	* @returns
4409	* @param pDevExt Device globals.
4410	* @param pSession Session data.
4411	* @param pReq The request buffer.
4412	*/
4413	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq)
4414	{
4415	int rc = VINF_SUCCESS;
4416	const char *pszSymbol = pReq->u.In.pszSymbol;
4417	const char *pszModule = pReq->u.In.pszModule;
4418	size_t cbSymbol;
4419	char const *pszEnd;
4420	uint32_t i;
4421
4422	/*
4423	* Input validation.
4424	*/
4425	AssertPtrReturn(pszSymbol, VERR_INVALID_POINTER);
4426	pszEnd = RTStrEnd(pszSymbol, 512);
4427	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4428	cbSymbol = pszEnd - pszSymbol + 1;
4429
4430	if (pszModule)
4431	{
4432	AssertPtrReturn(pszModule, VERR_INVALID_POINTER);
4433	pszEnd = RTStrEnd(pszModule, 64);
4434	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4435	}
4436	Log3(("supdrvIDC_LdrGetSymbol: pszModule=%p:{%s} pszSymbol=%p:{%s}\n", pszModule, pszModule, pszSymbol, pszSymbol));
4437
4438
4439	if ( !pszModule
4440	\|\| !strcmp(pszModule, "SupDrv"))
4441	{
4442	/*
4443	* Search the support driver export table.
4444	*/
4445	for (i = 0; i < RT_ELEMENTS(g_aFunctions); i++)
4446	if (!strcmp(g_aFunctions[i].szName, pszSymbol))
4447	{
4448	pReq->u.Out.pfnSymbol = g_aFunctions[i].pfn;
4449	break;
4450	}
4451	}
4452	else
4453	{
4454	/*
4455	* Find the loader image.
4456	*/
4457	PSUPDRVLDRIMAGE pImage;
4458
4459	supdrvLdrLock(pDevExt);
4460
4461	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
4462	if (!strcmp(pImage->szName, pszModule))
4463	break;
4464	if (pImage && pImage->uState == SUP_IOCTL_LDR_LOAD)
4465	{
4466	/*
4467	* Search the symbol strings.
4468	*/
4469	const char *pchStrings = pImage->pachStrTab;
4470	PCSUPLDRSYM paSyms = pImage->paSymbols;
4471	for (i = 0; i < pImage->cSymbols; i++)
4472	{
4473	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4474	&& !memcmp(pchStrings + paSyms[i].offName, pszSymbol, cbSymbol))
4475	{
4476	/*
4477	* Found it! Calc the symbol address and add a reference to the module.
4478	*/
4479	pReq->u.Out.pfnSymbol = (PFNRT)((uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol);
4480	rc = supdrvLdrAddUsage(pSession, pImage);
4481	break;
4482	}
4483	}
4484	}
4485	else
4486	rc = pImage ? VERR_WRONG_ORDER : VERR_MODULE_NOT_FOUND;
4487
4488	supdrvLdrUnlock(pDevExt);
4489	}
4490	return rc;
4491	}
4492
4493
4494	/**
4495	* Updates the VMMR0 entry point pointers.
4496	*
4497	* @returns IPRT status code.
4498	* @param pDevExt Device globals.
4499	* @param pSession Session data.
4500	* @param pVMMR0 VMMR0 image handle.
4501	* @param pvVMMR0EntryInt VMMR0EntryInt address.
4502	* @param pvVMMR0EntryFast VMMR0EntryFast address.
4503	* @param pvVMMR0EntryEx VMMR0EntryEx address.
4504	* @remark Caller must own the loader mutex.
4505	*/
4506	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx)
4507	{
4508	int rc = VINF_SUCCESS;
4509	LogFlow(("supdrvLdrSetR0EP pvVMMR0=%p pvVMMR0EntryInt=%p\n", pvVMMR0, pvVMMR0EntryInt));
4510
4511
4512	/*
4513	* Check if not yet set.
4514	*/
4515	if (!pDevExt->pvVMMR0)
4516	{
4517	pDevExt->pvVMMR0 = pvVMMR0;
4518	pDevExt->pfnVMMR0EntryInt = pvVMMR0EntryInt;
4519	pDevExt->pfnVMMR0EntryFast = pvVMMR0EntryFast;
4520	pDevExt->pfnVMMR0EntryEx = pvVMMR0EntryEx;
4521	}
4522	else
4523	{
4524	/*
4525	* Return failure or success depending on whether the values match or not.
4526	*/
4527	if ( pDevExt->pvVMMR0 != pvVMMR0
4528	\|\| (void *)pDevExt->pfnVMMR0EntryInt != pvVMMR0EntryInt
4529	\|\| (void *)pDevExt->pfnVMMR0EntryFast != pvVMMR0EntryFast
4530	\|\| (void *)pDevExt->pfnVMMR0EntryEx != pvVMMR0EntryEx)
4531	{
4532	AssertMsgFailed(("SUP_IOCTL_LDR_SETR0EP: Already set pointing to a different module!\n"));
4533	rc = VERR_INVALID_PARAMETER;
4534	}
4535	}
4536	return rc;
4537	}
4538
4539
4540	/**
4541	* Unsets the VMMR0 entry point installed by supdrvLdrSetR0EP.
4542	*
4543	* @param pDevExt Device globals.
4544	*/
4545	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt)
4546	{
4547	pDevExt->pvVMMR0 = NULL;
4548	pDevExt->pfnVMMR0EntryInt = NULL;
4549	pDevExt->pfnVMMR0EntryFast = NULL;
4550	pDevExt->pfnVMMR0EntryEx = NULL;
4551	}
4552
4553
4554	/**
4555	* Adds a usage reference in the specified session of an image.
4556	*
4557	* Called while owning the loader semaphore.
4558	*
4559	* @returns VINF_SUCCESS on success and VERR_NO_MEMORY on failure.
4560	* @param pSession Session in question.
4561	* @param pImage Image which the session is using.
4562	*/
4563	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage)
4564	{
4565	PSUPDRVLDRUSAGE pUsage;
4566	LogFlow(("supdrvLdrAddUsage: pImage=%p\n", pImage));
4567
4568	/*
4569	* Referenced it already?
4570	*/
4571	pUsage = pSession->pLdrUsage;
4572	while (pUsage)
4573	{
4574	if (pUsage->pImage == pImage)
4575	{
4576	pUsage->cUsage++;
4577	return VINF_SUCCESS;
4578	}
4579	pUsage = pUsage->pNext;
4580	}
4581
4582	/*
4583	* Allocate new usage record.
4584	*/
4585	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
4586	AssertReturn(pUsage, /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_5);
4587	pUsage->cUsage = 1;
4588	pUsage->pImage = pImage;
4589	pUsage->pNext = pSession->pLdrUsage;
4590	pSession->pLdrUsage = pUsage;
4591	return VINF_SUCCESS;
4592	}
4593
4594
4595	/**
4596	* Frees a load image.
4597	*
4598	* @param pDevExt Pointer to device extension.
4599	* @param pImage Pointer to the image we're gonna free.
4600	* This image must exit!
4601	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
4602	*/
4603	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
4604	{
4605	PSUPDRVLDRIMAGE pImagePrev;
4606	LogFlow(("supdrvLdrFree: pImage=%p\n", pImage));
4607
4608	/* find it - arg. should've used doubly linked list. */
4609	Assert(pDevExt->pLdrImages);
4610	pImagePrev = NULL;
4611	if (pDevExt->pLdrImages != pImage)
4612	{
4613	pImagePrev = pDevExt->pLdrImages;
4614	while (pImagePrev->pNext != pImage)
4615	pImagePrev = pImagePrev->pNext;
4616	Assert(pImagePrev->pNext == pImage);
4617	}
4618
4619	/* unlink */
4620	if (pImagePrev)
4621	pImagePrev->pNext = pImage->pNext;
4622	else
4623	pDevExt->pLdrImages = pImage->pNext;
4624
4625	/* check if this is VMMR0.r0 unset its entry point pointers. */
4626	if (pDevExt->pvVMMR0 == pImage->pvImage)
4627	supdrvLdrUnsetVMMR0EPs(pDevExt);
4628
4629	/* check for objects with destructors in this image. (Shouldn't happen.) */
4630	if (pDevExt->pObjs)
4631	{
4632	unsigned cObjs = 0;
4633	PSUPDRVOBJ pObj;
4634	RTSpinlockAcquire(pDevExt->Spinlock);
4635	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4636	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4637	{
4638	pObj->pfnDestructor = NULL;
4639	cObjs++;
4640	}
4641	RTSpinlockRelease(pDevExt->Spinlock);
4642	if (cObjs)
4643	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
4644	}
4645
4646	/* call termination function if fully loaded. */
4647	if ( pImage->pfnModuleTerm
4648	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
4649	{
4650	LogFlow(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
4651	pImage->pfnModuleTerm(pImage);
4652	}
4653
4654	/* Inform the tracing component. */
4655	supdrvTracerModuleUnloading(pDevExt, pImage);
4656
4657	/* do native unload if appropriate. */
4658	if (pImage->fNative)
4659	supdrvOSLdrUnload(pDevExt, pImage);
4660
4661	/* free the image */
4662	pImage->cUsage = 0;
4663	pImage->pDevExt = NULL;
4664	pImage->pNext = NULL;
4665	pImage->uState = SUP_IOCTL_LDR_FREE;
4666	RTMemExecFree(pImage->pvImageAlloc, pImage->cbImageBits + 31);
4667	pImage->pvImageAlloc = NULL;
4668	RTMemFree(pImage->pachStrTab);
4669	pImage->pachStrTab = NULL;
4670	RTMemFree(pImage->paSymbols);
4671	pImage->paSymbols = NULL;
4672	RTMemFree(pImage);
4673	}
4674
4675
4676	/**
4677	* Acquires the loader lock.
4678	*
4679	* @returns IPRT status code.
4680	* @param pDevExt The device extension.
4681	*/
4682	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt)
4683	{
4684	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4685	int rc = RTSemMutexRequest(pDevExt->mtxLdr, RT_INDEFINITE_WAIT);
4686	#else
4687	int rc = RTSemFastMutexRequest(pDevExt->mtxLdr);
4688	#endif
4689	AssertRC(rc);
4690	return rc;
4691	}
4692
4693
4694	/**
4695	* Releases the loader lock.
4696	*
4697	* @returns IPRT status code.
4698	* @param pDevExt The device extension.
4699	*/
4700	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt)
4701	{
4702	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4703	return RTSemMutexRelease(pDevExt->mtxLdr);
4704	#else
4705	return RTSemFastMutexRelease(pDevExt->mtxLdr);
4706	#endif
4707	}
4708
4709
4710	/**
4711	* Implements the service call request.
4712	*
4713	* @returns VBox status code.
4714	* @param pDevExt The device extension.
4715	* @param pSession The calling session.
4716	* @param pReq The request packet, valid.
4717	*/
4718	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq)
4719	{
4720	#if !defined(RT_OS_WINDOWS) \|\| defined(DEBUG)
4721	int rc;
4722
4723	/*
4724	* Find the module first in the module referenced by the calling session.
4725	*/
4726	rc = supdrvLdrLock(pDevExt);
4727	if (RT_SUCCESS(rc))
4728	{
4729	PFNSUPR0SERVICEREQHANDLER pfnServiceReqHandler = NULL;
4730	PSUPDRVLDRUSAGE pUsage;
4731
4732	for (pUsage = pSession->pLdrUsage; pUsage; pUsage = pUsage->pNext)
4733	if ( pUsage->pImage->pfnServiceReqHandler
4734	&& !strcmp(pUsage->pImage->szName, pReq->u.In.szName))
4735	{
4736	pfnServiceReqHandler = pUsage->pImage->pfnServiceReqHandler;
4737	break;
4738	}
4739	supdrvLdrUnlock(pDevExt);
4740
4741	if (pfnServiceReqHandler)
4742	{
4743	/*
4744	* Call it.
4745	*/
4746	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
4747	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, NULL);
4748	else
4749	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0]);
4750	}
4751	else
4752	rc = VERR_SUPDRV_SERVICE_NOT_FOUND;
4753	}
4754
4755	/* log it */
4756	if ( RT_FAILURE(rc)
4757	&& rc != VERR_INTERRUPTED
4758	&& rc != VERR_TIMEOUT)
4759	Log(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4760	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4761	else
4762	Log4(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4763	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4764	return rc;
4765	#else /* RT_OS_WINDOWS && !DEBUG */
4766	return VERR_NOT_IMPLEMENTED;
4767	#endif /* RT_OS_WINDOWS && !DEBUG */
4768	}
4769
4770
4771	/**
4772	* Implements the logger settings request.
4773	*
4774	* @returns VBox status code.
4775	* @param pDevExt The device extension.
4776	* @param pSession The caller's session.
4777	* @param pReq The request.
4778	*/
4779	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq)
4780	{
4781	const char *pszGroup = &pReq->u.In.szStrings[pReq->u.In.offGroups];
4782	const char *pszFlags = &pReq->u.In.szStrings[pReq->u.In.offFlags];
4783	const char *pszDest = &pReq->u.In.szStrings[pReq->u.In.offDestination];
4784	PRTLOGGER pLogger = NULL;
4785	int rc;
4786
4787	/*
4788	* Some further validation.
4789	*/
4790	switch (pReq->u.In.fWhat)
4791	{
4792	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4793	case SUPLOGGERSETTINGS_WHAT_CREATE:
4794	break;
4795
4796	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4797	if (pszGroup \|\| pszFlags \|\| *pszDest)
4798	return VERR_INVALID_PARAMETER;
4799	if (pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_RELEASE)
4800	return VERR_ACCESS_DENIED;
4801	break;
4802
4803	default:
4804	return VERR_INTERNAL_ERROR;
4805	}
4806
4807	/*
4808	* Get the logger.
4809	*/
4810	switch (pReq->u.In.fWhich)
4811	{
4812	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4813	pLogger = RTLogGetDefaultInstance();
4814	break;
4815
4816	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4817	pLogger = RTLogRelDefaultInstance();
4818	break;
4819
4820	default:
4821	return VERR_INTERNAL_ERROR;
4822	}
4823
4824	/*
4825	* Do the job.
4826	*/
4827	switch (pReq->u.In.fWhat)
4828	{
4829	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4830	if (pLogger)
4831	{
4832	rc = RTLogFlags(pLogger, pszFlags);
4833	if (RT_SUCCESS(rc))
4834	rc = RTLogGroupSettings(pLogger, pszGroup);
4835	NOREF(pszDest);
4836	}
4837	else
4838	rc = VERR_NOT_FOUND;
4839	break;
4840
4841	case SUPLOGGERSETTINGS_WHAT_CREATE:
4842	{
4843	if (pLogger)
4844	rc = VERR_ALREADY_EXISTS;
4845	else
4846	{
4847	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
4848
4849	rc = RTLogCreate(&pLogger,
4850	0 /* fFlags */,
4851	pszGroup,
4852	pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_DEBUG
4853	? "VBOX_LOG"
4854	: "VBOX_RELEASE_LOG",
4855	RT_ELEMENTS(s_apszGroups),
4856	s_apszGroups,
4857	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER,
4858	NULL);
4859	if (RT_SUCCESS(rc))
4860	{
4861	rc = RTLogFlags(pLogger, pszFlags);
4862	NOREF(pszDest);
4863	if (RT_SUCCESS(rc))
4864	{
4865	switch (pReq->u.In.fWhich)
4866	{
4867	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4868	pLogger = RTLogSetDefaultInstance(pLogger);
4869	break;
4870	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4871	pLogger = RTLogRelSetDefaultInstance(pLogger);
4872	break;
4873	}
4874	}
4875	RTLogDestroy(pLogger);
4876	}
4877	}
4878	break;
4879	}
4880
4881	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4882	switch (pReq->u.In.fWhich)
4883	{
4884	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4885	pLogger = RTLogSetDefaultInstance(NULL);
4886	break;
4887	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4888	pLogger = RTLogRelSetDefaultInstance(NULL);
4889	break;
4890	}
4891	rc = RTLogDestroy(pLogger);
4892	break;
4893
4894	default:
4895	{
4896	rc = VERR_INTERNAL_ERROR;
4897	break;
4898	}
4899	}
4900
4901	return rc;
4902	}
4903
4904
4905	/**
4906	* Creates the GIP.
4907	*
4908	* @returns VBox status code.
4909	* @param pDevExt Instance data. GIP stuff may be updated.
4910	*/
4911	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
4912	{
4913	PSUPGLOBALINFOPAGE pGip;
4914	RTHCPHYS HCPhysGip;
4915	uint32_t u32SystemResolution;
4916	uint32_t u32Interval;
4917	unsigned cCpus;
4918	int rc;
4919
4920
4921	LogFlow(("supdrvGipCreate:\n"));
4922
4923	/* assert order */
4924	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
4925	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
4926	Assert(!pDevExt->pGipTimer);
4927
4928	/*
4929	* Check the CPU count.
4930	*/
4931	cCpus = RTMpGetArraySize();
4932	if ( cCpus > RTCPUSET_MAX_CPUS
4933	\|\| cCpus > 256 /ApicId is used for the mappings/)
4934	{
4935	SUPR0Printf("VBoxDrv: Too many CPUs (%u) for the GIP (max %u)\n", cCpus, RT_MIN(RTCPUSET_MAX_CPUS, 256));
4936	return VERR_TOO_MANY_CPUS;
4937	}
4938
4939	/*
4940	* Allocate a contiguous set of pages with a default kernel mapping.
4941	*/
4942	rc = RTR0MemObjAllocCont(&pDevExt->GipMemObj, RT_UOFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), false /fExecutable/);
4943	if (RT_FAILURE(rc))
4944	{
4945	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
4946	return rc;
4947	}
4948	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
4949	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
4950
4951	/*
4952	* Find a reasonable update interval and initialize the structure.
4953	*/
4954	u32Interval = u32SystemResolution = RTTimerGetSystemGranularity();
4955	while (u32Interval < 10000000 /* 10 ms */)
4956	u32Interval += u32SystemResolution;
4957
4958	supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), 1000000000 / u32Interval /=Hz/, cCpus);
4959
4960	/*
4961	* Create the timer.
4962	* If CPU_ALL isn't supported we'll have to fall back to synchronous mode.
4963	*/
4964	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
4965	{
4966	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, RTTIMER_FLAGS_CPU_ALL, supdrvGipAsyncTimer, pDevExt);
4967	if (rc == VERR_NOT_SUPPORTED)
4968	{
4969	OSDBGPRINT(("supdrvGipCreate: omni timer not supported, falling back to synchronous mode\n"));
4970	pGip->u32Mode = SUPGIPMODE_SYNC_TSC;
4971	}
4972	}
4973	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
4974	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0, supdrvGipSyncTimer, pDevExt);
4975	if (RT_SUCCESS(rc))
4976	{
4977	rc = RTMpNotificationRegister(supdrvGipMpEvent, pDevExt);
4978	if (RT_SUCCESS(rc))
4979	{
4980	rc = RTMpOnAll(supdrvGipInitOnCpu, pDevExt, pGip);
4981	if (RT_SUCCESS(rc))
4982	{
4983	/*
4984	* We're good.
4985	*/
4986	Log(("supdrvGipCreate: %u ns interval.\n", u32Interval));
4987	g_pSUPGlobalInfoPage = pGip;
4988	return VINF_SUCCESS;
4989	}
4990
4991	OSDBGPRINT(("supdrvGipCreate: RTMpOnAll failed with rc=%Rrc\n", rc));
4992	RTMpNotificationDeregister(supdrvGipMpEvent, pDevExt);
4993
4994	}
4995	else
4996	OSDBGPRINT(("supdrvGipCreate: failed to register MP event notfication. rc=%Rrc\n", rc));
4997	}
4998	else
4999	{
5000	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %u ns interval. rc=%Rrc\n", u32Interval, rc));
5001	Assert(!pDevExt->pGipTimer);
5002	}
5003	supdrvGipDestroy(pDevExt);
5004	return rc;
5005	}
5006
5007
5008	/**
5009	* Terminates the GIP.
5010	*
5011	* @param pDevExt Instance data. GIP stuff may be updated.
5012	*/
5013	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
5014	{
5015	int rc;
5016	#ifdef DEBUG_DARWIN_GIP
5017	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
5018	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
5019	pDevExt->pGipTimer, pDevExt->GipMemObj));
5020	#endif
5021
5022	/*
5023	* Invalid the GIP data.
5024	*/
5025	if (pDevExt->pGip)
5026	{
5027	supdrvGipTerm(pDevExt->pGip);
5028	pDevExt->pGip = NULL;
5029	}
5030	g_pSUPGlobalInfoPage = NULL;
5031
5032	/*
5033	* Destroy the timer and free the GIP memory object.
5034	*/
5035	if (pDevExt->pGipTimer)
5036	{
5037	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
5038	pDevExt->pGipTimer = NULL;
5039	}
5040
5041	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
5042	{
5043	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
5044	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
5045	}
5046
5047	/*
5048	* Finally, make sure we've release the system timer resolution request
5049	* if one actually succeeded and is still pending.
5050	*/
5051	if (pDevExt->u32SystemTimerGranularityGrant)
5052	{
5053	rc = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
5054	pDevExt->u32SystemTimerGranularityGrant = 0;
5055	}
5056	}
5057
5058
5059	/**
5060	* Timer callback function sync GIP mode.
5061	* @param pTimer The timer.
5062	* @param pvUser The device extension.
5063	*/
5064	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
5065	{
5066	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
5067	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5068	uint64_t u64TSC = ASMReadTSC();
5069	uint64_t NanoTS = RTTimeSystemNanoTS();
5070
5071	supdrvGipUpdate(pDevExt, NanoTS, u64TSC, NIL_RTCPUID, iTick);
5072
5073	ASMSetFlags(fOldFlags);
5074	}
5075
5076
5077	/**
5078	* Timer callback function for async GIP mode.
5079	* @param pTimer The timer.
5080	* @param pvUser The device extension.
5081	*/
5082	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
5083	{
5084	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
5085	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5086	RTCPUID idCpu = RTMpCpuId();
5087	uint64_t u64TSC = ASMReadTSC();
5088	uint64_t NanoTS = RTTimeSystemNanoTS();
5089
5090	/** @todo reset the transaction number and whatnot when iTick == 1. */
5091	if (pDevExt->idGipMaster == idCpu)
5092	supdrvGipUpdate(pDevExt, NanoTS, u64TSC, idCpu, iTick);
5093	else
5094	supdrvGipUpdatePerCpu(pDevExt, NanoTS, u64TSC, idCpu, ASMGetApicId(), iTick);
5095
5096	ASMSetFlags(fOldFlags);
5097	}
5098
5099
5100	/**
5101	* Finds our (@a idCpu) entry, or allocates a new one if not found.
5102	*
5103	* @returns Index of the CPU in the cache set.
5104	* @param pGip The GIP.
5105	* @param idCpu The CPU ID.
5106	*/
5107	static uint32_t supdrvGipCpuIndexFromCpuId(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
5108	{
5109	uint32_t i, cTries;
5110
5111	/*
5112	* ASSUMES that CPU IDs are constant.
5113	*/
5114	for (i = 0; i < pGip->cCpus; i++)
5115	if (pGip->aCPUs[i].idCpu == idCpu)
5116	return i;
5117
5118	cTries = 0;
5119	do
5120	{
5121	for (i = 0; i < pGip->cCpus; i++)
5122	{
5123	bool fRc;
5124	ASMAtomicCmpXchgSize(&pGip->aCPUs[i].idCpu, idCpu, NIL_RTCPUID, fRc);
5125	if (fRc)
5126	return i;
5127	}
5128	} while (cTries++ < 32);
5129	AssertReleaseFailed();
5130	return i - 1;
5131	}
5132
5133
5134	/**
5135	* The calling CPU should be accounted as online, update GIP accordingly.
5136	*
5137	* This is used by supdrvGipMpEvent as well as the supdrvGipCreate.
5138	*
5139	* @param pDevExt The device extension.
5140	* @param idCpu The CPU ID.
5141	*/
5142	static void supdrvGipMpEventOnline(PSUPDRVDEVEXT pDevExt, RTCPUID idCpu)
5143	{
5144	int iCpuSet = 0;
5145	uint16_t idApic = UINT16_MAX;
5146	uint32_t i = 0;
5147	uint64_t u64NanoTS = 0;
5148	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5149
5150	AssertPtrReturnVoid(pGip);
5151	AssertRelease(idCpu == RTMpCpuId());
5152	Assert(pGip->cPossibleCpus == RTMpGetCount());
5153
5154	/*
5155	* Do this behind a spinlock with interrupts disabled as this can fire
5156	* on all CPUs simultaneously, see #6110.
5157	*/
5158	RTSpinlockAcquire(pDevExt->hGipSpinlock);
5159
5160	/*
5161	* Update the globals.
5162	*/
5163	ASMAtomicWriteU16(&pGip->cPresentCpus, RTMpGetPresentCount());
5164	ASMAtomicWriteU16(&pGip->cOnlineCpus, RTMpGetOnlineCount());
5165	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5166	if (iCpuSet >= 0)
5167	{
5168	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5169	RTCpuSetAddByIndex(&pGip->OnlineCpuSet, iCpuSet);
5170	RTCpuSetAddByIndex(&pGip->PresentCpuSet, iCpuSet);
5171	}
5172
5173	/*
5174	* Update the entry.
5175	*/
5176	u64NanoTS = RTTimeSystemNanoTS() - pGip->u32UpdateIntervalNS;
5177	i = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
5178	supdrvGipInitCpu(pGip, &pGip->aCPUs[i], u64NanoTS);
5179	idApic = ASMGetApicId();
5180	ASMAtomicWriteU16(&pGip->aCPUs[i].idApic, idApic);
5181	ASMAtomicWriteS16(&pGip->aCPUs[i].iCpuSet, (int16_t)iCpuSet);
5182	ASMAtomicWriteSize(&pGip->aCPUs[i].idCpu, idCpu);
5183
5184	/*
5185	* Update the APIC ID and CPU set index mappings.
5186	*/
5187	ASMAtomicWriteU16(&pGip->aiCpuFromApicId[idApic], i);
5188	ASMAtomicWriteU16(&pGip->aiCpuFromCpuSetIdx[iCpuSet], i);
5189
5190	/* commit it */
5191	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_ONLINE);
5192
5193	RTSpinlockReleaseNoInts(pDevExt->hGipSpinlock);
5194	}
5195
5196
5197	/**
5198	* The CPU should be accounted as offline, update the GIP accordingly.
5199	*
5200	* This is used by supdrvGipMpEvent.
5201	*
5202	* @param pDevExt The device extension.
5203	* @param idCpu The CPU ID.
5204	*/
5205	static void supdrvGipMpEventOffline(PSUPDRVDEVEXT pDevExt, RTCPUID idCpu)
5206	{
5207	int iCpuSet;
5208	unsigned i;
5209
5210	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5211
5212	AssertPtrReturnVoid(pGip);
5213	RTSpinlockAcquire(pDevExt->hGipSpinlock);
5214
5215	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5216	AssertReturnVoid(iCpuSet >= 0);
5217
5218	i = pGip->aiCpuFromCpuSetIdx[iCpuSet];
5219	AssertReturnVoid(i < pGip->cCpus);
5220	AssertReturnVoid(pGip->aCPUs[i].idCpu == idCpu);
5221
5222	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5223	RTCpuSetDelByIndex(&pGip->OnlineCpuSet, iCpuSet);
5224
5225	/* commit it */
5226	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_OFFLINE);
5227
5228	RTSpinlockReleaseNoInts(pDevExt->hGipSpinlock);
5229	}
5230
5231
5232	/**
5233	* Multiprocessor event notification callback.
5234	*
5235	* This is used to make sure that the GIP master gets passed on to
5236	* another CPU. It also updates the associated CPU data.
5237	*
5238	* @param enmEvent The event.
5239	* @param idCpu The cpu it applies to.
5240	* @param pvUser Pointer to the device extension.
5241	*
5242	* @remarks This function -must- fire on the newly online'd CPU for the
5243	* RTMPEVENT_ONLINE case and can fire on any CPU for the
5244	* RTMPEVENT_OFFLINE case.
5245	*/
5246	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser)
5247	{
5248	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5249	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5250
5251	AssertRelease(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
5252
5253	/*
5254	* Update the GIP CPU data.
5255	*/
5256	if (pGip)
5257	{
5258	switch (enmEvent)
5259	{
5260	case RTMPEVENT_ONLINE:
5261	AssertRelease(idCpu == RTMpCpuId());
5262	supdrvGipMpEventOnline(pDevExt, idCpu);
5263	break;
5264	case RTMPEVENT_OFFLINE:
5265	supdrvGipMpEventOffline(pDevExt, idCpu);
5266	break;
5267
5268	}
5269	}
5270
5271	/*
5272	* Make sure there is a master GIP.
5273	*/
5274	if (enmEvent == RTMPEVENT_OFFLINE)
5275	{
5276	RTCPUID idGipMaster = ASMAtomicReadU32(&pDevExt->idGipMaster);
5277	if (idGipMaster == idCpu)
5278	{
5279	/*
5280	* Find a new GIP master.
5281	*/
5282	bool fIgnored;
5283	unsigned i;
5284	RTCPUID idNewGipMaster = NIL_RTCPUID;
5285	RTCPUSET OnlineCpus;
5286	RTMpGetOnlineSet(&OnlineCpus);
5287
5288	for (i = 0; i < RTCPUSET_MAX_CPUS; i++)
5289	{
5290	RTCPUID idCurCpu = RTMpCpuIdFromSetIndex(i);
5291	if ( RTCpuSetIsMember(&OnlineCpus, idCurCpu)
5292	&& idCurCpu != idGipMaster)
5293	{
5294	idNewGipMaster = idCurCpu;
5295	break;
5296	}
5297	}
5298
5299	Log(("supdrvGipMpEvent: Gip master %#lx -> %#lx\n", (long)idGipMaster, (long)idNewGipMaster));
5300	ASMAtomicCmpXchgSize(&pDevExt->idGipMaster, idNewGipMaster, idGipMaster, fIgnored);
5301	NOREF(fIgnored);
5302	}
5303	}
5304	}
5305
5306
5307	/**
5308	* Callback used by supdrvDetermineAsyncTSC to read the TSC on a CPU.
5309	*
5310	* @param idCpu Ignored.
5311	* @param pvUser1 Where to put the TSC.
5312	* @param pvUser2 Ignored.
5313	*/
5314	static DECLCALLBACK(void) supdrvDetermineAsyncTscWorker(RTCPUID idCpu, void pvUser1, void pvUser2)
5315	{
5316	#if 1
5317	ASMAtomicWriteU64((uint64_t volatile *)pvUser1, ASMReadTSC());
5318	#else
5319	(uint64_t )pvUser1 = ASMReadTSC();
5320	#endif
5321	}
5322
5323
5324	/**
5325	* Determine if Async GIP mode is required because of TSC drift.
5326	*
5327	* When using the default/normal timer code it is essential that the time stamp counter
5328	* (TSC) runs never backwards, that is, a read operation to the counter should return
5329	* a bigger value than any previous read operation. This is guaranteed by the latest
5330	* AMD CPUs and by newer Intel CPUs which never enter the C2 state (P4). In any other
5331	* case we have to choose the asynchronous timer mode.
5332	*
5333	* @param poffMin Pointer to the determined difference between different cores.
5334	* @return false if the time stamp counters appear to be synchronized, true otherwise.
5335	*/
5336	static bool supdrvDetermineAsyncTsc(uint64_t *poffMin)
5337	{
5338	/*
5339	* Just iterate all the cpus 8 times and make sure that the TSC is
5340	* ever increasing. We don't bother taking TSC rollover into account.
5341	*/
5342	int iEndCpu = RTMpGetArraySize();
5343	int iCpu;
5344	int cLoops = 8;
5345	bool fAsync = false;
5346	int rc = VINF_SUCCESS;
5347	uint64_t offMax = 0;
5348	uint64_t offMin = ~(uint64_t)0;
5349	uint64_t PrevTsc = ASMReadTSC();
5350
5351	while (cLoops-- > 0)
5352	{
5353	for (iCpu = 0; iCpu < iEndCpu; iCpu++)
5354	{
5355	uint64_t CurTsc;
5356	rc = RTMpOnSpecific(RTMpCpuIdFromSetIndex(iCpu), supdrvDetermineAsyncTscWorker, &CurTsc, NULL);
5357	if (RT_SUCCESS(rc))
5358	{
5359	if (CurTsc <= PrevTsc)
5360	{
5361	fAsync = true;
5362	offMin = offMax = PrevTsc - CurTsc;
5363	Log(("supdrvDetermineAsyncTsc: iCpu=%d cLoops=%d CurTsc=%llx PrevTsc=%llx\n",
5364	iCpu, cLoops, CurTsc, PrevTsc));
5365	break;
5366	}
5367
5368	/* Gather statistics (except the first time). */
5369	if (iCpu != 0 \|\| cLoops != 7)
5370	{
5371	uint64_t off = CurTsc - PrevTsc;
5372	if (off < offMin)
5373	offMin = off;
5374	if (off > offMax)
5375	offMax = off;
5376	Log2(("%d/%d: off=%llx\n", cLoops, iCpu, off));
5377	}
5378
5379	/* Next */
5380	PrevTsc = CurTsc;
5381	}
5382	else if (rc == VERR_NOT_SUPPORTED)
5383	break;
5384	else
5385	AssertMsg(rc == VERR_CPU_NOT_FOUND \|\| rc == VERR_CPU_OFFLINE, ("%d\n", rc));
5386	}
5387
5388	/* broke out of the loop. */
5389	if (iCpu < iEndCpu)
5390	break;
5391	}
5392
5393	poffMin = offMin; / Almost RTMpOnSpecific profiling. */
5394	Log(("supdrvDetermineAsyncTsc: returns %d; iEndCpu=%d rc=%d offMin=%llx offMax=%llx\n",
5395	fAsync, iEndCpu, rc, offMin, offMax));
5396	#if !defined(RT_OS_SOLARIS) && !defined(RT_OS_OS2) && !defined(RT_OS_WINDOWS)
5397	OSDBGPRINT(("vboxdrv: fAsync=%d offMin=%#lx offMax=%#lx\n", fAsync, (long)offMin, (long)offMax));
5398	#endif
5399	return fAsync;
5400	}
5401
5402
5403	/**
5404	* Determine the GIP TSC mode.
5405	*
5406	* @returns The most suitable TSC mode.
5407	* @param pDevExt Pointer to the device instance data.
5408	*/
5409	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt)
5410	{
5411	/*
5412	* On SMP we're faced with two problems:
5413	* (1) There might be a skew between the CPU, so that cpu0
5414	* returns a TSC that is slightly different from cpu1.
5415	* (2) Power management (and other things) may cause the TSC
5416	* to run at a non-constant speed, and cause the speed
5417	* to be different on the cpus. This will result in (1).
5418	*
5419	* So, on SMP systems we'll have to select the ASYNC update method
5420	* if there are symptoms of these problems.
5421	*/
5422	if (RTMpGetCount() > 1)
5423	{
5424	uint32_t uEAX, uEBX, uECX, uEDX;
5425	uint64_t u64DiffCoresIgnored;
5426
5427	/* Permit the user and/or the OS specific bits to force async mode. */
5428	if (supdrvOSGetForcedAsyncTscMode(pDevExt))
5429	return SUPGIPMODE_ASYNC_TSC;
5430
5431	/* Try check for current differences between the cpus. */
5432	if (supdrvDetermineAsyncTsc(&u64DiffCoresIgnored))
5433	return SUPGIPMODE_ASYNC_TSC;
5434
5435	/*
5436	* If the CPU supports power management and is an AMD one we
5437	* won't trust it unless it has the TscInvariant bit is set.
5438	*/
5439	/* Check for "AuthenticAMD" */
5440	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
5441	if ( uEAX >= 1
5442	&& uEBX == X86_CPUID_VENDOR_AMD_EBX
5443	&& uECX == X86_CPUID_VENDOR_AMD_ECX
5444	&& uEDX == X86_CPUID_VENDOR_AMD_EDX)
5445	{
5446	/* Check for APM support and that TscInvariant is cleared. */
5447	ASMCpuId(0x80000000, &uEAX, &uEBX, &uECX, &uEDX);
5448	if (uEAX >= 0x80000007)
5449	{
5450	ASMCpuId(0x80000007, &uEAX, &uEBX, &uECX, &uEDX);
5451	if ( !(uEDX & RT_BIT(8))/* TscInvariant */
5452	&& (uEDX & 0x3e)) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
5453	return SUPGIPMODE_ASYNC_TSC;
5454	}
5455	}
5456	}
5457	return SUPGIPMODE_SYNC_TSC;
5458	}
5459
5460
5461	/**
5462	* Initializes per-CPU GIP information.
5463	*
5464	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5465	* @param pCpu Pointer to which GIP CPU to initalize.
5466	* @param u64NanoTS The current nanosecond timestamp.
5467	*/
5468	static void supdrvGipInitCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pCpu, uint64_t u64NanoTS)
5469	{
5470	pCpu->u32TransactionId = 2;
5471	pCpu->u64NanoTS = u64NanoTS;
5472	pCpu->u64TSC = ASMReadTSC();
5473
5474	ASMAtomicWriteSize(&pCpu->enmState, SUPGIPCPUSTATE_INVALID);
5475	ASMAtomicWriteSize(&pCpu->idCpu, NIL_RTCPUID);
5476	ASMAtomicWriteS16(&pCpu->iCpuSet, -1);
5477	ASMAtomicWriteU16(&pCpu->idApic, UINT16_MAX);
5478
5479	/*
5480	* We don't know the following values until we've executed updates.
5481	* So, we'll just pretend it's a 4 GHz CPU and adjust the history it on
5482	* the 2nd timer callout.
5483	*/
5484	pCpu->u64CpuHz = _4G + 1; /* tstGIP-2 depends on this. */
5485	pCpu->u32UpdateIntervalTSC
5486	= pCpu->au32TSCHistory[0]
5487	= pCpu->au32TSCHistory[1]
5488	= pCpu->au32TSCHistory[2]
5489	= pCpu->au32TSCHistory[3]
5490	= pCpu->au32TSCHistory[4]
5491	= pCpu->au32TSCHistory[5]
5492	= pCpu->au32TSCHistory[6]
5493	= pCpu->au32TSCHistory[7]
5494	= (uint32_t)(_4G / pGip->u32UpdateHz);
5495	}
5496
5497
5498	/**
5499	* Initializes the GIP data.
5500	*
5501	* @param pDevExt Pointer to the device instance data.
5502	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5503	* @param HCPhys The physical address of the GIP.
5504	* @param u64NanoTS The current nanosecond timestamp.
5505	* @param uUpdateHz The update frequency.
5506	* @param cCpus The CPU count.
5507	*/
5508	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
5509	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned cCpus)
5510	{
5511	size_t const cbGip = RT_ALIGN_Z(RT_OFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), PAGE_SIZE);
5512	unsigned i;
5513	#ifdef DEBUG_DARWIN_GIP
5514	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5515	#else
5516	LogFlow(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5517	#endif
5518
5519	/*
5520	* Initialize the structure.
5521	*/
5522	memset(pGip, 0, cbGip);
5523	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
5524	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
5525	pGip->u32Mode = supdrvGipDeterminTscMode(pDevExt);
5526	pGip->cCpus = (uint16_t)cCpus;
5527	pGip->cPages = (uint16_t)(cbGip / PAGE_SIZE);
5528	pGip->u32UpdateHz = uUpdateHz;
5529	pGip->u32UpdateIntervalNS = 1000000000 / uUpdateHz;
5530	pGip->u64NanoTSLastUpdateHz = u64NanoTS;
5531	RTCpuSetEmpty(&pGip->OnlineCpuSet);
5532	RTCpuSetEmpty(&pGip->PresentCpuSet);
5533	RTMpGetSet(&pGip->PossibleCpuSet);
5534	pGip->cOnlineCpus = RTMpGetOnlineCount();
5535	pGip->cPresentCpus = RTMpGetPresentCount();
5536	pGip->cPossibleCpus = RTMpGetCount();
5537	pGip->idCpuMax = RTMpGetMaxCpuId();
5538	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromApicId); i++)
5539	pGip->aiCpuFromApicId[i] = 0;
5540	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx); i++)
5541	pGip->aiCpuFromCpuSetIdx[i] = UINT16_MAX;
5542
5543	for (i = 0; i < cCpus; i++)
5544	supdrvGipInitCpu(pGip, &pGip->aCPUs[i], u64NanoTS);
5545
5546	/*
5547	* Link it to the device extension.
5548	*/
5549	pDevExt->pGip = pGip;
5550	pDevExt->HCPhysGip = HCPhys;
5551	pDevExt->cGipUsers = 0;
5552	}
5553
5554
5555	/**
5556	* On CPU initialization callback for RTMpOnAll.
5557	*
5558	* @param idCpu The CPU ID.
5559	* @param pvUser1 The device extension.
5560	* @param pvUser2 The GIP.
5561	*/
5562	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2)
5563	{
5564	/* This is good enough, even though it will update some of the globals a
5565	bit to much. */
5566	supdrvGipMpEventOnline((PSUPDRVDEVEXT)pvUser1, idCpu);
5567	}
5568
5569
5570	/**
5571	* Invalidates the GIP data upon termination.
5572	*
5573	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5574	*/
5575	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
5576	{
5577	unsigned i;
5578	pGip->u32Magic = 0;
5579	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
5580	{
5581	pGip->aCPUs[i].u64NanoTS = 0;
5582	pGip->aCPUs[i].u64TSC = 0;
5583	pGip->aCPUs[i].iTSCHistoryHead = 0;
5584	}
5585	}
5586
5587
5588	/**
5589	* Worker routine for supdrvGipUpdate and supdrvGipUpdatePerCpu that
5590	* updates all the per cpu data except the transaction id.
5591	*
5592	* @param pDevExt The device extension.
5593	* @param pGipCpu Pointer to the per cpu data.
5594	* @param u64NanoTS The current time stamp.
5595	* @param u64TSC The current TSC.
5596	* @param iTick The current timer tick.
5597	*/
5598	static void supdrvGipDoUpdateCpu(PSUPDRVDEVEXT pDevExt, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick)
5599	{
5600	uint64_t u64TSCDelta;
5601	uint32_t u32UpdateIntervalTSC;
5602	uint32_t u32UpdateIntervalTSCSlack;
5603	unsigned iTSCHistoryHead;
5604	uint64_t u64CpuHz;
5605	uint32_t u32TransactionId;
5606
5607	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5608	AssertPtrReturnVoid(pGip);
5609
5610	/* Delta between this and the previous update. */
5611	ASMAtomicUoWriteU32(&pGipCpu->u32PrevUpdateIntervalNS, (uint32_t)(u64NanoTS - pGipCpu->u64NanoTS));
5612
5613	/*
5614	* Update the NanoTS.
5615	*/
5616	ASMAtomicWriteU64(&pGipCpu->u64NanoTS, u64NanoTS);
5617
5618	/*
5619	* Calc TSC delta.
5620	*/
5621	/** @todo validate the NanoTS delta, don't trust the OS to call us when it should... */
5622	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
5623	ASMAtomicWriteU64(&pGipCpu->u64TSC, u64TSC);
5624
5625	if (u64TSCDelta >> 32)
5626	{
5627	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
5628	pGipCpu->cErrors++;
5629	}
5630
5631	/*
5632	* On the 2nd and 3rd callout, reset the history with the current TSC
5633	* interval since the values entered by supdrvGipInit are totally off.
5634	* The interval on the 1st callout completely unreliable, the 2nd is a bit
5635	* better, while the 3rd should be most reliable.
5636	*/
5637	u32TransactionId = pGipCpu->u32TransactionId;
5638	if (RT_UNLIKELY( ( u32TransactionId == 5
5639	\|\| u32TransactionId == 7)
5640	&& ( iTick == 2
5641	\|\| iTick == 3) ))
5642	{
5643	unsigned i;
5644	for (i = 0; i < RT_ELEMENTS(pGipCpu->au32TSCHistory); i++)
5645	ASMAtomicUoWriteU32(&pGipCpu->au32TSCHistory[i], (uint32_t)u64TSCDelta);
5646	}
5647
5648	/*
5649	* TSC History.
5650	*/
5651	Assert(RT_ELEMENTS(pGipCpu->au32TSCHistory) == 8);
5652	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
5653	ASMAtomicWriteU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
5654	ASMAtomicWriteU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
5655
5656	/*
5657	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
5658	*/
5659	if (pGip->u32UpdateHz >= 1000)
5660	{
5661	uint32_t u32;
5662	u32 = pGipCpu->au32TSCHistory[0];
5663	u32 += pGipCpu->au32TSCHistory[1];
5664	u32 += pGipCpu->au32TSCHistory[2];
5665	u32 += pGipCpu->au32TSCHistory[3];
5666	u32 >>= 2;
5667	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
5668	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
5669	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
5670	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
5671	u32UpdateIntervalTSC >>= 2;
5672	u32UpdateIntervalTSC += u32;
5673	u32UpdateIntervalTSC >>= 1;
5674
5675	/* Value chosen for a 2GHz Athlon64 running linux 2.6.10/11, . */
5676	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
5677	}
5678	else if (pGip->u32UpdateHz >= 90)
5679	{
5680	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5681	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
5682	u32UpdateIntervalTSC >>= 1;
5683
5684	/* value chosen on a 2GHz thinkpad running windows */
5685	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
5686	}
5687	else
5688	{
5689	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5690
5691	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
5692	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
5693	}
5694	ASMAtomicWriteU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
5695
5696	/*
5697	* CpuHz.
5698	*/
5699	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, pGip->u32UpdateHz);
5700	ASMAtomicWriteU64(&pGipCpu->u64CpuHz, u64CpuHz);
5701	}
5702
5703
5704	/**
5705	* Updates the GIP.
5706	*
5707	* @param pDevExt The device extension.
5708	* @param u64NanoTS The current nanosecond timesamp.
5709	* @param u64TSC The current TSC timesamp.
5710	* @param idCpu The CPU ID.
5711	* @param iTick The current timer tick.
5712	*/
5713	static void supdrvGipUpdate(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick)
5714	{
5715	/*
5716	* Determine the relevant CPU data.
5717	*/
5718	PSUPGIPCPU pGipCpu;
5719	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5720	AssertPtrReturnVoid(pGip);
5721
5722	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
5723	pGipCpu = &pGip->aCPUs[0];
5724	else
5725	{
5726	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
5727	if (RT_UNLIKELY(iCpu >= pGip->cCpus))
5728	return;
5729	pGipCpu = &pGip->aCPUs[iCpu];
5730	if (RT_UNLIKELY(pGipCpu->idCpu != idCpu))
5731	return;
5732	}
5733
5734	/*
5735	* Start update transaction.
5736	*/
5737	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5738	{
5739	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
5740	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5741	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5742	pGipCpu->cErrors++;
5743	return;
5744	}
5745
5746	/*
5747	* Recalc the update frequency every 0x800th time.
5748	*/
5749	if (!(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
5750	{
5751	if (pGip->u64NanoTSLastUpdateHz)
5752	{
5753	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
5754	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
5755	uint32_t u32UpdateHz = (uint32_t)((UINT64_C(1000000000) * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
5756	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
5757	{
5758	ASMAtomicWriteU32(&pGip->u32UpdateHz, u32UpdateHz);
5759	ASMAtomicWriteU32(&pGip->u32UpdateIntervalNS, 1000000000 / u32UpdateHz);
5760	}
5761	#endif
5762	}
5763	ASMAtomicWriteU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS);
5764	}
5765
5766	/*
5767	* Update the data.
5768	*/
5769	supdrvGipDoUpdateCpu(pDevExt, pGipCpu, u64NanoTS, u64TSC, iTick);
5770
5771	/*
5772	* Complete transaction.
5773	*/
5774	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5775	}
5776
5777
5778	/**
5779	* Updates the per cpu GIP data for the calling cpu.
5780	*
5781	* @param pDevExt The device extension.
5782	* @param u64NanoTS The current nanosecond timesamp.
5783	* @param u64TSC The current TSC timesamp.
5784	* @param idCpu The CPU ID.
5785	* @param idApic The APIC id for the CPU index.
5786	* @param iTick The current timer tick.
5787	*/
5788	static void supdrvGipUpdatePerCpu(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC,
5789	RTCPUID idCpu, uint8_t idApic, uint64_t iTick)
5790	{
5791	uint32_t iCpu;
5792	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5793
5794	/*
5795	* Avoid a potential race when a CPU online notification doesn't fire on
5796	* the onlined CPU but the tick creeps in before the event notification is
5797	* run.
5798	*/
5799	if (RT_UNLIKELY(iTick == 1))
5800	{
5801	iCpu = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
5802	if (pGip->aCPUs[iCpu].enmState == SUPGIPCPUSTATE_OFFLINE)
5803	supdrvGipMpEventOnline(pDevExt, idCpu);
5804	}
5805
5806	iCpu = pGip->aiCpuFromApicId[idApic];
5807	if (RT_LIKELY(iCpu < pGip->cCpus))
5808	{
5809	PSUPGIPCPU pGipCpu = &pGip->aCPUs[iCpu];
5810	if (pGipCpu->idCpu == idCpu)
5811	{
5812	/*
5813	* Start update transaction.
5814	*/
5815	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5816	{
5817	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5818	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5819	pGipCpu->cErrors++;
5820	return;
5821	}
5822
5823	/*
5824	* Update the data.
5825	*/
5826	supdrvGipDoUpdateCpu(pDevExt, pGipCpu, u64NanoTS, u64TSC, iTick);
5827
5828	/*
5829	* Complete transaction.
5830	*/
5831	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5832	}
5833	}
5834	}
5835

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

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDrv.c@ 41147

Download in other formats: