SUPDrv.c@ 41254

Last change on this file since 41254 was 41254, checked in by vboxsync, 13 years ago
SUPDrv: Change SUPR0Printf to debug Log statement.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 217.7 KB

Line
1	/* $Id: SUPDrv.c 41254 2012-05-11 13:20:47Z 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	{ "RTR0MemKernelCopyFrom", (void *)RTR0MemKernelCopyFrom },
295	{ "RTR0MemKernelCopyTo", (void *)RTR0MemKernelCopyTo },
296	{ "RTR0MemObjAddress", (void *)RTR0MemObjAddress },
297	{ "RTR0MemObjAddressR3", (void *)RTR0MemObjAddressR3 },
298	{ "RTR0MemObjAllocContTag", (void *)RTR0MemObjAllocContTag },
299	{ "RTR0MemObjAllocLowTag", (void *)RTR0MemObjAllocLowTag },
300	{ "RTR0MemObjAllocPageTag", (void *)RTR0MemObjAllocPageTag },
301	{ "RTR0MemObjAllocPhysExTag", (void *)RTR0MemObjAllocPhysExTag },
302	{ "RTR0MemObjAllocPhysNCTag", (void *)RTR0MemObjAllocPhysNCTag },
303	{ "RTR0MemObjAllocPhysTag", (void *)RTR0MemObjAllocPhysTag },
304	{ "RTR0MemObjEnterPhysTag", (void *)RTR0MemObjEnterPhysTag },
305	{ "RTR0MemObjFree", (void *)RTR0MemObjFree },
306	{ "RTR0MemObjGetPagePhysAddr", (void *)RTR0MemObjGetPagePhysAddr },
307	{ "RTR0MemObjIsMapping", (void *)RTR0MemObjIsMapping },
308	{ "RTR0MemObjLockUserTag", (void *)RTR0MemObjLockUserTag },
309	{ "RTR0MemObjMapKernelExTag", (void *)RTR0MemObjMapKernelExTag },
310	{ "RTR0MemObjMapKernelTag", (void *)RTR0MemObjMapKernelTag },
311	{ "RTR0MemObjMapUserTag", (void *)RTR0MemObjMapUserTag },
312	{ "RTR0MemObjProtect", (void *)RTR0MemObjProtect },
313	{ "RTR0MemObjSize", (void *)RTR0MemObjSize },
314	{ "RTR0MemUserCopyFrom", (void *)RTR0MemUserCopyFrom },
315	{ "RTR0MemUserCopyTo", (void *)RTR0MemUserCopyTo },
316	{ "RTR0MemUserIsValidAddr", (void *)RTR0MemUserIsValidAddr },
317	{ "RTR0ProcHandleSelf", (void *)RTR0ProcHandleSelf },
318	{ "RTSemEventCreate", (void *)RTSemEventCreate },
319	{ "RTSemEventDestroy", (void *)RTSemEventDestroy },
320	{ "RTSemEventGetResolution", (void *)RTSemEventGetResolution },
321	{ "RTSemEventMultiCreate", (void *)RTSemEventMultiCreate },
322	{ "RTSemEventMultiDestroy", (void *)RTSemEventMultiDestroy },
323	{ "RTSemEventMultiGetResolution", (void *)RTSemEventMultiGetResolution },
324	{ "RTSemEventMultiReset", (void *)RTSemEventMultiReset },
325	{ "RTSemEventMultiSignal", (void *)RTSemEventMultiSignal },
326	{ "RTSemEventMultiWait", (void *)RTSemEventMultiWait },
327	{ "RTSemEventMultiWaitEx", (void *)RTSemEventMultiWaitEx },
328	{ "RTSemEventMultiWaitExDebug", (void *)RTSemEventMultiWaitExDebug },
329	{ "RTSemEventMultiWaitNoResume", (void *)RTSemEventMultiWaitNoResume },
330	{ "RTSemEventSignal", (void *)RTSemEventSignal },
331	{ "RTSemEventWait", (void *)RTSemEventWait },
332	{ "RTSemEventWaitEx", (void *)RTSemEventWaitEx },
333	{ "RTSemEventWaitExDebug", (void *)RTSemEventWaitExDebug },
334	{ "RTSemEventWaitNoResume", (void *)RTSemEventWaitNoResume },
335	{ "RTSemFastMutexCreate", (void *)RTSemFastMutexCreate },
336	{ "RTSemFastMutexDestroy", (void *)RTSemFastMutexDestroy },
337	{ "RTSemFastMutexRelease", (void *)RTSemFastMutexRelease },
338	{ "RTSemFastMutexRequest", (void *)RTSemFastMutexRequest },
339	{ "RTSemMutexCreate", (void *)RTSemMutexCreate },
340	{ "RTSemMutexDestroy", (void *)RTSemMutexDestroy },
341	{ "RTSemMutexRelease", (void *)RTSemMutexRelease },
342	{ "RTSemMutexRequest", (void *)RTSemMutexRequest },
343	{ "RTSemMutexRequestDebug", (void *)RTSemMutexRequestDebug },
344	{ "RTSemMutexRequestNoResume", (void *)RTSemMutexRequestNoResume },
345	{ "RTSemMutexRequestNoResumeDebug", (void *)RTSemMutexRequestNoResumeDebug },
346	{ "RTSpinlockAcquire", (void *)RTSpinlockAcquire },
347	{ "RTSpinlockCreate", (void *)RTSpinlockCreate },
348	{ "RTSpinlockDestroy", (void *)RTSpinlockDestroy },
349	{ "RTSpinlockRelease", (void *)RTSpinlockRelease },
350	{ "RTSpinlockReleaseNoInts", (void *)RTSpinlockReleaseNoInts },
351	{ "RTStrCopy", (void *)RTStrCopy },
352	{ "RTStrDupTag", (void *)RTStrDupTag },
353	{ "RTStrFormat", (void *)RTStrFormat },
354	{ "RTStrFormatNumber", (void *)RTStrFormatNumber },
355	{ "RTStrFormatTypeDeregister", (void *)RTStrFormatTypeDeregister },
356	{ "RTStrFormatTypeRegister", (void *)RTStrFormatTypeRegister },
357	{ "RTStrFormatTypeSetUser", (void *)RTStrFormatTypeSetUser },
358	{ "RTStrFormatV", (void *)RTStrFormatV },
359	{ "RTStrFree", (void *)RTStrFree },
360	{ "RTStrNCmp", (void *)RTStrNCmp },
361	{ "RTStrPrintf", (void *)RTStrPrintf },
362	{ "RTStrPrintfEx", (void *)RTStrPrintfEx },
363	{ "RTStrPrintfExV", (void *)RTStrPrintfExV },
364	{ "RTStrPrintfV", (void *)RTStrPrintfV },
365	{ "RTThreadCreate", (void *)RTThreadCreate },
366	{ "RTThreadGetName", (void *)RTThreadGetName },
367	{ "RTThreadGetNative", (void *)RTThreadGetNative },
368	{ "RTThreadGetType", (void *)RTThreadGetType },
369	{ "RTThreadIsInInterrupt", (void *)RTThreadIsInInterrupt },
370	{ "RTThreadNativeSelf", (void *)RTThreadNativeSelf },
371	{ "RTThreadPreemptDisable", (void *)RTThreadPreemptDisable },
372	{ "RTThreadPreemptIsEnabled", (void *)RTThreadPreemptIsEnabled },
373	{ "RTThreadPreemptIsPending", (void *)RTThreadPreemptIsPending },
374	{ "RTThreadPreemptIsPendingTrusty", (void *)RTThreadPreemptIsPendingTrusty },
375	{ "RTThreadPreemptIsPossible", (void *)RTThreadPreemptIsPossible },
376	{ "RTThreadPreemptRestore", (void *)RTThreadPreemptRestore },
377	{ "RTThreadSelf", (void *)RTThreadSelf },
378	{ "RTThreadSelfName", (void *)RTThreadSelfName },
379	{ "RTThreadSleep", (void *)RTThreadSleep },
380	{ "RTThreadUserReset", (void *)RTThreadUserReset },
381	{ "RTThreadUserSignal", (void *)RTThreadUserSignal },
382	{ "RTThreadUserWait", (void *)RTThreadUserWait },
383	{ "RTThreadUserWaitNoResume", (void *)RTThreadUserWaitNoResume },
384	{ "RTThreadWait", (void *)RTThreadWait },
385	{ "RTThreadWaitNoResume", (void *)RTThreadWaitNoResume },
386	{ "RTThreadYield", (void *)RTThreadYield },
387	{ "RTTimeMilliTS", (void *)RTTimeMilliTS },
388	{ "RTTimeNanoTS", (void *)RTTimeNanoTS },
389	{ "RTTimeNow", (void *)RTTimeNow },
390	{ "RTTimerCanDoHighResolution", (void *)RTTimerCanDoHighResolution },
391	{ "RTTimerChangeInterval", (void *)RTTimerChangeInterval },
392	{ "RTTimerCreate", (void *)RTTimerCreate },
393	{ "RTTimerCreateEx", (void *)RTTimerCreateEx },
394	{ "RTTimerDestroy", (void *)RTTimerDestroy },
395	{ "RTTimerGetSystemGranularity", (void *)RTTimerGetSystemGranularity },
396	{ "RTTimerReleaseSystemGranularity", (void *)RTTimerReleaseSystemGranularity },
397	{ "RTTimerRequestSystemGranularity", (void *)RTTimerRequestSystemGranularity },
398	{ "RTTimerStart", (void *)RTTimerStart },
399	{ "RTTimerStop", (void *)RTTimerStop },
400	{ "RTTimeSystemMilliTS", (void *)RTTimeSystemMilliTS },
401	{ "RTTimeSystemNanoTS", (void *)RTTimeSystemNanoTS },
402	{ "RTUuidCompare", (void *)RTUuidCompare },
403	{ "RTUuidCompareStr", (void *)RTUuidCompareStr },
404	{ "RTUuidFromStr", (void *)RTUuidFromStr },
405	/* SED: END */
406	};
407
408	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
409	/**
410	* Drag in the rest of IRPT since we share it with the
411	* rest of the kernel modules on darwin.
412	*/
413	PFNRT g_apfnVBoxDrvIPRTDeps[] =
414	{
415	/* VBoxNetAdp */
416	(PFNRT)RTRandBytes,
417	/* VBoxUSB */
418	(PFNRT)RTPathStripFilename,
419	NULL
420	};
421	#endif /* RT_OS_DARWIN \|\| RT_OS_SOLARIS \|\| RT_OS_SOLARIS */
422
423
424	/**
425	* Initializes the device extentsion structure.
426	*
427	* @returns IPRT status code.
428	* @param pDevExt The device extension to initialize.
429	* @param cbSession The size of the session structure. The size of
430	* SUPDRVSESSION may be smaller when SUPDRV_AGNOSTIC is
431	* defined because we're skipping the OS specific members
432	* then.
433	*/
434	int VBOXCALL supdrvInitDevExt(PSUPDRVDEVEXT pDevExt, size_t cbSession)
435	{
436	int rc;
437
438	#ifdef SUPDRV_WITH_RELEASE_LOGGER
439	/*
440	* Create the release log.
441	*/
442	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
443	PRTLOGGER pRelLogger;
444	rc = RTLogCreate(&pRelLogger, 0 /* fFlags */, "all",
445	"VBOX_RELEASE_LOG", RT_ELEMENTS(s_apszGroups), s_apszGroups, RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER, NULL);
446	if (RT_SUCCESS(rc))
447	RTLogRelSetDefaultInstance(pRelLogger);
448	/** @todo Add native hook for getting logger config parameters and setting
449	* them. On linux we should use the module parameter stuff... */
450	#endif
451
452	/*
453	* Initialize it.
454	*/
455	memset(pDevExt, 0, sizeof(*pDevExt));
456	rc = RTSpinlockCreate(&pDevExt->Spinlock, RTSPINLOCK_FLAGS_INTERRUPT_UNSAFE, "SUPDrvDevExt");
457	if (RT_SUCCESS(rc))
458	{
459	rc = RTSpinlockCreate(&pDevExt->hGipSpinlock, RTSPINLOCK_FLAGS_INTERRUPT_SAFE, "SUPDrvGip");
460	if (RT_SUCCESS(rc))
461	{
462	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
463	rc = RTSemMutexCreate(&pDevExt->mtxLdr);
464	#else
465	rc = RTSemFastMutexCreate(&pDevExt->mtxLdr);
466	#endif
467	if (RT_SUCCESS(rc))
468	{
469	rc = RTSemFastMutexCreate(&pDevExt->mtxComponentFactory);
470	if (RT_SUCCESS(rc))
471	{
472	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
473	rc = RTSemMutexCreate(&pDevExt->mtxGip);
474	#else
475	rc = RTSemFastMutexCreate(&pDevExt->mtxGip);
476	#endif
477	if (RT_SUCCESS(rc))
478	{
479	rc = supdrvGipCreate(pDevExt);
480	if (RT_SUCCESS(rc))
481	{
482	rc = supdrvTracerInit(pDevExt);
483	if (RT_SUCCESS(rc))
484	{
485	pDevExt->pLdrInitImage = NULL;
486	pDevExt->hLdrInitThread = NIL_RTNATIVETHREAD;
487	pDevExt->u32Cookie = BIRD; /** @todo make this random? */
488	pDevExt->cbSession = (uint32_t)cbSession;
489
490	/*
491	* Fixup the absolute symbols.
492	*
493	* Because of the table indexing assumptions we'll have a little #ifdef orgy
494	* here rather than distributing this to OS specific files. At least for now.
495	*/
496	#ifdef RT_OS_DARWIN
497	# if ARCH_BITS == 32
498	if (SUPR0GetPagingMode() >= SUPPAGINGMODE_AMD64)
499	{
500	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
501	g_aFunctions[1].pfn = (void )0x80; / SUPR0Abs64bitKernelCS - KERNEL64_CS, seg.h */
502	g_aFunctions[2].pfn = (void )0x88; / SUPR0Abs64bitKernelSS - KERNEL64_SS, seg.h */
503	g_aFunctions[3].pfn = (void )0x88; / SUPR0Abs64bitKernelDS - KERNEL64_SS, seg.h */
504	}
505	else
506	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
507	g_aFunctions[4].pfn = (void )0x08; / SUPR0AbsKernelCS - KERNEL_CS, seg.h */
508	g_aFunctions[5].pfn = (void )0x10; / SUPR0AbsKernelSS - KERNEL_DS, seg.h */
509	g_aFunctions[6].pfn = (void )0x10; / SUPR0AbsKernelDS - KERNEL_DS, seg.h */
510	g_aFunctions[7].pfn = (void )0x10; / SUPR0AbsKernelES - KERNEL_DS, seg.h */
511	g_aFunctions[8].pfn = (void )0x10; / SUPR0AbsKernelFS - KERNEL_DS, seg.h */
512	g_aFunctions[9].pfn = (void )0x48; / SUPR0AbsKernelGS - CPU_DATA_GS, seg.h */
513	# else /* 64-bit darwin: */
514	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
515	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
516	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
517	g_aFunctions[3].pfn = (void )0; / SUPR0Abs64bitKernelDS */
518	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
519	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
520	g_aFunctions[6].pfn = (void )0; / SUPR0AbsKernelDS */
521	g_aFunctions[7].pfn = (void )0; / SUPR0AbsKernelES */
522	g_aFunctions[8].pfn = (void )0; / SUPR0AbsKernelFS */
523	g_aFunctions[9].pfn = (void )0; / SUPR0AbsKernelGS */
524
525	# endif
526	#else /* !RT_OS_DARWIN */
527	# if ARCH_BITS == 64
528	g_aFunctions[0].pfn = (void )1; / SUPR0AbsIs64bit */
529	g_aFunctions[1].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0Abs64bitKernelCS */
530	g_aFunctions[2].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0Abs64bitKernelSS */
531	g_aFunctions[3].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0Abs64bitKernelDS */
532	# else
533	g_aFunctions[0].pfn = g_aFunctions[1].pfn = g_aFunctions[2].pfn = g_aFunctions[4].pfn = (void *)0;
534	# endif
535	g_aFunctions[4].pfn = (void )(uintptr_t)ASMGetCS(); / SUPR0AbsKernelCS */
536	g_aFunctions[5].pfn = (void )(uintptr_t)ASMGetSS(); / SUPR0AbsKernelSS */
537	g_aFunctions[6].pfn = (void )(uintptr_t)ASMGetDS(); / SUPR0AbsKernelDS */
538	g_aFunctions[7].pfn = (void )(uintptr_t)ASMGetES(); / SUPR0AbsKernelES */
539	g_aFunctions[8].pfn = (void )(uintptr_t)ASMGetFS(); / SUPR0AbsKernelFS */
540	g_aFunctions[9].pfn = (void )(uintptr_t)ASMGetGS(); / SUPR0AbsKernelGS */
541	#endif /* !RT_OS_DARWIN */
542	return VINF_SUCCESS;
543	}
544
545	supdrvGipDestroy(pDevExt);
546	}
547
548	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
549	RTSemMutexDestroy(pDevExt->mtxGip);
550	pDevExt->mtxGip = NIL_RTSEMMUTEX;
551	#else
552	RTSemFastMutexDestroy(pDevExt->mtxGip);
553	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
554	#endif
555	}
556	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
557	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
558	}
559	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
560	RTSemMutexDestroy(pDevExt->mtxLdr);
561	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
562	#else
563	RTSemFastMutexDestroy(pDevExt->mtxLdr);
564	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
565	#endif
566	}
567	RTSpinlockDestroy(pDevExt->hGipSpinlock);
568	pDevExt->hGipSpinlock = NIL_RTSPINLOCK;
569	}
570	RTSpinlockDestroy(pDevExt->Spinlock);
571	pDevExt->Spinlock = NIL_RTSPINLOCK;
572	}
573	#ifdef SUPDRV_WITH_RELEASE_LOGGER
574	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
575	RTLogDestroy(RTLogSetDefaultInstance(NULL));
576	#endif
577
578	return rc;
579	}
580
581
582	/**
583	* Delete the device extension (e.g. cleanup members).
584	*
585	* @param pDevExt The device extension to delete.
586	*/
587	void VBOXCALL supdrvDeleteDevExt(PSUPDRVDEVEXT pDevExt)
588	{
589	PSUPDRVOBJ pObj;
590	PSUPDRVUSAGE pUsage;
591
592	/*
593	* Kill mutexes and spinlocks.
594	*/
595	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
596	RTSemMutexDestroy(pDevExt->mtxGip);
597	pDevExt->mtxGip = NIL_RTSEMMUTEX;
598	#else
599	RTSemFastMutexDestroy(pDevExt->mtxGip);
600	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
601	#endif
602	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
603	RTSemMutexDestroy(pDevExt->mtxLdr);
604	pDevExt->mtxLdr = NIL_RTSEMMUTEX;
605	#else
606	RTSemFastMutexDestroy(pDevExt->mtxLdr);
607	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
608	#endif
609	RTSpinlockDestroy(pDevExt->Spinlock);
610	pDevExt->Spinlock = NIL_RTSPINLOCK;
611	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
612	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
613
614	/*
615	* Free lists.
616	*/
617	/* objects. */
618	pObj = pDevExt->pObjs;
619	Assert(!pObj); /* (can trigger on forced unloads) */
620	pDevExt->pObjs = NULL;
621	while (pObj)
622	{
623	void *pvFree = pObj;
624	pObj = pObj->pNext;
625	RTMemFree(pvFree);
626	}
627
628	/* usage records. */
629	pUsage = pDevExt->pUsageFree;
630	pDevExt->pUsageFree = NULL;
631	while (pUsage)
632	{
633	void *pvFree = pUsage;
634	pUsage = pUsage->pNext;
635	RTMemFree(pvFree);
636	}
637
638	/* kill the GIP. */
639	supdrvGipDestroy(pDevExt);
640	RTSpinlockDestroy(pDevExt->hGipSpinlock);
641	pDevExt->hGipSpinlock = NIL_RTSPINLOCK;
642
643	supdrvTracerTerm(pDevExt);
644
645	#ifdef SUPDRV_WITH_RELEASE_LOGGER
646	/* destroy the loggers. */
647	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
648	RTLogDestroy(RTLogSetDefaultInstance(NULL));
649	#endif
650	}
651
652
653	/**
654	* Create session.
655	*
656	* @returns IPRT status code.
657	* @param pDevExt Device extension.
658	* @param fUser Flag indicating whether this is a user or kernel session.
659	* @param ppSession Where to store the pointer to the session data.
660	*/
661	int VBOXCALL supdrvCreateSession(PSUPDRVDEVEXT pDevExt, bool fUser, PSUPDRVSESSION *ppSession)
662	{
663	/*
664	* Allocate memory for the session data.
665	*/
666	int rc;
667	PSUPDRVSESSION pSession = *ppSession = (PSUPDRVSESSION)RTMemAllocZ(pDevExt->cbSession);
668	if (pSession)
669	{
670	/* Initialize session data. */
671	rc = RTSpinlockCreate(&pSession->Spinlock, RTSPINLOCK_FLAGS_INTERRUPT_UNSAFE, "SUPDrvSession");
672	if (!rc)
673	{
674	rc = RTHandleTableCreateEx(&pSession->hHandleTable,
675	RTHANDLETABLE_FLAGS_LOCKED \| RTHANDLETABLE_FLAGS_CONTEXT,
676	1 /uBase/, 32768 /cMax/, supdrvSessionObjHandleRetain, pSession);
677	if (RT_SUCCESS(rc))
678	{
679	Assert(pSession->Spinlock != NIL_RTSPINLOCK);
680	pSession->pDevExt = pDevExt;
681	pSession->u32Cookie = BIRD_INV;
682	/*pSession->pLdrUsage = NULL;
683	pSession->pVM = NULL;
684	pSession->pUsage = NULL;
685	pSession->pGip = NULL;
686	pSession->fGipReferenced = false;
687	pSession->Bundle.cUsed = 0; */
688	pSession->Uid = NIL_RTUID;
689	pSession->Gid = NIL_RTGID;
690	if (fUser)
691	{
692	pSession->Process = RTProcSelf();
693	pSession->R0Process = RTR0ProcHandleSelf();
694	}
695	else
696	{
697	pSession->Process = NIL_RTPROCESS;
698	pSession->R0Process = NIL_RTR0PROCESS;
699	}
700	/pSession->uTracerData = 0;/
701	pSession->hTracerCaller = NIL_RTNATIVETHREAD;
702	RTListInit(&pSession->TpProviders);
703	/pSession->cTpProviders = 0;/
704	/pSession->cTpProbesFiring = 0;/
705	RTListInit(&pSession->TpUmods);
706	/RT_ZERO(pSession->apTpLookupTable);/
707
708	VBOXDRV_SESSION_CREATE(pSession, fUser);
709	LogFlow(("Created session %p initial cookie=%#x\n", pSession, pSession->u32Cookie));
710	return VINF_SUCCESS;
711	}
712
713	RTSpinlockDestroy(pSession->Spinlock);
714	}
715	RTMemFree(pSession);
716	*ppSession = NULL;
717	Log(("Failed to create spinlock, rc=%d!\n", rc));
718	}
719	else
720	rc = VERR_NO_MEMORY;
721
722	return rc;
723	}
724
725
726	/**
727	* Shared code for cleaning up a session.
728	*
729	* @param pDevExt Device extension.
730	* @param pSession Session data.
731	* This data will be freed by this routine.
732	*/
733	void VBOXCALL supdrvCloseSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
734	{
735	VBOXDRV_SESSION_CLOSE(pSession);
736
737	/*
738	* Cleanup the session first.
739	*/
740	supdrvCleanupSession(pDevExt, pSession);
741
742	/*
743	* Free the rest of the session stuff.
744	*/
745	RTSpinlockDestroy(pSession->Spinlock);
746	pSession->Spinlock = NIL_RTSPINLOCK;
747	pSession->pDevExt = NULL;
748	RTMemFree(pSession);
749	LogFlow(("supdrvCloseSession: returns\n"));
750	}
751
752
753	/**
754	* Shared code for cleaning up a session (but not quite freeing it).
755	*
756	* This is primarily intended for MAC OS X where we have to clean up the memory
757	* stuff before the file handle is closed.
758	*
759	* @param pDevExt Device extension.
760	* @param pSession Session data.
761	* This data will be freed by this routine.
762	*/
763	void VBOXCALL supdrvCleanupSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
764	{
765	int rc;
766	PSUPDRVBUNDLE pBundle;
767	LogFlow(("supdrvCleanupSession: pSession=%p\n", pSession));
768
769	/*
770	* Remove logger instances related to this session.
771	*/
772	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pSession);
773
774	/*
775	* Destroy the handle table.
776	*/
777	rc = RTHandleTableDestroy(pSession->hHandleTable, supdrvSessionObjHandleDelete, pSession);
778	AssertRC(rc);
779	pSession->hHandleTable = NIL_RTHANDLETABLE;
780
781	/*
782	* Release object references made in this session.
783	* In theory there should be noone racing us in this session.
784	*/
785	Log2(("release objects - start\n"));
786	if (pSession->pUsage)
787	{
788	PSUPDRVUSAGE pUsage;
789	RTSpinlockAcquire(pDevExt->Spinlock);
790
791	while ((pUsage = pSession->pUsage) != NULL)
792	{
793	PSUPDRVOBJ pObj = pUsage->pObj;
794	pSession->pUsage = pUsage->pNext;
795
796	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
797	if (pUsage->cUsage < pObj->cUsage)
798	{
799	pObj->cUsage -= pUsage->cUsage;
800	RTSpinlockRelease(pDevExt->Spinlock);
801	}
802	else
803	{
804	/* Destroy the object and free the record. */
805	if (pDevExt->pObjs == pObj)
806	pDevExt->pObjs = pObj->pNext;
807	else
808	{
809	PSUPDRVOBJ pObjPrev;
810	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
811	if (pObjPrev->pNext == pObj)
812	{
813	pObjPrev->pNext = pObj->pNext;
814	break;
815	}
816	Assert(pObjPrev);
817	}
818	RTSpinlockRelease(pDevExt->Spinlock);
819
820	Log(("supdrvCleanupSession: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
821	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
822	if (pObj->pfnDestructor)
823	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
824	RTMemFree(pObj);
825	}
826
827	/* free it and continue. */
828	RTMemFree(pUsage);
829
830	RTSpinlockAcquire(pDevExt->Spinlock);
831	}
832
833	RTSpinlockRelease(pDevExt->Spinlock);
834	AssertMsg(!pSession->pUsage, ("Some buster reregistered an object during desturction!\n"));
835	}
836	Log2(("release objects - done\n"));
837
838	/*
839	* Do tracer cleanups related to this session.
840	*/
841	Log2(("release tracer stuff - start\n"));
842	supdrvTracerCleanupSession(pDevExt, pSession);
843	Log2(("release tracer stuff - end\n"));
844
845	/*
846	* Release memory allocated in the session.
847	*
848	* We do not serialize this as we assume that the application will
849	* not allocated memory while closing the file handle object.
850	*/
851	Log2(("freeing memory:\n"));
852	pBundle = &pSession->Bundle;
853	while (pBundle)
854	{
855	PSUPDRVBUNDLE pToFree;
856	unsigned i;
857
858	/*
859	* Check and unlock all entries in the bundle.
860	*/
861	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
862	{
863	if (pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ)
864	{
865	Log2(("eType=%d pvR0=%p pvR3=%p cb=%ld\n", pBundle->aMem[i].eType, RTR0MemObjAddress(pBundle->aMem[i].MemObj),
866	(void *)RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3), (long)RTR0MemObjSize(pBundle->aMem[i].MemObj)));
867	if (pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ)
868	{
869	rc = RTR0MemObjFree(pBundle->aMem[i].MapObjR3, false);
870	AssertRC(rc); /** @todo figure out how to handle this. */
871	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
872	}
873	rc = RTR0MemObjFree(pBundle->aMem[i].MemObj, true /* fFreeMappings */);
874	AssertRC(rc); /** @todo figure out how to handle this. */
875	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
876	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
877	}
878	}
879
880	/*
881	* Advance and free previous bundle.
882	*/
883	pToFree = pBundle;
884	pBundle = pBundle->pNext;
885
886	pToFree->pNext = NULL;
887	pToFree->cUsed = 0;
888	if (pToFree != &pSession->Bundle)
889	RTMemFree(pToFree);
890	}
891	Log2(("freeing memory - done\n"));
892
893	/*
894	* Deregister component factories.
895	*/
896	RTSemFastMutexRequest(pDevExt->mtxComponentFactory);
897	Log2(("deregistering component factories:\n"));
898	if (pDevExt->pComponentFactoryHead)
899	{
900	PSUPDRVFACTORYREG pPrev = NULL;
901	PSUPDRVFACTORYREG pCur = pDevExt->pComponentFactoryHead;
902	while (pCur)
903	{
904	if (pCur->pSession == pSession)
905	{
906	/* unlink it */
907	PSUPDRVFACTORYREG pNext = pCur->pNext;
908	if (pPrev)
909	pPrev->pNext = pNext;
910	else
911	pDevExt->pComponentFactoryHead = pNext;
912
913	/* free it */
914	pCur->pNext = NULL;
915	pCur->pSession = NULL;
916	pCur->pFactory = NULL;
917	RTMemFree(pCur);
918
919	/* next */
920	pCur = pNext;
921	}
922	else
923	{
924	/* next */
925	pPrev = pCur;
926	pCur = pCur->pNext;
927	}
928	}
929	}
930	RTSemFastMutexRelease(pDevExt->mtxComponentFactory);
931	Log2(("deregistering component factories - done\n"));
932
933	/*
934	* Loaded images needs to be dereferenced and possibly freed up.
935	*/
936	supdrvLdrLock(pDevExt);
937	Log2(("freeing images:\n"));
938	if (pSession->pLdrUsage)
939	{
940	PSUPDRVLDRUSAGE pUsage = pSession->pLdrUsage;
941	pSession->pLdrUsage = NULL;
942	while (pUsage)
943	{
944	void *pvFree = pUsage;
945	PSUPDRVLDRIMAGE pImage = pUsage->pImage;
946	if (pImage->cUsage > pUsage->cUsage)
947	pImage->cUsage -= pUsage->cUsage;
948	else
949	supdrvLdrFree(pDevExt, pImage);
950	pUsage->pImage = NULL;
951	pUsage = pUsage->pNext;
952	RTMemFree(pvFree);
953	}
954	}
955	supdrvLdrUnlock(pDevExt);
956	Log2(("freeing images - done\n"));
957
958	/*
959	* Unmap the GIP.
960	*/
961	Log2(("umapping GIP:\n"));
962	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
963	{
964	SUPR0GipUnmap(pSession);
965	pSession->fGipReferenced = 0;
966	}
967	Log2(("umapping GIP - done\n"));
968	}
969
970
971	/**
972	* RTHandleTableDestroy callback used by supdrvCleanupSession.
973	*
974	* @returns IPRT status code, see SUPR0ObjAddRef.
975	* @param hHandleTable The handle table handle. Ignored.
976	* @param pvObj The object pointer.
977	* @param pvCtx Context, the handle type. Ignored.
978	* @param pvUser Session pointer.
979	*/
980	static DECLCALLBACK(int) supdrvSessionObjHandleRetain(RTHANDLETABLE hHandleTable, void pvObj, void pvCtx, void *pvUser)
981	{
982	NOREF(pvCtx);
983	NOREF(hHandleTable);
984	return SUPR0ObjAddRefEx(pvObj, (PSUPDRVSESSION)pvUser, true /fNoBlocking/);
985	}
986
987
988	/**
989	* RTHandleTableDestroy callback used by supdrvCleanupSession.
990	*
991	* @param hHandleTable The handle table handle. Ignored.
992	* @param h The handle value. Ignored.
993	* @param pvObj The object pointer.
994	* @param pvCtx Context, the handle type. Ignored.
995	* @param pvUser Session pointer.
996	*/
997	static DECLCALLBACK(void) supdrvSessionObjHandleDelete(RTHANDLETABLE hHandleTable, uint32_t h, void pvObj, void pvCtx, void *pvUser)
998	{
999	NOREF(pvCtx);
1000	NOREF(h);
1001	NOREF(hHandleTable);
1002	SUPR0ObjRelease(pvObj, (PSUPDRVSESSION)pvUser);
1003	}
1004
1005
1006	/**
1007	* Fast path I/O Control worker.
1008	*
1009	* @returns VBox status code that should be passed down to ring-3 unchanged.
1010	* @param uIOCtl Function number.
1011	* @param idCpu VMCPU id.
1012	* @param pDevExt Device extention.
1013	* @param pSession Session data.
1014	*/
1015	int VBOXCALL supdrvIOCtlFast(uintptr_t uIOCtl, VMCPUID idCpu, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
1016	{
1017	/*
1018	* We check the two prereqs after doing this only to allow the compiler to optimize things better.
1019	*/
1020	if (RT_LIKELY( RT_VALID_PTR(pSession)
1021	&& pSession->pVM
1022	&& pDevExt->pfnVMMR0EntryFast))
1023	{
1024	switch (uIOCtl)
1025	{
1026	case SUP_IOCTL_FAST_DO_RAW_RUN:
1027	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_RAW_RUN);
1028	break;
1029	case SUP_IOCTL_FAST_DO_HWACC_RUN:
1030	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_HWACC_RUN);
1031	break;
1032	case SUP_IOCTL_FAST_DO_NOP:
1033	pDevExt->pfnVMMR0EntryFast(pSession->pVM, idCpu, SUP_VMMR0_DO_NOP);
1034	break;
1035	default:
1036	return VERR_INTERNAL_ERROR;
1037	}
1038	return VINF_SUCCESS;
1039	}
1040	return VERR_INTERNAL_ERROR;
1041	}
1042
1043
1044	/**
1045	* Helper for supdrvIOCtl. Check if pszStr contains any character of pszChars.
1046	* We would use strpbrk here if this function would be contained in the RedHat kABI white
1047	* list, see http://www.kerneldrivers.org/RHEL5.
1048	*
1049	* @returns 1 if pszStr does contain any character of pszChars, 0 otherwise.
1050	* @param pszStr String to check
1051	* @param pszChars Character set
1052	*/
1053	static int supdrvCheckInvalidChar(const char pszStr, const char pszChars)
1054	{
1055	int chCur;
1056	while ((chCur = *pszStr++) != '\0')
1057	{
1058	int ch;
1059	const char *psz = pszChars;
1060	while ((ch = *psz++) != '\0')
1061	if (ch == chCur)
1062	return 1;
1063
1064	}
1065	return 0;
1066	}
1067
1068
1069
1070	/**
1071	* I/O Control inner worker (tracing reasons).
1072	*
1073	* @returns IPRT status code.
1074	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1075	*
1076	* @param uIOCtl Function number.
1077	* @param pDevExt Device extention.
1078	* @param pSession Session data.
1079	* @param pReqHdr The request header.
1080	*/
1081	static int supdrvIOCtlInner(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1082	{
1083	/*
1084	* Validation macros
1085	*/
1086	#define REQ_CHECK_SIZES_EX(Name, cbInExpect, cbOutExpect) \
1087	do { \
1088	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect) \|\| pReqHdr->cbOut != (cbOutExpect))) \
1089	{ \
1090	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld. cbOut=%ld expected %ld.\n", \
1091	(long)pReqHdr->cbIn, (long)(cbInExpect), (long)pReqHdr->cbOut, (long)(cbOutExpect))); \
1092	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1093	} \
1094	} while (0)
1095
1096	#define REQ_CHECK_SIZES(Name) REQ_CHECK_SIZES_EX(Name, Name ## _SIZE_IN, Name ## _SIZE_OUT)
1097
1098	#define REQ_CHECK_SIZE_IN(Name, cbInExpect) \
1099	do { \
1100	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect))) \
1101	{ \
1102	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld.\n", \
1103	(long)pReqHdr->cbIn, (long)(cbInExpect))); \
1104	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1105	} \
1106	} while (0)
1107
1108	#define REQ_CHECK_SIZE_OUT(Name, cbOutExpect) \
1109	do { \
1110	if (RT_UNLIKELY(pReqHdr->cbOut != (cbOutExpect))) \
1111	{ \
1112	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbOut=%ld expected %ld.\n", \
1113	(long)pReqHdr->cbOut, (long)(cbOutExpect))); \
1114	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1115	} \
1116	} while (0)
1117
1118	#define REQ_CHECK_EXPR(Name, expr) \
1119	do { \
1120	if (RT_UNLIKELY(!(expr))) \
1121	{ \
1122	OSDBGPRINT(( #Name ": %s\n", #expr)); \
1123	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1124	} \
1125	} while (0)
1126
1127	#define REQ_CHECK_EXPR_FMT(expr, fmt) \
1128	do { \
1129	if (RT_UNLIKELY(!(expr))) \
1130	{ \
1131	OSDBGPRINT( fmt ); \
1132	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1133	} \
1134	} while (0)
1135
1136	/*
1137	* The switch.
1138	*/
1139	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
1140	{
1141	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
1142	{
1143	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
1144	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
1145	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
1146	{
1147	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
1148	pReq->Hdr.rc = VERR_INVALID_MAGIC;
1149	return 0;
1150	}
1151
1152	#if 0
1153	/*
1154	* Call out to the OS specific code and let it do permission checks on the
1155	* client process.
1156	*/
1157	if (!supdrvOSValidateClientProcess(pDevExt, pSession))
1158	{
1159	pReq->u.Out.u32Cookie = 0xffffffff;
1160	pReq->u.Out.u32SessionCookie = 0xffffffff;
1161	pReq->u.Out.u32SessionVersion = 0xffffffff;
1162	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1163	pReq->u.Out.pSession = NULL;
1164	pReq->u.Out.cFunctions = 0;
1165	pReq->Hdr.rc = VERR_PERMISSION_DENIED;
1166	return 0;
1167	}
1168	#endif
1169
1170	/*
1171	* Match the version.
1172	* The current logic is very simple, match the major interface version.
1173	*/
1174	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
1175	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
1176	{
1177	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1178	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
1179	pReq->u.Out.u32Cookie = 0xffffffff;
1180	pReq->u.Out.u32SessionCookie = 0xffffffff;
1181	pReq->u.Out.u32SessionVersion = 0xffffffff;
1182	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1183	pReq->u.Out.pSession = NULL;
1184	pReq->u.Out.cFunctions = 0;
1185	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1186	return 0;
1187	}
1188
1189	/*
1190	* Fill in return data and be gone.
1191	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
1192	* u32SessionVersion <= u32ReqVersion!
1193	*/
1194	/** @todo Somehow validate the client and negotiate a secure cookie... */
1195	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
1196	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
1197	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
1198	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
1199	pReq->u.Out.pSession = pSession;
1200	pReq->u.Out.cFunctions = sizeof(g_aFunctions) / sizeof(g_aFunctions[0]);
1201	pReq->Hdr.rc = VINF_SUCCESS;
1202	return 0;
1203	}
1204
1205	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_QUERY_FUNCS(0)):
1206	{
1207	/* validate */
1208	PSUPQUERYFUNCS pReq = (PSUPQUERYFUNCS)pReqHdr;
1209	REQ_CHECK_SIZES_EX(SUP_IOCTL_QUERY_FUNCS, SUP_IOCTL_QUERY_FUNCS_SIZE_IN, SUP_IOCTL_QUERY_FUNCS_SIZE_OUT(RT_ELEMENTS(g_aFunctions)));
1210
1211	/* execute */
1212	pReq->u.Out.cFunctions = RT_ELEMENTS(g_aFunctions);
1213	memcpy(&pReq->u.Out.aFunctions[0], g_aFunctions, sizeof(g_aFunctions));
1214	pReq->Hdr.rc = VINF_SUCCESS;
1215	return 0;
1216	}
1217
1218	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_LOCK):
1219	{
1220	/* validate */
1221	PSUPPAGELOCK pReq = (PSUPPAGELOCK)pReqHdr;
1222	REQ_CHECK_SIZE_IN(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_IN);
1223	REQ_CHECK_SIZE_OUT(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_OUT(pReq->u.In.cPages));
1224	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.cPages > 0);
1225	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.pvR3 >= PAGE_SIZE);
1226
1227	/* execute */
1228	pReq->Hdr.rc = SUPR0LockMem(pSession, pReq->u.In.pvR3, pReq->u.In.cPages, &pReq->u.Out.aPages[0]);
1229	if (RT_FAILURE(pReq->Hdr.rc))
1230	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1231	return 0;
1232	}
1233
1234	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_UNLOCK):
1235	{
1236	/* validate */
1237	PSUPPAGEUNLOCK pReq = (PSUPPAGEUNLOCK)pReqHdr;
1238	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_UNLOCK);
1239
1240	/* execute */
1241	pReq->Hdr.rc = SUPR0UnlockMem(pSession, pReq->u.In.pvR3);
1242	return 0;
1243	}
1244
1245	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_ALLOC):
1246	{
1247	/* validate */
1248	PSUPCONTALLOC pReq = (PSUPCONTALLOC)pReqHdr;
1249	REQ_CHECK_SIZES(SUP_IOCTL_CONT_ALLOC);
1250
1251	/* execute */
1252	pReq->Hdr.rc = SUPR0ContAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.HCPhys);
1253	if (RT_FAILURE(pReq->Hdr.rc))
1254	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1255	return 0;
1256	}
1257
1258	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_FREE):
1259	{
1260	/* validate */
1261	PSUPCONTFREE pReq = (PSUPCONTFREE)pReqHdr;
1262	REQ_CHECK_SIZES(SUP_IOCTL_CONT_FREE);
1263
1264	/* execute */
1265	pReq->Hdr.rc = SUPR0ContFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1266	return 0;
1267	}
1268
1269	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_OPEN):
1270	{
1271	/* validate */
1272	PSUPLDROPEN pReq = (PSUPLDROPEN)pReqHdr;
1273	REQ_CHECK_SIZES(SUP_IOCTL_LDR_OPEN);
1274	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs > 0);
1275	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageWithTabs < 16*_1M);
1276	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1277	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits > 0);
1278	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImageBits < pReq->u.In.cbImageWithTabs);
1279	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.szName[0]);
1280	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1281	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, !supdrvCheckInvalidChar(pReq->u.In.szName, ";:()[]{}/\\\|&*%#@!~`\"'"));
1282	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, RTStrEnd(pReq->u.In.szFilename, sizeof(pReq->u.In.szFilename)));
1283
1284	/* execute */
1285	pReq->Hdr.rc = supdrvIOCtl_LdrOpen(pDevExt, pSession, pReq);
1286	return 0;
1287	}
1288
1289	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOAD):
1290	{
1291	/* validate */
1292	PSUPLDRLOAD pReq = (PSUPLDRLOAD)pReqHdr;
1293	REQ_CHECK_EXPR(Name, pReq->Hdr.cbIn >= sizeof(*pReq));
1294	REQ_CHECK_SIZES_EX(SUP_IOCTL_LDR_LOAD, SUP_IOCTL_LDR_LOAD_SIZE_IN(pReq->u.In.cbImageWithTabs), SUP_IOCTL_LDR_LOAD_SIZE_OUT);
1295	REQ_CHECK_EXPR(SUP_IOCTL_LDR_LOAD, pReq->u.In.cSymbols <= 16384);
1296	REQ_CHECK_EXPR_FMT( !pReq->u.In.cSymbols
1297	\|\| ( pReq->u.In.offSymbols < pReq->u.In.cbImageWithTabs
1298	&& pReq->u.In.offSymbols + pReq->u.In.cSymbols * sizeof(SUPLDRSYM) <= pReq->u.In.cbImageWithTabs),
1299	("SUP_IOCTL_LDR_LOAD: offSymbols=%#lx cSymbols=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offSymbols,
1300	(long)pReq->u.In.cSymbols, (long)pReq->u.In.cbImageWithTabs));
1301	REQ_CHECK_EXPR_FMT( !pReq->u.In.cbStrTab
1302	\|\| ( pReq->u.In.offStrTab < pReq->u.In.cbImageWithTabs
1303	&& pReq->u.In.offStrTab + pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs
1304	&& pReq->u.In.cbStrTab <= pReq->u.In.cbImageWithTabs),
1305	("SUP_IOCTL_LDR_LOAD: offStrTab=%#lx cbStrTab=%#lx cbImageWithTabs=%#lx\n", (long)pReq->u.In.offStrTab,
1306	(long)pReq->u.In.cbStrTab, (long)pReq->u.In.cbImageWithTabs));
1307
1308	if (pReq->u.In.cSymbols)
1309	{
1310	uint32_t i;
1311	PSUPLDRSYM paSyms = (PSUPLDRSYM)&pReq->u.In.abImage[pReq->u.In.offSymbols];
1312	for (i = 0; i < pReq->u.In.cSymbols; i++)
1313	{
1314	REQ_CHECK_EXPR_FMT(paSyms[i].offSymbol < pReq->u.In.cbImageWithTabs,
1315	("SUP_IOCTL_LDR_LOAD: sym #%ld: symb off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offSymbol, (long)pReq->u.In.cbImageWithTabs));
1316	REQ_CHECK_EXPR_FMT(paSyms[i].offName < pReq->u.In.cbStrTab,
1317	("SUP_IOCTL_LDR_LOAD: sym #%ld: name off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1318	REQ_CHECK_EXPR_FMT(RTStrEnd((char const *)&pReq->u.In.abImage[pReq->u.In.offStrTab + paSyms[i].offName],
1319	pReq->u.In.cbStrTab - paSyms[i].offName),
1320	("SUP_IOCTL_LDR_LOAD: sym #%ld: unterminated name! (%#lx / %#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImageWithTabs));
1321	}
1322	}
1323
1324	/* execute */
1325	pReq->Hdr.rc = supdrvIOCtl_LdrLoad(pDevExt, pSession, pReq);
1326	return 0;
1327	}
1328
1329	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_FREE):
1330	{
1331	/* validate */
1332	PSUPLDRFREE pReq = (PSUPLDRFREE)pReqHdr;
1333	REQ_CHECK_SIZES(SUP_IOCTL_LDR_FREE);
1334
1335	/* execute */
1336	pReq->Hdr.rc = supdrvIOCtl_LdrFree(pDevExt, pSession, pReq);
1337	return 0;
1338	}
1339
1340	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_GET_SYMBOL):
1341	{
1342	/* validate */
1343	PSUPLDRGETSYMBOL pReq = (PSUPLDRGETSYMBOL)pReqHdr;
1344	REQ_CHECK_SIZES(SUP_IOCTL_LDR_GET_SYMBOL);
1345	REQ_CHECK_EXPR(SUP_IOCTL_LDR_GET_SYMBOL, RTStrEnd(pReq->u.In.szSymbol, sizeof(pReq->u.In.szSymbol)));
1346
1347	/* execute */
1348	pReq->Hdr.rc = supdrvIOCtl_LdrGetSymbol(pDevExt, pSession, pReq);
1349	return 0;
1350	}
1351
1352	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0(0)):
1353	{
1354	/* validate */
1355	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1356	Log4(("SUP_IOCTL_CALL_VMMR0: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1357	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1358
1359	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_VMMR0_SIZE(0))
1360	{
1361	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(0), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(0));
1362
1363	/* execute */
1364	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1365	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg, pSession);
1366	else
1367	pReq->Hdr.rc = VERR_WRONG_ORDER;
1368	}
1369	else
1370	{
1371	PSUPVMMR0REQHDR pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1372	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR)),
1373	("SUP_IOCTL_CALL_VMMR0: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR))));
1374	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1375	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(pVMMReq->cbReq));
1376
1377	/* execute */
1378	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1379	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1380	else
1381	pReq->Hdr.rc = VERR_WRONG_ORDER;
1382	}
1383
1384	if ( RT_FAILURE(pReq->Hdr.rc)
1385	&& pReq->Hdr.rc != VERR_INTERRUPTED
1386	&& pReq->Hdr.rc != VERR_TIMEOUT)
1387	Log(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1388	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1389	else
1390	Log4(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1391	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1392	return 0;
1393	}
1394
1395	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0_BIG):
1396	{
1397	/* validate */
1398	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1399	PSUPVMMR0REQHDR pVMMReq;
1400	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1401	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1402
1403	pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1404	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR)),
1405	("SUP_IOCTL_CALL_VMMR0_BIG: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_BIG_SIZE(sizeof(SUPVMMR0REQHDR))));
1406	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0_BIG, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1407	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));
1408
1409	/* execute */
1410	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1411	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.idCpu, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1412	else
1413	pReq->Hdr.rc = VERR_WRONG_ORDER;
1414
1415	if ( RT_FAILURE(pReq->Hdr.rc)
1416	&& pReq->Hdr.rc != VERR_INTERRUPTED
1417	&& pReq->Hdr.rc != VERR_TIMEOUT)
1418	Log(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1419	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1420	else
1421	Log4(("SUP_IOCTL_CALL_VMMR0_BIG: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1422	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1423	return 0;
1424	}
1425
1426	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_PAGING_MODE):
1427	{
1428	/* validate */
1429	PSUPGETPAGINGMODE pReq = (PSUPGETPAGINGMODE)pReqHdr;
1430	REQ_CHECK_SIZES(SUP_IOCTL_GET_PAGING_MODE);
1431
1432	/* execute */
1433	pReq->Hdr.rc = VINF_SUCCESS;
1434	pReq->u.Out.enmMode = SUPR0GetPagingMode();
1435	return 0;
1436	}
1437
1438	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_ALLOC):
1439	{
1440	/* validate */
1441	PSUPLOWALLOC pReq = (PSUPLOWALLOC)pReqHdr;
1442	REQ_CHECK_EXPR(SUP_IOCTL_LOW_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_LOW_ALLOC_SIZE_IN);
1443	REQ_CHECK_SIZES_EX(SUP_IOCTL_LOW_ALLOC, SUP_IOCTL_LOW_ALLOC_SIZE_IN, SUP_IOCTL_LOW_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1444
1445	/* execute */
1446	pReq->Hdr.rc = SUPR0LowAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1447	if (RT_FAILURE(pReq->Hdr.rc))
1448	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1449	return 0;
1450	}
1451
1452	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_FREE):
1453	{
1454	/* validate */
1455	PSUPLOWFREE pReq = (PSUPLOWFREE)pReqHdr;
1456	REQ_CHECK_SIZES(SUP_IOCTL_LOW_FREE);
1457
1458	/* execute */
1459	pReq->Hdr.rc = SUPR0LowFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1460	return 0;
1461	}
1462
1463	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_MAP):
1464	{
1465	/* validate */
1466	PSUPGIPMAP pReq = (PSUPGIPMAP)pReqHdr;
1467	REQ_CHECK_SIZES(SUP_IOCTL_GIP_MAP);
1468
1469	/* execute */
1470	pReq->Hdr.rc = SUPR0GipMap(pSession, &pReq->u.Out.pGipR3, &pReq->u.Out.HCPhysGip);
1471	if (RT_SUCCESS(pReq->Hdr.rc))
1472	pReq->u.Out.pGipR0 = pDevExt->pGip;
1473	return 0;
1474	}
1475
1476	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_UNMAP):
1477	{
1478	/* validate */
1479	PSUPGIPUNMAP pReq = (PSUPGIPUNMAP)pReqHdr;
1480	REQ_CHECK_SIZES(SUP_IOCTL_GIP_UNMAP);
1481
1482	/* execute */
1483	pReq->Hdr.rc = SUPR0GipUnmap(pSession);
1484	return 0;
1485	}
1486
1487	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SET_VM_FOR_FAST):
1488	{
1489	/* validate */
1490	PSUPSETVMFORFAST pReq = (PSUPSETVMFORFAST)pReqHdr;
1491	REQ_CHECK_SIZES(SUP_IOCTL_SET_VM_FOR_FAST);
1492	REQ_CHECK_EXPR_FMT( !pReq->u.In.pVMR0
1493	\|\| ( VALID_PTR(pReq->u.In.pVMR0)
1494	&& !((uintptr_t)pReq->u.In.pVMR0 & (PAGE_SIZE - 1))),
1495	("SUP_IOCTL_SET_VM_FOR_FAST: pVMR0=%p!\n", pReq->u.In.pVMR0));
1496	/* execute */
1497	pSession->pVM = pReq->u.In.pVMR0;
1498	pReq->Hdr.rc = VINF_SUCCESS;
1499	return 0;
1500	}
1501
1502	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_ALLOC_EX):
1503	{
1504	/* validate */
1505	PSUPPAGEALLOCEX pReq = (PSUPPAGEALLOCEX)pReqHdr;
1506	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC_EX, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_EX_SIZE_IN);
1507	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));
1508	REQ_CHECK_EXPR_FMT(pReq->u.In.fKernelMapping \|\| pReq->u.In.fUserMapping,
1509	("SUP_IOCTL_PAGE_ALLOC_EX: No mapping requested!\n"));
1510	REQ_CHECK_EXPR_FMT(pReq->u.In.fUserMapping,
1511	("SUP_IOCTL_PAGE_ALLOC_EX: Must have user mapping!\n"));
1512	REQ_CHECK_EXPR_FMT(!pReq->u.In.fReserved0 && !pReq->u.In.fReserved1,
1513	("SUP_IOCTL_PAGE_ALLOC_EX: fReserved0=%d fReserved1=%d\n", pReq->u.In.fReserved0, pReq->u.In.fReserved1));
1514
1515	/* execute */
1516	pReq->Hdr.rc = SUPR0PageAllocEx(pSession, pReq->u.In.cPages, 0 /* fFlags */,
1517	pReq->u.In.fUserMapping ? &pReq->u.Out.pvR3 : NULL,
1518	pReq->u.In.fKernelMapping ? &pReq->u.Out.pvR0 : NULL,
1519	&pReq->u.Out.aPages[0]);
1520	if (RT_FAILURE(pReq->Hdr.rc))
1521	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1522	return 0;
1523	}
1524
1525	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_MAP_KERNEL):
1526	{
1527	/* validate */
1528	PSUPPAGEMAPKERNEL pReq = (PSUPPAGEMAPKERNEL)pReqHdr;
1529	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_MAP_KERNEL);
1530	REQ_CHECK_EXPR_FMT(!pReq->u.In.fFlags, ("SUP_IOCTL_PAGE_MAP_KERNEL: fFlags=%#x! MBZ\n", pReq->u.In.fFlags));
1531	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_MAP_KERNEL: offSub=%#x\n", pReq->u.In.offSub));
1532	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1533	("SUP_IOCTL_PAGE_MAP_KERNEL: cbSub=%#x\n", pReq->u.In.cbSub));
1534
1535	/* execute */
1536	pReq->Hdr.rc = SUPR0PageMapKernel(pSession, pReq->u.In.pvR3, pReq->u.In.offSub, pReq->u.In.cbSub,
1537	pReq->u.In.fFlags, &pReq->u.Out.pvR0);
1538	if (RT_FAILURE(pReq->Hdr.rc))
1539	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1540	return 0;
1541	}
1542
1543	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_PROTECT):
1544	{
1545	/* validate */
1546	PSUPPAGEPROTECT pReq = (PSUPPAGEPROTECT)pReqHdr;
1547	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_PROTECT);
1548	REQ_CHECK_EXPR_FMT(!(pReq->u.In.fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)),
1549	("SUP_IOCTL_PAGE_PROTECT: fProt=%#x!\n", pReq->u.In.fProt));
1550	REQ_CHECK_EXPR_FMT(!(pReq->u.In.offSub & PAGE_OFFSET_MASK), ("SUP_IOCTL_PAGE_PROTECT: offSub=%#x\n", pReq->u.In.offSub));
1551	REQ_CHECK_EXPR_FMT(pReq->u.In.cbSub && !(pReq->u.In.cbSub & PAGE_OFFSET_MASK),
1552	("SUP_IOCTL_PAGE_PROTECT: cbSub=%#x\n", pReq->u.In.cbSub));
1553
1554	/* execute */
1555	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);
1556	return 0;
1557	}
1558
1559	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_FREE):
1560	{
1561	/* validate */
1562	PSUPPAGEFREE pReq = (PSUPPAGEFREE)pReqHdr;
1563	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_FREE);
1564
1565	/* execute */
1566	pReq->Hdr.rc = SUPR0PageFree(pSession, pReq->u.In.pvR3);
1567	return 0;
1568	}
1569
1570	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_SERVICE(0)):
1571	{
1572	/* validate */
1573	PSUPCALLSERVICE pReq = (PSUPCALLSERVICE)pReqHdr;
1574	Log4(("SUP_IOCTL_CALL_SERVICE: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1575	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1576
1577	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
1578	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(0), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(0));
1579	else
1580	{
1581	PSUPR0SERVICEREQHDR pSrvReq = (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0];
1582	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR)),
1583	("SUP_IOCTL_CALL_SERVICE: cbIn=%#x < %#lx\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_SERVICE_SIZE(sizeof(SUPR0SERVICEREQHDR))));
1584	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, pSrvReq->u32Magic == SUPR0SERVICEREQHDR_MAGIC);
1585	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_SERVICE, SUP_IOCTL_CALL_SERVICE_SIZE_IN(pSrvReq->cbReq), SUP_IOCTL_CALL_SERVICE_SIZE_OUT(pSrvReq->cbReq));
1586	}
1587	REQ_CHECK_EXPR(SUP_IOCTL_CALL_SERVICE, RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)));
1588
1589	/* execute */
1590	pReq->Hdr.rc = supdrvIOCtl_CallServiceModule(pDevExt, pSession, pReq);
1591	return 0;
1592	}
1593
1594	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOGGER_SETTINGS(0)):
1595	{
1596	/* validate */
1597	PSUPLOGGERSETTINGS pReq = (PSUPLOGGERSETTINGS)pReqHdr;
1598	size_t cbStrTab;
1599	REQ_CHECK_SIZE_OUT(SUP_IOCTL_LOGGER_SETTINGS, SUP_IOCTL_LOGGER_SETTINGS_SIZE_OUT);
1600	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->Hdr.cbIn >= SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(1));
1601	cbStrTab = pReq->Hdr.cbIn - SUP_IOCTL_LOGGER_SETTINGS_SIZE_IN(0);
1602	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offGroups < cbStrTab);
1603	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offFlags < cbStrTab);
1604	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.offDestination < cbStrTab);
1605	REQ_CHECK_EXPR_FMT(pReq->u.In.szStrings[cbStrTab - 1] == '\0',
1606	("SUP_IOCTL_LOGGER_SETTINGS: cbIn=%#x cbStrTab=%#zx LastChar=%d\n",
1607	pReq->Hdr.cbIn, cbStrTab, pReq->u.In.szStrings[cbStrTab - 1]));
1608	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhich <= SUPLOGGERSETTINGS_WHICH_RELEASE);
1609	REQ_CHECK_EXPR(SUP_IOCTL_LOGGER_SETTINGS, pReq->u.In.fWhat <= SUPLOGGERSETTINGS_WHAT_DESTROY);
1610
1611	/* execute */
1612	pReq->Hdr.rc = supdrvIOCtl_LoggerSettings(pDevExt, pSession, pReq);
1613	return 0;
1614	}
1615
1616	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP2):
1617	{
1618	/* validate */
1619	PSUPSEMOP2 pReq = (PSUPSEMOP2)pReqHdr;
1620	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP2, SUP_IOCTL_SEM_OP2_SIZE_IN, SUP_IOCTL_SEM_OP2_SIZE_OUT);
1621	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP2, pReq->u.In.uReserved == 0);
1622
1623	/* execute */
1624	switch (pReq->u.In.uType)
1625	{
1626	case SUP_SEM_TYPE_EVENT:
1627	{
1628	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1629	switch (pReq->u.In.uOp)
1630	{
1631	case SUPSEMOP2_WAIT_MS_REL:
1632	pReq->Hdr.rc = SUPSemEventWaitNoResume(pSession, hEvent, pReq->u.In.uArg.cRelMsTimeout);
1633	break;
1634	case SUPSEMOP2_WAIT_NS_ABS:
1635	pReq->Hdr.rc = SUPSemEventWaitNsAbsIntr(pSession, hEvent, pReq->u.In.uArg.uAbsNsTimeout);
1636	break;
1637	case SUPSEMOP2_WAIT_NS_REL:
1638	pReq->Hdr.rc = SUPSemEventWaitNsRelIntr(pSession, hEvent, pReq->u.In.uArg.cRelNsTimeout);
1639	break;
1640	case SUPSEMOP2_SIGNAL:
1641	pReq->Hdr.rc = SUPSemEventSignal(pSession, hEvent);
1642	break;
1643	case SUPSEMOP2_CLOSE:
1644	pReq->Hdr.rc = SUPSemEventClose(pSession, hEvent);
1645	break;
1646	case SUPSEMOP2_RESET:
1647	default:
1648	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1649	break;
1650	}
1651	break;
1652	}
1653
1654	case SUP_SEM_TYPE_EVENT_MULTI:
1655	{
1656	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1657	switch (pReq->u.In.uOp)
1658	{
1659	case SUPSEMOP2_WAIT_MS_REL:
1660	pReq->Hdr.rc = SUPSemEventMultiWaitNoResume(pSession, hEventMulti, pReq->u.In.uArg.cRelMsTimeout);
1661	break;
1662	case SUPSEMOP2_WAIT_NS_ABS:
1663	pReq->Hdr.rc = SUPSemEventMultiWaitNsAbsIntr(pSession, hEventMulti, pReq->u.In.uArg.uAbsNsTimeout);
1664	break;
1665	case SUPSEMOP2_WAIT_NS_REL:
1666	pReq->Hdr.rc = SUPSemEventMultiWaitNsRelIntr(pSession, hEventMulti, pReq->u.In.uArg.cRelNsTimeout);
1667	break;
1668	case SUPSEMOP2_SIGNAL:
1669	pReq->Hdr.rc = SUPSemEventMultiSignal(pSession, hEventMulti);
1670	break;
1671	case SUPSEMOP2_CLOSE:
1672	pReq->Hdr.rc = SUPSemEventMultiClose(pSession, hEventMulti);
1673	break;
1674	case SUPSEMOP2_RESET:
1675	pReq->Hdr.rc = SUPSemEventMultiReset(pSession, hEventMulti);
1676	break;
1677	default:
1678	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1679	break;
1680	}
1681	break;
1682	}
1683
1684	default:
1685	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1686	break;
1687	}
1688	return 0;
1689	}
1690
1691	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SEM_OP3):
1692	{
1693	/* validate */
1694	PSUPSEMOP3 pReq = (PSUPSEMOP3)pReqHdr;
1695	REQ_CHECK_SIZES_EX(SUP_IOCTL_SEM_OP3, SUP_IOCTL_SEM_OP3_SIZE_IN, SUP_IOCTL_SEM_OP3_SIZE_OUT);
1696	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, pReq->u.In.u32Reserved == 0 && pReq->u.In.u64Reserved == 0);
1697
1698	/* execute */
1699	switch (pReq->u.In.uType)
1700	{
1701	case SUP_SEM_TYPE_EVENT:
1702	{
1703	SUPSEMEVENT hEvent = (SUPSEMEVENT)(uintptr_t)pReq->u.In.hSem;
1704	switch (pReq->u.In.uOp)
1705	{
1706	case SUPSEMOP3_CREATE:
1707	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1708	pReq->Hdr.rc = SUPSemEventCreate(pSession, &hEvent);
1709	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEvent;
1710	break;
1711	case SUPSEMOP3_GET_RESOLUTION:
1712	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEvent == NIL_SUPSEMEVENT);
1713	pReq->Hdr.rc = VINF_SUCCESS;
1714	pReq->Hdr.cbOut = sizeof(*pReq);
1715	pReq->u.Out.cNsResolution = SUPSemEventGetResolution(pSession);
1716	break;
1717	default:
1718	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1719	break;
1720	}
1721	break;
1722	}
1723
1724	case SUP_SEM_TYPE_EVENT_MULTI:
1725	{
1726	SUPSEMEVENTMULTI hEventMulti = (SUPSEMEVENTMULTI)(uintptr_t)pReq->u.In.hSem;
1727	switch (pReq->u.In.uOp)
1728	{
1729	case SUPSEMOP3_CREATE:
1730	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1731	pReq->Hdr.rc = SUPSemEventMultiCreate(pSession, &hEventMulti);
1732	pReq->u.Out.hSem = (uint32_t)(uintptr_t)hEventMulti;
1733	break;
1734	case SUPSEMOP3_GET_RESOLUTION:
1735	REQ_CHECK_EXPR(SUP_IOCTL_SEM_OP3, hEventMulti == NIL_SUPSEMEVENTMULTI);
1736	pReq->Hdr.rc = VINF_SUCCESS;
1737	pReq->u.Out.cNsResolution = SUPSemEventMultiGetResolution(pSession);
1738	break;
1739	default:
1740	pReq->Hdr.rc = VERR_INVALID_FUNCTION;
1741	break;
1742	}
1743	break;
1744	}
1745
1746	default:
1747	pReq->Hdr.rc = VERR_INVALID_PARAMETER;
1748	break;
1749	}
1750	return 0;
1751	}
1752
1753	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_VT_CAPS):
1754	{
1755	/* validate */
1756	PSUPVTCAPS pReq = (PSUPVTCAPS)pReqHdr;
1757	REQ_CHECK_SIZES(SUP_IOCTL_VT_CAPS);
1758	REQ_CHECK_EXPR(SUP_IOCTL_VT_CAPS, pReq->Hdr.cbIn <= SUP_IOCTL_VT_CAPS_SIZE_IN);
1759
1760	/* execute */
1761	pReq->Hdr.rc = SUPR0QueryVTCaps(pSession, &pReq->u.Out.Caps);
1762	if (RT_FAILURE(pReq->Hdr.rc))
1763	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1764	return 0;
1765	}
1766
1767	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_OPEN):
1768	{
1769	/* validate */
1770	PSUPTRACEROPEN pReq = (PSUPTRACEROPEN)pReqHdr;
1771	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_OPEN);
1772
1773	/* execute */
1774	pReq->Hdr.rc = supdrvIOCtl_TracerOpen(pDevExt, pSession, pReq->u.In.uCookie, pReq->u.In.uArg);
1775	return 0;
1776	}
1777
1778	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_CLOSE):
1779	{
1780	/* validate */
1781	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_CLOSE);
1782
1783	/* execute */
1784	pReqHdr->rc = supdrvIOCtl_TracerClose(pDevExt, pSession);
1785	return 0;
1786	}
1787
1788	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_IOCTL):
1789	{
1790	/* validate */
1791	PSUPTRACERIOCTL pReq = (PSUPTRACERIOCTL)pReqHdr;
1792	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_IOCTL);
1793
1794	/* execute */
1795	pReqHdr->rc = supdrvIOCtl_TracerIOCtl(pDevExt, pSession, pReq->u.In.uCmd, pReq->u.In.uArg, &pReq->u.Out.iRetVal);
1796	return 0;
1797	}
1798
1799	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_REG):
1800	{
1801	/* validate */
1802	PSUPTRACERUMODREG pReq = (PSUPTRACERUMODREG)pReqHdr;
1803	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_REG);
1804	if (!RTStrEnd(pReq->u.In.szName, sizeof(pReq->u.In.szName)))
1805	return VERR_INVALID_PARAMETER;
1806
1807	/* execute */
1808	pReqHdr->rc = supdrvIOCtl_TracerUmodRegister(pDevExt, pSession,
1809	pReq->u.In.R3PtrVtgHdr, pReq->u.In.uVtgHdrAddr,
1810	pReq->u.In.R3PtrStrTab, pReq->u.In.cbStrTab,
1811	pReq->u.In.szName, pReq->u.In.fFlags);
1812	return 0;
1813	}
1814
1815	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_DEREG):
1816	{
1817	/* validate */
1818	PSUPTRACERUMODDEREG pReq = (PSUPTRACERUMODDEREG)pReqHdr;
1819	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_DEREG);
1820
1821	/* execute */
1822	pReqHdr->rc = supdrvIOCtl_TracerUmodDeregister(pDevExt, pSession, pReq->u.In.pVtgHdr);
1823	return 0;
1824	}
1825
1826	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_TRACER_UMOD_FIRE_PROBE):
1827	{
1828	/* validate */
1829	PSUPTRACERUMODFIREPROBE pReq = (PSUPTRACERUMODFIREPROBE)pReqHdr;
1830	REQ_CHECK_SIZES(SUP_IOCTL_TRACER_UMOD_FIRE_PROBE);
1831
1832	supdrvIOCtl_TracerUmodProbeFire(pDevExt, pSession, &pReq->u.In);
1833	pReqHdr->rc = VINF_SUCCESS;
1834	return 0;
1835	}
1836
1837	default:
1838	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1839	break;
1840	}
1841	return VERR_GENERAL_FAILURE;
1842	}
1843
1844
1845	/**
1846	* I/O Control worker.
1847	*
1848	* @returns IPRT status code.
1849	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1850	*
1851	* @param uIOCtl Function number.
1852	* @param pDevExt Device extention.
1853	* @param pSession Session data.
1854	* @param pReqHdr The request header.
1855	*/
1856	int VBOXCALL supdrvIOCtl(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
1857	{
1858	int rc;
1859	VBOXDRV_IOCTL_ENTRY(pSession, uIOCtl, pReqHdr);
1860
1861	/*
1862	* Validate the request.
1863	*/
1864	/* this first check could probably be omitted as its also done by the OS specific code... */
1865	if (RT_UNLIKELY( (pReqHdr->fFlags & SUPREQHDR_FLAGS_MAGIC_MASK) != SUPREQHDR_FLAGS_MAGIC
1866	\|\| pReqHdr->cbIn < sizeof(*pReqHdr)
1867	\|\| pReqHdr->cbOut < sizeof(*pReqHdr)))
1868	{
1869	OSDBGPRINT(("vboxdrv: Bad ioctl request header; cbIn=%#lx cbOut=%#lx fFlags=%#lx\n",
1870	(long)pReqHdr->cbIn, (long)pReqHdr->cbOut, (long)pReqHdr->fFlags));
1871	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1872	return VERR_INVALID_PARAMETER;
1873	}
1874	if (RT_UNLIKELY(!RT_VALID_PTR(pSession)))
1875	{
1876	OSDBGPRINT(("vboxdrv: Invalid pSession valud %p (ioctl=%p)\n", pSession, (void *)uIOCtl));
1877	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1878	return VERR_INVALID_PARAMETER;
1879	}
1880	if (RT_UNLIKELY(uIOCtl == SUP_IOCTL_COOKIE))
1881	{
1882	if (pReqHdr->u32Cookie != SUPCOOKIE_INITIAL_COOKIE)
1883	{
1884	OSDBGPRINT(("SUP_IOCTL_COOKIE: bad cookie %#lx\n", (long)pReqHdr->u32Cookie));
1885	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1886	return VERR_INVALID_PARAMETER;
1887	}
1888	}
1889	else if (RT_UNLIKELY( pReqHdr->u32Cookie != pDevExt->u32Cookie
1890	\|\| pReqHdr->u32SessionCookie != pSession->u32Cookie))
1891	{
1892	OSDBGPRINT(("vboxdrv: bad cookie %#lx / %#lx.\n", (long)pReqHdr->u32Cookie, (long)pReqHdr->u32SessionCookie));
1893	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, VERR_INVALID_PARAMETER, VINF_SUCCESS);
1894	return VERR_INVALID_PARAMETER;
1895	}
1896
1897	/*
1898	* Hand it to an inner function to avoid lots of unnecessary return tracepoints
1899	*/
1900	rc = supdrvIOCtlInner(uIOCtl, pDevExt, pSession, pReqHdr);
1901
1902	VBOXDRV_IOCTL_RETURN(pSession, uIOCtl, pReqHdr, pReqHdr->rc, rc);
1903	return rc;
1904	}
1905
1906
1907	/**
1908	* Inter-Driver Communication (IDC) worker.
1909	*
1910	* @returns VBox status code.
1911	* @retval VINF_SUCCESS on success.
1912	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1913	* @retval VERR_NOT_SUPPORTED if the request isn't supported.
1914	*
1915	* @param uReq The request (function) code.
1916	* @param pDevExt Device extention.
1917	* @param pSession Session data.
1918	* @param pReqHdr The request header.
1919	*/
1920	int VBOXCALL supdrvIDC(uintptr_t uReq, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQHDR pReqHdr)
1921	{
1922	/*
1923	* The OS specific code has already validated the pSession
1924	* pointer, and the request size being greater or equal to
1925	* size of the header.
1926	*
1927	* So, just check that pSession is a kernel context session.
1928	*/
1929	if (RT_UNLIKELY( pSession
1930	&& pSession->R0Process != NIL_RTR0PROCESS))
1931	return VERR_INVALID_PARAMETER;
1932
1933	/*
1934	* Validation macro.
1935	*/
1936	#define REQ_CHECK_IDC_SIZE(Name, cbExpect) \
1937	do { \
1938	if (RT_UNLIKELY(pReqHdr->cb != (cbExpect))) \
1939	{ \
1940	OSDBGPRINT(( #Name ": Invalid input/output sizes. cb=%ld expected %ld.\n", \
1941	(long)pReqHdr->cb, (long)(cbExpect))); \
1942	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1943	} \
1944	} while (0)
1945
1946	switch (uReq)
1947	{
1948	case SUPDRV_IDC_REQ_CONNECT:
1949	{
1950	PSUPDRVIDCREQCONNECT pReq = (PSUPDRVIDCREQCONNECT)pReqHdr;
1951	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_CONNECT, sizeof(*pReq));
1952
1953	/*
1954	* Validate the cookie and other input.
1955	*/
1956	if (pReq->Hdr.pSession != NULL)
1957	{
1958	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Hdr.pSession=%p expected NULL!\n", pReq->Hdr.pSession));
1959	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1960	}
1961	if (pReq->u.In.u32MagicCookie != SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE)
1962	{
1963	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: u32MagicCookie=%#x expected %#x!\n",
1964	(unsigned)pReq->u.In.u32MagicCookie, (unsigned)SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE));
1965	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1966	}
1967	if ( pReq->u.In.uMinVersion > pReq->u.In.uReqVersion
1968	\|\| (pReq->u.In.uMinVersion & UINT32_C(0xffff0000)) != (pReq->u.In.uReqVersion & UINT32_C(0xffff0000)))
1969	{
1970	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: uMinVersion=%#x uMaxVersion=%#x doesn't match!\n",
1971	pReq->u.In.uMinVersion, pReq->u.In.uReqVersion));
1972	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1973	}
1974	if (pSession != NULL)
1975	{
1976	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: pSession=%p expected NULL!\n", pSession));
1977	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1978	}
1979
1980	/*
1981	* Match the version.
1982	* The current logic is very simple, match the major interface version.
1983	*/
1984	if ( pReq->u.In.uMinVersion > SUPDRV_IDC_VERSION
1985	\|\| (pReq->u.In.uMinVersion & 0xffff0000) != (SUPDRV_IDC_VERSION & 0xffff0000))
1986	{
1987	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1988	pReq->u.In.uReqVersion, pReq->u.In.uMinVersion, (unsigned)SUPDRV_IDC_VERSION));
1989	pReq->u.Out.pSession = NULL;
1990	pReq->u.Out.uSessionVersion = 0xffffffff;
1991	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1992	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1993	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1994	return VINF_SUCCESS;
1995	}
1996
1997	pReq->u.Out.pSession = NULL;
1998	pReq->u.Out.uSessionVersion = SUPDRV_IDC_VERSION;
1999	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
2000	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
2001
2002	pReq->Hdr.rc = supdrvCreateSession(pDevExt, false /* fUser */, &pSession);
2003	if (RT_FAILURE(pReq->Hdr.rc))
2004	{
2005	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: failed to create session, rc=%d\n", pReq->Hdr.rc));
2006	return VINF_SUCCESS;
2007	}
2008
2009	pReq->u.Out.pSession = pSession;
2010	pReq->Hdr.pSession = pSession;
2011
2012	return VINF_SUCCESS;
2013	}
2014
2015	case SUPDRV_IDC_REQ_DISCONNECT:
2016	{
2017	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_DISCONNECT, sizeof(*pReqHdr));
2018
2019	supdrvCloseSession(pDevExt, pSession);
2020	return pReqHdr->rc = VINF_SUCCESS;
2021	}
2022
2023	case SUPDRV_IDC_REQ_GET_SYMBOL:
2024	{
2025	PSUPDRVIDCREQGETSYM pReq = (PSUPDRVIDCREQGETSYM)pReqHdr;
2026	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_GET_SYMBOL, sizeof(*pReq));
2027
2028	pReq->Hdr.rc = supdrvIDC_LdrGetSymbol(pDevExt, pSession, pReq);
2029	return VINF_SUCCESS;
2030	}
2031
2032	case SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY:
2033	{
2034	PSUPDRVIDCREQCOMPREGFACTORY pReq = (PSUPDRVIDCREQCOMPREGFACTORY)pReqHdr;
2035	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY, sizeof(*pReq));
2036
2037	pReq->Hdr.rc = SUPR0ComponentRegisterFactory(pSession, pReq->u.In.pFactory);
2038	return VINF_SUCCESS;
2039	}
2040
2041	case SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY:
2042	{
2043	PSUPDRVIDCREQCOMPDEREGFACTORY pReq = (PSUPDRVIDCREQCOMPDEREGFACTORY)pReqHdr;
2044	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY, sizeof(*pReq));
2045
2046	pReq->Hdr.rc = SUPR0ComponentDeregisterFactory(pSession, pReq->u.In.pFactory);
2047	return VINF_SUCCESS;
2048	}
2049
2050	default:
2051	Log(("Unknown IDC %#lx\n", (long)uReq));
2052	break;
2053	}
2054
2055	#undef REQ_CHECK_IDC_SIZE
2056	return VERR_NOT_SUPPORTED;
2057	}
2058
2059
2060	/**
2061	* Register a object for reference counting.
2062	* The object is registered with one reference in the specified session.
2063	*
2064	* @returns Unique identifier on success (pointer).
2065	* All future reference must use this identifier.
2066	* @returns NULL on failure.
2067	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
2068	* @param pvUser1 The first user argument.
2069	* @param pvUser2 The second user argument.
2070	*/
2071	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
2072	{
2073	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2074	PSUPDRVOBJ pObj;
2075	PSUPDRVUSAGE pUsage;
2076
2077	/*
2078	* Validate the input.
2079	*/
2080	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
2081	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
2082	AssertPtrReturn(pfnDestructor, NULL);
2083
2084	/*
2085	* Allocate and initialize the object.
2086	*/
2087	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
2088	if (!pObj)
2089	return NULL;
2090	pObj->u32Magic = SUPDRVOBJ_MAGIC;
2091	pObj->enmType = enmType;
2092	pObj->pNext = NULL;
2093	pObj->cUsage = 1;
2094	pObj->pfnDestructor = pfnDestructor;
2095	pObj->pvUser1 = pvUser1;
2096	pObj->pvUser2 = pvUser2;
2097	pObj->CreatorUid = pSession->Uid;
2098	pObj->CreatorGid = pSession->Gid;
2099	pObj->CreatorProcess= pSession->Process;
2100	supdrvOSObjInitCreator(pObj, pSession);
2101
2102	/*
2103	* Allocate the usage record.
2104	* (We keep freed usage records around to simplify SUPR0ObjAddRefEx().)
2105	*/
2106	RTSpinlockAcquire(pDevExt->Spinlock);
2107
2108	pUsage = pDevExt->pUsageFree;
2109	if (pUsage)
2110	pDevExt->pUsageFree = pUsage->pNext;
2111	else
2112	{
2113	RTSpinlockRelease(pDevExt->Spinlock);
2114	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
2115	if (!pUsage)
2116	{
2117	RTMemFree(pObj);
2118	return NULL;
2119	}
2120	RTSpinlockAcquire(pDevExt->Spinlock);
2121	}
2122
2123	/*
2124	* Insert the object and create the session usage record.
2125	*/
2126	/* The object. */
2127	pObj->pNext = pDevExt->pObjs;
2128	pDevExt->pObjs = pObj;
2129
2130	/* The session record. */
2131	pUsage->cUsage = 1;
2132	pUsage->pObj = pObj;
2133	pUsage->pNext = pSession->pUsage;
2134	/* Log2(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext)); */
2135	pSession->pUsage = pUsage;
2136
2137	RTSpinlockRelease(pDevExt->Spinlock);
2138
2139	Log(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
2140	return pObj;
2141	}
2142
2143
2144	/**
2145	* Increment the reference counter for the object associating the reference
2146	* with the specified session.
2147	*
2148	* @returns IPRT status code.
2149	* @param pvObj The identifier returned by SUPR0ObjRegister().
2150	* @param pSession The session which is referencing the object.
2151	*
2152	* @remarks The caller should not own any spinlocks and must carefully protect
2153	* itself against potential race with the destructor so freed memory
2154	* isn't accessed here.
2155	*/
2156	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
2157	{
2158	return SUPR0ObjAddRefEx(pvObj, pSession, false /* fNoBlocking */);
2159	}
2160
2161
2162	/**
2163	* Increment the reference counter for the object associating the reference
2164	* with the specified session.
2165	*
2166	* @returns IPRT status code.
2167	* @retval VERR_TRY_AGAIN if fNoBlocking was set and a new usage record
2168	* couldn't be allocated. (If you see this you're not doing the right
2169	* thing and it won't ever work reliably.)
2170	*
2171	* @param pvObj The identifier returned by SUPR0ObjRegister().
2172	* @param pSession The session which is referencing the object.
2173	* @param fNoBlocking Set if it's not OK to block. Never try to make the
2174	* first reference to an object in a session with this
2175	* argument set.
2176	*
2177	* @remarks The caller should not own any spinlocks and must carefully protect
2178	* itself against potential race with the destructor so freed memory
2179	* isn't accessed here.
2180	*/
2181	SUPR0DECL(int) SUPR0ObjAddRefEx(void *pvObj, PSUPDRVSESSION pSession, bool fNoBlocking)
2182	{
2183	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2184	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2185	int rc = VINF_SUCCESS;
2186	PSUPDRVUSAGE pUsagePre;
2187	PSUPDRVUSAGE pUsage;
2188
2189	/*
2190	* Validate the input.
2191	* Be ready for the destruction race (someone might be stuck in the
2192	* destructor waiting a lock we own).
2193	*/
2194	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2195	AssertPtrReturn(pObj, VERR_INVALID_POINTER);
2196	AssertMsgReturn(pObj->u32Magic == SUPDRVOBJ_MAGIC \|\| pObj->u32Magic == SUPDRVOBJ_MAGIC_DEAD,
2197	("Invalid pvObj=%p magic=%#x (expected %#x or %#x)\n", pvObj, pObj->u32Magic, SUPDRVOBJ_MAGIC, SUPDRVOBJ_MAGIC_DEAD),
2198	VERR_INVALID_PARAMETER);
2199
2200	RTSpinlockAcquire(pDevExt->Spinlock);
2201
2202	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2203	{
2204	RTSpinlockRelease(pDevExt->Spinlock);
2205
2206	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2207	return VERR_WRONG_ORDER;
2208	}
2209
2210	/*
2211	* Preallocate the usage record if we can.
2212	*/
2213	pUsagePre = pDevExt->pUsageFree;
2214	if (pUsagePre)
2215	pDevExt->pUsageFree = pUsagePre->pNext;
2216	else if (!fNoBlocking)
2217	{
2218	RTSpinlockRelease(pDevExt->Spinlock);
2219	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
2220	if (!pUsagePre)
2221	return VERR_NO_MEMORY;
2222
2223	RTSpinlockAcquire(pDevExt->Spinlock);
2224	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
2225	{
2226	RTSpinlockRelease(pDevExt->Spinlock);
2227
2228	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
2229	return VERR_WRONG_ORDER;
2230	}
2231	}
2232
2233	/*
2234	* Reference the object.
2235	*/
2236	pObj->cUsage++;
2237
2238	/*
2239	* Look for the session record.
2240	*/
2241	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
2242	{
2243	/Log(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2244	if (pUsage->pObj == pObj)
2245	break;
2246	}
2247	if (pUsage)
2248	pUsage->cUsage++;
2249	else if (pUsagePre)
2250	{
2251	/* create a new session record. */
2252	pUsagePre->cUsage = 1;
2253	pUsagePre->pObj = pObj;
2254	pUsagePre->pNext = pSession->pUsage;
2255	pSession->pUsage = pUsagePre;
2256	/Log(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));/
2257
2258	pUsagePre = NULL;
2259	}
2260	else
2261	{
2262	pObj->cUsage--;
2263	rc = VERR_TRY_AGAIN;
2264	}
2265
2266	/*
2267	* Put any unused usage record into the free list..
2268	*/
2269	if (pUsagePre)
2270	{
2271	pUsagePre->pNext = pDevExt->pUsageFree;
2272	pDevExt->pUsageFree = pUsagePre;
2273	}
2274
2275	RTSpinlockRelease(pDevExt->Spinlock);
2276
2277	return rc;
2278	}
2279
2280
2281	/**
2282	* Decrement / destroy a reference counter record for an object.
2283	*
2284	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
2285	*
2286	* @returns IPRT status code.
2287	* @retval VINF_SUCCESS if not destroyed.
2288	* @retval VINF_OBJECT_DESTROYED if it's destroyed by this release call.
2289	* @retval VERR_INVALID_PARAMETER if the object isn't valid. Will assert in
2290	* string builds.
2291	*
2292	* @param pvObj The identifier returned by SUPR0ObjRegister().
2293	* @param pSession The session which is referencing the object.
2294	*/
2295	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
2296	{
2297	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2298	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2299	int rc = VERR_INVALID_PARAMETER;
2300	PSUPDRVUSAGE pUsage;
2301	PSUPDRVUSAGE pUsagePrev;
2302
2303	/*
2304	* Validate the input.
2305	*/
2306	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2307	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2308	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2309	VERR_INVALID_PARAMETER);
2310
2311	/*
2312	* Acquire the spinlock and look for the usage record.
2313	*/
2314	RTSpinlockAcquire(pDevExt->Spinlock);
2315
2316	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
2317	pUsage;
2318	pUsagePrev = pUsage, pUsage = pUsage->pNext)
2319	{
2320	/Log2(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
2321	if (pUsage->pObj == pObj)
2322	{
2323	rc = VINF_SUCCESS;
2324	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
2325	if (pUsage->cUsage > 1)
2326	{
2327	pObj->cUsage--;
2328	pUsage->cUsage--;
2329	}
2330	else
2331	{
2332	/*
2333	* Free the session record.
2334	*/
2335	if (pUsagePrev)
2336	pUsagePrev->pNext = pUsage->pNext;
2337	else
2338	pSession->pUsage = pUsage->pNext;
2339	pUsage->pNext = pDevExt->pUsageFree;
2340	pDevExt->pUsageFree = pUsage;
2341
2342	/* What about the object? */
2343	if (pObj->cUsage > 1)
2344	pObj->cUsage--;
2345	else
2346	{
2347	/*
2348	* Object is to be destroyed, unlink it.
2349	*/
2350	pObj->u32Magic = SUPDRVOBJ_MAGIC_DEAD;
2351	rc = VINF_OBJECT_DESTROYED;
2352	if (pDevExt->pObjs == pObj)
2353	pDevExt->pObjs = pObj->pNext;
2354	else
2355	{
2356	PSUPDRVOBJ pObjPrev;
2357	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
2358	if (pObjPrev->pNext == pObj)
2359	{
2360	pObjPrev->pNext = pObj->pNext;
2361	break;
2362	}
2363	Assert(pObjPrev);
2364	}
2365	}
2366	}
2367	break;
2368	}
2369	}
2370
2371	RTSpinlockRelease(pDevExt->Spinlock);
2372
2373	/*
2374	* Call the destructor and free the object if required.
2375	*/
2376	if (rc == VINF_OBJECT_DESTROYED)
2377	{
2378	Log(("SUPR0ObjRelease: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
2379	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
2380	if (pObj->pfnDestructor)
2381	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
2382	RTMemFree(pObj);
2383	}
2384
2385	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
2386	return rc;
2387	}
2388
2389
2390	/**
2391	* Verifies that the current process can access the specified object.
2392	*
2393	* @returns The following IPRT status code:
2394	* @retval VINF_SUCCESS if access was granted.
2395	* @retval VERR_PERMISSION_DENIED if denied access.
2396	* @retval VERR_INVALID_PARAMETER if invalid parameter.
2397	*
2398	* @param pvObj The identifier returned by SUPR0ObjRegister().
2399	* @param pSession The session which wishes to access the object.
2400	* @param pszObjName Object string name. This is optional and depends on the object type.
2401	*
2402	* @remark The caller is responsible for making sure the object isn't removed while
2403	* we're inside this function. If uncertain about this, just call AddRef before calling us.
2404	*/
2405	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
2406	{
2407	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
2408	int rc;
2409
2410	/*
2411	* Validate the input.
2412	*/
2413	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2414	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
2415	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
2416	VERR_INVALID_PARAMETER);
2417
2418	/*
2419	* Check access. (returns true if a decision has been made.)
2420	*/
2421	rc = VERR_INTERNAL_ERROR;
2422	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
2423	return rc;
2424
2425	/*
2426	* Default policy is to allow the user to access his own
2427	* stuff but nothing else.
2428	*/
2429	if (pObj->CreatorUid == pSession->Uid)
2430	return VINF_SUCCESS;
2431	return VERR_PERMISSION_DENIED;
2432	}
2433
2434
2435	/**
2436	* Lock pages.
2437	*
2438	* @returns IPRT status code.
2439	* @param pSession Session to which the locked memory should be associated.
2440	* @param pvR3 Start of the memory range to lock.
2441	* This must be page aligned.
2442	* @param cPages Number of pages to lock.
2443	* @param paPages Where to put the physical addresses of locked memory.
2444	*/
2445	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
2446	{
2447	int rc;
2448	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2449	const size_t cb = (size_t)cPages << PAGE_SHIFT;
2450	LogFlow(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
2451
2452	/*
2453	* Verify input.
2454	*/
2455	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2456	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
2457	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
2458	\|\| !pvR3)
2459	{
2460	Log(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
2461	return VERR_INVALID_PARAMETER;
2462	}
2463
2464	/*
2465	* Let IPRT do the job.
2466	*/
2467	Mem.eType = MEMREF_TYPE_LOCKED;
2468	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTMEM_PROT_READ \| RTMEM_PROT_WRITE, RTR0ProcHandleSelf());
2469	if (RT_SUCCESS(rc))
2470	{
2471	uint32_t iPage = cPages;
2472	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
2473	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
2474
2475	while (iPage-- > 0)
2476	{
2477	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2478	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
2479	{
2480	AssertMsgFailed(("iPage=%d\n", iPage));
2481	rc = VERR_INTERNAL_ERROR;
2482	break;
2483	}
2484	}
2485	if (RT_SUCCESS(rc))
2486	rc = supdrvMemAdd(&Mem, pSession);
2487	if (RT_FAILURE(rc))
2488	{
2489	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
2490	AssertRC(rc2);
2491	}
2492	}
2493
2494	return rc;
2495	}
2496
2497
2498	/**
2499	* Unlocks the memory pointed to by pv.
2500	*
2501	* @returns IPRT status code.
2502	* @param pSession Session to which the memory was locked.
2503	* @param pvR3 Memory to unlock.
2504	*/
2505	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2506	{
2507	LogFlow(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2508	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2509	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
2510	}
2511
2512
2513	/**
2514	* Allocates a chunk of page aligned memory with contiguous and fixed physical
2515	* backing.
2516	*
2517	* @returns IPRT status code.
2518	* @param pSession Session data.
2519	* @param cPages Number of pages to allocate.
2520	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
2521	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
2522	* @param pHCPhys Where to put the physical address of allocated memory.
2523	*/
2524	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
2525	{
2526	int rc;
2527	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2528	LogFlow(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
2529
2530	/*
2531	* Validate input.
2532	*/
2533	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2534	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
2535	{
2536	Log(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
2537	pSession, ppvR0, ppvR3, pHCPhys));
2538	return VERR_INVALID_PARAMETER;
2539
2540	}
2541	if (cPages < 1 \|\| cPages >= 256)
2542	{
2543	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2544	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2545	}
2546
2547	/*
2548	* Let IPRT do the job.
2549	*/
2550	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
2551	if (RT_SUCCESS(rc))
2552	{
2553	int rc2;
2554	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2555	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2556	if (RT_SUCCESS(rc))
2557	{
2558	Mem.eType = MEMREF_TYPE_CONT;
2559	rc = supdrvMemAdd(&Mem, pSession);
2560	if (!rc)
2561	{
2562	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2563	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2564	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
2565	return 0;
2566	}
2567
2568	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2569	AssertRC(rc2);
2570	}
2571	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2572	AssertRC(rc2);
2573	}
2574
2575	return rc;
2576	}
2577
2578
2579	/**
2580	* Frees memory allocated using SUPR0ContAlloc().
2581	*
2582	* @returns IPRT status code.
2583	* @param pSession The session to which the memory was allocated.
2584	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2585	*/
2586	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2587	{
2588	LogFlow(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2589	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2590	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
2591	}
2592
2593
2594	/**
2595	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
2596	*
2597	* The memory isn't zeroed.
2598	*
2599	* @returns IPRT status code.
2600	* @param pSession Session data.
2601	* @param cPages Number of pages to allocate.
2602	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
2603	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
2604	* @param paPages Where to put the physical addresses of allocated memory.
2605	*/
2606	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
2607	{
2608	unsigned iPage;
2609	int rc;
2610	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2611	LogFlow(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
2612
2613	/*
2614	* Validate input.
2615	*/
2616	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2617	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
2618	{
2619	Log(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
2620	pSession, ppvR3, ppvR0, paPages));
2621	return VERR_INVALID_PARAMETER;
2622
2623	}
2624	if (cPages < 1 \|\| cPages >= 256)
2625	{
2626	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
2627	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2628	}
2629
2630	/*
2631	* Let IPRT do the work.
2632	*/
2633	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
2634	if (RT_SUCCESS(rc))
2635	{
2636	int rc2;
2637	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2638	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2639	if (RT_SUCCESS(rc))
2640	{
2641	Mem.eType = MEMREF_TYPE_LOW;
2642	rc = supdrvMemAdd(&Mem, pSession);
2643	if (!rc)
2644	{
2645	for (iPage = 0; iPage < cPages; iPage++)
2646	{
2647	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
2648	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%RHp\n", paPages[iPage]));
2649	}
2650	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2651	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2652	return 0;
2653	}
2654
2655	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2656	AssertRC(rc2);
2657	}
2658
2659	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2660	AssertRC(rc2);
2661	}
2662
2663	return rc;
2664	}
2665
2666
2667	/**
2668	* Frees memory allocated using SUPR0LowAlloc().
2669	*
2670	* @returns IPRT status code.
2671	* @param pSession The session to which the memory was allocated.
2672	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2673	*/
2674	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2675	{
2676	LogFlow(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2677	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2678	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
2679	}
2680
2681
2682
2683	/**
2684	* Allocates a chunk of memory with both R0 and R3 mappings.
2685	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
2686	*
2687	* @returns IPRT status code.
2688	* @param pSession The session to associated the allocation with.
2689	* @param cb Number of bytes to allocate.
2690	* @param ppvR0 Where to store the address of the Ring-0 mapping.
2691	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2692	*/
2693	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
2694	{
2695	int rc;
2696	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2697	LogFlow(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
2698
2699	/*
2700	* Validate input.
2701	*/
2702	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2703	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
2704	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
2705	if (cb < 1 \|\| cb >= _4M)
2706	{
2707	Log(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
2708	return VERR_INVALID_PARAMETER;
2709	}
2710
2711	/*
2712	* Let IPRT do the work.
2713	*/
2714	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
2715	if (RT_SUCCESS(rc))
2716	{
2717	int rc2;
2718	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2719	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2720	if (RT_SUCCESS(rc))
2721	{
2722	Mem.eType = MEMREF_TYPE_MEM;
2723	rc = supdrvMemAdd(&Mem, pSession);
2724	if (!rc)
2725	{
2726	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2727	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2728	return VINF_SUCCESS;
2729	}
2730
2731	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2732	AssertRC(rc2);
2733	}
2734
2735	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2736	AssertRC(rc2);
2737	}
2738
2739	return rc;
2740	}
2741
2742
2743	/**
2744	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
2745	*
2746	* @returns IPRT status code.
2747	* @param pSession The session to which the memory was allocated.
2748	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2749	* @param paPages Where to store the physical addresses.
2750	*/
2751	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
2752	{
2753	PSUPDRVBUNDLE pBundle;
2754	LogFlow(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
2755
2756	/*
2757	* Validate input.
2758	*/
2759	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2760	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
2761	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
2762
2763	/*
2764	* Search for the address.
2765	*/
2766	RTSpinlockAcquire(pSession->Spinlock);
2767	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2768	{
2769	if (pBundle->cUsed > 0)
2770	{
2771	unsigned i;
2772	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2773	{
2774	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
2775	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2776	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2777	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2778	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
2779	)
2780	)
2781	{
2782	const size_t cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2783	size_t iPage;
2784	for (iPage = 0; iPage < cPages; iPage++)
2785	{
2786	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2787	paPages[iPage].uReserved = 0;
2788	}
2789	RTSpinlockRelease(pSession->Spinlock);
2790	return VINF_SUCCESS;
2791	}
2792	}
2793	}
2794	}
2795	RTSpinlockRelease(pSession->Spinlock);
2796	Log(("Failed to find %p!!!\n", (void *)uPtr));
2797	return VERR_INVALID_PARAMETER;
2798	}
2799
2800
2801	/**
2802	* Free memory allocated by SUPR0MemAlloc().
2803	*
2804	* @returns IPRT status code.
2805	* @param pSession The session owning the allocation.
2806	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2807	*/
2808	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2809	{
2810	LogFlow(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2811	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2812	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
2813	}
2814
2815
2816	/**
2817	* Allocates a chunk of memory with a kernel or/and a user mode mapping.
2818	*
2819	* The memory is fixed and it's possible to query the physical addresses using
2820	* SUPR0MemGetPhys().
2821	*
2822	* @returns IPRT status code.
2823	* @param pSession The session to associated the allocation with.
2824	* @param cPages The number of pages to allocate.
2825	* @param fFlags Flags, reserved for the future. Must be zero.
2826	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2827	* NULL if no ring-3 mapping.
2828	* @param ppvR3 Where to store the address of the Ring-0 mapping.
2829	* NULL if no ring-0 mapping.
2830	* @param paPages Where to store the addresses of the pages. Optional.
2831	*/
2832	SUPR0DECL(int) SUPR0PageAllocEx(PSUPDRVSESSION pSession, uint32_t cPages, uint32_t fFlags, PRTR3PTR ppvR3, PRTR0PTR ppvR0, PRTHCPHYS paPages)
2833	{
2834	int rc;
2835	SUPDRVMEMREF Mem = { NIL_RTR0MEMOBJ, NIL_RTR0MEMOBJ, MEMREF_TYPE_UNUSED };
2836	LogFlow(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
2837
2838	/*
2839	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2840	*/
2841	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2842	AssertPtrNullReturn(ppvR3, VERR_INVALID_POINTER);
2843	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2844	AssertReturn(ppvR3 \|\| ppvR0, VERR_INVALID_PARAMETER);
2845	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2846	if (cPages < 1 \|\| cPages > VBOX_MAX_ALLOC_PAGE_COUNT)
2847	{
2848	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)));
2849	return VERR_PAGE_COUNT_OUT_OF_RANGE;
2850	}
2851
2852	/*
2853	* Let IPRT do the work.
2854	*/
2855	if (ppvR0)
2856	rc = RTR0MemObjAllocPage(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, true /* fExecutable */);
2857	else
2858	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
2859	if (RT_SUCCESS(rc))
2860	{
2861	int rc2;
2862	if (ppvR3)
2863	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2864	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2865	else
2866	Mem.MapObjR3 = NIL_RTR0MEMOBJ;
2867	if (RT_SUCCESS(rc))
2868	{
2869	Mem.eType = MEMREF_TYPE_PAGE;
2870	rc = supdrvMemAdd(&Mem, pSession);
2871	if (!rc)
2872	{
2873	if (ppvR3)
2874	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2875	if (ppvR0)
2876	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2877	if (paPages)
2878	{
2879	uint32_t iPage = cPages;
2880	while (iPage-- > 0)
2881	{
2882	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
2883	Assert(paPages[iPage] != NIL_RTHCPHYS);
2884	}
2885	}
2886	return VINF_SUCCESS;
2887	}
2888
2889	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2890	AssertRC(rc2);
2891	}
2892
2893	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2894	AssertRC(rc2);
2895	}
2896	return rc;
2897	}
2898
2899
2900	/**
2901	* Maps a chunk of memory previously allocated by SUPR0PageAllocEx into kernel
2902	* space.
2903	*
2904	* @returns IPRT status code.
2905	* @param pSession The session to associated the allocation with.
2906	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
2907	* @param offSub Where to start mapping. Must be page aligned.
2908	* @param cbSub How much to map. Must be page aligned.
2909	* @param fFlags Flags, MBZ.
2910	* @param ppvR0 Where to return the address of the ring-0 mapping on
2911	* success.
2912	*/
2913	SUPR0DECL(int) SUPR0PageMapKernel(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t offSub, uint32_t cbSub,
2914	uint32_t fFlags, PRTR0PTR ppvR0)
2915	{
2916	int rc;
2917	PSUPDRVBUNDLE pBundle;
2918	RTR0MEMOBJ hMemObj = NIL_RTR0MEMOBJ;
2919	LogFlow(("SUPR0PageMapKernel: pSession=%p pvR3=%p offSub=%#x cbSub=%#x\n", pSession, pvR3, offSub, cbSub));
2920
2921	/*
2922	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2923	*/
2924	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2925	AssertPtrNullReturn(ppvR0, VERR_INVALID_POINTER);
2926	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2927	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2928	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2929	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
2930
2931	/*
2932	* Find the memory object.
2933	*/
2934	RTSpinlockAcquire(pSession->Spinlock);
2935	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2936	{
2937	if (pBundle->cUsed > 0)
2938	{
2939	unsigned i;
2940	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2941	{
2942	if ( ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
2943	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2944	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2945	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
2946	\|\| ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED
2947	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2948	&& pBundle->aMem[i].MapObjR3 == NIL_RTR0MEMOBJ
2949	&& RTR0MemObjAddressR3(pBundle->aMem[i].MemObj) == pvR3))
2950	{
2951	hMemObj = pBundle->aMem[i].MemObj;
2952	break;
2953	}
2954	}
2955	}
2956	}
2957	RTSpinlockRelease(pSession->Spinlock);
2958
2959	rc = VERR_INVALID_PARAMETER;
2960	if (hMemObj != NIL_RTR0MEMOBJ)
2961	{
2962	/*
2963	* Do some further input validations before calling IPRT.
2964	* (Cleanup is done indirectly by telling RTR0MemObjFree to include mappings.)
2965	*/
2966	size_t cbMemObj = RTR0MemObjSize(hMemObj);
2967	if ( offSub < cbMemObj
2968	&& cbSub <= cbMemObj
2969	&& offSub + cbSub <= cbMemObj)
2970	{
2971	RTR0MEMOBJ hMapObj;
2972	rc = RTR0MemObjMapKernelEx(&hMapObj, hMemObj, (void *)-1, 0,
2973	RTMEM_PROT_READ \| RTMEM_PROT_WRITE, offSub, cbSub);
2974	if (RT_SUCCESS(rc))
2975	*ppvR0 = RTR0MemObjAddress(hMapObj);
2976	}
2977	else
2978	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
2979
2980	}
2981	return rc;
2982	}
2983
2984
2985	/**
2986	* Changes the page level protection of one or more pages previously allocated
2987	* by SUPR0PageAllocEx.
2988	*
2989	* @returns IPRT status code.
2990	* @param pSession The session to associated the allocation with.
2991	* @param pvR3 The ring-3 address returned by SUPR0PageAllocEx.
2992	* NIL_RTR3PTR if the ring-3 mapping should be unaffected.
2993	* @param pvR0 The ring-0 address returned by SUPR0PageAllocEx.
2994	* NIL_RTR0PTR if the ring-0 mapping should be unaffected.
2995	* @param offSub Where to start changing. Must be page aligned.
2996	* @param cbSub How much to change. Must be page aligned.
2997	* @param fProt The new page level protection, see RTMEM_PROT_*.
2998	*/
2999	SUPR0DECL(int) SUPR0PageProtect(PSUPDRVSESSION pSession, RTR3PTR pvR3, RTR0PTR pvR0, uint32_t offSub, uint32_t cbSub, uint32_t fProt)
3000	{
3001	int rc;
3002	PSUPDRVBUNDLE pBundle;
3003	RTR0MEMOBJ hMemObjR0 = NIL_RTR0MEMOBJ;
3004	RTR0MEMOBJ hMemObjR3 = NIL_RTR0MEMOBJ;
3005	LogFlow(("SUPR0PageProtect: pSession=%p pvR3=%p pvR0=%p offSub=%#x cbSub=%#x fProt-%#x\n", pSession, pvR3, pvR0, offSub, cbSub, fProt));
3006
3007	/*
3008	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
3009	*/
3010	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3011	AssertReturn(!(fProt & ~(RTMEM_PROT_READ \| RTMEM_PROT_WRITE \| RTMEM_PROT_EXEC \| RTMEM_PROT_NONE)), VERR_INVALID_PARAMETER);
3012	AssertReturn(!(offSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3013	AssertReturn(!(cbSub & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3014	AssertReturn(cbSub, VERR_INVALID_PARAMETER);
3015
3016	/*
3017	* Find the memory object.
3018	*/
3019	RTSpinlockAcquire(pSession->Spinlock);
3020	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3021	{
3022	if (pBundle->cUsed > 0)
3023	{
3024	unsigned i;
3025	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3026	{
3027	if ( pBundle->aMem[i].eType == MEMREF_TYPE_PAGE
3028	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3029	&& ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3030	\|\| pvR3 == NIL_RTR3PTR)
3031	&& ( pvR0 == NIL_RTR0PTR
3032	\|\| RTR0MemObjAddress(pBundle->aMem[i].MemObj) == pvR0)
3033	&& ( pvR3 == NIL_RTR3PTR
3034	\|\| RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3))
3035	{
3036	if (pvR0 != NIL_RTR0PTR)
3037	hMemObjR0 = pBundle->aMem[i].MemObj;
3038	if (pvR3 != NIL_RTR3PTR)
3039	hMemObjR3 = pBundle->aMem[i].MapObjR3;
3040	break;
3041	}
3042	}
3043	}
3044	}
3045	RTSpinlockRelease(pSession->Spinlock);
3046
3047	rc = VERR_INVALID_PARAMETER;
3048	if ( hMemObjR0 != NIL_RTR0MEMOBJ
3049	\|\| hMemObjR3 != NIL_RTR0MEMOBJ)
3050	{
3051	/*
3052	* Do some further input validations before calling IPRT.
3053	*/
3054	size_t cbMemObj = hMemObjR0 != NIL_RTR0PTR ? RTR0MemObjSize(hMemObjR0) : RTR0MemObjSize(hMemObjR3);
3055	if ( offSub < cbMemObj
3056	&& cbSub <= cbMemObj
3057	&& offSub + cbSub <= cbMemObj)
3058	{
3059	rc = VINF_SUCCESS;
3060	if (hMemObjR3 != NIL_RTR0PTR)
3061	rc = RTR0MemObjProtect(hMemObjR3, offSub, cbSub, fProt);
3062	if (hMemObjR0 != NIL_RTR0PTR && RT_SUCCESS(rc))
3063	rc = RTR0MemObjProtect(hMemObjR0, offSub, cbSub, fProt);
3064	}
3065	else
3066	SUPR0Printf("SUPR0PageMapKernel: cbMemObj=%#x offSub=%#x cbSub=%#x\n", cbMemObj, offSub, cbSub);
3067
3068	}
3069	return rc;
3070
3071	}
3072
3073
3074	/**
3075	* Free memory allocated by SUPR0PageAlloc() and SUPR0PageAllocEx().
3076	*
3077	* @returns IPRT status code.
3078	* @param pSession The session owning the allocation.
3079	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc() or
3080	* SUPR0PageAllocEx().
3081	*/
3082	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
3083	{
3084	LogFlow(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
3085	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3086	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_PAGE);
3087	}
3088
3089
3090	/**
3091	* Gets the paging mode of the current CPU.
3092	*
3093	* @returns Paging mode, SUPPAGEINGMODE_INVALID on error.
3094	*/
3095	SUPR0DECL(SUPPAGINGMODE) SUPR0GetPagingMode(void)
3096	{
3097	SUPPAGINGMODE enmMode;
3098
3099	RTR0UINTREG cr0 = ASMGetCR0();
3100	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
3101	enmMode = SUPPAGINGMODE_INVALID;
3102	else
3103	{
3104	RTR0UINTREG cr4 = ASMGetCR4();
3105	uint32_t fNXEPlusLMA = 0;
3106	if (cr4 & X86_CR4_PAE)
3107	{
3108	uint32_t fAmdFeatures = ASMCpuId_EDX(0x80000001);
3109	if (fAmdFeatures & (X86_CPUID_AMD_FEATURE_EDX_NX \| X86_CPUID_AMD_FEATURE_EDX_LONG_MODE))
3110	{
3111	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
3112	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
3113	fNXEPlusLMA \|= RT_BIT(0);
3114	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
3115	fNXEPlusLMA \|= RT_BIT(1);
3116	}
3117	}
3118
3119	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
3120	{
3121	case 0:
3122	enmMode = SUPPAGINGMODE_32_BIT;
3123	break;
3124
3125	case X86_CR4_PGE:
3126	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
3127	break;
3128
3129	case X86_CR4_PAE:
3130	enmMode = SUPPAGINGMODE_PAE;
3131	break;
3132
3133	case X86_CR4_PAE \| RT_BIT(0):
3134	enmMode = SUPPAGINGMODE_PAE_NX;
3135	break;
3136
3137	case X86_CR4_PAE \| X86_CR4_PGE:
3138	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3139	break;
3140
3141	case X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3142	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3143	break;
3144
3145	case RT_BIT(1) \| X86_CR4_PAE:
3146	enmMode = SUPPAGINGMODE_AMD64;
3147	break;
3148
3149	case RT_BIT(1) \| X86_CR4_PAE \| RT_BIT(0):
3150	enmMode = SUPPAGINGMODE_AMD64_NX;
3151	break;
3152
3153	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
3154	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
3155	break;
3156
3157	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3158	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
3159	break;
3160
3161	default:
3162	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
3163	enmMode = SUPPAGINGMODE_INVALID;
3164	break;
3165	}
3166	}
3167	return enmMode;
3168	}
3169
3170
3171	/**
3172	* Enables or disabled hardware virtualization extensions using native OS APIs.
3173	*
3174	* @returns VBox status code.
3175	* @retval VINF_SUCCESS on success.
3176	* @retval VERR_NOT_SUPPORTED if not supported by the native OS.
3177	*
3178	* @param fEnable Whether to enable or disable.
3179	*/
3180	SUPR0DECL(int) SUPR0EnableVTx(bool fEnable)
3181	{
3182	#ifdef RT_OS_DARWIN
3183	return supdrvOSEnableVTx(fEnable);
3184	#else
3185	return VERR_NOT_SUPPORTED;
3186	#endif
3187	}
3188
3189
3190	/** @todo document me */
3191	SUPR0DECL(int) SUPR0QueryVTCaps(PSUPDRVSESSION pSession, uint32_t *pfCaps)
3192	{
3193	/*
3194	* Input validation.
3195	*/
3196	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3197	AssertPtrReturn(pfCaps, VERR_INVALID_POINTER);
3198
3199	*pfCaps = 0;
3200
3201	if (ASMHasCpuId())
3202	{
3203	uint32_t u32FeaturesECX;
3204	uint32_t u32Dummy;
3205	uint32_t u32FeaturesEDX;
3206	uint32_t u32VendorEBX, u32VendorECX, u32VendorEDX, u32AMDFeatureEDX, u32AMDFeatureECX;
3207	uint64_t val;
3208
3209	ASMCpuId(0, &u32Dummy, &u32VendorEBX, &u32VendorECX, &u32VendorEDX);
3210	ASMCpuId(1, &u32Dummy, &u32Dummy, &u32FeaturesECX, &u32FeaturesEDX);
3211	/* Query AMD features. */
3212	ASMCpuId(0x80000001, &u32Dummy, &u32Dummy, &u32AMDFeatureECX, &u32AMDFeatureEDX);
3213
3214	if ( u32VendorEBX == X86_CPUID_VENDOR_INTEL_EBX
3215	&& u32VendorECX == X86_CPUID_VENDOR_INTEL_ECX
3216	&& u32VendorEDX == X86_CPUID_VENDOR_INTEL_EDX
3217	)
3218	{
3219	if ( (u32FeaturesECX & X86_CPUID_FEATURE_ECX_VMX)
3220	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3221	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3222	)
3223	{
3224	val = ASMRdMsr(MSR_IA32_FEATURE_CONTROL);
3225	/*
3226	* Both the LOCK and VMXON bit must be set; otherwise VMXON will generate a #GP.
3227	* Once the lock bit is set, this MSR can no longer be modified.
3228	*/
3229	if ( (val & (MSR_IA32_FEATURE_CONTROL_VMXON\|MSR_IA32_FEATURE_CONTROL_LOCK))
3230	== (MSR_IA32_FEATURE_CONTROL_VMXON\|MSR_IA32_FEATURE_CONTROL_LOCK) /* enabled and locked */
3231	\|\| !(val & MSR_IA32_FEATURE_CONTROL_LOCK) /* not enabled, but not locked either */
3232	)
3233	{
3234	VMX_CAPABILITY vtCaps;
3235
3236	*pfCaps \|= SUPVTCAPS_VT_X;
3237
3238	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS);
3239	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC_USE_SECONDARY_EXEC_CTRL)
3240	{
3241	vtCaps.u = ASMRdMsr(MSR_IA32_VMX_PROCBASED_CTLS2);
3242	if (vtCaps.n.allowed1 & VMX_VMCS_CTRL_PROC_EXEC2_EPT)
3243	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3244	}
3245	return VINF_SUCCESS;
3246	}
3247	return VERR_VMX_MSR_LOCKED_OR_DISABLED;
3248	}
3249	return VERR_VMX_NO_VMX;
3250	}
3251
3252	if ( u32VendorEBX == X86_CPUID_VENDOR_AMD_EBX
3253	&& u32VendorECX == X86_CPUID_VENDOR_AMD_ECX
3254	&& u32VendorEDX == X86_CPUID_VENDOR_AMD_EDX
3255	)
3256	{
3257	if ( (u32AMDFeatureECX & X86_CPUID_AMD_FEATURE_ECX_SVM)
3258	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_MSR)
3259	&& (u32FeaturesEDX & X86_CPUID_FEATURE_EDX_FXSR)
3260	)
3261	{
3262	/* Check if SVM is disabled */
3263	val = ASMRdMsr(MSR_K8_VM_CR);
3264	if (!(val & MSR_K8_VM_CR_SVM_DISABLE))
3265	{
3266	*pfCaps \|= SUPVTCAPS_AMD_V;
3267
3268	/* Query AMD features. */
3269	ASMCpuId(0x8000000A, &u32Dummy, &u32Dummy, &u32Dummy, &u32FeaturesEDX);
3270
3271	if (u32FeaturesEDX & AMD_CPUID_SVM_FEATURE_EDX_NESTED_PAGING)
3272	*pfCaps \|= SUPVTCAPS_NESTED_PAGING;
3273
3274	return VINF_SUCCESS;
3275	}
3276	return VERR_SVM_DISABLED;
3277	}
3278	return VERR_SVM_NO_SVM;
3279	}
3280	}
3281
3282	return VERR_UNSUPPORTED_CPU;
3283	}
3284
3285
3286	/**
3287	* (Re-)initializes the per-cpu structure prior to starting or resuming the GIP
3288	* updating.
3289	*
3290	* @param pGipCpu The per CPU structure for this CPU.
3291	* @param u64NanoTS The current time.
3292	*/
3293	static void supdrvGipReInitCpu(PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
3294	{
3295	pGipCpu->u64TSC = ASMReadTSC() - pGipCpu->u32UpdateIntervalTSC;
3296	pGipCpu->u64NanoTS = u64NanoTS;
3297	}
3298
3299
3300	/**
3301	* Set the current TSC and NanoTS value for the CPU.
3302	*
3303	* @param idCpu The CPU ID. Unused - we have to use the APIC ID.
3304	* @param pvUser1 Pointer to the ring-0 GIP mapping.
3305	* @param pvUser2 Pointer to the variable holding the current time.
3306	*/
3307	static DECLCALLBACK(void) supdrvGipReInitCpuCallback(RTCPUID idCpu, void pvUser1, void pvUser2)
3308	{
3309	PSUPGLOBALINFOPAGE pGip = (PSUPGLOBALINFOPAGE)pvUser1;
3310	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
3311
3312	if (RT_LIKELY(iCpu < pGip->cCpus && pGip->aCPUs[iCpu].idCpu == idCpu))
3313	supdrvGipReInitCpu(&pGip->aCPUs[iCpu], (uint64_t )pvUser2);
3314
3315	NOREF(pvUser2);
3316	NOREF(idCpu);
3317	}
3318
3319
3320	/**
3321	* Maps the GIP into userspace and/or get the physical address of the GIP.
3322	*
3323	* @returns IPRT status code.
3324	* @param pSession Session to which the GIP mapping should belong.
3325	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
3326	* @param pHCPhysGip Where to store the physical address. (optional)
3327	*
3328	* @remark There is no reference counting on the mapping, so one call to this function
3329	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
3330	* and remove the session as a GIP user.
3331	*/
3332	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
3333	{
3334	int rc;
3335	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3336	RTR3PTR pGipR3 = NIL_RTR3PTR;
3337	RTHCPHYS HCPhys = NIL_RTHCPHYS;
3338	LogFlow(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
3339
3340	/*
3341	* Validate
3342	*/
3343	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3344	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
3345	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
3346
3347	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3348	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3349	#else
3350	RTSemFastMutexRequest(pDevExt->mtxGip);
3351	#endif
3352	if (pDevExt->pGip)
3353	{
3354	/*
3355	* Map it?
3356	*/
3357	rc = VINF_SUCCESS;
3358	if (ppGipR3)
3359	{
3360	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
3361	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
3362	RTMEM_PROT_READ, RTR0ProcHandleSelf());
3363	if (RT_SUCCESS(rc))
3364	pGipR3 = RTR0MemObjAddressR3(pSession->GipMapObjR3);
3365	}
3366
3367	/*
3368	* Get physical address.
3369	*/
3370	if (pHCPhysGip && RT_SUCCESS(rc))
3371	HCPhys = pDevExt->HCPhysGip;
3372
3373	/*
3374	* Reference globally.
3375	*/
3376	if (!pSession->fGipReferenced && RT_SUCCESS(rc))
3377	{
3378	pSession->fGipReferenced = 1;
3379	pDevExt->cGipUsers++;
3380	if (pDevExt->cGipUsers == 1)
3381	{
3382	PSUPGLOBALINFOPAGE pGipR0 = pDevExt->pGip;
3383	uint64_t u64NanoTS;
3384	uint32_t u32SystemResolution;
3385	unsigned i;
3386
3387	LogFlow(("SUPR0GipMap: Resumes GIP updating\n"));
3388
3389	/*
3390	* Try bump up the system timer resolution.
3391	* The more interrupts the better...
3392	*/
3393	if ( RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
3394	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
3395	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1953125 /* 512 HZ */, &u32SystemResolution))
3396	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 2000000 /* 500 HZ */, &u32SystemResolution))
3397	)
3398	{
3399	Assert(RTTimerGetSystemGranularity() <= u32SystemResolution);
3400	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
3401	}
3402
3403	if (pGipR0->aCPUs[0].u32TransactionId != 2 /* not the first time */)
3404	{
3405	for (i = 0; i < RT_ELEMENTS(pGipR0->aCPUs); i++)
3406	ASMAtomicUoWriteU32(&pGipR0->aCPUs[i].u32TransactionId,
3407	(pGipR0->aCPUs[i].u32TransactionId + GIP_UPDATEHZ_RECALC_FREQ * 2)
3408	& ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
3409	ASMAtomicWriteU64(&pGipR0->u64NanoTSLastUpdateHz, 0);
3410	}
3411
3412	u64NanoTS = RTTimeSystemNanoTS() - pGipR0->u32UpdateIntervalNS;
3413	if ( pGipR0->u32Mode == SUPGIPMODE_SYNC_TSC
3414	\|\| RTMpGetOnlineCount() == 1)
3415	supdrvGipReInitCpu(&pGipR0->aCPUs[0], u64NanoTS);
3416	else
3417	RTMpOnAll(supdrvGipReInitCpuCallback, pGipR0, &u64NanoTS);
3418
3419	#ifndef DO_NOT_START_GIP
3420	rc = RTTimerStart(pDevExt->pGipTimer, 0); AssertRC(rc);
3421	#endif
3422	rc = VINF_SUCCESS;
3423	}
3424	}
3425	}
3426	else
3427	{
3428	rc = VERR_GENERAL_FAILURE;
3429	Log(("SUPR0GipMap: GIP is not available!\n"));
3430	}
3431	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3432	RTSemMutexRelease(pDevExt->mtxGip);
3433	#else
3434	RTSemFastMutexRelease(pDevExt->mtxGip);
3435	#endif
3436
3437	/*
3438	* Write returns.
3439	*/
3440	if (pHCPhysGip)
3441	*pHCPhysGip = HCPhys;
3442	if (ppGipR3)
3443	*ppGipR3 = pGipR3;
3444
3445	#ifdef DEBUG_DARWIN_GIP
3446	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3447	#else
3448	LogFlow(( "SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
3449	#endif
3450	return rc;
3451	}
3452
3453
3454	/**
3455	* Unmaps any user mapping of the GIP and terminates all GIP access
3456	* from this session.
3457	*
3458	* @returns IPRT status code.
3459	* @param pSession Session to which the GIP mapping should belong.
3460	*/
3461	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
3462	{
3463	int rc = VINF_SUCCESS;
3464	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
3465	#ifdef DEBUG_DARWIN_GIP
3466	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
3467	pSession,
3468	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
3469	pSession->GipMapObjR3));
3470	#else
3471	LogFlow(("SUPR0GipUnmap: pSession=%p\n", pSession));
3472	#endif
3473	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3474
3475	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3476	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
3477	#else
3478	RTSemFastMutexRequest(pDevExt->mtxGip);
3479	#endif
3480
3481	/*
3482	* Unmap anything?
3483	*/
3484	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
3485	{
3486	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
3487	AssertRC(rc);
3488	if (RT_SUCCESS(rc))
3489	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
3490	}
3491
3492	/*
3493	* Dereference global GIP.
3494	*/
3495	if (pSession->fGipReferenced && !rc)
3496	{
3497	pSession->fGipReferenced = 0;
3498	if ( pDevExt->cGipUsers > 0
3499	&& !--pDevExt->cGipUsers)
3500	{
3501	LogFlow(("SUPR0GipUnmap: Suspends GIP updating\n"));
3502	#ifndef DO_NOT_START_GIP
3503	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = VINF_SUCCESS;
3504	#endif
3505
3506	if (pDevExt->u32SystemTimerGranularityGrant)
3507	{
3508	int rc2 = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant);
3509	AssertRC(rc2);
3510	pDevExt->u32SystemTimerGranularityGrant = 0;
3511	}
3512	}
3513	}
3514
3515	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3516	RTSemMutexRelease(pDevExt->mtxGip);
3517	#else
3518	RTSemFastMutexRelease(pDevExt->mtxGip);
3519	#endif
3520
3521	return rc;
3522	}
3523
3524
3525	/**
3526	* Gets the GIP pointer.
3527	*
3528	* @returns Pointer to the GIP or NULL.
3529	*/
3530	SUPDECL(PSUPGLOBALINFOPAGE) SUPGetGIP(void)
3531	{
3532	return g_pSUPGlobalInfoPage;
3533	}
3534
3535
3536	/**
3537	* Register a component factory with the support driver.
3538	*
3539	* This is currently restricted to kernel sessions only.
3540	*
3541	* @returns VBox status code.
3542	* @retval VINF_SUCCESS on success.
3543	* @retval VERR_NO_MEMORY if we're out of memory.
3544	* @retval VERR_ALREADY_EXISTS if the factory has already been registered.
3545	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3546	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3547	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3548	*
3549	* @param pSession The SUPDRV session (must be a ring-0 session).
3550	* @param pFactory Pointer to the component factory registration structure.
3551	*
3552	* @remarks This interface is also available via SUPR0IdcComponentRegisterFactory.
3553	*/
3554	SUPR0DECL(int) SUPR0ComponentRegisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3555	{
3556	PSUPDRVFACTORYREG pNewReg;
3557	const char *psz;
3558	int rc;
3559
3560	/*
3561	* Validate parameters.
3562	*/
3563	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3564	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3565	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3566	AssertPtrReturn(pFactory->pfnQueryFactoryInterface, VERR_INVALID_POINTER);
3567	psz = RTStrEnd(pFactory->szName, sizeof(pFactory->szName));
3568	AssertReturn(psz, VERR_INVALID_PARAMETER);
3569
3570	/*
3571	* Allocate and initialize a new registration structure.
3572	*/
3573	pNewReg = (PSUPDRVFACTORYREG)RTMemAlloc(sizeof(SUPDRVFACTORYREG));
3574	if (pNewReg)
3575	{
3576	pNewReg->pNext = NULL;
3577	pNewReg->pFactory = pFactory;
3578	pNewReg->pSession = pSession;
3579	pNewReg->cchName = psz - &pFactory->szName[0];
3580
3581	/*
3582	* Add it to the tail of the list after checking for prior registration.
3583	*/
3584	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3585	if (RT_SUCCESS(rc))
3586	{
3587	PSUPDRVFACTORYREG pPrev = NULL;
3588	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3589	while (pCur && pCur->pFactory != pFactory)
3590	{
3591	pPrev = pCur;
3592	pCur = pCur->pNext;
3593	}
3594	if (!pCur)
3595	{
3596	if (pPrev)
3597	pPrev->pNext = pNewReg;
3598	else
3599	pSession->pDevExt->pComponentFactoryHead = pNewReg;
3600	rc = VINF_SUCCESS;
3601	}
3602	else
3603	rc = VERR_ALREADY_EXISTS;
3604
3605	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3606	}
3607
3608	if (RT_FAILURE(rc))
3609	RTMemFree(pNewReg);
3610	}
3611	else
3612	rc = VERR_NO_MEMORY;
3613	return rc;
3614	}
3615
3616
3617	/**
3618	* Deregister a component factory.
3619	*
3620	* @returns VBox status code.
3621	* @retval VINF_SUCCESS on success.
3622	* @retval VERR_NOT_FOUND if the factory wasn't registered.
3623	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
3624	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3625	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3626	*
3627	* @param pSession The SUPDRV session (must be a ring-0 session).
3628	* @param pFactory Pointer to the component factory registration structure
3629	* previously passed SUPR0ComponentRegisterFactory().
3630	*
3631	* @remarks This interface is also available via SUPR0IdcComponentDeregisterFactory.
3632	*/
3633	SUPR0DECL(int) SUPR0ComponentDeregisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
3634	{
3635	int rc;
3636
3637	/*
3638	* Validate parameters.
3639	*/
3640	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3641	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
3642	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
3643
3644	/*
3645	* Take the lock and look for the registration record.
3646	*/
3647	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3648	if (RT_SUCCESS(rc))
3649	{
3650	PSUPDRVFACTORYREG pPrev = NULL;
3651	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3652	while (pCur && pCur->pFactory != pFactory)
3653	{
3654	pPrev = pCur;
3655	pCur = pCur->pNext;
3656	}
3657	if (pCur)
3658	{
3659	if (!pPrev)
3660	pSession->pDevExt->pComponentFactoryHead = pCur->pNext;
3661	else
3662	pPrev->pNext = pCur->pNext;
3663
3664	pCur->pNext = NULL;
3665	pCur->pFactory = NULL;
3666	pCur->pSession = NULL;
3667	rc = VINF_SUCCESS;
3668	}
3669	else
3670	rc = VERR_NOT_FOUND;
3671
3672	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3673
3674	RTMemFree(pCur);
3675	}
3676	return rc;
3677	}
3678
3679
3680	/**
3681	* Queries a component factory.
3682	*
3683	* @returns VBox status code.
3684	* @retval VERR_INVALID_PARAMETER on invalid parameter.
3685	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
3686	* @retval VERR_SUPDRV_COMPONENT_NOT_FOUND if the component factory wasn't found.
3687	* @retval VERR_SUPDRV_INTERFACE_NOT_SUPPORTED if the interface wasn't supported.
3688	*
3689	* @param pSession The SUPDRV session.
3690	* @param pszName The name of the component factory.
3691	* @param pszInterfaceUuid The UUID of the factory interface (stringified).
3692	* @param ppvFactoryIf Where to store the factory interface.
3693	*/
3694	SUPR0DECL(int) SUPR0ComponentQueryFactory(PSUPDRVSESSION pSession, const char pszName, const char pszInterfaceUuid, void **ppvFactoryIf)
3695	{
3696	const char *pszEnd;
3697	size_t cchName;
3698	int rc;
3699
3700	/*
3701	* Validate parameters.
3702	*/
3703	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
3704
3705	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
3706	pszEnd = RTStrEnd(pszName, RT_SIZEOFMEMB(SUPDRVFACTORY, szName));
3707	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3708	cchName = pszEnd - pszName;
3709
3710	AssertPtrReturn(pszInterfaceUuid, VERR_INVALID_POINTER);
3711	pszEnd = RTStrEnd(pszInterfaceUuid, RTUUID_STR_LENGTH);
3712	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3713
3714	AssertPtrReturn(ppvFactoryIf, VERR_INVALID_POINTER);
3715	*ppvFactoryIf = NULL;
3716
3717	/*
3718	* Take the lock and try all factories by this name.
3719	*/
3720	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
3721	if (RT_SUCCESS(rc))
3722	{
3723	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
3724	rc = VERR_SUPDRV_COMPONENT_NOT_FOUND;
3725	while (pCur)
3726	{
3727	if ( pCur->cchName == cchName
3728	&& !memcmp(pCur->pFactory->szName, pszName, cchName))
3729	{
3730	void *pvFactory = pCur->pFactory->pfnQueryFactoryInterface(pCur->pFactory, pSession, pszInterfaceUuid);
3731	if (pvFactory)
3732	{
3733	*ppvFactoryIf = pvFactory;
3734	rc = VINF_SUCCESS;
3735	break;
3736	}
3737	rc = VERR_SUPDRV_INTERFACE_NOT_SUPPORTED;
3738	}
3739
3740	/* next */
3741	pCur = pCur->pNext;
3742	}
3743
3744	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
3745	}
3746	return rc;
3747	}
3748
3749
3750	/**
3751	* Adds a memory object to the session.
3752	*
3753	* @returns IPRT status code.
3754	* @param pMem Memory tracking structure containing the
3755	* information to track.
3756	* @param pSession The session.
3757	*/
3758	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
3759	{
3760	PSUPDRVBUNDLE pBundle;
3761
3762	/*
3763	* Find free entry and record the allocation.
3764	*/
3765	RTSpinlockAcquire(pSession->Spinlock);
3766	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3767	{
3768	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
3769	{
3770	unsigned i;
3771	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3772	{
3773	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
3774	{
3775	pBundle->cUsed++;
3776	pBundle->aMem[i] = *pMem;
3777	RTSpinlockRelease(pSession->Spinlock);
3778	return VINF_SUCCESS;
3779	}
3780	}
3781	AssertFailed(); /* !!this can't be happening!!! */
3782	}
3783	}
3784	RTSpinlockRelease(pSession->Spinlock);
3785
3786	/*
3787	* Need to allocate a new bundle.
3788	* Insert into the last entry in the bundle.
3789	*/
3790	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
3791	if (!pBundle)
3792	return VERR_NO_MEMORY;
3793
3794	/* take last entry. */
3795	pBundle->cUsed++;
3796	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
3797
3798	/* insert into list. */
3799	RTSpinlockAcquire(pSession->Spinlock);
3800	pBundle->pNext = pSession->Bundle.pNext;
3801	pSession->Bundle.pNext = pBundle;
3802	RTSpinlockRelease(pSession->Spinlock);
3803
3804	return VINF_SUCCESS;
3805	}
3806
3807
3808	/**
3809	* Releases a memory object referenced by pointer and type.
3810	*
3811	* @returns IPRT status code.
3812	* @param pSession Session data.
3813	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
3814	* @param eType Memory type.
3815	*/
3816	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
3817	{
3818	PSUPDRVBUNDLE pBundle;
3819
3820	/*
3821	* Validate input.
3822	*/
3823	if (!uPtr)
3824	{
3825	Log(("Illegal address %p\n", (void *)uPtr));
3826	return VERR_INVALID_PARAMETER;
3827	}
3828
3829	/*
3830	* Search for the address.
3831	*/
3832	RTSpinlockAcquire(pSession->Spinlock);
3833	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
3834	{
3835	if (pBundle->cUsed > 0)
3836	{
3837	unsigned i;
3838	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
3839	{
3840	if ( pBundle->aMem[i].eType == eType
3841	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
3842	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
3843	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
3844	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
3845	)
3846	{
3847	/* Make a copy of it and release it outside the spinlock. */
3848	SUPDRVMEMREF Mem = pBundle->aMem[i];
3849	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
3850	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
3851	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
3852	RTSpinlockRelease(pSession->Spinlock);
3853
3854	if (Mem.MapObjR3 != NIL_RTR0MEMOBJ)
3855	{
3856	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
3857	AssertRC(rc); /** @todo figure out how to handle this. */
3858	}
3859	if (Mem.MemObj != NIL_RTR0MEMOBJ)
3860	{
3861	int rc = RTR0MemObjFree(Mem.MemObj, true /* fFreeMappings */);
3862	AssertRC(rc); /** @todo figure out how to handle this. */
3863	}
3864	return VINF_SUCCESS;
3865	}
3866	}
3867	}
3868	}
3869	RTSpinlockRelease(pSession->Spinlock);
3870	Log(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
3871	return VERR_INVALID_PARAMETER;
3872	}
3873
3874
3875	/**
3876	* Opens an image. If it's the first time it's opened the call must upload
3877	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
3878	*
3879	* This is the 1st step of the loading.
3880	*
3881	* @returns IPRT status code.
3882	* @param pDevExt Device globals.
3883	* @param pSession Session data.
3884	* @param pReq The open request.
3885	*/
3886	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
3887	{
3888	int rc;
3889	PSUPDRVLDRIMAGE pImage;
3890	void *pv;
3891	size_t cchName = strlen(pReq->u.In.szName); /* (caller checked < 32). */
3892	LogFlow(("supdrvIOCtl_LdrOpen: szName=%s cbImageWithTabs=%d\n", pReq->u.In.szName, pReq->u.In.cbImageWithTabs));
3893
3894	/*
3895	* Check if we got an instance of the image already.
3896	*/
3897	supdrvLdrLock(pDevExt);
3898	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
3899	{
3900	if ( pImage->szName[cchName] == '\0'
3901	&& !memcmp(pImage->szName, pReq->u.In.szName, cchName))
3902	{
3903	/** @todo check cbImageBits and cbImageWithTabs here, if they differs that indicates that the images are different. */
3904	pImage->cUsage++;
3905	pReq->u.Out.pvImageBase = pImage->pvImage;
3906	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
3907	pReq->u.Out.fNativeLoader = pImage->fNative;
3908	supdrvLdrAddUsage(pSession, pImage);
3909	supdrvLdrUnlock(pDevExt);
3910	return VINF_SUCCESS;
3911	}
3912	}
3913	/* (not found - add it!) */
3914
3915	/*
3916	* Allocate memory.
3917	*/
3918	Assert(cchName < sizeof(pImage->szName));
3919	pv = RTMemAlloc(sizeof(SUPDRVLDRIMAGE));
3920	if (!pv)
3921	{
3922	supdrvLdrUnlock(pDevExt);
3923	Log(("supdrvIOCtl_LdrOpen: RTMemAlloc() failed\n"));
3924	return /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_2;
3925	}
3926
3927	/*
3928	* Setup and link in the LDR stuff.
3929	*/
3930	pImage = (PSUPDRVLDRIMAGE)pv;
3931	pImage->pvImage = NULL;
3932	pImage->pvImageAlloc = NULL;
3933	pImage->cbImageWithTabs = pReq->u.In.cbImageWithTabs;
3934	pImage->cbImageBits = pReq->u.In.cbImageBits;
3935	pImage->cSymbols = 0;
3936	pImage->paSymbols = NULL;
3937	pImage->pachStrTab = NULL;
3938	pImage->cbStrTab = 0;
3939	pImage->pfnModuleInit = NULL;
3940	pImage->pfnModuleTerm = NULL;
3941	pImage->pfnServiceReqHandler = NULL;
3942	pImage->uState = SUP_IOCTL_LDR_OPEN;
3943	pImage->cUsage = 1;
3944	pImage->pDevExt = pDevExt;
3945	memcpy(pImage->szName, pReq->u.In.szName, cchName + 1);
3946
3947	/*
3948	* Try load it using the native loader, if that isn't supported, fall back
3949	* on the older method.
3950	*/
3951	pImage->fNative = true;
3952	rc = supdrvOSLdrOpen(pDevExt, pImage, pReq->u.In.szFilename);
3953	if (rc == VERR_NOT_SUPPORTED)
3954	{
3955	pImage->pvImageAlloc = RTMemExecAlloc(pImage->cbImageBits + 31);
3956	pImage->pvImage = RT_ALIGN_P(pImage->pvImageAlloc, 32);
3957	pImage->fNative = false;
3958	rc = pImage->pvImageAlloc ? VINF_SUCCESS : VERR_NO_EXEC_MEMORY;
3959	}
3960	if (RT_FAILURE(rc))
3961	{
3962	supdrvLdrUnlock(pDevExt);
3963	RTMemFree(pImage);
3964	Log(("supdrvIOCtl_LdrOpen(%s): failed - %Rrc\n", pReq->u.In.szName, rc));
3965	return rc;
3966	}
3967	Assert(VALID_PTR(pImage->pvImage) \|\| RT_FAILURE(rc));
3968
3969	/*
3970	* Link it.
3971	*/
3972	pImage->pNext = pDevExt->pLdrImages;
3973	pDevExt->pLdrImages = pImage;
3974
3975	supdrvLdrAddUsage(pSession, pImage);
3976
3977	pReq->u.Out.pvImageBase = pImage->pvImage;
3978	pReq->u.Out.fNeedsLoading = true;
3979	pReq->u.Out.fNativeLoader = pImage->fNative;
3980	supdrvOSLdrNotifyOpened(pDevExt, pImage);
3981
3982	supdrvLdrUnlock(pDevExt);
3983	return VINF_SUCCESS;
3984	}
3985
3986
3987	/**
3988	* Worker that validates a pointer to an image entrypoint.
3989	*
3990	* @returns IPRT status code.
3991	* @param pDevExt The device globals.
3992	* @param pImage The loader image.
3993	* @param pv The pointer into the image.
3994	* @param fMayBeNull Whether it may be NULL.
3995	* @param pszWhat What is this entrypoint? (for logging)
3996	* @param pbImageBits The image bits prepared by ring-3.
3997	*
3998	* @remarks Will leave the lock on failure.
3999	*/
4000	static int supdrvLdrValidatePointer(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage, void *pv,
4001	bool fMayBeNull, const uint8_t pbImageBits, const char pszWhat)
4002	{
4003	if (!fMayBeNull \|\| pv)
4004	{
4005	if ((uintptr_t)pv - (uintptr_t)pImage->pvImage >= pImage->cbImageBits)
4006	{
4007	supdrvLdrUnlock(pDevExt);
4008	Log(("Out of range (%p LB %#x): %s=%p\n", pImage->pvImage, pImage->cbImageBits, pszWhat, pv));
4009	return VERR_INVALID_PARAMETER;
4010	}
4011
4012	if (pImage->fNative)
4013	{
4014	int rc = supdrvOSLdrValidatePointer(pDevExt, pImage, pv, pbImageBits);
4015	if (RT_FAILURE(rc))
4016	{
4017	supdrvLdrUnlock(pDevExt);
4018	Log(("Bad entry point address: %s=%p (rc=%Rrc)\n", pszWhat, pv, rc));
4019	return rc;
4020	}
4021	}
4022	}
4023	return VINF_SUCCESS;
4024	}
4025
4026
4027	/**
4028	* Loads the image bits.
4029	*
4030	* This is the 2nd step of the loading.
4031	*
4032	* @returns IPRT status code.
4033	* @param pDevExt Device globals.
4034	* @param pSession Session data.
4035	* @param pReq The request.
4036	*/
4037	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
4038	{
4039	PSUPDRVLDRUSAGE pUsage;
4040	PSUPDRVLDRIMAGE pImage;
4041	int rc;
4042	LogFlow(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImageWithBits=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImageWithTabs));
4043
4044	/*
4045	* Find the ldr image.
4046	*/
4047	supdrvLdrLock(pDevExt);
4048	pUsage = pSession->pLdrUsage;
4049	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4050	pUsage = pUsage->pNext;
4051	if (!pUsage)
4052	{
4053	supdrvLdrUnlock(pDevExt);
4054	Log(("SUP_IOCTL_LDR_LOAD: couldn't find image!\n"));
4055	return VERR_INVALID_HANDLE;
4056	}
4057	pImage = pUsage->pImage;
4058
4059	/*
4060	* Validate input.
4061	*/
4062	if ( pImage->cbImageWithTabs != pReq->u.In.cbImageWithTabs
4063	\|\| pImage->cbImageBits != pReq->u.In.cbImageBits)
4064	{
4065	supdrvLdrUnlock(pDevExt);
4066	Log(("SUP_IOCTL_LDR_LOAD: image size mismatch!! %d(prep) != %d(load) or %d != %d\n",
4067	pImage->cbImageWithTabs, pReq->u.In.cbImageWithTabs, pImage->cbImageBits, pReq->u.In.cbImageBits));
4068	return VERR_INVALID_HANDLE;
4069	}
4070
4071	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
4072	{
4073	unsigned uState = pImage->uState;
4074	supdrvLdrUnlock(pDevExt);
4075	if (uState != SUP_IOCTL_LDR_LOAD)
4076	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
4077	return VERR_ALREADY_LOADED;
4078	}
4079
4080	switch (pReq->u.In.eEPType)
4081	{
4082	case SUPLDRLOADEP_NOTHING:
4083	break;
4084
4085	case SUPLDRLOADEP_VMMR0:
4086	rc = supdrvLdrValidatePointer( pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0, false, pReq->u.In.abImage, "pvVMMR0");
4087	if (RT_SUCCESS(rc))
4088	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt, false, pReq->u.In.abImage, "pvVMMR0EntryInt");
4089	if (RT_SUCCESS(rc))
4090	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, false, pReq->u.In.abImage, "pvVMMR0EntryFast");
4091	if (RT_SUCCESS(rc))
4092	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx, false, pReq->u.In.abImage, "pvVMMR0EntryEx");
4093	if (RT_FAILURE(rc))
4094	return rc;
4095	break;
4096
4097	case SUPLDRLOADEP_SERVICE:
4098	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.EP.Service.pfnServiceReq, false, pReq->u.In.abImage, "pfnServiceReq");
4099	if (RT_FAILURE(rc))
4100	return rc;
4101	if ( pReq->u.In.EP.Service.apvReserved[0] != NIL_RTR0PTR
4102	\|\| pReq->u.In.EP.Service.apvReserved[1] != NIL_RTR0PTR
4103	\|\| pReq->u.In.EP.Service.apvReserved[2] != NIL_RTR0PTR)
4104	{
4105	supdrvLdrUnlock(pDevExt);
4106	Log(("Out of range (%p LB %#x): apvReserved={%p,%p,%p} MBZ!\n",
4107	pImage->pvImage, pReq->u.In.cbImageWithTabs,
4108	pReq->u.In.EP.Service.apvReserved[0],
4109	pReq->u.In.EP.Service.apvReserved[1],
4110	pReq->u.In.EP.Service.apvReserved[2]));
4111	return VERR_INVALID_PARAMETER;
4112	}
4113	break;
4114
4115	default:
4116	supdrvLdrUnlock(pDevExt);
4117	Log(("Invalid eEPType=%d\n", pReq->u.In.eEPType));
4118	return VERR_INVALID_PARAMETER;
4119	}
4120
4121	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleInit, true, pReq->u.In.abImage, "pfnModuleInit");
4122	if (RT_FAILURE(rc))
4123	return rc;
4124	rc = supdrvLdrValidatePointer(pDevExt, pImage, pReq->u.In.pfnModuleTerm, true, pReq->u.In.abImage, "pfnModuleTerm");
4125	if (RT_FAILURE(rc))
4126	return rc;
4127
4128	/*
4129	* Allocate and copy the tables.
4130	* (No need to do try/except as this is a buffered request.)
4131	*/
4132	pImage->cbStrTab = pReq->u.In.cbStrTab;
4133	if (pImage->cbStrTab)
4134	{
4135	pImage->pachStrTab = (char *)RTMemAlloc(pImage->cbStrTab);
4136	if (pImage->pachStrTab)
4137	memcpy(pImage->pachStrTab, &pReq->u.In.abImage[pReq->u.In.offStrTab], pImage->cbStrTab);
4138	else
4139	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_3;
4140	}
4141
4142	pImage->cSymbols = pReq->u.In.cSymbols;
4143	if (RT_SUCCESS(rc) && pImage->cSymbols)
4144	{
4145	size_t cbSymbols = pImage->cSymbols * sizeof(SUPLDRSYM);
4146	pImage->paSymbols = (PSUPLDRSYM)RTMemAlloc(cbSymbols);
4147	if (pImage->paSymbols)
4148	memcpy(pImage->paSymbols, &pReq->u.In.abImage[pReq->u.In.offSymbols], cbSymbols);
4149	else
4150	rc = /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_4;
4151	}
4152
4153	/*
4154	* Copy the bits / complete native loading.
4155	*/
4156	if (RT_SUCCESS(rc))
4157	{
4158	pImage->uState = SUP_IOCTL_LDR_LOAD;
4159	pImage->pfnModuleInit = pReq->u.In.pfnModuleInit;
4160	pImage->pfnModuleTerm = pReq->u.In.pfnModuleTerm;
4161
4162	if (pImage->fNative)
4163	rc = supdrvOSLdrLoad(pDevExt, pImage, pReq->u.In.abImage, pReq);
4164	else
4165	{
4166	memcpy(pImage->pvImage, &pReq->u.In.abImage[0], pImage->cbImageBits);
4167	Log(("vboxdrv: Loaded '%s' at %p\n", pImage->szName, pImage->pvImage));
4168	}
4169	}
4170
4171	/*
4172	* Update any entry points.
4173	*/
4174	if (RT_SUCCESS(rc))
4175	{
4176	switch (pReq->u.In.eEPType)
4177	{
4178	default:
4179	case SUPLDRLOADEP_NOTHING:
4180	rc = VINF_SUCCESS;
4181	break;
4182	case SUPLDRLOADEP_VMMR0:
4183	rc = supdrvLdrSetVMMR0EPs(pDevExt, pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
4184	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
4185	break;
4186	case SUPLDRLOADEP_SERVICE:
4187	pImage->pfnServiceReqHandler = pReq->u.In.EP.Service.pfnServiceReq;
4188	rc = VINF_SUCCESS;
4189	break;
4190	}
4191	}
4192
4193	/*
4194	* On success call the module initialization.
4195	*/
4196	LogFlow(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
4197	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
4198	{
4199	Log(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
4200	pDevExt->pLdrInitImage = pImage;
4201	pDevExt->hLdrInitThread = RTThreadNativeSelf();
4202	rc = pImage->pfnModuleInit(pImage);
4203	pDevExt->pLdrInitImage = NULL;
4204	pDevExt->hLdrInitThread = NIL_RTNATIVETHREAD;
4205	if (RT_FAILURE(rc) && pDevExt->pvVMMR0 == pImage->pvImage)
4206	supdrvLdrUnsetVMMR0EPs(pDevExt);
4207	}
4208
4209	if (RT_FAILURE(rc))
4210	{
4211	/* Inform the tracing component in case ModuleInit registered TPs. */
4212	supdrvTracerModuleUnloading(pDevExt, pImage);
4213
4214	pImage->uState = SUP_IOCTL_LDR_OPEN;
4215	pImage->pfnModuleInit = NULL;
4216	pImage->pfnModuleTerm = NULL;
4217	pImage->pfnServiceReqHandler= NULL;
4218	pImage->cbStrTab = 0;
4219	RTMemFree(pImage->pachStrTab);
4220	pImage->pachStrTab = NULL;
4221	RTMemFree(pImage->paSymbols);
4222	pImage->paSymbols = NULL;
4223	pImage->cSymbols = 0;
4224	}
4225
4226	supdrvLdrUnlock(pDevExt);
4227	return rc;
4228	}
4229
4230
4231	/**
4232	* Frees a previously loaded (prep'ed) image.
4233	*
4234	* @returns IPRT status code.
4235	* @param pDevExt Device globals.
4236	* @param pSession Session data.
4237	* @param pReq The request.
4238	*/
4239	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
4240	{
4241	int rc;
4242	PSUPDRVLDRUSAGE pUsagePrev;
4243	PSUPDRVLDRUSAGE pUsage;
4244	PSUPDRVLDRIMAGE pImage;
4245	LogFlow(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
4246
4247	/*
4248	* Find the ldr image.
4249	*/
4250	supdrvLdrLock(pDevExt);
4251	pUsagePrev = NULL;
4252	pUsage = pSession->pLdrUsage;
4253	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4254	{
4255	pUsagePrev = pUsage;
4256	pUsage = pUsage->pNext;
4257	}
4258	if (!pUsage)
4259	{
4260	supdrvLdrUnlock(pDevExt);
4261	Log(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
4262	return VERR_INVALID_HANDLE;
4263	}
4264
4265	/*
4266	* Check if we can remove anything.
4267	*/
4268	rc = VINF_SUCCESS;
4269	pImage = pUsage->pImage;
4270	if (pImage->cUsage <= 1 \|\| pUsage->cUsage <= 1)
4271	{
4272	/*
4273	* Check if there are any objects with destructors in the image, if
4274	* so leave it for the session cleanup routine so we get a chance to
4275	* clean things up in the right order and not leave them all dangling.
4276	*/
4277	RTSpinlockAcquire(pDevExt->Spinlock);
4278	if (pImage->cUsage <= 1)
4279	{
4280	PSUPDRVOBJ pObj;
4281	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4282	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4283	{
4284	rc = VERR_DANGLING_OBJECTS;
4285	break;
4286	}
4287	}
4288	else
4289	{
4290	PSUPDRVUSAGE pGenUsage;
4291	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
4292	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4293	{
4294	rc = VERR_DANGLING_OBJECTS;
4295	break;
4296	}
4297	}
4298	RTSpinlockRelease(pDevExt->Spinlock);
4299	if (rc == VINF_SUCCESS)
4300	{
4301	/* unlink it */
4302	if (pUsagePrev)
4303	pUsagePrev->pNext = pUsage->pNext;
4304	else
4305	pSession->pLdrUsage = pUsage->pNext;
4306
4307	/* free it */
4308	pUsage->pImage = NULL;
4309	pUsage->pNext = NULL;
4310	RTMemFree(pUsage);
4311
4312	/*
4313	* Dereference the image.
4314	*/
4315	if (pImage->cUsage <= 1)
4316	supdrvLdrFree(pDevExt, pImage);
4317	else
4318	pImage->cUsage--;
4319	}
4320	else
4321	{
4322	Log(("supdrvIOCtl_LdrFree: Dangling objects in %p/%s!\n", pImage->pvImage, pImage->szName));
4323	rc = VINF_SUCCESS; /** @todo BRANCH-2.1: remove this after branching. */
4324	}
4325	}
4326	else
4327	{
4328	/*
4329	* Dereference both image and usage.
4330	*/
4331	pImage->cUsage--;
4332	pUsage->cUsage--;
4333	}
4334
4335	supdrvLdrUnlock(pDevExt);
4336	return rc;
4337	}
4338
4339
4340	/**
4341	* Gets the address of a symbol in an open image.
4342	*
4343	* @returns IPRT status code.
4344	* @param pDevExt Device globals.
4345	* @param pSession Session data.
4346	* @param pReq The request buffer.
4347	*/
4348	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
4349	{
4350	PSUPDRVLDRIMAGE pImage;
4351	PSUPDRVLDRUSAGE pUsage;
4352	uint32_t i;
4353	PSUPLDRSYM paSyms;
4354	const char *pchStrings;
4355	const size_t cbSymbol = strlen(pReq->u.In.szSymbol) + 1;
4356	void *pvSymbol = NULL;
4357	int rc = VERR_GENERAL_FAILURE;
4358	Log3(("supdrvIOCtl_LdrGetSymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
4359
4360	/*
4361	* Find the ldr image.
4362	*/
4363	supdrvLdrLock(pDevExt);
4364	pUsage = pSession->pLdrUsage;
4365	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
4366	pUsage = pUsage->pNext;
4367	if (!pUsage)
4368	{
4369	supdrvLdrUnlock(pDevExt);
4370	Log(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
4371	return VERR_INVALID_HANDLE;
4372	}
4373	pImage = pUsage->pImage;
4374	if (pImage->uState != SUP_IOCTL_LDR_LOAD)
4375	{
4376	unsigned uState = pImage->uState;
4377	supdrvLdrUnlock(pDevExt);
4378	Log(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", uState, uState)); NOREF(uState);
4379	return VERR_ALREADY_LOADED;
4380	}
4381
4382	/*
4383	* Search the symbol strings.
4384	*
4385	* Note! The int32_t is for native loading on solaris where the data
4386	* and text segments are in very different places.
4387	*/
4388	pchStrings = pImage->pachStrTab;
4389	paSyms = pImage->paSymbols;
4390	for (i = 0; i < pImage->cSymbols; i++)
4391	{
4392	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4393	&& !memcmp(pchStrings + paSyms[i].offName, pReq->u.In.szSymbol, cbSymbol))
4394	{
4395	pvSymbol = (uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol;
4396	rc = VINF_SUCCESS;
4397	break;
4398	}
4399	}
4400	supdrvLdrUnlock(pDevExt);
4401	pReq->u.Out.pvSymbol = pvSymbol;
4402	return rc;
4403	}
4404
4405
4406	/**
4407	* Gets the address of a symbol in an open image or the support driver.
4408	*
4409	* @returns VINF_SUCCESS on success.
4410	* @returns
4411	* @param pDevExt Device globals.
4412	* @param pSession Session data.
4413	* @param pReq The request buffer.
4414	*/
4415	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq)
4416	{
4417	int rc = VINF_SUCCESS;
4418	const char *pszSymbol = pReq->u.In.pszSymbol;
4419	const char *pszModule = pReq->u.In.pszModule;
4420	size_t cbSymbol;
4421	char const *pszEnd;
4422	uint32_t i;
4423
4424	/*
4425	* Input validation.
4426	*/
4427	AssertPtrReturn(pszSymbol, VERR_INVALID_POINTER);
4428	pszEnd = RTStrEnd(pszSymbol, 512);
4429	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4430	cbSymbol = pszEnd - pszSymbol + 1;
4431
4432	if (pszModule)
4433	{
4434	AssertPtrReturn(pszModule, VERR_INVALID_POINTER);
4435	pszEnd = RTStrEnd(pszModule, 64);
4436	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
4437	}
4438	Log3(("supdrvIDC_LdrGetSymbol: pszModule=%p:{%s} pszSymbol=%p:{%s}\n", pszModule, pszModule, pszSymbol, pszSymbol));
4439
4440
4441	if ( !pszModule
4442	\|\| !strcmp(pszModule, "SupDrv"))
4443	{
4444	/*
4445	* Search the support driver export table.
4446	*/
4447	for (i = 0; i < RT_ELEMENTS(g_aFunctions); i++)
4448	if (!strcmp(g_aFunctions[i].szName, pszSymbol))
4449	{
4450	pReq->u.Out.pfnSymbol = g_aFunctions[i].pfn;
4451	break;
4452	}
4453	}
4454	else
4455	{
4456	/*
4457	* Find the loader image.
4458	*/
4459	PSUPDRVLDRIMAGE pImage;
4460
4461	supdrvLdrLock(pDevExt);
4462
4463	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
4464	if (!strcmp(pImage->szName, pszModule))
4465	break;
4466	if (pImage && pImage->uState == SUP_IOCTL_LDR_LOAD)
4467	{
4468	/*
4469	* Search the symbol strings.
4470	*/
4471	const char *pchStrings = pImage->pachStrTab;
4472	PCSUPLDRSYM paSyms = pImage->paSymbols;
4473	for (i = 0; i < pImage->cSymbols; i++)
4474	{
4475	if ( paSyms[i].offName + cbSymbol <= pImage->cbStrTab
4476	&& !memcmp(pchStrings + paSyms[i].offName, pszSymbol, cbSymbol))
4477	{
4478	/*
4479	* Found it! Calc the symbol address and add a reference to the module.
4480	*/
4481	pReq->u.Out.pfnSymbol = (PFNRT)((uint8_t *)pImage->pvImage + (int32_t)paSyms[i].offSymbol);
4482	rc = supdrvLdrAddUsage(pSession, pImage);
4483	break;
4484	}
4485	}
4486	}
4487	else
4488	rc = pImage ? VERR_WRONG_ORDER : VERR_MODULE_NOT_FOUND;
4489
4490	supdrvLdrUnlock(pDevExt);
4491	}
4492	return rc;
4493	}
4494
4495
4496	/**
4497	* Updates the VMMR0 entry point pointers.
4498	*
4499	* @returns IPRT status code.
4500	* @param pDevExt Device globals.
4501	* @param pSession Session data.
4502	* @param pVMMR0 VMMR0 image handle.
4503	* @param pvVMMR0EntryInt VMMR0EntryInt address.
4504	* @param pvVMMR0EntryFast VMMR0EntryFast address.
4505	* @param pvVMMR0EntryEx VMMR0EntryEx address.
4506	* @remark Caller must own the loader mutex.
4507	*/
4508	static int supdrvLdrSetVMMR0EPs(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx)
4509	{
4510	int rc = VINF_SUCCESS;
4511	LogFlow(("supdrvLdrSetR0EP pvVMMR0=%p pvVMMR0EntryInt=%p\n", pvVMMR0, pvVMMR0EntryInt));
4512
4513
4514	/*
4515	* Check if not yet set.
4516	*/
4517	if (!pDevExt->pvVMMR0)
4518	{
4519	pDevExt->pvVMMR0 = pvVMMR0;
4520	pDevExt->pfnVMMR0EntryInt = pvVMMR0EntryInt;
4521	pDevExt->pfnVMMR0EntryFast = pvVMMR0EntryFast;
4522	pDevExt->pfnVMMR0EntryEx = pvVMMR0EntryEx;
4523	}
4524	else
4525	{
4526	/*
4527	* Return failure or success depending on whether the values match or not.
4528	*/
4529	if ( pDevExt->pvVMMR0 != pvVMMR0
4530	\|\| (void *)pDevExt->pfnVMMR0EntryInt != pvVMMR0EntryInt
4531	\|\| (void *)pDevExt->pfnVMMR0EntryFast != pvVMMR0EntryFast
4532	\|\| (void *)pDevExt->pfnVMMR0EntryEx != pvVMMR0EntryEx)
4533	{
4534	AssertMsgFailed(("SUP_IOCTL_LDR_SETR0EP: Already set pointing to a different module!\n"));
4535	rc = VERR_INVALID_PARAMETER;
4536	}
4537	}
4538	return rc;
4539	}
4540
4541
4542	/**
4543	* Unsets the VMMR0 entry point installed by supdrvLdrSetR0EP.
4544	*
4545	* @param pDevExt Device globals.
4546	*/
4547	static void supdrvLdrUnsetVMMR0EPs(PSUPDRVDEVEXT pDevExt)
4548	{
4549	pDevExt->pvVMMR0 = NULL;
4550	pDevExt->pfnVMMR0EntryInt = NULL;
4551	pDevExt->pfnVMMR0EntryFast = NULL;
4552	pDevExt->pfnVMMR0EntryEx = NULL;
4553	}
4554
4555
4556	/**
4557	* Adds a usage reference in the specified session of an image.
4558	*
4559	* Called while owning the loader semaphore.
4560	*
4561	* @returns VINF_SUCCESS on success and VERR_NO_MEMORY on failure.
4562	* @param pSession Session in question.
4563	* @param pImage Image which the session is using.
4564	*/
4565	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage)
4566	{
4567	PSUPDRVLDRUSAGE pUsage;
4568	LogFlow(("supdrvLdrAddUsage: pImage=%p\n", pImage));
4569
4570	/*
4571	* Referenced it already?
4572	*/
4573	pUsage = pSession->pLdrUsage;
4574	while (pUsage)
4575	{
4576	if (pUsage->pImage == pImage)
4577	{
4578	pUsage->cUsage++;
4579	return VINF_SUCCESS;
4580	}
4581	pUsage = pUsage->pNext;
4582	}
4583
4584	/*
4585	* Allocate new usage record.
4586	*/
4587	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
4588	AssertReturn(pUsage, /VERR_NO_MEMORY/ VERR_INTERNAL_ERROR_5);
4589	pUsage->cUsage = 1;
4590	pUsage->pImage = pImage;
4591	pUsage->pNext = pSession->pLdrUsage;
4592	pSession->pLdrUsage = pUsage;
4593	return VINF_SUCCESS;
4594	}
4595
4596
4597	/**
4598	* Frees a load image.
4599	*
4600	* @param pDevExt Pointer to device extension.
4601	* @param pImage Pointer to the image we're gonna free.
4602	* This image must exit!
4603	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
4604	*/
4605	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
4606	{
4607	PSUPDRVLDRIMAGE pImagePrev;
4608	LogFlow(("supdrvLdrFree: pImage=%p\n", pImage));
4609
4610	/* find it - arg. should've used doubly linked list. */
4611	Assert(pDevExt->pLdrImages);
4612	pImagePrev = NULL;
4613	if (pDevExt->pLdrImages != pImage)
4614	{
4615	pImagePrev = pDevExt->pLdrImages;
4616	while (pImagePrev->pNext != pImage)
4617	pImagePrev = pImagePrev->pNext;
4618	Assert(pImagePrev->pNext == pImage);
4619	}
4620
4621	/* unlink */
4622	if (pImagePrev)
4623	pImagePrev->pNext = pImage->pNext;
4624	else
4625	pDevExt->pLdrImages = pImage->pNext;
4626
4627	/* check if this is VMMR0.r0 unset its entry point pointers. */
4628	if (pDevExt->pvVMMR0 == pImage->pvImage)
4629	supdrvLdrUnsetVMMR0EPs(pDevExt);
4630
4631	/* check for objects with destructors in this image. (Shouldn't happen.) */
4632	if (pDevExt->pObjs)
4633	{
4634	unsigned cObjs = 0;
4635	PSUPDRVOBJ pObj;
4636	RTSpinlockAcquire(pDevExt->Spinlock);
4637	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4638	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImageBits))
4639	{
4640	pObj->pfnDestructor = NULL;
4641	cObjs++;
4642	}
4643	RTSpinlockRelease(pDevExt->Spinlock);
4644	if (cObjs)
4645	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
4646	}
4647
4648	/* call termination function if fully loaded. */
4649	if ( pImage->pfnModuleTerm
4650	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
4651	{
4652	LogFlow(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
4653	pImage->pfnModuleTerm(pImage);
4654	}
4655
4656	/* Inform the tracing component. */
4657	supdrvTracerModuleUnloading(pDevExt, pImage);
4658
4659	/* do native unload if appropriate. */
4660	if (pImage->fNative)
4661	supdrvOSLdrUnload(pDevExt, pImage);
4662
4663	/* free the image */
4664	pImage->cUsage = 0;
4665	pImage->pDevExt = NULL;
4666	pImage->pNext = NULL;
4667	pImage->uState = SUP_IOCTL_LDR_FREE;
4668	RTMemExecFree(pImage->pvImageAlloc, pImage->cbImageBits + 31);
4669	pImage->pvImageAlloc = NULL;
4670	RTMemFree(pImage->pachStrTab);
4671	pImage->pachStrTab = NULL;
4672	RTMemFree(pImage->paSymbols);
4673	pImage->paSymbols = NULL;
4674	RTMemFree(pImage);
4675	}
4676
4677
4678	/**
4679	* Acquires the loader lock.
4680	*
4681	* @returns IPRT status code.
4682	* @param pDevExt The device extension.
4683	*/
4684	DECLINLINE(int) supdrvLdrLock(PSUPDRVDEVEXT pDevExt)
4685	{
4686	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4687	int rc = RTSemMutexRequest(pDevExt->mtxLdr, RT_INDEFINITE_WAIT);
4688	#else
4689	int rc = RTSemFastMutexRequest(pDevExt->mtxLdr);
4690	#endif
4691	AssertRC(rc);
4692	return rc;
4693	}
4694
4695
4696	/**
4697	* Releases the loader lock.
4698	*
4699	* @returns IPRT status code.
4700	* @param pDevExt The device extension.
4701	*/
4702	DECLINLINE(int) supdrvLdrUnlock(PSUPDRVDEVEXT pDevExt)
4703	{
4704	#ifdef SUPDRV_USE_MUTEX_FOR_LDR
4705	return RTSemMutexRelease(pDevExt->mtxLdr);
4706	#else
4707	return RTSemFastMutexRelease(pDevExt->mtxLdr);
4708	#endif
4709	}
4710
4711
4712	/**
4713	* Implements the service call request.
4714	*
4715	* @returns VBox status code.
4716	* @param pDevExt The device extension.
4717	* @param pSession The calling session.
4718	* @param pReq The request packet, valid.
4719	*/
4720	static int supdrvIOCtl_CallServiceModule(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPCALLSERVICE pReq)
4721	{
4722	#if !defined(RT_OS_WINDOWS) \|\| defined(DEBUG)
4723	int rc;
4724
4725	/*
4726	* Find the module first in the module referenced by the calling session.
4727	*/
4728	rc = supdrvLdrLock(pDevExt);
4729	if (RT_SUCCESS(rc))
4730	{
4731	PFNSUPR0SERVICEREQHANDLER pfnServiceReqHandler = NULL;
4732	PSUPDRVLDRUSAGE pUsage;
4733
4734	for (pUsage = pSession->pLdrUsage; pUsage; pUsage = pUsage->pNext)
4735	if ( pUsage->pImage->pfnServiceReqHandler
4736	&& !strcmp(pUsage->pImage->szName, pReq->u.In.szName))
4737	{
4738	pfnServiceReqHandler = pUsage->pImage->pfnServiceReqHandler;
4739	break;
4740	}
4741	supdrvLdrUnlock(pDevExt);
4742
4743	if (pfnServiceReqHandler)
4744	{
4745	/*
4746	* Call it.
4747	*/
4748	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_SERVICE_SIZE(0))
4749	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, NULL);
4750	else
4751	rc = pfnServiceReqHandler(pSession, pReq->u.In.uOperation, pReq->u.In.u64Arg, (PSUPR0SERVICEREQHDR)&pReq->abReqPkt[0]);
4752	}
4753	else
4754	rc = VERR_SUPDRV_SERVICE_NOT_FOUND;
4755	}
4756
4757	/* log it */
4758	if ( RT_FAILURE(rc)
4759	&& rc != VERR_INTERRUPTED
4760	&& rc != VERR_TIMEOUT)
4761	Log(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4762	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4763	else
4764	Log4(("SUP_IOCTL_CALL_SERVICE: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
4765	rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
4766	return rc;
4767	#else /* RT_OS_WINDOWS && !DEBUG */
4768	return VERR_NOT_IMPLEMENTED;
4769	#endif /* RT_OS_WINDOWS && !DEBUG */
4770	}
4771
4772
4773	/**
4774	* Implements the logger settings request.
4775	*
4776	* @returns VBox status code.
4777	* @param pDevExt The device extension.
4778	* @param pSession The caller's session.
4779	* @param pReq The request.
4780	*/
4781	static int supdrvIOCtl_LoggerSettings(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLOGGERSETTINGS pReq)
4782	{
4783	const char *pszGroup = &pReq->u.In.szStrings[pReq->u.In.offGroups];
4784	const char *pszFlags = &pReq->u.In.szStrings[pReq->u.In.offFlags];
4785	const char *pszDest = &pReq->u.In.szStrings[pReq->u.In.offDestination];
4786	PRTLOGGER pLogger = NULL;
4787	int rc;
4788
4789	/*
4790	* Some further validation.
4791	*/
4792	switch (pReq->u.In.fWhat)
4793	{
4794	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4795	case SUPLOGGERSETTINGS_WHAT_CREATE:
4796	break;
4797
4798	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4799	if (pszGroup \|\| pszFlags \|\| *pszDest)
4800	return VERR_INVALID_PARAMETER;
4801	if (pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_RELEASE)
4802	return VERR_ACCESS_DENIED;
4803	break;
4804
4805	default:
4806	return VERR_INTERNAL_ERROR;
4807	}
4808
4809	/*
4810	* Get the logger.
4811	*/
4812	switch (pReq->u.In.fWhich)
4813	{
4814	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4815	pLogger = RTLogGetDefaultInstance();
4816	break;
4817
4818	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4819	pLogger = RTLogRelDefaultInstance();
4820	break;
4821
4822	default:
4823	return VERR_INTERNAL_ERROR;
4824	}
4825
4826	/*
4827	* Do the job.
4828	*/
4829	switch (pReq->u.In.fWhat)
4830	{
4831	case SUPLOGGERSETTINGS_WHAT_SETTINGS:
4832	if (pLogger)
4833	{
4834	rc = RTLogFlags(pLogger, pszFlags);
4835	if (RT_SUCCESS(rc))
4836	rc = RTLogGroupSettings(pLogger, pszGroup);
4837	NOREF(pszDest);
4838	}
4839	else
4840	rc = VERR_NOT_FOUND;
4841	break;
4842
4843	case SUPLOGGERSETTINGS_WHAT_CREATE:
4844	{
4845	if (pLogger)
4846	rc = VERR_ALREADY_EXISTS;
4847	else
4848	{
4849	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
4850
4851	rc = RTLogCreate(&pLogger,
4852	0 /* fFlags */,
4853	pszGroup,
4854	pReq->u.In.fWhich == SUPLOGGERSETTINGS_WHICH_DEBUG
4855	? "VBOX_LOG"
4856	: "VBOX_RELEASE_LOG",
4857	RT_ELEMENTS(s_apszGroups),
4858	s_apszGroups,
4859	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER,
4860	NULL);
4861	if (RT_SUCCESS(rc))
4862	{
4863	rc = RTLogFlags(pLogger, pszFlags);
4864	NOREF(pszDest);
4865	if (RT_SUCCESS(rc))
4866	{
4867	switch (pReq->u.In.fWhich)
4868	{
4869	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4870	pLogger = RTLogSetDefaultInstance(pLogger);
4871	break;
4872	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4873	pLogger = RTLogRelSetDefaultInstance(pLogger);
4874	break;
4875	}
4876	}
4877	RTLogDestroy(pLogger);
4878	}
4879	}
4880	break;
4881	}
4882
4883	case SUPLOGGERSETTINGS_WHAT_DESTROY:
4884	switch (pReq->u.In.fWhich)
4885	{
4886	case SUPLOGGERSETTINGS_WHICH_DEBUG:
4887	pLogger = RTLogSetDefaultInstance(NULL);
4888	break;
4889	case SUPLOGGERSETTINGS_WHICH_RELEASE:
4890	pLogger = RTLogRelSetDefaultInstance(NULL);
4891	break;
4892	}
4893	rc = RTLogDestroy(pLogger);
4894	break;
4895
4896	default:
4897	{
4898	rc = VERR_INTERNAL_ERROR;
4899	break;
4900	}
4901	}
4902
4903	return rc;
4904	}
4905
4906
4907	/**
4908	* Creates the GIP.
4909	*
4910	* @returns VBox status code.
4911	* @param pDevExt Instance data. GIP stuff may be updated.
4912	*/
4913	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
4914	{
4915	PSUPGLOBALINFOPAGE pGip;
4916	RTHCPHYS HCPhysGip;
4917	uint32_t u32SystemResolution;
4918	uint32_t u32Interval;
4919	unsigned cCpus;
4920	int rc;
4921
4922
4923	LogFlow(("supdrvGipCreate:\n"));
4924
4925	/* assert order */
4926	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
4927	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
4928	Assert(!pDevExt->pGipTimer);
4929
4930	/*
4931	* Check the CPU count.
4932	*/
4933	cCpus = RTMpGetArraySize();
4934	if ( cCpus > RTCPUSET_MAX_CPUS
4935	\|\| cCpus > 256 /ApicId is used for the mappings/)
4936	{
4937	SUPR0Printf("VBoxDrv: Too many CPUs (%u) for the GIP (max %u)\n", cCpus, RT_MIN(RTCPUSET_MAX_CPUS, 256));
4938	return VERR_TOO_MANY_CPUS;
4939	}
4940
4941	/*
4942	* Allocate a contiguous set of pages with a default kernel mapping.
4943	*/
4944	rc = RTR0MemObjAllocCont(&pDevExt->GipMemObj, RT_UOFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), false /fExecutable/);
4945	if (RT_FAILURE(rc))
4946	{
4947	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
4948	return rc;
4949	}
4950	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
4951	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
4952
4953	/*
4954	* Find a reasonable update interval and initialize the structure.
4955	*/
4956	u32Interval = u32SystemResolution = RTTimerGetSystemGranularity();
4957	while (u32Interval < 10000000 /* 10 ms */)
4958	u32Interval += u32SystemResolution;
4959
4960	supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), 1000000000 / u32Interval /=Hz/, cCpus);
4961
4962	/*
4963	* Create the timer.
4964	* If CPU_ALL isn't supported we'll have to fall back to synchronous mode.
4965	*/
4966	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
4967	{
4968	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, RTTIMER_FLAGS_CPU_ALL, supdrvGipAsyncTimer, pDevExt);
4969	if (rc == VERR_NOT_SUPPORTED)
4970	{
4971	OSDBGPRINT(("supdrvGipCreate: omni timer not supported, falling back to synchronous mode\n"));
4972	pGip->u32Mode = SUPGIPMODE_SYNC_TSC;
4973	}
4974	}
4975	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
4976	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0, supdrvGipSyncTimer, pDevExt);
4977	if (RT_SUCCESS(rc))
4978	{
4979	rc = RTMpNotificationRegister(supdrvGipMpEvent, pDevExt);
4980	if (RT_SUCCESS(rc))
4981	{
4982	rc = RTMpOnAll(supdrvGipInitOnCpu, pDevExt, pGip);
4983	if (RT_SUCCESS(rc))
4984	{
4985	/*
4986	* We're good.
4987	*/
4988	Log(("supdrvGipCreate: %u ns interval.\n", u32Interval));
4989	g_pSUPGlobalInfoPage = pGip;
4990	return VINF_SUCCESS;
4991	}
4992
4993	OSDBGPRINT(("supdrvGipCreate: RTMpOnAll failed with rc=%Rrc\n", rc));
4994	RTMpNotificationDeregister(supdrvGipMpEvent, pDevExt);
4995
4996	}
4997	else
4998	OSDBGPRINT(("supdrvGipCreate: failed to register MP event notfication. rc=%Rrc\n", rc));
4999	}
5000	else
5001	{
5002	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %u ns interval. rc=%Rrc\n", u32Interval, rc));
5003	Assert(!pDevExt->pGipTimer);
5004	}
5005	supdrvGipDestroy(pDevExt);
5006	return rc;
5007	}
5008
5009
5010	/**
5011	* Terminates the GIP.
5012	*
5013	* @param pDevExt Instance data. GIP stuff may be updated.
5014	*/
5015	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
5016	{
5017	int rc;
5018	#ifdef DEBUG_DARWIN_GIP
5019	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
5020	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
5021	pDevExt->pGipTimer, pDevExt->GipMemObj));
5022	#endif
5023
5024	/*
5025	* Invalid the GIP data.
5026	*/
5027	if (pDevExt->pGip)
5028	{
5029	supdrvGipTerm(pDevExt->pGip);
5030	pDevExt->pGip = NULL;
5031	}
5032	g_pSUPGlobalInfoPage = NULL;
5033
5034	/*
5035	* Destroy the timer and free the GIP memory object.
5036	*/
5037	if (pDevExt->pGipTimer)
5038	{
5039	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
5040	pDevExt->pGipTimer = NULL;
5041	}
5042
5043	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
5044	{
5045	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
5046	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
5047	}
5048
5049	/*
5050	* Finally, make sure we've release the system timer resolution request
5051	* if one actually succeeded and is still pending.
5052	*/
5053	if (pDevExt->u32SystemTimerGranularityGrant)
5054	{
5055	rc = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
5056	pDevExt->u32SystemTimerGranularityGrant = 0;
5057	}
5058	}
5059
5060
5061	/**
5062	* Timer callback function sync GIP mode.
5063	* @param pTimer The timer.
5064	* @param pvUser The device extension.
5065	*/
5066	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
5067	{
5068	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
5069	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5070	uint64_t u64TSC = ASMReadTSC();
5071	uint64_t NanoTS = RTTimeSystemNanoTS();
5072
5073	supdrvGipUpdate(pDevExt, NanoTS, u64TSC, NIL_RTCPUID, iTick);
5074
5075	ASMSetFlags(fOldFlags);
5076	}
5077
5078
5079	/**
5080	* Timer callback function for async GIP mode.
5081	* @param pTimer The timer.
5082	* @param pvUser The device extension.
5083	*/
5084	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
5085	{
5086	RTCCUINTREG fOldFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
5087	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5088	RTCPUID idCpu = RTMpCpuId();
5089	uint64_t u64TSC = ASMReadTSC();
5090	uint64_t NanoTS = RTTimeSystemNanoTS();
5091
5092	/** @todo reset the transaction number and whatnot when iTick == 1. */
5093	if (pDevExt->idGipMaster == idCpu)
5094	supdrvGipUpdate(pDevExt, NanoTS, u64TSC, idCpu, iTick);
5095	else
5096	supdrvGipUpdatePerCpu(pDevExt, NanoTS, u64TSC, idCpu, ASMGetApicId(), iTick);
5097
5098	ASMSetFlags(fOldFlags);
5099	}
5100
5101
5102	/**
5103	* Finds our (@a idCpu) entry, or allocates a new one if not found.
5104	*
5105	* @returns Index of the CPU in the cache set.
5106	* @param pGip The GIP.
5107	* @param idCpu The CPU ID.
5108	*/
5109	static uint32_t supdrvGipCpuIndexFromCpuId(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
5110	{
5111	uint32_t i, cTries;
5112
5113	/*
5114	* ASSUMES that CPU IDs are constant.
5115	*/
5116	for (i = 0; i < pGip->cCpus; i++)
5117	if (pGip->aCPUs[i].idCpu == idCpu)
5118	return i;
5119
5120	cTries = 0;
5121	do
5122	{
5123	for (i = 0; i < pGip->cCpus; i++)
5124	{
5125	bool fRc;
5126	ASMAtomicCmpXchgSize(&pGip->aCPUs[i].idCpu, idCpu, NIL_RTCPUID, fRc);
5127	if (fRc)
5128	return i;
5129	}
5130	} while (cTries++ < 32);
5131	AssertReleaseFailed();
5132	return i - 1;
5133	}
5134
5135
5136	/**
5137	* The calling CPU should be accounted as online, update GIP accordingly.
5138	*
5139	* This is used by supdrvGipMpEvent as well as the supdrvGipCreate.
5140	*
5141	* @param pDevExt The device extension.
5142	* @param idCpu The CPU ID.
5143	*/
5144	static void supdrvGipMpEventOnline(PSUPDRVDEVEXT pDevExt, RTCPUID idCpu)
5145	{
5146	int iCpuSet = 0;
5147	uint16_t idApic = UINT16_MAX;
5148	uint32_t i = 0;
5149	uint64_t u64NanoTS = 0;
5150	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5151
5152	AssertPtrReturnVoid(pGip);
5153	AssertRelease(idCpu == RTMpCpuId());
5154	Assert(pGip->cPossibleCpus == RTMpGetCount());
5155
5156	/*
5157	* Do this behind a spinlock with interrupts disabled as this can fire
5158	* on all CPUs simultaneously, see #6110.
5159	*/
5160	RTSpinlockAcquire(pDevExt->hGipSpinlock);
5161
5162	/*
5163	* Update the globals.
5164	*/
5165	ASMAtomicWriteU16(&pGip->cPresentCpus, RTMpGetPresentCount());
5166	ASMAtomicWriteU16(&pGip->cOnlineCpus, RTMpGetOnlineCount());
5167	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5168	if (iCpuSet >= 0)
5169	{
5170	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5171	RTCpuSetAddByIndex(&pGip->OnlineCpuSet, iCpuSet);
5172	RTCpuSetAddByIndex(&pGip->PresentCpuSet, iCpuSet);
5173	}
5174
5175	/*
5176	* Update the entry.
5177	*/
5178	u64NanoTS = RTTimeSystemNanoTS() - pGip->u32UpdateIntervalNS;
5179	i = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
5180	supdrvGipInitCpu(pGip, &pGip->aCPUs[i], u64NanoTS);
5181	idApic = ASMGetApicId();
5182	ASMAtomicWriteU16(&pGip->aCPUs[i].idApic, idApic);
5183	ASMAtomicWriteS16(&pGip->aCPUs[i].iCpuSet, (int16_t)iCpuSet);
5184	ASMAtomicWriteSize(&pGip->aCPUs[i].idCpu, idCpu);
5185
5186	/*
5187	* Update the APIC ID and CPU set index mappings.
5188	*/
5189	ASMAtomicWriteU16(&pGip->aiCpuFromApicId[idApic], i);
5190	ASMAtomicWriteU16(&pGip->aiCpuFromCpuSetIdx[iCpuSet], i);
5191
5192	/* commit it */
5193	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_ONLINE);
5194
5195	RTSpinlockReleaseNoInts(pDevExt->hGipSpinlock);
5196	}
5197
5198
5199	/**
5200	* The CPU should be accounted as offline, update the GIP accordingly.
5201	*
5202	* This is used by supdrvGipMpEvent.
5203	*
5204	* @param pDevExt The device extension.
5205	* @param idCpu The CPU ID.
5206	*/
5207	static void supdrvGipMpEventOffline(PSUPDRVDEVEXT pDevExt, RTCPUID idCpu)
5208	{
5209	int iCpuSet;
5210	unsigned i;
5211
5212	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5213
5214	AssertPtrReturnVoid(pGip);
5215	RTSpinlockAcquire(pDevExt->hGipSpinlock);
5216
5217	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
5218	AssertReturnVoid(iCpuSet >= 0);
5219
5220	i = pGip->aiCpuFromCpuSetIdx[iCpuSet];
5221	AssertReturnVoid(i < pGip->cCpus);
5222	AssertReturnVoid(pGip->aCPUs[i].idCpu == idCpu);
5223
5224	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
5225	RTCpuSetDelByIndex(&pGip->OnlineCpuSet, iCpuSet);
5226
5227	/* commit it */
5228	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_OFFLINE);
5229
5230	RTSpinlockReleaseNoInts(pDevExt->hGipSpinlock);
5231	}
5232
5233
5234	/**
5235	* Multiprocessor event notification callback.
5236	*
5237	* This is used to make sure that the GIP master gets passed on to
5238	* another CPU. It also updates the associated CPU data.
5239	*
5240	* @param enmEvent The event.
5241	* @param idCpu The cpu it applies to.
5242	* @param pvUser Pointer to the device extension.
5243	*
5244	* @remarks This function -must- fire on the newly online'd CPU for the
5245	* RTMPEVENT_ONLINE case and can fire on any CPU for the
5246	* RTMPEVENT_OFFLINE case.
5247	*/
5248	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser)
5249	{
5250	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
5251	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5252
5253	AssertRelease(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
5254
5255	/*
5256	* Update the GIP CPU data.
5257	*/
5258	if (pGip)
5259	{
5260	switch (enmEvent)
5261	{
5262	case RTMPEVENT_ONLINE:
5263	AssertRelease(idCpu == RTMpCpuId());
5264	supdrvGipMpEventOnline(pDevExt, idCpu);
5265	break;
5266	case RTMPEVENT_OFFLINE:
5267	supdrvGipMpEventOffline(pDevExt, idCpu);
5268	break;
5269
5270	}
5271	}
5272
5273	/*
5274	* Make sure there is a master GIP.
5275	*/
5276	if (enmEvent == RTMPEVENT_OFFLINE)
5277	{
5278	RTCPUID idGipMaster = ASMAtomicReadU32(&pDevExt->idGipMaster);
5279	if (idGipMaster == idCpu)
5280	{
5281	/*
5282	* Find a new GIP master.
5283	*/
5284	bool fIgnored;
5285	unsigned i;
5286	RTCPUID idNewGipMaster = NIL_RTCPUID;
5287	RTCPUSET OnlineCpus;
5288	RTMpGetOnlineSet(&OnlineCpus);
5289
5290	for (i = 0; i < RTCPUSET_MAX_CPUS; i++)
5291	{
5292	RTCPUID idCurCpu = RTMpCpuIdFromSetIndex(i);
5293	if ( RTCpuSetIsMember(&OnlineCpus, idCurCpu)
5294	&& idCurCpu != idGipMaster)
5295	{
5296	idNewGipMaster = idCurCpu;
5297	break;
5298	}
5299	}
5300
5301	Log(("supdrvGipMpEvent: Gip master %#lx -> %#lx\n", (long)idGipMaster, (long)idNewGipMaster));
5302	ASMAtomicCmpXchgSize(&pDevExt->idGipMaster, idNewGipMaster, idGipMaster, fIgnored);
5303	NOREF(fIgnored);
5304	}
5305	}
5306	}
5307
5308
5309	/**
5310	* Callback used by supdrvDetermineAsyncTSC to read the TSC on a CPU.
5311	*
5312	* @param idCpu Ignored.
5313	* @param pvUser1 Where to put the TSC.
5314	* @param pvUser2 Ignored.
5315	*/
5316	static DECLCALLBACK(void) supdrvDetermineAsyncTscWorker(RTCPUID idCpu, void pvUser1, void pvUser2)
5317	{
5318	#if 1
5319	ASMAtomicWriteU64((uint64_t volatile *)pvUser1, ASMReadTSC());
5320	#else
5321	(uint64_t )pvUser1 = ASMReadTSC();
5322	#endif
5323	}
5324
5325
5326	/**
5327	* Determine if Async GIP mode is required because of TSC drift.
5328	*
5329	* When using the default/normal timer code it is essential that the time stamp counter
5330	* (TSC) runs never backwards, that is, a read operation to the counter should return
5331	* a bigger value than any previous read operation. This is guaranteed by the latest
5332	* AMD CPUs and by newer Intel CPUs which never enter the C2 state (P4). In any other
5333	* case we have to choose the asynchronous timer mode.
5334	*
5335	* @param poffMin Pointer to the determined difference between different cores.
5336	* @return false if the time stamp counters appear to be synchronized, true otherwise.
5337	*/
5338	static bool supdrvDetermineAsyncTsc(uint64_t *poffMin)
5339	{
5340	/*
5341	* Just iterate all the cpus 8 times and make sure that the TSC is
5342	* ever increasing. We don't bother taking TSC rollover into account.
5343	*/
5344	int iEndCpu = RTMpGetArraySize();
5345	int iCpu;
5346	int cLoops = 8;
5347	bool fAsync = false;
5348	int rc = VINF_SUCCESS;
5349	uint64_t offMax = 0;
5350	uint64_t offMin = ~(uint64_t)0;
5351	uint64_t PrevTsc = ASMReadTSC();
5352
5353	while (cLoops-- > 0)
5354	{
5355	for (iCpu = 0; iCpu < iEndCpu; iCpu++)
5356	{
5357	uint64_t CurTsc;
5358	rc = RTMpOnSpecific(RTMpCpuIdFromSetIndex(iCpu), supdrvDetermineAsyncTscWorker, &CurTsc, NULL);
5359	if (RT_SUCCESS(rc))
5360	{
5361	if (CurTsc <= PrevTsc)
5362	{
5363	fAsync = true;
5364	offMin = offMax = PrevTsc - CurTsc;
5365	Log(("supdrvDetermineAsyncTsc: iCpu=%d cLoops=%d CurTsc=%llx PrevTsc=%llx\n",
5366	iCpu, cLoops, CurTsc, PrevTsc));
5367	break;
5368	}
5369
5370	/* Gather statistics (except the first time). */
5371	if (iCpu != 0 \|\| cLoops != 7)
5372	{
5373	uint64_t off = CurTsc - PrevTsc;
5374	if (off < offMin)
5375	offMin = off;
5376	if (off > offMax)
5377	offMax = off;
5378	Log2(("%d/%d: off=%llx\n", cLoops, iCpu, off));
5379	}
5380
5381	/* Next */
5382	PrevTsc = CurTsc;
5383	}
5384	else if (rc == VERR_NOT_SUPPORTED)
5385	break;
5386	else
5387	AssertMsg(rc == VERR_CPU_NOT_FOUND \|\| rc == VERR_CPU_OFFLINE, ("%d\n", rc));
5388	}
5389
5390	/* broke out of the loop. */
5391	if (iCpu < iEndCpu)
5392	break;
5393	}
5394
5395	poffMin = offMin; / Almost RTMpOnSpecific profiling. */
5396	Log(("supdrvDetermineAsyncTsc: returns %d; iEndCpu=%d rc=%d offMin=%llx offMax=%llx\n",
5397	fAsync, iEndCpu, rc, offMin, offMax));
5398	#if !defined(RT_OS_SOLARIS) && !defined(RT_OS_OS2) && !defined(RT_OS_WINDOWS)
5399	OSDBGPRINT(("vboxdrv: fAsync=%d offMin=%#lx offMax=%#lx\n", fAsync, (long)offMin, (long)offMax));
5400	#endif
5401	return fAsync;
5402	}
5403
5404
5405	/**
5406	* Determine the GIP TSC mode.
5407	*
5408	* @returns The most suitable TSC mode.
5409	* @param pDevExt Pointer to the device instance data.
5410	*/
5411	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt)
5412	{
5413	/*
5414	* On SMP we're faced with two problems:
5415	* (1) There might be a skew between the CPU, so that cpu0
5416	* returns a TSC that is slightly different from cpu1.
5417	* (2) Power management (and other things) may cause the TSC
5418	* to run at a non-constant speed, and cause the speed
5419	* to be different on the cpus. This will result in (1).
5420	*
5421	* So, on SMP systems we'll have to select the ASYNC update method
5422	* if there are symptoms of these problems.
5423	*/
5424	if (RTMpGetCount() > 1)
5425	{
5426	uint32_t uEAX, uEBX, uECX, uEDX;
5427	uint64_t u64DiffCoresIgnored;
5428
5429	/* Permit the user and/or the OS specific bits to force async mode. */
5430	if (supdrvOSGetForcedAsyncTscMode(pDevExt))
5431	return SUPGIPMODE_ASYNC_TSC;
5432
5433	/* Try check for current differences between the cpus. */
5434	if (supdrvDetermineAsyncTsc(&u64DiffCoresIgnored))
5435	return SUPGIPMODE_ASYNC_TSC;
5436
5437	/*
5438	* If the CPU supports power management and is an AMD one we
5439	* won't trust it unless it has the TscInvariant bit is set.
5440	*/
5441	/* Check for "AuthenticAMD" */
5442	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
5443	if ( uEAX >= 1
5444	&& uEBX == X86_CPUID_VENDOR_AMD_EBX
5445	&& uECX == X86_CPUID_VENDOR_AMD_ECX
5446	&& uEDX == X86_CPUID_VENDOR_AMD_EDX)
5447	{
5448	/* Check for APM support and that TscInvariant is cleared. */
5449	ASMCpuId(0x80000000, &uEAX, &uEBX, &uECX, &uEDX);
5450	if (uEAX >= 0x80000007)
5451	{
5452	ASMCpuId(0x80000007, &uEAX, &uEBX, &uECX, &uEDX);
5453	if ( !(uEDX & RT_BIT(8))/* TscInvariant */
5454	&& (uEDX & 0x3e)) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
5455	return SUPGIPMODE_ASYNC_TSC;
5456	}
5457	}
5458	}
5459	return SUPGIPMODE_SYNC_TSC;
5460	}
5461
5462
5463	/**
5464	* Initializes per-CPU GIP information.
5465	*
5466	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5467	* @param pCpu Pointer to which GIP CPU to initalize.
5468	* @param u64NanoTS The current nanosecond timestamp.
5469	*/
5470	static void supdrvGipInitCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pCpu, uint64_t u64NanoTS)
5471	{
5472	pCpu->u32TransactionId = 2;
5473	pCpu->u64NanoTS = u64NanoTS;
5474	pCpu->u64TSC = ASMReadTSC();
5475
5476	ASMAtomicWriteSize(&pCpu->enmState, SUPGIPCPUSTATE_INVALID);
5477	ASMAtomicWriteSize(&pCpu->idCpu, NIL_RTCPUID);
5478	ASMAtomicWriteS16(&pCpu->iCpuSet, -1);
5479	ASMAtomicWriteU16(&pCpu->idApic, UINT16_MAX);
5480
5481	/*
5482	* We don't know the following values until we've executed updates.
5483	* So, we'll just pretend it's a 4 GHz CPU and adjust the history it on
5484	* the 2nd timer callout.
5485	*/
5486	pCpu->u64CpuHz = _4G + 1; /* tstGIP-2 depends on this. */
5487	pCpu->u32UpdateIntervalTSC
5488	= pCpu->au32TSCHistory[0]
5489	= pCpu->au32TSCHistory[1]
5490	= pCpu->au32TSCHistory[2]
5491	= pCpu->au32TSCHistory[3]
5492	= pCpu->au32TSCHistory[4]
5493	= pCpu->au32TSCHistory[5]
5494	= pCpu->au32TSCHistory[6]
5495	= pCpu->au32TSCHistory[7]
5496	= (uint32_t)(_4G / pGip->u32UpdateHz);
5497	}
5498
5499
5500	/**
5501	* Initializes the GIP data.
5502	*
5503	* @param pDevExt Pointer to the device instance data.
5504	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5505	* @param HCPhys The physical address of the GIP.
5506	* @param u64NanoTS The current nanosecond timestamp.
5507	* @param uUpdateHz The update frequency.
5508	* @param cCpus The CPU count.
5509	*/
5510	static void supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
5511	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned cCpus)
5512	{
5513	size_t const cbGip = RT_ALIGN_Z(RT_OFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]), PAGE_SIZE);
5514	unsigned i;
5515	#ifdef DEBUG_DARWIN_GIP
5516	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5517	#else
5518	LogFlow(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
5519	#endif
5520
5521	/*
5522	* Initialize the structure.
5523	*/
5524	memset(pGip, 0, cbGip);
5525	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
5526	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
5527	pGip->u32Mode = supdrvGipDeterminTscMode(pDevExt);
5528	pGip->cCpus = (uint16_t)cCpus;
5529	pGip->cPages = (uint16_t)(cbGip / PAGE_SIZE);
5530	pGip->u32UpdateHz = uUpdateHz;
5531	pGip->u32UpdateIntervalNS = 1000000000 / uUpdateHz;
5532	pGip->u64NanoTSLastUpdateHz = u64NanoTS;
5533	RTCpuSetEmpty(&pGip->OnlineCpuSet);
5534	RTCpuSetEmpty(&pGip->PresentCpuSet);
5535	RTMpGetSet(&pGip->PossibleCpuSet);
5536	pGip->cOnlineCpus = RTMpGetOnlineCount();
5537	pGip->cPresentCpus = RTMpGetPresentCount();
5538	pGip->cPossibleCpus = RTMpGetCount();
5539	pGip->idCpuMax = RTMpGetMaxCpuId();
5540	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromApicId); i++)
5541	pGip->aiCpuFromApicId[i] = 0;
5542	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx); i++)
5543	pGip->aiCpuFromCpuSetIdx[i] = UINT16_MAX;
5544
5545	for (i = 0; i < cCpus; i++)
5546	supdrvGipInitCpu(pGip, &pGip->aCPUs[i], u64NanoTS);
5547
5548	/*
5549	* Link it to the device extension.
5550	*/
5551	pDevExt->pGip = pGip;
5552	pDevExt->HCPhysGip = HCPhys;
5553	pDevExt->cGipUsers = 0;
5554	}
5555
5556
5557	/**
5558	* On CPU initialization callback for RTMpOnAll.
5559	*
5560	* @param idCpu The CPU ID.
5561	* @param pvUser1 The device extension.
5562	* @param pvUser2 The GIP.
5563	*/
5564	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2)
5565	{
5566	/* This is good enough, even though it will update some of the globals a
5567	bit to much. */
5568	supdrvGipMpEventOnline((PSUPDRVDEVEXT)pvUser1, idCpu);
5569	}
5570
5571
5572	/**
5573	* Invalidates the GIP data upon termination.
5574	*
5575	* @param pGip Pointer to the read-write kernel mapping of the GIP.
5576	*/
5577	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
5578	{
5579	unsigned i;
5580	pGip->u32Magic = 0;
5581	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
5582	{
5583	pGip->aCPUs[i].u64NanoTS = 0;
5584	pGip->aCPUs[i].u64TSC = 0;
5585	pGip->aCPUs[i].iTSCHistoryHead = 0;
5586	}
5587	}
5588
5589
5590	/**
5591	* Worker routine for supdrvGipUpdate and supdrvGipUpdatePerCpu that
5592	* updates all the per cpu data except the transaction id.
5593	*
5594	* @param pDevExt The device extension.
5595	* @param pGipCpu Pointer to the per cpu data.
5596	* @param u64NanoTS The current time stamp.
5597	* @param u64TSC The current TSC.
5598	* @param iTick The current timer tick.
5599	*/
5600	static void supdrvGipDoUpdateCpu(PSUPDRVDEVEXT pDevExt, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick)
5601	{
5602	uint64_t u64TSCDelta;
5603	uint32_t u32UpdateIntervalTSC;
5604	uint32_t u32UpdateIntervalTSCSlack;
5605	unsigned iTSCHistoryHead;
5606	uint64_t u64CpuHz;
5607	uint32_t u32TransactionId;
5608
5609	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5610	AssertPtrReturnVoid(pGip);
5611
5612	/* Delta between this and the previous update. */
5613	ASMAtomicUoWriteU32(&pGipCpu->u32PrevUpdateIntervalNS, (uint32_t)(u64NanoTS - pGipCpu->u64NanoTS));
5614
5615	/*
5616	* Update the NanoTS.
5617	*/
5618	ASMAtomicWriteU64(&pGipCpu->u64NanoTS, u64NanoTS);
5619
5620	/*
5621	* Calc TSC delta.
5622	*/
5623	/** @todo validate the NanoTS delta, don't trust the OS to call us when it should... */
5624	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
5625	ASMAtomicWriteU64(&pGipCpu->u64TSC, u64TSC);
5626
5627	if (u64TSCDelta >> 32)
5628	{
5629	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
5630	pGipCpu->cErrors++;
5631	}
5632
5633	/*
5634	* On the 2nd and 3rd callout, reset the history with the current TSC
5635	* interval since the values entered by supdrvGipInit are totally off.
5636	* The interval on the 1st callout completely unreliable, the 2nd is a bit
5637	* better, while the 3rd should be most reliable.
5638	*/
5639	u32TransactionId = pGipCpu->u32TransactionId;
5640	if (RT_UNLIKELY( ( u32TransactionId == 5
5641	\|\| u32TransactionId == 7)
5642	&& ( iTick == 2
5643	\|\| iTick == 3) ))
5644	{
5645	unsigned i;
5646	for (i = 0; i < RT_ELEMENTS(pGipCpu->au32TSCHistory); i++)
5647	ASMAtomicUoWriteU32(&pGipCpu->au32TSCHistory[i], (uint32_t)u64TSCDelta);
5648	}
5649
5650	/*
5651	* TSC History.
5652	*/
5653	Assert(RT_ELEMENTS(pGipCpu->au32TSCHistory) == 8);
5654	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
5655	ASMAtomicWriteU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
5656	ASMAtomicWriteU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
5657
5658	/*
5659	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
5660	*/
5661	if (pGip->u32UpdateHz >= 1000)
5662	{
5663	uint32_t u32;
5664	u32 = pGipCpu->au32TSCHistory[0];
5665	u32 += pGipCpu->au32TSCHistory[1];
5666	u32 += pGipCpu->au32TSCHistory[2];
5667	u32 += pGipCpu->au32TSCHistory[3];
5668	u32 >>= 2;
5669	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
5670	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
5671	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
5672	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
5673	u32UpdateIntervalTSC >>= 2;
5674	u32UpdateIntervalTSC += u32;
5675	u32UpdateIntervalTSC >>= 1;
5676
5677	/* Value chosen for a 2GHz Athlon64 running linux 2.6.10/11, . */
5678	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
5679	}
5680	else if (pGip->u32UpdateHz >= 90)
5681	{
5682	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5683	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
5684	u32UpdateIntervalTSC >>= 1;
5685
5686	/* value chosen on a 2GHz thinkpad running windows */
5687	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
5688	}
5689	else
5690	{
5691	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
5692
5693	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
5694	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
5695	}
5696	ASMAtomicWriteU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
5697
5698	/*
5699	* CpuHz.
5700	*/
5701	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, pGip->u32UpdateHz);
5702	ASMAtomicWriteU64(&pGipCpu->u64CpuHz, u64CpuHz);
5703	}
5704
5705
5706	/**
5707	* Updates the GIP.
5708	*
5709	* @param pDevExt The device extension.
5710	* @param u64NanoTS The current nanosecond timesamp.
5711	* @param u64TSC The current TSC timesamp.
5712	* @param idCpu The CPU ID.
5713	* @param iTick The current timer tick.
5714	*/
5715	static void supdrvGipUpdate(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick)
5716	{
5717	/*
5718	* Determine the relevant CPU data.
5719	*/
5720	PSUPGIPCPU pGipCpu;
5721	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5722	AssertPtrReturnVoid(pGip);
5723
5724	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
5725	pGipCpu = &pGip->aCPUs[0];
5726	else
5727	{
5728	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
5729	if (RT_UNLIKELY(iCpu >= pGip->cCpus))
5730	return;
5731	pGipCpu = &pGip->aCPUs[iCpu];
5732	if (RT_UNLIKELY(pGipCpu->idCpu != idCpu))
5733	return;
5734	}
5735
5736	/*
5737	* Start update transaction.
5738	*/
5739	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5740	{
5741	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
5742	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5743	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5744	pGipCpu->cErrors++;
5745	return;
5746	}
5747
5748	/*
5749	* Recalc the update frequency every 0x800th time.
5750	*/
5751	if (!(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
5752	{
5753	if (pGip->u64NanoTSLastUpdateHz)
5754	{
5755	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
5756	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
5757	uint32_t u32UpdateHz = (uint32_t)((UINT64_C(1000000000) * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
5758	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
5759	{
5760	ASMAtomicWriteU32(&pGip->u32UpdateHz, u32UpdateHz);
5761	ASMAtomicWriteU32(&pGip->u32UpdateIntervalNS, 1000000000 / u32UpdateHz);
5762	}
5763	#endif
5764	}
5765	ASMAtomicWriteU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS);
5766	}
5767
5768	/*
5769	* Update the data.
5770	*/
5771	supdrvGipDoUpdateCpu(pDevExt, pGipCpu, u64NanoTS, u64TSC, iTick);
5772
5773	/*
5774	* Complete transaction.
5775	*/
5776	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5777	}
5778
5779
5780	/**
5781	* Updates the per cpu GIP data for the calling cpu.
5782	*
5783	* @param pDevExt The device extension.
5784	* @param u64NanoTS The current nanosecond timesamp.
5785	* @param u64TSC The current TSC timesamp.
5786	* @param idCpu The CPU ID.
5787	* @param idApic The APIC id for the CPU index.
5788	* @param iTick The current timer tick.
5789	*/
5790	static void supdrvGipUpdatePerCpu(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC,
5791	RTCPUID idCpu, uint8_t idApic, uint64_t iTick)
5792	{
5793	uint32_t iCpu;
5794	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
5795
5796	/*
5797	* Avoid a potential race when a CPU online notification doesn't fire on
5798	* the onlined CPU but the tick creeps in before the event notification is
5799	* run.
5800	*/
5801	if (RT_UNLIKELY(iTick == 1))
5802	{
5803	iCpu = supdrvGipCpuIndexFromCpuId(pGip, idCpu);
5804	if (pGip->aCPUs[iCpu].enmState == SUPGIPCPUSTATE_OFFLINE)
5805	supdrvGipMpEventOnline(pDevExt, idCpu);
5806	}
5807
5808	iCpu = pGip->aiCpuFromApicId[idApic];
5809	if (RT_LIKELY(iCpu < pGip->cCpus))
5810	{
5811	PSUPGIPCPU pGipCpu = &pGip->aCPUs[iCpu];
5812	if (pGipCpu->idCpu == idCpu)
5813	{
5814	/*
5815	* Start update transaction.
5816	*/
5817	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
5818	{
5819	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
5820	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5821	pGipCpu->cErrors++;
5822	return;
5823	}
5824
5825	/*
5826	* Update the data.
5827	*/
5828	supdrvGipDoUpdateCpu(pDevExt, pGipCpu, u64NanoTS, u64TSC, iTick);
5829
5830	/*
5831	* Complete transaction.
5832	*/
5833	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
5834	}
5835	}
5836	}
5837

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

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDrv.c@ 41254

Download in other formats: