SUPDrv.c@ 10682

Last change on this file since 10682 was 10680, checked in by vboxsync, 17 years ago
Increase minor version, needed some timestamp functions.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 171.1 KB

Line
1	/* $Revision: 10680 $ */
2	/** @file
3	* VBoxDrv - The VirtualBox Support Driver - Common code.
4	*/
5
6	/*
7	* Copyright (C) 2006-2007 Sun Microsystems, Inc.
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	* Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa
27	* Clara, CA 95054 USA or visit http://www.sun.com if you need
28	* additional information or have any questions.
29	*/
30
31	/*******************************************************************************
32	* Header Files *
33	*******************************************************************************/
34	#define LOG_GROUP LOG_GROUP_SUP_DRV
35	#include "SUPDrvInternal.h"
36	#ifndef PAGE_SHIFT
37	# include <iprt/param.h>
38	#endif
39	#include <iprt/alloc.h>
40	#include <iprt/semaphore.h>
41	#include <iprt/spinlock.h>
42	#include <iprt/thread.h>
43	#include <iprt/process.h>
44	#include <iprt/mp.h>
45	#include <iprt/cpuset.h>
46	#include <iprt/uuid.h>
47	#include <VBox/log.h>
48	#include <VBox/err.h>
49	/* VBox/x86.h not compatible with the Linux kernel sources */
50	#ifdef RT_OS_LINUX
51	# define X86_CPUID_VENDOR_AMD_EBX 0x68747541
52	# define X86_CPUID_VENDOR_AMD_ECX 0x444d4163
53	# define X86_CPUID_VENDOR_AMD_EDX 0x69746e65
54	#else
55	# include <VBox/x86.h>
56	#endif
57
58	/*
59	* Logging assignments:
60	* Log - useful stuff, like failures.
61	* LogFlow - program flow, except the really noisy bits.
62	* Log2 - Cleanup and IDTE
63	* Log3 - Loader flow noise.
64	* Log4 - Call VMMR0 flow noise.
65	* Log5 - Native yet-to-be-defined noise.
66	* Log6 - Native ioctl flow noise.
67	*
68	* Logging requires BUILD_TYPE=debug and possibly changes to the logger
69	* instanciation in log-vbox.c(pp).
70	*/
71
72
73	/*******************************************************************************
74	* Defined Constants And Macros *
75	*******************************************************************************/
76	/* from x86.h - clashes with linux thus this duplication */
77	#undef X86_CR0_PG
78	#define X86_CR0_PG RT_BIT(31)
79	#undef X86_CR0_PE
80	#define X86_CR0_PE RT_BIT(0)
81	#undef X86_CPUID_AMD_FEATURE_EDX_NX
82	#define X86_CPUID_AMD_FEATURE_EDX_NX RT_BIT(20)
83	#undef MSR_K6_EFER
84	#define MSR_K6_EFER 0xc0000080
85	#undef MSR_K6_EFER_NXE
86	#define MSR_K6_EFER_NXE RT_BIT(11)
87	#undef MSR_K6_EFER_LMA
88	#define MSR_K6_EFER_LMA RT_BIT(10)
89	#undef X86_CR4_PGE
90	#define X86_CR4_PGE RT_BIT(7)
91	#undef X86_CR4_PAE
92	#define X86_CR4_PAE RT_BIT(5)
93	#undef X86_CPUID_AMD_FEATURE_EDX_LONG_MODE
94	#define X86_CPUID_AMD_FEATURE_EDX_LONG_MODE RT_BIT(29)
95
96
97	/** The frequency by which we recalculate the u32UpdateHz and
98	* u32UpdateIntervalNS GIP members. The value must be a power of 2. */
99	#define GIP_UPDATEHZ_RECALC_FREQ 0x800
100
101	/**
102	* Validates a session pointer.
103	*
104	* @returns true/false accordingly.
105	* @param pSession The session.
106	*/
107	#define SUP_IS_SESSION_VALID(pSession) \
108	( VALID_PTR(pSession) \
109	&& pSession->u32Cookie == BIRD_INV)
110
111	/** @def VBOX_SVN_REV
112	* The makefile should define this if it can. */
113	#ifndef VBOX_SVN_REV
114	# define VBOX_SVN_REV 0
115	#endif
116
117
118	/*******************************************************************************
119	* Global Variables *
120	*******************************************************************************/
121	/**
122	* Array of the R0 SUP API.
123	*/
124	static SUPFUNC g_aFunctions[] =
125	{
126	/* name function */
127	{ "SUPR0ComponentRegisterFactory", (void *)SUPR0ComponentRegisterFactory },
128	{ "SUPR0ComponentDeregisterFactory", (void *)SUPR0ComponentDeregisterFactory },
129	{ "SUPR0ComponentQueryFactory", (void *)SUPR0ComponentQueryFactory },
130	{ "SUPR0ObjRegister", (void *)SUPR0ObjRegister },
131	{ "SUPR0ObjAddRef", (void *)SUPR0ObjAddRef },
132	{ "SUPR0ObjRelease", (void *)SUPR0ObjRelease },
133	{ "SUPR0ObjVerifyAccess", (void *)SUPR0ObjVerifyAccess },
134	{ "SUPR0LockMem", (void *)SUPR0LockMem },
135	{ "SUPR0UnlockMem", (void *)SUPR0UnlockMem },
136	{ "SUPR0ContAlloc", (void *)SUPR0ContAlloc },
137	{ "SUPR0ContFree", (void *)SUPR0ContFree },
138	{ "SUPR0LowAlloc", (void *)SUPR0LowAlloc },
139	{ "SUPR0LowFree", (void *)SUPR0LowFree },
140	{ "SUPR0MemAlloc", (void *)SUPR0MemAlloc },
141	{ "SUPR0MemGetPhys", (void *)SUPR0MemGetPhys },
142	{ "SUPR0MemFree", (void *)SUPR0MemFree },
143	{ "SUPR0PageAlloc", (void *)SUPR0PageAlloc },
144	{ "SUPR0PageFree", (void *)SUPR0PageFree },
145	{ "SUPR0Printf", (void *)SUPR0Printf },
146	{ "RTMemAlloc", (void *)RTMemAlloc },
147	{ "RTMemAllocZ", (void *)RTMemAllocZ },
148	{ "RTMemFree", (void *)RTMemFree },
149	/{ "RTMemDup", (void )RTMemDup },*/
150	{ "RTMemRealloc", (void *)RTMemRealloc },
151	{ "RTR0MemObjAllocLow", (void *)RTR0MemObjAllocLow },
152	{ "RTR0MemObjAllocPage", (void *)RTR0MemObjAllocPage },
153	{ "RTR0MemObjAllocPhys", (void *)RTR0MemObjAllocPhys },
154	{ "RTR0MemObjAllocPhysNC", (void *)RTR0MemObjAllocPhysNC },
155	{ "RTR0MemObjAllocCont", (void *)RTR0MemObjAllocCont },
156	{ "RTR0MemObjLockUser", (void *)RTR0MemObjLockUser },
157	{ "RTR0MemObjMapKernel", (void *)RTR0MemObjMapKernel },
158	{ "RTR0MemObjMapUser", (void *)RTR0MemObjMapUser },
159	{ "RTR0MemObjAddress", (void *)RTR0MemObjAddress },
160	{ "RTR0MemObjAddressR3", (void *)RTR0MemObjAddressR3 },
161	{ "RTR0MemObjSize", (void *)RTR0MemObjSize },
162	{ "RTR0MemObjIsMapping", (void *)RTR0MemObjIsMapping },
163	{ "RTR0MemObjGetPagePhysAddr", (void *)RTR0MemObjGetPagePhysAddr },
164	{ "RTR0MemObjFree", (void *)RTR0MemObjFree },
165	/* These don't work yet on linux - use fast mutexes!
166	{ "RTSemMutexCreate", (void *)RTSemMutexCreate },
167	{ "RTSemMutexRequest", (void *)RTSemMutexRequest },
168	{ "RTSemMutexRelease", (void *)RTSemMutexRelease },
169	{ "RTSemMutexDestroy", (void *)RTSemMutexDestroy },
170	*/
171	{ "RTProcSelf", (void *)RTProcSelf },
172	{ "RTR0ProcHandleSelf", (void *)RTR0ProcHandleSelf },
173	{ "RTSemFastMutexCreate", (void *)RTSemFastMutexCreate },
174	{ "RTSemFastMutexDestroy", (void *)RTSemFastMutexDestroy },
175	{ "RTSemFastMutexRequest", (void *)RTSemFastMutexRequest },
176	{ "RTSemFastMutexRelease", (void *)RTSemFastMutexRelease },
177	{ "RTSemEventCreate", (void *)RTSemEventCreate },
178	{ "RTSemEventSignal", (void *)RTSemEventSignal },
179	{ "RTSemEventWait", (void *)RTSemEventWait },
180	{ "RTSemEventWaitNoResume", (void *)RTSemEventWaitNoResume },
181	{ "RTSemEventDestroy", (void *)RTSemEventDestroy },
182	{ "RTSemEventMultiCreate", (void *)RTSemEventMultiCreate },
183	{ "RTSemEventMultiSignal", (void *)RTSemEventMultiSignal },
184	{ "RTSemEventMultiReset", (void *)RTSemEventMultiReset },
185	{ "RTSemEventMultiWait", (void *)RTSemEventMultiWait },
186	{ "RTSemEventMultiWaitNoResume", (void *)RTSemEventMultiWaitNoResume },
187	{ "RTSemEventMultiDestroy", (void *)RTSemEventMultiDestroy },
188	{ "RTSpinlockCreate", (void *)RTSpinlockCreate },
189	{ "RTSpinlockDestroy", (void *)RTSpinlockDestroy },
190	{ "RTSpinlockAcquire", (void *)RTSpinlockAcquire },
191	{ "RTSpinlockRelease", (void *)RTSpinlockRelease },
192	{ "RTSpinlockAcquireNoInts", (void *)RTSpinlockAcquireNoInts },
193	{ "RTSpinlockReleaseNoInts", (void *)RTSpinlockReleaseNoInts },
194	{ "RTTimeNanoTS", (void *)RTTimeNanoTS },
195	{ "RTTimeMillieTS", (void *)RTTimeMilliTS },
196	{ "RTTimeSystemNanoTS", (void *)RTTimeSystemNanoTS },
197	{ "RTTimeSystemMillieTS", (void *)RTTimeSystemMilliTS },
198	{ "RTThreadNativeSelf", (void *)RTThreadNativeSelf },
199	{ "RTThreadSleep", (void *)RTThreadSleep },
200	{ "RTThreadYield", (void *)RTThreadYield },
201	#if 0 /* Thread APIs, Part 2. */
202	{ "RTThreadSelf", (void *)RTThreadSelf },
203	{ "RTThreadCreate", (void *)RTThreadCreate },
204	{ "RTThreadGetNative", (void *)RTThreadGetNative },
205	{ "RTThreadWait", (void *)RTThreadWait },
206	{ "RTThreadWaitNoResume", (void *)RTThreadWaitNoResume },
207	{ "RTThreadGetName", (void *)RTThreadGetName },
208	{ "RTThreadSelfName", (void *)RTThreadSelfName },
209	{ "RTThreadGetType", (void *)RTThreadGetType },
210	{ "RTThreadUserSignal", (void *)RTThreadUserSignal },
211	{ "RTThreadUserReset", (void *)RTThreadUserReset },
212	{ "RTThreadUserWait", (void *)RTThreadUserWait },
213	{ "RTThreadUserWaitNoResume", (void *)RTThreadUserWaitNoResume },
214	#endif
215	{ "RTLogDefaultInstance", (void *)RTLogDefaultInstance },
216	{ "RTMpCpuId", (void *)RTMpCpuId },
217	{ "RTMpCpuIdFromSetIndex", (void *)RTMpCpuIdFromSetIndex },
218	{ "RTMpCpuIdToSetIndex", (void *)RTMpCpuIdToSetIndex },
219	{ "RTMpIsCpuPossible", (void *)RTMpIsCpuPossible },
220	{ "RTMpGetCount", (void *)RTMpGetCount },
221	{ "RTMpGetMaxCpuId", (void *)RTMpGetMaxCpuId },
222	{ "RTMpGetOnlineCount", (void *)RTMpGetOnlineCount },
223	{ "RTMpGetOnlineSet", (void *)RTMpGetOnlineSet },
224	{ "RTMpGetSet", (void *)RTMpGetSet },
225	{ "RTMpIsCpuOnline", (void *)RTMpIsCpuOnline },
226	{ "RTMpOnAll", (void *)RTMpOnAll },
227	{ "RTMpOnOthers", (void *)RTMpOnOthers },
228	{ "RTMpOnSpecific", (void *)RTMpOnSpecific },
229	{ "RTLogRelDefaultInstance", (void *)RTLogRelDefaultInstance },
230	{ "RTLogSetDefaultInstanceThread", (void *)RTLogSetDefaultInstanceThread },
231	{ "RTLogLogger", (void *)RTLogLogger },
232	{ "RTLogLoggerEx", (void *)RTLogLoggerEx },
233	{ "RTLogLoggerExV", (void *)RTLogLoggerExV },
234	{ "RTLogPrintf", (void *)RTLogPrintf },
235	{ "RTLogPrintfV", (void *)RTLogPrintfV },
236	{ "AssertMsg1", (void *)AssertMsg1 },
237	{ "AssertMsg2", (void *)AssertMsg2 },
238	};
239
240
241	/*******************************************************************************
242	* Internal Functions *
243	*******************************************************************************/
244	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession);
245	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType);
246	#ifdef VBOX_WITH_IDT_PATCHING
247	static int supdrvIOCtl_IdtInstall(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPIDTINSTALL pReq);
248	static PSUPDRVPATCH supdrvIdtPatchOne(PSUPDRVDEVEXT pDevExt, PSUPDRVPATCH pPatch);
249	static int supdrvIOCtl_IdtRemoveAll(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession);
250	static void supdrvIdtRemoveOne(PSUPDRVDEVEXT pDevExt, PSUPDRVPATCH pPatch);
251	static void supdrvIdtWrite(volatile void pvIdtEntry, const SUPDRVIDTE pNewIDTEntry);
252	#endif /* VBOX_WITH_IDT_PATCHING */
253	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq);
254	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq);
255	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq);
256	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq);
257	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq);
258	static int supdrvLdrSetR0EP(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx);
259	static void supdrvLdrUnsetR0EP(PSUPDRVDEVEXT pDevExt);
260	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage);
261	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage);
262	static SUPPAGINGMODE supdrvIOCtl_GetPagingMode(void);
263	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt);
264	#ifdef RT_OS_WINDOWS
265	static int supdrvPageGetPhys(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages);
266	static bool supdrvPageWasLockedByPageAlloc(PSUPDRVSESSION pSession, RTR3PTR pvR3);
267	#endif /* RT_OS_WINDOWS */
268	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt);
269	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt);
270	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
271	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
272	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser);
273
274
275	/**
276	* Initializes the device extentsion structure.
277	*
278	* @returns IPRT status code.
279	* @param pDevExt The device extension to initialize.
280	*/
281	int VBOXCALL supdrvInitDevExt(PSUPDRVDEVEXT pDevExt)
282	{
283	int rc;
284
285	#ifdef SUPDRV_WITH_RELEASE_LOGGER
286	/*
287	* Create the release log.
288	*/
289	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
290	PRTLOGGER pRelLogger;
291	rc = RTLogCreate(&pRelLogger, 0 /* fFlags */, "all",
292	"VBOX_RELEASE_LOG", RT_ELEMENTS(s_apszGroups), s_apszGroups,
293	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER, NULL);
294	if (RT_SUCCESS(rc))
295	RTLogRelSetDefaultInstance(pRelLogger);
296	#endif
297
298	/*
299	* Initialize it.
300	*/
301	memset(pDevExt, 0, sizeof(*pDevExt));
302	rc = RTSpinlockCreate(&pDevExt->Spinlock);
303	if (!rc)
304	{
305	rc = RTSemFastMutexCreate(&pDevExt->mtxLdr);
306	if (!rc)
307	{
308	rc = RTSemFastMutexCreate(&pDevExt->mtxComponentFactory);
309	if (!rc)
310	{
311	rc = RTSemFastMutexCreate(&pDevExt->mtxGip);
312	if (!rc)
313	{
314	rc = supdrvGipCreate(pDevExt);
315	if (RT_SUCCESS(rc))
316	{
317	pDevExt->u32Cookie = BIRD; /** @todo make this random? */
318	return VINF_SUCCESS;
319	}
320
321	RTSemFastMutexDestroy(pDevExt->mtxGip);
322	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
323	}
324	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
325	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
326	}
327	RTSemFastMutexDestroy(pDevExt->mtxLdr);
328	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
329	}
330	RTSpinlockDestroy(pDevExt->Spinlock);
331	pDevExt->Spinlock = NIL_RTSPINLOCK;
332	}
333	#ifdef SUPDRV_WITH_RELEASE_LOGGER
334	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
335	RTLogDestroy(RTLogSetDefaultInstance(NULL));
336	#endif
337
338	return rc;
339	}
340
341
342	/**
343	* Delete the device extension (e.g. cleanup members).
344	*
345	* @param pDevExt The device extension to delete.
346	*/
347	void VBOXCALL supdrvDeleteDevExt(PSUPDRVDEVEXT pDevExt)
348	{
349	#ifdef VBOX_WITH_IDT_PATCHING
350	PSUPDRVPATCH pPatch;
351	#endif
352	PSUPDRVOBJ pObj;
353	PSUPDRVUSAGE pUsage;
354
355	/*
356	* Kill mutexes and spinlocks.
357	*/
358	RTSemFastMutexDestroy(pDevExt->mtxGip);
359	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
360	RTSemFastMutexDestroy(pDevExt->mtxLdr);
361	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
362	RTSpinlockDestroy(pDevExt->Spinlock);
363	pDevExt->Spinlock = NIL_RTSPINLOCK;
364
365	/*
366	* Free lists.
367	*/
368	#ifdef VBOX_WITH_IDT_PATCHING
369	/* patches */
370	/** @todo make sure we don't uninstall patches which has been patched by someone else. */
371	pPatch = pDevExt->pIdtPatchesFree;
372	pDevExt->pIdtPatchesFree = NULL;
373	while (pPatch)
374	{
375	void *pvFree = pPatch;
376	pPatch = pPatch->pNext;
377	RTMemExecFree(pvFree);
378	}
379	#endif /* VBOX_WITH_IDT_PATCHING */
380
381	/* objects. */
382	pObj = pDevExt->pObjs;
383	#if !defined(DEBUG_bird) \|\| !defined(RT_OS_LINUX) /* breaks unloading, temporary, remove me! */
384	Assert(!pObj); /* (can trigger on forced unloads) */
385	#endif
386	pDevExt->pObjs = NULL;
387	while (pObj)
388	{
389	void *pvFree = pObj;
390	pObj = pObj->pNext;
391	RTMemFree(pvFree);
392	}
393
394	/* usage records. */
395	pUsage = pDevExt->pUsageFree;
396	pDevExt->pUsageFree = NULL;
397	while (pUsage)
398	{
399	void *pvFree = pUsage;
400	pUsage = pUsage->pNext;
401	RTMemFree(pvFree);
402	}
403
404	/* kill the GIP. */
405	supdrvGipDestroy(pDevExt);
406
407	#ifdef SUPDRV_WITH_RELEASE_LOGGER
408	/* destroy the loggers. */
409	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
410	RTLogDestroy(RTLogSetDefaultInstance(NULL));
411	#endif
412	}
413
414
415	/**
416	* Create session.
417	*
418	* @returns IPRT status code.
419	* @param pDevExt Device extension.
420	* @param fUser Flag indicating whether this is a user or kernel session.
421	* @param ppSession Where to store the pointer to the session data.
422	*/
423	int VBOXCALL supdrvCreateSession(PSUPDRVDEVEXT pDevExt, bool fUser, PSUPDRVSESSION *ppSession)
424	{
425	/*
426	* Allocate memory for the session data.
427	*/
428	int rc = VERR_NO_MEMORY;
429	PSUPDRVSESSION pSession = ppSession = (PSUPDRVSESSION)RTMemAllocZ(sizeof(pSession));
430	if (pSession)
431	{
432	/* Initialize session data. */
433	rc = RTSpinlockCreate(&pSession->Spinlock);
434	if (!rc)
435	{
436	Assert(pSession->Spinlock != NIL_RTSPINLOCK);
437	pSession->pDevExt = pDevExt;
438	pSession->u32Cookie = BIRD_INV;
439	/*pSession->pLdrUsage = NULL;
440	pSession->pPatchUsage = NULL;
441	pSession->pVM = NULL;
442	pSession->pUsage = NULL;
443	pSession->pGip = NULL;
444	pSession->fGipReferenced = false;
445	pSession->Bundle.cUsed = 0; */
446	pSession->Uid = NIL_RTUID;
447	pSession->Gid = NIL_RTGID;
448	if (fUser)
449	{
450	pSession->Process = RTProcSelf();
451	pSession->R0Process = RTR0ProcHandleSelf();
452	}
453	else
454	{
455	pSession->Process = NIL_RTPROCESS;
456	pSession->R0Process = NIL_RTR0PROCESS;
457	}
458
459	LogFlow(("Created session %p initial cookie=%#x\n", pSession, pSession->u32Cookie));
460	return VINF_SUCCESS;
461	}
462
463	RTMemFree(pSession);
464	*ppSession = NULL;
465	Log(("Failed to create spinlock, rc=%d!\n", rc));
466	}
467
468	return rc;
469	}
470
471
472	/**
473	* Shared code for cleaning up a session.
474	*
475	* @param pDevExt Device extension.
476	* @param pSession Session data.
477	* This data will be freed by this routine.
478	*/
479	void VBOXCALL supdrvCloseSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
480	{
481	/*
482	* Cleanup the session first.
483	*/
484	supdrvCleanupSession(pDevExt, pSession);
485
486	/*
487	* Free the rest of the session stuff.
488	*/
489	RTSpinlockDestroy(pSession->Spinlock);
490	pSession->Spinlock = NIL_RTSPINLOCK;
491	pSession->pDevExt = NULL;
492	RTMemFree(pSession);
493	LogFlow(("supdrvCloseSession: returns\n"));
494	}
495
496
497	/**
498	* Shared code for cleaning up a session (but not quite freeing it).
499	*
500	* This is primarily intended for MAC OS X where we have to clean up the memory
501	* stuff before the file handle is closed.
502	*
503	* @param pDevExt Device extension.
504	* @param pSession Session data.
505	* This data will be freed by this routine.
506	*/
507	void VBOXCALL supdrvCleanupSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
508	{
509	PSUPDRVBUNDLE pBundle;
510	LogFlow(("supdrvCleanupSession: pSession=%p\n", pSession));
511
512	/*
513	* Remove logger instances related to this session.
514	*/
515	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pSession);
516
517	#ifdef VBOX_WITH_IDT_PATCHING
518	/*
519	* Uninstall any IDT patches installed for this session.
520	*/
521	supdrvIOCtl_IdtRemoveAll(pDevExt, pSession);
522	#endif
523
524	/*
525	* Release object references made in this session.
526	* In theory there should be noone racing us in this session.
527	*/
528	Log2(("release objects - start\n"));
529	if (pSession->pUsage)
530	{
531	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
532	PSUPDRVUSAGE pUsage;
533	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
534
535	while ((pUsage = pSession->pUsage) != NULL)
536	{
537	PSUPDRVOBJ pObj = pUsage->pObj;
538	pSession->pUsage = pUsage->pNext;
539
540	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
541	if (pUsage->cUsage < pObj->cUsage)
542	{
543	pObj->cUsage -= pUsage->cUsage;
544	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
545	}
546	else
547	{
548	/* Destroy the object and free the record. */
549	if (pDevExt->pObjs == pObj)
550	pDevExt->pObjs = pObj->pNext;
551	else
552	{
553	PSUPDRVOBJ pObjPrev;
554	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
555	if (pObjPrev->pNext == pObj)
556	{
557	pObjPrev->pNext = pObj->pNext;
558	break;
559	}
560	Assert(pObjPrev);
561	}
562	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
563
564	Log(("supdrvCleanupSession: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
565	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
566	if (pObj->pfnDestructor)
567	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
568	RTMemFree(pObj);
569	}
570
571	/* free it and continue. */
572	RTMemFree(pUsage);
573
574	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
575	}
576
577	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
578	AssertMsg(!pSession->pUsage, ("Some buster reregistered an object during desturction!\n"));
579	}
580	Log2(("release objects - done\n"));
581
582	/*
583	* Release memory allocated in the session.
584	*
585	* We do not serialize this as we assume that the application will
586	* not allocated memory while closing the file handle object.
587	*/
588	Log2(("freeing memory:\n"));
589	pBundle = &pSession->Bundle;
590	while (pBundle)
591	{
592	PSUPDRVBUNDLE pToFree;
593	unsigned i;
594
595	/*
596	* Check and unlock all entries in the bundle.
597	*/
598	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
599	{
600	if (pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ)
601	{
602	int rc;
603	Log2(("eType=%d pvR0=%p pvR3=%p cb=%ld\n", pBundle->aMem[i].eType, RTR0MemObjAddress(pBundle->aMem[i].MemObj),
604	(void *)RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3), (long)RTR0MemObjSize(pBundle->aMem[i].MemObj)));
605	if (pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ)
606	{
607	rc = RTR0MemObjFree(pBundle->aMem[i].MapObjR3, false);
608	AssertRC(rc); /** @todo figure out how to handle this. */
609	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
610	}
611	rc = RTR0MemObjFree(pBundle->aMem[i].MemObj, false);
612	AssertRC(rc); /** @todo figure out how to handle this. */
613	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
614	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
615	}
616	}
617
618	/*
619	* Advance and free previous bundle.
620	*/
621	pToFree = pBundle;
622	pBundle = pBundle->pNext;
623
624	pToFree->pNext = NULL;
625	pToFree->cUsed = 0;
626	if (pToFree != &pSession->Bundle)
627	RTMemFree(pToFree);
628	}
629	Log2(("freeing memory - done\n"));
630
631	/*
632	* Deregister component factories.
633	*/
634	RTSemFastMutexRequest(pDevExt->mtxComponentFactory);
635	Log2(("deregistering component factories:\n"));
636	if (pDevExt->pComponentFactoryHead)
637	{
638	PSUPDRVFACTORYREG pPrev = NULL;
639	PSUPDRVFACTORYREG pCur = pDevExt->pComponentFactoryHead;
640	while (pCur)
641	{
642	if (pCur->pSession == pSession)
643	{
644	/* unlink it */
645	PSUPDRVFACTORYREG pNext = pCur->pNext;
646	if (pPrev)
647	pPrev->pNext = pNext;
648	else
649	pDevExt->pComponentFactoryHead = pNext;
650
651	/* free it */
652	pCur->pNext = NULL;
653	pCur->pSession = NULL;
654	pCur->pFactory = NULL;
655	RTMemFree(pCur);
656
657	/* next */
658	pCur = pNext;
659	}
660	else
661	{
662	/* next */
663	pPrev = pCur;
664	pCur = pCur->pNext;
665	}
666	}
667	}
668	RTSemFastMutexRelease(pDevExt->mtxComponentFactory);
669	Log2(("deregistering component factories - done\n"));
670
671	/*
672	* Loaded images needs to be dereferenced and possibly freed up.
673	*/
674	RTSemFastMutexRequest(pDevExt->mtxLdr);
675	Log2(("freeing images:\n"));
676	if (pSession->pLdrUsage)
677	{
678	PSUPDRVLDRUSAGE pUsage = pSession->pLdrUsage;
679	pSession->pLdrUsage = NULL;
680	while (pUsage)
681	{
682	void *pvFree = pUsage;
683	PSUPDRVLDRIMAGE pImage = pUsage->pImage;
684	if (pImage->cUsage > pUsage->cUsage)
685	pImage->cUsage -= pUsage->cUsage;
686	else
687	supdrvLdrFree(pDevExt, pImage);
688	pUsage->pImage = NULL;
689	pUsage = pUsage->pNext;
690	RTMemFree(pvFree);
691	}
692	}
693	RTSemFastMutexRelease(pDevExt->mtxLdr);
694	Log2(("freeing images - done\n"));
695
696	/*
697	* Unmap the GIP.
698	*/
699	Log2(("umapping GIP:\n"));
700	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
701	{
702	SUPR0GipUnmap(pSession);
703	pSession->fGipReferenced = 0;
704	}
705	Log2(("umapping GIP - done\n"));
706	}
707
708
709	/**
710	* Fast path I/O Control worker.
711	*
712	* @returns VBox status code that should be passed down to ring-3 unchanged.
713	* @param uIOCtl Function number.
714	* @param pDevExt Device extention.
715	* @param pSession Session data.
716	*/
717	int VBOXCALL supdrvIOCtlFast(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
718	{
719	int rc;
720
721	/*
722	* We check the two prereqs after doing this only to allow the compiler to optimize things better.
723	*/
724	if (RT_LIKELY(pSession->pVM && pDevExt->pfnVMMR0EntryFast))
725	{
726	switch (uIOCtl)
727	{
728	case SUP_IOCTL_FAST_DO_RAW_RUN:
729	rc = pDevExt->pfnVMMR0EntryFast(pSession->pVM, SUP_VMMR0_DO_RAW_RUN);
730	break;
731	case SUP_IOCTL_FAST_DO_HWACC_RUN:
732	rc = pDevExt->pfnVMMR0EntryFast(pSession->pVM, SUP_VMMR0_DO_HWACC_RUN);
733	break;
734	case SUP_IOCTL_FAST_DO_NOP:
735	rc = pDevExt->pfnVMMR0EntryFast(pSession->pVM, SUP_VMMR0_DO_NOP);
736	break;
737	default:
738	rc = VERR_INTERNAL_ERROR;
739	break;
740	}
741	}
742	else
743	rc = VERR_INTERNAL_ERROR;
744
745	return rc;
746	}
747
748
749	/**
750	* Helper for supdrvIOCtl. Check if pszStr contains any character of pszChars.
751	* We would use strpbrk here if this function would be contained in the RedHat kABI white
752	* list, see http://www.kerneldrivers.org/RHEL5.
753	*
754	* @return 1 if pszStr does contain any character of pszChars, 0 otherwise.
755	* @param pszStr String to check
756	* @param pszChars Character set
757	*/
758	static int supdrvCheckInvalidChar(const char pszStr, const char pszChars)
759	{
760	int chCur;
761	while ((chCur = *pszStr++) != '\0')
762	{
763	int ch;
764	const char *psz = pszChars;
765	while ((ch = *psz++) != '\0')
766	if (ch == chCur)
767	return 1;
768
769	}
770	return 0;
771	}
772
773
774	/**
775	* I/O Control worker.
776	*
777	* @returns 0 on success.
778	* @returns VERR_INVALID_PARAMETER if the request is invalid.
779	*
780	* @param uIOCtl Function number.
781	* @param pDevExt Device extention.
782	* @param pSession Session data.
783	* @param pReqHdr The request header.
784	*/
785	int VBOXCALL supdrvIOCtl(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
786	{
787	/*
788	* Validate the request.
789	*/
790	/* this first check could probably be omitted as its also done by the OS specific code... */
791	if (RT_UNLIKELY( (pReqHdr->fFlags & SUPREQHDR_FLAGS_MAGIC_MASK) != SUPREQHDR_FLAGS_MAGIC
792	\|\| pReqHdr->cbIn < sizeof(*pReqHdr)
793	\|\| pReqHdr->cbOut < sizeof(*pReqHdr)))
794	{
795	OSDBGPRINT(("vboxdrv: Bad ioctl request header; cbIn=%#lx cbOut=%#lx fFlags=%#lx\n",
796	(long)pReqHdr->cbIn, (long)pReqHdr->cbOut, (long)pReqHdr->fFlags));
797	return VERR_INVALID_PARAMETER;
798	}
799	if (RT_UNLIKELY(uIOCtl == SUP_IOCTL_COOKIE))
800	{
801	if (pReqHdr->u32Cookie != SUPCOOKIE_INITIAL_COOKIE)
802	{
803	OSDBGPRINT(("SUP_IOCTL_COOKIE: bad cookie %#lx\n", (long)pReqHdr->u32Cookie));
804	return VERR_INVALID_PARAMETER;
805	}
806	}
807	else if (RT_UNLIKELY( pReqHdr->u32Cookie != pDevExt->u32Cookie
808	\|\| pReqHdr->u32SessionCookie != pSession->u32Cookie))
809	{
810	OSDBGPRINT(("vboxdrv: bad cookie %#lx / %#lx.\n", (long)pReqHdr->u32Cookie, (long)pReqHdr->u32SessionCookie));
811	return VERR_INVALID_PARAMETER;
812	}
813
814	/*
815	* Validation macros
816	*/
817	#define REQ_CHECK_SIZES_EX(Name, cbInExpect, cbOutExpect) \
818	do { \
819	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect) \|\| pReqHdr->cbOut != (cbOutExpect))) \
820	{ \
821	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld. cbOut=%ld expected %ld.\n", \
822	(long)pReq->Hdr.cbIn, (long)(cbInExpect), (long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
823	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
824	} \
825	} while (0)
826
827	#define REQ_CHECK_SIZES(Name) REQ_CHECK_SIZES_EX(Name, Name ## _SIZE_IN, Name ## _SIZE_OUT)
828
829	#define REQ_CHECK_SIZE_IN(Name, cbInExpect) \
830	do { \
831	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect))) \
832	{ \
833	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld.\n", \
834	(long)pReq->Hdr.cbIn, (long)(cbInExpect))); \
835	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
836	} \
837	} while (0)
838
839	#define REQ_CHECK_SIZE_OUT(Name, cbOutExpect) \
840	do { \
841	if (RT_UNLIKELY(pReqHdr->cbOut != (cbOutExpect))) \
842	{ \
843	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbOut=%ld expected %ld.\n", \
844	(long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
845	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
846	} \
847	} while (0)
848
849	#define REQ_CHECK_EXPR(Name, expr) \
850	do { \
851	if (RT_UNLIKELY(!(expr))) \
852	{ \
853	OSDBGPRINT(( #Name ": %s\n", #expr)); \
854	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
855	} \
856	} while (0)
857
858	#define REQ_CHECK_EXPR_FMT(expr, fmt) \
859	do { \
860	if (RT_UNLIKELY(!(expr))) \
861	{ \
862	OSDBGPRINT( fmt ); \
863	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
864	} \
865	} while (0)
866
867
868	/*
869	* The switch.
870	*/
871	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
872	{
873	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
874	{
875	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
876	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
877	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
878	{
879	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
880	pReq->Hdr.rc = VERR_INVALID_MAGIC;
881	return 0;
882	}
883
884	#if 0
885	/*
886	* Call out to the OS specific code and let it do permission checks on the
887	* client process.
888	*/
889	if (!supdrvOSValidateClientProcess(pDevExt, pSession))
890	{
891	pReq->u.Out.u32Cookie = 0xffffffff;
892	pReq->u.Out.u32SessionCookie = 0xffffffff;
893	pReq->u.Out.u32SessionVersion = 0xffffffff;
894	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
895	pReq->u.Out.pSession = NULL;
896	pReq->u.Out.cFunctions = 0;
897	pReq->Hdr.rc = VERR_PERMISSION_DENIED;
898	return 0;
899	}
900	#endif
901
902	/*
903	* Match the version.
904	* The current logic is very simple, match the major interface version.
905	*/
906	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
907	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
908	{
909	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
910	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
911	pReq->u.Out.u32Cookie = 0xffffffff;
912	pReq->u.Out.u32SessionCookie = 0xffffffff;
913	pReq->u.Out.u32SessionVersion = 0xffffffff;
914	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
915	pReq->u.Out.pSession = NULL;
916	pReq->u.Out.cFunctions = 0;
917	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
918	return 0;
919	}
920
921	/*
922	* Fill in return data and be gone.
923	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
924	* u32SessionVersion <= u32ReqVersion!
925	*/
926	/** @todo Somehow validate the client and negotiate a secure cookie... */
927	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
928	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
929	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
930	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
931	pReq->u.Out.pSession = pSession;
932	pReq->u.Out.cFunctions = sizeof(g_aFunctions) / sizeof(g_aFunctions[0]);
933	pReq->Hdr.rc = VINF_SUCCESS;
934	return 0;
935	}
936
937	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_QUERY_FUNCS(0)):
938	{
939	/* validate */
940	PSUPQUERYFUNCS pReq = (PSUPQUERYFUNCS)pReqHdr;
941	REQ_CHECK_SIZES_EX(SUP_IOCTL_QUERY_FUNCS, SUP_IOCTL_QUERY_FUNCS_SIZE_IN, SUP_IOCTL_QUERY_FUNCS_SIZE_OUT(RT_ELEMENTS(g_aFunctions)));
942
943	/* execute */
944	pReq->u.Out.cFunctions = RT_ELEMENTS(g_aFunctions);
945	memcpy(&pReq->u.Out.aFunctions[0], g_aFunctions, sizeof(g_aFunctions));
946	pReq->Hdr.rc = VINF_SUCCESS;
947	return 0;
948	}
949
950	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_IDT_INSTALL):
951	{
952	/* validate */
953	PSUPIDTINSTALL pReq = (PSUPIDTINSTALL)pReqHdr;
954	REQ_CHECK_SIZES(SUP_IOCTL_IDT_INSTALL);
955
956	/* execute */
957	#ifdef VBOX_WITH_IDT_PATCHING
958	pReq->Hdr.rc = supdrvIOCtl_IdtInstall(pDevExt, pSession, pReq);
959	#else
960	pReq->u.Out.u8Idt = 3;
961	pReq->Hdr.rc = VERR_NOT_SUPPORTED;
962	#endif
963	return 0;
964	}
965
966	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_IDT_REMOVE):
967	{
968	/* validate */
969	PSUPIDTREMOVE pReq = (PSUPIDTREMOVE)pReqHdr;
970	REQ_CHECK_SIZES(SUP_IOCTL_IDT_REMOVE);
971
972	/* execute */
973	#ifdef VBOX_WITH_IDT_PATCHING
974	pReq->Hdr.rc = supdrvIOCtl_IdtRemoveAll(pDevExt, pSession);
975	#else
976	pReq->Hdr.rc = VERR_NOT_SUPPORTED;
977	#endif
978	return 0;
979	}
980
981	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_LOCK):
982	{
983	/* validate */
984	PSUPPAGELOCK pReq = (PSUPPAGELOCK)pReqHdr;
985	REQ_CHECK_SIZE_IN(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_IN);
986	REQ_CHECK_SIZE_OUT(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_OUT(pReq->u.In.cPages));
987	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.cPages > 0);
988	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.pvR3 >= PAGE_SIZE);
989
990	/* execute */
991	pReq->Hdr.rc = SUPR0LockMem(pSession, pReq->u.In.pvR3, pReq->u.In.cPages, &pReq->u.Out.aPages[0]);
992	if (RT_FAILURE(pReq->Hdr.rc))
993	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
994	return 0;
995	}
996
997	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_UNLOCK):
998	{
999	/* validate */
1000	PSUPPAGEUNLOCK pReq = (PSUPPAGEUNLOCK)pReqHdr;
1001	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_UNLOCK);
1002
1003	/* execute */
1004	pReq->Hdr.rc = SUPR0UnlockMem(pSession, pReq->u.In.pvR3);
1005	return 0;
1006	}
1007
1008	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_ALLOC):
1009	{
1010	/* validate */
1011	PSUPCONTALLOC pReq = (PSUPCONTALLOC)pReqHdr;
1012	REQ_CHECK_SIZES(SUP_IOCTL_CONT_ALLOC);
1013
1014	/* execute */
1015	pReq->Hdr.rc = SUPR0ContAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.HCPhys);
1016	if (RT_FAILURE(pReq->Hdr.rc))
1017	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1018	return 0;
1019	}
1020
1021	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_FREE):
1022	{
1023	/* validate */
1024	PSUPCONTFREE pReq = (PSUPCONTFREE)pReqHdr;
1025	REQ_CHECK_SIZES(SUP_IOCTL_CONT_FREE);
1026
1027	/* execute */
1028	pReq->Hdr.rc = SUPR0ContFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1029	return 0;
1030	}
1031
1032	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_OPEN):
1033	{
1034	/* validate */
1035	PSUPLDROPEN pReq = (PSUPLDROPEN)pReqHdr;
1036	REQ_CHECK_SIZES(SUP_IOCTL_LDR_OPEN);
1037	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImage > 0);
1038	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImage < _1M*16);
1039	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.szName[0]);
1040	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, memchr(pReq->u.In.szName, '\0', sizeof(pReq->u.In.szName)));
1041	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, !supdrvCheckInvalidChar(pReq->u.In.szName, ";:()[]{}/\\\|&*%#@!~`\"'"));
1042
1043	/* execute */
1044	pReq->Hdr.rc = supdrvIOCtl_LdrOpen(pDevExt, pSession, pReq);
1045	return 0;
1046	}
1047
1048	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOAD):
1049	{
1050	/* validate */
1051	PSUPLDRLOAD pReq = (PSUPLDRLOAD)pReqHdr;
1052	REQ_CHECK_EXPR(Name, pReq->Hdr.cbIn >= sizeof(*pReq));
1053	REQ_CHECK_SIZES_EX(SUP_IOCTL_LDR_LOAD, SUP_IOCTL_LDR_LOAD_SIZE_IN(pReq->u.In.cbImage), SUP_IOCTL_LDR_LOAD_SIZE_OUT);
1054	REQ_CHECK_EXPR(SUP_IOCTL_LDR_LOAD, pReq->u.In.cSymbols <= 16384);
1055	REQ_CHECK_EXPR_FMT( !pReq->u.In.cSymbols
1056	\|\| ( pReq->u.In.offSymbols < pReq->u.In.cbImage
1057	&& pReq->u.In.offSymbols + pReq->u.In.cSymbols * sizeof(SUPLDRSYM) <= pReq->u.In.cbImage),
1058	("SUP_IOCTL_LDR_LOAD: offSymbols=%#lx cSymbols=%#lx cbImage=%#lx\n", (long)pReq->u.In.offSymbols,
1059	(long)pReq->u.In.cSymbols, (long)pReq->u.In.cbImage));
1060	REQ_CHECK_EXPR_FMT( !pReq->u.In.cbStrTab
1061	\|\| ( pReq->u.In.offStrTab < pReq->u.In.cbImage
1062	&& pReq->u.In.offStrTab + pReq->u.In.cbStrTab <= pReq->u.In.cbImage
1063	&& pReq->u.In.cbStrTab <= pReq->u.In.cbImage),
1064	("SUP_IOCTL_LDR_LOAD: offStrTab=%#lx cbStrTab=%#lx cbImage=%#lx\n", (long)pReq->u.In.offStrTab,
1065	(long)pReq->u.In.cbStrTab, (long)pReq->u.In.cbImage));
1066
1067	if (pReq->u.In.cSymbols)
1068	{
1069	uint32_t i;
1070	PSUPLDRSYM paSyms = (PSUPLDRSYM)&pReq->u.In.achImage[pReq->u.In.offSymbols];
1071	for (i = 0; i < pReq->u.In.cSymbols; i++)
1072	{
1073	REQ_CHECK_EXPR_FMT(paSyms[i].offSymbol < pReq->u.In.cbImage,
1074	("SUP_IOCTL_LDR_LOAD: sym #%ld: symb off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offSymbol, (long)pReq->u.In.cbImage));
1075	REQ_CHECK_EXPR_FMT(paSyms[i].offName < pReq->u.In.cbStrTab,
1076	("SUP_IOCTL_LDR_LOAD: sym #%ld: name off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImage));
1077	REQ_CHECK_EXPR_FMT(memchr(&pReq->u.In.achImage[pReq->u.In.offStrTab + paSyms[i].offName], '\0', pReq->u.In.cbStrTab - paSyms[i].offName),
1078	("SUP_IOCTL_LDR_LOAD: sym #%ld: unterminated name! (%#lx / %#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImage));
1079	}
1080	}
1081
1082	/* execute */
1083	pReq->Hdr.rc = supdrvIOCtl_LdrLoad(pDevExt, pSession, pReq);
1084	return 0;
1085	}
1086
1087	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_FREE):
1088	{
1089	/* validate */
1090	PSUPLDRFREE pReq = (PSUPLDRFREE)pReqHdr;
1091	REQ_CHECK_SIZES(SUP_IOCTL_LDR_FREE);
1092
1093	/* execute */
1094	pReq->Hdr.rc = supdrvIOCtl_LdrFree(pDevExt, pSession, pReq);
1095	return 0;
1096	}
1097
1098	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_GET_SYMBOL):
1099	{
1100	/* validate */
1101	PSUPLDRGETSYMBOL pReq = (PSUPLDRGETSYMBOL)pReqHdr;
1102	REQ_CHECK_SIZES(SUP_IOCTL_LDR_GET_SYMBOL);
1103	REQ_CHECK_EXPR(SUP_IOCTL_LDR_GET_SYMBOL, memchr(pReq->u.In.szSymbol, '\0', sizeof(pReq->u.In.szSymbol)));
1104
1105	/* execute */
1106	pReq->Hdr.rc = supdrvIOCtl_LdrGetSymbol(pDevExt, pSession, pReq);
1107	return 0;
1108	}
1109
1110	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0(0)):
1111	{
1112	/* validate */
1113	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1114	Log4(("SUP_IOCTL_CALL_VMMR0: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1115	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1116
1117	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_VMMR0_SIZE(0))
1118	{
1119	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(0), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(0));
1120
1121	/* execute */
1122	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1123	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg);
1124	else
1125	pReq->Hdr.rc = VERR_WRONG_ORDER;
1126	}
1127	else
1128	{
1129	PSUPVMMR0REQHDR pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1130	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR)),
1131	("SUP_IOCTL_CALL_VMMR0: cbIn=%#x < %#x\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR))));
1132	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1133	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(pVMMReq->cbReq));
1134
1135	/* execute */
1136	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1137	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg);
1138	else
1139	pReq->Hdr.rc = VERR_WRONG_ORDER;
1140	}
1141
1142	if ( RT_FAILURE(pReq->Hdr.rc)
1143	&& pReq->Hdr.rc != VERR_INTERRUPTED
1144	&& pReq->Hdr.rc != VERR_TIMEOUT)
1145	Log(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1146	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1147	else
1148	Log4(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1149	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1150	return 0;
1151	}
1152
1153	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_PAGING_MODE):
1154	{
1155	/* validate */
1156	PSUPGETPAGINGMODE pReq = (PSUPGETPAGINGMODE)pReqHdr;
1157	REQ_CHECK_SIZES(SUP_IOCTL_GET_PAGING_MODE);
1158
1159	/* execute */
1160	pReq->Hdr.rc = VINF_SUCCESS;
1161	pReq->u.Out.enmMode = supdrvIOCtl_GetPagingMode();
1162	return 0;
1163	}
1164
1165	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_ALLOC):
1166	{
1167	/* validate */
1168	PSUPLOWALLOC pReq = (PSUPLOWALLOC)pReqHdr;
1169	REQ_CHECK_EXPR(SUP_IOCTL_LOW_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_LOW_ALLOC_SIZE_IN);
1170	REQ_CHECK_SIZES_EX(SUP_IOCTL_LOW_ALLOC, SUP_IOCTL_LOW_ALLOC_SIZE_IN, SUP_IOCTL_LOW_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1171
1172	/* execute */
1173	pReq->Hdr.rc = SUPR0LowAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1174	if (RT_FAILURE(pReq->Hdr.rc))
1175	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1176	return 0;
1177	}
1178
1179	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_FREE):
1180	{
1181	/* validate */
1182	PSUPLOWFREE pReq = (PSUPLOWFREE)pReqHdr;
1183	REQ_CHECK_SIZES(SUP_IOCTL_LOW_FREE);
1184
1185	/* execute */
1186	pReq->Hdr.rc = SUPR0LowFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1187	return 0;
1188	}
1189
1190	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_MAP):
1191	{
1192	/* validate */
1193	PSUPGIPMAP pReq = (PSUPGIPMAP)pReqHdr;
1194	REQ_CHECK_SIZES(SUP_IOCTL_GIP_MAP);
1195
1196	/* execute */
1197	pReq->Hdr.rc = SUPR0GipMap(pSession, &pReq->u.Out.pGipR3, &pReq->u.Out.HCPhysGip);
1198	if (RT_SUCCESS(pReq->Hdr.rc))
1199	pReq->u.Out.pGipR0 = pDevExt->pGip;
1200	return 0;
1201	}
1202
1203	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_UNMAP):
1204	{
1205	/* validate */
1206	PSUPGIPUNMAP pReq = (PSUPGIPUNMAP)pReqHdr;
1207	REQ_CHECK_SIZES(SUP_IOCTL_GIP_UNMAP);
1208
1209	/* execute */
1210	pReq->Hdr.rc = SUPR0GipUnmap(pSession);
1211	return 0;
1212	}
1213
1214	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SET_VM_FOR_FAST):
1215	{
1216	/* validate */
1217	PSUPSETVMFORFAST pReq = (PSUPSETVMFORFAST)pReqHdr;
1218	REQ_CHECK_SIZES(SUP_IOCTL_SET_VM_FOR_FAST);
1219	REQ_CHECK_EXPR_FMT( !pReq->u.In.pVMR0
1220	\|\| ( VALID_PTR(pReq->u.In.pVMR0)
1221	&& !((uintptr_t)pReq->u.In.pVMR0 & (PAGE_SIZE - 1))),
1222	("SUP_IOCTL_SET_VM_FOR_FAST: pVMR0=%p!\n", pReq->u.In.pVMR0));
1223	/* execute */
1224	pSession->pVM = pReq->u.In.pVMR0;
1225	pReq->Hdr.rc = VINF_SUCCESS;
1226	return 0;
1227	}
1228
1229	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_ALLOC):
1230	{
1231	/* validate */
1232	PSUPPAGEALLOC pReq = (PSUPPAGEALLOC)pReqHdr;
1233	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_SIZE_IN);
1234	REQ_CHECK_SIZES_EX(SUP_IOCTL_PAGE_ALLOC, SUP_IOCTL_PAGE_ALLOC_SIZE_IN, SUP_IOCTL_PAGE_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1235
1236	/* execute */
1237	pReq->Hdr.rc = SUPR0PageAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1238	if (RT_FAILURE(pReq->Hdr.rc))
1239	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1240	return 0;
1241	}
1242
1243	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_FREE):
1244	{
1245	/* validate */
1246	PSUPPAGEFREE pReq = (PSUPPAGEFREE)pReqHdr;
1247	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_FREE);
1248
1249	/* execute */
1250	pReq->Hdr.rc = SUPR0PageFree(pSession, pReq->u.In.pvR3);
1251	return 0;
1252	}
1253
1254	default:
1255	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1256	break;
1257	}
1258	return SUPDRV_ERR_GENERAL_FAILURE;
1259	}
1260
1261
1262	/**
1263	* Inter-Driver Communcation (IDC) worker.
1264	*
1265	* @returns VBox status code.
1266	* @retval VINF_SUCCESS on success.
1267	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1268	* @retval VERR_NOT_SUPPORTED if the request isn't supported.
1269	*
1270	* @param uReq The request (function) code.
1271	* @param pDevExt Device extention.
1272	* @param pSession Session data.
1273	* @param pReqHdr The request header.
1274	*/
1275	int VBOXCALL supdrvIDC(uintptr_t uReq, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQHDR pReqHdr)
1276	{
1277	/*
1278	* The OS specific code has already validated the pSession
1279	* pointer, and the request size being greater or equal to
1280	* size of the header.
1281	*
1282	* So, just check that pSession is a kernel context session.
1283	*/
1284	if (RT_UNLIKELY( pSession
1285	&& pSession->R0Process != NIL_RTR0PROCESS))
1286	return VERR_INVALID_PARAMETER;
1287
1288	/*
1289	* Validation macro.
1290	*/
1291	#define REQ_CHECK_IDC_SIZE(Name, cbExpect) \
1292	do { \
1293	if (RT_UNLIKELY(pReqHdr->cb != (cbExpect))) \
1294	{ \
1295	OSDBGPRINT(( #Name ": Invalid input/output sizes. cb=%ld expected %ld.\n", \
1296	(long)pReqHdr->cb, (long)(cbExpect))); \
1297	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1298	} \
1299	} while (0)
1300
1301	switch (uReq)
1302	{
1303	case SUPDRV_IDC_REQ_CONNECT:
1304	{
1305	PSUPDRVIDCREQCONNECT pReq = (PSUPDRVIDCREQCONNECT)pReqHdr;
1306	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_CONNECT, sizeof(*pReq));
1307
1308	/*
1309	* Validate the cookie and other input.
1310	*/
1311	if (pReq->Hdr.pSession != NULL)
1312	{
1313	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: pSession=%p expected NULL!\n", pReq->Hdr.pSession));
1314	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1315	}
1316	if (pReq->u.In.u32MagicCookie != SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE)
1317	{
1318	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: u32MagicCookie=%#x expected %#x!\n",
1319	pReq->u.In.u32MagicCookie, SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE));
1320	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1321	}
1322	if ( pReq->u.In.uMinVersion > pReq->u.In.uReqVersion
1323	\|\| (pReq->u.In.uMinVersion & UINT32_C(0xffff0000)) != (pReq->u.In.uReqVersion & UINT32_C(0xffff0000)))
1324	{
1325	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: uMinVersion=%#x uMaxVersion=%#x doesn't match!\n",
1326	pReq->u.In.uMinVersion, pReq->u.In.uReqVersion));
1327	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1328	}
1329
1330	/*
1331	* Match the version.
1332	* The current logic is very simple, match the major interface version.
1333	*/
1334	if ( pReq->u.In.uMinVersion > SUPDRV_IDC_VERSION
1335	\|\| (pReq->u.In.uMinVersion & 0xffff0000) != (SUPDRV_IDC_VERSION & 0xffff0000))
1336	{
1337	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1338	pReq->u.In.uReqVersion, pReq->u.In.uMinVersion, SUPDRV_IDC_VERSION));
1339	pReq->u.Out.pSession = NULL;
1340	pReq->u.Out.uSessionVersion = 0xffffffff;
1341	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1342	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1343	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1344	return VINF_SUCCESS;
1345	}
1346
1347	pReq->u.Out.pSession = NULL;
1348	pReq->u.Out.uSessionVersion = SUPDRV_IDC_VERSION;
1349	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1350	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1351
1352	/*
1353	* On NT we will already have a session associated with the
1354	* client, just like with the SUP_IOCTL_COOKIE request, while
1355	* the other doesn't.
1356	*/
1357	#ifdef RT_OS_WINDOWS
1358	pReq->Hdr.rc = VINF_SUCCESS;
1359	#else
1360	AssertReturn(!pSession, VERR_INTERNAL_ERROR);
1361	pReq->Hdr.rc = supdrvCreateSession(pDevExt, false /* fUser */, &pSession);
1362	if (RT_FAILURE(pReq->Hdr.rc))
1363	{
1364	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: failed to create session, rc=%d\n", pReq->Hdr.rc));
1365	return VINF_SUCCESS;
1366	}
1367	#endif
1368
1369	pReq->u.Out.pSession = pSession;
1370	pReq->Hdr.pSession = pSession;
1371
1372	return VINF_SUCCESS;
1373	}
1374
1375	case SUPDRV_IDC_REQ_DISCONNECT:
1376	{
1377	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_DISCONNECT, sizeof(*pReqHdr));
1378
1379	#ifdef RT_OS_WINDOWS
1380	/* Windows will destroy the session when the file object is destroyed. */
1381	#else
1382	supdrvCloseSession(pDevExt, pSession);
1383	#endif
1384	return pReqHdr->rc = VINF_SUCCESS;
1385	}
1386
1387	case SUPDRV_IDC_REQ_GET_SYMBOL:
1388	{
1389	PSUPDRVIDCREQGETSYM pReq = (PSUPDRVIDCREQGETSYM)pReqHdr;
1390	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_GET_SYMBOL, sizeof(*pReq));
1391
1392	pReq->Hdr.rc = supdrvIDC_LdrGetSymbol(pDevExt, pSession, pReq);
1393	return VINF_SUCCESS;
1394	}
1395
1396	case SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY:
1397	{
1398	PSUPDRVIDCREQCOMPREGFACTORY pReq = (PSUPDRVIDCREQCOMPREGFACTORY)pReqHdr;
1399	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY, sizeof(*pReq));
1400
1401	pReq->Hdr.rc = SUPR0ComponentRegisterFactory(pSession, pReq->u.In.pFactory);
1402	return VINF_SUCCESS;
1403	}
1404
1405	case SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY:
1406	{
1407	PSUPDRVIDCREQCOMPDEREGFACTORY pReq = (PSUPDRVIDCREQCOMPDEREGFACTORY)pReqHdr;
1408	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY, sizeof(*pReq));
1409
1410	pReq->Hdr.rc = SUPR0ComponentDeregisterFactory(pSession, pReq->u.In.pFactory);
1411	return VINF_SUCCESS;
1412	}
1413
1414	default:
1415	Log(("Unknown IDC %#lx\n", (long)uReq));
1416	break;
1417	}
1418
1419	#undef REQ_CHECK_IDC_SIZE
1420	return VERR_NOT_SUPPORTED;
1421	}
1422
1423
1424	/**
1425	* Register a object for reference counting.
1426	* The object is registered with one reference in the specified session.
1427	*
1428	* @returns Unique identifier on success (pointer).
1429	* All future reference must use this identifier.
1430	* @returns NULL on failure.
1431	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
1432	* @param pvUser1 The first user argument.
1433	* @param pvUser2 The second user argument.
1434	*/
1435	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
1436	{
1437	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1438	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1439	PSUPDRVOBJ pObj;
1440	PSUPDRVUSAGE pUsage;
1441
1442	/*
1443	* Validate the input.
1444	*/
1445	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
1446	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
1447	AssertPtrReturn(pfnDestructor, NULL);
1448
1449	/*
1450	* Allocate and initialize the object.
1451	*/
1452	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
1453	if (!pObj)
1454	return NULL;
1455	pObj->u32Magic = SUPDRVOBJ_MAGIC;
1456	pObj->enmType = enmType;
1457	pObj->pNext = NULL;
1458	pObj->cUsage = 1;
1459	pObj->pfnDestructor = pfnDestructor;
1460	pObj->pvUser1 = pvUser1;
1461	pObj->pvUser2 = pvUser2;
1462	pObj->CreatorUid = pSession->Uid;
1463	pObj->CreatorGid = pSession->Gid;
1464	pObj->CreatorProcess= pSession->Process;
1465	supdrvOSObjInitCreator(pObj, pSession);
1466
1467	/*
1468	* Allocate the usage record.
1469	* (We keep freed usage records around to simplify SUPR0ObjAddRef().)
1470	*/
1471	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1472
1473	pUsage = pDevExt->pUsageFree;
1474	if (pUsage)
1475	pDevExt->pUsageFree = pUsage->pNext;
1476	else
1477	{
1478	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1479	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
1480	if (!pUsage)
1481	{
1482	RTMemFree(pObj);
1483	return NULL;
1484	}
1485	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1486	}
1487
1488	/*
1489	* Insert the object and create the session usage record.
1490	*/
1491	/* The object. */
1492	pObj->pNext = pDevExt->pObjs;
1493	pDevExt->pObjs = pObj;
1494
1495	/* The session record. */
1496	pUsage->cUsage = 1;
1497	pUsage->pObj = pObj;
1498	pUsage->pNext = pSession->pUsage;
1499	Log2(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));
1500	pSession->pUsage = pUsage;
1501
1502	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1503
1504	Log(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
1505	return pObj;
1506	}
1507
1508
1509	/**
1510	* Increment the reference counter for the object associating the reference
1511	* with the specified session.
1512	*
1513	* @returns IPRT status code.
1514	* @param pvObj The identifier returned by SUPR0ObjRegister().
1515	* @param pSession The session which is referencing the object.
1516	*
1517	* @remarks The caller should not own any spinlocks and must carefully protect
1518	* itself against potential race with the destructor so freed memory
1519	* isn't accessed here.
1520	*/
1521	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
1522	{
1523	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1524	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1525	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
1526	PSUPDRVUSAGE pUsagePre;
1527	PSUPDRVUSAGE pUsage;
1528
1529	/*
1530	* Validate the input.
1531	* Be ready for the destruction race (someone might be stuck in the
1532	* destructor waiting a lock we own).
1533	*/
1534	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1535	AssertPtrReturn(pObj, VERR_INVALID_POINTER);
1536	AssertMsgReturn(pObj->u32Magic == SUPDRVOBJ_MAGIC \|\| pObj->u32Magic == SUPDRVOBJ_MAGIC + 1,
1537	("Invalid pvObj=%p magic=%#x (expected %#x or %#x)\n", pvObj, pObj->u32Magic, SUPDRVOBJ_MAGIC, SUPDRVOBJ_MAGIC + 1),
1538	VERR_INVALID_PARAMETER);
1539
1540	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1541
1542	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
1543	{
1544	AssertMsg(pObj->u32Magic == SUPDRVOBJ_MAGIC + 1, ("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
1545
1546	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1547	return VERR_WRONG_ORDER;
1548	}
1549
1550	/*
1551	* Preallocate the usage record.
1552	*/
1553	pUsagePre = pDevExt->pUsageFree;
1554	if (pUsagePre)
1555	pDevExt->pUsageFree = pUsagePre->pNext;
1556	else
1557	{
1558	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1559	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
1560	if (!pUsagePre)
1561	return VERR_NO_MEMORY;
1562
1563	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1564	}
1565
1566	/*
1567	* Reference the object.
1568	*/
1569	pObj->cUsage++;
1570
1571	/*
1572	* Look for the session record.
1573	*/
1574	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
1575	{
1576	Log(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));
1577	if (pUsage->pObj == pObj)
1578	break;
1579	}
1580	if (pUsage)
1581	pUsage->cUsage++;
1582	else
1583	{
1584	/* create a new session record. */
1585	pUsagePre->cUsage = 1;
1586	pUsagePre->pObj = pObj;
1587	pUsagePre->pNext = pSession->pUsage;
1588	pSession->pUsage = pUsagePre;
1589	Log(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));
1590
1591	pUsagePre = NULL;
1592	}
1593
1594	/*
1595	* Put any unused usage record into the free list..
1596	*/
1597	if (pUsagePre)
1598	{
1599	pUsagePre->pNext = pDevExt->pUsageFree;
1600	pDevExt->pUsageFree = pUsagePre;
1601	}
1602
1603	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1604
1605	return VINF_SUCCESS;
1606	}
1607
1608
1609	/**
1610	* Decrement / destroy a reference counter record for an object.
1611	*
1612	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
1613	*
1614	* @returns IPRT status code.
1615	* @param pvObj The identifier returned by SUPR0ObjRegister().
1616	* @param pSession The session which is referencing the object.
1617	*/
1618	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
1619	{
1620	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1621	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1622	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
1623	bool fDestroy = false;
1624	PSUPDRVUSAGE pUsage;
1625	PSUPDRVUSAGE pUsagePrev;
1626
1627	/*
1628	* Validate the input.
1629	*/
1630	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1631	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
1632	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
1633	VERR_INVALID_PARAMETER);
1634
1635	/*
1636	* Acquire the spinlock and look for the usage record.
1637	*/
1638	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1639
1640	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
1641	pUsage;
1642	pUsagePrev = pUsage, pUsage = pUsage->pNext)
1643	{
1644	Log2(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));
1645	if (pUsage->pObj == pObj)
1646	{
1647	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
1648	if (pUsage->cUsage > 1)
1649	{
1650	pObj->cUsage--;
1651	pUsage->cUsage--;
1652	}
1653	else
1654	{
1655	/*
1656	* Free the session record.
1657	*/
1658	if (pUsagePrev)
1659	pUsagePrev->pNext = pUsage->pNext;
1660	else
1661	pSession->pUsage = pUsage->pNext;
1662	pUsage->pNext = pDevExt->pUsageFree;
1663	pDevExt->pUsageFree = pUsage;
1664
1665	/* What about the object? */
1666	if (pObj->cUsage > 1)
1667	pObj->cUsage--;
1668	else
1669	{
1670	/*
1671	* Object is to be destroyed, unlink it.
1672	*/
1673	pObj->u32Magic = SUPDRVOBJ_MAGIC + 1;
1674	fDestroy = true;
1675	if (pDevExt->pObjs == pObj)
1676	pDevExt->pObjs = pObj->pNext;
1677	else
1678	{
1679	PSUPDRVOBJ pObjPrev;
1680	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
1681	if (pObjPrev->pNext == pObj)
1682	{
1683	pObjPrev->pNext = pObj->pNext;
1684	break;
1685	}
1686	Assert(pObjPrev);
1687	}
1688	}
1689	}
1690	break;
1691	}
1692	}
1693
1694	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1695
1696	/*
1697	* Call the destructor and free the object if required.
1698	*/
1699	if (fDestroy)
1700	{
1701	Log(("SUPR0ObjRelease: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
1702	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
1703	if (pObj->pfnDestructor)
1704	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
1705	RTMemFree(pObj);
1706	}
1707
1708	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
1709	return pUsage ? VINF_SUCCESS : VERR_INVALID_PARAMETER;
1710	}
1711
1712	/**
1713	* Verifies that the current process can access the specified object.
1714	*
1715	* @returns The following IPRT status code:
1716	* @retval VINF_SUCCESS if access was granted.
1717	* @retval VERR_PERMISSION_DENIED if denied access.
1718	* @retval VERR_INVALID_PARAMETER if invalid parameter.
1719	*
1720	* @param pvObj The identifier returned by SUPR0ObjRegister().
1721	* @param pSession The session which wishes to access the object.
1722	* @param pszObjName Object string name. This is optional and depends on the object type.
1723	*
1724	* @remark The caller is responsible for making sure the object isn't removed while
1725	* we're inside this function. If uncertain about this, just call AddRef before calling us.
1726	*/
1727	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
1728	{
1729	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
1730	int rc;
1731
1732	/*
1733	* Validate the input.
1734	*/
1735	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1736	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
1737	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
1738	VERR_INVALID_PARAMETER);
1739
1740	/*
1741	* Check access. (returns true if a decision has been made.)
1742	*/
1743	rc = VERR_INTERNAL_ERROR;
1744	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
1745	return rc;
1746
1747	/*
1748	* Default policy is to allow the user to access his own
1749	* stuff but nothing else.
1750	*/
1751	if (pObj->CreatorUid == pSession->Uid)
1752	return VINF_SUCCESS;
1753	return VERR_PERMISSION_DENIED;
1754	}
1755
1756
1757	/**
1758	* Lock pages.
1759	*
1760	* @returns IPRT status code.
1761	* @param pSession Session to which the locked memory should be associated.
1762	* @param pvR3 Start of the memory range to lock.
1763	* This must be page aligned.
1764	* @param cb Size of the memory range to lock.
1765	* This must be page aligned.
1766	*/
1767	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
1768	{
1769	int rc;
1770	SUPDRVMEMREF Mem = {0};
1771	const size_t cb = (size_t)cPages << PAGE_SHIFT;
1772	LogFlow(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
1773
1774	/*
1775	* Verify input.
1776	*/
1777	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1778	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
1779	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
1780	\|\| !pvR3)
1781	{
1782	Log(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
1783	return VERR_INVALID_PARAMETER;
1784	}
1785
1786	#ifdef RT_OS_WINDOWS /* A temporary hack for windows, will be removed once all ring-3 code has been cleaned up. */
1787	/* First check if we allocated it using SUPPageAlloc; if so then we don't need to lock it again */
1788	rc = supdrvPageGetPhys(pSession, pvR3, cPages, paPages);
1789	if (RT_SUCCESS(rc))
1790	return rc;
1791	#endif
1792
1793	/*
1794	* Let IPRT do the job.
1795	*/
1796	Mem.eType = MEMREF_TYPE_LOCKED;
1797	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTR0ProcHandleSelf());
1798	if (RT_SUCCESS(rc))
1799	{
1800	uint32_t iPage = cPages;
1801	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
1802	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
1803
1804	while (iPage-- > 0)
1805	{
1806	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
1807	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
1808	{
1809	AssertMsgFailed(("iPage=%d\n", iPage));
1810	rc = VERR_INTERNAL_ERROR;
1811	break;
1812	}
1813	}
1814	if (RT_SUCCESS(rc))
1815	rc = supdrvMemAdd(&Mem, pSession);
1816	if (RT_FAILURE(rc))
1817	{
1818	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
1819	AssertRC(rc2);
1820	}
1821	}
1822
1823	return rc;
1824	}
1825
1826
1827	/**
1828	* Unlocks the memory pointed to by pv.
1829	*
1830	* @returns IPRT status code.
1831	* @param pSession Session to which the memory was locked.
1832	* @param pvR3 Memory to unlock.
1833	*/
1834	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
1835	{
1836	LogFlow(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
1837	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1838	#ifdef RT_OS_WINDOWS
1839	/*
1840	* Temporary hack for windows - SUPR0PageFree will unlock SUPR0PageAlloc
1841	* allocations; ignore this call.
1842	*/
1843	if (supdrvPageWasLockedByPageAlloc(pSession, pvR3))
1844	{
1845	LogFlow(("Page will be unlocked in SUPR0PageFree -> ignore\n"));
1846	return VINF_SUCCESS;
1847	}
1848	#endif
1849	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
1850	}
1851
1852
1853	/**
1854	* Allocates a chunk of page aligned memory with contiguous and fixed physical
1855	* backing.
1856	*
1857	* @returns IPRT status code.
1858	* @param pSession Session data.
1859	* @param cb Number of bytes to allocate.
1860	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
1861	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
1862	* @param pHCPhys Where to put the physical address of allocated memory.
1863	*/
1864	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
1865	{
1866	int rc;
1867	SUPDRVMEMREF Mem = {0};
1868	LogFlow(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
1869
1870	/*
1871	* Validate input.
1872	*/
1873	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1874	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
1875	{
1876	Log(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
1877	pSession, ppvR0, ppvR3, pHCPhys));
1878	return VERR_INVALID_PARAMETER;
1879
1880	}
1881	if (cPages < 1 \|\| cPages >= 256)
1882	{
1883	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256\n", cPages));
1884	return VERR_INVALID_PARAMETER;
1885	}
1886
1887	/*
1888	* Let IPRT do the job.
1889	*/
1890	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
1891	if (RT_SUCCESS(rc))
1892	{
1893	int rc2;
1894	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
1895	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
1896	if (RT_SUCCESS(rc))
1897	{
1898	Mem.eType = MEMREF_TYPE_CONT;
1899	rc = supdrvMemAdd(&Mem, pSession);
1900	if (!rc)
1901	{
1902	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
1903	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
1904	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
1905	return 0;
1906	}
1907
1908	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
1909	AssertRC(rc2);
1910	}
1911	rc2 = RTR0MemObjFree(Mem.MemObj, false);
1912	AssertRC(rc2);
1913	}
1914
1915	return rc;
1916	}
1917
1918
1919	/**
1920	* Frees memory allocated using SUPR0ContAlloc().
1921	*
1922	* @returns IPRT status code.
1923	* @param pSession The session to which the memory was allocated.
1924	* @param uPtr Pointer to the memory (ring-3 or ring-0).
1925	*/
1926	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
1927	{
1928	LogFlow(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
1929	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1930	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
1931	}
1932
1933
1934	/**
1935	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
1936	*
1937	* The memory isn't zeroed.
1938	*
1939	* @returns IPRT status code.
1940	* @param pSession Session data.
1941	* @param cPages Number of pages to allocate.
1942	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
1943	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
1944	* @param paPages Where to put the physical addresses of allocated memory.
1945	*/
1946	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
1947	{
1948	unsigned iPage;
1949	int rc;
1950	SUPDRVMEMREF Mem = {0};
1951	LogFlow(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
1952
1953	/*
1954	* Validate input.
1955	*/
1956	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1957	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
1958	{
1959	Log(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
1960	pSession, ppvR3, ppvR0, paPages));
1961	return VERR_INVALID_PARAMETER;
1962
1963	}
1964	if (cPages < 1 \|\| cPages > 256)
1965	{
1966	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
1967	return VERR_INVALID_PARAMETER;
1968	}
1969
1970	/*
1971	* Let IPRT do the work.
1972	*/
1973	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
1974	if (RT_SUCCESS(rc))
1975	{
1976	int rc2;
1977	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
1978	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
1979	if (RT_SUCCESS(rc))
1980	{
1981	Mem.eType = MEMREF_TYPE_LOW;
1982	rc = supdrvMemAdd(&Mem, pSession);
1983	if (!rc)
1984	{
1985	for (iPage = 0; iPage < cPages; iPage++)
1986	{
1987	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
1988	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%VHp\n", paPages[iPage]));
1989	}
1990	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
1991	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
1992	return 0;
1993	}
1994
1995	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
1996	AssertRC(rc2);
1997	}
1998
1999	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2000	AssertRC(rc2);
2001	}
2002
2003	return rc;
2004	}
2005
2006
2007	/**
2008	* Frees memory allocated using SUPR0LowAlloc().
2009	*
2010	* @returns IPRT status code.
2011	* @param pSession The session to which the memory was allocated.
2012	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2013	*/
2014	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2015	{
2016	LogFlow(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2017	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2018	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
2019	}
2020
2021
2022
2023	/**
2024	* Allocates a chunk of memory with both R0 and R3 mappings.
2025	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
2026	*
2027	* @returns IPRT status code.
2028	* @param pSession The session to associated the allocation with.
2029	* @param cb Number of bytes to allocate.
2030	* @param ppvR0 Where to store the address of the Ring-0 mapping.
2031	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2032	*/
2033	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
2034	{
2035	int rc;
2036	SUPDRVMEMREF Mem = {0};
2037	LogFlow(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
2038
2039	/*
2040	* Validate input.
2041	*/
2042	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2043	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
2044	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
2045	if (cb < 1 \|\| cb >= _4M)
2046	{
2047	Log(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
2048	return VERR_INVALID_PARAMETER;
2049	}
2050
2051	/*
2052	* Let IPRT do the work.
2053	*/
2054	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
2055	if (RT_SUCCESS(rc))
2056	{
2057	int rc2;
2058	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2059	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2060	if (RT_SUCCESS(rc))
2061	{
2062	Mem.eType = MEMREF_TYPE_MEM;
2063	rc = supdrvMemAdd(&Mem, pSession);
2064	if (!rc)
2065	{
2066	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2067	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2068	return VINF_SUCCESS;
2069	}
2070	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2071	AssertRC(rc2);
2072	}
2073
2074	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2075	AssertRC(rc2);
2076	}
2077
2078	return rc;
2079	}
2080
2081
2082	/**
2083	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
2084	*
2085	* @returns IPRT status code.
2086	* @param pSession The session to which the memory was allocated.
2087	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2088	* @param paPages Where to store the physical addresses.
2089	*/
2090	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
2091	{
2092	PSUPDRVBUNDLE pBundle;
2093	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2094	LogFlow(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
2095
2096	/*
2097	* Validate input.
2098	*/
2099	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2100	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
2101	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
2102
2103	/*
2104	* Search for the address.
2105	*/
2106	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2107	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2108	{
2109	if (pBundle->cUsed > 0)
2110	{
2111	unsigned i;
2112	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2113	{
2114	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
2115	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2116	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2117	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2118	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
2119	)
2120	)
2121	{
2122	const unsigned cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2123	unsigned iPage;
2124	for (iPage = 0; iPage < cPages; iPage++)
2125	{
2126	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2127	paPages[iPage].uReserved = 0;
2128	}
2129	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2130	return VINF_SUCCESS;
2131	}
2132	}
2133	}
2134	}
2135	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2136	Log(("Failed to find %p!!!\n", (void *)uPtr));
2137	return VERR_INVALID_PARAMETER;
2138	}
2139
2140
2141	/**
2142	* Free memory allocated by SUPR0MemAlloc().
2143	*
2144	* @returns IPRT status code.
2145	* @param pSession The session owning the allocation.
2146	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2147	*/
2148	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2149	{
2150	LogFlow(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2151	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2152	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
2153	}
2154
2155
2156	/**
2157	* Allocates a chunk of memory with only a R3 mappings.
2158	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
2159	*
2160	* @returns IPRT status code.
2161	* @param pSession The session to associated the allocation with.
2162	* @param cPages The number of pages to allocate.
2163	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2164	* @param paPages Where to store the addresses of the pages. Optional.
2165	*/
2166	SUPR0DECL(int) SUPR0PageAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR3PTR ppvR3, PRTHCPHYS paPages)
2167	{
2168	int rc;
2169	SUPDRVMEMREF Mem = {0};
2170	LogFlow(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
2171
2172	/*
2173	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2174	*/
2175	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2176	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
2177	if (cPages < 1 \|\| cPages > (128 * _1M)/PAGE_SIZE)
2178	{
2179	Log(("SUPR0PageAlloc: Illegal request cb=%u; must be greater than 0 and smaller than 128MB.\n", cPages));
2180	return VERR_INVALID_PARAMETER;
2181	}
2182
2183	/*
2184	* Let IPRT do the work.
2185	*/
2186	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
2187	if (RT_SUCCESS(rc))
2188	{
2189	int rc2;
2190	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2191	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2192	if (RT_SUCCESS(rc))
2193	{
2194	Mem.eType = MEMREF_TYPE_LOCKED_SUP;
2195	rc = supdrvMemAdd(&Mem, pSession);
2196	if (!rc)
2197	{
2198	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2199	if (paPages)
2200	{
2201	uint32_t iPage = cPages;
2202	while (iPage-- > 0)
2203	{
2204	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
2205	Assert(paPages[iPage] != NIL_RTHCPHYS);
2206	}
2207	}
2208	return VINF_SUCCESS;
2209	}
2210	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2211	AssertRC(rc2);
2212	}
2213
2214	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2215	AssertRC(rc2);
2216	}
2217	return rc;
2218	}
2219
2220
2221	#ifdef RT_OS_WINDOWS
2222	/**
2223	* Check if the pages were locked by SUPR0PageAlloc
2224	*
2225	* This function will be removed along with the lock/unlock hacks when
2226	* we've cleaned up the ring-3 code properly.
2227	*
2228	* @returns boolean
2229	* @param pSession The session to which the memory was allocated.
2230	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc().
2231	*/
2232	static bool supdrvPageWasLockedByPageAlloc(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2233	{
2234	PSUPDRVBUNDLE pBundle;
2235	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2236	LogFlow(("SUPR0PageIsLockedByPageAlloc: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2237
2238	/*
2239	* Search for the address.
2240	*/
2241	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2242	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2243	{
2244	if (pBundle->cUsed > 0)
2245	{
2246	unsigned i;
2247	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2248	{
2249	if ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED_SUP
2250	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2251	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2252	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
2253	{
2254	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2255	return true;
2256	}
2257	}
2258	}
2259	}
2260	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2261	return false;
2262	}
2263
2264
2265	/**
2266	* Get the physical addresses of memory allocated using SUPR0PageAlloc().
2267	*
2268	* This function will be removed along with the lock/unlock hacks when
2269	* we've cleaned up the ring-3 code properly.
2270	*
2271	* @returns IPRT status code.
2272	* @param pSession The session to which the memory was allocated.
2273	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc().
2274	* @param cPages Number of pages in paPages
2275	* @param paPages Where to store the physical addresses.
2276	*/
2277	static int supdrvPageGetPhys(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
2278	{
2279	PSUPDRVBUNDLE pBundle;
2280	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2281	LogFlow(("supdrvPageGetPhys: pSession=%p pvR3=%p cPages=%#lx paPages=%p\n", pSession, (void *)pvR3, (long)cPages, paPages));
2282
2283	/*
2284	* Search for the address.
2285	*/
2286	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2287	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2288	{
2289	if (pBundle->cUsed > 0)
2290	{
2291	unsigned i;
2292	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2293	{
2294	if ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED_SUP
2295	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2296	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2297	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
2298	{
2299	uint32_t iPage = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2300	cPages = RT_MIN(iPage, cPages);
2301	for (iPage = 0; iPage < cPages; iPage++)
2302	paPages[iPage] = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2303	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2304	return VINF_SUCCESS;
2305	}
2306	}
2307	}
2308	}
2309	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2310	return VERR_INVALID_PARAMETER;
2311	}
2312	#endif /* RT_OS_WINDOWS */
2313
2314
2315	/**
2316	* Free memory allocated by SUPR0PageAlloc().
2317	*
2318	* @returns IPRT status code.
2319	* @param pSession The session owning the allocation.
2320	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc().
2321	*/
2322	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2323	{
2324	LogFlow(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2325	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2326	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED_SUP);
2327	}
2328
2329
2330	/**
2331	* Maps the GIP into userspace and/or get the physical address of the GIP.
2332	*
2333	* @returns IPRT status code.
2334	* @param pSession Session to which the GIP mapping should belong.
2335	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
2336	* @param pHCPhysGip Where to store the physical address. (optional)
2337	*
2338	* @remark There is no reference counting on the mapping, so one call to this function
2339	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
2340	* and remove the session as a GIP user.
2341	*/
2342	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
2343	{
2344	int rc = 0;
2345	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2346	RTR3PTR pGip = NIL_RTR3PTR;
2347	RTHCPHYS HCPhys = NIL_RTHCPHYS;
2348	LogFlow(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
2349
2350	/*
2351	* Validate
2352	*/
2353	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2354	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
2355	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
2356
2357	RTSemFastMutexRequest(pDevExt->mtxGip);
2358	if (pDevExt->pGip)
2359	{
2360	/*
2361	* Map it?
2362	*/
2363	if (ppGipR3)
2364	{
2365	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
2366	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
2367	RTMEM_PROT_READ, RTR0ProcHandleSelf());
2368	if (RT_SUCCESS(rc))
2369	{
2370	pGip = RTR0MemObjAddressR3(pSession->GipMapObjR3);
2371	rc = VINF_SUCCESS; /** @todo remove this and replace the !rc below with RT_SUCCESS(rc). */
2372	}
2373	}
2374
2375	/*
2376	* Get physical address.
2377	*/
2378	if (pHCPhysGip && !rc)
2379	HCPhys = pDevExt->HCPhysGip;
2380
2381	/*
2382	* Reference globally.
2383	*/
2384	if (!pSession->fGipReferenced && !rc)
2385	{
2386	pSession->fGipReferenced = 1;
2387	pDevExt->cGipUsers++;
2388	if (pDevExt->cGipUsers == 1)
2389	{
2390	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
2391	unsigned i;
2392
2393	LogFlow(("SUPR0GipMap: Resumes GIP updating\n"));
2394
2395	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
2396	ASMAtomicXchgU32(&pGip->aCPUs[i].u32TransactionId, pGip->aCPUs[i].u32TransactionId & ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
2397	ASMAtomicXchgU64(&pGip->u64NanoTSLastUpdateHz, 0);
2398
2399	rc = RTTimerStart(pDevExt->pGipTimer, 0);
2400	AssertRC(rc); rc = VINF_SUCCESS;
2401	}
2402	}
2403	}
2404	else
2405	{
2406	rc = SUPDRV_ERR_GENERAL_FAILURE;
2407	Log(("SUPR0GipMap: GIP is not available!\n"));
2408	}
2409	RTSemFastMutexRelease(pDevExt->mtxGip);
2410
2411	/*
2412	* Write returns.
2413	*/
2414	if (pHCPhysGip)
2415	*pHCPhysGip = HCPhys;
2416	if (ppGipR3)
2417	*ppGipR3 = pGip;
2418
2419	#ifdef DEBUG_DARWIN_GIP
2420	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx ppGip=%p GipMapObjR3\n", rc, (unsigned long)HCPhys, pGip, pSession->GipMapObjR3));
2421	#else
2422	LogFlow(("SUPR0GipMap: returns %d pHCPhysGip=%lx ppGipR3=%p\n", rc, (unsigned long)HCPhys, (void *)(uintptr_t)pGip));
2423	#endif
2424	return rc;
2425	}
2426
2427
2428	/**
2429	* Unmaps any user mapping of the GIP and terminates all GIP access
2430	* from this session.
2431	*
2432	* @returns IPRT status code.
2433	* @param pSession Session to which the GIP mapping should belong.
2434	*/
2435	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
2436	{
2437	int rc = VINF_SUCCESS;
2438	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2439	#ifdef DEBUG_DARWIN_GIP
2440	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
2441	pSession,
2442	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
2443	pSession->GipMapObjR3));
2444	#else
2445	LogFlow(("SUPR0GipUnmap: pSession=%p\n", pSession));
2446	#endif
2447	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2448
2449	RTSemFastMutexRequest(pDevExt->mtxGip);
2450
2451	/*
2452	* Unmap anything?
2453	*/
2454	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
2455	{
2456	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
2457	AssertRC(rc);
2458	if (RT_SUCCESS(rc))
2459	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
2460	}
2461
2462	/*
2463	* Dereference global GIP.
2464	*/
2465	if (pSession->fGipReferenced && !rc)
2466	{
2467	pSession->fGipReferenced = 0;
2468	if ( pDevExt->cGipUsers > 0
2469	&& !--pDevExt->cGipUsers)
2470	{
2471	LogFlow(("SUPR0GipUnmap: Suspends GIP updating\n"));
2472	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = 0;
2473	}
2474	}
2475
2476	RTSemFastMutexRelease(pDevExt->mtxGip);
2477
2478	return rc;
2479	}
2480
2481
2482	/**
2483	* Register a component factory with the support driver.
2484	*
2485	* This is currently restricted to kernel sessions only.
2486	*
2487	* @returns VBox status code.
2488	* @retval VINF_SUCCESS on success.
2489	* @retval VERR_NO_MEMORY if we're out of memory.
2490	* @retval VERR_ALREADY_EXISTS if the factory has already been registered.
2491	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
2492	* @retval VERR_INVALID_PARAMETER on invalid parameter.
2493	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
2494	*
2495	* @param pSession The SUPDRV session (must be a ring-0 session).
2496	* @param pFactory Pointer to the component factory registration structure.
2497	*
2498	* @remarks This interface is also available via SUPR0IdcComponentRegisterFactory.
2499	*/
2500	SUPR0DECL(int) SUPR0ComponentRegisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
2501	{
2502	PSUPDRVFACTORYREG pNewReg;
2503	const char *psz;
2504	int rc;
2505
2506	/*
2507	* Validate parameters.
2508	*/
2509	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2510	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
2511	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
2512	AssertPtrReturn(pFactory->pfnQueryFactoryInterface, VERR_INVALID_POINTER);
2513	psz = (const char *)memchr(pFactory->szName, '\0', sizeof(pFactory->szName));
2514	AssertReturn(psz, VERR_INVALID_PARAMETER);
2515
2516	/*
2517	* Allocate and initialize a new registration structure.
2518	*/
2519	pNewReg = (PSUPDRVFACTORYREG)RTMemAlloc(sizeof(SUPDRVFACTORYREG));
2520	if (pNewReg)
2521	{
2522	pNewReg->pNext = NULL;
2523	pNewReg->pFactory = pFactory;
2524	pNewReg->pSession = pSession;
2525	pNewReg->cchName = psz - &pFactory->szName[0];
2526
2527	/*
2528	* Add it to the tail of the list after checking for prior registration.
2529	*/
2530	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
2531	if (RT_SUCCESS(rc))
2532	{
2533	PSUPDRVFACTORYREG pPrev = NULL;
2534	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
2535	while (pCur && pCur->pFactory != pFactory)
2536	{
2537	pPrev = pCur;
2538	pCur = pCur->pNext;
2539	}
2540	if (!pCur)
2541	{
2542	if (pPrev)
2543	pPrev->pNext = pNewReg;
2544	else
2545	pSession->pDevExt->pComponentFactoryHead = pNewReg;
2546	rc = VINF_SUCCESS;
2547	}
2548	else
2549	rc = VERR_ALREADY_EXISTS;
2550
2551	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
2552	}
2553
2554	if (RT_FAILURE(rc))
2555	RTMemFree(pNewReg);
2556	}
2557	else
2558	rc = VERR_NO_MEMORY;
2559	return rc;
2560	}
2561
2562
2563	/**
2564	* Deregister a component factory.
2565	*
2566	* @returns VBox status code.
2567	* @retval VINF_SUCCESS on success.
2568	* @retval VERR_NOT_FOUND if the factory wasn't registered.
2569	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
2570	* @retval VERR_INVALID_PARAMETER on invalid parameter.
2571	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
2572	*
2573	* @param pSession The SUPDRV session (must be a ring-0 session).
2574	* @param pFactory Pointer to the component factory registration structure
2575	* previously passed SUPR0ComponentRegisterFactory().
2576	*
2577	* @remarks This interface is also available via SUPR0IdcComponentDeregisterFactory.
2578	*/
2579	SUPR0DECL(int) SUPR0ComponentDeregisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
2580	{
2581	int rc;
2582
2583	/*
2584	* Validate parameters.
2585	*/
2586	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2587	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
2588	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
2589
2590	/*
2591	* Take the lock and look for the registration record.
2592	*/
2593	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
2594	if (RT_SUCCESS(rc))
2595	{
2596	PSUPDRVFACTORYREG pPrev = NULL;
2597	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
2598	while (pCur && pCur->pFactory != pFactory)
2599	{
2600	pPrev = pCur;
2601	pCur = pCur->pNext;
2602	}
2603	if (pCur)
2604	{
2605	if (!pPrev)
2606	pSession->pDevExt->pComponentFactoryHead = pCur->pNext;
2607	else
2608	pPrev->pNext = pCur->pNext;
2609
2610	pCur->pNext = NULL;
2611	pCur->pFactory = NULL;
2612	pCur->pSession = NULL;
2613	rc = VINF_SUCCESS;
2614	}
2615	else
2616	rc = VERR_NOT_FOUND;
2617
2618	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
2619
2620	RTMemFree(pCur);
2621	}
2622	return rc;
2623	}
2624
2625
2626	/**
2627	* Queries a component factory.
2628	*
2629	* @returns VBox status code.
2630	* @retval VERR_INVALID_PARAMETER on invalid parameter.
2631	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
2632	* @retval VERR_SUPDRV_COMPONENT_NOT_FOUND if the component factory wasn't found.
2633	* @retval VERR_SUPDRV_INTERFACE_NOT_SUPPORTED if the interface wasn't supported.
2634	*
2635	* @param pSession The SUPDRV session.
2636	* @param pszName The name of the component factory.
2637	* @param pszInterfaceUuid The UUID of the factory interface (stringified).
2638	* @param ppvFactoryIf Where to store the factory interface.
2639	*/
2640	SUPR0DECL(int) SUPR0ComponentQueryFactory(PSUPDRVSESSION pSession, const char pszName, const char pszInterfaceUuid, void **ppvFactoryIf)
2641	{
2642	const char *pszEnd;
2643	size_t cchName;
2644	int rc;
2645
2646	/*
2647	* Validate parameters.
2648	*/
2649	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2650
2651	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
2652	pszEnd = memchr(pszName, '\0', RT_SIZEOFMEMB(SUPDRVFACTORY, szName));
2653	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
2654	cchName = pszEnd - pszName;
2655
2656	AssertPtrReturn(pszInterfaceUuid, VERR_INVALID_POINTER);
2657	pszEnd = memchr(pszInterfaceUuid, '\0', RTUUID_STR_LENGTH);
2658	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
2659
2660	AssertPtrReturn(ppvFactoryIf, VERR_INVALID_POINTER);
2661	*ppvFactoryIf = NULL;
2662
2663	/*
2664	* Take the lock and try all factories by this name.
2665	*/
2666	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
2667	if (RT_SUCCESS(rc))
2668	{
2669	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
2670	rc = VERR_SUPDRV_COMPONENT_NOT_FOUND;
2671	while (pCur)
2672	{
2673	if ( pCur->cchName == cchName
2674	&& !memcmp(pCur->pFactory->szName, pszName, cchName))
2675	{
2676	void *pvFactory = pCur->pFactory->pfnQueryFactoryInterface(pCur->pFactory, pSession, pszInterfaceUuid);
2677	if (pvFactory)
2678	{
2679	*ppvFactoryIf = pvFactory;
2680	rc = VINF_SUCCESS;
2681	break;
2682	}
2683	rc = VERR_SUPDRV_INTERFACE_NOT_SUPPORTED;
2684	}
2685
2686	/* next */
2687	pCur = pCur->pNext;
2688	}
2689
2690	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
2691	}
2692	return rc;
2693	}
2694
2695
2696	/**
2697	* Adds a memory object to the session.
2698	*
2699	* @returns IPRT status code.
2700	* @param pMem Memory tracking structure containing the
2701	* information to track.
2702	* @param pSession The session.
2703	*/
2704	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
2705	{
2706	PSUPDRVBUNDLE pBundle;
2707	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2708
2709	/*
2710	* Find free entry and record the allocation.
2711	*/
2712	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2713	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2714	{
2715	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
2716	{
2717	unsigned i;
2718	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2719	{
2720	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
2721	{
2722	pBundle->cUsed++;
2723	pBundle->aMem[i] = *pMem;
2724	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2725	return VINF_SUCCESS;
2726	}
2727	}
2728	AssertFailed(); /* !!this can't be happening!!! */
2729	}
2730	}
2731	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2732
2733	/*
2734	* Need to allocate a new bundle.
2735	* Insert into the last entry in the bundle.
2736	*/
2737	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
2738	if (!pBundle)
2739	return VERR_NO_MEMORY;
2740
2741	/* take last entry. */
2742	pBundle->cUsed++;
2743	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
2744
2745	/* insert into list. */
2746	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2747	pBundle->pNext = pSession->Bundle.pNext;
2748	pSession->Bundle.pNext = pBundle;
2749	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2750
2751	return VINF_SUCCESS;
2752	}
2753
2754
2755	/**
2756	* Releases a memory object referenced by pointer and type.
2757	*
2758	* @returns IPRT status code.
2759	* @param pSession Session data.
2760	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
2761	* @param eType Memory type.
2762	*/
2763	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
2764	{
2765	PSUPDRVBUNDLE pBundle;
2766	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2767
2768	/*
2769	* Validate input.
2770	*/
2771	if (!uPtr)
2772	{
2773	Log(("Illegal address %p\n", (void *)uPtr));
2774	return VERR_INVALID_PARAMETER;
2775	}
2776
2777	/*
2778	* Search for the address.
2779	*/
2780	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2781	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2782	{
2783	if (pBundle->cUsed > 0)
2784	{
2785	unsigned i;
2786	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2787	{
2788	if ( pBundle->aMem[i].eType == eType
2789	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2790	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2791	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2792	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
2793	)
2794	{
2795	/* Make a copy of it and release it outside the spinlock. */
2796	SUPDRVMEMREF Mem = pBundle->aMem[i];
2797	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
2798	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
2799	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
2800	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2801
2802	if (Mem.MapObjR3)
2803	{
2804	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
2805	AssertRC(rc); /** @todo figure out how to handle this. */
2806	}
2807	if (Mem.MemObj)
2808	{
2809	int rc = RTR0MemObjFree(Mem.MemObj, false);
2810	AssertRC(rc); /** @todo figure out how to handle this. */
2811	}
2812	return VINF_SUCCESS;
2813	}
2814	}
2815	}
2816	}
2817	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2818	Log(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
2819	return VERR_INVALID_PARAMETER;
2820	}
2821
2822
2823	#ifdef VBOX_WITH_IDT_PATCHING
2824	/**
2825	* Install IDT for the current CPU.
2826	*
2827	* @returns One of the following IPRT status codes:
2828	* @retval VINF_SUCCESS on success.
2829	* @retval VERR_IDT_FAILED.
2830	* @retval VERR_NO_MEMORY.
2831	* @param pDevExt The device extension.
2832	* @param pSession The session data.
2833	* @param pReq The request.
2834	*/
2835	static int supdrvIOCtl_IdtInstall(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPIDTINSTALL pReq)
2836	{
2837	PSUPDRVPATCHUSAGE pUsagePre;
2838	PSUPDRVPATCH pPatchPre;
2839	RTIDTR Idtr;
2840	PSUPDRVPATCH pPatch;
2841	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2842	LogFlow(("supdrvIOCtl_IdtInstall\n"));
2843
2844	/*
2845	* Preallocate entry for this CPU cause we don't wanna do
2846	* that inside the spinlock!
2847	*/
2848	pUsagePre = (PSUPDRVPATCHUSAGE)RTMemAlloc(sizeof(*pUsagePre));
2849	if (!pUsagePre)
2850	return VERR_NO_MEMORY;
2851
2852	/*
2853	* Take the spinlock and see what we need to do.
2854	*/
2855	RTSpinlockAcquireNoInts(pDevExt->Spinlock, &SpinlockTmp);
2856
2857	/* check if we already got a free patch. */
2858	if (!pDevExt->pIdtPatchesFree)
2859	{
2860	/*
2861	* Allocate a patch - outside the spinlock of course.
2862	*/
2863	RTSpinlockReleaseNoInts(pDevExt->Spinlock, &SpinlockTmp);
2864
2865	pPatchPre = (PSUPDRVPATCH)RTMemExecAlloc(sizeof(*pPatchPre));
2866	if (!pPatchPre)
2867	return VERR_NO_MEMORY;
2868
2869	RTSpinlockAcquireNoInts(pDevExt->Spinlock, &SpinlockTmp);
2870	}
2871	else
2872	{
2873	pPatchPre = pDevExt->pIdtPatchesFree;
2874	pDevExt->pIdtPatchesFree = pPatchPre->pNext;
2875	}
2876
2877	/* look for matching patch entry */
2878	ASMGetIDTR(&Idtr);
2879	pPatch = pDevExt->pIdtPatches;
2880	while (pPatch && pPatch->pvIdt != (void *)Idtr.pIdt)
2881	pPatch = pPatch->pNext;
2882
2883	if (!pPatch)
2884	{
2885	/*
2886	* Create patch.
2887	*/
2888	pPatch = supdrvIdtPatchOne(pDevExt, pPatchPre);
2889	if (pPatch)
2890	pPatchPre = NULL; /* mark as used. */
2891	}
2892	else
2893	{
2894	/*
2895	* Simply increment patch usage.
2896	*/
2897	pPatch->cUsage++;
2898	}
2899
2900	if (pPatch)
2901	{
2902	/*
2903	* Increment and add if need be the session usage record for this patch.
2904	*/
2905	PSUPDRVPATCHUSAGE pUsage = pSession->pPatchUsage;
2906	while (pUsage && pUsage->pPatch != pPatch)
2907	pUsage = pUsage->pNext;
2908
2909	if (!pUsage)
2910	{
2911	/*
2912	* Add usage record.
2913	*/
2914	pUsagePre->cUsage = 1;
2915	pUsagePre->pPatch = pPatch;
2916	pUsagePre->pNext = pSession->pPatchUsage;
2917	pSession->pPatchUsage = pUsagePre;
2918	pUsagePre = NULL; /* mark as used. */
2919	}
2920	else
2921	{
2922	/*
2923	* Increment usage count.
2924	*/
2925	pUsage->cUsage++;
2926	}
2927	}
2928
2929	/* free patch - we accumulate them for paranoid saftly reasons. */
2930	if (pPatchPre)
2931	{
2932	pPatchPre->pNext = pDevExt->pIdtPatchesFree;
2933	pDevExt->pIdtPatchesFree = pPatchPre;
2934	}
2935
2936	RTSpinlockReleaseNoInts(pDevExt->Spinlock, &SpinlockTmp);
2937
2938	/*
2939	* Free unused preallocated buffers.
2940	*/
2941	if (pUsagePre)
2942	RTMemFree(pUsagePre);
2943
2944	pReq->u.Out.u8Idt = pDevExt->u8Idt;
2945
2946	return pPatch ? VINF_SUCCESS : VERR_IDT_FAILED;
2947	}
2948
2949
2950	/**
2951	* This creates a IDT patch entry.
2952	* If the first patch being installed it'll also determin the IDT entry
2953	* to use.
2954	*
2955	* @returns pPatch on success.
2956	* @returns NULL on failure.
2957	* @param pDevExt Pointer to globals.
2958	* @param pPatch Patch entry to use.
2959	* This will be linked into SUPDRVDEVEXT::pIdtPatches on
2960	* successful return.
2961	* @remark Call must be owning the SUPDRVDEVEXT::Spinlock!
2962	*/
2963	static PSUPDRVPATCH supdrvIdtPatchOne(PSUPDRVDEVEXT pDevExt, PSUPDRVPATCH pPatch)
2964	{
2965	RTIDTR Idtr;
2966	PSUPDRVIDTE paIdt;
2967	LogFlow(("supdrvIOCtl_IdtPatchOne: pPatch=%p\n", pPatch));
2968
2969	/*
2970	* Get IDT.
2971	*/
2972	ASMGetIDTR(&Idtr);
2973	paIdt = (PSUPDRVIDTE)Idtr.pIdt;
2974	/*
2975	* Recent Linux kernels can be configured to 1G user /3G kernel.
2976	*/
2977	if ((uintptr_t)paIdt < 0x40000000)
2978	{
2979	AssertMsgFailed(("bad paIdt=%p\n", paIdt));
2980	return NULL;
2981	}
2982
2983	if (!pDevExt->u8Idt)
2984	{
2985	/*
2986	* Test out the alternatives.
2987	*
2988	* At the moment we do not support chaining thus we ASSUME that one of
2989	* these 48 entries is unused (which is not a problem on Win32 and
2990	* Linux to my knowledge).
2991	*/
2992	/** @todo we MUST change this detection to try grab an entry which is NOT in use. This can be
2993	* combined with gathering info about which guest system call gates we can hook up directly. */
2994	unsigned i;
2995	uint8_t u8Idt = 0;
2996	static uint8_t au8Ints[] =
2997	{
2998	#ifdef RT_OS_WINDOWS /* We don't use 0xef and above because they are system stuff on linux (ef is IPI,
2999	* local apic timer, or some other frequently fireing thing). */
3000	0xef, 0xee, 0xed, 0xec,
3001	#endif
3002	0xeb, 0xea, 0xe9, 0xe8,
3003	0xdf, 0xde, 0xdd, 0xdc,
3004	0x7b, 0x7a, 0x79, 0x78,
3005	0xbf, 0xbe, 0xbd, 0xbc,
3006	};
3007	#if defined(RT_ARCH_AMD64) && defined(DEBUG)
3008	static int s_iWobble = 0;
3009	unsigned iMax = !(s_iWobble++ % 2) ? 0x80 : 0x100;
3010	Log2(("IDT: Idtr=%p:%#x\n", (void *)Idtr.pIdt, (unsigned)Idtr.cbIdt));
3011	for (i = iMax - 0x80; i*16+15 < Idtr.cbIdt && i < iMax; i++)
3012	{
3013	Log2(("%#x: %04x:%08x%04x%04x P=%d DPL=%d IST=%d Type1=%#x u32Reserved=%#x u5Reserved=%#x\n",
3014	i, paIdt[i].u16SegSel, paIdt[i].u32OffsetTop, paIdt[i].u16OffsetHigh, paIdt[i].u16OffsetLow,
3015	paIdt[i].u1Present, paIdt[i].u2DPL, paIdt[i].u3IST, paIdt[i].u5Type2,
3016	paIdt[i].u32Reserved, paIdt[i].u5Reserved));
3017	}
3018	#endif
3019	/* look for entries which are not present or otherwise unused. */
3020	for (i = 0; i < sizeof(au8Ints) / sizeof(au8Ints[0]); i++)
3021	{
3022	u8Idt = au8Ints[i];
3023	if ( u8Idt * sizeof(SUPDRVIDTE) < Idtr.cbIdt
3024	&& ( !paIdt[u8Idt].u1Present
3025	\|\| paIdt[u8Idt].u5Type2 == 0))
3026	break;
3027	u8Idt = 0;
3028	}
3029	if (!u8Idt)
3030	{
3031	/* try again, look for a compatible entry .*/
3032	for (i = 0; i < sizeof(au8Ints) / sizeof(au8Ints[0]); i++)
3033	{
3034	u8Idt = au8Ints[i];
3035	if ( u8Idt * sizeof(SUPDRVIDTE) < Idtr.cbIdt
3036	&& paIdt[u8Idt].u1Present
3037	&& paIdt[u8Idt].u5Type2 == SUPDRV_IDTE_TYPE2_INTERRUPT_GATE
3038	&& !(paIdt[u8Idt].u16SegSel & 3))
3039	break;
3040	u8Idt = 0;
3041	}
3042	if (!u8Idt)
3043	{
3044	Log(("Failed to find appropirate IDT entry!!\n"));
3045	return NULL;
3046	}
3047	}
3048	pDevExt->u8Idt = u8Idt;
3049	LogFlow(("supdrvIOCtl_IdtPatchOne: u8Idt=%x\n", u8Idt));
3050	}
3051
3052	/*
3053	* Prepare the patch
3054	*/
3055	memset(pPatch, 0, sizeof(*pPatch));
3056	pPatch->pvIdt = paIdt;
3057	pPatch->cUsage = 1;
3058	pPatch->pIdtEntry = &paIdt[pDevExt->u8Idt];
3059	pPatch->SavedIdt = paIdt[pDevExt->u8Idt];
3060	pPatch->ChangedIdt.u16OffsetLow = (uint32_t)((uintptr_t)&pPatch->auCode[0] & 0xffff);
3061	pPatch->ChangedIdt.u16OffsetHigh = (uint32_t)((uintptr_t)&pPatch->auCode[0] >> 16);
3062	#ifdef RT_ARCH_AMD64
3063	pPatch->ChangedIdt.u32OffsetTop = (uint32_t)((uintptr_t)&pPatch->auCode[0] >> 32);
3064	#endif
3065	pPatch->ChangedIdt.u16SegSel = ASMGetCS();
3066	#ifdef RT_ARCH_AMD64
3067	pPatch->ChangedIdt.u3IST = 0;
3068	pPatch->ChangedIdt.u5Reserved = 0;
3069	#else /* x86 */
3070	pPatch->ChangedIdt.u5Reserved = 0;
3071	pPatch->ChangedIdt.u3Type1 = 0;
3072	#endif /* x86 */
3073	pPatch->ChangedIdt.u5Type2 = SUPDRV_IDTE_TYPE2_INTERRUPT_GATE;
3074	pPatch->ChangedIdt.u2DPL = 3;
3075	pPatch->ChangedIdt.u1Present = 1;
3076
3077	/*
3078	* Generate the patch code.
3079	*/
3080	{
3081	#ifdef RT_ARCH_AMD64
3082	union
3083	{
3084	uint8_t *pb;
3085	uint32_t *pu32;
3086	uint64_t *pu64;
3087	} u, uFixJmp, uFixCall, uNotNested;
3088	u.pb = &pPatch->auCode[0];
3089
3090	/* check the cookie */
3091	*u.pb++ = 0x3d; // cmp eax, GLOBALCOOKIE
3092	*u.pu32++ = pDevExt->u32Cookie;
3093
3094	*u.pb++ = 0x74; // jz @VBoxCall
3095	*u.pb++ = 2;
3096
3097	/* jump to forwarder code. */
3098	*u.pb++ = 0xeb;
3099	uFixJmp = u;
3100	*u.pb++ = 0xfe;
3101
3102	// @VBoxCall:
3103	*u.pb++ = 0x0f; // swapgs
3104	*u.pb++ = 0x01;
3105	*u.pb++ = 0xf8;
3106
3107	/*
3108	* Call VMMR0Entry
3109	* We don't have to push the arguments here, but we have top
3110	* reserve some stack space for the interrupt forwarding.
3111	*/
3112	# ifdef RT_OS_WINDOWS
3113	*u.pb++ = 0x50; // push rax ; alignment filler.
3114	*u.pb++ = 0x41; // push r8 ; uArg
3115	*u.pb++ = 0x50;
3116	*u.pb++ = 0x52; // push rdx ; uOperation
3117	*u.pb++ = 0x51; // push rcx ; pVM
3118	# else
3119	*u.pb++ = 0x51; // push rcx ; alignment filler.
3120	*u.pb++ = 0x52; // push rdx ; uArg
3121	*u.pb++ = 0x56; // push rsi ; uOperation
3122	*u.pb++ = 0x57; // push rdi ; pVM
3123	# endif
3124
3125	*u.pb++ = 0xff; // call qword [pfnVMMR0EntryInt wrt rip]
3126	*u.pb++ = 0x15;
3127	uFixCall = u;
3128	*u.pu32++ = 0;
3129
3130	*u.pb++ = 0x48; // add rsp, 20h ; remove call frame.
3131	*u.pb++ = 0x81;
3132	*u.pb++ = 0xc4;
3133	*u.pu32++ = 0x20;
3134
3135	*u.pb++ = 0x0f; // swapgs
3136	*u.pb++ = 0x01;
3137	*u.pb++ = 0xf8;
3138
3139	/* Return to R3. */
3140	uNotNested = u;
3141	*u.pb++ = 0x48; // iretq
3142	*u.pb++ = 0xcf;
3143
3144	while ((uintptr_t)u.pb & 0x7) // align 8
3145	*u.pb++ = 0xcc;
3146
3147	/* Pointer to the VMMR0Entry. */ // pfnVMMR0EntryInt dq StubVMMR0Entry
3148	*uFixCall.pu32 = (uint32_t)(u.pb - uFixCall.pb - 4); uFixCall.pb = NULL;
3149	pPatch->offVMMR0EntryFixup = (uint16_t)(u.pb - &pPatch->auCode[0]);
3150	*u.pu64++ = pDevExt->pvVMMR0 ? (uint64_t)pDevExt->pfnVMMR0EntryInt : (uint64_t)u.pb + 8;
3151
3152	/* stub entry. */ // StubVMMR0Entry:
3153	pPatch->offStub = (uint16_t)(u.pb - &pPatch->auCode[0]);
3154	*u.pb++ = 0x33; // xor eax, eax
3155	*u.pb++ = 0xc0;
3156
3157	*u.pb++ = 0x48; // dec rax
3158	*u.pb++ = 0xff;
3159	*u.pb++ = 0xc8;
3160
3161	*u.pb++ = 0xc3; // ret
3162
3163	/* forward to the original handler using a retf. */
3164	*uFixJmp.pb = (uint8_t)(u.pb - uFixJmp.pb - 1); uFixJmp.pb = NULL;
3165
3166	*u.pb++ = 0x68; // push <target cs>
3167	*u.pu32++ = !pPatch->SavedIdt.u5Type2 ? ASMGetCS() : pPatch->SavedIdt.u16SegSel;
3168
3169	*u.pb++ = 0x68; // push <low target rip>
3170	*u.pu32++ = !pPatch->SavedIdt.u5Type2
3171	? (uint32_t)(uintptr_t)uNotNested.pb
3172	: (uint32_t)pPatch->SavedIdt.u16OffsetLow
3173	\| (uint32_t)pPatch->SavedIdt.u16OffsetHigh << 16;
3174
3175	*u.pb++ = 0xc7; // mov dword [rsp + 4], <high target rip>
3176	*u.pb++ = 0x44;
3177	*u.pb++ = 0x24;
3178	*u.pb++ = 0x04;
3179	*u.pu32++ = !pPatch->SavedIdt.u5Type2
3180	? (uint32_t)((uint64_t)uNotNested.pb >> 32)
3181	: pPatch->SavedIdt.u32OffsetTop;
3182
3183	*u.pb++ = 0x48; // retf ; does this require prefix?
3184	*u.pb++ = 0xcb;
3185
3186	#else /* RT_ARCH_X86 */
3187
3188	union
3189	{
3190	uint8_t *pb;
3191	uint16_t *pu16;
3192	uint32_t *pu32;
3193	} u, uFixJmpNotNested, uFixJmp, uFixCall, uNotNested;
3194	u.pb = &pPatch->auCode[0];
3195
3196	/* check the cookie */
3197	*u.pb++ = 0x81; // cmp esi, GLOBALCOOKIE
3198	*u.pb++ = 0xfe;
3199	*u.pu32++ = pDevExt->u32Cookie;
3200
3201	*u.pb++ = 0x74; // jz VBoxCall
3202	uFixJmp = u;
3203	*u.pb++ = 0;
3204
3205	/* jump (far) to the original handler / not-nested-stub. */
3206	*u.pb++ = 0xea; // jmp far NotNested
3207	uFixJmpNotNested = u;
3208	*u.pu32++ = 0;
3209	*u.pu16++ = 0;
3210
3211	/* save selector registers. */ // VBoxCall:
3212	*uFixJmp.pb = (uint8_t)(u.pb - uFixJmp.pb - 1);
3213	*u.pb++ = 0x0f; // push fs
3214	*u.pb++ = 0xa0;
3215
3216	*u.pb++ = 0x1e; // push ds
3217
3218	*u.pb++ = 0x06; // push es
3219
3220	/* call frame */
3221	*u.pb++ = 0x51; // push ecx
3222
3223	*u.pb++ = 0x52; // push edx
3224
3225	*u.pb++ = 0x50; // push eax
3226
3227	/* load ds, es and perhaps fs before call. */
3228	*u.pb++ = 0xb8; // mov eax, KernelDS
3229	*u.pu32++ = ASMGetDS();
3230
3231	*u.pb++ = 0x8e; // mov ds, eax
3232	*u.pb++ = 0xd8;
3233
3234	*u.pb++ = 0x8e; // mov es, eax
3235	*u.pb++ = 0xc0;
3236
3237	#ifdef RT_OS_WINDOWS
3238	*u.pb++ = 0xb8; // mov eax, KernelFS
3239	*u.pu32++ = ASMGetFS();
3240
3241	*u.pb++ = 0x8e; // mov fs, eax
3242	*u.pb++ = 0xe0;
3243	#endif
3244
3245	/* do the call. */
3246	*u.pb++ = 0xe8; // call _VMMR0Entry / StubVMMR0Entry
3247	uFixCall = u;
3248	pPatch->offVMMR0EntryFixup = (uint16_t)(u.pb - &pPatch->auCode[0]);
3249	*u.pu32++ = 0xfffffffb;
3250
3251	*u.pb++ = 0x83; // add esp, 0ch ; cdecl
3252	*u.pb++ = 0xc4;
3253	*u.pb++ = 0x0c;
3254
3255	/* restore selector registers. */
3256	*u.pb++ = 0x07; // pop es
3257	//
3258	*u.pb++ = 0x1f; // pop ds
3259
3260	*u.pb++ = 0x0f; // pop fs
3261	*u.pb++ = 0xa1;
3262
3263	uNotNested = u; // NotNested:
3264	*u.pb++ = 0xcf; // iretd
3265
3266	/* the stub VMMR0Entry. */ // StubVMMR0Entry:
3267	pPatch->offStub = (uint16_t)(u.pb - &pPatch->auCode[0]);
3268	*u.pb++ = 0x33; // xor eax, eax
3269	*u.pb++ = 0xc0;
3270
3271	*u.pb++ = 0x48; // dec eax
3272
3273	*u.pb++ = 0xc3; // ret
3274
3275	/* Fixup the VMMR0Entry call. */
3276	if (pDevExt->pvVMMR0)
3277	*uFixCall.pu32 = (uint32_t)pDevExt->pfnVMMR0EntryInt - (uint32_t)(uFixCall.pu32 + 1);
3278	else
3279	*uFixCall.pu32 = (uint32_t)&pPatch->auCode[pPatch->offStub] - (uint32_t)(uFixCall.pu32 + 1);
3280
3281	/* Fixup the forward / nested far jump. */
3282	if (!pPatch->SavedIdt.u5Type2)
3283	{
3284	*uFixJmpNotNested.pu32++ = (uint32_t)uNotNested.pb;
3285	*uFixJmpNotNested.pu16++ = ASMGetCS();
3286	}
3287	else
3288	{
3289	*uFixJmpNotNested.pu32++ = ((uint32_t)pPatch->SavedIdt.u16OffsetHigh << 16) \| pPatch->SavedIdt.u16OffsetLow;
3290	*uFixJmpNotNested.pu16++ = pPatch->SavedIdt.u16SegSel;
3291	}
3292	#endif /* RT_ARCH_X86 */
3293	Assert(u.pb <= &pPatch->auCode[sizeof(pPatch->auCode)]);
3294	#if 0
3295	/* dump the patch code */
3296	Log2(("patch code: %p\n", &pPatch->auCode[0]));
3297	for (uFixCall.pb = &pPatch->auCode[0]; uFixCall.pb < u.pb; uFixCall.pb++)
3298	Log2(("0x%02x,\n", *uFixCall.pb));
3299	#endif
3300	}
3301
3302	/*
3303	* Install the patch.
3304	*/
3305	supdrvIdtWrite(pPatch->pIdtEntry, &pPatch->ChangedIdt);
3306	AssertMsg(!memcmp((void *)pPatch->pIdtEntry, &pPatch->ChangedIdt, sizeof(pPatch->ChangedIdt)), ("The stupid change code didn't work!!!!!\n"));
3307
3308	/*
3309	* Link in the patch.
3310	*/
3311	pPatch->pNext = pDevExt->pIdtPatches;
3312	pDevExt->pIdtPatches = pPatch;
3313
3314	return pPatch;
3315	}
3316
3317
3318	/**
3319	* Removes the sessions IDT references.
3320	* This will uninstall our IDT patch if we left unreferenced.
3321	*
3322	* @returns VINF_SUCCESS.
3323	* @param pDevExt Device globals.
3324	* @param pSession Session data.
3325	*/
3326	static int supdrvIOCtl_IdtRemoveAll(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
3327	{
3328	PSUPDRVPATCHUSAGE pUsage;
3329	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3330	LogFlow(("supdrvIOCtl_IdtRemoveAll: pSession=%p\n", pSession));
3331
3332	/*
3333	* Take the spinlock.
3334	*/
3335	RTSpinlockAcquireNoInts(pDevExt->Spinlock, &SpinlockTmp);
3336
3337	/*
3338	* Walk usage list, removing patches as their usage count reaches zero.
3339	*/
3340	pUsage = pSession->pPatchUsage;
3341	while (pUsage)
3342	{
3343	if (pUsage->pPatch->cUsage <= pUsage->cUsage)
3344	supdrvIdtRemoveOne(pDevExt, pUsage->pPatch);
3345	else
3346	pUsage->pPatch->cUsage -= pUsage->cUsage;
3347
3348	/* next */
3349	pUsage = pUsage->pNext;
3350	}
3351
3352	/*
3353	* Empty the usage chain and we're done inside the spinlock.
3354	*/
3355	pUsage = pSession->pPatchUsage;
3356	pSession->pPatchUsage = NULL;
3357
3358	RTSpinlockReleaseNoInts(pDevExt->Spinlock, &SpinlockTmp);
3359
3360	/*
3361	* Free usage entries.
3362	*/
3363	while (pUsage)
3364	{
3365	void *pvToFree = pUsage;
3366	pUsage->cUsage = 0;
3367	pUsage->pPatch = NULL;
3368	pUsage = pUsage->pNext;
3369	RTMemFree(pvToFree);
3370	}
3371
3372	return VINF_SUCCESS;
3373	}
3374
3375
3376	/**
3377	* Remove one patch.
3378	*
3379	* Worker for supdrvIOCtl_IdtRemoveAll.
3380	*
3381	* @param pDevExt Device globals.
3382	* @param pPatch Patch entry to remove.
3383	* @remark Caller must own SUPDRVDEVEXT::Spinlock!
3384	*/
3385	static void supdrvIdtRemoveOne(PSUPDRVDEVEXT pDevExt, PSUPDRVPATCH pPatch)
3386	{
3387	LogFlow(("supdrvIdtRemoveOne: pPatch=%p\n", pPatch));
3388
3389	pPatch->cUsage = 0;
3390
3391	/*
3392	* If the IDT entry was changed it have to kick around for ever!
3393	* This will be attempted freed again, perhaps next time we'll succeed :-)
3394	*/
3395	if (memcmp((void *)pPatch->pIdtEntry, &pPatch->ChangedIdt, sizeof(pPatch->ChangedIdt)))
3396	{
3397	AssertMsgFailed(("The hijacked IDT entry has CHANGED!!!\n"));
3398	return;
3399	}
3400
3401	/*
3402	* Unlink it.
3403	*/
3404	if (pDevExt->pIdtPatches != pPatch)
3405	{
3406	PSUPDRVPATCH pPatchPrev = pDevExt->pIdtPatches;
3407	while (pPatchPrev)
3408	{
3409	if (pPatchPrev->pNext == pPatch)
3410	{
3411	pPatchPrev->pNext = pPatch->pNext;
3412	break;
3413	}
3414	pPatchPrev = pPatchPrev->pNext;
3415	}
3416	Assert(!pPatchPrev);
3417	}
3418	else
3419	pDevExt->pIdtPatches = pPatch->pNext;
3420	pPatch->pNext = NULL;
3421
3422
3423	/*
3424	* Verify and restore the IDT.
3425	*/
3426	AssertMsg(!memcmp((void *)pPatch->pIdtEntry, &pPatch->ChangedIdt, sizeof(pPatch->ChangedIdt)), ("The hijacked IDT entry has CHANGED!!!\n"));
3427	supdrvIdtWrite(pPatch->pIdtEntry, &pPatch->SavedIdt);
3428	AssertMsg(!memcmp((void *)pPatch->pIdtEntry, &pPatch->SavedIdt, sizeof(pPatch->SavedIdt)), ("The hijacked IDT entry has CHANGED!!!\n"));
3429
3430	/*
3431	* Put it in the free list.
3432	* (This free list stuff is to calm my paranoia.)
3433	*/
3434	pPatch->pvIdt = NULL;
3435	pPatch->pIdtEntry = NULL;
3436
3437	pPatch->pNext = pDevExt->pIdtPatchesFree;
3438	pDevExt->pIdtPatchesFree = pPatch;
3439	}
3440
3441
3442	/**
3443	* Write to an IDT entry.
3444	*
3445	* @param pvIdtEntry Where to write.
3446	* @param pNewIDTEntry What to write.
3447	*/
3448	static void supdrvIdtWrite(volatile void pvIdtEntry, const SUPDRVIDTE pNewIDTEntry)
3449	{
3450	RTR0UINTREG uCR0;
3451	RTR0UINTREG uFlags;
3452
3453	/*
3454	* On SMP machines (P4 hyperthreading included) we must preform a
3455	* 64-bit locked write when updating the IDT entry.
3456	*
3457	* The F00F bugfix for linux (and probably other OSes) causes
3458	* the IDT to be pointing to an readonly mapping. We get around that
3459	* by temporarily turning of WP. Since we're inside a spinlock at this
3460	* point, interrupts are disabled and there isn't any way the WP bit
3461	* flipping can cause any trouble.
3462	*/
3463
3464	/* Save & Clear interrupt flag; Save & clear WP. */
3465	uFlags = ASMGetFlags();
3466	ASMSetFlags(uFlags & ~(RTR0UINTREG)(1 << 9)); /X86_EFL_IF/
3467	Assert(!(ASMGetFlags() & (1 << 9)));
3468	uCR0 = ASMGetCR0();
3469	ASMSetCR0(uCR0 & ~(RTR0UINTREG)(1 << 16)); /X86_CR0_WP/
3470
3471	/* Update IDT Entry */
3472	#ifdef RT_ARCH_AMD64
3473	ASMAtomicXchgU128((volatile uint128_t )pvIdtEntry, (uint128_t *)(uintptr_t)pNewIDTEntry);
3474	#else
3475	ASMAtomicXchgU64((volatile uint64_t )pvIdtEntry, (uint64_t *)(uintptr_t)pNewIDTEntry);
3476	#endif
3477
3478	/* Restore CR0 & Flags */
3479	ASMSetCR0(uCR0);
3480	ASMSetFlags(uFlags);
3481	}
3482	#endif /* VBOX_WITH_IDT_PATCHING */
3483
3484
3485	/**
3486	* Opens an image. If it's the first time it's opened the call must upload
3487	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
3488	*
3489	* This is the 1st step of the loading.
3490	*
3491	* @returns IPRT status code.
3492	* @param pDevExt Device globals.
3493	* @param pSession Session data.
3494	* @param pReq The open request.
3495	*/
3496	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
3497	{
3498	PSUPDRVLDRIMAGE pImage;
3499	unsigned cb;
3500	void *pv;
3501	LogFlow(("supdrvIOCtl_LdrOpen: szName=%s cbImage=%d\n", pReq->u.In.szName, pReq->u.In.cbImage));
3502
3503	/*
3504	* Check if we got an instance of the image already.
3505	*/
3506	RTSemFastMutexRequest(pDevExt->mtxLdr);
3507	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
3508	{
3509	if (!strcmp(pImage->szName, pReq->u.In.szName))
3510	{
3511	pImage->cUsage++;
3512	pReq->u.Out.pvImageBase = pImage->pvImage;
3513	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
3514	supdrvLdrAddUsage(pSession, pImage);
3515	RTSemFastMutexRelease(pDevExt->mtxLdr);
3516	return VINF_SUCCESS;
3517	}
3518	}
3519	/* (not found - add it!) */
3520
3521	/*
3522	* Allocate memory.
3523	*/
3524	cb = pReq->u.In.cbImage + sizeof(SUPDRVLDRIMAGE) + 31;
3525	pv = RTMemExecAlloc(cb);
3526	if (!pv)
3527	{
3528	RTSemFastMutexRelease(pDevExt->mtxLdr);
3529	Log(("supdrvIOCtl_LdrOpen: RTMemExecAlloc(%u) failed\n", cb));
3530	return VERR_NO_MEMORY;
3531	}
3532
3533	/*
3534	* Setup and link in the LDR stuff.
3535	*/
3536	pImage = (PSUPDRVLDRIMAGE)pv;
3537	pImage->pvImage = RT_ALIGN_P(pImage + 1, 32);
3538	pImage->cbImage = pReq->u.In.cbImage;
3539	pImage->pfnModuleInit = NULL;
3540	pImage->pfnModuleTerm = NULL;
3541	pImage->uState = SUP_IOCTL_LDR_OPEN;
3542	pImage->cUsage = 1;
3543	strcpy(pImage->szName, pReq->u.In.szName);
3544
3545	pImage->pNext = pDevExt->pLdrImages;
3546	pDevExt->pLdrImages = pImage;
3547
3548	supdrvLdrAddUsage(pSession, pImage);
3549
3550	pReq->u.Out.pvImageBase = pImage->pvImage;
3551	pReq->u.Out.fNeedsLoading = true;
3552	RTSemFastMutexRelease(pDevExt->mtxLdr);
3553	return VINF_SUCCESS;
3554	}
3555
3556
3557	/**
3558	* Loads the image bits.
3559	*
3560	* This is the 2nd step of the loading.
3561	*
3562	* @returns IPRT status code.
3563	* @param pDevExt Device globals.
3564	* @param pSession Session data.
3565	* @param pReq The request.
3566	*/
3567	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
3568	{
3569	PSUPDRVLDRUSAGE pUsage;
3570	PSUPDRVLDRIMAGE pImage;
3571	int rc;
3572	LogFlow(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImage=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImage));
3573
3574	/*
3575	* Find the ldr image.
3576	*/
3577	RTSemFastMutexRequest(pDevExt->mtxLdr);
3578	pUsage = pSession->pLdrUsage;
3579	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3580	pUsage = pUsage->pNext;
3581	if (!pUsage)
3582	{
3583	RTSemFastMutexRelease(pDevExt->mtxLdr);
3584	Log(("SUP_IOCTL_LDR_LOAD: couldn't find image!\n"));
3585	return VERR_INVALID_HANDLE;
3586	}
3587	pImage = pUsage->pImage;
3588	if (pImage->cbImage != pReq->u.In.cbImage)
3589	{
3590	RTSemFastMutexRelease(pDevExt->mtxLdr);
3591	Log(("SUP_IOCTL_LDR_LOAD: image size mismatch!! %d(prep) != %d(load)\n", pImage->cbImage, pReq->u.In.cbImage));
3592	return VERR_INVALID_HANDLE;
3593	}
3594	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
3595	{
3596	unsigned uState = pImage->uState;
3597	RTSemFastMutexRelease(pDevExt->mtxLdr);
3598	if (uState != SUP_IOCTL_LDR_LOAD)
3599	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
3600	return SUPDRV_ERR_ALREADY_LOADED;
3601	}
3602	switch (pReq->u.In.eEPType)
3603	{
3604	case SUPLDRLOADEP_NOTHING:
3605	break;
3606	case SUPLDRLOADEP_VMMR0:
3607	if ( !pReq->u.In.EP.VMMR0.pvVMMR0
3608	\|\| !pReq->u.In.EP.VMMR0.pvVMMR0EntryInt
3609	\|\| !pReq->u.In.EP.VMMR0.pvVMMR0EntryFast
3610	\|\| !pReq->u.In.EP.VMMR0.pvVMMR0EntryEx)
3611	{
3612	RTSemFastMutexRelease(pDevExt->mtxLdr);
3613	Log(("NULL pointer: pvVMMR0=%p pvVMMR0EntryInt=%p pvVMMR0EntryFast=%p pvVMMR0EntryEx=%p!\n",
3614	pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
3615	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx));
3616	return VERR_INVALID_PARAMETER;
3617	}
3618	if ( (uintptr_t)pReq->u.In.EP.VMMR0.pvVMMR0EntryInt - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage
3619	\|\| (uintptr_t)pReq->u.In.EP.VMMR0.pvVMMR0EntryFast - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage
3620	\|\| (uintptr_t)pReq->u.In.EP.VMMR0.pvVMMR0EntryEx - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage)
3621	{
3622	RTSemFastMutexRelease(pDevExt->mtxLdr);
3623	Log(("Out of range (%p LB %#x): pvVMMR0EntryInt=%p, pvVMMR0EntryFast=%p or pvVMMR0EntryEx=%p is NULL!\n",
3624	pImage->pvImage, pReq->u.In.cbImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
3625	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx));
3626	return VERR_INVALID_PARAMETER;
3627	}
3628	break;
3629	default:
3630	RTSemFastMutexRelease(pDevExt->mtxLdr);
3631	Log(("Invalid eEPType=%d\n", pReq->u.In.eEPType));
3632	return VERR_INVALID_PARAMETER;
3633	}
3634	if ( pReq->u.In.pfnModuleInit
3635	&& (uintptr_t)pReq->u.In.pfnModuleInit - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage)
3636	{
3637	RTSemFastMutexRelease(pDevExt->mtxLdr);
3638	Log(("SUP_IOCTL_LDR_LOAD: pfnModuleInit=%p is outside the image (%p %d bytes)\n",
3639	pReq->u.In.pfnModuleInit, pImage->pvImage, pReq->u.In.cbImage));
3640	return VERR_INVALID_PARAMETER;
3641	}
3642	if ( pReq->u.In.pfnModuleTerm
3643	&& (uintptr_t)pReq->u.In.pfnModuleTerm - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage)
3644	{
3645	RTSemFastMutexRelease(pDevExt->mtxLdr);
3646	Log(("SUP_IOCTL_LDR_LOAD: pfnModuleTerm=%p is outside the image (%p %d bytes)\n",
3647	pReq->u.In.pfnModuleTerm, pImage->pvImage, pReq->u.In.cbImage));
3648	return VERR_INVALID_PARAMETER;
3649	}
3650
3651	/*
3652	* Copy the memory.
3653	*/
3654	/* no need to do try/except as this is a buffered request. */
3655	memcpy(pImage->pvImage, &pReq->u.In.achImage[0], pImage->cbImage);
3656	pImage->uState = SUP_IOCTL_LDR_LOAD;
3657	pImage->pfnModuleInit = pReq->u.In.pfnModuleInit;
3658	pImage->pfnModuleTerm = pReq->u.In.pfnModuleTerm;
3659	pImage->offSymbols = pReq->u.In.offSymbols;
3660	pImage->cSymbols = pReq->u.In.cSymbols;
3661	pImage->offStrTab = pReq->u.In.offStrTab;
3662	pImage->cbStrTab = pReq->u.In.cbStrTab;
3663
3664	/*
3665	* Update any entry points.
3666	*/
3667	switch (pReq->u.In.eEPType)
3668	{
3669	default:
3670	case SUPLDRLOADEP_NOTHING:
3671	rc = VINF_SUCCESS;
3672	break;
3673	case SUPLDRLOADEP_VMMR0:
3674	rc = supdrvLdrSetR0EP(pDevExt, pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
3675	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
3676	break;
3677	}
3678
3679	/*
3680	* On success call the module initialization.
3681	*/
3682	LogFlow(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
3683	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
3684	{
3685	Log(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
3686	rc = pImage->pfnModuleInit();
3687	if (rc && pDevExt->pvVMMR0 == pImage->pvImage)
3688	supdrvLdrUnsetR0EP(pDevExt);
3689	}
3690
3691	if (rc)
3692	pImage->uState = SUP_IOCTL_LDR_OPEN;
3693
3694	RTSemFastMutexRelease(pDevExt->mtxLdr);
3695	return rc;
3696	}
3697
3698
3699	/**
3700	* Frees a previously loaded (prep'ed) image.
3701	*
3702	* @returns IPRT status code.
3703	* @param pDevExt Device globals.
3704	* @param pSession Session data.
3705	* @param pReq The request.
3706	*/
3707	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
3708	{
3709	int rc;
3710	PSUPDRVLDRUSAGE pUsagePrev;
3711	PSUPDRVLDRUSAGE pUsage;
3712	PSUPDRVLDRIMAGE pImage;
3713	LogFlow(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
3714
3715	/*
3716	* Find the ldr image.
3717	*/
3718	RTSemFastMutexRequest(pDevExt->mtxLdr);
3719	pUsagePrev = NULL;
3720	pUsage = pSession->pLdrUsage;
3721	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3722	{
3723	pUsagePrev = pUsage;
3724	pUsage = pUsage->pNext;
3725	}
3726	if (!pUsage)
3727	{
3728	RTSemFastMutexRelease(pDevExt->mtxLdr);
3729	Log(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
3730	return VERR_INVALID_HANDLE;
3731	}
3732
3733	/*
3734	* Check if we can remove anything.
3735	*/
3736	rc = VINF_SUCCESS;
3737	pImage = pUsage->pImage;
3738	if (pImage->cUsage <= 1 \|\| pUsage->cUsage <= 1)
3739	{
3740	/*
3741	* Check if there are any objects with destructors in the image, if
3742	* so leave it for the session cleanup routine so we get a chance to
3743	* clean things up in the right order and not leave them all dangling.
3744	*/
3745	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3746	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
3747	if (pImage->cUsage <= 1)
3748	{
3749	PSUPDRVOBJ pObj;
3750	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
3751	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImage))
3752	{
3753	rc = VERR_SHARING_VIOLATION; /** @todo VERR_DANGLING_OBJECTS */
3754	break;
3755	}
3756	}
3757	else
3758	{
3759	PSUPDRVUSAGE pGenUsage;
3760	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
3761	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImage))
3762	{
3763	rc = VERR_SHARING_VIOLATION; /** @todo VERR_DANGLING_OBJECTS */
3764	break;
3765	}
3766	}
3767	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
3768	if (rc == VINF_SUCCESS)
3769	{
3770	/* unlink it */
3771	if (pUsagePrev)
3772	pUsagePrev->pNext = pUsage->pNext;
3773	else
3774	pSession->pLdrUsage = pUsage->pNext;
3775
3776	/* free it */
3777	pUsage->pImage = NULL;
3778	pUsage->pNext = NULL;
3779	RTMemFree(pUsage);
3780
3781	/*
3782	* Derefrence the image.
3783	*/
3784	if (pImage->cUsage <= 1)
3785	supdrvLdrFree(pDevExt, pImage);
3786	else
3787	pImage->cUsage--;
3788	}
3789	else
3790	Log(("supdrvIOCtl_LdrFree: Dangling objects in %p/%s!\n", pImage->pvImage, pImage->szName));
3791	}
3792	else
3793	{
3794	/*
3795	* Dereference both image and usage.
3796	*/
3797	pImage->cUsage--;
3798	pUsage->cUsage--;
3799	}
3800
3801	RTSemFastMutexRelease(pDevExt->mtxLdr);
3802	return VINF_SUCCESS;
3803	}
3804
3805
3806	/**
3807	* Gets the address of a symbol in an open image.
3808	*
3809	* @returns 0 on success.
3810	* @returns SUPDRV_ERR_* on failure.
3811	* @param pDevExt Device globals.
3812	* @param pSession Session data.
3813	* @param pReq The request buffer.
3814	*/
3815	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
3816	{
3817	PSUPDRVLDRIMAGE pImage;
3818	PSUPDRVLDRUSAGE pUsage;
3819	uint32_t i;
3820	PSUPLDRSYM paSyms;
3821	const char *pchStrings;
3822	const size_t cbSymbol = strlen(pReq->u.In.szSymbol) + 1;
3823	void *pvSymbol = NULL;
3824	int rc = VERR_GENERAL_FAILURE;
3825	Log3(("supdrvIOCtl_LdrGetSymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
3826
3827	/*
3828	* Find the ldr image.
3829	*/
3830	RTSemFastMutexRequest(pDevExt->mtxLdr);
3831	pUsage = pSession->pLdrUsage;
3832	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3833	pUsage = pUsage->pNext;
3834	if (!pUsage)
3835	{
3836	RTSemFastMutexRelease(pDevExt->mtxLdr);
3837	Log(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
3838	return VERR_INVALID_HANDLE;
3839	}
3840	pImage = pUsage->pImage;
3841	if (pImage->uState != SUP_IOCTL_LDR_LOAD)
3842	{
3843	unsigned uState = pImage->uState;
3844	RTSemFastMutexRelease(pDevExt->mtxLdr);
3845	Log(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", uState, uState)); NOREF(uState);
3846	return VERR_ALREADY_LOADED;
3847	}
3848
3849	/*
3850	* Search the symbol strings.
3851	*/
3852	pchStrings = (const char )((uint8_t )pImage->pvImage + pImage->offStrTab);
3853	paSyms = (PSUPLDRSYM)((uint8_t *)pImage->pvImage + pImage->offSymbols);
3854	for (i = 0; i < pImage->cSymbols; i++)
3855	{
3856	if ( paSyms[i].offSymbol < pImage->cbImage /* paranoia */
3857	&& paSyms[i].offName + cbSymbol <= pImage->cbStrTab
3858	&& !memcmp(pchStrings + paSyms[i].offName, pReq->u.In.szSymbol, cbSymbol))
3859	{
3860	pvSymbol = (uint8_t *)pImage->pvImage + paSyms[i].offSymbol;
3861	rc = VINF_SUCCESS;
3862	break;
3863	}
3864	}
3865	RTSemFastMutexRelease(pDevExt->mtxLdr);
3866	pReq->u.Out.pvSymbol = pvSymbol;
3867	return rc;
3868	}
3869
3870
3871	/**
3872	* Gets the address of a symbol in an open image or the support driver.
3873	*
3874	* @returns VINF_SUCCESS on success.
3875	* @returns
3876	* @param pDevExt Device globals.
3877	* @param pSession Session data.
3878	* @param pReq The request buffer.
3879	*/
3880	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq)
3881	{
3882	int rc = VINF_SUCCESS;
3883	const char *pszSymbol = pReq->u.In.pszSymbol;
3884	const char *pszModule = pReq->u.In.pszModule;
3885	size_t cbSymbol;
3886	char const *pszEnd;
3887	uint32_t i;
3888
3889	/*
3890	* Input validation.
3891	*/
3892	AssertPtrReturn(pszSymbol, VERR_INVALID_POINTER);
3893	pszEnd = (char *)memchr(pszSymbol, '\0', 512);
3894	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3895	cbSymbol = pszEnd - pszSymbol + 1;
3896
3897	if (pszModule)
3898	{
3899	AssertPtrReturn(pszModule, VERR_INVALID_POINTER);
3900	pszEnd = (char *)memchr(pszModule, '\0', 64);
3901	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3902	}
3903	Log3(("supdrvIDC_LdrGetSymbol: pszModule=%p:{%s} pszSymbol=%p:{%s}\n", pszModule, pszModule, pszSymbol, pszSymbol));
3904
3905
3906	if ( !pszModule
3907	\|\| !strcmp(pszModule, "SupDrv"))
3908	{
3909	/*
3910	* Search the support driver export table.
3911	*/
3912	for (i = 0; i < RT_ELEMENTS(g_aFunctions); i++)
3913	if (!strcmp(g_aFunctions[i].szName, pszSymbol))
3914	{
3915	pReq->u.Out.pfnSymbol = g_aFunctions[i].pfn;
3916	break;
3917	}
3918	}
3919	else
3920	{
3921	/*
3922	* Find the loader image.
3923	*/
3924	PSUPDRVLDRIMAGE pImage;
3925
3926	RTSemFastMutexRequest(pDevExt->mtxLdr);
3927
3928	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
3929	if (!strcmp(pImage->szName, pszModule))
3930	break;
3931	if (pImage && pImage->uState == SUP_IOCTL_LDR_LOAD)
3932	{
3933	/*
3934	* Search the symbol strings.
3935	*/
3936	const char pchStrings = (const char )((uint8_t *)pImage->pvImage + pImage->offStrTab);
3937	PCSUPLDRSYM paSyms = (PCSUPLDRSYM)((uint8_t *)pImage->pvImage + pImage->offSymbols);
3938	for (i = 0; i < pImage->cSymbols; i++)
3939	{
3940	if ( paSyms[i].offSymbol < pImage->cbImage /* paranoia */
3941	&& paSyms[i].offName + cbSymbol <= pImage->cbStrTab
3942	&& !memcmp(pchStrings + paSyms[i].offName, pszSymbol, cbSymbol))
3943	{
3944	/*
3945	* Found it! Calc the symbol address and add a reference to the module.
3946	*/
3947	pReq->u.Out.pfnSymbol = (PFNRT)((uint8_t *)pImage->pvImage + paSyms[i].offSymbol);
3948	rc = supdrvLdrAddUsage(pSession, pImage);
3949	break;
3950	}
3951	}
3952	}
3953	else
3954	rc = pImage ? VERR_WRONG_ORDER : VERR_MODULE_NOT_FOUND;
3955
3956	RTSemFastMutexRelease(pDevExt->mtxLdr);
3957	}
3958	return rc;
3959	}
3960
3961
3962	/**
3963	* Updates the IDT patches to point to the specified VMM R0 entry
3964	* point (i.e. VMMR0Enter()).
3965	*
3966	* @returns IPRT status code.
3967	* @param pDevExt Device globals.
3968	* @param pSession Session data.
3969	* @param pVMMR0 VMMR0 image handle.
3970	* @param pvVMMR0EntryInt VMMR0EntryInt address.
3971	* @param pvVMMR0EntryFast VMMR0EntryFast address.
3972	* @param pvVMMR0EntryEx VMMR0EntryEx address.
3973	* @remark Caller must own the loader mutex.
3974	*/
3975	static int supdrvLdrSetR0EP(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx)
3976	{
3977	int rc = VINF_SUCCESS;
3978	LogFlow(("supdrvLdrSetR0EP pvVMMR0=%p pvVMMR0EntryInt=%p\n", pvVMMR0, pvVMMR0EntryInt));
3979
3980
3981	/*
3982	* Check if not yet set.
3983	*/
3984	if (!pDevExt->pvVMMR0)
3985	{
3986	#ifdef VBOX_WITH_IDT_PATCHING
3987	PSUPDRVPATCH pPatch;
3988	#endif
3989
3990	/*
3991	* Set it and update IDT patch code.
3992	*/
3993	pDevExt->pvVMMR0 = pvVMMR0;
3994	pDevExt->pfnVMMR0EntryInt = pvVMMR0EntryInt;
3995	pDevExt->pfnVMMR0EntryFast = pvVMMR0EntryFast;
3996	pDevExt->pfnVMMR0EntryEx = pvVMMR0EntryEx;
3997	#ifdef VBOX_WITH_IDT_PATCHING
3998	for (pPatch = pDevExt->pIdtPatches; pPatch; pPatch = pPatch->pNext)
3999	{
4000	# ifdef RT_ARCH_AMD64
4001	ASMAtomicXchgU64((volatile uint64_t *)&pPatch->auCode[pPatch->offVMMR0EntryFixup], (uint64_t)pvVMMR0);
4002	# else /* RT_ARCH_X86 */
4003	ASMAtomicXchgU32((volatile uint32_t *)&pPatch->auCode[pPatch->offVMMR0EntryFixup],
4004	(uint32_t)pvVMMR0 - (uint32_t)&pPatch->auCode[pPatch->offVMMR0EntryFixup + 4]);
4005	# endif
4006	}
4007	#endif /* VBOX_WITH_IDT_PATCHING */
4008	}
4009	else
4010	{
4011	/*
4012	* Return failure or success depending on whether the values match or not.
4013	*/
4014	if ( pDevExt->pvVMMR0 != pvVMMR0
4015	\|\| (void *)pDevExt->pfnVMMR0EntryInt != pvVMMR0EntryInt
4016	\|\| (void *)pDevExt->pfnVMMR0EntryFast != pvVMMR0EntryFast
4017	\|\| (void *)pDevExt->pfnVMMR0EntryEx != pvVMMR0EntryEx)
4018	{
4019	AssertMsgFailed(("SUP_IOCTL_LDR_SETR0EP: Already set pointing to a different module!\n"));
4020	rc = VERR_INVALID_PARAMETER;
4021	}
4022	}
4023	return rc;
4024	}
4025
4026
4027	/**
4028	* Unsets the R0 entry point installed by supdrvLdrSetR0EP.
4029	*
4030	* @param pDevExt Device globals.
4031	*/
4032	static void supdrvLdrUnsetR0EP(PSUPDRVDEVEXT pDevExt)
4033	{
4034	#ifdef VBOX_WITH_IDT_PATCHING
4035	PSUPDRVPATCH pPatch;
4036	#endif
4037
4038	pDevExt->pvVMMR0 = NULL;
4039	pDevExt->pfnVMMR0EntryInt = NULL;
4040	pDevExt->pfnVMMR0EntryFast = NULL;
4041	pDevExt->pfnVMMR0EntryEx = NULL;
4042
4043	#ifdef VBOX_WITH_IDT_PATCHING
4044	for (pPatch = pDevExt->pIdtPatches; pPatch; pPatch = pPatch->pNext)
4045	{
4046	# ifdef RT_ARCH_AMD64
4047	ASMAtomicXchgU64((volatile uint64_t *)&pPatch->auCode[pPatch->offVMMR0EntryFixup],
4048	(uint64_t)&pPatch->auCode[pPatch->offStub]);
4049	# else /* RT_ARCH_X86 */
4050	ASMAtomicXchgU32((volatile uint32_t *)&pPatch->auCode[pPatch->offVMMR0EntryFixup],
4051	(uint32_t)&pPatch->auCode[pPatch->offStub] - (uint32_t)&pPatch->auCode[pPatch->offVMMR0EntryFixup + 4]);
4052	# endif
4053	}
4054	#endif /* VBOX_WITH_IDT_PATCHING */
4055	}
4056
4057
4058	/**
4059	* Adds a usage reference in the specified session of an image.
4060	*
4061	* Called while owning the loader semaphore.
4062	*
4063	* @returns VINF_SUCCESS on success and VERR_NO_MEMORY on failure.
4064	* @param pSession Session in question.
4065	* @param pImage Image which the session is using.
4066	*/
4067	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage)
4068	{
4069	PSUPDRVLDRUSAGE pUsage;
4070	LogFlow(("supdrvLdrAddUsage: pImage=%p\n", pImage));
4071
4072	/*
4073	* Referenced it already?
4074	*/
4075	pUsage = pSession->pLdrUsage;
4076	while (pUsage)
4077	{
4078	if (pUsage->pImage == pImage)
4079	{
4080	pUsage->cUsage++;
4081	return VINF_SUCCESS;
4082	}
4083	pUsage = pUsage->pNext;
4084	}
4085
4086	/*
4087	* Allocate new usage record.
4088	*/
4089	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
4090	AssertReturn(pUsage, VERR_NO_MEMORY);
4091	pUsage->cUsage = 1;
4092	pUsage->pImage = pImage;
4093	pUsage->pNext = pSession->pLdrUsage;
4094	pSession->pLdrUsage = pUsage;
4095	return VINF_SUCCESS;
4096	}
4097
4098
4099	/**
4100	* Frees a load image.
4101	*
4102	* @param pDevExt Pointer to device extension.
4103	* @param pImage Pointer to the image we're gonna free.
4104	* This image must exit!
4105	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
4106	*/
4107	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
4108	{
4109	PSUPDRVLDRIMAGE pImagePrev;
4110	LogFlow(("supdrvLdrFree: pImage=%p\n", pImage));
4111
4112	/* find it - arg. should've used doubly linked list. */
4113	Assert(pDevExt->pLdrImages);
4114	pImagePrev = NULL;
4115	if (pDevExt->pLdrImages != pImage)
4116	{
4117	pImagePrev = pDevExt->pLdrImages;
4118	while (pImagePrev->pNext != pImage)
4119	pImagePrev = pImagePrev->pNext;
4120	Assert(pImagePrev->pNext == pImage);
4121	}
4122
4123	/* unlink */
4124	if (pImagePrev)
4125	pImagePrev->pNext = pImage->pNext;
4126	else
4127	pDevExt->pLdrImages = pImage->pNext;
4128
4129	/* check if this is VMMR0.r0 and fix the Idt patches if it is. */
4130	if (pDevExt->pvVMMR0 == pImage->pvImage)
4131	supdrvLdrUnsetR0EP(pDevExt);
4132
4133	/* check for objects with destructors in this image. (Shouldn't happen.) */
4134	if (pDevExt->pObjs)
4135	{
4136	unsigned cObjs = 0;
4137	PSUPDRVOBJ pObj;
4138	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
4139	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
4140	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4141	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImage))
4142	{
4143	pObj->pfnDestructor = NULL;
4144	cObjs++;
4145	}
4146	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
4147	if (cObjs)
4148	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
4149	}
4150
4151	/* call termination function if fully loaded. */
4152	if ( pImage->pfnModuleTerm
4153	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
4154	{
4155	LogFlow(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
4156	pImage->pfnModuleTerm();
4157	}
4158
4159	/* free the image */
4160	pImage->cUsage = 0;
4161	pImage->pNext = 0;
4162	pImage->uState = SUP_IOCTL_LDR_FREE;
4163	RTMemExecFree(pImage);
4164	}
4165
4166
4167	/**
4168	* Gets the current paging mode of the CPU and stores in in pOut.
4169	*/
4170	static SUPPAGINGMODE supdrvIOCtl_GetPagingMode(void)
4171	{
4172	SUPPAGINGMODE enmMode;
4173
4174	RTR0UINTREG cr0 = ASMGetCR0();
4175	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
4176	enmMode = SUPPAGINGMODE_INVALID;
4177	else
4178	{
4179	RTR0UINTREG cr4 = ASMGetCR4();
4180	uint32_t fNXEPlusLMA = 0;
4181	if (cr4 & X86_CR4_PAE)
4182	{
4183	uint32_t fAmdFeatures = ASMCpuId_EDX(0x80000001);
4184	if (fAmdFeatures & (X86_CPUID_AMD_FEATURE_EDX_NX \| X86_CPUID_AMD_FEATURE_EDX_LONG_MODE))
4185	{
4186	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
4187	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
4188	fNXEPlusLMA \|= RT_BIT(0);
4189	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
4190	fNXEPlusLMA \|= RT_BIT(1);
4191	}
4192	}
4193
4194	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
4195	{
4196	case 0:
4197	enmMode = SUPPAGINGMODE_32_BIT;
4198	break;
4199
4200	case X86_CR4_PGE:
4201	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
4202	break;
4203
4204	case X86_CR4_PAE:
4205	enmMode = SUPPAGINGMODE_PAE;
4206	break;
4207
4208	case X86_CR4_PAE \| RT_BIT(0):
4209	enmMode = SUPPAGINGMODE_PAE_NX;
4210	break;
4211
4212	case X86_CR4_PAE \| X86_CR4_PGE:
4213	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
4214	break;
4215
4216	case X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
4217	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
4218	break;
4219
4220	case RT_BIT(1) \| X86_CR4_PAE:
4221	enmMode = SUPPAGINGMODE_AMD64;
4222	break;
4223
4224	case RT_BIT(1) \| X86_CR4_PAE \| RT_BIT(0):
4225	enmMode = SUPPAGINGMODE_AMD64_NX;
4226	break;
4227
4228	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
4229	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
4230	break;
4231
4232	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
4233	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
4234	break;
4235
4236	default:
4237	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
4238	enmMode = SUPPAGINGMODE_INVALID;
4239	break;
4240	}
4241	}
4242	return enmMode;
4243	}
4244
4245
4246	/**
4247	* Creates the GIP.
4248	*
4249	* @returns negative errno.
4250	* @param pDevExt Instance data. GIP stuff may be updated.
4251	*/
4252	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
4253	{
4254	PSUPGLOBALINFOPAGE pGip;
4255	RTHCPHYS HCPhysGip;
4256	uint32_t u32SystemResolution;
4257	uint32_t u32Interval;
4258	int rc;
4259
4260	LogFlow(("supdrvGipCreate:\n"));
4261
4262	/* assert order */
4263	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
4264	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
4265	Assert(!pDevExt->pGipTimer);
4266
4267	/*
4268	* Allocate a suitable page with a default kernel mapping.
4269	*/
4270	rc = RTR0MemObjAllocLow(&pDevExt->GipMemObj, PAGE_SIZE, false);
4271	if (RT_FAILURE(rc))
4272	{
4273	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
4274	return rc;
4275	}
4276	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
4277	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
4278
4279	#if 0 /** @todo Disabled this as we didn't used to do it before and causes unnecessary stress on laptops.
4280	* It only applies to Windows and should probably revisited later, if possible made part of the
4281	* timer code (return min granularity in RTTimerGetSystemGranularity and set it in RTTimerStart). */
4282	/*
4283	* Try bump up the system timer resolution.
4284	* The more interrupts the better...
4285	*/
4286	if ( RT_SUCCESS(RTTimerRequestSystemGranularity( 488281 /* 2048 HZ */, &u32SystemResolution))
4287	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 500000 /* 2000 HZ */, &u32SystemResolution))
4288	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
4289	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
4290	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 1953125 /* 512 HZ */, &u32SystemResolution))
4291	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 2000000 /* 500 HZ */, &u32SystemResolution))
4292	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 3906250 /* 256 HZ */, &u32SystemResolution))
4293	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 4000000 /* 250 HZ */, &u32SystemResolution))
4294	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 7812500 /* 128 HZ */, &u32SystemResolution))
4295	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity(10000000 /* 100 HZ */, &u32SystemResolution))
4296	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity(15625000 /* 64 HZ */, &u32SystemResolution))
4297	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity(31250000 /* 32 HZ */, &u32SystemResolution))
4298	)
4299	{
4300	Assert(RTTimerGetSystemGranularity() <= u32SystemResolution);
4301	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
4302	}
4303	#endif
4304
4305	/*
4306	* Find a reasonable update interval and initialize the structure.
4307	*/
4308	u32Interval = u32SystemResolution = RTTimerGetSystemGranularity();
4309	while (u32Interval < 10000000 /* 10 ms */)
4310	u32Interval += u32SystemResolution;
4311
4312	supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), 1000000000 / u32Interval /=Hz/);
4313
4314	/*
4315	* Create the timer.
4316	* If CPU_ALL isn't supported we'll have to fall back to synchronous mode.
4317	*/
4318	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
4319	{
4320	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, RTTIMER_FLAGS_CPU_ALL, supdrvGipAsyncTimer, pDevExt);
4321	if (rc == VERR_NOT_SUPPORTED)
4322	{
4323	OSDBGPRINT(("supdrvGipCreate: omni timer not supported, falling back to synchronous mode\n"));
4324	pGip->u32Mode = SUPGIPMODE_SYNC_TSC;
4325	}
4326	}
4327	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
4328	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0, supdrvGipSyncTimer, pDevExt);
4329	if (RT_SUCCESS(rc))
4330	{
4331	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
4332	rc = RTMpNotificationRegister(supdrvGipMpEvent, pDevExt);
4333	if (RT_SUCCESS(rc))
4334	{
4335	/*
4336	* We're good.
4337	*/
4338	dprintf(("supdrvGipCreate: %ld ns interval.\n", (long)u32Interval));
4339	return VINF_SUCCESS;
4340	}
4341
4342	OSDBGPRINT(("supdrvGipCreate: failed register MP event notfication. rc=%d\n", rc));
4343	}
4344	else
4345	{
4346	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %ld ns interval. rc=%d\n", (long)u32Interval, rc));
4347	Assert(!pDevExt->pGipTimer);
4348	}
4349	supdrvGipDestroy(pDevExt);
4350	return rc;
4351	}
4352
4353
4354	/**
4355	* Terminates the GIP.
4356	*
4357	* @param pDevExt Instance data. GIP stuff may be updated.
4358	*/
4359	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
4360	{
4361	int rc;
4362	#ifdef DEBUG_DARWIN_GIP
4363	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
4364	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
4365	pDevExt->pGipTimer, pDevExt->GipMemObj));
4366	#endif
4367
4368	/*
4369	* Invalid the GIP data.
4370	*/
4371	if (pDevExt->pGip)
4372	{
4373	supdrvGipTerm(pDevExt->pGip);
4374	pDevExt->pGip = NULL;
4375	}
4376
4377	/*
4378	* Destroy the timer and free the GIP memory object.
4379	*/
4380	if (pDevExt->pGipTimer)
4381	{
4382	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
4383	pDevExt->pGipTimer = NULL;
4384	}
4385
4386	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
4387	{
4388	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
4389	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
4390	}
4391
4392	/*
4393	* Finally, release the system timer resolution request if one succeeded.
4394	*/
4395	if (pDevExt->u32SystemTimerGranularityGrant)
4396	{
4397	rc = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
4398	pDevExt->u32SystemTimerGranularityGrant = 0;
4399	}
4400	}
4401
4402
4403	/**
4404	* Timer callback function sync GIP mode.
4405	* @param pTimer The timer.
4406	* @param pvUser The device extension.
4407	*/
4408	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
4409	{
4410	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4411	supdrvGipUpdate(pDevExt->pGip, RTTimeSystemNanoTS());
4412	}
4413
4414
4415	/**
4416	* Timer callback function for async GIP mode.
4417	* @param pTimer The timer.
4418	* @param pvUser The device extension.
4419	*/
4420	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
4421	{
4422	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4423	RTCPUID idCpu = RTMpCpuId();
4424	uint64_t NanoTS = RTTimeSystemNanoTS();
4425
4426	/** @todo reset the transaction number and whatnot when iTick == 1. */
4427	if (pDevExt->idGipMaster == idCpu)
4428	supdrvGipUpdate(pDevExt->pGip, NanoTS);
4429	else
4430	supdrvGipUpdatePerCpu(pDevExt->pGip, NanoTS, ASMGetApicId());
4431	}
4432
4433
4434	/**
4435	* Multiprocessor event notification callback.
4436	*
4437	* This is used to make sue that the GIP master gets passed on to
4438	* another CPU.
4439	*
4440	* @param enmEvent The event.
4441	* @param idCpu The cpu it applies to.
4442	* @param pvUser Pointer to the device extension.
4443	*/
4444	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser)
4445	{
4446	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4447	if (enmEvent == RTMPEVENT_OFFLINE)
4448	{
4449	RTCPUID idGipMaster;
4450	ASMAtomicReadSize(&pDevExt->idGipMaster, &idGipMaster);
4451	if (idGipMaster == idCpu)
4452	{
4453	/*
4454	* Find a new GIP master.
4455	*/
4456	bool fIgnored;
4457	unsigned i;
4458	RTCPUID idNewGipMaster = NIL_RTCPUID;
4459	RTCPUSET OnlineCpus;
4460	RTMpGetOnlineSet(&OnlineCpus);
4461
4462	for (i = 0; i < RTCPUSET_MAX_CPUS; i++)
4463	{
4464	RTCPUID idCurCpu = RTMpCpuIdFromSetIndex(i);
4465	if ( RTCpuSetIsMember(&OnlineCpus, idCurCpu)
4466	&& idCurCpu != idGipMaster)
4467	{
4468	idNewGipMaster = idCurCpu;
4469	break;
4470	}
4471	}
4472
4473	dprintf(("supdrvGipMpEvent: Gip master %#lx -> %#lx\n", (long)idGipMaster, (long)idNewGipMaster));
4474	ASMAtomicCmpXchgSize(&pDevExt->idGipMaster, idNewGipMaster, idGipMaster, fIgnored);
4475	NOREF(fIgnored);
4476	}
4477	}
4478	}
4479
4480
4481	/**
4482	* Initializes the GIP data.
4483	*
4484	* @returns IPRT status code.
4485	* @param pDevExt Pointer to the device instance data.
4486	* @param pGip Pointer to the read-write kernel mapping of the GIP.
4487	* @param HCPhys The physical address of the GIP.
4488	* @param u64NanoTS The current nanosecond timestamp.
4489	* @param uUpdateHz The update freqence.
4490	*/
4491	int VBOXCALL supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys, uint64_t u64NanoTS, unsigned uUpdateHz)
4492	{
4493	unsigned i;
4494	#ifdef DEBUG_DARWIN_GIP
4495	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz));
4496	#else
4497	LogFlow(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz));
4498	#endif
4499
4500	/*
4501	* Initialize the structure.
4502	*/
4503	memset(pGip, 0, PAGE_SIZE);
4504	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
4505	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
4506	pGip->u32Mode = supdrvGipDeterminTscMode(pDevExt);
4507	pGip->u32UpdateHz = uUpdateHz;
4508	pGip->u32UpdateIntervalNS = 1000000000 / uUpdateHz;
4509	pGip->u64NanoTSLastUpdateHz = u64NanoTS;
4510
4511	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
4512	{
4513	pGip->aCPUs[i].u32TransactionId = 2;
4514	pGip->aCPUs[i].u64NanoTS = u64NanoTS;
4515	pGip->aCPUs[i].u64TSC = ASMReadTSC();
4516
4517	/*
4518	* We don't know the following values until we've executed updates.
4519	* So, we'll just insert very high values.
4520	*/
4521	pGip->aCPUs[i].u64CpuHz = _4G + 1;
4522	pGip->aCPUs[i].u32UpdateIntervalTSC = _2G / 4;
4523	pGip->aCPUs[i].au32TSCHistory[0] = _2G / 4;
4524	pGip->aCPUs[i].au32TSCHistory[1] = _2G / 4;
4525	pGip->aCPUs[i].au32TSCHistory[2] = _2G / 4;
4526	pGip->aCPUs[i].au32TSCHistory[3] = _2G / 4;
4527	pGip->aCPUs[i].au32TSCHistory[4] = _2G / 4;
4528	pGip->aCPUs[i].au32TSCHistory[5] = _2G / 4;
4529	pGip->aCPUs[i].au32TSCHistory[6] = _2G / 4;
4530	pGip->aCPUs[i].au32TSCHistory[7] = _2G / 4;
4531	}
4532
4533	/*
4534	* Link it to the device extension.
4535	*/
4536	pDevExt->pGip = pGip;
4537	pDevExt->HCPhysGip = HCPhys;
4538	pDevExt->cGipUsers = 0;
4539
4540	return VINF_SUCCESS;
4541	}
4542
4543
4544	/**
4545	* Callback used by supdrvDetermineAsyncTSC to read the TSC on a CPU.
4546	*
4547	* @param idCpu Ignored.
4548	* @param pvUser1 Where to put the TSC.
4549	* @param pvUser2 Ignored.
4550	*/
4551	static DECLCALLBACK(void) supdrvDetermineAsyncTscWorker(RTCPUID idCpu, void pvUser1, void pvUser2)
4552	{
4553	#if 1
4554	ASMAtomicWriteU64((uint64_t volatile *)pvUser1, ASMReadTSC());
4555	#else
4556	(uint64_t )pvUser1 = ASMReadTSC();
4557	#endif
4558	}
4559
4560
4561	/**
4562	* Determine if Async GIP mode is required because of TSC drift.
4563	*
4564	* When using the default/normal timer code it is essential that the time stamp counter
4565	* (TSC) runs never backwards, that is, a read operation to the counter should return
4566	* a bigger value than any previous read operation. This is guaranteed by the latest
4567	* AMD CPUs and by newer Intel CPUs which never enter the C2 state (P4). In any other
4568	* case we have to choose the asynchronous timer mode.
4569	*
4570	* @param poffMin Pointer to the determined difference between different cores.
4571	* @return false if the time stamp counters appear to be synchron, true otherwise.
4572	*/
4573	bool VBOXCALL supdrvDetermineAsyncTsc(uint64_t *poffMin)
4574	{
4575	/*
4576	* Just iterate all the cpus 8 times and make sure that the TSC is
4577	* ever increasing. We don't bother taking TSC rollover into account.
4578	*/
4579	RTCPUSET CpuSet;
4580	int iLastCpu = RTCpuLastIndex(RTMpGetSet(&CpuSet));
4581	int iCpu;
4582	int cLoops = 8;
4583	bool fAsync = false;
4584	int rc = VINF_SUCCESS;
4585	uint64_t offMax = 0;
4586	uint64_t offMin = ~(uint64_t)0;
4587	uint64_t PrevTsc = ASMReadTSC();
4588
4589	while (cLoops-- > 0)
4590	{
4591	for (iCpu = 0; iCpu <= iLastCpu; iCpu++)
4592	{
4593	uint64_t CurTsc;
4594	rc = RTMpOnSpecific(RTMpCpuIdFromSetIndex(iCpu), supdrvDetermineAsyncTscWorker, &CurTsc, NULL);
4595	if (RT_SUCCESS(rc))
4596	{
4597	if (CurTsc <= PrevTsc)
4598	{
4599	fAsync = true;
4600	offMin = offMax = PrevTsc - CurTsc;
4601	dprintf(("supdrvDetermineAsyncTsc: iCpu=%d cLoops=%d CurTsc=%llx PrevTsc=%llx\n",
4602	iCpu, cLoops, CurTsc, PrevTsc));
4603	break;
4604	}
4605
4606	/* Gather statistics (except the first time). */
4607	if (iCpu != 0 \|\| cLoops != 7)
4608	{
4609	uint64_t off = CurTsc - PrevTsc;
4610	if (off < offMin)
4611	offMin = off;
4612	if (off > offMax)
4613	offMax = off;
4614	dprintf2(("%d/%d: off=%llx\n", cLoops, iCpu, off));
4615	}
4616
4617	/* Next */
4618	PrevTsc = CurTsc;
4619	}
4620	else if (rc == VERR_NOT_SUPPORTED)
4621	break;
4622	else
4623	AssertMsg(rc == VERR_CPU_NOT_FOUND \|\| rc == VERR_CPU_OFFLINE, ("%d\n", rc));
4624	}
4625
4626	/* broke out of the loop. */
4627	if (iCpu <= iLastCpu)
4628	break;
4629	}
4630
4631	poffMin = offMin; / Almost RTMpOnSpecific profiling. */
4632	dprintf(("supdrvDetermineAsyncTsc: returns %d; iLastCpu=%d rc=%d offMin=%llx offMax=%llx\n",
4633	fAsync, iLastCpu, rc, offMin, offMax));
4634	#if !defined(RT_OS_SOLARIS) && !defined(RT_OS_OS2) && !defined(RT_OS_WINDOWS)
4635	OSDBGPRINT(("vboxdrv: fAsync=%d offMin=%#lx offMax=%#lx\n", fAsync, (long)offMin, (long)offMax));
4636	#endif
4637	return fAsync;
4638	}
4639
4640
4641	/**
4642	* Determin the GIP TSC mode.
4643	*
4644	* @returns The most suitable TSC mode.
4645	* @param pDevExt Pointer to the device instance data.
4646	*/
4647	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt)
4648	{
4649	/*
4650	* On SMP we're faced with two problems:
4651	* (1) There might be a skew between the CPU, so that cpu0
4652	* returns a TSC that is sligtly different from cpu1.
4653	* (2) Power management (and other things) may cause the TSC
4654	* to run at a non-constant speed, and cause the speed
4655	* to be different on the cpus. This will result in (1).
4656	*
4657	* So, on SMP systems we'll have to select the ASYNC update method
4658	* if there are symphoms of these problems.
4659	*/
4660	if (RTMpGetCount() > 1)
4661	{
4662	uint32_t uEAX, uEBX, uECX, uEDX;
4663	uint64_t u64DiffCoresIgnored;
4664
4665	/* Permit the user and/or the OS specfic bits to force async mode. */
4666	if (supdrvOSGetForcedAsyncTscMode(pDevExt))
4667	return SUPGIPMODE_ASYNC_TSC;
4668
4669	/* Try check for current differences between the cpus. */
4670	if (supdrvDetermineAsyncTsc(&u64DiffCoresIgnored))
4671	return SUPGIPMODE_ASYNC_TSC;
4672
4673	/*
4674	* If the CPU supports power management and is an AMD one we
4675	* won't trust it unless it has the TscInvariant bit is set.
4676	*/
4677	/* Check for "AuthenticAMD" */
4678	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
4679	if ( uEAX >= 1
4680	&& uEBX == X86_CPUID_VENDOR_AMD_EBX
4681	&& uECX == X86_CPUID_VENDOR_AMD_ECX
4682	&& uEDX == X86_CPUID_VENDOR_AMD_EDX)
4683	{
4684	/* Check for APM support and that TscInvariant is cleared. */
4685	ASMCpuId(0x80000000, &uEAX, &uEBX, &uECX, &uEDX);
4686	if (uEAX >= 0x80000007)
4687	{
4688	ASMCpuId(0x80000007, &uEAX, &uEBX, &uECX, &uEDX);
4689	if ( !(uEDX & RT_BIT(8))/* TscInvariant */
4690	&& (uEDX & 0x3e)) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
4691	return SUPGIPMODE_ASYNC_TSC;
4692	}
4693	}
4694	}
4695	return SUPGIPMODE_SYNC_TSC;
4696	}
4697
4698
4699	/**
4700	* Invalidates the GIP data upon termination.
4701	*
4702	* @param pGip Pointer to the read-write kernel mapping of the GIP.
4703	*/
4704	void VBOXCALL supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
4705	{
4706	unsigned i;
4707	pGip->u32Magic = 0;
4708	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
4709	{
4710	pGip->aCPUs[i].u64NanoTS = 0;
4711	pGip->aCPUs[i].u64TSC = 0;
4712	pGip->aCPUs[i].iTSCHistoryHead = 0;
4713	}
4714	}
4715
4716
4717	/**
4718	* Worker routine for supdrvGipUpdate and supdrvGipUpdatePerCpu that
4719	* updates all the per cpu data except the transaction id.
4720	*
4721	* @param pGip The GIP.
4722	* @param pGipCpu Pointer to the per cpu data.
4723	* @param u64NanoTS The current time stamp.
4724	*/
4725	static void supdrvGipDoUpdateCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
4726	{
4727	uint64_t u64TSC;
4728	uint64_t u64TSCDelta;
4729	uint32_t u32UpdateIntervalTSC;
4730	uint32_t u32UpdateIntervalTSCSlack;
4731	unsigned iTSCHistoryHead;
4732	uint64_t u64CpuHz;
4733
4734	/*
4735	* Update the NanoTS.
4736	*/
4737	ASMAtomicXchgU64(&pGipCpu->u64NanoTS, u64NanoTS);
4738
4739	/*
4740	* Calc TSC delta.
4741	*/
4742	/** @todo validate the NanoTS delta, don't trust the OS to call us when it should... */
4743	u64TSC = ASMReadTSC();
4744	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
4745	ASMAtomicXchgU64(&pGipCpu->u64TSC, u64TSC);
4746
4747	if (u64TSCDelta >> 32)
4748	{
4749	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
4750	pGipCpu->cErrors++;
4751	}
4752
4753	/*
4754	* TSC History.
4755	*/
4756	Assert(ELEMENTS(pGipCpu->au32TSCHistory) == 8);
4757
4758	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
4759	ASMAtomicXchgU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
4760	ASMAtomicXchgU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
4761
4762	/*
4763	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
4764	*/
4765	if (pGip->u32UpdateHz >= 1000)
4766	{
4767	uint32_t u32;
4768	u32 = pGipCpu->au32TSCHistory[0];
4769	u32 += pGipCpu->au32TSCHistory[1];
4770	u32 += pGipCpu->au32TSCHistory[2];
4771	u32 += pGipCpu->au32TSCHistory[3];
4772	u32 >>= 2;
4773	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
4774	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
4775	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
4776	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
4777	u32UpdateIntervalTSC >>= 2;
4778	u32UpdateIntervalTSC += u32;
4779	u32UpdateIntervalTSC >>= 1;
4780
4781	/* Value choosen for a 2GHz Athlon64 running linux 2.6.10/11, . */
4782	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
4783	}
4784	else if (pGip->u32UpdateHz >= 90)
4785	{
4786	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
4787	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
4788	u32UpdateIntervalTSC >>= 1;
4789
4790	/* value choosen on a 2GHz thinkpad running windows */
4791	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
4792	}
4793	else
4794	{
4795	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
4796
4797	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
4798	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
4799	}
4800	ASMAtomicXchgU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
4801
4802	/*
4803	* CpuHz.
4804	*/
4805	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, pGip->u32UpdateHz);
4806	ASMAtomicXchgU64(&pGipCpu->u64CpuHz, u64CpuHz);
4807	}
4808
4809
4810	/**
4811	* Updates the GIP.
4812	*
4813	* @param pGip Pointer to the GIP.
4814	* @param u64NanoTS The current nanosecond timesamp.
4815	*/
4816	void VBOXCALL supdrvGipUpdate(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS)
4817	{
4818	/*
4819	* Determin the relevant CPU data.
4820	*/
4821	PSUPGIPCPU pGipCpu;
4822	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
4823	pGipCpu = &pGip->aCPUs[0];
4824	else
4825	{
4826	unsigned iCpu = ASMGetApicId();
4827	if (RT_LIKELY(iCpu >= RT_ELEMENTS(pGip->aCPUs)))
4828	return;
4829	pGipCpu = &pGip->aCPUs[iCpu];
4830	}
4831
4832	/*
4833	* Start update transaction.
4834	*/
4835	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
4836	{
4837	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
4838	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
4839	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
4840	pGipCpu->cErrors++;
4841	return;
4842	}
4843
4844	/*
4845	* Recalc the update frequency every 0x800th time.
4846	*/
4847	if (!(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
4848	{
4849	if (pGip->u64NanoTSLastUpdateHz)
4850	{
4851	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
4852	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
4853	uint32_t u32UpdateHz = (uint32_t)((UINT64_C(1000000000) * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
4854	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
4855	{
4856	ASMAtomicXchgU32(&pGip->u32UpdateHz, u32UpdateHz);
4857	ASMAtomicXchgU32(&pGip->u32UpdateIntervalNS, 1000000000 / u32UpdateHz);
4858	}
4859	#endif
4860	}
4861	ASMAtomicXchgU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS);
4862	}
4863
4864	/*
4865	* Update the data.
4866	*/
4867	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS);
4868
4869	/*
4870	* Complete transaction.
4871	*/
4872	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
4873	}
4874
4875
4876	/**
4877	* Updates the per cpu GIP data for the calling cpu.
4878	*
4879	* @param pGip Pointer to the GIP.
4880	* @param u64NanoTS The current nanosecond timesamp.
4881	* @param iCpu The CPU index.
4882	*/
4883	void VBOXCALL supdrvGipUpdatePerCpu(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, unsigned iCpu)
4884	{
4885	PSUPGIPCPU pGipCpu;
4886
4887	if (RT_LIKELY(iCpu < RT_ELEMENTS(pGip->aCPUs)))
4888	{
4889	pGipCpu = &pGip->aCPUs[iCpu];
4890
4891	/*
4892	* Start update transaction.
4893	*/
4894	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
4895	{
4896	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
4897	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
4898	pGipCpu->cErrors++;
4899	return;
4900	}
4901
4902	/*
4903	* Update the data.
4904	*/
4905	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS);
4906
4907	/*
4908	* Complete transaction.
4909	*/
4910	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
4911	}
4912	}
4913
4914
4915	#ifndef SUPDRV_WITH_RELEASE_LOGGER
4916	# ifndef DEBUG /** @todo change #ifndef DEBUG -> #ifdef LOG_ENABLED */
4917	/**
4918	* Stub function for non-debug builds.
4919	*/
4920	RTDECL(PRTLOGGER) RTLogDefaultInstance(void)
4921	{
4922	return NULL;
4923	}
4924
4925	RTDECL(PRTLOGGER) RTLogRelDefaultInstance(void)
4926	{
4927	return NULL;
4928	}
4929
4930	/**
4931	* Stub function for non-debug builds.
4932	*/
4933	RTDECL(int) RTLogSetDefaultInstanceThread(PRTLOGGER pLogger, uintptr_t uKey)
4934	{
4935	return 0;
4936	}
4937
4938	/**
4939	* Stub function for non-debug builds.
4940	*/
4941	RTDECL(void) RTLogLogger(PRTLOGGER pLogger, void pvCallerRet, const char pszFormat, ...)
4942	{
4943	}
4944
4945	/**
4946	* Stub function for non-debug builds.
4947	*/
4948	RTDECL(void) RTLogLoggerEx(PRTLOGGER pLogger, unsigned fFlags, unsigned iGroup, const char *pszFormat, ...)
4949	{
4950	}
4951
4952	/**
4953	* Stub function for non-debug builds.
4954	*/
4955	RTDECL(void) RTLogLoggerExV(PRTLOGGER pLogger, unsigned fFlags, unsigned iGroup, const char *pszFormat, va_list args)
4956	{
4957	}
4958
4959	/**
4960	* Stub function for non-debug builds.
4961	*/
4962	RTDECL(void) RTLogPrintf(const char *pszFormat, ...)
4963	{
4964	}
4965
4966	/**
4967	* Stub function for non-debug builds.
4968	*/
4969	RTDECL(void) RTLogPrintfV(const char *pszFormat, va_list args)
4970	{
4971	}
4972	# endif /* !DEBUG */
4973	#endif /* !SUPDRV_WITH_RELEASE_LOGGER */

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

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDrv.c@ 10682

Download in other formats: