SUPDrv.c@ 12000

最後變更在這個檔案從12000是 12000,由 vboxsync 提交於 17 年前
supdrv: Prototyped the fix for the PspGetSetContextInternal / RtlVirtualUnwind crashes on Vista64. (disabled)
屬性 svn:eol-style 設為 `native` 屬性 svn:keywords 設為 `Author Date Id Revision`
檔案大小: 171.7 KB

行
1	/* $Revision: 12000 $ */
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.alldomusa.eu.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	#ifdef SUPDRV_WITH_UNWIND_HACK
187	{ "RTSemEventMultiWaitNoResume", (void *)supdrvNtWrapRTSemEventMultiWaitNoResume },
188	#else
189	{ "RTSemEventMultiWaitNoResume", (void *)RTSemEventMultiWaitNoResume },
190	#endif
191	{ "RTSemEventMultiDestroy", (void *)RTSemEventMultiDestroy },
192	{ "RTSpinlockCreate", (void *)RTSpinlockCreate },
193	{ "RTSpinlockDestroy", (void *)RTSpinlockDestroy },
194	{ "RTSpinlockAcquire", (void *)RTSpinlockAcquire },
195	{ "RTSpinlockRelease", (void *)RTSpinlockRelease },
196	{ "RTSpinlockAcquireNoInts", (void *)RTSpinlockAcquireNoInts },
197	{ "RTSpinlockReleaseNoInts", (void *)RTSpinlockReleaseNoInts },
198	{ "RTTimeNanoTS", (void *)RTTimeNanoTS },
199	{ "RTTimeMillieTS", (void *)RTTimeMilliTS },
200	{ "RTTimeSystemNanoTS", (void *)RTTimeSystemNanoTS },
201	{ "RTTimeSystemMillieTS", (void *)RTTimeSystemMilliTS },
202	{ "RTThreadNativeSelf", (void *)RTThreadNativeSelf },
203	{ "RTThreadSleep", (void *)RTThreadSleep },
204	{ "RTThreadYield", (void *)RTThreadYield },
205	#if 0 /* Thread APIs, Part 2. */
206	{ "RTThreadSelf", (void *)RTThreadSelf },
207	{ "RTThreadCreate", (void *)RTThreadCreate },
208	{ "RTThreadGetNative", (void *)RTThreadGetNative },
209	{ "RTThreadWait", (void *)RTThreadWait },
210	{ "RTThreadWaitNoResume", (void *)RTThreadWaitNoResume },
211	{ "RTThreadGetName", (void *)RTThreadGetName },
212	{ "RTThreadSelfName", (void *)RTThreadSelfName },
213	{ "RTThreadGetType", (void *)RTThreadGetType },
214	{ "RTThreadUserSignal", (void *)RTThreadUserSignal },
215	{ "RTThreadUserReset", (void *)RTThreadUserReset },
216	{ "RTThreadUserWait", (void *)RTThreadUserWait },
217	{ "RTThreadUserWaitNoResume", (void *)RTThreadUserWaitNoResume },
218	#endif
219	{ "RTLogDefaultInstance", (void *)RTLogDefaultInstance },
220	{ "RTMpCpuId", (void *)RTMpCpuId },
221	{ "RTMpCpuIdFromSetIndex", (void *)RTMpCpuIdFromSetIndex },
222	{ "RTMpCpuIdToSetIndex", (void *)RTMpCpuIdToSetIndex },
223	{ "RTMpIsCpuPossible", (void *)RTMpIsCpuPossible },
224	{ "RTMpGetCount", (void *)RTMpGetCount },
225	{ "RTMpGetMaxCpuId", (void *)RTMpGetMaxCpuId },
226	{ "RTMpGetOnlineCount", (void *)RTMpGetOnlineCount },
227	{ "RTMpGetOnlineSet", (void *)RTMpGetOnlineSet },
228	{ "RTMpGetSet", (void *)RTMpGetSet },
229	{ "RTMpIsCpuOnline", (void *)RTMpIsCpuOnline },
230	{ "RTMpOnAll", (void *)RTMpOnAll },
231	{ "RTMpOnOthers", (void *)RTMpOnOthers },
232	{ "RTMpOnSpecific", (void *)RTMpOnSpecific },
233	{ "RTLogRelDefaultInstance", (void *)RTLogRelDefaultInstance },
234	{ "RTLogSetDefaultInstanceThread", (void *)RTLogSetDefaultInstanceThread },
235	{ "RTLogLogger", (void *)RTLogLogger },
236	{ "RTLogLoggerEx", (void *)RTLogLoggerEx },
237	{ "RTLogLoggerExV", (void *)RTLogLoggerExV },
238	{ "RTLogPrintf", (void *)RTLogPrintf },
239	{ "RTLogPrintfV", (void *)RTLogPrintfV },
240	{ "AssertMsg1", (void *)AssertMsg1 },
241	{ "AssertMsg2", (void *)AssertMsg2 },
242	};
243
244
245	/*******************************************************************************
246	* Internal Functions *
247	*******************************************************************************/
248	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession);
249	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType);
250	#ifdef VBOX_WITH_IDT_PATCHING
251	static int supdrvIOCtl_IdtInstall(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPIDTINSTALL pReq);
252	static PSUPDRVPATCH supdrvIdtPatchOne(PSUPDRVDEVEXT pDevExt, PSUPDRVPATCH pPatch);
253	static int supdrvIOCtl_IdtRemoveAll(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession);
254	static void supdrvIdtRemoveOne(PSUPDRVDEVEXT pDevExt, PSUPDRVPATCH pPatch);
255	static void supdrvIdtWrite(volatile void pvIdtEntry, const SUPDRVIDTE pNewIDTEntry);
256	#endif /* VBOX_WITH_IDT_PATCHING */
257	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq);
258	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq);
259	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq);
260	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq);
261	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq);
262	static int supdrvLdrSetR0EP(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx);
263	static void supdrvLdrUnsetR0EP(PSUPDRVDEVEXT pDevExt);
264	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage);
265	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage);
266	static SUPPAGINGMODE supdrvIOCtl_GetPagingMode(void);
267	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt);
268	#ifdef RT_OS_WINDOWS
269	static int supdrvPageGetPhys(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages);
270	static bool supdrvPageWasLockedByPageAlloc(PSUPDRVSESSION pSession, RTR3PTR pvR3);
271	#endif /* RT_OS_WINDOWS */
272	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt);
273	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt);
274	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
275	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
276	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser);
277
278
279	/**
280	* Initializes the device extentsion structure.
281	*
282	* @returns IPRT status code.
283	* @param pDevExt The device extension to initialize.
284	*/
285	int VBOXCALL supdrvInitDevExt(PSUPDRVDEVEXT pDevExt)
286	{
287	int rc;
288
289	#ifdef SUPDRV_WITH_RELEASE_LOGGER
290	/*
291	* Create the release log.
292	*/
293	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
294	PRTLOGGER pRelLogger;
295	rc = RTLogCreate(&pRelLogger, 0 /* fFlags */, "all",
296	"VBOX_RELEASE_LOG", RT_ELEMENTS(s_apszGroups), s_apszGroups,
297	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER, NULL);
298	if (RT_SUCCESS(rc))
299	RTLogRelSetDefaultInstance(pRelLogger);
300	#endif
301
302	/*
303	* Initialize it.
304	*/
305	memset(pDevExt, 0, sizeof(*pDevExt));
306	rc = RTSpinlockCreate(&pDevExt->Spinlock);
307	if (!rc)
308	{
309	rc = RTSemFastMutexCreate(&pDevExt->mtxLdr);
310	if (!rc)
311	{
312	rc = RTSemFastMutexCreate(&pDevExt->mtxComponentFactory);
313	if (!rc)
314	{
315	rc = RTSemFastMutexCreate(&pDevExt->mtxGip);
316	if (!rc)
317	{
318	rc = supdrvGipCreate(pDevExt);
319	if (RT_SUCCESS(rc))
320	{
321	pDevExt->u32Cookie = BIRD; /** @todo make this random? */
322	return VINF_SUCCESS;
323	}
324
325	RTSemFastMutexDestroy(pDevExt->mtxGip);
326	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
327	}
328	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
329	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
330	}
331	RTSemFastMutexDestroy(pDevExt->mtxLdr);
332	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
333	}
334	RTSpinlockDestroy(pDevExt->Spinlock);
335	pDevExt->Spinlock = NIL_RTSPINLOCK;
336	}
337	#ifdef SUPDRV_WITH_RELEASE_LOGGER
338	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
339	RTLogDestroy(RTLogSetDefaultInstance(NULL));
340	#endif
341
342	return rc;
343	}
344
345
346	/**
347	* Delete the device extension (e.g. cleanup members).
348	*
349	* @param pDevExt The device extension to delete.
350	*/
351	void VBOXCALL supdrvDeleteDevExt(PSUPDRVDEVEXT pDevExt)
352	{
353	#ifdef VBOX_WITH_IDT_PATCHING
354	PSUPDRVPATCH pPatch;
355	#endif
356	PSUPDRVOBJ pObj;
357	PSUPDRVUSAGE pUsage;
358
359	/*
360	* Kill mutexes and spinlocks.
361	*/
362	RTSemFastMutexDestroy(pDevExt->mtxGip);
363	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
364	RTSemFastMutexDestroy(pDevExt->mtxLdr);
365	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
366	RTSpinlockDestroy(pDevExt->Spinlock);
367	pDevExt->Spinlock = NIL_RTSPINLOCK;
368
369	/*
370	* Free lists.
371	*/
372	#ifdef VBOX_WITH_IDT_PATCHING
373	/* patches */
374	/** @todo make sure we don't uninstall patches which has been patched by someone else. */
375	pPatch = pDevExt->pIdtPatchesFree;
376	pDevExt->pIdtPatchesFree = NULL;
377	while (pPatch)
378	{
379	void *pvFree = pPatch;
380	pPatch = pPatch->pNext;
381	RTMemExecFree(pvFree);
382	}
383	#endif /* VBOX_WITH_IDT_PATCHING */
384
385	/* objects. */
386	pObj = pDevExt->pObjs;
387	#if !defined(DEBUG_bird) \|\| !defined(RT_OS_LINUX) /* breaks unloading, temporary, remove me! */
388	Assert(!pObj); /* (can trigger on forced unloads) */
389	#endif
390	pDevExt->pObjs = NULL;
391	while (pObj)
392	{
393	void *pvFree = pObj;
394	pObj = pObj->pNext;
395	RTMemFree(pvFree);
396	}
397
398	/* usage records. */
399	pUsage = pDevExt->pUsageFree;
400	pDevExt->pUsageFree = NULL;
401	while (pUsage)
402	{
403	void *pvFree = pUsage;
404	pUsage = pUsage->pNext;
405	RTMemFree(pvFree);
406	}
407
408	/* kill the GIP. */
409	supdrvGipDestroy(pDevExt);
410
411	#ifdef SUPDRV_WITH_RELEASE_LOGGER
412	/* destroy the loggers. */
413	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
414	RTLogDestroy(RTLogSetDefaultInstance(NULL));
415	#endif
416	}
417
418
419	/**
420	* Create session.
421	*
422	* @returns IPRT status code.
423	* @param pDevExt Device extension.
424	* @param fUser Flag indicating whether this is a user or kernel session.
425	* @param ppSession Where to store the pointer to the session data.
426	*/
427	int VBOXCALL supdrvCreateSession(PSUPDRVDEVEXT pDevExt, bool fUser, PSUPDRVSESSION *ppSession)
428	{
429	/*
430	* Allocate memory for the session data.
431	*/
432	int rc = VERR_NO_MEMORY;
433	PSUPDRVSESSION pSession = ppSession = (PSUPDRVSESSION)RTMemAllocZ(sizeof(pSession));
434	if (pSession)
435	{
436	/* Initialize session data. */
437	rc = RTSpinlockCreate(&pSession->Spinlock);
438	if (!rc)
439	{
440	Assert(pSession->Spinlock != NIL_RTSPINLOCK);
441	pSession->pDevExt = pDevExt;
442	pSession->u32Cookie = BIRD_INV;
443	/*pSession->pLdrUsage = NULL;
444	pSession->pPatchUsage = NULL;
445	pSession->pVM = NULL;
446	pSession->pUsage = NULL;
447	pSession->pGip = NULL;
448	pSession->fGipReferenced = false;
449	pSession->Bundle.cUsed = 0; */
450	pSession->Uid = NIL_RTUID;
451	pSession->Gid = NIL_RTGID;
452	if (fUser)
453	{
454	pSession->Process = RTProcSelf();
455	pSession->R0Process = RTR0ProcHandleSelf();
456	}
457	else
458	{
459	pSession->Process = NIL_RTPROCESS;
460	pSession->R0Process = NIL_RTR0PROCESS;
461	}
462
463	LogFlow(("Created session %p initial cookie=%#x\n", pSession, pSession->u32Cookie));
464	return VINF_SUCCESS;
465	}
466
467	RTMemFree(pSession);
468	*ppSession = NULL;
469	Log(("Failed to create spinlock, rc=%d!\n", rc));
470	}
471
472	return rc;
473	}
474
475
476	/**
477	* Shared code for cleaning up a session.
478	*
479	* @param pDevExt Device extension.
480	* @param pSession Session data.
481	* This data will be freed by this routine.
482	*/
483	void VBOXCALL supdrvCloseSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
484	{
485	/*
486	* Cleanup the session first.
487	*/
488	supdrvCleanupSession(pDevExt, pSession);
489
490	/*
491	* Free the rest of the session stuff.
492	*/
493	RTSpinlockDestroy(pSession->Spinlock);
494	pSession->Spinlock = NIL_RTSPINLOCK;
495	pSession->pDevExt = NULL;
496	RTMemFree(pSession);
497	LogFlow(("supdrvCloseSession: returns\n"));
498	}
499
500
501	/**
502	* Shared code for cleaning up a session (but not quite freeing it).
503	*
504	* This is primarily intended for MAC OS X where we have to clean up the memory
505	* stuff before the file handle is closed.
506	*
507	* @param pDevExt Device extension.
508	* @param pSession Session data.
509	* This data will be freed by this routine.
510	*/
511	void VBOXCALL supdrvCleanupSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
512	{
513	PSUPDRVBUNDLE pBundle;
514	LogFlow(("supdrvCleanupSession: pSession=%p\n", pSession));
515
516	/*
517	* Remove logger instances related to this session.
518	*/
519	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pSession);
520
521	#ifdef VBOX_WITH_IDT_PATCHING
522	/*
523	* Uninstall any IDT patches installed for this session.
524	*/
525	supdrvIOCtl_IdtRemoveAll(pDevExt, pSession);
526	#endif
527
528	/*
529	* Release object references made in this session.
530	* In theory there should be noone racing us in this session.
531	*/
532	Log2(("release objects - start\n"));
533	if (pSession->pUsage)
534	{
535	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
536	PSUPDRVUSAGE pUsage;
537	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
538
539	while ((pUsage = pSession->pUsage) != NULL)
540	{
541	PSUPDRVOBJ pObj = pUsage->pObj;
542	pSession->pUsage = pUsage->pNext;
543
544	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
545	if (pUsage->cUsage < pObj->cUsage)
546	{
547	pObj->cUsage -= pUsage->cUsage;
548	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
549	}
550	else
551	{
552	/* Destroy the object and free the record. */
553	if (pDevExt->pObjs == pObj)
554	pDevExt->pObjs = pObj->pNext;
555	else
556	{
557	PSUPDRVOBJ pObjPrev;
558	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
559	if (pObjPrev->pNext == pObj)
560	{
561	pObjPrev->pNext = pObj->pNext;
562	break;
563	}
564	Assert(pObjPrev);
565	}
566	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
567
568	Log(("supdrvCleanupSession: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
569	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
570	if (pObj->pfnDestructor)
571	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
572	RTMemFree(pObj);
573	}
574
575	/* free it and continue. */
576	RTMemFree(pUsage);
577
578	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
579	}
580
581	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
582	AssertMsg(!pSession->pUsage, ("Some buster reregistered an object during desturction!\n"));
583	}
584	Log2(("release objects - done\n"));
585
586	/*
587	* Release memory allocated in the session.
588	*
589	* We do not serialize this as we assume that the application will
590	* not allocated memory while closing the file handle object.
591	*/
592	Log2(("freeing memory:\n"));
593	pBundle = &pSession->Bundle;
594	while (pBundle)
595	{
596	PSUPDRVBUNDLE pToFree;
597	unsigned i;
598
599	/*
600	* Check and unlock all entries in the bundle.
601	*/
602	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
603	{
604	if (pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ)
605	{
606	int rc;
607	Log2(("eType=%d pvR0=%p pvR3=%p cb=%ld\n", pBundle->aMem[i].eType, RTR0MemObjAddress(pBundle->aMem[i].MemObj),
608	(void *)RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3), (long)RTR0MemObjSize(pBundle->aMem[i].MemObj)));
609	if (pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ)
610	{
611	rc = RTR0MemObjFree(pBundle->aMem[i].MapObjR3, false);
612	AssertRC(rc); /** @todo figure out how to handle this. */
613	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
614	}
615	rc = RTR0MemObjFree(pBundle->aMem[i].MemObj, false);
616	AssertRC(rc); /** @todo figure out how to handle this. */
617	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
618	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
619	}
620	}
621
622	/*
623	* Advance and free previous bundle.
624	*/
625	pToFree = pBundle;
626	pBundle = pBundle->pNext;
627
628	pToFree->pNext = NULL;
629	pToFree->cUsed = 0;
630	if (pToFree != &pSession->Bundle)
631	RTMemFree(pToFree);
632	}
633	Log2(("freeing memory - done\n"));
634
635	/*
636	* Deregister component factories.
637	*/
638	RTSemFastMutexRequest(pDevExt->mtxComponentFactory);
639	Log2(("deregistering component factories:\n"));
640	if (pDevExt->pComponentFactoryHead)
641	{
642	PSUPDRVFACTORYREG pPrev = NULL;
643	PSUPDRVFACTORYREG pCur = pDevExt->pComponentFactoryHead;
644	while (pCur)
645	{
646	if (pCur->pSession == pSession)
647	{
648	/* unlink it */
649	PSUPDRVFACTORYREG pNext = pCur->pNext;
650	if (pPrev)
651	pPrev->pNext = pNext;
652	else
653	pDevExt->pComponentFactoryHead = pNext;
654
655	/* free it */
656	pCur->pNext = NULL;
657	pCur->pSession = NULL;
658	pCur->pFactory = NULL;
659	RTMemFree(pCur);
660
661	/* next */
662	pCur = pNext;
663	}
664	else
665	{
666	/* next */
667	pPrev = pCur;
668	pCur = pCur->pNext;
669	}
670	}
671	}
672	RTSemFastMutexRelease(pDevExt->mtxComponentFactory);
673	Log2(("deregistering component factories - done\n"));
674
675	/*
676	* Loaded images needs to be dereferenced and possibly freed up.
677	*/
678	RTSemFastMutexRequest(pDevExt->mtxLdr);
679	Log2(("freeing images:\n"));
680	if (pSession->pLdrUsage)
681	{
682	PSUPDRVLDRUSAGE pUsage = pSession->pLdrUsage;
683	pSession->pLdrUsage = NULL;
684	while (pUsage)
685	{
686	void *pvFree = pUsage;
687	PSUPDRVLDRIMAGE pImage = pUsage->pImage;
688	if (pImage->cUsage > pUsage->cUsage)
689	pImage->cUsage -= pUsage->cUsage;
690	else
691	supdrvLdrFree(pDevExt, pImage);
692	pUsage->pImage = NULL;
693	pUsage = pUsage->pNext;
694	RTMemFree(pvFree);
695	}
696	}
697	RTSemFastMutexRelease(pDevExt->mtxLdr);
698	Log2(("freeing images - done\n"));
699
700	/*
701	* Unmap the GIP.
702	*/
703	Log2(("umapping GIP:\n"));
704	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
705	{
706	SUPR0GipUnmap(pSession);
707	pSession->fGipReferenced = 0;
708	}
709	Log2(("umapping GIP - done\n"));
710	}
711
712
713	/**
714	* Fast path I/O Control worker.
715	*
716	* @returns VBox status code that should be passed down to ring-3 unchanged.
717	* @param uIOCtl Function number.
718	* @param pDevExt Device extention.
719	* @param pSession Session data.
720	*/
721	int VBOXCALL supdrvIOCtlFast(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
722	{
723	/*
724	* We check the two prereqs after doing this only to allow the compiler to optimize things better.
725	*/
726	if (RT_LIKELY(pSession->pVM && pDevExt->pfnVMMR0EntryFast))
727	{
728	switch (uIOCtl)
729	{
730	case SUP_IOCTL_FAST_DO_RAW_RUN:
731	pDevExt->pfnVMMR0EntryFast(pSession->pVM, SUP_VMMR0_DO_RAW_RUN);
732	break;
733	case SUP_IOCTL_FAST_DO_HWACC_RUN:
734	pDevExt->pfnVMMR0EntryFast(pSession->pVM, SUP_VMMR0_DO_HWACC_RUN);
735	break;
736	case SUP_IOCTL_FAST_DO_NOP:
737	pDevExt->pfnVMMR0EntryFast(pSession->pVM, SUP_VMMR0_DO_NOP);
738	break;
739	default:
740	return VERR_INTERNAL_ERROR;
741	}
742	return VINF_SUCCESS;
743	}
744	return VERR_INTERNAL_ERROR;
745	}
746
747
748	/**
749	* Helper for supdrvIOCtl. Check if pszStr contains any character of pszChars.
750	* We would use strpbrk here if this function would be contained in the RedHat kABI white
751	* list, see http://www.kerneldrivers.org/RHEL5.
752	*
753	* @return 1 if pszStr does contain any character of pszChars, 0 otherwise.
754	* @param pszStr String to check
755	* @param pszChars Character set
756	*/
757	static int supdrvCheckInvalidChar(const char pszStr, const char pszChars)
758	{
759	int chCur;
760	while ((chCur = *pszStr++) != '\0')
761	{
762	int ch;
763	const char *psz = pszChars;
764	while ((ch = *psz++) != '\0')
765	if (ch == chCur)
766	return 1;
767
768	}
769	return 0;
770	}
771
772
773	/**
774	* I/O Control worker.
775	*
776	* @returns 0 on success.
777	* @returns VERR_INVALID_PARAMETER if the request is invalid.
778	*
779	* @param uIOCtl Function number.
780	* @param pDevExt Device extention.
781	* @param pSession Session data.
782	* @param pReqHdr The request header.
783	*/
784	int VBOXCALL supdrvIOCtl(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
785	{
786	/*
787	* Validate the request.
788	*/
789	/* this first check could probably be omitted as its also done by the OS specific code... */
790	if (RT_UNLIKELY( (pReqHdr->fFlags & SUPREQHDR_FLAGS_MAGIC_MASK) != SUPREQHDR_FLAGS_MAGIC
791	\|\| pReqHdr->cbIn < sizeof(*pReqHdr)
792	\|\| pReqHdr->cbOut < sizeof(*pReqHdr)))
793	{
794	OSDBGPRINT(("vboxdrv: Bad ioctl request header; cbIn=%#lx cbOut=%#lx fFlags=%#lx\n",
795	(long)pReqHdr->cbIn, (long)pReqHdr->cbOut, (long)pReqHdr->fFlags));
796	return VERR_INVALID_PARAMETER;
797	}
798	if (RT_UNLIKELY(uIOCtl == SUP_IOCTL_COOKIE))
799	{
800	if (pReqHdr->u32Cookie != SUPCOOKIE_INITIAL_COOKIE)
801	{
802	OSDBGPRINT(("SUP_IOCTL_COOKIE: bad cookie %#lx\n", (long)pReqHdr->u32Cookie));
803	return VERR_INVALID_PARAMETER;
804	}
805	}
806	else if (RT_UNLIKELY( pReqHdr->u32Cookie != pDevExt->u32Cookie
807	\|\| pReqHdr->u32SessionCookie != pSession->u32Cookie))
808	{
809	OSDBGPRINT(("vboxdrv: bad cookie %#lx / %#lx.\n", (long)pReqHdr->u32Cookie, (long)pReqHdr->u32SessionCookie));
810	return VERR_INVALID_PARAMETER;
811	}
812
813	/*
814	* Validation macros
815	*/
816	#define REQ_CHECK_SIZES_EX(Name, cbInExpect, cbOutExpect) \
817	do { \
818	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect) \|\| pReqHdr->cbOut != (cbOutExpect))) \
819	{ \
820	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld. cbOut=%ld expected %ld.\n", \
821	(long)pReq->Hdr.cbIn, (long)(cbInExpect), (long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
822	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
823	} \
824	} while (0)
825
826	#define REQ_CHECK_SIZES(Name) REQ_CHECK_SIZES_EX(Name, Name ## _SIZE_IN, Name ## _SIZE_OUT)
827
828	#define REQ_CHECK_SIZE_IN(Name, cbInExpect) \
829	do { \
830	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect))) \
831	{ \
832	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld.\n", \
833	(long)pReq->Hdr.cbIn, (long)(cbInExpect))); \
834	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
835	} \
836	} while (0)
837
838	#define REQ_CHECK_SIZE_OUT(Name, cbOutExpect) \
839	do { \
840	if (RT_UNLIKELY(pReqHdr->cbOut != (cbOutExpect))) \
841	{ \
842	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbOut=%ld expected %ld.\n", \
843	(long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
844	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
845	} \
846	} while (0)
847
848	#define REQ_CHECK_EXPR(Name, expr) \
849	do { \
850	if (RT_UNLIKELY(!(expr))) \
851	{ \
852	OSDBGPRINT(( #Name ": %s\n", #expr)); \
853	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
854	} \
855	} while (0)
856
857	#define REQ_CHECK_EXPR_FMT(expr, fmt) \
858	do { \
859	if (RT_UNLIKELY(!(expr))) \
860	{ \
861	OSDBGPRINT( fmt ); \
862	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
863	} \
864	} while (0)
865
866
867	/*
868	* The switch.
869	*/
870	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
871	{
872	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
873	{
874	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
875	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
876	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
877	{
878	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
879	pReq->Hdr.rc = VERR_INVALID_MAGIC;
880	return 0;
881	}
882
883	#if 0
884	/*
885	* Call out to the OS specific code and let it do permission checks on the
886	* client process.
887	*/
888	if (!supdrvOSValidateClientProcess(pDevExt, pSession))
889	{
890	pReq->u.Out.u32Cookie = 0xffffffff;
891	pReq->u.Out.u32SessionCookie = 0xffffffff;
892	pReq->u.Out.u32SessionVersion = 0xffffffff;
893	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
894	pReq->u.Out.pSession = NULL;
895	pReq->u.Out.cFunctions = 0;
896	pReq->Hdr.rc = VERR_PERMISSION_DENIED;
897	return 0;
898	}
899	#endif
900
901	/*
902	* Match the version.
903	* The current logic is very simple, match the major interface version.
904	*/
905	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
906	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
907	{
908	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
909	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
910	pReq->u.Out.u32Cookie = 0xffffffff;
911	pReq->u.Out.u32SessionCookie = 0xffffffff;
912	pReq->u.Out.u32SessionVersion = 0xffffffff;
913	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
914	pReq->u.Out.pSession = NULL;
915	pReq->u.Out.cFunctions = 0;
916	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
917	return 0;
918	}
919
920	/*
921	* Fill in return data and be gone.
922	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
923	* u32SessionVersion <= u32ReqVersion!
924	*/
925	/** @todo Somehow validate the client and negotiate a secure cookie... */
926	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
927	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
928	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
929	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
930	pReq->u.Out.pSession = pSession;
931	pReq->u.Out.cFunctions = sizeof(g_aFunctions) / sizeof(g_aFunctions[0]);
932	pReq->Hdr.rc = VINF_SUCCESS;
933	return 0;
934	}
935
936	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_QUERY_FUNCS(0)):
937	{
938	/* validate */
939	PSUPQUERYFUNCS pReq = (PSUPQUERYFUNCS)pReqHdr;
940	REQ_CHECK_SIZES_EX(SUP_IOCTL_QUERY_FUNCS, SUP_IOCTL_QUERY_FUNCS_SIZE_IN, SUP_IOCTL_QUERY_FUNCS_SIZE_OUT(RT_ELEMENTS(g_aFunctions)));
941
942	/* execute */
943	pReq->u.Out.cFunctions = RT_ELEMENTS(g_aFunctions);
944	memcpy(&pReq->u.Out.aFunctions[0], g_aFunctions, sizeof(g_aFunctions));
945	pReq->Hdr.rc = VINF_SUCCESS;
946	return 0;
947	}
948
949	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_IDT_INSTALL):
950	{
951	/* validate */
952	PSUPIDTINSTALL pReq = (PSUPIDTINSTALL)pReqHdr;
953	REQ_CHECK_SIZES(SUP_IOCTL_IDT_INSTALL);
954
955	/* execute */
956	#ifdef VBOX_WITH_IDT_PATCHING
957	pReq->Hdr.rc = supdrvIOCtl_IdtInstall(pDevExt, pSession, pReq);
958	#else
959	pReq->u.Out.u8Idt = 3;
960	pReq->Hdr.rc = VERR_NOT_SUPPORTED;
961	#endif
962	return 0;
963	}
964
965	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_IDT_REMOVE):
966	{
967	/* validate */
968	PSUPIDTREMOVE pReq = (PSUPIDTREMOVE)pReqHdr;
969	REQ_CHECK_SIZES(SUP_IOCTL_IDT_REMOVE);
970
971	/* execute */
972	#ifdef VBOX_WITH_IDT_PATCHING
973	pReq->Hdr.rc = supdrvIOCtl_IdtRemoveAll(pDevExt, pSession);
974	#else
975	pReq->Hdr.rc = VERR_NOT_SUPPORTED;
976	#endif
977	return 0;
978	}
979
980	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_LOCK):
981	{
982	/* validate */
983	PSUPPAGELOCK pReq = (PSUPPAGELOCK)pReqHdr;
984	REQ_CHECK_SIZE_IN(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_IN);
985	REQ_CHECK_SIZE_OUT(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_OUT(pReq->u.In.cPages));
986	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.cPages > 0);
987	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.pvR3 >= PAGE_SIZE);
988
989	/* execute */
990	pReq->Hdr.rc = SUPR0LockMem(pSession, pReq->u.In.pvR3, pReq->u.In.cPages, &pReq->u.Out.aPages[0]);
991	if (RT_FAILURE(pReq->Hdr.rc))
992	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
993	return 0;
994	}
995
996	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_UNLOCK):
997	{
998	/* validate */
999	PSUPPAGEUNLOCK pReq = (PSUPPAGEUNLOCK)pReqHdr;
1000	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_UNLOCK);
1001
1002	/* execute */
1003	pReq->Hdr.rc = SUPR0UnlockMem(pSession, pReq->u.In.pvR3);
1004	return 0;
1005	}
1006
1007	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_ALLOC):
1008	{
1009	/* validate */
1010	PSUPCONTALLOC pReq = (PSUPCONTALLOC)pReqHdr;
1011	REQ_CHECK_SIZES(SUP_IOCTL_CONT_ALLOC);
1012
1013	/* execute */
1014	pReq->Hdr.rc = SUPR0ContAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.HCPhys);
1015	if (RT_FAILURE(pReq->Hdr.rc))
1016	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1017	return 0;
1018	}
1019
1020	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_FREE):
1021	{
1022	/* validate */
1023	PSUPCONTFREE pReq = (PSUPCONTFREE)pReqHdr;
1024	REQ_CHECK_SIZES(SUP_IOCTL_CONT_FREE);
1025
1026	/* execute */
1027	pReq->Hdr.rc = SUPR0ContFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1028	return 0;
1029	}
1030
1031	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_OPEN):
1032	{
1033	/* validate */
1034	PSUPLDROPEN pReq = (PSUPLDROPEN)pReqHdr;
1035	REQ_CHECK_SIZES(SUP_IOCTL_LDR_OPEN);
1036	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImage > 0);
1037	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImage < _1M*16);
1038	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.szName[0]);
1039	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, memchr(pReq->u.In.szName, '\0', sizeof(pReq->u.In.szName)));
1040	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, !supdrvCheckInvalidChar(pReq->u.In.szName, ";:()[]{}/\\\|&*%#@!~`\"'"));
1041
1042	/* execute */
1043	pReq->Hdr.rc = supdrvIOCtl_LdrOpen(pDevExt, pSession, pReq);
1044	return 0;
1045	}
1046
1047	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOAD):
1048	{
1049	/* validate */
1050	PSUPLDRLOAD pReq = (PSUPLDRLOAD)pReqHdr;
1051	REQ_CHECK_EXPR(Name, pReq->Hdr.cbIn >= sizeof(*pReq));
1052	REQ_CHECK_SIZES_EX(SUP_IOCTL_LDR_LOAD, SUP_IOCTL_LDR_LOAD_SIZE_IN(pReq->u.In.cbImage), SUP_IOCTL_LDR_LOAD_SIZE_OUT);
1053	REQ_CHECK_EXPR(SUP_IOCTL_LDR_LOAD, pReq->u.In.cSymbols <= 16384);
1054	REQ_CHECK_EXPR_FMT( !pReq->u.In.cSymbols
1055	\|\| ( pReq->u.In.offSymbols < pReq->u.In.cbImage
1056	&& pReq->u.In.offSymbols + pReq->u.In.cSymbols * sizeof(SUPLDRSYM) <= pReq->u.In.cbImage),
1057	("SUP_IOCTL_LDR_LOAD: offSymbols=%#lx cSymbols=%#lx cbImage=%#lx\n", (long)pReq->u.In.offSymbols,
1058	(long)pReq->u.In.cSymbols, (long)pReq->u.In.cbImage));
1059	REQ_CHECK_EXPR_FMT( !pReq->u.In.cbStrTab
1060	\|\| ( pReq->u.In.offStrTab < pReq->u.In.cbImage
1061	&& pReq->u.In.offStrTab + pReq->u.In.cbStrTab <= pReq->u.In.cbImage
1062	&& pReq->u.In.cbStrTab <= pReq->u.In.cbImage),
1063	("SUP_IOCTL_LDR_LOAD: offStrTab=%#lx cbStrTab=%#lx cbImage=%#lx\n", (long)pReq->u.In.offStrTab,
1064	(long)pReq->u.In.cbStrTab, (long)pReq->u.In.cbImage));
1065
1066	if (pReq->u.In.cSymbols)
1067	{
1068	uint32_t i;
1069	PSUPLDRSYM paSyms = (PSUPLDRSYM)&pReq->u.In.achImage[pReq->u.In.offSymbols];
1070	for (i = 0; i < pReq->u.In.cSymbols; i++)
1071	{
1072	REQ_CHECK_EXPR_FMT(paSyms[i].offSymbol < pReq->u.In.cbImage,
1073	("SUP_IOCTL_LDR_LOAD: sym #%ld: symb off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offSymbol, (long)pReq->u.In.cbImage));
1074	REQ_CHECK_EXPR_FMT(paSyms[i].offName < pReq->u.In.cbStrTab,
1075	("SUP_IOCTL_LDR_LOAD: sym #%ld: name off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImage));
1076	REQ_CHECK_EXPR_FMT(memchr(&pReq->u.In.achImage[pReq->u.In.offStrTab + paSyms[i].offName], '\0', pReq->u.In.cbStrTab - paSyms[i].offName),
1077	("SUP_IOCTL_LDR_LOAD: sym #%ld: unterminated name! (%#lx / %#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImage));
1078	}
1079	}
1080
1081	/* execute */
1082	pReq->Hdr.rc = supdrvIOCtl_LdrLoad(pDevExt, pSession, pReq);
1083	return 0;
1084	}
1085
1086	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_FREE):
1087	{
1088	/* validate */
1089	PSUPLDRFREE pReq = (PSUPLDRFREE)pReqHdr;
1090	REQ_CHECK_SIZES(SUP_IOCTL_LDR_FREE);
1091
1092	/* execute */
1093	pReq->Hdr.rc = supdrvIOCtl_LdrFree(pDevExt, pSession, pReq);
1094	return 0;
1095	}
1096
1097	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_GET_SYMBOL):
1098	{
1099	/* validate */
1100	PSUPLDRGETSYMBOL pReq = (PSUPLDRGETSYMBOL)pReqHdr;
1101	REQ_CHECK_SIZES(SUP_IOCTL_LDR_GET_SYMBOL);
1102	REQ_CHECK_EXPR(SUP_IOCTL_LDR_GET_SYMBOL, memchr(pReq->u.In.szSymbol, '\0', sizeof(pReq->u.In.szSymbol)));
1103
1104	/* execute */
1105	pReq->Hdr.rc = supdrvIOCtl_LdrGetSymbol(pDevExt, pSession, pReq);
1106	return 0;
1107	}
1108
1109	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0(0)):
1110	{
1111	/* validate */
1112	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1113	Log4(("SUP_IOCTL_CALL_VMMR0: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1114	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1115
1116	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_VMMR0_SIZE(0))
1117	{
1118	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(0), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(0));
1119
1120	/* execute */
1121	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1122	#ifdef SUPDRV_WITH_UNWIND_HACK
1123	pReq->Hdr.rc = supdrvNtWrapVMMR0EntryEx((PFNRT)pDevExt->pfnVMMR0EntryEx, pReq->u.In.pVMR0, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg, pSession);
1124	#else
1125	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg, pSession);
1126	#endif
1127	else
1128	pReq->Hdr.rc = VERR_WRONG_ORDER;
1129	}
1130	else
1131	{
1132	PSUPVMMR0REQHDR pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1133	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR)),
1134	("SUP_IOCTL_CALL_VMMR0: cbIn=%#x < %#x\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR))));
1135	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1136	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(pVMMReq->cbReq));
1137
1138	/* execute */
1139	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1140	#ifdef SUPDRV_WITH_UNWIND_HACK
1141	pReq->Hdr.rc = supdrvNtWrapVMMR0EntryEx((PFNRT)pDevExt->pfnVMMR0EntryEx, pReq->u.In.pVMR0, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1142	#else
1143	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1144	#endif
1145	else
1146	pReq->Hdr.rc = VERR_WRONG_ORDER;
1147	}
1148
1149	if ( RT_FAILURE(pReq->Hdr.rc)
1150	&& pReq->Hdr.rc != VERR_INTERRUPTED
1151	&& pReq->Hdr.rc != VERR_TIMEOUT)
1152	Log(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1153	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1154	else
1155	Log4(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1156	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1157	return 0;
1158	}
1159
1160	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_PAGING_MODE):
1161	{
1162	/* validate */
1163	PSUPGETPAGINGMODE pReq = (PSUPGETPAGINGMODE)pReqHdr;
1164	REQ_CHECK_SIZES(SUP_IOCTL_GET_PAGING_MODE);
1165
1166	/* execute */
1167	pReq->Hdr.rc = VINF_SUCCESS;
1168	pReq->u.Out.enmMode = supdrvIOCtl_GetPagingMode();
1169	return 0;
1170	}
1171
1172	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_ALLOC):
1173	{
1174	/* validate */
1175	PSUPLOWALLOC pReq = (PSUPLOWALLOC)pReqHdr;
1176	REQ_CHECK_EXPR(SUP_IOCTL_LOW_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_LOW_ALLOC_SIZE_IN);
1177	REQ_CHECK_SIZES_EX(SUP_IOCTL_LOW_ALLOC, SUP_IOCTL_LOW_ALLOC_SIZE_IN, SUP_IOCTL_LOW_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1178
1179	/* execute */
1180	pReq->Hdr.rc = SUPR0LowAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1181	if (RT_FAILURE(pReq->Hdr.rc))
1182	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1183	return 0;
1184	}
1185
1186	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_FREE):
1187	{
1188	/* validate */
1189	PSUPLOWFREE pReq = (PSUPLOWFREE)pReqHdr;
1190	REQ_CHECK_SIZES(SUP_IOCTL_LOW_FREE);
1191
1192	/* execute */
1193	pReq->Hdr.rc = SUPR0LowFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1194	return 0;
1195	}
1196
1197	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_MAP):
1198	{
1199	/* validate */
1200	PSUPGIPMAP pReq = (PSUPGIPMAP)pReqHdr;
1201	REQ_CHECK_SIZES(SUP_IOCTL_GIP_MAP);
1202
1203	/* execute */
1204	pReq->Hdr.rc = SUPR0GipMap(pSession, &pReq->u.Out.pGipR3, &pReq->u.Out.HCPhysGip);
1205	if (RT_SUCCESS(pReq->Hdr.rc))
1206	pReq->u.Out.pGipR0 = pDevExt->pGip;
1207	return 0;
1208	}
1209
1210	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_UNMAP):
1211	{
1212	/* validate */
1213	PSUPGIPUNMAP pReq = (PSUPGIPUNMAP)pReqHdr;
1214	REQ_CHECK_SIZES(SUP_IOCTL_GIP_UNMAP);
1215
1216	/* execute */
1217	pReq->Hdr.rc = SUPR0GipUnmap(pSession);
1218	return 0;
1219	}
1220
1221	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SET_VM_FOR_FAST):
1222	{
1223	/* validate */
1224	PSUPSETVMFORFAST pReq = (PSUPSETVMFORFAST)pReqHdr;
1225	REQ_CHECK_SIZES(SUP_IOCTL_SET_VM_FOR_FAST);
1226	REQ_CHECK_EXPR_FMT( !pReq->u.In.pVMR0
1227	\|\| ( VALID_PTR(pReq->u.In.pVMR0)
1228	&& !((uintptr_t)pReq->u.In.pVMR0 & (PAGE_SIZE - 1))),
1229	("SUP_IOCTL_SET_VM_FOR_FAST: pVMR0=%p!\n", pReq->u.In.pVMR0));
1230	/* execute */
1231	pSession->pVM = pReq->u.In.pVMR0;
1232	pReq->Hdr.rc = VINF_SUCCESS;
1233	return 0;
1234	}
1235
1236	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_ALLOC):
1237	{
1238	/* validate */
1239	PSUPPAGEALLOC pReq = (PSUPPAGEALLOC)pReqHdr;
1240	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_SIZE_IN);
1241	REQ_CHECK_SIZES_EX(SUP_IOCTL_PAGE_ALLOC, SUP_IOCTL_PAGE_ALLOC_SIZE_IN, SUP_IOCTL_PAGE_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1242
1243	/* execute */
1244	pReq->Hdr.rc = SUPR0PageAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1245	if (RT_FAILURE(pReq->Hdr.rc))
1246	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1247	return 0;
1248	}
1249
1250	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_FREE):
1251	{
1252	/* validate */
1253	PSUPPAGEFREE pReq = (PSUPPAGEFREE)pReqHdr;
1254	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_FREE);
1255
1256	/* execute */
1257	pReq->Hdr.rc = SUPR0PageFree(pSession, pReq->u.In.pvR3);
1258	return 0;
1259	}
1260
1261	default:
1262	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1263	break;
1264	}
1265	return SUPDRV_ERR_GENERAL_FAILURE;
1266	}
1267
1268
1269	/**
1270	* Inter-Driver Communcation (IDC) worker.
1271	*
1272	* @returns VBox status code.
1273	* @retval VINF_SUCCESS on success.
1274	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1275	* @retval VERR_NOT_SUPPORTED if the request isn't supported.
1276	*
1277	* @param uReq The request (function) code.
1278	* @param pDevExt Device extention.
1279	* @param pSession Session data.
1280	* @param pReqHdr The request header.
1281	*/
1282	int VBOXCALL supdrvIDC(uintptr_t uReq, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQHDR pReqHdr)
1283	{
1284	/*
1285	* The OS specific code has already validated the pSession
1286	* pointer, and the request size being greater or equal to
1287	* size of the header.
1288	*
1289	* So, just check that pSession is a kernel context session.
1290	*/
1291	if (RT_UNLIKELY( pSession
1292	&& pSession->R0Process != NIL_RTR0PROCESS))
1293	return VERR_INVALID_PARAMETER;
1294
1295	/*
1296	* Validation macro.
1297	*/
1298	#define REQ_CHECK_IDC_SIZE(Name, cbExpect) \
1299	do { \
1300	if (RT_UNLIKELY(pReqHdr->cb != (cbExpect))) \
1301	{ \
1302	OSDBGPRINT(( #Name ": Invalid input/output sizes. cb=%ld expected %ld.\n", \
1303	(long)pReqHdr->cb, (long)(cbExpect))); \
1304	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1305	} \
1306	} while (0)
1307
1308	switch (uReq)
1309	{
1310	case SUPDRV_IDC_REQ_CONNECT:
1311	{
1312	PSUPDRVIDCREQCONNECT pReq = (PSUPDRVIDCREQCONNECT)pReqHdr;
1313	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_CONNECT, sizeof(*pReq));
1314
1315	/*
1316	* Validate the cookie and other input.
1317	*/
1318	if (pReq->Hdr.pSession != NULL)
1319	{
1320	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: pSession=%p expected NULL!\n", pReq->Hdr.pSession));
1321	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1322	}
1323	if (pReq->u.In.u32MagicCookie != SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE)
1324	{
1325	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: u32MagicCookie=%#x expected %#x!\n",
1326	pReq->u.In.u32MagicCookie, SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE));
1327	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1328	}
1329	if ( pReq->u.In.uMinVersion > pReq->u.In.uReqVersion
1330	\|\| (pReq->u.In.uMinVersion & UINT32_C(0xffff0000)) != (pReq->u.In.uReqVersion & UINT32_C(0xffff0000)))
1331	{
1332	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: uMinVersion=%#x uMaxVersion=%#x doesn't match!\n",
1333	pReq->u.In.uMinVersion, pReq->u.In.uReqVersion));
1334	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1335	}
1336
1337	/*
1338	* Match the version.
1339	* The current logic is very simple, match the major interface version.
1340	*/
1341	if ( pReq->u.In.uMinVersion > SUPDRV_IDC_VERSION
1342	\|\| (pReq->u.In.uMinVersion & 0xffff0000) != (SUPDRV_IDC_VERSION & 0xffff0000))
1343	{
1344	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1345	pReq->u.In.uReqVersion, pReq->u.In.uMinVersion, SUPDRV_IDC_VERSION));
1346	pReq->u.Out.pSession = NULL;
1347	pReq->u.Out.uSessionVersion = 0xffffffff;
1348	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1349	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1350	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1351	return VINF_SUCCESS;
1352	}
1353
1354	pReq->u.Out.pSession = NULL;
1355	pReq->u.Out.uSessionVersion = SUPDRV_IDC_VERSION;
1356	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1357	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1358
1359	/*
1360	* On NT we will already have a session associated with the
1361	* client, just like with the SUP_IOCTL_COOKIE request, while
1362	* the other doesn't.
1363	*/
1364	#ifdef RT_OS_WINDOWS
1365	pReq->Hdr.rc = VINF_SUCCESS;
1366	#else
1367	AssertReturn(!pSession, VERR_INTERNAL_ERROR);
1368	pReq->Hdr.rc = supdrvCreateSession(pDevExt, false /* fUser */, &pSession);
1369	if (RT_FAILURE(pReq->Hdr.rc))
1370	{
1371	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: failed to create session, rc=%d\n", pReq->Hdr.rc));
1372	return VINF_SUCCESS;
1373	}
1374	#endif
1375
1376	pReq->u.Out.pSession = pSession;
1377	pReq->Hdr.pSession = pSession;
1378
1379	return VINF_SUCCESS;
1380	}
1381
1382	case SUPDRV_IDC_REQ_DISCONNECT:
1383	{
1384	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_DISCONNECT, sizeof(*pReqHdr));
1385
1386	#ifdef RT_OS_WINDOWS
1387	/* Windows will destroy the session when the file object is destroyed. */
1388	#else
1389	supdrvCloseSession(pDevExt, pSession);
1390	#endif
1391	return pReqHdr->rc = VINF_SUCCESS;
1392	}
1393
1394	case SUPDRV_IDC_REQ_GET_SYMBOL:
1395	{
1396	PSUPDRVIDCREQGETSYM pReq = (PSUPDRVIDCREQGETSYM)pReqHdr;
1397	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_GET_SYMBOL, sizeof(*pReq));
1398
1399	pReq->Hdr.rc = supdrvIDC_LdrGetSymbol(pDevExt, pSession, pReq);
1400	return VINF_SUCCESS;
1401	}
1402
1403	case SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY:
1404	{
1405	PSUPDRVIDCREQCOMPREGFACTORY pReq = (PSUPDRVIDCREQCOMPREGFACTORY)pReqHdr;
1406	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY, sizeof(*pReq));
1407
1408	pReq->Hdr.rc = SUPR0ComponentRegisterFactory(pSession, pReq->u.In.pFactory);
1409	return VINF_SUCCESS;
1410	}
1411
1412	case SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY:
1413	{
1414	PSUPDRVIDCREQCOMPDEREGFACTORY pReq = (PSUPDRVIDCREQCOMPDEREGFACTORY)pReqHdr;
1415	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY, sizeof(*pReq));
1416
1417	pReq->Hdr.rc = SUPR0ComponentDeregisterFactory(pSession, pReq->u.In.pFactory);
1418	return VINF_SUCCESS;
1419	}
1420
1421	default:
1422	Log(("Unknown IDC %#lx\n", (long)uReq));
1423	break;
1424	}
1425
1426	#undef REQ_CHECK_IDC_SIZE
1427	return VERR_NOT_SUPPORTED;
1428	}
1429
1430
1431	/**
1432	* Register a object for reference counting.
1433	* The object is registered with one reference in the specified session.
1434	*
1435	* @returns Unique identifier on success (pointer).
1436	* All future reference must use this identifier.
1437	* @returns NULL on failure.
1438	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
1439	* @param pvUser1 The first user argument.
1440	* @param pvUser2 The second user argument.
1441	*/
1442	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
1443	{
1444	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1445	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1446	PSUPDRVOBJ pObj;
1447	PSUPDRVUSAGE pUsage;
1448
1449	/*
1450	* Validate the input.
1451	*/
1452	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
1453	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
1454	AssertPtrReturn(pfnDestructor, NULL);
1455
1456	/*
1457	* Allocate and initialize the object.
1458	*/
1459	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
1460	if (!pObj)
1461	return NULL;
1462	pObj->u32Magic = SUPDRVOBJ_MAGIC;
1463	pObj->enmType = enmType;
1464	pObj->pNext = NULL;
1465	pObj->cUsage = 1;
1466	pObj->pfnDestructor = pfnDestructor;
1467	pObj->pvUser1 = pvUser1;
1468	pObj->pvUser2 = pvUser2;
1469	pObj->CreatorUid = pSession->Uid;
1470	pObj->CreatorGid = pSession->Gid;
1471	pObj->CreatorProcess= pSession->Process;
1472	supdrvOSObjInitCreator(pObj, pSession);
1473
1474	/*
1475	* Allocate the usage record.
1476	* (We keep freed usage records around to simplify SUPR0ObjAddRef().)
1477	*/
1478	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1479
1480	pUsage = pDevExt->pUsageFree;
1481	if (pUsage)
1482	pDevExt->pUsageFree = pUsage->pNext;
1483	else
1484	{
1485	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1486	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
1487	if (!pUsage)
1488	{
1489	RTMemFree(pObj);
1490	return NULL;
1491	}
1492	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1493	}
1494
1495	/*
1496	* Insert the object and create the session usage record.
1497	*/
1498	/* The object. */
1499	pObj->pNext = pDevExt->pObjs;
1500	pDevExt->pObjs = pObj;
1501
1502	/* The session record. */
1503	pUsage->cUsage = 1;
1504	pUsage->pObj = pObj;
1505	pUsage->pNext = pSession->pUsage;
1506	/* Log2(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext)); */
1507	pSession->pUsage = pUsage;
1508
1509	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1510
1511	Log(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
1512	return pObj;
1513	}
1514
1515
1516	/**
1517	* Increment the reference counter for the object associating the reference
1518	* with the specified session.
1519	*
1520	* @returns IPRT status code.
1521	* @param pvObj The identifier returned by SUPR0ObjRegister().
1522	* @param pSession The session which is referencing the object.
1523	*
1524	* @remarks The caller should not own any spinlocks and must carefully protect
1525	* itself against potential race with the destructor so freed memory
1526	* isn't accessed here.
1527	*/
1528	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
1529	{
1530	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1531	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1532	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
1533	PSUPDRVUSAGE pUsagePre;
1534	PSUPDRVUSAGE pUsage;
1535
1536	/*
1537	* Validate the input.
1538	* Be ready for the destruction race (someone might be stuck in the
1539	* destructor waiting a lock we own).
1540	*/
1541	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1542	AssertPtrReturn(pObj, VERR_INVALID_POINTER);
1543	AssertMsgReturn(pObj->u32Magic == SUPDRVOBJ_MAGIC \|\| pObj->u32Magic == SUPDRVOBJ_MAGIC + 1,
1544	("Invalid pvObj=%p magic=%#x (expected %#x or %#x)\n", pvObj, pObj->u32Magic, SUPDRVOBJ_MAGIC, SUPDRVOBJ_MAGIC + 1),
1545	VERR_INVALID_PARAMETER);
1546
1547	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1548
1549	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
1550	{
1551	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1552
1553	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
1554	return VERR_WRONG_ORDER;
1555	}
1556
1557	/*
1558	* Preallocate the usage record.
1559	*/
1560	pUsagePre = pDevExt->pUsageFree;
1561	if (pUsagePre)
1562	pDevExt->pUsageFree = pUsagePre->pNext;
1563	else
1564	{
1565	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1566	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
1567	if (!pUsagePre)
1568	return VERR_NO_MEMORY;
1569
1570	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1571	}
1572
1573	/*
1574	* Reference the object.
1575	*/
1576	pObj->cUsage++;
1577
1578	/*
1579	* Look for the session record.
1580	*/
1581	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
1582	{
1583	/Log(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
1584	if (pUsage->pObj == pObj)
1585	break;
1586	}
1587	if (pUsage)
1588	pUsage->cUsage++;
1589	else
1590	{
1591	/* create a new session record. */
1592	pUsagePre->cUsage = 1;
1593	pUsagePre->pObj = pObj;
1594	pUsagePre->pNext = pSession->pUsage;
1595	pSession->pUsage = pUsagePre;
1596	/Log(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));/
1597
1598	pUsagePre = NULL;
1599	}
1600
1601	/*
1602	* Put any unused usage record into the free list..
1603	*/
1604	if (pUsagePre)
1605	{
1606	pUsagePre->pNext = pDevExt->pUsageFree;
1607	pDevExt->pUsageFree = pUsagePre;
1608	}
1609
1610	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1611
1612	return VINF_SUCCESS;
1613	}
1614
1615
1616	/**
1617	* Decrement / destroy a reference counter record for an object.
1618	*
1619	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
1620	*
1621	* @returns IPRT status code.
1622	* @param pvObj The identifier returned by SUPR0ObjRegister().
1623	* @param pSession The session which is referencing the object.
1624	*/
1625	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
1626	{
1627	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1628	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1629	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
1630	bool fDestroy = false;
1631	PSUPDRVUSAGE pUsage;
1632	PSUPDRVUSAGE pUsagePrev;
1633
1634	/*
1635	* Validate the input.
1636	*/
1637	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1638	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
1639	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
1640	VERR_INVALID_PARAMETER);
1641
1642	/*
1643	* Acquire the spinlock and look for the usage record.
1644	*/
1645	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1646
1647	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
1648	pUsage;
1649	pUsagePrev = pUsage, pUsage = pUsage->pNext)
1650	{
1651	/Log2(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
1652	if (pUsage->pObj == pObj)
1653	{
1654	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
1655	if (pUsage->cUsage > 1)
1656	{
1657	pObj->cUsage--;
1658	pUsage->cUsage--;
1659	}
1660	else
1661	{
1662	/*
1663	* Free the session record.
1664	*/
1665	if (pUsagePrev)
1666	pUsagePrev->pNext = pUsage->pNext;
1667	else
1668	pSession->pUsage = pUsage->pNext;
1669	pUsage->pNext = pDevExt->pUsageFree;
1670	pDevExt->pUsageFree = pUsage;
1671
1672	/* What about the object? */
1673	if (pObj->cUsage > 1)
1674	pObj->cUsage--;
1675	else
1676	{
1677	/*
1678	* Object is to be destroyed, unlink it.
1679	*/
1680	pObj->u32Magic = SUPDRVOBJ_MAGIC + 1;
1681	fDestroy = true;
1682	if (pDevExt->pObjs == pObj)
1683	pDevExt->pObjs = pObj->pNext;
1684	else
1685	{
1686	PSUPDRVOBJ pObjPrev;
1687	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
1688	if (pObjPrev->pNext == pObj)
1689	{
1690	pObjPrev->pNext = pObj->pNext;
1691	break;
1692	}
1693	Assert(pObjPrev);
1694	}
1695	}
1696	}
1697	break;
1698	}
1699	}
1700
1701	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1702
1703	/*
1704	* Call the destructor and free the object if required.
1705	*/
1706	if (fDestroy)
1707	{
1708	Log(("SUPR0ObjRelease: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
1709	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
1710	if (pObj->pfnDestructor)
1711	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
1712	RTMemFree(pObj);
1713	}
1714
1715	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
1716	return pUsage ? VINF_SUCCESS : VERR_INVALID_PARAMETER;
1717	}
1718
1719	/**
1720	* Verifies that the current process can access the specified object.
1721	*
1722	* @returns The following IPRT status code:
1723	* @retval VINF_SUCCESS if access was granted.
1724	* @retval VERR_PERMISSION_DENIED if denied access.
1725	* @retval VERR_INVALID_PARAMETER if invalid parameter.
1726	*
1727	* @param pvObj The identifier returned by SUPR0ObjRegister().
1728	* @param pSession The session which wishes to access the object.
1729	* @param pszObjName Object string name. This is optional and depends on the object type.
1730	*
1731	* @remark The caller is responsible for making sure the object isn't removed while
1732	* we're inside this function. If uncertain about this, just call AddRef before calling us.
1733	*/
1734	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
1735	{
1736	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
1737	int rc;
1738
1739	/*
1740	* Validate the input.
1741	*/
1742	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1743	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
1744	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
1745	VERR_INVALID_PARAMETER);
1746
1747	/*
1748	* Check access. (returns true if a decision has been made.)
1749	*/
1750	rc = VERR_INTERNAL_ERROR;
1751	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
1752	return rc;
1753
1754	/*
1755	* Default policy is to allow the user to access his own
1756	* stuff but nothing else.
1757	*/
1758	if (pObj->CreatorUid == pSession->Uid)
1759	return VINF_SUCCESS;
1760	return VERR_PERMISSION_DENIED;
1761	}
1762
1763
1764	/**
1765	* Lock pages.
1766	*
1767	* @returns IPRT status code.
1768	* @param pSession Session to which the locked memory should be associated.
1769	* @param pvR3 Start of the memory range to lock.
1770	* This must be page aligned.
1771	* @param cb Size of the memory range to lock.
1772	* This must be page aligned.
1773	*/
1774	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
1775	{
1776	int rc;
1777	SUPDRVMEMREF Mem = {0};
1778	const size_t cb = (size_t)cPages << PAGE_SHIFT;
1779	LogFlow(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
1780
1781	/*
1782	* Verify input.
1783	*/
1784	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1785	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
1786	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
1787	\|\| !pvR3)
1788	{
1789	Log(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
1790	return VERR_INVALID_PARAMETER;
1791	}
1792
1793	#ifdef RT_OS_WINDOWS /* A temporary hack for windows, will be removed once all ring-3 code has been cleaned up. */
1794	/* First check if we allocated it using SUPPageAlloc; if so then we don't need to lock it again */
1795	rc = supdrvPageGetPhys(pSession, pvR3, cPages, paPages);
1796	if (RT_SUCCESS(rc))
1797	return rc;
1798	#endif
1799
1800	/*
1801	* Let IPRT do the job.
1802	*/
1803	Mem.eType = MEMREF_TYPE_LOCKED;
1804	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTR0ProcHandleSelf());
1805	if (RT_SUCCESS(rc))
1806	{
1807	uint32_t iPage = cPages;
1808	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
1809	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
1810
1811	while (iPage-- > 0)
1812	{
1813	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
1814	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
1815	{
1816	AssertMsgFailed(("iPage=%d\n", iPage));
1817	rc = VERR_INTERNAL_ERROR;
1818	break;
1819	}
1820	}
1821	if (RT_SUCCESS(rc))
1822	rc = supdrvMemAdd(&Mem, pSession);
1823	if (RT_FAILURE(rc))
1824	{
1825	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
1826	AssertRC(rc2);
1827	}
1828	}
1829
1830	return rc;
1831	}
1832
1833
1834	/**
1835	* Unlocks the memory pointed to by pv.
1836	*
1837	* @returns IPRT status code.
1838	* @param pSession Session to which the memory was locked.
1839	* @param pvR3 Memory to unlock.
1840	*/
1841	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
1842	{
1843	LogFlow(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
1844	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1845	#ifdef RT_OS_WINDOWS
1846	/*
1847	* Temporary hack for windows - SUPR0PageFree will unlock SUPR0PageAlloc
1848	* allocations; ignore this call.
1849	*/
1850	if (supdrvPageWasLockedByPageAlloc(pSession, pvR3))
1851	{
1852	LogFlow(("Page will be unlocked in SUPR0PageFree -> ignore\n"));
1853	return VINF_SUCCESS;
1854	}
1855	#endif
1856	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
1857	}
1858
1859
1860	/**
1861	* Allocates a chunk of page aligned memory with contiguous and fixed physical
1862	* backing.
1863	*
1864	* @returns IPRT status code.
1865	* @param pSession Session data.
1866	* @param cb Number of bytes to allocate.
1867	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
1868	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
1869	* @param pHCPhys Where to put the physical address of allocated memory.
1870	*/
1871	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
1872	{
1873	int rc;
1874	SUPDRVMEMREF Mem = {0};
1875	LogFlow(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
1876
1877	/*
1878	* Validate input.
1879	*/
1880	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1881	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
1882	{
1883	Log(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
1884	pSession, ppvR0, ppvR3, pHCPhys));
1885	return VERR_INVALID_PARAMETER;
1886
1887	}
1888	if (cPages < 1 \|\| cPages >= 256)
1889	{
1890	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256\n", cPages));
1891	return VERR_INVALID_PARAMETER;
1892	}
1893
1894	/*
1895	* Let IPRT do the job.
1896	*/
1897	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
1898	if (RT_SUCCESS(rc))
1899	{
1900	int rc2;
1901	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
1902	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
1903	if (RT_SUCCESS(rc))
1904	{
1905	Mem.eType = MEMREF_TYPE_CONT;
1906	rc = supdrvMemAdd(&Mem, pSession);
1907	if (!rc)
1908	{
1909	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
1910	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
1911	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
1912	return 0;
1913	}
1914
1915	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
1916	AssertRC(rc2);
1917	}
1918	rc2 = RTR0MemObjFree(Mem.MemObj, false);
1919	AssertRC(rc2);
1920	}
1921
1922	return rc;
1923	}
1924
1925
1926	/**
1927	* Frees memory allocated using SUPR0ContAlloc().
1928	*
1929	* @returns IPRT status code.
1930	* @param pSession The session to which the memory was allocated.
1931	* @param uPtr Pointer to the memory (ring-3 or ring-0).
1932	*/
1933	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
1934	{
1935	LogFlow(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
1936	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1937	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
1938	}
1939
1940
1941	/**
1942	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
1943	*
1944	* The memory isn't zeroed.
1945	*
1946	* @returns IPRT status code.
1947	* @param pSession Session data.
1948	* @param cPages Number of pages to allocate.
1949	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
1950	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
1951	* @param paPages Where to put the physical addresses of allocated memory.
1952	*/
1953	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
1954	{
1955	unsigned iPage;
1956	int rc;
1957	SUPDRVMEMREF Mem = {0};
1958	LogFlow(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
1959
1960	/*
1961	* Validate input.
1962	*/
1963	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1964	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
1965	{
1966	Log(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
1967	pSession, ppvR3, ppvR0, paPages));
1968	return VERR_INVALID_PARAMETER;
1969
1970	}
1971	if (cPages < 1 \|\| cPages > 256)
1972	{
1973	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
1974	return VERR_INVALID_PARAMETER;
1975	}
1976
1977	/*
1978	* Let IPRT do the work.
1979	*/
1980	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
1981	if (RT_SUCCESS(rc))
1982	{
1983	int rc2;
1984	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
1985	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
1986	if (RT_SUCCESS(rc))
1987	{
1988	Mem.eType = MEMREF_TYPE_LOW;
1989	rc = supdrvMemAdd(&Mem, pSession);
1990	if (!rc)
1991	{
1992	for (iPage = 0; iPage < cPages; iPage++)
1993	{
1994	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
1995	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%VHp\n", paPages[iPage]));
1996	}
1997	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
1998	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
1999	return 0;
2000	}
2001
2002	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2003	AssertRC(rc2);
2004	}
2005
2006	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2007	AssertRC(rc2);
2008	}
2009
2010	return rc;
2011	}
2012
2013
2014	/**
2015	* Frees memory allocated using SUPR0LowAlloc().
2016	*
2017	* @returns IPRT status code.
2018	* @param pSession The session to which the memory was allocated.
2019	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2020	*/
2021	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2022	{
2023	LogFlow(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2024	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2025	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
2026	}
2027
2028
2029
2030	/**
2031	* Allocates a chunk of memory with both R0 and R3 mappings.
2032	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
2033	*
2034	* @returns IPRT status code.
2035	* @param pSession The session to associated the allocation with.
2036	* @param cb Number of bytes to allocate.
2037	* @param ppvR0 Where to store the address of the Ring-0 mapping.
2038	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2039	*/
2040	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
2041	{
2042	int rc;
2043	SUPDRVMEMREF Mem = {0};
2044	LogFlow(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
2045
2046	/*
2047	* Validate input.
2048	*/
2049	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2050	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
2051	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
2052	if (cb < 1 \|\| cb >= _4M)
2053	{
2054	Log(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
2055	return VERR_INVALID_PARAMETER;
2056	}
2057
2058	/*
2059	* Let IPRT do the work.
2060	*/
2061	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
2062	if (RT_SUCCESS(rc))
2063	{
2064	int rc2;
2065	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2066	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2067	if (RT_SUCCESS(rc))
2068	{
2069	Mem.eType = MEMREF_TYPE_MEM;
2070	rc = supdrvMemAdd(&Mem, pSession);
2071	if (!rc)
2072	{
2073	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2074	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2075	return VINF_SUCCESS;
2076	}
2077	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2078	AssertRC(rc2);
2079	}
2080
2081	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2082	AssertRC(rc2);
2083	}
2084
2085	return rc;
2086	}
2087
2088
2089	/**
2090	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
2091	*
2092	* @returns IPRT status code.
2093	* @param pSession The session to which the memory was allocated.
2094	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2095	* @param paPages Where to store the physical addresses.
2096	*/
2097	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
2098	{
2099	PSUPDRVBUNDLE pBundle;
2100	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2101	LogFlow(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
2102
2103	/*
2104	* Validate input.
2105	*/
2106	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2107	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
2108	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
2109
2110	/*
2111	* Search for the address.
2112	*/
2113	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2114	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2115	{
2116	if (pBundle->cUsed > 0)
2117	{
2118	unsigned i;
2119	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2120	{
2121	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
2122	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2123	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2124	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2125	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
2126	)
2127	)
2128	{
2129	const unsigned cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2130	unsigned iPage;
2131	for (iPage = 0; iPage < cPages; iPage++)
2132	{
2133	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2134	paPages[iPage].uReserved = 0;
2135	}
2136	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2137	return VINF_SUCCESS;
2138	}
2139	}
2140	}
2141	}
2142	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2143	Log(("Failed to find %p!!!\n", (void *)uPtr));
2144	return VERR_INVALID_PARAMETER;
2145	}
2146
2147
2148	/**
2149	* Free memory allocated by SUPR0MemAlloc().
2150	*
2151	* @returns IPRT status code.
2152	* @param pSession The session owning the allocation.
2153	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2154	*/
2155	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2156	{
2157	LogFlow(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2158	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2159	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
2160	}
2161
2162
2163	/**
2164	* Allocates a chunk of memory with only a R3 mappings.
2165	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
2166	*
2167	* @returns IPRT status code.
2168	* @param pSession The session to associated the allocation with.
2169	* @param cPages The number of pages to allocate.
2170	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2171	* @param paPages Where to store the addresses of the pages. Optional.
2172	*/
2173	SUPR0DECL(int) SUPR0PageAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR3PTR ppvR3, PRTHCPHYS paPages)
2174	{
2175	int rc;
2176	SUPDRVMEMREF Mem = {0};
2177	LogFlow(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
2178
2179	/*
2180	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2181	*/
2182	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2183	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
2184	if (cPages < 1 \|\| cPages > (128 * _1M)/PAGE_SIZE)
2185	{
2186	Log(("SUPR0PageAlloc: Illegal request cb=%u; must be greater than 0 and smaller than 128MB.\n", cPages));
2187	return VERR_INVALID_PARAMETER;
2188	}
2189
2190	/*
2191	* Let IPRT do the work.
2192	*/
2193	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
2194	if (RT_SUCCESS(rc))
2195	{
2196	int rc2;
2197	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2198	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2199	if (RT_SUCCESS(rc))
2200	{
2201	Mem.eType = MEMREF_TYPE_LOCKED_SUP;
2202	rc = supdrvMemAdd(&Mem, pSession);
2203	if (!rc)
2204	{
2205	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2206	if (paPages)
2207	{
2208	uint32_t iPage = cPages;
2209	while (iPage-- > 0)
2210	{
2211	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
2212	Assert(paPages[iPage] != NIL_RTHCPHYS);
2213	}
2214	}
2215	return VINF_SUCCESS;
2216	}
2217	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2218	AssertRC(rc2);
2219	}
2220
2221	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2222	AssertRC(rc2);
2223	}
2224	return rc;
2225	}
2226
2227
2228	#ifdef RT_OS_WINDOWS
2229	/**
2230	* Check if the pages were locked by SUPR0PageAlloc
2231	*
2232	* This function will be removed along with the lock/unlock hacks when
2233	* we've cleaned up the ring-3 code properly.
2234	*
2235	* @returns boolean
2236	* @param pSession The session to which the memory was allocated.
2237	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc().
2238	*/
2239	static bool supdrvPageWasLockedByPageAlloc(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2240	{
2241	PSUPDRVBUNDLE pBundle;
2242	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2243	LogFlow(("SUPR0PageIsLockedByPageAlloc: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2244
2245	/*
2246	* Search for the address.
2247	*/
2248	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2249	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2250	{
2251	if (pBundle->cUsed > 0)
2252	{
2253	unsigned i;
2254	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2255	{
2256	if ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED_SUP
2257	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2258	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2259	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
2260	{
2261	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2262	return true;
2263	}
2264	}
2265	}
2266	}
2267	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2268	return false;
2269	}
2270
2271
2272	/**
2273	* Get the physical addresses of memory allocated using SUPR0PageAlloc().
2274	*
2275	* This function will be removed along with the lock/unlock hacks when
2276	* we've cleaned up the ring-3 code properly.
2277	*
2278	* @returns IPRT status code.
2279	* @param pSession The session to which the memory was allocated.
2280	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc().
2281	* @param cPages Number of pages in paPages
2282	* @param paPages Where to store the physical addresses.
2283	*/
2284	static int supdrvPageGetPhys(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
2285	{
2286	PSUPDRVBUNDLE pBundle;
2287	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2288	LogFlow(("supdrvPageGetPhys: pSession=%p pvR3=%p cPages=%#lx paPages=%p\n", pSession, (void *)pvR3, (long)cPages, paPages));
2289
2290	/*
2291	* Search for the address.
2292	*/
2293	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2294	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2295	{
2296	if (pBundle->cUsed > 0)
2297	{
2298	unsigned i;
2299	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2300	{
2301	if ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED_SUP
2302	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2303	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2304	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
2305	{
2306	uint32_t iPage = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2307	cPages = RT_MIN(iPage, cPages);
2308	for (iPage = 0; iPage < cPages; iPage++)
2309	paPages[iPage] = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2310	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2311	return VINF_SUCCESS;
2312	}
2313	}
2314	}
2315	}
2316	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2317	return VERR_INVALID_PARAMETER;
2318	}
2319	#endif /* RT_OS_WINDOWS */
2320
2321
2322	/**
2323	* Free memory allocated by SUPR0PageAlloc().
2324	*
2325	* @returns IPRT status code.
2326	* @param pSession The session owning the allocation.
2327	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc().
2328	*/
2329	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2330	{
2331	LogFlow(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2332	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2333	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED_SUP);
2334	}
2335
2336
2337	/**
2338	* Maps the GIP into userspace and/or get the physical address of the GIP.
2339	*
2340	* @returns IPRT status code.
2341	* @param pSession Session to which the GIP mapping should belong.
2342	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
2343	* @param pHCPhysGip Where to store the physical address. (optional)
2344	*
2345	* @remark There is no reference counting on the mapping, so one call to this function
2346	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
2347	* and remove the session as a GIP user.
2348	*/
2349	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
2350	{
2351	int rc = 0;
2352	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2353	RTR3PTR pGip = NIL_RTR3PTR;
2354	RTHCPHYS HCPhys = NIL_RTHCPHYS;
2355	LogFlow(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
2356
2357	/*
2358	* Validate
2359	*/
2360	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2361	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
2362	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
2363
2364	RTSemFastMutexRequest(pDevExt->mtxGip);
2365	if (pDevExt->pGip)
2366	{
2367	/*
2368	* Map it?
2369	*/
2370	if (ppGipR3)
2371	{
2372	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
2373	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
2374	RTMEM_PROT_READ, RTR0ProcHandleSelf());
2375	if (RT_SUCCESS(rc))
2376	{
2377	pGip = RTR0MemObjAddressR3(pSession->GipMapObjR3);
2378	rc = VINF_SUCCESS; /** @todo remove this and replace the !rc below with RT_SUCCESS(rc). */
2379	}
2380	}
2381
2382	/*
2383	* Get physical address.
2384	*/
2385	if (pHCPhysGip && !rc)
2386	HCPhys = pDevExt->HCPhysGip;
2387
2388	/*
2389	* Reference globally.
2390	*/
2391	if (!pSession->fGipReferenced && !rc)
2392	{
2393	pSession->fGipReferenced = 1;
2394	pDevExt->cGipUsers++;
2395	if (pDevExt->cGipUsers == 1)
2396	{
2397	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
2398	unsigned i;
2399
2400	LogFlow(("SUPR0GipMap: Resumes GIP updating\n"));
2401
2402	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
2403	ASMAtomicXchgU32(&pGip->aCPUs[i].u32TransactionId, pGip->aCPUs[i].u32TransactionId & ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
2404	ASMAtomicXchgU64(&pGip->u64NanoTSLastUpdateHz, 0);
2405
2406	rc = RTTimerStart(pDevExt->pGipTimer, 0);
2407	AssertRC(rc); rc = VINF_SUCCESS;
2408	}
2409	}
2410	}
2411	else
2412	{
2413	rc = SUPDRV_ERR_GENERAL_FAILURE;
2414	Log(("SUPR0GipMap: GIP is not available!\n"));
2415	}
2416	RTSemFastMutexRelease(pDevExt->mtxGip);
2417
2418	/*
2419	* Write returns.
2420	*/
2421	if (pHCPhysGip)
2422	*pHCPhysGip = HCPhys;
2423	if (ppGipR3)
2424	*ppGipR3 = pGip;
2425
2426	#ifdef DEBUG_DARWIN_GIP
2427	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx ppGip=%p GipMapObjR3\n", rc, (unsigned long)HCPhys, pGip, pSession->GipMapObjR3));
2428	#else
2429	LogFlow(("SUPR0GipMap: returns %d pHCPhysGip=%lx ppGipR3=%p\n", rc, (unsigned long)HCPhys, (void *)(uintptr_t)pGip));
2430	#endif
2431	return rc;
2432	}
2433
2434
2435	/**
2436	* Unmaps any user mapping of the GIP and terminates all GIP access
2437	* from this session.
2438	*
2439	* @returns IPRT status code.
2440	* @param pSession Session to which the GIP mapping should belong.
2441	*/
2442	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
2443	{
2444	int rc = VINF_SUCCESS;
2445	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2446	#ifdef DEBUG_DARWIN_GIP
2447	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
2448	pSession,
2449	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
2450	pSession->GipMapObjR3));
2451	#else
2452	LogFlow(("SUPR0GipUnmap: pSession=%p\n", pSession));
2453	#endif
2454	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2455
2456	RTSemFastMutexRequest(pDevExt->mtxGip);
2457
2458	/*
2459	* Unmap anything?
2460	*/
2461	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
2462	{
2463	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
2464	AssertRC(rc);
2465	if (RT_SUCCESS(rc))
2466	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
2467	}
2468
2469	/*
2470	* Dereference global GIP.
2471	*/
2472	if (pSession->fGipReferenced && !rc)
2473	{
2474	pSession->fGipReferenced = 0;
2475	if ( pDevExt->cGipUsers > 0
2476	&& !--pDevExt->cGipUsers)
2477	{
2478	LogFlow(("SUPR0GipUnmap: Suspends GIP updating\n"));
2479	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = 0;
2480	}
2481	}
2482
2483	RTSemFastMutexRelease(pDevExt->mtxGip);
2484
2485	return rc;
2486	}
2487
2488
2489	/**
2490	* Register a component factory with the support driver.
2491	*
2492	* This is currently restricted to kernel sessions only.
2493	*
2494	* @returns VBox status code.
2495	* @retval VINF_SUCCESS on success.
2496	* @retval VERR_NO_MEMORY if we're out of memory.
2497	* @retval VERR_ALREADY_EXISTS if the factory has already been registered.
2498	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
2499	* @retval VERR_INVALID_PARAMETER on invalid parameter.
2500	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
2501	*
2502	* @param pSession The SUPDRV session (must be a ring-0 session).
2503	* @param pFactory Pointer to the component factory registration structure.
2504	*
2505	* @remarks This interface is also available via SUPR0IdcComponentRegisterFactory.
2506	*/
2507	SUPR0DECL(int) SUPR0ComponentRegisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
2508	{
2509	PSUPDRVFACTORYREG pNewReg;
2510	const char *psz;
2511	int rc;
2512
2513	/*
2514	* Validate parameters.
2515	*/
2516	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2517	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
2518	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
2519	AssertPtrReturn(pFactory->pfnQueryFactoryInterface, VERR_INVALID_POINTER);
2520	psz = (const char *)memchr(pFactory->szName, '\0', sizeof(pFactory->szName));
2521	AssertReturn(psz, VERR_INVALID_PARAMETER);
2522
2523	/*
2524	* Allocate and initialize a new registration structure.
2525	*/
2526	pNewReg = (PSUPDRVFACTORYREG)RTMemAlloc(sizeof(SUPDRVFACTORYREG));
2527	if (pNewReg)
2528	{
2529	pNewReg->pNext = NULL;
2530	pNewReg->pFactory = pFactory;
2531	pNewReg->pSession = pSession;
2532	pNewReg->cchName = psz - &pFactory->szName[0];
2533
2534	/*
2535	* Add it to the tail of the list after checking for prior registration.
2536	*/
2537	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
2538	if (RT_SUCCESS(rc))
2539	{
2540	PSUPDRVFACTORYREG pPrev = NULL;
2541	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
2542	while (pCur && pCur->pFactory != pFactory)
2543	{
2544	pPrev = pCur;
2545	pCur = pCur->pNext;
2546	}
2547	if (!pCur)
2548	{
2549	if (pPrev)
2550	pPrev->pNext = pNewReg;
2551	else
2552	pSession->pDevExt->pComponentFactoryHead = pNewReg;
2553	rc = VINF_SUCCESS;
2554	}
2555	else
2556	rc = VERR_ALREADY_EXISTS;
2557
2558	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
2559	}
2560
2561	if (RT_FAILURE(rc))
2562	RTMemFree(pNewReg);
2563	}
2564	else
2565	rc = VERR_NO_MEMORY;
2566	return rc;
2567	}
2568
2569
2570	/**
2571	* Deregister a component factory.
2572	*
2573	* @returns VBox status code.
2574	* @retval VINF_SUCCESS on success.
2575	* @retval VERR_NOT_FOUND if the factory wasn't registered.
2576	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
2577	* @retval VERR_INVALID_PARAMETER on invalid parameter.
2578	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
2579	*
2580	* @param pSession The SUPDRV session (must be a ring-0 session).
2581	* @param pFactory Pointer to the component factory registration structure
2582	* previously passed SUPR0ComponentRegisterFactory().
2583	*
2584	* @remarks This interface is also available via SUPR0IdcComponentDeregisterFactory.
2585	*/
2586	SUPR0DECL(int) SUPR0ComponentDeregisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
2587	{
2588	int rc;
2589
2590	/*
2591	* Validate parameters.
2592	*/
2593	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2594	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
2595	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
2596
2597	/*
2598	* Take the lock and look for the registration record.
2599	*/
2600	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
2601	if (RT_SUCCESS(rc))
2602	{
2603	PSUPDRVFACTORYREG pPrev = NULL;
2604	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
2605	while (pCur && pCur->pFactory != pFactory)
2606	{
2607	pPrev = pCur;
2608	pCur = pCur->pNext;
2609	}
2610	if (pCur)
2611	{
2612	if (!pPrev)
2613	pSession->pDevExt->pComponentFactoryHead = pCur->pNext;
2614	else
2615	pPrev->pNext = pCur->pNext;
2616
2617	pCur->pNext = NULL;
2618	pCur->pFactory = NULL;
2619	pCur->pSession = NULL;
2620	rc = VINF_SUCCESS;
2621	}
2622	else
2623	rc = VERR_NOT_FOUND;
2624
2625	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
2626
2627	RTMemFree(pCur);
2628	}
2629	return rc;
2630	}
2631
2632
2633	/**
2634	* Queries a component factory.
2635	*
2636	* @returns VBox status code.
2637	* @retval VERR_INVALID_PARAMETER on invalid parameter.
2638	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
2639	* @retval VERR_SUPDRV_COMPONENT_NOT_FOUND if the component factory wasn't found.
2640	* @retval VERR_SUPDRV_INTERFACE_NOT_SUPPORTED if the interface wasn't supported.
2641	*
2642	* @param pSession The SUPDRV session.
2643	* @param pszName The name of the component factory.
2644	* @param pszInterfaceUuid The UUID of the factory interface (stringified).
2645	* @param ppvFactoryIf Where to store the factory interface.
2646	*/
2647	SUPR0DECL(int) SUPR0ComponentQueryFactory(PSUPDRVSESSION pSession, const char pszName, const char pszInterfaceUuid, void **ppvFactoryIf)
2648	{
2649	const char *pszEnd;
2650	size_t cchName;
2651	int rc;
2652
2653	/*
2654	* Validate parameters.
2655	*/
2656	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2657
2658	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
2659	pszEnd = memchr(pszName, '\0', RT_SIZEOFMEMB(SUPDRVFACTORY, szName));
2660	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
2661	cchName = pszEnd - pszName;
2662
2663	AssertPtrReturn(pszInterfaceUuid, VERR_INVALID_POINTER);
2664	pszEnd = memchr(pszInterfaceUuid, '\0', RTUUID_STR_LENGTH);
2665	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
2666
2667	AssertPtrReturn(ppvFactoryIf, VERR_INVALID_POINTER);
2668	*ppvFactoryIf = NULL;
2669
2670	/*
2671	* Take the lock and try all factories by this name.
2672	*/
2673	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
2674	if (RT_SUCCESS(rc))
2675	{
2676	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
2677	rc = VERR_SUPDRV_COMPONENT_NOT_FOUND;
2678	while (pCur)
2679	{
2680	if ( pCur->cchName == cchName
2681	&& !memcmp(pCur->pFactory->szName, pszName, cchName))
2682	{
2683	void *pvFactory = pCur->pFactory->pfnQueryFactoryInterface(pCur->pFactory, pSession, pszInterfaceUuid);
2684	if (pvFactory)
2685	{
2686	*ppvFactoryIf = pvFactory;
2687	rc = VINF_SUCCESS;
2688	break;
2689	}
2690	rc = VERR_SUPDRV_INTERFACE_NOT_SUPPORTED;
2691	}
2692
2693	/* next */
2694	pCur = pCur->pNext;
2695	}
2696
2697	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
2698	}
2699	return rc;
2700	}
2701
2702
2703	/**
2704	* Adds a memory object to the session.
2705	*
2706	* @returns IPRT status code.
2707	* @param pMem Memory tracking structure containing the
2708	* information to track.
2709	* @param pSession The session.
2710	*/
2711	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
2712	{
2713	PSUPDRVBUNDLE pBundle;
2714	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2715
2716	/*
2717	* Find free entry and record the allocation.
2718	*/
2719	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2720	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2721	{
2722	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
2723	{
2724	unsigned i;
2725	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2726	{
2727	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
2728	{
2729	pBundle->cUsed++;
2730	pBundle->aMem[i] = *pMem;
2731	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2732	return VINF_SUCCESS;
2733	}
2734	}
2735	AssertFailed(); /* !!this can't be happening!!! */
2736	}
2737	}
2738	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2739
2740	/*
2741	* Need to allocate a new bundle.
2742	* Insert into the last entry in the bundle.
2743	*/
2744	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
2745	if (!pBundle)
2746	return VERR_NO_MEMORY;
2747
2748	/* take last entry. */
2749	pBundle->cUsed++;
2750	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
2751
2752	/* insert into list. */
2753	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2754	pBundle->pNext = pSession->Bundle.pNext;
2755	pSession->Bundle.pNext = pBundle;
2756	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2757
2758	return VINF_SUCCESS;
2759	}
2760
2761
2762	/**
2763	* Releases a memory object referenced by pointer and type.
2764	*
2765	* @returns IPRT status code.
2766	* @param pSession Session data.
2767	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
2768	* @param eType Memory type.
2769	*/
2770	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
2771	{
2772	PSUPDRVBUNDLE pBundle;
2773	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2774
2775	/*
2776	* Validate input.
2777	*/
2778	if (!uPtr)
2779	{
2780	Log(("Illegal address %p\n", (void *)uPtr));
2781	return VERR_INVALID_PARAMETER;
2782	}
2783
2784	/*
2785	* Search for the address.
2786	*/
2787	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2788	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2789	{
2790	if (pBundle->cUsed > 0)
2791	{
2792	unsigned i;
2793	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2794	{
2795	if ( pBundle->aMem[i].eType == eType
2796	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2797	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2798	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2799	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
2800	)
2801	{
2802	/* Make a copy of it and release it outside the spinlock. */
2803	SUPDRVMEMREF Mem = pBundle->aMem[i];
2804	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
2805	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
2806	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
2807	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2808
2809	if (Mem.MapObjR3)
2810	{
2811	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
2812	AssertRC(rc); /** @todo figure out how to handle this. */
2813	}
2814	if (Mem.MemObj)
2815	{
2816	int rc = RTR0MemObjFree(Mem.MemObj, false);
2817	AssertRC(rc); /** @todo figure out how to handle this. */
2818	}
2819	return VINF_SUCCESS;
2820	}
2821	}
2822	}
2823	}
2824	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2825	Log(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
2826	return VERR_INVALID_PARAMETER;
2827	}
2828
2829
2830	#ifdef VBOX_WITH_IDT_PATCHING
2831	/**
2832	* Install IDT for the current CPU.
2833	*
2834	* @returns One of the following IPRT status codes:
2835	* @retval VINF_SUCCESS on success.
2836	* @retval VERR_IDT_FAILED.
2837	* @retval VERR_NO_MEMORY.
2838	* @param pDevExt The device extension.
2839	* @param pSession The session data.
2840	* @param pReq The request.
2841	*/
2842	static int supdrvIOCtl_IdtInstall(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPIDTINSTALL pReq)
2843	{
2844	PSUPDRVPATCHUSAGE pUsagePre;
2845	PSUPDRVPATCH pPatchPre;
2846	RTIDTR Idtr;
2847	PSUPDRVPATCH pPatch;
2848	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2849	LogFlow(("supdrvIOCtl_IdtInstall\n"));
2850
2851	/*
2852	* Preallocate entry for this CPU cause we don't wanna do
2853	* that inside the spinlock!
2854	*/
2855	pUsagePre = (PSUPDRVPATCHUSAGE)RTMemAlloc(sizeof(*pUsagePre));
2856	if (!pUsagePre)
2857	return VERR_NO_MEMORY;
2858
2859	/*
2860	* Take the spinlock and see what we need to do.
2861	*/
2862	RTSpinlockAcquireNoInts(pDevExt->Spinlock, &SpinlockTmp);
2863
2864	/* check if we already got a free patch. */
2865	if (!pDevExt->pIdtPatchesFree)
2866	{
2867	/*
2868	* Allocate a patch - outside the spinlock of course.
2869	*/
2870	RTSpinlockReleaseNoInts(pDevExt->Spinlock, &SpinlockTmp);
2871
2872	pPatchPre = (PSUPDRVPATCH)RTMemExecAlloc(sizeof(*pPatchPre));
2873	if (!pPatchPre)
2874	return VERR_NO_MEMORY;
2875
2876	RTSpinlockAcquireNoInts(pDevExt->Spinlock, &SpinlockTmp);
2877	}
2878	else
2879	{
2880	pPatchPre = pDevExt->pIdtPatchesFree;
2881	pDevExt->pIdtPatchesFree = pPatchPre->pNext;
2882	}
2883
2884	/* look for matching patch entry */
2885	ASMGetIDTR(&Idtr);
2886	pPatch = pDevExt->pIdtPatches;
2887	while (pPatch && pPatch->pvIdt != (void *)Idtr.pIdt)
2888	pPatch = pPatch->pNext;
2889
2890	if (!pPatch)
2891	{
2892	/*
2893	* Create patch.
2894	*/
2895	pPatch = supdrvIdtPatchOne(pDevExt, pPatchPre);
2896	if (pPatch)
2897	pPatchPre = NULL; /* mark as used. */
2898	}
2899	else
2900	{
2901	/*
2902	* Simply increment patch usage.
2903	*/
2904	pPatch->cUsage++;
2905	}
2906
2907	if (pPatch)
2908	{
2909	/*
2910	* Increment and add if need be the session usage record for this patch.
2911	*/
2912	PSUPDRVPATCHUSAGE pUsage = pSession->pPatchUsage;
2913	while (pUsage && pUsage->pPatch != pPatch)
2914	pUsage = pUsage->pNext;
2915
2916	if (!pUsage)
2917	{
2918	/*
2919	* Add usage record.
2920	*/
2921	pUsagePre->cUsage = 1;
2922	pUsagePre->pPatch = pPatch;
2923	pUsagePre->pNext = pSession->pPatchUsage;
2924	pSession->pPatchUsage = pUsagePre;
2925	pUsagePre = NULL; /* mark as used. */
2926	}
2927	else
2928	{
2929	/*
2930	* Increment usage count.
2931	*/
2932	pUsage->cUsage++;
2933	}
2934	}
2935
2936	/* free patch - we accumulate them for paranoid saftly reasons. */
2937	if (pPatchPre)
2938	{
2939	pPatchPre->pNext = pDevExt->pIdtPatchesFree;
2940	pDevExt->pIdtPatchesFree = pPatchPre;
2941	}
2942
2943	RTSpinlockReleaseNoInts(pDevExt->Spinlock, &SpinlockTmp);
2944
2945	/*
2946	* Free unused preallocated buffers.
2947	*/
2948	if (pUsagePre)
2949	RTMemFree(pUsagePre);
2950
2951	pReq->u.Out.u8Idt = pDevExt->u8Idt;
2952
2953	return pPatch ? VINF_SUCCESS : VERR_IDT_FAILED;
2954	}
2955
2956
2957	/**
2958	* This creates a IDT patch entry.
2959	* If the first patch being installed it'll also determin the IDT entry
2960	* to use.
2961	*
2962	* @returns pPatch on success.
2963	* @returns NULL on failure.
2964	* @param pDevExt Pointer to globals.
2965	* @param pPatch Patch entry to use.
2966	* This will be linked into SUPDRVDEVEXT::pIdtPatches on
2967	* successful return.
2968	* @remark Call must be owning the SUPDRVDEVEXT::Spinlock!
2969	*/
2970	static PSUPDRVPATCH supdrvIdtPatchOne(PSUPDRVDEVEXT pDevExt, PSUPDRVPATCH pPatch)
2971	{
2972	RTIDTR Idtr;
2973	PSUPDRVIDTE paIdt;
2974	LogFlow(("supdrvIOCtl_IdtPatchOne: pPatch=%p\n", pPatch));
2975
2976	/*
2977	* Get IDT.
2978	*/
2979	ASMGetIDTR(&Idtr);
2980	paIdt = (PSUPDRVIDTE)Idtr.pIdt;
2981	/*
2982	* Recent Linux kernels can be configured to 1G user /3G kernel.
2983	*/
2984	if ((uintptr_t)paIdt < 0x40000000)
2985	{
2986	AssertMsgFailed(("bad paIdt=%p\n", paIdt));
2987	return NULL;
2988	}
2989
2990	if (!pDevExt->u8Idt)
2991	{
2992	/*
2993	* Test out the alternatives.
2994	*
2995	* At the moment we do not support chaining thus we ASSUME that one of
2996	* these 48 entries is unused (which is not a problem on Win32 and
2997	* Linux to my knowledge).
2998	*/
2999	/** @todo we MUST change this detection to try grab an entry which is NOT in use. This can be
3000	* combined with gathering info about which guest system call gates we can hook up directly. */
3001	unsigned i;
3002	uint8_t u8Idt = 0;
3003	static uint8_t au8Ints[] =
3004	{
3005	#ifdef RT_OS_WINDOWS /* We don't use 0xef and above because they are system stuff on linux (ef is IPI,
3006	* local apic timer, or some other frequently fireing thing). */
3007	0xef, 0xee, 0xed, 0xec,
3008	#endif
3009	0xeb, 0xea, 0xe9, 0xe8,
3010	0xdf, 0xde, 0xdd, 0xdc,
3011	0x7b, 0x7a, 0x79, 0x78,
3012	0xbf, 0xbe, 0xbd, 0xbc,
3013	};
3014	#if defined(RT_ARCH_AMD64) && defined(DEBUG)
3015	static int s_iWobble = 0;
3016	unsigned iMax = !(s_iWobble++ % 2) ? 0x80 : 0x100;
3017	Log2(("IDT: Idtr=%p:%#x\n", (void *)Idtr.pIdt, (unsigned)Idtr.cbIdt));
3018	for (i = iMax - 0x80; i*16+15 < Idtr.cbIdt && i < iMax; i++)
3019	{
3020	Log2(("%#x: %04x:%08x%04x%04x P=%d DPL=%d IST=%d Type1=%#x u32Reserved=%#x u5Reserved=%#x\n",
3021	i, paIdt[i].u16SegSel, paIdt[i].u32OffsetTop, paIdt[i].u16OffsetHigh, paIdt[i].u16OffsetLow,
3022	paIdt[i].u1Present, paIdt[i].u2DPL, paIdt[i].u3IST, paIdt[i].u5Type2,
3023	paIdt[i].u32Reserved, paIdt[i].u5Reserved));
3024	}
3025	#endif
3026	/* look for entries which are not present or otherwise unused. */
3027	for (i = 0; i < sizeof(au8Ints) / sizeof(au8Ints[0]); i++)
3028	{
3029	u8Idt = au8Ints[i];
3030	if ( u8Idt * sizeof(SUPDRVIDTE) < Idtr.cbIdt
3031	&& ( !paIdt[u8Idt].u1Present
3032	\|\| paIdt[u8Idt].u5Type2 == 0))
3033	break;
3034	u8Idt = 0;
3035	}
3036	if (!u8Idt)
3037	{
3038	/* try again, look for a compatible entry .*/
3039	for (i = 0; i < sizeof(au8Ints) / sizeof(au8Ints[0]); i++)
3040	{
3041	u8Idt = au8Ints[i];
3042	if ( u8Idt * sizeof(SUPDRVIDTE) < Idtr.cbIdt
3043	&& paIdt[u8Idt].u1Present
3044	&& paIdt[u8Idt].u5Type2 == SUPDRV_IDTE_TYPE2_INTERRUPT_GATE
3045	&& !(paIdt[u8Idt].u16SegSel & 3))
3046	break;
3047	u8Idt = 0;
3048	}
3049	if (!u8Idt)
3050	{
3051	Log(("Failed to find appropirate IDT entry!!\n"));
3052	return NULL;
3053	}
3054	}
3055	pDevExt->u8Idt = u8Idt;
3056	LogFlow(("supdrvIOCtl_IdtPatchOne: u8Idt=%x\n", u8Idt));
3057	}
3058
3059	/*
3060	* Prepare the patch
3061	*/
3062	memset(pPatch, 0, sizeof(*pPatch));
3063	pPatch->pvIdt = paIdt;
3064	pPatch->cUsage = 1;
3065	pPatch->pIdtEntry = &paIdt[pDevExt->u8Idt];
3066	pPatch->SavedIdt = paIdt[pDevExt->u8Idt];
3067	pPatch->ChangedIdt.u16OffsetLow = (uint32_t)((uintptr_t)&pPatch->auCode[0] & 0xffff);
3068	pPatch->ChangedIdt.u16OffsetHigh = (uint32_t)((uintptr_t)&pPatch->auCode[0] >> 16);
3069	#ifdef RT_ARCH_AMD64
3070	pPatch->ChangedIdt.u32OffsetTop = (uint32_t)((uintptr_t)&pPatch->auCode[0] >> 32);
3071	#endif
3072	pPatch->ChangedIdt.u16SegSel = ASMGetCS();
3073	#ifdef RT_ARCH_AMD64
3074	pPatch->ChangedIdt.u3IST = 0;
3075	pPatch->ChangedIdt.u5Reserved = 0;
3076	#else /* x86 */
3077	pPatch->ChangedIdt.u5Reserved = 0;
3078	pPatch->ChangedIdt.u3Type1 = 0;
3079	#endif /* x86 */
3080	pPatch->ChangedIdt.u5Type2 = SUPDRV_IDTE_TYPE2_INTERRUPT_GATE;
3081	pPatch->ChangedIdt.u2DPL = 3;
3082	pPatch->ChangedIdt.u1Present = 1;
3083
3084	/*
3085	* Generate the patch code.
3086	*/
3087	{
3088	#ifdef RT_ARCH_AMD64
3089	union
3090	{
3091	uint8_t *pb;
3092	uint32_t *pu32;
3093	uint64_t *pu64;
3094	} u, uFixJmp, uFixCall, uNotNested;
3095	u.pb = &pPatch->auCode[0];
3096
3097	/* check the cookie */
3098	*u.pb++ = 0x3d; // cmp eax, GLOBALCOOKIE
3099	*u.pu32++ = pDevExt->u32Cookie;
3100
3101	*u.pb++ = 0x74; // jz @VBoxCall
3102	*u.pb++ = 2;
3103
3104	/* jump to forwarder code. */
3105	*u.pb++ = 0xeb;
3106	uFixJmp = u;
3107	*u.pb++ = 0xfe;
3108
3109	// @VBoxCall:
3110	*u.pb++ = 0x0f; // swapgs
3111	*u.pb++ = 0x01;
3112	*u.pb++ = 0xf8;
3113
3114	/*
3115	* Call VMMR0Entry
3116	* We don't have to push the arguments here, but we have top
3117	* reserve some stack space for the interrupt forwarding.
3118	*/
3119	# ifdef RT_OS_WINDOWS
3120	*u.pb++ = 0x50; // push rax ; alignment filler.
3121	*u.pb++ = 0x41; // push r8 ; uArg
3122	*u.pb++ = 0x50;
3123	*u.pb++ = 0x52; // push rdx ; uOperation
3124	*u.pb++ = 0x51; // push rcx ; pVM
3125	# else
3126	*u.pb++ = 0x51; // push rcx ; alignment filler.
3127	*u.pb++ = 0x52; // push rdx ; uArg
3128	*u.pb++ = 0x56; // push rsi ; uOperation
3129	*u.pb++ = 0x57; // push rdi ; pVM
3130	# endif
3131
3132	*u.pb++ = 0xff; // call qword [pfnVMMR0EntryInt wrt rip]
3133	*u.pb++ = 0x15;
3134	uFixCall = u;
3135	*u.pu32++ = 0;
3136
3137	*u.pb++ = 0x48; // add rsp, 20h ; remove call frame.
3138	*u.pb++ = 0x81;
3139	*u.pb++ = 0xc4;
3140	*u.pu32++ = 0x20;
3141
3142	*u.pb++ = 0x0f; // swapgs
3143	*u.pb++ = 0x01;
3144	*u.pb++ = 0xf8;
3145
3146	/* Return to R3. */
3147	uNotNested = u;
3148	*u.pb++ = 0x48; // iretq
3149	*u.pb++ = 0xcf;
3150
3151	while ((uintptr_t)u.pb & 0x7) // align 8
3152	*u.pb++ = 0xcc;
3153
3154	/* Pointer to the VMMR0Entry. */ // pfnVMMR0EntryInt dq StubVMMR0Entry
3155	*uFixCall.pu32 = (uint32_t)(u.pb - uFixCall.pb - 4); uFixCall.pb = NULL;
3156	pPatch->offVMMR0EntryFixup = (uint16_t)(u.pb - &pPatch->auCode[0]);
3157	*u.pu64++ = pDevExt->pvVMMR0 ? (uint64_t)pDevExt->pfnVMMR0EntryInt : (uint64_t)u.pb + 8;
3158
3159	/* stub entry. */ // StubVMMR0Entry:
3160	pPatch->offStub = (uint16_t)(u.pb - &pPatch->auCode[0]);
3161	*u.pb++ = 0x33; // xor eax, eax
3162	*u.pb++ = 0xc0;
3163
3164	*u.pb++ = 0x48; // dec rax
3165	*u.pb++ = 0xff;
3166	*u.pb++ = 0xc8;
3167
3168	*u.pb++ = 0xc3; // ret
3169
3170	/* forward to the original handler using a retf. */
3171	*uFixJmp.pb = (uint8_t)(u.pb - uFixJmp.pb - 1); uFixJmp.pb = NULL;
3172
3173	*u.pb++ = 0x68; // push <target cs>
3174	*u.pu32++ = !pPatch->SavedIdt.u5Type2 ? ASMGetCS() : pPatch->SavedIdt.u16SegSel;
3175
3176	*u.pb++ = 0x68; // push <low target rip>
3177	*u.pu32++ = !pPatch->SavedIdt.u5Type2
3178	? (uint32_t)(uintptr_t)uNotNested.pb
3179	: (uint32_t)pPatch->SavedIdt.u16OffsetLow
3180	\| (uint32_t)pPatch->SavedIdt.u16OffsetHigh << 16;
3181
3182	*u.pb++ = 0xc7; // mov dword [rsp + 4], <high target rip>
3183	*u.pb++ = 0x44;
3184	*u.pb++ = 0x24;
3185	*u.pb++ = 0x04;
3186	*u.pu32++ = !pPatch->SavedIdt.u5Type2
3187	? (uint32_t)((uint64_t)uNotNested.pb >> 32)
3188	: pPatch->SavedIdt.u32OffsetTop;
3189
3190	*u.pb++ = 0x48; // retf ; does this require prefix?
3191	*u.pb++ = 0xcb;
3192
3193	#else /* RT_ARCH_X86 */
3194
3195	union
3196	{
3197	uint8_t *pb;
3198	uint16_t *pu16;
3199	uint32_t *pu32;
3200	} u, uFixJmpNotNested, uFixJmp, uFixCall, uNotNested;
3201	u.pb = &pPatch->auCode[0];
3202
3203	/* check the cookie */
3204	*u.pb++ = 0x81; // cmp esi, GLOBALCOOKIE
3205	*u.pb++ = 0xfe;
3206	*u.pu32++ = pDevExt->u32Cookie;
3207
3208	*u.pb++ = 0x74; // jz VBoxCall
3209	uFixJmp = u;
3210	*u.pb++ = 0;
3211
3212	/* jump (far) to the original handler / not-nested-stub. */
3213	*u.pb++ = 0xea; // jmp far NotNested
3214	uFixJmpNotNested = u;
3215	*u.pu32++ = 0;
3216	*u.pu16++ = 0;
3217
3218	/* save selector registers. */ // VBoxCall:
3219	*uFixJmp.pb = (uint8_t)(u.pb - uFixJmp.pb - 1);
3220	*u.pb++ = 0x0f; // push fs
3221	*u.pb++ = 0xa0;
3222
3223	*u.pb++ = 0x1e; // push ds
3224
3225	*u.pb++ = 0x06; // push es
3226
3227	/* call frame */
3228	*u.pb++ = 0x51; // push ecx
3229
3230	*u.pb++ = 0x52; // push edx
3231
3232	*u.pb++ = 0x50; // push eax
3233
3234	/* load ds, es and perhaps fs before call. */
3235	*u.pb++ = 0xb8; // mov eax, KernelDS
3236	*u.pu32++ = ASMGetDS();
3237
3238	*u.pb++ = 0x8e; // mov ds, eax
3239	*u.pb++ = 0xd8;
3240
3241	*u.pb++ = 0x8e; // mov es, eax
3242	*u.pb++ = 0xc0;
3243
3244	#ifdef RT_OS_WINDOWS
3245	*u.pb++ = 0xb8; // mov eax, KernelFS
3246	*u.pu32++ = ASMGetFS();
3247
3248	*u.pb++ = 0x8e; // mov fs, eax
3249	*u.pb++ = 0xe0;
3250	#endif
3251
3252	/* do the call. */
3253	*u.pb++ = 0xe8; // call _VMMR0Entry / StubVMMR0Entry
3254	uFixCall = u;
3255	pPatch->offVMMR0EntryFixup = (uint16_t)(u.pb - &pPatch->auCode[0]);
3256	*u.pu32++ = 0xfffffffb;
3257
3258	*u.pb++ = 0x83; // add esp, 0ch ; cdecl
3259	*u.pb++ = 0xc4;
3260	*u.pb++ = 0x0c;
3261
3262	/* restore selector registers. */
3263	*u.pb++ = 0x07; // pop es
3264	//
3265	*u.pb++ = 0x1f; // pop ds
3266
3267	*u.pb++ = 0x0f; // pop fs
3268	*u.pb++ = 0xa1;
3269
3270	uNotNested = u; // NotNested:
3271	*u.pb++ = 0xcf; // iretd
3272
3273	/* the stub VMMR0Entry. */ // StubVMMR0Entry:
3274	pPatch->offStub = (uint16_t)(u.pb - &pPatch->auCode[0]);
3275	*u.pb++ = 0x33; // xor eax, eax
3276	*u.pb++ = 0xc0;
3277
3278	*u.pb++ = 0x48; // dec eax
3279
3280	*u.pb++ = 0xc3; // ret
3281
3282	/* Fixup the VMMR0Entry call. */
3283	if (pDevExt->pvVMMR0)
3284	*uFixCall.pu32 = (uint32_t)pDevExt->pfnVMMR0EntryInt - (uint32_t)(uFixCall.pu32 + 1);
3285	else
3286	*uFixCall.pu32 = (uint32_t)&pPatch->auCode[pPatch->offStub] - (uint32_t)(uFixCall.pu32 + 1);
3287
3288	/* Fixup the forward / nested far jump. */
3289	if (!pPatch->SavedIdt.u5Type2)
3290	{
3291	*uFixJmpNotNested.pu32++ = (uint32_t)uNotNested.pb;
3292	*uFixJmpNotNested.pu16++ = ASMGetCS();
3293	}
3294	else
3295	{
3296	*uFixJmpNotNested.pu32++ = ((uint32_t)pPatch->SavedIdt.u16OffsetHigh << 16) \| pPatch->SavedIdt.u16OffsetLow;
3297	*uFixJmpNotNested.pu16++ = pPatch->SavedIdt.u16SegSel;
3298	}
3299	#endif /* RT_ARCH_X86 */
3300	Assert(u.pb <= &pPatch->auCode[sizeof(pPatch->auCode)]);
3301	#if 0
3302	/* dump the patch code */
3303	Log2(("patch code: %p\n", &pPatch->auCode[0]));
3304	for (uFixCall.pb = &pPatch->auCode[0]; uFixCall.pb < u.pb; uFixCall.pb++)
3305	Log2(("0x%02x,\n", *uFixCall.pb));
3306	#endif
3307	}
3308
3309	/*
3310	* Install the patch.
3311	*/
3312	supdrvIdtWrite(pPatch->pIdtEntry, &pPatch->ChangedIdt);
3313	AssertMsg(!memcmp((void *)pPatch->pIdtEntry, &pPatch->ChangedIdt, sizeof(pPatch->ChangedIdt)), ("The stupid change code didn't work!!!!!\n"));
3314
3315	/*
3316	* Link in the patch.
3317	*/
3318	pPatch->pNext = pDevExt->pIdtPatches;
3319	pDevExt->pIdtPatches = pPatch;
3320
3321	return pPatch;
3322	}
3323
3324
3325	/**
3326	* Removes the sessions IDT references.
3327	* This will uninstall our IDT patch if we left unreferenced.
3328	*
3329	* @returns VINF_SUCCESS.
3330	* @param pDevExt Device globals.
3331	* @param pSession Session data.
3332	*/
3333	static int supdrvIOCtl_IdtRemoveAll(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
3334	{
3335	PSUPDRVPATCHUSAGE pUsage;
3336	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3337	LogFlow(("supdrvIOCtl_IdtRemoveAll: pSession=%p\n", pSession));
3338
3339	/*
3340	* Take the spinlock.
3341	*/
3342	RTSpinlockAcquireNoInts(pDevExt->Spinlock, &SpinlockTmp);
3343
3344	/*
3345	* Walk usage list, removing patches as their usage count reaches zero.
3346	*/
3347	pUsage = pSession->pPatchUsage;
3348	while (pUsage)
3349	{
3350	if (pUsage->pPatch->cUsage <= pUsage->cUsage)
3351	supdrvIdtRemoveOne(pDevExt, pUsage->pPatch);
3352	else
3353	pUsage->pPatch->cUsage -= pUsage->cUsage;
3354
3355	/* next */
3356	pUsage = pUsage->pNext;
3357	}
3358
3359	/*
3360	* Empty the usage chain and we're done inside the spinlock.
3361	*/
3362	pUsage = pSession->pPatchUsage;
3363	pSession->pPatchUsage = NULL;
3364
3365	RTSpinlockReleaseNoInts(pDevExt->Spinlock, &SpinlockTmp);
3366
3367	/*
3368	* Free usage entries.
3369	*/
3370	while (pUsage)
3371	{
3372	void *pvToFree = pUsage;
3373	pUsage->cUsage = 0;
3374	pUsage->pPatch = NULL;
3375	pUsage = pUsage->pNext;
3376	RTMemFree(pvToFree);
3377	}
3378
3379	return VINF_SUCCESS;
3380	}
3381
3382
3383	/**
3384	* Remove one patch.
3385	*
3386	* Worker for supdrvIOCtl_IdtRemoveAll.
3387	*
3388	* @param pDevExt Device globals.
3389	* @param pPatch Patch entry to remove.
3390	* @remark Caller must own SUPDRVDEVEXT::Spinlock!
3391	*/
3392	static void supdrvIdtRemoveOne(PSUPDRVDEVEXT pDevExt, PSUPDRVPATCH pPatch)
3393	{
3394	LogFlow(("supdrvIdtRemoveOne: pPatch=%p\n", pPatch));
3395
3396	pPatch->cUsage = 0;
3397
3398	/*
3399	* If the IDT entry was changed it have to kick around for ever!
3400	* This will be attempted freed again, perhaps next time we'll succeed :-)
3401	*/
3402	if (memcmp((void *)pPatch->pIdtEntry, &pPatch->ChangedIdt, sizeof(pPatch->ChangedIdt)))
3403	{
3404	AssertMsgFailed(("The hijacked IDT entry has CHANGED!!!\n"));
3405	return;
3406	}
3407
3408	/*
3409	* Unlink it.
3410	*/
3411	if (pDevExt->pIdtPatches != pPatch)
3412	{
3413	PSUPDRVPATCH pPatchPrev = pDevExt->pIdtPatches;
3414	while (pPatchPrev)
3415	{
3416	if (pPatchPrev->pNext == pPatch)
3417	{
3418	pPatchPrev->pNext = pPatch->pNext;
3419	break;
3420	}
3421	pPatchPrev = pPatchPrev->pNext;
3422	}
3423	Assert(!pPatchPrev);
3424	}
3425	else
3426	pDevExt->pIdtPatches = pPatch->pNext;
3427	pPatch->pNext = NULL;
3428
3429
3430	/*
3431	* Verify and restore the IDT.
3432	*/
3433	AssertMsg(!memcmp((void *)pPatch->pIdtEntry, &pPatch->ChangedIdt, sizeof(pPatch->ChangedIdt)), ("The hijacked IDT entry has CHANGED!!!\n"));
3434	supdrvIdtWrite(pPatch->pIdtEntry, &pPatch->SavedIdt);
3435	AssertMsg(!memcmp((void *)pPatch->pIdtEntry, &pPatch->SavedIdt, sizeof(pPatch->SavedIdt)), ("The hijacked IDT entry has CHANGED!!!\n"));
3436
3437	/*
3438	* Put it in the free list.
3439	* (This free list stuff is to calm my paranoia.)
3440	*/
3441	pPatch->pvIdt = NULL;
3442	pPatch->pIdtEntry = NULL;
3443
3444	pPatch->pNext = pDevExt->pIdtPatchesFree;
3445	pDevExt->pIdtPatchesFree = pPatch;
3446	}
3447
3448
3449	/**
3450	* Write to an IDT entry.
3451	*
3452	* @param pvIdtEntry Where to write.
3453	* @param pNewIDTEntry What to write.
3454	*/
3455	static void supdrvIdtWrite(volatile void pvIdtEntry, const SUPDRVIDTE pNewIDTEntry)
3456	{
3457	RTR0UINTREG uCR0;
3458	RTR0UINTREG uFlags;
3459
3460	/*
3461	* On SMP machines (P4 hyperthreading included) we must preform a
3462	* 64-bit locked write when updating the IDT entry.
3463	*
3464	* The F00F bugfix for linux (and probably other OSes) causes
3465	* the IDT to be pointing to an readonly mapping. We get around that
3466	* by temporarily turning of WP. Since we're inside a spinlock at this
3467	* point, interrupts are disabled and there isn't any way the WP bit
3468	* flipping can cause any trouble.
3469	*/
3470
3471	/* Save & Clear interrupt flag; Save & clear WP. */
3472	uFlags = ASMGetFlags();
3473	ASMSetFlags(uFlags & ~(RTR0UINTREG)(1 << 9)); /X86_EFL_IF/
3474	Assert(!(ASMGetFlags() & (1 << 9)));
3475	uCR0 = ASMGetCR0();
3476	ASMSetCR0(uCR0 & ~(RTR0UINTREG)(1 << 16)); /X86_CR0_WP/
3477
3478	/* Update IDT Entry */
3479	#ifdef RT_ARCH_AMD64
3480	ASMAtomicXchgU128((volatile uint128_t )pvIdtEntry, (uint128_t *)(uintptr_t)pNewIDTEntry);
3481	#else
3482	ASMAtomicXchgU64((volatile uint64_t )pvIdtEntry, (uint64_t *)(uintptr_t)pNewIDTEntry);
3483	#endif
3484
3485	/* Restore CR0 & Flags */
3486	ASMSetCR0(uCR0);
3487	ASMSetFlags(uFlags);
3488	}
3489	#endif /* VBOX_WITH_IDT_PATCHING */
3490
3491
3492	/**
3493	* Opens an image. If it's the first time it's opened the call must upload
3494	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
3495	*
3496	* This is the 1st step of the loading.
3497	*
3498	* @returns IPRT status code.
3499	* @param pDevExt Device globals.
3500	* @param pSession Session data.
3501	* @param pReq The open request.
3502	*/
3503	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
3504	{
3505	PSUPDRVLDRIMAGE pImage;
3506	unsigned cb;
3507	void *pv;
3508	LogFlow(("supdrvIOCtl_LdrOpen: szName=%s cbImage=%d\n", pReq->u.In.szName, pReq->u.In.cbImage));
3509
3510	/*
3511	* Check if we got an instance of the image already.
3512	*/
3513	RTSemFastMutexRequest(pDevExt->mtxLdr);
3514	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
3515	{
3516	if (!strcmp(pImage->szName, pReq->u.In.szName))
3517	{
3518	pImage->cUsage++;
3519	pReq->u.Out.pvImageBase = pImage->pvImage;
3520	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
3521	supdrvLdrAddUsage(pSession, pImage);
3522	RTSemFastMutexRelease(pDevExt->mtxLdr);
3523	return VINF_SUCCESS;
3524	}
3525	}
3526	/* (not found - add it!) */
3527
3528	/*
3529	* Allocate memory.
3530	*/
3531	cb = pReq->u.In.cbImage + sizeof(SUPDRVLDRIMAGE) + 31;
3532	pv = RTMemExecAlloc(cb);
3533	if (!pv)
3534	{
3535	RTSemFastMutexRelease(pDevExt->mtxLdr);
3536	Log(("supdrvIOCtl_LdrOpen: RTMemExecAlloc(%u) failed\n", cb));
3537	return VERR_NO_MEMORY;
3538	}
3539
3540	/*
3541	* Setup and link in the LDR stuff.
3542	*/
3543	pImage = (PSUPDRVLDRIMAGE)pv;
3544	pImage->pvImage = RT_ALIGN_P(pImage + 1, 32);
3545	pImage->cbImage = pReq->u.In.cbImage;
3546	pImage->pfnModuleInit = NULL;
3547	pImage->pfnModuleTerm = NULL;
3548	pImage->uState = SUP_IOCTL_LDR_OPEN;
3549	pImage->cUsage = 1;
3550	strcpy(pImage->szName, pReq->u.In.szName);
3551
3552	pImage->pNext = pDevExt->pLdrImages;
3553	pDevExt->pLdrImages = pImage;
3554
3555	supdrvLdrAddUsage(pSession, pImage);
3556
3557	pReq->u.Out.pvImageBase = pImage->pvImage;
3558	pReq->u.Out.fNeedsLoading = true;
3559	RTSemFastMutexRelease(pDevExt->mtxLdr);
3560	return VINF_SUCCESS;
3561	}
3562
3563
3564	/**
3565	* Loads the image bits.
3566	*
3567	* This is the 2nd step of the loading.
3568	*
3569	* @returns IPRT status code.
3570	* @param pDevExt Device globals.
3571	* @param pSession Session data.
3572	* @param pReq The request.
3573	*/
3574	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
3575	{
3576	PSUPDRVLDRUSAGE pUsage;
3577	PSUPDRVLDRIMAGE pImage;
3578	int rc;
3579	LogFlow(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImage=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImage));
3580
3581	/*
3582	* Find the ldr image.
3583	*/
3584	RTSemFastMutexRequest(pDevExt->mtxLdr);
3585	pUsage = pSession->pLdrUsage;
3586	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3587	pUsage = pUsage->pNext;
3588	if (!pUsage)
3589	{
3590	RTSemFastMutexRelease(pDevExt->mtxLdr);
3591	Log(("SUP_IOCTL_LDR_LOAD: couldn't find image!\n"));
3592	return VERR_INVALID_HANDLE;
3593	}
3594	pImage = pUsage->pImage;
3595	if (pImage->cbImage != pReq->u.In.cbImage)
3596	{
3597	RTSemFastMutexRelease(pDevExt->mtxLdr);
3598	Log(("SUP_IOCTL_LDR_LOAD: image size mismatch!! %d(prep) != %d(load)\n", pImage->cbImage, pReq->u.In.cbImage));
3599	return VERR_INVALID_HANDLE;
3600	}
3601	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
3602	{
3603	unsigned uState = pImage->uState;
3604	RTSemFastMutexRelease(pDevExt->mtxLdr);
3605	if (uState != SUP_IOCTL_LDR_LOAD)
3606	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
3607	return SUPDRV_ERR_ALREADY_LOADED;
3608	}
3609	switch (pReq->u.In.eEPType)
3610	{
3611	case SUPLDRLOADEP_NOTHING:
3612	break;
3613	case SUPLDRLOADEP_VMMR0:
3614	if ( !pReq->u.In.EP.VMMR0.pvVMMR0
3615	\|\| !pReq->u.In.EP.VMMR0.pvVMMR0EntryInt
3616	\|\| !pReq->u.In.EP.VMMR0.pvVMMR0EntryFast
3617	\|\| !pReq->u.In.EP.VMMR0.pvVMMR0EntryEx)
3618	{
3619	RTSemFastMutexRelease(pDevExt->mtxLdr);
3620	Log(("NULL pointer: pvVMMR0=%p pvVMMR0EntryInt=%p pvVMMR0EntryFast=%p pvVMMR0EntryEx=%p!\n",
3621	pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
3622	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx));
3623	return VERR_INVALID_PARAMETER;
3624	}
3625	/** @todo validate pReq->u.In.EP.VMMR0.pvVMMR0 against pvImage! */
3626	if ( (uintptr_t)pReq->u.In.EP.VMMR0.pvVMMR0EntryInt - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage
3627	\|\| (uintptr_t)pReq->u.In.EP.VMMR0.pvVMMR0EntryFast - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage
3628	\|\| (uintptr_t)pReq->u.In.EP.VMMR0.pvVMMR0EntryEx - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage)
3629	{
3630	RTSemFastMutexRelease(pDevExt->mtxLdr);
3631	Log(("Out of range (%p LB %#x): pvVMMR0EntryInt=%p, pvVMMR0EntryFast=%p or pvVMMR0EntryEx=%p is NULL!\n",
3632	pImage->pvImage, pReq->u.In.cbImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
3633	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx));
3634	return VERR_INVALID_PARAMETER;
3635	}
3636	break;
3637	default:
3638	RTSemFastMutexRelease(pDevExt->mtxLdr);
3639	Log(("Invalid eEPType=%d\n", pReq->u.In.eEPType));
3640	return VERR_INVALID_PARAMETER;
3641	}
3642	if ( pReq->u.In.pfnModuleInit
3643	&& (uintptr_t)pReq->u.In.pfnModuleInit - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage)
3644	{
3645	RTSemFastMutexRelease(pDevExt->mtxLdr);
3646	Log(("SUP_IOCTL_LDR_LOAD: pfnModuleInit=%p is outside the image (%p %d bytes)\n",
3647	pReq->u.In.pfnModuleInit, pImage->pvImage, pReq->u.In.cbImage));
3648	return VERR_INVALID_PARAMETER;
3649	}
3650	if ( pReq->u.In.pfnModuleTerm
3651	&& (uintptr_t)pReq->u.In.pfnModuleTerm - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage)
3652	{
3653	RTSemFastMutexRelease(pDevExt->mtxLdr);
3654	Log(("SUP_IOCTL_LDR_LOAD: pfnModuleTerm=%p is outside the image (%p %d bytes)\n",
3655	pReq->u.In.pfnModuleTerm, pImage->pvImage, pReq->u.In.cbImage));
3656	return VERR_INVALID_PARAMETER;
3657	}
3658
3659	/*
3660	* Copy the memory.
3661	*/
3662	/* no need to do try/except as this is a buffered request. */
3663	memcpy(pImage->pvImage, &pReq->u.In.achImage[0], pImage->cbImage);
3664	pImage->uState = SUP_IOCTL_LDR_LOAD;
3665	pImage->pfnModuleInit = pReq->u.In.pfnModuleInit;
3666	pImage->pfnModuleTerm = pReq->u.In.pfnModuleTerm;
3667	pImage->offSymbols = pReq->u.In.offSymbols;
3668	pImage->cSymbols = pReq->u.In.cSymbols;
3669	pImage->offStrTab = pReq->u.In.offStrTab;
3670	pImage->cbStrTab = pReq->u.In.cbStrTab;
3671
3672	/*
3673	* Update any entry points.
3674	*/
3675	switch (pReq->u.In.eEPType)
3676	{
3677	default:
3678	case SUPLDRLOADEP_NOTHING:
3679	rc = VINF_SUCCESS;
3680	break;
3681	case SUPLDRLOADEP_VMMR0:
3682	rc = supdrvLdrSetR0EP(pDevExt, pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
3683	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
3684	break;
3685	}
3686
3687	/*
3688	* On success call the module initialization.
3689	*/
3690	LogFlow(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
3691	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
3692	{
3693	Log(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
3694	rc = pImage->pfnModuleInit();
3695	if (rc && pDevExt->pvVMMR0 == pImage->pvImage)
3696	supdrvLdrUnsetR0EP(pDevExt);
3697	}
3698
3699	if (rc)
3700	pImage->uState = SUP_IOCTL_LDR_OPEN;
3701
3702	RTSemFastMutexRelease(pDevExt->mtxLdr);
3703	return rc;
3704	}
3705
3706
3707	/**
3708	* Frees a previously loaded (prep'ed) image.
3709	*
3710	* @returns IPRT status code.
3711	* @param pDevExt Device globals.
3712	* @param pSession Session data.
3713	* @param pReq The request.
3714	*/
3715	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
3716	{
3717	int rc;
3718	PSUPDRVLDRUSAGE pUsagePrev;
3719	PSUPDRVLDRUSAGE pUsage;
3720	PSUPDRVLDRIMAGE pImage;
3721	LogFlow(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
3722
3723	/*
3724	* Find the ldr image.
3725	*/
3726	RTSemFastMutexRequest(pDevExt->mtxLdr);
3727	pUsagePrev = NULL;
3728	pUsage = pSession->pLdrUsage;
3729	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3730	{
3731	pUsagePrev = pUsage;
3732	pUsage = pUsage->pNext;
3733	}
3734	if (!pUsage)
3735	{
3736	RTSemFastMutexRelease(pDevExt->mtxLdr);
3737	Log(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
3738	return VERR_INVALID_HANDLE;
3739	}
3740
3741	/*
3742	* Check if we can remove anything.
3743	*/
3744	rc = VINF_SUCCESS;
3745	pImage = pUsage->pImage;
3746	if (pImage->cUsage <= 1 \|\| pUsage->cUsage <= 1)
3747	{
3748	/*
3749	* Check if there are any objects with destructors in the image, if
3750	* so leave it for the session cleanup routine so we get a chance to
3751	* clean things up in the right order and not leave them all dangling.
3752	*/
3753	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3754	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
3755	if (pImage->cUsage <= 1)
3756	{
3757	PSUPDRVOBJ pObj;
3758	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
3759	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImage))
3760	{
3761	rc = VERR_SHARING_VIOLATION; /** @todo VERR_DANGLING_OBJECTS */
3762	break;
3763	}
3764	}
3765	else
3766	{
3767	PSUPDRVUSAGE pGenUsage;
3768	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
3769	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImage))
3770	{
3771	rc = VERR_SHARING_VIOLATION; /** @todo VERR_DANGLING_OBJECTS */
3772	break;
3773	}
3774	}
3775	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
3776	if (rc == VINF_SUCCESS)
3777	{
3778	/* unlink it */
3779	if (pUsagePrev)
3780	pUsagePrev->pNext = pUsage->pNext;
3781	else
3782	pSession->pLdrUsage = pUsage->pNext;
3783
3784	/* free it */
3785	pUsage->pImage = NULL;
3786	pUsage->pNext = NULL;
3787	RTMemFree(pUsage);
3788
3789	/*
3790	* Derefrence the image.
3791	*/
3792	if (pImage->cUsage <= 1)
3793	supdrvLdrFree(pDevExt, pImage);
3794	else
3795	pImage->cUsage--;
3796	}
3797	else
3798	Log(("supdrvIOCtl_LdrFree: Dangling objects in %p/%s!\n", pImage->pvImage, pImage->szName));
3799	}
3800	else
3801	{
3802	/*
3803	* Dereference both image and usage.
3804	*/
3805	pImage->cUsage--;
3806	pUsage->cUsage--;
3807	}
3808
3809	RTSemFastMutexRelease(pDevExt->mtxLdr);
3810	return VINF_SUCCESS;
3811	}
3812
3813
3814	/**
3815	* Gets the address of a symbol in an open image.
3816	*
3817	* @returns 0 on success.
3818	* @returns SUPDRV_ERR_* on failure.
3819	* @param pDevExt Device globals.
3820	* @param pSession Session data.
3821	* @param pReq The request buffer.
3822	*/
3823	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
3824	{
3825	PSUPDRVLDRIMAGE pImage;
3826	PSUPDRVLDRUSAGE pUsage;
3827	uint32_t i;
3828	PSUPLDRSYM paSyms;
3829	const char *pchStrings;
3830	const size_t cbSymbol = strlen(pReq->u.In.szSymbol) + 1;
3831	void *pvSymbol = NULL;
3832	int rc = VERR_GENERAL_FAILURE;
3833	Log3(("supdrvIOCtl_LdrGetSymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
3834
3835	/*
3836	* Find the ldr image.
3837	*/
3838	RTSemFastMutexRequest(pDevExt->mtxLdr);
3839	pUsage = pSession->pLdrUsage;
3840	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3841	pUsage = pUsage->pNext;
3842	if (!pUsage)
3843	{
3844	RTSemFastMutexRelease(pDevExt->mtxLdr);
3845	Log(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
3846	return VERR_INVALID_HANDLE;
3847	}
3848	pImage = pUsage->pImage;
3849	if (pImage->uState != SUP_IOCTL_LDR_LOAD)
3850	{
3851	unsigned uState = pImage->uState;
3852	RTSemFastMutexRelease(pDevExt->mtxLdr);
3853	Log(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", uState, uState)); NOREF(uState);
3854	return VERR_ALREADY_LOADED;
3855	}
3856
3857	/*
3858	* Search the symbol strings.
3859	*/
3860	pchStrings = (const char )((uint8_t )pImage->pvImage + pImage->offStrTab);
3861	paSyms = (PSUPLDRSYM)((uint8_t *)pImage->pvImage + pImage->offSymbols);
3862	for (i = 0; i < pImage->cSymbols; i++)
3863	{
3864	if ( paSyms[i].offSymbol < pImage->cbImage /* paranoia */
3865	&& paSyms[i].offName + cbSymbol <= pImage->cbStrTab
3866	&& !memcmp(pchStrings + paSyms[i].offName, pReq->u.In.szSymbol, cbSymbol))
3867	{
3868	pvSymbol = (uint8_t *)pImage->pvImage + paSyms[i].offSymbol;
3869	rc = VINF_SUCCESS;
3870	break;
3871	}
3872	}
3873	RTSemFastMutexRelease(pDevExt->mtxLdr);
3874	pReq->u.Out.pvSymbol = pvSymbol;
3875	return rc;
3876	}
3877
3878
3879	/**
3880	* Gets the address of a symbol in an open image or the support driver.
3881	*
3882	* @returns VINF_SUCCESS on success.
3883	* @returns
3884	* @param pDevExt Device globals.
3885	* @param pSession Session data.
3886	* @param pReq The request buffer.
3887	*/
3888	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq)
3889	{
3890	int rc = VINF_SUCCESS;
3891	const char *pszSymbol = pReq->u.In.pszSymbol;
3892	const char *pszModule = pReq->u.In.pszModule;
3893	size_t cbSymbol;
3894	char const *pszEnd;
3895	uint32_t i;
3896
3897	/*
3898	* Input validation.
3899	*/
3900	AssertPtrReturn(pszSymbol, VERR_INVALID_POINTER);
3901	pszEnd = (char *)memchr(pszSymbol, '\0', 512);
3902	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3903	cbSymbol = pszEnd - pszSymbol + 1;
3904
3905	if (pszModule)
3906	{
3907	AssertPtrReturn(pszModule, VERR_INVALID_POINTER);
3908	pszEnd = (char *)memchr(pszModule, '\0', 64);
3909	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3910	}
3911	Log3(("supdrvIDC_LdrGetSymbol: pszModule=%p:{%s} pszSymbol=%p:{%s}\n", pszModule, pszModule, pszSymbol, pszSymbol));
3912
3913
3914	if ( !pszModule
3915	\|\| !strcmp(pszModule, "SupDrv"))
3916	{
3917	/*
3918	* Search the support driver export table.
3919	*/
3920	for (i = 0; i < RT_ELEMENTS(g_aFunctions); i++)
3921	if (!strcmp(g_aFunctions[i].szName, pszSymbol))
3922	{
3923	pReq->u.Out.pfnSymbol = g_aFunctions[i].pfn;
3924	break;
3925	}
3926	}
3927	else
3928	{
3929	/*
3930	* Find the loader image.
3931	*/
3932	PSUPDRVLDRIMAGE pImage;
3933
3934	RTSemFastMutexRequest(pDevExt->mtxLdr);
3935
3936	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
3937	if (!strcmp(pImage->szName, pszModule))
3938	break;
3939	if (pImage && pImage->uState == SUP_IOCTL_LDR_LOAD)
3940	{
3941	/*
3942	* Search the symbol strings.
3943	*/
3944	const char pchStrings = (const char )((uint8_t *)pImage->pvImage + pImage->offStrTab);
3945	PCSUPLDRSYM paSyms = (PCSUPLDRSYM)((uint8_t *)pImage->pvImage + pImage->offSymbols);
3946	for (i = 0; i < pImage->cSymbols; i++)
3947	{
3948	if ( paSyms[i].offSymbol < pImage->cbImage /* paranoia */
3949	&& paSyms[i].offName + cbSymbol <= pImage->cbStrTab
3950	&& !memcmp(pchStrings + paSyms[i].offName, pszSymbol, cbSymbol))
3951	{
3952	/*
3953	* Found it! Calc the symbol address and add a reference to the module.
3954	*/
3955	pReq->u.Out.pfnSymbol = (PFNRT)((uint8_t *)pImage->pvImage + paSyms[i].offSymbol);
3956	rc = supdrvLdrAddUsage(pSession, pImage);
3957	break;
3958	}
3959	}
3960	}
3961	else
3962	rc = pImage ? VERR_WRONG_ORDER : VERR_MODULE_NOT_FOUND;
3963
3964	RTSemFastMutexRelease(pDevExt->mtxLdr);
3965	}
3966	return rc;
3967	}
3968
3969
3970	/**
3971	* Updates the IDT patches to point to the specified VMM R0 entry
3972	* point (i.e. VMMR0Enter()).
3973	*
3974	* @returns IPRT status code.
3975	* @param pDevExt Device globals.
3976	* @param pSession Session data.
3977	* @param pVMMR0 VMMR0 image handle.
3978	* @param pvVMMR0EntryInt VMMR0EntryInt address.
3979	* @param pvVMMR0EntryFast VMMR0EntryFast address.
3980	* @param pvVMMR0EntryEx VMMR0EntryEx address.
3981	* @remark Caller must own the loader mutex.
3982	*/
3983	static int supdrvLdrSetR0EP(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx)
3984	{
3985	int rc = VINF_SUCCESS;
3986	LogFlow(("supdrvLdrSetR0EP pvVMMR0=%p pvVMMR0EntryInt=%p\n", pvVMMR0, pvVMMR0EntryInt));
3987
3988
3989	/*
3990	* Check if not yet set.
3991	*/
3992	if (!pDevExt->pvVMMR0)
3993	{
3994	#ifdef VBOX_WITH_IDT_PATCHING
3995	PSUPDRVPATCH pPatch;
3996	#endif
3997
3998	/*
3999	* Set it and update IDT patch code.
4000	*/
4001	pDevExt->pvVMMR0 = pvVMMR0;
4002	pDevExt->pfnVMMR0EntryInt = pvVMMR0EntryInt;
4003	pDevExt->pfnVMMR0EntryFast = pvVMMR0EntryFast;
4004	pDevExt->pfnVMMR0EntryEx = pvVMMR0EntryEx;
4005	#ifdef VBOX_WITH_IDT_PATCHING
4006	for (pPatch = pDevExt->pIdtPatches; pPatch; pPatch = pPatch->pNext)
4007	{
4008	# ifdef RT_ARCH_AMD64
4009	ASMAtomicXchgU64((volatile uint64_t *)&pPatch->auCode[pPatch->offVMMR0EntryFixup], (uint64_t)pvVMMR0);
4010	# else /* RT_ARCH_X86 */
4011	ASMAtomicXchgU32((volatile uint32_t *)&pPatch->auCode[pPatch->offVMMR0EntryFixup],
4012	(uint32_t)pvVMMR0 - (uint32_t)&pPatch->auCode[pPatch->offVMMR0EntryFixup + 4]);
4013	# endif
4014	}
4015	#endif /* VBOX_WITH_IDT_PATCHING */
4016	}
4017	else
4018	{
4019	/*
4020	* Return failure or success depending on whether the values match or not.
4021	*/
4022	if ( pDevExt->pvVMMR0 != pvVMMR0
4023	\|\| (void *)pDevExt->pfnVMMR0EntryInt != pvVMMR0EntryInt
4024	\|\| (void *)pDevExt->pfnVMMR0EntryFast != pvVMMR0EntryFast
4025	\|\| (void *)pDevExt->pfnVMMR0EntryEx != pvVMMR0EntryEx)
4026	{
4027	AssertMsgFailed(("SUP_IOCTL_LDR_SETR0EP: Already set pointing to a different module!\n"));
4028	rc = VERR_INVALID_PARAMETER;
4029	}
4030	}
4031	return rc;
4032	}
4033
4034
4035	/**
4036	* Unsets the R0 entry point installed by supdrvLdrSetR0EP.
4037	*
4038	* @param pDevExt Device globals.
4039	*/
4040	static void supdrvLdrUnsetR0EP(PSUPDRVDEVEXT pDevExt)
4041	{
4042	#ifdef VBOX_WITH_IDT_PATCHING
4043	PSUPDRVPATCH pPatch;
4044	#endif
4045
4046	pDevExt->pvVMMR0 = NULL;
4047	pDevExt->pfnVMMR0EntryInt = NULL;
4048	pDevExt->pfnVMMR0EntryFast = NULL;
4049	pDevExt->pfnVMMR0EntryEx = NULL;
4050
4051	#ifdef VBOX_WITH_IDT_PATCHING
4052	for (pPatch = pDevExt->pIdtPatches; pPatch; pPatch = pPatch->pNext)
4053	{
4054	# ifdef RT_ARCH_AMD64
4055	ASMAtomicXchgU64((volatile uint64_t *)&pPatch->auCode[pPatch->offVMMR0EntryFixup],
4056	(uint64_t)&pPatch->auCode[pPatch->offStub]);
4057	# else /* RT_ARCH_X86 */
4058	ASMAtomicXchgU32((volatile uint32_t *)&pPatch->auCode[pPatch->offVMMR0EntryFixup],
4059	(uint32_t)&pPatch->auCode[pPatch->offStub] - (uint32_t)&pPatch->auCode[pPatch->offVMMR0EntryFixup + 4]);
4060	# endif
4061	}
4062	#endif /* VBOX_WITH_IDT_PATCHING */
4063	}
4064
4065
4066	/**
4067	* Adds a usage reference in the specified session of an image.
4068	*
4069	* Called while owning the loader semaphore.
4070	*
4071	* @returns VINF_SUCCESS on success and VERR_NO_MEMORY on failure.
4072	* @param pSession Session in question.
4073	* @param pImage Image which the session is using.
4074	*/
4075	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage)
4076	{
4077	PSUPDRVLDRUSAGE pUsage;
4078	LogFlow(("supdrvLdrAddUsage: pImage=%p\n", pImage));
4079
4080	/*
4081	* Referenced it already?
4082	*/
4083	pUsage = pSession->pLdrUsage;
4084	while (pUsage)
4085	{
4086	if (pUsage->pImage == pImage)
4087	{
4088	pUsage->cUsage++;
4089	return VINF_SUCCESS;
4090	}
4091	pUsage = pUsage->pNext;
4092	}
4093
4094	/*
4095	* Allocate new usage record.
4096	*/
4097	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
4098	AssertReturn(pUsage, VERR_NO_MEMORY);
4099	pUsage->cUsage = 1;
4100	pUsage->pImage = pImage;
4101	pUsage->pNext = pSession->pLdrUsage;
4102	pSession->pLdrUsage = pUsage;
4103	return VINF_SUCCESS;
4104	}
4105
4106
4107	/**
4108	* Frees a load image.
4109	*
4110	* @param pDevExt Pointer to device extension.
4111	* @param pImage Pointer to the image we're gonna free.
4112	* This image must exit!
4113	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
4114	*/
4115	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
4116	{
4117	PSUPDRVLDRIMAGE pImagePrev;
4118	LogFlow(("supdrvLdrFree: pImage=%p\n", pImage));
4119
4120	/* find it - arg. should've used doubly linked list. */
4121	Assert(pDevExt->pLdrImages);
4122	pImagePrev = NULL;
4123	if (pDevExt->pLdrImages != pImage)
4124	{
4125	pImagePrev = pDevExt->pLdrImages;
4126	while (pImagePrev->pNext != pImage)
4127	pImagePrev = pImagePrev->pNext;
4128	Assert(pImagePrev->pNext == pImage);
4129	}
4130
4131	/* unlink */
4132	if (pImagePrev)
4133	pImagePrev->pNext = pImage->pNext;
4134	else
4135	pDevExt->pLdrImages = pImage->pNext;
4136
4137	/* check if this is VMMR0.r0 and fix the Idt patches if it is. */
4138	if (pDevExt->pvVMMR0 == pImage->pvImage)
4139	supdrvLdrUnsetR0EP(pDevExt);
4140
4141	/* check for objects with destructors in this image. (Shouldn't happen.) */
4142	if (pDevExt->pObjs)
4143	{
4144	unsigned cObjs = 0;
4145	PSUPDRVOBJ pObj;
4146	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
4147	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
4148	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4149	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImage))
4150	{
4151	pObj->pfnDestructor = NULL;
4152	cObjs++;
4153	}
4154	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
4155	if (cObjs)
4156	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
4157	}
4158
4159	/* call termination function if fully loaded. */
4160	if ( pImage->pfnModuleTerm
4161	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
4162	{
4163	LogFlow(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
4164	pImage->pfnModuleTerm();
4165	}
4166
4167	/* free the image */
4168	pImage->cUsage = 0;
4169	pImage->pNext = 0;
4170	pImage->uState = SUP_IOCTL_LDR_FREE;
4171	RTMemExecFree(pImage);
4172	}
4173
4174
4175	/**
4176	* Gets the current paging mode of the CPU and stores in in pOut.
4177	*/
4178	static SUPPAGINGMODE supdrvIOCtl_GetPagingMode(void)
4179	{
4180	SUPPAGINGMODE enmMode;
4181
4182	RTR0UINTREG cr0 = ASMGetCR0();
4183	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
4184	enmMode = SUPPAGINGMODE_INVALID;
4185	else
4186	{
4187	RTR0UINTREG cr4 = ASMGetCR4();
4188	uint32_t fNXEPlusLMA = 0;
4189	if (cr4 & X86_CR4_PAE)
4190	{
4191	uint32_t fAmdFeatures = ASMCpuId_EDX(0x80000001);
4192	if (fAmdFeatures & (X86_CPUID_AMD_FEATURE_EDX_NX \| X86_CPUID_AMD_FEATURE_EDX_LONG_MODE))
4193	{
4194	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
4195	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
4196	fNXEPlusLMA \|= RT_BIT(0);
4197	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
4198	fNXEPlusLMA \|= RT_BIT(1);
4199	}
4200	}
4201
4202	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
4203	{
4204	case 0:
4205	enmMode = SUPPAGINGMODE_32_BIT;
4206	break;
4207
4208	case X86_CR4_PGE:
4209	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
4210	break;
4211
4212	case X86_CR4_PAE:
4213	enmMode = SUPPAGINGMODE_PAE;
4214	break;
4215
4216	case X86_CR4_PAE \| RT_BIT(0):
4217	enmMode = SUPPAGINGMODE_PAE_NX;
4218	break;
4219
4220	case X86_CR4_PAE \| X86_CR4_PGE:
4221	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
4222	break;
4223
4224	case X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
4225	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
4226	break;
4227
4228	case RT_BIT(1) \| X86_CR4_PAE:
4229	enmMode = SUPPAGINGMODE_AMD64;
4230	break;
4231
4232	case RT_BIT(1) \| X86_CR4_PAE \| RT_BIT(0):
4233	enmMode = SUPPAGINGMODE_AMD64_NX;
4234	break;
4235
4236	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
4237	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
4238	break;
4239
4240	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
4241	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
4242	break;
4243
4244	default:
4245	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
4246	enmMode = SUPPAGINGMODE_INVALID;
4247	break;
4248	}
4249	}
4250	return enmMode;
4251	}
4252
4253
4254	/**
4255	* Creates the GIP.
4256	*
4257	* @returns negative errno.
4258	* @param pDevExt Instance data. GIP stuff may be updated.
4259	*/
4260	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
4261	{
4262	PSUPGLOBALINFOPAGE pGip;
4263	RTHCPHYS HCPhysGip;
4264	uint32_t u32SystemResolution;
4265	uint32_t u32Interval;
4266	int rc;
4267
4268	LogFlow(("supdrvGipCreate:\n"));
4269
4270	/* assert order */
4271	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
4272	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
4273	Assert(!pDevExt->pGipTimer);
4274
4275	/*
4276	* Allocate a suitable page with a default kernel mapping.
4277	*/
4278	rc = RTR0MemObjAllocLow(&pDevExt->GipMemObj, PAGE_SIZE, false);
4279	if (RT_FAILURE(rc))
4280	{
4281	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
4282	return rc;
4283	}
4284	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
4285	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
4286
4287	#if 0 /** @todo Disabled this as we didn't used to do it before and causes unnecessary stress on laptops.
4288	* It only applies to Windows and should probably revisited later, if possible made part of the
4289	* timer code (return min granularity in RTTimerGetSystemGranularity and set it in RTTimerStart). */
4290	/*
4291	* Try bump up the system timer resolution.
4292	* The more interrupts the better...
4293	*/
4294	if ( RT_SUCCESS(RTTimerRequestSystemGranularity( 488281 /* 2048 HZ */, &u32SystemResolution))
4295	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 500000 /* 2000 HZ */, &u32SystemResolution))
4296	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
4297	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
4298	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 1953125 /* 512 HZ */, &u32SystemResolution))
4299	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 2000000 /* 500 HZ */, &u32SystemResolution))
4300	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 3906250 /* 256 HZ */, &u32SystemResolution))
4301	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 4000000 /* 250 HZ */, &u32SystemResolution))
4302	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 7812500 /* 128 HZ */, &u32SystemResolution))
4303	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity(10000000 /* 100 HZ */, &u32SystemResolution))
4304	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity(15625000 /* 64 HZ */, &u32SystemResolution))
4305	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity(31250000 /* 32 HZ */, &u32SystemResolution))
4306	)
4307	{
4308	Assert(RTTimerGetSystemGranularity() <= u32SystemResolution);
4309	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
4310	}
4311	#endif
4312
4313	/*
4314	* Find a reasonable update interval and initialize the structure.
4315	*/
4316	u32Interval = u32SystemResolution = RTTimerGetSystemGranularity();
4317	while (u32Interval < 10000000 /* 10 ms */)
4318	u32Interval += u32SystemResolution;
4319
4320	supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), 1000000000 / u32Interval /=Hz/);
4321
4322	/*
4323	* Create the timer.
4324	* If CPU_ALL isn't supported we'll have to fall back to synchronous mode.
4325	*/
4326	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
4327	{
4328	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, RTTIMER_FLAGS_CPU_ALL, supdrvGipAsyncTimer, pDevExt);
4329	if (rc == VERR_NOT_SUPPORTED)
4330	{
4331	OSDBGPRINT(("supdrvGipCreate: omni timer not supported, falling back to synchronous mode\n"));
4332	pGip->u32Mode = SUPGIPMODE_SYNC_TSC;
4333	}
4334	}
4335	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
4336	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0, supdrvGipSyncTimer, pDevExt);
4337	if (RT_SUCCESS(rc))
4338	{
4339	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
4340	rc = RTMpNotificationRegister(supdrvGipMpEvent, pDevExt);
4341	if (RT_SUCCESS(rc))
4342	{
4343	/*
4344	* We're good.
4345	*/
4346	dprintf(("supdrvGipCreate: %ld ns interval.\n", (long)u32Interval));
4347	return VINF_SUCCESS;
4348	}
4349
4350	OSDBGPRINT(("supdrvGipCreate: failed register MP event notfication. rc=%d\n", rc));
4351	}
4352	else
4353	{
4354	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %ld ns interval. rc=%d\n", (long)u32Interval, rc));
4355	Assert(!pDevExt->pGipTimer);
4356	}
4357	supdrvGipDestroy(pDevExt);
4358	return rc;
4359	}
4360
4361
4362	/**
4363	* Terminates the GIP.
4364	*
4365	* @param pDevExt Instance data. GIP stuff may be updated.
4366	*/
4367	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
4368	{
4369	int rc;
4370	#ifdef DEBUG_DARWIN_GIP
4371	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
4372	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
4373	pDevExt->pGipTimer, pDevExt->GipMemObj));
4374	#endif
4375
4376	/*
4377	* Invalid the GIP data.
4378	*/
4379	if (pDevExt->pGip)
4380	{
4381	supdrvGipTerm(pDevExt->pGip);
4382	pDevExt->pGip = NULL;
4383	}
4384
4385	/*
4386	* Destroy the timer and free the GIP memory object.
4387	*/
4388	if (pDevExt->pGipTimer)
4389	{
4390	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
4391	pDevExt->pGipTimer = NULL;
4392	}
4393
4394	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
4395	{
4396	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
4397	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
4398	}
4399
4400	/*
4401	* Finally, release the system timer resolution request if one succeeded.
4402	*/
4403	if (pDevExt->u32SystemTimerGranularityGrant)
4404	{
4405	rc = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
4406	pDevExt->u32SystemTimerGranularityGrant = 0;
4407	}
4408	}
4409
4410
4411	/**
4412	* Timer callback function sync GIP mode.
4413	* @param pTimer The timer.
4414	* @param pvUser The device extension.
4415	*/
4416	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
4417	{
4418	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4419	supdrvGipUpdate(pDevExt->pGip, RTTimeSystemNanoTS());
4420	}
4421
4422
4423	/**
4424	* Timer callback function for async GIP mode.
4425	* @param pTimer The timer.
4426	* @param pvUser The device extension.
4427	*/
4428	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
4429	{
4430	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4431	RTCPUID idCpu = RTMpCpuId();
4432	uint64_t NanoTS = RTTimeSystemNanoTS();
4433
4434	/** @todo reset the transaction number and whatnot when iTick == 1. */
4435	if (pDevExt->idGipMaster == idCpu)
4436	supdrvGipUpdate(pDevExt->pGip, NanoTS);
4437	else
4438	supdrvGipUpdatePerCpu(pDevExt->pGip, NanoTS, ASMGetApicId());
4439	}
4440
4441
4442	/**
4443	* Multiprocessor event notification callback.
4444	*
4445	* This is used to make sue that the GIP master gets passed on to
4446	* another CPU.
4447	*
4448	* @param enmEvent The event.
4449	* @param idCpu The cpu it applies to.
4450	* @param pvUser Pointer to the device extension.
4451	*/
4452	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser)
4453	{
4454	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4455	if (enmEvent == RTMPEVENT_OFFLINE)
4456	{
4457	RTCPUID idGipMaster;
4458	ASMAtomicReadSize(&pDevExt->idGipMaster, &idGipMaster);
4459	if (idGipMaster == idCpu)
4460	{
4461	/*
4462	* Find a new GIP master.
4463	*/
4464	bool fIgnored;
4465	unsigned i;
4466	RTCPUID idNewGipMaster = NIL_RTCPUID;
4467	RTCPUSET OnlineCpus;
4468	RTMpGetOnlineSet(&OnlineCpus);
4469
4470	for (i = 0; i < RTCPUSET_MAX_CPUS; i++)
4471	{
4472	RTCPUID idCurCpu = RTMpCpuIdFromSetIndex(i);
4473	if ( RTCpuSetIsMember(&OnlineCpus, idCurCpu)
4474	&& idCurCpu != idGipMaster)
4475	{
4476	idNewGipMaster = idCurCpu;
4477	break;
4478	}
4479	}
4480
4481	dprintf(("supdrvGipMpEvent: Gip master %#lx -> %#lx\n", (long)idGipMaster, (long)idNewGipMaster));
4482	ASMAtomicCmpXchgSize(&pDevExt->idGipMaster, idNewGipMaster, idGipMaster, fIgnored);
4483	NOREF(fIgnored);
4484	}
4485	}
4486	}
4487
4488
4489	/**
4490	* Initializes the GIP data.
4491	*
4492	* @returns IPRT status code.
4493	* @param pDevExt Pointer to the device instance data.
4494	* @param pGip Pointer to the read-write kernel mapping of the GIP.
4495	* @param HCPhys The physical address of the GIP.
4496	* @param u64NanoTS The current nanosecond timestamp.
4497	* @param uUpdateHz The update freqence.
4498	*/
4499	int VBOXCALL supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys, uint64_t u64NanoTS, unsigned uUpdateHz)
4500	{
4501	unsigned i;
4502	#ifdef DEBUG_DARWIN_GIP
4503	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz));
4504	#else
4505	LogFlow(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz));
4506	#endif
4507
4508	/*
4509	* Initialize the structure.
4510	*/
4511	memset(pGip, 0, PAGE_SIZE);
4512	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
4513	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
4514	pGip->u32Mode = supdrvGipDeterminTscMode(pDevExt);
4515	pGip->u32UpdateHz = uUpdateHz;
4516	pGip->u32UpdateIntervalNS = 1000000000 / uUpdateHz;
4517	pGip->u64NanoTSLastUpdateHz = u64NanoTS;
4518
4519	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
4520	{
4521	pGip->aCPUs[i].u32TransactionId = 2;
4522	pGip->aCPUs[i].u64NanoTS = u64NanoTS;
4523	pGip->aCPUs[i].u64TSC = ASMReadTSC();
4524
4525	/*
4526	* We don't know the following values until we've executed updates.
4527	* So, we'll just insert very high values.
4528	*/
4529	pGip->aCPUs[i].u64CpuHz = _4G + 1;
4530	pGip->aCPUs[i].u32UpdateIntervalTSC = _2G / 4;
4531	pGip->aCPUs[i].au32TSCHistory[0] = _2G / 4;
4532	pGip->aCPUs[i].au32TSCHistory[1] = _2G / 4;
4533	pGip->aCPUs[i].au32TSCHistory[2] = _2G / 4;
4534	pGip->aCPUs[i].au32TSCHistory[3] = _2G / 4;
4535	pGip->aCPUs[i].au32TSCHistory[4] = _2G / 4;
4536	pGip->aCPUs[i].au32TSCHistory[5] = _2G / 4;
4537	pGip->aCPUs[i].au32TSCHistory[6] = _2G / 4;
4538	pGip->aCPUs[i].au32TSCHistory[7] = _2G / 4;
4539	}
4540
4541	/*
4542	* Link it to the device extension.
4543	*/
4544	pDevExt->pGip = pGip;
4545	pDevExt->HCPhysGip = HCPhys;
4546	pDevExt->cGipUsers = 0;
4547
4548	return VINF_SUCCESS;
4549	}
4550
4551
4552	/**
4553	* Callback used by supdrvDetermineAsyncTSC to read the TSC on a CPU.
4554	*
4555	* @param idCpu Ignored.
4556	* @param pvUser1 Where to put the TSC.
4557	* @param pvUser2 Ignored.
4558	*/
4559	static DECLCALLBACK(void) supdrvDetermineAsyncTscWorker(RTCPUID idCpu, void pvUser1, void pvUser2)
4560	{
4561	#if 1
4562	ASMAtomicWriteU64((uint64_t volatile *)pvUser1, ASMReadTSC());
4563	#else
4564	(uint64_t )pvUser1 = ASMReadTSC();
4565	#endif
4566	}
4567
4568
4569	/**
4570	* Determine if Async GIP mode is required because of TSC drift.
4571	*
4572	* When using the default/normal timer code it is essential that the time stamp counter
4573	* (TSC) runs never backwards, that is, a read operation to the counter should return
4574	* a bigger value than any previous read operation. This is guaranteed by the latest
4575	* AMD CPUs and by newer Intel CPUs which never enter the C2 state (P4). In any other
4576	* case we have to choose the asynchronous timer mode.
4577	*
4578	* @param poffMin Pointer to the determined difference between different cores.
4579	* @return false if the time stamp counters appear to be synchron, true otherwise.
4580	*/
4581	bool VBOXCALL supdrvDetermineAsyncTsc(uint64_t *poffMin)
4582	{
4583	/*
4584	* Just iterate all the cpus 8 times and make sure that the TSC is
4585	* ever increasing. We don't bother taking TSC rollover into account.
4586	*/
4587	RTCPUSET CpuSet;
4588	int iLastCpu = RTCpuLastIndex(RTMpGetSet(&CpuSet));
4589	int iCpu;
4590	int cLoops = 8;
4591	bool fAsync = false;
4592	int rc = VINF_SUCCESS;
4593	uint64_t offMax = 0;
4594	uint64_t offMin = ~(uint64_t)0;
4595	uint64_t PrevTsc = ASMReadTSC();
4596
4597	while (cLoops-- > 0)
4598	{
4599	for (iCpu = 0; iCpu <= iLastCpu; iCpu++)
4600	{
4601	uint64_t CurTsc;
4602	rc = RTMpOnSpecific(RTMpCpuIdFromSetIndex(iCpu), supdrvDetermineAsyncTscWorker, &CurTsc, NULL);
4603	if (RT_SUCCESS(rc))
4604	{
4605	if (CurTsc <= PrevTsc)
4606	{
4607	fAsync = true;
4608	offMin = offMax = PrevTsc - CurTsc;
4609	dprintf(("supdrvDetermineAsyncTsc: iCpu=%d cLoops=%d CurTsc=%llx PrevTsc=%llx\n",
4610	iCpu, cLoops, CurTsc, PrevTsc));
4611	break;
4612	}
4613
4614	/* Gather statistics (except the first time). */
4615	if (iCpu != 0 \|\| cLoops != 7)
4616	{
4617	uint64_t off = CurTsc - PrevTsc;
4618	if (off < offMin)
4619	offMin = off;
4620	if (off > offMax)
4621	offMax = off;
4622	dprintf2(("%d/%d: off=%llx\n", cLoops, iCpu, off));
4623	}
4624
4625	/* Next */
4626	PrevTsc = CurTsc;
4627	}
4628	else if (rc == VERR_NOT_SUPPORTED)
4629	break;
4630	else
4631	AssertMsg(rc == VERR_CPU_NOT_FOUND \|\| rc == VERR_CPU_OFFLINE, ("%d\n", rc));
4632	}
4633
4634	/* broke out of the loop. */
4635	if (iCpu <= iLastCpu)
4636	break;
4637	}
4638
4639	poffMin = offMin; / Almost RTMpOnSpecific profiling. */
4640	dprintf(("supdrvDetermineAsyncTsc: returns %d; iLastCpu=%d rc=%d offMin=%llx offMax=%llx\n",
4641	fAsync, iLastCpu, rc, offMin, offMax));
4642	#if !defined(RT_OS_SOLARIS) && !defined(RT_OS_OS2) && !defined(RT_OS_WINDOWS)
4643	OSDBGPRINT(("vboxdrv: fAsync=%d offMin=%#lx offMax=%#lx\n", fAsync, (long)offMin, (long)offMax));
4644	#endif
4645	return fAsync;
4646	}
4647
4648
4649	/**
4650	* Determin the GIP TSC mode.
4651	*
4652	* @returns The most suitable TSC mode.
4653	* @param pDevExt Pointer to the device instance data.
4654	*/
4655	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt)
4656	{
4657	/*
4658	* On SMP we're faced with two problems:
4659	* (1) There might be a skew between the CPU, so that cpu0
4660	* returns a TSC that is sligtly different from cpu1.
4661	* (2) Power management (and other things) may cause the TSC
4662	* to run at a non-constant speed, and cause the speed
4663	* to be different on the cpus. This will result in (1).
4664	*
4665	* So, on SMP systems we'll have to select the ASYNC update method
4666	* if there are symphoms of these problems.
4667	*/
4668	if (RTMpGetCount() > 1)
4669	{
4670	uint32_t uEAX, uEBX, uECX, uEDX;
4671	uint64_t u64DiffCoresIgnored;
4672
4673	/* Permit the user and/or the OS specfic bits to force async mode. */
4674	if (supdrvOSGetForcedAsyncTscMode(pDevExt))
4675	return SUPGIPMODE_ASYNC_TSC;
4676
4677	/* Try check for current differences between the cpus. */
4678	if (supdrvDetermineAsyncTsc(&u64DiffCoresIgnored))
4679	return SUPGIPMODE_ASYNC_TSC;
4680
4681	/*
4682	* If the CPU supports power management and is an AMD one we
4683	* won't trust it unless it has the TscInvariant bit is set.
4684	*/
4685	/* Check for "AuthenticAMD" */
4686	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
4687	if ( uEAX >= 1
4688	&& uEBX == X86_CPUID_VENDOR_AMD_EBX
4689	&& uECX == X86_CPUID_VENDOR_AMD_ECX
4690	&& uEDX == X86_CPUID_VENDOR_AMD_EDX)
4691	{
4692	/* Check for APM support and that TscInvariant is cleared. */
4693	ASMCpuId(0x80000000, &uEAX, &uEBX, &uECX, &uEDX);
4694	if (uEAX >= 0x80000007)
4695	{
4696	ASMCpuId(0x80000007, &uEAX, &uEBX, &uECX, &uEDX);
4697	if ( !(uEDX & RT_BIT(8))/* TscInvariant */
4698	&& (uEDX & 0x3e)) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
4699	return SUPGIPMODE_ASYNC_TSC;
4700	}
4701	}
4702	}
4703	return SUPGIPMODE_SYNC_TSC;
4704	}
4705
4706
4707	/**
4708	* Invalidates the GIP data upon termination.
4709	*
4710	* @param pGip Pointer to the read-write kernel mapping of the GIP.
4711	*/
4712	void VBOXCALL supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
4713	{
4714	unsigned i;
4715	pGip->u32Magic = 0;
4716	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
4717	{
4718	pGip->aCPUs[i].u64NanoTS = 0;
4719	pGip->aCPUs[i].u64TSC = 0;
4720	pGip->aCPUs[i].iTSCHistoryHead = 0;
4721	}
4722	}
4723
4724
4725	/**
4726	* Worker routine for supdrvGipUpdate and supdrvGipUpdatePerCpu that
4727	* updates all the per cpu data except the transaction id.
4728	*
4729	* @param pGip The GIP.
4730	* @param pGipCpu Pointer to the per cpu data.
4731	* @param u64NanoTS The current time stamp.
4732	*/
4733	static void supdrvGipDoUpdateCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
4734	{
4735	uint64_t u64TSC;
4736	uint64_t u64TSCDelta;
4737	uint32_t u32UpdateIntervalTSC;
4738	uint32_t u32UpdateIntervalTSCSlack;
4739	unsigned iTSCHistoryHead;
4740	uint64_t u64CpuHz;
4741
4742	/*
4743	* Update the NanoTS.
4744	*/
4745	ASMAtomicXchgU64(&pGipCpu->u64NanoTS, u64NanoTS);
4746
4747	/*
4748	* Calc TSC delta.
4749	*/
4750	/** @todo validate the NanoTS delta, don't trust the OS to call us when it should... */
4751	u64TSC = ASMReadTSC();
4752	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
4753	ASMAtomicXchgU64(&pGipCpu->u64TSC, u64TSC);
4754
4755	if (u64TSCDelta >> 32)
4756	{
4757	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
4758	pGipCpu->cErrors++;
4759	}
4760
4761	/*
4762	* TSC History.
4763	*/
4764	Assert(ELEMENTS(pGipCpu->au32TSCHistory) == 8);
4765
4766	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
4767	ASMAtomicXchgU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
4768	ASMAtomicXchgU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
4769
4770	/*
4771	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
4772	*/
4773	if (pGip->u32UpdateHz >= 1000)
4774	{
4775	uint32_t u32;
4776	u32 = pGipCpu->au32TSCHistory[0];
4777	u32 += pGipCpu->au32TSCHistory[1];
4778	u32 += pGipCpu->au32TSCHistory[2];
4779	u32 += pGipCpu->au32TSCHistory[3];
4780	u32 >>= 2;
4781	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
4782	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
4783	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
4784	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
4785	u32UpdateIntervalTSC >>= 2;
4786	u32UpdateIntervalTSC += u32;
4787	u32UpdateIntervalTSC >>= 1;
4788
4789	/* Value choosen for a 2GHz Athlon64 running linux 2.6.10/11, . */
4790	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
4791	}
4792	else if (pGip->u32UpdateHz >= 90)
4793	{
4794	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
4795	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
4796	u32UpdateIntervalTSC >>= 1;
4797
4798	/* value choosen on a 2GHz thinkpad running windows */
4799	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
4800	}
4801	else
4802	{
4803	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
4804
4805	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
4806	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
4807	}
4808	ASMAtomicXchgU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
4809
4810	/*
4811	* CpuHz.
4812	*/
4813	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, pGip->u32UpdateHz);
4814	ASMAtomicXchgU64(&pGipCpu->u64CpuHz, u64CpuHz);
4815	}
4816
4817
4818	/**
4819	* Updates the GIP.
4820	*
4821	* @param pGip Pointer to the GIP.
4822	* @param u64NanoTS The current nanosecond timesamp.
4823	*/
4824	void VBOXCALL supdrvGipUpdate(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS)
4825	{
4826	/*
4827	* Determin the relevant CPU data.
4828	*/
4829	PSUPGIPCPU pGipCpu;
4830	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
4831	pGipCpu = &pGip->aCPUs[0];
4832	else
4833	{
4834	unsigned iCpu = ASMGetApicId();
4835	if (RT_LIKELY(iCpu >= RT_ELEMENTS(pGip->aCPUs)))
4836	return;
4837	pGipCpu = &pGip->aCPUs[iCpu];
4838	}
4839
4840	/*
4841	* Start update transaction.
4842	*/
4843	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
4844	{
4845	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
4846	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
4847	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
4848	pGipCpu->cErrors++;
4849	return;
4850	}
4851
4852	/*
4853	* Recalc the update frequency every 0x800th time.
4854	*/
4855	if (!(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
4856	{
4857	if (pGip->u64NanoTSLastUpdateHz)
4858	{
4859	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
4860	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
4861	uint32_t u32UpdateHz = (uint32_t)((UINT64_C(1000000000) * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
4862	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
4863	{
4864	ASMAtomicXchgU32(&pGip->u32UpdateHz, u32UpdateHz);
4865	ASMAtomicXchgU32(&pGip->u32UpdateIntervalNS, 1000000000 / u32UpdateHz);
4866	}
4867	#endif
4868	}
4869	ASMAtomicXchgU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS);
4870	}
4871
4872	/*
4873	* Update the data.
4874	*/
4875	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS);
4876
4877	/*
4878	* Complete transaction.
4879	*/
4880	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
4881	}
4882
4883
4884	/**
4885	* Updates the per cpu GIP data for the calling cpu.
4886	*
4887	* @param pGip Pointer to the GIP.
4888	* @param u64NanoTS The current nanosecond timesamp.
4889	* @param iCpu The CPU index.
4890	*/
4891	void VBOXCALL supdrvGipUpdatePerCpu(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, unsigned iCpu)
4892	{
4893	PSUPGIPCPU pGipCpu;
4894
4895	if (RT_LIKELY(iCpu < RT_ELEMENTS(pGip->aCPUs)))
4896	{
4897	pGipCpu = &pGip->aCPUs[iCpu];
4898
4899	/*
4900	* Start update transaction.
4901	*/
4902	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
4903	{
4904	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
4905	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
4906	pGipCpu->cErrors++;
4907	return;
4908	}
4909
4910	/*
4911	* Update the data.
4912	*/
4913	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS);
4914
4915	/*
4916	* Complete transaction.
4917	*/
4918	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
4919	}
4920	}
4921
4922
4923	#ifndef SUPDRV_WITH_RELEASE_LOGGER
4924	# ifndef DEBUG /** @todo change #ifndef DEBUG -> #ifdef LOG_ENABLED */
4925	/**
4926	* Stub function for non-debug builds.
4927	*/
4928	RTDECL(PRTLOGGER) RTLogDefaultInstance(void)
4929	{
4930	return NULL;
4931	}
4932
4933	RTDECL(PRTLOGGER) RTLogRelDefaultInstance(void)
4934	{
4935	return NULL;
4936	}
4937
4938	/**
4939	* Stub function for non-debug builds.
4940	*/
4941	RTDECL(int) RTLogSetDefaultInstanceThread(PRTLOGGER pLogger, uintptr_t uKey)
4942	{
4943	return 0;
4944	}
4945
4946	/**
4947	* Stub function for non-debug builds.
4948	*/
4949	RTDECL(void) RTLogLogger(PRTLOGGER pLogger, void pvCallerRet, const char pszFormat, ...)
4950	{
4951	}
4952
4953	/**
4954	* Stub function for non-debug builds.
4955	*/
4956	RTDECL(void) RTLogLoggerEx(PRTLOGGER pLogger, unsigned fFlags, unsigned iGroup, const char *pszFormat, ...)
4957	{
4958	}
4959
4960	/**
4961	* Stub function for non-debug builds.
4962	*/
4963	RTDECL(void) RTLogLoggerExV(PRTLOGGER pLogger, unsigned fFlags, unsigned iGroup, const char *pszFormat, va_list args)
4964	{
4965	}
4966
4967	/**
4968	* Stub function for non-debug builds.
4969	*/
4970	RTDECL(void) RTLogPrintf(const char *pszFormat, ...)
4971	{
4972	}
4973
4974	/**
4975	* Stub function for non-debug builds.
4976	*/
4977	RTDECL(void) RTLogPrintfV(const char *pszFormat, va_list args)
4978	{
4979	}
4980	# endif /* !DEBUG */
4981	#endif /* !SUPDRV_WITH_RELEASE_LOGGER */

注意: 瀏覽 TracBrowser 來幫助您使用儲存庫瀏覽器

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDrv.c@ 12000

以其他格式下載: