SUPDrv.c@ 11836

Last change on this file since 11836 was 10805, checked in by vboxsync, 16 years ago
VMM+SUPDrv: Changed the VMMR0EntryEx interface to also take the session as an argument, this means breaking the current IOC interface and thus a major driver version change.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 171.2 KB

Line
1	/* $Revision: 10805 $ */
2	/** @file
3	* VBoxDrv - The VirtualBox Support Driver - Common code.
4	*/
5
6	/*
7	* Copyright (C) 2006-2007 Sun Microsystems, Inc.
8	*
9	* This file is part of VirtualBox Open Source Edition (OSE), as
10	* available from http://www.virtualbox.org. This file is free software;
11	* you can redistribute it and/or modify it under the terms of the GNU
12	* General Public License (GPL) as published by the Free Software
13	* Foundation, in version 2 as it comes in the "COPYING" file of the
14	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16	*
17	* The contents of this file may alternatively be used under the terms
18	* of the Common Development and Distribution License Version 1.0
19	* (CDDL) only, as it comes in the "COPYING.CDDL" file of the
20	* VirtualBox OSE distribution, in which case the provisions of the
21	* CDDL are applicable instead of those of the GPL.
22	*
23	* You may elect to license modified versions of this file under the
24	* terms and conditions of either the GPL or the CDDL or both.
25	*
26	* Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa
27	* Clara, CA 95054 USA or visit http://www.sun.com if you need
28	* additional information or have any questions.
29	*/
30
31	/*******************************************************************************
32	* Header Files *
33	*******************************************************************************/
34	#define LOG_GROUP LOG_GROUP_SUP_DRV
35	#include "SUPDrvInternal.h"
36	#ifndef PAGE_SHIFT
37	# include <iprt/param.h>
38	#endif
39	#include <iprt/alloc.h>
40	#include <iprt/semaphore.h>
41	#include <iprt/spinlock.h>
42	#include <iprt/thread.h>
43	#include <iprt/process.h>
44	#include <iprt/mp.h>
45	#include <iprt/cpuset.h>
46	#include <iprt/uuid.h>
47	#include <VBox/log.h>
48	#include <VBox/err.h>
49	/* VBox/x86.h not compatible with the Linux kernel sources */
50	#ifdef RT_OS_LINUX
51	# define X86_CPUID_VENDOR_AMD_EBX 0x68747541
52	# define X86_CPUID_VENDOR_AMD_ECX 0x444d4163
53	# define X86_CPUID_VENDOR_AMD_EDX 0x69746e65
54	#else
55	# include <VBox/x86.h>
56	#endif
57
58	/*
59	* Logging assignments:
60	* Log - useful stuff, like failures.
61	* LogFlow - program flow, except the really noisy bits.
62	* Log2 - Cleanup and IDTE
63	* Log3 - Loader flow noise.
64	* Log4 - Call VMMR0 flow noise.
65	* Log5 - Native yet-to-be-defined noise.
66	* Log6 - Native ioctl flow noise.
67	*
68	* Logging requires BUILD_TYPE=debug and possibly changes to the logger
69	* instanciation in log-vbox.c(pp).
70	*/
71
72
73	/*******************************************************************************
74	* Defined Constants And Macros *
75	*******************************************************************************/
76	/* from x86.h - clashes with linux thus this duplication */
77	#undef X86_CR0_PG
78	#define X86_CR0_PG RT_BIT(31)
79	#undef X86_CR0_PE
80	#define X86_CR0_PE RT_BIT(0)
81	#undef X86_CPUID_AMD_FEATURE_EDX_NX
82	#define X86_CPUID_AMD_FEATURE_EDX_NX RT_BIT(20)
83	#undef MSR_K6_EFER
84	#define MSR_K6_EFER 0xc0000080
85	#undef MSR_K6_EFER_NXE
86	#define MSR_K6_EFER_NXE RT_BIT(11)
87	#undef MSR_K6_EFER_LMA
88	#define MSR_K6_EFER_LMA RT_BIT(10)
89	#undef X86_CR4_PGE
90	#define X86_CR4_PGE RT_BIT(7)
91	#undef X86_CR4_PAE
92	#define X86_CR4_PAE RT_BIT(5)
93	#undef X86_CPUID_AMD_FEATURE_EDX_LONG_MODE
94	#define X86_CPUID_AMD_FEATURE_EDX_LONG_MODE RT_BIT(29)
95
96
97	/** The frequency by which we recalculate the u32UpdateHz and
98	* u32UpdateIntervalNS GIP members. The value must be a power of 2. */
99	#define GIP_UPDATEHZ_RECALC_FREQ 0x800
100
101	/**
102	* Validates a session pointer.
103	*
104	* @returns true/false accordingly.
105	* @param pSession The session.
106	*/
107	#define SUP_IS_SESSION_VALID(pSession) \
108	( VALID_PTR(pSession) \
109	&& pSession->u32Cookie == BIRD_INV)
110
111	/** @def VBOX_SVN_REV
112	* The makefile should define this if it can. */
113	#ifndef VBOX_SVN_REV
114	# define VBOX_SVN_REV 0
115	#endif
116
117
118	/*******************************************************************************
119	* Global Variables *
120	*******************************************************************************/
121	/**
122	* Array of the R0 SUP API.
123	*/
124	static SUPFUNC g_aFunctions[] =
125	{
126	/* name function */
127	{ "SUPR0ComponentRegisterFactory", (void *)SUPR0ComponentRegisterFactory },
128	{ "SUPR0ComponentDeregisterFactory", (void *)SUPR0ComponentDeregisterFactory },
129	{ "SUPR0ComponentQueryFactory", (void *)SUPR0ComponentQueryFactory },
130	{ "SUPR0ObjRegister", (void *)SUPR0ObjRegister },
131	{ "SUPR0ObjAddRef", (void *)SUPR0ObjAddRef },
132	{ "SUPR0ObjRelease", (void *)SUPR0ObjRelease },
133	{ "SUPR0ObjVerifyAccess", (void *)SUPR0ObjVerifyAccess },
134	{ "SUPR0LockMem", (void *)SUPR0LockMem },
135	{ "SUPR0UnlockMem", (void *)SUPR0UnlockMem },
136	{ "SUPR0ContAlloc", (void *)SUPR0ContAlloc },
137	{ "SUPR0ContFree", (void *)SUPR0ContFree },
138	{ "SUPR0LowAlloc", (void *)SUPR0LowAlloc },
139	{ "SUPR0LowFree", (void *)SUPR0LowFree },
140	{ "SUPR0MemAlloc", (void *)SUPR0MemAlloc },
141	{ "SUPR0MemGetPhys", (void *)SUPR0MemGetPhys },
142	{ "SUPR0MemFree", (void *)SUPR0MemFree },
143	{ "SUPR0PageAlloc", (void *)SUPR0PageAlloc },
144	{ "SUPR0PageFree", (void *)SUPR0PageFree },
145	{ "SUPR0Printf", (void *)SUPR0Printf },
146	{ "RTMemAlloc", (void *)RTMemAlloc },
147	{ "RTMemAllocZ", (void *)RTMemAllocZ },
148	{ "RTMemFree", (void *)RTMemFree },
149	/{ "RTMemDup", (void )RTMemDup },*/
150	{ "RTMemRealloc", (void *)RTMemRealloc },
151	{ "RTR0MemObjAllocLow", (void *)RTR0MemObjAllocLow },
152	{ "RTR0MemObjAllocPage", (void *)RTR0MemObjAllocPage },
153	{ "RTR0MemObjAllocPhys", (void *)RTR0MemObjAllocPhys },
154	{ "RTR0MemObjAllocPhysNC", (void *)RTR0MemObjAllocPhysNC },
155	{ "RTR0MemObjAllocCont", (void *)RTR0MemObjAllocCont },
156	{ "RTR0MemObjLockUser", (void *)RTR0MemObjLockUser },
157	{ "RTR0MemObjMapKernel", (void *)RTR0MemObjMapKernel },
158	{ "RTR0MemObjMapUser", (void *)RTR0MemObjMapUser },
159	{ "RTR0MemObjAddress", (void *)RTR0MemObjAddress },
160	{ "RTR0MemObjAddressR3", (void *)RTR0MemObjAddressR3 },
161	{ "RTR0MemObjSize", (void *)RTR0MemObjSize },
162	{ "RTR0MemObjIsMapping", (void *)RTR0MemObjIsMapping },
163	{ "RTR0MemObjGetPagePhysAddr", (void *)RTR0MemObjGetPagePhysAddr },
164	{ "RTR0MemObjFree", (void *)RTR0MemObjFree },
165	/* These don't work yet on linux - use fast mutexes!
166	{ "RTSemMutexCreate", (void *)RTSemMutexCreate },
167	{ "RTSemMutexRequest", (void *)RTSemMutexRequest },
168	{ "RTSemMutexRelease", (void *)RTSemMutexRelease },
169	{ "RTSemMutexDestroy", (void *)RTSemMutexDestroy },
170	*/
171	{ "RTProcSelf", (void *)RTProcSelf },
172	{ "RTR0ProcHandleSelf", (void *)RTR0ProcHandleSelf },
173	{ "RTSemFastMutexCreate", (void *)RTSemFastMutexCreate },
174	{ "RTSemFastMutexDestroy", (void *)RTSemFastMutexDestroy },
175	{ "RTSemFastMutexRequest", (void *)RTSemFastMutexRequest },
176	{ "RTSemFastMutexRelease", (void *)RTSemFastMutexRelease },
177	{ "RTSemEventCreate", (void *)RTSemEventCreate },
178	{ "RTSemEventSignal", (void *)RTSemEventSignal },
179	{ "RTSemEventWait", (void *)RTSemEventWait },
180	{ "RTSemEventWaitNoResume", (void *)RTSemEventWaitNoResume },
181	{ "RTSemEventDestroy", (void *)RTSemEventDestroy },
182	{ "RTSemEventMultiCreate", (void *)RTSemEventMultiCreate },
183	{ "RTSemEventMultiSignal", (void *)RTSemEventMultiSignal },
184	{ "RTSemEventMultiReset", (void *)RTSemEventMultiReset },
185	{ "RTSemEventMultiWait", (void *)RTSemEventMultiWait },
186	{ "RTSemEventMultiWaitNoResume", (void *)RTSemEventMultiWaitNoResume },
187	{ "RTSemEventMultiDestroy", (void *)RTSemEventMultiDestroy },
188	{ "RTSpinlockCreate", (void *)RTSpinlockCreate },
189	{ "RTSpinlockDestroy", (void *)RTSpinlockDestroy },
190	{ "RTSpinlockAcquire", (void *)RTSpinlockAcquire },
191	{ "RTSpinlockRelease", (void *)RTSpinlockRelease },
192	{ "RTSpinlockAcquireNoInts", (void *)RTSpinlockAcquireNoInts },
193	{ "RTSpinlockReleaseNoInts", (void *)RTSpinlockReleaseNoInts },
194	{ "RTTimeNanoTS", (void *)RTTimeNanoTS },
195	{ "RTTimeMillieTS", (void *)RTTimeMilliTS },
196	{ "RTTimeSystemNanoTS", (void *)RTTimeSystemNanoTS },
197	{ "RTTimeSystemMillieTS", (void *)RTTimeSystemMilliTS },
198	{ "RTThreadNativeSelf", (void *)RTThreadNativeSelf },
199	{ "RTThreadSleep", (void *)RTThreadSleep },
200	{ "RTThreadYield", (void *)RTThreadYield },
201	#if 0 /* Thread APIs, Part 2. */
202	{ "RTThreadSelf", (void *)RTThreadSelf },
203	{ "RTThreadCreate", (void *)RTThreadCreate },
204	{ "RTThreadGetNative", (void *)RTThreadGetNative },
205	{ "RTThreadWait", (void *)RTThreadWait },
206	{ "RTThreadWaitNoResume", (void *)RTThreadWaitNoResume },
207	{ "RTThreadGetName", (void *)RTThreadGetName },
208	{ "RTThreadSelfName", (void *)RTThreadSelfName },
209	{ "RTThreadGetType", (void *)RTThreadGetType },
210	{ "RTThreadUserSignal", (void *)RTThreadUserSignal },
211	{ "RTThreadUserReset", (void *)RTThreadUserReset },
212	{ "RTThreadUserWait", (void *)RTThreadUserWait },
213	{ "RTThreadUserWaitNoResume", (void *)RTThreadUserWaitNoResume },
214	#endif
215	{ "RTLogDefaultInstance", (void *)RTLogDefaultInstance },
216	{ "RTMpCpuId", (void *)RTMpCpuId },
217	{ "RTMpCpuIdFromSetIndex", (void *)RTMpCpuIdFromSetIndex },
218	{ "RTMpCpuIdToSetIndex", (void *)RTMpCpuIdToSetIndex },
219	{ "RTMpIsCpuPossible", (void *)RTMpIsCpuPossible },
220	{ "RTMpGetCount", (void *)RTMpGetCount },
221	{ "RTMpGetMaxCpuId", (void *)RTMpGetMaxCpuId },
222	{ "RTMpGetOnlineCount", (void *)RTMpGetOnlineCount },
223	{ "RTMpGetOnlineSet", (void *)RTMpGetOnlineSet },
224	{ "RTMpGetSet", (void *)RTMpGetSet },
225	{ "RTMpIsCpuOnline", (void *)RTMpIsCpuOnline },
226	{ "RTMpOnAll", (void *)RTMpOnAll },
227	{ "RTMpOnOthers", (void *)RTMpOnOthers },
228	{ "RTMpOnSpecific", (void *)RTMpOnSpecific },
229	{ "RTLogRelDefaultInstance", (void *)RTLogRelDefaultInstance },
230	{ "RTLogSetDefaultInstanceThread", (void *)RTLogSetDefaultInstanceThread },
231	{ "RTLogLogger", (void *)RTLogLogger },
232	{ "RTLogLoggerEx", (void *)RTLogLoggerEx },
233	{ "RTLogLoggerExV", (void *)RTLogLoggerExV },
234	{ "RTLogPrintf", (void *)RTLogPrintf },
235	{ "RTLogPrintfV", (void *)RTLogPrintfV },
236	{ "AssertMsg1", (void *)AssertMsg1 },
237	{ "AssertMsg2", (void *)AssertMsg2 },
238	};
239
240
241	/*******************************************************************************
242	* Internal Functions *
243	*******************************************************************************/
244	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession);
245	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType);
246	#ifdef VBOX_WITH_IDT_PATCHING
247	static int supdrvIOCtl_IdtInstall(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPIDTINSTALL pReq);
248	static PSUPDRVPATCH supdrvIdtPatchOne(PSUPDRVDEVEXT pDevExt, PSUPDRVPATCH pPatch);
249	static int supdrvIOCtl_IdtRemoveAll(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession);
250	static void supdrvIdtRemoveOne(PSUPDRVDEVEXT pDevExt, PSUPDRVPATCH pPatch);
251	static void supdrvIdtWrite(volatile void pvIdtEntry, const SUPDRVIDTE pNewIDTEntry);
252	#endif /* VBOX_WITH_IDT_PATCHING */
253	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq);
254	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq);
255	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq);
256	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq);
257	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq);
258	static int supdrvLdrSetR0EP(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx);
259	static void supdrvLdrUnsetR0EP(PSUPDRVDEVEXT pDevExt);
260	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage);
261	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage);
262	static SUPPAGINGMODE supdrvIOCtl_GetPagingMode(void);
263	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt);
264	#ifdef RT_OS_WINDOWS
265	static int supdrvPageGetPhys(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages);
266	static bool supdrvPageWasLockedByPageAlloc(PSUPDRVSESSION pSession, RTR3PTR pvR3);
267	#endif /* RT_OS_WINDOWS */
268	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt);
269	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt);
270	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
271	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
272	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser);
273
274
275	/**
276	* Initializes the device extentsion structure.
277	*
278	* @returns IPRT status code.
279	* @param pDevExt The device extension to initialize.
280	*/
281	int VBOXCALL supdrvInitDevExt(PSUPDRVDEVEXT pDevExt)
282	{
283	int rc;
284
285	#ifdef SUPDRV_WITH_RELEASE_LOGGER
286	/*
287	* Create the release log.
288	*/
289	static const char * const s_apszGroups[] = VBOX_LOGGROUP_NAMES;
290	PRTLOGGER pRelLogger;
291	rc = RTLogCreate(&pRelLogger, 0 /* fFlags */, "all",
292	"VBOX_RELEASE_LOG", RT_ELEMENTS(s_apszGroups), s_apszGroups,
293	RTLOGDEST_STDOUT \| RTLOGDEST_DEBUGGER, NULL);
294	if (RT_SUCCESS(rc))
295	RTLogRelSetDefaultInstance(pRelLogger);
296	#endif
297
298	/*
299	* Initialize it.
300	*/
301	memset(pDevExt, 0, sizeof(*pDevExt));
302	rc = RTSpinlockCreate(&pDevExt->Spinlock);
303	if (!rc)
304	{
305	rc = RTSemFastMutexCreate(&pDevExt->mtxLdr);
306	if (!rc)
307	{
308	rc = RTSemFastMutexCreate(&pDevExt->mtxComponentFactory);
309	if (!rc)
310	{
311	rc = RTSemFastMutexCreate(&pDevExt->mtxGip);
312	if (!rc)
313	{
314	rc = supdrvGipCreate(pDevExt);
315	if (RT_SUCCESS(rc))
316	{
317	pDevExt->u32Cookie = BIRD; /** @todo make this random? */
318	return VINF_SUCCESS;
319	}
320
321	RTSemFastMutexDestroy(pDevExt->mtxGip);
322	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
323	}
324	RTSemFastMutexDestroy(pDevExt->mtxComponentFactory);
325	pDevExt->mtxComponentFactory = NIL_RTSEMFASTMUTEX;
326	}
327	RTSemFastMutexDestroy(pDevExt->mtxLdr);
328	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
329	}
330	RTSpinlockDestroy(pDevExt->Spinlock);
331	pDevExt->Spinlock = NIL_RTSPINLOCK;
332	}
333	#ifdef SUPDRV_WITH_RELEASE_LOGGER
334	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
335	RTLogDestroy(RTLogSetDefaultInstance(NULL));
336	#endif
337
338	return rc;
339	}
340
341
342	/**
343	* Delete the device extension (e.g. cleanup members).
344	*
345	* @param pDevExt The device extension to delete.
346	*/
347	void VBOXCALL supdrvDeleteDevExt(PSUPDRVDEVEXT pDevExt)
348	{
349	#ifdef VBOX_WITH_IDT_PATCHING
350	PSUPDRVPATCH pPatch;
351	#endif
352	PSUPDRVOBJ pObj;
353	PSUPDRVUSAGE pUsage;
354
355	/*
356	* Kill mutexes and spinlocks.
357	*/
358	RTSemFastMutexDestroy(pDevExt->mtxGip);
359	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
360	RTSemFastMutexDestroy(pDevExt->mtxLdr);
361	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
362	RTSpinlockDestroy(pDevExt->Spinlock);
363	pDevExt->Spinlock = NIL_RTSPINLOCK;
364
365	/*
366	* Free lists.
367	*/
368	#ifdef VBOX_WITH_IDT_PATCHING
369	/* patches */
370	/** @todo make sure we don't uninstall patches which has been patched by someone else. */
371	pPatch = pDevExt->pIdtPatchesFree;
372	pDevExt->pIdtPatchesFree = NULL;
373	while (pPatch)
374	{
375	void *pvFree = pPatch;
376	pPatch = pPatch->pNext;
377	RTMemExecFree(pvFree);
378	}
379	#endif /* VBOX_WITH_IDT_PATCHING */
380
381	/* objects. */
382	pObj = pDevExt->pObjs;
383	#if !defined(DEBUG_bird) \|\| !defined(RT_OS_LINUX) /* breaks unloading, temporary, remove me! */
384	Assert(!pObj); /* (can trigger on forced unloads) */
385	#endif
386	pDevExt->pObjs = NULL;
387	while (pObj)
388	{
389	void *pvFree = pObj;
390	pObj = pObj->pNext;
391	RTMemFree(pvFree);
392	}
393
394	/* usage records. */
395	pUsage = pDevExt->pUsageFree;
396	pDevExt->pUsageFree = NULL;
397	while (pUsage)
398	{
399	void *pvFree = pUsage;
400	pUsage = pUsage->pNext;
401	RTMemFree(pvFree);
402	}
403
404	/* kill the GIP. */
405	supdrvGipDestroy(pDevExt);
406
407	#ifdef SUPDRV_WITH_RELEASE_LOGGER
408	/* destroy the loggers. */
409	RTLogDestroy(RTLogRelSetDefaultInstance(NULL));
410	RTLogDestroy(RTLogSetDefaultInstance(NULL));
411	#endif
412	}
413
414
415	/**
416	* Create session.
417	*
418	* @returns IPRT status code.
419	* @param pDevExt Device extension.
420	* @param fUser Flag indicating whether this is a user or kernel session.
421	* @param ppSession Where to store the pointer to the session data.
422	*/
423	int VBOXCALL supdrvCreateSession(PSUPDRVDEVEXT pDevExt, bool fUser, PSUPDRVSESSION *ppSession)
424	{
425	/*
426	* Allocate memory for the session data.
427	*/
428	int rc = VERR_NO_MEMORY;
429	PSUPDRVSESSION pSession = ppSession = (PSUPDRVSESSION)RTMemAllocZ(sizeof(pSession));
430	if (pSession)
431	{
432	/* Initialize session data. */
433	rc = RTSpinlockCreate(&pSession->Spinlock);
434	if (!rc)
435	{
436	Assert(pSession->Spinlock != NIL_RTSPINLOCK);
437	pSession->pDevExt = pDevExt;
438	pSession->u32Cookie = BIRD_INV;
439	/*pSession->pLdrUsage = NULL;
440	pSession->pPatchUsage = NULL;
441	pSession->pVM = NULL;
442	pSession->pUsage = NULL;
443	pSession->pGip = NULL;
444	pSession->fGipReferenced = false;
445	pSession->Bundle.cUsed = 0; */
446	pSession->Uid = NIL_RTUID;
447	pSession->Gid = NIL_RTGID;
448	if (fUser)
449	{
450	pSession->Process = RTProcSelf();
451	pSession->R0Process = RTR0ProcHandleSelf();
452	}
453	else
454	{
455	pSession->Process = NIL_RTPROCESS;
456	pSession->R0Process = NIL_RTR0PROCESS;
457	}
458
459	LogFlow(("Created session %p initial cookie=%#x\n", pSession, pSession->u32Cookie));
460	return VINF_SUCCESS;
461	}
462
463	RTMemFree(pSession);
464	*ppSession = NULL;
465	Log(("Failed to create spinlock, rc=%d!\n", rc));
466	}
467
468	return rc;
469	}
470
471
472	/**
473	* Shared code for cleaning up a session.
474	*
475	* @param pDevExt Device extension.
476	* @param pSession Session data.
477	* This data will be freed by this routine.
478	*/
479	void VBOXCALL supdrvCloseSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
480	{
481	/*
482	* Cleanup the session first.
483	*/
484	supdrvCleanupSession(pDevExt, pSession);
485
486	/*
487	* Free the rest of the session stuff.
488	*/
489	RTSpinlockDestroy(pSession->Spinlock);
490	pSession->Spinlock = NIL_RTSPINLOCK;
491	pSession->pDevExt = NULL;
492	RTMemFree(pSession);
493	LogFlow(("supdrvCloseSession: returns\n"));
494	}
495
496
497	/**
498	* Shared code for cleaning up a session (but not quite freeing it).
499	*
500	* This is primarily intended for MAC OS X where we have to clean up the memory
501	* stuff before the file handle is closed.
502	*
503	* @param pDevExt Device extension.
504	* @param pSession Session data.
505	* This data will be freed by this routine.
506	*/
507	void VBOXCALL supdrvCleanupSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
508	{
509	PSUPDRVBUNDLE pBundle;
510	LogFlow(("supdrvCleanupSession: pSession=%p\n", pSession));
511
512	/*
513	* Remove logger instances related to this session.
514	*/
515	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pSession);
516
517	#ifdef VBOX_WITH_IDT_PATCHING
518	/*
519	* Uninstall any IDT patches installed for this session.
520	*/
521	supdrvIOCtl_IdtRemoveAll(pDevExt, pSession);
522	#endif
523
524	/*
525	* Release object references made in this session.
526	* In theory there should be noone racing us in this session.
527	*/
528	Log2(("release objects - start\n"));
529	if (pSession->pUsage)
530	{
531	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
532	PSUPDRVUSAGE pUsage;
533	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
534
535	while ((pUsage = pSession->pUsage) != NULL)
536	{
537	PSUPDRVOBJ pObj = pUsage->pObj;
538	pSession->pUsage = pUsage->pNext;
539
540	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
541	if (pUsage->cUsage < pObj->cUsage)
542	{
543	pObj->cUsage -= pUsage->cUsage;
544	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
545	}
546	else
547	{
548	/* Destroy the object and free the record. */
549	if (pDevExt->pObjs == pObj)
550	pDevExt->pObjs = pObj->pNext;
551	else
552	{
553	PSUPDRVOBJ pObjPrev;
554	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
555	if (pObjPrev->pNext == pObj)
556	{
557	pObjPrev->pNext = pObj->pNext;
558	break;
559	}
560	Assert(pObjPrev);
561	}
562	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
563
564	Log(("supdrvCleanupSession: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
565	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
566	if (pObj->pfnDestructor)
567	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
568	RTMemFree(pObj);
569	}
570
571	/* free it and continue. */
572	RTMemFree(pUsage);
573
574	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
575	}
576
577	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
578	AssertMsg(!pSession->pUsage, ("Some buster reregistered an object during desturction!\n"));
579	}
580	Log2(("release objects - done\n"));
581
582	/*
583	* Release memory allocated in the session.
584	*
585	* We do not serialize this as we assume that the application will
586	* not allocated memory while closing the file handle object.
587	*/
588	Log2(("freeing memory:\n"));
589	pBundle = &pSession->Bundle;
590	while (pBundle)
591	{
592	PSUPDRVBUNDLE pToFree;
593	unsigned i;
594
595	/*
596	* Check and unlock all entries in the bundle.
597	*/
598	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
599	{
600	if (pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ)
601	{
602	int rc;
603	Log2(("eType=%d pvR0=%p pvR3=%p cb=%ld\n", pBundle->aMem[i].eType, RTR0MemObjAddress(pBundle->aMem[i].MemObj),
604	(void *)RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3), (long)RTR0MemObjSize(pBundle->aMem[i].MemObj)));
605	if (pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ)
606	{
607	rc = RTR0MemObjFree(pBundle->aMem[i].MapObjR3, false);
608	AssertRC(rc); /** @todo figure out how to handle this. */
609	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
610	}
611	rc = RTR0MemObjFree(pBundle->aMem[i].MemObj, false);
612	AssertRC(rc); /** @todo figure out how to handle this. */
613	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
614	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
615	}
616	}
617
618	/*
619	* Advance and free previous bundle.
620	*/
621	pToFree = pBundle;
622	pBundle = pBundle->pNext;
623
624	pToFree->pNext = NULL;
625	pToFree->cUsed = 0;
626	if (pToFree != &pSession->Bundle)
627	RTMemFree(pToFree);
628	}
629	Log2(("freeing memory - done\n"));
630
631	/*
632	* Deregister component factories.
633	*/
634	RTSemFastMutexRequest(pDevExt->mtxComponentFactory);
635	Log2(("deregistering component factories:\n"));
636	if (pDevExt->pComponentFactoryHead)
637	{
638	PSUPDRVFACTORYREG pPrev = NULL;
639	PSUPDRVFACTORYREG pCur = pDevExt->pComponentFactoryHead;
640	while (pCur)
641	{
642	if (pCur->pSession == pSession)
643	{
644	/* unlink it */
645	PSUPDRVFACTORYREG pNext = pCur->pNext;
646	if (pPrev)
647	pPrev->pNext = pNext;
648	else
649	pDevExt->pComponentFactoryHead = pNext;
650
651	/* free it */
652	pCur->pNext = NULL;
653	pCur->pSession = NULL;
654	pCur->pFactory = NULL;
655	RTMemFree(pCur);
656
657	/* next */
658	pCur = pNext;
659	}
660	else
661	{
662	/* next */
663	pPrev = pCur;
664	pCur = pCur->pNext;
665	}
666	}
667	}
668	RTSemFastMutexRelease(pDevExt->mtxComponentFactory);
669	Log2(("deregistering component factories - done\n"));
670
671	/*
672	* Loaded images needs to be dereferenced and possibly freed up.
673	*/
674	RTSemFastMutexRequest(pDevExt->mtxLdr);
675	Log2(("freeing images:\n"));
676	if (pSession->pLdrUsage)
677	{
678	PSUPDRVLDRUSAGE pUsage = pSession->pLdrUsage;
679	pSession->pLdrUsage = NULL;
680	while (pUsage)
681	{
682	void *pvFree = pUsage;
683	PSUPDRVLDRIMAGE pImage = pUsage->pImage;
684	if (pImage->cUsage > pUsage->cUsage)
685	pImage->cUsage -= pUsage->cUsage;
686	else
687	supdrvLdrFree(pDevExt, pImage);
688	pUsage->pImage = NULL;
689	pUsage = pUsage->pNext;
690	RTMemFree(pvFree);
691	}
692	}
693	RTSemFastMutexRelease(pDevExt->mtxLdr);
694	Log2(("freeing images - done\n"));
695
696	/*
697	* Unmap the GIP.
698	*/
699	Log2(("umapping GIP:\n"));
700	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
701	{
702	SUPR0GipUnmap(pSession);
703	pSession->fGipReferenced = 0;
704	}
705	Log2(("umapping GIP - done\n"));
706	}
707
708
709	/**
710	* Fast path I/O Control worker.
711	*
712	* @returns VBox status code that should be passed down to ring-3 unchanged.
713	* @param uIOCtl Function number.
714	* @param pDevExt Device extention.
715	* @param pSession Session data.
716	*/
717	int VBOXCALL supdrvIOCtlFast(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
718	{
719	/*
720	* We check the two prereqs after doing this only to allow the compiler to optimize things better.
721	*/
722	if (RT_LIKELY(pSession->pVM && pDevExt->pfnVMMR0EntryFast))
723	{
724	switch (uIOCtl)
725	{
726	case SUP_IOCTL_FAST_DO_RAW_RUN:
727	pDevExt->pfnVMMR0EntryFast(pSession->pVM, SUP_VMMR0_DO_RAW_RUN);
728	break;
729	case SUP_IOCTL_FAST_DO_HWACC_RUN:
730	pDevExt->pfnVMMR0EntryFast(pSession->pVM, SUP_VMMR0_DO_HWACC_RUN);
731	break;
732	case SUP_IOCTL_FAST_DO_NOP:
733	pDevExt->pfnVMMR0EntryFast(pSession->pVM, SUP_VMMR0_DO_NOP);
734	break;
735	default:
736	return VERR_INTERNAL_ERROR;
737	}
738	return VINF_SUCCESS;
739	}
740	return VERR_INTERNAL_ERROR;
741	}
742
743
744	/**
745	* Helper for supdrvIOCtl. Check if pszStr contains any character of pszChars.
746	* We would use strpbrk here if this function would be contained in the RedHat kABI white
747	* list, see http://www.kerneldrivers.org/RHEL5.
748	*
749	* @return 1 if pszStr does contain any character of pszChars, 0 otherwise.
750	* @param pszStr String to check
751	* @param pszChars Character set
752	*/
753	static int supdrvCheckInvalidChar(const char pszStr, const char pszChars)
754	{
755	int chCur;
756	while ((chCur = *pszStr++) != '\0')
757	{
758	int ch;
759	const char *psz = pszChars;
760	while ((ch = *psz++) != '\0')
761	if (ch == chCur)
762	return 1;
763
764	}
765	return 0;
766	}
767
768
769	/**
770	* I/O Control worker.
771	*
772	* @returns 0 on success.
773	* @returns VERR_INVALID_PARAMETER if the request is invalid.
774	*
775	* @param uIOCtl Function number.
776	* @param pDevExt Device extention.
777	* @param pSession Session data.
778	* @param pReqHdr The request header.
779	*/
780	int VBOXCALL supdrvIOCtl(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
781	{
782	/*
783	* Validate the request.
784	*/
785	/* this first check could probably be omitted as its also done by the OS specific code... */
786	if (RT_UNLIKELY( (pReqHdr->fFlags & SUPREQHDR_FLAGS_MAGIC_MASK) != SUPREQHDR_FLAGS_MAGIC
787	\|\| pReqHdr->cbIn < sizeof(*pReqHdr)
788	\|\| pReqHdr->cbOut < sizeof(*pReqHdr)))
789	{
790	OSDBGPRINT(("vboxdrv: Bad ioctl request header; cbIn=%#lx cbOut=%#lx fFlags=%#lx\n",
791	(long)pReqHdr->cbIn, (long)pReqHdr->cbOut, (long)pReqHdr->fFlags));
792	return VERR_INVALID_PARAMETER;
793	}
794	if (RT_UNLIKELY(uIOCtl == SUP_IOCTL_COOKIE))
795	{
796	if (pReqHdr->u32Cookie != SUPCOOKIE_INITIAL_COOKIE)
797	{
798	OSDBGPRINT(("SUP_IOCTL_COOKIE: bad cookie %#lx\n", (long)pReqHdr->u32Cookie));
799	return VERR_INVALID_PARAMETER;
800	}
801	}
802	else if (RT_UNLIKELY( pReqHdr->u32Cookie != pDevExt->u32Cookie
803	\|\| pReqHdr->u32SessionCookie != pSession->u32Cookie))
804	{
805	OSDBGPRINT(("vboxdrv: bad cookie %#lx / %#lx.\n", (long)pReqHdr->u32Cookie, (long)pReqHdr->u32SessionCookie));
806	return VERR_INVALID_PARAMETER;
807	}
808
809	/*
810	* Validation macros
811	*/
812	#define REQ_CHECK_SIZES_EX(Name, cbInExpect, cbOutExpect) \
813	do { \
814	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect) \|\| pReqHdr->cbOut != (cbOutExpect))) \
815	{ \
816	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld. cbOut=%ld expected %ld.\n", \
817	(long)pReq->Hdr.cbIn, (long)(cbInExpect), (long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
818	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
819	} \
820	} while (0)
821
822	#define REQ_CHECK_SIZES(Name) REQ_CHECK_SIZES_EX(Name, Name ## _SIZE_IN, Name ## _SIZE_OUT)
823
824	#define REQ_CHECK_SIZE_IN(Name, cbInExpect) \
825	do { \
826	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect))) \
827	{ \
828	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld.\n", \
829	(long)pReq->Hdr.cbIn, (long)(cbInExpect))); \
830	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
831	} \
832	} while (0)
833
834	#define REQ_CHECK_SIZE_OUT(Name, cbOutExpect) \
835	do { \
836	if (RT_UNLIKELY(pReqHdr->cbOut != (cbOutExpect))) \
837	{ \
838	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbOut=%ld expected %ld.\n", \
839	(long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
840	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
841	} \
842	} while (0)
843
844	#define REQ_CHECK_EXPR(Name, expr) \
845	do { \
846	if (RT_UNLIKELY(!(expr))) \
847	{ \
848	OSDBGPRINT(( #Name ": %s\n", #expr)); \
849	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
850	} \
851	} while (0)
852
853	#define REQ_CHECK_EXPR_FMT(expr, fmt) \
854	do { \
855	if (RT_UNLIKELY(!(expr))) \
856	{ \
857	OSDBGPRINT( fmt ); \
858	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
859	} \
860	} while (0)
861
862
863	/*
864	* The switch.
865	*/
866	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
867	{
868	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
869	{
870	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
871	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
872	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
873	{
874	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
875	pReq->Hdr.rc = VERR_INVALID_MAGIC;
876	return 0;
877	}
878
879	#if 0
880	/*
881	* Call out to the OS specific code and let it do permission checks on the
882	* client process.
883	*/
884	if (!supdrvOSValidateClientProcess(pDevExt, pSession))
885	{
886	pReq->u.Out.u32Cookie = 0xffffffff;
887	pReq->u.Out.u32SessionCookie = 0xffffffff;
888	pReq->u.Out.u32SessionVersion = 0xffffffff;
889	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
890	pReq->u.Out.pSession = NULL;
891	pReq->u.Out.cFunctions = 0;
892	pReq->Hdr.rc = VERR_PERMISSION_DENIED;
893	return 0;
894	}
895	#endif
896
897	/*
898	* Match the version.
899	* The current logic is very simple, match the major interface version.
900	*/
901	if ( pReq->u.In.u32MinVersion > SUPDRV_IOC_VERSION
902	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRV_IOC_VERSION & 0xffff0000))
903	{
904	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
905	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRV_IOC_VERSION));
906	pReq->u.Out.u32Cookie = 0xffffffff;
907	pReq->u.Out.u32SessionCookie = 0xffffffff;
908	pReq->u.Out.u32SessionVersion = 0xffffffff;
909	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
910	pReq->u.Out.pSession = NULL;
911	pReq->u.Out.cFunctions = 0;
912	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
913	return 0;
914	}
915
916	/*
917	* Fill in return data and be gone.
918	* N.B. The first one to change SUPDRV_IOC_VERSION shall makes sure that
919	* u32SessionVersion <= u32ReqVersion!
920	*/
921	/** @todo Somehow validate the client and negotiate a secure cookie... */
922	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
923	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
924	pReq->u.Out.u32SessionVersion = SUPDRV_IOC_VERSION;
925	pReq->u.Out.u32DriverVersion = SUPDRV_IOC_VERSION;
926	pReq->u.Out.pSession = pSession;
927	pReq->u.Out.cFunctions = sizeof(g_aFunctions) / sizeof(g_aFunctions[0]);
928	pReq->Hdr.rc = VINF_SUCCESS;
929	return 0;
930	}
931
932	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_QUERY_FUNCS(0)):
933	{
934	/* validate */
935	PSUPQUERYFUNCS pReq = (PSUPQUERYFUNCS)pReqHdr;
936	REQ_CHECK_SIZES_EX(SUP_IOCTL_QUERY_FUNCS, SUP_IOCTL_QUERY_FUNCS_SIZE_IN, SUP_IOCTL_QUERY_FUNCS_SIZE_OUT(RT_ELEMENTS(g_aFunctions)));
937
938	/* execute */
939	pReq->u.Out.cFunctions = RT_ELEMENTS(g_aFunctions);
940	memcpy(&pReq->u.Out.aFunctions[0], g_aFunctions, sizeof(g_aFunctions));
941	pReq->Hdr.rc = VINF_SUCCESS;
942	return 0;
943	}
944
945	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_IDT_INSTALL):
946	{
947	/* validate */
948	PSUPIDTINSTALL pReq = (PSUPIDTINSTALL)pReqHdr;
949	REQ_CHECK_SIZES(SUP_IOCTL_IDT_INSTALL);
950
951	/* execute */
952	#ifdef VBOX_WITH_IDT_PATCHING
953	pReq->Hdr.rc = supdrvIOCtl_IdtInstall(pDevExt, pSession, pReq);
954	#else
955	pReq->u.Out.u8Idt = 3;
956	pReq->Hdr.rc = VERR_NOT_SUPPORTED;
957	#endif
958	return 0;
959	}
960
961	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_IDT_REMOVE):
962	{
963	/* validate */
964	PSUPIDTREMOVE pReq = (PSUPIDTREMOVE)pReqHdr;
965	REQ_CHECK_SIZES(SUP_IOCTL_IDT_REMOVE);
966
967	/* execute */
968	#ifdef VBOX_WITH_IDT_PATCHING
969	pReq->Hdr.rc = supdrvIOCtl_IdtRemoveAll(pDevExt, pSession);
970	#else
971	pReq->Hdr.rc = VERR_NOT_SUPPORTED;
972	#endif
973	return 0;
974	}
975
976	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_LOCK):
977	{
978	/* validate */
979	PSUPPAGELOCK pReq = (PSUPPAGELOCK)pReqHdr;
980	REQ_CHECK_SIZE_IN(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_IN);
981	REQ_CHECK_SIZE_OUT(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_OUT(pReq->u.In.cPages));
982	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.cPages > 0);
983	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.pvR3 >= PAGE_SIZE);
984
985	/* execute */
986	pReq->Hdr.rc = SUPR0LockMem(pSession, pReq->u.In.pvR3, pReq->u.In.cPages, &pReq->u.Out.aPages[0]);
987	if (RT_FAILURE(pReq->Hdr.rc))
988	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
989	return 0;
990	}
991
992	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_UNLOCK):
993	{
994	/* validate */
995	PSUPPAGEUNLOCK pReq = (PSUPPAGEUNLOCK)pReqHdr;
996	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_UNLOCK);
997
998	/* execute */
999	pReq->Hdr.rc = SUPR0UnlockMem(pSession, pReq->u.In.pvR3);
1000	return 0;
1001	}
1002
1003	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_ALLOC):
1004	{
1005	/* validate */
1006	PSUPCONTALLOC pReq = (PSUPCONTALLOC)pReqHdr;
1007	REQ_CHECK_SIZES(SUP_IOCTL_CONT_ALLOC);
1008
1009	/* execute */
1010	pReq->Hdr.rc = SUPR0ContAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.HCPhys);
1011	if (RT_FAILURE(pReq->Hdr.rc))
1012	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1013	return 0;
1014	}
1015
1016	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_FREE):
1017	{
1018	/* validate */
1019	PSUPCONTFREE pReq = (PSUPCONTFREE)pReqHdr;
1020	REQ_CHECK_SIZES(SUP_IOCTL_CONT_FREE);
1021
1022	/* execute */
1023	pReq->Hdr.rc = SUPR0ContFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1024	return 0;
1025	}
1026
1027	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_OPEN):
1028	{
1029	/* validate */
1030	PSUPLDROPEN pReq = (PSUPLDROPEN)pReqHdr;
1031	REQ_CHECK_SIZES(SUP_IOCTL_LDR_OPEN);
1032	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImage > 0);
1033	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImage < _1M*16);
1034	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.szName[0]);
1035	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, memchr(pReq->u.In.szName, '\0', sizeof(pReq->u.In.szName)));
1036	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, !supdrvCheckInvalidChar(pReq->u.In.szName, ";:()[]{}/\\\|&*%#@!~`\"'"));
1037
1038	/* execute */
1039	pReq->Hdr.rc = supdrvIOCtl_LdrOpen(pDevExt, pSession, pReq);
1040	return 0;
1041	}
1042
1043	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOAD):
1044	{
1045	/* validate */
1046	PSUPLDRLOAD pReq = (PSUPLDRLOAD)pReqHdr;
1047	REQ_CHECK_EXPR(Name, pReq->Hdr.cbIn >= sizeof(*pReq));
1048	REQ_CHECK_SIZES_EX(SUP_IOCTL_LDR_LOAD, SUP_IOCTL_LDR_LOAD_SIZE_IN(pReq->u.In.cbImage), SUP_IOCTL_LDR_LOAD_SIZE_OUT);
1049	REQ_CHECK_EXPR(SUP_IOCTL_LDR_LOAD, pReq->u.In.cSymbols <= 16384);
1050	REQ_CHECK_EXPR_FMT( !pReq->u.In.cSymbols
1051	\|\| ( pReq->u.In.offSymbols < pReq->u.In.cbImage
1052	&& pReq->u.In.offSymbols + pReq->u.In.cSymbols * sizeof(SUPLDRSYM) <= pReq->u.In.cbImage),
1053	("SUP_IOCTL_LDR_LOAD: offSymbols=%#lx cSymbols=%#lx cbImage=%#lx\n", (long)pReq->u.In.offSymbols,
1054	(long)pReq->u.In.cSymbols, (long)pReq->u.In.cbImage));
1055	REQ_CHECK_EXPR_FMT( !pReq->u.In.cbStrTab
1056	\|\| ( pReq->u.In.offStrTab < pReq->u.In.cbImage
1057	&& pReq->u.In.offStrTab + pReq->u.In.cbStrTab <= pReq->u.In.cbImage
1058	&& pReq->u.In.cbStrTab <= pReq->u.In.cbImage),
1059	("SUP_IOCTL_LDR_LOAD: offStrTab=%#lx cbStrTab=%#lx cbImage=%#lx\n", (long)pReq->u.In.offStrTab,
1060	(long)pReq->u.In.cbStrTab, (long)pReq->u.In.cbImage));
1061
1062	if (pReq->u.In.cSymbols)
1063	{
1064	uint32_t i;
1065	PSUPLDRSYM paSyms = (PSUPLDRSYM)&pReq->u.In.achImage[pReq->u.In.offSymbols];
1066	for (i = 0; i < pReq->u.In.cSymbols; i++)
1067	{
1068	REQ_CHECK_EXPR_FMT(paSyms[i].offSymbol < pReq->u.In.cbImage,
1069	("SUP_IOCTL_LDR_LOAD: sym #%ld: symb off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offSymbol, (long)pReq->u.In.cbImage));
1070	REQ_CHECK_EXPR_FMT(paSyms[i].offName < pReq->u.In.cbStrTab,
1071	("SUP_IOCTL_LDR_LOAD: sym #%ld: name off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImage));
1072	REQ_CHECK_EXPR_FMT(memchr(&pReq->u.In.achImage[pReq->u.In.offStrTab + paSyms[i].offName], '\0', pReq->u.In.cbStrTab - paSyms[i].offName),
1073	("SUP_IOCTL_LDR_LOAD: sym #%ld: unterminated name! (%#lx / %#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImage));
1074	}
1075	}
1076
1077	/* execute */
1078	pReq->Hdr.rc = supdrvIOCtl_LdrLoad(pDevExt, pSession, pReq);
1079	return 0;
1080	}
1081
1082	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_FREE):
1083	{
1084	/* validate */
1085	PSUPLDRFREE pReq = (PSUPLDRFREE)pReqHdr;
1086	REQ_CHECK_SIZES(SUP_IOCTL_LDR_FREE);
1087
1088	/* execute */
1089	pReq->Hdr.rc = supdrvIOCtl_LdrFree(pDevExt, pSession, pReq);
1090	return 0;
1091	}
1092
1093	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_GET_SYMBOL):
1094	{
1095	/* validate */
1096	PSUPLDRGETSYMBOL pReq = (PSUPLDRGETSYMBOL)pReqHdr;
1097	REQ_CHECK_SIZES(SUP_IOCTL_LDR_GET_SYMBOL);
1098	REQ_CHECK_EXPR(SUP_IOCTL_LDR_GET_SYMBOL, memchr(pReq->u.In.szSymbol, '\0', sizeof(pReq->u.In.szSymbol)));
1099
1100	/* execute */
1101	pReq->Hdr.rc = supdrvIOCtl_LdrGetSymbol(pDevExt, pSession, pReq);
1102	return 0;
1103	}
1104
1105	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0(0)):
1106	{
1107	/* validate */
1108	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1109	Log4(("SUP_IOCTL_CALL_VMMR0: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1110	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1111
1112	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_VMMR0_SIZE(0))
1113	{
1114	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(0), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(0));
1115
1116	/* execute */
1117	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1118	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg, pSession);
1119	else
1120	pReq->Hdr.rc = VERR_WRONG_ORDER;
1121	}
1122	else
1123	{
1124	PSUPVMMR0REQHDR pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1125	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR)),
1126	("SUP_IOCTL_CALL_VMMR0: cbIn=%#x < %#x\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR))));
1127	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1128	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(pVMMReq->cbReq));
1129
1130	/* execute */
1131	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1132	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg, pSession);
1133	else
1134	pReq->Hdr.rc = VERR_WRONG_ORDER;
1135	}
1136
1137	if ( RT_FAILURE(pReq->Hdr.rc)
1138	&& pReq->Hdr.rc != VERR_INTERRUPTED
1139	&& pReq->Hdr.rc != VERR_TIMEOUT)
1140	Log(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1141	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1142	else
1143	Log4(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1144	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1145	return 0;
1146	}
1147
1148	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_PAGING_MODE):
1149	{
1150	/* validate */
1151	PSUPGETPAGINGMODE pReq = (PSUPGETPAGINGMODE)pReqHdr;
1152	REQ_CHECK_SIZES(SUP_IOCTL_GET_PAGING_MODE);
1153
1154	/* execute */
1155	pReq->Hdr.rc = VINF_SUCCESS;
1156	pReq->u.Out.enmMode = supdrvIOCtl_GetPagingMode();
1157	return 0;
1158	}
1159
1160	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_ALLOC):
1161	{
1162	/* validate */
1163	PSUPLOWALLOC pReq = (PSUPLOWALLOC)pReqHdr;
1164	REQ_CHECK_EXPR(SUP_IOCTL_LOW_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_LOW_ALLOC_SIZE_IN);
1165	REQ_CHECK_SIZES_EX(SUP_IOCTL_LOW_ALLOC, SUP_IOCTL_LOW_ALLOC_SIZE_IN, SUP_IOCTL_LOW_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1166
1167	/* execute */
1168	pReq->Hdr.rc = SUPR0LowAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1169	if (RT_FAILURE(pReq->Hdr.rc))
1170	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1171	return 0;
1172	}
1173
1174	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_FREE):
1175	{
1176	/* validate */
1177	PSUPLOWFREE pReq = (PSUPLOWFREE)pReqHdr;
1178	REQ_CHECK_SIZES(SUP_IOCTL_LOW_FREE);
1179
1180	/* execute */
1181	pReq->Hdr.rc = SUPR0LowFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1182	return 0;
1183	}
1184
1185	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_MAP):
1186	{
1187	/* validate */
1188	PSUPGIPMAP pReq = (PSUPGIPMAP)pReqHdr;
1189	REQ_CHECK_SIZES(SUP_IOCTL_GIP_MAP);
1190
1191	/* execute */
1192	pReq->Hdr.rc = SUPR0GipMap(pSession, &pReq->u.Out.pGipR3, &pReq->u.Out.HCPhysGip);
1193	if (RT_SUCCESS(pReq->Hdr.rc))
1194	pReq->u.Out.pGipR0 = pDevExt->pGip;
1195	return 0;
1196	}
1197
1198	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_UNMAP):
1199	{
1200	/* validate */
1201	PSUPGIPUNMAP pReq = (PSUPGIPUNMAP)pReqHdr;
1202	REQ_CHECK_SIZES(SUP_IOCTL_GIP_UNMAP);
1203
1204	/* execute */
1205	pReq->Hdr.rc = SUPR0GipUnmap(pSession);
1206	return 0;
1207	}
1208
1209	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SET_VM_FOR_FAST):
1210	{
1211	/* validate */
1212	PSUPSETVMFORFAST pReq = (PSUPSETVMFORFAST)pReqHdr;
1213	REQ_CHECK_SIZES(SUP_IOCTL_SET_VM_FOR_FAST);
1214	REQ_CHECK_EXPR_FMT( !pReq->u.In.pVMR0
1215	\|\| ( VALID_PTR(pReq->u.In.pVMR0)
1216	&& !((uintptr_t)pReq->u.In.pVMR0 & (PAGE_SIZE - 1))),
1217	("SUP_IOCTL_SET_VM_FOR_FAST: pVMR0=%p!\n", pReq->u.In.pVMR0));
1218	/* execute */
1219	pSession->pVM = pReq->u.In.pVMR0;
1220	pReq->Hdr.rc = VINF_SUCCESS;
1221	return 0;
1222	}
1223
1224	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_ALLOC):
1225	{
1226	/* validate */
1227	PSUPPAGEALLOC pReq = (PSUPPAGEALLOC)pReqHdr;
1228	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_SIZE_IN);
1229	REQ_CHECK_SIZES_EX(SUP_IOCTL_PAGE_ALLOC, SUP_IOCTL_PAGE_ALLOC_SIZE_IN, SUP_IOCTL_PAGE_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1230
1231	/* execute */
1232	pReq->Hdr.rc = SUPR0PageAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1233	if (RT_FAILURE(pReq->Hdr.rc))
1234	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1235	return 0;
1236	}
1237
1238	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_FREE):
1239	{
1240	/* validate */
1241	PSUPPAGEFREE pReq = (PSUPPAGEFREE)pReqHdr;
1242	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_FREE);
1243
1244	/* execute */
1245	pReq->Hdr.rc = SUPR0PageFree(pSession, pReq->u.In.pvR3);
1246	return 0;
1247	}
1248
1249	default:
1250	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1251	break;
1252	}
1253	return SUPDRV_ERR_GENERAL_FAILURE;
1254	}
1255
1256
1257	/**
1258	* Inter-Driver Communcation (IDC) worker.
1259	*
1260	* @returns VBox status code.
1261	* @retval VINF_SUCCESS on success.
1262	* @retval VERR_INVALID_PARAMETER if the request is invalid.
1263	* @retval VERR_NOT_SUPPORTED if the request isn't supported.
1264	*
1265	* @param uReq The request (function) code.
1266	* @param pDevExt Device extention.
1267	* @param pSession Session data.
1268	* @param pReqHdr The request header.
1269	*/
1270	int VBOXCALL supdrvIDC(uintptr_t uReq, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQHDR pReqHdr)
1271	{
1272	/*
1273	* The OS specific code has already validated the pSession
1274	* pointer, and the request size being greater or equal to
1275	* size of the header.
1276	*
1277	* So, just check that pSession is a kernel context session.
1278	*/
1279	if (RT_UNLIKELY( pSession
1280	&& pSession->R0Process != NIL_RTR0PROCESS))
1281	return VERR_INVALID_PARAMETER;
1282
1283	/*
1284	* Validation macro.
1285	*/
1286	#define REQ_CHECK_IDC_SIZE(Name, cbExpect) \
1287	do { \
1288	if (RT_UNLIKELY(pReqHdr->cb != (cbExpect))) \
1289	{ \
1290	OSDBGPRINT(( #Name ": Invalid input/output sizes. cb=%ld expected %ld.\n", \
1291	(long)pReqHdr->cb, (long)(cbExpect))); \
1292	return pReqHdr->rc = VERR_INVALID_PARAMETER; \
1293	} \
1294	} while (0)
1295
1296	switch (uReq)
1297	{
1298	case SUPDRV_IDC_REQ_CONNECT:
1299	{
1300	PSUPDRVIDCREQCONNECT pReq = (PSUPDRVIDCREQCONNECT)pReqHdr;
1301	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_CONNECT, sizeof(*pReq));
1302
1303	/*
1304	* Validate the cookie and other input.
1305	*/
1306	if (pReq->Hdr.pSession != NULL)
1307	{
1308	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: pSession=%p expected NULL!\n", pReq->Hdr.pSession));
1309	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1310	}
1311	if (pReq->u.In.u32MagicCookie != SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE)
1312	{
1313	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: u32MagicCookie=%#x expected %#x!\n",
1314	pReq->u.In.u32MagicCookie, SUPDRVIDCREQ_CONNECT_MAGIC_COOKIE));
1315	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1316	}
1317	if ( pReq->u.In.uMinVersion > pReq->u.In.uReqVersion
1318	\|\| (pReq->u.In.uMinVersion & UINT32_C(0xffff0000)) != (pReq->u.In.uReqVersion & UINT32_C(0xffff0000)))
1319	{
1320	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: uMinVersion=%#x uMaxVersion=%#x doesn't match!\n",
1321	pReq->u.In.uMinVersion, pReq->u.In.uReqVersion));
1322	return pReqHdr->rc = VERR_INVALID_PARAMETER;
1323	}
1324
1325	/*
1326	* Match the version.
1327	* The current logic is very simple, match the major interface version.
1328	*/
1329	if ( pReq->u.In.uMinVersion > SUPDRV_IDC_VERSION
1330	\|\| (pReq->u.In.uMinVersion & 0xffff0000) != (SUPDRV_IDC_VERSION & 0xffff0000))
1331	{
1332	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
1333	pReq->u.In.uReqVersion, pReq->u.In.uMinVersion, SUPDRV_IDC_VERSION));
1334	pReq->u.Out.pSession = NULL;
1335	pReq->u.Out.uSessionVersion = 0xffffffff;
1336	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1337	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1338	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
1339	return VINF_SUCCESS;
1340	}
1341
1342	pReq->u.Out.pSession = NULL;
1343	pReq->u.Out.uSessionVersion = SUPDRV_IDC_VERSION;
1344	pReq->u.Out.uDriverVersion = SUPDRV_IDC_VERSION;
1345	pReq->u.Out.uDriverRevision = VBOX_SVN_REV;
1346
1347	/*
1348	* On NT we will already have a session associated with the
1349	* client, just like with the SUP_IOCTL_COOKIE request, while
1350	* the other doesn't.
1351	*/
1352	#ifdef RT_OS_WINDOWS
1353	pReq->Hdr.rc = VINF_SUCCESS;
1354	#else
1355	AssertReturn(!pSession, VERR_INTERNAL_ERROR);
1356	pReq->Hdr.rc = supdrvCreateSession(pDevExt, false /* fUser */, &pSession);
1357	if (RT_FAILURE(pReq->Hdr.rc))
1358	{
1359	OSDBGPRINT(("SUPDRV_IDC_REQ_CONNECT: failed to create session, rc=%d\n", pReq->Hdr.rc));
1360	return VINF_SUCCESS;
1361	}
1362	#endif
1363
1364	pReq->u.Out.pSession = pSession;
1365	pReq->Hdr.pSession = pSession;
1366
1367	return VINF_SUCCESS;
1368	}
1369
1370	case SUPDRV_IDC_REQ_DISCONNECT:
1371	{
1372	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_DISCONNECT, sizeof(*pReqHdr));
1373
1374	#ifdef RT_OS_WINDOWS
1375	/* Windows will destroy the session when the file object is destroyed. */
1376	#else
1377	supdrvCloseSession(pDevExt, pSession);
1378	#endif
1379	return pReqHdr->rc = VINF_SUCCESS;
1380	}
1381
1382	case SUPDRV_IDC_REQ_GET_SYMBOL:
1383	{
1384	PSUPDRVIDCREQGETSYM pReq = (PSUPDRVIDCREQGETSYM)pReqHdr;
1385	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_GET_SYMBOL, sizeof(*pReq));
1386
1387	pReq->Hdr.rc = supdrvIDC_LdrGetSymbol(pDevExt, pSession, pReq);
1388	return VINF_SUCCESS;
1389	}
1390
1391	case SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY:
1392	{
1393	PSUPDRVIDCREQCOMPREGFACTORY pReq = (PSUPDRVIDCREQCOMPREGFACTORY)pReqHdr;
1394	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_REGISTER_FACTORY, sizeof(*pReq));
1395
1396	pReq->Hdr.rc = SUPR0ComponentRegisterFactory(pSession, pReq->u.In.pFactory);
1397	return VINF_SUCCESS;
1398	}
1399
1400	case SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY:
1401	{
1402	PSUPDRVIDCREQCOMPDEREGFACTORY pReq = (PSUPDRVIDCREQCOMPDEREGFACTORY)pReqHdr;
1403	REQ_CHECK_IDC_SIZE(SUPDRV_IDC_REQ_COMPONENT_DEREGISTER_FACTORY, sizeof(*pReq));
1404
1405	pReq->Hdr.rc = SUPR0ComponentDeregisterFactory(pSession, pReq->u.In.pFactory);
1406	return VINF_SUCCESS;
1407	}
1408
1409	default:
1410	Log(("Unknown IDC %#lx\n", (long)uReq));
1411	break;
1412	}
1413
1414	#undef REQ_CHECK_IDC_SIZE
1415	return VERR_NOT_SUPPORTED;
1416	}
1417
1418
1419	/**
1420	* Register a object for reference counting.
1421	* The object is registered with one reference in the specified session.
1422	*
1423	* @returns Unique identifier on success (pointer).
1424	* All future reference must use this identifier.
1425	* @returns NULL on failure.
1426	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
1427	* @param pvUser1 The first user argument.
1428	* @param pvUser2 The second user argument.
1429	*/
1430	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
1431	{
1432	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1433	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1434	PSUPDRVOBJ pObj;
1435	PSUPDRVUSAGE pUsage;
1436
1437	/*
1438	* Validate the input.
1439	*/
1440	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
1441	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
1442	AssertPtrReturn(pfnDestructor, NULL);
1443
1444	/*
1445	* Allocate and initialize the object.
1446	*/
1447	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
1448	if (!pObj)
1449	return NULL;
1450	pObj->u32Magic = SUPDRVOBJ_MAGIC;
1451	pObj->enmType = enmType;
1452	pObj->pNext = NULL;
1453	pObj->cUsage = 1;
1454	pObj->pfnDestructor = pfnDestructor;
1455	pObj->pvUser1 = pvUser1;
1456	pObj->pvUser2 = pvUser2;
1457	pObj->CreatorUid = pSession->Uid;
1458	pObj->CreatorGid = pSession->Gid;
1459	pObj->CreatorProcess= pSession->Process;
1460	supdrvOSObjInitCreator(pObj, pSession);
1461
1462	/*
1463	* Allocate the usage record.
1464	* (We keep freed usage records around to simplify SUPR0ObjAddRef().)
1465	*/
1466	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1467
1468	pUsage = pDevExt->pUsageFree;
1469	if (pUsage)
1470	pDevExt->pUsageFree = pUsage->pNext;
1471	else
1472	{
1473	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1474	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
1475	if (!pUsage)
1476	{
1477	RTMemFree(pObj);
1478	return NULL;
1479	}
1480	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1481	}
1482
1483	/*
1484	* Insert the object and create the session usage record.
1485	*/
1486	/* The object. */
1487	pObj->pNext = pDevExt->pObjs;
1488	pDevExt->pObjs = pObj;
1489
1490	/* The session record. */
1491	pUsage->cUsage = 1;
1492	pUsage->pObj = pObj;
1493	pUsage->pNext = pSession->pUsage;
1494	/* Log2(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext)); */
1495	pSession->pUsage = pUsage;
1496
1497	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1498
1499	Log(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
1500	return pObj;
1501	}
1502
1503
1504	/**
1505	* Increment the reference counter for the object associating the reference
1506	* with the specified session.
1507	*
1508	* @returns IPRT status code.
1509	* @param pvObj The identifier returned by SUPR0ObjRegister().
1510	* @param pSession The session which is referencing the object.
1511	*
1512	* @remarks The caller should not own any spinlocks and must carefully protect
1513	* itself against potential race with the destructor so freed memory
1514	* isn't accessed here.
1515	*/
1516	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
1517	{
1518	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1519	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1520	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
1521	PSUPDRVUSAGE pUsagePre;
1522	PSUPDRVUSAGE pUsage;
1523
1524	/*
1525	* Validate the input.
1526	* Be ready for the destruction race (someone might be stuck in the
1527	* destructor waiting a lock we own).
1528	*/
1529	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1530	AssertPtrReturn(pObj, VERR_INVALID_POINTER);
1531	AssertMsgReturn(pObj->u32Magic == SUPDRVOBJ_MAGIC \|\| pObj->u32Magic == SUPDRVOBJ_MAGIC + 1,
1532	("Invalid pvObj=%p magic=%#x (expected %#x or %#x)\n", pvObj, pObj->u32Magic, SUPDRVOBJ_MAGIC, SUPDRVOBJ_MAGIC + 1),
1533	VERR_INVALID_PARAMETER);
1534
1535	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1536
1537	if (RT_UNLIKELY(pObj->u32Magic != SUPDRVOBJ_MAGIC))
1538	{
1539	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1540
1541	AssertMsgFailed(("pvObj=%p magic=%#x\n", pvObj, pObj->u32Magic));
1542	return VERR_WRONG_ORDER;
1543	}
1544
1545	/*
1546	* Preallocate the usage record.
1547	*/
1548	pUsagePre = pDevExt->pUsageFree;
1549	if (pUsagePre)
1550	pDevExt->pUsageFree = pUsagePre->pNext;
1551	else
1552	{
1553	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1554	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
1555	if (!pUsagePre)
1556	return VERR_NO_MEMORY;
1557
1558	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1559	}
1560
1561	/*
1562	* Reference the object.
1563	*/
1564	pObj->cUsage++;
1565
1566	/*
1567	* Look for the session record.
1568	*/
1569	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
1570	{
1571	/Log(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
1572	if (pUsage->pObj == pObj)
1573	break;
1574	}
1575	if (pUsage)
1576	pUsage->cUsage++;
1577	else
1578	{
1579	/* create a new session record. */
1580	pUsagePre->cUsage = 1;
1581	pUsagePre->pObj = pObj;
1582	pUsagePre->pNext = pSession->pUsage;
1583	pSession->pUsage = pUsagePre;
1584	/Log(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));/
1585
1586	pUsagePre = NULL;
1587	}
1588
1589	/*
1590	* Put any unused usage record into the free list..
1591	*/
1592	if (pUsagePre)
1593	{
1594	pUsagePre->pNext = pDevExt->pUsageFree;
1595	pDevExt->pUsageFree = pUsagePre;
1596	}
1597
1598	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1599
1600	return VINF_SUCCESS;
1601	}
1602
1603
1604	/**
1605	* Decrement / destroy a reference counter record for an object.
1606	*
1607	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
1608	*
1609	* @returns IPRT status code.
1610	* @param pvObj The identifier returned by SUPR0ObjRegister().
1611	* @param pSession The session which is referencing the object.
1612	*/
1613	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
1614	{
1615	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1616	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1617	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
1618	bool fDestroy = false;
1619	PSUPDRVUSAGE pUsage;
1620	PSUPDRVUSAGE pUsagePrev;
1621
1622	/*
1623	* Validate the input.
1624	*/
1625	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1626	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
1627	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
1628	VERR_INVALID_PARAMETER);
1629
1630	/*
1631	* Acquire the spinlock and look for the usage record.
1632	*/
1633	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1634
1635	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
1636	pUsage;
1637	pUsagePrev = pUsage, pUsage = pUsage->pNext)
1638	{
1639	/Log2(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));/
1640	if (pUsage->pObj == pObj)
1641	{
1642	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
1643	if (pUsage->cUsage > 1)
1644	{
1645	pObj->cUsage--;
1646	pUsage->cUsage--;
1647	}
1648	else
1649	{
1650	/*
1651	* Free the session record.
1652	*/
1653	if (pUsagePrev)
1654	pUsagePrev->pNext = pUsage->pNext;
1655	else
1656	pSession->pUsage = pUsage->pNext;
1657	pUsage->pNext = pDevExt->pUsageFree;
1658	pDevExt->pUsageFree = pUsage;
1659
1660	/* What about the object? */
1661	if (pObj->cUsage > 1)
1662	pObj->cUsage--;
1663	else
1664	{
1665	/*
1666	* Object is to be destroyed, unlink it.
1667	*/
1668	pObj->u32Magic = SUPDRVOBJ_MAGIC + 1;
1669	fDestroy = true;
1670	if (pDevExt->pObjs == pObj)
1671	pDevExt->pObjs = pObj->pNext;
1672	else
1673	{
1674	PSUPDRVOBJ pObjPrev;
1675	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
1676	if (pObjPrev->pNext == pObj)
1677	{
1678	pObjPrev->pNext = pObj->pNext;
1679	break;
1680	}
1681	Assert(pObjPrev);
1682	}
1683	}
1684	}
1685	break;
1686	}
1687	}
1688
1689	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1690
1691	/*
1692	* Call the destructor and free the object if required.
1693	*/
1694	if (fDestroy)
1695	{
1696	Log(("SUPR0ObjRelease: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
1697	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
1698	if (pObj->pfnDestructor)
1699	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
1700	RTMemFree(pObj);
1701	}
1702
1703	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
1704	return pUsage ? VINF_SUCCESS : VERR_INVALID_PARAMETER;
1705	}
1706
1707	/**
1708	* Verifies that the current process can access the specified object.
1709	*
1710	* @returns The following IPRT status code:
1711	* @retval VINF_SUCCESS if access was granted.
1712	* @retval VERR_PERMISSION_DENIED if denied access.
1713	* @retval VERR_INVALID_PARAMETER if invalid parameter.
1714	*
1715	* @param pvObj The identifier returned by SUPR0ObjRegister().
1716	* @param pSession The session which wishes to access the object.
1717	* @param pszObjName Object string name. This is optional and depends on the object type.
1718	*
1719	* @remark The caller is responsible for making sure the object isn't removed while
1720	* we're inside this function. If uncertain about this, just call AddRef before calling us.
1721	*/
1722	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
1723	{
1724	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
1725	int rc;
1726
1727	/*
1728	* Validate the input.
1729	*/
1730	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1731	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
1732	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
1733	VERR_INVALID_PARAMETER);
1734
1735	/*
1736	* Check access. (returns true if a decision has been made.)
1737	*/
1738	rc = VERR_INTERNAL_ERROR;
1739	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
1740	return rc;
1741
1742	/*
1743	* Default policy is to allow the user to access his own
1744	* stuff but nothing else.
1745	*/
1746	if (pObj->CreatorUid == pSession->Uid)
1747	return VINF_SUCCESS;
1748	return VERR_PERMISSION_DENIED;
1749	}
1750
1751
1752	/**
1753	* Lock pages.
1754	*
1755	* @returns IPRT status code.
1756	* @param pSession Session to which the locked memory should be associated.
1757	* @param pvR3 Start of the memory range to lock.
1758	* This must be page aligned.
1759	* @param cb Size of the memory range to lock.
1760	* This must be page aligned.
1761	*/
1762	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
1763	{
1764	int rc;
1765	SUPDRVMEMREF Mem = {0};
1766	const size_t cb = (size_t)cPages << PAGE_SHIFT;
1767	LogFlow(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
1768
1769	/*
1770	* Verify input.
1771	*/
1772	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1773	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
1774	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
1775	\|\| !pvR3)
1776	{
1777	Log(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
1778	return VERR_INVALID_PARAMETER;
1779	}
1780
1781	#ifdef RT_OS_WINDOWS /* A temporary hack for windows, will be removed once all ring-3 code has been cleaned up. */
1782	/* First check if we allocated it using SUPPageAlloc; if so then we don't need to lock it again */
1783	rc = supdrvPageGetPhys(pSession, pvR3, cPages, paPages);
1784	if (RT_SUCCESS(rc))
1785	return rc;
1786	#endif
1787
1788	/*
1789	* Let IPRT do the job.
1790	*/
1791	Mem.eType = MEMREF_TYPE_LOCKED;
1792	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTR0ProcHandleSelf());
1793	if (RT_SUCCESS(rc))
1794	{
1795	uint32_t iPage = cPages;
1796	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
1797	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
1798
1799	while (iPage-- > 0)
1800	{
1801	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
1802	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
1803	{
1804	AssertMsgFailed(("iPage=%d\n", iPage));
1805	rc = VERR_INTERNAL_ERROR;
1806	break;
1807	}
1808	}
1809	if (RT_SUCCESS(rc))
1810	rc = supdrvMemAdd(&Mem, pSession);
1811	if (RT_FAILURE(rc))
1812	{
1813	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
1814	AssertRC(rc2);
1815	}
1816	}
1817
1818	return rc;
1819	}
1820
1821
1822	/**
1823	* Unlocks the memory pointed to by pv.
1824	*
1825	* @returns IPRT status code.
1826	* @param pSession Session to which the memory was locked.
1827	* @param pvR3 Memory to unlock.
1828	*/
1829	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
1830	{
1831	LogFlow(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
1832	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1833	#ifdef RT_OS_WINDOWS
1834	/*
1835	* Temporary hack for windows - SUPR0PageFree will unlock SUPR0PageAlloc
1836	* allocations; ignore this call.
1837	*/
1838	if (supdrvPageWasLockedByPageAlloc(pSession, pvR3))
1839	{
1840	LogFlow(("Page will be unlocked in SUPR0PageFree -> ignore\n"));
1841	return VINF_SUCCESS;
1842	}
1843	#endif
1844	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
1845	}
1846
1847
1848	/**
1849	* Allocates a chunk of page aligned memory with contiguous and fixed physical
1850	* backing.
1851	*
1852	* @returns IPRT status code.
1853	* @param pSession Session data.
1854	* @param cb Number of bytes to allocate.
1855	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
1856	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
1857	* @param pHCPhys Where to put the physical address of allocated memory.
1858	*/
1859	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
1860	{
1861	int rc;
1862	SUPDRVMEMREF Mem = {0};
1863	LogFlow(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
1864
1865	/*
1866	* Validate input.
1867	*/
1868	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1869	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
1870	{
1871	Log(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
1872	pSession, ppvR0, ppvR3, pHCPhys));
1873	return VERR_INVALID_PARAMETER;
1874
1875	}
1876	if (cPages < 1 \|\| cPages >= 256)
1877	{
1878	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256\n", cPages));
1879	return VERR_INVALID_PARAMETER;
1880	}
1881
1882	/*
1883	* Let IPRT do the job.
1884	*/
1885	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
1886	if (RT_SUCCESS(rc))
1887	{
1888	int rc2;
1889	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
1890	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
1891	if (RT_SUCCESS(rc))
1892	{
1893	Mem.eType = MEMREF_TYPE_CONT;
1894	rc = supdrvMemAdd(&Mem, pSession);
1895	if (!rc)
1896	{
1897	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
1898	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
1899	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
1900	return 0;
1901	}
1902
1903	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
1904	AssertRC(rc2);
1905	}
1906	rc2 = RTR0MemObjFree(Mem.MemObj, false);
1907	AssertRC(rc2);
1908	}
1909
1910	return rc;
1911	}
1912
1913
1914	/**
1915	* Frees memory allocated using SUPR0ContAlloc().
1916	*
1917	* @returns IPRT status code.
1918	* @param pSession The session to which the memory was allocated.
1919	* @param uPtr Pointer to the memory (ring-3 or ring-0).
1920	*/
1921	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
1922	{
1923	LogFlow(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
1924	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1925	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
1926	}
1927
1928
1929	/**
1930	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
1931	*
1932	* The memory isn't zeroed.
1933	*
1934	* @returns IPRT status code.
1935	* @param pSession Session data.
1936	* @param cPages Number of pages to allocate.
1937	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
1938	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
1939	* @param paPages Where to put the physical addresses of allocated memory.
1940	*/
1941	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
1942	{
1943	unsigned iPage;
1944	int rc;
1945	SUPDRVMEMREF Mem = {0};
1946	LogFlow(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
1947
1948	/*
1949	* Validate input.
1950	*/
1951	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1952	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
1953	{
1954	Log(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
1955	pSession, ppvR3, ppvR0, paPages));
1956	return VERR_INVALID_PARAMETER;
1957
1958	}
1959	if (cPages < 1 \|\| cPages > 256)
1960	{
1961	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
1962	return VERR_INVALID_PARAMETER;
1963	}
1964
1965	/*
1966	* Let IPRT do the work.
1967	*/
1968	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
1969	if (RT_SUCCESS(rc))
1970	{
1971	int rc2;
1972	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
1973	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
1974	if (RT_SUCCESS(rc))
1975	{
1976	Mem.eType = MEMREF_TYPE_LOW;
1977	rc = supdrvMemAdd(&Mem, pSession);
1978	if (!rc)
1979	{
1980	for (iPage = 0; iPage < cPages; iPage++)
1981	{
1982	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
1983	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%VHp\n", paPages[iPage]));
1984	}
1985	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
1986	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
1987	return 0;
1988	}
1989
1990	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
1991	AssertRC(rc2);
1992	}
1993
1994	rc2 = RTR0MemObjFree(Mem.MemObj, false);
1995	AssertRC(rc2);
1996	}
1997
1998	return rc;
1999	}
2000
2001
2002	/**
2003	* Frees memory allocated using SUPR0LowAlloc().
2004	*
2005	* @returns IPRT status code.
2006	* @param pSession The session to which the memory was allocated.
2007	* @param uPtr Pointer to the memory (ring-3 or ring-0).
2008	*/
2009	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2010	{
2011	LogFlow(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2012	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2013	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
2014	}
2015
2016
2017
2018	/**
2019	* Allocates a chunk of memory with both R0 and R3 mappings.
2020	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
2021	*
2022	* @returns IPRT status code.
2023	* @param pSession The session to associated the allocation with.
2024	* @param cb Number of bytes to allocate.
2025	* @param ppvR0 Where to store the address of the Ring-0 mapping.
2026	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2027	*/
2028	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
2029	{
2030	int rc;
2031	SUPDRVMEMREF Mem = {0};
2032	LogFlow(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
2033
2034	/*
2035	* Validate input.
2036	*/
2037	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2038	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
2039	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
2040	if (cb < 1 \|\| cb >= _4M)
2041	{
2042	Log(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
2043	return VERR_INVALID_PARAMETER;
2044	}
2045
2046	/*
2047	* Let IPRT do the work.
2048	*/
2049	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
2050	if (RT_SUCCESS(rc))
2051	{
2052	int rc2;
2053	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2054	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2055	if (RT_SUCCESS(rc))
2056	{
2057	Mem.eType = MEMREF_TYPE_MEM;
2058	rc = supdrvMemAdd(&Mem, pSession);
2059	if (!rc)
2060	{
2061	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
2062	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2063	return VINF_SUCCESS;
2064	}
2065	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2066	AssertRC(rc2);
2067	}
2068
2069	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2070	AssertRC(rc2);
2071	}
2072
2073	return rc;
2074	}
2075
2076
2077	/**
2078	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
2079	*
2080	* @returns IPRT status code.
2081	* @param pSession The session to which the memory was allocated.
2082	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2083	* @param paPages Where to store the physical addresses.
2084	*/
2085	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
2086	{
2087	PSUPDRVBUNDLE pBundle;
2088	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2089	LogFlow(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
2090
2091	/*
2092	* Validate input.
2093	*/
2094	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2095	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
2096	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
2097
2098	/*
2099	* Search for the address.
2100	*/
2101	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2102	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2103	{
2104	if (pBundle->cUsed > 0)
2105	{
2106	unsigned i;
2107	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2108	{
2109	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
2110	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2111	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2112	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2113	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
2114	)
2115	)
2116	{
2117	const unsigned cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2118	unsigned iPage;
2119	for (iPage = 0; iPage < cPages; iPage++)
2120	{
2121	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2122	paPages[iPage].uReserved = 0;
2123	}
2124	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2125	return VINF_SUCCESS;
2126	}
2127	}
2128	}
2129	}
2130	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2131	Log(("Failed to find %p!!!\n", (void *)uPtr));
2132	return VERR_INVALID_PARAMETER;
2133	}
2134
2135
2136	/**
2137	* Free memory allocated by SUPR0MemAlloc().
2138	*
2139	* @returns IPRT status code.
2140	* @param pSession The session owning the allocation.
2141	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
2142	*/
2143	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
2144	{
2145	LogFlow(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
2146	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2147	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
2148	}
2149
2150
2151	/**
2152	* Allocates a chunk of memory with only a R3 mappings.
2153	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
2154	*
2155	* @returns IPRT status code.
2156	* @param pSession The session to associated the allocation with.
2157	* @param cPages The number of pages to allocate.
2158	* @param ppvR3 Where to store the address of the Ring-3 mapping.
2159	* @param paPages Where to store the addresses of the pages. Optional.
2160	*/
2161	SUPR0DECL(int) SUPR0PageAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR3PTR ppvR3, PRTHCPHYS paPages)
2162	{
2163	int rc;
2164	SUPDRVMEMREF Mem = {0};
2165	LogFlow(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
2166
2167	/*
2168	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
2169	*/
2170	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2171	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
2172	if (cPages < 1 \|\| cPages > (128 * _1M)/PAGE_SIZE)
2173	{
2174	Log(("SUPR0PageAlloc: Illegal request cb=%u; must be greater than 0 and smaller than 128MB.\n", cPages));
2175	return VERR_INVALID_PARAMETER;
2176	}
2177
2178	/*
2179	* Let IPRT do the work.
2180	*/
2181	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
2182	if (RT_SUCCESS(rc))
2183	{
2184	int rc2;
2185	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
2186	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
2187	if (RT_SUCCESS(rc))
2188	{
2189	Mem.eType = MEMREF_TYPE_LOCKED_SUP;
2190	rc = supdrvMemAdd(&Mem, pSession);
2191	if (!rc)
2192	{
2193	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
2194	if (paPages)
2195	{
2196	uint32_t iPage = cPages;
2197	while (iPage-- > 0)
2198	{
2199	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
2200	Assert(paPages[iPage] != NIL_RTHCPHYS);
2201	}
2202	}
2203	return VINF_SUCCESS;
2204	}
2205	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
2206	AssertRC(rc2);
2207	}
2208
2209	rc2 = RTR0MemObjFree(Mem.MemObj, false);
2210	AssertRC(rc2);
2211	}
2212	return rc;
2213	}
2214
2215
2216	#ifdef RT_OS_WINDOWS
2217	/**
2218	* Check if the pages were locked by SUPR0PageAlloc
2219	*
2220	* This function will be removed along with the lock/unlock hacks when
2221	* we've cleaned up the ring-3 code properly.
2222	*
2223	* @returns boolean
2224	* @param pSession The session to which the memory was allocated.
2225	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc().
2226	*/
2227	static bool supdrvPageWasLockedByPageAlloc(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2228	{
2229	PSUPDRVBUNDLE pBundle;
2230	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2231	LogFlow(("SUPR0PageIsLockedByPageAlloc: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2232
2233	/*
2234	* Search for the address.
2235	*/
2236	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2237	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2238	{
2239	if (pBundle->cUsed > 0)
2240	{
2241	unsigned i;
2242	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2243	{
2244	if ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED_SUP
2245	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2246	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2247	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
2248	{
2249	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2250	return true;
2251	}
2252	}
2253	}
2254	}
2255	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2256	return false;
2257	}
2258
2259
2260	/**
2261	* Get the physical addresses of memory allocated using SUPR0PageAlloc().
2262	*
2263	* This function will be removed along with the lock/unlock hacks when
2264	* we've cleaned up the ring-3 code properly.
2265	*
2266	* @returns IPRT status code.
2267	* @param pSession The session to which the memory was allocated.
2268	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc().
2269	* @param cPages Number of pages in paPages
2270	* @param paPages Where to store the physical addresses.
2271	*/
2272	static int supdrvPageGetPhys(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
2273	{
2274	PSUPDRVBUNDLE pBundle;
2275	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2276	LogFlow(("supdrvPageGetPhys: pSession=%p pvR3=%p cPages=%#lx paPages=%p\n", pSession, (void *)pvR3, (long)cPages, paPages));
2277
2278	/*
2279	* Search for the address.
2280	*/
2281	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2282	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2283	{
2284	if (pBundle->cUsed > 0)
2285	{
2286	unsigned i;
2287	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2288	{
2289	if ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED_SUP
2290	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2291	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2292	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
2293	{
2294	uint32_t iPage = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2295	cPages = RT_MIN(iPage, cPages);
2296	for (iPage = 0; iPage < cPages; iPage++)
2297	paPages[iPage] = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2298	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2299	return VINF_SUCCESS;
2300	}
2301	}
2302	}
2303	}
2304	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2305	return VERR_INVALID_PARAMETER;
2306	}
2307	#endif /* RT_OS_WINDOWS */
2308
2309
2310	/**
2311	* Free memory allocated by SUPR0PageAlloc().
2312	*
2313	* @returns IPRT status code.
2314	* @param pSession The session owning the allocation.
2315	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc().
2316	*/
2317	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2318	{
2319	LogFlow(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2320	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2321	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED_SUP);
2322	}
2323
2324
2325	/**
2326	* Maps the GIP into userspace and/or get the physical address of the GIP.
2327	*
2328	* @returns IPRT status code.
2329	* @param pSession Session to which the GIP mapping should belong.
2330	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
2331	* @param pHCPhysGip Where to store the physical address. (optional)
2332	*
2333	* @remark There is no reference counting on the mapping, so one call to this function
2334	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
2335	* and remove the session as a GIP user.
2336	*/
2337	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
2338	{
2339	int rc = 0;
2340	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2341	RTR3PTR pGip = NIL_RTR3PTR;
2342	RTHCPHYS HCPhys = NIL_RTHCPHYS;
2343	LogFlow(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
2344
2345	/*
2346	* Validate
2347	*/
2348	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2349	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
2350	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
2351
2352	RTSemFastMutexRequest(pDevExt->mtxGip);
2353	if (pDevExt->pGip)
2354	{
2355	/*
2356	* Map it?
2357	*/
2358	if (ppGipR3)
2359	{
2360	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
2361	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
2362	RTMEM_PROT_READ, RTR0ProcHandleSelf());
2363	if (RT_SUCCESS(rc))
2364	{
2365	pGip = RTR0MemObjAddressR3(pSession->GipMapObjR3);
2366	rc = VINF_SUCCESS; /** @todo remove this and replace the !rc below with RT_SUCCESS(rc). */
2367	}
2368	}
2369
2370	/*
2371	* Get physical address.
2372	*/
2373	if (pHCPhysGip && !rc)
2374	HCPhys = pDevExt->HCPhysGip;
2375
2376	/*
2377	* Reference globally.
2378	*/
2379	if (!pSession->fGipReferenced && !rc)
2380	{
2381	pSession->fGipReferenced = 1;
2382	pDevExt->cGipUsers++;
2383	if (pDevExt->cGipUsers == 1)
2384	{
2385	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
2386	unsigned i;
2387
2388	LogFlow(("SUPR0GipMap: Resumes GIP updating\n"));
2389
2390	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
2391	ASMAtomicXchgU32(&pGip->aCPUs[i].u32TransactionId, pGip->aCPUs[i].u32TransactionId & ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
2392	ASMAtomicXchgU64(&pGip->u64NanoTSLastUpdateHz, 0);
2393
2394	rc = RTTimerStart(pDevExt->pGipTimer, 0);
2395	AssertRC(rc); rc = VINF_SUCCESS;
2396	}
2397	}
2398	}
2399	else
2400	{
2401	rc = SUPDRV_ERR_GENERAL_FAILURE;
2402	Log(("SUPR0GipMap: GIP is not available!\n"));
2403	}
2404	RTSemFastMutexRelease(pDevExt->mtxGip);
2405
2406	/*
2407	* Write returns.
2408	*/
2409	if (pHCPhysGip)
2410	*pHCPhysGip = HCPhys;
2411	if (ppGipR3)
2412	*ppGipR3 = pGip;
2413
2414	#ifdef DEBUG_DARWIN_GIP
2415	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx ppGip=%p GipMapObjR3\n", rc, (unsigned long)HCPhys, pGip, pSession->GipMapObjR3));
2416	#else
2417	LogFlow(("SUPR0GipMap: returns %d pHCPhysGip=%lx ppGipR3=%p\n", rc, (unsigned long)HCPhys, (void *)(uintptr_t)pGip));
2418	#endif
2419	return rc;
2420	}
2421
2422
2423	/**
2424	* Unmaps any user mapping of the GIP and terminates all GIP access
2425	* from this session.
2426	*
2427	* @returns IPRT status code.
2428	* @param pSession Session to which the GIP mapping should belong.
2429	*/
2430	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
2431	{
2432	int rc = VINF_SUCCESS;
2433	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2434	#ifdef DEBUG_DARWIN_GIP
2435	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
2436	pSession,
2437	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
2438	pSession->GipMapObjR3));
2439	#else
2440	LogFlow(("SUPR0GipUnmap: pSession=%p\n", pSession));
2441	#endif
2442	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2443
2444	RTSemFastMutexRequest(pDevExt->mtxGip);
2445
2446	/*
2447	* Unmap anything?
2448	*/
2449	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
2450	{
2451	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
2452	AssertRC(rc);
2453	if (RT_SUCCESS(rc))
2454	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
2455	}
2456
2457	/*
2458	* Dereference global GIP.
2459	*/
2460	if (pSession->fGipReferenced && !rc)
2461	{
2462	pSession->fGipReferenced = 0;
2463	if ( pDevExt->cGipUsers > 0
2464	&& !--pDevExt->cGipUsers)
2465	{
2466	LogFlow(("SUPR0GipUnmap: Suspends GIP updating\n"));
2467	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = 0;
2468	}
2469	}
2470
2471	RTSemFastMutexRelease(pDevExt->mtxGip);
2472
2473	return rc;
2474	}
2475
2476
2477	/**
2478	* Register a component factory with the support driver.
2479	*
2480	* This is currently restricted to kernel sessions only.
2481	*
2482	* @returns VBox status code.
2483	* @retval VINF_SUCCESS on success.
2484	* @retval VERR_NO_MEMORY if we're out of memory.
2485	* @retval VERR_ALREADY_EXISTS if the factory has already been registered.
2486	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
2487	* @retval VERR_INVALID_PARAMETER on invalid parameter.
2488	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
2489	*
2490	* @param pSession The SUPDRV session (must be a ring-0 session).
2491	* @param pFactory Pointer to the component factory registration structure.
2492	*
2493	* @remarks This interface is also available via SUPR0IdcComponentRegisterFactory.
2494	*/
2495	SUPR0DECL(int) SUPR0ComponentRegisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
2496	{
2497	PSUPDRVFACTORYREG pNewReg;
2498	const char *psz;
2499	int rc;
2500
2501	/*
2502	* Validate parameters.
2503	*/
2504	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2505	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
2506	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
2507	AssertPtrReturn(pFactory->pfnQueryFactoryInterface, VERR_INVALID_POINTER);
2508	psz = (const char *)memchr(pFactory->szName, '\0', sizeof(pFactory->szName));
2509	AssertReturn(psz, VERR_INVALID_PARAMETER);
2510
2511	/*
2512	* Allocate and initialize a new registration structure.
2513	*/
2514	pNewReg = (PSUPDRVFACTORYREG)RTMemAlloc(sizeof(SUPDRVFACTORYREG));
2515	if (pNewReg)
2516	{
2517	pNewReg->pNext = NULL;
2518	pNewReg->pFactory = pFactory;
2519	pNewReg->pSession = pSession;
2520	pNewReg->cchName = psz - &pFactory->szName[0];
2521
2522	/*
2523	* Add it to the tail of the list after checking for prior registration.
2524	*/
2525	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
2526	if (RT_SUCCESS(rc))
2527	{
2528	PSUPDRVFACTORYREG pPrev = NULL;
2529	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
2530	while (pCur && pCur->pFactory != pFactory)
2531	{
2532	pPrev = pCur;
2533	pCur = pCur->pNext;
2534	}
2535	if (!pCur)
2536	{
2537	if (pPrev)
2538	pPrev->pNext = pNewReg;
2539	else
2540	pSession->pDevExt->pComponentFactoryHead = pNewReg;
2541	rc = VINF_SUCCESS;
2542	}
2543	else
2544	rc = VERR_ALREADY_EXISTS;
2545
2546	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
2547	}
2548
2549	if (RT_FAILURE(rc))
2550	RTMemFree(pNewReg);
2551	}
2552	else
2553	rc = VERR_NO_MEMORY;
2554	return rc;
2555	}
2556
2557
2558	/**
2559	* Deregister a component factory.
2560	*
2561	* @returns VBox status code.
2562	* @retval VINF_SUCCESS on success.
2563	* @retval VERR_NOT_FOUND if the factory wasn't registered.
2564	* @retval VERR_ACCESS_DENIED if it isn't a kernel session.
2565	* @retval VERR_INVALID_PARAMETER on invalid parameter.
2566	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
2567	*
2568	* @param pSession The SUPDRV session (must be a ring-0 session).
2569	* @param pFactory Pointer to the component factory registration structure
2570	* previously passed SUPR0ComponentRegisterFactory().
2571	*
2572	* @remarks This interface is also available via SUPR0IdcComponentDeregisterFactory.
2573	*/
2574	SUPR0DECL(int) SUPR0ComponentDeregisterFactory(PSUPDRVSESSION pSession, PCSUPDRVFACTORY pFactory)
2575	{
2576	int rc;
2577
2578	/*
2579	* Validate parameters.
2580	*/
2581	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2582	AssertReturn(pSession->R0Process == NIL_RTR0PROCESS, VERR_ACCESS_DENIED);
2583	AssertPtrReturn(pFactory, VERR_INVALID_POINTER);
2584
2585	/*
2586	* Take the lock and look for the registration record.
2587	*/
2588	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
2589	if (RT_SUCCESS(rc))
2590	{
2591	PSUPDRVFACTORYREG pPrev = NULL;
2592	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
2593	while (pCur && pCur->pFactory != pFactory)
2594	{
2595	pPrev = pCur;
2596	pCur = pCur->pNext;
2597	}
2598	if (pCur)
2599	{
2600	if (!pPrev)
2601	pSession->pDevExt->pComponentFactoryHead = pCur->pNext;
2602	else
2603	pPrev->pNext = pCur->pNext;
2604
2605	pCur->pNext = NULL;
2606	pCur->pFactory = NULL;
2607	pCur->pSession = NULL;
2608	rc = VINF_SUCCESS;
2609	}
2610	else
2611	rc = VERR_NOT_FOUND;
2612
2613	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
2614
2615	RTMemFree(pCur);
2616	}
2617	return rc;
2618	}
2619
2620
2621	/**
2622	* Queries a component factory.
2623	*
2624	* @returns VBox status code.
2625	* @retval VERR_INVALID_PARAMETER on invalid parameter.
2626	* @retval VERR_INVALID_POINTER on invalid pointer parameter.
2627	* @retval VERR_SUPDRV_COMPONENT_NOT_FOUND if the component factory wasn't found.
2628	* @retval VERR_SUPDRV_INTERFACE_NOT_SUPPORTED if the interface wasn't supported.
2629	*
2630	* @param pSession The SUPDRV session.
2631	* @param pszName The name of the component factory.
2632	* @param pszInterfaceUuid The UUID of the factory interface (stringified).
2633	* @param ppvFactoryIf Where to store the factory interface.
2634	*/
2635	SUPR0DECL(int) SUPR0ComponentQueryFactory(PSUPDRVSESSION pSession, const char pszName, const char pszInterfaceUuid, void **ppvFactoryIf)
2636	{
2637	const char *pszEnd;
2638	size_t cchName;
2639	int rc;
2640
2641	/*
2642	* Validate parameters.
2643	*/
2644	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2645
2646	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
2647	pszEnd = memchr(pszName, '\0', RT_SIZEOFMEMB(SUPDRVFACTORY, szName));
2648	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
2649	cchName = pszEnd - pszName;
2650
2651	AssertPtrReturn(pszInterfaceUuid, VERR_INVALID_POINTER);
2652	pszEnd = memchr(pszInterfaceUuid, '\0', RTUUID_STR_LENGTH);
2653	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
2654
2655	AssertPtrReturn(ppvFactoryIf, VERR_INVALID_POINTER);
2656	*ppvFactoryIf = NULL;
2657
2658	/*
2659	* Take the lock and try all factories by this name.
2660	*/
2661	rc = RTSemFastMutexRequest(pSession->pDevExt->mtxComponentFactory);
2662	if (RT_SUCCESS(rc))
2663	{
2664	PSUPDRVFACTORYREG pCur = pSession->pDevExt->pComponentFactoryHead;
2665	rc = VERR_SUPDRV_COMPONENT_NOT_FOUND;
2666	while (pCur)
2667	{
2668	if ( pCur->cchName == cchName
2669	&& !memcmp(pCur->pFactory->szName, pszName, cchName))
2670	{
2671	void *pvFactory = pCur->pFactory->pfnQueryFactoryInterface(pCur->pFactory, pSession, pszInterfaceUuid);
2672	if (pvFactory)
2673	{
2674	*ppvFactoryIf = pvFactory;
2675	rc = VINF_SUCCESS;
2676	break;
2677	}
2678	rc = VERR_SUPDRV_INTERFACE_NOT_SUPPORTED;
2679	}
2680
2681	/* next */
2682	pCur = pCur->pNext;
2683	}
2684
2685	RTSemFastMutexRelease(pSession->pDevExt->mtxComponentFactory);
2686	}
2687	return rc;
2688	}
2689
2690
2691	/**
2692	* Adds a memory object to the session.
2693	*
2694	* @returns IPRT status code.
2695	* @param pMem Memory tracking structure containing the
2696	* information to track.
2697	* @param pSession The session.
2698	*/
2699	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
2700	{
2701	PSUPDRVBUNDLE pBundle;
2702	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2703
2704	/*
2705	* Find free entry and record the allocation.
2706	*/
2707	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2708	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2709	{
2710	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
2711	{
2712	unsigned i;
2713	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2714	{
2715	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
2716	{
2717	pBundle->cUsed++;
2718	pBundle->aMem[i] = *pMem;
2719	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2720	return VINF_SUCCESS;
2721	}
2722	}
2723	AssertFailed(); /* !!this can't be happening!!! */
2724	}
2725	}
2726	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2727
2728	/*
2729	* Need to allocate a new bundle.
2730	* Insert into the last entry in the bundle.
2731	*/
2732	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
2733	if (!pBundle)
2734	return VERR_NO_MEMORY;
2735
2736	/* take last entry. */
2737	pBundle->cUsed++;
2738	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
2739
2740	/* insert into list. */
2741	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2742	pBundle->pNext = pSession->Bundle.pNext;
2743	pSession->Bundle.pNext = pBundle;
2744	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2745
2746	return VINF_SUCCESS;
2747	}
2748
2749
2750	/**
2751	* Releases a memory object referenced by pointer and type.
2752	*
2753	* @returns IPRT status code.
2754	* @param pSession Session data.
2755	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
2756	* @param eType Memory type.
2757	*/
2758	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
2759	{
2760	PSUPDRVBUNDLE pBundle;
2761	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2762
2763	/*
2764	* Validate input.
2765	*/
2766	if (!uPtr)
2767	{
2768	Log(("Illegal address %p\n", (void *)uPtr));
2769	return VERR_INVALID_PARAMETER;
2770	}
2771
2772	/*
2773	* Search for the address.
2774	*/
2775	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2776	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2777	{
2778	if (pBundle->cUsed > 0)
2779	{
2780	unsigned i;
2781	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2782	{
2783	if ( pBundle->aMem[i].eType == eType
2784	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2785	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2786	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2787	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
2788	)
2789	{
2790	/* Make a copy of it and release it outside the spinlock. */
2791	SUPDRVMEMREF Mem = pBundle->aMem[i];
2792	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
2793	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
2794	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
2795	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2796
2797	if (Mem.MapObjR3)
2798	{
2799	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
2800	AssertRC(rc); /** @todo figure out how to handle this. */
2801	}
2802	if (Mem.MemObj)
2803	{
2804	int rc = RTR0MemObjFree(Mem.MemObj, false);
2805	AssertRC(rc); /** @todo figure out how to handle this. */
2806	}
2807	return VINF_SUCCESS;
2808	}
2809	}
2810	}
2811	}
2812	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2813	Log(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
2814	return VERR_INVALID_PARAMETER;
2815	}
2816
2817
2818	#ifdef VBOX_WITH_IDT_PATCHING
2819	/**
2820	* Install IDT for the current CPU.
2821	*
2822	* @returns One of the following IPRT status codes:
2823	* @retval VINF_SUCCESS on success.
2824	* @retval VERR_IDT_FAILED.
2825	* @retval VERR_NO_MEMORY.
2826	* @param pDevExt The device extension.
2827	* @param pSession The session data.
2828	* @param pReq The request.
2829	*/
2830	static int supdrvIOCtl_IdtInstall(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPIDTINSTALL pReq)
2831	{
2832	PSUPDRVPATCHUSAGE pUsagePre;
2833	PSUPDRVPATCH pPatchPre;
2834	RTIDTR Idtr;
2835	PSUPDRVPATCH pPatch;
2836	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2837	LogFlow(("supdrvIOCtl_IdtInstall\n"));
2838
2839	/*
2840	* Preallocate entry for this CPU cause we don't wanna do
2841	* that inside the spinlock!
2842	*/
2843	pUsagePre = (PSUPDRVPATCHUSAGE)RTMemAlloc(sizeof(*pUsagePre));
2844	if (!pUsagePre)
2845	return VERR_NO_MEMORY;
2846
2847	/*
2848	* Take the spinlock and see what we need to do.
2849	*/
2850	RTSpinlockAcquireNoInts(pDevExt->Spinlock, &SpinlockTmp);
2851
2852	/* check if we already got a free patch. */
2853	if (!pDevExt->pIdtPatchesFree)
2854	{
2855	/*
2856	* Allocate a patch - outside the spinlock of course.
2857	*/
2858	RTSpinlockReleaseNoInts(pDevExt->Spinlock, &SpinlockTmp);
2859
2860	pPatchPre = (PSUPDRVPATCH)RTMemExecAlloc(sizeof(*pPatchPre));
2861	if (!pPatchPre)
2862	return VERR_NO_MEMORY;
2863
2864	RTSpinlockAcquireNoInts(pDevExt->Spinlock, &SpinlockTmp);
2865	}
2866	else
2867	{
2868	pPatchPre = pDevExt->pIdtPatchesFree;
2869	pDevExt->pIdtPatchesFree = pPatchPre->pNext;
2870	}
2871
2872	/* look for matching patch entry */
2873	ASMGetIDTR(&Idtr);
2874	pPatch = pDevExt->pIdtPatches;
2875	while (pPatch && pPatch->pvIdt != (void *)Idtr.pIdt)
2876	pPatch = pPatch->pNext;
2877
2878	if (!pPatch)
2879	{
2880	/*
2881	* Create patch.
2882	*/
2883	pPatch = supdrvIdtPatchOne(pDevExt, pPatchPre);
2884	if (pPatch)
2885	pPatchPre = NULL; /* mark as used. */
2886	}
2887	else
2888	{
2889	/*
2890	* Simply increment patch usage.
2891	*/
2892	pPatch->cUsage++;
2893	}
2894
2895	if (pPatch)
2896	{
2897	/*
2898	* Increment and add if need be the session usage record for this patch.
2899	*/
2900	PSUPDRVPATCHUSAGE pUsage = pSession->pPatchUsage;
2901	while (pUsage && pUsage->pPatch != pPatch)
2902	pUsage = pUsage->pNext;
2903
2904	if (!pUsage)
2905	{
2906	/*
2907	* Add usage record.
2908	*/
2909	pUsagePre->cUsage = 1;
2910	pUsagePre->pPatch = pPatch;
2911	pUsagePre->pNext = pSession->pPatchUsage;
2912	pSession->pPatchUsage = pUsagePre;
2913	pUsagePre = NULL; /* mark as used. */
2914	}
2915	else
2916	{
2917	/*
2918	* Increment usage count.
2919	*/
2920	pUsage->cUsage++;
2921	}
2922	}
2923
2924	/* free patch - we accumulate them for paranoid saftly reasons. */
2925	if (pPatchPre)
2926	{
2927	pPatchPre->pNext = pDevExt->pIdtPatchesFree;
2928	pDevExt->pIdtPatchesFree = pPatchPre;
2929	}
2930
2931	RTSpinlockReleaseNoInts(pDevExt->Spinlock, &SpinlockTmp);
2932
2933	/*
2934	* Free unused preallocated buffers.
2935	*/
2936	if (pUsagePre)
2937	RTMemFree(pUsagePre);
2938
2939	pReq->u.Out.u8Idt = pDevExt->u8Idt;
2940
2941	return pPatch ? VINF_SUCCESS : VERR_IDT_FAILED;
2942	}
2943
2944
2945	/**
2946	* This creates a IDT patch entry.
2947	* If the first patch being installed it'll also determin the IDT entry
2948	* to use.
2949	*
2950	* @returns pPatch on success.
2951	* @returns NULL on failure.
2952	* @param pDevExt Pointer to globals.
2953	* @param pPatch Patch entry to use.
2954	* This will be linked into SUPDRVDEVEXT::pIdtPatches on
2955	* successful return.
2956	* @remark Call must be owning the SUPDRVDEVEXT::Spinlock!
2957	*/
2958	static PSUPDRVPATCH supdrvIdtPatchOne(PSUPDRVDEVEXT pDevExt, PSUPDRVPATCH pPatch)
2959	{
2960	RTIDTR Idtr;
2961	PSUPDRVIDTE paIdt;
2962	LogFlow(("supdrvIOCtl_IdtPatchOne: pPatch=%p\n", pPatch));
2963
2964	/*
2965	* Get IDT.
2966	*/
2967	ASMGetIDTR(&Idtr);
2968	paIdt = (PSUPDRVIDTE)Idtr.pIdt;
2969	/*
2970	* Recent Linux kernels can be configured to 1G user /3G kernel.
2971	*/
2972	if ((uintptr_t)paIdt < 0x40000000)
2973	{
2974	AssertMsgFailed(("bad paIdt=%p\n", paIdt));
2975	return NULL;
2976	}
2977
2978	if (!pDevExt->u8Idt)
2979	{
2980	/*
2981	* Test out the alternatives.
2982	*
2983	* At the moment we do not support chaining thus we ASSUME that one of
2984	* these 48 entries is unused (which is not a problem on Win32 and
2985	* Linux to my knowledge).
2986	*/
2987	/** @todo we MUST change this detection to try grab an entry which is NOT in use. This can be
2988	* combined with gathering info about which guest system call gates we can hook up directly. */
2989	unsigned i;
2990	uint8_t u8Idt = 0;
2991	static uint8_t au8Ints[] =
2992	{
2993	#ifdef RT_OS_WINDOWS /* We don't use 0xef and above because they are system stuff on linux (ef is IPI,
2994	* local apic timer, or some other frequently fireing thing). */
2995	0xef, 0xee, 0xed, 0xec,
2996	#endif
2997	0xeb, 0xea, 0xe9, 0xe8,
2998	0xdf, 0xde, 0xdd, 0xdc,
2999	0x7b, 0x7a, 0x79, 0x78,
3000	0xbf, 0xbe, 0xbd, 0xbc,
3001	};
3002	#if defined(RT_ARCH_AMD64) && defined(DEBUG)
3003	static int s_iWobble = 0;
3004	unsigned iMax = !(s_iWobble++ % 2) ? 0x80 : 0x100;
3005	Log2(("IDT: Idtr=%p:%#x\n", (void *)Idtr.pIdt, (unsigned)Idtr.cbIdt));
3006	for (i = iMax - 0x80; i*16+15 < Idtr.cbIdt && i < iMax; i++)
3007	{
3008	Log2(("%#x: %04x:%08x%04x%04x P=%d DPL=%d IST=%d Type1=%#x u32Reserved=%#x u5Reserved=%#x\n",
3009	i, paIdt[i].u16SegSel, paIdt[i].u32OffsetTop, paIdt[i].u16OffsetHigh, paIdt[i].u16OffsetLow,
3010	paIdt[i].u1Present, paIdt[i].u2DPL, paIdt[i].u3IST, paIdt[i].u5Type2,
3011	paIdt[i].u32Reserved, paIdt[i].u5Reserved));
3012	}
3013	#endif
3014	/* look for entries which are not present or otherwise unused. */
3015	for (i = 0; i < sizeof(au8Ints) / sizeof(au8Ints[0]); i++)
3016	{
3017	u8Idt = au8Ints[i];
3018	if ( u8Idt * sizeof(SUPDRVIDTE) < Idtr.cbIdt
3019	&& ( !paIdt[u8Idt].u1Present
3020	\|\| paIdt[u8Idt].u5Type2 == 0))
3021	break;
3022	u8Idt = 0;
3023	}
3024	if (!u8Idt)
3025	{
3026	/* try again, look for a compatible entry .*/
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 == SUPDRV_IDTE_TYPE2_INTERRUPT_GATE
3033	&& !(paIdt[u8Idt].u16SegSel & 3))
3034	break;
3035	u8Idt = 0;
3036	}
3037	if (!u8Idt)
3038	{
3039	Log(("Failed to find appropirate IDT entry!!\n"));
3040	return NULL;
3041	}
3042	}
3043	pDevExt->u8Idt = u8Idt;
3044	LogFlow(("supdrvIOCtl_IdtPatchOne: u8Idt=%x\n", u8Idt));
3045	}
3046
3047	/*
3048	* Prepare the patch
3049	*/
3050	memset(pPatch, 0, sizeof(*pPatch));
3051	pPatch->pvIdt = paIdt;
3052	pPatch->cUsage = 1;
3053	pPatch->pIdtEntry = &paIdt[pDevExt->u8Idt];
3054	pPatch->SavedIdt = paIdt[pDevExt->u8Idt];
3055	pPatch->ChangedIdt.u16OffsetLow = (uint32_t)((uintptr_t)&pPatch->auCode[0] & 0xffff);
3056	pPatch->ChangedIdt.u16OffsetHigh = (uint32_t)((uintptr_t)&pPatch->auCode[0] >> 16);
3057	#ifdef RT_ARCH_AMD64
3058	pPatch->ChangedIdt.u32OffsetTop = (uint32_t)((uintptr_t)&pPatch->auCode[0] >> 32);
3059	#endif
3060	pPatch->ChangedIdt.u16SegSel = ASMGetCS();
3061	#ifdef RT_ARCH_AMD64
3062	pPatch->ChangedIdt.u3IST = 0;
3063	pPatch->ChangedIdt.u5Reserved = 0;
3064	#else /* x86 */
3065	pPatch->ChangedIdt.u5Reserved = 0;
3066	pPatch->ChangedIdt.u3Type1 = 0;
3067	#endif /* x86 */
3068	pPatch->ChangedIdt.u5Type2 = SUPDRV_IDTE_TYPE2_INTERRUPT_GATE;
3069	pPatch->ChangedIdt.u2DPL = 3;
3070	pPatch->ChangedIdt.u1Present = 1;
3071
3072	/*
3073	* Generate the patch code.
3074	*/
3075	{
3076	#ifdef RT_ARCH_AMD64
3077	union
3078	{
3079	uint8_t *pb;
3080	uint32_t *pu32;
3081	uint64_t *pu64;
3082	} u, uFixJmp, uFixCall, uNotNested;
3083	u.pb = &pPatch->auCode[0];
3084
3085	/* check the cookie */
3086	*u.pb++ = 0x3d; // cmp eax, GLOBALCOOKIE
3087	*u.pu32++ = pDevExt->u32Cookie;
3088
3089	*u.pb++ = 0x74; // jz @VBoxCall
3090	*u.pb++ = 2;
3091
3092	/* jump to forwarder code. */
3093	*u.pb++ = 0xeb;
3094	uFixJmp = u;
3095	*u.pb++ = 0xfe;
3096
3097	// @VBoxCall:
3098	*u.pb++ = 0x0f; // swapgs
3099	*u.pb++ = 0x01;
3100	*u.pb++ = 0xf8;
3101
3102	/*
3103	* Call VMMR0Entry
3104	* We don't have to push the arguments here, but we have top
3105	* reserve some stack space for the interrupt forwarding.
3106	*/
3107	# ifdef RT_OS_WINDOWS
3108	*u.pb++ = 0x50; // push rax ; alignment filler.
3109	*u.pb++ = 0x41; // push r8 ; uArg
3110	*u.pb++ = 0x50;
3111	*u.pb++ = 0x52; // push rdx ; uOperation
3112	*u.pb++ = 0x51; // push rcx ; pVM
3113	# else
3114	*u.pb++ = 0x51; // push rcx ; alignment filler.
3115	*u.pb++ = 0x52; // push rdx ; uArg
3116	*u.pb++ = 0x56; // push rsi ; uOperation
3117	*u.pb++ = 0x57; // push rdi ; pVM
3118	# endif
3119
3120	*u.pb++ = 0xff; // call qword [pfnVMMR0EntryInt wrt rip]
3121	*u.pb++ = 0x15;
3122	uFixCall = u;
3123	*u.pu32++ = 0;
3124
3125	*u.pb++ = 0x48; // add rsp, 20h ; remove call frame.
3126	*u.pb++ = 0x81;
3127	*u.pb++ = 0xc4;
3128	*u.pu32++ = 0x20;
3129
3130	*u.pb++ = 0x0f; // swapgs
3131	*u.pb++ = 0x01;
3132	*u.pb++ = 0xf8;
3133
3134	/* Return to R3. */
3135	uNotNested = u;
3136	*u.pb++ = 0x48; // iretq
3137	*u.pb++ = 0xcf;
3138
3139	while ((uintptr_t)u.pb & 0x7) // align 8
3140	*u.pb++ = 0xcc;
3141
3142	/* Pointer to the VMMR0Entry. */ // pfnVMMR0EntryInt dq StubVMMR0Entry
3143	*uFixCall.pu32 = (uint32_t)(u.pb - uFixCall.pb - 4); uFixCall.pb = NULL;
3144	pPatch->offVMMR0EntryFixup = (uint16_t)(u.pb - &pPatch->auCode[0]);
3145	*u.pu64++ = pDevExt->pvVMMR0 ? (uint64_t)pDevExt->pfnVMMR0EntryInt : (uint64_t)u.pb + 8;
3146
3147	/* stub entry. */ // StubVMMR0Entry:
3148	pPatch->offStub = (uint16_t)(u.pb - &pPatch->auCode[0]);
3149	*u.pb++ = 0x33; // xor eax, eax
3150	*u.pb++ = 0xc0;
3151
3152	*u.pb++ = 0x48; // dec rax
3153	*u.pb++ = 0xff;
3154	*u.pb++ = 0xc8;
3155
3156	*u.pb++ = 0xc3; // ret
3157
3158	/* forward to the original handler using a retf. */
3159	*uFixJmp.pb = (uint8_t)(u.pb - uFixJmp.pb - 1); uFixJmp.pb = NULL;
3160
3161	*u.pb++ = 0x68; // push <target cs>
3162	*u.pu32++ = !pPatch->SavedIdt.u5Type2 ? ASMGetCS() : pPatch->SavedIdt.u16SegSel;
3163
3164	*u.pb++ = 0x68; // push <low target rip>
3165	*u.pu32++ = !pPatch->SavedIdt.u5Type2
3166	? (uint32_t)(uintptr_t)uNotNested.pb
3167	: (uint32_t)pPatch->SavedIdt.u16OffsetLow
3168	\| (uint32_t)pPatch->SavedIdt.u16OffsetHigh << 16;
3169
3170	*u.pb++ = 0xc7; // mov dword [rsp + 4], <high target rip>
3171	*u.pb++ = 0x44;
3172	*u.pb++ = 0x24;
3173	*u.pb++ = 0x04;
3174	*u.pu32++ = !pPatch->SavedIdt.u5Type2
3175	? (uint32_t)((uint64_t)uNotNested.pb >> 32)
3176	: pPatch->SavedIdt.u32OffsetTop;
3177
3178	*u.pb++ = 0x48; // retf ; does this require prefix?
3179	*u.pb++ = 0xcb;
3180
3181	#else /* RT_ARCH_X86 */
3182
3183	union
3184	{
3185	uint8_t *pb;
3186	uint16_t *pu16;
3187	uint32_t *pu32;
3188	} u, uFixJmpNotNested, uFixJmp, uFixCall, uNotNested;
3189	u.pb = &pPatch->auCode[0];
3190
3191	/* check the cookie */
3192	*u.pb++ = 0x81; // cmp esi, GLOBALCOOKIE
3193	*u.pb++ = 0xfe;
3194	*u.pu32++ = pDevExt->u32Cookie;
3195
3196	*u.pb++ = 0x74; // jz VBoxCall
3197	uFixJmp = u;
3198	*u.pb++ = 0;
3199
3200	/* jump (far) to the original handler / not-nested-stub. */
3201	*u.pb++ = 0xea; // jmp far NotNested
3202	uFixJmpNotNested = u;
3203	*u.pu32++ = 0;
3204	*u.pu16++ = 0;
3205
3206	/* save selector registers. */ // VBoxCall:
3207	*uFixJmp.pb = (uint8_t)(u.pb - uFixJmp.pb - 1);
3208	*u.pb++ = 0x0f; // push fs
3209	*u.pb++ = 0xa0;
3210
3211	*u.pb++ = 0x1e; // push ds
3212
3213	*u.pb++ = 0x06; // push es
3214
3215	/* call frame */
3216	*u.pb++ = 0x51; // push ecx
3217
3218	*u.pb++ = 0x52; // push edx
3219
3220	*u.pb++ = 0x50; // push eax
3221
3222	/* load ds, es and perhaps fs before call. */
3223	*u.pb++ = 0xb8; // mov eax, KernelDS
3224	*u.pu32++ = ASMGetDS();
3225
3226	*u.pb++ = 0x8e; // mov ds, eax
3227	*u.pb++ = 0xd8;
3228
3229	*u.pb++ = 0x8e; // mov es, eax
3230	*u.pb++ = 0xc0;
3231
3232	#ifdef RT_OS_WINDOWS
3233	*u.pb++ = 0xb8; // mov eax, KernelFS
3234	*u.pu32++ = ASMGetFS();
3235
3236	*u.pb++ = 0x8e; // mov fs, eax
3237	*u.pb++ = 0xe0;
3238	#endif
3239
3240	/* do the call. */
3241	*u.pb++ = 0xe8; // call _VMMR0Entry / StubVMMR0Entry
3242	uFixCall = u;
3243	pPatch->offVMMR0EntryFixup = (uint16_t)(u.pb - &pPatch->auCode[0]);
3244	*u.pu32++ = 0xfffffffb;
3245
3246	*u.pb++ = 0x83; // add esp, 0ch ; cdecl
3247	*u.pb++ = 0xc4;
3248	*u.pb++ = 0x0c;
3249
3250	/* restore selector registers. */
3251	*u.pb++ = 0x07; // pop es
3252	//
3253	*u.pb++ = 0x1f; // pop ds
3254
3255	*u.pb++ = 0x0f; // pop fs
3256	*u.pb++ = 0xa1;
3257
3258	uNotNested = u; // NotNested:
3259	*u.pb++ = 0xcf; // iretd
3260
3261	/* the stub VMMR0Entry. */ // StubVMMR0Entry:
3262	pPatch->offStub = (uint16_t)(u.pb - &pPatch->auCode[0]);
3263	*u.pb++ = 0x33; // xor eax, eax
3264	*u.pb++ = 0xc0;
3265
3266	*u.pb++ = 0x48; // dec eax
3267
3268	*u.pb++ = 0xc3; // ret
3269
3270	/* Fixup the VMMR0Entry call. */
3271	if (pDevExt->pvVMMR0)
3272	*uFixCall.pu32 = (uint32_t)pDevExt->pfnVMMR0EntryInt - (uint32_t)(uFixCall.pu32 + 1);
3273	else
3274	*uFixCall.pu32 = (uint32_t)&pPatch->auCode[pPatch->offStub] - (uint32_t)(uFixCall.pu32 + 1);
3275
3276	/* Fixup the forward / nested far jump. */
3277	if (!pPatch->SavedIdt.u5Type2)
3278	{
3279	*uFixJmpNotNested.pu32++ = (uint32_t)uNotNested.pb;
3280	*uFixJmpNotNested.pu16++ = ASMGetCS();
3281	}
3282	else
3283	{
3284	*uFixJmpNotNested.pu32++ = ((uint32_t)pPatch->SavedIdt.u16OffsetHigh << 16) \| pPatch->SavedIdt.u16OffsetLow;
3285	*uFixJmpNotNested.pu16++ = pPatch->SavedIdt.u16SegSel;
3286	}
3287	#endif /* RT_ARCH_X86 */
3288	Assert(u.pb <= &pPatch->auCode[sizeof(pPatch->auCode)]);
3289	#if 0
3290	/* dump the patch code */
3291	Log2(("patch code: %p\n", &pPatch->auCode[0]));
3292	for (uFixCall.pb = &pPatch->auCode[0]; uFixCall.pb < u.pb; uFixCall.pb++)
3293	Log2(("0x%02x,\n", *uFixCall.pb));
3294	#endif
3295	}
3296
3297	/*
3298	* Install the patch.
3299	*/
3300	supdrvIdtWrite(pPatch->pIdtEntry, &pPatch->ChangedIdt);
3301	AssertMsg(!memcmp((void *)pPatch->pIdtEntry, &pPatch->ChangedIdt, sizeof(pPatch->ChangedIdt)), ("The stupid change code didn't work!!!!!\n"));
3302
3303	/*
3304	* Link in the patch.
3305	*/
3306	pPatch->pNext = pDevExt->pIdtPatches;
3307	pDevExt->pIdtPatches = pPatch;
3308
3309	return pPatch;
3310	}
3311
3312
3313	/**
3314	* Removes the sessions IDT references.
3315	* This will uninstall our IDT patch if we left unreferenced.
3316	*
3317	* @returns VINF_SUCCESS.
3318	* @param pDevExt Device globals.
3319	* @param pSession Session data.
3320	*/
3321	static int supdrvIOCtl_IdtRemoveAll(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
3322	{
3323	PSUPDRVPATCHUSAGE pUsage;
3324	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3325	LogFlow(("supdrvIOCtl_IdtRemoveAll: pSession=%p\n", pSession));
3326
3327	/*
3328	* Take the spinlock.
3329	*/
3330	RTSpinlockAcquireNoInts(pDevExt->Spinlock, &SpinlockTmp);
3331
3332	/*
3333	* Walk usage list, removing patches as their usage count reaches zero.
3334	*/
3335	pUsage = pSession->pPatchUsage;
3336	while (pUsage)
3337	{
3338	if (pUsage->pPatch->cUsage <= pUsage->cUsage)
3339	supdrvIdtRemoveOne(pDevExt, pUsage->pPatch);
3340	else
3341	pUsage->pPatch->cUsage -= pUsage->cUsage;
3342
3343	/* next */
3344	pUsage = pUsage->pNext;
3345	}
3346
3347	/*
3348	* Empty the usage chain and we're done inside the spinlock.
3349	*/
3350	pUsage = pSession->pPatchUsage;
3351	pSession->pPatchUsage = NULL;
3352
3353	RTSpinlockReleaseNoInts(pDevExt->Spinlock, &SpinlockTmp);
3354
3355	/*
3356	* Free usage entries.
3357	*/
3358	while (pUsage)
3359	{
3360	void *pvToFree = pUsage;
3361	pUsage->cUsage = 0;
3362	pUsage->pPatch = NULL;
3363	pUsage = pUsage->pNext;
3364	RTMemFree(pvToFree);
3365	}
3366
3367	return VINF_SUCCESS;
3368	}
3369
3370
3371	/**
3372	* Remove one patch.
3373	*
3374	* Worker for supdrvIOCtl_IdtRemoveAll.
3375	*
3376	* @param pDevExt Device globals.
3377	* @param pPatch Patch entry to remove.
3378	* @remark Caller must own SUPDRVDEVEXT::Spinlock!
3379	*/
3380	static void supdrvIdtRemoveOne(PSUPDRVDEVEXT pDevExt, PSUPDRVPATCH pPatch)
3381	{
3382	LogFlow(("supdrvIdtRemoveOne: pPatch=%p\n", pPatch));
3383
3384	pPatch->cUsage = 0;
3385
3386	/*
3387	* If the IDT entry was changed it have to kick around for ever!
3388	* This will be attempted freed again, perhaps next time we'll succeed :-)
3389	*/
3390	if (memcmp((void *)pPatch->pIdtEntry, &pPatch->ChangedIdt, sizeof(pPatch->ChangedIdt)))
3391	{
3392	AssertMsgFailed(("The hijacked IDT entry has CHANGED!!!\n"));
3393	return;
3394	}
3395
3396	/*
3397	* Unlink it.
3398	*/
3399	if (pDevExt->pIdtPatches != pPatch)
3400	{
3401	PSUPDRVPATCH pPatchPrev = pDevExt->pIdtPatches;
3402	while (pPatchPrev)
3403	{
3404	if (pPatchPrev->pNext == pPatch)
3405	{
3406	pPatchPrev->pNext = pPatch->pNext;
3407	break;
3408	}
3409	pPatchPrev = pPatchPrev->pNext;
3410	}
3411	Assert(!pPatchPrev);
3412	}
3413	else
3414	pDevExt->pIdtPatches = pPatch->pNext;
3415	pPatch->pNext = NULL;
3416
3417
3418	/*
3419	* Verify and restore the IDT.
3420	*/
3421	AssertMsg(!memcmp((void *)pPatch->pIdtEntry, &pPatch->ChangedIdt, sizeof(pPatch->ChangedIdt)), ("The hijacked IDT entry has CHANGED!!!\n"));
3422	supdrvIdtWrite(pPatch->pIdtEntry, &pPatch->SavedIdt);
3423	AssertMsg(!memcmp((void *)pPatch->pIdtEntry, &pPatch->SavedIdt, sizeof(pPatch->SavedIdt)), ("The hijacked IDT entry has CHANGED!!!\n"));
3424
3425	/*
3426	* Put it in the free list.
3427	* (This free list stuff is to calm my paranoia.)
3428	*/
3429	pPatch->pvIdt = NULL;
3430	pPatch->pIdtEntry = NULL;
3431
3432	pPatch->pNext = pDevExt->pIdtPatchesFree;
3433	pDevExt->pIdtPatchesFree = pPatch;
3434	}
3435
3436
3437	/**
3438	* Write to an IDT entry.
3439	*
3440	* @param pvIdtEntry Where to write.
3441	* @param pNewIDTEntry What to write.
3442	*/
3443	static void supdrvIdtWrite(volatile void pvIdtEntry, const SUPDRVIDTE pNewIDTEntry)
3444	{
3445	RTR0UINTREG uCR0;
3446	RTR0UINTREG uFlags;
3447
3448	/*
3449	* On SMP machines (P4 hyperthreading included) we must preform a
3450	* 64-bit locked write when updating the IDT entry.
3451	*
3452	* The F00F bugfix for linux (and probably other OSes) causes
3453	* the IDT to be pointing to an readonly mapping. We get around that
3454	* by temporarily turning of WP. Since we're inside a spinlock at this
3455	* point, interrupts are disabled and there isn't any way the WP bit
3456	* flipping can cause any trouble.
3457	*/
3458
3459	/* Save & Clear interrupt flag; Save & clear WP. */
3460	uFlags = ASMGetFlags();
3461	ASMSetFlags(uFlags & ~(RTR0UINTREG)(1 << 9)); /X86_EFL_IF/
3462	Assert(!(ASMGetFlags() & (1 << 9)));
3463	uCR0 = ASMGetCR0();
3464	ASMSetCR0(uCR0 & ~(RTR0UINTREG)(1 << 16)); /X86_CR0_WP/
3465
3466	/* Update IDT Entry */
3467	#ifdef RT_ARCH_AMD64
3468	ASMAtomicXchgU128((volatile uint128_t )pvIdtEntry, (uint128_t *)(uintptr_t)pNewIDTEntry);
3469	#else
3470	ASMAtomicXchgU64((volatile uint64_t )pvIdtEntry, (uint64_t *)(uintptr_t)pNewIDTEntry);
3471	#endif
3472
3473	/* Restore CR0 & Flags */
3474	ASMSetCR0(uCR0);
3475	ASMSetFlags(uFlags);
3476	}
3477	#endif /* VBOX_WITH_IDT_PATCHING */
3478
3479
3480	/**
3481	* Opens an image. If it's the first time it's opened the call must upload
3482	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
3483	*
3484	* This is the 1st step of the loading.
3485	*
3486	* @returns IPRT status code.
3487	* @param pDevExt Device globals.
3488	* @param pSession Session data.
3489	* @param pReq The open request.
3490	*/
3491	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
3492	{
3493	PSUPDRVLDRIMAGE pImage;
3494	unsigned cb;
3495	void *pv;
3496	LogFlow(("supdrvIOCtl_LdrOpen: szName=%s cbImage=%d\n", pReq->u.In.szName, pReq->u.In.cbImage));
3497
3498	/*
3499	* Check if we got an instance of the image already.
3500	*/
3501	RTSemFastMutexRequest(pDevExt->mtxLdr);
3502	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
3503	{
3504	if (!strcmp(pImage->szName, pReq->u.In.szName))
3505	{
3506	pImage->cUsage++;
3507	pReq->u.Out.pvImageBase = pImage->pvImage;
3508	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
3509	supdrvLdrAddUsage(pSession, pImage);
3510	RTSemFastMutexRelease(pDevExt->mtxLdr);
3511	return VINF_SUCCESS;
3512	}
3513	}
3514	/* (not found - add it!) */
3515
3516	/*
3517	* Allocate memory.
3518	*/
3519	cb = pReq->u.In.cbImage + sizeof(SUPDRVLDRIMAGE) + 31;
3520	pv = RTMemExecAlloc(cb);
3521	if (!pv)
3522	{
3523	RTSemFastMutexRelease(pDevExt->mtxLdr);
3524	Log(("supdrvIOCtl_LdrOpen: RTMemExecAlloc(%u) failed\n", cb));
3525	return VERR_NO_MEMORY;
3526	}
3527
3528	/*
3529	* Setup and link in the LDR stuff.
3530	*/
3531	pImage = (PSUPDRVLDRIMAGE)pv;
3532	pImage->pvImage = RT_ALIGN_P(pImage + 1, 32);
3533	pImage->cbImage = pReq->u.In.cbImage;
3534	pImage->pfnModuleInit = NULL;
3535	pImage->pfnModuleTerm = NULL;
3536	pImage->uState = SUP_IOCTL_LDR_OPEN;
3537	pImage->cUsage = 1;
3538	strcpy(pImage->szName, pReq->u.In.szName);
3539
3540	pImage->pNext = pDevExt->pLdrImages;
3541	pDevExt->pLdrImages = pImage;
3542
3543	supdrvLdrAddUsage(pSession, pImage);
3544
3545	pReq->u.Out.pvImageBase = pImage->pvImage;
3546	pReq->u.Out.fNeedsLoading = true;
3547	RTSemFastMutexRelease(pDevExt->mtxLdr);
3548	return VINF_SUCCESS;
3549	}
3550
3551
3552	/**
3553	* Loads the image bits.
3554	*
3555	* This is the 2nd step of the loading.
3556	*
3557	* @returns IPRT status code.
3558	* @param pDevExt Device globals.
3559	* @param pSession Session data.
3560	* @param pReq The request.
3561	*/
3562	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
3563	{
3564	PSUPDRVLDRUSAGE pUsage;
3565	PSUPDRVLDRIMAGE pImage;
3566	int rc;
3567	LogFlow(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImage=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImage));
3568
3569	/*
3570	* Find the ldr image.
3571	*/
3572	RTSemFastMutexRequest(pDevExt->mtxLdr);
3573	pUsage = pSession->pLdrUsage;
3574	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3575	pUsage = pUsage->pNext;
3576	if (!pUsage)
3577	{
3578	RTSemFastMutexRelease(pDevExt->mtxLdr);
3579	Log(("SUP_IOCTL_LDR_LOAD: couldn't find image!\n"));
3580	return VERR_INVALID_HANDLE;
3581	}
3582	pImage = pUsage->pImage;
3583	if (pImage->cbImage != pReq->u.In.cbImage)
3584	{
3585	RTSemFastMutexRelease(pDevExt->mtxLdr);
3586	Log(("SUP_IOCTL_LDR_LOAD: image size mismatch!! %d(prep) != %d(load)\n", pImage->cbImage, pReq->u.In.cbImage));
3587	return VERR_INVALID_HANDLE;
3588	}
3589	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
3590	{
3591	unsigned uState = pImage->uState;
3592	RTSemFastMutexRelease(pDevExt->mtxLdr);
3593	if (uState != SUP_IOCTL_LDR_LOAD)
3594	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
3595	return SUPDRV_ERR_ALREADY_LOADED;
3596	}
3597	switch (pReq->u.In.eEPType)
3598	{
3599	case SUPLDRLOADEP_NOTHING:
3600	break;
3601	case SUPLDRLOADEP_VMMR0:
3602	if ( !pReq->u.In.EP.VMMR0.pvVMMR0
3603	\|\| !pReq->u.In.EP.VMMR0.pvVMMR0EntryInt
3604	\|\| !pReq->u.In.EP.VMMR0.pvVMMR0EntryFast
3605	\|\| !pReq->u.In.EP.VMMR0.pvVMMR0EntryEx)
3606	{
3607	RTSemFastMutexRelease(pDevExt->mtxLdr);
3608	Log(("NULL pointer: pvVMMR0=%p pvVMMR0EntryInt=%p pvVMMR0EntryFast=%p pvVMMR0EntryEx=%p!\n",
3609	pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
3610	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx));
3611	return VERR_INVALID_PARAMETER;
3612	}
3613	/** @todo validate pReq->u.In.EP.VMMR0.pvVMMR0 against pvImage! */
3614	if ( (uintptr_t)pReq->u.In.EP.VMMR0.pvVMMR0EntryInt - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage
3615	\|\| (uintptr_t)pReq->u.In.EP.VMMR0.pvVMMR0EntryFast - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage
3616	\|\| (uintptr_t)pReq->u.In.EP.VMMR0.pvVMMR0EntryEx - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage)
3617	{
3618	RTSemFastMutexRelease(pDevExt->mtxLdr);
3619	Log(("Out of range (%p LB %#x): pvVMMR0EntryInt=%p, pvVMMR0EntryFast=%p or pvVMMR0EntryEx=%p is NULL!\n",
3620	pImage->pvImage, pReq->u.In.cbImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
3621	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx));
3622	return VERR_INVALID_PARAMETER;
3623	}
3624	break;
3625	default:
3626	RTSemFastMutexRelease(pDevExt->mtxLdr);
3627	Log(("Invalid eEPType=%d\n", pReq->u.In.eEPType));
3628	return VERR_INVALID_PARAMETER;
3629	}
3630	if ( pReq->u.In.pfnModuleInit
3631	&& (uintptr_t)pReq->u.In.pfnModuleInit - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage)
3632	{
3633	RTSemFastMutexRelease(pDevExt->mtxLdr);
3634	Log(("SUP_IOCTL_LDR_LOAD: pfnModuleInit=%p is outside the image (%p %d bytes)\n",
3635	pReq->u.In.pfnModuleInit, pImage->pvImage, pReq->u.In.cbImage));
3636	return VERR_INVALID_PARAMETER;
3637	}
3638	if ( pReq->u.In.pfnModuleTerm
3639	&& (uintptr_t)pReq->u.In.pfnModuleTerm - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage)
3640	{
3641	RTSemFastMutexRelease(pDevExt->mtxLdr);
3642	Log(("SUP_IOCTL_LDR_LOAD: pfnModuleTerm=%p is outside the image (%p %d bytes)\n",
3643	pReq->u.In.pfnModuleTerm, pImage->pvImage, pReq->u.In.cbImage));
3644	return VERR_INVALID_PARAMETER;
3645	}
3646
3647	/*
3648	* Copy the memory.
3649	*/
3650	/* no need to do try/except as this is a buffered request. */
3651	memcpy(pImage->pvImage, &pReq->u.In.achImage[0], pImage->cbImage);
3652	pImage->uState = SUP_IOCTL_LDR_LOAD;
3653	pImage->pfnModuleInit = pReq->u.In.pfnModuleInit;
3654	pImage->pfnModuleTerm = pReq->u.In.pfnModuleTerm;
3655	pImage->offSymbols = pReq->u.In.offSymbols;
3656	pImage->cSymbols = pReq->u.In.cSymbols;
3657	pImage->offStrTab = pReq->u.In.offStrTab;
3658	pImage->cbStrTab = pReq->u.In.cbStrTab;
3659
3660	/*
3661	* Update any entry points.
3662	*/
3663	switch (pReq->u.In.eEPType)
3664	{
3665	default:
3666	case SUPLDRLOADEP_NOTHING:
3667	rc = VINF_SUCCESS;
3668	break;
3669	case SUPLDRLOADEP_VMMR0:
3670	rc = supdrvLdrSetR0EP(pDevExt, pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
3671	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
3672	break;
3673	}
3674
3675	/*
3676	* On success call the module initialization.
3677	*/
3678	LogFlow(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
3679	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
3680	{
3681	Log(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
3682	rc = pImage->pfnModuleInit();
3683	if (rc && pDevExt->pvVMMR0 == pImage->pvImage)
3684	supdrvLdrUnsetR0EP(pDevExt);
3685	}
3686
3687	if (rc)
3688	pImage->uState = SUP_IOCTL_LDR_OPEN;
3689
3690	RTSemFastMutexRelease(pDevExt->mtxLdr);
3691	return rc;
3692	}
3693
3694
3695	/**
3696	* Frees a previously loaded (prep'ed) image.
3697	*
3698	* @returns IPRT status code.
3699	* @param pDevExt Device globals.
3700	* @param pSession Session data.
3701	* @param pReq The request.
3702	*/
3703	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
3704	{
3705	int rc;
3706	PSUPDRVLDRUSAGE pUsagePrev;
3707	PSUPDRVLDRUSAGE pUsage;
3708	PSUPDRVLDRIMAGE pImage;
3709	LogFlow(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
3710
3711	/*
3712	* Find the ldr image.
3713	*/
3714	RTSemFastMutexRequest(pDevExt->mtxLdr);
3715	pUsagePrev = NULL;
3716	pUsage = pSession->pLdrUsage;
3717	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3718	{
3719	pUsagePrev = pUsage;
3720	pUsage = pUsage->pNext;
3721	}
3722	if (!pUsage)
3723	{
3724	RTSemFastMutexRelease(pDevExt->mtxLdr);
3725	Log(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
3726	return VERR_INVALID_HANDLE;
3727	}
3728
3729	/*
3730	* Check if we can remove anything.
3731	*/
3732	rc = VINF_SUCCESS;
3733	pImage = pUsage->pImage;
3734	if (pImage->cUsage <= 1 \|\| pUsage->cUsage <= 1)
3735	{
3736	/*
3737	* Check if there are any objects with destructors in the image, if
3738	* so leave it for the session cleanup routine so we get a chance to
3739	* clean things up in the right order and not leave them all dangling.
3740	*/
3741	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3742	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
3743	if (pImage->cUsage <= 1)
3744	{
3745	PSUPDRVOBJ pObj;
3746	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
3747	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImage))
3748	{
3749	rc = VERR_SHARING_VIOLATION; /** @todo VERR_DANGLING_OBJECTS */
3750	break;
3751	}
3752	}
3753	else
3754	{
3755	PSUPDRVUSAGE pGenUsage;
3756	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
3757	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImage))
3758	{
3759	rc = VERR_SHARING_VIOLATION; /** @todo VERR_DANGLING_OBJECTS */
3760	break;
3761	}
3762	}
3763	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
3764	if (rc == VINF_SUCCESS)
3765	{
3766	/* unlink it */
3767	if (pUsagePrev)
3768	pUsagePrev->pNext = pUsage->pNext;
3769	else
3770	pSession->pLdrUsage = pUsage->pNext;
3771
3772	/* free it */
3773	pUsage->pImage = NULL;
3774	pUsage->pNext = NULL;
3775	RTMemFree(pUsage);
3776
3777	/*
3778	* Derefrence the image.
3779	*/
3780	if (pImage->cUsage <= 1)
3781	supdrvLdrFree(pDevExt, pImage);
3782	else
3783	pImage->cUsage--;
3784	}
3785	else
3786	Log(("supdrvIOCtl_LdrFree: Dangling objects in %p/%s!\n", pImage->pvImage, pImage->szName));
3787	}
3788	else
3789	{
3790	/*
3791	* Dereference both image and usage.
3792	*/
3793	pImage->cUsage--;
3794	pUsage->cUsage--;
3795	}
3796
3797	RTSemFastMutexRelease(pDevExt->mtxLdr);
3798	return VINF_SUCCESS;
3799	}
3800
3801
3802	/**
3803	* Gets the address of a symbol in an open image.
3804	*
3805	* @returns 0 on success.
3806	* @returns SUPDRV_ERR_* on failure.
3807	* @param pDevExt Device globals.
3808	* @param pSession Session data.
3809	* @param pReq The request buffer.
3810	*/
3811	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
3812	{
3813	PSUPDRVLDRIMAGE pImage;
3814	PSUPDRVLDRUSAGE pUsage;
3815	uint32_t i;
3816	PSUPLDRSYM paSyms;
3817	const char *pchStrings;
3818	const size_t cbSymbol = strlen(pReq->u.In.szSymbol) + 1;
3819	void *pvSymbol = NULL;
3820	int rc = VERR_GENERAL_FAILURE;
3821	Log3(("supdrvIOCtl_LdrGetSymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
3822
3823	/*
3824	* Find the ldr image.
3825	*/
3826	RTSemFastMutexRequest(pDevExt->mtxLdr);
3827	pUsage = pSession->pLdrUsage;
3828	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3829	pUsage = pUsage->pNext;
3830	if (!pUsage)
3831	{
3832	RTSemFastMutexRelease(pDevExt->mtxLdr);
3833	Log(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
3834	return VERR_INVALID_HANDLE;
3835	}
3836	pImage = pUsage->pImage;
3837	if (pImage->uState != SUP_IOCTL_LDR_LOAD)
3838	{
3839	unsigned uState = pImage->uState;
3840	RTSemFastMutexRelease(pDevExt->mtxLdr);
3841	Log(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", uState, uState)); NOREF(uState);
3842	return VERR_ALREADY_LOADED;
3843	}
3844
3845	/*
3846	* Search the symbol strings.
3847	*/
3848	pchStrings = (const char )((uint8_t )pImage->pvImage + pImage->offStrTab);
3849	paSyms = (PSUPLDRSYM)((uint8_t *)pImage->pvImage + pImage->offSymbols);
3850	for (i = 0; i < pImage->cSymbols; i++)
3851	{
3852	if ( paSyms[i].offSymbol < pImage->cbImage /* paranoia */
3853	&& paSyms[i].offName + cbSymbol <= pImage->cbStrTab
3854	&& !memcmp(pchStrings + paSyms[i].offName, pReq->u.In.szSymbol, cbSymbol))
3855	{
3856	pvSymbol = (uint8_t *)pImage->pvImage + paSyms[i].offSymbol;
3857	rc = VINF_SUCCESS;
3858	break;
3859	}
3860	}
3861	RTSemFastMutexRelease(pDevExt->mtxLdr);
3862	pReq->u.Out.pvSymbol = pvSymbol;
3863	return rc;
3864	}
3865
3866
3867	/**
3868	* Gets the address of a symbol in an open image or the support driver.
3869	*
3870	* @returns VINF_SUCCESS on success.
3871	* @returns
3872	* @param pDevExt Device globals.
3873	* @param pSession Session data.
3874	* @param pReq The request buffer.
3875	*/
3876	static int supdrvIDC_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPDRVIDCREQGETSYM pReq)
3877	{
3878	int rc = VINF_SUCCESS;
3879	const char *pszSymbol = pReq->u.In.pszSymbol;
3880	const char *pszModule = pReq->u.In.pszModule;
3881	size_t cbSymbol;
3882	char const *pszEnd;
3883	uint32_t i;
3884
3885	/*
3886	* Input validation.
3887	*/
3888	AssertPtrReturn(pszSymbol, VERR_INVALID_POINTER);
3889	pszEnd = (char *)memchr(pszSymbol, '\0', 512);
3890	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3891	cbSymbol = pszEnd - pszSymbol + 1;
3892
3893	if (pszModule)
3894	{
3895	AssertPtrReturn(pszModule, VERR_INVALID_POINTER);
3896	pszEnd = (char *)memchr(pszModule, '\0', 64);
3897	AssertReturn(pszEnd, VERR_INVALID_PARAMETER);
3898	}
3899	Log3(("supdrvIDC_LdrGetSymbol: pszModule=%p:{%s} pszSymbol=%p:{%s}\n", pszModule, pszModule, pszSymbol, pszSymbol));
3900
3901
3902	if ( !pszModule
3903	\|\| !strcmp(pszModule, "SupDrv"))
3904	{
3905	/*
3906	* Search the support driver export table.
3907	*/
3908	for (i = 0; i < RT_ELEMENTS(g_aFunctions); i++)
3909	if (!strcmp(g_aFunctions[i].szName, pszSymbol))
3910	{
3911	pReq->u.Out.pfnSymbol = g_aFunctions[i].pfn;
3912	break;
3913	}
3914	}
3915	else
3916	{
3917	/*
3918	* Find the loader image.
3919	*/
3920	PSUPDRVLDRIMAGE pImage;
3921
3922	RTSemFastMutexRequest(pDevExt->mtxLdr);
3923
3924	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
3925	if (!strcmp(pImage->szName, pszModule))
3926	break;
3927	if (pImage && pImage->uState == SUP_IOCTL_LDR_LOAD)
3928	{
3929	/*
3930	* Search the symbol strings.
3931	*/
3932	const char pchStrings = (const char )((uint8_t *)pImage->pvImage + pImage->offStrTab);
3933	PCSUPLDRSYM paSyms = (PCSUPLDRSYM)((uint8_t *)pImage->pvImage + pImage->offSymbols);
3934	for (i = 0; i < pImage->cSymbols; i++)
3935	{
3936	if ( paSyms[i].offSymbol < pImage->cbImage /* paranoia */
3937	&& paSyms[i].offName + cbSymbol <= pImage->cbStrTab
3938	&& !memcmp(pchStrings + paSyms[i].offName, pszSymbol, cbSymbol))
3939	{
3940	/*
3941	* Found it! Calc the symbol address and add a reference to the module.
3942	*/
3943	pReq->u.Out.pfnSymbol = (PFNRT)((uint8_t *)pImage->pvImage + paSyms[i].offSymbol);
3944	rc = supdrvLdrAddUsage(pSession, pImage);
3945	break;
3946	}
3947	}
3948	}
3949	else
3950	rc = pImage ? VERR_WRONG_ORDER : VERR_MODULE_NOT_FOUND;
3951
3952	RTSemFastMutexRelease(pDevExt->mtxLdr);
3953	}
3954	return rc;
3955	}
3956
3957
3958	/**
3959	* Updates the IDT patches to point to the specified VMM R0 entry
3960	* point (i.e. VMMR0Enter()).
3961	*
3962	* @returns IPRT status code.
3963	* @param pDevExt Device globals.
3964	* @param pSession Session data.
3965	* @param pVMMR0 VMMR0 image handle.
3966	* @param pvVMMR0EntryInt VMMR0EntryInt address.
3967	* @param pvVMMR0EntryFast VMMR0EntryFast address.
3968	* @param pvVMMR0EntryEx VMMR0EntryEx address.
3969	* @remark Caller must own the loader mutex.
3970	*/
3971	static int supdrvLdrSetR0EP(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx)
3972	{
3973	int rc = VINF_SUCCESS;
3974	LogFlow(("supdrvLdrSetR0EP pvVMMR0=%p pvVMMR0EntryInt=%p\n", pvVMMR0, pvVMMR0EntryInt));
3975
3976
3977	/*
3978	* Check if not yet set.
3979	*/
3980	if (!pDevExt->pvVMMR0)
3981	{
3982	#ifdef VBOX_WITH_IDT_PATCHING
3983	PSUPDRVPATCH pPatch;
3984	#endif
3985
3986	/*
3987	* Set it and update IDT patch code.
3988	*/
3989	pDevExt->pvVMMR0 = pvVMMR0;
3990	pDevExt->pfnVMMR0EntryInt = pvVMMR0EntryInt;
3991	pDevExt->pfnVMMR0EntryFast = pvVMMR0EntryFast;
3992	pDevExt->pfnVMMR0EntryEx = pvVMMR0EntryEx;
3993	#ifdef VBOX_WITH_IDT_PATCHING
3994	for (pPatch = pDevExt->pIdtPatches; pPatch; pPatch = pPatch->pNext)
3995	{
3996	# ifdef RT_ARCH_AMD64
3997	ASMAtomicXchgU64((volatile uint64_t *)&pPatch->auCode[pPatch->offVMMR0EntryFixup], (uint64_t)pvVMMR0);
3998	# else /* RT_ARCH_X86 */
3999	ASMAtomicXchgU32((volatile uint32_t *)&pPatch->auCode[pPatch->offVMMR0EntryFixup],
4000	(uint32_t)pvVMMR0 - (uint32_t)&pPatch->auCode[pPatch->offVMMR0EntryFixup + 4]);
4001	# endif
4002	}
4003	#endif /* VBOX_WITH_IDT_PATCHING */
4004	}
4005	else
4006	{
4007	/*
4008	* Return failure or success depending on whether the values match or not.
4009	*/
4010	if ( pDevExt->pvVMMR0 != pvVMMR0
4011	\|\| (void *)pDevExt->pfnVMMR0EntryInt != pvVMMR0EntryInt
4012	\|\| (void *)pDevExt->pfnVMMR0EntryFast != pvVMMR0EntryFast
4013	\|\| (void *)pDevExt->pfnVMMR0EntryEx != pvVMMR0EntryEx)
4014	{
4015	AssertMsgFailed(("SUP_IOCTL_LDR_SETR0EP: Already set pointing to a different module!\n"));
4016	rc = VERR_INVALID_PARAMETER;
4017	}
4018	}
4019	return rc;
4020	}
4021
4022
4023	/**
4024	* Unsets the R0 entry point installed by supdrvLdrSetR0EP.
4025	*
4026	* @param pDevExt Device globals.
4027	*/
4028	static void supdrvLdrUnsetR0EP(PSUPDRVDEVEXT pDevExt)
4029	{
4030	#ifdef VBOX_WITH_IDT_PATCHING
4031	PSUPDRVPATCH pPatch;
4032	#endif
4033
4034	pDevExt->pvVMMR0 = NULL;
4035	pDevExt->pfnVMMR0EntryInt = NULL;
4036	pDevExt->pfnVMMR0EntryFast = NULL;
4037	pDevExt->pfnVMMR0EntryEx = NULL;
4038
4039	#ifdef VBOX_WITH_IDT_PATCHING
4040	for (pPatch = pDevExt->pIdtPatches; pPatch; pPatch = pPatch->pNext)
4041	{
4042	# ifdef RT_ARCH_AMD64
4043	ASMAtomicXchgU64((volatile uint64_t *)&pPatch->auCode[pPatch->offVMMR0EntryFixup],
4044	(uint64_t)&pPatch->auCode[pPatch->offStub]);
4045	# else /* RT_ARCH_X86 */
4046	ASMAtomicXchgU32((volatile uint32_t *)&pPatch->auCode[pPatch->offVMMR0EntryFixup],
4047	(uint32_t)&pPatch->auCode[pPatch->offStub] - (uint32_t)&pPatch->auCode[pPatch->offVMMR0EntryFixup + 4]);
4048	# endif
4049	}
4050	#endif /* VBOX_WITH_IDT_PATCHING */
4051	}
4052
4053
4054	/**
4055	* Adds a usage reference in the specified session of an image.
4056	*
4057	* Called while owning the loader semaphore.
4058	*
4059	* @returns VINF_SUCCESS on success and VERR_NO_MEMORY on failure.
4060	* @param pSession Session in question.
4061	* @param pImage Image which the session is using.
4062	*/
4063	static int supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage)
4064	{
4065	PSUPDRVLDRUSAGE pUsage;
4066	LogFlow(("supdrvLdrAddUsage: pImage=%p\n", pImage));
4067
4068	/*
4069	* Referenced it already?
4070	*/
4071	pUsage = pSession->pLdrUsage;
4072	while (pUsage)
4073	{
4074	if (pUsage->pImage == pImage)
4075	{
4076	pUsage->cUsage++;
4077	return VINF_SUCCESS;
4078	}
4079	pUsage = pUsage->pNext;
4080	}
4081
4082	/*
4083	* Allocate new usage record.
4084	*/
4085	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
4086	AssertReturn(pUsage, VERR_NO_MEMORY);
4087	pUsage->cUsage = 1;
4088	pUsage->pImage = pImage;
4089	pUsage->pNext = pSession->pLdrUsage;
4090	pSession->pLdrUsage = pUsage;
4091	return VINF_SUCCESS;
4092	}
4093
4094
4095	/**
4096	* Frees a load image.
4097	*
4098	* @param pDevExt Pointer to device extension.
4099	* @param pImage Pointer to the image we're gonna free.
4100	* This image must exit!
4101	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
4102	*/
4103	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
4104	{
4105	PSUPDRVLDRIMAGE pImagePrev;
4106	LogFlow(("supdrvLdrFree: pImage=%p\n", pImage));
4107
4108	/* find it - arg. should've used doubly linked list. */
4109	Assert(pDevExt->pLdrImages);
4110	pImagePrev = NULL;
4111	if (pDevExt->pLdrImages != pImage)
4112	{
4113	pImagePrev = pDevExt->pLdrImages;
4114	while (pImagePrev->pNext != pImage)
4115	pImagePrev = pImagePrev->pNext;
4116	Assert(pImagePrev->pNext == pImage);
4117	}
4118
4119	/* unlink */
4120	if (pImagePrev)
4121	pImagePrev->pNext = pImage->pNext;
4122	else
4123	pDevExt->pLdrImages = pImage->pNext;
4124
4125	/* check if this is VMMR0.r0 and fix the Idt patches if it is. */
4126	if (pDevExt->pvVMMR0 == pImage->pvImage)
4127	supdrvLdrUnsetR0EP(pDevExt);
4128
4129	/* check for objects with destructors in this image. (Shouldn't happen.) */
4130	if (pDevExt->pObjs)
4131	{
4132	unsigned cObjs = 0;
4133	PSUPDRVOBJ pObj;
4134	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
4135	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
4136	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
4137	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImage))
4138	{
4139	pObj->pfnDestructor = NULL;
4140	cObjs++;
4141	}
4142	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
4143	if (cObjs)
4144	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
4145	}
4146
4147	/* call termination function if fully loaded. */
4148	if ( pImage->pfnModuleTerm
4149	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
4150	{
4151	LogFlow(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
4152	pImage->pfnModuleTerm();
4153	}
4154
4155	/* free the image */
4156	pImage->cUsage = 0;
4157	pImage->pNext = 0;
4158	pImage->uState = SUP_IOCTL_LDR_FREE;
4159	RTMemExecFree(pImage);
4160	}
4161
4162
4163	/**
4164	* Gets the current paging mode of the CPU and stores in in pOut.
4165	*/
4166	static SUPPAGINGMODE supdrvIOCtl_GetPagingMode(void)
4167	{
4168	SUPPAGINGMODE enmMode;
4169
4170	RTR0UINTREG cr0 = ASMGetCR0();
4171	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
4172	enmMode = SUPPAGINGMODE_INVALID;
4173	else
4174	{
4175	RTR0UINTREG cr4 = ASMGetCR4();
4176	uint32_t fNXEPlusLMA = 0;
4177	if (cr4 & X86_CR4_PAE)
4178	{
4179	uint32_t fAmdFeatures = ASMCpuId_EDX(0x80000001);
4180	if (fAmdFeatures & (X86_CPUID_AMD_FEATURE_EDX_NX \| X86_CPUID_AMD_FEATURE_EDX_LONG_MODE))
4181	{
4182	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
4183	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
4184	fNXEPlusLMA \|= RT_BIT(0);
4185	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
4186	fNXEPlusLMA \|= RT_BIT(1);
4187	}
4188	}
4189
4190	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
4191	{
4192	case 0:
4193	enmMode = SUPPAGINGMODE_32_BIT;
4194	break;
4195
4196	case X86_CR4_PGE:
4197	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
4198	break;
4199
4200	case X86_CR4_PAE:
4201	enmMode = SUPPAGINGMODE_PAE;
4202	break;
4203
4204	case X86_CR4_PAE \| RT_BIT(0):
4205	enmMode = SUPPAGINGMODE_PAE_NX;
4206	break;
4207
4208	case X86_CR4_PAE \| X86_CR4_PGE:
4209	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
4210	break;
4211
4212	case X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
4213	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
4214	break;
4215
4216	case RT_BIT(1) \| X86_CR4_PAE:
4217	enmMode = SUPPAGINGMODE_AMD64;
4218	break;
4219
4220	case RT_BIT(1) \| X86_CR4_PAE \| RT_BIT(0):
4221	enmMode = SUPPAGINGMODE_AMD64_NX;
4222	break;
4223
4224	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
4225	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
4226	break;
4227
4228	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
4229	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
4230	break;
4231
4232	default:
4233	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
4234	enmMode = SUPPAGINGMODE_INVALID;
4235	break;
4236	}
4237	}
4238	return enmMode;
4239	}
4240
4241
4242	/**
4243	* Creates the GIP.
4244	*
4245	* @returns negative errno.
4246	* @param pDevExt Instance data. GIP stuff may be updated.
4247	*/
4248	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
4249	{
4250	PSUPGLOBALINFOPAGE pGip;
4251	RTHCPHYS HCPhysGip;
4252	uint32_t u32SystemResolution;
4253	uint32_t u32Interval;
4254	int rc;
4255
4256	LogFlow(("supdrvGipCreate:\n"));
4257
4258	/* assert order */
4259	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
4260	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
4261	Assert(!pDevExt->pGipTimer);
4262
4263	/*
4264	* Allocate a suitable page with a default kernel mapping.
4265	*/
4266	rc = RTR0MemObjAllocLow(&pDevExt->GipMemObj, PAGE_SIZE, false);
4267	if (RT_FAILURE(rc))
4268	{
4269	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
4270	return rc;
4271	}
4272	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
4273	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
4274
4275	#if 0 /** @todo Disabled this as we didn't used to do it before and causes unnecessary stress on laptops.
4276	* It only applies to Windows and should probably revisited later, if possible made part of the
4277	* timer code (return min granularity in RTTimerGetSystemGranularity and set it in RTTimerStart). */
4278	/*
4279	* Try bump up the system timer resolution.
4280	* The more interrupts the better...
4281	*/
4282	if ( RT_SUCCESS(RTTimerRequestSystemGranularity( 488281 /* 2048 HZ */, &u32SystemResolution))
4283	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 500000 /* 2000 HZ */, &u32SystemResolution))
4284	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
4285	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
4286	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 1953125 /* 512 HZ */, &u32SystemResolution))
4287	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 2000000 /* 500 HZ */, &u32SystemResolution))
4288	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 3906250 /* 256 HZ */, &u32SystemResolution))
4289	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 4000000 /* 250 HZ */, &u32SystemResolution))
4290	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 7812500 /* 128 HZ */, &u32SystemResolution))
4291	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity(10000000 /* 100 HZ */, &u32SystemResolution))
4292	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity(15625000 /* 64 HZ */, &u32SystemResolution))
4293	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity(31250000 /* 32 HZ */, &u32SystemResolution))
4294	)
4295	{
4296	Assert(RTTimerGetSystemGranularity() <= u32SystemResolution);
4297	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
4298	}
4299	#endif
4300
4301	/*
4302	* Find a reasonable update interval and initialize the structure.
4303	*/
4304	u32Interval = u32SystemResolution = RTTimerGetSystemGranularity();
4305	while (u32Interval < 10000000 /* 10 ms */)
4306	u32Interval += u32SystemResolution;
4307
4308	supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), 1000000000 / u32Interval /=Hz/);
4309
4310	/*
4311	* Create the timer.
4312	* If CPU_ALL isn't supported we'll have to fall back to synchronous mode.
4313	*/
4314	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
4315	{
4316	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, RTTIMER_FLAGS_CPU_ALL, supdrvGipAsyncTimer, pDevExt);
4317	if (rc == VERR_NOT_SUPPORTED)
4318	{
4319	OSDBGPRINT(("supdrvGipCreate: omni timer not supported, falling back to synchronous mode\n"));
4320	pGip->u32Mode = SUPGIPMODE_SYNC_TSC;
4321	}
4322	}
4323	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
4324	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0, supdrvGipSyncTimer, pDevExt);
4325	if (RT_SUCCESS(rc))
4326	{
4327	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
4328	rc = RTMpNotificationRegister(supdrvGipMpEvent, pDevExt);
4329	if (RT_SUCCESS(rc))
4330	{
4331	/*
4332	* We're good.
4333	*/
4334	dprintf(("supdrvGipCreate: %ld ns interval.\n", (long)u32Interval));
4335	return VINF_SUCCESS;
4336	}
4337
4338	OSDBGPRINT(("supdrvGipCreate: failed register MP event notfication. rc=%d\n", rc));
4339	}
4340	else
4341	{
4342	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %ld ns interval. rc=%d\n", (long)u32Interval, rc));
4343	Assert(!pDevExt->pGipTimer);
4344	}
4345	supdrvGipDestroy(pDevExt);
4346	return rc;
4347	}
4348
4349
4350	/**
4351	* Terminates the GIP.
4352	*
4353	* @param pDevExt Instance data. GIP stuff may be updated.
4354	*/
4355	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
4356	{
4357	int rc;
4358	#ifdef DEBUG_DARWIN_GIP
4359	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
4360	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
4361	pDevExt->pGipTimer, pDevExt->GipMemObj));
4362	#endif
4363
4364	/*
4365	* Invalid the GIP data.
4366	*/
4367	if (pDevExt->pGip)
4368	{
4369	supdrvGipTerm(pDevExt->pGip);
4370	pDevExt->pGip = NULL;
4371	}
4372
4373	/*
4374	* Destroy the timer and free the GIP memory object.
4375	*/
4376	if (pDevExt->pGipTimer)
4377	{
4378	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
4379	pDevExt->pGipTimer = NULL;
4380	}
4381
4382	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
4383	{
4384	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
4385	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
4386	}
4387
4388	/*
4389	* Finally, release the system timer resolution request if one succeeded.
4390	*/
4391	if (pDevExt->u32SystemTimerGranularityGrant)
4392	{
4393	rc = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
4394	pDevExt->u32SystemTimerGranularityGrant = 0;
4395	}
4396	}
4397
4398
4399	/**
4400	* Timer callback function sync GIP mode.
4401	* @param pTimer The timer.
4402	* @param pvUser The device extension.
4403	*/
4404	static DECLCALLBACK(void) supdrvGipSyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
4405	{
4406	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4407	supdrvGipUpdate(pDevExt->pGip, RTTimeSystemNanoTS());
4408	}
4409
4410
4411	/**
4412	* Timer callback function for async GIP mode.
4413	* @param pTimer The timer.
4414	* @param pvUser The device extension.
4415	*/
4416	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
4417	{
4418	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4419	RTCPUID idCpu = RTMpCpuId();
4420	uint64_t NanoTS = RTTimeSystemNanoTS();
4421
4422	/** @todo reset the transaction number and whatnot when iTick == 1. */
4423	if (pDevExt->idGipMaster == idCpu)
4424	supdrvGipUpdate(pDevExt->pGip, NanoTS);
4425	else
4426	supdrvGipUpdatePerCpu(pDevExt->pGip, NanoTS, ASMGetApicId());
4427	}
4428
4429
4430	/**
4431	* Multiprocessor event notification callback.
4432	*
4433	* This is used to make sue that the GIP master gets passed on to
4434	* another CPU.
4435	*
4436	* @param enmEvent The event.
4437	* @param idCpu The cpu it applies to.
4438	* @param pvUser Pointer to the device extension.
4439	*/
4440	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser)
4441	{
4442	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4443	if (enmEvent == RTMPEVENT_OFFLINE)
4444	{
4445	RTCPUID idGipMaster;
4446	ASMAtomicReadSize(&pDevExt->idGipMaster, &idGipMaster);
4447	if (idGipMaster == idCpu)
4448	{
4449	/*
4450	* Find a new GIP master.
4451	*/
4452	bool fIgnored;
4453	unsigned i;
4454	RTCPUID idNewGipMaster = NIL_RTCPUID;
4455	RTCPUSET OnlineCpus;
4456	RTMpGetOnlineSet(&OnlineCpus);
4457
4458	for (i = 0; i < RTCPUSET_MAX_CPUS; i++)
4459	{
4460	RTCPUID idCurCpu = RTMpCpuIdFromSetIndex(i);
4461	if ( RTCpuSetIsMember(&OnlineCpus, idCurCpu)
4462	&& idCurCpu != idGipMaster)
4463	{
4464	idNewGipMaster = idCurCpu;
4465	break;
4466	}
4467	}
4468
4469	dprintf(("supdrvGipMpEvent: Gip master %#lx -> %#lx\n", (long)idGipMaster, (long)idNewGipMaster));
4470	ASMAtomicCmpXchgSize(&pDevExt->idGipMaster, idNewGipMaster, idGipMaster, fIgnored);
4471	NOREF(fIgnored);
4472	}
4473	}
4474	}
4475
4476
4477	/**
4478	* Initializes the GIP data.
4479	*
4480	* @returns IPRT status code.
4481	* @param pDevExt Pointer to the device instance data.
4482	* @param pGip Pointer to the read-write kernel mapping of the GIP.
4483	* @param HCPhys The physical address of the GIP.
4484	* @param u64NanoTS The current nanosecond timestamp.
4485	* @param uUpdateHz The update freqence.
4486	*/
4487	int VBOXCALL supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys, uint64_t u64NanoTS, unsigned uUpdateHz)
4488	{
4489	unsigned i;
4490	#ifdef DEBUG_DARWIN_GIP
4491	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz));
4492	#else
4493	LogFlow(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz));
4494	#endif
4495
4496	/*
4497	* Initialize the structure.
4498	*/
4499	memset(pGip, 0, PAGE_SIZE);
4500	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
4501	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
4502	pGip->u32Mode = supdrvGipDeterminTscMode(pDevExt);
4503	pGip->u32UpdateHz = uUpdateHz;
4504	pGip->u32UpdateIntervalNS = 1000000000 / uUpdateHz;
4505	pGip->u64NanoTSLastUpdateHz = u64NanoTS;
4506
4507	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
4508	{
4509	pGip->aCPUs[i].u32TransactionId = 2;
4510	pGip->aCPUs[i].u64NanoTS = u64NanoTS;
4511	pGip->aCPUs[i].u64TSC = ASMReadTSC();
4512
4513	/*
4514	* We don't know the following values until we've executed updates.
4515	* So, we'll just insert very high values.
4516	*/
4517	pGip->aCPUs[i].u64CpuHz = _4G + 1;
4518	pGip->aCPUs[i].u32UpdateIntervalTSC = _2G / 4;
4519	pGip->aCPUs[i].au32TSCHistory[0] = _2G / 4;
4520	pGip->aCPUs[i].au32TSCHistory[1] = _2G / 4;
4521	pGip->aCPUs[i].au32TSCHistory[2] = _2G / 4;
4522	pGip->aCPUs[i].au32TSCHistory[3] = _2G / 4;
4523	pGip->aCPUs[i].au32TSCHistory[4] = _2G / 4;
4524	pGip->aCPUs[i].au32TSCHistory[5] = _2G / 4;
4525	pGip->aCPUs[i].au32TSCHistory[6] = _2G / 4;
4526	pGip->aCPUs[i].au32TSCHistory[7] = _2G / 4;
4527	}
4528
4529	/*
4530	* Link it to the device extension.
4531	*/
4532	pDevExt->pGip = pGip;
4533	pDevExt->HCPhysGip = HCPhys;
4534	pDevExt->cGipUsers = 0;
4535
4536	return VINF_SUCCESS;
4537	}
4538
4539
4540	/**
4541	* Callback used by supdrvDetermineAsyncTSC to read the TSC on a CPU.
4542	*
4543	* @param idCpu Ignored.
4544	* @param pvUser1 Where to put the TSC.
4545	* @param pvUser2 Ignored.
4546	*/
4547	static DECLCALLBACK(void) supdrvDetermineAsyncTscWorker(RTCPUID idCpu, void pvUser1, void pvUser2)
4548	{
4549	#if 1
4550	ASMAtomicWriteU64((uint64_t volatile *)pvUser1, ASMReadTSC());
4551	#else
4552	(uint64_t )pvUser1 = ASMReadTSC();
4553	#endif
4554	}
4555
4556
4557	/**
4558	* Determine if Async GIP mode is required because of TSC drift.
4559	*
4560	* When using the default/normal timer code it is essential that the time stamp counter
4561	* (TSC) runs never backwards, that is, a read operation to the counter should return
4562	* a bigger value than any previous read operation. This is guaranteed by the latest
4563	* AMD CPUs and by newer Intel CPUs which never enter the C2 state (P4). In any other
4564	* case we have to choose the asynchronous timer mode.
4565	*
4566	* @param poffMin Pointer to the determined difference between different cores.
4567	* @return false if the time stamp counters appear to be synchron, true otherwise.
4568	*/
4569	bool VBOXCALL supdrvDetermineAsyncTsc(uint64_t *poffMin)
4570	{
4571	/*
4572	* Just iterate all the cpus 8 times and make sure that the TSC is
4573	* ever increasing. We don't bother taking TSC rollover into account.
4574	*/
4575	RTCPUSET CpuSet;
4576	int iLastCpu = RTCpuLastIndex(RTMpGetSet(&CpuSet));
4577	int iCpu;
4578	int cLoops = 8;
4579	bool fAsync = false;
4580	int rc = VINF_SUCCESS;
4581	uint64_t offMax = 0;
4582	uint64_t offMin = ~(uint64_t)0;
4583	uint64_t PrevTsc = ASMReadTSC();
4584
4585	while (cLoops-- > 0)
4586	{
4587	for (iCpu = 0; iCpu <= iLastCpu; iCpu++)
4588	{
4589	uint64_t CurTsc;
4590	rc = RTMpOnSpecific(RTMpCpuIdFromSetIndex(iCpu), supdrvDetermineAsyncTscWorker, &CurTsc, NULL);
4591	if (RT_SUCCESS(rc))
4592	{
4593	if (CurTsc <= PrevTsc)
4594	{
4595	fAsync = true;
4596	offMin = offMax = PrevTsc - CurTsc;
4597	dprintf(("supdrvDetermineAsyncTsc: iCpu=%d cLoops=%d CurTsc=%llx PrevTsc=%llx\n",
4598	iCpu, cLoops, CurTsc, PrevTsc));
4599	break;
4600	}
4601
4602	/* Gather statistics (except the first time). */
4603	if (iCpu != 0 \|\| cLoops != 7)
4604	{
4605	uint64_t off = CurTsc - PrevTsc;
4606	if (off < offMin)
4607	offMin = off;
4608	if (off > offMax)
4609	offMax = off;
4610	dprintf2(("%d/%d: off=%llx\n", cLoops, iCpu, off));
4611	}
4612
4613	/* Next */
4614	PrevTsc = CurTsc;
4615	}
4616	else if (rc == VERR_NOT_SUPPORTED)
4617	break;
4618	else
4619	AssertMsg(rc == VERR_CPU_NOT_FOUND \|\| rc == VERR_CPU_OFFLINE, ("%d\n", rc));
4620	}
4621
4622	/* broke out of the loop. */
4623	if (iCpu <= iLastCpu)
4624	break;
4625	}
4626
4627	poffMin = offMin; / Almost RTMpOnSpecific profiling. */
4628	dprintf(("supdrvDetermineAsyncTsc: returns %d; iLastCpu=%d rc=%d offMin=%llx offMax=%llx\n",
4629	fAsync, iLastCpu, rc, offMin, offMax));
4630	#if !defined(RT_OS_SOLARIS) && !defined(RT_OS_OS2) && !defined(RT_OS_WINDOWS)
4631	OSDBGPRINT(("vboxdrv: fAsync=%d offMin=%#lx offMax=%#lx\n", fAsync, (long)offMin, (long)offMax));
4632	#endif
4633	return fAsync;
4634	}
4635
4636
4637	/**
4638	* Determin the GIP TSC mode.
4639	*
4640	* @returns The most suitable TSC mode.
4641	* @param pDevExt Pointer to the device instance data.
4642	*/
4643	static SUPGIPMODE supdrvGipDeterminTscMode(PSUPDRVDEVEXT pDevExt)
4644	{
4645	/*
4646	* On SMP we're faced with two problems:
4647	* (1) There might be a skew between the CPU, so that cpu0
4648	* returns a TSC that is sligtly different from cpu1.
4649	* (2) Power management (and other things) may cause the TSC
4650	* to run at a non-constant speed, and cause the speed
4651	* to be different on the cpus. This will result in (1).
4652	*
4653	* So, on SMP systems we'll have to select the ASYNC update method
4654	* if there are symphoms of these problems.
4655	*/
4656	if (RTMpGetCount() > 1)
4657	{
4658	uint32_t uEAX, uEBX, uECX, uEDX;
4659	uint64_t u64DiffCoresIgnored;
4660
4661	/* Permit the user and/or the OS specfic bits to force async mode. */
4662	if (supdrvOSGetForcedAsyncTscMode(pDevExt))
4663	return SUPGIPMODE_ASYNC_TSC;
4664
4665	/* Try check for current differences between the cpus. */
4666	if (supdrvDetermineAsyncTsc(&u64DiffCoresIgnored))
4667	return SUPGIPMODE_ASYNC_TSC;
4668
4669	/*
4670	* If the CPU supports power management and is an AMD one we
4671	* won't trust it unless it has the TscInvariant bit is set.
4672	*/
4673	/* Check for "AuthenticAMD" */
4674	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
4675	if ( uEAX >= 1
4676	&& uEBX == X86_CPUID_VENDOR_AMD_EBX
4677	&& uECX == X86_CPUID_VENDOR_AMD_ECX
4678	&& uEDX == X86_CPUID_VENDOR_AMD_EDX)
4679	{
4680	/* Check for APM support and that TscInvariant is cleared. */
4681	ASMCpuId(0x80000000, &uEAX, &uEBX, &uECX, &uEDX);
4682	if (uEAX >= 0x80000007)
4683	{
4684	ASMCpuId(0x80000007, &uEAX, &uEBX, &uECX, &uEDX);
4685	if ( !(uEDX & RT_BIT(8))/* TscInvariant */
4686	&& (uEDX & 0x3e)) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
4687	return SUPGIPMODE_ASYNC_TSC;
4688	}
4689	}
4690	}
4691	return SUPGIPMODE_SYNC_TSC;
4692	}
4693
4694
4695	/**
4696	* Invalidates the GIP data upon termination.
4697	*
4698	* @param pGip Pointer to the read-write kernel mapping of the GIP.
4699	*/
4700	void VBOXCALL supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
4701	{
4702	unsigned i;
4703	pGip->u32Magic = 0;
4704	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
4705	{
4706	pGip->aCPUs[i].u64NanoTS = 0;
4707	pGip->aCPUs[i].u64TSC = 0;
4708	pGip->aCPUs[i].iTSCHistoryHead = 0;
4709	}
4710	}
4711
4712
4713	/**
4714	* Worker routine for supdrvGipUpdate and supdrvGipUpdatePerCpu that
4715	* updates all the per cpu data except the transaction id.
4716	*
4717	* @param pGip The GIP.
4718	* @param pGipCpu Pointer to the per cpu data.
4719	* @param u64NanoTS The current time stamp.
4720	*/
4721	static void supdrvGipDoUpdateCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
4722	{
4723	uint64_t u64TSC;
4724	uint64_t u64TSCDelta;
4725	uint32_t u32UpdateIntervalTSC;
4726	uint32_t u32UpdateIntervalTSCSlack;
4727	unsigned iTSCHistoryHead;
4728	uint64_t u64CpuHz;
4729
4730	/*
4731	* Update the NanoTS.
4732	*/
4733	ASMAtomicXchgU64(&pGipCpu->u64NanoTS, u64NanoTS);
4734
4735	/*
4736	* Calc TSC delta.
4737	*/
4738	/** @todo validate the NanoTS delta, don't trust the OS to call us when it should... */
4739	u64TSC = ASMReadTSC();
4740	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
4741	ASMAtomicXchgU64(&pGipCpu->u64TSC, u64TSC);
4742
4743	if (u64TSCDelta >> 32)
4744	{
4745	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
4746	pGipCpu->cErrors++;
4747	}
4748
4749	/*
4750	* TSC History.
4751	*/
4752	Assert(ELEMENTS(pGipCpu->au32TSCHistory) == 8);
4753
4754	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
4755	ASMAtomicXchgU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
4756	ASMAtomicXchgU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
4757
4758	/*
4759	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
4760	*/
4761	if (pGip->u32UpdateHz >= 1000)
4762	{
4763	uint32_t u32;
4764	u32 = pGipCpu->au32TSCHistory[0];
4765	u32 += pGipCpu->au32TSCHistory[1];
4766	u32 += pGipCpu->au32TSCHistory[2];
4767	u32 += pGipCpu->au32TSCHistory[3];
4768	u32 >>= 2;
4769	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
4770	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
4771	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
4772	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
4773	u32UpdateIntervalTSC >>= 2;
4774	u32UpdateIntervalTSC += u32;
4775	u32UpdateIntervalTSC >>= 1;
4776
4777	/* Value choosen for a 2GHz Athlon64 running linux 2.6.10/11, . */
4778	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
4779	}
4780	else if (pGip->u32UpdateHz >= 90)
4781	{
4782	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
4783	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
4784	u32UpdateIntervalTSC >>= 1;
4785
4786	/* value choosen on a 2GHz thinkpad running windows */
4787	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
4788	}
4789	else
4790	{
4791	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
4792
4793	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
4794	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
4795	}
4796	ASMAtomicXchgU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
4797
4798	/*
4799	* CpuHz.
4800	*/
4801	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, pGip->u32UpdateHz);
4802	ASMAtomicXchgU64(&pGipCpu->u64CpuHz, u64CpuHz);
4803	}
4804
4805
4806	/**
4807	* Updates the GIP.
4808	*
4809	* @param pGip Pointer to the GIP.
4810	* @param u64NanoTS The current nanosecond timesamp.
4811	*/
4812	void VBOXCALL supdrvGipUpdate(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS)
4813	{
4814	/*
4815	* Determin the relevant CPU data.
4816	*/
4817	PSUPGIPCPU pGipCpu;
4818	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
4819	pGipCpu = &pGip->aCPUs[0];
4820	else
4821	{
4822	unsigned iCpu = ASMGetApicId();
4823	if (RT_LIKELY(iCpu >= RT_ELEMENTS(pGip->aCPUs)))
4824	return;
4825	pGipCpu = &pGip->aCPUs[iCpu];
4826	}
4827
4828	/*
4829	* Start update transaction.
4830	*/
4831	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
4832	{
4833	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
4834	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
4835	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
4836	pGipCpu->cErrors++;
4837	return;
4838	}
4839
4840	/*
4841	* Recalc the update frequency every 0x800th time.
4842	*/
4843	if (!(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
4844	{
4845	if (pGip->u64NanoTSLastUpdateHz)
4846	{
4847	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
4848	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
4849	uint32_t u32UpdateHz = (uint32_t)((UINT64_C(1000000000) * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
4850	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
4851	{
4852	ASMAtomicXchgU32(&pGip->u32UpdateHz, u32UpdateHz);
4853	ASMAtomicXchgU32(&pGip->u32UpdateIntervalNS, 1000000000 / u32UpdateHz);
4854	}
4855	#endif
4856	}
4857	ASMAtomicXchgU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS);
4858	}
4859
4860	/*
4861	* Update the data.
4862	*/
4863	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS);
4864
4865	/*
4866	* Complete transaction.
4867	*/
4868	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
4869	}
4870
4871
4872	/**
4873	* Updates the per cpu GIP data for the calling cpu.
4874	*
4875	* @param pGip Pointer to the GIP.
4876	* @param u64NanoTS The current nanosecond timesamp.
4877	* @param iCpu The CPU index.
4878	*/
4879	void VBOXCALL supdrvGipUpdatePerCpu(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, unsigned iCpu)
4880	{
4881	PSUPGIPCPU pGipCpu;
4882
4883	if (RT_LIKELY(iCpu < RT_ELEMENTS(pGip->aCPUs)))
4884	{
4885	pGipCpu = &pGip->aCPUs[iCpu];
4886
4887	/*
4888	* Start update transaction.
4889	*/
4890	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
4891	{
4892	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
4893	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
4894	pGipCpu->cErrors++;
4895	return;
4896	}
4897
4898	/*
4899	* Update the data.
4900	*/
4901	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS);
4902
4903	/*
4904	* Complete transaction.
4905	*/
4906	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
4907	}
4908	}
4909
4910
4911	#ifndef SUPDRV_WITH_RELEASE_LOGGER
4912	# ifndef DEBUG /** @todo change #ifndef DEBUG -> #ifdef LOG_ENABLED */
4913	/**
4914	* Stub function for non-debug builds.
4915	*/
4916	RTDECL(PRTLOGGER) RTLogDefaultInstance(void)
4917	{
4918	return NULL;
4919	}
4920
4921	RTDECL(PRTLOGGER) RTLogRelDefaultInstance(void)
4922	{
4923	return NULL;
4924	}
4925
4926	/**
4927	* Stub function for non-debug builds.
4928	*/
4929	RTDECL(int) RTLogSetDefaultInstanceThread(PRTLOGGER pLogger, uintptr_t uKey)
4930	{
4931	return 0;
4932	}
4933
4934	/**
4935	* Stub function for non-debug builds.
4936	*/
4937	RTDECL(void) RTLogLogger(PRTLOGGER pLogger, void pvCallerRet, const char pszFormat, ...)
4938	{
4939	}
4940
4941	/**
4942	* Stub function for non-debug builds.
4943	*/
4944	RTDECL(void) RTLogLoggerEx(PRTLOGGER pLogger, unsigned fFlags, unsigned iGroup, const char *pszFormat, ...)
4945	{
4946	}
4947
4948	/**
4949	* Stub function for non-debug builds.
4950	*/
4951	RTDECL(void) RTLogLoggerExV(PRTLOGGER pLogger, unsigned fFlags, unsigned iGroup, const char *pszFormat, va_list args)
4952	{
4953	}
4954
4955	/**
4956	* Stub function for non-debug builds.
4957	*/
4958	RTDECL(void) RTLogPrintf(const char *pszFormat, ...)
4959	{
4960	}
4961
4962	/**
4963	* Stub function for non-debug builds.
4964	*/
4965	RTDECL(void) RTLogPrintfV(const char *pszFormat, va_list args)
4966	{
4967	}
4968	# endif /* !DEBUG */
4969	#endif /* !SUPDRV_WITH_RELEASE_LOGGER */

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

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDrv.c@ 11836

Download in other formats: