SUPDRVShared.c@ 5999

Last change on this file since 5999 was 5999, checked in by vboxsync, 17 years ago
The Giant CDDL Dual-License Header Change.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 140.0 KB

Line
1	/* $Revision: 5999 $ */
2	/** @file
3	* VirtualBox Support Driver - Shared code.
4	*/
5
6	/*
7	* Copyright (C) 2006-2007 innotek GmbH
8	*
9	* This file is part of VirtualBox Open Source Edition (OSE), as
10	* available from http://www.virtualbox.org. This file is free software;
11	* you can redistribute it and/or modify it under the terms of the GNU
12	* General Public License (GPL) as published by the Free Software
13	* Foundation, in version 2 as it comes in the "COPYING" file of the
14	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16	*
17	* The contents of this file may alternatively be used under the terms
18	* of the Common Development and Distribution License Version 1.0
19	* (CDDL) only, as it comes in the "COPYING.CDDL" file of the
20	* VirtualBox OSE distribution, in which case the provisions of the
21	* CDDL are applicable instead of those of the GPL.
22	*
23	* You may elect to license modified versions of this file under the
24	* terms and conditions of either the GPL or the CDDL or both.
25	*/
26
27
28	/*******************************************************************************
29	* Header Files *
30	*******************************************************************************/
31	#include "SUPDRV.h"
32	#ifndef PAGE_SHIFT
33	# include <iprt/param.h>
34	#endif
35	#include <iprt/alloc.h>
36	#include <iprt/semaphore.h>
37	#include <iprt/spinlock.h>
38	#include <iprt/thread.h>
39	#include <iprt/process.h>
40	#include <iprt/log.h>
41
42
43	/*******************************************************************************
44	* Defined Constants And Macros *
45	*******************************************************************************/
46	/* from x86.h - clashes with linux thus this duplication */
47	#undef X86_CR0_PG
48	#define X86_CR0_PG RT_BIT(31)
49	#undef X86_CR0_PE
50	#define X86_CR0_PE RT_BIT(0)
51	#undef X86_CPUID_AMD_FEATURE_EDX_NX
52	#define X86_CPUID_AMD_FEATURE_EDX_NX RT_BIT(20)
53	#undef MSR_K6_EFER
54	#define MSR_K6_EFER 0xc0000080
55	#undef MSR_K6_EFER_NXE
56	#define MSR_K6_EFER_NXE RT_BIT(11)
57	#undef MSR_K6_EFER_LMA
58	#define MSR_K6_EFER_LMA RT_BIT(10)
59	#undef X86_CR4_PGE
60	#define X86_CR4_PGE RT_BIT(7)
61	#undef X86_CR4_PAE
62	#define X86_CR4_PAE RT_BIT(5)
63	#undef X86_CPUID_AMD_FEATURE_EDX_LONG_MODE
64	#define X86_CPUID_AMD_FEATURE_EDX_LONG_MODE RT_BIT(29)
65
66
67	/** The frequency by which we recalculate the u32UpdateHz and
68	* u32UpdateIntervalNS GIP members. The value must be a power of 2. */
69	#define GIP_UPDATEHZ_RECALC_FREQ 0x800
70
71	/**
72	* Validates a session pointer.
73	*
74	* @returns true/false accordingly.
75	* @param pSession The session.
76	*/
77	#define SUP_IS_SESSION_VALID(pSession) \
78	( VALID_PTR(pSession) \
79	&& pSession->u32Cookie == BIRD_INV)
80
81
82	/*******************************************************************************
83	* Global Variables *
84	*******************************************************************************/
85	/**
86	* Array of the R0 SUP API.
87	*/
88	static SUPFUNC g_aFunctions[] =
89	{
90	/* name function */
91	{ "SUPR0ObjRegister", (void *)SUPR0ObjRegister },
92	{ "SUPR0ObjAddRef", (void *)SUPR0ObjAddRef },
93	{ "SUPR0ObjRelease", (void *)SUPR0ObjRelease },
94	{ "SUPR0ObjVerifyAccess", (void *)SUPR0ObjVerifyAccess },
95	{ "SUPR0LockMem", (void *)SUPR0LockMem },
96	{ "SUPR0UnlockMem", (void *)SUPR0UnlockMem },
97	{ "SUPR0ContAlloc", (void *)SUPR0ContAlloc },
98	{ "SUPR0ContFree", (void *)SUPR0ContFree },
99	{ "SUPR0LowAlloc", (void *)SUPR0LowAlloc },
100	{ "SUPR0LowFree", (void *)SUPR0LowFree },
101	{ "SUPR0MemAlloc", (void *)SUPR0MemAlloc },
102	{ "SUPR0MemGetPhys", (void *)SUPR0MemGetPhys },
103	{ "SUPR0MemFree", (void *)SUPR0MemFree },
104	{ "SUPR0PageAlloc", (void *)SUPR0PageAlloc },
105	{ "SUPR0PageFree", (void *)SUPR0PageFree },
106	{ "SUPR0Printf", (void *)SUPR0Printf },
107	{ "RTMemAlloc", (void *)RTMemAlloc },
108	{ "RTMemAllocZ", (void *)RTMemAllocZ },
109	{ "RTMemFree", (void *)RTMemFree },
110	/{ "RTMemDup", (void )RTMemDup },*/
111	{ "RTMemRealloc", (void *)RTMemRealloc },
112	{ "RTR0MemObjAllocLow", (void *)RTR0MemObjAllocLow },
113	{ "RTR0MemObjAllocPage", (void *)RTR0MemObjAllocPage },
114	{ "RTR0MemObjAllocPhys", (void *)RTR0MemObjAllocPhys },
115	{ "RTR0MemObjAllocPhysNC", (void *)RTR0MemObjAllocPhysNC },
116	{ "RTR0MemObjLockUser", (void *)RTR0MemObjLockUser },
117	{ "RTR0MemObjMapKernel", (void *)RTR0MemObjMapKernel },
118	{ "RTR0MemObjMapUser", (void *)RTR0MemObjMapUser },
119	{ "RTR0MemObjAddress", (void *)RTR0MemObjAddress },
120	{ "RTR0MemObjAddressR3", (void *)RTR0MemObjAddressR3 },
121	{ "RTR0MemObjSize", (void *)RTR0MemObjSize },
122	{ "RTR0MemObjIsMapping", (void *)RTR0MemObjIsMapping },
123	{ "RTR0MemObjGetPagePhysAddr", (void *)RTR0MemObjGetPagePhysAddr },
124	{ "RTR0MemObjFree", (void *)RTR0MemObjFree },
125	/* These doesn't work yet on linux - use fast mutexes!
126	{ "RTSemMutexCreate", (void *)RTSemMutexCreate },
127	{ "RTSemMutexRequest", (void *)RTSemMutexRequest },
128	{ "RTSemMutexRelease", (void *)RTSemMutexRelease },
129	{ "RTSemMutexDestroy", (void *)RTSemMutexDestroy },
130	*/
131	{ "RTSemFastMutexCreate", (void *)RTSemFastMutexCreate },
132	{ "RTSemFastMutexDestroy", (void *)RTSemFastMutexDestroy },
133	{ "RTSemFastMutexRequest", (void *)RTSemFastMutexRequest },
134	{ "RTSemFastMutexRelease", (void *)RTSemFastMutexRelease },
135	{ "RTSemEventCreate", (void *)RTSemEventCreate },
136	{ "RTSemEventSignal", (void *)RTSemEventSignal },
137	{ "RTSemEventWait", (void *)RTSemEventWait },
138	{ "RTSemEventWaitNoResume", (void *)RTSemEventWaitNoResume },
139	{ "RTSemEventDestroy", (void *)RTSemEventDestroy },
140	{ "RTSemEventMultiCreate", (void *)RTSemEventMultiCreate },
141	{ "RTSemEventMultiSignal", (void *)RTSemEventMultiSignal },
142	{ "RTSemEventMultiReset", (void *)RTSemEventMultiReset },
143	{ "RTSemEventMultiWait", (void *)RTSemEventMultiWait },
144	{ "RTSemEventMultiWaitNoResume", (void *)RTSemEventMultiWaitNoResume },
145	{ "RTSemEventMultiDestroy", (void *)RTSemEventMultiDestroy },
146	{ "RTSpinlockCreate", (void *)RTSpinlockCreate },
147	{ "RTSpinlockDestroy", (void *)RTSpinlockDestroy },
148	{ "RTSpinlockAcquire", (void *)RTSpinlockAcquire },
149	{ "RTSpinlockRelease", (void *)RTSpinlockRelease },
150	{ "RTSpinlockAcquireNoInts", (void *)RTSpinlockAcquireNoInts },
151	{ "RTSpinlockReleaseNoInts", (void *)RTSpinlockReleaseNoInts },
152	{ "RTThreadNativeSelf", (void *)RTThreadNativeSelf },
153	{ "RTThreadSleep", (void *)RTThreadSleep },
154	{ "RTThreadYield", (void *)RTThreadYield },
155	#if 0 /* Thread APIs, Part 2. */
156	{ "RTThreadSelf", (void *)RTThreadSelf },
157	{ "RTThreadCreate", (void *)RTThreadCreate },
158	{ "RTThreadGetNative", (void *)RTThreadGetNative },
159	{ "RTThreadWait", (void *)RTThreadWait },
160	{ "RTThreadWaitNoResume", (void *)RTThreadWaitNoResume },
161	{ "RTThreadGetName", (void *)RTThreadGetName },
162	{ "RTThreadSelfName", (void *)RTThreadSelfName },
163	{ "RTThreadGetType", (void *)RTThreadGetType },
164	{ "RTThreadUserSignal", (void *)RTThreadUserSignal },
165	{ "RTThreadUserReset", (void *)RTThreadUserReset },
166	{ "RTThreadUserWait", (void *)RTThreadUserWait },
167	{ "RTThreadUserWaitNoResume", (void *)RTThreadUserWaitNoResume },
168	#endif
169	{ "RTLogDefaultInstance", (void *)RTLogDefaultInstance },
170	{ "RTLogRelDefaultInstance", (void *)RTLogRelDefaultInstance },
171	{ "RTLogSetDefaultInstanceThread", (void *)RTLogSetDefaultInstanceThread },
172	{ "RTLogLogger", (void *)RTLogLogger },
173	{ "RTLogLoggerEx", (void *)RTLogLoggerEx },
174	{ "RTLogLoggerExV", (void *)RTLogLoggerExV },
175	{ "RTLogPrintf", (void *)RTLogPrintf },
176	{ "RTLogPrintfV", (void *)RTLogPrintfV },
177	{ "AssertMsg1", (void *)AssertMsg1 },
178	{ "AssertMsg2", (void *)AssertMsg2 },
179	};
180
181
182	/*******************************************************************************
183	* Internal Functions *
184	*******************************************************************************/
185	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession);
186	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType);
187	#ifdef VBOX_WITH_IDT_PATCHING
188	static int supdrvIOCtl_IdtInstall(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPIDTINSTALL pReq);
189	static PSUPDRVPATCH supdrvIdtPatchOne(PSUPDRVDEVEXT pDevExt, PSUPDRVPATCH pPatch);
190	static int supdrvIOCtl_IdtRemoveAll(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession);
191	static void supdrvIdtRemoveOne(PSUPDRVDEVEXT pDevExt, PSUPDRVPATCH pPatch);
192	static void supdrvIdtWrite(volatile void pvIdtEntry, const SUPDRVIDTE pNewIDTEntry);
193	#endif /* VBOX_WITH_IDT_PATCHING */
194	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq);
195	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq);
196	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq);
197	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq);
198	static int supdrvLdrSetR0EP(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx);
199	static void supdrvLdrUnsetR0EP(PSUPDRVDEVEXT pDevExt);
200	static void supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage);
201	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage);
202	static SUPPAGINGMODE supdrvIOCtl_GetPagingMode(void);
203	static SUPGIPMODE supdrvGipDeterminTscMode(void);
204	#ifdef RT_OS_WINDOWS
205	static int supdrvPageGetPhys(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages);
206	static bool supdrvPageWasLockedByPageAlloc(PSUPDRVSESSION pSession, RTR3PTR pvR3);
207	#endif
208	#ifdef USE_NEW_OS_INTERFACE_FOR_GIP
209	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt);
210	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt);
211	static DECLCALLBACK(void) supdrvGipTimer(PRTTIMER pTimer, void *pvUser);
212	#endif
213
214
215	/**
216	* Initializes the device extentsion structure.
217	*
218	* @returns IPRT status code.
219	* @param pDevExt The device extension to initialize.
220	*/
221	int VBOXCALL supdrvInitDevExt(PSUPDRVDEVEXT pDevExt)
222	{
223	/*
224	* Initialize it.
225	*/
226	int rc;
227	memset(pDevExt, 0, sizeof(*pDevExt));
228	rc = RTSpinlockCreate(&pDevExt->Spinlock);
229	if (!rc)
230	{
231	rc = RTSemFastMutexCreate(&pDevExt->mtxLdr);
232	if (!rc)
233	{
234	rc = RTSemFastMutexCreate(&pDevExt->mtxGip);
235	if (!rc)
236	{
237	#ifdef USE_NEW_OS_INTERFACE_FOR_GIP
238	rc = supdrvGipCreate(pDevExt);
239	if (RT_SUCCESS(rc))
240	{
241	pDevExt->u32Cookie = BIRD; /** @todo make this random? */
242	return VINF_SUCCESS;
243	}
244	#else
245	pDevExt->u32Cookie = BIRD;
246	return VINF_SUCCESS;
247	#endif
248	}
249	RTSemFastMutexDestroy(pDevExt->mtxLdr);
250	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
251	}
252	RTSpinlockDestroy(pDevExt->Spinlock);
253	pDevExt->Spinlock = NIL_RTSPINLOCK;
254	}
255	return rc;
256	}
257
258
259	/**
260	* Delete the device extension (e.g. cleanup members).
261	*
262	* @param pDevExt The device extension to delete.
263	*/
264	void VBOXCALL supdrvDeleteDevExt(PSUPDRVDEVEXT pDevExt)
265	{
266	#ifdef VBOX_WITH_IDT_PATCHING
267	PSUPDRVPATCH pPatch;
268	#endif
269	PSUPDRVOBJ pObj;
270	PSUPDRVUSAGE pUsage;
271
272	/*
273	* Kill mutexes and spinlocks.
274	*/
275	RTSemFastMutexDestroy(pDevExt->mtxGip);
276	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
277	RTSemFastMutexDestroy(pDevExt->mtxLdr);
278	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
279	RTSpinlockDestroy(pDevExt->Spinlock);
280	pDevExt->Spinlock = NIL_RTSPINLOCK;
281
282	/*
283	* Free lists.
284	*/
285	#ifdef VBOX_WITH_IDT_PATCHING
286	/* patches */
287	/** @todo make sure we don't uninstall patches which has been patched by someone else. */
288	pPatch = pDevExt->pIdtPatchesFree;
289	pDevExt->pIdtPatchesFree = NULL;
290	while (pPatch)
291	{
292	void *pvFree = pPatch;
293	pPatch = pPatch->pNext;
294	RTMemExecFree(pvFree);
295	}
296	#endif /* VBOX_WITH_IDT_PATCHING */
297
298	/* objects. */
299	pObj = pDevExt->pObjs;
300	#if !defined(DEBUG_bird) \|\| !defined(RT_OS_LINUX) /* breaks unloading, temporary, remove me! */
301	Assert(!pObj); /* (can trigger on forced unloads) */
302	#endif
303	pDevExt->pObjs = NULL;
304	while (pObj)
305	{
306	void *pvFree = pObj;
307	pObj = pObj->pNext;
308	RTMemFree(pvFree);
309	}
310
311	/* usage records. */
312	pUsage = pDevExt->pUsageFree;
313	pDevExt->pUsageFree = NULL;
314	while (pUsage)
315	{
316	void *pvFree = pUsage;
317	pUsage = pUsage->pNext;
318	RTMemFree(pvFree);
319	}
320
321	#ifdef USE_NEW_OS_INTERFACE_FOR_GIP
322	/* kill the GIP */
323	supdrvGipDestroy(pDevExt);
324	#endif
325	}
326
327
328	/**
329	* Create session.
330	*
331	* @returns IPRT status code.
332	* @param pDevExt Device extension.
333	* @param ppSession Where to store the pointer to the session data.
334	*/
335	int VBOXCALL supdrvCreateSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION *ppSession)
336	{
337	/*
338	* Allocate memory for the session data.
339	*/
340	int rc = VERR_NO_MEMORY;
341	PSUPDRVSESSION pSession = ppSession = (PSUPDRVSESSION)RTMemAllocZ(sizeof(pSession));
342	if (pSession)
343	{
344	/* Initialize session data. */
345	rc = RTSpinlockCreate(&pSession->Spinlock);
346	if (!rc)
347	{
348	Assert(pSession->Spinlock != NIL_RTSPINLOCK);
349	pSession->pDevExt = pDevExt;
350	pSession->u32Cookie = BIRD_INV;
351	/*pSession->pLdrUsage = NULL;
352	pSession->pPatchUsage = NULL;
353	pSession->pUsage = NULL;
354	pSession->pGip = NULL;
355	pSession->fGipReferenced = false;
356	pSession->Bundle.cUsed = 0 */
357
358	dprintf(("Created session %p initial cookie=%#x\n", pSession, pSession->u32Cookie));
359	return VINF_SUCCESS;
360	}
361
362	RTMemFree(pSession);
363	*ppSession = NULL;
364	}
365
366	dprintf(("Failed to create spinlock, rc=%d!\n", rc));
367	return rc;
368	}
369
370
371	/**
372	* Shared code for cleaning up a session.
373	*
374	* @param pDevExt Device extension.
375	* @param pSession Session data.
376	* This data will be freed by this routine.
377	*/
378	void VBOXCALL supdrvCloseSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
379	{
380	/*
381	* Cleanup the session first.
382	*/
383	supdrvCleanupSession(pDevExt, pSession);
384
385	/*
386	* Free the rest of the session stuff.
387	*/
388	RTSpinlockDestroy(pSession->Spinlock);
389	pSession->Spinlock = NIL_RTSPINLOCK;
390	pSession->pDevExt = NULL;
391	RTMemFree(pSession);
392	dprintf2(("supdrvCloseSession: returns\n"));
393	}
394
395
396	/**
397	* Shared code for cleaning up a session (but not quite freeing it).
398	*
399	* This is primarily intended for MAC OS X where we have to clean up the memory
400	* stuff before the file handle is closed.
401	*
402	* @param pDevExt Device extension.
403	* @param pSession Session data.
404	* This data will be freed by this routine.
405	*/
406	void VBOXCALL supdrvCleanupSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
407	{
408	PSUPDRVBUNDLE pBundle;
409	dprintf(("supdrvCleanupSession: pSession=%p\n", pSession));
410
411	/*
412	* Remove logger instances related to this session.
413	* (This assumes the dprintf and dprintf2 macros doesn't use the normal logging.)
414	*/
415	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pSession);
416
417	#ifdef VBOX_WITH_IDT_PATCHING
418	/*
419	* Uninstall any IDT patches installed for this session.
420	*/
421	supdrvIOCtl_IdtRemoveAll(pDevExt, pSession);
422	#endif
423
424	/*
425	* Release object references made in this session.
426	* In theory there should be noone racing us in this session.
427	*/
428	dprintf2(("release objects - start\n"));
429	if (pSession->pUsage)
430	{
431	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
432	PSUPDRVUSAGE pUsage;
433	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
434
435	while ((pUsage = pSession->pUsage) != NULL)
436	{
437	PSUPDRVOBJ pObj = pUsage->pObj;
438	pSession->pUsage = pUsage->pNext;
439
440	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
441	if (pUsage->cUsage < pObj->cUsage)
442	{
443	pObj->cUsage -= pUsage->cUsage;
444	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
445	}
446	else
447	{
448	/* Destroy the object and free the record. */
449	if (pDevExt->pObjs == pObj)
450	pDevExt->pObjs = pObj->pNext;
451	else
452	{
453	PSUPDRVOBJ pObjPrev;
454	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
455	if (pObjPrev->pNext == pObj)
456	{
457	pObjPrev->pNext = pObj->pNext;
458	break;
459	}
460	Assert(pObjPrev);
461	}
462	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
463
464	if (pObj->pfnDestructor)
465	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
466	RTMemFree(pObj);
467	}
468
469	/* free it and continue. */
470	RTMemFree(pUsage);
471
472	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
473	}
474
475	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
476	AssertMsg(!pSession->pUsage, ("Some buster reregistered an object during desturction!\n"));
477	}
478	dprintf2(("release objects - done\n"));
479
480	/*
481	* Release memory allocated in the session.
482	*
483	* We do not serialize this as we assume that the application will
484	* not allocated memory while closing the file handle object.
485	*/
486	dprintf2(("freeing memory:\n"));
487	pBundle = &pSession->Bundle;
488	while (pBundle)
489	{
490	PSUPDRVBUNDLE pToFree;
491	unsigned i;
492
493	/*
494	* Check and unlock all entries in the bundle.
495	*/
496	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
497	{
498	if (pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ)
499	{
500	int rc;
501	dprintf2(("eType=%d pvR0=%p pvR3=%p cb=%ld\n", pBundle->aMem[i].eType, RTR0MemObjAddress(pBundle->aMem[i].MemObj),
502	(void *)RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3), (long)RTR0MemObjSize(pBundle->aMem[i].MemObj)));
503	if (pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ)
504	{
505	rc = RTR0MemObjFree(pBundle->aMem[i].MapObjR3, false);
506	AssertRC(rc); /** @todo figure out how to handle this. */
507	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
508	}
509	rc = RTR0MemObjFree(pBundle->aMem[i].MemObj, false);
510	AssertRC(rc); /** @todo figure out how to handle this. */
511	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
512	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
513	}
514	}
515
516	/*
517	* Advance and free previous bundle.
518	*/
519	pToFree = pBundle;
520	pBundle = pBundle->pNext;
521
522	pToFree->pNext = NULL;
523	pToFree->cUsed = 0;
524	if (pToFree != &pSession->Bundle)
525	RTMemFree(pToFree);
526	}
527	dprintf2(("freeing memory - done\n"));
528
529	/*
530	* Loaded images needs to be dereferenced and possibly freed up.
531	*/
532	RTSemFastMutexRequest(pDevExt->mtxLdr);
533	dprintf2(("freeing images:\n"));
534	if (pSession->pLdrUsage)
535	{
536	PSUPDRVLDRUSAGE pUsage = pSession->pLdrUsage;
537	pSession->pLdrUsage = NULL;
538	while (pUsage)
539	{
540	void *pvFree = pUsage;
541	PSUPDRVLDRIMAGE pImage = pUsage->pImage;
542	if (pImage->cUsage > pUsage->cUsage)
543	pImage->cUsage -= pUsage->cUsage;
544	else
545	supdrvLdrFree(pDevExt, pImage);
546	pUsage->pImage = NULL;
547	pUsage = pUsage->pNext;
548	RTMemFree(pvFree);
549	}
550	}
551	RTSemFastMutexRelease(pDevExt->mtxLdr);
552	dprintf2(("freeing images - done\n"));
553
554	/*
555	* Unmap the GIP.
556	*/
557	dprintf2(("umapping GIP:\n"));
558	#ifdef USE_NEW_OS_INTERFACE_FOR_GIP
559	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
560	#else
561	if (pSession->pGip)
562	#endif
563	{
564	SUPR0GipUnmap(pSession);
565	#ifndef USE_NEW_OS_INTERFACE_FOR_GIP
566	pSession->pGip = NULL;
567	#endif
568	pSession->fGipReferenced = 0;
569	}
570	dprintf2(("umapping GIP - done\n"));
571	}
572
573
574	/**
575	* Fast path I/O Control worker.
576	*
577	* @returns VBox status code that should be passed down to ring-3 unchanged.
578	* @param uIOCtl Function number.
579	* @param pDevExt Device extention.
580	* @param pSession Session data.
581	*/
582	int VBOXCALL supdrvIOCtlFast(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
583	{
584	int rc;
585
586	/*
587	* We check the two prereqs after doing this only to allow the compiler to optimize things better.
588	*/
589	if (RT_LIKELY(pSession->pVM && pDevExt->pfnVMMR0EntryFast))
590	{
591	switch (uIOCtl)
592	{
593	case SUP_IOCTL_FAST_DO_RAW_RUN:
594	rc = pDevExt->pfnVMMR0EntryFast(pSession->pVM, SUP_VMMR0_DO_RAW_RUN);
595	break;
596	case SUP_IOCTL_FAST_DO_HWACC_RUN:
597	rc = pDevExt->pfnVMMR0EntryFast(pSession->pVM, SUP_VMMR0_DO_HWACC_RUN);
598	break;
599	case SUP_IOCTL_FAST_DO_NOP:
600	rc = pDevExt->pfnVMMR0EntryFast(pSession->pVM, SUP_VMMR0_DO_NOP);
601	break;
602	default:
603	rc = VERR_INTERNAL_ERROR;
604	break;
605	}
606	}
607	else
608	rc = VERR_INTERNAL_ERROR;
609
610	return rc;
611	}
612
613
614	/**
615	* I/O Control worker.
616	*
617	* @returns 0 on success.
618	* @returns VERR_INVALID_PARAMETER if the request is invalid.
619	*
620	* @param uIOCtl Function number.
621	* @param pDevExt Device extention.
622	* @param pSession Session data.
623	* @param pReqHdr The request header.
624	*/
625	int VBOXCALL supdrvIOCtl(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
626	{
627	/*
628	* Validate the request.
629	*/
630	/* this first check could probably be omitted as its also done by the OS specific code... */
631	if (RT_UNLIKELY( (pReqHdr->fFlags & SUPREQHDR_FLAGS_MAGIC_MASK) != SUPREQHDR_FLAGS_MAGIC
632	\|\| pReqHdr->cbIn < sizeof(*pReqHdr)
633	\|\| pReqHdr->cbOut < sizeof(*pReqHdr)))
634	{
635	OSDBGPRINT(("vboxdrv: Bad ioctl request header; cbIn=%#lx cbOut=%#lx fFlags=%#lx\n",
636	(long)pReqHdr->cbIn, (long)pReqHdr->cbOut, (long)pReqHdr->fFlags));
637	return VERR_INVALID_PARAMETER;
638	}
639	if (RT_UNLIKELY(uIOCtl == SUP_IOCTL_COOKIE))
640	{
641	if (pReqHdr->u32Cookie != SUPCOOKIE_INITIAL_COOKIE)
642	{
643	OSDBGPRINT(("SUP_IOCTL_COOKIE: bad cookie %#lx\n", (long)pReqHdr->u32Cookie));
644	return VERR_INVALID_PARAMETER;
645	}
646	}
647	else if (RT_UNLIKELY( pReqHdr->u32Cookie != pDevExt->u32Cookie
648	\|\| pReqHdr->u32SessionCookie != pSession->u32Cookie))
649	{
650	OSDBGPRINT(("vboxdrv: bad cookie %#lx / %#lx.\n", (long)pReqHdr->u32Cookie, (long)pReqHdr->u32SessionCookie));
651	return VERR_INVALID_PARAMETER;
652	}
653
654	/*
655	* Validation macros
656	*/
657	#define REQ_CHECK_SIZES_EX(Name, cbInExpect, cbOutExpect) \
658	do { \
659	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect) \|\| pReqHdr->cbOut != (cbOutExpect))) \
660	{ \
661	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld. cbOut=%ld expected %ld.\n", \
662	(long)pReq->Hdr.cbIn, (long)(cbInExpect), (long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
663	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
664	} \
665	} while (0)
666
667	#define REQ_CHECK_SIZES(Name) REQ_CHECK_SIZES_EX(Name, Name ## _SIZE_IN, Name ## _SIZE_OUT)
668
669	#define REQ_CHECK_SIZE_IN(Name, cbInExpect) \
670	do { \
671	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect))) \
672	{ \
673	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld.\n", \
674	(long)pReq->Hdr.cbIn, (long)(cbInExpect))); \
675	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
676	} \
677	} while (0)
678
679	#define REQ_CHECK_SIZE_OUT(Name, cbOutExpect) \
680	do { \
681	if (RT_UNLIKELY(pReqHdr->cbOut != (cbOutExpect))) \
682	{ \
683	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbOut=%ld expected %ld.\n", \
684	(long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
685	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
686	} \
687	} while (0)
688
689	#define REQ_CHECK_EXPR(Name, expr) \
690	do { \
691	if (RT_UNLIKELY(!(expr))) \
692	{ \
693	OSDBGPRINT(( #Name ": %s\n", #expr)); \
694	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
695	} \
696	} while (0)
697
698	#define REQ_CHECK_EXPR_FMT(expr, fmt) \
699	do { \
700	if (RT_UNLIKELY(!(expr))) \
701	{ \
702	OSDBGPRINT( fmt ); \
703	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
704	} \
705	} while (0)
706
707
708	/*
709	* The switch.
710	*/
711	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
712	{
713	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
714	{
715	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
716	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
717	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
718	{
719	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
720	pReq->Hdr.rc = VERR_INVALID_MAGIC;
721	return 0;
722	}
723
724	#if 0
725	/*
726	* Call out to the OS specific code and let it do permission checks on the
727	* client process.
728	*/
729	if (!supdrvOSValidateClientProcess(pDevExt, pSession))
730	{
731	pReq->u.Out.u32Cookie = 0xffffffff;
732	pReq->u.Out.u32SessionCookie = 0xffffffff;
733	pReq->u.Out.u32SessionVersion = 0xffffffff;
734	pReq->u.Out.u32DriverVersion = SUPDRVIOC_VERSION;
735	pReq->u.Out.pSession = NULL;
736	pReq->u.Out.cFunctions = 0;
737	pReq->Hdr.rc = VERR_PERMISSION_DENIED;
738	return 0;
739	}
740	#endif
741
742	/*
743	* Match the version.
744	* The current logic is very simple, match the major interface version.
745	*/
746	if ( pReq->u.In.u32MinVersion > SUPDRVIOC_VERSION
747	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRVIOC_VERSION & 0xffff0000))
748	{
749	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
750	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRVIOC_VERSION));
751	pReq->u.Out.u32Cookie = 0xffffffff;
752	pReq->u.Out.u32SessionCookie = 0xffffffff;
753	pReq->u.Out.u32SessionVersion = 0xffffffff;
754	pReq->u.Out.u32DriverVersion = SUPDRVIOC_VERSION;
755	pReq->u.Out.pSession = NULL;
756	pReq->u.Out.cFunctions = 0;
757	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
758	return 0;
759	}
760
761	/*
762	* Fill in return data and be gone.
763	* N.B. The first one to change SUPDRVIOC_VERSION shall makes sure that
764	* u32SessionVersion <= u32ReqVersion!
765	*/
766	/** @todo Somehow validate the client and negotiate a secure cookie... */
767	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
768	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
769	pReq->u.Out.u32SessionVersion = SUPDRVIOC_VERSION;
770	pReq->u.Out.u32DriverVersion = SUPDRVIOC_VERSION;
771	pReq->u.Out.pSession = pSession;
772	pReq->u.Out.cFunctions = sizeof(g_aFunctions) / sizeof(g_aFunctions[0]);
773	pReq->Hdr.rc = VINF_SUCCESS;
774	return 0;
775	}
776
777	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_QUERY_FUNCS(0)):
778	{
779	/* validate */
780	PSUPQUERYFUNCS pReq = (PSUPQUERYFUNCS)pReqHdr;
781	REQ_CHECK_SIZES_EX(SUP_IOCTL_QUERY_FUNCS, SUP_IOCTL_QUERY_FUNCS_SIZE_IN, SUP_IOCTL_QUERY_FUNCS_SIZE_OUT(RT_ELEMENTS(g_aFunctions)));
782
783	/* execute */
784	pReq->u.Out.cFunctions = RT_ELEMENTS(g_aFunctions);
785	memcpy(&pReq->u.Out.aFunctions[0], g_aFunctions, sizeof(g_aFunctions));
786	pReq->Hdr.rc = VINF_SUCCESS;
787	return 0;
788	}
789
790	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_IDT_INSTALL):
791	{
792	/* validate */
793	PSUPIDTINSTALL pReq = (PSUPIDTINSTALL)pReqHdr;
794	REQ_CHECK_SIZES(SUP_IOCTL_IDT_INSTALL);
795
796	/* execute */
797	#ifdef VBOX_WITH_IDT_PATCHING
798	pReq->Hdr.rc = supdrvIOCtl_IdtInstall(pDevExt, pSession, pReq);
799	#else
800	pReq->u.Out.u8Idt = 3;
801	pReq->Hdr.rc = VERR_NOT_SUPPORTED;
802	#endif
803	return 0;
804	}
805
806	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_IDT_REMOVE):
807	{
808	/* validate */
809	PSUPIDTREMOVE pReq = (PSUPIDTREMOVE)pReqHdr;
810	REQ_CHECK_SIZES(SUP_IOCTL_IDT_REMOVE);
811
812	/* execute */
813	#ifdef VBOX_WITH_IDT_PATCHING
814	pReq->Hdr.rc = supdrvIOCtl_IdtRemoveAll(pDevExt, pSession);
815	#else
816	pReq->Hdr.rc = VERR_NOT_SUPPORTED;
817	#endif
818	return 0;
819	}
820
821	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_LOCK):
822	{
823	/* validate */
824	PSUPPAGELOCK pReq = (PSUPPAGELOCK)pReqHdr;
825	REQ_CHECK_SIZE_IN(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_IN);
826	REQ_CHECK_SIZE_OUT(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_OUT(pReq->u.In.cPages));
827	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.cPages > 0);
828	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.pvR3 >= PAGE_SIZE);
829
830	/* execute */
831	pReq->Hdr.rc = SUPR0LockMem(pSession, pReq->u.In.pvR3, pReq->u.In.cPages, &pReq->u.Out.aPages[0]);
832	if (RT_FAILURE(pReq->Hdr.rc))
833	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
834	return 0;
835	}
836
837	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_UNLOCK):
838	{
839	/* validate */
840	PSUPPAGEUNLOCK pReq = (PSUPPAGEUNLOCK)pReqHdr;
841	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_UNLOCK);
842
843	/* execute */
844	pReq->Hdr.rc = SUPR0UnlockMem(pSession, pReq->u.In.pvR3);
845	return 0;
846	}
847
848	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_ALLOC):
849	{
850	/* validate */
851	PSUPCONTALLOC pReq = (PSUPCONTALLOC)pReqHdr;
852	REQ_CHECK_SIZES(SUP_IOCTL_CONT_ALLOC);
853
854	/* execute */
855	pReq->Hdr.rc = SUPR0ContAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.HCPhys);
856	if (RT_FAILURE(pReq->Hdr.rc))
857	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
858	return 0;
859	}
860
861	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_FREE):
862	{
863	/* validate */
864	PSUPCONTFREE pReq = (PSUPCONTFREE)pReqHdr;
865	REQ_CHECK_SIZES(SUP_IOCTL_CONT_FREE);
866
867	/* execute */
868	pReq->Hdr.rc = SUPR0ContFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
869	return 0;
870	}
871
872	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_OPEN):
873	{
874	/* validate */
875	PSUPLDROPEN pReq = (PSUPLDROPEN)pReqHdr;
876	REQ_CHECK_SIZES(SUP_IOCTL_LDR_OPEN);
877	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImage > 0);
878	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImage < _1M*16);
879	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.szName[0]);
880	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, memchr(pReq->u.In.szName, '\0', sizeof(pReq->u.In.szName)));
881	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, !strpbrk(pReq->u.In.szName, ";:()[]{}/\\\|&*%#@!~`\"'"));
882
883	/* execute */
884	pReq->Hdr.rc = supdrvIOCtl_LdrOpen(pDevExt, pSession, pReq);
885	return 0;
886	}
887
888	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOAD):
889	{
890	/* validate */
891	PSUPLDRLOAD pReq = (PSUPLDRLOAD)pReqHdr;
892	REQ_CHECK_EXPR(Name, pReq->Hdr.cbIn >= sizeof(*pReq));
893	REQ_CHECK_SIZES_EX(SUP_IOCTL_LDR_LOAD, SUP_IOCTL_LDR_LOAD_SIZE_IN(pReq->u.In.cbImage), SUP_IOCTL_LDR_LOAD_SIZE_OUT);
894	REQ_CHECK_EXPR(SUP_IOCTL_LDR_LOAD, pReq->u.In.cSymbols <= 16384);
895	REQ_CHECK_EXPR_FMT( !pReq->u.In.cSymbols
896	\|\| ( pReq->u.In.offSymbols < pReq->u.In.cbImage
897	&& pReq->u.In.offSymbols + pReq->u.In.cSymbols * sizeof(SUPLDRSYM) <= pReq->u.In.cbImage),
898	("SUP_IOCTL_LDR_LOAD: offSymbols=%#lx cSymbols=%#lx cbImage=%#lx\n", (long)pReq->u.In.offSymbols,
899	(long)pReq->u.In.cSymbols, (long)pReq->u.In.cbImage));
900	REQ_CHECK_EXPR_FMT( !pReq->u.In.cbStrTab
901	\|\| ( pReq->u.In.offStrTab < pReq->u.In.cbImage
902	&& pReq->u.In.offStrTab + pReq->u.In.cbStrTab <= pReq->u.In.cbImage
903	&& pReq->u.In.cbStrTab <= pReq->u.In.cbImage),
904	("SUP_IOCTL_LDR_LOAD: offStrTab=%#lx cbStrTab=%#lx cbImage=%#lx\n", (long)pReq->u.In.offStrTab,
905	(long)pReq->u.In.cbStrTab, (long)pReq->u.In.cbImage));
906
907	if (pReq->u.In.cSymbols)
908	{
909	uint32_t i;
910	PSUPLDRSYM paSyms = (PSUPLDRSYM)&pReq->u.In.achImage[pReq->u.In.offSymbols];
911	for (i = 0; i < pReq->u.In.cSymbols; i++)
912	{
913	REQ_CHECK_EXPR_FMT(paSyms[i].offSymbol < pReq->u.In.cbImage,
914	("SUP_IOCTL_LDR_LOAD: sym #%ld: symb off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offSymbol, (long)pReq->u.In.cbImage));
915	REQ_CHECK_EXPR_FMT(paSyms[i].offName < pReq->u.In.cbStrTab,
916	("SUP_IOCTL_LDR_LOAD: sym #%ld: name off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImage));
917	REQ_CHECK_EXPR_FMT(memchr(&pReq->u.In.achImage[pReq->u.In.offStrTab + paSyms[i].offName], '\0', pReq->u.In.cbStrTab - paSyms[i].offName),
918	("SUP_IOCTL_LDR_LOAD: sym #%ld: unterminated name! (%#lx / %#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImage));
919	}
920	}
921
922	/* execute */
923	pReq->Hdr.rc = supdrvIOCtl_LdrLoad(pDevExt, pSession, pReq);
924	return 0;
925	}
926
927	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_FREE):
928	{
929	/* validate */
930	PSUPLDRFREE pReq = (PSUPLDRFREE)pReqHdr;
931	REQ_CHECK_SIZES(SUP_IOCTL_LDR_FREE);
932
933	/* execute */
934	pReq->Hdr.rc = supdrvIOCtl_LdrFree(pDevExt, pSession, pReq);
935	return 0;
936	}
937
938	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_GET_SYMBOL):
939	{
940	/* validate */
941	PSUPLDRGETSYMBOL pReq = (PSUPLDRGETSYMBOL)pReqHdr;
942	REQ_CHECK_SIZES(SUP_IOCTL_LDR_GET_SYMBOL);
943	REQ_CHECK_EXPR(SUP_IOCTL_LDR_GET_SYMBOL, memchr(pReq->u.In.szSymbol, '\0', sizeof(pReq->u.In.szSymbol)));
944
945	/* execute */
946	pReq->Hdr.rc = supdrvIOCtl_LdrGetSymbol(pDevExt, pSession, pReq);
947	return 0;
948	}
949
950	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0(0)):
951	{
952	/* validate */
953	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
954	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_VMMR0_SIZE(0))
955	{
956	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(0), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(0));
957
958	/* execute */
959	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
960	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg);
961	else
962	pReq->Hdr.rc = VERR_WRONG_ORDER;
963	}
964	else
965	{
966	PSUPVMMR0REQHDR pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
967	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR)),
968	("SUP_IOCTL_CALL_VMMR0: cbIn=%#x < %#x\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR))));
969	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
970	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(pVMMReq->cbReq));
971
972	/* execute */
973	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
974	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg);
975	else
976	pReq->Hdr.rc = VERR_WRONG_ORDER;
977	}
978	return 0;
979	}
980
981	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_PAGING_MODE):
982	{
983	/* validate */
984	PSUPGETPAGINGMODE pReq = (PSUPGETPAGINGMODE)pReqHdr;
985	REQ_CHECK_SIZES(SUP_IOCTL_GET_PAGING_MODE);
986
987	/* execute */
988	pReq->Hdr.rc = VINF_SUCCESS;
989	pReq->u.Out.enmMode = supdrvIOCtl_GetPagingMode();
990	return 0;
991	}
992
993	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_ALLOC):
994	{
995	/* validate */
996	PSUPLOWALLOC pReq = (PSUPLOWALLOC)pReqHdr;
997	REQ_CHECK_EXPR(SUP_IOCTL_LOW_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_LOW_ALLOC_SIZE_IN);
998	REQ_CHECK_SIZES_EX(SUP_IOCTL_LOW_ALLOC, SUP_IOCTL_LOW_ALLOC_SIZE_IN, SUP_IOCTL_LOW_ALLOC_SIZE_OUT(pReq->u.In.cPages));
999
1000	/* execute */
1001	pReq->Hdr.rc = SUPR0LowAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1002	if (RT_FAILURE(pReq->Hdr.rc))
1003	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1004	return 0;
1005	}
1006
1007	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_FREE):
1008	{
1009	/* validate */
1010	PSUPLOWFREE pReq = (PSUPLOWFREE)pReqHdr;
1011	REQ_CHECK_SIZES(SUP_IOCTL_LOW_FREE);
1012
1013	/* execute */
1014	pReq->Hdr.rc = SUPR0LowFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1015	return 0;
1016	}
1017
1018	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_MAP):
1019	{
1020	/* validate */
1021	PSUPGIPMAP pReq = (PSUPGIPMAP)pReqHdr;
1022	REQ_CHECK_SIZES(SUP_IOCTL_GIP_MAP);
1023
1024	/* execute */
1025	pReq->Hdr.rc = SUPR0GipMap(pSession, &pReq->u.Out.pGipR3, &pReq->u.Out.HCPhysGip);
1026	if (RT_SUCCESS(pReq->Hdr.rc))
1027	pReq->u.Out.pGipR0 = pDevExt->pGip;
1028	return 0;
1029	}
1030
1031	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_UNMAP):
1032	{
1033	/* validate */
1034	PSUPGIPUNMAP pReq = (PSUPGIPUNMAP)pReqHdr;
1035	REQ_CHECK_SIZES(SUP_IOCTL_GIP_UNMAP);
1036
1037	/* execute */
1038	pReq->Hdr.rc = SUPR0GipUnmap(pSession);
1039	return 0;
1040	}
1041
1042	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SET_VM_FOR_FAST):
1043	{
1044	/* validate */
1045	PSUPSETVMFORFAST pReq = (PSUPSETVMFORFAST)pReqHdr;
1046	REQ_CHECK_SIZES(SUP_IOCTL_SET_VM_FOR_FAST);
1047	REQ_CHECK_EXPR_FMT( !pReq->u.In.pVMR0
1048	\|\| ( VALID_PTR(pReq->u.In.pVMR0)
1049	&& !((uintptr_t)pReq->u.In.pVMR0 & (PAGE_SIZE - 1))),
1050	("SUP_IOCTL_SET_VM_FOR_FAST: pVMR0=%p!\n", pReq->u.In.pVMR0));
1051	/* execute */
1052	pSession->pVM = pReq->u.In.pVMR0;
1053	pReq->Hdr.rc = VINF_SUCCESS;
1054	return 0;
1055	}
1056
1057	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_ALLOC):
1058	{
1059	/* validate */
1060	PSUPPAGEALLOC pReq = (PSUPPAGEALLOC)pReqHdr;
1061	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_SIZE_IN);
1062	REQ_CHECK_SIZES_EX(SUP_IOCTL_PAGE_ALLOC, SUP_IOCTL_PAGE_ALLOC_SIZE_IN, SUP_IOCTL_PAGE_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1063
1064	/* execute */
1065	pReq->Hdr.rc = SUPR0PageAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1066	if (RT_FAILURE(pReq->Hdr.rc))
1067	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1068	return 0;
1069	}
1070
1071	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_FREE):
1072	{
1073	/* validate */
1074	PSUPPAGEFREE pReq = (PSUPPAGEFREE)pReqHdr;
1075	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_FREE);
1076
1077	/* execute */
1078	pReq->Hdr.rc = SUPR0PageFree(pSession, pReq->u.In.pvR3);
1079	return 0;
1080	}
1081
1082	default:
1083	dprintf(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1084	break;
1085	}
1086	return SUPDRV_ERR_GENERAL_FAILURE;
1087	}
1088
1089
1090	/**
1091	* Register a object for reference counting.
1092	* The object is registered with one reference in the specified session.
1093	*
1094	* @returns Unique identifier on success (pointer).
1095	* All future reference must use this identifier.
1096	* @returns NULL on failure.
1097	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
1098	* @param pvUser1 The first user argument.
1099	* @param pvUser2 The second user argument.
1100	*/
1101	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
1102	{
1103	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1104	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1105	PSUPDRVOBJ pObj;
1106	PSUPDRVUSAGE pUsage;
1107
1108	/*
1109	* Validate the input.
1110	*/
1111	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
1112	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
1113	AssertPtrReturn(pfnDestructor, NULL);
1114
1115	/*
1116	* Allocate and initialize the object.
1117	*/
1118	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
1119	if (!pObj)
1120	return NULL;
1121	pObj->u32Magic = SUPDRVOBJ_MAGIC;
1122	pObj->enmType = enmType;
1123	pObj->pNext = NULL;
1124	pObj->cUsage = 1;
1125	pObj->pfnDestructor = pfnDestructor;
1126	pObj->pvUser1 = pvUser1;
1127	pObj->pvUser2 = pvUser2;
1128	pObj->CreatorUid = pSession->Uid;
1129	pObj->CreatorGid = pSession->Gid;
1130	pObj->CreatorProcess= pSession->Process;
1131	supdrvOSObjInitCreator(pObj, pSession);
1132
1133	/*
1134	* Allocate the usage record.
1135	* (We keep freed usage records around to simplity SUPR0ObjAddRef().)
1136	*/
1137	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1138
1139	pUsage = pDevExt->pUsageFree;
1140	if (pUsage)
1141	pDevExt->pUsageFree = pUsage->pNext;
1142	else
1143	{
1144	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1145	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
1146	if (!pUsage)
1147	{
1148	RTMemFree(pObj);
1149	return NULL;
1150	}
1151	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1152	}
1153
1154	/*
1155	* Insert the object and create the session usage record.
1156	*/
1157	/* The object. */
1158	pObj->pNext = pDevExt->pObjs;
1159	pDevExt->pObjs = pObj;
1160
1161	/* The session record. */
1162	pUsage->cUsage = 1;
1163	pUsage->pObj = pObj;
1164	pUsage->pNext = pSession->pUsage;
1165	dprintf(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));
1166	pSession->pUsage = pUsage;
1167
1168	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1169
1170	dprintf(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
1171	return pObj;
1172	}
1173
1174
1175	/**
1176	* Increment the reference counter for the object associating the reference
1177	* with the specified session.
1178	*
1179	* @returns IPRT status code.
1180	* @param pvObj The identifier returned by SUPR0ObjRegister().
1181	* @param pSession The session which is referencing the object.
1182	*/
1183	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
1184	{
1185	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1186	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1187	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
1188	PSUPDRVUSAGE pUsagePre;
1189	PSUPDRVUSAGE pUsage;
1190
1191	/*
1192	* Validate the input.
1193	*/
1194	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1195	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
1196	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
1197	VERR_INVALID_PARAMETER);
1198
1199	/*
1200	* Preallocate the usage record.
1201	*/
1202	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1203
1204	pUsagePre = pDevExt->pUsageFree;
1205	if (pUsagePre)
1206	pDevExt->pUsageFree = pUsagePre->pNext;
1207	else
1208	{
1209	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1210	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
1211	if (!pUsagePre)
1212	return VERR_NO_MEMORY;
1213	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1214	}
1215
1216	/*
1217	* Reference the object.
1218	*/
1219	pObj->cUsage++;
1220
1221	/*
1222	* Look for the session record.
1223	*/
1224	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
1225	{
1226	dprintf(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));
1227	if (pUsage->pObj == pObj)
1228	break;
1229	}
1230	if (pUsage)
1231	pUsage->cUsage++;
1232	else
1233	{
1234	/* create a new session record. */
1235	pUsagePre->cUsage = 1;
1236	pUsagePre->pObj = pObj;
1237	pUsagePre->pNext = pSession->pUsage;
1238	pSession->pUsage = pUsagePre;
1239	dprintf(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));
1240
1241	pUsagePre = NULL;
1242	}
1243
1244	/*
1245	* Put any unused usage record into the free list..
1246	*/
1247	if (pUsagePre)
1248	{
1249	pUsagePre->pNext = pDevExt->pUsageFree;
1250	pDevExt->pUsageFree = pUsagePre;
1251	}
1252
1253	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1254
1255	return VINF_SUCCESS;
1256	}
1257
1258
1259	/**
1260	* Decrement / destroy a reference counter record for an object.
1261	*
1262	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
1263	*
1264	* @returns IPRT status code.
1265	* @param pvObj The identifier returned by SUPR0ObjRegister().
1266	* @param pSession The session which is referencing the object.
1267	*/
1268	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
1269	{
1270	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1271	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1272	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
1273	bool fDestroy = false;
1274	PSUPDRVUSAGE pUsage;
1275	PSUPDRVUSAGE pUsagePrev;
1276
1277	/*
1278	* Validate the input.
1279	*/
1280	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1281	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
1282	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
1283	VERR_INVALID_PARAMETER);
1284
1285	/*
1286	* Acquire the spinlock and look for the usage record.
1287	*/
1288	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1289
1290	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
1291	pUsage;
1292	pUsagePrev = pUsage, pUsage = pUsage->pNext)
1293	{
1294	dprintf(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));
1295	if (pUsage->pObj == pObj)
1296	{
1297	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
1298	if (pUsage->cUsage > 1)
1299	{
1300	pObj->cUsage--;
1301	pUsage->cUsage--;
1302	}
1303	else
1304	{
1305	/*
1306	* Free the session record.
1307	*/
1308	if (pUsagePrev)
1309	pUsagePrev->pNext = pUsage->pNext;
1310	else
1311	pSession->pUsage = pUsage->pNext;
1312	pUsage->pNext = pDevExt->pUsageFree;
1313	pDevExt->pUsageFree = pUsage;
1314
1315	/* What about the object? */
1316	if (pObj->cUsage > 1)
1317	pObj->cUsage--;
1318	else
1319	{
1320	/*
1321	* Object is to be destroyed, unlink it.
1322	*/
1323	fDestroy = true;
1324	if (pDevExt->pObjs == pObj)
1325	pDevExt->pObjs = pObj->pNext;
1326	else
1327	{
1328	PSUPDRVOBJ pObjPrev;
1329	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
1330	if (pObjPrev->pNext == pObj)
1331	{
1332	pObjPrev->pNext = pObj->pNext;
1333	break;
1334	}
1335	Assert(pObjPrev);
1336	}
1337	}
1338	}
1339	break;
1340	}
1341	}
1342
1343	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1344
1345	/*
1346	* Call the destructor and free the object if required.
1347	*/
1348	if (fDestroy)
1349	{
1350	pObj->u32Magic++;
1351	if (pObj->pfnDestructor)
1352	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
1353	RTMemFree(pObj);
1354	}
1355
1356	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
1357	return pUsage ? VINF_SUCCESS : VERR_INVALID_PARAMETER;
1358	}
1359
1360	/**
1361	* Verifies that the current process can access the specified object.
1362	*
1363	* @returns The following IPRT status code:
1364	* @retval VINF_SUCCESS if access was granted.
1365	* @retval VERR_PERMISSION_DENIED if denied access.
1366	* @retval VERR_INVALID_PARAMETER if invalid parameter.
1367	*
1368	* @param pvObj The identifier returned by SUPR0ObjRegister().
1369	* @param pSession The session which wishes to access the object.
1370	* @param pszObjName Object string name. This is optional and depends on the object type.
1371	*
1372	* @remark The caller is responsible for making sure the object isn't removed while
1373	* we're inside this function. If uncertain about this, just call AddRef before calling us.
1374	*/
1375	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
1376	{
1377	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
1378	int rc;
1379
1380	/*
1381	* Validate the input.
1382	*/
1383	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1384	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
1385	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
1386	VERR_INVALID_PARAMETER);
1387
1388	/*
1389	* Check access. (returns true if a decision has been made.)
1390	*/
1391	rc = VERR_INTERNAL_ERROR;
1392	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
1393	return rc;
1394
1395	/*
1396	* Default policy is to allow the user to access his own
1397	* stuff but nothing else.
1398	*/
1399	if (pObj->CreatorUid == pSession->Uid)
1400	return VINF_SUCCESS;
1401	return VERR_PERMISSION_DENIED;
1402	}
1403
1404
1405	/**
1406	* Lock pages.
1407	*
1408	* @returns IPRT status code.
1409	* @param pSession Session to which the locked memory should be associated.
1410	* @param pvR3 Start of the memory range to lock.
1411	* This must be page aligned.
1412	* @param cb Size of the memory range to lock.
1413	* This must be page aligned.
1414	*/
1415	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
1416	{
1417	int rc;
1418	SUPDRVMEMREF Mem = {0};
1419	const size_t cb = (size_t)cPages << PAGE_SHIFT;
1420	dprintf(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
1421
1422	/*
1423	* Verify input.
1424	*/
1425	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1426	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
1427	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
1428	\|\| !pvR3)
1429	{
1430	dprintf(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
1431	return VERR_INVALID_PARAMETER;
1432	}
1433
1434	#ifdef RT_OS_WINDOWS /* A temporary hack for windows, will be removed once all ring-3 code has been cleaned up. */
1435	/* First check if we allocated it using SUPPageAlloc; if so then we don't need to lock it again */
1436	rc = supdrvPageGetPhys(pSession, pvR3, cPages, paPages);
1437	if (RT_SUCCESS(rc))
1438	return rc;
1439	#endif
1440
1441	/*
1442	* Let IPRT do the job.
1443	*/
1444	Mem.eType = MEMREF_TYPE_LOCKED;
1445	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTR0ProcHandleSelf());
1446	if (RT_SUCCESS(rc))
1447	{
1448	uint32_t iPage = cPages;
1449	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
1450	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
1451
1452	while (iPage-- > 0)
1453	{
1454	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
1455	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
1456	{
1457	AssertMsgFailed(("iPage=%d\n", iPage));
1458	rc = VERR_INTERNAL_ERROR;
1459	break;
1460	}
1461	}
1462	if (RT_SUCCESS(rc))
1463	rc = supdrvMemAdd(&Mem, pSession);
1464	if (RT_FAILURE(rc))
1465	{
1466	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
1467	AssertRC(rc2);
1468	}
1469	}
1470
1471	return rc;
1472	}
1473
1474
1475	/**
1476	* Unlocks the memory pointed to by pv.
1477	*
1478	* @returns IPRT status code.
1479	* @param pSession Session to which the memory was locked.
1480	* @param pvR3 Memory to unlock.
1481	*/
1482	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
1483	{
1484	dprintf(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
1485	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1486	#ifdef RT_OS_WINDOWS
1487	/*
1488	* Temporary hack for windows - SUPR0PageFree will unlock SUPR0PageAlloc
1489	* allocations; ignore this call.
1490	*/
1491	if (supdrvPageWasLockedByPageAlloc(pSession, pvR3))
1492	{
1493	dprintf(("Page will be unlocked in SUPR0PageFree -> ignore\n"));
1494	return VINF_SUCCESS;
1495	}
1496	#endif
1497	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
1498	}
1499
1500
1501	/**
1502	* Allocates a chunk of page aligned memory with contiguous and fixed physical
1503	* backing.
1504	*
1505	* @returns IPRT status code.
1506	* @param pSession Session data.
1507	* @param cb Number of bytes to allocate.
1508	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
1509	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
1510	* @param pHCPhys Where to put the physical address of allocated memory.
1511	*/
1512	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
1513	{
1514	int rc;
1515	SUPDRVMEMREF Mem = {0};
1516	dprintf(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
1517
1518	/*
1519	* Validate input.
1520	*/
1521	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1522	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
1523	{
1524	dprintf(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
1525	pSession, ppvR0, ppvR3, pHCPhys));
1526	return VERR_INVALID_PARAMETER;
1527
1528	}
1529	if (cPages < 1 \|\| cPages >= 256)
1530	{
1531	dprintf(("Illegal request cPages=%d, must be greater than 0 and smaller than 256\n", cPages));
1532	return VERR_INVALID_PARAMETER;
1533	}
1534
1535	/*
1536	* Let IPRT do the job.
1537	*/
1538	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
1539	if (RT_SUCCESS(rc))
1540	{
1541	int rc2;
1542	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
1543	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
1544	if (RT_SUCCESS(rc))
1545	{
1546	Mem.eType = MEMREF_TYPE_CONT;
1547	rc = supdrvMemAdd(&Mem, pSession);
1548	if (!rc)
1549	{
1550	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
1551	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
1552	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
1553	return 0;
1554	}
1555
1556	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
1557	AssertRC(rc2);
1558	}
1559	rc2 = RTR0MemObjFree(Mem.MemObj, false);
1560	AssertRC(rc2);
1561	}
1562
1563	return rc;
1564	}
1565
1566
1567	/**
1568	* Frees memory allocated using SUPR0ContAlloc().
1569	*
1570	* @returns IPRT status code.
1571	* @param pSession The session to which the memory was allocated.
1572	* @param uPtr Pointer to the memory (ring-3 or ring-0).
1573	*/
1574	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
1575	{
1576	dprintf(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
1577	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1578	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
1579	}
1580
1581
1582	/**
1583	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
1584	*
1585	* @returns IPRT status code.
1586	* @param pSession Session data.
1587	* @param cPages Number of pages to allocate.
1588	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
1589	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
1590	* @param paPages Where to put the physical addresses of allocated memory.
1591	*/
1592	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
1593	{
1594	unsigned iPage;
1595	int rc;
1596	SUPDRVMEMREF Mem = {0};
1597	dprintf(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
1598
1599	/*
1600	* Validate input.
1601	*/
1602	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1603	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
1604	{
1605	dprintf(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
1606	pSession, ppvR3, ppvR0, paPages));
1607	return VERR_INVALID_PARAMETER;
1608
1609	}
1610	if (cPages < 1 \|\| cPages > 256)
1611	{
1612	dprintf(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
1613	return VERR_INVALID_PARAMETER;
1614	}
1615
1616	/*
1617	* Let IPRT do the work.
1618	*/
1619	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
1620	if (RT_SUCCESS(rc))
1621	{
1622	int rc2;
1623	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
1624	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
1625	if (RT_SUCCESS(rc))
1626	{
1627	Mem.eType = MEMREF_TYPE_LOW;
1628	rc = supdrvMemAdd(&Mem, pSession);
1629	if (!rc)
1630	{
1631	for (iPage = 0; iPage < cPages; iPage++)
1632	{
1633	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
1634	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%VHp\n", paPages[iPage]));
1635	}
1636	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
1637	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
1638	return 0;
1639	}
1640
1641	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
1642	AssertRC(rc2);
1643	}
1644
1645	rc2 = RTR0MemObjFree(Mem.MemObj, false);
1646	AssertRC(rc2);
1647	}
1648
1649	return rc;
1650	}
1651
1652
1653	/**
1654	* Frees memory allocated using SUPR0LowAlloc().
1655	*
1656	* @returns IPRT status code.
1657	* @param pSession The session to which the memory was allocated.
1658	* @param uPtr Pointer to the memory (ring-3 or ring-0).
1659	*/
1660	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
1661	{
1662	dprintf(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
1663	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1664	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
1665	}
1666
1667
1668
1669	/**
1670	* Allocates a chunk of memory with both R0 and R3 mappings.
1671	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
1672	*
1673	* @returns IPRT status code.
1674	* @param pSession The session to associated the allocation with.
1675	* @param cb Number of bytes to allocate.
1676	* @param ppvR0 Where to store the address of the Ring-0 mapping.
1677	* @param ppvR3 Where to store the address of the Ring-3 mapping.
1678	*/
1679	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
1680	{
1681	int rc;
1682	SUPDRVMEMREF Mem = {0};
1683	dprintf(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
1684
1685	/*
1686	* Validate input.
1687	*/
1688	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1689	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
1690	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
1691	if (cb < 1 \|\| cb >= _4M)
1692	{
1693	dprintf(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
1694	return VERR_INVALID_PARAMETER;
1695	}
1696
1697	/*
1698	* Let IPRT do the work.
1699	*/
1700	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
1701	if (RT_SUCCESS(rc))
1702	{
1703	int rc2;
1704	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
1705	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
1706	if (RT_SUCCESS(rc))
1707	{
1708	Mem.eType = MEMREF_TYPE_MEM;
1709	rc = supdrvMemAdd(&Mem, pSession);
1710	if (!rc)
1711	{
1712	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
1713	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
1714	return VINF_SUCCESS;
1715	}
1716	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
1717	AssertRC(rc2);
1718	}
1719
1720	rc2 = RTR0MemObjFree(Mem.MemObj, false);
1721	AssertRC(rc2);
1722	}
1723
1724	return rc;
1725	}
1726
1727
1728	/**
1729	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
1730	*
1731	* @returns IPRT status code.
1732	* @param pSession The session to which the memory was allocated.
1733	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
1734	* @param paPages Where to store the physical addresses.
1735	*/
1736	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
1737	{
1738	PSUPDRVBUNDLE pBundle;
1739	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1740	dprintf(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
1741
1742	/*
1743	* Validate input.
1744	*/
1745	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1746	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
1747	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
1748
1749	/*
1750	* Search for the address.
1751	*/
1752	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
1753	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
1754	{
1755	if (pBundle->cUsed > 0)
1756	{
1757	unsigned i;
1758	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
1759	{
1760	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
1761	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
1762	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
1763	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
1764	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
1765	)
1766	)
1767	{
1768	const unsigned cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
1769	unsigned iPage;
1770	for (iPage = 0; iPage < cPages; iPage++)
1771	{
1772	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
1773	paPages[iPage].uReserved = 0;
1774	}
1775	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
1776	return VINF_SUCCESS;
1777	}
1778	}
1779	}
1780	}
1781	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
1782	dprintf(("Failed to find %p!!!\n", (void *)uPtr));
1783	return VERR_INVALID_PARAMETER;
1784	}
1785
1786
1787	/**
1788	* Free memory allocated by SUPR0MemAlloc().
1789	*
1790	* @returns IPRT status code.
1791	* @param pSession The session owning the allocation.
1792	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
1793	*/
1794	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
1795	{
1796	dprintf(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
1797	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1798	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
1799	}
1800
1801
1802	/**
1803	* Allocates a chunk of memory with only a R3 mappings.
1804	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
1805	*
1806	* @returns IPRT status code.
1807	* @param pSession The session to associated the allocation with.
1808	* @param cPages The number of pages to allocate.
1809	* @param ppvR3 Where to store the address of the Ring-3 mapping.
1810	* @param paPages Where to store the addresses of the pages. Optional.
1811	*/
1812	SUPR0DECL(int) SUPR0PageAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR3PTR ppvR3, PRTHCPHYS paPages)
1813	{
1814	int rc;
1815	SUPDRVMEMREF Mem = {0};
1816	dprintf(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
1817
1818	/*
1819	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
1820	*/
1821	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1822	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
1823	if (cPages < 1 \|\| cPages > (128 * _1M)/PAGE_SIZE)
1824	{
1825	dprintf(("SUPR0PageAlloc: Illegal request cb=%u; must be greater than 0 and smaller than 128MB.\n", cPages));
1826	return VERR_INVALID_PARAMETER;
1827	}
1828
1829	/*
1830	* Let IPRT do the work.
1831	*/
1832	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
1833	if (RT_SUCCESS(rc))
1834	{
1835	int rc2;
1836	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
1837	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
1838	if (RT_SUCCESS(rc))
1839	{
1840	Mem.eType = MEMREF_TYPE_LOCKED_SUP;
1841	rc = supdrvMemAdd(&Mem, pSession);
1842	if (!rc)
1843	{
1844	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
1845	if (paPages)
1846	{
1847	uint32_t iPage = cPages;
1848	while (iPage-- > 0)
1849	{
1850	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
1851	Assert(paPages[iPage] != NIL_RTHCPHYS);
1852	}
1853	}
1854	return VINF_SUCCESS;
1855	}
1856	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
1857	AssertRC(rc2);
1858	}
1859
1860	rc2 = RTR0MemObjFree(Mem.MemObj, false);
1861	AssertRC(rc2);
1862	}
1863	return rc;
1864	}
1865
1866
1867	#ifdef RT_OS_WINDOWS
1868	/**
1869	* Check if the pages were locked by SUPR0PageAlloc
1870	*
1871	* This function will be removed along with the lock/unlock hacks when
1872	* we've cleaned up the ring-3 code properly.
1873	*
1874	* @returns boolean
1875	* @param pSession The session to which the memory was allocated.
1876	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc().
1877	*/
1878	static bool supdrvPageWasLockedByPageAlloc(PSUPDRVSESSION pSession, RTR3PTR pvR3)
1879	{
1880	PSUPDRVBUNDLE pBundle;
1881	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1882	dprintf(("SUPR0PageIsLockedByPageAlloc: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
1883
1884	/*
1885	* Search for the address.
1886	*/
1887	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
1888	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
1889	{
1890	if (pBundle->cUsed > 0)
1891	{
1892	unsigned i;
1893	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
1894	{
1895	if ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED_SUP
1896	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
1897	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
1898	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
1899	{
1900	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
1901	return true;
1902	}
1903	}
1904	}
1905	}
1906	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
1907	return false;
1908	}
1909
1910
1911	/**
1912	* Get the physical addresses of memory allocated using SUPR0PageAlloc().
1913	*
1914	* This function will be removed along with the lock/unlock hacks when
1915	* we've cleaned up the ring-3 code properly.
1916	*
1917	* @returns IPRT status code.
1918	* @param pSession The session to which the memory was allocated.
1919	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc().
1920	* @param cPages Number of pages in paPages
1921	* @param paPages Where to store the physical addresses.
1922	*/
1923	static int supdrvPageGetPhys(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
1924	{
1925	PSUPDRVBUNDLE pBundle;
1926	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1927	dprintf(("supdrvPageGetPhys: pSession=%p pvR3=%p cPages=%#lx paPages=%p\n", pSession, (void *)pvR3, (long)cPages, paPages));
1928
1929	/*
1930	* Search for the address.
1931	*/
1932	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
1933	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
1934	{
1935	if (pBundle->cUsed > 0)
1936	{
1937	unsigned i;
1938	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
1939	{
1940	if ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED_SUP
1941	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
1942	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
1943	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
1944	{
1945	uint32_t iPage = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
1946	cPages = RT_MIN(iPage, cPages);
1947	for (iPage = 0; iPage < cPages; iPage++)
1948	paPages[iPage] = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
1949	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
1950	return VINF_SUCCESS;
1951	}
1952	}
1953	}
1954	}
1955	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
1956	return VERR_INVALID_PARAMETER;
1957	}
1958	#endif /* RT_OS_WINDOWS */
1959
1960
1961	/**
1962	* Free memory allocated by SUPR0PageAlloc().
1963	*
1964	* @returns IPRT status code.
1965	* @param pSession The session owning the allocation.
1966	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc().
1967	*/
1968	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
1969	{
1970	dprintf(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
1971	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1972	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED_SUP);
1973	}
1974
1975
1976	/**
1977	* Maps the GIP into userspace and/or get the physical address of the GIP.
1978	*
1979	* @returns IPRT status code.
1980	* @param pSession Session to which the GIP mapping should belong.
1981	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
1982	* @param pHCPhysGip Where to store the physical address. (optional)
1983	*
1984	* @remark There is no reference counting on the mapping, so one call to this function
1985	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
1986	* and remove the session as a GIP user.
1987	*/
1988	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
1989	{
1990	int rc = 0;
1991	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1992	RTR3PTR pGip = NIL_RTR3PTR;
1993	RTHCPHYS HCPhys = NIL_RTHCPHYS;
1994	dprintf(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
1995
1996	/*
1997	* Validate
1998	*/
1999	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2000	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
2001	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
2002
2003	RTSemFastMutexRequest(pDevExt->mtxGip);
2004	if (pDevExt->pGip)
2005	{
2006	/*
2007	* Map it?
2008	*/
2009	if (ppGipR3)
2010	{
2011	#ifdef USE_NEW_OS_INTERFACE_FOR_GIP
2012	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
2013	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
2014	RTMEM_PROT_READ, RTR0ProcHandleSelf());
2015	if (RT_SUCCESS(rc))
2016	{
2017	pGip = RTR0MemObjAddressR3(pSession->GipMapObjR3);
2018	rc = VINF_SUCCESS; /** @todo remove this and replace the !rc below with RT_SUCCESS(rc). */
2019	}
2020	#else /* !USE_NEW_OS_INTERFACE_FOR_GIP */
2021	if (!pSession->pGip)
2022	rc = supdrvOSGipMap(pSession->pDevExt, &pSession->pGip);
2023	if (!rc)
2024	pGip = (RTR3PTR)pSession->pGip;
2025	#endif /* !USE_NEW_OS_INTERFACE_FOR_GIP */
2026	}
2027
2028	/*
2029	* Get physical address.
2030	*/
2031	if (pHCPhysGip && !rc)
2032	HCPhys = pDevExt->HCPhysGip;
2033
2034	/*
2035	* Reference globally.
2036	*/
2037	if (!pSession->fGipReferenced && !rc)
2038	{
2039	pSession->fGipReferenced = 1;
2040	pDevExt->cGipUsers++;
2041	if (pDevExt->cGipUsers == 1)
2042	{
2043	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
2044	unsigned i;
2045
2046	dprintf(("SUPR0GipMap: Resumes GIP updating\n"));
2047
2048	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
2049	ASMAtomicXchgU32(&pGip->aCPUs[i].u32TransactionId, pGip->aCPUs[i].u32TransactionId & ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
2050	ASMAtomicXchgU64(&pGip->u64NanoTSLastUpdateHz, 0);
2051
2052	#ifdef USE_NEW_OS_INTERFACE_FOR_GIP
2053	rc = RTTimerStart(pDevExt->pGipTimer, 0);
2054	AssertRC(rc); rc = VINF_SUCCESS;
2055	#else
2056	supdrvOSGipResume(pDevExt);
2057	#endif
2058	}
2059	}
2060	}
2061	else
2062	{
2063	rc = SUPDRV_ERR_GENERAL_FAILURE;
2064	dprintf(("SUPR0GipMap: GIP is not available!\n"));
2065	}
2066	RTSemFastMutexRelease(pDevExt->mtxGip);
2067
2068	/*
2069	* Write returns.
2070	*/
2071	if (pHCPhysGip)
2072	*pHCPhysGip = HCPhys;
2073	if (ppGipR3)
2074	*ppGipR3 = pGip;
2075
2076	#ifdef DEBUG_DARWIN_GIP
2077	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx ppGip=%p GipMapObjR3\n", rc, (unsigned long)HCPhys, pGip, pSession->GipMapObjR3));
2078	#else
2079	dprintf(("SUPR0GipMap: returns %d pHCPhysGip=%lx ppGipR3=%p\n", rc, (unsigned long)HCPhys, (void *)(uintptr_t)pGip));
2080	#endif
2081	return rc;
2082	}
2083
2084
2085	/**
2086	* Unmaps any user mapping of the GIP and terminates all GIP access
2087	* from this session.
2088	*
2089	* @returns IPRT status code.
2090	* @param pSession Session to which the GIP mapping should belong.
2091	*/
2092	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
2093	{
2094	int rc = VINF_SUCCESS;
2095	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2096	#ifdef DEBUG_DARWIN_GIP
2097	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
2098	pSession,
2099	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
2100	pSession->GipMapObjR3));
2101	#else
2102	dprintf(("SUPR0GipUnmap: pSession=%p\n", pSession));
2103	#endif
2104	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2105
2106	RTSemFastMutexRequest(pDevExt->mtxGip);
2107
2108	/*
2109	* Unmap anything?
2110	*/
2111	#ifdef USE_NEW_OS_INTERFACE_FOR_GIP
2112	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
2113	{
2114	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
2115	AssertRC(rc);
2116	if (RT_SUCCESS(rc))
2117	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
2118	}
2119	#else
2120	if (pSession->pGip)
2121	{
2122	rc = supdrvOSGipUnmap(pDevExt, pSession->pGip);
2123	if (!rc)
2124	pSession->pGip = NULL;
2125	}
2126	#endif
2127
2128	/*
2129	* Dereference global GIP.
2130	*/
2131	if (pSession->fGipReferenced && !rc)
2132	{
2133	pSession->fGipReferenced = 0;
2134	if ( pDevExt->cGipUsers > 0
2135	&& !--pDevExt->cGipUsers)
2136	{
2137	dprintf(("SUPR0GipUnmap: Suspends GIP updating\n"));
2138	#ifdef USE_NEW_OS_INTERFACE_FOR_GIP
2139	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = 0;
2140	#else
2141	supdrvOSGipSuspend(pDevExt);
2142	#endif
2143	}
2144	}
2145
2146	RTSemFastMutexRelease(pDevExt->mtxGip);
2147
2148	return rc;
2149	}
2150
2151
2152	/**
2153	* Adds a memory object to the session.
2154	*
2155	* @returns IPRT status code.
2156	* @param pMem Memory tracking structure containing the
2157	* information to track.
2158	* @param pSession The session.
2159	*/
2160	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
2161	{
2162	PSUPDRVBUNDLE pBundle;
2163	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2164
2165	/*
2166	* Find free entry and record the allocation.
2167	*/
2168	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2169	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2170	{
2171	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
2172	{
2173	unsigned i;
2174	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2175	{
2176	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
2177	{
2178	pBundle->cUsed++;
2179	pBundle->aMem[i] = *pMem;
2180	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2181	return VINF_SUCCESS;
2182	}
2183	}
2184	AssertFailed(); /* !!this can't be happening!!! */
2185	}
2186	}
2187	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2188
2189	/*
2190	* Need to allocate a new bundle.
2191	* Insert into the last entry in the bundle.
2192	*/
2193	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
2194	if (!pBundle)
2195	return VERR_NO_MEMORY;
2196
2197	/* take last entry. */
2198	pBundle->cUsed++;
2199	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
2200
2201	/* insert into list. */
2202	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2203	pBundle->pNext = pSession->Bundle.pNext;
2204	pSession->Bundle.pNext = pBundle;
2205	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2206
2207	return VINF_SUCCESS;
2208	}
2209
2210
2211	/**
2212	* Releases a memory object referenced by pointer and type.
2213	*
2214	* @returns IPRT status code.
2215	* @param pSession Session data.
2216	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
2217	* @param eType Memory type.
2218	*/
2219	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
2220	{
2221	PSUPDRVBUNDLE pBundle;
2222	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2223
2224	/*
2225	* Validate input.
2226	*/
2227	if (!uPtr)
2228	{
2229	dprintf(("Illegal address %p\n", (void *)uPtr));
2230	return VERR_INVALID_PARAMETER;
2231	}
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 == eType
2245	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2246	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2247	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2248	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
2249	)
2250	{
2251	/* Make a copy of it and release it outside the spinlock. */
2252	SUPDRVMEMREF Mem = pBundle->aMem[i];
2253	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
2254	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
2255	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
2256	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2257
2258	if (Mem.MapObjR3)
2259	{
2260	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
2261	AssertRC(rc); /** @todo figure out how to handle this. */
2262	}
2263	if (Mem.MemObj)
2264	{
2265	int rc = RTR0MemObjFree(Mem.MemObj, false);
2266	AssertRC(rc); /** @todo figure out how to handle this. */
2267	}
2268	return VINF_SUCCESS;
2269	}
2270	}
2271	}
2272	}
2273	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2274	dprintf(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
2275	return VERR_INVALID_PARAMETER;
2276	}
2277
2278
2279	#ifdef VBOX_WITH_IDT_PATCHING
2280	/**
2281	* Install IDT for the current CPU.
2282	*
2283	* @returns One of the following IPRT status codes:
2284	* @retval VINF_SUCCESS on success.
2285	* @retval VERR_IDT_FAILED.
2286	* @retval VERR_NO_MEMORY.
2287	* @param pDevExt The device extension.
2288	* @param pSession The session data.
2289	* @param pReq The request.
2290	*/
2291	static int supdrvIOCtl_IdtInstall(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPIDTINSTALL pReq)
2292	{
2293	PSUPDRVPATCHUSAGE pUsagePre;
2294	PSUPDRVPATCH pPatchPre;
2295	RTIDTR Idtr;
2296	PSUPDRVPATCH pPatch;
2297	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2298	dprintf(("supdrvIOCtl_IdtInstall\n"));
2299
2300	/*
2301	* Preallocate entry for this CPU cause we don't wanna do
2302	* that inside the spinlock!
2303	*/
2304	pUsagePre = (PSUPDRVPATCHUSAGE)RTMemAlloc(sizeof(*pUsagePre));
2305	if (!pUsagePre)
2306	return VERR_NO_MEMORY;
2307
2308	/*
2309	* Take the spinlock and see what we need to do.
2310	*/
2311	RTSpinlockAcquireNoInts(pDevExt->Spinlock, &SpinlockTmp);
2312
2313	/* check if we already got a free patch. */
2314	if (!pDevExt->pIdtPatchesFree)
2315	{
2316	/*
2317	* Allocate a patch - outside the spinlock of course.
2318	*/
2319	RTSpinlockReleaseNoInts(pDevExt->Spinlock, &SpinlockTmp);
2320
2321	pPatchPre = (PSUPDRVPATCH)RTMemExecAlloc(sizeof(*pPatchPre));
2322	if (!pPatchPre)
2323	return VERR_NO_MEMORY;
2324
2325	RTSpinlockAcquireNoInts(pDevExt->Spinlock, &SpinlockTmp);
2326	}
2327	else
2328	{
2329	pPatchPre = pDevExt->pIdtPatchesFree;
2330	pDevExt->pIdtPatchesFree = pPatchPre->pNext;
2331	}
2332
2333	/* look for matching patch entry */
2334	ASMGetIDTR(&Idtr);
2335	pPatch = pDevExt->pIdtPatches;
2336	while (pPatch && pPatch->pvIdt != (void *)Idtr.pIdt)
2337	pPatch = pPatch->pNext;
2338
2339	if (!pPatch)
2340	{
2341	/*
2342	* Create patch.
2343	*/
2344	pPatch = supdrvIdtPatchOne(pDevExt, pPatchPre);
2345	if (pPatch)
2346	pPatchPre = NULL; /* mark as used. */
2347	}
2348	else
2349	{
2350	/*
2351	* Simply increment patch usage.
2352	*/
2353	pPatch->cUsage++;
2354	}
2355
2356	if (pPatch)
2357	{
2358	/*
2359	* Increment and add if need be the session usage record for this patch.
2360	*/
2361	PSUPDRVPATCHUSAGE pUsage = pSession->pPatchUsage;
2362	while (pUsage && pUsage->pPatch != pPatch)
2363	pUsage = pUsage->pNext;
2364
2365	if (!pUsage)
2366	{
2367	/*
2368	* Add usage record.
2369	*/
2370	pUsagePre->cUsage = 1;
2371	pUsagePre->pPatch = pPatch;
2372	pUsagePre->pNext = pSession->pPatchUsage;
2373	pSession->pPatchUsage = pUsagePre;
2374	pUsagePre = NULL; /* mark as used. */
2375	}
2376	else
2377	{
2378	/*
2379	* Increment usage count.
2380	*/
2381	pUsage->cUsage++;
2382	}
2383	}
2384
2385	/* free patch - we accumulate them for paranoid saftly reasons. */
2386	if (pPatchPre)
2387	{
2388	pPatchPre->pNext = pDevExt->pIdtPatchesFree;
2389	pDevExt->pIdtPatchesFree = pPatchPre;
2390	}
2391
2392	RTSpinlockReleaseNoInts(pDevExt->Spinlock, &SpinlockTmp);
2393
2394	/*
2395	* Free unused preallocated buffers.
2396	*/
2397	if (pUsagePre)
2398	RTMemFree(pUsagePre);
2399
2400	pReq->u.Out.u8Idt = pDevExt->u8Idt;
2401
2402	return pPatch ? VINF_SUCCESS : VERR_IDT_FAILED;
2403	}
2404
2405
2406	/**
2407	* This creates a IDT patch entry.
2408	* If the first patch being installed it'll also determin the IDT entry
2409	* to use.
2410	*
2411	* @returns pPatch on success.
2412	* @returns NULL on failure.
2413	* @param pDevExt Pointer to globals.
2414	* @param pPatch Patch entry to use.
2415	* This will be linked into SUPDRVDEVEXT::pIdtPatches on
2416	* successful return.
2417	* @remark Call must be owning the SUPDRVDEVEXT::Spinlock!
2418	*/
2419	static PSUPDRVPATCH supdrvIdtPatchOne(PSUPDRVDEVEXT pDevExt, PSUPDRVPATCH pPatch)
2420	{
2421	RTIDTR Idtr;
2422	PSUPDRVIDTE paIdt;
2423	dprintf(("supdrvIOCtl_IdtPatchOne: pPatch=%p\n", pPatch));
2424
2425	/*
2426	* Get IDT.
2427	*/
2428	ASMGetIDTR(&Idtr);
2429	paIdt = (PSUPDRVIDTE)Idtr.pIdt;
2430	/*
2431	* Recent Linux kernels can be configured to 1G user /3G kernel.
2432	*/
2433	if ((uintptr_t)paIdt < 0x40000000)
2434	{
2435	AssertMsgFailed(("bad paIdt=%p\n", paIdt));
2436	return NULL;
2437	}
2438
2439	if (!pDevExt->u8Idt)
2440	{
2441	/*
2442	* Test out the alternatives.
2443	*
2444	* At the moment we do not support chaining thus we ASSUME that one of
2445	* these 48 entries is unused (which is not a problem on Win32 and
2446	* Linux to my knowledge).
2447	*/
2448	/** @todo we MUST change this detection to try grab an entry which is NOT in use. This can be
2449	* combined with gathering info about which guest system call gates we can hook up directly. */
2450	unsigned i;
2451	uint8_t u8Idt = 0;
2452	static uint8_t au8Ints[] =
2453	{
2454	#ifdef RT_OS_WINDOWS /* We don't use 0xef and above because they are system stuff on linux (ef is IPI,
2455	* local apic timer, or some other frequently fireing thing). */
2456	0xef, 0xee, 0xed, 0xec,
2457	#endif
2458	0xeb, 0xea, 0xe9, 0xe8,
2459	0xdf, 0xde, 0xdd, 0xdc,
2460	0x7b, 0x7a, 0x79, 0x78,
2461	0xbf, 0xbe, 0xbd, 0xbc,
2462	};
2463	#if defined(RT_ARCH_AMD64) && defined(DEBUG)
2464	static int s_iWobble = 0;
2465	unsigned iMax = !(s_iWobble++ % 2) ? 0x80 : 0x100;
2466	dprintf(("IDT: Idtr=%p:%#x\n", (void *)Idtr.pIdt, (unsigned)Idtr.cbIdt));
2467	for (i = iMax - 0x80; i*16+15 < Idtr.cbIdt && i < iMax; i++)
2468	{
2469	dprintf(("%#x: %04x:%08x%04x%04x P=%d DPL=%d IST=%d Type1=%#x u32Reserved=%#x u5Reserved=%#x\n",
2470	i, paIdt[i].u16SegSel, paIdt[i].u32OffsetTop, paIdt[i].u16OffsetHigh, paIdt[i].u16OffsetLow,
2471	paIdt[i].u1Present, paIdt[i].u2DPL, paIdt[i].u3IST, paIdt[i].u5Type2,
2472	paIdt[i].u32Reserved, paIdt[i].u5Reserved));
2473	}
2474	#endif
2475	/* look for entries which are not present or otherwise unused. */
2476	for (i = 0; i < sizeof(au8Ints) / sizeof(au8Ints[0]); i++)
2477	{
2478	u8Idt = au8Ints[i];
2479	if ( u8Idt * sizeof(SUPDRVIDTE) < Idtr.cbIdt
2480	&& ( !paIdt[u8Idt].u1Present
2481	\|\| paIdt[u8Idt].u5Type2 == 0))
2482	break;
2483	u8Idt = 0;
2484	}
2485	if (!u8Idt)
2486	{
2487	/* try again, look for a compatible entry .*/
2488	for (i = 0; i < sizeof(au8Ints) / sizeof(au8Ints[0]); i++)
2489	{
2490	u8Idt = au8Ints[i];
2491	if ( u8Idt * sizeof(SUPDRVIDTE) < Idtr.cbIdt
2492	&& paIdt[u8Idt].u1Present
2493	&& paIdt[u8Idt].u5Type2 == SUPDRV_IDTE_TYPE2_INTERRUPT_GATE
2494	&& !(paIdt[u8Idt].u16SegSel & 3))
2495	break;
2496	u8Idt = 0;
2497	}
2498	if (!u8Idt)
2499	{
2500	dprintf(("Failed to find appropirate IDT entry!!\n"));
2501	return NULL;
2502	}
2503	}
2504	pDevExt->u8Idt = u8Idt;
2505	dprintf(("supdrvIOCtl_IdtPatchOne: u8Idt=%x\n", u8Idt));
2506	}
2507
2508	/*
2509	* Prepare the patch
2510	*/
2511	memset(pPatch, 0, sizeof(*pPatch));
2512	pPatch->pvIdt = paIdt;
2513	pPatch->cUsage = 1;
2514	pPatch->pIdtEntry = &paIdt[pDevExt->u8Idt];
2515	pPatch->SavedIdt = paIdt[pDevExt->u8Idt];
2516	pPatch->ChangedIdt.u16OffsetLow = (uint32_t)((uintptr_t)&pPatch->auCode[0] & 0xffff);
2517	pPatch->ChangedIdt.u16OffsetHigh = (uint32_t)((uintptr_t)&pPatch->auCode[0] >> 16);
2518	#ifdef RT_ARCH_AMD64
2519	pPatch->ChangedIdt.u32OffsetTop = (uint32_t)((uintptr_t)&pPatch->auCode[0] >> 32);
2520	#endif
2521	pPatch->ChangedIdt.u16SegSel = ASMGetCS();
2522	#ifdef RT_ARCH_AMD64
2523	pPatch->ChangedIdt.u3IST = 0;
2524	pPatch->ChangedIdt.u5Reserved = 0;
2525	#else /* x86 */
2526	pPatch->ChangedIdt.u5Reserved = 0;
2527	pPatch->ChangedIdt.u3Type1 = 0;
2528	#endif /* x86 */
2529	pPatch->ChangedIdt.u5Type2 = SUPDRV_IDTE_TYPE2_INTERRUPT_GATE;
2530	pPatch->ChangedIdt.u2DPL = 3;
2531	pPatch->ChangedIdt.u1Present = 1;
2532
2533	/*
2534	* Generate the patch code.
2535	*/
2536	{
2537	#ifdef RT_ARCH_AMD64
2538	union
2539	{
2540	uint8_t *pb;
2541	uint32_t *pu32;
2542	uint64_t *pu64;
2543	} u, uFixJmp, uFixCall, uNotNested;
2544	u.pb = &pPatch->auCode[0];
2545
2546	/* check the cookie */
2547	*u.pb++ = 0x3d; // cmp eax, GLOBALCOOKIE
2548	*u.pu32++ = pDevExt->u32Cookie;
2549
2550	*u.pb++ = 0x74; // jz @VBoxCall
2551	*u.pb++ = 2;
2552
2553	/* jump to forwarder code. */
2554	*u.pb++ = 0xeb;
2555	uFixJmp = u;
2556	*u.pb++ = 0xfe;
2557
2558	// @VBoxCall:
2559	*u.pb++ = 0x0f; // swapgs
2560	*u.pb++ = 0x01;
2561	*u.pb++ = 0xf8;
2562
2563	/*
2564	* Call VMMR0Entry
2565	* We don't have to push the arguments here, but we have top
2566	* reserve some stack space for the interrupt forwarding.
2567	*/
2568	# ifdef RT_OS_WINDOWS
2569	*u.pb++ = 0x50; // push rax ; alignment filler.
2570	*u.pb++ = 0x41; // push r8 ; uArg
2571	*u.pb++ = 0x50;
2572	*u.pb++ = 0x52; // push rdx ; uOperation
2573	*u.pb++ = 0x51; // push rcx ; pVM
2574	# else
2575	*u.pb++ = 0x51; // push rcx ; alignment filler.
2576	*u.pb++ = 0x52; // push rdx ; uArg
2577	*u.pb++ = 0x56; // push rsi ; uOperation
2578	*u.pb++ = 0x57; // push rdi ; pVM
2579	# endif
2580
2581	*u.pb++ = 0xff; // call qword [pfnVMMR0EntryInt wrt rip]
2582	*u.pb++ = 0x15;
2583	uFixCall = u;
2584	*u.pu32++ = 0;
2585
2586	*u.pb++ = 0x48; // add rsp, 20h ; remove call frame.
2587	*u.pb++ = 0x81;
2588	*u.pb++ = 0xc4;
2589	*u.pu32++ = 0x20;
2590
2591	*u.pb++ = 0x0f; // swapgs
2592	*u.pb++ = 0x01;
2593	*u.pb++ = 0xf8;
2594
2595	/* Return to R3. */
2596	uNotNested = u;
2597	*u.pb++ = 0x48; // iretq
2598	*u.pb++ = 0xcf;
2599
2600	while ((uintptr_t)u.pb & 0x7) // align 8
2601	*u.pb++ = 0xcc;
2602
2603	/* Pointer to the VMMR0Entry. */ // pfnVMMR0EntryInt dq StubVMMR0Entry
2604	*uFixCall.pu32 = (uint32_t)(u.pb - uFixCall.pb - 4); uFixCall.pb = NULL;
2605	pPatch->offVMMR0EntryFixup = (uint16_t)(u.pb - &pPatch->auCode[0]);
2606	*u.pu64++ = pDevExt->pvVMMR0 ? (uint64_t)pDevExt->pfnVMMR0EntryInt : (uint64_t)u.pb + 8;
2607
2608	/* stub entry. */ // StubVMMR0Entry:
2609	pPatch->offStub = (uint16_t)(u.pb - &pPatch->auCode[0]);
2610	*u.pb++ = 0x33; // xor eax, eax
2611	*u.pb++ = 0xc0;
2612
2613	*u.pb++ = 0x48; // dec rax
2614	*u.pb++ = 0xff;
2615	*u.pb++ = 0xc8;
2616
2617	*u.pb++ = 0xc3; // ret
2618
2619	/* forward to the original handler using a retf. */
2620	*uFixJmp.pb = (uint8_t)(u.pb - uFixJmp.pb - 1); uFixJmp.pb = NULL;
2621
2622	*u.pb++ = 0x68; // push <target cs>
2623	*u.pu32++ = !pPatch->SavedIdt.u5Type2 ? ASMGetCS() : pPatch->SavedIdt.u16SegSel;
2624
2625	*u.pb++ = 0x68; // push <low target rip>
2626	*u.pu32++ = !pPatch->SavedIdt.u5Type2
2627	? (uint32_t)(uintptr_t)uNotNested.pb
2628	: (uint32_t)pPatch->SavedIdt.u16OffsetLow
2629	\| (uint32_t)pPatch->SavedIdt.u16OffsetHigh << 16;
2630
2631	*u.pb++ = 0xc7; // mov dword [rsp + 4], <high target rip>
2632	*u.pb++ = 0x44;
2633	*u.pb++ = 0x24;
2634	*u.pb++ = 0x04;
2635	*u.pu32++ = !pPatch->SavedIdt.u5Type2
2636	? (uint32_t)((uint64_t)uNotNested.pb >> 32)
2637	: pPatch->SavedIdt.u32OffsetTop;
2638
2639	*u.pb++ = 0x48; // retf ; does this require prefix?
2640	*u.pb++ = 0xcb;
2641
2642	#else /* RT_ARCH_X86 */
2643
2644	union
2645	{
2646	uint8_t *pb;
2647	uint16_t *pu16;
2648	uint32_t *pu32;
2649	} u, uFixJmpNotNested, uFixJmp, uFixCall, uNotNested;
2650	u.pb = &pPatch->auCode[0];
2651
2652	/* check the cookie */
2653	*u.pb++ = 0x81; // cmp esi, GLOBALCOOKIE
2654	*u.pb++ = 0xfe;
2655	*u.pu32++ = pDevExt->u32Cookie;
2656
2657	*u.pb++ = 0x74; // jz VBoxCall
2658	uFixJmp = u;
2659	*u.pb++ = 0;
2660
2661	/* jump (far) to the original handler / not-nested-stub. */
2662	*u.pb++ = 0xea; // jmp far NotNested
2663	uFixJmpNotNested = u;
2664	*u.pu32++ = 0;
2665	*u.pu16++ = 0;
2666
2667	/* save selector registers. */ // VBoxCall:
2668	*uFixJmp.pb = (uint8_t)(u.pb - uFixJmp.pb - 1);
2669	*u.pb++ = 0x0f; // push fs
2670	*u.pb++ = 0xa0;
2671
2672	*u.pb++ = 0x1e; // push ds
2673
2674	*u.pb++ = 0x06; // push es
2675
2676	/* call frame */
2677	*u.pb++ = 0x51; // push ecx
2678
2679	*u.pb++ = 0x52; // push edx
2680
2681	*u.pb++ = 0x50; // push eax
2682
2683	/* load ds, es and perhaps fs before call. */
2684	*u.pb++ = 0xb8; // mov eax, KernelDS
2685	*u.pu32++ = ASMGetDS();
2686
2687	*u.pb++ = 0x8e; // mov ds, eax
2688	*u.pb++ = 0xd8;
2689
2690	*u.pb++ = 0x8e; // mov es, eax
2691	*u.pb++ = 0xc0;
2692
2693	#ifdef RT_OS_WINDOWS
2694	*u.pb++ = 0xb8; // mov eax, KernelFS
2695	*u.pu32++ = ASMGetFS();
2696
2697	*u.pb++ = 0x8e; // mov fs, eax
2698	*u.pb++ = 0xe0;
2699	#endif
2700
2701	/* do the call. */
2702	*u.pb++ = 0xe8; // call _VMMR0Entry / StubVMMR0Entry
2703	uFixCall = u;
2704	pPatch->offVMMR0EntryFixup = (uint16_t)(u.pb - &pPatch->auCode[0]);
2705	*u.pu32++ = 0xfffffffb;
2706
2707	*u.pb++ = 0x83; // add esp, 0ch ; cdecl
2708	*u.pb++ = 0xc4;
2709	*u.pb++ = 0x0c;
2710
2711	/* restore selector registers. */
2712	*u.pb++ = 0x07; // pop es
2713	//
2714	*u.pb++ = 0x1f; // pop ds
2715
2716	*u.pb++ = 0x0f; // pop fs
2717	*u.pb++ = 0xa1;
2718
2719	uNotNested = u; // NotNested:
2720	*u.pb++ = 0xcf; // iretd
2721
2722	/* the stub VMMR0Entry. */ // StubVMMR0Entry:
2723	pPatch->offStub = (uint16_t)(u.pb - &pPatch->auCode[0]);
2724	*u.pb++ = 0x33; // xor eax, eax
2725	*u.pb++ = 0xc0;
2726
2727	*u.pb++ = 0x48; // dec eax
2728
2729	*u.pb++ = 0xc3; // ret
2730
2731	/* Fixup the VMMR0Entry call. */
2732	if (pDevExt->pvVMMR0)
2733	*uFixCall.pu32 = (uint32_t)pDevExt->pfnVMMR0EntryInt - (uint32_t)(uFixCall.pu32 + 1);
2734	else
2735	*uFixCall.pu32 = (uint32_t)&pPatch->auCode[pPatch->offStub] - (uint32_t)(uFixCall.pu32 + 1);
2736
2737	/* Fixup the forward / nested far jump. */
2738	if (!pPatch->SavedIdt.u5Type2)
2739	{
2740	*uFixJmpNotNested.pu32++ = (uint32_t)uNotNested.pb;
2741	*uFixJmpNotNested.pu16++ = ASMGetCS();
2742	}
2743	else
2744	{
2745	*uFixJmpNotNested.pu32++ = ((uint32_t)pPatch->SavedIdt.u16OffsetHigh << 16) \| pPatch->SavedIdt.u16OffsetLow;
2746	*uFixJmpNotNested.pu16++ = pPatch->SavedIdt.u16SegSel;
2747	}
2748	#endif /* RT_ARCH_X86 */
2749	Assert(u.pb <= &pPatch->auCode[sizeof(pPatch->auCode)]);
2750	#if 0
2751	/* dump the patch code */
2752	dprintf(("patch code: %p\n", &pPatch->auCode[0]));
2753	for (uFixCall.pb = &pPatch->auCode[0]; uFixCall.pb < u.pb; uFixCall.pb++)
2754	dprintf(("0x%02x,\n", *uFixCall.pb));
2755	#endif
2756	}
2757
2758	/*
2759	* Install the patch.
2760	*/
2761	supdrvIdtWrite(pPatch->pIdtEntry, &pPatch->ChangedIdt);
2762	AssertMsg(!memcmp((void *)pPatch->pIdtEntry, &pPatch->ChangedIdt, sizeof(pPatch->ChangedIdt)), ("The stupid change code didn't work!!!!!\n"));
2763
2764	/*
2765	* Link in the patch.
2766	*/
2767	pPatch->pNext = pDevExt->pIdtPatches;
2768	pDevExt->pIdtPatches = pPatch;
2769
2770	return pPatch;
2771	}
2772
2773
2774	/**
2775	* Removes the sessions IDT references.
2776	* This will uninstall our IDT patch if we left unreferenced.
2777	*
2778	* @returns VINF_SUCCESS.
2779	* @param pDevExt Device globals.
2780	* @param pSession Session data.
2781	*/
2782	static int supdrvIOCtl_IdtRemoveAll(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
2783	{
2784	PSUPDRVPATCHUSAGE pUsage;
2785	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2786	dprintf(("supdrvIOCtl_IdtRemoveAll: pSession=%p\n", pSession));
2787
2788	/*
2789	* Take the spinlock.
2790	*/
2791	RTSpinlockAcquireNoInts(pDevExt->Spinlock, &SpinlockTmp);
2792
2793	/*
2794	* Walk usage list, removing patches as their usage count reaches zero.
2795	*/
2796	pUsage = pSession->pPatchUsage;
2797	while (pUsage)
2798	{
2799	if (pUsage->pPatch->cUsage <= pUsage->cUsage)
2800	supdrvIdtRemoveOne(pDevExt, pUsage->pPatch);
2801	else
2802	pUsage->pPatch->cUsage -= pUsage->cUsage;
2803
2804	/* next */
2805	pUsage = pUsage->pNext;
2806	}
2807
2808	/*
2809	* Empty the usage chain and we're done inside the spinlock.
2810	*/
2811	pUsage = pSession->pPatchUsage;
2812	pSession->pPatchUsage = NULL;
2813
2814	RTSpinlockReleaseNoInts(pDevExt->Spinlock, &SpinlockTmp);
2815
2816	/*
2817	* Free usage entries.
2818	*/
2819	while (pUsage)
2820	{
2821	void *pvToFree = pUsage;
2822	pUsage->cUsage = 0;
2823	pUsage->pPatch = NULL;
2824	pUsage = pUsage->pNext;
2825	RTMemFree(pvToFree);
2826	}
2827
2828	return VINF_SUCCESS;
2829	}
2830
2831
2832	/**
2833	* Remove one patch.
2834	*
2835	* Worker for supdrvIOCtl_IdtRemoveAll.
2836	*
2837	* @param pDevExt Device globals.
2838	* @param pPatch Patch entry to remove.
2839	* @remark Caller must own SUPDRVDEVEXT::Spinlock!
2840	*/
2841	static void supdrvIdtRemoveOne(PSUPDRVDEVEXT pDevExt, PSUPDRVPATCH pPatch)
2842	{
2843	dprintf(("supdrvIdtRemoveOne: pPatch=%p\n", pPatch));
2844
2845	pPatch->cUsage = 0;
2846
2847	/*
2848	* If the IDT entry was changed it have to kick around for ever!
2849	* This will be attempted freed again, perhaps next time we'll succeed :-)
2850	*/
2851	if (memcmp((void *)pPatch->pIdtEntry, &pPatch->ChangedIdt, sizeof(pPatch->ChangedIdt)))
2852	{
2853	AssertMsgFailed(("The hijacked IDT entry has CHANGED!!!\n"));
2854	return;
2855	}
2856
2857	/*
2858	* Unlink it.
2859	*/
2860	if (pDevExt->pIdtPatches != pPatch)
2861	{
2862	PSUPDRVPATCH pPatchPrev = pDevExt->pIdtPatches;
2863	while (pPatchPrev)
2864	{
2865	if (pPatchPrev->pNext == pPatch)
2866	{
2867	pPatchPrev->pNext = pPatch->pNext;
2868	break;
2869	}
2870	pPatchPrev = pPatchPrev->pNext;
2871	}
2872	Assert(!pPatchPrev);
2873	}
2874	else
2875	pDevExt->pIdtPatches = pPatch->pNext;
2876	pPatch->pNext = NULL;
2877
2878
2879	/*
2880	* Verify and restore the IDT.
2881	*/
2882	AssertMsg(!memcmp((void *)pPatch->pIdtEntry, &pPatch->ChangedIdt, sizeof(pPatch->ChangedIdt)), ("The hijacked IDT entry has CHANGED!!!\n"));
2883	supdrvIdtWrite(pPatch->pIdtEntry, &pPatch->SavedIdt);
2884	AssertMsg(!memcmp((void *)pPatch->pIdtEntry, &pPatch->SavedIdt, sizeof(pPatch->SavedIdt)), ("The hijacked IDT entry has CHANGED!!!\n"));
2885
2886	/*
2887	* Put it in the free list.
2888	* (This free list stuff is to calm my paranoia.)
2889	*/
2890	pPatch->pvIdt = NULL;
2891	pPatch->pIdtEntry = NULL;
2892
2893	pPatch->pNext = pDevExt->pIdtPatchesFree;
2894	pDevExt->pIdtPatchesFree = pPatch;
2895	}
2896
2897
2898	/**
2899	* Write to an IDT entry.
2900	*
2901	* @param pvIdtEntry Where to write.
2902	* @param pNewIDTEntry What to write.
2903	*/
2904	static void supdrvIdtWrite(volatile void pvIdtEntry, const SUPDRVIDTE pNewIDTEntry)
2905	{
2906	RTUINTREG uCR0;
2907	RTUINTREG uFlags;
2908
2909	/*
2910	* On SMP machines (P4 hyperthreading included) we must preform a
2911	* 64-bit locked write when updating the IDT entry.
2912	*
2913	* The F00F bugfix for linux (and probably other OSes) causes
2914	* the IDT to be pointing to an readonly mapping. We get around that
2915	* by temporarily turning of WP. Since we're inside a spinlock at this
2916	* point, interrupts are disabled and there isn't any way the WP bit
2917	* flipping can cause any trouble.
2918	*/
2919
2920	/* Save & Clear interrupt flag; Save & clear WP. */
2921	uFlags = ASMGetFlags();
2922	ASMSetFlags(uFlags & ~(RTUINTREG)(1 << 9)); /X86_EFL_IF/
2923	Assert(!(ASMGetFlags() & (1 << 9)));
2924	uCR0 = ASMGetCR0();
2925	ASMSetCR0(uCR0 & ~(RTUINTREG)(1 << 16)); /X86_CR0_WP/
2926
2927	/* Update IDT Entry */
2928	#ifdef RT_ARCH_AMD64
2929	ASMAtomicXchgU128((volatile uint128_t )pvIdtEntry, (uint128_t *)(uintptr_t)pNewIDTEntry);
2930	#else
2931	ASMAtomicXchgU64((volatile uint64_t )pvIdtEntry, (uint64_t *)(uintptr_t)pNewIDTEntry);
2932	#endif
2933
2934	/* Restore CR0 & Flags */
2935	ASMSetCR0(uCR0);
2936	ASMSetFlags(uFlags);
2937	}
2938	#endif /* VBOX_WITH_IDT_PATCHING */
2939
2940
2941	/**
2942	* Opens an image. If it's the first time it's opened the call must upload
2943	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
2944	*
2945	* This is the 1st step of the loading.
2946	*
2947	* @returns IPRT status code.
2948	* @param pDevExt Device globals.
2949	* @param pSession Session data.
2950	* @param pReq The open request.
2951	*/
2952	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
2953	{
2954	PSUPDRVLDRIMAGE pImage;
2955	unsigned cb;
2956	void *pv;
2957	dprintf(("supdrvIOCtl_LdrOpen: szName=%s cbImage=%d\n", pReq->u.In.szName, pReq->u.In.cbImage));
2958
2959	/*
2960	* Check if we got an instance of the image already.
2961	*/
2962	RTSemFastMutexRequest(pDevExt->mtxLdr);
2963	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
2964	{
2965	if (!strcmp(pImage->szName, pReq->u.In.szName))
2966	{
2967	pImage->cUsage++;
2968	pReq->u.Out.pvImageBase = pImage->pvImage;
2969	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
2970	supdrvLdrAddUsage(pSession, pImage);
2971	RTSemFastMutexRelease(pDevExt->mtxLdr);
2972	return VINF_SUCCESS;
2973	}
2974	}
2975	/* (not found - add it!) */
2976
2977	/*
2978	* Allocate memory.
2979	*/
2980	cb = pReq->u.In.cbImage + sizeof(SUPDRVLDRIMAGE) + 31;
2981	pv = RTMemExecAlloc(cb);
2982	if (!pv)
2983	{
2984	RTSemFastMutexRelease(pDevExt->mtxLdr);
2985	return VERR_NO_MEMORY;
2986	}
2987
2988	/*
2989	* Setup and link in the LDR stuff.
2990	*/
2991	pImage = (PSUPDRVLDRIMAGE)pv;
2992	pImage->pvImage = RT_ALIGN_P(pImage + 1, 32);
2993	pImage->cbImage = pReq->u.In.cbImage;
2994	pImage->pfnModuleInit = NULL;
2995	pImage->pfnModuleTerm = NULL;
2996	pImage->uState = SUP_IOCTL_LDR_OPEN;
2997	pImage->cUsage = 1;
2998	strcpy(pImage->szName, pReq->u.In.szName);
2999
3000	pImage->pNext = pDevExt->pLdrImages;
3001	pDevExt->pLdrImages = pImage;
3002
3003	supdrvLdrAddUsage(pSession, pImage);
3004
3005	pReq->u.Out.pvImageBase = pImage->pvImage;
3006	pReq->u.Out.fNeedsLoading = true;
3007	RTSemFastMutexRelease(pDevExt->mtxLdr);
3008	return VINF_SUCCESS;
3009	}
3010
3011
3012	/**
3013	* Loads the image bits.
3014	*
3015	* This is the 2nd step of the loading.
3016	*
3017	* @returns IPRT status code.
3018	* @param pDevExt Device globals.
3019	* @param pSession Session data.
3020	* @param pReq The request.
3021	*/
3022	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
3023	{
3024	PSUPDRVLDRUSAGE pUsage;
3025	PSUPDRVLDRIMAGE pImage;
3026	int rc;
3027	dprintf(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImage=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImage));
3028
3029	/*
3030	* Find the ldr image.
3031	*/
3032	RTSemFastMutexRequest(pDevExt->mtxLdr);
3033	pUsage = pSession->pLdrUsage;
3034	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3035	pUsage = pUsage->pNext;
3036	if (!pUsage)
3037	{
3038	RTSemFastMutexRelease(pDevExt->mtxLdr);
3039	dprintf(("SUP_IOCTL_LDR_LOAD: couldn't find image!\n"));
3040	return VERR_INVALID_HANDLE;
3041	}
3042	pImage = pUsage->pImage;
3043	if (pImage->cbImage != pReq->u.In.cbImage)
3044	{
3045	RTSemFastMutexRelease(pDevExt->mtxLdr);
3046	dprintf(("SUP_IOCTL_LDR_LOAD: image size mismatch!! %d(prep) != %d(load)\n", pImage->cbImage, pReq->u.In.cbImage));
3047	return VERR_INVALID_HANDLE;
3048	}
3049	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
3050	{
3051	unsigned uState = pImage->uState;
3052	RTSemFastMutexRelease(pDevExt->mtxLdr);
3053	if (uState != SUP_IOCTL_LDR_LOAD)
3054	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
3055	return SUPDRV_ERR_ALREADY_LOADED;
3056	}
3057	switch (pReq->u.In.eEPType)
3058	{
3059	case SUPLDRLOADEP_NOTHING:
3060	break;
3061	case SUPLDRLOADEP_VMMR0:
3062	if ( !pReq->u.In.EP.VMMR0.pvVMMR0
3063	\|\| !pReq->u.In.EP.VMMR0.pvVMMR0EntryInt
3064	\|\| !pReq->u.In.EP.VMMR0.pvVMMR0EntryFast
3065	\|\| !pReq->u.In.EP.VMMR0.pvVMMR0EntryEx)
3066	{
3067	RTSemFastMutexRelease(pDevExt->mtxLdr);
3068	dprintf(("NULL pointer: pvVMMR0=%p pvVMMR0EntryInt=%p pvVMMR0EntryFast=%p pvVMMR0EntryEx=%p!\n",
3069	pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
3070	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx));
3071	return VERR_INVALID_PARAMETER;
3072	}
3073	if ( (uintptr_t)pReq->u.In.EP.VMMR0.pvVMMR0EntryInt - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage
3074	\|\| (uintptr_t)pReq->u.In.EP.VMMR0.pvVMMR0EntryFast - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage
3075	\|\| (uintptr_t)pReq->u.In.EP.VMMR0.pvVMMR0EntryEx - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage)
3076	{
3077	RTSemFastMutexRelease(pDevExt->mtxLdr);
3078	dprintf(("Out of range (%p LB %#x): pvVMMR0EntryInt=%p, pvVMMR0EntryFast=%p or pvVMMR0EntryEx=%p is NULL!\n",
3079	pImage->pvImage, pReq->u.In.cbImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
3080	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx));
3081	return VERR_INVALID_PARAMETER;
3082	}
3083	break;
3084	default:
3085	RTSemFastMutexRelease(pDevExt->mtxLdr);
3086	dprintf(("Invalid eEPType=%d\n", pReq->u.In.eEPType));
3087	return VERR_INVALID_PARAMETER;
3088	}
3089	if ( pReq->u.In.pfnModuleInit
3090	&& (uintptr_t)pReq->u.In.pfnModuleInit - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage)
3091	{
3092	RTSemFastMutexRelease(pDevExt->mtxLdr);
3093	dprintf(("SUP_IOCTL_LDR_LOAD: pfnModuleInit=%p is outside the image (%p %d bytes)\n",
3094	pReq->u.In.pfnModuleInit, pImage->pvImage, pReq->u.In.cbImage));
3095	return VERR_INVALID_PARAMETER;
3096	}
3097	if ( pReq->u.In.pfnModuleTerm
3098	&& (uintptr_t)pReq->u.In.pfnModuleTerm - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage)
3099	{
3100	RTSemFastMutexRelease(pDevExt->mtxLdr);
3101	dprintf(("SUP_IOCTL_LDR_LOAD: pfnModuleTerm=%p is outside the image (%p %d bytes)\n",
3102	pReq->u.In.pfnModuleTerm, pImage->pvImage, pReq->u.In.cbImage));
3103	return VERR_INVALID_PARAMETER;
3104	}
3105
3106	/*
3107	* Copy the memory.
3108	*/
3109	/* no need to do try/except as this is a buffered request. */
3110	memcpy(pImage->pvImage, &pReq->u.In.achImage[0], pImage->cbImage);
3111	pImage->uState = SUP_IOCTL_LDR_LOAD;
3112	pImage->pfnModuleInit = pReq->u.In.pfnModuleInit;
3113	pImage->pfnModuleTerm = pReq->u.In.pfnModuleTerm;
3114	pImage->offSymbols = pReq->u.In.offSymbols;
3115	pImage->cSymbols = pReq->u.In.cSymbols;
3116	pImage->offStrTab = pReq->u.In.offStrTab;
3117	pImage->cbStrTab = pReq->u.In.cbStrTab;
3118
3119	/*
3120	* Update any entry points.
3121	*/
3122	switch (pReq->u.In.eEPType)
3123	{
3124	default:
3125	case SUPLDRLOADEP_NOTHING:
3126	rc = VINF_SUCCESS;
3127	break;
3128	case SUPLDRLOADEP_VMMR0:
3129	rc = supdrvLdrSetR0EP(pDevExt, pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
3130	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
3131	break;
3132	}
3133
3134	/*
3135	* On success call the module initialization.
3136	*/
3137	dprintf(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
3138	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
3139	{
3140	dprintf(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
3141	rc = pImage->pfnModuleInit();
3142	if (rc && pDevExt->pvVMMR0 == pImage->pvImage)
3143	supdrvLdrUnsetR0EP(pDevExt);
3144	}
3145
3146	if (rc)
3147	pImage->uState = SUP_IOCTL_LDR_OPEN;
3148
3149	RTSemFastMutexRelease(pDevExt->mtxLdr);
3150	return rc;
3151	}
3152
3153
3154	/**
3155	* Frees a previously loaded (prep'ed) image.
3156	*
3157	* @returns IPRT status code.
3158	* @param pDevExt Device globals.
3159	* @param pSession Session data.
3160	* @param pReq The request.
3161	*/
3162	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
3163	{
3164	int rc;
3165	PSUPDRVLDRUSAGE pUsagePrev;
3166	PSUPDRVLDRUSAGE pUsage;
3167	PSUPDRVLDRIMAGE pImage;
3168	dprintf(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
3169
3170	/*
3171	* Find the ldr image.
3172	*/
3173	RTSemFastMutexRequest(pDevExt->mtxLdr);
3174	pUsagePrev = NULL;
3175	pUsage = pSession->pLdrUsage;
3176	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3177	{
3178	pUsagePrev = pUsage;
3179	pUsage = pUsage->pNext;
3180	}
3181	if (!pUsage)
3182	{
3183	RTSemFastMutexRelease(pDevExt->mtxLdr);
3184	dprintf(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
3185	return VERR_INVALID_HANDLE;
3186	}
3187
3188	/*
3189	* Check if we can remove anything.
3190	*/
3191	rc = VINF_SUCCESS;
3192	pImage = pUsage->pImage;
3193	if (pImage->cUsage <= 1 \|\| pUsage->cUsage <= 1)
3194	{
3195	/*
3196	* Check if there are any objects with destructors in the image, if
3197	* so leave it for the session cleanup routine so we get a chance to
3198	* clean things up in the right order and not leave them all dangling.
3199	*/
3200	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3201	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
3202	if (pImage->cUsage <= 1)
3203	{
3204	PSUPDRVOBJ pObj;
3205	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
3206	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImage))
3207	{
3208	rc = VERR_SHARING_VIOLATION; /** @todo VERR_DANGLING_OBJECTS */
3209	break;
3210	}
3211	}
3212	else
3213	{
3214	PSUPDRVUSAGE pGenUsage;
3215	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
3216	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImage))
3217	{
3218	rc = VERR_SHARING_VIOLATION; /** @todo VERR_DANGLING_OBJECTS */
3219	break;
3220	}
3221	}
3222	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
3223	if (rc == VINF_SUCCESS)
3224	{
3225	/* unlink it */
3226	if (pUsagePrev)
3227	pUsagePrev->pNext = pUsage->pNext;
3228	else
3229	pSession->pLdrUsage = pUsage->pNext;
3230
3231	/* free it */
3232	pUsage->pImage = NULL;
3233	pUsage->pNext = NULL;
3234	RTMemFree(pUsage);
3235
3236	/*
3237	* Derefrence the image.
3238	*/
3239	if (pImage->cUsage <= 1)
3240	supdrvLdrFree(pDevExt, pImage);
3241	else
3242	pImage->cUsage--;
3243	}
3244	}
3245	else
3246	{
3247	/*
3248	* Dereference both image and usage.
3249	*/
3250	pImage->cUsage--;
3251	pUsage->cUsage--;
3252	}
3253
3254	RTSemFastMutexRelease(pDevExt->mtxLdr);
3255	return VINF_SUCCESS;
3256	}
3257
3258
3259	/**
3260	* Gets the address of a symbol in an open image.
3261	*
3262	* @returns 0 on success.
3263	* @returns SUPDRV_ERR_* on failure.
3264	* @param pDevExt Device globals.
3265	* @param pSession Session data.
3266	* @param pReq The request buffer.
3267	*/
3268	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
3269	{
3270	PSUPDRVLDRIMAGE pImage;
3271	PSUPDRVLDRUSAGE pUsage;
3272	uint32_t i;
3273	PSUPLDRSYM paSyms;
3274	const char *pchStrings;
3275	const size_t cbSymbol = strlen(pReq->u.In.szSymbol) + 1;
3276	void *pvSymbol = NULL;
3277	int rc = VERR_GENERAL_FAILURE;
3278	dprintf2(("supdrvIOCtl_LdrGetSymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
3279
3280	/*
3281	* Find the ldr image.
3282	*/
3283	RTSemFastMutexRequest(pDevExt->mtxLdr);
3284	pUsage = pSession->pLdrUsage;
3285	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3286	pUsage = pUsage->pNext;
3287	if (!pUsage)
3288	{
3289	RTSemFastMutexRelease(pDevExt->mtxLdr);
3290	dprintf(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
3291	return VERR_INVALID_HANDLE;
3292	}
3293	pImage = pUsage->pImage;
3294	if (pImage->uState != SUP_IOCTL_LDR_LOAD)
3295	{
3296	unsigned uState = pImage->uState;
3297	RTSemFastMutexRelease(pDevExt->mtxLdr);
3298	dprintf(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", uState, uState)); NOREF(uState);
3299	return VERR_ALREADY_LOADED;
3300	}
3301
3302	/*
3303	* Search the symbol string.
3304	*/
3305	pchStrings = (const char )((uint8_t )pImage->pvImage + pImage->offStrTab);
3306	paSyms = (PSUPLDRSYM)((uint8_t *)pImage->pvImage + pImage->offSymbols);
3307	for (i = 0; i < pImage->cSymbols; i++)
3308	{
3309	if ( paSyms[i].offSymbol < pImage->cbImage /* paranoia */
3310	&& paSyms[i].offName + cbSymbol <= pImage->cbStrTab
3311	&& !memcmp(pchStrings + paSyms[i].offName, pReq->u.In.szSymbol, cbSymbol))
3312	{
3313	pvSymbol = (uint8_t *)pImage->pvImage + paSyms[i].offSymbol;
3314	rc = VINF_SUCCESS;
3315	break;
3316	}
3317	}
3318	RTSemFastMutexRelease(pDevExt->mtxLdr);
3319	pReq->u.Out.pvSymbol = pvSymbol;
3320	return rc;
3321	}
3322
3323
3324	/**
3325	* Updates the IDT patches to point to the specified VMM R0 entry
3326	* point (i.e. VMMR0Enter()).
3327	*
3328	* @returns IPRT status code.
3329	* @param pDevExt Device globals.
3330	* @param pSession Session data.
3331	* @param pVMMR0 VMMR0 image handle.
3332	* @param pvVMMR0EntryInt VMMR0EntryInt address.
3333	* @param pvVMMR0EntryFast VMMR0EntryFast address.
3334	* @param pvVMMR0EntryEx VMMR0EntryEx address.
3335	* @remark Caller must own the loader mutex.
3336	*/
3337	static int supdrvLdrSetR0EP(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx)
3338	{
3339	int rc = VINF_SUCCESS;
3340	dprintf(("supdrvLdrSetR0EP pvVMMR0=%p pvVMMR0EntryInt=%p\n", pvVMMR0, pvVMMR0EntryInt));
3341
3342
3343	/*
3344	* Check if not yet set.
3345	*/
3346	if (!pDevExt->pvVMMR0)
3347	{
3348	#ifdef VBOX_WITH_IDT_PATCHING
3349	PSUPDRVPATCH pPatch;
3350	#endif
3351
3352	/*
3353	* Set it and update IDT patch code.
3354	*/
3355	pDevExt->pvVMMR0 = pvVMMR0;
3356	pDevExt->pfnVMMR0EntryInt = pvVMMR0EntryInt;
3357	pDevExt->pfnVMMR0EntryFast = pvVMMR0EntryFast;
3358	pDevExt->pfnVMMR0EntryEx = pvVMMR0EntryEx;
3359	#ifdef VBOX_WITH_IDT_PATCHING
3360	for (pPatch = pDevExt->pIdtPatches; pPatch; pPatch = pPatch->pNext)
3361	{
3362	# ifdef RT_ARCH_AMD64
3363	ASMAtomicXchgU64((volatile uint64_t *)&pPatch->auCode[pPatch->offVMMR0EntryFixup], (uint64_t)pvVMMR0);
3364	# else /* RT_ARCH_X86 */
3365	ASMAtomicXchgU32((volatile uint32_t *)&pPatch->auCode[pPatch->offVMMR0EntryFixup],
3366	(uint32_t)pvVMMR0 - (uint32_t)&pPatch->auCode[pPatch->offVMMR0EntryFixup + 4]);
3367	# endif
3368	}
3369	#endif /* VBOX_WITH_IDT_PATCHING */
3370	}
3371	else
3372	{
3373	/*
3374	* Return failure or success depending on whether the values match or not.
3375	*/
3376	if ( pDevExt->pvVMMR0 != pvVMMR0
3377	\|\| (void *)pDevExt->pfnVMMR0EntryInt != pvVMMR0EntryInt
3378	\|\| (void *)pDevExt->pfnVMMR0EntryFast != pvVMMR0EntryFast
3379	\|\| (void *)pDevExt->pfnVMMR0EntryEx != pvVMMR0EntryEx)
3380	{
3381	AssertMsgFailed(("SUP_IOCTL_LDR_SETR0EP: Already set pointing to a different module!\n"));
3382	rc = VERR_INVALID_PARAMETER;
3383	}
3384	}
3385	return rc;
3386	}
3387
3388
3389	/**
3390	* Unsets the R0 entry point installed by supdrvLdrSetR0EP.
3391	*
3392	* @param pDevExt Device globals.
3393	*/
3394	static void supdrvLdrUnsetR0EP(PSUPDRVDEVEXT pDevExt)
3395	{
3396	#ifdef VBOX_WITH_IDT_PATCHING
3397	PSUPDRVPATCH pPatch;
3398	#endif
3399
3400	pDevExt->pvVMMR0 = NULL;
3401	pDevExt->pfnVMMR0EntryInt = NULL;
3402	pDevExt->pfnVMMR0EntryFast = NULL;
3403	pDevExt->pfnVMMR0EntryEx = NULL;
3404
3405	#ifdef VBOX_WITH_IDT_PATCHING
3406	for (pPatch = pDevExt->pIdtPatches; pPatch; pPatch = pPatch->pNext)
3407	{
3408	# ifdef RT_ARCH_AMD64
3409	ASMAtomicXchgU64((volatile uint64_t *)&pPatch->auCode[pPatch->offVMMR0EntryFixup],
3410	(uint64_t)&pPatch->auCode[pPatch->offStub]);
3411	# else /* RT_ARCH_X86 */
3412	ASMAtomicXchgU32((volatile uint32_t *)&pPatch->auCode[pPatch->offVMMR0EntryFixup],
3413	(uint32_t)&pPatch->auCode[pPatch->offStub] - (uint32_t)&pPatch->auCode[pPatch->offVMMR0EntryFixup + 4]);
3414	# endif
3415	}
3416	#endif /* VBOX_WITH_IDT_PATCHING */
3417	}
3418
3419
3420	/**
3421	* Adds a usage reference in the specified session of an image.
3422	*
3423	* @param pSession Session in question.
3424	* @param pImage Image which the session is using.
3425	*/
3426	static void supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage)
3427	{
3428	PSUPDRVLDRUSAGE pUsage;
3429	dprintf(("supdrvLdrAddUsage: pImage=%p\n", pImage));
3430
3431	/*
3432	* Referenced it already?
3433	*/
3434	pUsage = pSession->pLdrUsage;
3435	while (pUsage)
3436	{
3437	if (pUsage->pImage == pImage)
3438	{
3439	pUsage->cUsage++;
3440	return;
3441	}
3442	pUsage = pUsage->pNext;
3443	}
3444
3445	/*
3446	* Allocate new usage record.
3447	*/
3448	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
3449	Assert(pUsage);
3450	if (pUsage)
3451	{
3452	pUsage->cUsage = 1;
3453	pUsage->pImage = pImage;
3454	pUsage->pNext = pSession->pLdrUsage;
3455	pSession->pLdrUsage = pUsage;
3456	}
3457	/* ignore errors... */
3458	}
3459
3460
3461	/**
3462	* Frees a load image.
3463	*
3464	* @param pDevExt Pointer to device extension.
3465	* @param pImage Pointer to the image we're gonna free.
3466	* This image must exit!
3467	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
3468	*/
3469	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
3470	{
3471	PSUPDRVLDRIMAGE pImagePrev;
3472	dprintf(("supdrvLdrFree: pImage=%p\n", pImage));
3473
3474	/* find it - arg. should've used doubly linked list. */
3475	Assert(pDevExt->pLdrImages);
3476	pImagePrev = NULL;
3477	if (pDevExt->pLdrImages != pImage)
3478	{
3479	pImagePrev = pDevExt->pLdrImages;
3480	while (pImagePrev->pNext != pImage)
3481	pImagePrev = pImagePrev->pNext;
3482	Assert(pImagePrev->pNext == pImage);
3483	}
3484
3485	/* unlink */
3486	if (pImagePrev)
3487	pImagePrev->pNext = pImage->pNext;
3488	else
3489	pDevExt->pLdrImages = pImage->pNext;
3490
3491	/* check if this is VMMR0.r0 and fix the Idt patches if it is. */
3492	if (pDevExt->pvVMMR0 == pImage->pvImage)
3493	supdrvLdrUnsetR0EP(pDevExt);
3494
3495	/* check for objects with destructors in this image. (Shouldn't happen.) */
3496	if (pDevExt->pObjs)
3497	{
3498	unsigned cObjs = 0;
3499	PSUPDRVOBJ pObj;
3500	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3501	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
3502	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
3503	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImage))
3504	{
3505	pObj->pfnDestructor = NULL;
3506	cObjs++;
3507	}
3508	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
3509	if (cObjs)
3510	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
3511	}
3512
3513	/* call termination function if fully loaded. */
3514	if ( pImage->pfnModuleTerm
3515	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
3516	{
3517	dprintf(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
3518	pImage->pfnModuleTerm();
3519	}
3520
3521	/* free the image */
3522	pImage->cUsage = 0;
3523	pImage->pNext = 0;
3524	pImage->uState = SUP_IOCTL_LDR_FREE;
3525	RTMemExecFree(pImage);
3526	}
3527
3528
3529	/**
3530	* Gets the current paging mode of the CPU and stores in in pOut.
3531	*/
3532	static SUPPAGINGMODE supdrvIOCtl_GetPagingMode(void)
3533	{
3534	SUPPAGINGMODE enmMode;
3535
3536	RTUINTREG cr0 = ASMGetCR0();
3537	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
3538	enmMode = SUPPAGINGMODE_INVALID;
3539	else
3540	{
3541	RTUINTREG cr4 = ASMGetCR4();
3542	uint32_t fNXEPlusLMA = 0;
3543	if (cr4 & X86_CR4_PAE)
3544	{
3545	uint32_t fAmdFeatures = ASMCpuId_EDX(0x80000001);
3546	if (fAmdFeatures & (X86_CPUID_AMD_FEATURE_EDX_NX \| X86_CPUID_AMD_FEATURE_EDX_LONG_MODE))
3547	{
3548	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
3549	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
3550	fNXEPlusLMA \|= RT_BIT(0);
3551	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
3552	fNXEPlusLMA \|= RT_BIT(1);
3553	}
3554	}
3555
3556	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
3557	{
3558	case 0:
3559	enmMode = SUPPAGINGMODE_32_BIT;
3560	break;
3561
3562	case X86_CR4_PGE:
3563	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
3564	break;
3565
3566	case X86_CR4_PAE:
3567	enmMode = SUPPAGINGMODE_PAE;
3568	break;
3569
3570	case X86_CR4_PAE \| RT_BIT(0):
3571	enmMode = SUPPAGINGMODE_PAE_NX;
3572	break;
3573
3574	case X86_CR4_PAE \| X86_CR4_PGE:
3575	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3576	break;
3577
3578	case X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3579	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3580	break;
3581
3582	case RT_BIT(1) \| X86_CR4_PAE:
3583	enmMode = SUPPAGINGMODE_AMD64;
3584	break;
3585
3586	case RT_BIT(1) \| X86_CR4_PAE \| RT_BIT(0):
3587	enmMode = SUPPAGINGMODE_AMD64_NX;
3588	break;
3589
3590	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
3591	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
3592	break;
3593
3594	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3595	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
3596	break;
3597
3598	default:
3599	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
3600	enmMode = SUPPAGINGMODE_INVALID;
3601	break;
3602	}
3603	}
3604	return enmMode;
3605	}
3606
3607
3608	#ifdef USE_NEW_OS_INTERFACE_FOR_GIP
3609	/**
3610	* Creates the GIP.
3611	*
3612	* @returns negative errno.
3613	* @param pDevExt Instance data. GIP stuff may be updated.
3614	*/
3615	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
3616	{
3617	PSUPGLOBALINFOPAGE pGip;
3618	RTHCPHYS HCPhysGip;
3619	uint32_t u32SystemResolution;
3620	uint32_t u32Interval;
3621	int rc;
3622
3623	dprintf(("supdrvGipCreate:\n"));
3624
3625	/* assert order */
3626	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
3627	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
3628	Assert(!pDevExt->pGipTimer);
3629
3630	/*
3631	* Allocate a suitable page with a default kernel mapping.
3632	*/
3633	rc = RTR0MemObjAllocLow(&pDevExt->GipMemObj, PAGE_SIZE, false);
3634	if (RT_FAILURE(rc))
3635	{
3636	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
3637	return rc;
3638	}
3639	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
3640	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
3641
3642	/*
3643	* Try bump up the system timer resolution.
3644	* The more interrupts the better...
3645	*/
3646	if ( RT_SUCCESS(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
3647	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
3648	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 3906250 /* 256 HZ */, &u32SystemResolution))
3649	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 4000000 /* 250 HZ */, &u32SystemResolution))
3650	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 7812500 /* 128 HZ */, &u32SystemResolution))
3651	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity(10000000 /* 100 HZ */, &u32SystemResolution))
3652	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity(15625000 /* 64 HZ */, &u32SystemResolution))
3653	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity(31250000 /* 32 HZ */, &u32SystemResolution))
3654	)
3655	{
3656	Assert(RTTimerGetSystemGranularity() <= u32SystemResolution);
3657	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
3658	}
3659
3660	/*
3661	* Find a reasonable update interval, something close to 10ms would be nice,
3662	* and create a recurring timer.
3663	*/
3664	u32Interval = u32SystemResolution = RTTimerGetSystemGranularity();
3665	while (u32Interval < 10000000 /* 10 ms */)
3666	u32Interval += u32SystemResolution;
3667
3668	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0, supdrvGipTimer, pDevExt);
3669	if (RT_FAILURE(rc))
3670	{
3671	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %RU32 ns interval. rc=%d\n", u32Interval, rc));
3672	Assert(!pDevExt->pGipTimer);
3673	supdrvGipDestroy(pDevExt);
3674	return rc;
3675	}
3676
3677	/*
3678	* We're good.
3679	*/
3680	supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), 1000000000 / u32Interval /=Hz/);
3681	return VINF_SUCCESS;
3682	}
3683
3684
3685	/**
3686	* Terminates the GIP.
3687	*
3688	* @param pDevExt Instance data. GIP stuff may be updated.
3689	*/
3690	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
3691	{
3692	int rc;
3693	#ifdef DEBUG_DARWIN_GIP
3694	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
3695	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
3696	pDevExt->pGipTimer, pDevExt->GipMemObj));
3697	#endif
3698
3699	/*
3700	* Invalid the GIP data.
3701	*/
3702	if (pDevExt->pGip)
3703	{
3704	supdrvGipTerm(pDevExt->pGip);
3705	pDevExt->pGip = NULL;
3706	}
3707
3708	/*
3709	* Destroy the timer and free the GIP memory object.
3710	*/
3711	if (pDevExt->pGipTimer)
3712	{
3713	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
3714	pDevExt->pGipTimer = NULL;
3715	}
3716
3717	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
3718	{
3719	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
3720	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
3721	}
3722
3723	/*
3724	* Finally, release the system timer resolution request if one succeeded.
3725	*/
3726	if (pDevExt->u32SystemTimerGranularityGrant)
3727	{
3728	rc = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
3729	pDevExt->u32SystemTimerGranularityGrant = 0;
3730	}
3731	}
3732
3733
3734	/**
3735	* Timer callback function.
3736	* @param pTimer The timer.
3737	* @param pvUser The device extension.
3738	*/
3739	static DECLCALLBACK(void) supdrvGipTimer(PRTTIMER pTimer, void *pvUser)
3740	{
3741	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
3742	supdrvGipUpdate(pDevExt->pGip, RTTimeSystemNanoTS());
3743	}
3744	#endif /* USE_NEW_OS_INTERFACE_FOR_GIP */
3745
3746
3747	/**
3748	* Initializes the GIP data.
3749	*
3750	* @returns IPRT status code.
3751	* @param pDevExt Pointer to the device instance data.
3752	* @param pGip Pointer to the read-write kernel mapping of the GIP.
3753	* @param HCPhys The physical address of the GIP.
3754	* @param u64NanoTS The current nanosecond timestamp.
3755	* @param uUpdateHz The update freqence.
3756	*/
3757	int VBOXCALL supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys, uint64_t u64NanoTS, unsigned uUpdateHz)
3758	{
3759	unsigned i;
3760	#ifdef DEBUG_DARWIN_GIP
3761	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz));
3762	#else
3763	dprintf(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz));
3764	#endif
3765
3766	/*
3767	* Initialize the structure.
3768	*/
3769	memset(pGip, 0, PAGE_SIZE);
3770	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
3771	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
3772	pGip->u32Mode = supdrvGipDeterminTscMode();
3773	pGip->u32UpdateHz = uUpdateHz;
3774	pGip->u32UpdateIntervalNS = 1000000000 / uUpdateHz;
3775	pGip->u64NanoTSLastUpdateHz = u64NanoTS;
3776
3777	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
3778	{
3779	pGip->aCPUs[i].u32TransactionId = 2;
3780	pGip->aCPUs[i].u64NanoTS = u64NanoTS;
3781	pGip->aCPUs[i].u64TSC = ASMReadTSC();
3782
3783	/*
3784	* We don't know the following values until we've executed updates.
3785	* So, we'll just insert very high values.
3786	*/
3787	pGip->aCPUs[i].u64CpuHz = _4G + 1;
3788	pGip->aCPUs[i].u32UpdateIntervalTSC = _2G / 4;
3789	pGip->aCPUs[i].au32TSCHistory[0] = _2G / 4;
3790	pGip->aCPUs[i].au32TSCHistory[1] = _2G / 4;
3791	pGip->aCPUs[i].au32TSCHistory[2] = _2G / 4;
3792	pGip->aCPUs[i].au32TSCHistory[3] = _2G / 4;
3793	pGip->aCPUs[i].au32TSCHistory[4] = _2G / 4;
3794	pGip->aCPUs[i].au32TSCHistory[5] = _2G / 4;
3795	pGip->aCPUs[i].au32TSCHistory[6] = _2G / 4;
3796	pGip->aCPUs[i].au32TSCHistory[7] = _2G / 4;
3797	}
3798
3799	/*
3800	* Link it to the device extension.
3801	*/
3802	pDevExt->pGip = pGip;
3803	pDevExt->HCPhysGip = HCPhys;
3804	pDevExt->cGipUsers = 0;
3805
3806	return VINF_SUCCESS;
3807	}
3808
3809
3810	/**
3811	* Determin the GIP TSC mode.
3812	*
3813	* @returns The most suitable TSC mode.
3814	*/
3815	static SUPGIPMODE supdrvGipDeterminTscMode(void)
3816	{
3817	#ifndef USE_NEW_OS_INTERFACE_FOR_GIP
3818	/*
3819	* The problem here is that AMD processors with power management features
3820	* may easily end up with different TSCs because the CPUs or even cores
3821	* on the same physical chip run at different frequencies to save power.
3822	*
3823	* It is rumoured that this will be corrected with Barcelona and it's
3824	* expected that this will be indicated by the TscInvariant bit in
3825	* cpuid(0x80000007). So, the "difficult" bit here is to correctly
3826	* identify the older CPUs which don't do different frequency and
3827	* can be relied upon to have somewhat uniform TSC between the cpus.
3828	*/
3829	if (supdrvOSGetCPUCount() > 1)
3830	{
3831	uint32_t uEAX, uEBX, uECX, uEDX;
3832
3833	/* Permit user users override. */
3834	if (supdrvOSGetForcedAsyncTscMode())
3835	return SUPGIPMODE_ASYNC_TSC;
3836
3837	/* Check for "AuthenticAMD" */
3838	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
3839	if (uEAX >= 1 && uEBX == 0x68747541 && uECX == 0x444d4163 && uEDX == 0x69746e65)
3840	{
3841	/* Check for APM support and that TscInvariant is cleared. */
3842	ASMCpuId(0x80000000, &uEAX, &uEBX, &uECX, &uEDX);
3843	if (uEAX >= 0x80000007)
3844	{
3845	ASMCpuId(0x80000007, &uEAX, &uEBX, &uECX, &uEDX);
3846	if ( !(uEDX & RT_BIT(8))/* TscInvariant */
3847	&& (uEDX & 0x3e)) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
3848	return SUPGIPMODE_ASYNC_TSC;
3849	}
3850	}
3851	}
3852	#endif
3853	return SUPGIPMODE_SYNC_TSC;
3854	}
3855
3856
3857	/**
3858	* Invalidates the GIP data upon termination.
3859	*
3860	* @param pGip Pointer to the read-write kernel mapping of the GIP.
3861	*/
3862	void VBOXCALL supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
3863	{
3864	unsigned i;
3865	pGip->u32Magic = 0;
3866	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
3867	{
3868	pGip->aCPUs[i].u64NanoTS = 0;
3869	pGip->aCPUs[i].u64TSC = 0;
3870	pGip->aCPUs[i].iTSCHistoryHead = 0;
3871	}
3872	}
3873
3874
3875	/**
3876	* Worker routine for supdrvGipUpdate and supdrvGipUpdatePerCpu that
3877	* updates all the per cpu data except the transaction id.
3878	*
3879	* @param pGip The GIP.
3880	* @param pGipCpu Pointer to the per cpu data.
3881	* @param u64NanoTS The current time stamp.
3882	*/
3883	static void supdrvGipDoUpdateCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
3884	{
3885	uint64_t u64TSC;
3886	uint64_t u64TSCDelta;
3887	uint32_t u32UpdateIntervalTSC;
3888	uint32_t u32UpdateIntervalTSCSlack;
3889	unsigned iTSCHistoryHead;
3890	uint64_t u64CpuHz;
3891
3892	/*
3893	* Update the NanoTS.
3894	*/
3895	ASMAtomicXchgU64(&pGipCpu->u64NanoTS, u64NanoTS);
3896
3897	/*
3898	* Calc TSC delta.
3899	*/
3900	/** @todo validate the NanoTS delta, don't trust the OS to call us when it should... */
3901	u64TSC = ASMReadTSC();
3902	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
3903	ASMAtomicXchgU64(&pGipCpu->u64TSC, u64TSC);
3904
3905	if (u64TSCDelta >> 32)
3906	{
3907	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
3908	pGipCpu->cErrors++;
3909	}
3910
3911	/*
3912	* TSC History.
3913	*/
3914	Assert(ELEMENTS(pGipCpu->au32TSCHistory) == 8);
3915
3916	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
3917	ASMAtomicXchgU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
3918	ASMAtomicXchgU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
3919
3920	/*
3921	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
3922	*/
3923	if (pGip->u32UpdateHz >= 1000)
3924	{
3925	uint32_t u32;
3926	u32 = pGipCpu->au32TSCHistory[0];
3927	u32 += pGipCpu->au32TSCHistory[1];
3928	u32 += pGipCpu->au32TSCHistory[2];
3929	u32 += pGipCpu->au32TSCHistory[3];
3930	u32 >>= 2;
3931	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
3932	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
3933	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
3934	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
3935	u32UpdateIntervalTSC >>= 2;
3936	u32UpdateIntervalTSC += u32;
3937	u32UpdateIntervalTSC >>= 1;
3938
3939	/* Value choosen for a 2GHz Athlon64 running linux 2.6.10/11, . */
3940	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
3941	}
3942	else if (pGip->u32UpdateHz >= 90)
3943	{
3944	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
3945	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
3946	u32UpdateIntervalTSC >>= 1;
3947
3948	/* value choosen on a 2GHz thinkpad running windows */
3949	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
3950	}
3951	else
3952	{
3953	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
3954
3955	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
3956	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
3957	}
3958	ASMAtomicXchgU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
3959
3960	/*
3961	* CpuHz.
3962	*/
3963	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, pGip->u32UpdateHz);
3964	ASMAtomicXchgU64(&pGipCpu->u64CpuHz, u64CpuHz);
3965	}
3966
3967
3968	/**
3969	* Updates the GIP.
3970	*
3971	* @param pGip Pointer to the GIP.
3972	* @param u64NanoTS The current nanosecond timesamp.
3973	*/
3974	void VBOXCALL supdrvGipUpdate(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS)
3975	{
3976	/*
3977	* Determin the relevant CPU data.
3978	*/
3979	PSUPGIPCPU pGipCpu;
3980	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
3981	pGipCpu = &pGip->aCPUs[0];
3982	else
3983	{
3984	unsigned iCpu = ASMGetApicId();
3985	if (RT_LIKELY(iCpu >= RT_ELEMENTS(pGip->aCPUs)))
3986	return;
3987	pGipCpu = &pGip->aCPUs[iCpu];
3988	}
3989
3990	/*
3991	* Start update transaction.
3992	*/
3993	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
3994	{
3995	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
3996	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
3997	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
3998	pGipCpu->cErrors++;
3999	return;
4000	}
4001
4002	/*
4003	* Recalc the update frequency every 0x800th time.
4004	*/
4005	if (!(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
4006	{
4007	if (pGip->u64NanoTSLastUpdateHz)
4008	{
4009	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
4010	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
4011	uint32_t u32UpdateHz = (uint32_t)((UINT64_C(1000000000) * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
4012	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
4013	{
4014	ASMAtomicXchgU32(&pGip->u32UpdateHz, u32UpdateHz);
4015	ASMAtomicXchgU32(&pGip->u32UpdateIntervalNS, 1000000000 / u32UpdateHz);
4016	}
4017	#endif
4018	}
4019	ASMAtomicXchgU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS);
4020	}
4021
4022	/*
4023	* Update the data.
4024	*/
4025	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS);
4026
4027	/*
4028	* Complete transaction.
4029	*/
4030	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
4031	}
4032
4033
4034	/**
4035	* Updates the per cpu GIP data for the calling cpu.
4036	*
4037	* @param pGip Pointer to the GIP.
4038	* @param u64NanoTS The current nanosecond timesamp.
4039	* @param iCpu The CPU index.
4040	*/
4041	void VBOXCALL supdrvGipUpdatePerCpu(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, unsigned iCpu)
4042	{
4043	PSUPGIPCPU pGipCpu;
4044
4045	if (RT_LIKELY(iCpu <= RT_ELEMENTS(pGip->aCPUs)))
4046	{
4047	pGipCpu = &pGip->aCPUs[iCpu];
4048
4049	/*
4050	* Start update transaction.
4051	*/
4052	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
4053	{
4054	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
4055	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
4056	pGipCpu->cErrors++;
4057	return;
4058	}
4059
4060	/*
4061	* Update the data.
4062	*/
4063	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS);
4064
4065	/*
4066	* Complete transaction.
4067	*/
4068	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
4069	}
4070	}
4071
4072
4073	#ifndef DEBUG /** @todo change #ifndef DEBUG -> #ifdef LOG_ENABLED */
4074	/**
4075	* Stub function for non-debug builds.
4076	*/
4077	RTDECL(PRTLOGGER) RTLogDefaultInstance(void)
4078	{
4079	return NULL;
4080	}
4081
4082	RTDECL(PRTLOGGER) RTLogRelDefaultInstance(void)
4083	{
4084	return NULL;
4085	}
4086
4087	/**
4088	* Stub function for non-debug builds.
4089	*/
4090	RTDECL(int) RTLogSetDefaultInstanceThread(PRTLOGGER pLogger, uintptr_t uKey)
4091	{
4092	return 0;
4093	}
4094
4095	/**
4096	* Stub function for non-debug builds.
4097	*/
4098	RTDECL(void) RTLogLogger(PRTLOGGER pLogger, void pvCallerRet, const char pszFormat, ...)
4099	{
4100	}
4101
4102	/**
4103	* Stub function for non-debug builds.
4104	*/
4105	RTDECL(void) RTLogLoggerEx(PRTLOGGER pLogger, unsigned fFlags, unsigned iGroup, const char *pszFormat, ...)
4106	{
4107	}
4108
4109	/**
4110	* Stub function for non-debug builds.
4111	*/
4112	RTDECL(void) RTLogLoggerExV(PRTLOGGER pLogger, unsigned fFlags, unsigned iGroup, const char *pszFormat, va_list args)
4113	{
4114	}
4115
4116	/**
4117	* Stub function for non-debug builds.
4118	*/
4119	RTDECL(void) RTLogPrintf(const char *pszFormat, ...)
4120	{
4121	}
4122
4123	/**
4124	* Stub function for non-debug builds.
4125	*/
4126	RTDECL(void) RTLogPrintfV(const char *pszFormat, va_list args)
4127	{
4128	}
4129	#endif /* !DEBUG */
4130

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

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDRVShared.c@ 5999

Download in other formats: