SUPDRVShared.c@ 4985

Last change on this file since 4985 was 4971, checked in by vboxsync, 17 years ago
GVM.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 137.8 KB

Line
1	/* $Revision: 4971 $ */
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 as published by the Free Software Foundation,
13	* in version 2 as it comes in the "COPYING" file of the VirtualBox OSE
14	* distribution. VirtualBox OSE is distributed in the hope that it will
15	* be useful, but WITHOUT ANY WARRANTY of any kind.
16	*/
17
18
19	/*******************************************************************************
20	* Header Files *
21	*******************************************************************************/
22	#include "SUPDRV.h"
23	#ifndef PAGE_SHIFT
24	# include <iprt/param.h>
25	#endif
26	#include <iprt/alloc.h>
27	#include <iprt/semaphore.h>
28	#include <iprt/spinlock.h>
29	#include <iprt/thread.h>
30	#include <iprt/process.h>
31	#include <iprt/log.h>
32
33
34	/*******************************************************************************
35	* Defined Constants And Macros *
36	*******************************************************************************/
37	/* from x86.h - clashes with linux thus this duplication */
38	#undef X86_CR0_PG
39	#define X86_CR0_PG BIT(31)
40	#undef X86_CR0_PE
41	#define X86_CR0_PE BIT(0)
42	#undef X86_CPUID_AMD_FEATURE_EDX_NX
43	#define X86_CPUID_AMD_FEATURE_EDX_NX BIT(20)
44	#undef MSR_K6_EFER
45	#define MSR_K6_EFER 0xc0000080
46	#undef MSR_K6_EFER_NXE
47	#define MSR_K6_EFER_NXE BIT(11)
48	#undef MSR_K6_EFER_LMA
49	#define MSR_K6_EFER_LMA BIT(10)
50	#undef X86_CR4_PGE
51	#define X86_CR4_PGE BIT(7)
52	#undef X86_CR4_PAE
53	#define X86_CR4_PAE BIT(5)
54	#undef X86_CPUID_AMD_FEATURE_EDX_LONG_MODE
55	#define X86_CPUID_AMD_FEATURE_EDX_LONG_MODE BIT(29)
56
57
58	/** The frequency by which we recalculate the u32UpdateHz and
59	* u32UpdateIntervalNS GIP members. The value must be a power of 2. */
60	#define GIP_UPDATEHZ_RECALC_FREQ 0x800
61
62	/**
63	* Validates a session pointer.
64	*
65	* @returns true/false accordingly.
66	* @param pSession The session.
67	*/
68	#define SUP_IS_SESSION_VALID(pSession) \
69	( VALID_PTR(pSession) \
70	&& pSession->u32Cookie == BIRD_INV)
71
72
73	/*******************************************************************************
74	* Global Variables *
75	*******************************************************************************/
76	/**
77	* Array of the R0 SUP API.
78	*/
79	static SUPFUNC g_aFunctions[] =
80	{
81	/* name function */
82	{ "SUPR0ObjRegister", (void *)SUPR0ObjRegister },
83	{ "SUPR0ObjAddRef", (void *)SUPR0ObjAddRef },
84	{ "SUPR0ObjRelease", (void *)SUPR0ObjRelease },
85	{ "SUPR0ObjVerifyAccess", (void *)SUPR0ObjVerifyAccess },
86	{ "SUPR0LockMem", (void *)SUPR0LockMem },
87	{ "SUPR0UnlockMem", (void *)SUPR0UnlockMem },
88	{ "SUPR0ContAlloc", (void *)SUPR0ContAlloc },
89	{ "SUPR0ContFree", (void *)SUPR0ContFree },
90	{ "SUPR0LowAlloc", (void *)SUPR0LowAlloc },
91	{ "SUPR0LowFree", (void *)SUPR0LowFree },
92	{ "SUPR0MemAlloc", (void *)SUPR0MemAlloc },
93	{ "SUPR0MemGetPhys", (void *)SUPR0MemGetPhys },
94	{ "SUPR0MemFree", (void *)SUPR0MemFree },
95	{ "SUPR0PageAlloc", (void *)SUPR0PageAlloc },
96	{ "SUPR0PageFree", (void *)SUPR0PageFree },
97	{ "SUPR0Printf", (void *)SUPR0Printf },
98	{ "RTMemAlloc", (void *)RTMemAlloc },
99	{ "RTMemAllocZ", (void *)RTMemAllocZ },
100	{ "RTMemFree", (void *)RTMemFree },
101	/* These doesn't work yet on linux - use fast mutexes!
102	{ "RTSemMutexCreate", (void *)RTSemMutexCreate },
103	{ "RTSemMutexRequest", (void *)RTSemMutexRequest },
104	{ "RTSemMutexRelease", (void *)RTSemMutexRelease },
105	{ "RTSemMutexDestroy", (void *)RTSemMutexDestroy },
106	*/
107	{ "RTSemFastMutexCreate", (void *)RTSemFastMutexCreate },
108	{ "RTSemFastMutexDestroy", (void *)RTSemFastMutexDestroy },
109	{ "RTSemFastMutexRequest", (void *)RTSemFastMutexRequest },
110	{ "RTSemFastMutexRelease", (void *)RTSemFastMutexRelease },
111	{ "RTSemEventCreate", (void *)RTSemEventCreate },
112	{ "RTSemEventSignal", (void *)RTSemEventSignal },
113	{ "RTSemEventWait", (void *)RTSemEventWait },
114	{ "RTSemEventDestroy", (void *)RTSemEventDestroy },
115	{ "RTSpinlockCreate", (void *)RTSpinlockCreate },
116	{ "RTSpinlockDestroy", (void *)RTSpinlockDestroy },
117	{ "RTSpinlockAcquire", (void *)RTSpinlockAcquire },
118	{ "RTSpinlockRelease", (void *)RTSpinlockRelease },
119	{ "RTSpinlockAcquireNoInts", (void *)RTSpinlockAcquireNoInts },
120	{ "RTSpinlockReleaseNoInts", (void *)RTSpinlockReleaseNoInts },
121	{ "RTThreadNativeSelf", (void *)RTThreadNativeSelf },
122	{ "RTThreadSleep", (void *)RTThreadSleep },
123	{ "RTThreadYield", (void *)RTThreadYield },
124	#if 0 /* Thread APIs, Part 2. */
125	{ "RTThreadSelf", (void *)RTThreadSelf },
126	{ "RTThreadCreate", (void *)RTThreadCreate },
127	{ "RTThreadGetNative", (void *)RTThreadGetNative },
128	{ "RTThreadWait", (void *)RTThreadWait },
129	{ "RTThreadWaitNoResume", (void *)RTThreadWaitNoResume },
130	{ "RTThreadGetName", (void *)RTThreadGetName },
131	{ "RTThreadSelfName", (void *)RTThreadSelfName },
132	{ "RTThreadGetType", (void *)RTThreadGetType },
133	{ "RTThreadUserSignal", (void *)RTThreadUserSignal },
134	{ "RTThreadUserReset", (void *)RTThreadUserReset },
135	{ "RTThreadUserWait", (void *)RTThreadUserWait },
136	{ "RTThreadUserWaitNoResume", (void *)RTThreadUserWaitNoResume },
137	#endif
138	{ "RTLogDefaultInstance", (void *)RTLogDefaultInstance },
139	{ "RTLogRelDefaultInstance", (void *)RTLogRelDefaultInstance },
140	{ "RTLogSetDefaultInstanceThread", (void *)RTLogSetDefaultInstanceThread },
141	{ "RTLogLogger", (void *)RTLogLogger },
142	{ "RTLogLoggerEx", (void *)RTLogLoggerEx },
143	{ "RTLogLoggerExV", (void *)RTLogLoggerExV },
144	{ "RTLogPrintf", (void *)RTLogPrintf },
145	{ "RTLogPrintfV", (void *)RTLogPrintfV },
146	{ "AssertMsg1", (void *)AssertMsg1 },
147	{ "AssertMsg2", (void *)AssertMsg2 },
148	};
149
150
151	/*******************************************************************************
152	* Internal Functions *
153	*******************************************************************************/
154	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession);
155	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType);
156	#ifdef VBOX_WITH_IDT_PATCHING
157	static int supdrvIOCtl_IdtInstall(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPIDTINSTALL pReq);
158	static PSUPDRVPATCH supdrvIdtPatchOne(PSUPDRVDEVEXT pDevExt, PSUPDRVPATCH pPatch);
159	static int supdrvIOCtl_IdtRemoveAll(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession);
160	static void supdrvIdtRemoveOne(PSUPDRVDEVEXT pDevExt, PSUPDRVPATCH pPatch);
161	static void supdrvIdtWrite(volatile void pvIdtEntry, const SUPDRVIDTE pNewIDTEntry);
162	#endif /* VBOX_WITH_IDT_PATCHING */
163	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq);
164	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq);
165	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq);
166	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq);
167	static int supdrvLdrSetR0EP(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx);
168	static void supdrvLdrUnsetR0EP(PSUPDRVDEVEXT pDevExt);
169	static void supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage);
170	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage);
171	static SUPPAGINGMODE supdrvIOCtl_GetPagingMode(void);
172	static SUPGIPMODE supdrvGipDeterminTscMode(void);
173	#ifdef RT_OS_WINDOWS
174	static int supdrvPageGetPhys(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages);
175	static bool supdrvPageWasLockedByPageAlloc(PSUPDRVSESSION pSession, RTR3PTR pvR3);
176	#endif
177	#ifdef USE_NEW_OS_INTERFACE_FOR_GIP
178	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt);
179	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt);
180	static DECLCALLBACK(void) supdrvGipTimer(PRTTIMER pTimer, void *pvUser);
181	#endif
182
183
184	/**
185	* Initializes the device extentsion structure.
186	*
187	* @returns IPRT status code.
188	* @param pDevExt The device extension to initialize.
189	*/
190	int VBOXCALL supdrvInitDevExt(PSUPDRVDEVEXT pDevExt)
191	{
192	/*
193	* Initialize it.
194	*/
195	int rc;
196	memset(pDevExt, 0, sizeof(*pDevExt));
197	rc = RTSpinlockCreate(&pDevExt->Spinlock);
198	if (!rc)
199	{
200	rc = RTSemFastMutexCreate(&pDevExt->mtxLdr);
201	if (!rc)
202	{
203	rc = RTSemFastMutexCreate(&pDevExt->mtxGip);
204	if (!rc)
205	{
206	#ifdef USE_NEW_OS_INTERFACE_FOR_GIP
207	rc = supdrvGipCreate(pDevExt);
208	if (RT_SUCCESS(rc))
209	{
210	pDevExt->u32Cookie = BIRD; /** @todo make this random? */
211	return VINF_SUCCESS;
212	}
213	#else
214	pDevExt->u32Cookie = BIRD;
215	return VINF_SUCCESS;
216	#endif
217	}
218	RTSemFastMutexDestroy(pDevExt->mtxLdr);
219	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
220	}
221	RTSpinlockDestroy(pDevExt->Spinlock);
222	pDevExt->Spinlock = NIL_RTSPINLOCK;
223	}
224	return rc;
225	}
226
227
228	/**
229	* Delete the device extension (e.g. cleanup members).
230	*
231	* @param pDevExt The device extension to delete.
232	*/
233	void VBOXCALL supdrvDeleteDevExt(PSUPDRVDEVEXT pDevExt)
234	{
235	#ifdef VBOX_WITH_IDT_PATCHING
236	PSUPDRVPATCH pPatch;
237	#endif
238	PSUPDRVOBJ pObj;
239	PSUPDRVUSAGE pUsage;
240
241	/*
242	* Kill mutexes and spinlocks.
243	*/
244	RTSemFastMutexDestroy(pDevExt->mtxGip);
245	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
246	RTSemFastMutexDestroy(pDevExt->mtxLdr);
247	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
248	RTSpinlockDestroy(pDevExt->Spinlock);
249	pDevExt->Spinlock = NIL_RTSPINLOCK;
250
251	/*
252	* Free lists.
253	*/
254	#ifdef VBOX_WITH_IDT_PATCHING
255	/* patches */
256	/** @todo make sure we don't uninstall patches which has been patched by someone else. */
257	pPatch = pDevExt->pIdtPatchesFree;
258	pDevExt->pIdtPatchesFree = NULL;
259	while (pPatch)
260	{
261	void *pvFree = pPatch;
262	pPatch = pPatch->pNext;
263	RTMemExecFree(pvFree);
264	}
265	#endif /* VBOX_WITH_IDT_PATCHING */
266
267	/* objects. */
268	pObj = pDevExt->pObjs;
269	#if !defined(DEBUG_bird) \|\| !defined(RT_OS_LINUX) /* breaks unloading, temporary, remove me! */
270	Assert(!pObj); /* (can trigger on forced unloads) */
271	#endif
272	pDevExt->pObjs = NULL;
273	while (pObj)
274	{
275	void *pvFree = pObj;
276	pObj = pObj->pNext;
277	RTMemFree(pvFree);
278	}
279
280	/* usage records. */
281	pUsage = pDevExt->pUsageFree;
282	pDevExt->pUsageFree = NULL;
283	while (pUsage)
284	{
285	void *pvFree = pUsage;
286	pUsage = pUsage->pNext;
287	RTMemFree(pvFree);
288	}
289
290	#ifdef USE_NEW_OS_INTERFACE_FOR_GIP
291	/* kill the GIP */
292	supdrvGipDestroy(pDevExt);
293	#endif
294	}
295
296
297	/**
298	* Create session.
299	*
300	* @returns IPRT status code.
301	* @param pDevExt Device extension.
302	* @param ppSession Where to store the pointer to the session data.
303	*/
304	int VBOXCALL supdrvCreateSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION *ppSession)
305	{
306	/*
307	* Allocate memory for the session data.
308	*/
309	int rc = VERR_NO_MEMORY;
310	PSUPDRVSESSION pSession = ppSession = (PSUPDRVSESSION)RTMemAllocZ(sizeof(pSession));
311	if (pSession)
312	{
313	/* Initialize session data. */
314	rc = RTSpinlockCreate(&pSession->Spinlock);
315	if (!rc)
316	{
317	Assert(pSession->Spinlock != NIL_RTSPINLOCK);
318	pSession->pDevExt = pDevExt;
319	pSession->u32Cookie = BIRD_INV;
320	/*pSession->pLdrUsage = NULL;
321	pSession->pPatchUsage = NULL;
322	pSession->pUsage = NULL;
323	pSession->pGip = NULL;
324	pSession->fGipReferenced = false;
325	pSession->Bundle.cUsed = 0 */
326
327	dprintf(("Created session %p initial cookie=%#x\n", pSession, pSession->u32Cookie));
328	return VINF_SUCCESS;
329	}
330
331	RTMemFree(pSession);
332	*ppSession = NULL;
333	}
334
335	dprintf(("Failed to create spinlock, rc=%d!\n", rc));
336	return rc;
337	}
338
339
340	/**
341	* Shared code for cleaning up a session.
342	*
343	* @param pDevExt Device extension.
344	* @param pSession Session data.
345	* This data will be freed by this routine.
346	*/
347	void VBOXCALL supdrvCloseSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
348	{
349	/*
350	* Cleanup the session first.
351	*/
352	supdrvCleanupSession(pDevExt, pSession);
353
354	/*
355	* Free the rest of the session stuff.
356	*/
357	RTSpinlockDestroy(pSession->Spinlock);
358	pSession->Spinlock = NIL_RTSPINLOCK;
359	pSession->pDevExt = NULL;
360	RTMemFree(pSession);
361	dprintf2(("supdrvCloseSession: returns\n"));
362	}
363
364
365	/**
366	* Shared code for cleaning up a session (but not quite freeing it).
367	*
368	* This is primarily intended for MAC OS X where we have to clean up the memory
369	* stuff before the file handle is closed.
370	*
371	* @param pDevExt Device extension.
372	* @param pSession Session data.
373	* This data will be freed by this routine.
374	*/
375	void VBOXCALL supdrvCleanupSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
376	{
377	PSUPDRVBUNDLE pBundle;
378	dprintf(("supdrvCleanupSession: pSession=%p\n", pSession));
379
380	/*
381	* Remove logger instances related to this session.
382	* (This assumes the dprintf and dprintf2 macros doesn't use the normal logging.)
383	*/
384	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pSession);
385
386	#ifdef VBOX_WITH_IDT_PATCHING
387	/*
388	* Uninstall any IDT patches installed for this session.
389	*/
390	supdrvIOCtl_IdtRemoveAll(pDevExt, pSession);
391	#endif
392
393	/*
394	* Release object references made in this session.
395	* In theory there should be noone racing us in this session.
396	*/
397	dprintf2(("release objects - start\n"));
398	if (pSession->pUsage)
399	{
400	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
401	PSUPDRVUSAGE pUsage;
402	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
403
404	while ((pUsage = pSession->pUsage) != NULL)
405	{
406	PSUPDRVOBJ pObj = pUsage->pObj;
407	pSession->pUsage = pUsage->pNext;
408
409	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
410	if (pUsage->cUsage < pObj->cUsage)
411	{
412	pObj->cUsage -= pUsage->cUsage;
413	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
414	}
415	else
416	{
417	/* Destroy the object and free the record. */
418	if (pDevExt->pObjs == pObj)
419	pDevExt->pObjs = pObj->pNext;
420	else
421	{
422	PSUPDRVOBJ pObjPrev;
423	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
424	if (pObjPrev->pNext == pObj)
425	{
426	pObjPrev->pNext = pObj->pNext;
427	break;
428	}
429	Assert(pObjPrev);
430	}
431	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
432
433	if (pObj->pfnDestructor)
434	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
435	RTMemFree(pObj);
436	}
437
438	/* free it and continue. */
439	RTMemFree(pUsage);
440
441	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
442	}
443
444	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
445	AssertMsg(!pSession->pUsage, ("Some buster reregistered an object during desturction!\n"));
446	}
447	dprintf2(("release objects - done\n"));
448
449	/*
450	* Release memory allocated in the session.
451	*
452	* We do not serialize this as we assume that the application will
453	* not allocated memory while closing the file handle object.
454	*/
455	dprintf2(("freeing memory:\n"));
456	pBundle = &pSession->Bundle;
457	while (pBundle)
458	{
459	PSUPDRVBUNDLE pToFree;
460	unsigned i;
461
462	/*
463	* Check and unlock all entries in the bundle.
464	*/
465	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
466	{
467	if (pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ)
468	{
469	int rc;
470	dprintf2(("eType=%d pvR0=%p pvR3=%p cb=%ld\n", pBundle->aMem[i].eType, RTR0MemObjAddress(pBundle->aMem[i].MemObj),
471	(void *)RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3), (long)RTR0MemObjSize(pBundle->aMem[i].MemObj)));
472	if (pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ)
473	{
474	rc = RTR0MemObjFree(pBundle->aMem[i].MapObjR3, false);
475	AssertRC(rc); /** @todo figure out how to handle this. */
476	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
477	}
478	rc = RTR0MemObjFree(pBundle->aMem[i].MemObj, false);
479	AssertRC(rc); /** @todo figure out how to handle this. */
480	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
481	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
482	}
483	}
484
485	/*
486	* Advance and free previous bundle.
487	*/
488	pToFree = pBundle;
489	pBundle = pBundle->pNext;
490
491	pToFree->pNext = NULL;
492	pToFree->cUsed = 0;
493	if (pToFree != &pSession->Bundle)
494	RTMemFree(pToFree);
495	}
496	dprintf2(("freeing memory - done\n"));
497
498	/*
499	* Loaded images needs to be dereferenced and possibly freed up.
500	*/
501	RTSemFastMutexRequest(pDevExt->mtxLdr);
502	dprintf2(("freeing images:\n"));
503	if (pSession->pLdrUsage)
504	{
505	PSUPDRVLDRUSAGE pUsage = pSession->pLdrUsage;
506	pSession->pLdrUsage = NULL;
507	while (pUsage)
508	{
509	void *pvFree = pUsage;
510	PSUPDRVLDRIMAGE pImage = pUsage->pImage;
511	if (pImage->cUsage > pUsage->cUsage)
512	pImage->cUsage -= pUsage->cUsage;
513	else
514	supdrvLdrFree(pDevExt, pImage);
515	pUsage->pImage = NULL;
516	pUsage = pUsage->pNext;
517	RTMemFree(pvFree);
518	}
519	}
520	RTSemFastMutexRelease(pDevExt->mtxLdr);
521	dprintf2(("freeing images - done\n"));
522
523	/*
524	* Unmap the GIP.
525	*/
526	dprintf2(("umapping GIP:\n"));
527	#ifdef USE_NEW_OS_INTERFACE_FOR_GIP
528	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
529	#else
530	if (pSession->pGip)
531	#endif
532	{
533	SUPR0GipUnmap(pSession);
534	#ifndef USE_NEW_OS_INTERFACE_FOR_GIP
535	pSession->pGip = NULL;
536	#endif
537	pSession->fGipReferenced = 0;
538	}
539	dprintf2(("umapping GIP - done\n"));
540	}
541
542
543	/**
544	* Fast path I/O Control worker.
545	*
546	* @returns VBox status code that should be passed down to ring-3 unchanged.
547	* @param uIOCtl Function number.
548	* @param pDevExt Device extention.
549	* @param pSession Session data.
550	*/
551	int VBOXCALL supdrvIOCtlFast(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
552	{
553	int rc;
554
555	/*
556	* We check the two prereqs after doing this only to allow the compiler to optimize things better.
557	*/
558	if (RT_LIKELY(pSession->pVM && pDevExt->pfnVMMR0EntryFast))
559	{
560	switch (uIOCtl)
561	{
562	case SUP_IOCTL_FAST_DO_RAW_RUN:
563	rc = pDevExt->pfnVMMR0EntryFast(pSession->pVM, SUP_VMMR0_DO_RAW_RUN);
564	break;
565	case SUP_IOCTL_FAST_DO_HWACC_RUN:
566	rc = pDevExt->pfnVMMR0EntryFast(pSession->pVM, SUP_VMMR0_DO_HWACC_RUN);
567	break;
568	case SUP_IOCTL_FAST_DO_NOP:
569	rc = pDevExt->pfnVMMR0EntryFast(pSession->pVM, SUP_VMMR0_DO_NOP);
570	break;
571	default:
572	rc = VERR_INTERNAL_ERROR;
573	break;
574	}
575	}
576	else
577	rc = VERR_INTERNAL_ERROR;
578
579	return rc;
580	}
581
582
583	/**
584	* I/O Control worker.
585	*
586	* @returns 0 on success.
587	* @returns VERR_INVALID_PARAMETER if the request is invalid.
588	*
589	* @param uIOCtl Function number.
590	* @param pDevExt Device extention.
591	* @param pSession Session data.
592	* @param pReqHdr The request header.
593	*/
594	int VBOXCALL supdrvIOCtl(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
595	{
596	/*
597	* Validate the request.
598	*/
599	/* this first check could probably be omitted as its also done by the OS specific code... */
600	if (RT_UNLIKELY( (pReqHdr->fFlags & SUPREQHDR_FLAGS_MAGIC_MASK) != SUPREQHDR_FLAGS_MAGIC
601	\|\| pReqHdr->cbIn < sizeof(*pReqHdr)
602	\|\| pReqHdr->cbOut < sizeof(*pReqHdr)))
603	{
604	OSDBGPRINT(("vboxdrv: Bad ioctl request header; cbIn=%#lx cbOut=%#lx fFlags=%#lx\n",
605	(long)pReqHdr->cbIn, (long)pReqHdr->cbOut, (long)pReqHdr->fFlags));
606	return VERR_INVALID_PARAMETER;
607	}
608	if (RT_UNLIKELY(uIOCtl == SUP_IOCTL_COOKIE))
609	{
610	if (pReqHdr->u32Cookie != SUPCOOKIE_INITIAL_COOKIE)
611	{
612	OSDBGPRINT(("SUP_IOCTL_COOKIE: bad cookie %#lx\n", (long)pReqHdr->u32Cookie));
613	return VERR_INVALID_PARAMETER;
614	}
615	}
616	else if (RT_UNLIKELY( pReqHdr->u32Cookie != pDevExt->u32Cookie
617	\|\| pReqHdr->u32SessionCookie != pSession->u32Cookie))
618	{
619	OSDBGPRINT(("vboxdrv: bad cookie %#lx / %#lx.\n", (long)pReqHdr->u32Cookie, (long)pReqHdr->u32SessionCookie));
620	return VERR_INVALID_PARAMETER;
621	}
622
623	/*
624	* Validation macros
625	*/
626	#define REQ_CHECK_SIZES_EX(Name, cbInExpect, cbOutExpect) \
627	do { \
628	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect) \|\| pReqHdr->cbOut != (cbOutExpect))) \
629	{ \
630	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld. cbOut=%ld expected %ld.\n", \
631	(long)pReq->Hdr.cbIn, (long)(cbInExpect), (long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
632	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
633	} \
634	} while (0)
635
636	#define REQ_CHECK_SIZES(Name) REQ_CHECK_SIZES_EX(Name, Name ## _SIZE_IN, Name ## _SIZE_OUT)
637
638	#define REQ_CHECK_SIZE_IN(Name, cbInExpect) \
639	do { \
640	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect))) \
641	{ \
642	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld.\n", \
643	(long)pReq->Hdr.cbIn, (long)(cbInExpect))); \
644	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
645	} \
646	} while (0)
647
648	#define REQ_CHECK_SIZE_OUT(Name, cbOutExpect) \
649	do { \
650	if (RT_UNLIKELY(pReqHdr->cbOut != (cbOutExpect))) \
651	{ \
652	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbOut=%ld expected %ld.\n", \
653	(long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
654	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
655	} \
656	} while (0)
657
658	#define REQ_CHECK_EXPR(Name, expr) \
659	do { \
660	if (RT_UNLIKELY(!(expr))) \
661	{ \
662	OSDBGPRINT(( #Name ": %s\n", #expr)); \
663	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
664	} \
665	} while (0)
666
667	#define REQ_CHECK_EXPR_FMT(expr, fmt) \
668	do { \
669	if (RT_UNLIKELY(!(expr))) \
670	{ \
671	OSDBGPRINT( fmt ); \
672	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
673	} \
674	} while (0)
675
676
677	/*
678	* The switch.
679	*/
680	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
681	{
682	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
683	{
684	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
685	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
686	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
687	{
688	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
689	pReq->Hdr.rc = VERR_INVALID_MAGIC;
690	return 0;
691	}
692
693	#if 0
694	/*
695	* Call out to the OS specific code and let it do permission checks on the
696	* client process.
697	*/
698	if (!supdrvOSValidateClientProcess(pDevExt, pSession))
699	{
700	pReq->u.Out.u32Cookie = 0xffffffff;
701	pReq->u.Out.u32SessionCookie = 0xffffffff;
702	pReq->u.Out.u32SessionVersion = 0xffffffff;
703	pReq->u.Out.u32DriverVersion = SUPDRVIOC_VERSION;
704	pReq->u.Out.pSession = NULL;
705	pReq->u.Out.cFunctions = 0;
706	pReq->Hdr.rc = VERR_PERMISSION_DENIED;
707	return 0;
708	}
709	#endif
710
711	/*
712	* Match the version.
713	* The current logic is very simple, match the major interface version.
714	*/
715	if ( pReq->u.In.u32MinVersion > SUPDRVIOC_VERSION
716	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRVIOC_VERSION & 0xffff0000))
717	{
718	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
719	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRVIOC_VERSION));
720	pReq->u.Out.u32Cookie = 0xffffffff;
721	pReq->u.Out.u32SessionCookie = 0xffffffff;
722	pReq->u.Out.u32SessionVersion = 0xffffffff;
723	pReq->u.Out.u32DriverVersion = SUPDRVIOC_VERSION;
724	pReq->u.Out.pSession = NULL;
725	pReq->u.Out.cFunctions = 0;
726	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
727	return 0;
728	}
729
730	/*
731	* Fill in return data and be gone.
732	* N.B. The first one to change SUPDRVIOC_VERSION shall makes sure that
733	* u32SessionVersion <= u32ReqVersion!
734	*/
735	/** @todo Somehow validate the client and negotiate a secure cookie... */
736	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
737	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
738	pReq->u.Out.u32SessionVersion = SUPDRVIOC_VERSION;
739	pReq->u.Out.u32DriverVersion = SUPDRVIOC_VERSION;
740	pReq->u.Out.pSession = pSession;
741	pReq->u.Out.cFunctions = sizeof(g_aFunctions) / sizeof(g_aFunctions[0]);
742	pReq->Hdr.rc = VINF_SUCCESS;
743	return 0;
744	}
745
746	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_QUERY_FUNCS(0)):
747	{
748	/* validate */
749	PSUPQUERYFUNCS pReq = (PSUPQUERYFUNCS)pReqHdr;
750	REQ_CHECK_SIZES_EX(SUP_IOCTL_QUERY_FUNCS, SUP_IOCTL_QUERY_FUNCS_SIZE_IN, SUP_IOCTL_QUERY_FUNCS_SIZE_OUT(RT_ELEMENTS(g_aFunctions)));
751
752	/* execute */
753	pReq->u.Out.cFunctions = RT_ELEMENTS(g_aFunctions);
754	memcpy(&pReq->u.Out.aFunctions[0], g_aFunctions, sizeof(g_aFunctions));
755	pReq->Hdr.rc = VINF_SUCCESS;
756	return 0;
757	}
758
759	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_IDT_INSTALL):
760	{
761	/* validate */
762	PSUPIDTINSTALL pReq = (PSUPIDTINSTALL)pReqHdr;
763	REQ_CHECK_SIZES(SUP_IOCTL_IDT_INSTALL);
764
765	/* execute */
766	#ifdef VBOX_WITH_IDT_PATCHING
767	pReq->Hdr.rc = supdrvIOCtl_IdtInstall(pDevExt, pSession, pReq);
768	#else
769	pReq->u.Out.u8Idt = 3;
770	pReq->Hdr.rc = VERR_NOT_SUPPORTED;
771	#endif
772	return 0;
773	}
774
775	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_IDT_REMOVE):
776	{
777	/* validate */
778	PSUPIDTREMOVE pReq = (PSUPIDTREMOVE)pReqHdr;
779	REQ_CHECK_SIZES(SUP_IOCTL_IDT_REMOVE);
780
781	/* execute */
782	#ifdef VBOX_WITH_IDT_PATCHING
783	pReq->Hdr.rc = supdrvIOCtl_IdtRemoveAll(pDevExt, pSession);
784	#else
785	pReq->Hdr.rc = VERR_NOT_SUPPORTED;
786	#endif
787	return 0;
788	}
789
790	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_LOCK):
791	{
792	/* validate */
793	PSUPPAGELOCK pReq = (PSUPPAGELOCK)pReqHdr;
794	REQ_CHECK_SIZE_IN(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_IN);
795	REQ_CHECK_SIZE_OUT(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_OUT(pReq->u.In.cPages));
796	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.cPages > 0);
797	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.pvR3 >= PAGE_SIZE);
798
799	/* execute */
800	pReq->Hdr.rc = SUPR0LockMem(pSession, pReq->u.In.pvR3, pReq->u.In.cPages, &pReq->u.Out.aPages[0]);
801	if (RT_FAILURE(pReq->Hdr.rc))
802	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
803	return 0;
804	}
805
806	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_UNLOCK):
807	{
808	/* validate */
809	PSUPPAGEUNLOCK pReq = (PSUPPAGEUNLOCK)pReqHdr;
810	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_UNLOCK);
811
812	/* execute */
813	pReq->Hdr.rc = SUPR0UnlockMem(pSession, pReq->u.In.pvR3);
814	return 0;
815	}
816
817	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_ALLOC):
818	{
819	/* validate */
820	PSUPCONTALLOC pReq = (PSUPCONTALLOC)pReqHdr;
821	REQ_CHECK_SIZES(SUP_IOCTL_CONT_ALLOC);
822
823	/* execute */
824	pReq->Hdr.rc = SUPR0ContAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.HCPhys);
825	if (RT_FAILURE(pReq->Hdr.rc))
826	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
827	return 0;
828	}
829
830	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_FREE):
831	{
832	/* validate */
833	PSUPCONTFREE pReq = (PSUPCONTFREE)pReqHdr;
834	REQ_CHECK_SIZES(SUP_IOCTL_CONT_FREE);
835
836	/* execute */
837	pReq->Hdr.rc = SUPR0ContFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
838	return 0;
839	}
840
841	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_OPEN):
842	{
843	/* validate */
844	PSUPLDROPEN pReq = (PSUPLDROPEN)pReqHdr;
845	REQ_CHECK_SIZES(SUP_IOCTL_LDR_OPEN);
846	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImage > 0);
847	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImage < _1M*16);
848	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.szName[0]);
849	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, memchr(pReq->u.In.szName, '\0', sizeof(pReq->u.In.szName)));
850	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, !strpbrk(pReq->u.In.szName, ";:()[]{}/\\\|&*%#@!~`\"'"));
851
852	/* execute */
853	pReq->Hdr.rc = supdrvIOCtl_LdrOpen(pDevExt, pSession, pReq);
854	return 0;
855	}
856
857	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOAD):
858	{
859	/* validate */
860	PSUPLDRLOAD pReq = (PSUPLDRLOAD)pReqHdr;
861	REQ_CHECK_EXPR(Name, pReq->Hdr.cbIn >= sizeof(*pReq));
862	REQ_CHECK_SIZES_EX(SUP_IOCTL_LDR_LOAD, SUP_IOCTL_LDR_LOAD_SIZE_IN(pReq->u.In.cbImage), SUP_IOCTL_LDR_LOAD_SIZE_OUT);
863	REQ_CHECK_EXPR(SUP_IOCTL_LDR_LOAD, pReq->u.In.cSymbols <= 16384);
864	REQ_CHECK_EXPR_FMT( !pReq->u.In.cSymbols
865	\|\| ( pReq->u.In.offSymbols < pReq->u.In.cbImage
866	&& pReq->u.In.offSymbols + pReq->u.In.cSymbols * sizeof(SUPLDRSYM) <= pReq->u.In.cbImage),
867	("SUP_IOCTL_LDR_LOAD: offSymbols=%#lx cSymbols=%#lx cbImage=%#lx\n", (long)pReq->u.In.offSymbols,
868	(long)pReq->u.In.cSymbols, (long)pReq->u.In.cbImage));
869	REQ_CHECK_EXPR_FMT( !pReq->u.In.cbStrTab
870	\|\| ( pReq->u.In.offStrTab < pReq->u.In.cbImage
871	&& pReq->u.In.offStrTab + pReq->u.In.cbStrTab <= pReq->u.In.cbImage
872	&& pReq->u.In.cbStrTab <= pReq->u.In.cbImage),
873	("SUP_IOCTL_LDR_LOAD: offStrTab=%#lx cbStrTab=%#lx cbImage=%#lx\n", (long)pReq->u.In.offStrTab,
874	(long)pReq->u.In.cbStrTab, (long)pReq->u.In.cbImage));
875
876	if (pReq->u.In.cSymbols)
877	{
878	uint32_t i;
879	PSUPLDRSYM paSyms = (PSUPLDRSYM)&pReq->u.In.achImage[pReq->u.In.offSymbols];
880	for (i = 0; i < pReq->u.In.cSymbols; i++)
881	{
882	REQ_CHECK_EXPR_FMT(paSyms[i].offSymbol < pReq->u.In.cbImage,
883	("SUP_IOCTL_LDR_LOAD: sym #%ld: symb off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offSymbol, (long)pReq->u.In.cbImage));
884	REQ_CHECK_EXPR_FMT(paSyms[i].offName < pReq->u.In.cbStrTab,
885	("SUP_IOCTL_LDR_LOAD: sym #%ld: name off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImage));
886	REQ_CHECK_EXPR_FMT(memchr(&pReq->u.In.achImage[pReq->u.In.offStrTab + paSyms[i].offName], '\0', pReq->u.In.cbStrTab - paSyms[i].offName),
887	("SUP_IOCTL_LDR_LOAD: sym #%ld: unterminated name! (%#lx / %#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImage));
888	}
889	}
890
891	/* execute */
892	pReq->Hdr.rc = supdrvIOCtl_LdrLoad(pDevExt, pSession, pReq);
893	return 0;
894	}
895
896	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_FREE):
897	{
898	/* validate */
899	PSUPLDRFREE pReq = (PSUPLDRFREE)pReqHdr;
900	REQ_CHECK_SIZES(SUP_IOCTL_LDR_FREE);
901
902	/* execute */
903	pReq->Hdr.rc = supdrvIOCtl_LdrFree(pDevExt, pSession, pReq);
904	return 0;
905	}
906
907	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_GET_SYMBOL):
908	{
909	/* validate */
910	PSUPLDRGETSYMBOL pReq = (PSUPLDRGETSYMBOL)pReqHdr;
911	REQ_CHECK_SIZES(SUP_IOCTL_LDR_GET_SYMBOL);
912	REQ_CHECK_EXPR(SUP_IOCTL_LDR_GET_SYMBOL, memchr(pReq->u.In.szSymbol, '\0', sizeof(pReq->u.In.szSymbol)));
913
914	/* execute */
915	pReq->Hdr.rc = supdrvIOCtl_LdrGetSymbol(pDevExt, pSession, pReq);
916	return 0;
917	}
918
919	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0(0)):
920	{
921	/* validate */
922	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
923	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_VMMR0_SIZE(0))
924	{
925	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(0), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(0));
926
927	/* execute */
928	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
929	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg);
930	else
931	pReq->Hdr.rc = VERR_WRONG_ORDER;
932	}
933	else
934	{
935	PSUPVMMR0REQHDR pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
936	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR)),
937	("SUP_IOCTL_CALL_VMMR0: cbIn=%#x < %#x\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR))));
938	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
939	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(pVMMReq->cbReq));
940
941	/* execute */
942	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
943	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg);
944	else
945	pReq->Hdr.rc = VERR_WRONG_ORDER;
946	}
947	return 0;
948	}
949
950	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_PAGING_MODE):
951	{
952	/* validate */
953	PSUPGETPAGINGMODE pReq = (PSUPGETPAGINGMODE)pReqHdr;
954	REQ_CHECK_SIZES(SUP_IOCTL_GET_PAGING_MODE);
955
956	/* execute */
957	pReq->Hdr.rc = VINF_SUCCESS;
958	pReq->u.Out.enmMode = supdrvIOCtl_GetPagingMode();
959	return 0;
960	}
961
962	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_ALLOC):
963	{
964	/* validate */
965	PSUPLOWALLOC pReq = (PSUPLOWALLOC)pReqHdr;
966	REQ_CHECK_EXPR(SUP_IOCTL_LOW_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_LOW_ALLOC_SIZE_IN);
967	REQ_CHECK_SIZES_EX(SUP_IOCTL_LOW_ALLOC, SUP_IOCTL_LOW_ALLOC_SIZE_IN, SUP_IOCTL_LOW_ALLOC_SIZE_OUT(pReq->u.In.cPages));
968
969	/* execute */
970	pReq->Hdr.rc = SUPR0LowAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
971	if (RT_FAILURE(pReq->Hdr.rc))
972	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
973	return 0;
974	}
975
976	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_FREE):
977	{
978	/* validate */
979	PSUPLOWFREE pReq = (PSUPLOWFREE)pReqHdr;
980	REQ_CHECK_SIZES(SUP_IOCTL_LOW_FREE);
981
982	/* execute */
983	pReq->Hdr.rc = SUPR0LowFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
984	return 0;
985	}
986
987	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_MAP):
988	{
989	/* validate */
990	PSUPGIPMAP pReq = (PSUPGIPMAP)pReqHdr;
991	REQ_CHECK_SIZES(SUP_IOCTL_GIP_MAP);
992
993	/* execute */
994	pReq->Hdr.rc = SUPR0GipMap(pSession, &pReq->u.Out.pGipR3, &pReq->u.Out.HCPhysGip);
995	if (RT_SUCCESS(pReq->Hdr.rc))
996	pReq->u.Out.pGipR0 = pDevExt->pGip;
997	return 0;
998	}
999
1000	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_UNMAP):
1001	{
1002	/* validate */
1003	PSUPGIPUNMAP pReq = (PSUPGIPUNMAP)pReqHdr;
1004	REQ_CHECK_SIZES(SUP_IOCTL_GIP_UNMAP);
1005
1006	/* execute */
1007	pReq->Hdr.rc = SUPR0GipUnmap(pSession);
1008	return 0;
1009	}
1010
1011	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SET_VM_FOR_FAST):
1012	{
1013	/* validate */
1014	PSUPSETVMFORFAST pReq = (PSUPSETVMFORFAST)pReqHdr;
1015	REQ_CHECK_SIZES(SUP_IOCTL_SET_VM_FOR_FAST);
1016	REQ_CHECK_EXPR_FMT( !pReq->u.In.pVMR0
1017	\|\| ( VALID_PTR(pReq->u.In.pVMR0)
1018	&& !((uintptr_t)pReq->u.In.pVMR0 & (PAGE_SIZE - 1))),
1019	("SUP_IOCTL_SET_VM_FOR_FAST: pVMR0=%p!\n", pReq->u.In.pVMR0));
1020	/* execute */
1021	pSession->pVM = pReq->u.In.pVMR0;
1022	pReq->Hdr.rc = VINF_SUCCESS;
1023	return 0;
1024	}
1025
1026	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_ALLOC):
1027	{
1028	/* validate */
1029	PSUPPAGEALLOC pReq = (PSUPPAGEALLOC)pReqHdr;
1030	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_SIZE_IN);
1031	REQ_CHECK_SIZES_EX(SUP_IOCTL_PAGE_ALLOC, SUP_IOCTL_PAGE_ALLOC_SIZE_IN, SUP_IOCTL_PAGE_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1032
1033	/* execute */
1034	pReq->Hdr.rc = SUPR0PageAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1035	if (RT_FAILURE(pReq->Hdr.rc))
1036	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1037	return 0;
1038	}
1039
1040	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_FREE):
1041	{
1042	/* validate */
1043	PSUPPAGEFREE pReq = (PSUPPAGEFREE)pReqHdr;
1044	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_FREE);
1045
1046	/* execute */
1047	pReq->Hdr.rc = SUPR0PageFree(pSession, pReq->u.In.pvR3);
1048	return 0;
1049	}
1050
1051	default:
1052	dprintf(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1053	break;
1054	}
1055	return SUPDRV_ERR_GENERAL_FAILURE;
1056	}
1057
1058
1059	/**
1060	* Register a object for reference counting.
1061	* The object is registered with one reference in the specified session.
1062	*
1063	* @returns Unique identifier on success (pointer).
1064	* All future reference must use this identifier.
1065	* @returns NULL on failure.
1066	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
1067	* @param pvUser1 The first user argument.
1068	* @param pvUser2 The second user argument.
1069	*/
1070	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
1071	{
1072	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1073	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1074	PSUPDRVOBJ pObj;
1075	PSUPDRVUSAGE pUsage;
1076
1077	/*
1078	* Validate the input.
1079	*/
1080	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
1081	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
1082	AssertPtrReturn(pfnDestructor, NULL);
1083
1084	/*
1085	* Allocate and initialize the object.
1086	*/
1087	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
1088	if (!pObj)
1089	return NULL;
1090	pObj->u32Magic = SUPDRVOBJ_MAGIC;
1091	pObj->enmType = enmType;
1092	pObj->pNext = NULL;
1093	pObj->cUsage = 1;
1094	pObj->pfnDestructor = pfnDestructor;
1095	pObj->pvUser1 = pvUser1;
1096	pObj->pvUser2 = pvUser2;
1097	pObj->CreatorUid = pSession->Uid;
1098	pObj->CreatorGid = pSession->Gid;
1099	pObj->CreatorProcess= pSession->Process;
1100	supdrvOSObjInitCreator(pObj, pSession);
1101
1102	/*
1103	* Allocate the usage record.
1104	* (We keep freed usage records around to simplity SUPR0ObjAddRef().)
1105	*/
1106	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1107
1108	pUsage = pDevExt->pUsageFree;
1109	if (pUsage)
1110	pDevExt->pUsageFree = pUsage->pNext;
1111	else
1112	{
1113	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1114	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
1115	if (!pUsage)
1116	{
1117	RTMemFree(pObj);
1118	return NULL;
1119	}
1120	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1121	}
1122
1123	/*
1124	* Insert the object and create the session usage record.
1125	*/
1126	/* The object. */
1127	pObj->pNext = pDevExt->pObjs;
1128	pDevExt->pObjs = pObj;
1129
1130	/* The session record. */
1131	pUsage->cUsage = 1;
1132	pUsage->pObj = pObj;
1133	pUsage->pNext = pSession->pUsage;
1134	dprintf(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));
1135	pSession->pUsage = pUsage;
1136
1137	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1138
1139	dprintf(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
1140	return pObj;
1141	}
1142
1143
1144	/**
1145	* Increment the reference counter for the object associating the reference
1146	* with the specified session.
1147	*
1148	* @returns IPRT status code.
1149	* @param pvObj The identifier returned by SUPR0ObjRegister().
1150	* @param pSession The session which is referencing the object.
1151	*/
1152	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
1153	{
1154	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1155	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1156	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
1157	PSUPDRVUSAGE pUsagePre;
1158	PSUPDRVUSAGE pUsage;
1159
1160	/*
1161	* Validate the input.
1162	*/
1163	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1164	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
1165	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
1166	VERR_INVALID_PARAMETER);
1167
1168	/*
1169	* Preallocate the usage record.
1170	*/
1171	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1172
1173	pUsagePre = pDevExt->pUsageFree;
1174	if (pUsagePre)
1175	pDevExt->pUsageFree = pUsagePre->pNext;
1176	else
1177	{
1178	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1179	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
1180	if (!pUsagePre)
1181	return VERR_NO_MEMORY;
1182	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1183	}
1184
1185	/*
1186	* Reference the object.
1187	*/
1188	pObj->cUsage++;
1189
1190	/*
1191	* Look for the session record.
1192	*/
1193	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
1194	{
1195	dprintf(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));
1196	if (pUsage->pObj == pObj)
1197	break;
1198	}
1199	if (pUsage)
1200	pUsage->cUsage++;
1201	else
1202	{
1203	/* create a new session record. */
1204	pUsagePre->cUsage = 1;
1205	pUsagePre->pObj = pObj;
1206	pUsagePre->pNext = pSession->pUsage;
1207	pSession->pUsage = pUsagePre;
1208	dprintf(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));
1209
1210	pUsagePre = NULL;
1211	}
1212
1213	/*
1214	* Put any unused usage record into the free list..
1215	*/
1216	if (pUsagePre)
1217	{
1218	pUsagePre->pNext = pDevExt->pUsageFree;
1219	pDevExt->pUsageFree = pUsagePre;
1220	}
1221
1222	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1223
1224	return VINF_SUCCESS;
1225	}
1226
1227
1228	/**
1229	* Decrement / destroy a reference counter record for an object.
1230	*
1231	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
1232	*
1233	* @returns IPRT status code.
1234	* @param pvObj The identifier returned by SUPR0ObjRegister().
1235	* @param pSession The session which is referencing the object.
1236	*/
1237	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
1238	{
1239	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1240	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1241	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
1242	bool fDestroy = false;
1243	PSUPDRVUSAGE pUsage;
1244	PSUPDRVUSAGE pUsagePrev;
1245
1246	/*
1247	* Validate the input.
1248	*/
1249	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1250	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
1251	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
1252	VERR_INVALID_PARAMETER);
1253
1254	/*
1255	* Acquire the spinlock and look for the usage record.
1256	*/
1257	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1258
1259	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
1260	pUsage;
1261	pUsagePrev = pUsage, pUsage = pUsage->pNext)
1262	{
1263	dprintf(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));
1264	if (pUsage->pObj == pObj)
1265	{
1266	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
1267	if (pUsage->cUsage > 1)
1268	{
1269	pObj->cUsage--;
1270	pUsage->cUsage--;
1271	}
1272	else
1273	{
1274	/*
1275	* Free the session record.
1276	*/
1277	if (pUsagePrev)
1278	pUsagePrev->pNext = pUsage->pNext;
1279	else
1280	pSession->pUsage = pUsage->pNext;
1281	pUsage->pNext = pDevExt->pUsageFree;
1282	pDevExt->pUsageFree = pUsage;
1283
1284	/* What about the object? */
1285	if (pObj->cUsage > 1)
1286	pObj->cUsage--;
1287	else
1288	{
1289	/*
1290	* Object is to be destroyed, unlink it.
1291	*/
1292	fDestroy = true;
1293	if (pDevExt->pObjs == pObj)
1294	pDevExt->pObjs = pObj->pNext;
1295	else
1296	{
1297	PSUPDRVOBJ pObjPrev;
1298	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
1299	if (pObjPrev->pNext == pObj)
1300	{
1301	pObjPrev->pNext = pObj->pNext;
1302	break;
1303	}
1304	Assert(pObjPrev);
1305	}
1306	}
1307	}
1308	break;
1309	}
1310	}
1311
1312	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1313
1314	/*
1315	* Call the destructor and free the object if required.
1316	*/
1317	if (fDestroy)
1318	{
1319	pObj->u32Magic++;
1320	if (pObj->pfnDestructor)
1321	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
1322	RTMemFree(pObj);
1323	}
1324
1325	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
1326	return pUsage ? VINF_SUCCESS : VERR_INVALID_PARAMETER;
1327	}
1328
1329	/**
1330	* Verifies that the current process can access the specified object.
1331	*
1332	* @returns The following IPRT status code:
1333	* @retval VINF_SUCCESS if access was granted.
1334	* @retval VERR_PERMISSION_DENIED if denied access.
1335	* @retval VERR_INVALID_PARAMETER if invalid parameter.
1336	*
1337	* @param pvObj The identifier returned by SUPR0ObjRegister().
1338	* @param pSession The session which wishes to access the object.
1339	* @param pszObjName Object string name. This is optional and depends on the object type.
1340	*
1341	* @remark The caller is responsible for making sure the object isn't removed while
1342	* we're inside this function. If uncertain about this, just call AddRef before calling us.
1343	*/
1344	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
1345	{
1346	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
1347	int rc;
1348
1349	/*
1350	* Validate the input.
1351	*/
1352	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1353	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
1354	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
1355	VERR_INVALID_PARAMETER);
1356
1357	/*
1358	* Check access. (returns true if a decision has been made.)
1359	*/
1360	rc = VERR_INTERNAL_ERROR;
1361	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
1362	return rc;
1363
1364	/*
1365	* Default policy is to allow the user to access his own
1366	* stuff but nothing else.
1367	*/
1368	if (pObj->CreatorUid == pSession->Uid)
1369	return VINF_SUCCESS;
1370	return VERR_PERMISSION_DENIED;
1371	}
1372
1373
1374	/**
1375	* Lock pages.
1376	*
1377	* @returns IPRT status code.
1378	* @param pSession Session to which the locked memory should be associated.
1379	* @param pvR3 Start of the memory range to lock.
1380	* This must be page aligned.
1381	* @param cb Size of the memory range to lock.
1382	* This must be page aligned.
1383	*/
1384	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
1385	{
1386	int rc;
1387	SUPDRVMEMREF Mem = {0};
1388	const size_t cb = (size_t)cPages << PAGE_SHIFT;
1389	dprintf(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
1390
1391	/*
1392	* Verify input.
1393	*/
1394	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1395	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
1396	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
1397	\|\| !pvR3)
1398	{
1399	dprintf(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
1400	return VERR_INVALID_PARAMETER;
1401	}
1402
1403	#ifdef RT_OS_WINDOWS /* A temporary hack for windows, will be removed once all ring-3 code has been cleaned up. */
1404	/* First check if we allocated it using SUPPageAlloc; if so then we don't need to lock it again */
1405	rc = supdrvPageGetPhys(pSession, pvR3, cPages, paPages);
1406	if (RT_SUCCESS(rc))
1407	return rc;
1408	#endif
1409
1410	/*
1411	* Let IPRT do the job.
1412	*/
1413	Mem.eType = MEMREF_TYPE_LOCKED;
1414	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTR0ProcHandleSelf());
1415	if (RT_SUCCESS(rc))
1416	{
1417	uint32_t iPage = cPages;
1418	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
1419	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
1420
1421	while (iPage-- > 0)
1422	{
1423	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
1424	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
1425	{
1426	AssertMsgFailed(("iPage=%d\n", iPage));
1427	rc = VERR_INTERNAL_ERROR;
1428	break;
1429	}
1430	}
1431	if (RT_SUCCESS(rc))
1432	rc = supdrvMemAdd(&Mem, pSession);
1433	if (RT_FAILURE(rc))
1434	{
1435	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
1436	AssertRC(rc2);
1437	}
1438	}
1439
1440	return rc;
1441	}
1442
1443
1444	/**
1445	* Unlocks the memory pointed to by pv.
1446	*
1447	* @returns IPRT status code.
1448	* @param pSession Session to which the memory was locked.
1449	* @param pvR3 Memory to unlock.
1450	*/
1451	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
1452	{
1453	dprintf(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
1454	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1455	#ifdef RT_OS_WINDOWS
1456	/*
1457	* Temporary hack for windows - SUPR0PageFree will unlock SUPR0PageAlloc
1458	* allocations; ignore this call.
1459	*/
1460	if (supdrvPageWasLockedByPageAlloc(pSession, pvR3))
1461	{
1462	dprintf(("Page will be unlocked in SUPR0PageFree -> ignore\n"));
1463	return VINF_SUCCESS;
1464	}
1465	#endif
1466	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
1467	}
1468
1469
1470	/**
1471	* Allocates a chunk of page aligned memory with contiguous and fixed physical
1472	* backing.
1473	*
1474	* @returns IPRT status code.
1475	* @param pSession Session data.
1476	* @param cb Number of bytes to allocate.
1477	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
1478	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
1479	* @param pHCPhys Where to put the physical address of allocated memory.
1480	*/
1481	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
1482	{
1483	int rc;
1484	SUPDRVMEMREF Mem = {0};
1485	dprintf(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
1486
1487	/*
1488	* Validate input.
1489	*/
1490	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1491	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
1492	{
1493	dprintf(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
1494	pSession, ppvR0, ppvR3, pHCPhys));
1495	return VERR_INVALID_PARAMETER;
1496
1497	}
1498	if (cPages < 1 \|\| cPages >= 256)
1499	{
1500	dprintf(("Illegal request cPages=%d, must be greater than 0 and smaller than 256\n", cPages));
1501	return VERR_INVALID_PARAMETER;
1502	}
1503
1504	/*
1505	* Let IPRT do the job.
1506	*/
1507	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
1508	if (RT_SUCCESS(rc))
1509	{
1510	int rc2;
1511	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
1512	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
1513	if (RT_SUCCESS(rc))
1514	{
1515	Mem.eType = MEMREF_TYPE_CONT;
1516	rc = supdrvMemAdd(&Mem, pSession);
1517	if (!rc)
1518	{
1519	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
1520	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
1521	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
1522	return 0;
1523	}
1524
1525	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
1526	AssertRC(rc2);
1527	}
1528	rc2 = RTR0MemObjFree(Mem.MemObj, false);
1529	AssertRC(rc2);
1530	}
1531
1532	return rc;
1533	}
1534
1535
1536	/**
1537	* Frees memory allocated using SUPR0ContAlloc().
1538	*
1539	* @returns IPRT status code.
1540	* @param pSession The session to which the memory was allocated.
1541	* @param uPtr Pointer to the memory (ring-3 or ring-0).
1542	*/
1543	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
1544	{
1545	dprintf(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
1546	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1547	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
1548	}
1549
1550
1551	/**
1552	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
1553	*
1554	* @returns IPRT status code.
1555	* @param pSession Session data.
1556	* @param cPages Number of pages to allocate.
1557	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
1558	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
1559	* @param paPages Where to put the physical addresses of allocated memory.
1560	*/
1561	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
1562	{
1563	unsigned iPage;
1564	int rc;
1565	SUPDRVMEMREF Mem = {0};
1566	dprintf(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
1567
1568	/*
1569	* Validate input.
1570	*/
1571	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1572	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
1573	{
1574	dprintf(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
1575	pSession, ppvR3, ppvR0, paPages));
1576	return VERR_INVALID_PARAMETER;
1577
1578	}
1579	if (cPages < 1 \|\| cPages > 256)
1580	{
1581	dprintf(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
1582	return VERR_INVALID_PARAMETER;
1583	}
1584
1585	/*
1586	* Let IPRT do the work.
1587	*/
1588	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
1589	if (RT_SUCCESS(rc))
1590	{
1591	int rc2;
1592	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
1593	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
1594	if (RT_SUCCESS(rc))
1595	{
1596	Mem.eType = MEMREF_TYPE_LOW;
1597	rc = supdrvMemAdd(&Mem, pSession);
1598	if (!rc)
1599	{
1600	for (iPage = 0; iPage < cPages; iPage++)
1601	{
1602	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
1603	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%VHp\n", paPages[iPage]));
1604	}
1605	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
1606	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
1607	return 0;
1608	}
1609
1610	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
1611	AssertRC(rc2);
1612	}
1613
1614	rc2 = RTR0MemObjFree(Mem.MemObj, false);
1615	AssertRC(rc2);
1616	}
1617
1618	return rc;
1619	}
1620
1621
1622	/**
1623	* Frees memory allocated using SUPR0LowAlloc().
1624	*
1625	* @returns IPRT status code.
1626	* @param pSession The session to which the memory was allocated.
1627	* @param uPtr Pointer to the memory (ring-3 or ring-0).
1628	*/
1629	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
1630	{
1631	dprintf(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
1632	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1633	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
1634	}
1635
1636
1637
1638	/**
1639	* Allocates a chunk of memory with both R0 and R3 mappings.
1640	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
1641	*
1642	* @returns IPRT status code.
1643	* @param pSession The session to associated the allocation with.
1644	* @param cb Number of bytes to allocate.
1645	* @param ppvR0 Where to store the address of the Ring-0 mapping.
1646	* @param ppvR3 Where to store the address of the Ring-3 mapping.
1647	*/
1648	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
1649	{
1650	int rc;
1651	SUPDRVMEMREF Mem = {0};
1652	dprintf(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
1653
1654	/*
1655	* Validate input.
1656	*/
1657	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1658	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
1659	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
1660	if (cb < 1 \|\| cb >= _4M)
1661	{
1662	dprintf(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
1663	return VERR_INVALID_PARAMETER;
1664	}
1665
1666	/*
1667	* Let IPRT do the work.
1668	*/
1669	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
1670	if (RT_SUCCESS(rc))
1671	{
1672	int rc2;
1673	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
1674	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
1675	if (RT_SUCCESS(rc))
1676	{
1677	Mem.eType = MEMREF_TYPE_MEM;
1678	rc = supdrvMemAdd(&Mem, pSession);
1679	if (!rc)
1680	{
1681	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
1682	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
1683	return VINF_SUCCESS;
1684	}
1685	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
1686	AssertRC(rc2);
1687	}
1688
1689	rc2 = RTR0MemObjFree(Mem.MemObj, false);
1690	AssertRC(rc2);
1691	}
1692
1693	return rc;
1694	}
1695
1696
1697	/**
1698	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
1699	*
1700	* @returns IPRT status code.
1701	* @param pSession The session to which the memory was allocated.
1702	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
1703	* @param paPages Where to store the physical addresses.
1704	*/
1705	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
1706	{
1707	PSUPDRVBUNDLE pBundle;
1708	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1709	dprintf(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
1710
1711	/*
1712	* Validate input.
1713	*/
1714	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1715	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
1716	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
1717
1718	/*
1719	* Search for the address.
1720	*/
1721	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
1722	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
1723	{
1724	if (pBundle->cUsed > 0)
1725	{
1726	unsigned i;
1727	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
1728	{
1729	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
1730	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
1731	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
1732	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
1733	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
1734	)
1735	)
1736	{
1737	const unsigned cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
1738	unsigned iPage;
1739	for (iPage = 0; iPage < cPages; iPage++)
1740	{
1741	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
1742	paPages[iPage].uReserved = 0;
1743	}
1744	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
1745	return VINF_SUCCESS;
1746	}
1747	}
1748	}
1749	}
1750	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
1751	dprintf(("Failed to find %p!!!\n", (void *)uPtr));
1752	return VERR_INVALID_PARAMETER;
1753	}
1754
1755
1756	/**
1757	* Free memory allocated by SUPR0MemAlloc().
1758	*
1759	* @returns IPRT status code.
1760	* @param pSession The session owning the allocation.
1761	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
1762	*/
1763	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
1764	{
1765	dprintf(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
1766	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1767	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
1768	}
1769
1770
1771	/**
1772	* Allocates a chunk of memory with only a R3 mappings.
1773	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
1774	*
1775	* @returns IPRT status code.
1776	* @param pSession The session to associated the allocation with.
1777	* @param cPages The number of pages to allocate.
1778	* @param ppvR3 Where to store the address of the Ring-3 mapping.
1779	* @param paPages Where to store the addresses of the pages. Optional.
1780	*/
1781	SUPR0DECL(int) SUPR0PageAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR3PTR ppvR3, PRTHCPHYS paPages)
1782	{
1783	int rc;
1784	SUPDRVMEMREF Mem = {0};
1785	dprintf(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
1786
1787	/*
1788	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
1789	*/
1790	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1791	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
1792	if (cPages < 1 \|\| cPages > (128 * _1M)/PAGE_SIZE)
1793	{
1794	dprintf(("SUPR0PageAlloc: Illegal request cb=%u; must be greater than 0 and smaller than 128MB.\n", cPages));
1795	return VERR_INVALID_PARAMETER;
1796	}
1797
1798	/*
1799	* Let IPRT do the work.
1800	*/
1801	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
1802	if (RT_SUCCESS(rc))
1803	{
1804	int rc2;
1805	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
1806	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
1807	if (RT_SUCCESS(rc))
1808	{
1809	Mem.eType = MEMREF_TYPE_LOCKED_SUP;
1810	rc = supdrvMemAdd(&Mem, pSession);
1811	if (!rc)
1812	{
1813	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
1814	if (paPages)
1815	{
1816	uint32_t iPage = cPages;
1817	while (iPage-- > 0)
1818	{
1819	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
1820	Assert(paPages[iPage] != NIL_RTHCPHYS);
1821	}
1822	}
1823	return VINF_SUCCESS;
1824	}
1825	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
1826	AssertRC(rc2);
1827	}
1828
1829	rc2 = RTR0MemObjFree(Mem.MemObj, false);
1830	AssertRC(rc2);
1831	}
1832	return rc;
1833	}
1834
1835
1836	#ifdef RT_OS_WINDOWS
1837	/**
1838	* Check if the pages were locked by SUPR0PageAlloc
1839	*
1840	* This function will be removed along with the lock/unlock hacks when
1841	* we've cleaned up the ring-3 code properly.
1842	*
1843	* @returns boolean
1844	* @param pSession The session to which the memory was allocated.
1845	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc().
1846	*/
1847	static bool supdrvPageWasLockedByPageAlloc(PSUPDRVSESSION pSession, RTR3PTR pvR3)
1848	{
1849	PSUPDRVBUNDLE pBundle;
1850	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1851	dprintf(("SUPR0PageIsLockedByPageAlloc: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
1852
1853	/*
1854	* Search for the address.
1855	*/
1856	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
1857	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
1858	{
1859	if (pBundle->cUsed > 0)
1860	{
1861	unsigned i;
1862	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
1863	{
1864	if ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED_SUP
1865	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
1866	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
1867	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
1868	{
1869	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
1870	return true;
1871	}
1872	}
1873	}
1874	}
1875	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
1876	return false;
1877	}
1878
1879
1880	/**
1881	* Get the physical addresses of memory allocated using SUPR0PageAlloc().
1882	*
1883	* This function will be removed along with the lock/unlock hacks when
1884	* we've cleaned up the ring-3 code properly.
1885	*
1886	* @returns IPRT status code.
1887	* @param pSession The session to which the memory was allocated.
1888	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc().
1889	* @param cPages Number of pages in paPages
1890	* @param paPages Where to store the physical addresses.
1891	*/
1892	static int supdrvPageGetPhys(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
1893	{
1894	PSUPDRVBUNDLE pBundle;
1895	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1896	dprintf(("supdrvPageGetPhys: pSession=%p pvR3=%p cPages=%#lx paPages=%p\n", pSession, (void *)pvR3, (long)cPages, paPages));
1897
1898	/*
1899	* Search for the address.
1900	*/
1901	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
1902	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
1903	{
1904	if (pBundle->cUsed > 0)
1905	{
1906	unsigned i;
1907	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
1908	{
1909	if ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED_SUP
1910	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
1911	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
1912	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
1913	{
1914	uint32_t iPage = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
1915	cPages = RT_MIN(iPage, cPages);
1916	for (iPage = 0; iPage < cPages; iPage++)
1917	paPages[iPage] = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
1918	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
1919	return VINF_SUCCESS;
1920	}
1921	}
1922	}
1923	}
1924	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
1925	return VERR_INVALID_PARAMETER;
1926	}
1927	#endif /* RT_OS_WINDOWS */
1928
1929
1930	/**
1931	* Free memory allocated by SUPR0PageAlloc().
1932	*
1933	* @returns IPRT status code.
1934	* @param pSession The session owning the allocation.
1935	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc().
1936	*/
1937	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
1938	{
1939	dprintf(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
1940	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1941	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED_SUP);
1942	}
1943
1944
1945	/**
1946	* Maps the GIP into userspace and/or get the physical address of the GIP.
1947	*
1948	* @returns IPRT status code.
1949	* @param pSession Session to which the GIP mapping should belong.
1950	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
1951	* @param pHCPhysGip Where to store the physical address. (optional)
1952	*
1953	* @remark There is no reference counting on the mapping, so one call to this function
1954	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
1955	* and remove the session as a GIP user.
1956	*/
1957	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
1958	{
1959	int rc = 0;
1960	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1961	RTR3PTR pGip = NIL_RTR3PTR;
1962	RTHCPHYS HCPhys = NIL_RTHCPHYS;
1963	dprintf(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
1964
1965	/*
1966	* Validate
1967	*/
1968	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1969	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
1970	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
1971
1972	RTSemFastMutexRequest(pDevExt->mtxGip);
1973	if (pDevExt->pGip)
1974	{
1975	/*
1976	* Map it?
1977	*/
1978	if (ppGipR3)
1979	{
1980	#ifdef USE_NEW_OS_INTERFACE_FOR_GIP
1981	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
1982	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
1983	RTMEM_PROT_READ, RTR0ProcHandleSelf());
1984	if (RT_SUCCESS(rc))
1985	{
1986	pGip = RTR0MemObjAddressR3(pSession->GipMapObjR3);
1987	rc = VINF_SUCCESS; /** @todo remove this and replace the !rc below with RT_SUCCESS(rc). */
1988	}
1989	#else /* !USE_NEW_OS_INTERFACE_FOR_GIP */
1990	if (!pSession->pGip)
1991	rc = supdrvOSGipMap(pSession->pDevExt, &pSession->pGip);
1992	if (!rc)
1993	pGip = (RTR3PTR)pSession->pGip;
1994	#endif /* !USE_NEW_OS_INTERFACE_FOR_GIP */
1995	}
1996
1997	/*
1998	* Get physical address.
1999	*/
2000	if (pHCPhysGip && !rc)
2001	HCPhys = pDevExt->HCPhysGip;
2002
2003	/*
2004	* Reference globally.
2005	*/
2006	if (!pSession->fGipReferenced && !rc)
2007	{
2008	pSession->fGipReferenced = 1;
2009	pDevExt->cGipUsers++;
2010	if (pDevExt->cGipUsers == 1)
2011	{
2012	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
2013	unsigned i;
2014
2015	dprintf(("SUPR0GipMap: Resumes GIP updating\n"));
2016
2017	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
2018	ASMAtomicXchgU32(&pGip->aCPUs[i].u32TransactionId, pGip->aCPUs[i].u32TransactionId & ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
2019	ASMAtomicXchgU64(&pGip->u64NanoTSLastUpdateHz, 0);
2020
2021	#ifdef USE_NEW_OS_INTERFACE_FOR_GIP
2022	rc = RTTimerStart(pDevExt->pGipTimer, 0);
2023	AssertRC(rc); rc = VINF_SUCCESS;
2024	#else
2025	supdrvOSGipResume(pDevExt);
2026	#endif
2027	}
2028	}
2029	}
2030	else
2031	{
2032	rc = SUPDRV_ERR_GENERAL_FAILURE;
2033	dprintf(("SUPR0GipMap: GIP is not available!\n"));
2034	}
2035	RTSemFastMutexRelease(pDevExt->mtxGip);
2036
2037	/*
2038	* Write returns.
2039	*/
2040	if (pHCPhysGip)
2041	*pHCPhysGip = HCPhys;
2042	if (ppGipR3)
2043	*ppGipR3 = pGip;
2044
2045	#ifdef DEBUG_DARWIN_GIP
2046	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx ppGip=%p GipMapObjR3\n", rc, (unsigned long)HCPhys, pGip, pSession->GipMapObjR3));
2047	#else
2048	dprintf(("SUPR0GipMap: returns %d pHCPhysGip=%lx ppGipR3=%p\n", rc, (unsigned long)HCPhys, (void *)(uintptr_t)pGip));
2049	#endif
2050	return rc;
2051	}
2052
2053
2054	/**
2055	* Unmaps any user mapping of the GIP and terminates all GIP access
2056	* from this session.
2057	*
2058	* @returns IPRT status code.
2059	* @param pSession Session to which the GIP mapping should belong.
2060	*/
2061	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
2062	{
2063	int rc = VINF_SUCCESS;
2064	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2065	#ifdef DEBUG_DARWIN_GIP
2066	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
2067	pSession,
2068	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
2069	pSession->GipMapObjR3));
2070	#else
2071	dprintf(("SUPR0GipUnmap: pSession=%p\n", pSession));
2072	#endif
2073	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2074
2075	RTSemFastMutexRequest(pDevExt->mtxGip);
2076
2077	/*
2078	* Unmap anything?
2079	*/
2080	#ifdef USE_NEW_OS_INTERFACE_FOR_GIP
2081	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
2082	{
2083	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
2084	AssertRC(rc);
2085	if (RT_SUCCESS(rc))
2086	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
2087	}
2088	#else
2089	if (pSession->pGip)
2090	{
2091	rc = supdrvOSGipUnmap(pDevExt, pSession->pGip);
2092	if (!rc)
2093	pSession->pGip = NULL;
2094	}
2095	#endif
2096
2097	/*
2098	* Dereference global GIP.
2099	*/
2100	if (pSession->fGipReferenced && !rc)
2101	{
2102	pSession->fGipReferenced = 0;
2103	if ( pDevExt->cGipUsers > 0
2104	&& !--pDevExt->cGipUsers)
2105	{
2106	dprintf(("SUPR0GipUnmap: Suspends GIP updating\n"));
2107	#ifdef USE_NEW_OS_INTERFACE_FOR_GIP
2108	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = 0;
2109	#else
2110	supdrvOSGipSuspend(pDevExt);
2111	#endif
2112	}
2113	}
2114
2115	RTSemFastMutexRelease(pDevExt->mtxGip);
2116
2117	return rc;
2118	}
2119
2120
2121	/**
2122	* Adds a memory object to the session.
2123	*
2124	* @returns IPRT status code.
2125	* @param pMem Memory tracking structure containing the
2126	* information to track.
2127	* @param pSession The session.
2128	*/
2129	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
2130	{
2131	PSUPDRVBUNDLE pBundle;
2132	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2133
2134	/*
2135	* Find free entry and record the allocation.
2136	*/
2137	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2138	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2139	{
2140	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
2141	{
2142	unsigned i;
2143	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2144	{
2145	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
2146	{
2147	pBundle->cUsed++;
2148	pBundle->aMem[i] = *pMem;
2149	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2150	return VINF_SUCCESS;
2151	}
2152	}
2153	AssertFailed(); /* !!this can't be happening!!! */
2154	}
2155	}
2156	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2157
2158	/*
2159	* Need to allocate a new bundle.
2160	* Insert into the last entry in the bundle.
2161	*/
2162	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
2163	if (!pBundle)
2164	return VERR_NO_MEMORY;
2165
2166	/* take last entry. */
2167	pBundle->cUsed++;
2168	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
2169
2170	/* insert into list. */
2171	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2172	pBundle->pNext = pSession->Bundle.pNext;
2173	pSession->Bundle.pNext = pBundle;
2174	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2175
2176	return VINF_SUCCESS;
2177	}
2178
2179
2180	/**
2181	* Releases a memory object referenced by pointer and type.
2182	*
2183	* @returns IPRT status code.
2184	* @param pSession Session data.
2185	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
2186	* @param eType Memory type.
2187	*/
2188	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
2189	{
2190	PSUPDRVBUNDLE pBundle;
2191	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2192
2193	/*
2194	* Validate input.
2195	*/
2196	if (!uPtr)
2197	{
2198	dprintf(("Illegal address %p\n", (void *)uPtr));
2199	return VERR_INVALID_PARAMETER;
2200	}
2201
2202	/*
2203	* Search for the address.
2204	*/
2205	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2206	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2207	{
2208	if (pBundle->cUsed > 0)
2209	{
2210	unsigned i;
2211	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2212	{
2213	if ( pBundle->aMem[i].eType == eType
2214	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2215	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2216	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2217	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
2218	)
2219	{
2220	/* Make a copy of it and release it outside the spinlock. */
2221	SUPDRVMEMREF Mem = pBundle->aMem[i];
2222	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
2223	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
2224	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
2225	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2226
2227	if (Mem.MapObjR3)
2228	{
2229	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
2230	AssertRC(rc); /** @todo figure out how to handle this. */
2231	}
2232	if (Mem.MemObj)
2233	{
2234	int rc = RTR0MemObjFree(Mem.MemObj, false);
2235	AssertRC(rc); /** @todo figure out how to handle this. */
2236	}
2237	return VINF_SUCCESS;
2238	}
2239	}
2240	}
2241	}
2242	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2243	dprintf(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
2244	return VERR_INVALID_PARAMETER;
2245	}
2246
2247
2248	#ifdef VBOX_WITH_IDT_PATCHING
2249	/**
2250	* Install IDT for the current CPU.
2251	*
2252	* @returns One of the following IPRT status codes:
2253	* @retval VINF_SUCCESS on success.
2254	* @retval VERR_IDT_FAILED.
2255	* @retval VERR_NO_MEMORY.
2256	* @param pDevExt The device extension.
2257	* @param pSession The session data.
2258	* @param pReq The request.
2259	*/
2260	static int supdrvIOCtl_IdtInstall(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPIDTINSTALL pReq)
2261	{
2262	PSUPDRVPATCHUSAGE pUsagePre;
2263	PSUPDRVPATCH pPatchPre;
2264	RTIDTR Idtr;
2265	PSUPDRVPATCH pPatch;
2266	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2267	dprintf(("supdrvIOCtl_IdtInstall\n"));
2268
2269	/*
2270	* Preallocate entry for this CPU cause we don't wanna do
2271	* that inside the spinlock!
2272	*/
2273	pUsagePre = (PSUPDRVPATCHUSAGE)RTMemAlloc(sizeof(*pUsagePre));
2274	if (!pUsagePre)
2275	return VERR_NO_MEMORY;
2276
2277	/*
2278	* Take the spinlock and see what we need to do.
2279	*/
2280	RTSpinlockAcquireNoInts(pDevExt->Spinlock, &SpinlockTmp);
2281
2282	/* check if we already got a free patch. */
2283	if (!pDevExt->pIdtPatchesFree)
2284	{
2285	/*
2286	* Allocate a patch - outside the spinlock of course.
2287	*/
2288	RTSpinlockReleaseNoInts(pDevExt->Spinlock, &SpinlockTmp);
2289
2290	pPatchPre = (PSUPDRVPATCH)RTMemExecAlloc(sizeof(*pPatchPre));
2291	if (!pPatchPre)
2292	return VERR_NO_MEMORY;
2293
2294	RTSpinlockAcquireNoInts(pDevExt->Spinlock, &SpinlockTmp);
2295	}
2296	else
2297	{
2298	pPatchPre = pDevExt->pIdtPatchesFree;
2299	pDevExt->pIdtPatchesFree = pPatchPre->pNext;
2300	}
2301
2302	/* look for matching patch entry */
2303	ASMGetIDTR(&Idtr);
2304	pPatch = pDevExt->pIdtPatches;
2305	while (pPatch && pPatch->pvIdt != (void *)Idtr.pIdt)
2306	pPatch = pPatch->pNext;
2307
2308	if (!pPatch)
2309	{
2310	/*
2311	* Create patch.
2312	*/
2313	pPatch = supdrvIdtPatchOne(pDevExt, pPatchPre);
2314	if (pPatch)
2315	pPatchPre = NULL; /* mark as used. */
2316	}
2317	else
2318	{
2319	/*
2320	* Simply increment patch usage.
2321	*/
2322	pPatch->cUsage++;
2323	}
2324
2325	if (pPatch)
2326	{
2327	/*
2328	* Increment and add if need be the session usage record for this patch.
2329	*/
2330	PSUPDRVPATCHUSAGE pUsage = pSession->pPatchUsage;
2331	while (pUsage && pUsage->pPatch != pPatch)
2332	pUsage = pUsage->pNext;
2333
2334	if (!pUsage)
2335	{
2336	/*
2337	* Add usage record.
2338	*/
2339	pUsagePre->cUsage = 1;
2340	pUsagePre->pPatch = pPatch;
2341	pUsagePre->pNext = pSession->pPatchUsage;
2342	pSession->pPatchUsage = pUsagePre;
2343	pUsagePre = NULL; /* mark as used. */
2344	}
2345	else
2346	{
2347	/*
2348	* Increment usage count.
2349	*/
2350	pUsage->cUsage++;
2351	}
2352	}
2353
2354	/* free patch - we accumulate them for paranoid saftly reasons. */
2355	if (pPatchPre)
2356	{
2357	pPatchPre->pNext = pDevExt->pIdtPatchesFree;
2358	pDevExt->pIdtPatchesFree = pPatchPre;
2359	}
2360
2361	RTSpinlockReleaseNoInts(pDevExt->Spinlock, &SpinlockTmp);
2362
2363	/*
2364	* Free unused preallocated buffers.
2365	*/
2366	if (pUsagePre)
2367	RTMemFree(pUsagePre);
2368
2369	pReq->u.Out.u8Idt = pDevExt->u8Idt;
2370
2371	return pPatch ? VINF_SUCCESS : VERR_IDT_FAILED;
2372	}
2373
2374
2375	/**
2376	* This creates a IDT patch entry.
2377	* If the first patch being installed it'll also determin the IDT entry
2378	* to use.
2379	*
2380	* @returns pPatch on success.
2381	* @returns NULL on failure.
2382	* @param pDevExt Pointer to globals.
2383	* @param pPatch Patch entry to use.
2384	* This will be linked into SUPDRVDEVEXT::pIdtPatches on
2385	* successful return.
2386	* @remark Call must be owning the SUPDRVDEVEXT::Spinlock!
2387	*/
2388	static PSUPDRVPATCH supdrvIdtPatchOne(PSUPDRVDEVEXT pDevExt, PSUPDRVPATCH pPatch)
2389	{
2390	RTIDTR Idtr;
2391	PSUPDRVIDTE paIdt;
2392	dprintf(("supdrvIOCtl_IdtPatchOne: pPatch=%p\n", pPatch));
2393
2394	/*
2395	* Get IDT.
2396	*/
2397	ASMGetIDTR(&Idtr);
2398	paIdt = (PSUPDRVIDTE)Idtr.pIdt;
2399	/*
2400	* Recent Linux kernels can be configured to 1G user /3G kernel.
2401	*/
2402	if ((uintptr_t)paIdt < 0x40000000)
2403	{
2404	AssertMsgFailed(("bad paIdt=%p\n", paIdt));
2405	return NULL;
2406	}
2407
2408	if (!pDevExt->u8Idt)
2409	{
2410	/*
2411	* Test out the alternatives.
2412	*
2413	* At the moment we do not support chaining thus we ASSUME that one of
2414	* these 48 entries is unused (which is not a problem on Win32 and
2415	* Linux to my knowledge).
2416	*/
2417	/** @todo we MUST change this detection to try grab an entry which is NOT in use. This can be
2418	* combined with gathering info about which guest system call gates we can hook up directly. */
2419	unsigned i;
2420	uint8_t u8Idt = 0;
2421	static uint8_t au8Ints[] =
2422	{
2423	#ifdef RT_OS_WINDOWS /* We don't use 0xef and above because they are system stuff on linux (ef is IPI,
2424	* local apic timer, or some other frequently fireing thing). */
2425	0xef, 0xee, 0xed, 0xec,
2426	#endif
2427	0xeb, 0xea, 0xe9, 0xe8,
2428	0xdf, 0xde, 0xdd, 0xdc,
2429	0x7b, 0x7a, 0x79, 0x78,
2430	0xbf, 0xbe, 0xbd, 0xbc,
2431	};
2432	#if defined(RT_ARCH_AMD64) && defined(DEBUG)
2433	static int s_iWobble = 0;
2434	unsigned iMax = !(s_iWobble++ % 2) ? 0x80 : 0x100;
2435	dprintf(("IDT: Idtr=%p:%#x\n", (void *)Idtr.pIdt, (unsigned)Idtr.cbIdt));
2436	for (i = iMax - 0x80; i*16+15 < Idtr.cbIdt && i < iMax; i++)
2437	{
2438	dprintf(("%#x: %04x:%08x%04x%04x P=%d DPL=%d IST=%d Type1=%#x u32Reserved=%#x u5Reserved=%#x\n",
2439	i, paIdt[i].u16SegSel, paIdt[i].u32OffsetTop, paIdt[i].u16OffsetHigh, paIdt[i].u16OffsetLow,
2440	paIdt[i].u1Present, paIdt[i].u2DPL, paIdt[i].u3IST, paIdt[i].u5Type2,
2441	paIdt[i].u32Reserved, paIdt[i].u5Reserved));
2442	}
2443	#endif
2444	/* look for entries which are not present or otherwise unused. */
2445	for (i = 0; i < sizeof(au8Ints) / sizeof(au8Ints[0]); i++)
2446	{
2447	u8Idt = au8Ints[i];
2448	if ( u8Idt * sizeof(SUPDRVIDTE) < Idtr.cbIdt
2449	&& ( !paIdt[u8Idt].u1Present
2450	\|\| paIdt[u8Idt].u5Type2 == 0))
2451	break;
2452	u8Idt = 0;
2453	}
2454	if (!u8Idt)
2455	{
2456	/* try again, look for a compatible entry .*/
2457	for (i = 0; i < sizeof(au8Ints) / sizeof(au8Ints[0]); i++)
2458	{
2459	u8Idt = au8Ints[i];
2460	if ( u8Idt * sizeof(SUPDRVIDTE) < Idtr.cbIdt
2461	&& paIdt[u8Idt].u1Present
2462	&& paIdt[u8Idt].u5Type2 == SUPDRV_IDTE_TYPE2_INTERRUPT_GATE
2463	&& !(paIdt[u8Idt].u16SegSel & 3))
2464	break;
2465	u8Idt = 0;
2466	}
2467	if (!u8Idt)
2468	{
2469	dprintf(("Failed to find appropirate IDT entry!!\n"));
2470	return NULL;
2471	}
2472	}
2473	pDevExt->u8Idt = u8Idt;
2474	dprintf(("supdrvIOCtl_IdtPatchOne: u8Idt=%x\n", u8Idt));
2475	}
2476
2477	/*
2478	* Prepare the patch
2479	*/
2480	memset(pPatch, 0, sizeof(*pPatch));
2481	pPatch->pvIdt = paIdt;
2482	pPatch->cUsage = 1;
2483	pPatch->pIdtEntry = &paIdt[pDevExt->u8Idt];
2484	pPatch->SavedIdt = paIdt[pDevExt->u8Idt];
2485	pPatch->ChangedIdt.u16OffsetLow = (uint32_t)((uintptr_t)&pPatch->auCode[0] & 0xffff);
2486	pPatch->ChangedIdt.u16OffsetHigh = (uint32_t)((uintptr_t)&pPatch->auCode[0] >> 16);
2487	#ifdef RT_ARCH_AMD64
2488	pPatch->ChangedIdt.u32OffsetTop = (uint32_t)((uintptr_t)&pPatch->auCode[0] >> 32);
2489	#endif
2490	pPatch->ChangedIdt.u16SegSel = ASMGetCS();
2491	#ifdef RT_ARCH_AMD64
2492	pPatch->ChangedIdt.u3IST = 0;
2493	pPatch->ChangedIdt.u5Reserved = 0;
2494	#else /* x86 */
2495	pPatch->ChangedIdt.u5Reserved = 0;
2496	pPatch->ChangedIdt.u3Type1 = 0;
2497	#endif /* x86 */
2498	pPatch->ChangedIdt.u5Type2 = SUPDRV_IDTE_TYPE2_INTERRUPT_GATE;
2499	pPatch->ChangedIdt.u2DPL = 3;
2500	pPatch->ChangedIdt.u1Present = 1;
2501
2502	/*
2503	* Generate the patch code.
2504	*/
2505	{
2506	#ifdef RT_ARCH_AMD64
2507	union
2508	{
2509	uint8_t *pb;
2510	uint32_t *pu32;
2511	uint64_t *pu64;
2512	} u, uFixJmp, uFixCall, uNotNested;
2513	u.pb = &pPatch->auCode[0];
2514
2515	/* check the cookie */
2516	*u.pb++ = 0x3d; // cmp eax, GLOBALCOOKIE
2517	*u.pu32++ = pDevExt->u32Cookie;
2518
2519	*u.pb++ = 0x74; // jz @VBoxCall
2520	*u.pb++ = 2;
2521
2522	/* jump to forwarder code. */
2523	*u.pb++ = 0xeb;
2524	uFixJmp = u;
2525	*u.pb++ = 0xfe;
2526
2527	// @VBoxCall:
2528	*u.pb++ = 0x0f; // swapgs
2529	*u.pb++ = 0x01;
2530	*u.pb++ = 0xf8;
2531
2532	/*
2533	* Call VMMR0Entry
2534	* We don't have to push the arguments here, but we have top
2535	* reserve some stack space for the interrupt forwarding.
2536	*/
2537	# ifdef RT_OS_WINDOWS
2538	*u.pb++ = 0x50; // push rax ; alignment filler.
2539	*u.pb++ = 0x41; // push r8 ; uArg
2540	*u.pb++ = 0x50;
2541	*u.pb++ = 0x52; // push rdx ; uOperation
2542	*u.pb++ = 0x51; // push rcx ; pVM
2543	# else
2544	*u.pb++ = 0x51; // push rcx ; alignment filler.
2545	*u.pb++ = 0x52; // push rdx ; uArg
2546	*u.pb++ = 0x56; // push rsi ; uOperation
2547	*u.pb++ = 0x57; // push rdi ; pVM
2548	# endif
2549
2550	*u.pb++ = 0xff; // call qword [pfnVMMR0EntryInt wrt rip]
2551	*u.pb++ = 0x15;
2552	uFixCall = u;
2553	*u.pu32++ = 0;
2554
2555	*u.pb++ = 0x48; // add rsp, 20h ; remove call frame.
2556	*u.pb++ = 0x81;
2557	*u.pb++ = 0xc4;
2558	*u.pu32++ = 0x20;
2559
2560	*u.pb++ = 0x0f; // swapgs
2561	*u.pb++ = 0x01;
2562	*u.pb++ = 0xf8;
2563
2564	/* Return to R3. */
2565	uNotNested = u;
2566	*u.pb++ = 0x48; // iretq
2567	*u.pb++ = 0xcf;
2568
2569	while ((uintptr_t)u.pb & 0x7) // align 8
2570	*u.pb++ = 0xcc;
2571
2572	/* Pointer to the VMMR0Entry. */ // pfnVMMR0EntryInt dq StubVMMR0Entry
2573	*uFixCall.pu32 = (uint32_t)(u.pb - uFixCall.pb - 4); uFixCall.pb = NULL;
2574	pPatch->offVMMR0EntryFixup = (uint16_t)(u.pb - &pPatch->auCode[0]);
2575	*u.pu64++ = pDevExt->pvVMMR0 ? (uint64_t)pDevExt->pfnVMMR0EntryInt : (uint64_t)u.pb + 8;
2576
2577	/* stub entry. */ // StubVMMR0Entry:
2578	pPatch->offStub = (uint16_t)(u.pb - &pPatch->auCode[0]);
2579	*u.pb++ = 0x33; // xor eax, eax
2580	*u.pb++ = 0xc0;
2581
2582	*u.pb++ = 0x48; // dec rax
2583	*u.pb++ = 0xff;
2584	*u.pb++ = 0xc8;
2585
2586	*u.pb++ = 0xc3; // ret
2587
2588	/* forward to the original handler using a retf. */
2589	*uFixJmp.pb = (uint8_t)(u.pb - uFixJmp.pb - 1); uFixJmp.pb = NULL;
2590
2591	*u.pb++ = 0x68; // push <target cs>
2592	*u.pu32++ = !pPatch->SavedIdt.u5Type2 ? ASMGetCS() : pPatch->SavedIdt.u16SegSel;
2593
2594	*u.pb++ = 0x68; // push <low target rip>
2595	*u.pu32++ = !pPatch->SavedIdt.u5Type2
2596	? (uint32_t)(uintptr_t)uNotNested.pb
2597	: (uint32_t)pPatch->SavedIdt.u16OffsetLow
2598	\| (uint32_t)pPatch->SavedIdt.u16OffsetHigh << 16;
2599
2600	*u.pb++ = 0xc7; // mov dword [rsp + 4], <high target rip>
2601	*u.pb++ = 0x44;
2602	*u.pb++ = 0x24;
2603	*u.pb++ = 0x04;
2604	*u.pu32++ = !pPatch->SavedIdt.u5Type2
2605	? (uint32_t)((uint64_t)uNotNested.pb >> 32)
2606	: pPatch->SavedIdt.u32OffsetTop;
2607
2608	*u.pb++ = 0x48; // retf ; does this require prefix?
2609	*u.pb++ = 0xcb;
2610
2611	#else /* RT_ARCH_X86 */
2612
2613	union
2614	{
2615	uint8_t *pb;
2616	uint16_t *pu16;
2617	uint32_t *pu32;
2618	} u, uFixJmpNotNested, uFixJmp, uFixCall, uNotNested;
2619	u.pb = &pPatch->auCode[0];
2620
2621	/* check the cookie */
2622	*u.pb++ = 0x81; // cmp esi, GLOBALCOOKIE
2623	*u.pb++ = 0xfe;
2624	*u.pu32++ = pDevExt->u32Cookie;
2625
2626	*u.pb++ = 0x74; // jz VBoxCall
2627	uFixJmp = u;
2628	*u.pb++ = 0;
2629
2630	/* jump (far) to the original handler / not-nested-stub. */
2631	*u.pb++ = 0xea; // jmp far NotNested
2632	uFixJmpNotNested = u;
2633	*u.pu32++ = 0;
2634	*u.pu16++ = 0;
2635
2636	/* save selector registers. */ // VBoxCall:
2637	*uFixJmp.pb = (uint8_t)(u.pb - uFixJmp.pb - 1);
2638	*u.pb++ = 0x0f; // push fs
2639	*u.pb++ = 0xa0;
2640
2641	*u.pb++ = 0x1e; // push ds
2642
2643	*u.pb++ = 0x06; // push es
2644
2645	/* call frame */
2646	*u.pb++ = 0x51; // push ecx
2647
2648	*u.pb++ = 0x52; // push edx
2649
2650	*u.pb++ = 0x50; // push eax
2651
2652	/* load ds, es and perhaps fs before call. */
2653	*u.pb++ = 0xb8; // mov eax, KernelDS
2654	*u.pu32++ = ASMGetDS();
2655
2656	*u.pb++ = 0x8e; // mov ds, eax
2657	*u.pb++ = 0xd8;
2658
2659	*u.pb++ = 0x8e; // mov es, eax
2660	*u.pb++ = 0xc0;
2661
2662	#ifdef RT_OS_WINDOWS
2663	*u.pb++ = 0xb8; // mov eax, KernelFS
2664	*u.pu32++ = ASMGetFS();
2665
2666	*u.pb++ = 0x8e; // mov fs, eax
2667	*u.pb++ = 0xe0;
2668	#endif
2669
2670	/* do the call. */
2671	*u.pb++ = 0xe8; // call _VMMR0Entry / StubVMMR0Entry
2672	uFixCall = u;
2673	pPatch->offVMMR0EntryFixup = (uint16_t)(u.pb - &pPatch->auCode[0]);
2674	*u.pu32++ = 0xfffffffb;
2675
2676	*u.pb++ = 0x83; // add esp, 0ch ; cdecl
2677	*u.pb++ = 0xc4;
2678	*u.pb++ = 0x0c;
2679
2680	/* restore selector registers. */
2681	*u.pb++ = 0x07; // pop es
2682	//
2683	*u.pb++ = 0x1f; // pop ds
2684
2685	*u.pb++ = 0x0f; // pop fs
2686	*u.pb++ = 0xa1;
2687
2688	uNotNested = u; // NotNested:
2689	*u.pb++ = 0xcf; // iretd
2690
2691	/* the stub VMMR0Entry. */ // StubVMMR0Entry:
2692	pPatch->offStub = (uint16_t)(u.pb - &pPatch->auCode[0]);
2693	*u.pb++ = 0x33; // xor eax, eax
2694	*u.pb++ = 0xc0;
2695
2696	*u.pb++ = 0x48; // dec eax
2697
2698	*u.pb++ = 0xc3; // ret
2699
2700	/* Fixup the VMMR0Entry call. */
2701	if (pDevExt->pvVMMR0)
2702	*uFixCall.pu32 = (uint32_t)pDevExt->pfnVMMR0EntryInt - (uint32_t)(uFixCall.pu32 + 1);
2703	else
2704	*uFixCall.pu32 = (uint32_t)&pPatch->auCode[pPatch->offStub] - (uint32_t)(uFixCall.pu32 + 1);
2705
2706	/* Fixup the forward / nested far jump. */
2707	if (!pPatch->SavedIdt.u5Type2)
2708	{
2709	*uFixJmpNotNested.pu32++ = (uint32_t)uNotNested.pb;
2710	*uFixJmpNotNested.pu16++ = ASMGetCS();
2711	}
2712	else
2713	{
2714	*uFixJmpNotNested.pu32++ = ((uint32_t)pPatch->SavedIdt.u16OffsetHigh << 16) \| pPatch->SavedIdt.u16OffsetLow;
2715	*uFixJmpNotNested.pu16++ = pPatch->SavedIdt.u16SegSel;
2716	}
2717	#endif /* RT_ARCH_X86 */
2718	Assert(u.pb <= &pPatch->auCode[sizeof(pPatch->auCode)]);
2719	#if 0
2720	/* dump the patch code */
2721	dprintf(("patch code: %p\n", &pPatch->auCode[0]));
2722	for (uFixCall.pb = &pPatch->auCode[0]; uFixCall.pb < u.pb; uFixCall.pb++)
2723	dprintf(("0x%02x,\n", *uFixCall.pb));
2724	#endif
2725	}
2726
2727	/*
2728	* Install the patch.
2729	*/
2730	supdrvIdtWrite(pPatch->pIdtEntry, &pPatch->ChangedIdt);
2731	AssertMsg(!memcmp((void *)pPatch->pIdtEntry, &pPatch->ChangedIdt, sizeof(pPatch->ChangedIdt)), ("The stupid change code didn't work!!!!!\n"));
2732
2733	/*
2734	* Link in the patch.
2735	*/
2736	pPatch->pNext = pDevExt->pIdtPatches;
2737	pDevExt->pIdtPatches = pPatch;
2738
2739	return pPatch;
2740	}
2741
2742
2743	/**
2744	* Removes the sessions IDT references.
2745	* This will uninstall our IDT patch if we left unreferenced.
2746	*
2747	* @returns VINF_SUCCESS.
2748	* @param pDevExt Device globals.
2749	* @param pSession Session data.
2750	*/
2751	static int supdrvIOCtl_IdtRemoveAll(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
2752	{
2753	PSUPDRVPATCHUSAGE pUsage;
2754	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2755	dprintf(("supdrvIOCtl_IdtRemoveAll: pSession=%p\n", pSession));
2756
2757	/*
2758	* Take the spinlock.
2759	*/
2760	RTSpinlockAcquireNoInts(pDevExt->Spinlock, &SpinlockTmp);
2761
2762	/*
2763	* Walk usage list, removing patches as their usage count reaches zero.
2764	*/
2765	pUsage = pSession->pPatchUsage;
2766	while (pUsage)
2767	{
2768	if (pUsage->pPatch->cUsage <= pUsage->cUsage)
2769	supdrvIdtRemoveOne(pDevExt, pUsage->pPatch);
2770	else
2771	pUsage->pPatch->cUsage -= pUsage->cUsage;
2772
2773	/* next */
2774	pUsage = pUsage->pNext;
2775	}
2776
2777	/*
2778	* Empty the usage chain and we're done inside the spinlock.
2779	*/
2780	pUsage = pSession->pPatchUsage;
2781	pSession->pPatchUsage = NULL;
2782
2783	RTSpinlockReleaseNoInts(pDevExt->Spinlock, &SpinlockTmp);
2784
2785	/*
2786	* Free usage entries.
2787	*/
2788	while (pUsage)
2789	{
2790	void *pvToFree = pUsage;
2791	pUsage->cUsage = 0;
2792	pUsage->pPatch = NULL;
2793	pUsage = pUsage->pNext;
2794	RTMemFree(pvToFree);
2795	}
2796
2797	return VINF_SUCCESS;
2798	}
2799
2800
2801	/**
2802	* Remove one patch.
2803	*
2804	* Worker for supdrvIOCtl_IdtRemoveAll.
2805	*
2806	* @param pDevExt Device globals.
2807	* @param pPatch Patch entry to remove.
2808	* @remark Caller must own SUPDRVDEVEXT::Spinlock!
2809	*/
2810	static void supdrvIdtRemoveOne(PSUPDRVDEVEXT pDevExt, PSUPDRVPATCH pPatch)
2811	{
2812	dprintf(("supdrvIdtRemoveOne: pPatch=%p\n", pPatch));
2813
2814	pPatch->cUsage = 0;
2815
2816	/*
2817	* If the IDT entry was changed it have to kick around for ever!
2818	* This will be attempted freed again, perhaps next time we'll succeed :-)
2819	*/
2820	if (memcmp((void *)pPatch->pIdtEntry, &pPatch->ChangedIdt, sizeof(pPatch->ChangedIdt)))
2821	{
2822	AssertMsgFailed(("The hijacked IDT entry has CHANGED!!!\n"));
2823	return;
2824	}
2825
2826	/*
2827	* Unlink it.
2828	*/
2829	if (pDevExt->pIdtPatches != pPatch)
2830	{
2831	PSUPDRVPATCH pPatchPrev = pDevExt->pIdtPatches;
2832	while (pPatchPrev)
2833	{
2834	if (pPatchPrev->pNext == pPatch)
2835	{
2836	pPatchPrev->pNext = pPatch->pNext;
2837	break;
2838	}
2839	pPatchPrev = pPatchPrev->pNext;
2840	}
2841	Assert(!pPatchPrev);
2842	}
2843	else
2844	pDevExt->pIdtPatches = pPatch->pNext;
2845	pPatch->pNext = NULL;
2846
2847
2848	/*
2849	* Verify and restore the IDT.
2850	*/
2851	AssertMsg(!memcmp((void *)pPatch->pIdtEntry, &pPatch->ChangedIdt, sizeof(pPatch->ChangedIdt)), ("The hijacked IDT entry has CHANGED!!!\n"));
2852	supdrvIdtWrite(pPatch->pIdtEntry, &pPatch->SavedIdt);
2853	AssertMsg(!memcmp((void *)pPatch->pIdtEntry, &pPatch->SavedIdt, sizeof(pPatch->SavedIdt)), ("The hijacked IDT entry has CHANGED!!!\n"));
2854
2855	/*
2856	* Put it in the free list.
2857	* (This free list stuff is to calm my paranoia.)
2858	*/
2859	pPatch->pvIdt = NULL;
2860	pPatch->pIdtEntry = NULL;
2861
2862	pPatch->pNext = pDevExt->pIdtPatchesFree;
2863	pDevExt->pIdtPatchesFree = pPatch;
2864	}
2865
2866
2867	/**
2868	* Write to an IDT entry.
2869	*
2870	* @param pvIdtEntry Where to write.
2871	* @param pNewIDTEntry What to write.
2872	*/
2873	static void supdrvIdtWrite(volatile void pvIdtEntry, const SUPDRVIDTE pNewIDTEntry)
2874	{
2875	RTUINTREG uCR0;
2876	RTUINTREG uFlags;
2877
2878	/*
2879	* On SMP machines (P4 hyperthreading included) we must preform a
2880	* 64-bit locked write when updating the IDT entry.
2881	*
2882	* The F00F bugfix for linux (and probably other OSes) causes
2883	* the IDT to be pointing to an readonly mapping. We get around that
2884	* by temporarily turning of WP. Since we're inside a spinlock at this
2885	* point, interrupts are disabled and there isn't any way the WP bit
2886	* flipping can cause any trouble.
2887	*/
2888
2889	/* Save & Clear interrupt flag; Save & clear WP. */
2890	uFlags = ASMGetFlags();
2891	ASMSetFlags(uFlags & ~(RTUINTREG)(1 << 9)); /X86_EFL_IF/
2892	Assert(!(ASMGetFlags() & (1 << 9)));
2893	uCR0 = ASMGetCR0();
2894	ASMSetCR0(uCR0 & ~(RTUINTREG)(1 << 16)); /X86_CR0_WP/
2895
2896	/* Update IDT Entry */
2897	#ifdef RT_ARCH_AMD64
2898	ASMAtomicXchgU128((volatile uint128_t )pvIdtEntry, (uint128_t *)(uintptr_t)pNewIDTEntry);
2899	#else
2900	ASMAtomicXchgU64((volatile uint64_t )pvIdtEntry, (uint64_t *)(uintptr_t)pNewIDTEntry);
2901	#endif
2902
2903	/* Restore CR0 & Flags */
2904	ASMSetCR0(uCR0);
2905	ASMSetFlags(uFlags);
2906	}
2907	#endif /* VBOX_WITH_IDT_PATCHING */
2908
2909
2910	/**
2911	* Opens an image. If it's the first time it's opened the call must upload
2912	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
2913	*
2914	* This is the 1st step of the loading.
2915	*
2916	* @returns IPRT status code.
2917	* @param pDevExt Device globals.
2918	* @param pSession Session data.
2919	* @param pReq The open request.
2920	*/
2921	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
2922	{
2923	PSUPDRVLDRIMAGE pImage;
2924	unsigned cb;
2925	void *pv;
2926	dprintf(("supdrvIOCtl_LdrOpen: szName=%s cbImage=%d\n", pReq->u.In.szName, pReq->u.In.cbImage));
2927
2928	/*
2929	* Check if we got an instance of the image already.
2930	*/
2931	RTSemFastMutexRequest(pDevExt->mtxLdr);
2932	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
2933	{
2934	if (!strcmp(pImage->szName, pReq->u.In.szName))
2935	{
2936	pImage->cUsage++;
2937	pReq->u.Out.pvImageBase = pImage->pvImage;
2938	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
2939	supdrvLdrAddUsage(pSession, pImage);
2940	RTSemFastMutexRelease(pDevExt->mtxLdr);
2941	return VINF_SUCCESS;
2942	}
2943	}
2944	/* (not found - add it!) */
2945
2946	/*
2947	* Allocate memory.
2948	*/
2949	cb = pReq->u.In.cbImage + sizeof(SUPDRVLDRIMAGE) + 31;
2950	pv = RTMemExecAlloc(cb);
2951	if (!pv)
2952	{
2953	RTSemFastMutexRelease(pDevExt->mtxLdr);
2954	return VERR_NO_MEMORY;
2955	}
2956
2957	/*
2958	* Setup and link in the LDR stuff.
2959	*/
2960	pImage = (PSUPDRVLDRIMAGE)pv;
2961	pImage->pvImage = ALIGNP(pImage + 1, 32);
2962	pImage->cbImage = pReq->u.In.cbImage;
2963	pImage->pfnModuleInit = NULL;
2964	pImage->pfnModuleTerm = NULL;
2965	pImage->uState = SUP_IOCTL_LDR_OPEN;
2966	pImage->cUsage = 1;
2967	strcpy(pImage->szName, pReq->u.In.szName);
2968
2969	pImage->pNext = pDevExt->pLdrImages;
2970	pDevExt->pLdrImages = pImage;
2971
2972	supdrvLdrAddUsage(pSession, pImage);
2973
2974	pReq->u.Out.pvImageBase = pImage->pvImage;
2975	pReq->u.Out.fNeedsLoading = true;
2976	RTSemFastMutexRelease(pDevExt->mtxLdr);
2977	return VINF_SUCCESS;
2978	}
2979
2980
2981	/**
2982	* Loads the image bits.
2983	*
2984	* This is the 2nd step of the loading.
2985	*
2986	* @returns IPRT status code.
2987	* @param pDevExt Device globals.
2988	* @param pSession Session data.
2989	* @param pReq The request.
2990	*/
2991	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
2992	{
2993	PSUPDRVLDRUSAGE pUsage;
2994	PSUPDRVLDRIMAGE pImage;
2995	int rc;
2996	dprintf(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImage=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImage));
2997
2998	/*
2999	* Find the ldr image.
3000	*/
3001	RTSemFastMutexRequest(pDevExt->mtxLdr);
3002	pUsage = pSession->pLdrUsage;
3003	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3004	pUsage = pUsage->pNext;
3005	if (!pUsage)
3006	{
3007	RTSemFastMutexRelease(pDevExt->mtxLdr);
3008	dprintf(("SUP_IOCTL_LDR_LOAD: couldn't find image!\n"));
3009	return VERR_INVALID_HANDLE;
3010	}
3011	pImage = pUsage->pImage;
3012	if (pImage->cbImage != pReq->u.In.cbImage)
3013	{
3014	RTSemFastMutexRelease(pDevExt->mtxLdr);
3015	dprintf(("SUP_IOCTL_LDR_LOAD: image size mismatch!! %d(prep) != %d(load)\n", pImage->cbImage, pReq->u.In.cbImage));
3016	return VERR_INVALID_HANDLE;
3017	}
3018	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
3019	{
3020	unsigned uState = pImage->uState;
3021	RTSemFastMutexRelease(pDevExt->mtxLdr);
3022	if (uState != SUP_IOCTL_LDR_LOAD)
3023	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
3024	return SUPDRV_ERR_ALREADY_LOADED;
3025	}
3026	switch (pReq->u.In.eEPType)
3027	{
3028	case SUPLDRLOADEP_NOTHING:
3029	break;
3030	case SUPLDRLOADEP_VMMR0:
3031	if ( !pReq->u.In.EP.VMMR0.pvVMMR0
3032	\|\| !pReq->u.In.EP.VMMR0.pvVMMR0EntryInt
3033	\|\| !pReq->u.In.EP.VMMR0.pvVMMR0EntryFast
3034	\|\| !pReq->u.In.EP.VMMR0.pvVMMR0EntryEx)
3035	{
3036	RTSemFastMutexRelease(pDevExt->mtxLdr);
3037	dprintf(("NULL pointer: pvVMMR0=%p pvVMMR0EntryInt=%p pvVMMR0EntryFast=%p pvVMMR0EntryEx=%p!\n",
3038	pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
3039	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx));
3040	return VERR_INVALID_PARAMETER;
3041	}
3042	if ( (uintptr_t)pReq->u.In.EP.VMMR0.pvVMMR0EntryInt - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage
3043	\|\| (uintptr_t)pReq->u.In.EP.VMMR0.pvVMMR0EntryFast - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage
3044	\|\| (uintptr_t)pReq->u.In.EP.VMMR0.pvVMMR0EntryEx - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage)
3045	{
3046	RTSemFastMutexRelease(pDevExt->mtxLdr);
3047	dprintf(("Out of range (%p LB %#x): pvVMMR0EntryInt=%p, pvVMMR0EntryFast=%p or pvVMMR0EntryEx=%p is NULL!\n",
3048	pImage->pvImage, pReq->u.In.cbImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
3049	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx));
3050	return VERR_INVALID_PARAMETER;
3051	}
3052	break;
3053	default:
3054	RTSemFastMutexRelease(pDevExt->mtxLdr);
3055	dprintf(("Invalid eEPType=%d\n", pReq->u.In.eEPType));
3056	return VERR_INVALID_PARAMETER;
3057	}
3058	if ( pReq->u.In.pfnModuleInit
3059	&& (uintptr_t)pReq->u.In.pfnModuleInit - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage)
3060	{
3061	RTSemFastMutexRelease(pDevExt->mtxLdr);
3062	dprintf(("SUP_IOCTL_LDR_LOAD: pfnModuleInit=%p is outside the image (%p %d bytes)\n",
3063	pReq->u.In.pfnModuleInit, pImage->pvImage, pReq->u.In.cbImage));
3064	return VERR_INVALID_PARAMETER;
3065	}
3066	if ( pReq->u.In.pfnModuleTerm
3067	&& (uintptr_t)pReq->u.In.pfnModuleTerm - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage)
3068	{
3069	RTSemFastMutexRelease(pDevExt->mtxLdr);
3070	dprintf(("SUP_IOCTL_LDR_LOAD: pfnModuleTerm=%p is outside the image (%p %d bytes)\n",
3071	pReq->u.In.pfnModuleTerm, pImage->pvImage, pReq->u.In.cbImage));
3072	return VERR_INVALID_PARAMETER;
3073	}
3074
3075	/*
3076	* Copy the memory.
3077	*/
3078	/* no need to do try/except as this is a buffered request. */
3079	memcpy(pImage->pvImage, &pReq->u.In.achImage[0], pImage->cbImage);
3080	pImage->uState = SUP_IOCTL_LDR_LOAD;
3081	pImage->pfnModuleInit = pReq->u.In.pfnModuleInit;
3082	pImage->pfnModuleTerm = pReq->u.In.pfnModuleTerm;
3083	pImage->offSymbols = pReq->u.In.offSymbols;
3084	pImage->cSymbols = pReq->u.In.cSymbols;
3085	pImage->offStrTab = pReq->u.In.offStrTab;
3086	pImage->cbStrTab = pReq->u.In.cbStrTab;
3087
3088	/*
3089	* Update any entry points.
3090	*/
3091	switch (pReq->u.In.eEPType)
3092	{
3093	default:
3094	case SUPLDRLOADEP_NOTHING:
3095	rc = VINF_SUCCESS;
3096	break;
3097	case SUPLDRLOADEP_VMMR0:
3098	rc = supdrvLdrSetR0EP(pDevExt, pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
3099	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
3100	break;
3101	}
3102
3103	/*
3104	* On success call the module initialization.
3105	*/
3106	dprintf(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
3107	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
3108	{
3109	dprintf(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
3110	rc = pImage->pfnModuleInit();
3111	if (rc && pDevExt->pvVMMR0 == pImage->pvImage)
3112	supdrvLdrUnsetR0EP(pDevExt);
3113	}
3114
3115	if (rc)
3116	pImage->uState = SUP_IOCTL_LDR_OPEN;
3117
3118	RTSemFastMutexRelease(pDevExt->mtxLdr);
3119	return rc;
3120	}
3121
3122
3123	/**
3124	* Frees a previously loaded (prep'ed) image.
3125	*
3126	* @returns IPRT status code.
3127	* @param pDevExt Device globals.
3128	* @param pSession Session data.
3129	* @param pReq The request.
3130	*/
3131	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
3132	{
3133	int rc;
3134	PSUPDRVLDRUSAGE pUsagePrev;
3135	PSUPDRVLDRUSAGE pUsage;
3136	PSUPDRVLDRIMAGE pImage;
3137	dprintf(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
3138
3139	/*
3140	* Find the ldr image.
3141	*/
3142	RTSemFastMutexRequest(pDevExt->mtxLdr);
3143	pUsagePrev = NULL;
3144	pUsage = pSession->pLdrUsage;
3145	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3146	{
3147	pUsagePrev = pUsage;
3148	pUsage = pUsage->pNext;
3149	}
3150	if (!pUsage)
3151	{
3152	RTSemFastMutexRelease(pDevExt->mtxLdr);
3153	dprintf(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
3154	return VERR_INVALID_HANDLE;
3155	}
3156
3157	/*
3158	* Check if we can remove anything.
3159	*/
3160	rc = VINF_SUCCESS;
3161	pImage = pUsage->pImage;
3162	if (pImage->cUsage <= 1 \|\| pUsage->cUsage <= 1)
3163	{
3164	/*
3165	* Check if there are any objects with destructors in the image, if
3166	* so leave it for the session cleanup routine so we get a chance to
3167	* clean things up in the right order and not leave them all dangling.
3168	*/
3169	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3170	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
3171	if (pImage->cUsage <= 1)
3172	{
3173	PSUPDRVOBJ pObj;
3174	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
3175	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImage))
3176	{
3177	rc = VERR_SHARING_VIOLATION; /** @todo VERR_DANGLING_OBJECTS */
3178	break;
3179	}
3180	}
3181	else
3182	{
3183	PSUPDRVUSAGE pGenUsage;
3184	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
3185	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImage))
3186	{
3187	rc = VERR_SHARING_VIOLATION; /** @todo VERR_DANGLING_OBJECTS */
3188	break;
3189	}
3190	}
3191	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
3192	if (rc == VINF_SUCCESS)
3193	{
3194	/* unlink it */
3195	if (pUsagePrev)
3196	pUsagePrev->pNext = pUsage->pNext;
3197	else
3198	pSession->pLdrUsage = pUsage->pNext;
3199
3200	/* free it */
3201	pUsage->pImage = NULL;
3202	pUsage->pNext = NULL;
3203	RTMemFree(pUsage);
3204
3205	/*
3206	* Derefrence the image.
3207	*/
3208	if (pImage->cUsage <= 1)
3209	supdrvLdrFree(pDevExt, pImage);
3210	else
3211	pImage->cUsage--;
3212	}
3213	}
3214	else
3215	{
3216	/*
3217	* Dereference both image and usage.
3218	*/
3219	pImage->cUsage--;
3220	pUsage->cUsage--;
3221	}
3222
3223	RTSemFastMutexRelease(pDevExt->mtxLdr);
3224	return VINF_SUCCESS;
3225	}
3226
3227
3228	/**
3229	* Gets the address of a symbol in an open image.
3230	*
3231	* @returns 0 on success.
3232	* @returns SUPDRV_ERR_* on failure.
3233	* @param pDevExt Device globals.
3234	* @param pSession Session data.
3235	* @param pReq The request buffer.
3236	*/
3237	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
3238	{
3239	PSUPDRVLDRIMAGE pImage;
3240	PSUPDRVLDRUSAGE pUsage;
3241	uint32_t i;
3242	PSUPLDRSYM paSyms;
3243	const char *pchStrings;
3244	const size_t cbSymbol = strlen(pReq->u.In.szSymbol) + 1;
3245	void *pvSymbol = NULL;
3246	int rc = VERR_GENERAL_FAILURE;
3247	dprintf2(("supdrvIOCtl_LdrGetSymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
3248
3249	/*
3250	* Find the ldr image.
3251	*/
3252	RTSemFastMutexRequest(pDevExt->mtxLdr);
3253	pUsage = pSession->pLdrUsage;
3254	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3255	pUsage = pUsage->pNext;
3256	if (!pUsage)
3257	{
3258	RTSemFastMutexRelease(pDevExt->mtxLdr);
3259	dprintf(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
3260	return VERR_INVALID_HANDLE;
3261	}
3262	pImage = pUsage->pImage;
3263	if (pImage->uState != SUP_IOCTL_LDR_LOAD)
3264	{
3265	unsigned uState = pImage->uState;
3266	RTSemFastMutexRelease(pDevExt->mtxLdr);
3267	dprintf(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", uState, uState)); NOREF(uState);
3268	return VERR_ALREADY_LOADED;
3269	}
3270
3271	/*
3272	* Search the symbol string.
3273	*/
3274	pchStrings = (const char )((uint8_t )pImage->pvImage + pImage->offStrTab);
3275	paSyms = (PSUPLDRSYM)((uint8_t *)pImage->pvImage + pImage->offSymbols);
3276	for (i = 0; i < pImage->cSymbols; i++)
3277	{
3278	if ( paSyms[i].offSymbol < pImage->cbImage /* paranoia */
3279	&& paSyms[i].offName + cbSymbol <= pImage->cbStrTab
3280	&& !memcmp(pchStrings + paSyms[i].offName, pReq->u.In.szSymbol, cbSymbol))
3281	{
3282	pvSymbol = (uint8_t *)pImage->pvImage + paSyms[i].offSymbol;
3283	rc = VINF_SUCCESS;
3284	break;
3285	}
3286	}
3287	RTSemFastMutexRelease(pDevExt->mtxLdr);
3288	pReq->u.Out.pvSymbol = pvSymbol;
3289	return rc;
3290	}
3291
3292
3293	/**
3294	* Updates the IDT patches to point to the specified VMM R0 entry
3295	* point (i.e. VMMR0Enter()).
3296	*
3297	* @returns IPRT status code.
3298	* @param pDevExt Device globals.
3299	* @param pSession Session data.
3300	* @param pVMMR0 VMMR0 image handle.
3301	* @param pvVMMR0EntryInt VMMR0EntryInt address.
3302	* @param pvVMMR0EntryFast VMMR0EntryFast address.
3303	* @param pvVMMR0EntryEx VMMR0EntryEx address.
3304	* @remark Caller must own the loader mutex.
3305	*/
3306	static int supdrvLdrSetR0EP(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx)
3307	{
3308	int rc = VINF_SUCCESS;
3309	dprintf(("supdrvLdrSetR0EP pvVMMR0=%p pvVMMR0EntryInt=%p\n", pvVMMR0, pvVMMR0EntryInt));
3310
3311
3312	/*
3313	* Check if not yet set.
3314	*/
3315	if (!pDevExt->pvVMMR0)
3316	{
3317	#ifdef VBOX_WITH_IDT_PATCHING
3318	PSUPDRVPATCH pPatch;
3319	#endif
3320
3321	/*
3322	* Set it and update IDT patch code.
3323	*/
3324	pDevExt->pvVMMR0 = pvVMMR0;
3325	pDevExt->pfnVMMR0EntryInt = pvVMMR0EntryInt;
3326	pDevExt->pfnVMMR0EntryFast = pvVMMR0EntryFast;
3327	pDevExt->pfnVMMR0EntryEx = pvVMMR0EntryEx;
3328	#ifdef VBOX_WITH_IDT_PATCHING
3329	for (pPatch = pDevExt->pIdtPatches; pPatch; pPatch = pPatch->pNext)
3330	{
3331	# ifdef RT_ARCH_AMD64
3332	ASMAtomicXchgU64((volatile uint64_t *)&pPatch->auCode[pPatch->offVMMR0EntryFixup], (uint64_t)pvVMMR0);
3333	# else /* RT_ARCH_X86 */
3334	ASMAtomicXchgU32((volatile uint32_t *)&pPatch->auCode[pPatch->offVMMR0EntryFixup],
3335	(uint32_t)pvVMMR0 - (uint32_t)&pPatch->auCode[pPatch->offVMMR0EntryFixup + 4]);
3336	# endif
3337	}
3338	#endif /* VBOX_WITH_IDT_PATCHING */
3339	}
3340	else
3341	{
3342	/*
3343	* Return failure or success depending on whether the values match or not.
3344	*/
3345	if ( pDevExt->pvVMMR0 != pvVMMR0
3346	\|\| (void *)pDevExt->pfnVMMR0EntryInt != pvVMMR0EntryInt
3347	\|\| (void *)pDevExt->pfnVMMR0EntryFast != pvVMMR0EntryFast
3348	\|\| (void *)pDevExt->pfnVMMR0EntryEx != pvVMMR0EntryEx)
3349	{
3350	AssertMsgFailed(("SUP_IOCTL_LDR_SETR0EP: Already set pointing to a different module!\n"));
3351	rc = VERR_INVALID_PARAMETER;
3352	}
3353	}
3354	return rc;
3355	}
3356
3357
3358	/**
3359	* Unsets the R0 entry point installed by supdrvLdrSetR0EP.
3360	*
3361	* @param pDevExt Device globals.
3362	*/
3363	static void supdrvLdrUnsetR0EP(PSUPDRVDEVEXT pDevExt)
3364	{
3365	#ifdef VBOX_WITH_IDT_PATCHING
3366	PSUPDRVPATCH pPatch;
3367	#endif
3368
3369	pDevExt->pvVMMR0 = NULL;
3370	pDevExt->pfnVMMR0EntryInt = NULL;
3371	pDevExt->pfnVMMR0EntryFast = NULL;
3372	pDevExt->pfnVMMR0EntryEx = NULL;
3373
3374	#ifdef VBOX_WITH_IDT_PATCHING
3375	for (pPatch = pDevExt->pIdtPatches; pPatch; pPatch = pPatch->pNext)
3376	{
3377	# ifdef RT_ARCH_AMD64
3378	ASMAtomicXchgU64((volatile uint64_t *)&pPatch->auCode[pPatch->offVMMR0EntryFixup],
3379	(uint64_t)&pPatch->auCode[pPatch->offStub]);
3380	# else /* RT_ARCH_X86 */
3381	ASMAtomicXchgU32((volatile uint32_t *)&pPatch->auCode[pPatch->offVMMR0EntryFixup],
3382	(uint32_t)&pPatch->auCode[pPatch->offStub] - (uint32_t)&pPatch->auCode[pPatch->offVMMR0EntryFixup + 4]);
3383	# endif
3384	}
3385	#endif /* VBOX_WITH_IDT_PATCHING */
3386	}
3387
3388
3389	/**
3390	* Adds a usage reference in the specified session of an image.
3391	*
3392	* @param pSession Session in question.
3393	* @param pImage Image which the session is using.
3394	*/
3395	static void supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage)
3396	{
3397	PSUPDRVLDRUSAGE pUsage;
3398	dprintf(("supdrvLdrAddUsage: pImage=%p\n", pImage));
3399
3400	/*
3401	* Referenced it already?
3402	*/
3403	pUsage = pSession->pLdrUsage;
3404	while (pUsage)
3405	{
3406	if (pUsage->pImage == pImage)
3407	{
3408	pUsage->cUsage++;
3409	return;
3410	}
3411	pUsage = pUsage->pNext;
3412	}
3413
3414	/*
3415	* Allocate new usage record.
3416	*/
3417	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
3418	Assert(pUsage);
3419	if (pUsage)
3420	{
3421	pUsage->cUsage = 1;
3422	pUsage->pImage = pImage;
3423	pUsage->pNext = pSession->pLdrUsage;
3424	pSession->pLdrUsage = pUsage;
3425	}
3426	/* ignore errors... */
3427	}
3428
3429
3430	/**
3431	* Frees a load image.
3432	*
3433	* @param pDevExt Pointer to device extension.
3434	* @param pImage Pointer to the image we're gonna free.
3435	* This image must exit!
3436	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
3437	*/
3438	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
3439	{
3440	PSUPDRVLDRIMAGE pImagePrev;
3441	dprintf(("supdrvLdrFree: pImage=%p\n", pImage));
3442
3443	/* find it - arg. should've used doubly linked list. */
3444	Assert(pDevExt->pLdrImages);
3445	pImagePrev = NULL;
3446	if (pDevExt->pLdrImages != pImage)
3447	{
3448	pImagePrev = pDevExt->pLdrImages;
3449	while (pImagePrev->pNext != pImage)
3450	pImagePrev = pImagePrev->pNext;
3451	Assert(pImagePrev->pNext == pImage);
3452	}
3453
3454	/* unlink */
3455	if (pImagePrev)
3456	pImagePrev->pNext = pImage->pNext;
3457	else
3458	pDevExt->pLdrImages = pImage->pNext;
3459
3460	/* check if this is VMMR0.r0 and fix the Idt patches if it is. */
3461	if (pDevExt->pvVMMR0 == pImage->pvImage)
3462	supdrvLdrUnsetR0EP(pDevExt);
3463
3464	/* check for objects with destructors in this image. (Shouldn't happen.) */
3465	if (pDevExt->pObjs)
3466	{
3467	unsigned cObjs = 0;
3468	PSUPDRVOBJ pObj;
3469	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3470	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
3471	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
3472	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImage))
3473	{
3474	pObj->pfnDestructor = NULL;
3475	cObjs++;
3476	}
3477	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
3478	if (cObjs)
3479	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
3480	}
3481
3482	/* call termination function if fully loaded. */
3483	if ( pImage->pfnModuleTerm
3484	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
3485	{
3486	dprintf(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
3487	pImage->pfnModuleTerm();
3488	}
3489
3490	/* free the image */
3491	pImage->cUsage = 0;
3492	pImage->pNext = 0;
3493	pImage->uState = SUP_IOCTL_LDR_FREE;
3494	RTMemExecFree(pImage);
3495	}
3496
3497
3498	/**
3499	* Gets the current paging mode of the CPU and stores in in pOut.
3500	*/
3501	static SUPPAGINGMODE supdrvIOCtl_GetPagingMode(void)
3502	{
3503	SUPPAGINGMODE enmMode;
3504
3505	RTUINTREG cr0 = ASMGetCR0();
3506	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
3507	enmMode = SUPPAGINGMODE_INVALID;
3508	else
3509	{
3510	RTUINTREG cr4 = ASMGetCR4();
3511	uint32_t fNXEPlusLMA = 0;
3512	if (cr4 & X86_CR4_PAE)
3513	{
3514	uint32_t fAmdFeatures = ASMCpuId_EDX(0x80000001);
3515	if (fAmdFeatures & (X86_CPUID_AMD_FEATURE_EDX_NX \| X86_CPUID_AMD_FEATURE_EDX_LONG_MODE))
3516	{
3517	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
3518	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
3519	fNXEPlusLMA \|= BIT(0);
3520	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
3521	fNXEPlusLMA \|= BIT(1);
3522	}
3523	}
3524
3525	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
3526	{
3527	case 0:
3528	enmMode = SUPPAGINGMODE_32_BIT;
3529	break;
3530
3531	case X86_CR4_PGE:
3532	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
3533	break;
3534
3535	case X86_CR4_PAE:
3536	enmMode = SUPPAGINGMODE_PAE;
3537	break;
3538
3539	case X86_CR4_PAE \| BIT(0):
3540	enmMode = SUPPAGINGMODE_PAE_NX;
3541	break;
3542
3543	case X86_CR4_PAE \| X86_CR4_PGE:
3544	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3545	break;
3546
3547	case X86_CR4_PAE \| X86_CR4_PGE \| BIT(0):
3548	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3549	break;
3550
3551	case BIT(1) \| X86_CR4_PAE:
3552	enmMode = SUPPAGINGMODE_AMD64;
3553	break;
3554
3555	case BIT(1) \| X86_CR4_PAE \| BIT(0):
3556	enmMode = SUPPAGINGMODE_AMD64_NX;
3557	break;
3558
3559	case BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
3560	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
3561	break;
3562
3563	case BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| BIT(0):
3564	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
3565	break;
3566
3567	default:
3568	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
3569	enmMode = SUPPAGINGMODE_INVALID;
3570	break;
3571	}
3572	}
3573	return enmMode;
3574	}
3575
3576
3577	#ifdef USE_NEW_OS_INTERFACE_FOR_GIP
3578	/**
3579	* Creates the GIP.
3580	*
3581	* @returns negative errno.
3582	* @param pDevExt Instance data. GIP stuff may be updated.
3583	*/
3584	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
3585	{
3586	PSUPGLOBALINFOPAGE pGip;
3587	RTHCPHYS HCPhysGip;
3588	uint32_t u32SystemResolution;
3589	uint32_t u32Interval;
3590	int rc;
3591
3592	dprintf(("supdrvGipCreate:\n"));
3593
3594	/* assert order */
3595	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
3596	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
3597	Assert(!pDevExt->pGipTimer);
3598
3599	/*
3600	* Allocate a suitable page with a default kernel mapping.
3601	*/
3602	rc = RTR0MemObjAllocLow(&pDevExt->GipMemObj, PAGE_SIZE, false);
3603	if (RT_FAILURE(rc))
3604	{
3605	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
3606	return rc;
3607	}
3608	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
3609	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
3610
3611	/*
3612	* Try bump up the system timer resolution.
3613	* The more interrupts the better...
3614	*/
3615	if ( RT_SUCCESS(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
3616	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
3617	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 3906250 /* 256 HZ */, &u32SystemResolution))
3618	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 4000000 /* 250 HZ */, &u32SystemResolution))
3619	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 7812500 /* 128 HZ */, &u32SystemResolution))
3620	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity(10000000 /* 100 HZ */, &u32SystemResolution))
3621	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity(15625000 /* 64 HZ */, &u32SystemResolution))
3622	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity(31250000 /* 32 HZ */, &u32SystemResolution))
3623	)
3624	{
3625	Assert(RTTimerGetSystemGranularity() <= u32SystemResolution);
3626	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
3627	}
3628
3629	/*
3630	* Find a reasonable update interval, something close to 10ms would be nice,
3631	* and create a recurring timer.
3632	*/
3633	u32Interval = u32SystemResolution = RTTimerGetSystemGranularity();
3634	while (u32Interval < 10000000 /* 10 ms */)
3635	u32Interval += u32SystemResolution;
3636
3637	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0, supdrvGipTimer, pDevExt);
3638	if (RT_FAILURE(rc))
3639	{
3640	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %RU32 ns interval. rc=%d\n", u32Interval, rc));
3641	Assert(!pDevExt->pGipTimer);
3642	supdrvGipDestroy(pDevExt);
3643	return rc;
3644	}
3645
3646	/*
3647	* We're good.
3648	*/
3649	supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), 1000000000 / u32Interval /=Hz/);
3650	return VINF_SUCCESS;
3651	}
3652
3653
3654	/**
3655	* Terminates the GIP.
3656	*
3657	* @param pDevExt Instance data. GIP stuff may be updated.
3658	*/
3659	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
3660	{
3661	int rc;
3662	#ifdef DEBUG_DARWIN_GIP
3663	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
3664	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
3665	pDevExt->pGipTimer, pDevExt->GipMemObj));
3666	#endif
3667
3668	/*
3669	* Invalid the GIP data.
3670	*/
3671	if (pDevExt->pGip)
3672	{
3673	supdrvGipTerm(pDevExt->pGip);
3674	pDevExt->pGip = NULL;
3675	}
3676
3677	/*
3678	* Destroy the timer and free the GIP memory object.
3679	*/
3680	if (pDevExt->pGipTimer)
3681	{
3682	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
3683	pDevExt->pGipTimer = NULL;
3684	}
3685
3686	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
3687	{
3688	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
3689	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
3690	}
3691
3692	/*
3693	* Finally, release the system timer resolution request if one succeeded.
3694	*/
3695	if (pDevExt->u32SystemTimerGranularityGrant)
3696	{
3697	rc = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
3698	pDevExt->u32SystemTimerGranularityGrant = 0;
3699	}
3700	}
3701
3702
3703	/**
3704	* Timer callback function.
3705	* @param pTimer The timer.
3706	* @param pvUser The device extension.
3707	*/
3708	static DECLCALLBACK(void) supdrvGipTimer(PRTTIMER pTimer, void *pvUser)
3709	{
3710	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
3711	supdrvGipUpdate(pDevExt->pGip, RTTimeSystemNanoTS());
3712	}
3713	#endif /* USE_NEW_OS_INTERFACE_FOR_GIP */
3714
3715
3716	/**
3717	* Initializes the GIP data.
3718	*
3719	* @returns IPRT status code.
3720	* @param pDevExt Pointer to the device instance data.
3721	* @param pGip Pointer to the read-write kernel mapping of the GIP.
3722	* @param HCPhys The physical address of the GIP.
3723	* @param u64NanoTS The current nanosecond timestamp.
3724	* @param uUpdateHz The update freqence.
3725	*/
3726	int VBOXCALL supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys, uint64_t u64NanoTS, unsigned uUpdateHz)
3727	{
3728	unsigned i;
3729	#ifdef DEBUG_DARWIN_GIP
3730	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz));
3731	#else
3732	dprintf(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz));
3733	#endif
3734
3735	/*
3736	* Initialize the structure.
3737	*/
3738	memset(pGip, 0, PAGE_SIZE);
3739	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
3740	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
3741	pGip->u32Mode = supdrvGipDeterminTscMode();
3742	pGip->u32UpdateHz = uUpdateHz;
3743	pGip->u32UpdateIntervalNS = 1000000000 / uUpdateHz;
3744	pGip->u64NanoTSLastUpdateHz = u64NanoTS;
3745
3746	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
3747	{
3748	pGip->aCPUs[i].u32TransactionId = 2;
3749	pGip->aCPUs[i].u64NanoTS = u64NanoTS;
3750	pGip->aCPUs[i].u64TSC = ASMReadTSC();
3751
3752	/*
3753	* We don't know the following values until we've executed updates.
3754	* So, we'll just insert very high values.
3755	*/
3756	pGip->aCPUs[i].u64CpuHz = _4G + 1;
3757	pGip->aCPUs[i].u32UpdateIntervalTSC = _2G / 4;
3758	pGip->aCPUs[i].au32TSCHistory[0] = _2G / 4;
3759	pGip->aCPUs[i].au32TSCHistory[1] = _2G / 4;
3760	pGip->aCPUs[i].au32TSCHistory[2] = _2G / 4;
3761	pGip->aCPUs[i].au32TSCHistory[3] = _2G / 4;
3762	pGip->aCPUs[i].au32TSCHistory[4] = _2G / 4;
3763	pGip->aCPUs[i].au32TSCHistory[5] = _2G / 4;
3764	pGip->aCPUs[i].au32TSCHistory[6] = _2G / 4;
3765	pGip->aCPUs[i].au32TSCHistory[7] = _2G / 4;
3766	}
3767
3768	/*
3769	* Link it to the device extension.
3770	*/
3771	pDevExt->pGip = pGip;
3772	pDevExt->HCPhysGip = HCPhys;
3773	pDevExt->cGipUsers = 0;
3774
3775	return VINF_SUCCESS;
3776	}
3777
3778
3779	/**
3780	* Determin the GIP TSC mode.
3781	*
3782	* @returns The most suitable TSC mode.
3783	*/
3784	static SUPGIPMODE supdrvGipDeterminTscMode(void)
3785	{
3786	#ifndef USE_NEW_OS_INTERFACE_FOR_GIP
3787	/*
3788	* The problem here is that AMD processors with power management features
3789	* may easily end up with different TSCs because the CPUs or even cores
3790	* on the same physical chip run at different frequencies to save power.
3791	*
3792	* It is rumoured that this will be corrected with Barcelona and it's
3793	* expected that this will be indicated by the TscInvariant bit in
3794	* cpuid(0x80000007). So, the "difficult" bit here is to correctly
3795	* identify the older CPUs which don't do different frequency and
3796	* can be relied upon to have somewhat uniform TSC between the cpus.
3797	*/
3798	if (supdrvOSGetCPUCount() > 1)
3799	{
3800	uint32_t uEAX, uEBX, uECX, uEDX;
3801
3802	/* Permit user users override. */
3803	if (supdrvOSGetForcedAsyncTscMode())
3804	return SUPGIPMODE_ASYNC_TSC;
3805
3806	/* Check for "AuthenticAMD" */
3807	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
3808	if (uEAX >= 1 && uEBX == 0x68747541 && uECX == 0x444d4163 && uEDX == 0x69746e65)
3809	{
3810	/* Check for APM support and that TscInvariant is cleared. */
3811	ASMCpuId(0x80000000, &uEAX, &uEBX, &uECX, &uEDX);
3812	if (uEAX >= 0x80000007)
3813	{
3814	ASMCpuId(0x80000007, &uEAX, &uEBX, &uECX, &uEDX);
3815	if ( !(uEDX & BIT(8))/* TscInvariant */
3816	&& (uEDX & 0x3e)) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
3817	return SUPGIPMODE_ASYNC_TSC;
3818	}
3819	}
3820	}
3821	#endif
3822	return SUPGIPMODE_SYNC_TSC;
3823	}
3824
3825
3826	/**
3827	* Invalidates the GIP data upon termination.
3828	*
3829	* @param pGip Pointer to the read-write kernel mapping of the GIP.
3830	*/
3831	void VBOXCALL supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
3832	{
3833	unsigned i;
3834	pGip->u32Magic = 0;
3835	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
3836	{
3837	pGip->aCPUs[i].u64NanoTS = 0;
3838	pGip->aCPUs[i].u64TSC = 0;
3839	pGip->aCPUs[i].iTSCHistoryHead = 0;
3840	}
3841	}
3842
3843
3844	/**
3845	* Worker routine for supdrvGipUpdate and supdrvGipUpdatePerCpu that
3846	* updates all the per cpu data except the transaction id.
3847	*
3848	* @param pGip The GIP.
3849	* @param pGipCpu Pointer to the per cpu data.
3850	* @param u64NanoTS The current time stamp.
3851	*/
3852	static void supdrvGipDoUpdateCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
3853	{
3854	uint64_t u64TSC;
3855	uint64_t u64TSCDelta;
3856	uint32_t u32UpdateIntervalTSC;
3857	uint32_t u32UpdateIntervalTSCSlack;
3858	unsigned iTSCHistoryHead;
3859	uint64_t u64CpuHz;
3860
3861	/*
3862	* Update the NanoTS.
3863	*/
3864	ASMAtomicXchgU64(&pGipCpu->u64NanoTS, u64NanoTS);
3865
3866	/*
3867	* Calc TSC delta.
3868	*/
3869	/** @todo validate the NanoTS delta, don't trust the OS to call us when it should... */
3870	u64TSC = ASMReadTSC();
3871	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
3872	ASMAtomicXchgU64(&pGipCpu->u64TSC, u64TSC);
3873
3874	if (u64TSCDelta >> 32)
3875	{
3876	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
3877	pGipCpu->cErrors++;
3878	}
3879
3880	/*
3881	* TSC History.
3882	*/
3883	Assert(ELEMENTS(pGipCpu->au32TSCHistory) == 8);
3884
3885	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
3886	ASMAtomicXchgU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
3887	ASMAtomicXchgU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
3888
3889	/*
3890	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
3891	*/
3892	if (pGip->u32UpdateHz >= 1000)
3893	{
3894	uint32_t u32;
3895	u32 = pGipCpu->au32TSCHistory[0];
3896	u32 += pGipCpu->au32TSCHistory[1];
3897	u32 += pGipCpu->au32TSCHistory[2];
3898	u32 += pGipCpu->au32TSCHistory[3];
3899	u32 >>= 2;
3900	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
3901	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
3902	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
3903	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
3904	u32UpdateIntervalTSC >>= 2;
3905	u32UpdateIntervalTSC += u32;
3906	u32UpdateIntervalTSC >>= 1;
3907
3908	/* Value choosen for a 2GHz Athlon64 running linux 2.6.10/11, . */
3909	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
3910	}
3911	else if (pGip->u32UpdateHz >= 90)
3912	{
3913	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
3914	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
3915	u32UpdateIntervalTSC >>= 1;
3916
3917	/* value choosen on a 2GHz thinkpad running windows */
3918	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
3919	}
3920	else
3921	{
3922	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
3923
3924	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
3925	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
3926	}
3927	ASMAtomicXchgU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
3928
3929	/*
3930	* CpuHz.
3931	*/
3932	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, pGip->u32UpdateHz);
3933	ASMAtomicXchgU64(&pGipCpu->u64CpuHz, u64CpuHz);
3934	}
3935
3936
3937	/**
3938	* Updates the GIP.
3939	*
3940	* @param pGip Pointer to the GIP.
3941	* @param u64NanoTS The current nanosecond timesamp.
3942	*/
3943	void VBOXCALL supdrvGipUpdate(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS)
3944	{
3945	/*
3946	* Determin the relevant CPU data.
3947	*/
3948	PSUPGIPCPU pGipCpu;
3949	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
3950	pGipCpu = &pGip->aCPUs[0];
3951	else
3952	{
3953	unsigned iCpu = ASMGetApicId();
3954	if (RT_LIKELY(iCpu >= RT_ELEMENTS(pGip->aCPUs)))
3955	return;
3956	pGipCpu = &pGip->aCPUs[iCpu];
3957	}
3958
3959	/*
3960	* Start update transaction.
3961	*/
3962	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
3963	{
3964	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
3965	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
3966	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
3967	pGipCpu->cErrors++;
3968	return;
3969	}
3970
3971	/*
3972	* Recalc the update frequency every 0x800th time.
3973	*/
3974	if (!(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
3975	{
3976	if (pGip->u64NanoTSLastUpdateHz)
3977	{
3978	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
3979	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
3980	uint32_t u32UpdateHz = (uint32_t)((UINT64_C(1000000000) * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
3981	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
3982	{
3983	ASMAtomicXchgU32(&pGip->u32UpdateHz, u32UpdateHz);
3984	ASMAtomicXchgU32(&pGip->u32UpdateIntervalNS, 1000000000 / u32UpdateHz);
3985	}
3986	#endif
3987	}
3988	ASMAtomicXchgU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS);
3989	}
3990
3991	/*
3992	* Update the data.
3993	*/
3994	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS);
3995
3996	/*
3997	* Complete transaction.
3998	*/
3999	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
4000	}
4001
4002
4003	/**
4004	* Updates the per cpu GIP data for the calling cpu.
4005	*
4006	* @param pGip Pointer to the GIP.
4007	* @param u64NanoTS The current nanosecond timesamp.
4008	* @param iCpu The CPU index.
4009	*/
4010	void VBOXCALL supdrvGipUpdatePerCpu(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, unsigned iCpu)
4011	{
4012	PSUPGIPCPU pGipCpu;
4013
4014	if (RT_LIKELY(iCpu <= RT_ELEMENTS(pGip->aCPUs)))
4015	{
4016	pGipCpu = &pGip->aCPUs[iCpu];
4017
4018	/*
4019	* Start update transaction.
4020	*/
4021	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
4022	{
4023	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
4024	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
4025	pGipCpu->cErrors++;
4026	return;
4027	}
4028
4029	/*
4030	* Update the data.
4031	*/
4032	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS);
4033
4034	/*
4035	* Complete transaction.
4036	*/
4037	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
4038	}
4039	}
4040
4041
4042	#ifndef DEBUG /** @todo change #ifndef DEBUG -> #ifdef LOG_ENABLED */
4043	/**
4044	* Stub function for non-debug builds.
4045	*/
4046	RTDECL(PRTLOGGER) RTLogDefaultInstance(void)
4047	{
4048	return NULL;
4049	}
4050
4051	RTDECL(PRTLOGGER) RTLogRelDefaultInstance(void)
4052	{
4053	return NULL;
4054	}
4055
4056	/**
4057	* Stub function for non-debug builds.
4058	*/
4059	RTDECL(int) RTLogSetDefaultInstanceThread(PRTLOGGER pLogger, uintptr_t uKey)
4060	{
4061	return 0;
4062	}
4063
4064	/**
4065	* Stub function for non-debug builds.
4066	*/
4067	RTDECL(void) RTLogLogger(PRTLOGGER pLogger, void pvCallerRet, const char pszFormat, ...)
4068	{
4069	}
4070
4071	/**
4072	* Stub function for non-debug builds.
4073	*/
4074	RTDECL(void) RTLogLoggerEx(PRTLOGGER pLogger, unsigned fFlags, unsigned iGroup, const char *pszFormat, ...)
4075	{
4076	}
4077
4078	/**
4079	* Stub function for non-debug builds.
4080	*/
4081	RTDECL(void) RTLogLoggerExV(PRTLOGGER pLogger, unsigned fFlags, unsigned iGroup, const char *pszFormat, va_list args)
4082	{
4083	}
4084
4085	/**
4086	* Stub function for non-debug builds.
4087	*/
4088	RTDECL(void) RTLogPrintf(const char *pszFormat, ...)
4089	{
4090	}
4091
4092	/**
4093	* Stub function for non-debug builds.
4094	*/
4095	RTDECL(void) RTLogPrintfV(const char *pszFormat, va_list args)
4096	{
4097	}
4098	#endif /* !DEBUG */
4099

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

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDRVShared.c@ 4985

Download in other formats: