SUPDRVShared.c@ 8760

Last change on this file since 8760 was 8155, checked in by vboxsync, 17 years ago
The Big Sun Rebranding Header Change
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 146.5 KB

Line
1	/* $Revision: 8155 $ */
2	/** @file
3	* VirtualBox Support Driver - Shared code.
4	*/
5
6	/*
7	* Copyright (C) 2006-2007 Sun Microsystems, Inc.
8	*
9	* This file is part of VirtualBox Open Source Edition (OSE), as
10	* available from http://www.virtualbox.org. This file is free software;
11	* you can redistribute it and/or modify it under the terms of the GNU
12	* General Public License (GPL) as published by the Free Software
13	* Foundation, in version 2 as it comes in the "COPYING" file of the
14	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16	*
17	* The contents of this file may alternatively be used under the terms
18	* of the Common Development and Distribution License Version 1.0
19	* (CDDL) only, as it comes in the "COPYING.CDDL" file of the
20	* VirtualBox OSE distribution, in which case the provisions of the
21	* CDDL are applicable instead of those of the GPL.
22	*
23	* You may elect to license modified versions of this file under the
24	* terms and conditions of either the GPL or the CDDL or both.
25	*
26	* Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa
27	* Clara, CA 95054 USA or visit http://www.sun.com if you need
28	* additional information or have any questions.
29	*/
30
31
32	/*******************************************************************************
33	* Header Files *
34	*******************************************************************************/
35	#include "SUPDRV.h"
36	#ifndef PAGE_SHIFT
37	# include <iprt/param.h>
38	#endif
39	#include <iprt/alloc.h>
40	#include <iprt/semaphore.h>
41	#include <iprt/spinlock.h>
42	#include <iprt/thread.h>
43	#include <iprt/process.h>
44	#include <iprt/mp.h>
45	#include <iprt/cpuset.h>
46	#include <iprt/log.h>
47
48	/*
49	* Logging assignments:
50	* Log - useful stuff, like failures.
51	* LogFlow - program flow, except the really noisy bits.
52	* Log2 - Cleanup and IDTE
53	* Log3 - Loader flow noise.
54	* Log4 - Call VMMR0 flow noise.
55	* Log5 - Native yet-to-be-defined noise.
56	* Log6 - Native ioctl flow noise.
57	*
58	* Logging requires BUILD_TYPE=debug and possibly changes to the logger
59	* instanciation in log-vbox.c(pp).
60	*/
61
62
63	/*******************************************************************************
64	* Defined Constants And Macros *
65	*******************************************************************************/
66	/* from x86.h - clashes with linux thus this duplication */
67	#undef X86_CR0_PG
68	#define X86_CR0_PG RT_BIT(31)
69	#undef X86_CR0_PE
70	#define X86_CR0_PE RT_BIT(0)
71	#undef X86_CPUID_AMD_FEATURE_EDX_NX
72	#define X86_CPUID_AMD_FEATURE_EDX_NX RT_BIT(20)
73	#undef MSR_K6_EFER
74	#define MSR_K6_EFER 0xc0000080
75	#undef MSR_K6_EFER_NXE
76	#define MSR_K6_EFER_NXE RT_BIT(11)
77	#undef MSR_K6_EFER_LMA
78	#define MSR_K6_EFER_LMA RT_BIT(10)
79	#undef X86_CR4_PGE
80	#define X86_CR4_PGE RT_BIT(7)
81	#undef X86_CR4_PAE
82	#define X86_CR4_PAE RT_BIT(5)
83	#undef X86_CPUID_AMD_FEATURE_EDX_LONG_MODE
84	#define X86_CPUID_AMD_FEATURE_EDX_LONG_MODE RT_BIT(29)
85
86
87	/** The frequency by which we recalculate the u32UpdateHz and
88	* u32UpdateIntervalNS GIP members. The value must be a power of 2. */
89	#define GIP_UPDATEHZ_RECALC_FREQ 0x800
90
91	/**
92	* Validates a session pointer.
93	*
94	* @returns true/false accordingly.
95	* @param pSession The session.
96	*/
97	#define SUP_IS_SESSION_VALID(pSession) \
98	( VALID_PTR(pSession) \
99	&& pSession->u32Cookie == BIRD_INV)
100
101
102	/*******************************************************************************
103	* Global Variables *
104	*******************************************************************************/
105	/**
106	* Array of the R0 SUP API.
107	*/
108	static SUPFUNC g_aFunctions[] =
109	{
110	/* name function */
111	{ "SUPR0ObjRegister", (void *)SUPR0ObjRegister },
112	{ "SUPR0ObjAddRef", (void *)SUPR0ObjAddRef },
113	{ "SUPR0ObjRelease", (void *)SUPR0ObjRelease },
114	{ "SUPR0ObjVerifyAccess", (void *)SUPR0ObjVerifyAccess },
115	{ "SUPR0LockMem", (void *)SUPR0LockMem },
116	{ "SUPR0UnlockMem", (void *)SUPR0UnlockMem },
117	{ "SUPR0ContAlloc", (void *)SUPR0ContAlloc },
118	{ "SUPR0ContFree", (void *)SUPR0ContFree },
119	{ "SUPR0LowAlloc", (void *)SUPR0LowAlloc },
120	{ "SUPR0LowFree", (void *)SUPR0LowFree },
121	{ "SUPR0MemAlloc", (void *)SUPR0MemAlloc },
122	{ "SUPR0MemGetPhys", (void *)SUPR0MemGetPhys },
123	{ "SUPR0MemFree", (void *)SUPR0MemFree },
124	{ "SUPR0PageAlloc", (void *)SUPR0PageAlloc },
125	{ "SUPR0PageFree", (void *)SUPR0PageFree },
126	{ "SUPR0Printf", (void *)SUPR0Printf },
127	{ "RTMemAlloc", (void *)RTMemAlloc },
128	{ "RTMemAllocZ", (void *)RTMemAllocZ },
129	{ "RTMemFree", (void *)RTMemFree },
130	/{ "RTMemDup", (void )RTMemDup },*/
131	{ "RTMemRealloc", (void *)RTMemRealloc },
132	{ "RTR0MemObjAllocLow", (void *)RTR0MemObjAllocLow },
133	{ "RTR0MemObjAllocPage", (void *)RTR0MemObjAllocPage },
134	{ "RTR0MemObjAllocPhys", (void *)RTR0MemObjAllocPhys },
135	{ "RTR0MemObjAllocPhysNC", (void *)RTR0MemObjAllocPhysNC },
136	{ "RTR0MemObjAllocCont", (void *)RTR0MemObjAllocCont },
137	{ "RTR0MemObjLockUser", (void *)RTR0MemObjLockUser },
138	{ "RTR0MemObjMapKernel", (void *)RTR0MemObjMapKernel },
139	{ "RTR0MemObjMapUser", (void *)RTR0MemObjMapUser },
140	{ "RTR0MemObjAddress", (void *)RTR0MemObjAddress },
141	{ "RTR0MemObjAddressR3", (void *)RTR0MemObjAddressR3 },
142	{ "RTR0MemObjSize", (void *)RTR0MemObjSize },
143	{ "RTR0MemObjIsMapping", (void *)RTR0MemObjIsMapping },
144	{ "RTR0MemObjGetPagePhysAddr", (void *)RTR0MemObjGetPagePhysAddr },
145	{ "RTR0MemObjFree", (void *)RTR0MemObjFree },
146	/* These don't work yet on linux - use fast mutexes!
147	{ "RTSemMutexCreate", (void *)RTSemMutexCreate },
148	{ "RTSemMutexRequest", (void *)RTSemMutexRequest },
149	{ "RTSemMutexRelease", (void *)RTSemMutexRelease },
150	{ "RTSemMutexDestroy", (void *)RTSemMutexDestroy },
151	*/
152	{ "RTProcSelf", (void *)RTProcSelf },
153	{ "RTR0ProcHandleSelf", (void *)RTR0ProcHandleSelf },
154	{ "RTSemFastMutexCreate", (void *)RTSemFastMutexCreate },
155	{ "RTSemFastMutexDestroy", (void *)RTSemFastMutexDestroy },
156	{ "RTSemFastMutexRequest", (void *)RTSemFastMutexRequest },
157	{ "RTSemFastMutexRelease", (void *)RTSemFastMutexRelease },
158	{ "RTSemEventCreate", (void *)RTSemEventCreate },
159	{ "RTSemEventSignal", (void *)RTSemEventSignal },
160	{ "RTSemEventWait", (void *)RTSemEventWait },
161	{ "RTSemEventWaitNoResume", (void *)RTSemEventWaitNoResume },
162	{ "RTSemEventDestroy", (void *)RTSemEventDestroy },
163	{ "RTSemEventMultiCreate", (void *)RTSemEventMultiCreate },
164	{ "RTSemEventMultiSignal", (void *)RTSemEventMultiSignal },
165	{ "RTSemEventMultiReset", (void *)RTSemEventMultiReset },
166	{ "RTSemEventMultiWait", (void *)RTSemEventMultiWait },
167	{ "RTSemEventMultiWaitNoResume", (void *)RTSemEventMultiWaitNoResume },
168	{ "RTSemEventMultiDestroy", (void *)RTSemEventMultiDestroy },
169	{ "RTSpinlockCreate", (void *)RTSpinlockCreate },
170	{ "RTSpinlockDestroy", (void *)RTSpinlockDestroy },
171	{ "RTSpinlockAcquire", (void *)RTSpinlockAcquire },
172	{ "RTSpinlockRelease", (void *)RTSpinlockRelease },
173	{ "RTSpinlockAcquireNoInts", (void *)RTSpinlockAcquireNoInts },
174	{ "RTSpinlockReleaseNoInts", (void *)RTSpinlockReleaseNoInts },
175	{ "RTThreadNativeSelf", (void *)RTThreadNativeSelf },
176	{ "RTThreadSleep", (void *)RTThreadSleep },
177	{ "RTThreadYield", (void *)RTThreadYield },
178	#if 0 /* Thread APIs, Part 2. */
179	{ "RTThreadSelf", (void *)RTThreadSelf },
180	{ "RTThreadCreate", (void *)RTThreadCreate },
181	{ "RTThreadGetNative", (void *)RTThreadGetNative },
182	{ "RTThreadWait", (void *)RTThreadWait },
183	{ "RTThreadWaitNoResume", (void *)RTThreadWaitNoResume },
184	{ "RTThreadGetName", (void *)RTThreadGetName },
185	{ "RTThreadSelfName", (void *)RTThreadSelfName },
186	{ "RTThreadGetType", (void *)RTThreadGetType },
187	{ "RTThreadUserSignal", (void *)RTThreadUserSignal },
188	{ "RTThreadUserReset", (void *)RTThreadUserReset },
189	{ "RTThreadUserWait", (void *)RTThreadUserWait },
190	{ "RTThreadUserWaitNoResume", (void *)RTThreadUserWaitNoResume },
191	#endif
192	{ "RTLogDefaultInstance", (void *)RTLogDefaultInstance },
193	{ "RTMpCpuIdFromSetIndex", (void *)RTMpCpuIdFromSetIndex },
194	{ "RTMpCpuIdToSetIndex", (void *)RTMpCpuIdToSetIndex },
195	{ "RTMpDoesCpuExist", (void *)RTMpDoesCpuExist },
196	{ "RTMpGetCount", (void *)RTMpGetCount },
197	{ "RTMpGetMaxCpuId", (void *)RTMpGetMaxCpuId },
198	{ "RTMpGetOnlineCount", (void *)RTMpGetOnlineCount },
199	{ "RTMpGetOnlineSet", (void *)RTMpGetOnlineSet },
200	{ "RTMpGetSet", (void *)RTMpGetSet },
201	{ "RTMpIsCpuOnline", (void *)RTMpIsCpuOnline },
202	{ "RTMpOnAll", (void *)RTMpOnAll },
203	{ "RTMpOnOthers", (void *)RTMpOnOthers },
204	{ "RTMpOnSpecific", (void *)RTMpOnSpecific },
205	{ "RTLogRelDefaultInstance", (void *)RTLogRelDefaultInstance },
206	{ "RTLogSetDefaultInstanceThread", (void *)RTLogSetDefaultInstanceThread },
207	{ "RTLogLogger", (void *)RTLogLogger },
208	{ "RTLogLoggerEx", (void *)RTLogLoggerEx },
209	{ "RTLogLoggerExV", (void *)RTLogLoggerExV },
210	{ "RTLogPrintf", (void *)RTLogPrintf },
211	{ "RTLogPrintfV", (void *)RTLogPrintfV },
212	{ "AssertMsg1", (void *)AssertMsg1 },
213	{ "AssertMsg2", (void *)AssertMsg2 },
214	};
215
216
217	/*******************************************************************************
218	* Internal Functions *
219	*******************************************************************************/
220	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession);
221	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType);
222	#ifdef VBOX_WITH_IDT_PATCHING
223	static int supdrvIOCtl_IdtInstall(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPIDTINSTALL pReq);
224	static PSUPDRVPATCH supdrvIdtPatchOne(PSUPDRVDEVEXT pDevExt, PSUPDRVPATCH pPatch);
225	static int supdrvIOCtl_IdtRemoveAll(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession);
226	static void supdrvIdtRemoveOne(PSUPDRVDEVEXT pDevExt, PSUPDRVPATCH pPatch);
227	static void supdrvIdtWrite(volatile void pvIdtEntry, const SUPDRVIDTE pNewIDTEntry);
228	#endif /* VBOX_WITH_IDT_PATCHING */
229	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq);
230	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq);
231	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq);
232	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq);
233	static int supdrvLdrSetR0EP(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx);
234	static void supdrvLdrUnsetR0EP(PSUPDRVDEVEXT pDevExt);
235	static void supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage);
236	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage);
237	static SUPPAGINGMODE supdrvIOCtl_GetPagingMode(void);
238	static SUPGIPMODE supdrvGipDeterminTscMode(void);
239	#ifdef RT_OS_WINDOWS
240	static int supdrvPageGetPhys(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages);
241	static bool supdrvPageWasLockedByPageAlloc(PSUPDRVSESSION pSession, RTR3PTR pvR3);
242	#endif
243	#ifdef USE_NEW_OS_INTERFACE_FOR_GIP
244	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt);
245	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt);
246	static DECLCALLBACK(void) supdrvGipTimer(PRTTIMER pTimer, void *pvUser);
247	#endif
248
249
250	/**
251	* Initializes the device extentsion structure.
252	*
253	* @returns IPRT status code.
254	* @param pDevExt The device extension to initialize.
255	*/
256	int VBOXCALL supdrvInitDevExt(PSUPDRVDEVEXT pDevExt)
257	{
258	/*
259	* Initialize it.
260	*/
261	int rc;
262	memset(pDevExt, 0, sizeof(*pDevExt));
263	rc = RTSpinlockCreate(&pDevExt->Spinlock);
264	if (!rc)
265	{
266	rc = RTSemFastMutexCreate(&pDevExt->mtxLdr);
267	if (!rc)
268	{
269	rc = RTSemFastMutexCreate(&pDevExt->mtxGip);
270	if (!rc)
271	{
272	#ifdef USE_NEW_OS_INTERFACE_FOR_GIP
273	rc = supdrvGipCreate(pDevExt);
274	if (RT_SUCCESS(rc))
275	{
276	pDevExt->u32Cookie = BIRD; /** @todo make this random? */
277	return VINF_SUCCESS;
278	}
279	#else
280	pDevExt->u32Cookie = BIRD;
281	return VINF_SUCCESS;
282	#endif
283	}
284	RTSemFastMutexDestroy(pDevExt->mtxLdr);
285	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
286	}
287	RTSpinlockDestroy(pDevExt->Spinlock);
288	pDevExt->Spinlock = NIL_RTSPINLOCK;
289	}
290	return rc;
291	}
292
293
294	/**
295	* Delete the device extension (e.g. cleanup members).
296	*
297	* @param pDevExt The device extension to delete.
298	*/
299	void VBOXCALL supdrvDeleteDevExt(PSUPDRVDEVEXT pDevExt)
300	{
301	#ifdef VBOX_WITH_IDT_PATCHING
302	PSUPDRVPATCH pPatch;
303	#endif
304	PSUPDRVOBJ pObj;
305	PSUPDRVUSAGE pUsage;
306
307	/*
308	* Kill mutexes and spinlocks.
309	*/
310	RTSemFastMutexDestroy(pDevExt->mtxGip);
311	pDevExt->mtxGip = NIL_RTSEMFASTMUTEX;
312	RTSemFastMutexDestroy(pDevExt->mtxLdr);
313	pDevExt->mtxLdr = NIL_RTSEMFASTMUTEX;
314	RTSpinlockDestroy(pDevExt->Spinlock);
315	pDevExt->Spinlock = NIL_RTSPINLOCK;
316
317	/*
318	* Free lists.
319	*/
320	#ifdef VBOX_WITH_IDT_PATCHING
321	/* patches */
322	/** @todo make sure we don't uninstall patches which has been patched by someone else. */
323	pPatch = pDevExt->pIdtPatchesFree;
324	pDevExt->pIdtPatchesFree = NULL;
325	while (pPatch)
326	{
327	void *pvFree = pPatch;
328	pPatch = pPatch->pNext;
329	RTMemExecFree(pvFree);
330	}
331	#endif /* VBOX_WITH_IDT_PATCHING */
332
333	/* objects. */
334	pObj = pDevExt->pObjs;
335	#if !defined(DEBUG_bird) \|\| !defined(RT_OS_LINUX) /* breaks unloading, temporary, remove me! */
336	Assert(!pObj); /* (can trigger on forced unloads) */
337	#endif
338	pDevExt->pObjs = NULL;
339	while (pObj)
340	{
341	void *pvFree = pObj;
342	pObj = pObj->pNext;
343	RTMemFree(pvFree);
344	}
345
346	/* usage records. */
347	pUsage = pDevExt->pUsageFree;
348	pDevExt->pUsageFree = NULL;
349	while (pUsage)
350	{
351	void *pvFree = pUsage;
352	pUsage = pUsage->pNext;
353	RTMemFree(pvFree);
354	}
355
356	#ifdef USE_NEW_OS_INTERFACE_FOR_GIP
357	/* kill the GIP */
358	supdrvGipDestroy(pDevExt);
359	#endif
360	}
361
362
363	/**
364	* Create session.
365	*
366	* @returns IPRT status code.
367	* @param pDevExt Device extension.
368	* @param ppSession Where to store the pointer to the session data.
369	*/
370	int VBOXCALL supdrvCreateSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION *ppSession)
371	{
372	/*
373	* Allocate memory for the session data.
374	*/
375	int rc = VERR_NO_MEMORY;
376	PSUPDRVSESSION pSession = ppSession = (PSUPDRVSESSION)RTMemAllocZ(sizeof(pSession));
377	if (pSession)
378	{
379	/* Initialize session data. */
380	rc = RTSpinlockCreate(&pSession->Spinlock);
381	if (!rc)
382	{
383	Assert(pSession->Spinlock != NIL_RTSPINLOCK);
384	pSession->pDevExt = pDevExt;
385	pSession->u32Cookie = BIRD_INV;
386	/*pSession->pLdrUsage = NULL;
387	pSession->pPatchUsage = NULL;
388	pSession->pUsage = NULL;
389	pSession->pGip = NULL;
390	pSession->fGipReferenced = false;
391	pSession->Bundle.cUsed = 0 */
392
393	LogFlow(("Created session %p initial cookie=%#x\n", pSession, pSession->u32Cookie));
394	return VINF_SUCCESS;
395	}
396
397	RTMemFree(pSession);
398	*ppSession = NULL;
399	Log(("Failed to create spinlock, rc=%d!\n", rc));
400	}
401
402	return rc;
403	}
404
405
406	/**
407	* Shared code for cleaning up a session.
408	*
409	* @param pDevExt Device extension.
410	* @param pSession Session data.
411	* This data will be freed by this routine.
412	*/
413	void VBOXCALL supdrvCloseSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
414	{
415	/*
416	* Cleanup the session first.
417	*/
418	supdrvCleanupSession(pDevExt, pSession);
419
420	/*
421	* Free the rest of the session stuff.
422	*/
423	RTSpinlockDestroy(pSession->Spinlock);
424	pSession->Spinlock = NIL_RTSPINLOCK;
425	pSession->pDevExt = NULL;
426	RTMemFree(pSession);
427	LogFlow(("supdrvCloseSession: returns\n"));
428	}
429
430
431	/**
432	* Shared code for cleaning up a session (but not quite freeing it).
433	*
434	* This is primarily intended for MAC OS X where we have to clean up the memory
435	* stuff before the file handle is closed.
436	*
437	* @param pDevExt Device extension.
438	* @param pSession Session data.
439	* This data will be freed by this routine.
440	*/
441	void VBOXCALL supdrvCleanupSession(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
442	{
443	PSUPDRVBUNDLE pBundle;
444	LogFlow(("supdrvCleanupSession: pSession=%p\n", pSession));
445
446	/*
447	* Remove logger instances related to this session.
448	*/
449	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pSession);
450
451	#ifdef VBOX_WITH_IDT_PATCHING
452	/*
453	* Uninstall any IDT patches installed for this session.
454	*/
455	supdrvIOCtl_IdtRemoveAll(pDevExt, pSession);
456	#endif
457
458	/*
459	* Release object references made in this session.
460	* In theory there should be noone racing us in this session.
461	*/
462	Log2(("release objects - start\n"));
463	if (pSession->pUsage)
464	{
465	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
466	PSUPDRVUSAGE pUsage;
467	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
468
469	while ((pUsage = pSession->pUsage) != NULL)
470	{
471	PSUPDRVOBJ pObj = pUsage->pObj;
472	pSession->pUsage = pUsage->pNext;
473
474	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
475	if (pUsage->cUsage < pObj->cUsage)
476	{
477	pObj->cUsage -= pUsage->cUsage;
478	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
479	}
480	else
481	{
482	/* Destroy the object and free the record. */
483	if (pDevExt->pObjs == pObj)
484	pDevExt->pObjs = pObj->pNext;
485	else
486	{
487	PSUPDRVOBJ pObjPrev;
488	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
489	if (pObjPrev->pNext == pObj)
490	{
491	pObjPrev->pNext = pObj->pNext;
492	break;
493	}
494	Assert(pObjPrev);
495	}
496	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
497
498	Log(("supdrvCleanupSession: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
499	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
500	if (pObj->pfnDestructor)
501	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
502	RTMemFree(pObj);
503	}
504
505	/* free it and continue. */
506	RTMemFree(pUsage);
507
508	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
509	}
510
511	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
512	AssertMsg(!pSession->pUsage, ("Some buster reregistered an object during desturction!\n"));
513	}
514	Log2(("release objects - done\n"));
515
516	/*
517	* Release memory allocated in the session.
518	*
519	* We do not serialize this as we assume that the application will
520	* not allocated memory while closing the file handle object.
521	*/
522	Log2(("freeing memory:\n"));
523	pBundle = &pSession->Bundle;
524	while (pBundle)
525	{
526	PSUPDRVBUNDLE pToFree;
527	unsigned i;
528
529	/*
530	* Check and unlock all entries in the bundle.
531	*/
532	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
533	{
534	if (pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ)
535	{
536	int rc;
537	Log2(("eType=%d pvR0=%p pvR3=%p cb=%ld\n", pBundle->aMem[i].eType, RTR0MemObjAddress(pBundle->aMem[i].MemObj),
538	(void *)RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3), (long)RTR0MemObjSize(pBundle->aMem[i].MemObj)));
539	if (pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ)
540	{
541	rc = RTR0MemObjFree(pBundle->aMem[i].MapObjR3, false);
542	AssertRC(rc); /** @todo figure out how to handle this. */
543	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
544	}
545	rc = RTR0MemObjFree(pBundle->aMem[i].MemObj, false);
546	AssertRC(rc); /** @todo figure out how to handle this. */
547	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
548	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
549	}
550	}
551
552	/*
553	* Advance and free previous bundle.
554	*/
555	pToFree = pBundle;
556	pBundle = pBundle->pNext;
557
558	pToFree->pNext = NULL;
559	pToFree->cUsed = 0;
560	if (pToFree != &pSession->Bundle)
561	RTMemFree(pToFree);
562	}
563	Log2(("freeing memory - done\n"));
564
565	/*
566	* Loaded images needs to be dereferenced and possibly freed up.
567	*/
568	RTSemFastMutexRequest(pDevExt->mtxLdr);
569	Log2(("freeing images:\n"));
570	if (pSession->pLdrUsage)
571	{
572	PSUPDRVLDRUSAGE pUsage = pSession->pLdrUsage;
573	pSession->pLdrUsage = NULL;
574	while (pUsage)
575	{
576	void *pvFree = pUsage;
577	PSUPDRVLDRIMAGE pImage = pUsage->pImage;
578	if (pImage->cUsage > pUsage->cUsage)
579	pImage->cUsage -= pUsage->cUsage;
580	else
581	supdrvLdrFree(pDevExt, pImage);
582	pUsage->pImage = NULL;
583	pUsage = pUsage->pNext;
584	RTMemFree(pvFree);
585	}
586	}
587	RTSemFastMutexRelease(pDevExt->mtxLdr);
588	Log2(("freeing images - done\n"));
589
590	/*
591	* Unmap the GIP.
592	*/
593	Log2(("umapping GIP:\n"));
594	#ifdef USE_NEW_OS_INTERFACE_FOR_GIP
595	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
596	#else
597	if (pSession->pGip)
598	#endif
599	{
600	SUPR0GipUnmap(pSession);
601	#ifndef USE_NEW_OS_INTERFACE_FOR_GIP
602	pSession->pGip = NULL;
603	#endif
604	pSession->fGipReferenced = 0;
605	}
606	Log2(("umapping GIP - done\n"));
607	}
608
609
610	/**
611	* Fast path I/O Control worker.
612	*
613	* @returns VBox status code that should be passed down to ring-3 unchanged.
614	* @param uIOCtl Function number.
615	* @param pDevExt Device extention.
616	* @param pSession Session data.
617	*/
618	int VBOXCALL supdrvIOCtlFast(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
619	{
620	int rc;
621
622	/*
623	* We check the two prereqs after doing this only to allow the compiler to optimize things better.
624	*/
625	if (RT_LIKELY(pSession->pVM && pDevExt->pfnVMMR0EntryFast))
626	{
627	switch (uIOCtl)
628	{
629	case SUP_IOCTL_FAST_DO_RAW_RUN:
630	rc = pDevExt->pfnVMMR0EntryFast(pSession->pVM, SUP_VMMR0_DO_RAW_RUN);
631	break;
632	case SUP_IOCTL_FAST_DO_HWACC_RUN:
633	rc = pDevExt->pfnVMMR0EntryFast(pSession->pVM, SUP_VMMR0_DO_HWACC_RUN);
634	break;
635	case SUP_IOCTL_FAST_DO_NOP:
636	rc = pDevExt->pfnVMMR0EntryFast(pSession->pVM, SUP_VMMR0_DO_NOP);
637	break;
638	default:
639	rc = VERR_INTERNAL_ERROR;
640	break;
641	}
642	}
643	else
644	rc = VERR_INTERNAL_ERROR;
645
646	return rc;
647	}
648
649
650	/**
651	* Helper for supdrvIOCtl. Check if pszStr contains any character of pszChars.
652	* We would use strpbrk here if this function would be contained in the RedHat kABI white
653	* list, see http://www.kerneldrivers.org/RHEL5.
654	*
655	* @return 1 if pszStr does contain any character of pszChars, 0 otherwise.
656	* @param pszStr String to check
657	* @param pszChars Character set
658	*/
659	static int supdrvCheckInvalidChar(const char pszStr, const char pszChars)
660	{
661	int chCur;
662	while ((chCur = *pszStr++) != '\0')
663	{
664	int ch;
665	const char *psz = pszChars;
666	while ((ch = *psz++) != '\0')
667	if (ch == chCur)
668	return 1;
669
670	}
671	return 0;
672	}
673
674
675	/**
676	* I/O Control worker.
677	*
678	* @returns 0 on success.
679	* @returns VERR_INVALID_PARAMETER if the request is invalid.
680	*
681	* @param uIOCtl Function number.
682	* @param pDevExt Device extention.
683	* @param pSession Session data.
684	* @param pReqHdr The request header.
685	*/
686	int VBOXCALL supdrvIOCtl(uintptr_t uIOCtl, PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPREQHDR pReqHdr)
687	{
688	/*
689	* Validate the request.
690	*/
691	/* this first check could probably be omitted as its also done by the OS specific code... */
692	if (RT_UNLIKELY( (pReqHdr->fFlags & SUPREQHDR_FLAGS_MAGIC_MASK) != SUPREQHDR_FLAGS_MAGIC
693	\|\| pReqHdr->cbIn < sizeof(*pReqHdr)
694	\|\| pReqHdr->cbOut < sizeof(*pReqHdr)))
695	{
696	OSDBGPRINT(("vboxdrv: Bad ioctl request header; cbIn=%#lx cbOut=%#lx fFlags=%#lx\n",
697	(long)pReqHdr->cbIn, (long)pReqHdr->cbOut, (long)pReqHdr->fFlags));
698	return VERR_INVALID_PARAMETER;
699	}
700	if (RT_UNLIKELY(uIOCtl == SUP_IOCTL_COOKIE))
701	{
702	if (pReqHdr->u32Cookie != SUPCOOKIE_INITIAL_COOKIE)
703	{
704	OSDBGPRINT(("SUP_IOCTL_COOKIE: bad cookie %#lx\n", (long)pReqHdr->u32Cookie));
705	return VERR_INVALID_PARAMETER;
706	}
707	}
708	else if (RT_UNLIKELY( pReqHdr->u32Cookie != pDevExt->u32Cookie
709	\|\| pReqHdr->u32SessionCookie != pSession->u32Cookie))
710	{
711	OSDBGPRINT(("vboxdrv: bad cookie %#lx / %#lx.\n", (long)pReqHdr->u32Cookie, (long)pReqHdr->u32SessionCookie));
712	return VERR_INVALID_PARAMETER;
713	}
714
715	/*
716	* Validation macros
717	*/
718	#define REQ_CHECK_SIZES_EX(Name, cbInExpect, cbOutExpect) \
719	do { \
720	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect) \|\| pReqHdr->cbOut != (cbOutExpect))) \
721	{ \
722	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld. cbOut=%ld expected %ld.\n", \
723	(long)pReq->Hdr.cbIn, (long)(cbInExpect), (long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
724	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
725	} \
726	} while (0)
727
728	#define REQ_CHECK_SIZES(Name) REQ_CHECK_SIZES_EX(Name, Name ## _SIZE_IN, Name ## _SIZE_OUT)
729
730	#define REQ_CHECK_SIZE_IN(Name, cbInExpect) \
731	do { \
732	if (RT_UNLIKELY(pReqHdr->cbIn != (cbInExpect))) \
733	{ \
734	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbIn=%ld expected %ld.\n", \
735	(long)pReq->Hdr.cbIn, (long)(cbInExpect))); \
736	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
737	} \
738	} while (0)
739
740	#define REQ_CHECK_SIZE_OUT(Name, cbOutExpect) \
741	do { \
742	if (RT_UNLIKELY(pReqHdr->cbOut != (cbOutExpect))) \
743	{ \
744	OSDBGPRINT(( #Name ": Invalid input/output sizes. cbOut=%ld expected %ld.\n", \
745	(long)pReq->Hdr.cbOut, (long)(cbOutExpect))); \
746	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
747	} \
748	} while (0)
749
750	#define REQ_CHECK_EXPR(Name, expr) \
751	do { \
752	if (RT_UNLIKELY(!(expr))) \
753	{ \
754	OSDBGPRINT(( #Name ": %s\n", #expr)); \
755	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
756	} \
757	} while (0)
758
759	#define REQ_CHECK_EXPR_FMT(expr, fmt) \
760	do { \
761	if (RT_UNLIKELY(!(expr))) \
762	{ \
763	OSDBGPRINT( fmt ); \
764	return pReq->Hdr.rc = VERR_INVALID_PARAMETER; \
765	} \
766	} while (0)
767
768
769	/*
770	* The switch.
771	*/
772	switch (SUP_CTL_CODE_NO_SIZE(uIOCtl))
773	{
774	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_COOKIE):
775	{
776	PSUPCOOKIE pReq = (PSUPCOOKIE)pReqHdr;
777	REQ_CHECK_SIZES(SUP_IOCTL_COOKIE);
778	if (strncmp(pReq->u.In.szMagic, SUPCOOKIE_MAGIC, sizeof(pReq->u.In.szMagic)))
779	{
780	OSDBGPRINT(("SUP_IOCTL_COOKIE: invalid magic %.16s\n", pReq->u.In.szMagic));
781	pReq->Hdr.rc = VERR_INVALID_MAGIC;
782	return 0;
783	}
784
785	#if 0
786	/*
787	* Call out to the OS specific code and let it do permission checks on the
788	* client process.
789	*/
790	if (!supdrvOSValidateClientProcess(pDevExt, pSession))
791	{
792	pReq->u.Out.u32Cookie = 0xffffffff;
793	pReq->u.Out.u32SessionCookie = 0xffffffff;
794	pReq->u.Out.u32SessionVersion = 0xffffffff;
795	pReq->u.Out.u32DriverVersion = SUPDRVIOC_VERSION;
796	pReq->u.Out.pSession = NULL;
797	pReq->u.Out.cFunctions = 0;
798	pReq->Hdr.rc = VERR_PERMISSION_DENIED;
799	return 0;
800	}
801	#endif
802
803	/*
804	* Match the version.
805	* The current logic is very simple, match the major interface version.
806	*/
807	if ( pReq->u.In.u32MinVersion > SUPDRVIOC_VERSION
808	\|\| (pReq->u.In.u32MinVersion & 0xffff0000) != (SUPDRVIOC_VERSION & 0xffff0000))
809	{
810	OSDBGPRINT(("SUP_IOCTL_COOKIE: Version mismatch. Requested: %#x Min: %#x Current: %#x\n",
811	pReq->u.In.u32ReqVersion, pReq->u.In.u32MinVersion, SUPDRVIOC_VERSION));
812	pReq->u.Out.u32Cookie = 0xffffffff;
813	pReq->u.Out.u32SessionCookie = 0xffffffff;
814	pReq->u.Out.u32SessionVersion = 0xffffffff;
815	pReq->u.Out.u32DriverVersion = SUPDRVIOC_VERSION;
816	pReq->u.Out.pSession = NULL;
817	pReq->u.Out.cFunctions = 0;
818	pReq->Hdr.rc = VERR_VERSION_MISMATCH;
819	return 0;
820	}
821
822	/*
823	* Fill in return data and be gone.
824	* N.B. The first one to change SUPDRVIOC_VERSION shall makes sure that
825	* u32SessionVersion <= u32ReqVersion!
826	*/
827	/** @todo Somehow validate the client and negotiate a secure cookie... */
828	pReq->u.Out.u32Cookie = pDevExt->u32Cookie;
829	pReq->u.Out.u32SessionCookie = pSession->u32Cookie;
830	pReq->u.Out.u32SessionVersion = SUPDRVIOC_VERSION;
831	pReq->u.Out.u32DriverVersion = SUPDRVIOC_VERSION;
832	pReq->u.Out.pSession = pSession;
833	pReq->u.Out.cFunctions = sizeof(g_aFunctions) / sizeof(g_aFunctions[0]);
834	pReq->Hdr.rc = VINF_SUCCESS;
835	return 0;
836	}
837
838	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_QUERY_FUNCS(0)):
839	{
840	/* validate */
841	PSUPQUERYFUNCS pReq = (PSUPQUERYFUNCS)pReqHdr;
842	REQ_CHECK_SIZES_EX(SUP_IOCTL_QUERY_FUNCS, SUP_IOCTL_QUERY_FUNCS_SIZE_IN, SUP_IOCTL_QUERY_FUNCS_SIZE_OUT(RT_ELEMENTS(g_aFunctions)));
843
844	/* execute */
845	pReq->u.Out.cFunctions = RT_ELEMENTS(g_aFunctions);
846	memcpy(&pReq->u.Out.aFunctions[0], g_aFunctions, sizeof(g_aFunctions));
847	pReq->Hdr.rc = VINF_SUCCESS;
848	return 0;
849	}
850
851	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_IDT_INSTALL):
852	{
853	/* validate */
854	PSUPIDTINSTALL pReq = (PSUPIDTINSTALL)pReqHdr;
855	REQ_CHECK_SIZES(SUP_IOCTL_IDT_INSTALL);
856
857	/* execute */
858	#ifdef VBOX_WITH_IDT_PATCHING
859	pReq->Hdr.rc = supdrvIOCtl_IdtInstall(pDevExt, pSession, pReq);
860	#else
861	pReq->u.Out.u8Idt = 3;
862	pReq->Hdr.rc = VERR_NOT_SUPPORTED;
863	#endif
864	return 0;
865	}
866
867	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_IDT_REMOVE):
868	{
869	/* validate */
870	PSUPIDTREMOVE pReq = (PSUPIDTREMOVE)pReqHdr;
871	REQ_CHECK_SIZES(SUP_IOCTL_IDT_REMOVE);
872
873	/* execute */
874	#ifdef VBOX_WITH_IDT_PATCHING
875	pReq->Hdr.rc = supdrvIOCtl_IdtRemoveAll(pDevExt, pSession);
876	#else
877	pReq->Hdr.rc = VERR_NOT_SUPPORTED;
878	#endif
879	return 0;
880	}
881
882	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_LOCK):
883	{
884	/* validate */
885	PSUPPAGELOCK pReq = (PSUPPAGELOCK)pReqHdr;
886	REQ_CHECK_SIZE_IN(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_IN);
887	REQ_CHECK_SIZE_OUT(SUP_IOCTL_PAGE_LOCK, SUP_IOCTL_PAGE_LOCK_SIZE_OUT(pReq->u.In.cPages));
888	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.cPages > 0);
889	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_LOCK, pReq->u.In.pvR3 >= PAGE_SIZE);
890
891	/* execute */
892	pReq->Hdr.rc = SUPR0LockMem(pSession, pReq->u.In.pvR3, pReq->u.In.cPages, &pReq->u.Out.aPages[0]);
893	if (RT_FAILURE(pReq->Hdr.rc))
894	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
895	return 0;
896	}
897
898	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_UNLOCK):
899	{
900	/* validate */
901	PSUPPAGEUNLOCK pReq = (PSUPPAGEUNLOCK)pReqHdr;
902	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_UNLOCK);
903
904	/* execute */
905	pReq->Hdr.rc = SUPR0UnlockMem(pSession, pReq->u.In.pvR3);
906	return 0;
907	}
908
909	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_ALLOC):
910	{
911	/* validate */
912	PSUPCONTALLOC pReq = (PSUPCONTALLOC)pReqHdr;
913	REQ_CHECK_SIZES(SUP_IOCTL_CONT_ALLOC);
914
915	/* execute */
916	pReq->Hdr.rc = SUPR0ContAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.HCPhys);
917	if (RT_FAILURE(pReq->Hdr.rc))
918	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
919	return 0;
920	}
921
922	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CONT_FREE):
923	{
924	/* validate */
925	PSUPCONTFREE pReq = (PSUPCONTFREE)pReqHdr;
926	REQ_CHECK_SIZES(SUP_IOCTL_CONT_FREE);
927
928	/* execute */
929	pReq->Hdr.rc = SUPR0ContFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
930	return 0;
931	}
932
933	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_OPEN):
934	{
935	/* validate */
936	PSUPLDROPEN pReq = (PSUPLDROPEN)pReqHdr;
937	REQ_CHECK_SIZES(SUP_IOCTL_LDR_OPEN);
938	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImage > 0);
939	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.cbImage < _1M*16);
940	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, pReq->u.In.szName[0]);
941	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, memchr(pReq->u.In.szName, '\0', sizeof(pReq->u.In.szName)));
942	REQ_CHECK_EXPR(SUP_IOCTL_LDR_OPEN, !supdrvCheckInvalidChar(pReq->u.In.szName, ";:()[]{}/\\\|&*%#@!~`\"'"));
943
944	/* execute */
945	pReq->Hdr.rc = supdrvIOCtl_LdrOpen(pDevExt, pSession, pReq);
946	return 0;
947	}
948
949	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_LOAD):
950	{
951	/* validate */
952	PSUPLDRLOAD pReq = (PSUPLDRLOAD)pReqHdr;
953	REQ_CHECK_EXPR(Name, pReq->Hdr.cbIn >= sizeof(*pReq));
954	REQ_CHECK_SIZES_EX(SUP_IOCTL_LDR_LOAD, SUP_IOCTL_LDR_LOAD_SIZE_IN(pReq->u.In.cbImage), SUP_IOCTL_LDR_LOAD_SIZE_OUT);
955	REQ_CHECK_EXPR(SUP_IOCTL_LDR_LOAD, pReq->u.In.cSymbols <= 16384);
956	REQ_CHECK_EXPR_FMT( !pReq->u.In.cSymbols
957	\|\| ( pReq->u.In.offSymbols < pReq->u.In.cbImage
958	&& pReq->u.In.offSymbols + pReq->u.In.cSymbols * sizeof(SUPLDRSYM) <= pReq->u.In.cbImage),
959	("SUP_IOCTL_LDR_LOAD: offSymbols=%#lx cSymbols=%#lx cbImage=%#lx\n", (long)pReq->u.In.offSymbols,
960	(long)pReq->u.In.cSymbols, (long)pReq->u.In.cbImage));
961	REQ_CHECK_EXPR_FMT( !pReq->u.In.cbStrTab
962	\|\| ( pReq->u.In.offStrTab < pReq->u.In.cbImage
963	&& pReq->u.In.offStrTab + pReq->u.In.cbStrTab <= pReq->u.In.cbImage
964	&& pReq->u.In.cbStrTab <= pReq->u.In.cbImage),
965	("SUP_IOCTL_LDR_LOAD: offStrTab=%#lx cbStrTab=%#lx cbImage=%#lx\n", (long)pReq->u.In.offStrTab,
966	(long)pReq->u.In.cbStrTab, (long)pReq->u.In.cbImage));
967
968	if (pReq->u.In.cSymbols)
969	{
970	uint32_t i;
971	PSUPLDRSYM paSyms = (PSUPLDRSYM)&pReq->u.In.achImage[pReq->u.In.offSymbols];
972	for (i = 0; i < pReq->u.In.cSymbols; i++)
973	{
974	REQ_CHECK_EXPR_FMT(paSyms[i].offSymbol < pReq->u.In.cbImage,
975	("SUP_IOCTL_LDR_LOAD: sym #%ld: symb off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offSymbol, (long)pReq->u.In.cbImage));
976	REQ_CHECK_EXPR_FMT(paSyms[i].offName < pReq->u.In.cbStrTab,
977	("SUP_IOCTL_LDR_LOAD: sym #%ld: name off %#lx (max=%#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImage));
978	REQ_CHECK_EXPR_FMT(memchr(&pReq->u.In.achImage[pReq->u.In.offStrTab + paSyms[i].offName], '\0', pReq->u.In.cbStrTab - paSyms[i].offName),
979	("SUP_IOCTL_LDR_LOAD: sym #%ld: unterminated name! (%#lx / %#lx)\n", (long)i, (long)paSyms[i].offName, (long)pReq->u.In.cbImage));
980	}
981	}
982
983	/* execute */
984	pReq->Hdr.rc = supdrvIOCtl_LdrLoad(pDevExt, pSession, pReq);
985	return 0;
986	}
987
988	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_FREE):
989	{
990	/* validate */
991	PSUPLDRFREE pReq = (PSUPLDRFREE)pReqHdr;
992	REQ_CHECK_SIZES(SUP_IOCTL_LDR_FREE);
993
994	/* execute */
995	pReq->Hdr.rc = supdrvIOCtl_LdrFree(pDevExt, pSession, pReq);
996	return 0;
997	}
998
999	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LDR_GET_SYMBOL):
1000	{
1001	/* validate */
1002	PSUPLDRGETSYMBOL pReq = (PSUPLDRGETSYMBOL)pReqHdr;
1003	REQ_CHECK_SIZES(SUP_IOCTL_LDR_GET_SYMBOL);
1004	REQ_CHECK_EXPR(SUP_IOCTL_LDR_GET_SYMBOL, memchr(pReq->u.In.szSymbol, '\0', sizeof(pReq->u.In.szSymbol)));
1005
1006	/* execute */
1007	pReq->Hdr.rc = supdrvIOCtl_LdrGetSymbol(pDevExt, pSession, pReq);
1008	return 0;
1009	}
1010
1011	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_CALL_VMMR0(0)):
1012	{
1013	/* validate */
1014	PSUPCALLVMMR0 pReq = (PSUPCALLVMMR0)pReqHdr;
1015	Log4(("SUP_IOCTL_CALL_VMMR0: op=%u in=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1016	pReq->u.In.uOperation, pReq->Hdr.cbIn, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1017
1018	if (pReq->Hdr.cbIn == SUP_IOCTL_CALL_VMMR0_SIZE(0))
1019	{
1020	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(0), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(0));
1021
1022	/* execute */
1023	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1024	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.uOperation, NULL, pReq->u.In.u64Arg);
1025	else
1026	pReq->Hdr.rc = VERR_WRONG_ORDER;
1027	}
1028	else
1029	{
1030	PSUPVMMR0REQHDR pVMMReq = (PSUPVMMR0REQHDR)&pReq->abReqPkt[0];
1031	REQ_CHECK_EXPR_FMT(pReq->Hdr.cbIn >= SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR)),
1032	("SUP_IOCTL_CALL_VMMR0: cbIn=%#x < %#x\n", pReq->Hdr.cbIn, SUP_IOCTL_CALL_VMMR0_SIZE(sizeof(SUPVMMR0REQHDR))));
1033	REQ_CHECK_EXPR(SUP_IOCTL_CALL_VMMR0, pVMMReq->u32Magic == SUPVMMR0REQHDR_MAGIC);
1034	REQ_CHECK_SIZES_EX(SUP_IOCTL_CALL_VMMR0, SUP_IOCTL_CALL_VMMR0_SIZE_IN(pVMMReq->cbReq), SUP_IOCTL_CALL_VMMR0_SIZE_OUT(pVMMReq->cbReq));
1035
1036	/* execute */
1037	if (RT_LIKELY(pDevExt->pfnVMMR0EntryEx))
1038	pReq->Hdr.rc = pDevExt->pfnVMMR0EntryEx(pReq->u.In.pVMR0, pReq->u.In.uOperation, pVMMReq, pReq->u.In.u64Arg);
1039	else
1040	pReq->Hdr.rc = VERR_WRONG_ORDER;
1041	}
1042
1043	if ( RT_FAILURE(pReq->Hdr.rc)
1044	&& pReq->Hdr.rc != VERR_INTERRUPTED
1045	&& pReq->Hdr.rc != VERR_TIMEOUT)
1046	Log(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1047	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1048	else
1049	Log4(("SUP_IOCTL_CALL_VMMR0: rc=%Rrc op=%u out=%u arg=%RX64 p/t=%RTproc/%RTthrd\n",
1050	pReq->Hdr.rc, pReq->u.In.uOperation, pReq->Hdr.cbOut, pReq->u.In.u64Arg, RTProcSelf(), RTThreadNativeSelf()));
1051	return 0;
1052	}
1053
1054	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GET_PAGING_MODE):
1055	{
1056	/* validate */
1057	PSUPGETPAGINGMODE pReq = (PSUPGETPAGINGMODE)pReqHdr;
1058	REQ_CHECK_SIZES(SUP_IOCTL_GET_PAGING_MODE);
1059
1060	/* execute */
1061	pReq->Hdr.rc = VINF_SUCCESS;
1062	pReq->u.Out.enmMode = supdrvIOCtl_GetPagingMode();
1063	return 0;
1064	}
1065
1066	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_ALLOC):
1067	{
1068	/* validate */
1069	PSUPLOWALLOC pReq = (PSUPLOWALLOC)pReqHdr;
1070	REQ_CHECK_EXPR(SUP_IOCTL_LOW_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_LOW_ALLOC_SIZE_IN);
1071	REQ_CHECK_SIZES_EX(SUP_IOCTL_LOW_ALLOC, SUP_IOCTL_LOW_ALLOC_SIZE_IN, SUP_IOCTL_LOW_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1072
1073	/* execute */
1074	pReq->Hdr.rc = SUPR0LowAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR0, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1075	if (RT_FAILURE(pReq->Hdr.rc))
1076	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1077	return 0;
1078	}
1079
1080	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_LOW_FREE):
1081	{
1082	/* validate */
1083	PSUPLOWFREE pReq = (PSUPLOWFREE)pReqHdr;
1084	REQ_CHECK_SIZES(SUP_IOCTL_LOW_FREE);
1085
1086	/* execute */
1087	pReq->Hdr.rc = SUPR0LowFree(pSession, (RTHCUINTPTR)pReq->u.In.pvR3);
1088	return 0;
1089	}
1090
1091	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_MAP):
1092	{
1093	/* validate */
1094	PSUPGIPMAP pReq = (PSUPGIPMAP)pReqHdr;
1095	REQ_CHECK_SIZES(SUP_IOCTL_GIP_MAP);
1096
1097	/* execute */
1098	pReq->Hdr.rc = SUPR0GipMap(pSession, &pReq->u.Out.pGipR3, &pReq->u.Out.HCPhysGip);
1099	if (RT_SUCCESS(pReq->Hdr.rc))
1100	pReq->u.Out.pGipR0 = pDevExt->pGip;
1101	return 0;
1102	}
1103
1104	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_GIP_UNMAP):
1105	{
1106	/* validate */
1107	PSUPGIPUNMAP pReq = (PSUPGIPUNMAP)pReqHdr;
1108	REQ_CHECK_SIZES(SUP_IOCTL_GIP_UNMAP);
1109
1110	/* execute */
1111	pReq->Hdr.rc = SUPR0GipUnmap(pSession);
1112	return 0;
1113	}
1114
1115	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_SET_VM_FOR_FAST):
1116	{
1117	/* validate */
1118	PSUPSETVMFORFAST pReq = (PSUPSETVMFORFAST)pReqHdr;
1119	REQ_CHECK_SIZES(SUP_IOCTL_SET_VM_FOR_FAST);
1120	REQ_CHECK_EXPR_FMT( !pReq->u.In.pVMR0
1121	\|\| ( VALID_PTR(pReq->u.In.pVMR0)
1122	&& !((uintptr_t)pReq->u.In.pVMR0 & (PAGE_SIZE - 1))),
1123	("SUP_IOCTL_SET_VM_FOR_FAST: pVMR0=%p!\n", pReq->u.In.pVMR0));
1124	/* execute */
1125	pSession->pVM = pReq->u.In.pVMR0;
1126	pReq->Hdr.rc = VINF_SUCCESS;
1127	return 0;
1128	}
1129
1130	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_ALLOC):
1131	{
1132	/* validate */
1133	PSUPPAGEALLOC pReq = (PSUPPAGEALLOC)pReqHdr;
1134	REQ_CHECK_EXPR(SUP_IOCTL_PAGE_ALLOC, pReq->Hdr.cbIn <= SUP_IOCTL_PAGE_ALLOC_SIZE_IN);
1135	REQ_CHECK_SIZES_EX(SUP_IOCTL_PAGE_ALLOC, SUP_IOCTL_PAGE_ALLOC_SIZE_IN, SUP_IOCTL_PAGE_ALLOC_SIZE_OUT(pReq->u.In.cPages));
1136
1137	/* execute */
1138	pReq->Hdr.rc = SUPR0PageAlloc(pSession, pReq->u.In.cPages, &pReq->u.Out.pvR3, &pReq->u.Out.aPages[0]);
1139	if (RT_FAILURE(pReq->Hdr.rc))
1140	pReq->Hdr.cbOut = sizeof(pReq->Hdr);
1141	return 0;
1142	}
1143
1144	case SUP_CTL_CODE_NO_SIZE(SUP_IOCTL_PAGE_FREE):
1145	{
1146	/* validate */
1147	PSUPPAGEFREE pReq = (PSUPPAGEFREE)pReqHdr;
1148	REQ_CHECK_SIZES(SUP_IOCTL_PAGE_FREE);
1149
1150	/* execute */
1151	pReq->Hdr.rc = SUPR0PageFree(pSession, pReq->u.In.pvR3);
1152	return 0;
1153	}
1154
1155	default:
1156	Log(("Unknown IOCTL %#lx\n", (long)uIOCtl));
1157	break;
1158	}
1159	return SUPDRV_ERR_GENERAL_FAILURE;
1160	}
1161
1162
1163	/**
1164	* Register a object for reference counting.
1165	* The object is registered with one reference in the specified session.
1166	*
1167	* @returns Unique identifier on success (pointer).
1168	* All future reference must use this identifier.
1169	* @returns NULL on failure.
1170	* @param pfnDestructor The destructore function which will be called when the reference count reaches 0.
1171	* @param pvUser1 The first user argument.
1172	* @param pvUser2 The second user argument.
1173	*/
1174	SUPR0DECL(void ) SUPR0ObjRegister(PSUPDRVSESSION pSession, SUPDRVOBJTYPE enmType, PFNSUPDRVDESTRUCTOR pfnDestructor, void pvUser1, void *pvUser2)
1175	{
1176	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1177	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1178	PSUPDRVOBJ pObj;
1179	PSUPDRVUSAGE pUsage;
1180
1181	/*
1182	* Validate the input.
1183	*/
1184	AssertReturn(SUP_IS_SESSION_VALID(pSession), NULL);
1185	AssertReturn(enmType > SUPDRVOBJTYPE_INVALID && enmType < SUPDRVOBJTYPE_END, NULL);
1186	AssertPtrReturn(pfnDestructor, NULL);
1187
1188	/*
1189	* Allocate and initialize the object.
1190	*/
1191	pObj = (PSUPDRVOBJ)RTMemAlloc(sizeof(*pObj));
1192	if (!pObj)
1193	return NULL;
1194	pObj->u32Magic = SUPDRVOBJ_MAGIC;
1195	pObj->enmType = enmType;
1196	pObj->pNext = NULL;
1197	pObj->cUsage = 1;
1198	pObj->pfnDestructor = pfnDestructor;
1199	pObj->pvUser1 = pvUser1;
1200	pObj->pvUser2 = pvUser2;
1201	pObj->CreatorUid = pSession->Uid;
1202	pObj->CreatorGid = pSession->Gid;
1203	pObj->CreatorProcess= pSession->Process;
1204	supdrvOSObjInitCreator(pObj, pSession);
1205
1206	/*
1207	* Allocate the usage record.
1208	* (We keep freed usage records around to simplity SUPR0ObjAddRef().)
1209	*/
1210	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1211
1212	pUsage = pDevExt->pUsageFree;
1213	if (pUsage)
1214	pDevExt->pUsageFree = pUsage->pNext;
1215	else
1216	{
1217	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1218	pUsage = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsage));
1219	if (!pUsage)
1220	{
1221	RTMemFree(pObj);
1222	return NULL;
1223	}
1224	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1225	}
1226
1227	/*
1228	* Insert the object and create the session usage record.
1229	*/
1230	/* The object. */
1231	pObj->pNext = pDevExt->pObjs;
1232	pDevExt->pObjs = pObj;
1233
1234	/* The session record. */
1235	pUsage->cUsage = 1;
1236	pUsage->pObj = pObj;
1237	pUsage->pNext = pSession->pUsage;
1238	Log2(("SUPR0ObjRegister: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));
1239	pSession->pUsage = pUsage;
1240
1241	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1242
1243	Log(("SUPR0ObjRegister: returns %p (pvUser1=%p, pvUser=%p)\n", pObj, pvUser1, pvUser2));
1244	return pObj;
1245	}
1246
1247
1248	/**
1249	* Increment the reference counter for the object associating the reference
1250	* with the specified session.
1251	*
1252	* @returns IPRT status code.
1253	* @param pvObj The identifier returned by SUPR0ObjRegister().
1254	* @param pSession The session which is referencing the object.
1255	*/
1256	SUPR0DECL(int) SUPR0ObjAddRef(void *pvObj, PSUPDRVSESSION pSession)
1257	{
1258	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1259	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1260	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
1261	PSUPDRVUSAGE pUsagePre;
1262	PSUPDRVUSAGE pUsage;
1263
1264	/*
1265	* Validate the input.
1266	*/
1267	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1268	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
1269	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
1270	VERR_INVALID_PARAMETER);
1271
1272	/*
1273	* Preallocate the usage record.
1274	*/
1275	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1276
1277	pUsagePre = pDevExt->pUsageFree;
1278	if (pUsagePre)
1279	pDevExt->pUsageFree = pUsagePre->pNext;
1280	else
1281	{
1282	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1283	pUsagePre = (PSUPDRVUSAGE)RTMemAlloc(sizeof(*pUsagePre));
1284	if (!pUsagePre)
1285	return VERR_NO_MEMORY;
1286	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1287	}
1288
1289	/*
1290	* Reference the object.
1291	*/
1292	pObj->cUsage++;
1293
1294	/*
1295	* Look for the session record.
1296	*/
1297	for (pUsage = pSession->pUsage; pUsage; pUsage = pUsage->pNext)
1298	{
1299	Log(("SUPR0AddRef: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));
1300	if (pUsage->pObj == pObj)
1301	break;
1302	}
1303	if (pUsage)
1304	pUsage->cUsage++;
1305	else
1306	{
1307	/* create a new session record. */
1308	pUsagePre->cUsage = 1;
1309	pUsagePre->pObj = pObj;
1310	pUsagePre->pNext = pSession->pUsage;
1311	pSession->pUsage = pUsagePre;
1312	Log(("SUPR0AddRef: pUsagePre=%p:{.pObj=%p, .pNext=%p}\n", pUsagePre, pUsagePre->pObj, pUsagePre->pNext));
1313
1314	pUsagePre = NULL;
1315	}
1316
1317	/*
1318	* Put any unused usage record into the free list..
1319	*/
1320	if (pUsagePre)
1321	{
1322	pUsagePre->pNext = pDevExt->pUsageFree;
1323	pDevExt->pUsageFree = pUsagePre;
1324	}
1325
1326	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1327
1328	return VINF_SUCCESS;
1329	}
1330
1331
1332	/**
1333	* Decrement / destroy a reference counter record for an object.
1334	*
1335	* The object is uniquely identified by pfnDestructor+pvUser1+pvUser2.
1336	*
1337	* @returns IPRT status code.
1338	* @param pvObj The identifier returned by SUPR0ObjRegister().
1339	* @param pSession The session which is referencing the object.
1340	*/
1341	SUPR0DECL(int) SUPR0ObjRelease(void *pvObj, PSUPDRVSESSION pSession)
1342	{
1343	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1344	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
1345	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
1346	bool fDestroy = false;
1347	PSUPDRVUSAGE pUsage;
1348	PSUPDRVUSAGE pUsagePrev;
1349
1350	/*
1351	* Validate the input.
1352	*/
1353	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1354	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
1355	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
1356	VERR_INVALID_PARAMETER);
1357
1358	/*
1359	* Acquire the spinlock and look for the usage record.
1360	*/
1361	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
1362
1363	for (pUsagePrev = NULL, pUsage = pSession->pUsage;
1364	pUsage;
1365	pUsagePrev = pUsage, pUsage = pUsage->pNext)
1366	{
1367	Log2(("SUPR0ObjRelease: pUsage=%p:{.pObj=%p, .pNext=%p}\n", pUsage, pUsage->pObj, pUsage->pNext));
1368	if (pUsage->pObj == pObj)
1369	{
1370	AssertMsg(pUsage->cUsage >= 1 && pObj->cUsage >= pUsage->cUsage, ("glob %d; sess %d\n", pObj->cUsage, pUsage->cUsage));
1371	if (pUsage->cUsage > 1)
1372	{
1373	pObj->cUsage--;
1374	pUsage->cUsage--;
1375	}
1376	else
1377	{
1378	/*
1379	* Free the session record.
1380	*/
1381	if (pUsagePrev)
1382	pUsagePrev->pNext = pUsage->pNext;
1383	else
1384	pSession->pUsage = pUsage->pNext;
1385	pUsage->pNext = pDevExt->pUsageFree;
1386	pDevExt->pUsageFree = pUsage;
1387
1388	/* What about the object? */
1389	if (pObj->cUsage > 1)
1390	pObj->cUsage--;
1391	else
1392	{
1393	/*
1394	* Object is to be destroyed, unlink it.
1395	*/
1396	pObj->u32Magic = SUPDRVOBJ_MAGIC + 1;
1397	fDestroy = true;
1398	if (pDevExt->pObjs == pObj)
1399	pDevExt->pObjs = pObj->pNext;
1400	else
1401	{
1402	PSUPDRVOBJ pObjPrev;
1403	for (pObjPrev = pDevExt->pObjs; pObjPrev; pObjPrev = pObjPrev->pNext)
1404	if (pObjPrev->pNext == pObj)
1405	{
1406	pObjPrev->pNext = pObj->pNext;
1407	break;
1408	}
1409	Assert(pObjPrev);
1410	}
1411	}
1412	}
1413	break;
1414	}
1415	}
1416
1417	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
1418
1419	/*
1420	* Call the destructor and free the object if required.
1421	*/
1422	if (fDestroy)
1423	{
1424	Log(("SUPR0ObjRelease: destroying %p/%d (%p/%p) cpid=%RTproc pid=%RTproc dtor=%p\n",
1425	pObj, pObj->enmType, pObj->pvUser1, pObj->pvUser2, pObj->CreatorProcess, RTProcSelf(), pObj->pfnDestructor));
1426	if (pObj->pfnDestructor)
1427	pObj->pfnDestructor(pObj, pObj->pvUser1, pObj->pvUser2);
1428	RTMemFree(pObj);
1429	}
1430
1431	AssertMsg(pUsage, ("pvObj=%p\n", pvObj));
1432	return pUsage ? VINF_SUCCESS : VERR_INVALID_PARAMETER;
1433	}
1434
1435	/**
1436	* Verifies that the current process can access the specified object.
1437	*
1438	* @returns The following IPRT status code:
1439	* @retval VINF_SUCCESS if access was granted.
1440	* @retval VERR_PERMISSION_DENIED if denied access.
1441	* @retval VERR_INVALID_PARAMETER if invalid parameter.
1442	*
1443	* @param pvObj The identifier returned by SUPR0ObjRegister().
1444	* @param pSession The session which wishes to access the object.
1445	* @param pszObjName Object string name. This is optional and depends on the object type.
1446	*
1447	* @remark The caller is responsible for making sure the object isn't removed while
1448	* we're inside this function. If uncertain about this, just call AddRef before calling us.
1449	*/
1450	SUPR0DECL(int) SUPR0ObjVerifyAccess(void pvObj, PSUPDRVSESSION pSession, const char pszObjName)
1451	{
1452	PSUPDRVOBJ pObj = (PSUPDRVOBJ)pvObj;
1453	int rc;
1454
1455	/*
1456	* Validate the input.
1457	*/
1458	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1459	AssertMsgReturn(VALID_PTR(pObj) && pObj->u32Magic == SUPDRVOBJ_MAGIC,
1460	("Invalid pvObj=%p magic=%#x (exepcted %#x)\n", pvObj, pObj ? pObj->u32Magic : 0, SUPDRVOBJ_MAGIC),
1461	VERR_INVALID_PARAMETER);
1462
1463	/*
1464	* Check access. (returns true if a decision has been made.)
1465	*/
1466	rc = VERR_INTERNAL_ERROR;
1467	if (supdrvOSObjCanAccess(pObj, pSession, pszObjName, &rc))
1468	return rc;
1469
1470	/*
1471	* Default policy is to allow the user to access his own
1472	* stuff but nothing else.
1473	*/
1474	if (pObj->CreatorUid == pSession->Uid)
1475	return VINF_SUCCESS;
1476	return VERR_PERMISSION_DENIED;
1477	}
1478
1479
1480	/**
1481	* Lock pages.
1482	*
1483	* @returns IPRT status code.
1484	* @param pSession Session to which the locked memory should be associated.
1485	* @param pvR3 Start of the memory range to lock.
1486	* This must be page aligned.
1487	* @param cb Size of the memory range to lock.
1488	* This must be page aligned.
1489	*/
1490	SUPR0DECL(int) SUPR0LockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
1491	{
1492	int rc;
1493	SUPDRVMEMREF Mem = {0};
1494	const size_t cb = (size_t)cPages << PAGE_SHIFT;
1495	LogFlow(("SUPR0LockMem: pSession=%p pvR3=%p cPages=%d paPages=%p\n", pSession, (void *)pvR3, cPages, paPages));
1496
1497	/*
1498	* Verify input.
1499	*/
1500	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1501	AssertPtrReturn(paPages, VERR_INVALID_PARAMETER);
1502	if ( RT_ALIGN_R3PT(pvR3, PAGE_SIZE, RTR3PTR) != pvR3
1503	\|\| !pvR3)
1504	{
1505	Log(("pvR3 (%p) must be page aligned and not NULL!\n", (void *)pvR3));
1506	return VERR_INVALID_PARAMETER;
1507	}
1508
1509	#ifdef RT_OS_WINDOWS /* A temporary hack for windows, will be removed once all ring-3 code has been cleaned up. */
1510	/* First check if we allocated it using SUPPageAlloc; if so then we don't need to lock it again */
1511	rc = supdrvPageGetPhys(pSession, pvR3, cPages, paPages);
1512	if (RT_SUCCESS(rc))
1513	return rc;
1514	#endif
1515
1516	/*
1517	* Let IPRT do the job.
1518	*/
1519	Mem.eType = MEMREF_TYPE_LOCKED;
1520	rc = RTR0MemObjLockUser(&Mem.MemObj, pvR3, cb, RTR0ProcHandleSelf());
1521	if (RT_SUCCESS(rc))
1522	{
1523	uint32_t iPage = cPages;
1524	AssertMsg(RTR0MemObjAddressR3(Mem.MemObj) == pvR3, ("%p == %p\n", RTR0MemObjAddressR3(Mem.MemObj), pvR3));
1525	AssertMsg(RTR0MemObjSize(Mem.MemObj) == cb, ("%x == %x\n", RTR0MemObjSize(Mem.MemObj), cb));
1526
1527	while (iPage-- > 0)
1528	{
1529	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
1530	if (RT_UNLIKELY(paPages[iPage] == NIL_RTCCPHYS))
1531	{
1532	AssertMsgFailed(("iPage=%d\n", iPage));
1533	rc = VERR_INTERNAL_ERROR;
1534	break;
1535	}
1536	}
1537	if (RT_SUCCESS(rc))
1538	rc = supdrvMemAdd(&Mem, pSession);
1539	if (RT_FAILURE(rc))
1540	{
1541	int rc2 = RTR0MemObjFree(Mem.MemObj, false);
1542	AssertRC(rc2);
1543	}
1544	}
1545
1546	return rc;
1547	}
1548
1549
1550	/**
1551	* Unlocks the memory pointed to by pv.
1552	*
1553	* @returns IPRT status code.
1554	* @param pSession Session to which the memory was locked.
1555	* @param pvR3 Memory to unlock.
1556	*/
1557	SUPR0DECL(int) SUPR0UnlockMem(PSUPDRVSESSION pSession, RTR3PTR pvR3)
1558	{
1559	LogFlow(("SUPR0UnlockMem: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
1560	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1561	#ifdef RT_OS_WINDOWS
1562	/*
1563	* Temporary hack for windows - SUPR0PageFree will unlock SUPR0PageAlloc
1564	* allocations; ignore this call.
1565	*/
1566	if (supdrvPageWasLockedByPageAlloc(pSession, pvR3))
1567	{
1568	LogFlow(("Page will be unlocked in SUPR0PageFree -> ignore\n"));
1569	return VINF_SUCCESS;
1570	}
1571	#endif
1572	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED);
1573	}
1574
1575
1576	/**
1577	* Allocates a chunk of page aligned memory with contiguous and fixed physical
1578	* backing.
1579	*
1580	* @returns IPRT status code.
1581	* @param pSession Session data.
1582	* @param cb Number of bytes to allocate.
1583	* @param ppvR0 Where to put the address of Ring-0 mapping the allocated memory.
1584	* @param ppvR3 Where to put the address of Ring-3 mapping the allocated memory.
1585	* @param pHCPhys Where to put the physical address of allocated memory.
1586	*/
1587	SUPR0DECL(int) SUPR0ContAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS pHCPhys)
1588	{
1589	int rc;
1590	SUPDRVMEMREF Mem = {0};
1591	LogFlow(("SUPR0ContAlloc: pSession=%p cPages=%d ppvR0=%p ppvR3=%p pHCPhys=%p\n", pSession, cPages, ppvR0, ppvR3, pHCPhys));
1592
1593	/*
1594	* Validate input.
1595	*/
1596	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1597	if (!ppvR3 \|\| !ppvR0 \|\| !pHCPhys)
1598	{
1599	Log(("Null pointer. All of these should be set: pSession=%p ppvR0=%p ppvR3=%p pHCPhys=%p\n",
1600	pSession, ppvR0, ppvR3, pHCPhys));
1601	return VERR_INVALID_PARAMETER;
1602
1603	}
1604	if (cPages < 1 \|\| cPages >= 256)
1605	{
1606	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256\n", cPages));
1607	return VERR_INVALID_PARAMETER;
1608	}
1609
1610	/*
1611	* Let IPRT do the job.
1612	*/
1613	rc = RTR0MemObjAllocCont(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable R0 mapping */);
1614	if (RT_SUCCESS(rc))
1615	{
1616	int rc2;
1617	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
1618	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
1619	if (RT_SUCCESS(rc))
1620	{
1621	Mem.eType = MEMREF_TYPE_CONT;
1622	rc = supdrvMemAdd(&Mem, pSession);
1623	if (!rc)
1624	{
1625	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
1626	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
1627	*pHCPhys = RTR0MemObjGetPagePhysAddr(Mem.MemObj, 0);
1628	return 0;
1629	}
1630
1631	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
1632	AssertRC(rc2);
1633	}
1634	rc2 = RTR0MemObjFree(Mem.MemObj, false);
1635	AssertRC(rc2);
1636	}
1637
1638	return rc;
1639	}
1640
1641
1642	/**
1643	* Frees memory allocated using SUPR0ContAlloc().
1644	*
1645	* @returns IPRT status code.
1646	* @param pSession The session to which the memory was allocated.
1647	* @param uPtr Pointer to the memory (ring-3 or ring-0).
1648	*/
1649	SUPR0DECL(int) SUPR0ContFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
1650	{
1651	LogFlow(("SUPR0ContFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
1652	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1653	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_CONT);
1654	}
1655
1656
1657	/**
1658	* Allocates a chunk of page aligned memory with fixed physical backing below 4GB.
1659	*
1660	* The memory isn't zeroed.
1661	*
1662	* @returns IPRT status code.
1663	* @param pSession Session data.
1664	* @param cPages Number of pages to allocate.
1665	* @param ppvR0 Where to put the address of Ring-0 mapping of the allocated memory.
1666	* @param ppvR3 Where to put the address of Ring-3 mapping of the allocated memory.
1667	* @param paPages Where to put the physical addresses of allocated memory.
1668	*/
1669	SUPR0DECL(int) SUPR0LowAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR0PTR ppvR0, PRTR3PTR ppvR3, PRTHCPHYS paPages)
1670	{
1671	unsigned iPage;
1672	int rc;
1673	SUPDRVMEMREF Mem = {0};
1674	LogFlow(("SUPR0LowAlloc: pSession=%p cPages=%d ppvR3=%p ppvR0=%p paPages=%p\n", pSession, cPages, ppvR3, ppvR0, paPages));
1675
1676	/*
1677	* Validate input.
1678	*/
1679	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1680	if (!ppvR3 \|\| !ppvR0 \|\| !paPages)
1681	{
1682	Log(("Null pointer. All of these should be set: pSession=%p ppvR3=%p ppvR0=%p paPages=%p\n",
1683	pSession, ppvR3, ppvR0, paPages));
1684	return VERR_INVALID_PARAMETER;
1685
1686	}
1687	if (cPages < 1 \|\| cPages > 256)
1688	{
1689	Log(("Illegal request cPages=%d, must be greater than 0 and smaller than 256.\n", cPages));
1690	return VERR_INVALID_PARAMETER;
1691	}
1692
1693	/*
1694	* Let IPRT do the work.
1695	*/
1696	rc = RTR0MemObjAllocLow(&Mem.MemObj, cPages << PAGE_SHIFT, true /* executable ring-0 mapping */);
1697	if (RT_SUCCESS(rc))
1698	{
1699	int rc2;
1700	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
1701	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
1702	if (RT_SUCCESS(rc))
1703	{
1704	Mem.eType = MEMREF_TYPE_LOW;
1705	rc = supdrvMemAdd(&Mem, pSession);
1706	if (!rc)
1707	{
1708	for (iPage = 0; iPage < cPages; iPage++)
1709	{
1710	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MemObj, iPage);
1711	AssertMsg(!(paPages[iPage] & (PAGE_SIZE - 1)), ("iPage=%d Phys=%VHp\n", paPages[iPage]));
1712	}
1713	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
1714	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
1715	return 0;
1716	}
1717
1718	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
1719	AssertRC(rc2);
1720	}
1721
1722	rc2 = RTR0MemObjFree(Mem.MemObj, false);
1723	AssertRC(rc2);
1724	}
1725
1726	return rc;
1727	}
1728
1729
1730	/**
1731	* Frees memory allocated using SUPR0LowAlloc().
1732	*
1733	* @returns IPRT status code.
1734	* @param pSession The session to which the memory was allocated.
1735	* @param uPtr Pointer to the memory (ring-3 or ring-0).
1736	*/
1737	SUPR0DECL(int) SUPR0LowFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
1738	{
1739	LogFlow(("SUPR0LowFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
1740	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1741	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_LOW);
1742	}
1743
1744
1745
1746	/**
1747	* Allocates a chunk of memory with both R0 and R3 mappings.
1748	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
1749	*
1750	* @returns IPRT status code.
1751	* @param pSession The session to associated the allocation with.
1752	* @param cb Number of bytes to allocate.
1753	* @param ppvR0 Where to store the address of the Ring-0 mapping.
1754	* @param ppvR3 Where to store the address of the Ring-3 mapping.
1755	*/
1756	SUPR0DECL(int) SUPR0MemAlloc(PSUPDRVSESSION pSession, uint32_t cb, PRTR0PTR ppvR0, PRTR3PTR ppvR3)
1757	{
1758	int rc;
1759	SUPDRVMEMREF Mem = {0};
1760	LogFlow(("SUPR0MemAlloc: pSession=%p cb=%d ppvR0=%p ppvR3=%p\n", pSession, cb, ppvR0, ppvR3));
1761
1762	/*
1763	* Validate input.
1764	*/
1765	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1766	AssertPtrReturn(ppvR0, VERR_INVALID_POINTER);
1767	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
1768	if (cb < 1 \|\| cb >= _4M)
1769	{
1770	Log(("Illegal request cb=%u; must be greater than 0 and smaller than 4MB.\n", cb));
1771	return VERR_INVALID_PARAMETER;
1772	}
1773
1774	/*
1775	* Let IPRT do the work.
1776	*/
1777	rc = RTR0MemObjAllocPage(&Mem.MemObj, cb, true /* executable ring-0 mapping */);
1778	if (RT_SUCCESS(rc))
1779	{
1780	int rc2;
1781	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
1782	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
1783	if (RT_SUCCESS(rc))
1784	{
1785	Mem.eType = MEMREF_TYPE_MEM;
1786	rc = supdrvMemAdd(&Mem, pSession);
1787	if (!rc)
1788	{
1789	*ppvR0 = RTR0MemObjAddress(Mem.MemObj);
1790	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
1791	return VINF_SUCCESS;
1792	}
1793	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
1794	AssertRC(rc2);
1795	}
1796
1797	rc2 = RTR0MemObjFree(Mem.MemObj, false);
1798	AssertRC(rc2);
1799	}
1800
1801	return rc;
1802	}
1803
1804
1805	/**
1806	* Get the physical addresses of memory allocated using SUPR0MemAlloc().
1807	*
1808	* @returns IPRT status code.
1809	* @param pSession The session to which the memory was allocated.
1810	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
1811	* @param paPages Where to store the physical addresses.
1812	*/
1813	SUPR0DECL(int) SUPR0MemGetPhys(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, PSUPPAGE paPages) /** @todo switch this bugger to RTHCPHYS */
1814	{
1815	PSUPDRVBUNDLE pBundle;
1816	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1817	LogFlow(("SUPR0MemGetPhys: pSession=%p uPtr=%p paPages=%p\n", pSession, (void *)uPtr, paPages));
1818
1819	/*
1820	* Validate input.
1821	*/
1822	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1823	AssertPtrReturn(paPages, VERR_INVALID_POINTER);
1824	AssertReturn(uPtr, VERR_INVALID_PARAMETER);
1825
1826	/*
1827	* Search for the address.
1828	*/
1829	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
1830	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
1831	{
1832	if (pBundle->cUsed > 0)
1833	{
1834	unsigned i;
1835	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
1836	{
1837	if ( pBundle->aMem[i].eType == MEMREF_TYPE_MEM
1838	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
1839	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
1840	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
1841	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr)
1842	)
1843	)
1844	{
1845	const unsigned cPages = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
1846	unsigned iPage;
1847	for (iPage = 0; iPage < cPages; iPage++)
1848	{
1849	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
1850	paPages[iPage].uReserved = 0;
1851	}
1852	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
1853	return VINF_SUCCESS;
1854	}
1855	}
1856	}
1857	}
1858	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
1859	Log(("Failed to find %p!!!\n", (void *)uPtr));
1860	return VERR_INVALID_PARAMETER;
1861	}
1862
1863
1864	/**
1865	* Free memory allocated by SUPR0MemAlloc().
1866	*
1867	* @returns IPRT status code.
1868	* @param pSession The session owning the allocation.
1869	* @param uPtr The Ring-0 or Ring-3 address returned by SUPR0MemAlloc().
1870	*/
1871	SUPR0DECL(int) SUPR0MemFree(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr)
1872	{
1873	LogFlow(("SUPR0MemFree: pSession=%p uPtr=%p\n", pSession, (void *)uPtr));
1874	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1875	return supdrvMemRelease(pSession, uPtr, MEMREF_TYPE_MEM);
1876	}
1877
1878
1879	/**
1880	* Allocates a chunk of memory with only a R3 mappings.
1881	* The memory is fixed and it's possible to query the physical addresses using SUPR0MemGetPhys().
1882	*
1883	* @returns IPRT status code.
1884	* @param pSession The session to associated the allocation with.
1885	* @param cPages The number of pages to allocate.
1886	* @param ppvR3 Where to store the address of the Ring-3 mapping.
1887	* @param paPages Where to store the addresses of the pages. Optional.
1888	*/
1889	SUPR0DECL(int) SUPR0PageAlloc(PSUPDRVSESSION pSession, uint32_t cPages, PRTR3PTR ppvR3, PRTHCPHYS paPages)
1890	{
1891	int rc;
1892	SUPDRVMEMREF Mem = {0};
1893	LogFlow(("SUPR0PageAlloc: pSession=%p cb=%d ppvR3=%p\n", pSession, cPages, ppvR3));
1894
1895	/*
1896	* Validate input. The allowed allocation size must be at least equal to the maximum guest VRAM size.
1897	*/
1898	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
1899	AssertPtrReturn(ppvR3, VERR_INVALID_POINTER);
1900	if (cPages < 1 \|\| cPages > (128 * _1M)/PAGE_SIZE)
1901	{
1902	Log(("SUPR0PageAlloc: Illegal request cb=%u; must be greater than 0 and smaller than 128MB.\n", cPages));
1903	return VERR_INVALID_PARAMETER;
1904	}
1905
1906	/*
1907	* Let IPRT do the work.
1908	*/
1909	rc = RTR0MemObjAllocPhysNC(&Mem.MemObj, (size_t)cPages * PAGE_SIZE, NIL_RTHCPHYS);
1910	if (RT_SUCCESS(rc))
1911	{
1912	int rc2;
1913	rc = RTR0MemObjMapUser(&Mem.MapObjR3, Mem.MemObj, (RTR3PTR)-1, 0,
1914	RTMEM_PROT_EXEC \| RTMEM_PROT_WRITE \| RTMEM_PROT_READ, RTR0ProcHandleSelf());
1915	if (RT_SUCCESS(rc))
1916	{
1917	Mem.eType = MEMREF_TYPE_LOCKED_SUP;
1918	rc = supdrvMemAdd(&Mem, pSession);
1919	if (!rc)
1920	{
1921	*ppvR3 = RTR0MemObjAddressR3(Mem.MapObjR3);
1922	if (paPages)
1923	{
1924	uint32_t iPage = cPages;
1925	while (iPage-- > 0)
1926	{
1927	paPages[iPage] = RTR0MemObjGetPagePhysAddr(Mem.MapObjR3, iPage);
1928	Assert(paPages[iPage] != NIL_RTHCPHYS);
1929	}
1930	}
1931	return VINF_SUCCESS;
1932	}
1933	rc2 = RTR0MemObjFree(Mem.MapObjR3, false);
1934	AssertRC(rc2);
1935	}
1936
1937	rc2 = RTR0MemObjFree(Mem.MemObj, false);
1938	AssertRC(rc2);
1939	}
1940	return rc;
1941	}
1942
1943
1944	#ifdef RT_OS_WINDOWS
1945	/**
1946	* Check if the pages were locked by SUPR0PageAlloc
1947	*
1948	* This function will be removed along with the lock/unlock hacks when
1949	* we've cleaned up the ring-3 code properly.
1950	*
1951	* @returns boolean
1952	* @param pSession The session to which the memory was allocated.
1953	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc().
1954	*/
1955	static bool supdrvPageWasLockedByPageAlloc(PSUPDRVSESSION pSession, RTR3PTR pvR3)
1956	{
1957	PSUPDRVBUNDLE pBundle;
1958	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
1959	LogFlow(("SUPR0PageIsLockedByPageAlloc: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
1960
1961	/*
1962	* Search for the address.
1963	*/
1964	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
1965	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
1966	{
1967	if (pBundle->cUsed > 0)
1968	{
1969	unsigned i;
1970	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
1971	{
1972	if ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED_SUP
1973	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
1974	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
1975	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
1976	{
1977	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
1978	return true;
1979	}
1980	}
1981	}
1982	}
1983	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
1984	return false;
1985	}
1986
1987
1988	/**
1989	* Get the physical addresses of memory allocated using SUPR0PageAlloc().
1990	*
1991	* This function will be removed along with the lock/unlock hacks when
1992	* we've cleaned up the ring-3 code properly.
1993	*
1994	* @returns IPRT status code.
1995	* @param pSession The session to which the memory was allocated.
1996	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc().
1997	* @param cPages Number of pages in paPages
1998	* @param paPages Where to store the physical addresses.
1999	*/
2000	static int supdrvPageGetPhys(PSUPDRVSESSION pSession, RTR3PTR pvR3, uint32_t cPages, PRTHCPHYS paPages)
2001	{
2002	PSUPDRVBUNDLE pBundle;
2003	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2004	LogFlow(("supdrvPageGetPhys: pSession=%p pvR3=%p cPages=%#lx paPages=%p\n", pSession, (void *)pvR3, (long)cPages, paPages));
2005
2006	/*
2007	* Search for the address.
2008	*/
2009	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2010	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2011	{
2012	if (pBundle->cUsed > 0)
2013	{
2014	unsigned i;
2015	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2016	{
2017	if ( pBundle->aMem[i].eType == MEMREF_TYPE_LOCKED_SUP
2018	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2019	&& pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2020	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == pvR3)
2021	{
2022	uint32_t iPage = RTR0MemObjSize(pBundle->aMem[i].MemObj) >> PAGE_SHIFT;
2023	cPages = RT_MIN(iPage, cPages);
2024	for (iPage = 0; iPage < cPages; iPage++)
2025	paPages[iPage] = RTR0MemObjGetPagePhysAddr(pBundle->aMem[i].MemObj, iPage);
2026	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2027	return VINF_SUCCESS;
2028	}
2029	}
2030	}
2031	}
2032	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2033	return VERR_INVALID_PARAMETER;
2034	}
2035	#endif /* RT_OS_WINDOWS */
2036
2037
2038	/**
2039	* Free memory allocated by SUPR0PageAlloc().
2040	*
2041	* @returns IPRT status code.
2042	* @param pSession The session owning the allocation.
2043	* @param pvR3 The Ring-3 address returned by SUPR0PageAlloc().
2044	*/
2045	SUPR0DECL(int) SUPR0PageFree(PSUPDRVSESSION pSession, RTR3PTR pvR3)
2046	{
2047	LogFlow(("SUPR0PageFree: pSession=%p pvR3=%p\n", pSession, (void *)pvR3));
2048	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2049	return supdrvMemRelease(pSession, (RTHCUINTPTR)pvR3, MEMREF_TYPE_LOCKED_SUP);
2050	}
2051
2052
2053	/**
2054	* Maps the GIP into userspace and/or get the physical address of the GIP.
2055	*
2056	* @returns IPRT status code.
2057	* @param pSession Session to which the GIP mapping should belong.
2058	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
2059	* @param pHCPhysGip Where to store the physical address. (optional)
2060	*
2061	* @remark There is no reference counting on the mapping, so one call to this function
2062	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
2063	* and remove the session as a GIP user.
2064	*/
2065	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
2066	{
2067	int rc = 0;
2068	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2069	RTR3PTR pGip = NIL_RTR3PTR;
2070	RTHCPHYS HCPhys = NIL_RTHCPHYS;
2071	LogFlow(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
2072
2073	/*
2074	* Validate
2075	*/
2076	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2077	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
2078	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
2079
2080	RTSemFastMutexRequest(pDevExt->mtxGip);
2081	if (pDevExt->pGip)
2082	{
2083	/*
2084	* Map it?
2085	*/
2086	if (ppGipR3)
2087	{
2088	#ifdef USE_NEW_OS_INTERFACE_FOR_GIP
2089	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
2090	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
2091	RTMEM_PROT_READ, RTR0ProcHandleSelf());
2092	if (RT_SUCCESS(rc))
2093	{
2094	pGip = RTR0MemObjAddressR3(pSession->GipMapObjR3);
2095	rc = VINF_SUCCESS; /** @todo remove this and replace the !rc below with RT_SUCCESS(rc). */
2096	}
2097	#else /* !USE_NEW_OS_INTERFACE_FOR_GIP */
2098	if (!pSession->pGip)
2099	rc = supdrvOSGipMap(pSession->pDevExt, &pSession->pGip);
2100	if (!rc)
2101	pGip = (RTR3PTR)pSession->pGip;
2102	#endif /* !USE_NEW_OS_INTERFACE_FOR_GIP */
2103	}
2104
2105	/*
2106	* Get physical address.
2107	*/
2108	if (pHCPhysGip && !rc)
2109	HCPhys = pDevExt->HCPhysGip;
2110
2111	/*
2112	* Reference globally.
2113	*/
2114	if (!pSession->fGipReferenced && !rc)
2115	{
2116	pSession->fGipReferenced = 1;
2117	pDevExt->cGipUsers++;
2118	if (pDevExt->cGipUsers == 1)
2119	{
2120	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
2121	unsigned i;
2122
2123	LogFlow(("SUPR0GipMap: Resumes GIP updating\n"));
2124
2125	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
2126	ASMAtomicXchgU32(&pGip->aCPUs[i].u32TransactionId, pGip->aCPUs[i].u32TransactionId & ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
2127	ASMAtomicXchgU64(&pGip->u64NanoTSLastUpdateHz, 0);
2128
2129	#ifdef USE_NEW_OS_INTERFACE_FOR_GIP
2130	rc = RTTimerStart(pDevExt->pGipTimer, 0);
2131	AssertRC(rc); rc = VINF_SUCCESS;
2132	#else
2133	supdrvOSGipResume(pDevExt);
2134	#endif
2135	}
2136	}
2137	}
2138	else
2139	{
2140	rc = SUPDRV_ERR_GENERAL_FAILURE;
2141	Log(("SUPR0GipMap: GIP is not available!\n"));
2142	}
2143	RTSemFastMutexRelease(pDevExt->mtxGip);
2144
2145	/*
2146	* Write returns.
2147	*/
2148	if (pHCPhysGip)
2149	*pHCPhysGip = HCPhys;
2150	if (ppGipR3)
2151	*ppGipR3 = pGip;
2152
2153	#ifdef DEBUG_DARWIN_GIP
2154	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx ppGip=%p GipMapObjR3\n", rc, (unsigned long)HCPhys, pGip, pSession->GipMapObjR3));
2155	#else
2156	LogFlow(("SUPR0GipMap: returns %d pHCPhysGip=%lx ppGipR3=%p\n", rc, (unsigned long)HCPhys, (void *)(uintptr_t)pGip));
2157	#endif
2158	return rc;
2159	}
2160
2161
2162	/**
2163	* Unmaps any user mapping of the GIP and terminates all GIP access
2164	* from this session.
2165	*
2166	* @returns IPRT status code.
2167	* @param pSession Session to which the GIP mapping should belong.
2168	*/
2169	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
2170	{
2171	int rc = VINF_SUCCESS;
2172	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
2173	#ifdef DEBUG_DARWIN_GIP
2174	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
2175	pSession,
2176	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
2177	pSession->GipMapObjR3));
2178	#else
2179	LogFlow(("SUPR0GipUnmap: pSession=%p\n", pSession));
2180	#endif
2181	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
2182
2183	RTSemFastMutexRequest(pDevExt->mtxGip);
2184
2185	/*
2186	* Unmap anything?
2187	*/
2188	#ifdef USE_NEW_OS_INTERFACE_FOR_GIP
2189	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
2190	{
2191	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
2192	AssertRC(rc);
2193	if (RT_SUCCESS(rc))
2194	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
2195	}
2196	#else
2197	if (pSession->pGip)
2198	{
2199	rc = supdrvOSGipUnmap(pDevExt, pSession->pGip);
2200	if (!rc)
2201	pSession->pGip = NULL;
2202	}
2203	#endif
2204
2205	/*
2206	* Dereference global GIP.
2207	*/
2208	if (pSession->fGipReferenced && !rc)
2209	{
2210	pSession->fGipReferenced = 0;
2211	if ( pDevExt->cGipUsers > 0
2212	&& !--pDevExt->cGipUsers)
2213	{
2214	LogFlow(("SUPR0GipUnmap: Suspends GIP updating\n"));
2215	#ifdef USE_NEW_OS_INTERFACE_FOR_GIP
2216	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = 0;
2217	#else
2218	supdrvOSGipSuspend(pDevExt);
2219	#endif
2220	}
2221	}
2222
2223	RTSemFastMutexRelease(pDevExt->mtxGip);
2224
2225	return rc;
2226	}
2227
2228
2229	/**
2230	* Adds a memory object to the session.
2231	*
2232	* @returns IPRT status code.
2233	* @param pMem Memory tracking structure containing the
2234	* information to track.
2235	* @param pSession The session.
2236	*/
2237	static int supdrvMemAdd(PSUPDRVMEMREF pMem, PSUPDRVSESSION pSession)
2238	{
2239	PSUPDRVBUNDLE pBundle;
2240	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2241
2242	/*
2243	* Find free entry and record the allocation.
2244	*/
2245	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2246	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2247	{
2248	if (pBundle->cUsed < RT_ELEMENTS(pBundle->aMem))
2249	{
2250	unsigned i;
2251	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2252	{
2253	if (pBundle->aMem[i].MemObj == NIL_RTR0MEMOBJ)
2254	{
2255	pBundle->cUsed++;
2256	pBundle->aMem[i] = *pMem;
2257	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2258	return VINF_SUCCESS;
2259	}
2260	}
2261	AssertFailed(); /* !!this can't be happening!!! */
2262	}
2263	}
2264	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2265
2266	/*
2267	* Need to allocate a new bundle.
2268	* Insert into the last entry in the bundle.
2269	*/
2270	pBundle = (PSUPDRVBUNDLE)RTMemAllocZ(sizeof(*pBundle));
2271	if (!pBundle)
2272	return VERR_NO_MEMORY;
2273
2274	/* take last entry. */
2275	pBundle->cUsed++;
2276	pBundle->aMem[RT_ELEMENTS(pBundle->aMem) - 1] = *pMem;
2277
2278	/* insert into list. */
2279	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2280	pBundle->pNext = pSession->Bundle.pNext;
2281	pSession->Bundle.pNext = pBundle;
2282	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2283
2284	return VINF_SUCCESS;
2285	}
2286
2287
2288	/**
2289	* Releases a memory object referenced by pointer and type.
2290	*
2291	* @returns IPRT status code.
2292	* @param pSession Session data.
2293	* @param uPtr Pointer to memory. This is matched against both the R0 and R3 addresses.
2294	* @param eType Memory type.
2295	*/
2296	static int supdrvMemRelease(PSUPDRVSESSION pSession, RTHCUINTPTR uPtr, SUPDRVMEMREFTYPE eType)
2297	{
2298	PSUPDRVBUNDLE pBundle;
2299	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2300
2301	/*
2302	* Validate input.
2303	*/
2304	if (!uPtr)
2305	{
2306	Log(("Illegal address %p\n", (void *)uPtr));
2307	return VERR_INVALID_PARAMETER;
2308	}
2309
2310	/*
2311	* Search for the address.
2312	*/
2313	RTSpinlockAcquire(pSession->Spinlock, &SpinlockTmp);
2314	for (pBundle = &pSession->Bundle; pBundle; pBundle = pBundle->pNext)
2315	{
2316	if (pBundle->cUsed > 0)
2317	{
2318	unsigned i;
2319	for (i = 0; i < RT_ELEMENTS(pBundle->aMem); i++)
2320	{
2321	if ( pBundle->aMem[i].eType == eType
2322	&& pBundle->aMem[i].MemObj != NIL_RTR0MEMOBJ
2323	&& ( (RTHCUINTPTR)RTR0MemObjAddress(pBundle->aMem[i].MemObj) == uPtr
2324	\|\| ( pBundle->aMem[i].MapObjR3 != NIL_RTR0MEMOBJ
2325	&& RTR0MemObjAddressR3(pBundle->aMem[i].MapObjR3) == uPtr))
2326	)
2327	{
2328	/* Make a copy of it and release it outside the spinlock. */
2329	SUPDRVMEMREF Mem = pBundle->aMem[i];
2330	pBundle->aMem[i].eType = MEMREF_TYPE_UNUSED;
2331	pBundle->aMem[i].MemObj = NIL_RTR0MEMOBJ;
2332	pBundle->aMem[i].MapObjR3 = NIL_RTR0MEMOBJ;
2333	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2334
2335	if (Mem.MapObjR3)
2336	{
2337	int rc = RTR0MemObjFree(Mem.MapObjR3, false);
2338	AssertRC(rc); /** @todo figure out how to handle this. */
2339	}
2340	if (Mem.MemObj)
2341	{
2342	int rc = RTR0MemObjFree(Mem.MemObj, false);
2343	AssertRC(rc); /** @todo figure out how to handle this. */
2344	}
2345	return VINF_SUCCESS;
2346	}
2347	}
2348	}
2349	}
2350	RTSpinlockRelease(pSession->Spinlock, &SpinlockTmp);
2351	Log(("Failed to find %p!!! (eType=%d)\n", (void *)uPtr, eType));
2352	return VERR_INVALID_PARAMETER;
2353	}
2354
2355
2356	#ifdef VBOX_WITH_IDT_PATCHING
2357	/**
2358	* Install IDT for the current CPU.
2359	*
2360	* @returns One of the following IPRT status codes:
2361	* @retval VINF_SUCCESS on success.
2362	* @retval VERR_IDT_FAILED.
2363	* @retval VERR_NO_MEMORY.
2364	* @param pDevExt The device extension.
2365	* @param pSession The session data.
2366	* @param pReq The request.
2367	*/
2368	static int supdrvIOCtl_IdtInstall(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPIDTINSTALL pReq)
2369	{
2370	PSUPDRVPATCHUSAGE pUsagePre;
2371	PSUPDRVPATCH pPatchPre;
2372	RTIDTR Idtr;
2373	PSUPDRVPATCH pPatch;
2374	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2375	LogFlow(("supdrvIOCtl_IdtInstall\n"));
2376
2377	/*
2378	* Preallocate entry for this CPU cause we don't wanna do
2379	* that inside the spinlock!
2380	*/
2381	pUsagePre = (PSUPDRVPATCHUSAGE)RTMemAlloc(sizeof(*pUsagePre));
2382	if (!pUsagePre)
2383	return VERR_NO_MEMORY;
2384
2385	/*
2386	* Take the spinlock and see what we need to do.
2387	*/
2388	RTSpinlockAcquireNoInts(pDevExt->Spinlock, &SpinlockTmp);
2389
2390	/* check if we already got a free patch. */
2391	if (!pDevExt->pIdtPatchesFree)
2392	{
2393	/*
2394	* Allocate a patch - outside the spinlock of course.
2395	*/
2396	RTSpinlockReleaseNoInts(pDevExt->Spinlock, &SpinlockTmp);
2397
2398	pPatchPre = (PSUPDRVPATCH)RTMemExecAlloc(sizeof(*pPatchPre));
2399	if (!pPatchPre)
2400	return VERR_NO_MEMORY;
2401
2402	RTSpinlockAcquireNoInts(pDevExt->Spinlock, &SpinlockTmp);
2403	}
2404	else
2405	{
2406	pPatchPre = pDevExt->pIdtPatchesFree;
2407	pDevExt->pIdtPatchesFree = pPatchPre->pNext;
2408	}
2409
2410	/* look for matching patch entry */
2411	ASMGetIDTR(&Idtr);
2412	pPatch = pDevExt->pIdtPatches;
2413	while (pPatch && pPatch->pvIdt != (void *)Idtr.pIdt)
2414	pPatch = pPatch->pNext;
2415
2416	if (!pPatch)
2417	{
2418	/*
2419	* Create patch.
2420	*/
2421	pPatch = supdrvIdtPatchOne(pDevExt, pPatchPre);
2422	if (pPatch)
2423	pPatchPre = NULL; /* mark as used. */
2424	}
2425	else
2426	{
2427	/*
2428	* Simply increment patch usage.
2429	*/
2430	pPatch->cUsage++;
2431	}
2432
2433	if (pPatch)
2434	{
2435	/*
2436	* Increment and add if need be the session usage record for this patch.
2437	*/
2438	PSUPDRVPATCHUSAGE pUsage = pSession->pPatchUsage;
2439	while (pUsage && pUsage->pPatch != pPatch)
2440	pUsage = pUsage->pNext;
2441
2442	if (!pUsage)
2443	{
2444	/*
2445	* Add usage record.
2446	*/
2447	pUsagePre->cUsage = 1;
2448	pUsagePre->pPatch = pPatch;
2449	pUsagePre->pNext = pSession->pPatchUsage;
2450	pSession->pPatchUsage = pUsagePre;
2451	pUsagePre = NULL; /* mark as used. */
2452	}
2453	else
2454	{
2455	/*
2456	* Increment usage count.
2457	*/
2458	pUsage->cUsage++;
2459	}
2460	}
2461
2462	/* free patch - we accumulate them for paranoid saftly reasons. */
2463	if (pPatchPre)
2464	{
2465	pPatchPre->pNext = pDevExt->pIdtPatchesFree;
2466	pDevExt->pIdtPatchesFree = pPatchPre;
2467	}
2468
2469	RTSpinlockReleaseNoInts(pDevExt->Spinlock, &SpinlockTmp);
2470
2471	/*
2472	* Free unused preallocated buffers.
2473	*/
2474	if (pUsagePre)
2475	RTMemFree(pUsagePre);
2476
2477	pReq->u.Out.u8Idt = pDevExt->u8Idt;
2478
2479	return pPatch ? VINF_SUCCESS : VERR_IDT_FAILED;
2480	}
2481
2482
2483	/**
2484	* This creates a IDT patch entry.
2485	* If the first patch being installed it'll also determin the IDT entry
2486	* to use.
2487	*
2488	* @returns pPatch on success.
2489	* @returns NULL on failure.
2490	* @param pDevExt Pointer to globals.
2491	* @param pPatch Patch entry to use.
2492	* This will be linked into SUPDRVDEVEXT::pIdtPatches on
2493	* successful return.
2494	* @remark Call must be owning the SUPDRVDEVEXT::Spinlock!
2495	*/
2496	static PSUPDRVPATCH supdrvIdtPatchOne(PSUPDRVDEVEXT pDevExt, PSUPDRVPATCH pPatch)
2497	{
2498	RTIDTR Idtr;
2499	PSUPDRVIDTE paIdt;
2500	LogFlow(("supdrvIOCtl_IdtPatchOne: pPatch=%p\n", pPatch));
2501
2502	/*
2503	* Get IDT.
2504	*/
2505	ASMGetIDTR(&Idtr);
2506	paIdt = (PSUPDRVIDTE)Idtr.pIdt;
2507	/*
2508	* Recent Linux kernels can be configured to 1G user /3G kernel.
2509	*/
2510	if ((uintptr_t)paIdt < 0x40000000)
2511	{
2512	AssertMsgFailed(("bad paIdt=%p\n", paIdt));
2513	return NULL;
2514	}
2515
2516	if (!pDevExt->u8Idt)
2517	{
2518	/*
2519	* Test out the alternatives.
2520	*
2521	* At the moment we do not support chaining thus we ASSUME that one of
2522	* these 48 entries is unused (which is not a problem on Win32 and
2523	* Linux to my knowledge).
2524	*/
2525	/** @todo we MUST change this detection to try grab an entry which is NOT in use. This can be
2526	* combined with gathering info about which guest system call gates we can hook up directly. */
2527	unsigned i;
2528	uint8_t u8Idt = 0;
2529	static uint8_t au8Ints[] =
2530	{
2531	#ifdef RT_OS_WINDOWS /* We don't use 0xef and above because they are system stuff on linux (ef is IPI,
2532	* local apic timer, or some other frequently fireing thing). */
2533	0xef, 0xee, 0xed, 0xec,
2534	#endif
2535	0xeb, 0xea, 0xe9, 0xe8,
2536	0xdf, 0xde, 0xdd, 0xdc,
2537	0x7b, 0x7a, 0x79, 0x78,
2538	0xbf, 0xbe, 0xbd, 0xbc,
2539	};
2540	#if defined(RT_ARCH_AMD64) && defined(DEBUG)
2541	static int s_iWobble = 0;
2542	unsigned iMax = !(s_iWobble++ % 2) ? 0x80 : 0x100;
2543	Log2(("IDT: Idtr=%p:%#x\n", (void *)Idtr.pIdt, (unsigned)Idtr.cbIdt));
2544	for (i = iMax - 0x80; i*16+15 < Idtr.cbIdt && i < iMax; i++)
2545	{
2546	Log2(("%#x: %04x:%08x%04x%04x P=%d DPL=%d IST=%d Type1=%#x u32Reserved=%#x u5Reserved=%#x\n",
2547	i, paIdt[i].u16SegSel, paIdt[i].u32OffsetTop, paIdt[i].u16OffsetHigh, paIdt[i].u16OffsetLow,
2548	paIdt[i].u1Present, paIdt[i].u2DPL, paIdt[i].u3IST, paIdt[i].u5Type2,
2549	paIdt[i].u32Reserved, paIdt[i].u5Reserved));
2550	}
2551	#endif
2552	/* look for entries which are not present or otherwise unused. */
2553	for (i = 0; i < sizeof(au8Ints) / sizeof(au8Ints[0]); i++)
2554	{
2555	u8Idt = au8Ints[i];
2556	if ( u8Idt * sizeof(SUPDRVIDTE) < Idtr.cbIdt
2557	&& ( !paIdt[u8Idt].u1Present
2558	\|\| paIdt[u8Idt].u5Type2 == 0))
2559	break;
2560	u8Idt = 0;
2561	}
2562	if (!u8Idt)
2563	{
2564	/* try again, look for a compatible entry .*/
2565	for (i = 0; i < sizeof(au8Ints) / sizeof(au8Ints[0]); i++)
2566	{
2567	u8Idt = au8Ints[i];
2568	if ( u8Idt * sizeof(SUPDRVIDTE) < Idtr.cbIdt
2569	&& paIdt[u8Idt].u1Present
2570	&& paIdt[u8Idt].u5Type2 == SUPDRV_IDTE_TYPE2_INTERRUPT_GATE
2571	&& !(paIdt[u8Idt].u16SegSel & 3))
2572	break;
2573	u8Idt = 0;
2574	}
2575	if (!u8Idt)
2576	{
2577	Log(("Failed to find appropirate IDT entry!!\n"));
2578	return NULL;
2579	}
2580	}
2581	pDevExt->u8Idt = u8Idt;
2582	LogFlow(("supdrvIOCtl_IdtPatchOne: u8Idt=%x\n", u8Idt));
2583	}
2584
2585	/*
2586	* Prepare the patch
2587	*/
2588	memset(pPatch, 0, sizeof(*pPatch));
2589	pPatch->pvIdt = paIdt;
2590	pPatch->cUsage = 1;
2591	pPatch->pIdtEntry = &paIdt[pDevExt->u8Idt];
2592	pPatch->SavedIdt = paIdt[pDevExt->u8Idt];
2593	pPatch->ChangedIdt.u16OffsetLow = (uint32_t)((uintptr_t)&pPatch->auCode[0] & 0xffff);
2594	pPatch->ChangedIdt.u16OffsetHigh = (uint32_t)((uintptr_t)&pPatch->auCode[0] >> 16);
2595	#ifdef RT_ARCH_AMD64
2596	pPatch->ChangedIdt.u32OffsetTop = (uint32_t)((uintptr_t)&pPatch->auCode[0] >> 32);
2597	#endif
2598	pPatch->ChangedIdt.u16SegSel = ASMGetCS();
2599	#ifdef RT_ARCH_AMD64
2600	pPatch->ChangedIdt.u3IST = 0;
2601	pPatch->ChangedIdt.u5Reserved = 0;
2602	#else /* x86 */
2603	pPatch->ChangedIdt.u5Reserved = 0;
2604	pPatch->ChangedIdt.u3Type1 = 0;
2605	#endif /* x86 */
2606	pPatch->ChangedIdt.u5Type2 = SUPDRV_IDTE_TYPE2_INTERRUPT_GATE;
2607	pPatch->ChangedIdt.u2DPL = 3;
2608	pPatch->ChangedIdt.u1Present = 1;
2609
2610	/*
2611	* Generate the patch code.
2612	*/
2613	{
2614	#ifdef RT_ARCH_AMD64
2615	union
2616	{
2617	uint8_t *pb;
2618	uint32_t *pu32;
2619	uint64_t *pu64;
2620	} u, uFixJmp, uFixCall, uNotNested;
2621	u.pb = &pPatch->auCode[0];
2622
2623	/* check the cookie */
2624	*u.pb++ = 0x3d; // cmp eax, GLOBALCOOKIE
2625	*u.pu32++ = pDevExt->u32Cookie;
2626
2627	*u.pb++ = 0x74; // jz @VBoxCall
2628	*u.pb++ = 2;
2629
2630	/* jump to forwarder code. */
2631	*u.pb++ = 0xeb;
2632	uFixJmp = u;
2633	*u.pb++ = 0xfe;
2634
2635	// @VBoxCall:
2636	*u.pb++ = 0x0f; // swapgs
2637	*u.pb++ = 0x01;
2638	*u.pb++ = 0xf8;
2639
2640	/*
2641	* Call VMMR0Entry
2642	* We don't have to push the arguments here, but we have top
2643	* reserve some stack space for the interrupt forwarding.
2644	*/
2645	# ifdef RT_OS_WINDOWS
2646	*u.pb++ = 0x50; // push rax ; alignment filler.
2647	*u.pb++ = 0x41; // push r8 ; uArg
2648	*u.pb++ = 0x50;
2649	*u.pb++ = 0x52; // push rdx ; uOperation
2650	*u.pb++ = 0x51; // push rcx ; pVM
2651	# else
2652	*u.pb++ = 0x51; // push rcx ; alignment filler.
2653	*u.pb++ = 0x52; // push rdx ; uArg
2654	*u.pb++ = 0x56; // push rsi ; uOperation
2655	*u.pb++ = 0x57; // push rdi ; pVM
2656	# endif
2657
2658	*u.pb++ = 0xff; // call qword [pfnVMMR0EntryInt wrt rip]
2659	*u.pb++ = 0x15;
2660	uFixCall = u;
2661	*u.pu32++ = 0;
2662
2663	*u.pb++ = 0x48; // add rsp, 20h ; remove call frame.
2664	*u.pb++ = 0x81;
2665	*u.pb++ = 0xc4;
2666	*u.pu32++ = 0x20;
2667
2668	*u.pb++ = 0x0f; // swapgs
2669	*u.pb++ = 0x01;
2670	*u.pb++ = 0xf8;
2671
2672	/* Return to R3. */
2673	uNotNested = u;
2674	*u.pb++ = 0x48; // iretq
2675	*u.pb++ = 0xcf;
2676
2677	while ((uintptr_t)u.pb & 0x7) // align 8
2678	*u.pb++ = 0xcc;
2679
2680	/* Pointer to the VMMR0Entry. */ // pfnVMMR0EntryInt dq StubVMMR0Entry
2681	*uFixCall.pu32 = (uint32_t)(u.pb - uFixCall.pb - 4); uFixCall.pb = NULL;
2682	pPatch->offVMMR0EntryFixup = (uint16_t)(u.pb - &pPatch->auCode[0]);
2683	*u.pu64++ = pDevExt->pvVMMR0 ? (uint64_t)pDevExt->pfnVMMR0EntryInt : (uint64_t)u.pb + 8;
2684
2685	/* stub entry. */ // StubVMMR0Entry:
2686	pPatch->offStub = (uint16_t)(u.pb - &pPatch->auCode[0]);
2687	*u.pb++ = 0x33; // xor eax, eax
2688	*u.pb++ = 0xc0;
2689
2690	*u.pb++ = 0x48; // dec rax
2691	*u.pb++ = 0xff;
2692	*u.pb++ = 0xc8;
2693
2694	*u.pb++ = 0xc3; // ret
2695
2696	/* forward to the original handler using a retf. */
2697	*uFixJmp.pb = (uint8_t)(u.pb - uFixJmp.pb - 1); uFixJmp.pb = NULL;
2698
2699	*u.pb++ = 0x68; // push <target cs>
2700	*u.pu32++ = !pPatch->SavedIdt.u5Type2 ? ASMGetCS() : pPatch->SavedIdt.u16SegSel;
2701
2702	*u.pb++ = 0x68; // push <low target rip>
2703	*u.pu32++ = !pPatch->SavedIdt.u5Type2
2704	? (uint32_t)(uintptr_t)uNotNested.pb
2705	: (uint32_t)pPatch->SavedIdt.u16OffsetLow
2706	\| (uint32_t)pPatch->SavedIdt.u16OffsetHigh << 16;
2707
2708	*u.pb++ = 0xc7; // mov dword [rsp + 4], <high target rip>
2709	*u.pb++ = 0x44;
2710	*u.pb++ = 0x24;
2711	*u.pb++ = 0x04;
2712	*u.pu32++ = !pPatch->SavedIdt.u5Type2
2713	? (uint32_t)((uint64_t)uNotNested.pb >> 32)
2714	: pPatch->SavedIdt.u32OffsetTop;
2715
2716	*u.pb++ = 0x48; // retf ; does this require prefix?
2717	*u.pb++ = 0xcb;
2718
2719	#else /* RT_ARCH_X86 */
2720
2721	union
2722	{
2723	uint8_t *pb;
2724	uint16_t *pu16;
2725	uint32_t *pu32;
2726	} u, uFixJmpNotNested, uFixJmp, uFixCall, uNotNested;
2727	u.pb = &pPatch->auCode[0];
2728
2729	/* check the cookie */
2730	*u.pb++ = 0x81; // cmp esi, GLOBALCOOKIE
2731	*u.pb++ = 0xfe;
2732	*u.pu32++ = pDevExt->u32Cookie;
2733
2734	*u.pb++ = 0x74; // jz VBoxCall
2735	uFixJmp = u;
2736	*u.pb++ = 0;
2737
2738	/* jump (far) to the original handler / not-nested-stub. */
2739	*u.pb++ = 0xea; // jmp far NotNested
2740	uFixJmpNotNested = u;
2741	*u.pu32++ = 0;
2742	*u.pu16++ = 0;
2743
2744	/* save selector registers. */ // VBoxCall:
2745	*uFixJmp.pb = (uint8_t)(u.pb - uFixJmp.pb - 1);
2746	*u.pb++ = 0x0f; // push fs
2747	*u.pb++ = 0xa0;
2748
2749	*u.pb++ = 0x1e; // push ds
2750
2751	*u.pb++ = 0x06; // push es
2752
2753	/* call frame */
2754	*u.pb++ = 0x51; // push ecx
2755
2756	*u.pb++ = 0x52; // push edx
2757
2758	*u.pb++ = 0x50; // push eax
2759
2760	/* load ds, es and perhaps fs before call. */
2761	*u.pb++ = 0xb8; // mov eax, KernelDS
2762	*u.pu32++ = ASMGetDS();
2763
2764	*u.pb++ = 0x8e; // mov ds, eax
2765	*u.pb++ = 0xd8;
2766
2767	*u.pb++ = 0x8e; // mov es, eax
2768	*u.pb++ = 0xc0;
2769
2770	#ifdef RT_OS_WINDOWS
2771	*u.pb++ = 0xb8; // mov eax, KernelFS
2772	*u.pu32++ = ASMGetFS();
2773
2774	*u.pb++ = 0x8e; // mov fs, eax
2775	*u.pb++ = 0xe0;
2776	#endif
2777
2778	/* do the call. */
2779	*u.pb++ = 0xe8; // call _VMMR0Entry / StubVMMR0Entry
2780	uFixCall = u;
2781	pPatch->offVMMR0EntryFixup = (uint16_t)(u.pb - &pPatch->auCode[0]);
2782	*u.pu32++ = 0xfffffffb;
2783
2784	*u.pb++ = 0x83; // add esp, 0ch ; cdecl
2785	*u.pb++ = 0xc4;
2786	*u.pb++ = 0x0c;
2787
2788	/* restore selector registers. */
2789	*u.pb++ = 0x07; // pop es
2790	//
2791	*u.pb++ = 0x1f; // pop ds
2792
2793	*u.pb++ = 0x0f; // pop fs
2794	*u.pb++ = 0xa1;
2795
2796	uNotNested = u; // NotNested:
2797	*u.pb++ = 0xcf; // iretd
2798
2799	/* the stub VMMR0Entry. */ // StubVMMR0Entry:
2800	pPatch->offStub = (uint16_t)(u.pb - &pPatch->auCode[0]);
2801	*u.pb++ = 0x33; // xor eax, eax
2802	*u.pb++ = 0xc0;
2803
2804	*u.pb++ = 0x48; // dec eax
2805
2806	*u.pb++ = 0xc3; // ret
2807
2808	/* Fixup the VMMR0Entry call. */
2809	if (pDevExt->pvVMMR0)
2810	*uFixCall.pu32 = (uint32_t)pDevExt->pfnVMMR0EntryInt - (uint32_t)(uFixCall.pu32 + 1);
2811	else
2812	*uFixCall.pu32 = (uint32_t)&pPatch->auCode[pPatch->offStub] - (uint32_t)(uFixCall.pu32 + 1);
2813
2814	/* Fixup the forward / nested far jump. */
2815	if (!pPatch->SavedIdt.u5Type2)
2816	{
2817	*uFixJmpNotNested.pu32++ = (uint32_t)uNotNested.pb;
2818	*uFixJmpNotNested.pu16++ = ASMGetCS();
2819	}
2820	else
2821	{
2822	*uFixJmpNotNested.pu32++ = ((uint32_t)pPatch->SavedIdt.u16OffsetHigh << 16) \| pPatch->SavedIdt.u16OffsetLow;
2823	*uFixJmpNotNested.pu16++ = pPatch->SavedIdt.u16SegSel;
2824	}
2825	#endif /* RT_ARCH_X86 */
2826	Assert(u.pb <= &pPatch->auCode[sizeof(pPatch->auCode)]);
2827	#if 0
2828	/* dump the patch code */
2829	Log2(("patch code: %p\n", &pPatch->auCode[0]));
2830	for (uFixCall.pb = &pPatch->auCode[0]; uFixCall.pb < u.pb; uFixCall.pb++)
2831	Log2(("0x%02x,\n", *uFixCall.pb));
2832	#endif
2833	}
2834
2835	/*
2836	* Install the patch.
2837	*/
2838	supdrvIdtWrite(pPatch->pIdtEntry, &pPatch->ChangedIdt);
2839	AssertMsg(!memcmp((void *)pPatch->pIdtEntry, &pPatch->ChangedIdt, sizeof(pPatch->ChangedIdt)), ("The stupid change code didn't work!!!!!\n"));
2840
2841	/*
2842	* Link in the patch.
2843	*/
2844	pPatch->pNext = pDevExt->pIdtPatches;
2845	pDevExt->pIdtPatches = pPatch;
2846
2847	return pPatch;
2848	}
2849
2850
2851	/**
2852	* Removes the sessions IDT references.
2853	* This will uninstall our IDT patch if we left unreferenced.
2854	*
2855	* @returns VINF_SUCCESS.
2856	* @param pDevExt Device globals.
2857	* @param pSession Session data.
2858	*/
2859	static int supdrvIOCtl_IdtRemoveAll(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession)
2860	{
2861	PSUPDRVPATCHUSAGE pUsage;
2862	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
2863	LogFlow(("supdrvIOCtl_IdtRemoveAll: pSession=%p\n", pSession));
2864
2865	/*
2866	* Take the spinlock.
2867	*/
2868	RTSpinlockAcquireNoInts(pDevExt->Spinlock, &SpinlockTmp);
2869
2870	/*
2871	* Walk usage list, removing patches as their usage count reaches zero.
2872	*/
2873	pUsage = pSession->pPatchUsage;
2874	while (pUsage)
2875	{
2876	if (pUsage->pPatch->cUsage <= pUsage->cUsage)
2877	supdrvIdtRemoveOne(pDevExt, pUsage->pPatch);
2878	else
2879	pUsage->pPatch->cUsage -= pUsage->cUsage;
2880
2881	/* next */
2882	pUsage = pUsage->pNext;
2883	}
2884
2885	/*
2886	* Empty the usage chain and we're done inside the spinlock.
2887	*/
2888	pUsage = pSession->pPatchUsage;
2889	pSession->pPatchUsage = NULL;
2890
2891	RTSpinlockReleaseNoInts(pDevExt->Spinlock, &SpinlockTmp);
2892
2893	/*
2894	* Free usage entries.
2895	*/
2896	while (pUsage)
2897	{
2898	void *pvToFree = pUsage;
2899	pUsage->cUsage = 0;
2900	pUsage->pPatch = NULL;
2901	pUsage = pUsage->pNext;
2902	RTMemFree(pvToFree);
2903	}
2904
2905	return VINF_SUCCESS;
2906	}
2907
2908
2909	/**
2910	* Remove one patch.
2911	*
2912	* Worker for supdrvIOCtl_IdtRemoveAll.
2913	*
2914	* @param pDevExt Device globals.
2915	* @param pPatch Patch entry to remove.
2916	* @remark Caller must own SUPDRVDEVEXT::Spinlock!
2917	*/
2918	static void supdrvIdtRemoveOne(PSUPDRVDEVEXT pDevExt, PSUPDRVPATCH pPatch)
2919	{
2920	LogFlow(("supdrvIdtRemoveOne: pPatch=%p\n", pPatch));
2921
2922	pPatch->cUsage = 0;
2923
2924	/*
2925	* If the IDT entry was changed it have to kick around for ever!
2926	* This will be attempted freed again, perhaps next time we'll succeed :-)
2927	*/
2928	if (memcmp((void *)pPatch->pIdtEntry, &pPatch->ChangedIdt, sizeof(pPatch->ChangedIdt)))
2929	{
2930	AssertMsgFailed(("The hijacked IDT entry has CHANGED!!!\n"));
2931	return;
2932	}
2933
2934	/*
2935	* Unlink it.
2936	*/
2937	if (pDevExt->pIdtPatches != pPatch)
2938	{
2939	PSUPDRVPATCH pPatchPrev = pDevExt->pIdtPatches;
2940	while (pPatchPrev)
2941	{
2942	if (pPatchPrev->pNext == pPatch)
2943	{
2944	pPatchPrev->pNext = pPatch->pNext;
2945	break;
2946	}
2947	pPatchPrev = pPatchPrev->pNext;
2948	}
2949	Assert(!pPatchPrev);
2950	}
2951	else
2952	pDevExt->pIdtPatches = pPatch->pNext;
2953	pPatch->pNext = NULL;
2954
2955
2956	/*
2957	* Verify and restore the IDT.
2958	*/
2959	AssertMsg(!memcmp((void *)pPatch->pIdtEntry, &pPatch->ChangedIdt, sizeof(pPatch->ChangedIdt)), ("The hijacked IDT entry has CHANGED!!!\n"));
2960	supdrvIdtWrite(pPatch->pIdtEntry, &pPatch->SavedIdt);
2961	AssertMsg(!memcmp((void *)pPatch->pIdtEntry, &pPatch->SavedIdt, sizeof(pPatch->SavedIdt)), ("The hijacked IDT entry has CHANGED!!!\n"));
2962
2963	/*
2964	* Put it in the free list.
2965	* (This free list stuff is to calm my paranoia.)
2966	*/
2967	pPatch->pvIdt = NULL;
2968	pPatch->pIdtEntry = NULL;
2969
2970	pPatch->pNext = pDevExt->pIdtPatchesFree;
2971	pDevExt->pIdtPatchesFree = pPatch;
2972	}
2973
2974
2975	/**
2976	* Write to an IDT entry.
2977	*
2978	* @param pvIdtEntry Where to write.
2979	* @param pNewIDTEntry What to write.
2980	*/
2981	static void supdrvIdtWrite(volatile void pvIdtEntry, const SUPDRVIDTE pNewIDTEntry)
2982	{
2983	RTUINTREG uCR0;
2984	RTUINTREG uFlags;
2985
2986	/*
2987	* On SMP machines (P4 hyperthreading included) we must preform a
2988	* 64-bit locked write when updating the IDT entry.
2989	*
2990	* The F00F bugfix for linux (and probably other OSes) causes
2991	* the IDT to be pointing to an readonly mapping. We get around that
2992	* by temporarily turning of WP. Since we're inside a spinlock at this
2993	* point, interrupts are disabled and there isn't any way the WP bit
2994	* flipping can cause any trouble.
2995	*/
2996
2997	/* Save & Clear interrupt flag; Save & clear WP. */
2998	uFlags = ASMGetFlags();
2999	ASMSetFlags(uFlags & ~(RTUINTREG)(1 << 9)); /X86_EFL_IF/
3000	Assert(!(ASMGetFlags() & (1 << 9)));
3001	uCR0 = ASMGetCR0();
3002	ASMSetCR0(uCR0 & ~(RTUINTREG)(1 << 16)); /X86_CR0_WP/
3003
3004	/* Update IDT Entry */
3005	#ifdef RT_ARCH_AMD64
3006	ASMAtomicXchgU128((volatile uint128_t )pvIdtEntry, (uint128_t *)(uintptr_t)pNewIDTEntry);
3007	#else
3008	ASMAtomicXchgU64((volatile uint64_t )pvIdtEntry, (uint64_t *)(uintptr_t)pNewIDTEntry);
3009	#endif
3010
3011	/* Restore CR0 & Flags */
3012	ASMSetCR0(uCR0);
3013	ASMSetFlags(uFlags);
3014	}
3015	#endif /* VBOX_WITH_IDT_PATCHING */
3016
3017
3018	/**
3019	* Opens an image. If it's the first time it's opened the call must upload
3020	* the bits using the supdrvIOCtl_LdrLoad() / SUPDRV_IOCTL_LDR_LOAD function.
3021	*
3022	* This is the 1st step of the loading.
3023	*
3024	* @returns IPRT status code.
3025	* @param pDevExt Device globals.
3026	* @param pSession Session data.
3027	* @param pReq The open request.
3028	*/
3029	static int supdrvIOCtl_LdrOpen(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDROPEN pReq)
3030	{
3031	PSUPDRVLDRIMAGE pImage;
3032	unsigned cb;
3033	void *pv;
3034	LogFlow(("supdrvIOCtl_LdrOpen: szName=%s cbImage=%d\n", pReq->u.In.szName, pReq->u.In.cbImage));
3035
3036	/*
3037	* Check if we got an instance of the image already.
3038	*/
3039	RTSemFastMutexRequest(pDevExt->mtxLdr);
3040	for (pImage = pDevExt->pLdrImages; pImage; pImage = pImage->pNext)
3041	{
3042	if (!strcmp(pImage->szName, pReq->u.In.szName))
3043	{
3044	pImage->cUsage++;
3045	pReq->u.Out.pvImageBase = pImage->pvImage;
3046	pReq->u.Out.fNeedsLoading = pImage->uState == SUP_IOCTL_LDR_OPEN;
3047	supdrvLdrAddUsage(pSession, pImage);
3048	RTSemFastMutexRelease(pDevExt->mtxLdr);
3049	return VINF_SUCCESS;
3050	}
3051	}
3052	/* (not found - add it!) */
3053
3054	/*
3055	* Allocate memory.
3056	*/
3057	cb = pReq->u.In.cbImage + sizeof(SUPDRVLDRIMAGE) + 31;
3058	pv = RTMemExecAlloc(cb);
3059	if (!pv)
3060	{
3061	RTSemFastMutexRelease(pDevExt->mtxLdr);
3062	Log(("supdrvIOCtl_LdrOpen: RTMemExecAlloc(%u) failed\n", cb));
3063	return VERR_NO_MEMORY;
3064	}
3065
3066	/*
3067	* Setup and link in the LDR stuff.
3068	*/
3069	pImage = (PSUPDRVLDRIMAGE)pv;
3070	pImage->pvImage = RT_ALIGN_P(pImage + 1, 32);
3071	pImage->cbImage = pReq->u.In.cbImage;
3072	pImage->pfnModuleInit = NULL;
3073	pImage->pfnModuleTerm = NULL;
3074	pImage->uState = SUP_IOCTL_LDR_OPEN;
3075	pImage->cUsage = 1;
3076	strcpy(pImage->szName, pReq->u.In.szName);
3077
3078	pImage->pNext = pDevExt->pLdrImages;
3079	pDevExt->pLdrImages = pImage;
3080
3081	supdrvLdrAddUsage(pSession, pImage);
3082
3083	pReq->u.Out.pvImageBase = pImage->pvImage;
3084	pReq->u.Out.fNeedsLoading = true;
3085	RTSemFastMutexRelease(pDevExt->mtxLdr);
3086	return VINF_SUCCESS;
3087	}
3088
3089
3090	/**
3091	* Loads the image bits.
3092	*
3093	* This is the 2nd step of the loading.
3094	*
3095	* @returns IPRT status code.
3096	* @param pDevExt Device globals.
3097	* @param pSession Session data.
3098	* @param pReq The request.
3099	*/
3100	static int supdrvIOCtl_LdrLoad(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRLOAD pReq)
3101	{
3102	PSUPDRVLDRUSAGE pUsage;
3103	PSUPDRVLDRIMAGE pImage;
3104	int rc;
3105	LogFlow(("supdrvIOCtl_LdrLoad: pvImageBase=%p cbImage=%d\n", pReq->u.In.pvImageBase, pReq->u.In.cbImage));
3106
3107	/*
3108	* Find the ldr image.
3109	*/
3110	RTSemFastMutexRequest(pDevExt->mtxLdr);
3111	pUsage = pSession->pLdrUsage;
3112	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3113	pUsage = pUsage->pNext;
3114	if (!pUsage)
3115	{
3116	RTSemFastMutexRelease(pDevExt->mtxLdr);
3117	Log(("SUP_IOCTL_LDR_LOAD: couldn't find image!\n"));
3118	return VERR_INVALID_HANDLE;
3119	}
3120	pImage = pUsage->pImage;
3121	if (pImage->cbImage != pReq->u.In.cbImage)
3122	{
3123	RTSemFastMutexRelease(pDevExt->mtxLdr);
3124	Log(("SUP_IOCTL_LDR_LOAD: image size mismatch!! %d(prep) != %d(load)\n", pImage->cbImage, pReq->u.In.cbImage));
3125	return VERR_INVALID_HANDLE;
3126	}
3127	if (pImage->uState != SUP_IOCTL_LDR_OPEN)
3128	{
3129	unsigned uState = pImage->uState;
3130	RTSemFastMutexRelease(pDevExt->mtxLdr);
3131	if (uState != SUP_IOCTL_LDR_LOAD)
3132	AssertMsgFailed(("SUP_IOCTL_LDR_LOAD: invalid image state %d (%#x)!\n", uState, uState));
3133	return SUPDRV_ERR_ALREADY_LOADED;
3134	}
3135	switch (pReq->u.In.eEPType)
3136	{
3137	case SUPLDRLOADEP_NOTHING:
3138	break;
3139	case SUPLDRLOADEP_VMMR0:
3140	if ( !pReq->u.In.EP.VMMR0.pvVMMR0
3141	\|\| !pReq->u.In.EP.VMMR0.pvVMMR0EntryInt
3142	\|\| !pReq->u.In.EP.VMMR0.pvVMMR0EntryFast
3143	\|\| !pReq->u.In.EP.VMMR0.pvVMMR0EntryEx)
3144	{
3145	RTSemFastMutexRelease(pDevExt->mtxLdr);
3146	Log(("NULL pointer: pvVMMR0=%p pvVMMR0EntryInt=%p pvVMMR0EntryFast=%p pvVMMR0EntryEx=%p!\n",
3147	pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
3148	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx));
3149	return VERR_INVALID_PARAMETER;
3150	}
3151	if ( (uintptr_t)pReq->u.In.EP.VMMR0.pvVMMR0EntryInt - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage
3152	\|\| (uintptr_t)pReq->u.In.EP.VMMR0.pvVMMR0EntryFast - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage
3153	\|\| (uintptr_t)pReq->u.In.EP.VMMR0.pvVMMR0EntryEx - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage)
3154	{
3155	RTSemFastMutexRelease(pDevExt->mtxLdr);
3156	Log(("Out of range (%p LB %#x): pvVMMR0EntryInt=%p, pvVMMR0EntryFast=%p or pvVMMR0EntryEx=%p is NULL!\n",
3157	pImage->pvImage, pReq->u.In.cbImage, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
3158	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx));
3159	return VERR_INVALID_PARAMETER;
3160	}
3161	break;
3162	default:
3163	RTSemFastMutexRelease(pDevExt->mtxLdr);
3164	Log(("Invalid eEPType=%d\n", pReq->u.In.eEPType));
3165	return VERR_INVALID_PARAMETER;
3166	}
3167	if ( pReq->u.In.pfnModuleInit
3168	&& (uintptr_t)pReq->u.In.pfnModuleInit - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage)
3169	{
3170	RTSemFastMutexRelease(pDevExt->mtxLdr);
3171	Log(("SUP_IOCTL_LDR_LOAD: pfnModuleInit=%p is outside the image (%p %d bytes)\n",
3172	pReq->u.In.pfnModuleInit, pImage->pvImage, pReq->u.In.cbImage));
3173	return VERR_INVALID_PARAMETER;
3174	}
3175	if ( pReq->u.In.pfnModuleTerm
3176	&& (uintptr_t)pReq->u.In.pfnModuleTerm - (uintptr_t)pImage->pvImage >= pReq->u.In.cbImage)
3177	{
3178	RTSemFastMutexRelease(pDevExt->mtxLdr);
3179	Log(("SUP_IOCTL_LDR_LOAD: pfnModuleTerm=%p is outside the image (%p %d bytes)\n",
3180	pReq->u.In.pfnModuleTerm, pImage->pvImage, pReq->u.In.cbImage));
3181	return VERR_INVALID_PARAMETER;
3182	}
3183
3184	/*
3185	* Copy the memory.
3186	*/
3187	/* no need to do try/except as this is a buffered request. */
3188	memcpy(pImage->pvImage, &pReq->u.In.achImage[0], pImage->cbImage);
3189	pImage->uState = SUP_IOCTL_LDR_LOAD;
3190	pImage->pfnModuleInit = pReq->u.In.pfnModuleInit;
3191	pImage->pfnModuleTerm = pReq->u.In.pfnModuleTerm;
3192	pImage->offSymbols = pReq->u.In.offSymbols;
3193	pImage->cSymbols = pReq->u.In.cSymbols;
3194	pImage->offStrTab = pReq->u.In.offStrTab;
3195	pImage->cbStrTab = pReq->u.In.cbStrTab;
3196
3197	/*
3198	* Update any entry points.
3199	*/
3200	switch (pReq->u.In.eEPType)
3201	{
3202	default:
3203	case SUPLDRLOADEP_NOTHING:
3204	rc = VINF_SUCCESS;
3205	break;
3206	case SUPLDRLOADEP_VMMR0:
3207	rc = supdrvLdrSetR0EP(pDevExt, pReq->u.In.EP.VMMR0.pvVMMR0, pReq->u.In.EP.VMMR0.pvVMMR0EntryInt,
3208	pReq->u.In.EP.VMMR0.pvVMMR0EntryFast, pReq->u.In.EP.VMMR0.pvVMMR0EntryEx);
3209	break;
3210	}
3211
3212	/*
3213	* On success call the module initialization.
3214	*/
3215	LogFlow(("supdrvIOCtl_LdrLoad: pfnModuleInit=%p\n", pImage->pfnModuleInit));
3216	if (RT_SUCCESS(rc) && pImage->pfnModuleInit)
3217	{
3218	Log(("supdrvIOCtl_LdrLoad: calling pfnModuleInit=%p\n", pImage->pfnModuleInit));
3219	rc = pImage->pfnModuleInit();
3220	if (rc && pDevExt->pvVMMR0 == pImage->pvImage)
3221	supdrvLdrUnsetR0EP(pDevExt);
3222	}
3223
3224	if (rc)
3225	pImage->uState = SUP_IOCTL_LDR_OPEN;
3226
3227	RTSemFastMutexRelease(pDevExt->mtxLdr);
3228	return rc;
3229	}
3230
3231
3232	/**
3233	* Frees a previously loaded (prep'ed) image.
3234	*
3235	* @returns IPRT status code.
3236	* @param pDevExt Device globals.
3237	* @param pSession Session data.
3238	* @param pReq The request.
3239	*/
3240	static int supdrvIOCtl_LdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRFREE pReq)
3241	{
3242	int rc;
3243	PSUPDRVLDRUSAGE pUsagePrev;
3244	PSUPDRVLDRUSAGE pUsage;
3245	PSUPDRVLDRIMAGE pImage;
3246	LogFlow(("supdrvIOCtl_LdrFree: pvImageBase=%p\n", pReq->u.In.pvImageBase));
3247
3248	/*
3249	* Find the ldr image.
3250	*/
3251	RTSemFastMutexRequest(pDevExt->mtxLdr);
3252	pUsagePrev = NULL;
3253	pUsage = pSession->pLdrUsage;
3254	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3255	{
3256	pUsagePrev = pUsage;
3257	pUsage = pUsage->pNext;
3258	}
3259	if (!pUsage)
3260	{
3261	RTSemFastMutexRelease(pDevExt->mtxLdr);
3262	Log(("SUP_IOCTL_LDR_FREE: couldn't find image!\n"));
3263	return VERR_INVALID_HANDLE;
3264	}
3265
3266	/*
3267	* Check if we can remove anything.
3268	*/
3269	rc = VINF_SUCCESS;
3270	pImage = pUsage->pImage;
3271	if (pImage->cUsage <= 1 \|\| pUsage->cUsage <= 1)
3272	{
3273	/*
3274	* Check if there are any objects with destructors in the image, if
3275	* so leave it for the session cleanup routine so we get a chance to
3276	* clean things up in the right order and not leave them all dangling.
3277	*/
3278	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3279	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
3280	if (pImage->cUsage <= 1)
3281	{
3282	PSUPDRVOBJ pObj;
3283	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
3284	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImage))
3285	{
3286	rc = VERR_SHARING_VIOLATION; /** @todo VERR_DANGLING_OBJECTS */
3287	break;
3288	}
3289	}
3290	else
3291	{
3292	PSUPDRVUSAGE pGenUsage;
3293	for (pGenUsage = pSession->pUsage; pGenUsage; pGenUsage = pGenUsage->pNext)
3294	if (RT_UNLIKELY((uintptr_t)pGenUsage->pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImage))
3295	{
3296	rc = VERR_SHARING_VIOLATION; /** @todo VERR_DANGLING_OBJECTS */
3297	break;
3298	}
3299	}
3300	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
3301	if (rc == VINF_SUCCESS)
3302	{
3303	/* unlink it */
3304	if (pUsagePrev)
3305	pUsagePrev->pNext = pUsage->pNext;
3306	else
3307	pSession->pLdrUsage = pUsage->pNext;
3308
3309	/* free it */
3310	pUsage->pImage = NULL;
3311	pUsage->pNext = NULL;
3312	RTMemFree(pUsage);
3313
3314	/*
3315	* Derefrence the image.
3316	*/
3317	if (pImage->cUsage <= 1)
3318	supdrvLdrFree(pDevExt, pImage);
3319	else
3320	pImage->cUsage--;
3321	}
3322	else
3323	Log(("supdrvIOCtl_LdrFree: Dangling objects in %p/%s!\n", pImage->pvImage, pImage->szName));
3324	}
3325	else
3326	{
3327	/*
3328	* Dereference both image and usage.
3329	*/
3330	pImage->cUsage--;
3331	pUsage->cUsage--;
3332	}
3333
3334	RTSemFastMutexRelease(pDevExt->mtxLdr);
3335	return VINF_SUCCESS;
3336	}
3337
3338
3339	/**
3340	* Gets the address of a symbol in an open image.
3341	*
3342	* @returns 0 on success.
3343	* @returns SUPDRV_ERR_* on failure.
3344	* @param pDevExt Device globals.
3345	* @param pSession Session data.
3346	* @param pReq The request buffer.
3347	*/
3348	static int supdrvIOCtl_LdrGetSymbol(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPLDRGETSYMBOL pReq)
3349	{
3350	PSUPDRVLDRIMAGE pImage;
3351	PSUPDRVLDRUSAGE pUsage;
3352	uint32_t i;
3353	PSUPLDRSYM paSyms;
3354	const char *pchStrings;
3355	const size_t cbSymbol = strlen(pReq->u.In.szSymbol) + 1;
3356	void *pvSymbol = NULL;
3357	int rc = VERR_GENERAL_FAILURE;
3358	Log3(("supdrvIOCtl_LdrGetSymbol: pvImageBase=%p szSymbol=\"%s\"\n", pReq->u.In.pvImageBase, pReq->u.In.szSymbol));
3359
3360	/*
3361	* Find the ldr image.
3362	*/
3363	RTSemFastMutexRequest(pDevExt->mtxLdr);
3364	pUsage = pSession->pLdrUsage;
3365	while (pUsage && pUsage->pImage->pvImage != pReq->u.In.pvImageBase)
3366	pUsage = pUsage->pNext;
3367	if (!pUsage)
3368	{
3369	RTSemFastMutexRelease(pDevExt->mtxLdr);
3370	Log(("SUP_IOCTL_LDR_GET_SYMBOL: couldn't find image!\n"));
3371	return VERR_INVALID_HANDLE;
3372	}
3373	pImage = pUsage->pImage;
3374	if (pImage->uState != SUP_IOCTL_LDR_LOAD)
3375	{
3376	unsigned uState = pImage->uState;
3377	RTSemFastMutexRelease(pDevExt->mtxLdr);
3378	Log(("SUP_IOCTL_LDR_GET_SYMBOL: invalid image state %d (%#x)!\n", uState, uState)); NOREF(uState);
3379	return VERR_ALREADY_LOADED;
3380	}
3381
3382	/*
3383	* Search the symbol string.
3384	*/
3385	pchStrings = (const char )((uint8_t )pImage->pvImage + pImage->offStrTab);
3386	paSyms = (PSUPLDRSYM)((uint8_t *)pImage->pvImage + pImage->offSymbols);
3387	for (i = 0; i < pImage->cSymbols; i++)
3388	{
3389	if ( paSyms[i].offSymbol < pImage->cbImage /* paranoia */
3390	&& paSyms[i].offName + cbSymbol <= pImage->cbStrTab
3391	&& !memcmp(pchStrings + paSyms[i].offName, pReq->u.In.szSymbol, cbSymbol))
3392	{
3393	pvSymbol = (uint8_t *)pImage->pvImage + paSyms[i].offSymbol;
3394	rc = VINF_SUCCESS;
3395	break;
3396	}
3397	}
3398	RTSemFastMutexRelease(pDevExt->mtxLdr);
3399	pReq->u.Out.pvSymbol = pvSymbol;
3400	return rc;
3401	}
3402
3403
3404	/**
3405	* Updates the IDT patches to point to the specified VMM R0 entry
3406	* point (i.e. VMMR0Enter()).
3407	*
3408	* @returns IPRT status code.
3409	* @param pDevExt Device globals.
3410	* @param pSession Session data.
3411	* @param pVMMR0 VMMR0 image handle.
3412	* @param pvVMMR0EntryInt VMMR0EntryInt address.
3413	* @param pvVMMR0EntryFast VMMR0EntryFast address.
3414	* @param pvVMMR0EntryEx VMMR0EntryEx address.
3415	* @remark Caller must own the loader mutex.
3416	*/
3417	static int supdrvLdrSetR0EP(PSUPDRVDEVEXT pDevExt, void pvVMMR0, void pvVMMR0EntryInt, void pvVMMR0EntryFast, void pvVMMR0EntryEx)
3418	{
3419	int rc = VINF_SUCCESS;
3420	LogFlow(("supdrvLdrSetR0EP pvVMMR0=%p pvVMMR0EntryInt=%p\n", pvVMMR0, pvVMMR0EntryInt));
3421
3422
3423	/*
3424	* Check if not yet set.
3425	*/
3426	if (!pDevExt->pvVMMR0)
3427	{
3428	#ifdef VBOX_WITH_IDT_PATCHING
3429	PSUPDRVPATCH pPatch;
3430	#endif
3431
3432	/*
3433	* Set it and update IDT patch code.
3434	*/
3435	pDevExt->pvVMMR0 = pvVMMR0;
3436	pDevExt->pfnVMMR0EntryInt = pvVMMR0EntryInt;
3437	pDevExt->pfnVMMR0EntryFast = pvVMMR0EntryFast;
3438	pDevExt->pfnVMMR0EntryEx = pvVMMR0EntryEx;
3439	#ifdef VBOX_WITH_IDT_PATCHING
3440	for (pPatch = pDevExt->pIdtPatches; pPatch; pPatch = pPatch->pNext)
3441	{
3442	# ifdef RT_ARCH_AMD64
3443	ASMAtomicXchgU64((volatile uint64_t *)&pPatch->auCode[pPatch->offVMMR0EntryFixup], (uint64_t)pvVMMR0);
3444	# else /* RT_ARCH_X86 */
3445	ASMAtomicXchgU32((volatile uint32_t *)&pPatch->auCode[pPatch->offVMMR0EntryFixup],
3446	(uint32_t)pvVMMR0 - (uint32_t)&pPatch->auCode[pPatch->offVMMR0EntryFixup + 4]);
3447	# endif
3448	}
3449	#endif /* VBOX_WITH_IDT_PATCHING */
3450	}
3451	else
3452	{
3453	/*
3454	* Return failure or success depending on whether the values match or not.
3455	*/
3456	if ( pDevExt->pvVMMR0 != pvVMMR0
3457	\|\| (void *)pDevExt->pfnVMMR0EntryInt != pvVMMR0EntryInt
3458	\|\| (void *)pDevExt->pfnVMMR0EntryFast != pvVMMR0EntryFast
3459	\|\| (void *)pDevExt->pfnVMMR0EntryEx != pvVMMR0EntryEx)
3460	{
3461	AssertMsgFailed(("SUP_IOCTL_LDR_SETR0EP: Already set pointing to a different module!\n"));
3462	rc = VERR_INVALID_PARAMETER;
3463	}
3464	}
3465	return rc;
3466	}
3467
3468
3469	/**
3470	* Unsets the R0 entry point installed by supdrvLdrSetR0EP.
3471	*
3472	* @param pDevExt Device globals.
3473	*/
3474	static void supdrvLdrUnsetR0EP(PSUPDRVDEVEXT pDevExt)
3475	{
3476	#ifdef VBOX_WITH_IDT_PATCHING
3477	PSUPDRVPATCH pPatch;
3478	#endif
3479
3480	pDevExt->pvVMMR0 = NULL;
3481	pDevExt->pfnVMMR0EntryInt = NULL;
3482	pDevExt->pfnVMMR0EntryFast = NULL;
3483	pDevExt->pfnVMMR0EntryEx = NULL;
3484
3485	#ifdef VBOX_WITH_IDT_PATCHING
3486	for (pPatch = pDevExt->pIdtPatches; pPatch; pPatch = pPatch->pNext)
3487	{
3488	# ifdef RT_ARCH_AMD64
3489	ASMAtomicXchgU64((volatile uint64_t *)&pPatch->auCode[pPatch->offVMMR0EntryFixup],
3490	(uint64_t)&pPatch->auCode[pPatch->offStub]);
3491	# else /* RT_ARCH_X86 */
3492	ASMAtomicXchgU32((volatile uint32_t *)&pPatch->auCode[pPatch->offVMMR0EntryFixup],
3493	(uint32_t)&pPatch->auCode[pPatch->offStub] - (uint32_t)&pPatch->auCode[pPatch->offVMMR0EntryFixup + 4]);
3494	# endif
3495	}
3496	#endif /* VBOX_WITH_IDT_PATCHING */
3497	}
3498
3499
3500	/**
3501	* Adds a usage reference in the specified session of an image.
3502	*
3503	* @param pSession Session in question.
3504	* @param pImage Image which the session is using.
3505	*/
3506	static void supdrvLdrAddUsage(PSUPDRVSESSION pSession, PSUPDRVLDRIMAGE pImage)
3507	{
3508	PSUPDRVLDRUSAGE pUsage;
3509	LogFlow(("supdrvLdrAddUsage: pImage=%p\n", pImage));
3510
3511	/*
3512	* Referenced it already?
3513	*/
3514	pUsage = pSession->pLdrUsage;
3515	while (pUsage)
3516	{
3517	if (pUsage->pImage == pImage)
3518	{
3519	pUsage->cUsage++;
3520	return;
3521	}
3522	pUsage = pUsage->pNext;
3523	}
3524
3525	/*
3526	* Allocate new usage record.
3527	*/
3528	pUsage = (PSUPDRVLDRUSAGE)RTMemAlloc(sizeof(*pUsage));
3529	Assert(pUsage);
3530	if (pUsage)
3531	{
3532	pUsage->cUsage = 1;
3533	pUsage->pImage = pImage;
3534	pUsage->pNext = pSession->pLdrUsage;
3535	pSession->pLdrUsage = pUsage;
3536	}
3537	/* ignore errors... */
3538	}
3539
3540
3541	/**
3542	* Frees a load image.
3543	*
3544	* @param pDevExt Pointer to device extension.
3545	* @param pImage Pointer to the image we're gonna free.
3546	* This image must exit!
3547	* @remark The caller MUST own SUPDRVDEVEXT::mtxLdr!
3548	*/
3549	static void supdrvLdrFree(PSUPDRVDEVEXT pDevExt, PSUPDRVLDRIMAGE pImage)
3550	{
3551	PSUPDRVLDRIMAGE pImagePrev;
3552	LogFlow(("supdrvLdrFree: pImage=%p\n", pImage));
3553
3554	/* find it - arg. should've used doubly linked list. */
3555	Assert(pDevExt->pLdrImages);
3556	pImagePrev = NULL;
3557	if (pDevExt->pLdrImages != pImage)
3558	{
3559	pImagePrev = pDevExt->pLdrImages;
3560	while (pImagePrev->pNext != pImage)
3561	pImagePrev = pImagePrev->pNext;
3562	Assert(pImagePrev->pNext == pImage);
3563	}
3564
3565	/* unlink */
3566	if (pImagePrev)
3567	pImagePrev->pNext = pImage->pNext;
3568	else
3569	pDevExt->pLdrImages = pImage->pNext;
3570
3571	/* check if this is VMMR0.r0 and fix the Idt patches if it is. */
3572	if (pDevExt->pvVMMR0 == pImage->pvImage)
3573	supdrvLdrUnsetR0EP(pDevExt);
3574
3575	/* check for objects with destructors in this image. (Shouldn't happen.) */
3576	if (pDevExt->pObjs)
3577	{
3578	unsigned cObjs = 0;
3579	PSUPDRVOBJ pObj;
3580	RTSPINLOCKTMP SpinlockTmp = RTSPINLOCKTMP_INITIALIZER;
3581	RTSpinlockAcquire(pDevExt->Spinlock, &SpinlockTmp);
3582	for (pObj = pDevExt->pObjs; pObj; pObj = pObj->pNext)
3583	if (RT_UNLIKELY((uintptr_t)pObj->pfnDestructor - (uintptr_t)pImage->pvImage < pImage->cbImage))
3584	{
3585	pObj->pfnDestructor = NULL;
3586	cObjs++;
3587	}
3588	RTSpinlockRelease(pDevExt->Spinlock, &SpinlockTmp);
3589	if (cObjs)
3590	OSDBGPRINT(("supdrvLdrFree: Image '%s' has %d dangling objects!\n", pImage->szName, cObjs));
3591	}
3592
3593	/* call termination function if fully loaded. */
3594	if ( pImage->pfnModuleTerm
3595	&& pImage->uState == SUP_IOCTL_LDR_LOAD)
3596	{
3597	LogFlow(("supdrvIOCtl_LdrLoad: calling pfnModuleTerm=%p\n", pImage->pfnModuleTerm));
3598	pImage->pfnModuleTerm();
3599	}
3600
3601	/* free the image */
3602	pImage->cUsage = 0;
3603	pImage->pNext = 0;
3604	pImage->uState = SUP_IOCTL_LDR_FREE;
3605	RTMemExecFree(pImage);
3606	}
3607
3608
3609	/**
3610	* Gets the current paging mode of the CPU and stores in in pOut.
3611	*/
3612	static SUPPAGINGMODE supdrvIOCtl_GetPagingMode(void)
3613	{
3614	SUPPAGINGMODE enmMode;
3615
3616	RTUINTREG cr0 = ASMGetCR0();
3617	if ((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE))
3618	enmMode = SUPPAGINGMODE_INVALID;
3619	else
3620	{
3621	RTUINTREG cr4 = ASMGetCR4();
3622	uint32_t fNXEPlusLMA = 0;
3623	if (cr4 & X86_CR4_PAE)
3624	{
3625	uint32_t fAmdFeatures = ASMCpuId_EDX(0x80000001);
3626	if (fAmdFeatures & (X86_CPUID_AMD_FEATURE_EDX_NX \| X86_CPUID_AMD_FEATURE_EDX_LONG_MODE))
3627	{
3628	uint64_t efer = ASMRdMsr(MSR_K6_EFER);
3629	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_NX) && (efer & MSR_K6_EFER_NXE))
3630	fNXEPlusLMA \|= RT_BIT(0);
3631	if ((fAmdFeatures & X86_CPUID_AMD_FEATURE_EDX_LONG_MODE) && (efer & MSR_K6_EFER_LMA))
3632	fNXEPlusLMA \|= RT_BIT(1);
3633	}
3634	}
3635
3636	switch ((cr4 & (X86_CR4_PAE \| X86_CR4_PGE)) \| fNXEPlusLMA)
3637	{
3638	case 0:
3639	enmMode = SUPPAGINGMODE_32_BIT;
3640	break;
3641
3642	case X86_CR4_PGE:
3643	enmMode = SUPPAGINGMODE_32_BIT_GLOBAL;
3644	break;
3645
3646	case X86_CR4_PAE:
3647	enmMode = SUPPAGINGMODE_PAE;
3648	break;
3649
3650	case X86_CR4_PAE \| RT_BIT(0):
3651	enmMode = SUPPAGINGMODE_PAE_NX;
3652	break;
3653
3654	case X86_CR4_PAE \| X86_CR4_PGE:
3655	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3656	break;
3657
3658	case X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3659	enmMode = SUPPAGINGMODE_PAE_GLOBAL;
3660	break;
3661
3662	case RT_BIT(1) \| X86_CR4_PAE:
3663	enmMode = SUPPAGINGMODE_AMD64;
3664	break;
3665
3666	case RT_BIT(1) \| X86_CR4_PAE \| RT_BIT(0):
3667	enmMode = SUPPAGINGMODE_AMD64_NX;
3668	break;
3669
3670	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE:
3671	enmMode = SUPPAGINGMODE_AMD64_GLOBAL;
3672	break;
3673
3674	case RT_BIT(1) \| X86_CR4_PAE \| X86_CR4_PGE \| RT_BIT(0):
3675	enmMode = SUPPAGINGMODE_AMD64_GLOBAL_NX;
3676	break;
3677
3678	default:
3679	AssertMsgFailed(("Cannot happen! cr4=%#x fNXEPlusLMA=%d\n", cr4, fNXEPlusLMA));
3680	enmMode = SUPPAGINGMODE_INVALID;
3681	break;
3682	}
3683	}
3684	return enmMode;
3685	}
3686
3687
3688	#ifdef USE_NEW_OS_INTERFACE_FOR_GIP
3689	/**
3690	* Creates the GIP.
3691	*
3692	* @returns negative errno.
3693	* @param pDevExt Instance data. GIP stuff may be updated.
3694	*/
3695	static int supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
3696	{
3697	PSUPGLOBALINFOPAGE pGip;
3698	RTHCPHYS HCPhysGip;
3699	uint32_t u32SystemResolution;
3700	uint32_t u32Interval;
3701	int rc;
3702
3703	LogFlow(("supdrvGipCreate:\n"));
3704
3705	/* assert order */
3706	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
3707	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
3708	Assert(!pDevExt->pGipTimer);
3709
3710	/*
3711	* Allocate a suitable page with a default kernel mapping.
3712	*/
3713	rc = RTR0MemObjAllocLow(&pDevExt->GipMemObj, PAGE_SIZE, false);
3714	if (RT_FAILURE(rc))
3715	{
3716	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
3717	return rc;
3718	}
3719	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
3720	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
3721
3722	/*
3723	* Try bump up the system timer resolution.
3724	* The more interrupts the better...
3725	*/
3726	if ( RT_SUCCESS(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
3727	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
3728	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 3906250 /* 256 HZ */, &u32SystemResolution))
3729	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 4000000 /* 250 HZ */, &u32SystemResolution))
3730	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity( 7812500 /* 128 HZ */, &u32SystemResolution))
3731	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity(10000000 /* 100 HZ */, &u32SystemResolution))
3732	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity(15625000 /* 64 HZ */, &u32SystemResolution))
3733	\|\| RT_SUCCESS(RTTimerRequestSystemGranularity(31250000 /* 32 HZ */, &u32SystemResolution))
3734	)
3735	{
3736	Assert(RTTimerGetSystemGranularity() <= u32SystemResolution);
3737	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
3738	}
3739
3740	/*
3741	* Find a reasonable update interval, something close to 10ms would be nice,
3742	* and create a recurring timer.
3743	*/
3744	u32Interval = u32SystemResolution = RTTimerGetSystemGranularity();
3745	while (u32Interval < 10000000 /* 10 ms */)
3746	u32Interval += u32SystemResolution;
3747
3748	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0, supdrvGipTimer, pDevExt);
3749	if (RT_FAILURE(rc))
3750	{
3751	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %RU32 ns interval. rc=%d\n", u32Interval, rc));
3752	Assert(!pDevExt->pGipTimer);
3753	supdrvGipDestroy(pDevExt);
3754	return rc;
3755	}
3756
3757	/*
3758	* We're good.
3759	*/
3760	supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), 1000000000 / u32Interval /=Hz/);
3761	return VINF_SUCCESS;
3762	}
3763
3764
3765	/**
3766	* Terminates the GIP.
3767	*
3768	* @param pDevExt Instance data. GIP stuff may be updated.
3769	*/
3770	static void supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
3771	{
3772	int rc;
3773	#ifdef DEBUG_DARWIN_GIP
3774	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
3775	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
3776	pDevExt->pGipTimer, pDevExt->GipMemObj));
3777	#endif
3778
3779	/*
3780	* Invalid the GIP data.
3781	*/
3782	if (pDevExt->pGip)
3783	{
3784	supdrvGipTerm(pDevExt->pGip);
3785	pDevExt->pGip = NULL;
3786	}
3787
3788	/*
3789	* Destroy the timer and free the GIP memory object.
3790	*/
3791	if (pDevExt->pGipTimer)
3792	{
3793	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
3794	pDevExt->pGipTimer = NULL;
3795	}
3796
3797	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
3798	{
3799	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
3800	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
3801	}
3802
3803	/*
3804	* Finally, release the system timer resolution request if one succeeded.
3805	*/
3806	if (pDevExt->u32SystemTimerGranularityGrant)
3807	{
3808	rc = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant); AssertRC(rc);
3809	pDevExt->u32SystemTimerGranularityGrant = 0;
3810	}
3811	}
3812
3813
3814	/**
3815	* Timer callback function.
3816	* @param pTimer The timer.
3817	* @param pvUser The device extension.
3818	*/
3819	static DECLCALLBACK(void) supdrvGipTimer(PRTTIMER pTimer, void *pvUser)
3820	{
3821	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
3822	supdrvGipUpdate(pDevExt->pGip, RTTimeSystemNanoTS());
3823	}
3824	#endif /* USE_NEW_OS_INTERFACE_FOR_GIP */
3825
3826
3827	/**
3828	* Initializes the GIP data.
3829	*
3830	* @returns IPRT status code.
3831	* @param pDevExt Pointer to the device instance data.
3832	* @param pGip Pointer to the read-write kernel mapping of the GIP.
3833	* @param HCPhys The physical address of the GIP.
3834	* @param u64NanoTS The current nanosecond timestamp.
3835	* @param uUpdateHz The update freqence.
3836	*/
3837	int VBOXCALL supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys, uint64_t u64NanoTS, unsigned uUpdateHz)
3838	{
3839	unsigned i;
3840	#ifdef DEBUG_DARWIN_GIP
3841	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz));
3842	#else
3843	LogFlow(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz));
3844	#endif
3845
3846	/*
3847	* Initialize the structure.
3848	*/
3849	memset(pGip, 0, PAGE_SIZE);
3850	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
3851	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
3852	pGip->u32Mode = supdrvGipDeterminTscMode();
3853	pGip->u32UpdateHz = uUpdateHz;
3854	pGip->u32UpdateIntervalNS = 1000000000 / uUpdateHz;
3855	pGip->u64NanoTSLastUpdateHz = u64NanoTS;
3856
3857	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
3858	{
3859	pGip->aCPUs[i].u32TransactionId = 2;
3860	pGip->aCPUs[i].u64NanoTS = u64NanoTS;
3861	pGip->aCPUs[i].u64TSC = ASMReadTSC();
3862
3863	/*
3864	* We don't know the following values until we've executed updates.
3865	* So, we'll just insert very high values.
3866	*/
3867	pGip->aCPUs[i].u64CpuHz = _4G + 1;
3868	pGip->aCPUs[i].u32UpdateIntervalTSC = _2G / 4;
3869	pGip->aCPUs[i].au32TSCHistory[0] = _2G / 4;
3870	pGip->aCPUs[i].au32TSCHistory[1] = _2G / 4;
3871	pGip->aCPUs[i].au32TSCHistory[2] = _2G / 4;
3872	pGip->aCPUs[i].au32TSCHistory[3] = _2G / 4;
3873	pGip->aCPUs[i].au32TSCHistory[4] = _2G / 4;
3874	pGip->aCPUs[i].au32TSCHistory[5] = _2G / 4;
3875	pGip->aCPUs[i].au32TSCHistory[6] = _2G / 4;
3876	pGip->aCPUs[i].au32TSCHistory[7] = _2G / 4;
3877	}
3878
3879	/*
3880	* Link it to the device extension.
3881	*/
3882	pDevExt->pGip = pGip;
3883	pDevExt->HCPhysGip = HCPhys;
3884	pDevExt->cGipUsers = 0;
3885
3886	return VINF_SUCCESS;
3887	}
3888
3889
3890	/**
3891	* Determin the GIP TSC mode.
3892	*
3893	* @returns The most suitable TSC mode.
3894	*/
3895	static SUPGIPMODE supdrvGipDeterminTscMode(void)
3896	{
3897	#ifndef USE_NEW_OS_INTERFACE_FOR_GIP
3898	/*
3899	* The problem here is that AMD processors with power management features
3900	* may easily end up with different TSCs because the CPUs or even cores
3901	* on the same physical chip run at different frequencies to save power.
3902	*
3903	* It is rumoured that this will be corrected with Barcelona and it's
3904	* expected that this will be indicated by the TscInvariant bit in
3905	* cpuid(0x80000007). So, the "difficult" bit here is to correctly
3906	* identify the older CPUs which don't do different frequency and
3907	* can be relied upon to have somewhat uniform TSC between the cpus.
3908	*/
3909	if (supdrvOSGetCPUCount() > 1)
3910	{
3911	uint32_t uEAX, uEBX, uECX, uEDX;
3912
3913	/* Permit user users override. */
3914	if (supdrvOSGetForcedAsyncTscMode())
3915	return SUPGIPMODE_ASYNC_TSC;
3916
3917	/* Check for "AuthenticAMD" */
3918	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
3919	if (uEAX >= 1 && uEBX == 0x68747541 && uECX == 0x444d4163 && uEDX == 0x69746e65)
3920	{
3921	/* Check for APM support and that TscInvariant is cleared. */
3922	ASMCpuId(0x80000000, &uEAX, &uEBX, &uECX, &uEDX);
3923	if (uEAX >= 0x80000007)
3924	{
3925	ASMCpuId(0x80000007, &uEAX, &uEBX, &uECX, &uEDX);
3926	if ( !(uEDX & RT_BIT(8))/* TscInvariant */
3927	&& (uEDX & 0x3e)) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
3928	return SUPGIPMODE_ASYNC_TSC;
3929	}
3930	}
3931	}
3932	#endif
3933	return SUPGIPMODE_SYNC_TSC;
3934	}
3935
3936
3937	/**
3938	* Invalidates the GIP data upon termination.
3939	*
3940	* @param pGip Pointer to the read-write kernel mapping of the GIP.
3941	*/
3942	void VBOXCALL supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
3943	{
3944	unsigned i;
3945	pGip->u32Magic = 0;
3946	for (i = 0; i < RT_ELEMENTS(pGip->aCPUs); i++)
3947	{
3948	pGip->aCPUs[i].u64NanoTS = 0;
3949	pGip->aCPUs[i].u64TSC = 0;
3950	pGip->aCPUs[i].iTSCHistoryHead = 0;
3951	}
3952	}
3953
3954
3955	/**
3956	* Worker routine for supdrvGipUpdate and supdrvGipUpdatePerCpu that
3957	* updates all the per cpu data except the transaction id.
3958	*
3959	* @param pGip The GIP.
3960	* @param pGipCpu Pointer to the per cpu data.
3961	* @param u64NanoTS The current time stamp.
3962	*/
3963	static void supdrvGipDoUpdateCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
3964	{
3965	uint64_t u64TSC;
3966	uint64_t u64TSCDelta;
3967	uint32_t u32UpdateIntervalTSC;
3968	uint32_t u32UpdateIntervalTSCSlack;
3969	unsigned iTSCHistoryHead;
3970	uint64_t u64CpuHz;
3971
3972	/*
3973	* Update the NanoTS.
3974	*/
3975	ASMAtomicXchgU64(&pGipCpu->u64NanoTS, u64NanoTS);
3976
3977	/*
3978	* Calc TSC delta.
3979	*/
3980	/** @todo validate the NanoTS delta, don't trust the OS to call us when it should... */
3981	u64TSC = ASMReadTSC();
3982	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
3983	ASMAtomicXchgU64(&pGipCpu->u64TSC, u64TSC);
3984
3985	if (u64TSCDelta >> 32)
3986	{
3987	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
3988	pGipCpu->cErrors++;
3989	}
3990
3991	/*
3992	* TSC History.
3993	*/
3994	Assert(ELEMENTS(pGipCpu->au32TSCHistory) == 8);
3995
3996	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
3997	ASMAtomicXchgU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
3998	ASMAtomicXchgU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
3999
4000	/*
4001	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
4002	*/
4003	if (pGip->u32UpdateHz >= 1000)
4004	{
4005	uint32_t u32;
4006	u32 = pGipCpu->au32TSCHistory[0];
4007	u32 += pGipCpu->au32TSCHistory[1];
4008	u32 += pGipCpu->au32TSCHistory[2];
4009	u32 += pGipCpu->au32TSCHistory[3];
4010	u32 >>= 2;
4011	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
4012	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
4013	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
4014	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
4015	u32UpdateIntervalTSC >>= 2;
4016	u32UpdateIntervalTSC += u32;
4017	u32UpdateIntervalTSC >>= 1;
4018
4019	/* Value choosen for a 2GHz Athlon64 running linux 2.6.10/11, . */
4020	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
4021	}
4022	else if (pGip->u32UpdateHz >= 90)
4023	{
4024	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
4025	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
4026	u32UpdateIntervalTSC >>= 1;
4027
4028	/* value choosen on a 2GHz thinkpad running windows */
4029	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
4030	}
4031	else
4032	{
4033	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
4034
4035	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
4036	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
4037	}
4038	ASMAtomicXchgU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
4039
4040	/*
4041	* CpuHz.
4042	*/
4043	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, pGip->u32UpdateHz);
4044	ASMAtomicXchgU64(&pGipCpu->u64CpuHz, u64CpuHz);
4045	}
4046
4047
4048	/**
4049	* Updates the GIP.
4050	*
4051	* @param pGip Pointer to the GIP.
4052	* @param u64NanoTS The current nanosecond timesamp.
4053	*/
4054	void VBOXCALL supdrvGipUpdate(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS)
4055	{
4056	/*
4057	* Determin the relevant CPU data.
4058	*/
4059	PSUPGIPCPU pGipCpu;
4060	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
4061	pGipCpu = &pGip->aCPUs[0];
4062	else
4063	{
4064	unsigned iCpu = ASMGetApicId();
4065	if (RT_LIKELY(iCpu >= RT_ELEMENTS(pGip->aCPUs)))
4066	return;
4067	pGipCpu = &pGip->aCPUs[iCpu];
4068	}
4069
4070	/*
4071	* Start update transaction.
4072	*/
4073	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
4074	{
4075	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
4076	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
4077	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
4078	pGipCpu->cErrors++;
4079	return;
4080	}
4081
4082	/*
4083	* Recalc the update frequency every 0x800th time.
4084	*/
4085	if (!(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
4086	{
4087	if (pGip->u64NanoTSLastUpdateHz)
4088	{
4089	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
4090	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
4091	uint32_t u32UpdateHz = (uint32_t)((UINT64_C(1000000000) * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
4092	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
4093	{
4094	ASMAtomicXchgU32(&pGip->u32UpdateHz, u32UpdateHz);
4095	ASMAtomicXchgU32(&pGip->u32UpdateIntervalNS, 1000000000 / u32UpdateHz);
4096	}
4097	#endif
4098	}
4099	ASMAtomicXchgU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS);
4100	}
4101
4102	/*
4103	* Update the data.
4104	*/
4105	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS);
4106
4107	/*
4108	* Complete transaction.
4109	*/
4110	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
4111	}
4112
4113
4114	/**
4115	* Updates the per cpu GIP data for the calling cpu.
4116	*
4117	* @param pGip Pointer to the GIP.
4118	* @param u64NanoTS The current nanosecond timesamp.
4119	* @param iCpu The CPU index.
4120	*/
4121	void VBOXCALL supdrvGipUpdatePerCpu(PSUPGLOBALINFOPAGE pGip, uint64_t u64NanoTS, unsigned iCpu)
4122	{
4123	PSUPGIPCPU pGipCpu;
4124
4125	if (RT_LIKELY(iCpu < RT_ELEMENTS(pGip->aCPUs)))
4126	{
4127	pGipCpu = &pGip->aCPUs[iCpu];
4128
4129	/*
4130	* Start update transaction.
4131	*/
4132	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
4133	{
4134	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
4135	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
4136	pGipCpu->cErrors++;
4137	return;
4138	}
4139
4140	/*
4141	* Update the data.
4142	*/
4143	supdrvGipDoUpdateCpu(pGip, pGipCpu, u64NanoTS);
4144
4145	/*
4146	* Complete transaction.
4147	*/
4148	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
4149	}
4150	}
4151
4152
4153	/**
4154	* Callback used by supdrvDetermineAsyncTSC to read the TSC on a CPU.
4155	*
4156	* @param idCpu Ignored.
4157	* @param pvUser1 Where to put the TSC.
4158	* @param pvUser2 Ignored.
4159	*/
4160	static DECLCALLBACK(void) supdrvDetermineAsyncTscWorker(RTCPUID idCpu, void pvUser1, void pvUser2)
4161	{
4162	(uint64_t )pvUser1 = ASMReadTSC();
4163	}
4164
4165
4166	/**
4167	* Determine if Async GIP mode is required because of TSC drift.
4168	*
4169	* When using the default/normal timer code it is essential that the time stamp counter
4170	* (TSC) runs never backwards, that is, a read operation to the counter should return
4171	* a bigger value than any previous read operation. This is guaranteed by the latest
4172	* AMD CPUs and by newer Intel CPUs which never enter the C2 state (P4). In any other
4173	* case we have to choose the asynchronous timer mode.
4174	*
4175	* @param pu64Diff pointer to the determined difference between different cores.
4176	* @return false if the time stamp counters appear to be synchron, true otherwise.
4177	*/
4178	bool VBOXCALL supdrvDetermineAsyncTsc(uint64_t *pu64DiffCores)
4179	{
4180	static uint64_t s_aTsc[8][RTCPUSET_MAX_CPUS];
4181	uint64_t u64Diff, u64DiffMin, u64DiffMax, u64TscLast;
4182	int iSlot, iCpu, cCpus;
4183	bool fBackwards;
4184	RTCPUSET OnlineCpus;
4185	int rc;
4186
4187	*pu64DiffCores = 1;
4188
4189	RTMpGetOnlineSet(&OnlineCpus);
4190	cCpus = RTCpuSetCount(&OnlineCpus);
4191	if (cCpus < 2)
4192	return false;
4193	Assert(cCpus <= RT_ELEMENTS(s_aTsc[0]));
4194
4195	/*
4196	* Collect data from the online CPUs.
4197	*/
4198	for (iSlot = 0; iSlot < RT_ELEMENTS(s_aTsc); iSlot++)
4199	{
4200	RTCPUID iCpuSet = 0;
4201	for (iCpu = 0; iCpu < cCpus; iCpu++)
4202	{
4203	while (!RTCpuSetIsMember(&OnlineCpus, iCpuSet))
4204	iCpuSet++; /* skip offline CPU */
4205	rc = RTMpOnSpecific(RTMpCpuIdFromSetIndex(iCpuSet), supdrvDetermineAsyncTscWorker, &s_aTsc[iSlot][iCpu], NULL);
4206	if (rc == VERR_NOT_SUPPORTED)
4207	return false;
4208	iCpuSet++;
4209	}
4210	}
4211
4212	/*
4213	* Check that the TSC reads are strictly ascending.
4214	*/
4215	fBackwards = false;
4216	u64DiffMin = (uint64_t)~0;
4217	u64TscLast = 0;
4218	for (iSlot = 0; iSlot < RT_ELEMENTS(s_aTsc); iSlot++)
4219	{
4220	uint64_t u64Tsc0 = s_aTsc[iSlot][0];
4221	u64DiffMax = 0;
4222	if (u64Tsc0 <= u64TscLast)
4223	fBackwards = true;
4224	u64TscLast = u64Tsc0;
4225	for (iCpu = 1; iCpu < cCpus; iCpu++)
4226	{
4227	uint64_t u64TscN = s_aTsc[iSlot][iCpu];
4228	if (u64TscN <= u64TscLast)
4229	fBackwards = true;
4230	u64TscLast = u64TscN;
4231
4232	u64Diff = u64TscN > u64Tsc0 ? u64TscN - u64Tsc0 : u64Tsc0 - u64TscN;
4233	if (u64DiffMax < u64Diff)
4234	u64DiffMax = u64Diff;
4235	}
4236	if (u64DiffMin > u64DiffMax)
4237	u64DiffMin = u64DiffMax;
4238	}
4239	/* informational */
4240	*pu64DiffCores = u64DiffMin;
4241
4242	return fBackwards;
4243	}
4244
4245
4246	#ifndef DEBUG /** @todo change #ifndef DEBUG -> #ifdef LOG_ENABLED */
4247	/**
4248	* Stub function for non-debug builds.
4249	*/
4250	RTDECL(PRTLOGGER) RTLogDefaultInstance(void)
4251	{
4252	return NULL;
4253	}
4254
4255	RTDECL(PRTLOGGER) RTLogRelDefaultInstance(void)
4256	{
4257	return NULL;
4258	}
4259
4260	/**
4261	* Stub function for non-debug builds.
4262	*/
4263	RTDECL(int) RTLogSetDefaultInstanceThread(PRTLOGGER pLogger, uintptr_t uKey)
4264	{
4265	return 0;
4266	}
4267
4268	/**
4269	* Stub function for non-debug builds.
4270	*/
4271	RTDECL(void) RTLogLogger(PRTLOGGER pLogger, void pvCallerRet, const char pszFormat, ...)
4272	{
4273	}
4274
4275	/**
4276	* Stub function for non-debug builds.
4277	*/
4278	RTDECL(void) RTLogLoggerEx(PRTLOGGER pLogger, unsigned fFlags, unsigned iGroup, const char *pszFormat, ...)
4279	{
4280	}
4281
4282	/**
4283	* Stub function for non-debug builds.
4284	*/
4285	RTDECL(void) RTLogLoggerExV(PRTLOGGER pLogger, unsigned fFlags, unsigned iGroup, const char *pszFormat, va_list args)
4286	{
4287	}
4288
4289	/**
4290	* Stub function for non-debug builds.
4291	*/
4292	RTDECL(void) RTLogPrintf(const char *pszFormat, ...)
4293	{
4294	}
4295
4296	/**
4297	* Stub function for non-debug builds.
4298	*/
4299	RTDECL(void) RTLogPrintfV(const char *pszFormat, va_list args)
4300	{
4301	}
4302	#endif /* !DEBUG */
4303

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

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDRVShared.c@ 8760

Download in other formats: