VMMR0.cpp@ 78003

Last change on this file since 78003 was 76553, checked in by vboxsync, 6 years ago
scm --update-copyright-year
Property svn:eol-style set to `native` Property svn:keywords set to `Id Revision`
File size: 103.9 KB

Line
1	/* $Id: VMMR0.cpp 76553 2019-01-01 01:45:53Z vboxsync $ */
2	/** @file
3	* VMM - Host Context Ring 0.
4	*/
5
6	/*
7	* Copyright (C) 2006-2019 Oracle Corporation
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
18
19	/*********************************************************************************************************************************
20	* Header Files *
21	*********************************************************************************************************************************/
22	#define LOG_GROUP LOG_GROUP_VMM
23	#include <VBox/vmm/vmm.h>
24	#include <VBox/sup.h>
25	#include <VBox/vmm/trpm.h>
26	#include <VBox/vmm/cpum.h>
27	#include <VBox/vmm/pdmapi.h>
28	#include <VBox/vmm/pgm.h>
29	#ifdef VBOX_WITH_NEM_R0
30	# include <VBox/vmm/nem.h>
31	#endif
32	#include <VBox/vmm/em.h>
33	#include <VBox/vmm/stam.h>
34	#include <VBox/vmm/tm.h>
35	#include "VMMInternal.h"
36	#include <VBox/vmm/vm.h>
37	#include <VBox/vmm/gvm.h>
38	#ifdef VBOX_WITH_PCI_PASSTHROUGH
39	# include <VBox/vmm/pdmpci.h>
40	#endif
41	#include <VBox/vmm/apic.h>
42
43	#include <VBox/vmm/gvmm.h>
44	#include <VBox/vmm/gmm.h>
45	#include <VBox/vmm/gim.h>
46	#include <VBox/intnet.h>
47	#include <VBox/vmm/hm.h>
48	#include <VBox/param.h>
49	#include <VBox/err.h>
50	#include <VBox/version.h>
51	#include <VBox/log.h>
52
53	#include <iprt/asm-amd64-x86.h>
54	#include <iprt/assert.h>
55	#include <iprt/crc.h>
56	#include <iprt/mp.h>
57	#include <iprt/once.h>
58	#include <iprt/stdarg.h>
59	#include <iprt/string.h>
60	#include <iprt/thread.h>
61	#include <iprt/timer.h>
62	#include <iprt/time.h>
63
64	#include "dtrace/VBoxVMM.h"
65
66
67	#if defined(_MSC_VER) && defined(RT_ARCH_AMD64) /** @todo check this with with VC7! */
68	# pragma intrinsic(_AddressOfReturnAddress)
69	#endif
70
71	#if defined(RT_OS_DARWIN) && ARCH_BITS == 32
72	# error "32-bit darwin is no longer supported. Go back to 4.3 or earlier!"
73	#endif
74
75
76
77	/*********************************************************************************************************************************
78	* Defined Constants And Macros *
79	*********************************************************************************************************************************/
80	/** @def VMM_CHECK_SMAP_SETUP
81	* SMAP check setup. */
82	/** @def VMM_CHECK_SMAP_CHECK
83	* Checks that the AC flag is set if SMAP is enabled. If AC is not set,
84	* it will be logged and @a a_BadExpr is executed. */
85	/** @def VMM_CHECK_SMAP_CHECK2
86	* Checks that the AC flag is set if SMAP is enabled. If AC is not set, it will
87	* be logged, written to the VMs assertion text buffer, and @a a_BadExpr is
88	* executed. */
89	#if defined(VBOX_STRICT) \|\| 1
90	# define VMM_CHECK_SMAP_SETUP() uint32_t const fKernelFeatures = SUPR0GetKernelFeatures()
91	# define VMM_CHECK_SMAP_CHECK(a_BadExpr) \
92	do { \
93	if (fKernelFeatures & SUPKERNELFEATURES_SMAP) \
94	{ \
95	RTCCUINTREG fEflCheck = ASMGetFlags(); \
96	if (RT_LIKELY(fEflCheck & X86_EFL_AC)) \
97	{ /* likely */ } \
98	else \
99	{ \
100	SUPR0Printf("%s, line %d: EFLAGS.AC is clear! (%#x)\n", __FUNCTION__, __LINE__, (uint32_t)fEflCheck); \
101	a_BadExpr; \
102	} \
103	} \
104	} while (0)
105	# define VMM_CHECK_SMAP_CHECK2(a_pVM, a_BadExpr) \
106	do { \
107	if (fKernelFeatures & SUPKERNELFEATURES_SMAP) \
108	{ \
109	RTCCUINTREG fEflCheck = ASMGetFlags(); \
110	if (RT_LIKELY(fEflCheck & X86_EFL_AC)) \
111	{ /* likely */ } \
112	else \
113	{ \
114	SUPR0BadContext((a_pVM) ? (a_pVM)->pSession : NULL, __FILE__, __LINE__, "EFLAGS.AC is zero!"); \
115	RTStrPrintf(pVM->vmm.s.szRing0AssertMsg1, sizeof(pVM->vmm.s.szRing0AssertMsg1), \
116	"%s, line %d: EFLAGS.AC is clear! (%#x)\n", __FUNCTION__, __LINE__, (uint32_t)fEflCheck); \
117	a_BadExpr; \
118	} \
119	} \
120	} while (0)
121	#else
122	# define VMM_CHECK_SMAP_SETUP() uint32_t const fKernelFeatures = 0
123	# define VMM_CHECK_SMAP_CHECK(a_BadExpr) NOREF(fKernelFeatures)
124	# define VMM_CHECK_SMAP_CHECK2(a_pVM, a_BadExpr) NOREF(fKernelFeatures)
125	#endif
126
127
128	/*********************************************************************************************************************************
129	* Internal Functions *
130	*********************************************************************************************************************************/
131	RT_C_DECLS_BEGIN
132	#if defined(RT_ARCH_X86) && (defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD))
133	extern uint64_t __udivdi3(uint64_t, uint64_t);
134	extern uint64_t __umoddi3(uint64_t, uint64_t);
135	#endif
136	RT_C_DECLS_END
137
138
139	/*********************************************************************************************************************************
140	* Global Variables *
141	*********************************************************************************************************************************/
142	/** Drag in necessary library bits.
143	* The runtime lives here (in VMMR0.r0) and VBoxDDR0.r0 links against us. /
144	PFNRT g_VMMR0Deps[] =
145	{
146	(PFNRT)RTCrc32,
147	(PFNRT)RTOnce,
148	#if defined(RT_ARCH_X86) && (defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD))
149	(PFNRT)__udivdi3,
150	(PFNRT)__umoddi3,
151	#endif
152	NULL
153	};
154
155	#ifdef RT_OS_SOLARIS
156	/* Dependency information for the native solaris loader. */
157	extern "C" { char _depends_on[] = "vboxdrv"; }
158	#endif
159
160	/** The result of SUPR0GetRawModeUsability(), set by ModuleInit(). */
161	int g_rcRawModeUsability = VINF_SUCCESS;
162
163
164	/**
165	* Initialize the module.
166	* This is called when we're first loaded.
167	*
168	* @returns 0 on success.
169	* @returns VBox status on failure.
170	* @param hMod Image handle for use in APIs.
171	*/
172	DECLEXPORT(int) ModuleInit(void *hMod)
173	{
174	VMM_CHECK_SMAP_SETUP();
175	VMM_CHECK_SMAP_CHECK(RT_NOTHING);
176
177	#ifdef VBOX_WITH_DTRACE_R0
178	/*
179	* The first thing to do is register the static tracepoints.
180	* (Deregistration is automatic.)
181	*/
182	int rc2 = SUPR0TracerRegisterModule(hMod, &g_VTGObjHeader);
183	if (RT_FAILURE(rc2))
184	return rc2;
185	#endif
186	LogFlow(("ModuleInit:\n"));
187
188	#ifdef VBOX_WITH_64ON32_CMOS_DEBUG
189	/*
190	* Display the CMOS debug code.
191	*/
192	ASMOutU8(0x72, 0x03);
193	uint8_t bDebugCode = ASMInU8(0x73);
194	LogRel(("CMOS Debug Code: %#x (%d)\n", bDebugCode, bDebugCode));
195	RTLogComPrintf("CMOS Debug Code: %#x (%d)\n", bDebugCode, bDebugCode);
196	#endif
197
198	/*
199	* Initialize the VMM, GVMM, GMM, HM, PGM (Darwin) and INTNET.
200	*/
201	int rc = vmmInitFormatTypes();
202	if (RT_SUCCESS(rc))
203	{
204	VMM_CHECK_SMAP_CHECK(RT_NOTHING);
205	rc = GVMMR0Init();
206	if (RT_SUCCESS(rc))
207	{
208	VMM_CHECK_SMAP_CHECK(RT_NOTHING);
209	rc = GMMR0Init();
210	if (RT_SUCCESS(rc))
211	{
212	VMM_CHECK_SMAP_CHECK(RT_NOTHING);
213	rc = HMR0Init();
214	if (RT_SUCCESS(rc))
215	{
216	VMM_CHECK_SMAP_CHECK(RT_NOTHING);
217	rc = PGMRegisterStringFormatTypes();
218	if (RT_SUCCESS(rc))
219	{
220	VMM_CHECK_SMAP_CHECK(RT_NOTHING);
221	#ifdef VBOX_WITH_2X_4GB_ADDR_SPACE
222	rc = PGMR0DynMapInit();
223	#endif
224	if (RT_SUCCESS(rc))
225	{
226	VMM_CHECK_SMAP_CHECK(RT_NOTHING);
227	rc = IntNetR0Init();
228	if (RT_SUCCESS(rc))
229	{
230	#ifdef VBOX_WITH_PCI_PASSTHROUGH
231	VMM_CHECK_SMAP_CHECK(RT_NOTHING);
232	rc = PciRawR0Init();
233	#endif
234	if (RT_SUCCESS(rc))
235	{
236	VMM_CHECK_SMAP_CHECK(RT_NOTHING);
237	rc = CPUMR0ModuleInit();
238	if (RT_SUCCESS(rc))
239	{
240	#ifdef VBOX_WITH_TRIPLE_FAULT_HACK
241	VMM_CHECK_SMAP_CHECK(RT_NOTHING);
242	rc = vmmR0TripleFaultHackInit();
243	if (RT_SUCCESS(rc))
244	#endif
245	{
246	VMM_CHECK_SMAP_CHECK(rc = VERR_VMM_SMAP_BUT_AC_CLEAR);
247	if (RT_SUCCESS(rc))
248	{
249	g_rcRawModeUsability = SUPR0GetRawModeUsability();
250	if (g_rcRawModeUsability != VINF_SUCCESS)
251	SUPR0Printf("VMMR0!ModuleInit: SUPR0GetRawModeUsability -> %Rrc\n",
252	g_rcRawModeUsability);
253	LogFlow(("ModuleInit: returns success\n"));
254	return VINF_SUCCESS;
255	}
256	}
257
258	/*
259	* Bail out.
260	*/
261	#ifdef VBOX_WITH_TRIPLE_FAULT_HACK
262	vmmR0TripleFaultHackTerm();
263	#endif
264	}
265	else
266	LogRel(("ModuleInit: CPUMR0ModuleInit -> %Rrc\n", rc));
267	#ifdef VBOX_WITH_PCI_PASSTHROUGH
268	PciRawR0Term();
269	#endif
270	}
271	else
272	LogRel(("ModuleInit: PciRawR0Init -> %Rrc\n", rc));
273	IntNetR0Term();
274	}
275	else
276	LogRel(("ModuleInit: IntNetR0Init -> %Rrc\n", rc));
277	#ifdef VBOX_WITH_2X_4GB_ADDR_SPACE
278	PGMR0DynMapTerm();
279	#endif
280	}
281	else
282	LogRel(("ModuleInit: PGMR0DynMapInit -> %Rrc\n", rc));
283	PGMDeregisterStringFormatTypes();
284	}
285	else
286	LogRel(("ModuleInit: PGMRegisterStringFormatTypes -> %Rrc\n", rc));
287	HMR0Term();
288	}
289	else
290	LogRel(("ModuleInit: HMR0Init -> %Rrc\n", rc));
291	GMMR0Term();
292	}
293	else
294	LogRel(("ModuleInit: GMMR0Init -> %Rrc\n", rc));
295	GVMMR0Term();
296	}
297	else
298	LogRel(("ModuleInit: GVMMR0Init -> %Rrc\n", rc));
299	vmmTermFormatTypes();
300	}
301	else
302	LogRel(("ModuleInit: vmmInitFormatTypes -> %Rrc\n", rc));
303
304	LogFlow(("ModuleInit: failed %Rrc\n", rc));
305	return rc;
306	}
307
308
309	/**
310	* Terminate the module.
311	* This is called when we're finally unloaded.
312	*
313	* @param hMod Image handle for use in APIs.
314	*/
315	DECLEXPORT(void) ModuleTerm(void *hMod)
316	{
317	NOREF(hMod);
318	LogFlow(("ModuleTerm:\n"));
319
320	/*
321	* Terminate the CPUM module (Local APIC cleanup).
322	*/
323	CPUMR0ModuleTerm();
324
325	/*
326	* Terminate the internal network service.
327	*/
328	IntNetR0Term();
329
330	/*
331	* PGM (Darwin), HM and PciRaw global cleanup.
332	*/
333	#ifdef VBOX_WITH_2X_4GB_ADDR_SPACE
334	PGMR0DynMapTerm();
335	#endif
336	#ifdef VBOX_WITH_PCI_PASSTHROUGH
337	PciRawR0Term();
338	#endif
339	PGMDeregisterStringFormatTypes();
340	HMR0Term();
341	#ifdef VBOX_WITH_TRIPLE_FAULT_HACK
342	vmmR0TripleFaultHackTerm();
343	#endif
344
345	/*
346	* Destroy the GMM and GVMM instances.
347	*/
348	GMMR0Term();
349	GVMMR0Term();
350
351	vmmTermFormatTypes();
352
353	LogFlow(("ModuleTerm: returns\n"));
354	}
355
356
357	/**
358	* Initiates the R0 driver for a particular VM instance.
359	*
360	* @returns VBox status code.
361	*
362	* @param pGVM The global (ring-0) VM structure.
363	* @param pVM The cross context VM structure.
364	* @param uSvnRev The SVN revision of the ring-3 part.
365	* @param uBuildType Build type indicator.
366	* @thread EMT(0)
367	*/
368	static int vmmR0InitVM(PGVM pGVM, PVM pVM, uint32_t uSvnRev, uint32_t uBuildType)
369	{
370	VMM_CHECK_SMAP_SETUP();
371	VMM_CHECK_SMAP_CHECK(return VERR_VMM_SMAP_BUT_AC_CLEAR);
372
373	/*
374	* Match the SVN revisions and build type.
375	*/
376	if (uSvnRev != VMMGetSvnRev())
377	{
378	LogRel(("VMMR0InitVM: Revision mismatch, r3=%d r0=%d\n", uSvnRev, VMMGetSvnRev()));
379	SUPR0Printf("VMMR0InitVM: Revision mismatch, r3=%d r0=%d\n", uSvnRev, VMMGetSvnRev());
380	return VERR_VMM_R0_VERSION_MISMATCH;
381	}
382	if (uBuildType != vmmGetBuildType())
383	{
384	LogRel(("VMMR0InitVM: Build type mismatch, r3=%#x r0=%#x\n", uBuildType, vmmGetBuildType()));
385	SUPR0Printf("VMMR0InitVM: Build type mismatch, r3=%#x r0=%#x\n", uBuildType, vmmGetBuildType());
386	return VERR_VMM_R0_VERSION_MISMATCH;
387	}
388
389	int rc = GVMMR0ValidateGVMandVMandEMT(pGVM, pVM, 0 /idCpu/);
390	if (RT_FAILURE(rc))
391	return rc;
392
393	#ifdef LOG_ENABLED
394	/*
395	* Register the EMT R0 logger instance for VCPU 0.
396	*/
397	PVMCPU pVCpu = &pVM->aCpus[0];
398
399	PVMMR0LOGGER pR0Logger = pVCpu->vmm.s.pR0LoggerR0;
400	if (pR0Logger)
401	{
402	# if 0 /* testing of the logger. */
403	LogCom(("vmmR0InitVM: before %p\n", RTLogDefaultInstance()));
404	LogCom(("vmmR0InitVM: pfnFlush=%p actual=%p\n", pR0Logger->Logger.pfnFlush, vmmR0LoggerFlush));
405	LogCom(("vmmR0InitVM: pfnLogger=%p actual=%p\n", pR0Logger->Logger.pfnLogger, vmmR0LoggerWrapper));
406	LogCom(("vmmR0InitVM: offScratch=%d fFlags=%#x fDestFlags=%#x\n", pR0Logger->Logger.offScratch, pR0Logger->Logger.fFlags, pR0Logger->Logger.fDestFlags));
407
408	RTLogSetDefaultInstanceThread(&pR0Logger->Logger, (uintptr_t)pVM->pSession);
409	LogCom(("vmmR0InitVM: after %p reg\n", RTLogDefaultInstance()));
410	RTLogSetDefaultInstanceThread(NULL, pVM->pSession);
411	LogCom(("vmmR0InitVM: after %p dereg\n", RTLogDefaultInstance()));
412
413	pR0Logger->Logger.pfnLogger("hello ring-0 logger\n");
414	LogCom(("vmmR0InitVM: returned successfully from direct logger call.\n"));
415	pR0Logger->Logger.pfnFlush(&pR0Logger->Logger);
416	LogCom(("vmmR0InitVM: returned successfully from direct flush call.\n"));
417
418	RTLogSetDefaultInstanceThread(&pR0Logger->Logger, (uintptr_t)pVM->pSession);
419	LogCom(("vmmR0InitVM: after %p reg2\n", RTLogDefaultInstance()));
420	pR0Logger->Logger.pfnLogger("hello ring-0 logger\n");
421	LogCom(("vmmR0InitVM: returned successfully from direct logger call (2). offScratch=%d\n", pR0Logger->Logger.offScratch));
422	RTLogSetDefaultInstanceThread(NULL, pVM->pSession);
423	LogCom(("vmmR0InitVM: after %p dereg2\n", RTLogDefaultInstance()));
424
425	RTLogLoggerEx(&pR0Logger->Logger, 0, ~0U, "hello ring-0 logger (RTLogLoggerEx)\n");
426	LogCom(("vmmR0InitVM: RTLogLoggerEx returned fine offScratch=%d\n", pR0Logger->Logger.offScratch));
427
428	RTLogSetDefaultInstanceThread(&pR0Logger->Logger, (uintptr_t)pVM->pSession);
429	RTLogPrintf("hello ring-0 logger (RTLogPrintf)\n");
430	LogCom(("vmmR0InitVM: RTLogPrintf returned fine offScratch=%d\n", pR0Logger->Logger.offScratch));
431	# endif
432	Log(("Switching to per-thread logging instance %p (key=%p)\n", &pR0Logger->Logger, pVM->pSession));
433	RTLogSetDefaultInstanceThread(&pR0Logger->Logger, (uintptr_t)pVM->pSession);
434	pR0Logger->fRegistered = true;
435	}
436	#endif /* LOG_ENABLED */
437
438	/*
439	* Check if the host supports high resolution timers or not.
440	*/
441	if ( pVM->vmm.s.fUsePeriodicPreemptionTimers
442	&& !RTTimerCanDoHighResolution())
443	pVM->vmm.s.fUsePeriodicPreemptionTimers = false;
444
445	/*
446	* Initialize the per VM data for GVMM and GMM.
447	*/
448	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
449	rc = GVMMR0InitVM(pGVM);
450	// if (RT_SUCCESS(rc))
451	// rc = GMMR0InitPerVMData(pVM);
452	if (RT_SUCCESS(rc))
453	{
454	/*
455	* Init HM, CPUM and PGM (Darwin only).
456	*/
457	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
458	rc = HMR0InitVM(pVM);
459	if (RT_SUCCESS(rc))
460	VMM_CHECK_SMAP_CHECK2(pVM, rc = VERR_VMM_RING0_ASSERTION); /* CPUR0InitVM will otherwise panic the host */
461	if (RT_SUCCESS(rc))
462	{
463	rc = CPUMR0InitVM(pVM);
464	if (RT_SUCCESS(rc))
465	{
466	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
467	#ifdef VBOX_WITH_2X_4GB_ADDR_SPACE
468	rc = PGMR0DynMapInitVM(pVM);
469	#endif
470	if (RT_SUCCESS(rc))
471	{
472	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
473	rc = EMR0InitVM(pGVM, pVM);
474	if (RT_SUCCESS(rc))
475	{
476	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
477	#ifdef VBOX_WITH_PCI_PASSTHROUGH
478	rc = PciRawR0InitVM(pGVM, pVM);
479	#endif
480	if (RT_SUCCESS(rc))
481	{
482	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
483	rc = GIMR0InitVM(pVM);
484	if (RT_SUCCESS(rc))
485	{
486	VMM_CHECK_SMAP_CHECK2(pVM, rc = VERR_VMM_RING0_ASSERTION);
487	if (RT_SUCCESS(rc))
488	{
489	GVMMR0DoneInitVM(pGVM);
490
491	/*
492	* Collect a bit of info for the VM release log.
493	*/
494	pVM->vmm.s.fIsPreemptPendingApiTrusty = RTThreadPreemptIsPendingTrusty();
495	pVM->vmm.s.fIsPreemptPossible = RTThreadPreemptIsPossible();;
496
497	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
498	return rc;
499	}
500
501	/* bail out*/
502	GIMR0TermVM(pVM);
503	}
504	#ifdef VBOX_WITH_PCI_PASSTHROUGH
505	PciRawR0TermVM(pGVM, pVM);
506	#endif
507	}
508	}
509	}
510	}
511	HMR0TermVM(pVM);
512	}
513	}
514
515	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pVM->pSession);
516	return rc;
517	}
518
519
520	/**
521	* Does EMT specific VM initialization.
522	*
523	* @returns VBox status code.
524	* @param pGVM The ring-0 VM structure.
525	* @param pVM The cross context VM structure.
526	* @param idCpu The EMT that's calling.
527	*/
528	static int vmmR0InitVMEmt(PGVM pGVM, PVM pVM, VMCPUID idCpu)
529	{
530	/* Paranoia (caller checked these already). */
531	AssertReturn(idCpu < pGVM->cCpus, VERR_INVALID_CPU_ID);
532	AssertReturn(pGVM->aCpus[idCpu].hEMT == RTThreadNativeSelf(), VERR_INVALID_CPU_ID);
533
534	#ifdef LOG_ENABLED
535	/*
536	* Registration of ring 0 loggers.
537	*/
538	PVMCPU pVCpu = &pVM->aCpus[idCpu];
539	PVMMR0LOGGER pR0Logger = pVCpu->vmm.s.pR0LoggerR0;
540	if ( pR0Logger
541	&& !pR0Logger->fRegistered)
542	{
543	RTLogSetDefaultInstanceThread(&pR0Logger->Logger, (uintptr_t)pVM->pSession);
544	pR0Logger->fRegistered = true;
545	}
546	#endif
547	RT_NOREF(pVM);
548
549	return VINF_SUCCESS;
550	}
551
552
553
554	/**
555	* Terminates the R0 bits for a particular VM instance.
556	*
557	* This is normally called by ring-3 as part of the VM termination process, but
558	* may alternatively be called during the support driver session cleanup when
559	* the VM object is destroyed (see GVMM).
560	*
561	* @returns VBox status code.
562	*
563	* @param pGVM The global (ring-0) VM structure.
564	* @param pVM The cross context VM structure.
565	* @param idCpu Set to 0 if EMT(0) or NIL_VMCPUID if session cleanup
566	* thread.
567	* @thread EMT(0) or session clean up thread.
568	*/
569	VMMR0_INT_DECL(int) VMMR0TermVM(PGVM pGVM, PVM pVM, VMCPUID idCpu)
570	{
571	/*
572	* Check EMT(0) claim if we're called from userland.
573	*/
574	if (idCpu != NIL_VMCPUID)
575	{
576	AssertReturn(idCpu == 0, VERR_INVALID_CPU_ID);
577	int rc = GVMMR0ValidateGVMandVMandEMT(pGVM, pVM, idCpu);
578	if (RT_FAILURE(rc))
579	return rc;
580	}
581
582	#ifdef VBOX_WITH_PCI_PASSTHROUGH
583	PciRawR0TermVM(pGVM, pVM);
584	#endif
585
586	/*
587	* Tell GVMM what we're up to and check that we only do this once.
588	*/
589	if (GVMMR0DoingTermVM(pGVM))
590	{
591	GIMR0TermVM(pVM);
592
593	/** @todo I wish to call PGMR0PhysFlushHandyPages(pVM, &pVM->aCpus[idCpu])
594	* here to make sure we don't leak any shared pages if we crash... */
595	#ifdef VBOX_WITH_2X_4GB_ADDR_SPACE
596	PGMR0DynMapTermVM(pVM);
597	#endif
598	HMR0TermVM(pVM);
599	}
600
601	/*
602	* Deregister the logger.
603	*/
604	RTLogSetDefaultInstanceThread(NULL, (uintptr_t)pVM->pSession);
605	return VINF_SUCCESS;
606	}
607
608
609	/**
610	* An interrupt or unhalt force flag is set, deal with it.
611	*
612	* @returns VINF_SUCCESS (or VINF_EM_HALT).
613	* @param pVCpu The cross context virtual CPU structure.
614	* @param uMWait Result from EMMonitorWaitIsActive().
615	* @param enmInterruptibility Guest CPU interruptbility level.
616	*/
617	static int vmmR0DoHaltInterrupt(PVMCPU pVCpu, unsigned uMWait, CPUMINTERRUPTIBILITY enmInterruptibility)
618	{
619	Assert(!TRPMHasTrap(pVCpu));
620	Assert( enmInterruptibility > CPUMINTERRUPTIBILITY_INVALID
621	&& enmInterruptibility < CPUMINTERRUPTIBILITY_END);
622
623	/*
624	* Pending interrupts w/o any SMIs or NMIs? That the usual case.
625	*/
626	if ( VMCPU_FF_IS_ANY_SET(pVCpu, VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC)
627	&& !VMCPU_FF_IS_ANY_SET(pVCpu, VMCPU_FF_INTERRUPT_SMI \| VMCPU_FF_INTERRUPT_NMI))
628	{
629	if (enmInterruptibility <= CPUMINTERRUPTIBILITY_UNRESTRAINED)
630	{
631	uint8_t u8Interrupt = 0;
632	int rc = PDMGetInterrupt(pVCpu, &u8Interrupt);
633	Log(("vmmR0DoHaltInterrupt: CPU%d u8Interrupt=%d (%#x) rc=%Rrc\n", pVCpu->idCpu, u8Interrupt, u8Interrupt, rc));
634	if (RT_SUCCESS(rc))
635	{
636	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_UNHALT);
637
638	rc = TRPMAssertTrap(pVCpu, u8Interrupt, TRPM_HARDWARE_INT);
639	AssertRCSuccess(rc);
640	STAM_REL_COUNTER_INC(&pVCpu->vmm.s.StatR0HaltExec);
641	return rc;
642	}
643	}
644	}
645	/*
646	* SMI is not implemented yet, at least not here.
647	*/
648	else if (VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_INTERRUPT_SMI))
649	{
650	return VINF_EM_HALT;
651	}
652	/*
653	* NMI.
654	*/
655	else if (VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_INTERRUPT_NMI))
656	{
657	if (enmInterruptibility < CPUMINTERRUPTIBILITY_NMI_INHIBIT)
658	{
659	/** @todo later. */
660	return VINF_EM_HALT;
661	}
662	}
663	/*
664	* Nested-guest virtual interrupt.
665	*/
666	else if (VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_INTERRUPT_NESTED_GUEST))
667	{
668	if (enmInterruptibility < CPUMINTERRUPTIBILITY_VIRT_INT_DISABLED)
669	{
670	/** @todo NSTVMX: NSTSVM: Remember, we might have to check and perform VM-exits
671	* here before injecting the virtual interrupt. See emR3ForcedActions
672	* for details. */
673	return VINF_EM_HALT;
674	}
675	}
676
677	if (VMCPU_FF_TEST_AND_CLEAR(pVCpu, VMCPU_FF_UNHALT))
678	{
679	STAM_REL_COUNTER_INC(&pVCpu->vmm.s.StatR0HaltExec);
680	return VINF_SUCCESS;
681	}
682	if (uMWait > 1)
683	{
684	STAM_REL_COUNTER_INC(&pVCpu->vmm.s.StatR0HaltExec);
685	return VINF_SUCCESS;
686	}
687
688	return VINF_EM_HALT;
689	}
690
691
692	/**
693	* This does one round of vmR3HaltGlobal1Halt().
694	*
695	* The rational here is that we'll reduce latency in interrupt situations if we
696	* don't go to ring-3 immediately on a VINF_EM_HALT (guest executed HLT or
697	* MWAIT), but do one round of blocking here instead and hope the interrupt is
698	* raised in the meanwhile.
699	*
700	* If we go to ring-3 we'll quit the inner HM/NEM loop in EM and end up in the
701	* outer loop, which will then call VMR3WaitHalted() and that in turn will do a
702	* ring-0 call (unless we're too close to a timer event). When the interrupt
703	* wakes us up, we'll return from ring-0 and EM will by instinct do a
704	* rescheduling (because of raw-mode) before it resumes the HM/NEM loop and gets
705	* back to VMMR0EntryFast().
706	*
707	* @returns VINF_SUCCESS or VINF_EM_HALT.
708	* @param pGVM The ring-0 VM structure.
709	* @param pVM The cross context VM structure.
710	* @param pGVCpu The ring-0 virtual CPU structure.
711	* @param pVCpu The cross context virtual CPU structure.
712	*
713	* @todo r=bird: All the blocking/waiting and EMT managment should move out of
714	* the VM module, probably to VMM. Then this would be more weird wrt
715	* parameters and statistics.
716	*/
717	static int vmmR0DoHalt(PGVM pGVM, PVM pVM, PGVMCPU pGVCpu, PVMCPU pVCpu)
718	{
719	Assert(pVCpu == pGVCpu->pVCpu);
720
721	/*
722	* Do spin stat historization.
723	*/
724	if (++pVCpu->vmm.s.cR0Halts & 0xff)
725	{ /* likely */ }
726	else if (pVCpu->vmm.s.cR0HaltsSucceeded > pVCpu->vmm.s.cR0HaltsToRing3)
727	{
728	pVCpu->vmm.s.cR0HaltsSucceeded = 2;
729	pVCpu->vmm.s.cR0HaltsToRing3 = 0;
730	}
731	else
732	{
733	pVCpu->vmm.s.cR0HaltsSucceeded = 0;
734	pVCpu->vmm.s.cR0HaltsToRing3 = 2;
735	}
736
737	/*
738	* Flags that makes us go to ring-3.
739	*/
740	uint32_t const fVmFFs = VM_FF_TM_VIRTUAL_SYNC \| VM_FF_PDM_QUEUES \| VM_FF_PDM_DMA
741	\| VM_FF_DBGF \| VM_FF_REQUEST \| VM_FF_CHECK_VM_STATE
742	\| VM_FF_RESET \| VM_FF_EMT_RENDEZVOUS \| VM_FF_PGM_NEED_HANDY_PAGES
743	\| VM_FF_PGM_NO_MEMORY \| VM_FF_REM_HANDLER_NOTIFY \| VM_FF_DEBUG_SUSPEND;
744	uint64_t const fCpuFFs = VMCPU_FF_TIMER \| VMCPU_FF_PDM_CRITSECT \| VMCPU_FF_IEM
745	\| VMCPU_FF_REQUEST \| VMCPU_FF_DBGF \| VMCPU_FF_HM_UPDATE_CR3
746	\| VMCPU_FF_HM_UPDATE_PAE_PDPES \| VMCPU_FF_PGM_SYNC_CR3 \| VMCPU_FF_PGM_SYNC_CR3_NON_GLOBAL
747	\| VMCPU_FF_TO_R3 \| VMCPU_FF_IOM
748	#ifdef VBOX_WITH_RAW_MODE
749	\| VMCPU_FF_TRPM_SYNC_IDT \| VMCPU_FF_SELM_SYNC_TSS \| VMCPU_FF_SELM_SYNC_GDT
750	\| VMCPU_FF_SELM_SYNC_LDT \| VMCPU_FF_CSAM_SCAN_PAGE \| VMCPU_FF_CSAM_PENDING_ACTION
751	\| VMCPU_FF_CPUM
752	#endif
753	;
754
755	/*
756	* Check preconditions.
757	*/
758	unsigned const uMWait = EMMonitorWaitIsActive(pVCpu);
759	CPUMINTERRUPTIBILITY const enmInterruptibility = CPUMGetGuestInterruptibility(pVCpu);
760	if ( pVCpu->vmm.s.fMayHaltInRing0
761	&& !TRPMHasTrap(pVCpu)
762	&& ( enmInterruptibility == CPUMINTERRUPTIBILITY_UNRESTRAINED
763	\|\| uMWait > 1))
764	{
765	if ( !VM_FF_IS_ANY_SET(pVM, fVmFFs)
766	&& !VMCPU_FF_IS_ANY_SET(pVCpu, fCpuFFs))
767	{
768	/*
769	* Interrupts pending already?
770	*/
771	if (VMCPU_FF_TEST_AND_CLEAR(pVCpu, VMCPU_FF_UPDATE_APIC))
772	APICUpdatePendingInterrupts(pVCpu);
773
774	/*
775	* Flags that wake up from the halted state.
776	*/
777	uint64_t const fIntMask = VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC \| VMCPU_FF_INTERRUPT_NESTED_GUEST
778	\| VMCPU_FF_INTERRUPT_NMI \| VMCPU_FF_INTERRUPT_SMI \| VMCPU_FF_UNHALT;
779
780	if (VMCPU_FF_IS_ANY_SET(pVCpu, fIntMask))
781	return vmmR0DoHaltInterrupt(pVCpu, uMWait, enmInterruptibility);
782	ASMNopPause();
783
784	/*
785	* Check out how long till the next timer event.
786	*/
787	uint64_t u64Delta;
788	uint64_t u64GipTime = TMTimerPollGIP(pVM, pVCpu, &u64Delta);
789
790	if ( !VM_FF_IS_ANY_SET(pVM, fVmFFs)
791	&& !VMCPU_FF_IS_ANY_SET(pVCpu, fCpuFFs))
792	{
793	if (VMCPU_FF_TEST_AND_CLEAR(pVCpu, VMCPU_FF_UPDATE_APIC))
794	APICUpdatePendingInterrupts(pVCpu);
795
796	if (VMCPU_FF_IS_ANY_SET(pVCpu, fIntMask))
797	return vmmR0DoHaltInterrupt(pVCpu, uMWait, enmInterruptibility);
798
799	/*
800	* Wait if there is enough time to the next timer event.
801	*/
802	if (u64Delta >= pVCpu->vmm.s.cNsSpinBlockThreshold)
803	{
804	/* If there are few other CPU cores around, we will procrastinate a
805	little before going to sleep, hoping for some device raising an
806	interrupt or similar. Though, the best thing here would be to
807	dynamically adjust the spin count according to its usfulness or
808	something... */
809	if ( pVCpu->vmm.s.cR0HaltsSucceeded > pVCpu->vmm.s.cR0HaltsToRing3
810	&& RTMpGetOnlineCount() >= 4)
811	{
812	/** @todo Figure out how we can skip this if it hasn't help recently...
813	* @bugref{9172#c12} */
814	uint32_t cSpinLoops = 42;
815	while (cSpinLoops-- > 0)
816	{
817	ASMNopPause();
818	if (VMCPU_FF_TEST_AND_CLEAR(pVCpu, VMCPU_FF_UPDATE_APIC))
819	APICUpdatePendingInterrupts(pVCpu);
820	ASMNopPause();
821	if (VM_FF_IS_ANY_SET(pVM, fVmFFs))
822	{
823	STAM_REL_COUNTER_INC(&pVCpu->vmm.s.StatR0HaltToR3FromSpin);
824	return VINF_EM_HALT;
825	}
826	ASMNopPause();
827	if (VMCPU_FF_IS_ANY_SET(pVCpu, fCpuFFs))
828	{
829	STAM_REL_COUNTER_INC(&pVCpu->vmm.s.StatR0HaltToR3FromSpin);
830	return VINF_EM_HALT;
831	}
832	ASMNopPause();
833	if (VMCPU_FF_IS_ANY_SET(pVCpu, fIntMask))
834	{
835	STAM_REL_COUNTER_INC(&pVCpu->vmm.s.StatR0HaltExecFromSpin);
836	return vmmR0DoHaltInterrupt(pVCpu, uMWait, enmInterruptibility);
837	}
838	ASMNopPause();
839	}
840	}
841
842	/* Block. We have to set the state to VMCPUSTATE_STARTED_HALTED here so ring-3
843	knows when to notify us (cannot access VMINTUSERPERVMCPU::fWait from here). */
844	VMCPU_CMPXCHG_STATE(pVCpu, VMCPUSTATE_STARTED_HALTED, VMCPUSTATE_STARTED);
845	uint64_t const u64StartSchedHalt = RTTimeNanoTS();
846	int rc = GVMMR0SchedHalt(pGVM, pVM, pGVCpu, u64GipTime);
847	uint64_t const u64EndSchedHalt = RTTimeNanoTS();
848	uint64_t const cNsElapsedSchedHalt = u64EndSchedHalt - u64StartSchedHalt;
849	VMCPU_CMPXCHG_STATE(pVCpu, VMCPUSTATE_STARTED, VMCPUSTATE_STARTED_HALTED);
850	STAM_REL_PROFILE_ADD_PERIOD(&pVCpu->vmm.s.StatR0HaltBlock, cNsElapsedSchedHalt);
851	if ( rc == VINF_SUCCESS
852	\|\| rc == VERR_INTERRUPTED)
853
854	{
855	/* Keep some stats like ring-3 does. */
856	int64_t const cNsOverslept = u64EndSchedHalt - u64GipTime;
857	if (cNsOverslept > 50000)
858	STAM_REL_PROFILE_ADD_PERIOD(&pVCpu->vmm.s.StatR0HaltBlockOverslept, cNsOverslept);
859	else if (cNsOverslept < -50000)
860	STAM_REL_PROFILE_ADD_PERIOD(&pVCpu->vmm.s.StatR0HaltBlockInsomnia, cNsElapsedSchedHalt);
861	else
862	STAM_REL_PROFILE_ADD_PERIOD(&pVCpu->vmm.s.StatR0HaltBlockOnTime, cNsElapsedSchedHalt);
863
864	/*
865	* Recheck whether we can resume execution or have to go to ring-3.
866	*/
867	if ( !VM_FF_IS_ANY_SET(pVM, fVmFFs)
868	&& !VMCPU_FF_IS_ANY_SET(pVCpu, fCpuFFs))
869	{
870	if (VMCPU_FF_TEST_AND_CLEAR(pVCpu, VMCPU_FF_UPDATE_APIC))
871	APICUpdatePendingInterrupts(pVCpu);
872	if (VMCPU_FF_IS_ANY_SET(pVCpu, fIntMask))
873	{
874	STAM_REL_COUNTER_INC(&pVCpu->vmm.s.StatR0HaltExecFromBlock);
875	return vmmR0DoHaltInterrupt(pVCpu, uMWait, enmInterruptibility);
876	}
877	}
878	}
879	}
880	}
881	}
882	}
883	return VINF_EM_HALT;
884	}
885
886
887	/**
888	* VMM ring-0 thread-context callback.
889	*
890	* This does common HM state updating and calls the HM-specific thread-context
891	* callback.
892	*
893	* @param enmEvent The thread-context event.
894	* @param pvUser Opaque pointer to the VMCPU.
895	*
896	* @thread EMT(pvUser)
897	*/
898	static DECLCALLBACK(void) vmmR0ThreadCtxCallback(RTTHREADCTXEVENT enmEvent, void *pvUser)
899	{
900	PVMCPU pVCpu = (PVMCPU)pvUser;
901
902	switch (enmEvent)
903	{
904	case RTTHREADCTXEVENT_IN:
905	{
906	/*
907	* Linux may call us with preemption enabled (really!) but technically we
908	* cannot get preempted here, otherwise we end up in an infinite recursion
909	* scenario (i.e. preempted in resume hook -> preempt hook -> resume hook...
910	* ad infinitum). Let's just disable preemption for now...
911	*/
912	/** @todo r=bird: I don't believe the above. The linux code is clearly enabling
913	* preemption after doing the callout (one or two functions up the
914	* call chain). */
915	/** @todo r=ramshankar: See @bugref{5313#c30}. */
916	RTTHREADPREEMPTSTATE ParanoidPreemptState = RTTHREADPREEMPTSTATE_INITIALIZER;
917	RTThreadPreemptDisable(&ParanoidPreemptState);
918
919	/* We need to update the VCPU <-> host CPU mapping. */
920	RTCPUID idHostCpu;
921	uint32_t iHostCpuSet = RTMpCurSetIndexAndId(&idHostCpu);
922	pVCpu->iHostCpuSet = iHostCpuSet;
923	ASMAtomicWriteU32(&pVCpu->idHostCpu, idHostCpu);
924
925	/* In the very unlikely event that the GIP delta for the CPU we're
926	rescheduled needs calculating, try force a return to ring-3.
927	We unfortunately cannot do the measurements right here. */
928	if (RT_UNLIKELY(SUPIsTscDeltaAvailableForCpuSetIndex(iHostCpuSet)))
929	VMCPU_FF_SET(pVCpu, VMCPU_FF_TO_R3);
930
931	/* Invoke the HM-specific thread-context callback. */
932	HMR0ThreadCtxCallback(enmEvent, pvUser);
933
934	/* Restore preemption. */
935	RTThreadPreemptRestore(&ParanoidPreemptState);
936	break;
937	}
938
939	case RTTHREADCTXEVENT_OUT:
940	{
941	/* Invoke the HM-specific thread-context callback. */
942	HMR0ThreadCtxCallback(enmEvent, pvUser);
943
944	/*
945	* Sigh. See VMMGetCpu() used by VMCPU_ASSERT_EMT(). We cannot let several VCPUs
946	* have the same host CPU associated with it.
947	*/
948	pVCpu->iHostCpuSet = UINT32_MAX;
949	ASMAtomicWriteU32(&pVCpu->idHostCpu, NIL_RTCPUID);
950	break;
951	}
952
953	default:
954	/* Invoke the HM-specific thread-context callback. */
955	HMR0ThreadCtxCallback(enmEvent, pvUser);
956	break;
957	}
958	}
959
960
961	/**
962	* Creates thread switching hook for the current EMT thread.
963	*
964	* This is called by GVMMR0CreateVM and GVMMR0RegisterVCpu. If the host
965	* platform does not implement switcher hooks, no hooks will be create and the
966	* member set to NIL_RTTHREADCTXHOOK.
967	*
968	* @returns VBox status code.
969	* @param pVCpu The cross context virtual CPU structure.
970	* @thread EMT(pVCpu)
971	*/
972	VMMR0_INT_DECL(int) VMMR0ThreadCtxHookCreateForEmt(PVMCPU pVCpu)
973	{
974	VMCPU_ASSERT_EMT(pVCpu);
975	Assert(pVCpu->vmm.s.hCtxHook == NIL_RTTHREADCTXHOOK);
976
977	#if 1 /* To disable this stuff change to zero. */
978	int rc = RTThreadCtxHookCreate(&pVCpu->vmm.s.hCtxHook, 0, vmmR0ThreadCtxCallback, pVCpu);
979	if (RT_SUCCESS(rc))
980	return rc;
981	#else
982	RT_NOREF(vmmR0ThreadCtxCallback);
983	int rc = VERR_NOT_SUPPORTED;
984	#endif
985
986	pVCpu->vmm.s.hCtxHook = NIL_RTTHREADCTXHOOK;
987	if (rc == VERR_NOT_SUPPORTED)
988	return VINF_SUCCESS;
989
990	LogRelMax(32, ("RTThreadCtxHookCreate failed! rc=%Rrc pVCpu=%p idCpu=%RU32\n", rc, pVCpu, pVCpu->idCpu));
991	return VINF_SUCCESS; /* Just ignore it, we can live without context hooks. */
992	}
993
994
995	/**
996	* Destroys the thread switching hook for the specified VCPU.
997	*
998	* @param pVCpu The cross context virtual CPU structure.
999	* @remarks Can be called from any thread.
1000	*/
1001	VMMR0_INT_DECL(void) VMMR0ThreadCtxHookDestroyForEmt(PVMCPU pVCpu)
1002	{
1003	int rc = RTThreadCtxHookDestroy(pVCpu->vmm.s.hCtxHook);
1004	AssertRC(rc);
1005	pVCpu->vmm.s.hCtxHook = NIL_RTTHREADCTXHOOK;
1006	}
1007
1008
1009	/**
1010	* Disables the thread switching hook for this VCPU (if we got one).
1011	*
1012	* @param pVCpu The cross context virtual CPU structure.
1013	* @thread EMT(pVCpu)
1014	*
1015	* @remarks This also clears VMCPU::idHostCpu, so the mapping is invalid after
1016	* this call. This means you have to be careful with what you do!
1017	*/
1018	VMMR0_INT_DECL(void) VMMR0ThreadCtxHookDisable(PVMCPU pVCpu)
1019	{
1020	/*
1021	* Clear the VCPU <-> host CPU mapping as we've left HM context.
1022	* @bugref{7726#c19} explains the need for this trick:
1023	*
1024	* hmR0VmxCallRing3Callback/hmR0SvmCallRing3Callback &
1025	* hmR0VmxLeaveSession/hmR0SvmLeaveSession disables context hooks during
1026	* longjmp & normal return to ring-3, which opens a window where we may be
1027	* rescheduled without changing VMCPUID::idHostCpu and cause confusion if
1028	* the CPU starts executing a different EMT. Both functions first disables
1029	* preemption and then calls HMR0LeaveCpu which invalids idHostCpu, leaving
1030	* an opening for getting preempted.
1031	*/
1032	/** @todo Make HM not need this API! Then we could leave the hooks enabled
1033	* all the time. */
1034	/** @todo move this into the context hook disabling if(). */
1035	ASMAtomicWriteU32(&pVCpu->idHostCpu, NIL_RTCPUID);
1036
1037	/*
1038	* Disable the context hook, if we got one.
1039	*/
1040	if (pVCpu->vmm.s.hCtxHook != NIL_RTTHREADCTXHOOK)
1041	{
1042	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1043	int rc = RTThreadCtxHookDisable(pVCpu->vmm.s.hCtxHook);
1044	AssertRC(rc);
1045	}
1046	}
1047
1048
1049	/**
1050	* Internal version of VMMR0ThreadCtxHooksAreRegistered.
1051	*
1052	* @returns true if registered, false otherwise.
1053	* @param pVCpu The cross context virtual CPU structure.
1054	*/
1055	DECLINLINE(bool) vmmR0ThreadCtxHookIsEnabled(PVMCPU pVCpu)
1056	{
1057	return RTThreadCtxHookIsEnabled(pVCpu->vmm.s.hCtxHook);
1058	}
1059
1060
1061	/**
1062	* Whether thread-context hooks are registered for this VCPU.
1063	*
1064	* @returns true if registered, false otherwise.
1065	* @param pVCpu The cross context virtual CPU structure.
1066	*/
1067	VMMR0_INT_DECL(bool) VMMR0ThreadCtxHookIsEnabled(PVMCPU pVCpu)
1068	{
1069	return vmmR0ThreadCtxHookIsEnabled(pVCpu);
1070	}
1071
1072
1073	#ifdef VBOX_WITH_STATISTICS
1074	/**
1075	* Record return code statistics
1076	* @param pVM The cross context VM structure.
1077	* @param pVCpu The cross context virtual CPU structure.
1078	* @param rc The status code.
1079	*/
1080	static void vmmR0RecordRC(PVM pVM, PVMCPU pVCpu, int rc)
1081	{
1082	/*
1083	* Collect statistics.
1084	*/
1085	switch (rc)
1086	{
1087	case VINF_SUCCESS:
1088	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetNormal);
1089	break;
1090	case VINF_EM_RAW_INTERRUPT:
1091	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetInterrupt);
1092	break;
1093	case VINF_EM_RAW_INTERRUPT_HYPER:
1094	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetInterruptHyper);
1095	break;
1096	case VINF_EM_RAW_GUEST_TRAP:
1097	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetGuestTrap);
1098	break;
1099	case VINF_EM_RAW_RING_SWITCH:
1100	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetRingSwitch);
1101	break;
1102	case VINF_EM_RAW_RING_SWITCH_INT:
1103	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetRingSwitchInt);
1104	break;
1105	case VINF_EM_RAW_STALE_SELECTOR:
1106	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetStaleSelector);
1107	break;
1108	case VINF_EM_RAW_IRET_TRAP:
1109	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetIRETTrap);
1110	break;
1111	case VINF_IOM_R3_IOPORT_READ:
1112	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetIORead);
1113	break;
1114	case VINF_IOM_R3_IOPORT_WRITE:
1115	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetIOWrite);
1116	break;
1117	case VINF_IOM_R3_IOPORT_COMMIT_WRITE:
1118	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetIOCommitWrite);
1119	break;
1120	case VINF_IOM_R3_MMIO_READ:
1121	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetMMIORead);
1122	break;
1123	case VINF_IOM_R3_MMIO_WRITE:
1124	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetMMIOWrite);
1125	break;
1126	case VINF_IOM_R3_MMIO_COMMIT_WRITE:
1127	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetMMIOCommitWrite);
1128	break;
1129	case VINF_IOM_R3_MMIO_READ_WRITE:
1130	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetMMIOReadWrite);
1131	break;
1132	case VINF_PATM_HC_MMIO_PATCH_READ:
1133	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetMMIOPatchRead);
1134	break;
1135	case VINF_PATM_HC_MMIO_PATCH_WRITE:
1136	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetMMIOPatchWrite);
1137	break;
1138	case VINF_CPUM_R3_MSR_READ:
1139	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetMSRRead);
1140	break;
1141	case VINF_CPUM_R3_MSR_WRITE:
1142	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetMSRWrite);
1143	break;
1144	case VINF_EM_RAW_EMULATE_INSTR:
1145	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetEmulate);
1146	break;
1147	case VINF_PATCH_EMULATE_INSTR:
1148	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetPatchEmulate);
1149	break;
1150	case VINF_EM_RAW_EMULATE_INSTR_LDT_FAULT:
1151	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetLDTFault);
1152	break;
1153	case VINF_EM_RAW_EMULATE_INSTR_GDT_FAULT:
1154	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetGDTFault);
1155	break;
1156	case VINF_EM_RAW_EMULATE_INSTR_IDT_FAULT:
1157	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetIDTFault);
1158	break;
1159	case VINF_EM_RAW_EMULATE_INSTR_TSS_FAULT:
1160	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetTSSFault);
1161	break;
1162	case VINF_CSAM_PENDING_ACTION:
1163	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetCSAMTask);
1164	break;
1165	case VINF_PGM_SYNC_CR3:
1166	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetSyncCR3);
1167	break;
1168	case VINF_PATM_PATCH_INT3:
1169	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetPatchInt3);
1170	break;
1171	case VINF_PATM_PATCH_TRAP_PF:
1172	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetPatchPF);
1173	break;
1174	case VINF_PATM_PATCH_TRAP_GP:
1175	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetPatchGP);
1176	break;
1177	case VINF_PATM_PENDING_IRQ_AFTER_IRET:
1178	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetPatchIretIRQ);
1179	break;
1180	case VINF_EM_RESCHEDULE_REM:
1181	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetRescheduleREM);
1182	break;
1183	case VINF_EM_RAW_TO_R3:
1184	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetToR3Total);
1185	if (VM_FF_IS_SET(pVM, VM_FF_TM_VIRTUAL_SYNC))
1186	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetToR3TMVirt);
1187	else if (VM_FF_IS_SET(pVM, VM_FF_PGM_NEED_HANDY_PAGES))
1188	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetToR3HandyPages);
1189	else if (VM_FF_IS_SET(pVM, VM_FF_PDM_QUEUES))
1190	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetToR3PDMQueues);
1191	else if (VM_FF_IS_SET(pVM, VM_FF_EMT_RENDEZVOUS))
1192	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetToR3Rendezvous);
1193	else if (VM_FF_IS_SET(pVM, VM_FF_PDM_DMA))
1194	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetToR3DMA);
1195	else if (VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_TIMER))
1196	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetToR3Timer);
1197	else if (VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_PDM_CRITSECT))
1198	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetToR3CritSect);
1199	else if (VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_TO_R3))
1200	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetToR3FF);
1201	else if (VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_IEM))
1202	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetToR3Iem);
1203	else if (VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_IOM))
1204	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetToR3Iom);
1205	else
1206	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetToR3Unknown);
1207	break;
1208
1209	case VINF_EM_RAW_TIMER_PENDING:
1210	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetTimerPending);
1211	break;
1212	case VINF_EM_RAW_INTERRUPT_PENDING:
1213	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetInterruptPending);
1214	break;
1215	case VINF_VMM_CALL_HOST:
1216	switch (pVCpu->vmm.s.enmCallRing3Operation)
1217	{
1218	case VMMCALLRING3_PDM_CRIT_SECT_ENTER:
1219	STAM_COUNTER_INC(&pVM->vmm.s.StatRZCallPDMCritSectEnter);
1220	break;
1221	case VMMCALLRING3_PDM_LOCK:
1222	STAM_COUNTER_INC(&pVM->vmm.s.StatRZCallPDMLock);
1223	break;
1224	case VMMCALLRING3_PGM_POOL_GROW:
1225	STAM_COUNTER_INC(&pVM->vmm.s.StatRZCallPGMPoolGrow);
1226	break;
1227	case VMMCALLRING3_PGM_LOCK:
1228	STAM_COUNTER_INC(&pVM->vmm.s.StatRZCallPGMLock);
1229	break;
1230	case VMMCALLRING3_PGM_MAP_CHUNK:
1231	STAM_COUNTER_INC(&pVM->vmm.s.StatRZCallPGMMapChunk);
1232	break;
1233	case VMMCALLRING3_PGM_ALLOCATE_HANDY_PAGES:
1234	STAM_COUNTER_INC(&pVM->vmm.s.StatRZCallPGMAllocHandy);
1235	break;
1236	case VMMCALLRING3_REM_REPLAY_HANDLER_NOTIFICATIONS:
1237	STAM_COUNTER_INC(&pVM->vmm.s.StatRZCallRemReplay);
1238	break;
1239	case VMMCALLRING3_VMM_LOGGER_FLUSH:
1240	STAM_COUNTER_INC(&pVM->vmm.s.StatRZCallLogFlush);
1241	break;
1242	case VMMCALLRING3_VM_SET_ERROR:
1243	STAM_COUNTER_INC(&pVM->vmm.s.StatRZCallVMSetError);
1244	break;
1245	case VMMCALLRING3_VM_SET_RUNTIME_ERROR:
1246	STAM_COUNTER_INC(&pVM->vmm.s.StatRZCallVMSetRuntimeError);
1247	break;
1248	case VMMCALLRING3_VM_R0_ASSERTION:
1249	default:
1250	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetCallRing3);
1251	break;
1252	}
1253	break;
1254	case VINF_PATM_DUPLICATE_FUNCTION:
1255	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetPATMDuplicateFn);
1256	break;
1257	case VINF_PGM_CHANGE_MODE:
1258	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetPGMChangeMode);
1259	break;
1260	case VINF_PGM_POOL_FLUSH_PENDING:
1261	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetPGMFlushPending);
1262	break;
1263	case VINF_EM_PENDING_REQUEST:
1264	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetPendingRequest);
1265	break;
1266	case VINF_EM_HM_PATCH_TPR_INSTR:
1267	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetPatchTPR);
1268	break;
1269	default:
1270	STAM_COUNTER_INC(&pVM->vmm.s.StatRZRetMisc);
1271	break;
1272	}
1273	}
1274	#endif /* VBOX_WITH_STATISTICS */
1275
1276
1277	/**
1278	* The Ring 0 entry point, called by the fast-ioctl path.
1279	*
1280	* @param pGVM The global (ring-0) VM structure.
1281	* @param pVM The cross context VM structure.
1282	* The return code is stored in pVM->vmm.s.iLastGZRc.
1283	* @param idCpu The Virtual CPU ID of the calling EMT.
1284	* @param enmOperation Which operation to execute.
1285	* @remarks Assume called with interrupts _enabled_.
1286	*/
1287	VMMR0DECL(void) VMMR0EntryFast(PGVM pGVM, PVM pVM, VMCPUID idCpu, VMMR0OPERATION enmOperation)
1288	{
1289	/*
1290	* Validation.
1291	*/
1292	if ( idCpu < pGVM->cCpus
1293	&& pGVM->cCpus == pVM->cCpus)
1294	{ /likely/ }
1295	else
1296	{
1297	SUPR0Printf("VMMR0EntryFast: Bad idCpu=%#x cCpus=%#x/%#x\n", idCpu, pGVM->cCpus, pVM->cCpus);
1298	return;
1299	}
1300
1301	PGVMCPU pGVCpu = &pGVM->aCpus[idCpu];
1302	PVMCPU pVCpu = &pVM->aCpus[idCpu];
1303	RTNATIVETHREAD const hNativeThread = RTThreadNativeSelf();
1304	if (RT_LIKELY( pGVCpu->hEMT == hNativeThread
1305	&& pVCpu->hNativeThreadR0 == hNativeThread))
1306	{ /* likely */ }
1307	else
1308	{
1309	SUPR0Printf("VMMR0EntryFast: Bad thread idCpu=%#x hNativeSelf=%p pGVCpu->hEmt=%p pVCpu->hNativeThreadR0=%p\n",
1310	idCpu, hNativeThread, pGVCpu->hEMT, pVCpu->hNativeThreadR0);
1311	return;
1312	}
1313
1314	/*
1315	* SMAP fun.
1316	*/
1317	VMM_CHECK_SMAP_SETUP();
1318	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1319
1320	/*
1321	* Perform requested operation.
1322	*/
1323	switch (enmOperation)
1324	{
1325	/*
1326	* Switch to GC and run guest raw mode code.
1327	* Disable interrupts before doing the world switch.
1328	*/
1329	case VMMR0_DO_RAW_RUN:
1330	{
1331	#ifdef VBOX_WITH_RAW_MODE
1332	# ifndef VBOX_WITH_2X_4GB_ADDR_SPACE_IN_R0
1333	/* Some safety precautions first. */
1334	if (RT_UNLIKELY(!PGMGetHyperCR3(pVCpu)))
1335	{
1336	pVCpu->vmm.s.iLastGZRc = VERR_PGM_NO_CR3_SHADOW_ROOT;
1337	break;
1338	}
1339	# endif
1340	if (RT_SUCCESS(g_rcRawModeUsability))
1341	{ /* likely */ }
1342	else
1343	{
1344	pVCpu->vmm.s.iLastGZRc = g_rcRawModeUsability;
1345	break;
1346	}
1347
1348	/*
1349	* Disable preemption.
1350	*/
1351	RTTHREADPREEMPTSTATE PreemptState = RTTHREADPREEMPTSTATE_INITIALIZER;
1352	RTThreadPreemptDisable(&PreemptState);
1353
1354	/*
1355	* Get the host CPU identifiers, make sure they are valid and that
1356	* we've got a TSC delta for the CPU.
1357	*/
1358	RTCPUID idHostCpu;
1359	uint32_t iHostCpuSet = RTMpCurSetIndexAndId(&idHostCpu);
1360	if (RT_LIKELY( iHostCpuSet < RTCPUSET_MAX_CPUS
1361	&& SUPIsTscDeltaAvailableForCpuSetIndex(iHostCpuSet)))
1362	{
1363	/*
1364	* Commit the CPU identifiers and update the periodict preemption timer if it's active.
1365	*/
1366	# ifdef VBOX_WITH_VMMR0_DISABLE_LAPIC_NMI
1367	CPUMR0SetLApic(pVCpu, iHostCpuSet);
1368	# endif
1369	pVCpu->iHostCpuSet = iHostCpuSet;
1370	ASMAtomicWriteU32(&pVCpu->idHostCpu, idHostCpu);
1371
1372	if (pVM->vmm.s.fUsePeriodicPreemptionTimers)
1373	GVMMR0SchedUpdatePeriodicPreemptionTimer(pVM, pVCpu->idHostCpu, TMCalcHostTimerFrequency(pVM, pVCpu));
1374
1375	/*
1376	* We might need to disable VT-x if the active switcher turns off paging.
1377	*/
1378	bool fVTxDisabled;
1379	int rc = HMR0EnterSwitcher(pVM, pVM->vmm.s.enmSwitcher, &fVTxDisabled);
1380	if (RT_SUCCESS(rc))
1381	{
1382	/*
1383	* Disable interrupts and run raw-mode code. The loop is for efficiently
1384	* dispatching tracepoints that fired in raw-mode context.
1385	*/
1386	RTCCUINTREG uFlags = ASMIntDisableFlags();
1387
1388	for (;;)
1389	{
1390	VMCPU_SET_STATE(pVCpu, VMCPUSTATE_STARTED_EXEC);
1391	TMNotifyStartOfExecution(pVCpu);
1392
1393	rc = pVM->vmm.s.pfnR0ToRawMode(pVM);
1394	pVCpu->vmm.s.iLastGZRc = rc;
1395
1396	TMNotifyEndOfExecution(pVCpu);
1397	VMCPU_SET_STATE(pVCpu, VMCPUSTATE_STARTED);
1398
1399	if (rc != VINF_VMM_CALL_TRACER)
1400	break;
1401	SUPR0TracerUmodProbeFire(pVM->pSession, &pVCpu->vmm.s.TracerCtx);
1402	}
1403
1404	/*
1405	* Re-enable VT-x before we dispatch any pending host interrupts and
1406	* re-enables interrupts.
1407	*/
1408	HMR0LeaveSwitcher(pVM, fVTxDisabled);
1409
1410	if ( rc == VINF_EM_RAW_INTERRUPT
1411	\|\| rc == VINF_EM_RAW_INTERRUPT_HYPER)
1412	TRPMR0DispatchHostInterrupt(pVM);
1413
1414	ASMSetFlags(uFlags);
1415
1416	/* Fire dtrace probe and collect statistics. */
1417	VBOXVMM_R0_VMM_RETURN_TO_RING3_RC(pVCpu, CPUMQueryGuestCtxPtr(pVCpu), rc);
1418	# ifdef VBOX_WITH_STATISTICS
1419	STAM_COUNTER_INC(&pVM->vmm.s.StatRunRC);
1420	vmmR0RecordRC(pVM, pVCpu, rc);
1421	# endif
1422	}
1423	else
1424	pVCpu->vmm.s.iLastGZRc = rc;
1425
1426	/*
1427	* Invalidate the host CPU identifiers as we restore preemption.
1428	*/
1429	pVCpu->iHostCpuSet = UINT32_MAX;
1430	ASMAtomicWriteU32(&pVCpu->idHostCpu, NIL_RTCPUID);
1431
1432	RTThreadPreemptRestore(&PreemptState);
1433	}
1434	/*
1435	* Invalid CPU set index or TSC delta in need of measuring.
1436	*/
1437	else
1438	{
1439	RTThreadPreemptRestore(&PreemptState);
1440	if (iHostCpuSet < RTCPUSET_MAX_CPUS)
1441	{
1442	int rc = SUPR0TscDeltaMeasureBySetIndex(pVM->pSession, iHostCpuSet, 0 /fFlags/,
1443	2 /cMsWaitRetry/, 5RT_MS_1SEC /cMsWaitThread*/,
1444	0 /default cTries/);
1445	if (RT_SUCCESS(rc) \|\| rc == VERR_CPU_OFFLINE)
1446	pVCpu->vmm.s.iLastGZRc = VINF_EM_RAW_TO_R3;
1447	else
1448	pVCpu->vmm.s.iLastGZRc = rc;
1449	}
1450	else
1451	pVCpu->vmm.s.iLastGZRc = VERR_INVALID_CPU_INDEX;
1452	}
1453
1454	#else /* !VBOX_WITH_RAW_MODE */
1455	pVCpu->vmm.s.iLastGZRc = VERR_RAW_MODE_NOT_SUPPORTED;
1456	#endif
1457	break;
1458	}
1459
1460	/*
1461	* Run guest code using the available hardware acceleration technology.
1462	*/
1463	case VMMR0_DO_HM_RUN:
1464	{
1465	for (;;) /* hlt loop */
1466	{
1467	/*
1468	* Disable preemption.
1469	*/
1470	Assert(!vmmR0ThreadCtxHookIsEnabled(pVCpu));
1471	RTTHREADPREEMPTSTATE PreemptState = RTTHREADPREEMPTSTATE_INITIALIZER;
1472	RTThreadPreemptDisable(&PreemptState);
1473
1474	/*
1475	* Get the host CPU identifiers, make sure they are valid and that
1476	* we've got a TSC delta for the CPU.
1477	*/
1478	RTCPUID idHostCpu;
1479	uint32_t iHostCpuSet = RTMpCurSetIndexAndId(&idHostCpu);
1480	if (RT_LIKELY( iHostCpuSet < RTCPUSET_MAX_CPUS
1481	&& SUPIsTscDeltaAvailableForCpuSetIndex(iHostCpuSet)))
1482	{
1483	pVCpu->iHostCpuSet = iHostCpuSet;
1484	ASMAtomicWriteU32(&pVCpu->idHostCpu, idHostCpu);
1485
1486	/*
1487	* Update the periodic preemption timer if it's active.
1488	*/
1489	if (pVM->vmm.s.fUsePeriodicPreemptionTimers)
1490	GVMMR0SchedUpdatePeriodicPreemptionTimer(pVM, pVCpu->idHostCpu, TMCalcHostTimerFrequency(pVM, pVCpu));
1491	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1492
1493	#ifdef VMM_R0_TOUCH_FPU
1494	/*
1495	* Make sure we've got the FPU state loaded so and we don't need to clear
1496	* CR0.TS and get out of sync with the host kernel when loading the guest
1497	* FPU state. @ref sec_cpum_fpu (CPUM.cpp) and @bugref{4053}.
1498	*/
1499	CPUMR0TouchHostFpu();
1500	#endif
1501	int rc;
1502	bool fPreemptRestored = false;
1503	if (!HMR0SuspendPending())
1504	{
1505	/*
1506	* Enable the context switching hook.
1507	*/
1508	if (pVCpu->vmm.s.hCtxHook != NIL_RTTHREADCTXHOOK)
1509	{
1510	Assert(!RTThreadCtxHookIsEnabled(pVCpu->vmm.s.hCtxHook));
1511	int rc2 = RTThreadCtxHookEnable(pVCpu->vmm.s.hCtxHook); AssertRC(rc2);
1512	}
1513
1514	/*
1515	* Enter HM context.
1516	*/
1517	rc = HMR0Enter(pVCpu);
1518	if (RT_SUCCESS(rc))
1519	{
1520	VMCPU_SET_STATE(pVCpu, VMCPUSTATE_STARTED_HM);
1521
1522	/*
1523	* When preemption hooks are in place, enable preemption now that
1524	* we're in HM context.
1525	*/
1526	if (vmmR0ThreadCtxHookIsEnabled(pVCpu))
1527	{
1528	fPreemptRestored = true;
1529	RTThreadPreemptRestore(&PreemptState);
1530	}
1531
1532	/*
1533	* Setup the longjmp machinery and execute guest code (calls HMR0RunGuestCode).
1534	*/
1535	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1536	rc = vmmR0CallRing3SetJmp(&pVCpu->vmm.s.CallRing3JmpBufR0, HMR0RunGuestCode, pVM, pVCpu);
1537	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1538
1539	/*
1540	* Assert sanity on the way out. Using manual assertions code here as normal
1541	* assertions are going to panic the host since we're outside the setjmp/longjmp zone.
1542	*/
1543	if (RT_UNLIKELY( VMCPU_GET_STATE(pVCpu) != VMCPUSTATE_STARTED_HM
1544	&& RT_SUCCESS_NP(rc) && rc != VINF_VMM_CALL_HOST ))
1545	{
1546	pVM->vmm.s.szRing0AssertMsg1[0] = '\0';
1547	RTStrPrintf(pVM->vmm.s.szRing0AssertMsg2, sizeof(pVM->vmm.s.szRing0AssertMsg2),
1548	"Got VMCPU state %d expected %d.\n", VMCPU_GET_STATE(pVCpu), VMCPUSTATE_STARTED_HM);
1549	rc = VERR_VMM_WRONG_HM_VMCPU_STATE;
1550	}
1551	/** @todo Get rid of this. HM shouldn't disable the context hook. */
1552	else if (RT_UNLIKELY(vmmR0ThreadCtxHookIsEnabled(pVCpu)))
1553	{
1554	pVM->vmm.s.szRing0AssertMsg1[0] = '\0';
1555	RTStrPrintf(pVM->vmm.s.szRing0AssertMsg2, sizeof(pVM->vmm.s.szRing0AssertMsg2),
1556	"Thread-context hooks still enabled! VCPU=%p Id=%u rc=%d.\n", pVCpu, pVCpu->idCpu, rc);
1557	rc = VERR_INVALID_STATE;
1558	}
1559
1560	VMCPU_SET_STATE(pVCpu, VMCPUSTATE_STARTED);
1561	}
1562	STAM_COUNTER_INC(&pVM->vmm.s.StatRunRC);
1563
1564	/*
1565	* Invalidate the host CPU identifiers before we disable the context
1566	* hook / restore preemption.
1567	*/
1568	pVCpu->iHostCpuSet = UINT32_MAX;
1569	ASMAtomicWriteU32(&pVCpu->idHostCpu, NIL_RTCPUID);
1570
1571	/*
1572	* Disable context hooks. Due to unresolved cleanup issues, we
1573	* cannot leave the hooks enabled when we return to ring-3.
1574	*
1575	* Note! At the moment HM may also have disabled the hook
1576	* when we get here, but the IPRT API handles that.
1577	*/
1578	if (pVCpu->vmm.s.hCtxHook != NIL_RTTHREADCTXHOOK)
1579	{
1580	ASMAtomicWriteU32(&pVCpu->idHostCpu, NIL_RTCPUID);
1581	RTThreadCtxHookDisable(pVCpu->vmm.s.hCtxHook);
1582	}
1583	}
1584	/*
1585	* The system is about to go into suspend mode; go back to ring 3.
1586	*/
1587	else
1588	{
1589	rc = VINF_EM_RAW_INTERRUPT;
1590	pVCpu->iHostCpuSet = UINT32_MAX;
1591	ASMAtomicWriteU32(&pVCpu->idHostCpu, NIL_RTCPUID);
1592	}
1593
1594	/** @todo When HM stops messing with the context hook state, we'll disable
1595	* preemption again before the RTThreadCtxHookDisable call. */
1596	if (!fPreemptRestored)
1597	RTThreadPreemptRestore(&PreemptState);
1598
1599	pVCpu->vmm.s.iLastGZRc = rc;
1600
1601	/* Fire dtrace probe and collect statistics. */
1602	VBOXVMM_R0_VMM_RETURN_TO_RING3_HM(pVCpu, CPUMQueryGuestCtxPtr(pVCpu), rc);
1603	#ifdef VBOX_WITH_STATISTICS
1604	vmmR0RecordRC(pVM, pVCpu, rc);
1605	#endif
1606	#if 1
1607	/*
1608	* If this is a halt.
1609	*/
1610	if (rc != VINF_EM_HALT)
1611	{ /* we're not in a hurry for a HLT, so prefer this path */ }
1612	else
1613	{
1614	pVCpu->vmm.s.iLastGZRc = rc = vmmR0DoHalt(pGVM, pVM, pGVCpu, pVCpu);
1615	if (rc == VINF_SUCCESS)
1616	{
1617	pVCpu->vmm.s.cR0HaltsSucceeded++;
1618	continue;
1619	}
1620	pVCpu->vmm.s.cR0HaltsToRing3++;
1621	}
1622	#endif
1623	}
1624	/*
1625	* Invalid CPU set index or TSC delta in need of measuring.
1626	*/
1627	else
1628	{
1629	pVCpu->iHostCpuSet = UINT32_MAX;
1630	ASMAtomicWriteU32(&pVCpu->idHostCpu, NIL_RTCPUID);
1631	RTThreadPreemptRestore(&PreemptState);
1632	if (iHostCpuSet < RTCPUSET_MAX_CPUS)
1633	{
1634	int rc = SUPR0TscDeltaMeasureBySetIndex(pVM->pSession, iHostCpuSet, 0 /fFlags/,
1635	2 /cMsWaitRetry/, 5RT_MS_1SEC /cMsWaitThread*/,
1636	0 /default cTries/);
1637	if (RT_SUCCESS(rc) \|\| rc == VERR_CPU_OFFLINE)
1638	pVCpu->vmm.s.iLastGZRc = VINF_EM_RAW_TO_R3;
1639	else
1640	pVCpu->vmm.s.iLastGZRc = rc;
1641	}
1642	else
1643	pVCpu->vmm.s.iLastGZRc = VERR_INVALID_CPU_INDEX;
1644	}
1645	break;
1646
1647	} /* halt loop. */
1648	break;
1649	}
1650
1651	#ifdef VBOX_WITH_NEM_R0
1652	# if defined(RT_ARCH_AMD64) && defined(RT_OS_WINDOWS)
1653	case VMMR0_DO_NEM_RUN:
1654	{
1655	/*
1656	* Setup the longjmp machinery and execute guest code (calls NEMR0RunGuestCode).
1657	*/
1658	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1659	int rc = vmmR0CallRing3SetJmp2(&pVCpu->vmm.s.CallRing3JmpBufR0, NEMR0RunGuestCode, pGVM, idCpu);
1660	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1661	STAM_COUNTER_INC(&pVM->vmm.s.StatRunRC);
1662
1663	pVCpu->vmm.s.iLastGZRc = rc;
1664
1665	/*
1666	* Fire dtrace probe and collect statistics.
1667	*/
1668	VBOXVMM_R0_VMM_RETURN_TO_RING3_NEM(pVCpu, CPUMQueryGuestCtxPtr(pVCpu), rc);
1669	# ifdef VBOX_WITH_STATISTICS
1670	vmmR0RecordRC(pVM, pVCpu, rc);
1671	# endif
1672	break;
1673	}
1674	# endif
1675	#endif
1676
1677
1678	/*
1679	* For profiling.
1680	*/
1681	case VMMR0_DO_NOP:
1682	pVCpu->vmm.s.iLastGZRc = VINF_SUCCESS;
1683	break;
1684
1685	/*
1686	* Shouldn't happen.
1687	*/
1688	default:
1689	AssertMsgFailed(("%#x\n", enmOperation));
1690	pVCpu->vmm.s.iLastGZRc = VERR_NOT_SUPPORTED;
1691	break;
1692	}
1693	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1694	}
1695
1696
1697	/**
1698	* Validates a session or VM session argument.
1699	*
1700	* @returns true / false accordingly.
1701	* @param pVM The cross context VM structure.
1702	* @param pClaimedSession The session claim to validate.
1703	* @param pSession The session argument.
1704	*/
1705	DECLINLINE(bool) vmmR0IsValidSession(PVM pVM, PSUPDRVSESSION pClaimedSession, PSUPDRVSESSION pSession)
1706	{
1707	/* This must be set! */
1708	if (!pSession)
1709	return false;
1710
1711	/* Only one out of the two. */
1712	if (pVM && pClaimedSession)
1713	return false;
1714	if (pVM)
1715	pClaimedSession = pVM->pSession;
1716	return pClaimedSession == pSession;
1717	}
1718
1719
1720	/**
1721	* VMMR0EntryEx worker function, either called directly or when ever possible
1722	* called thru a longjmp so we can exit safely on failure.
1723	*
1724	* @returns VBox status code.
1725	* @param pGVM The global (ring-0) VM structure.
1726	* @param pVM The cross context VM structure.
1727	* @param idCpu Virtual CPU ID argument. Must be NIL_VMCPUID if pVM
1728	* is NIL_RTR0PTR, and may be NIL_VMCPUID if it isn't
1729	* @param enmOperation Which operation to execute.
1730	* @param pReqHdr This points to a SUPVMMR0REQHDR packet. Optional.
1731	* The support driver validates this if it's present.
1732	* @param u64Arg Some simple constant argument.
1733	* @param pSession The session of the caller.
1734	*
1735	* @remarks Assume called with interrupts _enabled_.
1736	*/
1737	static int vmmR0EntryExWorker(PGVM pGVM, PVM pVM, VMCPUID idCpu, VMMR0OPERATION enmOperation,
1738	PSUPVMMR0REQHDR pReqHdr, uint64_t u64Arg, PSUPDRVSESSION pSession)
1739	{
1740	/*
1741	* Validate pGVM, pVM and idCpu for consistency and validity.
1742	*/
1743	if ( pGVM != NULL
1744	\|\| pVM != NULL)
1745	{
1746	if (RT_LIKELY( RT_VALID_PTR(pGVM)
1747	&& RT_VALID_PTR(pVM)
1748	&& ((uintptr_t)pVM & PAGE_OFFSET_MASK) == 0))
1749	{ /* likely */ }
1750	else
1751	{
1752	SUPR0Printf("vmmR0EntryExWorker: Invalid pGVM=%p and/or pVM=%p! (op=%d)\n", pGVM, pVM, enmOperation);
1753	return VERR_INVALID_POINTER;
1754	}
1755
1756	if (RT_LIKELY(pGVM->pVM == pVM))
1757	{ /* likely */ }
1758	else
1759	{
1760	SUPR0Printf("vmmR0EntryExWorker: pVM mismatch: got %p, pGVM->pVM=%p\n", pVM, pGVM->pVM);
1761	return VERR_INVALID_PARAMETER;
1762	}
1763
1764	if (RT_LIKELY(idCpu == NIL_VMCPUID \|\| idCpu < pGVM->cCpus))
1765	{ /* likely */ }
1766	else
1767	{
1768	SUPR0Printf("vmmR0EntryExWorker: Invalid idCpu %#x (cCpus=%#x)\n", idCpu, pGVM->cCpus);
1769	return VERR_INVALID_PARAMETER;
1770	}
1771
1772	if (RT_LIKELY( pVM->enmVMState >= VMSTATE_CREATING
1773	&& pVM->enmVMState <= VMSTATE_TERMINATED
1774	&& pVM->cCpus == pGVM->cCpus
1775	&& pVM->pSession == pSession
1776	&& pVM->pVMR0 == pVM))
1777	{ /* likely */ }
1778	else
1779	{
1780	SUPR0Printf("vmmR0EntryExWorker: Invalid pVM=%p:{.enmVMState=%d, .cCpus=%#x(==%#x), .pSession=%p(==%p), .pVMR0=%p(==%p)}! (op=%d)\n",
1781	pVM, pVM->enmVMState, pVM->cCpus, pGVM->cCpus, pVM->pSession, pSession, pVM->pVMR0, pVM, enmOperation);
1782	return VERR_INVALID_POINTER;
1783	}
1784	}
1785	else if (RT_LIKELY(idCpu == NIL_VMCPUID))
1786	{ /* likely */ }
1787	else
1788	{
1789	SUPR0Printf("vmmR0EntryExWorker: Invalid idCpu=%u\n", idCpu);
1790	return VERR_INVALID_PARAMETER;
1791	}
1792
1793	/*
1794	* SMAP fun.
1795	*/
1796	VMM_CHECK_SMAP_SETUP();
1797	VMM_CHECK_SMAP_CHECK(RT_NOTHING);
1798
1799	/*
1800	* Process the request.
1801	*/
1802	int rc;
1803	switch (enmOperation)
1804	{
1805	/*
1806	* GVM requests
1807	*/
1808	case VMMR0_DO_GVMM_CREATE_VM:
1809	if (pGVM == NULL && pVM == NULL && u64Arg == 0 && idCpu == NIL_VMCPUID)
1810	rc = GVMMR0CreateVMReq((PGVMMCREATEVMREQ)pReqHdr, pSession);
1811	else
1812	rc = VERR_INVALID_PARAMETER;
1813	VMM_CHECK_SMAP_CHECK(RT_NOTHING);
1814	break;
1815
1816	case VMMR0_DO_GVMM_DESTROY_VM:
1817	if (pReqHdr == NULL && u64Arg == 0)
1818	rc = GVMMR0DestroyVM(pGVM, pVM);
1819	else
1820	rc = VERR_INVALID_PARAMETER;
1821	VMM_CHECK_SMAP_CHECK(RT_NOTHING);
1822	break;
1823
1824	case VMMR0_DO_GVMM_REGISTER_VMCPU:
1825	if (pGVM != NULL && pVM != NULL)
1826	rc = GVMMR0RegisterVCpu(pGVM, pVM, idCpu);
1827	else
1828	rc = VERR_INVALID_PARAMETER;
1829	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1830	break;
1831
1832	case VMMR0_DO_GVMM_DEREGISTER_VMCPU:
1833	if (pGVM != NULL && pVM != NULL)
1834	rc = GVMMR0DeregisterVCpu(pGVM, pVM, idCpu);
1835	else
1836	rc = VERR_INVALID_PARAMETER;
1837	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1838	break;
1839
1840	case VMMR0_DO_GVMM_SCHED_HALT:
1841	if (pReqHdr)
1842	return VERR_INVALID_PARAMETER;
1843	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1844	rc = GVMMR0SchedHaltReq(pGVM, pVM, idCpu, u64Arg);
1845	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1846	break;
1847
1848	case VMMR0_DO_GVMM_SCHED_WAKE_UP:
1849	if (pReqHdr \|\| u64Arg)
1850	return VERR_INVALID_PARAMETER;
1851	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1852	rc = GVMMR0SchedWakeUp(pGVM, pVM, idCpu);
1853	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1854	break;
1855
1856	case VMMR0_DO_GVMM_SCHED_POKE:
1857	if (pReqHdr \|\| u64Arg)
1858	return VERR_INVALID_PARAMETER;
1859	rc = GVMMR0SchedPoke(pGVM, pVM, idCpu);
1860	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1861	break;
1862
1863	case VMMR0_DO_GVMM_SCHED_WAKE_UP_AND_POKE_CPUS:
1864	if (u64Arg)
1865	return VERR_INVALID_PARAMETER;
1866	rc = GVMMR0SchedWakeUpAndPokeCpusReq(pGVM, pVM, (PGVMMSCHEDWAKEUPANDPOKECPUSREQ)pReqHdr);
1867	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1868	break;
1869
1870	case VMMR0_DO_GVMM_SCHED_POLL:
1871	if (pReqHdr \|\| u64Arg > 1)
1872	return VERR_INVALID_PARAMETER;
1873	rc = GVMMR0SchedPoll(pGVM, pVM, idCpu, !!u64Arg);
1874	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1875	break;
1876
1877	case VMMR0_DO_GVMM_QUERY_STATISTICS:
1878	if (u64Arg)
1879	return VERR_INVALID_PARAMETER;
1880	rc = GVMMR0QueryStatisticsReq(pGVM, pVM, (PGVMMQUERYSTATISTICSSREQ)pReqHdr, pSession);
1881	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1882	break;
1883
1884	case VMMR0_DO_GVMM_RESET_STATISTICS:
1885	if (u64Arg)
1886	return VERR_INVALID_PARAMETER;
1887	rc = GVMMR0ResetStatisticsReq(pGVM, pVM, (PGVMMRESETSTATISTICSSREQ)pReqHdr, pSession);
1888	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1889	break;
1890
1891	/*
1892	* Initialize the R0 part of a VM instance.
1893	*/
1894	case VMMR0_DO_VMMR0_INIT:
1895	rc = vmmR0InitVM(pGVM, pVM, RT_LODWORD(u64Arg), RT_HIDWORD(u64Arg));
1896	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1897	break;
1898
1899	/*
1900	* Does EMT specific ring-0 init.
1901	*/
1902	case VMMR0_DO_VMMR0_INIT_EMT:
1903	rc = vmmR0InitVMEmt(pGVM, pVM, idCpu);
1904	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1905	break;
1906
1907	/*
1908	* Terminate the R0 part of a VM instance.
1909	*/
1910	case VMMR0_DO_VMMR0_TERM:
1911	rc = VMMR0TermVM(pGVM, pVM, 0 /idCpu/);
1912	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1913	break;
1914
1915	/*
1916	* Attempt to enable hm mode and check the current setting.
1917	*/
1918	case VMMR0_DO_HM_ENABLE:
1919	rc = HMR0EnableAllCpus(pVM);
1920	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1921	break;
1922
1923	/*
1924	* Setup the hardware accelerated session.
1925	*/
1926	case VMMR0_DO_HM_SETUP_VM:
1927	rc = HMR0SetupVM(pVM);
1928	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1929	break;
1930
1931	/*
1932	* Switch to RC to execute Hypervisor function.
1933	*/
1934	case VMMR0_DO_CALL_HYPERVISOR:
1935	{
1936	#ifdef VBOX_WITH_RAW_MODE
1937	/*
1938	* Validate input / context.
1939	*/
1940	if (RT_UNLIKELY(idCpu != 0))
1941	return VERR_INVALID_CPU_ID;
1942	if (RT_UNLIKELY(pVM->cCpus != 1))
1943	return VERR_INVALID_PARAMETER;
1944	PVMCPU pVCpu = &pVM->aCpus[idCpu];
1945	# ifndef VBOX_WITH_2X_4GB_ADDR_SPACE_IN_R0
1946	if (RT_UNLIKELY(!PGMGetHyperCR3(pVCpu)))
1947	return VERR_PGM_NO_CR3_SHADOW_ROOT;
1948	# endif
1949	if (RT_FAILURE(g_rcRawModeUsability))
1950	return g_rcRawModeUsability;
1951
1952	/*
1953	* Disable interrupts.
1954	*/
1955	RTCCUINTREG fFlags = ASMIntDisableFlags();
1956
1957	/*
1958	* Get the host CPU identifiers, make sure they are valid and that
1959	* we've got a TSC delta for the CPU.
1960	*/
1961	RTCPUID idHostCpu;
1962	uint32_t iHostCpuSet = RTMpCurSetIndexAndId(&idHostCpu);
1963	if (RT_UNLIKELY(iHostCpuSet >= RTCPUSET_MAX_CPUS))
1964	{
1965	ASMSetFlags(fFlags);
1966	return VERR_INVALID_CPU_INDEX;
1967	}
1968	if (RT_UNLIKELY(!SUPIsTscDeltaAvailableForCpuSetIndex(iHostCpuSet)))
1969	{
1970	ASMSetFlags(fFlags);
1971	rc = SUPR0TscDeltaMeasureBySetIndex(pVM->pSession, iHostCpuSet, 0 /fFlags/,
1972	2 /cMsWaitRetry/, 5RT_MS_1SEC /cMsWaitThread*/,
1973	0 /default cTries/);
1974	if (RT_FAILURE(rc) && rc != VERR_CPU_OFFLINE)
1975	{
1976	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1977	return rc;
1978	}
1979	}
1980
1981	/*
1982	* Commit the CPU identifiers.
1983	*/
1984	# ifdef VBOX_WITH_VMMR0_DISABLE_LAPIC_NMI
1985	CPUMR0SetLApic(pVCpu, iHostCpuSet);
1986	# endif
1987	pVCpu->iHostCpuSet = iHostCpuSet;
1988	ASMAtomicWriteU32(&pVCpu->idHostCpu, idHostCpu);
1989
1990	/*
1991	* We might need to disable VT-x if the active switcher turns off paging.
1992	*/
1993	bool fVTxDisabled;
1994	rc = HMR0EnterSwitcher(pVM, pVM->vmm.s.enmSwitcher, &fVTxDisabled);
1995	if (RT_SUCCESS(rc))
1996	{
1997	/*
1998	* Go through the wormhole...
1999	*/
2000	rc = pVM->vmm.s.pfnR0ToRawMode(pVM);
2001
2002	/*
2003	* Re-enable VT-x before we dispatch any pending host interrupts.
2004	*/
2005	HMR0LeaveSwitcher(pVM, fVTxDisabled);
2006
2007	if ( rc == VINF_EM_RAW_INTERRUPT
2008	\|\| rc == VINF_EM_RAW_INTERRUPT_HYPER)
2009	TRPMR0DispatchHostInterrupt(pVM);
2010	}
2011
2012	/*
2013	* Invalidate the host CPU identifiers as we restore interrupts.
2014	*/
2015	pVCpu->iHostCpuSet = UINT32_MAX;
2016	ASMAtomicWriteU32(&pVCpu->idHostCpu, NIL_RTCPUID);
2017	ASMSetFlags(fFlags);
2018
2019	#else /* !VBOX_WITH_RAW_MODE */
2020	rc = VERR_RAW_MODE_NOT_SUPPORTED;
2021	#endif
2022	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2023	break;
2024	}
2025
2026	/*
2027	* PGM wrappers.
2028	*/
2029	case VMMR0_DO_PGM_ALLOCATE_HANDY_PAGES:
2030	if (idCpu == NIL_VMCPUID)
2031	return VERR_INVALID_CPU_ID;
2032	rc = PGMR0PhysAllocateHandyPages(pGVM, pVM, idCpu);
2033	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2034	break;
2035
2036	case VMMR0_DO_PGM_FLUSH_HANDY_PAGES:
2037	if (idCpu == NIL_VMCPUID)
2038	return VERR_INVALID_CPU_ID;
2039	rc = PGMR0PhysFlushHandyPages(pGVM, pVM, idCpu);
2040	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2041	break;
2042
2043	case VMMR0_DO_PGM_ALLOCATE_LARGE_HANDY_PAGE:
2044	if (idCpu == NIL_VMCPUID)
2045	return VERR_INVALID_CPU_ID;
2046	rc = PGMR0PhysAllocateLargeHandyPage(pGVM, pVM, idCpu);
2047	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2048	break;
2049
2050	case VMMR0_DO_PGM_PHYS_SETUP_IOMMU:
2051	if (idCpu != 0)
2052	return VERR_INVALID_CPU_ID;
2053	rc = PGMR0PhysSetupIoMmu(pGVM, pVM);
2054	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2055	break;
2056
2057	/*
2058	* GMM wrappers.
2059	*/
2060	case VMMR0_DO_GMM_INITIAL_RESERVATION:
2061	if (u64Arg)
2062	return VERR_INVALID_PARAMETER;
2063	rc = GMMR0InitialReservationReq(pGVM, pVM, idCpu, (PGMMINITIALRESERVATIONREQ)pReqHdr);
2064	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2065	break;
2066
2067	case VMMR0_DO_GMM_UPDATE_RESERVATION:
2068	if (u64Arg)
2069	return VERR_INVALID_PARAMETER;
2070	rc = GMMR0UpdateReservationReq(pGVM, pVM, idCpu, (PGMMUPDATERESERVATIONREQ)pReqHdr);
2071	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2072	break;
2073
2074	case VMMR0_DO_GMM_ALLOCATE_PAGES:
2075	if (u64Arg)
2076	return VERR_INVALID_PARAMETER;
2077	rc = GMMR0AllocatePagesReq(pGVM, pVM, idCpu, (PGMMALLOCATEPAGESREQ)pReqHdr);
2078	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2079	break;
2080
2081	case VMMR0_DO_GMM_FREE_PAGES:
2082	if (u64Arg)
2083	return VERR_INVALID_PARAMETER;
2084	rc = GMMR0FreePagesReq(pGVM, pVM, idCpu, (PGMMFREEPAGESREQ)pReqHdr);
2085	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2086	break;
2087
2088	case VMMR0_DO_GMM_FREE_LARGE_PAGE:
2089	if (u64Arg)
2090	return VERR_INVALID_PARAMETER;
2091	rc = GMMR0FreeLargePageReq(pGVM, pVM, idCpu, (PGMMFREELARGEPAGEREQ)pReqHdr);
2092	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2093	break;
2094
2095	case VMMR0_DO_GMM_QUERY_HYPERVISOR_MEM_STATS:
2096	if (u64Arg)
2097	return VERR_INVALID_PARAMETER;
2098	rc = GMMR0QueryHypervisorMemoryStatsReq((PGMMMEMSTATSREQ)pReqHdr);
2099	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2100	break;
2101
2102	case VMMR0_DO_GMM_QUERY_MEM_STATS:
2103	if (idCpu == NIL_VMCPUID)
2104	return VERR_INVALID_CPU_ID;
2105	if (u64Arg)
2106	return VERR_INVALID_PARAMETER;
2107	rc = GMMR0QueryMemoryStatsReq(pGVM, pVM, idCpu, (PGMMMEMSTATSREQ)pReqHdr);
2108	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2109	break;
2110
2111	case VMMR0_DO_GMM_BALLOONED_PAGES:
2112	if (u64Arg)
2113	return VERR_INVALID_PARAMETER;
2114	rc = GMMR0BalloonedPagesReq(pGVM, pVM, idCpu, (PGMMBALLOONEDPAGESREQ)pReqHdr);
2115	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2116	break;
2117
2118	case VMMR0_DO_GMM_MAP_UNMAP_CHUNK:
2119	if (u64Arg)
2120	return VERR_INVALID_PARAMETER;
2121	rc = GMMR0MapUnmapChunkReq(pGVM, pVM, (PGMMMAPUNMAPCHUNKREQ)pReqHdr);
2122	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2123	break;
2124
2125	case VMMR0_DO_GMM_SEED_CHUNK:
2126	if (pReqHdr)
2127	return VERR_INVALID_PARAMETER;
2128	rc = GMMR0SeedChunk(pGVM, pVM, idCpu, (RTR3PTR)u64Arg);
2129	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2130	break;
2131
2132	case VMMR0_DO_GMM_REGISTER_SHARED_MODULE:
2133	if (idCpu == NIL_VMCPUID)
2134	return VERR_INVALID_CPU_ID;
2135	if (u64Arg)
2136	return VERR_INVALID_PARAMETER;
2137	rc = GMMR0RegisterSharedModuleReq(pGVM, pVM, idCpu, (PGMMREGISTERSHAREDMODULEREQ)pReqHdr);
2138	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2139	break;
2140
2141	case VMMR0_DO_GMM_UNREGISTER_SHARED_MODULE:
2142	if (idCpu == NIL_VMCPUID)
2143	return VERR_INVALID_CPU_ID;
2144	if (u64Arg)
2145	return VERR_INVALID_PARAMETER;
2146	rc = GMMR0UnregisterSharedModuleReq(pGVM, pVM, idCpu, (PGMMUNREGISTERSHAREDMODULEREQ)pReqHdr);
2147	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2148	break;
2149
2150	case VMMR0_DO_GMM_RESET_SHARED_MODULES:
2151	if (idCpu == NIL_VMCPUID)
2152	return VERR_INVALID_CPU_ID;
2153	if ( u64Arg
2154	\|\| pReqHdr)
2155	return VERR_INVALID_PARAMETER;
2156	rc = GMMR0ResetSharedModules(pGVM, pVM, idCpu);
2157	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2158	break;
2159
2160	#ifdef VBOX_WITH_PAGE_SHARING
2161	case VMMR0_DO_GMM_CHECK_SHARED_MODULES:
2162	{
2163	if (idCpu == NIL_VMCPUID)
2164	return VERR_INVALID_CPU_ID;
2165	if ( u64Arg
2166	\|\| pReqHdr)
2167	return VERR_INVALID_PARAMETER;
2168	rc = GMMR0CheckSharedModules(pGVM, pVM, idCpu);
2169	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2170	break;
2171	}
2172	#endif
2173
2174	#if defined(VBOX_STRICT) && HC_ARCH_BITS == 64
2175	case VMMR0_DO_GMM_FIND_DUPLICATE_PAGE:
2176	if (u64Arg)
2177	return VERR_INVALID_PARAMETER;
2178	rc = GMMR0FindDuplicatePageReq(pGVM, pVM, (PGMMFINDDUPLICATEPAGEREQ)pReqHdr);
2179	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2180	break;
2181	#endif
2182
2183	case VMMR0_DO_GMM_QUERY_STATISTICS:
2184	if (u64Arg)
2185	return VERR_INVALID_PARAMETER;
2186	rc = GMMR0QueryStatisticsReq(pGVM, pVM, (PGMMQUERYSTATISTICSSREQ)pReqHdr);
2187	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2188	break;
2189
2190	case VMMR0_DO_GMM_RESET_STATISTICS:
2191	if (u64Arg)
2192	return VERR_INVALID_PARAMETER;
2193	rc = GMMR0ResetStatisticsReq(pGVM, pVM, (PGMMRESETSTATISTICSSREQ)pReqHdr);
2194	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2195	break;
2196
2197	/*
2198	* A quick GCFGM mock-up.
2199	*/
2200	/** @todo GCFGM with proper access control, ring-3 management interface and all that. */
2201	case VMMR0_DO_GCFGM_SET_VALUE:
2202	case VMMR0_DO_GCFGM_QUERY_VALUE:
2203	{
2204	if (pGVM \|\| pVM \|\| !pReqHdr \|\| u64Arg \|\| idCpu != NIL_VMCPUID)
2205	return VERR_INVALID_PARAMETER;
2206	PGCFGMVALUEREQ pReq = (PGCFGMVALUEREQ)pReqHdr;
2207	if (pReq->Hdr.cbReq != sizeof(*pReq))
2208	return VERR_INVALID_PARAMETER;
2209	if (enmOperation == VMMR0_DO_GCFGM_SET_VALUE)
2210	{
2211	rc = GVMMR0SetConfig(pReq->pSession, &pReq->szName[0], pReq->u64Value);
2212	//if (rc == VERR_CFGM_VALUE_NOT_FOUND)
2213	// rc = GMMR0SetConfig(pReq->pSession, &pReq->szName[0], pReq->u64Value);
2214	}
2215	else
2216	{
2217	rc = GVMMR0QueryConfig(pReq->pSession, &pReq->szName[0], &pReq->u64Value);
2218	//if (rc == VERR_CFGM_VALUE_NOT_FOUND)
2219	// rc = GMMR0QueryConfig(pReq->pSession, &pReq->szName[0], &pReq->u64Value);
2220	}
2221	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2222	break;
2223	}
2224
2225	/*
2226	* PDM Wrappers.
2227	*/
2228	case VMMR0_DO_PDM_DRIVER_CALL_REQ_HANDLER:
2229	{
2230	if (!pReqHdr \|\| u64Arg \|\| idCpu != NIL_VMCPUID)
2231	return VERR_INVALID_PARAMETER;
2232	rc = PDMR0DriverCallReqHandler(pGVM, pVM, (PPDMDRIVERCALLREQHANDLERREQ)pReqHdr);
2233	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2234	break;
2235	}
2236
2237	case VMMR0_DO_PDM_DEVICE_CALL_REQ_HANDLER:
2238	{
2239	if (!pReqHdr \|\| u64Arg \|\| idCpu != NIL_VMCPUID)
2240	return VERR_INVALID_PARAMETER;
2241	rc = PDMR0DeviceCallReqHandler(pGVM, pVM, (PPDMDEVICECALLREQHANDLERREQ)pReqHdr);
2242	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2243	break;
2244	}
2245
2246	/*
2247	* Requests to the internal networking service.
2248	*/
2249	case VMMR0_DO_INTNET_OPEN:
2250	{
2251	PINTNETOPENREQ pReq = (PINTNETOPENREQ)pReqHdr;
2252	if (u64Arg \|\| !pReq \|\| !vmmR0IsValidSession(pVM, pReq->pSession, pSession) \|\| idCpu != NIL_VMCPUID)
2253	return VERR_INVALID_PARAMETER;
2254	rc = IntNetR0OpenReq(pSession, pReq);
2255	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2256	break;
2257	}
2258
2259	case VMMR0_DO_INTNET_IF_CLOSE:
2260	if (u64Arg \|\| !pReqHdr \|\| !vmmR0IsValidSession(pVM, ((PINTNETIFCLOSEREQ)pReqHdr)->pSession, pSession) \|\| idCpu != NIL_VMCPUID)
2261	return VERR_INVALID_PARAMETER;
2262	rc = IntNetR0IfCloseReq(pSession, (PINTNETIFCLOSEREQ)pReqHdr);
2263	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2264	break;
2265
2266
2267	case VMMR0_DO_INTNET_IF_GET_BUFFER_PTRS:
2268	if (u64Arg \|\| !pReqHdr \|\| !vmmR0IsValidSession(pVM, ((PINTNETIFGETBUFFERPTRSREQ)pReqHdr)->pSession, pSession) \|\| idCpu != NIL_VMCPUID)
2269	return VERR_INVALID_PARAMETER;
2270	rc = IntNetR0IfGetBufferPtrsReq(pSession, (PINTNETIFGETBUFFERPTRSREQ)pReqHdr);
2271	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2272	break;
2273
2274	case VMMR0_DO_INTNET_IF_SET_PROMISCUOUS_MODE:
2275	if (u64Arg \|\| !pReqHdr \|\| !vmmR0IsValidSession(pVM, ((PINTNETIFSETPROMISCUOUSMODEREQ)pReqHdr)->pSession, pSession) \|\| idCpu != NIL_VMCPUID)
2276	return VERR_INVALID_PARAMETER;
2277	rc = IntNetR0IfSetPromiscuousModeReq(pSession, (PINTNETIFSETPROMISCUOUSMODEREQ)pReqHdr);
2278	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2279	break;
2280
2281	case VMMR0_DO_INTNET_IF_SET_MAC_ADDRESS:
2282	if (u64Arg \|\| !pReqHdr \|\| !vmmR0IsValidSession(pVM, ((PINTNETIFSETMACADDRESSREQ)pReqHdr)->pSession, pSession) \|\| idCpu != NIL_VMCPUID)
2283	return VERR_INVALID_PARAMETER;
2284	rc = IntNetR0IfSetMacAddressReq(pSession, (PINTNETIFSETMACADDRESSREQ)pReqHdr);
2285	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2286	break;
2287
2288	case VMMR0_DO_INTNET_IF_SET_ACTIVE:
2289	if (u64Arg \|\| !pReqHdr \|\| !vmmR0IsValidSession(pVM, ((PINTNETIFSETACTIVEREQ)pReqHdr)->pSession, pSession) \|\| idCpu != NIL_VMCPUID)
2290	return VERR_INVALID_PARAMETER;
2291	rc = IntNetR0IfSetActiveReq(pSession, (PINTNETIFSETACTIVEREQ)pReqHdr);
2292	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2293	break;
2294
2295	case VMMR0_DO_INTNET_IF_SEND:
2296	if (u64Arg \|\| !pReqHdr \|\| !vmmR0IsValidSession(pVM, ((PINTNETIFSENDREQ)pReqHdr)->pSession, pSession) \|\| idCpu != NIL_VMCPUID)
2297	return VERR_INVALID_PARAMETER;
2298	rc = IntNetR0IfSendReq(pSession, (PINTNETIFSENDREQ)pReqHdr);
2299	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2300	break;
2301
2302	case VMMR0_DO_INTNET_IF_WAIT:
2303	if (u64Arg \|\| !pReqHdr \|\| !vmmR0IsValidSession(pVM, ((PINTNETIFWAITREQ)pReqHdr)->pSession, pSession) \|\| idCpu != NIL_VMCPUID)
2304	return VERR_INVALID_PARAMETER;
2305	rc = IntNetR0IfWaitReq(pSession, (PINTNETIFWAITREQ)pReqHdr);
2306	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2307	break;
2308
2309	case VMMR0_DO_INTNET_IF_ABORT_WAIT:
2310	if (u64Arg \|\| !pReqHdr \|\| !vmmR0IsValidSession(pVM, ((PINTNETIFWAITREQ)pReqHdr)->pSession, pSession) \|\| idCpu != NIL_VMCPUID)
2311	return VERR_INVALID_PARAMETER;
2312	rc = IntNetR0IfAbortWaitReq(pSession, (PINTNETIFABORTWAITREQ)pReqHdr);
2313	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2314	break;
2315
2316	#ifdef VBOX_WITH_PCI_PASSTHROUGH
2317	/*
2318	* Requests to host PCI driver service.
2319	*/
2320	case VMMR0_DO_PCIRAW_REQ:
2321	if (u64Arg \|\| !pReqHdr \|\| !vmmR0IsValidSession(pVM, ((PPCIRAWSENDREQ)pReqHdr)->pSession, pSession) \|\| idCpu != NIL_VMCPUID)
2322	return VERR_INVALID_PARAMETER;
2323	rc = PciRawR0ProcessReq(pGVM, pVM, pSession, (PPCIRAWSENDREQ)pReqHdr);
2324	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2325	break;
2326	#endif
2327
2328	/*
2329	* NEM requests.
2330	*/
2331	#ifdef VBOX_WITH_NEM_R0
2332	# if defined(RT_ARCH_AMD64) && defined(RT_OS_WINDOWS)
2333	case VMMR0_DO_NEM_INIT_VM:
2334	if (u64Arg \|\| pReqHdr \|\| idCpu != 0)
2335	return VERR_INVALID_PARAMETER;
2336	rc = NEMR0InitVM(pGVM, pVM);
2337	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2338	break;
2339
2340	case VMMR0_DO_NEM_INIT_VM_PART_2:
2341	if (u64Arg \|\| pReqHdr \|\| idCpu != 0)
2342	return VERR_INVALID_PARAMETER;
2343	rc = NEMR0InitVMPart2(pGVM, pVM);
2344	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2345	break;
2346
2347	case VMMR0_DO_NEM_MAP_PAGES:
2348	if (u64Arg \|\| pReqHdr \|\| idCpu == NIL_VMCPUID)
2349	return VERR_INVALID_PARAMETER;
2350	rc = NEMR0MapPages(pGVM, pVM, idCpu);
2351	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2352	break;
2353
2354	case VMMR0_DO_NEM_UNMAP_PAGES:
2355	if (u64Arg \|\| pReqHdr \|\| idCpu == NIL_VMCPUID)
2356	return VERR_INVALID_PARAMETER;
2357	rc = NEMR0UnmapPages(pGVM, pVM, idCpu);
2358	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2359	break;
2360
2361	case VMMR0_DO_NEM_EXPORT_STATE:
2362	if (u64Arg \|\| pReqHdr \|\| idCpu == NIL_VMCPUID)
2363	return VERR_INVALID_PARAMETER;
2364	rc = NEMR0ExportState(pGVM, pVM, idCpu);
2365	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2366	break;
2367
2368	case VMMR0_DO_NEM_IMPORT_STATE:
2369	if (pReqHdr \|\| idCpu == NIL_VMCPUID)
2370	return VERR_INVALID_PARAMETER;
2371	rc = NEMR0ImportState(pGVM, pVM, idCpu, u64Arg);
2372	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2373	break;
2374
2375	case VMMR0_DO_NEM_QUERY_CPU_TICK:
2376	if (u64Arg \|\| pReqHdr \|\| idCpu == NIL_VMCPUID)
2377	return VERR_INVALID_PARAMETER;
2378	rc = NEMR0QueryCpuTick(pGVM, pVM, idCpu);
2379	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2380	break;
2381
2382	case VMMR0_DO_NEM_RESUME_CPU_TICK_ON_ALL:
2383	if (pReqHdr \|\| idCpu == NIL_VMCPUID)
2384	return VERR_INVALID_PARAMETER;
2385	rc = NEMR0ResumeCpuTickOnAll(pGVM, pVM, idCpu, u64Arg);
2386	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2387	break;
2388
2389	case VMMR0_DO_NEM_UPDATE_STATISTICS:
2390	if (u64Arg \|\| pReqHdr)
2391	return VERR_INVALID_PARAMETER;
2392	rc = NEMR0UpdateStatistics(pGVM, pVM, idCpu);
2393	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2394	break;
2395
2396	# if 1 && defined(DEBUG_bird)
2397	case VMMR0_DO_NEM_EXPERIMENT:
2398	if (pReqHdr)
2399	return VERR_INVALID_PARAMETER;
2400	rc = NEMR0DoExperiment(pGVM, pVM, idCpu, u64Arg);
2401	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2402	break;
2403	# endif
2404	# endif
2405	#endif
2406
2407	/*
2408	* For profiling.
2409	*/
2410	case VMMR0_DO_NOP:
2411	case VMMR0_DO_SLOW_NOP:
2412	return VINF_SUCCESS;
2413
2414	/*
2415	* For testing Ring-0 APIs invoked in this environment.
2416	*/
2417	case VMMR0_DO_TESTS:
2418	/** @todo make new test */
2419	return VINF_SUCCESS;
2420
2421
2422	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS)
2423	case VMMR0_DO_TEST_SWITCHER3264:
2424	if (idCpu == NIL_VMCPUID)
2425	return VERR_INVALID_CPU_ID;
2426	rc = HMR0TestSwitcher3264(pVM);
2427	VMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2428	break;
2429	#endif
2430	default:
2431	/*
2432	* We're returning VERR_NOT_SUPPORT here so we've got something else
2433	* than -1 which the interrupt gate glue code might return.
2434	*/
2435	Log(("operation %#x is not supported\n", enmOperation));
2436	return VERR_NOT_SUPPORTED;
2437	}
2438	return rc;
2439	}
2440
2441
2442	/**
2443	* Argument for vmmR0EntryExWrapper containing the arguments for VMMR0EntryEx.
2444	*/
2445	typedef struct VMMR0ENTRYEXARGS
2446	{
2447	PGVM pGVM;
2448	PVM pVM;
2449	VMCPUID idCpu;
2450	VMMR0OPERATION enmOperation;
2451	PSUPVMMR0REQHDR pReq;
2452	uint64_t u64Arg;
2453	PSUPDRVSESSION pSession;
2454	} VMMR0ENTRYEXARGS;
2455	/** Pointer to a vmmR0EntryExWrapper argument package. */
2456	typedef VMMR0ENTRYEXARGS *PVMMR0ENTRYEXARGS;
2457
2458	/**
2459	* This is just a longjmp wrapper function for VMMR0EntryEx calls.
2460	*
2461	* @returns VBox status code.
2462	* @param pvArgs The argument package
2463	*/
2464	static DECLCALLBACK(int) vmmR0EntryExWrapper(void *pvArgs)
2465	{
2466	return vmmR0EntryExWorker(((PVMMR0ENTRYEXARGS)pvArgs)->pGVM,
2467	((PVMMR0ENTRYEXARGS)pvArgs)->pVM,
2468	((PVMMR0ENTRYEXARGS)pvArgs)->idCpu,
2469	((PVMMR0ENTRYEXARGS)pvArgs)->enmOperation,
2470	((PVMMR0ENTRYEXARGS)pvArgs)->pReq,
2471	((PVMMR0ENTRYEXARGS)pvArgs)->u64Arg,
2472	((PVMMR0ENTRYEXARGS)pvArgs)->pSession);
2473	}
2474
2475
2476	/**
2477	* The Ring 0 entry point, called by the support library (SUP).
2478	*
2479	* @returns VBox status code.
2480	* @param pGVM The global (ring-0) VM structure.
2481	* @param pVM The cross context VM structure.
2482	* @param idCpu Virtual CPU ID argument. Must be NIL_VMCPUID if pVM
2483	* is NIL_RTR0PTR, and may be NIL_VMCPUID if it isn't
2484	* @param enmOperation Which operation to execute.
2485	* @param pReq Pointer to the SUPVMMR0REQHDR packet. Optional.
2486	* @param u64Arg Some simple constant argument.
2487	* @param pSession The session of the caller.
2488	* @remarks Assume called with interrupts _enabled_.
2489	*/
2490	VMMR0DECL(int) VMMR0EntryEx(PGVM pGVM, PVM pVM, VMCPUID idCpu, VMMR0OPERATION enmOperation,
2491	PSUPVMMR0REQHDR pReq, uint64_t u64Arg, PSUPDRVSESSION pSession)
2492	{
2493	/*
2494	* Requests that should only happen on the EMT thread will be
2495	* wrapped in a setjmp so we can assert without causing trouble.
2496	*/
2497	if ( pVM != NULL
2498	&& pGVM != NULL
2499	&& idCpu < pGVM->cCpus
2500	&& pVM->pVMR0 != NULL)
2501	{
2502	switch (enmOperation)
2503	{
2504	/* These might/will be called before VMMR3Init. */
2505	case VMMR0_DO_GMM_INITIAL_RESERVATION:
2506	case VMMR0_DO_GMM_UPDATE_RESERVATION:
2507	case VMMR0_DO_GMM_ALLOCATE_PAGES:
2508	case VMMR0_DO_GMM_FREE_PAGES:
2509	case VMMR0_DO_GMM_BALLOONED_PAGES:
2510	/* On the mac we might not have a valid jmp buf, so check these as well. */
2511	case VMMR0_DO_VMMR0_INIT:
2512	case VMMR0_DO_VMMR0_TERM:
2513	{
2514	PGVMCPU pGVCpu = &pGVM->aCpus[idCpu];
2515	PVMCPU pVCpu = &pVM->aCpus[idCpu];
2516	RTNATIVETHREAD hNativeThread = RTThreadNativeSelf();
2517	if (RT_LIKELY( pGVCpu->hEMT == hNativeThread
2518	&& pVCpu->hNativeThreadR0 == hNativeThread))
2519	{
2520	if (!pVCpu->vmm.s.CallRing3JmpBufR0.pvSavedStack)
2521	break;
2522
2523	/** @todo validate this EMT claim... GVM knows. */
2524	VMMR0ENTRYEXARGS Args;
2525	Args.pGVM = pGVM;
2526	Args.pVM = pVM;
2527	Args.idCpu = idCpu;
2528	Args.enmOperation = enmOperation;
2529	Args.pReq = pReq;
2530	Args.u64Arg = u64Arg;
2531	Args.pSession = pSession;
2532	return vmmR0CallRing3SetJmpEx(&pVCpu->vmm.s.CallRing3JmpBufR0, vmmR0EntryExWrapper, &Args);
2533	}
2534	return VERR_VM_THREAD_NOT_EMT;
2535	}
2536
2537	default:
2538	break;
2539	}
2540	}
2541	return vmmR0EntryExWorker(pGVM, pVM, idCpu, enmOperation, pReq, u64Arg, pSession);
2542	}
2543
2544
2545	/**
2546	* Checks whether we've armed the ring-0 long jump machinery.
2547	*
2548	* @returns @c true / @c false
2549	* @param pVCpu The cross context virtual CPU structure.
2550	* @thread EMT
2551	* @sa VMMIsLongJumpArmed
2552	*/
2553	VMMR0_INT_DECL(bool) VMMR0IsLongJumpArmed(PVMCPU pVCpu)
2554	{
2555	#ifdef RT_ARCH_X86
2556	return pVCpu->vmm.s.CallRing3JmpBufR0.eip
2557	&& !pVCpu->vmm.s.CallRing3JmpBufR0.fInRing3Call;
2558	#else
2559	return pVCpu->vmm.s.CallRing3JmpBufR0.rip
2560	&& !pVCpu->vmm.s.CallRing3JmpBufR0.fInRing3Call;
2561	#endif
2562	}
2563
2564
2565	/**
2566	* Checks whether we've done a ring-3 long jump.
2567	*
2568	* @returns @c true / @c false
2569	* @param pVCpu The cross context virtual CPU structure.
2570	* @thread EMT
2571	*/
2572	VMMR0_INT_DECL(bool) VMMR0IsInRing3LongJump(PVMCPU pVCpu)
2573	{
2574	return pVCpu->vmm.s.CallRing3JmpBufR0.fInRing3Call;
2575	}
2576
2577
2578	/**
2579	* Internal R0 logger worker: Flush logger.
2580	*
2581	* @param pLogger The logger instance to flush.
2582	* @remark This function must be exported!
2583	*/
2584	VMMR0DECL(void) vmmR0LoggerFlush(PRTLOGGER pLogger)
2585	{
2586	#ifdef LOG_ENABLED
2587	/*
2588	* Convert the pLogger into a VM handle and 'call' back to Ring-3.
2589	* (This is a bit paranoid code.)
2590	*/
2591	PVMMR0LOGGER pR0Logger = (PVMMR0LOGGER)((uintptr_t)pLogger - RT_UOFFSETOF(VMMR0LOGGER, Logger));
2592	if ( !VALID_PTR(pR0Logger)
2593	\|\| !VALID_PTR(pR0Logger + 1)
2594	\|\| pLogger->u32Magic != RTLOGGER_MAGIC)
2595	{
2596	# ifdef DEBUG
2597	SUPR0Printf("vmmR0LoggerFlush: pLogger=%p!\n", pLogger);
2598	# endif
2599	return;
2600	}
2601	if (pR0Logger->fFlushingDisabled)
2602	return; /* quietly */
2603
2604	PVM pVM = pR0Logger->pVM;
2605	if ( !VALID_PTR(pVM)
2606	\|\| pVM->pVMR0 != pVM)
2607	{
2608	# ifdef DEBUG
2609	SUPR0Printf("vmmR0LoggerFlush: pVM=%p! pVMR0=%p! pLogger=%p\n", pVM, pVM->pVMR0, pLogger);
2610	# endif
2611	return;
2612	}
2613
2614	PVMCPU pVCpu = VMMGetCpu(pVM);
2615	if (pVCpu)
2616	{
2617	/*
2618	* Check that the jump buffer is armed.
2619	*/
2620	# ifdef RT_ARCH_X86
2621	if ( !pVCpu->vmm.s.CallRing3JmpBufR0.eip
2622	\|\| pVCpu->vmm.s.CallRing3JmpBufR0.fInRing3Call)
2623	# else
2624	if ( !pVCpu->vmm.s.CallRing3JmpBufR0.rip
2625	\|\| pVCpu->vmm.s.CallRing3JmpBufR0.fInRing3Call)
2626	# endif
2627	{
2628	# ifdef DEBUG
2629	SUPR0Printf("vmmR0LoggerFlush: Jump buffer isn't armed!\n");
2630	# endif
2631	return;
2632	}
2633	VMMRZCallRing3(pVM, pVCpu, VMMCALLRING3_VMM_LOGGER_FLUSH, 0);
2634	}
2635	# ifdef DEBUG
2636	else
2637	SUPR0Printf("vmmR0LoggerFlush: invalid VCPU context!\n");
2638	# endif
2639	#else
2640	NOREF(pLogger);
2641	#endif /* LOG_ENABLED */
2642	}
2643
2644	#ifdef LOG_ENABLED
2645
2646	/**
2647	* Disables flushing of the ring-0 debug log.
2648	*
2649	* @param pVCpu The cross context virtual CPU structure.
2650	*/
2651	VMMR0_INT_DECL(void) VMMR0LogFlushDisable(PVMCPU pVCpu)
2652	{
2653	if (pVCpu->vmm.s.pR0LoggerR0)
2654	pVCpu->vmm.s.pR0LoggerR0->fFlushingDisabled = true;
2655	if (pVCpu->vmm.s.pR0RelLoggerR0)
2656	pVCpu->vmm.s.pR0RelLoggerR0->fFlushingDisabled = true;
2657	}
2658
2659
2660	/**
2661	* Enables flushing of the ring-0 debug log.
2662	*
2663	* @param pVCpu The cross context virtual CPU structure.
2664	*/
2665	VMMR0_INT_DECL(void) VMMR0LogFlushEnable(PVMCPU pVCpu)
2666	{
2667	if (pVCpu->vmm.s.pR0LoggerR0)
2668	pVCpu->vmm.s.pR0LoggerR0->fFlushingDisabled = false;
2669	if (pVCpu->vmm.s.pR0RelLoggerR0)
2670	pVCpu->vmm.s.pR0RelLoggerR0->fFlushingDisabled = false;
2671	}
2672
2673
2674	/**
2675	* Checks if log flushing is disabled or not.
2676	*
2677	* @param pVCpu The cross context virtual CPU structure.
2678	*/
2679	VMMR0_INT_DECL(bool) VMMR0IsLogFlushDisabled(PVMCPU pVCpu)
2680	{
2681	if (pVCpu->vmm.s.pR0LoggerR0)
2682	return pVCpu->vmm.s.pR0LoggerR0->fFlushingDisabled;
2683	if (pVCpu->vmm.s.pR0RelLoggerR0)
2684	return pVCpu->vmm.s.pR0RelLoggerR0->fFlushingDisabled;
2685	return true;
2686	}
2687
2688	#endif /* LOG_ENABLED */
2689
2690	/**
2691	* Override RTLogRelGetDefaultInstanceEx so we can do LogRel to VBox.log from EMTs in ring-0.
2692	*/
2693	DECLEXPORT(PRTLOGGER) RTLogRelGetDefaultInstanceEx(uint32_t fFlagsAndGroup)
2694	{
2695	PGVMCPU pGVCpu = GVMMR0GetGVCpuByEMT(NIL_RTNATIVETHREAD);
2696	if (pGVCpu)
2697	{
2698	PVMCPU pVCpu = pGVCpu->pVCpu;
2699	if (RT_VALID_PTR(pVCpu))
2700	{
2701	PVMMR0LOGGER pVmmLogger = pVCpu->vmm.s.pR0RelLoggerR0;
2702	if (RT_VALID_PTR(pVmmLogger))
2703	{
2704	if ( pVmmLogger->fCreated
2705	&& pVmmLogger->pVM == pGVCpu->pVM)
2706	{
2707	if (pVmmLogger->Logger.fFlags & RTLOGFLAGS_DISABLED)
2708	return NULL;
2709	uint16_t const fFlags = RT_LO_U16(fFlagsAndGroup);
2710	uint16_t const iGroup = RT_HI_U16(fFlagsAndGroup);
2711	if ( iGroup != UINT16_MAX
2712	&& ( ( pVmmLogger->Logger.afGroups[iGroup < pVmmLogger->Logger.cGroups ? iGroup : 0]
2713	& (fFlags \| (uint32_t)RTLOGGRPFLAGS_ENABLED))
2714	!= (fFlags \| (uint32_t)RTLOGGRPFLAGS_ENABLED)))
2715	return NULL;
2716	return &pVmmLogger->Logger;
2717	}
2718	}
2719	}
2720	}
2721	return SUPR0GetDefaultLogRelInstanceEx(fFlagsAndGroup);
2722	}
2723
2724
2725	/**
2726	* Jump back to ring-3 if we're the EMT and the longjmp is armed.
2727	*
2728	* @returns true if the breakpoint should be hit, false if it should be ignored.
2729	*/
2730	DECLEXPORT(bool) RTCALL RTAssertShouldPanic(void)
2731	{
2732	#if 0
2733	return true;
2734	#else
2735	PVM pVM = GVMMR0GetVMByEMT(NIL_RTNATIVETHREAD);
2736	if (pVM)
2737	{
2738	PVMCPU pVCpu = VMMGetCpu(pVM);
2739
2740	if (pVCpu)
2741	{
2742	#ifdef RT_ARCH_X86
2743	if ( pVCpu->vmm.s.CallRing3JmpBufR0.eip
2744	&& !pVCpu->vmm.s.CallRing3JmpBufR0.fInRing3Call)
2745	#else
2746	if ( pVCpu->vmm.s.CallRing3JmpBufR0.rip
2747	&& !pVCpu->vmm.s.CallRing3JmpBufR0.fInRing3Call)
2748	#endif
2749	{
2750	int rc = VMMRZCallRing3(pVM, pVCpu, VMMCALLRING3_VM_R0_ASSERTION, 0);
2751	return RT_FAILURE_NP(rc);
2752	}
2753	}
2754	}
2755	#ifdef RT_OS_LINUX
2756	return true;
2757	#else
2758	return false;
2759	#endif
2760	#endif
2761	}
2762
2763
2764	/**
2765	* Override this so we can push it up to ring-3.
2766	*
2767	* @param pszExpr Expression. Can be NULL.
2768	* @param uLine Location line number.
2769	* @param pszFile Location file name.
2770	* @param pszFunction Location function name.
2771	*/
2772	DECLEXPORT(void) RTCALL RTAssertMsg1Weak(const char pszExpr, unsigned uLine, const char pszFile, const char *pszFunction)
2773	{
2774	/*
2775	* To the log.
2776	*/
2777	LogAlways(("\n!!R0-Assertion Failed!!\n"
2778	"Expression: %s\n"
2779	"Location : %s(%d) %s\n",
2780	pszExpr, pszFile, uLine, pszFunction));
2781
2782	/*
2783	* To the global VMM buffer.
2784	*/
2785	PVM pVM = GVMMR0GetVMByEMT(NIL_RTNATIVETHREAD);
2786	if (pVM)
2787	RTStrPrintf(pVM->vmm.s.szRing0AssertMsg1, sizeof(pVM->vmm.s.szRing0AssertMsg1),
2788	"\n!!R0-Assertion Failed!!\n"
2789	"Expression: %.*s\n"
2790	"Location : %s(%d) %s\n",
2791	sizeof(pVM->vmm.s.szRing0AssertMsg1) / 4 * 3, pszExpr,
2792	pszFile, uLine, pszFunction);
2793
2794	/*
2795	* Continue the normal way.
2796	*/
2797	RTAssertMsg1(pszExpr, uLine, pszFile, pszFunction);
2798	}
2799
2800
2801	/**
2802	* Callback for RTLogFormatV which writes to the ring-3 log port.
2803	* See PFNLOGOUTPUT() for details.
2804	*/
2805	static DECLCALLBACK(size_t) rtLogOutput(void pv, const char pachChars, size_t cbChars)
2806	{
2807	for (size_t i = 0; i < cbChars; i++)
2808	{
2809	LogAlways(("%c", pachChars[i])); NOREF(pachChars);
2810	}
2811
2812	NOREF(pv);
2813	return cbChars;
2814	}
2815
2816
2817	/**
2818	* Override this so we can push it up to ring-3.
2819	*
2820	* @param pszFormat The format string.
2821	* @param va Arguments.
2822	*/
2823	DECLEXPORT(void) RTCALL RTAssertMsg2WeakV(const char *pszFormat, va_list va)
2824	{
2825	va_list vaCopy;
2826
2827	/*
2828	* Push the message to the loggers.
2829	*/
2830	PRTLOGGER pLog = RTLogGetDefaultInstance(); /* Don't initialize it here... */
2831	if (pLog)
2832	{
2833	va_copy(vaCopy, va);
2834	RTLogFormatV(rtLogOutput, pLog, pszFormat, vaCopy);
2835	va_end(vaCopy);
2836	}
2837	pLog = RTLogRelGetDefaultInstance();
2838	if (pLog)
2839	{
2840	va_copy(vaCopy, va);
2841	RTLogFormatV(rtLogOutput, pLog, pszFormat, vaCopy);
2842	va_end(vaCopy);
2843	}
2844
2845	/*
2846	* Push it to the global VMM buffer.
2847	*/
2848	PVM pVM = GVMMR0GetVMByEMT(NIL_RTNATIVETHREAD);
2849	if (pVM)
2850	{
2851	va_copy(vaCopy, va);
2852	RTStrPrintfV(pVM->vmm.s.szRing0AssertMsg2, sizeof(pVM->vmm.s.szRing0AssertMsg2), pszFormat, vaCopy);
2853	va_end(vaCopy);
2854	}
2855
2856	/*
2857	* Continue the normal way.
2858	*/
2859	RTAssertMsg2V(pszFormat, va);
2860	}
2861

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

source: vbox/trunk/src/VBox/VMM/VMMR0/VMMR0.cpp@ 78003

Download in other formats: