HMSVMR0.cpp@ 52023

Last change on this file since 52023 was 52009, checked in by vboxsync, 10 years ago
Comments and nits.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 194.5 KB

Line
1	/* $Id: HMSVMR0.cpp 52009 2014-07-14 05:27:31Z vboxsync $ */
2	/** @file
3	* HM SVM (AMD-V) - Host Context Ring-0.
4	*/
5
6	/*
7	* Copyright (C) 2013-2014 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	* Header Files *
20	*******************************************************************************/
21	#define LOG_GROUP LOG_GROUP_HM
22	#include <iprt/asm-amd64-x86.h>
23	#include <iprt/thread.h>
24
25	#include "HMInternal.h"
26	#include <VBox/vmm/vm.h>
27	#include "HMSVMR0.h"
28	#include <VBox/vmm/pdmapi.h>
29	#include <VBox/vmm/dbgf.h>
30	#include <VBox/vmm/iom.h>
31	#include <VBox/vmm/tm.h>
32	#include <VBox/vmm/gim.h>
33
34	#ifdef DEBUG_ramshankar
35	# define HMSVM_SYNC_FULL_GUEST_STATE
36	# define HMSVM_ALWAYS_TRAP_ALL_XCPTS
37	# define HMSVM_ALWAYS_TRAP_PF
38	# define HMSVM_ALWAYS_TRAP_TASK_SWITCH
39	#endif
40
41
42	/*******************************************************************************
43	* Defined Constants And Macros *
44	*******************************************************************************/
45	#ifdef VBOX_WITH_STATISTICS
46	# define HMSVM_EXITCODE_STAM_COUNTER_INC(u64ExitCode) do { \
47	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitAll); \
48	if ((u64ExitCode) == SVM_EXIT_NPF) \
49	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitReasonNpf); \
50	else \
51	STAM_COUNTER_INC(&pVCpu->hm.s.paStatExitReasonR0[(u64ExitCode) & MASK_EXITREASON_STAT]); \
52	} while (0)
53	#else
54	# define HMSVM_EXITCODE_STAM_COUNTER_INC(u64ExitCode) do { } while (0)
55	#endif
56
57	/** If we decide to use a function table approach this can be useful to
58	* switch to a "static DECLCALLBACK(int)". */
59	#define HMSVM_EXIT_DECL static int
60
61	/** @name Segment attribute conversion between CPU and AMD-V VMCB format.
62	*
63	* The CPU format of the segment attribute is described in X86DESCATTRBITS
64	* which is 16-bits (i.e. includes 4 bits of the segment limit).
65	*
66	* The AMD-V VMCB format the segment attribute is compact 12-bits (strictly
67	* only the attribute bits and nothing else). Upper 4-bits are unused.
68	*
69	* @{ */
70	#define HMSVM_CPU_2_VMCB_SEG_ATTR(a) ( ((a) & 0xff) \| (((a) & 0xf000) >> 4) )
71	#define HMSVM_VMCB_2_CPU_SEG_ATTR(a) ( ((a) & 0xff) \| (((a) & 0x0f00) << 4) )
72	/** @} */
73
74	/** @name Macros for loading, storing segment registers to/from the VMCB.
75	* @{ */
76	#define HMSVM_LOAD_SEG_REG(REG, reg) \
77	do \
78	{ \
79	Assert(pCtx->reg.fFlags & CPUMSELREG_FLAGS_VALID); \
80	Assert(pCtx->reg.ValidSel == pCtx->reg.Sel); \
81	pVmcb->guest.REG.u16Sel = pCtx->reg.Sel; \
82	pVmcb->guest.REG.u32Limit = pCtx->reg.u32Limit; \
83	pVmcb->guest.REG.u64Base = pCtx->reg.u64Base; \
84	pVmcb->guest.REG.u16Attr = HMSVM_CPU_2_VMCB_SEG_ATTR(pCtx->reg.Attr.u); \
85	} while (0)
86
87	#define HMSVM_SAVE_SEG_REG(REG, reg) \
88	do \
89	{ \
90	pMixedCtx->reg.Sel = pVmcb->guest.REG.u16Sel; \
91	pMixedCtx->reg.ValidSel = pVmcb->guest.REG.u16Sel; \
92	pMixedCtx->reg.fFlags = CPUMSELREG_FLAGS_VALID; \
93	pMixedCtx->reg.u32Limit = pVmcb->guest.REG.u32Limit; \
94	pMixedCtx->reg.u64Base = pVmcb->guest.REG.u64Base; \
95	pMixedCtx->reg.Attr.u = HMSVM_VMCB_2_CPU_SEG_ATTR(pVmcb->guest.REG.u16Attr); \
96	} while (0)
97	/** @} */
98
99	/** Macro for checking and returning from the using function for
100	* \#VMEXIT intercepts that maybe caused during delivering of another
101	* event in the guest. */
102	#define HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY() \
103	do \
104	{ \
105	int rc = hmR0SvmCheckExitDueToEventDelivery(pVCpu, pCtx, pSvmTransient); \
106	if (RT_UNLIKELY(rc == VINF_HM_DOUBLE_FAULT)) \
107	return VINF_SUCCESS; \
108	else if (RT_UNLIKELY(rc == VINF_EM_RESET)) \
109	return rc; \
110	} while (0)
111
112	/** Macro for upgrading a @a a_rc to VINF_EM_DBG_STEPPED after emulating an
113	* instruction that exited. */
114	#define HMSVM_CHECK_SINGLE_STEP(a_pVCpu, a_rc) \
115	do { \
116	if ((a_pVCpu)->hm.s.fSingleInstruction && (a_rc) == VINF_SUCCESS) \
117	(a_rc) = VINF_EM_DBG_STEPPED; \
118	} while (0)
119
120	/** Assert that preemption is disabled or covered by thread-context hooks. */
121	#define HMSVM_ASSERT_PREEMPT_SAFE() Assert( VMMR0ThreadCtxHooksAreRegistered(pVCpu) \
122	\|\| !RTThreadPreemptIsEnabled(NIL_RTTHREAD));
123
124	/** Assert that we haven't migrated CPUs when thread-context hooks are not
125	* used. */
126	#define HMSVM_ASSERT_CPU_SAFE() AssertMsg( VMMR0ThreadCtxHooksAreRegistered(pVCpu) \
127	\|\| pVCpu->hm.s.idEnteredCpu == RTMpCpuId(), \
128	("Illegal migration! Entered on CPU %u Current %u\n", \
129	pVCpu->hm.s.idEnteredCpu, RTMpCpuId()));
130
131	/** Exception bitmap mask for all contributory exceptions.
132	*
133	* Page fault is deliberately excluded here as it's conditional as to whether
134	* it's contributory or benign. Page faults are handled separately.
135	*/
136	#define HMSVM_CONTRIBUTORY_XCPT_MASK ( RT_BIT(X86_XCPT_GP) \| RT_BIT(X86_XCPT_NP) \| RT_BIT(X86_XCPT_SS) \| RT_BIT(X86_XCPT_TS) \
137	\| RT_BIT(X86_XCPT_DE))
138
139	/** @name VMCB Clean Bits.
140	*
141	* These flags are used for VMCB-state caching. A set VMCB Clean bit indicates
142	* AMD-V doesn't need to reload the corresponding value(s) from the VMCB in
143	* memory.
144	*
145	* @{ */
146	/** All intercepts vectors, TSC offset, PAUSE filter counter. */
147	#define HMSVM_VMCB_CLEAN_INTERCEPTS RT_BIT(0)
148	/** I/O permission bitmap, MSR permission bitmap. */
149	#define HMSVM_VMCB_CLEAN_IOPM_MSRPM RT_BIT(1)
150	/** ASID. */
151	#define HMSVM_VMCB_CLEAN_ASID RT_BIT(2)
152	/** TRP: V_TPR, V_IRQ, V_INTR_PRIO, V_IGN_TPR, V_INTR_MASKING,
153	V_INTR_VECTOR. */
154	#define HMSVM_VMCB_CLEAN_TPR RT_BIT(3)
155	/** Nested Paging: Nested CR3 (nCR3), PAT. */
156	#define HMSVM_VMCB_CLEAN_NP RT_BIT(4)
157	/** Control registers (CR0, CR3, CR4, EFER). */
158	#define HMSVM_VMCB_CLEAN_CRX_EFER RT_BIT(5)
159	/** Debug registers (DR6, DR7). */
160	#define HMSVM_VMCB_CLEAN_DRX RT_BIT(6)
161	/** GDT, IDT limit and base. */
162	#define HMSVM_VMCB_CLEAN_DT RT_BIT(7)
163	/** Segment register: CS, SS, DS, ES limit and base. */
164	#define HMSVM_VMCB_CLEAN_SEG RT_BIT(8)
165	/** CR2.*/
166	#define HMSVM_VMCB_CLEAN_CR2 RT_BIT(9)
167	/** Last-branch record (DbgCtlMsr, br_from, br_to, lastint_from, lastint_to) */
168	#define HMSVM_VMCB_CLEAN_LBR RT_BIT(10)
169	/** AVIC (AVIC APIC_BAR; AVIC APIC_BACKING_PAGE, AVIC
170	PHYSICAL_TABLE and AVIC LOGICAL_TABLE Pointers). */
171	#define HMSVM_VMCB_CLEAN_AVIC RT_BIT(11)
172	/** Mask of all valid VMCB Clean bits. */
173	#define HMSVM_VMCB_CLEAN_ALL ( HMSVM_VMCB_CLEAN_INTERCEPTS \
174	\| HMSVM_VMCB_CLEAN_IOPM_MSRPM \
175	\| HMSVM_VMCB_CLEAN_ASID \
176	\| HMSVM_VMCB_CLEAN_TPR \
177	\| HMSVM_VMCB_CLEAN_NP \
178	\| HMSVM_VMCB_CLEAN_CRX_EFER \
179	\| HMSVM_VMCB_CLEAN_DRX \
180	\| HMSVM_VMCB_CLEAN_DT \
181	\| HMSVM_VMCB_CLEAN_SEG \
182	\| HMSVM_VMCB_CLEAN_CR2 \
183	\| HMSVM_VMCB_CLEAN_LBR \
184	\| HMSVM_VMCB_CLEAN_AVIC)
185	/** @} */
186
187	/** @name SVM transient.
188	*
189	* A state structure for holding miscellaneous information across AMD-V
190	* VMRUN/#VMEXIT operation, restored after the transition.
191	*
192	* @{ */
193	typedef struct SVMTRANSIENT
194	{
195	/** The host's rflags/eflags. */
196	RTCCUINTREG uEflags;
197	#if HC_ARCH_BITS == 32
198	uint32_t u32Alignment0;
199	#endif
200
201	/** The #VMEXIT exit code (the EXITCODE field in the VMCB). */
202	uint64_t u64ExitCode;
203	/** The guest's TPR value used for TPR shadowing. */
204	uint8_t u8GuestTpr;
205	/** Alignment. */
206	uint8_t abAlignment0[7];
207
208	/** Whether the guest FPU state was active at the time of #VMEXIT. */
209	bool fWasGuestFPUStateActive;
210	/** Whether the guest debug state was active at the time of #VMEXIT. */
211	bool fWasGuestDebugStateActive;
212	/** Whether the hyper debug state was active at the time of #VMEXIT. */
213	bool fWasHyperDebugStateActive;
214	/** Whether the TSC offset mode needs to be updated. */
215	bool fUpdateTscOffsetting;
216	/** Whether the TSC_AUX MSR needs restoring on #VMEXIT. */
217	bool fRestoreTscAuxMsr;
218	/** Whether the #VMEXIT was caused by a page-fault during delivery of a
219	* contributary exception or a page-fault. */
220	bool fVectoringPF;
221	} SVMTRANSIENT, *PSVMTRANSIENT;
222	AssertCompileMemberAlignment(SVMTRANSIENT, u64ExitCode, sizeof(uint64_t));
223	AssertCompileMemberAlignment(SVMTRANSIENT, fWasGuestFPUStateActive, sizeof(uint64_t));
224	/** @} */
225
226	/**
227	* MSRPM (MSR permission bitmap) read permissions (for guest RDMSR).
228	*/
229	typedef enum SVMMSREXITREAD
230	{
231	/** Reading this MSR causes a VM-exit. */
232	SVMMSREXIT_INTERCEPT_READ = 0xb,
233	/** Reading this MSR does not cause a VM-exit. */
234	SVMMSREXIT_PASSTHRU_READ
235	} SVMMSREXITREAD;
236
237	/**
238	* MSRPM (MSR permission bitmap) write permissions (for guest WRMSR).
239	*/
240	typedef enum SVMMSREXITWRITE
241	{
242	/** Writing to this MSR causes a VM-exit. */
243	SVMMSREXIT_INTERCEPT_WRITE = 0xd,
244	/** Writing to this MSR does not cause a VM-exit. */
245	SVMMSREXIT_PASSTHRU_WRITE
246	} SVMMSREXITWRITE;
247
248	/**
249	* SVM VM-exit handler.
250	*
251	* @returns VBox status code.
252	* @param pVCpu Pointer to the VMCPU.
253	* @param pMixedCtx Pointer to the guest-CPU context.
254	* @param pSvmTransient Pointer to the SVM-transient structure.
255	*/
256	typedef int FNSVMEXITHANDLER(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
257
258	/*******************************************************************************
259	* Internal Functions *
260	*******************************************************************************/
261	static void hmR0SvmSetMsrPermission(PVMCPU pVCpu, unsigned uMsr, SVMMSREXITREAD enmRead, SVMMSREXITWRITE enmWrite);
262	static void hmR0SvmPendingEventToTrpmTrap(PVMCPU pVCpu);
263	static void hmR0SvmLeave(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx);
264
265	/** @name VM-exit handlers.
266	* @{
267	*/
268	static FNSVMEXITHANDLER hmR0SvmExitIntr;
269	static FNSVMEXITHANDLER hmR0SvmExitWbinvd;
270	static FNSVMEXITHANDLER hmR0SvmExitInvd;
271	static FNSVMEXITHANDLER hmR0SvmExitCpuid;
272	static FNSVMEXITHANDLER hmR0SvmExitRdtsc;
273	static FNSVMEXITHANDLER hmR0SvmExitRdtscp;
274	static FNSVMEXITHANDLER hmR0SvmExitRdpmc;
275	static FNSVMEXITHANDLER hmR0SvmExitInvlpg;
276	static FNSVMEXITHANDLER hmR0SvmExitHlt;
277	static FNSVMEXITHANDLER hmR0SvmExitMonitor;
278	static FNSVMEXITHANDLER hmR0SvmExitMwait;
279	static FNSVMEXITHANDLER hmR0SvmExitShutdown;
280	static FNSVMEXITHANDLER hmR0SvmExitReadCRx;
281	static FNSVMEXITHANDLER hmR0SvmExitWriteCRx;
282	static FNSVMEXITHANDLER hmR0SvmExitSetPendingXcptUD;
283	static FNSVMEXITHANDLER hmR0SvmExitMsr;
284	static FNSVMEXITHANDLER hmR0SvmExitReadDRx;
285	static FNSVMEXITHANDLER hmR0SvmExitWriteDRx;
286	static FNSVMEXITHANDLER hmR0SvmExitIOInstr;
287	static FNSVMEXITHANDLER hmR0SvmExitNestedPF;
288	static FNSVMEXITHANDLER hmR0SvmExitVIntr;
289	static FNSVMEXITHANDLER hmR0SvmExitTaskSwitch;
290	static FNSVMEXITHANDLER hmR0SvmExitVmmCall;
291	static FNSVMEXITHANDLER hmR0SvmExitXcptPF;
292	static FNSVMEXITHANDLER hmR0SvmExitXcptNM;
293	static FNSVMEXITHANDLER hmR0SvmExitXcptMF;
294	static FNSVMEXITHANDLER hmR0SvmExitXcptDB;
295	/** @} */
296
297	DECLINLINE(int) hmR0SvmHandleExit(PVMCPU pVCpu, PCPUMCTX pMixedCtx, PSVMTRANSIENT pSvmTransient);
298
299	/*******************************************************************************
300	* Global Variables *
301	*******************************************************************************/
302	/** Ring-0 memory object for the IO bitmap. */
303	RTR0MEMOBJ g_hMemObjIOBitmap = NIL_RTR0MEMOBJ;
304	/** Physical address of the IO bitmap. */
305	RTHCPHYS g_HCPhysIOBitmap = 0;
306	/** Virtual address of the IO bitmap. */
307	R0PTRTYPE(void *) g_pvIOBitmap = NULL;
308
309
310	/**
311	* Sets up and activates AMD-V on the current CPU.
312	*
313	* @returns VBox status code.
314	* @param pCpu Pointer to the CPU info struct.
315	* @param pVM Pointer to the VM (can be NULL after a resume!).
316	* @param pvCpuPage Pointer to the global CPU page.
317	* @param HCPhysCpuPage Physical address of the global CPU page.
318	* @param fEnabledByHost Whether the host OS has already initialized AMD-V.
319	* @param pvArg Unused on AMD-V.
320	*/
321	VMMR0DECL(int) SVMR0EnableCpu(PHMGLOBALCPUINFO pCpu, PVM pVM, void *pvCpuPage, RTHCPHYS HCPhysCpuPage, bool fEnabledByHost,
322	void *pvArg)
323	{
324	Assert(!fEnabledByHost);
325	Assert(HCPhysCpuPage && HCPhysCpuPage != NIL_RTHCPHYS);
326	Assert(RT_ALIGN_T(HCPhysCpuPage, _4K, RTHCPHYS) == HCPhysCpuPage);
327	Assert(pvCpuPage);
328	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
329
330	NOREF(pvArg);
331	NOREF(fEnabledByHost);
332
333	/* Paranoid: Disable interrupt as, in theory, interrupt handlers might mess with EFER. */
334	RTCCUINTREG uEflags = ASMIntDisableFlags();
335
336	/*
337	* We must turn on AMD-V and setup the host state physical address, as those MSRs are per CPU.
338	*/
339	uint64_t u64HostEfer = ASMRdMsr(MSR_K6_EFER);
340	if (u64HostEfer & MSR_K6_EFER_SVME)
341	{
342	/* If the VBOX_HWVIRTEX_IGNORE_SVM_IN_USE is active, then we blindly use AMD-V. */
343	if ( pVM
344	&& pVM->hm.s.svm.fIgnoreInUseError)
345	{
346	pCpu->fIgnoreAMDVInUseError = true;
347	}
348
349	if (!pCpu->fIgnoreAMDVInUseError)
350	{
351	ASMSetFlags(uEflags);
352	return VERR_SVM_IN_USE;
353	}
354	}
355
356	/* Turn on AMD-V in the EFER MSR. */
357	ASMWrMsr(MSR_K6_EFER, u64HostEfer \| MSR_K6_EFER_SVME);
358
359	/* Write the physical page address where the CPU will store the host state while executing the VM. */
360	ASMWrMsr(MSR_K8_VM_HSAVE_PA, HCPhysCpuPage);
361
362	/* Restore interrupts. */
363	ASMSetFlags(uEflags);
364
365	/*
366	* Theoretically, other hypervisors may have used ASIDs, ideally we should flush all non-zero ASIDs
367	* when enabling SVM. AMD doesn't have an SVM instruction to flush all ASIDs (flushing is done
368	* upon VMRUN). Therefore, just set the fFlushAsidBeforeUse flag which instructs hmR0SvmSetupTLB()
369	* to flush the TLB with before using a new ASID.
370	*/
371	pCpu->fFlushAsidBeforeUse = true;
372
373	/*
374	* Ensure each VCPU scheduled on this CPU gets a new VPID on resume. See @bugref{6255}.
375	*/
376	++pCpu->cTlbFlushes;
377
378	return VINF_SUCCESS;
379	}
380
381
382	/**
383	* Deactivates AMD-V on the current CPU.
384	*
385	* @returns VBox status code.
386	* @param pCpu Pointer to the CPU info struct.
387	* @param pvCpuPage Pointer to the global CPU page.
388	* @param HCPhysCpuPage Physical address of the global CPU page.
389	*/
390	VMMR0DECL(int) SVMR0DisableCpu(PHMGLOBALCPUINFO pCpu, void *pvCpuPage, RTHCPHYS HCPhysCpuPage)
391	{
392	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
393	AssertReturn( HCPhysCpuPage
394	&& HCPhysCpuPage != NIL_RTHCPHYS, VERR_INVALID_PARAMETER);
395	AssertReturn(pvCpuPage, VERR_INVALID_PARAMETER);
396	NOREF(pCpu);
397
398	/* Paranoid: Disable interrupts as, in theory, interrupt handlers might mess with EFER. */
399	RTCCUINTREG uEflags = ASMIntDisableFlags();
400
401	/* Turn off AMD-V in the EFER MSR. */
402	uint64_t u64HostEfer = ASMRdMsr(MSR_K6_EFER);
403	ASMWrMsr(MSR_K6_EFER, u64HostEfer & ~MSR_K6_EFER_SVME);
404
405	/* Invalidate host state physical address. */
406	ASMWrMsr(MSR_K8_VM_HSAVE_PA, 0);
407
408	/* Restore interrupts. */
409	ASMSetFlags(uEflags);
410
411	return VINF_SUCCESS;
412	}
413
414
415	/**
416	* Does global AMD-V initialization (called during module initialization).
417	*
418	* @returns VBox status code.
419	*/
420	VMMR0DECL(int) SVMR0GlobalInit(void)
421	{
422	/*
423	* Allocate 12 KB for the IO bitmap. Since this is non-optional and we always intercept all IO accesses, it's done
424	* once globally here instead of per-VM.
425	*/
426	Assert(g_hMemObjIOBitmap == NIL_RTR0MEMOBJ);
427	int rc = RTR0MemObjAllocCont(&g_hMemObjIOBitmap, 3 << PAGE_SHIFT, false /* fExecutable */);
428	if (RT_FAILURE(rc))
429	return rc;
430
431	g_pvIOBitmap = RTR0MemObjAddress(g_hMemObjIOBitmap);
432	g_HCPhysIOBitmap = RTR0MemObjGetPagePhysAddr(g_hMemObjIOBitmap, 0 /* iPage */);
433
434	/* Set all bits to intercept all IO accesses. */
435	ASMMemFill32(g_pvIOBitmap, 3 << PAGE_SHIFT, UINT32_C(0xffffffff));
436	return VINF_SUCCESS;
437	}
438
439
440	/**
441	* Does global AMD-V termination (called during module termination).
442	*/
443	VMMR0DECL(void) SVMR0GlobalTerm(void)
444	{
445	if (g_hMemObjIOBitmap != NIL_RTR0MEMOBJ)
446	{
447	RTR0MemObjFree(g_hMemObjIOBitmap, true /* fFreeMappings */);
448	g_pvIOBitmap = NULL;
449	g_HCPhysIOBitmap = 0;
450	g_hMemObjIOBitmap = NIL_RTR0MEMOBJ;
451	}
452	}
453
454
455	/**
456	* Frees any allocated per-VCPU structures for a VM.
457	*
458	* @param pVM Pointer to the VM.
459	*/
460	DECLINLINE(void) hmR0SvmFreeStructs(PVM pVM)
461	{
462	for (uint32_t i = 0; i < pVM->cCpus; i++)
463	{
464	PVMCPU pVCpu = &pVM->aCpus[i];
465	AssertPtr(pVCpu);
466
467	if (pVCpu->hm.s.svm.hMemObjVmcbHost != NIL_RTR0MEMOBJ)
468	{
469	RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjVmcbHost, false);
470	pVCpu->hm.s.svm.pvVmcbHost = 0;
471	pVCpu->hm.s.svm.HCPhysVmcbHost = 0;
472	pVCpu->hm.s.svm.hMemObjVmcbHost = NIL_RTR0MEMOBJ;
473	}
474
475	if (pVCpu->hm.s.svm.hMemObjVmcb != NIL_RTR0MEMOBJ)
476	{
477	RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjVmcb, false);
478	pVCpu->hm.s.svm.pvVmcb = 0;
479	pVCpu->hm.s.svm.HCPhysVmcb = 0;
480	pVCpu->hm.s.svm.hMemObjVmcb = NIL_RTR0MEMOBJ;
481	}
482
483	if (pVCpu->hm.s.svm.hMemObjMsrBitmap != NIL_RTR0MEMOBJ)
484	{
485	RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjMsrBitmap, false);
486	pVCpu->hm.s.svm.pvMsrBitmap = 0;
487	pVCpu->hm.s.svm.HCPhysMsrBitmap = 0;
488	pVCpu->hm.s.svm.hMemObjMsrBitmap = NIL_RTR0MEMOBJ;
489	}
490	}
491	}
492
493
494	/**
495	* Does per-VM AMD-V initialization.
496	*
497	* @returns VBox status code.
498	* @param pVM Pointer to the VM.
499	*/
500	VMMR0DECL(int) SVMR0InitVM(PVM pVM)
501	{
502	int rc = VERR_INTERNAL_ERROR_5;
503
504	/*
505	* Check for an AMD CPU erratum which requires us to flush the TLB before every world-switch.
506	*/
507	uint32_t u32Family;
508	uint32_t u32Model;
509	uint32_t u32Stepping;
510	if (HMAmdIsSubjectToErratum170(&u32Family, &u32Model, &u32Stepping))
511	{
512	Log4(("SVMR0InitVM: AMD cpu with erratum 170 family %#x model %#x stepping %#x\n", u32Family, u32Model, u32Stepping));
513	pVM->hm.s.svm.fAlwaysFlushTLB = true;
514	}
515
516	/*
517	* Initialize the R0 memory objects up-front so we can properly cleanup on allocation failures.
518	*/
519	for (VMCPUID i = 0; i < pVM->cCpus; i++)
520	{
521	PVMCPU pVCpu = &pVM->aCpus[i];
522	pVCpu->hm.s.svm.hMemObjVmcbHost = NIL_RTR0MEMOBJ;
523	pVCpu->hm.s.svm.hMemObjVmcb = NIL_RTR0MEMOBJ;
524	pVCpu->hm.s.svm.hMemObjMsrBitmap = NIL_RTR0MEMOBJ;
525	}
526
527	for (VMCPUID i = 0; i < pVM->cCpus; i++)
528	{
529	PVMCPU pVCpu = &pVM->aCpus[i];
530
531	/*
532	* Allocate one page for the host-context VM control block (VMCB). This is used for additional host-state (such as
533	* FS, GS, Kernel GS Base, etc.) apart from the host-state save area specified in MSR_K8_VM_HSAVE_PA.
534	*/
535	rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjVmcbHost, 1 << PAGE_SHIFT, false /* fExecutable */);
536	if (RT_FAILURE(rc))
537	goto failure_cleanup;
538
539	pVCpu->hm.s.svm.pvVmcbHost = RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjVmcbHost);
540	pVCpu->hm.s.svm.HCPhysVmcbHost = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjVmcbHost, 0 /* iPage */);
541	Assert(pVCpu->hm.s.svm.HCPhysVmcbHost < _4G);
542	ASMMemZeroPage(pVCpu->hm.s.svm.pvVmcbHost);
543
544	/*
545	* Allocate one page for the guest-state VMCB.
546	*/
547	rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjVmcb, 1 << PAGE_SHIFT, false /* fExecutable */);
548	if (RT_FAILURE(rc))
549	goto failure_cleanup;
550
551	pVCpu->hm.s.svm.pvVmcb = RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjVmcb);
552	pVCpu->hm.s.svm.HCPhysVmcb = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjVmcb, 0 /* iPage */);
553	Assert(pVCpu->hm.s.svm.HCPhysVmcb < _4G);
554	ASMMemZeroPage(pVCpu->hm.s.svm.pvVmcb);
555
556	/*
557	* Allocate two pages (8 KB) for the MSR permission bitmap. There doesn't seem to be a way to convince
558	* SVM to not require one.
559	*/
560	rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjMsrBitmap, 2 << PAGE_SHIFT, false /* fExecutable */);
561	if (RT_FAILURE(rc))
562	goto failure_cleanup;
563
564	pVCpu->hm.s.svm.pvMsrBitmap = RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjMsrBitmap);
565	pVCpu->hm.s.svm.HCPhysMsrBitmap = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjMsrBitmap, 0 /* iPage */);
566	/* Set all bits to intercept all MSR accesses (changed later on). */
567	ASMMemFill32(pVCpu->hm.s.svm.pvMsrBitmap, 2 << PAGE_SHIFT, UINT32_C(0xffffffff));
568	}
569
570	return VINF_SUCCESS;
571
572	failure_cleanup:
573	hmR0SvmFreeStructs(pVM);
574	return rc;
575	}
576
577
578	/**
579	* Does per-VM AMD-V termination.
580	*
581	* @returns VBox status code.
582	* @param pVM Pointer to the VM.
583	*/
584	VMMR0DECL(int) SVMR0TermVM(PVM pVM)
585	{
586	hmR0SvmFreeStructs(pVM);
587	return VINF_SUCCESS;
588	}
589
590
591	/**
592	* Sets the permission bits for the specified MSR in the MSRPM.
593	*
594	* @param pVCpu Pointer to the VMCPU.
595	* @param uMsr The MSR for which the access permissions are being set.
596	* @param enmRead MSR read permissions.
597	* @param enmWrite MSR write permissions.
598	*/
599	static void hmR0SvmSetMsrPermission(PVMCPU pVCpu, unsigned uMsr, SVMMSREXITREAD enmRead, SVMMSREXITWRITE enmWrite)
600	{
601	unsigned ulBit;
602	uint8_t pbMsrBitmap = (uint8_t )pVCpu->hm.s.svm.pvMsrBitmap;
603
604	/*
605	* Layout:
606	* Byte offset MSR range
607	* 0x000 - 0x7ff 0x00000000 - 0x00001fff
608	* 0x800 - 0xfff 0xc0000000 - 0xc0001fff
609	* 0x1000 - 0x17ff 0xc0010000 - 0xc0011fff
610	* 0x1800 - 0x1fff Reserved
611	*/
612	if (uMsr <= 0x00001FFF)
613	{
614	/* Pentium-compatible MSRs. */
615	ulBit = uMsr * 2;
616	}
617	else if ( uMsr >= 0xC0000000
618	&& uMsr <= 0xC0001FFF)
619	{
620	/* AMD Sixth Generation x86 Processor MSRs. */
621	ulBit = (uMsr - 0xC0000000) * 2;
622	pbMsrBitmap += 0x800;
623	}
624	else if ( uMsr >= 0xC0010000
625	&& uMsr <= 0xC0011FFF)
626	{
627	/* AMD Seventh and Eighth Generation Processor MSRs. */
628	ulBit = (uMsr - 0xC0001000) * 2;
629	pbMsrBitmap += 0x1000;
630	}
631	else
632	{
633	AssertFailed();
634	return;
635	}
636
637	Assert(ulBit < 0x3fff /* 16 * 1024 - 1 */);
638	if (enmRead == SVMMSREXIT_INTERCEPT_READ)
639	ASMBitSet(pbMsrBitmap, ulBit);
640	else
641	ASMBitClear(pbMsrBitmap, ulBit);
642
643	if (enmWrite == SVMMSREXIT_INTERCEPT_WRITE)
644	ASMBitSet(pbMsrBitmap, ulBit + 1);
645	else
646	ASMBitClear(pbMsrBitmap, ulBit + 1);
647
648	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
649	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_IOPM_MSRPM;
650	}
651
652
653	/**
654	* Sets up AMD-V for the specified VM.
655	* This function is only called once per-VM during initalization.
656	*
657	* @returns VBox status code.
658	* @param pVM Pointer to the VM.
659	*/
660	VMMR0DECL(int) SVMR0SetupVM(PVM pVM)
661	{
662	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
663	AssertReturn(pVM, VERR_INVALID_PARAMETER);
664	Assert(pVM->hm.s.svm.fSupported);
665
666	for (VMCPUID i = 0; i < pVM->cCpus; i++)
667	{
668	PVMCPU pVCpu = &pVM->aCpus[i];
669	PSVMVMCB pVmcb = (PSVMVMCB)pVM->aCpus[i].hm.s.svm.pvVmcb;
670
671	AssertMsgReturn(pVmcb, ("Invalid pVmcb for vcpu[%u]\n", i), VERR_SVM_INVALID_PVMCB);
672
673	/* Trap exceptions unconditionally (debug purposes). */
674	#ifdef HMSVM_ALWAYS_TRAP_PF
675	pVmcb->ctrl.u32InterceptException \|= RT_BIT(X86_XCPT_PF);
676	#endif
677	#ifdef HMSVM_ALWAYS_TRAP_ALL_XCPTS
678	/* If you add any exceptions here, make sure to update hmR0SvmHandleExit(). */
679	pVmcb->ctrl.u32InterceptException \|= 0
680	\| RT_BIT(X86_XCPT_BP)
681	\| RT_BIT(X86_XCPT_DB)
682	\| RT_BIT(X86_XCPT_DE)
683	\| RT_BIT(X86_XCPT_NM)
684	\| RT_BIT(X86_XCPT_UD)
685	\| RT_BIT(X86_XCPT_NP)
686	\| RT_BIT(X86_XCPT_SS)
687	\| RT_BIT(X86_XCPT_GP)
688	\| RT_BIT(X86_XCPT_PF)
689	\| RT_BIT(X86_XCPT_MF)
690	;
691	#endif
692
693	/* Set up unconditional intercepts and conditions. */
694	pVmcb->ctrl.u32InterceptCtrl1 = SVM_CTRL1_INTERCEPT_INTR /* External interrupt causes a VM-exit. */
695	\| SVM_CTRL1_INTERCEPT_NMI /* Non-Maskable Interrupts causes a VM-exit. */
696	\| SVM_CTRL1_INTERCEPT_INIT /* INIT signal causes a VM-exit. */
697	\| SVM_CTRL1_INTERCEPT_RDPMC /* RDPMC causes a VM-exit. */
698	\| SVM_CTRL1_INTERCEPT_CPUID /* CPUID causes a VM-exit. */
699	\| SVM_CTRL1_INTERCEPT_RSM /* RSM causes a VM-exit. */
700	\| SVM_CTRL1_INTERCEPT_HLT /* HLT causes a VM-exit. */
701	\| SVM_CTRL1_INTERCEPT_INOUT_BITMAP /* Use the IOPM to cause IOIO VM-exits. */
702	\| SVM_CTRL1_INTERCEPT_MSR_SHADOW /* MSR access not covered by MSRPM causes a VM-exit.*/
703	\| SVM_CTRL1_INTERCEPT_INVLPGA /* INVLPGA causes a VM-exit. */
704	\| SVM_CTRL1_INTERCEPT_SHUTDOWN /* Shutdown events causes a VM-exit. */
705	\| SVM_CTRL1_INTERCEPT_FERR_FREEZE; /* Intercept "freezing" during legacy FPU handling. */
706
707	pVmcb->ctrl.u32InterceptCtrl2 = SVM_CTRL2_INTERCEPT_VMRUN /* VMRUN causes a VM-exit. */
708	\| SVM_CTRL2_INTERCEPT_VMMCALL /* VMMCALL causes a VM-exit. */
709	\| SVM_CTRL2_INTERCEPT_VMLOAD /* VMLOAD causes a VM-exit. */
710	\| SVM_CTRL2_INTERCEPT_VMSAVE /* VMSAVE causes a VM-exit. */
711	\| SVM_CTRL2_INTERCEPT_STGI /* STGI causes a VM-exit. */
712	\| SVM_CTRL2_INTERCEPT_CLGI /* CLGI causes a VM-exit. */
713	\| SVM_CTRL2_INTERCEPT_SKINIT /* SKINIT causes a VM-exit. */
714	\| SVM_CTRL2_INTERCEPT_WBINVD /* WBINVD causes a VM-exit. */
715	\| SVM_CTRL2_INTERCEPT_MONITOR /* MONITOR causes a VM-exit. */
716	\| SVM_CTRL2_INTERCEPT_MWAIT; /* MWAIT causes a VM-exit. */
717
718	/* CR0, CR4 reads must be intercepted, our shadow values are not necessarily the same as the guest's. */
719	pVmcb->ctrl.u16InterceptRdCRx = RT_BIT(0) \| RT_BIT(4);
720
721	/* CR0, CR4 writes must be intercepted for the same reasons as above. */
722	pVmcb->ctrl.u16InterceptWrCRx = RT_BIT(0) \| RT_BIT(4);
723
724	/* Intercept all DRx reads and writes by default. Changed later on. */
725	pVmcb->ctrl.u16InterceptRdDRx = 0xffff;
726	pVmcb->ctrl.u16InterceptWrDRx = 0xffff;
727
728	/* Virtualize masking of INTR interrupts. (reads/writes from/to CR8 go to the V_TPR register) */
729	pVmcb->ctrl.IntCtrl.n.u1VIrqMasking = 1;
730
731	/* Ignore the priority in the TPR. This is necessary for delivering PIC style (ExtInt) interrupts and we currently
732	deliver both PIC and APIC interrupts alike. See hmR0SvmInjectPendingEvent() */
733	pVmcb->ctrl.IntCtrl.n.u1IgnoreTPR = 1;
734
735	/* Set IO and MSR bitmap permission bitmap physical addresses. */
736	pVmcb->ctrl.u64IOPMPhysAddr = g_HCPhysIOBitmap;
737	pVmcb->ctrl.u64MSRPMPhysAddr = pVCpu->hm.s.svm.HCPhysMsrBitmap;
738
739	/* No LBR virtualization. */
740	pVmcb->ctrl.u64LBRVirt = 0;
741
742	/* Initially set all VMCB clean bits to 0 indicating that everything should be loaded from the VMCB in memory. */
743	pVmcb->ctrl.u64VmcbCleanBits = 0;
744
745	/* The host ASID MBZ, for the guest start with 1. */
746	pVmcb->ctrl.TLBCtrl.n.u32ASID = 1;
747
748	/*
749	* Setup the PAT MSR (applicable for Nested Paging only).
750	* The default value should be 0x0007040600070406ULL, but we want to treat all guest memory as WB,
751	* so choose type 6 for all PAT slots.
752	*/
753	pVmcb->guest.u64GPAT = UINT64_C(0x0006060606060606);
754
755	/* Setup Nested Paging. This doesn't change throughout the execution time of the VM. */
756	pVmcb->ctrl.NestedPaging.n.u1NestedPaging = pVM->hm.s.fNestedPaging;
757
758	/* Without Nested Paging, we need additionally intercepts. */
759	if (!pVM->hm.s.fNestedPaging)
760	{
761	/* CR3 reads/writes must be intercepted; our shadow values differ from the guest values. */
762	pVmcb->ctrl.u16InterceptRdCRx \|= RT_BIT(3);
763	pVmcb->ctrl.u16InterceptWrCRx \|= RT_BIT(3);
764
765	/* Intercept INVLPG and task switches (may change CR3, EFLAGS, LDT). */
766	pVmcb->ctrl.u32InterceptCtrl1 \|= SVM_CTRL1_INTERCEPT_INVLPG
767	\| SVM_CTRL1_INTERCEPT_TASK_SWITCH;
768
769	/* Page faults must be intercepted to implement shadow paging. */
770	pVmcb->ctrl.u32InterceptException \|= RT_BIT(X86_XCPT_PF);
771	}
772
773	#ifdef HMSVM_ALWAYS_TRAP_TASK_SWITCH
774	pVmcb->ctrl.u32InterceptCtrl1 \|= SVM_CTRL1_INTERCEPT_TASK_SWITCH;
775	#endif
776
777	/*
778	* The following MSRs are saved/restored automatically during the world-switch.
779	* Don't intercept guest read/write accesses to these MSRs.
780	*/
781	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
782	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_CSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
783	hmR0SvmSetMsrPermission(pVCpu, MSR_K6_STAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
784	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_SF_MASK, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
785	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_FS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
786	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_GS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
787	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_KERNEL_GS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
788	hmR0SvmSetMsrPermission(pVCpu, MSR_IA32_SYSENTER_CS, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
789	hmR0SvmSetMsrPermission(pVCpu, MSR_IA32_SYSENTER_ESP, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
790	hmR0SvmSetMsrPermission(pVCpu, MSR_IA32_SYSENTER_EIP, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
791	}
792
793	return VINF_SUCCESS;
794	}
795
796
797	/**
798	* Invalidates a guest page by guest virtual address.
799	*
800	* @returns VBox status code.
801	* @param pVM Pointer to the VM.
802	* @param pVCpu Pointer to the VMCPU.
803	* @param GCVirt Guest virtual address of the page to invalidate.
804	*/
805	VMMR0DECL(int) SVMR0InvalidatePage(PVM pVM, PVMCPU pVCpu, RTGCPTR GCVirt)
806	{
807	AssertReturn(pVM, VERR_INVALID_PARAMETER);
808	Assert(pVM->hm.s.svm.fSupported);
809
810	bool fFlushPending = pVM->hm.s.svm.fAlwaysFlushTLB \|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_TLB_FLUSH);
811
812	/* Skip it if a TLB flush is already pending. */
813	if (!fFlushPending)
814	{
815	Log4(("SVMR0InvalidatePage %RGv\n", GCVirt));
816
817	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
818	AssertMsgReturn(pVmcb, ("Invalid pVmcb!\n"), VERR_SVM_INVALID_PVMCB);
819
820	#if HC_ARCH_BITS == 32
821	/* If we get a flush in 64-bit guest mode, then force a full TLB flush. INVLPGA takes only 32-bit addresses. */
822	if (CPUMIsGuestInLongMode(pVCpu))
823	VMCPU_FF_SET(pVCpu, VMCPU_FF_TLB_FLUSH);
824	else
825	#endif
826	{
827	SVMR0InvlpgA(GCVirt, pVmcb->ctrl.TLBCtrl.n.u32ASID);
828	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlbInvlpgVirt);
829	}
830	}
831	return VINF_SUCCESS;
832	}
833
834
835	/**
836	* Flushes the appropriate tagged-TLB entries.
837	*
838	* @param pVM Pointer to the VM.
839	* @param pVCpu Pointer to the VMCPU.
840	*/
841	static void hmR0SvmFlushTaggedTlb(PVMCPU pVCpu)
842	{
843	PVM pVM = pVCpu->CTX_SUFF(pVM);
844	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
845	PHMGLOBALCPUINFO pCpu = HMR0GetCurrentCpu();
846
847	/*
848	* Force a TLB flush for the first world switch if the current CPU differs from the one we ran on last.
849	* This can happen both for start & resume due to long jumps back to ring-3.
850	* If the TLB flush count changed, another VM (VCPU rather) has hit the ASID limit while flushing the TLB,
851	* so we cannot reuse the ASIDs without flushing.
852	*/
853	bool fNewAsid = false;
854	Assert(pCpu->idCpu != NIL_RTCPUID);
855	if ( pVCpu->hm.s.idLastCpu != pCpu->idCpu
856	\|\| pVCpu->hm.s.cTlbFlushes != pCpu->cTlbFlushes)
857	{
858	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlbWorldSwitch);
859	pVCpu->hm.s.fForceTLBFlush = true;
860	fNewAsid = true;
861	}
862
863	/* Set TLB flush state as checked until we return from the world switch. */
864	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, true);
865
866	/* Check for explicit TLB shootdowns. */
867	if (VMCPU_FF_TEST_AND_CLEAR(pVCpu, VMCPU_FF_TLB_FLUSH))
868	{
869	pVCpu->hm.s.fForceTLBFlush = true;
870	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlb);
871	}
872
873	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_NOTHING;
874
875	if (pVM->hm.s.svm.fAlwaysFlushTLB)
876	{
877	/*
878	* This is the AMD erratum 170. We need to flush the entire TLB for each world switch. Sad.
879	*/
880	pCpu->uCurrentAsid = 1;
881	pVCpu->hm.s.uCurrentAsid = 1;
882	pVCpu->hm.s.cTlbFlushes = pCpu->cTlbFlushes;
883	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
884
885	/* Clear the VMCB Clean Bit for NP while flushing the TLB. See @bugref{7152}. */
886	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
887	}
888	else if (pVCpu->hm.s.fForceTLBFlush)
889	{
890	/* Clear the VMCB Clean Bit for NP while flushing the TLB. See @bugref{7152}. */
891	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
892
893	if (fNewAsid)
894	{
895	++pCpu->uCurrentAsid;
896	bool fHitASIDLimit = false;
897	if (pCpu->uCurrentAsid >= pVM->hm.s.uMaxAsid)
898	{
899	pCpu->uCurrentAsid = 1; /* Wraparound at 1; host uses 0 */
900	pCpu->cTlbFlushes++; /* All VCPUs that run on this host CPU must use a new VPID. */
901	fHitASIDLimit = true;
902
903	if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_FLUSH_BY_ASID)
904	{
905	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_SINGLE_CONTEXT;
906	pCpu->fFlushAsidBeforeUse = true;
907	}
908	else
909	{
910	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
911	pCpu->fFlushAsidBeforeUse = false;
912	}
913	}
914
915	if ( !fHitASIDLimit
916	&& pCpu->fFlushAsidBeforeUse)
917	{
918	if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_FLUSH_BY_ASID)
919	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_SINGLE_CONTEXT;
920	else
921	{
922	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
923	pCpu->fFlushAsidBeforeUse = false;
924	}
925	}
926
927	pVCpu->hm.s.uCurrentAsid = pCpu->uCurrentAsid;
928	pVCpu->hm.s.idLastCpu = pCpu->idCpu;
929	pVCpu->hm.s.cTlbFlushes = pCpu->cTlbFlushes;
930	}
931	else
932	{
933	if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_FLUSH_BY_ASID)
934	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_SINGLE_CONTEXT;
935	else
936	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
937	}
938
939	pVCpu->hm.s.fForceTLBFlush = false;
940	}
941	/** @todo We never set VMCPU_FF_TLB_SHOOTDOWN anywhere so this path should
942	* not be executed. See hmQueueInvlPage() where it is commented
943	* out. Support individual entry flushing someday. */
944	#if 0
945	else
946	{
947	if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_TLB_SHOOTDOWN))
948	{
949	/* Deal with pending TLB shootdown actions which were queued when we were not executing code. */
950	STAM_COUNTER_INC(&pVCpu->hm.s.StatTlbShootdown);
951	for (uint32_t i = 0; i < pVCpu->hm.s.TlbShootdown.cPages; i++)
952	SVMR0InvlpgA(pVCpu->hm.s.TlbShootdown.aPages[i], pVmcb->ctrl.TLBCtrl.n.u32ASID);
953
954	pVCpu->hm.s.TlbShootdown.cPages = 0;
955	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_TLB_SHOOTDOWN);
956	}
957	}
958	#endif
959
960
961	/* Update VMCB with the ASID. */
962	if (pVmcb->ctrl.TLBCtrl.n.u32ASID != pVCpu->hm.s.uCurrentAsid)
963	{
964	pVmcb->ctrl.TLBCtrl.n.u32ASID = pVCpu->hm.s.uCurrentAsid;
965	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_ASID;
966	}
967
968	AssertMsg(pVCpu->hm.s.idLastCpu == pCpu->idCpu,
969	("vcpu idLastCpu=%x pcpu idCpu=%x\n", pVCpu->hm.s.idLastCpu, pCpu->idCpu));
970	AssertMsg(pVCpu->hm.s.cTlbFlushes == pCpu->cTlbFlushes,
971	("Flush count mismatch for cpu %d (%x vs %x)\n", pCpu->idCpu, pVCpu->hm.s.cTlbFlushes, pCpu->cTlbFlushes));
972	AssertMsg(pCpu->uCurrentAsid >= 1 && pCpu->uCurrentAsid < pVM->hm.s.uMaxAsid,
973	("cpu%d uCurrentAsid = %x\n", pCpu->idCpu, pCpu->uCurrentAsid));
974	AssertMsg(pVCpu->hm.s.uCurrentAsid >= 1 && pVCpu->hm.s.uCurrentAsid < pVM->hm.s.uMaxAsid,
975	("cpu%d VM uCurrentAsid = %x\n", pCpu->idCpu, pVCpu->hm.s.uCurrentAsid));
976
977	#ifdef VBOX_WITH_STATISTICS
978	if (pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_NOTHING)
979	STAM_COUNTER_INC(&pVCpu->hm.s.StatNoFlushTlbWorldSwitch);
980	else if ( pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_SINGLE_CONTEXT
981	\|\| pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_SINGLE_CONTEXT_RETAIN_GLOBALS)
982	{
983	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushAsid);
984	}
985	else
986	{
987	Assert(pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_ENTIRE);
988	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushEntire);
989	}
990	#endif
991	}
992
993
994	/** @name 64-bit guest on 32-bit host OS helper functions.
995	*
996	* The host CPU is still 64-bit capable but the host OS is running in 32-bit
997	* mode (code segment, paging). These wrappers/helpers perform the necessary
998	* bits for the 32->64 switcher.
999	*
1000	* @{ */
1001	#if HC_ARCH_BITS == 32 && defined(VBOX_ENABLE_64_BITS_GUESTS) && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
1002	/**
1003	* Prepares for and executes VMRUN (64-bit guests on a 32-bit host).
1004	*
1005	* @returns VBox status code.
1006	* @param HCPhysVmcbHost Physical address of host VMCB.
1007	* @param HCPhysVmcb Physical address of the VMCB.
1008	* @param pCtx Pointer to the guest-CPU context.
1009	* @param pVM Pointer to the VM.
1010	* @param pVCpu Pointer to the VMCPU.
1011	*/
1012	DECLASM(int) SVMR0VMSwitcherRun64(RTHCPHYS HCPhysVmcbHost, RTHCPHYS HCPhysVmcb, PCPUMCTX pCtx, PVM pVM, PVMCPU pVCpu)
1013	{
1014	uint32_t aParam[4];
1015	aParam[0] = (uint32_t)(HCPhysVmcbHost); /* Param 1: HCPhysVmcbHost - Lo. */
1016	aParam[1] = (uint32_t)(HCPhysVmcbHost >> 32); /* Param 1: HCPhysVmcbHost - Hi. */
1017	aParam[2] = (uint32_t)(HCPhysVmcb); /* Param 2: HCPhysVmcb - Lo. */
1018	aParam[3] = (uint32_t)(HCPhysVmcb >> 32); /* Param 2: HCPhysVmcb - Hi. */
1019
1020	return SVMR0Execute64BitsHandler(pVM, pVCpu, pCtx, HM64ON32OP_SVMRCVMRun64, 4, &aParam[0]);
1021	}
1022
1023
1024	/**
1025	* Executes the specified VMRUN handler in 64-bit mode.
1026	*
1027	* @returns VBox status code.
1028	* @param pVM Pointer to the VM.
1029	* @param pVCpu Pointer to the VMCPU.
1030	* @param pCtx Pointer to the guest-CPU context.
1031	* @param enmOp The operation to perform.
1032	* @param cbParam Number of parameters.
1033	* @param paParam Array of 32-bit parameters.
1034	*/
1035	VMMR0DECL(int) SVMR0Execute64BitsHandler(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, HM64ON32OP enmOp, uint32_t cbParam,
1036	uint32_t *paParam)
1037	{
1038	AssertReturn(pVM->hm.s.pfnHost32ToGuest64R0, VERR_HM_NO_32_TO_64_SWITCHER);
1039	Assert(enmOp > HM64ON32OP_INVALID && enmOp < HM64ON32OP_END);
1040
1041	/* Disable interrupts. */
1042	RTHCUINTREG uOldEFlags = ASMIntDisableFlags();
1043
1044	#ifdef VBOX_WITH_VMMR0_DISABLE_LAPIC_NMI
1045	RTCPUID idHostCpu = RTMpCpuId();
1046	CPUMR0SetLApic(pVCpu, idHostCpu);
1047	#endif
1048
1049	CPUMSetHyperESP(pVCpu, VMMGetStackRC(pVCpu));
1050	CPUMSetHyperEIP(pVCpu, enmOp);
1051	for (int i = (int)cbParam - 1; i >= 0; i--)
1052	CPUMPushHyper(pVCpu, paParam[i]);
1053
1054	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatWorldSwitch3264, z);
1055	/* Call the switcher. */
1056	int rc = pVM->hm.s.pfnHost32ToGuest64R0(pVM, RT_OFFSETOF(VM, aCpus[pVCpu->idCpu].cpum) - RT_OFFSETOF(VM, cpum));
1057	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatWorldSwitch3264, z);
1058
1059	/* Restore interrupts. */
1060	ASMSetFlags(uOldEFlags);
1061	return rc;
1062	}
1063
1064	#endif /* HC_ARCH_BITS == 32 && defined(VBOX_ENABLE_64_BITS_GUESTS) */
1065	/** @} */
1066
1067
1068	/**
1069	* Adds an exception to the intercept exception bitmap in the VMCB and updates
1070	* the corresponding VMCB Clean bit.
1071	*
1072	* @param pVmcb Pointer to the VM control block.
1073	* @param u32Xcpt The value of the exception (X86_XCPT_*).
1074	*/
1075	DECLINLINE(void) hmR0SvmAddXcptIntercept(PSVMVMCB pVmcb, uint32_t u32Xcpt)
1076	{
1077	if (!(pVmcb->ctrl.u32InterceptException & RT_BIT(u32Xcpt)))
1078	{
1079	pVmcb->ctrl.u32InterceptException \|= RT_BIT(u32Xcpt);
1080	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1081	}
1082	}
1083
1084
1085	/**
1086	* Removes an exception from the intercept-exception bitmap in the VMCB and
1087	* updates the corresponding VMCB Clean bit.
1088	*
1089	* @param pVmcb Pointer to the VM control block.
1090	* @param u32Xcpt The value of the exception (X86_XCPT_*).
1091	*/
1092	DECLINLINE(void) hmR0SvmRemoveXcptIntercept(PSVMVMCB pVmcb, uint32_t u32Xcpt)
1093	{
1094	#ifndef HMSVM_ALWAYS_TRAP_ALL_XCPTS
1095	if (pVmcb->ctrl.u32InterceptException & RT_BIT(u32Xcpt))
1096	{
1097	pVmcb->ctrl.u32InterceptException &= ~RT_BIT(u32Xcpt);
1098	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1099	}
1100	#endif
1101	}
1102
1103
1104	/**
1105	* Loads the guest CR0 control register into the guest-state area in the VMCB.
1106	* Although the guest CR0 is a separate field in the VMCB we have to consider
1107	* the FPU state itself which is shared between the host and the guest.
1108	*
1109	* @returns VBox status code.
1110	* @param pVM Pointer to the VMCPU.
1111	* @param pVmcb Pointer to the VM control block.
1112	* @param pCtx Pointer to the guest-CPU context.
1113	*
1114	* @remarks No-long-jump zone!!!
1115	*/
1116	static void hmR0SvmLoadSharedCR0(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1117	{
1118	/*
1119	* Guest CR0.
1120	*/
1121	PVM pVM = pVCpu->CTX_SUFF(pVM);
1122	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR0))
1123	{
1124	uint64_t u64GuestCR0 = pCtx->cr0;
1125
1126	/* Always enable caching. */
1127	u64GuestCR0 &= ~(X86_CR0_CD \| X86_CR0_NW);
1128
1129	/*
1130	* When Nested Paging is not available use shadow page tables and intercept #PFs (the latter done in SVMR0SetupVM()).
1131	*/
1132	if (!pVM->hm.s.fNestedPaging)
1133	{
1134	u64GuestCR0 \|= X86_CR0_PG; /* When Nested Paging is not available, use shadow page tables. */
1135	u64GuestCR0 \|= X86_CR0_WP; /* Guest CPL 0 writes to its read-only pages should cause a #PF VM-exit. */
1136	}
1137
1138	/*
1139	* Guest FPU bits.
1140	*/
1141	bool fInterceptNM = false;
1142	bool fInterceptMF = false;
1143	u64GuestCR0 \|= X86_CR0_NE; /* Use internal x87 FPU exceptions handling rather than external interrupts. */
1144	if (CPUMIsGuestFPUStateActive(pVCpu))
1145	{
1146	/* Catch floating point exceptions if we need to report them to the guest in a different way. */
1147	if (!(u64GuestCR0 & X86_CR0_NE))
1148	{
1149	Log4(("hmR0SvmLoadGuestControlRegs: Intercepting Guest CR0.MP Old-style FPU handling!!!\n"));
1150	fInterceptMF = true;
1151	}
1152	}
1153	else
1154	{
1155	fInterceptNM = true; /* Guest FPU inactive, VM-exit on #NM for lazy FPU loading. */
1156	u64GuestCR0 \|= X86_CR0_TS /* Guest can task switch quickly and do lazy FPU syncing. */
1157	\| X86_CR0_MP; /* FWAIT/WAIT should not ignore CR0.TS and should generate #NM. */
1158	}
1159
1160	/*
1161	* Update the exception intercept bitmap.
1162	*/
1163	if (fInterceptNM)
1164	hmR0SvmAddXcptIntercept(pVmcb, X86_XCPT_NM);
1165	else
1166	hmR0SvmRemoveXcptIntercept(pVmcb, X86_XCPT_NM);
1167
1168	if (fInterceptMF)
1169	hmR0SvmAddXcptIntercept(pVmcb, X86_XCPT_MF);
1170	else
1171	hmR0SvmRemoveXcptIntercept(pVmcb, X86_XCPT_MF);
1172
1173	pVmcb->guest.u64CR0 = u64GuestCR0;
1174	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1175	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_CR0);
1176	}
1177	}
1178
1179
1180	/**
1181	* Loads the guest control registers (CR2, CR3, CR4) into the VMCB.
1182	*
1183	* @returns VBox status code.
1184	* @param pVCpu Pointer to the VMCPU.
1185	* @param pVmcb Pointer to the VM control block.
1186	* @param pCtx Pointer to the guest-CPU context.
1187	*
1188	* @remarks No-long-jump zone!!!
1189	*/
1190	static int hmR0SvmLoadGuestControlRegs(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1191	{
1192	PVM pVM = pVCpu->CTX_SUFF(pVM);
1193
1194	/*
1195	* Guest CR2.
1196	*/
1197	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR2))
1198	{
1199	pVmcb->guest.u64CR2 = pCtx->cr2;
1200	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CR2;
1201	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_CR2);
1202	}
1203
1204	/*
1205	* Guest CR3.
1206	*/
1207	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR3))
1208	{
1209	if (pVM->hm.s.fNestedPaging)
1210	{
1211	PGMMODE enmShwPagingMode;
1212	#if HC_ARCH_BITS == 32
1213	if (CPUMIsGuestInLongModeEx(pCtx))
1214	enmShwPagingMode = PGMMODE_AMD64_NX;
1215	else
1216	#endif
1217	enmShwPagingMode = PGMGetHostMode(pVM);
1218
1219	pVmcb->ctrl.u64NestedPagingCR3 = PGMGetNestedCR3(pVCpu, enmShwPagingMode);
1220	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
1221	Assert(pVmcb->ctrl.u64NestedPagingCR3);
1222	pVmcb->guest.u64CR3 = pCtx->cr3;
1223	}
1224	else
1225	pVmcb->guest.u64CR3 = PGMGetHyperCR3(pVCpu);
1226
1227	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1228	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_CR3);
1229	}
1230
1231	/*
1232	* Guest CR4.
1233	*/
1234	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR4))
1235	{
1236	uint64_t u64GuestCR4 = pCtx->cr4;
1237	if (!pVM->hm.s.fNestedPaging)
1238	{
1239	switch (pVCpu->hm.s.enmShadowMode)
1240	{
1241	case PGMMODE_REAL:
1242	case PGMMODE_PROTECTED: /* Protected mode, no paging. */
1243	AssertFailed();
1244	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1245
1246	case PGMMODE_32_BIT: /* 32-bit paging. */
1247	u64GuestCR4 &= ~X86_CR4_PAE;
1248	break;
1249
1250	case PGMMODE_PAE: /* PAE paging. */
1251	case PGMMODE_PAE_NX: /* PAE paging with NX enabled. */
1252	/** Must use PAE paging as we could use physical memory > 4 GB */
1253	u64GuestCR4 \|= X86_CR4_PAE;
1254	break;
1255
1256	case PGMMODE_AMD64: /* 64-bit AMD paging (long mode). */
1257	case PGMMODE_AMD64_NX: /* 64-bit AMD paging (long mode) with NX enabled. */
1258	#ifdef VBOX_ENABLE_64_BITS_GUESTS
1259	break;
1260	#else
1261	AssertFailed();
1262	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1263	#endif
1264
1265	default: /* shut up gcc */
1266	AssertFailed();
1267	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1268	}
1269	}
1270
1271	pVmcb->guest.u64CR4 = u64GuestCR4;
1272	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1273	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_CR4);
1274	}
1275
1276	return VINF_SUCCESS;
1277	}
1278
1279
1280	/**
1281	* Loads the guest segment registers into the VMCB.
1282	*
1283	* @returns VBox status code.
1284	* @param pVCpu Pointer to the VMCPU.
1285	* @param pVmcb Pointer to the VM control block.
1286	* @param pCtx Pointer to the guest-CPU context.
1287	*
1288	* @remarks No-long-jump zone!!!
1289	*/
1290	static void hmR0SvmLoadGuestSegmentRegs(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1291	{
1292	/* Guest Segment registers: CS, SS, DS, ES, FS, GS. */
1293	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_SEGMENT_REGS))
1294	{
1295	HMSVM_LOAD_SEG_REG(CS, cs);
1296	HMSVM_LOAD_SEG_REG(SS, ss);
1297	HMSVM_LOAD_SEG_REG(DS, ds);
1298	HMSVM_LOAD_SEG_REG(ES, es);
1299	HMSVM_LOAD_SEG_REG(FS, fs);
1300	HMSVM_LOAD_SEG_REG(GS, gs);
1301
1302	pVmcb->guest.u8CPL = pCtx->ss.Attr.n.u2Dpl;
1303	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_SEG;
1304	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_SEGMENT_REGS);
1305	}
1306
1307	/* Guest TR. */
1308	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_TR))
1309	{
1310	HMSVM_LOAD_SEG_REG(TR, tr);
1311	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_TR);
1312	}
1313
1314	/* Guest LDTR. */
1315	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_LDTR))
1316	{
1317	HMSVM_LOAD_SEG_REG(LDTR, ldtr);
1318	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_LDTR);
1319	}
1320
1321	/* Guest GDTR. */
1322	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_GDTR))
1323	{
1324	pVmcb->guest.GDTR.u32Limit = pCtx->gdtr.cbGdt;
1325	pVmcb->guest.GDTR.u64Base = pCtx->gdtr.pGdt;
1326	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DT;
1327	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_GDTR);
1328	}
1329
1330	/* Guest IDTR. */
1331	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_IDTR))
1332	{
1333	pVmcb->guest.IDTR.u32Limit = pCtx->idtr.cbIdt;
1334	pVmcb->guest.IDTR.u64Base = pCtx->idtr.pIdt;
1335	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DT;
1336	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_IDTR);
1337	}
1338	}
1339
1340
1341	/**
1342	* Loads the guest MSRs into the VMCB.
1343	*
1344	* @param pVCpu Pointer to the VMCPU.
1345	* @param pVmcb Pointer to the VM control block.
1346	* @param pCtx Pointer to the guest-CPU context.
1347	*
1348	* @remarks No-long-jump zone!!!
1349	*/
1350	static void hmR0SvmLoadGuestMsrs(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1351	{
1352	/* Guest Sysenter MSRs. */
1353	pVmcb->guest.u64SysEnterCS = pCtx->SysEnter.cs;
1354	pVmcb->guest.u64SysEnterEIP = pCtx->SysEnter.eip;
1355	pVmcb->guest.u64SysEnterESP = pCtx->SysEnter.esp;
1356
1357	/*
1358	* Guest EFER MSR.
1359	* AMD-V requires guest EFER.SVME to be set. Weird.
1360	* See AMD spec. 15.5.1 "Basic Operation" \| "Canonicalization and Consistency Checks".
1361	*/
1362	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_EFER_MSR))
1363	{
1364	pVmcb->guest.u64EFER = pCtx->msrEFER \| MSR_K6_EFER_SVME;
1365	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1366	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_EFER_MSR);
1367	}
1368
1369	/* 64-bit MSRs. */
1370	if (CPUMIsGuestInLongModeEx(pCtx))
1371	{
1372	pVmcb->guest.FS.u64Base = pCtx->fs.u64Base;
1373	pVmcb->guest.GS.u64Base = pCtx->gs.u64Base;
1374	}
1375	else
1376	{
1377	/* If the guest isn't in 64-bit mode, clear MSR_K6_LME bit from guest EFER otherwise AMD-V expects amd64 shadow paging. */
1378	if (pCtx->msrEFER & MSR_K6_EFER_LME)
1379	{
1380	pVmcb->guest.u64EFER &= ~MSR_K6_EFER_LME;
1381	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1382	}
1383	}
1384
1385
1386	/** @todo The following are used in 64-bit only (SYSCALL/SYSRET) but they might
1387	* be writable in 32-bit mode. Clarify with AMD spec. */
1388	pVmcb->guest.u64STAR = pCtx->msrSTAR;
1389	pVmcb->guest.u64LSTAR = pCtx->msrLSTAR;
1390	pVmcb->guest.u64CSTAR = pCtx->msrCSTAR;
1391	pVmcb->guest.u64SFMASK = pCtx->msrSFMASK;
1392	pVmcb->guest.u64KernelGSBase = pCtx->msrKERNELGSBASE;
1393	}
1394
1395
1396	/**
1397	* Loads the guest state into the VMCB and programs the necessary intercepts
1398	* accordingly.
1399	*
1400	* @param pVCpu Pointer to the VMCPU.
1401	* @param pVmcb Pointer to the VM control block.
1402	* @param pCtx Pointer to the guest-CPU context.
1403	*
1404	* @remarks No-long-jump zone!!!
1405	* @remarks Requires EFLAGS to be up-to-date in the VMCB!
1406	*/
1407	static void hmR0SvmLoadSharedDebugState(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1408	{
1409	if (!HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_DEBUG))
1410	return;
1411	Assert((pCtx->dr[6] & X86_DR6_RA1_MASK) == X86_DR6_RA1_MASK); Assert((pCtx->dr[6] & X86_DR6_RAZ_MASK) == 0);
1412	Assert((pCtx->dr[7] & X86_DR7_RA1_MASK) == X86_DR7_RA1_MASK); Assert((pCtx->dr[7] & X86_DR7_RAZ_MASK) == 0);
1413
1414	bool fInterceptDB = false;
1415	bool fInterceptMovDRx = false;
1416
1417	/*
1418	* Anyone single stepping on the host side? If so, we'll have to use the
1419	* trap flag in the guest EFLAGS since AMD-V doesn't have a trap flag on
1420	* the VMM level like VT-x implementations does.
1421	*/
1422	bool const fStepping = pVCpu->hm.s.fSingleInstruction \|\| DBGFIsStepping(pVCpu);
1423	if (fStepping)
1424	{
1425	pVCpu->hm.s.fClearTrapFlag = true;
1426	pVmcb->guest.u64RFlags \|= X86_EFL_TF;
1427	fInterceptDB = true;
1428	fInterceptMovDRx = true; /* Need clean DR6, no guest mess. */
1429	}
1430
1431	if ( fStepping
1432	\|\| (CPUMGetHyperDR7(pVCpu) & X86_DR7_ENABLED_MASK))
1433	{
1434	/*
1435	* Use the combined guest and host DRx values found in the hypervisor
1436	* register set because the debugger has breakpoints active or someone
1437	* is single stepping on the host side.
1438	*
1439	* Note! DBGF expects a clean DR6 state before executing guest code.
1440	*/
1441	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS) && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
1442	if ( CPUMIsGuestInLongModeEx(pCtx)
1443	&& !CPUMIsHyperDebugStateActivePending(pVCpu))
1444	{
1445	CPUMR0LoadHyperDebugState(pVCpu, false /* include DR6 */);
1446	Assert(!CPUMIsGuestDebugStateActivePending(pVCpu));
1447	Assert(CPUMIsHyperDebugStateActivePending(pVCpu));
1448	}
1449	else
1450	#endif
1451	if (!CPUMIsHyperDebugStateActive(pVCpu))
1452	{
1453	CPUMR0LoadHyperDebugState(pVCpu, false /* include DR6 */);
1454	Assert(!CPUMIsGuestDebugStateActive(pVCpu));
1455	Assert(CPUMIsHyperDebugStateActive(pVCpu));
1456	}
1457
1458	/* Update DR6 & DR7. (The other DRx values are handled by CPUM one way or the other.) */
1459	if ( pVmcb->guest.u64DR6 != X86_DR6_INIT_VAL
1460	\|\| pVmcb->guest.u64DR7 != CPUMGetHyperDR7(pVCpu))
1461	{
1462	pVmcb->guest.u64DR7 = CPUMGetHyperDR7(pVCpu);
1463	pVmcb->guest.u64DR6 = X86_DR6_INIT_VAL;
1464	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
1465	pVCpu->hm.s.fUsingHyperDR7 = true;
1466	}
1467
1468	/** @todo If we cared, we could optimize to allow the guest to read registers
1469	* with the same values. */
1470	fInterceptDB = true;
1471	fInterceptMovDRx = true;
1472	Log5(("hmR0SvmLoadSharedDebugState: Loaded hyper DRx\n"));
1473	}
1474	else
1475	{
1476	/*
1477	* Update DR6, DR7 with the guest values if necessary.
1478	*/
1479	if ( pVmcb->guest.u64DR7 != pCtx->dr[7]
1480	\|\| pVmcb->guest.u64DR6 != pCtx->dr[6])
1481	{
1482	pVmcb->guest.u64DR7 = pCtx->dr[7];
1483	pVmcb->guest.u64DR6 = pCtx->dr[6];
1484	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
1485	pVCpu->hm.s.fUsingHyperDR7 = false;
1486	}
1487
1488	/*
1489	* If the guest has enabled debug registers, we need to load them prior to
1490	* executing guest code so they'll trigger at the right time.
1491	*/
1492	if (pCtx->dr[7] & (X86_DR7_ENABLED_MASK \| X86_DR7_GD)) /** @todo Why GD? */
1493	{
1494	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS) && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
1495	if ( CPUMIsGuestInLongModeEx(pCtx)
1496	&& !CPUMIsGuestDebugStateActivePending(pVCpu))
1497	{
1498	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
1499	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxArmed);
1500	Assert(!CPUMIsHyperDebugStateActivePending(pVCpu));
1501	Assert(CPUMIsGuestDebugStateActivePending(pVCpu));
1502	}
1503	else
1504	#endif
1505	if (!CPUMIsGuestDebugStateActive(pVCpu))
1506	{
1507	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
1508	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxArmed);
1509	Assert(!CPUMIsHyperDebugStateActive(pVCpu));
1510	Assert(CPUMIsGuestDebugStateActive(pVCpu));
1511	}
1512	Log5(("hmR0SvmLoadSharedDebugState: Loaded guest DRx\n"));
1513	}
1514	/*
1515	* If no debugging enabled, we'll lazy load DR0-3. We don't need to
1516	* intercept #DB as DR6 is updated in the VMCB.
1517	*/
1518	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS) && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
1519	else if ( !CPUMIsGuestDebugStateActivePending(pVCpu)
1520	&& !CPUMIsGuestDebugStateActive(pVCpu))
1521	#else
1522	else if (!CPUMIsGuestDebugStateActive(pVCpu))
1523	#endif
1524	{
1525	fInterceptMovDRx = true;
1526	}
1527	}
1528
1529	/*
1530	* Set up the intercepts.
1531	*/
1532	if (fInterceptDB)
1533	hmR0SvmAddXcptIntercept(pVmcb, X86_XCPT_DB);
1534	else
1535	hmR0SvmRemoveXcptIntercept(pVmcb, X86_XCPT_DB);
1536
1537	if (fInterceptMovDRx)
1538	{
1539	if ( pVmcb->ctrl.u16InterceptRdDRx != 0xffff
1540	\|\| pVmcb->ctrl.u16InterceptWrDRx != 0xffff)
1541	{
1542	pVmcb->ctrl.u16InterceptRdDRx = 0xffff;
1543	pVmcb->ctrl.u16InterceptWrDRx = 0xffff;
1544	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1545	}
1546	}
1547	else
1548	{
1549	if ( pVmcb->ctrl.u16InterceptRdDRx
1550	\|\| pVmcb->ctrl.u16InterceptWrDRx)
1551	{
1552	pVmcb->ctrl.u16InterceptRdDRx = 0;
1553	pVmcb->ctrl.u16InterceptWrDRx = 0;
1554	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1555	}
1556	}
1557
1558	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_DEBUG);
1559	}
1560
1561
1562	/**
1563	* Loads the guest APIC state (currently just the TPR).
1564	*
1565	* @returns VBox status code.
1566	* @param pVCpu Pointer to the VMCPU.
1567	* @param pVmcb Pointer to the VM control block.
1568	* @param pCtx Pointer to the guest-CPU context.
1569	*/
1570	static int hmR0SvmLoadGuestApicState(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1571	{
1572	if (!HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE))
1573	return VINF_SUCCESS;
1574
1575	bool fPendingIntr;
1576	uint8_t u8Tpr;
1577	int rc = PDMApicGetTPR(pVCpu, &u8Tpr, &fPendingIntr, NULL /* pu8PendingIrq */);
1578	AssertRCReturn(rc, rc);
1579
1580	/* Assume that we need to trap all TPR accesses and thus need not check on
1581	every #VMEXIT if we should update the TPR. */
1582	Assert(pVmcb->ctrl.IntCtrl.n.u1VIrqMasking);
1583	pVCpu->hm.s.svm.fSyncVTpr = false;
1584
1585	/* 32-bit guests uses LSTAR MSR for patching guest code which touches the TPR. */
1586	if (pVCpu->CTX_SUFF(pVM)->hm.s.fTPRPatchingActive)
1587	{
1588	pCtx->msrLSTAR = u8Tpr;
1589
1590	/* If there are interrupts pending, intercept LSTAR writes, otherwise don't intercept reads or writes. */
1591	if (fPendingIntr)
1592	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_INTERCEPT_WRITE);
1593	else
1594	{
1595	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1596	pVCpu->hm.s.svm.fSyncVTpr = true;
1597	}
1598	}
1599	else
1600	{
1601	/* Bits 3-0 of the VTPR field correspond to bits 7-4 of the TPR (which is the Task-Priority Class). */
1602	pVmcb->ctrl.IntCtrl.n.u8VTPR = (u8Tpr >> 4);
1603
1604	/* If there are interrupts pending, intercept CR8 writes to evaluate ASAP if we can deliver the interrupt to the guest. */
1605	if (fPendingIntr)
1606	pVmcb->ctrl.u16InterceptWrCRx \|= RT_BIT(8);
1607	else
1608	{
1609	pVmcb->ctrl.u16InterceptWrCRx &= ~RT_BIT(8);
1610	pVCpu->hm.s.svm.fSyncVTpr = true;
1611	}
1612
1613	pVmcb->ctrl.u64VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS \| HMSVM_VMCB_CLEAN_TPR);
1614	}
1615
1616	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
1617	return rc;
1618	}
1619
1620
1621	/**
1622	* Sets up the appropriate function to run guest code.
1623	*
1624	* @returns VBox status code.
1625	* @param pVCpu Pointer to the VMCPU.
1626	* @param pCtx Pointer to the guest-CPU context.
1627	*
1628	* @remarks No-long-jump zone!!!
1629	*/
1630	static int hmR0SvmSetupVMRunHandler(PVMCPU pVCpu, PCPUMCTX pCtx)
1631	{
1632	if (CPUMIsGuestInLongModeEx(pCtx))
1633	{
1634	#ifndef VBOX_ENABLE_64_BITS_GUESTS
1635	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1636	#endif
1637	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.fAllow64BitGuests); /* Guaranteed by hmR3InitFinalizeR0(). */
1638	#if HC_ARCH_BITS == 32 && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
1639	/* 32-bit host. We need to switch to 64-bit before running the 64-bit guest. */
1640	pVCpu->hm.s.svm.pfnVMRun = SVMR0VMSwitcherRun64;
1641	#else
1642	/* 64-bit host or hybrid host. */
1643	pVCpu->hm.s.svm.pfnVMRun = SVMR0VMRun64;
1644	#endif
1645	}
1646	else
1647	{
1648	/* Guest is not in long mode, use the 32-bit handler. */
1649	pVCpu->hm.s.svm.pfnVMRun = SVMR0VMRun;
1650	}
1651	return VINF_SUCCESS;
1652	}
1653
1654
1655	/**
1656	* Enters the AMD-V session.
1657	*
1658	* @returns VBox status code.
1659	* @param pVM Pointer to the VM.
1660	* @param pVCpu Pointer to the VMCPU.
1661	* @param pCpu Pointer to the CPU info struct.
1662	*/
1663	VMMR0DECL(int) SVMR0Enter(PVM pVM, PVMCPU pVCpu, PHMGLOBALCPUINFO pCpu)
1664	{
1665	AssertPtr(pVM);
1666	AssertPtr(pVCpu);
1667	Assert(pVM->hm.s.svm.fSupported);
1668	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1669	NOREF(pVM); NOREF(pCpu);
1670
1671	LogFlowFunc(("pVM=%p pVCpu=%p\n", pVM, pVCpu));
1672	Assert(HMCPU_CF_IS_SET(pVCpu, HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_HOST_GUEST_SHARED_STATE));
1673
1674	pVCpu->hm.s.fLeaveDone = false;
1675	return VINF_SUCCESS;
1676	}
1677
1678
1679	/**
1680	* Thread-context callback for AMD-V.
1681	*
1682	* @param enmEvent The thread-context event.
1683	* @param pVCpu Pointer to the VMCPU.
1684	* @param fGlobalInit Whether global VT-x/AMD-V init. is used.
1685	* @thread EMT(pVCpu)
1686	*/
1687	VMMR0DECL(void) SVMR0ThreadCtxCallback(RTTHREADCTXEVENT enmEvent, PVMCPU pVCpu, bool fGlobalInit)
1688	{
1689	NOREF(fGlobalInit);
1690
1691	switch (enmEvent)
1692	{
1693	case RTTHREADCTXEVENT_PREEMPTING:
1694	{
1695	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1696	Assert(VMMR0ThreadCtxHooksAreRegistered(pVCpu));
1697	VMCPU_ASSERT_EMT(pVCpu);
1698
1699	PVM pVM = pVCpu->CTX_SUFF(pVM);
1700	PCPUMCTX pCtx = CPUMQueryGuestCtxPtr(pVCpu);
1701
1702	/* No longjmps (log-flush, locks) in this fragile context. */
1703	VMMRZCallRing3Disable(pVCpu);
1704
1705	if (!pVCpu->hm.s.fLeaveDone)
1706	{
1707	hmR0SvmLeave(pVM, pVCpu, pCtx);
1708	pVCpu->hm.s.fLeaveDone = true;
1709	}
1710
1711	/* Leave HM context, takes care of local init (term). */
1712	int rc = HMR0LeaveCpu(pVCpu);
1713	AssertRC(rc); NOREF(rc);
1714
1715	/* Restore longjmp state. */
1716	VMMRZCallRing3Enable(pVCpu);
1717	STAM_COUNTER_INC(&pVCpu->hm.s.StatPreemptPreempting);
1718	break;
1719	}
1720
1721	case RTTHREADCTXEVENT_RESUMED:
1722	{
1723	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1724	Assert(VMMR0ThreadCtxHooksAreRegistered(pVCpu));
1725	VMCPU_ASSERT_EMT(pVCpu);
1726
1727	/* No longjmps (log-flush, locks) in this fragile context. */
1728	VMMRZCallRing3Disable(pVCpu);
1729
1730	/*
1731	* Initialize the bare minimum state required for HM. This takes care of
1732	* initializing AMD-V if necessary (onlined CPUs, local init etc.)
1733	*/
1734	int rc = HMR0EnterCpu(pVCpu);
1735	AssertRC(rc); NOREF(rc);
1736	Assert(HMCPU_CF_IS_SET(pVCpu, HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_HOST_GUEST_SHARED_STATE));
1737
1738	pVCpu->hm.s.fLeaveDone = false;
1739
1740	/* Restore longjmp state. */
1741	VMMRZCallRing3Enable(pVCpu);
1742	break;
1743	}
1744
1745	default:
1746	break;
1747	}
1748	}
1749
1750
1751	/**
1752	* Saves the host state.
1753	*
1754	* @returns VBox status code.
1755	* @param pVM Pointer to the VM.
1756	* @param pVCpu Pointer to the VMCPU.
1757	*
1758	* @remarks No-long-jump zone!!!
1759	*/
1760	VMMR0DECL(int) SVMR0SaveHostState(PVM pVM, PVMCPU pVCpu)
1761	{
1762	NOREF(pVM);
1763	NOREF(pVCpu);
1764	/* Nothing to do here. AMD-V does this for us automatically during the world-switch. */
1765	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_HOST_CONTEXT);
1766	return VINF_SUCCESS;
1767	}
1768
1769
1770	/**
1771	* Loads the guest state into the VMCB. The CPU state will be loaded from these
1772	* fields on every successful VM-entry.
1773	*
1774	* Also sets up the appropriate VMRUN function to execute guest code based on
1775	* the guest CPU mode.
1776	*
1777	* @returns VBox status code.
1778	* @param pVM Pointer to the VM.
1779	* @param pVCpu Pointer to the VMCPU.
1780	* @param pCtx Pointer to the guest-CPU context.
1781	*
1782	* @remarks No-long-jump zone!!!
1783	*/
1784	static int hmR0SvmLoadGuestState(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
1785	{
1786	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
1787	AssertMsgReturn(pVmcb, ("Invalid pVmcb\n"), VERR_SVM_INVALID_PVMCB);
1788
1789	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatLoadGuestState, x);
1790
1791	int rc = hmR0SvmLoadGuestControlRegs(pVCpu, pVmcb, pCtx);
1792	AssertLogRelMsgRCReturn(rc, ("hmR0SvmLoadGuestControlRegs! rc=%Rrc (pVM=%p pVCpu=%p)\n", rc, pVM, pVCpu), rc);
1793
1794	hmR0SvmLoadGuestSegmentRegs(pVCpu, pVmcb, pCtx);
1795	hmR0SvmLoadGuestMsrs(pVCpu, pVmcb, pCtx);
1796
1797	pVmcb->guest.u64RIP = pCtx->rip;
1798	pVmcb->guest.u64RSP = pCtx->rsp;
1799	pVmcb->guest.u64RFlags = pCtx->eflags.u32;
1800	pVmcb->guest.u64RAX = pCtx->rax;
1801
1802	rc = hmR0SvmLoadGuestApicState(pVCpu, pVmcb, pCtx);
1803	AssertLogRelMsgRCReturn(rc, ("hmR0SvmLoadGuestApicState! rc=%Rrc (pVM=%p pVCpu=%p)\n", rc, pVM, pVCpu), rc);
1804
1805	rc = hmR0SvmSetupVMRunHandler(pVCpu, pCtx);
1806	AssertLogRelMsgRCReturn(rc, ("hmR0SvmSetupVMRunHandler! rc=%Rrc (pVM=%p pVCpu=%p)\n", rc, pVM, pVCpu), rc);
1807
1808	/* Clear any unused and reserved bits. */
1809	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_RIP /* Unused (loaded unconditionally). */
1810	\| HM_CHANGED_GUEST_RSP
1811	\| HM_CHANGED_GUEST_RFLAGS
1812	\| HM_CHANGED_GUEST_SYSENTER_CS_MSR
1813	\| HM_CHANGED_GUEST_SYSENTER_EIP_MSR
1814	\| HM_CHANGED_GUEST_SYSENTER_ESP_MSR
1815	\| HM_CHANGED_GUEST_LAZY_MSRS /* Unused. */
1816	\| HM_CHANGED_SVM_RESERVED1 /* Reserved. */
1817	\| HM_CHANGED_SVM_RESERVED2
1818	\| HM_CHANGED_SVM_RESERVED3
1819	\| HM_CHANGED_SVM_RESERVED4);
1820
1821	/* All the guest state bits should be loaded except maybe the host context and/or shared host/guest bits. */
1822	AssertMsg( !HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_ALL_GUEST)
1823	\|\| HMCPU_CF_IS_PENDING_ONLY(pVCpu, HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_HOST_GUEST_SHARED_STATE),
1824	("fContextUseFlags=%#RX32\n", HMCPU_CF_VALUE(pVCpu)));
1825
1826	Log4(("Load: CS:RIP=%04x:%RX64 EFL=%#x SS:RSP=%04x:%RX64\n", pCtx->cs.Sel, pCtx->rip, pCtx->eflags.u, pCtx->ss, pCtx->rsp));
1827	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatLoadGuestState, x);
1828	return rc;
1829	}
1830
1831
1832	/**
1833	* Loads the state shared between the host and guest into the
1834	* VMCB.
1835	*
1836	* @param pVCpu Pointer to the VMCPU.
1837	* @param pVmcb Pointer to the VM control block.
1838	* @param pCtx Pointer to the guest-CPU context.
1839	*
1840	* @remarks No-long-jump zone!!!
1841	*/
1842	static void hmR0SvmLoadSharedState(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1843	{
1844	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1845	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
1846
1847	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR0))
1848	hmR0SvmLoadSharedCR0(pVCpu, pVmcb, pCtx);
1849
1850	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_DEBUG))
1851	hmR0SvmLoadSharedDebugState(pVCpu, pVmcb, pCtx);
1852
1853	/* Unused on AMD-V. */
1854	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_LAZY_MSRS);
1855
1856	AssertMsg(!HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_HOST_GUEST_SHARED_STATE),
1857	("fContextUseFlags=%#RX32\n", HMCPU_CF_VALUE(pVCpu)));
1858	}
1859
1860
1861	/**
1862	* Saves the entire guest state from the VMCB into the
1863	* guest-CPU context. Currently there is no residual state left in the CPU that
1864	* is not updated in the VMCB.
1865	*
1866	* @returns VBox status code.
1867	* @param pVCpu Pointer to the VMCPU.
1868	* @param pMixedCtx Pointer to the guest-CPU context. The data may be
1869	* out-of-sync. Make sure to update the required fields
1870	* before using them.
1871	*/
1872	static void hmR0SvmSaveGuestState(PVMCPU pVCpu, PCPUMCTX pMixedCtx)
1873	{
1874	Assert(VMMRZCallRing3IsEnabled(pVCpu));
1875
1876	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
1877
1878	pMixedCtx->rip = pVmcb->guest.u64RIP;
1879	pMixedCtx->rsp = pVmcb->guest.u64RSP;
1880	pMixedCtx->eflags.u32 = pVmcb->guest.u64RFlags;
1881	pMixedCtx->rax = pVmcb->guest.u64RAX;
1882
1883	/*
1884	* Guest interrupt shadow.
1885	*/
1886	if (pVmcb->ctrl.u64IntShadow & SVM_INTERRUPT_SHADOW_ACTIVE)
1887	EMSetInhibitInterruptsPC(pVCpu, pMixedCtx->rip);
1888	else
1889	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS);
1890
1891	/*
1892	* Guest Control registers: CR2, CR3 (handled at the end) - accesses to other control registers are always intercepted.
1893	*/
1894	pMixedCtx->cr2 = pVmcb->guest.u64CR2;
1895
1896	/*
1897	* Guest MSRs.
1898	*/
1899	pMixedCtx->msrSTAR = pVmcb->guest.u64STAR; /* legacy syscall eip, cs & ss */
1900	pMixedCtx->msrLSTAR = pVmcb->guest.u64LSTAR; /* 64-bit mode syscall rip */
1901	pMixedCtx->msrCSTAR = pVmcb->guest.u64CSTAR; /* compatibility mode syscall rip */
1902	pMixedCtx->msrSFMASK = pVmcb->guest.u64SFMASK; /* syscall flag mask */
1903	pMixedCtx->msrKERNELGSBASE = pVmcb->guest.u64KernelGSBase; /* swapgs exchange value */
1904	pMixedCtx->SysEnter.cs = pVmcb->guest.u64SysEnterCS;
1905	pMixedCtx->SysEnter.eip = pVmcb->guest.u64SysEnterEIP;
1906	pMixedCtx->SysEnter.esp = pVmcb->guest.u64SysEnterESP;
1907
1908	/*
1909	* Guest segment registers (includes FS, GS base MSRs for 64-bit guests).
1910	*/
1911	HMSVM_SAVE_SEG_REG(CS, cs);
1912	HMSVM_SAVE_SEG_REG(SS, ss);
1913	HMSVM_SAVE_SEG_REG(DS, ds);
1914	HMSVM_SAVE_SEG_REG(ES, es);
1915	HMSVM_SAVE_SEG_REG(FS, fs);
1916	HMSVM_SAVE_SEG_REG(GS, gs);
1917
1918	/*
1919	* Correct the hidden CS granularity bit. Haven't seen it being wrong in any other
1920	* register (yet).
1921	*/
1922	/** @todo SELM might need to be fixed as it too should not care about the
1923	* granularity bit. See @bugref{6785}. */
1924	if ( !pMixedCtx->cs.Attr.n.u1Granularity
1925	&& pMixedCtx->cs.Attr.n.u1Present
1926	&& pMixedCtx->cs.u32Limit > UINT32_C(0xfffff))
1927	{
1928	Assert((pMixedCtx->cs.u32Limit & 0xfff) == 0xfff);
1929	pMixedCtx->cs.Attr.n.u1Granularity = 1;
1930	}
1931
1932	#ifdef VBOX_STRICT
1933	# define HMSVM_ASSERT_SEG_GRANULARITY(reg) \
1934	AssertMsg( !pMixedCtx->reg.Attr.n.u1Present \
1935	\|\| ( pMixedCtx->reg.Attr.n.u1Granularity \
1936	? (pMixedCtx->reg.u32Limit & 0xfff) == 0xfff \
1937	: pMixedCtx->reg.u32Limit <= UINT32_C(0xfffff)), \
1938	("Invalid Segment Attributes Limit=%#RX32 Attr=%#RX32 Base=%#RX64\n", pMixedCtx->reg.u32Limit, \
1939	pMixedCtx->reg.Attr.u, pMixedCtx->reg.u64Base))
1940
1941	HMSVM_ASSERT_SEG_GRANULARITY(cs);
1942	HMSVM_ASSERT_SEG_GRANULARITY(ss);
1943	HMSVM_ASSERT_SEG_GRANULARITY(ds);
1944	HMSVM_ASSERT_SEG_GRANULARITY(es);
1945	HMSVM_ASSERT_SEG_GRANULARITY(fs);
1946	HMSVM_ASSERT_SEG_GRANULARITY(gs);
1947
1948	# undef HMSVM_ASSERT_SEL_GRANULARITY
1949	#endif
1950
1951	/*
1952	* Sync the hidden SS DPL field. AMD CPUs have a separate CPL field in the VMCB and uses that
1953	* and thus it's possible that when the CPL changes during guest execution that the SS DPL
1954	* isn't updated by AMD-V. Observed on some AMD Fusion CPUs with 64-bit guests.
1955	* See AMD spec. 15.5.1 "Basic operation".
1956	*/
1957	Assert(!(pVmcb->guest.u8CPL & ~0x3));
1958	pMixedCtx->ss.Attr.n.u2Dpl = pVmcb->guest.u8CPL & 0x3;
1959
1960	/*
1961	* Guest TR.
1962	* Fixup TR attributes so it's compatible with Intel. Important when saved-states are used
1963	* between Intel and AMD. See @bugref{6208} comment #39.
1964	*/
1965	HMSVM_SAVE_SEG_REG(TR, tr);
1966	if (CPUMIsGuestInLongModeEx(pMixedCtx))
1967	pMixedCtx->tr.Attr.n.u4Type = X86_SEL_TYPE_SYS_386_TSS_BUSY;
1968
1969	/*
1970	* Guest Descriptor-Table registers.
1971	*/
1972	HMSVM_SAVE_SEG_REG(LDTR, ldtr);
1973	pMixedCtx->gdtr.cbGdt = pVmcb->guest.GDTR.u32Limit;
1974	pMixedCtx->gdtr.pGdt = pVmcb->guest.GDTR.u64Base;
1975
1976	pMixedCtx->idtr.cbIdt = pVmcb->guest.IDTR.u32Limit;
1977	pMixedCtx->idtr.pIdt = pVmcb->guest.IDTR.u64Base;
1978
1979	/*
1980	* Guest Debug registers.
1981	*/
1982	if (!pVCpu->hm.s.fUsingHyperDR7)
1983	{
1984	pMixedCtx->dr[6] = pVmcb->guest.u64DR6;
1985	pMixedCtx->dr[7] = pVmcb->guest.u64DR7;
1986	}
1987	else
1988	{
1989	Assert(pVmcb->guest.u64DR7 == CPUMGetHyperDR7(pVCpu));
1990	CPUMSetHyperDR6(pVCpu, pVmcb->guest.u64DR6);
1991	}
1992
1993	/*
1994	* With Nested Paging, CR3 changes are not intercepted. Therefore, sync. it now.
1995	* This is done as the very last step of syncing the guest state, as PGMUpdateCR3() may cause longjmp's to ring-3.
1996	*/
1997	if ( pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging
1998	&& pMixedCtx->cr3 != pVmcb->guest.u64CR3)
1999	{
2000	CPUMSetGuestCR3(pVCpu, pVmcb->guest.u64CR3);
2001	PGMUpdateCR3(pVCpu, pVmcb->guest.u64CR3);
2002	}
2003	}
2004
2005
2006	/**
2007	* Does the necessary state syncing before returning to ring-3 for any reason
2008	* (longjmp, preemption, voluntary exits to ring-3) from AMD-V.
2009	*
2010	* @param pVM Pointer to the VM.
2011	* @param pVCpu Pointer to the VMCPU.
2012	* @param pMixedCtx Pointer to the guest-CPU context.
2013	*
2014	* @remarks No-long-jmp zone!!!
2015	*/
2016	static void hmR0SvmLeave(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
2017	{
2018	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2019	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2020	Assert(VMMR0IsLogFlushDisabled(pVCpu));
2021
2022	/*
2023	* !!! IMPORTANT !!!
2024	* If you modify code here, make sure to check whether hmR0SvmCallRing3Callback() needs to be updated too.
2025	*/
2026
2027	/* Restore host FPU state if necessary and resync on next R0 reentry .*/
2028	if (CPUMIsGuestFPUStateActive(pVCpu))
2029	{
2030	CPUMR0SaveGuestFPU(pVM, pVCpu, pCtx);
2031	Assert(!CPUMIsGuestFPUStateActive(pVCpu));
2032	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR0);
2033	}
2034
2035	/*
2036	* Restore host debug registers if necessary and resync on next R0 reentry.
2037	*/
2038	#ifdef VBOX_STRICT
2039	if (CPUMIsHyperDebugStateActive(pVCpu))
2040	{
2041	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2042	Assert(pVmcb->ctrl.u16InterceptRdDRx == 0xffff);
2043	Assert(pVmcb->ctrl.u16InterceptWrDRx == 0xffff);
2044	}
2045	#endif
2046	if (CPUMR0DebugStateMaybeSaveGuestAndRestoreHost(pVCpu, false /* save DR6 */))
2047	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_DEBUG);
2048
2049	Assert(!CPUMIsHyperDebugStateActive(pVCpu));
2050	Assert(!CPUMIsGuestDebugStateActive(pVCpu));
2051
2052	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatEntry);
2053	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatLoadGuestState);
2054	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatExit1);
2055	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatExit2);
2056	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchLongJmpToR3);
2057
2058	VMCPU_CMPXCHG_STATE(pVCpu, VMCPUSTATE_STARTED_HM, VMCPUSTATE_STARTED_EXEC);
2059	}
2060
2061
2062	/**
2063	* Leaves the AMD-V session.
2064	*
2065	* @returns VBox status code.
2066	* @param pVM Pointer to the VM.
2067	* @param pVCpu Pointer to the VMCPU.
2068	* @param pCtx Pointer to the guest-CPU context.
2069	*/
2070	static int hmR0SvmLeaveSession(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
2071	{
2072	HM_DISABLE_PREEMPT_IF_NEEDED();
2073	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2074	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2075
2076	/* When thread-context hooks are used, we can avoid doing the leave again if we had been preempted before
2077	and done this from the SVMR0ThreadCtxCallback(). */
2078	if (!pVCpu->hm.s.fLeaveDone)
2079	{
2080	hmR0SvmLeave(pVM, pVCpu, pCtx);
2081	pVCpu->hm.s.fLeaveDone = true;
2082	}
2083
2084	/*
2085	* !!! IMPORTANT !!!
2086	* If you modify code here, make sure to check whether hmR0SvmCallRing3Callback() needs to be updated too.
2087	*/
2088
2089	/* Deregister hook now that we've left HM context before re-enabling preemption. */
2090	if (VMMR0ThreadCtxHooksAreRegistered(pVCpu))
2091	VMMR0ThreadCtxHooksDeregister(pVCpu);
2092
2093	/* Leave HM context. This takes care of local init (term). */
2094	int rc = HMR0LeaveCpu(pVCpu);
2095
2096	HM_RESTORE_PREEMPT_IF_NEEDED();
2097	return rc;
2098	}
2099
2100
2101	/**
2102	* Does the necessary state syncing before doing a longjmp to ring-3.
2103	*
2104	* @returns VBox status code.
2105	* @param pVM Pointer to the VM.
2106	* @param pVCpu Pointer to the VMCPU.
2107	* @param pCtx Pointer to the guest-CPU context.
2108	*
2109	* @remarks No-long-jmp zone!!!
2110	*/
2111	static int hmR0SvmLongJmpToRing3(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
2112	{
2113	return hmR0SvmLeaveSession(pVM, pVCpu, pCtx);
2114	}
2115
2116
2117	/**
2118	* VMMRZCallRing3() callback wrapper which saves the guest state (or restores
2119	* any remaining host state) before we longjump to ring-3 and possibly get
2120	* preempted.
2121	*
2122	* @param pVCpu Pointer to the VMCPU.
2123	* @param enmOperation The operation causing the ring-3 longjump.
2124	* @param pvUser The user argument (pointer to the possibly
2125	* out-of-date guest-CPU context).
2126	*/
2127	DECLCALLBACK(int) hmR0SvmCallRing3Callback(PVMCPU pVCpu, VMMCALLRING3 enmOperation, void *pvUser)
2128	{
2129	if (enmOperation == VMMCALLRING3_VM_R0_ASSERTION)
2130	{
2131	/*
2132	* !!! IMPORTANT !!!
2133	* If you modify code here, make sure to check whether hmR0SvmLeave() and hmR0SvmLeaveSession() needs
2134	* to be updated too. This is a stripped down version which gets out ASAP trying to not trigger any assertion.
2135	*/
2136	VMMRZCallRing3RemoveNotification(pVCpu);
2137	VMMRZCallRing3Disable(pVCpu);
2138	HM_DISABLE_PREEMPT_IF_NEEDED();
2139
2140	/* Restore host FPU state if necessary and resync on next R0 reentry .*/
2141	if (CPUMIsGuestFPUStateActive(pVCpu))
2142	CPUMR0SaveGuestFPU(pVCpu->CTX_SUFF(pVM), pVCpu, (PCPUMCTX)pvUser);
2143
2144	/* Restore host debug registers if necessary and resync on next R0 reentry. */
2145	CPUMR0DebugStateMaybeSaveGuestAndRestoreHost(pVCpu, false /* save DR6 */);
2146
2147	/* Deregister hook now that we've left HM context before re-enabling preemption. */
2148	if (VMMR0ThreadCtxHooksAreRegistered(pVCpu))
2149	VMMR0ThreadCtxHooksDeregister(pVCpu);
2150
2151	/* Leave HM context. This takes care of local init (term). */
2152	HMR0LeaveCpu(pVCpu);
2153
2154	HM_RESTORE_PREEMPT_IF_NEEDED();
2155	return VINF_SUCCESS;
2156	}
2157
2158	Assert(pVCpu);
2159	Assert(pvUser);
2160	Assert(VMMRZCallRing3IsEnabled(pVCpu));
2161	HMSVM_ASSERT_PREEMPT_SAFE();
2162
2163	VMMRZCallRing3Disable(pVCpu);
2164	Assert(VMMR0IsLogFlushDisabled(pVCpu));
2165
2166	Log4(("hmR0SvmCallRing3Callback->hmR0SvmLongJmpToRing3\n"));
2167	int rc = hmR0SvmLongJmpToRing3(pVCpu->CTX_SUFF(pVM), pVCpu, (PCPUMCTX)pvUser);
2168	AssertRCReturn(rc, rc);
2169
2170	VMMRZCallRing3Enable(pVCpu);
2171	return VINF_SUCCESS;
2172	}
2173
2174
2175	/**
2176	* Take necessary actions before going back to ring-3.
2177	*
2178	* An action requires us to go back to ring-3. This function does the necessary
2179	* steps before we can safely return to ring-3. This is not the same as longjmps
2180	* to ring-3, this is voluntary.
2181	*
2182	* @param pVM Pointer to the VM.
2183	* @param pVCpu Pointer to the VMCPU.
2184	* @param pCtx Pointer to the guest-CPU context.
2185	* @param rcExit The reason for exiting to ring-3. Can be
2186	* VINF_VMM_UNKNOWN_RING3_CALL.
2187	*/
2188	static void hmR0SvmExitToRing3(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, int rcExit)
2189	{
2190	Assert(pVM);
2191	Assert(pVCpu);
2192	Assert(pCtx);
2193	HMSVM_ASSERT_PREEMPT_SAFE();
2194
2195	/* Please, no longjumps here (any logging shouldn't flush jump back to ring-3). NO LOGGING BEFORE THIS POINT! */
2196	VMMRZCallRing3Disable(pVCpu);
2197	Log4(("hmR0SvmExitToRing3: rcExit=%d\n", rcExit));
2198
2199	/* We need to do this only while truly exiting the "inner loop" back to ring-3 and -not- for any longjmp to ring3. */
2200	if (pVCpu->hm.s.Event.fPending)
2201	{
2202	hmR0SvmPendingEventToTrpmTrap(pVCpu);
2203	Assert(!pVCpu->hm.s.Event.fPending);
2204	}
2205
2206	/* Sync. the necessary state for going back to ring-3. */
2207	hmR0SvmLeaveSession(pVM, pVCpu, pCtx);
2208	STAM_COUNTER_DEC(&pVCpu->hm.s.StatSwitchLongJmpToR3);
2209
2210	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_TO_R3);
2211	CPUMSetChangedFlags(pVCpu, CPUM_CHANGED_SYSENTER_MSR
2212	\| CPUM_CHANGED_LDTR
2213	\| CPUM_CHANGED_GDTR
2214	\| CPUM_CHANGED_IDTR
2215	\| CPUM_CHANGED_TR
2216	\| CPUM_CHANGED_HIDDEN_SEL_REGS);
2217	if ( pVM->hm.s.fNestedPaging
2218	&& CPUMIsGuestPagingEnabledEx(pCtx))
2219	{
2220	CPUMSetChangedFlags(pVCpu, CPUM_CHANGED_GLOBAL_TLB_FLUSH);
2221	}
2222
2223	/* On our way back from ring-3 reload the guest state if there is a possibility of it being changed. */
2224	if (rcExit != VINF_EM_RAW_INTERRUPT)
2225	HMCPU_CF_SET(pVCpu, HM_CHANGED_ALL_GUEST);
2226
2227	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchExitToR3);
2228
2229	/* We do -not- want any longjmp notifications after this! We must return to ring-3 ASAP. */
2230	VMMRZCallRing3RemoveNotification(pVCpu);
2231	VMMRZCallRing3Enable(pVCpu);
2232	}
2233
2234
2235	/**
2236	* Updates the use of TSC offsetting mode for the CPU and adjusts the necessary
2237	* intercepts.
2238	*
2239	* @param pVCpu Pointer to the VMCPU.
2240	*
2241	* @remarks No-long-jump zone!!!
2242	*/
2243	static void hmR0SvmUpdateTscOffsetting(PVMCPU pVCpu)
2244	{
2245	bool fParavirtTsc = false;
2246	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2247	if (TMCpuTickCanUseRealTSC(pVCpu, &pVmcb->ctrl.u64TSCOffset, &fParavirtTsc))
2248	{
2249	uint64_t u64CurTSC = ASMReadTSC();
2250	uint64_t u64LastTick = TMCpuTickGetLastSeen(pVCpu);
2251	if (fParavirtTsc)
2252	{
2253	#if 0
2254	if (u64CurTSC + pVmcb->ctrl.u64TSCOffset > u64LastTick)
2255	{
2256	pVmcb->ctrl.u64TSCOffset = u64LastTick - u64CurTSC;
2257	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscOffsetAdjusted);
2258	}
2259	int rc = GIMR0UpdateParavirtTsc(pVCpu->CTX_SUFF(pVM), pVmcb->ctrl.u64TSCOffset);
2260	AssertRC(rc);
2261	#endif
2262	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscParavirt);
2263	}
2264
2265	if (u64CurTSC + pVmcb->ctrl.u64TSCOffset >= TMCpuTickGetLastSeen(pVCpu))
2266	{
2267	pVmcb->ctrl.u32InterceptCtrl1 &= ~SVM_CTRL1_INTERCEPT_RDTSC;
2268	pVmcb->ctrl.u32InterceptCtrl2 &= ~SVM_CTRL2_INTERCEPT_RDTSCP;
2269	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscOffset);
2270	}
2271	else
2272	{
2273	pVmcb->ctrl.u32InterceptCtrl1 \|= SVM_CTRL1_INTERCEPT_RDTSC;
2274	pVmcb->ctrl.u32InterceptCtrl2 \|= SVM_CTRL2_INTERCEPT_RDTSCP;
2275	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscInterceptOverFlow);
2276	}
2277	}
2278	else
2279	{
2280	Assert(!fParavirtTsc);
2281	pVmcb->ctrl.u32InterceptCtrl1 \|= SVM_CTRL1_INTERCEPT_RDTSC;
2282	pVmcb->ctrl.u32InterceptCtrl2 \|= SVM_CTRL2_INTERCEPT_RDTSCP;
2283	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscIntercept);
2284	}
2285
2286	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2287	}
2288
2289
2290	/**
2291	* Sets an event as a pending event to be injected into the guest.
2292	*
2293	* @param pVCpu Pointer to the VMCPU.
2294	* @param pEvent Pointer to the SVM event.
2295	* @param GCPtrFaultAddress The fault-address (CR2) in case it's a
2296	* page-fault.
2297	*
2298	* @remarks Statistics counter assumes this is a guest event being reflected to
2299	* the guest i.e. 'StatInjectPendingReflect' is incremented always.
2300	*/
2301	DECLINLINE(void) hmR0SvmSetPendingEvent(PVMCPU pVCpu, PSVMEVENT pEvent, RTGCUINTPTR GCPtrFaultAddress)
2302	{
2303	Assert(!pVCpu->hm.s.Event.fPending);
2304	Assert(pEvent->n.u1Valid);
2305
2306	pVCpu->hm.s.Event.u64IntInfo = pEvent->u;
2307	pVCpu->hm.s.Event.fPending = true;
2308	pVCpu->hm.s.Event.GCPtrFaultAddress = GCPtrFaultAddress;
2309
2310	Log4(("hmR0SvmSetPendingEvent: u=%#RX64 u8Vector=%#x Type=%#x ErrorCodeValid=%RTbool ErrorCode=%#RX32\n", pEvent->u,
2311	pEvent->n.u8Vector, (uint8_t)pEvent->n.u3Type, !!pEvent->n.u1ErrorCodeValid, pEvent->n.u32ErrorCode));
2312
2313	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectPendingReflect);
2314	}
2315
2316
2317	/**
2318	* Injects an event into the guest upon VMRUN by updating the relevant field
2319	* in the VMCB.
2320	*
2321	* @param pVCpu Pointer to the VMCPU.
2322	* @param pVmcb Pointer to the guest VM control block.
2323	* @param pCtx Pointer to the guest-CPU context.
2324	* @param pEvent Pointer to the event.
2325	*
2326	* @remarks No-long-jump zone!!!
2327	* @remarks Requires CR0!
2328	*/
2329	DECLINLINE(void) hmR0SvmInjectEventVmcb(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx, PSVMEVENT pEvent)
2330	{
2331	NOREF(pVCpu); NOREF(pCtx);
2332
2333	pVmcb->ctrl.EventInject.u = pEvent->u;
2334	STAM_COUNTER_INC(&pVCpu->hm.s.paStatInjectedIrqsR0[pEvent->n.u8Vector & MASK_INJECT_IRQ_STAT]);
2335
2336	Log4(("hmR0SvmInjectEventVmcb: u=%#RX64 u8Vector=%#x Type=%#x ErrorCodeValid=%RTbool ErrorCode=%#RX32\n", pEvent->u,
2337	pEvent->n.u8Vector, (uint8_t)pEvent->n.u3Type, !!pEvent->n.u1ErrorCodeValid, pEvent->n.u32ErrorCode));
2338	}
2339
2340
2341
2342	/**
2343	* Converts any TRPM trap into a pending HM event. This is typically used when
2344	* entering from ring-3 (not longjmp returns).
2345	*
2346	* @param pVCpu Pointer to the VMCPU.
2347	*/
2348	static void hmR0SvmTrpmTrapToPendingEvent(PVMCPU pVCpu)
2349	{
2350	Assert(TRPMHasTrap(pVCpu));
2351	Assert(!pVCpu->hm.s.Event.fPending);
2352
2353	uint8_t uVector;
2354	TRPMEVENT enmTrpmEvent;
2355	RTGCUINT uErrCode;
2356	RTGCUINTPTR GCPtrFaultAddress;
2357	uint8_t cbInstr;
2358
2359	int rc = TRPMQueryTrapAll(pVCpu, &uVector, &enmTrpmEvent, &uErrCode, &GCPtrFaultAddress, &cbInstr);
2360	AssertRC(rc);
2361
2362	SVMEVENT Event;
2363	Event.u = 0;
2364	Event.n.u1Valid = 1;
2365	Event.n.u8Vector = uVector;
2366
2367	/* Refer AMD spec. 15.20 "Event Injection" for the format. */
2368	if (enmTrpmEvent == TRPM_TRAP)
2369	{
2370	Event.n.u3Type = SVM_EVENT_EXCEPTION;
2371	switch (uVector)
2372	{
2373	case X86_XCPT_PF:
2374	case X86_XCPT_DF:
2375	case X86_XCPT_TS:
2376	case X86_XCPT_NP:
2377	case X86_XCPT_SS:
2378	case X86_XCPT_GP:
2379	case X86_XCPT_AC:
2380	{
2381	Event.n.u1ErrorCodeValid = 1;
2382	Event.n.u32ErrorCode = uErrCode;
2383	break;
2384	}
2385	}
2386	}
2387	else if (enmTrpmEvent == TRPM_HARDWARE_INT)
2388	{
2389	if (uVector == X86_XCPT_NMI)
2390	Event.n.u3Type = SVM_EVENT_NMI;
2391	else
2392	Event.n.u3Type = SVM_EVENT_EXTERNAL_IRQ;
2393	}
2394	else if (enmTrpmEvent == TRPM_SOFTWARE_INT)
2395	Event.n.u3Type = SVM_EVENT_SOFTWARE_INT;
2396	else
2397	AssertMsgFailed(("Invalid TRPM event type %d\n", enmTrpmEvent));
2398
2399	rc = TRPMResetTrap(pVCpu);
2400	AssertRC(rc);
2401
2402	Log4(("TRPM->HM event: u=%#RX64 u8Vector=%#x uErrorCodeValid=%RTbool uErrorCode=%#RX32\n", Event.u, Event.n.u8Vector,
2403	!!Event.n.u1ErrorCodeValid, Event.n.u32ErrorCode));
2404
2405	hmR0SvmSetPendingEvent(pVCpu, &Event, GCPtrFaultAddress);
2406	STAM_COUNTER_DEC(&pVCpu->hm.s.StatInjectPendingReflect);
2407	}
2408
2409
2410	/**
2411	* Converts any pending SVM event into a TRPM trap. Typically used when leaving
2412	* AMD-V to execute any instruction.
2413	*
2414	* @param pvCpu Pointer to the VMCPU.
2415	*/
2416	static void hmR0SvmPendingEventToTrpmTrap(PVMCPU pVCpu)
2417	{
2418	Assert(pVCpu->hm.s.Event.fPending);
2419	Assert(TRPMQueryTrap(pVCpu, NULL /* pu8TrapNo /, NULL / pEnmType */) == VERR_TRPM_NO_ACTIVE_TRAP);
2420
2421	SVMEVENT Event;
2422	Event.u = pVCpu->hm.s.Event.u64IntInfo;
2423
2424	uint8_t uVector = Event.n.u8Vector;
2425	uint8_t uVectorType = Event.n.u3Type;
2426
2427	TRPMEVENT enmTrapType;
2428	switch (uVectorType)
2429	{
2430	case SVM_EVENT_EXTERNAL_IRQ:
2431	case SVM_EVENT_NMI:
2432	enmTrapType = TRPM_HARDWARE_INT;
2433	break;
2434	case SVM_EVENT_SOFTWARE_INT:
2435	enmTrapType = TRPM_SOFTWARE_INT;
2436	break;
2437	case SVM_EVENT_EXCEPTION:
2438	enmTrapType = TRPM_TRAP;
2439	break;
2440	default:
2441	AssertMsgFailed(("Invalid pending-event type %#x\n", uVectorType));
2442	enmTrapType = TRPM_32BIT_HACK;
2443	break;
2444	}
2445
2446	Log4(("HM event->TRPM: uVector=%#x enmTrapType=%d\n", uVector, uVectorType));
2447
2448	int rc = TRPMAssertTrap(pVCpu, uVector, enmTrapType);
2449	AssertRC(rc);
2450
2451	if (Event.n.u1ErrorCodeValid)
2452	TRPMSetErrorCode(pVCpu, Event.n.u32ErrorCode);
2453
2454	if ( uVectorType == SVM_EVENT_EXCEPTION
2455	&& uVector == X86_XCPT_PF)
2456	{
2457	TRPMSetFaultAddress(pVCpu, pVCpu->hm.s.Event.GCPtrFaultAddress);
2458	Assert(pVCpu->hm.s.Event.GCPtrFaultAddress == CPUMGetGuestCR2(pVCpu));
2459	}
2460	else if (uVectorType == SVM_EVENT_SOFTWARE_INT)
2461	{
2462	AssertMsg( uVectorType == SVM_EVENT_SOFTWARE_INT
2463	\|\| (uVector == X86_XCPT_BP \|\| uVector == X86_XCPT_OF),
2464	("Invalid vector: uVector=%#x uVectorType=%#x\n", uVector, uVectorType));
2465	TRPMSetInstrLength(pVCpu, pVCpu->hm.s.Event.cbInstr);
2466	}
2467	pVCpu->hm.s.Event.fPending = false;
2468	}
2469
2470
2471	/**
2472	* Gets the guest's interrupt-shadow.
2473	*
2474	* @returns The guest's interrupt-shadow.
2475	* @param pVCpu Pointer to the VMCPU.
2476	* @param pCtx Pointer to the guest-CPU context.
2477	*
2478	* @remarks No-long-jump zone!!!
2479	* @remarks Has side-effects with VMCPU_FF_INHIBIT_INTERRUPTS force-flag.
2480	*/
2481	DECLINLINE(uint32_t) hmR0SvmGetGuestIntrShadow(PVMCPU pVCpu, PCPUMCTX pCtx)
2482	{
2483	/*
2484	* Instructions like STI and MOV SS inhibit interrupts till the next instruction completes. Check if we should
2485	* inhibit interrupts or clear any existing interrupt-inhibition.
2486	*/
2487	uint32_t uIntrState = 0;
2488	if (VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS))
2489	{
2490	if (pCtx->rip != EMGetInhibitInterruptsPC(pVCpu))
2491	{
2492	/*
2493	* We can clear the inhibit force flag as even if we go back to the recompiler without executing guest code in
2494	* AMD-V, the flag's condition to be cleared is met and thus the cleared state is correct.
2495	*/
2496	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS);
2497	}
2498	else
2499	uIntrState = SVM_INTERRUPT_SHADOW_ACTIVE;
2500	}
2501	return uIntrState;
2502	}
2503
2504
2505	/**
2506	* Sets the virtual interrupt intercept control in the VMCB which
2507	* instructs AMD-V to cause a #VMEXIT as soon as the guest is in a state to
2508	* receive interrupts.
2509	*
2510	* @param pVmcb Pointer to the VM control block.
2511	*/
2512	DECLINLINE(void) hmR0SvmSetVirtIntrIntercept(PSVMVMCB pVmcb)
2513	{
2514	if (!(pVmcb->ctrl.u32InterceptCtrl1 & SVM_CTRL1_INTERCEPT_VINTR))
2515	{
2516	pVmcb->ctrl.IntCtrl.n.u1VIrqValid = 1; /* A virtual interrupt is pending. */
2517	pVmcb->ctrl.IntCtrl.n.u8VIrqVector = 0; /* Not necessary as we #VMEXIT for delivering the interrupt. */
2518	pVmcb->ctrl.u32InterceptCtrl1 \|= SVM_CTRL1_INTERCEPT_VINTR;
2519	pVmcb->ctrl.u64VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS \| HMSVM_VMCB_CLEAN_TPR);
2520
2521	Log4(("Setting VINTR intercept\n"));
2522	}
2523	}
2524
2525
2526	/**
2527	* Evaluates the event to be delivered to the guest and sets it as the pending
2528	* event.
2529	*
2530	* @param pVCpu Pointer to the VMCPU.
2531	* @param pCtx Pointer to the guest-CPU context.
2532	*/
2533	static void hmR0SvmEvaluatePendingEvent(PVMCPU pVCpu, PCPUMCTX pCtx)
2534	{
2535	Assert(!pVCpu->hm.s.Event.fPending);
2536	Log4Func(("\n"));
2537
2538	const bool fIntShadow = RT_BOOL(hmR0SvmGetGuestIntrShadow(pVCpu, pCtx));
2539	const bool fBlockInt = !(pCtx->eflags.u32 & X86_EFL_IF);
2540	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2541
2542	SVMEVENT Event;
2543	Event.u = 0;
2544	/** @todo SMI. SMIs take priority over NMIs. */
2545	if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_NMI)) /* NMI. NMIs take priority over regular interrupts . */
2546	{
2547	if (!fIntShadow)
2548	{
2549	Log4(("Pending NMI\n"));
2550
2551	Event.n.u1Valid = 1;
2552	Event.n.u8Vector = X86_XCPT_NMI;
2553	Event.n.u3Type = SVM_EVENT_NMI;
2554
2555	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
2556	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INTERRUPT_NMI);
2557	}
2558	else
2559	hmR0SvmSetVirtIntrIntercept(pVmcb);
2560	}
2561	else if (VMCPU_FF_IS_PENDING(pVCpu, (VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC)))
2562	{
2563	/*
2564	* Check if the guest can receive external interrupts (PIC/APIC). Once we do PDMGetInterrupt() we -must- deliver
2565	* the interrupt ASAP. We must not execute any guest code until we inject the interrupt which is why it is
2566	* evaluated here and not set as pending, solely based on the force-flags.
2567	*/
2568	if ( !fBlockInt
2569	&& !fIntShadow)
2570	{
2571	uint8_t u8Interrupt;
2572	int rc = PDMGetInterrupt(pVCpu, &u8Interrupt);
2573	if (RT_SUCCESS(rc))
2574	{
2575	Log4(("Injecting external interrupt u8Interrupt=%#x\n", u8Interrupt));
2576
2577	Event.n.u1Valid = 1;
2578	Event.n.u8Vector = u8Interrupt;
2579	Event.n.u3Type = SVM_EVENT_EXTERNAL_IRQ;
2580
2581	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
2582	}
2583	else
2584	{
2585	/** @todo Does this actually happen? If not turn it into an assertion. */
2586	Assert(!VMCPU_FF_IS_PENDING(pVCpu, (VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC)));
2587	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchGuestIrq);
2588	}
2589	}
2590	else
2591	hmR0SvmSetVirtIntrIntercept(pVmcb);
2592	}
2593	}
2594
2595
2596	/**
2597	* Injects any pending events into the guest if the guest is in a state to
2598	* receive them.
2599	*
2600	* @param pVCpu Pointer to the VMCPU.
2601	* @param pCtx Pointer to the guest-CPU context.
2602	*/
2603	static void hmR0SvmInjectPendingEvent(PVMCPU pVCpu, PCPUMCTX pCtx)
2604	{
2605	Assert(!TRPMHasTrap(pVCpu));
2606	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2607	Log4Func(("\n"));
2608
2609	const bool fIntShadow = !!hmR0SvmGetGuestIntrShadow(pVCpu, pCtx);
2610	const bool fBlockInt = !(pCtx->eflags.u32 & X86_EFL_IF);
2611	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2612
2613	if (pVCpu->hm.s.Event.fPending) /* First, inject any pending HM events. */
2614	{
2615	SVMEVENT Event;
2616	Event.u = pVCpu->hm.s.Event.u64IntInfo;
2617	Assert(Event.n.u1Valid);
2618	#ifdef VBOX_STRICT
2619	if (Event.n.u3Type == SVM_EVENT_EXTERNAL_IRQ)
2620	{
2621	Assert(!fBlockInt);
2622	Assert(!fIntShadow);
2623	}
2624	else if (Event.n.u3Type == SVM_EVENT_NMI)
2625	Assert(!fIntShadow);
2626	#endif
2627
2628	Log4(("Injecting pending HM event.\n"));
2629	hmR0SvmInjectEventVmcb(pVCpu, pVmcb, pCtx, &Event);
2630	pVCpu->hm.s.Event.fPending = false;
2631
2632	#ifdef VBOX_WITH_STATISTICS
2633	if (Event.n.u3Type == SVM_EVENT_EXTERNAL_IRQ)
2634	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectInterrupt);
2635	else
2636	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectXcpt);
2637	#endif
2638	}
2639
2640	/* Update the guest interrupt shadow in the VMCB. */
2641	pVmcb->ctrl.u64IntShadow = !!fIntShadow;
2642	NOREF(fBlockInt);
2643	}
2644
2645
2646	/**
2647	* Reports world-switch error and dumps some useful debug info.
2648	*
2649	* @param pVM Pointer to the VM.
2650	* @param pVCpu Pointer to the VMCPU.
2651	* @param rcVMRun The return code from VMRUN (or
2652	* VERR_SVM_INVALID_GUEST_STATE for invalid
2653	* guest-state).
2654	* @param pCtx Pointer to the guest-CPU context.
2655	*/
2656	static void hmR0SvmReportWorldSwitchError(PVM pVM, PVMCPU pVCpu, int rcVMRun, PCPUMCTX pCtx)
2657	{
2658	NOREF(pCtx);
2659	HMSVM_ASSERT_PREEMPT_SAFE();
2660	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2661
2662	if (rcVMRun == VERR_SVM_INVALID_GUEST_STATE)
2663	{
2664	HMDumpRegs(pVM, pVCpu, pCtx); NOREF(pVM);
2665	#ifdef VBOX_STRICT
2666	Log4(("ctrl.u64VmcbCleanBits %#RX64\n", pVmcb->ctrl.u64VmcbCleanBits));
2667	Log4(("ctrl.u16InterceptRdCRx %#x\n", pVmcb->ctrl.u16InterceptRdCRx));
2668	Log4(("ctrl.u16InterceptWrCRx %#x\n", pVmcb->ctrl.u16InterceptWrCRx));
2669	Log4(("ctrl.u16InterceptRdDRx %#x\n", pVmcb->ctrl.u16InterceptRdDRx));
2670	Log4(("ctrl.u16InterceptWrDRx %#x\n", pVmcb->ctrl.u16InterceptWrDRx));
2671	Log4(("ctrl.u32InterceptException %#x\n", pVmcb->ctrl.u32InterceptException));
2672	Log4(("ctrl.u32InterceptCtrl1 %#x\n", pVmcb->ctrl.u32InterceptCtrl1));
2673	Log4(("ctrl.u32InterceptCtrl2 %#x\n", pVmcb->ctrl.u32InterceptCtrl2));
2674	Log4(("ctrl.u64IOPMPhysAddr %#RX64\n", pVmcb->ctrl.u64IOPMPhysAddr));
2675	Log4(("ctrl.u64MSRPMPhysAddr %#RX64\n", pVmcb->ctrl.u64MSRPMPhysAddr));
2676	Log4(("ctrl.u64TSCOffset %#RX64\n", pVmcb->ctrl.u64TSCOffset));
2677
2678	Log4(("ctrl.TLBCtrl.u32ASID %#x\n", pVmcb->ctrl.TLBCtrl.n.u32ASID));
2679	Log4(("ctrl.TLBCtrl.u8TLBFlush %#x\n", pVmcb->ctrl.TLBCtrl.n.u8TLBFlush));
2680	Log4(("ctrl.TLBCtrl.u24Reserved %#x\n", pVmcb->ctrl.TLBCtrl.n.u24Reserved));
2681
2682	Log4(("ctrl.IntCtrl.u8VTPR %#x\n", pVmcb->ctrl.IntCtrl.n.u8VTPR));
2683	Log4(("ctrl.IntCtrl.u1VIrqValid %#x\n", pVmcb->ctrl.IntCtrl.n.u1VIrqValid));
2684	Log4(("ctrl.IntCtrl.u7Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u7Reserved));
2685	Log4(("ctrl.IntCtrl.u4VIrqPriority %#x\n", pVmcb->ctrl.IntCtrl.n.u4VIrqPriority));
2686	Log4(("ctrl.IntCtrl.u1IgnoreTPR %#x\n", pVmcb->ctrl.IntCtrl.n.u1IgnoreTPR));
2687	Log4(("ctrl.IntCtrl.u3Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u3Reserved));
2688	Log4(("ctrl.IntCtrl.u1VIrqMasking %#x\n", pVmcb->ctrl.IntCtrl.n.u1VIrqMasking));
2689	Log4(("ctrl.IntCtrl.u6Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u6Reserved));
2690	Log4(("ctrl.IntCtrl.u8VIrqVector %#x\n", pVmcb->ctrl.IntCtrl.n.u8VIrqVector));
2691	Log4(("ctrl.IntCtrl.u24Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u24Reserved));
2692
2693	Log4(("ctrl.u64IntShadow %#RX64\n", pVmcb->ctrl.u64IntShadow));
2694	Log4(("ctrl.u64ExitCode %#RX64\n", pVmcb->ctrl.u64ExitCode));
2695	Log4(("ctrl.u64ExitInfo1 %#RX64\n", pVmcb->ctrl.u64ExitInfo1));
2696	Log4(("ctrl.u64ExitInfo2 %#RX64\n", pVmcb->ctrl.u64ExitInfo2));
2697	Log4(("ctrl.ExitIntInfo.u8Vector %#x\n", pVmcb->ctrl.ExitIntInfo.n.u8Vector));
2698	Log4(("ctrl.ExitIntInfo.u3Type %#x\n", pVmcb->ctrl.ExitIntInfo.n.u3Type));
2699	Log4(("ctrl.ExitIntInfo.u1ErrorCodeValid %#x\n", pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid));
2700	Log4(("ctrl.ExitIntInfo.u19Reserved %#x\n", pVmcb->ctrl.ExitIntInfo.n.u19Reserved));
2701	Log4(("ctrl.ExitIntInfo.u1Valid %#x\n", pVmcb->ctrl.ExitIntInfo.n.u1Valid));
2702	Log4(("ctrl.ExitIntInfo.u32ErrorCode %#x\n", pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode));
2703	Log4(("ctrl.NestedPaging %#RX64\n", pVmcb->ctrl.NestedPaging.u));
2704	Log4(("ctrl.EventInject.u8Vector %#x\n", pVmcb->ctrl.EventInject.n.u8Vector));
2705	Log4(("ctrl.EventInject.u3Type %#x\n", pVmcb->ctrl.EventInject.n.u3Type));
2706	Log4(("ctrl.EventInject.u1ErrorCodeValid %#x\n", pVmcb->ctrl.EventInject.n.u1ErrorCodeValid));
2707	Log4(("ctrl.EventInject.u19Reserved %#x\n", pVmcb->ctrl.EventInject.n.u19Reserved));
2708	Log4(("ctrl.EventInject.u1Valid %#x\n", pVmcb->ctrl.EventInject.n.u1Valid));
2709	Log4(("ctrl.EventInject.u32ErrorCode %#x\n", pVmcb->ctrl.EventInject.n.u32ErrorCode));
2710
2711	Log4(("ctrl.u64NestedPagingCR3 %#RX64\n", pVmcb->ctrl.u64NestedPagingCR3));
2712	Log4(("ctrl.u64LBRVirt %#RX64\n", pVmcb->ctrl.u64LBRVirt));
2713
2714	Log4(("guest.CS.u16Sel %RTsel\n", pVmcb->guest.CS.u16Sel));
2715	Log4(("guest.CS.u16Attr %#x\n", pVmcb->guest.CS.u16Attr));
2716	Log4(("guest.CS.u32Limit %#RX32\n", pVmcb->guest.CS.u32Limit));
2717	Log4(("guest.CS.u64Base %#RX64\n", pVmcb->guest.CS.u64Base));
2718	Log4(("guest.DS.u16Sel %#RTsel\n", pVmcb->guest.DS.u16Sel));
2719	Log4(("guest.DS.u16Attr %#x\n", pVmcb->guest.DS.u16Attr));
2720	Log4(("guest.DS.u32Limit %#RX32\n", pVmcb->guest.DS.u32Limit));
2721	Log4(("guest.DS.u64Base %#RX64\n", pVmcb->guest.DS.u64Base));
2722	Log4(("guest.ES.u16Sel %RTsel\n", pVmcb->guest.ES.u16Sel));
2723	Log4(("guest.ES.u16Attr %#x\n", pVmcb->guest.ES.u16Attr));
2724	Log4(("guest.ES.u32Limit %#RX32\n", pVmcb->guest.ES.u32Limit));
2725	Log4(("guest.ES.u64Base %#RX64\n", pVmcb->guest.ES.u64Base));
2726	Log4(("guest.FS.u16Sel %RTsel\n", pVmcb->guest.FS.u16Sel));
2727	Log4(("guest.FS.u16Attr %#x\n", pVmcb->guest.FS.u16Attr));
2728	Log4(("guest.FS.u32Limit %#RX32\n", pVmcb->guest.FS.u32Limit));
2729	Log4(("guest.FS.u64Base %#RX64\n", pVmcb->guest.FS.u64Base));
2730	Log4(("guest.GS.u16Sel %RTsel\n", pVmcb->guest.GS.u16Sel));
2731	Log4(("guest.GS.u16Attr %#x\n", pVmcb->guest.GS.u16Attr));
2732	Log4(("guest.GS.u32Limit %#RX32\n", pVmcb->guest.GS.u32Limit));
2733	Log4(("guest.GS.u64Base %#RX64\n", pVmcb->guest.GS.u64Base));
2734
2735	Log4(("guest.GDTR.u32Limit %#RX32\n", pVmcb->guest.GDTR.u32Limit));
2736	Log4(("guest.GDTR.u64Base %#RX64\n", pVmcb->guest.GDTR.u64Base));
2737
2738	Log4(("guest.LDTR.u16Sel %RTsel\n", pVmcb->guest.LDTR.u16Sel));
2739	Log4(("guest.LDTR.u16Attr %#x\n", pVmcb->guest.LDTR.u16Attr));
2740	Log4(("guest.LDTR.u32Limit %#RX32\n", pVmcb->guest.LDTR.u32Limit));
2741	Log4(("guest.LDTR.u64Base %#RX64\n", pVmcb->guest.LDTR.u64Base));
2742
2743	Log4(("guest.IDTR.u32Limit %#RX32\n", pVmcb->guest.IDTR.u32Limit));
2744	Log4(("guest.IDTR.u64Base %#RX64\n", pVmcb->guest.IDTR.u64Base));
2745
2746	Log4(("guest.TR.u16Sel %RTsel\n", pVmcb->guest.TR.u16Sel));
2747	Log4(("guest.TR.u16Attr %#x\n", pVmcb->guest.TR.u16Attr));
2748	Log4(("guest.TR.u32Limit %#RX32\n", pVmcb->guest.TR.u32Limit));
2749	Log4(("guest.TR.u64Base %#RX64\n", pVmcb->guest.TR.u64Base));
2750
2751	Log4(("guest.u8CPL %#x\n", pVmcb->guest.u8CPL));
2752	Log4(("guest.u64CR0 %#RX64\n", pVmcb->guest.u64CR0));
2753	Log4(("guest.u64CR2 %#RX64\n", pVmcb->guest.u64CR2));
2754	Log4(("guest.u64CR3 %#RX64\n", pVmcb->guest.u64CR3));
2755	Log4(("guest.u64CR4 %#RX64\n", pVmcb->guest.u64CR4));
2756	Log4(("guest.u64DR6 %#RX64\n", pVmcb->guest.u64DR6));
2757	Log4(("guest.u64DR7 %#RX64\n", pVmcb->guest.u64DR7));
2758
2759	Log4(("guest.u64RIP %#RX64\n", pVmcb->guest.u64RIP));
2760	Log4(("guest.u64RSP %#RX64\n", pVmcb->guest.u64RSP));
2761	Log4(("guest.u64RAX %#RX64\n", pVmcb->guest.u64RAX));
2762	Log4(("guest.u64RFlags %#RX64\n", pVmcb->guest.u64RFlags));
2763
2764	Log4(("guest.u64SysEnterCS %#RX64\n", pVmcb->guest.u64SysEnterCS));
2765	Log4(("guest.u64SysEnterEIP %#RX64\n", pVmcb->guest.u64SysEnterEIP));
2766	Log4(("guest.u64SysEnterESP %#RX64\n", pVmcb->guest.u64SysEnterESP));
2767
2768	Log4(("guest.u64EFER %#RX64\n", pVmcb->guest.u64EFER));
2769	Log4(("guest.u64STAR %#RX64\n", pVmcb->guest.u64STAR));
2770	Log4(("guest.u64LSTAR %#RX64\n", pVmcb->guest.u64LSTAR));
2771	Log4(("guest.u64CSTAR %#RX64\n", pVmcb->guest.u64CSTAR));
2772	Log4(("guest.u64SFMASK %#RX64\n", pVmcb->guest.u64SFMASK));
2773	Log4(("guest.u64KernelGSBase %#RX64\n", pVmcb->guest.u64KernelGSBase));
2774	Log4(("guest.u64GPAT %#RX64\n", pVmcb->guest.u64GPAT));
2775	Log4(("guest.u64DBGCTL %#RX64\n", pVmcb->guest.u64DBGCTL));
2776	Log4(("guest.u64BR_FROM %#RX64\n", pVmcb->guest.u64BR_FROM));
2777	Log4(("guest.u64BR_TO %#RX64\n", pVmcb->guest.u64BR_TO));
2778	Log4(("guest.u64LASTEXCPFROM %#RX64\n", pVmcb->guest.u64LASTEXCPFROM));
2779	Log4(("guest.u64LASTEXCPTO %#RX64\n", pVmcb->guest.u64LASTEXCPTO));
2780	#else
2781	NOREF(pVmcb);
2782	#endif /* VBOX_STRICT */
2783	}
2784	else
2785	Log4(("hmR0SvmReportWorldSwitchError: rcVMRun=%d\n", rcVMRun));
2786	}
2787
2788
2789	/**
2790	* Check per-VM and per-VCPU force flag actions that require us to go back to
2791	* ring-3 for one reason or another.
2792	*
2793	* @returns VBox status code (information status code included).
2794	* @retval VINF_SUCCESS if we don't have any actions that require going back to
2795	* ring-3.
2796	* @retval VINF_PGM_SYNC_CR3 if we have pending PGM CR3 sync.
2797	* @retval VINF_EM_PENDING_REQUEST if we have pending requests (like hardware
2798	* interrupts)
2799	* @retval VINF_PGM_POOL_FLUSH_PENDING if PGM is doing a pool flush and requires
2800	* all EMTs to be in ring-3.
2801	* @retval VINF_EM_RAW_TO_R3 if there is pending DMA requests.
2802	* @retval VINF_EM_NO_MEMORY PGM is out of memory, we need to return
2803	* to the EM loop.
2804	*
2805	* @param pVM Pointer to the VM.
2806	* @param pVCpu Pointer to the VMCPU.
2807	* @param pCtx Pointer to the guest-CPU context.
2808	*/
2809	static int hmR0SvmCheckForceFlags(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
2810	{
2811	Assert(VMMRZCallRing3IsEnabled(pVCpu));
2812
2813	/* On AMD-V we don't need to update CR3, PAE PDPES lazily. See hmR0SvmSaveGuestState(). */
2814	Assert(!VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_UPDATE_CR3));
2815	Assert(!VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_UPDATE_PAE_PDPES));
2816
2817	if ( VM_FF_IS_PENDING(pVM, !pVCpu->hm.s.fSingleInstruction
2818	? VM_FF_HP_R0_PRE_HM_MASK : VM_FF_HP_R0_PRE_HM_STEP_MASK)
2819	\|\| VMCPU_FF_IS_PENDING(pVCpu, !pVCpu->hm.s.fSingleInstruction
2820	? VMCPU_FF_HP_R0_PRE_HM_MASK : VMCPU_FF_HP_R0_PRE_HM_STEP_MASK) )
2821	{
2822	/* Pending PGM C3 sync. */
2823	if (VMCPU_FF_IS_PENDING(pVCpu,VMCPU_FF_PGM_SYNC_CR3 \| VMCPU_FF_PGM_SYNC_CR3_NON_GLOBAL))
2824	{
2825	int rc = PGMSyncCR3(pVCpu, pCtx->cr0, pCtx->cr3, pCtx->cr4, VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_PGM_SYNC_CR3));
2826	if (rc != VINF_SUCCESS)
2827	{
2828	Log4(("hmR0SvmCheckForceFlags: PGMSyncCR3 forcing us back to ring-3. rc=%d\n", rc));
2829	return rc;
2830	}
2831	}
2832
2833	/* Pending HM-to-R3 operations (critsects, timers, EMT rendezvous etc.) */
2834	/* -XXX- what was that about single stepping? */
2835	if ( VM_FF_IS_PENDING(pVM, VM_FF_HM_TO_R3_MASK)
2836	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_TO_R3_MASK))
2837	{
2838	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHmToR3FF);
2839	int rc = RT_UNLIKELY(VM_FF_IS_PENDING(pVM, VM_FF_PGM_NO_MEMORY)) ? VINF_EM_NO_MEMORY : VINF_EM_RAW_TO_R3;
2840	Log4(("hmR0SvmCheckForceFlags: HM_TO_R3 forcing us back to ring-3. rc=%d\n", rc));
2841	return rc;
2842	}
2843
2844	/* Pending VM request packets, such as hardware interrupts. */
2845	if ( VM_FF_IS_PENDING(pVM, VM_FF_REQUEST)
2846	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_REQUEST))
2847	{
2848	Log4(("hmR0SvmCheckForceFlags: Pending VM request forcing us back to ring-3\n"));
2849	return VINF_EM_PENDING_REQUEST;
2850	}
2851
2852	/* Pending PGM pool flushes. */
2853	if (VM_FF_IS_PENDING(pVM, VM_FF_PGM_POOL_FLUSH_PENDING))
2854	{
2855	Log4(("hmR0SvmCheckForceFlags: PGM pool flush pending forcing us back to ring-3\n"));
2856	return VINF_PGM_POOL_FLUSH_PENDING;
2857	}
2858
2859	/* Pending DMA requests. */
2860	if (VM_FF_IS_PENDING(pVM, VM_FF_PDM_DMA))
2861	{
2862	Log4(("hmR0SvmCheckForceFlags: Pending DMA request forcing us back to ring-3\n"));
2863	return VINF_EM_RAW_TO_R3;
2864	}
2865	}
2866
2867	return VINF_SUCCESS;
2868	}
2869
2870
2871	/**
2872	* Does the preparations before executing guest code in AMD-V.
2873	*
2874	* This may cause longjmps to ring-3 and may even result in rescheduling to the
2875	* recompiler. We must be cautious what we do here regarding committing
2876	* guest-state information into the the VMCB assuming we assuredly execute the
2877	* guest in AMD-V. If we fall back to the recompiler after updating the VMCB and
2878	* clearing the common-state (TRPM/forceflags), we must undo those changes so
2879	* that the recompiler can (and should) use them when it resumes guest
2880	* execution. Otherwise such operations must be done when we can no longer
2881	* exit to ring-3.
2882	*
2883	* @returns VBox status code (informational status codes included).
2884	* @retval VINF_SUCCESS if we can proceed with running the guest.
2885	* @retval VINF_* scheduling changes, we have to go back to ring-3.
2886	*
2887	* @param pVM Pointer to the VM.
2888	* @param pVCpu Pointer to the VMCPU.
2889	* @param pCtx Pointer to the guest-CPU context.
2890	* @param pSvmTransient Pointer to the SVM transient structure.
2891	*/
2892	static int hmR0SvmPreRunGuest(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
2893	{
2894	HMSVM_ASSERT_PREEMPT_SAFE();
2895
2896	/* Check force flag actions that might require us to go back to ring-3. */
2897	int rc = hmR0SvmCheckForceFlags(pVM, pVCpu, pCtx);
2898	if (rc != VINF_SUCCESS)
2899	return rc;
2900
2901	if (TRPMHasTrap(pVCpu))
2902	hmR0SvmTrpmTrapToPendingEvent(pVCpu);
2903	else if (!pVCpu->hm.s.Event.fPending)
2904	hmR0SvmEvaluatePendingEvent(pVCpu, pCtx);
2905
2906	#ifdef HMSVM_SYNC_FULL_GUEST_STATE
2907	HMCPU_CF_SET(pVCpu, HM_CHANGED_ALL_GUEST);
2908	#endif
2909
2910	/* Load the guest bits that are not shared with the host in any way since we can longjmp or get preempted. */
2911	rc = hmR0SvmLoadGuestState(pVM, pVCpu, pCtx);
2912	AssertRCReturn(rc, rc);
2913	STAM_COUNTER_INC(&pVCpu->hm.s.StatLoadFull);
2914
2915	/*
2916	* If we're not intercepting TPR changes in the guest, save the guest TPR before the world-switch
2917	* so we can update it on the way back if the guest changed the TPR.
2918	*/
2919	if (pVCpu->hm.s.svm.fSyncVTpr)
2920	{
2921	if (pVM->hm.s.fTPRPatchingActive)
2922	pSvmTransient->u8GuestTpr = pCtx->msrLSTAR;
2923	else
2924	{
2925	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2926	pSvmTransient->u8GuestTpr = pVmcb->ctrl.IntCtrl.n.u8VTPR;
2927	}
2928	}
2929
2930	/*
2931	* No longjmps to ring-3 from this point on!!!
2932	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional, better than a kernel panic.
2933	* This also disables flushing of the R0-logger instance (if any).
2934	*/
2935	VMMRZCallRing3Disable(pVCpu);
2936
2937	/*
2938	* We disable interrupts so that we don't miss any interrupts that would flag preemption (IPI/timers etc.)
2939	* when thread-context hooks aren't used and we've been running with preemption disabled for a while.
2940	*
2941	* We need to check for force-flags that could've possible been altered since we last checked them (e.g.
2942	* by PDMGetInterrupt() leaving the PDM critical section, see @bugref{6398}).
2943	*
2944	* We also check a couple of other force-flags as a last opportunity to get the EMT back to ring-3 before
2945	* executing guest code.
2946	*/
2947	pSvmTransient->uEflags = ASMIntDisableFlags();
2948	if ( VM_FF_IS_PENDING(pVM, VM_FF_EMT_RENDEZVOUS \| VM_FF_TM_VIRTUAL_SYNC)
2949	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_TO_R3_MASK))
2950	{
2951	ASMSetFlags(pSvmTransient->uEflags);
2952	VMMRZCallRing3Enable(pVCpu);
2953	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHmToR3FF);
2954	return VINF_EM_RAW_TO_R3;
2955	}
2956	if (RTThreadPreemptIsPending(NIL_RTTHREAD))
2957	{
2958	ASMSetFlags(pSvmTransient->uEflags);
2959	VMMRZCallRing3Enable(pVCpu);
2960	STAM_COUNTER_INC(&pVCpu->hm.s.StatPendingHostIrq);
2961	return VINF_EM_RAW_INTERRUPT;
2962	}
2963
2964	return VINF_SUCCESS;
2965	}
2966
2967
2968	/**
2969	* Prepares to run guest code in AMD-V and we've committed to doing so. This
2970	* means there is no backing out to ring-3 or anywhere else at this
2971	* point.
2972	*
2973	* @param pVM Pointer to the VM.
2974	* @param pVCpu Pointer to the VMCPU.
2975	* @param pCtx Pointer to the guest-CPU context.
2976	* @param pSvmTransient Pointer to the SVM transient structure.
2977	*
2978	* @remarks Called with preemption disabled.
2979	* @remarks No-long-jump zone!!!
2980	*/
2981	static void hmR0SvmPreRunGuestCommitted(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
2982	{
2983	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2984	Assert(VMMR0IsLogFlushDisabled(pVCpu));
2985	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2986
2987	VMCPU_ASSERT_STATE(pVCpu, VMCPUSTATE_STARTED_HM);
2988	VMCPU_SET_STATE(pVCpu, VMCPUSTATE_STARTED_EXEC); /* Indicate the start of guest execution. */
2989
2990	hmR0SvmInjectPendingEvent(pVCpu, pCtx);
2991
2992	if ( pVCpu->hm.s.fUseGuestFpu
2993	&& !CPUMIsGuestFPUStateActive(pVCpu))
2994	{
2995	CPUMR0LoadGuestFPU(pVM, pVCpu, pCtx);
2996	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR0);
2997	}
2998
2999	/* Load the state shared between host and guest (FPU, debug). */
3000	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
3001	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_HOST_GUEST_SHARED_STATE))
3002	hmR0SvmLoadSharedState(pVCpu, pVmcb, pCtx);
3003	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_HOST_CONTEXT); /* Preemption might set this, nothing to do on AMD-V. */
3004	AssertMsg(!HMCPU_CF_VALUE(pVCpu), ("fContextUseFlags=%#RX32\n", HMCPU_CF_VALUE(pVCpu)));
3005
3006	/* Setup TSC offsetting. */
3007	RTCPUID idCurrentCpu = HMR0GetCurrentCpu()->idCpu;
3008	if ( pSvmTransient->fUpdateTscOffsetting
3009	\|\| idCurrentCpu != pVCpu->hm.s.idLastCpu)
3010	{
3011	hmR0SvmUpdateTscOffsetting(pVCpu);
3012	pSvmTransient->fUpdateTscOffsetting = false;
3013	}
3014
3015	/* If we've migrating CPUs, mark the VMCB Clean bits as dirty. */
3016	if (idCurrentCpu != pVCpu->hm.s.idLastCpu)
3017	pVmcb->ctrl.u64VmcbCleanBits = 0;
3018
3019	/* Store status of the shared guest-host state at the time of VMRUN. */
3020	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS) && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
3021	if (CPUMIsGuestInLongModeEx(pCtx))
3022	{
3023	pSvmTransient->fWasGuestDebugStateActive = CPUMIsGuestDebugStateActivePending(pVCpu);
3024	pSvmTransient->fWasHyperDebugStateActive = CPUMIsHyperDebugStateActivePending(pVCpu);
3025	}
3026	else
3027	#endif
3028	{
3029	pSvmTransient->fWasGuestDebugStateActive = CPUMIsGuestDebugStateActive(pVCpu);
3030	pSvmTransient->fWasHyperDebugStateActive = CPUMIsHyperDebugStateActive(pVCpu);
3031	}
3032	pSvmTransient->fWasGuestFPUStateActive = CPUMIsGuestFPUStateActive(pVCpu);
3033
3034	/* Flush the appropriate tagged-TLB entries. */
3035	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, true); /* Used for TLB-shootdowns, set this across the world switch. */
3036	hmR0SvmFlushTaggedTlb(pVCpu);
3037	Assert(HMR0GetCurrentCpu()->idCpu == pVCpu->hm.s.idLastCpu);
3038
3039	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatEntry, &pVCpu->hm.s.StatInGC, x);
3040
3041	TMNotifyStartOfExecution(pVCpu); /* Finally, notify TM to resume its clocks as we're about
3042	to start executing. */
3043
3044	/*
3045	* Save the current Host TSC_AUX and write the guest TSC_AUX to the host, so that
3046	* RDTSCPs (that don't cause exits) reads the guest MSR. See @bugref{3324}.
3047	*
3048	* This should be done -after- any RDTSCPs for obtaining the host timestamp (TM, STAM etc).
3049	*/
3050	if ( (pVM->hm.s.cpuid.u32AMDFeatureEDX & X86_CPUID_EXT_FEATURE_EDX_RDTSCP)
3051	&& !(pVmcb->ctrl.u32InterceptCtrl2 & SVM_CTRL2_INTERCEPT_RDTSCP))
3052	{
3053	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_TSC_AUX, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
3054	pVCpu->hm.s.u64HostTscAux = ASMRdMsr(MSR_K8_TSC_AUX);
3055	uint64_t u64GuestTscAux = CPUMR0GetGuestTscAux(pVCpu);
3056	if (u64GuestTscAux != pVCpu->hm.s.u64HostTscAux)
3057	ASMWrMsr(MSR_K8_TSC_AUX, u64GuestTscAux);
3058	pSvmTransient->fRestoreTscAuxMsr = true;
3059	}
3060	else
3061	{
3062	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_TSC_AUX, SVMMSREXIT_INTERCEPT_READ, SVMMSREXIT_INTERCEPT_WRITE);
3063	pSvmTransient->fRestoreTscAuxMsr = false;
3064	}
3065
3066	/* If VMCB Clean bits isn't supported by the CPU, simply mark all state-bits as dirty, indicating (re)load-from-VMCB. */
3067	if (!(pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_VMCB_CLEAN))
3068	pVmcb->ctrl.u64VmcbCleanBits = 0;
3069	}
3070
3071
3072	/**
3073	* Wrapper for running the guest code in AMD-V.
3074	*
3075	* @returns VBox strict status code.
3076	* @param pVM Pointer to the VM.
3077	* @param pVCpu Pointer to the VMCPU.
3078	* @param pCtx Pointer to the guest-CPU context.
3079	*
3080	* @remarks No-long-jump zone!!!
3081	*/
3082	DECLINLINE(int) hmR0SvmRunGuest(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3083	{
3084	/*
3085	* 64-bit Windows uses XMM registers in the kernel as the Microsoft compiler expresses floating-point operations
3086	* using SSE instructions. Some XMM registers (XMM6-XMM15) are callee-saved and thus the need for this XMM wrapper.
3087	* Refer MSDN docs. "Configuring Programs for 64-bit / x64 Software Conventions / Register Usage" for details.
3088	*/
3089	#ifdef VBOX_WITH_KERNEL_USING_XMM
3090	return HMR0SVMRunWrapXMM(pVCpu->hm.s.svm.HCPhysVmcbHost, pVCpu->hm.s.svm.HCPhysVmcb, pCtx, pVM, pVCpu,
3091	pVCpu->hm.s.svm.pfnVMRun);
3092	#else
3093	return pVCpu->hm.s.svm.pfnVMRun(pVCpu->hm.s.svm.HCPhysVmcbHost, pVCpu->hm.s.svm.HCPhysVmcb, pCtx, pVM, pVCpu);
3094	#endif
3095	}
3096
3097
3098	/**
3099	* Performs some essential restoration of state after running guest code in
3100	* AMD-V.
3101	*
3102	* @param pVM Pointer to the VM.
3103	* @param pVCpu Pointer to the VMCPU.
3104	* @param pMixedCtx Pointer to the guest-CPU context. The data maybe
3105	* out-of-sync. Make sure to update the required fields
3106	* before using them.
3107	* @param pSvmTransient Pointer to the SVM transient structure.
3108	* @param rcVMRun Return code of VMRUN.
3109	*
3110	* @remarks Called with interrupts disabled.
3111	* @remarks No-long-jump zone!!! This function will however re-enable longjmps
3112	* unconditionally when it is safe to do so.
3113	*/
3114	static void hmR0SvmPostRunGuest(PVM pVM, PVMCPU pVCpu, PCPUMCTX pMixedCtx, PSVMTRANSIENT pSvmTransient, int rcVMRun)
3115	{
3116	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
3117
3118	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, false); /* See HMInvalidatePageOnAllVCpus(): used for TLB-shootdowns. */
3119	ASMAtomicIncU32(&pVCpu->hm.s.cWorldSwitchExits); /* Initialized in vmR3CreateUVM(): used for TLB-shootdowns. */
3120
3121	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
3122	pVmcb->ctrl.u64VmcbCleanBits = HMSVM_VMCB_CLEAN_ALL; /* Mark the VMCB-state cache as unmodified by VMM. */
3123
3124	if (pSvmTransient->fRestoreTscAuxMsr)
3125	{
3126	uint64_t u64GuestTscAuxMsr = ASMRdMsr(MSR_K8_TSC_AUX);
3127	CPUMR0SetGuestTscAux(pVCpu, u64GuestTscAuxMsr);
3128	if (u64GuestTscAuxMsr != pVCpu->hm.s.u64HostTscAux)
3129	ASMWrMsr(MSR_K8_TSC_AUX, pVCpu->hm.s.u64HostTscAux);
3130	}
3131
3132	if (!(pVmcb->ctrl.u32InterceptCtrl1 & SVM_CTRL1_INTERCEPT_RDTSC))
3133	{
3134	/** @todo Find a way to fix hardcoding a guestimate. */
3135	TMCpuTickSetLastSeen(pVCpu, ASMReadTSC() + pVmcb->ctrl.u64TSCOffset - 0x400);
3136	}
3137
3138	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatInGC, &pVCpu->hm.s.StatExit1, x);
3139	TMNotifyEndOfExecution(pVCpu); /* Notify TM that the guest is no longer running. */
3140	VMCPU_SET_STATE(pVCpu, VMCPUSTATE_STARTED_HM);
3141
3142	Assert(!(ASMGetFlags() & X86_EFL_IF));
3143	ASMSetFlags(pSvmTransient->uEflags); /* Enable interrupts. */
3144	VMMRZCallRing3Enable(pVCpu); /* It is now safe to do longjmps to ring-3!!! */
3145
3146	/* If VMRUN failed, we can bail out early. This does -not- cover SVM_EXIT_INVALID. */
3147	if (RT_UNLIKELY(rcVMRun != VINF_SUCCESS))
3148	{
3149	Log4(("VMRUN failure: rcVMRun=%Rrc\n", rcVMRun));
3150	return;
3151	}
3152
3153	pSvmTransient->u64ExitCode = pVmcb->ctrl.u64ExitCode; /* Save the #VMEXIT reason. */
3154	pSvmTransient->fVectoringPF = false; /* Vectoring page-fault needs to be determined later. */
3155	hmR0SvmSaveGuestState(pVCpu, pMixedCtx); /* Save the guest state from the VMCB to the guest-CPU context. */
3156
3157	if (RT_LIKELY(pSvmTransient->u64ExitCode != (uint64_t)SVM_EXIT_INVALID))
3158	{
3159	if (pVCpu->hm.s.svm.fSyncVTpr)
3160	{
3161	/* TPR patching (for 32-bit guests) uses LSTAR MSR for holding the TPR value, otherwise uses the VTPR. */
3162	if ( pVM->hm.s.fTPRPatchingActive
3163	&& (pMixedCtx->msrLSTAR & 0xff) != pSvmTransient->u8GuestTpr)
3164	{
3165	int rc = PDMApicSetTPR(pVCpu, pMixedCtx->msrLSTAR & 0xff);
3166	AssertRC(rc);
3167	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
3168	}
3169	else if (pSvmTransient->u8GuestTpr != pVmcb->ctrl.IntCtrl.n.u8VTPR)
3170	{
3171	int rc = PDMApicSetTPR(pVCpu, pVmcb->ctrl.IntCtrl.n.u8VTPR << 4);
3172	AssertRC(rc);
3173	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
3174	}
3175	}
3176	}
3177	}
3178
3179
3180	/**
3181	* Runs the guest code using AMD-V.
3182	*
3183	* @returns VBox status code.
3184	* @param pVM Pointer to the VM.
3185	* @param pVCpu Pointer to the VMCPU.
3186	*/
3187	static int hmR0SvmRunGuestCodeNormal(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3188	{
3189	SVMTRANSIENT SvmTransient;
3190	SvmTransient.fUpdateTscOffsetting = true;
3191	uint32_t cLoops = 0;
3192	int rc = VERR_INTERNAL_ERROR_5;
3193
3194	for (;; cLoops++)
3195	{
3196	Assert(!HMR0SuspendPending());
3197	HMSVM_ASSERT_CPU_SAFE();
3198
3199	/* Preparatory work for running guest code, this may force us to return
3200	to ring-3. This bugger disables interrupts on VINF_SUCCESS! */
3201	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatEntry, x);
3202	rc = hmR0SvmPreRunGuest(pVM, pVCpu, pCtx, &SvmTransient);
3203	if (rc != VINF_SUCCESS)
3204	break;
3205
3206	/*
3207	* No longjmps to ring-3 from this point on!!!
3208	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional, better than a kernel panic.
3209	* This also disables flushing of the R0-logger instance (if any).
3210	*/
3211	hmR0SvmPreRunGuestCommitted(pVM, pVCpu, pCtx, &SvmTransient);
3212	rc = hmR0SvmRunGuest(pVM, pVCpu, pCtx);
3213
3214	/* Restore any residual host-state and save any bits shared between host
3215	and guest into the guest-CPU state. Re-enables interrupts! */
3216	hmR0SvmPostRunGuest(pVM, pVCpu, pCtx, &SvmTransient, rc);
3217
3218	if (RT_UNLIKELY( rc != VINF_SUCCESS /* Check for VMRUN errors. */
3219	\|\| SvmTransient.u64ExitCode == (uint64_t)SVM_EXIT_INVALID)) /* Check for invalid guest-state errors. */
3220	{
3221	if (rc == VINF_SUCCESS)
3222	rc = VERR_SVM_INVALID_GUEST_STATE;
3223	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit1, x);
3224	hmR0SvmReportWorldSwitchError(pVM, pVCpu, rc, pCtx);
3225	break;
3226	}
3227
3228	/* Handle the #VMEXIT. */
3229	HMSVM_EXITCODE_STAM_COUNTER_INC(SvmTransient.u64ExitCode);
3230	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatExit1, &pVCpu->hm.s.StatExit2, x);
3231	rc = hmR0SvmHandleExit(pVCpu, pCtx, &SvmTransient);
3232	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit2, x);
3233	if (rc != VINF_SUCCESS)
3234	break;
3235	else if (cLoops > pVM->hm.s.cMaxResumeLoops)
3236	{
3237	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitMaxResume);
3238	rc = VINF_EM_RAW_INTERRUPT;
3239	break;
3240	}
3241	}
3242
3243	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatEntry, x);
3244	return rc;
3245	}
3246
3247
3248	/**
3249	* Runs the guest code using AMD-V in single step mode.
3250	*
3251	* @returns VBox status code.
3252	* @param pVM Pointer to the VM.
3253	* @param pVCpu Pointer to the VMCPU.
3254	* @param pCtx Pointer to the guest-CPU context.
3255	*/
3256	static int hmR0SvmRunGuestCodeStep(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3257	{
3258	SVMTRANSIENT SvmTransient;
3259	SvmTransient.fUpdateTscOffsetting = true;
3260	uint32_t cLoops = 0;
3261	int rc = VERR_INTERNAL_ERROR_5;
3262	uint16_t uCsStart = pCtx->cs.Sel;
3263	uint64_t uRipStart = pCtx->rip;
3264
3265	for (;; cLoops++)
3266	{
3267	Assert(!HMR0SuspendPending());
3268	AssertMsg(pVCpu->hm.s.idEnteredCpu == RTMpCpuId(),
3269	("Illegal migration! Entered on CPU %u Current %u cLoops=%u\n", (unsigned)pVCpu->hm.s.idEnteredCpu,
3270	(unsigned)RTMpCpuId(), cLoops));
3271
3272	/* Preparatory work for running guest code, this may force us to return
3273	to ring-3. This bugger disables interrupts on VINF_SUCCESS! */
3274	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatEntry, x);
3275	rc = hmR0SvmPreRunGuest(pVM, pVCpu, pCtx, &SvmTransient);
3276	if (rc != VINF_SUCCESS)
3277	break;
3278
3279	/*
3280	* No longjmps to ring-3 from this point on!!!
3281	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional, better than a kernel panic.
3282	* This also disables flushing of the R0-logger instance (if any).
3283	*/
3284	VMMRZCallRing3Disable(pVCpu);
3285	VMMRZCallRing3RemoveNotification(pVCpu);
3286	hmR0SvmPreRunGuestCommitted(pVM, pVCpu, pCtx, &SvmTransient);
3287
3288	rc = hmR0SvmRunGuest(pVM, pVCpu, pCtx);
3289
3290	/*
3291	* Restore any residual host-state and save any bits shared between host and guest into the guest-CPU state.
3292	* This will also re-enable longjmps to ring-3 when it has reached a safe point!!!
3293	*/
3294	hmR0SvmPostRunGuest(pVM, pVCpu, pCtx, &SvmTransient, rc);
3295	if (RT_UNLIKELY( rc != VINF_SUCCESS /* Check for VMRUN errors. */
3296	\|\| SvmTransient.u64ExitCode == (uint64_t)SVM_EXIT_INVALID)) /* Check for invalid guest-state errors. */
3297	{
3298	if (rc == VINF_SUCCESS)
3299	rc = VERR_SVM_INVALID_GUEST_STATE;
3300	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit1, x);
3301	hmR0SvmReportWorldSwitchError(pVM, pVCpu, rc, pCtx);
3302	return rc;
3303	}
3304
3305	/* Handle the #VMEXIT. */
3306	HMSVM_EXITCODE_STAM_COUNTER_INC(SvmTransient.u64ExitCode);
3307	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatExit1, &pVCpu->hm.s.StatExit2, x);
3308	rc = hmR0SvmHandleExit(pVCpu, pCtx, &SvmTransient);
3309	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit2, x);
3310	if (rc != VINF_SUCCESS)
3311	break;
3312	else if (cLoops > pVM->hm.s.cMaxResumeLoops)
3313	{
3314	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitMaxResume);
3315	rc = VINF_EM_RAW_INTERRUPT;
3316	break;
3317	}
3318
3319	/*
3320	* Did the RIP change, if so, consider it a single step.
3321	* Otherwise, make sure one of the TFs gets set.
3322	*/
3323	if ( pCtx->rip != uRipStart
3324	\|\| pCtx->cs.Sel != uCsStart)
3325	{
3326	rc = VINF_EM_DBG_STEPPED;
3327	break;
3328	}
3329	pVCpu->hm.s.fContextUseFlags \|= HM_CHANGED_GUEST_DEBUG;
3330	}
3331
3332	/*
3333	* Clear the X86_EFL_TF if necessary.
3334	*/
3335	if (pVCpu->hm.s.fClearTrapFlag)
3336	{
3337	pVCpu->hm.s.fClearTrapFlag = false;
3338	pCtx->eflags.Bits.u1TF = 0;
3339	}
3340
3341	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatEntry, x);
3342	return rc;
3343	}
3344
3345
3346	/**
3347	* Runs the guest code using AMD-V.
3348	*
3349	* @returns VBox status code.
3350	* @param pVM Pointer to the VM.
3351	* @param pVCpu Pointer to the VMCPU.
3352	* @param pCtx Pointer to the guest-CPU context.
3353	*/
3354	VMMR0DECL(int) SVMR0RunGuestCode(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3355	{
3356	Assert(VMMRZCallRing3IsEnabled(pVCpu));
3357	HMSVM_ASSERT_PREEMPT_SAFE();
3358	VMMRZCallRing3SetNotification(pVCpu, hmR0SvmCallRing3Callback, pCtx);
3359
3360	int rc;
3361	if (!pVCpu->hm.s.fSingleInstruction && !DBGFIsStepping(pVCpu))
3362	rc = hmR0SvmRunGuestCodeNormal(pVM, pVCpu, pCtx);
3363	else
3364	rc = hmR0SvmRunGuestCodeStep(pVM, pVCpu, pCtx);
3365
3366	if (rc == VERR_EM_INTERPRETER)
3367	rc = VINF_EM_RAW_EMULATE_INSTR;
3368	else if (rc == VINF_EM_RESET)
3369	rc = VINF_EM_TRIPLE_FAULT;
3370
3371	/* Prepare to return to ring-3. This will remove longjmp notifications. */
3372	hmR0SvmExitToRing3(pVM, pVCpu, pCtx, rc);
3373	Assert(!VMMRZCallRing3IsNotificationSet(pVCpu));
3374	return rc;
3375	}
3376
3377
3378	/**
3379	* Handles a #VMEXIT (for all EXITCODE values except SVM_EXIT_INVALID).
3380	*
3381	* @returns VBox status code (informational status codes included).
3382	* @param pVCpu Pointer to the VMCPU.
3383	* @param pCtx Pointer to the guest-CPU context.
3384	* @param pSvmTransient Pointer to the SVM transient structure.
3385	*/
3386	DECLINLINE(int) hmR0SvmHandleExit(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
3387	{
3388	Assert(pSvmTransient->u64ExitCode != (uint64_t)SVM_EXIT_INVALID);
3389	Assert(pSvmTransient->u64ExitCode <= SVM_EXIT_MAX);
3390
3391	/*
3392	* The ordering of the case labels is based on most-frequently-occurring VM-exits for most guests under
3393	* normal workloads (for some definition of "normal").
3394	*/
3395	uint32_t u32ExitCode = pSvmTransient->u64ExitCode;
3396	switch (pSvmTransient->u64ExitCode)
3397	{
3398	case SVM_EXIT_NPF:
3399	return hmR0SvmExitNestedPF(pVCpu, pCtx, pSvmTransient);
3400
3401	case SVM_EXIT_IOIO:
3402	return hmR0SvmExitIOInstr(pVCpu, pCtx, pSvmTransient);
3403
3404	case SVM_EXIT_RDTSC:
3405	return hmR0SvmExitRdtsc(pVCpu, pCtx, pSvmTransient);
3406
3407	case SVM_EXIT_RDTSCP:
3408	return hmR0SvmExitRdtscp(pVCpu, pCtx, pSvmTransient);
3409
3410	case SVM_EXIT_CPUID:
3411	return hmR0SvmExitCpuid(pVCpu, pCtx, pSvmTransient);
3412
3413	case SVM_EXIT_EXCEPTION_E: /* X86_XCPT_PF */
3414	return hmR0SvmExitXcptPF(pVCpu, pCtx, pSvmTransient);
3415
3416	case SVM_EXIT_EXCEPTION_7: /* X86_XCPT_NM */
3417	return hmR0SvmExitXcptNM(pVCpu, pCtx, pSvmTransient);
3418
3419	case SVM_EXIT_EXCEPTION_10: /* X86_XCPT_MF */
3420	return hmR0SvmExitXcptMF(pVCpu, pCtx, pSvmTransient);
3421
3422	case SVM_EXIT_EXCEPTION_1: /* X86_XCPT_DB */
3423	return hmR0SvmExitXcptDB(pVCpu, pCtx, pSvmTransient);
3424
3425	case SVM_EXIT_MONITOR:
3426	return hmR0SvmExitMonitor(pVCpu, pCtx, pSvmTransient);
3427
3428	case SVM_EXIT_MWAIT:
3429	return hmR0SvmExitMwait(pVCpu, pCtx, pSvmTransient);
3430
3431	case SVM_EXIT_HLT:
3432	return hmR0SvmExitHlt(pVCpu, pCtx, pSvmTransient);
3433
3434	case SVM_EXIT_READ_CR0:
3435	case SVM_EXIT_READ_CR3:
3436	case SVM_EXIT_READ_CR4:
3437	return hmR0SvmExitReadCRx(pVCpu, pCtx, pSvmTransient);
3438
3439	case SVM_EXIT_WRITE_CR0:
3440	case SVM_EXIT_WRITE_CR3:
3441	case SVM_EXIT_WRITE_CR4:
3442	case SVM_EXIT_WRITE_CR8:
3443	return hmR0SvmExitWriteCRx(pVCpu, pCtx, pSvmTransient);
3444
3445	case SVM_EXIT_VINTR:
3446	return hmR0SvmExitVIntr(pVCpu, pCtx, pSvmTransient);
3447
3448	case SVM_EXIT_INTR:
3449	case SVM_EXIT_FERR_FREEZE:
3450	case SVM_EXIT_NMI:
3451	return hmR0SvmExitIntr(pVCpu, pCtx, pSvmTransient);
3452
3453	case SVM_EXIT_MSR:
3454	return hmR0SvmExitMsr(pVCpu, pCtx, pSvmTransient);
3455
3456	case SVM_EXIT_INVLPG:
3457	return hmR0SvmExitInvlpg(pVCpu, pCtx, pSvmTransient);
3458
3459	case SVM_EXIT_WBINVD:
3460	return hmR0SvmExitWbinvd(pVCpu, pCtx, pSvmTransient);
3461
3462	case SVM_EXIT_INVD:
3463	return hmR0SvmExitInvd(pVCpu, pCtx, pSvmTransient);
3464
3465	case SVM_EXIT_RDPMC:
3466	return hmR0SvmExitRdpmc(pVCpu, pCtx, pSvmTransient);
3467
3468	default:
3469	{
3470	switch (pSvmTransient->u64ExitCode)
3471	{
3472	case SVM_EXIT_READ_DR0: case SVM_EXIT_READ_DR1: case SVM_EXIT_READ_DR2: case SVM_EXIT_READ_DR3:
3473	case SVM_EXIT_READ_DR6: case SVM_EXIT_READ_DR7: case SVM_EXIT_READ_DR8: case SVM_EXIT_READ_DR9:
3474	case SVM_EXIT_READ_DR10: case SVM_EXIT_READ_DR11: case SVM_EXIT_READ_DR12: case SVM_EXIT_READ_DR13:
3475	case SVM_EXIT_READ_DR14: case SVM_EXIT_READ_DR15:
3476	return hmR0SvmExitReadDRx(pVCpu, pCtx, pSvmTransient);
3477
3478	case SVM_EXIT_WRITE_DR0: case SVM_EXIT_WRITE_DR1: case SVM_EXIT_WRITE_DR2: case SVM_EXIT_WRITE_DR3:
3479	case SVM_EXIT_WRITE_DR6: case SVM_EXIT_WRITE_DR7: case SVM_EXIT_WRITE_DR8: case SVM_EXIT_WRITE_DR9:
3480	case SVM_EXIT_WRITE_DR10: case SVM_EXIT_WRITE_DR11: case SVM_EXIT_WRITE_DR12: case SVM_EXIT_WRITE_DR13:
3481	case SVM_EXIT_WRITE_DR14: case SVM_EXIT_WRITE_DR15:
3482	return hmR0SvmExitWriteDRx(pVCpu, pCtx, pSvmTransient);
3483
3484	case SVM_EXIT_TASK_SWITCH:
3485	return hmR0SvmExitTaskSwitch(pVCpu, pCtx, pSvmTransient);
3486
3487	case SVM_EXIT_VMMCALL:
3488	return hmR0SvmExitVmmCall(pVCpu, pCtx, pSvmTransient);
3489
3490	case SVM_EXIT_SHUTDOWN:
3491	return hmR0SvmExitShutdown(pVCpu, pCtx, pSvmTransient);
3492
3493	case SVM_EXIT_SMI:
3494	case SVM_EXIT_INIT:
3495	{
3496	/*
3497	* We don't intercept NMIs. As for INIT signals, it really shouldn't ever happen here. If it ever does,
3498	* we want to know about it so log the exit code and bail.
3499	*/
3500	AssertMsgFailed(("hmR0SvmHandleExit: Unexpected exit %#RX32\n", (uint32_t)pSvmTransient->u64ExitCode));
3501	pVCpu->hm.s.u32HMError = (uint32_t)pSvmTransient->u64ExitCode;
3502	return VERR_SVM_UNEXPECTED_EXIT;
3503	}
3504
3505	case SVM_EXIT_INVLPGA:
3506	case SVM_EXIT_RSM:
3507	case SVM_EXIT_VMRUN:
3508	case SVM_EXIT_VMLOAD:
3509	case SVM_EXIT_VMSAVE:
3510	case SVM_EXIT_STGI:
3511	case SVM_EXIT_CLGI:
3512	case SVM_EXIT_SKINIT:
3513	return hmR0SvmExitSetPendingXcptUD(pVCpu, pCtx, pSvmTransient);
3514
3515	#ifdef HMSVM_ALWAYS_TRAP_ALL_XCPTS
3516	case SVM_EXIT_EXCEPTION_0: /* X86_XCPT_DE */
3517	/* SVM_EXIT_EXCEPTION_1: / / X86_XCPT_DB - Handled above. */
3518	case SVM_EXIT_EXCEPTION_2: /* X86_XCPT_NMI */
3519	case SVM_EXIT_EXCEPTION_3: /* X86_XCPT_BP */
3520	case SVM_EXIT_EXCEPTION_4: /* X86_XCPT_OF */
3521	case SVM_EXIT_EXCEPTION_5: /* X86_XCPT_BR */
3522	case SVM_EXIT_EXCEPTION_6: /* X86_XCPT_UD */
3523	/* SVM_EXIT_EXCEPTION_7: / / X86_XCPT_NM - Handled above. */
3524	case SVM_EXIT_EXCEPTION_8: /* X86_XCPT_DF */
3525	case SVM_EXIT_EXCEPTION_9: /* X86_XCPT_CO_SEG_OVERRUN */
3526	case SVM_EXIT_EXCEPTION_A: /* X86_XCPT_TS */
3527	case SVM_EXIT_EXCEPTION_B: /* X86_XCPT_NP */
3528	case SVM_EXIT_EXCEPTION_C: /* X86_XCPT_SS */
3529	case SVM_EXIT_EXCEPTION_D: /* X86_XCPT_GP */
3530	/* SVM_EXIT_EXCEPTION_E: / / X86_XCPT_PF - Handled above. */
3531	/* SVM_EXIT_EXCEPTION_10: / / X86_XCPT_MF - Handled above. */
3532	case SVM_EXIT_EXCEPTION_11: /* X86_XCPT_AC */
3533	case SVM_EXIT_EXCEPTION_12: /* X86_XCPT_MC */
3534	case SVM_EXIT_EXCEPTION_13: /* X86_XCPT_XF */
3535	case SVM_EXIT_EXCEPTION_F: /* Reserved */
3536	case SVM_EXIT_EXCEPTION_14: case SVM_EXIT_EXCEPTION_15: case SVM_EXIT_EXCEPTION_16:
3537	case SVM_EXIT_EXCEPTION_17: case SVM_EXIT_EXCEPTION_18: case SVM_EXIT_EXCEPTION_19:
3538	case SVM_EXIT_EXCEPTION_1A: case SVM_EXIT_EXCEPTION_1B: case SVM_EXIT_EXCEPTION_1C:
3539	case SVM_EXIT_EXCEPTION_1D: case SVM_EXIT_EXCEPTION_1E: case SVM_EXIT_EXCEPTION_1F:
3540	{
3541	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
3542	SVMEVENT Event;
3543	Event.u = 0;
3544	Event.n.u1Valid = 1;
3545	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3546	Event.n.u8Vector = pSvmTransient->u64ExitCode - SVM_EXIT_EXCEPTION_0;
3547
3548	switch (Event.n.u8Vector)
3549	{
3550	case X86_XCPT_DE:
3551	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestDE);
3552	break;
3553
3554	case X86_XCPT_BP:
3555	/** Saves the wrong EIP on the stack (pointing to the int3) instead of the
3556	* next instruction. */
3557	/** @todo Investigate this later. */
3558	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestBP);
3559	break;
3560
3561	case X86_XCPT_UD:
3562	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestUD);
3563	break;
3564
3565	case X86_XCPT_NP:
3566	Event.n.u1ErrorCodeValid = 1;
3567	Event.n.u32ErrorCode = pVmcb->ctrl.u64ExitInfo1;
3568	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestNP);
3569	break;
3570
3571	case X86_XCPT_SS:
3572	Event.n.u1ErrorCodeValid = 1;
3573	Event.n.u32ErrorCode = pVmcb->ctrl.u64ExitInfo1;
3574	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestSS);
3575	break;
3576
3577	case X86_XCPT_GP:
3578	Event.n.u1ErrorCodeValid = 1;
3579	Event.n.u32ErrorCode = pVmcb->ctrl.u64ExitInfo1;
3580	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestGP);
3581	break;
3582
3583	default:
3584	AssertMsgFailed(("hmR0SvmHandleExit: Unexpected exit caused by exception %#x\n", Event.n.u8Vector));
3585	pVCpu->hm.s.u32HMError = Event.n.u8Vector;
3586	return VERR_SVM_UNEXPECTED_XCPT_EXIT;
3587	}
3588
3589	Log4(("#Xcpt: Vector=%#x at CS:RIP=%04x:%RGv\n", Event.n.u8Vector, pCtx->cs.Sel, (RTGCPTR)pCtx->rip));
3590	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3591	return VINF_SUCCESS;
3592	}
3593	#endif /* HMSVM_ALWAYS_TRAP_ALL_XCPTS */
3594
3595	default:
3596	{
3597	AssertMsgFailed(("hmR0SvmHandleExit: Unknown exit code %#x\n", u32ExitCode));
3598	pVCpu->hm.s.u32HMError = u32ExitCode;
3599	return VERR_SVM_UNKNOWN_EXIT;
3600	}
3601	}
3602	}
3603	}
3604	return VERR_INTERNAL_ERROR_5; /* Should never happen. */
3605	}
3606
3607
3608	#ifdef DEBUG
3609	/* Is there some generic IPRT define for this that are not in Runtime/internal/\* ?? */
3610	# define HMSVM_ASSERT_PREEMPT_CPUID_VAR() \
3611	RTCPUID const idAssertCpu = RTThreadPreemptIsEnabled(NIL_RTTHREAD) ? NIL_RTCPUID : RTMpCpuId()
3612
3613	# define HMSVM_ASSERT_PREEMPT_CPUID() \
3614	do \
3615	{ \
3616	RTCPUID const idAssertCpuNow = RTThreadPreemptIsEnabled(NIL_RTTHREAD) ? NIL_RTCPUID : RTMpCpuId(); \
3617	AssertMsg(idAssertCpu == idAssertCpuNow, ("SVM %#x, %#x\n", idAssertCpu, idAssertCpuNow)); \
3618	} while (0)
3619
3620	# define HMSVM_VALIDATE_EXIT_HANDLER_PARAMS() \
3621	do { \
3622	AssertPtr(pVCpu); \
3623	AssertPtr(pCtx); \
3624	AssertPtr(pSvmTransient); \
3625	Assert(ASMIntAreEnabled()); \
3626	HMSVM_ASSERT_PREEMPT_SAFE(); \
3627	HMSVM_ASSERT_PREEMPT_CPUID_VAR(); \
3628	Log4Func(("vcpu[%u] -v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-\n", (uint32_t)pVCpu->idCpu)); \
3629	HMSVM_ASSERT_PREEMPT_SAFE(); \
3630	if (VMMR0IsLogFlushDisabled(pVCpu)) \
3631	HMSVM_ASSERT_PREEMPT_CPUID(); \
3632	} while (0)
3633	#else /* Release builds */
3634	# define HMSVM_VALIDATE_EXIT_HANDLER_PARAMS() do { NOREF(pVCpu); NOREF(pCtx); NOREF(pSvmTransient); } while (0)
3635	#endif
3636
3637
3638	/**
3639	* Worker for hmR0SvmInterpretInvlpg().
3640	*
3641	* @return VBox status code.
3642	* @param pVCpu Pointer to the VMCPU.
3643	* @param pCpu Pointer to the disassembler state.
3644	* @param pRegFrame Pointer to the register frame.
3645	*/
3646	static int hmR0SvmInterpretInvlPgEx(PVMCPU pVCpu, PDISCPUSTATE pCpu, PCPUMCTXCORE pRegFrame)
3647	{
3648	DISQPVPARAMVAL Param1;
3649	RTGCPTR GCPtrPage;
3650
3651	int rc = DISQueryParamVal(pRegFrame, pCpu, &pCpu->Param1, &Param1, DISQPVWHICH_SRC);
3652	if (RT_FAILURE(rc))
3653	return VERR_EM_INTERPRETER;
3654
3655	if ( Param1.type == DISQPV_TYPE_IMMEDIATE
3656	\|\| Param1.type == DISQPV_TYPE_ADDRESS)
3657	{
3658	if (!(Param1.flags & (DISQPV_FLAG_32 \| DISQPV_FLAG_64)))
3659	return VERR_EM_INTERPRETER;
3660
3661	GCPtrPage = Param1.val.val64;
3662	VBOXSTRICTRC rc2 = EMInterpretInvlpg(pVCpu->CTX_SUFF(pVM), pVCpu, pRegFrame, GCPtrPage);
3663	rc = VBOXSTRICTRC_VAL(rc2);
3664	}
3665	else
3666	{
3667	Log4(("hmR0SvmInterpretInvlPgEx invalid parameter type %#x\n", Param1.type));
3668	rc = VERR_EM_INTERPRETER;
3669	}
3670
3671	return rc;
3672	}
3673
3674
3675	/**
3676	* Interprets INVLPG.
3677	*
3678	* @returns VBox status code.
3679	* @retval VINF_* Scheduling instructions.
3680	* @retval VERR_EM_INTERPRETER Something we can't cope with.
3681	* @retval VERR_* Fatal errors.
3682	*
3683	* @param pVM Pointer to the VM.
3684	* @param pRegFrame Pointer to the register frame.
3685	*
3686	* @remarks Updates the RIP if the instruction was executed successfully.
3687	*/
3688	static int hmR0SvmInterpretInvlpg(PVM pVM, PVMCPU pVCpu, PCPUMCTXCORE pRegFrame)
3689	{
3690	/* Only allow 32 & 64 bit code. */
3691	if (CPUMGetGuestCodeBits(pVCpu) != 16)
3692	{
3693	PDISSTATE pDis = &pVCpu->hm.s.DisState;
3694	int rc = EMInterpretDisasCurrent(pVM, pVCpu, pDis, NULL /* pcbInstr */);
3695	if ( RT_SUCCESS(rc)
3696	&& pDis->pCurInstr->uOpcode == OP_INVLPG)
3697	{
3698	rc = hmR0SvmInterpretInvlPgEx(pVCpu, pDis, pRegFrame);
3699	if (RT_SUCCESS(rc))
3700	pRegFrame->rip += pDis->cbInstr;
3701	return rc;
3702	}
3703	else
3704	Log4(("hmR0SvmInterpretInvlpg: EMInterpretDisasCurrent returned %Rrc uOpCode=%#x\n", rc, pDis->pCurInstr->uOpcode));
3705	}
3706	return VERR_EM_INTERPRETER;
3707	}
3708
3709
3710	/**
3711	* Sets an invalid-opcode (#UD) exception as pending-for-injection into the VM.
3712	*
3713	* @param pVCpu Pointer to the VMCPU.
3714	*/
3715	DECLINLINE(void) hmR0SvmSetPendingXcptUD(PVMCPU pVCpu)
3716	{
3717	SVMEVENT Event;
3718	Event.u = 0;
3719	Event.n.u1Valid = 1;
3720	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3721	Event.n.u8Vector = X86_XCPT_UD;
3722	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3723	}
3724
3725
3726	/**
3727	* Sets a debug (#DB) exception as pending-for-injection into the VM.
3728	*
3729	* @param pVCpu Pointer to the VMCPU.
3730	*/
3731	DECLINLINE(void) hmR0SvmSetPendingXcptDB(PVMCPU pVCpu)
3732	{
3733	SVMEVENT Event;
3734	Event.u = 0;
3735	Event.n.u1Valid = 1;
3736	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3737	Event.n.u8Vector = X86_XCPT_DB;
3738	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3739	}
3740
3741
3742	/**
3743	* Sets a page fault (#PF) exception as pending-for-injection into the VM.
3744	*
3745	* @param pVCpu Pointer to the VMCPU.
3746	* @param pCtx Pointer to the guest-CPU context.
3747	* @param u32ErrCode The error-code for the page-fault.
3748	* @param uFaultAddress The page fault address (CR2).
3749	*
3750	* @remarks This updates the guest CR2 with @a uFaultAddress!
3751	*/
3752	DECLINLINE(void) hmR0SvmSetPendingXcptPF(PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t u32ErrCode, RTGCUINTPTR uFaultAddress)
3753	{
3754	SVMEVENT Event;
3755	Event.u = 0;
3756	Event.n.u1Valid = 1;
3757	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3758	Event.n.u8Vector = X86_XCPT_PF;
3759	Event.n.u1ErrorCodeValid = 1;
3760	Event.n.u32ErrorCode = u32ErrCode;
3761
3762	/* Update CR2 of the guest. */
3763	if (pCtx->cr2 != uFaultAddress)
3764	{
3765	pCtx->cr2 = uFaultAddress;
3766	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR2);
3767	}
3768
3769	hmR0SvmSetPendingEvent(pVCpu, &Event, uFaultAddress);
3770	}
3771
3772
3773	/**
3774	* Sets a device-not-available (#NM) exception as pending-for-injection into the
3775	* VM.
3776	*
3777	* @param pVCpu Pointer to the VMCPU.
3778	*/
3779	DECLINLINE(void) hmR0SvmSetPendingXcptNM(PVMCPU pVCpu)
3780	{
3781	SVMEVENT Event;
3782	Event.u = 0;
3783	Event.n.u1Valid = 1;
3784	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3785	Event.n.u8Vector = X86_XCPT_NM;
3786	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3787	}
3788
3789
3790	/**
3791	* Sets a math-fault (#MF) exception as pending-for-injection into the VM.
3792	*
3793	* @param pVCpu Pointer to the VMCPU.
3794	*/
3795	DECLINLINE(void) hmR0SvmSetPendingXcptMF(PVMCPU pVCpu)
3796	{
3797	SVMEVENT Event;
3798	Event.u = 0;
3799	Event.n.u1Valid = 1;
3800	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3801	Event.n.u8Vector = X86_XCPT_MF;
3802	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3803	}
3804
3805
3806	/**
3807	* Sets a double fault (#DF) exception as pending-for-injection into the VM.
3808	*
3809	* @param pVCpu Pointer to the VMCPU.
3810	*/
3811	DECLINLINE(void) hmR0SvmSetPendingXcptDF(PVMCPU pVCpu)
3812	{
3813	SVMEVENT Event;
3814	Event.u = 0;
3815	Event.n.u1Valid = 1;
3816	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3817	Event.n.u8Vector = X86_XCPT_DF;
3818	Event.n.u1ErrorCodeValid = 1;
3819	Event.n.u32ErrorCode = 0;
3820	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3821	}
3822
3823
3824	/**
3825	* Emulates a simple MOV TPR (CR8) instruction, used for TPR patching on 32-bit
3826	* guests. This simply looks up the patch record at EIP and does the required.
3827	*
3828	* This VMMCALL is used a fallback mechanism when mov to/from cr8 isn't exactly
3829	* like how we want it to be (e.g. not followed by shr 4 as is usually done for
3830	* TPR). See hmR3ReplaceTprInstr() for the details.
3831	*
3832	* @returns VBox status code.
3833	* @retval VINF_SUCCESS if the access was handled successfully.
3834	* @retval VERR_NOT_FOUND if no patch record for this eip could be found.
3835	* @retval VERR_SVM_UNEXPECTED_PATCH_TYPE if the found patch type is invalid.
3836	*
3837	* @param pVM Pointer to the VM.
3838	* @param pVCpu Pointer to the VMCPU.
3839	* @param pCtx Pointer to the guest-CPU context.
3840	*/
3841	static int hmR0SvmEmulateMovTpr(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3842	{
3843	Log4(("Emulated VMMCall TPR access replacement at RIP=%RGv\n", pCtx->rip));
3844	bool fPatchFound = false;
3845	for (;;)
3846	{
3847	bool fPending;
3848	uint8_t u8Tpr;
3849
3850	PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
3851	if (!pPatch)
3852	break;
3853
3854	fPatchFound = true;
3855	switch (pPatch->enmType)
3856	{
3857	case HMTPRINSTR_READ:
3858	{
3859	int rc = PDMApicGetTPR(pVCpu, &u8Tpr, &fPending, NULL /* pu8PendingIrq */);
3860	AssertRC(rc);
3861
3862	rc = DISWriteReg32(CPUMCTX2CORE(pCtx), pPatch->uDstOperand, u8Tpr);
3863	AssertRC(rc);
3864	pCtx->rip += pPatch->cbOp;
3865	break;
3866	}
3867
3868	case HMTPRINSTR_WRITE_REG:
3869	case HMTPRINSTR_WRITE_IMM:
3870	{
3871	if (pPatch->enmType == HMTPRINSTR_WRITE_REG)
3872	{
3873	uint32_t u32Val;
3874	int rc = DISFetchReg32(CPUMCTX2CORE(pCtx), pPatch->uSrcOperand, &u32Val);
3875	AssertRC(rc);
3876	u8Tpr = u32Val;
3877	}
3878	else
3879	u8Tpr = (uint8_t)pPatch->uSrcOperand;
3880
3881	int rc2 = PDMApicSetTPR(pVCpu, u8Tpr);
3882	AssertRC(rc2);
3883	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
3884
3885	pCtx->rip += pPatch->cbOp;
3886	break;
3887	}
3888
3889	default:
3890	AssertMsgFailed(("Unexpected patch type %d\n", pPatch->enmType));
3891	pVCpu->hm.s.u32HMError = pPatch->enmType;
3892	return VERR_SVM_UNEXPECTED_PATCH_TYPE;
3893	}
3894	}
3895
3896	if (fPatchFound)
3897	return VINF_SUCCESS;
3898	return VERR_NOT_FOUND;
3899	}
3900
3901
3902	/**
3903	* Determines if an exception is a contributory exception. Contributory
3904	* exceptions are ones which can cause double-faults. Page-fault is
3905	* intentionally not included here as it's a conditional contributory exception.
3906	*
3907	* @returns true if the exception is contributory, false otherwise.
3908	* @param uVector The exception vector.
3909	*/
3910	DECLINLINE(bool) hmR0SvmIsContributoryXcpt(const uint32_t uVector)
3911	{
3912	switch (uVector)
3913	{
3914	case X86_XCPT_GP:
3915	case X86_XCPT_SS:
3916	case X86_XCPT_NP:
3917	case X86_XCPT_TS:
3918	case X86_XCPT_DE:
3919	return true;
3920	default:
3921	break;
3922	}
3923	return false;
3924	}
3925
3926
3927	/**
3928	* Handle a condition that occurred while delivering an event through the guest
3929	* IDT.
3930	*
3931	* @returns VBox status code (informational error codes included).
3932	* @retval VINF_SUCCESS if we should continue handling the VM-exit.
3933	* @retval VINF_HM_DOUBLE_FAULT if a #DF condition was detected and we ought to
3934	* continue execution of the guest which will delivery the #DF.
3935	* @retval VINF_EM_RESET if we detected a triple-fault condition.
3936	*
3937	* @param pVCpu Pointer to the VMCPU.
3938	* @param pCtx Pointer to the guest-CPU context.
3939	* @param pSvmTransient Pointer to the SVM transient structure.
3940	*
3941	* @remarks No-long-jump zone!!!
3942	*/
3943	static int hmR0SvmCheckExitDueToEventDelivery(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
3944	{
3945	int rc = VINF_SUCCESS;
3946	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
3947
3948	/* See AMD spec. 15.7.3 "EXITINFO Pseudo-Code". The EXITINTINFO (if valid) contains the prior exception (IDT vector)
3949	* that was trying to be delivered to the guest which caused a #VMEXIT which was intercepted (Exit vector). */
3950	if (pVmcb->ctrl.ExitIntInfo.n.u1Valid)
3951	{
3952	uint8_t uIdtVector = pVmcb->ctrl.ExitIntInfo.n.u8Vector;
3953
3954	typedef enum
3955	{
3956	SVMREFLECTXCPT_XCPT, /* Reflect the exception to the guest or for further evaluation by VMM. */
3957	SVMREFLECTXCPT_DF, /* Reflect the exception as a double-fault to the guest. */
3958	SVMREFLECTXCPT_TF, /* Indicate a triple faulted state to the VMM. */
3959	SVMREFLECTXCPT_NONE /* Nothing to reflect. */
3960	} SVMREFLECTXCPT;
3961
3962	SVMREFLECTXCPT enmReflect = SVMREFLECTXCPT_NONE;
3963	if (pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_EXCEPTION)
3964	{
3965	if (pSvmTransient->u64ExitCode - SVM_EXIT_EXCEPTION_0 <= SVM_EXIT_EXCEPTION_1F)
3966	{
3967	uint8_t uExitVector = (uint8_t)(pSvmTransient->u64ExitCode - SVM_EXIT_EXCEPTION_0);
3968
3969	#ifdef VBOX_STRICT
3970	if ( hmR0SvmIsContributoryXcpt(uIdtVector)
3971	&& uExitVector == X86_XCPT_PF)
3972	{
3973	Log4(("IDT: Contributory #PF uCR2=%#RX64\n", pVCpu->idCpu, pCtx->cr2));
3974	}
3975	#endif
3976	if ( uExitVector == X86_XCPT_PF
3977	&& uIdtVector == X86_XCPT_PF)
3978	{
3979	pSvmTransient->fVectoringPF = true;
3980	Log4(("IDT: Vectoring #PF uCR2=%#RX64\n", pCtx->cr2));
3981	}
3982	else if ( (pVmcb->ctrl.u32InterceptException & HMSVM_CONTRIBUTORY_XCPT_MASK)
3983	&& hmR0SvmIsContributoryXcpt(uExitVector)
3984	&& ( hmR0SvmIsContributoryXcpt(uIdtVector)
3985	\|\| uIdtVector == X86_XCPT_PF))
3986	{
3987	enmReflect = SVMREFLECTXCPT_DF;
3988	Log4(("IDT: Pending vectoring #DF %#RX64 uIdtVector=%#x uExitVector=%#x\n", pVCpu->hm.s.Event.u64IntInfo,
3989	uIdtVector, uExitVector));
3990	}
3991	else if (uIdtVector == X86_XCPT_DF)
3992	{
3993	enmReflect = SVMREFLECTXCPT_TF;
3994	Log4(("IDT: Pending vectoring triple-fault %#RX64 uIdtVector=%#x uExitVector=%#x\n",
3995	pVCpu->hm.s.Event.u64IntInfo, uIdtVector, uExitVector));
3996	}
3997	else
3998	enmReflect = SVMREFLECTXCPT_XCPT;
3999	}
4000	else
4001	{
4002	/*
4003	* If event delivery caused an #VMEXIT that is not an exception (e.g. #NPF) then reflect the original
4004	* exception to the guest after handling the VM-exit.
4005	*/
4006	enmReflect = SVMREFLECTXCPT_XCPT;
4007	}
4008	}
4009	else if (pVmcb->ctrl.ExitIntInfo.n.u3Type != SVM_EVENT_SOFTWARE_INT)
4010	{
4011	/* Ignore software interrupts (INT n) as they reoccur when restarting the instruction. */
4012	enmReflect = SVMREFLECTXCPT_XCPT;
4013	}
4014
4015	switch (enmReflect)
4016	{
4017	case SVMREFLECTXCPT_XCPT:
4018	{
4019	Assert(pVmcb->ctrl.ExitIntInfo.n.u3Type != SVM_EVENT_SOFTWARE_INT);
4020	hmR0SvmSetPendingEvent(pVCpu, &pVmcb->ctrl.ExitIntInfo, 0 /* GCPtrFaultAddress */);
4021
4022	/* If uExitVector is #PF, CR2 value will be updated from the VMCB if it's a guest #PF. See hmR0SvmExitXcptPF(). */
4023	Log4(("IDT: Pending vectoring event %#RX64 ErrValid=%RTbool Err=%#RX32\n", pVmcb->ctrl.ExitIntInfo.u,
4024	!!pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid, pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode));
4025	break;
4026	}
4027
4028	case SVMREFLECTXCPT_DF:
4029	{
4030	hmR0SvmSetPendingXcptDF(pVCpu);
4031	rc = VINF_HM_DOUBLE_FAULT;
4032	break;
4033	}
4034
4035	case SVMREFLECTXCPT_TF:
4036	{
4037	rc = VINF_EM_RESET;
4038	break;
4039	}
4040
4041	default:
4042	Assert(rc == VINF_SUCCESS);
4043	break;
4044	}
4045	}
4046	Assert(rc == VINF_SUCCESS \|\| rc == VINF_HM_DOUBLE_FAULT \|\| rc == VINF_EM_RESET);
4047	NOREF(pCtx);
4048	return rc;
4049	}
4050
4051
4052	/**
4053	* Advances the guest RIP in the if the NRIP_SAVE feature is supported by the
4054	* CPU, otherwise advances the RIP by @a cb bytes.
4055	*
4056	* @param pVCpu Pointer to the VMCPU.
4057	* @param pCtx Pointer to the guest-CPU context.
4058	* @param cb RIP increment value in bytes.
4059	*
4060	* @remarks Use this function only from #VMEXIT's where the NRIP value is valid
4061	* when NRIP_SAVE is supported by the CPU!
4062	*/
4063	DECLINLINE(void) hmR0SvmUpdateRip(PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t cb)
4064	{
4065	if (pVCpu->CTX_SUFF(pVM)->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_NRIP_SAVE)
4066	{
4067	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4068	pCtx->rip = pVmcb->ctrl.u64NextRIP;
4069	}
4070	else
4071	pCtx->rip += cb;
4072	}
4073
4074
4075	/* -=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= */
4076	/* -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- #VMEXIT handlers -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- */
4077	/* -=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= */
4078
4079	/** @name VM-exit handlers.
4080	* @{
4081	*/
4082
4083	/**
4084	* #VMEXIT handler for external interrupts, NMIs, FPU assertion freeze and INIT
4085	* signals (SVM_EXIT_INTR, SVM_EXIT_NMI, SVM_EXIT_FERR_FREEZE, SVM_EXIT_INIT).
4086	*/
4087	HMSVM_EXIT_DECL hmR0SvmExitIntr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4088	{
4089	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4090
4091	if (pSvmTransient->u64ExitCode == SVM_EXIT_NMI)
4092	STAM_REL_COUNTER_INC(&pVCpu->hm.s.StatExitHostNmiInGC);
4093	else if (pSvmTransient->u64ExitCode == SVM_EXIT_INTR)
4094	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitExtInt);
4095
4096	/*
4097	* AMD-V has no preemption timer and the generic periodic preemption timer has no way to signal -before- the timer
4098	* fires if the current interrupt is our own timer or a some other host interrupt. We also cannot examine what
4099	* interrupt it is until the host actually take the interrupt.
4100	*
4101	* Going back to executing guest code here unconditionally causes random scheduling problems (observed on an
4102	* AMD Phenom 9850 Quad-Core on Windows 64-bit host).
4103	*/
4104	return VINF_EM_RAW_INTERRUPT;
4105	}
4106
4107
4108	/**
4109	* #VMEXIT handler for WBINVD (SVM_EXIT_WBINVD). Conditional #VMEXIT.
4110	*/
4111	HMSVM_EXIT_DECL hmR0SvmExitWbinvd(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4112	{
4113	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4114
4115	hmR0SvmUpdateRip(pVCpu, pCtx, 2);
4116	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitWbinvd);
4117	int rc = VINF_SUCCESS;
4118	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4119	return rc;
4120	}
4121
4122
4123	/**
4124	* #VMEXIT handler for INVD (SVM_EXIT_INVD). Unconditional #VMEXIT.
4125	*/
4126	HMSVM_EXIT_DECL hmR0SvmExitInvd(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4127	{
4128	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4129
4130	hmR0SvmUpdateRip(pVCpu, pCtx, 2);
4131	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitInvd);
4132	int rc = VINF_SUCCESS;
4133	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4134	return rc;
4135	}
4136
4137
4138	/**
4139	* #VMEXIT handler for INVD (SVM_EXIT_CPUID). Conditional #VMEXIT.
4140	*/
4141	HMSVM_EXIT_DECL hmR0SvmExitCpuid(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4142	{
4143	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4144	PVM pVM = pVCpu->CTX_SUFF(pVM);
4145	int rc = EMInterpretCpuId(pVM, pVCpu, CPUMCTX2CORE(pCtx));
4146	if (RT_LIKELY(rc == VINF_SUCCESS))
4147	{
4148	hmR0SvmUpdateRip(pVCpu, pCtx, 2);
4149	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4150	}
4151	else
4152	{
4153	AssertMsgFailed(("hmR0SvmExitCpuid: EMInterpretCpuId failed with %Rrc\n", rc));
4154	rc = VERR_EM_INTERPRETER;
4155	}
4156	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCpuid);
4157	return rc;
4158	}
4159
4160
4161	/**
4162	* #VMEXIT handler for RDTSC (SVM_EXIT_RDTSC). Conditional #VMEXIT.
4163	*/
4164	HMSVM_EXIT_DECL hmR0SvmExitRdtsc(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4165	{
4166	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4167	PVM pVM = pVCpu->CTX_SUFF(pVM);
4168	int rc = EMInterpretRdtsc(pVM, pVCpu, CPUMCTX2CORE(pCtx));
4169	if (RT_LIKELY(rc == VINF_SUCCESS))
4170	{
4171	hmR0SvmUpdateRip(pVCpu, pCtx, 2);
4172	pSvmTransient->fUpdateTscOffsetting = true;
4173
4174	/* Single step check. */
4175	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4176	}
4177	else
4178	{
4179	AssertMsgFailed(("hmR0SvmExitRdtsc: EMInterpretRdtsc failed with %Rrc\n", rc));
4180	rc = VERR_EM_INTERPRETER;
4181	}
4182	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdtsc);
4183	return rc;
4184	}
4185
4186
4187	/**
4188	* #VMEXIT handler for RDTSCP (SVM_EXIT_RDTSCP). Conditional #VMEXIT.
4189	*/
4190	HMSVM_EXIT_DECL hmR0SvmExitRdtscp(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4191	{
4192	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4193	int rc = EMInterpretRdtscp(pVCpu->CTX_SUFF(pVM), pVCpu, pCtx);
4194	if (RT_LIKELY(rc == VINF_SUCCESS))
4195	{
4196	hmR0SvmUpdateRip(pVCpu, pCtx, 3);
4197	pSvmTransient->fUpdateTscOffsetting = true;
4198	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4199	}
4200	else
4201	{
4202	AssertMsgFailed(("hmR0SvmExitRdtsc: EMInterpretRdtscp failed with %Rrc\n", rc));
4203	rc = VERR_EM_INTERPRETER;
4204	}
4205	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdtscp);
4206	return rc;
4207	}
4208
4209
4210	/**
4211	* #VMEXIT handler for RDPMC (SVM_EXIT_RDPMC). Conditional #VMEXIT.
4212	*/
4213	HMSVM_EXIT_DECL hmR0SvmExitRdpmc(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4214	{
4215	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4216	int rc = EMInterpretRdpmc(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
4217	if (RT_LIKELY(rc == VINF_SUCCESS))
4218	{
4219	hmR0SvmUpdateRip(pVCpu, pCtx, 2);
4220	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4221	}
4222	else
4223	{
4224	AssertMsgFailed(("hmR0SvmExitRdpmc: EMInterpretRdpmc failed with %Rrc\n", rc));
4225	rc = VERR_EM_INTERPRETER;
4226	}
4227	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdpmc);
4228	return rc;
4229	}
4230
4231
4232	/**
4233	* #VMEXIT handler for INVLPG (SVM_EXIT_INVLPG). Conditional #VMEXIT.
4234	*/
4235	HMSVM_EXIT_DECL hmR0SvmExitInvlpg(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4236	{
4237	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4238	PVM pVM = pVCpu->CTX_SUFF(pVM);
4239	Assert(!pVM->hm.s.fNestedPaging);
4240
4241	/** @todo Decode Assist. */
4242	int rc = hmR0SvmInterpretInvlpg(pVM, pVCpu, CPUMCTX2CORE(pCtx)); /* Updates RIP if successful. */
4243	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitInvlpg);
4244	Assert(rc == VINF_SUCCESS \|\| rc == VERR_EM_INTERPRETER);
4245	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4246	return rc;
4247	}
4248
4249
4250	/**
4251	* #VMEXIT handler for HLT (SVM_EXIT_HLT). Conditional #VMEXIT.
4252	*/
4253	HMSVM_EXIT_DECL hmR0SvmExitHlt(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4254	{
4255	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4256	hmR0SvmUpdateRip(pVCpu, pCtx, 1);
4257	int rc = EMShouldContinueAfterHalt(pVCpu, pCtx) ? VINF_SUCCESS : VINF_EM_HALT;
4258	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4259	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitHlt);
4260	return rc;
4261	}
4262
4263
4264	/**
4265	* #VMEXIT handler for MONITOR (SVM_EXIT_MONITOR). Conditional #VMEXIT.
4266	*/
4267	HMSVM_EXIT_DECL hmR0SvmExitMonitor(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4268	{
4269	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4270	int rc = EMInterpretMonitor(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
4271	if (RT_LIKELY(rc == VINF_SUCCESS))
4272	{
4273	hmR0SvmUpdateRip(pVCpu, pCtx, 3);
4274	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4275	}
4276	else
4277	{
4278	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMonitor: EMInterpretMonitor failed with %Rrc\n", rc));
4279	rc = VERR_EM_INTERPRETER;
4280	}
4281	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitMonitor);
4282	return rc;
4283	}
4284
4285
4286	/**
4287	* #VMEXIT handler for MWAIT (SVM_EXIT_MWAIT). Conditional #VMEXIT.
4288	*/
4289	HMSVM_EXIT_DECL hmR0SvmExitMwait(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4290	{
4291	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4292	VBOXSTRICTRC rc2 = EMInterpretMWait(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
4293	int rc = VBOXSTRICTRC_VAL(rc2);
4294	if ( rc == VINF_EM_HALT
4295	\|\| rc == VINF_SUCCESS)
4296	{
4297	hmR0SvmUpdateRip(pVCpu, pCtx, 3);
4298
4299	if ( rc == VINF_EM_HALT
4300	&& EMMonitorWaitShouldContinue(pVCpu, pCtx))
4301	{
4302	rc = VINF_SUCCESS;
4303	}
4304	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4305	}
4306	else
4307	{
4308	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMwait: EMInterpretMWait failed with %Rrc\n", rc));
4309	rc = VERR_EM_INTERPRETER;
4310	}
4311	AssertMsg(rc == VINF_SUCCESS \|\| rc == VINF_EM_HALT \|\| rc == VERR_EM_INTERPRETER,
4312	("hmR0SvmExitMwait: EMInterpretMWait failed rc=%Rrc\n", rc));
4313	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitMwait);
4314	return rc;
4315	}
4316
4317
4318	/**
4319	* #VMEXIT handler for shutdown (triple-fault) (SVM_EXIT_SHUTDOWN).
4320	* Conditional #VMEXIT.
4321	*/
4322	HMSVM_EXIT_DECL hmR0SvmExitShutdown(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4323	{
4324	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4325	return VINF_EM_RESET;
4326	}
4327
4328
4329	/**
4330	* #VMEXIT handler for CRx reads (SVM_EXIT_READ_CR*). Conditional #VMEXIT.
4331	*/
4332	HMSVM_EXIT_DECL hmR0SvmExitReadCRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4333	{
4334	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4335
4336	Log4(("hmR0SvmExitReadCRx: CS:RIP=%04x:%#RX64\n", pCtx->cs.Sel, pCtx->rip));
4337
4338	/** @todo Decode Assist. */
4339	VBOXSTRICTRC rc2 = EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */);
4340	int rc = VBOXSTRICTRC_VAL(rc2);
4341	AssertMsg(rc == VINF_SUCCESS \|\| rc == VERR_EM_INTERPRETER \|\| rc == VINF_PGM_CHANGE_MODE \|\| rc == VINF_PGM_SYNC_CR3,
4342	("hmR0SvmExitReadCRx: EMInterpretInstruction failed rc=%Rrc\n", rc));
4343	Assert((pSvmTransient->u64ExitCode - SVM_EXIT_READ_CR0) <= 15);
4344	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCRxRead[pSvmTransient->u64ExitCode - SVM_EXIT_READ_CR0]);
4345	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4346	return rc;
4347	}
4348
4349
4350	/**
4351	* #VMEXIT handler for CRx writes (SVM_EXIT_WRITE_CR*). Conditional #VMEXIT.
4352	*/
4353	HMSVM_EXIT_DECL hmR0SvmExitWriteCRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4354	{
4355	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4356	/** @todo Decode Assist. */
4357	VBOXSTRICTRC rc2 = EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */);
4358	int rc = VBOXSTRICTRC_VAL(rc2);
4359	if (rc == VINF_SUCCESS)
4360	{
4361	/* RIP has been updated by EMInterpretInstruction(). */
4362	Assert((pSvmTransient->u64ExitCode - SVM_EXIT_WRITE_CR0) <= 15);
4363	switch (pSvmTransient->u64ExitCode - SVM_EXIT_WRITE_CR0)
4364	{
4365	case 0: /* CR0. */
4366	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR0);
4367	break;
4368
4369	case 3: /* CR3. */
4370	Assert(!pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
4371	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR3);
4372	break;
4373
4374	case 4: /* CR4. */
4375	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR4);
4376	break;
4377
4378	case 8: /* CR8 (TPR). */
4379	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
4380	break;
4381
4382	default:
4383	AssertMsgFailed(("hmR0SvmExitWriteCRx: Invalid/Unexpected Write-CRx exit. u64ExitCode=%#RX64 %#x CRx=%#RX64\n",
4384	pSvmTransient->u64ExitCode, pSvmTransient->u64ExitCode - SVM_EXIT_WRITE_CR0));
4385	break;
4386	}
4387	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4388	}
4389	else
4390	Assert(rc == VERR_EM_INTERPRETER \|\| rc == VINF_PGM_CHANGE_MODE \|\| rc == VINF_PGM_SYNC_CR3);
4391	return rc;
4392	}
4393
4394
4395	/**
4396	* #VMEXIT handler for instructions that result in a #UD exception delivered to
4397	* the guest.
4398	*/
4399	HMSVM_EXIT_DECL hmR0SvmExitSetPendingXcptUD(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4400	{
4401	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4402	hmR0SvmSetPendingXcptUD(pVCpu);
4403	return VINF_SUCCESS;
4404	}
4405
4406
4407	/**
4408	* #VMEXIT handler for MSR read and writes (SVM_EXIT_MSR). Conditional #VMEXIT.
4409	*/
4410	HMSVM_EXIT_DECL hmR0SvmExitMsr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4411	{
4412	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4413	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4414	PVM pVM = pVCpu->CTX_SUFF(pVM);
4415
4416	int rc;
4417	if (pVmcb->ctrl.u64ExitInfo1 == SVM_EXIT1_MSR_WRITE)
4418	{
4419	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitWrmsr);
4420
4421	/* Handle TPR patching; intercepted LSTAR write. */
4422	if ( pVM->hm.s.fTPRPatchingActive
4423	&& pCtx->ecx == MSR_K8_LSTAR)
4424	{
4425	if ((pCtx->eax & 0xff) != pSvmTransient->u8GuestTpr)
4426	{
4427	/* Our patch code uses LSTAR for TPR caching for 32-bit guests. */
4428	int rc2 = PDMApicSetTPR(pVCpu, pCtx->eax & 0xff);
4429	AssertRC(rc2);
4430	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
4431	}
4432	hmR0SvmUpdateRip(pVCpu, pCtx, 2);
4433	rc = VINF_SUCCESS;
4434	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4435	return rc;
4436	}
4437
4438	if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_NRIP_SAVE)
4439	{
4440	rc = EMInterpretWrmsr(pVM, pVCpu, CPUMCTX2CORE(pCtx));
4441	if (RT_LIKELY(rc == VINF_SUCCESS))
4442	{
4443	pCtx->rip = pVmcb->ctrl.u64NextRIP;
4444	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4445	}
4446	else
4447	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMsr: EMInterpretWrmsr failed rc=%Rrc\n", rc));
4448	}
4449	else
4450	{
4451	rc = VBOXSTRICTRC_TODO(EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */));
4452	if (RT_LIKELY(rc == VINF_SUCCESS))
4453	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc); /* RIP updated by EMInterpretInstruction(). */
4454	else
4455	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMsr: WrMsr. EMInterpretInstruction failed rc=%Rrc\n", rc));
4456	}
4457
4458	if (rc == VINF_SUCCESS)
4459	{
4460	/* If this is an X2APIC WRMSR access, update the APIC state as well. */
4461	if ( pCtx->ecx >= MSR_IA32_X2APIC_START
4462	&& pCtx->ecx <= MSR_IA32_X2APIC_END)
4463	{
4464	/*
4465	* We've already saved the APIC related guest-state (TPR) in hmR0SvmPostRunGuest(). When full APIC register
4466	* virtualization is implemented we'll have to make sure APIC state is saved from the VMCB before
4467	* EMInterpretWrmsr() changes it.
4468	*/
4469	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
4470	}
4471	else if (pCtx->ecx == MSR_K6_EFER)
4472	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_EFER_MSR);
4473	else if (pCtx->ecx == MSR_IA32_TSC)
4474	pSvmTransient->fUpdateTscOffsetting = true;
4475	}
4476	}
4477	else
4478	{
4479	/* MSR Read access. */
4480	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdmsr);
4481	Assert(pVmcb->ctrl.u64ExitInfo1 == SVM_EXIT1_MSR_READ);
4482
4483	if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_NRIP_SAVE)
4484	{
4485	rc = EMInterpretRdmsr(pVM, pVCpu, CPUMCTX2CORE(pCtx));
4486	if (RT_LIKELY(rc == VINF_SUCCESS))
4487	{
4488	pCtx->rip = pVmcb->ctrl.u64NextRIP;
4489	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4490	}
4491	else
4492	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMsr: EMInterpretRdmsr failed rc=%Rrc\n", rc));
4493	}
4494	else
4495	{
4496	rc = VBOXSTRICTRC_TODO(EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0));
4497	if (RT_UNLIKELY(rc != VINF_SUCCESS))
4498	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMsr: RdMsr. EMInterpretInstruction failed rc=%Rrc\n", rc));
4499	/* RIP updated by EMInterpretInstruction(). */
4500	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4501	}
4502	}
4503
4504	/* RIP has been updated by EMInterpret[Rd\|Wr]msr(). */
4505	return rc;
4506	}
4507
4508
4509	/**
4510	* #VMEXIT handler for DRx read (SVM_EXIT_READ_DRx). Conditional #VMEXIT.
4511	*/
4512	HMSVM_EXIT_DECL hmR0SvmExitReadDRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4513	{
4514	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4515	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitDRxRead);
4516
4517	/* We should -not- get this VM-exit if the guest's debug registers were active. */
4518	if (pSvmTransient->fWasGuestDebugStateActive)
4519	{
4520	AssertMsgFailed(("hmR0SvmHandleExit: Unexpected exit %#RX32\n", (uint32_t)pSvmTransient->u64ExitCode));
4521	pVCpu->hm.s.u32HMError = (uint32_t)pSvmTransient->u64ExitCode;
4522	return VERR_SVM_UNEXPECTED_EXIT;
4523	}
4524
4525	/*
4526	* Lazy DR0-3 loading.
4527	*/
4528	if (!pSvmTransient->fWasHyperDebugStateActive)
4529	{
4530	Assert(!DBGFIsStepping(pVCpu)); Assert(!pVCpu->hm.s.fSingleInstruction);
4531	Log5(("hmR0SvmExitReadDRx: Lazy loading guest debug registers\n"));
4532
4533	/* Don't intercept DRx read and writes. */
4534	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4535	pVmcb->ctrl.u16InterceptRdDRx = 0;
4536	pVmcb->ctrl.u16InterceptWrDRx = 0;
4537	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
4538
4539	/* We're playing with the host CPU state here, make sure we don't preempt or longjmp. */
4540	VMMRZCallRing3Disable(pVCpu);
4541	HM_DISABLE_PREEMPT_IF_NEEDED();
4542
4543	/* Save the host & load the guest debug state, restart execution of the MOV DRx instruction. */
4544	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
4545	Assert(CPUMIsGuestDebugStateActive(pVCpu) \|\| HC_ARCH_BITS == 32);
4546
4547	HM_RESTORE_PREEMPT_IF_NEEDED();
4548	VMMRZCallRing3Enable(pVCpu);
4549
4550	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxContextSwitch);
4551	return VINF_SUCCESS;
4552	}
4553
4554	/*
4555	* Interpret the read/writing of DRx.
4556	*/
4557	/** @todo Decode assist. */
4558	VBOXSTRICTRC rc = EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */);
4559	Log5(("hmR0SvmExitReadDRx: Emulated DRx access: rc=%Rrc\n", VBOXSTRICTRC_VAL(rc)));
4560	if (RT_LIKELY(rc == VINF_SUCCESS))
4561	{
4562	/* Not necessary for read accesses but whatever doesn't hurt for now, will be fixed with decode assist. */
4563	/** @todo CPUM should set this flag! */
4564	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_DEBUG);
4565	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4566	}
4567	else
4568	Assert(rc == VERR_EM_INTERPRETER);
4569	return VBOXSTRICTRC_TODO(rc);
4570	}
4571
4572
4573	/**
4574	* #VMEXIT handler for DRx write (SVM_EXIT_WRITE_DRx). Conditional #VMEXIT.
4575	*/
4576	HMSVM_EXIT_DECL hmR0SvmExitWriteDRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4577	{
4578	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4579	/* For now it's the same since we interpret the instruction anyway. Will change when using of Decode Assist is implemented. */
4580	int rc = hmR0SvmExitReadDRx(pVCpu, pCtx, pSvmTransient);
4581	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitDRxWrite);
4582	STAM_COUNTER_DEC(&pVCpu->hm.s.StatExitDRxRead);
4583	return rc;
4584	}
4585
4586
4587	/**
4588	* #VMEXIT handler for I/O instructions (SVM_EXIT_IOIO). Conditional #VMEXIT.
4589	*/
4590	HMSVM_EXIT_DECL hmR0SvmExitIOInstr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4591	{
4592	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4593
4594	/* I/O operation lookup arrays. */
4595	static uint32_t const s_aIOSize[8] = { 0, 1, 2, 0, 4, 0, 0, 0 }; /* Size of the I/O accesses in bytes. */
4596	static uint32_t const s_aIOOpAnd[8] = { 0, 0xff, 0xffff, 0, 0xffffffff, 0, 0, 0 }; /* AND masks for saving
4597	the result (in AL/AX/EAX). */
4598	Log4(("hmR0SvmExitIOInstr: CS:RIP=%04x:%#RX64\n", pCtx->cs.Sel, pCtx->rip));
4599
4600	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4601	PVM pVM = pVCpu->CTX_SUFF(pVM);
4602
4603	/* Refer AMD spec. 15.10.2 "IN and OUT Behaviour" and Figure 15-2. "EXITINFO1 for IOIO Intercept" for the format. */
4604	SVMIOIOEXIT IoExitInfo;
4605	IoExitInfo.u = (uint32_t)pVmcb->ctrl.u64ExitInfo1;
4606	uint32_t uIOWidth = (IoExitInfo.u >> 4) & 0x7;
4607	uint32_t cbValue = s_aIOSize[uIOWidth];
4608	uint32_t uAndVal = s_aIOOpAnd[uIOWidth];
4609
4610	if (RT_UNLIKELY(!cbValue))
4611	{
4612	AssertMsgFailed(("hmR0SvmExitIOInstr: Invalid IO operation. uIOWidth=%u\n", uIOWidth));
4613	return VERR_EM_INTERPRETER;
4614	}
4615
4616	VBOXSTRICTRC rcStrict;
4617	if (IoExitInfo.n.u1STR)
4618	{
4619	/* INS/OUTS - I/O String instruction. */
4620	PDISCPUSTATE pDis = &pVCpu->hm.s.DisState;
4621
4622	/** @todo Huh? why can't we use the segment prefix information given by AMD-V
4623	* in EXITINFO1? Investigate once this thing is up and running. */
4624
4625	rcStrict = EMInterpretDisasCurrent(pVM, pVCpu, pDis, NULL);
4626	if (rcStrict == VINF_SUCCESS)
4627	{
4628	if (IoExitInfo.n.u1Type == SVM_IOIO_WRITE)
4629	{
4630	rcStrict = IOMInterpretOUTSEx(pVM, pVCpu, CPUMCTX2CORE(pCtx), IoExitInfo.n.u16Port, pDis->fPrefix,
4631	(DISCPUMODE)pDis->uAddrMode, cbValue);
4632	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOStringWrite);
4633	}
4634	else
4635	{
4636	rcStrict = IOMInterpretINSEx(pVM, pVCpu, CPUMCTX2CORE(pCtx), IoExitInfo.n.u16Port, pDis->fPrefix,
4637	(DISCPUMODE)pDis->uAddrMode, cbValue);
4638	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOStringRead);
4639	}
4640	}
4641	else
4642	rcStrict = VINF_EM_RAW_EMULATE_INSTR;
4643	}
4644	else
4645	{
4646	/* IN/OUT - I/O instruction. */
4647	Assert(!IoExitInfo.n.u1REP);
4648
4649	if (IoExitInfo.n.u1Type == SVM_IOIO_WRITE)
4650	{
4651	rcStrict = IOMIOPortWrite(pVM, pVCpu, IoExitInfo.n.u16Port, pCtx->eax & uAndVal, cbValue);
4652	if (rcStrict == VINF_IOM_R3_IOPORT_WRITE)
4653	HMR0SavePendingIOPortWrite(pVCpu, pCtx->rip, pVmcb->ctrl.u64ExitInfo2, IoExitInfo.n.u16Port, uAndVal, cbValue);
4654
4655	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOWrite);
4656	}
4657	else
4658	{
4659	uint32_t u32Val = 0;
4660
4661	rcStrict = IOMIOPortRead(pVM, pVCpu, IoExitInfo.n.u16Port, &u32Val, cbValue);
4662	if (IOM_SUCCESS(rcStrict))
4663	{
4664	/* Save result of I/O IN instr. in AL/AX/EAX. */
4665	pCtx->eax = (pCtx->eax & ~uAndVal) \| (u32Val & uAndVal);
4666	}
4667	else if (rcStrict == VINF_IOM_R3_IOPORT_READ)
4668	HMR0SavePendingIOPortRead(pVCpu, pCtx->rip, pVmcb->ctrl.u64ExitInfo2, IoExitInfo.n.u16Port, uAndVal, cbValue);
4669
4670	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIORead);
4671	}
4672	}
4673
4674	if (IOM_SUCCESS(rcStrict))
4675	{
4676	/* AMD-V saves the RIP of the instruction following the IO instruction in EXITINFO2. */
4677	pCtx->rip = pVmcb->ctrl.u64ExitInfo2;
4678
4679	/*
4680	* If any I/O breakpoints are armed, we need to check if one triggered
4681	* and take appropriate action.
4682	* Note that the I/O breakpoint type is undefined if CR4.DE is 0.
4683	*/
4684	/** @todo Optimize away the DBGFBpIsHwIoArmed call by having DBGF tell the
4685	* execution engines about whether hyper BPs and such are pending. */
4686	uint32_t const uDr7 = pCtx->dr[7];
4687	if (RT_UNLIKELY( ( (uDr7 & X86_DR7_ENABLED_MASK)
4688	&& X86_DR7_ANY_RW_IO(uDr7)
4689	&& (pCtx->cr4 & X86_CR4_DE))
4690	\|\| DBGFBpIsHwIoArmed(pVM)))
4691	{
4692	/* We're playing with the host CPU state here, make sure we don't preempt or longjmp. */
4693	VMMRZCallRing3Disable(pVCpu);
4694	HM_DISABLE_PREEMPT_IF_NEEDED();
4695
4696	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxIoCheck);
4697	CPUMR0DebugStateMaybeSaveGuest(pVCpu, false /fDr6/);
4698
4699	VBOXSTRICTRC rcStrict2 = DBGFBpCheckIo(pVM, pVCpu, pCtx, IoExitInfo.n.u16Port, cbValue);
4700	if (rcStrict2 == VINF_EM_RAW_GUEST_TRAP)
4701	{
4702	/* Raise #DB. */
4703	pVmcb->guest.u64DR6 = pCtx->dr[6];
4704	pVmcb->guest.u64DR7 = pCtx->dr[7];
4705	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
4706	hmR0SvmSetPendingXcptDB(pVCpu);
4707	}
4708	/* rcStrict is VINF_SUCCESS or in [VINF_EM_FIRST..VINF_EM_LAST]. */
4709	else if ( rcStrict2 != VINF_SUCCESS
4710	&& (rcStrict == VINF_SUCCESS \|\| rcStrict2 < rcStrict))
4711	rcStrict = rcStrict2;
4712
4713	HM_RESTORE_PREEMPT_IF_NEEDED();
4714	VMMRZCallRing3Enable(pVCpu);
4715	}
4716
4717	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
4718	}
4719
4720	#ifdef VBOX_STRICT
4721	if (rcStrict == VINF_IOM_R3_IOPORT_READ)
4722	Assert(IoExitInfo.n.u1Type == SVM_IOIO_READ);
4723	else if (rcStrict == VINF_IOM_R3_IOPORT_WRITE)
4724	Assert(IoExitInfo.n.u1Type == SVM_IOIO_WRITE);
4725	else
4726	{
4727	/** @todo r=bird: This is missing a bunch of VINF_EM_FIRST..VINF_EM_LAST
4728	* statuses, that the VMM device and some others may return. See
4729	* IOM_SUCCESS() for guidance. */
4730	AssertMsg( RT_FAILURE(rcStrict)
4731	\|\| rcStrict == VINF_SUCCESS
4732	\|\| rcStrict == VINF_EM_RAW_EMULATE_INSTR
4733	\|\| rcStrict == VINF_EM_DBG_BREAKPOINT
4734	\|\| rcStrict == VINF_EM_RAW_GUEST_TRAP
4735	\|\| rcStrict == VINF_TRPM_XCPT_DISPATCHED, ("%Rrc\n", VBOXSTRICTRC_VAL(rcStrict)));
4736	}
4737	#endif
4738	return VBOXSTRICTRC_TODO(rcStrict);
4739	}
4740
4741
4742	/**
4743	* #VMEXIT handler for Nested Page-faults (SVM_EXIT_NPF). Conditional
4744	* #VMEXIT.
4745	*/
4746	HMSVM_EXIT_DECL hmR0SvmExitNestedPF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4747	{
4748	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4749	PVM pVM = pVCpu->CTX_SUFF(pVM);
4750	Assert(pVM->hm.s.fNestedPaging);
4751
4752	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
4753
4754	/* See AMD spec. 15.25.6 "Nested versus Guest Page Faults, Fault Ordering" for VMCB details for #NPF. */
4755	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4756	uint32_t u32ErrCode = pVmcb->ctrl.u64ExitInfo1;
4757	RTGCPHYS GCPhysFaultAddr = pVmcb->ctrl.u64ExitInfo2;
4758
4759	Log4(("#NPF at CS:RIP=%04x:%#RX64 faultaddr=%RGp errcode=%#x \n", pCtx->cs.Sel, pCtx->rip, GCPhysFaultAddr, u32ErrCode));
4760
4761	#ifdef VBOX_HM_WITH_GUEST_PATCHING
4762	/* TPR patching for 32-bit guests, using the reserved bit in the page tables for MMIO regions. */
4763	if ( pVM->hm.s.fTprPatchingAllowed
4764	&& (GCPhysFaultAddr & PAGE_OFFSET_MASK) == 0x80 /* TPR offset. */
4765	&& ( !(u32ErrCode & X86_TRAP_PF_P) /* Not present */
4766	\|\| (u32ErrCode & (X86_TRAP_PF_P \| X86_TRAP_PF_RSVD)) == (X86_TRAP_PF_P \| X86_TRAP_PF_RSVD)) /* MMIO page. */
4767	&& !CPUMIsGuestInLongModeEx(pCtx)
4768	&& !CPUMGetGuestCPL(pVCpu)
4769	&& pVM->hm.s.cPatches < RT_ELEMENTS(pVM->hm.s.aPatches))
4770	{
4771	RTGCPHYS GCPhysApicBase = pCtx->msrApicBase;
4772	GCPhysApicBase &= PAGE_BASE_GC_MASK;
4773
4774	if (GCPhysFaultAddr == GCPhysApicBase + 0x80)
4775	{
4776	/* Only attempt to patch the instruction once. */
4777	PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
4778	if (!pPatch)
4779	return VINF_EM_HM_PATCH_TPR_INSTR;
4780	}
4781	}
4782	#endif
4783
4784	/*
4785	* Determine the nested paging mode.
4786	*/
4787	PGMMODE enmNestedPagingMode;
4788	#if HC_ARCH_BITS == 32
4789	if (CPUMIsGuestInLongModeEx(pCtx))
4790	enmNestedPagingMode = PGMMODE_AMD64_NX;
4791	else
4792	#endif
4793	enmNestedPagingMode = PGMGetHostMode(pVM);
4794
4795	/*
4796	* MMIO optimization using the reserved (RSVD) bit in the guest page tables for MMIO pages.
4797	*/
4798	int rc;
4799	Assert((u32ErrCode & (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P)) != X86_TRAP_PF_RSVD);
4800	if ((u32ErrCode & (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P)) == (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P))
4801	{
4802	VBOXSTRICTRC rc2 = PGMR0Trap0eHandlerNPMisconfig(pVM, pVCpu, enmNestedPagingMode, CPUMCTX2CORE(pCtx), GCPhysFaultAddr,
4803	u32ErrCode);
4804	rc = VBOXSTRICTRC_VAL(rc2);
4805
4806	/*
4807	* If we succeed, resume guest execution.
4808	* If we fail in interpreting the instruction because we couldn't get the guest physical address
4809	* of the page containing the instruction via the guest's page tables (we would invalidate the guest page
4810	* in the host TLB), resume execution which would cause a guest page fault to let the guest handle this
4811	* weird case. See @bugref{6043}.
4812	*/
4813	if ( rc == VINF_SUCCESS
4814	\|\| rc == VERR_PAGE_TABLE_NOT_PRESENT
4815	\|\| rc == VERR_PAGE_NOT_PRESENT)
4816	{
4817	/* Successfully handled MMIO operation. */
4818	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
4819	rc = VINF_SUCCESS;
4820	}
4821	return rc;
4822	}
4823
4824	TRPMAssertXcptPF(pVCpu, GCPhysFaultAddr, u32ErrCode);
4825	rc = PGMR0Trap0eHandlerNestedPaging(pVM, pVCpu, enmNestedPagingMode, u32ErrCode, CPUMCTX2CORE(pCtx), GCPhysFaultAddr);
4826	TRPMResetTrap(pVCpu);
4827
4828	Log4(("#NPF: PGMR0Trap0eHandlerNestedPaging returned %Rrc CS:RIP=%04x:%#RX64\n", rc, pCtx->cs.Sel, pCtx->rip));
4829
4830	/*
4831	* Same case as PGMR0Trap0eHandlerNPMisconfig(). See comment above, @bugref{6043}.
4832	*/
4833	if ( rc == VINF_SUCCESS
4834	\|\| rc == VERR_PAGE_TABLE_NOT_PRESENT
4835	\|\| rc == VERR_PAGE_NOT_PRESENT)
4836	{
4837	/* We've successfully synced our shadow page tables. */
4838	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPF);
4839	rc = VINF_SUCCESS;
4840	}
4841
4842	return rc;
4843	}
4844
4845
4846	/**
4847	* #VMEXIT handler for virtual interrupt (SVM_EXIT_VINTR). Conditional #VMEXIT.
4848	*/
4849	HMSVM_EXIT_DECL hmR0SvmExitVIntr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4850	{
4851	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4852
4853	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4854	pVmcb->ctrl.IntCtrl.n.u1VIrqValid = 0; /* No virtual interrupts pending, we'll inject the current one before reentry. */
4855	pVmcb->ctrl.IntCtrl.n.u8VIrqVector = 0;
4856
4857	/* Indicate that we no longer need to VM-exit when the guest is ready to receive interrupts, it is now ready. */
4858	pVmcb->ctrl.u32InterceptCtrl1 &= ~SVM_CTRL1_INTERCEPT_VINTR;
4859	pVmcb->ctrl.u64VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS \| HMSVM_VMCB_CLEAN_TPR);
4860
4861	/* Deliver the pending interrupt via hmR0SvmPreRunGuest()->hmR0SvmInjectEventVmcb() and resume guest execution. */
4862	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIntWindow);
4863	return VINF_SUCCESS;
4864	}
4865
4866
4867	/**
4868	* #VMEXIT handler for task switches (SVM_EXIT_TASK_SWITCH). Conditional #VMEXIT.
4869	*/
4870	HMSVM_EXIT_DECL hmR0SvmExitTaskSwitch(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4871	{
4872	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4873
4874	#ifndef HMSVM_ALWAYS_TRAP_TASK_SWITCH
4875	Assert(!pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
4876	#endif
4877
4878	/* Check if this task-switch occurred while delivery an event through the guest IDT. */
4879	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4880	if ( !(pVmcb->ctrl.u64ExitInfo2 & (SVM_EXIT2_TASK_SWITCH_IRET \| SVM_EXIT2_TASK_SWITCH_JMP))
4881	&& pVCpu->hm.s.Event.fPending) /** @todo fPending cannot be 'true', see hmR0SvmInjectPendingEvent(). See @bugref{7362}.*/
4882	{
4883	/*
4884	* AMD-V does not provide us with the original exception but we have it in u64IntInfo since we
4885	* injected the event during VM-entry.
4886	*/
4887	Log4(("hmR0SvmExitTaskSwitch: TS occurred during event delivery.\n"));
4888	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitTaskSwitch);
4889	return VINF_EM_RAW_INJECT_TRPM_EVENT;
4890	}
4891
4892	/** @todo Emulate task switch someday, currently just going back to ring-3 for
4893	* emulation. */
4894	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitTaskSwitch);
4895	return VERR_EM_INTERPRETER;
4896	}
4897
4898
4899	/**
4900	* #VMEXIT handler for VMMCALL (SVM_EXIT_VMMCALL). Conditional #VMEXIT.
4901	*/
4902	HMSVM_EXIT_DECL hmR0SvmExitVmmCall(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4903	{
4904	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4905
4906	/* First check if this is a patched VMMCALL for mov TPR */
4907	int rc = hmR0SvmEmulateMovTpr(pVCpu->CTX_SUFF(pVM), pVCpu, pCtx);
4908	if (rc == VINF_SUCCESS)
4909	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4910	else if (rc == VERR_NOT_FOUND)
4911	{
4912	/* Handle GIM provider hypercalls. */
4913	rc = VERR_NOT_SUPPORTED;
4914	if (GIMAreHypercallsEnabled(pVCpu))
4915	rc = GIMHypercall(pVCpu, pCtx);
4916	}
4917
4918	if (rc != VINF_SUCCESS)
4919	hmR0SvmSetPendingXcptUD(pVCpu);
4920	return VINF_SUCCESS;
4921	}
4922
4923
4924	/**
4925	* #VMEXIT handler for page-fault exceptions (SVM_EXIT_EXCEPTION_E). Conditional
4926	* #VMEXIT.
4927	*/
4928	HMSVM_EXIT_DECL hmR0SvmExitXcptPF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4929	{
4930	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4931
4932	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
4933
4934	/* See AMD spec. 15.12.15 "#PF (Page Fault)". */
4935	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4936	uint32_t u32ErrCode = pVmcb->ctrl.u64ExitInfo1;
4937	RTGCUINTPTR uFaultAddress = pVmcb->ctrl.u64ExitInfo2;
4938	PVM pVM = pVCpu->CTX_SUFF(pVM);
4939
4940	#if defined(HMSVM_ALWAYS_TRAP_ALL_XCPTS) \|\| defined(HMSVM_ALWAYS_TRAP_PF)
4941	if (pVM->hm.s.fNestedPaging)
4942	{
4943	pVCpu->hm.s.Event.fPending = false; /* In case it's a contributory or vectoring #PF. */
4944	if (!pSvmTransient->fVectoringPF)
4945	{
4946	/* A genuine guest #PF, reflect it to the guest. */
4947	hmR0SvmSetPendingXcptPF(pVCpu, pCtx, u32ErrCode, uFaultAddress);
4948	Log4(("#PF: Guest page fault at %04X:%RGv FaultAddr=%RGv ErrCode=%#x\n", pCtx->cs.Sel, (RTGCPTR)pCtx->rip,
4949	uFaultAddress, u32ErrCode));
4950	}
4951	else
4952	{
4953	/* A guest page-fault occurred during delivery of a page-fault. Inject #DF. */
4954	hmR0SvmSetPendingXcptDF(pVCpu);
4955	Log4(("Pending #DF due to vectoring #PF. NP\n"));
4956	}
4957	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestPF);
4958	return VINF_SUCCESS;
4959	}
4960	#endif
4961
4962	Assert(!pVM->hm.s.fNestedPaging);
4963
4964	#ifdef VBOX_HM_WITH_GUEST_PATCHING
4965	/* Shortcut for APIC TPR reads and writes; only applicable to 32-bit guests. */
4966	if ( pVM->hm.s.fTprPatchingAllowed
4967	&& (uFaultAddress & 0xfff) == 0x80 /* TPR offset. */
4968	&& !(u32ErrCode & X86_TRAP_PF_P) /* Not present. */
4969	&& !CPUMIsGuestInLongModeEx(pCtx)
4970	&& !CPUMGetGuestCPL(pVCpu)
4971	&& pVM->hm.s.cPatches < RT_ELEMENTS(pVM->hm.s.aPatches))
4972	{
4973	RTGCPHYS GCPhysApicBase;
4974	GCPhysApicBase = pCtx->msrApicBase;
4975	GCPhysApicBase &= PAGE_BASE_GC_MASK;
4976
4977	/* Check if the page at the fault-address is the APIC base. */
4978	RTGCPHYS GCPhysPage;
4979	int rc2 = PGMGstGetPage(pVCpu, (RTGCPTR)uFaultAddress, NULL /* pfFlags */, &GCPhysPage);
4980	if ( rc2 == VINF_SUCCESS
4981	&& GCPhysPage == GCPhysApicBase)
4982	{
4983	/* Only attempt to patch the instruction once. */
4984	PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
4985	if (!pPatch)
4986	return VINF_EM_HM_PATCH_TPR_INSTR;
4987	}
4988	}
4989	#endif
4990
4991	Log4(("#PF: uFaultAddress=%#RX64 CS:RIP=%#04x:%#RX64 u32ErrCode %#RX32 cr3=%#RX64\n", uFaultAddress, pCtx->cs.Sel,
4992	pCtx->rip, u32ErrCode, pCtx->cr3));
4993
4994	TRPMAssertXcptPF(pVCpu, uFaultAddress, u32ErrCode);
4995	int rc = PGMTrap0eHandler(pVCpu, u32ErrCode, CPUMCTX2CORE(pCtx), (RTGCPTR)uFaultAddress);
4996
4997	Log4(("#PF rc=%Rrc\n", rc));
4998
4999	if (rc == VINF_SUCCESS)
5000	{
5001	/* Successfully synced shadow pages tables or emulated an MMIO instruction. */
5002	TRPMResetTrap(pVCpu);
5003	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPF);
5004	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
5005	return rc;
5006	}
5007	else if (rc == VINF_EM_RAW_GUEST_TRAP)
5008	{
5009	pVCpu->hm.s.Event.fPending = false; /* In case it's a contributory or vectoring #PF. */
5010
5011	if (!pSvmTransient->fVectoringPF)
5012	{
5013	/* It's a guest page fault and needs to be reflected to the guest. */
5014	u32ErrCode = TRPMGetErrorCode(pVCpu); /* The error code might have been changed. */
5015	TRPMResetTrap(pVCpu);
5016	hmR0SvmSetPendingXcptPF(pVCpu, pCtx, u32ErrCode, uFaultAddress);
5017	}
5018	else
5019	{
5020	/* A guest page-fault occurred during delivery of a page-fault. Inject #DF. */
5021	TRPMResetTrap(pVCpu);
5022	hmR0SvmSetPendingXcptDF(pVCpu);
5023	Log4(("#PF: Pending #DF due to vectoring #PF\n"));
5024	}
5025
5026	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestPF);
5027	return VINF_SUCCESS;
5028	}
5029
5030	TRPMResetTrap(pVCpu);
5031	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPFEM);
5032	return rc;
5033	}
5034
5035
5036	/**
5037	* #VMEXIT handler for device-not-available exceptions (SVM_EXIT_EXCEPTION_7).
5038	* Conditional #VMEXIT.
5039	*/
5040	HMSVM_EXIT_DECL hmR0SvmExitXcptNM(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5041	{
5042	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5043
5044	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
5045
5046	/* We're playing with the host CPU state here, make sure we don't preempt or longjmp. */
5047	VMMRZCallRing3Disable(pVCpu);
5048	HM_DISABLE_PREEMPT_IF_NEEDED();
5049
5050	int rc;
5051	/* If the guest FPU was active at the time of the #NM exit, then it's a guest fault. */
5052	if (pSvmTransient->fWasGuestFPUStateActive)
5053	{
5054	rc = VINF_EM_RAW_GUEST_TRAP;
5055	Assert(CPUMIsGuestFPUStateActive(pVCpu) \|\| HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR0));
5056	}
5057	else
5058	{
5059	#ifndef HMSVM_ALWAYS_TRAP_ALL_XCPTS
5060	Assert(!pSvmTransient->fWasGuestFPUStateActive);
5061	#endif
5062	rc = CPUMR0Trap07Handler(pVCpu->CTX_SUFF(pVM), pVCpu, pCtx);
5063	Assert(rc == VINF_EM_RAW_GUEST_TRAP \|\| (rc == VINF_SUCCESS && CPUMIsGuestFPUStateActive(pVCpu)));
5064	}
5065
5066	HM_RESTORE_PREEMPT_IF_NEEDED();
5067	VMMRZCallRing3Enable(pVCpu);
5068
5069	if (rc == VINF_SUCCESS)
5070	{
5071	/* Guest FPU state was activated, we'll want to change CR0 FPU intercepts before the next VM-reentry. */
5072	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR0);
5073	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowNM);
5074	pVCpu->hm.s.fUseGuestFpu = true;
5075	}
5076	else
5077	{
5078	/* Forward #NM to the guest. */
5079	Assert(rc == VINF_EM_RAW_GUEST_TRAP);
5080	hmR0SvmSetPendingXcptNM(pVCpu);
5081	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestNM);
5082	}
5083	return VINF_SUCCESS;
5084	}
5085
5086
5087	/**
5088	* #VMEXIT handler for math-fault exceptions (SVM_EXIT_EXCEPTION_10).
5089	* Conditional #VMEXIT.
5090	*/
5091	HMSVM_EXIT_DECL hmR0SvmExitXcptMF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5092	{
5093	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5094
5095	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
5096
5097	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestMF);
5098
5099	if (!(pCtx->cr0 & X86_CR0_NE))
5100	{
5101	PVM pVM = pVCpu->CTX_SUFF(pVM);
5102	PDISSTATE pDis = &pVCpu->hm.s.DisState;
5103	unsigned cbOp;
5104	int rc = EMInterpretDisasCurrent(pVM, pVCpu, pDis, &cbOp);
5105	if (RT_SUCCESS(rc))
5106	{
5107	/* Convert a #MF into a FERR -> IRQ 13. See @bugref{6117}. */
5108	rc = PDMIsaSetIrq(pVCpu->CTX_SUFF(pVM), 13, 1, 0 /* uTagSrc */);
5109	if (RT_SUCCESS(rc))
5110	pCtx->rip += cbOp;
5111	}
5112	else
5113	Log4(("hmR0SvmExitXcptMF: EMInterpretDisasCurrent returned %Rrc uOpCode=%#x\n", rc, pDis->pCurInstr->uOpcode));
5114	return rc;
5115	}
5116
5117	hmR0SvmSetPendingXcptMF(pVCpu);
5118	return VINF_SUCCESS;
5119	}
5120
5121
5122	/**
5123	* #VMEXIT handler for debug exceptions (SVM_EXIT_EXCEPTION_1). Conditional
5124	* #VMEXIT.
5125	*/
5126	HMSVM_EXIT_DECL hmR0SvmExitXcptDB(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5127	{
5128	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5129
5130	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
5131
5132	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestDB);
5133
5134	/* This can be a fault-type #DB (instruction breakpoint) or a trap-type #DB (data breakpoint). However, for both cases
5135	DR6 and DR7 are updated to what the exception handler expects. See AMD spec. 15.12.2 "#DB (Debug)". */
5136	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
5137	PVM pVM = pVCpu->CTX_SUFF(pVM);
5138	int rc = DBGFRZTrap01Handler(pVM, pVCpu, CPUMCTX2CORE(pCtx), pVmcb->guest.u64DR6, pVCpu->hm.s.fSingleInstruction);
5139	if (rc == VINF_EM_RAW_GUEST_TRAP)
5140	{
5141	Log5(("hmR0SvmExitXcptDB: DR6=%#RX64 -> guest trap\n", pVmcb->guest.u64DR6));
5142	if (CPUMIsHyperDebugStateActive(pVCpu))
5143	CPUMSetGuestDR6(pVCpu, CPUMGetGuestDR6(pVCpu) \| pVmcb->guest.u64DR6);
5144
5145	/* Reflect the exception back to the guest. */
5146	hmR0SvmSetPendingXcptDB(pVCpu);
5147	rc = VINF_SUCCESS;
5148	}
5149
5150	/*
5151	* Update DR6.
5152	*/
5153	if (CPUMIsHyperDebugStateActive(pVCpu))
5154	{
5155	Log5(("hmR0SvmExitXcptDB: DR6=%#RX64 -> %Rrc\n", pVmcb->guest.u64DR6, rc));
5156	pVmcb->guest.u64DR6 = X86_DR6_INIT_VAL;
5157	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
5158	}
5159	else
5160	{
5161	AssertMsg(rc == VINF_SUCCESS, ("rc=%Rrc\n", rc));
5162	Assert(!pVCpu->hm.s.fSingleInstruction && !DBGFIsStepping(pVCpu));
5163	}
5164
5165	return rc;
5166	}
5167
5168	/** @} */
5169

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

source: vbox/trunk/src/VBox/VMM/VMMR0/HMSVMR0.cpp@ 52023

Download in other formats: