HMSVMR0.cpp@ 48174

Last change on this file since 48174 was 48174, checked in by vboxsync, 11 years ago
HM: Adjusted some CPUMIsXxxDebugStateActive() assertion for 64-on-32-bit conditions.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 181.5 KB

Line
1	/* $Id: HMSVMR0.cpp 48174 2013-08-30 01:55:50Z vboxsync $ */
2	/** @file
3	* HM SVM (AMD-V) - Host Context Ring-0.
4	*/
5
6	/*
7	* Copyright (C) 2013 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
33	#ifdef DEBUG_ramshankar
34	# define HMSVM_SYNC_FULL_GUEST_STATE
35	# define HMSVM_ALWAYS_TRAP_ALL_XCPTS
36	# define HMSVM_ALWAYS_TRAP_PF
37	# define HMSVM_ALWAYS_TRAP_TASK_SWITCH
38	#endif
39
40
41	/*******************************************************************************
42	* Defined Constants And Macros *
43	*******************************************************************************/
44	#ifdef VBOX_WITH_STATISTICS
45	# define HMSVM_EXITCODE_STAM_COUNTER_INC(u64ExitCode) do { \
46	if ((u64ExitCode) == SVM_EXIT_NPF) \
47	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitReasonNpf); \
48	else \
49	STAM_COUNTER_INC(&pVCpu->hm.s.paStatExitReasonR0[(u64ExitCode) & MASK_EXITREASON_STAT]); \
50	} while (0)
51	#else
52	# define HMSVM_EXITCODE_STAM_COUNTER_INC(u64ExitCode) do { } while (0)
53	#endif
54
55	/** If we decide to use a function table approach this can be useful to
56	* switch to a "static DECLCALLBACK(int)". */
57	#define HMSVM_EXIT_DECL static int
58
59
60	/** @name Segment attribute conversion between CPU and AMD-V VMCB format.
61	*
62	* The CPU format of the segment attribute is described in X86DESCATTRBITS
63	* which is 16-bits (i.e. includes 4 bits of the segment limit).
64	*
65	* The AMD-V VMCB format the segment attribute is compact 12-bits (strictly
66	* only the attribute bits and nothing else). Upper 4-bits are unused.
67	*
68	* @{ */
69	#define HMSVM_CPU_2_VMCB_SEG_ATTR(a) ( ((a) & 0xff) \| (((a) & 0xf000) >> 4) )
70	#define HMSVM_VMCB_2_CPU_SEG_ATTR(a) ( ((a) & 0xff) \| (((a) & 0x0f00) << 4) )
71	/** @} */
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
100	/** Macro for checking and returning from the using function for
101	* \#VMEXIT intercepts that maybe caused during delivering of another
102	* event in the guest. */
103	#define HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY() \
104	do \
105	{ \
106	int rc = hmR0SvmCheckExitDueToEventDelivery(pVCpu, pCtx, pSvmTransient); \
107	if (RT_UNLIKELY(rc == VINF_HM_DOUBLE_FAULT)) \
108	return VINF_SUCCESS; \
109	else if (RT_UNLIKELY(rc == VINF_EM_RESET)) \
110	return rc; \
111	} while (0)
112
113	/** Macro for upgrading a @a a_rc to VINF_EM_DBG_STEPPED after emulating an
114	* instruction that exited. */
115	#define HMSVM_CHECK_SINGLE_STEP(a_pVCpu, a_rc) \
116	do { \
117	if ((a_pVCpu)->hm.s.fSingleInstruction && (a_rc) == VINF_SUCCESS) \
118	(a_rc) = VINF_EM_DBG_STEPPED; \
119	} while (0)
120
121	/** Assert that preemption is disabled or covered by thread-context hooks. */
122	#define HMSVM_ASSERT_PREEMPT_SAFE() Assert( VMMR0ThreadCtxHooksAreRegistered(pVCpu) \
123	\|\| !RTThreadPreemptIsEnabled(NIL_RTTHREAD));
124
125	/** Assert that we haven't migrated CPUs when thread-context hooks are not
126	* used. */
127	#define HMSVM_ASSERT_CPU_SAFE() AssertMsg( VMMR0ThreadCtxHooksAreRegistered(pVCpu) \
128	\|\| pVCpu->hm.s.idEnteredCpu == RTMpCpuId(), \
129	("Illegal migration! Entered on CPU %u Current %u\n", \
130	pVCpu->hm.s.idEnteredCpu, RTMpCpuId()));
131
132	/** Exception bitmap mask for all contributory exceptions.
133	*
134	* Page fault is deliberately excluded here as it's conditional as to whether
135	* it's contributory or benign. Page faults are handled separately.
136	*/
137	#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) \
138	\| RT_BIT(X86_XCPT_DE))
139
140
141	/** @name VMCB Clean Bits.
142	*
143	* These flags are used for VMCB-state caching. A set VMCB Clean Bit indicates
144	* AMD-V doesn't need to reload the corresponding value(s) from the VMCB in
145	* memory.
146	*
147	* @{ */
148	/** All intercepts vectors, TSC offset, PAUSE filter counter. */
149	#define HMSVM_VMCB_CLEAN_INTERCEPTS RT_BIT(0)
150	/** I/O permission bitmap, MSR permission bitmap. */
151	#define HMSVM_VMCB_CLEAN_IOPM_MSRPM RT_BIT(1)
152	/** ASID. */
153	#define HMSVM_VMCB_CLEAN_ASID RT_BIT(2)
154	/** TRP: V_TPR, V_IRQ, V_INTR_PRIO, V_IGN_TPR, V_INTR_MASKING,
155	V_INTR_VECTOR. */
156	#define HMSVM_VMCB_CLEAN_TPR RT_BIT(3)
157	/** Nested Paging: Nested CR3 (nCR3), PAT. */
158	#define HMSVM_VMCB_CLEAN_NP RT_BIT(4)
159	/** Control registers (CR0, CR3, CR4, EFER). */
160	#define HMSVM_VMCB_CLEAN_CRX_EFER RT_BIT(5)
161	/** Debug registers (DR6, DR7). */
162	#define HMSVM_VMCB_CLEAN_DRX RT_BIT(6)
163	/** GDT, IDT limit and base. */
164	#define HMSVM_VMCB_CLEAN_DT RT_BIT(7)
165	/** Segment register: CS, SS, DS, ES limit and base. */
166	#define HMSVM_VMCB_CLEAN_SEG RT_BIT(8)
167	/** CR2.*/
168	#define HMSVM_VMCB_CLEAN_CR2 RT_BIT(9)
169	/** Last-branch record (DbgCtlMsr, br_from, br_to, lastint_from, lastint_to) */
170	#define HMSVM_VMCB_CLEAN_LBR RT_BIT(10)
171	/** AVIC (AVIC APIC_BAR; AVIC APIC_BACKING_PAGE, AVIC
172	PHYSICAL_TABLE and AVIC LOGICAL_TABLE Pointers). */
173	#define HMSVM_VMCB_CLEAN_AVIC RT_BIT(11)
174	/** Mask of all valid VMCB Clean bits. */
175	#define HMSVM_VMCB_CLEAN_ALL ( HMSVM_VMCB_CLEAN_INTERCEPTS \
176	\| HMSVM_VMCB_CLEAN_IOPM_MSRPM \
177	\| HMSVM_VMCB_CLEAN_ASID \
178	\| HMSVM_VMCB_CLEAN_TPR \
179	\| HMSVM_VMCB_CLEAN_NP \
180	\| HMSVM_VMCB_CLEAN_CRX_EFER \
181	\| HMSVM_VMCB_CLEAN_DRX \
182	\| HMSVM_VMCB_CLEAN_DT \
183	\| HMSVM_VMCB_CLEAN_SEG \
184	\| HMSVM_VMCB_CLEAN_CR2 \
185	\| HMSVM_VMCB_CLEAN_LBR \
186	\| HMSVM_VMCB_CLEAN_AVIC)
187	/** @} */
188
189	/** @name SVM transient.
190	*
191	* A state structure for holding miscellaneous information across AMD-V
192	* VMRUN/#VMEXIT operation, restored after the transition.
193	*
194	* @{ */
195	typedef struct SVMTRANSIENT
196	{
197	/** The host's rflags/eflags. */
198	RTCCUINTREG uEflags;
199	#if HC_ARCH_BITS == 32
200	uint32_t u32Alignment0;
201	#endif
202
203	/** The #VMEXIT exit code (the EXITCODE field in the VMCB). */
204	uint64_t u64ExitCode;
205	/** The guest's TPR value used for TPR shadowing. */
206	uint8_t u8GuestTpr;
207	/** Alignment. */
208	uint8_t abAlignment0[7];
209
210	/** Whether the TSC_AUX MSR needs restoring on #VMEXIT. */
211	bool fRestoreTscAuxMsr;
212	/** Whether the #VMEXIT was caused by a page-fault during delivery of a
213	* contributary exception or a page-fault. */
214	bool fVectoringPF;
215	/** Whether the TSC offset mode needs to be updated. */
216	bool fUpdateTscOffsetting;
217	} SVMTRANSIENT, *PSVMTRANSIENT;
218	AssertCompileMemberAlignment(SVMTRANSIENT, u64ExitCode, sizeof(uint64_t));
219	AssertCompileMemberAlignment(SVMTRANSIENT, fRestoreTscAuxMsr, sizeof(uint64_t));
220	/** @} */
221
222
223	/**
224	* MSRPM (MSR permission bitmap) read permissions (for guest RDMSR).
225	*/
226	typedef enum SVMMSREXITREAD
227	{
228	/** Reading this MSR causes a VM-exit. */
229	SVMMSREXIT_INTERCEPT_READ = 0xb,
230	/** Reading this MSR does not cause a VM-exit. */
231	SVMMSREXIT_PASSTHRU_READ
232	} SVMMSREXITREAD;
233
234	/**
235	* MSRPM (MSR permission bitmap) write permissions (for guest WRMSR).
236	*/
237	typedef enum SVMMSREXITWRITE
238	{
239	/** Writing to this MSR causes a VM-exit. */
240	SVMMSREXIT_INTERCEPT_WRITE = 0xd,
241	/** Writing to this MSR does not cause a VM-exit. */
242	SVMMSREXIT_PASSTHRU_WRITE
243	} SVMMSREXITWRITE;
244
245
246	/*******************************************************************************
247	* Internal Functions *
248	*******************************************************************************/
249	static void hmR0SvmSetMsrPermission(PVMCPU pVCpu, unsigned uMsr, SVMMSREXITREAD enmRead, SVMMSREXITWRITE enmWrite);
250	static void hmR0SvmPendingEventToTrpmTrap(PVMCPU pVCpu);
251	static void hmR0SvmLeave(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx);
252
253	HMSVM_EXIT_DECL hmR0SvmExitIntr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
254	HMSVM_EXIT_DECL hmR0SvmExitWbinvd(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
255	HMSVM_EXIT_DECL hmR0SvmExitInvd(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
256	HMSVM_EXIT_DECL hmR0SvmExitCpuid(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
257	HMSVM_EXIT_DECL hmR0SvmExitRdtsc(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
258	HMSVM_EXIT_DECL hmR0SvmExitRdtscp(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
259	HMSVM_EXIT_DECL hmR0SvmExitRdpmc(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
260	HMSVM_EXIT_DECL hmR0SvmExitInvlpg(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
261	HMSVM_EXIT_DECL hmR0SvmExitHlt(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
262	HMSVM_EXIT_DECL hmR0SvmExitMonitor(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
263	HMSVM_EXIT_DECL hmR0SvmExitMwait(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
264	HMSVM_EXIT_DECL hmR0SvmExitShutdown(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
265	HMSVM_EXIT_DECL hmR0SvmExitReadCRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
266	HMSVM_EXIT_DECL hmR0SvmExitWriteCRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
267	HMSVM_EXIT_DECL hmR0SvmExitSetPendingXcptUD(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
268	HMSVM_EXIT_DECL hmR0SvmExitMsr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
269	HMSVM_EXIT_DECL hmR0SvmExitReadDRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
270	HMSVM_EXIT_DECL hmR0SvmExitWriteDRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
271	HMSVM_EXIT_DECL hmR0SvmExitIOInstr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
272	HMSVM_EXIT_DECL hmR0SvmExitNestedPF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
273	HMSVM_EXIT_DECL hmR0SvmExitVIntr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
274	HMSVM_EXIT_DECL hmR0SvmExitTaskSwitch(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
275	HMSVM_EXIT_DECL hmR0SvmExitVmmCall(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
276	HMSVM_EXIT_DECL hmR0SvmExitXcptPF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
277	HMSVM_EXIT_DECL hmR0SvmExitXcptNM(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
278	HMSVM_EXIT_DECL hmR0SvmExitXcptMF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
279	HMSVM_EXIT_DECL hmR0SvmExitXcptDB(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
280
281	DECLINLINE(int) hmR0SvmHandleExit(PVMCPU pVCpu, PCPUMCTX pMixedCtx, PSVMTRANSIENT pSvmTransient);
282
283
284	/*******************************************************************************
285	* Global Variables *
286	*******************************************************************************/
287	/** Ring-0 memory object for the IO bitmap. */
288	RTR0MEMOBJ g_hMemObjIOBitmap = NIL_RTR0MEMOBJ;
289	/** Physical address of the IO bitmap. */
290	RTHCPHYS g_HCPhysIOBitmap = 0;
291	/** Virtual address of the IO bitmap. */
292	R0PTRTYPE(void *) g_pvIOBitmap = NULL;
293
294
295	/**
296	* Sets up and activates AMD-V on the current CPU.
297	*
298	* @returns VBox status code.
299	* @param pCpu Pointer to the CPU info struct.
300	* @param pVM Pointer to the VM (can be NULL after a resume!).
301	* @param pvCpuPage Pointer to the global CPU page.
302	* @param HCPhysCpuPage Physical address of the global CPU page.
303	*/
304	VMMR0DECL(int) SVMR0EnableCpu(PHMGLOBALCPUINFO pCpu, PVM pVM, void *pvCpuPage, RTHCPHYS HCPhysCpuPage, bool fEnabledByHost)
305	{
306	AssertReturn(!fEnabledByHost, VERR_INVALID_PARAMETER);
307	AssertReturn( HCPhysCpuPage
308	&& HCPhysCpuPage != NIL_RTHCPHYS, VERR_INVALID_PARAMETER);
309	AssertReturn(pvCpuPage, VERR_INVALID_PARAMETER);
310
311	/*
312	* We must turn on AMD-V and setup the host state physical address, as those MSRs are per CPU.
313	*/
314	uint64_t u64HostEfer = ASMRdMsr(MSR_K6_EFER);
315	if (u64HostEfer & MSR_K6_EFER_SVME)
316	{
317	/* If the VBOX_HWVIRTEX_IGNORE_SVM_IN_USE is active, then we blindly use AMD-V. */
318	if ( pVM
319	&& pVM->hm.s.svm.fIgnoreInUseError)
320	{
321	pCpu->fIgnoreAMDVInUseError = true;
322	}
323
324	if (!pCpu->fIgnoreAMDVInUseError)
325	return VERR_SVM_IN_USE;
326	}
327
328	/* Turn on AMD-V in the EFER MSR. */
329	ASMWrMsr(MSR_K6_EFER, u64HostEfer \| MSR_K6_EFER_SVME);
330
331	/* Write the physical page address where the CPU will store the host state while executing the VM. */
332	ASMWrMsr(MSR_K8_VM_HSAVE_PA, HCPhysCpuPage);
333
334	/*
335	* Theoretically, other hypervisors may have used ASIDs, ideally we should flush all non-zero ASIDs
336	* when enabling SVM. AMD doesn't have an SVM instruction to flush all ASIDs (flushing is done
337	* upon VMRUN). Therefore, just set the fFlushAsidBeforeUse flag which instructs hmR0SvmSetupTLB()
338	* to flush the TLB with before using a new ASID.
339	*/
340	pCpu->fFlushAsidBeforeUse = true;
341
342	/*
343	* Ensure each VCPU scheduled on this CPU gets a new VPID on resume. See @bugref{6255}.
344	*/
345	++pCpu->cTlbFlushes;
346
347	return VINF_SUCCESS;
348	}
349
350
351	/**
352	* Deactivates AMD-V on the current CPU.
353	*
354	* @returns VBox status code.
355	* @param pCpu Pointer to the CPU info struct.
356	* @param pvCpuPage Pointer to the global CPU page.
357	* @param HCPhysCpuPage Physical address of the global CPU page.
358	*/
359	VMMR0DECL(int) SVMR0DisableCpu(PHMGLOBALCPUINFO pCpu, void *pvCpuPage, RTHCPHYS HCPhysCpuPage)
360	{
361	AssertReturn( HCPhysCpuPage
362	&& HCPhysCpuPage != NIL_RTHCPHYS, VERR_INVALID_PARAMETER);
363	AssertReturn(pvCpuPage, VERR_INVALID_PARAMETER);
364	NOREF(pCpu);
365
366	/* Turn off AMD-V in the EFER MSR. */
367	uint64_t u64HostEfer = ASMRdMsr(MSR_K6_EFER);
368	ASMWrMsr(MSR_K6_EFER, u64HostEfer & ~MSR_K6_EFER_SVME);
369
370	/* Invalidate host state physical address. */
371	ASMWrMsr(MSR_K8_VM_HSAVE_PA, 0);
372
373	return VINF_SUCCESS;
374	}
375
376
377	/**
378	* Does global AMD-V initialization (called during module initialization).
379	*
380	* @returns VBox status code.
381	*/
382	VMMR0DECL(int) SVMR0GlobalInit(void)
383	{
384	/*
385	* Allocate 12 KB for the IO bitmap. Since this is non-optional and we always intercept all IO accesses, it's done
386	* once globally here instead of per-VM.
387	*/
388	Assert(g_hMemObjIOBitmap == NIL_RTR0MEMOBJ);
389	int rc = RTR0MemObjAllocCont(&g_hMemObjIOBitmap, 3 << PAGE_SHIFT, false /* fExecutable */);
390	if (RT_FAILURE(rc))
391	return rc;
392
393	g_pvIOBitmap = RTR0MemObjAddress(g_hMemObjIOBitmap);
394	g_HCPhysIOBitmap = RTR0MemObjGetPagePhysAddr(g_hMemObjIOBitmap, 0 /* iPage */);
395
396	/* Set all bits to intercept all IO accesses. */
397	ASMMemFill32(g_pvIOBitmap, 3 << PAGE_SHIFT, UINT32_C(0xffffffff));
398	return VINF_SUCCESS;
399	}
400
401
402	/**
403	* Does global AMD-V termination (called during module termination).
404	*/
405	VMMR0DECL(void) SVMR0GlobalTerm(void)
406	{
407	if (g_hMemObjIOBitmap != NIL_RTR0MEMOBJ)
408	{
409	RTR0MemObjFree(g_hMemObjIOBitmap, false /* fFreeMappings */);
410	g_pvIOBitmap = NULL;
411	g_HCPhysIOBitmap = 0;
412	g_hMemObjIOBitmap = NIL_RTR0MEMOBJ;
413	}
414	}
415
416
417	/**
418	* Frees any allocated per-VCPU structures for a VM.
419	*
420	* @param pVM Pointer to the VM.
421	*/
422	DECLINLINE(void) hmR0SvmFreeStructs(PVM pVM)
423	{
424	for (uint32_t i = 0; i < pVM->cCpus; i++)
425	{
426	PVMCPU pVCpu = &pVM->aCpus[i];
427	AssertPtr(pVCpu);
428
429	if (pVCpu->hm.s.svm.hMemObjVmcbHost != NIL_RTR0MEMOBJ)
430	{
431	RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjVmcbHost, false);
432	pVCpu->hm.s.svm.pvVmcbHost = 0;
433	pVCpu->hm.s.svm.HCPhysVmcbHost = 0;
434	pVCpu->hm.s.svm.hMemObjVmcbHost = NIL_RTR0MEMOBJ;
435	}
436
437	if (pVCpu->hm.s.svm.hMemObjVmcb != NIL_RTR0MEMOBJ)
438	{
439	RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjVmcb, false);
440	pVCpu->hm.s.svm.pvVmcb = 0;
441	pVCpu->hm.s.svm.HCPhysVmcb = 0;
442	pVCpu->hm.s.svm.hMemObjVmcb = NIL_RTR0MEMOBJ;
443	}
444
445	if (pVCpu->hm.s.svm.hMemObjMsrBitmap != NIL_RTR0MEMOBJ)
446	{
447	RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjMsrBitmap, false);
448	pVCpu->hm.s.svm.pvMsrBitmap = 0;
449	pVCpu->hm.s.svm.HCPhysMsrBitmap = 0;
450	pVCpu->hm.s.svm.hMemObjMsrBitmap = NIL_RTR0MEMOBJ;
451	}
452	}
453	}
454
455
456	/**
457	* Does per-VM AMD-V initialization.
458	*
459	* @returns VBox status code.
460	* @param pVM Pointer to the VM.
461	*/
462	VMMR0DECL(int) SVMR0InitVM(PVM pVM)
463	{
464	int rc = VERR_INTERNAL_ERROR_5;
465
466	/*
467	* Check for an AMD CPU erratum which requires us to flush the TLB before every world-switch.
468	*/
469	uint32_t u32Family;
470	uint32_t u32Model;
471	uint32_t u32Stepping;
472	if (HMAmdIsSubjectToErratum170(&u32Family, &u32Model, &u32Stepping))
473	{
474	Log4(("SVMR0InitVM: AMD cpu with erratum 170 family %#x model %#x stepping %#x\n", u32Family, u32Model, u32Stepping));
475	pVM->hm.s.svm.fAlwaysFlushTLB = true;
476	}
477
478	/*
479	* Initialize the R0 memory objects up-front so we can properly cleanup on allocation failures.
480	*/
481	for (VMCPUID i = 0; i < pVM->cCpus; i++)
482	{
483	PVMCPU pVCpu = &pVM->aCpus[i];
484	pVCpu->hm.s.svm.hMemObjVmcbHost = NIL_RTR0MEMOBJ;
485	pVCpu->hm.s.svm.hMemObjVmcb = NIL_RTR0MEMOBJ;
486	pVCpu->hm.s.svm.hMemObjMsrBitmap = NIL_RTR0MEMOBJ;
487	}
488
489	for (VMCPUID i = 0; i < pVM->cCpus; i++)
490	{
491	PVMCPU pVCpu = &pVM->aCpus[i];
492
493	/*
494	* Allocate one page for the host-context VM control block (VMCB). This is used for additional host-state (such as
495	* FS, GS, Kernel GS Base, etc.) apart from the host-state save area specified in MSR_K8_VM_HSAVE_PA.
496	*/
497	rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjVmcbHost, 1 << PAGE_SHIFT, false /* fExecutable */);
498	if (RT_FAILURE(rc))
499	goto failure_cleanup;
500
501	pVCpu->hm.s.svm.pvVmcbHost = RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjVmcbHost);
502	pVCpu->hm.s.svm.HCPhysVmcbHost = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjVmcbHost, 0 /* iPage */);
503	Assert(pVCpu->hm.s.svm.HCPhysVmcbHost < _4G);
504	ASMMemZeroPage(pVCpu->hm.s.svm.pvVmcbHost);
505
506	/*
507	* Allocate one page for the guest-state VMCB.
508	*/
509	rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjVmcb, 1 << PAGE_SHIFT, false /* fExecutable */);
510	if (RT_FAILURE(rc))
511	goto failure_cleanup;
512
513	pVCpu->hm.s.svm.pvVmcb = RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjVmcb);
514	pVCpu->hm.s.svm.HCPhysVmcb = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjVmcb, 0 /* iPage */);
515	Assert(pVCpu->hm.s.svm.HCPhysVmcb < _4G);
516	ASMMemZeroPage(pVCpu->hm.s.svm.pvVmcb);
517
518	/*
519	* Allocate two pages (8 KB) for the MSR permission bitmap. There doesn't seem to be a way to convince
520	* SVM to not require one.
521	*/
522	rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjMsrBitmap, 2 << PAGE_SHIFT, false /* fExecutable */);
523	if (RT_FAILURE(rc))
524	goto failure_cleanup;
525
526	pVCpu->hm.s.svm.pvMsrBitmap = RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjMsrBitmap);
527	pVCpu->hm.s.svm.HCPhysMsrBitmap = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjMsrBitmap, 0 /* iPage */);
528	/* Set all bits to intercept all MSR accesses (changed later on). */
529	ASMMemFill32(pVCpu->hm.s.svm.pvMsrBitmap, 2 << PAGE_SHIFT, 0xffffffff);
530	}
531
532	return VINF_SUCCESS;
533
534	failure_cleanup:
535	hmR0SvmFreeStructs(pVM);
536	return rc;
537	}
538
539
540	/**
541	* Does per-VM AMD-V termination.
542	*
543	* @returns VBox status code.
544	* @param pVM Pointer to the VM.
545	*/
546	VMMR0DECL(int) SVMR0TermVM(PVM pVM)
547	{
548	hmR0SvmFreeStructs(pVM);
549	return VINF_SUCCESS;
550	}
551
552
553	/**
554	* Sets the permission bits for the specified MSR in the MSRPM.
555	*
556	* @param pVCpu Pointer to the VMCPU.
557	* @param uMsr The MSR for which the access permissions are being set.
558	* @param enmRead MSR read permissions.
559	* @param enmWrite MSR write permissions.
560	*/
561	static void hmR0SvmSetMsrPermission(PVMCPU pVCpu, unsigned uMsr, SVMMSREXITREAD enmRead, SVMMSREXITWRITE enmWrite)
562	{
563	unsigned ulBit;
564	uint8_t pbMsrBitmap = (uint8_t )pVCpu->hm.s.svm.pvMsrBitmap;
565
566	/*
567	* Layout:
568	* Byte offset MSR range
569	* 0x000 - 0x7ff 0x00000000 - 0x00001fff
570	* 0x800 - 0xfff 0xc0000000 - 0xc0001fff
571	* 0x1000 - 0x17ff 0xc0010000 - 0xc0011fff
572	* 0x1800 - 0x1fff Reserved
573	*/
574	if (uMsr <= 0x00001FFF)
575	{
576	/* Pentium-compatible MSRs. */
577	ulBit = uMsr * 2;
578	}
579	else if ( uMsr >= 0xC0000000
580	&& uMsr <= 0xC0001FFF)
581	{
582	/* AMD Sixth Generation x86 Processor MSRs. */
583	ulBit = (uMsr - 0xC0000000) * 2;
584	pbMsrBitmap += 0x800;
585	}
586	else if ( uMsr >= 0xC0010000
587	&& uMsr <= 0xC0011FFF)
588	{
589	/* AMD Seventh and Eighth Generation Processor MSRs. */
590	ulBit = (uMsr - 0xC0001000) * 2;
591	pbMsrBitmap += 0x1000;
592	}
593	else
594	{
595	AssertFailed();
596	return;
597	}
598
599	Assert(ulBit < 0x3fff /* 16 * 1024 - 1 */);
600	if (enmRead == SVMMSREXIT_INTERCEPT_READ)
601	ASMBitSet(pbMsrBitmap, ulBit);
602	else
603	ASMBitClear(pbMsrBitmap, ulBit);
604
605	if (enmWrite == SVMMSREXIT_INTERCEPT_WRITE)
606	ASMBitSet(pbMsrBitmap, ulBit + 1);
607	else
608	ASMBitClear(pbMsrBitmap, ulBit + 1);
609
610	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
611	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_IOPM_MSRPM;
612	}
613
614
615	/**
616	* Sets up AMD-V for the specified VM.
617	* This function is only called once per-VM during initalization.
618	*
619	* @returns VBox status code.
620	* @param pVM Pointer to the VM.
621	*/
622	VMMR0DECL(int) SVMR0SetupVM(PVM pVM)
623	{
624	int rc = VINF_SUCCESS;
625
626	AssertReturn(pVM, VERR_INVALID_PARAMETER);
627	Assert(pVM->hm.s.svm.fSupported);
628
629	for (VMCPUID i = 0; i < pVM->cCpus; i++)
630	{
631	PVMCPU pVCpu = &pVM->aCpus[i];
632	PSVMVMCB pVmcb = (PSVMVMCB)pVM->aCpus[i].hm.s.svm.pvVmcb;
633
634	AssertMsgReturn(pVmcb, ("Invalid pVmcb\n"), VERR_SVM_INVALID_PVMCB);
635
636	/* Trap exceptions unconditionally (debug purposes). */
637	#ifdef HMSVM_ALWAYS_TRAP_PF
638	pVmcb->ctrl.u32InterceptException \|= RT_BIT(X86_XCPT_PF);
639	#endif
640	#ifdef HMSVM_ALWAYS_TRAP_ALL_XCPTS
641	/* If you add any exceptions here, make sure to update hmR0SvmHandleExit(). */
642	pVmcb->ctrl.u32InterceptException \|= 0
643	\| RT_BIT(X86_XCPT_BP)
644	\| RT_BIT(X86_XCPT_DB)
645	\| RT_BIT(X86_XCPT_DE)
646	\| RT_BIT(X86_XCPT_NM)
647	\| RT_BIT(X86_XCPT_UD)
648	\| RT_BIT(X86_XCPT_NP)
649	\| RT_BIT(X86_XCPT_SS)
650	\| RT_BIT(X86_XCPT_GP)
651	\| RT_BIT(X86_XCPT_PF)
652	\| RT_BIT(X86_XCPT_MF)
653	;
654	#endif
655
656	/* Set up unconditional intercepts and conditions. */
657	pVmcb->ctrl.u32InterceptCtrl1 = SVM_CTRL1_INTERCEPT_INTR /* External interrupt causes a VM-exit. */
658	\| SVM_CTRL1_INTERCEPT_NMI /* Non-Maskable Interrupts causes a VM-exit. */
659	\| SVM_CTRL1_INTERCEPT_INIT /* INIT signal causes a VM-exit. */
660	\| SVM_CTRL1_INTERCEPT_RDPMC /* RDPMC causes a VM-exit. */
661	\| SVM_CTRL1_INTERCEPT_CPUID /* CPUID causes a VM-exit. */
662	\| SVM_CTRL1_INTERCEPT_RSM /* RSM causes a VM-exit. */
663	\| SVM_CTRL1_INTERCEPT_HLT /* HLT causes a VM-exit. */
664	\| SVM_CTRL1_INTERCEPT_INOUT_BITMAP /* Use the IOPM to cause IOIO VM-exits. */
665	\| SVM_CTRL1_INTERCEPT_MSR_SHADOW /* MSR access not covered by MSRPM causes a VM-exit.*/
666	\| SVM_CTRL1_INTERCEPT_INVLPGA /* INVLPGA causes a VM-exit. */
667	\| SVM_CTRL1_INTERCEPT_SHUTDOWN /* Shutdown events causes a VM-exit. */
668	\| SVM_CTRL1_INTERCEPT_FERR_FREEZE; /* Intercept "freezing" during legacy FPU handling. */
669
670	pVmcb->ctrl.u32InterceptCtrl2 = SVM_CTRL2_INTERCEPT_VMRUN /* VMRUN causes a VM-exit. */
671	\| SVM_CTRL2_INTERCEPT_VMMCALL /* VMMCALL causes a VM-exit. */
672	\| SVM_CTRL2_INTERCEPT_VMLOAD /* VMLOAD causes a VM-exit. */
673	\| SVM_CTRL2_INTERCEPT_VMSAVE /* VMSAVE causes a VM-exit. */
674	\| SVM_CTRL2_INTERCEPT_STGI /* STGI causes a VM-exit. */
675	\| SVM_CTRL2_INTERCEPT_CLGI /* CLGI causes a VM-exit. */
676	\| SVM_CTRL2_INTERCEPT_SKINIT /* SKINIT causes a VM-exit. */
677	\| SVM_CTRL2_INTERCEPT_WBINVD /* WBINVD causes a VM-exit. */
678	\| SVM_CTRL2_INTERCEPT_MONITOR /* MONITOR causes a VM-exit. */
679	\| SVM_CTRL2_INTERCEPT_MWAIT; /* MWAIT causes a VM-exit. */
680
681	/* CR0, CR4 reads must be intercepted, our shadow values are not necessarily the same as the guest's. */
682	pVmcb->ctrl.u16InterceptRdCRx = RT_BIT(0) \| RT_BIT(4);
683
684	/* CR0, CR4 writes must be intercepted for the same reasons as above. */
685	pVmcb->ctrl.u16InterceptWrCRx = RT_BIT(0) \| RT_BIT(4);
686
687	/* Intercept all DRx reads and writes by default. Changed later on. */
688	pVmcb->ctrl.u16InterceptRdDRx = 0xffff;
689	pVmcb->ctrl.u16InterceptWrDRx = 0xffff;
690
691	/* Virtualize masking of INTR interrupts. (reads/writes from/to CR8 go to the V_TPR register) */
692	pVmcb->ctrl.IntCtrl.n.u1VIrqMasking = 1;
693
694	/* Ignore the priority in the TPR. This is necessary for delivering PIC style (ExtInt) interrupts and we currently
695	deliver both PIC and APIC interrupts alike. See hmR0SvmInjectPendingEvent() */
696	pVmcb->ctrl.IntCtrl.n.u1IgnoreTPR = 1;
697
698	/* Set IO and MSR bitmap permission bitmap physical addresses. */
699	pVmcb->ctrl.u64IOPMPhysAddr = g_HCPhysIOBitmap;
700	pVmcb->ctrl.u64MSRPMPhysAddr = pVCpu->hm.s.svm.HCPhysMsrBitmap;
701
702	/* No LBR virtualization. */
703	pVmcb->ctrl.u64LBRVirt = 0;
704
705	/* Initially set all VMCB clean bits to 0 indicating that everything should be loaded from the VMCB in memory. */
706	pVmcb->ctrl.u64VmcbCleanBits = 0;
707
708	/* The host ASID MBZ, for the guest start with 1. */
709	pVmcb->ctrl.TLBCtrl.n.u32ASID = 1;
710
711	/*
712	* Setup the PAT MSR (applicable for Nested Paging only).
713	* The default value should be 0x0007040600070406ULL, but we want to treat all guest memory as WB,
714	* so choose type 6 for all PAT slots.
715	*/
716	pVmcb->guest.u64GPAT = UINT64_C(0x0006060606060606);
717
718	/* Without Nested Paging, we need additionally intercepts. */
719	if (!pVM->hm.s.fNestedPaging)
720	{
721	/* CR3 reads/writes must be intercepted; our shadow values differ from the guest values. */
722	pVmcb->ctrl.u16InterceptRdCRx \|= RT_BIT(3);
723	pVmcb->ctrl.u16InterceptWrCRx \|= RT_BIT(3);
724
725	/* Intercept INVLPG and task switches (may change CR3, EFLAGS, LDT). */
726	pVmcb->ctrl.u32InterceptCtrl1 \|= SVM_CTRL1_INTERCEPT_INVLPG
727	\| SVM_CTRL1_INTERCEPT_TASK_SWITCH;
728
729	/* Page faults must be intercepted to implement shadow paging. */
730	pVmcb->ctrl.u32InterceptException \|= RT_BIT(X86_XCPT_PF);
731	}
732
733	#ifdef HMSVM_ALWAYS_TRAP_TASK_SWITCH
734	pVmcb->ctrl.u32InterceptCtrl1 \|= SVM_CTRL1_INTERCEPT_TASK_SWITCH;
735	#endif
736
737	/*
738	* The following MSRs are saved/restored automatically during the world-switch.
739	* Don't intercept guest read/write accesses to these MSRs.
740	*/
741	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
742	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_CSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
743	hmR0SvmSetMsrPermission(pVCpu, MSR_K6_STAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
744	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_SF_MASK, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
745	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_FS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
746	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_GS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
747	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_KERNEL_GS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
748	hmR0SvmSetMsrPermission(pVCpu, MSR_IA32_SYSENTER_CS, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
749	hmR0SvmSetMsrPermission(pVCpu, MSR_IA32_SYSENTER_ESP, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
750	hmR0SvmSetMsrPermission(pVCpu, MSR_IA32_SYSENTER_EIP, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
751	}
752
753	return rc;
754	}
755
756
757	/**
758	* Invalidates a guest page by guest virtual address.
759	*
760	* @returns VBox status code.
761	* @param pVM Pointer to the VM.
762	* @param pVCpu Pointer to the VMCPU.
763	* @param GCVirt Guest virtual address of the page to invalidate.
764	*/
765	VMMR0DECL(int) SVMR0InvalidatePage(PVM pVM, PVMCPU pVCpu, RTGCPTR GCVirt)
766	{
767	AssertReturn(pVM, VERR_INVALID_PARAMETER);
768	Assert(pVM->hm.s.svm.fSupported);
769
770	bool fFlushPending = pVM->hm.s.svm.fAlwaysFlushTLB \|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_TLB_FLUSH);
771
772	/* Skip it if a TLB flush is already pending. */
773	if (!fFlushPending)
774	{
775	Log4(("SVMR0InvalidatePage %RGv\n", GCVirt));
776
777	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
778	AssertMsgReturn(pVmcb, ("Invalid pVmcb!\n"), VERR_SVM_INVALID_PVMCB);
779
780	#if HC_ARCH_BITS == 32
781	/* If we get a flush in 64-bit guest mode, then force a full TLB flush. INVLPGA takes only 32-bit addresses. */
782	if (CPUMIsGuestInLongMode(pVCpu))
783	VMCPU_FF_SET(pVCpu, VMCPU_FF_TLB_FLUSH);
784	else
785	#endif
786	{
787	SVMR0InvlpgA(GCVirt, pVmcb->ctrl.TLBCtrl.n.u32ASID);
788	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlbInvlpgVirt);
789	}
790	}
791	return VINF_SUCCESS;
792	}
793
794
795	/**
796	* Flushes the appropriate tagged-TLB entries.
797	*
798	* @param pVM Pointer to the VM.
799	* @param pVCpu Pointer to the VMCPU.
800	*/
801	static void hmR0SvmFlushTaggedTlb(PVMCPU pVCpu)
802	{
803	PVM pVM = pVCpu->CTX_SUFF(pVM);
804	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
805	PHMGLOBALCPUINFO pCpu = HMR0GetCurrentCpu();
806
807	/*
808	* Force a TLB flush for the first world switch if the current CPU differs from the one we ran on last.
809	* This can happen both for start & resume due to long jumps back to ring-3.
810	* If the TLB flush count changed, another VM (VCPU rather) has hit the ASID limit while flushing the TLB,
811	* so we cannot reuse the ASIDs without flushing.
812	*/
813	bool fNewAsid = false;
814	if ( pVCpu->hm.s.idLastCpu != pCpu->idCpu
815	\|\| pVCpu->hm.s.cTlbFlushes != pCpu->cTlbFlushes)
816	{
817	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlbWorldSwitch);
818	pVCpu->hm.s.fForceTLBFlush = true;
819	fNewAsid = true;
820	}
821
822	/* Set TLB flush state as checked until we return from the world switch. */
823	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, true);
824
825	/* Check for explicit TLB shootdowns. */
826	if (VMCPU_FF_TEST_AND_CLEAR(pVCpu, VMCPU_FF_TLB_FLUSH))
827	{
828	pVCpu->hm.s.fForceTLBFlush = true;
829	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlb);
830	}
831
832	pVCpu->hm.s.idLastCpu = pCpu->idCpu;
833	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_NOTHING;
834
835	if (pVM->hm.s.svm.fAlwaysFlushTLB)
836	{
837	/*
838	* This is the AMD erratum 170. We need to flush the entire TLB for each world switch. Sad.
839	*/
840	pCpu->uCurrentAsid = 1;
841	pVCpu->hm.s.uCurrentAsid = 1;
842	pVCpu->hm.s.cTlbFlushes = pCpu->cTlbFlushes;
843	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
844	}
845	else if (pVCpu->hm.s.fForceTLBFlush)
846	{
847	if (fNewAsid)
848	{
849	++pCpu->uCurrentAsid;
850	bool fHitASIDLimit = false;
851	if (pCpu->uCurrentAsid >= pVM->hm.s.uMaxAsid)
852	{
853	pCpu->uCurrentAsid = 1; /* Wraparound at 1; host uses 0 */
854	pCpu->cTlbFlushes++; /* All VCPUs that run on this host CPU must use a new VPID. */
855	fHitASIDLimit = true;
856
857	if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_FLUSH_BY_ASID)
858	{
859	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_SINGLE_CONTEXT;
860	pCpu->fFlushAsidBeforeUse = true;
861	}
862	else
863	{
864	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
865	pCpu->fFlushAsidBeforeUse = false;
866	}
867	}
868
869	if ( !fHitASIDLimit
870	&& pCpu->fFlushAsidBeforeUse)
871	{
872	if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_FLUSH_BY_ASID)
873	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_SINGLE_CONTEXT;
874	else
875	{
876	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
877	pCpu->fFlushAsidBeforeUse = false;
878	}
879	}
880
881	pVCpu->hm.s.uCurrentAsid = pCpu->uCurrentAsid;
882	pVCpu->hm.s.cTlbFlushes = pCpu->cTlbFlushes;
883	}
884	else
885	{
886	if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_FLUSH_BY_ASID)
887	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_SINGLE_CONTEXT;
888	else
889	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
890	}
891
892	pVCpu->hm.s.fForceTLBFlush = false;
893	}
894	else
895	{
896	/** @todo We never set VMCPU_FF_TLB_SHOOTDOWN anywhere so this path should
897	* not be executed. See hmQueueInvlPage() where it is commented
898	* out. Support individual entry flushing someday. */
899	if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_TLB_SHOOTDOWN))
900	{
901	/* Deal with pending TLB shootdown actions which were queued when we were not executing code. */
902	STAM_COUNTER_INC(&pVCpu->hm.s.StatTlbShootdown);
903	for (uint32_t i = 0; i < pVCpu->hm.s.TlbShootdown.cPages; i++)
904	SVMR0InvlpgA(pVCpu->hm.s.TlbShootdown.aPages[i], pVmcb->ctrl.TLBCtrl.n.u32ASID);
905	}
906	}
907
908	pVCpu->hm.s.TlbShootdown.cPages = 0;
909	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_TLB_SHOOTDOWN);
910
911	/* Update VMCB with the ASID. */
912	if (pVmcb->ctrl.TLBCtrl.n.u32ASID != pVCpu->hm.s.uCurrentAsid)
913	{
914	pVmcb->ctrl.TLBCtrl.n.u32ASID = pVCpu->hm.s.uCurrentAsid;
915	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_ASID;
916	}
917
918	AssertMsg(pVCpu->hm.s.cTlbFlushes == pCpu->cTlbFlushes,
919	("Flush count mismatch for cpu %d (%x vs %x)\n", pCpu->idCpu, pVCpu->hm.s.cTlbFlushes, pCpu->cTlbFlushes));
920	AssertMsg(pCpu->uCurrentAsid >= 1 && pCpu->uCurrentAsid < pVM->hm.s.uMaxAsid,
921	("cpu%d uCurrentAsid = %x\n", pCpu->idCpu, pCpu->uCurrentAsid));
922	AssertMsg(pVCpu->hm.s.uCurrentAsid >= 1 && pVCpu->hm.s.uCurrentAsid < pVM->hm.s.uMaxAsid,
923	("cpu%d VM uCurrentAsid = %x\n", pCpu->idCpu, pVCpu->hm.s.uCurrentAsid));
924
925	#ifdef VBOX_WITH_STATISTICS
926	if (pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_NOTHING)
927	STAM_COUNTER_INC(&pVCpu->hm.s.StatNoFlushTlbWorldSwitch);
928	else if ( pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_SINGLE_CONTEXT
929	\|\| pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_SINGLE_CONTEXT_RETAIN_GLOBALS)
930	{
931	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushAsid);
932	}
933	else
934	{
935	Assert(pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_ENTIRE);
936	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushEntire);
937	}
938	#endif
939	}
940
941
942	/** @name 64-bit guest on 32-bit host OS helper functions.
943	*
944	* The host CPU is still 64-bit capable but the host OS is running in 32-bit
945	* mode (code segment, paging). These wrappers/helpers perform the necessary
946	* bits for the 32->64 switcher.
947	*
948	* @{ */
949	#if HC_ARCH_BITS == 32 && defined(VBOX_ENABLE_64_BITS_GUESTS) && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
950	/**
951	* Prepares for and executes VMRUN (64-bit guests on a 32-bit host).
952	*
953	* @returns VBox status code.
954	* @param HCPhysVmcbHost Physical address of host VMCB.
955	* @param HCPhysVmcb Physical address of the VMCB.
956	* @param pCtx Pointer to the guest-CPU context.
957	* @param pVM Pointer to the VM.
958	* @param pVCpu Pointer to the VMCPU.
959	*/
960	DECLASM(int) SVMR0VMSwitcherRun64(RTHCPHYS HCPhysVmcbHost, RTHCPHYS HCPhysVmcb, PCPUMCTX pCtx, PVM pVM, PVMCPU pVCpu)
961	{
962	uint32_t aParam[4];
963	aParam[0] = (uint32_t)(HCPhysVmcbHost); /* Param 1: HCPhysVmcbHost - Lo. */
964	aParam[1] = (uint32_t)(HCPhysVmcbHost >> 32); /* Param 1: HCPhysVmcbHost - Hi. */
965	aParam[2] = (uint32_t)(HCPhysVmcb); /* Param 2: HCPhysVmcb - Lo. */
966	aParam[3] = (uint32_t)(HCPhysVmcb >> 32); /* Param 2: HCPhysVmcb - Hi. */
967
968	return SVMR0Execute64BitsHandler(pVM, pVCpu, pCtx, HM64ON32OP_SVMRCVMRun64, 4, &aParam[0]);
969	}
970
971
972	/**
973	* Executes the specified VMRUN handler in 64-bit mode.
974	*
975	* @returns VBox status code.
976	* @param pVM Pointer to the VM.
977	* @param pVCpu Pointer to the VMCPU.
978	* @param pCtx Pointer to the guest-CPU context.
979	* @param enmOp The operation to perform.
980	* @param cbParam Number of parameters.
981	* @param paParam Array of 32-bit parameters.
982	*/
983	VMMR0DECL(int) SVMR0Execute64BitsHandler(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, HM64ON32OP enmOp, uint32_t cbParam,
984	uint32_t *paParam)
985	{
986	AssertReturn(pVM->hm.s.pfnHost32ToGuest64R0, VERR_HM_NO_32_TO_64_SWITCHER);
987	Assert(enmOp > HM64ON32OP_INVALID && enmOp < HM64ON32OP_END);
988
989	/* Disable interrupts. */
990	RTHCUINTREG uOldEFlags = ASMIntDisableFlags();
991
992	#ifdef VBOX_WITH_VMMR0_DISABLE_LAPIC_NMI
993	RTCPUID idHostCpu = RTMpCpuId();
994	CPUMR0SetLApic(pVCpu, idHostCpu);
995	#endif
996
997	CPUMSetHyperESP(pVCpu, VMMGetStackRC(pVCpu));
998	CPUMSetHyperEIP(pVCpu, enmOp);
999	for (int i = (int)cbParam - 1; i >= 0; i--)
1000	CPUMPushHyper(pVCpu, paParam[i]);
1001
1002	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatWorldSwitch3264, z);
1003	/* Call the switcher. */
1004	int rc = pVM->hm.s.pfnHost32ToGuest64R0(pVM, RT_OFFSETOF(VM, aCpus[pVCpu->idCpu].cpum) - RT_OFFSETOF(VM, cpum));
1005	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatWorldSwitch3264, z);
1006
1007	/* Restore interrupts. */
1008	ASMSetFlags(uOldEFlags);
1009	return rc;
1010	}
1011
1012	#endif /* HC_ARCH_BITS == 32 && defined(VBOX_ENABLE_64_BITS_GUESTS) */
1013	/** @} */
1014
1015
1016	/**
1017	* Adds an exception to the intercept exception bitmap in the VMCB and updates
1018	* the corresponding VMCB Clean Bit.
1019	*
1020	* @param pVmcb Pointer to the VMCB.
1021	* @param u32Xcpt The value of the exception (X86_XCPT_*).
1022	*/
1023	DECLINLINE(void) hmR0SvmAddXcptIntercept(PSVMVMCB pVmcb, uint32_t u32Xcpt)
1024	{
1025	if (!(pVmcb->ctrl.u32InterceptException & RT_BIT(u32Xcpt)))
1026	{
1027	pVmcb->ctrl.u32InterceptException \|= RT_BIT(u32Xcpt);
1028	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1029	}
1030	}
1031
1032
1033	/**
1034	* Removes an exception from the intercept-exception bitmap in the VMCB and
1035	* updates the corresponding VMCB Clean Bit.
1036	*
1037	* @param pVmcb Pointer to the VMCB.
1038	* @param u32Xcpt The value of the exception (X86_XCPT_*).
1039	*/
1040	DECLINLINE(void) hmR0SvmRemoveXcptIntercept(PSVMVMCB pVmcb, uint32_t u32Xcpt)
1041	{
1042	#ifndef HMSVM_ALWAYS_TRAP_ALL_XCPTS
1043	if (pVmcb->ctrl.u32InterceptException & RT_BIT(u32Xcpt))
1044	{
1045	pVmcb->ctrl.u32InterceptException &= ~RT_BIT(u32Xcpt);
1046	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1047	}
1048	#endif
1049	}
1050
1051
1052	/**
1053	* Loads the guest control registers (CR0, CR2, CR3, CR4) into the VMCB.
1054	*
1055	* @returns VBox status code.
1056	* @param pVCpu Pointer to the VMCPU.
1057	* @param pVmcb Pointer to the VMCB.
1058	* @param pCtx Pointer the guest-CPU context.
1059	*
1060	* @remarks No-long-jump zone!!!
1061	*/
1062	DECLINLINE(int) hmR0SvmLoadGuestControlRegs(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1063	{
1064	/*
1065	* Guest CR0.
1066	*/
1067	PVM pVM = pVCpu->CTX_SUFF(pVM);
1068	if (pVCpu->hm.s.fContextUseFlags & HM_CHANGED_GUEST_CR0)
1069	{
1070	uint64_t u64GuestCR0 = pCtx->cr0;
1071
1072	/* Always enable caching. */
1073	u64GuestCR0 &= ~(X86_CR0_CD \| X86_CR0_NW);
1074
1075	/*
1076	* When Nested Paging is not available use shadow page tables and intercept #PFs (the latter done in SVMR0SetupVM()).
1077	*/
1078	if (!pVM->hm.s.fNestedPaging)
1079	{
1080	u64GuestCR0 \|= X86_CR0_PG; /* When Nested Paging is not available, use shadow page tables. */
1081	u64GuestCR0 \|= X86_CR0_WP; /* Guest CPL 0 writes to its read-only pages should cause a #PF VM-exit. */
1082	}
1083
1084	/*
1085	* Guest FPU bits.
1086	*/
1087	bool fInterceptNM = false;
1088	bool fInterceptMF = false;
1089	u64GuestCR0 \|= X86_CR0_NE; /* Use internal x87 FPU exceptions handling rather than external interrupts. */
1090	if (CPUMIsGuestFPUStateActive(pVCpu))
1091	{
1092	/* Catch floating point exceptions if we need to report them to the guest in a different way. */
1093	if (!(u64GuestCR0 & X86_CR0_NE))
1094	{
1095	Log4(("hmR0SvmLoadGuestControlRegs: Intercepting Guest CR0.MP Old-style FPU handling!!!\n"));
1096	fInterceptMF = true;
1097	}
1098	}
1099	else
1100	{
1101	fInterceptNM = true; /* Guest FPU inactive, VM-exit on #NM for lazy FPU loading. */
1102	u64GuestCR0 \|= X86_CR0_TS /* Guest can task switch quickly and do lazy FPU syncing. */
1103	\| X86_CR0_MP; /* FWAIT/WAIT should not ignore CR0.TS and should generate #NM. */
1104	}
1105
1106	/*
1107	* Update the exception intercept bitmap.
1108	*/
1109	if (fInterceptNM)
1110	hmR0SvmAddXcptIntercept(pVmcb, X86_XCPT_NM);
1111	else
1112	hmR0SvmRemoveXcptIntercept(pVmcb, X86_XCPT_NM);
1113
1114	if (fInterceptMF)
1115	hmR0SvmAddXcptIntercept(pVmcb, X86_XCPT_MF);
1116	else
1117	hmR0SvmRemoveXcptIntercept(pVmcb, X86_XCPT_MF);
1118
1119	pVmcb->guest.u64CR0 = u64GuestCR0;
1120	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1121	pVCpu->hm.s.fContextUseFlags &= ~HM_CHANGED_GUEST_CR0;
1122	}
1123
1124	/*
1125	* Guest CR2.
1126	*/
1127	if (pVCpu->hm.s.fContextUseFlags & HM_CHANGED_GUEST_CR2)
1128	{
1129	pVmcb->guest.u64CR2 = pCtx->cr2;
1130	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CR2;
1131	pVCpu->hm.s.fContextUseFlags &= ~HM_CHANGED_GUEST_CR2;
1132	}
1133
1134	/*
1135	* Guest CR3.
1136	*/
1137	if (pVCpu->hm.s.fContextUseFlags & HM_CHANGED_GUEST_CR3)
1138	{
1139	if (pVM->hm.s.fNestedPaging)
1140	{
1141	PGMMODE enmShwPagingMode;
1142	#if HC_ARCH_BITS == 32
1143	if (CPUMIsGuestInLongModeEx(pCtx))
1144	enmShwPagingMode = PGMMODE_AMD64_NX;
1145	else
1146	#endif
1147	enmShwPagingMode = PGMGetHostMode(pVM);
1148
1149	pVmcb->ctrl.u64NestedPagingCR3 = PGMGetNestedCR3(pVCpu, enmShwPagingMode);
1150	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
1151	Assert(pVmcb->ctrl.u64NestedPagingCR3);
1152	pVmcb->guest.u64CR3 = pCtx->cr3;
1153	}
1154	else
1155	pVmcb->guest.u64CR3 = PGMGetHyperCR3(pVCpu);
1156
1157	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1158	pVCpu->hm.s.fContextUseFlags &= ~HM_CHANGED_GUEST_CR3;
1159	}
1160
1161	/*
1162	* Guest CR4.
1163	*/
1164	if (pVCpu->hm.s.fContextUseFlags & HM_CHANGED_GUEST_CR4)
1165	{
1166	uint64_t u64GuestCR4 = pCtx->cr4;
1167	if (!pVM->hm.s.fNestedPaging)
1168	{
1169	switch (pVCpu->hm.s.enmShadowMode)
1170	{
1171	case PGMMODE_REAL:
1172	case PGMMODE_PROTECTED: /* Protected mode, no paging. */
1173	AssertFailed();
1174	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1175
1176	case PGMMODE_32_BIT: /* 32-bit paging. */
1177	u64GuestCR4 &= ~X86_CR4_PAE;
1178	break;
1179
1180	case PGMMODE_PAE: /* PAE paging. */
1181	case PGMMODE_PAE_NX: /* PAE paging with NX enabled. */
1182	/** Must use PAE paging as we could use physical memory > 4 GB */
1183	u64GuestCR4 \|= X86_CR4_PAE;
1184	break;
1185
1186	case PGMMODE_AMD64: /* 64-bit AMD paging (long mode). */
1187	case PGMMODE_AMD64_NX: /* 64-bit AMD paging (long mode) with NX enabled. */
1188	#ifdef VBOX_ENABLE_64_BITS_GUESTS
1189	break;
1190	#else
1191	AssertFailed();
1192	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1193	#endif
1194
1195	default: /* shut up gcc */
1196	AssertFailed();
1197	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1198	}
1199	}
1200
1201	pVmcb->guest.u64CR4 = u64GuestCR4;
1202	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1203	pVCpu->hm.s.fContextUseFlags &= ~HM_CHANGED_GUEST_CR4;
1204	}
1205
1206	return VINF_SUCCESS;
1207	}
1208
1209
1210	/**
1211	* Loads the guest segment registers into the VMCB.
1212	*
1213	* @returns VBox status code.
1214	* @param pVCpu Pointer to the VMCPU.
1215	* @param pVmcb Pointer to the VMCB.
1216	* @param pCtx Pointer to the guest-CPU context.
1217	*
1218	* @remarks No-long-jump zone!!!
1219	*/
1220	DECLINLINE(void) hmR0SvmLoadGuestSegmentRegs(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1221	{
1222	/* Guest Segment registers: CS, SS, DS, ES, FS, GS. */
1223	if (pVCpu->hm.s.fContextUseFlags & HM_CHANGED_GUEST_SEGMENT_REGS)
1224	{
1225	HMSVM_LOAD_SEG_REG(CS, cs);
1226	HMSVM_LOAD_SEG_REG(SS, ss);
1227	HMSVM_LOAD_SEG_REG(DS, ds);
1228	HMSVM_LOAD_SEG_REG(ES, es);
1229	HMSVM_LOAD_SEG_REG(FS, fs);
1230	HMSVM_LOAD_SEG_REG(GS, gs);
1231
1232	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_SEG;
1233	pVCpu->hm.s.fContextUseFlags &= ~HM_CHANGED_GUEST_SEGMENT_REGS;
1234	}
1235
1236	/* Guest TR. */
1237	if (pVCpu->hm.s.fContextUseFlags & HM_CHANGED_GUEST_TR)
1238	{
1239	HMSVM_LOAD_SEG_REG(TR, tr);
1240	pVCpu->hm.s.fContextUseFlags &= ~HM_CHANGED_GUEST_TR;
1241	}
1242
1243	/* Guest LDTR. */
1244	if (pVCpu->hm.s.fContextUseFlags & HM_CHANGED_GUEST_LDTR)
1245	{
1246	HMSVM_LOAD_SEG_REG(LDTR, ldtr);
1247	pVCpu->hm.s.fContextUseFlags &= ~HM_CHANGED_GUEST_LDTR;
1248	}
1249
1250	/* Guest GDTR. */
1251	if (pVCpu->hm.s.fContextUseFlags & HM_CHANGED_GUEST_GDTR)
1252	{
1253	pVmcb->guest.GDTR.u32Limit = pCtx->gdtr.cbGdt;
1254	pVmcb->guest.GDTR.u64Base = pCtx->gdtr.pGdt;
1255	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DT;
1256	pVCpu->hm.s.fContextUseFlags &= ~HM_CHANGED_GUEST_GDTR;
1257	}
1258
1259	/* Guest IDTR. */
1260	if (pVCpu->hm.s.fContextUseFlags & HM_CHANGED_GUEST_IDTR)
1261	{
1262	pVmcb->guest.IDTR.u32Limit = pCtx->idtr.cbIdt;
1263	pVmcb->guest.IDTR.u64Base = pCtx->idtr.pIdt;
1264	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DT;
1265	pVCpu->hm.s.fContextUseFlags &= ~HM_CHANGED_GUEST_IDTR;
1266	}
1267	}
1268
1269
1270	/**
1271	* Loads the guest MSRs into the VMCB.
1272	*
1273	* @param pVCpu Pointer to the VMCPU.
1274	* @param pVmcb Pointer to the VMCB.
1275	* @param pCtx Pointer to the guest-CPU context.
1276	*
1277	* @remarks No-long-jump zone!!!
1278	*/
1279	DECLINLINE(void) hmR0SvmLoadGuestMsrs(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1280	{
1281	/* Guest Sysenter MSRs. */
1282	pVmcb->guest.u64SysEnterCS = pCtx->SysEnter.cs;
1283	pVmcb->guest.u64SysEnterEIP = pCtx->SysEnter.eip;
1284	pVmcb->guest.u64SysEnterESP = pCtx->SysEnter.esp;
1285
1286	/*
1287	* Guest EFER MSR.
1288	* AMD-V requires guest EFER.SVME to be set. Weird. .
1289	* See AMD spec. 15.5.1 "Basic Operation" \| "Canonicalization and Consistency Checks".
1290	*/
1291	if (pVCpu->hm.s.fContextUseFlags & HM_CHANGED_SVM_GUEST_EFER_MSR)
1292	{
1293	pVmcb->guest.u64EFER = pCtx->msrEFER \| MSR_K6_EFER_SVME;
1294	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1295	pVCpu->hm.s.fContextUseFlags &= ~HM_CHANGED_SVM_GUEST_EFER_MSR;
1296	}
1297
1298	/* 64-bit MSRs. */
1299	if (CPUMIsGuestInLongModeEx(pCtx))
1300	{
1301	pVmcb->guest.FS.u64Base = pCtx->fs.u64Base;
1302	pVmcb->guest.GS.u64Base = pCtx->gs.u64Base;
1303	}
1304	else
1305	{
1306	/* If the guest isn't in 64-bit mode, clear MSR_K6_LME bit from guest EFER otherwise AMD-V expects amd64 shadow paging. */
1307	if (pCtx->msrEFER & MSR_K6_EFER_LME)
1308	{
1309	pVmcb->guest.u64EFER &= ~MSR_K6_EFER_LME;
1310	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1311	}
1312	}
1313
1314
1315	/** @todo The following are used in 64-bit only (SYSCALL/SYSRET) but they might
1316	* be writable in 32-bit mode. Clarify with AMD spec. */
1317	pVmcb->guest.u64STAR = pCtx->msrSTAR;
1318	pVmcb->guest.u64LSTAR = pCtx->msrLSTAR;
1319	pVmcb->guest.u64CSTAR = pCtx->msrCSTAR;
1320	pVmcb->guest.u64SFMASK = pCtx->msrSFMASK;
1321	pVmcb->guest.u64KernelGSBase = pCtx->msrKERNELGSBASE;
1322	}
1323
1324
1325	/**
1326	* Loads the guest debug registers into the VMCB.
1327	*
1328	* @param pVCpu Pointer to the VMCPU.
1329	* @param pVmcb Pointer to the VMCB.
1330	* @param pCtx Pointer to the guest-CPU context.
1331	*
1332	* @remarks No-long-jump zone!!!
1333	* @remarks Requires EFLAGS to be up-to-date in the VMCB!
1334	*/
1335	DECLINLINE(void) hmR0SvmLoadGuestDebugRegs(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1336	{
1337	if (!(pVCpu->hm.s.fContextUseFlags & HM_CHANGED_GUEST_DEBUG))
1338	return;
1339	Assert((pCtx->dr[6] & X86_DR6_RA1_MASK) == X86_DR6_RA1_MASK); Assert((pCtx->dr[6] & X86_DR6_RAZ_MASK) == 0);
1340	Assert((pCtx->dr[7] & X86_DR7_RA1_MASK) == X86_DR7_RA1_MASK); Assert((pCtx->dr[7] & X86_DR7_RAZ_MASK) == 0);
1341
1342	bool fInterceptDB = false;
1343	bool fInterceptMovDRx = false;
1344
1345	/*
1346	* Anyone single stepping on the host side? If so, we'll have to use the
1347	* trap flag in the guest EFLAGS since AMD-V doesn't have a trap flag on
1348	* the VMM level like VT-x implementations does.
1349	*/
1350	bool const fStepping = pVCpu->hm.s.fSingleInstruction \|\| DBGFIsStepping(pVCpu);
1351	if (fStepping)
1352	{
1353	pVCpu->hm.s.fClearTrapFlag = true;
1354	pVmcb->guest.u64RFlags \|= X86_EFL_TF;
1355	fInterceptDB = true;
1356	fInterceptMovDRx = true; /* Need clean DR6, no guest mess. */
1357	}
1358
1359	PVM pVM = pVCpu->CTX_SUFF(pVM);
1360	if (fStepping \|\| (CPUMGetHyperDR7(pVCpu) & X86_DR7_ENABLED_MASK))
1361	{
1362	/*
1363	* Use the combined guest and host DRx values found in the hypervisor
1364	* register set because the debugger has breakpoints active or someone
1365	* is single stepping on the host side.
1366	*
1367	* Note! DBGF expects a clean DR6 state before executing guest code.
1368	*/
1369	if (!CPUMIsHyperDebugStateActive(pVCpu))
1370	CPUMR0LoadHyperDebugState(pVCpu, false /* include DR6 */);
1371	Assert(!CPUMIsGuestDebugStateActive(pVCpu));
1372	Assert(CPUMIsHyperDebugStateActive(pVCpu) \|\| HC_ARCH_BITS == 32);
1373
1374	/* Update DR6 & DR7. (The other DRx values are handled by CPUM one way or the other.) */
1375	if ( pVmcb->guest.u64DR6 != X86_DR6_INIT_VAL
1376	\|\| pVmcb->guest.u64DR7 != CPUMGetHyperDR7(pVCpu) )
1377	{
1378	pVmcb->guest.u64DR7 = CPUMGetHyperDR7(pVCpu);
1379	pVmcb->guest.u64DR6 = X86_DR6_INIT_VAL;
1380	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
1381	}
1382
1383	/** @todo If we cared, we could optimize to allow the guest to read registers
1384	* with the same values. */
1385	fInterceptDB = true;
1386	fInterceptMovDRx = true;
1387	Log5(("hm: Loaded hyper DRx\n"));
1388	}
1389	else
1390	{
1391	/*
1392	* Update DR6, DR7 with the guest values if necessary.
1393	*/
1394	if ( pVmcb->guest.u64DR7 != pCtx->dr[7]
1395	\|\| pVmcb->guest.u64DR6 != pCtx->dr[6])
1396	{
1397	pVmcb->guest.u64DR7 = pCtx->dr[7];
1398	pVmcb->guest.u64DR6 = pCtx->dr[6];
1399	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
1400	}
1401
1402	/*
1403	* If the guest has enabled debug registers, we need to load them prior to
1404	* executing guest code so they'll trigger at the right time.
1405	*/
1406	if (pCtx->dr[7] & (X86_DR7_ENABLED_MASK \| X86_DR7_GD)) /** @todo Why GD? */
1407	{
1408	if (!CPUMIsGuestDebugStateActive(pVCpu))
1409	{
1410	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
1411	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxArmed);
1412	}
1413	Assert(!CPUMIsHyperDebugStateActive(pVCpu));
1414	Assert(CPUMIsGuestDebugStateActive(pVCpu) \|\| HC_ARCH_BITS == 32);
1415	Log5(("hm: Loaded guest DRx\n"));
1416	}
1417	/*
1418	* If no debugging enabled, we'll lazy load DR0-3.
1419	*/
1420	else if (!CPUMIsGuestDebugStateActive(pVCpu))
1421	fInterceptMovDRx = true;
1422	}
1423
1424	/*
1425	* Set up the intercepts.
1426	*/
1427	if (fInterceptDB)
1428	hmR0SvmAddXcptIntercept(pVmcb, X86_XCPT_DB);
1429	else
1430	hmR0SvmRemoveXcptIntercept(pVmcb, X86_XCPT_DB);
1431
1432	if (fInterceptMovDRx)
1433	{
1434	if ( pVmcb->ctrl.u16InterceptRdDRx != 0xffff
1435	\|\| pVmcb->ctrl.u16InterceptWrDRx != 0xffff)
1436	{
1437	pVmcb->ctrl.u16InterceptRdDRx = 0xffff;
1438	pVmcb->ctrl.u16InterceptWrDRx = 0xffff;
1439	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1440	}
1441	}
1442	else
1443	{
1444	if ( pVmcb->ctrl.u16InterceptRdDRx
1445	\|\| pVmcb->ctrl.u16InterceptWrDRx)
1446	{
1447	pVmcb->ctrl.u16InterceptRdDRx = 0;
1448	pVmcb->ctrl.u16InterceptWrDRx = 0;
1449	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1450	}
1451	}
1452
1453	pVCpu->hm.s.fContextUseFlags &= ~HM_CHANGED_GUEST_DEBUG;
1454	}
1455
1456
1457	/**
1458	* Loads the guest APIC state (currently just the TPR).
1459	*
1460	* @returns VBox status code.
1461	* @param pVCpu Pointer to the VMCPU.
1462	* @param pVmcb Pointer to the VMCB.
1463	* @param pCtx Pointer to the guest-CPU context.
1464	*/
1465	DECLINLINE(int) hmR0SvmLoadGuestApicState(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1466	{
1467	if (!(pVCpu->hm.s.fContextUseFlags & HM_CHANGED_SVM_GUEST_APIC_STATE))
1468	return VINF_SUCCESS;
1469
1470	bool fPendingIntr;
1471	uint8_t u8Tpr;
1472	int rc = PDMApicGetTPR(pVCpu, &u8Tpr, &fPendingIntr, NULL /* pu8PendingIrq */);
1473	AssertRCReturn(rc, rc);
1474
1475	/** Assume that we need to trap all TPR accesses and thus need not check on
1476	* every #VMEXIT if we should update the TPR. */
1477	Assert(pVmcb->ctrl.IntCtrl.n.u1VIrqMasking);
1478	pVCpu->hm.s.svm.fSyncVTpr = false;
1479
1480	/* 32-bit guests uses LSTAR MSR for patching guest code which touches the TPR. */
1481	if (pVCpu->CTX_SUFF(pVM)->hm.s.fTPRPatchingActive)
1482	{
1483	pCtx->msrLSTAR = u8Tpr;
1484
1485	/* If there are interrupts pending, intercept LSTAR writes, otherwise don't intercept reads or writes. */
1486	if (fPendingIntr)
1487	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_INTERCEPT_WRITE);
1488	else
1489	{
1490	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1491	pVCpu->hm.s.svm.fSyncVTpr = true;
1492	}
1493	}
1494	else
1495	{
1496	/* Bits 3-0 of the VTPR field correspond to bits 7-4 of the TPR (which is the Task-Priority Class). */
1497	pVmcb->ctrl.IntCtrl.n.u8VTPR = (u8Tpr >> 4);
1498
1499	/* If there are interrupts pending, intercept CR8 writes to evaluate ASAP if we can deliver the interrupt to the guest. */
1500	if (fPendingIntr)
1501	pVmcb->ctrl.u16InterceptWrCRx \|= RT_BIT(8);
1502	else
1503	{
1504	pVmcb->ctrl.u16InterceptWrCRx &= ~RT_BIT(8);
1505	pVCpu->hm.s.svm.fSyncVTpr = true;
1506	}
1507
1508	pVmcb->ctrl.u64VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS \| HMSVM_VMCB_CLEAN_TPR);
1509	}
1510
1511	pVCpu->hm.s.fContextUseFlags &= ~HM_CHANGED_SVM_GUEST_APIC_STATE;
1512	return rc;
1513	}
1514
1515
1516	/**
1517	* Sets up the appropriate function to run guest code.
1518	*
1519	* @returns VBox status code.
1520	* @param pVCpu Pointer to the VMCPU.
1521	* @param pCtx Pointer to the guest-CPU context.
1522	*
1523	* @remarks No-long-jump zone!!!
1524	*/
1525	static int hmR0SvmSetupVMRunHandler(PVMCPU pVCpu, PCPUMCTX pCtx)
1526	{
1527	if (CPUMIsGuestInLongModeEx(pCtx))
1528	{
1529	#ifndef VBOX_ENABLE_64_BITS_GUESTS
1530	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1531	#endif
1532	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.fAllow64BitGuests); /* Guaranteed by hmR3InitFinalizeR0(). */
1533	#if HC_ARCH_BITS == 32 && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
1534	/* 32-bit host. We need to switch to 64-bit before running the 64-bit guest. */
1535	pVCpu->hm.s.svm.pfnVMRun = SVMR0VMSwitcherRun64;
1536	#else
1537	/* 64-bit host or hybrid host. */
1538	pVCpu->hm.s.svm.pfnVMRun = SVMR0VMRun64;
1539	#endif
1540	}
1541	else
1542	{
1543	/* Guest is not in long mode, use the 32-bit handler. */
1544	pVCpu->hm.s.svm.pfnVMRun = SVMR0VMRun;
1545	}
1546	return VINF_SUCCESS;
1547	}
1548
1549
1550	/**
1551	* Enters the AMD-V session.
1552	*
1553	* @returns VBox status code.
1554	* @param pVM Pointer to the VM.
1555	* @param pVCpu Pointer to the VMCPU.
1556	* @param pCpu Pointer to the CPU info struct.
1557	*/
1558	VMMR0DECL(int) SVMR0Enter(PVM pVM, PVMCPU pVCpu, PHMGLOBALCPUINFO pCpu)
1559	{
1560	AssertPtr(pVM);
1561	AssertPtr(pVCpu);
1562	Assert(pVM->hm.s.svm.fSupported);
1563	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1564	NOREF(pCpu);
1565
1566	LogFlowFunc(("pVM=%p pVCpu=%p\n", pVM, pVCpu));
1567
1568	pVCpu->hm.s.fLeaveDone = false;
1569	return VINF_SUCCESS;
1570	}
1571
1572
1573	/**
1574	* Thread-context callback for AMD-V.
1575	*
1576	* @param enmEvent The thread-context event.
1577	* @param pVCpu Pointer to the VMCPU.
1578	* @param fGlobalInit Whether global VT-x/AMD-V init. is used.
1579	*/
1580	VMMR0DECL(void) SVMR0ThreadCtxCallback(RTTHREADCTXEVENT enmEvent, PVMCPU pVCpu, bool fGlobalInit)
1581	{
1582	switch (enmEvent)
1583	{
1584	case RTTHREADCTXEVENT_PREEMPTING:
1585	{
1586	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1587	Assert(VMMR0ThreadCtxHooksAreRegistered(pVCpu));
1588	VMCPU_ASSERT_EMT(pVCpu);
1589
1590	PVM pVM = pVCpu->CTX_SUFF(pVM);
1591	PCPUMCTX pCtx = CPUMQueryGuestCtxPtr(pVCpu);
1592	VMMRZCallRing3Disable(pVCpu); /* No longjmps (log-flush, locks) in this fragile context. */
1593
1594	if (!pVCpu->hm.s.fLeaveDone)
1595	{
1596	hmR0SvmLeave(pVM, pVCpu, pCtx);
1597	pVCpu->hm.s.fLeaveDone = true;
1598	}
1599
1600	int rc = HMR0LeaveCpu(pVCpu); /* Leave HM context, takes care of local init (term). */
1601	AssertRC(rc); NOREF(rc);
1602
1603	VMMRZCallRing3Enable(pVCpu); /* Restore longjmp state. */
1604	break;
1605	}
1606
1607	case RTTHREADCTXEVENT_RESUMED:
1608	{
1609	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1610	Assert(VMMR0ThreadCtxHooksAreRegistered(pVCpu));
1611	VMCPU_ASSERT_EMT(pVCpu);
1612
1613	VMMRZCallRing3Disable(pVCpu); /* No longjmps (log-flush, locks) in this fragile context. */
1614
1615	/*
1616	* Initialize the bare minimum state required for HM. This takes care of
1617	* initializing AMD-V if necessary (onlined CPUs, local init etc.)
1618	*/
1619	int rc = HMR0EnterCpu(pVCpu);
1620	AssertRC(rc); NOREF(rc);
1621	Assert(pVCpu->hm.s.fContextUseFlags & (HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_HOST_GUEST_SHARED_STATE));
1622
1623	pVCpu->hm.s.fLeaveDone = false;
1624	VMMRZCallRing3Enable(pVCpu); /* Restore longjmp state. */
1625	break;
1626	}
1627
1628	default:
1629	break;
1630	}
1631	}
1632
1633
1634	/**
1635	* Saves the host state.
1636	*
1637	* @returns VBox status code.
1638	* @param pVM Pointer to the VM.
1639	* @param pVCpu Pointer to the VMCPU.
1640	*
1641	* @remarks No-long-jump zone!!!
1642	*/
1643	VMMR0DECL(int) SVMR0SaveHostState(PVM pVM, PVMCPU pVCpu)
1644	{
1645	NOREF(pVM);
1646	NOREF(pVCpu);
1647	/* Nothing to do here. AMD-V does this for us automatically during the world-switch. */
1648	pVCpu->hm.s.fContextUseFlags &= ~HM_CHANGED_HOST_CONTEXT;
1649	return VINF_SUCCESS;
1650	}
1651
1652
1653	/**
1654	* Loads the guest state into the VMCB. The CPU state will be loaded from these
1655	* fields on every successful VM-entry.
1656	*
1657	* Sets up the appropriate VMRUN function to execute guest code based
1658	* on the guest CPU mode.
1659	*
1660	* @returns VBox status code.
1661	* @param pVM Pointer to the VM.
1662	* @param pVCpu Pointer to the VMCPU.
1663	* @param pMixedCtx Pointer to the guest-CPU context.
1664	*
1665	* @remarks No-long-jump zone!!!
1666	*/
1667	static int hmR0SvmLoadGuestState(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
1668	{
1669	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
1670	AssertMsgReturn(pVmcb, ("Invalid pVmcb\n"), VERR_SVM_INVALID_PVMCB);
1671
1672	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatLoadGuestState, x);
1673
1674	int rc = hmR0SvmLoadGuestControlRegs(pVCpu, pVmcb, pCtx);
1675	AssertLogRelMsgRCReturn(rc, ("hmR0SvmLoadGuestControlRegs! rc=%Rrc (pVM=%p pVCpu=%p)\n", rc, pVM, pVCpu), rc);
1676
1677	hmR0SvmLoadGuestSegmentRegs(pVCpu, pVmcb, pCtx);
1678	hmR0SvmLoadGuestMsrs(pVCpu, pVmcb, pCtx);
1679
1680	pVmcb->guest.u64RIP = pCtx->rip;
1681	pVmcb->guest.u64RSP = pCtx->rsp;
1682	pVmcb->guest.u64RFlags = pCtx->eflags.u32;
1683	pVmcb->guest.u8CPL = pCtx->ss.Attr.n.u2Dpl;
1684	pVmcb->guest.u64RAX = pCtx->rax;
1685
1686	/* hmR0SvmLoadGuestDebugRegs() must be called -after- updating guest RFLAGS as the RFLAGS may need to be changed. */
1687	hmR0SvmLoadGuestDebugRegs(pVCpu, pVmcb, pCtx);
1688
1689	rc = hmR0SvmLoadGuestApicState(pVCpu, pVmcb, pCtx);
1690	AssertLogRelMsgRCReturn(rc, ("hmR0SvmLoadGuestApicState! rc=%Rrc (pVM=%p pVCpu=%p)\n", rc, pVM, pVCpu), rc);
1691
1692	rc = hmR0SvmSetupVMRunHandler(pVCpu, pCtx);
1693	AssertLogRelMsgRCReturn(rc, ("hmR0SvmSetupVMRunHandler! rc=%Rrc (pVM=%p pVCpu=%p)\n", rc, pVM, pVCpu), rc);
1694
1695	/* Clear any unused and reserved bits. */
1696	pVCpu->hm.s.fContextUseFlags &= ~( HM_CHANGED_GUEST_RIP /* Unused (loaded unconditionally). */
1697	\| HM_CHANGED_GUEST_RSP
1698	\| HM_CHANGED_GUEST_RFLAGS
1699	\| HM_CHANGED_GUEST_SYSENTER_CS_MSR
1700	\| HM_CHANGED_GUEST_SYSENTER_EIP_MSR
1701	\| HM_CHANGED_GUEST_SYSENTER_ESP_MSR
1702	\| HM_CHANGED_SVM_RESERVED1 /* Reserved. */
1703	\| HM_CHANGED_SVM_RESERVED2
1704	\| HM_CHANGED_SVM_RESERVED3);
1705
1706	/* All the guest state bits should be loaded except maybe the host context and shared host/guest bits. */
1707	AssertMsg( !(pVCpu->hm.s.fContextUseFlags & HM_CHANGED_ALL_GUEST)
1708	\|\| !(pVCpu->hm.s.fContextUseFlags & ~(HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_HOST_GUEST_SHARED_STATE)),
1709	("Missed updating flags while loading guest state. pVM=%p pVCpu=%p fContextUseFlags=%#RX32\n",
1710	pVM, pVCpu, pVCpu->hm.s.fContextUseFlags));
1711
1712	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));
1713
1714	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatLoadGuestState, x);
1715	return rc;
1716	}
1717
1718
1719	/**
1720	* Saves the entire guest state from the VMCB into the
1721	* guest-CPU context. Currently there is no residual state left in the CPU that
1722	* is not updated in the VMCB.
1723	*
1724	* @returns VBox status code.
1725	* @param pVCpu Pointer to the VMCPU.
1726	* @param pMixedCtx Pointer to the guest-CPU context. The data may be
1727	* out-of-sync. Make sure to update the required fields
1728	* before using them.
1729	*/
1730	static void hmR0SvmSaveGuestState(PVMCPU pVCpu, PCPUMCTX pMixedCtx)
1731	{
1732	Assert(VMMRZCallRing3IsEnabled(pVCpu));
1733
1734	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
1735
1736	pMixedCtx->rip = pVmcb->guest.u64RIP;
1737	pMixedCtx->rsp = pVmcb->guest.u64RSP;
1738	pMixedCtx->eflags.u32 = pVmcb->guest.u64RFlags;
1739	pMixedCtx->rax = pVmcb->guest.u64RAX;
1740
1741	/*
1742	* Guest interrupt shadow.
1743	*/
1744	if (pVmcb->ctrl.u64IntShadow & SVM_INTERRUPT_SHADOW_ACTIVE)
1745	EMSetInhibitInterruptsPC(pVCpu, pMixedCtx->rip);
1746	else
1747	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS);
1748
1749	/*
1750	* Guest Control registers: CR2, CR3 (handled at the end) - accesses to other control registers are always intercepted.
1751	*/
1752	pMixedCtx->cr2 = pVmcb->guest.u64CR2;
1753
1754	/*
1755	* Guest MSRs.
1756	*/
1757	pMixedCtx->msrSTAR = pVmcb->guest.u64STAR; /* legacy syscall eip, cs & ss */
1758	pMixedCtx->msrLSTAR = pVmcb->guest.u64LSTAR; /* 64-bit mode syscall rip */
1759	pMixedCtx->msrCSTAR = pVmcb->guest.u64CSTAR; /* compatibility mode syscall rip */
1760	pMixedCtx->msrSFMASK = pVmcb->guest.u64SFMASK; /* syscall flag mask */
1761	pMixedCtx->msrKERNELGSBASE = pVmcb->guest.u64KernelGSBase; /* swapgs exchange value */
1762	pMixedCtx->SysEnter.cs = pVmcb->guest.u64SysEnterCS;
1763	pMixedCtx->SysEnter.eip = pVmcb->guest.u64SysEnterEIP;
1764	pMixedCtx->SysEnter.esp = pVmcb->guest.u64SysEnterESP;
1765
1766	/*
1767	* Guest segment registers (includes FS, GS base MSRs for 64-bit guests).
1768	*/
1769	HMSVM_SAVE_SEG_REG(CS, cs);
1770	HMSVM_SAVE_SEG_REG(SS, ss);
1771	HMSVM_SAVE_SEG_REG(DS, ds);
1772	HMSVM_SAVE_SEG_REG(ES, es);
1773	HMSVM_SAVE_SEG_REG(FS, fs);
1774	HMSVM_SAVE_SEG_REG(GS, gs);
1775
1776	/*
1777	* Correct the hidden CS granularity bit. Haven't seen it being wrong in any other
1778	* register (yet).
1779	*/
1780	/** @todo SELM might need to be fixed as it too should not care about the
1781	* granularity bit. See @bugref{6785}. */
1782	if ( !pMixedCtx->cs.Attr.n.u1Granularity
1783	&& pMixedCtx->cs.Attr.n.u1Present
1784	&& pMixedCtx->cs.u32Limit > UINT32_C(0xfffff))
1785	{
1786	Assert((pMixedCtx->cs.u32Limit & 0xfff) == 0xfff);
1787	pMixedCtx->cs.Attr.n.u1Granularity = 1;
1788	}
1789
1790	#ifdef VBOX_STRICT
1791	# define HMSVM_ASSERT_SEG_GRANULARITY(reg) \
1792	AssertMsg( !pMixedCtx->reg.Attr.n.u1Present \
1793	\|\| ( pMixedCtx->reg.Attr.n.u1Granularity \
1794	? (pMixedCtx->reg.u32Limit & 0xfff) == 0xfff \
1795	: pMixedCtx->reg.u32Limit <= UINT32_C(0xfffff)), \
1796	("Invalid Segment Attributes Limit=%#RX32 Attr=%#RX32 Base=%#RX64\n", pMixedCtx->reg.u32Limit, \
1797	pMixedCtx->reg.Attr.u, pMixedCtx->reg.u64Base))
1798
1799	HMSVM_ASSERT_SEG_GRANULARITY(cs);
1800	HMSVM_ASSERT_SEG_GRANULARITY(ss);
1801	HMSVM_ASSERT_SEG_GRANULARITY(ds);
1802	HMSVM_ASSERT_SEG_GRANULARITY(es);
1803	HMSVM_ASSERT_SEG_GRANULARITY(fs);
1804	HMSVM_ASSERT_SEG_GRANULARITY(gs);
1805
1806	# undef HMSVM_ASSERT_SEL_GRANULARITY
1807	#endif
1808
1809	/*
1810	* Sync the hidden SS DPL field. AMD CPUs have a separate CPL field in the VMCB and uses that
1811	* and thus it's possible that when the CPL changes during guest execution that the SS DPL
1812	* isn't updated by AMD-V. Observed on some AMD Fusion CPUs with 64-bit guests.
1813	* See AMD spec. 15.5.1 "Basic operation".
1814	*/
1815	Assert(!(pVmcb->guest.u8CPL & ~0x3));
1816	pMixedCtx->ss.Attr.n.u2Dpl = pVmcb->guest.u8CPL & 0x3;
1817
1818	/*
1819	* Guest Descriptor-Table registers.
1820	*/
1821	HMSVM_SAVE_SEG_REG(TR, tr);
1822	HMSVM_SAVE_SEG_REG(LDTR, ldtr);
1823	pMixedCtx->gdtr.cbGdt = pVmcb->guest.GDTR.u32Limit;
1824	pMixedCtx->gdtr.pGdt = pVmcb->guest.GDTR.u64Base;
1825
1826	pMixedCtx->idtr.cbIdt = pVmcb->guest.IDTR.u32Limit;
1827	pMixedCtx->idtr.pIdt = pVmcb->guest.IDTR.u64Base;
1828
1829	/*
1830	* Guest Debug registers.
1831	*/
1832	if (!CPUMIsHyperDebugStateActive(pVCpu))
1833	{
1834	pMixedCtx->dr[6] = pVmcb->guest.u64DR6;
1835	pMixedCtx->dr[7] = pVmcb->guest.u64DR7;
1836	}
1837	else
1838	{
1839	Assert(pVmcb->guest.u64DR7 == CPUMGetHyperDR7(pVCpu));
1840	CPUMSetHyperDR6(pVCpu, pVmcb->guest.u64DR6);
1841	}
1842
1843	/*
1844	* With Nested Paging, CR3 changes are not intercepted. Therefore, sync. it now.
1845	* This is done as the very last step of syncing the guest state, as PGMUpdateCR3() may cause longjmp's to ring-3.
1846	*/
1847	if ( pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging
1848	&& pMixedCtx->cr3 != pVmcb->guest.u64CR3)
1849	{
1850	CPUMSetGuestCR3(pVCpu, pVmcb->guest.u64CR3);
1851	PGMUpdateCR3(pVCpu, pVmcb->guest.u64CR3);
1852	}
1853	}
1854
1855
1856	/**
1857	* Does the necessary state syncing before returning to ring-3 for any reason
1858	* (longjmp, preemption, voluntary exits to ring-3) from AMD-V.
1859	*
1860	* @param pVM Pointer to the VM.
1861	* @param pVCpu Pointer to the VMCPU.
1862	* @param pMixedCtx Pointer to the guest-CPU context.
1863	*
1864	* @remarks No-long-jmp zone!!!
1865	*/
1866	static void hmR0SvmLeave(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
1867	{
1868	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1869	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
1870	Assert(VMMR0IsLogFlushDisabled(pVCpu));
1871
1872	/* Restore host FPU state if necessary and resync on next R0 reentry .*/
1873	if (CPUMIsGuestFPUStateActive(pVCpu))
1874	{
1875	CPUMR0SaveGuestFPU(pVM, pVCpu, pCtx);
1876	Assert(!CPUMIsGuestFPUStateActive(pVCpu));
1877	pVCpu->hm.s.fContextUseFlags \|= HM_CHANGED_GUEST_CR0;
1878	}
1879
1880	/*
1881	* Restore host debug registers if necessary and resync on next R0 reentry.
1882	*/
1883	#ifdef VBOX_STRICT
1884	if (CPUMIsHyperDebugStateActive(pVCpu))
1885	{
1886	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
1887	Assert(pVmcb->ctrl.u16InterceptRdDRx == 0xffff);
1888	Assert(pVmcb->ctrl.u16InterceptWrDRx == 0xffff);
1889	}
1890	#endif
1891	if (CPUMR0DebugStateMaybeSaveGuestAndRestoreHost(pVCpu, false /* save DR6 */))
1892	pVCpu->hm.s.fContextUseFlags \|= HM_CHANGED_GUEST_DEBUG;
1893
1894	Assert(!CPUMIsHyperDebugStateActive(pVCpu));
1895	Assert(!CPUMIsGuestDebugStateActive(pVCpu));
1896
1897	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatEntry);
1898	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatLoadGuestState);
1899	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatExit1);
1900	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatExit2);
1901	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchLongJmpToR3);
1902
1903	VMCPU_CMPXCHG_STATE(pVCpu, VMCPUSTATE_STARTED_HM, VMCPUSTATE_STARTED_EXEC);
1904	}
1905
1906
1907	/**
1908	* Leaves the AMD-V session.
1909	*
1910	* @returns VBox status code.
1911	* @param pVM Pointer to the VM.
1912	* @param pVCpu Pointer to the VMCPU.
1913	* @param pCtx Pointer to the guest-CPU context.
1914	*/
1915	DECLINLINE(void) hmR0SvmLeaveSession(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
1916	{
1917	HM_DISABLE_PREEMPT_IF_NEEDED();
1918	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
1919	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1920
1921	/* When thread-context hooks are used, we can avoid doing the leave again if we had been preempted before
1922	and done this from the VMXR0ThreadCtxCallback(). */
1923	if (!pVCpu->hm.s.fLeaveDone)
1924	{
1925	hmR0SvmLeave(pVM, pVCpu, pCtx);
1926	pVCpu->hm.s.fLeaveDone = true;
1927	}
1928
1929	/* Deregister hook now that we've left HM context before re-enabling preemption. */
1930	/** @todo This is bad. Deregistering here means we need to VMCLEAR always
1931	* (longjmp/exit-to-r3) in VT-x which is not efficient. */
1932	if (VMMR0ThreadCtxHooksAreRegistered(pVCpu))
1933	VMMR0ThreadCtxHooksDeregister(pVCpu);
1934
1935	/* Leave HM context. This takes care of local init (term). */
1936	int rc = HMR0LeaveCpu(pVCpu);
1937	AssertRC(rc); NOREF(rc);
1938
1939	HM_RESTORE_PREEMPT_IF_NEEDED();
1940	}
1941
1942
1943	/**
1944	* Does the necessary state syncing before doing a longjmp to ring-3.
1945	*
1946	* @param pVM Pointer to the VM.
1947	* @param pVCpu Pointer to the VMCPU.
1948	* @param pCtx Pointer to the guest-CPU context.
1949	*
1950	* @remarks No-long-jmp zone!!!
1951	*/
1952	static void hmR0SvmLongJmpToRing3(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
1953	{
1954	hmR0SvmLeaveSession(pVM, pVCpu, pCtx);
1955	}
1956
1957
1958	/**
1959	* VMMRZCallRing3() callback wrapper which saves the guest state (or restores
1960	* any remaining host state) before we longjump to ring-3 and possibly get
1961	* preempted.
1962	*
1963	* @param pVCpu Pointer to the VMCPU.
1964	* @param enmOperation The operation causing the ring-3 longjump.
1965	* @param pvUser The user argument (pointer to the possibly
1966	* out-of-date guest-CPU context).
1967	*
1968	* @remarks Must never be called with @a enmOperation ==
1969	* VMMCALLRING3_VM_R0_ASSERTION.
1970	*/
1971	DECLCALLBACK(void) hmR0SvmCallRing3Callback(PVMCPU pVCpu, VMMCALLRING3 enmOperation, void *pvUser)
1972	{
1973	/* VMMRZCallRing3() already makes sure we never get called as a result of an longjmp due to an assertion, */
1974	Assert(pVCpu);
1975	Assert(pvUser);
1976	Assert(VMMRZCallRing3IsEnabled(pVCpu));
1977	HMSVM_ASSERT_PREEMPT_SAFE();
1978
1979	VMMRZCallRing3Disable(pVCpu);
1980	Assert(VMMR0IsLogFlushDisabled(pVCpu));
1981
1982	Log4(("hmR0SvmCallRing3Callback->hmR0SvmLongJmpToRing3\n"));
1983	hmR0SvmLongJmpToRing3(pVCpu->CTX_SUFF(pVM), pVCpu, (PCPUMCTX)pvUser);
1984
1985	VMMRZCallRing3Enable(pVCpu);
1986	}
1987
1988
1989	/**
1990	* Take necessary actions before going back to ring-3.
1991	*
1992	* An action requires us to go back to ring-3. This function does the necessary
1993	* steps before we can safely return to ring-3. This is not the same as longjmps
1994	* to ring-3, this is voluntary.
1995	*
1996	* @param pVM Pointer to the VM.
1997	* @param pVCpu Pointer to the VMCPU.
1998	* @param pCtx Pointer to the guest-CPU context.
1999	* @param rcExit The reason for exiting to ring-3. Can be
2000	* VINF_VMM_UNKNOWN_RING3_CALL.
2001	*/
2002	static void hmR0SvmExitToRing3(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, int rcExit)
2003	{
2004	Assert(pVM);
2005	Assert(pVCpu);
2006	Assert(pCtx);
2007	HMSVM_ASSERT_PREEMPT_SAFE();
2008
2009	if (RT_UNLIKELY(rcExit == VERR_SVM_INVALID_GUEST_STATE))
2010	{
2011	/* We don't need to do any syncing here, we're not going to come back to execute anything again. */
2012	return;
2013	}
2014
2015	/* Please, no longjumps here (any logging shouldn't flush jump back to ring-3). NO LOGGING BEFORE THIS POINT! */
2016	VMMRZCallRing3Disable(pVCpu);
2017	Log4(("hmR0SvmExitToRing3: rcExit=%d\n", rcExit));
2018
2019	/* We need to do this only while truly exiting the "inner loop" back to ring-3 and -not- for any longjmp to ring3. */
2020	if (pVCpu->hm.s.Event.fPending)
2021	{
2022	hmR0SvmPendingEventToTrpmTrap(pVCpu);
2023	Assert(!pVCpu->hm.s.Event.fPending);
2024	}
2025
2026	/* Sync. the necessary state for going back to ring-3. */
2027	hmR0SvmLeaveSession(pVM, pVCpu, pCtx);
2028	STAM_COUNTER_DEC(&pVCpu->hm.s.StatSwitchLongJmpToR3);
2029
2030	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_TO_R3);
2031	CPUMSetChangedFlags(pVCpu, CPUM_CHANGED_SYSENTER_MSR
2032	\| CPUM_CHANGED_LDTR
2033	\| CPUM_CHANGED_GDTR
2034	\| CPUM_CHANGED_IDTR
2035	\| CPUM_CHANGED_TR
2036	\| CPUM_CHANGED_HIDDEN_SEL_REGS);
2037	if ( pVM->hm.s.fNestedPaging
2038	&& CPUMIsGuestPagingEnabledEx(pCtx))
2039	{
2040	CPUMSetChangedFlags(pVCpu, CPUM_CHANGED_GLOBAL_TLB_FLUSH);
2041	}
2042
2043	/* Make sure we've undo the trap flag if we tried to single step something. */
2044	if (pVCpu->hm.s.fClearTrapFlag)
2045	{
2046	pCtx->eflags.Bits.u1TF = 0;
2047	pVCpu->hm.s.fClearTrapFlag = false;
2048	}
2049
2050	/* On our way back from ring-3 the following needs to be done. */
2051	/** @todo This can change with preemption hooks. */
2052	if (rcExit == VINF_EM_RAW_INTERRUPT)
2053	pVCpu->hm.s.fContextUseFlags \|= HM_CHANGED_HOST_CONTEXT;
2054	else
2055	pVCpu->hm.s.fContextUseFlags \|= HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_ALL_GUEST;
2056
2057	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchExitToR3);
2058	VMMRZCallRing3Enable(pVCpu);
2059	}
2060
2061
2062	/**
2063	* Updates the use of TSC offsetting mode for the CPU and adjusts the necessary
2064	* intercepts.
2065	*
2066	* @param pVCpu Pointer to the VMCPU.
2067	*
2068	* @remarks No-long-jump zone!!!
2069	*/
2070	static void hmR0SvmUpdateTscOffsetting(PVMCPU pVCpu)
2071	{
2072	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2073	if (TMCpuTickCanUseRealTSC(pVCpu, &pVmcb->ctrl.u64TSCOffset))
2074	{
2075	uint64_t u64CurTSC = ASMReadTSC();
2076	if (u64CurTSC + pVmcb->ctrl.u64TSCOffset > TMCpuTickGetLastSeen(pVCpu))
2077	{
2078	pVmcb->ctrl.u32InterceptCtrl1 &= ~SVM_CTRL1_INTERCEPT_RDTSC;
2079	pVmcb->ctrl.u32InterceptCtrl2 &= ~SVM_CTRL2_INTERCEPT_RDTSCP;
2080	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscOffset);
2081	}
2082	else
2083	{
2084	pVmcb->ctrl.u32InterceptCtrl1 \|= SVM_CTRL1_INTERCEPT_RDTSC;
2085	pVmcb->ctrl.u32InterceptCtrl2 \|= SVM_CTRL2_INTERCEPT_RDTSCP;
2086	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscInterceptOverFlow);
2087	}
2088	}
2089	else
2090	{
2091	pVmcb->ctrl.u32InterceptCtrl1 \|= SVM_CTRL1_INTERCEPT_RDTSC;
2092	pVmcb->ctrl.u32InterceptCtrl2 \|= SVM_CTRL2_INTERCEPT_RDTSCP;
2093	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscIntercept);
2094	}
2095
2096	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2097	}
2098
2099
2100	/**
2101	* Sets an event as a pending event to be injected into the guest.
2102	*
2103	* @param pVCpu Pointer to the VMCPU.
2104	* @param pEvent Pointer to the SVM event.
2105	* @param GCPtrFaultAddress The fault-address (CR2) in case it's a
2106	* page-fault.
2107	*
2108	* @remarks Statistics counter assumes this is a guest event being reflected to
2109	* the guest i.e. 'StatInjectPendingReflect' is incremented always.
2110	*/
2111	DECLINLINE(void) hmR0SvmSetPendingEvent(PVMCPU pVCpu, PSVMEVENT pEvent, RTGCUINTPTR GCPtrFaultAddress)
2112	{
2113	Assert(!pVCpu->hm.s.Event.fPending);
2114	Assert(pEvent->n.u1Valid);
2115
2116	pVCpu->hm.s.Event.u64IntrInfo = pEvent->u;
2117	pVCpu->hm.s.Event.fPending = true;
2118	pVCpu->hm.s.Event.GCPtrFaultAddress = GCPtrFaultAddress;
2119
2120	Log4(("hmR0SvmSetPendingEvent: u=%#RX64 u8Vector=%#x Type=%#x ErrorCodeValid=%RTbool ErrorCode=%#RX32\n", pEvent->u,
2121	pEvent->n.u8Vector, (uint8_t)pEvent->n.u3Type, !!pEvent->n.u1ErrorCodeValid, pEvent->n.u32ErrorCode));
2122
2123	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectPendingReflect);
2124	}
2125
2126
2127	/**
2128	* Injects an event into the guest upon VMRUN by updating the relevant field
2129	* in the VMCB.
2130	*
2131	* @param pVCpu Pointer to the VMCPU.
2132	* @param pVmcb Pointer to the guest VMCB.
2133	* @param pCtx Pointer to the guest-CPU context.
2134	* @param pEvent Pointer to the event.
2135	*
2136	* @remarks No-long-jump zone!!!
2137	* @remarks Requires CR0!
2138	*/
2139	DECLINLINE(void) hmR0SvmInjectEventVmcb(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx, PSVMEVENT pEvent)
2140	{
2141	pVmcb->ctrl.EventInject.u = pEvent->u;
2142	STAM_COUNTER_INC(&pVCpu->hm.s.paStatInjectedIrqsR0[pEvent->n.u8Vector & MASK_INJECT_IRQ_STAT]);
2143
2144	Log4(("hmR0SvmInjectEventVmcb: u=%#RX64 u8Vector=%#x Type=%#x ErrorCodeValid=%RTbool ErrorCode=%#RX32\n", pEvent->u,
2145	pEvent->n.u8Vector, (uint8_t)pEvent->n.u3Type, !!pEvent->n.u1ErrorCodeValid, pEvent->n.u32ErrorCode));
2146	}
2147
2148
2149
2150	/**
2151	* Converts any TRPM trap into a pending HM event. This is typically used when
2152	* entering from ring-3 (not longjmp returns).
2153	*
2154	* @param pVCpu Pointer to the VMCPU.
2155	*/
2156	static void hmR0SvmTrpmTrapToPendingEvent(PVMCPU pVCpu)
2157	{
2158	Assert(TRPMHasTrap(pVCpu));
2159	Assert(!pVCpu->hm.s.Event.fPending);
2160
2161	uint8_t uVector;
2162	TRPMEVENT enmTrpmEvent;
2163	RTGCUINT uErrCode;
2164	RTGCUINTPTR GCPtrFaultAddress;
2165	uint8_t cbInstr;
2166
2167	int rc = TRPMQueryTrapAll(pVCpu, &uVector, &enmTrpmEvent, &uErrCode, &GCPtrFaultAddress, &cbInstr);
2168	AssertRC(rc);
2169
2170	SVMEVENT Event;
2171	Event.u = 0;
2172	Event.n.u1Valid = 1;
2173	Event.n.u8Vector = uVector;
2174
2175	/* Refer AMD spec. 15.20 "Event Injection" for the format. */
2176	if (enmTrpmEvent == TRPM_TRAP)
2177	{
2178	Event.n.u3Type = SVM_EVENT_EXCEPTION;
2179	switch (uVector)
2180	{
2181	case X86_XCPT_PF:
2182	case X86_XCPT_DF:
2183	case X86_XCPT_TS:
2184	case X86_XCPT_NP:
2185	case X86_XCPT_SS:
2186	case X86_XCPT_GP:
2187	case X86_XCPT_AC:
2188	{
2189	Event.n.u1ErrorCodeValid = 1;
2190	Event.n.u32ErrorCode = uErrCode;
2191	break;
2192	}
2193	}
2194	}
2195	else if (enmTrpmEvent == TRPM_HARDWARE_INT)
2196	{
2197	if (uVector == X86_XCPT_NMI)
2198	Event.n.u3Type = SVM_EVENT_NMI;
2199	else
2200	Event.n.u3Type = SVM_EVENT_EXTERNAL_IRQ;
2201	}
2202	else if (enmTrpmEvent == TRPM_SOFTWARE_INT)
2203	Event.n.u3Type = SVM_EVENT_SOFTWARE_INT;
2204	else
2205	AssertMsgFailed(("Invalid TRPM event type %d\n", enmTrpmEvent));
2206
2207	rc = TRPMResetTrap(pVCpu);
2208	AssertRC(rc);
2209
2210	Log4(("TRPM->HM event: u=%#RX64 u8Vector=%#x uErrorCodeValid=%RTbool uErrorCode=%#RX32\n", Event.u, Event.n.u8Vector,
2211	!!Event.n.u1ErrorCodeValid, Event.n.u32ErrorCode));
2212
2213	hmR0SvmSetPendingEvent(pVCpu, &Event, GCPtrFaultAddress);
2214	STAM_COUNTER_DEC(&pVCpu->hm.s.StatInjectPendingReflect);
2215	}
2216
2217
2218	/**
2219	* Converts any pending SVM event into a TRPM trap. Typically used when leaving
2220	* AMD-V to execute any instruction.
2221	*
2222	* @param pvCpu Pointer to the VMCPU.
2223	*/
2224	static void hmR0SvmPendingEventToTrpmTrap(PVMCPU pVCpu)
2225	{
2226	Assert(pVCpu->hm.s.Event.fPending);
2227	Assert(TRPMQueryTrap(pVCpu, NULL /* pu8TrapNo /, NULL / pEnmType */) == VERR_TRPM_NO_ACTIVE_TRAP);
2228
2229	SVMEVENT Event;
2230	Event.u = pVCpu->hm.s.Event.u64IntrInfo;
2231
2232	uint8_t uVector = Event.n.u8Vector;
2233	uint8_t uVectorType = Event.n.u3Type;
2234
2235	TRPMEVENT enmTrapType;
2236	switch (uVectorType)
2237	{
2238	case SVM_EVENT_EXTERNAL_IRQ:
2239	case SVM_EVENT_NMI:
2240	enmTrapType = TRPM_HARDWARE_INT;
2241	break;
2242	case SVM_EVENT_SOFTWARE_INT:
2243	enmTrapType = TRPM_SOFTWARE_INT;
2244	break;
2245	case SVM_EVENT_EXCEPTION:
2246	enmTrapType = TRPM_TRAP;
2247	break;
2248	default:
2249	AssertMsgFailed(("Invalid pending-event type %#x\n", uVectorType));
2250	enmTrapType = TRPM_32BIT_HACK;
2251	break;
2252	}
2253
2254	Log4(("HM event->TRPM: uVector=%#x enmTrapType=%d\n", uVector, uVectorType));
2255
2256	int rc = TRPMAssertTrap(pVCpu, uVector, enmTrapType);
2257	AssertRC(rc);
2258
2259	if (Event.n.u1ErrorCodeValid)
2260	TRPMSetErrorCode(pVCpu, Event.n.u32ErrorCode);
2261
2262	if ( uVectorType == SVM_EVENT_EXCEPTION
2263	&& uVector == X86_XCPT_PF)
2264	{
2265	TRPMSetFaultAddress(pVCpu, pVCpu->hm.s.Event.GCPtrFaultAddress);
2266	Assert(pVCpu->hm.s.Event.GCPtrFaultAddress == CPUMGetGuestCR2(pVCpu));
2267	}
2268	else if (uVectorType == SVM_EVENT_SOFTWARE_INT)
2269	{
2270	AssertMsg( uVectorType == SVM_EVENT_SOFTWARE_INT
2271	\|\| (uVector == X86_XCPT_BP \|\| uVector == X86_XCPT_OF),
2272	("Invalid vector: uVector=%#x uVectorType=%#x\n", uVector, uVectorType));
2273	TRPMSetInstrLength(pVCpu, pVCpu->hm.s.Event.cbInstr);
2274	}
2275	pVCpu->hm.s.Event.fPending = false;
2276	}
2277
2278
2279	/**
2280	* Gets the guest's interrupt-shadow.
2281	*
2282	* @returns The guest's interrupt-shadow.
2283	* @param pVCpu Pointer to the VMCPU.
2284	* @param pCtx Pointer to the guest-CPU context.
2285	*
2286	* @remarks No-long-jump zone!!!
2287	* @remarks Has side-effects with VMCPU_FF_INHIBIT_INTERRUPTS force-flag.
2288	*/
2289	DECLINLINE(uint32_t) hmR0SvmGetGuestIntrShadow(PVMCPU pVCpu, PCPUMCTX pCtx)
2290	{
2291	/*
2292	* Instructions like STI and MOV SS inhibit interrupts till the next instruction completes. Check if we should
2293	* inhibit interrupts or clear any existing interrupt-inhibition.
2294	*/
2295	uint32_t uIntrState = 0;
2296	if (VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS))
2297	{
2298	if (pCtx->rip != EMGetInhibitInterruptsPC(pVCpu))
2299	{
2300	/*
2301	* We can clear the inhibit force flag as even if we go back to the recompiler without executing guest code in
2302	* AMD-V, the flag's condition to be cleared is met and thus the cleared state is correct.
2303	*/
2304	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS);
2305	}
2306	else
2307	uIntrState = SVM_INTERRUPT_SHADOW_ACTIVE;
2308	}
2309	return uIntrState;
2310	}
2311
2312
2313	/**
2314	* Sets the virtual interrupt intercept control in the VMCB which
2315	* instructs AMD-V to cause a #VMEXIT as soon as the guest is in a state to
2316	* receive interrupts.
2317	*
2318	* @param pVmcb Pointer to the VMCB.
2319	*/
2320	DECLINLINE(void) hmR0SvmSetVirtIntrIntercept(PSVMVMCB pVmcb)
2321	{
2322	if (!(pVmcb->ctrl.u32InterceptCtrl1 & SVM_CTRL1_INTERCEPT_VINTR))
2323	{
2324	pVmcb->ctrl.IntCtrl.n.u1VIrqValid = 1; /* A virtual interrupt is pending. */
2325	pVmcb->ctrl.IntCtrl.n.u8VIrqVector = 0; /* Not necessary as we #VMEXIT for delivering the interrupt. */
2326	pVmcb->ctrl.u32InterceptCtrl1 \|= SVM_CTRL1_INTERCEPT_VINTR;
2327	pVmcb->ctrl.u64VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS \| HMSVM_VMCB_CLEAN_TPR);
2328
2329	Log4(("Setting VINTR intercept\n"));
2330	}
2331	}
2332
2333
2334	/**
2335	* Evaluates the event to be delivered to the guest and sets it as the pending
2336	* event.
2337	*
2338	* @param pVCpu Pointer to the VMCPU.
2339	* @param pCtx Pointer to the guest-CPU context.
2340	*/
2341	static void hmR0SvmEvaluatePendingEvent(PVMCPU pVCpu, PCPUMCTX pCtx)
2342	{
2343	Assert(!pVCpu->hm.s.Event.fPending);
2344	Log4Func(("\n"));
2345
2346	const bool fIntShadow = !!hmR0SvmGetGuestIntrShadow(pVCpu, pCtx);
2347	const bool fBlockInt = !(pCtx->eflags.u32 & X86_EFL_IF);
2348	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2349
2350	SVMEVENT Event;
2351	Event.u = 0;
2352	/** @todo SMI. SMIs take priority over NMIs. */
2353	if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_NMI)) /* NMI. NMIs take priority over regular interrupts . */
2354	{
2355	if (!fIntShadow)
2356	{
2357	Log4(("Pending NMI\n"));
2358
2359	Event.n.u1Valid = 1;
2360	Event.n.u8Vector = X86_XCPT_NMI;
2361	Event.n.u3Type = SVM_EVENT_NMI;
2362
2363	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
2364	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INTERRUPT_NMI);
2365	}
2366	else
2367	hmR0SvmSetVirtIntrIntercept(pVmcb);
2368	}
2369	else if (VMCPU_FF_IS_PENDING(pVCpu, (VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC)))
2370	{
2371	/*
2372	* Check if the guest can receive external interrupts (PIC/APIC). Once we do PDMGetInterrupt() we -must- deliver
2373	* the interrupt ASAP. We must not execute any guest code until we inject the interrupt which is why it is
2374	* evaluated here and not set as pending, solely based on the force-flags.
2375	*/
2376	if ( !fBlockInt
2377	&& !fIntShadow)
2378	{
2379	uint8_t u8Interrupt;
2380	int rc = PDMGetInterrupt(pVCpu, &u8Interrupt);
2381	if (RT_SUCCESS(rc))
2382	{
2383	Log4(("Injecting external interrupt u8Interrupt=%#x\n", u8Interrupt));
2384
2385	Event.n.u1Valid = 1;
2386	Event.n.u8Vector = u8Interrupt;
2387	Event.n.u3Type = SVM_EVENT_EXTERNAL_IRQ;
2388
2389	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
2390	}
2391	else
2392	{
2393	/** @todo Does this actually happen? If not turn it into an assertion. */
2394	Assert(!VMCPU_FF_IS_PENDING(pVCpu, (VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC)));
2395	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchGuestIrq);
2396	}
2397	}
2398	else
2399	hmR0SvmSetVirtIntrIntercept(pVmcb);
2400	}
2401	}
2402
2403
2404	/**
2405	* Injects any pending events into the guest if the guest is in a state to
2406	* receive them.
2407	*
2408	* @param pVCpu Pointer to the VMCPU.
2409	* @param pCtx Pointer to the guest-CPU context.
2410	*/
2411	static void hmR0SvmInjectPendingEvent(PVMCPU pVCpu, PCPUMCTX pCtx)
2412	{
2413	Assert(!TRPMHasTrap(pVCpu));
2414	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2415	Log4Func(("\n"));
2416
2417	const bool fIntShadow = !!hmR0SvmGetGuestIntrShadow(pVCpu, pCtx);
2418	const bool fBlockInt = !(pCtx->eflags.u32 & X86_EFL_IF);
2419	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2420
2421	if (pVCpu->hm.s.Event.fPending) /* First, inject any pending HM events. */
2422	{
2423	SVMEVENT Event;
2424	Event.u = pVCpu->hm.s.Event.u64IntrInfo;
2425	Assert(Event.n.u1Valid);
2426	#ifdef VBOX_STRICT
2427	if (Event.n.u3Type == SVM_EVENT_EXTERNAL_IRQ)
2428	{
2429	Assert(!fBlockInt);
2430	Assert(!fIntShadow);
2431	}
2432	else if (Event.n.u3Type == SVM_EVENT_NMI)
2433	Assert(!fIntShadow);
2434	#endif
2435
2436	Log4(("Injecting pending HM event.\n"));
2437	hmR0SvmInjectEventVmcb(pVCpu, pVmcb, pCtx, &Event);
2438	pVCpu->hm.s.Event.fPending = false;
2439
2440	#ifdef VBOX_WITH_STATISTICS
2441	if (Event.n.u3Type == SVM_EVENT_EXTERNAL_IRQ)
2442	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectInterrupt);
2443	else
2444	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectXcpt);
2445	#endif
2446	}
2447
2448	/* Update the guest interrupt shadow in the VMCB. */
2449	pVmcb->ctrl.u64IntShadow = !!fIntShadow;
2450	}
2451
2452
2453	/**
2454	* Reports world-switch error and dumps some useful debug info.
2455	*
2456	* @param pVM Pointer to the VM.
2457	* @param pVCpu Pointer to the VMCPU.
2458	* @param rcVMRun The return code from VMRUN (or
2459	* VERR_SVM_INVALID_GUEST_STATE for invalid
2460	* guest-state).
2461	* @param pCtx Pointer to the guest-CPU context.
2462	*/
2463	static void hmR0SvmReportWorldSwitchError(PVM pVM, PVMCPU pVCpu, int rcVMRun, PCPUMCTX pCtx)
2464	{
2465	HMSVM_ASSERT_PREEMPT_SAFE();
2466	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2467
2468	if (rcVMRun == VERR_SVM_INVALID_GUEST_STATE)
2469	{
2470	HMDumpRegs(pVM, pVCpu, pCtx);
2471	#ifdef VBOX_STRICT
2472	Log4(("ctrl.u64VmcbCleanBits %#RX64\n", pVmcb->ctrl.u64VmcbCleanBits));
2473	Log4(("ctrl.u16InterceptRdCRx %#x\n", pVmcb->ctrl.u16InterceptRdCRx));
2474	Log4(("ctrl.u16InterceptWrCRx %#x\n", pVmcb->ctrl.u16InterceptWrCRx));
2475	Log4(("ctrl.u16InterceptRdDRx %#x\n", pVmcb->ctrl.u16InterceptRdDRx));
2476	Log4(("ctrl.u16InterceptWrDRx %#x\n", pVmcb->ctrl.u16InterceptWrDRx));
2477	Log4(("ctrl.u32InterceptException %#x\n", pVmcb->ctrl.u32InterceptException));
2478	Log4(("ctrl.u32InterceptCtrl1 %#x\n", pVmcb->ctrl.u32InterceptCtrl1));
2479	Log4(("ctrl.u32InterceptCtrl2 %#x\n", pVmcb->ctrl.u32InterceptCtrl2));
2480	Log4(("ctrl.u64IOPMPhysAddr %#RX64\n", pVmcb->ctrl.u64IOPMPhysAddr));
2481	Log4(("ctrl.u64MSRPMPhysAddr %#RX64\n", pVmcb->ctrl.u64MSRPMPhysAddr));
2482	Log4(("ctrl.u64TSCOffset %#RX64\n", pVmcb->ctrl.u64TSCOffset));
2483
2484	Log4(("ctrl.TLBCtrl.u32ASID %#x\n", pVmcb->ctrl.TLBCtrl.n.u32ASID));
2485	Log4(("ctrl.TLBCtrl.u8TLBFlush %#x\n", pVmcb->ctrl.TLBCtrl.n.u8TLBFlush));
2486	Log4(("ctrl.TLBCtrl.u24Reserved %#x\n", pVmcb->ctrl.TLBCtrl.n.u24Reserved));
2487
2488	Log4(("ctrl.IntCtrl.u8VTPR %#x\n", pVmcb->ctrl.IntCtrl.n.u8VTPR));
2489	Log4(("ctrl.IntCtrl.u1VIrqValid %#x\n", pVmcb->ctrl.IntCtrl.n.u1VIrqValid));
2490	Log4(("ctrl.IntCtrl.u7Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u7Reserved));
2491	Log4(("ctrl.IntCtrl.u4VIrqPriority %#x\n", pVmcb->ctrl.IntCtrl.n.u4VIrqPriority));
2492	Log4(("ctrl.IntCtrl.u1IgnoreTPR %#x\n", pVmcb->ctrl.IntCtrl.n.u1IgnoreTPR));
2493	Log4(("ctrl.IntCtrl.u3Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u3Reserved));
2494	Log4(("ctrl.IntCtrl.u1VIrqMasking %#x\n", pVmcb->ctrl.IntCtrl.n.u1VIrqMasking));
2495	Log4(("ctrl.IntCtrl.u6Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u6Reserved));
2496	Log4(("ctrl.IntCtrl.u8VIrqVector %#x\n", pVmcb->ctrl.IntCtrl.n.u8VIrqVector));
2497	Log4(("ctrl.IntCtrl.u24Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u24Reserved));
2498
2499	Log4(("ctrl.u64IntShadow %#RX64\n", pVmcb->ctrl.u64IntShadow));
2500	Log4(("ctrl.u64ExitCode %#RX64\n", pVmcb->ctrl.u64ExitCode));
2501	Log4(("ctrl.u64ExitInfo1 %#RX64\n", pVmcb->ctrl.u64ExitInfo1));
2502	Log4(("ctrl.u64ExitInfo2 %#RX64\n", pVmcb->ctrl.u64ExitInfo2));
2503	Log4(("ctrl.ExitIntInfo.u8Vector %#x\n", pVmcb->ctrl.ExitIntInfo.n.u8Vector));
2504	Log4(("ctrl.ExitIntInfo.u3Type %#x\n", pVmcb->ctrl.ExitIntInfo.n.u3Type));
2505	Log4(("ctrl.ExitIntInfo.u1ErrorCodeValid %#x\n", pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid));
2506	Log4(("ctrl.ExitIntInfo.u19Reserved %#x\n", pVmcb->ctrl.ExitIntInfo.n.u19Reserved));
2507	Log4(("ctrl.ExitIntInfo.u1Valid %#x\n", pVmcb->ctrl.ExitIntInfo.n.u1Valid));
2508	Log4(("ctrl.ExitIntInfo.u32ErrorCode %#x\n", pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode));
2509	Log4(("ctrl.NestedPaging %#RX64\n", pVmcb->ctrl.NestedPaging.u));
2510	Log4(("ctrl.EventInject.u8Vector %#x\n", pVmcb->ctrl.EventInject.n.u8Vector));
2511	Log4(("ctrl.EventInject.u3Type %#x\n", pVmcb->ctrl.EventInject.n.u3Type));
2512	Log4(("ctrl.EventInject.u1ErrorCodeValid %#x\n", pVmcb->ctrl.EventInject.n.u1ErrorCodeValid));
2513	Log4(("ctrl.EventInject.u19Reserved %#x\n", pVmcb->ctrl.EventInject.n.u19Reserved));
2514	Log4(("ctrl.EventInject.u1Valid %#x\n", pVmcb->ctrl.EventInject.n.u1Valid));
2515	Log4(("ctrl.EventInject.u32ErrorCode %#x\n", pVmcb->ctrl.EventInject.n.u32ErrorCode));
2516
2517	Log4(("ctrl.u64NestedPagingCR3 %#RX64\n", pVmcb->ctrl.u64NestedPagingCR3));
2518	Log4(("ctrl.u64LBRVirt %#RX64\n", pVmcb->ctrl.u64LBRVirt));
2519
2520	Log4(("guest.CS.u16Sel %RTsel\n", pVmcb->guest.CS.u16Sel));
2521	Log4(("guest.CS.u16Attr %#x\n", pVmcb->guest.CS.u16Attr));
2522	Log4(("guest.CS.u32Limit %#RX32\n", pVmcb->guest.CS.u32Limit));
2523	Log4(("guest.CS.u64Base %#RX64\n", pVmcb->guest.CS.u64Base));
2524	Log4(("guest.DS.u16Sel %#RTsel\n", pVmcb->guest.DS.u16Sel));
2525	Log4(("guest.DS.u16Attr %#x\n", pVmcb->guest.DS.u16Attr));
2526	Log4(("guest.DS.u32Limit %#RX32\n", pVmcb->guest.DS.u32Limit));
2527	Log4(("guest.DS.u64Base %#RX64\n", pVmcb->guest.DS.u64Base));
2528	Log4(("guest.ES.u16Sel %RTsel\n", pVmcb->guest.ES.u16Sel));
2529	Log4(("guest.ES.u16Attr %#x\n", pVmcb->guest.ES.u16Attr));
2530	Log4(("guest.ES.u32Limit %#RX32\n", pVmcb->guest.ES.u32Limit));
2531	Log4(("guest.ES.u64Base %#RX64\n", pVmcb->guest.ES.u64Base));
2532	Log4(("guest.FS.u16Sel %RTsel\n", pVmcb->guest.FS.u16Sel));
2533	Log4(("guest.FS.u16Attr %#x\n", pVmcb->guest.FS.u16Attr));
2534	Log4(("guest.FS.u32Limit %#RX32\n", pVmcb->guest.FS.u32Limit));
2535	Log4(("guest.FS.u64Base %#RX64\n", pVmcb->guest.FS.u64Base));
2536	Log4(("guest.GS.u16Sel %RTsel\n", pVmcb->guest.GS.u16Sel));
2537	Log4(("guest.GS.u16Attr %#x\n", pVmcb->guest.GS.u16Attr));
2538	Log4(("guest.GS.u32Limit %#RX32\n", pVmcb->guest.GS.u32Limit));
2539	Log4(("guest.GS.u64Base %#RX64\n", pVmcb->guest.GS.u64Base));
2540
2541	Log4(("guest.GDTR.u32Limit %#RX32\n", pVmcb->guest.GDTR.u32Limit));
2542	Log4(("guest.GDTR.u64Base %#RX64\n", pVmcb->guest.GDTR.u64Base));
2543
2544	Log4(("guest.LDTR.u16Sel %RTsel\n", pVmcb->guest.LDTR.u16Sel));
2545	Log4(("guest.LDTR.u16Attr %#x\n", pVmcb->guest.LDTR.u16Attr));
2546	Log4(("guest.LDTR.u32Limit %#RX32\n", pVmcb->guest.LDTR.u32Limit));
2547	Log4(("guest.LDTR.u64Base %#RX64\n", pVmcb->guest.LDTR.u64Base));
2548
2549	Log4(("guest.IDTR.u32Limit %#RX32\n", pVmcb->guest.IDTR.u32Limit));
2550	Log4(("guest.IDTR.u64Base %#RX64\n", pVmcb->guest.IDTR.u64Base));
2551
2552	Log4(("guest.TR.u16Sel %RTsel\n", pVmcb->guest.TR.u16Sel));
2553	Log4(("guest.TR.u16Attr %#x\n", pVmcb->guest.TR.u16Attr));
2554	Log4(("guest.TR.u32Limit %#RX32\n", pVmcb->guest.TR.u32Limit));
2555	Log4(("guest.TR.u64Base %#RX64\n", pVmcb->guest.TR.u64Base));
2556
2557	Log4(("guest.u8CPL %#x\n", pVmcb->guest.u8CPL));
2558	Log4(("guest.u64CR0 %#RX64\n", pVmcb->guest.u64CR0));
2559	Log4(("guest.u64CR2 %#RX64\n", pVmcb->guest.u64CR2));
2560	Log4(("guest.u64CR3 %#RX64\n", pVmcb->guest.u64CR3));
2561	Log4(("guest.u64CR4 %#RX64\n", pVmcb->guest.u64CR4));
2562	Log4(("guest.u64DR6 %#RX64\n", pVmcb->guest.u64DR6));
2563	Log4(("guest.u64DR7 %#RX64\n", pVmcb->guest.u64DR7));
2564
2565	Log4(("guest.u64RIP %#RX64\n", pVmcb->guest.u64RIP));
2566	Log4(("guest.u64RSP %#RX64\n", pVmcb->guest.u64RSP));
2567	Log4(("guest.u64RAX %#RX64\n", pVmcb->guest.u64RAX));
2568	Log4(("guest.u64RFlags %#RX64\n", pVmcb->guest.u64RFlags));
2569
2570	Log4(("guest.u64SysEnterCS %#RX64\n", pVmcb->guest.u64SysEnterCS));
2571	Log4(("guest.u64SysEnterEIP %#RX64\n", pVmcb->guest.u64SysEnterEIP));
2572	Log4(("guest.u64SysEnterESP %#RX64\n", pVmcb->guest.u64SysEnterESP));
2573
2574	Log4(("guest.u64EFER %#RX64\n", pVmcb->guest.u64EFER));
2575	Log4(("guest.u64STAR %#RX64\n", pVmcb->guest.u64STAR));
2576	Log4(("guest.u64LSTAR %#RX64\n", pVmcb->guest.u64LSTAR));
2577	Log4(("guest.u64CSTAR %#RX64\n", pVmcb->guest.u64CSTAR));
2578	Log4(("guest.u64SFMASK %#RX64\n", pVmcb->guest.u64SFMASK));
2579	Log4(("guest.u64KernelGSBase %#RX64\n", pVmcb->guest.u64KernelGSBase));
2580	Log4(("guest.u64GPAT %#RX64\n", pVmcb->guest.u64GPAT));
2581	Log4(("guest.u64DBGCTL %#RX64\n", pVmcb->guest.u64DBGCTL));
2582	Log4(("guest.u64BR_FROM %#RX64\n", pVmcb->guest.u64BR_FROM));
2583	Log4(("guest.u64BR_TO %#RX64\n", pVmcb->guest.u64BR_TO));
2584	Log4(("guest.u64LASTEXCPFROM %#RX64\n", pVmcb->guest.u64LASTEXCPFROM));
2585	Log4(("guest.u64LASTEXCPTO %#RX64\n", pVmcb->guest.u64LASTEXCPTO));
2586	#endif
2587	}
2588	else
2589	Log4(("hmR0SvmReportWorldSwitchError: rcVMRun=%d\n", rcVMRun));
2590	}
2591
2592
2593	/**
2594	* Check per-VM and per-VCPU force flag actions that require us to go back to
2595	* ring-3 for one reason or another.
2596	*
2597	* @returns VBox status code (information status code included).
2598	* @retval VINF_SUCCESS if we don't have any actions that require going back to
2599	* ring-3.
2600	* @retval VINF_PGM_SYNC_CR3 if we have pending PGM CR3 sync.
2601	* @retval VINF_EM_PENDING_REQUEST if we have pending requests (like hardware
2602	* interrupts)
2603	* @retval VINF_PGM_POOL_FLUSH_PENDING if PGM is doing a pool flush and requires
2604	* all EMTs to be in ring-3.
2605	* @retval VINF_EM_RAW_TO_R3 if there is pending DMA requests.
2606	* @retval VINF_EM_NO_MEMORY PGM is out of memory, we need to return
2607	* to the EM loop.
2608	*
2609	* @param pVM Pointer to the VM.
2610	* @param pVCpu Pointer to the VMCPU.
2611	* @param pCtx Pointer to the guest-CPU context.
2612	*/
2613	static int hmR0SvmCheckForceFlags(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
2614	{
2615	Assert(VMMRZCallRing3IsEnabled(pVCpu));
2616
2617	/* On AMD-V we don't need to update CR3, PAE PDPES lazily. See hmR0SvmSaveGuestState(). */
2618	Assert(!VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_UPDATE_CR3));
2619	Assert(!VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_UPDATE_PAE_PDPES));
2620
2621	if ( VM_FF_IS_PENDING(pVM, !pVCpu->hm.s.fSingleInstruction
2622	? VM_FF_HP_R0_PRE_HM_MASK : VM_FF_HP_R0_PRE_HM_STEP_MASK)
2623	\|\| VMCPU_FF_IS_PENDING(pVCpu, !pVCpu->hm.s.fSingleInstruction
2624	? VMCPU_FF_HP_R0_PRE_HM_MASK : VMCPU_FF_HP_R0_PRE_HM_STEP_MASK) )
2625	{
2626	/* Pending PGM C3 sync. */
2627	if (VMCPU_FF_IS_PENDING(pVCpu,VMCPU_FF_PGM_SYNC_CR3 \| VMCPU_FF_PGM_SYNC_CR3_NON_GLOBAL))
2628	{
2629	int rc = PGMSyncCR3(pVCpu, pCtx->cr0, pCtx->cr3, pCtx->cr4, VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_PGM_SYNC_CR3));
2630	if (rc != VINF_SUCCESS)
2631	{
2632	Log4(("hmR0SvmCheckForceFlags: PGMSyncCR3 forcing us back to ring-3. rc=%d\n", rc));
2633	return rc;
2634	}
2635	}
2636
2637	/* Pending HM-to-R3 operations (critsects, timers, EMT rendezvous etc.) */
2638	/* -XXX- what was that about single stepping? */
2639	if ( VM_FF_IS_PENDING(pVM, VM_FF_HM_TO_R3_MASK)
2640	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_TO_R3_MASK))
2641	{
2642	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHmToR3FF);
2643	int rc = RT_UNLIKELY(VM_FF_IS_PENDING(pVM, VM_FF_PGM_NO_MEMORY)) ? VINF_EM_NO_MEMORY : VINF_EM_RAW_TO_R3;
2644	Log4(("hmR0SvmCheckForceFlags: HM_TO_R3 forcing us back to ring-3. rc=%d\n", rc));
2645	return rc;
2646	}
2647
2648	/* Pending VM request packets, such as hardware interrupts. */
2649	if ( VM_FF_IS_PENDING(pVM, VM_FF_REQUEST)
2650	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_REQUEST))
2651	{
2652	Log4(("hmR0SvmCheckForceFlags: Pending VM request forcing us back to ring-3\n"));
2653	return VINF_EM_PENDING_REQUEST;
2654	}
2655
2656	/* Pending PGM pool flushes. */
2657	if (VM_FF_IS_PENDING(pVM, VM_FF_PGM_POOL_FLUSH_PENDING))
2658	{
2659	Log4(("hmR0SvmCheckForceFlags: PGM pool flush pending forcing us back to ring-3\n"));
2660	return VINF_PGM_POOL_FLUSH_PENDING;
2661	}
2662
2663	/* Pending DMA requests. */
2664	if (VM_FF_IS_PENDING(pVM, VM_FF_PDM_DMA))
2665	{
2666	Log4(("hmR0SvmCheckForceFlags: Pending DMA request forcing us back to ring-3\n"));
2667	return VINF_EM_RAW_TO_R3;
2668	}
2669	}
2670
2671	return VINF_SUCCESS;
2672	}
2673
2674
2675	/**
2676	* Does the preparations before executing guest code in AMD-V.
2677	*
2678	* This may cause longjmps to ring-3 and may even result in rescheduling to the
2679	* recompiler. We must be cautious what we do here regarding committing
2680	* guest-state information into the the VMCB assuming we assuredly execute the
2681	* guest in AMD-V. If we fall back to the recompiler after updating the VMCB and
2682	* clearing the common-state (TRPM/forceflags), we must undo those changes so
2683	* that the recompiler can (and should) use them when it resumes guest
2684	* execution. Otherwise such operations must be done when we can no longer
2685	* exit to ring-3.
2686	*
2687	* @returns VBox status code (informational status codes included).
2688	* @retval VINF_SUCCESS if we can proceed with running the guest.
2689	* @retval VINF_* scheduling changes, we have to go back to ring-3.
2690	*
2691	* @param pVM Pointer to the VM.
2692	* @param pVCpu Pointer to the VMCPU.
2693	* @param pCtx Pointer to the guest-CPU context.
2694	* @param pSvmTransient Pointer to the SVM transient structure.
2695	*/
2696	DECLINLINE(int) hmR0SvmPreRunGuest(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
2697	{
2698	/* Check force flag actions that might require us to go back to ring-3. */
2699	int rc = hmR0SvmCheckForceFlags(pVM, pVCpu, pCtx);
2700	if (rc != VINF_SUCCESS)
2701	return rc;
2702
2703	if (TRPMHasTrap(pVCpu))
2704	hmR0SvmTrpmTrapToPendingEvent(pVCpu);
2705	else if (!pVCpu->hm.s.Event.fPending)
2706	hmR0SvmEvaluatePendingEvent(pVCpu, pCtx);
2707
2708	/*
2709	* We disable interrupts so that we don't miss any interrupts that would flag preemption (IPI/timers etc.)
2710	* when thread-context hooks aren't used and we've been running with preemption disabled for a while.
2711	*
2712	* We need to check for force-flags that could've possible been altered since we last checked them (e.g.
2713	* by PDMGetInterrupt() leaving the PDM critical section, see @bugref{6398}).
2714	*
2715	* We also check a couple of other force-flags as a last opportunity to get the EMT back to ring-3 before
2716	* executing guest code.
2717	*/
2718	pSvmTransient->uEflags = ASMIntDisableFlags();
2719	if ( VM_FF_IS_PENDING(pVM, VM_FF_EMT_RENDEZVOUS \| VM_FF_TM_VIRTUAL_SYNC)
2720	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_TO_R3_MASK))
2721	{
2722	ASMSetFlags(pSvmTransient->uEflags);
2723	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHmToR3FF);
2724	return VINF_EM_RAW_TO_R3;
2725	}
2726	else if (RTThreadPreemptIsPending(NIL_RTTHREAD))
2727	{
2728	ASMSetFlags(pSvmTransient->uEflags);
2729	STAM_COUNTER_INC(&pVCpu->hm.s.StatPendingHostIrq);
2730	return VINF_EM_RAW_INTERRUPT;
2731	}
2732
2733	/* Indicate the start of guest execution. No more longjmps or returns to ring-3 from this point!!! */
2734	VMCPU_ASSERT_STATE(pVCpu, VMCPUSTATE_STARTED_HM);
2735	VMCPU_SET_STATE(pVCpu, VMCPUSTATE_STARTED_EXEC);
2736
2737	return VINF_SUCCESS;
2738	}
2739
2740
2741	/**
2742	* Prepares to run guest code in AMD-V and we've committed to doing so. This
2743	* means there is no backing out to ring-3 or anywhere else at this
2744	* point.
2745	*
2746	* @param pVM Pointer to the VM.
2747	* @param pVCpu Pointer to the VMCPU.
2748	* @param pCtx Pointer to the guest-CPU context.
2749	* @param pSvmTransient Pointer to the SVM transient structure.
2750	*
2751	* @remarks Called with preemption disabled.
2752	* @remarks No-long-jump zone!!!
2753	*/
2754	DECLINLINE(void) hmR0SvmPreRunGuestCommitted(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
2755	{
2756	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2757	Assert(VMMR0IsLogFlushDisabled(pVCpu));
2758
2759	hmR0SvmInjectPendingEvent(pVCpu, pCtx);
2760
2761	/*
2762	* Re-enable nested paging (automatically disabled on every VM-exit). See AMD spec. 15.25.3 "Enabling Nested Paging".
2763	* We avoid changing the corresponding VMCB Clean Bit as we're not changing it to a different value since the previous run.
2764	*/
2765	/** @todo The above assumption could be wrong. It's not documented what
2766	* should be done wrt to the VMCB Clean Bit, but we'll find out the
2767	* hard way. */
2768	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2769	pVmcb->ctrl.NestedPaging.n.u1NestedPaging = pVM->hm.s.fNestedPaging;
2770
2771	#ifdef HMSVM_SYNC_FULL_GUEST_STATE
2772	pVCpu->hm.s.fContextUseFlags \|= HM_CHANGED_ALL_GUEST;
2773	#endif
2774
2775	/* Load the guest state. */
2776	int rc = hmR0SvmLoadGuestState(pVM, pVCpu, pCtx);
2777	AssertRC(rc);
2778	pVCpu->hm.s.fContextUseFlags &= ~HM_CHANGED_HOST_CONTEXT; /* Preemption might set this, nothing to do on AMD-V. */
2779	AssertMsg(!pVCpu->hm.s.fContextUseFlags, ("fContextUseFlags =%#x\n", pVCpu->hm.s.fContextUseFlags));
2780	STAM_COUNTER_INC(&pVCpu->hm.s.StatLoadFull);
2781
2782	/* If VMCB Clean Bits isn't supported by the CPU, simply mark all state-bits as dirty, indicating (re)load-from-VMCB. */
2783	if (!(pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_VMCB_CLEAN))
2784	pVmcb->ctrl.u64VmcbCleanBits = 0;
2785
2786	/*
2787	* If we're not intercepting TPR changes in the guest, save the guest TPR before the world-switch
2788	* so we can update it on the way back if the guest changed the TPR.
2789	*/
2790	if (pVCpu->hm.s.svm.fSyncVTpr)
2791	{
2792	if (pVM->hm.s.fTPRPatchingActive)
2793	pSvmTransient->u8GuestTpr = pCtx->msrLSTAR;
2794	else
2795	pSvmTransient->u8GuestTpr = pVmcb->ctrl.IntCtrl.n.u8VTPR;
2796	}
2797
2798	/* Setup TSC offsetting. */
2799	if ( pSvmTransient->fUpdateTscOffsetting
2800	\|\| HMR0GetCurrentCpu()->idCpu != pVCpu->hm.s.idLastCpu)
2801	{
2802	hmR0SvmUpdateTscOffsetting(pVCpu);
2803	pSvmTransient->fUpdateTscOffsetting = false;
2804	}
2805
2806	/* Flush the appropriate tagged-TLB entries. */
2807	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, true); /* Used for TLB-shootdowns, set this across the world switch. */
2808	hmR0SvmFlushTaggedTlb(pVCpu);
2809	Assert(HMR0GetCurrentCpu()->idCpu == pVCpu->hm.s.idLastCpu);
2810
2811	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatEntry, &pVCpu->hm.s.StatInGC, x);
2812
2813	TMNotifyStartOfExecution(pVCpu); /* Finally, notify TM to resume its clocks as we're about
2814	to start executing. */
2815
2816	/*
2817	* Save the current Host TSC_AUX and write the guest TSC_AUX to the host, so that
2818	* RDTSCPs (that don't cause exits) reads the guest MSR. See @bugref{3324}.
2819	*
2820	* This should be done -after- any RDTSCPs for obtaining the host timestamp (TM, STAM etc).
2821	*/
2822	pSvmTransient->fRestoreTscAuxMsr = false;
2823	if ( (pVM->hm.s.cpuid.u32AMDFeatureEDX & X86_CPUID_EXT_FEATURE_EDX_RDTSCP)
2824	&& !(pVmcb->ctrl.u32InterceptCtrl2 & SVM_CTRL2_INTERCEPT_RDTSCP))
2825	{
2826	pVCpu->hm.s.u64HostTscAux = ASMRdMsr(MSR_K8_TSC_AUX);
2827	uint64_t u64GuestTscAux = 0;
2828	int rc2 = CPUMQueryGuestMsr(pVCpu, MSR_K8_TSC_AUX, &u64GuestTscAux);
2829	AssertRC(rc2);
2830	if (u64GuestTscAux != pVCpu->hm.s.u64HostTscAux)
2831	{
2832	ASMWrMsr(MSR_K8_TSC_AUX, u64GuestTscAux);
2833	pSvmTransient->fRestoreTscAuxMsr = true;
2834	}
2835	}
2836	}
2837
2838
2839	/**
2840	* Wrapper for running the guest code in AMD-V.
2841	*
2842	* @returns VBox strict status code.
2843	* @param pVM Pointer to the VM.
2844	* @param pVCpu Pointer to the VMCPU.
2845	* @param pCtx Pointer to the guest-CPU context.
2846	*
2847	* @remarks No-long-jump zone!!!
2848	*/
2849	DECLINLINE(int) hmR0SvmRunGuest(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
2850	{
2851	/*
2852	* 64-bit Windows uses XMM registers in the kernel as the Microsoft compiler expresses floating-point operations
2853	* using SSE instructions. Some XMM registers (XMM6-XMM15) are callee-saved and thus the need for this XMM wrapper.
2854	* Refer MSDN docs. "Configuring Programs for 64-bit / x64 Software Conventions / Register Usage" for details.
2855	*/
2856	#ifdef VBOX_WITH_KERNEL_USING_XMM
2857	return HMR0SVMRunWrapXMM(pVCpu->hm.s.svm.HCPhysVmcbHost, pVCpu->hm.s.svm.HCPhysVmcb, pCtx, pVM, pVCpu,
2858	pVCpu->hm.s.svm.pfnVMRun);
2859	#else
2860	return pVCpu->hm.s.svm.pfnVMRun(pVCpu->hm.s.svm.HCPhysVmcbHost, pVCpu->hm.s.svm.HCPhysVmcb, pCtx, pVM, pVCpu);
2861	#endif
2862	}
2863
2864
2865	/**
2866	* Performs some essential restoration of state after running guest code in
2867	* AMD-V.
2868	*
2869	* @param pVM Pointer to the VM.
2870	* @param pVCpu Pointer to the VMCPU.
2871	* @param pMixedCtx Pointer to the guest-CPU context. The data maybe
2872	* out-of-sync. Make sure to update the required fields
2873	* before using them.
2874	* @param pSvmTransient Pointer to the SVM transient structure.
2875	* @param rcVMRun Return code of VMRUN.
2876	*
2877	* @remarks Called with interrupts disabled.
2878	* @remarks No-long-jump zone!!! This function will however re-enable longjmps
2879	* unconditionally when it is safe to do so.
2880	*/
2881	DECLINLINE(void) hmR0SvmPostRunGuest(PVM pVM, PVMCPU pVCpu, PCPUMCTX pMixedCtx, PSVMTRANSIENT pSvmTransient, int rcVMRun)
2882	{
2883	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2884
2885	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, false); /* See HMInvalidatePageOnAllVCpus(): used for TLB-shootdowns. */
2886	ASMAtomicIncU32(&pVCpu->hm.s.cWorldSwitchExits); /* Initialized in vmR3CreateUVM(): used for TLB-shootdowns. */
2887
2888	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2889	pVmcb->ctrl.u64VmcbCleanBits = HMSVM_VMCB_CLEAN_ALL; /* Mark the VMCB-state cache as unmodified by VMM. */
2890
2891	if (pSvmTransient->fRestoreTscAuxMsr)
2892	ASMWrMsr(MSR_K8_TSC_AUX, pVCpu->hm.s.u64HostTscAux);
2893
2894	if (!(pVmcb->ctrl.u32InterceptCtrl1 & SVM_CTRL1_INTERCEPT_RDTSC))
2895	{
2896	/** @todo Find a way to fix hardcoding a guestimate. */
2897	TMCpuTickSetLastSeen(pVCpu, ASMReadTSC() + pVmcb->ctrl.u64TSCOffset - 0x400);
2898	}
2899
2900	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatInGC, &pVCpu->hm.s.StatExit1, x);
2901	TMNotifyEndOfExecution(pVCpu); /* Notify TM that the guest is no longer running. */
2902	VMCPU_SET_STATE(pVCpu, VMCPUSTATE_STARTED_HM);
2903
2904	Assert(!(ASMGetFlags() & X86_EFL_IF));
2905	ASMSetFlags(pSvmTransient->uEflags); /* Enable interrupts. */
2906
2907	VMMRZCallRing3Enable(pVCpu); /* It is now safe to do longjmps to ring-3!!! */
2908
2909	/* If VMRUN failed, we can bail out early. This does -not- cover SVM_EXIT_INVALID. */
2910	if (RT_UNLIKELY(rcVMRun != VINF_SUCCESS))
2911	{
2912	Log4(("VMRUN failure: rcVMRun=%Rrc\n", rcVMRun));
2913	return;
2914	}
2915
2916	pSvmTransient->u64ExitCode = pVmcb->ctrl.u64ExitCode; /* Save the #VMEXIT reason. */
2917	pSvmTransient->fVectoringPF = false; /* Vectoring page-fault needs to be determined later. */
2918	hmR0SvmSaveGuestState(pVCpu, pMixedCtx); /* Save the guest state from the VMCB to the guest-CPU context. */
2919
2920	if (RT_LIKELY(pSvmTransient->u64ExitCode != (uint64_t)SVM_EXIT_INVALID))
2921	{
2922	if (pVCpu->hm.s.svm.fSyncVTpr)
2923	{
2924	/* TPR patching (for 32-bit guests) uses LSTAR MSR for holding the TPR value, otherwise uses the VTPR. */
2925	if ( pVM->hm.s.fTPRPatchingActive
2926	&& (pMixedCtx->msrLSTAR & 0xff) != pSvmTransient->u8GuestTpr)
2927	{
2928	int rc = PDMApicSetTPR(pVCpu, pMixedCtx->msrLSTAR & 0xff);
2929	AssertRC(rc);
2930	pVCpu->hm.s.fContextUseFlags \|= HM_CHANGED_SVM_GUEST_APIC_STATE;
2931	}
2932	else if (pSvmTransient->u8GuestTpr != pVmcb->ctrl.IntCtrl.n.u8VTPR)
2933	{
2934	int rc = PDMApicSetTPR(pVCpu, pVmcb->ctrl.IntCtrl.n.u8VTPR << 4);
2935	AssertRC(rc);
2936	pVCpu->hm.s.fContextUseFlags \|= HM_CHANGED_SVM_GUEST_APIC_STATE;
2937	}
2938	}
2939	}
2940	}
2941
2942
2943	/**
2944	* Runs the guest code using AMD-V.
2945	*
2946	* @returns VBox status code.
2947	* @param pVM Pointer to the VM.
2948	* @param pVCpu Pointer to the VMCPU.
2949	* @param pCtx Pointer to the guest-CPU context.
2950	*/
2951	VMMR0DECL(int) SVMR0RunGuestCode(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
2952	{
2953	Assert(VMMRZCallRing3IsEnabled(pVCpu));
2954	HMSVM_ASSERT_PREEMPT_SAFE();
2955	VMMRZCallRing3SetNotification(pVCpu, hmR0SvmCallRing3Callback, pCtx);
2956
2957	SVMTRANSIENT SvmTransient;
2958	SvmTransient.fUpdateTscOffsetting = true;
2959	uint32_t cLoops = 0;
2960	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2961	int rc = VERR_INTERNAL_ERROR_5;
2962
2963	for (;; cLoops++)
2964	{
2965	Assert(!HMR0SuspendPending());
2966	HMSVM_ASSERT_CPU_SAFE();
2967
2968	/* Preparatory work for running guest code, this may return to ring-3 for some last minute updates. */
2969	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatEntry, x);
2970	rc = hmR0SvmPreRunGuest(pVM, pVCpu, pCtx, &SvmTransient);
2971	if (rc != VINF_SUCCESS)
2972	break;
2973
2974	/*
2975	* No longjmps to ring-3 from this point on!!!
2976	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional, better than a kernel panic.
2977	* This also disables flushing of the R0-logger instance (if any).
2978	*/
2979	VMMRZCallRing3Disable(pVCpu);
2980	hmR0SvmPreRunGuestCommitted(pVM, pVCpu, pCtx, &SvmTransient);
2981
2982	rc = hmR0SvmRunGuest(pVM, pVCpu, pCtx);
2983
2984	/*
2985	* Restore any residual host-state and save any bits shared between host and guest into the guest-CPU state.
2986	* This will also re-enable longjmps to ring-3 when it has reached a safe point!!!
2987	*/
2988	hmR0SvmPostRunGuest(pVM, pVCpu, pCtx, &SvmTransient, rc);
2989	if (RT_UNLIKELY( rc != VINF_SUCCESS /* Check for VMRUN errors. */
2990	\|\| SvmTransient.u64ExitCode == (uint64_t)SVM_EXIT_INVALID)) /* Check for invalid guest-state errors. */
2991	{
2992	if (rc == VINF_SUCCESS)
2993	rc = VERR_SVM_INVALID_GUEST_STATE;
2994	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit1, x);
2995	hmR0SvmReportWorldSwitchError(pVM, pVCpu, rc, pCtx);
2996	break;
2997	}
2998
2999	/* Handle the #VMEXIT. */
3000	HMSVM_EXITCODE_STAM_COUNTER_INC(SvmTransient.u64ExitCode);
3001	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatExit1, &pVCpu->hm.s.StatExit2, x);
3002	rc = hmR0SvmHandleExit(pVCpu, pCtx, &SvmTransient);
3003	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit2, x);
3004	if (rc != VINF_SUCCESS)
3005	break;
3006	else if (cLoops > pVM->hm.s.cMaxResumeLoops)
3007	{
3008	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitMaxResume);
3009	rc = VINF_EM_RAW_INTERRUPT;
3010	break;
3011	}
3012	}
3013
3014	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatEntry, x);
3015	if (rc == VERR_EM_INTERPRETER)
3016	rc = VINF_EM_RAW_EMULATE_INSTR;
3017	else if (rc == VINF_EM_RESET)
3018	rc = VINF_EM_TRIPLE_FAULT;
3019
3020	hmR0SvmExitToRing3(pVM, pVCpu, pCtx, rc);
3021	VMMRZCallRing3RemoveNotification(pVCpu);
3022	return rc;
3023	}
3024
3025
3026	/**
3027	* Handles a #VMEXIT (for all EXITCODE values except SVM_EXIT_INVALID).
3028	*
3029	* @returns VBox status code (informational status codes included).
3030	* @param pVCpu Pointer to the VMCPU.
3031	* @param pCtx Pointer to the guest-CPU context.
3032	* @param pSvmTransient Pointer to the SVM transient structure.
3033	*/
3034	DECLINLINE(int) hmR0SvmHandleExit(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
3035	{
3036	Assert(pSvmTransient->u64ExitCode != (uint64_t)SVM_EXIT_INVALID);
3037	Assert(pSvmTransient->u64ExitCode <= SVM_EXIT_MAX);
3038
3039	/*
3040	* The ordering of the case labels is based on most-frequently-occurring VM-exits for most guests under
3041	* normal workloads (for some definition of "normal").
3042	*/
3043	uint32_t u32ExitCode = pSvmTransient->u64ExitCode;
3044	switch (pSvmTransient->u64ExitCode)
3045	{
3046	case SVM_EXIT_NPF:
3047	return hmR0SvmExitNestedPF(pVCpu, pCtx, pSvmTransient);
3048
3049	case SVM_EXIT_IOIO:
3050	return hmR0SvmExitIOInstr(pVCpu, pCtx, pSvmTransient);
3051
3052	case SVM_EXIT_RDTSC:
3053	return hmR0SvmExitRdtsc(pVCpu, pCtx, pSvmTransient);
3054
3055	case SVM_EXIT_RDTSCP:
3056	return hmR0SvmExitRdtscp(pVCpu, pCtx, pSvmTransient);
3057
3058	case SVM_EXIT_CPUID:
3059	return hmR0SvmExitCpuid(pVCpu, pCtx, pSvmTransient);
3060
3061	case SVM_EXIT_EXCEPTION_E: /* X86_XCPT_PF */
3062	return hmR0SvmExitXcptPF(pVCpu, pCtx, pSvmTransient);
3063
3064	case SVM_EXIT_EXCEPTION_7: /* X86_XCPT_NM */
3065	return hmR0SvmExitXcptNM(pVCpu, pCtx, pSvmTransient);
3066
3067	case SVM_EXIT_EXCEPTION_10: /* X86_XCPT_MF */
3068	return hmR0SvmExitXcptMF(pVCpu, pCtx, pSvmTransient);
3069
3070	case SVM_EXIT_EXCEPTION_1: /* X86_XCPT_DB */
3071	return hmR0SvmExitXcptDB(pVCpu, pCtx, pSvmTransient);
3072
3073	case SVM_EXIT_MONITOR:
3074	return hmR0SvmExitMonitor(pVCpu, pCtx, pSvmTransient);
3075
3076	case SVM_EXIT_MWAIT:
3077	return hmR0SvmExitMwait(pVCpu, pCtx, pSvmTransient);
3078
3079	case SVM_EXIT_HLT:
3080	return hmR0SvmExitHlt(pVCpu, pCtx, pSvmTransient);
3081
3082	case SVM_EXIT_READ_CR0:
3083	case SVM_EXIT_READ_CR3:
3084	case SVM_EXIT_READ_CR4:
3085	return hmR0SvmExitReadCRx(pVCpu, pCtx, pSvmTransient);
3086
3087	case SVM_EXIT_WRITE_CR0:
3088	case SVM_EXIT_WRITE_CR3:
3089	case SVM_EXIT_WRITE_CR4:
3090	case SVM_EXIT_WRITE_CR8:
3091	return hmR0SvmExitWriteCRx(pVCpu, pCtx, pSvmTransient);
3092
3093	case SVM_EXIT_VINTR:
3094	return hmR0SvmExitVIntr(pVCpu, pCtx, pSvmTransient);
3095
3096	case SVM_EXIT_INTR:
3097	case SVM_EXIT_FERR_FREEZE:
3098	case SVM_EXIT_NMI:
3099	return hmR0SvmExitIntr(pVCpu, pCtx, pSvmTransient);
3100
3101	case SVM_EXIT_MSR:
3102	return hmR0SvmExitMsr(pVCpu, pCtx, pSvmTransient);
3103
3104	case SVM_EXIT_INVLPG:
3105	return hmR0SvmExitInvlpg(pVCpu, pCtx, pSvmTransient);
3106
3107	case SVM_EXIT_WBINVD:
3108	return hmR0SvmExitWbinvd(pVCpu, pCtx, pSvmTransient);
3109
3110	case SVM_EXIT_INVD:
3111	return hmR0SvmExitInvd(pVCpu, pCtx, pSvmTransient);
3112
3113	case SVM_EXIT_RDPMC:
3114	return hmR0SvmExitRdpmc(pVCpu, pCtx, pSvmTransient);
3115
3116	default:
3117	{
3118	switch (pSvmTransient->u64ExitCode)
3119	{
3120	case SVM_EXIT_READ_DR0: case SVM_EXIT_READ_DR1: case SVM_EXIT_READ_DR2: case SVM_EXIT_READ_DR3:
3121	case SVM_EXIT_READ_DR6: case SVM_EXIT_READ_DR7: case SVM_EXIT_READ_DR8: case SVM_EXIT_READ_DR9:
3122	case SVM_EXIT_READ_DR10: case SVM_EXIT_READ_DR11: case SVM_EXIT_READ_DR12: case SVM_EXIT_READ_DR13:
3123	case SVM_EXIT_READ_DR14: case SVM_EXIT_READ_DR15:
3124	return hmR0SvmExitReadDRx(pVCpu, pCtx, pSvmTransient);
3125
3126	case SVM_EXIT_WRITE_DR0: case SVM_EXIT_WRITE_DR1: case SVM_EXIT_WRITE_DR2: case SVM_EXIT_WRITE_DR3:
3127	case SVM_EXIT_WRITE_DR6: case SVM_EXIT_WRITE_DR7: case SVM_EXIT_WRITE_DR8: case SVM_EXIT_WRITE_DR9:
3128	case SVM_EXIT_WRITE_DR10: case SVM_EXIT_WRITE_DR11: case SVM_EXIT_WRITE_DR12: case SVM_EXIT_WRITE_DR13:
3129	case SVM_EXIT_WRITE_DR14: case SVM_EXIT_WRITE_DR15:
3130	return hmR0SvmExitWriteDRx(pVCpu, pCtx, pSvmTransient);
3131
3132	case SVM_EXIT_TASK_SWITCH:
3133	return hmR0SvmExitTaskSwitch(pVCpu, pCtx, pSvmTransient);
3134
3135	case SVM_EXIT_VMMCALL:
3136	return hmR0SvmExitVmmCall(pVCpu, pCtx, pSvmTransient);
3137
3138	case SVM_EXIT_SHUTDOWN:
3139	return hmR0SvmExitShutdown(pVCpu, pCtx, pSvmTransient);
3140
3141	case SVM_EXIT_SMI:
3142	case SVM_EXIT_INIT:
3143	{
3144	/*
3145	* We don't intercept NMIs. As for INIT signals, it really shouldn't ever happen here. If it ever does,
3146	* we want to know about it so log the exit code and bail.
3147	*/
3148	AssertMsgFailed(("hmR0SvmHandleExit: Unexpected exit %#RX32\n", (uint32_t)pSvmTransient->u64ExitCode));
3149	pVCpu->hm.s.u32HMError = (uint32_t)pSvmTransient->u64ExitCode;
3150	return VERR_SVM_UNEXPECTED_EXIT;
3151	}
3152
3153	case SVM_EXIT_INVLPGA:
3154	case SVM_EXIT_RSM:
3155	case SVM_EXIT_VMRUN:
3156	case SVM_EXIT_VMLOAD:
3157	case SVM_EXIT_VMSAVE:
3158	case SVM_EXIT_STGI:
3159	case SVM_EXIT_CLGI:
3160	case SVM_EXIT_SKINIT:
3161	return hmR0SvmExitSetPendingXcptUD(pVCpu, pCtx, pSvmTransient);
3162
3163	#ifdef HMSVM_ALWAYS_TRAP_ALL_XCPTS
3164	case SVM_EXIT_EXCEPTION_0: /* X86_XCPT_DE */
3165	/* SVM_EXIT_EXCEPTION_1: / / X86_XCPT_DB - Handled above. */
3166	case SVM_EXIT_EXCEPTION_2: /* X86_XCPT_NMI */
3167	case SVM_EXIT_EXCEPTION_3: /* X86_XCPT_BP */
3168	case SVM_EXIT_EXCEPTION_4: /* X86_XCPT_OF */
3169	case SVM_EXIT_EXCEPTION_5: /* X86_XCPT_BR */
3170	case SVM_EXIT_EXCEPTION_6: /* X86_XCPT_UD */
3171	/* SVM_EXIT_EXCEPTION_7: / / X86_XCPT_NM - Handled above. */
3172	case SVM_EXIT_EXCEPTION_8: /* X86_XCPT_DF */
3173	case SVM_EXIT_EXCEPTION_9: /* X86_XCPT_CO_SEG_OVERRUN */
3174	case SVM_EXIT_EXCEPTION_A: /* X86_XCPT_TS */
3175	case SVM_EXIT_EXCEPTION_B: /* X86_XCPT_NP */
3176	case SVM_EXIT_EXCEPTION_C: /* X86_XCPT_SS */
3177	case SVM_EXIT_EXCEPTION_D: /* X86_XCPT_GP */
3178	/* SVM_EXIT_EXCEPTION_E: / / X86_XCPT_PF - Handled above. */
3179	/* SVM_EXIT_EXCEPTION_10: / / X86_XCPT_MF - Handled above. */
3180	case SVM_EXIT_EXCEPTION_11: /* X86_XCPT_AC */
3181	case SVM_EXIT_EXCEPTION_12: /* X86_XCPT_MC */
3182	case SVM_EXIT_EXCEPTION_13: /* X86_XCPT_XF */
3183
3184	case SVM_EXIT_EXCEPTION_F: /* Reserved */
3185	case SVM_EXIT_EXCEPTION_14: case SVM_EXIT_EXCEPTION_15: case SVM_EXIT_EXCEPTION_16:
3186	case SVM_EXIT_EXCEPTION_17: case SVM_EXIT_EXCEPTION_18: case SVM_EXIT_EXCEPTION_19:
3187	case SVM_EXIT_EXCEPTION_1A: case SVM_EXIT_EXCEPTION_1B: case SVM_EXIT_EXCEPTION_1C:
3188	case SVM_EXIT_EXCEPTION_1D: case SVM_EXIT_EXCEPTION_1E: case SVM_EXIT_EXCEPTION_1F:
3189	{
3190	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
3191	SVMEVENT Event;
3192	Event.u = 0;
3193	Event.n.u1Valid = 1;
3194	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3195	Event.n.u8Vector = pSvmTransient->u64ExitCode - SVM_EXIT_EXCEPTION_0;
3196
3197	switch (Event.n.u8Vector)
3198	{
3199	case X86_XCPT_DE:
3200	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestDE);
3201	break;
3202
3203	case X86_XCPT_BP:
3204	/** Saves the wrong EIP on the stack (pointing to the int3) instead of the
3205	* next instruction. */
3206	/** @todo Investigate this later. */
3207	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestBP);
3208	break;
3209
3210	case X86_XCPT_UD:
3211	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestUD);
3212	break;
3213
3214	case X86_XCPT_NP:
3215	Event.n.u1ErrorCodeValid = 1;
3216	Event.n.u32ErrorCode = pVmcb->ctrl.u64ExitInfo1;
3217	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestNP);
3218	break;
3219
3220	case X86_XCPT_SS:
3221	Event.n.u1ErrorCodeValid = 1;
3222	Event.n.u32ErrorCode = pVmcb->ctrl.u64ExitInfo1;
3223	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestSS);
3224	break;
3225
3226	case X86_XCPT_GP:
3227	Event.n.u1ErrorCodeValid = 1;
3228	Event.n.u32ErrorCode = pVmcb->ctrl.u64ExitInfo1;
3229	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestGP);
3230	break;
3231
3232	default:
3233	AssertMsgFailed(("hmR0SvmHandleExit: Unexpected exit caused by exception %#x\n", Event.n.u8Vector));
3234	pVCpu->hm.s.u32HMError = Event.n.u8Vector;
3235	return VERR_SVM_UNEXPECTED_XCPT_EXIT;
3236	}
3237
3238	Log4(("#Xcpt: Vector=%#x at CS:RIP=%04x:%RGv\n", Event.n.u8Vector, pCtx->cs.Sel, (RTGCPTR)pCtx->rip));
3239	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3240	return VINF_SUCCESS;
3241	}
3242	#endif /* HMSVM_ALWAYS_TRAP_ALL_XCPTS */
3243
3244	default:
3245	{
3246	AssertMsgFailed(("hmR0SvmHandleExit: Unknown exit code %#x\n", u32ExitCode));
3247	pVCpu->hm.s.u32HMError = u32ExitCode;
3248	return VERR_SVM_UNKNOWN_EXIT;
3249	}
3250	}
3251	}
3252	}
3253	return VERR_INTERNAL_ERROR_5; /* Should never happen. */
3254	}
3255
3256
3257	#ifdef DEBUG
3258	/* Is there some generic IPRT define for this that are not in Runtime/internal/\* ?? */
3259	# define HMSVM_ASSERT_PREEMPT_CPUID_VAR() \
3260	RTCPUID const idAssertCpu = RTThreadPreemptIsEnabled(NIL_RTTHREAD) ? NIL_RTCPUID : RTMpCpuId()
3261
3262	# define HMSVM_ASSERT_PREEMPT_CPUID() \
3263	do \
3264	{ \
3265	RTCPUID const idAssertCpuNow = RTThreadPreemptIsEnabled(NIL_RTTHREAD) ? NIL_RTCPUID : RTMpCpuId(); \
3266	AssertMsg(idAssertCpu == idAssertCpuNow, ("SVM %#x, %#x\n", idAssertCpu, idAssertCpuNow)); \
3267	} while (0)
3268
3269	# define HMSVM_VALIDATE_EXIT_HANDLER_PARAMS() \
3270	do { \
3271	AssertPtr(pVCpu); \
3272	AssertPtr(pCtx); \
3273	AssertPtr(pSvmTransient); \
3274	Assert(ASMIntAreEnabled()); \
3275	HMSVM_ASSERT_PREEMPT_SAFE(); \
3276	HMSVM_ASSERT_PREEMPT_CPUID_VAR(); \
3277	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)); \
3278	HMSVM_ASSERT_PREEMPT_SAFE(); \
3279	if (VMMR0IsLogFlushDisabled(pVCpu)) \
3280	HMSVM_ASSERT_PREEMPT_CPUID(); \
3281	} while (0)
3282	#else /* Release builds */
3283	# define HMSVM_VALIDATE_EXIT_HANDLER_PARAMS() do { } while(0)
3284	#endif
3285
3286
3287	/**
3288	* Worker for hmR0SvmInterpretInvlpg().
3289	*
3290	* @return VBox status code.
3291	* @param pVCpu Pointer to the VMCPU.
3292	* @param pCpu Pointer to the disassembler state.
3293	* @param pRegFrame Pointer to the register frame.
3294	*/
3295	static int hmR0SvmInterpretInvlPgEx(PVMCPU pVCpu, PDISCPUSTATE pCpu, PCPUMCTXCORE pRegFrame)
3296	{
3297	DISQPVPARAMVAL Param1;
3298	RTGCPTR GCPtrPage;
3299
3300	int rc = DISQueryParamVal(pRegFrame, pCpu, &pCpu->Param1, &Param1, DISQPVWHICH_SRC);
3301	if (RT_FAILURE(rc))
3302	return VERR_EM_INTERPRETER;
3303
3304	if ( Param1.type == DISQPV_TYPE_IMMEDIATE
3305	\|\| Param1.type == DISQPV_TYPE_ADDRESS)
3306	{
3307	if (!(Param1.flags & (DISQPV_FLAG_32 \| DISQPV_FLAG_64)))
3308	return VERR_EM_INTERPRETER;
3309
3310	GCPtrPage = Param1.val.val64;
3311	VBOXSTRICTRC rc2 = EMInterpretInvlpg(pVCpu->CTX_SUFF(pVM), pVCpu, pRegFrame, GCPtrPage);
3312	rc = VBOXSTRICTRC_VAL(rc2);
3313	}
3314	else
3315	{
3316	Log4(("hmR0SvmInterpretInvlPgEx invalid parameter type %#x\n", Param1.type));
3317	rc = VERR_EM_INTERPRETER;
3318	}
3319
3320	return rc;
3321	}
3322
3323
3324	/**
3325	* Interprets INVLPG.
3326	*
3327	* @returns VBox status code.
3328	* @retval VINF_* Scheduling instructions.
3329	* @retval VERR_EM_INTERPRETER Something we can't cope with.
3330	* @retval VERR_* Fatal errors.
3331	*
3332	* @param pVM Pointer to the VM.
3333	* @param pRegFrame Pointer to the register frame.
3334	*
3335	* @remarks Updates the RIP if the instruction was executed successfully.
3336	*/
3337	static int hmR0SvmInterpretInvlpg(PVM pVM, PVMCPU pVCpu, PCPUMCTXCORE pRegFrame)
3338	{
3339	/* Only allow 32 & 64 bit code. */
3340	if (CPUMGetGuestCodeBits(pVCpu) != 16)
3341	{
3342	PDISSTATE pDis = &pVCpu->hm.s.DisState;
3343	int rc = EMInterpretDisasCurrent(pVM, pVCpu, pDis, NULL /* pcbInstr */);
3344	if ( RT_SUCCESS(rc)
3345	&& pDis->pCurInstr->uOpcode == OP_INVLPG)
3346	{
3347	rc = hmR0SvmInterpretInvlPgEx(pVCpu, pDis, pRegFrame);
3348	if (RT_SUCCESS(rc))
3349	pRegFrame->rip += pDis->cbInstr;
3350	return rc;
3351	}
3352	else
3353	Log4(("hmR0SvmInterpretInvlpg: EMInterpretDisasCurrent returned %Rrc uOpCode=%#x\n", rc, pDis->pCurInstr->uOpcode));
3354	}
3355	return VERR_EM_INTERPRETER;
3356	}
3357
3358
3359	/**
3360	* Sets an invalid-opcode (#UD) exception as pending-for-injection into the VM.
3361	*
3362	* @param pVCpu Pointer to the VMCPU.
3363	*/
3364	DECLINLINE(void) hmR0SvmSetPendingXcptUD(PVMCPU pVCpu)
3365	{
3366	SVMEVENT Event;
3367	Event.u = 0;
3368	Event.n.u1Valid = 1;
3369	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3370	Event.n.u8Vector = X86_XCPT_UD;
3371	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3372	}
3373
3374
3375	/**
3376	* Sets a debug (#DB) exception as pending-for-injection into the VM.
3377	*
3378	* @param pVCpu Pointer to the VMCPU.
3379	*/
3380	DECLINLINE(void) hmR0SvmSetPendingXcptDB(PVMCPU pVCpu)
3381	{
3382	SVMEVENT Event;
3383	Event.u = 0;
3384	Event.n.u1Valid = 1;
3385	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3386	Event.n.u8Vector = X86_XCPT_DB;
3387	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3388	}
3389
3390
3391	/**
3392	* Sets a page fault (#PF) exception as pending-for-injection into the VM.
3393	*
3394	* @param pVCpu Pointer to the VMCPU.
3395	* @param pCtx Pointer to the guest-CPU context.
3396	* @param u32ErrCode The error-code for the page-fault.
3397	* @param uFaultAddress The page fault address (CR2).
3398	*
3399	* @remarks This updates the guest CR2 with @a uFaultAddress!
3400	*/
3401	DECLINLINE(void) hmR0SvmSetPendingXcptPF(PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t u32ErrCode, RTGCUINTPTR uFaultAddress)
3402	{
3403	SVMEVENT Event;
3404	Event.u = 0;
3405	Event.n.u1Valid = 1;
3406	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3407	Event.n.u8Vector = X86_XCPT_PF;
3408	Event.n.u1ErrorCodeValid = 1;
3409	Event.n.u32ErrorCode = u32ErrCode;
3410
3411	/* Update CR2 of the guest. */
3412	if (pCtx->cr2 != uFaultAddress)
3413	{
3414	pCtx->cr2 = uFaultAddress;
3415	pVCpu->hm.s.fContextUseFlags \|= HM_CHANGED_GUEST_CR2;
3416	}
3417
3418	hmR0SvmSetPendingEvent(pVCpu, &Event, uFaultAddress);
3419	}
3420
3421
3422	/**
3423	* Sets a device-not-available (#NM) exception as pending-for-injection into the
3424	* VM.
3425	*
3426	* @param pVCpu Pointer to the VMCPU.
3427	*/
3428	DECLINLINE(void) hmR0SvmSetPendingXcptNM(PVMCPU pVCpu)
3429	{
3430	SVMEVENT Event;
3431	Event.u = 0;
3432	Event.n.u1Valid = 1;
3433	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3434	Event.n.u8Vector = X86_XCPT_NM;
3435	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3436	}
3437
3438
3439	/**
3440	* Sets a math-fault (#MF) exception as pending-for-injection into the VM.
3441	*
3442	* @param pVCpu Pointer to the VMCPU.
3443	*/
3444	DECLINLINE(void) hmR0SvmSetPendingXcptMF(PVMCPU pVCpu)
3445	{
3446	SVMEVENT Event;
3447	Event.u = 0;
3448	Event.n.u1Valid = 1;
3449	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3450	Event.n.u8Vector = X86_XCPT_MF;
3451	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3452	}
3453
3454
3455	/**
3456	* Sets a double fault (#DF) exception as pending-for-injection into the VM.
3457	*
3458	* @param pVCpu Pointer to the VMCPU.
3459	*/
3460	DECLINLINE(void) hmR0SvmSetPendingXcptDF(PVMCPU pVCpu)
3461	{
3462	SVMEVENT Event;
3463	Event.u = 0;
3464	Event.n.u1Valid = 1;
3465	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3466	Event.n.u8Vector = X86_XCPT_DF;
3467	Event.n.u1ErrorCodeValid = 1;
3468	Event.n.u32ErrorCode = 0;
3469	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3470	}
3471
3472
3473	/**
3474	* Emulates a simple MOV TPR (CR8) instruction, used for TPR patching on 32-bit
3475	* guests. This simply looks up the patch record at EIP and does the required.
3476	*
3477	* This VMMCALL is used a fallback mechanism when mov to/from cr8 isn't exactly
3478	* like how we want it to be (e.g. not followed by shr 4 as is usually done for
3479	* TPR). See hmR3ReplaceTprInstr() for the details.
3480	*
3481	* @returns VBox status code.
3482	* @param pVM Pointer to the VM.
3483	* @param pVCpu Pointer to the VMCPU.
3484	* @param pCtx Pointer to the guest-CPU context.
3485	*/
3486	static int hmR0SvmEmulateMovTpr(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3487	{
3488	Log4(("Emulated VMMCall TPR access replacement at RIP=%RGv\n", pCtx->rip));
3489	for (;;)
3490	{
3491	bool fPending;
3492	uint8_t u8Tpr;
3493
3494	PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
3495	if (!pPatch)
3496	break;
3497
3498	switch (pPatch->enmType)
3499	{
3500	case HMTPRINSTR_READ:
3501	{
3502	int rc = PDMApicGetTPR(pVCpu, &u8Tpr, &fPending, NULL /* pu8PendingIrq */);
3503	AssertRC(rc);
3504
3505	rc = DISWriteReg32(CPUMCTX2CORE(pCtx), pPatch->uDstOperand, u8Tpr);
3506	AssertRC(rc);
3507	pCtx->rip += pPatch->cbOp;
3508	break;
3509	}
3510
3511	case HMTPRINSTR_WRITE_REG:
3512	case HMTPRINSTR_WRITE_IMM:
3513	{
3514	if (pPatch->enmType == HMTPRINSTR_WRITE_REG)
3515	{
3516	uint32_t u32Val;
3517	int rc = DISFetchReg32(CPUMCTX2CORE(pCtx), pPatch->uSrcOperand, &u32Val);
3518	AssertRC(rc);
3519	u8Tpr = u32Val;
3520	}
3521	else
3522	u8Tpr = (uint8_t)pPatch->uSrcOperand;
3523
3524	int rc2 = PDMApicSetTPR(pVCpu, u8Tpr);
3525	AssertRC(rc2);
3526	pVCpu->hm.s.fContextUseFlags \|= HM_CHANGED_SVM_GUEST_APIC_STATE;
3527
3528	pCtx->rip += pPatch->cbOp;
3529	break;
3530	}
3531
3532	default:
3533	AssertMsgFailed(("Unexpected patch type %d\n", pPatch->enmType));
3534	pVCpu->hm.s.u32HMError = pPatch->enmType;
3535	return VERR_SVM_UNEXPECTED_PATCH_TYPE;
3536	}
3537	}
3538
3539	return VINF_SUCCESS;
3540	}
3541
3542	/**
3543	* Determines if an exception is a contributory exception. Contributory
3544	* exceptions are ones which can cause double-faults. Page-fault is
3545	* intentionally not included here as it's a conditional contributory exception.
3546	*
3547	* @returns true if the exception is contributory, false otherwise.
3548	* @param uVector The exception vector.
3549	*/
3550	DECLINLINE(bool) hmR0SvmIsContributoryXcpt(const uint32_t uVector)
3551	{
3552	switch (uVector)
3553	{
3554	case X86_XCPT_GP:
3555	case X86_XCPT_SS:
3556	case X86_XCPT_NP:
3557	case X86_XCPT_TS:
3558	case X86_XCPT_DE:
3559	return true;
3560	default:
3561	break;
3562	}
3563	return false;
3564	}
3565
3566
3567	/**
3568	* Handle a condition that occurred while delivering an event through the guest
3569	* IDT.
3570	*
3571	* @returns VBox status code (informational error codes included).
3572	* @retval VINF_SUCCESS if we should continue handling the VM-exit.
3573	* @retval VINF_HM_DOUBLE_FAULT if a #DF condition was detected and we ought to
3574	* continue execution of the guest which will delivery the #DF.
3575	* @retval VINF_EM_RESET if we detected a triple-fault condition.
3576	*
3577	* @param pVCpu Pointer to the VMCPU.
3578	* @param pCtx Pointer to the guest-CPU context.
3579	* @param pSvmTransient Pointer to the SVM transient structure.
3580	*
3581	* @remarks No-long-jump zone!!!
3582	*/
3583	static int hmR0SvmCheckExitDueToEventDelivery(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
3584	{
3585	int rc = VINF_SUCCESS;
3586	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
3587
3588	/* See AMD spec. 15.7.3 "EXITINFO Pseudo-Code". The EXITINTINFO (if valid) contains the prior exception (IDT vector)
3589	* that was trying to be delivered to the guest which caused a #VMEXIT which was intercepted (Exit vector). */
3590	if (pVmcb->ctrl.ExitIntInfo.n.u1Valid)
3591	{
3592	uint8_t uIdtVector = pVmcb->ctrl.ExitIntInfo.n.u8Vector;
3593
3594	typedef enum
3595	{
3596	SVMREFLECTXCPT_XCPT, /* Reflect the exception to the guest or for further evaluation by VMM. */
3597	SVMREFLECTXCPT_DF, /* Reflect the exception as a double-fault to the guest. */
3598	SVMREFLECTXCPT_TF, /* Indicate a triple faulted state to the VMM. */
3599	SVMREFLECTXCPT_NONE /* Nothing to reflect. */
3600	} SVMREFLECTXCPT;
3601
3602	SVMREFLECTXCPT enmReflect = SVMREFLECTXCPT_NONE;
3603	if (pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_EXCEPTION)
3604	{
3605	if (pSvmTransient->u64ExitCode - SVM_EXIT_EXCEPTION_0 <= SVM_EXIT_EXCEPTION_1F)
3606	{
3607	uint8_t uExitVector = (uint8_t)(pSvmTransient->u64ExitCode - SVM_EXIT_EXCEPTION_0);
3608
3609	#ifdef VBOX_STRICT
3610	if ( hmR0SvmIsContributoryXcpt(uIdtVector)
3611	&& uExitVector == X86_XCPT_PF)
3612	{
3613	Log4(("IDT: Contributory #PF uCR2=%#RX64\n", pVCpu->idCpu, pCtx->cr2));
3614	}
3615	#endif
3616	if ( uExitVector == X86_XCPT_PF
3617	&& uIdtVector == X86_XCPT_PF)
3618	{
3619	pSvmTransient->fVectoringPF = true;
3620	Log4(("IDT: Vectoring #PF uCR2=%#RX64\n", pCtx->cr2));
3621	}
3622	else if ( (pVmcb->ctrl.u32InterceptException & HMSVM_CONTRIBUTORY_XCPT_MASK)
3623	&& hmR0SvmIsContributoryXcpt(uExitVector)
3624	&& ( hmR0SvmIsContributoryXcpt(uIdtVector)
3625	\|\| uIdtVector == X86_XCPT_PF))
3626	{
3627	enmReflect = SVMREFLECTXCPT_DF;
3628	Log4(("IDT: Pending vectoring #DF %#RX64 uIdtVector=%#x uExitVector=%#x\n", pVCpu->hm.s.Event.u64IntrInfo,
3629	uIdtVector, uExitVector));
3630	}
3631	else if (uIdtVector == X86_XCPT_DF)
3632	{
3633	enmReflect = SVMREFLECTXCPT_TF;
3634	Log4(("IDT: Pending vectoring triple-fault %#RX64 uIdtVector=%#x uExitVector=%#x\n", pVCpu->hm.s.Event.u64IntrInfo,
3635	uIdtVector, uExitVector));
3636	}
3637	else
3638	enmReflect = SVMREFLECTXCPT_XCPT;
3639	}
3640	else
3641	{
3642	/*
3643	* If event delivery caused an #VMEXIT that is not an exception (e.g. #NPF) then reflect the original
3644	* exception to the guest after handling the VM-exit.
3645	*/
3646	enmReflect = SVMREFLECTXCPT_XCPT;
3647	}
3648	}
3649	else if (pVmcb->ctrl.ExitIntInfo.n.u3Type != SVM_EVENT_SOFTWARE_INT)
3650	{
3651	/* Ignore software interrupts (INT n) as they reoccur when restarting the instruction. */
3652	enmReflect = SVMREFLECTXCPT_XCPT;
3653	}
3654
3655	switch (enmReflect)
3656	{
3657	case SVMREFLECTXCPT_XCPT:
3658	{
3659	Assert(pVmcb->ctrl.ExitIntInfo.n.u3Type != SVM_EVENT_SOFTWARE_INT);
3660	hmR0SvmSetPendingEvent(pVCpu, &pVmcb->ctrl.ExitIntInfo, 0 /* GCPtrFaultAddress */);
3661
3662	/* If uExitVector is #PF, CR2 value will be updated from the VMCB if it's a guest #PF. See hmR0SvmExitXcptPF(). */
3663	Log4(("IDT: Pending vectoring event %#RX64 ErrValid=%RTbool Err=%#RX32\n", pVmcb->ctrl.ExitIntInfo.u,
3664	!!pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid, pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode));
3665	break;
3666	}
3667
3668	case SVMREFLECTXCPT_DF:
3669	{
3670	hmR0SvmSetPendingXcptDF(pVCpu);
3671	rc = VINF_HM_DOUBLE_FAULT;
3672	break;
3673	}
3674
3675	case SVMREFLECTXCPT_TF:
3676	{
3677	rc = VINF_EM_RESET;
3678	break;
3679	}
3680
3681	default:
3682	Assert(rc == VINF_SUCCESS);
3683	break;
3684	}
3685	}
3686	Assert(rc == VINF_SUCCESS \|\| rc == VINF_HM_DOUBLE_FAULT \|\| rc == VINF_EM_RESET);
3687	return rc;
3688	}
3689
3690
3691	/**
3692	* Advances the guest RIP in the if the NRIP_SAVE feature is supported by the
3693	* CPU, otherwise advances the RIP by @a cb bytes.
3694	*
3695	* @param pVCpu Pointer to the VMCPU.
3696	* @param pCtx Pointer to the guest-CPU context.
3697	* @param cb RIP increment value in bytes.
3698	*
3699	* @remarks Use this function only from #VMEXIT's where the NRIP value is valid
3700	* when NRIP_SAVE is supported by the CPU!
3701	*/
3702	DECLINLINE(void) hmR0SvmUpdateRip(PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t cb)
3703	{
3704	if (pVCpu->CTX_SUFF(pVM)->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_NRIP_SAVE)
3705	{
3706	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
3707	pCtx->rip = pVmcb->ctrl.u64NextRIP;
3708	}
3709	else
3710	pCtx->rip += cb;
3711	}
3712
3713
3714	/* -=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= */
3715	/* -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- #VMEXIT handlers -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- */
3716	/* -=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= */
3717
3718	/** @name VM-exit handlers.
3719	* @{
3720	*/
3721
3722	/**
3723	* #VMEXIT handler for external interrupts, NMIs, FPU assertion freeze and INIT
3724	* signals (SVM_EXIT_INTR, SVM_EXIT_NMI, SVM_EXIT_FERR_FREEZE, SVM_EXIT_INIT).
3725	*/
3726	HMSVM_EXIT_DECL hmR0SvmExitIntr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
3727	{
3728	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
3729
3730	if (pSvmTransient->u64ExitCode == SVM_EXIT_NMI)
3731	STAM_REL_COUNTER_INC(&pVCpu->hm.s.StatExitHostNmi);
3732	else if (pSvmTransient->u64ExitCode == SVM_EXIT_INTR)
3733	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitExtInt);
3734
3735	/*
3736	* AMD-V has no preemption timer and the generic periodic preemption timer has no way to signal -before- the timer
3737	* fires if the current interrupt is our own timer or a some other host interrupt. We also cannot examine what
3738	* interrupt it is until the host actually take the interrupt.
3739	*
3740	* Going back to executing guest code here unconditionally causes random scheduling problems (observed on an
3741	* AMD Phenom 9850 Quad-Core on Windows 64-bit host).
3742	*/
3743	return VINF_EM_RAW_INTERRUPT;
3744	}
3745
3746
3747	/**
3748	* #VMEXIT handler for WBINVD (SVM_EXIT_WBINVD). Conditional #VMEXIT.
3749	*/
3750	HMSVM_EXIT_DECL hmR0SvmExitWbinvd(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
3751	{
3752	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
3753
3754	hmR0SvmUpdateRip(pVCpu, pCtx, 2);
3755	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitWbinvd);
3756	int rc = VINF_SUCCESS;
3757	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
3758	return rc;
3759	}
3760
3761
3762	/**
3763	* #VMEXIT handler for INVD (SVM_EXIT_INVD). Unconditional #VMEXIT.
3764	*/
3765	HMSVM_EXIT_DECL hmR0SvmExitInvd(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
3766	{
3767	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
3768
3769	hmR0SvmUpdateRip(pVCpu, pCtx, 2);
3770	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitInvd);
3771	int rc = VINF_SUCCESS;
3772	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
3773	return rc;
3774	}
3775
3776
3777	/**
3778	* #VMEXIT handler for INVD (SVM_EXIT_CPUID). Conditional #VMEXIT.
3779	*/
3780	HMSVM_EXIT_DECL hmR0SvmExitCpuid(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
3781	{
3782	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
3783	PVM pVM = pVCpu->CTX_SUFF(pVM);
3784	int rc = EMInterpretCpuId(pVM, pVCpu, CPUMCTX2CORE(pCtx));
3785	if (RT_LIKELY(rc == VINF_SUCCESS))
3786	{
3787	hmR0SvmUpdateRip(pVCpu, pCtx, 2);
3788	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
3789	}
3790	else
3791	{
3792	AssertMsgFailed(("hmR0SvmExitCpuid: EMInterpretCpuId failed with %Rrc\n", rc));
3793	rc = VERR_EM_INTERPRETER;
3794	}
3795	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCpuid);
3796	return rc;
3797	}
3798
3799
3800	/**
3801	* #VMEXIT handler for RDTSC (SVM_EXIT_RDTSC). Conditional #VMEXIT.
3802	*/
3803	HMSVM_EXIT_DECL hmR0SvmExitRdtsc(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
3804	{
3805	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
3806	PVM pVM = pVCpu->CTX_SUFF(pVM);
3807	int rc = EMInterpretRdtsc(pVM, pVCpu, CPUMCTX2CORE(pCtx));
3808	if (RT_LIKELY(rc == VINF_SUCCESS))
3809	{
3810	hmR0SvmUpdateRip(pVCpu, pCtx, 2);
3811	pSvmTransient->fUpdateTscOffsetting = true;
3812
3813	/* Single step check. */
3814	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
3815	}
3816	else
3817	{
3818	AssertMsgFailed(("hmR0SvmExitRdtsc: EMInterpretRdtsc failed with %Rrc\n", rc));
3819	rc = VERR_EM_INTERPRETER;
3820	}
3821	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdtsc);
3822	return rc;
3823	}
3824
3825
3826	/**
3827	* #VMEXIT handler for RDTSCP (SVM_EXIT_RDTSCP). Conditional #VMEXIT.
3828	*/
3829	HMSVM_EXIT_DECL hmR0SvmExitRdtscp(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
3830	{
3831	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
3832	int rc = EMInterpretRdtscp(pVCpu->CTX_SUFF(pVM), pVCpu, pCtx);
3833	if (RT_LIKELY(rc == VINF_SUCCESS))
3834	{
3835	hmR0SvmUpdateRip(pVCpu, pCtx, 3);
3836	pSvmTransient->fUpdateTscOffsetting = true;
3837	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
3838	}
3839	else
3840	{
3841	AssertMsgFailed(("hmR0SvmExitRdtsc: EMInterpretRdtscp failed with %Rrc\n", rc));
3842	rc = VERR_EM_INTERPRETER;
3843	}
3844	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdtscp);
3845	return rc;
3846	}
3847
3848
3849	/**
3850	* #VMEXIT handler for RDPMC (SVM_EXIT_RDPMC). Conditional #VMEXIT.
3851	*/
3852	HMSVM_EXIT_DECL hmR0SvmExitRdpmc(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
3853	{
3854	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
3855	int rc = EMInterpretRdpmc(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
3856	if (RT_LIKELY(rc == VINF_SUCCESS))
3857	{
3858	hmR0SvmUpdateRip(pVCpu, pCtx, 2);
3859	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
3860	}
3861	else
3862	{
3863	AssertMsgFailed(("hmR0SvmExitRdpmc: EMInterpretRdpmc failed with %Rrc\n", rc));
3864	rc = VERR_EM_INTERPRETER;
3865	}
3866	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdpmc);
3867	return rc;
3868	}
3869
3870
3871	/**
3872	* #VMEXIT handler for INVLPG (SVM_EXIT_INVLPG). Conditional #VMEXIT.
3873	*/
3874	HMSVM_EXIT_DECL hmR0SvmExitInvlpg(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
3875	{
3876	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
3877	PVM pVM = pVCpu->CTX_SUFF(pVM);
3878	Assert(!pVM->hm.s.fNestedPaging);
3879
3880	/** @todo Decode Assist. */
3881	int rc = hmR0SvmInterpretInvlpg(pVM, pVCpu, CPUMCTX2CORE(pCtx)); /* Updates RIP if successful. */
3882	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitInvlpg);
3883	Assert(rc == VINF_SUCCESS \|\| rc == VERR_EM_INTERPRETER);
3884	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
3885	return rc;
3886	}
3887
3888
3889	/**
3890	* #VMEXIT handler for HLT (SVM_EXIT_HLT). Conditional #VMEXIT.
3891	*/
3892	HMSVM_EXIT_DECL hmR0SvmExitHlt(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
3893	{
3894	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
3895	hmR0SvmUpdateRip(pVCpu, pCtx, 1);
3896	int rc = EMShouldContinueAfterHalt(pVCpu, pCtx) ? VINF_SUCCESS : VINF_EM_HALT;
3897	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
3898	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitHlt);
3899	return rc;
3900	}
3901
3902
3903	/**
3904	* #VMEXIT handler for MONITOR (SVM_EXIT_MONITOR). Conditional #VMEXIT.
3905	*/
3906	HMSVM_EXIT_DECL hmR0SvmExitMonitor(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
3907	{
3908	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
3909	int rc = EMInterpretMonitor(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
3910	if (RT_LIKELY(rc == VINF_SUCCESS))
3911	{
3912	hmR0SvmUpdateRip(pVCpu, pCtx, 3);
3913	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
3914	}
3915	else
3916	{
3917	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMonitor: EMInterpretMonitor failed with %Rrc\n", rc));
3918	rc = VERR_EM_INTERPRETER;
3919	}
3920	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitMonitor);
3921	return rc;
3922	}
3923
3924
3925	/**
3926	* #VMEXIT handler for MWAIT (SVM_EXIT_MWAIT). Conditional #VMEXIT.
3927	*/
3928	HMSVM_EXIT_DECL hmR0SvmExitMwait(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
3929	{
3930	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
3931	VBOXSTRICTRC rc2 = EMInterpretMWait(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
3932	int rc = VBOXSTRICTRC_VAL(rc2);
3933	if ( rc == VINF_EM_HALT
3934	\|\| rc == VINF_SUCCESS)
3935	{
3936	hmR0SvmUpdateRip(pVCpu, pCtx, 3);
3937
3938	if ( rc == VINF_EM_HALT
3939	&& EMShouldContinueAfterHalt(pVCpu, pCtx))
3940	{
3941	rc = VINF_SUCCESS;
3942	}
3943	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
3944	}
3945	else
3946	{
3947	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMwait: EMInterpretMWait failed with %Rrc\n", rc));
3948	rc = VERR_EM_INTERPRETER;
3949	}
3950	AssertMsg(rc == VINF_SUCCESS \|\| rc == VINF_EM_HALT \|\| rc == VERR_EM_INTERPRETER,
3951	("hmR0SvmExitMwait: EMInterpretMWait failed rc=%Rrc\n", rc));
3952	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitMwait);
3953	return rc;
3954	}
3955
3956
3957	/**
3958	* #VMEXIT handler for shutdown (triple-fault) (SVM_EXIT_SHUTDOWN).
3959	* Conditional #VMEXIT.
3960	*/
3961	HMSVM_EXIT_DECL hmR0SvmExitShutdown(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
3962	{
3963	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
3964	return VINF_EM_RESET;
3965	}
3966
3967
3968	/**
3969	* #VMEXIT handler for CRx reads (SVM_EXIT_READ_CR*). Conditional #VMEXIT.
3970	*/
3971	HMSVM_EXIT_DECL hmR0SvmExitReadCRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
3972	{
3973	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
3974
3975	Log4(("hmR0SvmExitReadCRx: CS:RIP=%04x:%#RX64\n", pCtx->cs.Sel, pCtx->rip));
3976
3977	/** @todo Decode Assist. */
3978	VBOXSTRICTRC rc2 = EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */);
3979	int rc = VBOXSTRICTRC_VAL(rc2);
3980	AssertMsg(rc == VINF_SUCCESS \|\| rc == VERR_EM_INTERPRETER \|\| rc == VINF_PGM_CHANGE_MODE \|\| rc == VINF_PGM_SYNC_CR3,
3981	("hmR0SvmExitReadCRx: EMInterpretInstruction failed rc=%Rrc\n", rc));
3982	Assert((pSvmTransient->u64ExitCode - SVM_EXIT_READ_CR0) <= 15);
3983	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCRxRead[pSvmTransient->u64ExitCode - SVM_EXIT_READ_CR0]);
3984	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
3985	return rc;
3986	}
3987
3988
3989	/**
3990	* #VMEXIT handler for CRx writes (SVM_EXIT_WRITE_CR*). Conditional #VMEXIT.
3991	*/
3992	HMSVM_EXIT_DECL hmR0SvmExitWriteCRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
3993	{
3994	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
3995	/** @todo Decode Assist. */
3996	VBOXSTRICTRC rc2 = EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */);
3997	int rc = VBOXSTRICTRC_VAL(rc2);
3998	if (rc == VINF_SUCCESS)
3999	{
4000	/* RIP has been updated by EMInterpretInstruction(). */
4001	Assert((pSvmTransient->u64ExitCode - SVM_EXIT_WRITE_CR0) <= 15);
4002	switch (pSvmTransient->u64ExitCode - SVM_EXIT_WRITE_CR0)
4003	{
4004	case 0: /* CR0. */
4005	pVCpu->hm.s.fContextUseFlags \|= HM_CHANGED_GUEST_CR0;
4006	break;
4007
4008	case 3: /* CR3. */
4009	Assert(!pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
4010	pVCpu->hm.s.fContextUseFlags \|= HM_CHANGED_GUEST_CR3;
4011	break;
4012
4013	case 4: /* CR4. */
4014	pVCpu->hm.s.fContextUseFlags \|= HM_CHANGED_GUEST_CR4;
4015	break;
4016
4017	case 8: /* CR8 (TPR). */
4018	pVCpu->hm.s.fContextUseFlags \|= HM_CHANGED_SVM_GUEST_APIC_STATE;
4019	break;
4020
4021	default:
4022	AssertMsgFailed(("hmR0SvmExitWriteCRx: Invalid/Unexpected Write-CRx exit. u64ExitCode=%#RX64 %#x CRx=%#RX64\n",
4023	pSvmTransient->u64ExitCode, pSvmTransient->u64ExitCode - SVM_EXIT_WRITE_CR0));
4024	break;
4025	}
4026	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4027	}
4028	else
4029	Assert(rc == VERR_EM_INTERPRETER \|\| rc == VINF_PGM_CHANGE_MODE \|\| rc == VINF_PGM_SYNC_CR3);
4030	return rc;
4031	}
4032
4033
4034	/**
4035	* #VMEXIT handler for instructions that result in a #UD exception delivered to
4036	* the guest.
4037	*/
4038	HMSVM_EXIT_DECL hmR0SvmExitSetPendingXcptUD(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4039	{
4040	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4041	hmR0SvmSetPendingXcptUD(pVCpu);
4042	return VINF_SUCCESS;
4043	}
4044
4045
4046	/**
4047	* #VMEXIT handler for MSR read and writes (SVM_EXIT_MSR). Conditional #VMEXIT.
4048	*/
4049	HMSVM_EXIT_DECL hmR0SvmExitMsr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4050	{
4051	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4052	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4053	PVM pVM = pVCpu->CTX_SUFF(pVM);
4054
4055	int rc;
4056	if (pVmcb->ctrl.u64ExitInfo1 == SVM_EXIT1_MSR_WRITE)
4057	{
4058	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitWrmsr);
4059
4060	/* Handle TPR patching; intercepted LSTAR write. */
4061	if ( pVM->hm.s.fTPRPatchingActive
4062	&& pCtx->ecx == MSR_K8_LSTAR)
4063	{
4064	if ((pCtx->eax & 0xff) != pSvmTransient->u8GuestTpr)
4065	{
4066	/* Our patch code uses LSTAR for TPR caching for 32-bit guests. */
4067	int rc2 = PDMApicSetTPR(pVCpu, pCtx->eax & 0xff);
4068	AssertRC(rc2);
4069	pVCpu->hm.s.fContextUseFlags \|= HM_CHANGED_SVM_GUEST_APIC_STATE;
4070	}
4071	hmR0SvmUpdateRip(pVCpu, pCtx, 2);
4072	rc = VINF_SUCCESS;
4073	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4074	return rc;
4075	}
4076
4077	if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_NRIP_SAVE)
4078	{
4079	rc = EMInterpretWrmsr(pVM, pVCpu, CPUMCTX2CORE(pCtx));
4080	if (RT_LIKELY(rc == VINF_SUCCESS))
4081	{
4082	pCtx->rip = pVmcb->ctrl.u64NextRIP;
4083	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4084	}
4085	else
4086	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMsr: EMInterpretWrmsr failed rc=%Rrc\n", rc));
4087	}
4088	else
4089	{
4090	rc = VBOXSTRICTRC_TODO(EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */));
4091	if (RT_UNLIKELY(rc != VINF_SUCCESS))
4092	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMsr: WrMsr. EMInterpretInstruction failed rc=%Rrc\n", rc));
4093	/* RIP updated by EMInterpretInstruction(). */
4094	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4095	}
4096
4097	/* If this is an X2APIC WRMSR access, update the APIC state as well. */
4098	if ( pCtx->ecx >= MSR_IA32_X2APIC_START
4099	&& pCtx->ecx <= MSR_IA32_X2APIC_END)
4100	{
4101	/* We've already saved the APIC related guest-state (TPR) in hmR0SvmPostRunGuest(). When full APIC register
4102	* virtualization is implemented we'll have to make sure APIC state is saved from the VMCB before
4103	EMInterpretWrmsr() changes it. */
4104	pVCpu->hm.s.fContextUseFlags \|= HM_CHANGED_SVM_GUEST_APIC_STATE;
4105	}
4106	else if (pCtx->ecx == MSR_K6_EFER)
4107	pVCpu->hm.s.fContextUseFlags \|= HM_CHANGED_SVM_GUEST_EFER_MSR;
4108	else if (pCtx->ecx == MSR_IA32_TSC)
4109	pSvmTransient->fUpdateTscOffsetting = true;
4110	}
4111	else
4112	{
4113	/* MSR Read access. */
4114	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdmsr);
4115	Assert(pVmcb->ctrl.u64ExitInfo1 == SVM_EXIT1_MSR_READ);
4116
4117	if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_NRIP_SAVE)
4118	{
4119	rc = EMInterpretRdmsr(pVM, pVCpu, CPUMCTX2CORE(pCtx));
4120	if (RT_LIKELY(rc == VINF_SUCCESS))
4121	{
4122	pCtx->rip = pVmcb->ctrl.u64NextRIP;
4123	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4124	}
4125	else
4126	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMsr: EMInterpretRdmsr failed rc=%Rrc\n", rc));
4127	}
4128	else
4129	{
4130	rc = VBOXSTRICTRC_TODO(EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0));
4131	if (RT_UNLIKELY(rc != VINF_SUCCESS))
4132	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMsr: RdMsr. EMInterpretInstruction failed rc=%Rrc\n", rc));
4133	/* RIP updated by EMInterpretInstruction(). */
4134	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4135	}
4136	}
4137
4138	/* RIP has been updated by EMInterpret[Rd\|Wr]msr(). */
4139	return rc;
4140	}
4141
4142
4143	/**
4144	* #VMEXIT handler for DRx read (SVM_EXIT_READ_DRx). Conditional #VMEXIT.
4145	*/
4146	HMSVM_EXIT_DECL hmR0SvmExitReadDRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4147	{
4148	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4149	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitDRxRead);
4150
4151	/* We should -not- get this VM-exit if the guest is debugging. */
4152	AssertMsgReturn(!CPUMIsGuestDebugStateActive(pVCpu),
4153	("hmR0SvmExitReadDRx: Unexpected exit. pVCpu=%p pCtx=%p\n", pVCpu, pCtx),
4154	VERR_SVM_UNEXPECTED_EXIT);
4155
4156	/*
4157	* Lazy DR0-3 loading?
4158	*/
4159	if (!CPUMIsHyperDebugStateActive(pVCpu))
4160	{
4161	Assert(!DBGFIsStepping(pVCpu)); Assert(!pVCpu->hm.s.fSingleInstruction);
4162	Log5(("hmR0SvmExitReadDRx: Lazy loading guest debug registers\n"));
4163
4164	/* Don't intercept DRx read and writes. */
4165	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4166	pVmcb->ctrl.u16InterceptRdDRx = 0;
4167	pVmcb->ctrl.u16InterceptWrDRx = 0;
4168	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
4169
4170	/* Save the host & load the guest debug state, restart execution of the MOV DRx instruction. */
4171	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
4172	Assert(CPUMIsGuestDebugStateActive(pVCpu) \|\| HC_ARCH_BITS == 32);
4173
4174	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxContextSwitch);
4175	return VINF_SUCCESS;
4176	}
4177
4178	/*
4179	* Interpret the read/writing of DRx.
4180	*/
4181	/** @todo Decode assist. */
4182	VBOXSTRICTRC rc = EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */);
4183	Log5(("hmR0SvmExitReadDRx: Emulatined DRx access: rc=%Rrc\n", VBOXSTRICTRC_VAL(rc)));
4184	if (RT_LIKELY(rc == VINF_SUCCESS))
4185	{
4186	/* Not necessary for read accesses but whatever doesn't hurt for now, will be fixed with decode assist. */
4187	/** @todo CPUM should set this flag! */
4188	pVCpu->hm.s.fContextUseFlags \|= HM_CHANGED_GUEST_DEBUG;
4189	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4190	}
4191	else
4192	Assert(rc == VERR_EM_INTERPRETER);
4193	return VBOXSTRICTRC_TODO(rc);
4194	}
4195
4196
4197	/**
4198	* #VMEXIT handler for DRx write (SVM_EXIT_WRITE_DRx). Conditional #VMEXIT.
4199	*/
4200	HMSVM_EXIT_DECL hmR0SvmExitWriteDRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4201	{
4202	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4203	/* For now it's the same since we interpret the instruction anyway. Will change when using of Decode Assist is implemented. */
4204	int rc = hmR0SvmExitReadDRx(pVCpu, pCtx, pSvmTransient);
4205	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitDRxWrite);
4206	STAM_COUNTER_DEC(&pVCpu->hm.s.StatExitDRxRead);
4207	return rc;
4208	}
4209
4210
4211	/**
4212	* #VMEXIT handler for I/O instructions (SVM_EXIT_IOIO). Conditional #VMEXIT.
4213	*/
4214	HMSVM_EXIT_DECL hmR0SvmExitIOInstr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4215	{
4216	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4217
4218	/* I/O operation lookup arrays. */
4219	static uint32_t const s_aIOSize[8] = { 0, 1, 2, 0, 4, 0, 0, 0 }; /* Size of the I/O accesses in bytes. */
4220	static uint32_t const s_aIOOpAnd[8] = { 0, 0xff, 0xffff, 0, 0xffffffff, 0, 0, 0 }; /* AND masks for saving
4221	the result (in AL/AX/EAX). */
4222	Log4(("hmR0SvmExitIOInstr: CS:RIP=%04x:%#RX64\n", pCtx->cs.Sel, pCtx->rip));
4223
4224	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4225	PVM pVM = pVCpu->CTX_SUFF(pVM);
4226
4227	/* Refer AMD spec. 15.10.2 "IN and OUT Behaviour" and Figure 15-2. "EXITINFO1 for IOIO Intercept" for the format. */
4228	SVMIOIOEXIT IoExitInfo;
4229	IoExitInfo.u = (uint32_t)pVmcb->ctrl.u64ExitInfo1;
4230	uint32_t uIOWidth = (IoExitInfo.u >> 4) & 0x7;
4231	uint32_t cbValue = s_aIOSize[uIOWidth];
4232	uint32_t uAndVal = s_aIOOpAnd[uIOWidth];
4233
4234	if (RT_UNLIKELY(!cbValue))
4235	{
4236	AssertMsgFailed(("hmR0SvmExitIOInstr: Invalid IO operation. uIOWidth=%u\n", uIOWidth));
4237	return VERR_EM_INTERPRETER;
4238	}
4239
4240	VBOXSTRICTRC rcStrict;
4241	if (IoExitInfo.n.u1STR)
4242	{
4243	/* INS/OUTS - I/O String instruction. */
4244	PDISCPUSTATE pDis = &pVCpu->hm.s.DisState;
4245
4246	/** @todo Huh? why can't we use the segment prefix information given by AMD-V
4247	* in EXITINFO1? Investigate once this thing is up and running. */
4248
4249	rcStrict = EMInterpretDisasCurrent(pVM, pVCpu, pDis, NULL);
4250	if (rcStrict == VINF_SUCCESS)
4251	{
4252	if (IoExitInfo.n.u1Type == SVM_IOIO_WRITE)
4253	{
4254	rcStrict = IOMInterpretOUTSEx(pVM, pVCpu, CPUMCTX2CORE(pCtx), IoExitInfo.n.u16Port, pDis->fPrefix,
4255	(DISCPUMODE)pDis->uAddrMode, cbValue);
4256	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOStringWrite);
4257	}
4258	else
4259	{
4260	rcStrict = IOMInterpretINSEx(pVM, pVCpu, CPUMCTX2CORE(pCtx), IoExitInfo.n.u16Port, pDis->fPrefix,
4261	(DISCPUMODE)pDis->uAddrMode, cbValue);
4262	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOStringRead);
4263	}
4264	}
4265	else
4266	rcStrict = VINF_EM_RAW_EMULATE_INSTR;
4267	}
4268	else
4269	{
4270	/* IN/OUT - I/O instruction. */
4271	Assert(!IoExitInfo.n.u1REP);
4272
4273	if (IoExitInfo.n.u1Type == SVM_IOIO_WRITE)
4274	{
4275	rcStrict = IOMIOPortWrite(pVM, pVCpu, IoExitInfo.n.u16Port, pCtx->eax & uAndVal, cbValue);
4276	if (rcStrict == VINF_IOM_R3_IOPORT_WRITE)
4277	HMR0SavePendingIOPortWrite(pVCpu, pCtx->rip, pVmcb->ctrl.u64ExitInfo2, IoExitInfo.n.u16Port, uAndVal, cbValue);
4278
4279	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOWrite);
4280	}
4281	else
4282	{
4283	uint32_t u32Val = 0;
4284
4285	rcStrict = IOMIOPortRead(pVM, pVCpu, IoExitInfo.n.u16Port, &u32Val, cbValue);
4286	if (IOM_SUCCESS(rcStrict))
4287	{
4288	/* Save result of I/O IN instr. in AL/AX/EAX. */
4289	pCtx->eax = (pCtx->eax & ~uAndVal) \| (u32Val & uAndVal);
4290	}
4291	else if (rcStrict == VINF_IOM_R3_IOPORT_READ)
4292	HMR0SavePendingIOPortRead(pVCpu, pCtx->rip, pVmcb->ctrl.u64ExitInfo2, IoExitInfo.n.u16Port, uAndVal, cbValue);
4293
4294	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIORead);
4295	}
4296	}
4297
4298	if (IOM_SUCCESS(rcStrict))
4299	{
4300	/* AMD-V saves the RIP of the instruction following the IO instruction in EXITINFO2. */
4301	pCtx->rip = pVmcb->ctrl.u64ExitInfo2;
4302
4303	/*
4304	* If any I/O breakpoints are armed, we need to check if one triggered
4305	* and take appropriate action.
4306	* Note that the I/O breakpoint type is undefined if CR4.DE is 0.
4307	*/
4308	/** @todo Optimize away the DBGFBpIsHwIoArmed call by having DBGF tell the
4309	* execution engines about whether hyper BPs and such are pending. */
4310	uint32_t const uDr7 = pCtx->dr[7];
4311	if (RT_UNLIKELY( ( (uDr7 & X86_DR7_ENABLED_MASK)
4312	&& X86_DR7_ANY_RW_IO(uDr7)
4313	&& (pCtx->cr4 & X86_CR4_DE))
4314	\|\| DBGFBpIsHwIoArmed(pVM)))
4315	{
4316	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxIoCheck);
4317	CPUMR0DebugStateMaybeSaveGuest(pVCpu, false /fDr6/);
4318
4319	VBOXSTRICTRC rcStrict2 = DBGFBpCheckIo(pVM, pVCpu, pCtx, IoExitInfo.n.u16Port, cbValue);
4320	if (rcStrict2 == VINF_EM_RAW_GUEST_TRAP)
4321	{
4322	/* Raise #DB. */
4323	pVmcb->guest.u64DR6 = pCtx->dr[6];
4324	pVmcb->guest.u64DR7 = pCtx->dr[7];
4325	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
4326	hmR0SvmSetPendingXcptDB(pVCpu);
4327	}
4328	/* rcStrict is VINF_SUCCESS or in [VINF_EM_FIRST..VINF_EM_LAST]. */
4329	else if ( rcStrict2 != VINF_SUCCESS
4330	&& (rcStrict == VINF_SUCCESS \|\| rcStrict2 < rcStrict))
4331	rcStrict = rcStrict2;
4332	}
4333
4334	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
4335	}
4336
4337	#ifdef VBOX_STRICT
4338	if (rcStrict == VINF_IOM_R3_IOPORT_READ)
4339	Assert(IoExitInfo.n.u1Type == SVM_IOIO_READ);
4340	else if (rcStrict == VINF_IOM_R3_IOPORT_WRITE)
4341	Assert(IoExitInfo.n.u1Type == SVM_IOIO_WRITE);
4342	else
4343	{
4344	/** @todo r=bird: This is missing a bunch of VINF_EM_FIRST..VINF_EM_LAST
4345	* statuses, that the VMM device and some others may return. See
4346	* IOM_SUCCESS() for guidance. */
4347	AssertMsg( RT_FAILURE(rcStrict)
4348	\|\| rcStrict == VINF_SUCCESS
4349	\|\| rcStrict == VINF_EM_RAW_EMULATE_INSTR
4350	\|\| rcStrict == VINF_EM_DBG_BREAKPOINT
4351	\|\| rcStrict == VINF_EM_RAW_GUEST_TRAP
4352	\|\| rcStrict == VINF_TRPM_XCPT_DISPATCHED, ("%Rrc\n", VBOXSTRICTRC_VAL(rcStrict)));
4353	}
4354	#endif
4355	return VBOXSTRICTRC_TODO(rcStrict);
4356	}
4357
4358
4359	/**
4360	* #VMEXIT handler for Nested Page-faults (SVM_EXIT_NPF). Conditional
4361	* #VMEXIT.
4362	*/
4363	HMSVM_EXIT_DECL hmR0SvmExitNestedPF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4364	{
4365	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4366	PVM pVM = pVCpu->CTX_SUFF(pVM);
4367	Assert(pVM->hm.s.fNestedPaging);
4368
4369	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
4370
4371	/* See AMD spec. 15.25.6 "Nested versus Guest Page Faults, Fault Ordering" for VMCB details for #NPF. */
4372	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4373	uint32_t u32ErrCode = pVmcb->ctrl.u64ExitInfo1;
4374	RTGCPHYS GCPhysFaultAddr = pVmcb->ctrl.u64ExitInfo2;
4375
4376	Log4(("#NPF at CS:RIP=%04x:%#RX64 faultaddr=%RGp errcode=%#x \n", pCtx->cs.Sel, pCtx->rip, GCPhysFaultAddr, u32ErrCode));
4377
4378	#ifdef VBOX_HM_WITH_GUEST_PATCHING
4379	/* TPR patching for 32-bit guests, using the reserved bit in the page tables for MMIO regions. */
4380	if ( pVM->hm.s.fTRPPatchingAllowed
4381	&& (GCPhysFaultAddr & PAGE_OFFSET_MASK) == 0x80 /* TPR offset. */
4382	&& ( !(u32ErrCode & X86_TRAP_PF_P) /* Not present */
4383	\|\| (u32ErrCode & (X86_TRAP_PF_P \| X86_TRAP_PF_RSVD)) == (X86_TRAP_PF_P \| X86_TRAP_PF_RSVD)) /* MMIO page. */
4384	&& !CPUMGetGuestCPL(pVCpu)
4385	&& !CPUMIsGuestInLongModeEx(pCtx)
4386	&& pVM->hm.s.cPatches < RT_ELEMENTS(pVM->hm.s.aPatches))
4387	{
4388	RTGCPHYS GCPhysApicBase = pCtx->msrApicBase;
4389	GCPhysApicBase &= PAGE_BASE_GC_MASK;
4390
4391	if (GCPhysFaultAddr == GCPhysApicBase + 0x80)
4392	{
4393	/* Only attempt to patch the instruction once. */
4394	PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
4395	if (!pPatch)
4396	return VINF_EM_HM_PATCH_TPR_INSTR;
4397	}
4398	}
4399	#endif
4400
4401	/*
4402	* Determine the nested paging mode.
4403	*/
4404	PGMMODE enmNestedPagingMode;
4405	#if HC_ARCH_BITS == 32
4406	if (CPUMIsGuestInLongModeEx(pCtx))
4407	enmNestedPagingMode = PGMMODE_AMD64_NX;
4408	else
4409	#endif
4410	enmNestedPagingMode = PGMGetHostMode(pVM);
4411
4412	/*
4413	* MMIO optimization using the reserved (RSVD) bit in the guest page tables for MMIO pages.
4414	*/
4415	int rc;
4416	Assert((u32ErrCode & (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P)) != X86_TRAP_PF_RSVD);
4417	if ((u32ErrCode & (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P)) == (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P))
4418	{
4419	VBOXSTRICTRC rc2 = PGMR0Trap0eHandlerNPMisconfig(pVM, pVCpu, enmNestedPagingMode, CPUMCTX2CORE(pCtx), GCPhysFaultAddr,
4420	u32ErrCode);
4421	rc = VBOXSTRICTRC_VAL(rc2);
4422
4423	/*
4424	* If we succeed, resume guest execution.
4425	* If we fail in interpreting the instruction because we couldn't get the guest physical address
4426	* of the page containing the instruction via the guest's page tables (we would invalidate the guest page
4427	* in the host TLB), resume execution which would cause a guest page fault to let the guest handle this
4428	* weird case. See @bugref{6043}.
4429	*/
4430	if ( rc == VINF_SUCCESS
4431	\|\| rc == VERR_PAGE_TABLE_NOT_PRESENT
4432	\|\| rc == VERR_PAGE_NOT_PRESENT)
4433	{
4434	/* Successfully handled MMIO operation. */
4435	pVCpu->hm.s.fContextUseFlags \|= HM_CHANGED_SVM_GUEST_APIC_STATE;
4436	rc = VINF_SUCCESS;
4437	}
4438	return rc;
4439	}
4440
4441	TRPMAssertXcptPF(pVCpu, GCPhysFaultAddr, u32ErrCode);
4442	rc = PGMR0Trap0eHandlerNestedPaging(pVM, pVCpu, enmNestedPagingMode, u32ErrCode, CPUMCTX2CORE(pCtx), GCPhysFaultAddr);
4443	TRPMResetTrap(pVCpu);
4444
4445	Log4(("#NPF: PGMR0Trap0eHandlerNestedPaging returned %Rrc CS:RIP=%04x:%#RX64\n", rc, pCtx->cs.Sel, pCtx->rip));
4446
4447	/*
4448	* Same case as PGMR0Trap0eHandlerNPMisconfig(). See comment above, @bugref{6043}.
4449	*/
4450	if ( rc == VINF_SUCCESS
4451	\|\| rc == VERR_PAGE_TABLE_NOT_PRESENT
4452	\|\| rc == VERR_PAGE_NOT_PRESENT)
4453	{
4454	/* We've successfully synced our shadow page tables. */
4455	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPF);
4456	rc = VINF_SUCCESS;
4457	}
4458
4459	return rc;
4460	}
4461
4462
4463	/**
4464	* #VMEXIT handler for virtual interrupt (SVM_EXIT_VINTR). Conditional #VMEXIT.
4465	*/
4466	HMSVM_EXIT_DECL hmR0SvmExitVIntr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4467	{
4468	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4469
4470	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4471	pVmcb->ctrl.IntCtrl.n.u1VIrqValid = 0; /* No virtual interrupts pending, we'll inject the current one before reentry. */
4472	pVmcb->ctrl.IntCtrl.n.u8VIrqVector = 0;
4473
4474	/* Indicate that we no longer need to VM-exit when the guest is ready to receive interrupts, it is now ready. */
4475	pVmcb->ctrl.u32InterceptCtrl1 &= ~SVM_CTRL1_INTERCEPT_VINTR;
4476	pVmcb->ctrl.u64VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS \| HMSVM_VMCB_CLEAN_TPR);
4477
4478	/* Deliver the pending interrupt via hmR0SvmPreRunGuest()->hmR0SvmInjectEventVmcb() and resume guest execution. */
4479	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIntWindow);
4480	return VINF_SUCCESS;
4481	}
4482
4483
4484	/**
4485	* #VMEXIT handler for task switches (SVM_EXIT_TASK_SWITCH). Conditional #VMEXIT.
4486	*/
4487	HMSVM_EXIT_DECL hmR0SvmExitTaskSwitch(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4488	{
4489	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4490
4491	#ifndef HMSVM_ALWAYS_TRAP_TASK_SWITCH
4492	Assert(!pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
4493	#endif
4494
4495	/* Check if this task-switch occurred while delivery an event through the guest IDT. */
4496	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4497	if ( !(pVmcb->ctrl.u64ExitInfo2 & (SVM_EXIT2_TASK_SWITCH_IRET \| SVM_EXIT2_TASK_SWITCH_JMP))
4498	&& pVCpu->hm.s.Event.fPending)
4499	{
4500	/*
4501	* AMD-V does not provide us with the original exception but we have it in u64IntrInfo since we
4502	* injected the event during VM-entry. Software interrupts and exceptions will be regenerated
4503	* when the recompiler restarts the instruction.
4504	*/
4505	SVMEVENT Event;
4506	Event.u = pVCpu->hm.s.Event.u64IntrInfo;
4507	if ( Event.n.u3Type == SVM_EVENT_EXCEPTION
4508	\|\| Event.n.u3Type == SVM_EVENT_SOFTWARE_INT)
4509	{
4510	pVCpu->hm.s.Event.fPending = false;
4511	}
4512	else
4513	Log4(("hmR0SvmExitTaskSwitch: TS occurred during event delivery. Kept pending u8Vector=%#x\n", Event.n.u8Vector));
4514	}
4515
4516	/** @todo Emulate task switch someday, currently just going back to ring-3 for
4517	* emulation. */
4518	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitTaskSwitch);
4519	return VERR_EM_INTERPRETER;
4520	}
4521
4522
4523	/**
4524	* #VMEXIT handler for VMMCALL (SVM_EXIT_VMMCALL). Conditional #VMEXIT.
4525	*/
4526	HMSVM_EXIT_DECL hmR0SvmExitVmmCall(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4527	{
4528	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4529
4530	int rc = hmR0SvmEmulateMovTpr(pVCpu->CTX_SUFF(pVM), pVCpu, pCtx);
4531	if (RT_LIKELY(rc == VINF_SUCCESS))
4532	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4533	else
4534	hmR0SvmSetPendingXcptUD(pVCpu);
4535	return VINF_SUCCESS;
4536	}
4537
4538
4539	/**
4540	* #VMEXIT handler for page-fault exceptions (SVM_EXIT_EXCEPTION_E). Conditional
4541	* #VMEXIT.
4542	*/
4543	HMSVM_EXIT_DECL hmR0SvmExitXcptPF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4544	{
4545	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4546
4547	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
4548
4549	/* See AMD spec. 15.12.15 "#PF (Page Fault)". */
4550	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4551	uint32_t u32ErrCode = pVmcb->ctrl.u64ExitInfo1;
4552	RTGCUINTPTR uFaultAddress = pVmcb->ctrl.u64ExitInfo2;
4553	PVM pVM = pVCpu->CTX_SUFF(pVM);
4554
4555	#if defined(HMSVM_ALWAYS_TRAP_ALL_XCPTS) \|\| defined(HMSVM_ALWAYS_TRAP_PF)
4556	if (pVM->hm.s.fNestedPaging)
4557	{
4558	pVCpu->hm.s.Event.fPending = false; /* In case it's a contributory or vectoring #PF. */
4559	if (!pSvmTransient->fVectoringPF)
4560	{
4561	/* A genuine guest #PF, reflect it to the guest. */
4562	hmR0SvmSetPendingXcptPF(pVCpu, pCtx, u32ErrCode, uFaultAddress);
4563	Log4(("#PF: Guest page fault at %04X:%RGv FaultAddr=%RGv ErrCode=%#x\n", pCtx->cs.Sel, (RTGCPTR)pCtx->rip,
4564	uFaultAddress, u32ErrCode));
4565	}
4566	else
4567	{
4568	/* A guest page-fault occurred during delivery of a page-fault. Inject #DF. */
4569	hmR0SvmSetPendingXcptDF(pVCpu);
4570	Log4(("Pending #DF due to vectoring #PF. NP\n"));
4571	}
4572	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestPF);
4573	return VINF_SUCCESS;
4574	}
4575	#endif
4576
4577	Assert(!pVM->hm.s.fNestedPaging);
4578
4579	#ifdef VBOX_HM_WITH_GUEST_PATCHING
4580	/* Shortcut for APIC TPR reads and writes; only applicable to 32-bit guests. */
4581	if ( pVM->hm.s.fTRPPatchingAllowed
4582	&& (uFaultAddress & 0xfff) == 0x80 /* TPR offset. */
4583	&& !(u32ErrCode & X86_TRAP_PF_P) /* Not present. */
4584	&& !CPUMGetGuestCPL(pVCpu)
4585	&& !CPUMIsGuestInLongModeEx(pCtx)
4586	&& pVM->hm.s.cPatches < RT_ELEMENTS(pVM->hm.s.aPatches))
4587	{
4588	RTGCPHYS GCPhysApicBase;
4589	GCPhysApicBase = pCtx->msrApicBase;
4590	GCPhysApicBase &= PAGE_BASE_GC_MASK;
4591
4592	/* Check if the page at the fault-address is the APIC base. */
4593	RTGCPHYS GCPhysPage;
4594	int rc2 = PGMGstGetPage(pVCpu, (RTGCPTR)uFaultAddress, NULL /* pfFlags */, &GCPhysPage);
4595	if ( rc2 == VINF_SUCCESS
4596	&& GCPhysPage == GCPhysApicBase)
4597	{
4598	/* Only attempt to patch the instruction once. */
4599	PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
4600	if (!pPatch)
4601	return VINF_EM_HM_PATCH_TPR_INSTR;
4602	}
4603	}
4604	#endif
4605
4606	Log4(("#PF: uFaultAddress=%#RX64 CS:RIP=%#04x:%#RX64 u32ErrCode %#RX32 cr3=%#RX64\n", uFaultAddress, pCtx->cs.Sel,
4607	pCtx->rip, u32ErrCode, pCtx->cr3));
4608
4609	TRPMAssertXcptPF(pVCpu, uFaultAddress, u32ErrCode);
4610	int rc = PGMTrap0eHandler(pVCpu, u32ErrCode, CPUMCTX2CORE(pCtx), (RTGCPTR)uFaultAddress);
4611
4612	Log4(("#PF rc=%Rrc\n", rc));
4613
4614	if (rc == VINF_SUCCESS)
4615	{
4616	/* Successfully synced shadow pages tables or emulated an MMIO instruction. */
4617	TRPMResetTrap(pVCpu);
4618	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPF);
4619	pVCpu->hm.s.fContextUseFlags \|= HM_CHANGED_SVM_GUEST_APIC_STATE;
4620	return rc;
4621	}
4622	else if (rc == VINF_EM_RAW_GUEST_TRAP)
4623	{
4624	pVCpu->hm.s.Event.fPending = false; /* In case it's a contributory or vectoring #PF. */
4625
4626	if (!pSvmTransient->fVectoringPF)
4627	{
4628	/* It's a guest page fault and needs to be reflected to the guest. */
4629	u32ErrCode = TRPMGetErrorCode(pVCpu); /* The error code might have been changed. */
4630	TRPMResetTrap(pVCpu);
4631	hmR0SvmSetPendingXcptPF(pVCpu, pCtx, u32ErrCode, uFaultAddress);
4632	}
4633	else
4634	{
4635	/* A guest page-fault occurred during delivery of a page-fault. Inject #DF. */
4636	TRPMResetTrap(pVCpu);
4637	hmR0SvmSetPendingXcptDF(pVCpu);
4638	Log4(("#PF: Pending #DF due to vectoring #PF\n"));
4639	}
4640
4641	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestPF);
4642	return VINF_SUCCESS;
4643	}
4644
4645	TRPMResetTrap(pVCpu);
4646	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPFEM);
4647	return rc;
4648	}
4649
4650
4651	/**
4652	* #VMEXIT handler for device-not-available exceptions (SVM_EXIT_EXCEPTION_7).
4653	* Conditional #VMEXIT.
4654	*/
4655	HMSVM_EXIT_DECL hmR0SvmExitXcptNM(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4656	{
4657	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4658
4659	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
4660
4661	#ifndef HMSVM_ALWAYS_TRAP_ALL_XCPTS
4662	Assert(!CPUMIsGuestFPUStateActive(pVCpu));
4663	#endif
4664
4665	/* Lazy FPU loading; load the guest-FPU state transparently and continue execution of the guest. */
4666	int rc = CPUMR0LoadGuestFPU(pVCpu->CTX_SUFF(pVM), pVCpu, pCtx);
4667	if (rc == VINF_SUCCESS)
4668	{
4669	Assert(CPUMIsGuestFPUStateActive(pVCpu));
4670	pVCpu->hm.s.fContextUseFlags \|= HM_CHANGED_GUEST_CR0;
4671	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowNM);
4672	return VINF_SUCCESS;
4673	}
4674
4675	/* Forward #NM to the guest. */
4676	Assert(rc == VINF_EM_RAW_GUEST_TRAP);
4677	hmR0SvmSetPendingXcptNM(pVCpu);
4678	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestNM);
4679	return VINF_SUCCESS;
4680	}
4681
4682
4683	/**
4684	* #VMEXIT handler for math-fault exceptions (SVM_EXIT_EXCEPTION_10).
4685	* Conditional #VMEXIT.
4686	*/
4687	HMSVM_EXIT_DECL hmR0SvmExitXcptMF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4688	{
4689	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4690
4691	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
4692
4693	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestMF);
4694
4695	if (!(pCtx->cr0 & X86_CR0_NE))
4696	{
4697	/* Old-style FPU error reporting needs some extra work. */
4698	/** @todo don't fall back to the recompiler, but do it manually. */
4699	return VERR_EM_INTERPRETER;
4700	}
4701
4702	hmR0SvmSetPendingXcptMF(pVCpu);
4703	return VINF_SUCCESS;
4704	}
4705
4706
4707	/**
4708	* #VMEXIT handler for debug exceptions (SVM_EXIT_EXCEPTION_1). Conditional
4709	* #VMEXIT.
4710	*/
4711	HMSVM_EXIT_DECL hmR0SvmExitXcptDB(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4712	{
4713	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4714
4715	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
4716
4717	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestDB);
4718
4719	/* If we set the trap flag above, we have to clear it. */
4720	if (pVCpu->hm.s.fClearTrapFlag)
4721	{
4722	pVCpu->hm.s.fClearTrapFlag = false;
4723	pCtx->eflags.Bits.u1TF = 0;
4724	}
4725
4726	/* This can be a fault-type #DB (instruction breakpoint) or a trap-type #DB (data breakpoint). However, for both cases
4727	DR6 and DR7 are updated to what the exception handler expects. See AMD spec. 15.12.2 "#DB (Debug)". */
4728	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4729	PVM pVM = pVCpu->CTX_SUFF(pVM);
4730	int rc = DBGFRZTrap01Handler(pVM, pVCpu, CPUMCTX2CORE(pCtx), pVmcb->guest.u64DR6, pVCpu->hm.s.fSingleInstruction);
4731	if (rc == VINF_EM_RAW_GUEST_TRAP)
4732	{
4733	Log5(("hmR0SvmExitXcptDB: DR6=%#RX64 -> guest trap\n", pVmcb->guest.u64DR6));
4734	if (CPUMIsHyperDebugStateActive(pVCpu))
4735	CPUMSetGuestDR6(pVCpu, CPUMGetGuestDR6(pVCpu) \| pVmcb->guest.u64DR6);
4736
4737	/* Reflect the exception back to the guest. */
4738	hmR0SvmSetPendingXcptDB(pVCpu);
4739	rc = VINF_SUCCESS;
4740	}
4741
4742	/*
4743	* Update DR6.
4744	*/
4745	if (CPUMIsHyperDebugStateActive(pVCpu))
4746	{
4747	Log5(("hmR0SvmExitXcptDB: DR6=%#RX64 -> %Rrc\n", pVmcb->guest.u64DR6, rc));
4748	pVmcb->guest.u64DR6 = X86_DR6_INIT_VAL;
4749	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
4750	}
4751	else
4752	{
4753	AssertMsg(rc == VINF_SUCCESS, ("rc=%Rrc\n", rc));
4754	Assert(!pVCpu->hm.s.fSingleInstruction && !DBGFIsStepping(pVCpu));
4755	}
4756
4757	return rc;
4758	}
4759
4760	/** @} */
4761

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

source: vbox/trunk/src/VBox/VMM/VMMR0/HMSVMR0.cpp@ 48174

Download in other formats: