HMSVMR0.cpp@ 49969

最後變更在這個檔案從49969是 49969,由 vboxsync 提交於 11 年前
VMM/HMSVMR0: todo addressed in r91390.
屬性 svn:eol-style 設為 `native` 屬性 svn:keywords 設為 `Author Date Id Revision`
檔案大小: 193.0 KB

行
1	/* $Id: HMSVMR0.cpp 49969 2013-12-17 22:04:18Z 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.alldomusa.eu.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	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitAll); \
47	if ((u64ExitCode) == SVM_EXIT_NPF) \
48	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitReasonNpf); \
49	else \
50	STAM_COUNTER_INC(&pVCpu->hm.s.paStatExitReasonR0[(u64ExitCode) & MASK_EXITREASON_STAT]); \
51	} while (0)
52	#else
53	# define HMSVM_EXITCODE_STAM_COUNTER_INC(u64ExitCode) do { } while (0)
54	#endif
55
56	/** If we decide to use a function table approach this can be useful to
57	* switch to a "static DECLCALLBACK(int)". */
58	#define HMSVM_EXIT_DECL static int
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	/** @name Macros for loading, storing segment registers to/from the VMCB.
74	* @{ */
75	#define HMSVM_LOAD_SEG_REG(REG, reg) \
76	do \
77	{ \
78	Assert(pCtx->reg.fFlags & CPUMSELREG_FLAGS_VALID); \
79	Assert(pCtx->reg.ValidSel == pCtx->reg.Sel); \
80	pVmcb->guest.REG.u16Sel = pCtx->reg.Sel; \
81	pVmcb->guest.REG.u32Limit = pCtx->reg.u32Limit; \
82	pVmcb->guest.REG.u64Base = pCtx->reg.u64Base; \
83	pVmcb->guest.REG.u16Attr = HMSVM_CPU_2_VMCB_SEG_ATTR(pCtx->reg.Attr.u); \
84	} while (0)
85
86	#define HMSVM_SAVE_SEG_REG(REG, reg) \
87	do \
88	{ \
89	pMixedCtx->reg.Sel = pVmcb->guest.REG.u16Sel; \
90	pMixedCtx->reg.ValidSel = pVmcb->guest.REG.u16Sel; \
91	pMixedCtx->reg.fFlags = CPUMSELREG_FLAGS_VALID; \
92	pMixedCtx->reg.u32Limit = pVmcb->guest.REG.u32Limit; \
93	pMixedCtx->reg.u64Base = pVmcb->guest.REG.u64Base; \
94	pMixedCtx->reg.Attr.u = HMSVM_VMCB_2_CPU_SEG_ATTR(pVmcb->guest.REG.u16Attr); \
95	} while (0)
96	/** @} */
97
98	/** Macro for checking and returning from the using function for
99	* \#VMEXIT intercepts that maybe caused during delivering of another
100	* event in the guest. */
101	#define HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY() \
102	do \
103	{ \
104	int rc = hmR0SvmCheckExitDueToEventDelivery(pVCpu, pCtx, pSvmTransient); \
105	if (RT_UNLIKELY(rc == VINF_HM_DOUBLE_FAULT)) \
106	return VINF_SUCCESS; \
107	else if (RT_UNLIKELY(rc == VINF_EM_RESET)) \
108	return rc; \
109	} while (0)
110
111	/** Macro for upgrading a @a a_rc to VINF_EM_DBG_STEPPED after emulating an
112	* instruction that exited. */
113	#define HMSVM_CHECK_SINGLE_STEP(a_pVCpu, a_rc) \
114	do { \
115	if ((a_pVCpu)->hm.s.fSingleInstruction && (a_rc) == VINF_SUCCESS) \
116	(a_rc) = VINF_EM_DBG_STEPPED; \
117	} while (0)
118
119	/** Assert that preemption is disabled or covered by thread-context hooks. */
120	#define HMSVM_ASSERT_PREEMPT_SAFE() Assert( VMMR0ThreadCtxHooksAreRegistered(pVCpu) \
121	\|\| !RTThreadPreemptIsEnabled(NIL_RTTHREAD));
122
123	/** Assert that we haven't migrated CPUs when thread-context hooks are not
124	* used. */
125	#define HMSVM_ASSERT_CPU_SAFE() AssertMsg( VMMR0ThreadCtxHooksAreRegistered(pVCpu) \
126	\|\| pVCpu->hm.s.idEnteredCpu == RTMpCpuId(), \
127	("Illegal migration! Entered on CPU %u Current %u\n", \
128	pVCpu->hm.s.idEnteredCpu, RTMpCpuId()));
129
130	/** Exception bitmap mask for all contributory exceptions.
131	*
132	* Page fault is deliberately excluded here as it's conditional as to whether
133	* it's contributory or benign. Page faults are handled separately.
134	*/
135	#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) \
136	\| RT_BIT(X86_XCPT_DE))
137
138	/** @name VMCB Clean Bits.
139	*
140	* These flags are used for VMCB-state caching. A set VMCB Clean bit indicates
141	* AMD-V doesn't need to reload the corresponding value(s) from the VMCB in
142	* memory.
143	*
144	* @{ */
145	/** All intercepts vectors, TSC offset, PAUSE filter counter. */
146	#define HMSVM_VMCB_CLEAN_INTERCEPTS RT_BIT(0)
147	/** I/O permission bitmap, MSR permission bitmap. */
148	#define HMSVM_VMCB_CLEAN_IOPM_MSRPM RT_BIT(1)
149	/** ASID. */
150	#define HMSVM_VMCB_CLEAN_ASID RT_BIT(2)
151	/** TRP: V_TPR, V_IRQ, V_INTR_PRIO, V_IGN_TPR, V_INTR_MASKING,
152	V_INTR_VECTOR. */
153	#define HMSVM_VMCB_CLEAN_TPR RT_BIT(3)
154	/** Nested Paging: Nested CR3 (nCR3), PAT. */
155	#define HMSVM_VMCB_CLEAN_NP RT_BIT(4)
156	/** Control registers (CR0, CR3, CR4, EFER). */
157	#define HMSVM_VMCB_CLEAN_CRX_EFER RT_BIT(5)
158	/** Debug registers (DR6, DR7). */
159	#define HMSVM_VMCB_CLEAN_DRX RT_BIT(6)
160	/** GDT, IDT limit and base. */
161	#define HMSVM_VMCB_CLEAN_DT RT_BIT(7)
162	/** Segment register: CS, SS, DS, ES limit and base. */
163	#define HMSVM_VMCB_CLEAN_SEG RT_BIT(8)
164	/** CR2.*/
165	#define HMSVM_VMCB_CLEAN_CR2 RT_BIT(9)
166	/** Last-branch record (DbgCtlMsr, br_from, br_to, lastint_from, lastint_to) */
167	#define HMSVM_VMCB_CLEAN_LBR RT_BIT(10)
168	/** AVIC (AVIC APIC_BAR; AVIC APIC_BACKING_PAGE, AVIC
169	PHYSICAL_TABLE and AVIC LOGICAL_TABLE Pointers). */
170	#define HMSVM_VMCB_CLEAN_AVIC RT_BIT(11)
171	/** Mask of all valid VMCB Clean bits. */
172	#define HMSVM_VMCB_CLEAN_ALL ( HMSVM_VMCB_CLEAN_INTERCEPTS \
173	\| HMSVM_VMCB_CLEAN_IOPM_MSRPM \
174	\| HMSVM_VMCB_CLEAN_ASID \
175	\| HMSVM_VMCB_CLEAN_TPR \
176	\| HMSVM_VMCB_CLEAN_NP \
177	\| HMSVM_VMCB_CLEAN_CRX_EFER \
178	\| HMSVM_VMCB_CLEAN_DRX \
179	\| HMSVM_VMCB_CLEAN_DT \
180	\| HMSVM_VMCB_CLEAN_SEG \
181	\| HMSVM_VMCB_CLEAN_CR2 \
182	\| HMSVM_VMCB_CLEAN_LBR \
183	\| HMSVM_VMCB_CLEAN_AVIC)
184	/** @} */
185
186	/** @name SVM transient.
187	*
188	* A state structure for holding miscellaneous information across AMD-V
189	* VMRUN/#VMEXIT operation, restored after the transition.
190	*
191	* @{ */
192	typedef struct SVMTRANSIENT
193	{
194	/** The host's rflags/eflags. */
195	RTCCUINTREG uEflags;
196	#if HC_ARCH_BITS == 32
197	uint32_t u32Alignment0;
198	#endif
199
200	/** The #VMEXIT exit code (the EXITCODE field in the VMCB). */
201	uint64_t u64ExitCode;
202	/** The guest's TPR value used for TPR shadowing. */
203	uint8_t u8GuestTpr;
204	/** Alignment. */
205	uint8_t abAlignment0[7];
206
207	/** Whether the guest FPU state was active at the time of #VMEXIT. */
208	bool fWasGuestFPUStateActive;
209	/** Whether the guest debug state was active at the time of #VMEXIT. */
210	bool fWasGuestDebugStateActive;
211	/** Whether the hyper debug state was active at the time of #VMEXIT. */
212	bool fWasHyperDebugStateActive;
213	/** Whether the TSC offset mode needs to be updated. */
214	bool fUpdateTscOffsetting;
215	/** Whether the TSC_AUX MSR needs restoring on #VMEXIT. */
216	bool fRestoreTscAuxMsr;
217	/** Whether the #VMEXIT was caused by a page-fault during delivery of a
218	* contributary exception or a page-fault. */
219	bool fVectoringPF;
220	} SVMTRANSIENT, *PSVMTRANSIENT;
221	AssertCompileMemberAlignment(SVMTRANSIENT, u64ExitCode, sizeof(uint64_t));
222	AssertCompileMemberAlignment(SVMTRANSIENT, fWasGuestFPUStateActive, sizeof(uint64_t));
223	/** @} */
224
225	/**
226	* MSRPM (MSR permission bitmap) read permissions (for guest RDMSR).
227	*/
228	typedef enum SVMMSREXITREAD
229	{
230	/** Reading this MSR causes a VM-exit. */
231	SVMMSREXIT_INTERCEPT_READ = 0xb,
232	/** Reading this MSR does not cause a VM-exit. */
233	SVMMSREXIT_PASSTHRU_READ
234	} SVMMSREXITREAD;
235
236	/**
237	* MSRPM (MSR permission bitmap) write permissions (for guest WRMSR).
238	*/
239	typedef enum SVMMSREXITWRITE
240	{
241	/** Writing to this MSR causes a VM-exit. */
242	SVMMSREXIT_INTERCEPT_WRITE = 0xd,
243	/** Writing to this MSR does not cause a VM-exit. */
244	SVMMSREXIT_PASSTHRU_WRITE
245	} SVMMSREXITWRITE;
246
247	/**
248	* SVM VM-exit handler.
249	*
250	* @returns VBox status code.
251	* @param pVCpu Pointer to the VMCPU.
252	* @param pMixedCtx Pointer to the guest-CPU context.
253	* @param pSvmTransient Pointer to the SVM-transient structure.
254	*/
255	typedef int FNSVMEXITHANDLER(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
256
257	/*******************************************************************************
258	* Internal Functions *
259	*******************************************************************************/
260	static void hmR0SvmSetMsrPermission(PVMCPU pVCpu, unsigned uMsr, SVMMSREXITREAD enmRead, SVMMSREXITWRITE enmWrite);
261	static void hmR0SvmPendingEventToTrpmTrap(PVMCPU pVCpu);
262	static void hmR0SvmLeave(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx);
263
264	/** @name VM-exit handlers.
265	* @{
266	*/
267	static FNSVMEXITHANDLER hmR0SvmExitIntr;
268	static FNSVMEXITHANDLER hmR0SvmExitWbinvd;
269	static FNSVMEXITHANDLER hmR0SvmExitInvd;
270	static FNSVMEXITHANDLER hmR0SvmExitCpuid;
271	static FNSVMEXITHANDLER hmR0SvmExitRdtsc;
272	static FNSVMEXITHANDLER hmR0SvmExitRdtscp;
273	static FNSVMEXITHANDLER hmR0SvmExitRdpmc;
274	static FNSVMEXITHANDLER hmR0SvmExitInvlpg;
275	static FNSVMEXITHANDLER hmR0SvmExitHlt;
276	static FNSVMEXITHANDLER hmR0SvmExitMonitor;
277	static FNSVMEXITHANDLER hmR0SvmExitMwait;
278	static FNSVMEXITHANDLER hmR0SvmExitShutdown;
279	static FNSVMEXITHANDLER hmR0SvmExitReadCRx;
280	static FNSVMEXITHANDLER hmR0SvmExitWriteCRx;
281	static FNSVMEXITHANDLER hmR0SvmExitSetPendingXcptUD;
282	static FNSVMEXITHANDLER hmR0SvmExitMsr;
283	static FNSVMEXITHANDLER hmR0SvmExitReadDRx;
284	static FNSVMEXITHANDLER hmR0SvmExitWriteDRx;
285	static FNSVMEXITHANDLER hmR0SvmExitIOInstr;
286	static FNSVMEXITHANDLER hmR0SvmExitNestedPF;
287	static FNSVMEXITHANDLER hmR0SvmExitVIntr;
288	static FNSVMEXITHANDLER hmR0SvmExitTaskSwitch;
289	static FNSVMEXITHANDLER hmR0SvmExitVmmCall;
290	static FNSVMEXITHANDLER hmR0SvmExitXcptPF;
291	static FNSVMEXITHANDLER hmR0SvmExitXcptNM;
292	static FNSVMEXITHANDLER hmR0SvmExitXcptMF;
293	static FNSVMEXITHANDLER hmR0SvmExitXcptDB;
294	/** @} */
295
296	DECLINLINE(int) hmR0SvmHandleExit(PVMCPU pVCpu, PCPUMCTX pMixedCtx, PSVMTRANSIENT pSvmTransient);
297
298	/*******************************************************************************
299	* Global Variables *
300	*******************************************************************************/
301	/** Ring-0 memory object for the IO bitmap. */
302	RTR0MEMOBJ g_hMemObjIOBitmap = NIL_RTR0MEMOBJ;
303	/** Physical address of the IO bitmap. */
304	RTHCPHYS g_HCPhysIOBitmap = 0;
305	/** Virtual address of the IO bitmap. */
306	R0PTRTYPE(void *) g_pvIOBitmap = NULL;
307
308
309	/**
310	* Sets up and activates AMD-V on the current CPU.
311	*
312	* @returns VBox status code.
313	* @param pCpu Pointer to the CPU info struct.
314	* @param pVM Pointer to the VM (can be NULL after a resume!).
315	* @param pvCpuPage Pointer to the global CPU page.
316	* @param HCPhysCpuPage Physical address of the global CPU page.
317	* @param fEnabledByHost Whether the host OS has already initialized AMD-V.
318	* @param pvArg Unused on AMD-V.
319	*/
320	VMMR0DECL(int) SVMR0EnableCpu(PHMGLOBALCPUINFO pCpu, PVM pVM, void *pvCpuPage, RTHCPHYS HCPhysCpuPage, bool fEnabledByHost,
321	void *pvArg)
322	{
323	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
324	AssertReturn(!fEnabledByHost, VERR_INVALID_PARAMETER);
325	AssertReturn( HCPhysCpuPage
326	&& HCPhysCpuPage != NIL_RTHCPHYS, VERR_INVALID_PARAMETER);
327	AssertReturn(pvCpuPage, VERR_INVALID_PARAMETER);
328	NOREF(pvArg);
329	NOREF(fEnabledByHost);
330
331	/* Paranoid: Disable interrupt as, in theory, interrupt handlers might mess with EFER. */
332	RTCCUINTREG uEflags = ASMIntDisableFlags();
333
334	/*
335	* We must turn on AMD-V and setup the host state physical address, as those MSRs are per CPU.
336	*/
337	uint64_t u64HostEfer = ASMRdMsr(MSR_K6_EFER);
338	if (u64HostEfer & MSR_K6_EFER_SVME)
339	{
340	/* If the VBOX_HWVIRTEX_IGNORE_SVM_IN_USE is active, then we blindly use AMD-V. */
341	if ( pVM
342	&& pVM->hm.s.svm.fIgnoreInUseError)
343	{
344	pCpu->fIgnoreAMDVInUseError = true;
345	}
346
347	if (!pCpu->fIgnoreAMDVInUseError)
348	{
349	ASMSetFlags(uEflags);
350	return VERR_SVM_IN_USE;
351	}
352	}
353
354	/* Turn on AMD-V in the EFER MSR. */
355	ASMWrMsr(MSR_K6_EFER, u64HostEfer \| MSR_K6_EFER_SVME);
356
357	/* Write the physical page address where the CPU will store the host state while executing the VM. */
358	ASMWrMsr(MSR_K8_VM_HSAVE_PA, HCPhysCpuPage);
359
360	/* Restore interrupts. */
361	ASMSetFlags(uEflags);
362
363	/*
364	* Theoretically, other hypervisors may have used ASIDs, ideally we should flush all non-zero ASIDs
365	* when enabling SVM. AMD doesn't have an SVM instruction to flush all ASIDs (flushing is done
366	* upon VMRUN). Therefore, just set the fFlushAsidBeforeUse flag which instructs hmR0SvmSetupTLB()
367	* to flush the TLB with before using a new ASID.
368	*/
369	pCpu->fFlushAsidBeforeUse = true;
370
371	/*
372	* Ensure each VCPU scheduled on this CPU gets a new VPID on resume. See @bugref{6255}.
373	*/
374	++pCpu->cTlbFlushes;
375
376	return VINF_SUCCESS;
377	}
378
379
380	/**
381	* Deactivates AMD-V on the current CPU.
382	*
383	* @returns VBox status code.
384	* @param pCpu Pointer to the CPU info struct.
385	* @param pvCpuPage Pointer to the global CPU page.
386	* @param HCPhysCpuPage Physical address of the global CPU page.
387	*/
388	VMMR0DECL(int) SVMR0DisableCpu(PHMGLOBALCPUINFO pCpu, void *pvCpuPage, RTHCPHYS HCPhysCpuPage)
389	{
390	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
391	AssertReturn( HCPhysCpuPage
392	&& HCPhysCpuPage != NIL_RTHCPHYS, VERR_INVALID_PARAMETER);
393	AssertReturn(pvCpuPage, VERR_INVALID_PARAMETER);
394	NOREF(pCpu);
395
396	/* Paranoid: Disable interrupts as, in theory, interrupt handlers might mess with EFER. */
397	RTCCUINTREG uEflags = ASMIntDisableFlags();
398
399	/* Turn off AMD-V in the EFER MSR. */
400	uint64_t u64HostEfer = ASMRdMsr(MSR_K6_EFER);
401	ASMWrMsr(MSR_K6_EFER, u64HostEfer & ~MSR_K6_EFER_SVME);
402
403	/* Invalidate host state physical address. */
404	ASMWrMsr(MSR_K8_VM_HSAVE_PA, 0);
405
406	/* Restore interrupts. */
407	ASMSetFlags(uEflags);
408
409	return VINF_SUCCESS;
410	}
411
412
413	/**
414	* Does global AMD-V initialization (called during module initialization).
415	*
416	* @returns VBox status code.
417	*/
418	VMMR0DECL(int) SVMR0GlobalInit(void)
419	{
420	/*
421	* Allocate 12 KB for the IO bitmap. Since this is non-optional and we always intercept all IO accesses, it's done
422	* once globally here instead of per-VM.
423	*/
424	Assert(g_hMemObjIOBitmap == NIL_RTR0MEMOBJ);
425	int rc = RTR0MemObjAllocCont(&g_hMemObjIOBitmap, 3 << PAGE_SHIFT, false /* fExecutable */);
426	if (RT_FAILURE(rc))
427	return rc;
428
429	g_pvIOBitmap = RTR0MemObjAddress(g_hMemObjIOBitmap);
430	g_HCPhysIOBitmap = RTR0MemObjGetPagePhysAddr(g_hMemObjIOBitmap, 0 /* iPage */);
431
432	/* Set all bits to intercept all IO accesses. */
433	ASMMemFill32(g_pvIOBitmap, 3 << PAGE_SHIFT, UINT32_C(0xffffffff));
434	return VINF_SUCCESS;
435	}
436
437
438	/**
439	* Does global AMD-V termination (called during module termination).
440	*/
441	VMMR0DECL(void) SVMR0GlobalTerm(void)
442	{
443	if (g_hMemObjIOBitmap != NIL_RTR0MEMOBJ)
444	{
445	RTR0MemObjFree(g_hMemObjIOBitmap, true /* fFreeMappings */);
446	g_pvIOBitmap = NULL;
447	g_HCPhysIOBitmap = 0;
448	g_hMemObjIOBitmap = NIL_RTR0MEMOBJ;
449	}
450	}
451
452
453	/**
454	* Frees any allocated per-VCPU structures for a VM.
455	*
456	* @param pVM Pointer to the VM.
457	*/
458	DECLINLINE(void) hmR0SvmFreeStructs(PVM pVM)
459	{
460	for (uint32_t i = 0; i < pVM->cCpus; i++)
461	{
462	PVMCPU pVCpu = &pVM->aCpus[i];
463	AssertPtr(pVCpu);
464
465	if (pVCpu->hm.s.svm.hMemObjVmcbHost != NIL_RTR0MEMOBJ)
466	{
467	RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjVmcbHost, false);
468	pVCpu->hm.s.svm.pvVmcbHost = 0;
469	pVCpu->hm.s.svm.HCPhysVmcbHost = 0;
470	pVCpu->hm.s.svm.hMemObjVmcbHost = NIL_RTR0MEMOBJ;
471	}
472
473	if (pVCpu->hm.s.svm.hMemObjVmcb != NIL_RTR0MEMOBJ)
474	{
475	RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjVmcb, false);
476	pVCpu->hm.s.svm.pvVmcb = 0;
477	pVCpu->hm.s.svm.HCPhysVmcb = 0;
478	pVCpu->hm.s.svm.hMemObjVmcb = NIL_RTR0MEMOBJ;
479	}
480
481	if (pVCpu->hm.s.svm.hMemObjMsrBitmap != NIL_RTR0MEMOBJ)
482	{
483	RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjMsrBitmap, false);
484	pVCpu->hm.s.svm.pvMsrBitmap = 0;
485	pVCpu->hm.s.svm.HCPhysMsrBitmap = 0;
486	pVCpu->hm.s.svm.hMemObjMsrBitmap = NIL_RTR0MEMOBJ;
487	}
488	}
489	}
490
491
492	/**
493	* Does per-VM AMD-V initialization.
494	*
495	* @returns VBox status code.
496	* @param pVM Pointer to the VM.
497	*/
498	VMMR0DECL(int) SVMR0InitVM(PVM pVM)
499	{
500	int rc = VERR_INTERNAL_ERROR_5;
501
502	/*
503	* Check for an AMD CPU erratum which requires us to flush the TLB before every world-switch.
504	*/
505	uint32_t u32Family;
506	uint32_t u32Model;
507	uint32_t u32Stepping;
508	if (HMAmdIsSubjectToErratum170(&u32Family, &u32Model, &u32Stepping))
509	{
510	Log4(("SVMR0InitVM: AMD cpu with erratum 170 family %#x model %#x stepping %#x\n", u32Family, u32Model, u32Stepping));
511	pVM->hm.s.svm.fAlwaysFlushTLB = true;
512	}
513
514	/*
515	* Initialize the R0 memory objects up-front so we can properly cleanup on allocation failures.
516	*/
517	for (VMCPUID i = 0; i < pVM->cCpus; i++)
518	{
519	PVMCPU pVCpu = &pVM->aCpus[i];
520	pVCpu->hm.s.svm.hMemObjVmcbHost = NIL_RTR0MEMOBJ;
521	pVCpu->hm.s.svm.hMemObjVmcb = NIL_RTR0MEMOBJ;
522	pVCpu->hm.s.svm.hMemObjMsrBitmap = NIL_RTR0MEMOBJ;
523	}
524
525	for (VMCPUID i = 0; i < pVM->cCpus; i++)
526	{
527	PVMCPU pVCpu = &pVM->aCpus[i];
528
529	/*
530	* Allocate one page for the host-context VM control block (VMCB). This is used for additional host-state (such as
531	* FS, GS, Kernel GS Base, etc.) apart from the host-state save area specified in MSR_K8_VM_HSAVE_PA.
532	*/
533	rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjVmcbHost, 1 << PAGE_SHIFT, false /* fExecutable */);
534	if (RT_FAILURE(rc))
535	goto failure_cleanup;
536
537	pVCpu->hm.s.svm.pvVmcbHost = RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjVmcbHost);
538	pVCpu->hm.s.svm.HCPhysVmcbHost = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjVmcbHost, 0 /* iPage */);
539	Assert(pVCpu->hm.s.svm.HCPhysVmcbHost < _4G);
540	ASMMemZeroPage(pVCpu->hm.s.svm.pvVmcbHost);
541
542	/*
543	* Allocate one page for the guest-state VMCB.
544	*/
545	rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjVmcb, 1 << PAGE_SHIFT, false /* fExecutable */);
546	if (RT_FAILURE(rc))
547	goto failure_cleanup;
548
549	pVCpu->hm.s.svm.pvVmcb = RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjVmcb);
550	pVCpu->hm.s.svm.HCPhysVmcb = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjVmcb, 0 /* iPage */);
551	Assert(pVCpu->hm.s.svm.HCPhysVmcb < _4G);
552	ASMMemZeroPage(pVCpu->hm.s.svm.pvVmcb);
553
554	/*
555	* Allocate two pages (8 KB) for the MSR permission bitmap. There doesn't seem to be a way to convince
556	* SVM to not require one.
557	*/
558	rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjMsrBitmap, 2 << PAGE_SHIFT, false /* fExecutable */);
559	if (RT_FAILURE(rc))
560	goto failure_cleanup;
561
562	pVCpu->hm.s.svm.pvMsrBitmap = RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjMsrBitmap);
563	pVCpu->hm.s.svm.HCPhysMsrBitmap = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjMsrBitmap, 0 /* iPage */);
564	/* Set all bits to intercept all MSR accesses (changed later on). */
565	ASMMemFill32(pVCpu->hm.s.svm.pvMsrBitmap, 2 << PAGE_SHIFT, UINT32_C(0xffffffff));
566	}
567
568	return VINF_SUCCESS;
569
570	failure_cleanup:
571	hmR0SvmFreeStructs(pVM);
572	return rc;
573	}
574
575
576	/**
577	* Does per-VM AMD-V termination.
578	*
579	* @returns VBox status code.
580	* @param pVM Pointer to the VM.
581	*/
582	VMMR0DECL(int) SVMR0TermVM(PVM pVM)
583	{
584	hmR0SvmFreeStructs(pVM);
585	return VINF_SUCCESS;
586	}
587
588
589	/**
590	* Sets the permission bits for the specified MSR in the MSRPM.
591	*
592	* @param pVCpu Pointer to the VMCPU.
593	* @param uMsr The MSR for which the access permissions are being set.
594	* @param enmRead MSR read permissions.
595	* @param enmWrite MSR write permissions.
596	*/
597	static void hmR0SvmSetMsrPermission(PVMCPU pVCpu, unsigned uMsr, SVMMSREXITREAD enmRead, SVMMSREXITWRITE enmWrite)
598	{
599	unsigned ulBit;
600	uint8_t pbMsrBitmap = (uint8_t )pVCpu->hm.s.svm.pvMsrBitmap;
601
602	/*
603	* Layout:
604	* Byte offset MSR range
605	* 0x000 - 0x7ff 0x00000000 - 0x00001fff
606	* 0x800 - 0xfff 0xc0000000 - 0xc0001fff
607	* 0x1000 - 0x17ff 0xc0010000 - 0xc0011fff
608	* 0x1800 - 0x1fff Reserved
609	*/
610	if (uMsr <= 0x00001FFF)
611	{
612	/* Pentium-compatible MSRs. */
613	ulBit = uMsr * 2;
614	}
615	else if ( uMsr >= 0xC0000000
616	&& uMsr <= 0xC0001FFF)
617	{
618	/* AMD Sixth Generation x86 Processor MSRs. */
619	ulBit = (uMsr - 0xC0000000) * 2;
620	pbMsrBitmap += 0x800;
621	}
622	else if ( uMsr >= 0xC0010000
623	&& uMsr <= 0xC0011FFF)
624	{
625	/* AMD Seventh and Eighth Generation Processor MSRs. */
626	ulBit = (uMsr - 0xC0001000) * 2;
627	pbMsrBitmap += 0x1000;
628	}
629	else
630	{
631	AssertFailed();
632	return;
633	}
634
635	Assert(ulBit < 0x3fff /* 16 * 1024 - 1 */);
636	if (enmRead == SVMMSREXIT_INTERCEPT_READ)
637	ASMBitSet(pbMsrBitmap, ulBit);
638	else
639	ASMBitClear(pbMsrBitmap, ulBit);
640
641	if (enmWrite == SVMMSREXIT_INTERCEPT_WRITE)
642	ASMBitSet(pbMsrBitmap, ulBit + 1);
643	else
644	ASMBitClear(pbMsrBitmap, ulBit + 1);
645
646	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
647	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_IOPM_MSRPM;
648	}
649
650
651	/**
652	* Sets up AMD-V for the specified VM.
653	* This function is only called once per-VM during initalization.
654	*
655	* @returns VBox status code.
656	* @param pVM Pointer to the VM.
657	*/
658	VMMR0DECL(int) SVMR0SetupVM(PVM pVM)
659	{
660	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
661	AssertReturn(pVM, VERR_INVALID_PARAMETER);
662	Assert(pVM->hm.s.svm.fSupported);
663
664	for (VMCPUID i = 0; i < pVM->cCpus; i++)
665	{
666	PVMCPU pVCpu = &pVM->aCpus[i];
667	PSVMVMCB pVmcb = (PSVMVMCB)pVM->aCpus[i].hm.s.svm.pvVmcb;
668
669	AssertMsgReturn(pVmcb, ("Invalid pVmcb for vcpu[%u]\n", i), VERR_SVM_INVALID_PVMCB);
670
671	/* Trap exceptions unconditionally (debug purposes). */
672	#ifdef HMSVM_ALWAYS_TRAP_PF
673	pVmcb->ctrl.u32InterceptException \|= RT_BIT(X86_XCPT_PF);
674	#endif
675	#ifdef HMSVM_ALWAYS_TRAP_ALL_XCPTS
676	/* If you add any exceptions here, make sure to update hmR0SvmHandleExit(). */
677	pVmcb->ctrl.u32InterceptException \|= 0
678	\| RT_BIT(X86_XCPT_BP)
679	\| RT_BIT(X86_XCPT_DB)
680	\| RT_BIT(X86_XCPT_DE)
681	\| RT_BIT(X86_XCPT_NM)
682	\| RT_BIT(X86_XCPT_UD)
683	\| RT_BIT(X86_XCPT_NP)
684	\| RT_BIT(X86_XCPT_SS)
685	\| RT_BIT(X86_XCPT_GP)
686	\| RT_BIT(X86_XCPT_PF)
687	\| RT_BIT(X86_XCPT_MF)
688	;
689	#endif
690
691	/* Set up unconditional intercepts and conditions. */
692	pVmcb->ctrl.u32InterceptCtrl1 = SVM_CTRL1_INTERCEPT_INTR /* External interrupt causes a VM-exit. */
693	\| SVM_CTRL1_INTERCEPT_NMI /* Non-Maskable Interrupts causes a VM-exit. */
694	\| SVM_CTRL1_INTERCEPT_INIT /* INIT signal causes a VM-exit. */
695	\| SVM_CTRL1_INTERCEPT_RDPMC /* RDPMC causes a VM-exit. */
696	\| SVM_CTRL1_INTERCEPT_CPUID /* CPUID causes a VM-exit. */
697	\| SVM_CTRL1_INTERCEPT_RSM /* RSM causes a VM-exit. */
698	\| SVM_CTRL1_INTERCEPT_HLT /* HLT causes a VM-exit. */
699	\| SVM_CTRL1_INTERCEPT_INOUT_BITMAP /* Use the IOPM to cause IOIO VM-exits. */
700	\| SVM_CTRL1_INTERCEPT_MSR_SHADOW /* MSR access not covered by MSRPM causes a VM-exit.*/
701	\| SVM_CTRL1_INTERCEPT_INVLPGA /* INVLPGA causes a VM-exit. */
702	\| SVM_CTRL1_INTERCEPT_SHUTDOWN /* Shutdown events causes a VM-exit. */
703	\| SVM_CTRL1_INTERCEPT_FERR_FREEZE; /* Intercept "freezing" during legacy FPU handling. */
704
705	pVmcb->ctrl.u32InterceptCtrl2 = SVM_CTRL2_INTERCEPT_VMRUN /* VMRUN causes a VM-exit. */
706	\| SVM_CTRL2_INTERCEPT_VMMCALL /* VMMCALL causes a VM-exit. */
707	\| SVM_CTRL2_INTERCEPT_VMLOAD /* VMLOAD causes a VM-exit. */
708	\| SVM_CTRL2_INTERCEPT_VMSAVE /* VMSAVE causes a VM-exit. */
709	\| SVM_CTRL2_INTERCEPT_STGI /* STGI causes a VM-exit. */
710	\| SVM_CTRL2_INTERCEPT_CLGI /* CLGI causes a VM-exit. */
711	\| SVM_CTRL2_INTERCEPT_SKINIT /* SKINIT causes a VM-exit. */
712	\| SVM_CTRL2_INTERCEPT_WBINVD /* WBINVD causes a VM-exit. */
713	\| SVM_CTRL2_INTERCEPT_MONITOR /* MONITOR causes a VM-exit. */
714	\| SVM_CTRL2_INTERCEPT_MWAIT; /* MWAIT causes a VM-exit. */
715
716	/* CR0, CR4 reads must be intercepted, our shadow values are not necessarily the same as the guest's. */
717	pVmcb->ctrl.u16InterceptRdCRx = RT_BIT(0) \| RT_BIT(4);
718
719	/* CR0, CR4 writes must be intercepted for the same reasons as above. */
720	pVmcb->ctrl.u16InterceptWrCRx = RT_BIT(0) \| RT_BIT(4);
721
722	/* Intercept all DRx reads and writes by default. Changed later on. */
723	pVmcb->ctrl.u16InterceptRdDRx = 0xffff;
724	pVmcb->ctrl.u16InterceptWrDRx = 0xffff;
725
726	/* Virtualize masking of INTR interrupts. (reads/writes from/to CR8 go to the V_TPR register) */
727	pVmcb->ctrl.IntCtrl.n.u1VIrqMasking = 1;
728
729	/* Ignore the priority in the TPR. This is necessary for delivering PIC style (ExtInt) interrupts and we currently
730	deliver both PIC and APIC interrupts alike. See hmR0SvmInjectPendingEvent() */
731	pVmcb->ctrl.IntCtrl.n.u1IgnoreTPR = 1;
732
733	/* Set IO and MSR bitmap permission bitmap physical addresses. */
734	pVmcb->ctrl.u64IOPMPhysAddr = g_HCPhysIOBitmap;
735	pVmcb->ctrl.u64MSRPMPhysAddr = pVCpu->hm.s.svm.HCPhysMsrBitmap;
736
737	/* No LBR virtualization. */
738	pVmcb->ctrl.u64LBRVirt = 0;
739
740	/* Initially set all VMCB clean bits to 0 indicating that everything should be loaded from the VMCB in memory. */
741	pVmcb->ctrl.u64VmcbCleanBits = 0;
742
743	/* The host ASID MBZ, for the guest start with 1. */
744	pVmcb->ctrl.TLBCtrl.n.u32ASID = 1;
745
746	/*
747	* Setup the PAT MSR (applicable for Nested Paging only).
748	* The default value should be 0x0007040600070406ULL, but we want to treat all guest memory as WB,
749	* so choose type 6 for all PAT slots.
750	*/
751	pVmcb->guest.u64GPAT = UINT64_C(0x0006060606060606);
752
753	/* Setup Nested Paging. This doesn't change throughout the execution time of the VM. */
754	pVmcb->ctrl.NestedPaging.n.u1NestedPaging = pVM->hm.s.fNestedPaging;
755
756	/* Without Nested Paging, we need additionally intercepts. */
757	if (!pVM->hm.s.fNestedPaging)
758	{
759	/* CR3 reads/writes must be intercepted; our shadow values differ from the guest values. */
760	pVmcb->ctrl.u16InterceptRdCRx \|= RT_BIT(3);
761	pVmcb->ctrl.u16InterceptWrCRx \|= RT_BIT(3);
762
763	/* Intercept INVLPG and task switches (may change CR3, EFLAGS, LDT). */
764	pVmcb->ctrl.u32InterceptCtrl1 \|= SVM_CTRL1_INTERCEPT_INVLPG
765	\| SVM_CTRL1_INTERCEPT_TASK_SWITCH;
766
767	/* Page faults must be intercepted to implement shadow paging. */
768	pVmcb->ctrl.u32InterceptException \|= RT_BIT(X86_XCPT_PF);
769	}
770
771	#ifdef HMSVM_ALWAYS_TRAP_TASK_SWITCH
772	pVmcb->ctrl.u32InterceptCtrl1 \|= SVM_CTRL1_INTERCEPT_TASK_SWITCH;
773	#endif
774
775	/*
776	* The following MSRs are saved/restored automatically during the world-switch.
777	* Don't intercept guest read/write accesses to these MSRs.
778	*/
779	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
780	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_CSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
781	hmR0SvmSetMsrPermission(pVCpu, MSR_K6_STAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
782	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_SF_MASK, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
783	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_FS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
784	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_GS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
785	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_KERNEL_GS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
786	hmR0SvmSetMsrPermission(pVCpu, MSR_IA32_SYSENTER_CS, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
787	hmR0SvmSetMsrPermission(pVCpu, MSR_IA32_SYSENTER_ESP, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
788	hmR0SvmSetMsrPermission(pVCpu, MSR_IA32_SYSENTER_EIP, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
789	}
790
791	return VINF_SUCCESS;
792	}
793
794
795	/**
796	* Invalidates a guest page by guest virtual address.
797	*
798	* @returns VBox status code.
799	* @param pVM Pointer to the VM.
800	* @param pVCpu Pointer to the VMCPU.
801	* @param GCVirt Guest virtual address of the page to invalidate.
802	*/
803	VMMR0DECL(int) SVMR0InvalidatePage(PVM pVM, PVMCPU pVCpu, RTGCPTR GCVirt)
804	{
805	AssertReturn(pVM, VERR_INVALID_PARAMETER);
806	Assert(pVM->hm.s.svm.fSupported);
807
808	bool fFlushPending = pVM->hm.s.svm.fAlwaysFlushTLB \|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_TLB_FLUSH);
809
810	/* Skip it if a TLB flush is already pending. */
811	if (!fFlushPending)
812	{
813	Log4(("SVMR0InvalidatePage %RGv\n", GCVirt));
814
815	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
816	AssertMsgReturn(pVmcb, ("Invalid pVmcb!\n"), VERR_SVM_INVALID_PVMCB);
817
818	#if HC_ARCH_BITS == 32
819	/* If we get a flush in 64-bit guest mode, then force a full TLB flush. INVLPGA takes only 32-bit addresses. */
820	if (CPUMIsGuestInLongMode(pVCpu))
821	VMCPU_FF_SET(pVCpu, VMCPU_FF_TLB_FLUSH);
822	else
823	#endif
824	{
825	SVMR0InvlpgA(GCVirt, pVmcb->ctrl.TLBCtrl.n.u32ASID);
826	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlbInvlpgVirt);
827	}
828	}
829	return VINF_SUCCESS;
830	}
831
832
833	/**
834	* Flushes the appropriate tagged-TLB entries.
835	*
836	* @param pVM Pointer to the VM.
837	* @param pVCpu Pointer to the VMCPU.
838	*/
839	static void hmR0SvmFlushTaggedTlb(PVMCPU pVCpu)
840	{
841	PVM pVM = pVCpu->CTX_SUFF(pVM);
842	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
843	PHMGLOBALCPUINFO pCpu = HMR0GetCurrentCpu();
844
845	/*
846	* Force a TLB flush for the first world switch if the current CPU differs from the one we ran on last.
847	* This can happen both for start & resume due to long jumps back to ring-3.
848	* If the TLB flush count changed, another VM (VCPU rather) has hit the ASID limit while flushing the TLB,
849	* so we cannot reuse the ASIDs without flushing.
850	*/
851	bool fNewAsid = false;
852	Assert(pCpu->idCpu != NIL_RTCPUID);
853	if ( pVCpu->hm.s.idLastCpu != pCpu->idCpu
854	\|\| pVCpu->hm.s.cTlbFlushes != pCpu->cTlbFlushes)
855	{
856	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlbWorldSwitch);
857	pVCpu->hm.s.fForceTLBFlush = true;
858	fNewAsid = true;
859	}
860
861	/* Set TLB flush state as checked until we return from the world switch. */
862	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, true);
863
864	/* Check for explicit TLB shootdowns. */
865	if (VMCPU_FF_TEST_AND_CLEAR(pVCpu, VMCPU_FF_TLB_FLUSH))
866	{
867	pVCpu->hm.s.fForceTLBFlush = true;
868	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlb);
869	}
870
871	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_NOTHING;
872
873	if (pVM->hm.s.svm.fAlwaysFlushTLB)
874	{
875	/*
876	* This is the AMD erratum 170. We need to flush the entire TLB for each world switch. Sad.
877	*/
878	pCpu->uCurrentAsid = 1;
879	pVCpu->hm.s.uCurrentAsid = 1;
880	pVCpu->hm.s.cTlbFlushes = pCpu->cTlbFlushes;
881	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
882
883	/* Clear the VMCB Clean Bit for NP while flushing the TLB. See @bugref{7152}. */
884	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
885	}
886	else if (pVCpu->hm.s.fForceTLBFlush)
887	{
888	/* Clear the VMCB Clean Bit for NP while flushing the TLB. See @bugref{7152}. */
889	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
890
891	if (fNewAsid)
892	{
893	++pCpu->uCurrentAsid;
894	bool fHitASIDLimit = false;
895	if (pCpu->uCurrentAsid >= pVM->hm.s.uMaxAsid)
896	{
897	pCpu->uCurrentAsid = 1; /* Wraparound at 1; host uses 0 */
898	pCpu->cTlbFlushes++; /* All VCPUs that run on this host CPU must use a new VPID. */
899	fHitASIDLimit = true;
900
901	if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_FLUSH_BY_ASID)
902	{
903	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_SINGLE_CONTEXT;
904	pCpu->fFlushAsidBeforeUse = true;
905	}
906	else
907	{
908	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
909	pCpu->fFlushAsidBeforeUse = false;
910	}
911	}
912
913	if ( !fHitASIDLimit
914	&& pCpu->fFlushAsidBeforeUse)
915	{
916	if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_FLUSH_BY_ASID)
917	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_SINGLE_CONTEXT;
918	else
919	{
920	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
921	pCpu->fFlushAsidBeforeUse = false;
922	}
923	}
924
925	pVCpu->hm.s.uCurrentAsid = pCpu->uCurrentAsid;
926	pVCpu->hm.s.idLastCpu = pCpu->idCpu;
927	pVCpu->hm.s.cTlbFlushes = pCpu->cTlbFlushes;
928	}
929	else
930	{
931	if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_FLUSH_BY_ASID)
932	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_SINGLE_CONTEXT;
933	else
934	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
935	}
936
937	pVCpu->hm.s.fForceTLBFlush = false;
938	}
939	/** @todo We never set VMCPU_FF_TLB_SHOOTDOWN anywhere so this path should
940	* not be executed. See hmQueueInvlPage() where it is commented
941	* out. Support individual entry flushing someday. */
942	#if 0
943	else
944	{
945	if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_TLB_SHOOTDOWN))
946	{
947	/* Deal with pending TLB shootdown actions which were queued when we were not executing code. */
948	STAM_COUNTER_INC(&pVCpu->hm.s.StatTlbShootdown);
949	for (uint32_t i = 0; i < pVCpu->hm.s.TlbShootdown.cPages; i++)
950	SVMR0InvlpgA(pVCpu->hm.s.TlbShootdown.aPages[i], pVmcb->ctrl.TLBCtrl.n.u32ASID);
951
952	pVCpu->hm.s.TlbShootdown.cPages = 0;
953	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_TLB_SHOOTDOWN);
954	}
955	}
956	#endif
957
958
959	/* Update VMCB with the ASID. */
960	if (pVmcb->ctrl.TLBCtrl.n.u32ASID != pVCpu->hm.s.uCurrentAsid)
961	{
962	pVmcb->ctrl.TLBCtrl.n.u32ASID = pVCpu->hm.s.uCurrentAsid;
963	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_ASID;
964	}
965
966	AssertMsg(pVCpu->hm.s.idLastCpu == pCpu->idCpu,
967	("vcpu idLastCpu=%x pcpu idCpu=%x\n", pVCpu->hm.s.idLastCpu, pCpu->idCpu));
968	AssertMsg(pVCpu->hm.s.cTlbFlushes == pCpu->cTlbFlushes,
969	("Flush count mismatch for cpu %d (%x vs %x)\n", pCpu->idCpu, pVCpu->hm.s.cTlbFlushes, pCpu->cTlbFlushes));
970	AssertMsg(pCpu->uCurrentAsid >= 1 && pCpu->uCurrentAsid < pVM->hm.s.uMaxAsid,
971	("cpu%d uCurrentAsid = %x\n", pCpu->idCpu, pCpu->uCurrentAsid));
972	AssertMsg(pVCpu->hm.s.uCurrentAsid >= 1 && pVCpu->hm.s.uCurrentAsid < pVM->hm.s.uMaxAsid,
973	("cpu%d VM uCurrentAsid = %x\n", pCpu->idCpu, pVCpu->hm.s.uCurrentAsid));
974
975	#ifdef VBOX_WITH_STATISTICS
976	if (pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_NOTHING)
977	STAM_COUNTER_INC(&pVCpu->hm.s.StatNoFlushTlbWorldSwitch);
978	else if ( pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_SINGLE_CONTEXT
979	\|\| pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_SINGLE_CONTEXT_RETAIN_GLOBALS)
980	{
981	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushAsid);
982	}
983	else
984	{
985	Assert(pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_ENTIRE);
986	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushEntire);
987	}
988	#endif
989	}
990
991
992	/** @name 64-bit guest on 32-bit host OS helper functions.
993	*
994	* The host CPU is still 64-bit capable but the host OS is running in 32-bit
995	* mode (code segment, paging). These wrappers/helpers perform the necessary
996	* bits for the 32->64 switcher.
997	*
998	* @{ */
999	#if HC_ARCH_BITS == 32 && defined(VBOX_ENABLE_64_BITS_GUESTS) && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
1000	/**
1001	* Prepares for and executes VMRUN (64-bit guests on a 32-bit host).
1002	*
1003	* @returns VBox status code.
1004	* @param HCPhysVmcbHost Physical address of host VMCB.
1005	* @param HCPhysVmcb Physical address of the VMCB.
1006	* @param pCtx Pointer to the guest-CPU context.
1007	* @param pVM Pointer to the VM.
1008	* @param pVCpu Pointer to the VMCPU.
1009	*/
1010	DECLASM(int) SVMR0VMSwitcherRun64(RTHCPHYS HCPhysVmcbHost, RTHCPHYS HCPhysVmcb, PCPUMCTX pCtx, PVM pVM, PVMCPU pVCpu)
1011	{
1012	uint32_t aParam[4];
1013	aParam[0] = (uint32_t)(HCPhysVmcbHost); /* Param 1: HCPhysVmcbHost - Lo. */
1014	aParam[1] = (uint32_t)(HCPhysVmcbHost >> 32); /* Param 1: HCPhysVmcbHost - Hi. */
1015	aParam[2] = (uint32_t)(HCPhysVmcb); /* Param 2: HCPhysVmcb - Lo. */
1016	aParam[3] = (uint32_t)(HCPhysVmcb >> 32); /* Param 2: HCPhysVmcb - Hi. */
1017
1018	return SVMR0Execute64BitsHandler(pVM, pVCpu, pCtx, HM64ON32OP_SVMRCVMRun64, 4, &aParam[0]);
1019	}
1020
1021
1022	/**
1023	* Executes the specified VMRUN handler in 64-bit mode.
1024	*
1025	* @returns VBox status code.
1026	* @param pVM Pointer to the VM.
1027	* @param pVCpu Pointer to the VMCPU.
1028	* @param pCtx Pointer to the guest-CPU context.
1029	* @param enmOp The operation to perform.
1030	* @param cbParam Number of parameters.
1031	* @param paParam Array of 32-bit parameters.
1032	*/
1033	VMMR0DECL(int) SVMR0Execute64BitsHandler(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, HM64ON32OP enmOp, uint32_t cbParam,
1034	uint32_t *paParam)
1035	{
1036	AssertReturn(pVM->hm.s.pfnHost32ToGuest64R0, VERR_HM_NO_32_TO_64_SWITCHER);
1037	Assert(enmOp > HM64ON32OP_INVALID && enmOp < HM64ON32OP_END);
1038
1039	/* Disable interrupts. */
1040	RTHCUINTREG uOldEFlags = ASMIntDisableFlags();
1041
1042	#ifdef VBOX_WITH_VMMR0_DISABLE_LAPIC_NMI
1043	RTCPUID idHostCpu = RTMpCpuId();
1044	CPUMR0SetLApic(pVCpu, idHostCpu);
1045	#endif
1046
1047	CPUMSetHyperESP(pVCpu, VMMGetStackRC(pVCpu));
1048	CPUMSetHyperEIP(pVCpu, enmOp);
1049	for (int i = (int)cbParam - 1; i >= 0; i--)
1050	CPUMPushHyper(pVCpu, paParam[i]);
1051
1052	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatWorldSwitch3264, z);
1053	/* Call the switcher. */
1054	int rc = pVM->hm.s.pfnHost32ToGuest64R0(pVM, RT_OFFSETOF(VM, aCpus[pVCpu->idCpu].cpum) - RT_OFFSETOF(VM, cpum));
1055	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatWorldSwitch3264, z);
1056
1057	/* Restore interrupts. */
1058	ASMSetFlags(uOldEFlags);
1059	return rc;
1060	}
1061
1062	#endif /* HC_ARCH_BITS == 32 && defined(VBOX_ENABLE_64_BITS_GUESTS) */
1063	/** @} */
1064
1065
1066	/**
1067	* Adds an exception to the intercept exception bitmap in the VMCB and updates
1068	* the corresponding VMCB Clean bit.
1069	*
1070	* @param pVmcb Pointer to the VM control block.
1071	* @param u32Xcpt The value of the exception (X86_XCPT_*).
1072	*/
1073	DECLINLINE(void) hmR0SvmAddXcptIntercept(PSVMVMCB pVmcb, uint32_t u32Xcpt)
1074	{
1075	if (!(pVmcb->ctrl.u32InterceptException & RT_BIT(u32Xcpt)))
1076	{
1077	pVmcb->ctrl.u32InterceptException \|= RT_BIT(u32Xcpt);
1078	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1079	}
1080	}
1081
1082
1083	/**
1084	* Removes an exception from the intercept-exception bitmap in the VMCB and
1085	* updates the corresponding VMCB Clean bit.
1086	*
1087	* @param pVmcb Pointer to the VM control block.
1088	* @param u32Xcpt The value of the exception (X86_XCPT_*).
1089	*/
1090	DECLINLINE(void) hmR0SvmRemoveXcptIntercept(PSVMVMCB pVmcb, uint32_t u32Xcpt)
1091	{
1092	#ifndef HMSVM_ALWAYS_TRAP_ALL_XCPTS
1093	if (pVmcb->ctrl.u32InterceptException & RT_BIT(u32Xcpt))
1094	{
1095	pVmcb->ctrl.u32InterceptException &= ~RT_BIT(u32Xcpt);
1096	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1097	}
1098	#endif
1099	}
1100
1101
1102	/**
1103	* Loads the guest CR0 control register into the guest-state area in the VMCB.
1104	* Although the guest CR0 is a separate field in the VMCB we have to consider
1105	* the FPU state itself which is shared between the host and the guest.
1106	*
1107	* @returns VBox status code.
1108	* @param pVM Pointer to the VMCPU.
1109	* @param pVmcb Pointer to the VM control block.
1110	* @param pCtx Pointer to the guest-CPU context.
1111	*
1112	* @remarks No-long-jump zone!!!
1113	*/
1114	static void hmR0SvmLoadSharedCR0(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1115	{
1116	/*
1117	* Guest CR0.
1118	*/
1119	PVM pVM = pVCpu->CTX_SUFF(pVM);
1120	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR0))
1121	{
1122	uint64_t u64GuestCR0 = pCtx->cr0;
1123
1124	/* Always enable caching. */
1125	u64GuestCR0 &= ~(X86_CR0_CD \| X86_CR0_NW);
1126
1127	/*
1128	* When Nested Paging is not available use shadow page tables and intercept #PFs (the latter done in SVMR0SetupVM()).
1129	*/
1130	if (!pVM->hm.s.fNestedPaging)
1131	{
1132	u64GuestCR0 \|= X86_CR0_PG; /* When Nested Paging is not available, use shadow page tables. */
1133	u64GuestCR0 \|= X86_CR0_WP; /* Guest CPL 0 writes to its read-only pages should cause a #PF VM-exit. */
1134	}
1135
1136	/*
1137	* Guest FPU bits.
1138	*/
1139	bool fInterceptNM = false;
1140	bool fInterceptMF = false;
1141	u64GuestCR0 \|= X86_CR0_NE; /* Use internal x87 FPU exceptions handling rather than external interrupts. */
1142	if (CPUMIsGuestFPUStateActive(pVCpu))
1143	{
1144	/* Catch floating point exceptions if we need to report them to the guest in a different way. */
1145	if (!(u64GuestCR0 & X86_CR0_NE))
1146	{
1147	Log4(("hmR0SvmLoadGuestControlRegs: Intercepting Guest CR0.MP Old-style FPU handling!!!\n"));
1148	fInterceptMF = true;
1149	}
1150	}
1151	else
1152	{
1153	fInterceptNM = true; /* Guest FPU inactive, VM-exit on #NM for lazy FPU loading. */
1154	u64GuestCR0 \|= X86_CR0_TS /* Guest can task switch quickly and do lazy FPU syncing. */
1155	\| X86_CR0_MP; /* FWAIT/WAIT should not ignore CR0.TS and should generate #NM. */
1156	}
1157
1158	/*
1159	* Update the exception intercept bitmap.
1160	*/
1161	if (fInterceptNM)
1162	hmR0SvmAddXcptIntercept(pVmcb, X86_XCPT_NM);
1163	else
1164	hmR0SvmRemoveXcptIntercept(pVmcb, X86_XCPT_NM);
1165
1166	if (fInterceptMF)
1167	hmR0SvmAddXcptIntercept(pVmcb, X86_XCPT_MF);
1168	else
1169	hmR0SvmRemoveXcptIntercept(pVmcb, X86_XCPT_MF);
1170
1171	pVmcb->guest.u64CR0 = u64GuestCR0;
1172	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1173	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_CR0);
1174	}
1175	}
1176
1177
1178	/**
1179	* Loads the guest control registers (CR2, CR3, CR4) into the VMCB.
1180	*
1181	* @returns VBox status code.
1182	* @param pVCpu Pointer to the VMCPU.
1183	* @param pVmcb Pointer to the VM control block.
1184	* @param pCtx Pointer to the guest-CPU context.
1185	*
1186	* @remarks No-long-jump zone!!!
1187	*/
1188	static int hmR0SvmLoadGuestControlRegs(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1189	{
1190	PVM pVM = pVCpu->CTX_SUFF(pVM);
1191
1192	/*
1193	* Guest CR2.
1194	*/
1195	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR2))
1196	{
1197	pVmcb->guest.u64CR2 = pCtx->cr2;
1198	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CR2;
1199	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_CR2);
1200	}
1201
1202	/*
1203	* Guest CR3.
1204	*/
1205	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR3))
1206	{
1207	if (pVM->hm.s.fNestedPaging)
1208	{
1209	PGMMODE enmShwPagingMode;
1210	#if HC_ARCH_BITS == 32
1211	if (CPUMIsGuestInLongModeEx(pCtx))
1212	enmShwPagingMode = PGMMODE_AMD64_NX;
1213	else
1214	#endif
1215	enmShwPagingMode = PGMGetHostMode(pVM);
1216
1217	pVmcb->ctrl.u64NestedPagingCR3 = PGMGetNestedCR3(pVCpu, enmShwPagingMode);
1218	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
1219	Assert(pVmcb->ctrl.u64NestedPagingCR3);
1220	pVmcb->guest.u64CR3 = pCtx->cr3;
1221	}
1222	else
1223	pVmcb->guest.u64CR3 = PGMGetHyperCR3(pVCpu);
1224
1225	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1226	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_CR3);
1227	}
1228
1229	/*
1230	* Guest CR4.
1231	*/
1232	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR4))
1233	{
1234	uint64_t u64GuestCR4 = pCtx->cr4;
1235	if (!pVM->hm.s.fNestedPaging)
1236	{
1237	switch (pVCpu->hm.s.enmShadowMode)
1238	{
1239	case PGMMODE_REAL:
1240	case PGMMODE_PROTECTED: /* Protected mode, no paging. */
1241	AssertFailed();
1242	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1243
1244	case PGMMODE_32_BIT: /* 32-bit paging. */
1245	u64GuestCR4 &= ~X86_CR4_PAE;
1246	break;
1247
1248	case PGMMODE_PAE: /* PAE paging. */
1249	case PGMMODE_PAE_NX: /* PAE paging with NX enabled. */
1250	/** Must use PAE paging as we could use physical memory > 4 GB */
1251	u64GuestCR4 \|= X86_CR4_PAE;
1252	break;
1253
1254	case PGMMODE_AMD64: /* 64-bit AMD paging (long mode). */
1255	case PGMMODE_AMD64_NX: /* 64-bit AMD paging (long mode) with NX enabled. */
1256	#ifdef VBOX_ENABLE_64_BITS_GUESTS
1257	break;
1258	#else
1259	AssertFailed();
1260	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1261	#endif
1262
1263	default: /* shut up gcc */
1264	AssertFailed();
1265	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1266	}
1267	}
1268
1269	pVmcb->guest.u64CR4 = u64GuestCR4;
1270	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1271	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_CR4);
1272	}
1273
1274	return VINF_SUCCESS;
1275	}
1276
1277
1278	/**
1279	* Loads the guest segment registers into the VMCB.
1280	*
1281	* @returns VBox status code.
1282	* @param pVCpu Pointer to the VMCPU.
1283	* @param pVmcb Pointer to the VM control block.
1284	* @param pCtx Pointer to the guest-CPU context.
1285	*
1286	* @remarks No-long-jump zone!!!
1287	*/
1288	static void hmR0SvmLoadGuestSegmentRegs(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1289	{
1290	/* Guest Segment registers: CS, SS, DS, ES, FS, GS. */
1291	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_SEGMENT_REGS))
1292	{
1293	HMSVM_LOAD_SEG_REG(CS, cs);
1294	HMSVM_LOAD_SEG_REG(SS, ss);
1295	HMSVM_LOAD_SEG_REG(DS, ds);
1296	HMSVM_LOAD_SEG_REG(ES, es);
1297	HMSVM_LOAD_SEG_REG(FS, fs);
1298	HMSVM_LOAD_SEG_REG(GS, gs);
1299
1300	pVmcb->guest.u8CPL = pCtx->ss.Attr.n.u2Dpl;
1301	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_SEG;
1302	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_SEGMENT_REGS);
1303	}
1304
1305	/* Guest TR. */
1306	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_TR))
1307	{
1308	HMSVM_LOAD_SEG_REG(TR, tr);
1309	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_TR);
1310	}
1311
1312	/* Guest LDTR. */
1313	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_LDTR))
1314	{
1315	HMSVM_LOAD_SEG_REG(LDTR, ldtr);
1316	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_LDTR);
1317	}
1318
1319	/* Guest GDTR. */
1320	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_GDTR))
1321	{
1322	pVmcb->guest.GDTR.u32Limit = pCtx->gdtr.cbGdt;
1323	pVmcb->guest.GDTR.u64Base = pCtx->gdtr.pGdt;
1324	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DT;
1325	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_GDTR);
1326	}
1327
1328	/* Guest IDTR. */
1329	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_IDTR))
1330	{
1331	pVmcb->guest.IDTR.u32Limit = pCtx->idtr.cbIdt;
1332	pVmcb->guest.IDTR.u64Base = pCtx->idtr.pIdt;
1333	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DT;
1334	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_IDTR);
1335	}
1336	}
1337
1338
1339	/**
1340	* Loads the guest MSRs into the VMCB.
1341	*
1342	* @param pVCpu Pointer to the VMCPU.
1343	* @param pVmcb Pointer to the VM control block.
1344	* @param pCtx Pointer to the guest-CPU context.
1345	*
1346	* @remarks No-long-jump zone!!!
1347	*/
1348	static void hmR0SvmLoadGuestMsrs(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1349	{
1350	/* Guest Sysenter MSRs. */
1351	pVmcb->guest.u64SysEnterCS = pCtx->SysEnter.cs;
1352	pVmcb->guest.u64SysEnterEIP = pCtx->SysEnter.eip;
1353	pVmcb->guest.u64SysEnterESP = pCtx->SysEnter.esp;
1354
1355	/*
1356	* Guest EFER MSR.
1357	* AMD-V requires guest EFER.SVME to be set. Weird. .
1358	* See AMD spec. 15.5.1 "Basic Operation" \| "Canonicalization and Consistency Checks".
1359	*/
1360	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_SVM_GUEST_EFER_MSR))
1361	{
1362	pVmcb->guest.u64EFER = pCtx->msrEFER \| MSR_K6_EFER_SVME;
1363	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1364	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_SVM_GUEST_EFER_MSR);
1365	}
1366
1367	/* 64-bit MSRs. */
1368	if (CPUMIsGuestInLongModeEx(pCtx))
1369	{
1370	pVmcb->guest.FS.u64Base = pCtx->fs.u64Base;
1371	pVmcb->guest.GS.u64Base = pCtx->gs.u64Base;
1372	}
1373	else
1374	{
1375	/* If the guest isn't in 64-bit mode, clear MSR_K6_LME bit from guest EFER otherwise AMD-V expects amd64 shadow paging. */
1376	if (pCtx->msrEFER & MSR_K6_EFER_LME)
1377	{
1378	pVmcb->guest.u64EFER &= ~MSR_K6_EFER_LME;
1379	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1380	}
1381	}
1382
1383
1384	/** @todo The following are used in 64-bit only (SYSCALL/SYSRET) but they might
1385	* be writable in 32-bit mode. Clarify with AMD spec. */
1386	pVmcb->guest.u64STAR = pCtx->msrSTAR;
1387	pVmcb->guest.u64LSTAR = pCtx->msrLSTAR;
1388	pVmcb->guest.u64CSTAR = pCtx->msrCSTAR;
1389	pVmcb->guest.u64SFMASK = pCtx->msrSFMASK;
1390	pVmcb->guest.u64KernelGSBase = pCtx->msrKERNELGSBASE;
1391	}
1392
1393
1394	/**
1395	* Loads the guest state into the VMCB and programs the necessary intercepts
1396	* accordingly.
1397	*
1398	* @param pVCpu Pointer to the VMCPU.
1399	* @param pVmcb Pointer to the VM control block.
1400	* @param pCtx Pointer to the guest-CPU context.
1401	*
1402	* @remarks No-long-jump zone!!!
1403	* @remarks Requires EFLAGS to be up-to-date in the VMCB!
1404	*/
1405	static void hmR0SvmLoadSharedDebugState(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1406	{
1407	if (!HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_DEBUG))
1408	return;
1409	Assert((pCtx->dr[6] & X86_DR6_RA1_MASK) == X86_DR6_RA1_MASK); Assert((pCtx->dr[6] & X86_DR6_RAZ_MASK) == 0);
1410	Assert((pCtx->dr[7] & X86_DR7_RA1_MASK) == X86_DR7_RA1_MASK); Assert((pCtx->dr[7] & X86_DR7_RAZ_MASK) == 0);
1411
1412	bool fInterceptDB = false;
1413	bool fInterceptMovDRx = false;
1414
1415	/*
1416	* Anyone single stepping on the host side? If so, we'll have to use the
1417	* trap flag in the guest EFLAGS since AMD-V doesn't have a trap flag on
1418	* the VMM level like VT-x implementations does.
1419	*/
1420	bool const fStepping = pVCpu->hm.s.fSingleInstruction \|\| DBGFIsStepping(pVCpu);
1421	if (fStepping)
1422	{
1423	pVCpu->hm.s.fClearTrapFlag = true;
1424	pVmcb->guest.u64RFlags \|= X86_EFL_TF;
1425	fInterceptDB = true;
1426	fInterceptMovDRx = true; /* Need clean DR6, no guest mess. */
1427	}
1428
1429	if ( fStepping
1430	\|\| (CPUMGetHyperDR7(pVCpu) & X86_DR7_ENABLED_MASK))
1431	{
1432	/*
1433	* Use the combined guest and host DRx values found in the hypervisor
1434	* register set because the debugger has breakpoints active or someone
1435	* is single stepping on the host side.
1436	*
1437	* Note! DBGF expects a clean DR6 state before executing guest code.
1438	*/
1439	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS) && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
1440	if ( CPUMIsGuestInLongModeEx(pCtx)
1441	&& !CPUMIsHyperDebugStateActivePending(pVCpu))
1442	{
1443	CPUMR0LoadHyperDebugState(pVCpu, false /* include DR6 */);
1444	Assert(!CPUMIsGuestDebugStateActivePending(pVCpu));
1445	Assert(CPUMIsHyperDebugStateActivePending(pVCpu));
1446	}
1447	else
1448	#endif
1449	if (!CPUMIsHyperDebugStateActive(pVCpu))
1450	{
1451	CPUMR0LoadHyperDebugState(pVCpu, false /* include DR6 */);
1452	Assert(!CPUMIsGuestDebugStateActive(pVCpu));
1453	Assert(CPUMIsHyperDebugStateActive(pVCpu));
1454	}
1455
1456	/* Update DR6 & DR7. (The other DRx values are handled by CPUM one way or the other.) */
1457	if ( pVmcb->guest.u64DR6 != X86_DR6_INIT_VAL
1458	\|\| pVmcb->guest.u64DR7 != CPUMGetHyperDR7(pVCpu))
1459	{
1460	pVmcb->guest.u64DR7 = CPUMGetHyperDR7(pVCpu);
1461	pVmcb->guest.u64DR6 = X86_DR6_INIT_VAL;
1462	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
1463	pVCpu->hm.s.fUsingHyperDR7 = true;
1464	}
1465
1466	/** @todo If we cared, we could optimize to allow the guest to read registers
1467	* with the same values. */
1468	fInterceptDB = true;
1469	fInterceptMovDRx = true;
1470	Log5(("hmR0SvmLoadSharedDebugState: Loaded hyper DRx\n"));
1471	}
1472	else
1473	{
1474	/*
1475	* Update DR6, DR7 with the guest values if necessary.
1476	*/
1477	if ( pVmcb->guest.u64DR7 != pCtx->dr[7]
1478	\|\| pVmcb->guest.u64DR6 != pCtx->dr[6])
1479	{
1480	pVmcb->guest.u64DR7 = pCtx->dr[7];
1481	pVmcb->guest.u64DR6 = pCtx->dr[6];
1482	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
1483	pVCpu->hm.s.fUsingHyperDR7 = false;
1484	}
1485
1486	/*
1487	* If the guest has enabled debug registers, we need to load them prior to
1488	* executing guest code so they'll trigger at the right time.
1489	*/
1490	if (pCtx->dr[7] & (X86_DR7_ENABLED_MASK \| X86_DR7_GD)) /** @todo Why GD? */
1491	{
1492	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS) && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
1493	if ( CPUMIsGuestInLongModeEx(pCtx)
1494	&& !CPUMIsGuestDebugStateActivePending(pVCpu))
1495	{
1496	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
1497	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxArmed);
1498	Assert(!CPUMIsHyperDebugStateActivePending(pVCpu));
1499	Assert(CPUMIsGuestDebugStateActivePending(pVCpu));
1500	}
1501	else
1502	#endif
1503	if (!CPUMIsGuestDebugStateActive(pVCpu))
1504	{
1505	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
1506	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxArmed);
1507	Assert(!CPUMIsHyperDebugStateActive(pVCpu));
1508	Assert(CPUMIsGuestDebugStateActive(pVCpu));
1509	}
1510	Log5(("hmR0SvmLoadSharedDebugState: Loaded guest DRx\n"));
1511	}
1512	/*
1513	* If no debugging enabled, we'll lazy load DR0-3. We don't need to
1514	* intercept #DB as DR6 is updated in the VMCB.
1515	*/
1516	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS) && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
1517	else if ( !CPUMIsGuestDebugStateActivePending(pVCpu)
1518	&& !CPUMIsGuestDebugStateActive(pVCpu))
1519	#else
1520	else if (!CPUMIsGuestDebugStateActive(pVCpu))
1521	#endif
1522	{
1523	fInterceptMovDRx = true;
1524	}
1525	}
1526
1527	/*
1528	* Set up the intercepts.
1529	*/
1530	if (fInterceptDB)
1531	hmR0SvmAddXcptIntercept(pVmcb, X86_XCPT_DB);
1532	else
1533	hmR0SvmRemoveXcptIntercept(pVmcb, X86_XCPT_DB);
1534
1535	if (fInterceptMovDRx)
1536	{
1537	if ( pVmcb->ctrl.u16InterceptRdDRx != 0xffff
1538	\|\| pVmcb->ctrl.u16InterceptWrDRx != 0xffff)
1539	{
1540	pVmcb->ctrl.u16InterceptRdDRx = 0xffff;
1541	pVmcb->ctrl.u16InterceptWrDRx = 0xffff;
1542	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1543	}
1544	}
1545	else
1546	{
1547	if ( pVmcb->ctrl.u16InterceptRdDRx
1548	\|\| pVmcb->ctrl.u16InterceptWrDRx)
1549	{
1550	pVmcb->ctrl.u16InterceptRdDRx = 0;
1551	pVmcb->ctrl.u16InterceptWrDRx = 0;
1552	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1553	}
1554	}
1555
1556	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_DEBUG);
1557	}
1558
1559
1560	/**
1561	* Loads the guest APIC state (currently just the TPR).
1562	*
1563	* @returns VBox status code.
1564	* @param pVCpu Pointer to the VMCPU.
1565	* @param pVmcb Pointer to the VM control block.
1566	* @param pCtx Pointer to the guest-CPU context.
1567	*/
1568	static int hmR0SvmLoadGuestApicState(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1569	{
1570	if (!HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE))
1571	return VINF_SUCCESS;
1572
1573	bool fPendingIntr;
1574	uint8_t u8Tpr;
1575	int rc = PDMApicGetTPR(pVCpu, &u8Tpr, &fPendingIntr, NULL /* pu8PendingIrq */);
1576	AssertRCReturn(rc, rc);
1577
1578	/* Assume that we need to trap all TPR accesses and thus need not check on
1579	every #VMEXIT if we should update the TPR. */
1580	Assert(pVmcb->ctrl.IntCtrl.n.u1VIrqMasking);
1581	pVCpu->hm.s.svm.fSyncVTpr = false;
1582
1583	/* 32-bit guests uses LSTAR MSR for patching guest code which touches the TPR. */
1584	if (pVCpu->CTX_SUFF(pVM)->hm.s.fTPRPatchingActive)
1585	{
1586	pCtx->msrLSTAR = u8Tpr;
1587
1588	/* If there are interrupts pending, intercept LSTAR writes, otherwise don't intercept reads or writes. */
1589	if (fPendingIntr)
1590	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_INTERCEPT_WRITE);
1591	else
1592	{
1593	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1594	pVCpu->hm.s.svm.fSyncVTpr = true;
1595	}
1596	}
1597	else
1598	{
1599	/* Bits 3-0 of the VTPR field correspond to bits 7-4 of the TPR (which is the Task-Priority Class). */
1600	pVmcb->ctrl.IntCtrl.n.u8VTPR = (u8Tpr >> 4);
1601
1602	/* If there are interrupts pending, intercept CR8 writes to evaluate ASAP if we can deliver the interrupt to the guest. */
1603	if (fPendingIntr)
1604	pVmcb->ctrl.u16InterceptWrCRx \|= RT_BIT(8);
1605	else
1606	{
1607	pVmcb->ctrl.u16InterceptWrCRx &= ~RT_BIT(8);
1608	pVCpu->hm.s.svm.fSyncVTpr = true;
1609	}
1610
1611	pVmcb->ctrl.u64VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS \| HMSVM_VMCB_CLEAN_TPR);
1612	}
1613
1614	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
1615	return rc;
1616	}
1617
1618
1619	/**
1620	* Sets up the appropriate function to run guest code.
1621	*
1622	* @returns VBox status code.
1623	* @param pVCpu Pointer to the VMCPU.
1624	* @param pCtx Pointer to the guest-CPU context.
1625	*
1626	* @remarks No-long-jump zone!!!
1627	*/
1628	static int hmR0SvmSetupVMRunHandler(PVMCPU pVCpu, PCPUMCTX pCtx)
1629	{
1630	if (CPUMIsGuestInLongModeEx(pCtx))
1631	{
1632	#ifndef VBOX_ENABLE_64_BITS_GUESTS
1633	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1634	#endif
1635	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.fAllow64BitGuests); /* Guaranteed by hmR3InitFinalizeR0(). */
1636	#if HC_ARCH_BITS == 32 && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
1637	/* 32-bit host. We need to switch to 64-bit before running the 64-bit guest. */
1638	pVCpu->hm.s.svm.pfnVMRun = SVMR0VMSwitcherRun64;
1639	#else
1640	/* 64-bit host or hybrid host. */
1641	pVCpu->hm.s.svm.pfnVMRun = SVMR0VMRun64;
1642	#endif
1643	}
1644	else
1645	{
1646	/* Guest is not in long mode, use the 32-bit handler. */
1647	pVCpu->hm.s.svm.pfnVMRun = SVMR0VMRun;
1648	}
1649	return VINF_SUCCESS;
1650	}
1651
1652
1653	/**
1654	* Enters the AMD-V session.
1655	*
1656	* @returns VBox status code.
1657	* @param pVM Pointer to the VM.
1658	* @param pVCpu Pointer to the VMCPU.
1659	* @param pCpu Pointer to the CPU info struct.
1660	*/
1661	VMMR0DECL(int) SVMR0Enter(PVM pVM, PVMCPU pVCpu, PHMGLOBALCPUINFO pCpu)
1662	{
1663	AssertPtr(pVM);
1664	AssertPtr(pVCpu);
1665	Assert(pVM->hm.s.svm.fSupported);
1666	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1667	NOREF(pVM); NOREF(pCpu);
1668
1669	LogFlowFunc(("pVM=%p pVCpu=%p\n", pVM, pVCpu));
1670	Assert(HMCPU_CF_IS_SET(pVCpu, HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_HOST_GUEST_SHARED_STATE));
1671
1672	pVCpu->hm.s.fLeaveDone = false;
1673	return VINF_SUCCESS;
1674	}
1675
1676
1677	/**
1678	* Thread-context callback for AMD-V.
1679	*
1680	* @param enmEvent The thread-context event.
1681	* @param pVCpu Pointer to the VMCPU.
1682	* @param fGlobalInit Whether global VT-x/AMD-V init. is used.
1683	* @thread EMT(pVCpu)
1684	*/
1685	VMMR0DECL(void) SVMR0ThreadCtxCallback(RTTHREADCTXEVENT enmEvent, PVMCPU pVCpu, bool fGlobalInit)
1686	{
1687	NOREF(fGlobalInit);
1688
1689	switch (enmEvent)
1690	{
1691	case RTTHREADCTXEVENT_PREEMPTING:
1692	{
1693	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1694	Assert(VMMR0ThreadCtxHooksAreRegistered(pVCpu));
1695	VMCPU_ASSERT_EMT(pVCpu);
1696
1697	PVM pVM = pVCpu->CTX_SUFF(pVM);
1698	PCPUMCTX pCtx = CPUMQueryGuestCtxPtr(pVCpu);
1699
1700	/* No longjmps (log-flush, locks) in this fragile context. */
1701	VMMRZCallRing3Disable(pVCpu);
1702
1703	if (!pVCpu->hm.s.fLeaveDone)
1704	{
1705	hmR0SvmLeave(pVM, pVCpu, pCtx);
1706	pVCpu->hm.s.fLeaveDone = true;
1707	}
1708
1709	/* Leave HM context, takes care of local init (term). */
1710	int rc = HMR0LeaveCpu(pVCpu);
1711	AssertRC(rc); NOREF(rc);
1712
1713	/* Restore longjmp state. */
1714	VMMRZCallRing3Enable(pVCpu);
1715	STAM_COUNTER_INC(&pVCpu->hm.s.StatPreemptPreempting);
1716	break;
1717	}
1718
1719	case RTTHREADCTXEVENT_RESUMED:
1720	{
1721	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1722	Assert(VMMR0ThreadCtxHooksAreRegistered(pVCpu));
1723	VMCPU_ASSERT_EMT(pVCpu);
1724
1725	/* No longjmps (log-flush, locks) in this fragile context. */
1726	VMMRZCallRing3Disable(pVCpu);
1727
1728	/*
1729	* Initialize the bare minimum state required for HM. This takes care of
1730	* initializing AMD-V if necessary (onlined CPUs, local init etc.)
1731	*/
1732	int rc = HMR0EnterCpu(pVCpu);
1733	AssertRC(rc); NOREF(rc);
1734	Assert(HMCPU_CF_IS_SET(pVCpu, HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_HOST_GUEST_SHARED_STATE));
1735
1736	pVCpu->hm.s.fLeaveDone = false;
1737
1738	/* Restore longjmp state. */
1739	VMMRZCallRing3Enable(pVCpu);
1740	break;
1741	}
1742
1743	default:
1744	break;
1745	}
1746	}
1747
1748
1749	/**
1750	* Saves the host state.
1751	*
1752	* @returns VBox status code.
1753	* @param pVM Pointer to the VM.
1754	* @param pVCpu Pointer to the VMCPU.
1755	*
1756	* @remarks No-long-jump zone!!!
1757	*/
1758	VMMR0DECL(int) SVMR0SaveHostState(PVM pVM, PVMCPU pVCpu)
1759	{
1760	NOREF(pVM);
1761	NOREF(pVCpu);
1762	/* Nothing to do here. AMD-V does this for us automatically during the world-switch. */
1763	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_HOST_CONTEXT);
1764	return VINF_SUCCESS;
1765	}
1766
1767
1768	/**
1769	* Loads the guest state into the VMCB. The CPU state will be loaded from these
1770	* fields on every successful VM-entry.
1771	*
1772	* Also sets up the appropriate VMRUN function to execute guest code based on
1773	* the guest CPU mode.
1774	*
1775	* @returns VBox status code.
1776	* @param pVM Pointer to the VM.
1777	* @param pVCpu Pointer to the VMCPU.
1778	* @param pCtx Pointer to the guest-CPU context.
1779	*
1780	* @remarks No-long-jump zone!!!
1781	*/
1782	static int hmR0SvmLoadGuestState(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
1783	{
1784	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
1785	AssertMsgReturn(pVmcb, ("Invalid pVmcb\n"), VERR_SVM_INVALID_PVMCB);
1786
1787	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatLoadGuestState, x);
1788
1789	int rc = hmR0SvmLoadGuestControlRegs(pVCpu, pVmcb, pCtx);
1790	AssertLogRelMsgRCReturn(rc, ("hmR0SvmLoadGuestControlRegs! rc=%Rrc (pVM=%p pVCpu=%p)\n", rc, pVM, pVCpu), rc);
1791
1792	hmR0SvmLoadGuestSegmentRegs(pVCpu, pVmcb, pCtx);
1793	hmR0SvmLoadGuestMsrs(pVCpu, pVmcb, pCtx);
1794
1795	pVmcb->guest.u64RIP = pCtx->rip;
1796	pVmcb->guest.u64RSP = pCtx->rsp;
1797	pVmcb->guest.u64RFlags = pCtx->eflags.u32;
1798	pVmcb->guest.u64RAX = pCtx->rax;
1799
1800	rc = hmR0SvmLoadGuestApicState(pVCpu, pVmcb, pCtx);
1801	AssertLogRelMsgRCReturn(rc, ("hmR0SvmLoadGuestApicState! rc=%Rrc (pVM=%p pVCpu=%p)\n", rc, pVM, pVCpu), rc);
1802
1803	rc = hmR0SvmSetupVMRunHandler(pVCpu, pCtx);
1804	AssertLogRelMsgRCReturn(rc, ("hmR0SvmSetupVMRunHandler! rc=%Rrc (pVM=%p pVCpu=%p)\n", rc, pVM, pVCpu), rc);
1805
1806	/* Clear any unused and reserved bits. */
1807	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_RIP /* Unused (loaded unconditionally). */
1808	\| HM_CHANGED_GUEST_RSP
1809	\| HM_CHANGED_GUEST_RFLAGS
1810	\| HM_CHANGED_GUEST_SYSENTER_CS_MSR
1811	\| HM_CHANGED_GUEST_SYSENTER_EIP_MSR
1812	\| HM_CHANGED_GUEST_SYSENTER_ESP_MSR
1813	\| HM_CHANGED_GUEST_LAZY_MSRS /* Unused. */
1814	\| HM_CHANGED_SVM_RESERVED1 /* Reserved. */
1815	\| HM_CHANGED_SVM_RESERVED2
1816	\| HM_CHANGED_SVM_RESERVED3);
1817
1818	/* All the guest state bits should be loaded except maybe the host context and/or shared host/guest bits. */
1819	AssertMsg( !HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_ALL_GUEST)
1820	\|\| HMCPU_CF_IS_PENDING_ONLY(pVCpu, HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_HOST_GUEST_SHARED_STATE),
1821	("fContextUseFlags=%#RX32\n", HMCPU_CF_VALUE(pVCpu)));
1822
1823	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));
1824	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatLoadGuestState, x);
1825	return rc;
1826	}
1827
1828
1829	/**
1830	* Loads the state shared between the host and guest into the
1831	* VMCB.
1832	*
1833	* @param pVCpu Pointer to the VMCPU.
1834	* @param pVmcb Pointer to the VM control block.
1835	* @param pCtx Pointer to the guest-CPU context.
1836	*
1837	* @remarks No-long-jump zone!!!
1838	*/
1839	static void hmR0SvmLoadSharedState(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1840	{
1841	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1842	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
1843
1844	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR0))
1845	hmR0SvmLoadSharedCR0(pVCpu, pVmcb, pCtx);
1846
1847	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_DEBUG))
1848	hmR0SvmLoadSharedDebugState(pVCpu, pVmcb, pCtx);
1849
1850	/* Unused on AMD-V. */
1851	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_LAZY_MSRS);
1852
1853	AssertMsg(!HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_HOST_GUEST_SHARED_STATE),
1854	("fContextUseFlags=%#RX32\n", HMCPU_CF_VALUE(pVCpu)));
1855	}
1856
1857
1858	/**
1859	* Saves the entire guest state from the VMCB into the
1860	* guest-CPU context. Currently there is no residual state left in the CPU that
1861	* is not updated in the VMCB.
1862	*
1863	* @returns VBox status code.
1864	* @param pVCpu Pointer to the VMCPU.
1865	* @param pMixedCtx Pointer to the guest-CPU context. The data may be
1866	* out-of-sync. Make sure to update the required fields
1867	* before using them.
1868	*/
1869	static void hmR0SvmSaveGuestState(PVMCPU pVCpu, PCPUMCTX pMixedCtx)
1870	{
1871	Assert(VMMRZCallRing3IsEnabled(pVCpu));
1872
1873	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
1874
1875	pMixedCtx->rip = pVmcb->guest.u64RIP;
1876	pMixedCtx->rsp = pVmcb->guest.u64RSP;
1877	pMixedCtx->eflags.u32 = pVmcb->guest.u64RFlags;
1878	pMixedCtx->rax = pVmcb->guest.u64RAX;
1879
1880	/*
1881	* Guest interrupt shadow.
1882	*/
1883	if (pVmcb->ctrl.u64IntShadow & SVM_INTERRUPT_SHADOW_ACTIVE)
1884	EMSetInhibitInterruptsPC(pVCpu, pMixedCtx->rip);
1885	else
1886	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS);
1887
1888	/*
1889	* Guest Control registers: CR2, CR3 (handled at the end) - accesses to other control registers are always intercepted.
1890	*/
1891	pMixedCtx->cr2 = pVmcb->guest.u64CR2;
1892
1893	/*
1894	* Guest MSRs.
1895	*/
1896	pMixedCtx->msrSTAR = pVmcb->guest.u64STAR; /* legacy syscall eip, cs & ss */
1897	pMixedCtx->msrLSTAR = pVmcb->guest.u64LSTAR; /* 64-bit mode syscall rip */
1898	pMixedCtx->msrCSTAR = pVmcb->guest.u64CSTAR; /* compatibility mode syscall rip */
1899	pMixedCtx->msrSFMASK = pVmcb->guest.u64SFMASK; /* syscall flag mask */
1900	pMixedCtx->msrKERNELGSBASE = pVmcb->guest.u64KernelGSBase; /* swapgs exchange value */
1901	pMixedCtx->SysEnter.cs = pVmcb->guest.u64SysEnterCS;
1902	pMixedCtx->SysEnter.eip = pVmcb->guest.u64SysEnterEIP;
1903	pMixedCtx->SysEnter.esp = pVmcb->guest.u64SysEnterESP;
1904
1905	/*
1906	* Guest segment registers (includes FS, GS base MSRs for 64-bit guests).
1907	*/
1908	HMSVM_SAVE_SEG_REG(CS, cs);
1909	HMSVM_SAVE_SEG_REG(SS, ss);
1910	HMSVM_SAVE_SEG_REG(DS, ds);
1911	HMSVM_SAVE_SEG_REG(ES, es);
1912	HMSVM_SAVE_SEG_REG(FS, fs);
1913	HMSVM_SAVE_SEG_REG(GS, gs);
1914
1915	/*
1916	* Correct the hidden CS granularity bit. Haven't seen it being wrong in any other
1917	* register (yet).
1918	*/
1919	/** @todo SELM might need to be fixed as it too should not care about the
1920	* granularity bit. See @bugref{6785}. */
1921	if ( !pMixedCtx->cs.Attr.n.u1Granularity
1922	&& pMixedCtx->cs.Attr.n.u1Present
1923	&& pMixedCtx->cs.u32Limit > UINT32_C(0xfffff))
1924	{
1925	Assert((pMixedCtx->cs.u32Limit & 0xfff) == 0xfff);
1926	pMixedCtx->cs.Attr.n.u1Granularity = 1;
1927	}
1928
1929	#ifdef VBOX_STRICT
1930	# define HMSVM_ASSERT_SEG_GRANULARITY(reg) \
1931	AssertMsg( !pMixedCtx->reg.Attr.n.u1Present \
1932	\|\| ( pMixedCtx->reg.Attr.n.u1Granularity \
1933	? (pMixedCtx->reg.u32Limit & 0xfff) == 0xfff \
1934	: pMixedCtx->reg.u32Limit <= UINT32_C(0xfffff)), \
1935	("Invalid Segment Attributes Limit=%#RX32 Attr=%#RX32 Base=%#RX64\n", pMixedCtx->reg.u32Limit, \
1936	pMixedCtx->reg.Attr.u, pMixedCtx->reg.u64Base))
1937
1938	HMSVM_ASSERT_SEG_GRANULARITY(cs);
1939	HMSVM_ASSERT_SEG_GRANULARITY(ss);
1940	HMSVM_ASSERT_SEG_GRANULARITY(ds);
1941	HMSVM_ASSERT_SEG_GRANULARITY(es);
1942	HMSVM_ASSERT_SEG_GRANULARITY(fs);
1943	HMSVM_ASSERT_SEG_GRANULARITY(gs);
1944
1945	# undef HMSVM_ASSERT_SEL_GRANULARITY
1946	#endif
1947
1948	/*
1949	* Sync the hidden SS DPL field. AMD CPUs have a separate CPL field in the VMCB and uses that
1950	* and thus it's possible that when the CPL changes during guest execution that the SS DPL
1951	* isn't updated by AMD-V. Observed on some AMD Fusion CPUs with 64-bit guests.
1952	* See AMD spec. 15.5.1 "Basic operation".
1953	*/
1954	Assert(!(pVmcb->guest.u8CPL & ~0x3));
1955	pMixedCtx->ss.Attr.n.u2Dpl = pVmcb->guest.u8CPL & 0x3;
1956
1957	/*
1958	* Guest Descriptor-Table registers.
1959	*/
1960	HMSVM_SAVE_SEG_REG(TR, tr);
1961	HMSVM_SAVE_SEG_REG(LDTR, ldtr);
1962	pMixedCtx->gdtr.cbGdt = pVmcb->guest.GDTR.u32Limit;
1963	pMixedCtx->gdtr.pGdt = pVmcb->guest.GDTR.u64Base;
1964
1965	pMixedCtx->idtr.cbIdt = pVmcb->guest.IDTR.u32Limit;
1966	pMixedCtx->idtr.pIdt = pVmcb->guest.IDTR.u64Base;
1967
1968	/*
1969	* Guest Debug registers.
1970	*/
1971	if (!pVCpu->hm.s.fUsingHyperDR7)
1972	{
1973	pMixedCtx->dr[6] = pVmcb->guest.u64DR6;
1974	pMixedCtx->dr[7] = pVmcb->guest.u64DR7;
1975	}
1976	else
1977	{
1978	Assert(pVmcb->guest.u64DR7 == CPUMGetHyperDR7(pVCpu));
1979	CPUMSetHyperDR6(pVCpu, pVmcb->guest.u64DR6);
1980	}
1981
1982	/*
1983	* With Nested Paging, CR3 changes are not intercepted. Therefore, sync. it now.
1984	* This is done as the very last step of syncing the guest state, as PGMUpdateCR3() may cause longjmp's to ring-3.
1985	*/
1986	if ( pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging
1987	&& pMixedCtx->cr3 != pVmcb->guest.u64CR3)
1988	{
1989	CPUMSetGuestCR3(pVCpu, pVmcb->guest.u64CR3);
1990	PGMUpdateCR3(pVCpu, pVmcb->guest.u64CR3);
1991	}
1992	}
1993
1994
1995	/**
1996	* Does the necessary state syncing before returning to ring-3 for any reason
1997	* (longjmp, preemption, voluntary exits to ring-3) from AMD-V.
1998	*
1999	* @param pVM Pointer to the VM.
2000	* @param pVCpu Pointer to the VMCPU.
2001	* @param pMixedCtx Pointer to the guest-CPU context.
2002	*
2003	* @remarks No-long-jmp zone!!!
2004	*/
2005	static void hmR0SvmLeave(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
2006	{
2007	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2008	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2009	Assert(VMMR0IsLogFlushDisabled(pVCpu));
2010
2011	/*
2012	* !!! IMPORTANT !!!
2013	* If you modify code here, make sure to check whether hmR0SvmCallRing3Callback() needs to be updated too.
2014	*/
2015
2016	/* Restore host FPU state if necessary and resync on next R0 reentry .*/
2017	if (CPUMIsGuestFPUStateActive(pVCpu))
2018	{
2019	CPUMR0SaveGuestFPU(pVM, pVCpu, pCtx);
2020	Assert(!CPUMIsGuestFPUStateActive(pVCpu));
2021	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR0);
2022	}
2023
2024	/*
2025	* Restore host debug registers if necessary and resync on next R0 reentry.
2026	*/
2027	#ifdef VBOX_STRICT
2028	if (CPUMIsHyperDebugStateActive(pVCpu))
2029	{
2030	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2031	Assert(pVmcb->ctrl.u16InterceptRdDRx == 0xffff);
2032	Assert(pVmcb->ctrl.u16InterceptWrDRx == 0xffff);
2033	}
2034	#endif
2035	if (CPUMR0DebugStateMaybeSaveGuestAndRestoreHost(pVCpu, false /* save DR6 */))
2036	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_DEBUG);
2037
2038	Assert(!CPUMIsHyperDebugStateActive(pVCpu));
2039	Assert(!CPUMIsGuestDebugStateActive(pVCpu));
2040
2041	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatEntry);
2042	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatLoadGuestState);
2043	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatExit1);
2044	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatExit2);
2045	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchLongJmpToR3);
2046
2047	VMCPU_CMPXCHG_STATE(pVCpu, VMCPUSTATE_STARTED_HM, VMCPUSTATE_STARTED_EXEC);
2048	}
2049
2050
2051	/**
2052	* Leaves the AMD-V session.
2053	*
2054	* @returns VBox status code.
2055	* @param pVM Pointer to the VM.
2056	* @param pVCpu Pointer to the VMCPU.
2057	* @param pCtx Pointer to the guest-CPU context.
2058	*/
2059	static int hmR0SvmLeaveSession(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
2060	{
2061	HM_DISABLE_PREEMPT_IF_NEEDED();
2062	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2063	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2064
2065	/* When thread-context hooks are used, we can avoid doing the leave again if we had been preempted before
2066	and done this from the SVMR0ThreadCtxCallback(). */
2067	if (!pVCpu->hm.s.fLeaveDone)
2068	{
2069	hmR0SvmLeave(pVM, pVCpu, pCtx);
2070	pVCpu->hm.s.fLeaveDone = true;
2071	}
2072
2073	/*
2074	* !!! IMPORTANT !!!
2075	* If you modify code here, make sure to check whether hmR0SvmCallRing3Callback() needs to be updated too.
2076	*/
2077
2078	/* Deregister hook now that we've left HM context before re-enabling preemption. */
2079	if (VMMR0ThreadCtxHooksAreRegistered(pVCpu))
2080	VMMR0ThreadCtxHooksDeregister(pVCpu);
2081
2082	/* Leave HM context. This takes care of local init (term). */
2083	int rc = HMR0LeaveCpu(pVCpu);
2084
2085	HM_RESTORE_PREEMPT_IF_NEEDED();
2086	return rc;
2087	}
2088
2089
2090	/**
2091	* Does the necessary state syncing before doing a longjmp to ring-3.
2092	*
2093	* @returns VBox status code.
2094	* @param pVM Pointer to the VM.
2095	* @param pVCpu Pointer to the VMCPU.
2096	* @param pCtx Pointer to the guest-CPU context.
2097	*
2098	* @remarks No-long-jmp zone!!!
2099	*/
2100	static int hmR0SvmLongJmpToRing3(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
2101	{
2102	return hmR0SvmLeaveSession(pVM, pVCpu, pCtx);
2103	}
2104
2105
2106	/**
2107	* VMMRZCallRing3() callback wrapper which saves the guest state (or restores
2108	* any remaining host state) before we longjump to ring-3 and possibly get
2109	* preempted.
2110	*
2111	* @param pVCpu Pointer to the VMCPU.
2112	* @param enmOperation The operation causing the ring-3 longjump.
2113	* @param pvUser The user argument (pointer to the possibly
2114	* out-of-date guest-CPU context).
2115	*/
2116	DECLCALLBACK(int) hmR0SvmCallRing3Callback(PVMCPU pVCpu, VMMCALLRING3 enmOperation, void *pvUser)
2117	{
2118	if (enmOperation == VMMCALLRING3_VM_R0_ASSERTION)
2119	{
2120	/*
2121	* !!! IMPORTANT !!!
2122	* If you modify code here, make sure to check whether hmR0SvmLeave() and hmR0SvmLeaveSession() needs
2123	* to be updated too. This is a stripped down version which gets out ASAP trying to not trigger any assertion.
2124	*/
2125	VMMRZCallRing3RemoveNotification(pVCpu);
2126	VMMRZCallRing3Disable(pVCpu);
2127	HM_DISABLE_PREEMPT_IF_NEEDED();
2128
2129	/* Restore host FPU state if necessary and resync on next R0 reentry .*/
2130	if (CPUMIsGuestFPUStateActive(pVCpu))
2131	CPUMR0SaveGuestFPU(pVCpu->CTX_SUFF(pVM), pVCpu, (PCPUMCTX)pvUser);
2132
2133	/* Restore host debug registers if necessary and resync on next R0 reentry. */
2134	CPUMR0DebugStateMaybeSaveGuestAndRestoreHost(pVCpu, false /* save DR6 */);
2135
2136	/* Deregister hook now that we've left HM context before re-enabling preemption. */
2137	if (VMMR0ThreadCtxHooksAreRegistered(pVCpu))
2138	VMMR0ThreadCtxHooksDeregister(pVCpu);
2139
2140	/* Leave HM context. This takes care of local init (term). */
2141	HMR0LeaveCpu(pVCpu);
2142
2143	HM_RESTORE_PREEMPT_IF_NEEDED();
2144	return VINF_SUCCESS;
2145	}
2146
2147	Assert(pVCpu);
2148	Assert(pvUser);
2149	Assert(VMMRZCallRing3IsEnabled(pVCpu));
2150	HMSVM_ASSERT_PREEMPT_SAFE();
2151
2152	VMMRZCallRing3Disable(pVCpu);
2153	Assert(VMMR0IsLogFlushDisabled(pVCpu));
2154
2155	Log4(("hmR0SvmCallRing3Callback->hmR0SvmLongJmpToRing3\n"));
2156	int rc = hmR0SvmLongJmpToRing3(pVCpu->CTX_SUFF(pVM), pVCpu, (PCPUMCTX)pvUser);
2157	AssertRCReturn(rc, rc);
2158
2159	VMMRZCallRing3Enable(pVCpu);
2160	return VINF_SUCCESS;
2161	}
2162
2163
2164	/**
2165	* Take necessary actions before going back to ring-3.
2166	*
2167	* An action requires us to go back to ring-3. This function does the necessary
2168	* steps before we can safely return to ring-3. This is not the same as longjmps
2169	* to ring-3, this is voluntary.
2170	*
2171	* @param pVM Pointer to the VM.
2172	* @param pVCpu Pointer to the VMCPU.
2173	* @param pCtx Pointer to the guest-CPU context.
2174	* @param rcExit The reason for exiting to ring-3. Can be
2175	* VINF_VMM_UNKNOWN_RING3_CALL.
2176	*/
2177	static void hmR0SvmExitToRing3(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, int rcExit)
2178	{
2179	Assert(pVM);
2180	Assert(pVCpu);
2181	Assert(pCtx);
2182	HMSVM_ASSERT_PREEMPT_SAFE();
2183
2184	/* Please, no longjumps here (any logging shouldn't flush jump back to ring-3). NO LOGGING BEFORE THIS POINT! */
2185	VMMRZCallRing3Disable(pVCpu);
2186	Log4(("hmR0SvmExitToRing3: rcExit=%d\n", rcExit));
2187
2188	/* We need to do this only while truly exiting the "inner loop" back to ring-3 and -not- for any longjmp to ring3. */
2189	if (pVCpu->hm.s.Event.fPending)
2190	{
2191	hmR0SvmPendingEventToTrpmTrap(pVCpu);
2192	Assert(!pVCpu->hm.s.Event.fPending);
2193	}
2194
2195	/* Sync. the necessary state for going back to ring-3. */
2196	hmR0SvmLeaveSession(pVM, pVCpu, pCtx);
2197	STAM_COUNTER_DEC(&pVCpu->hm.s.StatSwitchLongJmpToR3);
2198
2199	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_TO_R3);
2200	CPUMSetChangedFlags(pVCpu, CPUM_CHANGED_SYSENTER_MSR
2201	\| CPUM_CHANGED_LDTR
2202	\| CPUM_CHANGED_GDTR
2203	\| CPUM_CHANGED_IDTR
2204	\| CPUM_CHANGED_TR
2205	\| CPUM_CHANGED_HIDDEN_SEL_REGS);
2206	if ( pVM->hm.s.fNestedPaging
2207	&& CPUMIsGuestPagingEnabledEx(pCtx))
2208	{
2209	CPUMSetChangedFlags(pVCpu, CPUM_CHANGED_GLOBAL_TLB_FLUSH);
2210	}
2211
2212	/* On our way back from ring-3 reload the guest state if there is a possibility of it being changed. */
2213	if (rcExit != VINF_EM_RAW_INTERRUPT)
2214	HMCPU_CF_SET(pVCpu, HM_CHANGED_ALL_GUEST);
2215
2216	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchExitToR3);
2217
2218	/* We do -not- want any longjmp notifications after this! We must return to ring-3 ASAP. */
2219	VMMRZCallRing3RemoveNotification(pVCpu);
2220	VMMRZCallRing3Enable(pVCpu);
2221	}
2222
2223
2224	/**
2225	* Updates the use of TSC offsetting mode for the CPU and adjusts the necessary
2226	* intercepts.
2227	*
2228	* @param pVCpu Pointer to the VMCPU.
2229	*
2230	* @remarks No-long-jump zone!!!
2231	*/
2232	static void hmR0SvmUpdateTscOffsetting(PVMCPU pVCpu)
2233	{
2234	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2235	if (TMCpuTickCanUseRealTSC(pVCpu, &pVmcb->ctrl.u64TSCOffset))
2236	{
2237	uint64_t u64CurTSC = ASMReadTSC();
2238	if (u64CurTSC + pVmcb->ctrl.u64TSCOffset > TMCpuTickGetLastSeen(pVCpu))
2239	{
2240	pVmcb->ctrl.u32InterceptCtrl1 &= ~SVM_CTRL1_INTERCEPT_RDTSC;
2241	pVmcb->ctrl.u32InterceptCtrl2 &= ~SVM_CTRL2_INTERCEPT_RDTSCP;
2242	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscOffset);
2243	}
2244	else
2245	{
2246	pVmcb->ctrl.u32InterceptCtrl1 \|= SVM_CTRL1_INTERCEPT_RDTSC;
2247	pVmcb->ctrl.u32InterceptCtrl2 \|= SVM_CTRL2_INTERCEPT_RDTSCP;
2248	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscInterceptOverFlow);
2249	}
2250	}
2251	else
2252	{
2253	pVmcb->ctrl.u32InterceptCtrl1 \|= SVM_CTRL1_INTERCEPT_RDTSC;
2254	pVmcb->ctrl.u32InterceptCtrl2 \|= SVM_CTRL2_INTERCEPT_RDTSCP;
2255	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscIntercept);
2256	}
2257
2258	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2259	}
2260
2261
2262	/**
2263	* Sets an event as a pending event to be injected into the guest.
2264	*
2265	* @param pVCpu Pointer to the VMCPU.
2266	* @param pEvent Pointer to the SVM event.
2267	* @param GCPtrFaultAddress The fault-address (CR2) in case it's a
2268	* page-fault.
2269	*
2270	* @remarks Statistics counter assumes this is a guest event being reflected to
2271	* the guest i.e. 'StatInjectPendingReflect' is incremented always.
2272	*/
2273	DECLINLINE(void) hmR0SvmSetPendingEvent(PVMCPU pVCpu, PSVMEVENT pEvent, RTGCUINTPTR GCPtrFaultAddress)
2274	{
2275	Assert(!pVCpu->hm.s.Event.fPending);
2276	Assert(pEvent->n.u1Valid);
2277
2278	pVCpu->hm.s.Event.u64IntInfo = pEvent->u;
2279	pVCpu->hm.s.Event.fPending = true;
2280	pVCpu->hm.s.Event.GCPtrFaultAddress = GCPtrFaultAddress;
2281
2282	Log4(("hmR0SvmSetPendingEvent: u=%#RX64 u8Vector=%#x Type=%#x ErrorCodeValid=%RTbool ErrorCode=%#RX32\n", pEvent->u,
2283	pEvent->n.u8Vector, (uint8_t)pEvent->n.u3Type, !!pEvent->n.u1ErrorCodeValid, pEvent->n.u32ErrorCode));
2284
2285	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectPendingReflect);
2286	}
2287
2288
2289	/**
2290	* Injects an event into the guest upon VMRUN by updating the relevant field
2291	* in the VMCB.
2292	*
2293	* @param pVCpu Pointer to the VMCPU.
2294	* @param pVmcb Pointer to the guest VM control block.
2295	* @param pCtx Pointer to the guest-CPU context.
2296	* @param pEvent Pointer to the event.
2297	*
2298	* @remarks No-long-jump zone!!!
2299	* @remarks Requires CR0!
2300	*/
2301	DECLINLINE(void) hmR0SvmInjectEventVmcb(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx, PSVMEVENT pEvent)
2302	{
2303	NOREF(pVCpu); NOREF(pCtx);
2304
2305	pVmcb->ctrl.EventInject.u = pEvent->u;
2306	STAM_COUNTER_INC(&pVCpu->hm.s.paStatInjectedIrqsR0[pEvent->n.u8Vector & MASK_INJECT_IRQ_STAT]);
2307
2308	Log4(("hmR0SvmInjectEventVmcb: u=%#RX64 u8Vector=%#x Type=%#x ErrorCodeValid=%RTbool ErrorCode=%#RX32\n", pEvent->u,
2309	pEvent->n.u8Vector, (uint8_t)pEvent->n.u3Type, !!pEvent->n.u1ErrorCodeValid, pEvent->n.u32ErrorCode));
2310	}
2311
2312
2313
2314	/**
2315	* Converts any TRPM trap into a pending HM event. This is typically used when
2316	* entering from ring-3 (not longjmp returns).
2317	*
2318	* @param pVCpu Pointer to the VMCPU.
2319	*/
2320	static void hmR0SvmTrpmTrapToPendingEvent(PVMCPU pVCpu)
2321	{
2322	Assert(TRPMHasTrap(pVCpu));
2323	Assert(!pVCpu->hm.s.Event.fPending);
2324
2325	uint8_t uVector;
2326	TRPMEVENT enmTrpmEvent;
2327	RTGCUINT uErrCode;
2328	RTGCUINTPTR GCPtrFaultAddress;
2329	uint8_t cbInstr;
2330
2331	int rc = TRPMQueryTrapAll(pVCpu, &uVector, &enmTrpmEvent, &uErrCode, &GCPtrFaultAddress, &cbInstr);
2332	AssertRC(rc);
2333
2334	SVMEVENT Event;
2335	Event.u = 0;
2336	Event.n.u1Valid = 1;
2337	Event.n.u8Vector = uVector;
2338
2339	/* Refer AMD spec. 15.20 "Event Injection" for the format. */
2340	if (enmTrpmEvent == TRPM_TRAP)
2341	{
2342	Event.n.u3Type = SVM_EVENT_EXCEPTION;
2343	switch (uVector)
2344	{
2345	case X86_XCPT_PF:
2346	case X86_XCPT_DF:
2347	case X86_XCPT_TS:
2348	case X86_XCPT_NP:
2349	case X86_XCPT_SS:
2350	case X86_XCPT_GP:
2351	case X86_XCPT_AC:
2352	{
2353	Event.n.u1ErrorCodeValid = 1;
2354	Event.n.u32ErrorCode = uErrCode;
2355	break;
2356	}
2357	}
2358	}
2359	else if (enmTrpmEvent == TRPM_HARDWARE_INT)
2360	{
2361	if (uVector == X86_XCPT_NMI)
2362	Event.n.u3Type = SVM_EVENT_NMI;
2363	else
2364	Event.n.u3Type = SVM_EVENT_EXTERNAL_IRQ;
2365	}
2366	else if (enmTrpmEvent == TRPM_SOFTWARE_INT)
2367	Event.n.u3Type = SVM_EVENT_SOFTWARE_INT;
2368	else
2369	AssertMsgFailed(("Invalid TRPM event type %d\n", enmTrpmEvent));
2370
2371	rc = TRPMResetTrap(pVCpu);
2372	AssertRC(rc);
2373
2374	Log4(("TRPM->HM event: u=%#RX64 u8Vector=%#x uErrorCodeValid=%RTbool uErrorCode=%#RX32\n", Event.u, Event.n.u8Vector,
2375	!!Event.n.u1ErrorCodeValid, Event.n.u32ErrorCode));
2376
2377	hmR0SvmSetPendingEvent(pVCpu, &Event, GCPtrFaultAddress);
2378	STAM_COUNTER_DEC(&pVCpu->hm.s.StatInjectPendingReflect);
2379	}
2380
2381
2382	/**
2383	* Converts any pending SVM event into a TRPM trap. Typically used when leaving
2384	* AMD-V to execute any instruction.
2385	*
2386	* @param pvCpu Pointer to the VMCPU.
2387	*/
2388	static void hmR0SvmPendingEventToTrpmTrap(PVMCPU pVCpu)
2389	{
2390	Assert(pVCpu->hm.s.Event.fPending);
2391	Assert(TRPMQueryTrap(pVCpu, NULL /* pu8TrapNo /, NULL / pEnmType */) == VERR_TRPM_NO_ACTIVE_TRAP);
2392
2393	SVMEVENT Event;
2394	Event.u = pVCpu->hm.s.Event.u64IntInfo;
2395
2396	uint8_t uVector = Event.n.u8Vector;
2397	uint8_t uVectorType = Event.n.u3Type;
2398
2399	TRPMEVENT enmTrapType;
2400	switch (uVectorType)
2401	{
2402	case SVM_EVENT_EXTERNAL_IRQ:
2403	case SVM_EVENT_NMI:
2404	enmTrapType = TRPM_HARDWARE_INT;
2405	break;
2406	case SVM_EVENT_SOFTWARE_INT:
2407	enmTrapType = TRPM_SOFTWARE_INT;
2408	break;
2409	case SVM_EVENT_EXCEPTION:
2410	enmTrapType = TRPM_TRAP;
2411	break;
2412	default:
2413	AssertMsgFailed(("Invalid pending-event type %#x\n", uVectorType));
2414	enmTrapType = TRPM_32BIT_HACK;
2415	break;
2416	}
2417
2418	Log4(("HM event->TRPM: uVector=%#x enmTrapType=%d\n", uVector, uVectorType));
2419
2420	int rc = TRPMAssertTrap(pVCpu, uVector, enmTrapType);
2421	AssertRC(rc);
2422
2423	if (Event.n.u1ErrorCodeValid)
2424	TRPMSetErrorCode(pVCpu, Event.n.u32ErrorCode);
2425
2426	if ( uVectorType == SVM_EVENT_EXCEPTION
2427	&& uVector == X86_XCPT_PF)
2428	{
2429	TRPMSetFaultAddress(pVCpu, pVCpu->hm.s.Event.GCPtrFaultAddress);
2430	Assert(pVCpu->hm.s.Event.GCPtrFaultAddress == CPUMGetGuestCR2(pVCpu));
2431	}
2432	else if (uVectorType == SVM_EVENT_SOFTWARE_INT)
2433	{
2434	AssertMsg( uVectorType == SVM_EVENT_SOFTWARE_INT
2435	\|\| (uVector == X86_XCPT_BP \|\| uVector == X86_XCPT_OF),
2436	("Invalid vector: uVector=%#x uVectorType=%#x\n", uVector, uVectorType));
2437	TRPMSetInstrLength(pVCpu, pVCpu->hm.s.Event.cbInstr);
2438	}
2439	pVCpu->hm.s.Event.fPending = false;
2440	}
2441
2442
2443	/**
2444	* Gets the guest's interrupt-shadow.
2445	*
2446	* @returns The guest's interrupt-shadow.
2447	* @param pVCpu Pointer to the VMCPU.
2448	* @param pCtx Pointer to the guest-CPU context.
2449	*
2450	* @remarks No-long-jump zone!!!
2451	* @remarks Has side-effects with VMCPU_FF_INHIBIT_INTERRUPTS force-flag.
2452	*/
2453	DECLINLINE(uint32_t) hmR0SvmGetGuestIntrShadow(PVMCPU pVCpu, PCPUMCTX pCtx)
2454	{
2455	/*
2456	* Instructions like STI and MOV SS inhibit interrupts till the next instruction completes. Check if we should
2457	* inhibit interrupts or clear any existing interrupt-inhibition.
2458	*/
2459	uint32_t uIntrState = 0;
2460	if (VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS))
2461	{
2462	if (pCtx->rip != EMGetInhibitInterruptsPC(pVCpu))
2463	{
2464	/*
2465	* We can clear the inhibit force flag as even if we go back to the recompiler without executing guest code in
2466	* AMD-V, the flag's condition to be cleared is met and thus the cleared state is correct.
2467	*/
2468	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS);
2469	}
2470	else
2471	uIntrState = SVM_INTERRUPT_SHADOW_ACTIVE;
2472	}
2473	return uIntrState;
2474	}
2475
2476
2477	/**
2478	* Sets the virtual interrupt intercept control in the VMCB which
2479	* instructs AMD-V to cause a #VMEXIT as soon as the guest is in a state to
2480	* receive interrupts.
2481	*
2482	* @param pVmcb Pointer to the VM control block.
2483	*/
2484	DECLINLINE(void) hmR0SvmSetVirtIntrIntercept(PSVMVMCB pVmcb)
2485	{
2486	if (!(pVmcb->ctrl.u32InterceptCtrl1 & SVM_CTRL1_INTERCEPT_VINTR))
2487	{
2488	pVmcb->ctrl.IntCtrl.n.u1VIrqValid = 1; /* A virtual interrupt is pending. */
2489	pVmcb->ctrl.IntCtrl.n.u8VIrqVector = 0; /* Not necessary as we #VMEXIT for delivering the interrupt. */
2490	pVmcb->ctrl.u32InterceptCtrl1 \|= SVM_CTRL1_INTERCEPT_VINTR;
2491	pVmcb->ctrl.u64VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS \| HMSVM_VMCB_CLEAN_TPR);
2492
2493	Log4(("Setting VINTR intercept\n"));
2494	}
2495	}
2496
2497
2498	/**
2499	* Evaluates the event to be delivered to the guest and sets it as the pending
2500	* event.
2501	*
2502	* @param pVCpu Pointer to the VMCPU.
2503	* @param pCtx Pointer to the guest-CPU context.
2504	*/
2505	static void hmR0SvmEvaluatePendingEvent(PVMCPU pVCpu, PCPUMCTX pCtx)
2506	{
2507	Assert(!pVCpu->hm.s.Event.fPending);
2508	Log4Func(("\n"));
2509
2510	const bool fIntShadow = !!hmR0SvmGetGuestIntrShadow(pVCpu, pCtx);
2511	const bool fBlockInt = !(pCtx->eflags.u32 & X86_EFL_IF);
2512	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2513
2514	SVMEVENT Event;
2515	Event.u = 0;
2516	/** @todo SMI. SMIs take priority over NMIs. */
2517	if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_NMI)) /* NMI. NMIs take priority over regular interrupts . */
2518	{
2519	if (!fIntShadow)
2520	{
2521	Log4(("Pending NMI\n"));
2522
2523	Event.n.u1Valid = 1;
2524	Event.n.u8Vector = X86_XCPT_NMI;
2525	Event.n.u3Type = SVM_EVENT_NMI;
2526
2527	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
2528	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INTERRUPT_NMI);
2529	}
2530	else
2531	hmR0SvmSetVirtIntrIntercept(pVmcb);
2532	}
2533	else if (VMCPU_FF_IS_PENDING(pVCpu, (VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC)))
2534	{
2535	/*
2536	* Check if the guest can receive external interrupts (PIC/APIC). Once we do PDMGetInterrupt() we -must- deliver
2537	* the interrupt ASAP. We must not execute any guest code until we inject the interrupt which is why it is
2538	* evaluated here and not set as pending, solely based on the force-flags.
2539	*/
2540	if ( !fBlockInt
2541	&& !fIntShadow)
2542	{
2543	uint8_t u8Interrupt;
2544	int rc = PDMGetInterrupt(pVCpu, &u8Interrupt);
2545	if (RT_SUCCESS(rc))
2546	{
2547	Log4(("Injecting external interrupt u8Interrupt=%#x\n", u8Interrupt));
2548
2549	Event.n.u1Valid = 1;
2550	Event.n.u8Vector = u8Interrupt;
2551	Event.n.u3Type = SVM_EVENT_EXTERNAL_IRQ;
2552
2553	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
2554	}
2555	else
2556	{
2557	/** @todo Does this actually happen? If not turn it into an assertion. */
2558	Assert(!VMCPU_FF_IS_PENDING(pVCpu, (VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC)));
2559	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchGuestIrq);
2560	}
2561	}
2562	else
2563	hmR0SvmSetVirtIntrIntercept(pVmcb);
2564	}
2565	}
2566
2567
2568	/**
2569	* Injects any pending events into the guest if the guest is in a state to
2570	* receive them.
2571	*
2572	* @param pVCpu Pointer to the VMCPU.
2573	* @param pCtx Pointer to the guest-CPU context.
2574	*/
2575	static void hmR0SvmInjectPendingEvent(PVMCPU pVCpu, PCPUMCTX pCtx)
2576	{
2577	Assert(!TRPMHasTrap(pVCpu));
2578	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2579	Log4Func(("\n"));
2580
2581	const bool fIntShadow = !!hmR0SvmGetGuestIntrShadow(pVCpu, pCtx);
2582	const bool fBlockInt = !(pCtx->eflags.u32 & X86_EFL_IF);
2583	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2584
2585	if (pVCpu->hm.s.Event.fPending) /* First, inject any pending HM events. */
2586	{
2587	SVMEVENT Event;
2588	Event.u = pVCpu->hm.s.Event.u64IntInfo;
2589	Assert(Event.n.u1Valid);
2590	#ifdef VBOX_STRICT
2591	if (Event.n.u3Type == SVM_EVENT_EXTERNAL_IRQ)
2592	{
2593	Assert(!fBlockInt);
2594	Assert(!fIntShadow);
2595	}
2596	else if (Event.n.u3Type == SVM_EVENT_NMI)
2597	Assert(!fIntShadow);
2598	#endif
2599
2600	Log4(("Injecting pending HM event.\n"));
2601	hmR0SvmInjectEventVmcb(pVCpu, pVmcb, pCtx, &Event);
2602	pVCpu->hm.s.Event.fPending = false;
2603
2604	#ifdef VBOX_WITH_STATISTICS
2605	if (Event.n.u3Type == SVM_EVENT_EXTERNAL_IRQ)
2606	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectInterrupt);
2607	else
2608	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectXcpt);
2609	#endif
2610	}
2611
2612	/* Update the guest interrupt shadow in the VMCB. */
2613	pVmcb->ctrl.u64IntShadow = !!fIntShadow;
2614	NOREF(fBlockInt);
2615	}
2616
2617
2618	/**
2619	* Reports world-switch error and dumps some useful debug info.
2620	*
2621	* @param pVM Pointer to the VM.
2622	* @param pVCpu Pointer to the VMCPU.
2623	* @param rcVMRun The return code from VMRUN (or
2624	* VERR_SVM_INVALID_GUEST_STATE for invalid
2625	* guest-state).
2626	* @param pCtx Pointer to the guest-CPU context.
2627	*/
2628	static void hmR0SvmReportWorldSwitchError(PVM pVM, PVMCPU pVCpu, int rcVMRun, PCPUMCTX pCtx)
2629	{
2630	NOREF(pCtx);
2631	HMSVM_ASSERT_PREEMPT_SAFE();
2632	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2633
2634	if (rcVMRun == VERR_SVM_INVALID_GUEST_STATE)
2635	{
2636	HMDumpRegs(pVM, pVCpu, pCtx); NOREF(pVM);
2637	#ifdef VBOX_STRICT
2638	Log4(("ctrl.u64VmcbCleanBits %#RX64\n", pVmcb->ctrl.u64VmcbCleanBits));
2639	Log4(("ctrl.u16InterceptRdCRx %#x\n", pVmcb->ctrl.u16InterceptRdCRx));
2640	Log4(("ctrl.u16InterceptWrCRx %#x\n", pVmcb->ctrl.u16InterceptWrCRx));
2641	Log4(("ctrl.u16InterceptRdDRx %#x\n", pVmcb->ctrl.u16InterceptRdDRx));
2642	Log4(("ctrl.u16InterceptWrDRx %#x\n", pVmcb->ctrl.u16InterceptWrDRx));
2643	Log4(("ctrl.u32InterceptException %#x\n", pVmcb->ctrl.u32InterceptException));
2644	Log4(("ctrl.u32InterceptCtrl1 %#x\n", pVmcb->ctrl.u32InterceptCtrl1));
2645	Log4(("ctrl.u32InterceptCtrl2 %#x\n", pVmcb->ctrl.u32InterceptCtrl2));
2646	Log4(("ctrl.u64IOPMPhysAddr %#RX64\n", pVmcb->ctrl.u64IOPMPhysAddr));
2647	Log4(("ctrl.u64MSRPMPhysAddr %#RX64\n", pVmcb->ctrl.u64MSRPMPhysAddr));
2648	Log4(("ctrl.u64TSCOffset %#RX64\n", pVmcb->ctrl.u64TSCOffset));
2649
2650	Log4(("ctrl.TLBCtrl.u32ASID %#x\n", pVmcb->ctrl.TLBCtrl.n.u32ASID));
2651	Log4(("ctrl.TLBCtrl.u8TLBFlush %#x\n", pVmcb->ctrl.TLBCtrl.n.u8TLBFlush));
2652	Log4(("ctrl.TLBCtrl.u24Reserved %#x\n", pVmcb->ctrl.TLBCtrl.n.u24Reserved));
2653
2654	Log4(("ctrl.IntCtrl.u8VTPR %#x\n", pVmcb->ctrl.IntCtrl.n.u8VTPR));
2655	Log4(("ctrl.IntCtrl.u1VIrqValid %#x\n", pVmcb->ctrl.IntCtrl.n.u1VIrqValid));
2656	Log4(("ctrl.IntCtrl.u7Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u7Reserved));
2657	Log4(("ctrl.IntCtrl.u4VIrqPriority %#x\n", pVmcb->ctrl.IntCtrl.n.u4VIrqPriority));
2658	Log4(("ctrl.IntCtrl.u1IgnoreTPR %#x\n", pVmcb->ctrl.IntCtrl.n.u1IgnoreTPR));
2659	Log4(("ctrl.IntCtrl.u3Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u3Reserved));
2660	Log4(("ctrl.IntCtrl.u1VIrqMasking %#x\n", pVmcb->ctrl.IntCtrl.n.u1VIrqMasking));
2661	Log4(("ctrl.IntCtrl.u6Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u6Reserved));
2662	Log4(("ctrl.IntCtrl.u8VIrqVector %#x\n", pVmcb->ctrl.IntCtrl.n.u8VIrqVector));
2663	Log4(("ctrl.IntCtrl.u24Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u24Reserved));
2664
2665	Log4(("ctrl.u64IntShadow %#RX64\n", pVmcb->ctrl.u64IntShadow));
2666	Log4(("ctrl.u64ExitCode %#RX64\n", pVmcb->ctrl.u64ExitCode));
2667	Log4(("ctrl.u64ExitInfo1 %#RX64\n", pVmcb->ctrl.u64ExitInfo1));
2668	Log4(("ctrl.u64ExitInfo2 %#RX64\n", pVmcb->ctrl.u64ExitInfo2));
2669	Log4(("ctrl.ExitIntInfo.u8Vector %#x\n", pVmcb->ctrl.ExitIntInfo.n.u8Vector));
2670	Log4(("ctrl.ExitIntInfo.u3Type %#x\n", pVmcb->ctrl.ExitIntInfo.n.u3Type));
2671	Log4(("ctrl.ExitIntInfo.u1ErrorCodeValid %#x\n", pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid));
2672	Log4(("ctrl.ExitIntInfo.u19Reserved %#x\n", pVmcb->ctrl.ExitIntInfo.n.u19Reserved));
2673	Log4(("ctrl.ExitIntInfo.u1Valid %#x\n", pVmcb->ctrl.ExitIntInfo.n.u1Valid));
2674	Log4(("ctrl.ExitIntInfo.u32ErrorCode %#x\n", pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode));
2675	Log4(("ctrl.NestedPaging %#RX64\n", pVmcb->ctrl.NestedPaging.u));
2676	Log4(("ctrl.EventInject.u8Vector %#x\n", pVmcb->ctrl.EventInject.n.u8Vector));
2677	Log4(("ctrl.EventInject.u3Type %#x\n", pVmcb->ctrl.EventInject.n.u3Type));
2678	Log4(("ctrl.EventInject.u1ErrorCodeValid %#x\n", pVmcb->ctrl.EventInject.n.u1ErrorCodeValid));
2679	Log4(("ctrl.EventInject.u19Reserved %#x\n", pVmcb->ctrl.EventInject.n.u19Reserved));
2680	Log4(("ctrl.EventInject.u1Valid %#x\n", pVmcb->ctrl.EventInject.n.u1Valid));
2681	Log4(("ctrl.EventInject.u32ErrorCode %#x\n", pVmcb->ctrl.EventInject.n.u32ErrorCode));
2682
2683	Log4(("ctrl.u64NestedPagingCR3 %#RX64\n", pVmcb->ctrl.u64NestedPagingCR3));
2684	Log4(("ctrl.u64LBRVirt %#RX64\n", pVmcb->ctrl.u64LBRVirt));
2685
2686	Log4(("guest.CS.u16Sel %RTsel\n", pVmcb->guest.CS.u16Sel));
2687	Log4(("guest.CS.u16Attr %#x\n", pVmcb->guest.CS.u16Attr));
2688	Log4(("guest.CS.u32Limit %#RX32\n", pVmcb->guest.CS.u32Limit));
2689	Log4(("guest.CS.u64Base %#RX64\n", pVmcb->guest.CS.u64Base));
2690	Log4(("guest.DS.u16Sel %#RTsel\n", pVmcb->guest.DS.u16Sel));
2691	Log4(("guest.DS.u16Attr %#x\n", pVmcb->guest.DS.u16Attr));
2692	Log4(("guest.DS.u32Limit %#RX32\n", pVmcb->guest.DS.u32Limit));
2693	Log4(("guest.DS.u64Base %#RX64\n", pVmcb->guest.DS.u64Base));
2694	Log4(("guest.ES.u16Sel %RTsel\n", pVmcb->guest.ES.u16Sel));
2695	Log4(("guest.ES.u16Attr %#x\n", pVmcb->guest.ES.u16Attr));
2696	Log4(("guest.ES.u32Limit %#RX32\n", pVmcb->guest.ES.u32Limit));
2697	Log4(("guest.ES.u64Base %#RX64\n", pVmcb->guest.ES.u64Base));
2698	Log4(("guest.FS.u16Sel %RTsel\n", pVmcb->guest.FS.u16Sel));
2699	Log4(("guest.FS.u16Attr %#x\n", pVmcb->guest.FS.u16Attr));
2700	Log4(("guest.FS.u32Limit %#RX32\n", pVmcb->guest.FS.u32Limit));
2701	Log4(("guest.FS.u64Base %#RX64\n", pVmcb->guest.FS.u64Base));
2702	Log4(("guest.GS.u16Sel %RTsel\n", pVmcb->guest.GS.u16Sel));
2703	Log4(("guest.GS.u16Attr %#x\n", pVmcb->guest.GS.u16Attr));
2704	Log4(("guest.GS.u32Limit %#RX32\n", pVmcb->guest.GS.u32Limit));
2705	Log4(("guest.GS.u64Base %#RX64\n", pVmcb->guest.GS.u64Base));
2706
2707	Log4(("guest.GDTR.u32Limit %#RX32\n", pVmcb->guest.GDTR.u32Limit));
2708	Log4(("guest.GDTR.u64Base %#RX64\n", pVmcb->guest.GDTR.u64Base));
2709
2710	Log4(("guest.LDTR.u16Sel %RTsel\n", pVmcb->guest.LDTR.u16Sel));
2711	Log4(("guest.LDTR.u16Attr %#x\n", pVmcb->guest.LDTR.u16Attr));
2712	Log4(("guest.LDTR.u32Limit %#RX32\n", pVmcb->guest.LDTR.u32Limit));
2713	Log4(("guest.LDTR.u64Base %#RX64\n", pVmcb->guest.LDTR.u64Base));
2714
2715	Log4(("guest.IDTR.u32Limit %#RX32\n", pVmcb->guest.IDTR.u32Limit));
2716	Log4(("guest.IDTR.u64Base %#RX64\n", pVmcb->guest.IDTR.u64Base));
2717
2718	Log4(("guest.TR.u16Sel %RTsel\n", pVmcb->guest.TR.u16Sel));
2719	Log4(("guest.TR.u16Attr %#x\n", pVmcb->guest.TR.u16Attr));
2720	Log4(("guest.TR.u32Limit %#RX32\n", pVmcb->guest.TR.u32Limit));
2721	Log4(("guest.TR.u64Base %#RX64\n", pVmcb->guest.TR.u64Base));
2722
2723	Log4(("guest.u8CPL %#x\n", pVmcb->guest.u8CPL));
2724	Log4(("guest.u64CR0 %#RX64\n", pVmcb->guest.u64CR0));
2725	Log4(("guest.u64CR2 %#RX64\n", pVmcb->guest.u64CR2));
2726	Log4(("guest.u64CR3 %#RX64\n", pVmcb->guest.u64CR3));
2727	Log4(("guest.u64CR4 %#RX64\n", pVmcb->guest.u64CR4));
2728	Log4(("guest.u64DR6 %#RX64\n", pVmcb->guest.u64DR6));
2729	Log4(("guest.u64DR7 %#RX64\n", pVmcb->guest.u64DR7));
2730
2731	Log4(("guest.u64RIP %#RX64\n", pVmcb->guest.u64RIP));
2732	Log4(("guest.u64RSP %#RX64\n", pVmcb->guest.u64RSP));
2733	Log4(("guest.u64RAX %#RX64\n", pVmcb->guest.u64RAX));
2734	Log4(("guest.u64RFlags %#RX64\n", pVmcb->guest.u64RFlags));
2735
2736	Log4(("guest.u64SysEnterCS %#RX64\n", pVmcb->guest.u64SysEnterCS));
2737	Log4(("guest.u64SysEnterEIP %#RX64\n", pVmcb->guest.u64SysEnterEIP));
2738	Log4(("guest.u64SysEnterESP %#RX64\n", pVmcb->guest.u64SysEnterESP));
2739
2740	Log4(("guest.u64EFER %#RX64\n", pVmcb->guest.u64EFER));
2741	Log4(("guest.u64STAR %#RX64\n", pVmcb->guest.u64STAR));
2742	Log4(("guest.u64LSTAR %#RX64\n", pVmcb->guest.u64LSTAR));
2743	Log4(("guest.u64CSTAR %#RX64\n", pVmcb->guest.u64CSTAR));
2744	Log4(("guest.u64SFMASK %#RX64\n", pVmcb->guest.u64SFMASK));
2745	Log4(("guest.u64KernelGSBase %#RX64\n", pVmcb->guest.u64KernelGSBase));
2746	Log4(("guest.u64GPAT %#RX64\n", pVmcb->guest.u64GPAT));
2747	Log4(("guest.u64DBGCTL %#RX64\n", pVmcb->guest.u64DBGCTL));
2748	Log4(("guest.u64BR_FROM %#RX64\n", pVmcb->guest.u64BR_FROM));
2749	Log4(("guest.u64BR_TO %#RX64\n", pVmcb->guest.u64BR_TO));
2750	Log4(("guest.u64LASTEXCPFROM %#RX64\n", pVmcb->guest.u64LASTEXCPFROM));
2751	Log4(("guest.u64LASTEXCPTO %#RX64\n", pVmcb->guest.u64LASTEXCPTO));
2752	#else
2753	NOREF(pVmcb);
2754	#endif /* VBOX_STRICT */
2755	}
2756	else
2757	Log4(("hmR0SvmReportWorldSwitchError: rcVMRun=%d\n", rcVMRun));
2758	}
2759
2760
2761	/**
2762	* Check per-VM and per-VCPU force flag actions that require us to go back to
2763	* ring-3 for one reason or another.
2764	*
2765	* @returns VBox status code (information status code included).
2766	* @retval VINF_SUCCESS if we don't have any actions that require going back to
2767	* ring-3.
2768	* @retval VINF_PGM_SYNC_CR3 if we have pending PGM CR3 sync.
2769	* @retval VINF_EM_PENDING_REQUEST if we have pending requests (like hardware
2770	* interrupts)
2771	* @retval VINF_PGM_POOL_FLUSH_PENDING if PGM is doing a pool flush and requires
2772	* all EMTs to be in ring-3.
2773	* @retval VINF_EM_RAW_TO_R3 if there is pending DMA requests.
2774	* @retval VINF_EM_NO_MEMORY PGM is out of memory, we need to return
2775	* to the EM loop.
2776	*
2777	* @param pVM Pointer to the VM.
2778	* @param pVCpu Pointer to the VMCPU.
2779	* @param pCtx Pointer to the guest-CPU context.
2780	*/
2781	static int hmR0SvmCheckForceFlags(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
2782	{
2783	Assert(VMMRZCallRing3IsEnabled(pVCpu));
2784
2785	/* On AMD-V we don't need to update CR3, PAE PDPES lazily. See hmR0SvmSaveGuestState(). */
2786	Assert(!VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_UPDATE_CR3));
2787	Assert(!VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_UPDATE_PAE_PDPES));
2788
2789	if ( VM_FF_IS_PENDING(pVM, !pVCpu->hm.s.fSingleInstruction
2790	? VM_FF_HP_R0_PRE_HM_MASK : VM_FF_HP_R0_PRE_HM_STEP_MASK)
2791	\|\| VMCPU_FF_IS_PENDING(pVCpu, !pVCpu->hm.s.fSingleInstruction
2792	? VMCPU_FF_HP_R0_PRE_HM_MASK : VMCPU_FF_HP_R0_PRE_HM_STEP_MASK) )
2793	{
2794	/* Pending PGM C3 sync. */
2795	if (VMCPU_FF_IS_PENDING(pVCpu,VMCPU_FF_PGM_SYNC_CR3 \| VMCPU_FF_PGM_SYNC_CR3_NON_GLOBAL))
2796	{
2797	int rc = PGMSyncCR3(pVCpu, pCtx->cr0, pCtx->cr3, pCtx->cr4, VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_PGM_SYNC_CR3));
2798	if (rc != VINF_SUCCESS)
2799	{
2800	Log4(("hmR0SvmCheckForceFlags: PGMSyncCR3 forcing us back to ring-3. rc=%d\n", rc));
2801	return rc;
2802	}
2803	}
2804
2805	/* Pending HM-to-R3 operations (critsects, timers, EMT rendezvous etc.) */
2806	/* -XXX- what was that about single stepping? */
2807	if ( VM_FF_IS_PENDING(pVM, VM_FF_HM_TO_R3_MASK)
2808	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_TO_R3_MASK))
2809	{
2810	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHmToR3FF);
2811	int rc = RT_UNLIKELY(VM_FF_IS_PENDING(pVM, VM_FF_PGM_NO_MEMORY)) ? VINF_EM_NO_MEMORY : VINF_EM_RAW_TO_R3;
2812	Log4(("hmR0SvmCheckForceFlags: HM_TO_R3 forcing us back to ring-3. rc=%d\n", rc));
2813	return rc;
2814	}
2815
2816	/* Pending VM request packets, such as hardware interrupts. */
2817	if ( VM_FF_IS_PENDING(pVM, VM_FF_REQUEST)
2818	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_REQUEST))
2819	{
2820	Log4(("hmR0SvmCheckForceFlags: Pending VM request forcing us back to ring-3\n"));
2821	return VINF_EM_PENDING_REQUEST;
2822	}
2823
2824	/* Pending PGM pool flushes. */
2825	if (VM_FF_IS_PENDING(pVM, VM_FF_PGM_POOL_FLUSH_PENDING))
2826	{
2827	Log4(("hmR0SvmCheckForceFlags: PGM pool flush pending forcing us back to ring-3\n"));
2828	return VINF_PGM_POOL_FLUSH_PENDING;
2829	}
2830
2831	/* Pending DMA requests. */
2832	if (VM_FF_IS_PENDING(pVM, VM_FF_PDM_DMA))
2833	{
2834	Log4(("hmR0SvmCheckForceFlags: Pending DMA request forcing us back to ring-3\n"));
2835	return VINF_EM_RAW_TO_R3;
2836	}
2837	}
2838
2839	return VINF_SUCCESS;
2840	}
2841
2842
2843	/**
2844	* Does the preparations before executing guest code in AMD-V.
2845	*
2846	* This may cause longjmps to ring-3 and may even result in rescheduling to the
2847	* recompiler. We must be cautious what we do here regarding committing
2848	* guest-state information into the the VMCB assuming we assuredly execute the
2849	* guest in AMD-V. If we fall back to the recompiler after updating the VMCB and
2850	* clearing the common-state (TRPM/forceflags), we must undo those changes so
2851	* that the recompiler can (and should) use them when it resumes guest
2852	* execution. Otherwise such operations must be done when we can no longer
2853	* exit to ring-3.
2854	*
2855	* @returns VBox status code (informational status codes included).
2856	* @retval VINF_SUCCESS if we can proceed with running the guest.
2857	* @retval VINF_* scheduling changes, we have to go back to ring-3.
2858	*
2859	* @param pVM Pointer to the VM.
2860	* @param pVCpu Pointer to the VMCPU.
2861	* @param pCtx Pointer to the guest-CPU context.
2862	* @param pSvmTransient Pointer to the SVM transient structure.
2863	*/
2864	static int hmR0SvmPreRunGuest(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
2865	{
2866	HMSVM_ASSERT_PREEMPT_SAFE();
2867
2868	/* Check force flag actions that might require us to go back to ring-3. */
2869	int rc = hmR0SvmCheckForceFlags(pVM, pVCpu, pCtx);
2870	if (rc != VINF_SUCCESS)
2871	return rc;
2872
2873	if (TRPMHasTrap(pVCpu))
2874	hmR0SvmTrpmTrapToPendingEvent(pVCpu);
2875	else if (!pVCpu->hm.s.Event.fPending)
2876	hmR0SvmEvaluatePendingEvent(pVCpu, pCtx);
2877
2878	#ifdef HMSVM_SYNC_FULL_GUEST_STATE
2879	HMCPU_CF_SET(pVCpu, HM_CHANGED_ALL_GUEST);
2880	#endif
2881
2882	/* Load the guest bits that are not shared with the host in any way since we can longjmp or get preempted. */
2883	rc = hmR0SvmLoadGuestState(pVM, pVCpu, pCtx);
2884	AssertRCReturn(rc, rc);
2885	STAM_COUNTER_INC(&pVCpu->hm.s.StatLoadFull);
2886
2887	/*
2888	* If we're not intercepting TPR changes in the guest, save the guest TPR before the world-switch
2889	* so we can update it on the way back if the guest changed the TPR.
2890	*/
2891	if (pVCpu->hm.s.svm.fSyncVTpr)
2892	{
2893	if (pVM->hm.s.fTPRPatchingActive)
2894	pSvmTransient->u8GuestTpr = pCtx->msrLSTAR;
2895	else
2896	{
2897	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2898	pSvmTransient->u8GuestTpr = pVmcb->ctrl.IntCtrl.n.u8VTPR;
2899	}
2900	}
2901
2902	/*
2903	* No longjmps to ring-3 from this point on!!!
2904	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional, better than a kernel panic.
2905	* This also disables flushing of the R0-logger instance (if any).
2906	*/
2907	VMMRZCallRing3Disable(pVCpu);
2908
2909	/*
2910	* We disable interrupts so that we don't miss any interrupts that would flag preemption (IPI/timers etc.)
2911	* when thread-context hooks aren't used and we've been running with preemption disabled for a while.
2912	*
2913	* We need to check for force-flags that could've possible been altered since we last checked them (e.g.
2914	* by PDMGetInterrupt() leaving the PDM critical section, see @bugref{6398}).
2915	*
2916	* We also check a couple of other force-flags as a last opportunity to get the EMT back to ring-3 before
2917	* executing guest code.
2918	*/
2919	pSvmTransient->uEflags = ASMIntDisableFlags();
2920	if ( VM_FF_IS_PENDING(pVM, VM_FF_EMT_RENDEZVOUS \| VM_FF_TM_VIRTUAL_SYNC)
2921	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_TO_R3_MASK))
2922	{
2923	ASMSetFlags(pSvmTransient->uEflags);
2924	VMMRZCallRing3Enable(pVCpu);
2925	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHmToR3FF);
2926	return VINF_EM_RAW_TO_R3;
2927	}
2928	if (RTThreadPreemptIsPending(NIL_RTTHREAD))
2929	{
2930	ASMSetFlags(pSvmTransient->uEflags);
2931	VMMRZCallRing3Enable(pVCpu);
2932	STAM_COUNTER_INC(&pVCpu->hm.s.StatPendingHostIrq);
2933	return VINF_EM_RAW_INTERRUPT;
2934	}
2935
2936	return VINF_SUCCESS;
2937	}
2938
2939
2940	/**
2941	* Prepares to run guest code in AMD-V and we've committed to doing so. This
2942	* means there is no backing out to ring-3 or anywhere else at this
2943	* point.
2944	*
2945	* @param pVM Pointer to the VM.
2946	* @param pVCpu Pointer to the VMCPU.
2947	* @param pCtx Pointer to the guest-CPU context.
2948	* @param pSvmTransient Pointer to the SVM transient structure.
2949	*
2950	* @remarks Called with preemption disabled.
2951	* @remarks No-long-jump zone!!!
2952	*/
2953	static void hmR0SvmPreRunGuestCommitted(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
2954	{
2955	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2956	Assert(VMMR0IsLogFlushDisabled(pVCpu));
2957	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2958
2959	VMCPU_ASSERT_STATE(pVCpu, VMCPUSTATE_STARTED_HM);
2960	VMCPU_SET_STATE(pVCpu, VMCPUSTATE_STARTED_EXEC); /* Indicate the start of guest execution. */
2961
2962	hmR0SvmInjectPendingEvent(pVCpu, pCtx);
2963
2964	if ( pVCpu->hm.s.fUseGuestFpu
2965	&& !CPUMIsGuestFPUStateActive(pVCpu))
2966	{
2967	CPUMR0LoadGuestFPU(pVM, pVCpu, pCtx);
2968	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR0);
2969	}
2970
2971	/* Load the state shared between host and guest (FPU, debug). */
2972	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2973	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_HOST_GUEST_SHARED_STATE))
2974	hmR0SvmLoadSharedState(pVCpu, pVmcb, pCtx);
2975	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_HOST_CONTEXT); /* Preemption might set this, nothing to do on AMD-V. */
2976	AssertMsg(!HMCPU_CF_VALUE(pVCpu), ("fContextUseFlags=%#RX32\n", HMCPU_CF_VALUE(pVCpu)));
2977
2978	/* Setup TSC offsetting. */
2979	if ( pSvmTransient->fUpdateTscOffsetting
2980	\|\| HMR0GetCurrentCpu()->idCpu != pVCpu->hm.s.idLastCpu)
2981	{
2982	hmR0SvmUpdateTscOffsetting(pVCpu);
2983	pSvmTransient->fUpdateTscOffsetting = false;
2984	}
2985
2986	/* If we've migrating CPUs, mark the VMCB Clean bits as dirty. */
2987	if (HMR0GetCurrentCpu()->idCpu != pVCpu->hm.s.idLastCpu)
2988	pVmcb->ctrl.u64VmcbCleanBits = 0;
2989
2990	/* Store status of the shared guest-host state at the time of VMRUN. */
2991	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS) && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
2992	if (CPUMIsGuestInLongModeEx(pCtx))
2993	{
2994	pSvmTransient->fWasGuestDebugStateActive = CPUMIsGuestDebugStateActivePending(pVCpu);
2995	pSvmTransient->fWasHyperDebugStateActive = CPUMIsHyperDebugStateActivePending(pVCpu);
2996	}
2997	else
2998	#endif
2999	{
3000	pSvmTransient->fWasGuestDebugStateActive = CPUMIsGuestDebugStateActive(pVCpu);
3001	pSvmTransient->fWasHyperDebugStateActive = CPUMIsHyperDebugStateActive(pVCpu);
3002	}
3003	pSvmTransient->fWasGuestFPUStateActive = CPUMIsGuestFPUStateActive(pVCpu);
3004
3005	/* Flush the appropriate tagged-TLB entries. */
3006	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, true); /* Used for TLB-shootdowns, set this across the world switch. */
3007	hmR0SvmFlushTaggedTlb(pVCpu);
3008	Assert(HMR0GetCurrentCpu()->idCpu == pVCpu->hm.s.idLastCpu);
3009
3010	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatEntry, &pVCpu->hm.s.StatInGC, x);
3011
3012	TMNotifyStartOfExecution(pVCpu); /* Finally, notify TM to resume its clocks as we're about
3013	to start executing. */
3014
3015	/*
3016	* Save the current Host TSC_AUX and write the guest TSC_AUX to the host, so that
3017	* RDTSCPs (that don't cause exits) reads the guest MSR. See @bugref{3324}.
3018	*
3019	* This should be done -after- any RDTSCPs for obtaining the host timestamp (TM, STAM etc).
3020	*/
3021	pSvmTransient->fRestoreTscAuxMsr = false;
3022	if ( (pVM->hm.s.cpuid.u32AMDFeatureEDX & X86_CPUID_EXT_FEATURE_EDX_RDTSCP)
3023	&& !(pVmcb->ctrl.u32InterceptCtrl2 & SVM_CTRL2_INTERCEPT_RDTSCP))
3024	{
3025	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_TSC_AUX, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
3026	pVCpu->hm.s.u64HostTscAux = ASMRdMsr(MSR_K8_TSC_AUX);
3027	uint64_t u64GuestTscAux = CPUMR0GetGuestTscAux(pVCpu);
3028	if (u64GuestTscAux != pVCpu->hm.s.u64HostTscAux)
3029	ASMWrMsr(MSR_K8_TSC_AUX, u64GuestTscAux);
3030	pSvmTransient->fRestoreTscAuxMsr = true;
3031	}
3032	else
3033	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_TSC_AUX, SVMMSREXIT_INTERCEPT_READ, SVMMSREXIT_INTERCEPT_WRITE);
3034
3035	/* If VMCB Clean bits isn't supported by the CPU, simply mark all state-bits as dirty, indicating (re)load-from-VMCB. */
3036	if (!(pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_VMCB_CLEAN))
3037	pVmcb->ctrl.u64VmcbCleanBits = 0;
3038	}
3039
3040
3041	/**
3042	* Wrapper for running the guest code in AMD-V.
3043	*
3044	* @returns VBox strict status code.
3045	* @param pVM Pointer to the VM.
3046	* @param pVCpu Pointer to the VMCPU.
3047	* @param pCtx Pointer to the guest-CPU context.
3048	*
3049	* @remarks No-long-jump zone!!!
3050	*/
3051	DECLINLINE(int) hmR0SvmRunGuest(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3052	{
3053	/*
3054	* 64-bit Windows uses XMM registers in the kernel as the Microsoft compiler expresses floating-point operations
3055	* using SSE instructions. Some XMM registers (XMM6-XMM15) are callee-saved and thus the need for this XMM wrapper.
3056	* Refer MSDN docs. "Configuring Programs for 64-bit / x64 Software Conventions / Register Usage" for details.
3057	*/
3058	#ifdef VBOX_WITH_KERNEL_USING_XMM
3059	return HMR0SVMRunWrapXMM(pVCpu->hm.s.svm.HCPhysVmcbHost, pVCpu->hm.s.svm.HCPhysVmcb, pCtx, pVM, pVCpu,
3060	pVCpu->hm.s.svm.pfnVMRun);
3061	#else
3062	return pVCpu->hm.s.svm.pfnVMRun(pVCpu->hm.s.svm.HCPhysVmcbHost, pVCpu->hm.s.svm.HCPhysVmcb, pCtx, pVM, pVCpu);
3063	#endif
3064	}
3065
3066
3067	/**
3068	* Performs some essential restoration of state after running guest code in
3069	* AMD-V.
3070	*
3071	* @param pVM Pointer to the VM.
3072	* @param pVCpu Pointer to the VMCPU.
3073	* @param pMixedCtx Pointer to the guest-CPU context. The data maybe
3074	* out-of-sync. Make sure to update the required fields
3075	* before using them.
3076	* @param pSvmTransient Pointer to the SVM transient structure.
3077	* @param rcVMRun Return code of VMRUN.
3078	*
3079	* @remarks Called with interrupts disabled.
3080	* @remarks No-long-jump zone!!! This function will however re-enable longjmps
3081	* unconditionally when it is safe to do so.
3082	*/
3083	static void hmR0SvmPostRunGuest(PVM pVM, PVMCPU pVCpu, PCPUMCTX pMixedCtx, PSVMTRANSIENT pSvmTransient, int rcVMRun)
3084	{
3085	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
3086
3087	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, false); /* See HMInvalidatePageOnAllVCpus(): used for TLB-shootdowns. */
3088	ASMAtomicIncU32(&pVCpu->hm.s.cWorldSwitchExits); /* Initialized in vmR3CreateUVM(): used for TLB-shootdowns. */
3089
3090	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
3091	pVmcb->ctrl.u64VmcbCleanBits = HMSVM_VMCB_CLEAN_ALL; /* Mark the VMCB-state cache as unmodified by VMM. */
3092
3093	if (pSvmTransient->fRestoreTscAuxMsr)
3094	{
3095	CPUMR0SetGuestTscAux(pVCpu, ASMRdMsr(MSR_K8_TSC_AUX));
3096	ASMWrMsr(MSR_K8_TSC_AUX, pVCpu->hm.s.u64HostTscAux);
3097	}
3098
3099	if (!(pVmcb->ctrl.u32InterceptCtrl1 & SVM_CTRL1_INTERCEPT_RDTSC))
3100	{
3101	/** @todo Find a way to fix hardcoding a guestimate. */
3102	TMCpuTickSetLastSeen(pVCpu, ASMReadTSC() + pVmcb->ctrl.u64TSCOffset - 0x400);
3103	}
3104
3105	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatInGC, &pVCpu->hm.s.StatExit1, x);
3106	TMNotifyEndOfExecution(pVCpu); /* Notify TM that the guest is no longer running. */
3107	VMCPU_SET_STATE(pVCpu, VMCPUSTATE_STARTED_HM);
3108
3109	Assert(!(ASMGetFlags() & X86_EFL_IF));
3110	ASMSetFlags(pSvmTransient->uEflags); /* Enable interrupts. */
3111	VMMRZCallRing3Enable(pVCpu); /* It is now safe to do longjmps to ring-3!!! */
3112
3113	/* If VMRUN failed, we can bail out early. This does -not- cover SVM_EXIT_INVALID. */
3114	if (RT_UNLIKELY(rcVMRun != VINF_SUCCESS))
3115	{
3116	Log4(("VMRUN failure: rcVMRun=%Rrc\n", rcVMRun));
3117	return;
3118	}
3119
3120	pSvmTransient->u64ExitCode = pVmcb->ctrl.u64ExitCode; /* Save the #VMEXIT reason. */
3121	pSvmTransient->fVectoringPF = false; /* Vectoring page-fault needs to be determined later. */
3122	hmR0SvmSaveGuestState(pVCpu, pMixedCtx); /* Save the guest state from the VMCB to the guest-CPU context. */
3123
3124	if (RT_LIKELY(pSvmTransient->u64ExitCode != (uint64_t)SVM_EXIT_INVALID))
3125	{
3126	if (pVCpu->hm.s.svm.fSyncVTpr)
3127	{
3128	/* TPR patching (for 32-bit guests) uses LSTAR MSR for holding the TPR value, otherwise uses the VTPR. */
3129	if ( pVM->hm.s.fTPRPatchingActive
3130	&& (pMixedCtx->msrLSTAR & 0xff) != pSvmTransient->u8GuestTpr)
3131	{
3132	int rc = PDMApicSetTPR(pVCpu, pMixedCtx->msrLSTAR & 0xff);
3133	AssertRC(rc);
3134	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
3135	}
3136	else if (pSvmTransient->u8GuestTpr != pVmcb->ctrl.IntCtrl.n.u8VTPR)
3137	{
3138	int rc = PDMApicSetTPR(pVCpu, pVmcb->ctrl.IntCtrl.n.u8VTPR << 4);
3139	AssertRC(rc);
3140	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
3141	}
3142	}
3143	}
3144	}
3145
3146
3147	/**
3148	* Runs the guest code using AMD-V.
3149	*
3150	* @returns VBox status code.
3151	* @param pVM Pointer to the VM.
3152	* @param pVCpu Pointer to the VMCPU.
3153	*/
3154	static int hmR0SvmRunGuestCodeNormal(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3155	{
3156	SVMTRANSIENT SvmTransient;
3157	SvmTransient.fUpdateTscOffsetting = true;
3158	uint32_t cLoops = 0;
3159	int rc = VERR_INTERNAL_ERROR_5;
3160
3161	for (;; cLoops++)
3162	{
3163	Assert(!HMR0SuspendPending());
3164	HMSVM_ASSERT_CPU_SAFE();
3165
3166	/* Preparatory work for running guest code, this may force us to return
3167	to ring-3. This bugger disables interrupts on VINF_SUCCESS! */
3168	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatEntry, x);
3169	rc = hmR0SvmPreRunGuest(pVM, pVCpu, pCtx, &SvmTransient);
3170	if (rc != VINF_SUCCESS)
3171	break;
3172
3173	/*
3174	* No longjmps to ring-3 from this point on!!!
3175	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional, better than a kernel panic.
3176	* This also disables flushing of the R0-logger instance (if any).
3177	*/
3178	hmR0SvmPreRunGuestCommitted(pVM, pVCpu, pCtx, &SvmTransient);
3179	rc = hmR0SvmRunGuest(pVM, pVCpu, pCtx);
3180
3181	/* Restore any residual host-state and save any bits shared between host
3182	and guest into the guest-CPU state. Re-enables interrupts! */
3183	hmR0SvmPostRunGuest(pVM, pVCpu, pCtx, &SvmTransient, rc);
3184
3185	if (RT_UNLIKELY( rc != VINF_SUCCESS /* Check for VMRUN errors. */
3186	\|\| SvmTransient.u64ExitCode == (uint64_t)SVM_EXIT_INVALID)) /* Check for invalid guest-state errors. */
3187	{
3188	if (rc == VINF_SUCCESS)
3189	rc = VERR_SVM_INVALID_GUEST_STATE;
3190	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit1, x);
3191	hmR0SvmReportWorldSwitchError(pVM, pVCpu, rc, pCtx);
3192	break;
3193	}
3194
3195	/* Handle the #VMEXIT. */
3196	HMSVM_EXITCODE_STAM_COUNTER_INC(SvmTransient.u64ExitCode);
3197	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatExit1, &pVCpu->hm.s.StatExit2, x);
3198	rc = hmR0SvmHandleExit(pVCpu, pCtx, &SvmTransient);
3199	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit2, x);
3200	if (rc != VINF_SUCCESS)
3201	break;
3202	else if (cLoops > pVM->hm.s.cMaxResumeLoops)
3203	{
3204	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitMaxResume);
3205	rc = VINF_EM_RAW_INTERRUPT;
3206	break;
3207	}
3208	}
3209
3210	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatEntry, x);
3211	return rc;
3212	}
3213
3214
3215	/**
3216	* Runs the guest code using AMD-V in single step mode.
3217	*
3218	* @returns VBox status code.
3219	* @param pVM Pointer to the VM.
3220	* @param pVCpu Pointer to the VMCPU.
3221	* @param pCtx Pointer to the guest-CPU context.
3222	*/
3223	static int hmR0SvmRunGuestCodeStep(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3224	{
3225	SVMTRANSIENT SvmTransient;
3226	SvmTransient.fUpdateTscOffsetting = true;
3227	uint32_t cLoops = 0;
3228	int rc = VERR_INTERNAL_ERROR_5;
3229	uint16_t uCsStart = pCtx->cs.Sel;
3230	uint64_t uRipStart = pCtx->rip;
3231
3232	for (;; cLoops++)
3233	{
3234	Assert(!HMR0SuspendPending());
3235	AssertMsg(pVCpu->hm.s.idEnteredCpu == RTMpCpuId(),
3236	("Illegal migration! Entered on CPU %u Current %u cLoops=%u\n", (unsigned)pVCpu->hm.s.idEnteredCpu,
3237	(unsigned)RTMpCpuId(), cLoops));
3238
3239	/* Preparatory work for running guest code, this may force us to return
3240	to ring-3. This bugger disables interrupts on VINF_SUCCESS! */
3241	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatEntry, x);
3242	rc = hmR0SvmPreRunGuest(pVM, pVCpu, pCtx, &SvmTransient);
3243	if (rc != VINF_SUCCESS)
3244	break;
3245
3246	/*
3247	* No longjmps to ring-3 from this point on!!!
3248	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional, better than a kernel panic.
3249	* This also disables flushing of the R0-logger instance (if any).
3250	*/
3251	VMMRZCallRing3Disable(pVCpu);
3252	VMMRZCallRing3RemoveNotification(pVCpu);
3253	hmR0SvmPreRunGuestCommitted(pVM, pVCpu, pCtx, &SvmTransient);
3254
3255	rc = hmR0SvmRunGuest(pVM, pVCpu, pCtx);
3256
3257	/*
3258	* Restore any residual host-state and save any bits shared between host and guest into the guest-CPU state.
3259	* This will also re-enable longjmps to ring-3 when it has reached a safe point!!!
3260	*/
3261	hmR0SvmPostRunGuest(pVM, pVCpu, pCtx, &SvmTransient, rc);
3262	if (RT_UNLIKELY( rc != VINF_SUCCESS /* Check for VMRUN errors. */
3263	\|\| SvmTransient.u64ExitCode == (uint64_t)SVM_EXIT_INVALID)) /* Check for invalid guest-state errors. */
3264	{
3265	if (rc == VINF_SUCCESS)
3266	rc = VERR_SVM_INVALID_GUEST_STATE;
3267	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit1, x);
3268	hmR0SvmReportWorldSwitchError(pVM, pVCpu, rc, pCtx);
3269	return rc;
3270	}
3271
3272	/* Handle the #VMEXIT. */
3273	HMSVM_EXITCODE_STAM_COUNTER_INC(SvmTransient.u64ExitCode);
3274	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatExit1, &pVCpu->hm.s.StatExit2, x);
3275	rc = hmR0SvmHandleExit(pVCpu, pCtx, &SvmTransient);
3276	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit2, x);
3277	if (rc != VINF_SUCCESS)
3278	break;
3279	else if (cLoops > pVM->hm.s.cMaxResumeLoops)
3280	{
3281	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitMaxResume);
3282	rc = VINF_EM_RAW_INTERRUPT;
3283	break;
3284	}
3285
3286	/*
3287	* Did the RIP change, if so, consider it a single step.
3288	* Otherwise, make sure one of the TFs gets set.
3289	*/
3290	if ( pCtx->rip != uRipStart
3291	\|\| pCtx->cs.Sel != uCsStart)
3292	{
3293	rc = VINF_EM_DBG_STEPPED;
3294	break;
3295	}
3296	pVCpu->hm.s.fContextUseFlags \|= HM_CHANGED_GUEST_DEBUG;
3297	}
3298
3299	/*
3300	* Clear the X86_EFL_TF if necessary.
3301	*/
3302	if (pVCpu->hm.s.fClearTrapFlag)
3303	{
3304	pVCpu->hm.s.fClearTrapFlag = false;
3305	pCtx->eflags.Bits.u1TF = 0;
3306	}
3307
3308	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatEntry, x);
3309	return rc;
3310	}
3311
3312
3313	/**
3314	* Runs the guest code using AMD-V.
3315	*
3316	* @returns VBox status code.
3317	* @param pVM Pointer to the VM.
3318	* @param pVCpu Pointer to the VMCPU.
3319	* @param pCtx Pointer to the guest-CPU context.
3320	*/
3321	VMMR0DECL(int) SVMR0RunGuestCode(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3322	{
3323	Assert(VMMRZCallRing3IsEnabled(pVCpu));
3324	HMSVM_ASSERT_PREEMPT_SAFE();
3325	VMMRZCallRing3SetNotification(pVCpu, hmR0SvmCallRing3Callback, pCtx);
3326
3327	int rc;
3328	if (!pVCpu->hm.s.fSingleInstruction && !DBGFIsStepping(pVCpu))
3329	rc = hmR0SvmRunGuestCodeNormal(pVM, pVCpu, pCtx);
3330	else
3331	rc = hmR0SvmRunGuestCodeStep(pVM, pVCpu, pCtx);
3332
3333	if (rc == VERR_EM_INTERPRETER)
3334	rc = VINF_EM_RAW_EMULATE_INSTR;
3335	else if (rc == VINF_EM_RESET)
3336	rc = VINF_EM_TRIPLE_FAULT;
3337
3338	/* Prepare to return to ring-3. This will remove longjmp notifications. */
3339	hmR0SvmExitToRing3(pVM, pVCpu, pCtx, rc);
3340	Assert(!VMMRZCallRing3IsNotificationSet(pVCpu));
3341	return rc;
3342	}
3343
3344
3345	/**
3346	* Handles a #VMEXIT (for all EXITCODE values except SVM_EXIT_INVALID).
3347	*
3348	* @returns VBox status code (informational status codes included).
3349	* @param pVCpu Pointer to the VMCPU.
3350	* @param pCtx Pointer to the guest-CPU context.
3351	* @param pSvmTransient Pointer to the SVM transient structure.
3352	*/
3353	DECLINLINE(int) hmR0SvmHandleExit(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
3354	{
3355	Assert(pSvmTransient->u64ExitCode != (uint64_t)SVM_EXIT_INVALID);
3356	Assert(pSvmTransient->u64ExitCode <= SVM_EXIT_MAX);
3357
3358	/*
3359	* The ordering of the case labels is based on most-frequently-occurring VM-exits for most guests under
3360	* normal workloads (for some definition of "normal").
3361	*/
3362	uint32_t u32ExitCode = pSvmTransient->u64ExitCode;
3363	switch (pSvmTransient->u64ExitCode)
3364	{
3365	case SVM_EXIT_NPF:
3366	return hmR0SvmExitNestedPF(pVCpu, pCtx, pSvmTransient);
3367
3368	case SVM_EXIT_IOIO:
3369	return hmR0SvmExitIOInstr(pVCpu, pCtx, pSvmTransient);
3370
3371	case SVM_EXIT_RDTSC:
3372	return hmR0SvmExitRdtsc(pVCpu, pCtx, pSvmTransient);
3373
3374	case SVM_EXIT_RDTSCP:
3375	return hmR0SvmExitRdtscp(pVCpu, pCtx, pSvmTransient);
3376
3377	case SVM_EXIT_CPUID:
3378	return hmR0SvmExitCpuid(pVCpu, pCtx, pSvmTransient);
3379
3380	case SVM_EXIT_EXCEPTION_E: /* X86_XCPT_PF */
3381	return hmR0SvmExitXcptPF(pVCpu, pCtx, pSvmTransient);
3382
3383	case SVM_EXIT_EXCEPTION_7: /* X86_XCPT_NM */
3384	return hmR0SvmExitXcptNM(pVCpu, pCtx, pSvmTransient);
3385
3386	case SVM_EXIT_EXCEPTION_10: /* X86_XCPT_MF */
3387	return hmR0SvmExitXcptMF(pVCpu, pCtx, pSvmTransient);
3388
3389	case SVM_EXIT_EXCEPTION_1: /* X86_XCPT_DB */
3390	return hmR0SvmExitXcptDB(pVCpu, pCtx, pSvmTransient);
3391
3392	case SVM_EXIT_MONITOR:
3393	return hmR0SvmExitMonitor(pVCpu, pCtx, pSvmTransient);
3394
3395	case SVM_EXIT_MWAIT:
3396	return hmR0SvmExitMwait(pVCpu, pCtx, pSvmTransient);
3397
3398	case SVM_EXIT_HLT:
3399	return hmR0SvmExitHlt(pVCpu, pCtx, pSvmTransient);
3400
3401	case SVM_EXIT_READ_CR0:
3402	case SVM_EXIT_READ_CR3:
3403	case SVM_EXIT_READ_CR4:
3404	return hmR0SvmExitReadCRx(pVCpu, pCtx, pSvmTransient);
3405
3406	case SVM_EXIT_WRITE_CR0:
3407	case SVM_EXIT_WRITE_CR3:
3408	case SVM_EXIT_WRITE_CR4:
3409	case SVM_EXIT_WRITE_CR8:
3410	return hmR0SvmExitWriteCRx(pVCpu, pCtx, pSvmTransient);
3411
3412	case SVM_EXIT_VINTR:
3413	return hmR0SvmExitVIntr(pVCpu, pCtx, pSvmTransient);
3414
3415	case SVM_EXIT_INTR:
3416	case SVM_EXIT_FERR_FREEZE:
3417	case SVM_EXIT_NMI:
3418	return hmR0SvmExitIntr(pVCpu, pCtx, pSvmTransient);
3419
3420	case SVM_EXIT_MSR:
3421	return hmR0SvmExitMsr(pVCpu, pCtx, pSvmTransient);
3422
3423	case SVM_EXIT_INVLPG:
3424	return hmR0SvmExitInvlpg(pVCpu, pCtx, pSvmTransient);
3425
3426	case SVM_EXIT_WBINVD:
3427	return hmR0SvmExitWbinvd(pVCpu, pCtx, pSvmTransient);
3428
3429	case SVM_EXIT_INVD:
3430	return hmR0SvmExitInvd(pVCpu, pCtx, pSvmTransient);
3431
3432	case SVM_EXIT_RDPMC:
3433	return hmR0SvmExitRdpmc(pVCpu, pCtx, pSvmTransient);
3434
3435	default:
3436	{
3437	switch (pSvmTransient->u64ExitCode)
3438	{
3439	case SVM_EXIT_READ_DR0: case SVM_EXIT_READ_DR1: case SVM_EXIT_READ_DR2: case SVM_EXIT_READ_DR3:
3440	case SVM_EXIT_READ_DR6: case SVM_EXIT_READ_DR7: case SVM_EXIT_READ_DR8: case SVM_EXIT_READ_DR9:
3441	case SVM_EXIT_READ_DR10: case SVM_EXIT_READ_DR11: case SVM_EXIT_READ_DR12: case SVM_EXIT_READ_DR13:
3442	case SVM_EXIT_READ_DR14: case SVM_EXIT_READ_DR15:
3443	return hmR0SvmExitReadDRx(pVCpu, pCtx, pSvmTransient);
3444
3445	case SVM_EXIT_WRITE_DR0: case SVM_EXIT_WRITE_DR1: case SVM_EXIT_WRITE_DR2: case SVM_EXIT_WRITE_DR3:
3446	case SVM_EXIT_WRITE_DR6: case SVM_EXIT_WRITE_DR7: case SVM_EXIT_WRITE_DR8: case SVM_EXIT_WRITE_DR9:
3447	case SVM_EXIT_WRITE_DR10: case SVM_EXIT_WRITE_DR11: case SVM_EXIT_WRITE_DR12: case SVM_EXIT_WRITE_DR13:
3448	case SVM_EXIT_WRITE_DR14: case SVM_EXIT_WRITE_DR15:
3449	return hmR0SvmExitWriteDRx(pVCpu, pCtx, pSvmTransient);
3450
3451	case SVM_EXIT_TASK_SWITCH:
3452	return hmR0SvmExitTaskSwitch(pVCpu, pCtx, pSvmTransient);
3453
3454	case SVM_EXIT_VMMCALL:
3455	return hmR0SvmExitVmmCall(pVCpu, pCtx, pSvmTransient);
3456
3457	case SVM_EXIT_SHUTDOWN:
3458	return hmR0SvmExitShutdown(pVCpu, pCtx, pSvmTransient);
3459
3460	case SVM_EXIT_SMI:
3461	case SVM_EXIT_INIT:
3462	{
3463	/*
3464	* We don't intercept NMIs. As for INIT signals, it really shouldn't ever happen here. If it ever does,
3465	* we want to know about it so log the exit code and bail.
3466	*/
3467	AssertMsgFailed(("hmR0SvmHandleExit: Unexpected exit %#RX32\n", (uint32_t)pSvmTransient->u64ExitCode));
3468	pVCpu->hm.s.u32HMError = (uint32_t)pSvmTransient->u64ExitCode;
3469	return VERR_SVM_UNEXPECTED_EXIT;
3470	}
3471
3472	case SVM_EXIT_INVLPGA:
3473	case SVM_EXIT_RSM:
3474	case SVM_EXIT_VMRUN:
3475	case SVM_EXIT_VMLOAD:
3476	case SVM_EXIT_VMSAVE:
3477	case SVM_EXIT_STGI:
3478	case SVM_EXIT_CLGI:
3479	case SVM_EXIT_SKINIT:
3480	return hmR0SvmExitSetPendingXcptUD(pVCpu, pCtx, pSvmTransient);
3481
3482	#ifdef HMSVM_ALWAYS_TRAP_ALL_XCPTS
3483	case SVM_EXIT_EXCEPTION_0: /* X86_XCPT_DE */
3484	/* SVM_EXIT_EXCEPTION_1: / / X86_XCPT_DB - Handled above. */
3485	case SVM_EXIT_EXCEPTION_2: /* X86_XCPT_NMI */
3486	case SVM_EXIT_EXCEPTION_3: /* X86_XCPT_BP */
3487	case SVM_EXIT_EXCEPTION_4: /* X86_XCPT_OF */
3488	case SVM_EXIT_EXCEPTION_5: /* X86_XCPT_BR */
3489	case SVM_EXIT_EXCEPTION_6: /* X86_XCPT_UD */
3490	/* SVM_EXIT_EXCEPTION_7: / / X86_XCPT_NM - Handled above. */
3491	case SVM_EXIT_EXCEPTION_8: /* X86_XCPT_DF */
3492	case SVM_EXIT_EXCEPTION_9: /* X86_XCPT_CO_SEG_OVERRUN */
3493	case SVM_EXIT_EXCEPTION_A: /* X86_XCPT_TS */
3494	case SVM_EXIT_EXCEPTION_B: /* X86_XCPT_NP */
3495	case SVM_EXIT_EXCEPTION_C: /* X86_XCPT_SS */
3496	case SVM_EXIT_EXCEPTION_D: /* X86_XCPT_GP */
3497	/* SVM_EXIT_EXCEPTION_E: / / X86_XCPT_PF - Handled above. */
3498	/* SVM_EXIT_EXCEPTION_10: / / X86_XCPT_MF - Handled above. */
3499	case SVM_EXIT_EXCEPTION_11: /* X86_XCPT_AC */
3500	case SVM_EXIT_EXCEPTION_12: /* X86_XCPT_MC */
3501	case SVM_EXIT_EXCEPTION_13: /* X86_XCPT_XF */
3502	case SVM_EXIT_EXCEPTION_F: /* Reserved */
3503	case SVM_EXIT_EXCEPTION_14: case SVM_EXIT_EXCEPTION_15: case SVM_EXIT_EXCEPTION_16:
3504	case SVM_EXIT_EXCEPTION_17: case SVM_EXIT_EXCEPTION_18: case SVM_EXIT_EXCEPTION_19:
3505	case SVM_EXIT_EXCEPTION_1A: case SVM_EXIT_EXCEPTION_1B: case SVM_EXIT_EXCEPTION_1C:
3506	case SVM_EXIT_EXCEPTION_1D: case SVM_EXIT_EXCEPTION_1E: case SVM_EXIT_EXCEPTION_1F:
3507	{
3508	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
3509	SVMEVENT Event;
3510	Event.u = 0;
3511	Event.n.u1Valid = 1;
3512	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3513	Event.n.u8Vector = pSvmTransient->u64ExitCode - SVM_EXIT_EXCEPTION_0;
3514
3515	switch (Event.n.u8Vector)
3516	{
3517	case X86_XCPT_DE:
3518	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestDE);
3519	break;
3520
3521	case X86_XCPT_BP:
3522	/** Saves the wrong EIP on the stack (pointing to the int3) instead of the
3523	* next instruction. */
3524	/** @todo Investigate this later. */
3525	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestBP);
3526	break;
3527
3528	case X86_XCPT_UD:
3529	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestUD);
3530	break;
3531
3532	case X86_XCPT_NP:
3533	Event.n.u1ErrorCodeValid = 1;
3534	Event.n.u32ErrorCode = pVmcb->ctrl.u64ExitInfo1;
3535	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestNP);
3536	break;
3537
3538	case X86_XCPT_SS:
3539	Event.n.u1ErrorCodeValid = 1;
3540	Event.n.u32ErrorCode = pVmcb->ctrl.u64ExitInfo1;
3541	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestSS);
3542	break;
3543
3544	case X86_XCPT_GP:
3545	Event.n.u1ErrorCodeValid = 1;
3546	Event.n.u32ErrorCode = pVmcb->ctrl.u64ExitInfo1;
3547	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestGP);
3548	break;
3549
3550	default:
3551	AssertMsgFailed(("hmR0SvmHandleExit: Unexpected exit caused by exception %#x\n", Event.n.u8Vector));
3552	pVCpu->hm.s.u32HMError = Event.n.u8Vector;
3553	return VERR_SVM_UNEXPECTED_XCPT_EXIT;
3554	}
3555
3556	Log4(("#Xcpt: Vector=%#x at CS:RIP=%04x:%RGv\n", Event.n.u8Vector, pCtx->cs.Sel, (RTGCPTR)pCtx->rip));
3557	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3558	return VINF_SUCCESS;
3559	}
3560	#endif /* HMSVM_ALWAYS_TRAP_ALL_XCPTS */
3561
3562	default:
3563	{
3564	AssertMsgFailed(("hmR0SvmHandleExit: Unknown exit code %#x\n", u32ExitCode));
3565	pVCpu->hm.s.u32HMError = u32ExitCode;
3566	return VERR_SVM_UNKNOWN_EXIT;
3567	}
3568	}
3569	}
3570	}
3571	return VERR_INTERNAL_ERROR_5; /* Should never happen. */
3572	}
3573
3574
3575	#ifdef DEBUG
3576	/* Is there some generic IPRT define for this that are not in Runtime/internal/\* ?? */
3577	# define HMSVM_ASSERT_PREEMPT_CPUID_VAR() \
3578	RTCPUID const idAssertCpu = RTThreadPreemptIsEnabled(NIL_RTTHREAD) ? NIL_RTCPUID : RTMpCpuId()
3579
3580	# define HMSVM_ASSERT_PREEMPT_CPUID() \
3581	do \
3582	{ \
3583	RTCPUID const idAssertCpuNow = RTThreadPreemptIsEnabled(NIL_RTTHREAD) ? NIL_RTCPUID : RTMpCpuId(); \
3584	AssertMsg(idAssertCpu == idAssertCpuNow, ("SVM %#x, %#x\n", idAssertCpu, idAssertCpuNow)); \
3585	} while (0)
3586
3587	# define HMSVM_VALIDATE_EXIT_HANDLER_PARAMS() \
3588	do { \
3589	AssertPtr(pVCpu); \
3590	AssertPtr(pCtx); \
3591	AssertPtr(pSvmTransient); \
3592	Assert(ASMIntAreEnabled()); \
3593	HMSVM_ASSERT_PREEMPT_SAFE(); \
3594	HMSVM_ASSERT_PREEMPT_CPUID_VAR(); \
3595	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)); \
3596	HMSVM_ASSERT_PREEMPT_SAFE(); \
3597	if (VMMR0IsLogFlushDisabled(pVCpu)) \
3598	HMSVM_ASSERT_PREEMPT_CPUID(); \
3599	} while (0)
3600	#else /* Release builds */
3601	# define HMSVM_VALIDATE_EXIT_HANDLER_PARAMS() do { NOREF(pVCpu); NOREF(pCtx); NOREF(pSvmTransient); } while (0)
3602	#endif
3603
3604
3605	/**
3606	* Worker for hmR0SvmInterpretInvlpg().
3607	*
3608	* @return VBox status code.
3609	* @param pVCpu Pointer to the VMCPU.
3610	* @param pCpu Pointer to the disassembler state.
3611	* @param pRegFrame Pointer to the register frame.
3612	*/
3613	static int hmR0SvmInterpretInvlPgEx(PVMCPU pVCpu, PDISCPUSTATE pCpu, PCPUMCTXCORE pRegFrame)
3614	{
3615	DISQPVPARAMVAL Param1;
3616	RTGCPTR GCPtrPage;
3617
3618	int rc = DISQueryParamVal(pRegFrame, pCpu, &pCpu->Param1, &Param1, DISQPVWHICH_SRC);
3619	if (RT_FAILURE(rc))
3620	return VERR_EM_INTERPRETER;
3621
3622	if ( Param1.type == DISQPV_TYPE_IMMEDIATE
3623	\|\| Param1.type == DISQPV_TYPE_ADDRESS)
3624	{
3625	if (!(Param1.flags & (DISQPV_FLAG_32 \| DISQPV_FLAG_64)))
3626	return VERR_EM_INTERPRETER;
3627
3628	GCPtrPage = Param1.val.val64;
3629	VBOXSTRICTRC rc2 = EMInterpretInvlpg(pVCpu->CTX_SUFF(pVM), pVCpu, pRegFrame, GCPtrPage);
3630	rc = VBOXSTRICTRC_VAL(rc2);
3631	}
3632	else
3633	{
3634	Log4(("hmR0SvmInterpretInvlPgEx invalid parameter type %#x\n", Param1.type));
3635	rc = VERR_EM_INTERPRETER;
3636	}
3637
3638	return rc;
3639	}
3640
3641
3642	/**
3643	* Interprets INVLPG.
3644	*
3645	* @returns VBox status code.
3646	* @retval VINF_* Scheduling instructions.
3647	* @retval VERR_EM_INTERPRETER Something we can't cope with.
3648	* @retval VERR_* Fatal errors.
3649	*
3650	* @param pVM Pointer to the VM.
3651	* @param pRegFrame Pointer to the register frame.
3652	*
3653	* @remarks Updates the RIP if the instruction was executed successfully.
3654	*/
3655	static int hmR0SvmInterpretInvlpg(PVM pVM, PVMCPU pVCpu, PCPUMCTXCORE pRegFrame)
3656	{
3657	/* Only allow 32 & 64 bit code. */
3658	if (CPUMGetGuestCodeBits(pVCpu) != 16)
3659	{
3660	PDISSTATE pDis = &pVCpu->hm.s.DisState;
3661	int rc = EMInterpretDisasCurrent(pVM, pVCpu, pDis, NULL /* pcbInstr */);
3662	if ( RT_SUCCESS(rc)
3663	&& pDis->pCurInstr->uOpcode == OP_INVLPG)
3664	{
3665	rc = hmR0SvmInterpretInvlPgEx(pVCpu, pDis, pRegFrame);
3666	if (RT_SUCCESS(rc))
3667	pRegFrame->rip += pDis->cbInstr;
3668	return rc;
3669	}
3670	else
3671	Log4(("hmR0SvmInterpretInvlpg: EMInterpretDisasCurrent returned %Rrc uOpCode=%#x\n", rc, pDis->pCurInstr->uOpcode));
3672	}
3673	return VERR_EM_INTERPRETER;
3674	}
3675
3676
3677	/**
3678	* Sets an invalid-opcode (#UD) exception as pending-for-injection into the VM.
3679	*
3680	* @param pVCpu Pointer to the VMCPU.
3681	*/
3682	DECLINLINE(void) hmR0SvmSetPendingXcptUD(PVMCPU pVCpu)
3683	{
3684	SVMEVENT Event;
3685	Event.u = 0;
3686	Event.n.u1Valid = 1;
3687	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3688	Event.n.u8Vector = X86_XCPT_UD;
3689	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3690	}
3691
3692
3693	/**
3694	* Sets a debug (#DB) exception as pending-for-injection into the VM.
3695	*
3696	* @param pVCpu Pointer to the VMCPU.
3697	*/
3698	DECLINLINE(void) hmR0SvmSetPendingXcptDB(PVMCPU pVCpu)
3699	{
3700	SVMEVENT Event;
3701	Event.u = 0;
3702	Event.n.u1Valid = 1;
3703	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3704	Event.n.u8Vector = X86_XCPT_DB;
3705	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3706	}
3707
3708
3709	/**
3710	* Sets a page fault (#PF) exception as pending-for-injection into the VM.
3711	*
3712	* @param pVCpu Pointer to the VMCPU.
3713	* @param pCtx Pointer to the guest-CPU context.
3714	* @param u32ErrCode The error-code for the page-fault.
3715	* @param uFaultAddress The page fault address (CR2).
3716	*
3717	* @remarks This updates the guest CR2 with @a uFaultAddress!
3718	*/
3719	DECLINLINE(void) hmR0SvmSetPendingXcptPF(PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t u32ErrCode, RTGCUINTPTR uFaultAddress)
3720	{
3721	SVMEVENT Event;
3722	Event.u = 0;
3723	Event.n.u1Valid = 1;
3724	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3725	Event.n.u8Vector = X86_XCPT_PF;
3726	Event.n.u1ErrorCodeValid = 1;
3727	Event.n.u32ErrorCode = u32ErrCode;
3728
3729	/* Update CR2 of the guest. */
3730	if (pCtx->cr2 != uFaultAddress)
3731	{
3732	pCtx->cr2 = uFaultAddress;
3733	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR2);
3734	}
3735
3736	hmR0SvmSetPendingEvent(pVCpu, &Event, uFaultAddress);
3737	}
3738
3739
3740	/**
3741	* Sets a device-not-available (#NM) exception as pending-for-injection into the
3742	* VM.
3743	*
3744	* @param pVCpu Pointer to the VMCPU.
3745	*/
3746	DECLINLINE(void) hmR0SvmSetPendingXcptNM(PVMCPU pVCpu)
3747	{
3748	SVMEVENT Event;
3749	Event.u = 0;
3750	Event.n.u1Valid = 1;
3751	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3752	Event.n.u8Vector = X86_XCPT_NM;
3753	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3754	}
3755
3756
3757	/**
3758	* Sets a math-fault (#MF) exception as pending-for-injection into the VM.
3759	*
3760	* @param pVCpu Pointer to the VMCPU.
3761	*/
3762	DECLINLINE(void) hmR0SvmSetPendingXcptMF(PVMCPU pVCpu)
3763	{
3764	SVMEVENT Event;
3765	Event.u = 0;
3766	Event.n.u1Valid = 1;
3767	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3768	Event.n.u8Vector = X86_XCPT_MF;
3769	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3770	}
3771
3772
3773	/**
3774	* Sets a double fault (#DF) exception as pending-for-injection into the VM.
3775	*
3776	* @param pVCpu Pointer to the VMCPU.
3777	*/
3778	DECLINLINE(void) hmR0SvmSetPendingXcptDF(PVMCPU pVCpu)
3779	{
3780	SVMEVENT Event;
3781	Event.u = 0;
3782	Event.n.u1Valid = 1;
3783	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3784	Event.n.u8Vector = X86_XCPT_DF;
3785	Event.n.u1ErrorCodeValid = 1;
3786	Event.n.u32ErrorCode = 0;
3787	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3788	}
3789
3790
3791	/**
3792	* Emulates a simple MOV TPR (CR8) instruction, used for TPR patching on 32-bit
3793	* guests. This simply looks up the patch record at EIP and does the required.
3794	*
3795	* This VMMCALL is used a fallback mechanism when mov to/from cr8 isn't exactly
3796	* like how we want it to be (e.g. not followed by shr 4 as is usually done for
3797	* TPR). See hmR3ReplaceTprInstr() for the details.
3798	*
3799	* @returns VBox status code.
3800	* @param pVM Pointer to the VM.
3801	* @param pVCpu Pointer to the VMCPU.
3802	* @param pCtx Pointer to the guest-CPU context.
3803	*/
3804	static int hmR0SvmEmulateMovTpr(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3805	{
3806	Log4(("Emulated VMMCall TPR access replacement at RIP=%RGv\n", pCtx->rip));
3807	for (;;)
3808	{
3809	bool fPending;
3810	uint8_t u8Tpr;
3811
3812	PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
3813	if (!pPatch)
3814	break;
3815
3816	switch (pPatch->enmType)
3817	{
3818	case HMTPRINSTR_READ:
3819	{
3820	int rc = PDMApicGetTPR(pVCpu, &u8Tpr, &fPending, NULL /* pu8PendingIrq */);
3821	AssertRC(rc);
3822
3823	rc = DISWriteReg32(CPUMCTX2CORE(pCtx), pPatch->uDstOperand, u8Tpr);
3824	AssertRC(rc);
3825	pCtx->rip += pPatch->cbOp;
3826	break;
3827	}
3828
3829	case HMTPRINSTR_WRITE_REG:
3830	case HMTPRINSTR_WRITE_IMM:
3831	{
3832	if (pPatch->enmType == HMTPRINSTR_WRITE_REG)
3833	{
3834	uint32_t u32Val;
3835	int rc = DISFetchReg32(CPUMCTX2CORE(pCtx), pPatch->uSrcOperand, &u32Val);
3836	AssertRC(rc);
3837	u8Tpr = u32Val;
3838	}
3839	else
3840	u8Tpr = (uint8_t)pPatch->uSrcOperand;
3841
3842	int rc2 = PDMApicSetTPR(pVCpu, u8Tpr);
3843	AssertRC(rc2);
3844	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
3845
3846	pCtx->rip += pPatch->cbOp;
3847	break;
3848	}
3849
3850	default:
3851	AssertMsgFailed(("Unexpected patch type %d\n", pPatch->enmType));
3852	pVCpu->hm.s.u32HMError = pPatch->enmType;
3853	return VERR_SVM_UNEXPECTED_PATCH_TYPE;
3854	}
3855	}
3856
3857	return VINF_SUCCESS;
3858	}
3859
3860
3861	/**
3862	* Determines if an exception is a contributory exception. Contributory
3863	* exceptions are ones which can cause double-faults. Page-fault is
3864	* intentionally not included here as it's a conditional contributory exception.
3865	*
3866	* @returns true if the exception is contributory, false otherwise.
3867	* @param uVector The exception vector.
3868	*/
3869	DECLINLINE(bool) hmR0SvmIsContributoryXcpt(const uint32_t uVector)
3870	{
3871	switch (uVector)
3872	{
3873	case X86_XCPT_GP:
3874	case X86_XCPT_SS:
3875	case X86_XCPT_NP:
3876	case X86_XCPT_TS:
3877	case X86_XCPT_DE:
3878	return true;
3879	default:
3880	break;
3881	}
3882	return false;
3883	}
3884
3885
3886	/**
3887	* Handle a condition that occurred while delivering an event through the guest
3888	* IDT.
3889	*
3890	* @returns VBox status code (informational error codes included).
3891	* @retval VINF_SUCCESS if we should continue handling the VM-exit.
3892	* @retval VINF_HM_DOUBLE_FAULT if a #DF condition was detected and we ought to
3893	* continue execution of the guest which will delivery the #DF.
3894	* @retval VINF_EM_RESET if we detected a triple-fault condition.
3895	*
3896	* @param pVCpu Pointer to the VMCPU.
3897	* @param pCtx Pointer to the guest-CPU context.
3898	* @param pSvmTransient Pointer to the SVM transient structure.
3899	*
3900	* @remarks No-long-jump zone!!!
3901	*/
3902	static int hmR0SvmCheckExitDueToEventDelivery(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
3903	{
3904	int rc = VINF_SUCCESS;
3905	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
3906
3907	/* See AMD spec. 15.7.3 "EXITINFO Pseudo-Code". The EXITINTINFO (if valid) contains the prior exception (IDT vector)
3908	* that was trying to be delivered to the guest which caused a #VMEXIT which was intercepted (Exit vector). */
3909	if (pVmcb->ctrl.ExitIntInfo.n.u1Valid)
3910	{
3911	uint8_t uIdtVector = pVmcb->ctrl.ExitIntInfo.n.u8Vector;
3912
3913	typedef enum
3914	{
3915	SVMREFLECTXCPT_XCPT, /* Reflect the exception to the guest or for further evaluation by VMM. */
3916	SVMREFLECTXCPT_DF, /* Reflect the exception as a double-fault to the guest. */
3917	SVMREFLECTXCPT_TF, /* Indicate a triple faulted state to the VMM. */
3918	SVMREFLECTXCPT_NONE /* Nothing to reflect. */
3919	} SVMREFLECTXCPT;
3920
3921	SVMREFLECTXCPT enmReflect = SVMREFLECTXCPT_NONE;
3922	if (pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_EXCEPTION)
3923	{
3924	if (pSvmTransient->u64ExitCode - SVM_EXIT_EXCEPTION_0 <= SVM_EXIT_EXCEPTION_1F)
3925	{
3926	uint8_t uExitVector = (uint8_t)(pSvmTransient->u64ExitCode - SVM_EXIT_EXCEPTION_0);
3927
3928	#ifdef VBOX_STRICT
3929	if ( hmR0SvmIsContributoryXcpt(uIdtVector)
3930	&& uExitVector == X86_XCPT_PF)
3931	{
3932	Log4(("IDT: Contributory #PF uCR2=%#RX64\n", pVCpu->idCpu, pCtx->cr2));
3933	}
3934	#endif
3935	if ( uExitVector == X86_XCPT_PF
3936	&& uIdtVector == X86_XCPT_PF)
3937	{
3938	pSvmTransient->fVectoringPF = true;
3939	Log4(("IDT: Vectoring #PF uCR2=%#RX64\n", pCtx->cr2));
3940	}
3941	else if ( (pVmcb->ctrl.u32InterceptException & HMSVM_CONTRIBUTORY_XCPT_MASK)
3942	&& hmR0SvmIsContributoryXcpt(uExitVector)
3943	&& ( hmR0SvmIsContributoryXcpt(uIdtVector)
3944	\|\| uIdtVector == X86_XCPT_PF))
3945	{
3946	enmReflect = SVMREFLECTXCPT_DF;
3947	Log4(("IDT: Pending vectoring #DF %#RX64 uIdtVector=%#x uExitVector=%#x\n", pVCpu->hm.s.Event.u64IntInfo,
3948	uIdtVector, uExitVector));
3949	}
3950	else if (uIdtVector == X86_XCPT_DF)
3951	{
3952	enmReflect = SVMREFLECTXCPT_TF;
3953	Log4(("IDT: Pending vectoring triple-fault %#RX64 uIdtVector=%#x uExitVector=%#x\n",
3954	pVCpu->hm.s.Event.u64IntInfo, uIdtVector, uExitVector));
3955	}
3956	else
3957	enmReflect = SVMREFLECTXCPT_XCPT;
3958	}
3959	else
3960	{
3961	/*
3962	* If event delivery caused an #VMEXIT that is not an exception (e.g. #NPF) then reflect the original
3963	* exception to the guest after handling the VM-exit.
3964	*/
3965	enmReflect = SVMREFLECTXCPT_XCPT;
3966	}
3967	}
3968	else if (pVmcb->ctrl.ExitIntInfo.n.u3Type != SVM_EVENT_SOFTWARE_INT)
3969	{
3970	/* Ignore software interrupts (INT n) as they reoccur when restarting the instruction. */
3971	enmReflect = SVMREFLECTXCPT_XCPT;
3972	}
3973
3974	switch (enmReflect)
3975	{
3976	case SVMREFLECTXCPT_XCPT:
3977	{
3978	Assert(pVmcb->ctrl.ExitIntInfo.n.u3Type != SVM_EVENT_SOFTWARE_INT);
3979	hmR0SvmSetPendingEvent(pVCpu, &pVmcb->ctrl.ExitIntInfo, 0 /* GCPtrFaultAddress */);
3980
3981	/* If uExitVector is #PF, CR2 value will be updated from the VMCB if it's a guest #PF. See hmR0SvmExitXcptPF(). */
3982	Log4(("IDT: Pending vectoring event %#RX64 ErrValid=%RTbool Err=%#RX32\n", pVmcb->ctrl.ExitIntInfo.u,
3983	!!pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid, pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode));
3984	break;
3985	}
3986
3987	case SVMREFLECTXCPT_DF:
3988	{
3989	hmR0SvmSetPendingXcptDF(pVCpu);
3990	rc = VINF_HM_DOUBLE_FAULT;
3991	break;
3992	}
3993
3994	case SVMREFLECTXCPT_TF:
3995	{
3996	rc = VINF_EM_RESET;
3997	break;
3998	}
3999
4000	default:
4001	Assert(rc == VINF_SUCCESS);
4002	break;
4003	}
4004	}
4005	Assert(rc == VINF_SUCCESS \|\| rc == VINF_HM_DOUBLE_FAULT \|\| rc == VINF_EM_RESET);
4006	NOREF(pCtx);
4007	return rc;
4008	}
4009
4010
4011	/**
4012	* Advances the guest RIP in the if the NRIP_SAVE feature is supported by the
4013	* CPU, otherwise advances the RIP by @a cb bytes.
4014	*
4015	* @param pVCpu Pointer to the VMCPU.
4016	* @param pCtx Pointer to the guest-CPU context.
4017	* @param cb RIP increment value in bytes.
4018	*
4019	* @remarks Use this function only from #VMEXIT's where the NRIP value is valid
4020	* when NRIP_SAVE is supported by the CPU!
4021	*/
4022	DECLINLINE(void) hmR0SvmUpdateRip(PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t cb)
4023	{
4024	if (pVCpu->CTX_SUFF(pVM)->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_NRIP_SAVE)
4025	{
4026	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4027	pCtx->rip = pVmcb->ctrl.u64NextRIP;
4028	}
4029	else
4030	pCtx->rip += cb;
4031	}
4032
4033
4034	/* -=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= */
4035	/* -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- #VMEXIT handlers -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- */
4036	/* -=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= */
4037
4038	/** @name VM-exit handlers.
4039	* @{
4040	*/
4041
4042	/**
4043	* #VMEXIT handler for external interrupts, NMIs, FPU assertion freeze and INIT
4044	* signals (SVM_EXIT_INTR, SVM_EXIT_NMI, SVM_EXIT_FERR_FREEZE, SVM_EXIT_INIT).
4045	*/
4046	HMSVM_EXIT_DECL hmR0SvmExitIntr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4047	{
4048	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4049
4050	if (pSvmTransient->u64ExitCode == SVM_EXIT_NMI)
4051	STAM_REL_COUNTER_INC(&pVCpu->hm.s.StatExitHostNmiInGC);
4052	else if (pSvmTransient->u64ExitCode == SVM_EXIT_INTR)
4053	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitExtInt);
4054
4055	/*
4056	* AMD-V has no preemption timer and the generic periodic preemption timer has no way to signal -before- the timer
4057	* fires if the current interrupt is our own timer or a some other host interrupt. We also cannot examine what
4058	* interrupt it is until the host actually take the interrupt.
4059	*
4060	* Going back to executing guest code here unconditionally causes random scheduling problems (observed on an
4061	* AMD Phenom 9850 Quad-Core on Windows 64-bit host).
4062	*/
4063	return VINF_EM_RAW_INTERRUPT;
4064	}
4065
4066
4067	/**
4068	* #VMEXIT handler for WBINVD (SVM_EXIT_WBINVD). Conditional #VMEXIT.
4069	*/
4070	HMSVM_EXIT_DECL hmR0SvmExitWbinvd(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4071	{
4072	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4073
4074	hmR0SvmUpdateRip(pVCpu, pCtx, 2);
4075	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitWbinvd);
4076	int rc = VINF_SUCCESS;
4077	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4078	return rc;
4079	}
4080
4081
4082	/**
4083	* #VMEXIT handler for INVD (SVM_EXIT_INVD). Unconditional #VMEXIT.
4084	*/
4085	HMSVM_EXIT_DECL hmR0SvmExitInvd(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4086	{
4087	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4088
4089	hmR0SvmUpdateRip(pVCpu, pCtx, 2);
4090	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitInvd);
4091	int rc = VINF_SUCCESS;
4092	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4093	return rc;
4094	}
4095
4096
4097	/**
4098	* #VMEXIT handler for INVD (SVM_EXIT_CPUID). Conditional #VMEXIT.
4099	*/
4100	HMSVM_EXIT_DECL hmR0SvmExitCpuid(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4101	{
4102	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4103	PVM pVM = pVCpu->CTX_SUFF(pVM);
4104	int rc = EMInterpretCpuId(pVM, pVCpu, CPUMCTX2CORE(pCtx));
4105	if (RT_LIKELY(rc == VINF_SUCCESS))
4106	{
4107	hmR0SvmUpdateRip(pVCpu, pCtx, 2);
4108	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4109	}
4110	else
4111	{
4112	AssertMsgFailed(("hmR0SvmExitCpuid: EMInterpretCpuId failed with %Rrc\n", rc));
4113	rc = VERR_EM_INTERPRETER;
4114	}
4115	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCpuid);
4116	return rc;
4117	}
4118
4119
4120	/**
4121	* #VMEXIT handler for RDTSC (SVM_EXIT_RDTSC). Conditional #VMEXIT.
4122	*/
4123	HMSVM_EXIT_DECL hmR0SvmExitRdtsc(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4124	{
4125	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4126	PVM pVM = pVCpu->CTX_SUFF(pVM);
4127	int rc = EMInterpretRdtsc(pVM, pVCpu, CPUMCTX2CORE(pCtx));
4128	if (RT_LIKELY(rc == VINF_SUCCESS))
4129	{
4130	hmR0SvmUpdateRip(pVCpu, pCtx, 2);
4131	pSvmTransient->fUpdateTscOffsetting = true;
4132
4133	/* Single step check. */
4134	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4135	}
4136	else
4137	{
4138	AssertMsgFailed(("hmR0SvmExitRdtsc: EMInterpretRdtsc failed with %Rrc\n", rc));
4139	rc = VERR_EM_INTERPRETER;
4140	}
4141	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdtsc);
4142	return rc;
4143	}
4144
4145
4146	/**
4147	* #VMEXIT handler for RDTSCP (SVM_EXIT_RDTSCP). Conditional #VMEXIT.
4148	*/
4149	HMSVM_EXIT_DECL hmR0SvmExitRdtscp(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4150	{
4151	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4152	int rc = EMInterpretRdtscp(pVCpu->CTX_SUFF(pVM), pVCpu, pCtx);
4153	if (RT_LIKELY(rc == VINF_SUCCESS))
4154	{
4155	hmR0SvmUpdateRip(pVCpu, pCtx, 3);
4156	pSvmTransient->fUpdateTscOffsetting = true;
4157	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4158	}
4159	else
4160	{
4161	AssertMsgFailed(("hmR0SvmExitRdtsc: EMInterpretRdtscp failed with %Rrc\n", rc));
4162	rc = VERR_EM_INTERPRETER;
4163	}
4164	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdtscp);
4165	return rc;
4166	}
4167
4168
4169	/**
4170	* #VMEXIT handler for RDPMC (SVM_EXIT_RDPMC). Conditional #VMEXIT.
4171	*/
4172	HMSVM_EXIT_DECL hmR0SvmExitRdpmc(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4173	{
4174	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4175	int rc = EMInterpretRdpmc(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
4176	if (RT_LIKELY(rc == VINF_SUCCESS))
4177	{
4178	hmR0SvmUpdateRip(pVCpu, pCtx, 2);
4179	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4180	}
4181	else
4182	{
4183	AssertMsgFailed(("hmR0SvmExitRdpmc: EMInterpretRdpmc failed with %Rrc\n", rc));
4184	rc = VERR_EM_INTERPRETER;
4185	}
4186	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdpmc);
4187	return rc;
4188	}
4189
4190
4191	/**
4192	* #VMEXIT handler for INVLPG (SVM_EXIT_INVLPG). Conditional #VMEXIT.
4193	*/
4194	HMSVM_EXIT_DECL hmR0SvmExitInvlpg(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4195	{
4196	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4197	PVM pVM = pVCpu->CTX_SUFF(pVM);
4198	Assert(!pVM->hm.s.fNestedPaging);
4199
4200	/** @todo Decode Assist. */
4201	int rc = hmR0SvmInterpretInvlpg(pVM, pVCpu, CPUMCTX2CORE(pCtx)); /* Updates RIP if successful. */
4202	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitInvlpg);
4203	Assert(rc == VINF_SUCCESS \|\| rc == VERR_EM_INTERPRETER);
4204	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4205	return rc;
4206	}
4207
4208
4209	/**
4210	* #VMEXIT handler for HLT (SVM_EXIT_HLT). Conditional #VMEXIT.
4211	*/
4212	HMSVM_EXIT_DECL hmR0SvmExitHlt(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4213	{
4214	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4215	hmR0SvmUpdateRip(pVCpu, pCtx, 1);
4216	int rc = EMShouldContinueAfterHalt(pVCpu, pCtx) ? VINF_SUCCESS : VINF_EM_HALT;
4217	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4218	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitHlt);
4219	return rc;
4220	}
4221
4222
4223	/**
4224	* #VMEXIT handler for MONITOR (SVM_EXIT_MONITOR). Conditional #VMEXIT.
4225	*/
4226	HMSVM_EXIT_DECL hmR0SvmExitMonitor(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4227	{
4228	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4229	int rc = EMInterpretMonitor(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
4230	if (RT_LIKELY(rc == VINF_SUCCESS))
4231	{
4232	hmR0SvmUpdateRip(pVCpu, pCtx, 3);
4233	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4234	}
4235	else
4236	{
4237	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMonitor: EMInterpretMonitor failed with %Rrc\n", rc));
4238	rc = VERR_EM_INTERPRETER;
4239	}
4240	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitMonitor);
4241	return rc;
4242	}
4243
4244
4245	/**
4246	* #VMEXIT handler for MWAIT (SVM_EXIT_MWAIT). Conditional #VMEXIT.
4247	*/
4248	HMSVM_EXIT_DECL hmR0SvmExitMwait(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4249	{
4250	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4251	VBOXSTRICTRC rc2 = EMInterpretMWait(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
4252	int rc = VBOXSTRICTRC_VAL(rc2);
4253	if ( rc == VINF_EM_HALT
4254	\|\| rc == VINF_SUCCESS)
4255	{
4256	hmR0SvmUpdateRip(pVCpu, pCtx, 3);
4257
4258	if ( rc == VINF_EM_HALT
4259	&& EMMonitorWaitShouldContinue(pVCpu, pCtx))
4260	{
4261	rc = VINF_SUCCESS;
4262	}
4263	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4264	}
4265	else
4266	{
4267	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMwait: EMInterpretMWait failed with %Rrc\n", rc));
4268	rc = VERR_EM_INTERPRETER;
4269	}
4270	AssertMsg(rc == VINF_SUCCESS \|\| rc == VINF_EM_HALT \|\| rc == VERR_EM_INTERPRETER,
4271	("hmR0SvmExitMwait: EMInterpretMWait failed rc=%Rrc\n", rc));
4272	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitMwait);
4273	return rc;
4274	}
4275
4276
4277	/**
4278	* #VMEXIT handler for shutdown (triple-fault) (SVM_EXIT_SHUTDOWN).
4279	* Conditional #VMEXIT.
4280	*/
4281	HMSVM_EXIT_DECL hmR0SvmExitShutdown(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4282	{
4283	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4284	return VINF_EM_RESET;
4285	}
4286
4287
4288	/**
4289	* #VMEXIT handler for CRx reads (SVM_EXIT_READ_CR*). Conditional #VMEXIT.
4290	*/
4291	HMSVM_EXIT_DECL hmR0SvmExitReadCRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4292	{
4293	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4294
4295	Log4(("hmR0SvmExitReadCRx: CS:RIP=%04x:%#RX64\n", pCtx->cs.Sel, pCtx->rip));
4296
4297	/** @todo Decode Assist. */
4298	VBOXSTRICTRC rc2 = EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */);
4299	int rc = VBOXSTRICTRC_VAL(rc2);
4300	AssertMsg(rc == VINF_SUCCESS \|\| rc == VERR_EM_INTERPRETER \|\| rc == VINF_PGM_CHANGE_MODE \|\| rc == VINF_PGM_SYNC_CR3,
4301	("hmR0SvmExitReadCRx: EMInterpretInstruction failed rc=%Rrc\n", rc));
4302	Assert((pSvmTransient->u64ExitCode - SVM_EXIT_READ_CR0) <= 15);
4303	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCRxRead[pSvmTransient->u64ExitCode - SVM_EXIT_READ_CR0]);
4304	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4305	return rc;
4306	}
4307
4308
4309	/**
4310	* #VMEXIT handler for CRx writes (SVM_EXIT_WRITE_CR*). Conditional #VMEXIT.
4311	*/
4312	HMSVM_EXIT_DECL hmR0SvmExitWriteCRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4313	{
4314	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4315	/** @todo Decode Assist. */
4316	VBOXSTRICTRC rc2 = EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */);
4317	int rc = VBOXSTRICTRC_VAL(rc2);
4318	if (rc == VINF_SUCCESS)
4319	{
4320	/* RIP has been updated by EMInterpretInstruction(). */
4321	Assert((pSvmTransient->u64ExitCode - SVM_EXIT_WRITE_CR0) <= 15);
4322	switch (pSvmTransient->u64ExitCode - SVM_EXIT_WRITE_CR0)
4323	{
4324	case 0: /* CR0. */
4325	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR0);
4326	break;
4327
4328	case 3: /* CR3. */
4329	Assert(!pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
4330	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR3);
4331	break;
4332
4333	case 4: /* CR4. */
4334	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR4);
4335	break;
4336
4337	case 8: /* CR8 (TPR). */
4338	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
4339	break;
4340
4341	default:
4342	AssertMsgFailed(("hmR0SvmExitWriteCRx: Invalid/Unexpected Write-CRx exit. u64ExitCode=%#RX64 %#x CRx=%#RX64\n",
4343	pSvmTransient->u64ExitCode, pSvmTransient->u64ExitCode - SVM_EXIT_WRITE_CR0));
4344	break;
4345	}
4346	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4347	}
4348	else
4349	Assert(rc == VERR_EM_INTERPRETER \|\| rc == VINF_PGM_CHANGE_MODE \|\| rc == VINF_PGM_SYNC_CR3);
4350	return rc;
4351	}
4352
4353
4354	/**
4355	* #VMEXIT handler for instructions that result in a #UD exception delivered to
4356	* the guest.
4357	*/
4358	HMSVM_EXIT_DECL hmR0SvmExitSetPendingXcptUD(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4359	{
4360	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4361	hmR0SvmSetPendingXcptUD(pVCpu);
4362	return VINF_SUCCESS;
4363	}
4364
4365
4366	/**
4367	* #VMEXIT handler for MSR read and writes (SVM_EXIT_MSR). Conditional #VMEXIT.
4368	*/
4369	HMSVM_EXIT_DECL hmR0SvmExitMsr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4370	{
4371	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4372	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4373	PVM pVM = pVCpu->CTX_SUFF(pVM);
4374
4375	int rc;
4376	if (pVmcb->ctrl.u64ExitInfo1 == SVM_EXIT1_MSR_WRITE)
4377	{
4378	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitWrmsr);
4379
4380	/* Handle TPR patching; intercepted LSTAR write. */
4381	if ( pVM->hm.s.fTPRPatchingActive
4382	&& pCtx->ecx == MSR_K8_LSTAR)
4383	{
4384	if ((pCtx->eax & 0xff) != pSvmTransient->u8GuestTpr)
4385	{
4386	/* Our patch code uses LSTAR for TPR caching for 32-bit guests. */
4387	int rc2 = PDMApicSetTPR(pVCpu, pCtx->eax & 0xff);
4388	AssertRC(rc2);
4389	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
4390	}
4391	hmR0SvmUpdateRip(pVCpu, pCtx, 2);
4392	rc = VINF_SUCCESS;
4393	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4394	return rc;
4395	}
4396
4397	if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_NRIP_SAVE)
4398	{
4399	rc = EMInterpretWrmsr(pVM, pVCpu, CPUMCTX2CORE(pCtx));
4400	if (RT_LIKELY(rc == VINF_SUCCESS))
4401	{
4402	pCtx->rip = pVmcb->ctrl.u64NextRIP;
4403	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4404	}
4405	else
4406	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMsr: EMInterpretWrmsr failed rc=%Rrc\n", rc));
4407	}
4408	else
4409	{
4410	rc = VBOXSTRICTRC_TODO(EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */));
4411	if (RT_UNLIKELY(rc != VINF_SUCCESS))
4412	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMsr: WrMsr. EMInterpretInstruction failed rc=%Rrc\n", rc));
4413	/* RIP updated by EMInterpretInstruction(). */
4414	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4415	}
4416
4417	/* If this is an X2APIC WRMSR access, update the APIC state as well. */
4418	if ( pCtx->ecx >= MSR_IA32_X2APIC_START
4419	&& pCtx->ecx <= MSR_IA32_X2APIC_END)
4420	{
4421	/*
4422	* We've already saved the APIC related guest-state (TPR) in hmR0SvmPostRunGuest(). When full APIC register
4423	* virtualization is implemented we'll have to make sure APIC state is saved from the VMCB before
4424	* EMInterpretWrmsr() changes it.
4425	*/
4426	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
4427	}
4428	else if (pCtx->ecx == MSR_K6_EFER)
4429	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_EFER_MSR);
4430	else if (pCtx->ecx == MSR_IA32_TSC)
4431	pSvmTransient->fUpdateTscOffsetting = true;
4432	}
4433	else
4434	{
4435	/* MSR Read access. */
4436	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdmsr);
4437	Assert(pVmcb->ctrl.u64ExitInfo1 == SVM_EXIT1_MSR_READ);
4438
4439	if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_NRIP_SAVE)
4440	{
4441	rc = EMInterpretRdmsr(pVM, pVCpu, CPUMCTX2CORE(pCtx));
4442	if (RT_LIKELY(rc == VINF_SUCCESS))
4443	{
4444	pCtx->rip = pVmcb->ctrl.u64NextRIP;
4445	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4446	}
4447	else
4448	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMsr: EMInterpretRdmsr failed rc=%Rrc\n", rc));
4449	}
4450	else
4451	{
4452	rc = VBOXSTRICTRC_TODO(EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0));
4453	if (RT_UNLIKELY(rc != VINF_SUCCESS))
4454	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMsr: RdMsr. EMInterpretInstruction failed rc=%Rrc\n", rc));
4455	/* RIP updated by EMInterpretInstruction(). */
4456	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4457	}
4458	}
4459
4460	/* RIP has been updated by EMInterpret[Rd\|Wr]msr(). */
4461	return rc;
4462	}
4463
4464
4465	/**
4466	* #VMEXIT handler for DRx read (SVM_EXIT_READ_DRx). Conditional #VMEXIT.
4467	*/
4468	HMSVM_EXIT_DECL hmR0SvmExitReadDRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4469	{
4470	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4471	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitDRxRead);
4472
4473	/* We should -not- get this VM-exit if the guest's debug registers were active. */
4474	if (pSvmTransient->fWasGuestDebugStateActive)
4475	{
4476	AssertMsgFailed(("hmR0SvmHandleExit: Unexpected exit %#RX32\n", (uint32_t)pSvmTransient->u64ExitCode));
4477	pVCpu->hm.s.u32HMError = (uint32_t)pSvmTransient->u64ExitCode;
4478	return VERR_SVM_UNEXPECTED_EXIT;
4479	}
4480
4481	/*
4482	* Lazy DR0-3 loading.
4483	*/
4484	if (!pSvmTransient->fWasHyperDebugStateActive)
4485	{
4486	Assert(!DBGFIsStepping(pVCpu)); Assert(!pVCpu->hm.s.fSingleInstruction);
4487	Log5(("hmR0SvmExitReadDRx: Lazy loading guest debug registers\n"));
4488
4489	/* Don't intercept DRx read and writes. */
4490	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4491	pVmcb->ctrl.u16InterceptRdDRx = 0;
4492	pVmcb->ctrl.u16InterceptWrDRx = 0;
4493	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
4494
4495	/* We're playing with the host CPU state here, make sure we don't preempt or longjmp. */
4496	VMMRZCallRing3Disable(pVCpu);
4497	HM_DISABLE_PREEMPT_IF_NEEDED();
4498
4499	/* Save the host & load the guest debug state, restart execution of the MOV DRx instruction. */
4500	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
4501	Assert(CPUMIsGuestDebugStateActive(pVCpu) \|\| HC_ARCH_BITS == 32);
4502
4503	HM_RESTORE_PREEMPT_IF_NEEDED();
4504	VMMRZCallRing3Enable(pVCpu);
4505
4506	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxContextSwitch);
4507	return VINF_SUCCESS;
4508	}
4509
4510	/*
4511	* Interpret the read/writing of DRx.
4512	*/
4513	/** @todo Decode assist. */
4514	VBOXSTRICTRC rc = EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */);
4515	Log5(("hmR0SvmExitReadDRx: Emulated DRx access: rc=%Rrc\n", VBOXSTRICTRC_VAL(rc)));
4516	if (RT_LIKELY(rc == VINF_SUCCESS))
4517	{
4518	/* Not necessary for read accesses but whatever doesn't hurt for now, will be fixed with decode assist. */
4519	/** @todo CPUM should set this flag! */
4520	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_DEBUG);
4521	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4522	}
4523	else
4524	Assert(rc == VERR_EM_INTERPRETER);
4525	return VBOXSTRICTRC_TODO(rc);
4526	}
4527
4528
4529	/**
4530	* #VMEXIT handler for DRx write (SVM_EXIT_WRITE_DRx). Conditional #VMEXIT.
4531	*/
4532	HMSVM_EXIT_DECL hmR0SvmExitWriteDRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4533	{
4534	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4535	/* For now it's the same since we interpret the instruction anyway. Will change when using of Decode Assist is implemented. */
4536	int rc = hmR0SvmExitReadDRx(pVCpu, pCtx, pSvmTransient);
4537	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitDRxWrite);
4538	STAM_COUNTER_DEC(&pVCpu->hm.s.StatExitDRxRead);
4539	return rc;
4540	}
4541
4542
4543	/**
4544	* #VMEXIT handler for I/O instructions (SVM_EXIT_IOIO). Conditional #VMEXIT.
4545	*/
4546	HMSVM_EXIT_DECL hmR0SvmExitIOInstr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4547	{
4548	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4549
4550	/* I/O operation lookup arrays. */
4551	static uint32_t const s_aIOSize[8] = { 0, 1, 2, 0, 4, 0, 0, 0 }; /* Size of the I/O accesses in bytes. */
4552	static uint32_t const s_aIOOpAnd[8] = { 0, 0xff, 0xffff, 0, 0xffffffff, 0, 0, 0 }; /* AND masks for saving
4553	the result (in AL/AX/EAX). */
4554	Log4(("hmR0SvmExitIOInstr: CS:RIP=%04x:%#RX64\n", pCtx->cs.Sel, pCtx->rip));
4555
4556	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4557	PVM pVM = pVCpu->CTX_SUFF(pVM);
4558
4559	/* Refer AMD spec. 15.10.2 "IN and OUT Behaviour" and Figure 15-2. "EXITINFO1 for IOIO Intercept" for the format. */
4560	SVMIOIOEXIT IoExitInfo;
4561	IoExitInfo.u = (uint32_t)pVmcb->ctrl.u64ExitInfo1;
4562	uint32_t uIOWidth = (IoExitInfo.u >> 4) & 0x7;
4563	uint32_t cbValue = s_aIOSize[uIOWidth];
4564	uint32_t uAndVal = s_aIOOpAnd[uIOWidth];
4565
4566	if (RT_UNLIKELY(!cbValue))
4567	{
4568	AssertMsgFailed(("hmR0SvmExitIOInstr: Invalid IO operation. uIOWidth=%u\n", uIOWidth));
4569	return VERR_EM_INTERPRETER;
4570	}
4571
4572	VBOXSTRICTRC rcStrict;
4573	if (IoExitInfo.n.u1STR)
4574	{
4575	/* INS/OUTS - I/O String instruction. */
4576	PDISCPUSTATE pDis = &pVCpu->hm.s.DisState;
4577
4578	/** @todo Huh? why can't we use the segment prefix information given by AMD-V
4579	* in EXITINFO1? Investigate once this thing is up and running. */
4580
4581	rcStrict = EMInterpretDisasCurrent(pVM, pVCpu, pDis, NULL);
4582	if (rcStrict == VINF_SUCCESS)
4583	{
4584	if (IoExitInfo.n.u1Type == SVM_IOIO_WRITE)
4585	{
4586	rcStrict = IOMInterpretOUTSEx(pVM, pVCpu, CPUMCTX2CORE(pCtx), IoExitInfo.n.u16Port, pDis->fPrefix,
4587	(DISCPUMODE)pDis->uAddrMode, cbValue);
4588	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOStringWrite);
4589	}
4590	else
4591	{
4592	rcStrict = IOMInterpretINSEx(pVM, pVCpu, CPUMCTX2CORE(pCtx), IoExitInfo.n.u16Port, pDis->fPrefix,
4593	(DISCPUMODE)pDis->uAddrMode, cbValue);
4594	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOStringRead);
4595	}
4596	}
4597	else
4598	rcStrict = VINF_EM_RAW_EMULATE_INSTR;
4599	}
4600	else
4601	{
4602	/* IN/OUT - I/O instruction. */
4603	Assert(!IoExitInfo.n.u1REP);
4604
4605	if (IoExitInfo.n.u1Type == SVM_IOIO_WRITE)
4606	{
4607	rcStrict = IOMIOPortWrite(pVM, pVCpu, IoExitInfo.n.u16Port, pCtx->eax & uAndVal, cbValue);
4608	if (rcStrict == VINF_IOM_R3_IOPORT_WRITE)
4609	HMR0SavePendingIOPortWrite(pVCpu, pCtx->rip, pVmcb->ctrl.u64ExitInfo2, IoExitInfo.n.u16Port, uAndVal, cbValue);
4610
4611	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOWrite);
4612	}
4613	else
4614	{
4615	uint32_t u32Val = 0;
4616
4617	rcStrict = IOMIOPortRead(pVM, pVCpu, IoExitInfo.n.u16Port, &u32Val, cbValue);
4618	if (IOM_SUCCESS(rcStrict))
4619	{
4620	/* Save result of I/O IN instr. in AL/AX/EAX. */
4621	pCtx->eax = (pCtx->eax & ~uAndVal) \| (u32Val & uAndVal);
4622	}
4623	else if (rcStrict == VINF_IOM_R3_IOPORT_READ)
4624	HMR0SavePendingIOPortRead(pVCpu, pCtx->rip, pVmcb->ctrl.u64ExitInfo2, IoExitInfo.n.u16Port, uAndVal, cbValue);
4625
4626	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIORead);
4627	}
4628	}
4629
4630	if (IOM_SUCCESS(rcStrict))
4631	{
4632	/* AMD-V saves the RIP of the instruction following the IO instruction in EXITINFO2. */
4633	pCtx->rip = pVmcb->ctrl.u64ExitInfo2;
4634
4635	/*
4636	* If any I/O breakpoints are armed, we need to check if one triggered
4637	* and take appropriate action.
4638	* Note that the I/O breakpoint type is undefined if CR4.DE is 0.
4639	*/
4640	/** @todo Optimize away the DBGFBpIsHwIoArmed call by having DBGF tell the
4641	* execution engines about whether hyper BPs and such are pending. */
4642	uint32_t const uDr7 = pCtx->dr[7];
4643	if (RT_UNLIKELY( ( (uDr7 & X86_DR7_ENABLED_MASK)
4644	&& X86_DR7_ANY_RW_IO(uDr7)
4645	&& (pCtx->cr4 & X86_CR4_DE))
4646	\|\| DBGFBpIsHwIoArmed(pVM)))
4647	{
4648	/* We're playing with the host CPU state here, make sure we don't preempt or longjmp. */
4649	VMMRZCallRing3Disable(pVCpu);
4650	HM_DISABLE_PREEMPT_IF_NEEDED();
4651
4652	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxIoCheck);
4653	CPUMR0DebugStateMaybeSaveGuest(pVCpu, false /fDr6/);
4654
4655	VBOXSTRICTRC rcStrict2 = DBGFBpCheckIo(pVM, pVCpu, pCtx, IoExitInfo.n.u16Port, cbValue);
4656	if (rcStrict2 == VINF_EM_RAW_GUEST_TRAP)
4657	{
4658	/* Raise #DB. */
4659	pVmcb->guest.u64DR6 = pCtx->dr[6];
4660	pVmcb->guest.u64DR7 = pCtx->dr[7];
4661	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
4662	hmR0SvmSetPendingXcptDB(pVCpu);
4663	}
4664	/* rcStrict is VINF_SUCCESS or in [VINF_EM_FIRST..VINF_EM_LAST]. */
4665	else if ( rcStrict2 != VINF_SUCCESS
4666	&& (rcStrict == VINF_SUCCESS \|\| rcStrict2 < rcStrict))
4667	rcStrict = rcStrict2;
4668
4669	HM_RESTORE_PREEMPT_IF_NEEDED();
4670	VMMRZCallRing3Enable(pVCpu);
4671	}
4672
4673	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
4674	}
4675
4676	#ifdef VBOX_STRICT
4677	if (rcStrict == VINF_IOM_R3_IOPORT_READ)
4678	Assert(IoExitInfo.n.u1Type == SVM_IOIO_READ);
4679	else if (rcStrict == VINF_IOM_R3_IOPORT_WRITE)
4680	Assert(IoExitInfo.n.u1Type == SVM_IOIO_WRITE);
4681	else
4682	{
4683	/** @todo r=bird: This is missing a bunch of VINF_EM_FIRST..VINF_EM_LAST
4684	* statuses, that the VMM device and some others may return. See
4685	* IOM_SUCCESS() for guidance. */
4686	AssertMsg( RT_FAILURE(rcStrict)
4687	\|\| rcStrict == VINF_SUCCESS
4688	\|\| rcStrict == VINF_EM_RAW_EMULATE_INSTR
4689	\|\| rcStrict == VINF_EM_DBG_BREAKPOINT
4690	\|\| rcStrict == VINF_EM_RAW_GUEST_TRAP
4691	\|\| rcStrict == VINF_TRPM_XCPT_DISPATCHED, ("%Rrc\n", VBOXSTRICTRC_VAL(rcStrict)));
4692	}
4693	#endif
4694	return VBOXSTRICTRC_TODO(rcStrict);
4695	}
4696
4697
4698	/**
4699	* #VMEXIT handler for Nested Page-faults (SVM_EXIT_NPF). Conditional
4700	* #VMEXIT.
4701	*/
4702	HMSVM_EXIT_DECL hmR0SvmExitNestedPF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4703	{
4704	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4705	PVM pVM = pVCpu->CTX_SUFF(pVM);
4706	Assert(pVM->hm.s.fNestedPaging);
4707
4708	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
4709
4710	/* See AMD spec. 15.25.6 "Nested versus Guest Page Faults, Fault Ordering" for VMCB details for #NPF. */
4711	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4712	uint32_t u32ErrCode = pVmcb->ctrl.u64ExitInfo1;
4713	RTGCPHYS GCPhysFaultAddr = pVmcb->ctrl.u64ExitInfo2;
4714
4715	Log4(("#NPF at CS:RIP=%04x:%#RX64 faultaddr=%RGp errcode=%#x \n", pCtx->cs.Sel, pCtx->rip, GCPhysFaultAddr, u32ErrCode));
4716
4717	#ifdef VBOX_HM_WITH_GUEST_PATCHING
4718	/* TPR patching for 32-bit guests, using the reserved bit in the page tables for MMIO regions. */
4719	if ( pVM->hm.s.fTprPatchingAllowed
4720	&& (GCPhysFaultAddr & PAGE_OFFSET_MASK) == 0x80 /* TPR offset. */
4721	&& ( !(u32ErrCode & X86_TRAP_PF_P) /* Not present */
4722	\|\| (u32ErrCode & (X86_TRAP_PF_P \| X86_TRAP_PF_RSVD)) == (X86_TRAP_PF_P \| X86_TRAP_PF_RSVD)) /* MMIO page. */
4723	&& !CPUMIsGuestInLongModeEx(pCtx)
4724	&& !CPUMGetGuestCPL(pVCpu)
4725	&& pVM->hm.s.cPatches < RT_ELEMENTS(pVM->hm.s.aPatches))
4726	{
4727	RTGCPHYS GCPhysApicBase = pCtx->msrApicBase;
4728	GCPhysApicBase &= PAGE_BASE_GC_MASK;
4729
4730	if (GCPhysFaultAddr == GCPhysApicBase + 0x80)
4731	{
4732	/* Only attempt to patch the instruction once. */
4733	PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
4734	if (!pPatch)
4735	return VINF_EM_HM_PATCH_TPR_INSTR;
4736	}
4737	}
4738	#endif
4739
4740	/*
4741	* Determine the nested paging mode.
4742	*/
4743	PGMMODE enmNestedPagingMode;
4744	#if HC_ARCH_BITS == 32
4745	if (CPUMIsGuestInLongModeEx(pCtx))
4746	enmNestedPagingMode = PGMMODE_AMD64_NX;
4747	else
4748	#endif
4749	enmNestedPagingMode = PGMGetHostMode(pVM);
4750
4751	/*
4752	* MMIO optimization using the reserved (RSVD) bit in the guest page tables for MMIO pages.
4753	*/
4754	int rc;
4755	Assert((u32ErrCode & (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P)) != X86_TRAP_PF_RSVD);
4756	if ((u32ErrCode & (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P)) == (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P))
4757	{
4758	VBOXSTRICTRC rc2 = PGMR0Trap0eHandlerNPMisconfig(pVM, pVCpu, enmNestedPagingMode, CPUMCTX2CORE(pCtx), GCPhysFaultAddr,
4759	u32ErrCode);
4760	rc = VBOXSTRICTRC_VAL(rc2);
4761
4762	/*
4763	* If we succeed, resume guest execution.
4764	* If we fail in interpreting the instruction because we couldn't get the guest physical address
4765	* of the page containing the instruction via the guest's page tables (we would invalidate the guest page
4766	* in the host TLB), resume execution which would cause a guest page fault to let the guest handle this
4767	* weird case. See @bugref{6043}.
4768	*/
4769	if ( rc == VINF_SUCCESS
4770	\|\| rc == VERR_PAGE_TABLE_NOT_PRESENT
4771	\|\| rc == VERR_PAGE_NOT_PRESENT)
4772	{
4773	/* Successfully handled MMIO operation. */
4774	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
4775	rc = VINF_SUCCESS;
4776	}
4777	return rc;
4778	}
4779
4780	TRPMAssertXcptPF(pVCpu, GCPhysFaultAddr, u32ErrCode);
4781	rc = PGMR0Trap0eHandlerNestedPaging(pVM, pVCpu, enmNestedPagingMode, u32ErrCode, CPUMCTX2CORE(pCtx), GCPhysFaultAddr);
4782	TRPMResetTrap(pVCpu);
4783
4784	Log4(("#NPF: PGMR0Trap0eHandlerNestedPaging returned %Rrc CS:RIP=%04x:%#RX64\n", rc, pCtx->cs.Sel, pCtx->rip));
4785
4786	/*
4787	* Same case as PGMR0Trap0eHandlerNPMisconfig(). See comment above, @bugref{6043}.
4788	*/
4789	if ( rc == VINF_SUCCESS
4790	\|\| rc == VERR_PAGE_TABLE_NOT_PRESENT
4791	\|\| rc == VERR_PAGE_NOT_PRESENT)
4792	{
4793	/* We've successfully synced our shadow page tables. */
4794	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPF);
4795	rc = VINF_SUCCESS;
4796	}
4797
4798	return rc;
4799	}
4800
4801
4802	/**
4803	* #VMEXIT handler for virtual interrupt (SVM_EXIT_VINTR). Conditional #VMEXIT.
4804	*/
4805	HMSVM_EXIT_DECL hmR0SvmExitVIntr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4806	{
4807	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4808
4809	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4810	pVmcb->ctrl.IntCtrl.n.u1VIrqValid = 0; /* No virtual interrupts pending, we'll inject the current one before reentry. */
4811	pVmcb->ctrl.IntCtrl.n.u8VIrqVector = 0;
4812
4813	/* Indicate that we no longer need to VM-exit when the guest is ready to receive interrupts, it is now ready. */
4814	pVmcb->ctrl.u32InterceptCtrl1 &= ~SVM_CTRL1_INTERCEPT_VINTR;
4815	pVmcb->ctrl.u64VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS \| HMSVM_VMCB_CLEAN_TPR);
4816
4817	/* Deliver the pending interrupt via hmR0SvmPreRunGuest()->hmR0SvmInjectEventVmcb() and resume guest execution. */
4818	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIntWindow);
4819	return VINF_SUCCESS;
4820	}
4821
4822
4823	/**
4824	* #VMEXIT handler for task switches (SVM_EXIT_TASK_SWITCH). Conditional #VMEXIT.
4825	*/
4826	HMSVM_EXIT_DECL hmR0SvmExitTaskSwitch(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4827	{
4828	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4829
4830	#ifndef HMSVM_ALWAYS_TRAP_TASK_SWITCH
4831	Assert(!pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
4832	#endif
4833
4834	/* Check if this task-switch occurred while delivery an event through the guest IDT. */
4835	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4836	if ( !(pVmcb->ctrl.u64ExitInfo2 & (SVM_EXIT2_TASK_SWITCH_IRET \| SVM_EXIT2_TASK_SWITCH_JMP))
4837	&& pVCpu->hm.s.Event.fPending)
4838	{
4839	/*
4840	* AMD-V does not provide us with the original exception but we have it in u64IntInfo since we
4841	* injected the event during VM-entry. Software interrupts and exceptions will be regenerated
4842	* when the recompiler restarts the instruction.
4843	*/
4844	SVMEVENT Event;
4845	Event.u = pVCpu->hm.s.Event.u64IntInfo;
4846	if ( Event.n.u3Type == SVM_EVENT_EXCEPTION
4847	\|\| Event.n.u3Type == SVM_EVENT_SOFTWARE_INT)
4848	{
4849	pVCpu->hm.s.Event.fPending = false;
4850	}
4851	else
4852	Log4(("hmR0SvmExitTaskSwitch: TS occurred during event delivery. Kept pending u8Vector=%#x\n", Event.n.u8Vector));
4853	}
4854
4855	/** @todo Emulate task switch someday, currently just going back to ring-3 for
4856	* emulation. */
4857	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitTaskSwitch);
4858	return VERR_EM_INTERPRETER;
4859	}
4860
4861
4862	/**
4863	* #VMEXIT handler for VMMCALL (SVM_EXIT_VMMCALL). Conditional #VMEXIT.
4864	*/
4865	HMSVM_EXIT_DECL hmR0SvmExitVmmCall(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4866	{
4867	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4868
4869	int rc = hmR0SvmEmulateMovTpr(pVCpu->CTX_SUFF(pVM), pVCpu, pCtx);
4870	if (RT_LIKELY(rc == VINF_SUCCESS))
4871	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4872	else
4873	hmR0SvmSetPendingXcptUD(pVCpu);
4874	return VINF_SUCCESS;
4875	}
4876
4877
4878	/**
4879	* #VMEXIT handler for page-fault exceptions (SVM_EXIT_EXCEPTION_E). Conditional
4880	* #VMEXIT.
4881	*/
4882	HMSVM_EXIT_DECL hmR0SvmExitXcptPF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4883	{
4884	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4885
4886	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
4887
4888	/* See AMD spec. 15.12.15 "#PF (Page Fault)". */
4889	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4890	uint32_t u32ErrCode = pVmcb->ctrl.u64ExitInfo1;
4891	RTGCUINTPTR uFaultAddress = pVmcb->ctrl.u64ExitInfo2;
4892	PVM pVM = pVCpu->CTX_SUFF(pVM);
4893
4894	#if defined(HMSVM_ALWAYS_TRAP_ALL_XCPTS) \|\| defined(HMSVM_ALWAYS_TRAP_PF)
4895	if (pVM->hm.s.fNestedPaging)
4896	{
4897	pVCpu->hm.s.Event.fPending = false; /* In case it's a contributory or vectoring #PF. */
4898	if (!pSvmTransient->fVectoringPF)
4899	{
4900	/* A genuine guest #PF, reflect it to the guest. */
4901	hmR0SvmSetPendingXcptPF(pVCpu, pCtx, u32ErrCode, uFaultAddress);
4902	Log4(("#PF: Guest page fault at %04X:%RGv FaultAddr=%RGv ErrCode=%#x\n", pCtx->cs.Sel, (RTGCPTR)pCtx->rip,
4903	uFaultAddress, u32ErrCode));
4904	}
4905	else
4906	{
4907	/* A guest page-fault occurred during delivery of a page-fault. Inject #DF. */
4908	hmR0SvmSetPendingXcptDF(pVCpu);
4909	Log4(("Pending #DF due to vectoring #PF. NP\n"));
4910	}
4911	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestPF);
4912	return VINF_SUCCESS;
4913	}
4914	#endif
4915
4916	Assert(!pVM->hm.s.fNestedPaging);
4917
4918	#ifdef VBOX_HM_WITH_GUEST_PATCHING
4919	/* Shortcut for APIC TPR reads and writes; only applicable to 32-bit guests. */
4920	if ( pVM->hm.s.fTprPatchingAllowed
4921	&& (uFaultAddress & 0xfff) == 0x80 /* TPR offset. */
4922	&& !(u32ErrCode & X86_TRAP_PF_P) /* Not present. */
4923	&& !CPUMIsGuestInLongModeEx(pCtx)
4924	&& !CPUMGetGuestCPL(pVCpu)
4925	&& pVM->hm.s.cPatches < RT_ELEMENTS(pVM->hm.s.aPatches))
4926	{
4927	RTGCPHYS GCPhysApicBase;
4928	GCPhysApicBase = pCtx->msrApicBase;
4929	GCPhysApicBase &= PAGE_BASE_GC_MASK;
4930
4931	/* Check if the page at the fault-address is the APIC base. */
4932	RTGCPHYS GCPhysPage;
4933	int rc2 = PGMGstGetPage(pVCpu, (RTGCPTR)uFaultAddress, NULL /* pfFlags */, &GCPhysPage);
4934	if ( rc2 == VINF_SUCCESS
4935	&& GCPhysPage == GCPhysApicBase)
4936	{
4937	/* Only attempt to patch the instruction once. */
4938	PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
4939	if (!pPatch)
4940	return VINF_EM_HM_PATCH_TPR_INSTR;
4941	}
4942	}
4943	#endif
4944
4945	Log4(("#PF: uFaultAddress=%#RX64 CS:RIP=%#04x:%#RX64 u32ErrCode %#RX32 cr3=%#RX64\n", uFaultAddress, pCtx->cs.Sel,
4946	pCtx->rip, u32ErrCode, pCtx->cr3));
4947
4948	TRPMAssertXcptPF(pVCpu, uFaultAddress, u32ErrCode);
4949	int rc = PGMTrap0eHandler(pVCpu, u32ErrCode, CPUMCTX2CORE(pCtx), (RTGCPTR)uFaultAddress);
4950
4951	Log4(("#PF rc=%Rrc\n", rc));
4952
4953	if (rc == VINF_SUCCESS)
4954	{
4955	/* Successfully synced shadow pages tables or emulated an MMIO instruction. */
4956	TRPMResetTrap(pVCpu);
4957	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPF);
4958	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
4959	return rc;
4960	}
4961	else if (rc == VINF_EM_RAW_GUEST_TRAP)
4962	{
4963	pVCpu->hm.s.Event.fPending = false; /* In case it's a contributory or vectoring #PF. */
4964
4965	if (!pSvmTransient->fVectoringPF)
4966	{
4967	/* It's a guest page fault and needs to be reflected to the guest. */
4968	u32ErrCode = TRPMGetErrorCode(pVCpu); /* The error code might have been changed. */
4969	TRPMResetTrap(pVCpu);
4970	hmR0SvmSetPendingXcptPF(pVCpu, pCtx, u32ErrCode, uFaultAddress);
4971	}
4972	else
4973	{
4974	/* A guest page-fault occurred during delivery of a page-fault. Inject #DF. */
4975	TRPMResetTrap(pVCpu);
4976	hmR0SvmSetPendingXcptDF(pVCpu);
4977	Log4(("#PF: Pending #DF due to vectoring #PF\n"));
4978	}
4979
4980	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestPF);
4981	return VINF_SUCCESS;
4982	}
4983
4984	TRPMResetTrap(pVCpu);
4985	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPFEM);
4986	return rc;
4987	}
4988
4989
4990	/**
4991	* #VMEXIT handler for device-not-available exceptions (SVM_EXIT_EXCEPTION_7).
4992	* Conditional #VMEXIT.
4993	*/
4994	HMSVM_EXIT_DECL hmR0SvmExitXcptNM(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4995	{
4996	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4997
4998	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
4999
5000	/* We're playing with the host CPU state here, make sure we don't preempt or longjmp. */
5001	VMMRZCallRing3Disable(pVCpu);
5002	HM_DISABLE_PREEMPT_IF_NEEDED();
5003
5004	int rc;
5005	/* If the guest FPU was active at the time of the #NM exit, then it's a guest fault. */
5006	if (pSvmTransient->fWasGuestFPUStateActive)
5007	{
5008	rc = VINF_EM_RAW_GUEST_TRAP;
5009	Assert(CPUMIsGuestFPUStateActive(pVCpu) \|\| HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR0));
5010	}
5011	else
5012	{
5013	#ifndef HMSVM_ALWAYS_TRAP_ALL_XCPTS
5014	Assert(!pSvmTransient->fWasGuestFPUStateActive);
5015	#endif
5016	rc = CPUMR0Trap07Handler(pVCpu->CTX_SUFF(pVM), pVCpu, pCtx);
5017	Assert(rc == VINF_EM_RAW_GUEST_TRAP \|\| (rc == VINF_SUCCESS && CPUMIsGuestFPUStateActive(pVCpu)));
5018	}
5019
5020	HM_RESTORE_PREEMPT_IF_NEEDED();
5021	VMMRZCallRing3Enable(pVCpu);
5022
5023	if (rc == VINF_SUCCESS)
5024	{
5025	/* Guest FPU state was activated, we'll want to change CR0 FPU intercepts before the next VM-reentry. */
5026	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR0);
5027	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowNM);
5028	pVCpu->hm.s.fUseGuestFpu = true;
5029	}
5030	else
5031	{
5032	/* Forward #NM to the guest. */
5033	Assert(rc == VINF_EM_RAW_GUEST_TRAP);
5034	hmR0SvmSetPendingXcptNM(pVCpu);
5035	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestNM);
5036	}
5037	return VINF_SUCCESS;
5038	}
5039
5040
5041	/**
5042	* #VMEXIT handler for math-fault exceptions (SVM_EXIT_EXCEPTION_10).
5043	* Conditional #VMEXIT.
5044	*/
5045	HMSVM_EXIT_DECL hmR0SvmExitXcptMF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5046	{
5047	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5048
5049	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
5050
5051	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestMF);
5052
5053	if (!(pCtx->cr0 & X86_CR0_NE))
5054	{
5055	PVM pVM = pVCpu->CTX_SUFF(pVM);
5056	PDISSTATE pDis = &pVCpu->hm.s.DisState;
5057	unsigned cbOp;
5058	int rc = EMInterpretDisasCurrent(pVM, pVCpu, pDis, &cbOp);
5059	if (RT_SUCCESS(rc))
5060	{
5061	/* Convert a #MF into a FERR -> IRQ 13. See @bugref{6117}. */
5062	rc = PDMIsaSetIrq(pVCpu->CTX_SUFF(pVM), 13, 1, 0 /* uTagSrc */);
5063	if (RT_SUCCESS(rc))
5064	pCtx->rip += cbOp;
5065	}
5066	else
5067	Log4(("hmR0SvmExitXcptMF: EMInterpretDisasCurrent returned %Rrc uOpCode=%#x\n", rc, pDis->pCurInstr->uOpcode));
5068	return rc;
5069	}
5070
5071	hmR0SvmSetPendingXcptMF(pVCpu);
5072	return VINF_SUCCESS;
5073	}
5074
5075
5076	/**
5077	* #VMEXIT handler for debug exceptions (SVM_EXIT_EXCEPTION_1). Conditional
5078	* #VMEXIT.
5079	*/
5080	HMSVM_EXIT_DECL hmR0SvmExitXcptDB(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5081	{
5082	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5083
5084	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
5085
5086	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestDB);
5087
5088
5089	/* This can be a fault-type #DB (instruction breakpoint) or a trap-type #DB (data breakpoint). However, for both cases
5090	DR6 and DR7 are updated to what the exception handler expects. See AMD spec. 15.12.2 "#DB (Debug)". */
5091	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
5092	PVM pVM = pVCpu->CTX_SUFF(pVM);
5093	int rc = DBGFRZTrap01Handler(pVM, pVCpu, CPUMCTX2CORE(pCtx), pVmcb->guest.u64DR6, pVCpu->hm.s.fSingleInstruction);
5094	if (rc == VINF_EM_RAW_GUEST_TRAP)
5095	{
5096	Log5(("hmR0SvmExitXcptDB: DR6=%#RX64 -> guest trap\n", pVmcb->guest.u64DR6));
5097	if (CPUMIsHyperDebugStateActive(pVCpu))
5098	CPUMSetGuestDR6(pVCpu, CPUMGetGuestDR6(pVCpu) \| pVmcb->guest.u64DR6);
5099
5100	/* Reflect the exception back to the guest. */
5101	hmR0SvmSetPendingXcptDB(pVCpu);
5102	rc = VINF_SUCCESS;
5103	}
5104
5105	/*
5106	* Update DR6.
5107	*/
5108	if (CPUMIsHyperDebugStateActive(pVCpu))
5109	{
5110	Log5(("hmR0SvmExitXcptDB: DR6=%#RX64 -> %Rrc\n", pVmcb->guest.u64DR6, rc));
5111	pVmcb->guest.u64DR6 = X86_DR6_INIT_VAL;
5112	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
5113	}
5114	else
5115	{
5116	AssertMsg(rc == VINF_SUCCESS, ("rc=%Rrc\n", rc));
5117	Assert(!pVCpu->hm.s.fSingleInstruction && !DBGFIsStepping(pVCpu));
5118	}
5119
5120	return rc;
5121	}
5122
5123	/** @} */
5124

注意: 瀏覽 TracBrowser 來幫助您使用儲存庫瀏覽器

source: vbox/trunk/src/VBox/VMM/VMMR0/HMSVMR0.cpp@ 49969

以其他格式下載: