HMSVMR0.cpp@ 54196

Last change on this file since 54196 was 54196, checked in by vboxsync, 10 years ago
VMM/HM: stat adjustments.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 199.3 KB

Line
1	/* $Id: HMSVMR0.cpp 54196 2015-02-13 12:25:55Z vboxsync $ */
2	/** @file
3	* HM SVM (AMD-V) - Host Context Ring-0.
4	*/
5
6	/*
7	* Copyright (C) 2013-2015 Oracle Corporation
8	*
9	* This file is part of VirtualBox Open Source Edition (OSE), as
10	* available from http://www.virtualbox.org. This file is free software;
11	* you can redistribute it and/or modify it under the terms of the GNU
12	* General Public License (GPL) as published by the Free Software
13	* Foundation, in version 2 as it comes in the "COPYING" file of the
14	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16	*/
17
18	/*******************************************************************************
19	* Header Files *
20	*******************************************************************************/
21	#define LOG_GROUP LOG_GROUP_HM
22	#include <iprt/asm-amd64-x86.h>
23	#include <iprt/thread.h>
24
25	#include "HMInternal.h"
26	#include <VBox/vmm/vm.h>
27	#include "HMSVMR0.h"
28	#include <VBox/vmm/pdmapi.h>
29	#include <VBox/vmm/dbgf.h>
30	#include <VBox/vmm/iom.h>
31	#include <VBox/vmm/tm.h>
32	#include <VBox/vmm/gim.h>
33	#include "dtrace/VBoxVMM.h"
34
35	#ifdef DEBUG_ramshankar
36	# define HMSVM_SYNC_FULL_GUEST_STATE
37	# define HMSVM_ALWAYS_TRAP_ALL_XCPTS
38	# define HMSVM_ALWAYS_TRAP_PF
39	# define HMSVM_ALWAYS_TRAP_TASK_SWITCH
40	#endif
41
42
43	/*******************************************************************************
44	* Defined Constants And Macros *
45	*******************************************************************************/
46	#ifdef VBOX_WITH_STATISTICS
47	# define HMSVM_EXITCODE_STAM_COUNTER_INC(u64ExitCode) do { \
48	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitAll); \
49	if ((u64ExitCode) == SVM_EXIT_NPF) \
50	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitReasonNpf); \
51	else \
52	STAM_COUNTER_INC(&pVCpu->hm.s.paStatExitReasonR0[(u64ExitCode) & MASK_EXITREASON_STAT]); \
53	} while (0)
54	#else
55	# define HMSVM_EXITCODE_STAM_COUNTER_INC(u64ExitCode) do { } while (0)
56	#endif
57
58	/** If we decide to use a function table approach this can be useful to
59	* switch to a "static DECLCALLBACK(int)". */
60	#define HMSVM_EXIT_DECL static int
61
62	/** @name Segment attribute conversion between CPU and AMD-V VMCB format.
63	*
64	* The CPU format of the segment attribute is described in X86DESCATTRBITS
65	* which is 16-bits (i.e. includes 4 bits of the segment limit).
66	*
67	* The AMD-V VMCB format the segment attribute is compact 12-bits (strictly
68	* only the attribute bits and nothing else). Upper 4-bits are unused.
69	*
70	* @{ */
71	#define HMSVM_CPU_2_VMCB_SEG_ATTR(a) ( ((a) & 0xff) \| (((a) & 0xf000) >> 4) )
72	#define HMSVM_VMCB_2_CPU_SEG_ATTR(a) ( ((a) & 0xff) \| (((a) & 0x0f00) << 4) )
73	/** @} */
74
75	/** @name Macros for loading, storing segment registers to/from the VMCB.
76	* @{ */
77	#define HMSVM_LOAD_SEG_REG(REG, reg) \
78	do \
79	{ \
80	Assert(pCtx->reg.fFlags & CPUMSELREG_FLAGS_VALID); \
81	Assert(pCtx->reg.ValidSel == pCtx->reg.Sel); \
82	pVmcb->guest.REG.u16Sel = pCtx->reg.Sel; \
83	pVmcb->guest.REG.u32Limit = pCtx->reg.u32Limit; \
84	pVmcb->guest.REG.u64Base = pCtx->reg.u64Base; \
85	pVmcb->guest.REG.u16Attr = HMSVM_CPU_2_VMCB_SEG_ATTR(pCtx->reg.Attr.u); \
86	} while (0)
87
88	#define HMSVM_SAVE_SEG_REG(REG, reg) \
89	do \
90	{ \
91	pMixedCtx->reg.Sel = pVmcb->guest.REG.u16Sel; \
92	pMixedCtx->reg.ValidSel = pVmcb->guest.REG.u16Sel; \
93	pMixedCtx->reg.fFlags = CPUMSELREG_FLAGS_VALID; \
94	pMixedCtx->reg.u32Limit = pVmcb->guest.REG.u32Limit; \
95	pMixedCtx->reg.u64Base = pVmcb->guest.REG.u64Base; \
96	pMixedCtx->reg.Attr.u = HMSVM_VMCB_2_CPU_SEG_ATTR(pVmcb->guest.REG.u16Attr); \
97	} while (0)
98	/** @} */
99
100	/** Macro for checking and returning from the using function for
101	* \#VMEXIT intercepts that maybe caused during delivering of another
102	* event in the guest. */
103	#define HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY() \
104	do \
105	{ \
106	int rc = hmR0SvmCheckExitDueToEventDelivery(pVCpu, pCtx, pSvmTransient); \
107	if (RT_UNLIKELY(rc == VINF_HM_DOUBLE_FAULT)) \
108	return VINF_SUCCESS; \
109	else if (RT_UNLIKELY(rc == VINF_EM_RESET)) \
110	return rc; \
111	} while (0)
112
113	/** Macro for upgrading a @a a_rc to VINF_EM_DBG_STEPPED after emulating an
114	* instruction that exited. */
115	#define HMSVM_CHECK_SINGLE_STEP(a_pVCpu, a_rc) \
116	do { \
117	if ((a_pVCpu)->hm.s.fSingleInstruction && (a_rc) == VINF_SUCCESS) \
118	(a_rc) = VINF_EM_DBG_STEPPED; \
119	} while (0)
120
121	/** Assert that preemption is disabled or covered by thread-context hooks. */
122	#define HMSVM_ASSERT_PREEMPT_SAFE() Assert( VMMR0ThreadCtxHooksAreRegistered(pVCpu) \
123	\|\| !RTThreadPreemptIsEnabled(NIL_RTTHREAD));
124
125	/** Assert that we haven't migrated CPUs when thread-context hooks are not
126	* used. */
127	#define HMSVM_ASSERT_CPU_SAFE() AssertMsg( VMMR0ThreadCtxHooksAreRegistered(pVCpu) \
128	\|\| pVCpu->hm.s.idEnteredCpu == RTMpCpuId(), \
129	("Illegal migration! Entered on CPU %u Current %u\n", \
130	pVCpu->hm.s.idEnteredCpu, RTMpCpuId()));
131
132	/** Exception bitmap mask for all contributory exceptions.
133	*
134	* Page fault is deliberately excluded here as it's conditional as to whether
135	* it's contributory or benign. Page faults are handled separately.
136	*/
137	#define HMSVM_CONTRIBUTORY_XCPT_MASK ( RT_BIT(X86_XCPT_GP) \| RT_BIT(X86_XCPT_NP) \| RT_BIT(X86_XCPT_SS) \| RT_BIT(X86_XCPT_TS) \
138	\| RT_BIT(X86_XCPT_DE))
139
140	/** @name VMCB Clean Bits.
141	*
142	* These flags are used for VMCB-state caching. A set VMCB Clean bit indicates
143	* AMD-V doesn't need to reload the corresponding value(s) from the VMCB in
144	* memory.
145	*
146	* @{ */
147	/** All intercepts vectors, TSC offset, PAUSE filter counter. */
148	#define HMSVM_VMCB_CLEAN_INTERCEPTS RT_BIT(0)
149	/** I/O permission bitmap, MSR permission bitmap. */
150	#define HMSVM_VMCB_CLEAN_IOPM_MSRPM RT_BIT(1)
151	/** ASID. */
152	#define HMSVM_VMCB_CLEAN_ASID RT_BIT(2)
153	/** TRP: V_TPR, V_IRQ, V_INTR_PRIO, V_IGN_TPR, V_INTR_MASKING,
154	V_INTR_VECTOR. */
155	#define HMSVM_VMCB_CLEAN_TPR RT_BIT(3)
156	/** Nested Paging: Nested CR3 (nCR3), PAT. */
157	#define HMSVM_VMCB_CLEAN_NP RT_BIT(4)
158	/** Control registers (CR0, CR3, CR4, EFER). */
159	#define HMSVM_VMCB_CLEAN_CRX_EFER RT_BIT(5)
160	/** Debug registers (DR6, DR7). */
161	#define HMSVM_VMCB_CLEAN_DRX RT_BIT(6)
162	/** GDT, IDT limit and base. */
163	#define HMSVM_VMCB_CLEAN_DT RT_BIT(7)
164	/** Segment register: CS, SS, DS, ES limit and base. */
165	#define HMSVM_VMCB_CLEAN_SEG RT_BIT(8)
166	/** CR2.*/
167	#define HMSVM_VMCB_CLEAN_CR2 RT_BIT(9)
168	/** Last-branch record (DbgCtlMsr, br_from, br_to, lastint_from, lastint_to) */
169	#define HMSVM_VMCB_CLEAN_LBR RT_BIT(10)
170	/** AVIC (AVIC APIC_BAR; AVIC APIC_BACKING_PAGE, AVIC
171	PHYSICAL_TABLE and AVIC LOGICAL_TABLE Pointers). */
172	#define HMSVM_VMCB_CLEAN_AVIC RT_BIT(11)
173	/** Mask of all valid VMCB Clean bits. */
174	#define HMSVM_VMCB_CLEAN_ALL ( HMSVM_VMCB_CLEAN_INTERCEPTS \
175	\| HMSVM_VMCB_CLEAN_IOPM_MSRPM \
176	\| HMSVM_VMCB_CLEAN_ASID \
177	\| HMSVM_VMCB_CLEAN_TPR \
178	\| HMSVM_VMCB_CLEAN_NP \
179	\| HMSVM_VMCB_CLEAN_CRX_EFER \
180	\| HMSVM_VMCB_CLEAN_DRX \
181	\| HMSVM_VMCB_CLEAN_DT \
182	\| HMSVM_VMCB_CLEAN_SEG \
183	\| HMSVM_VMCB_CLEAN_CR2 \
184	\| HMSVM_VMCB_CLEAN_LBR \
185	\| HMSVM_VMCB_CLEAN_AVIC)
186	/** @} */
187
188	/** @name SVM transient.
189	*
190	* A state structure for holding miscellaneous information across AMD-V
191	* VMRUN/#VMEXIT operation, restored after the transition.
192	*
193	* @{ */
194	typedef struct SVMTRANSIENT
195	{
196	/** The host's rflags/eflags. */
197	RTCCUINTREG uEflags;
198	#if HC_ARCH_BITS == 32
199	uint32_t u32Alignment0;
200	#endif
201
202	/** The #VMEXIT exit code (the EXITCODE field in the VMCB). */
203	uint64_t u64ExitCode;
204	/** The guest's TPR value used for TPR shadowing. */
205	uint8_t u8GuestTpr;
206	/** Alignment. */
207	uint8_t abAlignment0[7];
208
209	/** Whether the guest FPU state was active at the time of #VMEXIT. */
210	bool fWasGuestFPUStateActive;
211	/** Whether the guest debug state was active at the time of #VMEXIT. */
212	bool fWasGuestDebugStateActive;
213	/** Whether the hyper debug state was active at the time of #VMEXIT. */
214	bool fWasHyperDebugStateActive;
215	/** Whether the TSC offset mode needs to be updated. */
216	bool fUpdateTscOffsetting;
217	/** Whether the TSC_AUX MSR needs restoring on #VMEXIT. */
218	bool fRestoreTscAuxMsr;
219	/** Whether the #VMEXIT was caused by a page-fault during delivery of a
220	* contributary exception or a page-fault. */
221	bool fVectoringDoublePF;
222	/** Whether the #VMEXIT was caused by a page-fault during delivery of an
223	* external interrupt or NMI. */
224	bool fVectoringPF;
225	} SVMTRANSIENT, *PSVMTRANSIENT;
226	AssertCompileMemberAlignment(SVMTRANSIENT, u64ExitCode, sizeof(uint64_t));
227	AssertCompileMemberAlignment(SVMTRANSIENT, fWasGuestFPUStateActive, sizeof(uint64_t));
228	/** @} */
229
230	/**
231	* MSRPM (MSR permission bitmap) read permissions (for guest RDMSR).
232	*/
233	typedef enum SVMMSREXITREAD
234	{
235	/** Reading this MSR causes a #VMEXIT. */
236	SVMMSREXIT_INTERCEPT_READ = 0xb,
237	/** Reading this MSR does not cause a #VMEXIT. */
238	SVMMSREXIT_PASSTHRU_READ
239	} SVMMSREXITREAD;
240
241	/**
242	* MSRPM (MSR permission bitmap) write permissions (for guest WRMSR).
243	*/
244	typedef enum SVMMSREXITWRITE
245	{
246	/** Writing to this MSR causes a #VMEXIT. */
247	SVMMSREXIT_INTERCEPT_WRITE = 0xd,
248	/** Writing to this MSR does not cause a #VMEXIT. */
249	SVMMSREXIT_PASSTHRU_WRITE
250	} SVMMSREXITWRITE;
251
252	/**
253	* SVM #VMEXIT handler.
254	*
255	* @returns VBox status code.
256	* @param pVCpu Pointer to the VMCPU.
257	* @param pMixedCtx Pointer to the guest-CPU context.
258	* @param pSvmTransient Pointer to the SVM-transient structure.
259	*/
260	typedef int FNSVMEXITHANDLER(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
261
262	/*******************************************************************************
263	* Internal Functions *
264	*******************************************************************************/
265	static void hmR0SvmSetMsrPermission(PVMCPU pVCpu, unsigned uMsr, SVMMSREXITREAD enmRead, SVMMSREXITWRITE enmWrite);
266	static void hmR0SvmPendingEventToTrpmTrap(PVMCPU pVCpu);
267	static void hmR0SvmLeave(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx);
268
269	/** @name #VMEXIT handlers.
270	* @{
271	*/
272	static FNSVMEXITHANDLER hmR0SvmExitIntr;
273	static FNSVMEXITHANDLER hmR0SvmExitWbinvd;
274	static FNSVMEXITHANDLER hmR0SvmExitInvd;
275	static FNSVMEXITHANDLER hmR0SvmExitCpuid;
276	static FNSVMEXITHANDLER hmR0SvmExitRdtsc;
277	static FNSVMEXITHANDLER hmR0SvmExitRdtscp;
278	static FNSVMEXITHANDLER hmR0SvmExitRdpmc;
279	static FNSVMEXITHANDLER hmR0SvmExitInvlpg;
280	static FNSVMEXITHANDLER hmR0SvmExitHlt;
281	static FNSVMEXITHANDLER hmR0SvmExitMonitor;
282	static FNSVMEXITHANDLER hmR0SvmExitMwait;
283	static FNSVMEXITHANDLER hmR0SvmExitShutdown;
284	static FNSVMEXITHANDLER hmR0SvmExitReadCRx;
285	static FNSVMEXITHANDLER hmR0SvmExitWriteCRx;
286	static FNSVMEXITHANDLER hmR0SvmExitSetPendingXcptUD;
287	static FNSVMEXITHANDLER hmR0SvmExitMsr;
288	static FNSVMEXITHANDLER hmR0SvmExitReadDRx;
289	static FNSVMEXITHANDLER hmR0SvmExitWriteDRx;
290	static FNSVMEXITHANDLER hmR0SvmExitIOInstr;
291	static FNSVMEXITHANDLER hmR0SvmExitNestedPF;
292	static FNSVMEXITHANDLER hmR0SvmExitVIntr;
293	static FNSVMEXITHANDLER hmR0SvmExitTaskSwitch;
294	static FNSVMEXITHANDLER hmR0SvmExitVmmCall;
295	static FNSVMEXITHANDLER hmR0SvmExitIret;
296	static FNSVMEXITHANDLER hmR0SvmExitXcptPF;
297	static FNSVMEXITHANDLER hmR0SvmExitXcptNM;
298	static FNSVMEXITHANDLER hmR0SvmExitXcptMF;
299	static FNSVMEXITHANDLER hmR0SvmExitXcptDB;
300	/** @} */
301
302	DECLINLINE(int) hmR0SvmHandleExit(PVMCPU pVCpu, PCPUMCTX pMixedCtx, PSVMTRANSIENT pSvmTransient);
303
304	/*******************************************************************************
305	* Global Variables *
306	*******************************************************************************/
307	/** Ring-0 memory object for the IO bitmap. */
308	RTR0MEMOBJ g_hMemObjIOBitmap = NIL_RTR0MEMOBJ;
309	/** Physical address of the IO bitmap. */
310	RTHCPHYS g_HCPhysIOBitmap = 0;
311	/** Virtual address of the IO bitmap. */
312	R0PTRTYPE(void *) g_pvIOBitmap = NULL;
313
314
315	/**
316	* Sets up and activates AMD-V on the current CPU.
317	*
318	* @returns VBox status code.
319	* @param pCpu Pointer to the CPU info struct.
320	* @param pVM Pointer to the VM (can be NULL after a resume!).
321	* @param pvCpuPage Pointer to the global CPU page.
322	* @param HCPhysCpuPage Physical address of the global CPU page.
323	* @param fEnabledByHost Whether the host OS has already initialized AMD-V.
324	* @param pvArg Unused on AMD-V.
325	*/
326	VMMR0DECL(int) SVMR0EnableCpu(PHMGLOBALCPUINFO pCpu, PVM pVM, void *pvCpuPage, RTHCPHYS HCPhysCpuPage, bool fEnabledByHost,
327	void *pvArg)
328	{
329	Assert(!fEnabledByHost);
330	Assert(HCPhysCpuPage && HCPhysCpuPage != NIL_RTHCPHYS);
331	Assert(RT_ALIGN_T(HCPhysCpuPage, _4K, RTHCPHYS) == HCPhysCpuPage);
332	Assert(pvCpuPage);
333	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
334
335	NOREF(pvArg);
336	NOREF(fEnabledByHost);
337
338	/* Paranoid: Disable interrupt as, in theory, interrupt handlers might mess with EFER. */
339	RTCCUINTREG uEflags = ASMIntDisableFlags();
340
341	/*
342	* We must turn on AMD-V and setup the host state physical address, as those MSRs are per CPU.
343	*/
344	uint64_t u64HostEfer = ASMRdMsr(MSR_K6_EFER);
345	if (u64HostEfer & MSR_K6_EFER_SVME)
346	{
347	/* If the VBOX_HWVIRTEX_IGNORE_SVM_IN_USE is active, then we blindly use AMD-V. */
348	if ( pVM
349	&& pVM->hm.s.svm.fIgnoreInUseError)
350	{
351	pCpu->fIgnoreAMDVInUseError = true;
352	}
353
354	if (!pCpu->fIgnoreAMDVInUseError)
355	{
356	ASMSetFlags(uEflags);
357	return VERR_SVM_IN_USE;
358	}
359	}
360
361	/* Turn on AMD-V in the EFER MSR. */
362	ASMWrMsr(MSR_K6_EFER, u64HostEfer \| MSR_K6_EFER_SVME);
363
364	/* Write the physical page address where the CPU will store the host state while executing the VM. */
365	ASMWrMsr(MSR_K8_VM_HSAVE_PA, HCPhysCpuPage);
366
367	/* Restore interrupts. */
368	ASMSetFlags(uEflags);
369
370	/*
371	* Theoretically, other hypervisors may have used ASIDs, ideally we should flush all non-zero ASIDs
372	* when enabling SVM. AMD doesn't have an SVM instruction to flush all ASIDs (flushing is done
373	* upon VMRUN). Therefore, just set the fFlushAsidBeforeUse flag which instructs hmR0SvmSetupTLB()
374	* to flush the TLB with before using a new ASID.
375	*/
376	pCpu->fFlushAsidBeforeUse = true;
377
378	/*
379	* Ensure each VCPU scheduled on this CPU gets a new VPID on resume. See @bugref{6255}.
380	*/
381	++pCpu->cTlbFlushes;
382
383	return VINF_SUCCESS;
384	}
385
386
387	/**
388	* Deactivates AMD-V on the current CPU.
389	*
390	* @returns VBox status code.
391	* @param pCpu Pointer to the CPU info struct.
392	* @param pvCpuPage Pointer to the global CPU page.
393	* @param HCPhysCpuPage Physical address of the global CPU page.
394	*/
395	VMMR0DECL(int) SVMR0DisableCpu(PHMGLOBALCPUINFO pCpu, void *pvCpuPage, RTHCPHYS HCPhysCpuPage)
396	{
397	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
398	AssertReturn( HCPhysCpuPage
399	&& HCPhysCpuPage != NIL_RTHCPHYS, VERR_INVALID_PARAMETER);
400	AssertReturn(pvCpuPage, VERR_INVALID_PARAMETER);
401	NOREF(pCpu);
402
403	/* Paranoid: Disable interrupts as, in theory, interrupt handlers might mess with EFER. */
404	RTCCUINTREG uEflags = ASMIntDisableFlags();
405
406	/* Turn off AMD-V in the EFER MSR. */
407	uint64_t u64HostEfer = ASMRdMsr(MSR_K6_EFER);
408	ASMWrMsr(MSR_K6_EFER, u64HostEfer & ~MSR_K6_EFER_SVME);
409
410	/* Invalidate host state physical address. */
411	ASMWrMsr(MSR_K8_VM_HSAVE_PA, 0);
412
413	/* Restore interrupts. */
414	ASMSetFlags(uEflags);
415
416	return VINF_SUCCESS;
417	}
418
419
420	/**
421	* Does global AMD-V initialization (called during module initialization).
422	*
423	* @returns VBox status code.
424	*/
425	VMMR0DECL(int) SVMR0GlobalInit(void)
426	{
427	/*
428	* Allocate 12 KB for the IO bitmap. Since this is non-optional and we always intercept all IO accesses, it's done
429	* once globally here instead of per-VM.
430	*/
431	Assert(g_hMemObjIOBitmap == NIL_RTR0MEMOBJ);
432	int rc = RTR0MemObjAllocCont(&g_hMemObjIOBitmap, 3 << PAGE_SHIFT, false /* fExecutable */);
433	if (RT_FAILURE(rc))
434	return rc;
435
436	g_pvIOBitmap = RTR0MemObjAddress(g_hMemObjIOBitmap);
437	g_HCPhysIOBitmap = RTR0MemObjGetPagePhysAddr(g_hMemObjIOBitmap, 0 /* iPage */);
438
439	/* Set all bits to intercept all IO accesses. */
440	ASMMemFill32(g_pvIOBitmap, 3 << PAGE_SHIFT, UINT32_C(0xffffffff));
441	return VINF_SUCCESS;
442	}
443
444
445	/**
446	* Does global AMD-V termination (called during module termination).
447	*/
448	VMMR0DECL(void) SVMR0GlobalTerm(void)
449	{
450	if (g_hMemObjIOBitmap != NIL_RTR0MEMOBJ)
451	{
452	RTR0MemObjFree(g_hMemObjIOBitmap, true /* fFreeMappings */);
453	g_pvIOBitmap = NULL;
454	g_HCPhysIOBitmap = 0;
455	g_hMemObjIOBitmap = NIL_RTR0MEMOBJ;
456	}
457	}
458
459
460	/**
461	* Frees any allocated per-VCPU structures for a VM.
462	*
463	* @param pVM Pointer to the VM.
464	*/
465	DECLINLINE(void) hmR0SvmFreeStructs(PVM pVM)
466	{
467	for (uint32_t i = 0; i < pVM->cCpus; i++)
468	{
469	PVMCPU pVCpu = &pVM->aCpus[i];
470	AssertPtr(pVCpu);
471
472	if (pVCpu->hm.s.svm.hMemObjVmcbHost != NIL_RTR0MEMOBJ)
473	{
474	RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjVmcbHost, false);
475	pVCpu->hm.s.svm.pvVmcbHost = 0;
476	pVCpu->hm.s.svm.HCPhysVmcbHost = 0;
477	pVCpu->hm.s.svm.hMemObjVmcbHost = NIL_RTR0MEMOBJ;
478	}
479
480	if (pVCpu->hm.s.svm.hMemObjVmcb != NIL_RTR0MEMOBJ)
481	{
482	RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjVmcb, false);
483	pVCpu->hm.s.svm.pvVmcb = 0;
484	pVCpu->hm.s.svm.HCPhysVmcb = 0;
485	pVCpu->hm.s.svm.hMemObjVmcb = NIL_RTR0MEMOBJ;
486	}
487
488	if (pVCpu->hm.s.svm.hMemObjMsrBitmap != NIL_RTR0MEMOBJ)
489	{
490	RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjMsrBitmap, false);
491	pVCpu->hm.s.svm.pvMsrBitmap = 0;
492	pVCpu->hm.s.svm.HCPhysMsrBitmap = 0;
493	pVCpu->hm.s.svm.hMemObjMsrBitmap = NIL_RTR0MEMOBJ;
494	}
495	}
496	}
497
498
499	/**
500	* Does per-VM AMD-V initialization.
501	*
502	* @returns VBox status code.
503	* @param pVM Pointer to the VM.
504	*/
505	VMMR0DECL(int) SVMR0InitVM(PVM pVM)
506	{
507	int rc = VERR_INTERNAL_ERROR_5;
508
509	/*
510	* Check for an AMD CPU erratum which requires us to flush the TLB before every world-switch.
511	*/
512	uint32_t u32Family;
513	uint32_t u32Model;
514	uint32_t u32Stepping;
515	if (HMAmdIsSubjectToErratum170(&u32Family, &u32Model, &u32Stepping))
516	{
517	Log4(("SVMR0InitVM: AMD cpu with erratum 170 family %#x model %#x stepping %#x\n", u32Family, u32Model, u32Stepping));
518	pVM->hm.s.svm.fAlwaysFlushTLB = true;
519	}
520
521	/*
522	* Initialize the R0 memory objects up-front so we can properly cleanup on allocation failures.
523	*/
524	for (VMCPUID i = 0; i < pVM->cCpus; i++)
525	{
526	PVMCPU pVCpu = &pVM->aCpus[i];
527	pVCpu->hm.s.svm.hMemObjVmcbHost = NIL_RTR0MEMOBJ;
528	pVCpu->hm.s.svm.hMemObjVmcb = NIL_RTR0MEMOBJ;
529	pVCpu->hm.s.svm.hMemObjMsrBitmap = NIL_RTR0MEMOBJ;
530	}
531
532	for (VMCPUID i = 0; i < pVM->cCpus; i++)
533	{
534	PVMCPU pVCpu = &pVM->aCpus[i];
535
536	/*
537	* Allocate one page for the host-context VM control block (VMCB). This is used for additional host-state (such as
538	* FS, GS, Kernel GS Base, etc.) apart from the host-state save area specified in MSR_K8_VM_HSAVE_PA.
539	*/
540	rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjVmcbHost, 1 << PAGE_SHIFT, false /* fExecutable */);
541	if (RT_FAILURE(rc))
542	goto failure_cleanup;
543
544	pVCpu->hm.s.svm.pvVmcbHost = RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjVmcbHost);
545	pVCpu->hm.s.svm.HCPhysVmcbHost = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjVmcbHost, 0 /* iPage */);
546	Assert(pVCpu->hm.s.svm.HCPhysVmcbHost < _4G);
547	ASMMemZeroPage(pVCpu->hm.s.svm.pvVmcbHost);
548
549	/*
550	* Allocate one page for the guest-state VMCB.
551	*/
552	rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjVmcb, 1 << PAGE_SHIFT, false /* fExecutable */);
553	if (RT_FAILURE(rc))
554	goto failure_cleanup;
555
556	pVCpu->hm.s.svm.pvVmcb = RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjVmcb);
557	pVCpu->hm.s.svm.HCPhysVmcb = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjVmcb, 0 /* iPage */);
558	Assert(pVCpu->hm.s.svm.HCPhysVmcb < _4G);
559	ASMMemZeroPage(pVCpu->hm.s.svm.pvVmcb);
560
561	/*
562	* Allocate two pages (8 KB) for the MSR permission bitmap. There doesn't seem to be a way to convince
563	* SVM to not require one.
564	*/
565	rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjMsrBitmap, 2 << PAGE_SHIFT, false /* fExecutable */);
566	if (RT_FAILURE(rc))
567	goto failure_cleanup;
568
569	pVCpu->hm.s.svm.pvMsrBitmap = RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjMsrBitmap);
570	pVCpu->hm.s.svm.HCPhysMsrBitmap = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjMsrBitmap, 0 /* iPage */);
571	/* Set all bits to intercept all MSR accesses (changed later on). */
572	ASMMemFill32(pVCpu->hm.s.svm.pvMsrBitmap, 2 << PAGE_SHIFT, UINT32_C(0xffffffff));
573	}
574
575	return VINF_SUCCESS;
576
577	failure_cleanup:
578	hmR0SvmFreeStructs(pVM);
579	return rc;
580	}
581
582
583	/**
584	* Does per-VM AMD-V termination.
585	*
586	* @returns VBox status code.
587	* @param pVM Pointer to the VM.
588	*/
589	VMMR0DECL(int) SVMR0TermVM(PVM pVM)
590	{
591	hmR0SvmFreeStructs(pVM);
592	return VINF_SUCCESS;
593	}
594
595
596	/**
597	* Sets the permission bits for the specified MSR in the MSRPM.
598	*
599	* @param pVCpu Pointer to the VMCPU.
600	* @param uMsr The MSR for which the access permissions are being set.
601	* @param enmRead MSR read permissions.
602	* @param enmWrite MSR write permissions.
603	*/
604	static void hmR0SvmSetMsrPermission(PVMCPU pVCpu, unsigned uMsr, SVMMSREXITREAD enmRead, SVMMSREXITWRITE enmWrite)
605	{
606	unsigned ulBit;
607	uint8_t pbMsrBitmap = (uint8_t )pVCpu->hm.s.svm.pvMsrBitmap;
608
609	/*
610	* Layout:
611	* Byte offset MSR range
612	* 0x000 - 0x7ff 0x00000000 - 0x00001fff
613	* 0x800 - 0xfff 0xc0000000 - 0xc0001fff
614	* 0x1000 - 0x17ff 0xc0010000 - 0xc0011fff
615	* 0x1800 - 0x1fff Reserved
616	*/
617	if (uMsr <= 0x00001FFF)
618	{
619	/* Pentium-compatible MSRs. */
620	ulBit = uMsr * 2;
621	}
622	else if ( uMsr >= 0xC0000000
623	&& uMsr <= 0xC0001FFF)
624	{
625	/* AMD Sixth Generation x86 Processor MSRs. */
626	ulBit = (uMsr - 0xC0000000) * 2;
627	pbMsrBitmap += 0x800;
628	}
629	else if ( uMsr >= 0xC0010000
630	&& uMsr <= 0xC0011FFF)
631	{
632	/* AMD Seventh and Eighth Generation Processor MSRs. */
633	ulBit = (uMsr - 0xC0001000) * 2;
634	pbMsrBitmap += 0x1000;
635	}
636	else
637	{
638	AssertFailed();
639	return;
640	}
641
642	Assert(ulBit < 0x3fff /* 16 * 1024 - 1 */);
643	if (enmRead == SVMMSREXIT_INTERCEPT_READ)
644	ASMBitSet(pbMsrBitmap, ulBit);
645	else
646	ASMBitClear(pbMsrBitmap, ulBit);
647
648	if (enmWrite == SVMMSREXIT_INTERCEPT_WRITE)
649	ASMBitSet(pbMsrBitmap, ulBit + 1);
650	else
651	ASMBitClear(pbMsrBitmap, ulBit + 1);
652
653	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
654	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_IOPM_MSRPM;
655	}
656
657
658	/**
659	* Sets up AMD-V for the specified VM.
660	* This function is only called once per-VM during initalization.
661	*
662	* @returns VBox status code.
663	* @param pVM Pointer to the VM.
664	*/
665	VMMR0DECL(int) SVMR0SetupVM(PVM pVM)
666	{
667	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
668	AssertReturn(pVM, VERR_INVALID_PARAMETER);
669	Assert(pVM->hm.s.svm.fSupported);
670
671	for (VMCPUID i = 0; i < pVM->cCpus; i++)
672	{
673	PVMCPU pVCpu = &pVM->aCpus[i];
674	PSVMVMCB pVmcb = (PSVMVMCB)pVM->aCpus[i].hm.s.svm.pvVmcb;
675
676	AssertMsgReturn(pVmcb, ("Invalid pVmcb for vcpu[%u]\n", i), VERR_SVM_INVALID_PVMCB);
677
678	/* Initialize the #VMEXIT history array with end-of-array markers (UINT16_MAX). */
679	Assert(!pVCpu->hm.s.idxExitHistoryFree);
680	HMCPU_EXIT_HISTORY_RESET(pVCpu);
681
682	/* Trap exceptions unconditionally (debug purposes). */
683	#ifdef HMSVM_ALWAYS_TRAP_PF
684	pVmcb->ctrl.u32InterceptException \|= RT_BIT(X86_XCPT_PF);
685	#endif
686	#ifdef HMSVM_ALWAYS_TRAP_ALL_XCPTS
687	/* If you add any exceptions here, make sure to update hmR0SvmHandleExit(). */
688	pVmcb->ctrl.u32InterceptException \|= 0
689	\| RT_BIT(X86_XCPT_BP)
690	\| RT_BIT(X86_XCPT_DB)
691	\| RT_BIT(X86_XCPT_DE)
692	\| RT_BIT(X86_XCPT_NM)
693	\| RT_BIT(X86_XCPT_UD)
694	\| RT_BIT(X86_XCPT_NP)
695	\| RT_BIT(X86_XCPT_SS)
696	\| RT_BIT(X86_XCPT_GP)
697	\| RT_BIT(X86_XCPT_PF)
698	\| RT_BIT(X86_XCPT_MF)
699	;
700	#endif
701
702	/* Set up unconditional intercepts and conditions. */
703	pVmcb->ctrl.u32InterceptCtrl1 = SVM_CTRL1_INTERCEPT_INTR /* External interrupt causes a #VMEXIT. */
704	\| SVM_CTRL1_INTERCEPT_NMI /* Non-maskable interrupts causes a #VMEXIT. */
705	\| SVM_CTRL1_INTERCEPT_INIT /* INIT signal causes a #VMEXIT. */
706	\| SVM_CTRL1_INTERCEPT_RDPMC /* RDPMC causes a #VMEXIT. */
707	\| SVM_CTRL1_INTERCEPT_CPUID /* CPUID causes a #VMEXIT. */
708	\| SVM_CTRL1_INTERCEPT_RSM /* RSM causes a #VMEXIT. */
709	\| SVM_CTRL1_INTERCEPT_HLT /* HLT causes a #VMEXIT. */
710	\| SVM_CTRL1_INTERCEPT_INOUT_BITMAP /* Use the IOPM to cause IOIO #VMEXITs. */
711	\| SVM_CTRL1_INTERCEPT_MSR_SHADOW /* MSR access not covered by MSRPM causes a #VMEXIT.*/
712	\| SVM_CTRL1_INTERCEPT_INVLPGA /* INVLPGA causes a #VMEXIT. */
713	\| SVM_CTRL1_INTERCEPT_SHUTDOWN /* Shutdown events causes a #VMEXIT. */
714	\| SVM_CTRL1_INTERCEPT_FERR_FREEZE; /* Intercept "freezing" during legacy FPU handling. */
715
716	pVmcb->ctrl.u32InterceptCtrl2 = SVM_CTRL2_INTERCEPT_VMRUN /* VMRUN causes a #VMEXIT. */
717	\| SVM_CTRL2_INTERCEPT_VMMCALL /* VMMCALL causes a #VMEXIT. */
718	\| SVM_CTRL2_INTERCEPT_VMLOAD /* VMLOAD causes a #VMEXIT. */
719	\| SVM_CTRL2_INTERCEPT_VMSAVE /* VMSAVE causes a #VMEXIT. */
720	\| SVM_CTRL2_INTERCEPT_STGI /* STGI causes a #VMEXIT. */
721	\| SVM_CTRL2_INTERCEPT_CLGI /* CLGI causes a #VMEXIT. */
722	\| SVM_CTRL2_INTERCEPT_SKINIT /* SKINIT causes a #VMEXIT. */
723	\| SVM_CTRL2_INTERCEPT_WBINVD /* WBINVD causes a #VMEXIT. */
724	\| SVM_CTRL2_INTERCEPT_MONITOR /* MONITOR causes a #VMEXIT. */
725	\| SVM_CTRL2_INTERCEPT_MWAIT; /* MWAIT causes a #VMEXIT. */
726
727	/* CR0, CR4 reads must be intercepted, our shadow values are not necessarily the same as the guest's. */
728	pVmcb->ctrl.u16InterceptRdCRx = RT_BIT(0) \| RT_BIT(4);
729
730	/* CR0, CR4 writes must be intercepted for the same reasons as above. */
731	pVmcb->ctrl.u16InterceptWrCRx = RT_BIT(0) \| RT_BIT(4);
732
733	/* Intercept all DRx reads and writes by default. Changed later on. */
734	pVmcb->ctrl.u16InterceptRdDRx = 0xffff;
735	pVmcb->ctrl.u16InterceptWrDRx = 0xffff;
736
737	/* Virtualize masking of INTR interrupts. (reads/writes from/to CR8 go to the V_TPR register) */
738	pVmcb->ctrl.IntCtrl.n.u1VIrqMasking = 1;
739
740	/* Ignore the priority in the TPR. This is necessary for delivering PIC style (ExtInt) interrupts and we currently
741	deliver both PIC and APIC interrupts alike. See hmR0SvmInjectPendingEvent() */
742	pVmcb->ctrl.IntCtrl.n.u1IgnoreTPR = 1;
743
744	/* Set IO and MSR bitmap permission bitmap physical addresses. */
745	pVmcb->ctrl.u64IOPMPhysAddr = g_HCPhysIOBitmap;
746	pVmcb->ctrl.u64MSRPMPhysAddr = pVCpu->hm.s.svm.HCPhysMsrBitmap;
747
748	/* No LBR virtualization. */
749	pVmcb->ctrl.u64LBRVirt = 0;
750
751	/* Initially set all VMCB clean bits to 0 indicating that everything should be loaded from the VMCB in memory. */
752	pVmcb->ctrl.u64VmcbCleanBits = 0;
753
754	/* The host ASID MBZ, for the guest start with 1. */
755	pVmcb->ctrl.TLBCtrl.n.u32ASID = 1;
756
757	/*
758	* Setup the PAT MSR (applicable for Nested Paging only).
759	* The default value should be 0x0007040600070406ULL, but we want to treat all guest memory as WB,
760	* so choose type 6 for all PAT slots.
761	*/
762	pVmcb->guest.u64GPAT = UINT64_C(0x0006060606060606);
763
764	/* Setup Nested Paging. This doesn't change throughout the execution time of the VM. */
765	pVmcb->ctrl.NestedPaging.n.u1NestedPaging = pVM->hm.s.fNestedPaging;
766
767	/* Without Nested Paging, we need additionally intercepts. */
768	if (!pVM->hm.s.fNestedPaging)
769	{
770	/* CR3 reads/writes must be intercepted; our shadow values differ from the guest values. */
771	pVmcb->ctrl.u16InterceptRdCRx \|= RT_BIT(3);
772	pVmcb->ctrl.u16InterceptWrCRx \|= RT_BIT(3);
773
774	/* Intercept INVLPG and task switches (may change CR3, EFLAGS, LDT). */
775	pVmcb->ctrl.u32InterceptCtrl1 \|= SVM_CTRL1_INTERCEPT_INVLPG
776	\| SVM_CTRL1_INTERCEPT_TASK_SWITCH;
777
778	/* Page faults must be intercepted to implement shadow paging. */
779	pVmcb->ctrl.u32InterceptException \|= RT_BIT(X86_XCPT_PF);
780	}
781
782	#ifdef HMSVM_ALWAYS_TRAP_TASK_SWITCH
783	pVmcb->ctrl.u32InterceptCtrl1 \|= SVM_CTRL1_INTERCEPT_TASK_SWITCH;
784	#endif
785
786	/*
787	* The following MSRs are saved/restored automatically during the world-switch.
788	* Don't intercept guest read/write accesses to these MSRs.
789	*/
790	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
791	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_CSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
792	hmR0SvmSetMsrPermission(pVCpu, MSR_K6_STAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
793	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_SF_MASK, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
794	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_FS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
795	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_GS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
796	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_KERNEL_GS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
797	hmR0SvmSetMsrPermission(pVCpu, MSR_IA32_SYSENTER_CS, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
798	hmR0SvmSetMsrPermission(pVCpu, MSR_IA32_SYSENTER_ESP, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
799	hmR0SvmSetMsrPermission(pVCpu, MSR_IA32_SYSENTER_EIP, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
800	}
801
802	return VINF_SUCCESS;
803	}
804
805
806	/**
807	* Invalidates a guest page by guest virtual address.
808	*
809	* @returns VBox status code.
810	* @param pVM Pointer to the VM.
811	* @param pVCpu Pointer to the VMCPU.
812	* @param GCVirt Guest virtual address of the page to invalidate.
813	*/
814	VMMR0DECL(int) SVMR0InvalidatePage(PVM pVM, PVMCPU pVCpu, RTGCPTR GCVirt)
815	{
816	AssertReturn(pVM, VERR_INVALID_PARAMETER);
817	Assert(pVM->hm.s.svm.fSupported);
818
819	bool fFlushPending = pVM->hm.s.svm.fAlwaysFlushTLB \|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_TLB_FLUSH);
820
821	/* Skip it if a TLB flush is already pending. */
822	if (!fFlushPending)
823	{
824	Log4(("SVMR0InvalidatePage %RGv\n", GCVirt));
825
826	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
827	AssertMsgReturn(pVmcb, ("Invalid pVmcb!\n"), VERR_SVM_INVALID_PVMCB);
828
829	#if HC_ARCH_BITS == 32
830	/* If we get a flush in 64-bit guest mode, then force a full TLB flush. INVLPGA takes only 32-bit addresses. */
831	if (CPUMIsGuestInLongMode(pVCpu))
832	VMCPU_FF_SET(pVCpu, VMCPU_FF_TLB_FLUSH);
833	else
834	#endif
835	{
836	SVMR0InvlpgA(GCVirt, pVmcb->ctrl.TLBCtrl.n.u32ASID);
837	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlbInvlpgVirt);
838	}
839	}
840	return VINF_SUCCESS;
841	}
842
843
844	/**
845	* Flushes the appropriate tagged-TLB entries.
846	*
847	* @param pVM Pointer to the VM.
848	* @param pVCpu Pointer to the VMCPU.
849	*/
850	static void hmR0SvmFlushTaggedTlb(PVMCPU pVCpu)
851	{
852	PVM pVM = pVCpu->CTX_SUFF(pVM);
853	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
854	PHMGLOBALCPUINFO pCpu = HMR0GetCurrentCpu();
855
856	/*
857	* Force a TLB flush for the first world switch if the current CPU differs from the one we ran on last.
858	* This can happen both for start & resume due to long jumps back to ring-3.
859	* If the TLB flush count changed, another VM (VCPU rather) has hit the ASID limit while flushing the TLB,
860	* so we cannot reuse the ASIDs without flushing.
861	*/
862	bool fNewAsid = false;
863	Assert(pCpu->idCpu != NIL_RTCPUID);
864	if ( pVCpu->hm.s.idLastCpu != pCpu->idCpu
865	\|\| pVCpu->hm.s.cTlbFlushes != pCpu->cTlbFlushes)
866	{
867	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlbWorldSwitch);
868	pVCpu->hm.s.fForceTLBFlush = true;
869	fNewAsid = true;
870	}
871
872	/* Set TLB flush state as checked until we return from the world switch. */
873	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, true);
874
875	/* Check for explicit TLB shootdowns. */
876	if (VMCPU_FF_TEST_AND_CLEAR(pVCpu, VMCPU_FF_TLB_FLUSH))
877	{
878	pVCpu->hm.s.fForceTLBFlush = true;
879	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlb);
880	}
881
882	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_NOTHING;
883
884	if (pVM->hm.s.svm.fAlwaysFlushTLB)
885	{
886	/*
887	* This is the AMD erratum 170. We need to flush the entire TLB for each world switch. Sad.
888	*/
889	pCpu->uCurrentAsid = 1;
890	pVCpu->hm.s.uCurrentAsid = 1;
891	pVCpu->hm.s.cTlbFlushes = pCpu->cTlbFlushes;
892	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
893
894	/* Clear the VMCB Clean Bit for NP while flushing the TLB. See @bugref{7152}. */
895	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
896	}
897	else if (pVCpu->hm.s.fForceTLBFlush)
898	{
899	/* Clear the VMCB Clean Bit for NP while flushing the TLB. See @bugref{7152}. */
900	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
901
902	if (fNewAsid)
903	{
904	++pCpu->uCurrentAsid;
905	bool fHitASIDLimit = false;
906	if (pCpu->uCurrentAsid >= pVM->hm.s.uMaxAsid)
907	{
908	pCpu->uCurrentAsid = 1; /* Wraparound at 1; host uses 0 */
909	pCpu->cTlbFlushes++; /* All VCPUs that run on this host CPU must use a new VPID. */
910	fHitASIDLimit = true;
911
912	if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_FLUSH_BY_ASID)
913	{
914	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_SINGLE_CONTEXT;
915	pCpu->fFlushAsidBeforeUse = true;
916	}
917	else
918	{
919	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
920	pCpu->fFlushAsidBeforeUse = false;
921	}
922	}
923
924	if ( !fHitASIDLimit
925	&& pCpu->fFlushAsidBeforeUse)
926	{
927	if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_FLUSH_BY_ASID)
928	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_SINGLE_CONTEXT;
929	else
930	{
931	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
932	pCpu->fFlushAsidBeforeUse = false;
933	}
934	}
935
936	pVCpu->hm.s.uCurrentAsid = pCpu->uCurrentAsid;
937	pVCpu->hm.s.idLastCpu = pCpu->idCpu;
938	pVCpu->hm.s.cTlbFlushes = pCpu->cTlbFlushes;
939	}
940	else
941	{
942	if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_FLUSH_BY_ASID)
943	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_SINGLE_CONTEXT;
944	else
945	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
946	}
947
948	pVCpu->hm.s.fForceTLBFlush = false;
949	}
950	/** @todo We never set VMCPU_FF_TLB_SHOOTDOWN anywhere so this path should
951	* not be executed. See hmQueueInvlPage() where it is commented
952	* out. Support individual entry flushing someday. */
953	#if 0
954	else
955	{
956	if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_TLB_SHOOTDOWN))
957	{
958	/* Deal with pending TLB shootdown actions which were queued when we were not executing code. */
959	STAM_COUNTER_INC(&pVCpu->hm.s.StatTlbShootdown);
960	for (uint32_t i = 0; i < pVCpu->hm.s.TlbShootdown.cPages; i++)
961	SVMR0InvlpgA(pVCpu->hm.s.TlbShootdown.aPages[i], pVmcb->ctrl.TLBCtrl.n.u32ASID);
962
963	pVCpu->hm.s.TlbShootdown.cPages = 0;
964	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_TLB_SHOOTDOWN);
965	}
966	}
967	#endif
968
969
970	/* Update VMCB with the ASID. */
971	if (pVmcb->ctrl.TLBCtrl.n.u32ASID != pVCpu->hm.s.uCurrentAsid)
972	{
973	pVmcb->ctrl.TLBCtrl.n.u32ASID = pVCpu->hm.s.uCurrentAsid;
974	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_ASID;
975	}
976
977	AssertMsg(pVCpu->hm.s.idLastCpu == pCpu->idCpu,
978	("vcpu idLastCpu=%x pcpu idCpu=%x\n", pVCpu->hm.s.idLastCpu, pCpu->idCpu));
979	AssertMsg(pVCpu->hm.s.cTlbFlushes == pCpu->cTlbFlushes,
980	("Flush count mismatch for cpu %d (%x vs %x)\n", pCpu->idCpu, pVCpu->hm.s.cTlbFlushes, pCpu->cTlbFlushes));
981	AssertMsg(pCpu->uCurrentAsid >= 1 && pCpu->uCurrentAsid < pVM->hm.s.uMaxAsid,
982	("cpu%d uCurrentAsid = %x\n", pCpu->idCpu, pCpu->uCurrentAsid));
983	AssertMsg(pVCpu->hm.s.uCurrentAsid >= 1 && pVCpu->hm.s.uCurrentAsid < pVM->hm.s.uMaxAsid,
984	("cpu%d VM uCurrentAsid = %x\n", pCpu->idCpu, pVCpu->hm.s.uCurrentAsid));
985
986	#ifdef VBOX_WITH_STATISTICS
987	if (pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_NOTHING)
988	STAM_COUNTER_INC(&pVCpu->hm.s.StatNoFlushTlbWorldSwitch);
989	else if ( pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_SINGLE_CONTEXT
990	\|\| pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_SINGLE_CONTEXT_RETAIN_GLOBALS)
991	{
992	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushAsid);
993	}
994	else
995	{
996	Assert(pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_ENTIRE);
997	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushEntire);
998	}
999	#endif
1000	}
1001
1002
1003	/** @name 64-bit guest on 32-bit host OS helper functions.
1004	*
1005	* The host CPU is still 64-bit capable but the host OS is running in 32-bit
1006	* mode (code segment, paging). These wrappers/helpers perform the necessary
1007	* bits for the 32->64 switcher.
1008	*
1009	* @{ */
1010	#if HC_ARCH_BITS == 32 && defined(VBOX_ENABLE_64_BITS_GUESTS) && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
1011	/**
1012	* Prepares for and executes VMRUN (64-bit guests on a 32-bit host).
1013	*
1014	* @returns VBox status code.
1015	* @param HCPhysVmcbHost Physical address of host VMCB.
1016	* @param HCPhysVmcb Physical address of the VMCB.
1017	* @param pCtx Pointer to the guest-CPU context.
1018	* @param pVM Pointer to the VM.
1019	* @param pVCpu Pointer to the VMCPU.
1020	*/
1021	DECLASM(int) SVMR0VMSwitcherRun64(RTHCPHYS HCPhysVmcbHost, RTHCPHYS HCPhysVmcb, PCPUMCTX pCtx, PVM pVM, PVMCPU pVCpu)
1022	{
1023	uint32_t aParam[4];
1024	aParam[0] = (uint32_t)(HCPhysVmcbHost); /* Param 1: HCPhysVmcbHost - Lo. */
1025	aParam[1] = (uint32_t)(HCPhysVmcbHost >> 32); /* Param 1: HCPhysVmcbHost - Hi. */
1026	aParam[2] = (uint32_t)(HCPhysVmcb); /* Param 2: HCPhysVmcb - Lo. */
1027	aParam[3] = (uint32_t)(HCPhysVmcb >> 32); /* Param 2: HCPhysVmcb - Hi. */
1028
1029	return SVMR0Execute64BitsHandler(pVM, pVCpu, pCtx, HM64ON32OP_SVMRCVMRun64, 4, &aParam[0]);
1030	}
1031
1032
1033	/**
1034	* Executes the specified VMRUN handler in 64-bit mode.
1035	*
1036	* @returns VBox status code.
1037	* @param pVM Pointer to the VM.
1038	* @param pVCpu Pointer to the VMCPU.
1039	* @param pCtx Pointer to the guest-CPU context.
1040	* @param enmOp The operation to perform.
1041	* @param cbParam Number of parameters.
1042	* @param paParam Array of 32-bit parameters.
1043	*/
1044	VMMR0DECL(int) SVMR0Execute64BitsHandler(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, HM64ON32OP enmOp, uint32_t cbParam,
1045	uint32_t *paParam)
1046	{
1047	AssertReturn(pVM->hm.s.pfnHost32ToGuest64R0, VERR_HM_NO_32_TO_64_SWITCHER);
1048	Assert(enmOp > HM64ON32OP_INVALID && enmOp < HM64ON32OP_END);
1049
1050	/* Disable interrupts. */
1051	RTHCUINTREG uOldEFlags = ASMIntDisableFlags();
1052
1053	#ifdef VBOX_WITH_VMMR0_DISABLE_LAPIC_NMI
1054	RTCPUID idHostCpu = RTMpCpuId();
1055	CPUMR0SetLApic(pVCpu, idHostCpu);
1056	#endif
1057
1058	CPUMSetHyperESP(pVCpu, VMMGetStackRC(pVCpu));
1059	CPUMSetHyperEIP(pVCpu, enmOp);
1060	for (int i = (int)cbParam - 1; i >= 0; i--)
1061	CPUMPushHyper(pVCpu, paParam[i]);
1062
1063	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatWorldSwitch3264, z);
1064	/* Call the switcher. */
1065	int rc = pVM->hm.s.pfnHost32ToGuest64R0(pVM, RT_OFFSETOF(VM, aCpus[pVCpu->idCpu].cpum) - RT_OFFSETOF(VM, cpum));
1066	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatWorldSwitch3264, z);
1067
1068	/* Restore interrupts. */
1069	ASMSetFlags(uOldEFlags);
1070	return rc;
1071	}
1072
1073	#endif /* HC_ARCH_BITS == 32 && defined(VBOX_ENABLE_64_BITS_GUESTS) */
1074	/** @} */
1075
1076
1077	/**
1078	* Adds an exception to the intercept exception bitmap in the VMCB and updates
1079	* the corresponding VMCB Clean bit.
1080	*
1081	* @param pVmcb Pointer to the VM control block.
1082	* @param u32Xcpt The value of the exception (X86_XCPT_*).
1083	*/
1084	DECLINLINE(void) hmR0SvmAddXcptIntercept(PSVMVMCB pVmcb, uint32_t u32Xcpt)
1085	{
1086	if (!(pVmcb->ctrl.u32InterceptException & RT_BIT(u32Xcpt)))
1087	{
1088	pVmcb->ctrl.u32InterceptException \|= RT_BIT(u32Xcpt);
1089	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1090	}
1091	}
1092
1093
1094	/**
1095	* Removes an exception from the intercept-exception bitmap in the VMCB and
1096	* updates the corresponding VMCB Clean bit.
1097	*
1098	* @param pVmcb Pointer to the VM control block.
1099	* @param u32Xcpt The value of the exception (X86_XCPT_*).
1100	*/
1101	DECLINLINE(void) hmR0SvmRemoveXcptIntercept(PSVMVMCB pVmcb, uint32_t u32Xcpt)
1102	{
1103	#ifndef HMSVM_ALWAYS_TRAP_ALL_XCPTS
1104	if (pVmcb->ctrl.u32InterceptException & RT_BIT(u32Xcpt))
1105	{
1106	pVmcb->ctrl.u32InterceptException &= ~RT_BIT(u32Xcpt);
1107	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1108	}
1109	#endif
1110	}
1111
1112
1113	/**
1114	* Loads the guest CR0 control register into the guest-state area in the VMCB.
1115	* Although the guest CR0 is a separate field in the VMCB we have to consider
1116	* the FPU state itself which is shared between the host and the guest.
1117	*
1118	* @returns VBox status code.
1119	* @param pVM Pointer to the VMCPU.
1120	* @param pVmcb Pointer to the VM control block.
1121	* @param pCtx Pointer to the guest-CPU context.
1122	*
1123	* @remarks No-long-jump zone!!!
1124	*/
1125	static void hmR0SvmLoadSharedCR0(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1126	{
1127	/*
1128	* Guest CR0.
1129	*/
1130	PVM pVM = pVCpu->CTX_SUFF(pVM);
1131	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR0))
1132	{
1133	uint64_t u64GuestCR0 = pCtx->cr0;
1134
1135	/* Always enable caching. */
1136	u64GuestCR0 &= ~(X86_CR0_CD \| X86_CR0_NW);
1137
1138	/*
1139	* When Nested Paging is not available use shadow page tables and intercept #PFs (the latter done in SVMR0SetupVM()).
1140	*/
1141	if (!pVM->hm.s.fNestedPaging)
1142	{
1143	u64GuestCR0 \|= X86_CR0_PG; /* When Nested Paging is not available, use shadow page tables. */
1144	u64GuestCR0 \|= X86_CR0_WP; /* Guest CPL 0 writes to its read-only pages should cause a #PF #VMEXIT. */
1145	}
1146
1147	/*
1148	* Guest FPU bits.
1149	*/
1150	bool fInterceptNM = false;
1151	bool fInterceptMF = false;
1152	u64GuestCR0 \|= X86_CR0_NE; /* Use internal x87 FPU exceptions handling rather than external interrupts. */
1153	if (CPUMIsGuestFPUStateActive(pVCpu))
1154	{
1155	/* Catch floating point exceptions if we need to report them to the guest in a different way. */
1156	if (!(pCtx->cr0 & X86_CR0_NE))
1157	{
1158	Log4(("hmR0SvmLoadGuestControlRegs: Intercepting Guest CR0.MP Old-style FPU handling!!!\n"));
1159	fInterceptMF = true;
1160	}
1161	}
1162	else
1163	{
1164	fInterceptNM = true; /* Guest FPU inactive, #VMEXIT on #NM for lazy FPU loading. */
1165	u64GuestCR0 \|= X86_CR0_TS /* Guest can task switch quickly and do lazy FPU syncing. */
1166	\| X86_CR0_MP; /* FWAIT/WAIT should not ignore CR0.TS and should generate #NM. */
1167	}
1168
1169	/*
1170	* Update the exception intercept bitmap.
1171	*/
1172	if (fInterceptNM)
1173	hmR0SvmAddXcptIntercept(pVmcb, X86_XCPT_NM);
1174	else
1175	hmR0SvmRemoveXcptIntercept(pVmcb, X86_XCPT_NM);
1176
1177	if (fInterceptMF)
1178	hmR0SvmAddXcptIntercept(pVmcb, X86_XCPT_MF);
1179	else
1180	hmR0SvmRemoveXcptIntercept(pVmcb, X86_XCPT_MF);
1181
1182	pVmcb->guest.u64CR0 = u64GuestCR0;
1183	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1184	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_CR0);
1185	}
1186	}
1187
1188
1189	/**
1190	* Loads the guest control registers (CR2, CR3, CR4) into the VMCB.
1191	*
1192	* @returns VBox status code.
1193	* @param pVCpu Pointer to the VMCPU.
1194	* @param pVmcb Pointer to the VM control block.
1195	* @param pCtx Pointer to the guest-CPU context.
1196	*
1197	* @remarks No-long-jump zone!!!
1198	*/
1199	static int hmR0SvmLoadGuestControlRegs(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1200	{
1201	PVM pVM = pVCpu->CTX_SUFF(pVM);
1202
1203	/*
1204	* Guest CR2.
1205	*/
1206	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR2))
1207	{
1208	pVmcb->guest.u64CR2 = pCtx->cr2;
1209	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CR2;
1210	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_CR2);
1211	}
1212
1213	/*
1214	* Guest CR3.
1215	*/
1216	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR3))
1217	{
1218	if (pVM->hm.s.fNestedPaging)
1219	{
1220	PGMMODE enmShwPagingMode;
1221	#if HC_ARCH_BITS == 32
1222	if (CPUMIsGuestInLongModeEx(pCtx))
1223	enmShwPagingMode = PGMMODE_AMD64_NX;
1224	else
1225	#endif
1226	enmShwPagingMode = PGMGetHostMode(pVM);
1227
1228	pVmcb->ctrl.u64NestedPagingCR3 = PGMGetNestedCR3(pVCpu, enmShwPagingMode);
1229	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
1230	Assert(pVmcb->ctrl.u64NestedPagingCR3);
1231	pVmcb->guest.u64CR3 = pCtx->cr3;
1232	}
1233	else
1234	pVmcb->guest.u64CR3 = PGMGetHyperCR3(pVCpu);
1235
1236	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1237	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_CR3);
1238	}
1239
1240	/*
1241	* Guest CR4.
1242	*/
1243	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR4))
1244	{
1245	uint64_t u64GuestCR4 = pCtx->cr4;
1246	if (!pVM->hm.s.fNestedPaging)
1247	{
1248	switch (pVCpu->hm.s.enmShadowMode)
1249	{
1250	case PGMMODE_REAL:
1251	case PGMMODE_PROTECTED: /* Protected mode, no paging. */
1252	AssertFailed();
1253	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1254
1255	case PGMMODE_32_BIT: /* 32-bit paging. */
1256	u64GuestCR4 &= ~X86_CR4_PAE;
1257	break;
1258
1259	case PGMMODE_PAE: /* PAE paging. */
1260	case PGMMODE_PAE_NX: /* PAE paging with NX enabled. */
1261	/** Must use PAE paging as we could use physical memory > 4 GB */
1262	u64GuestCR4 \|= X86_CR4_PAE;
1263	break;
1264
1265	case PGMMODE_AMD64: /* 64-bit AMD paging (long mode). */
1266	case PGMMODE_AMD64_NX: /* 64-bit AMD paging (long mode) with NX enabled. */
1267	#ifdef VBOX_ENABLE_64_BITS_GUESTS
1268	break;
1269	#else
1270	AssertFailed();
1271	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1272	#endif
1273
1274	default: /* shut up gcc */
1275	AssertFailed();
1276	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1277	}
1278	}
1279
1280	pVmcb->guest.u64CR4 = u64GuestCR4;
1281	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1282	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_CR4);
1283	}
1284
1285	return VINF_SUCCESS;
1286	}
1287
1288
1289	/**
1290	* Loads the guest segment registers into the VMCB.
1291	*
1292	* @returns VBox status code.
1293	* @param pVCpu Pointer to the VMCPU.
1294	* @param pVmcb Pointer to the VM control block.
1295	* @param pCtx Pointer to the guest-CPU context.
1296	*
1297	* @remarks No-long-jump zone!!!
1298	*/
1299	static void hmR0SvmLoadGuestSegmentRegs(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1300	{
1301	/* Guest Segment registers: CS, SS, DS, ES, FS, GS. */
1302	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_SEGMENT_REGS))
1303	{
1304	HMSVM_LOAD_SEG_REG(CS, cs);
1305	HMSVM_LOAD_SEG_REG(SS, ss);
1306	HMSVM_LOAD_SEG_REG(DS, ds);
1307	HMSVM_LOAD_SEG_REG(ES, es);
1308	HMSVM_LOAD_SEG_REG(FS, fs);
1309	HMSVM_LOAD_SEG_REG(GS, gs);
1310
1311	pVmcb->guest.u8CPL = pCtx->ss.Attr.n.u2Dpl;
1312	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_SEG;
1313	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_SEGMENT_REGS);
1314	}
1315
1316	/* Guest TR. */
1317	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_TR))
1318	{
1319	HMSVM_LOAD_SEG_REG(TR, tr);
1320	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_TR);
1321	}
1322
1323	/* Guest LDTR. */
1324	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_LDTR))
1325	{
1326	HMSVM_LOAD_SEG_REG(LDTR, ldtr);
1327	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_LDTR);
1328	}
1329
1330	/* Guest GDTR. */
1331	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_GDTR))
1332	{
1333	pVmcb->guest.GDTR.u32Limit = pCtx->gdtr.cbGdt;
1334	pVmcb->guest.GDTR.u64Base = pCtx->gdtr.pGdt;
1335	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DT;
1336	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_GDTR);
1337	}
1338
1339	/* Guest IDTR. */
1340	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_IDTR))
1341	{
1342	pVmcb->guest.IDTR.u32Limit = pCtx->idtr.cbIdt;
1343	pVmcb->guest.IDTR.u64Base = pCtx->idtr.pIdt;
1344	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DT;
1345	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_IDTR);
1346	}
1347	}
1348
1349
1350	/**
1351	* Loads the guest MSRs into the VMCB.
1352	*
1353	* @param pVCpu Pointer to the VMCPU.
1354	* @param pVmcb Pointer to the VM control block.
1355	* @param pCtx Pointer to the guest-CPU context.
1356	*
1357	* @remarks No-long-jump zone!!!
1358	*/
1359	static void hmR0SvmLoadGuestMsrs(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1360	{
1361	/* Guest Sysenter MSRs. */
1362	pVmcb->guest.u64SysEnterCS = pCtx->SysEnter.cs;
1363	pVmcb->guest.u64SysEnterEIP = pCtx->SysEnter.eip;
1364	pVmcb->guest.u64SysEnterESP = pCtx->SysEnter.esp;
1365
1366	/*
1367	* Guest EFER MSR.
1368	* AMD-V requires guest EFER.SVME to be set. Weird.
1369	* See AMD spec. 15.5.1 "Basic Operation" \| "Canonicalization and Consistency Checks".
1370	*/
1371	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_EFER_MSR))
1372	{
1373	pVmcb->guest.u64EFER = pCtx->msrEFER \| MSR_K6_EFER_SVME;
1374	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1375	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_EFER_MSR);
1376	}
1377
1378	/* 64-bit MSRs. */
1379	if (CPUMIsGuestInLongModeEx(pCtx))
1380	{
1381	pVmcb->guest.FS.u64Base = pCtx->fs.u64Base;
1382	pVmcb->guest.GS.u64Base = pCtx->gs.u64Base;
1383	}
1384	else
1385	{
1386	/* If the guest isn't in 64-bit mode, clear MSR_K6_LME bit from guest EFER otherwise AMD-V expects amd64 shadow paging. */
1387	if (pCtx->msrEFER & MSR_K6_EFER_LME)
1388	{
1389	pVmcb->guest.u64EFER &= ~MSR_K6_EFER_LME;
1390	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1391	}
1392	}
1393
1394
1395	/** @todo The following are used in 64-bit only (SYSCALL/SYSRET) but they might
1396	* be writable in 32-bit mode. Clarify with AMD spec. */
1397	pVmcb->guest.u64STAR = pCtx->msrSTAR;
1398	pVmcb->guest.u64LSTAR = pCtx->msrLSTAR;
1399	pVmcb->guest.u64CSTAR = pCtx->msrCSTAR;
1400	pVmcb->guest.u64SFMASK = pCtx->msrSFMASK;
1401	pVmcb->guest.u64KernelGSBase = pCtx->msrKERNELGSBASE;
1402	}
1403
1404
1405	/**
1406	* Loads the guest state into the VMCB and programs the necessary intercepts
1407	* accordingly.
1408	*
1409	* @param pVCpu Pointer to the VMCPU.
1410	* @param pVmcb Pointer to the VM control block.
1411	* @param pCtx Pointer to the guest-CPU context.
1412	*
1413	* @remarks No-long-jump zone!!!
1414	* @remarks Requires EFLAGS to be up-to-date in the VMCB!
1415	*/
1416	static void hmR0SvmLoadSharedDebugState(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1417	{
1418	if (!HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_DEBUG))
1419	return;
1420	Assert((pCtx->dr[6] & X86_DR6_RA1_MASK) == X86_DR6_RA1_MASK); Assert((pCtx->dr[6] & X86_DR6_RAZ_MASK) == 0);
1421	Assert((pCtx->dr[7] & X86_DR7_RA1_MASK) == X86_DR7_RA1_MASK); Assert((pCtx->dr[7] & X86_DR7_RAZ_MASK) == 0);
1422
1423	bool fInterceptDB = false;
1424	bool fInterceptMovDRx = false;
1425
1426	/*
1427	* Anyone single stepping on the host side? If so, we'll have to use the
1428	* trap flag in the guest EFLAGS since AMD-V doesn't have a trap flag on
1429	* the VMM level like the VT-x implementations does.
1430	*/
1431	bool const fStepping = pVCpu->hm.s.fSingleInstruction \|\| DBGFIsStepping(pVCpu);
1432	if (fStepping)
1433	{
1434	pVCpu->hm.s.fClearTrapFlag = true;
1435	pVmcb->guest.u64RFlags \|= X86_EFL_TF;
1436	fInterceptDB = true;
1437	fInterceptMovDRx = true; /* Need clean DR6, no guest mess. */
1438	}
1439
1440	if ( fStepping
1441	\|\| (CPUMGetHyperDR7(pVCpu) & X86_DR7_ENABLED_MASK))
1442	{
1443	/*
1444	* Use the combined guest and host DRx values found in the hypervisor
1445	* register set because the debugger has breakpoints active or someone
1446	* is single stepping on the host side.
1447	*
1448	* Note! DBGF expects a clean DR6 state before executing guest code.
1449	*/
1450	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS) && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
1451	if ( CPUMIsGuestInLongModeEx(pCtx)
1452	&& !CPUMIsHyperDebugStateActivePending(pVCpu))
1453	{
1454	CPUMR0LoadHyperDebugState(pVCpu, false /* include DR6 */);
1455	Assert(!CPUMIsGuestDebugStateActivePending(pVCpu));
1456	Assert(CPUMIsHyperDebugStateActivePending(pVCpu));
1457	}
1458	else
1459	#endif
1460	if (!CPUMIsHyperDebugStateActive(pVCpu))
1461	{
1462	CPUMR0LoadHyperDebugState(pVCpu, false /* include DR6 */);
1463	Assert(!CPUMIsGuestDebugStateActive(pVCpu));
1464	Assert(CPUMIsHyperDebugStateActive(pVCpu));
1465	}
1466
1467	/* Update DR6 & DR7. (The other DRx values are handled by CPUM one way or the other.) */
1468	if ( pVmcb->guest.u64DR6 != X86_DR6_INIT_VAL
1469	\|\| pVmcb->guest.u64DR7 != CPUMGetHyperDR7(pVCpu))
1470	{
1471	pVmcb->guest.u64DR7 = CPUMGetHyperDR7(pVCpu);
1472	pVmcb->guest.u64DR6 = X86_DR6_INIT_VAL;
1473	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
1474	pVCpu->hm.s.fUsingHyperDR7 = true;
1475	}
1476
1477	/** @todo If we cared, we could optimize to allow the guest to read registers
1478	* with the same values. */
1479	fInterceptDB = true;
1480	fInterceptMovDRx = true;
1481	Log5(("hmR0SvmLoadSharedDebugState: Loaded hyper DRx\n"));
1482	}
1483	else
1484	{
1485	/*
1486	* Update DR6, DR7 with the guest values if necessary.
1487	*/
1488	if ( pVmcb->guest.u64DR7 != pCtx->dr[7]
1489	\|\| pVmcb->guest.u64DR6 != pCtx->dr[6])
1490	{
1491	pVmcb->guest.u64DR7 = pCtx->dr[7];
1492	pVmcb->guest.u64DR6 = pCtx->dr[6];
1493	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
1494	pVCpu->hm.s.fUsingHyperDR7 = false;
1495	}
1496
1497	/*
1498	* If the guest has enabled debug registers, we need to load them prior to
1499	* executing guest code so they'll trigger at the right time.
1500	*/
1501	if (pCtx->dr[7] & (X86_DR7_ENABLED_MASK \| X86_DR7_GD)) /** @todo Why GD? */
1502	{
1503	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS) && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
1504	if ( CPUMIsGuestInLongModeEx(pCtx)
1505	&& !CPUMIsGuestDebugStateActivePending(pVCpu))
1506	{
1507	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
1508	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxArmed);
1509	Assert(!CPUMIsHyperDebugStateActivePending(pVCpu));
1510	Assert(CPUMIsGuestDebugStateActivePending(pVCpu));
1511	}
1512	else
1513	#endif
1514	if (!CPUMIsGuestDebugStateActive(pVCpu))
1515	{
1516	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
1517	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxArmed);
1518	Assert(!CPUMIsHyperDebugStateActive(pVCpu));
1519	Assert(CPUMIsGuestDebugStateActive(pVCpu));
1520	}
1521	Log5(("hmR0SvmLoadSharedDebugState: Loaded guest DRx\n"));
1522	}
1523	/*
1524	* If no debugging enabled, we'll lazy load DR0-3. We don't need to
1525	* intercept #DB as DR6 is updated in the VMCB.
1526	*/
1527	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS) && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
1528	else if ( !CPUMIsGuestDebugStateActivePending(pVCpu)
1529	&& !CPUMIsGuestDebugStateActive(pVCpu))
1530	#else
1531	else if (!CPUMIsGuestDebugStateActive(pVCpu))
1532	#endif
1533	{
1534	fInterceptMovDRx = true;
1535	}
1536	}
1537
1538	/*
1539	* Set up the intercepts.
1540	*/
1541	if (fInterceptDB)
1542	hmR0SvmAddXcptIntercept(pVmcb, X86_XCPT_DB);
1543	else
1544	hmR0SvmRemoveXcptIntercept(pVmcb, X86_XCPT_DB);
1545
1546	if (fInterceptMovDRx)
1547	{
1548	if ( pVmcb->ctrl.u16InterceptRdDRx != 0xffff
1549	\|\| pVmcb->ctrl.u16InterceptWrDRx != 0xffff)
1550	{
1551	pVmcb->ctrl.u16InterceptRdDRx = 0xffff;
1552	pVmcb->ctrl.u16InterceptWrDRx = 0xffff;
1553	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1554	}
1555	}
1556	else
1557	{
1558	if ( pVmcb->ctrl.u16InterceptRdDRx
1559	\|\| pVmcb->ctrl.u16InterceptWrDRx)
1560	{
1561	pVmcb->ctrl.u16InterceptRdDRx = 0;
1562	pVmcb->ctrl.u16InterceptWrDRx = 0;
1563	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1564	}
1565	}
1566
1567	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_DEBUG);
1568	}
1569
1570
1571	/**
1572	* Loads the guest APIC state (currently just the TPR).
1573	*
1574	* @returns VBox status code.
1575	* @param pVCpu Pointer to the VMCPU.
1576	* @param pVmcb Pointer to the VM control block.
1577	* @param pCtx Pointer to the guest-CPU context.
1578	*/
1579	static int hmR0SvmLoadGuestApicState(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1580	{
1581	if (!HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE))
1582	return VINF_SUCCESS;
1583
1584	bool fPendingIntr;
1585	uint8_t u8Tpr;
1586	int rc = PDMApicGetTPR(pVCpu, &u8Tpr, &fPendingIntr, NULL /* pu8PendingIrq */);
1587	AssertRCReturn(rc, rc);
1588
1589	/* Assume that we need to trap all TPR accesses and thus need not check on
1590	every #VMEXIT if we should update the TPR. */
1591	Assert(pVmcb->ctrl.IntCtrl.n.u1VIrqMasking);
1592	pVCpu->hm.s.svm.fSyncVTpr = false;
1593
1594	/* 32-bit guests uses LSTAR MSR for patching guest code which touches the TPR. */
1595	if (pVCpu->CTX_SUFF(pVM)->hm.s.fTPRPatchingActive)
1596	{
1597	pCtx->msrLSTAR = u8Tpr;
1598
1599	/* If there are interrupts pending, intercept LSTAR writes, otherwise don't intercept reads or writes. */
1600	if (fPendingIntr)
1601	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_INTERCEPT_WRITE);
1602	else
1603	{
1604	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1605	pVCpu->hm.s.svm.fSyncVTpr = true;
1606	}
1607	}
1608	else
1609	{
1610	/* Bits 3-0 of the VTPR field correspond to bits 7-4 of the TPR (which is the Task-Priority Class). */
1611	pVmcb->ctrl.IntCtrl.n.u8VTPR = (u8Tpr >> 4);
1612
1613	/* If there are interrupts pending, intercept CR8 writes to evaluate ASAP if we can deliver the interrupt to the guest. */
1614	if (fPendingIntr)
1615	pVmcb->ctrl.u16InterceptWrCRx \|= RT_BIT(8);
1616	else
1617	{
1618	pVmcb->ctrl.u16InterceptWrCRx &= ~RT_BIT(8);
1619	pVCpu->hm.s.svm.fSyncVTpr = true;
1620	}
1621
1622	pVmcb->ctrl.u64VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS \| HMSVM_VMCB_CLEAN_TPR);
1623	}
1624
1625	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
1626	return rc;
1627	}
1628
1629
1630	/**
1631	* Sets up the appropriate function to run guest code.
1632	*
1633	* @returns VBox status code.
1634	* @param pVCpu Pointer to the VMCPU.
1635	* @param pCtx Pointer to the guest-CPU context.
1636	*
1637	* @remarks No-long-jump zone!!!
1638	*/
1639	static int hmR0SvmSetupVMRunHandler(PVMCPU pVCpu, PCPUMCTX pCtx)
1640	{
1641	if (CPUMIsGuestInLongModeEx(pCtx))
1642	{
1643	#ifndef VBOX_ENABLE_64_BITS_GUESTS
1644	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1645	#endif
1646	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.fAllow64BitGuests); /* Guaranteed by hmR3InitFinalizeR0(). */
1647	#if HC_ARCH_BITS == 32 && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
1648	/* 32-bit host. We need to switch to 64-bit before running the 64-bit guest. */
1649	pVCpu->hm.s.svm.pfnVMRun = SVMR0VMSwitcherRun64;
1650	#else
1651	/* 64-bit host or hybrid host. */
1652	pVCpu->hm.s.svm.pfnVMRun = SVMR0VMRun64;
1653	#endif
1654	}
1655	else
1656	{
1657	/* Guest is not in long mode, use the 32-bit handler. */
1658	pVCpu->hm.s.svm.pfnVMRun = SVMR0VMRun;
1659	}
1660	return VINF_SUCCESS;
1661	}
1662
1663
1664	/**
1665	* Enters the AMD-V session.
1666	*
1667	* @returns VBox status code.
1668	* @param pVM Pointer to the VM.
1669	* @param pVCpu Pointer to the VMCPU.
1670	* @param pCpu Pointer to the CPU info struct.
1671	*/
1672	VMMR0DECL(int) SVMR0Enter(PVM pVM, PVMCPU pVCpu, PHMGLOBALCPUINFO pCpu)
1673	{
1674	AssertPtr(pVM);
1675	AssertPtr(pVCpu);
1676	Assert(pVM->hm.s.svm.fSupported);
1677	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1678	NOREF(pVM); NOREF(pCpu);
1679
1680	LogFlowFunc(("pVM=%p pVCpu=%p\n", pVM, pVCpu));
1681	Assert(HMCPU_CF_IS_SET(pVCpu, HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_HOST_GUEST_SHARED_STATE));
1682
1683	pVCpu->hm.s.fLeaveDone = false;
1684	return VINF_SUCCESS;
1685	}
1686
1687
1688	/**
1689	* Thread-context callback for AMD-V.
1690	*
1691	* @param enmEvent The thread-context event.
1692	* @param pVCpu Pointer to the VMCPU.
1693	* @param fGlobalInit Whether global VT-x/AMD-V init. is used.
1694	* @thread EMT(pVCpu)
1695	*/
1696	VMMR0DECL(void) SVMR0ThreadCtxCallback(RTTHREADCTXEVENT enmEvent, PVMCPU pVCpu, bool fGlobalInit)
1697	{
1698	NOREF(fGlobalInit);
1699
1700	switch (enmEvent)
1701	{
1702	case RTTHREADCTXEVENT_PREEMPTING:
1703	{
1704	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1705	Assert(VMMR0ThreadCtxHooksAreRegistered(pVCpu));
1706	VMCPU_ASSERT_EMT(pVCpu);
1707
1708	PVM pVM = pVCpu->CTX_SUFF(pVM);
1709	PCPUMCTX pCtx = CPUMQueryGuestCtxPtr(pVCpu);
1710
1711	/* No longjmps (log-flush, locks) in this fragile context. */
1712	VMMRZCallRing3Disable(pVCpu);
1713
1714	if (!pVCpu->hm.s.fLeaveDone)
1715	{
1716	hmR0SvmLeave(pVM, pVCpu, pCtx);
1717	pVCpu->hm.s.fLeaveDone = true;
1718	}
1719
1720	/* Leave HM context, takes care of local init (term). */
1721	int rc = HMR0LeaveCpu(pVCpu);
1722	AssertRC(rc); NOREF(rc);
1723
1724	/* Restore longjmp state. */
1725	VMMRZCallRing3Enable(pVCpu);
1726	STAM_COUNTER_INC(&pVCpu->hm.s.StatPreemptPreempting);
1727	break;
1728	}
1729
1730	case RTTHREADCTXEVENT_RESUMED:
1731	{
1732	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1733	Assert(VMMR0ThreadCtxHooksAreRegistered(pVCpu));
1734	VMCPU_ASSERT_EMT(pVCpu);
1735
1736	/* No longjmps (log-flush, locks) in this fragile context. */
1737	VMMRZCallRing3Disable(pVCpu);
1738
1739	/*
1740	* Initialize the bare minimum state required for HM. This takes care of
1741	* initializing AMD-V if necessary (onlined CPUs, local init etc.)
1742	*/
1743	int rc = HMR0EnterCpu(pVCpu);
1744	AssertRC(rc); NOREF(rc);
1745	Assert(HMCPU_CF_IS_SET(pVCpu, HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_HOST_GUEST_SHARED_STATE));
1746
1747	pVCpu->hm.s.fLeaveDone = false;
1748
1749	/* Restore longjmp state. */
1750	VMMRZCallRing3Enable(pVCpu);
1751	break;
1752	}
1753
1754	default:
1755	break;
1756	}
1757	}
1758
1759
1760	/**
1761	* Saves the host state.
1762	*
1763	* @returns VBox status code.
1764	* @param pVM Pointer to the VM.
1765	* @param pVCpu Pointer to the VMCPU.
1766	*
1767	* @remarks No-long-jump zone!!!
1768	*/
1769	VMMR0DECL(int) SVMR0SaveHostState(PVM pVM, PVMCPU pVCpu)
1770	{
1771	NOREF(pVM);
1772	NOREF(pVCpu);
1773	/* Nothing to do here. AMD-V does this for us automatically during the world-switch. */
1774	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_HOST_CONTEXT);
1775	return VINF_SUCCESS;
1776	}
1777
1778
1779	/**
1780	* Loads the guest state into the VMCB.
1781	*
1782	* The CPU state will be loaded from these fields on every successful VM-entry.
1783	* Also sets up the appropriate VMRUN function to execute guest code based on
1784	* the guest CPU mode.
1785	*
1786	* @returns VBox status code.
1787	* @param pVM Pointer to the VM.
1788	* @param pVCpu Pointer to the VMCPU.
1789	* @param pCtx Pointer to the guest-CPU context.
1790	*
1791	* @remarks No-long-jump zone!!!
1792	*/
1793	static int hmR0SvmLoadGuestState(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
1794	{
1795	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
1796	AssertMsgReturn(pVmcb, ("Invalid pVmcb\n"), VERR_SVM_INVALID_PVMCB);
1797
1798	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatLoadGuestState, x);
1799
1800	int rc = hmR0SvmLoadGuestControlRegs(pVCpu, pVmcb, pCtx);
1801	AssertLogRelMsgRCReturn(rc, ("hmR0SvmLoadGuestControlRegs! rc=%Rrc (pVM=%p pVCpu=%p)\n", rc, pVM, pVCpu), rc);
1802
1803	hmR0SvmLoadGuestSegmentRegs(pVCpu, pVmcb, pCtx);
1804	hmR0SvmLoadGuestMsrs(pVCpu, pVmcb, pCtx);
1805
1806	pVmcb->guest.u64RIP = pCtx->rip;
1807	pVmcb->guest.u64RSP = pCtx->rsp;
1808	pVmcb->guest.u64RFlags = pCtx->eflags.u32;
1809	pVmcb->guest.u64RAX = pCtx->rax;
1810
1811	rc = hmR0SvmLoadGuestApicState(pVCpu, pVmcb, pCtx);
1812	AssertLogRelMsgRCReturn(rc, ("hmR0SvmLoadGuestApicState! rc=%Rrc (pVM=%p pVCpu=%p)\n", rc, pVM, pVCpu), rc);
1813
1814	rc = hmR0SvmSetupVMRunHandler(pVCpu, pCtx);
1815	AssertLogRelMsgRCReturn(rc, ("hmR0SvmSetupVMRunHandler! rc=%Rrc (pVM=%p pVCpu=%p)\n", rc, pVM, pVCpu), rc);
1816
1817	/* Clear any unused and reserved bits. */
1818	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_RIP /* Unused (loaded unconditionally). */
1819	\| HM_CHANGED_GUEST_RSP
1820	\| HM_CHANGED_GUEST_RFLAGS
1821	\| HM_CHANGED_GUEST_SYSENTER_CS_MSR
1822	\| HM_CHANGED_GUEST_SYSENTER_EIP_MSR
1823	\| HM_CHANGED_GUEST_SYSENTER_ESP_MSR
1824	\| HM_CHANGED_GUEST_LAZY_MSRS /* Unused. */
1825	\| HM_CHANGED_SVM_RESERVED1 /* Reserved. */
1826	\| HM_CHANGED_SVM_RESERVED2
1827	\| HM_CHANGED_SVM_RESERVED3
1828	\| HM_CHANGED_SVM_RESERVED4);
1829
1830	/* All the guest state bits should be loaded except maybe the host context and/or shared host/guest bits. */
1831	AssertMsg( !HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_ALL_GUEST)
1832	\|\| HMCPU_CF_IS_PENDING_ONLY(pVCpu, HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_HOST_GUEST_SHARED_STATE),
1833	("fContextUseFlags=%#RX32\n", HMCPU_CF_VALUE(pVCpu)));
1834
1835	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));
1836	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatLoadGuestState, x);
1837	return rc;
1838	}
1839
1840
1841	/**
1842	* Loads the state shared between the host and guest into the
1843	* VMCB.
1844	*
1845	* @param pVCpu Pointer to the VMCPU.
1846	* @param pVmcb Pointer to the VM control block.
1847	* @param pCtx Pointer to the guest-CPU context.
1848	*
1849	* @remarks No-long-jump zone!!!
1850	*/
1851	static void hmR0SvmLoadSharedState(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1852	{
1853	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1854	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
1855
1856	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR0))
1857	hmR0SvmLoadSharedCR0(pVCpu, pVmcb, pCtx);
1858
1859	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_DEBUG))
1860	hmR0SvmLoadSharedDebugState(pVCpu, pVmcb, pCtx);
1861
1862	/* Unused on AMD-V. */
1863	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_LAZY_MSRS);
1864
1865	AssertMsg(!HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_HOST_GUEST_SHARED_STATE),
1866	("fContextUseFlags=%#RX32\n", HMCPU_CF_VALUE(pVCpu)));
1867	}
1868
1869
1870	/**
1871	* Saves the entire guest state from the VMCB into the
1872	* guest-CPU context. Currently there is no residual state left in the CPU that
1873	* is not updated in the VMCB.
1874	*
1875	* @returns VBox status code.
1876	* @param pVCpu Pointer to the VMCPU.
1877	* @param pMixedCtx Pointer to the guest-CPU context. The data may be
1878	* out-of-sync. Make sure to update the required fields
1879	* before using them.
1880	*/
1881	static void hmR0SvmSaveGuestState(PVMCPU pVCpu, PCPUMCTX pMixedCtx)
1882	{
1883	Assert(VMMRZCallRing3IsEnabled(pVCpu));
1884
1885	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
1886
1887	pMixedCtx->rip = pVmcb->guest.u64RIP;
1888	pMixedCtx->rsp = pVmcb->guest.u64RSP;
1889	pMixedCtx->eflags.u32 = pVmcb->guest.u64RFlags;
1890	pMixedCtx->rax = pVmcb->guest.u64RAX;
1891
1892	/*
1893	* Guest interrupt shadow.
1894	*/
1895	if (pVmcb->ctrl.u64IntShadow & SVM_INTERRUPT_SHADOW_ACTIVE)
1896	EMSetInhibitInterruptsPC(pVCpu, pMixedCtx->rip);
1897	else if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS))
1898	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS);
1899
1900	/*
1901	* Guest Control registers: CR2, CR3 (handled at the end) - accesses to other control registers are always intercepted.
1902	*/
1903	pMixedCtx->cr2 = pVmcb->guest.u64CR2;
1904
1905	/*
1906	* Guest MSRs.
1907	*/
1908	pMixedCtx->msrSTAR = pVmcb->guest.u64STAR; /* legacy syscall eip, cs & ss */
1909	pMixedCtx->msrLSTAR = pVmcb->guest.u64LSTAR; /* 64-bit mode syscall rip */
1910	pMixedCtx->msrCSTAR = pVmcb->guest.u64CSTAR; /* compatibility mode syscall rip */
1911	pMixedCtx->msrSFMASK = pVmcb->guest.u64SFMASK; /* syscall flag mask */
1912	pMixedCtx->msrKERNELGSBASE = pVmcb->guest.u64KernelGSBase; /* swapgs exchange value */
1913	pMixedCtx->SysEnter.cs = pVmcb->guest.u64SysEnterCS;
1914	pMixedCtx->SysEnter.eip = pVmcb->guest.u64SysEnterEIP;
1915	pMixedCtx->SysEnter.esp = pVmcb->guest.u64SysEnterESP;
1916
1917	/*
1918	* Guest segment registers (includes FS, GS base MSRs for 64-bit guests).
1919	*/
1920	HMSVM_SAVE_SEG_REG(CS, cs);
1921	HMSVM_SAVE_SEG_REG(SS, ss);
1922	HMSVM_SAVE_SEG_REG(DS, ds);
1923	HMSVM_SAVE_SEG_REG(ES, es);
1924	HMSVM_SAVE_SEG_REG(FS, fs);
1925	HMSVM_SAVE_SEG_REG(GS, gs);
1926
1927	/*
1928	* Correct the hidden CS granularity bit. Haven't seen it being wrong in any other
1929	* register (yet).
1930	*/
1931	/** @todo SELM might need to be fixed as it too should not care about the
1932	* granularity bit. See @bugref{6785}. */
1933	if ( !pMixedCtx->cs.Attr.n.u1Granularity
1934	&& pMixedCtx->cs.Attr.n.u1Present
1935	&& pMixedCtx->cs.u32Limit > UINT32_C(0xfffff))
1936	{
1937	Assert((pMixedCtx->cs.u32Limit & 0xfff) == 0xfff);
1938	pMixedCtx->cs.Attr.n.u1Granularity = 1;
1939	}
1940
1941	#ifdef VBOX_STRICT
1942	# define HMSVM_ASSERT_SEG_GRANULARITY(reg) \
1943	AssertMsg( !pMixedCtx->reg.Attr.n.u1Present \
1944	\|\| ( pMixedCtx->reg.Attr.n.u1Granularity \
1945	? (pMixedCtx->reg.u32Limit & 0xfff) == 0xfff \
1946	: pMixedCtx->reg.u32Limit <= UINT32_C(0xfffff)), \
1947	("Invalid Segment Attributes Limit=%#RX32 Attr=%#RX32 Base=%#RX64\n", pMixedCtx->reg.u32Limit, \
1948	pMixedCtx->reg.Attr.u, pMixedCtx->reg.u64Base))
1949
1950	HMSVM_ASSERT_SEG_GRANULARITY(cs);
1951	HMSVM_ASSERT_SEG_GRANULARITY(ss);
1952	HMSVM_ASSERT_SEG_GRANULARITY(ds);
1953	HMSVM_ASSERT_SEG_GRANULARITY(es);
1954	HMSVM_ASSERT_SEG_GRANULARITY(fs);
1955	HMSVM_ASSERT_SEG_GRANULARITY(gs);
1956
1957	# undef HMSVM_ASSERT_SEL_GRANULARITY
1958	#endif
1959
1960	/*
1961	* Sync the hidden SS DPL field. AMD CPUs have a separate CPL field in the VMCB and uses that
1962	* and thus it's possible that when the CPL changes during guest execution that the SS DPL
1963	* isn't updated by AMD-V. Observed on some AMD Fusion CPUs with 64-bit guests.
1964	* See AMD spec. 15.5.1 "Basic operation".
1965	*/
1966	Assert(!(pVmcb->guest.u8CPL & ~0x3));
1967	pMixedCtx->ss.Attr.n.u2Dpl = pVmcb->guest.u8CPL & 0x3;
1968
1969	/*
1970	* Guest TR.
1971	* Fixup TR attributes so it's compatible with Intel. Important when saved-states are used
1972	* between Intel and AMD. See @bugref{6208} comment #39.
1973	*/
1974	HMSVM_SAVE_SEG_REG(TR, tr);
1975	if (CPUMIsGuestInLongModeEx(pMixedCtx))
1976	pMixedCtx->tr.Attr.n.u4Type = X86_SEL_TYPE_SYS_386_TSS_BUSY;
1977
1978	/*
1979	* Guest Descriptor-Table registers.
1980	*/
1981	HMSVM_SAVE_SEG_REG(LDTR, ldtr);
1982	pMixedCtx->gdtr.cbGdt = pVmcb->guest.GDTR.u32Limit;
1983	pMixedCtx->gdtr.pGdt = pVmcb->guest.GDTR.u64Base;
1984
1985	pMixedCtx->idtr.cbIdt = pVmcb->guest.IDTR.u32Limit;
1986	pMixedCtx->idtr.pIdt = pVmcb->guest.IDTR.u64Base;
1987
1988	/*
1989	* Guest Debug registers.
1990	*/
1991	if (!pVCpu->hm.s.fUsingHyperDR7)
1992	{
1993	pMixedCtx->dr[6] = pVmcb->guest.u64DR6;
1994	pMixedCtx->dr[7] = pVmcb->guest.u64DR7;
1995	}
1996	else
1997	{
1998	Assert(pVmcb->guest.u64DR7 == CPUMGetHyperDR7(pVCpu));
1999	CPUMSetHyperDR6(pVCpu, pVmcb->guest.u64DR6);
2000	}
2001
2002	/*
2003	* With Nested Paging, CR3 changes are not intercepted. Therefore, sync. it now.
2004	* This is done as the very last step of syncing the guest state, as PGMUpdateCR3() may cause longjmp's to ring-3.
2005	*/
2006	if ( pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging
2007	&& pMixedCtx->cr3 != pVmcb->guest.u64CR3)
2008	{
2009	CPUMSetGuestCR3(pVCpu, pVmcb->guest.u64CR3);
2010	PGMUpdateCR3(pVCpu, pVmcb->guest.u64CR3);
2011	}
2012	}
2013
2014
2015	/**
2016	* Does the necessary state syncing before returning to ring-3 for any reason
2017	* (longjmp, preemption, voluntary exits to ring-3) from AMD-V.
2018	*
2019	* @param pVM Pointer to the VM.
2020	* @param pVCpu Pointer to the VMCPU.
2021	* @param pMixedCtx Pointer to the guest-CPU context.
2022	*
2023	* @remarks No-long-jmp zone!!!
2024	*/
2025	static void hmR0SvmLeave(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
2026	{
2027	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2028	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2029	Assert(VMMR0IsLogFlushDisabled(pVCpu));
2030
2031	/*
2032	* !!! IMPORTANT !!!
2033	* If you modify code here, make sure to check whether hmR0SvmCallRing3Callback() needs to be updated too.
2034	*/
2035
2036	/* Restore host FPU state if necessary and resync on next R0 reentry .*/
2037	if (CPUMIsGuestFPUStateActive(pVCpu))
2038	{
2039	CPUMR0SaveGuestFPU(pVM, pVCpu, pCtx);
2040	Assert(!CPUMIsGuestFPUStateActive(pVCpu));
2041	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR0);
2042	}
2043
2044	/*
2045	* Restore host debug registers if necessary and resync on next R0 reentry.
2046	*/
2047	#ifdef VBOX_STRICT
2048	if (CPUMIsHyperDebugStateActive(pVCpu))
2049	{
2050	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2051	Assert(pVmcb->ctrl.u16InterceptRdDRx == 0xffff);
2052	Assert(pVmcb->ctrl.u16InterceptWrDRx == 0xffff);
2053	}
2054	#endif
2055	if (CPUMR0DebugStateMaybeSaveGuestAndRestoreHost(pVCpu, false /* save DR6 */))
2056	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_DEBUG);
2057
2058	Assert(!CPUMIsHyperDebugStateActive(pVCpu));
2059	Assert(!CPUMIsGuestDebugStateActive(pVCpu));
2060
2061	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatEntry);
2062	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatLoadGuestState);
2063	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatExit1);
2064	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatExit2);
2065	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchLongJmpToR3);
2066
2067	VMCPU_CMPXCHG_STATE(pVCpu, VMCPUSTATE_STARTED_HM, VMCPUSTATE_STARTED_EXEC);
2068	}
2069
2070
2071	/**
2072	* Leaves the AMD-V session.
2073	*
2074	* @returns VBox status code.
2075	* @param pVM Pointer to the VM.
2076	* @param pVCpu Pointer to the VMCPU.
2077	* @param pCtx Pointer to the guest-CPU context.
2078	*/
2079	static int hmR0SvmLeaveSession(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
2080	{
2081	HM_DISABLE_PREEMPT_IF_NEEDED();
2082	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2083	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2084
2085	/* When thread-context hooks are used, we can avoid doing the leave again if we had been preempted before
2086	and done this from the SVMR0ThreadCtxCallback(). */
2087	if (!pVCpu->hm.s.fLeaveDone)
2088	{
2089	hmR0SvmLeave(pVM, pVCpu, pCtx);
2090	pVCpu->hm.s.fLeaveDone = true;
2091	}
2092
2093	/*
2094	* !!! IMPORTANT !!!
2095	* If you modify code here, make sure to check whether hmR0SvmCallRing3Callback() needs to be updated too.
2096	*/
2097
2098	/* Deregister hook now that we've left HM context before re-enabling preemption. */
2099	if (VMMR0ThreadCtxHooksAreRegistered(pVCpu))
2100	VMMR0ThreadCtxHooksDeregister(pVCpu);
2101
2102	/* Leave HM context. This takes care of local init (term). */
2103	int rc = HMR0LeaveCpu(pVCpu);
2104
2105	HM_RESTORE_PREEMPT_IF_NEEDED();
2106	return rc;
2107	}
2108
2109
2110	/**
2111	* Does the necessary state syncing before doing a longjmp to ring-3.
2112	*
2113	* @returns VBox status code.
2114	* @param pVM Pointer to the VM.
2115	* @param pVCpu Pointer to the VMCPU.
2116	* @param pCtx Pointer to the guest-CPU context.
2117	*
2118	* @remarks No-long-jmp zone!!!
2119	*/
2120	static int hmR0SvmLongJmpToRing3(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
2121	{
2122	return hmR0SvmLeaveSession(pVM, pVCpu, pCtx);
2123	}
2124
2125
2126	/**
2127	* VMMRZCallRing3() callback wrapper which saves the guest state (or restores
2128	* any remaining host state) before we longjump to ring-3 and possibly get
2129	* preempted.
2130	*
2131	* @param pVCpu Pointer to the VMCPU.
2132	* @param enmOperation The operation causing the ring-3 longjump.
2133	* @param pvUser The user argument (pointer to the possibly
2134	* out-of-date guest-CPU context).
2135	*/
2136	DECLCALLBACK(int) hmR0SvmCallRing3Callback(PVMCPU pVCpu, VMMCALLRING3 enmOperation, void *pvUser)
2137	{
2138	if (enmOperation == VMMCALLRING3_VM_R0_ASSERTION)
2139	{
2140	/*
2141	* !!! IMPORTANT !!!
2142	* If you modify code here, make sure to check whether hmR0SvmLeave() and hmR0SvmLeaveSession() needs
2143	* to be updated too. This is a stripped down version which gets out ASAP trying to not trigger any assertion.
2144	*/
2145	VMMRZCallRing3RemoveNotification(pVCpu);
2146	VMMRZCallRing3Disable(pVCpu);
2147	HM_DISABLE_PREEMPT_IF_NEEDED();
2148
2149	/* Restore host FPU state if necessary and resync on next R0 reentry .*/
2150	if (CPUMIsGuestFPUStateActive(pVCpu))
2151	CPUMR0SaveGuestFPU(pVCpu->CTX_SUFF(pVM), pVCpu, (PCPUMCTX)pvUser);
2152
2153	/* Restore host debug registers if necessary and resync on next R0 reentry. */
2154	CPUMR0DebugStateMaybeSaveGuestAndRestoreHost(pVCpu, false /* save DR6 */);
2155
2156	/* Deregister hook now that we've left HM context before re-enabling preemption. */
2157	if (VMMR0ThreadCtxHooksAreRegistered(pVCpu))
2158	VMMR0ThreadCtxHooksDeregister(pVCpu);
2159
2160	/* Leave HM context. This takes care of local init (term). */
2161	HMR0LeaveCpu(pVCpu);
2162
2163	HM_RESTORE_PREEMPT_IF_NEEDED();
2164	return VINF_SUCCESS;
2165	}
2166
2167	Assert(pVCpu);
2168	Assert(pvUser);
2169	Assert(VMMRZCallRing3IsEnabled(pVCpu));
2170	HMSVM_ASSERT_PREEMPT_SAFE();
2171
2172	VMMRZCallRing3Disable(pVCpu);
2173	Assert(VMMR0IsLogFlushDisabled(pVCpu));
2174
2175	Log4(("hmR0SvmCallRing3Callback->hmR0SvmLongJmpToRing3\n"));
2176	int rc = hmR0SvmLongJmpToRing3(pVCpu->CTX_SUFF(pVM), pVCpu, (PCPUMCTX)pvUser);
2177	AssertRCReturn(rc, rc);
2178
2179	VMMRZCallRing3Enable(pVCpu);
2180	return VINF_SUCCESS;
2181	}
2182
2183
2184	/**
2185	* Take necessary actions before going back to ring-3.
2186	*
2187	* An action requires us to go back to ring-3. This function does the necessary
2188	* steps before we can safely return to ring-3. This is not the same as longjmps
2189	* to ring-3, this is voluntary.
2190	*
2191	* @param pVM Pointer to the VM.
2192	* @param pVCpu Pointer to the VMCPU.
2193	* @param pCtx Pointer to the guest-CPU context.
2194	* @param rcExit The reason for exiting to ring-3. Can be
2195	* VINF_VMM_UNKNOWN_RING3_CALL.
2196	*/
2197	static void hmR0SvmExitToRing3(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, int rcExit)
2198	{
2199	Assert(pVM);
2200	Assert(pVCpu);
2201	Assert(pCtx);
2202	HMSVM_ASSERT_PREEMPT_SAFE();
2203
2204	/* Please, no longjumps here (any logging shouldn't flush jump back to ring-3). NO LOGGING BEFORE THIS POINT! */
2205	VMMRZCallRing3Disable(pVCpu);
2206	Log4(("hmR0SvmExitToRing3: rcExit=%d\n", rcExit));
2207
2208	/* We need to do this only while truly exiting the "inner loop" back to ring-3 and -not- for any longjmp to ring3. */
2209	if (pVCpu->hm.s.Event.fPending)
2210	{
2211	hmR0SvmPendingEventToTrpmTrap(pVCpu);
2212	Assert(!pVCpu->hm.s.Event.fPending);
2213	}
2214
2215	/* If we're emulating an instruction, we shouldn't have any TRPM traps pending
2216	and if we're injecting an event we should have a TRPM trap pending. */
2217	Assert(rcExit != VINF_EM_RAW_INJECT_TRPM_EVENT \|\| TRPMHasTrap(pVCpu));
2218	Assert(rcExit != VINF_EM_RAW_EMULATE_INSTR \|\| !TRPMHasTrap(pVCpu));
2219
2220	/* Sync. the necessary state for going back to ring-3. */
2221	hmR0SvmLeaveSession(pVM, pVCpu, pCtx);
2222	STAM_COUNTER_DEC(&pVCpu->hm.s.StatSwitchLongJmpToR3);
2223
2224	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_TO_R3);
2225	CPUMSetChangedFlags(pVCpu, CPUM_CHANGED_SYSENTER_MSR
2226	\| CPUM_CHANGED_LDTR
2227	\| CPUM_CHANGED_GDTR
2228	\| CPUM_CHANGED_IDTR
2229	\| CPUM_CHANGED_TR
2230	\| CPUM_CHANGED_HIDDEN_SEL_REGS);
2231	if ( pVM->hm.s.fNestedPaging
2232	&& CPUMIsGuestPagingEnabledEx(pCtx))
2233	{
2234	CPUMSetChangedFlags(pVCpu, CPUM_CHANGED_GLOBAL_TLB_FLUSH);
2235	}
2236
2237	/* On our way back from ring-3 reload the guest state if there is a possibility of it being changed. */
2238	if (rcExit != VINF_EM_RAW_INTERRUPT)
2239	HMCPU_CF_SET(pVCpu, HM_CHANGED_ALL_GUEST);
2240
2241	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchExitToR3);
2242
2243	/* We do -not- want any longjmp notifications after this! We must return to ring-3 ASAP. */
2244	VMMRZCallRing3RemoveNotification(pVCpu);
2245	VMMRZCallRing3Enable(pVCpu);
2246	}
2247
2248
2249	/**
2250	* Updates the use of TSC offsetting mode for the CPU and adjusts the necessary
2251	* intercepts.
2252	*
2253	* @param pVCpu Pointer to the VMCPU.
2254	*
2255	* @remarks No-long-jump zone!!!
2256	*/
2257	static void hmR0SvmUpdateTscOffsetting(PVMCPU pVCpu)
2258	{
2259	bool fParavirtTsc;
2260	bool fCanUseRealTsc;
2261	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2262	fCanUseRealTsc = TMCpuTickCanUseRealTSC(pVCpu, &pVmcb->ctrl.u64TSCOffset, &fParavirtTsc);
2263	if (fCanUseRealTsc)
2264	{
2265	pVmcb->ctrl.u32InterceptCtrl1 &= ~SVM_CTRL1_INTERCEPT_RDTSC;
2266	pVmcb->ctrl.u32InterceptCtrl2 &= ~SVM_CTRL2_INTERCEPT_RDTSCP;
2267	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscOffset);
2268	}
2269	else
2270	{
2271	pVmcb->ctrl.u32InterceptCtrl1 \|= SVM_CTRL1_INTERCEPT_RDTSC;
2272	pVmcb->ctrl.u32InterceptCtrl2 \|= SVM_CTRL2_INTERCEPT_RDTSCP;
2273	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscIntercept);
2274	}
2275	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2276
2277	/** @todo later optimize this to be done elsewhere and not before every
2278	* VM-entry. */
2279	if (fParavirtTsc)
2280	{
2281	int rc = GIMR0UpdateParavirtTsc(pVCpu->CTX_SUFF(pVM), 0 /* u64Offset */);
2282	AssertRC(rc);
2283	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscParavirt);
2284	}
2285	}
2286
2287
2288	/**
2289	* Sets an event as a pending event to be injected into the guest.
2290	*
2291	* @param pVCpu Pointer to the VMCPU.
2292	* @param pEvent Pointer to the SVM event.
2293	* @param GCPtrFaultAddress The fault-address (CR2) in case it's a
2294	* page-fault.
2295	*
2296	* @remarks Statistics counter assumes this is a guest event being reflected to
2297	* the guest i.e. 'StatInjectPendingReflect' is incremented always.
2298	*/
2299	DECLINLINE(void) hmR0SvmSetPendingEvent(PVMCPU pVCpu, PSVMEVENT pEvent, RTGCUINTPTR GCPtrFaultAddress)
2300	{
2301	Assert(!pVCpu->hm.s.Event.fPending);
2302	Assert(pEvent->n.u1Valid);
2303
2304	pVCpu->hm.s.Event.u64IntInfo = pEvent->u;
2305	pVCpu->hm.s.Event.fPending = true;
2306	pVCpu->hm.s.Event.GCPtrFaultAddress = GCPtrFaultAddress;
2307
2308	Log4(("hmR0SvmSetPendingEvent: 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	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectPendingReflect);
2312	}
2313
2314
2315	/**
2316	* Injects an event into the guest upon VMRUN by updating the relevant field
2317	* in the VMCB.
2318	*
2319	* @param pVCpu Pointer to the VMCPU.
2320	* @param pVmcb Pointer to the guest VM control block.
2321	* @param pCtx Pointer to the guest-CPU context.
2322	* @param pEvent Pointer to the event.
2323	*
2324	* @remarks No-long-jump zone!!!
2325	* @remarks Requires CR0!
2326	*/
2327	DECLINLINE(void) hmR0SvmInjectEventVmcb(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx, PSVMEVENT pEvent)
2328	{
2329	NOREF(pVCpu); NOREF(pCtx);
2330
2331	pVmcb->ctrl.EventInject.u = pEvent->u;
2332	STAM_COUNTER_INC(&pVCpu->hm.s.paStatInjectedIrqsR0[pEvent->n.u8Vector & MASK_INJECT_IRQ_STAT]);
2333
2334	Log4(("hmR0SvmInjectEventVmcb: u=%#RX64 u8Vector=%#x Type=%#x ErrorCodeValid=%RTbool ErrorCode=%#RX32\n", pEvent->u,
2335	pEvent->n.u8Vector, (uint8_t)pEvent->n.u3Type, !!pEvent->n.u1ErrorCodeValid, pEvent->n.u32ErrorCode));
2336	}
2337
2338
2339
2340	/**
2341	* Converts any TRPM trap into a pending HM event. This is typically used when
2342	* entering from ring-3 (not longjmp returns).
2343	*
2344	* @param pVCpu Pointer to the VMCPU.
2345	*/
2346	static void hmR0SvmTrpmTrapToPendingEvent(PVMCPU pVCpu)
2347	{
2348	Assert(TRPMHasTrap(pVCpu));
2349	Assert(!pVCpu->hm.s.Event.fPending);
2350
2351	uint8_t uVector;
2352	TRPMEVENT enmTrpmEvent;
2353	RTGCUINT uErrCode;
2354	RTGCUINTPTR GCPtrFaultAddress;
2355	uint8_t cbInstr;
2356
2357	int rc = TRPMQueryTrapAll(pVCpu, &uVector, &enmTrpmEvent, &uErrCode, &GCPtrFaultAddress, &cbInstr);
2358	AssertRC(rc);
2359
2360	SVMEVENT Event;
2361	Event.u = 0;
2362	Event.n.u1Valid = 1;
2363	Event.n.u8Vector = uVector;
2364
2365	/* Refer AMD spec. 15.20 "Event Injection" for the format. */
2366	if (enmTrpmEvent == TRPM_TRAP)
2367	{
2368	Event.n.u3Type = SVM_EVENT_EXCEPTION;
2369	switch (uVector)
2370	{
2371	case X86_XCPT_NMI:
2372	{
2373	Event.n.u3Type = SVM_EVENT_NMI;
2374	break;
2375	}
2376
2377	case X86_XCPT_PF:
2378	case X86_XCPT_DF:
2379	case X86_XCPT_TS:
2380	case X86_XCPT_NP:
2381	case X86_XCPT_SS:
2382	case X86_XCPT_GP:
2383	case X86_XCPT_AC:
2384	{
2385	Event.n.u1ErrorCodeValid = 1;
2386	Event.n.u32ErrorCode = uErrCode;
2387	break;
2388	}
2389	}
2390	}
2391	else if (enmTrpmEvent == TRPM_HARDWARE_INT)
2392	Event.n.u3Type = SVM_EVENT_EXTERNAL_IRQ;
2393	else if (enmTrpmEvent == TRPM_SOFTWARE_INT)
2394	Event.n.u3Type = SVM_EVENT_SOFTWARE_INT;
2395	else
2396	AssertMsgFailed(("Invalid TRPM event type %d\n", enmTrpmEvent));
2397
2398	rc = TRPMResetTrap(pVCpu);
2399	AssertRC(rc);
2400
2401	Log4(("TRPM->HM event: u=%#RX64 u8Vector=%#x uErrorCodeValid=%RTbool uErrorCode=%#RX32\n", Event.u, Event.n.u8Vector,
2402	!!Event.n.u1ErrorCodeValid, Event.n.u32ErrorCode));
2403
2404	hmR0SvmSetPendingEvent(pVCpu, &Event, GCPtrFaultAddress);
2405	STAM_COUNTER_DEC(&pVCpu->hm.s.StatInjectPendingReflect);
2406	}
2407
2408
2409	/**
2410	* Converts any pending SVM event into a TRPM trap. Typically used when leaving
2411	* AMD-V to execute any instruction.
2412	*
2413	* @param pvCpu Pointer to the VMCPU.
2414	*/
2415	static void hmR0SvmPendingEventToTrpmTrap(PVMCPU pVCpu)
2416	{
2417	Assert(pVCpu->hm.s.Event.fPending);
2418	Assert(TRPMQueryTrap(pVCpu, NULL /* pu8TrapNo /, NULL / pEnmType */) == VERR_TRPM_NO_ACTIVE_TRAP);
2419
2420	SVMEVENT Event;
2421	Event.u = pVCpu->hm.s.Event.u64IntInfo;
2422
2423	uint8_t uVector = Event.n.u8Vector;
2424	uint8_t uVectorType = Event.n.u3Type;
2425
2426	TRPMEVENT enmTrapType;
2427	switch (uVectorType)
2428	{
2429	case SVM_EVENT_EXTERNAL_IRQ:
2430	enmTrapType = TRPM_HARDWARE_INT;
2431	break;
2432	case SVM_EVENT_SOFTWARE_INT:
2433	enmTrapType = TRPM_SOFTWARE_INT;
2434	break;
2435	case SVM_EVENT_EXCEPTION:
2436	case SVM_EVENT_NMI:
2437	enmTrapType = TRPM_TRAP;
2438	break;
2439	default:
2440	AssertMsgFailed(("Invalid pending-event type %#x\n", uVectorType));
2441	enmTrapType = TRPM_32BIT_HACK;
2442	break;
2443	}
2444
2445	Log4(("HM event->TRPM: uVector=%#x enmTrapType=%d\n", uVector, uVectorType));
2446
2447	int rc = TRPMAssertTrap(pVCpu, uVector, enmTrapType);
2448	AssertRC(rc);
2449
2450	if (Event.n.u1ErrorCodeValid)
2451	TRPMSetErrorCode(pVCpu, Event.n.u32ErrorCode);
2452
2453	if ( uVectorType == SVM_EVENT_EXCEPTION
2454	&& uVector == X86_XCPT_PF)
2455	{
2456	TRPMSetFaultAddress(pVCpu, pVCpu->hm.s.Event.GCPtrFaultAddress);
2457	Assert(pVCpu->hm.s.Event.GCPtrFaultAddress == CPUMGetGuestCR2(pVCpu));
2458	}
2459	else if (uVectorType == SVM_EVENT_SOFTWARE_INT)
2460	{
2461	AssertMsg( uVectorType == SVM_EVENT_SOFTWARE_INT
2462	\|\| (uVector == X86_XCPT_BP \|\| uVector == X86_XCPT_OF),
2463	("Invalid vector: uVector=%#x uVectorType=%#x\n", uVector, uVectorType));
2464	TRPMSetInstrLength(pVCpu, pVCpu->hm.s.Event.cbInstr);
2465	}
2466	pVCpu->hm.s.Event.fPending = false;
2467	}
2468
2469
2470	/**
2471	* Gets the guest's interrupt-shadow.
2472	*
2473	* @returns The guest's interrupt-shadow.
2474	* @param pVCpu Pointer to the VMCPU.
2475	* @param pCtx Pointer to the guest-CPU context.
2476	*
2477	* @remarks No-long-jump zone!!!
2478	* @remarks Has side-effects with VMCPU_FF_INHIBIT_INTERRUPTS force-flag.
2479	*/
2480	DECLINLINE(uint32_t) hmR0SvmGetGuestIntrShadow(PVMCPU pVCpu, PCPUMCTX pCtx)
2481	{
2482	/*
2483	* Instructions like STI and MOV SS inhibit interrupts till the next instruction completes. Check if we should
2484	* inhibit interrupts or clear any existing interrupt-inhibition.
2485	*/
2486	uint32_t uIntrState = 0;
2487	if (VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS))
2488	{
2489	if (pCtx->rip != EMGetInhibitInterruptsPC(pVCpu))
2490	{
2491	/*
2492	* We can clear the inhibit force flag as even if we go back to the recompiler without executing guest code in
2493	* AMD-V, the flag's condition to be cleared is met and thus the cleared state is correct.
2494	*/
2495	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS);
2496	}
2497	else
2498	uIntrState = SVM_INTERRUPT_SHADOW_ACTIVE;
2499	}
2500	return uIntrState;
2501	}
2502
2503
2504	/**
2505	* Sets the virtual interrupt intercept control in the VMCB which
2506	* instructs AMD-V to cause a #VMEXIT as soon as the guest is in a state to
2507	* receive interrupts.
2508	*
2509	* @param pVmcb Pointer to the VM control block.
2510	*/
2511	DECLINLINE(void) hmR0SvmSetVirtIntrIntercept(PSVMVMCB pVmcb)
2512	{
2513	if (!(pVmcb->ctrl.u32InterceptCtrl1 & SVM_CTRL1_INTERCEPT_VINTR))
2514	{
2515	pVmcb->ctrl.IntCtrl.n.u1VIrqValid = 1; /* A virtual interrupt is pending. */
2516	pVmcb->ctrl.IntCtrl.n.u8VIrqVector = 0; /* Not necessary as we #VMEXIT for delivering the interrupt. */
2517	pVmcb->ctrl.u32InterceptCtrl1 \|= SVM_CTRL1_INTERCEPT_VINTR;
2518	pVmcb->ctrl.u64VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS \| HMSVM_VMCB_CLEAN_TPR);
2519
2520	Log4(("Setting VINTR intercept\n"));
2521	}
2522	}
2523
2524
2525	/**
2526	* Sets the IRET intercept control in the VMCB which instructs AMD-V to cause a
2527	* #VMEXIT as soon as a guest starts executing an IRET. This is used to unblock
2528	* virtual NMIs.
2529	*
2530	* @param pVmcb Pointer to the VM control block.
2531	*/
2532	DECLINLINE(void) hmR0SvmSetIretIntercept(PSVMVMCB pVmcb)
2533	{
2534	if (!(pVmcb->ctrl.u32InterceptCtrl1 & SVM_CTRL1_INTERCEPT_IRET))
2535	{
2536	pVmcb->ctrl.u32InterceptCtrl1 \|= SVM_CTRL1_INTERCEPT_IRET;
2537	pVmcb->ctrl.u64VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS);
2538
2539	Log4(("Setting IRET intercept\n"));
2540	}
2541	}
2542
2543
2544	/**
2545	* Clears the IRET intercept control in the VMCB.
2546	*
2547	* @param pVmcb Pointer to the VM control block.
2548	*/
2549	DECLINLINE(void) hmR0SvmClearIretIntercept(PSVMVMCB pVmcb)
2550	{
2551	if (pVmcb->ctrl.u32InterceptCtrl1 & SVM_CTRL1_INTERCEPT_IRET)
2552	{
2553	pVmcb->ctrl.u32InterceptCtrl1 &= ~SVM_CTRL1_INTERCEPT_IRET;
2554	pVmcb->ctrl.u64VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS);
2555
2556	Log4(("Clearing IRET intercept\n"));
2557	}
2558	}
2559
2560
2561	/**
2562	* Evaluates the event to be delivered to the guest and sets it as the pending
2563	* event.
2564	*
2565	* @param pVCpu Pointer to the VMCPU.
2566	* @param pCtx Pointer to the guest-CPU context.
2567	*/
2568	static void hmR0SvmEvaluatePendingEvent(PVMCPU pVCpu, PCPUMCTX pCtx)
2569	{
2570	Assert(!pVCpu->hm.s.Event.fPending);
2571	Log4Func(("\n"));
2572
2573	bool const fIntShadow = RT_BOOL(hmR0SvmGetGuestIntrShadow(pVCpu, pCtx));
2574	bool const fBlockInt = !(pCtx->eflags.u32 & X86_EFL_IF);
2575	bool const fBlockNmi = RT_BOOL(VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_BLOCK_NMIS));
2576	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2577
2578	SVMEVENT Event;
2579	Event.u = 0;
2580	/** @todo SMI. SMIs take priority over NMIs. */
2581	if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_NMI)) /* NMI. NMIs take priority over regular interrupts . */
2582	{
2583	if (fBlockNmi)
2584	hmR0SvmSetIretIntercept(pVmcb);
2585	else if (fIntShadow)
2586	hmR0SvmSetVirtIntrIntercept(pVmcb);
2587	else
2588	{
2589	Log4(("Pending NMI\n"));
2590
2591	Event.n.u1Valid = 1;
2592	Event.n.u8Vector = X86_XCPT_NMI;
2593	Event.n.u3Type = SVM_EVENT_NMI;
2594
2595	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
2596	hmR0SvmSetIretIntercept(pVmcb);
2597	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INTERRUPT_NMI);
2598	}
2599	}
2600	else if (VMCPU_FF_IS_PENDING(pVCpu, (VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC)))
2601	{
2602	/*
2603	* Check if the guest can receive external interrupts (PIC/APIC). Once we do PDMGetInterrupt() we -must- deliver
2604	* the interrupt ASAP. We must not execute any guest code until we inject the interrupt which is why it is
2605	* evaluated here and not set as pending, solely based on the force-flags.
2606	*/
2607	if ( !fBlockInt
2608	&& !fIntShadow)
2609	{
2610	uint8_t u8Interrupt;
2611	int rc = PDMGetInterrupt(pVCpu, &u8Interrupt);
2612	if (RT_SUCCESS(rc))
2613	{
2614	Log4(("Injecting external interrupt u8Interrupt=%#x\n", u8Interrupt));
2615
2616	Event.n.u1Valid = 1;
2617	Event.n.u8Vector = u8Interrupt;
2618	Event.n.u3Type = SVM_EVENT_EXTERNAL_IRQ;
2619
2620	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
2621	}
2622	else
2623	{
2624	/** @todo Does this actually happen? If not turn it into an assertion. */
2625	Assert(!VMCPU_FF_IS_PENDING(pVCpu, (VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC)));
2626	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchGuestIrq);
2627	}
2628	}
2629	else
2630	hmR0SvmSetVirtIntrIntercept(pVmcb);
2631	}
2632	}
2633
2634
2635	/**
2636	* Injects any pending events into the guest if the guest is in a state to
2637	* receive them.
2638	*
2639	* @param pVCpu Pointer to the VMCPU.
2640	* @param pCtx Pointer to the guest-CPU context.
2641	*/
2642	static void hmR0SvmInjectPendingEvent(PVMCPU pVCpu, PCPUMCTX pCtx)
2643	{
2644	Assert(!TRPMHasTrap(pVCpu));
2645	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2646
2647	bool const fIntShadow = RT_BOOL(hmR0SvmGetGuestIntrShadow(pVCpu, pCtx));
2648	bool const fBlockInt = !(pCtx->eflags.u32 & X86_EFL_IF);
2649	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2650
2651	if (pVCpu->hm.s.Event.fPending) /* First, inject any pending HM events. */
2652	{
2653	SVMEVENT Event;
2654	Event.u = pVCpu->hm.s.Event.u64IntInfo;
2655	Assert(Event.n.u1Valid);
2656	#ifdef VBOX_STRICT
2657	if (Event.n.u3Type == SVM_EVENT_EXTERNAL_IRQ)
2658	{
2659	Assert(!fBlockInt);
2660	Assert(!fIntShadow);
2661	}
2662	else if (Event.n.u3Type == SVM_EVENT_NMI)
2663	Assert(!fIntShadow);
2664	#endif
2665
2666	Log4(("Injecting pending HM event.\n"));
2667	hmR0SvmInjectEventVmcb(pVCpu, pVmcb, pCtx, &Event);
2668	pVCpu->hm.s.Event.fPending = false;
2669
2670	#ifdef VBOX_WITH_STATISTICS
2671	if (Event.n.u3Type == SVM_EVENT_EXTERNAL_IRQ)
2672	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectInterrupt);
2673	else
2674	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectXcpt);
2675	#endif
2676	}
2677
2678	/* Update the guest interrupt shadow in the VMCB. */
2679	pVmcb->ctrl.u64IntShadow = !!fIntShadow;
2680	NOREF(fBlockInt);
2681	}
2682
2683
2684	/**
2685	* Reports world-switch error and dumps some useful debug info.
2686	*
2687	* @param pVM Pointer to the VM.
2688	* @param pVCpu Pointer to the VMCPU.
2689	* @param rcVMRun The return code from VMRUN (or
2690	* VERR_SVM_INVALID_GUEST_STATE for invalid
2691	* guest-state).
2692	* @param pCtx Pointer to the guest-CPU context.
2693	*/
2694	static void hmR0SvmReportWorldSwitchError(PVM pVM, PVMCPU pVCpu, int rcVMRun, PCPUMCTX pCtx)
2695	{
2696	NOREF(pCtx);
2697	HMSVM_ASSERT_PREEMPT_SAFE();
2698	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2699
2700	if (rcVMRun == VERR_SVM_INVALID_GUEST_STATE)
2701	{
2702	HMDumpRegs(pVM, pVCpu, pCtx); NOREF(pVM);
2703	#ifdef VBOX_STRICT
2704	Log4(("ctrl.u64VmcbCleanBits %#RX64\n", pVmcb->ctrl.u64VmcbCleanBits));
2705	Log4(("ctrl.u16InterceptRdCRx %#x\n", pVmcb->ctrl.u16InterceptRdCRx));
2706	Log4(("ctrl.u16InterceptWrCRx %#x\n", pVmcb->ctrl.u16InterceptWrCRx));
2707	Log4(("ctrl.u16InterceptRdDRx %#x\n", pVmcb->ctrl.u16InterceptRdDRx));
2708	Log4(("ctrl.u16InterceptWrDRx %#x\n", pVmcb->ctrl.u16InterceptWrDRx));
2709	Log4(("ctrl.u32InterceptException %#x\n", pVmcb->ctrl.u32InterceptException));
2710	Log4(("ctrl.u32InterceptCtrl1 %#x\n", pVmcb->ctrl.u32InterceptCtrl1));
2711	Log4(("ctrl.u32InterceptCtrl2 %#x\n", pVmcb->ctrl.u32InterceptCtrl2));
2712	Log4(("ctrl.u64IOPMPhysAddr %#RX64\n", pVmcb->ctrl.u64IOPMPhysAddr));
2713	Log4(("ctrl.u64MSRPMPhysAddr %#RX64\n", pVmcb->ctrl.u64MSRPMPhysAddr));
2714	Log4(("ctrl.u64TSCOffset %#RX64\n", pVmcb->ctrl.u64TSCOffset));
2715
2716	Log4(("ctrl.TLBCtrl.u32ASID %#x\n", pVmcb->ctrl.TLBCtrl.n.u32ASID));
2717	Log4(("ctrl.TLBCtrl.u8TLBFlush %#x\n", pVmcb->ctrl.TLBCtrl.n.u8TLBFlush));
2718	Log4(("ctrl.TLBCtrl.u24Reserved %#x\n", pVmcb->ctrl.TLBCtrl.n.u24Reserved));
2719
2720	Log4(("ctrl.IntCtrl.u8VTPR %#x\n", pVmcb->ctrl.IntCtrl.n.u8VTPR));
2721	Log4(("ctrl.IntCtrl.u1VIrqValid %#x\n", pVmcb->ctrl.IntCtrl.n.u1VIrqValid));
2722	Log4(("ctrl.IntCtrl.u7Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u7Reserved));
2723	Log4(("ctrl.IntCtrl.u4VIrqPriority %#x\n", pVmcb->ctrl.IntCtrl.n.u4VIrqPriority));
2724	Log4(("ctrl.IntCtrl.u1IgnoreTPR %#x\n", pVmcb->ctrl.IntCtrl.n.u1IgnoreTPR));
2725	Log4(("ctrl.IntCtrl.u3Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u3Reserved));
2726	Log4(("ctrl.IntCtrl.u1VIrqMasking %#x\n", pVmcb->ctrl.IntCtrl.n.u1VIrqMasking));
2727	Log4(("ctrl.IntCtrl.u6Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u6Reserved));
2728	Log4(("ctrl.IntCtrl.u8VIrqVector %#x\n", pVmcb->ctrl.IntCtrl.n.u8VIrqVector));
2729	Log4(("ctrl.IntCtrl.u24Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u24Reserved));
2730
2731	Log4(("ctrl.u64IntShadow %#RX64\n", pVmcb->ctrl.u64IntShadow));
2732	Log4(("ctrl.u64ExitCode %#RX64\n", pVmcb->ctrl.u64ExitCode));
2733	Log4(("ctrl.u64ExitInfo1 %#RX64\n", pVmcb->ctrl.u64ExitInfo1));
2734	Log4(("ctrl.u64ExitInfo2 %#RX64\n", pVmcb->ctrl.u64ExitInfo2));
2735	Log4(("ctrl.ExitIntInfo.u8Vector %#x\n", pVmcb->ctrl.ExitIntInfo.n.u8Vector));
2736	Log4(("ctrl.ExitIntInfo.u3Type %#x\n", pVmcb->ctrl.ExitIntInfo.n.u3Type));
2737	Log4(("ctrl.ExitIntInfo.u1ErrorCodeValid %#x\n", pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid));
2738	Log4(("ctrl.ExitIntInfo.u19Reserved %#x\n", pVmcb->ctrl.ExitIntInfo.n.u19Reserved));
2739	Log4(("ctrl.ExitIntInfo.u1Valid %#x\n", pVmcb->ctrl.ExitIntInfo.n.u1Valid));
2740	Log4(("ctrl.ExitIntInfo.u32ErrorCode %#x\n", pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode));
2741	Log4(("ctrl.NestedPaging %#RX64\n", pVmcb->ctrl.NestedPaging.u));
2742	Log4(("ctrl.EventInject.u8Vector %#x\n", pVmcb->ctrl.EventInject.n.u8Vector));
2743	Log4(("ctrl.EventInject.u3Type %#x\n", pVmcb->ctrl.EventInject.n.u3Type));
2744	Log4(("ctrl.EventInject.u1ErrorCodeValid %#x\n", pVmcb->ctrl.EventInject.n.u1ErrorCodeValid));
2745	Log4(("ctrl.EventInject.u19Reserved %#x\n", pVmcb->ctrl.EventInject.n.u19Reserved));
2746	Log4(("ctrl.EventInject.u1Valid %#x\n", pVmcb->ctrl.EventInject.n.u1Valid));
2747	Log4(("ctrl.EventInject.u32ErrorCode %#x\n", pVmcb->ctrl.EventInject.n.u32ErrorCode));
2748
2749	Log4(("ctrl.u64NestedPagingCR3 %#RX64\n", pVmcb->ctrl.u64NestedPagingCR3));
2750	Log4(("ctrl.u64LBRVirt %#RX64\n", pVmcb->ctrl.u64LBRVirt));
2751
2752	Log4(("guest.CS.u16Sel %RTsel\n", pVmcb->guest.CS.u16Sel));
2753	Log4(("guest.CS.u16Attr %#x\n", pVmcb->guest.CS.u16Attr));
2754	Log4(("guest.CS.u32Limit %#RX32\n", pVmcb->guest.CS.u32Limit));
2755	Log4(("guest.CS.u64Base %#RX64\n", pVmcb->guest.CS.u64Base));
2756	Log4(("guest.DS.u16Sel %#RTsel\n", pVmcb->guest.DS.u16Sel));
2757	Log4(("guest.DS.u16Attr %#x\n", pVmcb->guest.DS.u16Attr));
2758	Log4(("guest.DS.u32Limit %#RX32\n", pVmcb->guest.DS.u32Limit));
2759	Log4(("guest.DS.u64Base %#RX64\n", pVmcb->guest.DS.u64Base));
2760	Log4(("guest.ES.u16Sel %RTsel\n", pVmcb->guest.ES.u16Sel));
2761	Log4(("guest.ES.u16Attr %#x\n", pVmcb->guest.ES.u16Attr));
2762	Log4(("guest.ES.u32Limit %#RX32\n", pVmcb->guest.ES.u32Limit));
2763	Log4(("guest.ES.u64Base %#RX64\n", pVmcb->guest.ES.u64Base));
2764	Log4(("guest.FS.u16Sel %RTsel\n", pVmcb->guest.FS.u16Sel));
2765	Log4(("guest.FS.u16Attr %#x\n", pVmcb->guest.FS.u16Attr));
2766	Log4(("guest.FS.u32Limit %#RX32\n", pVmcb->guest.FS.u32Limit));
2767	Log4(("guest.FS.u64Base %#RX64\n", pVmcb->guest.FS.u64Base));
2768	Log4(("guest.GS.u16Sel %RTsel\n", pVmcb->guest.GS.u16Sel));
2769	Log4(("guest.GS.u16Attr %#x\n", pVmcb->guest.GS.u16Attr));
2770	Log4(("guest.GS.u32Limit %#RX32\n", pVmcb->guest.GS.u32Limit));
2771	Log4(("guest.GS.u64Base %#RX64\n", pVmcb->guest.GS.u64Base));
2772
2773	Log4(("guest.GDTR.u32Limit %#RX32\n", pVmcb->guest.GDTR.u32Limit));
2774	Log4(("guest.GDTR.u64Base %#RX64\n", pVmcb->guest.GDTR.u64Base));
2775
2776	Log4(("guest.LDTR.u16Sel %RTsel\n", pVmcb->guest.LDTR.u16Sel));
2777	Log4(("guest.LDTR.u16Attr %#x\n", pVmcb->guest.LDTR.u16Attr));
2778	Log4(("guest.LDTR.u32Limit %#RX32\n", pVmcb->guest.LDTR.u32Limit));
2779	Log4(("guest.LDTR.u64Base %#RX64\n", pVmcb->guest.LDTR.u64Base));
2780
2781	Log4(("guest.IDTR.u32Limit %#RX32\n", pVmcb->guest.IDTR.u32Limit));
2782	Log4(("guest.IDTR.u64Base %#RX64\n", pVmcb->guest.IDTR.u64Base));
2783
2784	Log4(("guest.TR.u16Sel %RTsel\n", pVmcb->guest.TR.u16Sel));
2785	Log4(("guest.TR.u16Attr %#x\n", pVmcb->guest.TR.u16Attr));
2786	Log4(("guest.TR.u32Limit %#RX32\n", pVmcb->guest.TR.u32Limit));
2787	Log4(("guest.TR.u64Base %#RX64\n", pVmcb->guest.TR.u64Base));
2788
2789	Log4(("guest.u8CPL %#x\n", pVmcb->guest.u8CPL));
2790	Log4(("guest.u64CR0 %#RX64\n", pVmcb->guest.u64CR0));
2791	Log4(("guest.u64CR2 %#RX64\n", pVmcb->guest.u64CR2));
2792	Log4(("guest.u64CR3 %#RX64\n", pVmcb->guest.u64CR3));
2793	Log4(("guest.u64CR4 %#RX64\n", pVmcb->guest.u64CR4));
2794	Log4(("guest.u64DR6 %#RX64\n", pVmcb->guest.u64DR6));
2795	Log4(("guest.u64DR7 %#RX64\n", pVmcb->guest.u64DR7));
2796
2797	Log4(("guest.u64RIP %#RX64\n", pVmcb->guest.u64RIP));
2798	Log4(("guest.u64RSP %#RX64\n", pVmcb->guest.u64RSP));
2799	Log4(("guest.u64RAX %#RX64\n", pVmcb->guest.u64RAX));
2800	Log4(("guest.u64RFlags %#RX64\n", pVmcb->guest.u64RFlags));
2801
2802	Log4(("guest.u64SysEnterCS %#RX64\n", pVmcb->guest.u64SysEnterCS));
2803	Log4(("guest.u64SysEnterEIP %#RX64\n", pVmcb->guest.u64SysEnterEIP));
2804	Log4(("guest.u64SysEnterESP %#RX64\n", pVmcb->guest.u64SysEnterESP));
2805
2806	Log4(("guest.u64EFER %#RX64\n", pVmcb->guest.u64EFER));
2807	Log4(("guest.u64STAR %#RX64\n", pVmcb->guest.u64STAR));
2808	Log4(("guest.u64LSTAR %#RX64\n", pVmcb->guest.u64LSTAR));
2809	Log4(("guest.u64CSTAR %#RX64\n", pVmcb->guest.u64CSTAR));
2810	Log4(("guest.u64SFMASK %#RX64\n", pVmcb->guest.u64SFMASK));
2811	Log4(("guest.u64KernelGSBase %#RX64\n", pVmcb->guest.u64KernelGSBase));
2812	Log4(("guest.u64GPAT %#RX64\n", pVmcb->guest.u64GPAT));
2813	Log4(("guest.u64DBGCTL %#RX64\n", pVmcb->guest.u64DBGCTL));
2814	Log4(("guest.u64BR_FROM %#RX64\n", pVmcb->guest.u64BR_FROM));
2815	Log4(("guest.u64BR_TO %#RX64\n", pVmcb->guest.u64BR_TO));
2816	Log4(("guest.u64LASTEXCPFROM %#RX64\n", pVmcb->guest.u64LASTEXCPFROM));
2817	Log4(("guest.u64LASTEXCPTO %#RX64\n", pVmcb->guest.u64LASTEXCPTO));
2818	#else
2819	NOREF(pVmcb);
2820	#endif /* VBOX_STRICT */
2821	}
2822	else
2823	Log4(("hmR0SvmReportWorldSwitchError: rcVMRun=%d\n", rcVMRun));
2824	}
2825
2826
2827	/**
2828	* Check per-VM and per-VCPU force flag actions that require us to go back to
2829	* ring-3 for one reason or another.
2830	*
2831	* @returns VBox status code (information status code included).
2832	* @retval VINF_SUCCESS if we don't have any actions that require going back to
2833	* ring-3.
2834	* @retval VINF_PGM_SYNC_CR3 if we have pending PGM CR3 sync.
2835	* @retval VINF_EM_PENDING_REQUEST if we have pending requests (like hardware
2836	* interrupts)
2837	* @retval VINF_PGM_POOL_FLUSH_PENDING if PGM is doing a pool flush and requires
2838	* all EMTs to be in ring-3.
2839	* @retval VINF_EM_RAW_TO_R3 if there is pending DMA requests.
2840	* @retval VINF_EM_NO_MEMORY PGM is out of memory, we need to return
2841	* to the EM loop.
2842	*
2843	* @param pVM Pointer to the VM.
2844	* @param pVCpu Pointer to the VMCPU.
2845	* @param pCtx Pointer to the guest-CPU context.
2846	*/
2847	static int hmR0SvmCheckForceFlags(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
2848	{
2849	Assert(VMMRZCallRing3IsEnabled(pVCpu));
2850
2851	/* On AMD-V we don't need to update CR3, PAE PDPES lazily. See hmR0SvmSaveGuestState(). */
2852	Assert(!VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_UPDATE_CR3));
2853	Assert(!VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_UPDATE_PAE_PDPES));
2854
2855	if ( VM_FF_IS_PENDING(pVM, !pVCpu->hm.s.fSingleInstruction
2856	? VM_FF_HP_R0_PRE_HM_MASK : VM_FF_HP_R0_PRE_HM_STEP_MASK)
2857	\|\| VMCPU_FF_IS_PENDING(pVCpu, !pVCpu->hm.s.fSingleInstruction
2858	? VMCPU_FF_HP_R0_PRE_HM_MASK : VMCPU_FF_HP_R0_PRE_HM_STEP_MASK) )
2859	{
2860	/* Pending PGM C3 sync. */
2861	if (VMCPU_FF_IS_PENDING(pVCpu,VMCPU_FF_PGM_SYNC_CR3 \| VMCPU_FF_PGM_SYNC_CR3_NON_GLOBAL))
2862	{
2863	int rc = PGMSyncCR3(pVCpu, pCtx->cr0, pCtx->cr3, pCtx->cr4, VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_PGM_SYNC_CR3));
2864	if (rc != VINF_SUCCESS)
2865	{
2866	Log4(("hmR0SvmCheckForceFlags: PGMSyncCR3 forcing us back to ring-3. rc=%d\n", rc));
2867	return rc;
2868	}
2869	}
2870
2871	/* Pending HM-to-R3 operations (critsects, timers, EMT rendezvous etc.) */
2872	/* -XXX- what was that about single stepping? */
2873	if ( VM_FF_IS_PENDING(pVM, VM_FF_HM_TO_R3_MASK)
2874	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_TO_R3_MASK))
2875	{
2876	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHmToR3FF);
2877	int rc = RT_UNLIKELY(VM_FF_IS_PENDING(pVM, VM_FF_PGM_NO_MEMORY)) ? VINF_EM_NO_MEMORY : VINF_EM_RAW_TO_R3;
2878	Log4(("hmR0SvmCheckForceFlags: HM_TO_R3 forcing us back to ring-3. rc=%d\n", rc));
2879	return rc;
2880	}
2881
2882	/* Pending VM request packets, such as hardware interrupts. */
2883	if ( VM_FF_IS_PENDING(pVM, VM_FF_REQUEST)
2884	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_REQUEST))
2885	{
2886	Log4(("hmR0SvmCheckForceFlags: Pending VM request forcing us back to ring-3\n"));
2887	return VINF_EM_PENDING_REQUEST;
2888	}
2889
2890	/* Pending PGM pool flushes. */
2891	if (VM_FF_IS_PENDING(pVM, VM_FF_PGM_POOL_FLUSH_PENDING))
2892	{
2893	Log4(("hmR0SvmCheckForceFlags: PGM pool flush pending forcing us back to ring-3\n"));
2894	return VINF_PGM_POOL_FLUSH_PENDING;
2895	}
2896
2897	/* Pending DMA requests. */
2898	if (VM_FF_IS_PENDING(pVM, VM_FF_PDM_DMA))
2899	{
2900	Log4(("hmR0SvmCheckForceFlags: Pending DMA request forcing us back to ring-3\n"));
2901	return VINF_EM_RAW_TO_R3;
2902	}
2903	}
2904
2905	return VINF_SUCCESS;
2906	}
2907
2908
2909	/**
2910	* Does the preparations before executing guest code in AMD-V.
2911	*
2912	* This may cause longjmps to ring-3 and may even result in rescheduling to the
2913	* recompiler. We must be cautious what we do here regarding committing
2914	* guest-state information into the the VMCB assuming we assuredly execute the
2915	* guest in AMD-V. If we fall back to the recompiler after updating the VMCB and
2916	* clearing the common-state (TRPM/forceflags), we must undo those changes so
2917	* that the recompiler can (and should) use them when it resumes guest
2918	* execution. Otherwise such operations must be done when we can no longer
2919	* exit to ring-3.
2920	*
2921	* @returns VBox status code (informational status codes included).
2922	* @retval VINF_SUCCESS if we can proceed with running the guest.
2923	* @retval VINF_* scheduling changes, we have to go back to ring-3.
2924	*
2925	* @param pVM Pointer to the VM.
2926	* @param pVCpu Pointer to the VMCPU.
2927	* @param pCtx Pointer to the guest-CPU context.
2928	* @param pSvmTransient Pointer to the SVM transient structure.
2929	*/
2930	static int hmR0SvmPreRunGuest(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
2931	{
2932	HMSVM_ASSERT_PREEMPT_SAFE();
2933
2934	/* Check force flag actions that might require us to go back to ring-3. */
2935	int rc = hmR0SvmCheckForceFlags(pVM, pVCpu, pCtx);
2936	if (rc != VINF_SUCCESS)
2937	return rc;
2938
2939	if (TRPMHasTrap(pVCpu))
2940	hmR0SvmTrpmTrapToPendingEvent(pVCpu);
2941	else if (!pVCpu->hm.s.Event.fPending)
2942	hmR0SvmEvaluatePendingEvent(pVCpu, pCtx);
2943
2944	#ifdef HMSVM_SYNC_FULL_GUEST_STATE
2945	HMCPU_CF_SET(pVCpu, HM_CHANGED_ALL_GUEST);
2946	#endif
2947
2948	/* Load the guest bits that are not shared with the host in any way since we can longjmp or get preempted. */
2949	rc = hmR0SvmLoadGuestState(pVM, pVCpu, pCtx);
2950	AssertRCReturn(rc, rc);
2951	STAM_COUNTER_INC(&pVCpu->hm.s.StatLoadFull);
2952
2953	/*
2954	* If we're not intercepting TPR changes in the guest, save the guest TPR before the world-switch
2955	* so we can update it on the way back if the guest changed the TPR.
2956	*/
2957	if (pVCpu->hm.s.svm.fSyncVTpr)
2958	{
2959	if (pVM->hm.s.fTPRPatchingActive)
2960	pSvmTransient->u8GuestTpr = pCtx->msrLSTAR;
2961	else
2962	{
2963	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2964	pSvmTransient->u8GuestTpr = pVmcb->ctrl.IntCtrl.n.u8VTPR;
2965	}
2966	}
2967
2968	/*
2969	* No longjmps to ring-3 from this point on!!!
2970	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional, better than a kernel panic.
2971	* This also disables flushing of the R0-logger instance (if any).
2972	*/
2973	VMMRZCallRing3Disable(pVCpu);
2974
2975	/*
2976	* We disable interrupts so that we don't miss any interrupts that would flag preemption (IPI/timers etc.)
2977	* when thread-context hooks aren't used and we've been running with preemption disabled for a while.
2978	*
2979	* We need to check for force-flags that could've possible been altered since we last checked them (e.g.
2980	* by PDMGetInterrupt() leaving the PDM critical section, see @bugref{6398}).
2981	*
2982	* We also check a couple of other force-flags as a last opportunity to get the EMT back to ring-3 before
2983	* executing guest code.
2984	*/
2985	pSvmTransient->uEflags = ASMIntDisableFlags();
2986	if ( VM_FF_IS_PENDING(pVM, VM_FF_EMT_RENDEZVOUS \| VM_FF_TM_VIRTUAL_SYNC)
2987	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_TO_R3_MASK))
2988	{
2989	ASMSetFlags(pSvmTransient->uEflags);
2990	VMMRZCallRing3Enable(pVCpu);
2991	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHmToR3FF);
2992	return VINF_EM_RAW_TO_R3;
2993	}
2994	if (RTThreadPreemptIsPending(NIL_RTTHREAD))
2995	{
2996	ASMSetFlags(pSvmTransient->uEflags);
2997	VMMRZCallRing3Enable(pVCpu);
2998	STAM_COUNTER_INC(&pVCpu->hm.s.StatPendingHostIrq);
2999	return VINF_EM_RAW_INTERRUPT;
3000	}
3001
3002	/*
3003	* If we are injecting an NMI, we must set VMCPU_FF_BLOCK_NMIS only when we are going to execute
3004	* guest code for certain (no exits to ring-3). Otherwise, we could re-read the flag on re-entry into
3005	* AMD-V and conclude that NMI inhibition is active when we have not even delivered the NMI.
3006	*
3007	* With VT-x, this is handled by the Guest interruptibility information VMCS field which will set the
3008	* VMCS field after actually delivering the NMI which we read on VM-exit to determine the state.
3009	*/
3010	if (pVCpu->hm.s.Event.fPending)
3011	{
3012	SVMEVENT Event;
3013	Event.u = pVCpu->hm.s.Event.u64IntInfo;
3014	if ( Event.n.u1Valid
3015	&& Event.n.u3Type == SVM_EVENT_NMI
3016	&& Event.n.u8Vector == X86_XCPT_NMI
3017	&& !VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_BLOCK_NMIS))
3018	{
3019	VMCPU_FF_SET(pVCpu, VMCPU_FF_BLOCK_NMIS);
3020	}
3021	}
3022
3023	return VINF_SUCCESS;
3024	}
3025
3026
3027	/**
3028	* Prepares to run guest code in AMD-V and we've committed to doing so. This
3029	* means there is no backing out to ring-3 or anywhere else at this
3030	* point.
3031	*
3032	* @param pVM Pointer to the VM.
3033	* @param pVCpu Pointer to the VMCPU.
3034	* @param pCtx Pointer to the guest-CPU context.
3035	* @param pSvmTransient Pointer to the SVM transient structure.
3036	*
3037	* @remarks Called with preemption disabled.
3038	* @remarks No-long-jump zone!!!
3039	*/
3040	static void hmR0SvmPreRunGuestCommitted(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
3041	{
3042	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
3043	Assert(VMMR0IsLogFlushDisabled(pVCpu));
3044	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
3045
3046	VMCPU_ASSERT_STATE(pVCpu, VMCPUSTATE_STARTED_HM);
3047	VMCPU_SET_STATE(pVCpu, VMCPUSTATE_STARTED_EXEC); /* Indicate the start of guest execution. */
3048
3049	hmR0SvmInjectPendingEvent(pVCpu, pCtx);
3050
3051	if ( pVCpu->hm.s.fUseGuestFpu
3052	&& !CPUMIsGuestFPUStateActive(pVCpu))
3053	{
3054	CPUMR0LoadGuestFPU(pVM, pVCpu, pCtx);
3055	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR0);
3056	}
3057
3058	/* Load the state shared between host and guest (FPU, debug). */
3059	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
3060	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_HOST_GUEST_SHARED_STATE))
3061	hmR0SvmLoadSharedState(pVCpu, pVmcb, pCtx);
3062	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_HOST_CONTEXT); /* Preemption might set this, nothing to do on AMD-V. */
3063	AssertMsg(!HMCPU_CF_VALUE(pVCpu), ("fContextUseFlags=%#RX32\n", HMCPU_CF_VALUE(pVCpu)));
3064
3065	/* Setup TSC offsetting. */
3066	RTCPUID idCurrentCpu = HMR0GetCurrentCpu()->idCpu;
3067	if ( pSvmTransient->fUpdateTscOffsetting
3068	\|\| idCurrentCpu != pVCpu->hm.s.idLastCpu)
3069	{
3070	hmR0SvmUpdateTscOffsetting(pVCpu);
3071	pSvmTransient->fUpdateTscOffsetting = false;
3072	}
3073
3074	/* If we've migrating CPUs, mark the VMCB Clean bits as dirty. */
3075	if (idCurrentCpu != pVCpu->hm.s.idLastCpu)
3076	pVmcb->ctrl.u64VmcbCleanBits = 0;
3077
3078	/* Store status of the shared guest-host state at the time of VMRUN. */
3079	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS) && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
3080	if (CPUMIsGuestInLongModeEx(pCtx))
3081	{
3082	pSvmTransient->fWasGuestDebugStateActive = CPUMIsGuestDebugStateActivePending(pVCpu);
3083	pSvmTransient->fWasHyperDebugStateActive = CPUMIsHyperDebugStateActivePending(pVCpu);
3084	}
3085	else
3086	#endif
3087	{
3088	pSvmTransient->fWasGuestDebugStateActive = CPUMIsGuestDebugStateActive(pVCpu);
3089	pSvmTransient->fWasHyperDebugStateActive = CPUMIsHyperDebugStateActive(pVCpu);
3090	}
3091	pSvmTransient->fWasGuestFPUStateActive = CPUMIsGuestFPUStateActive(pVCpu);
3092
3093	/* Flush the appropriate tagged-TLB entries. */
3094	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, true); /* Used for TLB-shootdowns, set this across the world switch. */
3095	hmR0SvmFlushTaggedTlb(pVCpu);
3096	Assert(HMR0GetCurrentCpu()->idCpu == pVCpu->hm.s.idLastCpu);
3097
3098	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatEntry, &pVCpu->hm.s.StatInGC, x);
3099
3100	TMNotifyStartOfExecution(pVCpu); /* Finally, notify TM to resume its clocks as we're about
3101	to start executing. */
3102
3103	/*
3104	* Save the current Host TSC_AUX and write the guest TSC_AUX to the host, so that
3105	* RDTSCPs (that don't cause exits) reads the guest MSR. See @bugref{3324}.
3106	*
3107	* This should be done -after- any RDTSCPs for obtaining the host timestamp (TM, STAM etc).
3108	*/
3109	if ( (pVM->hm.s.cpuid.u32AMDFeatureEDX & X86_CPUID_EXT_FEATURE_EDX_RDTSCP)
3110	&& !(pVmcb->ctrl.u32InterceptCtrl2 & SVM_CTRL2_INTERCEPT_RDTSCP))
3111	{
3112	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_TSC_AUX, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
3113	pVCpu->hm.s.u64HostTscAux = ASMRdMsr(MSR_K8_TSC_AUX);
3114	uint64_t u64GuestTscAux = CPUMR0GetGuestTscAux(pVCpu);
3115	if (u64GuestTscAux != pVCpu->hm.s.u64HostTscAux)
3116	ASMWrMsr(MSR_K8_TSC_AUX, u64GuestTscAux);
3117	pSvmTransient->fRestoreTscAuxMsr = true;
3118	}
3119	else
3120	{
3121	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_TSC_AUX, SVMMSREXIT_INTERCEPT_READ, SVMMSREXIT_INTERCEPT_WRITE);
3122	pSvmTransient->fRestoreTscAuxMsr = false;
3123	}
3124
3125	/* If VMCB Clean bits isn't supported by the CPU, simply mark all state-bits as dirty, indicating (re)load-from-VMCB. */
3126	if (!(pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_VMCB_CLEAN))
3127	pVmcb->ctrl.u64VmcbCleanBits = 0;
3128	}
3129
3130
3131	/**
3132	* Wrapper for running the guest code in AMD-V.
3133	*
3134	* @returns VBox strict status code.
3135	* @param pVM Pointer to the VM.
3136	* @param pVCpu Pointer to the VMCPU.
3137	* @param pCtx Pointer to the guest-CPU context.
3138	*
3139	* @remarks No-long-jump zone!!!
3140	*/
3141	DECLINLINE(int) hmR0SvmRunGuest(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3142	{
3143	/*
3144	* 64-bit Windows uses XMM registers in the kernel as the Microsoft compiler expresses floating-point operations
3145	* using SSE instructions. Some XMM registers (XMM6-XMM15) are callee-saved and thus the need for this XMM wrapper.
3146	* Refer MSDN docs. "Configuring Programs for 64-bit / x64 Software Conventions / Register Usage" for details.
3147	*/
3148	#ifdef VBOX_WITH_KERNEL_USING_XMM
3149	return HMR0SVMRunWrapXMM(pVCpu->hm.s.svm.HCPhysVmcbHost, pVCpu->hm.s.svm.HCPhysVmcb, pCtx, pVM, pVCpu,
3150	pVCpu->hm.s.svm.pfnVMRun);
3151	#else
3152	return pVCpu->hm.s.svm.pfnVMRun(pVCpu->hm.s.svm.HCPhysVmcbHost, pVCpu->hm.s.svm.HCPhysVmcb, pCtx, pVM, pVCpu);
3153	#endif
3154	}
3155
3156
3157	/**
3158	* Performs some essential restoration of state after running guest code in
3159	* AMD-V.
3160	*
3161	* @param pVM Pointer to the VM.
3162	* @param pVCpu Pointer to the VMCPU.
3163	* @param pMixedCtx Pointer to the guest-CPU context. The data maybe
3164	* out-of-sync. Make sure to update the required fields
3165	* before using them.
3166	* @param pSvmTransient Pointer to the SVM transient structure.
3167	* @param rcVMRun Return code of VMRUN.
3168	*
3169	* @remarks Called with interrupts disabled.
3170	* @remarks No-long-jump zone!!! This function will however re-enable longjmps
3171	* unconditionally when it is safe to do so.
3172	*/
3173	static void hmR0SvmPostRunGuest(PVM pVM, PVMCPU pVCpu, PCPUMCTX pMixedCtx, PSVMTRANSIENT pSvmTransient, int rcVMRun)
3174	{
3175	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
3176
3177	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, false); /* See HMInvalidatePageOnAllVCpus(): used for TLB-shootdowns. */
3178	ASMAtomicIncU32(&pVCpu->hm.s.cWorldSwitchExits); /* Initialized in vmR3CreateUVM(): used for TLB-shootdowns. */
3179
3180	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
3181	pVmcb->ctrl.u64VmcbCleanBits = HMSVM_VMCB_CLEAN_ALL; /* Mark the VMCB-state cache as unmodified by VMM. */
3182
3183	if (pSvmTransient->fRestoreTscAuxMsr)
3184	{
3185	uint64_t u64GuestTscAuxMsr = ASMRdMsr(MSR_K8_TSC_AUX);
3186	CPUMR0SetGuestTscAux(pVCpu, u64GuestTscAuxMsr);
3187	if (u64GuestTscAuxMsr != pVCpu->hm.s.u64HostTscAux)
3188	ASMWrMsr(MSR_K8_TSC_AUX, pVCpu->hm.s.u64HostTscAux);
3189	}
3190
3191	if (!(pVmcb->ctrl.u32InterceptCtrl1 & SVM_CTRL1_INTERCEPT_RDTSC))
3192	TMCpuTickSetLastSeen(pVCpu, ASMReadTSC() + pVmcb->ctrl.u64TSCOffset); /** @todo use SUPReadTSC() eventually. */
3193
3194	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatInGC, &pVCpu->hm.s.StatExit1, x);
3195	TMNotifyEndOfExecution(pVCpu); /* Notify TM that the guest is no longer running. */
3196	VMCPU_SET_STATE(pVCpu, VMCPUSTATE_STARTED_HM);
3197
3198	Assert(!(ASMGetFlags() & X86_EFL_IF));
3199	ASMSetFlags(pSvmTransient->uEflags); /* Enable interrupts. */
3200	VMMRZCallRing3Enable(pVCpu); /* It is now safe to do longjmps to ring-3!!! */
3201
3202	/* If VMRUN failed, we can bail out early. This does -not- cover SVM_EXIT_INVALID. */
3203	if (RT_UNLIKELY(rcVMRun != VINF_SUCCESS))
3204	{
3205	Log4(("VMRUN failure: rcVMRun=%Rrc\n", rcVMRun));
3206	return;
3207	}
3208
3209	pSvmTransient->u64ExitCode = pVmcb->ctrl.u64ExitCode; /* Save the #VMEXIT reason. */
3210	HMCPU_EXIT_HISTORY_ADD(pVCpu, pVmcb->ctrl.u64ExitCode); /* Update the #VMEXIT history array. */
3211	pSvmTransient->fVectoringDoublePF = false; /* Vectoring double page-fault needs to be determined later. */
3212	pSvmTransient->fVectoringPF = false; /* Vectoring page-fault needs to be determined later. */
3213
3214	hmR0SvmSaveGuestState(pVCpu, pMixedCtx); /* Save the guest state from the VMCB to the guest-CPU context. */
3215
3216	if (RT_LIKELY(pSvmTransient->u64ExitCode != (uint64_t)SVM_EXIT_INVALID))
3217	{
3218	if (pVCpu->hm.s.svm.fSyncVTpr)
3219	{
3220	/* TPR patching (for 32-bit guests) uses LSTAR MSR for holding the TPR value, otherwise uses the VTPR. */
3221	if ( pVM->hm.s.fTPRPatchingActive
3222	&& (pMixedCtx->msrLSTAR & 0xff) != pSvmTransient->u8GuestTpr)
3223	{
3224	int rc = PDMApicSetTPR(pVCpu, pMixedCtx->msrLSTAR & 0xff);
3225	AssertRC(rc);
3226	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
3227	}
3228	else if (pSvmTransient->u8GuestTpr != pVmcb->ctrl.IntCtrl.n.u8VTPR)
3229	{
3230	int rc = PDMApicSetTPR(pVCpu, pVmcb->ctrl.IntCtrl.n.u8VTPR << 4);
3231	AssertRC(rc);
3232	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
3233	}
3234	}
3235	}
3236	}
3237
3238
3239	/**
3240	* Runs the guest code using AMD-V.
3241	*
3242	* @returns VBox status code.
3243	* @param pVM Pointer to the VM.
3244	* @param pVCpu Pointer to the VMCPU.
3245	*/
3246	static int hmR0SvmRunGuestCodeNormal(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3247	{
3248	SVMTRANSIENT SvmTransient;
3249	SvmTransient.fUpdateTscOffsetting = true;
3250	uint32_t cLoops = 0;
3251	int rc = VERR_INTERNAL_ERROR_5;
3252
3253	for (;; cLoops++)
3254	{
3255	Assert(!HMR0SuspendPending());
3256	HMSVM_ASSERT_CPU_SAFE();
3257
3258	/* Preparatory work for running guest code, this may force us to return
3259	to ring-3. This bugger disables interrupts on VINF_SUCCESS! */
3260	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatEntry, x);
3261	rc = hmR0SvmPreRunGuest(pVM, pVCpu, pCtx, &SvmTransient);
3262	if (rc != VINF_SUCCESS)
3263	break;
3264
3265	/*
3266	* No longjmps to ring-3 from this point on!!!
3267	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional, better than a kernel panic.
3268	* This also disables flushing of the R0-logger instance (if any).
3269	*/
3270	hmR0SvmPreRunGuestCommitted(pVM, pVCpu, pCtx, &SvmTransient);
3271	rc = hmR0SvmRunGuest(pVM, pVCpu, pCtx);
3272
3273	/* Restore any residual host-state and save any bits shared between host
3274	and guest into the guest-CPU state. Re-enables interrupts! */
3275	hmR0SvmPostRunGuest(pVM, pVCpu, pCtx, &SvmTransient, rc);
3276
3277	if (RT_UNLIKELY( rc != VINF_SUCCESS /* Check for VMRUN errors. */
3278	\|\| SvmTransient.u64ExitCode == (uint64_t)SVM_EXIT_INVALID)) /* Check for invalid guest-state errors. */
3279	{
3280	if (rc == VINF_SUCCESS)
3281	rc = VERR_SVM_INVALID_GUEST_STATE;
3282	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit1, x);
3283	hmR0SvmReportWorldSwitchError(pVM, pVCpu, rc, pCtx);
3284	break;
3285	}
3286
3287	/* Handle the #VMEXIT. */
3288	HMSVM_EXITCODE_STAM_COUNTER_INC(SvmTransient.u64ExitCode);
3289	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatExit1, &pVCpu->hm.s.StatExit2, x);
3290	VBOXVMM_R0_HMSVM_VMEXIT(pVCpu, pCtx, SvmTransient.u64ExitCode, (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb);
3291	rc = hmR0SvmHandleExit(pVCpu, pCtx, &SvmTransient);
3292	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit2, x);
3293	if (rc != VINF_SUCCESS)
3294	break;
3295	if (cLoops > pVM->hm.s.cMaxResumeLoops)
3296	{
3297	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchMaxResumeLoops);
3298	rc = VINF_EM_RAW_INTERRUPT;
3299	break;
3300	}
3301	}
3302
3303	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatEntry, x);
3304	return rc;
3305	}
3306
3307
3308	/**
3309	* Runs the guest code using AMD-V in single step mode.
3310	*
3311	* @returns VBox status code.
3312	* @param pVM Pointer to the VM.
3313	* @param pVCpu Pointer to the VMCPU.
3314	* @param pCtx Pointer to the guest-CPU context.
3315	*/
3316	static int hmR0SvmRunGuestCodeStep(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3317	{
3318	SVMTRANSIENT SvmTransient;
3319	SvmTransient.fUpdateTscOffsetting = true;
3320	uint32_t cLoops = 0;
3321	int rc = VERR_INTERNAL_ERROR_5;
3322	uint16_t uCsStart = pCtx->cs.Sel;
3323	uint64_t uRipStart = pCtx->rip;
3324
3325	for (;; cLoops++)
3326	{
3327	Assert(!HMR0SuspendPending());
3328	AssertMsg(pVCpu->hm.s.idEnteredCpu == RTMpCpuId(),
3329	("Illegal migration! Entered on CPU %u Current %u cLoops=%u\n", (unsigned)pVCpu->hm.s.idEnteredCpu,
3330	(unsigned)RTMpCpuId(), cLoops));
3331
3332	/* Preparatory work for running guest code, this may force us to return
3333	to ring-3. This bugger disables interrupts on VINF_SUCCESS! */
3334	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatEntry, x);
3335	rc = hmR0SvmPreRunGuest(pVM, pVCpu, pCtx, &SvmTransient);
3336	if (rc != VINF_SUCCESS)
3337	break;
3338
3339	/*
3340	* No longjmps to ring-3 from this point on!!!
3341	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional, better than a kernel panic.
3342	* This also disables flushing of the R0-logger instance (if any).
3343	*/
3344	VMMRZCallRing3Disable(pVCpu);
3345	VMMRZCallRing3RemoveNotification(pVCpu);
3346	hmR0SvmPreRunGuestCommitted(pVM, pVCpu, pCtx, &SvmTransient);
3347
3348	rc = hmR0SvmRunGuest(pVM, pVCpu, pCtx);
3349
3350	/*
3351	* Restore any residual host-state and save any bits shared between host and guest into the guest-CPU state.
3352	* This will also re-enable longjmps to ring-3 when it has reached a safe point!!!
3353	*/
3354	hmR0SvmPostRunGuest(pVM, pVCpu, pCtx, &SvmTransient, rc);
3355	if (RT_UNLIKELY( rc != VINF_SUCCESS /* Check for VMRUN errors. */
3356	\|\| SvmTransient.u64ExitCode == (uint64_t)SVM_EXIT_INVALID)) /* Check for invalid guest-state errors. */
3357	{
3358	if (rc == VINF_SUCCESS)
3359	rc = VERR_SVM_INVALID_GUEST_STATE;
3360	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit1, x);
3361	hmR0SvmReportWorldSwitchError(pVM, pVCpu, rc, pCtx);
3362	return rc;
3363	}
3364
3365	/* Handle the #VMEXIT. */
3366	HMSVM_EXITCODE_STAM_COUNTER_INC(SvmTransient.u64ExitCode);
3367	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatExit1, &pVCpu->hm.s.StatExit2, x);
3368	VBOXVMM_R0_HMSVM_VMEXIT(pVCpu, pCtx, SvmTransient.u64ExitCode, (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb);
3369	rc = hmR0SvmHandleExit(pVCpu, pCtx, &SvmTransient);
3370	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit2, x);
3371	if (rc != VINF_SUCCESS)
3372	break;
3373	if (cLoops > pVM->hm.s.cMaxResumeLoops)
3374	{
3375	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchMaxResumeLoops);
3376	rc = VINF_EM_RAW_INTERRUPT;
3377	break;
3378	}
3379
3380	/*
3381	* Did the RIP change, if so, consider it a single step.
3382	* Otherwise, make sure one of the TFs gets set.
3383	*/
3384	if ( pCtx->rip != uRipStart
3385	\|\| pCtx->cs.Sel != uCsStart)
3386	{
3387	rc = VINF_EM_DBG_STEPPED;
3388	break;
3389	}
3390	pVCpu->hm.s.fContextUseFlags \|= HM_CHANGED_GUEST_DEBUG;
3391	}
3392
3393	/*
3394	* Clear the X86_EFL_TF if necessary.
3395	*/
3396	if (pVCpu->hm.s.fClearTrapFlag)
3397	{
3398	pVCpu->hm.s.fClearTrapFlag = false;
3399	pCtx->eflags.Bits.u1TF = 0;
3400	}
3401
3402	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatEntry, x);
3403	return rc;
3404	}
3405
3406
3407	/**
3408	* Runs the guest code using AMD-V.
3409	*
3410	* @returns VBox status code.
3411	* @param pVM Pointer to the VM.
3412	* @param pVCpu Pointer to the VMCPU.
3413	* @param pCtx Pointer to the guest-CPU context.
3414	*/
3415	VMMR0DECL(int) SVMR0RunGuestCode(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3416	{
3417	Assert(VMMRZCallRing3IsEnabled(pVCpu));
3418	HMSVM_ASSERT_PREEMPT_SAFE();
3419	VMMRZCallRing3SetNotification(pVCpu, hmR0SvmCallRing3Callback, pCtx);
3420
3421	int rc;
3422	if (!pVCpu->hm.s.fSingleInstruction && !DBGFIsStepping(pVCpu))
3423	rc = hmR0SvmRunGuestCodeNormal(pVM, pVCpu, pCtx);
3424	else
3425	rc = hmR0SvmRunGuestCodeStep(pVM, pVCpu, pCtx);
3426
3427	if (rc == VERR_EM_INTERPRETER)
3428	rc = VINF_EM_RAW_EMULATE_INSTR;
3429	else if (rc == VINF_EM_RESET)
3430	rc = VINF_EM_TRIPLE_FAULT;
3431
3432	/* Prepare to return to ring-3. This will remove longjmp notifications. */
3433	hmR0SvmExitToRing3(pVM, pVCpu, pCtx, rc);
3434	Assert(!VMMRZCallRing3IsNotificationSet(pVCpu));
3435	return rc;
3436	}
3437
3438
3439	/**
3440	* Handles a #VMEXIT (for all EXITCODE values except SVM_EXIT_INVALID).
3441	*
3442	* @returns VBox status code (informational status codes included).
3443	* @param pVCpu Pointer to the VMCPU.
3444	* @param pCtx Pointer to the guest-CPU context.
3445	* @param pSvmTransient Pointer to the SVM transient structure.
3446	*/
3447	DECLINLINE(int) hmR0SvmHandleExit(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
3448	{
3449	Assert(pSvmTransient->u64ExitCode != (uint64_t)SVM_EXIT_INVALID);
3450	Assert(pSvmTransient->u64ExitCode <= SVM_EXIT_MAX);
3451
3452	/*
3453	* The ordering of the case labels is based on most-frequently-occurring #VMEXITs for most guests under
3454	* normal workloads (for some definition of "normal").
3455	*/
3456	uint32_t u32ExitCode = pSvmTransient->u64ExitCode;
3457	switch (pSvmTransient->u64ExitCode)
3458	{
3459	case SVM_EXIT_NPF:
3460	return hmR0SvmExitNestedPF(pVCpu, pCtx, pSvmTransient);
3461
3462	case SVM_EXIT_IOIO:
3463	return hmR0SvmExitIOInstr(pVCpu, pCtx, pSvmTransient);
3464
3465	case SVM_EXIT_RDTSC:
3466	return hmR0SvmExitRdtsc(pVCpu, pCtx, pSvmTransient);
3467
3468	case SVM_EXIT_RDTSCP:
3469	return hmR0SvmExitRdtscp(pVCpu, pCtx, pSvmTransient);
3470
3471	case SVM_EXIT_CPUID:
3472	return hmR0SvmExitCpuid(pVCpu, pCtx, pSvmTransient);
3473
3474	case SVM_EXIT_EXCEPTION_E: /* X86_XCPT_PF */
3475	return hmR0SvmExitXcptPF(pVCpu, pCtx, pSvmTransient);
3476
3477	case SVM_EXIT_EXCEPTION_7: /* X86_XCPT_NM */
3478	return hmR0SvmExitXcptNM(pVCpu, pCtx, pSvmTransient);
3479
3480	case SVM_EXIT_EXCEPTION_10: /* X86_XCPT_MF */
3481	return hmR0SvmExitXcptMF(pVCpu, pCtx, pSvmTransient);
3482
3483	case SVM_EXIT_EXCEPTION_1: /* X86_XCPT_DB */
3484	return hmR0SvmExitXcptDB(pVCpu, pCtx, pSvmTransient);
3485
3486	case SVM_EXIT_MONITOR:
3487	return hmR0SvmExitMonitor(pVCpu, pCtx, pSvmTransient);
3488
3489	case SVM_EXIT_MWAIT:
3490	return hmR0SvmExitMwait(pVCpu, pCtx, pSvmTransient);
3491
3492	case SVM_EXIT_HLT:
3493	return hmR0SvmExitHlt(pVCpu, pCtx, pSvmTransient);
3494
3495	case SVM_EXIT_READ_CR0:
3496	case SVM_EXIT_READ_CR3:
3497	case SVM_EXIT_READ_CR4:
3498	return hmR0SvmExitReadCRx(pVCpu, pCtx, pSvmTransient);
3499
3500	case SVM_EXIT_WRITE_CR0:
3501	case SVM_EXIT_WRITE_CR3:
3502	case SVM_EXIT_WRITE_CR4:
3503	case SVM_EXIT_WRITE_CR8:
3504	return hmR0SvmExitWriteCRx(pVCpu, pCtx, pSvmTransient);
3505
3506	case SVM_EXIT_VINTR:
3507	return hmR0SvmExitVIntr(pVCpu, pCtx, pSvmTransient);
3508
3509	case SVM_EXIT_INTR:
3510	case SVM_EXIT_FERR_FREEZE:
3511	case SVM_EXIT_NMI:
3512	return hmR0SvmExitIntr(pVCpu, pCtx, pSvmTransient);
3513
3514	case SVM_EXIT_MSR:
3515	return hmR0SvmExitMsr(pVCpu, pCtx, pSvmTransient);
3516
3517	case SVM_EXIT_INVLPG:
3518	return hmR0SvmExitInvlpg(pVCpu, pCtx, pSvmTransient);
3519
3520	case SVM_EXIT_WBINVD:
3521	return hmR0SvmExitWbinvd(pVCpu, pCtx, pSvmTransient);
3522
3523	case SVM_EXIT_INVD:
3524	return hmR0SvmExitInvd(pVCpu, pCtx, pSvmTransient);
3525
3526	case SVM_EXIT_RDPMC:
3527	return hmR0SvmExitRdpmc(pVCpu, pCtx, pSvmTransient);
3528
3529	default:
3530	{
3531	switch (pSvmTransient->u64ExitCode)
3532	{
3533	case SVM_EXIT_READ_DR0: case SVM_EXIT_READ_DR1: case SVM_EXIT_READ_DR2: case SVM_EXIT_READ_DR3:
3534	case SVM_EXIT_READ_DR6: case SVM_EXIT_READ_DR7: case SVM_EXIT_READ_DR8: case SVM_EXIT_READ_DR9:
3535	case SVM_EXIT_READ_DR10: case SVM_EXIT_READ_DR11: case SVM_EXIT_READ_DR12: case SVM_EXIT_READ_DR13:
3536	case SVM_EXIT_READ_DR14: case SVM_EXIT_READ_DR15:
3537	return hmR0SvmExitReadDRx(pVCpu, pCtx, pSvmTransient);
3538
3539	case SVM_EXIT_WRITE_DR0: case SVM_EXIT_WRITE_DR1: case SVM_EXIT_WRITE_DR2: case SVM_EXIT_WRITE_DR3:
3540	case SVM_EXIT_WRITE_DR6: case SVM_EXIT_WRITE_DR7: case SVM_EXIT_WRITE_DR8: case SVM_EXIT_WRITE_DR9:
3541	case SVM_EXIT_WRITE_DR10: case SVM_EXIT_WRITE_DR11: case SVM_EXIT_WRITE_DR12: case SVM_EXIT_WRITE_DR13:
3542	case SVM_EXIT_WRITE_DR14: case SVM_EXIT_WRITE_DR15:
3543	return hmR0SvmExitWriteDRx(pVCpu, pCtx, pSvmTransient);
3544
3545	case SVM_EXIT_TASK_SWITCH:
3546	return hmR0SvmExitTaskSwitch(pVCpu, pCtx, pSvmTransient);
3547
3548	case SVM_EXIT_VMMCALL:
3549	return hmR0SvmExitVmmCall(pVCpu, pCtx, pSvmTransient);
3550
3551	case SVM_EXIT_IRET:
3552	return hmR0SvmExitIret(pVCpu, pCtx, pSvmTransient);
3553
3554	case SVM_EXIT_SHUTDOWN:
3555	return hmR0SvmExitShutdown(pVCpu, pCtx, pSvmTransient);
3556
3557	case SVM_EXIT_SMI:
3558	case SVM_EXIT_INIT:
3559	{
3560	/*
3561	* We don't intercept NMIs. As for INIT signals, it really shouldn't ever happen here. If it ever does,
3562	* we want to know about it so log the exit code and bail.
3563	*/
3564	AssertMsgFailed(("hmR0SvmHandleExit: Unexpected exit %#RX32\n", (uint32_t)pSvmTransient->u64ExitCode));
3565	pVCpu->hm.s.u32HMError = (uint32_t)pSvmTransient->u64ExitCode;
3566	return VERR_SVM_UNEXPECTED_EXIT;
3567	}
3568
3569	case SVM_EXIT_INVLPGA:
3570	case SVM_EXIT_RSM:
3571	case SVM_EXIT_VMRUN:
3572	case SVM_EXIT_VMLOAD:
3573	case SVM_EXIT_VMSAVE:
3574	case SVM_EXIT_STGI:
3575	case SVM_EXIT_CLGI:
3576	case SVM_EXIT_SKINIT:
3577	return hmR0SvmExitSetPendingXcptUD(pVCpu, pCtx, pSvmTransient);
3578
3579	#ifdef HMSVM_ALWAYS_TRAP_ALL_XCPTS
3580	case SVM_EXIT_EXCEPTION_0: /* X86_XCPT_DE */
3581	/* SVM_EXIT_EXCEPTION_1: / / X86_XCPT_DB - Handled above. */
3582	case SVM_EXIT_EXCEPTION_2: /* X86_XCPT_NMI */
3583	case SVM_EXIT_EXCEPTION_3: /* X86_XCPT_BP */
3584	case SVM_EXIT_EXCEPTION_4: /* X86_XCPT_OF */
3585	case SVM_EXIT_EXCEPTION_5: /* X86_XCPT_BR */
3586	case SVM_EXIT_EXCEPTION_6: /* X86_XCPT_UD */
3587	/* SVM_EXIT_EXCEPTION_7: / / X86_XCPT_NM - Handled above. */
3588	case SVM_EXIT_EXCEPTION_8: /* X86_XCPT_DF */
3589	case SVM_EXIT_EXCEPTION_9: /* X86_XCPT_CO_SEG_OVERRUN */
3590	case SVM_EXIT_EXCEPTION_A: /* X86_XCPT_TS */
3591	case SVM_EXIT_EXCEPTION_B: /* X86_XCPT_NP */
3592	case SVM_EXIT_EXCEPTION_C: /* X86_XCPT_SS */
3593	case SVM_EXIT_EXCEPTION_D: /* X86_XCPT_GP */
3594	/* SVM_EXIT_EXCEPTION_E: / / X86_XCPT_PF - Handled above. */
3595	/* SVM_EXIT_EXCEPTION_10: / / X86_XCPT_MF - Handled above. */
3596	case SVM_EXIT_EXCEPTION_11: /* X86_XCPT_AC */
3597	case SVM_EXIT_EXCEPTION_12: /* X86_XCPT_MC */
3598	case SVM_EXIT_EXCEPTION_13: /* X86_XCPT_XF */
3599	case SVM_EXIT_EXCEPTION_F: /* Reserved */
3600	case SVM_EXIT_EXCEPTION_14: case SVM_EXIT_EXCEPTION_15: case SVM_EXIT_EXCEPTION_16:
3601	case SVM_EXIT_EXCEPTION_17: case SVM_EXIT_EXCEPTION_18: case SVM_EXIT_EXCEPTION_19:
3602	case SVM_EXIT_EXCEPTION_1A: case SVM_EXIT_EXCEPTION_1B: case SVM_EXIT_EXCEPTION_1C:
3603	case SVM_EXIT_EXCEPTION_1D: case SVM_EXIT_EXCEPTION_1E: case SVM_EXIT_EXCEPTION_1F:
3604	{
3605	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
3606	SVMEVENT Event;
3607	Event.u = 0;
3608	Event.n.u1Valid = 1;
3609	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3610	Event.n.u8Vector = pSvmTransient->u64ExitCode - SVM_EXIT_EXCEPTION_0;
3611
3612	switch (Event.n.u8Vector)
3613	{
3614	case X86_XCPT_DE:
3615	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestDE);
3616	break;
3617
3618	case X86_XCPT_BP:
3619	/** Saves the wrong EIP on the stack (pointing to the int3) instead of the
3620	* next instruction. */
3621	/** @todo Investigate this later. */
3622	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestBP);
3623	break;
3624
3625	case X86_XCPT_UD:
3626	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestUD);
3627	break;
3628
3629	case X86_XCPT_NP:
3630	Event.n.u1ErrorCodeValid = 1;
3631	Event.n.u32ErrorCode = pVmcb->ctrl.u64ExitInfo1;
3632	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestNP);
3633	break;
3634
3635	case X86_XCPT_SS:
3636	Event.n.u1ErrorCodeValid = 1;
3637	Event.n.u32ErrorCode = pVmcb->ctrl.u64ExitInfo1;
3638	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestSS);
3639	break;
3640
3641	case X86_XCPT_GP:
3642	Event.n.u1ErrorCodeValid = 1;
3643	Event.n.u32ErrorCode = pVmcb->ctrl.u64ExitInfo1;
3644	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestGP);
3645	break;
3646
3647	default:
3648	AssertMsgFailed(("hmR0SvmHandleExit: Unexpected exit caused by exception %#x\n", Event.n.u8Vector));
3649	pVCpu->hm.s.u32HMError = Event.n.u8Vector;
3650	return VERR_SVM_UNEXPECTED_XCPT_EXIT;
3651	}
3652
3653	Log4(("#Xcpt: Vector=%#x at CS:RIP=%04x:%RGv\n", Event.n.u8Vector, pCtx->cs.Sel, (RTGCPTR)pCtx->rip));
3654	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3655	return VINF_SUCCESS;
3656	}
3657	#endif /* HMSVM_ALWAYS_TRAP_ALL_XCPTS */
3658
3659	default:
3660	{
3661	AssertMsgFailed(("hmR0SvmHandleExit: Unknown exit code %#x\n", u32ExitCode));
3662	pVCpu->hm.s.u32HMError = u32ExitCode;
3663	return VERR_SVM_UNKNOWN_EXIT;
3664	}
3665	}
3666	}
3667	}
3668	return VERR_INTERNAL_ERROR_5; /* Should never happen. */
3669	}
3670
3671
3672	#ifdef DEBUG
3673	/* Is there some generic IPRT define for this that are not in Runtime/internal/\* ?? */
3674	# define HMSVM_ASSERT_PREEMPT_CPUID_VAR() \
3675	RTCPUID const idAssertCpu = RTThreadPreemptIsEnabled(NIL_RTTHREAD) ? NIL_RTCPUID : RTMpCpuId()
3676
3677	# define HMSVM_ASSERT_PREEMPT_CPUID() \
3678	do \
3679	{ \
3680	RTCPUID const idAssertCpuNow = RTThreadPreemptIsEnabled(NIL_RTTHREAD) ? NIL_RTCPUID : RTMpCpuId(); \
3681	AssertMsg(idAssertCpu == idAssertCpuNow, ("SVM %#x, %#x\n", idAssertCpu, idAssertCpuNow)); \
3682	} while (0)
3683
3684	# define HMSVM_VALIDATE_EXIT_HANDLER_PARAMS() \
3685	do { \
3686	AssertPtr(pVCpu); \
3687	AssertPtr(pCtx); \
3688	AssertPtr(pSvmTransient); \
3689	Assert(ASMIntAreEnabled()); \
3690	HMSVM_ASSERT_PREEMPT_SAFE(); \
3691	HMSVM_ASSERT_PREEMPT_CPUID_VAR(); \
3692	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)); \
3693	HMSVM_ASSERT_PREEMPT_SAFE(); \
3694	if (VMMR0IsLogFlushDisabled(pVCpu)) \
3695	HMSVM_ASSERT_PREEMPT_CPUID(); \
3696	} while (0)
3697	#else /* Release builds */
3698	# define HMSVM_VALIDATE_EXIT_HANDLER_PARAMS() do { NOREF(pVCpu); NOREF(pCtx); NOREF(pSvmTransient); } while (0)
3699	#endif
3700
3701
3702	/**
3703	* Worker for hmR0SvmInterpretInvlpg().
3704	*
3705	* @return VBox status code.
3706	* @param pVCpu Pointer to the VMCPU.
3707	* @param pCpu Pointer to the disassembler state.
3708	* @param pRegFrame Pointer to the register frame.
3709	*/
3710	static int hmR0SvmInterpretInvlPgEx(PVMCPU pVCpu, PDISCPUSTATE pCpu, PCPUMCTXCORE pRegFrame)
3711	{
3712	DISQPVPARAMVAL Param1;
3713	RTGCPTR GCPtrPage;
3714
3715	int rc = DISQueryParamVal(pRegFrame, pCpu, &pCpu->Param1, &Param1, DISQPVWHICH_SRC);
3716	if (RT_FAILURE(rc))
3717	return VERR_EM_INTERPRETER;
3718
3719	if ( Param1.type == DISQPV_TYPE_IMMEDIATE
3720	\|\| Param1.type == DISQPV_TYPE_ADDRESS)
3721	{
3722	if (!(Param1.flags & (DISQPV_FLAG_32 \| DISQPV_FLAG_64)))
3723	return VERR_EM_INTERPRETER;
3724
3725	GCPtrPage = Param1.val.val64;
3726	VBOXSTRICTRC rc2 = EMInterpretInvlpg(pVCpu->CTX_SUFF(pVM), pVCpu, pRegFrame, GCPtrPage);
3727	rc = VBOXSTRICTRC_VAL(rc2);
3728	}
3729	else
3730	{
3731	Log4(("hmR0SvmInterpretInvlPgEx invalid parameter type %#x\n", Param1.type));
3732	rc = VERR_EM_INTERPRETER;
3733	}
3734
3735	return rc;
3736	}
3737
3738
3739	/**
3740	* Interprets INVLPG.
3741	*
3742	* @returns VBox status code.
3743	* @retval VINF_* Scheduling instructions.
3744	* @retval VERR_EM_INTERPRETER Something we can't cope with.
3745	* @retval VERR_* Fatal errors.
3746	*
3747	* @param pVM Pointer to the VM.
3748	* @param pRegFrame Pointer to the register frame.
3749	*
3750	* @remarks Updates the RIP if the instruction was executed successfully.
3751	*/
3752	static int hmR0SvmInterpretInvlpg(PVM pVM, PVMCPU pVCpu, PCPUMCTXCORE pRegFrame)
3753	{
3754	/* Only allow 32 & 64 bit code. */
3755	if (CPUMGetGuestCodeBits(pVCpu) != 16)
3756	{
3757	PDISSTATE pDis = &pVCpu->hm.s.DisState;
3758	int rc = EMInterpretDisasCurrent(pVM, pVCpu, pDis, NULL /* pcbInstr */);
3759	if ( RT_SUCCESS(rc)
3760	&& pDis->pCurInstr->uOpcode == OP_INVLPG)
3761	{
3762	rc = hmR0SvmInterpretInvlPgEx(pVCpu, pDis, pRegFrame);
3763	if (RT_SUCCESS(rc))
3764	pRegFrame->rip += pDis->cbInstr;
3765	return rc;
3766	}
3767	else
3768	Log4(("hmR0SvmInterpretInvlpg: EMInterpretDisasCurrent returned %Rrc uOpCode=%#x\n", rc, pDis->pCurInstr->uOpcode));
3769	}
3770	return VERR_EM_INTERPRETER;
3771	}
3772
3773
3774	/**
3775	* Sets an invalid-opcode (#UD) exception as pending-for-injection into the VM.
3776	*
3777	* @param pVCpu Pointer to the VMCPU.
3778	*/
3779	DECLINLINE(void) hmR0SvmSetPendingXcptUD(PVMCPU pVCpu)
3780	{
3781	SVMEVENT Event;
3782	Event.u = 0;
3783	Event.n.u1Valid = 1;
3784	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3785	Event.n.u8Vector = X86_XCPT_UD;
3786	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3787	}
3788
3789
3790	/**
3791	* Sets a debug (#DB) exception as pending-for-injection into the VM.
3792	*
3793	* @param pVCpu Pointer to the VMCPU.
3794	*/
3795	DECLINLINE(void) hmR0SvmSetPendingXcptDB(PVMCPU pVCpu)
3796	{
3797	SVMEVENT Event;
3798	Event.u = 0;
3799	Event.n.u1Valid = 1;
3800	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3801	Event.n.u8Vector = X86_XCPT_DB;
3802	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3803	}
3804
3805
3806	/**
3807	* Sets a page fault (#PF) exception as pending-for-injection into the VM.
3808	*
3809	* @param pVCpu Pointer to the VMCPU.
3810	* @param pCtx Pointer to the guest-CPU context.
3811	* @param u32ErrCode The error-code for the page-fault.
3812	* @param uFaultAddress The page fault address (CR2).
3813	*
3814	* @remarks This updates the guest CR2 with @a uFaultAddress!
3815	*/
3816	DECLINLINE(void) hmR0SvmSetPendingXcptPF(PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t u32ErrCode, RTGCUINTPTR uFaultAddress)
3817	{
3818	SVMEVENT Event;
3819	Event.u = 0;
3820	Event.n.u1Valid = 1;
3821	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3822	Event.n.u8Vector = X86_XCPT_PF;
3823	Event.n.u1ErrorCodeValid = 1;
3824	Event.n.u32ErrorCode = u32ErrCode;
3825
3826	/* Update CR2 of the guest. */
3827	if (pCtx->cr2 != uFaultAddress)
3828	{
3829	pCtx->cr2 = uFaultAddress;
3830	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR2);
3831	}
3832
3833	hmR0SvmSetPendingEvent(pVCpu, &Event, uFaultAddress);
3834	}
3835
3836
3837	/**
3838	* Sets a device-not-available (#NM) exception as pending-for-injection into the
3839	* VM.
3840	*
3841	* @param pVCpu Pointer to the VMCPU.
3842	*/
3843	DECLINLINE(void) hmR0SvmSetPendingXcptNM(PVMCPU pVCpu)
3844	{
3845	SVMEVENT Event;
3846	Event.u = 0;
3847	Event.n.u1Valid = 1;
3848	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3849	Event.n.u8Vector = X86_XCPT_NM;
3850	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3851	}
3852
3853
3854	/**
3855	* Sets a math-fault (#MF) exception as pending-for-injection into the VM.
3856	*
3857	* @param pVCpu Pointer to the VMCPU.
3858	*/
3859	DECLINLINE(void) hmR0SvmSetPendingXcptMF(PVMCPU pVCpu)
3860	{
3861	SVMEVENT Event;
3862	Event.u = 0;
3863	Event.n.u1Valid = 1;
3864	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3865	Event.n.u8Vector = X86_XCPT_MF;
3866	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3867	}
3868
3869
3870	/**
3871	* Sets a double fault (#DF) exception as pending-for-injection into the VM.
3872	*
3873	* @param pVCpu Pointer to the VMCPU.
3874	*/
3875	DECLINLINE(void) hmR0SvmSetPendingXcptDF(PVMCPU pVCpu)
3876	{
3877	SVMEVENT Event;
3878	Event.u = 0;
3879	Event.n.u1Valid = 1;
3880	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3881	Event.n.u8Vector = X86_XCPT_DF;
3882	Event.n.u1ErrorCodeValid = 1;
3883	Event.n.u32ErrorCode = 0;
3884	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3885	}
3886
3887
3888	/**
3889	* Emulates a simple MOV TPR (CR8) instruction, used for TPR patching on 32-bit
3890	* guests. This simply looks up the patch record at EIP and does the required.
3891	*
3892	* This VMMCALL is used a fallback mechanism when mov to/from cr8 isn't exactly
3893	* like how we want it to be (e.g. not followed by shr 4 as is usually done for
3894	* TPR). See hmR3ReplaceTprInstr() for the details.
3895	*
3896	* @returns VBox status code.
3897	* @retval VINF_SUCCESS if the access was handled successfully.
3898	* @retval VERR_NOT_FOUND if no patch record for this eip could be found.
3899	* @retval VERR_SVM_UNEXPECTED_PATCH_TYPE if the found patch type is invalid.
3900	*
3901	* @param pVM Pointer to the VM.
3902	* @param pVCpu Pointer to the VMCPU.
3903	* @param pCtx Pointer to the guest-CPU context.
3904	*/
3905	static int hmR0SvmEmulateMovTpr(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3906	{
3907	Log4(("Emulated VMMCall TPR access replacement at RIP=%RGv\n", pCtx->rip));
3908
3909	/*
3910	* We do this in a loop as we increment the RIP after a successful emulation
3911	* and the new RIP may be a patched instruction which needs emulation as well.
3912	*/
3913	bool fPatchFound = false;
3914	for (;;)
3915	{
3916	bool fPending;
3917	uint8_t u8Tpr;
3918
3919	PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
3920	if (!pPatch)
3921	break;
3922
3923	fPatchFound = true;
3924	switch (pPatch->enmType)
3925	{
3926	case HMTPRINSTR_READ:
3927	{
3928	int rc = PDMApicGetTPR(pVCpu, &u8Tpr, &fPending, NULL /* pu8PendingIrq */);
3929	AssertRC(rc);
3930
3931	rc = DISWriteReg32(CPUMCTX2CORE(pCtx), pPatch->uDstOperand, u8Tpr);
3932	AssertRC(rc);
3933	pCtx->rip += pPatch->cbOp;
3934	break;
3935	}
3936
3937	case HMTPRINSTR_WRITE_REG:
3938	case HMTPRINSTR_WRITE_IMM:
3939	{
3940	if (pPatch->enmType == HMTPRINSTR_WRITE_REG)
3941	{
3942	uint32_t u32Val;
3943	int rc = DISFetchReg32(CPUMCTX2CORE(pCtx), pPatch->uSrcOperand, &u32Val);
3944	AssertRC(rc);
3945	u8Tpr = u32Val;
3946	}
3947	else
3948	u8Tpr = (uint8_t)pPatch->uSrcOperand;
3949
3950	int rc2 = PDMApicSetTPR(pVCpu, u8Tpr);
3951	AssertRC(rc2);
3952	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
3953
3954	pCtx->rip += pPatch->cbOp;
3955	break;
3956	}
3957
3958	default:
3959	AssertMsgFailed(("Unexpected patch type %d\n", pPatch->enmType));
3960	pVCpu->hm.s.u32HMError = pPatch->enmType;
3961	return VERR_SVM_UNEXPECTED_PATCH_TYPE;
3962	}
3963	}
3964
3965	if (fPatchFound)
3966	return VINF_SUCCESS;
3967	return VERR_NOT_FOUND;
3968	}
3969
3970
3971	/**
3972	* Determines if an exception is a contributory exception.
3973	*
3974	* Contributory exceptions are ones which can cause double-faults unless the
3975	* original exception was a benign exception. Page-fault is intentionally not
3976	* included here as it's a conditional contributory exception.
3977	*
3978	* @returns true if the exception is contributory, false otherwise.
3979	* @param uVector The exception vector.
3980	*/
3981	DECLINLINE(bool) hmR0SvmIsContributoryXcpt(const uint32_t uVector)
3982	{
3983	switch (uVector)
3984	{
3985	case X86_XCPT_GP:
3986	case X86_XCPT_SS:
3987	case X86_XCPT_NP:
3988	case X86_XCPT_TS:
3989	case X86_XCPT_DE:
3990	return true;
3991	default:
3992	break;
3993	}
3994	return false;
3995	}
3996
3997
3998	/**
3999	* Handle a condition that occurred while delivering an event through the guest
4000	* IDT.
4001	*
4002	* @returns VBox status code (informational error codes included).
4003	* @retval VINF_SUCCESS if we should continue handling the #VMEXIT.
4004	* @retval VINF_HM_DOUBLE_FAULT if a #DF condition was detected and we ought to
4005	* continue execution of the guest which will delivery the #DF.
4006	* @retval VINF_EM_RESET if we detected a triple-fault condition.
4007	*
4008	* @param pVCpu Pointer to the VMCPU.
4009	* @param pCtx Pointer to the guest-CPU context.
4010	* @param pSvmTransient Pointer to the SVM transient structure.
4011	*
4012	* @remarks No-long-jump zone!!!
4013	*/
4014	static int hmR0SvmCheckExitDueToEventDelivery(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4015	{
4016	int rc = VINF_SUCCESS;
4017	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4018
4019	/* See AMD spec. 15.7.3 "EXITINFO Pseudo-Code". The EXITINTINFO (if valid) contains the prior exception (IDT vector)
4020	* that was trying to be delivered to the guest which caused a #VMEXIT which was intercepted (Exit vector). */
4021	if (pVmcb->ctrl.ExitIntInfo.n.u1Valid)
4022	{
4023	uint8_t uIdtVector = pVmcb->ctrl.ExitIntInfo.n.u8Vector;
4024
4025	typedef enum
4026	{
4027	SVMREFLECTXCPT_XCPT, /* Reflect the exception to the guest or for further evaluation by VMM. */
4028	SVMREFLECTXCPT_DF, /* Reflect the exception as a double-fault to the guest. */
4029	SVMREFLECTXCPT_TF, /* Indicate a triple faulted state to the VMM. */
4030	SVMREFLECTXCPT_NONE /* Nothing to reflect. */
4031	} SVMREFLECTXCPT;
4032
4033	SVMREFLECTXCPT enmReflect = SVMREFLECTXCPT_NONE;
4034	bool fReflectingNmi = false;
4035	if (pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_EXCEPTION)
4036	{
4037	if (pSvmTransient->u64ExitCode - SVM_EXIT_EXCEPTION_0 <= SVM_EXIT_EXCEPTION_1F)
4038	{
4039	uint8_t uExitVector = (uint8_t)(pSvmTransient->u64ExitCode - SVM_EXIT_EXCEPTION_0);
4040
4041	#ifdef VBOX_STRICT
4042	if ( hmR0SvmIsContributoryXcpt(uIdtVector)
4043	&& uExitVector == X86_XCPT_PF)
4044	{
4045	Log4(("IDT: Contributory #PF uCR2=%#RX64\n", pVCpu->idCpu, pCtx->cr2));
4046	}
4047	#endif
4048	if ( uExitVector == X86_XCPT_PF
4049	&& uIdtVector == X86_XCPT_PF)
4050	{
4051	pSvmTransient->fVectoringDoublePF = true;
4052	Log4(("IDT: Vectoring double #PF uCR2=%#RX64\n", pCtx->cr2));
4053	}
4054	else if ( (pVmcb->ctrl.u32InterceptException & HMSVM_CONTRIBUTORY_XCPT_MASK)
4055	&& hmR0SvmIsContributoryXcpt(uExitVector)
4056	&& ( hmR0SvmIsContributoryXcpt(uIdtVector)
4057	\|\| uIdtVector == X86_XCPT_PF))
4058	{
4059	enmReflect = SVMREFLECTXCPT_DF;
4060	Log4(("IDT: Pending vectoring #DF %#RX64 uIdtVector=%#x uExitVector=%#x\n", pVCpu->hm.s.Event.u64IntInfo,
4061	uIdtVector, uExitVector));
4062	}
4063	else if (uIdtVector == X86_XCPT_DF)
4064	{
4065	enmReflect = SVMREFLECTXCPT_TF;
4066	Log4(("IDT: Pending vectoring triple-fault %#RX64 uIdtVector=%#x uExitVector=%#x\n",
4067	pVCpu->hm.s.Event.u64IntInfo, uIdtVector, uExitVector));
4068	}
4069	else
4070	enmReflect = SVMREFLECTXCPT_XCPT;
4071	}
4072	else
4073	{
4074	/*
4075	* If event delivery caused an #VMEXIT that is not an exception (e.g. #NPF) then reflect the original
4076	* exception to the guest after handling the #VMEXIT.
4077	*/
4078	enmReflect = SVMREFLECTXCPT_XCPT;
4079	}
4080	}
4081	else if ( pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_EXTERNAL_IRQ
4082	\|\| pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_NMI)
4083	{
4084	enmReflect = SVMREFLECTXCPT_XCPT;
4085	fReflectingNmi = RT_BOOL(pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_NMI);
4086
4087	if (pSvmTransient->u64ExitCode - SVM_EXIT_EXCEPTION_0 <= SVM_EXIT_EXCEPTION_1F)
4088	{
4089	uint8_t uExitVector = (uint8_t)(pSvmTransient->u64ExitCode - SVM_EXIT_EXCEPTION_0);
4090	if (uExitVector == X86_XCPT_PF)
4091	{
4092	pSvmTransient->fVectoringPF = true;
4093	Log4(("IDT: Vectoring #PF due to Ext-Int/NMI. uCR2=%#RX64\n", pCtx->cr2));
4094	}
4095	}
4096	}
4097	/* else: Ignore software interrupts (INT n) as they reoccur when restarting the instruction. */
4098
4099	switch (enmReflect)
4100	{
4101	case SVMREFLECTXCPT_XCPT:
4102	{
4103	/* If we are re-injecting the NMI, clear NMI blocking. */
4104	if (fReflectingNmi)
4105	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_BLOCK_NMIS);
4106
4107	Assert(pVmcb->ctrl.ExitIntInfo.n.u3Type != SVM_EVENT_SOFTWARE_INT);
4108	hmR0SvmSetPendingEvent(pVCpu, &pVmcb->ctrl.ExitIntInfo, 0 /* GCPtrFaultAddress */);
4109
4110	/* If uExitVector is #PF, CR2 value will be updated from the VMCB if it's a guest #PF. See hmR0SvmExitXcptPF(). */
4111	Log4(("IDT: Pending vectoring event %#RX64 ErrValid=%RTbool Err=%#RX32\n", pVmcb->ctrl.ExitIntInfo.u,
4112	!!pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid, pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode));
4113	break;
4114	}
4115
4116	case SVMREFLECTXCPT_DF:
4117	{
4118	hmR0SvmSetPendingXcptDF(pVCpu);
4119	rc = VINF_HM_DOUBLE_FAULT;
4120	break;
4121	}
4122
4123	case SVMREFLECTXCPT_TF:
4124	{
4125	rc = VINF_EM_RESET;
4126	break;
4127	}
4128
4129	default:
4130	Assert(rc == VINF_SUCCESS);
4131	break;
4132	}
4133	}
4134	Assert(rc == VINF_SUCCESS \|\| rc == VINF_HM_DOUBLE_FAULT \|\| rc == VINF_EM_RESET);
4135	NOREF(pCtx);
4136	return rc;
4137	}
4138
4139
4140	/**
4141	* Advances the guest RIP in the if the NRIP_SAVE feature is supported by the
4142	* CPU, otherwise advances the RIP by @a cb bytes.
4143	*
4144	* @param pVCpu Pointer to the VMCPU.
4145	* @param pCtx Pointer to the guest-CPU context.
4146	* @param cb RIP increment value in bytes.
4147	*
4148	* @remarks Use this function only from #VMEXIT's where the NRIP value is valid
4149	* when NRIP_SAVE is supported by the CPU!
4150	*/
4151	DECLINLINE(void) hmR0SvmUpdateRip(PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t cb)
4152	{
4153	if (pVCpu->CTX_SUFF(pVM)->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_NRIP_SAVE)
4154	{
4155	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4156	pCtx->rip = pVmcb->ctrl.u64NextRIP;
4157	}
4158	else
4159	pCtx->rip += cb;
4160	}
4161
4162
4163	/* -=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= */
4164	/* -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- #VMEXIT handlers -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- */
4165	/* -=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= */
4166
4167	/** @name #VMEXIT handlers.
4168	* @{
4169	*/
4170
4171	/**
4172	* #VMEXIT handler for external interrupts, NMIs, FPU assertion freeze and INIT
4173	* signals (SVM_EXIT_INTR, SVM_EXIT_NMI, SVM_EXIT_FERR_FREEZE, SVM_EXIT_INIT).
4174	*/
4175	HMSVM_EXIT_DECL hmR0SvmExitIntr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4176	{
4177	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4178
4179	if (pSvmTransient->u64ExitCode == SVM_EXIT_NMI)
4180	STAM_REL_COUNTER_INC(&pVCpu->hm.s.StatExitHostNmiInGC);
4181	else if (pSvmTransient->u64ExitCode == SVM_EXIT_INTR)
4182	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitExtInt);
4183
4184	/*
4185	* AMD-V has no preemption timer and the generic periodic preemption timer has no way to signal -before- the timer
4186	* fires if the current interrupt is our own timer or a some other host interrupt. We also cannot examine what
4187	* interrupt it is until the host actually take the interrupt.
4188	*
4189	* Going back to executing guest code here unconditionally causes random scheduling problems (observed on an
4190	* AMD Phenom 9850 Quad-Core on Windows 64-bit host).
4191	*/
4192	return VINF_EM_RAW_INTERRUPT;
4193	}
4194
4195
4196	/**
4197	* #VMEXIT handler for WBINVD (SVM_EXIT_WBINVD). Conditional #VMEXIT.
4198	*/
4199	HMSVM_EXIT_DECL hmR0SvmExitWbinvd(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4200	{
4201	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4202
4203	hmR0SvmUpdateRip(pVCpu, pCtx, 2);
4204	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitWbinvd);
4205	int rc = VINF_SUCCESS;
4206	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4207	return rc;
4208	}
4209
4210
4211	/**
4212	* #VMEXIT handler for INVD (SVM_EXIT_INVD). Unconditional #VMEXIT.
4213	*/
4214	HMSVM_EXIT_DECL hmR0SvmExitInvd(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4215	{
4216	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4217
4218	hmR0SvmUpdateRip(pVCpu, pCtx, 2);
4219	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitInvd);
4220	int rc = VINF_SUCCESS;
4221	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4222	return rc;
4223	}
4224
4225
4226	/**
4227	* #VMEXIT handler for INVD (SVM_EXIT_CPUID). Conditional #VMEXIT.
4228	*/
4229	HMSVM_EXIT_DECL hmR0SvmExitCpuid(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4230	{
4231	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4232	PVM pVM = pVCpu->CTX_SUFF(pVM);
4233	int rc = EMInterpretCpuId(pVM, pVCpu, CPUMCTX2CORE(pCtx));
4234	if (RT_LIKELY(rc == VINF_SUCCESS))
4235	{
4236	hmR0SvmUpdateRip(pVCpu, pCtx, 2);
4237	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4238	}
4239	else
4240	{
4241	AssertMsgFailed(("hmR0SvmExitCpuid: EMInterpretCpuId failed with %Rrc\n", rc));
4242	rc = VERR_EM_INTERPRETER;
4243	}
4244	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCpuid);
4245	return rc;
4246	}
4247
4248
4249	/**
4250	* #VMEXIT handler for RDTSC (SVM_EXIT_RDTSC). Conditional #VMEXIT.
4251	*/
4252	HMSVM_EXIT_DECL hmR0SvmExitRdtsc(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4253	{
4254	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4255	PVM pVM = pVCpu->CTX_SUFF(pVM);
4256	int rc = EMInterpretRdtsc(pVM, pVCpu, CPUMCTX2CORE(pCtx));
4257	if (RT_LIKELY(rc == VINF_SUCCESS))
4258	{
4259	hmR0SvmUpdateRip(pVCpu, pCtx, 2);
4260	pSvmTransient->fUpdateTscOffsetting = true;
4261
4262	/* Single step check. */
4263	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4264	}
4265	else
4266	{
4267	AssertMsgFailed(("hmR0SvmExitRdtsc: EMInterpretRdtsc failed with %Rrc\n", rc));
4268	rc = VERR_EM_INTERPRETER;
4269	}
4270	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdtsc);
4271	return rc;
4272	}
4273
4274
4275	/**
4276	* #VMEXIT handler for RDTSCP (SVM_EXIT_RDTSCP). Conditional #VMEXIT.
4277	*/
4278	HMSVM_EXIT_DECL hmR0SvmExitRdtscp(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4279	{
4280	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4281	int rc = EMInterpretRdtscp(pVCpu->CTX_SUFF(pVM), pVCpu, pCtx);
4282	if (RT_LIKELY(rc == VINF_SUCCESS))
4283	{
4284	hmR0SvmUpdateRip(pVCpu, pCtx, 3);
4285	pSvmTransient->fUpdateTscOffsetting = true;
4286	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4287	}
4288	else
4289	{
4290	AssertMsgFailed(("hmR0SvmExitRdtsc: EMInterpretRdtscp failed with %Rrc\n", rc));
4291	rc = VERR_EM_INTERPRETER;
4292	}
4293	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdtscp);
4294	return rc;
4295	}
4296
4297
4298	/**
4299	* #VMEXIT handler for RDPMC (SVM_EXIT_RDPMC). Conditional #VMEXIT.
4300	*/
4301	HMSVM_EXIT_DECL hmR0SvmExitRdpmc(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4302	{
4303	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4304	int rc = EMInterpretRdpmc(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
4305	if (RT_LIKELY(rc == VINF_SUCCESS))
4306	{
4307	hmR0SvmUpdateRip(pVCpu, pCtx, 2);
4308	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4309	}
4310	else
4311	{
4312	AssertMsgFailed(("hmR0SvmExitRdpmc: EMInterpretRdpmc failed with %Rrc\n", rc));
4313	rc = VERR_EM_INTERPRETER;
4314	}
4315	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdpmc);
4316	return rc;
4317	}
4318
4319
4320	/**
4321	* #VMEXIT handler for INVLPG (SVM_EXIT_INVLPG). Conditional #VMEXIT.
4322	*/
4323	HMSVM_EXIT_DECL hmR0SvmExitInvlpg(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4324	{
4325	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4326	PVM pVM = pVCpu->CTX_SUFF(pVM);
4327	Assert(!pVM->hm.s.fNestedPaging);
4328
4329	/** @todo Decode Assist. */
4330	int rc = hmR0SvmInterpretInvlpg(pVM, pVCpu, CPUMCTX2CORE(pCtx)); /* Updates RIP if successful. */
4331	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitInvlpg);
4332	Assert(rc == VINF_SUCCESS \|\| rc == VERR_EM_INTERPRETER);
4333	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4334	return rc;
4335	}
4336
4337
4338	/**
4339	* #VMEXIT handler for HLT (SVM_EXIT_HLT). Conditional #VMEXIT.
4340	*/
4341	HMSVM_EXIT_DECL hmR0SvmExitHlt(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4342	{
4343	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4344
4345	hmR0SvmUpdateRip(pVCpu, pCtx, 1);
4346	int rc = EMShouldContinueAfterHalt(pVCpu, pCtx) ? VINF_SUCCESS : VINF_EM_HALT;
4347	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4348	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitHlt);
4349	if (rc != VINF_SUCCESS)
4350	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHltToR3);
4351	return rc;
4352	}
4353
4354
4355	/**
4356	* #VMEXIT handler for MONITOR (SVM_EXIT_MONITOR). Conditional #VMEXIT.
4357	*/
4358	HMSVM_EXIT_DECL hmR0SvmExitMonitor(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4359	{
4360	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4361	int rc = EMInterpretMonitor(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
4362	if (RT_LIKELY(rc == VINF_SUCCESS))
4363	{
4364	hmR0SvmUpdateRip(pVCpu, pCtx, 3);
4365	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4366	}
4367	else
4368	{
4369	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMonitor: EMInterpretMonitor failed with %Rrc\n", rc));
4370	rc = VERR_EM_INTERPRETER;
4371	}
4372	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitMonitor);
4373	return rc;
4374	}
4375
4376
4377	/**
4378	* #VMEXIT handler for MWAIT (SVM_EXIT_MWAIT). Conditional #VMEXIT.
4379	*/
4380	HMSVM_EXIT_DECL hmR0SvmExitMwait(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4381	{
4382	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4383	VBOXSTRICTRC rc2 = EMInterpretMWait(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
4384	int rc = VBOXSTRICTRC_VAL(rc2);
4385	if ( rc == VINF_EM_HALT
4386	\|\| rc == VINF_SUCCESS)
4387	{
4388	hmR0SvmUpdateRip(pVCpu, pCtx, 3);
4389
4390	if ( rc == VINF_EM_HALT
4391	&& EMMonitorWaitShouldContinue(pVCpu, pCtx))
4392	{
4393	rc = VINF_SUCCESS;
4394	}
4395	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4396	}
4397	else
4398	{
4399	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMwait: EMInterpretMWait failed with %Rrc\n", rc));
4400	rc = VERR_EM_INTERPRETER;
4401	}
4402	AssertMsg(rc == VINF_SUCCESS \|\| rc == VINF_EM_HALT \|\| rc == VERR_EM_INTERPRETER,
4403	("hmR0SvmExitMwait: EMInterpretMWait failed rc=%Rrc\n", rc));
4404	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitMwait);
4405	return rc;
4406	}
4407
4408
4409	/**
4410	* #VMEXIT handler for shutdown (triple-fault) (SVM_EXIT_SHUTDOWN).
4411	* Conditional #VMEXIT.
4412	*/
4413	HMSVM_EXIT_DECL hmR0SvmExitShutdown(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4414	{
4415	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4416	return VINF_EM_RESET;
4417	}
4418
4419
4420	/**
4421	* #VMEXIT handler for CRx reads (SVM_EXIT_READ_CR*). Conditional #VMEXIT.
4422	*/
4423	HMSVM_EXIT_DECL hmR0SvmExitReadCRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4424	{
4425	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4426
4427	Log4(("hmR0SvmExitReadCRx: CS:RIP=%04x:%#RX64\n", pCtx->cs.Sel, pCtx->rip));
4428
4429	/** @todo Decode Assist. */
4430	VBOXSTRICTRC rc2 = EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */);
4431	int rc = VBOXSTRICTRC_VAL(rc2);
4432	AssertMsg(rc == VINF_SUCCESS \|\| rc == VERR_EM_INTERPRETER \|\| rc == VINF_PGM_CHANGE_MODE \|\| rc == VINF_PGM_SYNC_CR3,
4433	("hmR0SvmExitReadCRx: EMInterpretInstruction failed rc=%Rrc\n", rc));
4434	Assert((pSvmTransient->u64ExitCode - SVM_EXIT_READ_CR0) <= 15);
4435	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCRxRead[pSvmTransient->u64ExitCode - SVM_EXIT_READ_CR0]);
4436	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4437	return rc;
4438	}
4439
4440
4441	/**
4442	* #VMEXIT handler for CRx writes (SVM_EXIT_WRITE_CR*). Conditional #VMEXIT.
4443	*/
4444	HMSVM_EXIT_DECL hmR0SvmExitWriteCRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4445	{
4446	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4447	/** @todo Decode Assist. */
4448	VBOXSTRICTRC rc2 = EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */);
4449	int rc = VBOXSTRICTRC_VAL(rc2);
4450	if (rc == VINF_SUCCESS)
4451	{
4452	/* RIP has been updated by EMInterpretInstruction(). */
4453	Assert((pSvmTransient->u64ExitCode - SVM_EXIT_WRITE_CR0) <= 15);
4454	switch (pSvmTransient->u64ExitCode - SVM_EXIT_WRITE_CR0)
4455	{
4456	case 0: /* CR0. */
4457	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR0);
4458	break;
4459
4460	case 3: /* CR3. */
4461	Assert(!pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
4462	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR3);
4463	break;
4464
4465	case 4: /* CR4. */
4466	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR4);
4467	break;
4468
4469	case 8: /* CR8 (TPR). */
4470	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
4471	break;
4472
4473	default:
4474	AssertMsgFailed(("hmR0SvmExitWriteCRx: Invalid/Unexpected Write-CRx exit. u64ExitCode=%#RX64 %#x CRx=%#RX64\n",
4475	pSvmTransient->u64ExitCode, pSvmTransient->u64ExitCode - SVM_EXIT_WRITE_CR0));
4476	break;
4477	}
4478	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4479	}
4480	else
4481	Assert(rc == VERR_EM_INTERPRETER \|\| rc == VINF_PGM_CHANGE_MODE \|\| rc == VINF_PGM_SYNC_CR3);
4482	return rc;
4483	}
4484
4485
4486	/**
4487	* #VMEXIT handler for instructions that result in a #UD exception delivered to
4488	* the guest.
4489	*/
4490	HMSVM_EXIT_DECL hmR0SvmExitSetPendingXcptUD(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4491	{
4492	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4493	hmR0SvmSetPendingXcptUD(pVCpu);
4494	return VINF_SUCCESS;
4495	}
4496
4497
4498	/**
4499	* #VMEXIT handler for MSR read and writes (SVM_EXIT_MSR). Conditional #VMEXIT.
4500	*/
4501	HMSVM_EXIT_DECL hmR0SvmExitMsr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4502	{
4503	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4504	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4505	PVM pVM = pVCpu->CTX_SUFF(pVM);
4506
4507	int rc;
4508	if (pVmcb->ctrl.u64ExitInfo1 == SVM_EXIT1_MSR_WRITE)
4509	{
4510	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitWrmsr);
4511
4512	/* Handle TPR patching; intercepted LSTAR write. */
4513	if ( pVM->hm.s.fTPRPatchingActive
4514	&& pCtx->ecx == MSR_K8_LSTAR)
4515	{
4516	if ((pCtx->eax & 0xff) != pSvmTransient->u8GuestTpr)
4517	{
4518	/* Our patch code uses LSTAR for TPR caching for 32-bit guests. */
4519	int rc2 = PDMApicSetTPR(pVCpu, pCtx->eax & 0xff);
4520	AssertRC(rc2);
4521	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
4522	}
4523	hmR0SvmUpdateRip(pVCpu, pCtx, 2);
4524	rc = VINF_SUCCESS;
4525	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4526	return rc;
4527	}
4528
4529	if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_NRIP_SAVE)
4530	{
4531	rc = EMInterpretWrmsr(pVM, pVCpu, CPUMCTX2CORE(pCtx));
4532	if (RT_LIKELY(rc == VINF_SUCCESS))
4533	{
4534	pCtx->rip = pVmcb->ctrl.u64NextRIP;
4535	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4536	}
4537	else
4538	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMsr: EMInterpretWrmsr failed rc=%Rrc\n", rc));
4539	}
4540	else
4541	{
4542	rc = VBOXSTRICTRC_TODO(EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */));
4543	if (RT_LIKELY(rc == VINF_SUCCESS))
4544	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc); /* RIP updated by EMInterpretInstruction(). */
4545	else
4546	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMsr: WrMsr. EMInterpretInstruction failed rc=%Rrc\n", rc));
4547	}
4548
4549	if (rc == VINF_SUCCESS)
4550	{
4551	/* If this is an X2APIC WRMSR access, update the APIC state as well. */
4552	if ( pCtx->ecx >= MSR_IA32_X2APIC_START
4553	&& pCtx->ecx <= MSR_IA32_X2APIC_END)
4554	{
4555	/*
4556	* We've already saved the APIC related guest-state (TPR) in hmR0SvmPostRunGuest(). When full APIC register
4557	* virtualization is implemented we'll have to make sure APIC state is saved from the VMCB before
4558	* EMInterpretWrmsr() changes it.
4559	*/
4560	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
4561	}
4562	else if (pCtx->ecx == MSR_K6_EFER)
4563	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_EFER_MSR);
4564	else if (pCtx->ecx == MSR_IA32_TSC)
4565	pSvmTransient->fUpdateTscOffsetting = true;
4566	}
4567	}
4568	else
4569	{
4570	/* MSR Read access. */
4571	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdmsr);
4572	Assert(pVmcb->ctrl.u64ExitInfo1 == SVM_EXIT1_MSR_READ);
4573
4574	if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_NRIP_SAVE)
4575	{
4576	rc = EMInterpretRdmsr(pVM, pVCpu, CPUMCTX2CORE(pCtx));
4577	if (RT_LIKELY(rc == VINF_SUCCESS))
4578	{
4579	pCtx->rip = pVmcb->ctrl.u64NextRIP;
4580	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4581	}
4582	else
4583	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMsr: EMInterpretRdmsr failed rc=%Rrc\n", rc));
4584	}
4585	else
4586	{
4587	rc = VBOXSTRICTRC_TODO(EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0));
4588	if (RT_UNLIKELY(rc != VINF_SUCCESS))
4589	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMsr: RdMsr. EMInterpretInstruction failed rc=%Rrc\n", rc));
4590	/* RIP updated by EMInterpretInstruction(). */
4591	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4592	}
4593	}
4594
4595	/* RIP has been updated by EMInterpret[Rd\|Wr]msr(). */
4596	return rc;
4597	}
4598
4599
4600	/**
4601	* #VMEXIT handler for DRx read (SVM_EXIT_READ_DRx). Conditional #VMEXIT.
4602	*/
4603	HMSVM_EXIT_DECL hmR0SvmExitReadDRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4604	{
4605	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4606	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitDRxRead);
4607
4608	/* We should -not- get this #VMEXIT if the guest's debug registers were active. */
4609	if (pSvmTransient->fWasGuestDebugStateActive)
4610	{
4611	AssertMsgFailed(("hmR0SvmHandleExit: Unexpected exit %#RX32\n", (uint32_t)pSvmTransient->u64ExitCode));
4612	pVCpu->hm.s.u32HMError = (uint32_t)pSvmTransient->u64ExitCode;
4613	return VERR_SVM_UNEXPECTED_EXIT;
4614	}
4615
4616	/*
4617	* Lazy DR0-3 loading.
4618	*/
4619	if (!pSvmTransient->fWasHyperDebugStateActive)
4620	{
4621	Assert(!DBGFIsStepping(pVCpu)); Assert(!pVCpu->hm.s.fSingleInstruction);
4622	Log5(("hmR0SvmExitReadDRx: Lazy loading guest debug registers\n"));
4623
4624	/* Don't intercept DRx read and writes. */
4625	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4626	pVmcb->ctrl.u16InterceptRdDRx = 0;
4627	pVmcb->ctrl.u16InterceptWrDRx = 0;
4628	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
4629
4630	/* We're playing with the host CPU state here, make sure we don't preempt or longjmp. */
4631	VMMRZCallRing3Disable(pVCpu);
4632	HM_DISABLE_PREEMPT_IF_NEEDED();
4633
4634	/* Save the host & load the guest debug state, restart execution of the MOV DRx instruction. */
4635	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
4636	Assert(CPUMIsGuestDebugStateActive(pVCpu) \|\| HC_ARCH_BITS == 32);
4637
4638	HM_RESTORE_PREEMPT_IF_NEEDED();
4639	VMMRZCallRing3Enable(pVCpu);
4640
4641	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxContextSwitch);
4642	return VINF_SUCCESS;
4643	}
4644
4645	/*
4646	* Interpret the read/writing of DRx.
4647	*/
4648	/** @todo Decode assist. */
4649	VBOXSTRICTRC rc = EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */);
4650	Log5(("hmR0SvmExitReadDRx: Emulated DRx access: rc=%Rrc\n", VBOXSTRICTRC_VAL(rc)));
4651	if (RT_LIKELY(rc == VINF_SUCCESS))
4652	{
4653	/* Not necessary for read accesses but whatever doesn't hurt for now, will be fixed with decode assist. */
4654	/** @todo CPUM should set this flag! */
4655	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_DEBUG);
4656	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4657	}
4658	else
4659	Assert(rc == VERR_EM_INTERPRETER);
4660	return VBOXSTRICTRC_TODO(rc);
4661	}
4662
4663
4664	/**
4665	* #VMEXIT handler for DRx write (SVM_EXIT_WRITE_DRx). Conditional #VMEXIT.
4666	*/
4667	HMSVM_EXIT_DECL hmR0SvmExitWriteDRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4668	{
4669	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4670	/* For now it's the same since we interpret the instruction anyway. Will change when using of Decode Assist is implemented. */
4671	int rc = hmR0SvmExitReadDRx(pVCpu, pCtx, pSvmTransient);
4672	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitDRxWrite);
4673	STAM_COUNTER_DEC(&pVCpu->hm.s.StatExitDRxRead);
4674	return rc;
4675	}
4676
4677
4678	/**
4679	* #VMEXIT handler for I/O instructions (SVM_EXIT_IOIO). Conditional #VMEXIT.
4680	*/
4681	HMSVM_EXIT_DECL hmR0SvmExitIOInstr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4682	{
4683	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4684
4685	/* I/O operation lookup arrays. */
4686	static uint32_t const s_aIOSize[8] = { 0, 1, 2, 0, 4, 0, 0, 0 }; /* Size of the I/O accesses in bytes. */
4687	static uint32_t const s_aIOOpAnd[8] = { 0, 0xff, 0xffff, 0, 0xffffffff, 0, 0, 0 }; /* AND masks for saving
4688	the result (in AL/AX/EAX). */
4689	Log4(("hmR0SvmExitIOInstr: CS:RIP=%04x:%#RX64\n", pCtx->cs.Sel, pCtx->rip));
4690
4691	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4692	PVM pVM = pVCpu->CTX_SUFF(pVM);
4693
4694	/* Refer AMD spec. 15.10.2 "IN and OUT Behaviour" and Figure 15-2. "EXITINFO1 for IOIO Intercept" for the format. */
4695	SVMIOIOEXIT IoExitInfo;
4696	IoExitInfo.u = (uint32_t)pVmcb->ctrl.u64ExitInfo1;
4697	uint32_t uIOWidth = (IoExitInfo.u >> 4) & 0x7;
4698	uint32_t cbValue = s_aIOSize[uIOWidth];
4699	uint32_t uAndVal = s_aIOOpAnd[uIOWidth];
4700
4701	if (RT_UNLIKELY(!cbValue))
4702	{
4703	AssertMsgFailed(("hmR0SvmExitIOInstr: Invalid IO operation. uIOWidth=%u\n", uIOWidth));
4704	return VERR_EM_INTERPRETER;
4705	}
4706
4707	VBOXSTRICTRC rcStrict;
4708	if (IoExitInfo.n.u1STR)
4709	{
4710	/* INS/OUTS - I/O String instruction. */
4711	PDISCPUSTATE pDis = &pVCpu->hm.s.DisState;
4712
4713	/** @todo Huh? why can't we use the segment prefix information given by AMD-V
4714	* in EXITINFO1? Investigate once this thing is up and running. */
4715
4716	rcStrict = EMInterpretDisasCurrent(pVM, pVCpu, pDis, NULL);
4717	if (rcStrict == VINF_SUCCESS)
4718	{
4719	if (IoExitInfo.n.u1Type == SVM_IOIO_WRITE)
4720	{
4721	rcStrict = IOMInterpretOUTSEx(pVM, pVCpu, CPUMCTX2CORE(pCtx), IoExitInfo.n.u16Port, pDis->fPrefix,
4722	(DISCPUMODE)pDis->uAddrMode, cbValue);
4723	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOStringWrite);
4724	}
4725	else
4726	{
4727	rcStrict = IOMInterpretINSEx(pVM, pVCpu, CPUMCTX2CORE(pCtx), IoExitInfo.n.u16Port, pDis->fPrefix,
4728	(DISCPUMODE)pDis->uAddrMode, cbValue);
4729	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOStringRead);
4730	}
4731	}
4732	else
4733	rcStrict = VINF_EM_RAW_EMULATE_INSTR;
4734	}
4735	else
4736	{
4737	/* IN/OUT - I/O instruction. */
4738	Assert(!IoExitInfo.n.u1REP);
4739
4740	if (IoExitInfo.n.u1Type == SVM_IOIO_WRITE)
4741	{
4742	rcStrict = IOMIOPortWrite(pVM, pVCpu, IoExitInfo.n.u16Port, pCtx->eax & uAndVal, cbValue);
4743	if (rcStrict == VINF_IOM_R3_IOPORT_WRITE)
4744	HMR0SavePendingIOPortWrite(pVCpu, pCtx->rip, pVmcb->ctrl.u64ExitInfo2, IoExitInfo.n.u16Port, uAndVal, cbValue);
4745
4746	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOWrite);
4747	}
4748	else
4749	{
4750	uint32_t u32Val = 0;
4751
4752	rcStrict = IOMIOPortRead(pVM, pVCpu, IoExitInfo.n.u16Port, &u32Val, cbValue);
4753	if (IOM_SUCCESS(rcStrict))
4754	{
4755	/* Save result of I/O IN instr. in AL/AX/EAX. */
4756	pCtx->eax = (pCtx->eax & ~uAndVal) \| (u32Val & uAndVal);
4757	}
4758	else if (rcStrict == VINF_IOM_R3_IOPORT_READ)
4759	HMR0SavePendingIOPortRead(pVCpu, pCtx->rip, pVmcb->ctrl.u64ExitInfo2, IoExitInfo.n.u16Port, uAndVal, cbValue);
4760
4761	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIORead);
4762	}
4763	}
4764
4765	if (IOM_SUCCESS(rcStrict))
4766	{
4767	/* AMD-V saves the RIP of the instruction following the IO instruction in EXITINFO2. */
4768	pCtx->rip = pVmcb->ctrl.u64ExitInfo2;
4769
4770	/*
4771	* If any I/O breakpoints are armed, we need to check if one triggered
4772	* and take appropriate action.
4773	* Note that the I/O breakpoint type is undefined if CR4.DE is 0.
4774	*/
4775	/** @todo Optimize away the DBGFBpIsHwIoArmed call by having DBGF tell the
4776	* execution engines about whether hyper BPs and such are pending. */
4777	uint32_t const uDr7 = pCtx->dr[7];
4778	if (RT_UNLIKELY( ( (uDr7 & X86_DR7_ENABLED_MASK)
4779	&& X86_DR7_ANY_RW_IO(uDr7)
4780	&& (pCtx->cr4 & X86_CR4_DE))
4781	\|\| DBGFBpIsHwIoArmed(pVM)))
4782	{
4783	/* We're playing with the host CPU state here, make sure we don't preempt or longjmp. */
4784	VMMRZCallRing3Disable(pVCpu);
4785	HM_DISABLE_PREEMPT_IF_NEEDED();
4786
4787	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxIoCheck);
4788	CPUMR0DebugStateMaybeSaveGuest(pVCpu, false /fDr6/);
4789
4790	VBOXSTRICTRC rcStrict2 = DBGFBpCheckIo(pVM, pVCpu, pCtx, IoExitInfo.n.u16Port, cbValue);
4791	if (rcStrict2 == VINF_EM_RAW_GUEST_TRAP)
4792	{
4793	/* Raise #DB. */
4794	pVmcb->guest.u64DR6 = pCtx->dr[6];
4795	pVmcb->guest.u64DR7 = pCtx->dr[7];
4796	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
4797	hmR0SvmSetPendingXcptDB(pVCpu);
4798	}
4799	/* rcStrict is VINF_SUCCESS or in [VINF_EM_FIRST..VINF_EM_LAST]. */
4800	else if ( rcStrict2 != VINF_SUCCESS
4801	&& (rcStrict == VINF_SUCCESS \|\| rcStrict2 < rcStrict))
4802	rcStrict = rcStrict2;
4803
4804	HM_RESTORE_PREEMPT_IF_NEEDED();
4805	VMMRZCallRing3Enable(pVCpu);
4806	}
4807
4808	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
4809	}
4810
4811	#ifdef VBOX_STRICT
4812	if (rcStrict == VINF_IOM_R3_IOPORT_READ)
4813	Assert(IoExitInfo.n.u1Type == SVM_IOIO_READ);
4814	else if (rcStrict == VINF_IOM_R3_IOPORT_WRITE)
4815	Assert(IoExitInfo.n.u1Type == SVM_IOIO_WRITE);
4816	else
4817	{
4818	/** @todo r=bird: This is missing a bunch of VINF_EM_FIRST..VINF_EM_LAST
4819	* statuses, that the VMM device and some others may return. See
4820	* IOM_SUCCESS() for guidance. */
4821	AssertMsg( RT_FAILURE(rcStrict)
4822	\|\| rcStrict == VINF_SUCCESS
4823	\|\| rcStrict == VINF_EM_RAW_EMULATE_INSTR
4824	\|\| rcStrict == VINF_EM_DBG_BREAKPOINT
4825	\|\| rcStrict == VINF_EM_RAW_GUEST_TRAP
4826	\|\| rcStrict == VINF_TRPM_XCPT_DISPATCHED, ("%Rrc\n", VBOXSTRICTRC_VAL(rcStrict)));
4827	}
4828	#endif
4829	return VBOXSTRICTRC_TODO(rcStrict);
4830	}
4831
4832
4833	/**
4834	* #VMEXIT handler for Nested Page-faults (SVM_EXIT_NPF). Conditional
4835	* #VMEXIT.
4836	*/
4837	HMSVM_EXIT_DECL hmR0SvmExitNestedPF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4838	{
4839	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4840	PVM pVM = pVCpu->CTX_SUFF(pVM);
4841	Assert(pVM->hm.s.fNestedPaging);
4842
4843	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
4844
4845	/* See AMD spec. 15.25.6 "Nested versus Guest Page Faults, Fault Ordering" for VMCB details for #NPF. */
4846	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4847	uint32_t u32ErrCode = pVmcb->ctrl.u64ExitInfo1;
4848	RTGCPHYS GCPhysFaultAddr = pVmcb->ctrl.u64ExitInfo2;
4849
4850	Log4(("#NPF at CS:RIP=%04x:%#RX64 faultaddr=%RGp errcode=%#x \n", pCtx->cs.Sel, pCtx->rip, GCPhysFaultAddr, u32ErrCode));
4851
4852	#ifdef VBOX_HM_WITH_GUEST_PATCHING
4853	/* TPR patching for 32-bit guests, using the reserved bit in the page tables for MMIO regions. */
4854	if ( pVM->hm.s.fTprPatchingAllowed
4855	&& (GCPhysFaultAddr & PAGE_OFFSET_MASK) == 0x80 /* TPR offset. */
4856	&& ( !(u32ErrCode & X86_TRAP_PF_P) /* Not present */
4857	\|\| (u32ErrCode & (X86_TRAP_PF_P \| X86_TRAP_PF_RSVD)) == (X86_TRAP_PF_P \| X86_TRAP_PF_RSVD)) /* MMIO page. */
4858	&& !CPUMIsGuestInLongModeEx(pCtx)
4859	&& !CPUMGetGuestCPL(pVCpu)
4860	&& pVM->hm.s.cPatches < RT_ELEMENTS(pVM->hm.s.aPatches))
4861	{
4862	RTGCPHYS GCPhysApicBase = pCtx->msrApicBase;
4863	GCPhysApicBase &= PAGE_BASE_GC_MASK;
4864
4865	if (GCPhysFaultAddr == GCPhysApicBase + 0x80)
4866	{
4867	/* Only attempt to patch the instruction once. */
4868	PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
4869	if (!pPatch)
4870	return VINF_EM_HM_PATCH_TPR_INSTR;
4871	}
4872	}
4873	#endif
4874
4875	/*
4876	* Determine the nested paging mode.
4877	*/
4878	PGMMODE enmNestedPagingMode;
4879	#if HC_ARCH_BITS == 32
4880	if (CPUMIsGuestInLongModeEx(pCtx))
4881	enmNestedPagingMode = PGMMODE_AMD64_NX;
4882	else
4883	#endif
4884	enmNestedPagingMode = PGMGetHostMode(pVM);
4885
4886	/*
4887	* MMIO optimization using the reserved (RSVD) bit in the guest page tables for MMIO pages.
4888	*/
4889	int rc;
4890	Assert((u32ErrCode & (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P)) != X86_TRAP_PF_RSVD);
4891	if ((u32ErrCode & (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P)) == (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P))
4892	{
4893	VBOXSTRICTRC rc2 = PGMR0Trap0eHandlerNPMisconfig(pVM, pVCpu, enmNestedPagingMode, CPUMCTX2CORE(pCtx), GCPhysFaultAddr,
4894	u32ErrCode);
4895	rc = VBOXSTRICTRC_VAL(rc2);
4896
4897	/*
4898	* If we succeed, resume guest execution.
4899	* If we fail in interpreting the instruction because we couldn't get the guest physical address
4900	* of the page containing the instruction via the guest's page tables (we would invalidate the guest page
4901	* in the host TLB), resume execution which would cause a guest page fault to let the guest handle this
4902	* weird case. See @bugref{6043}.
4903	*/
4904	if ( rc == VINF_SUCCESS
4905	\|\| rc == VERR_PAGE_TABLE_NOT_PRESENT
4906	\|\| rc == VERR_PAGE_NOT_PRESENT)
4907	{
4908	/* Successfully handled MMIO operation. */
4909	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
4910	rc = VINF_SUCCESS;
4911	}
4912	return rc;
4913	}
4914
4915	TRPMAssertXcptPF(pVCpu, GCPhysFaultAddr, u32ErrCode);
4916	rc = PGMR0Trap0eHandlerNestedPaging(pVM, pVCpu, enmNestedPagingMode, u32ErrCode, CPUMCTX2CORE(pCtx), GCPhysFaultAddr);
4917	TRPMResetTrap(pVCpu);
4918
4919	Log4(("#NPF: PGMR0Trap0eHandlerNestedPaging returned %Rrc CS:RIP=%04x:%#RX64\n", rc, pCtx->cs.Sel, pCtx->rip));
4920
4921	/*
4922	* Same case as PGMR0Trap0eHandlerNPMisconfig(). See comment above, @bugref{6043}.
4923	*/
4924	if ( rc == VINF_SUCCESS
4925	\|\| rc == VERR_PAGE_TABLE_NOT_PRESENT
4926	\|\| rc == VERR_PAGE_NOT_PRESENT)
4927	{
4928	/* We've successfully synced our shadow page tables. */
4929	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPF);
4930	rc = VINF_SUCCESS;
4931	}
4932
4933	return rc;
4934	}
4935
4936
4937	/**
4938	* #VMEXIT handler for virtual interrupt (SVM_EXIT_VINTR). Conditional #VMEXIT.
4939	*/
4940	HMSVM_EXIT_DECL hmR0SvmExitVIntr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4941	{
4942	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4943
4944	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4945	pVmcb->ctrl.IntCtrl.n.u1VIrqValid = 0; /* No virtual interrupts pending, we'll inject the current one/NMI before reentry. */
4946	pVmcb->ctrl.IntCtrl.n.u8VIrqVector = 0;
4947
4948	/* Indicate that we no longer need to #VMEXIT when the guest is ready to receive interrupts/NMIs, it is now ready. */
4949	pVmcb->ctrl.u32InterceptCtrl1 &= ~SVM_CTRL1_INTERCEPT_VINTR;
4950	pVmcb->ctrl.u64VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS \| HMSVM_VMCB_CLEAN_TPR);
4951
4952	/* Deliver the pending interrupt/NMI via hmR0SvmEvaluatePendingEvent() and resume guest execution. */
4953	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIntWindow);
4954	return VINF_SUCCESS;
4955	}
4956
4957
4958	/**
4959	* #VMEXIT handler for task switches (SVM_EXIT_TASK_SWITCH). Conditional #VMEXIT.
4960	*/
4961	HMSVM_EXIT_DECL hmR0SvmExitTaskSwitch(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4962	{
4963	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4964
4965	#ifndef HMSVM_ALWAYS_TRAP_TASK_SWITCH
4966	Assert(!pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
4967	#endif
4968
4969	/* Check if this task-switch occurred while delivery an event through the guest IDT. */
4970	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4971	if ( !(pVmcb->ctrl.u64ExitInfo2 & (SVM_EXIT2_TASK_SWITCH_IRET \| SVM_EXIT2_TASK_SWITCH_JMP))
4972	&& pVCpu->hm.s.Event.fPending) /** @todo fPending cannot be 'true', see hmR0SvmInjectPendingEvent(). See @bugref{7362}.*/
4973	{
4974	/*
4975	* AMD-V does not provide us with the original exception but we have it in u64IntInfo since we
4976	* injected the event during VM-entry.
4977	*/
4978	Log4(("hmR0SvmExitTaskSwitch: TS occurred during event delivery.\n"));
4979	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitTaskSwitch);
4980	return VINF_EM_RAW_INJECT_TRPM_EVENT;
4981	}
4982
4983	/** @todo Emulate task switch someday, currently just going back to ring-3 for
4984	* emulation. */
4985	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitTaskSwitch);
4986	return VERR_EM_INTERPRETER;
4987	}
4988
4989
4990	/**
4991	* #VMEXIT handler for VMMCALL (SVM_EXIT_VMMCALL). Conditional #VMEXIT.
4992	*/
4993	HMSVM_EXIT_DECL hmR0SvmExitVmmCall(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4994	{
4995	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4996
4997	/* First check if this is a patched VMMCALL for mov TPR */
4998	int rc = hmR0SvmEmulateMovTpr(pVCpu->CTX_SUFF(pVM), pVCpu, pCtx);
4999	if (rc == VINF_SUCCESS)
5000	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
5001	else if (rc == VERR_NOT_FOUND)
5002	{
5003	/* Handle GIM provider hypercalls. */
5004	rc = VERR_NOT_SUPPORTED;
5005	if (GIMAreHypercallsEnabled(pVCpu))
5006	rc = GIMHypercall(pVCpu, pCtx);
5007	}
5008
5009	if (rc != VINF_SUCCESS)
5010	hmR0SvmSetPendingXcptUD(pVCpu);
5011	return VINF_SUCCESS;
5012	}
5013
5014
5015	/**
5016	* #VMEXIT handler for IRET (SVM_EXIT_IRET). Conditional #VMEXIT.
5017	*/
5018	HMSVM_EXIT_DECL hmR0SvmExitIret(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5019	{
5020	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5021
5022	/* Clear NMI blocking. */
5023	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_BLOCK_NMIS);
5024
5025	/* Indicate that we no longer need to #VMEXIT when the guest is ready to receive NMIs, it is now ready. */
5026	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
5027	hmR0SvmClearIretIntercept(pVmcb);
5028
5029	/* Deliver the pending NMI via hmR0SvmEvaluatePendingEvent() and resume guest execution. */
5030	return VINF_SUCCESS;
5031	}
5032
5033
5034	/**
5035	* #VMEXIT handler for page-fault exceptions (SVM_EXIT_EXCEPTION_E). Conditional
5036	* #VMEXIT.
5037	*/
5038	HMSVM_EXIT_DECL hmR0SvmExitXcptPF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5039	{
5040	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5041
5042	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
5043
5044	/* See AMD spec. 15.12.15 "#PF (Page Fault)". */
5045	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
5046	uint32_t u32ErrCode = pVmcb->ctrl.u64ExitInfo1;
5047	RTGCUINTPTR uFaultAddress = pVmcb->ctrl.u64ExitInfo2;
5048	PVM pVM = pVCpu->CTX_SUFF(pVM);
5049
5050	#if defined(HMSVM_ALWAYS_TRAP_ALL_XCPTS) \|\| defined(HMSVM_ALWAYS_TRAP_PF)
5051	if (pVM->hm.s.fNestedPaging)
5052	{
5053	pVCpu->hm.s.Event.fPending = false; /* In case it's a contributory or vectoring #PF. */
5054	if (!pSvmTransient->fVectoringDoublePF)
5055	{
5056	/* A genuine guest #PF, reflect it to the guest. */
5057	hmR0SvmSetPendingXcptPF(pVCpu, pCtx, u32ErrCode, uFaultAddress);
5058	Log4(("#PF: Guest page fault at %04X:%RGv FaultAddr=%RGv ErrCode=%#x\n", pCtx->cs.Sel, (RTGCPTR)pCtx->rip,
5059	uFaultAddress, u32ErrCode));
5060	}
5061	else
5062	{
5063	/* A guest page-fault occurred during delivery of a page-fault. Inject #DF. */
5064	hmR0SvmSetPendingXcptDF(pVCpu);
5065	Log4(("Pending #DF due to vectoring #PF. NP\n"));
5066	}
5067	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestPF);
5068	return VINF_SUCCESS;
5069	}
5070	#endif
5071
5072	Assert(!pVM->hm.s.fNestedPaging);
5073
5074	#ifdef VBOX_HM_WITH_GUEST_PATCHING
5075	/* Shortcut for APIC TPR reads and writes; only applicable to 32-bit guests. */
5076	if ( pVM->hm.s.fTprPatchingAllowed
5077	&& (uFaultAddress & 0xfff) == 0x80 /* TPR offset. */
5078	&& !(u32ErrCode & X86_TRAP_PF_P) /* Not present. */
5079	&& !CPUMIsGuestInLongModeEx(pCtx)
5080	&& !CPUMGetGuestCPL(pVCpu)
5081	&& pVM->hm.s.cPatches < RT_ELEMENTS(pVM->hm.s.aPatches))
5082	{
5083	RTGCPHYS GCPhysApicBase;
5084	GCPhysApicBase = pCtx->msrApicBase;
5085	GCPhysApicBase &= PAGE_BASE_GC_MASK;
5086
5087	/* Check if the page at the fault-address is the APIC base. */
5088	RTGCPHYS GCPhysPage;
5089	int rc2 = PGMGstGetPage(pVCpu, (RTGCPTR)uFaultAddress, NULL /* pfFlags */, &GCPhysPage);
5090	if ( rc2 == VINF_SUCCESS
5091	&& GCPhysPage == GCPhysApicBase)
5092	{
5093	/* Only attempt to patch the instruction once. */
5094	PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
5095	if (!pPatch)
5096	return VINF_EM_HM_PATCH_TPR_INSTR;
5097	}
5098	}
5099	#endif
5100
5101	Log4(("#PF: uFaultAddress=%#RX64 CS:RIP=%#04x:%#RX64 u32ErrCode %#RX32 cr3=%#RX64\n", uFaultAddress, pCtx->cs.Sel,
5102	pCtx->rip, u32ErrCode, pCtx->cr3));
5103
5104	/* If it's a vectoring #PF, emulate injecting the original event injection as PGMTrap0eHandler() is incapable
5105	of differentiating between instruction emulation and event injection that caused a #PF. See @bugref{6607}. */
5106	if (pSvmTransient->fVectoringPF)
5107	{
5108	Assert(pVCpu->hm.s.Event.fPending);
5109	return VINF_EM_RAW_INJECT_TRPM_EVENT;
5110	}
5111
5112	TRPMAssertXcptPF(pVCpu, uFaultAddress, u32ErrCode);
5113	int rc = PGMTrap0eHandler(pVCpu, u32ErrCode, CPUMCTX2CORE(pCtx), (RTGCPTR)uFaultAddress);
5114
5115	Log4(("#PF rc=%Rrc\n", rc));
5116
5117	if (rc == VINF_SUCCESS)
5118	{
5119	/* Successfully synced shadow pages tables or emulated an MMIO instruction. */
5120	TRPMResetTrap(pVCpu);
5121	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPF);
5122	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
5123	return rc;
5124	}
5125	else if (rc == VINF_EM_RAW_GUEST_TRAP)
5126	{
5127	pVCpu->hm.s.Event.fPending = false; /* In case it's a contributory or vectoring #PF. */
5128
5129	if (!pSvmTransient->fVectoringDoublePF)
5130	{
5131	/* It's a guest page fault and needs to be reflected to the guest. */
5132	u32ErrCode = TRPMGetErrorCode(pVCpu); /* The error code might have been changed. */
5133	TRPMResetTrap(pVCpu);
5134	hmR0SvmSetPendingXcptPF(pVCpu, pCtx, u32ErrCode, uFaultAddress);
5135	}
5136	else
5137	{
5138	/* A guest page-fault occurred during delivery of a page-fault. Inject #DF. */
5139	TRPMResetTrap(pVCpu);
5140	hmR0SvmSetPendingXcptDF(pVCpu);
5141	Log4(("#PF: Pending #DF due to vectoring #PF\n"));
5142	}
5143
5144	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestPF);
5145	return VINF_SUCCESS;
5146	}
5147
5148	TRPMResetTrap(pVCpu);
5149	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPFEM);
5150	return rc;
5151	}
5152
5153
5154	/**
5155	* #VMEXIT handler for device-not-available exceptions (SVM_EXIT_EXCEPTION_7).
5156	* Conditional #VMEXIT.
5157	*/
5158	HMSVM_EXIT_DECL hmR0SvmExitXcptNM(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5159	{
5160	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5161
5162	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
5163
5164	/* We're playing with the host CPU state here, make sure we don't preempt or longjmp. */
5165	VMMRZCallRing3Disable(pVCpu);
5166	HM_DISABLE_PREEMPT_IF_NEEDED();
5167
5168	int rc;
5169	/* If the guest FPU was active at the time of the #NM exit, then it's a guest fault. */
5170	if (pSvmTransient->fWasGuestFPUStateActive)
5171	{
5172	rc = VINF_EM_RAW_GUEST_TRAP;
5173	Assert(CPUMIsGuestFPUStateActive(pVCpu) \|\| HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR0));
5174	}
5175	else
5176	{
5177	#ifndef HMSVM_ALWAYS_TRAP_ALL_XCPTS
5178	Assert(!pSvmTransient->fWasGuestFPUStateActive);
5179	#endif
5180	rc = CPUMR0Trap07Handler(pVCpu->CTX_SUFF(pVM), pVCpu, pCtx);
5181	Assert(rc == VINF_EM_RAW_GUEST_TRAP \|\| (rc == VINF_SUCCESS && CPUMIsGuestFPUStateActive(pVCpu)));
5182	}
5183
5184	HM_RESTORE_PREEMPT_IF_NEEDED();
5185	VMMRZCallRing3Enable(pVCpu);
5186
5187	if (rc == VINF_SUCCESS)
5188	{
5189	/* Guest FPU state was activated, we'll want to change CR0 FPU intercepts before the next VM-reentry. */
5190	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR0);
5191	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowNM);
5192	pVCpu->hm.s.fUseGuestFpu = true;
5193	}
5194	else
5195	{
5196	/* Forward #NM to the guest. */
5197	Assert(rc == VINF_EM_RAW_GUEST_TRAP);
5198	hmR0SvmSetPendingXcptNM(pVCpu);
5199	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestNM);
5200	}
5201	return VINF_SUCCESS;
5202	}
5203
5204
5205	/**
5206	* #VMEXIT handler for math-fault exceptions (SVM_EXIT_EXCEPTION_10).
5207	* Conditional #VMEXIT.
5208	*/
5209	HMSVM_EXIT_DECL hmR0SvmExitXcptMF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5210	{
5211	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5212
5213	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
5214
5215	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestMF);
5216
5217	if (!(pCtx->cr0 & X86_CR0_NE))
5218	{
5219	PVM pVM = pVCpu->CTX_SUFF(pVM);
5220	PDISSTATE pDis = &pVCpu->hm.s.DisState;
5221	unsigned cbOp;
5222	int rc = EMInterpretDisasCurrent(pVM, pVCpu, pDis, &cbOp);
5223	if (RT_SUCCESS(rc))
5224	{
5225	/* Convert a #MF into a FERR -> IRQ 13. See @bugref{6117}. */
5226	rc = PDMIsaSetIrq(pVCpu->CTX_SUFF(pVM), 13, 1, 0 /* uTagSrc */);
5227	if (RT_SUCCESS(rc))
5228	pCtx->rip += cbOp;
5229	}
5230	else
5231	Log4(("hmR0SvmExitXcptMF: EMInterpretDisasCurrent returned %Rrc uOpCode=%#x\n", rc, pDis->pCurInstr->uOpcode));
5232	return rc;
5233	}
5234
5235	hmR0SvmSetPendingXcptMF(pVCpu);
5236	return VINF_SUCCESS;
5237	}
5238
5239
5240	/**
5241	* #VMEXIT handler for debug exceptions (SVM_EXIT_EXCEPTION_1). Conditional
5242	* #VMEXIT.
5243	*/
5244	HMSVM_EXIT_DECL hmR0SvmExitXcptDB(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5245	{
5246	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5247
5248	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
5249
5250	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestDB);
5251
5252	/* This can be a fault-type #DB (instruction breakpoint) or a trap-type #DB (data breakpoint). However, for both cases
5253	DR6 and DR7 are updated to what the exception handler expects. See AMD spec. 15.12.2 "#DB (Debug)". */
5254	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
5255	PVM pVM = pVCpu->CTX_SUFF(pVM);
5256	int rc = DBGFRZTrap01Handler(pVM, pVCpu, CPUMCTX2CORE(pCtx), pVmcb->guest.u64DR6, pVCpu->hm.s.fSingleInstruction);
5257	if (rc == VINF_EM_RAW_GUEST_TRAP)
5258	{
5259	Log5(("hmR0SvmExitXcptDB: DR6=%#RX64 -> guest trap\n", pVmcb->guest.u64DR6));
5260	if (CPUMIsHyperDebugStateActive(pVCpu))
5261	CPUMSetGuestDR6(pVCpu, CPUMGetGuestDR6(pVCpu) \| pVmcb->guest.u64DR6);
5262
5263	/* Reflect the exception back to the guest. */
5264	hmR0SvmSetPendingXcptDB(pVCpu);
5265	rc = VINF_SUCCESS;
5266	}
5267
5268	/*
5269	* Update DR6.
5270	*/
5271	if (CPUMIsHyperDebugStateActive(pVCpu))
5272	{
5273	Log5(("hmR0SvmExitXcptDB: DR6=%#RX64 -> %Rrc\n", pVmcb->guest.u64DR6, rc));
5274	pVmcb->guest.u64DR6 = X86_DR6_INIT_VAL;
5275	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
5276	}
5277	else
5278	{
5279	AssertMsg(rc == VINF_SUCCESS, ("rc=%Rrc\n", rc));
5280	Assert(!pVCpu->hm.s.fSingleInstruction && !DBGFIsStepping(pVCpu));
5281	}
5282
5283	return rc;
5284	}
5285
5286	/** @} */
5287

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

source: vbox/trunk/src/VBox/VMM/VMMR0/HMSVMR0.cpp@ 54196

Download in other formats: