HMSVMR0.cpp@ 57229

Last change on this file since 57229 was 57062, checked in by vboxsync, 9 years ago
VMM/HMSVMR0: nit.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 206.8 KB

Line
1	/* $Id: HMSVMR0.cpp 57062 2015-07-23 13:49:20Z 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 <VBox/vmm/pdmapi.h>
26	#include <VBox/vmm/dbgf.h>
27	#include <VBox/vmm/iem.h>
28	#include <VBox/vmm/iom.h>
29	#include <VBox/vmm/tm.h>
30	#include <VBox/vmm/gim.h>
31	#include "HMInternal.h"
32	#include <VBox/vmm/vm.h>
33	#include "HMSVMR0.h"
34	#include "dtrace/VBoxVMM.h"
35
36	#ifdef DEBUG_ramshankar
37	# define HMSVM_SYNC_FULL_GUEST_STATE
38	# define HMSVM_ALWAYS_TRAP_ALL_XCPTS
39	# define HMSVM_ALWAYS_TRAP_PF
40	# define HMSVM_ALWAYS_TRAP_TASK_SWITCH
41	#endif
42
43
44	/*******************************************************************************
45	* Defined Constants And Macros *
46	*******************************************************************************/
47	#ifdef VBOX_WITH_STATISTICS
48	# define HMSVM_EXITCODE_STAM_COUNTER_INC(u64ExitCode) do { \
49	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitAll); \
50	if ((u64ExitCode) == SVM_EXIT_NPF) \
51	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitReasonNpf); \
52	else \
53	STAM_COUNTER_INC(&pVCpu->hm.s.paStatExitReasonR0[(u64ExitCode) & MASK_EXITREASON_STAT]); \
54	} while (0)
55	#else
56	# define HMSVM_EXITCODE_STAM_COUNTER_INC(u64ExitCode) do { } while (0)
57	#endif
58
59	/** If we decide to use a function table approach this can be useful to
60	* switch to a "static DECLCALLBACK(int)". */
61	#define HMSVM_EXIT_DECL static int
62
63	/** @name Segment attribute conversion between CPU and AMD-V VMCB format.
64	*
65	* The CPU format of the segment attribute is described in X86DESCATTRBITS
66	* which is 16-bits (i.e. includes 4 bits of the segment limit).
67	*
68	* The AMD-V VMCB format the segment attribute is compact 12-bits (strictly
69	* only the attribute bits and nothing else). Upper 4-bits are unused.
70	*
71	* @{ */
72	#define HMSVM_CPU_2_VMCB_SEG_ATTR(a) ( ((a) & 0xff) \| (((a) & 0xf000) >> 4) )
73	#define HMSVM_VMCB_2_CPU_SEG_ATTR(a) ( ((a) & 0xff) \| (((a) & 0x0f00) << 4) )
74	/** @} */
75
76	/** @name Macros for loading, storing segment registers to/from the VMCB.
77	* @{ */
78	#define HMSVM_LOAD_SEG_REG(REG, reg) \
79	do \
80	{ \
81	Assert(pCtx->reg.fFlags & CPUMSELREG_FLAGS_VALID); \
82	Assert(pCtx->reg.ValidSel == pCtx->reg.Sel); \
83	pVmcb->guest.REG.u16Sel = pCtx->reg.Sel; \
84	pVmcb->guest.REG.u32Limit = pCtx->reg.u32Limit; \
85	pVmcb->guest.REG.u64Base = pCtx->reg.u64Base; \
86	pVmcb->guest.REG.u16Attr = HMSVM_CPU_2_VMCB_SEG_ATTR(pCtx->reg.Attr.u); \
87	} while (0)
88
89	#define HMSVM_SAVE_SEG_REG(REG, reg) \
90	do \
91	{ \
92	pMixedCtx->reg.Sel = pVmcb->guest.REG.u16Sel; \
93	pMixedCtx->reg.ValidSel = pVmcb->guest.REG.u16Sel; \
94	pMixedCtx->reg.fFlags = CPUMSELREG_FLAGS_VALID; \
95	pMixedCtx->reg.u32Limit = pVmcb->guest.REG.u32Limit; \
96	pMixedCtx->reg.u64Base = pVmcb->guest.REG.u64Base; \
97	pMixedCtx->reg.Attr.u = HMSVM_VMCB_2_CPU_SEG_ATTR(pVmcb->guest.REG.u16Attr); \
98	} while (0)
99	/** @} */
100
101	/** Macro for checking and returning from the using function for
102	* \#VMEXIT intercepts that maybe caused during delivering of another
103	* event in the guest. */
104	#define HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY() \
105	do \
106	{ \
107	int rc = hmR0SvmCheckExitDueToEventDelivery(pVCpu, pCtx, pSvmTransient); \
108	if (RT_UNLIKELY(rc == VINF_HM_DOUBLE_FAULT)) \
109	return VINF_SUCCESS; \
110	else if (RT_UNLIKELY(rc == VINF_EM_RESET)) \
111	return rc; \
112	} while (0)
113
114	/** Macro for upgrading a @a a_rc to VINF_EM_DBG_STEPPED after emulating an
115	* instruction that exited. */
116	#define HMSVM_CHECK_SINGLE_STEP(a_pVCpu, a_rc) \
117	do { \
118	if ((a_pVCpu)->hm.s.fSingleInstruction && (a_rc) == VINF_SUCCESS) \
119	(a_rc) = VINF_EM_DBG_STEPPED; \
120	} while (0)
121
122	/** Assert that preemption is disabled or covered by thread-context hooks. */
123	#define HMSVM_ASSERT_PREEMPT_SAFE() Assert( VMMR0ThreadCtxHookIsEnabled(pVCpu) \
124	\|\| !RTThreadPreemptIsEnabled(NIL_RTTHREAD));
125
126	/** Assert that we haven't migrated CPUs when thread-context hooks are not
127	* used. */
128	#define HMSVM_ASSERT_CPU_SAFE() AssertMsg( VMMR0ThreadCtxHookIsEnabled(pVCpu) \
129	\|\| pVCpu->hm.s.idEnteredCpu == RTMpCpuId(), \
130	("Illegal migration! Entered on CPU %u Current %u\n", \
131	pVCpu->hm.s.idEnteredCpu, RTMpCpuId()));
132
133	/** Exception bitmap mask for all contributory exceptions.
134	*
135	* Page fault is deliberately excluded here as it's conditional as to whether
136	* it's contributory or benign. Page faults are handled separately.
137	*/
138	#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) \
139	\| RT_BIT(X86_XCPT_DE))
140
141	/** @name VMCB Clean Bits.
142	*
143	* These flags are used for VMCB-state caching. A set VMCB Clean bit indicates
144	* AMD-V doesn't need to reload the corresponding value(s) from the VMCB in
145	* memory.
146	*
147	* @{ */
148	/** All intercepts vectors, TSC offset, PAUSE filter counter. */
149	#define HMSVM_VMCB_CLEAN_INTERCEPTS RT_BIT(0)
150	/** I/O permission bitmap, MSR permission bitmap. */
151	#define HMSVM_VMCB_CLEAN_IOPM_MSRPM RT_BIT(1)
152	/** ASID. */
153	#define HMSVM_VMCB_CLEAN_ASID RT_BIT(2)
154	/** TRP: V_TPR, V_IRQ, V_INTR_PRIO, V_IGN_TPR, V_INTR_MASKING,
155	V_INTR_VECTOR. */
156	#define HMSVM_VMCB_CLEAN_TPR RT_BIT(3)
157	/** Nested Paging: Nested CR3 (nCR3), PAT. */
158	#define HMSVM_VMCB_CLEAN_NP RT_BIT(4)
159	/** Control registers (CR0, CR3, CR4, EFER). */
160	#define HMSVM_VMCB_CLEAN_CRX_EFER RT_BIT(5)
161	/** Debug registers (DR6, DR7). */
162	#define HMSVM_VMCB_CLEAN_DRX RT_BIT(6)
163	/** GDT, IDT limit and base. */
164	#define HMSVM_VMCB_CLEAN_DT RT_BIT(7)
165	/** Segment register: CS, SS, DS, ES limit and base. */
166	#define HMSVM_VMCB_CLEAN_SEG RT_BIT(8)
167	/** CR2.*/
168	#define HMSVM_VMCB_CLEAN_CR2 RT_BIT(9)
169	/** Last-branch record (DbgCtlMsr, br_from, br_to, lastint_from, lastint_to) */
170	#define HMSVM_VMCB_CLEAN_LBR RT_BIT(10)
171	/** AVIC (AVIC APIC_BAR; AVIC APIC_BACKING_PAGE, AVIC
172	PHYSICAL_TABLE and AVIC LOGICAL_TABLE Pointers). */
173	#define HMSVM_VMCB_CLEAN_AVIC RT_BIT(11)
174	/** Mask of all valid VMCB Clean bits. */
175	#define HMSVM_VMCB_CLEAN_ALL ( HMSVM_VMCB_CLEAN_INTERCEPTS \
176	\| HMSVM_VMCB_CLEAN_IOPM_MSRPM \
177	\| HMSVM_VMCB_CLEAN_ASID \
178	\| HMSVM_VMCB_CLEAN_TPR \
179	\| HMSVM_VMCB_CLEAN_NP \
180	\| HMSVM_VMCB_CLEAN_CRX_EFER \
181	\| HMSVM_VMCB_CLEAN_DRX \
182	\| HMSVM_VMCB_CLEAN_DT \
183	\| HMSVM_VMCB_CLEAN_SEG \
184	\| HMSVM_VMCB_CLEAN_CR2 \
185	\| HMSVM_VMCB_CLEAN_LBR \
186	\| HMSVM_VMCB_CLEAN_AVIC)
187	/** @} */
188
189	/** @name SVM transient.
190	*
191	* A state structure for holding miscellaneous information across AMD-V
192	* VMRUN/#VMEXIT operation, restored after the transition.
193	*
194	* @{ */
195	typedef struct SVMTRANSIENT
196	{
197	/** The host's rflags/eflags. */
198	RTCCUINTREG fEFlags;
199	#if HC_ARCH_BITS == 32
200	uint32_t u32Alignment0;
201	#endif
202
203	/** The #VMEXIT exit code (the EXITCODE field in the VMCB). */
204	uint64_t u64ExitCode;
205	/** The guest's TPR value used for TPR shadowing. */
206	uint8_t u8GuestTpr;
207	/** Alignment. */
208	uint8_t abAlignment0[7];
209
210	/** Whether the guest FPU state was active at the time of #VMEXIT. */
211	bool fWasGuestFPUStateActive;
212	/** Whether the guest debug state was active at the time of #VMEXIT. */
213	bool fWasGuestDebugStateActive;
214	/** Whether the hyper debug state was active at the time of #VMEXIT. */
215	bool fWasHyperDebugStateActive;
216	/** Whether the TSC offset mode needs to be updated. */
217	bool fUpdateTscOffsetting;
218	/** Whether the TSC_AUX MSR needs restoring on #VMEXIT. */
219	bool fRestoreTscAuxMsr;
220	/** Whether the #VMEXIT was caused by a page-fault during delivery of a
221	* contributary exception or a page-fault. */
222	bool fVectoringDoublePF;
223	/** Whether the #VMEXIT was caused by a page-fault during delivery of an
224	* external interrupt or NMI. */
225	bool fVectoringPF;
226	} SVMTRANSIENT, *PSVMTRANSIENT;
227	AssertCompileMemberAlignment(SVMTRANSIENT, u64ExitCode, sizeof(uint64_t));
228	AssertCompileMemberAlignment(SVMTRANSIENT, fWasGuestFPUStateActive, sizeof(uint64_t));
229	/** @} */
230
231	/**
232	* MSRPM (MSR permission bitmap) read permissions (for guest RDMSR).
233	*/
234	typedef enum SVMMSREXITREAD
235	{
236	/** Reading this MSR causes a #VMEXIT. */
237	SVMMSREXIT_INTERCEPT_READ = 0xb,
238	/** Reading this MSR does not cause a #VMEXIT. */
239	SVMMSREXIT_PASSTHRU_READ
240	} SVMMSREXITREAD;
241
242	/**
243	* MSRPM (MSR permission bitmap) write permissions (for guest WRMSR).
244	*/
245	typedef enum SVMMSREXITWRITE
246	{
247	/** Writing to this MSR causes a #VMEXIT. */
248	SVMMSREXIT_INTERCEPT_WRITE = 0xd,
249	/** Writing to this MSR does not cause a #VMEXIT. */
250	SVMMSREXIT_PASSTHRU_WRITE
251	} SVMMSREXITWRITE;
252
253	/**
254	* SVM #VMEXIT handler.
255	*
256	* @returns VBox status code.
257	* @param pVCpu Pointer to the VMCPU.
258	* @param pMixedCtx Pointer to the guest-CPU context.
259	* @param pSvmTransient Pointer to the SVM-transient structure.
260	*/
261	typedef int FNSVMEXITHANDLER(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
262
263	/*******************************************************************************
264	* Internal Functions *
265	*******************************************************************************/
266	static void hmR0SvmSetMsrPermission(PVMCPU pVCpu, unsigned uMsr, SVMMSREXITREAD enmRead, SVMMSREXITWRITE enmWrite);
267	static void hmR0SvmPendingEventToTrpmTrap(PVMCPU pVCpu);
268	static void hmR0SvmLeave(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx);
269
270	/** @name #VMEXIT handlers.
271	* @{
272	*/
273	static FNSVMEXITHANDLER hmR0SvmExitIntr;
274	static FNSVMEXITHANDLER hmR0SvmExitWbinvd;
275	static FNSVMEXITHANDLER hmR0SvmExitInvd;
276	static FNSVMEXITHANDLER hmR0SvmExitCpuid;
277	static FNSVMEXITHANDLER hmR0SvmExitRdtsc;
278	static FNSVMEXITHANDLER hmR0SvmExitRdtscp;
279	static FNSVMEXITHANDLER hmR0SvmExitRdpmc;
280	static FNSVMEXITHANDLER hmR0SvmExitInvlpg;
281	static FNSVMEXITHANDLER hmR0SvmExitHlt;
282	static FNSVMEXITHANDLER hmR0SvmExitMonitor;
283	static FNSVMEXITHANDLER hmR0SvmExitMwait;
284	static FNSVMEXITHANDLER hmR0SvmExitShutdown;
285	static FNSVMEXITHANDLER hmR0SvmExitReadCRx;
286	static FNSVMEXITHANDLER hmR0SvmExitWriteCRx;
287	static FNSVMEXITHANDLER hmR0SvmExitSetPendingXcptUD;
288	static FNSVMEXITHANDLER hmR0SvmExitMsr;
289	static FNSVMEXITHANDLER hmR0SvmExitReadDRx;
290	static FNSVMEXITHANDLER hmR0SvmExitWriteDRx;
291	static FNSVMEXITHANDLER hmR0SvmExitXsetbv;
292	static FNSVMEXITHANDLER hmR0SvmExitIOInstr;
293	static FNSVMEXITHANDLER hmR0SvmExitNestedPF;
294	static FNSVMEXITHANDLER hmR0SvmExitVIntr;
295	static FNSVMEXITHANDLER hmR0SvmExitTaskSwitch;
296	static FNSVMEXITHANDLER hmR0SvmExitVmmCall;
297	static FNSVMEXITHANDLER hmR0SvmExitIret;
298	static FNSVMEXITHANDLER hmR0SvmExitXcptPF;
299	static FNSVMEXITHANDLER hmR0SvmExitXcptNM;
300	static FNSVMEXITHANDLER hmR0SvmExitXcptUD;
301	static FNSVMEXITHANDLER hmR0SvmExitXcptMF;
302	static FNSVMEXITHANDLER hmR0SvmExitXcptDB;
303	/** @} */
304
305	DECLINLINE(int) hmR0SvmHandleExit(PVMCPU pVCpu, PCPUMCTX pMixedCtx, PSVMTRANSIENT pSvmTransient);
306
307	/*******************************************************************************
308	* Global Variables *
309	*******************************************************************************/
310	/** Ring-0 memory object for the IO bitmap. */
311	RTR0MEMOBJ g_hMemObjIOBitmap = NIL_RTR0MEMOBJ;
312	/** Physical address of the IO bitmap. */
313	RTHCPHYS g_HCPhysIOBitmap = 0;
314	/** Virtual address of the IO bitmap. */
315	R0PTRTYPE(void *) g_pvIOBitmap = NULL;
316
317
318	/**
319	* Sets up and activates AMD-V on the current CPU.
320	*
321	* @returns VBox status code.
322	* @param pCpu Pointer to the CPU info struct.
323	* @param pVM Pointer to the VM (can be NULL after a resume!).
324	* @param pvCpuPage Pointer to the global CPU page.
325	* @param HCPhysCpuPage Physical address of the global CPU page.
326	* @param fEnabledByHost Whether the host OS has already initialized AMD-V.
327	* @param pvArg Unused on AMD-V.
328	*/
329	VMMR0DECL(int) SVMR0EnableCpu(PHMGLOBALCPUINFO pCpu, PVM pVM, void *pvCpuPage, RTHCPHYS HCPhysCpuPage, bool fEnabledByHost,
330	void *pvArg)
331	{
332	Assert(!fEnabledByHost);
333	Assert(HCPhysCpuPage && HCPhysCpuPage != NIL_RTHCPHYS);
334	Assert(RT_ALIGN_T(HCPhysCpuPage, _4K, RTHCPHYS) == HCPhysCpuPage);
335	Assert(pvCpuPage);
336	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
337
338	NOREF(pvArg);
339	NOREF(fEnabledByHost);
340
341	/* Paranoid: Disable interrupt as, in theory, interrupt handlers might mess with EFER. */
342	RTCCUINTREG fEFlags = ASMIntDisableFlags();
343
344	/*
345	* We must turn on AMD-V and setup the host state physical address, as those MSRs are per CPU.
346	*/
347	uint64_t u64HostEfer = ASMRdMsr(MSR_K6_EFER);
348	if (u64HostEfer & MSR_K6_EFER_SVME)
349	{
350	/* If the VBOX_HWVIRTEX_IGNORE_SVM_IN_USE is active, then we blindly use AMD-V. */
351	if ( pVM
352	&& pVM->hm.s.svm.fIgnoreInUseError)
353	{
354	pCpu->fIgnoreAMDVInUseError = true;
355	}
356
357	if (!pCpu->fIgnoreAMDVInUseError)
358	{
359	ASMSetFlags(fEFlags);
360	return VERR_SVM_IN_USE;
361	}
362	}
363
364	/* Turn on AMD-V in the EFER MSR. */
365	ASMWrMsr(MSR_K6_EFER, u64HostEfer \| MSR_K6_EFER_SVME);
366
367	/* Write the physical page address where the CPU will store the host state while executing the VM. */
368	ASMWrMsr(MSR_K8_VM_HSAVE_PA, HCPhysCpuPage);
369
370	/* Restore interrupts. */
371	ASMSetFlags(fEFlags);
372
373	/*
374	* Theoretically, other hypervisors may have used ASIDs, ideally we should flush all non-zero ASIDs
375	* when enabling SVM. AMD doesn't have an SVM instruction to flush all ASIDs (flushing is done
376	* upon VMRUN). Therefore, just set the fFlushAsidBeforeUse flag which instructs hmR0SvmSetupTLB()
377	* to flush the TLB with before using a new ASID.
378	*/
379	pCpu->fFlushAsidBeforeUse = true;
380
381	/*
382	* Ensure each VCPU scheduled on this CPU gets a new VPID on resume. See @bugref{6255}.
383	*/
384	++pCpu->cTlbFlushes;
385
386	return VINF_SUCCESS;
387	}
388
389
390	/**
391	* Deactivates AMD-V on the current CPU.
392	*
393	* @returns VBox status code.
394	* @param pCpu Pointer to the CPU info struct.
395	* @param pvCpuPage Pointer to the global CPU page.
396	* @param HCPhysCpuPage Physical address of the global CPU page.
397	*/
398	VMMR0DECL(int) SVMR0DisableCpu(PHMGLOBALCPUINFO pCpu, void *pvCpuPage, RTHCPHYS HCPhysCpuPage)
399	{
400	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
401	AssertReturn( HCPhysCpuPage
402	&& HCPhysCpuPage != NIL_RTHCPHYS, VERR_INVALID_PARAMETER);
403	AssertReturn(pvCpuPage, VERR_INVALID_PARAMETER);
404	NOREF(pCpu);
405
406	/* Paranoid: Disable interrupts as, in theory, interrupt handlers might mess with EFER. */
407	RTCCUINTREG fEFlags = ASMIntDisableFlags();
408
409	/* Turn off AMD-V in the EFER MSR. */
410	uint64_t u64HostEfer = ASMRdMsr(MSR_K6_EFER);
411	ASMWrMsr(MSR_K6_EFER, u64HostEfer & ~MSR_K6_EFER_SVME);
412
413	/* Invalidate host state physical address. */
414	ASMWrMsr(MSR_K8_VM_HSAVE_PA, 0);
415
416	/* Restore interrupts. */
417	ASMSetFlags(fEFlags);
418
419	return VINF_SUCCESS;
420	}
421
422
423	/**
424	* Does global AMD-V initialization (called during module initialization).
425	*
426	* @returns VBox status code.
427	*/
428	VMMR0DECL(int) SVMR0GlobalInit(void)
429	{
430	/*
431	* Allocate 12 KB for the IO bitmap. Since this is non-optional and we always intercept all IO accesses, it's done
432	* once globally here instead of per-VM.
433	*/
434	Assert(g_hMemObjIOBitmap == NIL_RTR0MEMOBJ);
435	int rc = RTR0MemObjAllocCont(&g_hMemObjIOBitmap, 3 << PAGE_SHIFT, false /* fExecutable */);
436	if (RT_FAILURE(rc))
437	return rc;
438
439	g_pvIOBitmap = RTR0MemObjAddress(g_hMemObjIOBitmap);
440	g_HCPhysIOBitmap = RTR0MemObjGetPagePhysAddr(g_hMemObjIOBitmap, 0 /* iPage */);
441
442	/* Set all bits to intercept all IO accesses. */
443	ASMMemFill32(g_pvIOBitmap, 3 << PAGE_SHIFT, UINT32_C(0xffffffff));
444	return VINF_SUCCESS;
445	}
446
447
448	/**
449	* Does global AMD-V termination (called during module termination).
450	*/
451	VMMR0DECL(void) SVMR0GlobalTerm(void)
452	{
453	if (g_hMemObjIOBitmap != NIL_RTR0MEMOBJ)
454	{
455	RTR0MemObjFree(g_hMemObjIOBitmap, true /* fFreeMappings */);
456	g_pvIOBitmap = NULL;
457	g_HCPhysIOBitmap = 0;
458	g_hMemObjIOBitmap = NIL_RTR0MEMOBJ;
459	}
460	}
461
462
463	/**
464	* Frees any allocated per-VCPU structures for a VM.
465	*
466	* @param pVM Pointer to the VM.
467	*/
468	DECLINLINE(void) hmR0SvmFreeStructs(PVM pVM)
469	{
470	for (uint32_t i = 0; i < pVM->cCpus; i++)
471	{
472	PVMCPU pVCpu = &pVM->aCpus[i];
473	AssertPtr(pVCpu);
474
475	if (pVCpu->hm.s.svm.hMemObjVmcbHost != NIL_RTR0MEMOBJ)
476	{
477	RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjVmcbHost, false);
478	pVCpu->hm.s.svm.pvVmcbHost = 0;
479	pVCpu->hm.s.svm.HCPhysVmcbHost = 0;
480	pVCpu->hm.s.svm.hMemObjVmcbHost = NIL_RTR0MEMOBJ;
481	}
482
483	if (pVCpu->hm.s.svm.hMemObjVmcb != NIL_RTR0MEMOBJ)
484	{
485	RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjVmcb, false);
486	pVCpu->hm.s.svm.pvVmcb = 0;
487	pVCpu->hm.s.svm.HCPhysVmcb = 0;
488	pVCpu->hm.s.svm.hMemObjVmcb = NIL_RTR0MEMOBJ;
489	}
490
491	if (pVCpu->hm.s.svm.hMemObjMsrBitmap != NIL_RTR0MEMOBJ)
492	{
493	RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjMsrBitmap, false);
494	pVCpu->hm.s.svm.pvMsrBitmap = 0;
495	pVCpu->hm.s.svm.HCPhysMsrBitmap = 0;
496	pVCpu->hm.s.svm.hMemObjMsrBitmap = NIL_RTR0MEMOBJ;
497	}
498	}
499	}
500
501
502	/**
503	* Does per-VM AMD-V initialization.
504	*
505	* @returns VBox status code.
506	* @param pVM Pointer to the VM.
507	*/
508	VMMR0DECL(int) SVMR0InitVM(PVM pVM)
509	{
510	int rc = VERR_INTERNAL_ERROR_5;
511
512	/*
513	* Check for an AMD CPU erratum which requires us to flush the TLB before every world-switch.
514	*/
515	uint32_t u32Family;
516	uint32_t u32Model;
517	uint32_t u32Stepping;
518	if (HMAmdIsSubjectToErratum170(&u32Family, &u32Model, &u32Stepping))
519	{
520	Log4(("SVMR0InitVM: AMD cpu with erratum 170 family %#x model %#x stepping %#x\n", u32Family, u32Model, u32Stepping));
521	pVM->hm.s.svm.fAlwaysFlushTLB = true;
522	}
523
524	/*
525	* Initialize the R0 memory objects up-front so we can properly cleanup on allocation failures.
526	*/
527	for (VMCPUID i = 0; i < pVM->cCpus; i++)
528	{
529	PVMCPU pVCpu = &pVM->aCpus[i];
530	pVCpu->hm.s.svm.hMemObjVmcbHost = NIL_RTR0MEMOBJ;
531	pVCpu->hm.s.svm.hMemObjVmcb = NIL_RTR0MEMOBJ;
532	pVCpu->hm.s.svm.hMemObjMsrBitmap = NIL_RTR0MEMOBJ;
533	}
534
535	for (VMCPUID i = 0; i < pVM->cCpus; i++)
536	{
537	PVMCPU pVCpu = &pVM->aCpus[i];
538
539	/*
540	* Allocate one page for the host-context VM control block (VMCB). This is used for additional host-state (such as
541	* FS, GS, Kernel GS Base, etc.) apart from the host-state save area specified in MSR_K8_VM_HSAVE_PA.
542	*/
543	rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjVmcbHost, 1 << PAGE_SHIFT, false /* fExecutable */);
544	if (RT_FAILURE(rc))
545	goto failure_cleanup;
546
547	pVCpu->hm.s.svm.pvVmcbHost = RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjVmcbHost);
548	pVCpu->hm.s.svm.HCPhysVmcbHost = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjVmcbHost, 0 /* iPage */);
549	Assert(pVCpu->hm.s.svm.HCPhysVmcbHost < _4G);
550	ASMMemZeroPage(pVCpu->hm.s.svm.pvVmcbHost);
551
552	/*
553	* Allocate one page for the guest-state VMCB.
554	*/
555	rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjVmcb, 1 << PAGE_SHIFT, false /* fExecutable */);
556	if (RT_FAILURE(rc))
557	goto failure_cleanup;
558
559	pVCpu->hm.s.svm.pvVmcb = RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjVmcb);
560	pVCpu->hm.s.svm.HCPhysVmcb = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjVmcb, 0 /* iPage */);
561	Assert(pVCpu->hm.s.svm.HCPhysVmcb < _4G);
562	ASMMemZeroPage(pVCpu->hm.s.svm.pvVmcb);
563
564	/*
565	* Allocate two pages (8 KB) for the MSR permission bitmap. There doesn't seem to be a way to convince
566	* SVM to not require one.
567	*/
568	rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjMsrBitmap, 2 << PAGE_SHIFT, false /* fExecutable */);
569	if (RT_FAILURE(rc))
570	goto failure_cleanup;
571
572	pVCpu->hm.s.svm.pvMsrBitmap = RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjMsrBitmap);
573	pVCpu->hm.s.svm.HCPhysMsrBitmap = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjMsrBitmap, 0 /* iPage */);
574	/* Set all bits to intercept all MSR accesses (changed later on). */
575	ASMMemFill32(pVCpu->hm.s.svm.pvMsrBitmap, 2 << PAGE_SHIFT, UINT32_C(0xffffffff));
576	}
577
578	return VINF_SUCCESS;
579
580	failure_cleanup:
581	hmR0SvmFreeStructs(pVM);
582	return rc;
583	}
584
585
586	/**
587	* Does per-VM AMD-V termination.
588	*
589	* @returns VBox status code.
590	* @param pVM Pointer to the VM.
591	*/
592	VMMR0DECL(int) SVMR0TermVM(PVM pVM)
593	{
594	hmR0SvmFreeStructs(pVM);
595	return VINF_SUCCESS;
596	}
597
598
599	/**
600	* Sets the permission bits for the specified MSR in the MSRPM.
601	*
602	* @param pVCpu Pointer to the VMCPU.
603	* @param uMsr The MSR for which the access permissions are being set.
604	* @param enmRead MSR read permissions.
605	* @param enmWrite MSR write permissions.
606	*/
607	static void hmR0SvmSetMsrPermission(PVMCPU pVCpu, unsigned uMsr, SVMMSREXITREAD enmRead, SVMMSREXITWRITE enmWrite)
608	{
609	unsigned ulBit;
610	uint8_t pbMsrBitmap = (uint8_t )pVCpu->hm.s.svm.pvMsrBitmap;
611
612	/*
613	* Layout:
614	* Byte offset MSR range
615	* 0x000 - 0x7ff 0x00000000 - 0x00001fff
616	* 0x800 - 0xfff 0xc0000000 - 0xc0001fff
617	* 0x1000 - 0x17ff 0xc0010000 - 0xc0011fff
618	* 0x1800 - 0x1fff Reserved
619	*/
620	if (uMsr <= 0x00001FFF)
621	{
622	/* Pentium-compatible MSRs. */
623	ulBit = uMsr * 2;
624	}
625	else if ( uMsr >= 0xC0000000
626	&& uMsr <= 0xC0001FFF)
627	{
628	/* AMD Sixth Generation x86 Processor MSRs. */
629	ulBit = (uMsr - 0xC0000000) * 2;
630	pbMsrBitmap += 0x800;
631	}
632	else if ( uMsr >= 0xC0010000
633	&& uMsr <= 0xC0011FFF)
634	{
635	/* AMD Seventh and Eighth Generation Processor MSRs. */
636	ulBit = (uMsr - 0xC0001000) * 2;
637	pbMsrBitmap += 0x1000;
638	}
639	else
640	{
641	AssertFailed();
642	return;
643	}
644
645	Assert(ulBit < 0x3fff /* 16 * 1024 - 1 */);
646	if (enmRead == SVMMSREXIT_INTERCEPT_READ)
647	ASMBitSet(pbMsrBitmap, ulBit);
648	else
649	ASMBitClear(pbMsrBitmap, ulBit);
650
651	if (enmWrite == SVMMSREXIT_INTERCEPT_WRITE)
652	ASMBitSet(pbMsrBitmap, ulBit + 1);
653	else
654	ASMBitClear(pbMsrBitmap, ulBit + 1);
655
656	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
657	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_IOPM_MSRPM;
658	}
659
660
661	/**
662	* Sets up AMD-V for the specified VM.
663	* This function is only called once per-VM during initalization.
664	*
665	* @returns VBox status code.
666	* @param pVM Pointer to the VM.
667	*/
668	VMMR0DECL(int) SVMR0SetupVM(PVM pVM)
669	{
670	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
671	AssertReturn(pVM, VERR_INVALID_PARAMETER);
672	Assert(pVM->hm.s.svm.fSupported);
673
674	for (VMCPUID i = 0; i < pVM->cCpus; i++)
675	{
676	PVMCPU pVCpu = &pVM->aCpus[i];
677	PSVMVMCB pVmcb = (PSVMVMCB)pVM->aCpus[i].hm.s.svm.pvVmcb;
678
679	AssertMsgReturn(pVmcb, ("Invalid pVmcb for vcpu[%u]\n", i), VERR_SVM_INVALID_PVMCB);
680
681	/* Initialize the #VMEXIT history array with end-of-array markers (UINT16_MAX). */
682	Assert(!pVCpu->hm.s.idxExitHistoryFree);
683	HMCPU_EXIT_HISTORY_RESET(pVCpu);
684
685	/* Trap exceptions unconditionally (debug purposes). */
686	#ifdef HMSVM_ALWAYS_TRAP_PF
687	pVmcb->ctrl.u32InterceptException \|= RT_BIT(X86_XCPT_PF);
688	#endif
689	#ifdef HMSVM_ALWAYS_TRAP_ALL_XCPTS
690	/* If you add any exceptions here, make sure to update hmR0SvmHandleExit(). */
691	pVmcb->ctrl.u32InterceptException \|= 0
692	\| RT_BIT(X86_XCPT_BP)
693	\| RT_BIT(X86_XCPT_DB)
694	\| RT_BIT(X86_XCPT_DE)
695	\| RT_BIT(X86_XCPT_NM)
696	\| RT_BIT(X86_XCPT_UD)
697	\| RT_BIT(X86_XCPT_NP)
698	\| RT_BIT(X86_XCPT_SS)
699	\| RT_BIT(X86_XCPT_GP)
700	\| RT_BIT(X86_XCPT_PF)
701	\| RT_BIT(X86_XCPT_MF)
702	;
703	#endif
704
705	/* Set up unconditional intercepts and conditions. */
706	pVmcb->ctrl.u32InterceptCtrl1 = SVM_CTRL1_INTERCEPT_INTR /* External interrupt causes a #VMEXIT. */
707	\| SVM_CTRL1_INTERCEPT_NMI /* Non-maskable interrupts causes a #VMEXIT. */
708	\| SVM_CTRL1_INTERCEPT_INIT /* INIT signal causes a #VMEXIT. */
709	\| SVM_CTRL1_INTERCEPT_RDPMC /* RDPMC causes a #VMEXIT. */
710	\| SVM_CTRL1_INTERCEPT_CPUID /* CPUID causes a #VMEXIT. */
711	\| SVM_CTRL1_INTERCEPT_RSM /* RSM causes a #VMEXIT. */
712	\| SVM_CTRL1_INTERCEPT_HLT /* HLT causes a #VMEXIT. */
713	\| SVM_CTRL1_INTERCEPT_INOUT_BITMAP /* Use the IOPM to cause IOIO #VMEXITs. */
714	\| SVM_CTRL1_INTERCEPT_MSR_SHADOW /* MSR access not covered by MSRPM causes a #VMEXIT.*/
715	\| SVM_CTRL1_INTERCEPT_INVLPGA /* INVLPGA causes a #VMEXIT. */
716	\| SVM_CTRL1_INTERCEPT_SHUTDOWN /* Shutdown events causes a #VMEXIT. */
717	\| SVM_CTRL1_INTERCEPT_FERR_FREEZE; /* Intercept "freezing" during legacy FPU handling. */
718
719	pVmcb->ctrl.u32InterceptCtrl2 = SVM_CTRL2_INTERCEPT_VMRUN /* VMRUN causes a #VMEXIT. */
720	\| SVM_CTRL2_INTERCEPT_VMMCALL /* VMMCALL causes a #VMEXIT. */
721	\| SVM_CTRL2_INTERCEPT_VMLOAD /* VMLOAD causes a #VMEXIT. */
722	\| SVM_CTRL2_INTERCEPT_VMSAVE /* VMSAVE causes a #VMEXIT. */
723	\| SVM_CTRL2_INTERCEPT_STGI /* STGI causes a #VMEXIT. */
724	\| SVM_CTRL2_INTERCEPT_CLGI /* CLGI causes a #VMEXIT. */
725	\| SVM_CTRL2_INTERCEPT_SKINIT /* SKINIT causes a #VMEXIT. */
726	\| SVM_CTRL2_INTERCEPT_WBINVD /* WBINVD causes a #VMEXIT. */
727	\| SVM_CTRL2_INTERCEPT_MONITOR /* MONITOR causes a #VMEXIT. */
728	\| SVM_CTRL2_INTERCEPT_MWAIT /* MWAIT causes a #VMEXIT. */
729	\| SVM_CTRL2_INTERCEPT_XSETBV; /* XSETBV causes a #VMEXIT. */
730
731	/* CR0, CR4 reads must be intercepted, our shadow values are not necessarily the same as the guest's. */
732	pVmcb->ctrl.u16InterceptRdCRx = RT_BIT(0) \| RT_BIT(4);
733
734	/* CR0, CR4 writes must be intercepted for the same reasons as above. */
735	pVmcb->ctrl.u16InterceptWrCRx = RT_BIT(0) \| RT_BIT(4);
736
737	/* Intercept all DRx reads and writes by default. Changed later on. */
738	pVmcb->ctrl.u16InterceptRdDRx = 0xffff;
739	pVmcb->ctrl.u16InterceptWrDRx = 0xffff;
740
741	/* Virtualize masking of INTR interrupts. (reads/writes from/to CR8 go to the V_TPR register) */
742	pVmcb->ctrl.IntCtrl.n.u1VIrqMasking = 1;
743
744	/* Ignore the priority in the TPR. This is necessary for delivering PIC style (ExtInt) interrupts and we currently
745	deliver both PIC and APIC interrupts alike. See hmR0SvmInjectPendingEvent() */
746	pVmcb->ctrl.IntCtrl.n.u1IgnoreTPR = 1;
747
748	/* Set IO and MSR bitmap permission bitmap physical addresses. */
749	pVmcb->ctrl.u64IOPMPhysAddr = g_HCPhysIOBitmap;
750	pVmcb->ctrl.u64MSRPMPhysAddr = pVCpu->hm.s.svm.HCPhysMsrBitmap;
751
752	/* No LBR virtualization. */
753	pVmcb->ctrl.u64LBRVirt = 0;
754
755	/* Initially set all VMCB clean bits to 0 indicating that everything should be loaded from the VMCB in memory. */
756	pVmcb->ctrl.u64VmcbCleanBits = 0;
757
758	/* The host ASID MBZ, for the guest start with 1. */
759	pVmcb->ctrl.TLBCtrl.n.u32ASID = 1;
760
761	/*
762	* Setup the PAT MSR (applicable for Nested Paging only).
763	* The default value should be 0x0007040600070406ULL, but we want to treat all guest memory as WB,
764	* so choose type 6 for all PAT slots.
765	*/
766	pVmcb->guest.u64GPAT = UINT64_C(0x0006060606060606);
767
768	/* Setup Nested Paging. This doesn't change throughout the execution time of the VM. */
769	pVmcb->ctrl.NestedPaging.n.u1NestedPaging = pVM->hm.s.fNestedPaging;
770
771	/* Without Nested Paging, we need additionally intercepts. */
772	if (!pVM->hm.s.fNestedPaging)
773	{
774	/* CR3 reads/writes must be intercepted; our shadow values differ from the guest values. */
775	pVmcb->ctrl.u16InterceptRdCRx \|= RT_BIT(3);
776	pVmcb->ctrl.u16InterceptWrCRx \|= RT_BIT(3);
777
778	/* Intercept INVLPG and task switches (may change CR3, EFLAGS, LDT). */
779	pVmcb->ctrl.u32InterceptCtrl1 \|= SVM_CTRL1_INTERCEPT_INVLPG
780	\| SVM_CTRL1_INTERCEPT_TASK_SWITCH;
781
782	/* Page faults must be intercepted to implement shadow paging. */
783	pVmcb->ctrl.u32InterceptException \|= RT_BIT(X86_XCPT_PF);
784	}
785
786	#ifdef HMSVM_ALWAYS_TRAP_TASK_SWITCH
787	pVmcb->ctrl.u32InterceptCtrl1 \|= SVM_CTRL1_INTERCEPT_TASK_SWITCH;
788	#endif
789
790	/* Apply the exceptions intercepts needed by the GIM provider. */
791	if (pVCpu->hm.s.fGIMTrapXcptUD)
792	pVmcb->ctrl.u32InterceptException \|= RT_BIT(X86_XCPT_UD);
793
794	/*
795	* The following MSRs are saved/restored automatically during the world-switch.
796	* Don't intercept guest read/write accesses to these MSRs.
797	*/
798	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
799	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_CSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
800	hmR0SvmSetMsrPermission(pVCpu, MSR_K6_STAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
801	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_SF_MASK, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
802	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_FS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
803	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_GS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
804	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_KERNEL_GS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
805	hmR0SvmSetMsrPermission(pVCpu, MSR_IA32_SYSENTER_CS, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
806	hmR0SvmSetMsrPermission(pVCpu, MSR_IA32_SYSENTER_ESP, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
807	hmR0SvmSetMsrPermission(pVCpu, MSR_IA32_SYSENTER_EIP, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
808	}
809
810	return VINF_SUCCESS;
811	}
812
813
814	/**
815	* Invalidates a guest page by guest virtual address.
816	*
817	* @returns VBox status code.
818	* @param pVM Pointer to the VM.
819	* @param pVCpu Pointer to the VMCPU.
820	* @param GCVirt Guest virtual address of the page to invalidate.
821	*/
822	VMMR0DECL(int) SVMR0InvalidatePage(PVM pVM, PVMCPU pVCpu, RTGCPTR GCVirt)
823	{
824	AssertReturn(pVM, VERR_INVALID_PARAMETER);
825	Assert(pVM->hm.s.svm.fSupported);
826
827	bool fFlushPending = pVM->hm.s.svm.fAlwaysFlushTLB \|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_TLB_FLUSH);
828
829	/* Skip it if a TLB flush is already pending. */
830	if (!fFlushPending)
831	{
832	Log4(("SVMR0InvalidatePage %RGv\n", GCVirt));
833
834	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
835	AssertMsgReturn(pVmcb, ("Invalid pVmcb!\n"), VERR_SVM_INVALID_PVMCB);
836
837	#if HC_ARCH_BITS == 32
838	/* If we get a flush in 64-bit guest mode, then force a full TLB flush. INVLPGA takes only 32-bit addresses. */
839	if (CPUMIsGuestInLongMode(pVCpu))
840	VMCPU_FF_SET(pVCpu, VMCPU_FF_TLB_FLUSH);
841	else
842	#endif
843	{
844	SVMR0InvlpgA(GCVirt, pVmcb->ctrl.TLBCtrl.n.u32ASID);
845	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlbInvlpgVirt);
846	}
847	}
848	return VINF_SUCCESS;
849	}
850
851
852	/**
853	* Flushes the appropriate tagged-TLB entries.
854	*
855	* @param pVM Pointer to the VM.
856	* @param pVCpu Pointer to the VMCPU.
857	*/
858	static void hmR0SvmFlushTaggedTlb(PVMCPU pVCpu)
859	{
860	PVM pVM = pVCpu->CTX_SUFF(pVM);
861	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
862	PHMGLOBALCPUINFO pCpu = HMR0GetCurrentCpu();
863
864	/*
865	* Force a TLB flush for the first world switch if the current CPU differs from the one we ran on last.
866	* This can happen both for start & resume due to long jumps back to ring-3.
867	* If the TLB flush count changed, another VM (VCPU rather) has hit the ASID limit while flushing the TLB,
868	* so we cannot reuse the ASIDs without flushing.
869	*/
870	bool fNewAsid = false;
871	Assert(pCpu->idCpu != NIL_RTCPUID);
872	if ( pVCpu->hm.s.idLastCpu != pCpu->idCpu
873	\|\| pVCpu->hm.s.cTlbFlushes != pCpu->cTlbFlushes)
874	{
875	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlbWorldSwitch);
876	pVCpu->hm.s.fForceTLBFlush = true;
877	fNewAsid = true;
878	}
879
880	/* Set TLB flush state as checked until we return from the world switch. */
881	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, true);
882
883	/* Check for explicit TLB shootdowns. */
884	if (VMCPU_FF_TEST_AND_CLEAR(pVCpu, VMCPU_FF_TLB_FLUSH))
885	{
886	pVCpu->hm.s.fForceTLBFlush = true;
887	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlb);
888	}
889
890	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_NOTHING;
891
892	if (pVM->hm.s.svm.fAlwaysFlushTLB)
893	{
894	/*
895	* This is the AMD erratum 170. We need to flush the entire TLB for each world switch. Sad.
896	*/
897	pCpu->uCurrentAsid = 1;
898	pVCpu->hm.s.uCurrentAsid = 1;
899	pVCpu->hm.s.cTlbFlushes = pCpu->cTlbFlushes;
900	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
901
902	/* Clear the VMCB Clean Bit for NP while flushing the TLB. See @bugref{7152}. */
903	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
904	}
905	else if (pVCpu->hm.s.fForceTLBFlush)
906	{
907	/* Clear the VMCB Clean Bit for NP while flushing the TLB. See @bugref{7152}. */
908	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
909
910	if (fNewAsid)
911	{
912	++pCpu->uCurrentAsid;
913	bool fHitASIDLimit = false;
914	if (pCpu->uCurrentAsid >= pVM->hm.s.uMaxAsid)
915	{
916	pCpu->uCurrentAsid = 1; /* Wraparound at 1; host uses 0 */
917	pCpu->cTlbFlushes++; /* All VCPUs that run on this host CPU must use a new VPID. */
918	fHitASIDLimit = true;
919
920	if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_FLUSH_BY_ASID)
921	{
922	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_SINGLE_CONTEXT;
923	pCpu->fFlushAsidBeforeUse = true;
924	}
925	else
926	{
927	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
928	pCpu->fFlushAsidBeforeUse = false;
929	}
930	}
931
932	if ( !fHitASIDLimit
933	&& pCpu->fFlushAsidBeforeUse)
934	{
935	if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_FLUSH_BY_ASID)
936	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_SINGLE_CONTEXT;
937	else
938	{
939	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
940	pCpu->fFlushAsidBeforeUse = false;
941	}
942	}
943
944	pVCpu->hm.s.uCurrentAsid = pCpu->uCurrentAsid;
945	pVCpu->hm.s.idLastCpu = pCpu->idCpu;
946	pVCpu->hm.s.cTlbFlushes = pCpu->cTlbFlushes;
947	}
948	else
949	{
950	if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_FLUSH_BY_ASID)
951	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_SINGLE_CONTEXT;
952	else
953	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
954	}
955
956	pVCpu->hm.s.fForceTLBFlush = false;
957	}
958	/** @todo We never set VMCPU_FF_TLB_SHOOTDOWN anywhere so this path should
959	* not be executed. See hmQueueInvlPage() where it is commented
960	* out. Support individual entry flushing someday. */
961	#if 0
962	else
963	{
964	if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_TLB_SHOOTDOWN))
965	{
966	/* Deal with pending TLB shootdown actions which were queued when we were not executing code. */
967	STAM_COUNTER_INC(&pVCpu->hm.s.StatTlbShootdown);
968	for (uint32_t i = 0; i < pVCpu->hm.s.TlbShootdown.cPages; i++)
969	SVMR0InvlpgA(pVCpu->hm.s.TlbShootdown.aPages[i], pVmcb->ctrl.TLBCtrl.n.u32ASID);
970
971	pVCpu->hm.s.TlbShootdown.cPages = 0;
972	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_TLB_SHOOTDOWN);
973	}
974	}
975	#endif
976
977
978	/* Update VMCB with the ASID. */
979	if (pVmcb->ctrl.TLBCtrl.n.u32ASID != pVCpu->hm.s.uCurrentAsid)
980	{
981	pVmcb->ctrl.TLBCtrl.n.u32ASID = pVCpu->hm.s.uCurrentAsid;
982	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_ASID;
983	}
984
985	AssertMsg(pVCpu->hm.s.idLastCpu == pCpu->idCpu,
986	("vcpu idLastCpu=%u pcpu idCpu=%u\n", pVCpu->hm.s.idLastCpu, pCpu->idCpu));
987	AssertMsg(pVCpu->hm.s.cTlbFlushes == pCpu->cTlbFlushes,
988	("Flush count mismatch for cpu %u (%u vs %u)\n", pCpu->idCpu, pVCpu->hm.s.cTlbFlushes, pCpu->cTlbFlushes));
989	AssertMsg(pCpu->uCurrentAsid >= 1 && pCpu->uCurrentAsid < pVM->hm.s.uMaxAsid,
990	("cpu%d uCurrentAsid = %x\n", pCpu->idCpu, pCpu->uCurrentAsid));
991	AssertMsg(pVCpu->hm.s.uCurrentAsid >= 1 && pVCpu->hm.s.uCurrentAsid < pVM->hm.s.uMaxAsid,
992	("cpu%d VM uCurrentAsid = %x\n", pCpu->idCpu, pVCpu->hm.s.uCurrentAsid));
993
994	#ifdef VBOX_WITH_STATISTICS
995	if (pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_NOTHING)
996	STAM_COUNTER_INC(&pVCpu->hm.s.StatNoFlushTlbWorldSwitch);
997	else if ( pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_SINGLE_CONTEXT
998	\|\| pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_SINGLE_CONTEXT_RETAIN_GLOBALS)
999	{
1000	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushAsid);
1001	}
1002	else
1003	{
1004	Assert(pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_ENTIRE);
1005	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushEntire);
1006	}
1007	#endif
1008	}
1009
1010
1011	/** @name 64-bit guest on 32-bit host OS helper functions.
1012	*
1013	* The host CPU is still 64-bit capable but the host OS is running in 32-bit
1014	* mode (code segment, paging). These wrappers/helpers perform the necessary
1015	* bits for the 32->64 switcher.
1016	*
1017	* @{ */
1018	#if HC_ARCH_BITS == 32 && defined(VBOX_ENABLE_64_BITS_GUESTS) && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
1019	/**
1020	* Prepares for and executes VMRUN (64-bit guests on a 32-bit host).
1021	*
1022	* @returns VBox status code.
1023	* @param HCPhysVmcbHost Physical address of host VMCB.
1024	* @param HCPhysVmcb Physical address of the VMCB.
1025	* @param pCtx Pointer to the guest-CPU context.
1026	* @param pVM Pointer to the VM.
1027	* @param pVCpu Pointer to the VMCPU.
1028	*/
1029	DECLASM(int) SVMR0VMSwitcherRun64(RTHCPHYS HCPhysVmcbHost, RTHCPHYS HCPhysVmcb, PCPUMCTX pCtx, PVM pVM, PVMCPU pVCpu)
1030	{
1031	uint32_t aParam[8];
1032	aParam[0] = (uint32_t)(HCPhysVmcbHost); /* Param 1: HCPhysVmcbHost - Lo. */
1033	aParam[1] = (uint32_t)(HCPhysVmcbHost >> 32); /* Param 1: HCPhysVmcbHost - Hi. */
1034	aParam[2] = (uint32_t)(HCPhysVmcb); /* Param 2: HCPhysVmcb - Lo. */
1035	aParam[3] = (uint32_t)(HCPhysVmcb >> 32); /* Param 2: HCPhysVmcb - Hi. */
1036	aParam[4] = VM_RC_ADDR(pVM, pVM);
1037	aParam[5] = 0;
1038	aParam[6] = VM_RC_ADDR(pVM, pVCpu);
1039	aParam[7] = 0;
1040
1041	return SVMR0Execute64BitsHandler(pVM, pVCpu, pCtx, HM64ON32OP_SVMRCVMRun64, RT_ELEMENTS(aParam), &aParam[0]);
1042	}
1043
1044
1045	/**
1046	* Executes the specified VMRUN handler in 64-bit mode.
1047	*
1048	* @returns VBox status code.
1049	* @param pVM Pointer to the VM.
1050	* @param pVCpu Pointer to the VMCPU.
1051	* @param pCtx Pointer to the guest-CPU context.
1052	* @param enmOp The operation to perform.
1053	* @param cParams Number of parameters.
1054	* @param paParam Array of 32-bit parameters.
1055	*/
1056	VMMR0DECL(int) SVMR0Execute64BitsHandler(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, HM64ON32OP enmOp,
1057	uint32_t cParams, uint32_t *paParam)
1058	{
1059	AssertReturn(pVM->hm.s.pfnHost32ToGuest64R0, VERR_HM_NO_32_TO_64_SWITCHER);
1060	Assert(enmOp > HM64ON32OP_INVALID && enmOp < HM64ON32OP_END);
1061
1062	/* Disable interrupts. */
1063	RTHCUINTREG uOldEFlags = ASMIntDisableFlags();
1064
1065	#ifdef VBOX_WITH_VMMR0_DISABLE_LAPIC_NMI
1066	RTCPUID idHostCpu = RTMpCpuId();
1067	CPUMR0SetLApic(pVCpu, idHostCpu);
1068	#endif
1069
1070	CPUMSetHyperESP(pVCpu, VMMGetStackRC(pVCpu));
1071	CPUMSetHyperEIP(pVCpu, enmOp);
1072	for (int i = (int)cParams - 1; i >= 0; i--)
1073	CPUMPushHyper(pVCpu, paParam[i]);
1074
1075	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatWorldSwitch3264, z);
1076	/* Call the switcher. */
1077	int rc = pVM->hm.s.pfnHost32ToGuest64R0(pVM, RT_OFFSETOF(VM, aCpus[pVCpu->idCpu].cpum) - RT_OFFSETOF(VM, cpum));
1078	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatWorldSwitch3264, z);
1079
1080	/* Restore interrupts. */
1081	ASMSetFlags(uOldEFlags);
1082	return rc;
1083	}
1084
1085	#endif /* HC_ARCH_BITS == 32 && defined(VBOX_ENABLE_64_BITS_GUESTS) */
1086	/** @} */
1087
1088
1089	/**
1090	* Adds an exception to the intercept exception bitmap in the VMCB and updates
1091	* the corresponding VMCB Clean bit.
1092	*
1093	* @param pVmcb Pointer to the VM control block.
1094	* @param u32Xcpt The value of the exception (X86_XCPT_*).
1095	*/
1096	DECLINLINE(void) hmR0SvmAddXcptIntercept(PSVMVMCB pVmcb, uint32_t u32Xcpt)
1097	{
1098	if (!(pVmcb->ctrl.u32InterceptException & RT_BIT(u32Xcpt)))
1099	{
1100	pVmcb->ctrl.u32InterceptException \|= RT_BIT(u32Xcpt);
1101	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1102	}
1103	}
1104
1105
1106	/**
1107	* Removes an exception from the intercept-exception bitmap in the VMCB and
1108	* updates the corresponding VMCB Clean bit.
1109	*
1110	* @param pVmcb Pointer to the VM control block.
1111	* @param u32Xcpt The value of the exception (X86_XCPT_*).
1112	*/
1113	DECLINLINE(void) hmR0SvmRemoveXcptIntercept(PSVMVMCB pVmcb, uint32_t u32Xcpt)
1114	{
1115	#ifndef HMSVM_ALWAYS_TRAP_ALL_XCPTS
1116	if (pVmcb->ctrl.u32InterceptException & RT_BIT(u32Xcpt))
1117	{
1118	pVmcb->ctrl.u32InterceptException &= ~RT_BIT(u32Xcpt);
1119	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1120	}
1121	#endif
1122	}
1123
1124
1125	/**
1126	* Loads the guest CR0 control register into the guest-state area in the VMCB.
1127	* Although the guest CR0 is a separate field in the VMCB we have to consider
1128	* the FPU state itself which is shared between the host and the guest.
1129	*
1130	* @returns VBox status code.
1131	* @param pVM Pointer to the VMCPU.
1132	* @param pVmcb Pointer to the VM control block.
1133	* @param pCtx Pointer to the guest-CPU context.
1134	*
1135	* @remarks No-long-jump zone!!!
1136	*/
1137	static void hmR0SvmLoadSharedCR0(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1138	{
1139	/*
1140	* Guest CR0.
1141	*/
1142	PVM pVM = pVCpu->CTX_SUFF(pVM);
1143	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR0))
1144	{
1145	uint64_t u64GuestCR0 = pCtx->cr0;
1146
1147	/* Always enable caching. */
1148	u64GuestCR0 &= ~(X86_CR0_CD \| X86_CR0_NW);
1149
1150	/*
1151	* When Nested Paging is not available use shadow page tables and intercept #PFs (the latter done in SVMR0SetupVM()).
1152	*/
1153	if (!pVM->hm.s.fNestedPaging)
1154	{
1155	u64GuestCR0 \|= X86_CR0_PG; /* When Nested Paging is not available, use shadow page tables. */
1156	u64GuestCR0 \|= X86_CR0_WP; /* Guest CPL 0 writes to its read-only pages should cause a #PF #VMEXIT. */
1157	}
1158
1159	/*
1160	* Guest FPU bits.
1161	*/
1162	bool fInterceptNM = false;
1163	bool fInterceptMF = false;
1164	u64GuestCR0 \|= X86_CR0_NE; /* Use internal x87 FPU exceptions handling rather than external interrupts. */
1165	if (CPUMIsGuestFPUStateActive(pVCpu))
1166	{
1167	/* Catch floating point exceptions if we need to report them to the guest in a different way. */
1168	if (!(pCtx->cr0 & X86_CR0_NE))
1169	{
1170	Log4(("hmR0SvmLoadGuestControlRegs: Intercepting Guest CR0.MP Old-style FPU handling!!!\n"));
1171	fInterceptMF = true;
1172	}
1173	}
1174	else
1175	{
1176	fInterceptNM = true; /* Guest FPU inactive, #VMEXIT on #NM for lazy FPU loading. */
1177	u64GuestCR0 \|= X86_CR0_TS /* Guest can task switch quickly and do lazy FPU syncing. */
1178	\| X86_CR0_MP; /* FWAIT/WAIT should not ignore CR0.TS and should generate #NM. */
1179	}
1180
1181	/*
1182	* Update the exception intercept bitmap.
1183	*/
1184	if (fInterceptNM)
1185	hmR0SvmAddXcptIntercept(pVmcb, X86_XCPT_NM);
1186	else
1187	hmR0SvmRemoveXcptIntercept(pVmcb, X86_XCPT_NM);
1188
1189	if (fInterceptMF)
1190	hmR0SvmAddXcptIntercept(pVmcb, X86_XCPT_MF);
1191	else
1192	hmR0SvmRemoveXcptIntercept(pVmcb, X86_XCPT_MF);
1193
1194	pVmcb->guest.u64CR0 = u64GuestCR0;
1195	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1196	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_CR0);
1197	}
1198	}
1199
1200
1201	/**
1202	* Loads the guest control registers (CR2, CR3, CR4) into the VMCB.
1203	*
1204	* @returns VBox status code.
1205	* @param pVCpu Pointer to the VMCPU.
1206	* @param pVmcb Pointer to the VM control block.
1207	* @param pCtx Pointer to the guest-CPU context.
1208	*
1209	* @remarks No-long-jump zone!!!
1210	*/
1211	static int hmR0SvmLoadGuestControlRegs(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1212	{
1213	PVM pVM = pVCpu->CTX_SUFF(pVM);
1214
1215	/*
1216	* Guest CR2.
1217	*/
1218	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR2))
1219	{
1220	pVmcb->guest.u64CR2 = pCtx->cr2;
1221	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CR2;
1222	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_CR2);
1223	}
1224
1225	/*
1226	* Guest CR3.
1227	*/
1228	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR3))
1229	{
1230	if (pVM->hm.s.fNestedPaging)
1231	{
1232	PGMMODE enmShwPagingMode;
1233	#if HC_ARCH_BITS == 32
1234	if (CPUMIsGuestInLongModeEx(pCtx))
1235	enmShwPagingMode = PGMMODE_AMD64_NX;
1236	else
1237	#endif
1238	enmShwPagingMode = PGMGetHostMode(pVM);
1239
1240	pVmcb->ctrl.u64NestedPagingCR3 = PGMGetNestedCR3(pVCpu, enmShwPagingMode);
1241	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
1242	Assert(pVmcb->ctrl.u64NestedPagingCR3);
1243	pVmcb->guest.u64CR3 = pCtx->cr3;
1244	}
1245	else
1246	pVmcb->guest.u64CR3 = PGMGetHyperCR3(pVCpu);
1247
1248	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1249	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_CR3);
1250	}
1251
1252	/*
1253	* Guest CR4.
1254	* ASSUMES this is done everytime we get in from ring-3! (XCR0)
1255	*/
1256	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR4))
1257	{
1258	uint64_t u64GuestCR4 = pCtx->cr4;
1259	if (!pVM->hm.s.fNestedPaging)
1260	{
1261	switch (pVCpu->hm.s.enmShadowMode)
1262	{
1263	case PGMMODE_REAL:
1264	case PGMMODE_PROTECTED: /* Protected mode, no paging. */
1265	AssertFailed();
1266	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1267
1268	case PGMMODE_32_BIT: /* 32-bit paging. */
1269	u64GuestCR4 &= ~X86_CR4_PAE;
1270	break;
1271
1272	case PGMMODE_PAE: /* PAE paging. */
1273	case PGMMODE_PAE_NX: /* PAE paging with NX enabled. */
1274	/** Must use PAE paging as we could use physical memory > 4 GB */
1275	u64GuestCR4 \|= X86_CR4_PAE;
1276	break;
1277
1278	case PGMMODE_AMD64: /* 64-bit AMD paging (long mode). */
1279	case PGMMODE_AMD64_NX: /* 64-bit AMD paging (long mode) with NX enabled. */
1280	#ifdef VBOX_ENABLE_64_BITS_GUESTS
1281	break;
1282	#else
1283	AssertFailed();
1284	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1285	#endif
1286
1287	default: /* shut up gcc */
1288	AssertFailed();
1289	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1290	}
1291	}
1292
1293	pVmcb->guest.u64CR4 = u64GuestCR4;
1294	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1295
1296	/* Whether to save/load/restore XCR0 during world switch depends on CR4.OSXSAVE and host+guest XCR0. */
1297	pVCpu->hm.s.fLoadSaveGuestXcr0 = (u64GuestCR4 & X86_CR4_OSXSAVE) && pCtx->aXcr[0] != ASMGetXcr0();
1298
1299	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_CR4);
1300	}
1301
1302	return VINF_SUCCESS;
1303	}
1304
1305
1306	/**
1307	* Loads the guest segment registers into the VMCB.
1308	*
1309	* @returns VBox status code.
1310	* @param pVCpu Pointer to the VMCPU.
1311	* @param pVmcb Pointer to the VM control block.
1312	* @param pCtx Pointer to the guest-CPU context.
1313	*
1314	* @remarks No-long-jump zone!!!
1315	*/
1316	static void hmR0SvmLoadGuestSegmentRegs(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1317	{
1318	/* Guest Segment registers: CS, SS, DS, ES, FS, GS. */
1319	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_SEGMENT_REGS))
1320	{
1321	HMSVM_LOAD_SEG_REG(CS, cs);
1322	HMSVM_LOAD_SEG_REG(SS, ss);
1323	HMSVM_LOAD_SEG_REG(DS, ds);
1324	HMSVM_LOAD_SEG_REG(ES, es);
1325	HMSVM_LOAD_SEG_REG(FS, fs);
1326	HMSVM_LOAD_SEG_REG(GS, gs);
1327
1328	pVmcb->guest.u8CPL = pCtx->ss.Attr.n.u2Dpl;
1329	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_SEG;
1330	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_SEGMENT_REGS);
1331	}
1332
1333	/* Guest TR. */
1334	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_TR))
1335	{
1336	HMSVM_LOAD_SEG_REG(TR, tr);
1337	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_TR);
1338	}
1339
1340	/* Guest LDTR. */
1341	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_LDTR))
1342	{
1343	HMSVM_LOAD_SEG_REG(LDTR, ldtr);
1344	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_LDTR);
1345	}
1346
1347	/* Guest GDTR. */
1348	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_GDTR))
1349	{
1350	pVmcb->guest.GDTR.u32Limit = pCtx->gdtr.cbGdt;
1351	pVmcb->guest.GDTR.u64Base = pCtx->gdtr.pGdt;
1352	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DT;
1353	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_GDTR);
1354	}
1355
1356	/* Guest IDTR. */
1357	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_IDTR))
1358	{
1359	pVmcb->guest.IDTR.u32Limit = pCtx->idtr.cbIdt;
1360	pVmcb->guest.IDTR.u64Base = pCtx->idtr.pIdt;
1361	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DT;
1362	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_IDTR);
1363	}
1364	}
1365
1366
1367	/**
1368	* Loads the guest MSRs into the VMCB.
1369	*
1370	* @param pVCpu Pointer to the VMCPU.
1371	* @param pVmcb Pointer to the VM control block.
1372	* @param pCtx Pointer to the guest-CPU context.
1373	*
1374	* @remarks No-long-jump zone!!!
1375	*/
1376	static void hmR0SvmLoadGuestMsrs(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1377	{
1378	/* Guest Sysenter MSRs. */
1379	pVmcb->guest.u64SysEnterCS = pCtx->SysEnter.cs;
1380	pVmcb->guest.u64SysEnterEIP = pCtx->SysEnter.eip;
1381	pVmcb->guest.u64SysEnterESP = pCtx->SysEnter.esp;
1382
1383	/*
1384	* Guest EFER MSR.
1385	* AMD-V requires guest EFER.SVME to be set. Weird.
1386	* See AMD spec. 15.5.1 "Basic Operation" \| "Canonicalization and Consistency Checks".
1387	*/
1388	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_EFER_MSR))
1389	{
1390	pVmcb->guest.u64EFER = pCtx->msrEFER \| MSR_K6_EFER_SVME;
1391	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1392	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_EFER_MSR);
1393	}
1394
1395	/* 64-bit MSRs. */
1396	if (CPUMIsGuestInLongModeEx(pCtx))
1397	{
1398	pVmcb->guest.FS.u64Base = pCtx->fs.u64Base;
1399	pVmcb->guest.GS.u64Base = pCtx->gs.u64Base;
1400	}
1401	else
1402	{
1403	/* If the guest isn't in 64-bit mode, clear MSR_K6_LME bit from guest EFER otherwise AMD-V expects amd64 shadow paging. */
1404	if (pCtx->msrEFER & MSR_K6_EFER_LME)
1405	{
1406	pVmcb->guest.u64EFER &= ~MSR_K6_EFER_LME;
1407	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1408	}
1409	}
1410
1411
1412	/** @todo The following are used in 64-bit only (SYSCALL/SYSRET) but they might
1413	* be writable in 32-bit mode. Clarify with AMD spec. */
1414	pVmcb->guest.u64STAR = pCtx->msrSTAR;
1415	pVmcb->guest.u64LSTAR = pCtx->msrLSTAR;
1416	pVmcb->guest.u64CSTAR = pCtx->msrCSTAR;
1417	pVmcb->guest.u64SFMASK = pCtx->msrSFMASK;
1418	pVmcb->guest.u64KernelGSBase = pCtx->msrKERNELGSBASE;
1419	}
1420
1421
1422	/**
1423	* Loads the guest state into the VMCB and programs the necessary intercepts
1424	* accordingly.
1425	*
1426	* @param pVCpu Pointer to the VMCPU.
1427	* @param pVmcb Pointer to the VM control block.
1428	* @param pCtx Pointer to the guest-CPU context.
1429	*
1430	* @remarks No-long-jump zone!!!
1431	* @remarks Requires EFLAGS to be up-to-date in the VMCB!
1432	*/
1433	static void hmR0SvmLoadSharedDebugState(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1434	{
1435	if (!HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_DEBUG))
1436	return;
1437	Assert((pCtx->dr[6] & X86_DR6_RA1_MASK) == X86_DR6_RA1_MASK); Assert((pCtx->dr[6] & X86_DR6_RAZ_MASK) == 0);
1438	Assert((pCtx->dr[7] & X86_DR7_RA1_MASK) == X86_DR7_RA1_MASK); Assert((pCtx->dr[7] & X86_DR7_RAZ_MASK) == 0);
1439
1440	bool fInterceptDB = false;
1441	bool fInterceptMovDRx = false;
1442
1443	/*
1444	* Anyone single stepping on the host side? If so, we'll have to use the
1445	* trap flag in the guest EFLAGS since AMD-V doesn't have a trap flag on
1446	* the VMM level like the VT-x implementations does.
1447	*/
1448	bool const fStepping = pVCpu->hm.s.fSingleInstruction \|\| DBGFIsStepping(pVCpu);
1449	if (fStepping)
1450	{
1451	pVCpu->hm.s.fClearTrapFlag = true;
1452	pVmcb->guest.u64RFlags \|= X86_EFL_TF;
1453	fInterceptDB = true;
1454	fInterceptMovDRx = true; /* Need clean DR6, no guest mess. */
1455	}
1456
1457	if ( fStepping
1458	\|\| (CPUMGetHyperDR7(pVCpu) & X86_DR7_ENABLED_MASK))
1459	{
1460	/*
1461	* Use the combined guest and host DRx values found in the hypervisor
1462	* register set because the debugger has breakpoints active or someone
1463	* is single stepping on the host side.
1464	*
1465	* Note! DBGF expects a clean DR6 state before executing guest code.
1466	*/
1467	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS) && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
1468	if ( CPUMIsGuestInLongModeEx(pCtx)
1469	&& !CPUMIsHyperDebugStateActivePending(pVCpu))
1470	{
1471	CPUMR0LoadHyperDebugState(pVCpu, false /* include DR6 */);
1472	Assert(!CPUMIsGuestDebugStateActivePending(pVCpu));
1473	Assert(CPUMIsHyperDebugStateActivePending(pVCpu));
1474	}
1475	else
1476	#endif
1477	if (!CPUMIsHyperDebugStateActive(pVCpu))
1478	{
1479	CPUMR0LoadHyperDebugState(pVCpu, false /* include DR6 */);
1480	Assert(!CPUMIsGuestDebugStateActive(pVCpu));
1481	Assert(CPUMIsHyperDebugStateActive(pVCpu));
1482	}
1483
1484	/* Update DR6 & DR7. (The other DRx values are handled by CPUM one way or the other.) */
1485	if ( pVmcb->guest.u64DR6 != X86_DR6_INIT_VAL
1486	\|\| pVmcb->guest.u64DR7 != CPUMGetHyperDR7(pVCpu))
1487	{
1488	pVmcb->guest.u64DR7 = CPUMGetHyperDR7(pVCpu);
1489	pVmcb->guest.u64DR6 = X86_DR6_INIT_VAL;
1490	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
1491	pVCpu->hm.s.fUsingHyperDR7 = true;
1492	}
1493
1494	/** @todo If we cared, we could optimize to allow the guest to read registers
1495	* with the same values. */
1496	fInterceptDB = true;
1497	fInterceptMovDRx = true;
1498	Log5(("hmR0SvmLoadSharedDebugState: Loaded hyper DRx\n"));
1499	}
1500	else
1501	{
1502	/*
1503	* Update DR6, DR7 with the guest values if necessary.
1504	*/
1505	if ( pVmcb->guest.u64DR7 != pCtx->dr[7]
1506	\|\| pVmcb->guest.u64DR6 != pCtx->dr[6])
1507	{
1508	pVmcb->guest.u64DR7 = pCtx->dr[7];
1509	pVmcb->guest.u64DR6 = pCtx->dr[6];
1510	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
1511	pVCpu->hm.s.fUsingHyperDR7 = false;
1512	}
1513
1514	/*
1515	* If the guest has enabled debug registers, we need to load them prior to
1516	* executing guest code so they'll trigger at the right time.
1517	*/
1518	if (pCtx->dr[7] & (X86_DR7_ENABLED_MASK \| X86_DR7_GD)) /** @todo Why GD? */
1519	{
1520	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS) && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
1521	if ( CPUMIsGuestInLongModeEx(pCtx)
1522	&& !CPUMIsGuestDebugStateActivePending(pVCpu))
1523	{
1524	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
1525	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxArmed);
1526	Assert(!CPUMIsHyperDebugStateActivePending(pVCpu));
1527	Assert(CPUMIsGuestDebugStateActivePending(pVCpu));
1528	}
1529	else
1530	#endif
1531	if (!CPUMIsGuestDebugStateActive(pVCpu))
1532	{
1533	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
1534	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxArmed);
1535	Assert(!CPUMIsHyperDebugStateActive(pVCpu));
1536	Assert(CPUMIsGuestDebugStateActive(pVCpu));
1537	}
1538	Log5(("hmR0SvmLoadSharedDebugState: Loaded guest DRx\n"));
1539	}
1540	/*
1541	* If no debugging enabled, we'll lazy load DR0-3. We don't need to
1542	* intercept #DB as DR6 is updated in the VMCB.
1543	*/
1544	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS) && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
1545	else if ( !CPUMIsGuestDebugStateActivePending(pVCpu)
1546	&& !CPUMIsGuestDebugStateActive(pVCpu))
1547	#else
1548	else if (!CPUMIsGuestDebugStateActive(pVCpu))
1549	#endif
1550	{
1551	fInterceptMovDRx = true;
1552	}
1553	}
1554
1555	/*
1556	* Set up the intercepts.
1557	*/
1558	if (fInterceptDB)
1559	hmR0SvmAddXcptIntercept(pVmcb, X86_XCPT_DB);
1560	else
1561	hmR0SvmRemoveXcptIntercept(pVmcb, X86_XCPT_DB);
1562
1563	if (fInterceptMovDRx)
1564	{
1565	if ( pVmcb->ctrl.u16InterceptRdDRx != 0xffff
1566	\|\| pVmcb->ctrl.u16InterceptWrDRx != 0xffff)
1567	{
1568	pVmcb->ctrl.u16InterceptRdDRx = 0xffff;
1569	pVmcb->ctrl.u16InterceptWrDRx = 0xffff;
1570	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1571	}
1572	}
1573	else
1574	{
1575	if ( pVmcb->ctrl.u16InterceptRdDRx
1576	\|\| pVmcb->ctrl.u16InterceptWrDRx)
1577	{
1578	pVmcb->ctrl.u16InterceptRdDRx = 0;
1579	pVmcb->ctrl.u16InterceptWrDRx = 0;
1580	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1581	}
1582	}
1583
1584	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_DEBUG);
1585	}
1586
1587
1588	/**
1589	* Loads the guest APIC state (currently just the TPR).
1590	*
1591	* @returns VBox status code.
1592	* @param pVCpu Pointer to the VMCPU.
1593	* @param pVmcb Pointer to the VM control block.
1594	* @param pCtx Pointer to the guest-CPU context.
1595	*/
1596	static int hmR0SvmLoadGuestApicState(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1597	{
1598	if (!HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE))
1599	return VINF_SUCCESS;
1600
1601	bool fPendingIntr;
1602	uint8_t u8Tpr;
1603	int rc = PDMApicGetTPR(pVCpu, &u8Tpr, &fPendingIntr, NULL /* pu8PendingIrq */);
1604	AssertRCReturn(rc, rc);
1605
1606	/* Assume that we need to trap all TPR accesses and thus need not check on
1607	every #VMEXIT if we should update the TPR. */
1608	Assert(pVmcb->ctrl.IntCtrl.n.u1VIrqMasking);
1609	pVCpu->hm.s.svm.fSyncVTpr = false;
1610
1611	/* 32-bit guests uses LSTAR MSR for patching guest code which touches the TPR. */
1612	if (pVCpu->CTX_SUFF(pVM)->hm.s.fTPRPatchingActive)
1613	{
1614	pCtx->msrLSTAR = u8Tpr;
1615
1616	/* If there are interrupts pending, intercept LSTAR writes, otherwise don't intercept reads or writes. */
1617	if (fPendingIntr)
1618	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_INTERCEPT_WRITE);
1619	else
1620	{
1621	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1622	pVCpu->hm.s.svm.fSyncVTpr = true;
1623	}
1624	}
1625	else
1626	{
1627	/* Bits 3-0 of the VTPR field correspond to bits 7-4 of the TPR (which is the Task-Priority Class). */
1628	pVmcb->ctrl.IntCtrl.n.u8VTPR = (u8Tpr >> 4);
1629
1630	/* If there are interrupts pending, intercept CR8 writes to evaluate ASAP if we can deliver the interrupt to the guest. */
1631	if (fPendingIntr)
1632	pVmcb->ctrl.u16InterceptWrCRx \|= RT_BIT(8);
1633	else
1634	{
1635	pVmcb->ctrl.u16InterceptWrCRx &= ~RT_BIT(8);
1636	pVCpu->hm.s.svm.fSyncVTpr = true;
1637	}
1638
1639	pVmcb->ctrl.u64VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS \| HMSVM_VMCB_CLEAN_TPR);
1640	}
1641
1642	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
1643	return rc;
1644	}
1645
1646
1647	/**
1648	* Loads the exception interrupts required for guest execution in the VMCB.
1649	*
1650	* @returns VBox status code.
1651	* @param pVCpu Pointer to the VMCPU.
1652	* @param pVmcb Pointer to the VM control block.
1653	* @param pCtx Pointer to the guest-CPU context.
1654	*/
1655	static int hmR0SvmLoadGuestXcptIntercepts(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1656	{
1657	int rc = VINF_SUCCESS;
1658	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_XCPT_INTERCEPTS))
1659	{
1660	/* The remaining intercepts are handled elsewhere, e.g. in hmR0SvmLoadSharedCR0(). */
1661	if (pVCpu->hm.s.fGIMTrapXcptUD)
1662	hmR0SvmAddXcptIntercept(pVmcb, X86_XCPT_UD);
1663	else
1664	hmR0SvmRemoveXcptIntercept(pVmcb, X86_XCPT_UD);
1665	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_XCPT_INTERCEPTS);
1666	}
1667	return rc;
1668	}
1669
1670
1671	/**
1672	* Sets up the appropriate function to run guest code.
1673	*
1674	* @returns VBox status code.
1675	* @param pVCpu Pointer to the VMCPU.
1676	* @param pCtx Pointer to the guest-CPU context.
1677	*
1678	* @remarks No-long-jump zone!!!
1679	*/
1680	static int hmR0SvmSetupVMRunHandler(PVMCPU pVCpu, PCPUMCTX pCtx)
1681	{
1682	if (CPUMIsGuestInLongModeEx(pCtx))
1683	{
1684	#ifndef VBOX_ENABLE_64_BITS_GUESTS
1685	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1686	#endif
1687	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.fAllow64BitGuests); /* Guaranteed by hmR3InitFinalizeR0(). */
1688	#if HC_ARCH_BITS == 32 && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
1689	/* 32-bit host. We need to switch to 64-bit before running the 64-bit guest. */
1690	pVCpu->hm.s.svm.pfnVMRun = SVMR0VMSwitcherRun64;
1691	#else
1692	/* 64-bit host or hybrid host. */
1693	pVCpu->hm.s.svm.pfnVMRun = SVMR0VMRun64;
1694	#endif
1695	}
1696	else
1697	{
1698	/* Guest is not in long mode, use the 32-bit handler. */
1699	pVCpu->hm.s.svm.pfnVMRun = SVMR0VMRun;
1700	}
1701	return VINF_SUCCESS;
1702	}
1703
1704
1705	/**
1706	* Enters the AMD-V session.
1707	*
1708	* @returns VBox status code.
1709	* @param pVM Pointer to the VM.
1710	* @param pVCpu Pointer to the VMCPU.
1711	* @param pCpu Pointer to the CPU info struct.
1712	*/
1713	VMMR0DECL(int) SVMR0Enter(PVM pVM, PVMCPU pVCpu, PHMGLOBALCPUINFO pCpu)
1714	{
1715	AssertPtr(pVM);
1716	AssertPtr(pVCpu);
1717	Assert(pVM->hm.s.svm.fSupported);
1718	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1719	NOREF(pVM); NOREF(pCpu);
1720
1721	LogFlowFunc(("pVM=%p pVCpu=%p\n", pVM, pVCpu));
1722	Assert(HMCPU_CF_IS_SET(pVCpu, HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_HOST_GUEST_SHARED_STATE));
1723
1724	pVCpu->hm.s.fLeaveDone = false;
1725	return VINF_SUCCESS;
1726	}
1727
1728
1729	/**
1730	* Thread-context callback for AMD-V.
1731	*
1732	* @param enmEvent The thread-context event.
1733	* @param pVCpu Pointer to the VMCPU.
1734	* @param fGlobalInit Whether global VT-x/AMD-V init. is used.
1735	* @thread EMT(pVCpu)
1736	*/
1737	VMMR0DECL(void) SVMR0ThreadCtxCallback(RTTHREADCTXEVENT enmEvent, PVMCPU pVCpu, bool fGlobalInit)
1738	{
1739	NOREF(fGlobalInit);
1740
1741	switch (enmEvent)
1742	{
1743	case RTTHREADCTXEVENT_OUT:
1744	{
1745	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1746	Assert(VMMR0ThreadCtxHookIsEnabled(pVCpu));
1747	VMCPU_ASSERT_EMT(pVCpu);
1748
1749	PVM pVM = pVCpu->CTX_SUFF(pVM);
1750	PCPUMCTX pCtx = CPUMQueryGuestCtxPtr(pVCpu);
1751
1752	/* No longjmps (log-flush, locks) in this fragile context. */
1753	VMMRZCallRing3Disable(pVCpu);
1754
1755	if (!pVCpu->hm.s.fLeaveDone)
1756	{
1757	hmR0SvmLeave(pVM, pVCpu, pCtx);
1758	pVCpu->hm.s.fLeaveDone = true;
1759	}
1760
1761	/* Leave HM context, takes care of local init (term). */
1762	int rc = HMR0LeaveCpu(pVCpu);
1763	AssertRC(rc); NOREF(rc);
1764
1765	/* Restore longjmp state. */
1766	VMMRZCallRing3Enable(pVCpu);
1767	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchPreempt);
1768	break;
1769	}
1770
1771	case RTTHREADCTXEVENT_IN:
1772	{
1773	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1774	Assert(VMMR0ThreadCtxHookIsEnabled(pVCpu));
1775	VMCPU_ASSERT_EMT(pVCpu);
1776
1777	/* No longjmps (log-flush, locks) in this fragile context. */
1778	VMMRZCallRing3Disable(pVCpu);
1779
1780	/*
1781	* Initialize the bare minimum state required for HM. This takes care of
1782	* initializing AMD-V if necessary (onlined CPUs, local init etc.)
1783	*/
1784	int rc = HMR0EnterCpu(pVCpu);
1785	AssertRC(rc); NOREF(rc);
1786	Assert(HMCPU_CF_IS_SET(pVCpu, HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_HOST_GUEST_SHARED_STATE));
1787
1788	pVCpu->hm.s.fLeaveDone = false;
1789
1790	/* Restore longjmp state. */
1791	VMMRZCallRing3Enable(pVCpu);
1792	break;
1793	}
1794
1795	default:
1796	break;
1797	}
1798	}
1799
1800
1801	/**
1802	* Saves the host state.
1803	*
1804	* @returns VBox status code.
1805	* @param pVM Pointer to the VM.
1806	* @param pVCpu Pointer to the VMCPU.
1807	*
1808	* @remarks No-long-jump zone!!!
1809	*/
1810	VMMR0DECL(int) SVMR0SaveHostState(PVM pVM, PVMCPU pVCpu)
1811	{
1812	NOREF(pVM);
1813	NOREF(pVCpu);
1814	/* Nothing to do here. AMD-V does this for us automatically during the world-switch. */
1815	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_HOST_CONTEXT);
1816	return VINF_SUCCESS;
1817	}
1818
1819
1820	/**
1821	* Loads the guest state into the VMCB.
1822	*
1823	* The CPU state will be loaded from these fields on every successful VM-entry.
1824	* Also sets up the appropriate VMRUN function to execute guest code based on
1825	* the guest CPU mode.
1826	*
1827	* @returns VBox status code.
1828	* @param pVM Pointer to the VM.
1829	* @param pVCpu Pointer to the VMCPU.
1830	* @param pCtx Pointer to the guest-CPU context.
1831	*
1832	* @remarks No-long-jump zone!!!
1833	*/
1834	static int hmR0SvmLoadGuestState(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
1835	{
1836	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
1837	AssertMsgReturn(pVmcb, ("Invalid pVmcb\n"), VERR_SVM_INVALID_PVMCB);
1838
1839	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatLoadGuestState, x);
1840
1841	int rc = hmR0SvmLoadGuestControlRegs(pVCpu, pVmcb, pCtx);
1842	AssertLogRelMsgRCReturn(rc, ("hmR0SvmLoadGuestControlRegs! rc=%Rrc (pVM=%p pVCpu=%p)\n", rc, pVM, pVCpu), rc);
1843
1844	hmR0SvmLoadGuestSegmentRegs(pVCpu, pVmcb, pCtx);
1845	hmR0SvmLoadGuestMsrs(pVCpu, pVmcb, pCtx);
1846
1847	pVmcb->guest.u64RIP = pCtx->rip;
1848	pVmcb->guest.u64RSP = pCtx->rsp;
1849	pVmcb->guest.u64RFlags = pCtx->eflags.u32;
1850	pVmcb->guest.u64RAX = pCtx->rax;
1851
1852	rc = hmR0SvmLoadGuestApicState(pVCpu, pVmcb, pCtx);
1853	AssertLogRelMsgRCReturn(rc, ("hmR0SvmLoadGuestApicState! rc=%Rrc (pVM=%p pVCpu=%p)\n", rc, pVM, pVCpu), rc);
1854
1855	rc = hmR0SvmLoadGuestXcptIntercepts(pVCpu, pVmcb, pCtx);
1856	AssertLogRelMsgRCReturn(rc, ("hmR0SvmLoadGuestXcptIntercepts! rc=%Rrc (pVM=%p pVCpu=%p)\n", rc, pVM, pVCpu), rc);
1857
1858	rc = hmR0SvmSetupVMRunHandler(pVCpu, pCtx);
1859	AssertLogRelMsgRCReturn(rc, ("hmR0SvmSetupVMRunHandler! rc=%Rrc (pVM=%p pVCpu=%p)\n", rc, pVM, pVCpu), rc);
1860
1861	/* Clear any unused and reserved bits. */
1862	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_RIP /* Unused (loaded unconditionally). */
1863	\| HM_CHANGED_GUEST_RSP
1864	\| HM_CHANGED_GUEST_RFLAGS
1865	\| HM_CHANGED_GUEST_SYSENTER_CS_MSR
1866	\| HM_CHANGED_GUEST_SYSENTER_EIP_MSR
1867	\| HM_CHANGED_GUEST_SYSENTER_ESP_MSR
1868	\| HM_CHANGED_GUEST_LAZY_MSRS /* Unused. */
1869	\| HM_CHANGED_SVM_RESERVED1 /* Reserved. */
1870	\| HM_CHANGED_SVM_RESERVED2
1871	\| HM_CHANGED_SVM_RESERVED3
1872	\| HM_CHANGED_SVM_RESERVED4);
1873
1874	/* All the guest state bits should be loaded except maybe the host context and/or shared host/guest bits. */
1875	AssertMsg( !HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_ALL_GUEST)
1876	\|\| HMCPU_CF_IS_PENDING_ONLY(pVCpu, HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_HOST_GUEST_SHARED_STATE),
1877	("fContextUseFlags=%#RX32\n", HMCPU_CF_VALUE(pVCpu)));
1878
1879	Log4(("Load: CS:RIP=%04x:%RX64 EFL=%#x SS:RSP=%04x:%RX64\n", pCtx->cs.Sel, pCtx->rip, pCtx->eflags.u, pCtx->ss.Sel, pCtx->rsp));
1880	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatLoadGuestState, x);
1881	return rc;
1882	}
1883
1884
1885	/**
1886	* Loads the state shared between the host and guest into the
1887	* VMCB.
1888	*
1889	* @param pVCpu Pointer to the VMCPU.
1890	* @param pVmcb Pointer to the VM control block.
1891	* @param pCtx Pointer to the guest-CPU context.
1892	*
1893	* @remarks No-long-jump zone!!!
1894	*/
1895	static void hmR0SvmLoadSharedState(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1896	{
1897	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1898	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
1899
1900	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR0))
1901	hmR0SvmLoadSharedCR0(pVCpu, pVmcb, pCtx);
1902
1903	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_DEBUG))
1904	hmR0SvmLoadSharedDebugState(pVCpu, pVmcb, pCtx);
1905
1906	/* Unused on AMD-V. */
1907	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_LAZY_MSRS);
1908
1909	AssertMsg(!HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_HOST_GUEST_SHARED_STATE),
1910	("fContextUseFlags=%#RX32\n", HMCPU_CF_VALUE(pVCpu)));
1911	}
1912
1913
1914	/**
1915	* Saves the entire guest state from the VMCB into the
1916	* guest-CPU context. Currently there is no residual state left in the CPU that
1917	* is not updated in the VMCB.
1918	*
1919	* @returns VBox status code.
1920	* @param pVCpu Pointer to the VMCPU.
1921	* @param pMixedCtx Pointer to the guest-CPU context. The data may be
1922	* out-of-sync. Make sure to update the required fields
1923	* before using them.
1924	*/
1925	static void hmR0SvmSaveGuestState(PVMCPU pVCpu, PCPUMCTX pMixedCtx)
1926	{
1927	Assert(VMMRZCallRing3IsEnabled(pVCpu));
1928
1929	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
1930
1931	pMixedCtx->rip = pVmcb->guest.u64RIP;
1932	pMixedCtx->rsp = pVmcb->guest.u64RSP;
1933	pMixedCtx->eflags.u32 = pVmcb->guest.u64RFlags;
1934	pMixedCtx->rax = pVmcb->guest.u64RAX;
1935
1936	/*
1937	* Guest interrupt shadow.
1938	*/
1939	if (pVmcb->ctrl.u64IntShadow & SVM_INTERRUPT_SHADOW_ACTIVE)
1940	EMSetInhibitInterruptsPC(pVCpu, pMixedCtx->rip);
1941	else if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS))
1942	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS);
1943
1944	/*
1945	* Guest Control registers: CR2, CR3 (handled at the end) - accesses to other control registers are always intercepted.
1946	*/
1947	pMixedCtx->cr2 = pVmcb->guest.u64CR2;
1948
1949	/*
1950	* Guest MSRs.
1951	*/
1952	pMixedCtx->msrSTAR = pVmcb->guest.u64STAR; /* legacy syscall eip, cs & ss */
1953	pMixedCtx->msrLSTAR = pVmcb->guest.u64LSTAR; /* 64-bit mode syscall rip */
1954	pMixedCtx->msrCSTAR = pVmcb->guest.u64CSTAR; /* compatibility mode syscall rip */
1955	pMixedCtx->msrSFMASK = pVmcb->guest.u64SFMASK; /* syscall flag mask */
1956	pMixedCtx->msrKERNELGSBASE = pVmcb->guest.u64KernelGSBase; /* swapgs exchange value */
1957	pMixedCtx->SysEnter.cs = pVmcb->guest.u64SysEnterCS;
1958	pMixedCtx->SysEnter.eip = pVmcb->guest.u64SysEnterEIP;
1959	pMixedCtx->SysEnter.esp = pVmcb->guest.u64SysEnterESP;
1960
1961	/*
1962	* Guest segment registers (includes FS, GS base MSRs for 64-bit guests).
1963	*/
1964	HMSVM_SAVE_SEG_REG(CS, cs);
1965	HMSVM_SAVE_SEG_REG(SS, ss);
1966	HMSVM_SAVE_SEG_REG(DS, ds);
1967	HMSVM_SAVE_SEG_REG(ES, es);
1968	HMSVM_SAVE_SEG_REG(FS, fs);
1969	HMSVM_SAVE_SEG_REG(GS, gs);
1970
1971	/*
1972	* Correct the hidden CS granularity bit. Haven't seen it being wrong in any other
1973	* register (yet).
1974	*/
1975	/** @todo SELM might need to be fixed as it too should not care about the
1976	* granularity bit. See @bugref{6785}. */
1977	if ( !pMixedCtx->cs.Attr.n.u1Granularity
1978	&& pMixedCtx->cs.Attr.n.u1Present
1979	&& pMixedCtx->cs.u32Limit > UINT32_C(0xfffff))
1980	{
1981	Assert((pMixedCtx->cs.u32Limit & 0xfff) == 0xfff);
1982	pMixedCtx->cs.Attr.n.u1Granularity = 1;
1983	}
1984
1985	#ifdef VBOX_STRICT
1986	# define HMSVM_ASSERT_SEG_GRANULARITY(reg) \
1987	AssertMsg( !pMixedCtx->reg.Attr.n.u1Present \
1988	\|\| ( pMixedCtx->reg.Attr.n.u1Granularity \
1989	? (pMixedCtx->reg.u32Limit & 0xfff) == 0xfff \
1990	: pMixedCtx->reg.u32Limit <= UINT32_C(0xfffff)), \
1991	("Invalid Segment Attributes Limit=%#RX32 Attr=%#RX32 Base=%#RX64\n", pMixedCtx->reg.u32Limit, \
1992	pMixedCtx->reg.Attr.u, pMixedCtx->reg.u64Base))
1993
1994	HMSVM_ASSERT_SEG_GRANULARITY(cs);
1995	HMSVM_ASSERT_SEG_GRANULARITY(ss);
1996	HMSVM_ASSERT_SEG_GRANULARITY(ds);
1997	HMSVM_ASSERT_SEG_GRANULARITY(es);
1998	HMSVM_ASSERT_SEG_GRANULARITY(fs);
1999	HMSVM_ASSERT_SEG_GRANULARITY(gs);
2000
2001	# undef HMSVM_ASSERT_SEL_GRANULARITY
2002	#endif
2003
2004	/*
2005	* Sync the hidden SS DPL field. AMD CPUs have a separate CPL field in the VMCB and uses that
2006	* and thus it's possible that when the CPL changes during guest execution that the SS DPL
2007	* isn't updated by AMD-V. Observed on some AMD Fusion CPUs with 64-bit guests.
2008	* See AMD spec. 15.5.1 "Basic operation".
2009	*/
2010	Assert(!(pVmcb->guest.u8CPL & ~0x3));
2011	pMixedCtx->ss.Attr.n.u2Dpl = pVmcb->guest.u8CPL & 0x3;
2012
2013	/*
2014	* Guest TR.
2015	* Fixup TR attributes so it's compatible with Intel. Important when saved-states are used
2016	* between Intel and AMD. See @bugref{6208} comment #39.
2017	* ASSUME that it's normally correct and that we're in 32-bit or 64-bit mode.
2018	*/
2019	HMSVM_SAVE_SEG_REG(TR, tr);
2020	if (pMixedCtx->tr.Attr.n.u4Type != X86_SEL_TYPE_SYS_386_TSS_BUSY)
2021	{
2022	if ( pMixedCtx->tr.Attr.n.u4Type == X86_SEL_TYPE_SYS_386_TSS_AVAIL
2023	\|\| CPUMIsGuestInLongModeEx(pMixedCtx))
2024	pMixedCtx->tr.Attr.n.u4Type = X86_SEL_TYPE_SYS_386_TSS_BUSY;
2025	else if (pMixedCtx->tr.Attr.n.u4Type == X86_SEL_TYPE_SYS_286_TSS_AVAIL)
2026	pMixedCtx->tr.Attr.n.u4Type = X86_SEL_TYPE_SYS_286_TSS_BUSY;
2027	}
2028
2029	/*
2030	* Guest Descriptor-Table registers.
2031	*/
2032	HMSVM_SAVE_SEG_REG(LDTR, ldtr);
2033	pMixedCtx->gdtr.cbGdt = pVmcb->guest.GDTR.u32Limit;
2034	pMixedCtx->gdtr.pGdt = pVmcb->guest.GDTR.u64Base;
2035
2036	pMixedCtx->idtr.cbIdt = pVmcb->guest.IDTR.u32Limit;
2037	pMixedCtx->idtr.pIdt = pVmcb->guest.IDTR.u64Base;
2038
2039	/*
2040	* Guest Debug registers.
2041	*/
2042	if (!pVCpu->hm.s.fUsingHyperDR7)
2043	{
2044	pMixedCtx->dr[6] = pVmcb->guest.u64DR6;
2045	pMixedCtx->dr[7] = pVmcb->guest.u64DR7;
2046	}
2047	else
2048	{
2049	Assert(pVmcb->guest.u64DR7 == CPUMGetHyperDR7(pVCpu));
2050	CPUMSetHyperDR6(pVCpu, pVmcb->guest.u64DR6);
2051	}
2052
2053	/*
2054	* With Nested Paging, CR3 changes are not intercepted. Therefore, sync. it now.
2055	* This is done as the very last step of syncing the guest state, as PGMUpdateCR3() may cause longjmp's to ring-3.
2056	*/
2057	if ( pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging
2058	&& pMixedCtx->cr3 != pVmcb->guest.u64CR3)
2059	{
2060	CPUMSetGuestCR3(pVCpu, pVmcb->guest.u64CR3);
2061	PGMUpdateCR3(pVCpu, pVmcb->guest.u64CR3);
2062	}
2063	}
2064
2065
2066	/**
2067	* Does the necessary state syncing before returning to ring-3 for any reason
2068	* (longjmp, preemption, voluntary exits to ring-3) from AMD-V.
2069	*
2070	* @param pVM Pointer to the VM.
2071	* @param pVCpu Pointer to the VMCPU.
2072	* @param pMixedCtx Pointer to the guest-CPU context.
2073	*
2074	* @remarks No-long-jmp zone!!!
2075	*/
2076	static void hmR0SvmLeave(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
2077	{
2078	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2079	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2080	Assert(VMMR0IsLogFlushDisabled(pVCpu));
2081
2082	/*
2083	* !!! IMPORTANT !!!
2084	* If you modify code here, make sure to check whether hmR0SvmCallRing3Callback() needs to be updated too.
2085	*/
2086
2087	/* Restore host FPU state if necessary and resync on next R0 reentry .*/
2088	if (CPUMIsGuestFPUStateActive(pVCpu))
2089	{
2090	CPUMR0SaveGuestFPU(pVM, pVCpu, pCtx);
2091	Assert(!CPUMIsGuestFPUStateActive(pVCpu));
2092	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR0);
2093	}
2094
2095	/*
2096	* Restore host debug registers if necessary and resync on next R0 reentry.
2097	*/
2098	#ifdef VBOX_STRICT
2099	if (CPUMIsHyperDebugStateActive(pVCpu))
2100	{
2101	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2102	Assert(pVmcb->ctrl.u16InterceptRdDRx == 0xffff);
2103	Assert(pVmcb->ctrl.u16InterceptWrDRx == 0xffff);
2104	}
2105	#endif
2106	if (CPUMR0DebugStateMaybeSaveGuestAndRestoreHost(pVCpu, false /* save DR6 */))
2107	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_DEBUG);
2108
2109	Assert(!CPUMIsHyperDebugStateActive(pVCpu));
2110	Assert(!CPUMIsGuestDebugStateActive(pVCpu));
2111
2112	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatEntry);
2113	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatLoadGuestState);
2114	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatExit1);
2115	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatExit2);
2116	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchLongJmpToR3);
2117
2118	VMCPU_CMPXCHG_STATE(pVCpu, VMCPUSTATE_STARTED_HM, VMCPUSTATE_STARTED_EXEC);
2119	}
2120
2121
2122	/**
2123	* Leaves the AMD-V session.
2124	*
2125	* @returns VBox status code.
2126	* @param pVM Pointer to the VM.
2127	* @param pVCpu Pointer to the VMCPU.
2128	* @param pCtx Pointer to the guest-CPU context.
2129	*/
2130	static int hmR0SvmLeaveSession(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
2131	{
2132	HM_DISABLE_PREEMPT();
2133	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2134	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2135
2136	/* When thread-context hooks are used, we can avoid doing the leave again if we had been preempted before
2137	and done this from the SVMR0ThreadCtxCallback(). */
2138	if (!pVCpu->hm.s.fLeaveDone)
2139	{
2140	hmR0SvmLeave(pVM, pVCpu, pCtx);
2141	pVCpu->hm.s.fLeaveDone = true;
2142	}
2143
2144	/*
2145	* !!! IMPORTANT !!!
2146	* If you modify code here, make sure to check whether hmR0SvmCallRing3Callback() needs to be updated too.
2147	*/
2148
2149	/** @todo eliminate the need for calling VMMR0ThreadCtxHookDisable here! */
2150	/* Deregister hook now that we've left HM context before re-enabling preemption. */
2151	VMMR0ThreadCtxHookDisable(pVCpu);
2152
2153	/* Leave HM context. This takes care of local init (term). */
2154	int rc = HMR0LeaveCpu(pVCpu);
2155
2156	HM_RESTORE_PREEMPT();
2157	return rc;
2158	}
2159
2160
2161	/**
2162	* Does the necessary state syncing before doing a longjmp to ring-3.
2163	*
2164	* @returns VBox status code.
2165	* @param pVM Pointer to the VM.
2166	* @param pVCpu Pointer to the VMCPU.
2167	* @param pCtx Pointer to the guest-CPU context.
2168	*
2169	* @remarks No-long-jmp zone!!!
2170	*/
2171	static int hmR0SvmLongJmpToRing3(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
2172	{
2173	return hmR0SvmLeaveSession(pVM, pVCpu, pCtx);
2174	}
2175
2176
2177	/**
2178	* VMMRZCallRing3() callback wrapper which saves the guest state (or restores
2179	* any remaining host state) before we longjump to ring-3 and possibly get
2180	* preempted.
2181	*
2182	* @param pVCpu Pointer to the VMCPU.
2183	* @param enmOperation The operation causing the ring-3 longjump.
2184	* @param pvUser The user argument (pointer to the possibly
2185	* out-of-date guest-CPU context).
2186	*/
2187	DECLCALLBACK(int) hmR0SvmCallRing3Callback(PVMCPU pVCpu, VMMCALLRING3 enmOperation, void *pvUser)
2188	{
2189	if (enmOperation == VMMCALLRING3_VM_R0_ASSERTION)
2190	{
2191	/*
2192	* !!! IMPORTANT !!!
2193	* If you modify code here, make sure to check whether hmR0SvmLeave() and hmR0SvmLeaveSession() needs
2194	* to be updated too. This is a stripped down version which gets out ASAP trying to not trigger any assertion.
2195	*/
2196	VMMRZCallRing3RemoveNotification(pVCpu);
2197	VMMRZCallRing3Disable(pVCpu);
2198	HM_DISABLE_PREEMPT();
2199
2200	/* Restore host FPU state if necessary and resync on next R0 reentry .*/
2201	if (CPUMIsGuestFPUStateActive(pVCpu))
2202	CPUMR0SaveGuestFPU(pVCpu->CTX_SUFF(pVM), pVCpu, (PCPUMCTX)pvUser);
2203
2204	/* Restore host debug registers if necessary and resync on next R0 reentry. */
2205	CPUMR0DebugStateMaybeSaveGuestAndRestoreHost(pVCpu, false /* save DR6 */);
2206
2207	/* Deregister the hook now that we've left HM context before re-enabling preemption. */
2208	/** @todo eliminate the need for calling VMMR0ThreadCtxHookDisable here! */
2209	VMMR0ThreadCtxHookDisable(pVCpu);
2210
2211	/* Leave HM context. This takes care of local init (term). */
2212	HMR0LeaveCpu(pVCpu);
2213
2214	HM_RESTORE_PREEMPT();
2215	return VINF_SUCCESS;
2216	}
2217
2218	Assert(pVCpu);
2219	Assert(pvUser);
2220	Assert(VMMRZCallRing3IsEnabled(pVCpu));
2221	HMSVM_ASSERT_PREEMPT_SAFE();
2222
2223	VMMRZCallRing3Disable(pVCpu);
2224	Assert(VMMR0IsLogFlushDisabled(pVCpu));
2225
2226	Log4(("hmR0SvmCallRing3Callback->hmR0SvmLongJmpToRing3\n"));
2227	int rc = hmR0SvmLongJmpToRing3(pVCpu->CTX_SUFF(pVM), pVCpu, (PCPUMCTX)pvUser);
2228	AssertRCReturn(rc, rc);
2229
2230	VMMRZCallRing3Enable(pVCpu);
2231	return VINF_SUCCESS;
2232	}
2233
2234
2235	/**
2236	* Take necessary actions before going back to ring-3.
2237	*
2238	* An action requires us to go back to ring-3. This function does the necessary
2239	* steps before we can safely return to ring-3. This is not the same as longjmps
2240	* to ring-3, this is voluntary.
2241	*
2242	* @param pVM Pointer to the VM.
2243	* @param pVCpu Pointer to the VMCPU.
2244	* @param pCtx Pointer to the guest-CPU context.
2245	* @param rcExit The reason for exiting to ring-3. Can be
2246	* VINF_VMM_UNKNOWN_RING3_CALL.
2247	*/
2248	static void hmR0SvmExitToRing3(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, int rcExit)
2249	{
2250	Assert(pVM);
2251	Assert(pVCpu);
2252	Assert(pCtx);
2253	HMSVM_ASSERT_PREEMPT_SAFE();
2254
2255	/* Please, no longjumps here (any logging shouldn't flush jump back to ring-3). NO LOGGING BEFORE THIS POINT! */
2256	VMMRZCallRing3Disable(pVCpu);
2257	Log4(("hmR0SvmExitToRing3: rcExit=%d\n", rcExit));
2258
2259	/* We need to do this only while truly exiting the "inner loop" back to ring-3 and -not- for any longjmp to ring3. */
2260	if (pVCpu->hm.s.Event.fPending)
2261	{
2262	hmR0SvmPendingEventToTrpmTrap(pVCpu);
2263	Assert(!pVCpu->hm.s.Event.fPending);
2264	}
2265
2266	/* If we're emulating an instruction, we shouldn't have any TRPM traps pending
2267	and if we're injecting an event we should have a TRPM trap pending. */
2268	Assert(rcExit != VINF_EM_RAW_INJECT_TRPM_EVENT \|\| TRPMHasTrap(pVCpu));
2269	Assert(rcExit != VINF_EM_RAW_EMULATE_INSTR \|\| !TRPMHasTrap(pVCpu));
2270
2271	/* Sync. the necessary state for going back to ring-3. */
2272	hmR0SvmLeaveSession(pVM, pVCpu, pCtx);
2273	STAM_COUNTER_DEC(&pVCpu->hm.s.StatSwitchLongJmpToR3);
2274
2275	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_TO_R3);
2276	CPUMSetChangedFlags(pVCpu, CPUM_CHANGED_SYSENTER_MSR
2277	\| CPUM_CHANGED_LDTR
2278	\| CPUM_CHANGED_GDTR
2279	\| CPUM_CHANGED_IDTR
2280	\| CPUM_CHANGED_TR
2281	\| CPUM_CHANGED_HIDDEN_SEL_REGS);
2282	if ( pVM->hm.s.fNestedPaging
2283	&& CPUMIsGuestPagingEnabledEx(pCtx))
2284	{
2285	CPUMSetChangedFlags(pVCpu, CPUM_CHANGED_GLOBAL_TLB_FLUSH);
2286	}
2287
2288	/* On our way back from ring-3 reload the guest state if there is a possibility of it being changed. */
2289	if (rcExit != VINF_EM_RAW_INTERRUPT)
2290	HMCPU_CF_SET(pVCpu, HM_CHANGED_ALL_GUEST);
2291
2292	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchExitToR3);
2293
2294	/* We do -not- want any longjmp notifications after this! We must return to ring-3 ASAP. */
2295	VMMRZCallRing3RemoveNotification(pVCpu);
2296	VMMRZCallRing3Enable(pVCpu);
2297	}
2298
2299
2300	/**
2301	* Updates the use of TSC offsetting mode for the CPU and adjusts the necessary
2302	* intercepts.
2303	*
2304	* @param pVM The shared VM handle.
2305	* @param pVCpu Pointer to the VMCPU.
2306	*
2307	* @remarks No-long-jump zone!!!
2308	*/
2309	static void hmR0SvmUpdateTscOffsetting(PVM pVM, PVMCPU pVCpu)
2310	{
2311	bool fParavirtTsc;
2312	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2313	bool fCanUseRealTsc = TMCpuTickCanUseRealTSC(pVM, pVCpu, &pVmcb->ctrl.u64TSCOffset, &fParavirtTsc);
2314	if (fCanUseRealTsc)
2315	{
2316	pVmcb->ctrl.u32InterceptCtrl1 &= ~SVM_CTRL1_INTERCEPT_RDTSC;
2317	pVmcb->ctrl.u32InterceptCtrl2 &= ~SVM_CTRL2_INTERCEPT_RDTSCP;
2318	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscOffset);
2319	}
2320	else
2321	{
2322	pVmcb->ctrl.u32InterceptCtrl1 \|= SVM_CTRL1_INTERCEPT_RDTSC;
2323	pVmcb->ctrl.u32InterceptCtrl2 \|= SVM_CTRL2_INTERCEPT_RDTSCP;
2324	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscIntercept);
2325	}
2326	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2327
2328	/** @todo later optimize this to be done elsewhere and not before every
2329	* VM-entry. */
2330	if (fParavirtTsc)
2331	{
2332	/* Currently neither Hyper-V nor KVM need to update their paravirt. TSC
2333	information before every VM-entry, hence disable it for performance sake. */
2334	#if 0
2335	int rc = GIMR0UpdateParavirtTsc(pVM, 0 /* u64Offset */);
2336	AssertRC(rc);
2337	#endif
2338	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscParavirt);
2339	}
2340	}
2341
2342
2343	/**
2344	* Sets an event as a pending event to be injected into the guest.
2345	*
2346	* @param pVCpu Pointer to the VMCPU.
2347	* @param pEvent Pointer to the SVM event.
2348	* @param GCPtrFaultAddress The fault-address (CR2) in case it's a
2349	* page-fault.
2350	*
2351	* @remarks Statistics counter assumes this is a guest event being reflected to
2352	* the guest i.e. 'StatInjectPendingReflect' is incremented always.
2353	*/
2354	DECLINLINE(void) hmR0SvmSetPendingEvent(PVMCPU pVCpu, PSVMEVENT pEvent, RTGCUINTPTR GCPtrFaultAddress)
2355	{
2356	Assert(!pVCpu->hm.s.Event.fPending);
2357	Assert(pEvent->n.u1Valid);
2358
2359	pVCpu->hm.s.Event.u64IntInfo = pEvent->u;
2360	pVCpu->hm.s.Event.fPending = true;
2361	pVCpu->hm.s.Event.GCPtrFaultAddress = GCPtrFaultAddress;
2362
2363	Log4(("hmR0SvmSetPendingEvent: u=%#RX64 u8Vector=%#x Type=%#x ErrorCodeValid=%RTbool ErrorCode=%#RX32\n", pEvent->u,
2364	pEvent->n.u8Vector, (uint8_t)pEvent->n.u3Type, !!pEvent->n.u1ErrorCodeValid, pEvent->n.u32ErrorCode));
2365
2366	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectPendingReflect);
2367	}
2368
2369
2370	/**
2371	* Injects an event into the guest upon VMRUN by updating the relevant field
2372	* in the VMCB.
2373	*
2374	* @param pVCpu Pointer to the VMCPU.
2375	* @param pVmcb Pointer to the guest VM control block.
2376	* @param pCtx Pointer to the guest-CPU context.
2377	* @param pEvent Pointer to the event.
2378	*
2379	* @remarks No-long-jump zone!!!
2380	* @remarks Requires CR0!
2381	*/
2382	DECLINLINE(void) hmR0SvmInjectEventVmcb(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx, PSVMEVENT pEvent)
2383	{
2384	NOREF(pVCpu); NOREF(pCtx);
2385
2386	pVmcb->ctrl.EventInject.u = pEvent->u;
2387	STAM_COUNTER_INC(&pVCpu->hm.s.paStatInjectedIrqsR0[pEvent->n.u8Vector & MASK_INJECT_IRQ_STAT]);
2388
2389	Log4(("hmR0SvmInjectEventVmcb: u=%#RX64 u8Vector=%#x Type=%#x ErrorCodeValid=%RTbool ErrorCode=%#RX32\n", pEvent->u,
2390	pEvent->n.u8Vector, (uint8_t)pEvent->n.u3Type, !!pEvent->n.u1ErrorCodeValid, pEvent->n.u32ErrorCode));
2391	}
2392
2393
2394
2395	/**
2396	* Converts any TRPM trap into a pending HM event. This is typically used when
2397	* entering from ring-3 (not longjmp returns).
2398	*
2399	* @param pVCpu Pointer to the VMCPU.
2400	*/
2401	static void hmR0SvmTrpmTrapToPendingEvent(PVMCPU pVCpu)
2402	{
2403	Assert(TRPMHasTrap(pVCpu));
2404	Assert(!pVCpu->hm.s.Event.fPending);
2405
2406	uint8_t uVector;
2407	TRPMEVENT enmTrpmEvent;
2408	RTGCUINT uErrCode;
2409	RTGCUINTPTR GCPtrFaultAddress;
2410	uint8_t cbInstr;
2411
2412	int rc = TRPMQueryTrapAll(pVCpu, &uVector, &enmTrpmEvent, &uErrCode, &GCPtrFaultAddress, &cbInstr);
2413	AssertRC(rc);
2414
2415	SVMEVENT Event;
2416	Event.u = 0;
2417	Event.n.u1Valid = 1;
2418	Event.n.u8Vector = uVector;
2419
2420	/* Refer AMD spec. 15.20 "Event Injection" for the format. */
2421	if (enmTrpmEvent == TRPM_TRAP)
2422	{
2423	Event.n.u3Type = SVM_EVENT_EXCEPTION;
2424	switch (uVector)
2425	{
2426	case X86_XCPT_NMI:
2427	{
2428	Event.n.u3Type = SVM_EVENT_NMI;
2429	break;
2430	}
2431
2432	case X86_XCPT_PF:
2433	case X86_XCPT_DF:
2434	case X86_XCPT_TS:
2435	case X86_XCPT_NP:
2436	case X86_XCPT_SS:
2437	case X86_XCPT_GP:
2438	case X86_XCPT_AC:
2439	{
2440	Event.n.u1ErrorCodeValid = 1;
2441	Event.n.u32ErrorCode = uErrCode;
2442	break;
2443	}
2444	}
2445	}
2446	else if (enmTrpmEvent == TRPM_HARDWARE_INT)
2447	Event.n.u3Type = SVM_EVENT_EXTERNAL_IRQ;
2448	else if (enmTrpmEvent == TRPM_SOFTWARE_INT)
2449	Event.n.u3Type = SVM_EVENT_SOFTWARE_INT;
2450	else
2451	AssertMsgFailed(("Invalid TRPM event type %d\n", enmTrpmEvent));
2452
2453	rc = TRPMResetTrap(pVCpu);
2454	AssertRC(rc);
2455
2456	Log4(("TRPM->HM event: u=%#RX64 u8Vector=%#x uErrorCodeValid=%RTbool uErrorCode=%#RX32\n", Event.u, Event.n.u8Vector,
2457	!!Event.n.u1ErrorCodeValid, Event.n.u32ErrorCode));
2458
2459	hmR0SvmSetPendingEvent(pVCpu, &Event, GCPtrFaultAddress);
2460	STAM_COUNTER_DEC(&pVCpu->hm.s.StatInjectPendingReflect);
2461	}
2462
2463
2464	/**
2465	* Converts any pending SVM event into a TRPM trap. Typically used when leaving
2466	* AMD-V to execute any instruction.
2467	*
2468	* @param pvCpu Pointer to the VMCPU.
2469	*/
2470	static void hmR0SvmPendingEventToTrpmTrap(PVMCPU pVCpu)
2471	{
2472	Assert(pVCpu->hm.s.Event.fPending);
2473	Assert(TRPMQueryTrap(pVCpu, NULL /* pu8TrapNo /, NULL / pEnmType */) == VERR_TRPM_NO_ACTIVE_TRAP);
2474
2475	SVMEVENT Event;
2476	Event.u = pVCpu->hm.s.Event.u64IntInfo;
2477
2478	uint8_t uVector = Event.n.u8Vector;
2479	uint8_t uVectorType = Event.n.u3Type;
2480
2481	TRPMEVENT enmTrapType;
2482	switch (uVectorType)
2483	{
2484	case SVM_EVENT_EXTERNAL_IRQ:
2485	enmTrapType = TRPM_HARDWARE_INT;
2486	break;
2487	case SVM_EVENT_SOFTWARE_INT:
2488	enmTrapType = TRPM_SOFTWARE_INT;
2489	break;
2490	case SVM_EVENT_EXCEPTION:
2491	case SVM_EVENT_NMI:
2492	enmTrapType = TRPM_TRAP;
2493	break;
2494	default:
2495	AssertMsgFailed(("Invalid pending-event type %#x\n", uVectorType));
2496	enmTrapType = TRPM_32BIT_HACK;
2497	break;
2498	}
2499
2500	Log4(("HM event->TRPM: uVector=%#x enmTrapType=%d\n", uVector, uVectorType));
2501
2502	int rc = TRPMAssertTrap(pVCpu, uVector, enmTrapType);
2503	AssertRC(rc);
2504
2505	if (Event.n.u1ErrorCodeValid)
2506	TRPMSetErrorCode(pVCpu, Event.n.u32ErrorCode);
2507
2508	if ( uVectorType == SVM_EVENT_EXCEPTION
2509	&& uVector == X86_XCPT_PF)
2510	{
2511	TRPMSetFaultAddress(pVCpu, pVCpu->hm.s.Event.GCPtrFaultAddress);
2512	Assert(pVCpu->hm.s.Event.GCPtrFaultAddress == CPUMGetGuestCR2(pVCpu));
2513	}
2514	else if (uVectorType == SVM_EVENT_SOFTWARE_INT)
2515	{
2516	AssertMsg( uVectorType == SVM_EVENT_SOFTWARE_INT
2517	\|\| (uVector == X86_XCPT_BP \|\| uVector == X86_XCPT_OF),
2518	("Invalid vector: uVector=%#x uVectorType=%#x\n", uVector, uVectorType));
2519	TRPMSetInstrLength(pVCpu, pVCpu->hm.s.Event.cbInstr);
2520	}
2521	pVCpu->hm.s.Event.fPending = false;
2522	}
2523
2524
2525	/**
2526	* Gets the guest's interrupt-shadow.
2527	*
2528	* @returns The guest's interrupt-shadow.
2529	* @param pVCpu Pointer to the VMCPU.
2530	* @param pCtx Pointer to the guest-CPU context.
2531	*
2532	* @remarks No-long-jump zone!!!
2533	* @remarks Has side-effects with VMCPU_FF_INHIBIT_INTERRUPTS force-flag.
2534	*/
2535	DECLINLINE(uint32_t) hmR0SvmGetGuestIntrShadow(PVMCPU pVCpu, PCPUMCTX pCtx)
2536	{
2537	/*
2538	* Instructions like STI and MOV SS inhibit interrupts till the next instruction completes. Check if we should
2539	* inhibit interrupts or clear any existing interrupt-inhibition.
2540	*/
2541	uint32_t uIntrState = 0;
2542	if (VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS))
2543	{
2544	if (pCtx->rip != EMGetInhibitInterruptsPC(pVCpu))
2545	{
2546	/*
2547	* We can clear the inhibit force flag as even if we go back to the recompiler without executing guest code in
2548	* AMD-V, the flag's condition to be cleared is met and thus the cleared state is correct.
2549	*/
2550	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS);
2551	}
2552	else
2553	uIntrState = SVM_INTERRUPT_SHADOW_ACTIVE;
2554	}
2555	return uIntrState;
2556	}
2557
2558
2559	/**
2560	* Sets the virtual interrupt intercept control in the VMCB which
2561	* instructs AMD-V to cause a #VMEXIT as soon as the guest is in a state to
2562	* receive interrupts.
2563	*
2564	* @param pVmcb Pointer to the VM control block.
2565	*/
2566	DECLINLINE(void) hmR0SvmSetVirtIntrIntercept(PSVMVMCB pVmcb)
2567	{
2568	if (!(pVmcb->ctrl.u32InterceptCtrl1 & SVM_CTRL1_INTERCEPT_VINTR))
2569	{
2570	pVmcb->ctrl.IntCtrl.n.u1VIrqValid = 1; /* A virtual interrupt is pending. */
2571	pVmcb->ctrl.IntCtrl.n.u8VIrqVector = 0; /* Not necessary as we #VMEXIT for delivering the interrupt. */
2572	pVmcb->ctrl.u32InterceptCtrl1 \|= SVM_CTRL1_INTERCEPT_VINTR;
2573	pVmcb->ctrl.u64VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS \| HMSVM_VMCB_CLEAN_TPR);
2574
2575	Log4(("Setting VINTR intercept\n"));
2576	}
2577	}
2578
2579
2580	/**
2581	* Sets the IRET intercept control in the VMCB which instructs AMD-V to cause a
2582	* #VMEXIT as soon as a guest starts executing an IRET. This is used to unblock
2583	* virtual NMIs.
2584	*
2585	* @param pVmcb Pointer to the VM control block.
2586	*/
2587	DECLINLINE(void) hmR0SvmSetIretIntercept(PSVMVMCB pVmcb)
2588	{
2589	if (!(pVmcb->ctrl.u32InterceptCtrl1 & SVM_CTRL1_INTERCEPT_IRET))
2590	{
2591	pVmcb->ctrl.u32InterceptCtrl1 \|= SVM_CTRL1_INTERCEPT_IRET;
2592	pVmcb->ctrl.u64VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS);
2593
2594	Log4(("Setting IRET intercept\n"));
2595	}
2596	}
2597
2598
2599	/**
2600	* Clears the IRET intercept control in the VMCB.
2601	*
2602	* @param pVmcb Pointer to the VM control block.
2603	*/
2604	DECLINLINE(void) hmR0SvmClearIretIntercept(PSVMVMCB pVmcb)
2605	{
2606	if (pVmcb->ctrl.u32InterceptCtrl1 & SVM_CTRL1_INTERCEPT_IRET)
2607	{
2608	pVmcb->ctrl.u32InterceptCtrl1 &= ~SVM_CTRL1_INTERCEPT_IRET;
2609	pVmcb->ctrl.u64VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS);
2610
2611	Log4(("Clearing IRET intercept\n"));
2612	}
2613	}
2614
2615
2616	/**
2617	* Evaluates the event to be delivered to the guest and sets it as the pending
2618	* event.
2619	*
2620	* @param pVCpu Pointer to the VMCPU.
2621	* @param pCtx Pointer to the guest-CPU context.
2622	*/
2623	static void hmR0SvmEvaluatePendingEvent(PVMCPU pVCpu, PCPUMCTX pCtx)
2624	{
2625	Assert(!pVCpu->hm.s.Event.fPending);
2626	Log4Func(("\n"));
2627
2628	bool const fIntShadow = RT_BOOL(hmR0SvmGetGuestIntrShadow(pVCpu, pCtx));
2629	bool const fBlockInt = !(pCtx->eflags.u32 & X86_EFL_IF);
2630	bool const fBlockNmi = RT_BOOL(VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_BLOCK_NMIS));
2631	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2632
2633	SVMEVENT Event;
2634	Event.u = 0;
2635	/** @todo SMI. SMIs take priority over NMIs. */
2636	if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_NMI)) /* NMI. NMIs take priority over regular interrupts . */
2637	{
2638	if (fBlockNmi)
2639	hmR0SvmSetIretIntercept(pVmcb);
2640	else if (fIntShadow)
2641	hmR0SvmSetVirtIntrIntercept(pVmcb);
2642	else
2643	{
2644	Log4(("Pending NMI\n"));
2645
2646	Event.n.u1Valid = 1;
2647	Event.n.u8Vector = X86_XCPT_NMI;
2648	Event.n.u3Type = SVM_EVENT_NMI;
2649
2650	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
2651	hmR0SvmSetIretIntercept(pVmcb);
2652	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INTERRUPT_NMI);
2653	}
2654	}
2655	else if (VMCPU_FF_IS_PENDING(pVCpu, (VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC)))
2656	{
2657	/*
2658	* Check if the guest can receive external interrupts (PIC/APIC). Once we do PDMGetInterrupt() we -must- deliver
2659	* the interrupt ASAP. We must not execute any guest code until we inject the interrupt which is why it is
2660	* evaluated here and not set as pending, solely based on the force-flags.
2661	*/
2662	if ( !fBlockInt
2663	&& !fIntShadow)
2664	{
2665	uint8_t u8Interrupt;
2666	int rc = PDMGetInterrupt(pVCpu, &u8Interrupt);
2667	if (RT_SUCCESS(rc))
2668	{
2669	Log4(("Injecting external interrupt u8Interrupt=%#x\n", u8Interrupt));
2670
2671	Event.n.u1Valid = 1;
2672	Event.n.u8Vector = u8Interrupt;
2673	Event.n.u3Type = SVM_EVENT_EXTERNAL_IRQ;
2674
2675	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
2676	}
2677	else
2678	{
2679	/** @todo Does this actually happen? If not turn it into an assertion. */
2680	Assert(!VMCPU_FF_IS_PENDING(pVCpu, (VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC)));
2681	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchGuestIrq);
2682	}
2683	}
2684	else
2685	hmR0SvmSetVirtIntrIntercept(pVmcb);
2686	}
2687	}
2688
2689
2690	/**
2691	* Injects any pending events into the guest if the guest is in a state to
2692	* receive them.
2693	*
2694	* @param pVCpu Pointer to the VMCPU.
2695	* @param pCtx Pointer to the guest-CPU context.
2696	*/
2697	static void hmR0SvmInjectPendingEvent(PVMCPU pVCpu, PCPUMCTX pCtx)
2698	{
2699	Assert(!TRPMHasTrap(pVCpu));
2700	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2701
2702	bool const fIntShadow = RT_BOOL(hmR0SvmGetGuestIntrShadow(pVCpu, pCtx));
2703	bool const fBlockInt = !(pCtx->eflags.u32 & X86_EFL_IF);
2704	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2705
2706	if (pVCpu->hm.s.Event.fPending) /* First, inject any pending HM events. */
2707	{
2708	SVMEVENT Event;
2709	Event.u = pVCpu->hm.s.Event.u64IntInfo;
2710	Assert(Event.n.u1Valid);
2711	#ifdef VBOX_STRICT
2712	if (Event.n.u3Type == SVM_EVENT_EXTERNAL_IRQ)
2713	{
2714	Assert(!fBlockInt);
2715	Assert(!fIntShadow);
2716	}
2717	else if (Event.n.u3Type == SVM_EVENT_NMI)
2718	Assert(!fIntShadow);
2719	#endif
2720
2721	Log4(("Injecting pending HM event.\n"));
2722	hmR0SvmInjectEventVmcb(pVCpu, pVmcb, pCtx, &Event);
2723	pVCpu->hm.s.Event.fPending = false;
2724
2725	#ifdef VBOX_WITH_STATISTICS
2726	if (Event.n.u3Type == SVM_EVENT_EXTERNAL_IRQ)
2727	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectInterrupt);
2728	else
2729	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectXcpt);
2730	#endif
2731	}
2732
2733	/* Update the guest interrupt shadow in the VMCB. */
2734	pVmcb->ctrl.u64IntShadow = !!fIntShadow;
2735	NOREF(fBlockInt);
2736	}
2737
2738
2739	/**
2740	* Reports world-switch error and dumps some useful debug info.
2741	*
2742	* @param pVM Pointer to the VM.
2743	* @param pVCpu Pointer to the VMCPU.
2744	* @param rcVMRun The return code from VMRUN (or
2745	* VERR_SVM_INVALID_GUEST_STATE for invalid
2746	* guest-state).
2747	* @param pCtx Pointer to the guest-CPU context.
2748	*/
2749	static void hmR0SvmReportWorldSwitchError(PVM pVM, PVMCPU pVCpu, int rcVMRun, PCPUMCTX pCtx)
2750	{
2751	NOREF(pCtx);
2752	HMSVM_ASSERT_PREEMPT_SAFE();
2753	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2754
2755	if (rcVMRun == VERR_SVM_INVALID_GUEST_STATE)
2756	{
2757	HMDumpRegs(pVM, pVCpu, pCtx); NOREF(pVM);
2758	#ifdef VBOX_STRICT
2759	Log4(("ctrl.u64VmcbCleanBits %#RX64\n", pVmcb->ctrl.u64VmcbCleanBits));
2760	Log4(("ctrl.u16InterceptRdCRx %#x\n", pVmcb->ctrl.u16InterceptRdCRx));
2761	Log4(("ctrl.u16InterceptWrCRx %#x\n", pVmcb->ctrl.u16InterceptWrCRx));
2762	Log4(("ctrl.u16InterceptRdDRx %#x\n", pVmcb->ctrl.u16InterceptRdDRx));
2763	Log4(("ctrl.u16InterceptWrDRx %#x\n", pVmcb->ctrl.u16InterceptWrDRx));
2764	Log4(("ctrl.u32InterceptException %#x\n", pVmcb->ctrl.u32InterceptException));
2765	Log4(("ctrl.u32InterceptCtrl1 %#x\n", pVmcb->ctrl.u32InterceptCtrl1));
2766	Log4(("ctrl.u32InterceptCtrl2 %#x\n", pVmcb->ctrl.u32InterceptCtrl2));
2767	Log4(("ctrl.u64IOPMPhysAddr %#RX64\n", pVmcb->ctrl.u64IOPMPhysAddr));
2768	Log4(("ctrl.u64MSRPMPhysAddr %#RX64\n", pVmcb->ctrl.u64MSRPMPhysAddr));
2769	Log4(("ctrl.u64TSCOffset %#RX64\n", pVmcb->ctrl.u64TSCOffset));
2770
2771	Log4(("ctrl.TLBCtrl.u32ASID %#x\n", pVmcb->ctrl.TLBCtrl.n.u32ASID));
2772	Log4(("ctrl.TLBCtrl.u8TLBFlush %#x\n", pVmcb->ctrl.TLBCtrl.n.u8TLBFlush));
2773	Log4(("ctrl.TLBCtrl.u24Reserved %#x\n", pVmcb->ctrl.TLBCtrl.n.u24Reserved));
2774
2775	Log4(("ctrl.IntCtrl.u8VTPR %#x\n", pVmcb->ctrl.IntCtrl.n.u8VTPR));
2776	Log4(("ctrl.IntCtrl.u1VIrqValid %#x\n", pVmcb->ctrl.IntCtrl.n.u1VIrqValid));
2777	Log4(("ctrl.IntCtrl.u7Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u7Reserved));
2778	Log4(("ctrl.IntCtrl.u4VIrqPriority %#x\n", pVmcb->ctrl.IntCtrl.n.u4VIrqPriority));
2779	Log4(("ctrl.IntCtrl.u1IgnoreTPR %#x\n", pVmcb->ctrl.IntCtrl.n.u1IgnoreTPR));
2780	Log4(("ctrl.IntCtrl.u3Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u3Reserved));
2781	Log4(("ctrl.IntCtrl.u1VIrqMasking %#x\n", pVmcb->ctrl.IntCtrl.n.u1VIrqMasking));
2782	Log4(("ctrl.IntCtrl.u6Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u6Reserved));
2783	Log4(("ctrl.IntCtrl.u8VIrqVector %#x\n", pVmcb->ctrl.IntCtrl.n.u8VIrqVector));
2784	Log4(("ctrl.IntCtrl.u24Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u24Reserved));
2785
2786	Log4(("ctrl.u64IntShadow %#RX64\n", pVmcb->ctrl.u64IntShadow));
2787	Log4(("ctrl.u64ExitCode %#RX64\n", pVmcb->ctrl.u64ExitCode));
2788	Log4(("ctrl.u64ExitInfo1 %#RX64\n", pVmcb->ctrl.u64ExitInfo1));
2789	Log4(("ctrl.u64ExitInfo2 %#RX64\n", pVmcb->ctrl.u64ExitInfo2));
2790	Log4(("ctrl.ExitIntInfo.u8Vector %#x\n", pVmcb->ctrl.ExitIntInfo.n.u8Vector));
2791	Log4(("ctrl.ExitIntInfo.u3Type %#x\n", pVmcb->ctrl.ExitIntInfo.n.u3Type));
2792	Log4(("ctrl.ExitIntInfo.u1ErrorCodeValid %#x\n", pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid));
2793	Log4(("ctrl.ExitIntInfo.u19Reserved %#x\n", pVmcb->ctrl.ExitIntInfo.n.u19Reserved));
2794	Log4(("ctrl.ExitIntInfo.u1Valid %#x\n", pVmcb->ctrl.ExitIntInfo.n.u1Valid));
2795	Log4(("ctrl.ExitIntInfo.u32ErrorCode %#x\n", pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode));
2796	Log4(("ctrl.NestedPaging %#RX64\n", pVmcb->ctrl.NestedPaging.u));
2797	Log4(("ctrl.EventInject.u8Vector %#x\n", pVmcb->ctrl.EventInject.n.u8Vector));
2798	Log4(("ctrl.EventInject.u3Type %#x\n", pVmcb->ctrl.EventInject.n.u3Type));
2799	Log4(("ctrl.EventInject.u1ErrorCodeValid %#x\n", pVmcb->ctrl.EventInject.n.u1ErrorCodeValid));
2800	Log4(("ctrl.EventInject.u19Reserved %#x\n", pVmcb->ctrl.EventInject.n.u19Reserved));
2801	Log4(("ctrl.EventInject.u1Valid %#x\n", pVmcb->ctrl.EventInject.n.u1Valid));
2802	Log4(("ctrl.EventInject.u32ErrorCode %#x\n", pVmcb->ctrl.EventInject.n.u32ErrorCode));
2803
2804	Log4(("ctrl.u64NestedPagingCR3 %#RX64\n", pVmcb->ctrl.u64NestedPagingCR3));
2805	Log4(("ctrl.u64LBRVirt %#RX64\n", pVmcb->ctrl.u64LBRVirt));
2806
2807	Log4(("guest.CS.u16Sel %RTsel\n", pVmcb->guest.CS.u16Sel));
2808	Log4(("guest.CS.u16Attr %#x\n", pVmcb->guest.CS.u16Attr));
2809	Log4(("guest.CS.u32Limit %#RX32\n", pVmcb->guest.CS.u32Limit));
2810	Log4(("guest.CS.u64Base %#RX64\n", pVmcb->guest.CS.u64Base));
2811	Log4(("guest.DS.u16Sel %#RTsel\n", pVmcb->guest.DS.u16Sel));
2812	Log4(("guest.DS.u16Attr %#x\n", pVmcb->guest.DS.u16Attr));
2813	Log4(("guest.DS.u32Limit %#RX32\n", pVmcb->guest.DS.u32Limit));
2814	Log4(("guest.DS.u64Base %#RX64\n", pVmcb->guest.DS.u64Base));
2815	Log4(("guest.ES.u16Sel %RTsel\n", pVmcb->guest.ES.u16Sel));
2816	Log4(("guest.ES.u16Attr %#x\n", pVmcb->guest.ES.u16Attr));
2817	Log4(("guest.ES.u32Limit %#RX32\n", pVmcb->guest.ES.u32Limit));
2818	Log4(("guest.ES.u64Base %#RX64\n", pVmcb->guest.ES.u64Base));
2819	Log4(("guest.FS.u16Sel %RTsel\n", pVmcb->guest.FS.u16Sel));
2820	Log4(("guest.FS.u16Attr %#x\n", pVmcb->guest.FS.u16Attr));
2821	Log4(("guest.FS.u32Limit %#RX32\n", pVmcb->guest.FS.u32Limit));
2822	Log4(("guest.FS.u64Base %#RX64\n", pVmcb->guest.FS.u64Base));
2823	Log4(("guest.GS.u16Sel %RTsel\n", pVmcb->guest.GS.u16Sel));
2824	Log4(("guest.GS.u16Attr %#x\n", pVmcb->guest.GS.u16Attr));
2825	Log4(("guest.GS.u32Limit %#RX32\n", pVmcb->guest.GS.u32Limit));
2826	Log4(("guest.GS.u64Base %#RX64\n", pVmcb->guest.GS.u64Base));
2827
2828	Log4(("guest.GDTR.u32Limit %#RX32\n", pVmcb->guest.GDTR.u32Limit));
2829	Log4(("guest.GDTR.u64Base %#RX64\n", pVmcb->guest.GDTR.u64Base));
2830
2831	Log4(("guest.LDTR.u16Sel %RTsel\n", pVmcb->guest.LDTR.u16Sel));
2832	Log4(("guest.LDTR.u16Attr %#x\n", pVmcb->guest.LDTR.u16Attr));
2833	Log4(("guest.LDTR.u32Limit %#RX32\n", pVmcb->guest.LDTR.u32Limit));
2834	Log4(("guest.LDTR.u64Base %#RX64\n", pVmcb->guest.LDTR.u64Base));
2835
2836	Log4(("guest.IDTR.u32Limit %#RX32\n", pVmcb->guest.IDTR.u32Limit));
2837	Log4(("guest.IDTR.u64Base %#RX64\n", pVmcb->guest.IDTR.u64Base));
2838
2839	Log4(("guest.TR.u16Sel %RTsel\n", pVmcb->guest.TR.u16Sel));
2840	Log4(("guest.TR.u16Attr %#x\n", pVmcb->guest.TR.u16Attr));
2841	Log4(("guest.TR.u32Limit %#RX32\n", pVmcb->guest.TR.u32Limit));
2842	Log4(("guest.TR.u64Base %#RX64\n", pVmcb->guest.TR.u64Base));
2843
2844	Log4(("guest.u8CPL %#x\n", pVmcb->guest.u8CPL));
2845	Log4(("guest.u64CR0 %#RX64\n", pVmcb->guest.u64CR0));
2846	Log4(("guest.u64CR2 %#RX64\n", pVmcb->guest.u64CR2));
2847	Log4(("guest.u64CR3 %#RX64\n", pVmcb->guest.u64CR3));
2848	Log4(("guest.u64CR4 %#RX64\n", pVmcb->guest.u64CR4));
2849	Log4(("guest.u64DR6 %#RX64\n", pVmcb->guest.u64DR6));
2850	Log4(("guest.u64DR7 %#RX64\n", pVmcb->guest.u64DR7));
2851
2852	Log4(("guest.u64RIP %#RX64\n", pVmcb->guest.u64RIP));
2853	Log4(("guest.u64RSP %#RX64\n", pVmcb->guest.u64RSP));
2854	Log4(("guest.u64RAX %#RX64\n", pVmcb->guest.u64RAX));
2855	Log4(("guest.u64RFlags %#RX64\n", pVmcb->guest.u64RFlags));
2856
2857	Log4(("guest.u64SysEnterCS %#RX64\n", pVmcb->guest.u64SysEnterCS));
2858	Log4(("guest.u64SysEnterEIP %#RX64\n", pVmcb->guest.u64SysEnterEIP));
2859	Log4(("guest.u64SysEnterESP %#RX64\n", pVmcb->guest.u64SysEnterESP));
2860
2861	Log4(("guest.u64EFER %#RX64\n", pVmcb->guest.u64EFER));
2862	Log4(("guest.u64STAR %#RX64\n", pVmcb->guest.u64STAR));
2863	Log4(("guest.u64LSTAR %#RX64\n", pVmcb->guest.u64LSTAR));
2864	Log4(("guest.u64CSTAR %#RX64\n", pVmcb->guest.u64CSTAR));
2865	Log4(("guest.u64SFMASK %#RX64\n", pVmcb->guest.u64SFMASK));
2866	Log4(("guest.u64KernelGSBase %#RX64\n", pVmcb->guest.u64KernelGSBase));
2867	Log4(("guest.u64GPAT %#RX64\n", pVmcb->guest.u64GPAT));
2868	Log4(("guest.u64DBGCTL %#RX64\n", pVmcb->guest.u64DBGCTL));
2869	Log4(("guest.u64BR_FROM %#RX64\n", pVmcb->guest.u64BR_FROM));
2870	Log4(("guest.u64BR_TO %#RX64\n", pVmcb->guest.u64BR_TO));
2871	Log4(("guest.u64LASTEXCPFROM %#RX64\n", pVmcb->guest.u64LASTEXCPFROM));
2872	Log4(("guest.u64LASTEXCPTO %#RX64\n", pVmcb->guest.u64LASTEXCPTO));
2873	#else
2874	NOREF(pVmcb);
2875	#endif /* VBOX_STRICT */
2876	}
2877	else
2878	Log4(("hmR0SvmReportWorldSwitchError: rcVMRun=%d\n", rcVMRun));
2879	}
2880
2881
2882	/**
2883	* Check per-VM and per-VCPU force flag actions that require us to go back to
2884	* ring-3 for one reason or another.
2885	*
2886	* @returns VBox status code (information status code included).
2887	* @retval VINF_SUCCESS if we don't have any actions that require going back to
2888	* ring-3.
2889	* @retval VINF_PGM_SYNC_CR3 if we have pending PGM CR3 sync.
2890	* @retval VINF_EM_PENDING_REQUEST if we have pending requests (like hardware
2891	* interrupts)
2892	* @retval VINF_PGM_POOL_FLUSH_PENDING if PGM is doing a pool flush and requires
2893	* all EMTs to be in ring-3.
2894	* @retval VINF_EM_RAW_TO_R3 if there is pending DMA requests.
2895	* @retval VINF_EM_NO_MEMORY PGM is out of memory, we need to return
2896	* to the EM loop.
2897	*
2898	* @param pVM Pointer to the VM.
2899	* @param pVCpu Pointer to the VMCPU.
2900	* @param pCtx Pointer to the guest-CPU context.
2901	*/
2902	static int hmR0SvmCheckForceFlags(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
2903	{
2904	Assert(VMMRZCallRing3IsEnabled(pVCpu));
2905
2906	/* On AMD-V we don't need to update CR3, PAE PDPES lazily. See hmR0SvmSaveGuestState(). */
2907	Assert(!VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_UPDATE_CR3));
2908	Assert(!VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_UPDATE_PAE_PDPES));
2909
2910	if ( VM_FF_IS_PENDING(pVM, !pVCpu->hm.s.fSingleInstruction
2911	? VM_FF_HP_R0_PRE_HM_MASK : VM_FF_HP_R0_PRE_HM_STEP_MASK)
2912	\|\| VMCPU_FF_IS_PENDING(pVCpu, !pVCpu->hm.s.fSingleInstruction
2913	? VMCPU_FF_HP_R0_PRE_HM_MASK : VMCPU_FF_HP_R0_PRE_HM_STEP_MASK) )
2914	{
2915	/* Pending PGM C3 sync. */
2916	if (VMCPU_FF_IS_PENDING(pVCpu,VMCPU_FF_PGM_SYNC_CR3 \| VMCPU_FF_PGM_SYNC_CR3_NON_GLOBAL))
2917	{
2918	int rc = PGMSyncCR3(pVCpu, pCtx->cr0, pCtx->cr3, pCtx->cr4, VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_PGM_SYNC_CR3));
2919	if (rc != VINF_SUCCESS)
2920	{
2921	Log4(("hmR0SvmCheckForceFlags: PGMSyncCR3 forcing us back to ring-3. rc=%d\n", rc));
2922	return rc;
2923	}
2924	}
2925
2926	/* Pending HM-to-R3 operations (critsects, timers, EMT rendezvous etc.) */
2927	/* -XXX- what was that about single stepping? */
2928	if ( VM_FF_IS_PENDING(pVM, VM_FF_HM_TO_R3_MASK)
2929	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_TO_R3_MASK))
2930	{
2931	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHmToR3FF);
2932	int rc = RT_UNLIKELY(VM_FF_IS_PENDING(pVM, VM_FF_PGM_NO_MEMORY)) ? VINF_EM_NO_MEMORY : VINF_EM_RAW_TO_R3;
2933	Log4(("hmR0SvmCheckForceFlags: HM_TO_R3 forcing us back to ring-3. rc=%d\n", rc));
2934	return rc;
2935	}
2936
2937	/* Pending VM request packets, such as hardware interrupts. */
2938	if ( VM_FF_IS_PENDING(pVM, VM_FF_REQUEST)
2939	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_REQUEST))
2940	{
2941	Log4(("hmR0SvmCheckForceFlags: Pending VM request forcing us back to ring-3\n"));
2942	return VINF_EM_PENDING_REQUEST;
2943	}
2944
2945	/* Pending PGM pool flushes. */
2946	if (VM_FF_IS_PENDING(pVM, VM_FF_PGM_POOL_FLUSH_PENDING))
2947	{
2948	Log4(("hmR0SvmCheckForceFlags: PGM pool flush pending forcing us back to ring-3\n"));
2949	return VINF_PGM_POOL_FLUSH_PENDING;
2950	}
2951
2952	/* Pending DMA requests. */
2953	if (VM_FF_IS_PENDING(pVM, VM_FF_PDM_DMA))
2954	{
2955	Log4(("hmR0SvmCheckForceFlags: Pending DMA request forcing us back to ring-3\n"));
2956	return VINF_EM_RAW_TO_R3;
2957	}
2958	}
2959
2960	return VINF_SUCCESS;
2961	}
2962
2963
2964	/**
2965	* Does the preparations before executing guest code in AMD-V.
2966	*
2967	* This may cause longjmps to ring-3 and may even result in rescheduling to the
2968	* recompiler. We must be cautious what we do here regarding committing
2969	* guest-state information into the the VMCB assuming we assuredly execute the
2970	* guest in AMD-V. If we fall back to the recompiler after updating the VMCB and
2971	* clearing the common-state (TRPM/forceflags), we must undo those changes so
2972	* that the recompiler can (and should) use them when it resumes guest
2973	* execution. Otherwise such operations must be done when we can no longer
2974	* exit to ring-3.
2975	*
2976	* @returns VBox status code (informational status codes included).
2977	* @retval VINF_SUCCESS if we can proceed with running the guest.
2978	* @retval VINF_* scheduling changes, we have to go back to ring-3.
2979	*
2980	* @param pVM Pointer to the VM.
2981	* @param pVCpu Pointer to the VMCPU.
2982	* @param pCtx Pointer to the guest-CPU context.
2983	* @param pSvmTransient Pointer to the SVM transient structure.
2984	*/
2985	static int hmR0SvmPreRunGuest(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
2986	{
2987	HMSVM_ASSERT_PREEMPT_SAFE();
2988
2989	/* Check force flag actions that might require us to go back to ring-3. */
2990	int rc = hmR0SvmCheckForceFlags(pVM, pVCpu, pCtx);
2991	if (rc != VINF_SUCCESS)
2992	return rc;
2993
2994	if (TRPMHasTrap(pVCpu))
2995	hmR0SvmTrpmTrapToPendingEvent(pVCpu);
2996	else if (!pVCpu->hm.s.Event.fPending)
2997	hmR0SvmEvaluatePendingEvent(pVCpu, pCtx);
2998
2999	#ifdef HMSVM_SYNC_FULL_GUEST_STATE
3000	HMCPU_CF_SET(pVCpu, HM_CHANGED_ALL_GUEST);
3001	#endif
3002
3003	/* Load the guest bits that are not shared with the host in any way since we can longjmp or get preempted. */
3004	rc = hmR0SvmLoadGuestState(pVM, pVCpu, pCtx);
3005	AssertRCReturn(rc, rc);
3006	STAM_COUNTER_INC(&pVCpu->hm.s.StatLoadFull);
3007
3008	/*
3009	* If we're not intercepting TPR changes in the guest, save the guest TPR before the world-switch
3010	* so we can update it on the way back if the guest changed the TPR.
3011	*/
3012	if (pVCpu->hm.s.svm.fSyncVTpr)
3013	{
3014	if (pVM->hm.s.fTPRPatchingActive)
3015	pSvmTransient->u8GuestTpr = pCtx->msrLSTAR;
3016	else
3017	{
3018	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
3019	pSvmTransient->u8GuestTpr = pVmcb->ctrl.IntCtrl.n.u8VTPR;
3020	}
3021	}
3022
3023	/*
3024	* No longjmps to ring-3 from this point on!!!
3025	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional, better than a kernel panic.
3026	* This also disables flushing of the R0-logger instance (if any).
3027	*/
3028	VMMRZCallRing3Disable(pVCpu);
3029
3030	/*
3031	* We disable interrupts so that we don't miss any interrupts that would flag preemption (IPI/timers etc.)
3032	* when thread-context hooks aren't used and we've been running with preemption disabled for a while.
3033	*
3034	* We need to check for force-flags that could've possible been altered since we last checked them (e.g.
3035	* by PDMGetInterrupt() leaving the PDM critical section, see @bugref{6398}).
3036	*
3037	* We also check a couple of other force-flags as a last opportunity to get the EMT back to ring-3 before
3038	* executing guest code.
3039	*/
3040	pSvmTransient->fEFlags = ASMIntDisableFlags();
3041	if ( VM_FF_IS_PENDING(pVM, VM_FF_EMT_RENDEZVOUS \| VM_FF_TM_VIRTUAL_SYNC)
3042	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_TO_R3_MASK))
3043	{
3044	ASMSetFlags(pSvmTransient->fEFlags);
3045	VMMRZCallRing3Enable(pVCpu);
3046	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHmToR3FF);
3047	return VINF_EM_RAW_TO_R3;
3048	}
3049	if (RTThreadPreemptIsPending(NIL_RTTHREAD))
3050	{
3051	ASMSetFlags(pSvmTransient->fEFlags);
3052	VMMRZCallRing3Enable(pVCpu);
3053	STAM_COUNTER_INC(&pVCpu->hm.s.StatPendingHostIrq);
3054	return VINF_EM_RAW_INTERRUPT;
3055	}
3056
3057	/*
3058	* If we are injecting an NMI, we must set VMCPU_FF_BLOCK_NMIS only when we are going to execute
3059	* guest code for certain (no exits to ring-3). Otherwise, we could re-read the flag on re-entry into
3060	* AMD-V and conclude that NMI inhibition is active when we have not even delivered the NMI.
3061	*
3062	* With VT-x, this is handled by the Guest interruptibility information VMCS field which will set the
3063	* VMCS field after actually delivering the NMI which we read on VM-exit to determine the state.
3064	*/
3065	if (pVCpu->hm.s.Event.fPending)
3066	{
3067	SVMEVENT Event;
3068	Event.u = pVCpu->hm.s.Event.u64IntInfo;
3069	if ( Event.n.u1Valid
3070	&& Event.n.u3Type == SVM_EVENT_NMI
3071	&& Event.n.u8Vector == X86_XCPT_NMI
3072	&& !VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_BLOCK_NMIS))
3073	{
3074	VMCPU_FF_SET(pVCpu, VMCPU_FF_BLOCK_NMIS);
3075	}
3076	}
3077
3078	return VINF_SUCCESS;
3079	}
3080
3081
3082	/**
3083	* Prepares to run guest code in AMD-V and we've committed to doing so. This
3084	* means there is no backing out to ring-3 or anywhere else at this
3085	* point.
3086	*
3087	* @param pVM Pointer to the VM.
3088	* @param pVCpu Pointer to the VMCPU.
3089	* @param pCtx Pointer to the guest-CPU context.
3090	* @param pSvmTransient Pointer to the SVM transient structure.
3091	*
3092	* @remarks Called with preemption disabled.
3093	* @remarks No-long-jump zone!!!
3094	*/
3095	static void hmR0SvmPreRunGuestCommitted(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
3096	{
3097	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
3098	Assert(VMMR0IsLogFlushDisabled(pVCpu));
3099	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
3100
3101	VMCPU_ASSERT_STATE(pVCpu, VMCPUSTATE_STARTED_HM);
3102	VMCPU_SET_STATE(pVCpu, VMCPUSTATE_STARTED_EXEC); /* Indicate the start of guest execution. */
3103
3104	hmR0SvmInjectPendingEvent(pVCpu, pCtx);
3105
3106	if ( pVCpu->hm.s.fPreloadGuestFpu
3107	&& !CPUMIsGuestFPUStateActive(pVCpu))
3108	{
3109	CPUMR0LoadGuestFPU(pVM, pVCpu, pCtx);
3110	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR0);
3111	}
3112
3113	/* Load the state shared between host and guest (FPU, debug). */
3114	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
3115	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_HOST_GUEST_SHARED_STATE))
3116	hmR0SvmLoadSharedState(pVCpu, pVmcb, pCtx);
3117	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_HOST_CONTEXT); /* Preemption might set this, nothing to do on AMD-V. */
3118	AssertMsg(!HMCPU_CF_VALUE(pVCpu), ("fContextUseFlags=%#RX32\n", HMCPU_CF_VALUE(pVCpu)));
3119
3120	/* Setup TSC offsetting. */
3121	RTCPUID idCurrentCpu = HMR0GetCurrentCpu()->idCpu;
3122	if ( pSvmTransient->fUpdateTscOffsetting
3123	\|\| idCurrentCpu != pVCpu->hm.s.idLastCpu)
3124	{
3125	hmR0SvmUpdateTscOffsetting(pVM, pVCpu);
3126	pSvmTransient->fUpdateTscOffsetting = false;
3127	}
3128
3129	/* If we've migrating CPUs, mark the VMCB Clean bits as dirty. */
3130	if (idCurrentCpu != pVCpu->hm.s.idLastCpu)
3131	pVmcb->ctrl.u64VmcbCleanBits = 0;
3132
3133	/* Store status of the shared guest-host state at the time of VMRUN. */
3134	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS) && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
3135	if (CPUMIsGuestInLongModeEx(pCtx))
3136	{
3137	pSvmTransient->fWasGuestDebugStateActive = CPUMIsGuestDebugStateActivePending(pVCpu);
3138	pSvmTransient->fWasHyperDebugStateActive = CPUMIsHyperDebugStateActivePending(pVCpu);
3139	}
3140	else
3141	#endif
3142	{
3143	pSvmTransient->fWasGuestDebugStateActive = CPUMIsGuestDebugStateActive(pVCpu);
3144	pSvmTransient->fWasHyperDebugStateActive = CPUMIsHyperDebugStateActive(pVCpu);
3145	}
3146	pSvmTransient->fWasGuestFPUStateActive = CPUMIsGuestFPUStateActive(pVCpu);
3147
3148	/* Flush the appropriate tagged-TLB entries. */
3149	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, true); /* Used for TLB-shootdowns, set this across the world switch. */
3150	hmR0SvmFlushTaggedTlb(pVCpu);
3151	Assert(HMR0GetCurrentCpu()->idCpu == pVCpu->hm.s.idLastCpu);
3152
3153	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatEntry, &pVCpu->hm.s.StatInGC, x);
3154
3155	TMNotifyStartOfExecution(pVCpu); /* Finally, notify TM to resume its clocks as we're about
3156	to start executing. */
3157
3158	/*
3159	* Save the current Host TSC_AUX and write the guest TSC_AUX to the host, so that
3160	* RDTSCPs (that don't cause exits) reads the guest MSR. See @bugref{3324}.
3161	*
3162	* This should be done -after- any RDTSCPs for obtaining the host timestamp (TM, STAM etc).
3163	*/
3164	if ( (pVM->hm.s.cpuid.u32AMDFeatureEDX & X86_CPUID_EXT_FEATURE_EDX_RDTSCP)
3165	&& !(pVmcb->ctrl.u32InterceptCtrl2 & SVM_CTRL2_INTERCEPT_RDTSCP))
3166	{
3167	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_TSC_AUX, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
3168	pVCpu->hm.s.u64HostTscAux = ASMRdMsr(MSR_K8_TSC_AUX);
3169	uint64_t u64GuestTscAux = CPUMR0GetGuestTscAux(pVCpu);
3170	if (u64GuestTscAux != pVCpu->hm.s.u64HostTscAux)
3171	ASMWrMsr(MSR_K8_TSC_AUX, u64GuestTscAux);
3172	pSvmTransient->fRestoreTscAuxMsr = true;
3173	}
3174	else
3175	{
3176	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_TSC_AUX, SVMMSREXIT_INTERCEPT_READ, SVMMSREXIT_INTERCEPT_WRITE);
3177	pSvmTransient->fRestoreTscAuxMsr = false;
3178	}
3179
3180	/* If VMCB Clean bits isn't supported by the CPU, simply mark all state-bits as dirty, indicating (re)load-from-VMCB. */
3181	if (!(pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_VMCB_CLEAN))
3182	pVmcb->ctrl.u64VmcbCleanBits = 0;
3183	}
3184
3185
3186	/**
3187	* Wrapper for running the guest code in AMD-V.
3188	*
3189	* @returns VBox strict status code.
3190	* @param pVM Pointer to the VM.
3191	* @param pVCpu Pointer to the VMCPU.
3192	* @param pCtx Pointer to the guest-CPU context.
3193	*
3194	* @remarks No-long-jump zone!!!
3195	*/
3196	DECLINLINE(int) hmR0SvmRunGuest(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3197	{
3198	/*
3199	* 64-bit Windows uses XMM registers in the kernel as the Microsoft compiler expresses floating-point operations
3200	* using SSE instructions. Some XMM registers (XMM6-XMM15) are callee-saved and thus the need for this XMM wrapper.
3201	* Refer MSDN docs. "Configuring Programs for 64-bit / x64 Software Conventions / Register Usage" for details.
3202	*/
3203	#ifdef VBOX_WITH_KERNEL_USING_XMM
3204	return HMR0SVMRunWrapXMM(pVCpu->hm.s.svm.HCPhysVmcbHost, pVCpu->hm.s.svm.HCPhysVmcb, pCtx, pVM, pVCpu,
3205	pVCpu->hm.s.svm.pfnVMRun);
3206	#else
3207	return pVCpu->hm.s.svm.pfnVMRun(pVCpu->hm.s.svm.HCPhysVmcbHost, pVCpu->hm.s.svm.HCPhysVmcb, pCtx, pVM, pVCpu);
3208	#endif
3209	}
3210
3211
3212	/**
3213	* Performs some essential restoration of state after running guest code in
3214	* AMD-V.
3215	*
3216	* @param pVM Pointer to the VM.
3217	* @param pVCpu Pointer to the VMCPU.
3218	* @param pMixedCtx Pointer to the guest-CPU context. The data maybe
3219	* out-of-sync. Make sure to update the required fields
3220	* before using them.
3221	* @param pSvmTransient Pointer to the SVM transient structure.
3222	* @param rcVMRun Return code of VMRUN.
3223	*
3224	* @remarks Called with interrupts disabled.
3225	* @remarks No-long-jump zone!!! This function will however re-enable longjmps
3226	* unconditionally when it is safe to do so.
3227	*/
3228	static void hmR0SvmPostRunGuest(PVM pVM, PVMCPU pVCpu, PCPUMCTX pMixedCtx, PSVMTRANSIENT pSvmTransient, int rcVMRun)
3229	{
3230	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
3231
3232	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, false); /* See HMInvalidatePageOnAllVCpus(): used for TLB-shootdowns. */
3233	ASMAtomicIncU32(&pVCpu->hm.s.cWorldSwitchExits); /* Initialized in vmR3CreateUVM(): used for TLB-shootdowns. */
3234
3235	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
3236	pVmcb->ctrl.u64VmcbCleanBits = HMSVM_VMCB_CLEAN_ALL; /* Mark the VMCB-state cache as unmodified by VMM. */
3237
3238	if (pSvmTransient->fRestoreTscAuxMsr)
3239	{
3240	uint64_t u64GuestTscAuxMsr = ASMRdMsr(MSR_K8_TSC_AUX);
3241	CPUMR0SetGuestTscAux(pVCpu, u64GuestTscAuxMsr);
3242	if (u64GuestTscAuxMsr != pVCpu->hm.s.u64HostTscAux)
3243	ASMWrMsr(MSR_K8_TSC_AUX, pVCpu->hm.s.u64HostTscAux);
3244	}
3245
3246	if (!(pVmcb->ctrl.u32InterceptCtrl1 & SVM_CTRL1_INTERCEPT_RDTSC))
3247	TMCpuTickSetLastSeen(pVCpu, ASMReadTSC() + pVmcb->ctrl.u64TSCOffset);
3248
3249	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatInGC, &pVCpu->hm.s.StatExit1, x);
3250	TMNotifyEndOfExecution(pVCpu); /* Notify TM that the guest is no longer running. */
3251	VMCPU_SET_STATE(pVCpu, VMCPUSTATE_STARTED_HM);
3252
3253	Assert(!(ASMGetFlags() & X86_EFL_IF));
3254	ASMSetFlags(pSvmTransient->fEFlags); /* Enable interrupts. */
3255	VMMRZCallRing3Enable(pVCpu); /* It is now safe to do longjmps to ring-3!!! */
3256
3257	/* If VMRUN failed, we can bail out early. This does -not- cover SVM_EXIT_INVALID. */
3258	if (RT_UNLIKELY(rcVMRun != VINF_SUCCESS))
3259	{
3260	Log4(("VMRUN failure: rcVMRun=%Rrc\n", rcVMRun));
3261	return;
3262	}
3263
3264	pSvmTransient->u64ExitCode = pVmcb->ctrl.u64ExitCode; /* Save the #VMEXIT reason. */
3265	HMCPU_EXIT_HISTORY_ADD(pVCpu, pVmcb->ctrl.u64ExitCode); /* Update the #VMEXIT history array. */
3266	pSvmTransient->fVectoringDoublePF = false; /* Vectoring double page-fault needs to be determined later. */
3267	pSvmTransient->fVectoringPF = false; /* Vectoring page-fault needs to be determined later. */
3268
3269	hmR0SvmSaveGuestState(pVCpu, pMixedCtx); /* Save the guest state from the VMCB to the guest-CPU context. */
3270
3271	if (RT_LIKELY(pSvmTransient->u64ExitCode != (uint64_t)SVM_EXIT_INVALID))
3272	{
3273	if (pVCpu->hm.s.svm.fSyncVTpr)
3274	{
3275	/* TPR patching (for 32-bit guests) uses LSTAR MSR for holding the TPR value, otherwise uses the VTPR. */
3276	if ( pVM->hm.s.fTPRPatchingActive
3277	&& (pMixedCtx->msrLSTAR & 0xff) != pSvmTransient->u8GuestTpr)
3278	{
3279	int rc = PDMApicSetTPR(pVCpu, pMixedCtx->msrLSTAR & 0xff);
3280	AssertRC(rc);
3281	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
3282	}
3283	else if (pSvmTransient->u8GuestTpr != pVmcb->ctrl.IntCtrl.n.u8VTPR)
3284	{
3285	int rc = PDMApicSetTPR(pVCpu, pVmcb->ctrl.IntCtrl.n.u8VTPR << 4);
3286	AssertRC(rc);
3287	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
3288	}
3289	}
3290	}
3291	}
3292
3293
3294	/**
3295	* Runs the guest code using AMD-V.
3296	*
3297	* @returns VBox status code.
3298	* @param pVM Pointer to the VM.
3299	* @param pVCpu Pointer to the VMCPU.
3300	*/
3301	static int hmR0SvmRunGuestCodeNormal(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3302	{
3303	SVMTRANSIENT SvmTransient;
3304	SvmTransient.fUpdateTscOffsetting = true;
3305	uint32_t cLoops = 0;
3306	int rc = VERR_INTERNAL_ERROR_5;
3307
3308	for (;; cLoops++)
3309	{
3310	Assert(!HMR0SuspendPending());
3311	HMSVM_ASSERT_CPU_SAFE();
3312
3313	/* Preparatory work for running guest code, this may force us to return
3314	to ring-3. This bugger disables interrupts on VINF_SUCCESS! */
3315	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatEntry, x);
3316	rc = hmR0SvmPreRunGuest(pVM, pVCpu, pCtx, &SvmTransient);
3317	if (rc != VINF_SUCCESS)
3318	break;
3319
3320	/*
3321	* No longjmps to ring-3 from this point on!!!
3322	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional, better than a kernel panic.
3323	* This also disables flushing of the R0-logger instance (if any).
3324	*/
3325	hmR0SvmPreRunGuestCommitted(pVM, pVCpu, pCtx, &SvmTransient);
3326	rc = hmR0SvmRunGuest(pVM, pVCpu, pCtx);
3327
3328	/* Restore any residual host-state and save any bits shared between host
3329	and guest into the guest-CPU state. Re-enables interrupts! */
3330	hmR0SvmPostRunGuest(pVM, pVCpu, pCtx, &SvmTransient, rc);
3331
3332	if (RT_UNLIKELY( rc != VINF_SUCCESS /* Check for VMRUN errors. */
3333	\|\| SvmTransient.u64ExitCode == (uint64_t)SVM_EXIT_INVALID)) /* Check for invalid guest-state errors. */
3334	{
3335	if (rc == VINF_SUCCESS)
3336	rc = VERR_SVM_INVALID_GUEST_STATE;
3337	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit1, x);
3338	hmR0SvmReportWorldSwitchError(pVM, pVCpu, rc, pCtx);
3339	break;
3340	}
3341
3342	/* Handle the #VMEXIT. */
3343	HMSVM_EXITCODE_STAM_COUNTER_INC(SvmTransient.u64ExitCode);
3344	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatExit1, &pVCpu->hm.s.StatExit2, x);
3345	VBOXVMM_R0_HMSVM_VMEXIT(pVCpu, pCtx, SvmTransient.u64ExitCode, (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb);
3346	rc = hmR0SvmHandleExit(pVCpu, pCtx, &SvmTransient);
3347	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit2, x);
3348	if (rc != VINF_SUCCESS)
3349	break;
3350	if (cLoops > pVM->hm.s.cMaxResumeLoops)
3351	{
3352	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchMaxResumeLoops);
3353	rc = VINF_EM_RAW_INTERRUPT;
3354	break;
3355	}
3356	}
3357
3358	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatEntry, x);
3359	return rc;
3360	}
3361
3362
3363	/**
3364	* Runs the guest code using AMD-V in single step mode.
3365	*
3366	* @returns VBox status code.
3367	* @param pVM Pointer to the VM.
3368	* @param pVCpu Pointer to the VMCPU.
3369	* @param pCtx Pointer to the guest-CPU context.
3370	*/
3371	static int hmR0SvmRunGuestCodeStep(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3372	{
3373	SVMTRANSIENT SvmTransient;
3374	SvmTransient.fUpdateTscOffsetting = true;
3375	uint32_t cLoops = 0;
3376	int rc = VERR_INTERNAL_ERROR_5;
3377	uint16_t uCsStart = pCtx->cs.Sel;
3378	uint64_t uRipStart = pCtx->rip;
3379
3380	for (;; cLoops++)
3381	{
3382	Assert(!HMR0SuspendPending());
3383	AssertMsg(pVCpu->hm.s.idEnteredCpu == RTMpCpuId(),
3384	("Illegal migration! Entered on CPU %u Current %u cLoops=%u\n", (unsigned)pVCpu->hm.s.idEnteredCpu,
3385	(unsigned)RTMpCpuId(), cLoops));
3386
3387	/* Preparatory work for running guest code, this may force us to return
3388	to ring-3. This bugger disables interrupts on VINF_SUCCESS! */
3389	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatEntry, x);
3390	rc = hmR0SvmPreRunGuest(pVM, pVCpu, pCtx, &SvmTransient);
3391	if (rc != VINF_SUCCESS)
3392	break;
3393
3394	/*
3395	* No longjmps to ring-3 from this point on!!!
3396	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional, better than a kernel panic.
3397	* This also disables flushing of the R0-logger instance (if any).
3398	*/
3399	VMMRZCallRing3Disable(pVCpu);
3400	VMMRZCallRing3RemoveNotification(pVCpu);
3401	hmR0SvmPreRunGuestCommitted(pVM, pVCpu, pCtx, &SvmTransient);
3402
3403	rc = hmR0SvmRunGuest(pVM, pVCpu, pCtx);
3404
3405	/*
3406	* Restore any residual host-state and save any bits shared between host and guest into the guest-CPU state.
3407	* This will also re-enable longjmps to ring-3 when it has reached a safe point!!!
3408	*/
3409	hmR0SvmPostRunGuest(pVM, pVCpu, pCtx, &SvmTransient, rc);
3410	if (RT_UNLIKELY( rc != VINF_SUCCESS /* Check for VMRUN errors. */
3411	\|\| SvmTransient.u64ExitCode == (uint64_t)SVM_EXIT_INVALID)) /* Check for invalid guest-state errors. */
3412	{
3413	if (rc == VINF_SUCCESS)
3414	rc = VERR_SVM_INVALID_GUEST_STATE;
3415	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit1, x);
3416	hmR0SvmReportWorldSwitchError(pVM, pVCpu, rc, pCtx);
3417	return rc;
3418	}
3419
3420	/* Handle the #VMEXIT. */
3421	HMSVM_EXITCODE_STAM_COUNTER_INC(SvmTransient.u64ExitCode);
3422	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatExit1, &pVCpu->hm.s.StatExit2, x);
3423	VBOXVMM_R0_HMSVM_VMEXIT(pVCpu, pCtx, SvmTransient.u64ExitCode, (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb);
3424	rc = hmR0SvmHandleExit(pVCpu, pCtx, &SvmTransient);
3425	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit2, x);
3426	if (rc != VINF_SUCCESS)
3427	break;
3428	if (cLoops > pVM->hm.s.cMaxResumeLoops)
3429	{
3430	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchMaxResumeLoops);
3431	rc = VINF_EM_RAW_INTERRUPT;
3432	break;
3433	}
3434
3435	/*
3436	* Did the RIP change, if so, consider it a single step.
3437	* Otherwise, make sure one of the TFs gets set.
3438	*/
3439	if ( pCtx->rip != uRipStart
3440	\|\| pCtx->cs.Sel != uCsStart)
3441	{
3442	rc = VINF_EM_DBG_STEPPED;
3443	break;
3444	}
3445	pVCpu->hm.s.fContextUseFlags \|= HM_CHANGED_GUEST_DEBUG;
3446	}
3447
3448	/*
3449	* Clear the X86_EFL_TF if necessary.
3450	*/
3451	if (pVCpu->hm.s.fClearTrapFlag)
3452	{
3453	pVCpu->hm.s.fClearTrapFlag = false;
3454	pCtx->eflags.Bits.u1TF = 0;
3455	}
3456
3457	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatEntry, x);
3458	return rc;
3459	}
3460
3461
3462	/**
3463	* Runs the guest code using AMD-V.
3464	*
3465	* @returns VBox status code.
3466	* @param pVM Pointer to the VM.
3467	* @param pVCpu Pointer to the VMCPU.
3468	* @param pCtx Pointer to the guest-CPU context.
3469	*/
3470	VMMR0DECL(int) SVMR0RunGuestCode(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3471	{
3472	Assert(VMMRZCallRing3IsEnabled(pVCpu));
3473	HMSVM_ASSERT_PREEMPT_SAFE();
3474	VMMRZCallRing3SetNotification(pVCpu, hmR0SvmCallRing3Callback, pCtx);
3475
3476	int rc;
3477	if (!pVCpu->hm.s.fSingleInstruction && !DBGFIsStepping(pVCpu))
3478	rc = hmR0SvmRunGuestCodeNormal(pVM, pVCpu, pCtx);
3479	else
3480	rc = hmR0SvmRunGuestCodeStep(pVM, pVCpu, pCtx);
3481
3482	if (rc == VERR_EM_INTERPRETER)
3483	rc = VINF_EM_RAW_EMULATE_INSTR;
3484	else if (rc == VINF_EM_RESET)
3485	rc = VINF_EM_TRIPLE_FAULT;
3486
3487	/* Prepare to return to ring-3. This will remove longjmp notifications. */
3488	hmR0SvmExitToRing3(pVM, pVCpu, pCtx, rc);
3489	Assert(!VMMRZCallRing3IsNotificationSet(pVCpu));
3490	return rc;
3491	}
3492
3493
3494	/**
3495	* Handles a #VMEXIT (for all EXITCODE values except SVM_EXIT_INVALID).
3496	*
3497	* @returns VBox status code (informational status codes included).
3498	* @param pVCpu Pointer to the VMCPU.
3499	* @param pCtx Pointer to the guest-CPU context.
3500	* @param pSvmTransient Pointer to the SVM transient structure.
3501	*/
3502	DECLINLINE(int) hmR0SvmHandleExit(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
3503	{
3504	Assert(pSvmTransient->u64ExitCode != (uint64_t)SVM_EXIT_INVALID);
3505	Assert(pSvmTransient->u64ExitCode <= SVM_EXIT_MAX);
3506
3507	/*
3508	* The ordering of the case labels is based on most-frequently-occurring #VMEXITs for most guests under
3509	* normal workloads (for some definition of "normal").
3510	*/
3511	uint32_t u32ExitCode = pSvmTransient->u64ExitCode;
3512	switch (pSvmTransient->u64ExitCode)
3513	{
3514	case SVM_EXIT_NPF:
3515	return hmR0SvmExitNestedPF(pVCpu, pCtx, pSvmTransient);
3516
3517	case SVM_EXIT_IOIO:
3518	return hmR0SvmExitIOInstr(pVCpu, pCtx, pSvmTransient);
3519
3520	case SVM_EXIT_RDTSC:
3521	return hmR0SvmExitRdtsc(pVCpu, pCtx, pSvmTransient);
3522
3523	case SVM_EXIT_RDTSCP:
3524	return hmR0SvmExitRdtscp(pVCpu, pCtx, pSvmTransient);
3525
3526	case SVM_EXIT_CPUID:
3527	return hmR0SvmExitCpuid(pVCpu, pCtx, pSvmTransient);
3528
3529	case SVM_EXIT_EXCEPTION_E: /* X86_XCPT_PF */
3530	return hmR0SvmExitXcptPF(pVCpu, pCtx, pSvmTransient);
3531
3532	case SVM_EXIT_EXCEPTION_7: /* X86_XCPT_NM */
3533	return hmR0SvmExitXcptNM(pVCpu, pCtx, pSvmTransient);
3534
3535	case SVM_EXIT_EXCEPTION_6: /* X86_XCPT_UD */
3536	return hmR0SvmExitXcptUD(pVCpu, pCtx, pSvmTransient);
3537
3538	case SVM_EXIT_EXCEPTION_10: /* X86_XCPT_MF */
3539	return hmR0SvmExitXcptMF(pVCpu, pCtx, pSvmTransient);
3540
3541	case SVM_EXIT_EXCEPTION_1: /* X86_XCPT_DB */
3542	return hmR0SvmExitXcptDB(pVCpu, pCtx, pSvmTransient);
3543
3544	case SVM_EXIT_MONITOR:
3545	return hmR0SvmExitMonitor(pVCpu, pCtx, pSvmTransient);
3546
3547	case SVM_EXIT_MWAIT:
3548	return hmR0SvmExitMwait(pVCpu, pCtx, pSvmTransient);
3549
3550	case SVM_EXIT_HLT:
3551	return hmR0SvmExitHlt(pVCpu, pCtx, pSvmTransient);
3552
3553	case SVM_EXIT_READ_CR0:
3554	case SVM_EXIT_READ_CR3:
3555	case SVM_EXIT_READ_CR4:
3556	return hmR0SvmExitReadCRx(pVCpu, pCtx, pSvmTransient);
3557
3558	case SVM_EXIT_WRITE_CR0:
3559	case SVM_EXIT_WRITE_CR3:
3560	case SVM_EXIT_WRITE_CR4:
3561	case SVM_EXIT_WRITE_CR8:
3562	return hmR0SvmExitWriteCRx(pVCpu, pCtx, pSvmTransient);
3563
3564	case SVM_EXIT_VMMCALL:
3565	return hmR0SvmExitVmmCall(pVCpu, pCtx, pSvmTransient);
3566
3567	case SVM_EXIT_VINTR:
3568	return hmR0SvmExitVIntr(pVCpu, pCtx, pSvmTransient);
3569
3570	case SVM_EXIT_INTR:
3571	case SVM_EXIT_FERR_FREEZE:
3572	case SVM_EXIT_NMI:
3573	return hmR0SvmExitIntr(pVCpu, pCtx, pSvmTransient);
3574
3575	case SVM_EXIT_MSR:
3576	return hmR0SvmExitMsr(pVCpu, pCtx, pSvmTransient);
3577
3578	case SVM_EXIT_INVLPG:
3579	return hmR0SvmExitInvlpg(pVCpu, pCtx, pSvmTransient);
3580
3581	case SVM_EXIT_WBINVD:
3582	return hmR0SvmExitWbinvd(pVCpu, pCtx, pSvmTransient);
3583
3584	case SVM_EXIT_INVD:
3585	return hmR0SvmExitInvd(pVCpu, pCtx, pSvmTransient);
3586
3587	case SVM_EXIT_RDPMC:
3588	return hmR0SvmExitRdpmc(pVCpu, pCtx, pSvmTransient);
3589
3590	default:
3591	{
3592	switch (pSvmTransient->u64ExitCode)
3593	{
3594	case SVM_EXIT_READ_DR0: case SVM_EXIT_READ_DR1: case SVM_EXIT_READ_DR2: case SVM_EXIT_READ_DR3:
3595	case SVM_EXIT_READ_DR6: case SVM_EXIT_READ_DR7: case SVM_EXIT_READ_DR8: case SVM_EXIT_READ_DR9:
3596	case SVM_EXIT_READ_DR10: case SVM_EXIT_READ_DR11: case SVM_EXIT_READ_DR12: case SVM_EXIT_READ_DR13:
3597	case SVM_EXIT_READ_DR14: case SVM_EXIT_READ_DR15:
3598	return hmR0SvmExitReadDRx(pVCpu, pCtx, pSvmTransient);
3599
3600	case SVM_EXIT_WRITE_DR0: case SVM_EXIT_WRITE_DR1: case SVM_EXIT_WRITE_DR2: case SVM_EXIT_WRITE_DR3:
3601	case SVM_EXIT_WRITE_DR6: case SVM_EXIT_WRITE_DR7: case SVM_EXIT_WRITE_DR8: case SVM_EXIT_WRITE_DR9:
3602	case SVM_EXIT_WRITE_DR10: case SVM_EXIT_WRITE_DR11: case SVM_EXIT_WRITE_DR12: case SVM_EXIT_WRITE_DR13:
3603	case SVM_EXIT_WRITE_DR14: case SVM_EXIT_WRITE_DR15:
3604	return hmR0SvmExitWriteDRx(pVCpu, pCtx, pSvmTransient);
3605
3606	case SVM_EXIT_XSETBV:
3607	return hmR0SvmExitXsetbv(pVCpu, pCtx, pSvmTransient);
3608
3609	case SVM_EXIT_TASK_SWITCH:
3610	return hmR0SvmExitTaskSwitch(pVCpu, pCtx, pSvmTransient);
3611
3612	case SVM_EXIT_IRET:
3613	return hmR0SvmExitIret(pVCpu, pCtx, pSvmTransient);
3614
3615	case SVM_EXIT_SHUTDOWN:
3616	return hmR0SvmExitShutdown(pVCpu, pCtx, pSvmTransient);
3617
3618	case SVM_EXIT_SMI:
3619	case SVM_EXIT_INIT:
3620	{
3621	/*
3622	* We don't intercept NMIs. As for INIT signals, it really shouldn't ever happen here. If it ever does,
3623	* we want to know about it so log the exit code and bail.
3624	*/
3625	AssertMsgFailed(("hmR0SvmHandleExit: Unexpected exit %#RX32\n", (uint32_t)pSvmTransient->u64ExitCode));
3626	pVCpu->hm.s.u32HMError = (uint32_t)pSvmTransient->u64ExitCode;
3627	return VERR_SVM_UNEXPECTED_EXIT;
3628	}
3629
3630	case SVM_EXIT_INVLPGA:
3631	case SVM_EXIT_RSM:
3632	case SVM_EXIT_VMRUN:
3633	case SVM_EXIT_VMLOAD:
3634	case SVM_EXIT_VMSAVE:
3635	case SVM_EXIT_STGI:
3636	case SVM_EXIT_CLGI:
3637	case SVM_EXIT_SKINIT:
3638	return hmR0SvmExitSetPendingXcptUD(pVCpu, pCtx, pSvmTransient);
3639
3640	#ifdef HMSVM_ALWAYS_TRAP_ALL_XCPTS
3641	case SVM_EXIT_EXCEPTION_0: /* X86_XCPT_DE */
3642	/* SVM_EXIT_EXCEPTION_1: / / X86_XCPT_DB - Handled above. */
3643	case SVM_EXIT_EXCEPTION_2: /* X86_XCPT_NMI */
3644	case SVM_EXIT_EXCEPTION_3: /* X86_XCPT_BP */
3645	case SVM_EXIT_EXCEPTION_4: /* X86_XCPT_OF */
3646	case SVM_EXIT_EXCEPTION_5: /* X86_XCPT_BR */
3647	/* case SVM_EXIT_EXCEPTION_6: / / X86_XCPT_UD - Handled above. */
3648	/* SVM_EXIT_EXCEPTION_7: / / X86_XCPT_NM - Handled above. */
3649	case SVM_EXIT_EXCEPTION_8: /* X86_XCPT_DF */
3650	case SVM_EXIT_EXCEPTION_9: /* X86_XCPT_CO_SEG_OVERRUN */
3651	case SVM_EXIT_EXCEPTION_A: /* X86_XCPT_TS */
3652	case SVM_EXIT_EXCEPTION_B: /* X86_XCPT_NP */
3653	case SVM_EXIT_EXCEPTION_C: /* X86_XCPT_SS */
3654	case SVM_EXIT_EXCEPTION_D: /* X86_XCPT_GP */
3655	/* SVM_EXIT_EXCEPTION_E: / / X86_XCPT_PF - Handled above. */
3656	/* SVM_EXIT_EXCEPTION_10: / / X86_XCPT_MF - Handled above. */
3657	case SVM_EXIT_EXCEPTION_11: /* X86_XCPT_AC */
3658	case SVM_EXIT_EXCEPTION_12: /* X86_XCPT_MC */
3659	case SVM_EXIT_EXCEPTION_13: /* X86_XCPT_XF */
3660	case SVM_EXIT_EXCEPTION_F: /* Reserved */
3661	case SVM_EXIT_EXCEPTION_14: case SVM_EXIT_EXCEPTION_15: case SVM_EXIT_EXCEPTION_16:
3662	case SVM_EXIT_EXCEPTION_17: case SVM_EXIT_EXCEPTION_18: case SVM_EXIT_EXCEPTION_19:
3663	case SVM_EXIT_EXCEPTION_1A: case SVM_EXIT_EXCEPTION_1B: case SVM_EXIT_EXCEPTION_1C:
3664	case SVM_EXIT_EXCEPTION_1D: case SVM_EXIT_EXCEPTION_1E: case SVM_EXIT_EXCEPTION_1F:
3665	{
3666	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
3667	SVMEVENT Event;
3668	Event.u = 0;
3669	Event.n.u1Valid = 1;
3670	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3671	Event.n.u8Vector = pSvmTransient->u64ExitCode - SVM_EXIT_EXCEPTION_0;
3672
3673	switch (Event.n.u8Vector)
3674	{
3675	case X86_XCPT_DE:
3676	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestDE);
3677	break;
3678
3679	case X86_XCPT_BP:
3680	/** Saves the wrong EIP on the stack (pointing to the int3) instead of the
3681	* next instruction. */
3682	/** @todo Investigate this later. */
3683	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestBP);
3684	break;
3685
3686	case X86_XCPT_NP:
3687	Event.n.u1ErrorCodeValid = 1;
3688	Event.n.u32ErrorCode = pVmcb->ctrl.u64ExitInfo1;
3689	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestNP);
3690	break;
3691
3692	case X86_XCPT_SS:
3693	Event.n.u1ErrorCodeValid = 1;
3694	Event.n.u32ErrorCode = pVmcb->ctrl.u64ExitInfo1;
3695	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestSS);
3696	break;
3697
3698	case X86_XCPT_GP:
3699	Event.n.u1ErrorCodeValid = 1;
3700	Event.n.u32ErrorCode = pVmcb->ctrl.u64ExitInfo1;
3701	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestGP);
3702	break;
3703
3704	default:
3705	AssertMsgFailed(("hmR0SvmHandleExit: Unexpected exit caused by exception %#x\n", Event.n.u8Vector));
3706	pVCpu->hm.s.u32HMError = Event.n.u8Vector;
3707	return VERR_SVM_UNEXPECTED_XCPT_EXIT;
3708	}
3709
3710	Log4(("#Xcpt: Vector=%#x at CS:RIP=%04x:%RGv\n", Event.n.u8Vector, pCtx->cs.Sel, (RTGCPTR)pCtx->rip));
3711	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3712	return VINF_SUCCESS;
3713	}
3714	#endif /* HMSVM_ALWAYS_TRAP_ALL_XCPTS */
3715
3716	default:
3717	{
3718	AssertMsgFailed(("hmR0SvmHandleExit: Unknown exit code %#x\n", u32ExitCode));
3719	pVCpu->hm.s.u32HMError = u32ExitCode;
3720	return VERR_SVM_UNKNOWN_EXIT;
3721	}
3722	}
3723	}
3724	}
3725	return VERR_INTERNAL_ERROR_5; /* Should never happen. */
3726	}
3727
3728
3729	#ifdef DEBUG
3730	/* Is there some generic IPRT define for this that are not in Runtime/internal/\* ?? */
3731	# define HMSVM_ASSERT_PREEMPT_CPUID_VAR() \
3732	RTCPUID const idAssertCpu = RTThreadPreemptIsEnabled(NIL_RTTHREAD) ? NIL_RTCPUID : RTMpCpuId()
3733
3734	# define HMSVM_ASSERT_PREEMPT_CPUID() \
3735	do \
3736	{ \
3737	RTCPUID const idAssertCpuNow = RTThreadPreemptIsEnabled(NIL_RTTHREAD) ? NIL_RTCPUID : RTMpCpuId(); \
3738	AssertMsg(idAssertCpu == idAssertCpuNow, ("SVM %#x, %#x\n", idAssertCpu, idAssertCpuNow)); \
3739	} while (0)
3740
3741	# define HMSVM_VALIDATE_EXIT_HANDLER_PARAMS() \
3742	do { \
3743	AssertPtr(pVCpu); \
3744	AssertPtr(pCtx); \
3745	AssertPtr(pSvmTransient); \
3746	Assert(ASMIntAreEnabled()); \
3747	HMSVM_ASSERT_PREEMPT_SAFE(); \
3748	HMSVM_ASSERT_PREEMPT_CPUID_VAR(); \
3749	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)); \
3750	HMSVM_ASSERT_PREEMPT_SAFE(); \
3751	if (VMMR0IsLogFlushDisabled(pVCpu)) \
3752	HMSVM_ASSERT_PREEMPT_CPUID(); \
3753	} while (0)
3754	#else /* Release builds */
3755	# define HMSVM_VALIDATE_EXIT_HANDLER_PARAMS() do { NOREF(pVCpu); NOREF(pCtx); NOREF(pSvmTransient); } while (0)
3756	#endif
3757
3758
3759	/**
3760	* Worker for hmR0SvmInterpretInvlpg().
3761	*
3762	* @return VBox status code.
3763	* @param pVCpu Pointer to the VMCPU.
3764	* @param pCpu Pointer to the disassembler state.
3765	* @param pCtx The guest CPU context.
3766	*/
3767	static int hmR0SvmInterpretInvlPgEx(PVMCPU pVCpu, PDISCPUSTATE pCpu, PCPUMCTX pCtx)
3768	{
3769	DISQPVPARAMVAL Param1;
3770	RTGCPTR GCPtrPage;
3771
3772	int rc = DISQueryParamVal(CPUMCTX2CORE(pCtx), pCpu, &pCpu->Param1, &Param1, DISQPVWHICH_SRC);
3773	if (RT_FAILURE(rc))
3774	return VERR_EM_INTERPRETER;
3775
3776	if ( Param1.type == DISQPV_TYPE_IMMEDIATE
3777	\|\| Param1.type == DISQPV_TYPE_ADDRESS)
3778	{
3779	if (!(Param1.flags & (DISQPV_FLAG_32 \| DISQPV_FLAG_64)))
3780	return VERR_EM_INTERPRETER;
3781
3782	GCPtrPage = Param1.val.val64;
3783	VBOXSTRICTRC rc2 = EMInterpretInvlpg(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx), GCPtrPage);
3784	rc = VBOXSTRICTRC_VAL(rc2);
3785	}
3786	else
3787	{
3788	Log4(("hmR0SvmInterpretInvlPgEx invalid parameter type %#x\n", Param1.type));
3789	rc = VERR_EM_INTERPRETER;
3790	}
3791
3792	return rc;
3793	}
3794
3795
3796	/**
3797	* Interprets INVLPG.
3798	*
3799	* @returns VBox status code.
3800	* @retval VINF_* Scheduling instructions.
3801	* @retval VERR_EM_INTERPRETER Something we can't cope with.
3802	* @retval VERR_* Fatal errors.
3803	*
3804	* @param pVM Pointer to the VM.
3805	* @param pCtx The guest CPU context.
3806	*
3807	* @remarks Updates the RIP if the instruction was executed successfully.
3808	*/
3809	static int hmR0SvmInterpretInvlpg(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3810	{
3811	/* Only allow 32 & 64 bit code. */
3812	if (CPUMGetGuestCodeBits(pVCpu) != 16)
3813	{
3814	PDISSTATE pDis = &pVCpu->hm.s.DisState;
3815	int rc = EMInterpretDisasCurrent(pVM, pVCpu, pDis, NULL /* pcbInstr */);
3816	if ( RT_SUCCESS(rc)
3817	&& pDis->pCurInstr->uOpcode == OP_INVLPG)
3818	{
3819	rc = hmR0SvmInterpretInvlPgEx(pVCpu, pDis, pCtx);
3820	if (RT_SUCCESS(rc))
3821	pCtx->rip += pDis->cbInstr;
3822	return rc;
3823	}
3824	else
3825	Log4(("hmR0SvmInterpretInvlpg: EMInterpretDisasCurrent returned %Rrc uOpCode=%#x\n", rc, pDis->pCurInstr->uOpcode));
3826	}
3827	return VERR_EM_INTERPRETER;
3828	}
3829
3830
3831	/**
3832	* Sets an invalid-opcode (#UD) exception as pending-for-injection into the VM.
3833	*
3834	* @param pVCpu Pointer to the VMCPU.
3835	*/
3836	DECLINLINE(void) hmR0SvmSetPendingXcptUD(PVMCPU pVCpu)
3837	{
3838	SVMEVENT Event;
3839	Event.u = 0;
3840	Event.n.u1Valid = 1;
3841	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3842	Event.n.u8Vector = X86_XCPT_UD;
3843	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3844	}
3845
3846
3847	/**
3848	* Sets a debug (#DB) exception as pending-for-injection into the VM.
3849	*
3850	* @param pVCpu Pointer to the VMCPU.
3851	*/
3852	DECLINLINE(void) hmR0SvmSetPendingXcptDB(PVMCPU pVCpu)
3853	{
3854	SVMEVENT Event;
3855	Event.u = 0;
3856	Event.n.u1Valid = 1;
3857	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3858	Event.n.u8Vector = X86_XCPT_DB;
3859	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3860	}
3861
3862
3863	/**
3864	* Sets a page fault (#PF) exception as pending-for-injection into the VM.
3865	*
3866	* @param pVCpu Pointer to the VMCPU.
3867	* @param pCtx Pointer to the guest-CPU context.
3868	* @param u32ErrCode The error-code for the page-fault.
3869	* @param uFaultAddress The page fault address (CR2).
3870	*
3871	* @remarks This updates the guest CR2 with @a uFaultAddress!
3872	*/
3873	DECLINLINE(void) hmR0SvmSetPendingXcptPF(PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t u32ErrCode, RTGCUINTPTR uFaultAddress)
3874	{
3875	SVMEVENT Event;
3876	Event.u = 0;
3877	Event.n.u1Valid = 1;
3878	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3879	Event.n.u8Vector = X86_XCPT_PF;
3880	Event.n.u1ErrorCodeValid = 1;
3881	Event.n.u32ErrorCode = u32ErrCode;
3882
3883	/* Update CR2 of the guest. */
3884	if (pCtx->cr2 != uFaultAddress)
3885	{
3886	pCtx->cr2 = uFaultAddress;
3887	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR2);
3888	}
3889
3890	hmR0SvmSetPendingEvent(pVCpu, &Event, uFaultAddress);
3891	}
3892
3893
3894	/**
3895	* Sets a device-not-available (#NM) exception as pending-for-injection into the
3896	* VM.
3897	*
3898	* @param pVCpu Pointer to the VMCPU.
3899	*/
3900	DECLINLINE(void) hmR0SvmSetPendingXcptNM(PVMCPU pVCpu)
3901	{
3902	SVMEVENT Event;
3903	Event.u = 0;
3904	Event.n.u1Valid = 1;
3905	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3906	Event.n.u8Vector = X86_XCPT_NM;
3907	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3908	}
3909
3910
3911	/**
3912	* Sets a math-fault (#MF) exception as pending-for-injection into the VM.
3913	*
3914	* @param pVCpu Pointer to the VMCPU.
3915	*/
3916	DECLINLINE(void) hmR0SvmSetPendingXcptMF(PVMCPU pVCpu)
3917	{
3918	SVMEVENT Event;
3919	Event.u = 0;
3920	Event.n.u1Valid = 1;
3921	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3922	Event.n.u8Vector = X86_XCPT_MF;
3923	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3924	}
3925
3926
3927	/**
3928	* Sets a double fault (#DF) exception as pending-for-injection into the VM.
3929	*
3930	* @param pVCpu Pointer to the VMCPU.
3931	*/
3932	DECLINLINE(void) hmR0SvmSetPendingXcptDF(PVMCPU pVCpu)
3933	{
3934	SVMEVENT Event;
3935	Event.u = 0;
3936	Event.n.u1Valid = 1;
3937	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3938	Event.n.u8Vector = X86_XCPT_DF;
3939	Event.n.u1ErrorCodeValid = 1;
3940	Event.n.u32ErrorCode = 0;
3941	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3942	}
3943
3944
3945	/**
3946	* Emulates a simple MOV TPR (CR8) instruction, used for TPR patching on 32-bit
3947	* guests. This simply looks up the patch record at EIP and does the required.
3948	*
3949	* This VMMCALL is used a fallback mechanism when mov to/from cr8 isn't exactly
3950	* like how we want it to be (e.g. not followed by shr 4 as is usually done for
3951	* TPR). See hmR3ReplaceTprInstr() for the details.
3952	*
3953	* @returns VBox status code.
3954	* @retval VINF_SUCCESS if the access was handled successfully.
3955	* @retval VERR_NOT_FOUND if no patch record for this RIP could be found.
3956	* @retval VERR_SVM_UNEXPECTED_PATCH_TYPE if the found patch type is invalid.
3957	*
3958	* @param pVM Pointer to the VM.
3959	* @param pVCpu Pointer to the VMCPU.
3960	* @param pCtx Pointer to the guest-CPU context.
3961	*/
3962	static int hmR0SvmEmulateMovTpr(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3963	{
3964	Log4(("Emulated VMMCall TPR access replacement at RIP=%RGv\n", pCtx->rip));
3965
3966	/*
3967	* We do this in a loop as we increment the RIP after a successful emulation
3968	* and the new RIP may be a patched instruction which needs emulation as well.
3969	*/
3970	bool fPatchFound = false;
3971	for (;;)
3972	{
3973	bool fPending;
3974	uint8_t u8Tpr;
3975
3976	PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
3977	if (!pPatch)
3978	break;
3979
3980	fPatchFound = true;
3981	switch (pPatch->enmType)
3982	{
3983	case HMTPRINSTR_READ:
3984	{
3985	int rc = PDMApicGetTPR(pVCpu, &u8Tpr, &fPending, NULL /* pu8PendingIrq */);
3986	AssertRC(rc);
3987
3988	rc = DISWriteReg32(CPUMCTX2CORE(pCtx), pPatch->uDstOperand, u8Tpr);
3989	AssertRC(rc);
3990	pCtx->rip += pPatch->cbOp;
3991	break;
3992	}
3993
3994	case HMTPRINSTR_WRITE_REG:
3995	case HMTPRINSTR_WRITE_IMM:
3996	{
3997	if (pPatch->enmType == HMTPRINSTR_WRITE_REG)
3998	{
3999	uint32_t u32Val;
4000	int rc = DISFetchReg32(CPUMCTX2CORE(pCtx), pPatch->uSrcOperand, &u32Val);
4001	AssertRC(rc);
4002	u8Tpr = u32Val;
4003	}
4004	else
4005	u8Tpr = (uint8_t)pPatch->uSrcOperand;
4006
4007	int rc2 = PDMApicSetTPR(pVCpu, u8Tpr);
4008	AssertRC(rc2);
4009	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
4010
4011	pCtx->rip += pPatch->cbOp;
4012	break;
4013	}
4014
4015	default:
4016	AssertMsgFailed(("Unexpected patch type %d\n", pPatch->enmType));
4017	pVCpu->hm.s.u32HMError = pPatch->enmType;
4018	return VERR_SVM_UNEXPECTED_PATCH_TYPE;
4019	}
4020	}
4021
4022	if (fPatchFound)
4023	return VINF_SUCCESS;
4024	return VERR_NOT_FOUND;
4025	}
4026
4027
4028	/**
4029	* Determines if an exception is a contributory exception.
4030	*
4031	* Contributory exceptions are ones which can cause double-faults unless the
4032	* original exception was a benign exception. Page-fault is intentionally not
4033	* included here as it's a conditional contributory exception.
4034	*
4035	* @returns true if the exception is contributory, false otherwise.
4036	* @param uVector The exception vector.
4037	*/
4038	DECLINLINE(bool) hmR0SvmIsContributoryXcpt(const uint32_t uVector)
4039	{
4040	switch (uVector)
4041	{
4042	case X86_XCPT_GP:
4043	case X86_XCPT_SS:
4044	case X86_XCPT_NP:
4045	case X86_XCPT_TS:
4046	case X86_XCPT_DE:
4047	return true;
4048	default:
4049	break;
4050	}
4051	return false;
4052	}
4053
4054
4055	/**
4056	* Handle a condition that occurred while delivering an event through the guest
4057	* IDT.
4058	*
4059	* @returns VBox status code (informational error codes included).
4060	* @retval VINF_SUCCESS if we should continue handling the #VMEXIT.
4061	* @retval VINF_HM_DOUBLE_FAULT if a #DF condition was detected and we ought to
4062	* continue execution of the guest which will delivery the #DF.
4063	* @retval VINF_EM_RESET if we detected a triple-fault condition.
4064	*
4065	* @param pVCpu Pointer to the VMCPU.
4066	* @param pCtx Pointer to the guest-CPU context.
4067	* @param pSvmTransient Pointer to the SVM transient structure.
4068	*
4069	* @remarks No-long-jump zone!!!
4070	*/
4071	static int hmR0SvmCheckExitDueToEventDelivery(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4072	{
4073	int rc = VINF_SUCCESS;
4074	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4075
4076	/* See AMD spec. 15.7.3 "EXITINFO Pseudo-Code". The EXITINTINFO (if valid) contains the prior exception (IDT vector)
4077	* that was trying to be delivered to the guest which caused a #VMEXIT which was intercepted (Exit vector). */
4078	if (pVmcb->ctrl.ExitIntInfo.n.u1Valid)
4079	{
4080	uint8_t uIdtVector = pVmcb->ctrl.ExitIntInfo.n.u8Vector;
4081
4082	typedef enum
4083	{
4084	SVMREFLECTXCPT_XCPT, /* Reflect the exception to the guest or for further evaluation by VMM. */
4085	SVMREFLECTXCPT_DF, /* Reflect the exception as a double-fault to the guest. */
4086	SVMREFLECTXCPT_TF, /* Indicate a triple faulted state to the VMM. */
4087	SVMREFLECTXCPT_NONE /* Nothing to reflect. */
4088	} SVMREFLECTXCPT;
4089
4090	SVMREFLECTXCPT enmReflect = SVMREFLECTXCPT_NONE;
4091	bool fReflectingNmi = false;
4092	if (pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_EXCEPTION)
4093	{
4094	if (pSvmTransient->u64ExitCode - SVM_EXIT_EXCEPTION_0 <= SVM_EXIT_EXCEPTION_1F)
4095	{
4096	uint8_t uExitVector = (uint8_t)(pSvmTransient->u64ExitCode - SVM_EXIT_EXCEPTION_0);
4097
4098	#ifdef VBOX_STRICT
4099	if ( hmR0SvmIsContributoryXcpt(uIdtVector)
4100	&& uExitVector == X86_XCPT_PF)
4101	{
4102	Log4(("IDT: Contributory #PF idCpu=%u uCR2=%#RX64\n", pVCpu->idCpu, pCtx->cr2));
4103	}
4104	#endif
4105	if ( uExitVector == X86_XCPT_PF
4106	&& uIdtVector == X86_XCPT_PF)
4107	{
4108	pSvmTransient->fVectoringDoublePF = true;
4109	Log4(("IDT: Vectoring double #PF uCR2=%#RX64\n", pCtx->cr2));
4110	}
4111	else if ( (pVmcb->ctrl.u32InterceptException & HMSVM_CONTRIBUTORY_XCPT_MASK)
4112	&& hmR0SvmIsContributoryXcpt(uExitVector)
4113	&& ( hmR0SvmIsContributoryXcpt(uIdtVector)
4114	\|\| uIdtVector == X86_XCPT_PF))
4115	{
4116	enmReflect = SVMREFLECTXCPT_DF;
4117	Log4(("IDT: Pending vectoring #DF %#RX64 uIdtVector=%#x uExitVector=%#x\n", pVCpu->hm.s.Event.u64IntInfo,
4118	uIdtVector, uExitVector));
4119	}
4120	else if (uIdtVector == X86_XCPT_DF)
4121	{
4122	enmReflect = SVMREFLECTXCPT_TF;
4123	Log4(("IDT: Pending vectoring triple-fault %#RX64 uIdtVector=%#x uExitVector=%#x\n",
4124	pVCpu->hm.s.Event.u64IntInfo, uIdtVector, uExitVector));
4125	}
4126	else
4127	enmReflect = SVMREFLECTXCPT_XCPT;
4128	}
4129	else
4130	{
4131	/*
4132	* If event delivery caused an #VMEXIT that is not an exception (e.g. #NPF) then reflect the original
4133	* exception to the guest after handling the #VMEXIT.
4134	*/
4135	enmReflect = SVMREFLECTXCPT_XCPT;
4136	}
4137	}
4138	else if ( pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_EXTERNAL_IRQ
4139	\|\| pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_NMI)
4140	{
4141	enmReflect = SVMREFLECTXCPT_XCPT;
4142	fReflectingNmi = RT_BOOL(pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_NMI);
4143
4144	if (pSvmTransient->u64ExitCode - SVM_EXIT_EXCEPTION_0 <= SVM_EXIT_EXCEPTION_1F)
4145	{
4146	uint8_t uExitVector = (uint8_t)(pSvmTransient->u64ExitCode - SVM_EXIT_EXCEPTION_0);
4147	if (uExitVector == X86_XCPT_PF)
4148	{
4149	pSvmTransient->fVectoringPF = true;
4150	Log4(("IDT: Vectoring #PF due to Ext-Int/NMI. uCR2=%#RX64\n", pCtx->cr2));
4151	}
4152	}
4153	}
4154	/* else: Ignore software interrupts (INT n) as they reoccur when restarting the instruction. */
4155
4156	switch (enmReflect)
4157	{
4158	case SVMREFLECTXCPT_XCPT:
4159	{
4160	/* If we are re-injecting the NMI, clear NMI blocking. */
4161	if (fReflectingNmi)
4162	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_BLOCK_NMIS);
4163
4164	Assert(pVmcb->ctrl.ExitIntInfo.n.u3Type != SVM_EVENT_SOFTWARE_INT);
4165	hmR0SvmSetPendingEvent(pVCpu, &pVmcb->ctrl.ExitIntInfo, 0 /* GCPtrFaultAddress */);
4166
4167	/* If uExitVector is #PF, CR2 value will be updated from the VMCB if it's a guest #PF. See hmR0SvmExitXcptPF(). */
4168	Log4(("IDT: Pending vectoring event %#RX64 ErrValid=%RTbool Err=%#RX32\n", pVmcb->ctrl.ExitIntInfo.u,
4169	!!pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid, pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode));
4170	break;
4171	}
4172
4173	case SVMREFLECTXCPT_DF:
4174	{
4175	hmR0SvmSetPendingXcptDF(pVCpu);
4176	rc = VINF_HM_DOUBLE_FAULT;
4177	break;
4178	}
4179
4180	case SVMREFLECTXCPT_TF:
4181	{
4182	rc = VINF_EM_RESET;
4183	break;
4184	}
4185
4186	default:
4187	Assert(rc == VINF_SUCCESS);
4188	break;
4189	}
4190	}
4191	Assert(rc == VINF_SUCCESS \|\| rc == VINF_HM_DOUBLE_FAULT \|\| rc == VINF_EM_RESET);
4192	NOREF(pCtx);
4193	return rc;
4194	}
4195
4196
4197	/**
4198	* Advances the guest RIP in the if the NRIP_SAVE feature is supported by the
4199	* CPU, otherwise advances the RIP by @a cb bytes.
4200	*
4201	* @param pVCpu Pointer to the VMCPU.
4202	* @param pCtx Pointer to the guest-CPU context.
4203	* @param cb RIP increment value in bytes.
4204	*
4205	* @remarks Use this function only from #VMEXIT's where the NRIP value is valid
4206	* when NRIP_SAVE is supported by the CPU!
4207	*/
4208	DECLINLINE(void) hmR0SvmUpdateRip(PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t cb)
4209	{
4210	if (pVCpu->CTX_SUFF(pVM)->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_NRIP_SAVE)
4211	{
4212	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4213	Assert(pVmcb->ctrl.u64NextRIP - pCtx->rip == cb);
4214	pCtx->rip = pVmcb->ctrl.u64NextRIP;
4215	}
4216	else
4217	pCtx->rip += cb;
4218	}
4219
4220
4221	/* -=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= */
4222	/* -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- #VMEXIT handlers -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- */
4223	/* -=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= */
4224
4225	/** @name #VMEXIT handlers.
4226	* @{
4227	*/
4228
4229	/**
4230	* #VMEXIT handler for external interrupts, NMIs, FPU assertion freeze and INIT
4231	* signals (SVM_EXIT_INTR, SVM_EXIT_NMI, SVM_EXIT_FERR_FREEZE, SVM_EXIT_INIT).
4232	*/
4233	HMSVM_EXIT_DECL hmR0SvmExitIntr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4234	{
4235	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4236
4237	if (pSvmTransient->u64ExitCode == SVM_EXIT_NMI)
4238	STAM_REL_COUNTER_INC(&pVCpu->hm.s.StatExitHostNmiInGC);
4239	else if (pSvmTransient->u64ExitCode == SVM_EXIT_INTR)
4240	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitExtInt);
4241
4242	/*
4243	* AMD-V has no preemption timer and the generic periodic preemption timer has no way to signal -before- the timer
4244	* fires if the current interrupt is our own timer or a some other host interrupt. We also cannot examine what
4245	* interrupt it is until the host actually take the interrupt.
4246	*
4247	* Going back to executing guest code here unconditionally causes random scheduling problems (observed on an
4248	* AMD Phenom 9850 Quad-Core on Windows 64-bit host).
4249	*/
4250	return VINF_EM_RAW_INTERRUPT;
4251	}
4252
4253
4254	/**
4255	* #VMEXIT handler for WBINVD (SVM_EXIT_WBINVD). Conditional #VMEXIT.
4256	*/
4257	HMSVM_EXIT_DECL hmR0SvmExitWbinvd(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4258	{
4259	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4260
4261	hmR0SvmUpdateRip(pVCpu, pCtx, 2);
4262	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitWbinvd);
4263	int rc = VINF_SUCCESS;
4264	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4265	return rc;
4266	}
4267
4268
4269	/**
4270	* #VMEXIT handler for INVD (SVM_EXIT_INVD). Unconditional #VMEXIT.
4271	*/
4272	HMSVM_EXIT_DECL hmR0SvmExitInvd(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4273	{
4274	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4275
4276	hmR0SvmUpdateRip(pVCpu, pCtx, 2);
4277	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitInvd);
4278	int rc = VINF_SUCCESS;
4279	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4280	return rc;
4281	}
4282
4283
4284	/**
4285	* #VMEXIT handler for INVD (SVM_EXIT_CPUID). Conditional #VMEXIT.
4286	*/
4287	HMSVM_EXIT_DECL hmR0SvmExitCpuid(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4288	{
4289	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4290	PVM pVM = pVCpu->CTX_SUFF(pVM);
4291	int rc = EMInterpretCpuId(pVM, pVCpu, CPUMCTX2CORE(pCtx));
4292	if (RT_LIKELY(rc == VINF_SUCCESS))
4293	{
4294	hmR0SvmUpdateRip(pVCpu, pCtx, 2);
4295	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4296	}
4297	else
4298	{
4299	AssertMsgFailed(("hmR0SvmExitCpuid: EMInterpretCpuId failed with %Rrc\n", rc));
4300	rc = VERR_EM_INTERPRETER;
4301	}
4302	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCpuid);
4303	return rc;
4304	}
4305
4306
4307	/**
4308	* #VMEXIT handler for RDTSC (SVM_EXIT_RDTSC). Conditional #VMEXIT.
4309	*/
4310	HMSVM_EXIT_DECL hmR0SvmExitRdtsc(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4311	{
4312	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4313	PVM pVM = pVCpu->CTX_SUFF(pVM);
4314	int rc = EMInterpretRdtsc(pVM, pVCpu, CPUMCTX2CORE(pCtx));
4315	if (RT_LIKELY(rc == VINF_SUCCESS))
4316	{
4317	hmR0SvmUpdateRip(pVCpu, pCtx, 2);
4318	pSvmTransient->fUpdateTscOffsetting = true;
4319
4320	/* Single step check. */
4321	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4322	}
4323	else
4324	{
4325	AssertMsgFailed(("hmR0SvmExitRdtsc: EMInterpretRdtsc failed with %Rrc\n", rc));
4326	rc = VERR_EM_INTERPRETER;
4327	}
4328	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdtsc);
4329	return rc;
4330	}
4331
4332
4333	/**
4334	* #VMEXIT handler for RDTSCP (SVM_EXIT_RDTSCP). Conditional #VMEXIT.
4335	*/
4336	HMSVM_EXIT_DECL hmR0SvmExitRdtscp(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4337	{
4338	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4339	int rc = EMInterpretRdtscp(pVCpu->CTX_SUFF(pVM), pVCpu, pCtx);
4340	if (RT_LIKELY(rc == VINF_SUCCESS))
4341	{
4342	hmR0SvmUpdateRip(pVCpu, pCtx, 3);
4343	pSvmTransient->fUpdateTscOffsetting = true;
4344	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4345	}
4346	else
4347	{
4348	AssertMsgFailed(("hmR0SvmExitRdtsc: EMInterpretRdtscp failed with %Rrc\n", rc));
4349	rc = VERR_EM_INTERPRETER;
4350	}
4351	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdtscp);
4352	return rc;
4353	}
4354
4355
4356	/**
4357	* #VMEXIT handler for RDPMC (SVM_EXIT_RDPMC). Conditional #VMEXIT.
4358	*/
4359	HMSVM_EXIT_DECL hmR0SvmExitRdpmc(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4360	{
4361	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4362	int rc = EMInterpretRdpmc(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
4363	if (RT_LIKELY(rc == VINF_SUCCESS))
4364	{
4365	hmR0SvmUpdateRip(pVCpu, pCtx, 2);
4366	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4367	}
4368	else
4369	{
4370	AssertMsgFailed(("hmR0SvmExitRdpmc: EMInterpretRdpmc failed with %Rrc\n", rc));
4371	rc = VERR_EM_INTERPRETER;
4372	}
4373	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdpmc);
4374	return rc;
4375	}
4376
4377
4378	/**
4379	* #VMEXIT handler for INVLPG (SVM_EXIT_INVLPG). Conditional #VMEXIT.
4380	*/
4381	HMSVM_EXIT_DECL hmR0SvmExitInvlpg(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4382	{
4383	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4384	PVM pVM = pVCpu->CTX_SUFF(pVM);
4385	Assert(!pVM->hm.s.fNestedPaging);
4386
4387	/** @todo Decode Assist. */
4388	int rc = hmR0SvmInterpretInvlpg(pVM, pVCpu, pCtx); /* Updates RIP if successful. */
4389	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitInvlpg);
4390	Assert(rc == VINF_SUCCESS \|\| rc == VERR_EM_INTERPRETER);
4391	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4392	return rc;
4393	}
4394
4395
4396	/**
4397	* #VMEXIT handler for HLT (SVM_EXIT_HLT). Conditional #VMEXIT.
4398	*/
4399	HMSVM_EXIT_DECL hmR0SvmExitHlt(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4400	{
4401	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4402
4403	hmR0SvmUpdateRip(pVCpu, pCtx, 1);
4404	int rc = EMShouldContinueAfterHalt(pVCpu, pCtx) ? VINF_SUCCESS : VINF_EM_HALT;
4405	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4406	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitHlt);
4407	if (rc != VINF_SUCCESS)
4408	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHltToR3);
4409	return rc;
4410	}
4411
4412
4413	/**
4414	* #VMEXIT handler for MONITOR (SVM_EXIT_MONITOR). Conditional #VMEXIT.
4415	*/
4416	HMSVM_EXIT_DECL hmR0SvmExitMonitor(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4417	{
4418	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4419	int rc = EMInterpretMonitor(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
4420	if (RT_LIKELY(rc == VINF_SUCCESS))
4421	{
4422	hmR0SvmUpdateRip(pVCpu, pCtx, 3);
4423	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4424	}
4425	else
4426	{
4427	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMonitor: EMInterpretMonitor failed with %Rrc\n", rc));
4428	rc = VERR_EM_INTERPRETER;
4429	}
4430	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitMonitor);
4431	return rc;
4432	}
4433
4434
4435	/**
4436	* #VMEXIT handler for MWAIT (SVM_EXIT_MWAIT). Conditional #VMEXIT.
4437	*/
4438	HMSVM_EXIT_DECL hmR0SvmExitMwait(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4439	{
4440	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4441	VBOXSTRICTRC rc2 = EMInterpretMWait(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
4442	int rc = VBOXSTRICTRC_VAL(rc2);
4443	if ( rc == VINF_EM_HALT
4444	\|\| rc == VINF_SUCCESS)
4445	{
4446	hmR0SvmUpdateRip(pVCpu, pCtx, 3);
4447
4448	if ( rc == VINF_EM_HALT
4449	&& EMMonitorWaitShouldContinue(pVCpu, pCtx))
4450	{
4451	rc = VINF_SUCCESS;
4452	}
4453	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4454	}
4455	else
4456	{
4457	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMwait: EMInterpretMWait failed with %Rrc\n", rc));
4458	rc = VERR_EM_INTERPRETER;
4459	}
4460	AssertMsg(rc == VINF_SUCCESS \|\| rc == VINF_EM_HALT \|\| rc == VERR_EM_INTERPRETER,
4461	("hmR0SvmExitMwait: EMInterpretMWait failed rc=%Rrc\n", rc));
4462	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitMwait);
4463	return rc;
4464	}
4465
4466
4467	/**
4468	* #VMEXIT handler for shutdown (triple-fault) (SVM_EXIT_SHUTDOWN).
4469	* Conditional #VMEXIT.
4470	*/
4471	HMSVM_EXIT_DECL hmR0SvmExitShutdown(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4472	{
4473	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4474	return VINF_EM_RESET;
4475	}
4476
4477
4478	/**
4479	* #VMEXIT handler for CRx reads (SVM_EXIT_READ_CR*). Conditional #VMEXIT.
4480	*/
4481	HMSVM_EXIT_DECL hmR0SvmExitReadCRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4482	{
4483	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4484
4485	Log4(("hmR0SvmExitReadCRx: CS:RIP=%04x:%#RX64\n", pCtx->cs.Sel, pCtx->rip));
4486
4487	/** @todo Decode Assist. */
4488	VBOXSTRICTRC rc2 = EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */);
4489	int rc = VBOXSTRICTRC_VAL(rc2);
4490	AssertMsg(rc == VINF_SUCCESS \|\| rc == VERR_EM_INTERPRETER \|\| rc == VINF_PGM_CHANGE_MODE \|\| rc == VINF_PGM_SYNC_CR3,
4491	("hmR0SvmExitReadCRx: EMInterpretInstruction failed rc=%Rrc\n", rc));
4492	Assert((pSvmTransient->u64ExitCode - SVM_EXIT_READ_CR0) <= 15);
4493	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCRxRead[pSvmTransient->u64ExitCode - SVM_EXIT_READ_CR0]);
4494	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4495	return rc;
4496	}
4497
4498
4499	/**
4500	* #VMEXIT handler for CRx writes (SVM_EXIT_WRITE_CR*). Conditional #VMEXIT.
4501	*/
4502	HMSVM_EXIT_DECL hmR0SvmExitWriteCRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4503	{
4504	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4505
4506	/** @todo Decode Assist. */
4507	VBOXSTRICTRC rcStrict = IEMExecOneBypassEx(pVCpu, CPUMCTX2CORE(pCtx), NULL);
4508	if (RT_UNLIKELY( rcStrict == VERR_IEM_ASPECT_NOT_IMPLEMENTED
4509	\|\| rcStrict == VERR_IEM_INSTR_NOT_IMPLEMENTED))
4510	rcStrict = VERR_EM_INTERPRETER;
4511	if (rcStrict == VINF_SUCCESS)
4512	{
4513	/* RIP has been updated by EMInterpretInstruction(). */
4514	Assert((pSvmTransient->u64ExitCode - SVM_EXIT_WRITE_CR0) <= 15);
4515	switch (pSvmTransient->u64ExitCode - SVM_EXIT_WRITE_CR0)
4516	{
4517	case 0: /* CR0. */
4518	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR0);
4519	break;
4520
4521	case 3: /* CR3. */
4522	Assert(!pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
4523	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR3);
4524	break;
4525
4526	case 4: /* CR4. */
4527	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR4);
4528	break;
4529
4530	case 8: /* CR8 (TPR). */
4531	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
4532	break;
4533
4534	default:
4535	AssertMsgFailed(("hmR0SvmExitWriteCRx: Invalid/Unexpected Write-CRx exit. u64ExitCode=%#RX64 %#x\n",
4536	pSvmTransient->u64ExitCode, pSvmTransient->u64ExitCode - SVM_EXIT_WRITE_CR0));
4537	break;
4538	}
4539	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
4540	}
4541	else
4542	Assert(rcStrict == VERR_EM_INTERPRETER \|\| rcStrict == VINF_PGM_CHANGE_MODE \|\| rcStrict == VINF_PGM_SYNC_CR3);
4543	return VBOXSTRICTRC_TODO(rcStrict);
4544	}
4545
4546
4547	/**
4548	* #VMEXIT handler for instructions that result in a #UD exception delivered to
4549	* the guest.
4550	*/
4551	HMSVM_EXIT_DECL hmR0SvmExitSetPendingXcptUD(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4552	{
4553	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4554	hmR0SvmSetPendingXcptUD(pVCpu);
4555	return VINF_SUCCESS;
4556	}
4557
4558
4559	/**
4560	* #VMEXIT handler for MSR read and writes (SVM_EXIT_MSR). Conditional #VMEXIT.
4561	*/
4562	HMSVM_EXIT_DECL hmR0SvmExitMsr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4563	{
4564	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4565	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4566	PVM pVM = pVCpu->CTX_SUFF(pVM);
4567
4568	int rc;
4569	if (pVmcb->ctrl.u64ExitInfo1 == SVM_EXIT1_MSR_WRITE)
4570	{
4571	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitWrmsr);
4572
4573	/* Handle TPR patching; intercepted LSTAR write. */
4574	if ( pVM->hm.s.fTPRPatchingActive
4575	&& pCtx->ecx == MSR_K8_LSTAR)
4576	{
4577	if ((pCtx->eax & 0xff) != pSvmTransient->u8GuestTpr)
4578	{
4579	/* Our patch code uses LSTAR for TPR caching for 32-bit guests. */
4580	int rc2 = PDMApicSetTPR(pVCpu, pCtx->eax & 0xff);
4581	AssertRC(rc2);
4582	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
4583	}
4584	hmR0SvmUpdateRip(pVCpu, pCtx, 2);
4585	rc = VINF_SUCCESS;
4586	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4587	return rc;
4588	}
4589
4590	if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_NRIP_SAVE)
4591	{
4592	rc = EMInterpretWrmsr(pVM, pVCpu, CPUMCTX2CORE(pCtx));
4593	if (RT_LIKELY(rc == VINF_SUCCESS))
4594	{
4595	pCtx->rip = pVmcb->ctrl.u64NextRIP;
4596	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4597	}
4598	else
4599	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMsr: EMInterpretWrmsr failed rc=%Rrc\n", rc));
4600	}
4601	else
4602	{
4603	rc = VBOXSTRICTRC_TODO(EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */));
4604	if (RT_LIKELY(rc == VINF_SUCCESS))
4605	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc); /* RIP updated by EMInterpretInstruction(). */
4606	else
4607	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMsr: WrMsr. EMInterpretInstruction failed rc=%Rrc\n", rc));
4608	}
4609
4610	if (rc == VINF_SUCCESS)
4611	{
4612	/* If this is an X2APIC WRMSR access, update the APIC state as well. */
4613	if ( pCtx->ecx >= MSR_IA32_X2APIC_START
4614	&& pCtx->ecx <= MSR_IA32_X2APIC_END)
4615	{
4616	/*
4617	* We've already saved the APIC related guest-state (TPR) in hmR0SvmPostRunGuest(). When full APIC register
4618	* virtualization is implemented we'll have to make sure APIC state is saved from the VMCB before
4619	* EMInterpretWrmsr() changes it.
4620	*/
4621	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
4622	}
4623	else if (pCtx->ecx == MSR_K6_EFER)
4624	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_EFER_MSR);
4625	else if (pCtx->ecx == MSR_IA32_TSC)
4626	pSvmTransient->fUpdateTscOffsetting = true;
4627	}
4628	}
4629	else
4630	{
4631	/* MSR Read access. */
4632	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdmsr);
4633	Assert(pVmcb->ctrl.u64ExitInfo1 == SVM_EXIT1_MSR_READ);
4634
4635	if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_NRIP_SAVE)
4636	{
4637	rc = EMInterpretRdmsr(pVM, pVCpu, CPUMCTX2CORE(pCtx));
4638	if (RT_LIKELY(rc == VINF_SUCCESS))
4639	{
4640	pCtx->rip = pVmcb->ctrl.u64NextRIP;
4641	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4642	}
4643	else
4644	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMsr: EMInterpretRdmsr failed rc=%Rrc\n", rc));
4645	}
4646	else
4647	{
4648	rc = VBOXSTRICTRC_TODO(EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0));
4649	if (RT_UNLIKELY(rc != VINF_SUCCESS))
4650	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMsr: RdMsr. EMInterpretInstruction failed rc=%Rrc\n", rc));
4651	/* RIP updated by EMInterpretInstruction(). */
4652	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4653	}
4654	}
4655
4656	/* RIP has been updated by EMInterpret[Rd\|Wr]msr(). */
4657	return rc;
4658	}
4659
4660
4661	/**
4662	* #VMEXIT handler for DRx read (SVM_EXIT_READ_DRx). Conditional #VMEXIT.
4663	*/
4664	HMSVM_EXIT_DECL hmR0SvmExitReadDRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4665	{
4666	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4667	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitDRxRead);
4668
4669	/* We should -not- get this #VMEXIT if the guest's debug registers were active. */
4670	if (pSvmTransient->fWasGuestDebugStateActive)
4671	{
4672	AssertMsgFailed(("hmR0SvmHandleExit: Unexpected exit %#RX32\n", (uint32_t)pSvmTransient->u64ExitCode));
4673	pVCpu->hm.s.u32HMError = (uint32_t)pSvmTransient->u64ExitCode;
4674	return VERR_SVM_UNEXPECTED_EXIT;
4675	}
4676
4677	/*
4678	* Lazy DR0-3 loading.
4679	*/
4680	if (!pSvmTransient->fWasHyperDebugStateActive)
4681	{
4682	Assert(!DBGFIsStepping(pVCpu)); Assert(!pVCpu->hm.s.fSingleInstruction);
4683	Log5(("hmR0SvmExitReadDRx: Lazy loading guest debug registers\n"));
4684
4685	/* Don't intercept DRx read and writes. */
4686	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4687	pVmcb->ctrl.u16InterceptRdDRx = 0;
4688	pVmcb->ctrl.u16InterceptWrDRx = 0;
4689	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
4690
4691	/* We're playing with the host CPU state here, make sure we don't preempt or longjmp. */
4692	VMMRZCallRing3Disable(pVCpu);
4693	HM_DISABLE_PREEMPT();
4694
4695	/* Save the host & load the guest debug state, restart execution of the MOV DRx instruction. */
4696	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
4697	Assert(CPUMIsGuestDebugStateActive(pVCpu) \|\| HC_ARCH_BITS == 32);
4698
4699	HM_RESTORE_PREEMPT();
4700	VMMRZCallRing3Enable(pVCpu);
4701
4702	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxContextSwitch);
4703	return VINF_SUCCESS;
4704	}
4705
4706	/*
4707	* Interpret the read/writing of DRx.
4708	*/
4709	/** @todo Decode assist. */
4710	VBOXSTRICTRC rc = EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */);
4711	Log5(("hmR0SvmExitReadDRx: Emulated DRx access: rc=%Rrc\n", VBOXSTRICTRC_VAL(rc)));
4712	if (RT_LIKELY(rc == VINF_SUCCESS))
4713	{
4714	/* Not necessary for read accesses but whatever doesn't hurt for now, will be fixed with decode assist. */
4715	/** @todo CPUM should set this flag! */
4716	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_DEBUG);
4717	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4718	}
4719	else
4720	Assert(rc == VERR_EM_INTERPRETER);
4721	return VBOXSTRICTRC_TODO(rc);
4722	}
4723
4724
4725	/**
4726	* #VMEXIT handler for DRx write (SVM_EXIT_WRITE_DRx). Conditional #VMEXIT.
4727	*/
4728	HMSVM_EXIT_DECL hmR0SvmExitWriteDRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4729	{
4730	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4731	/* For now it's the same since we interpret the instruction anyway. Will change when using of Decode Assist is implemented. */
4732	int rc = hmR0SvmExitReadDRx(pVCpu, pCtx, pSvmTransient);
4733	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitDRxWrite);
4734	STAM_COUNTER_DEC(&pVCpu->hm.s.StatExitDRxRead);
4735	return rc;
4736	}
4737
4738
4739	/**
4740	* #VMEXIT handler for XCRx write (SVM_EXIT_XSETBV). Conditional #VMEXIT.
4741	*/
4742	HMSVM_EXIT_DECL hmR0SvmExitXsetbv(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4743	{
4744	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4745
4746	/** @todo decode assists... */
4747	VBOXSTRICTRC rcStrict = IEMExecOne(pVCpu);
4748	if (rcStrict == VINF_IEM_RAISED_XCPT)
4749	HMCPU_CF_SET(pVCpu, HM_CHANGED_ALL_GUEST);
4750
4751	pVCpu->hm.s.fLoadSaveGuestXcr0 = (pCtx->cr4 & X86_CR4_OSXSAVE) && pCtx->aXcr[0] != ASMGetXcr0();
4752	Log4(("hmR0SvmExitXsetbv: New XCR0=%#RX64 fLoadSaveGuestXcr0=%d (cr4=%RX64) rcStrict=%Rrc\n",
4753	pCtx->aXcr[0], pVCpu->hm.s.fLoadSaveGuestXcr0, pCtx->cr4, VBOXSTRICTRC_VAL(rcStrict)));
4754
4755	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
4756	return VBOXSTRICTRC_TODO(rcStrict);
4757	}
4758
4759
4760	/**
4761	* #VMEXIT handler for I/O instructions (SVM_EXIT_IOIO). Conditional #VMEXIT.
4762	*/
4763	HMSVM_EXIT_DECL hmR0SvmExitIOInstr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4764	{
4765	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4766
4767	/* I/O operation lookup arrays. */
4768	static uint32_t const s_aIOSize[8] = { 0, 1, 2, 0, 4, 0, 0, 0 }; /* Size of the I/O accesses in bytes. */
4769	static uint32_t const s_aIOOpAnd[8] = { 0, 0xff, 0xffff, 0, 0xffffffff, 0, 0, 0 }; /* AND masks for saving
4770	the result (in AL/AX/EAX). */
4771	Log4(("hmR0SvmExitIOInstr: CS:RIP=%04x:%#RX64\n", pCtx->cs.Sel, pCtx->rip));
4772
4773	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4774	PVM pVM = pVCpu->CTX_SUFF(pVM);
4775
4776	/* Refer AMD spec. 15.10.2 "IN and OUT Behaviour" and Figure 15-2. "EXITINFO1 for IOIO Intercept" for the format. */
4777	SVMIOIOEXIT IoExitInfo;
4778	IoExitInfo.u = (uint32_t)pVmcb->ctrl.u64ExitInfo1;
4779	uint32_t uIOWidth = (IoExitInfo.u >> 4) & 0x7;
4780	uint32_t cbValue = s_aIOSize[uIOWidth];
4781	uint32_t uAndVal = s_aIOOpAnd[uIOWidth];
4782
4783	if (RT_UNLIKELY(!cbValue))
4784	{
4785	AssertMsgFailed(("hmR0SvmExitIOInstr: Invalid IO operation. uIOWidth=%u\n", uIOWidth));
4786	return VERR_EM_INTERPRETER;
4787	}
4788
4789	VBOXSTRICTRC rcStrict;
4790	bool fUpdateRipAlready = false;
4791	if (IoExitInfo.n.u1STR)
4792	{
4793	#ifdef VBOX_WITH_2ND_IEM_STEP
4794	/* INS/OUTS - I/O String instruction. */
4795	/** @todo Huh? why can't we use the segment prefix information given by AMD-V
4796	* in EXITINFO1? Investigate once this thing is up and running. */
4797	Log4(("CS:RIP=%04x:%08RX64 %#06x/%u %c str\n", pCtx->cs.Sel, pCtx->rip, IoExitInfo.n.u16Port, cbValue,
4798	IoExitInfo.n.u1Type == SVM_IOIO_WRITE ? 'w' : 'r'));
4799	AssertReturn(pCtx->dx == IoExitInfo.n.u16Port, VERR_SVM_IPE_2);
4800	static IEMMODE const s_aenmAddrMode[8] =
4801	{
4802	(IEMMODE)-1, IEMMODE_16BIT, IEMMODE_32BIT, (IEMMODE)-1, IEMMODE_64BIT, (IEMMODE)-1, (IEMMODE)-1, (IEMMODE)-1
4803	};
4804	IEMMODE enmAddrMode = s_aenmAddrMode[(IoExitInfo.u >> 7) & 0x7];
4805	if (enmAddrMode != (IEMMODE)-1)
4806	{
4807	uint64_t cbInstr = pVmcb->ctrl.u64ExitInfo2 - pCtx->rip;
4808	if (cbInstr <= 15 && cbInstr >= 1)
4809	{
4810	Assert(cbInstr >= 1U + IoExitInfo.n.u1REP);
4811	if (IoExitInfo.n.u1Type == SVM_IOIO_WRITE)
4812	{
4813	/* Don't know exactly how to detect whether u3SEG is valid, currently
4814	only enabling it for Bulldozer and later with NRIP. OS/2 broke on
4815	2384 Opterons when only checking NRIP. */
4816	if ( (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_NRIP_SAVE)
4817	&& pVM->cpum.ro.GuestFeatures.enmMicroarch >= kCpumMicroarch_AMD_15h_First)
4818	{
4819	AssertMsg(IoExitInfo.n.u3SEG == X86_SREG_DS \|\| cbInstr > 1U + IoExitInfo.n.u1REP,
4820	("u32Seg=%d cbInstr=%d u1REP=%d", IoExitInfo.n.u3SEG, cbInstr, IoExitInfo.n.u1REP));
4821	rcStrict = IEMExecStringIoWrite(pVCpu, cbValue, enmAddrMode, IoExitInfo.n.u1REP, (uint8_t)cbInstr,
4822	IoExitInfo.n.u3SEG);
4823	}
4824	else if (cbInstr == 1U + IoExitInfo.n.u1REP)
4825	rcStrict = IEMExecStringIoWrite(pVCpu, cbValue, enmAddrMode, IoExitInfo.n.u1REP, (uint8_t)cbInstr,
4826	X86_SREG_DS);
4827	else
4828	rcStrict = IEMExecOne(pVCpu);
4829	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOStringWrite);
4830	}
4831	else
4832	{
4833	AssertMsg(IoExitInfo.n.u3SEG == X86_SREG_ES /=0/, ("%#x\n", IoExitInfo.n.u3SEG));
4834	rcStrict = IEMExecStringIoRead(pVCpu, cbValue, enmAddrMode, IoExitInfo.n.u1REP, (uint8_t)cbInstr);
4835	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOStringRead);
4836	}
4837	}
4838	else
4839	{
4840	AssertMsgFailed(("rip=%RX64 nrip=%#RX64 cbInstr=%#RX64\n", pCtx->rip, pVmcb->ctrl.u64ExitInfo2, cbInstr));
4841	rcStrict = IEMExecOne(pVCpu);
4842	}
4843	}
4844	else
4845	{
4846	AssertMsgFailed(("IoExitInfo=%RX64\n", IoExitInfo.u));
4847	rcStrict = IEMExecOne(pVCpu);
4848	}
4849	fUpdateRipAlready = true;
4850
4851	#else
4852	/* INS/OUTS - I/O String instruction. */
4853	PDISCPUSTATE pDis = &pVCpu->hm.s.DisState;
4854
4855	/** @todo Huh? why can't we use the segment prefix information given by AMD-V
4856	* in EXITINFO1? Investigate once this thing is up and running. */
4857
4858	rcStrict = EMInterpretDisasCurrent(pVM, pVCpu, pDis, NULL);
4859	if (rcStrict == VINF_SUCCESS)
4860	{
4861	if (IoExitInfo.n.u1Type == SVM_IOIO_WRITE)
4862	{
4863	rcStrict = IOMInterpretOUTSEx(pVM, pVCpu, CPUMCTX2CORE(pCtx), IoExitInfo.n.u16Port, pDis->fPrefix,
4864	(DISCPUMODE)pDis->uAddrMode, cbValue);
4865	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOStringWrite);
4866	}
4867	else
4868	{
4869	rcStrict = IOMInterpretINSEx(pVM, pVCpu, CPUMCTX2CORE(pCtx), IoExitInfo.n.u16Port, pDis->fPrefix,
4870	(DISCPUMODE)pDis->uAddrMode, cbValue);
4871	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOStringRead);
4872	}
4873	}
4874	else
4875	rcStrict = VINF_EM_RAW_EMULATE_INSTR;
4876	#endif
4877	}
4878	else
4879	{
4880	/* IN/OUT - I/O instruction. */
4881	Assert(!IoExitInfo.n.u1REP);
4882
4883	if (IoExitInfo.n.u1Type == SVM_IOIO_WRITE)
4884	{
4885	rcStrict = IOMIOPortWrite(pVM, pVCpu, IoExitInfo.n.u16Port, pCtx->eax & uAndVal, cbValue);
4886	if (rcStrict == VINF_IOM_R3_IOPORT_WRITE)
4887	HMR0SavePendingIOPortWrite(pVCpu, pCtx->rip, pVmcb->ctrl.u64ExitInfo2, IoExitInfo.n.u16Port, uAndVal, cbValue);
4888
4889	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOWrite);
4890	}
4891	else
4892	{
4893	uint32_t u32Val = 0;
4894	rcStrict = IOMIOPortRead(pVM, pVCpu, IoExitInfo.n.u16Port, &u32Val, cbValue);
4895	if (IOM_SUCCESS(rcStrict))
4896	{
4897	/* Save result of I/O IN instr. in AL/AX/EAX. */
4898	/** @todo r=bird: 32-bit op size should clear high bits of rax! */
4899	pCtx->eax = (pCtx->eax & ~uAndVal) \| (u32Val & uAndVal);
4900	}
4901	else if (rcStrict == VINF_IOM_R3_IOPORT_READ)
4902	HMR0SavePendingIOPortRead(pVCpu, pCtx->rip, pVmcb->ctrl.u64ExitInfo2, IoExitInfo.n.u16Port, uAndVal, cbValue);
4903
4904	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIORead);
4905	}
4906	}
4907
4908	if (IOM_SUCCESS(rcStrict))
4909	{
4910	/* AMD-V saves the RIP of the instruction following the IO instruction in EXITINFO2. */
4911	if (!fUpdateRipAlready)
4912	pCtx->rip = pVmcb->ctrl.u64ExitInfo2;
4913
4914	/*
4915	* If any I/O breakpoints are armed, we need to check if one triggered
4916	* and take appropriate action.
4917	* Note that the I/O breakpoint type is undefined if CR4.DE is 0.
4918	*/
4919	/** @todo Optimize away the DBGFBpIsHwIoArmed call by having DBGF tell the
4920	* execution engines about whether hyper BPs and such are pending. */
4921	uint32_t const uDr7 = pCtx->dr[7];
4922	if (RT_UNLIKELY( ( (uDr7 & X86_DR7_ENABLED_MASK)
4923	&& X86_DR7_ANY_RW_IO(uDr7)
4924	&& (pCtx->cr4 & X86_CR4_DE))
4925	\|\| DBGFBpIsHwIoArmed(pVM)))
4926	{
4927	/* We're playing with the host CPU state here, make sure we don't preempt or longjmp. */
4928	VMMRZCallRing3Disable(pVCpu);
4929	HM_DISABLE_PREEMPT();
4930
4931	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxIoCheck);
4932	CPUMR0DebugStateMaybeSaveGuest(pVCpu, false /fDr6/);
4933
4934	VBOXSTRICTRC rcStrict2 = DBGFBpCheckIo(pVM, pVCpu, pCtx, IoExitInfo.n.u16Port, cbValue);
4935	if (rcStrict2 == VINF_EM_RAW_GUEST_TRAP)
4936	{
4937	/* Raise #DB. */
4938	pVmcb->guest.u64DR6 = pCtx->dr[6];
4939	pVmcb->guest.u64DR7 = pCtx->dr[7];
4940	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
4941	hmR0SvmSetPendingXcptDB(pVCpu);
4942	}
4943	/* rcStrict is VINF_SUCCESS or in [VINF_EM_FIRST..VINF_EM_LAST]. */
4944	else if ( rcStrict2 != VINF_SUCCESS
4945	&& (rcStrict == VINF_SUCCESS \|\| rcStrict2 < rcStrict))
4946	rcStrict = rcStrict2;
4947
4948	HM_RESTORE_PREEMPT();
4949	VMMRZCallRing3Enable(pVCpu);
4950	}
4951
4952	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
4953	}
4954
4955	#ifdef VBOX_STRICT
4956	if (rcStrict == VINF_IOM_R3_IOPORT_READ)
4957	Assert(IoExitInfo.n.u1Type == SVM_IOIO_READ);
4958	else if (rcStrict == VINF_IOM_R3_IOPORT_WRITE)
4959	Assert(IoExitInfo.n.u1Type == SVM_IOIO_WRITE);
4960	else
4961	{
4962	/** @todo r=bird: This is missing a bunch of VINF_EM_FIRST..VINF_EM_LAST
4963	* statuses, that the VMM device and some others may return. See
4964	* IOM_SUCCESS() for guidance. */
4965	AssertMsg( RT_FAILURE(rcStrict)
4966	\|\| rcStrict == VINF_SUCCESS
4967	\|\| rcStrict == VINF_EM_RAW_EMULATE_INSTR
4968	\|\| rcStrict == VINF_EM_DBG_BREAKPOINT
4969	\|\| rcStrict == VINF_EM_RAW_GUEST_TRAP
4970	\|\| rcStrict == VINF_EM_RAW_TO_R3
4971	\|\| rcStrict == VINF_TRPM_XCPT_DISPATCHED, ("%Rrc\n", VBOXSTRICTRC_VAL(rcStrict)));
4972	}
4973	#endif
4974	return VBOXSTRICTRC_TODO(rcStrict);
4975	}
4976
4977
4978	/**
4979	* #VMEXIT handler for Nested Page-faults (SVM_EXIT_NPF). Conditional
4980	* #VMEXIT.
4981	*/
4982	HMSVM_EXIT_DECL hmR0SvmExitNestedPF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4983	{
4984	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4985	PVM pVM = pVCpu->CTX_SUFF(pVM);
4986	Assert(pVM->hm.s.fNestedPaging);
4987
4988	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
4989
4990	/* See AMD spec. 15.25.6 "Nested versus Guest Page Faults, Fault Ordering" for VMCB details for #NPF. */
4991	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4992	uint32_t u32ErrCode = pVmcb->ctrl.u64ExitInfo1;
4993	RTGCPHYS GCPhysFaultAddr = pVmcb->ctrl.u64ExitInfo2;
4994
4995	Log4(("#NPF at CS:RIP=%04x:%#RX64 faultaddr=%RGp errcode=%#x \n", pCtx->cs.Sel, pCtx->rip, GCPhysFaultAddr, u32ErrCode));
4996
4997	#ifdef VBOX_HM_WITH_GUEST_PATCHING
4998	/* TPR patching for 32-bit guests, using the reserved bit in the page tables for MMIO regions. */
4999	if ( pVM->hm.s.fTprPatchingAllowed
5000	&& (GCPhysFaultAddr & PAGE_OFFSET_MASK) == 0x80 /* TPR offset. */
5001	&& ( !(u32ErrCode & X86_TRAP_PF_P) /* Not present */
5002	\|\| (u32ErrCode & (X86_TRAP_PF_P \| X86_TRAP_PF_RSVD)) == (X86_TRAP_PF_P \| X86_TRAP_PF_RSVD)) /* MMIO page. */
5003	&& !CPUMIsGuestInLongModeEx(pCtx)
5004	&& !CPUMGetGuestCPL(pVCpu)
5005	&& pVM->hm.s.cPatches < RT_ELEMENTS(pVM->hm.s.aPatches))
5006	{
5007	RTGCPHYS GCPhysApicBase = pCtx->msrApicBase;
5008	GCPhysApicBase &= PAGE_BASE_GC_MASK;
5009
5010	if (GCPhysFaultAddr == GCPhysApicBase + 0x80)
5011	{
5012	/* Only attempt to patch the instruction once. */
5013	PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
5014	if (!pPatch)
5015	return VINF_EM_HM_PATCH_TPR_INSTR;
5016	}
5017	}
5018	#endif
5019
5020	/*
5021	* Determine the nested paging mode.
5022	*/
5023	PGMMODE enmNestedPagingMode;
5024	#if HC_ARCH_BITS == 32
5025	if (CPUMIsGuestInLongModeEx(pCtx))
5026	enmNestedPagingMode = PGMMODE_AMD64_NX;
5027	else
5028	#endif
5029	enmNestedPagingMode = PGMGetHostMode(pVM);
5030
5031	/*
5032	* MMIO optimization using the reserved (RSVD) bit in the guest page tables for MMIO pages.
5033	*/
5034	int rc;
5035	Assert((u32ErrCode & (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P)) != X86_TRAP_PF_RSVD);
5036	if ((u32ErrCode & (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P)) == (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P))
5037	{
5038	VBOXSTRICTRC rc2 = PGMR0Trap0eHandlerNPMisconfig(pVM, pVCpu, enmNestedPagingMode, CPUMCTX2CORE(pCtx), GCPhysFaultAddr,
5039	u32ErrCode);
5040	rc = VBOXSTRICTRC_VAL(rc2);
5041
5042	/*
5043	* If we succeed, resume guest execution.
5044	* If we fail in interpreting the instruction because we couldn't get the guest physical address
5045	* of the page containing the instruction via the guest's page tables (we would invalidate the guest page
5046	* in the host TLB), resume execution which would cause a guest page fault to let the guest handle this
5047	* weird case. See @bugref{6043}.
5048	*/
5049	if ( rc == VINF_SUCCESS
5050	\|\| rc == VERR_PAGE_TABLE_NOT_PRESENT
5051	\|\| rc == VERR_PAGE_NOT_PRESENT)
5052	{
5053	/* Successfully handled MMIO operation. */
5054	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
5055	rc = VINF_SUCCESS;
5056	}
5057	return rc;
5058	}
5059
5060	TRPMAssertXcptPF(pVCpu, GCPhysFaultAddr, u32ErrCode);
5061	rc = PGMR0Trap0eHandlerNestedPaging(pVM, pVCpu, enmNestedPagingMode, u32ErrCode, CPUMCTX2CORE(pCtx), GCPhysFaultAddr);
5062	TRPMResetTrap(pVCpu);
5063
5064	Log4(("#NPF: PGMR0Trap0eHandlerNestedPaging returned %Rrc CS:RIP=%04x:%#RX64\n", rc, pCtx->cs.Sel, pCtx->rip));
5065
5066	/*
5067	* Same case as PGMR0Trap0eHandlerNPMisconfig(). See comment above, @bugref{6043}.
5068	*/
5069	if ( rc == VINF_SUCCESS
5070	\|\| rc == VERR_PAGE_TABLE_NOT_PRESENT
5071	\|\| rc == VERR_PAGE_NOT_PRESENT)
5072	{
5073	/* We've successfully synced our shadow page tables. */
5074	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPF);
5075	rc = VINF_SUCCESS;
5076	}
5077
5078	return rc;
5079	}
5080
5081
5082	/**
5083	* #VMEXIT handler for virtual interrupt (SVM_EXIT_VINTR). Conditional #VMEXIT.
5084	*/
5085	HMSVM_EXIT_DECL hmR0SvmExitVIntr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5086	{
5087	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5088
5089	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
5090	pVmcb->ctrl.IntCtrl.n.u1VIrqValid = 0; /* No virtual interrupts pending, we'll inject the current one/NMI before reentry. */
5091	pVmcb->ctrl.IntCtrl.n.u8VIrqVector = 0;
5092
5093	/* Indicate that we no longer need to #VMEXIT when the guest is ready to receive interrupts/NMIs, it is now ready. */
5094	pVmcb->ctrl.u32InterceptCtrl1 &= ~SVM_CTRL1_INTERCEPT_VINTR;
5095	pVmcb->ctrl.u64VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS \| HMSVM_VMCB_CLEAN_TPR);
5096
5097	/* Deliver the pending interrupt/NMI via hmR0SvmEvaluatePendingEvent() and resume guest execution. */
5098	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIntWindow);
5099	return VINF_SUCCESS;
5100	}
5101
5102
5103	/**
5104	* #VMEXIT handler for task switches (SVM_EXIT_TASK_SWITCH). Conditional #VMEXIT.
5105	*/
5106	HMSVM_EXIT_DECL hmR0SvmExitTaskSwitch(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5107	{
5108	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5109
5110	#ifndef HMSVM_ALWAYS_TRAP_TASK_SWITCH
5111	Assert(!pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
5112	#endif
5113
5114	/* Check if this task-switch occurred while delivery an event through the guest IDT. */
5115	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
5116	if ( !(pVmcb->ctrl.u64ExitInfo2 & (SVM_EXIT2_TASK_SWITCH_IRET \| SVM_EXIT2_TASK_SWITCH_JMP))
5117	&& pVCpu->hm.s.Event.fPending) /** @todo fPending cannot be 'true', see hmR0SvmInjectPendingEvent(). See @bugref{7362}.*/
5118	{
5119	/*
5120	* AMD-V does not provide us with the original exception but we have it in u64IntInfo since we
5121	* injected the event during VM-entry.
5122	*/
5123	Log4(("hmR0SvmExitTaskSwitch: TS occurred during event delivery.\n"));
5124	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitTaskSwitch);
5125	return VINF_EM_RAW_INJECT_TRPM_EVENT;
5126	}
5127
5128	/** @todo Emulate task switch someday, currently just going back to ring-3 for
5129	* emulation. */
5130	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitTaskSwitch);
5131	return VERR_EM_INTERPRETER;
5132	}
5133
5134
5135	/**
5136	* #VMEXIT handler for VMMCALL (SVM_EXIT_VMMCALL). Conditional #VMEXIT.
5137	*/
5138	HMSVM_EXIT_DECL hmR0SvmExitVmmCall(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5139	{
5140	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5141	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitVmcall);
5142
5143	/* First check if this is a patched VMMCALL for mov TPR */
5144	int rc = hmR0SvmEmulateMovTpr(pVCpu->CTX_SUFF(pVM), pVCpu, pCtx);
5145	if (rc == VINF_SUCCESS)
5146	{
5147	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
5148	return VINF_SUCCESS;
5149	}
5150	else if (rc == VERR_NOT_FOUND)
5151	{
5152	if (pVCpu->hm.s.fHypercallsEnabled)
5153	{
5154	rc = GIMHypercall(pVCpu, pCtx);
5155	if (RT_SUCCESS(rc))
5156	{
5157	/* If the hypercall changes anything other than guest general-purpose registers,
5158	we would need to reload the guest changed bits here before VM-reentry. */
5159	hmR0SvmUpdateRip(pVCpu, pCtx, 3);
5160	return VINF_SUCCESS;
5161	}
5162	}
5163	}
5164
5165	hmR0SvmSetPendingXcptUD(pVCpu);
5166	return VINF_SUCCESS;
5167	}
5168
5169
5170	/**
5171	* #VMEXIT handler for IRET (SVM_EXIT_IRET). Conditional #VMEXIT.
5172	*/
5173	HMSVM_EXIT_DECL hmR0SvmExitIret(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5174	{
5175	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5176
5177	/* Clear NMI blocking. */
5178	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_BLOCK_NMIS);
5179
5180	/* Indicate that we no longer need to #VMEXIT when the guest is ready to receive NMIs, it is now ready. */
5181	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
5182	hmR0SvmClearIretIntercept(pVmcb);
5183
5184	/* Deliver the pending NMI via hmR0SvmEvaluatePendingEvent() and resume guest execution. */
5185	return VINF_SUCCESS;
5186	}
5187
5188
5189	/**
5190	* #VMEXIT handler for page-fault exceptions (SVM_EXIT_EXCEPTION_E). Conditional
5191	* #VMEXIT.
5192	*/
5193	HMSVM_EXIT_DECL hmR0SvmExitXcptPF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5194	{
5195	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5196
5197	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
5198
5199	/* See AMD spec. 15.12.15 "#PF (Page Fault)". */
5200	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
5201	uint32_t u32ErrCode = pVmcb->ctrl.u64ExitInfo1;
5202	RTGCUINTPTR uFaultAddress = pVmcb->ctrl.u64ExitInfo2;
5203	PVM pVM = pVCpu->CTX_SUFF(pVM);
5204
5205	#if defined(HMSVM_ALWAYS_TRAP_ALL_XCPTS) \|\| defined(HMSVM_ALWAYS_TRAP_PF)
5206	if (pVM->hm.s.fNestedPaging)
5207	{
5208	pVCpu->hm.s.Event.fPending = false; /* In case it's a contributory or vectoring #PF. */
5209	if (!pSvmTransient->fVectoringDoublePF)
5210	{
5211	/* A genuine guest #PF, reflect it to the guest. */
5212	hmR0SvmSetPendingXcptPF(pVCpu, pCtx, u32ErrCode, uFaultAddress);
5213	Log4(("#PF: Guest page fault at %04X:%RGv FaultAddr=%RGv ErrCode=%#x\n", pCtx->cs.Sel, (RTGCPTR)pCtx->rip,
5214	uFaultAddress, u32ErrCode));
5215	}
5216	else
5217	{
5218	/* A guest page-fault occurred during delivery of a page-fault. Inject #DF. */
5219	hmR0SvmSetPendingXcptDF(pVCpu);
5220	Log4(("Pending #DF due to vectoring #PF. NP\n"));
5221	}
5222	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestPF);
5223	return VINF_SUCCESS;
5224	}
5225	#endif
5226
5227	Assert(!pVM->hm.s.fNestedPaging);
5228
5229	#ifdef VBOX_HM_WITH_GUEST_PATCHING
5230	/* Shortcut for APIC TPR reads and writes; only applicable to 32-bit guests. */
5231	if ( pVM->hm.s.fTprPatchingAllowed
5232	&& (uFaultAddress & 0xfff) == 0x80 /* TPR offset. */
5233	&& !(u32ErrCode & X86_TRAP_PF_P) /* Not present. */
5234	&& !CPUMIsGuestInLongModeEx(pCtx)
5235	&& !CPUMGetGuestCPL(pVCpu)
5236	&& pVM->hm.s.cPatches < RT_ELEMENTS(pVM->hm.s.aPatches))
5237	{
5238	RTGCPHYS GCPhysApicBase;
5239	GCPhysApicBase = pCtx->msrApicBase;
5240	GCPhysApicBase &= PAGE_BASE_GC_MASK;
5241
5242	/* Check if the page at the fault-address is the APIC base. */
5243	RTGCPHYS GCPhysPage;
5244	int rc2 = PGMGstGetPage(pVCpu, (RTGCPTR)uFaultAddress, NULL /* pfFlags */, &GCPhysPage);
5245	if ( rc2 == VINF_SUCCESS
5246	&& GCPhysPage == GCPhysApicBase)
5247	{
5248	/* Only attempt to patch the instruction once. */
5249	PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
5250	if (!pPatch)
5251	return VINF_EM_HM_PATCH_TPR_INSTR;
5252	}
5253	}
5254	#endif
5255
5256	Log4(("#PF: uFaultAddress=%#RX64 CS:RIP=%#04x:%#RX64 u32ErrCode %#RX32 cr3=%#RX64\n", uFaultAddress, pCtx->cs.Sel,
5257	pCtx->rip, u32ErrCode, pCtx->cr3));
5258
5259	/* If it's a vectoring #PF, emulate injecting the original event injection as PGMTrap0eHandler() is incapable
5260	of differentiating between instruction emulation and event injection that caused a #PF. See @bugref{6607}. */
5261	if (pSvmTransient->fVectoringPF)
5262	{
5263	Assert(pVCpu->hm.s.Event.fPending);
5264	return VINF_EM_RAW_INJECT_TRPM_EVENT;
5265	}
5266
5267	TRPMAssertXcptPF(pVCpu, uFaultAddress, u32ErrCode);
5268	int rc = PGMTrap0eHandler(pVCpu, u32ErrCode, CPUMCTX2CORE(pCtx), (RTGCPTR)uFaultAddress);
5269
5270	Log4(("#PF rc=%Rrc\n", rc));
5271
5272	if (rc == VINF_SUCCESS)
5273	{
5274	/* Successfully synced shadow pages tables or emulated an MMIO instruction. */
5275	TRPMResetTrap(pVCpu);
5276	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPF);
5277	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
5278	return rc;
5279	}
5280	else if (rc == VINF_EM_RAW_GUEST_TRAP)
5281	{
5282	pVCpu->hm.s.Event.fPending = false; /* In case it's a contributory or vectoring #PF. */
5283
5284	if (!pSvmTransient->fVectoringDoublePF)
5285	{
5286	/* It's a guest page fault and needs to be reflected to the guest. */
5287	u32ErrCode = TRPMGetErrorCode(pVCpu); /* The error code might have been changed. */
5288	TRPMResetTrap(pVCpu);
5289	hmR0SvmSetPendingXcptPF(pVCpu, pCtx, u32ErrCode, uFaultAddress);
5290	}
5291	else
5292	{
5293	/* A guest page-fault occurred during delivery of a page-fault. Inject #DF. */
5294	TRPMResetTrap(pVCpu);
5295	hmR0SvmSetPendingXcptDF(pVCpu);
5296	Log4(("#PF: Pending #DF due to vectoring #PF\n"));
5297	}
5298
5299	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestPF);
5300	return VINF_SUCCESS;
5301	}
5302
5303	TRPMResetTrap(pVCpu);
5304	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPFEM);
5305	return rc;
5306	}
5307
5308
5309	/**
5310	* #VMEXIT handler for device-not-available exceptions (SVM_EXIT_EXCEPTION_7).
5311	* Conditional #VMEXIT.
5312	*/
5313	HMSVM_EXIT_DECL hmR0SvmExitXcptNM(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5314	{
5315	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5316
5317	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
5318
5319	/* We're playing with the host CPU state here, make sure we don't preempt or longjmp. */
5320	VMMRZCallRing3Disable(pVCpu);
5321	HM_DISABLE_PREEMPT();
5322
5323	int rc;
5324	/* If the guest FPU was active at the time of the #NM exit, then it's a guest fault. */
5325	if (pSvmTransient->fWasGuestFPUStateActive)
5326	{
5327	rc = VINF_EM_RAW_GUEST_TRAP;
5328	Assert(CPUMIsGuestFPUStateActive(pVCpu) \|\| HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR0));
5329	}
5330	else
5331	{
5332	#ifndef HMSVM_ALWAYS_TRAP_ALL_XCPTS
5333	Assert(!pSvmTransient->fWasGuestFPUStateActive);
5334	#endif
5335	rc = CPUMR0Trap07Handler(pVCpu->CTX_SUFF(pVM), pVCpu, pCtx);
5336	Assert(rc == VINF_EM_RAW_GUEST_TRAP \|\| (rc == VINF_SUCCESS && CPUMIsGuestFPUStateActive(pVCpu)));
5337	}
5338
5339	HM_RESTORE_PREEMPT();
5340	VMMRZCallRing3Enable(pVCpu);
5341
5342	if (rc == VINF_SUCCESS)
5343	{
5344	/* Guest FPU state was activated, we'll want to change CR0 FPU intercepts before the next VM-reentry. */
5345	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR0);
5346	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowNM);
5347	pVCpu->hm.s.fPreloadGuestFpu = true;
5348	}
5349	else
5350	{
5351	/* Forward #NM to the guest. */
5352	Assert(rc == VINF_EM_RAW_GUEST_TRAP);
5353	hmR0SvmSetPendingXcptNM(pVCpu);
5354	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestNM);
5355	}
5356	return VINF_SUCCESS;
5357	}
5358
5359
5360	/**
5361	* #VMEXIT handler for undefined opcode (SVM_EXIT_EXCEPTION_6).
5362	* Conditional #VMEXIT.
5363	*/
5364	HMSVM_EXIT_DECL hmR0SvmExitXcptUD(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5365	{
5366	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5367
5368	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
5369
5370	if (pVCpu->hm.s.fGIMTrapXcptUD)
5371	GIMXcptUD(pVCpu, pCtx, NULL /* pDis */);
5372	else
5373	hmR0SvmSetPendingXcptUD(pVCpu);
5374
5375	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestUD);
5376	return VINF_SUCCESS;
5377	}
5378
5379
5380	/**
5381	* #VMEXIT handler for math-fault exceptions (SVM_EXIT_EXCEPTION_10).
5382	* Conditional #VMEXIT.
5383	*/
5384	HMSVM_EXIT_DECL hmR0SvmExitXcptMF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5385	{
5386	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5387
5388	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
5389
5390	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestMF);
5391
5392	if (!(pCtx->cr0 & X86_CR0_NE))
5393	{
5394	PVM pVM = pVCpu->CTX_SUFF(pVM);
5395	PDISSTATE pDis = &pVCpu->hm.s.DisState;
5396	unsigned cbOp;
5397	int rc = EMInterpretDisasCurrent(pVM, pVCpu, pDis, &cbOp);
5398	if (RT_SUCCESS(rc))
5399	{
5400	/* Convert a #MF into a FERR -> IRQ 13. See @bugref{6117}. */
5401	rc = PDMIsaSetIrq(pVCpu->CTX_SUFF(pVM), 13, 1, 0 /* uTagSrc */);
5402	if (RT_SUCCESS(rc))
5403	pCtx->rip += cbOp;
5404	}
5405	else
5406	Log4(("hmR0SvmExitXcptMF: EMInterpretDisasCurrent returned %Rrc uOpCode=%#x\n", rc, pDis->pCurInstr->uOpcode));
5407	return rc;
5408	}
5409
5410	hmR0SvmSetPendingXcptMF(pVCpu);
5411	return VINF_SUCCESS;
5412	}
5413
5414
5415	/**
5416	* #VMEXIT handler for debug exceptions (SVM_EXIT_EXCEPTION_1). Conditional
5417	* #VMEXIT.
5418	*/
5419	HMSVM_EXIT_DECL hmR0SvmExitXcptDB(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5420	{
5421	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5422
5423	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
5424
5425	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestDB);
5426
5427	/* This can be a fault-type #DB (instruction breakpoint) or a trap-type #DB (data breakpoint). However, for both cases
5428	DR6 and DR7 are updated to what the exception handler expects. See AMD spec. 15.12.2 "#DB (Debug)". */
5429	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
5430	PVM pVM = pVCpu->CTX_SUFF(pVM);
5431	int rc = DBGFRZTrap01Handler(pVM, pVCpu, CPUMCTX2CORE(pCtx), pVmcb->guest.u64DR6, pVCpu->hm.s.fSingleInstruction);
5432	if (rc == VINF_EM_RAW_GUEST_TRAP)
5433	{
5434	Log5(("hmR0SvmExitXcptDB: DR6=%#RX64 -> guest trap\n", pVmcb->guest.u64DR6));
5435	if (CPUMIsHyperDebugStateActive(pVCpu))
5436	CPUMSetGuestDR6(pVCpu, CPUMGetGuestDR6(pVCpu) \| pVmcb->guest.u64DR6);
5437
5438	/* Reflect the exception back to the guest. */
5439	hmR0SvmSetPendingXcptDB(pVCpu);
5440	rc = VINF_SUCCESS;
5441	}
5442
5443	/*
5444	* Update DR6.
5445	*/
5446	if (CPUMIsHyperDebugStateActive(pVCpu))
5447	{
5448	Log5(("hmR0SvmExitXcptDB: DR6=%#RX64 -> %Rrc\n", pVmcb->guest.u64DR6, rc));
5449	pVmcb->guest.u64DR6 = X86_DR6_INIT_VAL;
5450	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
5451	}
5452	else
5453	{
5454	AssertMsg(rc == VINF_SUCCESS, ("rc=%Rrc\n", rc));
5455	Assert(!pVCpu->hm.s.fSingleInstruction && !DBGFIsStepping(pVCpu));
5456	}
5457
5458	return rc;
5459	}
5460
5461	/** @} */
5462

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

source: vbox/trunk/src/VBox/VMM/VMMR0/HMSVMR0.cpp@ 57229

Download in other formats: