HMSVMR0.cpp@ 65919

Last change on this file since 65919 was 65912, checked in by vboxsync, 8 years ago
VMM: Nested Hw.virt: unused function fix.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 218.3 KB

Line
1	/* $Id: HMSVMR0.cpp 65912 2017-03-01 12:05:29Z vboxsync $ */
2	/** @file
3	* HM SVM (AMD-V) - Host Context Ring-0.
4	*/
5
6	/*
7	* Copyright (C) 2013-2016 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	/*********************************************************************************************************************************
20	* Header Files *
21	*********************************************************************************************************************************/
22	#define LOG_GROUP LOG_GROUP_HM
23	#include <iprt/asm-amd64-x86.h>
24	#include <iprt/thread.h>
25
26	#include <VBox/vmm/pdmapi.h>
27	#include <VBox/vmm/dbgf.h>
28	#include <VBox/vmm/iem.h>
29	#include <VBox/vmm/iom.h>
30	#include <VBox/vmm/tm.h>
31	#include <VBox/vmm/gim.h>
32	#include <VBox/vmm/apic.h>
33	#include "HMInternal.h"
34	#include <VBox/vmm/vm.h>
35	#include "HMSVMR0.h"
36	#include "dtrace/VBoxVMM.h"
37
38	#ifdef DEBUG_ramshankar
39	# define HMSVM_SYNC_FULL_GUEST_STATE
40	# define HMSVM_ALWAYS_TRAP_ALL_XCPTS
41	# define HMSVM_ALWAYS_TRAP_PF
42	# define HMSVM_ALWAYS_TRAP_TASK_SWITCH
43	#endif
44
45
46	/*********************************************************************************************************************************
47	* Defined Constants And Macros *
48	*********************************************************************************************************************************/
49	#ifdef VBOX_WITH_STATISTICS
50	# define HMSVM_EXITCODE_STAM_COUNTER_INC(u64ExitCode) do { \
51	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitAll); \
52	if ((u64ExitCode) == SVM_EXIT_NPF) \
53	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitReasonNpf); \
54	else \
55	STAM_COUNTER_INC(&pVCpu->hm.s.paStatExitReasonR0[(u64ExitCode) & MASK_EXITREASON_STAT]); \
56	} while (0)
57	#else
58	# define HMSVM_EXITCODE_STAM_COUNTER_INC(u64ExitCode) do { } while (0)
59	#endif
60
61	/** If we decide to use a function table approach this can be useful to
62	* switch to a "static DECLCALLBACK(int)". */
63	#define HMSVM_EXIT_DECL static int
64
65	/** @name Segment attribute conversion between CPU and AMD-V VMCB format.
66	*
67	* The CPU format of the segment attribute is described in X86DESCATTRBITS
68	* which is 16-bits (i.e. includes 4 bits of the segment limit).
69	*
70	* The AMD-V VMCB format the segment attribute is compact 12-bits (strictly
71	* only the attribute bits and nothing else). Upper 4-bits are unused.
72	*
73	* @{ */
74	#define HMSVM_CPU_2_VMCB_SEG_ATTR(a) ( ((a) & 0xff) \| (((a) & 0xf000) >> 4) )
75	#define HMSVM_VMCB_2_CPU_SEG_ATTR(a) ( ((a) & 0xff) \| (((a) & 0x0f00) << 4) )
76	/** @} */
77
78	/** @name Macros for loading, storing segment registers to/from the VMCB.
79	* @{ */
80	#define HMSVM_LOAD_SEG_REG(REG, reg) \
81	do \
82	{ \
83	Assert(pCtx->reg.fFlags & CPUMSELREG_FLAGS_VALID); \
84	Assert(pCtx->reg.ValidSel == pCtx->reg.Sel); \
85	pVmcb->guest.REG.u16Sel = pCtx->reg.Sel; \
86	pVmcb->guest.REG.u32Limit = pCtx->reg.u32Limit; \
87	pVmcb->guest.REG.u64Base = pCtx->reg.u64Base; \
88	pVmcb->guest.REG.u16Attr = HMSVM_CPU_2_VMCB_SEG_ATTR(pCtx->reg.Attr.u); \
89	} while (0)
90
91	#define HMSVM_SAVE_SEG_REG(REG, reg) \
92	do \
93	{ \
94	pMixedCtx->reg.Sel = pVmcb->guest.REG.u16Sel; \
95	pMixedCtx->reg.ValidSel = pVmcb->guest.REG.u16Sel; \
96	pMixedCtx->reg.fFlags = CPUMSELREG_FLAGS_VALID; \
97	pMixedCtx->reg.u32Limit = pVmcb->guest.REG.u32Limit; \
98	pMixedCtx->reg.u64Base = pVmcb->guest.REG.u64Base; \
99	pMixedCtx->reg.Attr.u = HMSVM_VMCB_2_CPU_SEG_ATTR(pVmcb->guest.REG.u16Attr); \
100	} while (0)
101	/** @} */
102
103	/** Macro for checking and returning from the using function for
104	* \#VMEXIT intercepts that maybe caused during delivering of another
105	* event in the guest. */
106	#define HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY() \
107	do \
108	{ \
109	int rc = hmR0SvmCheckExitDueToEventDelivery(pVCpu, pCtx, pSvmTransient); \
110	if (RT_LIKELY(rc == VINF_SUCCESS)) { /* likely */ } \
111	else if (rc == VINF_HM_DOUBLE_FAULT) \
112	return VINF_SUCCESS; \
113	else \
114	return rc; \
115	} while (0)
116
117	/** Macro for upgrading a @a a_rc to VINF_EM_DBG_STEPPED after emulating an
118	* instruction that exited. */
119	#define HMSVM_CHECK_SINGLE_STEP(a_pVCpu, a_rc) \
120	do { \
121	if ((a_pVCpu)->hm.s.fSingleInstruction && (a_rc) == VINF_SUCCESS) \
122	(a_rc) = VINF_EM_DBG_STEPPED; \
123	} while (0)
124
125	/** Assert that preemption is disabled or covered by thread-context hooks. */
126	#define HMSVM_ASSERT_PREEMPT_SAFE() Assert( VMMR0ThreadCtxHookIsEnabled(pVCpu) \
127	\|\| !RTThreadPreemptIsEnabled(NIL_RTTHREAD));
128
129	/** Assert that we haven't migrated CPUs when thread-context hooks are not
130	* used. */
131	#define HMSVM_ASSERT_CPU_SAFE() AssertMsg( VMMR0ThreadCtxHookIsEnabled(pVCpu) \
132	\|\| pVCpu->hm.s.idEnteredCpu == RTMpCpuId(), \
133	("Illegal migration! Entered on CPU %u Current %u\n", \
134	pVCpu->hm.s.idEnteredCpu, RTMpCpuId()));
135
136	/** Exception bitmap mask for all contributory exceptions.
137	*
138	* Page fault is deliberately excluded here as it's conditional as to whether
139	* it's contributory or benign. Page faults are handled separately.
140	*/
141	#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) \
142	\| RT_BIT(X86_XCPT_DE))
143
144	/** @name VMCB Clean Bits.
145	*
146	* These flags are used for VMCB-state caching. A set VMCB Clean bit indicates
147	* AMD-V doesn't need to reload the corresponding value(s) from the VMCB in
148	* memory.
149	*
150	* @{ */
151	/** All intercepts vectors, TSC offset, PAUSE filter counter. */
152	#define HMSVM_VMCB_CLEAN_INTERCEPTS RT_BIT(0)
153	/** I/O permission bitmap, MSR permission bitmap. */
154	#define HMSVM_VMCB_CLEAN_IOPM_MSRPM RT_BIT(1)
155	/** ASID. */
156	#define HMSVM_VMCB_CLEAN_ASID RT_BIT(2)
157	/** TRP: V_TPR, V_IRQ, V_INTR_PRIO, V_IGN_TPR, V_INTR_MASKING,
158	V_INTR_VECTOR. */
159	#define HMSVM_VMCB_CLEAN_TPR RT_BIT(3)
160	/** Nested Paging: Nested CR3 (nCR3), PAT. */
161	#define HMSVM_VMCB_CLEAN_NP RT_BIT(4)
162	/** Control registers (CR0, CR3, CR4, EFER). */
163	#define HMSVM_VMCB_CLEAN_CRX_EFER RT_BIT(5)
164	/** Debug registers (DR6, DR7). */
165	#define HMSVM_VMCB_CLEAN_DRX RT_BIT(6)
166	/** GDT, IDT limit and base. */
167	#define HMSVM_VMCB_CLEAN_DT RT_BIT(7)
168	/** Segment register: CS, SS, DS, ES limit and base. */
169	#define HMSVM_VMCB_CLEAN_SEG RT_BIT(8)
170	/** CR2.*/
171	#define HMSVM_VMCB_CLEAN_CR2 RT_BIT(9)
172	/** Last-branch record (DbgCtlMsr, br_from, br_to, lastint_from, lastint_to) */
173	#define HMSVM_VMCB_CLEAN_LBR RT_BIT(10)
174	/** AVIC (AVIC APIC_BAR; AVIC APIC_BACKING_PAGE, AVIC
175	PHYSICAL_TABLE and AVIC LOGICAL_TABLE Pointers). */
176	#define HMSVM_VMCB_CLEAN_AVIC RT_BIT(11)
177	/** Mask of all valid VMCB Clean bits. */
178	#define HMSVM_VMCB_CLEAN_ALL ( HMSVM_VMCB_CLEAN_INTERCEPTS \
179	\| HMSVM_VMCB_CLEAN_IOPM_MSRPM \
180	\| HMSVM_VMCB_CLEAN_ASID \
181	\| HMSVM_VMCB_CLEAN_TPR \
182	\| HMSVM_VMCB_CLEAN_NP \
183	\| HMSVM_VMCB_CLEAN_CRX_EFER \
184	\| HMSVM_VMCB_CLEAN_DRX \
185	\| HMSVM_VMCB_CLEAN_DT \
186	\| HMSVM_VMCB_CLEAN_SEG \
187	\| HMSVM_VMCB_CLEAN_CR2 \
188	\| HMSVM_VMCB_CLEAN_LBR \
189	\| HMSVM_VMCB_CLEAN_AVIC)
190	/** @} */
191
192	/** @name SVM transient.
193	*
194	* A state structure for holding miscellaneous information across AMD-V
195	* VMRUN/\#VMEXIT operation, restored after the transition.
196	*
197	* @{ */
198	typedef struct SVMTRANSIENT
199	{
200	/** The host's rflags/eflags. */
201	RTCCUINTREG fEFlags;
202	#if HC_ARCH_BITS == 32
203	uint32_t u32Alignment0;
204	#endif
205
206	/** The \#VMEXIT exit code (the EXITCODE field in the VMCB). */
207	uint64_t u64ExitCode;
208	/** The guest's TPR value used for TPR shadowing. */
209	uint8_t u8GuestTpr;
210	/** Alignment. */
211	uint8_t abAlignment0[7];
212
213	/** Whether the guest FPU state was active at the time of \#VMEXIT. */
214	bool fWasGuestFPUStateActive;
215	/** Whether the guest debug state was active at the time of \#VMEXIT. */
216	bool fWasGuestDebugStateActive;
217	/** Whether the hyper debug state was active at the time of \#VMEXIT. */
218	bool fWasHyperDebugStateActive;
219	/** Whether the TSC offset mode needs to be updated. */
220	bool fUpdateTscOffsetting;
221	/** Whether the TSC_AUX MSR needs restoring on \#VMEXIT. */
222	bool fRestoreTscAuxMsr;
223	/** Whether the \#VMEXIT was caused by a page-fault during delivery of a
224	* contributary exception or a page-fault. */
225	bool fVectoringDoublePF;
226	/** Whether the \#VMEXIT was caused by a page-fault during delivery of an
227	* external interrupt or NMI. */
228	bool fVectoringPF;
229	} SVMTRANSIENT, *PSVMTRANSIENT;
230	AssertCompileMemberAlignment(SVMTRANSIENT, u64ExitCode, sizeof(uint64_t));
231	AssertCompileMemberAlignment(SVMTRANSIENT, fWasGuestFPUStateActive, sizeof(uint64_t));
232	/** @} */
233
234	/**
235	* MSRPM (MSR permission bitmap) read permissions (for guest RDMSR).
236	*/
237	typedef enum SVMMSREXITREAD
238	{
239	/** Reading this MSR causes a \#VMEXIT. */
240	SVMMSREXIT_INTERCEPT_READ = 0xb,
241	/** Reading this MSR does not cause a \#VMEXIT. */
242	SVMMSREXIT_PASSTHRU_READ
243	} SVMMSREXITREAD;
244
245	/**
246	* MSRPM (MSR permission bitmap) write permissions (for guest WRMSR).
247	*/
248	typedef enum SVMMSREXITWRITE
249	{
250	/** Writing to this MSR causes a \#VMEXIT. */
251	SVMMSREXIT_INTERCEPT_WRITE = 0xd,
252	/** Writing to this MSR does not cause a \#VMEXIT. */
253	SVMMSREXIT_PASSTHRU_WRITE
254	} SVMMSREXITWRITE;
255
256	/**
257	* SVM \#VMEXIT handler.
258	*
259	* @returns VBox status code.
260	* @param pVCpu The cross context virtual CPU structure.
261	* @param pMixedCtx Pointer to the guest-CPU context.
262	* @param pSvmTransient Pointer to the SVM-transient structure.
263	*/
264	typedef int FNSVMEXITHANDLER(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
265
266
267	/*********************************************************************************************************************************
268	* Internal Functions *
269	*********************************************************************************************************************************/
270	static void hmR0SvmSetMsrPermission(PVMCPU pVCpu, unsigned uMsr, SVMMSREXITREAD enmRead, SVMMSREXITWRITE enmWrite);
271	static void hmR0SvmPendingEventToTrpmTrap(PVMCPU pVCpu);
272	static void hmR0SvmLeave(PVMCPU pVCpu);
273
274	/** @name \#VMEXIT handlers.
275	* @{
276	*/
277	static FNSVMEXITHANDLER hmR0SvmExitIntr;
278	static FNSVMEXITHANDLER hmR0SvmExitWbinvd;
279	static FNSVMEXITHANDLER hmR0SvmExitInvd;
280	static FNSVMEXITHANDLER hmR0SvmExitCpuid;
281	static FNSVMEXITHANDLER hmR0SvmExitRdtsc;
282	static FNSVMEXITHANDLER hmR0SvmExitRdtscp;
283	static FNSVMEXITHANDLER hmR0SvmExitRdpmc;
284	static FNSVMEXITHANDLER hmR0SvmExitInvlpg;
285	static FNSVMEXITHANDLER hmR0SvmExitHlt;
286	static FNSVMEXITHANDLER hmR0SvmExitMonitor;
287	static FNSVMEXITHANDLER hmR0SvmExitMwait;
288	static FNSVMEXITHANDLER hmR0SvmExitShutdown;
289	static FNSVMEXITHANDLER hmR0SvmExitReadCRx;
290	static FNSVMEXITHANDLER hmR0SvmExitWriteCRx;
291	static FNSVMEXITHANDLER hmR0SvmExitSetPendingXcptUD;
292	static FNSVMEXITHANDLER hmR0SvmExitMsr;
293	static FNSVMEXITHANDLER hmR0SvmExitReadDRx;
294	static FNSVMEXITHANDLER hmR0SvmExitWriteDRx;
295	static FNSVMEXITHANDLER hmR0SvmExitXsetbv;
296	static FNSVMEXITHANDLER hmR0SvmExitIOInstr;
297	static FNSVMEXITHANDLER hmR0SvmExitNestedPF;
298	static FNSVMEXITHANDLER hmR0SvmExitVIntr;
299	static FNSVMEXITHANDLER hmR0SvmExitTaskSwitch;
300	static FNSVMEXITHANDLER hmR0SvmExitVmmCall;
301	static FNSVMEXITHANDLER hmR0SvmExitPause;
302	static FNSVMEXITHANDLER hmR0SvmExitIret;
303	static FNSVMEXITHANDLER hmR0SvmExitXcptPF;
304	static FNSVMEXITHANDLER hmR0SvmExitXcptNM;
305	static FNSVMEXITHANDLER hmR0SvmExitXcptUD;
306	static FNSVMEXITHANDLER hmR0SvmExitXcptMF;
307	static FNSVMEXITHANDLER hmR0SvmExitXcptDB;
308	static FNSVMEXITHANDLER hmR0SvmExitXcptAC;
309	static FNSVMEXITHANDLER hmR0SvmExitXcptBP;
310	#ifdef VBOX_WITH_NESTED_HWVIRT
311	static FNSVMEXITHANDLER hmR0SvmExitClgi;
312	static FNSVMEXITHANDLER hmR0SvmExitStgi;
313	#endif
314	/** @} */
315
316	DECLINLINE(int) hmR0SvmHandleExit(PVMCPU pVCpu, PCPUMCTX pMixedCtx, PSVMTRANSIENT pSvmTransient);
317
318
319	/*********************************************************************************************************************************
320	* Global Variables *
321	*********************************************************************************************************************************/
322	/** Ring-0 memory object for the IO bitmap. */
323	RTR0MEMOBJ g_hMemObjIOBitmap = NIL_RTR0MEMOBJ;
324	/** Physical address of the IO bitmap. */
325	RTHCPHYS g_HCPhysIOBitmap = 0;
326	/** Virtual address of the IO bitmap. */
327	R0PTRTYPE(void *) g_pvIOBitmap = NULL;
328
329
330	/**
331	* Sets up and activates AMD-V on the current CPU.
332	*
333	* @returns VBox status code.
334	* @param pCpu Pointer to the CPU info struct.
335	* @param pVM The cross context VM structure. Can be
336	* NULL after a resume!
337	* @param pvCpuPage Pointer to the global CPU page.
338	* @param HCPhysCpuPage Physical address of the global CPU page.
339	* @param fEnabledByHost Whether the host OS has already initialized AMD-V.
340	* @param pvArg Unused on AMD-V.
341	*/
342	VMMR0DECL(int) SVMR0EnableCpu(PHMGLOBALCPUINFO pCpu, PVM pVM, void *pvCpuPage, RTHCPHYS HCPhysCpuPage, bool fEnabledByHost,
343	void *pvArg)
344	{
345	Assert(!fEnabledByHost);
346	Assert(HCPhysCpuPage && HCPhysCpuPage != NIL_RTHCPHYS);
347	Assert(RT_ALIGN_T(HCPhysCpuPage, _4K, RTHCPHYS) == HCPhysCpuPage);
348	Assert(pvCpuPage); NOREF(pvCpuPage);
349	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
350
351	NOREF(pvArg);
352	NOREF(fEnabledByHost);
353
354	/* Paranoid: Disable interrupt as, in theory, interrupt handlers might mess with EFER. */
355	RTCCUINTREG fEFlags = ASMIntDisableFlags();
356
357	/*
358	* We must turn on AMD-V and setup the host state physical address, as those MSRs are per CPU.
359	*/
360	uint64_t u64HostEfer = ASMRdMsr(MSR_K6_EFER);
361	if (u64HostEfer & MSR_K6_EFER_SVME)
362	{
363	/* If the VBOX_HWVIRTEX_IGNORE_SVM_IN_USE is active, then we blindly use AMD-V. */
364	if ( pVM
365	&& pVM->hm.s.svm.fIgnoreInUseError)
366	{
367	pCpu->fIgnoreAMDVInUseError = true;
368	}
369
370	if (!pCpu->fIgnoreAMDVInUseError)
371	{
372	ASMSetFlags(fEFlags);
373	return VERR_SVM_IN_USE;
374	}
375	}
376
377	/* Turn on AMD-V in the EFER MSR. */
378	ASMWrMsr(MSR_K6_EFER, u64HostEfer \| MSR_K6_EFER_SVME);
379
380	/* Write the physical page address where the CPU will store the host state while executing the VM. */
381	ASMWrMsr(MSR_K8_VM_HSAVE_PA, HCPhysCpuPage);
382
383	/* Restore interrupts. */
384	ASMSetFlags(fEFlags);
385
386	/*
387	* Theoretically, other hypervisors may have used ASIDs, ideally we should flush all non-zero ASIDs
388	* when enabling SVM. AMD doesn't have an SVM instruction to flush all ASIDs (flushing is done
389	* upon VMRUN). Therefore, just set the fFlushAsidBeforeUse flag which instructs hmR0SvmSetupTLB()
390	* to flush the TLB with before using a new ASID.
391	*/
392	pCpu->fFlushAsidBeforeUse = true;
393
394	/*
395	* Ensure each VCPU scheduled on this CPU gets a new VPID on resume. See @bugref{6255}.
396	*/
397	++pCpu->cTlbFlushes;
398
399	return VINF_SUCCESS;
400	}
401
402
403	/**
404	* Deactivates AMD-V on the current CPU.
405	*
406	* @returns VBox status code.
407	* @param pCpu Pointer to the CPU info struct.
408	* @param pvCpuPage Pointer to the global CPU page.
409	* @param HCPhysCpuPage Physical address of the global CPU page.
410	*/
411	VMMR0DECL(int) SVMR0DisableCpu(PHMGLOBALCPUINFO pCpu, void *pvCpuPage, RTHCPHYS HCPhysCpuPage)
412	{
413	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
414	AssertReturn( HCPhysCpuPage
415	&& HCPhysCpuPage != NIL_RTHCPHYS, VERR_INVALID_PARAMETER);
416	AssertReturn(pvCpuPage, VERR_INVALID_PARAMETER);
417	NOREF(pCpu);
418
419	/* Paranoid: Disable interrupts as, in theory, interrupt handlers might mess with EFER. */
420	RTCCUINTREG fEFlags = ASMIntDisableFlags();
421
422	/* Turn off AMD-V in the EFER MSR. */
423	uint64_t u64HostEfer = ASMRdMsr(MSR_K6_EFER);
424	ASMWrMsr(MSR_K6_EFER, u64HostEfer & ~MSR_K6_EFER_SVME);
425
426	/* Invalidate host state physical address. */
427	ASMWrMsr(MSR_K8_VM_HSAVE_PA, 0);
428
429	/* Restore interrupts. */
430	ASMSetFlags(fEFlags);
431
432	return VINF_SUCCESS;
433	}
434
435
436	/**
437	* Does global AMD-V initialization (called during module initialization).
438	*
439	* @returns VBox status code.
440	*/
441	VMMR0DECL(int) SVMR0GlobalInit(void)
442	{
443	/*
444	* Allocate 12 KB for the IO bitmap. Since this is non-optional and we always intercept all IO accesses, it's done
445	* once globally here instead of per-VM.
446	*/
447	Assert(g_hMemObjIOBitmap == NIL_RTR0MEMOBJ);
448	int rc = RTR0MemObjAllocCont(&g_hMemObjIOBitmap, 3 << PAGE_SHIFT, false /* fExecutable */);
449	if (RT_FAILURE(rc))
450	return rc;
451
452	g_pvIOBitmap = RTR0MemObjAddress(g_hMemObjIOBitmap);
453	g_HCPhysIOBitmap = RTR0MemObjGetPagePhysAddr(g_hMemObjIOBitmap, 0 /* iPage */);
454
455	/* Set all bits to intercept all IO accesses. */
456	ASMMemFill32(g_pvIOBitmap, 3 << PAGE_SHIFT, UINT32_C(0xffffffff));
457	return VINF_SUCCESS;
458	}
459
460
461	/**
462	* Does global AMD-V termination (called during module termination).
463	*/
464	VMMR0DECL(void) SVMR0GlobalTerm(void)
465	{
466	if (g_hMemObjIOBitmap != NIL_RTR0MEMOBJ)
467	{
468	RTR0MemObjFree(g_hMemObjIOBitmap, true /* fFreeMappings */);
469	g_pvIOBitmap = NULL;
470	g_HCPhysIOBitmap = 0;
471	g_hMemObjIOBitmap = NIL_RTR0MEMOBJ;
472	}
473	}
474
475
476	/**
477	* Frees any allocated per-VCPU structures for a VM.
478	*
479	* @param pVM The cross context VM structure.
480	*/
481	DECLINLINE(void) hmR0SvmFreeStructs(PVM pVM)
482	{
483	for (uint32_t i = 0; i < pVM->cCpus; i++)
484	{
485	PVMCPU pVCpu = &pVM->aCpus[i];
486	AssertPtr(pVCpu);
487
488	if (pVCpu->hm.s.svm.hMemObjVmcbHost != NIL_RTR0MEMOBJ)
489	{
490	RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjVmcbHost, false);
491	pVCpu->hm.s.svm.pvVmcbHost = 0;
492	pVCpu->hm.s.svm.HCPhysVmcbHost = 0;
493	pVCpu->hm.s.svm.hMemObjVmcbHost = NIL_RTR0MEMOBJ;
494	}
495
496	if (pVCpu->hm.s.svm.hMemObjVmcb != NIL_RTR0MEMOBJ)
497	{
498	RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjVmcb, false);
499	pVCpu->hm.s.svm.pvVmcb = 0;
500	pVCpu->hm.s.svm.HCPhysVmcb = 0;
501	pVCpu->hm.s.svm.hMemObjVmcb = NIL_RTR0MEMOBJ;
502	}
503
504	if (pVCpu->hm.s.svm.hMemObjMsrBitmap != NIL_RTR0MEMOBJ)
505	{
506	RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjMsrBitmap, false);
507	pVCpu->hm.s.svm.pvMsrBitmap = 0;
508	pVCpu->hm.s.svm.HCPhysMsrBitmap = 0;
509	pVCpu->hm.s.svm.hMemObjMsrBitmap = NIL_RTR0MEMOBJ;
510	}
511	}
512	}
513
514
515	/**
516	* Does per-VM AMD-V initialization.
517	*
518	* @returns VBox status code.
519	* @param pVM The cross context VM structure.
520	*/
521	VMMR0DECL(int) SVMR0InitVM(PVM pVM)
522	{
523	int rc = VERR_INTERNAL_ERROR_5;
524
525	/*
526	* Check for an AMD CPU erratum which requires us to flush the TLB before every world-switch.
527	*/
528	uint32_t u32Family;
529	uint32_t u32Model;
530	uint32_t u32Stepping;
531	if (HMAmdIsSubjectToErratum170(&u32Family, &u32Model, &u32Stepping))
532	{
533	Log4(("SVMR0InitVM: AMD cpu with erratum 170 family %#x model %#x stepping %#x\n", u32Family, u32Model, u32Stepping));
534	pVM->hm.s.svm.fAlwaysFlushTLB = true;
535	}
536
537	/*
538	* Initialize the R0 memory objects up-front so we can properly cleanup on allocation failures.
539	*/
540	for (VMCPUID i = 0; i < pVM->cCpus; i++)
541	{
542	PVMCPU pVCpu = &pVM->aCpus[i];
543	pVCpu->hm.s.svm.hMemObjVmcbHost = NIL_RTR0MEMOBJ;
544	pVCpu->hm.s.svm.hMemObjVmcb = NIL_RTR0MEMOBJ;
545	pVCpu->hm.s.svm.hMemObjMsrBitmap = NIL_RTR0MEMOBJ;
546	}
547
548	for (VMCPUID i = 0; i < pVM->cCpus; i++)
549	{
550	PVMCPU pVCpu = &pVM->aCpus[i];
551
552	/*
553	* Allocate one page for the host-context VM control block (VMCB). This is used for additional host-state (such as
554	* FS, GS, Kernel GS Base, etc.) apart from the host-state save area specified in MSR_K8_VM_HSAVE_PA.
555	*/
556	rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjVmcbHost, 1 << PAGE_SHIFT, false /* fExecutable */);
557	if (RT_FAILURE(rc))
558	goto failure_cleanup;
559
560	pVCpu->hm.s.svm.pvVmcbHost = RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjVmcbHost);
561	pVCpu->hm.s.svm.HCPhysVmcbHost = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjVmcbHost, 0 /* iPage */);
562	Assert(pVCpu->hm.s.svm.HCPhysVmcbHost < _4G);
563	ASMMemZeroPage(pVCpu->hm.s.svm.pvVmcbHost);
564
565	/*
566	* Allocate one page for the guest-state VMCB.
567	*/
568	rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjVmcb, 1 << PAGE_SHIFT, false /* fExecutable */);
569	if (RT_FAILURE(rc))
570	goto failure_cleanup;
571
572	pVCpu->hm.s.svm.pvVmcb = RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjVmcb);
573	pVCpu->hm.s.svm.HCPhysVmcb = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjVmcb, 0 /* iPage */);
574	Assert(pVCpu->hm.s.svm.HCPhysVmcb < _4G);
575	ASMMemZeroPage(pVCpu->hm.s.svm.pvVmcb);
576
577	/*
578	* Allocate two pages (8 KB) for the MSR permission bitmap. There doesn't seem to be a way to convince
579	* SVM to not require one.
580	*/
581	rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjMsrBitmap, 2 << PAGE_SHIFT, false /* fExecutable */);
582	if (RT_FAILURE(rc))
583	goto failure_cleanup;
584
585	pVCpu->hm.s.svm.pvMsrBitmap = RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjMsrBitmap);
586	pVCpu->hm.s.svm.HCPhysMsrBitmap = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjMsrBitmap, 0 /* iPage */);
587	/* Set all bits to intercept all MSR accesses (changed later on). */
588	ASMMemFill32(pVCpu->hm.s.svm.pvMsrBitmap, 2 << PAGE_SHIFT, UINT32_C(0xffffffff));
589	}
590
591	return VINF_SUCCESS;
592
593	failure_cleanup:
594	hmR0SvmFreeStructs(pVM);
595	return rc;
596	}
597
598
599	/**
600	* Does per-VM AMD-V termination.
601	*
602	* @returns VBox status code.
603	* @param pVM The cross context VM structure.
604	*/
605	VMMR0DECL(int) SVMR0TermVM(PVM pVM)
606	{
607	hmR0SvmFreeStructs(pVM);
608	return VINF_SUCCESS;
609	}
610
611
612	/**
613	* Sets the permission bits for the specified MSR in the MSRPM.
614	*
615	* @param pVCpu The cross context virtual CPU structure.
616	* @param uMsr The MSR for which the access permissions are being set.
617	* @param enmRead MSR read permissions.
618	* @param enmWrite MSR write permissions.
619	*/
620	static void hmR0SvmSetMsrPermission(PVMCPU pVCpu, unsigned uMsr, SVMMSREXITREAD enmRead, SVMMSREXITWRITE enmWrite)
621	{
622	unsigned uBit;
623	uint8_t pbMsrBitmap = (uint8_t )pVCpu->hm.s.svm.pvMsrBitmap;
624
625	/*
626	* Layout:
627	* Byte offset MSR range
628	* 0x000 - 0x7ff 0x00000000 - 0x00001fff
629	* 0x800 - 0xfff 0xc0000000 - 0xc0001fff
630	* 0x1000 - 0x17ff 0xc0010000 - 0xc0011fff
631	* 0x1800 - 0x1fff Reserved
632	*/
633	if (uMsr <= 0x00001FFF)
634	{
635	/* Pentium-compatible MSRs. */
636	uBit = uMsr * 2;
637	}
638	else if ( uMsr >= 0xC0000000
639	&& uMsr <= 0xC0001FFF)
640	{
641	/* AMD Sixth Generation x86 Processor MSRs. */
642	uBit = (uMsr - 0xC0000000) * 2;
643	pbMsrBitmap += 0x800;
644	}
645	else if ( uMsr >= 0xC0010000
646	&& uMsr <= 0xC0011FFF)
647	{
648	/* AMD Seventh and Eighth Generation Processor MSRs. */
649	uBit = (uMsr - 0xC0001000) * 2;
650	pbMsrBitmap += 0x1000;
651	}
652	else
653	{
654	AssertFailed();
655	return;
656	}
657
658	Assert(uBit < 0x3fff /* 16 * 1024 - 1 */);
659	if (enmRead == SVMMSREXIT_INTERCEPT_READ)
660	ASMBitSet(pbMsrBitmap, uBit);
661	else
662	ASMBitClear(pbMsrBitmap, uBit);
663
664	if (enmWrite == SVMMSREXIT_INTERCEPT_WRITE)
665	ASMBitSet(pbMsrBitmap, uBit + 1);
666	else
667	ASMBitClear(pbMsrBitmap, uBit + 1);
668
669	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
670	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_IOPM_MSRPM;
671	}
672
673
674	/**
675	* Sets up AMD-V for the specified VM.
676	* This function is only called once per-VM during initalization.
677	*
678	* @returns VBox status code.
679	* @param pVM The cross context VM structure.
680	*/
681	VMMR0DECL(int) SVMR0SetupVM(PVM pVM)
682	{
683	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
684	AssertReturn(pVM, VERR_INVALID_PARAMETER);
685	Assert(pVM->hm.s.svm.fSupported);
686
687	bool const fPauseFilter = RT_BOOL(pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_PAUSE_FILTER);
688	bool const fPauseFilterThreshold = RT_BOOL(pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_PAUSE_FILTER_THRESHOLD);
689	bool const fUsePauseFilter = fPauseFilter && pVM->hm.s.svm.cPauseFilter && pVM->hm.s.svm.cPauseFilterThresholdTicks;
690
691	for (VMCPUID i = 0; i < pVM->cCpus; i++)
692	{
693	PVMCPU pVCpu = &pVM->aCpus[i];
694	PSVMVMCB pVmcb = (PSVMVMCB)pVM->aCpus[i].hm.s.svm.pvVmcb;
695
696	AssertMsgReturn(pVmcb, ("Invalid pVmcb for vcpu[%u]\n", i), VERR_SVM_INVALID_PVMCB);
697
698	/* Initialize the #VMEXIT history array with end-of-array markers (UINT16_MAX). */
699	Assert(!pVCpu->hm.s.idxExitHistoryFree);
700	HMCPU_EXIT_HISTORY_RESET(pVCpu);
701
702	/* Always trap #AC for reasons of security. */
703	pVmcb->ctrl.u32InterceptException \|= RT_BIT_32(X86_XCPT_AC);
704
705	/* Always trap #DB for reasons of security. */
706	pVmcb->ctrl.u32InterceptException \|= RT_BIT_32(X86_XCPT_DB);
707
708	/* Trap exceptions unconditionally (debug purposes). */
709	#ifdef HMSVM_ALWAYS_TRAP_PF
710	pVmcb->ctrl.u32InterceptException \|= RT_BIT(X86_XCPT_PF);
711	#endif
712	#ifdef HMSVM_ALWAYS_TRAP_ALL_XCPTS
713	/* If you add any exceptions here, make sure to update hmR0SvmHandleExit(). */
714	pVmcb->ctrl.u32InterceptException \|= 0
715	\| RT_BIT(X86_XCPT_BP)
716	\| RT_BIT(X86_XCPT_DE)
717	\| RT_BIT(X86_XCPT_NM)
718	\| RT_BIT(X86_XCPT_UD)
719	\| RT_BIT(X86_XCPT_NP)
720	\| RT_BIT(X86_XCPT_SS)
721	\| RT_BIT(X86_XCPT_GP)
722	\| RT_BIT(X86_XCPT_PF)
723	\| RT_BIT(X86_XCPT_MF)
724	;
725	#endif
726
727	/* Set up unconditional intercepts and conditions. */
728	pVmcb->ctrl.u64InterceptCtrl = SVM_CTRL_INTERCEPT_INTR /* External interrupt causes a #VMEXIT. */
729	\| SVM_CTRL_INTERCEPT_NMI /* Non-maskable interrupts causes a #VMEXIT. */
730	\| SVM_CTRL_INTERCEPT_INIT /* INIT signal causes a #VMEXIT. */
731	\| SVM_CTRL_INTERCEPT_RDPMC /* RDPMC causes a #VMEXIT. */
732	\| SVM_CTRL_INTERCEPT_CPUID /* CPUID causes a #VMEXIT. */
733	\| SVM_CTRL_INTERCEPT_RSM /* RSM causes a #VMEXIT. */
734	\| SVM_CTRL_INTERCEPT_HLT /* HLT causes a #VMEXIT. */
735	\| SVM_CTRL_INTERCEPT_INOUT_BITMAP /* Use the IOPM to cause IOIO #VMEXITs. */
736	\| SVM_CTRL_INTERCEPT_MSR_SHADOW /* MSR access not covered by MSRPM causes a #VMEXIT.*/
737	\| SVM_CTRL_INTERCEPT_INVLPGA /* INVLPGA causes a #VMEXIT. */
738	\| SVM_CTRL_INTERCEPT_SHUTDOWN /* Shutdown events causes a #VMEXIT. */
739	\| SVM_CTRL_INTERCEPT_FERR_FREEZE /* Intercept "freezing" during legacy FPU handling. */
740	\| SVM_CTRL_INTERCEPT_VMRUN /* VMRUN causes a #VMEXIT. */
741	\| SVM_CTRL_INTERCEPT_VMMCALL /* VMMCALL causes a #VMEXIT. */
742	\| SVM_CTRL_INTERCEPT_VMLOAD /* VMLOAD causes a #VMEXIT. */
743	\| SVM_CTRL_INTERCEPT_VMSAVE /* VMSAVE causes a #VMEXIT. */
744	\| SVM_CTRL_INTERCEPT_STGI /* STGI causes a #VMEXIT. */
745	\| SVM_CTRL_INTERCEPT_CLGI /* CLGI causes a #VMEXIT. */
746	\| SVM_CTRL_INTERCEPT_SKINIT /* SKINIT causes a #VMEXIT. */
747	\| SVM_CTRL_INTERCEPT_WBINVD /* WBINVD causes a #VMEXIT. */
748	\| SVM_CTRL_INTERCEPT_MONITOR /* MONITOR causes a #VMEXIT. */
749	\| SVM_CTRL_INTERCEPT_MWAIT /* MWAIT causes a #VMEXIT. */
750	\| SVM_CTRL_INTERCEPT_XSETBV; /* XSETBV causes a #VMEXIT. */
751
752	/* CR0, CR4 reads must be intercepted, our shadow values are not necessarily the same as the guest's. */
753	pVmcb->ctrl.u16InterceptRdCRx = RT_BIT(0) \| RT_BIT(4);
754
755	/* CR0, CR4 writes must be intercepted for the same reasons as above. */
756	pVmcb->ctrl.u16InterceptWrCRx = RT_BIT(0) \| RT_BIT(4);
757
758	/* Intercept all DRx reads and writes by default. Changed later on. */
759	pVmcb->ctrl.u16InterceptRdDRx = 0xffff;
760	pVmcb->ctrl.u16InterceptWrDRx = 0xffff;
761
762	/* Virtualize masking of INTR interrupts. (reads/writes from/to CR8 go to the V_TPR register) */
763	pVmcb->ctrl.IntCtrl.n.u1VIrqMasking = 1;
764
765	/* Ignore the priority in the TPR. This is necessary for delivering PIC style (ExtInt) interrupts and we currently
766	deliver both PIC and APIC interrupts alike. See hmR0SvmInjectPendingEvent() */
767	pVmcb->ctrl.IntCtrl.n.u1IgnoreTPR = 1;
768
769	/* Set IO and MSR bitmap permission bitmap physical addresses. */
770	pVmcb->ctrl.u64IOPMPhysAddr = g_HCPhysIOBitmap;
771	pVmcb->ctrl.u64MSRPMPhysAddr = pVCpu->hm.s.svm.HCPhysMsrBitmap;
772
773	/* No LBR virtualization. */
774	pVmcb->ctrl.u64LBRVirt = 0;
775
776	/* Initially set all VMCB clean bits to 0 indicating that everything should be loaded from the VMCB in memory. */
777	pVmcb->ctrl.u64VmcbCleanBits = 0;
778
779	/* The host ASID MBZ, for the guest start with 1. */
780	pVmcb->ctrl.TLBCtrl.n.u32ASID = 1;
781
782	/*
783	* Setup the PAT MSR (applicable for Nested Paging only).
784	* The default value should be 0x0007040600070406ULL, but we want to treat all guest memory as WB,
785	* so choose type 6 for all PAT slots.
786	*/
787	pVmcb->guest.u64GPAT = UINT64_C(0x0006060606060606);
788
789	/* Setup Nested Paging. This doesn't change throughout the execution time of the VM. */
790	pVmcb->ctrl.NestedPaging.n.u1NestedPaging = pVM->hm.s.fNestedPaging;
791
792	/* Without Nested Paging, we need additionally intercepts. */
793	if (!pVM->hm.s.fNestedPaging)
794	{
795	/* CR3 reads/writes must be intercepted; our shadow values differ from the guest values. */
796	pVmcb->ctrl.u16InterceptRdCRx \|= RT_BIT(3);
797	pVmcb->ctrl.u16InterceptWrCRx \|= RT_BIT(3);
798
799	/* Intercept INVLPG and task switches (may change CR3, EFLAGS, LDT). */
800	pVmcb->ctrl.u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_INVLPG
801	\| SVM_CTRL_INTERCEPT_TASK_SWITCH;
802
803	/* Page faults must be intercepted to implement shadow paging. */
804	pVmcb->ctrl.u32InterceptException \|= RT_BIT(X86_XCPT_PF);
805	}
806
807	#ifdef HMSVM_ALWAYS_TRAP_TASK_SWITCH
808	pVmcb->ctrl.u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_TASK_SWITCH;
809	#endif
810
811	/* Apply the exceptions intercepts needed by the GIM provider. */
812	if (pVCpu->hm.s.fGIMTrapXcptUD)
813	pVmcb->ctrl.u32InterceptException \|= RT_BIT(X86_XCPT_UD);
814
815	/* Setup Pause Filter for guest pause-loop (spinlock) exiting. */
816	if (fUsePauseFilter)
817	{
818	pVmcb->ctrl.u16PauseFilterCount = pVM->hm.s.svm.cPauseFilter;
819	if (fPauseFilterThreshold)
820	pVmcb->ctrl.u16PauseFilterThreshold = pVM->hm.s.svm.cPauseFilterThresholdTicks;
821	}
822
823	/*
824	* The following MSRs are saved/restored automatically during the world-switch.
825	* Don't intercept guest read/write accesses to these MSRs.
826	*/
827	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
828	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_CSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
829	hmR0SvmSetMsrPermission(pVCpu, MSR_K6_STAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
830	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_SF_MASK, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
831	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_FS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
832	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_GS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
833	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_KERNEL_GS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
834	hmR0SvmSetMsrPermission(pVCpu, MSR_IA32_SYSENTER_CS, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
835	hmR0SvmSetMsrPermission(pVCpu, MSR_IA32_SYSENTER_ESP, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
836	hmR0SvmSetMsrPermission(pVCpu, MSR_IA32_SYSENTER_EIP, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
837	}
838
839	return VINF_SUCCESS;
840	}
841
842
843	/**
844	* Invalidates a guest page by guest virtual address.
845	*
846	* @returns VBox status code.
847	* @param pVM The cross context VM structure.
848	* @param pVCpu The cross context virtual CPU structure.
849	* @param GCVirt Guest virtual address of the page to invalidate.
850	*/
851	VMMR0DECL(int) SVMR0InvalidatePage(PVM pVM, PVMCPU pVCpu, RTGCPTR GCVirt)
852	{
853	AssertReturn(pVM, VERR_INVALID_PARAMETER);
854	Assert(pVM->hm.s.svm.fSupported);
855
856	bool fFlushPending = pVM->hm.s.svm.fAlwaysFlushTLB \|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_TLB_FLUSH);
857
858	/* Skip it if a TLB flush is already pending. */
859	if (!fFlushPending)
860	{
861	Log4(("SVMR0InvalidatePage %RGv\n", GCVirt));
862
863	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
864	AssertMsgReturn(pVmcb, ("Invalid pVmcb!\n"), VERR_SVM_INVALID_PVMCB);
865
866	#if HC_ARCH_BITS == 32
867	/* If we get a flush in 64-bit guest mode, then force a full TLB flush. INVLPGA takes only 32-bit addresses. */
868	if (CPUMIsGuestInLongMode(pVCpu))
869	VMCPU_FF_SET(pVCpu, VMCPU_FF_TLB_FLUSH);
870	else
871	#endif
872	{
873	SVMR0InvlpgA(GCVirt, pVmcb->ctrl.TLBCtrl.n.u32ASID);
874	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlbInvlpgVirt);
875	}
876	}
877	return VINF_SUCCESS;
878	}
879
880
881	/**
882	* Flushes the appropriate tagged-TLB entries.
883	*
884	* @param pVCpu The cross context virtual CPU structure.
885	*/
886	static void hmR0SvmFlushTaggedTlb(PVMCPU pVCpu)
887	{
888	PVM pVM = pVCpu->CTX_SUFF(pVM);
889	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
890	PHMGLOBALCPUINFO pCpu = hmR0GetCurrentCpu();
891
892	/*
893	* Force a TLB flush for the first world switch if the current CPU differs from the one we ran on last.
894	* This can happen both for start & resume due to long jumps back to ring-3.
895	* If the TLB flush count changed, another VM (VCPU rather) has hit the ASID limit while flushing the TLB,
896	* so we cannot reuse the ASIDs without flushing.
897	*/
898	bool fNewAsid = false;
899	Assert(pCpu->idCpu != NIL_RTCPUID);
900	if ( pVCpu->hm.s.idLastCpu != pCpu->idCpu
901	\|\| pVCpu->hm.s.cTlbFlushes != pCpu->cTlbFlushes)
902	{
903	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlbWorldSwitch);
904	pVCpu->hm.s.fForceTLBFlush = true;
905	fNewAsid = true;
906	}
907
908	/* Set TLB flush state as checked until we return from the world switch. */
909	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, true);
910
911	/* Check for explicit TLB flushes. */
912	if (VMCPU_FF_TEST_AND_CLEAR(pVCpu, VMCPU_FF_TLB_FLUSH))
913	{
914	pVCpu->hm.s.fForceTLBFlush = true;
915	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlb);
916	}
917
918	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_NOTHING;
919
920	if (pVM->hm.s.svm.fAlwaysFlushTLB)
921	{
922	/*
923	* This is the AMD erratum 170. We need to flush the entire TLB for each world switch. Sad.
924	*/
925	pCpu->uCurrentAsid = 1;
926	pVCpu->hm.s.uCurrentAsid = 1;
927	pVCpu->hm.s.cTlbFlushes = pCpu->cTlbFlushes;
928	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
929
930	/* Clear the VMCB Clean Bit for NP while flushing the TLB. See @bugref{7152}. */
931	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
932
933	/* Keep track of last CPU ID even when flushing all the time. */
934	if (fNewAsid)
935	pVCpu->hm.s.idLastCpu = pCpu->idCpu;
936	}
937	else if (pVCpu->hm.s.fForceTLBFlush)
938	{
939	/* Clear the VMCB Clean Bit for NP while flushing the TLB. See @bugref{7152}. */
940	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
941
942	if (fNewAsid)
943	{
944	++pCpu->uCurrentAsid;
945	bool fHitASIDLimit = false;
946	if (pCpu->uCurrentAsid >= pVM->hm.s.uMaxAsid)
947	{
948	pCpu->uCurrentAsid = 1; /* Wraparound at 1; host uses 0 */
949	pCpu->cTlbFlushes++; /* All VCPUs that run on this host CPU must use a new VPID. */
950	fHitASIDLimit = true;
951
952	if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_FLUSH_BY_ASID)
953	{
954	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_SINGLE_CONTEXT;
955	pCpu->fFlushAsidBeforeUse = true;
956	}
957	else
958	{
959	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
960	pCpu->fFlushAsidBeforeUse = false;
961	}
962	}
963
964	if ( !fHitASIDLimit
965	&& pCpu->fFlushAsidBeforeUse)
966	{
967	if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_FLUSH_BY_ASID)
968	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_SINGLE_CONTEXT;
969	else
970	{
971	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
972	pCpu->fFlushAsidBeforeUse = false;
973	}
974	}
975
976	pVCpu->hm.s.uCurrentAsid = pCpu->uCurrentAsid;
977	pVCpu->hm.s.idLastCpu = pCpu->idCpu;
978	pVCpu->hm.s.cTlbFlushes = pCpu->cTlbFlushes;
979	}
980	else
981	{
982	if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_FLUSH_BY_ASID)
983	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_SINGLE_CONTEXT;
984	else
985	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
986	}
987
988	pVCpu->hm.s.fForceTLBFlush = false;
989	}
990
991	/* Update VMCB with the ASID. */
992	if (pVmcb->ctrl.TLBCtrl.n.u32ASID != pVCpu->hm.s.uCurrentAsid)
993	{
994	pVmcb->ctrl.TLBCtrl.n.u32ASID = pVCpu->hm.s.uCurrentAsid;
995	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_ASID;
996	}
997
998	AssertMsg(pVCpu->hm.s.idLastCpu == pCpu->idCpu,
999	("vcpu idLastCpu=%u pcpu idCpu=%u\n", pVCpu->hm.s.idLastCpu, pCpu->idCpu));
1000	AssertMsg(pVCpu->hm.s.cTlbFlushes == pCpu->cTlbFlushes,
1001	("Flush count mismatch for cpu %u (%u vs %u)\n", pCpu->idCpu, pVCpu->hm.s.cTlbFlushes, pCpu->cTlbFlushes));
1002	AssertMsg(pCpu->uCurrentAsid >= 1 && pCpu->uCurrentAsid < pVM->hm.s.uMaxAsid,
1003	("cpu%d uCurrentAsid = %x\n", pCpu->idCpu, pCpu->uCurrentAsid));
1004	AssertMsg(pVCpu->hm.s.uCurrentAsid >= 1 && pVCpu->hm.s.uCurrentAsid < pVM->hm.s.uMaxAsid,
1005	("cpu%d VM uCurrentAsid = %x\n", pCpu->idCpu, pVCpu->hm.s.uCurrentAsid));
1006
1007	#ifdef VBOX_WITH_STATISTICS
1008	if (pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_NOTHING)
1009	STAM_COUNTER_INC(&pVCpu->hm.s.StatNoFlushTlbWorldSwitch);
1010	else if ( pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_SINGLE_CONTEXT
1011	\|\| pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_SINGLE_CONTEXT_RETAIN_GLOBALS)
1012	{
1013	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushAsid);
1014	}
1015	else
1016	{
1017	Assert(pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_ENTIRE);
1018	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushEntire);
1019	}
1020	#endif
1021	}
1022
1023
1024	/** @name 64-bit guest on 32-bit host OS helper functions.
1025	*
1026	* The host CPU is still 64-bit capable but the host OS is running in 32-bit
1027	* mode (code segment, paging). These wrappers/helpers perform the necessary
1028	* bits for the 32->64 switcher.
1029	*
1030	* @{ */
1031	#if HC_ARCH_BITS == 32 && defined(VBOX_ENABLE_64_BITS_GUESTS)
1032	/**
1033	* Prepares for and executes VMRUN (64-bit guests on a 32-bit host).
1034	*
1035	* @returns VBox status code.
1036	* @param HCPhysVmcbHost Physical address of host VMCB.
1037	* @param HCPhysVmcb Physical address of the VMCB.
1038	* @param pCtx Pointer to the guest-CPU context.
1039	* @param pVM The cross context VM structure.
1040	* @param pVCpu The cross context virtual CPU structure.
1041	*/
1042	DECLASM(int) SVMR0VMSwitcherRun64(RTHCPHYS HCPhysVmcbHost, RTHCPHYS HCPhysVmcb, PCPUMCTX pCtx, PVM pVM, PVMCPU pVCpu)
1043	{
1044	uint32_t aParam[8];
1045	aParam[0] = (uint32_t)(HCPhysVmcbHost); /* Param 1: HCPhysVmcbHost - Lo. */
1046	aParam[1] = (uint32_t)(HCPhysVmcbHost >> 32); /* Param 1: HCPhysVmcbHost - Hi. */
1047	aParam[2] = (uint32_t)(HCPhysVmcb); /* Param 2: HCPhysVmcb - Lo. */
1048	aParam[3] = (uint32_t)(HCPhysVmcb >> 32); /* Param 2: HCPhysVmcb - Hi. */
1049	aParam[4] = VM_RC_ADDR(pVM, pVM);
1050	aParam[5] = 0;
1051	aParam[6] = VM_RC_ADDR(pVM, pVCpu);
1052	aParam[7] = 0;
1053
1054	return SVMR0Execute64BitsHandler(pVM, pVCpu, pCtx, HM64ON32OP_SVMRCVMRun64, RT_ELEMENTS(aParam), &aParam[0]);
1055	}
1056
1057
1058	/**
1059	* Executes the specified VMRUN handler in 64-bit mode.
1060	*
1061	* @returns VBox status code.
1062	* @param pVM The cross context VM structure.
1063	* @param pVCpu The cross context virtual CPU structure.
1064	* @param pCtx Pointer to the guest-CPU context.
1065	* @param enmOp The operation to perform.
1066	* @param cParams Number of parameters.
1067	* @param paParam Array of 32-bit parameters.
1068	*/
1069	VMMR0DECL(int) SVMR0Execute64BitsHandler(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, HM64ON32OP enmOp,
1070	uint32_t cParams, uint32_t *paParam)
1071	{
1072	AssertReturn(pVM->hm.s.pfnHost32ToGuest64R0, VERR_HM_NO_32_TO_64_SWITCHER);
1073	Assert(enmOp > HM64ON32OP_INVALID && enmOp < HM64ON32OP_END);
1074
1075	NOREF(pCtx);
1076
1077	/* Disable interrupts. */
1078	RTHCUINTREG uOldEFlags = ASMIntDisableFlags();
1079
1080	#ifdef VBOX_WITH_VMMR0_DISABLE_LAPIC_NMI
1081	RTCPUID idHostCpu = RTMpCpuId();
1082	CPUMR0SetLApic(pVCpu, idHostCpu);
1083	#endif
1084
1085	CPUMSetHyperESP(pVCpu, VMMGetStackRC(pVCpu));
1086	CPUMSetHyperEIP(pVCpu, enmOp);
1087	for (int i = (int)cParams - 1; i >= 0; i--)
1088	CPUMPushHyper(pVCpu, paParam[i]);
1089
1090	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatWorldSwitch3264, z);
1091	/* Call the switcher. */
1092	int rc = pVM->hm.s.pfnHost32ToGuest64R0(pVM, RT_OFFSETOF(VM, aCpus[pVCpu->idCpu].cpum) - RT_OFFSETOF(VM, cpum));
1093	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatWorldSwitch3264, z);
1094
1095	/* Restore interrupts. */
1096	ASMSetFlags(uOldEFlags);
1097	return rc;
1098	}
1099
1100	#endif /* HC_ARCH_BITS == 32 && defined(VBOX_ENABLE_64_BITS_GUESTS) */
1101	/** @} */
1102
1103
1104	/**
1105	* Adds an exception to the intercept exception bitmap in the VMCB and updates
1106	* the corresponding VMCB Clean bit.
1107	*
1108	* @param pVmcb Pointer to the VM control block.
1109	* @param u32Xcpt The value of the exception (X86_XCPT_*).
1110	*/
1111	DECLINLINE(void) hmR0SvmAddXcptIntercept(PSVMVMCB pVmcb, uint32_t u32Xcpt)
1112	{
1113	if (!(pVmcb->ctrl.u32InterceptException & RT_BIT(u32Xcpt)))
1114	{
1115	pVmcb->ctrl.u32InterceptException \|= RT_BIT(u32Xcpt);
1116	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1117	}
1118	}
1119
1120
1121	/**
1122	* Removes an exception from the intercept-exception bitmap in the VMCB and
1123	* updates the corresponding VMCB Clean bit.
1124	*
1125	* @param pVmcb Pointer to the VM control block.
1126	* @param u32Xcpt The value of the exception (X86_XCPT_*).
1127	*/
1128	DECLINLINE(void) hmR0SvmRemoveXcptIntercept(PSVMVMCB pVmcb, uint32_t u32Xcpt)
1129	{
1130	Assert(u32Xcpt != X86_XCPT_DB);
1131	Assert(u32Xcpt != X86_XCPT_AC);
1132	#ifndef HMSVM_ALWAYS_TRAP_ALL_XCPTS
1133	if (pVmcb->ctrl.u32InterceptException & RT_BIT(u32Xcpt))
1134	{
1135	pVmcb->ctrl.u32InterceptException &= ~RT_BIT(u32Xcpt);
1136	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1137	}
1138	#endif
1139	}
1140
1141
1142	/**
1143	* Loads the guest CR0 control register into the guest-state area in the VMCB.
1144	* Although the guest CR0 is a separate field in the VMCB we have to consider
1145	* the FPU state itself which is shared between the host and the guest.
1146	*
1147	* @returns VBox status code.
1148	* @param pVCpu The cross context virtual CPU structure.
1149	* @param pVmcb Pointer to the VM control block.
1150	* @param pCtx Pointer to the guest-CPU context.
1151	*
1152	* @remarks No-long-jump zone!!!
1153	*/
1154	static void hmR0SvmLoadSharedCR0(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1155	{
1156	/*
1157	* Guest CR0.
1158	*/
1159	PVM pVM = pVCpu->CTX_SUFF(pVM);
1160	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR0))
1161	{
1162	uint64_t u64GuestCR0 = pCtx->cr0;
1163
1164	/* Always enable caching. */
1165	u64GuestCR0 &= ~(X86_CR0_CD \| X86_CR0_NW);
1166
1167	/*
1168	* When Nested Paging is not available use shadow page tables and intercept #PFs (the latter done in SVMR0SetupVM()).
1169	*/
1170	if (!pVM->hm.s.fNestedPaging)
1171	{
1172	u64GuestCR0 \|= X86_CR0_PG; /* When Nested Paging is not available, use shadow page tables. */
1173	u64GuestCR0 \|= X86_CR0_WP; /* Guest CPL 0 writes to its read-only pages should cause a #PF #VMEXIT. */
1174	}
1175
1176	/*
1177	* Guest FPU bits.
1178	*/
1179	bool fInterceptNM = false;
1180	bool fInterceptMF = false;
1181	u64GuestCR0 \|= X86_CR0_NE; /* Use internal x87 FPU exceptions handling rather than external interrupts. */
1182	if (CPUMIsGuestFPUStateActive(pVCpu))
1183	{
1184	/* Catch floating point exceptions if we need to report them to the guest in a different way. */
1185	if (!(pCtx->cr0 & X86_CR0_NE))
1186	{
1187	Log4(("hmR0SvmLoadGuestControlRegs: Intercepting Guest CR0.MP Old-style FPU handling!!!\n"));
1188	fInterceptMF = true;
1189	}
1190	}
1191	else
1192	{
1193	fInterceptNM = true; /* Guest FPU inactive, #VMEXIT on #NM for lazy FPU loading. */
1194	u64GuestCR0 \|= X86_CR0_TS /* Guest can task switch quickly and do lazy FPU syncing. */
1195	\| X86_CR0_MP; /* FWAIT/WAIT should not ignore CR0.TS and should generate #NM. */
1196	}
1197
1198	/*
1199	* Update the exception intercept bitmap.
1200	*/
1201	if (fInterceptNM)
1202	hmR0SvmAddXcptIntercept(pVmcb, X86_XCPT_NM);
1203	else
1204	hmR0SvmRemoveXcptIntercept(pVmcb, X86_XCPT_NM);
1205
1206	if (fInterceptMF)
1207	hmR0SvmAddXcptIntercept(pVmcb, X86_XCPT_MF);
1208	else
1209	hmR0SvmRemoveXcptIntercept(pVmcb, X86_XCPT_MF);
1210
1211	pVmcb->guest.u64CR0 = u64GuestCR0;
1212	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1213	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_CR0);
1214	}
1215	}
1216
1217
1218	/**
1219	* Loads the guest control registers (CR2, CR3, CR4) into the VMCB.
1220	*
1221	* @returns VBox status code.
1222	* @param pVCpu The cross context virtual CPU structure.
1223	* @param pVmcb Pointer to the VM control block.
1224	* @param pCtx Pointer to the guest-CPU context.
1225	*
1226	* @remarks No-long-jump zone!!!
1227	*/
1228	static int hmR0SvmLoadGuestControlRegs(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1229	{
1230	PVM pVM = pVCpu->CTX_SUFF(pVM);
1231
1232	/*
1233	* Guest CR2.
1234	*/
1235	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR2))
1236	{
1237	pVmcb->guest.u64CR2 = pCtx->cr2;
1238	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CR2;
1239	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_CR2);
1240	}
1241
1242	/*
1243	* Guest CR3.
1244	*/
1245	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR3))
1246	{
1247	if (pVM->hm.s.fNestedPaging)
1248	{
1249	PGMMODE enmShwPagingMode;
1250	#if HC_ARCH_BITS == 32
1251	if (CPUMIsGuestInLongModeEx(pCtx))
1252	enmShwPagingMode = PGMMODE_AMD64_NX;
1253	else
1254	#endif
1255	enmShwPagingMode = PGMGetHostMode(pVM);
1256
1257	pVmcb->ctrl.u64NestedPagingCR3 = PGMGetNestedCR3(pVCpu, enmShwPagingMode);
1258	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
1259	Assert(pVmcb->ctrl.u64NestedPagingCR3);
1260	pVmcb->guest.u64CR3 = pCtx->cr3;
1261	}
1262	else
1263	pVmcb->guest.u64CR3 = PGMGetHyperCR3(pVCpu);
1264
1265	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1266	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_CR3);
1267	}
1268
1269	/*
1270	* Guest CR4.
1271	* ASSUMES this is done everytime we get in from ring-3! (XCR0)
1272	*/
1273	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR4))
1274	{
1275	uint64_t u64GuestCR4 = pCtx->cr4;
1276	if (!pVM->hm.s.fNestedPaging)
1277	{
1278	switch (pVCpu->hm.s.enmShadowMode)
1279	{
1280	case PGMMODE_REAL:
1281	case PGMMODE_PROTECTED: /* Protected mode, no paging. */
1282	AssertFailed();
1283	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1284
1285	case PGMMODE_32_BIT: /* 32-bit paging. */
1286	u64GuestCR4 &= ~X86_CR4_PAE;
1287	break;
1288
1289	case PGMMODE_PAE: /* PAE paging. */
1290	case PGMMODE_PAE_NX: /* PAE paging with NX enabled. */
1291	/** Must use PAE paging as we could use physical memory > 4 GB */
1292	u64GuestCR4 \|= X86_CR4_PAE;
1293	break;
1294
1295	case PGMMODE_AMD64: /* 64-bit AMD paging (long mode). */
1296	case PGMMODE_AMD64_NX: /* 64-bit AMD paging (long mode) with NX enabled. */
1297	#ifdef VBOX_ENABLE_64_BITS_GUESTS
1298	break;
1299	#else
1300	AssertFailed();
1301	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1302	#endif
1303
1304	default: /* shut up gcc */
1305	AssertFailed();
1306	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1307	}
1308	}
1309
1310	pVmcb->guest.u64CR4 = u64GuestCR4;
1311	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1312
1313	/* Whether to save/load/restore XCR0 during world switch depends on CR4.OSXSAVE and host+guest XCR0. */
1314	pVCpu->hm.s.fLoadSaveGuestXcr0 = (u64GuestCR4 & X86_CR4_OSXSAVE) && pCtx->aXcr[0] != ASMGetXcr0();
1315
1316	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_CR4);
1317	}
1318
1319	return VINF_SUCCESS;
1320	}
1321
1322
1323	/**
1324	* Loads the guest segment registers into the VMCB.
1325	*
1326	* @returns VBox status code.
1327	* @param pVCpu The cross context virtual CPU structure.
1328	* @param pVmcb Pointer to the VM control block.
1329	* @param pCtx Pointer to the guest-CPU context.
1330	*
1331	* @remarks No-long-jump zone!!!
1332	*/
1333	static void hmR0SvmLoadGuestSegmentRegs(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1334	{
1335	/* Guest Segment registers: CS, SS, DS, ES, FS, GS. */
1336	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_SEGMENT_REGS))
1337	{
1338	HMSVM_LOAD_SEG_REG(CS, cs);
1339	HMSVM_LOAD_SEG_REG(SS, ss);
1340	HMSVM_LOAD_SEG_REG(DS, ds);
1341	HMSVM_LOAD_SEG_REG(ES, es);
1342	HMSVM_LOAD_SEG_REG(FS, fs);
1343	HMSVM_LOAD_SEG_REG(GS, gs);
1344
1345	pVmcb->guest.u8CPL = pCtx->ss.Attr.n.u2Dpl;
1346	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_SEG;
1347	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_SEGMENT_REGS);
1348	}
1349
1350	/* Guest TR. */
1351	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_TR))
1352	{
1353	HMSVM_LOAD_SEG_REG(TR, tr);
1354	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_TR);
1355	}
1356
1357	/* Guest LDTR. */
1358	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_LDTR))
1359	{
1360	HMSVM_LOAD_SEG_REG(LDTR, ldtr);
1361	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_LDTR);
1362	}
1363
1364	/* Guest GDTR. */
1365	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_GDTR))
1366	{
1367	pVmcb->guest.GDTR.u32Limit = pCtx->gdtr.cbGdt;
1368	pVmcb->guest.GDTR.u64Base = pCtx->gdtr.pGdt;
1369	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DT;
1370	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_GDTR);
1371	}
1372
1373	/* Guest IDTR. */
1374	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_IDTR))
1375	{
1376	pVmcb->guest.IDTR.u32Limit = pCtx->idtr.cbIdt;
1377	pVmcb->guest.IDTR.u64Base = pCtx->idtr.pIdt;
1378	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DT;
1379	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_IDTR);
1380	}
1381	}
1382
1383
1384	/**
1385	* Loads the guest MSRs into the VMCB.
1386	*
1387	* @param pVCpu The cross context virtual CPU structure.
1388	* @param pVmcb Pointer to the VM control block.
1389	* @param pCtx Pointer to the guest-CPU context.
1390	*
1391	* @remarks No-long-jump zone!!!
1392	*/
1393	static void hmR0SvmLoadGuestMsrs(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1394	{
1395	/* Guest Sysenter MSRs. */
1396	pVmcb->guest.u64SysEnterCS = pCtx->SysEnter.cs;
1397	pVmcb->guest.u64SysEnterEIP = pCtx->SysEnter.eip;
1398	pVmcb->guest.u64SysEnterESP = pCtx->SysEnter.esp;
1399
1400	/*
1401	* Guest EFER MSR.
1402	* AMD-V requires guest EFER.SVME to be set. Weird.
1403	* See AMD spec. 15.5.1 "Basic Operation" \| "Canonicalization and Consistency Checks".
1404	*/
1405	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_EFER_MSR))
1406	{
1407	pVmcb->guest.u64EFER = pCtx->msrEFER \| MSR_K6_EFER_SVME;
1408	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1409	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_EFER_MSR);
1410	}
1411
1412	/* 64-bit MSRs. */
1413	if (CPUMIsGuestInLongModeEx(pCtx))
1414	{
1415	pVmcb->guest.FS.u64Base = pCtx->fs.u64Base;
1416	pVmcb->guest.GS.u64Base = pCtx->gs.u64Base;
1417	}
1418	else
1419	{
1420	/* If the guest isn't in 64-bit mode, clear MSR_K6_LME bit from guest EFER otherwise AMD-V expects amd64 shadow paging. */
1421	if (pCtx->msrEFER & MSR_K6_EFER_LME)
1422	{
1423	pVmcb->guest.u64EFER &= ~MSR_K6_EFER_LME;
1424	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1425	}
1426	}
1427
1428
1429	/** @todo The following are used in 64-bit only (SYSCALL/SYSRET) but they might
1430	* be writable in 32-bit mode. Clarify with AMD spec. */
1431	pVmcb->guest.u64STAR = pCtx->msrSTAR;
1432	pVmcb->guest.u64LSTAR = pCtx->msrLSTAR;
1433	pVmcb->guest.u64CSTAR = pCtx->msrCSTAR;
1434	pVmcb->guest.u64SFMASK = pCtx->msrSFMASK;
1435	pVmcb->guest.u64KernelGSBase = pCtx->msrKERNELGSBASE;
1436	}
1437
1438
1439	/**
1440	* Loads the guest state into the VMCB and programs the necessary intercepts
1441	* accordingly.
1442	*
1443	* @param pVCpu The cross context virtual CPU structure.
1444	* @param pVmcb Pointer to the VM control block.
1445	* @param pCtx Pointer to the guest-CPU context.
1446	*
1447	* @remarks No-long-jump zone!!!
1448	* @remarks Requires EFLAGS to be up-to-date in the VMCB!
1449	*/
1450	static void hmR0SvmLoadSharedDebugState(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1451	{
1452	if (!HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_DEBUG))
1453	return;
1454	Assert((pCtx->dr[6] & X86_DR6_RA1_MASK) == X86_DR6_RA1_MASK); Assert((pCtx->dr[6] & X86_DR6_RAZ_MASK) == 0);
1455	Assert((pCtx->dr[7] & X86_DR7_RA1_MASK) == X86_DR7_RA1_MASK); Assert((pCtx->dr[7] & X86_DR7_RAZ_MASK) == 0);
1456
1457	bool fInterceptMovDRx = false;
1458
1459	/*
1460	* Anyone single stepping on the host side? If so, we'll have to use the
1461	* trap flag in the guest EFLAGS since AMD-V doesn't have a trap flag on
1462	* the VMM level like the VT-x implementations does.
1463	*/
1464	bool const fStepping = pVCpu->hm.s.fSingleInstruction;
1465	if (fStepping)
1466	{
1467	pVCpu->hm.s.fClearTrapFlag = true;
1468	pVmcb->guest.u64RFlags \|= X86_EFL_TF;
1469	fInterceptMovDRx = true; /* Need clean DR6, no guest mess. */
1470	}
1471	else
1472	Assert(!DBGFIsStepping(pVCpu));
1473
1474	if ( fStepping
1475	\|\| (CPUMGetHyperDR7(pVCpu) & X86_DR7_ENABLED_MASK))
1476	{
1477	/*
1478	* Use the combined guest and host DRx values found in the hypervisor
1479	* register set because the debugger has breakpoints active or someone
1480	* is single stepping on the host side.
1481	*
1482	* Note! DBGF expects a clean DR6 state before executing guest code.
1483	*/
1484	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS)
1485	if ( CPUMIsGuestInLongModeEx(pCtx)
1486	&& !CPUMIsHyperDebugStateActivePending(pVCpu))
1487	{
1488	CPUMR0LoadHyperDebugState(pVCpu, false /* include DR6 */);
1489	Assert(!CPUMIsGuestDebugStateActivePending(pVCpu));
1490	Assert(CPUMIsHyperDebugStateActivePending(pVCpu));
1491	}
1492	else
1493	#endif
1494	if (!CPUMIsHyperDebugStateActive(pVCpu))
1495	{
1496	CPUMR0LoadHyperDebugState(pVCpu, false /* include DR6 */);
1497	Assert(!CPUMIsGuestDebugStateActive(pVCpu));
1498	Assert(CPUMIsHyperDebugStateActive(pVCpu));
1499	}
1500
1501	/* Update DR6 & DR7. (The other DRx values are handled by CPUM one way or the other.) */
1502	if ( pVmcb->guest.u64DR6 != X86_DR6_INIT_VAL
1503	\|\| pVmcb->guest.u64DR7 != CPUMGetHyperDR7(pVCpu))
1504	{
1505	pVmcb->guest.u64DR7 = CPUMGetHyperDR7(pVCpu);
1506	pVmcb->guest.u64DR6 = X86_DR6_INIT_VAL;
1507	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
1508	pVCpu->hm.s.fUsingHyperDR7 = true;
1509	}
1510
1511	/** @todo If we cared, we could optimize to allow the guest to read registers
1512	* with the same values. */
1513	fInterceptMovDRx = true;
1514	Log5(("hmR0SvmLoadSharedDebugState: Loaded hyper DRx\n"));
1515	}
1516	else
1517	{
1518	/*
1519	* Update DR6, DR7 with the guest values if necessary.
1520	*/
1521	if ( pVmcb->guest.u64DR7 != pCtx->dr[7]
1522	\|\| pVmcb->guest.u64DR6 != pCtx->dr[6])
1523	{
1524	pVmcb->guest.u64DR7 = pCtx->dr[7];
1525	pVmcb->guest.u64DR6 = pCtx->dr[6];
1526	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
1527	pVCpu->hm.s.fUsingHyperDR7 = false;
1528	}
1529
1530	/*
1531	* If the guest has enabled debug registers, we need to load them prior to
1532	* executing guest code so they'll trigger at the right time.
1533	*/
1534	if (pCtx->dr[7] & (X86_DR7_ENABLED_MASK \| X86_DR7_GD)) /** @todo Why GD? */
1535	{
1536	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS)
1537	if ( CPUMIsGuestInLongModeEx(pCtx)
1538	&& !CPUMIsGuestDebugStateActivePending(pVCpu))
1539	{
1540	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
1541	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxArmed);
1542	Assert(!CPUMIsHyperDebugStateActivePending(pVCpu));
1543	Assert(CPUMIsGuestDebugStateActivePending(pVCpu));
1544	}
1545	else
1546	#endif
1547	if (!CPUMIsGuestDebugStateActive(pVCpu))
1548	{
1549	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
1550	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxArmed);
1551	Assert(!CPUMIsHyperDebugStateActive(pVCpu));
1552	Assert(CPUMIsGuestDebugStateActive(pVCpu));
1553	}
1554	Log5(("hmR0SvmLoadSharedDebugState: Loaded guest DRx\n"));
1555	}
1556	/*
1557	* If no debugging enabled, we'll lazy load DR0-3. We don't need to
1558	* intercept #DB as DR6 is updated in the VMCB.
1559	*
1560	* Note! If we cared and dared, we could skip intercepting \#DB here.
1561	* However, \#DB shouldn't be performance critical, so we'll play safe
1562	* and keep the code similar to the VT-x code and always intercept it.
1563	*/
1564	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS)
1565	else if ( !CPUMIsGuestDebugStateActivePending(pVCpu)
1566	&& !CPUMIsGuestDebugStateActive(pVCpu))
1567	#else
1568	else if (!CPUMIsGuestDebugStateActive(pVCpu))
1569	#endif
1570	{
1571	fInterceptMovDRx = true;
1572	}
1573	}
1574
1575	Assert(pVmcb->ctrl.u32InterceptException & RT_BIT_32(X86_XCPT_DB));
1576	if (fInterceptMovDRx)
1577	{
1578	if ( pVmcb->ctrl.u16InterceptRdDRx != 0xffff
1579	\|\| pVmcb->ctrl.u16InterceptWrDRx != 0xffff)
1580	{
1581	pVmcb->ctrl.u16InterceptRdDRx = 0xffff;
1582	pVmcb->ctrl.u16InterceptWrDRx = 0xffff;
1583	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1584	}
1585	}
1586	else
1587	{
1588	if ( pVmcb->ctrl.u16InterceptRdDRx
1589	\|\| pVmcb->ctrl.u16InterceptWrDRx)
1590	{
1591	pVmcb->ctrl.u16InterceptRdDRx = 0;
1592	pVmcb->ctrl.u16InterceptWrDRx = 0;
1593	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1594	}
1595	}
1596
1597	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_DEBUG);
1598	}
1599
1600
1601	/**
1602	* Loads the guest APIC state (currently just the TPR).
1603	*
1604	* @returns VBox status code.
1605	* @param pVCpu The cross context virtual CPU structure.
1606	* @param pVmcb Pointer to the VM control block.
1607	* @param pCtx Pointer to the guest-CPU context.
1608	*/
1609	static int hmR0SvmLoadGuestApicState(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1610	{
1611	if (!HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE))
1612	return VINF_SUCCESS;
1613
1614	int rc = VINF_SUCCESS;
1615	PVM pVM = pVCpu->CTX_SUFF(pVM);
1616	if ( PDMHasApic(pVM)
1617	&& APICIsEnabled(pVCpu))
1618	{
1619	bool fPendingIntr;
1620	uint8_t u8Tpr;
1621	rc = APICGetTpr(pVCpu, &u8Tpr, &fPendingIntr, NULL /* pu8PendingIrq */);
1622	AssertRCReturn(rc, rc);
1623
1624	/* Assume that we need to trap all TPR accesses and thus need not check on
1625	every #VMEXIT if we should update the TPR. */
1626	Assert(pVmcb->ctrl.IntCtrl.n.u1VIrqMasking);
1627	pVCpu->hm.s.svm.fSyncVTpr = false;
1628
1629	/* 32-bit guests uses LSTAR MSR for patching guest code which touches the TPR. */
1630	if (pVM->hm.s.fTPRPatchingActive)
1631	{
1632	pCtx->msrLSTAR = u8Tpr;
1633
1634	/* If there are interrupts pending, intercept LSTAR writes, otherwise don't intercept reads or writes. */
1635	if (fPendingIntr)
1636	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_INTERCEPT_WRITE);
1637	else
1638	{
1639	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1640	pVCpu->hm.s.svm.fSyncVTpr = true;
1641	}
1642	}
1643	else
1644	{
1645	/* Bits 3-0 of the VTPR field correspond to bits 7-4 of the TPR (which is the Task-Priority Class). */
1646	pVmcb->ctrl.IntCtrl.n.u8VTPR = (u8Tpr >> 4);
1647
1648	/* If there are interrupts pending, intercept CR8 writes to evaluate ASAP if we can deliver the interrupt to the guest. */
1649	if (fPendingIntr)
1650	pVmcb->ctrl.u16InterceptWrCRx \|= RT_BIT(8);
1651	else
1652	{
1653	pVmcb->ctrl.u16InterceptWrCRx &= ~RT_BIT(8);
1654	pVCpu->hm.s.svm.fSyncVTpr = true;
1655	}
1656
1657	pVmcb->ctrl.u64VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS \| HMSVM_VMCB_CLEAN_TPR);
1658	}
1659	}
1660
1661	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
1662	return rc;
1663	}
1664
1665
1666	/**
1667	* Loads the exception interrupts required for guest execution in the VMCB.
1668	*
1669	* @returns VBox status code.
1670	* @param pVCpu The cross context virtual CPU structure.
1671	* @param pVmcb Pointer to the VM control block.
1672	* @param pCtx Pointer to the guest-CPU context.
1673	*/
1674	static int hmR0SvmLoadGuestXcptIntercepts(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1675	{
1676	NOREF(pCtx);
1677	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_XCPT_INTERCEPTS))
1678	{
1679	/* Trap #UD for GIM provider (e.g. for hypercalls). */
1680	if (pVCpu->hm.s.fGIMTrapXcptUD)
1681	hmR0SvmAddXcptIntercept(pVmcb, X86_XCPT_UD);
1682	else
1683	hmR0SvmRemoveXcptIntercept(pVmcb, X86_XCPT_UD);
1684
1685	/* Trap #BP for INT3 debug breakpoints set by the VM debugger. */
1686	if (pVCpu->CTX_SUFF(pVM)->dbgf.ro.cEnabledInt3Breakpoints)
1687	hmR0SvmAddXcptIntercept(pVmcb, X86_XCPT_BP);
1688	else
1689	hmR0SvmRemoveXcptIntercept(pVmcb, X86_XCPT_BP);
1690
1691	/* The remaining intercepts are handled elsewhere, e.g. in hmR0SvmLoadSharedCR0(). */
1692	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_XCPT_INTERCEPTS);
1693	}
1694	return VINF_SUCCESS;
1695	}
1696
1697
1698	/**
1699	* Sets up the appropriate function to run guest code.
1700	*
1701	* @returns VBox status code.
1702	* @param pVCpu The cross context virtual CPU structure.
1703	* @param pCtx Pointer to the guest-CPU context.
1704	*
1705	* @remarks No-long-jump zone!!!
1706	*/
1707	static int hmR0SvmSetupVMRunHandler(PVMCPU pVCpu, PCPUMCTX pCtx)
1708	{
1709	if (CPUMIsGuestInLongModeEx(pCtx))
1710	{
1711	#ifndef VBOX_ENABLE_64_BITS_GUESTS
1712	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1713	#endif
1714	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.fAllow64BitGuests); /* Guaranteed by hmR3InitFinalizeR0(). */
1715	#if HC_ARCH_BITS == 32
1716	/* 32-bit host. We need to switch to 64-bit before running the 64-bit guest. */
1717	pVCpu->hm.s.svm.pfnVMRun = SVMR0VMSwitcherRun64;
1718	#else
1719	/* 64-bit host or hybrid host. */
1720	pVCpu->hm.s.svm.pfnVMRun = SVMR0VMRun64;
1721	#endif
1722	}
1723	else
1724	{
1725	/* Guest is not in long mode, use the 32-bit handler. */
1726	pVCpu->hm.s.svm.pfnVMRun = SVMR0VMRun;
1727	}
1728	return VINF_SUCCESS;
1729	}
1730
1731
1732	/**
1733	* Enters the AMD-V session.
1734	*
1735	* @returns VBox status code.
1736	* @param pVM The cross context VM structure.
1737	* @param pVCpu The cross context virtual CPU structure.
1738	* @param pCpu Pointer to the CPU info struct.
1739	*/
1740	VMMR0DECL(int) SVMR0Enter(PVM pVM, PVMCPU pVCpu, PHMGLOBALCPUINFO pCpu)
1741	{
1742	AssertPtr(pVM);
1743	AssertPtr(pVCpu);
1744	Assert(pVM->hm.s.svm.fSupported);
1745	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1746	NOREF(pVM); NOREF(pCpu);
1747
1748	LogFlowFunc(("pVM=%p pVCpu=%p\n", pVM, pVCpu));
1749	Assert(HMCPU_CF_IS_SET(pVCpu, HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_HOST_GUEST_SHARED_STATE));
1750
1751	pVCpu->hm.s.fLeaveDone = false;
1752	return VINF_SUCCESS;
1753	}
1754
1755
1756	/**
1757	* Thread-context callback for AMD-V.
1758	*
1759	* @param enmEvent The thread-context event.
1760	* @param pVCpu The cross context virtual CPU structure.
1761	* @param fGlobalInit Whether global VT-x/AMD-V init. is used.
1762	* @thread EMT(pVCpu)
1763	*/
1764	VMMR0DECL(void) SVMR0ThreadCtxCallback(RTTHREADCTXEVENT enmEvent, PVMCPU pVCpu, bool fGlobalInit)
1765	{
1766	NOREF(fGlobalInit);
1767
1768	switch (enmEvent)
1769	{
1770	case RTTHREADCTXEVENT_OUT:
1771	{
1772	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1773	Assert(VMMR0ThreadCtxHookIsEnabled(pVCpu));
1774	VMCPU_ASSERT_EMT(pVCpu);
1775
1776	/* No longjmps (log-flush, locks) in this fragile context. */
1777	VMMRZCallRing3Disable(pVCpu);
1778
1779	if (!pVCpu->hm.s.fLeaveDone)
1780	{
1781	hmR0SvmLeave(pVCpu);
1782	pVCpu->hm.s.fLeaveDone = true;
1783	}
1784
1785	/* Leave HM context, takes care of local init (term). */
1786	int rc = HMR0LeaveCpu(pVCpu);
1787	AssertRC(rc); NOREF(rc);
1788
1789	/* Restore longjmp state. */
1790	VMMRZCallRing3Enable(pVCpu);
1791	STAM_REL_COUNTER_INC(&pVCpu->hm.s.StatSwitchPreempt);
1792	break;
1793	}
1794
1795	case RTTHREADCTXEVENT_IN:
1796	{
1797	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1798	Assert(VMMR0ThreadCtxHookIsEnabled(pVCpu));
1799	VMCPU_ASSERT_EMT(pVCpu);
1800
1801	/* No longjmps (log-flush, locks) in this fragile context. */
1802	VMMRZCallRing3Disable(pVCpu);
1803
1804	/*
1805	* Initialize the bare minimum state required for HM. This takes care of
1806	* initializing AMD-V if necessary (onlined CPUs, local init etc.)
1807	*/
1808	int rc = HMR0EnterCpu(pVCpu);
1809	AssertRC(rc); NOREF(rc);
1810	Assert(HMCPU_CF_IS_SET(pVCpu, HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_HOST_GUEST_SHARED_STATE));
1811
1812	pVCpu->hm.s.fLeaveDone = false;
1813
1814	/* Restore longjmp state. */
1815	VMMRZCallRing3Enable(pVCpu);
1816	break;
1817	}
1818
1819	default:
1820	break;
1821	}
1822	}
1823
1824
1825	/**
1826	* Saves the host state.
1827	*
1828	* @returns VBox status code.
1829	* @param pVM The cross context VM structure.
1830	* @param pVCpu The cross context virtual CPU structure.
1831	*
1832	* @remarks No-long-jump zone!!!
1833	*/
1834	VMMR0DECL(int) SVMR0SaveHostState(PVM pVM, PVMCPU pVCpu)
1835	{
1836	NOREF(pVM);
1837	NOREF(pVCpu);
1838	/* Nothing to do here. AMD-V does this for us automatically during the world-switch. */
1839	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_HOST_CONTEXT);
1840	return VINF_SUCCESS;
1841	}
1842
1843
1844	/**
1845	* Loads the guest state into the VMCB.
1846	*
1847	* The CPU state will be loaded from these fields on every successful VM-entry.
1848	* Also sets up the appropriate VMRUN function to execute guest code based on
1849	* the guest CPU mode.
1850	*
1851	* @returns VBox status code.
1852	* @param pVM The cross context VM structure.
1853	* @param pVCpu The cross context virtual CPU structure.
1854	* @param pCtx Pointer to the guest-CPU context.
1855	*
1856	* @remarks No-long-jump zone!!!
1857	*/
1858	static int hmR0SvmLoadGuestState(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
1859	{
1860	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
1861	AssertMsgReturn(pVmcb, ("Invalid pVmcb\n"), VERR_SVM_INVALID_PVMCB);
1862
1863	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatLoadGuestState, x);
1864
1865	int rc = hmR0SvmLoadGuestControlRegs(pVCpu, pVmcb, pCtx);
1866	AssertLogRelMsgRCReturn(rc, ("hmR0SvmLoadGuestControlRegs! rc=%Rrc (pVM=%p pVCpu=%p)\n", rc, pVM, pVCpu), rc);
1867
1868	hmR0SvmLoadGuestSegmentRegs(pVCpu, pVmcb, pCtx);
1869	hmR0SvmLoadGuestMsrs(pVCpu, pVmcb, pCtx);
1870
1871	pVmcb->guest.u64RIP = pCtx->rip;
1872	pVmcb->guest.u64RSP = pCtx->rsp;
1873	pVmcb->guest.u64RFlags = pCtx->eflags.u32;
1874	pVmcb->guest.u64RAX = pCtx->rax;
1875
1876	rc = hmR0SvmLoadGuestApicState(pVCpu, pVmcb, pCtx);
1877	AssertLogRelMsgRCReturn(rc, ("hmR0SvmLoadGuestApicState! rc=%Rrc (pVM=%p pVCpu=%p)\n", rc, pVM, pVCpu), rc);
1878
1879	rc = hmR0SvmLoadGuestXcptIntercepts(pVCpu, pVmcb, pCtx);
1880	AssertLogRelMsgRCReturn(rc, ("hmR0SvmLoadGuestXcptIntercepts! rc=%Rrc (pVM=%p pVCpu=%p)\n", rc, pVM, pVCpu), rc);
1881
1882	rc = hmR0SvmSetupVMRunHandler(pVCpu, pCtx);
1883	AssertLogRelMsgRCReturn(rc, ("hmR0SvmSetupVMRunHandler! rc=%Rrc (pVM=%p pVCpu=%p)\n", rc, pVM, pVCpu), rc);
1884
1885	/* Clear any unused and reserved bits. */
1886	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_RIP /* Unused (loaded unconditionally). */
1887	\| HM_CHANGED_GUEST_RSP
1888	\| HM_CHANGED_GUEST_RFLAGS
1889	\| HM_CHANGED_GUEST_SYSENTER_CS_MSR
1890	\| HM_CHANGED_GUEST_SYSENTER_EIP_MSR
1891	\| HM_CHANGED_GUEST_SYSENTER_ESP_MSR
1892	\| HM_CHANGED_GUEST_LAZY_MSRS /* Unused. */
1893	\| HM_CHANGED_SVM_RESERVED1 /* Reserved. */
1894	\| HM_CHANGED_SVM_RESERVED2
1895	\| HM_CHANGED_SVM_RESERVED3
1896	\| HM_CHANGED_SVM_RESERVED4);
1897
1898	/* All the guest state bits should be loaded except maybe the host context and/or shared host/guest bits. */
1899	AssertMsg( !HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_ALL_GUEST)
1900	\|\| HMCPU_CF_IS_PENDING_ONLY(pVCpu, HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_HOST_GUEST_SHARED_STATE),
1901	("fContextUseFlags=%#RX32\n", HMCPU_CF_VALUE(pVCpu)));
1902
1903	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));
1904	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatLoadGuestState, x);
1905	return rc;
1906	}
1907
1908
1909	/**
1910	* Loads the state shared between the host and guest into the
1911	* VMCB.
1912	*
1913	* @param pVCpu The cross context virtual CPU structure.
1914	* @param pVmcb Pointer to the VM control block.
1915	* @param pCtx Pointer to the guest-CPU context.
1916	*
1917	* @remarks No-long-jump zone!!!
1918	*/
1919	static void hmR0SvmLoadSharedState(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1920	{
1921	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1922	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
1923
1924	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR0))
1925	hmR0SvmLoadSharedCR0(pVCpu, pVmcb, pCtx);
1926
1927	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_DEBUG))
1928	hmR0SvmLoadSharedDebugState(pVCpu, pVmcb, pCtx);
1929
1930	/* Unused on AMD-V. */
1931	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_LAZY_MSRS);
1932
1933	AssertMsg(!HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_HOST_GUEST_SHARED_STATE),
1934	("fContextUseFlags=%#RX32\n", HMCPU_CF_VALUE(pVCpu)));
1935	}
1936
1937
1938	/**
1939	* Saves the entire guest state from the VMCB into the
1940	* guest-CPU context. Currently there is no residual state left in the CPU that
1941	* is not updated in the VMCB.
1942	*
1943	* @returns VBox status code.
1944	* @param pVCpu The cross context virtual CPU structure.
1945	* @param pMixedCtx Pointer to the guest-CPU context. The data may be
1946	* out-of-sync. Make sure to update the required fields
1947	* before using them.
1948	*/
1949	static void hmR0SvmSaveGuestState(PVMCPU pVCpu, PCPUMCTX pMixedCtx)
1950	{
1951	Assert(VMMRZCallRing3IsEnabled(pVCpu));
1952
1953	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
1954
1955	pMixedCtx->rip = pVmcb->guest.u64RIP;
1956	pMixedCtx->rsp = pVmcb->guest.u64RSP;
1957	pMixedCtx->eflags.u32 = pVmcb->guest.u64RFlags;
1958	pMixedCtx->rax = pVmcb->guest.u64RAX;
1959
1960	/*
1961	* Guest interrupt shadow.
1962	*/
1963	if (pVmcb->ctrl.u64IntShadow & SVM_INTERRUPT_SHADOW_ACTIVE)
1964	EMSetInhibitInterruptsPC(pVCpu, pMixedCtx->rip);
1965	else if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS))
1966	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS);
1967
1968	/*
1969	* Guest Control registers: CR2, CR3 (handled at the end) - accesses to other control registers are always intercepted.
1970	*/
1971	pMixedCtx->cr2 = pVmcb->guest.u64CR2;
1972
1973	/*
1974	* Guest MSRs.
1975	*/
1976	pMixedCtx->msrSTAR = pVmcb->guest.u64STAR; /* legacy syscall eip, cs & ss */
1977	pMixedCtx->msrLSTAR = pVmcb->guest.u64LSTAR; /* 64-bit mode syscall rip */
1978	pMixedCtx->msrCSTAR = pVmcb->guest.u64CSTAR; /* compatibility mode syscall rip */
1979	pMixedCtx->msrSFMASK = pVmcb->guest.u64SFMASK; /* syscall flag mask */
1980	pMixedCtx->msrKERNELGSBASE = pVmcb->guest.u64KernelGSBase; /* swapgs exchange value */
1981	pMixedCtx->SysEnter.cs = pVmcb->guest.u64SysEnterCS;
1982	pMixedCtx->SysEnter.eip = pVmcb->guest.u64SysEnterEIP;
1983	pMixedCtx->SysEnter.esp = pVmcb->guest.u64SysEnterESP;
1984
1985	/*
1986	* Guest segment registers (includes FS, GS base MSRs for 64-bit guests).
1987	*/
1988	HMSVM_SAVE_SEG_REG(CS, cs);
1989	HMSVM_SAVE_SEG_REG(SS, ss);
1990	HMSVM_SAVE_SEG_REG(DS, ds);
1991	HMSVM_SAVE_SEG_REG(ES, es);
1992	HMSVM_SAVE_SEG_REG(FS, fs);
1993	HMSVM_SAVE_SEG_REG(GS, gs);
1994
1995	/*
1996	* Correct the hidden CS granularity bit. Haven't seen it being wrong in any other
1997	* register (yet).
1998	*/
1999	/** @todo SELM might need to be fixed as it too should not care about the
2000	* granularity bit. See @bugref{6785}. */
2001	if ( !pMixedCtx->cs.Attr.n.u1Granularity
2002	&& pMixedCtx->cs.Attr.n.u1Present
2003	&& pMixedCtx->cs.u32Limit > UINT32_C(0xfffff))
2004	{
2005	Assert((pMixedCtx->cs.u32Limit & 0xfff) == 0xfff);
2006	pMixedCtx->cs.Attr.n.u1Granularity = 1;
2007	}
2008
2009	#ifdef VBOX_STRICT
2010	# define HMSVM_ASSERT_SEG_GRANULARITY(reg) \
2011	AssertMsg( !pMixedCtx->reg.Attr.n.u1Present \
2012	\|\| ( pMixedCtx->reg.Attr.n.u1Granularity \
2013	? (pMixedCtx->reg.u32Limit & 0xfff) == 0xfff \
2014	: pMixedCtx->reg.u32Limit <= UINT32_C(0xfffff)), \
2015	("Invalid Segment Attributes Limit=%#RX32 Attr=%#RX32 Base=%#RX64\n", pMixedCtx->reg.u32Limit, \
2016	pMixedCtx->reg.Attr.u, pMixedCtx->reg.u64Base))
2017
2018	HMSVM_ASSERT_SEG_GRANULARITY(cs);
2019	HMSVM_ASSERT_SEG_GRANULARITY(ss);
2020	HMSVM_ASSERT_SEG_GRANULARITY(ds);
2021	HMSVM_ASSERT_SEG_GRANULARITY(es);
2022	HMSVM_ASSERT_SEG_GRANULARITY(fs);
2023	HMSVM_ASSERT_SEG_GRANULARITY(gs);
2024
2025	# undef HMSVM_ASSERT_SEL_GRANULARITY
2026	#endif
2027
2028	/*
2029	* Sync the hidden SS DPL field. AMD CPUs have a separate CPL field in the VMCB and uses that
2030	* and thus it's possible that when the CPL changes during guest execution that the SS DPL
2031	* isn't updated by AMD-V. Observed on some AMD Fusion CPUs with 64-bit guests.
2032	* See AMD spec. 15.5.1 "Basic operation".
2033	*/
2034	Assert(!(pVmcb->guest.u8CPL & ~0x3));
2035	pMixedCtx->ss.Attr.n.u2Dpl = pVmcb->guest.u8CPL & 0x3;
2036
2037	/*
2038	* Guest TR.
2039	* Fixup TR attributes so it's compatible with Intel. Important when saved-states are used
2040	* between Intel and AMD. See @bugref{6208#c39}.
2041	* ASSUME that it's normally correct and that we're in 32-bit or 64-bit mode.
2042	*/
2043	HMSVM_SAVE_SEG_REG(TR, tr);
2044	if (pMixedCtx->tr.Attr.n.u4Type != X86_SEL_TYPE_SYS_386_TSS_BUSY)
2045	{
2046	if ( pMixedCtx->tr.Attr.n.u4Type == X86_SEL_TYPE_SYS_386_TSS_AVAIL
2047	\|\| CPUMIsGuestInLongModeEx(pMixedCtx))
2048	pMixedCtx->tr.Attr.n.u4Type = X86_SEL_TYPE_SYS_386_TSS_BUSY;
2049	else if (pMixedCtx->tr.Attr.n.u4Type == X86_SEL_TYPE_SYS_286_TSS_AVAIL)
2050	pMixedCtx->tr.Attr.n.u4Type = X86_SEL_TYPE_SYS_286_TSS_BUSY;
2051	}
2052
2053	/*
2054	* Guest Descriptor-Table registers.
2055	*/
2056	HMSVM_SAVE_SEG_REG(LDTR, ldtr);
2057	pMixedCtx->gdtr.cbGdt = pVmcb->guest.GDTR.u32Limit;
2058	pMixedCtx->gdtr.pGdt = pVmcb->guest.GDTR.u64Base;
2059
2060	pMixedCtx->idtr.cbIdt = pVmcb->guest.IDTR.u32Limit;
2061	pMixedCtx->idtr.pIdt = pVmcb->guest.IDTR.u64Base;
2062
2063	/*
2064	* Guest Debug registers.
2065	*/
2066	if (!pVCpu->hm.s.fUsingHyperDR7)
2067	{
2068	pMixedCtx->dr[6] = pVmcb->guest.u64DR6;
2069	pMixedCtx->dr[7] = pVmcb->guest.u64DR7;
2070	}
2071	else
2072	{
2073	Assert(pVmcb->guest.u64DR7 == CPUMGetHyperDR7(pVCpu));
2074	CPUMSetHyperDR6(pVCpu, pVmcb->guest.u64DR6);
2075	}
2076
2077	/*
2078	* With Nested Paging, CR3 changes are not intercepted. Therefore, sync. it now.
2079	* This is done as the very last step of syncing the guest state, as PGMUpdateCR3() may cause longjmp's to ring-3.
2080	*/
2081	if ( pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging
2082	&& pMixedCtx->cr3 != pVmcb->guest.u64CR3)
2083	{
2084	CPUMSetGuestCR3(pVCpu, pVmcb->guest.u64CR3);
2085	PGMUpdateCR3(pVCpu, pVmcb->guest.u64CR3);
2086	}
2087	}
2088
2089
2090	/**
2091	* Does the necessary state syncing before returning to ring-3 for any reason
2092	* (longjmp, preemption, voluntary exits to ring-3) from AMD-V.
2093	*
2094	* @param pVCpu The cross context virtual CPU structure.
2095	*
2096	* @remarks No-long-jmp zone!!!
2097	*/
2098	static void hmR0SvmLeave(PVMCPU pVCpu)
2099	{
2100	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2101	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2102	Assert(VMMR0IsLogFlushDisabled(pVCpu));
2103
2104	/*
2105	* !!! IMPORTANT !!!
2106	* If you modify code here, make sure to check whether hmR0SvmCallRing3Callback() needs to be updated too.
2107	*/
2108
2109	/* Restore host FPU state if necessary and resync on next R0 reentry .*/
2110	if (CPUMR0FpuStateMaybeSaveGuestAndRestoreHost(pVCpu))
2111	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR0);
2112
2113	/*
2114	* Restore host debug registers if necessary and resync on next R0 reentry.
2115	*/
2116	#ifdef VBOX_STRICT
2117	if (CPUMIsHyperDebugStateActive(pVCpu))
2118	{
2119	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2120	Assert(pVmcb->ctrl.u16InterceptRdDRx == 0xffff);
2121	Assert(pVmcb->ctrl.u16InterceptWrDRx == 0xffff);
2122	}
2123	#endif
2124	if (CPUMR0DebugStateMaybeSaveGuestAndRestoreHost(pVCpu, false /* save DR6 */))
2125	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_DEBUG);
2126
2127	Assert(!CPUMIsHyperDebugStateActive(pVCpu));
2128	Assert(!CPUMIsGuestDebugStateActive(pVCpu));
2129
2130	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatEntry);
2131	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatLoadGuestState);
2132	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatExit1);
2133	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatExit2);
2134	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchLongJmpToR3);
2135
2136	VMCPU_CMPXCHG_STATE(pVCpu, VMCPUSTATE_STARTED_HM, VMCPUSTATE_STARTED_EXEC);
2137	}
2138
2139
2140	/**
2141	* Leaves the AMD-V session.
2142	*
2143	* @returns VBox status code.
2144	* @param pVCpu The cross context virtual CPU structure.
2145	*/
2146	static int hmR0SvmLeaveSession(PVMCPU pVCpu)
2147	{
2148	HM_DISABLE_PREEMPT();
2149	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2150	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2151
2152	/* When thread-context hooks are used, we can avoid doing the leave again if we had been preempted before
2153	and done this from the SVMR0ThreadCtxCallback(). */
2154	if (!pVCpu->hm.s.fLeaveDone)
2155	{
2156	hmR0SvmLeave(pVCpu);
2157	pVCpu->hm.s.fLeaveDone = true;
2158	}
2159
2160	/*
2161	* !!! IMPORTANT !!!
2162	* If you modify code here, make sure to check whether hmR0SvmCallRing3Callback() needs to be updated too.
2163	*/
2164
2165	/** @todo eliminate the need for calling VMMR0ThreadCtxHookDisable here! */
2166	/* Deregister hook now that we've left HM context before re-enabling preemption. */
2167	VMMR0ThreadCtxHookDisable(pVCpu);
2168
2169	/* Leave HM context. This takes care of local init (term). */
2170	int rc = HMR0LeaveCpu(pVCpu);
2171
2172	HM_RESTORE_PREEMPT();
2173	return rc;
2174	}
2175
2176
2177	/**
2178	* Does the necessary state syncing before doing a longjmp to ring-3.
2179	*
2180	* @returns VBox status code.
2181	* @param pVCpu The cross context virtual CPU structure.
2182	*
2183	* @remarks No-long-jmp zone!!!
2184	*/
2185	static int hmR0SvmLongJmpToRing3(PVMCPU pVCpu)
2186	{
2187	return hmR0SvmLeaveSession(pVCpu);
2188	}
2189
2190
2191	/**
2192	* VMMRZCallRing3() callback wrapper which saves the guest state (or restores
2193	* any remaining host state) before we longjump to ring-3 and possibly get
2194	* preempted.
2195	*
2196	* @param pVCpu The cross context virtual CPU structure.
2197	* @param enmOperation The operation causing the ring-3 longjump.
2198	* @param pvUser The user argument (pointer to the possibly
2199	* out-of-date guest-CPU context).
2200	*/
2201	static DECLCALLBACK(int) hmR0SvmCallRing3Callback(PVMCPU pVCpu, VMMCALLRING3 enmOperation, void *pvUser)
2202	{
2203	RT_NOREF_PV(pvUser);
2204
2205	if (enmOperation == VMMCALLRING3_VM_R0_ASSERTION)
2206	{
2207	/*
2208	* !!! IMPORTANT !!!
2209	* If you modify code here, make sure to check whether hmR0SvmLeave() and hmR0SvmLeaveSession() needs
2210	* to be updated too. This is a stripped down version which gets out ASAP trying to not trigger any assertion.
2211	*/
2212	VMMRZCallRing3RemoveNotification(pVCpu);
2213	VMMRZCallRing3Disable(pVCpu);
2214	HM_DISABLE_PREEMPT();
2215
2216	/* Restore host FPU state if necessary and resync on next R0 reentry. */
2217	CPUMR0FpuStateMaybeSaveGuestAndRestoreHost(pVCpu);
2218
2219	/* Restore host debug registers if necessary and resync on next R0 reentry. */
2220	CPUMR0DebugStateMaybeSaveGuestAndRestoreHost(pVCpu, false /* save DR6 */);
2221
2222	/* Deregister the hook now that we've left HM context before re-enabling preemption. */
2223	/** @todo eliminate the need for calling VMMR0ThreadCtxHookDisable here! */
2224	VMMR0ThreadCtxHookDisable(pVCpu);
2225
2226	/* Leave HM context. This takes care of local init (term). */
2227	HMR0LeaveCpu(pVCpu);
2228
2229	HM_RESTORE_PREEMPT();
2230	return VINF_SUCCESS;
2231	}
2232
2233	Assert(pVCpu);
2234	Assert(pvUser);
2235	Assert(VMMRZCallRing3IsEnabled(pVCpu));
2236	HMSVM_ASSERT_PREEMPT_SAFE();
2237
2238	VMMRZCallRing3Disable(pVCpu);
2239	Assert(VMMR0IsLogFlushDisabled(pVCpu));
2240
2241	Log4(("hmR0SvmCallRing3Callback->hmR0SvmLongJmpToRing3\n"));
2242	int rc = hmR0SvmLongJmpToRing3(pVCpu);
2243	AssertRCReturn(rc, rc);
2244
2245	VMMRZCallRing3Enable(pVCpu);
2246	return VINF_SUCCESS;
2247	}
2248
2249
2250	/**
2251	* Take necessary actions before going back to ring-3.
2252	*
2253	* An action requires us to go back to ring-3. This function does the necessary
2254	* steps before we can safely return to ring-3. This is not the same as longjmps
2255	* to ring-3, this is voluntary.
2256	*
2257	* @param pVM The cross context VM structure.
2258	* @param pVCpu The cross context virtual CPU structure.
2259	* @param pCtx Pointer to the guest-CPU context.
2260	* @param rcExit The reason for exiting to ring-3. Can be
2261	* VINF_VMM_UNKNOWN_RING3_CALL.
2262	*/
2263	static void hmR0SvmExitToRing3(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, int rcExit)
2264	{
2265	Assert(pVM);
2266	Assert(pVCpu);
2267	Assert(pCtx);
2268	HMSVM_ASSERT_PREEMPT_SAFE();
2269
2270	/* Please, no longjumps here (any logging shouldn't flush jump back to ring-3). NO LOGGING BEFORE THIS POINT! */
2271	VMMRZCallRing3Disable(pVCpu);
2272	Log4(("hmR0SvmExitToRing3: rcExit=%d\n", rcExit));
2273
2274	/* We need to do this only while truly exiting the "inner loop" back to ring-3 and -not- for any longjmp to ring3. */
2275	if (pVCpu->hm.s.Event.fPending)
2276	{
2277	hmR0SvmPendingEventToTrpmTrap(pVCpu);
2278	Assert(!pVCpu->hm.s.Event.fPending);
2279	}
2280
2281	/* If we're emulating an instruction, we shouldn't have any TRPM traps pending
2282	and if we're injecting an event we should have a TRPM trap pending. */
2283	AssertMsg(rcExit != VINF_EM_RAW_INJECT_TRPM_EVENT \|\| TRPMHasTrap(pVCpu), ("rcExit=%Rrc\n", rcExit));
2284	AssertMsg(rcExit != VINF_EM_RAW_EMULATE_INSTR \|\| !TRPMHasTrap(pVCpu), ("rcExit=%Rrc\n", rcExit));
2285
2286	/* Sync. the necessary state for going back to ring-3. */
2287	hmR0SvmLeaveSession(pVCpu);
2288	STAM_COUNTER_DEC(&pVCpu->hm.s.StatSwitchLongJmpToR3);
2289
2290	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_TO_R3);
2291	CPUMSetChangedFlags(pVCpu, CPUM_CHANGED_SYSENTER_MSR
2292	\| CPUM_CHANGED_LDTR
2293	\| CPUM_CHANGED_GDTR
2294	\| CPUM_CHANGED_IDTR
2295	\| CPUM_CHANGED_TR
2296	\| CPUM_CHANGED_HIDDEN_SEL_REGS);
2297	if ( pVM->hm.s.fNestedPaging
2298	&& CPUMIsGuestPagingEnabledEx(pCtx))
2299	{
2300	CPUMSetChangedFlags(pVCpu, CPUM_CHANGED_GLOBAL_TLB_FLUSH);
2301	}
2302
2303	/* On our way back from ring-3 reload the guest state if there is a possibility of it being changed. */
2304	if (rcExit != VINF_EM_RAW_INTERRUPT)
2305	HMCPU_CF_SET(pVCpu, HM_CHANGED_ALL_GUEST);
2306
2307	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchExitToR3);
2308
2309	/* We do -not- want any longjmp notifications after this! We must return to ring-3 ASAP. */
2310	VMMRZCallRing3RemoveNotification(pVCpu);
2311	VMMRZCallRing3Enable(pVCpu);
2312	}
2313
2314
2315	/**
2316	* Updates the use of TSC offsetting mode for the CPU and adjusts the necessary
2317	* intercepts.
2318	*
2319	* @param pVM The cross context VM structure.
2320	* @param pVCpu The cross context virtual CPU structure.
2321	*
2322	* @remarks No-long-jump zone!!!
2323	*/
2324	static void hmR0SvmUpdateTscOffsetting(PVM pVM, PVMCPU pVCpu)
2325	{
2326	bool fParavirtTsc;
2327	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2328	bool fCanUseRealTsc = TMCpuTickCanUseRealTSC(pVM, pVCpu, &pVmcb->ctrl.u64TSCOffset, &fParavirtTsc);
2329	if (fCanUseRealTsc)
2330	{
2331	pVmcb->ctrl.u64InterceptCtrl &= ~SVM_CTRL_INTERCEPT_RDTSC;
2332	pVmcb->ctrl.u64InterceptCtrl &= ~SVM_CTRL_INTERCEPT_RDTSCP;
2333	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscOffset);
2334	}
2335	else
2336	{
2337	pVmcb->ctrl.u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_RDTSC;
2338	pVmcb->ctrl.u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_RDTSCP;
2339	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscIntercept);
2340	}
2341	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2342
2343	/** @todo later optimize this to be done elsewhere and not before every
2344	* VM-entry. */
2345	if (fParavirtTsc)
2346	{
2347	/* Currently neither Hyper-V nor KVM need to update their paravirt. TSC
2348	information before every VM-entry, hence disable it for performance sake. */
2349	#if 0
2350	int rc = GIMR0UpdateParavirtTsc(pVM, 0 /* u64Offset */);
2351	AssertRC(rc);
2352	#endif
2353	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscParavirt);
2354	}
2355	}
2356
2357
2358	/**
2359	* Sets an event as a pending event to be injected into the guest.
2360	*
2361	* @param pVCpu The cross context virtual CPU structure.
2362	* @param pEvent Pointer to the SVM event.
2363	* @param GCPtrFaultAddress The fault-address (CR2) in case it's a
2364	* page-fault.
2365	*
2366	* @remarks Statistics counter assumes this is a guest event being reflected to
2367	* the guest i.e. 'StatInjectPendingReflect' is incremented always.
2368	*/
2369	DECLINLINE(void) hmR0SvmSetPendingEvent(PVMCPU pVCpu, PSVMEVENT pEvent, RTGCUINTPTR GCPtrFaultAddress)
2370	{
2371	Assert(!pVCpu->hm.s.Event.fPending);
2372	Assert(pEvent->n.u1Valid);
2373
2374	pVCpu->hm.s.Event.u64IntInfo = pEvent->u;
2375	pVCpu->hm.s.Event.fPending = true;
2376	pVCpu->hm.s.Event.GCPtrFaultAddress = GCPtrFaultAddress;
2377
2378	Log4(("hmR0SvmSetPendingEvent: u=%#RX64 u8Vector=%#x Type=%#x ErrorCodeValid=%RTbool ErrorCode=%#RX32\n", pEvent->u,
2379	pEvent->n.u8Vector, (uint8_t)pEvent->n.u3Type, !!pEvent->n.u1ErrorCodeValid, pEvent->n.u32ErrorCode));
2380	}
2381
2382
2383	/**
2384	* Injects an event into the guest upon VMRUN by updating the relevant field
2385	* in the VMCB.
2386	*
2387	* @param pVCpu The cross context virtual CPU structure.
2388	* @param pVmcb Pointer to the guest VM control block.
2389	* @param pCtx Pointer to the guest-CPU context.
2390	* @param pEvent Pointer to the event.
2391	*
2392	* @remarks No-long-jump zone!!!
2393	* @remarks Requires CR0!
2394	*/
2395	DECLINLINE(void) hmR0SvmInjectEventVmcb(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx, PSVMEVENT pEvent)
2396	{
2397	NOREF(pVCpu); NOREF(pCtx);
2398
2399	pVmcb->ctrl.EventInject.u = pEvent->u;
2400	STAM_COUNTER_INC(&pVCpu->hm.s.paStatInjectedIrqsR0[pEvent->n.u8Vector & MASK_INJECT_IRQ_STAT]);
2401
2402	Log4(("hmR0SvmInjectEventVmcb: u=%#RX64 u8Vector=%#x Type=%#x ErrorCodeValid=%RTbool ErrorCode=%#RX32\n", pEvent->u,
2403	pEvent->n.u8Vector, (uint8_t)pEvent->n.u3Type, !!pEvent->n.u1ErrorCodeValid, pEvent->n.u32ErrorCode));
2404	}
2405
2406
2407
2408	/**
2409	* Converts any TRPM trap into a pending HM event. This is typically used when
2410	* entering from ring-3 (not longjmp returns).
2411	*
2412	* @param pVCpu The cross context virtual CPU structure.
2413	*/
2414	static void hmR0SvmTrpmTrapToPendingEvent(PVMCPU pVCpu)
2415	{
2416	Assert(TRPMHasTrap(pVCpu));
2417	Assert(!pVCpu->hm.s.Event.fPending);
2418
2419	uint8_t uVector;
2420	TRPMEVENT enmTrpmEvent;
2421	RTGCUINT uErrCode;
2422	RTGCUINTPTR GCPtrFaultAddress;
2423	uint8_t cbInstr;
2424
2425	int rc = TRPMQueryTrapAll(pVCpu, &uVector, &enmTrpmEvent, &uErrCode, &GCPtrFaultAddress, &cbInstr);
2426	AssertRC(rc);
2427
2428	SVMEVENT Event;
2429	Event.u = 0;
2430	Event.n.u1Valid = 1;
2431	Event.n.u8Vector = uVector;
2432
2433	/* Refer AMD spec. 15.20 "Event Injection" for the format. */
2434	if (enmTrpmEvent == TRPM_TRAP)
2435	{
2436	Event.n.u3Type = SVM_EVENT_EXCEPTION;
2437	switch (uVector)
2438	{
2439	case X86_XCPT_NMI:
2440	{
2441	Event.n.u3Type = SVM_EVENT_NMI;
2442	break;
2443	}
2444
2445	case X86_XCPT_PF:
2446	case X86_XCPT_DF:
2447	case X86_XCPT_TS:
2448	case X86_XCPT_NP:
2449	case X86_XCPT_SS:
2450	case X86_XCPT_GP:
2451	case X86_XCPT_AC:
2452	{
2453	Event.n.u1ErrorCodeValid = 1;
2454	Event.n.u32ErrorCode = uErrCode;
2455	break;
2456	}
2457	}
2458	}
2459	else if (enmTrpmEvent == TRPM_HARDWARE_INT)
2460	Event.n.u3Type = SVM_EVENT_EXTERNAL_IRQ;
2461	else if (enmTrpmEvent == TRPM_SOFTWARE_INT)
2462	Event.n.u3Type = SVM_EVENT_SOFTWARE_INT;
2463	else
2464	AssertMsgFailed(("Invalid TRPM event type %d\n", enmTrpmEvent));
2465
2466	rc = TRPMResetTrap(pVCpu);
2467	AssertRC(rc);
2468
2469	Log4(("TRPM->HM event: u=%#RX64 u8Vector=%#x uErrorCodeValid=%RTbool uErrorCode=%#RX32\n", Event.u, Event.n.u8Vector,
2470	!!Event.n.u1ErrorCodeValid, Event.n.u32ErrorCode));
2471
2472	hmR0SvmSetPendingEvent(pVCpu, &Event, GCPtrFaultAddress);
2473	}
2474
2475
2476	/**
2477	* Converts any pending SVM event into a TRPM trap. Typically used when leaving
2478	* AMD-V to execute any instruction.
2479	*
2480	* @param pVCpu The cross context virtual CPU structure.
2481	*/
2482	static void hmR0SvmPendingEventToTrpmTrap(PVMCPU pVCpu)
2483	{
2484	Assert(pVCpu->hm.s.Event.fPending);
2485	Assert(TRPMQueryTrap(pVCpu, NULL /* pu8TrapNo /, NULL / pEnmType */) == VERR_TRPM_NO_ACTIVE_TRAP);
2486
2487	SVMEVENT Event;
2488	Event.u = pVCpu->hm.s.Event.u64IntInfo;
2489
2490	uint8_t uVector = Event.n.u8Vector;
2491	uint8_t uVectorType = Event.n.u3Type;
2492
2493	TRPMEVENT enmTrapType;
2494	switch (uVectorType)
2495	{
2496	case SVM_EVENT_EXTERNAL_IRQ:
2497	enmTrapType = TRPM_HARDWARE_INT;
2498	break;
2499	case SVM_EVENT_SOFTWARE_INT:
2500	enmTrapType = TRPM_SOFTWARE_INT;
2501	break;
2502	case SVM_EVENT_EXCEPTION:
2503	case SVM_EVENT_NMI:
2504	enmTrapType = TRPM_TRAP;
2505	break;
2506	default:
2507	AssertMsgFailed(("Invalid pending-event type %#x\n", uVectorType));
2508	enmTrapType = TRPM_32BIT_HACK;
2509	break;
2510	}
2511
2512	Log4(("HM event->TRPM: uVector=%#x enmTrapType=%d\n", uVector, uVectorType));
2513
2514	int rc = TRPMAssertTrap(pVCpu, uVector, enmTrapType);
2515	AssertRC(rc);
2516
2517	if (Event.n.u1ErrorCodeValid)
2518	TRPMSetErrorCode(pVCpu, Event.n.u32ErrorCode);
2519
2520	if ( uVectorType == SVM_EVENT_EXCEPTION
2521	&& uVector == X86_XCPT_PF)
2522	{
2523	TRPMSetFaultAddress(pVCpu, pVCpu->hm.s.Event.GCPtrFaultAddress);
2524	Assert(pVCpu->hm.s.Event.GCPtrFaultAddress == CPUMGetGuestCR2(pVCpu));
2525	}
2526	else if (uVectorType == SVM_EVENT_SOFTWARE_INT)
2527	{
2528	AssertMsg( uVectorType == SVM_EVENT_SOFTWARE_INT
2529	\|\| (uVector == X86_XCPT_BP \|\| uVector == X86_XCPT_OF),
2530	("Invalid vector: uVector=%#x uVectorType=%#x\n", uVector, uVectorType));
2531	TRPMSetInstrLength(pVCpu, pVCpu->hm.s.Event.cbInstr);
2532	}
2533	pVCpu->hm.s.Event.fPending = false;
2534	}
2535
2536
2537	/**
2538	* Gets the guest's interrupt-shadow.
2539	*
2540	* @returns The guest's interrupt-shadow.
2541	* @param pVCpu The cross context virtual CPU structure.
2542	* @param pCtx Pointer to the guest-CPU context.
2543	*
2544	* @remarks No-long-jump zone!!!
2545	* @remarks Has side-effects with VMCPU_FF_INHIBIT_INTERRUPTS force-flag.
2546	*/
2547	DECLINLINE(uint32_t) hmR0SvmGetGuestIntrShadow(PVMCPU pVCpu, PCPUMCTX pCtx)
2548	{
2549	/*
2550	* Instructions like STI and MOV SS inhibit interrupts till the next instruction completes. Check if we should
2551	* inhibit interrupts or clear any existing interrupt-inhibition.
2552	*/
2553	uint32_t uIntrState = 0;
2554	if (VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS))
2555	{
2556	if (pCtx->rip != EMGetInhibitInterruptsPC(pVCpu))
2557	{
2558	/*
2559	* We can clear the inhibit force flag as even if we go back to the recompiler without executing guest code in
2560	* AMD-V, the flag's condition to be cleared is met and thus the cleared state is correct.
2561	*/
2562	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS);
2563	}
2564	else
2565	uIntrState = SVM_INTERRUPT_SHADOW_ACTIVE;
2566	}
2567	return uIntrState;
2568	}
2569
2570
2571	/**
2572	* Sets the virtual interrupt intercept control in the VMCB which
2573	* instructs AMD-V to cause a \#VMEXIT as soon as the guest is in a state to
2574	* receive interrupts.
2575	*
2576	* @param pVmcb Pointer to the VM control block.
2577	*/
2578	DECLINLINE(void) hmR0SvmSetVirtIntrIntercept(PSVMVMCB pVmcb)
2579	{
2580	if (!(pVmcb->ctrl.u64InterceptCtrl & SVM_CTRL_INTERCEPT_VINTR))
2581	{
2582	pVmcb->ctrl.IntCtrl.n.u1VIrqValid = 1; /* A virtual interrupt is pending. */
2583	pVmcb->ctrl.IntCtrl.n.u8VIrqVector = 0; /* Not necessary as we #VMEXIT for delivering the interrupt. */
2584	pVmcb->ctrl.u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_VINTR;
2585	pVmcb->ctrl.u64VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS \| HMSVM_VMCB_CLEAN_TPR);
2586
2587	Log4(("Setting VINTR intercept\n"));
2588	}
2589	}
2590
2591
2592	#if 0
2593	/**
2594	* Clears the virtual interrupt intercept control in the VMCB as
2595	* we are figured the guest is unable process any interrupts
2596	* at this point of time.
2597	*
2598	* @param pVmcb Pointer to the VM control block.
2599	*/
2600	DECLINLINE(void) hmR0SvmClearVirtIntrIntercept(PSVMVMCB pVmcb)
2601	{
2602	if (pVmcb->ctrl.u64InterceptCtrl & SVM_CTRL_INTERCEPT_VINTR)
2603	{
2604	pVmcb->ctrl.u64InterceptCtrl &= ~SVM_CTRL_INTERCEPT_VINTR;
2605	pVmcb->ctrl.u64VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS);
2606	Log4(("Clearing VINTR intercept\n"));
2607	}
2608	}
2609	#endif
2610
2611
2612	/**
2613	* Sets the IRET intercept control in the VMCB which instructs AMD-V to cause a
2614	* \#VMEXIT as soon as a guest starts executing an IRET. This is used to unblock
2615	* virtual NMIs.
2616	*
2617	* @param pVmcb Pointer to the VM control block.
2618	*/
2619	DECLINLINE(void) hmR0SvmSetIretIntercept(PSVMVMCB pVmcb)
2620	{
2621	if (!(pVmcb->ctrl.u64InterceptCtrl & SVM_CTRL_INTERCEPT_IRET))
2622	{
2623	pVmcb->ctrl.u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_IRET;
2624	pVmcb->ctrl.u64VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS);
2625
2626	Log4(("Setting IRET intercept\n"));
2627	}
2628	}
2629
2630
2631	/**
2632	* Clears the IRET intercept control in the VMCB.
2633	*
2634	* @param pVmcb Pointer to the VM control block.
2635	*/
2636	DECLINLINE(void) hmR0SvmClearIretIntercept(PSVMVMCB pVmcb)
2637	{
2638	if (pVmcb->ctrl.u64InterceptCtrl & SVM_CTRL_INTERCEPT_IRET)
2639	{
2640	pVmcb->ctrl.u64InterceptCtrl &= ~SVM_CTRL_INTERCEPT_IRET;
2641	pVmcb->ctrl.u64VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS);
2642
2643	Log4(("Clearing IRET intercept\n"));
2644	}
2645	}
2646
2647
2648	/**
2649	* Evaluates the event to be delivered to the guest and sets it as the pending
2650	* event.
2651	*
2652	* @param pVCpu The cross context virtual CPU structure.
2653	* @param pCtx Pointer to the guest-CPU context.
2654	*/
2655	static void hmR0SvmEvaluatePendingEvent(PVMCPU pVCpu, PCPUMCTX pCtx)
2656	{
2657	Assert(!pVCpu->hm.s.Event.fPending);
2658	Log4Func(("\n"));
2659
2660	bool const fIntShadow = RT_BOOL(hmR0SvmGetGuestIntrShadow(pVCpu, pCtx));
2661	bool const fBlockInt = !(pCtx->eflags.u32 & X86_EFL_IF);
2662	bool const fBlockNmi = RT_BOOL(VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_BLOCK_NMIS));
2663	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2664
2665	if (VMCPU_FF_TEST_AND_CLEAR(pVCpu, VMCPU_FF_UPDATE_APIC))
2666	APICUpdatePendingInterrupts(pVCpu);
2667
2668	SVMEVENT Event;
2669	Event.u = 0;
2670	/** @todo SMI. SMIs take priority over NMIs. */
2671	if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_NMI)) /* NMI. NMIs take priority over regular interrupts . */
2672	{
2673	if (fBlockNmi)
2674	hmR0SvmSetIretIntercept(pVmcb);
2675	else if (fIntShadow)
2676	hmR0SvmSetVirtIntrIntercept(pVmcb);
2677	else
2678	{
2679	Log4(("Pending NMI\n"));
2680
2681	Event.n.u1Valid = 1;
2682	Event.n.u8Vector = X86_XCPT_NMI;
2683	Event.n.u3Type = SVM_EVENT_NMI;
2684
2685	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
2686	hmR0SvmSetIretIntercept(pVmcb);
2687	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INTERRUPT_NMI);
2688	}
2689	}
2690	else if ( VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC)
2691	&& !pVCpu->hm.s.fSingleInstruction)
2692	{
2693	/*
2694	* Check if the guest can receive external interrupts (PIC/APIC). Once PDMGetInterrupt() returns
2695	* a valid interrupt we must- deliver the interrupt. We can no longer re-request it from the APIC.
2696	*/
2697	if ( !fBlockInt
2698	&& !fIntShadow)
2699	{
2700	uint8_t u8Interrupt;
2701	int rc = PDMGetInterrupt(pVCpu, &u8Interrupt);
2702	if (RT_SUCCESS(rc))
2703	{
2704	Log4(("Injecting external interrupt u8Interrupt=%#x\n", u8Interrupt));
2705
2706	Event.n.u1Valid = 1;
2707	Event.n.u8Vector = u8Interrupt;
2708	Event.n.u3Type = SVM_EVENT_EXTERNAL_IRQ;
2709
2710	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
2711	}
2712	else if (rc == VERR_APIC_INTR_MASKED_BY_TPR)
2713	{
2714	/*
2715	* AMD-V has no TPR thresholding feature. We just avoid posting the interrupt.
2716	* We just avoid delivering the TPR-masked interrupt here. TPR will be updated
2717	* always via hmR0SvmLoadGuestState() -> hmR0SvmLoadGuestApicState().
2718	*/
2719	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchTprMaskedIrq);
2720	}
2721	else
2722	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchGuestIrq);
2723	}
2724	else
2725	hmR0SvmSetVirtIntrIntercept(pVmcb);
2726	}
2727	}
2728
2729
2730	/**
2731	* Injects any pending events into the guest if the guest is in a state to
2732	* receive them.
2733	*
2734	* @param pVCpu The cross context virtual CPU structure.
2735	* @param pCtx Pointer to the guest-CPU context.
2736	*/
2737	static void hmR0SvmInjectPendingEvent(PVMCPU pVCpu, PCPUMCTX pCtx)
2738	{
2739	Assert(!TRPMHasTrap(pVCpu));
2740	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2741
2742	bool const fIntShadow = RT_BOOL(hmR0SvmGetGuestIntrShadow(pVCpu, pCtx));
2743	bool const fBlockInt = !(pCtx->eflags.u32 & X86_EFL_IF);
2744	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2745
2746	if (pVCpu->hm.s.Event.fPending) /* First, inject any pending HM events. */
2747	{
2748	SVMEVENT Event;
2749	Event.u = pVCpu->hm.s.Event.u64IntInfo;
2750	Assert(Event.n.u1Valid);
2751	#ifdef VBOX_STRICT
2752	if (Event.n.u3Type == SVM_EVENT_EXTERNAL_IRQ)
2753	{
2754	Assert(!fBlockInt);
2755	Assert(!fIntShadow);
2756	}
2757	else if (Event.n.u3Type == SVM_EVENT_NMI)
2758	Assert(!fIntShadow);
2759	#endif
2760
2761	Log4(("Injecting pending HM event.\n"));
2762	hmR0SvmInjectEventVmcb(pVCpu, pVmcb, pCtx, &Event);
2763	pVCpu->hm.s.Event.fPending = false;
2764
2765	#ifdef VBOX_WITH_STATISTICS
2766	if (Event.n.u3Type == SVM_EVENT_EXTERNAL_IRQ)
2767	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectInterrupt);
2768	else
2769	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectXcpt);
2770	#endif
2771	}
2772
2773	/* Update the guest interrupt shadow in the VMCB. */
2774	pVmcb->ctrl.u64IntShadow = !!fIntShadow;
2775	NOREF(fBlockInt);
2776	}
2777
2778
2779	/**
2780	* Reports world-switch error and dumps some useful debug info.
2781	*
2782	* @param pVM The cross context VM structure.
2783	* @param pVCpu The cross context virtual CPU structure.
2784	* @param rcVMRun The return code from VMRUN (or
2785	* VERR_SVM_INVALID_GUEST_STATE for invalid
2786	* guest-state).
2787	* @param pCtx Pointer to the guest-CPU context.
2788	*/
2789	static void hmR0SvmReportWorldSwitchError(PVM pVM, PVMCPU pVCpu, int rcVMRun, PCPUMCTX pCtx)
2790	{
2791	NOREF(pCtx);
2792	HMSVM_ASSERT_PREEMPT_SAFE();
2793	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2794
2795	if (rcVMRun == VERR_SVM_INVALID_GUEST_STATE)
2796	{
2797	hmDumpRegs(pVM, pVCpu, pCtx); NOREF(pVM);
2798	#ifdef VBOX_STRICT
2799	Log4(("ctrl.u64VmcbCleanBits %#RX64\n", pVmcb->ctrl.u64VmcbCleanBits));
2800	Log4(("ctrl.u16InterceptRdCRx %#x\n", pVmcb->ctrl.u16InterceptRdCRx));
2801	Log4(("ctrl.u16InterceptWrCRx %#x\n", pVmcb->ctrl.u16InterceptWrCRx));
2802	Log4(("ctrl.u16InterceptRdDRx %#x\n", pVmcb->ctrl.u16InterceptRdDRx));
2803	Log4(("ctrl.u16InterceptWrDRx %#x\n", pVmcb->ctrl.u16InterceptWrDRx));
2804	Log4(("ctrl.u32InterceptException %#x\n", pVmcb->ctrl.u32InterceptException));
2805	Log4(("ctrl.u64InterceptCtrl %#RX64\n", pVmcb->ctrl.u64InterceptCtrl));
2806	Log4(("ctrl.u64IOPMPhysAddr %#RX64\n", pVmcb->ctrl.u64IOPMPhysAddr));
2807	Log4(("ctrl.u64MSRPMPhysAddr %#RX64\n", pVmcb->ctrl.u64MSRPMPhysAddr));
2808	Log4(("ctrl.u64TSCOffset %#RX64\n", pVmcb->ctrl.u64TSCOffset));
2809
2810	Log4(("ctrl.TLBCtrl.u32ASID %#x\n", pVmcb->ctrl.TLBCtrl.n.u32ASID));
2811	Log4(("ctrl.TLBCtrl.u8TLBFlush %#x\n", pVmcb->ctrl.TLBCtrl.n.u8TLBFlush));
2812	Log4(("ctrl.TLBCtrl.u24Reserved %#x\n", pVmcb->ctrl.TLBCtrl.n.u24Reserved));
2813
2814	Log4(("ctrl.IntCtrl.u8VTPR %#x\n", pVmcb->ctrl.IntCtrl.n.u8VTPR));
2815	Log4(("ctrl.IntCtrl.u1VIrqValid %#x\n", pVmcb->ctrl.IntCtrl.n.u1VIrqValid));
2816	Log4(("ctrl.IntCtrl.u7Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u7Reserved));
2817	Log4(("ctrl.IntCtrl.u4VIrqPriority %#x\n", pVmcb->ctrl.IntCtrl.n.u4VIrqPriority));
2818	Log4(("ctrl.IntCtrl.u1IgnoreTPR %#x\n", pVmcb->ctrl.IntCtrl.n.u1IgnoreTPR));
2819	Log4(("ctrl.IntCtrl.u3Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u3Reserved));
2820	Log4(("ctrl.IntCtrl.u1VIrqMasking %#x\n", pVmcb->ctrl.IntCtrl.n.u1VIrqMasking));
2821	Log4(("ctrl.IntCtrl.u6Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u6Reserved));
2822	Log4(("ctrl.IntCtrl.u8VIrqVector %#x\n", pVmcb->ctrl.IntCtrl.n.u8VIrqVector));
2823	Log4(("ctrl.IntCtrl.u24Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u24Reserved));
2824
2825	Log4(("ctrl.u64IntShadow %#RX64\n", pVmcb->ctrl.u64IntShadow));
2826	Log4(("ctrl.u64ExitCode %#RX64\n", pVmcb->ctrl.u64ExitCode));
2827	Log4(("ctrl.u64ExitInfo1 %#RX64\n", pVmcb->ctrl.u64ExitInfo1));
2828	Log4(("ctrl.u64ExitInfo2 %#RX64\n", pVmcb->ctrl.u64ExitInfo2));
2829	Log4(("ctrl.ExitIntInfo.u8Vector %#x\n", pVmcb->ctrl.ExitIntInfo.n.u8Vector));
2830	Log4(("ctrl.ExitIntInfo.u3Type %#x\n", pVmcb->ctrl.ExitIntInfo.n.u3Type));
2831	Log4(("ctrl.ExitIntInfo.u1ErrorCodeValid %#x\n", pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid));
2832	Log4(("ctrl.ExitIntInfo.u19Reserved %#x\n", pVmcb->ctrl.ExitIntInfo.n.u19Reserved));
2833	Log4(("ctrl.ExitIntInfo.u1Valid %#x\n", pVmcb->ctrl.ExitIntInfo.n.u1Valid));
2834	Log4(("ctrl.ExitIntInfo.u32ErrorCode %#x\n", pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode));
2835	Log4(("ctrl.NestedPaging %#RX64\n", pVmcb->ctrl.NestedPaging.u));
2836	Log4(("ctrl.EventInject.u8Vector %#x\n", pVmcb->ctrl.EventInject.n.u8Vector));
2837	Log4(("ctrl.EventInject.u3Type %#x\n", pVmcb->ctrl.EventInject.n.u3Type));
2838	Log4(("ctrl.EventInject.u1ErrorCodeValid %#x\n", pVmcb->ctrl.EventInject.n.u1ErrorCodeValid));
2839	Log4(("ctrl.EventInject.u19Reserved %#x\n", pVmcb->ctrl.EventInject.n.u19Reserved));
2840	Log4(("ctrl.EventInject.u1Valid %#x\n", pVmcb->ctrl.EventInject.n.u1Valid));
2841	Log4(("ctrl.EventInject.u32ErrorCode %#x\n", pVmcb->ctrl.EventInject.n.u32ErrorCode));
2842
2843	Log4(("ctrl.u64NestedPagingCR3 %#RX64\n", pVmcb->ctrl.u64NestedPagingCR3));
2844	Log4(("ctrl.u64LBRVirt %#RX64\n", pVmcb->ctrl.u64LBRVirt));
2845
2846	Log4(("guest.CS.u16Sel %RTsel\n", pVmcb->guest.CS.u16Sel));
2847	Log4(("guest.CS.u16Attr %#x\n", pVmcb->guest.CS.u16Attr));
2848	Log4(("guest.CS.u32Limit %#RX32\n", pVmcb->guest.CS.u32Limit));
2849	Log4(("guest.CS.u64Base %#RX64\n", pVmcb->guest.CS.u64Base));
2850	Log4(("guest.DS.u16Sel %#RTsel\n", pVmcb->guest.DS.u16Sel));
2851	Log4(("guest.DS.u16Attr %#x\n", pVmcb->guest.DS.u16Attr));
2852	Log4(("guest.DS.u32Limit %#RX32\n", pVmcb->guest.DS.u32Limit));
2853	Log4(("guest.DS.u64Base %#RX64\n", pVmcb->guest.DS.u64Base));
2854	Log4(("guest.ES.u16Sel %RTsel\n", pVmcb->guest.ES.u16Sel));
2855	Log4(("guest.ES.u16Attr %#x\n", pVmcb->guest.ES.u16Attr));
2856	Log4(("guest.ES.u32Limit %#RX32\n", pVmcb->guest.ES.u32Limit));
2857	Log4(("guest.ES.u64Base %#RX64\n", pVmcb->guest.ES.u64Base));
2858	Log4(("guest.FS.u16Sel %RTsel\n", pVmcb->guest.FS.u16Sel));
2859	Log4(("guest.FS.u16Attr %#x\n", pVmcb->guest.FS.u16Attr));
2860	Log4(("guest.FS.u32Limit %#RX32\n", pVmcb->guest.FS.u32Limit));
2861	Log4(("guest.FS.u64Base %#RX64\n", pVmcb->guest.FS.u64Base));
2862	Log4(("guest.GS.u16Sel %RTsel\n", pVmcb->guest.GS.u16Sel));
2863	Log4(("guest.GS.u16Attr %#x\n", pVmcb->guest.GS.u16Attr));
2864	Log4(("guest.GS.u32Limit %#RX32\n", pVmcb->guest.GS.u32Limit));
2865	Log4(("guest.GS.u64Base %#RX64\n", pVmcb->guest.GS.u64Base));
2866
2867	Log4(("guest.GDTR.u32Limit %#RX32\n", pVmcb->guest.GDTR.u32Limit));
2868	Log4(("guest.GDTR.u64Base %#RX64\n", pVmcb->guest.GDTR.u64Base));
2869
2870	Log4(("guest.LDTR.u16Sel %RTsel\n", pVmcb->guest.LDTR.u16Sel));
2871	Log4(("guest.LDTR.u16Attr %#x\n", pVmcb->guest.LDTR.u16Attr));
2872	Log4(("guest.LDTR.u32Limit %#RX32\n", pVmcb->guest.LDTR.u32Limit));
2873	Log4(("guest.LDTR.u64Base %#RX64\n", pVmcb->guest.LDTR.u64Base));
2874
2875	Log4(("guest.IDTR.u32Limit %#RX32\n", pVmcb->guest.IDTR.u32Limit));
2876	Log4(("guest.IDTR.u64Base %#RX64\n", pVmcb->guest.IDTR.u64Base));
2877
2878	Log4(("guest.TR.u16Sel %RTsel\n", pVmcb->guest.TR.u16Sel));
2879	Log4(("guest.TR.u16Attr %#x\n", pVmcb->guest.TR.u16Attr));
2880	Log4(("guest.TR.u32Limit %#RX32\n", pVmcb->guest.TR.u32Limit));
2881	Log4(("guest.TR.u64Base %#RX64\n", pVmcb->guest.TR.u64Base));
2882
2883	Log4(("guest.u8CPL %#x\n", pVmcb->guest.u8CPL));
2884	Log4(("guest.u64CR0 %#RX64\n", pVmcb->guest.u64CR0));
2885	Log4(("guest.u64CR2 %#RX64\n", pVmcb->guest.u64CR2));
2886	Log4(("guest.u64CR3 %#RX64\n", pVmcb->guest.u64CR3));
2887	Log4(("guest.u64CR4 %#RX64\n", pVmcb->guest.u64CR4));
2888	Log4(("guest.u64DR6 %#RX64\n", pVmcb->guest.u64DR6));
2889	Log4(("guest.u64DR7 %#RX64\n", pVmcb->guest.u64DR7));
2890
2891	Log4(("guest.u64RIP %#RX64\n", pVmcb->guest.u64RIP));
2892	Log4(("guest.u64RSP %#RX64\n", pVmcb->guest.u64RSP));
2893	Log4(("guest.u64RAX %#RX64\n", pVmcb->guest.u64RAX));
2894	Log4(("guest.u64RFlags %#RX64\n", pVmcb->guest.u64RFlags));
2895
2896	Log4(("guest.u64SysEnterCS %#RX64\n", pVmcb->guest.u64SysEnterCS));
2897	Log4(("guest.u64SysEnterEIP %#RX64\n", pVmcb->guest.u64SysEnterEIP));
2898	Log4(("guest.u64SysEnterESP %#RX64\n", pVmcb->guest.u64SysEnterESP));
2899
2900	Log4(("guest.u64EFER %#RX64\n", pVmcb->guest.u64EFER));
2901	Log4(("guest.u64STAR %#RX64\n", pVmcb->guest.u64STAR));
2902	Log4(("guest.u64LSTAR %#RX64\n", pVmcb->guest.u64LSTAR));
2903	Log4(("guest.u64CSTAR %#RX64\n", pVmcb->guest.u64CSTAR));
2904	Log4(("guest.u64SFMASK %#RX64\n", pVmcb->guest.u64SFMASK));
2905	Log4(("guest.u64KernelGSBase %#RX64\n", pVmcb->guest.u64KernelGSBase));
2906	Log4(("guest.u64GPAT %#RX64\n", pVmcb->guest.u64GPAT));
2907	Log4(("guest.u64DBGCTL %#RX64\n", pVmcb->guest.u64DBGCTL));
2908	Log4(("guest.u64BR_FROM %#RX64\n", pVmcb->guest.u64BR_FROM));
2909	Log4(("guest.u64BR_TO %#RX64\n", pVmcb->guest.u64BR_TO));
2910	Log4(("guest.u64LASTEXCPFROM %#RX64\n", pVmcb->guest.u64LASTEXCPFROM));
2911	Log4(("guest.u64LASTEXCPTO %#RX64\n", pVmcb->guest.u64LASTEXCPTO));
2912	#endif /* VBOX_STRICT */
2913	}
2914	else
2915	Log4(("hmR0SvmReportWorldSwitchError: rcVMRun=%d\n", rcVMRun));
2916
2917	NOREF(pVmcb);
2918	}
2919
2920
2921	/**
2922	* Check per-VM and per-VCPU force flag actions that require us to go back to
2923	* ring-3 for one reason or another.
2924	*
2925	* @returns VBox status code (information status code included).
2926	* @retval VINF_SUCCESS if we don't have any actions that require going back to
2927	* ring-3.
2928	* @retval VINF_PGM_SYNC_CR3 if we have pending PGM CR3 sync.
2929	* @retval VINF_EM_PENDING_REQUEST if we have pending requests (like hardware
2930	* interrupts)
2931	* @retval VINF_PGM_POOL_FLUSH_PENDING if PGM is doing a pool flush and requires
2932	* all EMTs to be in ring-3.
2933	* @retval VINF_EM_RAW_TO_R3 if there is pending DMA requests.
2934	* @retval VINF_EM_NO_MEMORY PGM is out of memory, we need to return
2935	* to the EM loop.
2936	*
2937	* @param pVM The cross context VM structure.
2938	* @param pVCpu The cross context virtual CPU structure.
2939	* @param pCtx Pointer to the guest-CPU context.
2940	*/
2941	static int hmR0SvmCheckForceFlags(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
2942	{
2943	Assert(VMMRZCallRing3IsEnabled(pVCpu));
2944
2945	/* On AMD-V we don't need to update CR3, PAE PDPES lazily. See hmR0SvmSaveGuestState(). */
2946	Assert(!VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_UPDATE_CR3));
2947	Assert(!VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_UPDATE_PAE_PDPES));
2948
2949	if ( VM_FF_IS_PENDING(pVM, !pVCpu->hm.s.fSingleInstruction
2950	? VM_FF_HP_R0_PRE_HM_MASK : VM_FF_HP_R0_PRE_HM_STEP_MASK)
2951	\|\| VMCPU_FF_IS_PENDING(pVCpu, !pVCpu->hm.s.fSingleInstruction
2952	? VMCPU_FF_HP_R0_PRE_HM_MASK : VMCPU_FF_HP_R0_PRE_HM_STEP_MASK) )
2953	{
2954	/* Pending PGM C3 sync. */
2955	if (VMCPU_FF_IS_PENDING(pVCpu,VMCPU_FF_PGM_SYNC_CR3 \| VMCPU_FF_PGM_SYNC_CR3_NON_GLOBAL))
2956	{
2957	int rc = PGMSyncCR3(pVCpu, pCtx->cr0, pCtx->cr3, pCtx->cr4, VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_PGM_SYNC_CR3));
2958	if (rc != VINF_SUCCESS)
2959	{
2960	Log4(("hmR0SvmCheckForceFlags: PGMSyncCR3 forcing us back to ring-3. rc=%d\n", rc));
2961	return rc;
2962	}
2963	}
2964
2965	/* Pending HM-to-R3 operations (critsects, timers, EMT rendezvous etc.) */
2966	/* -XXX- what was that about single stepping? */
2967	if ( VM_FF_IS_PENDING(pVM, VM_FF_HM_TO_R3_MASK)
2968	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_TO_R3_MASK))
2969	{
2970	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHmToR3FF);
2971	int rc = RT_UNLIKELY(VM_FF_IS_PENDING(pVM, VM_FF_PGM_NO_MEMORY)) ? VINF_EM_NO_MEMORY : VINF_EM_RAW_TO_R3;
2972	Log4(("hmR0SvmCheckForceFlags: HM_TO_R3 forcing us back to ring-3. rc=%d\n", rc));
2973	return rc;
2974	}
2975
2976	/* Pending VM request packets, such as hardware interrupts. */
2977	if ( VM_FF_IS_PENDING(pVM, VM_FF_REQUEST)
2978	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_REQUEST))
2979	{
2980	Log4(("hmR0SvmCheckForceFlags: Pending VM request forcing us back to ring-3\n"));
2981	return VINF_EM_PENDING_REQUEST;
2982	}
2983
2984	/* Pending PGM pool flushes. */
2985	if (VM_FF_IS_PENDING(pVM, VM_FF_PGM_POOL_FLUSH_PENDING))
2986	{
2987	Log4(("hmR0SvmCheckForceFlags: PGM pool flush pending forcing us back to ring-3\n"));
2988	return VINF_PGM_POOL_FLUSH_PENDING;
2989	}
2990
2991	/* Pending DMA requests. */
2992	if (VM_FF_IS_PENDING(pVM, VM_FF_PDM_DMA))
2993	{
2994	Log4(("hmR0SvmCheckForceFlags: Pending DMA request forcing us back to ring-3\n"));
2995	return VINF_EM_RAW_TO_R3;
2996	}
2997	}
2998
2999	return VINF_SUCCESS;
3000	}
3001
3002
3003	/**
3004	* Does the preparations before executing guest code in AMD-V.
3005	*
3006	* This may cause longjmps to ring-3 and may even result in rescheduling to the
3007	* recompiler. We must be cautious what we do here regarding committing
3008	* guest-state information into the VMCB assuming we assuredly execute the guest
3009	* in AMD-V. If we fall back to the recompiler after updating the VMCB and
3010	* clearing the common-state (TRPM/forceflags), we must undo those changes so
3011	* that the recompiler can (and should) use them when it resumes guest
3012	* execution. Otherwise such operations must be done when we can no longer
3013	* exit to ring-3.
3014	*
3015	* @returns VBox status code (informational status codes included).
3016	* @retval VINF_SUCCESS if we can proceed with running the guest.
3017	* @retval VINF_* scheduling changes, we have to go back to ring-3.
3018	*
3019	* @param pVM The cross context VM structure.
3020	* @param pVCpu The cross context virtual CPU structure.
3021	* @param pCtx Pointer to the guest-CPU context.
3022	* @param pSvmTransient Pointer to the SVM transient structure.
3023	*/
3024	static int hmR0SvmPreRunGuest(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
3025	{
3026	HMSVM_ASSERT_PREEMPT_SAFE();
3027
3028	/* Check force flag actions that might require us to go back to ring-3. */
3029	int rc = hmR0SvmCheckForceFlags(pVM, pVCpu, pCtx);
3030	if (rc != VINF_SUCCESS)
3031	return rc;
3032
3033	if (TRPMHasTrap(pVCpu))
3034	hmR0SvmTrpmTrapToPendingEvent(pVCpu);
3035	else if (!pVCpu->hm.s.Event.fPending)
3036	hmR0SvmEvaluatePendingEvent(pVCpu, pCtx);
3037
3038	/*
3039	* On the oldest AMD-V systems, we may not get enough information to reinject an NMI.
3040	* Just do it in software, see @bugref{8411}.
3041	* NB: If we could continue a task switch exit we wouldn't need to do this.
3042	*/
3043	if (RT_UNLIKELY(pVCpu->hm.s.Event.fPending && (((pVCpu->hm.s.Event.u64IntInfo >> 8) & 7) == SVM_EVENT_NMI)))
3044	if (RT_UNLIKELY(!pVM->hm.s.svm.u32Features))
3045	return VINF_EM_RAW_INJECT_TRPM_EVENT;
3046
3047	#ifdef HMSVM_SYNC_FULL_GUEST_STATE
3048	HMCPU_CF_SET(pVCpu, HM_CHANGED_ALL_GUEST);
3049	#endif
3050
3051	/* Load the guest bits that are not shared with the host in any way since we can longjmp or get preempted. */
3052	rc = hmR0SvmLoadGuestState(pVM, pVCpu, pCtx);
3053	AssertRCReturn(rc, rc);
3054	STAM_COUNTER_INC(&pVCpu->hm.s.StatLoadFull);
3055
3056	/*
3057	* If we're not intercepting TPR changes in the guest, save the guest TPR before the world-switch
3058	* so we can update it on the way back if the guest changed the TPR.
3059	*/
3060	if (pVCpu->hm.s.svm.fSyncVTpr)
3061	{
3062	if (pVM->hm.s.fTPRPatchingActive)
3063	pSvmTransient->u8GuestTpr = pCtx->msrLSTAR;
3064	else
3065	{
3066	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
3067	pSvmTransient->u8GuestTpr = pVmcb->ctrl.IntCtrl.n.u8VTPR;
3068	}
3069	}
3070
3071	/*
3072	* No longjmps to ring-3 from this point on!!!
3073	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional, better than a kernel panic.
3074	* This also disables flushing of the R0-logger instance (if any).
3075	*/
3076	VMMRZCallRing3Disable(pVCpu);
3077
3078	/*
3079	* We disable interrupts so that we don't miss any interrupts that would flag preemption (IPI/timers etc.)
3080	* when thread-context hooks aren't used and we've been running with preemption disabled for a while.
3081	*
3082	* We need to check for force-flags that could've possible been altered since we last checked them (e.g.
3083	* by PDMGetInterrupt() leaving the PDM critical section, see @bugref{6398}).
3084	*
3085	* We also check a couple of other force-flags as a last opportunity to get the EMT back to ring-3 before
3086	* executing guest code.
3087	*/
3088	pSvmTransient->fEFlags = ASMIntDisableFlags();
3089	if ( VM_FF_IS_PENDING(pVM, VM_FF_EMT_RENDEZVOUS \| VM_FF_TM_VIRTUAL_SYNC)
3090	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_TO_R3_MASK))
3091	{
3092	ASMSetFlags(pSvmTransient->fEFlags);
3093	VMMRZCallRing3Enable(pVCpu);
3094	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHmToR3FF);
3095	return VINF_EM_RAW_TO_R3;
3096	}
3097	if (RTThreadPreemptIsPending(NIL_RTTHREAD))
3098	{
3099	ASMSetFlags(pSvmTransient->fEFlags);
3100	VMMRZCallRing3Enable(pVCpu);
3101	STAM_COUNTER_INC(&pVCpu->hm.s.StatPendingHostIrq);
3102	return VINF_EM_RAW_INTERRUPT;
3103	}
3104
3105	/*
3106	* If we are injecting an NMI, we must set VMCPU_FF_BLOCK_NMIS only when we are going to execute
3107	* guest code for certain (no exits to ring-3). Otherwise, we could re-read the flag on re-entry into
3108	* AMD-V and conclude that NMI inhibition is active when we have not even delivered the NMI.
3109	*
3110	* With VT-x, this is handled by the Guest interruptibility information VMCS field which will set the
3111	* VMCS field after actually delivering the NMI which we read on VM-exit to determine the state.
3112	*/
3113	if (pVCpu->hm.s.Event.fPending)
3114	{
3115	SVMEVENT Event;
3116	Event.u = pVCpu->hm.s.Event.u64IntInfo;
3117	if ( Event.n.u1Valid
3118	&& Event.n.u3Type == SVM_EVENT_NMI
3119	&& Event.n.u8Vector == X86_XCPT_NMI
3120	&& !VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_BLOCK_NMIS))
3121	{
3122	VMCPU_FF_SET(pVCpu, VMCPU_FF_BLOCK_NMIS);
3123	}
3124	}
3125
3126	return VINF_SUCCESS;
3127	}
3128
3129
3130	/**
3131	* Prepares to run guest code in AMD-V and we've committed to doing so. This
3132	* means there is no backing out to ring-3 or anywhere else at this
3133	* point.
3134	*
3135	* @param pVM The cross context VM structure.
3136	* @param pVCpu The cross context virtual CPU structure.
3137	* @param pCtx Pointer to the guest-CPU context.
3138	* @param pSvmTransient Pointer to the SVM transient structure.
3139	*
3140	* @remarks Called with preemption disabled.
3141	* @remarks No-long-jump zone!!!
3142	*/
3143	static void hmR0SvmPreRunGuestCommitted(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
3144	{
3145	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
3146	Assert(VMMR0IsLogFlushDisabled(pVCpu));
3147	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
3148
3149	VMCPU_ASSERT_STATE(pVCpu, VMCPUSTATE_STARTED_HM);
3150	VMCPU_SET_STATE(pVCpu, VMCPUSTATE_STARTED_EXEC); /* Indicate the start of guest execution. */
3151
3152	hmR0SvmInjectPendingEvent(pVCpu, pCtx);
3153
3154	if ( pVCpu->hm.s.fPreloadGuestFpu
3155	&& !CPUMIsGuestFPUStateActive(pVCpu))
3156	{
3157	CPUMR0LoadGuestFPU(pVM, pVCpu); /* (Ignore rc, no need to set HM_CHANGED_HOST_CONTEXT for SVM.) */
3158	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR0);
3159	}
3160
3161	/* Load the state shared between host and guest (FPU, debug). */
3162	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
3163	if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_HOST_GUEST_SHARED_STATE))
3164	hmR0SvmLoadSharedState(pVCpu, pVmcb, pCtx);
3165	HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_HOST_CONTEXT); /* Preemption might set this, nothing to do on AMD-V. */
3166	AssertMsg(!HMCPU_CF_VALUE(pVCpu), ("fContextUseFlags=%#RX32\n", HMCPU_CF_VALUE(pVCpu)));
3167
3168	/* Setup TSC offsetting. */
3169	RTCPUID idCurrentCpu = hmR0GetCurrentCpu()->idCpu;
3170	if ( pSvmTransient->fUpdateTscOffsetting
3171	\|\| idCurrentCpu != pVCpu->hm.s.idLastCpu)
3172	{
3173	hmR0SvmUpdateTscOffsetting(pVM, pVCpu);
3174	pSvmTransient->fUpdateTscOffsetting = false;
3175	}
3176
3177	/* If we've migrating CPUs, mark the VMCB Clean bits as dirty. */
3178	if (idCurrentCpu != pVCpu->hm.s.idLastCpu)
3179	pVmcb->ctrl.u64VmcbCleanBits = 0;
3180
3181	/* Store status of the shared guest-host state at the time of VMRUN. */
3182	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS)
3183	if (CPUMIsGuestInLongModeEx(pCtx))
3184	{
3185	pSvmTransient->fWasGuestDebugStateActive = CPUMIsGuestDebugStateActivePending(pVCpu);
3186	pSvmTransient->fWasHyperDebugStateActive = CPUMIsHyperDebugStateActivePending(pVCpu);
3187	}
3188	else
3189	#endif
3190	{
3191	pSvmTransient->fWasGuestDebugStateActive = CPUMIsGuestDebugStateActive(pVCpu);
3192	pSvmTransient->fWasHyperDebugStateActive = CPUMIsHyperDebugStateActive(pVCpu);
3193	}
3194	pSvmTransient->fWasGuestFPUStateActive = CPUMIsGuestFPUStateActive(pVCpu);
3195
3196	/* Flush the appropriate tagged-TLB entries. */
3197	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, true); /* Used for TLB flushing, set this across the world switch. */
3198	hmR0SvmFlushTaggedTlb(pVCpu);
3199	Assert(hmR0GetCurrentCpu()->idCpu == pVCpu->hm.s.idLastCpu);
3200
3201	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatEntry, &pVCpu->hm.s.StatInGC, x);
3202
3203	TMNotifyStartOfExecution(pVCpu); /* Finally, notify TM to resume its clocks as we're about
3204	to start executing. */
3205
3206	/*
3207	* Save the current Host TSC_AUX and write the guest TSC_AUX to the host, so that
3208	* RDTSCPs (that don't cause exits) reads the guest MSR. See @bugref{3324}.
3209	*
3210	* This should be done -after- any RDTSCPs for obtaining the host timestamp (TM, STAM etc).
3211	*/
3212	if ( (pVM->hm.s.cpuid.u32AMDFeatureEDX & X86_CPUID_EXT_FEATURE_EDX_RDTSCP)
3213	&& !(pVmcb->ctrl.u64InterceptCtrl & SVM_CTRL_INTERCEPT_RDTSCP))
3214	{
3215	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_TSC_AUX, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
3216	pVCpu->hm.s.u64HostTscAux = ASMRdMsr(MSR_K8_TSC_AUX);
3217	uint64_t u64GuestTscAux = CPUMR0GetGuestTscAux(pVCpu);
3218	if (u64GuestTscAux != pVCpu->hm.s.u64HostTscAux)
3219	ASMWrMsr(MSR_K8_TSC_AUX, u64GuestTscAux);
3220	pSvmTransient->fRestoreTscAuxMsr = true;
3221	}
3222	else
3223	{
3224	hmR0SvmSetMsrPermission(pVCpu, MSR_K8_TSC_AUX, SVMMSREXIT_INTERCEPT_READ, SVMMSREXIT_INTERCEPT_WRITE);
3225	pSvmTransient->fRestoreTscAuxMsr = false;
3226	}
3227
3228	/* If VMCB Clean bits isn't supported by the CPU, simply mark all state-bits as dirty, indicating (re)load-from-VMCB. */
3229	if (!(pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_VMCB_CLEAN))
3230	pVmcb->ctrl.u64VmcbCleanBits = 0;
3231	}
3232
3233
3234	/**
3235	* Wrapper for running the guest code in AMD-V.
3236	*
3237	* @returns VBox strict status code.
3238	* @param pVM The cross context VM structure.
3239	* @param pVCpu The cross context virtual CPU structure.
3240	* @param pCtx Pointer to the guest-CPU context.
3241	*
3242	* @remarks No-long-jump zone!!!
3243	*/
3244	DECLINLINE(int) hmR0SvmRunGuest(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3245	{
3246	/*
3247	* 64-bit Windows uses XMM registers in the kernel as the Microsoft compiler expresses floating-point operations
3248	* using SSE instructions. Some XMM registers (XMM6-XMM15) are callee-saved and thus the need for this XMM wrapper.
3249	* Refer MSDN docs. "Configuring Programs for 64-bit / x64 Software Conventions / Register Usage" for details.
3250	*/
3251	#ifdef VBOX_WITH_KERNEL_USING_XMM
3252	return hmR0SVMRunWrapXMM(pVCpu->hm.s.svm.HCPhysVmcbHost, pVCpu->hm.s.svm.HCPhysVmcb, pCtx, pVM, pVCpu,
3253	pVCpu->hm.s.svm.pfnVMRun);
3254	#else
3255	return pVCpu->hm.s.svm.pfnVMRun(pVCpu->hm.s.svm.HCPhysVmcbHost, pVCpu->hm.s.svm.HCPhysVmcb, pCtx, pVM, pVCpu);
3256	#endif
3257	}
3258
3259
3260	/**
3261	* Performs some essential restoration of state after running guest code in
3262	* AMD-V.
3263	*
3264	* @param pVM The cross context VM structure.
3265	* @param pVCpu The cross context virtual CPU structure.
3266	* @param pMixedCtx Pointer to the guest-CPU context. The data maybe
3267	* out-of-sync. Make sure to update the required fields
3268	* before using them.
3269	* @param pSvmTransient Pointer to the SVM transient structure.
3270	* @param rcVMRun Return code of VMRUN.
3271	*
3272	* @remarks Called with interrupts disabled.
3273	* @remarks No-long-jump zone!!! This function will however re-enable longjmps
3274	* unconditionally when it is safe to do so.
3275	*/
3276	static void hmR0SvmPostRunGuest(PVM pVM, PVMCPU pVCpu, PCPUMCTX pMixedCtx, PSVMTRANSIENT pSvmTransient, int rcVMRun)
3277	{
3278	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
3279
3280	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, false); /* See HMInvalidatePageOnAllVCpus(): used for TLB flushing. */
3281	ASMAtomicIncU32(&pVCpu->hm.s.cWorldSwitchExits); /* Initialized in vmR3CreateUVM(): used for EMT poking. */
3282
3283	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
3284	pVmcb->ctrl.u64VmcbCleanBits = HMSVM_VMCB_CLEAN_ALL; /* Mark the VMCB-state cache as unmodified by VMM. */
3285
3286	/* TSC read must be done early for maximum accuracy. */
3287	if (!(pVmcb->ctrl.u64InterceptCtrl & SVM_CTRL_INTERCEPT_RDTSC))
3288	TMCpuTickSetLastSeen(pVCpu, ASMReadTSC() + pVmcb->ctrl.u64TSCOffset);
3289
3290	if (pSvmTransient->fRestoreTscAuxMsr)
3291	{
3292	uint64_t u64GuestTscAuxMsr = ASMRdMsr(MSR_K8_TSC_AUX);
3293	CPUMR0SetGuestTscAux(pVCpu, u64GuestTscAuxMsr);
3294	if (u64GuestTscAuxMsr != pVCpu->hm.s.u64HostTscAux)
3295	ASMWrMsr(MSR_K8_TSC_AUX, pVCpu->hm.s.u64HostTscAux);
3296	}
3297
3298	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatInGC, &pVCpu->hm.s.StatExit1, x);
3299	TMNotifyEndOfExecution(pVCpu); /* Notify TM that the guest is no longer running. */
3300	VMCPU_SET_STATE(pVCpu, VMCPUSTATE_STARTED_HM);
3301
3302	Assert(!(ASMGetFlags() & X86_EFL_IF));
3303	ASMSetFlags(pSvmTransient->fEFlags); /* Enable interrupts. */
3304	VMMRZCallRing3Enable(pVCpu); /* It is now safe to do longjmps to ring-3!!! */
3305
3306	/* If VMRUN failed, we can bail out early. This does -not- cover SVM_EXIT_INVALID. */
3307	if (RT_UNLIKELY(rcVMRun != VINF_SUCCESS))
3308	{
3309	Log4(("VMRUN failure: rcVMRun=%Rrc\n", rcVMRun));
3310	return;
3311	}
3312
3313	pSvmTransient->u64ExitCode = pVmcb->ctrl.u64ExitCode; /* Save the #VMEXIT reason. */
3314	HMCPU_EXIT_HISTORY_ADD(pVCpu, pVmcb->ctrl.u64ExitCode); /* Update the #VMEXIT history array. */
3315	pSvmTransient->fVectoringDoublePF = false; /* Vectoring double page-fault needs to be determined later. */
3316	pSvmTransient->fVectoringPF = false; /* Vectoring page-fault needs to be determined later. */
3317
3318	hmR0SvmSaveGuestState(pVCpu, pMixedCtx); /* Save the guest state from the VMCB to the guest-CPU context. */
3319
3320	if (RT_LIKELY(pSvmTransient->u64ExitCode != (uint64_t)SVM_EXIT_INVALID))
3321	{
3322	if (pVCpu->hm.s.svm.fSyncVTpr)
3323	{
3324	/* TPR patching (for 32-bit guests) uses LSTAR MSR for holding the TPR value, otherwise uses the VTPR. */
3325	if ( pVM->hm.s.fTPRPatchingActive
3326	&& (pMixedCtx->msrLSTAR & 0xff) != pSvmTransient->u8GuestTpr)
3327	{
3328	int rc = APICSetTpr(pVCpu, pMixedCtx->msrLSTAR & 0xff);
3329	AssertRC(rc);
3330	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
3331	}
3332	else if (pSvmTransient->u8GuestTpr != pVmcb->ctrl.IntCtrl.n.u8VTPR)
3333	{
3334	int rc = APICSetTpr(pVCpu, pVmcb->ctrl.IntCtrl.n.u8VTPR << 4);
3335	AssertRC(rc);
3336	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
3337	}
3338	}
3339	}
3340	}
3341
3342
3343	/**
3344	* Runs the guest code using AMD-V.
3345	*
3346	* @returns VBox status code.
3347	* @param pVM The cross context VM structure.
3348	* @param pVCpu The cross context virtual CPU structure.
3349	* @param pCtx Pointer to the guest-CPU context.
3350	*/
3351	static int hmR0SvmRunGuestCodeNormal(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3352	{
3353	SVMTRANSIENT SvmTransient;
3354	SvmTransient.fUpdateTscOffsetting = true;
3355	uint32_t cLoops = 0;
3356	int rc = VERR_INTERNAL_ERROR_5;
3357
3358	for (;; cLoops++)
3359	{
3360	Assert(!HMR0SuspendPending());
3361	HMSVM_ASSERT_CPU_SAFE();
3362
3363	/* Preparatory work for running guest code, this may force us to return
3364	to ring-3. This bugger disables interrupts on VINF_SUCCESS! */
3365	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatEntry, x);
3366	rc = hmR0SvmPreRunGuest(pVM, pVCpu, pCtx, &SvmTransient);
3367	if (rc != VINF_SUCCESS)
3368	break;
3369
3370	/*
3371	* No longjmps to ring-3 from this point on!!!
3372	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional, better than a kernel panic.
3373	* This also disables flushing of the R0-logger instance (if any).
3374	*/
3375	hmR0SvmPreRunGuestCommitted(pVM, pVCpu, pCtx, &SvmTransient);
3376	rc = hmR0SvmRunGuest(pVM, pVCpu, pCtx);
3377
3378	/* Restore any residual host-state and save any bits shared between host
3379	and guest into the guest-CPU state. Re-enables interrupts! */
3380	hmR0SvmPostRunGuest(pVM, pVCpu, pCtx, &SvmTransient, rc);
3381
3382	if (RT_UNLIKELY( rc != VINF_SUCCESS /* Check for VMRUN errors. */
3383	\|\| SvmTransient.u64ExitCode == (uint64_t)SVM_EXIT_INVALID)) /* Check for invalid guest-state errors. */
3384	{
3385	if (rc == VINF_SUCCESS)
3386	rc = VERR_SVM_INVALID_GUEST_STATE;
3387	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit1, x);
3388	hmR0SvmReportWorldSwitchError(pVM, pVCpu, rc, pCtx);
3389	break;
3390	}
3391
3392	/* Handle the #VMEXIT. */
3393	HMSVM_EXITCODE_STAM_COUNTER_INC(SvmTransient.u64ExitCode);
3394	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatExit1, &pVCpu->hm.s.StatExit2, x);
3395	VBOXVMM_R0_HMSVM_VMEXIT(pVCpu, pCtx, SvmTransient.u64ExitCode, (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb);
3396	rc = hmR0SvmHandleExit(pVCpu, pCtx, &SvmTransient);
3397	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit2, x);
3398	if (rc != VINF_SUCCESS)
3399	break;
3400	if (cLoops > pVM->hm.s.cMaxResumeLoops)
3401	{
3402	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchMaxResumeLoops);
3403	rc = VINF_EM_RAW_INTERRUPT;
3404	break;
3405	}
3406	}
3407
3408	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatEntry, x);
3409	return rc;
3410	}
3411
3412
3413	/**
3414	* Runs the guest code using AMD-V in single step mode.
3415	*
3416	* @returns VBox status code.
3417	* @param pVM The cross context VM structure.
3418	* @param pVCpu The cross context virtual CPU structure.
3419	* @param pCtx Pointer to the guest-CPU context.
3420	*/
3421	static int hmR0SvmRunGuestCodeStep(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3422	{
3423	SVMTRANSIENT SvmTransient;
3424	SvmTransient.fUpdateTscOffsetting = true;
3425	uint32_t cLoops = 0;
3426	int rc = VERR_INTERNAL_ERROR_5;
3427	uint16_t uCsStart = pCtx->cs.Sel;
3428	uint64_t uRipStart = pCtx->rip;
3429
3430	for (;; cLoops++)
3431	{
3432	Assert(!HMR0SuspendPending());
3433	AssertMsg(pVCpu->hm.s.idEnteredCpu == RTMpCpuId(),
3434	("Illegal migration! Entered on CPU %u Current %u cLoops=%u\n", (unsigned)pVCpu->hm.s.idEnteredCpu,
3435	(unsigned)RTMpCpuId(), cLoops));
3436
3437	/* Preparatory work for running guest code, this may force us to return
3438	to ring-3. This bugger disables interrupts on VINF_SUCCESS! */
3439	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatEntry, x);
3440	rc = hmR0SvmPreRunGuest(pVM, pVCpu, pCtx, &SvmTransient);
3441	if (rc != VINF_SUCCESS)
3442	break;
3443
3444	/*
3445	* No longjmps to ring-3 from this point on!!!
3446	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional, better than a kernel panic.
3447	* This also disables flushing of the R0-logger instance (if any).
3448	*/
3449	VMMRZCallRing3Disable(pVCpu);
3450	VMMRZCallRing3RemoveNotification(pVCpu);
3451	hmR0SvmPreRunGuestCommitted(pVM, pVCpu, pCtx, &SvmTransient);
3452
3453	rc = hmR0SvmRunGuest(pVM, pVCpu, pCtx);
3454
3455	/*
3456	* Restore any residual host-state and save any bits shared between host and guest into the guest-CPU state.
3457	* This will also re-enable longjmps to ring-3 when it has reached a safe point!!!
3458	*/
3459	hmR0SvmPostRunGuest(pVM, pVCpu, pCtx, &SvmTransient, rc);
3460	if (RT_UNLIKELY( rc != VINF_SUCCESS /* Check for VMRUN errors. */
3461	\|\| SvmTransient.u64ExitCode == (uint64_t)SVM_EXIT_INVALID)) /* Check for invalid guest-state errors. */
3462	{
3463	if (rc == VINF_SUCCESS)
3464	rc = VERR_SVM_INVALID_GUEST_STATE;
3465	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit1, x);
3466	hmR0SvmReportWorldSwitchError(pVM, pVCpu, rc, pCtx);
3467	return rc;
3468	}
3469
3470	/* Handle the #VMEXIT. */
3471	HMSVM_EXITCODE_STAM_COUNTER_INC(SvmTransient.u64ExitCode);
3472	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatExit1, &pVCpu->hm.s.StatExit2, x);
3473	VBOXVMM_R0_HMSVM_VMEXIT(pVCpu, pCtx, SvmTransient.u64ExitCode, (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb);
3474	rc = hmR0SvmHandleExit(pVCpu, pCtx, &SvmTransient);
3475	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit2, x);
3476	if (rc != VINF_SUCCESS)
3477	break;
3478	if (cLoops > pVM->hm.s.cMaxResumeLoops)
3479	{
3480	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchMaxResumeLoops);
3481	rc = VINF_EM_RAW_INTERRUPT;
3482	break;
3483	}
3484
3485	/*
3486	* Did the RIP change, if so, consider it a single step.
3487	* Otherwise, make sure one of the TFs gets set.
3488	*/
3489	if ( pCtx->rip != uRipStart
3490	\|\| pCtx->cs.Sel != uCsStart)
3491	{
3492	rc = VINF_EM_DBG_STEPPED;
3493	break;
3494	}
3495	pVCpu->hm.s.fContextUseFlags \|= HM_CHANGED_GUEST_DEBUG;
3496	}
3497
3498	/*
3499	* Clear the X86_EFL_TF if necessary.
3500	*/
3501	if (pVCpu->hm.s.fClearTrapFlag)
3502	{
3503	pVCpu->hm.s.fClearTrapFlag = false;
3504	pCtx->eflags.Bits.u1TF = 0;
3505	}
3506
3507	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatEntry, x);
3508	return rc;
3509	}
3510
3511
3512	/**
3513	* Runs the guest code using AMD-V.
3514	*
3515	* @returns Strict VBox status code.
3516	* @param pVM The cross context VM structure.
3517	* @param pVCpu The cross context virtual CPU structure.
3518	* @param pCtx Pointer to the guest-CPU context.
3519	*/
3520	VMMR0DECL(VBOXSTRICTRC) SVMR0RunGuestCode(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3521	{
3522	Assert(VMMRZCallRing3IsEnabled(pVCpu));
3523	HMSVM_ASSERT_PREEMPT_SAFE();
3524	VMMRZCallRing3SetNotification(pVCpu, hmR0SvmCallRing3Callback, pCtx);
3525
3526	int rc;
3527	if (!pVCpu->hm.s.fSingleInstruction)
3528	rc = hmR0SvmRunGuestCodeNormal(pVM, pVCpu, pCtx);
3529	else
3530	rc = hmR0SvmRunGuestCodeStep(pVM, pVCpu, pCtx);
3531
3532	if (rc == VERR_EM_INTERPRETER)
3533	rc = VINF_EM_RAW_EMULATE_INSTR;
3534	else if (rc == VINF_EM_RESET)
3535	rc = VINF_EM_TRIPLE_FAULT;
3536
3537	/* Prepare to return to ring-3. This will remove longjmp notifications. */
3538	hmR0SvmExitToRing3(pVM, pVCpu, pCtx, rc);
3539	Assert(!VMMRZCallRing3IsNotificationSet(pVCpu));
3540	return rc;
3541	}
3542
3543
3544	/**
3545	* Handles a \#VMEXIT (for all EXITCODE values except SVM_EXIT_INVALID).
3546	*
3547	* @returns VBox status code (informational status codes included).
3548	* @param pVCpu The cross context virtual CPU structure.
3549	* @param pCtx Pointer to the guest-CPU context.
3550	* @param pSvmTransient Pointer to the SVM transient structure.
3551	*/
3552	DECLINLINE(int) hmR0SvmHandleExit(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
3553	{
3554	Assert(pSvmTransient->u64ExitCode != (uint64_t)SVM_EXIT_INVALID);
3555	Assert(pSvmTransient->u64ExitCode <= SVM_EXIT_MAX);
3556
3557	/*
3558	* The ordering of the case labels is based on most-frequently-occurring #VMEXITs for most guests under
3559	* normal workloads (for some definition of "normal").
3560	*/
3561	uint32_t u32ExitCode = pSvmTransient->u64ExitCode;
3562	switch (pSvmTransient->u64ExitCode)
3563	{
3564	case SVM_EXIT_NPF:
3565	return hmR0SvmExitNestedPF(pVCpu, pCtx, pSvmTransient);
3566
3567	case SVM_EXIT_IOIO:
3568	return hmR0SvmExitIOInstr(pVCpu, pCtx, pSvmTransient);
3569
3570	case SVM_EXIT_RDTSC:
3571	return hmR0SvmExitRdtsc(pVCpu, pCtx, pSvmTransient);
3572
3573	case SVM_EXIT_RDTSCP:
3574	return hmR0SvmExitRdtscp(pVCpu, pCtx, pSvmTransient);
3575
3576	case SVM_EXIT_CPUID:
3577	return hmR0SvmExitCpuid(pVCpu, pCtx, pSvmTransient);
3578
3579	case SVM_EXIT_EXCEPTION_E: /* X86_XCPT_PF */
3580	return hmR0SvmExitXcptPF(pVCpu, pCtx, pSvmTransient);
3581
3582	case SVM_EXIT_EXCEPTION_7: /* X86_XCPT_NM */
3583	return hmR0SvmExitXcptNM(pVCpu, pCtx, pSvmTransient);
3584
3585	case SVM_EXIT_EXCEPTION_6: /* X86_XCPT_UD */
3586	return hmR0SvmExitXcptUD(pVCpu, pCtx, pSvmTransient);
3587
3588	case SVM_EXIT_EXCEPTION_10: /* X86_XCPT_MF */
3589	return hmR0SvmExitXcptMF(pVCpu, pCtx, pSvmTransient);
3590
3591	case SVM_EXIT_EXCEPTION_1: /* X86_XCPT_DB */
3592	return hmR0SvmExitXcptDB(pVCpu, pCtx, pSvmTransient);
3593
3594	case SVM_EXIT_EXCEPTION_11: /* X86_XCPT_AC */
3595	return hmR0SvmExitXcptAC(pVCpu, pCtx, pSvmTransient);
3596
3597	case SVM_EXIT_EXCEPTION_3: /* X86_XCPT_BP */
3598	return hmR0SvmExitXcptBP(pVCpu, pCtx, pSvmTransient);
3599
3600	case SVM_EXIT_MONITOR:
3601	return hmR0SvmExitMonitor(pVCpu, pCtx, pSvmTransient);
3602
3603	case SVM_EXIT_MWAIT:
3604	return hmR0SvmExitMwait(pVCpu, pCtx, pSvmTransient);
3605
3606	case SVM_EXIT_HLT:
3607	return hmR0SvmExitHlt(pVCpu, pCtx, pSvmTransient);
3608
3609	case SVM_EXIT_READ_CR0:
3610	case SVM_EXIT_READ_CR3:
3611	case SVM_EXIT_READ_CR4:
3612	return hmR0SvmExitReadCRx(pVCpu, pCtx, pSvmTransient);
3613
3614	case SVM_EXIT_WRITE_CR0:
3615	case SVM_EXIT_WRITE_CR3:
3616	case SVM_EXIT_WRITE_CR4:
3617	case SVM_EXIT_WRITE_CR8:
3618	return hmR0SvmExitWriteCRx(pVCpu, pCtx, pSvmTransient);
3619
3620	case SVM_EXIT_PAUSE:
3621	return hmR0SvmExitPause(pVCpu, pCtx, pSvmTransient);
3622
3623	case SVM_EXIT_VMMCALL:
3624	return hmR0SvmExitVmmCall(pVCpu, pCtx, pSvmTransient);
3625
3626	case SVM_EXIT_VINTR:
3627	return hmR0SvmExitVIntr(pVCpu, pCtx, pSvmTransient);
3628
3629	case SVM_EXIT_INTR:
3630	case SVM_EXIT_FERR_FREEZE:
3631	case SVM_EXIT_NMI:
3632	return hmR0SvmExitIntr(pVCpu, pCtx, pSvmTransient);
3633
3634	case SVM_EXIT_MSR:
3635	return hmR0SvmExitMsr(pVCpu, pCtx, pSvmTransient);
3636
3637	case SVM_EXIT_INVLPG:
3638	return hmR0SvmExitInvlpg(pVCpu, pCtx, pSvmTransient);
3639
3640	case SVM_EXIT_WBINVD:
3641	return hmR0SvmExitWbinvd(pVCpu, pCtx, pSvmTransient);
3642
3643	case SVM_EXIT_INVD:
3644	return hmR0SvmExitInvd(pVCpu, pCtx, pSvmTransient);
3645
3646	case SVM_EXIT_RDPMC:
3647	return hmR0SvmExitRdpmc(pVCpu, pCtx, pSvmTransient);
3648
3649	default:
3650	{
3651	switch (pSvmTransient->u64ExitCode)
3652	{
3653	case SVM_EXIT_READ_DR0: case SVM_EXIT_READ_DR1: case SVM_EXIT_READ_DR2: case SVM_EXIT_READ_DR3:
3654	case SVM_EXIT_READ_DR6: case SVM_EXIT_READ_DR7: case SVM_EXIT_READ_DR8: case SVM_EXIT_READ_DR9:
3655	case SVM_EXIT_READ_DR10: case SVM_EXIT_READ_DR11: case SVM_EXIT_READ_DR12: case SVM_EXIT_READ_DR13:
3656	case SVM_EXIT_READ_DR14: case SVM_EXIT_READ_DR15:
3657	return hmR0SvmExitReadDRx(pVCpu, pCtx, pSvmTransient);
3658
3659	case SVM_EXIT_WRITE_DR0: case SVM_EXIT_WRITE_DR1: case SVM_EXIT_WRITE_DR2: case SVM_EXIT_WRITE_DR3:
3660	case SVM_EXIT_WRITE_DR6: case SVM_EXIT_WRITE_DR7: case SVM_EXIT_WRITE_DR8: case SVM_EXIT_WRITE_DR9:
3661	case SVM_EXIT_WRITE_DR10: case SVM_EXIT_WRITE_DR11: case SVM_EXIT_WRITE_DR12: case SVM_EXIT_WRITE_DR13:
3662	case SVM_EXIT_WRITE_DR14: case SVM_EXIT_WRITE_DR15:
3663	return hmR0SvmExitWriteDRx(pVCpu, pCtx, pSvmTransient);
3664
3665	case SVM_EXIT_XSETBV:
3666	return hmR0SvmExitXsetbv(pVCpu, pCtx, pSvmTransient);
3667
3668	case SVM_EXIT_TASK_SWITCH:
3669	return hmR0SvmExitTaskSwitch(pVCpu, pCtx, pSvmTransient);
3670
3671	case SVM_EXIT_IRET:
3672	return hmR0SvmExitIret(pVCpu, pCtx, pSvmTransient);
3673
3674	case SVM_EXIT_SHUTDOWN:
3675	return hmR0SvmExitShutdown(pVCpu, pCtx, pSvmTransient);
3676
3677	case SVM_EXIT_SMI:
3678	case SVM_EXIT_INIT:
3679	{
3680	/*
3681	* We don't intercept NMIs. As for INIT signals, it really shouldn't ever happen here. If it ever does,
3682	* we want to know about it so log the exit code and bail.
3683	*/
3684	AssertMsgFailed(("hmR0SvmHandleExit: Unexpected exit %#RX32\n", (uint32_t)pSvmTransient->u64ExitCode));
3685	pVCpu->hm.s.u32HMError = (uint32_t)pSvmTransient->u64ExitCode;
3686	return VERR_SVM_UNEXPECTED_EXIT;
3687	}
3688
3689	#ifdef VBOX_WITH_NESTED_HWVIRT
3690	case SVM_EXIT_CLGI: return hmR0SvmExitClgi(pVCpu, pCtx, pSvmTransient);
3691	case SVM_EXIT_STGI: return hmR0SvmExitStgi(pVCpu, pCtx, pSvmTransient);
3692	#else
3693	case SVM_EXIT_CLGI:
3694	case SVM_EXIT_STGI:
3695	#endif
3696	case SVM_EXIT_INVLPGA:
3697	case SVM_EXIT_RSM:
3698	case SVM_EXIT_VMRUN:
3699	case SVM_EXIT_VMLOAD:
3700	case SVM_EXIT_VMSAVE:
3701	case SVM_EXIT_SKINIT:
3702	return hmR0SvmExitSetPendingXcptUD(pVCpu, pCtx, pSvmTransient);
3703
3704	#ifdef HMSVM_ALWAYS_TRAP_ALL_XCPTS
3705	case SVM_EXIT_EXCEPTION_0: /* X86_XCPT_DE */
3706	/* SVM_EXIT_EXCEPTION_1: / / X86_XCPT_DB - Handled above. */
3707	case SVM_EXIT_EXCEPTION_2: /* X86_XCPT_NMI */
3708	/* case SVM_EXIT_EXCEPTION_3: / / X86_XCPT_BP - Handled above. */
3709	case SVM_EXIT_EXCEPTION_4: /* X86_XCPT_OF */
3710	case SVM_EXIT_EXCEPTION_5: /* X86_XCPT_BR */
3711	/* case SVM_EXIT_EXCEPTION_6: / / X86_XCPT_UD - Handled above. */
3712	/* SVM_EXIT_EXCEPTION_7: / / X86_XCPT_NM - Handled above. */
3713	case SVM_EXIT_EXCEPTION_8: /* X86_XCPT_DF */
3714	case SVM_EXIT_EXCEPTION_9: /* X86_XCPT_CO_SEG_OVERRUN */
3715	case SVM_EXIT_EXCEPTION_A: /* X86_XCPT_TS */
3716	case SVM_EXIT_EXCEPTION_B: /* X86_XCPT_NP */
3717	case SVM_EXIT_EXCEPTION_C: /* X86_XCPT_SS */
3718	case SVM_EXIT_EXCEPTION_D: /* X86_XCPT_GP */
3719	/* SVM_EXIT_EXCEPTION_E: / / X86_XCPT_PF - Handled above. */
3720	/* SVM_EXIT_EXCEPTION_10: / / X86_XCPT_MF - Handled above. */
3721	/* SVM_EXIT_EXCEPTION_11: / / X86_XCPT_AC - Handled above. */
3722	case SVM_EXIT_EXCEPTION_12: /* X86_XCPT_MC */
3723	case SVM_EXIT_EXCEPTION_13: /* X86_XCPT_XF */
3724	case SVM_EXIT_EXCEPTION_F: /* Reserved */
3725	case SVM_EXIT_EXCEPTION_14: case SVM_EXIT_EXCEPTION_15: case SVM_EXIT_EXCEPTION_16:
3726	case SVM_EXIT_EXCEPTION_17: case SVM_EXIT_EXCEPTION_18: case SVM_EXIT_EXCEPTION_19:
3727	case SVM_EXIT_EXCEPTION_1A: case SVM_EXIT_EXCEPTION_1B: case SVM_EXIT_EXCEPTION_1C:
3728	case SVM_EXIT_EXCEPTION_1D: case SVM_EXIT_EXCEPTION_1E: case SVM_EXIT_EXCEPTION_1F:
3729	{
3730	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
3731	SVMEVENT Event;
3732	Event.u = 0;
3733	Event.n.u1Valid = 1;
3734	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3735	Event.n.u8Vector = pSvmTransient->u64ExitCode - SVM_EXIT_EXCEPTION_0;
3736
3737	switch (Event.n.u8Vector)
3738	{
3739	case X86_XCPT_DE:
3740	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestDE);
3741	break;
3742
3743	case X86_XCPT_NP:
3744	Event.n.u1ErrorCodeValid = 1;
3745	Event.n.u32ErrorCode = pVmcb->ctrl.u64ExitInfo1;
3746	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestNP);
3747	break;
3748
3749	case X86_XCPT_SS:
3750	Event.n.u1ErrorCodeValid = 1;
3751	Event.n.u32ErrorCode = pVmcb->ctrl.u64ExitInfo1;
3752	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestSS);
3753	break;
3754
3755	case X86_XCPT_GP:
3756	Event.n.u1ErrorCodeValid = 1;
3757	Event.n.u32ErrorCode = pVmcb->ctrl.u64ExitInfo1;
3758	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestGP);
3759	break;
3760
3761	default:
3762	AssertMsgFailed(("hmR0SvmHandleExit: Unexpected exit caused by exception %#x\n", Event.n.u8Vector));
3763	pVCpu->hm.s.u32HMError = Event.n.u8Vector;
3764	return VERR_SVM_UNEXPECTED_XCPT_EXIT;
3765	}
3766
3767	Log4(("#Xcpt: Vector=%#x at CS:RIP=%04x:%RGv\n", Event.n.u8Vector, pCtx->cs.Sel, (RTGCPTR)pCtx->rip));
3768	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3769	return VINF_SUCCESS;
3770	}
3771	#endif /* HMSVM_ALWAYS_TRAP_ALL_XCPTS */
3772
3773	default:
3774	{
3775	AssertMsgFailed(("hmR0SvmHandleExit: Unknown exit code %#x\n", u32ExitCode));
3776	pVCpu->hm.s.u32HMError = u32ExitCode;
3777	return VERR_SVM_UNKNOWN_EXIT;
3778	}
3779	}
3780	}
3781	}
3782	/* not reached */
3783	}
3784
3785
3786	#ifdef DEBUG
3787	/* Is there some generic IPRT define for this that are not in Runtime/internal/\* ?? */
3788	# define HMSVM_ASSERT_PREEMPT_CPUID_VAR() \
3789	RTCPUID const idAssertCpu = RTThreadPreemptIsEnabled(NIL_RTTHREAD) ? NIL_RTCPUID : RTMpCpuId()
3790
3791	# define HMSVM_ASSERT_PREEMPT_CPUID() \
3792	do \
3793	{ \
3794	RTCPUID const idAssertCpuNow = RTThreadPreemptIsEnabled(NIL_RTTHREAD) ? NIL_RTCPUID : RTMpCpuId(); \
3795	AssertMsg(idAssertCpu == idAssertCpuNow, ("SVM %#x, %#x\n", idAssertCpu, idAssertCpuNow)); \
3796	} while (0)
3797
3798	# define HMSVM_VALIDATE_EXIT_HANDLER_PARAMS() \
3799	do { \
3800	AssertPtr(pVCpu); \
3801	AssertPtr(pCtx); \
3802	AssertPtr(pSvmTransient); \
3803	Assert(ASMIntAreEnabled()); \
3804	HMSVM_ASSERT_PREEMPT_SAFE(); \
3805	HMSVM_ASSERT_PREEMPT_CPUID_VAR(); \
3806	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)); \
3807	HMSVM_ASSERT_PREEMPT_SAFE(); \
3808	if (VMMR0IsLogFlushDisabled(pVCpu)) \
3809	HMSVM_ASSERT_PREEMPT_CPUID(); \
3810	} while (0)
3811	#else /* Release builds */
3812	# define HMSVM_VALIDATE_EXIT_HANDLER_PARAMS() do { NOREF(pVCpu); NOREF(pCtx); NOREF(pSvmTransient); } while (0)
3813	#endif
3814
3815
3816	/**
3817	* Worker for hmR0SvmInterpretInvlpg().
3818	*
3819	* @return VBox status code.
3820	* @param pVCpu The cross context virtual CPU structure.
3821	* @param pCpu Pointer to the disassembler state.
3822	* @param pCtx The guest CPU context.
3823	*/
3824	static int hmR0SvmInterpretInvlPgEx(PVMCPU pVCpu, PDISCPUSTATE pCpu, PCPUMCTX pCtx)
3825	{
3826	DISQPVPARAMVAL Param1;
3827	RTGCPTR GCPtrPage;
3828
3829	int rc = DISQueryParamVal(CPUMCTX2CORE(pCtx), pCpu, &pCpu->Param1, &Param1, DISQPVWHICH_SRC);
3830	if (RT_FAILURE(rc))
3831	return VERR_EM_INTERPRETER;
3832
3833	if ( Param1.type == DISQPV_TYPE_IMMEDIATE
3834	\|\| Param1.type == DISQPV_TYPE_ADDRESS)
3835	{
3836	if (!(Param1.flags & (DISQPV_FLAG_32 \| DISQPV_FLAG_64)))
3837	return VERR_EM_INTERPRETER;
3838
3839	GCPtrPage = Param1.val.val64;
3840	VBOXSTRICTRC rc2 = EMInterpretInvlpg(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx), GCPtrPage);
3841	rc = VBOXSTRICTRC_VAL(rc2);
3842	}
3843	else
3844	{
3845	Log4(("hmR0SvmInterpretInvlPgEx invalid parameter type %#x\n", Param1.type));
3846	rc = VERR_EM_INTERPRETER;
3847	}
3848
3849	return rc;
3850	}
3851
3852
3853	/**
3854	* Interprets INVLPG.
3855	*
3856	* @returns VBox status code.
3857	* @retval VINF_* Scheduling instructions.
3858	* @retval VERR_EM_INTERPRETER Something we can't cope with.
3859	* @retval VERR_* Fatal errors.
3860	*
3861	* @param pVM The cross context VM structure.
3862	* @param pVCpu The cross context virtual CPU structure.
3863	* @param pCtx The guest CPU context.
3864	*
3865	* @remarks Updates the RIP if the instruction was executed successfully.
3866	*/
3867	static int hmR0SvmInterpretInvlpg(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3868	{
3869	/* Only allow 32 & 64 bit code. */
3870	if (CPUMGetGuestCodeBits(pVCpu) != 16)
3871	{
3872	PDISSTATE pDis = &pVCpu->hm.s.DisState;
3873	int rc = EMInterpretDisasCurrent(pVM, pVCpu, pDis, NULL /* pcbInstr */);
3874	if ( RT_SUCCESS(rc)
3875	&& pDis->pCurInstr->uOpcode == OP_INVLPG)
3876	{
3877	rc = hmR0SvmInterpretInvlPgEx(pVCpu, pDis, pCtx);
3878	if (RT_SUCCESS(rc))
3879	pCtx->rip += pDis->cbInstr;
3880	return rc;
3881	}
3882	else
3883	Log4(("hmR0SvmInterpretInvlpg: EMInterpretDisasCurrent returned %Rrc uOpCode=%#x\n", rc, pDis->pCurInstr->uOpcode));
3884	}
3885	return VERR_EM_INTERPRETER;
3886	}
3887
3888
3889	/**
3890	* Sets an invalid-opcode (\#UD) exception as pending-for-injection into the VM.
3891	*
3892	* @param pVCpu The cross context virtual CPU structure.
3893	*/
3894	DECLINLINE(void) hmR0SvmSetPendingXcptUD(PVMCPU pVCpu)
3895	{
3896	SVMEVENT Event;
3897	Event.u = 0;
3898	Event.n.u1Valid = 1;
3899	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3900	Event.n.u8Vector = X86_XCPT_UD;
3901	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3902	}
3903
3904
3905	/**
3906	* Sets a debug (\#DB) exception as pending-for-injection into the VM.
3907	*
3908	* @param pVCpu The cross context virtual CPU structure.
3909	*/
3910	DECLINLINE(void) hmR0SvmSetPendingXcptDB(PVMCPU pVCpu)
3911	{
3912	SVMEVENT Event;
3913	Event.u = 0;
3914	Event.n.u1Valid = 1;
3915	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3916	Event.n.u8Vector = X86_XCPT_DB;
3917	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3918	}
3919
3920
3921	/**
3922	* Sets a page fault (\#PF) exception as pending-for-injection into the VM.
3923	*
3924	* @param pVCpu The cross context virtual CPU structure.
3925	* @param pCtx Pointer to the guest-CPU context.
3926	* @param u32ErrCode The error-code for the page-fault.
3927	* @param uFaultAddress The page fault address (CR2).
3928	*
3929	* @remarks This updates the guest CR2 with @a uFaultAddress!
3930	*/
3931	DECLINLINE(void) hmR0SvmSetPendingXcptPF(PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t u32ErrCode, RTGCUINTPTR uFaultAddress)
3932	{
3933	SVMEVENT Event;
3934	Event.u = 0;
3935	Event.n.u1Valid = 1;
3936	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3937	Event.n.u8Vector = X86_XCPT_PF;
3938	Event.n.u1ErrorCodeValid = 1;
3939	Event.n.u32ErrorCode = u32ErrCode;
3940
3941	/* Update CR2 of the guest. */
3942	if (pCtx->cr2 != uFaultAddress)
3943	{
3944	pCtx->cr2 = uFaultAddress;
3945	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR2);
3946	}
3947
3948	hmR0SvmSetPendingEvent(pVCpu, &Event, uFaultAddress);
3949	}
3950
3951
3952	/**
3953	* Sets a device-not-available (\#NM) exception as pending-for-injection into
3954	* the VM.
3955	*
3956	* @param pVCpu The cross context virtual CPU structure.
3957	*/
3958	DECLINLINE(void) hmR0SvmSetPendingXcptNM(PVMCPU pVCpu)
3959	{
3960	SVMEVENT Event;
3961	Event.u = 0;
3962	Event.n.u1Valid = 1;
3963	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3964	Event.n.u8Vector = X86_XCPT_NM;
3965	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3966	}
3967
3968
3969	/**
3970	* Sets a math-fault (\#MF) exception as pending-for-injection into the VM.
3971	*
3972	* @param pVCpu The cross context virtual CPU structure.
3973	*/
3974	DECLINLINE(void) hmR0SvmSetPendingXcptMF(PVMCPU pVCpu)
3975	{
3976	SVMEVENT Event;
3977	Event.u = 0;
3978	Event.n.u1Valid = 1;
3979	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3980	Event.n.u8Vector = X86_XCPT_MF;
3981	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3982	}
3983
3984
3985	/**
3986	* Sets a double fault (\#DF) exception as pending-for-injection into the VM.
3987	*
3988	* @param pVCpu The cross context virtual CPU structure.
3989	*/
3990	DECLINLINE(void) hmR0SvmSetPendingXcptDF(PVMCPU pVCpu)
3991	{
3992	SVMEVENT Event;
3993	Event.u = 0;
3994	Event.n.u1Valid = 1;
3995	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3996	Event.n.u8Vector = X86_XCPT_DF;
3997	Event.n.u1ErrorCodeValid = 1;
3998	Event.n.u32ErrorCode = 0;
3999	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
4000	}
4001
4002
4003	/**
4004	* Emulates a simple MOV TPR (CR8) instruction, used for TPR patching on 32-bit
4005	* guests. This simply looks up the patch record at EIP and does the required.
4006	*
4007	* This VMMCALL is used a fallback mechanism when mov to/from cr8 isn't exactly
4008	* like how we want it to be (e.g. not followed by shr 4 as is usually done for
4009	* TPR). See hmR3ReplaceTprInstr() for the details.
4010	*
4011	* @returns VBox status code.
4012	* @retval VINF_SUCCESS if the access was handled successfully.
4013	* @retval VERR_NOT_FOUND if no patch record for this RIP could be found.
4014	* @retval VERR_SVM_UNEXPECTED_PATCH_TYPE if the found patch type is invalid.
4015	*
4016	* @param pVM The cross context VM structure.
4017	* @param pVCpu The cross context virtual CPU structure.
4018	* @param pCtx Pointer to the guest-CPU context.
4019	*/
4020	static int hmR0SvmEmulateMovTpr(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
4021	{
4022	Log4(("Emulated VMMCall TPR access replacement at RIP=%RGv\n", pCtx->rip));
4023
4024	/*
4025	* We do this in a loop as we increment the RIP after a successful emulation
4026	* and the new RIP may be a patched instruction which needs emulation as well.
4027	*/
4028	bool fPatchFound = false;
4029	for (;;)
4030	{
4031	bool fPending;
4032	uint8_t u8Tpr;
4033
4034	PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
4035	if (!pPatch)
4036	break;
4037
4038	fPatchFound = true;
4039	switch (pPatch->enmType)
4040	{
4041	case HMTPRINSTR_READ:
4042	{
4043	int rc = APICGetTpr(pVCpu, &u8Tpr, &fPending, NULL /* pu8PendingIrq */);
4044	AssertRC(rc);
4045
4046	rc = DISWriteReg32(CPUMCTX2CORE(pCtx), pPatch->uDstOperand, u8Tpr);
4047	AssertRC(rc);
4048	pCtx->rip += pPatch->cbOp;
4049	break;
4050	}
4051
4052	case HMTPRINSTR_WRITE_REG:
4053	case HMTPRINSTR_WRITE_IMM:
4054	{
4055	if (pPatch->enmType == HMTPRINSTR_WRITE_REG)
4056	{
4057	uint32_t u32Val;
4058	int rc = DISFetchReg32(CPUMCTX2CORE(pCtx), pPatch->uSrcOperand, &u32Val);
4059	AssertRC(rc);
4060	u8Tpr = u32Val;
4061	}
4062	else
4063	u8Tpr = (uint8_t)pPatch->uSrcOperand;
4064
4065	int rc2 = APICSetTpr(pVCpu, u8Tpr);
4066	AssertRC(rc2);
4067	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
4068
4069	pCtx->rip += pPatch->cbOp;
4070	break;
4071	}
4072
4073	default:
4074	AssertMsgFailed(("Unexpected patch type %d\n", pPatch->enmType));
4075	pVCpu->hm.s.u32HMError = pPatch->enmType;
4076	return VERR_SVM_UNEXPECTED_PATCH_TYPE;
4077	}
4078	}
4079
4080	if (fPatchFound)
4081	return VINF_SUCCESS;
4082	return VERR_NOT_FOUND;
4083	}
4084
4085
4086	/**
4087	* Determines if an exception is a contributory exception.
4088	*
4089	* Contributory exceptions are ones which can cause double-faults unless the
4090	* original exception was a benign exception. Page-fault is intentionally not
4091	* included here as it's a conditional contributory exception.
4092	*
4093	* @returns true if the exception is contributory, false otherwise.
4094	* @param uVector The exception vector.
4095	*/
4096	DECLINLINE(bool) hmR0SvmIsContributoryXcpt(const uint32_t uVector)
4097	{
4098	switch (uVector)
4099	{
4100	case X86_XCPT_GP:
4101	case X86_XCPT_SS:
4102	case X86_XCPT_NP:
4103	case X86_XCPT_TS:
4104	case X86_XCPT_DE:
4105	return true;
4106	default:
4107	break;
4108	}
4109	return false;
4110	}
4111
4112
4113	/**
4114	* Handle a condition that occurred while delivering an event through the guest
4115	* IDT.
4116	*
4117	* @returns VBox status code (informational error codes included).
4118	* @retval VINF_SUCCESS if we should continue handling the \#VMEXIT.
4119	* @retval VINF_HM_DOUBLE_FAULT if a \#DF condition was detected and we ought to
4120	* continue execution of the guest which will delivery the \#DF.
4121	* @retval VINF_EM_RESET if we detected a triple-fault condition.
4122	* @retval VERR_EM_GUEST_CPU_HANG if we detected a guest CPU hang.
4123	*
4124	* @param pVCpu The cross context virtual CPU structure.
4125	* @param pCtx Pointer to the guest-CPU context.
4126	* @param pSvmTransient Pointer to the SVM transient structure.
4127	*
4128	* @remarks No-long-jump zone!!!
4129	*/
4130	static int hmR0SvmCheckExitDueToEventDelivery(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4131	{
4132	int rc = VINF_SUCCESS;
4133	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4134
4135	Log4(("EXITINTINFO: Pending vectoring event %#RX64 Valid=%RTbool ErrValid=%RTbool Err=%#RX32 Type=%u Vector=%u\n",
4136	pVmcb->ctrl.ExitIntInfo.u, !!pVmcb->ctrl.ExitIntInfo.n.u1Valid, !!pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid,
4137	pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode, pVmcb->ctrl.ExitIntInfo.n.u3Type, pVmcb->ctrl.ExitIntInfo.n.u8Vector));
4138
4139	/* See AMD spec. 15.7.3 "EXITINFO Pseudo-Code". The EXITINTINFO (if valid) contains the prior exception (IDT vector)
4140	* that was trying to be delivered to the guest which caused a #VMEXIT which was intercepted (Exit vector). */
4141	if (pVmcb->ctrl.ExitIntInfo.n.u1Valid)
4142	{
4143	uint8_t uIdtVector = pVmcb->ctrl.ExitIntInfo.n.u8Vector;
4144
4145	typedef enum
4146	{
4147	SVMREFLECTXCPT_XCPT, /* Reflect the exception to the guest or for further evaluation by VMM. */
4148	SVMREFLECTXCPT_DF, /* Reflect the exception as a double-fault to the guest. */
4149	SVMREFLECTXCPT_TF, /* Indicate a triple faulted state to the VMM. */
4150	SVMREFLECTXCPT_HANG, /* Indicate bad VM trying to deadlock the CPU. */
4151	SVMREFLECTXCPT_NONE /* Nothing to reflect. */
4152	} SVMREFLECTXCPT;
4153
4154	SVMREFLECTXCPT enmReflect = SVMREFLECTXCPT_NONE;
4155	bool fReflectingNmi = false;
4156	if (pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_EXCEPTION)
4157	{
4158	if (pSvmTransient->u64ExitCode - SVM_EXIT_EXCEPTION_0 <= SVM_EXIT_EXCEPTION_1F)
4159	{
4160	uint8_t uExitVector = (uint8_t)(pSvmTransient->u64ExitCode - SVM_EXIT_EXCEPTION_0);
4161
4162	#ifdef VBOX_STRICT
4163	if ( hmR0SvmIsContributoryXcpt(uIdtVector)
4164	&& uExitVector == X86_XCPT_PF)
4165	{
4166	Log4(("IDT: Contributory #PF idCpu=%u uCR2=%#RX64\n", pVCpu->idCpu, pCtx->cr2));
4167	}
4168	#endif
4169
4170	if ( uIdtVector == X86_XCPT_BP
4171	\|\| uIdtVector == X86_XCPT_OF)
4172	{
4173	/* Ignore INT3/INTO, just re-execute. See @bugref{8357}. */
4174	}
4175	else if ( uExitVector == X86_XCPT_PF
4176	&& uIdtVector == X86_XCPT_PF)
4177	{
4178	pSvmTransient->fVectoringDoublePF = true;
4179	Log4(("IDT: Vectoring double #PF uCR2=%#RX64\n", pCtx->cr2));
4180	}
4181	else if ( uExitVector == X86_XCPT_AC
4182	&& uIdtVector == X86_XCPT_AC)
4183	{
4184	enmReflect = SVMREFLECTXCPT_HANG;
4185	Log4(("IDT: Nested #AC - Bad guest\n"));
4186	}
4187	else if ( (pVmcb->ctrl.u32InterceptException & HMSVM_CONTRIBUTORY_XCPT_MASK)
4188	&& hmR0SvmIsContributoryXcpt(uExitVector)
4189	&& ( hmR0SvmIsContributoryXcpt(uIdtVector)
4190	\|\| uIdtVector == X86_XCPT_PF))
4191	{
4192	enmReflect = SVMREFLECTXCPT_DF;
4193	Log4(("IDT: Pending vectoring #DF %#RX64 uIdtVector=%#x uExitVector=%#x\n", pVCpu->hm.s.Event.u64IntInfo,
4194	uIdtVector, uExitVector));
4195	}
4196	else if (uIdtVector == X86_XCPT_DF)
4197	{
4198	enmReflect = SVMREFLECTXCPT_TF;
4199	Log4(("IDT: Pending vectoring triple-fault %#RX64 uIdtVector=%#x uExitVector=%#x\n",
4200	pVCpu->hm.s.Event.u64IntInfo, uIdtVector, uExitVector));
4201	}
4202	else
4203	enmReflect = SVMREFLECTXCPT_XCPT;
4204	}
4205	else
4206	{
4207	/*
4208	* If event delivery caused an #VMEXIT that is not an exception (e.g. #NPF) then reflect the original
4209	* exception to the guest after handling the #VMEXIT.
4210	*/
4211	enmReflect = SVMREFLECTXCPT_XCPT;
4212	}
4213	}
4214	else if ( pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_EXTERNAL_IRQ
4215	\|\| pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_NMI)
4216	{
4217	enmReflect = SVMREFLECTXCPT_XCPT;
4218	fReflectingNmi = RT_BOOL(pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_NMI);
4219
4220	if (pSvmTransient->u64ExitCode - SVM_EXIT_EXCEPTION_0 <= SVM_EXIT_EXCEPTION_1F)
4221	{
4222	uint8_t uExitVector = (uint8_t)(pSvmTransient->u64ExitCode - SVM_EXIT_EXCEPTION_0);
4223	if (uExitVector == X86_XCPT_PF)
4224	{
4225	pSvmTransient->fVectoringPF = true;
4226	Log4(("IDT: Vectoring #PF due to Ext-Int/NMI. uCR2=%#RX64\n", pCtx->cr2));
4227	}
4228	}
4229	}
4230	/* else: Ignore software interrupts (INT n) as they reoccur when restarting the instruction. */
4231
4232	switch (enmReflect)
4233	{
4234	case SVMREFLECTXCPT_XCPT:
4235	{
4236	/* If we are re-injecting the NMI, clear NMI blocking. */
4237	if (fReflectingNmi)
4238	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_BLOCK_NMIS);
4239
4240	Assert(pVmcb->ctrl.ExitIntInfo.n.u3Type != SVM_EVENT_SOFTWARE_INT);
4241	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectPendingReflect);
4242	hmR0SvmSetPendingEvent(pVCpu, &pVmcb->ctrl.ExitIntInfo, 0 /* GCPtrFaultAddress */);
4243
4244	/* If uExitVector is #PF, CR2 value will be updated from the VMCB if it's a guest #PF. See hmR0SvmExitXcptPF(). */
4245	Log4(("IDT: Pending vectoring event %#RX64 ErrValid=%RTbool Err=%#RX32\n", pVmcb->ctrl.ExitIntInfo.u,
4246	!!pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid, pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode));
4247	break;
4248	}
4249
4250	case SVMREFLECTXCPT_DF:
4251	{
4252	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectPendingReflect);
4253	hmR0SvmSetPendingXcptDF(pVCpu);
4254	rc = VINF_HM_DOUBLE_FAULT;
4255	break;
4256	}
4257
4258	case SVMREFLECTXCPT_TF:
4259	{
4260	rc = VINF_EM_RESET;
4261	break;
4262	}
4263
4264	case SVMREFLECTXCPT_HANG:
4265	{
4266	rc = VERR_EM_GUEST_CPU_HANG;
4267	break;
4268	}
4269
4270	default:
4271	Assert(rc == VINF_SUCCESS);
4272	break;
4273	}
4274	}
4275	Assert(rc == VINF_SUCCESS \|\| rc == VINF_HM_DOUBLE_FAULT \|\| rc == VINF_EM_RESET \|\| rc == VERR_EM_GUEST_CPU_HANG);
4276	NOREF(pCtx);
4277	return rc;
4278	}
4279
4280	/**
4281	* Updates interrupt shadow for the current RIP.
4282	*/
4283	#define HMSVM_UPDATE_INTR_SHADOW(pVCpu, pCtx) \
4284	do { \
4285	/* Update interrupt shadow. */ \
4286	if ( VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS) \
4287	&& pCtx->rip != EMGetInhibitInterruptsPC(pVCpu)) \
4288	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS); \
4289	} while (0)
4290
4291	/**
4292	* Advances the guest RIP making use of the CPU's NRIP_SAVE feature if
4293	* supported, otherwise advances the RIP by the number of bytes specified in
4294	* @a cb.
4295	*
4296	* @param pVCpu The cross context virtual CPU structure.
4297	* @param pCtx Pointer to the guest-CPU context.
4298	* @param cb RIP increment value in bytes.
4299	*
4300	* @remarks Use this function only from \#VMEXIT's where the NRIP value is valid
4301	* when NRIP_SAVE is supported by the CPU, otherwise use
4302	* hmR0SvmAdvanceRipDumb!
4303	*/
4304	DECLINLINE(void) hmR0SvmAdvanceRipHwAssist(PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t cb)
4305	{
4306	if (pVCpu->CTX_SUFF(pVM)->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_NRIP_SAVE)
4307	{
4308	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4309	Assert(pVmcb->ctrl.u64NextRIP);
4310	Assert(pVmcb->ctrl.u64NextRIP - pCtx->rip == cb);
4311	pCtx->rip = pVmcb->ctrl.u64NextRIP;
4312	}
4313	else
4314	pCtx->rip += cb;
4315
4316	HMSVM_UPDATE_INTR_SHADOW(pVCpu, pCtx);
4317	}
4318
4319
4320	/**
4321	* Advances the guest RIP by the number of bytes specified in @a cb. This does
4322	* not make use of any hardware features to determine the instruction length.
4323	*
4324	* @param pVCpu The cross context virtual CPU structure.
4325	* @param pCtx Pointer to the guest-CPU context.
4326	* @param cb RIP increment value in bytes.
4327	*/
4328	DECLINLINE(void) hmR0SvmAdvanceRipDumb(PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t cb)
4329	{
4330	pCtx->rip += cb;
4331	HMSVM_UPDATE_INTR_SHADOW(pVCpu, pCtx);
4332	}
4333	#undef HMSVM_UPDATE_INTR_SHADOW
4334
4335
4336	/* -=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= */
4337	/* -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- #VMEXIT handlers -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- */
4338	/* -=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= */
4339
4340	/** @name \#VMEXIT handlers.
4341	* @{
4342	*/
4343
4344	/**
4345	* \#VMEXIT handler for external interrupts, NMIs, FPU assertion freeze and INIT
4346	* signals (SVM_EXIT_INTR, SVM_EXIT_NMI, SVM_EXIT_FERR_FREEZE, SVM_EXIT_INIT).
4347	*/
4348	HMSVM_EXIT_DECL hmR0SvmExitIntr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4349	{
4350	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4351
4352	if (pSvmTransient->u64ExitCode == SVM_EXIT_NMI)
4353	STAM_REL_COUNTER_INC(&pVCpu->hm.s.StatExitHostNmiInGC);
4354	else if (pSvmTransient->u64ExitCode == SVM_EXIT_INTR)
4355	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitExtInt);
4356
4357	/*
4358	* AMD-V has no preemption timer and the generic periodic preemption timer has no way to signal -before- the timer
4359	* fires if the current interrupt is our own timer or a some other host interrupt. We also cannot examine what
4360	* interrupt it is until the host actually take the interrupt.
4361	*
4362	* Going back to executing guest code here unconditionally causes random scheduling problems (observed on an
4363	* AMD Phenom 9850 Quad-Core on Windows 64-bit host).
4364	*/
4365	return VINF_EM_RAW_INTERRUPT;
4366	}
4367
4368
4369	/**
4370	* \#VMEXIT handler for WBINVD (SVM_EXIT_WBINVD). Conditional \#VMEXIT.
4371	*/
4372	HMSVM_EXIT_DECL hmR0SvmExitWbinvd(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4373	{
4374	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4375
4376	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 2);
4377	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitWbinvd);
4378	int rc = VINF_SUCCESS;
4379	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4380	return rc;
4381	}
4382
4383
4384	/**
4385	* \#VMEXIT handler for INVD (SVM_EXIT_INVD). Unconditional \#VMEXIT.
4386	*/
4387	HMSVM_EXIT_DECL hmR0SvmExitInvd(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4388	{
4389	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4390
4391	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 2);
4392	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitInvd);
4393	int rc = VINF_SUCCESS;
4394	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4395	return rc;
4396	}
4397
4398
4399	/**
4400	* \#VMEXIT handler for INVD (SVM_EXIT_CPUID). Conditional \#VMEXIT.
4401	*/
4402	HMSVM_EXIT_DECL hmR0SvmExitCpuid(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4403	{
4404	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4405	PVM pVM = pVCpu->CTX_SUFF(pVM);
4406	int rc = EMInterpretCpuId(pVM, pVCpu, CPUMCTX2CORE(pCtx));
4407	if (RT_LIKELY(rc == VINF_SUCCESS))
4408	{
4409	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 2);
4410	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4411	}
4412	else
4413	{
4414	AssertMsgFailed(("hmR0SvmExitCpuid: EMInterpretCpuId failed with %Rrc\n", rc));
4415	rc = VERR_EM_INTERPRETER;
4416	}
4417	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCpuid);
4418	return rc;
4419	}
4420
4421
4422	/**
4423	* \#VMEXIT handler for RDTSC (SVM_EXIT_RDTSC). Conditional \#VMEXIT.
4424	*/
4425	HMSVM_EXIT_DECL hmR0SvmExitRdtsc(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4426	{
4427	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4428	PVM pVM = pVCpu->CTX_SUFF(pVM);
4429	int rc = EMInterpretRdtsc(pVM, pVCpu, CPUMCTX2CORE(pCtx));
4430	if (RT_LIKELY(rc == VINF_SUCCESS))
4431	{
4432	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 2);
4433	pSvmTransient->fUpdateTscOffsetting = true;
4434
4435	/* Single step check. */
4436	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4437	}
4438	else
4439	{
4440	AssertMsgFailed(("hmR0SvmExitRdtsc: EMInterpretRdtsc failed with %Rrc\n", rc));
4441	rc = VERR_EM_INTERPRETER;
4442	}
4443	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdtsc);
4444	return rc;
4445	}
4446
4447
4448	/**
4449	* \#VMEXIT handler for RDTSCP (SVM_EXIT_RDTSCP). Conditional \#VMEXIT.
4450	*/
4451	HMSVM_EXIT_DECL hmR0SvmExitRdtscp(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4452	{
4453	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4454	int rc = EMInterpretRdtscp(pVCpu->CTX_SUFF(pVM), pVCpu, pCtx);
4455	if (RT_LIKELY(rc == VINF_SUCCESS))
4456	{
4457	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 3);
4458	pSvmTransient->fUpdateTscOffsetting = true;
4459	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4460	}
4461	else
4462	{
4463	AssertMsgFailed(("hmR0SvmExitRdtsc: EMInterpretRdtscp failed with %Rrc\n", rc));
4464	rc = VERR_EM_INTERPRETER;
4465	}
4466	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdtscp);
4467	return rc;
4468	}
4469
4470
4471	/**
4472	* \#VMEXIT handler for RDPMC (SVM_EXIT_RDPMC). Conditional \#VMEXIT.
4473	*/
4474	HMSVM_EXIT_DECL hmR0SvmExitRdpmc(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4475	{
4476	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4477	int rc = EMInterpretRdpmc(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
4478	if (RT_LIKELY(rc == VINF_SUCCESS))
4479	{
4480	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 2);
4481	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4482	}
4483	else
4484	{
4485	AssertMsgFailed(("hmR0SvmExitRdpmc: EMInterpretRdpmc failed with %Rrc\n", rc));
4486	rc = VERR_EM_INTERPRETER;
4487	}
4488	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdpmc);
4489	return rc;
4490	}
4491
4492
4493	/**
4494	* \#VMEXIT handler for INVLPG (SVM_EXIT_INVLPG). Conditional \#VMEXIT.
4495	*/
4496	HMSVM_EXIT_DECL hmR0SvmExitInvlpg(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4497	{
4498	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4499	PVM pVM = pVCpu->CTX_SUFF(pVM);
4500	Assert(!pVM->hm.s.fNestedPaging);
4501
4502	/** @todo Decode Assist. */
4503	int rc = hmR0SvmInterpretInvlpg(pVM, pVCpu, pCtx); /* Updates RIP if successful. */
4504	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitInvlpg);
4505	Assert(rc == VINF_SUCCESS \|\| rc == VERR_EM_INTERPRETER);
4506	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4507	return rc;
4508	}
4509
4510
4511	/**
4512	* \#VMEXIT handler for HLT (SVM_EXIT_HLT). Conditional \#VMEXIT.
4513	*/
4514	HMSVM_EXIT_DECL hmR0SvmExitHlt(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4515	{
4516	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4517
4518	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 1);
4519	int rc = EMShouldContinueAfterHalt(pVCpu, pCtx) ? VINF_SUCCESS : VINF_EM_HALT;
4520	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4521	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitHlt);
4522	if (rc != VINF_SUCCESS)
4523	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHltToR3);
4524	return rc;
4525	}
4526
4527
4528	/**
4529	* \#VMEXIT handler for MONITOR (SVM_EXIT_MONITOR). Conditional \#VMEXIT.
4530	*/
4531	HMSVM_EXIT_DECL hmR0SvmExitMonitor(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4532	{
4533	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4534	int rc = EMInterpretMonitor(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
4535	if (RT_LIKELY(rc == VINF_SUCCESS))
4536	{
4537	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 3);
4538	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4539	}
4540	else
4541	{
4542	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMonitor: EMInterpretMonitor failed with %Rrc\n", rc));
4543	rc = VERR_EM_INTERPRETER;
4544	}
4545	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitMonitor);
4546	return rc;
4547	}
4548
4549
4550	/**
4551	* \#VMEXIT handler for MWAIT (SVM_EXIT_MWAIT). Conditional \#VMEXIT.
4552	*/
4553	HMSVM_EXIT_DECL hmR0SvmExitMwait(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4554	{
4555	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4556	VBOXSTRICTRC rc2 = EMInterpretMWait(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
4557	int rc = VBOXSTRICTRC_VAL(rc2);
4558	if ( rc == VINF_EM_HALT
4559	\|\| rc == VINF_SUCCESS)
4560	{
4561	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 3);
4562
4563	if ( rc == VINF_EM_HALT
4564	&& EMMonitorWaitShouldContinue(pVCpu, pCtx))
4565	{
4566	rc = VINF_SUCCESS;
4567	}
4568	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4569	}
4570	else
4571	{
4572	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMwait: EMInterpretMWait failed with %Rrc\n", rc));
4573	rc = VERR_EM_INTERPRETER;
4574	}
4575	AssertMsg(rc == VINF_SUCCESS \|\| rc == VINF_EM_HALT \|\| rc == VERR_EM_INTERPRETER,
4576	("hmR0SvmExitMwait: EMInterpretMWait failed rc=%Rrc\n", rc));
4577	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitMwait);
4578	return rc;
4579	}
4580
4581
4582	/**
4583	* \#VMEXIT handler for shutdown (triple-fault) (SVM_EXIT_SHUTDOWN). Conditional
4584	* \#VMEXIT.
4585	*/
4586	HMSVM_EXIT_DECL hmR0SvmExitShutdown(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4587	{
4588	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4589	return VINF_EM_RESET;
4590	}
4591
4592
4593	/**
4594	* \#VMEXIT handler for CRx reads (SVM_EXIT_READ_CR*). Conditional \#VMEXIT.
4595	*/
4596	HMSVM_EXIT_DECL hmR0SvmExitReadCRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4597	{
4598	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4599
4600	Log4(("hmR0SvmExitReadCRx: CS:RIP=%04x:%#RX64\n", pCtx->cs.Sel, pCtx->rip));
4601
4602	/** @todo Decode Assist. */
4603	VBOXSTRICTRC rc2 = EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */);
4604	int rc = VBOXSTRICTRC_VAL(rc2);
4605	AssertMsg(rc == VINF_SUCCESS \|\| rc == VERR_EM_INTERPRETER \|\| rc == VINF_PGM_CHANGE_MODE \|\| rc == VINF_PGM_SYNC_CR3,
4606	("hmR0SvmExitReadCRx: EMInterpretInstruction failed rc=%Rrc\n", rc));
4607	Assert((pSvmTransient->u64ExitCode - SVM_EXIT_READ_CR0) <= 15);
4608	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCRxRead[pSvmTransient->u64ExitCode - SVM_EXIT_READ_CR0]);
4609	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4610	return rc;
4611	}
4612
4613
4614	/**
4615	* \#VMEXIT handler for CRx writes (SVM_EXIT_WRITE_CR*). Conditional \#VMEXIT.
4616	*/
4617	HMSVM_EXIT_DECL hmR0SvmExitWriteCRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4618	{
4619	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4620
4621	/** @todo Decode Assist. */
4622	VBOXSTRICTRC rcStrict = IEMExecOneBypassEx(pVCpu, CPUMCTX2CORE(pCtx), NULL);
4623	if (RT_UNLIKELY( rcStrict == VERR_IEM_ASPECT_NOT_IMPLEMENTED
4624	\|\| rcStrict == VERR_IEM_INSTR_NOT_IMPLEMENTED))
4625	rcStrict = VERR_EM_INTERPRETER;
4626	if (rcStrict == VINF_SUCCESS)
4627	{
4628	/* RIP has been updated by EMInterpretInstruction(). */
4629	Assert((pSvmTransient->u64ExitCode - SVM_EXIT_WRITE_CR0) <= 15);
4630	switch (pSvmTransient->u64ExitCode - SVM_EXIT_WRITE_CR0)
4631	{
4632	case 0: /* CR0. */
4633	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR0);
4634	break;
4635
4636	case 3: /* CR3. */
4637	Assert(!pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
4638	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR3);
4639	break;
4640
4641	case 4: /* CR4. */
4642	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR4);
4643	break;
4644
4645	case 8: /* CR8 (TPR). */
4646	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
4647	break;
4648
4649	default:
4650	AssertMsgFailed(("hmR0SvmExitWriteCRx: Invalid/Unexpected Write-CRx exit. u64ExitCode=%#RX64 %#x\n",
4651	pSvmTransient->u64ExitCode, pSvmTransient->u64ExitCode - SVM_EXIT_WRITE_CR0));
4652	break;
4653	}
4654	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
4655	}
4656	else
4657	Assert(rcStrict == VERR_EM_INTERPRETER \|\| rcStrict == VINF_PGM_CHANGE_MODE \|\| rcStrict == VINF_PGM_SYNC_CR3);
4658	return VBOXSTRICTRC_TODO(rcStrict);
4659	}
4660
4661
4662	/**
4663	* \#VMEXIT handler for instructions that result in a \#UD exception delivered
4664	* to the guest.
4665	*/
4666	HMSVM_EXIT_DECL hmR0SvmExitSetPendingXcptUD(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4667	{
4668	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4669	hmR0SvmSetPendingXcptUD(pVCpu);
4670	return VINF_SUCCESS;
4671	}
4672
4673
4674	/**
4675	* \#VMEXIT handler for MSR read and writes (SVM_EXIT_MSR). Conditional
4676	* \#VMEXIT.
4677	*/
4678	HMSVM_EXIT_DECL hmR0SvmExitMsr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4679	{
4680	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4681	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4682	PVM pVM = pVCpu->CTX_SUFF(pVM);
4683
4684	int rc;
4685	if (pVmcb->ctrl.u64ExitInfo1 == SVM_EXIT1_MSR_WRITE)
4686	{
4687	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitWrmsr);
4688
4689	/* Handle TPR patching; intercepted LSTAR write. */
4690	if ( pVM->hm.s.fTPRPatchingActive
4691	&& pCtx->ecx == MSR_K8_LSTAR)
4692	{
4693	if ((pCtx->eax & 0xff) != pSvmTransient->u8GuestTpr)
4694	{
4695	/* Our patch code uses LSTAR for TPR caching for 32-bit guests. */
4696	int rc2 = APICSetTpr(pVCpu, pCtx->eax & 0xff);
4697	AssertRC(rc2);
4698	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
4699	}
4700	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 2);
4701	rc = VINF_SUCCESS;
4702	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4703	return rc;
4704	}
4705
4706	if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_NRIP_SAVE)
4707	{
4708	rc = EMInterpretWrmsr(pVM, pVCpu, CPUMCTX2CORE(pCtx));
4709	if (RT_LIKELY(rc == VINF_SUCCESS))
4710	{
4711	pCtx->rip = pVmcb->ctrl.u64NextRIP;
4712	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4713	}
4714	else
4715	AssertMsg( rc == VERR_EM_INTERPRETER
4716	\|\| rc == VINF_CPUM_R3_MSR_WRITE, ("hmR0SvmExitMsr: EMInterpretWrmsr failed rc=%Rrc\n", rc));
4717	}
4718	else
4719	{
4720	rc = VBOXSTRICTRC_TODO(EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */));
4721	if (RT_LIKELY(rc == VINF_SUCCESS))
4722	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc); /* RIP updated by EMInterpretInstruction(). */
4723	else
4724	AssertMsg( rc == VERR_EM_INTERPRETER
4725	\|\| rc == VINF_CPUM_R3_MSR_WRITE, ("hmR0SvmExitMsr: WrMsr. EMInterpretInstruction failed rc=%Rrc\n", rc));
4726	}
4727
4728	if (rc == VINF_SUCCESS)
4729	{
4730	/* If this is an X2APIC WRMSR access, update the APIC state as well. */
4731	if ( pCtx->ecx >= MSR_IA32_X2APIC_START
4732	&& pCtx->ecx <= MSR_IA32_X2APIC_END)
4733	{
4734	/*
4735	* We've already saved the APIC related guest-state (TPR) in hmR0SvmPostRunGuest(). When full APIC register
4736	* virtualization is implemented we'll have to make sure APIC state is saved from the VMCB before
4737	* EMInterpretWrmsr() changes it.
4738	*/
4739	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
4740	}
4741	else if (pCtx->ecx == MSR_K6_EFER)
4742	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_EFER_MSR);
4743	else if (pCtx->ecx == MSR_IA32_TSC)
4744	pSvmTransient->fUpdateTscOffsetting = true;
4745	}
4746	}
4747	else
4748	{
4749	/* MSR Read access. */
4750	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdmsr);
4751	Assert(pVmcb->ctrl.u64ExitInfo1 == SVM_EXIT1_MSR_READ);
4752
4753	if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_NRIP_SAVE)
4754	{
4755	rc = EMInterpretRdmsr(pVM, pVCpu, CPUMCTX2CORE(pCtx));
4756	if (RT_LIKELY(rc == VINF_SUCCESS))
4757	{
4758	pCtx->rip = pVmcb->ctrl.u64NextRIP;
4759	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4760	}
4761	else
4762	AssertMsg( rc == VERR_EM_INTERPRETER
4763	\|\| rc == VINF_CPUM_R3_MSR_READ, ("hmR0SvmExitMsr: EMInterpretRdmsr failed rc=%Rrc\n", rc));
4764	}
4765	else
4766	{
4767	rc = VBOXSTRICTRC_TODO(EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0));
4768	if (RT_UNLIKELY(rc != VINF_SUCCESS))
4769	{
4770	AssertMsg( rc == VERR_EM_INTERPRETER
4771	\|\| rc == VINF_CPUM_R3_MSR_READ, ("hmR0SvmExitMsr: RdMsr. EMInterpretInstruction failed rc=%Rrc\n", rc));
4772	}
4773	/* RIP updated by EMInterpretInstruction(). */
4774	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4775	}
4776	}
4777
4778	/* RIP has been updated by EMInterpret[Rd\|Wr]msr(). */
4779	return rc;
4780	}
4781
4782
4783	/**
4784	* \#VMEXIT handler for DRx read (SVM_EXIT_READ_DRx). Conditional \#VMEXIT.
4785	*/
4786	HMSVM_EXIT_DECL hmR0SvmExitReadDRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4787	{
4788	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4789	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitDRxRead);
4790
4791	/* We should -not- get this #VMEXIT if the guest's debug registers were active. */
4792	if (pSvmTransient->fWasGuestDebugStateActive)
4793	{
4794	AssertMsgFailed(("hmR0SvmHandleExit: Unexpected exit %#RX32\n", (uint32_t)pSvmTransient->u64ExitCode));
4795	pVCpu->hm.s.u32HMError = (uint32_t)pSvmTransient->u64ExitCode;
4796	return VERR_SVM_UNEXPECTED_EXIT;
4797	}
4798
4799	/*
4800	* Lazy DR0-3 loading.
4801	*/
4802	if (!pSvmTransient->fWasHyperDebugStateActive)
4803	{
4804	Assert(!DBGFIsStepping(pVCpu)); Assert(!pVCpu->hm.s.fSingleInstruction);
4805	Log5(("hmR0SvmExitReadDRx: Lazy loading guest debug registers\n"));
4806
4807	/* Don't intercept DRx read and writes. */
4808	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4809	pVmcb->ctrl.u16InterceptRdDRx = 0;
4810	pVmcb->ctrl.u16InterceptWrDRx = 0;
4811	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
4812
4813	/* We're playing with the host CPU state here, make sure we don't preempt or longjmp. */
4814	VMMRZCallRing3Disable(pVCpu);
4815	HM_DISABLE_PREEMPT();
4816
4817	/* Save the host & load the guest debug state, restart execution of the MOV DRx instruction. */
4818	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
4819	Assert(CPUMIsGuestDebugStateActive(pVCpu) \|\| HC_ARCH_BITS == 32);
4820
4821	HM_RESTORE_PREEMPT();
4822	VMMRZCallRing3Enable(pVCpu);
4823
4824	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxContextSwitch);
4825	return VINF_SUCCESS;
4826	}
4827
4828	/*
4829	* Interpret the read/writing of DRx.
4830	*/
4831	/** @todo Decode assist. */
4832	VBOXSTRICTRC rc = EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */);
4833	Log5(("hmR0SvmExitReadDRx: Emulated DRx access: rc=%Rrc\n", VBOXSTRICTRC_VAL(rc)));
4834	if (RT_LIKELY(rc == VINF_SUCCESS))
4835	{
4836	/* Not necessary for read accesses but whatever doesn't hurt for now, will be fixed with decode assist. */
4837	/** @todo CPUM should set this flag! */
4838	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_DEBUG);
4839	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4840	}
4841	else
4842	Assert(rc == VERR_EM_INTERPRETER);
4843	return VBOXSTRICTRC_TODO(rc);
4844	}
4845
4846
4847	/**
4848	* \#VMEXIT handler for DRx write (SVM_EXIT_WRITE_DRx). Conditional \#VMEXIT.
4849	*/
4850	HMSVM_EXIT_DECL hmR0SvmExitWriteDRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4851	{
4852	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4853	/* For now it's the same since we interpret the instruction anyway. Will change when using of Decode Assist is implemented. */
4854	int rc = hmR0SvmExitReadDRx(pVCpu, pCtx, pSvmTransient);
4855	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitDRxWrite);
4856	STAM_COUNTER_DEC(&pVCpu->hm.s.StatExitDRxRead);
4857	return rc;
4858	}
4859
4860
4861	/**
4862	* \#VMEXIT handler for XCRx write (SVM_EXIT_XSETBV). Conditional \#VMEXIT.
4863	*/
4864	HMSVM_EXIT_DECL hmR0SvmExitXsetbv(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4865	{
4866	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4867
4868	/** @todo decode assists... */
4869	VBOXSTRICTRC rcStrict = IEMExecOne(pVCpu);
4870	if (rcStrict == VINF_IEM_RAISED_XCPT)
4871	HMCPU_CF_SET(pVCpu, HM_CHANGED_ALL_GUEST);
4872
4873	pVCpu->hm.s.fLoadSaveGuestXcr0 = (pCtx->cr4 & X86_CR4_OSXSAVE) && pCtx->aXcr[0] != ASMGetXcr0();
4874	Log4(("hmR0SvmExitXsetbv: New XCR0=%#RX64 fLoadSaveGuestXcr0=%d (cr4=%RX64) rcStrict=%Rrc\n",
4875	pCtx->aXcr[0], pVCpu->hm.s.fLoadSaveGuestXcr0, pCtx->cr4, VBOXSTRICTRC_VAL(rcStrict)));
4876
4877	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
4878	return VBOXSTRICTRC_TODO(rcStrict);
4879	}
4880
4881
4882	/**
4883	* \#VMEXIT handler for I/O instructions (SVM_EXIT_IOIO). Conditional \#VMEXIT.
4884	*/
4885	HMSVM_EXIT_DECL hmR0SvmExitIOInstr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4886	{
4887	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4888
4889	/* I/O operation lookup arrays. */
4890	static uint32_t const s_aIOSize[8] = { 0, 1, 2, 0, 4, 0, 0, 0 }; /* Size of the I/O accesses in bytes. */
4891	static uint32_t const s_aIOOpAnd[8] = { 0, 0xff, 0xffff, 0, 0xffffffff, 0, 0, 0 }; /* AND masks for saving
4892	the result (in AL/AX/EAX). */
4893	Log4(("hmR0SvmExitIOInstr: CS:RIP=%04x:%#RX64\n", pCtx->cs.Sel, pCtx->rip));
4894
4895	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4896	PVM pVM = pVCpu->CTX_SUFF(pVM);
4897
4898	/* Refer AMD spec. 15.10.2 "IN and OUT Behaviour" and Figure 15-2. "EXITINFO1 for IOIO Intercept" for the format. */
4899	SVMIOIOEXIT IoExitInfo;
4900	IoExitInfo.u = (uint32_t)pVmcb->ctrl.u64ExitInfo1;
4901	uint32_t uIOWidth = (IoExitInfo.u >> 4) & 0x7;
4902	uint32_t cbValue = s_aIOSize[uIOWidth];
4903	uint32_t uAndVal = s_aIOOpAnd[uIOWidth];
4904
4905	if (RT_UNLIKELY(!cbValue))
4906	{
4907	AssertMsgFailed(("hmR0SvmExitIOInstr: Invalid IO operation. uIOWidth=%u\n", uIOWidth));
4908	return VERR_EM_INTERPRETER;
4909	}
4910
4911	VBOXSTRICTRC rcStrict;
4912	bool fUpdateRipAlready = false;
4913	if (IoExitInfo.n.u1STR)
4914	{
4915	#ifdef VBOX_WITH_2ND_IEM_STEP
4916	/* INS/OUTS - I/O String instruction. */
4917	/** @todo Huh? why can't we use the segment prefix information given by AMD-V
4918	* in EXITINFO1? Investigate once this thing is up and running. */
4919	Log4(("CS:RIP=%04x:%08RX64 %#06x/%u %c str\n", pCtx->cs.Sel, pCtx->rip, IoExitInfo.n.u16Port, cbValue,
4920	IoExitInfo.n.u1Type == SVM_IOIO_WRITE ? 'w' : 'r'));
4921	AssertReturn(pCtx->dx == IoExitInfo.n.u16Port, VERR_SVM_IPE_2);
4922	static IEMMODE const s_aenmAddrMode[8] =
4923	{
4924	(IEMMODE)-1, IEMMODE_16BIT, IEMMODE_32BIT, (IEMMODE)-1, IEMMODE_64BIT, (IEMMODE)-1, (IEMMODE)-1, (IEMMODE)-1
4925	};
4926	IEMMODE enmAddrMode = s_aenmAddrMode[(IoExitInfo.u >> 7) & 0x7];
4927	if (enmAddrMode != (IEMMODE)-1)
4928	{
4929	uint64_t cbInstr = pVmcb->ctrl.u64ExitInfo2 - pCtx->rip;
4930	if (cbInstr <= 15 && cbInstr >= 1)
4931	{
4932	Assert(cbInstr >= 1U + IoExitInfo.n.u1REP);
4933	if (IoExitInfo.n.u1Type == SVM_IOIO_WRITE)
4934	{
4935	/* Don't know exactly how to detect whether u3SEG is valid, currently
4936	only enabling it for Bulldozer and later with NRIP. OS/2 broke on
4937	2384 Opterons when only checking NRIP. */
4938	if ( (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_NRIP_SAVE)
4939	&& pVM->cpum.ro.GuestFeatures.enmMicroarch >= kCpumMicroarch_AMD_15h_First)
4940	{
4941	AssertMsg(IoExitInfo.n.u3SEG == X86_SREG_DS \|\| cbInstr > 1U + IoExitInfo.n.u1REP,
4942	("u32Seg=%d cbInstr=%d u1REP=%d", IoExitInfo.n.u3SEG, cbInstr, IoExitInfo.n.u1REP));
4943	rcStrict = IEMExecStringIoWrite(pVCpu, cbValue, enmAddrMode, IoExitInfo.n.u1REP, (uint8_t)cbInstr,
4944	IoExitInfo.n.u3SEG, true /fIoChecked/);
4945	}
4946	else if (cbInstr == 1U + IoExitInfo.n.u1REP)
4947	rcStrict = IEMExecStringIoWrite(pVCpu, cbValue, enmAddrMode, IoExitInfo.n.u1REP, (uint8_t)cbInstr,
4948	X86_SREG_DS, true /fIoChecked/);
4949	else
4950	rcStrict = IEMExecOne(pVCpu);
4951	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOStringWrite);
4952	}
4953	else
4954	{
4955	AssertMsg(IoExitInfo.n.u3SEG == X86_SREG_ES /=0/, ("%#x\n", IoExitInfo.n.u3SEG));
4956	rcStrict = IEMExecStringIoRead(pVCpu, cbValue, enmAddrMode, IoExitInfo.n.u1REP, (uint8_t)cbInstr,
4957	true /fIoChecked/);
4958	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOStringRead);
4959	}
4960	}
4961	else
4962	{
4963	AssertMsgFailed(("rip=%RX64 nrip=%#RX64 cbInstr=%#RX64\n", pCtx->rip, pVmcb->ctrl.u64ExitInfo2, cbInstr));
4964	rcStrict = IEMExecOne(pVCpu);
4965	}
4966	}
4967	else
4968	{
4969	AssertMsgFailed(("IoExitInfo=%RX64\n", IoExitInfo.u));
4970	rcStrict = IEMExecOne(pVCpu);
4971	}
4972	fUpdateRipAlready = true;
4973
4974	#else
4975	/* INS/OUTS - I/O String instruction. */
4976	PDISCPUSTATE pDis = &pVCpu->hm.s.DisState;
4977
4978	/** @todo Huh? why can't we use the segment prefix information given by AMD-V
4979	* in EXITINFO1? Investigate once this thing is up and running. */
4980
4981	rcStrict = EMInterpretDisasCurrent(pVM, pVCpu, pDis, NULL);
4982	if (rcStrict == VINF_SUCCESS)
4983	{
4984	if (IoExitInfo.n.u1Type == SVM_IOIO_WRITE)
4985	{
4986	rcStrict = IOMInterpretOUTSEx(pVM, pVCpu, CPUMCTX2CORE(pCtx), IoExitInfo.n.u16Port, pDis->fPrefix,
4987	(DISCPUMODE)pDis->uAddrMode, cbValue);
4988	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOStringWrite);
4989	}
4990	else
4991	{
4992	rcStrict = IOMInterpretINSEx(pVM, pVCpu, CPUMCTX2CORE(pCtx), IoExitInfo.n.u16Port, pDis->fPrefix,
4993	(DISCPUMODE)pDis->uAddrMode, cbValue);
4994	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOStringRead);
4995	}
4996	}
4997	else
4998	rcStrict = VINF_EM_RAW_EMULATE_INSTR;
4999	#endif
5000	}
5001	else
5002	{
5003	/* IN/OUT - I/O instruction. */
5004	Assert(!IoExitInfo.n.u1REP);
5005
5006	if (IoExitInfo.n.u1Type == SVM_IOIO_WRITE)
5007	{
5008	rcStrict = IOMIOPortWrite(pVM, pVCpu, IoExitInfo.n.u16Port, pCtx->eax & uAndVal, cbValue);
5009	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOWrite);
5010	}
5011	else
5012	{
5013	uint32_t u32Val = 0;
5014	rcStrict = IOMIOPortRead(pVM, pVCpu, IoExitInfo.n.u16Port, &u32Val, cbValue);
5015	if (IOM_SUCCESS(rcStrict))
5016	{
5017	/* Save result of I/O IN instr. in AL/AX/EAX. */
5018	/** @todo r=bird: 32-bit op size should clear high bits of rax! */
5019	pCtx->eax = (pCtx->eax & ~uAndVal) \| (u32Val & uAndVal);
5020	}
5021	else if (rcStrict == VINF_IOM_R3_IOPORT_READ)
5022	HMR0SavePendingIOPortRead(pVCpu, pCtx->rip, pVmcb->ctrl.u64ExitInfo2, IoExitInfo.n.u16Port, uAndVal, cbValue);
5023
5024	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIORead);
5025	}
5026	}
5027
5028	if (IOM_SUCCESS(rcStrict))
5029	{
5030	/* AMD-V saves the RIP of the instruction following the IO instruction in EXITINFO2. */
5031	if (!fUpdateRipAlready)
5032	pCtx->rip = pVmcb->ctrl.u64ExitInfo2;
5033
5034	/*
5035	* If any I/O breakpoints are armed, we need to check if one triggered
5036	* and take appropriate action.
5037	* Note that the I/O breakpoint type is undefined if CR4.DE is 0.
5038	*/
5039	/** @todo Optimize away the DBGFBpIsHwIoArmed call by having DBGF tell the
5040	* execution engines about whether hyper BPs and such are pending. */
5041	uint32_t const uDr7 = pCtx->dr[7];
5042	if (RT_UNLIKELY( ( (uDr7 & X86_DR7_ENABLED_MASK)
5043	&& X86_DR7_ANY_RW_IO(uDr7)
5044	&& (pCtx->cr4 & X86_CR4_DE))
5045	\|\| DBGFBpIsHwIoArmed(pVM)))
5046	{
5047	/* We're playing with the host CPU state here, make sure we don't preempt or longjmp. */
5048	VMMRZCallRing3Disable(pVCpu);
5049	HM_DISABLE_PREEMPT();
5050
5051	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxIoCheck);
5052	CPUMR0DebugStateMaybeSaveGuest(pVCpu, false /fDr6/);
5053
5054	VBOXSTRICTRC rcStrict2 = DBGFBpCheckIo(pVM, pVCpu, pCtx, IoExitInfo.n.u16Port, cbValue);
5055	if (rcStrict2 == VINF_EM_RAW_GUEST_TRAP)
5056	{
5057	/* Raise #DB. */
5058	pVmcb->guest.u64DR6 = pCtx->dr[6];
5059	pVmcb->guest.u64DR7 = pCtx->dr[7];
5060	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
5061	hmR0SvmSetPendingXcptDB(pVCpu);
5062	}
5063	/* rcStrict is VINF_SUCCESS, VINF_IOM_R3_IOPORT_COMMIT_WRITE, or in [VINF_EM_FIRST..VINF_EM_LAST],
5064	however we can ditch VINF_IOM_R3_IOPORT_COMMIT_WRITE as it has VMCPU_FF_IOM as backup. */
5065	else if ( rcStrict2 != VINF_SUCCESS
5066	&& (rcStrict == VINF_SUCCESS \|\| rcStrict2 < rcStrict))
5067	rcStrict = rcStrict2;
5068	AssertCompile(VINF_EM_LAST < VINF_IOM_R3_IOPORT_COMMIT_WRITE);
5069
5070	HM_RESTORE_PREEMPT();
5071	VMMRZCallRing3Enable(pVCpu);
5072	}
5073
5074	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
5075	}
5076
5077	#ifdef VBOX_STRICT
5078	if (rcStrict == VINF_IOM_R3_IOPORT_READ)
5079	Assert(IoExitInfo.n.u1Type == SVM_IOIO_READ);
5080	else if (rcStrict == VINF_IOM_R3_IOPORT_WRITE \|\| rcStrict == VINF_IOM_R3_IOPORT_COMMIT_WRITE)
5081	Assert(IoExitInfo.n.u1Type == SVM_IOIO_WRITE);
5082	else
5083	{
5084	/** @todo r=bird: This is missing a bunch of VINF_EM_FIRST..VINF_EM_LAST
5085	* statuses, that the VMM device and some others may return. See
5086	* IOM_SUCCESS() for guidance. */
5087	AssertMsg( RT_FAILURE(rcStrict)
5088	\|\| rcStrict == VINF_SUCCESS
5089	\|\| rcStrict == VINF_EM_RAW_EMULATE_INSTR
5090	\|\| rcStrict == VINF_EM_DBG_BREAKPOINT
5091	\|\| rcStrict == VINF_EM_RAW_GUEST_TRAP
5092	\|\| rcStrict == VINF_EM_RAW_TO_R3
5093	\|\| rcStrict == VINF_TRPM_XCPT_DISPATCHED, ("%Rrc\n", VBOXSTRICTRC_VAL(rcStrict)));
5094	}
5095	#endif
5096	return VBOXSTRICTRC_TODO(rcStrict);
5097	}
5098
5099
5100	/**
5101	* \#VMEXIT handler for Nested Page-faults (SVM_EXIT_NPF). Conditional \#VMEXIT.
5102	*/
5103	HMSVM_EXIT_DECL hmR0SvmExitNestedPF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5104	{
5105	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5106	PVM pVM = pVCpu->CTX_SUFF(pVM);
5107	Assert(pVM->hm.s.fNestedPaging);
5108
5109	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
5110
5111	/* See AMD spec. 15.25.6 "Nested versus Guest Page Faults, Fault Ordering" for VMCB details for #NPF. */
5112	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
5113	uint32_t u32ErrCode = pVmcb->ctrl.u64ExitInfo1;
5114	RTGCPHYS GCPhysFaultAddr = pVmcb->ctrl.u64ExitInfo2;
5115
5116	Log4(("#NPF at CS:RIP=%04x:%#RX64 faultaddr=%RGp errcode=%#x \n", pCtx->cs.Sel, pCtx->rip, GCPhysFaultAddr, u32ErrCode));
5117
5118	#ifdef VBOX_HM_WITH_GUEST_PATCHING
5119	/* TPR patching for 32-bit guests, using the reserved bit in the page tables for MMIO regions. */
5120	if ( pVM->hm.s.fTprPatchingAllowed
5121	&& (GCPhysFaultAddr & PAGE_OFFSET_MASK) == XAPIC_OFF_TPR
5122	&& ( !(u32ErrCode & X86_TRAP_PF_P) /* Not present */
5123	\|\| (u32ErrCode & (X86_TRAP_PF_P \| X86_TRAP_PF_RSVD)) == (X86_TRAP_PF_P \| X86_TRAP_PF_RSVD)) /* MMIO page. */
5124	&& !CPUMIsGuestInLongModeEx(pCtx)
5125	&& !CPUMGetGuestCPL(pVCpu)
5126	&& pVM->hm.s.cPatches < RT_ELEMENTS(pVM->hm.s.aPatches))
5127	{
5128	RTGCPHYS GCPhysApicBase = APICGetBaseMsrNoCheck(pVCpu);
5129	GCPhysApicBase &= PAGE_BASE_GC_MASK;
5130
5131	if (GCPhysFaultAddr == GCPhysApicBase + XAPIC_OFF_TPR)
5132	{
5133	/* Only attempt to patch the instruction once. */
5134	PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
5135	if (!pPatch)
5136	return VINF_EM_HM_PATCH_TPR_INSTR;
5137	}
5138	}
5139	#endif
5140
5141	/*
5142	* Determine the nested paging mode.
5143	*/
5144	PGMMODE enmNestedPagingMode;
5145	#if HC_ARCH_BITS == 32
5146	if (CPUMIsGuestInLongModeEx(pCtx))
5147	enmNestedPagingMode = PGMMODE_AMD64_NX;
5148	else
5149	#endif
5150	enmNestedPagingMode = PGMGetHostMode(pVM);
5151
5152	/*
5153	* MMIO optimization using the reserved (RSVD) bit in the guest page tables for MMIO pages.
5154	*/
5155	int rc;
5156	Assert((u32ErrCode & (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P)) != X86_TRAP_PF_RSVD);
5157	if ((u32ErrCode & (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P)) == (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P))
5158	{
5159	/* If event delivery causes an MMIO #NPF, go back to instruction emulation as
5160	otherwise injecting the original pending event would most likely cause the same MMIO #NPF. */
5161	if (RT_UNLIKELY(pVCpu->hm.s.Event.fPending))
5162	return VERR_EM_INTERPRETER;
5163
5164	VBOXSTRICTRC rc2 = PGMR0Trap0eHandlerNPMisconfig(pVM, pVCpu, enmNestedPagingMode, CPUMCTX2CORE(pCtx), GCPhysFaultAddr,
5165	u32ErrCode);
5166	rc = VBOXSTRICTRC_VAL(rc2);
5167
5168	/*
5169	* If we succeed, resume guest execution.
5170	* If we fail in interpreting the instruction because we couldn't get the guest physical address
5171	* of the page containing the instruction via the guest's page tables (we would invalidate the guest page
5172	* in the host TLB), resume execution which would cause a guest page fault to let the guest handle this
5173	* weird case. See @bugref{6043}.
5174	*/
5175	if ( rc == VINF_SUCCESS
5176	\|\| rc == VERR_PAGE_TABLE_NOT_PRESENT
5177	\|\| rc == VERR_PAGE_NOT_PRESENT)
5178	{
5179	/* Successfully handled MMIO operation. */
5180	HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
5181	rc = VINF_SUCCESS;
5182	}
5183	return rc;
5184	}
5185
5186	TRPMAssertXcptPF(pVCpu, GCPhysFaultAddr, u32ErrCode);
5187	rc = PGMR0Trap0eHandlerNestedPaging(pVM, pVCpu, enmNestedPagingMode, u32ErrCode, CPUMCTX2CORE(pCtx), GCPhysFaultAddr);
5188	TRPMResetTrap(pVCpu);
5189
5190	Log4(("#NPF: PGMR0Trap0eHandlerNestedPaging returned %Rrc CS:RIP=%04x:%#RX64\n", rc, pCtx->cs.Sel, pCtx->rip));
5191
5192	/*
5193	* Same case as PGMR0Trap0eHandlerNPMisconfig(). See comment above, @bugref{6043}.
5194	*/
5195	if ( rc == VINF_SUCCESS
5196	\|\| rc == VERR_PAGE_TABLE_NOT_PRESENT
5197	\|\| rc == VERR_PAGE_NOT_PRESENT)
5198	{
5199	/* We've successfully synced our shadow page tables. */
5200	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPF);
5201	rc = VINF_SUCCESS;
5202	}
5203
5204	return rc;
5205	}
5206
5207
5208	/**
5209	* \#VMEXIT handler for virtual interrupt (SVM_EXIT_VINTR). Conditional
5210	* \#VMEXIT.
5211	*/
5212	HMSVM_EXIT_DECL hmR0SvmExitVIntr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5213	{
5214	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5215
5216	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
5217	pVmcb->ctrl.IntCtrl.n.u1VIrqValid = 0; /* No virtual interrupts pending, we'll inject the current one/NMI before reentry. */
5218	pVmcb->ctrl.IntCtrl.n.u8VIrqVector = 0;
5219
5220	/* Indicate that we no longer need to #VMEXIT when the guest is ready to receive interrupts/NMIs, it is now ready. */
5221	pVmcb->ctrl.u64InterceptCtrl &= ~SVM_CTRL_INTERCEPT_VINTR;
5222	pVmcb->ctrl.u64VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS \| HMSVM_VMCB_CLEAN_TPR);
5223
5224	/* Deliver the pending interrupt/NMI via hmR0SvmEvaluatePendingEvent() and resume guest execution. */
5225	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIntWindow);
5226	return VINF_SUCCESS;
5227	}
5228
5229
5230	/**
5231	* \#VMEXIT handler for task switches (SVM_EXIT_TASK_SWITCH). Conditional
5232	* \#VMEXIT.
5233	*/
5234	HMSVM_EXIT_DECL hmR0SvmExitTaskSwitch(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5235	{
5236	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5237
5238	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
5239
5240	#ifndef HMSVM_ALWAYS_TRAP_TASK_SWITCH
5241	Assert(!pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
5242	#endif
5243
5244	/* Check if this task-switch occurred while delivering an event through the guest IDT. */
5245	if (pVCpu->hm.s.Event.fPending) /* Can happen with exceptions/NMI. See @bugref{8411}. */
5246	{
5247	/*
5248	* AMD-V provides us with the exception which caused the TS; we collect
5249	* the information in the call to hmR0SvmCheckExitDueToEventDelivery.
5250	*/
5251	Log4(("hmR0SvmExitTaskSwitch: TS occurred during event delivery.\n"));
5252	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitTaskSwitch);
5253	return VINF_EM_RAW_INJECT_TRPM_EVENT;
5254	}
5255
5256	/** @todo Emulate task switch someday, currently just going back to ring-3 for
5257	* emulation. */
5258	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitTaskSwitch);
5259	return VERR_EM_INTERPRETER;
5260	}
5261
5262
5263	/**
5264	* \#VMEXIT handler for VMMCALL (SVM_EXIT_VMMCALL). Conditional \#VMEXIT.
5265	*/
5266	HMSVM_EXIT_DECL hmR0SvmExitVmmCall(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5267	{
5268	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5269	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitVmcall);
5270
5271	/* First check if this is a patched VMMCALL for mov TPR */
5272	int rc = hmR0SvmEmulateMovTpr(pVCpu->CTX_SUFF(pVM), pVCpu, pCtx);
5273	if (rc == VINF_SUCCESS)
5274	{
5275	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
5276	return VINF_SUCCESS;
5277	}
5278
5279	if (rc == VERR_NOT_FOUND)
5280	{
5281	if (pVCpu->hm.s.fHypercallsEnabled)
5282	{
5283	VBOXSTRICTRC rcStrict = GIMHypercall(pVCpu, pCtx);
5284	if (RT_SUCCESS(VBOXSTRICTRC_VAL(rcStrict)))
5285	{
5286	if (rcStrict == VINF_SUCCESS)
5287	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 3 /* cbInstr */);
5288	else
5289	Assert( rcStrict == VINF_GIM_HYPERCALL_CONTINUING
5290	\|\| rcStrict == VINF_GIM_R3_HYPERCALL);
5291
5292	/* If the hypercall changes anything other than guest's general-purpose registers,
5293	we would need to reload the guest changed bits here before VM-entry. */
5294	}
5295	rc = VBOXSTRICTRC_VAL(rcStrict);
5296	}
5297	else
5298	Log4(("hmR0SvmExitVmmCall: Hypercalls not enabled\n"));
5299	}
5300
5301	/* If hypercalls are disabled or the hypercall failed for some reason, raise #UD and continue. */
5302	if (RT_FAILURE(rc))
5303	{
5304	hmR0SvmSetPendingXcptUD(pVCpu);
5305	rc = VINF_SUCCESS;
5306	}
5307
5308	return rc;
5309	}
5310
5311
5312	/**
5313	* \#VMEXIT handler for VMMCALL (SVM_EXIT_VMMCALL). Conditional \#VMEXIT.
5314	*/
5315	HMSVM_EXIT_DECL hmR0SvmExitPause(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5316	{
5317	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5318	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitPause);
5319	return VINF_EM_RAW_INTERRUPT;
5320	}
5321
5322
5323	/**
5324	* \#VMEXIT handler for IRET (SVM_EXIT_IRET). Conditional \#VMEXIT.
5325	*/
5326	HMSVM_EXIT_DECL hmR0SvmExitIret(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5327	{
5328	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5329
5330	/* Clear NMI blocking. */
5331	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_BLOCK_NMIS);
5332
5333	/* Indicate that we no longer need to #VMEXIT when the guest is ready to receive NMIs, it is now ready. */
5334	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
5335	hmR0SvmClearIretIntercept(pVmcb);
5336
5337	/* Deliver the pending NMI via hmR0SvmEvaluatePendingEvent() and resume guest execution. */
5338	return VINF_SUCCESS;
5339	}
5340
5341
5342	/**
5343	* \#VMEXIT handler for page-fault exceptions (SVM_EXIT_EXCEPTION_E).
5344	* Conditional \#VMEXIT.
5345	*/
5346	HMSVM_EXIT_DECL hmR0SvmExitXcptPF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5347	{
5348	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5349
5350	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
5351
5352	/* See AMD spec. 15.12.15 "#PF (Page Fault)". */
5353	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
5354	uint32_t u32ErrCode = pVmcb->ctrl.u64ExitInfo1;
5355	RTGCUINTPTR uFaultAddress = pVmcb->ctrl.u64ExitInfo2;
5356	PVM pVM = pVCpu->CTX_SUFF(pVM);
5357
5358	#if defined(HMSVM_ALWAYS_TRAP_ALL_XCPTS) \|\| defined(HMSVM_ALWAYS_TRAP_PF)
5359	if (pVM->hm.s.fNestedPaging)
5360	{
5361	pVCpu->hm.s.Event.fPending = false; /* In case it's a contributory or vectoring #PF. */
5362	if (!pSvmTransient->fVectoringDoublePF)
5363	{
5364	/* A genuine guest #PF, reflect it to the guest. */
5365	hmR0SvmSetPendingXcptPF(pVCpu, pCtx, u32ErrCode, uFaultAddress);
5366	Log4(("#PF: Guest page fault at %04X:%RGv FaultAddr=%RGv ErrCode=%#x\n", pCtx->cs.Sel, (RTGCPTR)pCtx->rip,
5367	uFaultAddress, u32ErrCode));
5368	}
5369	else
5370	{
5371	/* A guest page-fault occurred during delivery of a page-fault. Inject #DF. */
5372	hmR0SvmSetPendingXcptDF(pVCpu);
5373	Log4(("Pending #DF due to vectoring #PF. NP\n"));
5374	}
5375	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestPF);
5376	return VINF_SUCCESS;
5377	}
5378	#endif
5379
5380	Assert(!pVM->hm.s.fNestedPaging);
5381
5382	#ifdef VBOX_HM_WITH_GUEST_PATCHING
5383	/* Shortcut for APIC TPR reads and writes; only applicable to 32-bit guests. */
5384	if ( pVM->hm.s.fTprPatchingAllowed
5385	&& (uFaultAddress & 0xfff) == XAPIC_OFF_TPR
5386	&& !(u32ErrCode & X86_TRAP_PF_P) /* Not present. */
5387	&& !CPUMIsGuestInLongModeEx(pCtx)
5388	&& !CPUMGetGuestCPL(pVCpu)
5389	&& pVM->hm.s.cPatches < RT_ELEMENTS(pVM->hm.s.aPatches))
5390	{
5391	RTGCPHYS GCPhysApicBase;
5392	GCPhysApicBase = APICGetBaseMsrNoCheck(pVCpu);
5393	GCPhysApicBase &= PAGE_BASE_GC_MASK;
5394
5395	/* Check if the page at the fault-address is the APIC base. */
5396	RTGCPHYS GCPhysPage;
5397	int rc2 = PGMGstGetPage(pVCpu, (RTGCPTR)uFaultAddress, NULL /* pfFlags */, &GCPhysPage);
5398	if ( rc2 == VINF_SUCCESS
5399	&& GCPhysPage == GCPhysApicBase)
5400	{
5401	/* Only attempt to patch the instruction once. */
5402	PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
5403	if (!pPatch)
5404	return VINF_EM_HM_PATCH_TPR_INSTR;
5405	}
5406	}
5407	#endif
5408
5409	Log4(("#PF: uFaultAddress=%#RX64 CS:RIP=%#04x:%#RX64 u32ErrCode %#RX32 cr3=%#RX64\n", uFaultAddress, pCtx->cs.Sel,
5410	pCtx->rip, u32ErrCode, pCtx->cr3));
5411
5412	/* If it's a vectoring #PF, emulate injecting the original event injection as PGMTrap0eHandler() is incapable
5413	of differentiating between instruction emulation and event injection that caused a #PF. See @bugref{6607}. */
5414	if (pSvmTransient->fVectoringPF)
5415	{
5416	Assert(pVCpu->hm.s.Event.fPending);
5417	return VINF_EM_RAW_INJECT_TRPM_EVENT;
5418	}
5419
5420	TRPMAssertXcptPF(pVCpu, uFaultAddress, u32ErrCode);
5421	int rc = PGMTrap0eHandler(pVCpu, u32ErrCode, CPUMCTX2CORE(pCtx), (RTGCPTR)uFaultAddress);
5422
5423	Log4(("#PF rc=%Rrc\n", rc));
5424
5425	if (rc == VINF_SUCCESS)
5426	{
5427	/* Successfully synced shadow pages tables or emulated an MMIO instruction. */
5428	TRPMResetTrap(pVCpu);
5429	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPF);
5430	HMCPU_CF_SET(pVCpu, HM_CHANGED_ALL_GUEST);
5431	return rc;
5432	}
5433	else if (rc == VINF_EM_RAW_GUEST_TRAP)
5434	{
5435	pVCpu->hm.s.Event.fPending = false; /* In case it's a contributory or vectoring #PF. */
5436
5437	if (!pSvmTransient->fVectoringDoublePF)
5438	{
5439	/* It's a guest page fault and needs to be reflected to the guest. */
5440	u32ErrCode = TRPMGetErrorCode(pVCpu); /* The error code might have been changed. */
5441	TRPMResetTrap(pVCpu);
5442	hmR0SvmSetPendingXcptPF(pVCpu, pCtx, u32ErrCode, uFaultAddress);
5443	}
5444	else
5445	{
5446	/* A guest page-fault occurred during delivery of a page-fault. Inject #DF. */
5447	TRPMResetTrap(pVCpu);
5448	hmR0SvmSetPendingXcptDF(pVCpu);
5449	Log4(("#PF: Pending #DF due to vectoring #PF\n"));
5450	}
5451
5452	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestPF);
5453	return VINF_SUCCESS;
5454	}
5455
5456	TRPMResetTrap(pVCpu);
5457	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPFEM);
5458	return rc;
5459	}
5460
5461
5462	/**
5463	* \#VMEXIT handler for device-not-available exceptions (SVM_EXIT_EXCEPTION_7).
5464	* Conditional \#VMEXIT.
5465	*/
5466	HMSVM_EXIT_DECL hmR0SvmExitXcptNM(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5467	{
5468	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5469
5470	/* Paranoia; Ensure we cannot be called as a result of event delivery. */
5471	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb; NOREF(pVmcb);
5472	Assert(!pVmcb->ctrl.ExitIntInfo.n.u1Valid);
5473
5474	/* We're playing with the host CPU state here, make sure we don't preempt or longjmp. */
5475	VMMRZCallRing3Disable(pVCpu);
5476	HM_DISABLE_PREEMPT();
5477
5478	int rc;
5479	/* If the guest FPU was active at the time of the #NM exit, then it's a guest fault. */
5480	if (pSvmTransient->fWasGuestFPUStateActive)
5481	{
5482	rc = VINF_EM_RAW_GUEST_TRAP;
5483	Assert(CPUMIsGuestFPUStateActive(pVCpu) \|\| HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR0));
5484	}
5485	else
5486	{
5487	#ifndef HMSVM_ALWAYS_TRAP_ALL_XCPTS
5488	Assert(!pSvmTransient->fWasGuestFPUStateActive);
5489	#endif
5490	rc = CPUMR0Trap07Handler(pVCpu->CTX_SUFF(pVM), pVCpu); /* (No need to set HM_CHANGED_HOST_CONTEXT for SVM.) */
5491	Assert( rc == VINF_EM_RAW_GUEST_TRAP
5492	\|\| ((rc == VINF_SUCCESS \|\| rc == VINF_CPUM_HOST_CR0_MODIFIED) && CPUMIsGuestFPUStateActive(pVCpu)));
5493	}
5494
5495	HM_RESTORE_PREEMPT();
5496	VMMRZCallRing3Enable(pVCpu);
5497
5498	if (rc == VINF_SUCCESS \|\| rc == VINF_CPUM_HOST_CR0_MODIFIED)
5499	{
5500	/* Guest FPU state was activated, we'll want to change CR0 FPU intercepts before the next VM-reentry. */
5501	HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR0);
5502	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowNM);
5503	pVCpu->hm.s.fPreloadGuestFpu = true;
5504	}
5505	else
5506	{
5507	/* Forward #NM to the guest. */
5508	Assert(rc == VINF_EM_RAW_GUEST_TRAP);
5509	hmR0SvmSetPendingXcptNM(pVCpu);
5510	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestNM);
5511	}
5512	return VINF_SUCCESS;
5513	}
5514
5515
5516	/**
5517	* \#VMEXIT handler for undefined opcode (SVM_EXIT_EXCEPTION_6).
5518	* Conditional \#VMEXIT.
5519	*/
5520	HMSVM_EXIT_DECL hmR0SvmExitXcptUD(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5521	{
5522	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5523
5524	/* Paranoia; Ensure we cannot be called as a result of event delivery. */
5525	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb; NOREF(pVmcb);
5526	Assert(!pVmcb->ctrl.ExitIntInfo.n.u1Valid);
5527
5528	int rc = VERR_SVM_UNEXPECTED_XCPT_EXIT;
5529	if (pVCpu->hm.s.fGIMTrapXcptUD)
5530	{
5531	uint8_t cbInstr = 0;
5532	VBOXSTRICTRC rcStrict = GIMXcptUD(pVCpu, pCtx, NULL /* pDis */, &cbInstr);
5533	if (rcStrict == VINF_SUCCESS)
5534	{
5535	/* #UD #VMEXIT does not have valid NRIP information, manually advance RIP. See @bugref{7270#c170}. */
5536	hmR0SvmAdvanceRipDumb(pVCpu, pCtx, cbInstr);
5537	rc = VINF_SUCCESS;
5538	}
5539	else if (rcStrict == VINF_GIM_HYPERCALL_CONTINUING)
5540	rc = VINF_SUCCESS;
5541	else if (rcStrict == VINF_GIM_R3_HYPERCALL)
5542	rc = VINF_GIM_R3_HYPERCALL;
5543	else
5544	Assert(RT_FAILURE(VBOXSTRICTRC_VAL(rcStrict)));
5545	}
5546
5547	/* If the GIM #UD exception handler didn't succeed for some reason or wasn't needed, raise #UD. */
5548	if (RT_FAILURE(rc))
5549	{
5550	hmR0SvmSetPendingXcptUD(pVCpu);
5551	rc = VINF_SUCCESS;
5552	}
5553
5554	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestUD);
5555	return rc;
5556	}
5557
5558
5559	/**
5560	* \#VMEXIT handler for math-fault exceptions (SVM_EXIT_EXCEPTION_10).
5561	* Conditional \#VMEXIT.
5562	*/
5563	HMSVM_EXIT_DECL hmR0SvmExitXcptMF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5564	{
5565	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5566
5567	/* Paranoia; Ensure we cannot be called as a result of event delivery. */
5568	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb; NOREF(pVmcb);
5569	Assert(!pVmcb->ctrl.ExitIntInfo.n.u1Valid);
5570
5571	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestMF);
5572
5573	if (!(pCtx->cr0 & X86_CR0_NE))
5574	{
5575	PVM pVM = pVCpu->CTX_SUFF(pVM);
5576	PDISSTATE pDis = &pVCpu->hm.s.DisState;
5577	unsigned cbOp;
5578	int rc = EMInterpretDisasCurrent(pVM, pVCpu, pDis, &cbOp);
5579	if (RT_SUCCESS(rc))
5580	{
5581	/* Convert a #MF into a FERR -> IRQ 13. See @bugref{6117}. */
5582	rc = PDMIsaSetIrq(pVCpu->CTX_SUFF(pVM), 13, 1, 0 /* uTagSrc */);
5583	if (RT_SUCCESS(rc))
5584	pCtx->rip += cbOp;
5585	}
5586	else
5587	Log4(("hmR0SvmExitXcptMF: EMInterpretDisasCurrent returned %Rrc uOpCode=%#x\n", rc, pDis->pCurInstr->uOpcode));
5588	return rc;
5589	}
5590
5591	hmR0SvmSetPendingXcptMF(pVCpu);
5592	return VINF_SUCCESS;
5593	}
5594
5595
5596	/**
5597	* \#VMEXIT handler for debug exceptions (SVM_EXIT_EXCEPTION_1). Conditional
5598	* \#VMEXIT.
5599	*/
5600	HMSVM_EXIT_DECL hmR0SvmExitXcptDB(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5601	{
5602	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5603
5604	/* If this #DB is the result of delivering an event, go back to the interpreter. */
5605	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
5606	if (RT_UNLIKELY(pVCpu->hm.s.Event.fPending))
5607	{
5608	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectPendingInterpret);
5609	return VERR_EM_INTERPRETER;
5610	}
5611
5612	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestDB);
5613
5614	/* This can be a fault-type #DB (instruction breakpoint) or a trap-type #DB (data breakpoint). However, for both cases
5615	DR6 and DR7 are updated to what the exception handler expects. See AMD spec. 15.12.2 "#DB (Debug)". */
5616	PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
5617	PVM pVM = pVCpu->CTX_SUFF(pVM);
5618	int rc = DBGFRZTrap01Handler(pVM, pVCpu, CPUMCTX2CORE(pCtx), pVmcb->guest.u64DR6, pVCpu->hm.s.fSingleInstruction);
5619	if (rc == VINF_EM_RAW_GUEST_TRAP)
5620	{
5621	Log5(("hmR0SvmExitXcptDB: DR6=%#RX64 -> guest trap\n", pVmcb->guest.u64DR6));
5622	if (CPUMIsHyperDebugStateActive(pVCpu))
5623	CPUMSetGuestDR6(pVCpu, CPUMGetGuestDR6(pVCpu) \| pVmcb->guest.u64DR6);
5624
5625	/* Reflect the exception back to the guest. */
5626	hmR0SvmSetPendingXcptDB(pVCpu);
5627	rc = VINF_SUCCESS;
5628	}
5629
5630	/*
5631	* Update DR6.
5632	*/
5633	if (CPUMIsHyperDebugStateActive(pVCpu))
5634	{
5635	Log5(("hmR0SvmExitXcptDB: DR6=%#RX64 -> %Rrc\n", pVmcb->guest.u64DR6, rc));
5636	pVmcb->guest.u64DR6 = X86_DR6_INIT_VAL;
5637	pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
5638	}
5639	else
5640	{
5641	AssertMsg(rc == VINF_SUCCESS, ("rc=%Rrc\n", rc));
5642	Assert(!pVCpu->hm.s.fSingleInstruction && !DBGFIsStepping(pVCpu));
5643	}
5644
5645	return rc;
5646	}
5647
5648
5649	/**
5650	* \#VMEXIT handler for alignment check exceptions (SVM_EXIT_EXCEPTION_11).
5651	* Conditional \#VMEXIT.
5652	*/
5653	HMSVM_EXIT_DECL hmR0SvmExitXcptAC(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5654	{
5655	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5656
5657	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
5658
5659	SVMEVENT Event;
5660	Event.u = 0;
5661	Event.n.u1Valid = 1;
5662	Event.n.u3Type = SVM_EVENT_EXCEPTION;
5663	Event.n.u8Vector = X86_XCPT_AC;
5664	Event.n.u1ErrorCodeValid = 1;
5665	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
5666	return VINF_SUCCESS;
5667	}
5668
5669
5670	/**
5671	* \#VMEXIT handler for breakpoint exceptions (SVM_EXIT_EXCEPTION_3).
5672	* Conditional \#VMEXIT.
5673	*/
5674	HMSVM_EXIT_DECL hmR0SvmExitXcptBP(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5675	{
5676	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5677
5678	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
5679
5680	int rc = DBGFRZTrap03Handler(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
5681	if (rc == VINF_EM_RAW_GUEST_TRAP)
5682	{
5683	SVMEVENT Event;
5684	Event.u = 0;
5685	Event.n.u1Valid = 1;
5686	Event.n.u3Type = SVM_EVENT_EXCEPTION;
5687	Event.n.u8Vector = X86_XCPT_BP;
5688	Event.n.u1ErrorCodeValid = 0;
5689	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
5690	}
5691
5692	Assert(rc == VINF_SUCCESS \|\| rc == VINF_EM_RAW_GUEST_TRAP \|\| rc == VINF_EM_DBG_BREAKPOINT);
5693	return rc;
5694	}
5695
5696
5697	#ifdef VBOX_WITH_NESTED_HWVIRT
5698	/**
5699	* \#VMEXIT handler for RDPMC (SVM_EXIT_CLGI). Conditional
5700	* \#VMEXIT.
5701	*/
5702	HMSVM_EXIT_DECL hmR0SvmExitClgi(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5703	{
5704	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5705	PVM pVM = pVCpu->CTX_SUFF(pVM);
5706	if (pVM->cpum.ro.GuestFeatures.fSvm)
5707	{
5708	/** @todo Stat. */
5709	/* STAM_COUNTER_INC(&pVCpu->hm.s.StatExitClgi); */
5710	VBOXSTRICTRC rcStrict = IEMExecDecodedClgi(pVCpu, 3);
5711	return VBOXSTRICTRC_VAL(rcStrict);
5712	}
5713	return hmR0SvmExitXcptUD(pVCpu, pCtx, pSvmTransient);
5714	}
5715
5716
5717	/**
5718	* \#VMEXIT handler for RDPMC (SVM_EXIT_STGI). Conditional
5719	* \#VMEXIT.
5720	*/
5721	HMSVM_EXIT_DECL hmR0SvmExitStgi(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5722	{
5723	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5724	PVM pVM = pVCpu->CTX_SUFF(pVM);
5725	if (pVM->cpum.ro.GuestFeatures.fSvm)
5726	{
5727	/** @todo Stat. */
5728	/* STAM_COUNTER_INC(&pVCpu->hm.s.StatExitStgi); */
5729	VBOXSTRICTRC rcStrict = IEMExecDecodedStgi(pVCpu, 3);
5730	return VBOXSTRICTRC_VAL(rcStrict);
5731	}
5732	return hmR0SvmExitXcptUD(pVCpu, pCtx, pSvmTransient);
5733	}
5734	#endif /* VBOX_WITH_NESTED_HWVIRT */
5735
5736
5737	/** @} */
5738

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

source: vbox/trunk/src/VBox/VMM/VMMR0/HMSVMR0.cpp@ 65919

Download in other formats: