HMSVMR0.cpp@ 84948

Last change on this file since 84948 was 83485, checked in by vboxsync, 5 years ago
SVM: Emulate intercepted IRET to avoid injecting a pending NMI too early (see bugref:6208).
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 312.1 KB

Line
1	/* $Id: HMSVMR0.cpp 83485 2020-03-30 14:37:45Z vboxsync $ */
2	/** @file
3	* HM SVM (AMD-V) - Host Context Ring-0.
4	*/
5
6	/*
7	* Copyright (C) 2013-2020 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	#define VMCPU_INCL_CPUM_GST_CTX
24	#include <iprt/asm-amd64-x86.h>
25	#include <iprt/thread.h>
26
27	#include <VBox/vmm/pdmapi.h>
28	#include <VBox/vmm/dbgf.h>
29	#include <VBox/vmm/iem.h>
30	#include <VBox/vmm/iom.h>
31	#include <VBox/vmm/tm.h>
32	#include <VBox/vmm/em.h>
33	#include <VBox/vmm/gim.h>
34	#include <VBox/vmm/apic.h>
35	#include "HMInternal.h"
36	#include <VBox/vmm/vmcc.h>
37	#include <VBox/err.h>
38	#include "HMSVMR0.h"
39	#include "dtrace/VBoxVMM.h"
40
41	#ifdef DEBUG_ramshankar
42	# define HMSVM_SYNC_FULL_GUEST_STATE
43	# define HMSVM_ALWAYS_TRAP_ALL_XCPTS
44	# define HMSVM_ALWAYS_TRAP_PF
45	# define HMSVM_ALWAYS_TRAP_TASK_SWITCH
46	#endif
47
48
49	/*********************************************************************************************************************************
50	* Defined Constants And Macros *
51	*********************************************************************************************************************************/
52	#ifdef VBOX_WITH_STATISTICS
53	# define HMSVM_EXITCODE_STAM_COUNTER_INC(u64ExitCode) do { \
54	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitAll); \
55	if ((u64ExitCode) == SVM_EXIT_NPF) \
56	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitReasonNpf); \
57	else \
58	STAM_COUNTER_INC(&pVCpu->hm.s.paStatExitReasonR0[(u64ExitCode) & MASK_EXITREASON_STAT]); \
59	} while (0)
60
61	# ifdef VBOX_WITH_NESTED_HWVIRT_SVM
62	# define HMSVM_NESTED_EXITCODE_STAM_COUNTER_INC(u64ExitCode) do { \
63	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitAll); \
64	STAM_COUNTER_INC(&pVCpu->hm.s.StatNestedExitAll); \
65	if ((u64ExitCode) == SVM_EXIT_NPF) \
66	STAM_COUNTER_INC(&pVCpu->hm.s.StatNestedExitReasonNpf); \
67	else \
68	STAM_COUNTER_INC(&pVCpu->hm.s.paStatNestedExitReasonR0[(u64ExitCode) & MASK_EXITREASON_STAT]); \
69	} while (0)
70	# endif
71	#else
72	# define HMSVM_EXITCODE_STAM_COUNTER_INC(u64ExitCode) do { } while (0)
73	# ifdef VBOX_WITH_NESTED_HWVIRT_SVM
74	# define HMSVM_NESTED_EXITCODE_STAM_COUNTER_INC(u64ExitCode) do { } while (0)
75	# endif
76	#endif /* !VBOX_WITH_STATISTICS */
77
78	/** If we decide to use a function table approach this can be useful to
79	* switch to a "static DECLCALLBACK(int)". */
80	#define HMSVM_EXIT_DECL static int
81
82	/**
83	* Subset of the guest-CPU state that is kept by SVM R0 code while executing the
84	* guest using hardware-assisted SVM.
85	*
86	* This excludes state like TSC AUX, GPRs (other than RSP, RAX) which are always
87	* are swapped and restored across the world-switch and also registers like
88	* EFER, PAT MSR etc. which cannot be modified by the guest without causing a
89	* \#VMEXIT.
90	*/
91	#define HMSVM_CPUMCTX_EXTRN_ALL ( CPUMCTX_EXTRN_RIP \
92	\| CPUMCTX_EXTRN_RFLAGS \
93	\| CPUMCTX_EXTRN_RAX \
94	\| CPUMCTX_EXTRN_RSP \
95	\| CPUMCTX_EXTRN_SREG_MASK \
96	\| CPUMCTX_EXTRN_CR0 \
97	\| CPUMCTX_EXTRN_CR2 \
98	\| CPUMCTX_EXTRN_CR3 \
99	\| CPUMCTX_EXTRN_TABLE_MASK \
100	\| CPUMCTX_EXTRN_DR6 \
101	\| CPUMCTX_EXTRN_DR7 \
102	\| CPUMCTX_EXTRN_KERNEL_GS_BASE \
103	\| CPUMCTX_EXTRN_SYSCALL_MSRS \
104	\| CPUMCTX_EXTRN_SYSENTER_MSRS \
105	\| CPUMCTX_EXTRN_HWVIRT \
106	\| CPUMCTX_EXTRN_HM_SVM_MASK)
107
108	/**
109	* Subset of the guest-CPU state that is shared between the guest and host.
110	*/
111	#define HMSVM_CPUMCTX_SHARED_STATE CPUMCTX_EXTRN_DR_MASK
112
113	/** Macro for importing guest state from the VMCB back into CPUMCTX. */
114	#define HMSVM_CPUMCTX_IMPORT_STATE(a_pVCpu, a_fWhat) \
115	do { \
116	if ((a_pVCpu)->cpum.GstCtx.fExtrn & (a_fWhat)) \
117	hmR0SvmImportGuestState((a_pVCpu), (a_fWhat)); \
118	} while (0)
119
120	/** Assert that the required state bits are fetched. */
121	#define HMSVM_CPUMCTX_ASSERT(a_pVCpu, a_fExtrnMbz) AssertMsg(!((a_pVCpu)->cpum.GstCtx.fExtrn & (a_fExtrnMbz)), \
122	("fExtrn=%#RX64 fExtrnMbz=%#RX64\n", \
123	(a_pVCpu)->cpum.GstCtx.fExtrn, (a_fExtrnMbz)))
124
125	/** Assert that preemption is disabled or covered by thread-context hooks. */
126	#define HMSVM_ASSERT_PREEMPT_SAFE(a_pVCpu) Assert( VMMR0ThreadCtxHookIsEnabled((a_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(a_pVCpu) AssertMsg( VMMR0ThreadCtxHookIsEnabled((a_pVCpu)) \
132	\|\| (a_pVCpu)->hm.s.idEnteredCpu == RTMpCpuId(), \
133	("Illegal migration! Entered on CPU %u Current %u\n", \
134	(a_pVCpu)->hm.s.idEnteredCpu, RTMpCpuId()));
135
136	/** Assert that we're not executing a nested-guest. */
137	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
138	# define HMSVM_ASSERT_NOT_IN_NESTED_GUEST(a_pCtx) Assert(!CPUMIsGuestInSvmNestedHwVirtMode((a_pCtx)))
139	#else
140	# define HMSVM_ASSERT_NOT_IN_NESTED_GUEST(a_pCtx) do { NOREF((a_pCtx)); } while (0)
141	#endif
142
143	/** Assert that we're executing a nested-guest. */
144	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
145	# define HMSVM_ASSERT_IN_NESTED_GUEST(a_pCtx) Assert(CPUMIsGuestInSvmNestedHwVirtMode((a_pCtx)))
146	#else
147	# define HMSVM_ASSERT_IN_NESTED_GUEST(a_pCtx) do { NOREF((a_pCtx)); } while (0)
148	#endif
149
150	/** Macro for checking and returning from the using function for
151	* \#VMEXIT intercepts that maybe caused during delivering of another
152	* event in the guest. */
153	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
154	# define HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(a_pVCpu, a_pSvmTransient) \
155	do \
156	{ \
157	int rc = hmR0SvmCheckExitDueToEventDelivery((a_pVCpu), (a_pSvmTransient)); \
158	if (RT_LIKELY(rc == VINF_SUCCESS)) { /* continue #VMEXIT handling */ } \
159	else if ( rc == VINF_HM_DOUBLE_FAULT) { return VINF_SUCCESS; } \
160	else if ( rc == VINF_EM_RESET \
161	&& CPUMIsGuestSvmCtrlInterceptSet((a_pVCpu), &(a_pVCpu)->cpum.GstCtx, SVM_CTRL_INTERCEPT_SHUTDOWN)) \
162	{ \
163	HMSVM_CPUMCTX_IMPORT_STATE((a_pVCpu), HMSVM_CPUMCTX_EXTRN_ALL); \
164	return VBOXSTRICTRC_TODO(IEMExecSvmVmexit((a_pVCpu), SVM_EXIT_SHUTDOWN, 0, 0)); \
165	} \
166	else \
167	return rc; \
168	} while (0)
169	#else
170	# define HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(a_pVCpu, a_pSvmTransient) \
171	do \
172	{ \
173	int rc = hmR0SvmCheckExitDueToEventDelivery((a_pVCpu), (a_pSvmTransient)); \
174	if (RT_LIKELY(rc == VINF_SUCCESS)) { /* continue #VMEXIT handling */ } \
175	else if ( rc == VINF_HM_DOUBLE_FAULT) { return VINF_SUCCESS; } \
176	else \
177	return rc; \
178	} while (0)
179	#endif
180
181	/** Macro for upgrading a @a a_rc to VINF_EM_DBG_STEPPED after emulating an
182	* instruction that exited. */
183	#define HMSVM_CHECK_SINGLE_STEP(a_pVCpu, a_rc) \
184	do { \
185	if ((a_pVCpu)->hm.s.fSingleInstruction && (a_rc) == VINF_SUCCESS) \
186	(a_rc) = VINF_EM_DBG_STEPPED; \
187	} while (0)
188
189	/** Validate segment descriptor granularity bit. */
190	#ifdef VBOX_STRICT
191	# define HMSVM_ASSERT_SEG_GRANULARITY(a_pCtx, reg) \
192	AssertMsg( !(a_pCtx)->reg.Attr.n.u1Present \
193	\|\| ( (a_pCtx)->reg.Attr.n.u1Granularity \
194	? ((a_pCtx)->reg.u32Limit & 0xfff) == 0xfff \
195	: (a_pCtx)->reg.u32Limit <= UINT32_C(0xfffff)), \
196	("Invalid Segment Attributes Limit=%#RX32 Attr=%#RX32 Base=%#RX64\n", (a_pCtx)->reg.u32Limit, \
197	(a_pCtx)->reg.Attr.u, (a_pCtx)->reg.u64Base))
198	#else
199	# define HMSVM_ASSERT_SEG_GRANULARITY(a_pCtx, reg) do { } while (0)
200	#endif
201
202	/**
203	* Exception bitmap mask for all contributory exceptions.
204	*
205	* Page fault is deliberately excluded here as it's conditional as to whether
206	* it's contributory or benign. Page faults are handled separately.
207	*/
208	#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) \
209	\| RT_BIT(X86_XCPT_DE))
210
211	/**
212	* Mandatory/unconditional guest control intercepts.
213	*
214	* SMIs can and do happen in normal operation. We need not intercept them
215	* while executing the guest (or nested-guest).
216	*/
217	#define HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS ( SVM_CTRL_INTERCEPT_INTR \
218	\| SVM_CTRL_INTERCEPT_NMI \
219	\| SVM_CTRL_INTERCEPT_INIT \
220	\| SVM_CTRL_INTERCEPT_RDPMC \
221	\| SVM_CTRL_INTERCEPT_CPUID \
222	\| SVM_CTRL_INTERCEPT_RSM \
223	\| SVM_CTRL_INTERCEPT_HLT \
224	\| SVM_CTRL_INTERCEPT_IOIO_PROT \
225	\| SVM_CTRL_INTERCEPT_MSR_PROT \
226	\| SVM_CTRL_INTERCEPT_INVLPGA \
227	\| SVM_CTRL_INTERCEPT_SHUTDOWN \
228	\| SVM_CTRL_INTERCEPT_FERR_FREEZE \
229	\| SVM_CTRL_INTERCEPT_VMRUN \
230	\| SVM_CTRL_INTERCEPT_SKINIT \
231	\| SVM_CTRL_INTERCEPT_WBINVD \
232	\| SVM_CTRL_INTERCEPT_MONITOR \
233	\| SVM_CTRL_INTERCEPT_MWAIT \
234	\| SVM_CTRL_INTERCEPT_CR0_SEL_WRITE \
235	\| SVM_CTRL_INTERCEPT_XSETBV)
236
237	/** @name VMCB Clean Bits.
238	*
239	* These flags are used for VMCB-state caching. A set VMCB Clean bit indicates
240	* AMD-V doesn't need to reload the corresponding value(s) from the VMCB in
241	* memory.
242	*
243	* @{ */
244	/** All intercepts vectors, TSC offset, PAUSE filter counter. */
245	#define HMSVM_VMCB_CLEAN_INTERCEPTS RT_BIT(0)
246	/** I/O permission bitmap, MSR permission bitmap. */
247	#define HMSVM_VMCB_CLEAN_IOPM_MSRPM RT_BIT(1)
248	/** ASID. */
249	#define HMSVM_VMCB_CLEAN_ASID RT_BIT(2)
250	/** TRP: V_TPR, V_IRQ, V_INTR_PRIO, V_IGN_TPR, V_INTR_MASKING,
251	V_INTR_VECTOR. */
252	#define HMSVM_VMCB_CLEAN_INT_CTRL RT_BIT(3)
253	/** Nested Paging: Nested CR3 (nCR3), PAT. */
254	#define HMSVM_VMCB_CLEAN_NP RT_BIT(4)
255	/** Control registers (CR0, CR3, CR4, EFER). */
256	#define HMSVM_VMCB_CLEAN_CRX_EFER RT_BIT(5)
257	/** Debug registers (DR6, DR7). */
258	#define HMSVM_VMCB_CLEAN_DRX RT_BIT(6)
259	/** GDT, IDT limit and base. */
260	#define HMSVM_VMCB_CLEAN_DT RT_BIT(7)
261	/** Segment register: CS, SS, DS, ES limit and base. */
262	#define HMSVM_VMCB_CLEAN_SEG RT_BIT(8)
263	/** CR2.*/
264	#define HMSVM_VMCB_CLEAN_CR2 RT_BIT(9)
265	/** Last-branch record (DbgCtlMsr, br_from, br_to, lastint_from, lastint_to) */
266	#define HMSVM_VMCB_CLEAN_LBR RT_BIT(10)
267	/** AVIC (AVIC APIC_BAR; AVIC APIC_BACKING_PAGE, AVIC
268	PHYSICAL_TABLE and AVIC LOGICAL_TABLE Pointers). */
269	#define HMSVM_VMCB_CLEAN_AVIC RT_BIT(11)
270	/** Mask of all valid VMCB Clean bits. */
271	#define HMSVM_VMCB_CLEAN_ALL ( HMSVM_VMCB_CLEAN_INTERCEPTS \
272	\| HMSVM_VMCB_CLEAN_IOPM_MSRPM \
273	\| HMSVM_VMCB_CLEAN_ASID \
274	\| HMSVM_VMCB_CLEAN_INT_CTRL \
275	\| HMSVM_VMCB_CLEAN_NP \
276	\| HMSVM_VMCB_CLEAN_CRX_EFER \
277	\| HMSVM_VMCB_CLEAN_DRX \
278	\| HMSVM_VMCB_CLEAN_DT \
279	\| HMSVM_VMCB_CLEAN_SEG \
280	\| HMSVM_VMCB_CLEAN_CR2 \
281	\| HMSVM_VMCB_CLEAN_LBR \
282	\| HMSVM_VMCB_CLEAN_AVIC)
283	/** @} */
284
285	/** @name SVM transient.
286	*
287	* A state structure for holding miscellaneous information across AMD-V
288	* VMRUN/\#VMEXIT operation, restored after the transition.
289	*
290	* @{ */
291	typedef struct SVMTRANSIENT
292	{
293	/** The host's rflags/eflags. */
294	RTCCUINTREG fEFlags;
295	/** The \#VMEXIT exit code (the EXITCODE field in the VMCB). */
296	uint64_t u64ExitCode;
297
298	/** The guest's TPR value used for TPR shadowing. */
299	uint8_t u8GuestTpr;
300	/** Alignment. */
301	uint8_t abAlignment0[7];
302
303	/** Pointer to the currently executing VMCB. */
304	PSVMVMCB pVmcb;
305
306	/** Whether we are currently executing a nested-guest. */
307	bool fIsNestedGuest;
308	/** Whether the guest debug state was active at the time of \#VMEXIT. */
309	bool fWasGuestDebugStateActive;
310	/** Whether the hyper debug state was active at the time of \#VMEXIT. */
311	bool fWasHyperDebugStateActive;
312	/** Whether the TSC offset mode needs to be updated. */
313	bool fUpdateTscOffsetting;
314	/** Whether the TSC_AUX MSR needs restoring on \#VMEXIT. */
315	bool fRestoreTscAuxMsr;
316	/** Whether the \#VMEXIT was caused by a page-fault during delivery of a
317	* contributary exception or a page-fault. */
318	bool fVectoringDoublePF;
319	/** Whether the \#VMEXIT was caused by a page-fault during delivery of an
320	* external interrupt or NMI. */
321	bool fVectoringPF;
322	/** Padding. */
323	bool afPadding0;
324	} SVMTRANSIENT;
325	/** Pointer to SVM transient state. */
326	typedef SVMTRANSIENT *PSVMTRANSIENT;
327	/** Pointer to a const SVM transient state. */
328	typedef const SVMTRANSIENT *PCSVMTRANSIENT;
329
330	AssertCompileSizeAlignment(SVMTRANSIENT, sizeof(uint64_t));
331	AssertCompileMemberAlignment(SVMTRANSIENT, u64ExitCode, sizeof(uint64_t));
332	AssertCompileMemberAlignment(SVMTRANSIENT, pVmcb, sizeof(uint64_t));
333	/** @} */
334
335	/**
336	* MSRPM (MSR permission bitmap) read permissions (for guest RDMSR).
337	*/
338	typedef enum SVMMSREXITREAD
339	{
340	/** Reading this MSR causes a \#VMEXIT. */
341	SVMMSREXIT_INTERCEPT_READ = 0xb,
342	/** Reading this MSR does not cause a \#VMEXIT. */
343	SVMMSREXIT_PASSTHRU_READ
344	} SVMMSREXITREAD;
345
346	/**
347	* MSRPM (MSR permission bitmap) write permissions (for guest WRMSR).
348	*/
349	typedef enum SVMMSREXITWRITE
350	{
351	/** Writing to this MSR causes a \#VMEXIT. */
352	SVMMSREXIT_INTERCEPT_WRITE = 0xd,
353	/** Writing to this MSR does not cause a \#VMEXIT. */
354	SVMMSREXIT_PASSTHRU_WRITE
355	} SVMMSREXITWRITE;
356
357	/**
358	* SVM \#VMEXIT handler.
359	*
360	* @returns VBox status code.
361	* @param pVCpu The cross context virtual CPU structure.
362	* @param pSvmTransient Pointer to the SVM-transient structure.
363	*/
364	typedef int FNSVMEXITHANDLER(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient);
365
366
367	/*********************************************************************************************************************************
368	* Internal Functions *
369	*********************************************************************************************************************************/
370	static void hmR0SvmPendingEventToTrpmTrap(PVMCPUCC pVCpu);
371	static void hmR0SvmLeave(PVMCPUCC pVCpu, bool fImportState);
372
373
374	/** @name \#VMEXIT handlers.
375	* @{
376	*/
377	static FNSVMEXITHANDLER hmR0SvmExitIntr;
378	static FNSVMEXITHANDLER hmR0SvmExitWbinvd;
379	static FNSVMEXITHANDLER hmR0SvmExitInvd;
380	static FNSVMEXITHANDLER hmR0SvmExitCpuid;
381	static FNSVMEXITHANDLER hmR0SvmExitRdtsc;
382	static FNSVMEXITHANDLER hmR0SvmExitRdtscp;
383	static FNSVMEXITHANDLER hmR0SvmExitRdpmc;
384	static FNSVMEXITHANDLER hmR0SvmExitInvlpg;
385	static FNSVMEXITHANDLER hmR0SvmExitHlt;
386	static FNSVMEXITHANDLER hmR0SvmExitMonitor;
387	static FNSVMEXITHANDLER hmR0SvmExitMwait;
388	static FNSVMEXITHANDLER hmR0SvmExitShutdown;
389	static FNSVMEXITHANDLER hmR0SvmExitUnexpected;
390	static FNSVMEXITHANDLER hmR0SvmExitReadCRx;
391	static FNSVMEXITHANDLER hmR0SvmExitWriteCRx;
392	static FNSVMEXITHANDLER hmR0SvmExitMsr;
393	static FNSVMEXITHANDLER hmR0SvmExitReadDRx;
394	static FNSVMEXITHANDLER hmR0SvmExitWriteDRx;
395	static FNSVMEXITHANDLER hmR0SvmExitXsetbv;
396	static FNSVMEXITHANDLER hmR0SvmExitIOInstr;
397	static FNSVMEXITHANDLER hmR0SvmExitNestedPF;
398	static FNSVMEXITHANDLER hmR0SvmExitVIntr;
399	static FNSVMEXITHANDLER hmR0SvmExitTaskSwitch;
400	static FNSVMEXITHANDLER hmR0SvmExitVmmCall;
401	static FNSVMEXITHANDLER hmR0SvmExitPause;
402	static FNSVMEXITHANDLER hmR0SvmExitFerrFreeze;
403	static FNSVMEXITHANDLER hmR0SvmExitIret;
404	static FNSVMEXITHANDLER hmR0SvmExitXcptPF;
405	static FNSVMEXITHANDLER hmR0SvmExitXcptUD;
406	static FNSVMEXITHANDLER hmR0SvmExitXcptMF;
407	static FNSVMEXITHANDLER hmR0SvmExitXcptDB;
408	static FNSVMEXITHANDLER hmR0SvmExitXcptAC;
409	static FNSVMEXITHANDLER hmR0SvmExitXcptBP;
410	static FNSVMEXITHANDLER hmR0SvmExitXcptGP;
411	#if defined(HMSVM_ALWAYS_TRAP_ALL_XCPTS) \|\| defined(VBOX_WITH_NESTED_HWVIRT_SVM)
412	static FNSVMEXITHANDLER hmR0SvmExitXcptGeneric;
413	#endif
414	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
415	static FNSVMEXITHANDLER hmR0SvmExitClgi;
416	static FNSVMEXITHANDLER hmR0SvmExitStgi;
417	static FNSVMEXITHANDLER hmR0SvmExitVmload;
418	static FNSVMEXITHANDLER hmR0SvmExitVmsave;
419	static FNSVMEXITHANDLER hmR0SvmExitInvlpga;
420	static FNSVMEXITHANDLER hmR0SvmExitVmrun;
421	static FNSVMEXITHANDLER hmR0SvmNestedExitXcptDB;
422	static FNSVMEXITHANDLER hmR0SvmNestedExitXcptBP;
423	#endif
424	/** @} */
425
426	static int hmR0SvmHandleExit(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient);
427	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
428	static int hmR0SvmHandleExitNested(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient);
429	#endif
430
431
432	/*********************************************************************************************************************************
433	* Global Variables *
434	*********************************************************************************************************************************/
435	/** Ring-0 memory object for the IO bitmap. */
436	static RTR0MEMOBJ g_hMemObjIOBitmap = NIL_RTR0MEMOBJ;
437	/** Physical address of the IO bitmap. */
438	static RTHCPHYS g_HCPhysIOBitmap;
439	/** Pointer to the IO bitmap. */
440	static R0PTRTYPE(void *) g_pvIOBitmap;
441
442	#ifdef VBOX_STRICT
443	# define HMSVM_LOG_RBP_RSP RT_BIT_32(0)
444	# define HMSVM_LOG_CR_REGS RT_BIT_32(1)
445	# define HMSVM_LOG_CS RT_BIT_32(2)
446	# define HMSVM_LOG_SS RT_BIT_32(3)
447	# define HMSVM_LOG_FS RT_BIT_32(4)
448	# define HMSVM_LOG_GS RT_BIT_32(5)
449	# define HMSVM_LOG_LBR RT_BIT_32(6)
450	# define HMSVM_LOG_ALL ( HMSVM_LOG_RBP_RSP \
451	\| HMSVM_LOG_CR_REGS \
452	\| HMSVM_LOG_CS \
453	\| HMSVM_LOG_SS \
454	\| HMSVM_LOG_FS \
455	\| HMSVM_LOG_GS \
456	\| HMSVM_LOG_LBR)
457
458	/**
459	* Dumps virtual CPU state and additional info. to the logger for diagnostics.
460	*
461	* @param pVCpu The cross context virtual CPU structure.
462	* @param pVmcb Pointer to the VM control block.
463	* @param pszPrefix Log prefix.
464	* @param fFlags Log flags, see HMSVM_LOG_XXX.
465	* @param uVerbose The verbosity level, currently unused.
466	*/
467	static void hmR0SvmLogState(PVMCPUCC pVCpu, PCSVMVMCB pVmcb, const char *pszPrefix, uint32_t fFlags, uint8_t uVerbose)
468	{
469	RT_NOREF2(pVCpu, uVerbose);
470	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
471
472	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_CS \| CPUMCTX_EXTRN_RIP \| CPUMCTX_EXTRN_RFLAGS);
473	Log4(("%s: cs:rip=%04x:%RX64 efl=%#RX64\n", pszPrefix, pCtx->cs.Sel, pCtx->rip, pCtx->rflags.u));
474
475	if (fFlags & HMSVM_LOG_RBP_RSP)
476	{
477	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_RSP \| CPUMCTX_EXTRN_RBP);
478	Log4(("%s: rsp=%#RX64 rbp=%#RX64\n", pszPrefix, pCtx->rsp, pCtx->rbp));
479	}
480
481	if (fFlags & HMSVM_LOG_CR_REGS)
482	{
483	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_CR0 \| CPUMCTX_EXTRN_CR3 \| CPUMCTX_EXTRN_CR4);
484	Log4(("%s: cr0=%#RX64 cr3=%#RX64 cr4=%#RX64\n", pszPrefix, pCtx->cr0, pCtx->cr3, pCtx->cr4));
485	}
486
487	if (fFlags & HMSVM_LOG_CS)
488	{
489	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_CS);
490	Log4(("%s: cs={%04x base=%016RX64 limit=%08x flags=%08x}\n", pszPrefix, pCtx->cs.Sel, pCtx->cs.u64Base,
491	pCtx->cs.u32Limit, pCtx->cs.Attr.u));
492	}
493	if (fFlags & HMSVM_LOG_SS)
494	{
495	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_SS);
496	Log4(("%s: ss={%04x base=%016RX64 limit=%08x flags=%08x}\n", pszPrefix, pCtx->ss.Sel, pCtx->ss.u64Base,
497	pCtx->ss.u32Limit, pCtx->ss.Attr.u));
498	}
499	if (fFlags & HMSVM_LOG_FS)
500	{
501	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_FS);
502	Log4(("%s: fs={%04x base=%016RX64 limit=%08x flags=%08x}\n", pszPrefix, pCtx->fs.Sel, pCtx->fs.u64Base,
503	pCtx->fs.u32Limit, pCtx->fs.Attr.u));
504	}
505	if (fFlags & HMSVM_LOG_GS)
506	{
507	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_GS);
508	Log4(("%s: gs={%04x base=%016RX64 limit=%08x flags=%08x}\n", pszPrefix, pCtx->gs.Sel, pCtx->gs.u64Base,
509	pCtx->gs.u32Limit, pCtx->gs.Attr.u));
510	}
511
512	PCSVMVMCBSTATESAVE pVmcbGuest = &pVmcb->guest;
513	if (fFlags & HMSVM_LOG_LBR)
514	{
515	Log4(("%s: br_from=%#RX64 br_to=%#RX64 lastxcpt_from=%#RX64 lastxcpt_to=%#RX64\n", pszPrefix, pVmcbGuest->u64BR_FROM,
516	pVmcbGuest->u64BR_TO, pVmcbGuest->u64LASTEXCPFROM, pVmcbGuest->u64LASTEXCPTO));
517	}
518	NOREF(pszPrefix); NOREF(pVmcbGuest); NOREF(pCtx);
519	}
520	#endif /* VBOX_STRICT */
521
522
523	/**
524	* Sets up and activates AMD-V on the current CPU.
525	*
526	* @returns VBox status code.
527	* @param pHostCpu The HM physical-CPU structure.
528	* @param pVM The cross context VM structure. Can be
529	* NULL after a resume!
530	* @param pvCpuPage Pointer to the global CPU page.
531	* @param HCPhysCpuPage Physical address of the global CPU page.
532	* @param fEnabledByHost Whether the host OS has already initialized AMD-V.
533	* @param pHwvirtMsrs Pointer to the hardware-virtualization MSRs (currently
534	* unused).
535	*/
536	VMMR0DECL(int) SVMR0EnableCpu(PHMPHYSCPU pHostCpu, PVMCC pVM, void *pvCpuPage, RTHCPHYS HCPhysCpuPage, bool fEnabledByHost,
537	PCSUPHWVIRTMSRS pHwvirtMsrs)
538	{
539	Assert(!fEnabledByHost);
540	Assert(HCPhysCpuPage && HCPhysCpuPage != NIL_RTHCPHYS);
541	Assert(RT_ALIGN_T(HCPhysCpuPage, _4K, RTHCPHYS) == HCPhysCpuPage);
542	Assert(pvCpuPage); NOREF(pvCpuPage);
543	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
544
545	RT_NOREF2(fEnabledByHost, pHwvirtMsrs);
546
547	/* Paranoid: Disable interrupt as, in theory, interrupt handlers might mess with EFER. */
548	RTCCUINTREG const fEFlags = ASMIntDisableFlags();
549
550	/*
551	* We must turn on AMD-V and setup the host state physical address, as those MSRs are per CPU.
552	*/
553	uint64_t u64HostEfer = ASMRdMsr(MSR_K6_EFER);
554	if (u64HostEfer & MSR_K6_EFER_SVME)
555	{
556	/* If the VBOX_HWVIRTEX_IGNORE_SVM_IN_USE is active, then we blindly use AMD-V. */
557	if ( pVM
558	&& pVM->hm.s.svm.fIgnoreInUseError)
559	pHostCpu->fIgnoreAMDVInUseError = true;
560
561	if (!pHostCpu->fIgnoreAMDVInUseError)
562	{
563	ASMSetFlags(fEFlags);
564	return VERR_SVM_IN_USE;
565	}
566	}
567
568	/* Turn on AMD-V in the EFER MSR. */
569	ASMWrMsr(MSR_K6_EFER, u64HostEfer \| MSR_K6_EFER_SVME);
570
571	/* Write the physical page address where the CPU will store the host state while executing the VM. */
572	ASMWrMsr(MSR_K8_VM_HSAVE_PA, HCPhysCpuPage);
573
574	/* Restore interrupts. */
575	ASMSetFlags(fEFlags);
576
577	/*
578	* Theoretically, other hypervisors may have used ASIDs, ideally we should flush all
579	* non-zero ASIDs when enabling SVM. AMD doesn't have an SVM instruction to flush all
580	* ASIDs (flushing is done upon VMRUN). Therefore, flag that we need to flush the TLB
581	* entirely with before executing any guest code.
582	*/
583	pHostCpu->fFlushAsidBeforeUse = true;
584
585	/*
586	* Ensure each VCPU scheduled on this CPU gets a new ASID on resume. See @bugref{6255}.
587	*/
588	++pHostCpu->cTlbFlushes;
589
590	return VINF_SUCCESS;
591	}
592
593
594	/**
595	* Deactivates AMD-V on the current CPU.
596	*
597	* @returns VBox status code.
598	* @param pHostCpu The HM physical-CPU structure.
599	* @param pvCpuPage Pointer to the global CPU page.
600	* @param HCPhysCpuPage Physical address of the global CPU page.
601	*/
602	VMMR0DECL(int) SVMR0DisableCpu(PHMPHYSCPU pHostCpu, void *pvCpuPage, RTHCPHYS HCPhysCpuPage)
603	{
604	RT_NOREF1(pHostCpu);
605	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
606	AssertReturn( HCPhysCpuPage
607	&& HCPhysCpuPage != NIL_RTHCPHYS, VERR_INVALID_PARAMETER);
608	AssertReturn(pvCpuPage, VERR_INVALID_PARAMETER);
609
610	/* Paranoid: Disable interrupts as, in theory, interrupt handlers might mess with EFER. */
611	RTCCUINTREG const fEFlags = ASMIntDisableFlags();
612
613	/* Turn off AMD-V in the EFER MSR. */
614	uint64_t u64HostEfer = ASMRdMsr(MSR_K6_EFER);
615	ASMWrMsr(MSR_K6_EFER, u64HostEfer & ~MSR_K6_EFER_SVME);
616
617	/* Invalidate host state physical address. */
618	ASMWrMsr(MSR_K8_VM_HSAVE_PA, 0);
619
620	/* Restore interrupts. */
621	ASMSetFlags(fEFlags);
622
623	return VINF_SUCCESS;
624	}
625
626
627	/**
628	* Does global AMD-V initialization (called during module initialization).
629	*
630	* @returns VBox status code.
631	*/
632	VMMR0DECL(int) SVMR0GlobalInit(void)
633	{
634	/*
635	* Allocate 12 KB (3 pages) for the IO bitmap. Since this is non-optional and we always
636	* intercept all IO accesses, it's done once globally here instead of per-VM.
637	*/
638	Assert(g_hMemObjIOBitmap == NIL_RTR0MEMOBJ);
639	int rc = RTR0MemObjAllocCont(&g_hMemObjIOBitmap, SVM_IOPM_PAGES << X86_PAGE_4K_SHIFT, false /* fExecutable */);
640	if (RT_FAILURE(rc))
641	return rc;
642
643	g_pvIOBitmap = RTR0MemObjAddress(g_hMemObjIOBitmap);
644	g_HCPhysIOBitmap = RTR0MemObjGetPagePhysAddr(g_hMemObjIOBitmap, 0 /* iPage */);
645
646	/* Set all bits to intercept all IO accesses. */
647	ASMMemFill32(g_pvIOBitmap, SVM_IOPM_PAGES << X86_PAGE_4K_SHIFT, UINT32_C(0xffffffff));
648
649	return VINF_SUCCESS;
650	}
651
652
653	/**
654	* Does global AMD-V termination (called during module termination).
655	*/
656	VMMR0DECL(void) SVMR0GlobalTerm(void)
657	{
658	if (g_hMemObjIOBitmap != NIL_RTR0MEMOBJ)
659	{
660	RTR0MemObjFree(g_hMemObjIOBitmap, true /* fFreeMappings */);
661	g_pvIOBitmap = NULL;
662	g_HCPhysIOBitmap = 0;
663	g_hMemObjIOBitmap = NIL_RTR0MEMOBJ;
664	}
665	}
666
667
668	/**
669	* Frees any allocated per-VCPU structures for a VM.
670	*
671	* @param pVM The cross context VM structure.
672	*/
673	DECLINLINE(void) hmR0SvmFreeStructs(PVMCC pVM)
674	{
675	for (VMCPUID idCpu = 0; idCpu < pVM->cCpus; idCpu++)
676	{
677	PVMCPUCC pVCpu = VMCC_GET_CPU(pVM, idCpu);
678	AssertPtr(pVCpu);
679
680	if (pVCpu->hm.s.svm.hMemObjVmcbHost != NIL_RTR0MEMOBJ)
681	{
682	RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjVmcbHost, false);
683	pVCpu->hm.s.svm.HCPhysVmcbHost = 0;
684	pVCpu->hm.s.svm.hMemObjVmcbHost = NIL_RTR0MEMOBJ;
685	}
686
687	if (pVCpu->hm.s.svm.hMemObjVmcb != NIL_RTR0MEMOBJ)
688	{
689	RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjVmcb, false);
690	pVCpu->hm.s.svm.pVmcb = NULL;
691	pVCpu->hm.s.svm.HCPhysVmcb = 0;
692	pVCpu->hm.s.svm.hMemObjVmcb = NIL_RTR0MEMOBJ;
693	}
694
695	if (pVCpu->hm.s.svm.hMemObjMsrBitmap != NIL_RTR0MEMOBJ)
696	{
697	RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjMsrBitmap, false);
698	pVCpu->hm.s.svm.pvMsrBitmap = NULL;
699	pVCpu->hm.s.svm.HCPhysMsrBitmap = 0;
700	pVCpu->hm.s.svm.hMemObjMsrBitmap = NIL_RTR0MEMOBJ;
701	}
702	}
703	}
704
705
706	/**
707	* Does per-VM AMD-V initialization.
708	*
709	* @returns VBox status code.
710	* @param pVM The cross context VM structure.
711	*/
712	VMMR0DECL(int) SVMR0InitVM(PVMCC pVM)
713	{
714	int rc = VERR_INTERNAL_ERROR_5;
715
716	/*
717	* Check for an AMD CPU erratum which requires us to flush the TLB before every world-switch.
718	*/
719	uint32_t u32Family;
720	uint32_t u32Model;
721	uint32_t u32Stepping;
722	if (HMIsSubjectToSvmErratum170(&u32Family, &u32Model, &u32Stepping))
723	{
724	Log4Func(("AMD cpu with erratum 170 family %#x model %#x stepping %#x\n", u32Family, u32Model, u32Stepping));
725	pVM->hm.s.svm.fAlwaysFlushTLB = true;
726	}
727
728	/*
729	* Initialize the R0 memory objects up-front so we can properly cleanup on allocation failures.
730	*/
731	for (VMCPUID idCpu = 0; idCpu < pVM->cCpus; idCpu++)
732	{
733	PVMCPUCC pVCpu = VMCC_GET_CPU(pVM, idCpu);
734	pVCpu->hm.s.svm.hMemObjVmcbHost = NIL_RTR0MEMOBJ;
735	pVCpu->hm.s.svm.hMemObjVmcb = NIL_RTR0MEMOBJ;
736	pVCpu->hm.s.svm.hMemObjMsrBitmap = NIL_RTR0MEMOBJ;
737	}
738
739	for (VMCPUID idCpu = 0; idCpu < pVM->cCpus; idCpu++)
740	{
741	PVMCPUCC pVCpu = VMCC_GET_CPU(pVM, idCpu);
742
743	/*
744	* Initialize the hardware-assisted SVM guest-execution handler.
745	* We now use a single handler for both 32-bit and 64-bit guests, see @bugref{6208#c73}.
746	*/
747	pVCpu->hm.s.svm.pfnVMRun = SVMR0VMRun;
748
749	/*
750	* Allocate one page for the host-context VM control block (VMCB). This is used for additional host-state (such as
751	* FS, GS, Kernel GS Base, etc.) apart from the host-state save area specified in MSR_K8_VM_HSAVE_PA.
752	*/
753	rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjVmcbHost, SVM_VMCB_PAGES << PAGE_SHIFT, false /* fExecutable */);
754	if (RT_FAILURE(rc))
755	goto failure_cleanup;
756
757	void *pvVmcbHost = RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjVmcbHost);
758	pVCpu->hm.s.svm.HCPhysVmcbHost = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjVmcbHost, 0 /* iPage */);
759	Assert(pVCpu->hm.s.svm.HCPhysVmcbHost < _4G);
760	ASMMemZeroPage(pvVmcbHost);
761
762	/*
763	* Allocate one page for the guest-state VMCB.
764	*/
765	rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjVmcb, SVM_VMCB_PAGES << PAGE_SHIFT, false /* fExecutable */);
766	if (RT_FAILURE(rc))
767	goto failure_cleanup;
768
769	pVCpu->hm.s.svm.pVmcb = (PSVMVMCB)RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjVmcb);
770	pVCpu->hm.s.svm.HCPhysVmcb = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjVmcb, 0 /* iPage */);
771	Assert(pVCpu->hm.s.svm.HCPhysVmcb < _4G);
772	ASMMemZeroPage(pVCpu->hm.s.svm.pVmcb);
773
774	/*
775	* Allocate two pages (8 KB) for the MSR permission bitmap. There doesn't seem to be a way to convince
776	* SVM to not require one.
777	*/
778	rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjMsrBitmap, SVM_MSRPM_PAGES << X86_PAGE_4K_SHIFT,
779	false /* fExecutable */);
780	if (RT_FAILURE(rc))
781	goto failure_cleanup;
782
783	pVCpu->hm.s.svm.pvMsrBitmap = RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjMsrBitmap);
784	pVCpu->hm.s.svm.HCPhysMsrBitmap = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjMsrBitmap, 0 /* iPage */);
785	/* Set all bits to intercept all MSR accesses (changed later on). */
786	ASMMemFill32(pVCpu->hm.s.svm.pvMsrBitmap, SVM_MSRPM_PAGES << X86_PAGE_4K_SHIFT, UINT32_C(0xffffffff));
787	}
788
789	return VINF_SUCCESS;
790
791	failure_cleanup:
792	hmR0SvmFreeStructs(pVM);
793	return rc;
794	}
795
796
797	/**
798	* Does per-VM AMD-V termination.
799	*
800	* @returns VBox status code.
801	* @param pVM The cross context VM structure.
802	*/
803	VMMR0DECL(int) SVMR0TermVM(PVMCC pVM)
804	{
805	hmR0SvmFreeStructs(pVM);
806	return VINF_SUCCESS;
807	}
808
809
810	/**
811	* Returns whether the VMCB Clean Bits feature is supported.
812	*
813	* @returns @c true if supported, @c false otherwise.
814	* @param pVCpu The cross context virtual CPU structure.
815	* @param fIsNestedGuest Whether we are currently executing the nested-guest.
816	*/
817	DECL_FORCE_INLINE(bool) hmR0SvmSupportsVmcbCleanBits(PVMCPUCC pVCpu, bool fIsNestedGuest)
818	{
819	PCVMCC pVM = pVCpu->CTX_SUFF(pVM);
820	bool const fHostVmcbCleanBits = RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_VMCB_CLEAN);
821	if (!fIsNestedGuest)
822	return fHostVmcbCleanBits;
823	return fHostVmcbCleanBits && pVM->cpum.ro.GuestFeatures.fSvmVmcbClean;
824	}
825
826
827	/**
828	* Returns whether the decode assists feature is supported.
829	*
830	* @returns @c true if supported, @c false otherwise.
831	* @param pVCpu The cross context virtual CPU structure.
832	*/
833	DECLINLINE(bool) hmR0SvmSupportsDecodeAssists(PVMCPUCC pVCpu)
834	{
835	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
836	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
837	if (CPUMIsGuestInSvmNestedHwVirtMode(&pVCpu->cpum.GstCtx))
838	{
839	return (pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_DECODE_ASSISTS)
840	&& pVM->cpum.ro.GuestFeatures.fSvmDecodeAssists;
841	}
842	#endif
843	return RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_DECODE_ASSISTS);
844	}
845
846
847	/**
848	* Returns whether the NRIP_SAVE feature is supported.
849	*
850	* @returns @c true if supported, @c false otherwise.
851	* @param pVCpu The cross context virtual CPU structure.
852	*/
853	DECLINLINE(bool) hmR0SvmSupportsNextRipSave(PVMCPUCC pVCpu)
854	{
855	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
856	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
857	if (CPUMIsGuestInSvmNestedHwVirtMode(&pVCpu->cpum.GstCtx))
858	{
859	return (pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_NRIP_SAVE)
860	&& pVM->cpum.ro.GuestFeatures.fSvmNextRipSave;
861	}
862	#endif
863	return RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_NRIP_SAVE);
864	}
865
866
867	/**
868	* Sets the permission bits for the specified MSR in the MSRPM bitmap.
869	*
870	* @param pVCpu The cross context virtual CPU structure.
871	* @param pbMsrBitmap Pointer to the MSR bitmap.
872	* @param idMsr The MSR for which the permissions are being set.
873	* @param enmRead MSR read permissions.
874	* @param enmWrite MSR write permissions.
875	*
876	* @remarks This function does -not- clear the VMCB clean bits for MSRPM. The
877	* caller needs to take care of this.
878	*/
879	static void hmR0SvmSetMsrPermission(PVMCPUCC pVCpu, uint8_t *pbMsrBitmap, uint32_t idMsr, SVMMSREXITREAD enmRead,
880	SVMMSREXITWRITE enmWrite)
881	{
882	bool const fInNestedGuestMode = CPUMIsGuestInSvmNestedHwVirtMode(&pVCpu->cpum.GstCtx);
883	uint16_t offMsrpm;
884	uint8_t uMsrpmBit;
885	int rc = CPUMGetSvmMsrpmOffsetAndBit(idMsr, &offMsrpm, &uMsrpmBit);
886	AssertRC(rc);
887
888	Assert(uMsrpmBit == 0 \|\| uMsrpmBit == 2 \|\| uMsrpmBit == 4 \|\| uMsrpmBit == 6);
889	Assert(offMsrpm < SVM_MSRPM_PAGES << X86_PAGE_4K_SHIFT);
890
891	pbMsrBitmap += offMsrpm;
892	if (enmRead == SVMMSREXIT_INTERCEPT_READ)
893	*pbMsrBitmap \|= RT_BIT(uMsrpmBit);
894	else
895	{
896	if (!fInNestedGuestMode)
897	*pbMsrBitmap &= ~RT_BIT(uMsrpmBit);
898	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
899	else
900	{
901	/* Only clear the bit if the nested-guest is also not intercepting the MSR read.*/
902	uint8_t const pbNstGstMsrBitmap = (uint8_t )pVCpu->cpum.GstCtx.hwvirt.svm.CTX_SUFF(pvMsrBitmap);
903	pbNstGstMsrBitmap += offMsrpm;
904	if (!(*pbNstGstMsrBitmap & RT_BIT(uMsrpmBit)))
905	*pbMsrBitmap &= ~RT_BIT(uMsrpmBit);
906	else
907	Assert(*pbMsrBitmap & RT_BIT(uMsrpmBit));
908	}
909	#endif
910	}
911
912	if (enmWrite == SVMMSREXIT_INTERCEPT_WRITE)
913	*pbMsrBitmap \|= RT_BIT(uMsrpmBit + 1);
914	else
915	{
916	if (!fInNestedGuestMode)
917	*pbMsrBitmap &= ~RT_BIT(uMsrpmBit + 1);
918	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
919	else
920	{
921	/* Only clear the bit if the nested-guest is also not intercepting the MSR write.*/
922	uint8_t const pbNstGstMsrBitmap = (uint8_t )pVCpu->cpum.GstCtx.hwvirt.svm.CTX_SUFF(pvMsrBitmap);
923	pbNstGstMsrBitmap += offMsrpm;
924	if (!(*pbNstGstMsrBitmap & RT_BIT(uMsrpmBit + 1)))
925	*pbMsrBitmap &= ~RT_BIT(uMsrpmBit + 1);
926	else
927	Assert(*pbMsrBitmap & RT_BIT(uMsrpmBit + 1));
928	}
929	#endif
930	}
931	}
932
933
934	/**
935	* Sets up AMD-V for the specified VM.
936	* This function is only called once per-VM during initalization.
937	*
938	* @returns VBox status code.
939	* @param pVM The cross context VM structure.
940	*/
941	VMMR0DECL(int) SVMR0SetupVM(PVMCC pVM)
942	{
943	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
944	AssertReturn(pVM, VERR_INVALID_PARAMETER);
945	Assert(pVM->hm.s.svm.fSupported);
946
947	bool const fPauseFilter = RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_PAUSE_FILTER);
948	bool const fPauseFilterThreshold = RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_PAUSE_FILTER_THRESHOLD);
949	bool const fUsePauseFilter = fPauseFilter && pVM->hm.s.svm.cPauseFilter;
950
951	bool const fLbrVirt = RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_LBR_VIRT);
952	bool const fUseLbrVirt = fLbrVirt && pVM->hm.s.svm.fLbrVirt; /** @todo IEM implementation etc. */
953
954	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
955	bool const fVirtVmsaveVmload = RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_VIRT_VMSAVE_VMLOAD);
956	bool const fUseVirtVmsaveVmload = fVirtVmsaveVmload && pVM->hm.s.svm.fVirtVmsaveVmload && pVM->hm.s.fNestedPaging;
957
958	bool const fVGif = RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_VGIF);
959	bool const fUseVGif = fVGif && pVM->hm.s.svm.fVGif;
960	#endif
961
962	PVMCPUCC pVCpu0 = VMCC_GET_CPU_0(pVM);
963	PSVMVMCB pVmcb0 = pVCpu0->hm.s.svm.pVmcb;
964	AssertMsgReturn(RT_VALID_PTR(pVmcb0), ("Invalid pVmcb (%p) for vcpu[0]\n", pVmcb0), VERR_SVM_INVALID_PVMCB);
965	PSVMVMCBCTRL pVmcbCtrl0 = &pVmcb0->ctrl;
966
967	/* Always trap #AC for reasons of security. */
968	pVmcbCtrl0->u32InterceptXcpt \|= RT_BIT_32(X86_XCPT_AC);
969
970	/* Always trap #DB for reasons of security. */
971	pVmcbCtrl0->u32InterceptXcpt \|= RT_BIT_32(X86_XCPT_DB);
972
973	/* Trap exceptions unconditionally (debug purposes). */
974	#ifdef HMSVM_ALWAYS_TRAP_PF
975	pVmcbCtrl0->u32InterceptXcpt \|= RT_BIT_32(X86_XCPT_PF);
976	#endif
977	#ifdef HMSVM_ALWAYS_TRAP_ALL_XCPTS
978	/* If you add any exceptions here, make sure to update hmR0SvmHandleExit(). */
979	pVmcbCtrl0->u32InterceptXcpt \|= RT_BIT_32(X86_XCPT_BP)
980	\| RT_BIT_32(X86_XCPT_DE)
981	\| RT_BIT_32(X86_XCPT_NM)
982	\| RT_BIT_32(X86_XCPT_UD)
983	\| RT_BIT_32(X86_XCPT_NP)
984	\| RT_BIT_32(X86_XCPT_SS)
985	\| RT_BIT_32(X86_XCPT_GP)
986	\| RT_BIT_32(X86_XCPT_PF)
987	\| RT_BIT_32(X86_XCPT_MF)
988	;
989	#endif
990
991	/* Apply the exceptions intercepts needed by the GIM provider. */
992	if (pVCpu0->hm.s.fGIMTrapXcptUD)
993	pVmcbCtrl0->u32InterceptXcpt \|= RT_BIT(X86_XCPT_UD);
994
995	/* The mesa 3d driver hack needs #GP. */
996	if (pVCpu0->hm.s.fTrapXcptGpForLovelyMesaDrv)
997	pVmcbCtrl0->u32InterceptXcpt \|= RT_BIT(X86_XCPT_GP);
998
999	/* Set up unconditional intercepts and conditions. */
1000	pVmcbCtrl0->u64InterceptCtrl = HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS
1001	\| SVM_CTRL_INTERCEPT_VMMCALL;
1002
1003	#ifdef HMSVM_ALWAYS_TRAP_TASK_SWITCH
1004	pVmcbCtrl0->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_TASK_SWITCH;
1005	#endif
1006
1007	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1008	/* Virtualized VMSAVE/VMLOAD. */
1009	pVmcbCtrl0->LbrVirt.n.u1VirtVmsaveVmload = fUseVirtVmsaveVmload;
1010	if (!fUseVirtVmsaveVmload)
1011	pVmcbCtrl0->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_VMSAVE
1012	\| SVM_CTRL_INTERCEPT_VMLOAD;
1013
1014	/* Virtual GIF. */
1015	pVmcbCtrl0->IntCtrl.n.u1VGifEnable = fUseVGif;
1016	if (!fUseVGif)
1017	pVmcbCtrl0->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_CLGI
1018	\| SVM_CTRL_INTERCEPT_STGI;
1019	#endif
1020
1021	/* CR4 writes must always be intercepted for tracking PGM mode changes. */
1022	pVmcbCtrl0->u16InterceptWrCRx = RT_BIT(4);
1023
1024	/* Intercept all DRx reads and writes by default. Changed later on. */
1025	pVmcbCtrl0->u16InterceptRdDRx = 0xffff;
1026	pVmcbCtrl0->u16InterceptWrDRx = 0xffff;
1027
1028	/* Virtualize masking of INTR interrupts. (reads/writes from/to CR8 go to the V_TPR register) */
1029	pVmcbCtrl0->IntCtrl.n.u1VIntrMasking = 1;
1030
1031	/* Ignore the priority in the virtual TPR. This is necessary for delivering PIC style (ExtInt) interrupts
1032	and we currently deliver both PIC and APIC interrupts alike, see hmR0SvmEvaluatePendingEvent() */
1033	pVmcbCtrl0->IntCtrl.n.u1IgnoreTPR = 1;
1034
1035	/* Set the IO permission bitmap physical addresses. */
1036	pVmcbCtrl0->u64IOPMPhysAddr = g_HCPhysIOBitmap;
1037
1038	/* LBR virtualization. */
1039	pVmcbCtrl0->LbrVirt.n.u1LbrVirt = fUseLbrVirt;
1040
1041	/* The host ASID MBZ, for the guest start with 1. */
1042	pVmcbCtrl0->TLBCtrl.n.u32ASID = 1;
1043
1044	/* Setup Nested Paging. This doesn't change throughout the execution time of the VM. */
1045	pVmcbCtrl0->NestedPagingCtrl.n.u1NestedPaging = pVM->hm.s.fNestedPaging;
1046
1047	/* Without Nested Paging, we need additionally intercepts. */
1048	if (!pVM->hm.s.fNestedPaging)
1049	{
1050	/* CR3 reads/writes must be intercepted; our shadow values differ from the guest values. */
1051	pVmcbCtrl0->u16InterceptRdCRx \|= RT_BIT(3);
1052	pVmcbCtrl0->u16InterceptWrCRx \|= RT_BIT(3);
1053
1054	/* Intercept INVLPG and task switches (may change CR3, EFLAGS, LDT). */
1055	pVmcbCtrl0->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_INVLPG
1056	\| SVM_CTRL_INTERCEPT_TASK_SWITCH;
1057
1058	/* Page faults must be intercepted to implement shadow paging. */
1059	pVmcbCtrl0->u32InterceptXcpt \|= RT_BIT(X86_XCPT_PF);
1060	}
1061
1062	/* Setup Pause Filter for guest pause-loop (spinlock) exiting. */
1063	if (fUsePauseFilter)
1064	{
1065	Assert(pVM->hm.s.svm.cPauseFilter > 0);
1066	pVmcbCtrl0->u16PauseFilterCount = pVM->hm.s.svm.cPauseFilter;
1067	if (fPauseFilterThreshold)
1068	pVmcbCtrl0->u16PauseFilterThreshold = pVM->hm.s.svm.cPauseFilterThresholdTicks;
1069	pVmcbCtrl0->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_PAUSE;
1070	}
1071
1072	/*
1073	* Setup the MSR permission bitmap.
1074	* The following MSRs are saved/restored automatically during the world-switch.
1075	* Don't intercept guest read/write accesses to these MSRs.
1076	*/
1077	uint8_t pbMsrBitmap0 = (uint8_t )pVCpu0->hm.s.svm.pvMsrBitmap;
1078	hmR0SvmSetMsrPermission(pVCpu0, pbMsrBitmap0, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1079	hmR0SvmSetMsrPermission(pVCpu0, pbMsrBitmap0, MSR_K8_CSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1080	hmR0SvmSetMsrPermission(pVCpu0, pbMsrBitmap0, MSR_K6_STAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1081	hmR0SvmSetMsrPermission(pVCpu0, pbMsrBitmap0, MSR_K8_SF_MASK, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1082	hmR0SvmSetMsrPermission(pVCpu0, pbMsrBitmap0, MSR_K8_FS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1083	hmR0SvmSetMsrPermission(pVCpu0, pbMsrBitmap0, MSR_K8_GS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1084	hmR0SvmSetMsrPermission(pVCpu0, pbMsrBitmap0, MSR_K8_KERNEL_GS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1085	hmR0SvmSetMsrPermission(pVCpu0, pbMsrBitmap0, MSR_IA32_SYSENTER_CS, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1086	hmR0SvmSetMsrPermission(pVCpu0, pbMsrBitmap0, MSR_IA32_SYSENTER_ESP, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1087	hmR0SvmSetMsrPermission(pVCpu0, pbMsrBitmap0, MSR_IA32_SYSENTER_EIP, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1088	pVmcbCtrl0->u64MSRPMPhysAddr = pVCpu0->hm.s.svm.HCPhysMsrBitmap;
1089
1090	/* Initially all VMCB clean bits MBZ indicating that everything should be loaded from the VMCB in memory. */
1091	Assert(pVmcbCtrl0->u32VmcbCleanBits == 0);
1092
1093	for (VMCPUID idCpu = 1; idCpu < pVM->cCpus; idCpu++)
1094	{
1095	PVMCPUCC pVCpuCur = VMCC_GET_CPU(pVM, idCpu);
1096	PSVMVMCB pVmcbCur = pVCpuCur->hm.s.svm.pVmcb;
1097	AssertMsgReturn(RT_VALID_PTR(pVmcbCur), ("Invalid pVmcb (%p) for vcpu[%u]\n", pVmcbCur, idCpu), VERR_SVM_INVALID_PVMCB);
1098	PSVMVMCBCTRL pVmcbCtrlCur = &pVmcbCur->ctrl;
1099
1100	/* Copy the VMCB control area. */
1101	memcpy(pVmcbCtrlCur, pVmcbCtrl0, sizeof(*pVmcbCtrlCur));
1102
1103	/* Copy the MSR bitmap and setup the VCPU-specific host physical address. */
1104	uint8_t pbMsrBitmapCur = (uint8_t )pVCpuCur->hm.s.svm.pvMsrBitmap;
1105	memcpy(pbMsrBitmapCur, pbMsrBitmap0, SVM_MSRPM_PAGES << X86_PAGE_4K_SHIFT);
1106	pVmcbCtrlCur->u64MSRPMPhysAddr = pVCpuCur->hm.s.svm.HCPhysMsrBitmap;
1107
1108	/* Initially all VMCB clean bits MBZ indicating that everything should be loaded from the VMCB in memory. */
1109	Assert(pVmcbCtrlCur->u32VmcbCleanBits == 0);
1110
1111	/* Verify our assumption that GIM providers trap #UD uniformly across VCPUs initially. */
1112	Assert(pVCpuCur->hm.s.fGIMTrapXcptUD == pVCpu0->hm.s.fGIMTrapXcptUD);
1113	}
1114
1115	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1116	LogRel(("HM: fUsePauseFilter=%RTbool fUseLbrVirt=%RTbool fUseVGif=%RTbool fUseVirtVmsaveVmload=%RTbool\n", fUsePauseFilter,
1117	fUseLbrVirt, fUseVGif, fUseVirtVmsaveVmload));
1118	#else
1119	LogRel(("HM: fUsePauseFilter=%RTbool fUseLbrVirt=%RTbool\n", fUsePauseFilter, fUseLbrVirt));
1120	#endif
1121	return VINF_SUCCESS;
1122	}
1123
1124
1125	/**
1126	* Gets a pointer to the currently active guest (or nested-guest) VMCB.
1127	*
1128	* @returns Pointer to the current context VMCB.
1129	* @param pVCpu The cross context virtual CPU structure.
1130	*/
1131	DECLINLINE(PSVMVMCB) hmR0SvmGetCurrentVmcb(PVMCPUCC pVCpu)
1132	{
1133	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1134	if (CPUMIsGuestInSvmNestedHwVirtMode(&pVCpu->cpum.GstCtx))
1135	return pVCpu->cpum.GstCtx.hwvirt.svm.CTX_SUFF(pVmcb);
1136	#endif
1137	return pVCpu->hm.s.svm.pVmcb;
1138	}
1139
1140
1141	/**
1142	* Gets a pointer to the nested-guest VMCB cache.
1143	*
1144	* @returns Pointer to the nested-guest VMCB cache.
1145	* @param pVCpu The cross context virtual CPU structure.
1146	*/
1147	DECLINLINE(PSVMNESTEDVMCBCACHE) hmR0SvmGetNestedVmcbCache(PVMCPUCC pVCpu)
1148	{
1149	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1150	Assert(pVCpu->hm.s.svm.NstGstVmcbCache.fCacheValid);
1151	return &pVCpu->hm.s.svm.NstGstVmcbCache;
1152	#else
1153	RT_NOREF(pVCpu);
1154	return NULL;
1155	#endif
1156	}
1157
1158
1159	/**
1160	* Invalidates a guest page by guest virtual address.
1161	*
1162	* @returns VBox status code.
1163	* @param pVCpu The cross context virtual CPU structure.
1164	* @param GCVirt Guest virtual address of the page to invalidate.
1165	*/
1166	VMMR0DECL(int) SVMR0InvalidatePage(PVMCPUCC pVCpu, RTGCPTR GCVirt)
1167	{
1168	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.svm.fSupported);
1169
1170	bool const fFlushPending = pVCpu->CTX_SUFF(pVM)->hm.s.svm.fAlwaysFlushTLB \|\| VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_TLB_FLUSH);
1171
1172	/* Skip it if a TLB flush is already pending. */
1173	if (!fFlushPending)
1174	{
1175	Log4Func(("%#RGv\n", GCVirt));
1176
1177	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
1178	AssertMsgReturn(pVmcb, ("Invalid pVmcb!\n"), VERR_SVM_INVALID_PVMCB);
1179
1180	SVMR0InvlpgA(GCVirt, pVmcb->ctrl.TLBCtrl.n.u32ASID);
1181	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlbInvlpgVirt);
1182	}
1183	return VINF_SUCCESS;
1184	}
1185
1186
1187	/**
1188	* Flushes the appropriate tagged-TLB entries.
1189	*
1190	* @param pHostCpu The HM physical-CPU structure.
1191	* @param pVCpu The cross context virtual CPU structure.
1192	* @param pVmcb Pointer to the VM control block.
1193	*/
1194	static void hmR0SvmFlushTaggedTlb(PHMPHYSCPU pHostCpu, PVMCPUCC pVCpu, PSVMVMCB pVmcb)
1195	{
1196	/*
1197	* Force a TLB flush for the first world switch if the current CPU differs from the one
1198	* we ran on last. This can happen both for start & resume due to long jumps back to
1199	* ring-3.
1200	*
1201	* We also force a TLB flush every time when executing a nested-guest VCPU as there is no
1202	* correlation between it and the physical CPU.
1203	*
1204	* If the TLB flush count changed, another VM (VCPU rather) has hit the ASID limit while
1205	* flushing the TLB, so we cannot reuse the ASIDs without flushing.
1206	*/
1207	bool fNewAsid = false;
1208	Assert(pHostCpu->idCpu != NIL_RTCPUID);
1209	if ( pVCpu->hm.s.idLastCpu != pHostCpu->idCpu
1210	\|\| pVCpu->hm.s.cTlbFlushes != pHostCpu->cTlbFlushes
1211	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1212	\|\| CPUMIsGuestInSvmNestedHwVirtMode(&pVCpu->cpum.GstCtx)
1213	#endif
1214	)
1215	{
1216	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlbWorldSwitch);
1217	pVCpu->hm.s.fForceTLBFlush = true;
1218	fNewAsid = true;
1219	}
1220
1221	/* Set TLB flush state as checked until we return from the world switch. */
1222	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, true);
1223
1224	/* Check for explicit TLB flushes. */
1225	if (VMCPU_FF_TEST_AND_CLEAR(pVCpu, VMCPU_FF_TLB_FLUSH))
1226	{
1227	pVCpu->hm.s.fForceTLBFlush = true;
1228	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlb);
1229	}
1230
1231	/*
1232	* If the AMD CPU erratum 170, We need to flush the entire TLB for each world switch. Sad.
1233	* This Host CPU requirement takes precedence.
1234	*/
1235	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
1236	if (pVM->hm.s.svm.fAlwaysFlushTLB)
1237	{
1238	pHostCpu->uCurrentAsid = 1;
1239	pVCpu->hm.s.uCurrentAsid = 1;
1240	pVCpu->hm.s.cTlbFlushes = pHostCpu->cTlbFlushes;
1241	pVCpu->hm.s.idLastCpu = pHostCpu->idCpu;
1242	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
1243
1244	/* Clear the VMCB Clean Bit for NP while flushing the TLB. See @bugref{7152}. */
1245	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
1246	}
1247	else
1248	{
1249	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_NOTHING;
1250	if (pVCpu->hm.s.fForceTLBFlush)
1251	{
1252	/* Clear the VMCB Clean Bit for NP while flushing the TLB. See @bugref{7152}. */
1253	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
1254
1255	if (fNewAsid)
1256	{
1257	++pHostCpu->uCurrentAsid;
1258
1259	bool fHitASIDLimit = false;
1260	if (pHostCpu->uCurrentAsid >= pVM->hm.s.uMaxAsid)
1261	{
1262	pHostCpu->uCurrentAsid = 1; /* Wraparound at 1; host uses 0 */
1263	pHostCpu->cTlbFlushes++; /* All VCPUs that run on this host CPU must use a new ASID. */
1264	fHitASIDLimit = true;
1265	}
1266
1267	if ( fHitASIDLimit
1268	\|\| pHostCpu->fFlushAsidBeforeUse)
1269	{
1270	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
1271	pHostCpu->fFlushAsidBeforeUse = false;
1272	}
1273
1274	pVCpu->hm.s.uCurrentAsid = pHostCpu->uCurrentAsid;
1275	pVCpu->hm.s.idLastCpu = pHostCpu->idCpu;
1276	pVCpu->hm.s.cTlbFlushes = pHostCpu->cTlbFlushes;
1277	}
1278	else
1279	{
1280	if (pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_FLUSH_BY_ASID)
1281	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_SINGLE_CONTEXT;
1282	else
1283	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
1284	}
1285
1286	pVCpu->hm.s.fForceTLBFlush = false;
1287	}
1288	}
1289
1290	/* Update VMCB with the ASID. */
1291	if (pVmcb->ctrl.TLBCtrl.n.u32ASID != pVCpu->hm.s.uCurrentAsid)
1292	{
1293	pVmcb->ctrl.TLBCtrl.n.u32ASID = pVCpu->hm.s.uCurrentAsid;
1294	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_ASID;
1295	}
1296
1297	AssertMsg(pVCpu->hm.s.idLastCpu == pHostCpu->idCpu,
1298	("vcpu idLastCpu=%u hostcpu idCpu=%u\n", pVCpu->hm.s.idLastCpu, pHostCpu->idCpu));
1299	AssertMsg(pVCpu->hm.s.cTlbFlushes == pHostCpu->cTlbFlushes,
1300	("Flush count mismatch for cpu %u (%u vs %u)\n", pHostCpu->idCpu, pVCpu->hm.s.cTlbFlushes, pHostCpu->cTlbFlushes));
1301	AssertMsg(pHostCpu->uCurrentAsid >= 1 && pHostCpu->uCurrentAsid < pVM->hm.s.uMaxAsid,
1302	("cpu%d uCurrentAsid = %x\n", pHostCpu->idCpu, pHostCpu->uCurrentAsid));
1303	AssertMsg(pVCpu->hm.s.uCurrentAsid >= 1 && pVCpu->hm.s.uCurrentAsid < pVM->hm.s.uMaxAsid,
1304	("cpu%d VM uCurrentAsid = %x\n", pHostCpu->idCpu, pVCpu->hm.s.uCurrentAsid));
1305
1306	#ifdef VBOX_WITH_STATISTICS
1307	if (pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_NOTHING)
1308	STAM_COUNTER_INC(&pVCpu->hm.s.StatNoFlushTlbWorldSwitch);
1309	else if ( pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_SINGLE_CONTEXT
1310	\|\| pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_SINGLE_CONTEXT_RETAIN_GLOBALS)
1311	{
1312	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushAsid);
1313	}
1314	else
1315	{
1316	Assert(pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_ENTIRE);
1317	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushEntire);
1318	}
1319	#endif
1320	}
1321
1322
1323	/**
1324	* Sets an exception intercept in the specified VMCB.
1325	*
1326	* @param pVmcb Pointer to the VM control block.
1327	* @param uXcpt The exception (X86_XCPT_*).
1328	*/
1329	DECLINLINE(void) hmR0SvmSetXcptIntercept(PSVMVMCB pVmcb, uint8_t uXcpt)
1330	{
1331	if (!(pVmcb->ctrl.u32InterceptXcpt & RT_BIT(uXcpt)))
1332	{
1333	pVmcb->ctrl.u32InterceptXcpt \|= RT_BIT(uXcpt);
1334	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1335	}
1336	}
1337
1338
1339	/**
1340	* Clears an exception intercept in the specified VMCB.
1341	*
1342	* @param pVCpu The cross context virtual CPU structure.
1343	* @param pVmcb Pointer to the VM control block.
1344	* @param uXcpt The exception (X86_XCPT_*).
1345	*
1346	* @remarks This takes into account if we're executing a nested-guest and only
1347	* removes the exception intercept if both the guest -and- nested-guest
1348	* are not intercepting it.
1349	*/
1350	DECLINLINE(void) hmR0SvmClearXcptIntercept(PVMCPUCC pVCpu, PSVMVMCB pVmcb, uint8_t uXcpt)
1351	{
1352	Assert(uXcpt != X86_XCPT_DB);
1353	Assert(uXcpt != X86_XCPT_AC);
1354	Assert(uXcpt != X86_XCPT_GP);
1355	#ifndef HMSVM_ALWAYS_TRAP_ALL_XCPTS
1356	if (pVmcb->ctrl.u32InterceptXcpt & RT_BIT(uXcpt))
1357	{
1358	bool fRemove = true;
1359	# ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1360	/* Only remove the intercept if the nested-guest is also not intercepting it! */
1361	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1362	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
1363	{
1364	PCSVMNESTEDVMCBCACHE pVmcbNstGstCache = hmR0SvmGetNestedVmcbCache(pVCpu);
1365	fRemove = !(pVmcbNstGstCache->u32InterceptXcpt & RT_BIT(uXcpt));
1366	}
1367	# else
1368	RT_NOREF(pVCpu);
1369	# endif
1370	if (fRemove)
1371	{
1372	pVmcb->ctrl.u32InterceptXcpt &= ~RT_BIT(uXcpt);
1373	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1374	}
1375	}
1376	#else
1377	RT_NOREF3(pVCpu, pVmcb, uXcpt);
1378	#endif
1379	}
1380
1381
1382	/**
1383	* Sets a control intercept in the specified VMCB.
1384	*
1385	* @param pVmcb Pointer to the VM control block.
1386	* @param fCtrlIntercept The control intercept (SVM_CTRL_INTERCEPT_*).
1387	*/
1388	DECLINLINE(void) hmR0SvmSetCtrlIntercept(PSVMVMCB pVmcb, uint64_t fCtrlIntercept)
1389	{
1390	if (!(pVmcb->ctrl.u64InterceptCtrl & fCtrlIntercept))
1391	{
1392	pVmcb->ctrl.u64InterceptCtrl \|= fCtrlIntercept;
1393	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1394	}
1395	}
1396
1397
1398	/**
1399	* Clears a control intercept in the specified VMCB.
1400	*
1401	* @returns @c true if the intercept is still set, @c false otherwise.
1402	* @param pVCpu The cross context virtual CPU structure.
1403	* @param pVmcb Pointer to the VM control block.
1404	* @param fCtrlIntercept The control intercept (SVM_CTRL_INTERCEPT_*).
1405	*
1406	* @remarks This takes into account if we're executing a nested-guest and only
1407	* removes the control intercept if both the guest -and- nested-guest
1408	* are not intercepting it.
1409	*/
1410	static bool hmR0SvmClearCtrlIntercept(PVMCPUCC pVCpu, PSVMVMCB pVmcb, uint64_t fCtrlIntercept)
1411	{
1412	if (pVmcb->ctrl.u64InterceptCtrl & fCtrlIntercept)
1413	{
1414	bool fRemove = true;
1415	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1416	/* Only remove the control intercept if the nested-guest is also not intercepting it! */
1417	if (CPUMIsGuestInSvmNestedHwVirtMode(&pVCpu->cpum.GstCtx))
1418	{
1419	PCSVMNESTEDVMCBCACHE pVmcbNstGstCache = hmR0SvmGetNestedVmcbCache(pVCpu);
1420	fRemove = !(pVmcbNstGstCache->u64InterceptCtrl & fCtrlIntercept);
1421	}
1422	#else
1423	RT_NOREF(pVCpu);
1424	#endif
1425	if (fRemove)
1426	{
1427	pVmcb->ctrl.u64InterceptCtrl &= ~fCtrlIntercept;
1428	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1429	}
1430	}
1431
1432	return RT_BOOL(pVmcb->ctrl.u64InterceptCtrl & fCtrlIntercept);
1433	}
1434
1435
1436	/**
1437	* Exports the guest (or nested-guest) CR0 into the VMCB.
1438	*
1439	* @param pVCpu The cross context virtual CPU structure.
1440	* @param pVmcb Pointer to the VM control block.
1441	*
1442	* @remarks This assumes we always pre-load the guest FPU.
1443	* @remarks No-long-jump zone!!!
1444	*/
1445	static void hmR0SvmExportGuestCR0(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
1446	{
1447	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1448
1449	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1450	uint64_t const uGuestCr0 = pCtx->cr0;
1451	uint64_t uShadowCr0 = uGuestCr0;
1452
1453	/* Always enable caching. */
1454	uShadowCr0 &= ~(X86_CR0_CD \| X86_CR0_NW);
1455
1456	/* When Nested Paging is not available use shadow page tables and intercept #PFs (latter done in SVMR0SetupVM()). */
1457	if (!pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging)
1458	{
1459	uShadowCr0 \|= X86_CR0_PG /* Use shadow page tables. */
1460	\| X86_CR0_WP; /* Guest CPL 0 writes to its read-only pages should cause a #PF #VMEXIT. */
1461	}
1462
1463	/*
1464	* Use the #MF style of legacy-FPU error reporting for now. Although AMD-V has MSRs that
1465	* lets us isolate the host from it, IEM/REM still needs work to emulate it properly,
1466	* see @bugref{7243#c103}.
1467	*/
1468	if (!(uGuestCr0 & X86_CR0_NE))
1469	{
1470	uShadowCr0 \|= X86_CR0_NE;
1471	hmR0SvmSetXcptIntercept(pVmcb, X86_XCPT_MF);
1472	}
1473	else
1474	hmR0SvmClearXcptIntercept(pVCpu, pVmcb, X86_XCPT_MF);
1475
1476	/*
1477	* If the shadow and guest CR0 are identical we can avoid intercepting CR0 reads.
1478	*
1479	* CR0 writes still needs interception as PGM requires tracking paging mode changes,
1480	* see @bugref{6944}.
1481	*
1482	* We also don't ever want to honor weird things like cache disable from the guest.
1483	* However, we can avoid intercepting changes to the TS & MP bits by clearing the CR0
1484	* write intercept below and keeping SVM_CTRL_INTERCEPT_CR0_SEL_WRITE instead.
1485	*/
1486	if (uShadowCr0 == uGuestCr0)
1487	{
1488	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
1489	{
1490	pVmcb->ctrl.u16InterceptRdCRx &= ~RT_BIT(0);
1491	pVmcb->ctrl.u16InterceptWrCRx &= ~RT_BIT(0);
1492	Assert(pVmcb->ctrl.u64InterceptCtrl & SVM_CTRL_INTERCEPT_CR0_SEL_WRITE);
1493	}
1494	else
1495	{
1496	/* If the nested-hypervisor intercepts CR0 reads/writes, we need to continue intercepting them. */
1497	PCSVMNESTEDVMCBCACHE pVmcbNstGstCache = hmR0SvmGetNestedVmcbCache(pVCpu);
1498	pVmcb->ctrl.u16InterceptRdCRx = (pVmcb->ctrl.u16InterceptRdCRx & ~RT_BIT(0))
1499	\| (pVmcbNstGstCache->u16InterceptRdCRx & RT_BIT(0));
1500	pVmcb->ctrl.u16InterceptWrCRx = (pVmcb->ctrl.u16InterceptWrCRx & ~RT_BIT(0))
1501	\| (pVmcbNstGstCache->u16InterceptWrCRx & RT_BIT(0));
1502	}
1503	}
1504	else
1505	{
1506	pVmcb->ctrl.u16InterceptRdCRx \|= RT_BIT(0);
1507	pVmcb->ctrl.u16InterceptWrCRx \|= RT_BIT(0);
1508	}
1509	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1510
1511	Assert(!RT_HI_U32(uShadowCr0));
1512	if (pVmcb->guest.u64CR0 != uShadowCr0)
1513	{
1514	pVmcb->guest.u64CR0 = uShadowCr0;
1515	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1516	}
1517	}
1518
1519
1520	/**
1521	* Exports the guest (or nested-guest) CR3 into the VMCB.
1522	*
1523	* @param pVCpu The cross context virtual CPU structure.
1524	* @param pVmcb Pointer to the VM control block.
1525	*
1526	* @remarks No-long-jump zone!!!
1527	*/
1528	static void hmR0SvmExportGuestCR3(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
1529	{
1530	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1531
1532	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
1533	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1534	if (pVM->hm.s.fNestedPaging)
1535	{
1536	pVmcb->ctrl.u64NestedPagingCR3 = PGMGetHyperCR3(pVCpu);
1537	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
1538	pVmcb->guest.u64CR3 = pCtx->cr3;
1539	Assert(pVmcb->ctrl.u64NestedPagingCR3);
1540	}
1541	else
1542	pVmcb->guest.u64CR3 = PGMGetHyperCR3(pVCpu);
1543
1544	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1545	}
1546
1547
1548	/**
1549	* Exports the guest (or nested-guest) CR4 into the VMCB.
1550	*
1551	* @param pVCpu The cross context virtual CPU structure.
1552	* @param pVmcb Pointer to the VM control block.
1553	*
1554	* @remarks No-long-jump zone!!!
1555	*/
1556	static int hmR0SvmExportGuestCR4(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
1557	{
1558	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1559
1560	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1561	uint64_t uShadowCr4 = pCtx->cr4;
1562	if (!pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging)
1563	{
1564	switch (pVCpu->hm.s.enmShadowMode)
1565	{
1566	case PGMMODE_REAL:
1567	case PGMMODE_PROTECTED: /* Protected mode, no paging. */
1568	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1569
1570	case PGMMODE_32_BIT: /* 32-bit paging. */
1571	uShadowCr4 &= ~X86_CR4_PAE;
1572	break;
1573
1574	case PGMMODE_PAE: /* PAE paging. */
1575	case PGMMODE_PAE_NX: /* PAE paging with NX enabled. */
1576	/** Must use PAE paging as we could use physical memory > 4 GB */
1577	uShadowCr4 \|= X86_CR4_PAE;
1578	break;
1579
1580	case PGMMODE_AMD64: /* 64-bit AMD paging (long mode). */
1581	case PGMMODE_AMD64_NX: /* 64-bit AMD paging (long mode) with NX enabled. */
1582	#ifdef VBOX_WITH_64_BITS_GUESTS
1583	break;
1584	#else
1585	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1586	#endif
1587
1588	default: /* shut up gcc */
1589	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1590	}
1591	}
1592
1593	/* Whether to save/load/restore XCR0 during world switch depends on CR4.OSXSAVE and host+guest XCR0. */
1594	pVCpu->hm.s.fLoadSaveGuestXcr0 = (pCtx->cr4 & X86_CR4_OSXSAVE) && pCtx->aXcr[0] != ASMGetXcr0();
1595
1596	/* Avoid intercepting CR4 reads if the guest and shadow CR4 values are identical. */
1597	if (uShadowCr4 == pCtx->cr4)
1598	{
1599	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
1600	pVmcb->ctrl.u16InterceptRdCRx &= ~RT_BIT(4);
1601	else
1602	{
1603	/* If the nested-hypervisor intercepts CR4 reads, we need to continue intercepting them. */
1604	PCSVMNESTEDVMCBCACHE pVmcbNstGstCache = hmR0SvmGetNestedVmcbCache(pVCpu);
1605	pVmcb->ctrl.u16InterceptRdCRx = (pVmcb->ctrl.u16InterceptRdCRx & ~RT_BIT(4))
1606	\| (pVmcbNstGstCache->u16InterceptRdCRx & RT_BIT(4));
1607	}
1608	}
1609	else
1610	pVmcb->ctrl.u16InterceptRdCRx \|= RT_BIT(4);
1611
1612	/* CR4 writes are always intercepted (both guest, nested-guest) for tracking PGM mode changes. */
1613	Assert(pVmcb->ctrl.u16InterceptWrCRx & RT_BIT(4));
1614
1615	/* Update VMCB with the shadow CR4 the appropriate VMCB clean bits. */
1616	Assert(!RT_HI_U32(uShadowCr4));
1617	pVmcb->guest.u64CR4 = uShadowCr4;
1618	pVmcb->ctrl.u32VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_CRX_EFER \| HMSVM_VMCB_CLEAN_INTERCEPTS);
1619
1620	return VINF_SUCCESS;
1621	}
1622
1623
1624	/**
1625	* Exports the guest (or nested-guest) control registers into the VMCB.
1626	*
1627	* @returns VBox status code.
1628	* @param pVCpu The cross context virtual CPU structure.
1629	* @param pVmcb Pointer to the VM control block.
1630	*
1631	* @remarks No-long-jump zone!!!
1632	*/
1633	static int hmR0SvmExportGuestControlRegs(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
1634	{
1635	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1636
1637	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_CR_MASK)
1638	{
1639	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_CR0)
1640	hmR0SvmExportGuestCR0(pVCpu, pVmcb);
1641
1642	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_CR2)
1643	{
1644	pVmcb->guest.u64CR2 = pVCpu->cpum.GstCtx.cr2;
1645	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CR2;
1646	}
1647
1648	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_CR3)
1649	hmR0SvmExportGuestCR3(pVCpu, pVmcb);
1650
1651	/* CR4 re-loading is ASSUMED to be done everytime we get in from ring-3! (XCR0) */
1652	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_CR4)
1653	{
1654	int rc = hmR0SvmExportGuestCR4(pVCpu, pVmcb);
1655	if (RT_FAILURE(rc))
1656	return rc;
1657	}
1658
1659	pVCpu->hm.s.fCtxChanged &= ~HM_CHANGED_GUEST_CR_MASK;
1660	}
1661	return VINF_SUCCESS;
1662	}
1663
1664
1665	/**
1666	* Exports the guest (or nested-guest) segment registers into the VMCB.
1667	*
1668	* @returns VBox status code.
1669	* @param pVCpu The cross context virtual CPU structure.
1670	* @param pVmcb Pointer to the VM control block.
1671	*
1672	* @remarks No-long-jump zone!!!
1673	*/
1674	static void hmR0SvmExportGuestSegmentRegs(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
1675	{
1676	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1677	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1678
1679	/* Guest segment registers. */
1680	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SREG_MASK)
1681	{
1682	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_CS)
1683	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, CS, cs);
1684
1685	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SS)
1686	{
1687	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, SS, ss);
1688	pVmcb->guest.u8CPL = pCtx->ss.Attr.n.u2Dpl;
1689	}
1690
1691	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_DS)
1692	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, DS, ds);
1693
1694	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_ES)
1695	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, ES, es);
1696
1697	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_FS)
1698	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, FS, fs);
1699
1700	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_GS)
1701	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, GS, gs);
1702
1703	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_SEG;
1704	}
1705
1706	/* Guest TR. */
1707	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_TR)
1708	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, TR, tr);
1709
1710	/* Guest LDTR. */
1711	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_LDTR)
1712	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, LDTR, ldtr);
1713
1714	/* Guest GDTR. */
1715	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_GDTR)
1716	{
1717	pVmcb->guest.GDTR.u32Limit = pCtx->gdtr.cbGdt;
1718	pVmcb->guest.GDTR.u64Base = pCtx->gdtr.pGdt;
1719	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DT;
1720	}
1721
1722	/* Guest IDTR. */
1723	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_IDTR)
1724	{
1725	pVmcb->guest.IDTR.u32Limit = pCtx->idtr.cbIdt;
1726	pVmcb->guest.IDTR.u64Base = pCtx->idtr.pIdt;
1727	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DT;
1728	}
1729
1730	pVCpu->hm.s.fCtxChanged &= ~( HM_CHANGED_GUEST_SREG_MASK
1731	\| HM_CHANGED_GUEST_TABLE_MASK);
1732	}
1733
1734
1735	/**
1736	* Exports the guest (or nested-guest) MSRs into the VMCB.
1737	*
1738	* @param pVCpu The cross context virtual CPU structure.
1739	* @param pVmcb Pointer to the VM control block.
1740	*
1741	* @remarks No-long-jump zone!!!
1742	*/
1743	static void hmR0SvmExportGuestMsrs(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
1744	{
1745	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1746	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1747
1748	/* Guest Sysenter MSRs. */
1749	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SYSENTER_MSR_MASK)
1750	{
1751	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SYSENTER_CS_MSR)
1752	pVmcb->guest.u64SysEnterCS = pCtx->SysEnter.cs;
1753
1754	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SYSENTER_EIP_MSR)
1755	pVmcb->guest.u64SysEnterEIP = pCtx->SysEnter.eip;
1756
1757	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SYSENTER_ESP_MSR)
1758	pVmcb->guest.u64SysEnterESP = pCtx->SysEnter.esp;
1759	}
1760
1761	/*
1762	* Guest EFER MSR.
1763	* AMD-V requires guest EFER.SVME to be set. Weird.
1764	* See AMD spec. 15.5.1 "Basic Operation" \| "Canonicalization and Consistency Checks".
1765	*/
1766	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_EFER_MSR)
1767	{
1768	pVmcb->guest.u64EFER = pCtx->msrEFER \| MSR_K6_EFER_SVME;
1769	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1770	}
1771
1772	/* If the guest isn't in 64-bit mode, clear MSR_K6_LME bit, otherwise SVM expects amd64 shadow paging. */
1773	if ( !CPUMIsGuestInLongModeEx(pCtx)
1774	&& (pCtx->msrEFER & MSR_K6_EFER_LME))
1775	{
1776	pVmcb->guest.u64EFER &= ~MSR_K6_EFER_LME;
1777	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1778	}
1779
1780	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SYSCALL_MSRS)
1781	{
1782	pVmcb->guest.u64STAR = pCtx->msrSTAR;
1783	pVmcb->guest.u64LSTAR = pCtx->msrLSTAR;
1784	pVmcb->guest.u64CSTAR = pCtx->msrCSTAR;
1785	pVmcb->guest.u64SFMASK = pCtx->msrSFMASK;
1786	}
1787
1788	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_KERNEL_GS_BASE)
1789	pVmcb->guest.u64KernelGSBase = pCtx->msrKERNELGSBASE;
1790
1791	pVCpu->hm.s.fCtxChanged &= ~( HM_CHANGED_GUEST_SYSENTER_MSR_MASK
1792	\| HM_CHANGED_GUEST_EFER_MSR
1793	\| HM_CHANGED_GUEST_SYSCALL_MSRS
1794	\| HM_CHANGED_GUEST_KERNEL_GS_BASE);
1795
1796	/*
1797	* Setup the PAT MSR (applicable for Nested Paging only).
1798	*
1799	* The default value should be MSR_IA32_CR_PAT_INIT_VAL, but we treat all guest memory
1800	* as WB, so choose type 6 for all PAT slots, see @bugref{9634}.
1801	*
1802	* While guests can modify and see the modified values through the shadow values,
1803	* we shall not honor any guest modifications of this MSR to ensure caching is always
1804	* enabled similar to how we clear CR0.CD and NW bits.
1805	*
1806	* For nested-guests this needs to always be set as well, see @bugref{7243#c109}.
1807	*/
1808	pVmcb->guest.u64PAT = UINT64_C(0x0006060606060606);
1809
1810	/* Enable the last branch record bit if LBR virtualization is enabled. */
1811	if (pVmcb->ctrl.LbrVirt.n.u1LbrVirt)
1812	pVmcb->guest.u64DBGCTL = MSR_IA32_DEBUGCTL_LBR;
1813	}
1814
1815
1816	/**
1817	* Exports the guest (or nested-guest) debug state into the VMCB and programs
1818	* the necessary intercepts accordingly.
1819	*
1820	* @param pVCpu The cross context virtual CPU structure.
1821	* @param pVmcb Pointer to the VM control block.
1822	*
1823	* @remarks No-long-jump zone!!!
1824	* @remarks Requires EFLAGS to be up-to-date in the VMCB!
1825	*/
1826	static void hmR0SvmExportSharedDebugState(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
1827	{
1828	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1829
1830	/*
1831	* Anyone single stepping on the host side? If so, we'll have to use the
1832	* trap flag in the guest EFLAGS since AMD-V doesn't have a trap flag on
1833	* the VMM level like the VT-x implementations does.
1834	*/
1835	bool fInterceptMovDRx = false;
1836	bool const fStepping = pVCpu->hm.s.fSingleInstruction \|\| DBGFIsStepping(pVCpu);
1837	if (fStepping)
1838	{
1839	pVCpu->hm.s.fClearTrapFlag = true;
1840	pVmcb->guest.u64RFlags \|= X86_EFL_TF;
1841	fInterceptMovDRx = true; /* Need clean DR6, no guest mess. */
1842	}
1843
1844	if ( fStepping
1845	\|\| (CPUMGetHyperDR7(pVCpu) & X86_DR7_ENABLED_MASK))
1846	{
1847	/*
1848	* Use the combined guest and host DRx values found in the hypervisor
1849	* register set because the debugger has breakpoints active or someone
1850	* is single stepping on the host side.
1851	*
1852	* Note! DBGF expects a clean DR6 state before executing guest code.
1853	*/
1854	if (!CPUMIsHyperDebugStateActive(pVCpu))
1855	{
1856	CPUMR0LoadHyperDebugState(pVCpu, false /* include DR6 */);
1857	Assert(!CPUMIsGuestDebugStateActive(pVCpu));
1858	Assert(CPUMIsHyperDebugStateActive(pVCpu));
1859	}
1860
1861	/* Update DR6 & DR7. (The other DRx values are handled by CPUM one way or the other.) */
1862	if ( pVmcb->guest.u64DR6 != X86_DR6_INIT_VAL
1863	\|\| pVmcb->guest.u64DR7 != CPUMGetHyperDR7(pVCpu))
1864	{
1865	pVmcb->guest.u64DR7 = CPUMGetHyperDR7(pVCpu);
1866	pVmcb->guest.u64DR6 = X86_DR6_INIT_VAL;
1867	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
1868	}
1869
1870	/** @todo If we cared, we could optimize to allow the guest to read registers
1871	* with the same values. */
1872	fInterceptMovDRx = true;
1873	pVCpu->hm.s.fUsingHyperDR7 = true;
1874	Log5(("hmR0SvmExportSharedDebugState: Loaded hyper DRx\n"));
1875	}
1876	else
1877	{
1878	/*
1879	* Update DR6, DR7 with the guest values if necessary.
1880	*/
1881	if ( pVmcb->guest.u64DR7 != pCtx->dr[7]
1882	\|\| pVmcb->guest.u64DR6 != pCtx->dr[6])
1883	{
1884	pVmcb->guest.u64DR7 = pCtx->dr[7];
1885	pVmcb->guest.u64DR6 = pCtx->dr[6];
1886	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
1887	}
1888	pVCpu->hm.s.fUsingHyperDR7 = false;
1889
1890	/*
1891	* If the guest has enabled debug registers, we need to load them prior to
1892	* executing guest code so they'll trigger at the right time.
1893	*/
1894	if (pCtx->dr[7] & (X86_DR7_ENABLED_MASK \| X86_DR7_GD)) /** @todo Why GD? */
1895	{
1896	if (!CPUMIsGuestDebugStateActive(pVCpu))
1897	{
1898	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
1899	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxArmed);
1900	Assert(!CPUMIsHyperDebugStateActive(pVCpu));
1901	Assert(CPUMIsGuestDebugStateActive(pVCpu));
1902	}
1903	Log5(("hmR0SvmExportSharedDebugState: Loaded guest DRx\n"));
1904	}
1905	/*
1906	* If no debugging enabled, we'll lazy load DR0-3. We don't need to
1907	* intercept #DB as DR6 is updated in the VMCB.
1908	*
1909	* Note! If we cared and dared, we could skip intercepting \#DB here.
1910	* However, \#DB shouldn't be performance critical, so we'll play safe
1911	* and keep the code similar to the VT-x code and always intercept it.
1912	*/
1913	else if (!CPUMIsGuestDebugStateActive(pVCpu))
1914	fInterceptMovDRx = true;
1915	}
1916
1917	Assert(pVmcb->ctrl.u32InterceptXcpt & RT_BIT_32(X86_XCPT_DB));
1918	if (fInterceptMovDRx)
1919	{
1920	if ( pVmcb->ctrl.u16InterceptRdDRx != 0xffff
1921	\|\| pVmcb->ctrl.u16InterceptWrDRx != 0xffff)
1922	{
1923	pVmcb->ctrl.u16InterceptRdDRx = 0xffff;
1924	pVmcb->ctrl.u16InterceptWrDRx = 0xffff;
1925	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1926	}
1927	}
1928	else
1929	{
1930	if ( pVmcb->ctrl.u16InterceptRdDRx
1931	\|\| pVmcb->ctrl.u16InterceptWrDRx)
1932	{
1933	pVmcb->ctrl.u16InterceptRdDRx = 0;
1934	pVmcb->ctrl.u16InterceptWrDRx = 0;
1935	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1936	}
1937	}
1938	Log4Func(("DR6=%#RX64 DR7=%#RX64\n", pCtx->dr[6], pCtx->dr[7]));
1939	}
1940
1941	/**
1942	* Exports the hardware virtualization state into the nested-guest
1943	* VMCB.
1944	*
1945	* @param pVCpu The cross context virtual CPU structure.
1946	* @param pVmcb Pointer to the VM control block.
1947	*
1948	* @remarks No-long-jump zone!!!
1949	*/
1950	static void hmR0SvmExportGuestHwvirtState(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
1951	{
1952	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1953
1954	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_HWVIRT)
1955	{
1956	if (pVmcb->ctrl.IntCtrl.n.u1VGifEnable)
1957	{
1958	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1959	PCVM pVM = pVCpu->CTX_SUFF(pVM);
1960
1961	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(pCtx); /* Nested VGIF is not supported yet. */
1962	Assert(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_VGIF); /* Physical hardware supports VGIF. */
1963	Assert(HMIsSvmVGifActive(pVM)); /* Outer VM has enabled VGIF. */
1964	NOREF(pVM);
1965
1966	pVmcb->ctrl.IntCtrl.n.u1VGif = CPUMGetGuestGif(pCtx);
1967	}
1968
1969	/*
1970	* Ensure the nested-guest pause-filter counters don't exceed the outer guest values esp.
1971	* since SVM doesn't have a preemption timer.
1972	*
1973	* We do this here rather than in hmR0SvmSetupVmcbNested() as we may have been executing the
1974	* nested-guest in IEM incl. PAUSE instructions which would update the pause-filter counters
1975	* and may continue execution in SVM R0 without a nested-guest #VMEXIT in between.
1976	*/
1977	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
1978	PSVMVMCBCTRL pVmcbCtrl = &pVmcb->ctrl;
1979	uint16_t const uGuestPauseFilterCount = pVM->hm.s.svm.cPauseFilter;
1980	uint16_t const uGuestPauseFilterThreshold = pVM->hm.s.svm.cPauseFilterThresholdTicks;
1981	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, &pVCpu->cpum.GstCtx, SVM_CTRL_INTERCEPT_PAUSE))
1982	{
1983	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1984	pVmcbCtrl->u16PauseFilterCount = RT_MIN(pCtx->hwvirt.svm.cPauseFilter, uGuestPauseFilterCount);
1985	pVmcbCtrl->u16PauseFilterThreshold = RT_MIN(pCtx->hwvirt.svm.cPauseFilterThreshold, uGuestPauseFilterThreshold);
1986	}
1987	else
1988	{
1989	/** @todo r=ramshankar: We can turn these assignments into assertions. */
1990	pVmcbCtrl->u16PauseFilterCount = uGuestPauseFilterCount;
1991	pVmcbCtrl->u16PauseFilterThreshold = uGuestPauseFilterThreshold;
1992	}
1993	pVmcbCtrl->u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1994
1995	pVCpu->hm.s.fCtxChanged &= ~HM_CHANGED_GUEST_HWVIRT;
1996	}
1997	}
1998
1999
2000	/**
2001	* Exports the guest APIC TPR state into the VMCB.
2002	*
2003	* @returns VBox status code.
2004	* @param pVCpu The cross context virtual CPU structure.
2005	* @param pVmcb Pointer to the VM control block.
2006	*/
2007	static int hmR0SvmExportGuestApicTpr(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
2008	{
2009	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(&pVCpu->cpum.GstCtx);
2010
2011	if (ASMAtomicUoReadU64(&pVCpu->hm.s.fCtxChanged) & HM_CHANGED_GUEST_APIC_TPR)
2012	{
2013	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
2014	if ( PDMHasApic(pVM)
2015	&& APICIsEnabled(pVCpu))
2016	{
2017	bool fPendingIntr;
2018	uint8_t u8Tpr;
2019	int rc = APICGetTpr(pVCpu, &u8Tpr, &fPendingIntr, NULL /* pu8PendingIrq */);
2020	AssertRCReturn(rc, rc);
2021
2022	/* Assume that we need to trap all TPR accesses and thus need not check on
2023	every #VMEXIT if we should update the TPR. */
2024	Assert(pVmcb->ctrl.IntCtrl.n.u1VIntrMasking);
2025	pVCpu->hm.s.svm.fSyncVTpr = false;
2026
2027	if (!pVM->hm.s.fTPRPatchingActive)
2028	{
2029	/* Bits 3-0 of the VTPR field correspond to bits 7-4 of the TPR (which is the Task-Priority Class). */
2030	pVmcb->ctrl.IntCtrl.n.u8VTPR = (u8Tpr >> 4);
2031
2032	/* If there are interrupts pending, intercept CR8 writes to evaluate ASAP if we
2033	can deliver the interrupt to the guest. */
2034	if (fPendingIntr)
2035	pVmcb->ctrl.u16InterceptWrCRx \|= RT_BIT(8);
2036	else
2037	{
2038	pVmcb->ctrl.u16InterceptWrCRx &= ~RT_BIT(8);
2039	pVCpu->hm.s.svm.fSyncVTpr = true;
2040	}
2041
2042	pVmcb->ctrl.u32VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS \| HMSVM_VMCB_CLEAN_INT_CTRL);
2043	}
2044	else
2045	{
2046	/* 32-bit guests uses LSTAR MSR for patching guest code which touches the TPR. */
2047	pVmcb->guest.u64LSTAR = u8Tpr;
2048	uint8_t pbMsrBitmap = (uint8_t )pVCpu->hm.s.svm.pvMsrBitmap;
2049
2050	/* If there are interrupts pending, intercept LSTAR writes, otherwise don't intercept reads or writes. */
2051	if (fPendingIntr)
2052	hmR0SvmSetMsrPermission(pVCpu, pbMsrBitmap, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_INTERCEPT_WRITE);
2053	else
2054	{
2055	hmR0SvmSetMsrPermission(pVCpu, pbMsrBitmap, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
2056	pVCpu->hm.s.svm.fSyncVTpr = true;
2057	}
2058	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_IOPM_MSRPM;
2059	}
2060	}
2061	ASMAtomicUoAndU64(&pVCpu->hm.s.fCtxChanged, ~HM_CHANGED_GUEST_APIC_TPR);
2062	}
2063	return VINF_SUCCESS;
2064	}
2065
2066
2067	/**
2068	* Sets up the exception interrupts required for guest execution in the VMCB.
2069	*
2070	* @param pVCpu The cross context virtual CPU structure.
2071	* @param pVmcb Pointer to the VM control block.
2072	*
2073	* @remarks No-long-jump zone!!!
2074	*/
2075	static void hmR0SvmExportGuestXcptIntercepts(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
2076	{
2077	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(&pVCpu->cpum.GstCtx);
2078
2079	/* If we modify intercepts from here, please check & adjust hmR0SvmMergeVmcbCtrlsNested() if required. */
2080	if (ASMAtomicUoReadU64(&pVCpu->hm.s.fCtxChanged) & HM_CHANGED_SVM_XCPT_INTERCEPTS)
2081	{
2082	/* Trap #UD for GIM provider (e.g. for hypercalls). */
2083	if (pVCpu->hm.s.fGIMTrapXcptUD)
2084	hmR0SvmSetXcptIntercept(pVmcb, X86_XCPT_UD);
2085	else
2086	hmR0SvmClearXcptIntercept(pVCpu, pVmcb, X86_XCPT_UD);
2087
2088	/* Trap #BP for INT3 debug breakpoints set by the VM debugger. */
2089	if (pVCpu->CTX_SUFF(pVM)->dbgf.ro.cEnabledInt3Breakpoints)
2090	hmR0SvmSetXcptIntercept(pVmcb, X86_XCPT_BP);
2091	else
2092	hmR0SvmClearXcptIntercept(pVCpu, pVmcb, X86_XCPT_BP);
2093
2094	/* The remaining intercepts are handled elsewhere, e.g. in hmR0SvmExportGuestCR0(). */
2095	ASMAtomicUoAndU64(&pVCpu->hm.s.fCtxChanged, ~HM_CHANGED_SVM_XCPT_INTERCEPTS);
2096	}
2097	}
2098
2099
2100	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
2101	/**
2102	* Merges guest and nested-guest intercepts for executing the nested-guest using
2103	* hardware-assisted SVM.
2104	*
2105	* This merges the guest and nested-guest intercepts in a way that if the outer
2106	* guest intercept is set we need to intercept it in the nested-guest as
2107	* well.
2108	*
2109	* @param pVCpu The cross context virtual CPU structure.
2110	* @param pVmcbNstGst Pointer to the nested-guest VM control block.
2111	*/
2112	static void hmR0SvmMergeVmcbCtrlsNested(PVMCPUCC pVCpu)
2113	{
2114	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
2115	PCSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
2116	PSVMVMCB pVmcbNstGst = pVCpu->cpum.GstCtx.hwvirt.svm.CTX_SUFF(pVmcb);
2117	PSVMVMCBCTRL pVmcbNstGstCtrl = &pVmcbNstGst->ctrl;
2118
2119	/* Merge the guest's CR intercepts into the nested-guest VMCB. */
2120	pVmcbNstGstCtrl->u16InterceptRdCRx \|= pVmcb->ctrl.u16InterceptRdCRx;
2121	pVmcbNstGstCtrl->u16InterceptWrCRx \|= pVmcb->ctrl.u16InterceptWrCRx;
2122
2123	/* Always intercept CR4 writes for tracking PGM mode changes. */
2124	pVmcbNstGstCtrl->u16InterceptWrCRx \|= RT_BIT(4);
2125
2126	/* Without nested paging, intercept CR3 reads and writes as we load shadow page tables. */
2127	if (!pVM->hm.s.fNestedPaging)
2128	{
2129	pVmcbNstGstCtrl->u16InterceptRdCRx \|= RT_BIT(3);
2130	pVmcbNstGstCtrl->u16InterceptWrCRx \|= RT_BIT(3);
2131	}
2132
2133	/** @todo Figure out debugging with nested-guests, till then just intercept
2134	* all DR[0-15] accesses. */
2135	pVmcbNstGstCtrl->u16InterceptRdDRx \|= 0xffff;
2136	pVmcbNstGstCtrl->u16InterceptWrDRx \|= 0xffff;
2137
2138	/*
2139	* Merge the guest's exception intercepts into the nested-guest VMCB.
2140	*
2141	* - #UD: Exclude these as the outer guest's GIM hypercalls are not applicable
2142	* while executing the nested-guest.
2143	*
2144	* - #BP: Exclude breakpoints set by the VM debugger for the outer guest. This can
2145	* be tweaked later depending on how we wish to implement breakpoints.
2146	*
2147	* - #GP: Exclude these as it's the inner VMMs problem to get vmsvga 3d drivers
2148	* loaded into their guests, not ours.
2149	*
2150	* Warning!! This ASSUMES we only intercept \#UD for hypercall purposes and \#BP
2151	* for VM debugger breakpoints, see hmR0SvmExportGuestXcptIntercepts().
2152	*/
2153	#ifndef HMSVM_ALWAYS_TRAP_ALL_XCPTS
2154	pVmcbNstGstCtrl->u32InterceptXcpt \|= pVmcb->ctrl.u32InterceptXcpt
2155	& ~( RT_BIT(X86_XCPT_UD)
2156	\| RT_BIT(X86_XCPT_BP)
2157	\| (pVCpu->hm.s.fTrapXcptGpForLovelyMesaDrv ? RT_BIT(X86_XCPT_GP) : 0));
2158	#else
2159	pVmcbNstGstCtrl->u32InterceptXcpt \|= pVmcb->ctrl.u32InterceptXcpt;
2160	#endif
2161
2162	/*
2163	* Adjust intercepts while executing the nested-guest that differ from the
2164	* outer guest intercepts.
2165	*
2166	* - VINTR: Exclude the outer guest intercept as we don't need to cause VINTR #VMEXITs
2167	* that belong to the nested-guest to the outer guest.
2168	*
2169	* - VMMCALL: Exclude the outer guest intercept as when it's also not intercepted by
2170	* the nested-guest, the physical CPU raises a \#UD exception as expected.
2171	*/
2172	pVmcbNstGstCtrl->u64InterceptCtrl \|= (pVmcb->ctrl.u64InterceptCtrl & ~( SVM_CTRL_INTERCEPT_VINTR
2173	\| SVM_CTRL_INTERCEPT_VMMCALL))
2174	\| HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS;
2175
2176	Assert( (pVmcbNstGstCtrl->u64InterceptCtrl & HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS)
2177	== HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS);
2178
2179	/* Finally, update the VMCB clean bits. */
2180	pVmcbNstGstCtrl->u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2181	}
2182	#endif
2183
2184
2185	/**
2186	* Enters the AMD-V session.
2187	*
2188	* @returns VBox status code.
2189	* @param pVCpu The cross context virtual CPU structure.
2190	*/
2191	VMMR0DECL(int) SVMR0Enter(PVMCPUCC pVCpu)
2192	{
2193	AssertPtr(pVCpu);
2194	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.svm.fSupported);
2195	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2196
2197	LogFlowFunc(("pVCpu=%p\n", pVCpu));
2198	Assert((pVCpu->hm.s.fCtxChanged & (HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE))
2199	== (HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE));
2200
2201	pVCpu->hm.s.fLeaveDone = false;
2202	return VINF_SUCCESS;
2203	}
2204
2205
2206	/**
2207	* Thread-context callback for AMD-V.
2208	*
2209	* @param enmEvent The thread-context event.
2210	* @param pVCpu The cross context virtual CPU structure.
2211	* @param fGlobalInit Whether global VT-x/AMD-V init. is used.
2212	* @thread EMT(pVCpu)
2213	*/
2214	VMMR0DECL(void) SVMR0ThreadCtxCallback(RTTHREADCTXEVENT enmEvent, PVMCPUCC pVCpu, bool fGlobalInit)
2215	{
2216	NOREF(fGlobalInit);
2217
2218	switch (enmEvent)
2219	{
2220	case RTTHREADCTXEVENT_OUT:
2221	{
2222	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2223	Assert(VMMR0ThreadCtxHookIsEnabled(pVCpu));
2224	VMCPU_ASSERT_EMT(pVCpu);
2225
2226	/* No longjmps (log-flush, locks) in this fragile context. */
2227	VMMRZCallRing3Disable(pVCpu);
2228
2229	if (!pVCpu->hm.s.fLeaveDone)
2230	{
2231	hmR0SvmLeave(pVCpu, false /* fImportState */);
2232	pVCpu->hm.s.fLeaveDone = true;
2233	}
2234
2235	/* Leave HM context, takes care of local init (term). */
2236	int rc = HMR0LeaveCpu(pVCpu);
2237	AssertRC(rc); NOREF(rc);
2238
2239	/* Restore longjmp state. */
2240	VMMRZCallRing3Enable(pVCpu);
2241	STAM_REL_COUNTER_INC(&pVCpu->hm.s.StatSwitchPreempt);
2242	break;
2243	}
2244
2245	case RTTHREADCTXEVENT_IN:
2246	{
2247	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2248	Assert(VMMR0ThreadCtxHookIsEnabled(pVCpu));
2249	VMCPU_ASSERT_EMT(pVCpu);
2250
2251	/* No longjmps (log-flush, locks) in this fragile context. */
2252	VMMRZCallRing3Disable(pVCpu);
2253
2254	/*
2255	* Initialize the bare minimum state required for HM. This takes care of
2256	* initializing AMD-V if necessary (onlined CPUs, local init etc.)
2257	*/
2258	int rc = hmR0EnterCpu(pVCpu);
2259	AssertRC(rc); NOREF(rc);
2260	Assert((pVCpu->hm.s.fCtxChanged & (HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE))
2261	== (HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE));
2262
2263	pVCpu->hm.s.fLeaveDone = false;
2264
2265	/* Restore longjmp state. */
2266	VMMRZCallRing3Enable(pVCpu);
2267	break;
2268	}
2269
2270	default:
2271	break;
2272	}
2273	}
2274
2275
2276	/**
2277	* Saves the host state.
2278	*
2279	* @returns VBox status code.
2280	* @param pVCpu The cross context virtual CPU structure.
2281	*
2282	* @remarks No-long-jump zone!!!
2283	*/
2284	VMMR0DECL(int) SVMR0ExportHostState(PVMCPUCC pVCpu)
2285	{
2286	NOREF(pVCpu);
2287
2288	/* Nothing to do here. AMD-V does this for us automatically during the world-switch. */
2289	ASMAtomicUoAndU64(&pVCpu->hm.s.fCtxChanged, ~HM_CHANGED_HOST_CONTEXT);
2290	return VINF_SUCCESS;
2291	}
2292
2293
2294	/**
2295	* Exports the guest or nested-guest state from the virtual-CPU context into the
2296	* VMCB.
2297	*
2298	* Also sets up the appropriate VMRUN function to execute guest or nested-guest
2299	* code based on the virtual-CPU mode.
2300	*
2301	* @returns VBox status code.
2302	* @param pVCpu The cross context virtual CPU structure.
2303	* @param pSvmTransient Pointer to the SVM-transient structure.
2304	*
2305	* @remarks No-long-jump zone!!!
2306	*/
2307	static int hmR0SvmExportGuestState(PVMCPUCC pVCpu, PCSVMTRANSIENT pSvmTransient)
2308	{
2309	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatExportGuestState, x);
2310
2311	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
2312	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2313	Assert(pVmcb);
2314
2315	pVmcb->guest.u64RIP = pCtx->rip;
2316	pVmcb->guest.u64RSP = pCtx->rsp;
2317	pVmcb->guest.u64RFlags = pCtx->eflags.u32;
2318	pVmcb->guest.u64RAX = pCtx->rax;
2319
2320	bool const fIsNestedGuest = pSvmTransient->fIsNestedGuest;
2321	RTCCUINTREG const fEFlags = ASMIntDisableFlags();
2322
2323	int rc = hmR0SvmExportGuestControlRegs(pVCpu, pVmcb);
2324	AssertRCReturnStmt(rc, ASMSetFlags(fEFlags), rc);
2325	hmR0SvmExportGuestSegmentRegs(pVCpu, pVmcb);
2326	hmR0SvmExportGuestMsrs(pVCpu, pVmcb);
2327	hmR0SvmExportGuestHwvirtState(pVCpu, pVmcb);
2328
2329	ASMSetFlags(fEFlags);
2330
2331	if (!fIsNestedGuest)
2332	{
2333	/* hmR0SvmExportGuestApicTpr() must be called -after- hmR0SvmExportGuestMsrs() as we
2334	otherwise we would overwrite the LSTAR MSR that we use for TPR patching. */
2335	hmR0SvmExportGuestApicTpr(pVCpu, pVmcb);
2336	hmR0SvmExportGuestXcptIntercepts(pVCpu, pVmcb);
2337	}
2338
2339	/* Clear any bits that may be set but exported unconditionally or unused/reserved bits. */
2340	uint64_t fUnusedMask = HM_CHANGED_GUEST_RIP
2341	\| HM_CHANGED_GUEST_RFLAGS
2342	\| HM_CHANGED_GUEST_GPRS_MASK
2343	\| HM_CHANGED_GUEST_X87
2344	\| HM_CHANGED_GUEST_SSE_AVX
2345	\| HM_CHANGED_GUEST_OTHER_XSAVE
2346	\| HM_CHANGED_GUEST_XCRx
2347	\| HM_CHANGED_GUEST_TSC_AUX
2348	\| HM_CHANGED_GUEST_OTHER_MSRS;
2349	if (fIsNestedGuest)
2350	fUnusedMask \|= HM_CHANGED_SVM_XCPT_INTERCEPTS
2351	\| HM_CHANGED_GUEST_APIC_TPR;
2352
2353	ASMAtomicUoAndU64(&pVCpu->hm.s.fCtxChanged, ~( fUnusedMask
2354	\| (HM_CHANGED_KEEPER_STATE_MASK & ~HM_CHANGED_SVM_MASK)));
2355
2356	#ifdef VBOX_STRICT
2357	/*
2358	* All of the guest-CPU state and SVM keeper bits should be exported here by now,
2359	* except for the host-context and/or shared host-guest context bits.
2360	*/
2361	uint64_t const fCtxChanged = ASMAtomicUoReadU64(&pVCpu->hm.s.fCtxChanged);
2362	AssertMsg(!(fCtxChanged & (HM_CHANGED_ALL_GUEST & ~HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE)),
2363	("fCtxChanged=%#RX64\n", fCtxChanged));
2364
2365	/*
2366	* If we need to log state that isn't always imported, we'll need to import them here.
2367	* See hmR0SvmPostRunGuest() for which part of the state is imported uncondtionally.
2368	*/
2369	hmR0SvmLogState(pVCpu, pVmcb, "hmR0SvmExportGuestState", 0 /* fFlags /, 0 / uVerbose */);
2370	#endif
2371
2372	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExportGuestState, x);
2373	return VINF_SUCCESS;
2374	}
2375
2376
2377	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
2378	/**
2379	* Merges the guest and nested-guest MSR permission bitmap.
2380	*
2381	* If the guest is intercepting an MSR we need to intercept it regardless of
2382	* whether the nested-guest is intercepting it or not.
2383	*
2384	* @param pHostCpu The HM physical-CPU structure.
2385	* @param pVCpu The cross context virtual CPU structure.
2386	*
2387	* @remarks No-long-jmp zone!!!
2388	*/
2389	DECLINLINE(void) hmR0SvmMergeMsrpmNested(PHMPHYSCPU pHostCpu, PVMCPUCC pVCpu)
2390	{
2391	uint64_t const pu64GstMsrpm = (uint64_t const )pVCpu->hm.s.svm.pvMsrBitmap;
2392	uint64_t const pu64NstGstMsrpm = (uint64_t const )pVCpu->cpum.GstCtx.hwvirt.svm.CTX_SUFF(pvMsrBitmap);
2393	uint64_t pu64DstMsrpm = (uint64_t )pHostCpu->n.svm.pvNstGstMsrpm;
2394
2395	/* MSRPM bytes from offset 0x1800 are reserved, so we stop merging there. */
2396	uint32_t const offRsvdQwords = 0x1800 >> 3;
2397	for (uint32_t i = 0; i < offRsvdQwords; i++)
2398	pu64DstMsrpm[i] = pu64NstGstMsrpm[i] \| pu64GstMsrpm[i];
2399	}
2400
2401
2402	/**
2403	* Caches the nested-guest VMCB fields before we modify them for execution using
2404	* hardware-assisted SVM.
2405	*
2406	* @returns true if the VMCB was previously already cached, false otherwise.
2407	* @param pVCpu The cross context virtual CPU structure.
2408	*
2409	* @sa HMNotifySvmNstGstVmexit.
2410	*/
2411	static bool hmR0SvmCacheVmcbNested(PVMCPUCC pVCpu)
2412	{
2413	/*
2414	* Cache the nested-guest programmed VMCB fields if we have not cached it yet.
2415	* Otherwise we risk re-caching the values we may have modified, see @bugref{7243#c44}.
2416	*
2417	* Nested-paging CR3 is not saved back into the VMCB on #VMEXIT, hence no need to
2418	* cache and restore it, see AMD spec. 15.25.4 "Nested Paging and VMRUN/#VMEXIT".
2419	*/
2420	PSVMNESTEDVMCBCACHE pVmcbNstGstCache = &pVCpu->hm.s.svm.NstGstVmcbCache;
2421	bool const fWasCached = pVmcbNstGstCache->fCacheValid;
2422	if (!fWasCached)
2423	{
2424	PCSVMVMCB pVmcbNstGst = pVCpu->cpum.GstCtx.hwvirt.svm.CTX_SUFF(pVmcb);
2425	PCSVMVMCBCTRL pVmcbNstGstCtrl = &pVmcbNstGst->ctrl;
2426	pVmcbNstGstCache->u16InterceptRdCRx = pVmcbNstGstCtrl->u16InterceptRdCRx;
2427	pVmcbNstGstCache->u16InterceptWrCRx = pVmcbNstGstCtrl->u16InterceptWrCRx;
2428	pVmcbNstGstCache->u16InterceptRdDRx = pVmcbNstGstCtrl->u16InterceptRdDRx;
2429	pVmcbNstGstCache->u16InterceptWrDRx = pVmcbNstGstCtrl->u16InterceptWrDRx;
2430	pVmcbNstGstCache->u16PauseFilterThreshold = pVmcbNstGstCtrl->u16PauseFilterThreshold;
2431	pVmcbNstGstCache->u16PauseFilterCount = pVmcbNstGstCtrl->u16PauseFilterCount;
2432	pVmcbNstGstCache->u32InterceptXcpt = pVmcbNstGstCtrl->u32InterceptXcpt;
2433	pVmcbNstGstCache->u64InterceptCtrl = pVmcbNstGstCtrl->u64InterceptCtrl;
2434	pVmcbNstGstCache->u64TSCOffset = pVmcbNstGstCtrl->u64TSCOffset;
2435	pVmcbNstGstCache->fVIntrMasking = pVmcbNstGstCtrl->IntCtrl.n.u1VIntrMasking;
2436	pVmcbNstGstCache->fNestedPaging = pVmcbNstGstCtrl->NestedPagingCtrl.n.u1NestedPaging;
2437	pVmcbNstGstCache->fLbrVirt = pVmcbNstGstCtrl->LbrVirt.n.u1LbrVirt;
2438	pVmcbNstGstCache->fCacheValid = true;
2439	Log4Func(("Cached VMCB fields\n"));
2440	}
2441
2442	return fWasCached;
2443	}
2444
2445
2446	/**
2447	* Sets up the nested-guest VMCB for execution using hardware-assisted SVM.
2448	*
2449	* This is done the first time we enter nested-guest execution using SVM R0
2450	* until the nested-guest \#VMEXIT (not to be confused with physical CPU
2451	* \#VMEXITs which may or may not cause a corresponding nested-guest \#VMEXIT).
2452	*
2453	* @param pVCpu The cross context virtual CPU structure.
2454	*/
2455	static void hmR0SvmSetupVmcbNested(PVMCPUCC pVCpu)
2456	{
2457	PSVMVMCB pVmcbNstGst = pVCpu->cpum.GstCtx.hwvirt.svm.CTX_SUFF(pVmcb);
2458	PSVMVMCBCTRL pVmcbNstGstCtrl = &pVmcbNstGst->ctrl;
2459
2460	HMSVM_ASSERT_IN_NESTED_GUEST(&pVCpu->cpum.GstCtx);
2461
2462	/*
2463	* First cache the nested-guest VMCB fields we may potentially modify.
2464	*/
2465	bool const fVmcbCached = hmR0SvmCacheVmcbNested(pVCpu);
2466	if (!fVmcbCached)
2467	{
2468	/*
2469	* The IOPM of the nested-guest can be ignored because the the guest always
2470	* intercepts all IO port accesses. Thus, we'll swap to the guest IOPM rather
2471	* than the nested-guest IOPM and swap the field back on the #VMEXIT.
2472	*/
2473	pVmcbNstGstCtrl->u64IOPMPhysAddr = g_HCPhysIOBitmap;
2474
2475	/*
2476	* Use the same nested-paging as the outer guest. We can't dynamically switch off
2477	* nested-paging suddenly while executing a VM (see assertion at the end of
2478	* Trap0eHandler() in PGMAllBth.h).
2479	*/
2480	pVmcbNstGstCtrl->NestedPagingCtrl.n.u1NestedPaging = pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging;
2481
2482	/* Always enable V_INTR_MASKING as we do not want to allow access to the physical APIC TPR. */
2483	pVmcbNstGstCtrl->IntCtrl.n.u1VIntrMasking = 1;
2484
2485	/*
2486	* Turn off TPR syncing on #VMEXIT for nested-guests as CR8 intercepts are subject
2487	* to the nested-guest intercepts and we always run with V_INTR_MASKING.
2488	*/
2489	pVCpu->hm.s.svm.fSyncVTpr = false;
2490
2491	#ifdef DEBUG_ramshankar
2492	/* For debugging purposes - copy the LBR info. from outer guest VMCB. */
2493	pVmcbNstGstCtrl->LbrVirt.n.u1LbrVirt = pVmcb->ctrl.LbrVirt.n.u1LbrVirt;
2494	#endif
2495
2496	/*
2497	* If we don't expose Virtualized-VMSAVE/VMLOAD feature to the outer guest, we
2498	* need to intercept VMSAVE/VMLOAD instructions executed by the nested-guest.
2499	*/
2500	if (!pVCpu->CTX_SUFF(pVM)->cpum.ro.GuestFeatures.fSvmVirtVmsaveVmload)
2501	pVmcbNstGstCtrl->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_VMSAVE
2502	\| SVM_CTRL_INTERCEPT_VMLOAD;
2503
2504	/*
2505	* If we don't expose Virtual GIF feature to the outer guest, we need to intercept
2506	* CLGI/STGI instructions executed by the nested-guest.
2507	*/
2508	if (!pVCpu->CTX_SUFF(pVM)->cpum.ro.GuestFeatures.fSvmVGif)
2509	pVmcbNstGstCtrl->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_CLGI
2510	\| SVM_CTRL_INTERCEPT_STGI;
2511
2512	/* Merge the guest and nested-guest intercepts. */
2513	hmR0SvmMergeVmcbCtrlsNested(pVCpu);
2514
2515	/* Update the VMCB clean bits. */
2516	pVmcbNstGstCtrl->u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2517	}
2518	else
2519	{
2520	Assert(!pVCpu->hm.s.svm.fSyncVTpr);
2521	Assert(pVmcbNstGstCtrl->u64IOPMPhysAddr == g_HCPhysIOBitmap);
2522	Assert(RT_BOOL(pVmcbNstGstCtrl->NestedPagingCtrl.n.u1NestedPaging) == pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
2523	}
2524	}
2525	#endif /* VBOX_WITH_NESTED_HWVIRT_SVM */
2526
2527
2528	/**
2529	* Exports the state shared between the host and guest (or nested-guest) into
2530	* the VMCB.
2531	*
2532	* @param pVCpu The cross context virtual CPU structure.
2533	* @param pVmcb Pointer to the VM control block.
2534	*
2535	* @remarks No-long-jump zone!!!
2536	*/
2537	static void hmR0SvmExportSharedState(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
2538	{
2539	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2540	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2541
2542	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_DR_MASK)
2543	{
2544	/** @todo Figure out stepping with nested-guest. */
2545	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2546	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
2547	hmR0SvmExportSharedDebugState(pVCpu, pVmcb);
2548	else
2549	{
2550	pVmcb->guest.u64DR6 = pCtx->dr[6];
2551	pVmcb->guest.u64DR7 = pCtx->dr[7];
2552	}
2553	}
2554
2555	pVCpu->hm.s.fCtxChanged &= ~HM_CHANGED_GUEST_DR_MASK;
2556	AssertMsg(!(pVCpu->hm.s.fCtxChanged & HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE),
2557	("fCtxChanged=%#RX64\n", pVCpu->hm.s.fCtxChanged));
2558	}
2559
2560
2561	/**
2562	* Worker for SVMR0ImportStateOnDemand.
2563	*
2564	* @param pVCpu The cross context virtual CPU structure.
2565	* @param fWhat What to import, CPUMCTX_EXTRN_XXX.
2566	*/
2567	static void hmR0SvmImportGuestState(PVMCPUCC pVCpu, uint64_t fWhat)
2568	{
2569	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatImportGuestState, x);
2570
2571	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2572	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
2573	PCSVMVMCBSTATESAVE pVmcbGuest = &pVmcb->guest;
2574	PCSVMVMCBCTRL pVmcbCtrl = &pVmcb->ctrl;
2575
2576	/*
2577	* We disable interrupts to make the updating of the state and in particular
2578	* the fExtrn modification atomic wrt to preemption hooks.
2579	*/
2580	RTCCUINTREG const fEFlags = ASMIntDisableFlags();
2581
2582	fWhat &= pCtx->fExtrn;
2583	if (fWhat)
2584	{
2585	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
2586	if (fWhat & CPUMCTX_EXTRN_HWVIRT)
2587	{
2588	if (pVmcbCtrl->IntCtrl.n.u1VGifEnable)
2589	{
2590	Assert(!CPUMIsGuestInSvmNestedHwVirtMode(pCtx)); /* We don't yet support passing VGIF feature to the guest. */
2591	Assert(HMIsSvmVGifActive(pVCpu->CTX_SUFF(pVM))); /* VM has configured it. */
2592	CPUMSetGuestGif(pCtx, pVmcbCtrl->IntCtrl.n.u1VGif);
2593	}
2594	}
2595
2596	if (fWhat & CPUMCTX_EXTRN_HM_SVM_HWVIRT_VIRQ)
2597	{
2598	if ( !pVmcbCtrl->IntCtrl.n.u1VIrqPending
2599	&& VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_INTERRUPT_NESTED_GUEST))
2600	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INTERRUPT_NESTED_GUEST);
2601	}
2602	#endif
2603
2604	if (fWhat & CPUMCTX_EXTRN_HM_SVM_INT_SHADOW)
2605	{
2606	if (pVmcbCtrl->IntShadow.n.u1IntShadow)
2607	EMSetInhibitInterruptsPC(pVCpu, pVmcbGuest->u64RIP);
2608	else if (VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS))
2609	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS);
2610	}
2611
2612	if (fWhat & CPUMCTX_EXTRN_RIP)
2613	pCtx->rip = pVmcbGuest->u64RIP;
2614
2615	if (fWhat & CPUMCTX_EXTRN_RFLAGS)
2616	pCtx->eflags.u32 = pVmcbGuest->u64RFlags;
2617
2618	if (fWhat & CPUMCTX_EXTRN_RSP)
2619	pCtx->rsp = pVmcbGuest->u64RSP;
2620
2621	if (fWhat & CPUMCTX_EXTRN_RAX)
2622	pCtx->rax = pVmcbGuest->u64RAX;
2623
2624	if (fWhat & CPUMCTX_EXTRN_SREG_MASK)
2625	{
2626	if (fWhat & CPUMCTX_EXTRN_CS)
2627	{
2628	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, CS, cs);
2629	/* Correct the CS granularity bit. Haven't seen it being wrong in any other register (yet). */
2630	/** @todo SELM might need to be fixed as it too should not care about the
2631	* granularity bit. See @bugref{6785}. */
2632	if ( !pCtx->cs.Attr.n.u1Granularity
2633	&& pCtx->cs.Attr.n.u1Present
2634	&& pCtx->cs.u32Limit > UINT32_C(0xfffff))
2635	{
2636	Assert((pCtx->cs.u32Limit & 0xfff) == 0xfff);
2637	pCtx->cs.Attr.n.u1Granularity = 1;
2638	}
2639	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, cs);
2640	}
2641	if (fWhat & CPUMCTX_EXTRN_SS)
2642	{
2643	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, SS, ss);
2644	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, ss);
2645	/*
2646	* Sync the hidden SS DPL field. AMD CPUs have a separate CPL field in the
2647	* VMCB and uses that and thus it's possible that when the CPL changes during
2648	* guest execution that the SS DPL isn't updated by AMD-V. Observed on some
2649	* AMD Fusion CPUs with 64-bit guests.
2650	*
2651	* See AMD spec. 15.5.1 "Basic operation".
2652	*/
2653	Assert(!(pVmcbGuest->u8CPL & ~0x3));
2654	uint8_t const uCpl = pVmcbGuest->u8CPL;
2655	if (pCtx->ss.Attr.n.u2Dpl != uCpl)
2656	pCtx->ss.Attr.n.u2Dpl = uCpl & 0x3;
2657	}
2658	if (fWhat & CPUMCTX_EXTRN_DS)
2659	{
2660	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, DS, ds);
2661	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, ds);
2662	}
2663	if (fWhat & CPUMCTX_EXTRN_ES)
2664	{
2665	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, ES, es);
2666	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, es);
2667	}
2668	if (fWhat & CPUMCTX_EXTRN_FS)
2669	{
2670	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, FS, fs);
2671	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, fs);
2672	}
2673	if (fWhat & CPUMCTX_EXTRN_GS)
2674	{
2675	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, GS, gs);
2676	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, gs);
2677	}
2678	}
2679
2680	if (fWhat & CPUMCTX_EXTRN_TABLE_MASK)
2681	{
2682	if (fWhat & CPUMCTX_EXTRN_TR)
2683	{
2684	/*
2685	* Fixup TR attributes so it's compatible with Intel. Important when saved-states
2686	* are used between Intel and AMD, see @bugref{6208#c39}.
2687	* ASSUME that it's normally correct and that we're in 32-bit or 64-bit mode.
2688	*/
2689	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, TR, tr);
2690	if (pCtx->tr.Attr.n.u4Type != X86_SEL_TYPE_SYS_386_TSS_BUSY)
2691	{
2692	if ( pCtx->tr.Attr.n.u4Type == X86_SEL_TYPE_SYS_386_TSS_AVAIL
2693	\|\| CPUMIsGuestInLongModeEx(pCtx))
2694	pCtx->tr.Attr.n.u4Type = X86_SEL_TYPE_SYS_386_TSS_BUSY;
2695	else if (pCtx->tr.Attr.n.u4Type == X86_SEL_TYPE_SYS_286_TSS_AVAIL)
2696	pCtx->tr.Attr.n.u4Type = X86_SEL_TYPE_SYS_286_TSS_BUSY;
2697	}
2698	}
2699
2700	if (fWhat & CPUMCTX_EXTRN_LDTR)
2701	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, LDTR, ldtr);
2702
2703	if (fWhat & CPUMCTX_EXTRN_GDTR)
2704	{
2705	pCtx->gdtr.cbGdt = pVmcbGuest->GDTR.u32Limit;
2706	pCtx->gdtr.pGdt = pVmcbGuest->GDTR.u64Base;
2707	}
2708
2709	if (fWhat & CPUMCTX_EXTRN_IDTR)
2710	{
2711	pCtx->idtr.cbIdt = pVmcbGuest->IDTR.u32Limit;
2712	pCtx->idtr.pIdt = pVmcbGuest->IDTR.u64Base;
2713	}
2714	}
2715
2716	if (fWhat & CPUMCTX_EXTRN_SYSCALL_MSRS)
2717	{
2718	pCtx->msrSTAR = pVmcbGuest->u64STAR;
2719	pCtx->msrLSTAR = pVmcbGuest->u64LSTAR;
2720	pCtx->msrCSTAR = pVmcbGuest->u64CSTAR;
2721	pCtx->msrSFMASK = pVmcbGuest->u64SFMASK;
2722	}
2723
2724	if (fWhat & CPUMCTX_EXTRN_SYSENTER_MSRS)
2725	{
2726	pCtx->SysEnter.cs = pVmcbGuest->u64SysEnterCS;
2727	pCtx->SysEnter.eip = pVmcbGuest->u64SysEnterEIP;
2728	pCtx->SysEnter.esp = pVmcbGuest->u64SysEnterESP;
2729	}
2730
2731	if (fWhat & CPUMCTX_EXTRN_KERNEL_GS_BASE)
2732	pCtx->msrKERNELGSBASE = pVmcbGuest->u64KernelGSBase;
2733
2734	if (fWhat & CPUMCTX_EXTRN_DR_MASK)
2735	{
2736	if (fWhat & CPUMCTX_EXTRN_DR6)
2737	{
2738	if (!pVCpu->hm.s.fUsingHyperDR7)
2739	pCtx->dr[6] = pVmcbGuest->u64DR6;
2740	else
2741	CPUMSetHyperDR6(pVCpu, pVmcbGuest->u64DR6);
2742	}
2743
2744	if (fWhat & CPUMCTX_EXTRN_DR7)
2745	{
2746	if (!pVCpu->hm.s.fUsingHyperDR7)
2747	pCtx->dr[7] = pVmcbGuest->u64DR7;
2748	else
2749	Assert(pVmcbGuest->u64DR7 == CPUMGetHyperDR7(pVCpu));
2750	}
2751	}
2752
2753	if (fWhat & CPUMCTX_EXTRN_CR_MASK)
2754	{
2755	if (fWhat & CPUMCTX_EXTRN_CR0)
2756	{
2757	/* We intercept changes to all CR0 bits except maybe TS & MP bits. */
2758	uint64_t const uCr0 = (pCtx->cr0 & ~(X86_CR0_TS \| X86_CR0_MP))
2759	\| (pVmcbGuest->u64CR0 & (X86_CR0_TS \| X86_CR0_MP));
2760	VMMRZCallRing3Disable(pVCpu); /* Calls into PGM which has Log statements. */
2761	CPUMSetGuestCR0(pVCpu, uCr0);
2762	VMMRZCallRing3Enable(pVCpu);
2763	}
2764
2765	if (fWhat & CPUMCTX_EXTRN_CR2)
2766	pCtx->cr2 = pVmcbGuest->u64CR2;
2767
2768	if (fWhat & CPUMCTX_EXTRN_CR3)
2769	{
2770	if ( pVmcbCtrl->NestedPagingCtrl.n.u1NestedPaging
2771	&& pCtx->cr3 != pVmcbGuest->u64CR3)
2772	{
2773	CPUMSetGuestCR3(pVCpu, pVmcbGuest->u64CR3);
2774	VMCPU_FF_SET(pVCpu, VMCPU_FF_HM_UPDATE_CR3);
2775	}
2776	}
2777
2778	/* Changes to CR4 are always intercepted. */
2779	}
2780
2781	/* Update fExtrn. */
2782	pCtx->fExtrn &= ~fWhat;
2783
2784	/* If everything has been imported, clear the HM keeper bit. */
2785	if (!(pCtx->fExtrn & HMSVM_CPUMCTX_EXTRN_ALL))
2786	{
2787	pCtx->fExtrn &= ~CPUMCTX_EXTRN_KEEPER_HM;
2788	Assert(!pCtx->fExtrn);
2789	}
2790	}
2791	else
2792	Assert(!pCtx->fExtrn \|\| (pCtx->fExtrn & HMSVM_CPUMCTX_EXTRN_ALL));
2793
2794	ASMSetFlags(fEFlags);
2795
2796	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatImportGuestState, x);
2797
2798	/*
2799	* Honor any pending CR3 updates.
2800	*
2801	* Consider this scenario: #VMEXIT -> VMMRZCallRing3Enable() -> do stuff that causes a longjmp
2802	* -> SVMR0CallRing3Callback() -> VMMRZCallRing3Disable() -> hmR0SvmImportGuestState()
2803	* -> Sets VMCPU_FF_HM_UPDATE_CR3 pending -> return from the longjmp -> continue with #VMEXIT
2804	* handling -> hmR0SvmImportGuestState() and here we are.
2805	*
2806	* The reason for such complicated handling is because VM-exits that call into PGM expect
2807	* CR3 to be up-to-date and thus any CR3-saves -before- the VM-exit (longjmp) would've
2808	* postponed the CR3 update via the force-flag and cleared CR3 from fExtrn. Any SVM R0
2809	* VM-exit handler that requests CR3 to be saved will end up here and we call PGMUpdateCR3().
2810	*
2811	* The longjmp exit path can't check these CR3 force-flags and call code that takes a lock again,
2812	* and does not process force-flag like regular exits to ring-3 either, we cover for it here.
2813	*/
2814	if ( VMMRZCallRing3IsEnabled(pVCpu)
2815	&& VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_HM_UPDATE_CR3))
2816	{
2817	Assert(pCtx->cr3 == pVmcbGuest->u64CR3);
2818	PGMUpdateCR3(pVCpu, pCtx->cr3);
2819	}
2820	}
2821
2822
2823	/**
2824	* Saves the guest (or nested-guest) state from the VMCB into the guest-CPU
2825	* context.
2826	*
2827	* Currently there is no residual state left in the CPU that is not updated in the
2828	* VMCB.
2829	*
2830	* @returns VBox status code.
2831	* @param pVCpu The cross context virtual CPU structure.
2832	* @param fWhat What to import, CPUMCTX_EXTRN_XXX.
2833	*/
2834	VMMR0DECL(int) SVMR0ImportStateOnDemand(PVMCPUCC pVCpu, uint64_t fWhat)
2835	{
2836	hmR0SvmImportGuestState(pVCpu, fWhat);
2837	return VINF_SUCCESS;
2838	}
2839
2840
2841	/**
2842	* Does the necessary state syncing before returning to ring-3 for any reason
2843	* (longjmp, preemption, voluntary exits to ring-3) from AMD-V.
2844	*
2845	* @param pVCpu The cross context virtual CPU structure.
2846	* @param fImportState Whether to import the guest state from the VMCB back
2847	* to the guest-CPU context.
2848	*
2849	* @remarks No-long-jmp zone!!!
2850	*/
2851	static void hmR0SvmLeave(PVMCPUCC pVCpu, bool fImportState)
2852	{
2853	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2854	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2855	Assert(VMMR0IsLogFlushDisabled(pVCpu));
2856
2857	/*
2858	* !!! IMPORTANT !!!
2859	* If you modify code here, make sure to check whether SVMR0CallRing3Callback() needs to be updated too.
2860	*/
2861
2862	/* Save the guest state if necessary. */
2863	if (fImportState)
2864	hmR0SvmImportGuestState(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
2865
2866	/* Restore host FPU state if necessary and resync on next R0 reentry. */
2867	CPUMR0FpuStateMaybeSaveGuestAndRestoreHost(pVCpu);
2868	Assert(!CPUMIsGuestFPUStateActive(pVCpu));
2869
2870	/*
2871	* Restore host debug registers if necessary and resync on next R0 reentry.
2872	*/
2873	#ifdef VBOX_STRICT
2874	if (CPUMIsHyperDebugStateActive(pVCpu))
2875	{
2876	PSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb; /** @todo nested-guest. */
2877	Assert(pVmcb->ctrl.u16InterceptRdDRx == 0xffff);
2878	Assert(pVmcb->ctrl.u16InterceptWrDRx == 0xffff);
2879	}
2880	#endif
2881	CPUMR0DebugStateMaybeSaveGuestAndRestoreHost(pVCpu, false /* save DR6 */);
2882	Assert(!CPUMIsHyperDebugStateActive(pVCpu));
2883	Assert(!CPUMIsGuestDebugStateActive(pVCpu));
2884
2885	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatEntry);
2886	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatImportGuestState);
2887	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatExportGuestState);
2888	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatPreExit);
2889	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatExitHandling);
2890	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatExitVmentry);
2891	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchLongJmpToR3);
2892
2893	VMCPU_CMPXCHG_STATE(pVCpu, VMCPUSTATE_STARTED_HM, VMCPUSTATE_STARTED_EXEC);
2894	}
2895
2896
2897	/**
2898	* Leaves the AMD-V session.
2899	*
2900	* Only used while returning to ring-3 either due to longjump or exits to
2901	* ring-3.
2902	*
2903	* @returns VBox status code.
2904	* @param pVCpu The cross context virtual CPU structure.
2905	*/
2906	static int hmR0SvmLeaveSession(PVMCPUCC pVCpu)
2907	{
2908	HM_DISABLE_PREEMPT(pVCpu);
2909	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2910	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2911
2912	/* When thread-context hooks are used, we can avoid doing the leave again if we had been preempted before
2913	and done this from the SVMR0ThreadCtxCallback(). */
2914	if (!pVCpu->hm.s.fLeaveDone)
2915	{
2916	hmR0SvmLeave(pVCpu, true /* fImportState */);
2917	pVCpu->hm.s.fLeaveDone = true;
2918	}
2919
2920	/*
2921	* !!! IMPORTANT !!!
2922	* If you modify code here, make sure to check whether SVMR0CallRing3Callback() needs to be updated too.
2923	*/
2924
2925	/** @todo eliminate the need for calling VMMR0ThreadCtxHookDisable here! */
2926	/* Deregister hook now that we've left HM context before re-enabling preemption. */
2927	VMMR0ThreadCtxHookDisable(pVCpu);
2928
2929	/* Leave HM context. This takes care of local init (term). */
2930	int rc = HMR0LeaveCpu(pVCpu);
2931
2932	HM_RESTORE_PREEMPT();
2933	return rc;
2934	}
2935
2936
2937	/**
2938	* Does the necessary state syncing before doing a longjmp to ring-3.
2939	*
2940	* @returns VBox status code.
2941	* @param pVCpu The cross context virtual CPU structure.
2942	*
2943	* @remarks No-long-jmp zone!!!
2944	*/
2945	static int hmR0SvmLongJmpToRing3(PVMCPUCC pVCpu)
2946	{
2947	return hmR0SvmLeaveSession(pVCpu);
2948	}
2949
2950
2951	/**
2952	* VMMRZCallRing3() callback wrapper which saves the guest state (or restores
2953	* any remaining host state) before we longjump to ring-3 and possibly get
2954	* preempted.
2955	*
2956	* @param pVCpu The cross context virtual CPU structure.
2957	* @param enmOperation The operation causing the ring-3 longjump.
2958	*/
2959	VMMR0DECL(int) SVMR0CallRing3Callback(PVMCPUCC pVCpu, VMMCALLRING3 enmOperation)
2960	{
2961	if (enmOperation == VMMCALLRING3_VM_R0_ASSERTION)
2962	{
2963	/*
2964	* !!! IMPORTANT !!!
2965	* If you modify code here, make sure to check whether hmR0SvmLeave() and hmR0SvmLeaveSession() needs
2966	* to be updated too. This is a stripped down version which gets out ASAP trying to not trigger any assertion.
2967	*/
2968	VMMRZCallRing3RemoveNotification(pVCpu);
2969	VMMRZCallRing3Disable(pVCpu);
2970	HM_DISABLE_PREEMPT(pVCpu);
2971
2972	/* Import the entire guest state. */
2973	hmR0SvmImportGuestState(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
2974
2975	/* Restore host FPU state if necessary and resync on next R0 reentry. */
2976	CPUMR0FpuStateMaybeSaveGuestAndRestoreHost(pVCpu);
2977
2978	/* Restore host debug registers if necessary and resync on next R0 reentry. */
2979	CPUMR0DebugStateMaybeSaveGuestAndRestoreHost(pVCpu, false /* save DR6 */);
2980
2981	/* Deregister the hook now that we've left HM context before re-enabling preemption. */
2982	/** @todo eliminate the need for calling VMMR0ThreadCtxHookDisable here! */
2983	VMMR0ThreadCtxHookDisable(pVCpu);
2984
2985	/* Leave HM context. This takes care of local init (term). */
2986	HMR0LeaveCpu(pVCpu);
2987
2988	HM_RESTORE_PREEMPT();
2989	return VINF_SUCCESS;
2990	}
2991
2992	Assert(pVCpu);
2993	Assert(VMMRZCallRing3IsEnabled(pVCpu));
2994	HMSVM_ASSERT_PREEMPT_SAFE(pVCpu);
2995
2996	VMMRZCallRing3Disable(pVCpu);
2997	Assert(VMMR0IsLogFlushDisabled(pVCpu));
2998
2999	Log4Func(("Calling hmR0SvmLongJmpToRing3\n"));
3000	int rc = hmR0SvmLongJmpToRing3(pVCpu);
3001	AssertRCReturn(rc, rc);
3002
3003	VMMRZCallRing3Enable(pVCpu);
3004	return VINF_SUCCESS;
3005	}
3006
3007
3008	/**
3009	* Take necessary actions before going back to ring-3.
3010	*
3011	* An action requires us to go back to ring-3. This function does the necessary
3012	* steps before we can safely return to ring-3. This is not the same as longjmps
3013	* to ring-3, this is voluntary.
3014	*
3015	* @returns VBox status code.
3016	* @param pVCpu The cross context virtual CPU structure.
3017	* @param rcExit The reason for exiting to ring-3. Can be
3018	* VINF_VMM_UNKNOWN_RING3_CALL.
3019	*/
3020	static int hmR0SvmExitToRing3(PVMCPUCC pVCpu, int rcExit)
3021	{
3022	Assert(pVCpu);
3023	HMSVM_ASSERT_PREEMPT_SAFE(pVCpu);
3024
3025	/* Please, no longjumps here (any logging shouldn't flush jump back to ring-3). NO LOGGING BEFORE THIS POINT! */
3026	VMMRZCallRing3Disable(pVCpu);
3027	Log4Func(("rcExit=%d LocalFF=%#RX64 GlobalFF=%#RX32\n", rcExit, (uint64_t)pVCpu->fLocalForcedActions,
3028	pVCpu->CTX_SUFF(pVM)->fGlobalForcedActions));
3029
3030	/* We need to do this only while truly exiting the "inner loop" back to ring-3 and -not- for any longjmp to ring3. */
3031	if (pVCpu->hm.s.Event.fPending)
3032	{
3033	hmR0SvmPendingEventToTrpmTrap(pVCpu);
3034	Assert(!pVCpu->hm.s.Event.fPending);
3035	}
3036
3037	/* Sync. the necessary state for going back to ring-3. */
3038	hmR0SvmLeaveSession(pVCpu);
3039	STAM_COUNTER_DEC(&pVCpu->hm.s.StatSwitchLongJmpToR3);
3040
3041	/* Thread-context hooks are unregistered at this point!!! */
3042	/* Ring-3 callback notifications are unregistered at this point!!! */
3043
3044	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_TO_R3);
3045	CPUMSetChangedFlags(pVCpu, CPUM_CHANGED_SYSENTER_MSR
3046	\| CPUM_CHANGED_LDTR
3047	\| CPUM_CHANGED_GDTR
3048	\| CPUM_CHANGED_IDTR
3049	\| CPUM_CHANGED_TR
3050	\| CPUM_CHANGED_HIDDEN_SEL_REGS);
3051	if ( pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging
3052	&& CPUMIsGuestPagingEnabledEx(&pVCpu->cpum.GstCtx))
3053	{
3054	CPUMSetChangedFlags(pVCpu, CPUM_CHANGED_GLOBAL_TLB_FLUSH);
3055	}
3056
3057	/* Update the exit-to-ring 3 reason. */
3058	pVCpu->hm.s.rcLastExitToR3 = rcExit;
3059
3060	/* On our way back from ring-3, reload the guest-CPU state if it may change while in ring-3. */
3061	if ( rcExit != VINF_EM_RAW_INTERRUPT
3062	\|\| CPUMIsGuestInSvmNestedHwVirtMode(&pVCpu->cpum.GstCtx))
3063	{
3064	Assert(!(pVCpu->cpum.GstCtx.fExtrn & HMSVM_CPUMCTX_EXTRN_ALL));
3065	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
3066	}
3067
3068	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchExitToR3);
3069	VMMRZCallRing3Enable(pVCpu);
3070
3071	/*
3072	* If we're emulating an instruction, we shouldn't have any TRPM traps pending
3073	* and if we're injecting an event we should have a TRPM trap pending.
3074	*/
3075	AssertReturnStmt(rcExit != VINF_EM_RAW_INJECT_TRPM_EVENT \|\| TRPMHasTrap(pVCpu),
3076	pVCpu->hm.s.u32HMError = rcExit,
3077	VERR_SVM_IPE_5);
3078	AssertReturnStmt(rcExit != VINF_EM_RAW_EMULATE_INSTR \|\| !TRPMHasTrap(pVCpu),
3079	pVCpu->hm.s.u32HMError = rcExit,
3080	VERR_SVM_IPE_4);
3081
3082	return rcExit;
3083	}
3084
3085
3086	/**
3087	* Updates the use of TSC offsetting mode for the CPU and adjusts the necessary
3088	* intercepts.
3089	*
3090	* @param pVCpu The cross context virtual CPU structure.
3091	* @param pVmcb Pointer to the VM control block.
3092	*
3093	* @remarks No-long-jump zone!!!
3094	*/
3095	static void hmR0SvmUpdateTscOffsetting(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
3096	{
3097	/*
3098	* Avoid intercepting RDTSC/RDTSCP if we determined the host TSC (++) is stable
3099	* and in case of a nested-guest, if the nested-VMCB specifies it is not intercepting
3100	* RDTSC/RDTSCP as well.
3101	*/
3102	bool fParavirtTsc;
3103	uint64_t uTscOffset;
3104	bool const fCanUseRealTsc = TMCpuTickCanUseRealTSC(pVCpu->CTX_SUFF(pVM), pVCpu, &uTscOffset, &fParavirtTsc);
3105
3106	bool fIntercept;
3107	if (fCanUseRealTsc)
3108	fIntercept = hmR0SvmClearCtrlIntercept(pVCpu, pVmcb, SVM_CTRL_INTERCEPT_RDTSC \| SVM_CTRL_INTERCEPT_RDTSCP);
3109	else
3110	{
3111	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_RDTSC \| SVM_CTRL_INTERCEPT_RDTSCP);
3112	fIntercept = true;
3113	}
3114
3115	if (!fIntercept)
3116	{
3117	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
3118	/* Apply the nested-guest VMCB's TSC offset over the guest TSC offset. */
3119	if (CPUMIsGuestInSvmNestedHwVirtMode(&pVCpu->cpum.GstCtx))
3120	uTscOffset = CPUMApplyNestedGuestTscOffset(pVCpu, uTscOffset);
3121	#endif
3122
3123	/* Update the TSC offset in the VMCB and the relevant clean bits. */
3124	pVmcb->ctrl.u64TSCOffset = uTscOffset;
3125	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
3126	}
3127
3128	/* Currently neither Hyper-V nor KVM need to update their paravirt. TSC
3129	information before every VM-entry, hence we have nothing to do here at the moment. */
3130	if (fParavirtTsc)
3131	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscParavirt);
3132	}
3133
3134
3135	/**
3136	* Sets an event as a pending event to be injected into the guest.
3137	*
3138	* @param pVCpu The cross context virtual CPU structure.
3139	* @param pEvent Pointer to the SVM event.
3140	* @param GCPtrFaultAddress The fault-address (CR2) in case it's a
3141	* page-fault.
3142	*
3143	* @remarks Statistics counter assumes this is a guest event being reflected to
3144	* the guest i.e. 'StatInjectPendingReflect' is incremented always.
3145	*/
3146	DECLINLINE(void) hmR0SvmSetPendingEvent(PVMCPUCC pVCpu, PSVMEVENT pEvent, RTGCUINTPTR GCPtrFaultAddress)
3147	{
3148	Assert(!pVCpu->hm.s.Event.fPending);
3149	Assert(pEvent->n.u1Valid);
3150
3151	pVCpu->hm.s.Event.u64IntInfo = pEvent->u;
3152	pVCpu->hm.s.Event.fPending = true;
3153	pVCpu->hm.s.Event.GCPtrFaultAddress = GCPtrFaultAddress;
3154
3155	Log4Func(("u=%#RX64 u8Vector=%#x Type=%#x ErrorCodeValid=%RTbool ErrorCode=%#RX32\n", pEvent->u, pEvent->n.u8Vector,
3156	(uint8_t)pEvent->n.u3Type, !!pEvent->n.u1ErrorCodeValid, pEvent->n.u32ErrorCode));
3157	}
3158
3159
3160	/**
3161	* Sets an invalid-opcode (\#UD) exception as pending-for-injection into the VM.
3162	*
3163	* @param pVCpu The cross context virtual CPU structure.
3164	*/
3165	DECLINLINE(void) hmR0SvmSetPendingXcptUD(PVMCPUCC pVCpu)
3166	{
3167	SVMEVENT Event;
3168	Event.u = 0;
3169	Event.n.u1Valid = 1;
3170	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3171	Event.n.u8Vector = X86_XCPT_UD;
3172	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3173	}
3174
3175
3176	/**
3177	* Sets a debug (\#DB) exception as pending-for-injection into the VM.
3178	*
3179	* @param pVCpu The cross context virtual CPU structure.
3180	*/
3181	DECLINLINE(void) hmR0SvmSetPendingXcptDB(PVMCPUCC pVCpu)
3182	{
3183	SVMEVENT Event;
3184	Event.u = 0;
3185	Event.n.u1Valid = 1;
3186	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3187	Event.n.u8Vector = X86_XCPT_DB;
3188	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3189	}
3190
3191
3192	/**
3193	* Sets a page fault (\#PF) exception as pending-for-injection into the VM.
3194	*
3195	* @param pVCpu The cross context virtual CPU structure.
3196	* @param u32ErrCode The error-code for the page-fault.
3197	* @param uFaultAddress The page fault address (CR2).
3198	*
3199	* @remarks This updates the guest CR2 with @a uFaultAddress!
3200	*/
3201	DECLINLINE(void) hmR0SvmSetPendingXcptPF(PVMCPUCC pVCpu, uint32_t u32ErrCode, RTGCUINTPTR uFaultAddress)
3202	{
3203	SVMEVENT Event;
3204	Event.u = 0;
3205	Event.n.u1Valid = 1;
3206	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3207	Event.n.u8Vector = X86_XCPT_PF;
3208	Event.n.u1ErrorCodeValid = 1;
3209	Event.n.u32ErrorCode = u32ErrCode;
3210
3211	/* Update CR2 of the guest. */
3212	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_CR2);
3213	if (pVCpu->cpum.GstCtx.cr2 != uFaultAddress)
3214	{
3215	pVCpu->cpum.GstCtx.cr2 = uFaultAddress;
3216	/* The VMCB clean bit for CR2 will be updated while re-loading the guest state. */
3217	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_CR2);
3218	}
3219
3220	hmR0SvmSetPendingEvent(pVCpu, &Event, uFaultAddress);
3221	}
3222
3223
3224	/**
3225	* Sets a math-fault (\#MF) exception as pending-for-injection into the VM.
3226	*
3227	* @param pVCpu The cross context virtual CPU structure.
3228	*/
3229	DECLINLINE(void) hmR0SvmSetPendingXcptMF(PVMCPUCC pVCpu)
3230	{
3231	SVMEVENT Event;
3232	Event.u = 0;
3233	Event.n.u1Valid = 1;
3234	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3235	Event.n.u8Vector = X86_XCPT_MF;
3236	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3237	}
3238
3239
3240	/**
3241	* Sets a double fault (\#DF) exception as pending-for-injection into the VM.
3242	*
3243	* @param pVCpu The cross context virtual CPU structure.
3244	*/
3245	DECLINLINE(void) hmR0SvmSetPendingXcptDF(PVMCPUCC pVCpu)
3246	{
3247	SVMEVENT Event;
3248	Event.u = 0;
3249	Event.n.u1Valid = 1;
3250	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3251	Event.n.u8Vector = X86_XCPT_DF;
3252	Event.n.u1ErrorCodeValid = 1;
3253	Event.n.u32ErrorCode = 0;
3254	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3255	}
3256
3257
3258	/**
3259	* Injects an event into the guest upon VMRUN by updating the relevant field
3260	* in the VMCB.
3261	*
3262	* @param pVCpu The cross context virtual CPU structure.
3263	* @param pVmcb Pointer to the guest VM control block.
3264	* @param pEvent Pointer to the event.
3265	*
3266	* @remarks No-long-jump zone!!!
3267	* @remarks Requires CR0!
3268	*/
3269	DECLINLINE(void) hmR0SvmInjectEventVmcb(PVMCPUCC pVCpu, PSVMVMCB pVmcb, PSVMEVENT pEvent)
3270	{
3271	Assert(!pVmcb->ctrl.EventInject.n.u1Valid);
3272	pVmcb->ctrl.EventInject.u = pEvent->u;
3273	if ( pVmcb->ctrl.EventInject.n.u3Type == SVM_EVENT_EXCEPTION
3274	\|\| pVmcb->ctrl.EventInject.n.u3Type == SVM_EVENT_NMI)
3275	{
3276	Assert(pEvent->n.u8Vector <= X86_XCPT_LAST);
3277	STAM_COUNTER_INC(&pVCpu->hm.s.paStatInjectedXcptsR0[pEvent->n.u8Vector]);
3278	}
3279	else
3280	STAM_COUNTER_INC(&pVCpu->hm.s.paStatInjectedIrqsR0[pEvent->n.u8Vector & MASK_INJECT_IRQ_STAT]);
3281	RT_NOREF(pVCpu);
3282
3283	Log4Func(("u=%#RX64 u8Vector=%#x Type=%#x ErrorCodeValid=%RTbool ErrorCode=%#RX32\n", pEvent->u, pEvent->n.u8Vector,
3284	(uint8_t)pEvent->n.u3Type, !!pEvent->n.u1ErrorCodeValid, pEvent->n.u32ErrorCode));
3285	}
3286
3287
3288
3289	/**
3290	* Converts any TRPM trap into a pending HM event. This is typically used when
3291	* entering from ring-3 (not longjmp returns).
3292	*
3293	* @param pVCpu The cross context virtual CPU structure.
3294	*/
3295	static void hmR0SvmTrpmTrapToPendingEvent(PVMCPUCC pVCpu)
3296	{
3297	Assert(TRPMHasTrap(pVCpu));
3298	Assert(!pVCpu->hm.s.Event.fPending);
3299
3300	uint8_t uVector;
3301	TRPMEVENT enmTrpmEvent;
3302	uint32_t uErrCode;
3303	RTGCUINTPTR GCPtrFaultAddress;
3304	uint8_t cbInstr;
3305
3306	int rc = TRPMQueryTrapAll(pVCpu, &uVector, &enmTrpmEvent, &uErrCode, &GCPtrFaultAddress, &cbInstr, NULL /* pfIcebp */);
3307	AssertRC(rc);
3308
3309	SVMEVENT Event;
3310	Event.u = 0;
3311	Event.n.u1Valid = 1;
3312	Event.n.u8Vector = uVector;
3313
3314	/* Refer AMD spec. 15.20 "Event Injection" for the format. */
3315	if (enmTrpmEvent == TRPM_TRAP)
3316	{
3317	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3318	switch (uVector)
3319	{
3320	case X86_XCPT_NMI:
3321	{
3322	Event.n.u3Type = SVM_EVENT_NMI;
3323	break;
3324	}
3325
3326	case X86_XCPT_BP:
3327	case X86_XCPT_OF:
3328	AssertMsgFailed(("Invalid TRPM vector %d for event type %d\n", uVector, enmTrpmEvent));
3329	RT_FALL_THRU();
3330
3331	case X86_XCPT_PF:
3332	case X86_XCPT_DF:
3333	case X86_XCPT_TS:
3334	case X86_XCPT_NP:
3335	case X86_XCPT_SS:
3336	case X86_XCPT_GP:
3337	case X86_XCPT_AC:
3338	{
3339	Event.n.u1ErrorCodeValid = 1;
3340	Event.n.u32ErrorCode = uErrCode;
3341	break;
3342	}
3343	}
3344	}
3345	else if (enmTrpmEvent == TRPM_HARDWARE_INT)
3346	Event.n.u3Type = SVM_EVENT_EXTERNAL_IRQ;
3347	else if (enmTrpmEvent == TRPM_SOFTWARE_INT)
3348	Event.n.u3Type = SVM_EVENT_SOFTWARE_INT;
3349	else
3350	AssertMsgFailed(("Invalid TRPM event type %d\n", enmTrpmEvent));
3351
3352	rc = TRPMResetTrap(pVCpu);
3353	AssertRC(rc);
3354
3355	Log4(("TRPM->HM event: u=%#RX64 u8Vector=%#x uErrorCodeValid=%RTbool uErrorCode=%#RX32\n", Event.u, Event.n.u8Vector,
3356	!!Event.n.u1ErrorCodeValid, Event.n.u32ErrorCode));
3357
3358	hmR0SvmSetPendingEvent(pVCpu, &Event, GCPtrFaultAddress);
3359	}
3360
3361
3362	/**
3363	* Converts any pending SVM event into a TRPM trap. Typically used when leaving
3364	* AMD-V to execute any instruction.
3365	*
3366	* @param pVCpu The cross context virtual CPU structure.
3367	*/
3368	static void hmR0SvmPendingEventToTrpmTrap(PVMCPUCC pVCpu)
3369	{
3370	Assert(pVCpu->hm.s.Event.fPending);
3371	Assert(TRPMQueryTrap(pVCpu, NULL /* pu8TrapNo /, NULL / pEnmType */) == VERR_TRPM_NO_ACTIVE_TRAP);
3372
3373	SVMEVENT Event;
3374	Event.u = pVCpu->hm.s.Event.u64IntInfo;
3375
3376	uint8_t uVector = Event.n.u8Vector;
3377	TRPMEVENT enmTrapType = HMSvmEventToTrpmEventType(&Event, uVector);
3378
3379	Log4(("HM event->TRPM: uVector=%#x enmTrapType=%d\n", uVector, Event.n.u3Type));
3380
3381	int rc = TRPMAssertTrap(pVCpu, uVector, enmTrapType);
3382	AssertRC(rc);
3383
3384	if (Event.n.u1ErrorCodeValid)
3385	TRPMSetErrorCode(pVCpu, Event.n.u32ErrorCode);
3386
3387	if ( enmTrapType == TRPM_TRAP
3388	&& uVector == X86_XCPT_PF)
3389	{
3390	TRPMSetFaultAddress(pVCpu, pVCpu->hm.s.Event.GCPtrFaultAddress);
3391	Assert(pVCpu->hm.s.Event.GCPtrFaultAddress == CPUMGetGuestCR2(pVCpu));
3392	}
3393	else if (enmTrapType == TRPM_SOFTWARE_INT)
3394	TRPMSetInstrLength(pVCpu, pVCpu->hm.s.Event.cbInstr);
3395	pVCpu->hm.s.Event.fPending = false;
3396	}
3397
3398
3399	/**
3400	* Checks if the guest (or nested-guest) has an interrupt shadow active right
3401	* now.
3402	*
3403	* @returns @c true if the interrupt shadow is active, @c false otherwise.
3404	* @param pVCpu The cross context virtual CPU structure.
3405	*
3406	* @remarks No-long-jump zone!!!
3407	* @remarks Has side-effects with VMCPU_FF_INHIBIT_INTERRUPTS force-flag.
3408	*/
3409	static bool hmR0SvmIsIntrShadowActive(PVMCPUCC pVCpu)
3410	{
3411	/*
3412	* Instructions like STI and MOV SS inhibit interrupts till the next instruction
3413	* completes. Check if we should inhibit interrupts or clear any existing
3414	* interrupt inhibition.
3415	*/
3416	if (VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS))
3417	{
3418	if (pVCpu->cpum.GstCtx.rip != EMGetInhibitInterruptsPC(pVCpu))
3419	{
3420	/*
3421	* We can clear the inhibit force flag as even if we go back to the recompiler
3422	* without executing guest code in AMD-V, the flag's condition to be cleared is
3423	* met and thus the cleared state is correct.
3424	*/
3425	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS);
3426	return false;
3427	}
3428	return true;
3429	}
3430	return false;
3431	}
3432
3433
3434	/**
3435	* Sets the virtual interrupt intercept control in the VMCB.
3436	*
3437	* @param pVCpu The cross context virtual CPU structure.
3438	* @param pVmcb Pointer to the VM control block.
3439	*/
3440	static void hmR0SvmSetIntWindowExiting(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
3441	{
3442	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(&pVCpu->cpum.GstCtx); NOREF(pVCpu);
3443
3444	/*
3445	* When AVIC isn't supported, set up an interrupt window to cause a #VMEXIT when the guest
3446	* is ready to accept interrupts. At #VMEXIT, we then get the interrupt from the APIC
3447	* (updating ISR at the right time) and inject the interrupt.
3448	*
3449	* With AVIC is supported, we could make use of the asynchronously delivery without
3450	* #VMEXIT and we would be passing the AVIC page to SVM.
3451	*
3452	* In AMD-V, an interrupt window is achieved using a combination of V_IRQ (an interrupt
3453	* is pending), V_IGN_TPR (ignore TPR priorities) and the VINTR intercept all being set.
3454	*/
3455	Assert(pVmcb->ctrl.IntCtrl.n.u1IgnoreTPR);
3456	pVmcb->ctrl.IntCtrl.n.u1VIrqPending = 1;
3457	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INT_CTRL;
3458	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_VINTR);
3459	Log4(("Set VINTR intercept\n"));
3460	}
3461
3462
3463	/**
3464	* Clears the virtual interrupt intercept control in the VMCB as
3465	* we are figured the guest is unable process any interrupts
3466	* at this point of time.
3467	*
3468	* @param pVCpu The cross context virtual CPU structure.
3469	* @param pVmcb Pointer to the VM control block.
3470	*/
3471	static void hmR0SvmClearIntWindowExiting(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
3472	{
3473	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(&pVCpu->cpum.GstCtx); NOREF(pVCpu);
3474
3475	PSVMVMCBCTRL pVmcbCtrl = &pVmcb->ctrl;
3476	if ( pVmcbCtrl->IntCtrl.n.u1VIrqPending
3477	\|\| (pVmcbCtrl->u64InterceptCtrl & SVM_CTRL_INTERCEPT_VINTR))
3478	{
3479	pVmcbCtrl->IntCtrl.n.u1VIrqPending = 0;
3480	pVmcbCtrl->u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INT_CTRL;
3481	hmR0SvmClearCtrlIntercept(pVCpu, pVmcb, SVM_CTRL_INTERCEPT_VINTR);
3482	Log4(("Cleared VINTR intercept\n"));
3483	}
3484	}
3485
3486
3487	/**
3488	* Evaluates the event to be delivered to the guest and sets it as the pending
3489	* event.
3490	*
3491	* @returns Strict VBox status code.
3492	* @param pVCpu The cross context virtual CPU structure.
3493	* @param pSvmTransient Pointer to the SVM transient structure.
3494	*/
3495	static VBOXSTRICTRC hmR0SvmEvaluatePendingEvent(PVMCPUCC pVCpu, PCSVMTRANSIENT pSvmTransient)
3496	{
3497	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
3498	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_HWVIRT
3499	\| CPUMCTX_EXTRN_RFLAGS
3500	\| CPUMCTX_EXTRN_HM_SVM_INT_SHADOW
3501	\| CPUMCTX_EXTRN_HM_SVM_HWVIRT_VIRQ);
3502
3503	Assert(!pVCpu->hm.s.Event.fPending);
3504	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
3505	Assert(pVmcb);
3506
3507	bool const fGif = CPUMGetGuestGif(pCtx);
3508	bool const fIntShadow = hmR0SvmIsIntrShadowActive(pVCpu);
3509	bool const fBlockNmi = VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_BLOCK_NMIS);
3510
3511	Log4Func(("fGif=%RTbool fBlockNmi=%RTbool fIntShadow=%RTbool fIntPending=%RTbool fNmiPending=%RTbool\n",
3512	fGif, fBlockNmi, fIntShadow, VMCPU_FF_IS_ANY_SET(pVCpu, VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC),
3513	VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_INTERRUPT_NMI)));
3514
3515	/** @todo SMI. SMIs take priority over NMIs. */
3516
3517	/*
3518	* Check if the guest or nested-guest can receive NMIs.
3519	* Nested NMIs are not allowed, see AMD spec. 8.1.4 "Masking External Interrupts".
3520	* NMIs take priority over maskable interrupts, see AMD spec. 8.5 "Priorities".
3521	*/
3522	if ( VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_INTERRUPT_NMI)
3523	&& !fBlockNmi)
3524	{
3525	if ( fGif
3526	&& !fIntShadow)
3527	{
3528	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
3529	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_NMI))
3530	{
3531	Log4(("Intercepting NMI -> #VMEXIT\n"));
3532	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
3533	return IEMExecSvmVmexit(pVCpu, SVM_EXIT_NMI, 0, 0);
3534	}
3535	#endif
3536	Log4(("Setting NMI pending for injection\n"));
3537	SVMEVENT Event;
3538	Event.u = 0;
3539	Event.n.u1Valid = 1;
3540	Event.n.u8Vector = X86_XCPT_NMI;
3541	Event.n.u3Type = SVM_EVENT_NMI;
3542	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3543	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INTERRUPT_NMI);
3544	}
3545	else if (!fGif)
3546	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_STGI);
3547	else if (!pSvmTransient->fIsNestedGuest)
3548	hmR0SvmSetIntWindowExiting(pVCpu, pVmcb);
3549	/* else: for nested-guests, interrupt-window exiting will be picked up when merging VMCB controls. */
3550	}
3551	/*
3552	* Check if the guest can receive external interrupts (PIC/APIC). Once PDMGetInterrupt()
3553	* returns a valid interrupt we -must- deliver the interrupt. We can no longer re-request
3554	* it from the APIC device.
3555	*
3556	* For nested-guests, physical interrupts always take priority over virtual interrupts.
3557	* We don't need to inject nested-guest virtual interrupts here, we can let the hardware
3558	* do that work when we execute nested-guest code esp. since all the required information
3559	* is in the VMCB, unlike physical interrupts where we need to fetch the interrupt from
3560	* the virtual interrupt controller.
3561	*
3562	* See AMD spec. 15.21.4 "Injecting Virtual (INTR) Interrupts".
3563	*/
3564	else if ( VMCPU_FF_IS_ANY_SET(pVCpu, VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC)
3565	&& !pVCpu->hm.s.fSingleInstruction)
3566	{
3567	bool const fBlockInt = !pSvmTransient->fIsNestedGuest ? !(pCtx->eflags.u32 & X86_EFL_IF)
3568	: CPUMIsGuestSvmPhysIntrEnabled(pVCpu, pCtx);
3569	if ( fGif
3570	&& !fBlockInt
3571	&& !fIntShadow)
3572	{
3573	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
3574	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_INTR))
3575	{
3576	Log4(("Intercepting INTR -> #VMEXIT\n"));
3577	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
3578	return IEMExecSvmVmexit(pVCpu, SVM_EXIT_INTR, 0, 0);
3579	}
3580	#endif
3581	uint8_t u8Interrupt;
3582	int rc = PDMGetInterrupt(pVCpu, &u8Interrupt);
3583	if (RT_SUCCESS(rc))
3584	{
3585	Log4(("Setting external interrupt %#x pending for injection\n", u8Interrupt));
3586	SVMEVENT Event;
3587	Event.u = 0;
3588	Event.n.u1Valid = 1;
3589	Event.n.u8Vector = u8Interrupt;
3590	Event.n.u3Type = SVM_EVENT_EXTERNAL_IRQ;
3591	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3592	}
3593	else if (rc == VERR_APIC_INTR_MASKED_BY_TPR)
3594	{
3595	/*
3596	* AMD-V has no TPR thresholding feature. TPR and the force-flag will be
3597	* updated eventually when the TPR is written by the guest.
3598	*/
3599	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchTprMaskedIrq);
3600	}
3601	else
3602	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchGuestIrq);
3603	}
3604	else if (!fGif)
3605	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_STGI);
3606	else if (!pSvmTransient->fIsNestedGuest)
3607	hmR0SvmSetIntWindowExiting(pVCpu, pVmcb);
3608	/* else: for nested-guests, interrupt-window exiting will be picked up when merging VMCB controls. */
3609	}
3610
3611	return VINF_SUCCESS;
3612	}
3613
3614
3615	/**
3616	* Injects any pending events into the guest (or nested-guest).
3617	*
3618	* @param pVCpu The cross context virtual CPU structure.
3619	* @param pVmcb Pointer to the VM control block.
3620	*
3621	* @remarks Must only be called when we are guaranteed to enter
3622	* hardware-assisted SVM execution and not return to ring-3
3623	* prematurely.
3624	*/
3625	static void hmR0SvmInjectPendingEvent(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
3626	{
3627	Assert(!TRPMHasTrap(pVCpu));
3628	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
3629
3630	bool const fIntShadow = hmR0SvmIsIntrShadowActive(pVCpu);
3631	#ifdef VBOX_STRICT
3632	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
3633	bool const fGif = CPUMGetGuestGif(pCtx);
3634	bool fAllowInt = fGif;
3635	if (fGif)
3636	{
3637	/*
3638	* For nested-guests we have no way to determine if we're injecting a physical or
3639	* virtual interrupt at this point. Hence the partial verification below.
3640	*/
3641	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
3642	fAllowInt = CPUMIsGuestSvmPhysIntrEnabled(pVCpu, pCtx) \|\| CPUMIsGuestSvmVirtIntrEnabled(pVCpu, pCtx);
3643	else
3644	fAllowInt = RT_BOOL(pCtx->eflags.u32 & X86_EFL_IF);
3645	}
3646	#endif
3647
3648	if (pVCpu->hm.s.Event.fPending)
3649	{
3650	SVMEVENT Event;
3651	Event.u = pVCpu->hm.s.Event.u64IntInfo;
3652	Assert(Event.n.u1Valid);
3653
3654	/*
3655	* Validate event injection pre-conditions.
3656	*/
3657	if (Event.n.u3Type == SVM_EVENT_EXTERNAL_IRQ)
3658	{
3659	Assert(fAllowInt);
3660	Assert(!fIntShadow);
3661	}
3662	else if (Event.n.u3Type == SVM_EVENT_NMI)
3663	{
3664	Assert(fGif);
3665	Assert(!fIntShadow);
3666	}
3667
3668	/*
3669	* Before injecting an NMI we must set VMCPU_FF_BLOCK_NMIS to prevent nested NMIs. We
3670	* do this only when we are surely going to inject the NMI as otherwise if we return
3671	* to ring-3 prematurely we could leave NMIs blocked indefinitely upon re-entry into
3672	* SVM R0.
3673	*
3674	* With VT-x, this is handled by the Guest interruptibility information VMCS field
3675	* which will set the VMCS field after actually delivering the NMI which we read on
3676	* VM-exit to determine the state.
3677	*/
3678	if ( Event.n.u3Type == SVM_EVENT_NMI
3679	&& Event.n.u8Vector == X86_XCPT_NMI
3680	&& !VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_BLOCK_NMIS))
3681	{
3682	VMCPU_FF_SET(pVCpu, VMCPU_FF_BLOCK_NMIS);
3683	}
3684
3685	/*
3686	* Inject it (update VMCB for injection by the hardware).
3687	*/
3688	Log4(("Injecting pending HM event\n"));
3689	hmR0SvmInjectEventVmcb(pVCpu, pVmcb, &Event);
3690	pVCpu->hm.s.Event.fPending = false;
3691
3692	if (Event.n.u3Type == SVM_EVENT_EXTERNAL_IRQ)
3693	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectInterrupt);
3694	else
3695	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectXcpt);
3696	}
3697	else
3698	Assert(pVmcb->ctrl.EventInject.n.u1Valid == 0);
3699
3700	/*
3701	* We could have injected an NMI through IEM and continue guest execution using
3702	* hardware-assisted SVM. In which case, we would not have any events pending (above)
3703	* but we still need to intercept IRET in order to eventually clear NMI inhibition.
3704	*/
3705	if (VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_BLOCK_NMIS))
3706	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_IRET);
3707
3708	/*
3709	* Update the guest interrupt shadow in the guest (or nested-guest) VMCB.
3710	*
3711	* For nested-guests: We need to update it too for the scenario where IEM executes
3712	* the nested-guest but execution later continues here with an interrupt shadow active.
3713	*/
3714	pVmcb->ctrl.IntShadow.n.u1IntShadow = fIntShadow;
3715	}
3716
3717
3718	/**
3719	* Reports world-switch error and dumps some useful debug info.
3720	*
3721	* @param pVCpu The cross context virtual CPU structure.
3722	* @param rcVMRun The return code from VMRUN (or
3723	* VERR_SVM_INVALID_GUEST_STATE for invalid
3724	* guest-state).
3725	*/
3726	static void hmR0SvmReportWorldSwitchError(PVMCPUCC pVCpu, int rcVMRun)
3727	{
3728	HMSVM_ASSERT_PREEMPT_SAFE(pVCpu);
3729	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(&pVCpu->cpum.GstCtx);
3730	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
3731
3732	if (rcVMRun == VERR_SVM_INVALID_GUEST_STATE)
3733	{
3734	#ifdef VBOX_STRICT
3735	hmR0DumpRegs(pVCpu, HM_DUMP_REG_FLAGS_ALL);
3736	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
3737	Log4(("ctrl.u32VmcbCleanBits %#RX32\n", pVmcb->ctrl.u32VmcbCleanBits));
3738	Log4(("ctrl.u16InterceptRdCRx %#x\n", pVmcb->ctrl.u16InterceptRdCRx));
3739	Log4(("ctrl.u16InterceptWrCRx %#x\n", pVmcb->ctrl.u16InterceptWrCRx));
3740	Log4(("ctrl.u16InterceptRdDRx %#x\n", pVmcb->ctrl.u16InterceptRdDRx));
3741	Log4(("ctrl.u16InterceptWrDRx %#x\n", pVmcb->ctrl.u16InterceptWrDRx));
3742	Log4(("ctrl.u32InterceptXcpt %#x\n", pVmcb->ctrl.u32InterceptXcpt));
3743	Log4(("ctrl.u64InterceptCtrl %#RX64\n", pVmcb->ctrl.u64InterceptCtrl));
3744	Log4(("ctrl.u64IOPMPhysAddr %#RX64\n", pVmcb->ctrl.u64IOPMPhysAddr));
3745	Log4(("ctrl.u64MSRPMPhysAddr %#RX64\n", pVmcb->ctrl.u64MSRPMPhysAddr));
3746	Log4(("ctrl.u64TSCOffset %#RX64\n", pVmcb->ctrl.u64TSCOffset));
3747
3748	Log4(("ctrl.TLBCtrl.u32ASID %#x\n", pVmcb->ctrl.TLBCtrl.n.u32ASID));
3749	Log4(("ctrl.TLBCtrl.u8TLBFlush %#x\n", pVmcb->ctrl.TLBCtrl.n.u8TLBFlush));
3750	Log4(("ctrl.TLBCtrl.u24Reserved %#x\n", pVmcb->ctrl.TLBCtrl.n.u24Reserved));
3751
3752	Log4(("ctrl.IntCtrl.u8VTPR %#x\n", pVmcb->ctrl.IntCtrl.n.u8VTPR));
3753	Log4(("ctrl.IntCtrl.u1VIrqPending %#x\n", pVmcb->ctrl.IntCtrl.n.u1VIrqPending));
3754	Log4(("ctrl.IntCtrl.u1VGif %#x\n", pVmcb->ctrl.IntCtrl.n.u1VGif));
3755	Log4(("ctrl.IntCtrl.u6Reserved0 %#x\n", pVmcb->ctrl.IntCtrl.n.u6Reserved));
3756	Log4(("ctrl.IntCtrl.u4VIntrPrio %#x\n", pVmcb->ctrl.IntCtrl.n.u4VIntrPrio));
3757	Log4(("ctrl.IntCtrl.u1IgnoreTPR %#x\n", pVmcb->ctrl.IntCtrl.n.u1IgnoreTPR));
3758	Log4(("ctrl.IntCtrl.u3Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u3Reserved));
3759	Log4(("ctrl.IntCtrl.u1VIntrMasking %#x\n", pVmcb->ctrl.IntCtrl.n.u1VIntrMasking));
3760	Log4(("ctrl.IntCtrl.u1VGifEnable %#x\n", pVmcb->ctrl.IntCtrl.n.u1VGifEnable));
3761	Log4(("ctrl.IntCtrl.u5Reserved1 %#x\n", pVmcb->ctrl.IntCtrl.n.u5Reserved));
3762	Log4(("ctrl.IntCtrl.u8VIntrVector %#x\n", pVmcb->ctrl.IntCtrl.n.u8VIntrVector));
3763	Log4(("ctrl.IntCtrl.u24Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u24Reserved));
3764
3765	Log4(("ctrl.IntShadow.u1IntShadow %#x\n", pVmcb->ctrl.IntShadow.n.u1IntShadow));
3766	Log4(("ctrl.IntShadow.u1GuestIntMask %#x\n", pVmcb->ctrl.IntShadow.n.u1GuestIntMask));
3767	Log4(("ctrl.u64ExitCode %#RX64\n", pVmcb->ctrl.u64ExitCode));
3768	Log4(("ctrl.u64ExitInfo1 %#RX64\n", pVmcb->ctrl.u64ExitInfo1));
3769	Log4(("ctrl.u64ExitInfo2 %#RX64\n", pVmcb->ctrl.u64ExitInfo2));
3770	Log4(("ctrl.ExitIntInfo.u8Vector %#x\n", pVmcb->ctrl.ExitIntInfo.n.u8Vector));
3771	Log4(("ctrl.ExitIntInfo.u3Type %#x\n", pVmcb->ctrl.ExitIntInfo.n.u3Type));
3772	Log4(("ctrl.ExitIntInfo.u1ErrorCodeValid %#x\n", pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid));
3773	Log4(("ctrl.ExitIntInfo.u19Reserved %#x\n", pVmcb->ctrl.ExitIntInfo.n.u19Reserved));
3774	Log4(("ctrl.ExitIntInfo.u1Valid %#x\n", pVmcb->ctrl.ExitIntInfo.n.u1Valid));
3775	Log4(("ctrl.ExitIntInfo.u32ErrorCode %#x\n", pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode));
3776	Log4(("ctrl.NestedPagingCtrl.u1NestedPaging %#x\n", pVmcb->ctrl.NestedPagingCtrl.n.u1NestedPaging));
3777	Log4(("ctrl.NestedPagingCtrl.u1Sev %#x\n", pVmcb->ctrl.NestedPagingCtrl.n.u1Sev));
3778	Log4(("ctrl.NestedPagingCtrl.u1SevEs %#x\n", pVmcb->ctrl.NestedPagingCtrl.n.u1SevEs));
3779	Log4(("ctrl.EventInject.u8Vector %#x\n", pVmcb->ctrl.EventInject.n.u8Vector));
3780	Log4(("ctrl.EventInject.u3Type %#x\n", pVmcb->ctrl.EventInject.n.u3Type));
3781	Log4(("ctrl.EventInject.u1ErrorCodeValid %#x\n", pVmcb->ctrl.EventInject.n.u1ErrorCodeValid));
3782	Log4(("ctrl.EventInject.u19Reserved %#x\n", pVmcb->ctrl.EventInject.n.u19Reserved));
3783	Log4(("ctrl.EventInject.u1Valid %#x\n", pVmcb->ctrl.EventInject.n.u1Valid));
3784	Log4(("ctrl.EventInject.u32ErrorCode %#x\n", pVmcb->ctrl.EventInject.n.u32ErrorCode));
3785
3786	Log4(("ctrl.u64NestedPagingCR3 %#RX64\n", pVmcb->ctrl.u64NestedPagingCR3));
3787
3788	Log4(("ctrl.LbrVirt.u1LbrVirt %#x\n", pVmcb->ctrl.LbrVirt.n.u1LbrVirt));
3789	Log4(("ctrl.LbrVirt.u1VirtVmsaveVmload %#x\n", pVmcb->ctrl.LbrVirt.n.u1VirtVmsaveVmload));
3790
3791	Log4(("guest.CS.u16Sel %RTsel\n", pVmcb->guest.CS.u16Sel));
3792	Log4(("guest.CS.u16Attr %#x\n", pVmcb->guest.CS.u16Attr));
3793	Log4(("guest.CS.u32Limit %#RX32\n", pVmcb->guest.CS.u32Limit));
3794	Log4(("guest.CS.u64Base %#RX64\n", pVmcb->guest.CS.u64Base));
3795	Log4(("guest.DS.u16Sel %#RTsel\n", pVmcb->guest.DS.u16Sel));
3796	Log4(("guest.DS.u16Attr %#x\n", pVmcb->guest.DS.u16Attr));
3797	Log4(("guest.DS.u32Limit %#RX32\n", pVmcb->guest.DS.u32Limit));
3798	Log4(("guest.DS.u64Base %#RX64\n", pVmcb->guest.DS.u64Base));
3799	Log4(("guest.ES.u16Sel %RTsel\n", pVmcb->guest.ES.u16Sel));
3800	Log4(("guest.ES.u16Attr %#x\n", pVmcb->guest.ES.u16Attr));
3801	Log4(("guest.ES.u32Limit %#RX32\n", pVmcb->guest.ES.u32Limit));
3802	Log4(("guest.ES.u64Base %#RX64\n", pVmcb->guest.ES.u64Base));
3803	Log4(("guest.FS.u16Sel %RTsel\n", pVmcb->guest.FS.u16Sel));
3804	Log4(("guest.FS.u16Attr %#x\n", pVmcb->guest.FS.u16Attr));
3805	Log4(("guest.FS.u32Limit %#RX32\n", pVmcb->guest.FS.u32Limit));
3806	Log4(("guest.FS.u64Base %#RX64\n", pVmcb->guest.FS.u64Base));
3807	Log4(("guest.GS.u16Sel %RTsel\n", pVmcb->guest.GS.u16Sel));
3808	Log4(("guest.GS.u16Attr %#x\n", pVmcb->guest.GS.u16Attr));
3809	Log4(("guest.GS.u32Limit %#RX32\n", pVmcb->guest.GS.u32Limit));
3810	Log4(("guest.GS.u64Base %#RX64\n", pVmcb->guest.GS.u64Base));
3811
3812	Log4(("guest.GDTR.u32Limit %#RX32\n", pVmcb->guest.GDTR.u32Limit));
3813	Log4(("guest.GDTR.u64Base %#RX64\n", pVmcb->guest.GDTR.u64Base));
3814
3815	Log4(("guest.LDTR.u16Sel %RTsel\n", pVmcb->guest.LDTR.u16Sel));
3816	Log4(("guest.LDTR.u16Attr %#x\n", pVmcb->guest.LDTR.u16Attr));
3817	Log4(("guest.LDTR.u32Limit %#RX32\n", pVmcb->guest.LDTR.u32Limit));
3818	Log4(("guest.LDTR.u64Base %#RX64\n", pVmcb->guest.LDTR.u64Base));
3819
3820	Log4(("guest.IDTR.u32Limit %#RX32\n", pVmcb->guest.IDTR.u32Limit));
3821	Log4(("guest.IDTR.u64Base %#RX64\n", pVmcb->guest.IDTR.u64Base));
3822
3823	Log4(("guest.TR.u16Sel %RTsel\n", pVmcb->guest.TR.u16Sel));
3824	Log4(("guest.TR.u16Attr %#x\n", pVmcb->guest.TR.u16Attr));
3825	Log4(("guest.TR.u32Limit %#RX32\n", pVmcb->guest.TR.u32Limit));
3826	Log4(("guest.TR.u64Base %#RX64\n", pVmcb->guest.TR.u64Base));
3827
3828	Log4(("guest.u8CPL %#x\n", pVmcb->guest.u8CPL));
3829	Log4(("guest.u64CR0 %#RX64\n", pVmcb->guest.u64CR0));
3830	Log4(("guest.u64CR2 %#RX64\n", pVmcb->guest.u64CR2));
3831	Log4(("guest.u64CR3 %#RX64\n", pVmcb->guest.u64CR3));
3832	Log4(("guest.u64CR4 %#RX64\n", pVmcb->guest.u64CR4));
3833	Log4(("guest.u64DR6 %#RX64\n", pVmcb->guest.u64DR6));
3834	Log4(("guest.u64DR7 %#RX64\n", pVmcb->guest.u64DR7));
3835
3836	Log4(("guest.u64RIP %#RX64\n", pVmcb->guest.u64RIP));
3837	Log4(("guest.u64RSP %#RX64\n", pVmcb->guest.u64RSP));
3838	Log4(("guest.u64RAX %#RX64\n", pVmcb->guest.u64RAX));
3839	Log4(("guest.u64RFlags %#RX64\n", pVmcb->guest.u64RFlags));
3840
3841	Log4(("guest.u64SysEnterCS %#RX64\n", pVmcb->guest.u64SysEnterCS));
3842	Log4(("guest.u64SysEnterEIP %#RX64\n", pVmcb->guest.u64SysEnterEIP));
3843	Log4(("guest.u64SysEnterESP %#RX64\n", pVmcb->guest.u64SysEnterESP));
3844
3845	Log4(("guest.u64EFER %#RX64\n", pVmcb->guest.u64EFER));
3846	Log4(("guest.u64STAR %#RX64\n", pVmcb->guest.u64STAR));
3847	Log4(("guest.u64LSTAR %#RX64\n", pVmcb->guest.u64LSTAR));
3848	Log4(("guest.u64CSTAR %#RX64\n", pVmcb->guest.u64CSTAR));
3849	Log4(("guest.u64SFMASK %#RX64\n", pVmcb->guest.u64SFMASK));
3850	Log4(("guest.u64KernelGSBase %#RX64\n", pVmcb->guest.u64KernelGSBase));
3851	Log4(("guest.u64PAT %#RX64\n", pVmcb->guest.u64PAT));
3852	Log4(("guest.u64DBGCTL %#RX64\n", pVmcb->guest.u64DBGCTL));
3853	Log4(("guest.u64BR_FROM %#RX64\n", pVmcb->guest.u64BR_FROM));
3854	Log4(("guest.u64BR_TO %#RX64\n", pVmcb->guest.u64BR_TO));
3855	Log4(("guest.u64LASTEXCPFROM %#RX64\n", pVmcb->guest.u64LASTEXCPFROM));
3856	Log4(("guest.u64LASTEXCPTO %#RX64\n", pVmcb->guest.u64LASTEXCPTO));
3857
3858	NOREF(pVmcb);
3859	#endif /* VBOX_STRICT */
3860	}
3861	else
3862	Log4Func(("rcVMRun=%d\n", rcVMRun));
3863	}
3864
3865
3866	/**
3867	* Check per-VM and per-VCPU force flag actions that require us to go back to
3868	* ring-3 for one reason or another.
3869	*
3870	* @returns VBox status code (information status code included).
3871	* @retval VINF_SUCCESS if we don't have any actions that require going back to
3872	* ring-3.
3873	* @retval VINF_PGM_SYNC_CR3 if we have pending PGM CR3 sync.
3874	* @retval VINF_EM_PENDING_REQUEST if we have pending requests (like hardware
3875	* interrupts)
3876	* @retval VINF_PGM_POOL_FLUSH_PENDING if PGM is doing a pool flush and requires
3877	* all EMTs to be in ring-3.
3878	* @retval VINF_EM_RAW_TO_R3 if there is pending DMA requests.
3879	* @retval VINF_EM_NO_MEMORY PGM is out of memory, we need to return
3880	* to the EM loop.
3881	*
3882	* @param pVCpu The cross context virtual CPU structure.
3883	*/
3884	static int hmR0SvmCheckForceFlags(PVMCPUCC pVCpu)
3885	{
3886	Assert(VMMRZCallRing3IsEnabled(pVCpu));
3887	Assert(!VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_HM_UPDATE_PAE_PDPES));
3888
3889	/* Could happen as a result of longjump. */
3890	if (VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_HM_UPDATE_CR3))
3891	PGMUpdateCR3(pVCpu, CPUMGetGuestCR3(pVCpu));
3892
3893	/* Update pending interrupts into the APIC's IRR. */
3894	if (VMCPU_FF_TEST_AND_CLEAR(pVCpu, VMCPU_FF_UPDATE_APIC))
3895	APICUpdatePendingInterrupts(pVCpu);
3896
3897	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
3898	if ( VM_FF_IS_ANY_SET(pVM, !pVCpu->hm.s.fSingleInstruction
3899	? VM_FF_HP_R0_PRE_HM_MASK : VM_FF_HP_R0_PRE_HM_STEP_MASK)
3900	\|\| VMCPU_FF_IS_ANY_SET(pVCpu, !pVCpu->hm.s.fSingleInstruction
3901	? VMCPU_FF_HP_R0_PRE_HM_MASK : VMCPU_FF_HP_R0_PRE_HM_STEP_MASK) )
3902	{
3903	/* Pending PGM C3 sync. */
3904	if (VMCPU_FF_IS_ANY_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3 \| VMCPU_FF_PGM_SYNC_CR3_NON_GLOBAL))
3905	{
3906	int rc = PGMSyncCR3(pVCpu, pVCpu->cpum.GstCtx.cr0, pVCpu->cpum.GstCtx.cr3, pVCpu->cpum.GstCtx.cr4,
3907	VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3));
3908	if (rc != VINF_SUCCESS)
3909	{
3910	Log4Func(("PGMSyncCR3 forcing us back to ring-3. rc=%d\n", rc));
3911	return rc;
3912	}
3913	}
3914
3915	/* Pending HM-to-R3 operations (critsects, timers, EMT rendezvous etc.) */
3916	/* -XXX- what was that about single stepping? */
3917	if ( VM_FF_IS_ANY_SET(pVM, VM_FF_HM_TO_R3_MASK)
3918	\|\| VMCPU_FF_IS_ANY_SET(pVCpu, VMCPU_FF_HM_TO_R3_MASK))
3919	{
3920	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHmToR3FF);
3921	int rc = RT_LIKELY(!VM_FF_IS_SET(pVM, VM_FF_PGM_NO_MEMORY)) ? VINF_EM_RAW_TO_R3 : VINF_EM_NO_MEMORY;
3922	Log4Func(("HM_TO_R3 forcing us back to ring-3. rc=%d\n", rc));
3923	return rc;
3924	}
3925
3926	/* Pending VM request packets, such as hardware interrupts. */
3927	if ( VM_FF_IS_SET(pVM, VM_FF_REQUEST)
3928	\|\| VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_REQUEST))
3929	{
3930	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchVmReq);
3931	Log4Func(("Pending VM request forcing us back to ring-3\n"));
3932	return VINF_EM_PENDING_REQUEST;
3933	}
3934
3935	/* Pending PGM pool flushes. */
3936	if (VM_FF_IS_SET(pVM, VM_FF_PGM_POOL_FLUSH_PENDING))
3937	{
3938	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchPgmPoolFlush);
3939	Log4Func(("PGM pool flush pending forcing us back to ring-3\n"));
3940	return VINF_PGM_POOL_FLUSH_PENDING;
3941	}
3942
3943	/* Pending DMA requests. */
3944	if (VM_FF_IS_SET(pVM, VM_FF_PDM_DMA))
3945	{
3946	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchDma);
3947	Log4Func(("Pending DMA request forcing us back to ring-3\n"));
3948	return VINF_EM_RAW_TO_R3;
3949	}
3950	}
3951
3952	return VINF_SUCCESS;
3953	}
3954
3955
3956	/**
3957	* Does the preparations before executing guest code in AMD-V.
3958	*
3959	* This may cause longjmps to ring-3 and may even result in rescheduling to the
3960	* recompiler. We must be cautious what we do here regarding committing
3961	* guest-state information into the VMCB assuming we assuredly execute the guest
3962	* in AMD-V. If we fall back to the recompiler after updating the VMCB and
3963	* clearing the common-state (TRPM/forceflags), we must undo those changes so
3964	* that the recompiler can (and should) use them when it resumes guest
3965	* execution. Otherwise such operations must be done when we can no longer
3966	* exit to ring-3.
3967	*
3968	* @returns VBox status code (informational status codes included).
3969	* @retval VINF_SUCCESS if we can proceed with running the guest.
3970	* @retval VINF_* scheduling changes, we have to go back to ring-3.
3971	*
3972	* @param pVCpu The cross context virtual CPU structure.
3973	* @param pSvmTransient Pointer to the SVM transient structure.
3974	*/
3975	static int hmR0SvmPreRunGuest(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
3976	{
3977	HMSVM_ASSERT_PREEMPT_SAFE(pVCpu);
3978
3979	#ifdef VBOX_WITH_NESTED_HWVIRT_ONLY_IN_IEM
3980	if (pSvmTransient->fIsNestedGuest)
3981	{
3982	Log2(("hmR0SvmPreRunGuest: Rescheduling to IEM due to nested-hwvirt or forced IEM exec -> VINF_EM_RESCHEDULE_REM\n"));
3983	return VINF_EM_RESCHEDULE_REM;
3984	}
3985	#endif
3986
3987	/* Check force flag actions that might require us to go back to ring-3. */
3988	int rc = hmR0SvmCheckForceFlags(pVCpu);
3989	if (rc != VINF_SUCCESS)
3990	return rc;
3991
3992	if (TRPMHasTrap(pVCpu))
3993	hmR0SvmTrpmTrapToPendingEvent(pVCpu);
3994	else if (!pVCpu->hm.s.Event.fPending)
3995	{
3996	VBOXSTRICTRC rcStrict = hmR0SvmEvaluatePendingEvent(pVCpu, pSvmTransient);
3997	if ( rcStrict != VINF_SUCCESS
3998	\|\| pSvmTransient->fIsNestedGuest != CPUMIsGuestInSvmNestedHwVirtMode(&pVCpu->cpum.GstCtx))
3999	{
4000	/* If a nested-guest VM-exit occurred, bail. */
4001	if (pSvmTransient->fIsNestedGuest)
4002	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchNstGstVmexit);
4003	return VBOXSTRICTRC_VAL(rcStrict);
4004	}
4005	}
4006
4007	/*
4008	* On the oldest AMD-V systems, we may not get enough information to reinject an NMI.
4009	* Just do it in software, see @bugref{8411}.
4010	* NB: If we could continue a task switch exit we wouldn't need to do this.
4011	*/
4012	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
4013	if (RT_UNLIKELY( !pVM->hm.s.svm.u32Features
4014	&& pVCpu->hm.s.Event.fPending
4015	&& SVM_EVENT_GET_TYPE(pVCpu->hm.s.Event.u64IntInfo) == SVM_EVENT_NMI))
4016	return VINF_EM_RAW_INJECT_TRPM_EVENT;
4017
4018	#ifdef HMSVM_SYNC_FULL_GUEST_STATE
4019	Assert(!(pVCpu->cpum.GstCtx.fExtrn & HMSVM_CPUMCTX_EXTRN_ALL));
4020	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
4021	#endif
4022
4023	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
4024	/*
4025	* Set up the nested-guest VMCB for execution using hardware-assisted SVM.
4026	*/
4027	if (pSvmTransient->fIsNestedGuest)
4028	hmR0SvmSetupVmcbNested(pVCpu);
4029	#endif
4030
4031	/*
4032	* Export the guest state bits that are not shared with the host in any way as we can
4033	* longjmp or get preempted in the midst of exporting some of the state.
4034	*/
4035	rc = hmR0SvmExportGuestState(pVCpu, pSvmTransient);
4036	AssertRCReturn(rc, rc);
4037	STAM_COUNTER_INC(&pVCpu->hm.s.StatExportFull);
4038
4039	/* Ensure we've cached (and hopefully modified) the nested-guest VMCB for execution using hardware-assisted SVM. */
4040	Assert(!pSvmTransient->fIsNestedGuest \|\| pVCpu->hm.s.svm.NstGstVmcbCache.fCacheValid);
4041
4042	/*
4043	* If we're not intercepting TPR changes in the guest, save the guest TPR before the
4044	* world-switch so we can update it on the way back if the guest changed the TPR.
4045	*/
4046	if (pVCpu->hm.s.svm.fSyncVTpr)
4047	{
4048	Assert(!pSvmTransient->fIsNestedGuest);
4049	PCSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
4050	if (pVM->hm.s.fTPRPatchingActive)
4051	pSvmTransient->u8GuestTpr = pVmcb->guest.u64LSTAR;
4052	else
4053	pSvmTransient->u8GuestTpr = pVmcb->ctrl.IntCtrl.n.u8VTPR;
4054	}
4055
4056	/*
4057	* No longjmps to ring-3 from this point on!!!
4058	*
4059	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional,
4060	* better than a kernel panic. This also disables flushing of the R0-logger instance.
4061	*/
4062	VMMRZCallRing3Disable(pVCpu);
4063
4064	/*
4065	* We disable interrupts so that we don't miss any interrupts that would flag preemption
4066	* (IPI/timers etc.) when thread-context hooks aren't used and we've been running with
4067	* preemption disabled for a while. Since this is purly to aid the
4068	* RTThreadPreemptIsPending() code, it doesn't matter that it may temporarily reenable and
4069	* disable interrupt on NT.
4070	*
4071	* We need to check for force-flags that could've possible been altered since we last
4072	* checked them (e.g. by PDMGetInterrupt() leaving the PDM critical section,
4073	* see @bugref{6398}).
4074	*
4075	* We also check a couple of other force-flags as a last opportunity to get the EMT back
4076	* to ring-3 before executing guest code.
4077	*/
4078	pSvmTransient->fEFlags = ASMIntDisableFlags();
4079	if ( VM_FF_IS_ANY_SET(pVM, VM_FF_EMT_RENDEZVOUS \| VM_FF_TM_VIRTUAL_SYNC)
4080	\|\| VMCPU_FF_IS_ANY_SET(pVCpu, VMCPU_FF_HM_TO_R3_MASK))
4081	{
4082	ASMSetFlags(pSvmTransient->fEFlags);
4083	VMMRZCallRing3Enable(pVCpu);
4084	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHmToR3FF);
4085	return VINF_EM_RAW_TO_R3;
4086	}
4087	if (RTThreadPreemptIsPending(NIL_RTTHREAD))
4088	{
4089	ASMSetFlags(pSvmTransient->fEFlags);
4090	VMMRZCallRing3Enable(pVCpu);
4091	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchPendingHostIrq);
4092	return VINF_EM_RAW_INTERRUPT;
4093	}
4094
4095	return VINF_SUCCESS;
4096	}
4097
4098
4099	/**
4100	* Prepares to run guest (or nested-guest) code in AMD-V and we've committed to
4101	* doing so.
4102	*
4103	* This means there is no backing out to ring-3 or anywhere else at this point.
4104	*
4105	* @param pVCpu The cross context virtual CPU structure.
4106	* @param pSvmTransient Pointer to the SVM transient structure.
4107	*
4108	* @remarks Called with preemption disabled.
4109	* @remarks No-long-jump zone!!!
4110	*/
4111	static void hmR0SvmPreRunGuestCommitted(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
4112	{
4113	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
4114	Assert(VMMR0IsLogFlushDisabled(pVCpu));
4115	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
4116
4117	VMCPU_ASSERT_STATE(pVCpu, VMCPUSTATE_STARTED_HM);
4118	VMCPU_SET_STATE(pVCpu, VMCPUSTATE_STARTED_EXEC); /* Indicate the start of guest execution. */
4119
4120	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
4121	PSVMVMCB pVmcb = pSvmTransient->pVmcb;
4122
4123	hmR0SvmInjectPendingEvent(pVCpu, pVmcb);
4124
4125	if (!CPUMIsGuestFPUStateActive(pVCpu))
4126	{
4127	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatLoadGuestFpuState, x);
4128	CPUMR0LoadGuestFPU(pVM, pVCpu); /* (Ignore rc, no need to set HM_CHANGED_HOST_CONTEXT for SVM.) */
4129	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatLoadGuestFpuState, x);
4130	STAM_COUNTER_INC(&pVCpu->hm.s.StatLoadGuestFpu);
4131	}
4132
4133	/* Load the state shared between host and guest (FPU, debug). */
4134	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE)
4135	hmR0SvmExportSharedState(pVCpu, pVmcb);
4136
4137	pVCpu->hm.s.fCtxChanged &= ~HM_CHANGED_HOST_CONTEXT; /* Preemption might set this, nothing to do on AMD-V. */
4138	AssertMsg(!pVCpu->hm.s.fCtxChanged, ("fCtxChanged=%#RX64\n", pVCpu->hm.s.fCtxChanged));
4139
4140	PHMPHYSCPU pHostCpu = hmR0GetCurrentCpu();
4141	RTCPUID const idHostCpu = pHostCpu->idCpu;
4142	bool const fMigratedHostCpu = idHostCpu != pVCpu->hm.s.idLastCpu;
4143
4144	/* Setup TSC offsetting. */
4145	if ( pSvmTransient->fUpdateTscOffsetting
4146	\|\| fMigratedHostCpu)
4147	{
4148	hmR0SvmUpdateTscOffsetting(pVCpu, pVmcb);
4149	pSvmTransient->fUpdateTscOffsetting = false;
4150	}
4151
4152	/* Record statistics of how often we use TSC offsetting as opposed to intercepting RDTSC/P. */
4153	if (!(pVmcb->ctrl.u64InterceptCtrl & (SVM_CTRL_INTERCEPT_RDTSC \| SVM_CTRL_INTERCEPT_RDTSCP)))
4154	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscOffset);
4155	else
4156	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscIntercept);
4157
4158	/* If we've migrating CPUs, mark the VMCB Clean bits as dirty. */
4159	if (fMigratedHostCpu)
4160	pVmcb->ctrl.u32VmcbCleanBits = 0;
4161
4162	/* Store status of the shared guest-host state at the time of VMRUN. */
4163	pSvmTransient->fWasGuestDebugStateActive = CPUMIsGuestDebugStateActive(pVCpu);
4164	pSvmTransient->fWasHyperDebugStateActive = CPUMIsHyperDebugStateActive(pVCpu);
4165
4166	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
4167	uint8_t *pbMsrBitmap;
4168	if (!pSvmTransient->fIsNestedGuest)
4169	pbMsrBitmap = (uint8_t *)pVCpu->hm.s.svm.pvMsrBitmap;
4170	else
4171	{
4172	hmR0SvmMergeMsrpmNested(pHostCpu, pVCpu);
4173
4174	/* Update the nested-guest VMCB with the newly merged MSRPM (clean bits updated below). */
4175	pVmcb->ctrl.u64MSRPMPhysAddr = pHostCpu->n.svm.HCPhysNstGstMsrpm;
4176	pbMsrBitmap = (uint8_t *)pHostCpu->n.svm.pvNstGstMsrpm;
4177	}
4178	#else
4179	uint8_t pbMsrBitmap = (uint8_t )pVCpu->hm.s.svm.pvMsrBitmap;
4180	#endif
4181
4182	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, true); /* Used for TLB flushing, set this across the world switch. */
4183	/* Flush the appropriate tagged-TLB entries. */
4184	hmR0SvmFlushTaggedTlb(pHostCpu, pVCpu, pVmcb);
4185	Assert(pVCpu->hm.s.idLastCpu == idHostCpu);
4186
4187	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatEntry, &pVCpu->hm.s.StatInGC, x);
4188
4189	TMNotifyStartOfExecution(pVM, pVCpu); /* Finally, notify TM to resume its clocks as we're about
4190	to start executing. */
4191
4192	/*
4193	* Save the current Host TSC_AUX and write the guest TSC_AUX to the host, so that RDTSCPs
4194	* (that don't cause exits) reads the guest MSR, see @bugref{3324}.
4195	*
4196	* This should be done -after- any RDTSCPs for obtaining the host timestamp (TM, STAM etc).
4197	*/
4198	if ( pVM->cpum.ro.HostFeatures.fRdTscP
4199	&& !(pVmcb->ctrl.u64InterceptCtrl & SVM_CTRL_INTERCEPT_RDTSCP))
4200	{
4201	uint64_t const uGuestTscAux = CPUMGetGuestTscAux(pVCpu);
4202	pVCpu->hm.s.svm.u64HostTscAux = ASMRdMsr(MSR_K8_TSC_AUX);
4203	if (uGuestTscAux != pVCpu->hm.s.svm.u64HostTscAux)
4204	ASMWrMsr(MSR_K8_TSC_AUX, uGuestTscAux);
4205	hmR0SvmSetMsrPermission(pVCpu, pbMsrBitmap, MSR_K8_TSC_AUX, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
4206	pSvmTransient->fRestoreTscAuxMsr = true;
4207	}
4208	else
4209	{
4210	hmR0SvmSetMsrPermission(pVCpu, pbMsrBitmap, MSR_K8_TSC_AUX, SVMMSREXIT_INTERCEPT_READ, SVMMSREXIT_INTERCEPT_WRITE);
4211	pSvmTransient->fRestoreTscAuxMsr = false;
4212	}
4213	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_IOPM_MSRPM;
4214
4215	/*
4216	* If VMCB Clean bits isn't supported by the CPU or exposed to the guest in the nested
4217	* virtualization case, mark all state-bits as dirty indicating to the CPU to re-load
4218	* from the VMCB.
4219	*/
4220	bool const fSupportsVmcbCleanBits = hmR0SvmSupportsVmcbCleanBits(pVCpu, pSvmTransient->fIsNestedGuest);
4221	if (!fSupportsVmcbCleanBits)
4222	pVmcb->ctrl.u32VmcbCleanBits = 0;
4223	}
4224
4225
4226	/**
4227	* Wrapper for running the guest (or nested-guest) code in AMD-V.
4228	*
4229	* @returns VBox strict status code.
4230	* @param pVCpu The cross context virtual CPU structure.
4231	* @param HCPhysVmcb The host physical address of the VMCB.
4232	*
4233	* @remarks No-long-jump zone!!!
4234	*/
4235	DECLINLINE(int) hmR0SvmRunGuest(PVMCPUCC pVCpu, RTHCPHYS HCPhysVmcb)
4236	{
4237	/* Mark that HM is the keeper of all guest-CPU registers now that we're going to execute guest code. */
4238	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
4239	pCtx->fExtrn \|= HMSVM_CPUMCTX_EXTRN_ALL \| CPUMCTX_EXTRN_KEEPER_HM;
4240
4241	/*
4242	* 64-bit Windows uses XMM registers in the kernel as the Microsoft compiler expresses
4243	* floating-point operations using SSE instructions. Some XMM registers (XMM6-XMM15) are
4244	* callee-saved and thus the need for this XMM wrapper.
4245	*
4246	* Refer MSDN "Configuring Programs for 64-bit/x64 Software Conventions / Register Usage".
4247	*/
4248	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
4249	#ifdef VBOX_WITH_KERNEL_USING_XMM
4250	return hmR0SVMRunWrapXMM(pVCpu->hm.s.svm.HCPhysVmcbHost, HCPhysVmcb, pCtx, pVM, pVCpu, pVCpu->hm.s.svm.pfnVMRun);
4251	#else
4252	return pVCpu->hm.s.svm.pfnVMRun(pVCpu->hm.s.svm.HCPhysVmcbHost, HCPhysVmcb, pCtx, pVM, pVCpu);
4253	#endif
4254	}
4255
4256
4257	/**
4258	* Performs some essential restoration of state after running guest (or
4259	* nested-guest) code in AMD-V.
4260	*
4261	* @param pVCpu The cross context virtual CPU structure.
4262	* @param pSvmTransient Pointer to the SVM transient structure.
4263	* @param rcVMRun Return code of VMRUN.
4264	*
4265	* @remarks Called with interrupts disabled.
4266	* @remarks No-long-jump zone!!! This function will however re-enable longjmps
4267	* unconditionally when it is safe to do so.
4268	*/
4269	static void hmR0SvmPostRunGuest(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient, int rcVMRun)
4270	{
4271	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
4272
4273	uint64_t const uHostTsc = ASMReadTSC(); /* Read the TSC as soon as possible. */
4274	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, false); /* See HMInvalidatePageOnAllVCpus(): used for TLB flushing. */
4275	ASMAtomicIncU32(&pVCpu->hm.s.cWorldSwitchExits); /* Initialized in vmR3CreateUVM(): used for EMT poking. */
4276
4277	PSVMVMCB pVmcb = pSvmTransient->pVmcb;
4278	PSVMVMCBCTRL pVmcbCtrl = &pVmcb->ctrl;
4279
4280	/* TSC read must be done early for maximum accuracy. */
4281	if (!(pVmcbCtrl->u64InterceptCtrl & SVM_CTRL_INTERCEPT_RDTSC))
4282	{
4283	if (!pSvmTransient->fIsNestedGuest)
4284	TMCpuTickSetLastSeen(pVCpu, uHostTsc + pVmcbCtrl->u64TSCOffset);
4285	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
4286	else
4287	{
4288	/* The nested-guest VMCB TSC offset shall eventually be restored on #VMEXIT via HMNotifySvmNstGstVmexit(). */
4289	uint64_t const uGstTsc = CPUMRemoveNestedGuestTscOffset(pVCpu, uHostTsc + pVmcbCtrl->u64TSCOffset);
4290	TMCpuTickSetLastSeen(pVCpu, uGstTsc);
4291	}
4292	#endif
4293	}
4294
4295	if (pSvmTransient->fRestoreTscAuxMsr)
4296	{
4297	uint64_t u64GuestTscAuxMsr = ASMRdMsr(MSR_K8_TSC_AUX);
4298	CPUMSetGuestTscAux(pVCpu, u64GuestTscAuxMsr);
4299	if (u64GuestTscAuxMsr != pVCpu->hm.s.svm.u64HostTscAux)
4300	ASMWrMsr(MSR_K8_TSC_AUX, pVCpu->hm.s.svm.u64HostTscAux);
4301	}
4302
4303	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatInGC, &pVCpu->hm.s.StatPreExit, x);
4304	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
4305	TMNotifyEndOfExecution(pVM, pVCpu); /* Notify TM that the guest is no longer running. */
4306	VMCPU_SET_STATE(pVCpu, VMCPUSTATE_STARTED_HM);
4307
4308	Assert(!(ASMGetFlags() & X86_EFL_IF));
4309	ASMSetFlags(pSvmTransient->fEFlags); /* Enable interrupts. */
4310	VMMRZCallRing3Enable(pVCpu); /* It is now safe to do longjmps to ring-3!!! */
4311
4312	/* If VMRUN failed, we can bail out early. This does -not- cover SVM_EXIT_INVALID. */
4313	if (RT_UNLIKELY(rcVMRun != VINF_SUCCESS))
4314	{
4315	Log4Func(("VMRUN failure: rcVMRun=%Rrc\n", rcVMRun));
4316	return;
4317	}
4318
4319	pSvmTransient->u64ExitCode = pVmcbCtrl->u64ExitCode; /* Save the #VMEXIT reason. */
4320	pSvmTransient->fVectoringDoublePF = false; /* Vectoring double page-fault needs to be determined later. */
4321	pSvmTransient->fVectoringPF = false; /* Vectoring page-fault needs to be determined later. */
4322	pVmcbCtrl->u32VmcbCleanBits = HMSVM_VMCB_CLEAN_ALL; /* Mark the VMCB-state cache as unmodified by VMM. */
4323
4324	#ifdef HMSVM_SYNC_FULL_GUEST_STATE
4325	hmR0SvmImportGuestState(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
4326	Assert(!(pVCpu->cpum.GstCtx.fExtrn & HMSVM_CPUMCTX_EXTRN_ALL));
4327	#else
4328	/*
4329	* Always import the following:
4330	*
4331	* - RIP for exit optimizations and evaluating event injection on re-entry.
4332	* - RFLAGS for evaluating event injection on VM re-entry and for exporting shared debug
4333	* state on preemption.
4334	* - Interrupt shadow, GIF for evaluating event injection on VM re-entry.
4335	* - CS for exit optimizations.
4336	* - RAX, RSP for simplifying assumptions on GPRs. All other GPRs are swapped by the
4337	* assembly switcher code.
4338	* - Shared state (only DR7 currently) for exporting shared debug state on preemption.
4339	*/
4340	hmR0SvmImportGuestState(pVCpu, CPUMCTX_EXTRN_RIP
4341	\| CPUMCTX_EXTRN_RFLAGS
4342	\| CPUMCTX_EXTRN_RAX
4343	\| CPUMCTX_EXTRN_RSP
4344	\| CPUMCTX_EXTRN_CS
4345	\| CPUMCTX_EXTRN_HWVIRT
4346	\| CPUMCTX_EXTRN_HM_SVM_INT_SHADOW
4347	\| CPUMCTX_EXTRN_HM_SVM_HWVIRT_VIRQ
4348	\| HMSVM_CPUMCTX_SHARED_STATE);
4349	#endif
4350
4351	if ( pSvmTransient->u64ExitCode != SVM_EXIT_INVALID
4352	&& pVCpu->hm.s.svm.fSyncVTpr)
4353	{
4354	Assert(!pSvmTransient->fIsNestedGuest);
4355	/* TPR patching (for 32-bit guests) uses LSTAR MSR for holding the TPR value, otherwise uses the VTPR. */
4356	if ( pVM->hm.s.fTPRPatchingActive
4357	&& (pVmcb->guest.u64LSTAR & 0xff) != pSvmTransient->u8GuestTpr)
4358	{
4359	int rc = APICSetTpr(pVCpu, pVmcb->guest.u64LSTAR & 0xff);
4360	AssertRC(rc);
4361	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
4362	}
4363	/* Sync TPR when we aren't intercepting CR8 writes. */
4364	else if (pSvmTransient->u8GuestTpr != pVmcbCtrl->IntCtrl.n.u8VTPR)
4365	{
4366	int rc = APICSetTpr(pVCpu, pVmcbCtrl->IntCtrl.n.u8VTPR << 4);
4367	AssertRC(rc);
4368	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
4369	}
4370	}
4371
4372	#ifdef DEBUG_ramshankar
4373	if (CPUMIsGuestInSvmNestedHwVirtMode(&pVCpu->cpum.GstCtx))
4374	{
4375	hmR0SvmImportGuestState(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
4376	hmR0SvmLogState(pVCpu, pVmcb, pVCpu->cpum.GstCtx, "hmR0SvmPostRunGuestNested", HMSVM_LOG_ALL & ~HMSVM_LOG_LBR,
4377	0 /* uVerbose */);
4378	}
4379	#endif
4380
4381	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_CS \| CPUMCTX_EXTRN_RIP);
4382	EMHistoryAddExit(pVCpu, EMEXIT_MAKE_FT(EMEXIT_F_KIND_SVM, pSvmTransient->u64ExitCode & EMEXIT_F_TYPE_MASK),
4383	pVCpu->cpum.GstCtx.cs.u64Base + pVCpu->cpum.GstCtx.rip, uHostTsc);
4384	}
4385
4386
4387	/**
4388	* Runs the guest code using AMD-V.
4389	*
4390	* @returns VBox status code.
4391	* @param pVCpu The cross context virtual CPU structure.
4392	* @param pcLoops Pointer to the number of executed loops.
4393	*/
4394	static int hmR0SvmRunGuestCodeNormal(PVMCPUCC pVCpu, uint32_t *pcLoops)
4395	{
4396	uint32_t const cMaxResumeLoops = pVCpu->CTX_SUFF(pVM)->hm.s.cMaxResumeLoops;
4397	Assert(pcLoops);
4398	Assert(*pcLoops <= cMaxResumeLoops);
4399
4400	SVMTRANSIENT SvmTransient;
4401	RT_ZERO(SvmTransient);
4402	SvmTransient.fUpdateTscOffsetting = true;
4403	SvmTransient.pVmcb = pVCpu->hm.s.svm.pVmcb;
4404
4405	int rc = VERR_INTERNAL_ERROR_5;
4406	for (;;)
4407	{
4408	Assert(!HMR0SuspendPending());
4409	HMSVM_ASSERT_CPU_SAFE(pVCpu);
4410
4411	/* Preparatory work for running nested-guest code, this may force us to return to
4412	ring-3. This bugger disables interrupts on VINF_SUCCESS! */
4413	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatEntry, x);
4414	rc = hmR0SvmPreRunGuest(pVCpu, &SvmTransient);
4415	if (rc != VINF_SUCCESS)
4416	break;
4417
4418	/*
4419	* No longjmps to ring-3 from this point on!!!
4420	*
4421	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional,
4422	* better than a kernel panic. This also disables flushing of the R0-logger instance.
4423	*/
4424	hmR0SvmPreRunGuestCommitted(pVCpu, &SvmTransient);
4425	rc = hmR0SvmRunGuest(pVCpu, pVCpu->hm.s.svm.HCPhysVmcb);
4426
4427	/* Restore any residual host-state and save any bits shared between host and guest
4428	into the guest-CPU state. Re-enables interrupts! */
4429	hmR0SvmPostRunGuest(pVCpu, &SvmTransient, rc);
4430
4431	if (RT_UNLIKELY( rc != VINF_SUCCESS /* Check for VMRUN errors. */
4432	\|\| SvmTransient.u64ExitCode == SVM_EXIT_INVALID)) /* Check for invalid guest-state errors. */
4433	{
4434	if (rc == VINF_SUCCESS)
4435	rc = VERR_SVM_INVALID_GUEST_STATE;
4436	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatPreExit, x);
4437	hmR0SvmReportWorldSwitchError(pVCpu, rc);
4438	break;
4439	}
4440
4441	/* Handle the #VMEXIT. */
4442	HMSVM_EXITCODE_STAM_COUNTER_INC(SvmTransient.u64ExitCode);
4443	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatPreExit, &pVCpu->hm.s.StatExitHandling, x);
4444	VBOXVMM_R0_HMSVM_VMEXIT(pVCpu, &pVCpu->cpum.GstCtx, SvmTransient.u64ExitCode, pVCpu->hm.s.svm.pVmcb);
4445	rc = hmR0SvmHandleExit(pVCpu, &SvmTransient);
4446	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExitHandling, x);
4447	if (rc != VINF_SUCCESS)
4448	break;
4449	if (++(*pcLoops) >= cMaxResumeLoops)
4450	{
4451	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchMaxResumeLoops);
4452	rc = VINF_EM_RAW_INTERRUPT;
4453	break;
4454	}
4455	}
4456
4457	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatEntry, x);
4458	return rc;
4459	}
4460
4461
4462	/**
4463	* Runs the guest code using AMD-V in single step mode.
4464	*
4465	* @returns VBox status code.
4466	* @param pVCpu The cross context virtual CPU structure.
4467	* @param pcLoops Pointer to the number of executed loops.
4468	*/
4469	static int hmR0SvmRunGuestCodeStep(PVMCPUCC pVCpu, uint32_t *pcLoops)
4470	{
4471	uint32_t const cMaxResumeLoops = pVCpu->CTX_SUFF(pVM)->hm.s.cMaxResumeLoops;
4472	Assert(pcLoops);
4473	Assert(*pcLoops <= cMaxResumeLoops);
4474
4475	SVMTRANSIENT SvmTransient;
4476	RT_ZERO(SvmTransient);
4477	SvmTransient.fUpdateTscOffsetting = true;
4478	SvmTransient.pVmcb = pVCpu->hm.s.svm.pVmcb;
4479
4480	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
4481	uint16_t uCsStart = pCtx->cs.Sel;
4482	uint64_t uRipStart = pCtx->rip;
4483
4484	int rc = VERR_INTERNAL_ERROR_5;
4485	for (;;)
4486	{
4487	Assert(!HMR0SuspendPending());
4488	AssertMsg(pVCpu->hm.s.idEnteredCpu == RTMpCpuId(),
4489	("Illegal migration! Entered on CPU %u Current %u cLoops=%u\n", (unsigned)pVCpu->hm.s.idEnteredCpu,
4490	(unsigned)RTMpCpuId(), *pcLoops));
4491
4492	/* Preparatory work for running nested-guest code, this may force us to return to
4493	ring-3. This bugger disables interrupts on VINF_SUCCESS! */
4494	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatEntry, x);
4495	rc = hmR0SvmPreRunGuest(pVCpu, &SvmTransient);
4496	if (rc != VINF_SUCCESS)
4497	break;
4498
4499	/*
4500	* No longjmps to ring-3 from this point on!!!
4501	*
4502	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional,
4503	* better than a kernel panic. This also disables flushing of the R0-logger instance.
4504	*/
4505	hmR0SvmPreRunGuestCommitted(pVCpu, &SvmTransient);
4506
4507	rc = hmR0SvmRunGuest(pVCpu, pVCpu->hm.s.svm.HCPhysVmcb);
4508
4509	/* Restore any residual host-state and save any bits shared between host and guest
4510	into the guest-CPU state. Re-enables interrupts! */
4511	hmR0SvmPostRunGuest(pVCpu, &SvmTransient, rc);
4512
4513	if (RT_UNLIKELY( rc != VINF_SUCCESS /* Check for VMRUN errors. */
4514	\|\| SvmTransient.u64ExitCode == SVM_EXIT_INVALID)) /* Check for invalid guest-state errors. */
4515	{
4516	if (rc == VINF_SUCCESS)
4517	rc = VERR_SVM_INVALID_GUEST_STATE;
4518	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatPreExit, x);
4519	hmR0SvmReportWorldSwitchError(pVCpu, rc);
4520	return rc;
4521	}
4522
4523	/* Handle the #VMEXIT. */
4524	HMSVM_EXITCODE_STAM_COUNTER_INC(SvmTransient.u64ExitCode);
4525	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatPreExit, &pVCpu->hm.s.StatExitHandling, x);
4526	VBOXVMM_R0_HMSVM_VMEXIT(pVCpu, pCtx, SvmTransient.u64ExitCode, pVCpu->hm.s.svm.pVmcb);
4527	rc = hmR0SvmHandleExit(pVCpu, &SvmTransient);
4528	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExitHandling, x);
4529	if (rc != VINF_SUCCESS)
4530	break;
4531	if (++(*pcLoops) >= cMaxResumeLoops)
4532	{
4533	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchMaxResumeLoops);
4534	rc = VINF_EM_RAW_INTERRUPT;
4535	break;
4536	}
4537
4538	/*
4539	* Did the RIP change, if so, consider it a single step.
4540	* Otherwise, make sure one of the TFs gets set.
4541	*/
4542	if ( pCtx->rip != uRipStart
4543	\|\| pCtx->cs.Sel != uCsStart)
4544	{
4545	rc = VINF_EM_DBG_STEPPED;
4546	break;
4547	}
4548	pVCpu->hm.s.fCtxChanged \|= HM_CHANGED_GUEST_DR_MASK;
4549	}
4550
4551	/*
4552	* Clear the X86_EFL_TF if necessary.
4553	*/
4554	if (pVCpu->hm.s.fClearTrapFlag)
4555	{
4556	pVCpu->hm.s.fClearTrapFlag = false;
4557	pCtx->eflags.Bits.u1TF = 0;
4558	}
4559
4560	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatEntry, x);
4561	return rc;
4562	}
4563
4564	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
4565	/**
4566	* Runs the nested-guest code using AMD-V.
4567	*
4568	* @returns VBox status code.
4569	* @param pVCpu The cross context virtual CPU structure.
4570	* @param pcLoops Pointer to the number of executed loops. If we're switching
4571	* from the guest-code execution loop to this nested-guest
4572	* execution loop pass the remainder value, else pass 0.
4573	*/
4574	static int hmR0SvmRunGuestCodeNested(PVMCPUCC pVCpu, uint32_t *pcLoops)
4575	{
4576	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
4577	HMSVM_ASSERT_IN_NESTED_GUEST(pCtx);
4578	Assert(pcLoops);
4579	Assert(*pcLoops <= pVCpu->CTX_SUFF(pVM)->hm.s.cMaxResumeLoops);
4580
4581	SVMTRANSIENT SvmTransient;
4582	RT_ZERO(SvmTransient);
4583	SvmTransient.fUpdateTscOffsetting = true;
4584	SvmTransient.pVmcb = pCtx->hwvirt.svm.CTX_SUFF(pVmcb);
4585	SvmTransient.fIsNestedGuest = true;
4586
4587	int rc = VERR_INTERNAL_ERROR_4;
4588	for (;;)
4589	{
4590	Assert(!HMR0SuspendPending());
4591	HMSVM_ASSERT_CPU_SAFE(pVCpu);
4592
4593	/* Preparatory work for running nested-guest code, this may force us to return to
4594	ring-3. This bugger disables interrupts on VINF_SUCCESS! */
4595	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatEntry, x);
4596	rc = hmR0SvmPreRunGuest(pVCpu, &SvmTransient);
4597	if ( rc != VINF_SUCCESS
4598	\|\| !CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
4599	break;
4600
4601	/*
4602	* No longjmps to ring-3 from this point on!!!
4603	*
4604	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional,
4605	* better than a kernel panic. This also disables flushing of the R0-logger instance.
4606	*/
4607	hmR0SvmPreRunGuestCommitted(pVCpu, &SvmTransient);
4608
4609	rc = hmR0SvmRunGuest(pVCpu, pCtx->hwvirt.svm.HCPhysVmcb);
4610
4611	/* Restore any residual host-state and save any bits shared between host and guest
4612	into the guest-CPU state. Re-enables interrupts! */
4613	hmR0SvmPostRunGuest(pVCpu, &SvmTransient, rc);
4614
4615	if (RT_LIKELY( rc == VINF_SUCCESS
4616	&& SvmTransient.u64ExitCode != SVM_EXIT_INVALID))
4617	{ /* extremely likely */ }
4618	else
4619	{
4620	/* VMRUN failed, shouldn't really happen, Guru. */
4621	if (rc != VINF_SUCCESS)
4622	break;
4623
4624	/* Invalid nested-guest state. Cause a #VMEXIT but assert on strict builds. */
4625	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
4626	AssertMsgFailed(("Invalid nested-guest state. rc=%Rrc u64ExitCode=%#RX64\n", rc, SvmTransient.u64ExitCode));
4627	rc = VBOXSTRICTRC_TODO(IEMExecSvmVmexit(pVCpu, SVM_EXIT_INVALID, 0, 0));
4628	break;
4629	}
4630
4631	/* Handle the #VMEXIT. */
4632	HMSVM_NESTED_EXITCODE_STAM_COUNTER_INC(SvmTransient.u64ExitCode);
4633	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatPreExit, &pVCpu->hm.s.StatExitHandling, x);
4634	VBOXVMM_R0_HMSVM_VMEXIT(pVCpu, pCtx, SvmTransient.u64ExitCode, pCtx->hwvirt.svm.CTX_SUFF(pVmcb));
4635	rc = hmR0SvmHandleExitNested(pVCpu, &SvmTransient);
4636	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExitHandling, x);
4637	if (rc == VINF_SUCCESS)
4638	{
4639	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
4640	{
4641	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchNstGstVmexit);
4642	rc = VINF_SVM_VMEXIT;
4643	}
4644	else
4645	{
4646	if (++(*pcLoops) <= pVCpu->CTX_SUFF(pVM)->hm.s.cMaxResumeLoops)
4647	continue;
4648	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchMaxResumeLoops);
4649	rc = VINF_EM_RAW_INTERRUPT;
4650	}
4651	}
4652	else
4653	Assert(rc != VINF_SVM_VMEXIT);
4654	break;
4655	/** @todo NSTSVM: handle single-stepping. */
4656	}
4657
4658	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatEntry, x);
4659	return rc;
4660	}
4661	#endif
4662
4663
4664	/**
4665	* Runs the guest code using AMD-V.
4666	*
4667	* @returns Strict VBox status code.
4668	* @param pVCpu The cross context virtual CPU structure.
4669	*/
4670	VMMR0DECL(VBOXSTRICTRC) SVMR0RunGuestCode(PVMCPUCC pVCpu)
4671	{
4672	AssertPtr(pVCpu);
4673	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
4674	Assert(VMMRZCallRing3IsEnabled(pVCpu));
4675	Assert(!ASMAtomicUoReadU64(&pCtx->fExtrn));
4676	HMSVM_ASSERT_PREEMPT_SAFE(pVCpu);
4677
4678	uint32_t cLoops = 0;
4679	int rc;
4680	for (;;)
4681	{
4682	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
4683	bool const fInNestedGuestMode = CPUMIsGuestInSvmNestedHwVirtMode(pCtx);
4684	#else
4685	NOREF(pCtx);
4686	bool const fInNestedGuestMode = false;
4687	#endif
4688	if (!fInNestedGuestMode)
4689	{
4690	if (!pVCpu->hm.s.fSingleInstruction)
4691	rc = hmR0SvmRunGuestCodeNormal(pVCpu, &cLoops);
4692	else
4693	rc = hmR0SvmRunGuestCodeStep(pVCpu, &cLoops);
4694	}
4695	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
4696	else
4697	rc = hmR0SvmRunGuestCodeNested(pVCpu, &cLoops);
4698
4699	if (rc == VINF_SVM_VMRUN)
4700	{
4701	Assert(CPUMIsGuestInSvmNestedHwVirtMode(pCtx));
4702	continue;
4703	}
4704	if (rc == VINF_SVM_VMEXIT)
4705	{
4706	Assert(!CPUMIsGuestInSvmNestedHwVirtMode(pCtx));
4707	continue;
4708	}
4709	#endif
4710	break;
4711	}
4712
4713	/* Fixup error codes. */
4714	if (rc == VERR_EM_INTERPRETER)
4715	rc = VINF_EM_RAW_EMULATE_INSTR;
4716	else if (rc == VINF_EM_RESET)
4717	rc = VINF_EM_TRIPLE_FAULT;
4718
4719	/* Prepare to return to ring-3. This will remove longjmp notifications. */
4720	rc = hmR0SvmExitToRing3(pVCpu, rc);
4721	Assert(!ASMAtomicUoReadU64(&pCtx->fExtrn));
4722	Assert(!VMMRZCallRing3IsNotificationSet(pVCpu));
4723	return rc;
4724	}
4725
4726
4727	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
4728	/**
4729	* Determines whether the given I/O access should cause a nested-guest \#VMEXIT.
4730	*
4731	* @param pvIoBitmap Pointer to the nested-guest IO bitmap.
4732	* @param pIoExitInfo Pointer to the SVMIOIOEXITINFO.
4733	*/
4734	static bool hmR0SvmIsIoInterceptSet(void *pvIoBitmap, PSVMIOIOEXITINFO pIoExitInfo)
4735	{
4736	const uint16_t u16Port = pIoExitInfo->n.u16Port;
4737	const SVMIOIOTYPE enmIoType = (SVMIOIOTYPE)pIoExitInfo->n.u1Type;
4738	const uint8_t cbReg = (pIoExitInfo->u >> SVM_IOIO_OP_SIZE_SHIFT) & 7;
4739	const uint8_t cAddrSizeBits = ((pIoExitInfo->u >> SVM_IOIO_ADDR_SIZE_SHIFT) & 7) << 4;
4740	const uint8_t iEffSeg = pIoExitInfo->n.u3Seg;
4741	const bool fRep = pIoExitInfo->n.u1Rep;
4742	const bool fStrIo = pIoExitInfo->n.u1Str;
4743
4744	return CPUMIsSvmIoInterceptSet(pvIoBitmap, u16Port, enmIoType, cbReg, cAddrSizeBits, iEffSeg, fRep, fStrIo,
4745	NULL /* pIoExitInfo */);
4746	}
4747
4748
4749	/**
4750	* Handles a nested-guest \#VMEXIT (for all EXITCODE values except
4751	* SVM_EXIT_INVALID).
4752	*
4753	* @returns VBox status code (informational status codes included).
4754	* @param pVCpu The cross context virtual CPU structure.
4755	* @param pSvmTransient Pointer to the SVM transient structure.
4756	*/
4757	static int hmR0SvmHandleExitNested(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
4758	{
4759	HMSVM_ASSERT_IN_NESTED_GUEST(&pVCpu->cpum.GstCtx);
4760	Assert(pSvmTransient->u64ExitCode != SVM_EXIT_INVALID);
4761	Assert(pSvmTransient->u64ExitCode <= SVM_EXIT_MAX);
4762
4763	/*
4764	* We import the complete state here because we use separate VMCBs for the guest and the
4765	* nested-guest, and the guest's VMCB is used after the #VMEXIT. We can only save/restore
4766	* the #VMEXIT specific state if we used the same VMCB for both guest and nested-guest.
4767	*/
4768	#define NST_GST_VMEXIT_CALL_RET(a_pVCpu, a_uExitCode, a_uExitInfo1, a_uExitInfo2) \
4769	do { \
4770	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL); \
4771	return VBOXSTRICTRC_TODO(IEMExecSvmVmexit((a_pVCpu), (a_uExitCode), (a_uExitInfo1), (a_uExitInfo2))); \
4772	} while (0)
4773
4774	/*
4775	* For all the #VMEXITs here we primarily figure out if the #VMEXIT is expected by the
4776	* nested-guest. If it isn't, it should be handled by the (outer) guest.
4777	*/
4778	PSVMVMCB pVmcbNstGst = pVCpu->cpum.GstCtx.hwvirt.svm.CTX_SUFF(pVmcb);
4779	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
4780	PSVMVMCBCTRL pVmcbNstGstCtrl = &pVmcbNstGst->ctrl;
4781	uint64_t const uExitCode = pVmcbNstGstCtrl->u64ExitCode;
4782	uint64_t const uExitInfo1 = pVmcbNstGstCtrl->u64ExitInfo1;
4783	uint64_t const uExitInfo2 = pVmcbNstGstCtrl->u64ExitInfo2;
4784
4785	Assert(uExitCode == pVmcbNstGstCtrl->u64ExitCode);
4786	switch (uExitCode)
4787	{
4788	case SVM_EXIT_CPUID:
4789	{
4790	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_CPUID))
4791	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4792	return hmR0SvmExitCpuid(pVCpu, pSvmTransient);
4793	}
4794
4795	case SVM_EXIT_RDTSC:
4796	{
4797	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_RDTSC))
4798	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4799	return hmR0SvmExitRdtsc(pVCpu, pSvmTransient);
4800	}
4801
4802	case SVM_EXIT_RDTSCP:
4803	{
4804	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_RDTSCP))
4805	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4806	return hmR0SvmExitRdtscp(pVCpu, pSvmTransient);
4807	}
4808
4809	case SVM_EXIT_MONITOR:
4810	{
4811	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_MONITOR))
4812	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4813	return hmR0SvmExitMonitor(pVCpu, pSvmTransient);
4814	}
4815
4816	case SVM_EXIT_MWAIT:
4817	{
4818	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_MWAIT))
4819	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4820	return hmR0SvmExitMwait(pVCpu, pSvmTransient);
4821	}
4822
4823	case SVM_EXIT_HLT:
4824	{
4825	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_HLT))
4826	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4827	return hmR0SvmExitHlt(pVCpu, pSvmTransient);
4828	}
4829
4830	case SVM_EXIT_MSR:
4831	{
4832	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_MSR_PROT))
4833	{
4834	uint32_t const idMsr = pVCpu->cpum.GstCtx.ecx;
4835	uint16_t offMsrpm;
4836	uint8_t uMsrpmBit;
4837	int rc = CPUMGetSvmMsrpmOffsetAndBit(idMsr, &offMsrpm, &uMsrpmBit);
4838	if (RT_SUCCESS(rc))
4839	{
4840	Assert(uMsrpmBit == 0 \|\| uMsrpmBit == 2 \|\| uMsrpmBit == 4 \|\| uMsrpmBit == 6);
4841	Assert(offMsrpm < SVM_MSRPM_PAGES << X86_PAGE_4K_SHIFT);
4842
4843	uint8_t const pbMsrBitmap = (uint8_t const )pVCpu->cpum.GstCtx.hwvirt.svm.CTX_SUFF(pvMsrBitmap);
4844	pbMsrBitmap += offMsrpm;
4845	bool const fInterceptRead = RT_BOOL(*pbMsrBitmap & RT_BIT(uMsrpmBit));
4846	bool const fInterceptWrite = RT_BOOL(*pbMsrBitmap & RT_BIT(uMsrpmBit + 1));
4847
4848	if ( (fInterceptWrite && pVmcbNstGstCtrl->u64ExitInfo1 == SVM_EXIT1_MSR_WRITE)
4849	\|\| (fInterceptRead && pVmcbNstGstCtrl->u64ExitInfo1 == SVM_EXIT1_MSR_READ))
4850	{
4851	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4852	}
4853	}
4854	else
4855	{
4856	/*
4857	* MSRs not covered by the MSRPM automatically cause an #VMEXIT.
4858	* See AMD-V spec. "15.11 MSR Intercepts".
4859	*/
4860	Assert(rc == VERR_OUT_OF_RANGE);
4861	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4862	}
4863	}
4864	return hmR0SvmExitMsr(pVCpu, pSvmTransient);
4865	}
4866
4867	case SVM_EXIT_IOIO:
4868	{
4869	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_IOIO_PROT))
4870	{
4871	void *pvIoBitmap = pVCpu->cpum.GstCtx.hwvirt.svm.CTX_SUFF(pvIoBitmap);
4872	SVMIOIOEXITINFO IoExitInfo;
4873	IoExitInfo.u = pVmcbNstGst->ctrl.u64ExitInfo1;
4874	bool const fIntercept = hmR0SvmIsIoInterceptSet(pvIoBitmap, &IoExitInfo);
4875	if (fIntercept)
4876	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4877	}
4878	return hmR0SvmExitIOInstr(pVCpu, pSvmTransient);
4879	}
4880
4881	case SVM_EXIT_XCPT_PF:
4882	{
4883	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
4884	if (pVM->hm.s.fNestedPaging)
4885	{
4886	uint32_t const u32ErrCode = pVmcbNstGstCtrl->u64ExitInfo1;
4887	uint64_t const uFaultAddress = pVmcbNstGstCtrl->u64ExitInfo2;
4888
4889	/* If the nested-guest is intercepting #PFs, cause a #PF #VMEXIT. */
4890	if (CPUMIsGuestSvmXcptInterceptSet(pVCpu, pCtx, X86_XCPT_PF))
4891	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, u32ErrCode, uFaultAddress);
4892
4893	/* If the nested-guest is not intercepting #PFs, forward the #PF to the guest. */
4894	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR2);
4895	hmR0SvmSetPendingXcptPF(pVCpu, u32ErrCode, uFaultAddress);
4896	return VINF_SUCCESS;
4897	}
4898	return hmR0SvmExitXcptPF(pVCpu, pSvmTransient);
4899	}
4900
4901	case SVM_EXIT_XCPT_UD:
4902	{
4903	if (CPUMIsGuestSvmXcptInterceptSet(pVCpu, pCtx, X86_XCPT_UD))
4904	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4905	hmR0SvmSetPendingXcptUD(pVCpu);
4906	return VINF_SUCCESS;
4907	}
4908
4909	case SVM_EXIT_XCPT_MF:
4910	{
4911	if (CPUMIsGuestSvmXcptInterceptSet(pVCpu, pCtx, X86_XCPT_MF))
4912	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4913	return hmR0SvmExitXcptMF(pVCpu, pSvmTransient);
4914	}
4915
4916	case SVM_EXIT_XCPT_DB:
4917	{
4918	if (CPUMIsGuestSvmXcptInterceptSet(pVCpu, pCtx, X86_XCPT_DB))
4919	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4920	return hmR0SvmNestedExitXcptDB(pVCpu, pSvmTransient);
4921	}
4922
4923	case SVM_EXIT_XCPT_AC:
4924	{
4925	if (CPUMIsGuestSvmXcptInterceptSet(pVCpu, pCtx, X86_XCPT_AC))
4926	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4927	return hmR0SvmExitXcptAC(pVCpu, pSvmTransient);
4928	}
4929
4930	case SVM_EXIT_XCPT_BP:
4931	{
4932	if (CPUMIsGuestSvmXcptInterceptSet(pVCpu, pCtx, X86_XCPT_BP))
4933	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4934	return hmR0SvmNestedExitXcptBP(pVCpu, pSvmTransient);
4935	}
4936
4937	case SVM_EXIT_READ_CR0:
4938	case SVM_EXIT_READ_CR3:
4939	case SVM_EXIT_READ_CR4:
4940	{
4941	uint8_t const uCr = uExitCode - SVM_EXIT_READ_CR0;
4942	if (CPUMIsGuestSvmReadCRxInterceptSet(pVCpu, pCtx, uCr))
4943	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4944	return hmR0SvmExitReadCRx(pVCpu, pSvmTransient);
4945	}
4946
4947	case SVM_EXIT_CR0_SEL_WRITE:
4948	{
4949	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_CR0_SEL_WRITE))
4950	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4951	return hmR0SvmExitWriteCRx(pVCpu, pSvmTransient);
4952	}
4953
4954	case SVM_EXIT_WRITE_CR0:
4955	case SVM_EXIT_WRITE_CR3:
4956	case SVM_EXIT_WRITE_CR4:
4957	case SVM_EXIT_WRITE_CR8: /* CR8 writes would go to the V_TPR rather than here, since we run with V_INTR_MASKING. */
4958	{
4959	uint8_t const uCr = uExitCode - SVM_EXIT_WRITE_CR0;
4960	Log4Func(("Write CR%u: uExitInfo1=%#RX64 uExitInfo2=%#RX64\n", uCr, uExitInfo1, uExitInfo2));
4961
4962	if (CPUMIsGuestSvmWriteCRxInterceptSet(pVCpu, pCtx, uCr))
4963	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4964	return hmR0SvmExitWriteCRx(pVCpu, pSvmTransient);
4965	}
4966
4967	case SVM_EXIT_PAUSE:
4968	{
4969	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_PAUSE))
4970	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4971	return hmR0SvmExitPause(pVCpu, pSvmTransient);
4972	}
4973
4974	case SVM_EXIT_VINTR:
4975	{
4976	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_VINTR))
4977	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4978	return hmR0SvmExitUnexpected(pVCpu, pSvmTransient);
4979	}
4980
4981	case SVM_EXIT_INTR:
4982	case SVM_EXIT_NMI:
4983	case SVM_EXIT_SMI:
4984	case SVM_EXIT_XCPT_NMI: /* Should not occur, SVM_EXIT_NMI is used instead. */
4985	{
4986	/*
4987	* We shouldn't direct physical interrupts, NMIs, SMIs to the nested-guest.
4988	*
4989	* Although we don't intercept SMIs, the nested-guest might. Therefore, we might
4990	* get an SMI #VMEXIT here so simply ignore rather than causing a corresponding
4991	* nested-guest #VMEXIT.
4992	*
4993	* We shall import the complete state here as we may cause #VMEXITs from ring-3
4994	* while trying to inject interrupts, see comment at the top of this function.
4995	*/
4996	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_ALL);
4997	return hmR0SvmExitIntr(pVCpu, pSvmTransient);
4998	}
4999
5000	case SVM_EXIT_FERR_FREEZE:
5001	{
5002	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_FERR_FREEZE))
5003	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5004	return hmR0SvmExitFerrFreeze(pVCpu, pSvmTransient);
5005	}
5006
5007	case SVM_EXIT_INVLPG:
5008	{
5009	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_INVLPG))
5010	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5011	return hmR0SvmExitInvlpg(pVCpu, pSvmTransient);
5012	}
5013
5014	case SVM_EXIT_WBINVD:
5015	{
5016	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_WBINVD))
5017	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5018	return hmR0SvmExitWbinvd(pVCpu, pSvmTransient);
5019	}
5020
5021	case SVM_EXIT_INVD:
5022	{
5023	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_INVD))
5024	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5025	return hmR0SvmExitInvd(pVCpu, pSvmTransient);
5026	}
5027
5028	case SVM_EXIT_RDPMC:
5029	{
5030	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_RDPMC))
5031	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5032	return hmR0SvmExitRdpmc(pVCpu, pSvmTransient);
5033	}
5034
5035	default:
5036	{
5037	switch (uExitCode)
5038	{
5039	case SVM_EXIT_READ_DR0: case SVM_EXIT_READ_DR1: case SVM_EXIT_READ_DR2: case SVM_EXIT_READ_DR3:
5040	case SVM_EXIT_READ_DR6: case SVM_EXIT_READ_DR7: case SVM_EXIT_READ_DR8: case SVM_EXIT_READ_DR9:
5041	case SVM_EXIT_READ_DR10: case SVM_EXIT_READ_DR11: case SVM_EXIT_READ_DR12: case SVM_EXIT_READ_DR13:
5042	case SVM_EXIT_READ_DR14: case SVM_EXIT_READ_DR15:
5043	{
5044	uint8_t const uDr = uExitCode - SVM_EXIT_READ_DR0;
5045	if (CPUMIsGuestSvmReadDRxInterceptSet(pVCpu, pCtx, uDr))
5046	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5047	return hmR0SvmExitReadDRx(pVCpu, pSvmTransient);
5048	}
5049
5050	case SVM_EXIT_WRITE_DR0: case SVM_EXIT_WRITE_DR1: case SVM_EXIT_WRITE_DR2: case SVM_EXIT_WRITE_DR3:
5051	case SVM_EXIT_WRITE_DR6: case SVM_EXIT_WRITE_DR7: case SVM_EXIT_WRITE_DR8: case SVM_EXIT_WRITE_DR9:
5052	case SVM_EXIT_WRITE_DR10: case SVM_EXIT_WRITE_DR11: case SVM_EXIT_WRITE_DR12: case SVM_EXIT_WRITE_DR13:
5053	case SVM_EXIT_WRITE_DR14: case SVM_EXIT_WRITE_DR15:
5054	{
5055	uint8_t const uDr = uExitCode - SVM_EXIT_WRITE_DR0;
5056	if (CPUMIsGuestSvmWriteDRxInterceptSet(pVCpu, pCtx, uDr))
5057	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5058	return hmR0SvmExitWriteDRx(pVCpu, pSvmTransient);
5059	}
5060
5061	case SVM_EXIT_XCPT_DE:
5062	/* SVM_EXIT_XCPT_DB: / / Handled above. */
5063	/* SVM_EXIT_XCPT_NMI: / / Handled above. */
5064	/* SVM_EXIT_XCPT_BP: / / Handled above. */
5065	case SVM_EXIT_XCPT_OF:
5066	case SVM_EXIT_XCPT_BR:
5067	/* SVM_EXIT_XCPT_UD: / / Handled above. */
5068	case SVM_EXIT_XCPT_NM:
5069	case SVM_EXIT_XCPT_DF:
5070	case SVM_EXIT_XCPT_CO_SEG_OVERRUN:
5071	case SVM_EXIT_XCPT_TS:
5072	case SVM_EXIT_XCPT_NP:
5073	case SVM_EXIT_XCPT_SS:
5074	case SVM_EXIT_XCPT_GP:
5075	/* SVM_EXIT_XCPT_PF: / / Handled above. */
5076	case SVM_EXIT_XCPT_15: /* Reserved. */
5077	/* SVM_EXIT_XCPT_MF: / / Handled above. */
5078	/* SVM_EXIT_XCPT_AC: / / Handled above. */
5079	case SVM_EXIT_XCPT_MC:
5080	case SVM_EXIT_XCPT_XF:
5081	case SVM_EXIT_XCPT_20: case SVM_EXIT_XCPT_21: case SVM_EXIT_XCPT_22: case SVM_EXIT_XCPT_23:
5082	case SVM_EXIT_XCPT_24: case SVM_EXIT_XCPT_25: case SVM_EXIT_XCPT_26: case SVM_EXIT_XCPT_27:
5083	case SVM_EXIT_XCPT_28: case SVM_EXIT_XCPT_29: case SVM_EXIT_XCPT_30: case SVM_EXIT_XCPT_31:
5084	{
5085	uint8_t const uVector = uExitCode - SVM_EXIT_XCPT_0;
5086	if (CPUMIsGuestSvmXcptInterceptSet(pVCpu, pCtx, uVector))
5087	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5088	return hmR0SvmExitXcptGeneric(pVCpu, pSvmTransient);
5089	}
5090
5091	case SVM_EXIT_XSETBV:
5092	{
5093	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_XSETBV))
5094	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5095	return hmR0SvmExitXsetbv(pVCpu, pSvmTransient);
5096	}
5097
5098	case SVM_EXIT_TASK_SWITCH:
5099	{
5100	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_TASK_SWITCH))
5101	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5102	return hmR0SvmExitTaskSwitch(pVCpu, pSvmTransient);
5103	}
5104
5105	case SVM_EXIT_IRET:
5106	{
5107	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_IRET))
5108	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5109	return hmR0SvmExitIret(pVCpu, pSvmTransient);
5110	}
5111
5112	case SVM_EXIT_SHUTDOWN:
5113	{
5114	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_SHUTDOWN))
5115	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5116	return hmR0SvmExitShutdown(pVCpu, pSvmTransient);
5117	}
5118
5119	case SVM_EXIT_VMMCALL:
5120	{
5121	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_VMMCALL))
5122	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5123	return hmR0SvmExitVmmCall(pVCpu, pSvmTransient);
5124	}
5125
5126	case SVM_EXIT_CLGI:
5127	{
5128	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_CLGI))
5129	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5130	return hmR0SvmExitClgi(pVCpu, pSvmTransient);
5131	}
5132
5133	case SVM_EXIT_STGI:
5134	{
5135	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_STGI))
5136	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5137	return hmR0SvmExitStgi(pVCpu, pSvmTransient);
5138	}
5139
5140	case SVM_EXIT_VMLOAD:
5141	{
5142	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_VMLOAD))
5143	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5144	return hmR0SvmExitVmload(pVCpu, pSvmTransient);
5145	}
5146
5147	case SVM_EXIT_VMSAVE:
5148	{
5149	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_VMSAVE))
5150	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5151	return hmR0SvmExitVmsave(pVCpu, pSvmTransient);
5152	}
5153
5154	case SVM_EXIT_INVLPGA:
5155	{
5156	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_INVLPGA))
5157	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5158	return hmR0SvmExitInvlpga(pVCpu, pSvmTransient);
5159	}
5160
5161	case SVM_EXIT_VMRUN:
5162	{
5163	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_VMRUN))
5164	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5165	return hmR0SvmExitVmrun(pVCpu, pSvmTransient);
5166	}
5167
5168	case SVM_EXIT_RSM:
5169	{
5170	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_RSM))
5171	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5172	hmR0SvmSetPendingXcptUD(pVCpu);
5173	return VINF_SUCCESS;
5174	}
5175
5176	case SVM_EXIT_SKINIT:
5177	{
5178	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_SKINIT))
5179	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5180	hmR0SvmSetPendingXcptUD(pVCpu);
5181	return VINF_SUCCESS;
5182	}
5183
5184	case SVM_EXIT_NPF:
5185	{
5186	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
5187	return hmR0SvmExitNestedPF(pVCpu, pSvmTransient);
5188	}
5189
5190	case SVM_EXIT_INIT: /* We shouldn't get INIT signals while executing a nested-guest. */
5191	return hmR0SvmExitUnexpected(pVCpu, pSvmTransient);
5192
5193	default:
5194	{
5195	AssertMsgFailed(("hmR0SvmHandleExitNested: Unknown exit code %#x\n", pSvmTransient->u64ExitCode));
5196	pVCpu->hm.s.u32HMError = pSvmTransient->u64ExitCode;
5197	return VERR_SVM_UNKNOWN_EXIT;
5198	}
5199	}
5200	}
5201	}
5202	/* not reached */
5203
5204	#undef NST_GST_VMEXIT_CALL_RET
5205	}
5206	#endif
5207
5208
5209	/**
5210	* Handles a guest \#VMEXIT (for all EXITCODE values except SVM_EXIT_INVALID).
5211	*
5212	* @returns VBox status code (informational status codes included).
5213	* @param pVCpu The cross context virtual CPU structure.
5214	* @param pSvmTransient Pointer to the SVM transient structure.
5215	*/
5216	static int hmR0SvmHandleExit(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
5217	{
5218	Assert(pSvmTransient->u64ExitCode != SVM_EXIT_INVALID);
5219	Assert(pSvmTransient->u64ExitCode <= SVM_EXIT_MAX);
5220
5221	#ifdef DEBUG_ramshankar
5222	# define VMEXIT_CALL_RET(a_fDbg, a_CallExpr) \
5223	do { \
5224	if ((a_fDbg) == 1) \
5225	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL); \
5226	int rc = a_CallExpr; \
5227	if ((a_fDbg) == 1) \
5228	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST); \
5229	return rc; \
5230	} while (0)
5231	#else
5232	# define VMEXIT_CALL_RET(a_fDbg, a_CallExpr) return a_CallExpr
5233	#endif
5234
5235	/*
5236	* The ordering of the case labels is based on most-frequently-occurring #VMEXITs
5237	* for most guests under normal workloads (for some definition of "normal").
5238	*/
5239	uint64_t const uExitCode = pSvmTransient->u64ExitCode;
5240	switch (uExitCode)
5241	{
5242	case SVM_EXIT_NPF: VMEXIT_CALL_RET(0, hmR0SvmExitNestedPF(pVCpu, pSvmTransient));
5243	case SVM_EXIT_IOIO: VMEXIT_CALL_RET(0, hmR0SvmExitIOInstr(pVCpu, pSvmTransient));
5244	case SVM_EXIT_RDTSC: VMEXIT_CALL_RET(0, hmR0SvmExitRdtsc(pVCpu, pSvmTransient));
5245	case SVM_EXIT_RDTSCP: VMEXIT_CALL_RET(0, hmR0SvmExitRdtscp(pVCpu, pSvmTransient));
5246	case SVM_EXIT_CPUID: VMEXIT_CALL_RET(0, hmR0SvmExitCpuid(pVCpu, pSvmTransient));
5247	case SVM_EXIT_XCPT_PF: VMEXIT_CALL_RET(0, hmR0SvmExitXcptPF(pVCpu, pSvmTransient));
5248	case SVM_EXIT_MSR: VMEXIT_CALL_RET(0, hmR0SvmExitMsr(pVCpu, pSvmTransient));
5249	case SVM_EXIT_MONITOR: VMEXIT_CALL_RET(0, hmR0SvmExitMonitor(pVCpu, pSvmTransient));
5250	case SVM_EXIT_MWAIT: VMEXIT_CALL_RET(0, hmR0SvmExitMwait(pVCpu, pSvmTransient));
5251	case SVM_EXIT_HLT: VMEXIT_CALL_RET(0, hmR0SvmExitHlt(pVCpu, pSvmTransient));
5252
5253	case SVM_EXIT_XCPT_NMI: /* Should not occur, SVM_EXIT_NMI is used instead. */
5254	case SVM_EXIT_INTR:
5255	case SVM_EXIT_NMI: VMEXIT_CALL_RET(0, hmR0SvmExitIntr(pVCpu, pSvmTransient));
5256
5257	case SVM_EXIT_READ_CR0:
5258	case SVM_EXIT_READ_CR3:
5259	case SVM_EXIT_READ_CR4: VMEXIT_CALL_RET(0, hmR0SvmExitReadCRx(pVCpu, pSvmTransient));
5260
5261	case SVM_EXIT_CR0_SEL_WRITE:
5262	case SVM_EXIT_WRITE_CR0:
5263	case SVM_EXIT_WRITE_CR3:
5264	case SVM_EXIT_WRITE_CR4:
5265	case SVM_EXIT_WRITE_CR8: VMEXIT_CALL_RET(0, hmR0SvmExitWriteCRx(pVCpu, pSvmTransient));
5266
5267	case SVM_EXIT_VINTR: VMEXIT_CALL_RET(0, hmR0SvmExitVIntr(pVCpu, pSvmTransient));
5268	case SVM_EXIT_PAUSE: VMEXIT_CALL_RET(0, hmR0SvmExitPause(pVCpu, pSvmTransient));
5269	case SVM_EXIT_VMMCALL: VMEXIT_CALL_RET(0, hmR0SvmExitVmmCall(pVCpu, pSvmTransient));
5270	case SVM_EXIT_INVLPG: VMEXIT_CALL_RET(0, hmR0SvmExitInvlpg(pVCpu, pSvmTransient));
5271	case SVM_EXIT_WBINVD: VMEXIT_CALL_RET(0, hmR0SvmExitWbinvd(pVCpu, pSvmTransient));
5272	case SVM_EXIT_INVD: VMEXIT_CALL_RET(0, hmR0SvmExitInvd(pVCpu, pSvmTransient));
5273	case SVM_EXIT_RDPMC: VMEXIT_CALL_RET(0, hmR0SvmExitRdpmc(pVCpu, pSvmTransient));
5274	case SVM_EXIT_IRET: VMEXIT_CALL_RET(0, hmR0SvmExitIret(pVCpu, pSvmTransient));
5275	case SVM_EXIT_XCPT_UD: VMEXIT_CALL_RET(0, hmR0SvmExitXcptUD(pVCpu, pSvmTransient));
5276	case SVM_EXIT_XCPT_MF: VMEXIT_CALL_RET(0, hmR0SvmExitXcptMF(pVCpu, pSvmTransient));
5277	case SVM_EXIT_XCPT_DB: VMEXIT_CALL_RET(0, hmR0SvmExitXcptDB(pVCpu, pSvmTransient));
5278	case SVM_EXIT_XCPT_AC: VMEXIT_CALL_RET(0, hmR0SvmExitXcptAC(pVCpu, pSvmTransient));
5279	case SVM_EXIT_XCPT_BP: VMEXIT_CALL_RET(0, hmR0SvmExitXcptBP(pVCpu, pSvmTransient));
5280	case SVM_EXIT_XCPT_GP: VMEXIT_CALL_RET(0, hmR0SvmExitXcptGP(pVCpu, pSvmTransient));
5281	case SVM_EXIT_XSETBV: VMEXIT_CALL_RET(0, hmR0SvmExitXsetbv(pVCpu, pSvmTransient));
5282	case SVM_EXIT_FERR_FREEZE: VMEXIT_CALL_RET(0, hmR0SvmExitFerrFreeze(pVCpu, pSvmTransient));
5283
5284	default:
5285	{
5286	switch (pSvmTransient->u64ExitCode)
5287	{
5288	case SVM_EXIT_READ_DR0: case SVM_EXIT_READ_DR1: case SVM_EXIT_READ_DR2: case SVM_EXIT_READ_DR3:
5289	case SVM_EXIT_READ_DR6: case SVM_EXIT_READ_DR7: case SVM_EXIT_READ_DR8: case SVM_EXIT_READ_DR9:
5290	case SVM_EXIT_READ_DR10: case SVM_EXIT_READ_DR11: case SVM_EXIT_READ_DR12: case SVM_EXIT_READ_DR13:
5291	case SVM_EXIT_READ_DR14: case SVM_EXIT_READ_DR15:
5292	VMEXIT_CALL_RET(0, hmR0SvmExitReadDRx(pVCpu, pSvmTransient));
5293
5294	case SVM_EXIT_WRITE_DR0: case SVM_EXIT_WRITE_DR1: case SVM_EXIT_WRITE_DR2: case SVM_EXIT_WRITE_DR3:
5295	case SVM_EXIT_WRITE_DR6: case SVM_EXIT_WRITE_DR7: case SVM_EXIT_WRITE_DR8: case SVM_EXIT_WRITE_DR9:
5296	case SVM_EXIT_WRITE_DR10: case SVM_EXIT_WRITE_DR11: case SVM_EXIT_WRITE_DR12: case SVM_EXIT_WRITE_DR13:
5297	case SVM_EXIT_WRITE_DR14: case SVM_EXIT_WRITE_DR15:
5298	VMEXIT_CALL_RET(0, hmR0SvmExitWriteDRx(pVCpu, pSvmTransient));
5299
5300	case SVM_EXIT_TASK_SWITCH: VMEXIT_CALL_RET(0, hmR0SvmExitTaskSwitch(pVCpu, pSvmTransient));
5301	case SVM_EXIT_SHUTDOWN: VMEXIT_CALL_RET(0, hmR0SvmExitShutdown(pVCpu, pSvmTransient));
5302
5303	case SVM_EXIT_SMI:
5304	case SVM_EXIT_INIT:
5305	{
5306	/*
5307	* We don't intercept SMIs. As for INIT signals, it really shouldn't ever happen here.
5308	* If it ever does, we want to know about it so log the exit code and bail.
5309	*/
5310	VMEXIT_CALL_RET(0, hmR0SvmExitUnexpected(pVCpu, pSvmTransient));
5311	}
5312
5313	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
5314	case SVM_EXIT_CLGI: VMEXIT_CALL_RET(0, hmR0SvmExitClgi(pVCpu, pSvmTransient));
5315	case SVM_EXIT_STGI: VMEXIT_CALL_RET(0, hmR0SvmExitStgi(pVCpu, pSvmTransient));
5316	case SVM_EXIT_VMLOAD: VMEXIT_CALL_RET(0, hmR0SvmExitVmload(pVCpu, pSvmTransient));
5317	case SVM_EXIT_VMSAVE: VMEXIT_CALL_RET(0, hmR0SvmExitVmsave(pVCpu, pSvmTransient));
5318	case SVM_EXIT_INVLPGA: VMEXIT_CALL_RET(0, hmR0SvmExitInvlpga(pVCpu, pSvmTransient));
5319	case SVM_EXIT_VMRUN: VMEXIT_CALL_RET(0, hmR0SvmExitVmrun(pVCpu, pSvmTransient));
5320	#else
5321	case SVM_EXIT_CLGI:
5322	case SVM_EXIT_STGI:
5323	case SVM_EXIT_VMLOAD:
5324	case SVM_EXIT_VMSAVE:
5325	case SVM_EXIT_INVLPGA:
5326	case SVM_EXIT_VMRUN:
5327	#endif
5328	case SVM_EXIT_RSM:
5329	case SVM_EXIT_SKINIT:
5330	{
5331	hmR0SvmSetPendingXcptUD(pVCpu);
5332	return VINF_SUCCESS;
5333	}
5334
5335	#ifdef HMSVM_ALWAYS_TRAP_ALL_XCPTS
5336	case SVM_EXIT_XCPT_DE:
5337	/* SVM_EXIT_XCPT_DB: / / Handled above. */
5338	/* SVM_EXIT_XCPT_NMI: / / Handled above. */
5339	/* SVM_EXIT_XCPT_BP: / / Handled above. */
5340	case SVM_EXIT_XCPT_OF:
5341	case SVM_EXIT_XCPT_BR:
5342	/* SVM_EXIT_XCPT_UD: / / Handled above. */
5343	case SVM_EXIT_XCPT_NM:
5344	case SVM_EXIT_XCPT_DF:
5345	case SVM_EXIT_XCPT_CO_SEG_OVERRUN:
5346	case SVM_EXIT_XCPT_TS:
5347	case SVM_EXIT_XCPT_NP:
5348	case SVM_EXIT_XCPT_SS:
5349	/* SVM_EXIT_XCPT_GP: / / Handled above. */
5350	/* SVM_EXIT_XCPT_PF: */
5351	case SVM_EXIT_XCPT_15: /* Reserved. */
5352	/* SVM_EXIT_XCPT_MF: / / Handled above. */
5353	/* SVM_EXIT_XCPT_AC: / / Handled above. */
5354	case SVM_EXIT_XCPT_MC:
5355	case SVM_EXIT_XCPT_XF:
5356	case SVM_EXIT_XCPT_20: case SVM_EXIT_XCPT_21: case SVM_EXIT_XCPT_22: case SVM_EXIT_XCPT_23:
5357	case SVM_EXIT_XCPT_24: case SVM_EXIT_XCPT_25: case SVM_EXIT_XCPT_26: case SVM_EXIT_XCPT_27:
5358	case SVM_EXIT_XCPT_28: case SVM_EXIT_XCPT_29: case SVM_EXIT_XCPT_30: case SVM_EXIT_XCPT_31:
5359	VMEXIT_CALL_RET(0, hmR0SvmExitXcptGeneric(pVCpu, pSvmTransient));
5360	#endif /* HMSVM_ALWAYS_TRAP_ALL_XCPTS */
5361
5362	default:
5363	{
5364	AssertMsgFailed(("hmR0SvmHandleExit: Unknown exit code %#RX64\n", uExitCode));
5365	pVCpu->hm.s.u32HMError = uExitCode;
5366	return VERR_SVM_UNKNOWN_EXIT;
5367	}
5368	}
5369	}
5370	}
5371	/* not reached */
5372	#undef VMEXIT_CALL_RET
5373	}
5374
5375
5376	#ifdef VBOX_STRICT
5377	/* Is there some generic IPRT define for this that are not in Runtime/internal/\* ?? */
5378	# define HMSVM_ASSERT_PREEMPT_CPUID_VAR() \
5379	RTCPUID const idAssertCpu = RTThreadPreemptIsEnabled(NIL_RTTHREAD) ? NIL_RTCPUID : RTMpCpuId()
5380
5381	# define HMSVM_ASSERT_PREEMPT_CPUID() \
5382	do \
5383	{ \
5384	RTCPUID const idAssertCpuNow = RTThreadPreemptIsEnabled(NIL_RTTHREAD) ? NIL_RTCPUID : RTMpCpuId(); \
5385	AssertMsg(idAssertCpu == idAssertCpuNow, ("SVM %#x, %#x\n", idAssertCpu, idAssertCpuNow)); \
5386	} while (0)
5387
5388	# define HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(a_pVCpu, a_pSvmTransient) \
5389	do { \
5390	AssertPtr((a_pVCpu)); \
5391	AssertPtr((a_pSvmTransient)); \
5392	Assert(ASMIntAreEnabled()); \
5393	HMSVM_ASSERT_PREEMPT_SAFE((a_pVCpu)); \
5394	HMSVM_ASSERT_PREEMPT_CPUID_VAR(); \
5395	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", (a_pVCpu)->idCpu)); \
5396	HMSVM_ASSERT_PREEMPT_SAFE((a_pVCpu)); \
5397	if (VMMR0IsLogFlushDisabled((a_pVCpu))) \
5398	HMSVM_ASSERT_PREEMPT_CPUID(); \
5399	} while (0)
5400	#else
5401	# define HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(a_pVCpu, a_pSvmTransient) \
5402	do { \
5403	RT_NOREF2(a_pVCpu, a_pSvmTransient); \
5404	} while (0)
5405	#endif
5406
5407
5408	/**
5409	* Gets the IEM exception flags for the specified SVM event.
5410	*
5411	* @returns The IEM exception flags.
5412	* @param pEvent Pointer to the SVM event.
5413	*
5414	* @remarks This function currently only constructs flags required for
5415	* IEMEvaluateRecursiveXcpt and not the complete flags (e.g. error-code
5416	* and CR2 aspects of an exception are not included).
5417	*/
5418	static uint32_t hmR0SvmGetIemXcptFlags(PCSVMEVENT pEvent)
5419	{
5420	uint8_t const uEventType = pEvent->n.u3Type;
5421	uint32_t fIemXcptFlags;
5422	switch (uEventType)
5423	{
5424	case SVM_EVENT_EXCEPTION:
5425	/*
5426	* Only INT3 and INTO instructions can raise #BP and #OF exceptions.
5427	* See AMD spec. Table 8-1. "Interrupt Vector Source and Cause".
5428	*/
5429	if (pEvent->n.u8Vector == X86_XCPT_BP)
5430	{
5431	fIemXcptFlags = IEM_XCPT_FLAGS_T_SOFT_INT \| IEM_XCPT_FLAGS_BP_INSTR;
5432	break;
5433	}
5434	if (pEvent->n.u8Vector == X86_XCPT_OF)
5435	{
5436	fIemXcptFlags = IEM_XCPT_FLAGS_T_SOFT_INT \| IEM_XCPT_FLAGS_OF_INSTR;
5437	break;
5438	}
5439	/** @todo How do we distinguish ICEBP \#DB from the regular one? */
5440	RT_FALL_THRU();
5441	case SVM_EVENT_NMI:
5442	fIemXcptFlags = IEM_XCPT_FLAGS_T_CPU_XCPT;
5443	break;
5444
5445	case SVM_EVENT_EXTERNAL_IRQ:
5446	fIemXcptFlags = IEM_XCPT_FLAGS_T_EXT_INT;
5447	break;
5448
5449	case SVM_EVENT_SOFTWARE_INT:
5450	fIemXcptFlags = IEM_XCPT_FLAGS_T_SOFT_INT;
5451	break;
5452
5453	default:
5454	fIemXcptFlags = 0;
5455	AssertMsgFailed(("Unexpected event type! uEventType=%#x uVector=%#x", uEventType, pEvent->n.u8Vector));
5456	break;
5457	}
5458	return fIemXcptFlags;
5459	}
5460
5461
5462	/**
5463	* Handle a condition that occurred while delivering an event through the guest
5464	* IDT.
5465	*
5466	* @returns VBox status code (informational error codes included).
5467	* @retval VINF_SUCCESS if we should continue handling the \#VMEXIT.
5468	* @retval VINF_HM_DOUBLE_FAULT if a \#DF condition was detected and we ought to
5469	* continue execution of the guest which will delivery the \#DF.
5470	* @retval VINF_EM_RESET if we detected a triple-fault condition.
5471	* @retval VERR_EM_GUEST_CPU_HANG if we detected a guest CPU hang.
5472	*
5473	* @param pVCpu The cross context virtual CPU structure.
5474	* @param pSvmTransient Pointer to the SVM transient structure.
5475	*
5476	* @remarks No-long-jump zone!!!
5477	*/
5478	static int hmR0SvmCheckExitDueToEventDelivery(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
5479	{
5480	int rc = VINF_SUCCESS;
5481	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
5482	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR2);
5483
5484	Log4(("EXITINTINFO: Pending vectoring event %#RX64 Valid=%RTbool ErrValid=%RTbool Err=%#RX32 Type=%u Vector=%u\n",
5485	pVmcb->ctrl.ExitIntInfo.u, !!pVmcb->ctrl.ExitIntInfo.n.u1Valid, !!pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid,
5486	pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode, pVmcb->ctrl.ExitIntInfo.n.u3Type, pVmcb->ctrl.ExitIntInfo.n.u8Vector));
5487
5488	/*
5489	* The EXITINTINFO (if valid) contains the prior exception (IDT vector) that was trying to
5490	* be delivered to the guest which caused a #VMEXIT which was intercepted (Exit vector).
5491	*
5492	* See AMD spec. 15.7.3 "EXITINFO Pseudo-Code".
5493	*/
5494	if (pVmcb->ctrl.ExitIntInfo.n.u1Valid)
5495	{
5496	IEMXCPTRAISE enmRaise;
5497	IEMXCPTRAISEINFO fRaiseInfo;
5498	bool const fExitIsHwXcpt = pSvmTransient->u64ExitCode - SVM_EXIT_XCPT_0 <= SVM_EXIT_XCPT_31;
5499	uint8_t const uIdtVector = pVmcb->ctrl.ExitIntInfo.n.u8Vector;
5500	if (fExitIsHwXcpt)
5501	{
5502	uint8_t const uExitVector = pSvmTransient->u64ExitCode - SVM_EXIT_XCPT_0;
5503	uint32_t const fIdtVectorFlags = hmR0SvmGetIemXcptFlags(&pVmcb->ctrl.ExitIntInfo);
5504	uint32_t const fExitVectorFlags = IEM_XCPT_FLAGS_T_CPU_XCPT;
5505	enmRaise = IEMEvaluateRecursiveXcpt(pVCpu, fIdtVectorFlags, uIdtVector, fExitVectorFlags, uExitVector, &fRaiseInfo);
5506	}
5507	else
5508	{
5509	/*
5510	* If delivery of an event caused a #VMEXIT that is not an exception (e.g. #NPF)
5511	* then we end up here.
5512	*
5513	* If the event was:
5514	* - a software interrupt, we can re-execute the instruction which will
5515	* regenerate the event.
5516	* - an NMI, we need to clear NMI blocking and re-inject the NMI.
5517	* - a hardware exception or external interrupt, we re-inject it.
5518	*/
5519	fRaiseInfo = IEMXCPTRAISEINFO_NONE;
5520	if (pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_SOFTWARE_INT)
5521	enmRaise = IEMXCPTRAISE_REEXEC_INSTR;
5522	else
5523	enmRaise = IEMXCPTRAISE_PREV_EVENT;
5524	}
5525
5526	switch (enmRaise)
5527	{
5528	case IEMXCPTRAISE_CURRENT_XCPT:
5529	case IEMXCPTRAISE_PREV_EVENT:
5530	{
5531	/* For software interrupts, we shall re-execute the instruction. */
5532	if (!(fRaiseInfo & IEMXCPTRAISEINFO_SOFT_INT_XCPT))
5533	{
5534	RTGCUINTPTR GCPtrFaultAddress = 0;
5535
5536	/* If we are re-injecting an NMI, clear NMI blocking. */
5537	if (pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_NMI)
5538	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_BLOCK_NMIS);
5539
5540	/* Determine a vectoring #PF condition, see comment in hmR0SvmExitXcptPF(). */
5541	if (fRaiseInfo & (IEMXCPTRAISEINFO_EXT_INT_PF \| IEMXCPTRAISEINFO_NMI_PF))
5542	{
5543	pSvmTransient->fVectoringPF = true;
5544	Log4Func(("IDT: Pending vectoring #PF due to delivery of Ext-Int/NMI. uCR2=%#RX64\n",
5545	pVCpu->cpum.GstCtx.cr2));
5546	}
5547	else if ( pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_EXCEPTION
5548	&& uIdtVector == X86_XCPT_PF)
5549	{
5550	/*
5551	* If the previous exception was a #PF, we need to recover the CR2 value.
5552	* This can't happen with shadow paging.
5553	*/
5554	GCPtrFaultAddress = pVCpu->cpum.GstCtx.cr2;
5555	}
5556
5557	/*
5558	* Without nested paging, when uExitVector is #PF, CR2 value will be updated from the VMCB's
5559	* exit info. fields, if it's a guest #PF, see hmR0SvmExitXcptPF().
5560	*/
5561	Assert(pVmcb->ctrl.ExitIntInfo.n.u3Type != SVM_EVENT_SOFTWARE_INT);
5562	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectReflect);
5563	hmR0SvmSetPendingEvent(pVCpu, &pVmcb->ctrl.ExitIntInfo, GCPtrFaultAddress);
5564
5565	Log4Func(("IDT: Pending vectoring event %#RX64 ErrValid=%RTbool Err=%#RX32 GCPtrFaultAddress=%#RX64\n",
5566	pVmcb->ctrl.ExitIntInfo.u, RT_BOOL(pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid),
5567	pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode, GCPtrFaultAddress));
5568	}
5569	break;
5570	}
5571
5572	case IEMXCPTRAISE_REEXEC_INSTR:
5573	{
5574	Assert(rc == VINF_SUCCESS);
5575	break;
5576	}
5577
5578	case IEMXCPTRAISE_DOUBLE_FAULT:
5579	{
5580	/*
5581	* Determing a vectoring double #PF condition. Used later, when PGM evaluates
5582	* the second #PF as a guest #PF (and not a shadow #PF) and needs to be
5583	* converted into a #DF.
5584	*/
5585	if (fRaiseInfo & IEMXCPTRAISEINFO_PF_PF)
5586	{
5587	Log4Func(("IDT: Pending vectoring double #PF uCR2=%#RX64\n", pVCpu->cpum.GstCtx.cr2));
5588	pSvmTransient->fVectoringDoublePF = true;
5589	Assert(rc == VINF_SUCCESS);
5590	}
5591	else
5592	{
5593	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectConvertDF);
5594	hmR0SvmSetPendingXcptDF(pVCpu);
5595	rc = VINF_HM_DOUBLE_FAULT;
5596	}
5597	break;
5598	}
5599
5600	case IEMXCPTRAISE_TRIPLE_FAULT:
5601	{
5602	rc = VINF_EM_RESET;
5603	break;
5604	}
5605
5606	case IEMXCPTRAISE_CPU_HANG:
5607	{
5608	rc = VERR_EM_GUEST_CPU_HANG;
5609	break;
5610	}
5611
5612	default:
5613	AssertMsgFailedBreakStmt(("Bogus enmRaise value: %d (%#x)\n", enmRaise, enmRaise), rc = VERR_SVM_IPE_2);
5614	}
5615	}
5616	Assert(rc == VINF_SUCCESS \|\| rc == VINF_HM_DOUBLE_FAULT \|\| rc == VINF_EM_RESET \|\| rc == VERR_EM_GUEST_CPU_HANG);
5617	return rc;
5618	}
5619
5620
5621	/**
5622	* Advances the guest RIP by the number of bytes specified in @a cb.
5623	*
5624	* @param pVCpu The cross context virtual CPU structure.
5625	* @param cb RIP increment value in bytes.
5626	*/
5627	DECLINLINE(void) hmR0SvmAdvanceRip(PVMCPUCC pVCpu, uint32_t cb)
5628	{
5629	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
5630	pCtx->rip += cb;
5631
5632	/* Update interrupt shadow. */
5633	if ( VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS)
5634	&& pCtx->rip != EMGetInhibitInterruptsPC(pVCpu))
5635	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS);
5636	}
5637
5638
5639	/* -=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= */
5640	/* -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- #VMEXIT handlers -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- */
5641	/* -=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= */
5642
5643	/** @name \#VMEXIT handlers.
5644	* @{
5645	*/
5646
5647	/**
5648	* \#VMEXIT handler for external interrupts, NMIs, FPU assertion freeze and INIT
5649	* signals (SVM_EXIT_INTR, SVM_EXIT_NMI, SVM_EXIT_FERR_FREEZE, SVM_EXIT_INIT).
5650	*/
5651	HMSVM_EXIT_DECL hmR0SvmExitIntr(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
5652	{
5653	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
5654
5655	if (pSvmTransient->u64ExitCode == SVM_EXIT_NMI)
5656	STAM_REL_COUNTER_INC(&pVCpu->hm.s.StatExitHostNmiInGC);
5657	else if (pSvmTransient->u64ExitCode == SVM_EXIT_INTR)
5658	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitExtInt);
5659
5660	/*
5661	* AMD-V has no preemption timer and the generic periodic preemption timer has no way to
5662	* signal -before- the timer fires if the current interrupt is our own timer or a some
5663	* other host interrupt. We also cannot examine what interrupt it is until the host
5664	* actually take the interrupt.
5665	*
5666	* Going back to executing guest code here unconditionally causes random scheduling
5667	* problems (observed on an AMD Phenom 9850 Quad-Core on Windows 64-bit host).
5668	*/
5669	return VINF_EM_RAW_INTERRUPT;
5670	}
5671
5672
5673	/**
5674	* \#VMEXIT handler for WBINVD (SVM_EXIT_WBINVD). Conditional \#VMEXIT.
5675	*/
5676	HMSVM_EXIT_DECL hmR0SvmExitWbinvd(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
5677	{
5678	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
5679
5680	VBOXSTRICTRC rcStrict;
5681	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
5682	if (fSupportsNextRipSave)
5683	{
5684	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK);
5685	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
5686	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
5687	rcStrict = IEMExecDecodedWbinvd(pVCpu, cbInstr);
5688	}
5689	else
5690	{
5691	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
5692	rcStrict = IEMExecOne(pVCpu);
5693	}
5694
5695	if (rcStrict == VINF_IEM_RAISED_XCPT)
5696	{
5697	rcStrict = VINF_SUCCESS;
5698	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
5699	}
5700	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
5701	return VBOXSTRICTRC_TODO(rcStrict);
5702	}
5703
5704
5705	/**
5706	* \#VMEXIT handler for INVD (SVM_EXIT_INVD). Unconditional \#VMEXIT.
5707	*/
5708	HMSVM_EXIT_DECL hmR0SvmExitInvd(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
5709	{
5710	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
5711
5712	VBOXSTRICTRC rcStrict;
5713	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
5714	if (fSupportsNextRipSave)
5715	{
5716	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK);
5717	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
5718	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
5719	rcStrict = IEMExecDecodedInvd(pVCpu, cbInstr);
5720	}
5721	else
5722	{
5723	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
5724	rcStrict = IEMExecOne(pVCpu);
5725	}
5726
5727	if (rcStrict == VINF_IEM_RAISED_XCPT)
5728	{
5729	rcStrict = VINF_SUCCESS;
5730	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
5731	}
5732	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
5733	return VBOXSTRICTRC_TODO(rcStrict);
5734	}
5735
5736
5737	/**
5738	* \#VMEXIT handler for INVD (SVM_EXIT_CPUID). Conditional \#VMEXIT.
5739	*/
5740	HMSVM_EXIT_DECL hmR0SvmExitCpuid(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
5741	{
5742	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
5743
5744	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK \| CPUMCTX_EXTRN_RAX \| CPUMCTX_EXTRN_RCX);
5745	VBOXSTRICTRC rcStrict;
5746	PCEMEXITREC pExitRec = EMHistoryUpdateFlagsAndTypeAndPC(pVCpu,
5747	EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_CPUID),
5748	pVCpu->cpum.GstCtx.rip + pVCpu->cpum.GstCtx.cs.u64Base);
5749	if (!pExitRec)
5750	{
5751	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
5752	if (fSupportsNextRipSave)
5753	{
5754	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
5755	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
5756	rcStrict = IEMExecDecodedCpuid(pVCpu, cbInstr);
5757	}
5758	else
5759	{
5760	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
5761	rcStrict = IEMExecOne(pVCpu);
5762	}
5763
5764	if (rcStrict == VINF_IEM_RAISED_XCPT)
5765	{
5766	rcStrict = VINF_SUCCESS;
5767	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
5768	}
5769	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
5770	}
5771	else
5772	{
5773	/*
5774	* Frequent exit or something needing probing. Get state and call EMHistoryExec.
5775	*/
5776	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
5777
5778	Log4(("CpuIdExit/%u: %04x:%08RX64: %#x/%#x -> EMHistoryExec\n",
5779	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip, pVCpu->cpum.GstCtx.eax, pVCpu->cpum.GstCtx.ecx));
5780
5781	rcStrict = EMHistoryExec(pVCpu, pExitRec, 0);
5782
5783	Log4(("CpuIdExit/%u: %04x:%08RX64: EMHistoryExec -> %Rrc + %04x:%08RX64\n",
5784	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip,
5785	VBOXSTRICTRC_VAL(rcStrict), pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip));
5786	}
5787	return VBOXSTRICTRC_TODO(rcStrict);
5788	}
5789
5790
5791	/**
5792	* \#VMEXIT handler for RDTSC (SVM_EXIT_RDTSC). Conditional \#VMEXIT.
5793	*/
5794	HMSVM_EXIT_DECL hmR0SvmExitRdtsc(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
5795	{
5796	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
5797
5798	VBOXSTRICTRC rcStrict;
5799	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
5800	if (fSupportsNextRipSave)
5801	{
5802	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK \| CPUMCTX_EXTRN_CR4);
5803	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
5804	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
5805	rcStrict = IEMExecDecodedRdtsc(pVCpu, cbInstr);
5806	}
5807	else
5808	{
5809	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
5810	rcStrict = IEMExecOne(pVCpu);
5811	}
5812
5813	if (rcStrict == VINF_SUCCESS)
5814	pSvmTransient->fUpdateTscOffsetting = true;
5815	else if (rcStrict == VINF_IEM_RAISED_XCPT)
5816	{
5817	rcStrict = VINF_SUCCESS;
5818	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
5819	}
5820	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
5821	return VBOXSTRICTRC_TODO(rcStrict);
5822	}
5823
5824
5825	/**
5826	* \#VMEXIT handler for RDTSCP (SVM_EXIT_RDTSCP). Conditional \#VMEXIT.
5827	*/
5828	HMSVM_EXIT_DECL hmR0SvmExitRdtscp(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
5829	{
5830	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
5831
5832	VBOXSTRICTRC rcStrict;
5833	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
5834	if (fSupportsNextRipSave)
5835	{
5836	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK \| CPUMCTX_EXTRN_CR4 \| CPUMCTX_EXTRN_TSC_AUX);
5837	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
5838	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
5839	rcStrict = IEMExecDecodedRdtscp(pVCpu, cbInstr);
5840	}
5841	else
5842	{
5843	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
5844	rcStrict = IEMExecOne(pVCpu);
5845	}
5846
5847	if (rcStrict == VINF_SUCCESS)
5848	pSvmTransient->fUpdateTscOffsetting = true;
5849	else if (rcStrict == VINF_IEM_RAISED_XCPT)
5850	{
5851	rcStrict = VINF_SUCCESS;
5852	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
5853	}
5854	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
5855	return VBOXSTRICTRC_TODO(rcStrict);
5856	}
5857
5858
5859	/**
5860	* \#VMEXIT handler for RDPMC (SVM_EXIT_RDPMC). Conditional \#VMEXIT.
5861	*/
5862	HMSVM_EXIT_DECL hmR0SvmExitRdpmc(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
5863	{
5864	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
5865
5866	VBOXSTRICTRC rcStrict;
5867	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
5868	if (fSupportsNextRipSave)
5869	{
5870	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK \| CPUMCTX_EXTRN_CR4);
5871	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
5872	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
5873	rcStrict = IEMExecDecodedRdpmc(pVCpu, cbInstr);
5874	}
5875	else
5876	{
5877	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
5878	rcStrict = IEMExecOne(pVCpu);
5879	}
5880
5881	if (rcStrict == VINF_IEM_RAISED_XCPT)
5882	{
5883	rcStrict = VINF_SUCCESS;
5884	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
5885	}
5886	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
5887	return VBOXSTRICTRC_TODO(rcStrict);
5888	}
5889
5890
5891	/**
5892	* \#VMEXIT handler for INVLPG (SVM_EXIT_INVLPG). Conditional \#VMEXIT.
5893	*/
5894	HMSVM_EXIT_DECL hmR0SvmExitInvlpg(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
5895	{
5896	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
5897	Assert(!pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
5898
5899	VBOXSTRICTRC rcStrict;
5900	bool const fSupportsDecodeAssists = hmR0SvmSupportsDecodeAssists(pVCpu);
5901	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
5902	if ( fSupportsDecodeAssists
5903	&& fSupportsNextRipSave)
5904	{
5905	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_MEM_MASK);
5906	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
5907	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
5908	RTGCPTR const GCPtrPage = pVmcb->ctrl.u64ExitInfo1;
5909	rcStrict = IEMExecDecodedInvlpg(pVCpu, cbInstr, GCPtrPage);
5910	}
5911	else
5912	{
5913	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
5914	rcStrict = IEMExecOne(pVCpu);
5915	}
5916
5917	if (rcStrict == VINF_IEM_RAISED_XCPT)
5918	{
5919	rcStrict = VINF_SUCCESS;
5920	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
5921	}
5922	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
5923	return VBOXSTRICTRC_VAL(rcStrict);
5924	}
5925
5926
5927	/**
5928	* \#VMEXIT handler for HLT (SVM_EXIT_HLT). Conditional \#VMEXIT.
5929	*/
5930	HMSVM_EXIT_DECL hmR0SvmExitHlt(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
5931	{
5932	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
5933
5934	VBOXSTRICTRC rcStrict;
5935	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
5936	if (fSupportsNextRipSave)
5937	{
5938	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK);
5939	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
5940	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
5941	rcStrict = IEMExecDecodedHlt(pVCpu, cbInstr);
5942	}
5943	else
5944	{
5945	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
5946	rcStrict = IEMExecOne(pVCpu);
5947	}
5948
5949	if ( rcStrict == VINF_EM_HALT
5950	\|\| rcStrict == VINF_SUCCESS)
5951	rcStrict = EMShouldContinueAfterHalt(pVCpu, &pVCpu->cpum.GstCtx) ? VINF_SUCCESS : VINF_EM_HALT;
5952	else if (rcStrict == VINF_IEM_RAISED_XCPT)
5953	{
5954	rcStrict = VINF_SUCCESS;
5955	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
5956	}
5957	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
5958	if (rcStrict != VINF_SUCCESS)
5959	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHltToR3);
5960	return VBOXSTRICTRC_VAL(rcStrict);;
5961	}
5962
5963
5964	/**
5965	* \#VMEXIT handler for MONITOR (SVM_EXIT_MONITOR). Conditional \#VMEXIT.
5966	*/
5967	HMSVM_EXIT_DECL hmR0SvmExitMonitor(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
5968	{
5969	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
5970
5971	/*
5972	* If the instruction length is supplied by the CPU is 3 bytes, we can be certain that no
5973	* segment override prefix is present (and thus use the default segment DS). Otherwise, a
5974	* segment override prefix or other prefixes might be used, in which case we fallback to
5975	* IEMExecOne() to figure out.
5976	*/
5977	VBOXSTRICTRC rcStrict;
5978	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
5979	uint8_t const cbInstr = hmR0SvmSupportsNextRipSave(pVCpu) ? pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip : 0;
5980	if (cbInstr)
5981	{
5982	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_MEM_MASK \| CPUMCTX_EXTRN_DS);
5983	rcStrict = IEMExecDecodedMonitor(pVCpu, cbInstr);
5984	}
5985	else
5986	{
5987	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
5988	rcStrict = IEMExecOne(pVCpu);
5989	}
5990
5991	if (rcStrict == VINF_IEM_RAISED_XCPT)
5992	{
5993	rcStrict = VINF_SUCCESS;
5994	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
5995	}
5996	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
5997	return VBOXSTRICTRC_TODO(rcStrict);
5998	}
5999
6000
6001	/**
6002	* \#VMEXIT handler for MWAIT (SVM_EXIT_MWAIT). Conditional \#VMEXIT.
6003	*/
6004	HMSVM_EXIT_DECL hmR0SvmExitMwait(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
6005	{
6006	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6007
6008	VBOXSTRICTRC rcStrict;
6009	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
6010	if (fSupportsNextRipSave)
6011	{
6012	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK);
6013	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
6014	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
6015	rcStrict = IEMExecDecodedMwait(pVCpu, cbInstr);
6016	}
6017	else
6018	{
6019	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6020	rcStrict = IEMExecOne(pVCpu);
6021	}
6022
6023	if ( rcStrict == VINF_EM_HALT
6024	&& EMMonitorWaitShouldContinue(pVCpu, &pVCpu->cpum.GstCtx))
6025	rcStrict = VINF_SUCCESS;
6026	else if (rcStrict == VINF_IEM_RAISED_XCPT)
6027	{
6028	rcStrict = VINF_SUCCESS;
6029	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
6030	}
6031	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6032	return VBOXSTRICTRC_TODO(rcStrict);
6033	}
6034
6035
6036	/**
6037	* \#VMEXIT handler for shutdown (triple-fault) (SVM_EXIT_SHUTDOWN). Conditional
6038	* \#VMEXIT.
6039	*/
6040	HMSVM_EXIT_DECL hmR0SvmExitShutdown(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
6041	{
6042	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6043	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6044	return VINF_EM_RESET;
6045	}
6046
6047
6048	/**
6049	* \#VMEXIT handler for unexpected exits. Conditional \#VMEXIT.
6050	*/
6051	HMSVM_EXIT_DECL hmR0SvmExitUnexpected(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
6052	{
6053	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
6054	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6055	AssertMsgFailed(("hmR0SvmExitUnexpected: ExitCode=%#RX64 uExitInfo1=%#RX64 uExitInfo2=%#RX64\n", pSvmTransient->u64ExitCode,
6056	pVmcb->ctrl.u64ExitInfo1, pVmcb->ctrl.u64ExitInfo2));
6057	RT_NOREF(pVmcb);
6058	pVCpu->hm.s.u32HMError = (uint32_t)pSvmTransient->u64ExitCode;
6059	return VERR_SVM_UNEXPECTED_EXIT;
6060	}
6061
6062
6063	/**
6064	* \#VMEXIT handler for CRx reads (SVM_EXIT_READ_CR*). Conditional \#VMEXIT.
6065	*/
6066	HMSVM_EXIT_DECL hmR0SvmExitReadCRx(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
6067	{
6068	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6069
6070	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
6071	Log4Func(("CS:RIP=%04x:%#RX64\n", pCtx->cs.Sel, pCtx->rip));
6072	#ifdef VBOX_WITH_STATISTICS
6073	switch (pSvmTransient->u64ExitCode)
6074	{
6075	case SVM_EXIT_READ_CR0: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR0Read); break;
6076	case SVM_EXIT_READ_CR2: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR2Read); break;
6077	case SVM_EXIT_READ_CR3: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR3Read); break;
6078	case SVM_EXIT_READ_CR4: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR4Read); break;
6079	case SVM_EXIT_READ_CR8: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR8Read); break;
6080	}
6081	#endif
6082
6083	bool const fSupportsDecodeAssists = hmR0SvmSupportsDecodeAssists(pVCpu);
6084	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
6085	if ( fSupportsDecodeAssists
6086	&& fSupportsNextRipSave)
6087	{
6088	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
6089	bool const fMovCRx = RT_BOOL(pVmcb->ctrl.u64ExitInfo1 & SVM_EXIT1_MOV_CRX_MASK);
6090	if (fMovCRx)
6091	{
6092	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK \| CPUMCTX_EXTRN_CR_MASK
6093	\| CPUMCTX_EXTRN_APIC_TPR);
6094	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pCtx->rip;
6095	uint8_t const iCrReg = pSvmTransient->u64ExitCode - SVM_EXIT_READ_CR0;
6096	uint8_t const iGReg = pVmcb->ctrl.u64ExitInfo1 & SVM_EXIT1_MOV_CRX_GPR_NUMBER;
6097	VBOXSTRICTRC rcStrict = IEMExecDecodedMovCRxRead(pVCpu, cbInstr, iGReg, iCrReg);
6098	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6099	return VBOXSTRICTRC_VAL(rcStrict);
6100	}
6101	/* else: SMSW instruction, fall back below to IEM for this. */
6102	}
6103
6104	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6105	VBOXSTRICTRC rcStrict = IEMExecOne(pVCpu);
6106	AssertMsg( rcStrict == VINF_SUCCESS
6107	\|\| rcStrict == VINF_PGM_SYNC_CR3
6108	\|\| rcStrict == VINF_IEM_RAISED_XCPT,
6109	("hmR0SvmExitReadCRx: IEMExecOne failed rc=%Rrc\n", VBOXSTRICTRC_VAL(rcStrict)));
6110	Assert((pSvmTransient->u64ExitCode - SVM_EXIT_READ_CR0) <= 15);
6111	if (rcStrict == VINF_IEM_RAISED_XCPT)
6112	{
6113	rcStrict = VINF_SUCCESS;
6114	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
6115	}
6116	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6117	return VBOXSTRICTRC_TODO(rcStrict);
6118	}
6119
6120
6121	/**
6122	* \#VMEXIT handler for CRx writes (SVM_EXIT_WRITE_CR*). Conditional \#VMEXIT.
6123	*/
6124	HMSVM_EXIT_DECL hmR0SvmExitWriteCRx(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
6125	{
6126	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6127
6128	uint64_t const uExitCode = pSvmTransient->u64ExitCode;
6129	uint8_t const iCrReg = uExitCode == SVM_EXIT_CR0_SEL_WRITE ? 0 : (pSvmTransient->u64ExitCode - SVM_EXIT_WRITE_CR0);
6130	Assert(iCrReg <= 15);
6131
6132	VBOXSTRICTRC rcStrict = VERR_SVM_IPE_5;
6133	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
6134	bool fDecodedInstr = false;
6135	bool const fSupportsDecodeAssists = hmR0SvmSupportsDecodeAssists(pVCpu);
6136	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
6137	if ( fSupportsDecodeAssists
6138	&& fSupportsNextRipSave)
6139	{
6140	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
6141	bool const fMovCRx = RT_BOOL(pVmcb->ctrl.u64ExitInfo1 & SVM_EXIT1_MOV_CRX_MASK);
6142	if (fMovCRx)
6143	{
6144	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_MEM_MASK \| CPUMCTX_EXTRN_CR3 \| CPUMCTX_EXTRN_CR4
6145	\| CPUMCTX_EXTRN_APIC_TPR);
6146	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pCtx->rip;
6147	uint8_t const iGReg = pVmcb->ctrl.u64ExitInfo1 & SVM_EXIT1_MOV_CRX_GPR_NUMBER;
6148	Log4Func(("Mov CR%u w/ iGReg=%#x\n", iCrReg, iGReg));
6149	rcStrict = IEMExecDecodedMovCRxWrite(pVCpu, cbInstr, iCrReg, iGReg);
6150	fDecodedInstr = true;
6151	}
6152	/* else: LMSW or CLTS instruction, fall back below to IEM for this. */
6153	}
6154
6155	if (!fDecodedInstr)
6156	{
6157	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6158	Log4Func(("iCrReg=%#x\n", iCrReg));
6159	rcStrict = IEMExecOne(pVCpu);
6160	if (RT_UNLIKELY( rcStrict == VERR_IEM_ASPECT_NOT_IMPLEMENTED
6161	\|\| rcStrict == VERR_IEM_INSTR_NOT_IMPLEMENTED))
6162	rcStrict = VERR_EM_INTERPRETER;
6163	}
6164
6165	if (rcStrict == VINF_SUCCESS)
6166	{
6167	switch (iCrReg)
6168	{
6169	case 0:
6170	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_CR0);
6171	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR0Write);
6172	break;
6173
6174	case 2:
6175	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_CR2);
6176	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR2Write);
6177	break;
6178
6179	case 3:
6180	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_CR3);
6181	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR3Write);
6182	break;
6183
6184	case 4:
6185	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_CR4);
6186	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR4Write);
6187	break;
6188
6189	case 8:
6190	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
6191	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR8Write);
6192	break;
6193
6194	default:
6195	{
6196	AssertMsgFailed(("hmR0SvmExitWriteCRx: Invalid/Unexpected Write-CRx exit. u64ExitCode=%#RX64 %#x\n",
6197	pSvmTransient->u64ExitCode, iCrReg));
6198	break;
6199	}
6200	}
6201	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6202	}
6203	else if (rcStrict == VINF_IEM_RAISED_XCPT)
6204	{
6205	rcStrict = VINF_SUCCESS;
6206	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
6207	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6208	}
6209	else
6210	Assert(rcStrict == VERR_EM_INTERPRETER \|\| rcStrict == VINF_PGM_SYNC_CR3);
6211	return VBOXSTRICTRC_TODO(rcStrict);
6212	}
6213
6214
6215	/**
6216	* \#VMEXIT helper for read MSRs, see hmR0SvmExitMsr.
6217	*
6218	* @returns Strict VBox status code.
6219	* @param pVCpu The cross context virtual CPU structure.
6220	* @param pVmcb Pointer to the VM control block.
6221	*/
6222	static VBOXSTRICTRC hmR0SvmExitReadMsr(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
6223	{
6224	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdmsr);
6225	Log4Func(("idMsr=%#RX32\n", pVCpu->cpum.GstCtx.ecx));
6226
6227	VBOXSTRICTRC rcStrict;
6228	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
6229	if (fSupportsNextRipSave)
6230	{
6231	/** @todo Optimize this: Only retrieve the MSR bits we need here. CPUMAllMsrs.cpp
6232	* can ask for what it needs instead of using CPUMCTX_EXTRN_ALL_MSRS. */
6233	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK \| CPUMCTX_EXTRN_ALL_MSRS);
6234	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
6235	rcStrict = IEMExecDecodedRdmsr(pVCpu, cbInstr);
6236	}
6237	else
6238	{
6239	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK \| CPUMCTX_EXTRN_ALL_MSRS);
6240	rcStrict = IEMExecOne(pVCpu);
6241	}
6242
6243	AssertMsg( rcStrict == VINF_SUCCESS
6244	\|\| rcStrict == VINF_IEM_RAISED_XCPT
6245	\|\| rcStrict == VINF_CPUM_R3_MSR_READ,
6246	("hmR0SvmExitReadMsr: Unexpected status %Rrc\n", VBOXSTRICTRC_VAL(rcStrict)));
6247
6248	if (rcStrict == VINF_IEM_RAISED_XCPT)
6249	{
6250	rcStrict = VINF_SUCCESS;
6251	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
6252	}
6253	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6254	return rcStrict;
6255	}
6256
6257
6258	/**
6259	* \#VMEXIT helper for write MSRs, see hmR0SvmExitMsr.
6260	*
6261	* @returns Strict VBox status code.
6262	* @param pVCpu The cross context virtual CPU structure.
6263	* @param pVmcb Pointer to the VM control block.
6264	* @param pSvmTransient Pointer to the SVM-transient structure.
6265	*/
6266	static VBOXSTRICTRC hmR0SvmExitWriteMsr(PVMCPUCC pVCpu, PSVMVMCB pVmcb, PSVMTRANSIENT pSvmTransient)
6267	{
6268	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
6269	uint32_t const idMsr = pCtx->ecx;
6270	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitWrmsr);
6271	Log4Func(("idMsr=%#RX32\n", idMsr));
6272
6273	/*
6274	* Handle TPR patching MSR writes.
6275	* We utilitize the LSTAR MSR for patching.
6276	*/
6277	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
6278	if ( pVCpu->CTX_SUFF(pVM)->hm.s.fTPRPatchingActive
6279	&& idMsr == MSR_K8_LSTAR)
6280	{
6281	unsigned cbInstr;
6282	if (fSupportsNextRipSave)
6283	cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
6284	else
6285	{
6286	PDISCPUSTATE pDis = &pVCpu->hm.s.DisState;
6287	int rc = EMInterpretDisasCurrent(pVCpu->CTX_SUFF(pVM), pVCpu, pDis, &cbInstr);
6288	if ( rc == VINF_SUCCESS
6289	&& pDis->pCurInstr->uOpcode == OP_WRMSR)
6290	Assert(cbInstr > 0);
6291	else
6292	cbInstr = 0;
6293	}
6294
6295	/* Our patch code uses LSTAR for TPR caching for 32-bit guests. */
6296	if ((pCtx->eax & 0xff) != pSvmTransient->u8GuestTpr)
6297	{
6298	int rc = APICSetTpr(pVCpu, pCtx->eax & 0xff);
6299	AssertRCReturn(rc, rc);
6300	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
6301	}
6302
6303	int rc = VINF_SUCCESS;
6304	hmR0SvmAdvanceRip(pVCpu, cbInstr);
6305	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6306	return rc;
6307	}
6308
6309	/*
6310	* Handle regular MSR writes.
6311	*/
6312	VBOXSTRICTRC rcStrict;
6313	if (fSupportsNextRipSave)
6314	{
6315	/** @todo Optimize this: We don't need to get much of the MSR state here
6316	* since we're only updating. CPUMAllMsrs.cpp can ask for what it needs and
6317	* clear the applicable extern flags. */
6318	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK \| CPUMCTX_EXTRN_ALL_MSRS);
6319	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
6320	rcStrict = IEMExecDecodedWrmsr(pVCpu, cbInstr);
6321	}
6322	else
6323	{
6324	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK \| CPUMCTX_EXTRN_ALL_MSRS);
6325	rcStrict = IEMExecOne(pVCpu);
6326	}
6327
6328	AssertMsg( rcStrict == VINF_SUCCESS
6329	\|\| rcStrict == VINF_IEM_RAISED_XCPT
6330	\|\| rcStrict == VINF_CPUM_R3_MSR_WRITE,
6331	("hmR0SvmExitWriteMsr: Unexpected status %Rrc\n", VBOXSTRICTRC_VAL(rcStrict)));
6332
6333	if (rcStrict == VINF_SUCCESS)
6334	{
6335	/* If this is an X2APIC WRMSR access, update the APIC TPR state. */
6336	if ( idMsr >= MSR_IA32_X2APIC_START
6337	&& idMsr <= MSR_IA32_X2APIC_END)
6338	{
6339	/*
6340	* We've already saved the APIC related guest-state (TPR) in hmR0SvmPostRunGuest().
6341	* When full APIC register virtualization is implemented we'll have to make sure
6342	* APIC state is saved from the VMCB before IEM changes it.
6343	*/
6344	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
6345	}
6346	else
6347	{
6348	switch (idMsr)
6349	{
6350	case MSR_IA32_TSC: pSvmTransient->fUpdateTscOffsetting = true; break;
6351	case MSR_K6_EFER: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_EFER_MSR); break;
6352	case MSR_K8_FS_BASE: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_FS); break;
6353	case MSR_K8_GS_BASE: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_GS); break;
6354	case MSR_IA32_SYSENTER_CS: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_SYSENTER_CS_MSR); break;
6355	case MSR_IA32_SYSENTER_EIP: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_SYSENTER_EIP_MSR); break;
6356	case MSR_IA32_SYSENTER_ESP: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_SYSENTER_ESP_MSR); break;
6357	}
6358	}
6359	}
6360	else if (rcStrict == VINF_IEM_RAISED_XCPT)
6361	{
6362	rcStrict = VINF_SUCCESS;
6363	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
6364	}
6365	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6366	return rcStrict;
6367	}
6368
6369
6370	/**
6371	* \#VMEXIT handler for MSR read and writes (SVM_EXIT_MSR). Conditional
6372	* \#VMEXIT.
6373	*/
6374	HMSVM_EXIT_DECL hmR0SvmExitMsr(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
6375	{
6376	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6377
6378	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
6379	if (pVmcb->ctrl.u64ExitInfo1 == SVM_EXIT1_MSR_READ)
6380	return VBOXSTRICTRC_TODO(hmR0SvmExitReadMsr(pVCpu, pVmcb));
6381
6382	Assert(pVmcb->ctrl.u64ExitInfo1 == SVM_EXIT1_MSR_WRITE);
6383	return VBOXSTRICTRC_TODO(hmR0SvmExitWriteMsr(pVCpu, pVmcb, pSvmTransient));
6384	}
6385
6386
6387	/**
6388	* \#VMEXIT handler for DRx read (SVM_EXIT_READ_DRx). Conditional \#VMEXIT.
6389	*/
6390	HMSVM_EXIT_DECL hmR0SvmExitReadDRx(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
6391	{
6392	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6393	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6394
6395	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitDRxRead);
6396
6397	/** @todo Stepping with nested-guest. */
6398	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
6399	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
6400	{
6401	/* We should -not- get this #VMEXIT if the guest's debug registers were active. */
6402	if (pSvmTransient->fWasGuestDebugStateActive)
6403	{
6404	AssertMsgFailed(("hmR0SvmExitReadDRx: Unexpected exit %#RX32\n", (uint32_t)pSvmTransient->u64ExitCode));
6405	pVCpu->hm.s.u32HMError = (uint32_t)pSvmTransient->u64ExitCode;
6406	return VERR_SVM_UNEXPECTED_EXIT;
6407	}
6408
6409	/*
6410	* Lazy DR0-3 loading.
6411	*/
6412	if (!pSvmTransient->fWasHyperDebugStateActive)
6413	{
6414	Assert(!DBGFIsStepping(pVCpu)); Assert(!pVCpu->hm.s.fSingleInstruction);
6415	Log5(("hmR0SvmExitReadDRx: Lazy loading guest debug registers\n"));
6416
6417	/* Don't intercept DRx read and writes. */
6418	PSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
6419	pVmcb->ctrl.u16InterceptRdDRx = 0;
6420	pVmcb->ctrl.u16InterceptWrDRx = 0;
6421	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
6422
6423	/* We're playing with the host CPU state here, make sure we don't preempt or longjmp. */
6424	VMMRZCallRing3Disable(pVCpu);
6425	HM_DISABLE_PREEMPT(pVCpu);
6426
6427	/* Save the host & load the guest debug state, restart execution of the MOV DRx instruction. */
6428	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
6429	Assert(CPUMIsGuestDebugStateActive(pVCpu));
6430
6431	HM_RESTORE_PREEMPT();
6432	VMMRZCallRing3Enable(pVCpu);
6433
6434	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxContextSwitch);
6435	return VINF_SUCCESS;
6436	}
6437	}
6438
6439	/*
6440	* Interpret the read/writing of DRx.
6441	*/
6442	/** @todo Decode assist. */
6443	VBOXSTRICTRC rc = EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */);
6444	Log5(("hmR0SvmExitReadDRx: Emulated DRx access: rc=%Rrc\n", VBOXSTRICTRC_VAL(rc)));
6445	if (RT_LIKELY(rc == VINF_SUCCESS))
6446	{
6447	/* Not necessary for read accesses but whatever doesn't hurt for now, will be fixed with decode assist. */
6448	/** @todo CPUM should set this flag! */
6449	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_DR_MASK);
6450	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6451	}
6452	else
6453	Assert(rc == VERR_EM_INTERPRETER);
6454	return VBOXSTRICTRC_TODO(rc);
6455	}
6456
6457
6458	/**
6459	* \#VMEXIT handler for DRx write (SVM_EXIT_WRITE_DRx). Conditional \#VMEXIT.
6460	*/
6461	HMSVM_EXIT_DECL hmR0SvmExitWriteDRx(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
6462	{
6463	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6464	/* For now it's the same since we interpret the instruction anyway. Will change when using of Decode Assist is implemented. */
6465	int rc = hmR0SvmExitReadDRx(pVCpu, pSvmTransient);
6466	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitDRxWrite);
6467	STAM_COUNTER_DEC(&pVCpu->hm.s.StatExitDRxRead);
6468	return rc;
6469	}
6470
6471
6472	/**
6473	* \#VMEXIT handler for XCRx write (SVM_EXIT_XSETBV). Conditional \#VMEXIT.
6474	*/
6475	HMSVM_EXIT_DECL hmR0SvmExitXsetbv(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
6476	{
6477	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6478	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6479
6480	/** @todo decode assists... */
6481	VBOXSTRICTRC rcStrict = IEMExecOne(pVCpu);
6482	if (RT_LIKELY(rcStrict == VINF_SUCCESS))
6483	{
6484	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
6485	pVCpu->hm.s.fLoadSaveGuestXcr0 = (pCtx->cr4 & X86_CR4_OSXSAVE) && pCtx->aXcr[0] != ASMGetXcr0();
6486	Log4Func(("New XCR0=%#RX64 fLoadSaveGuestXcr0=%RTbool (cr4=%#RX64)\n", pCtx->aXcr[0], pVCpu->hm.s.fLoadSaveGuestXcr0,
6487	pCtx->cr4));
6488	}
6489	else if (rcStrict == VINF_IEM_RAISED_XCPT)
6490	{
6491	rcStrict = VINF_SUCCESS;
6492	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
6493	}
6494	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6495	return VBOXSTRICTRC_TODO(rcStrict);
6496	}
6497
6498
6499	/**
6500	* \#VMEXIT handler for I/O instructions (SVM_EXIT_IOIO). Conditional \#VMEXIT.
6501	*/
6502	HMSVM_EXIT_DECL hmR0SvmExitIOInstr(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
6503	{
6504	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6505	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK \| CPUMCTX_EXTRN_SREG_MASK);
6506
6507	/* I/O operation lookup arrays. */
6508	static uint32_t const s_aIOSize[8] = { 0, 1, 2, 0, 4, 0, 0, 0 }; /* Size of the I/O accesses in bytes. */
6509	static uint32_t const s_aIOOpAnd[8] = { 0, 0xff, 0xffff, 0, 0xffffffff, 0, 0, 0 }; /* AND masks for saving
6510	the result (in AL/AX/EAX). */
6511	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
6512	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
6513	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
6514
6515	Log4Func(("CS:RIP=%04x:%#RX64\n", pCtx->cs.Sel, pCtx->rip));
6516
6517	/* Refer AMD spec. 15.10.2 "IN and OUT Behaviour" and Figure 15-2. "EXITINFO1 for IOIO Intercept" for the format. */
6518	SVMIOIOEXITINFO IoExitInfo;
6519	IoExitInfo.u = (uint32_t)pVmcb->ctrl.u64ExitInfo1;
6520	uint32_t uIOWidth = (IoExitInfo.u >> 4) & 0x7;
6521	uint32_t cbValue = s_aIOSize[uIOWidth];
6522	uint32_t uAndVal = s_aIOOpAnd[uIOWidth];
6523
6524	if (RT_UNLIKELY(!cbValue))
6525	{
6526	AssertMsgFailed(("hmR0SvmExitIOInstr: Invalid IO operation. uIOWidth=%u\n", uIOWidth));
6527	return VERR_EM_INTERPRETER;
6528	}
6529
6530	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CS \| CPUMCTX_EXTRN_RIP \| CPUMCTX_EXTRN_RFLAGS);
6531	VBOXSTRICTRC rcStrict;
6532	PCEMEXITREC pExitRec = NULL;
6533	if ( !pVCpu->hm.s.fSingleInstruction
6534	&& !pVCpu->cpum.GstCtx.eflags.Bits.u1TF)
6535	pExitRec = EMHistoryUpdateFlagsAndTypeAndPC(pVCpu,
6536	!IoExitInfo.n.u1Str
6537	? IoExitInfo.n.u1Type == SVM_IOIO_READ
6538	? EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_IO_PORT_READ)
6539	: EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_IO_PORT_WRITE)
6540	: IoExitInfo.n.u1Type == SVM_IOIO_READ
6541	? EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_IO_PORT_STR_READ)
6542	: EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_IO_PORT_STR_WRITE),
6543	pVCpu->cpum.GstCtx.rip + pVCpu->cpum.GstCtx.cs.u64Base);
6544	if (!pExitRec)
6545	{
6546	bool fUpdateRipAlready = false;
6547	if (IoExitInfo.n.u1Str)
6548	{
6549	/* INS/OUTS - I/O String instruction. */
6550	/** @todo Huh? why can't we use the segment prefix information given by AMD-V
6551	* in EXITINFO1? Investigate once this thing is up and running. */
6552	Log4Func(("CS:RIP=%04x:%08RX64 %#06x/%u %c str\n", pCtx->cs.Sel, pCtx->rip, IoExitInfo.n.u16Port, cbValue,
6553	IoExitInfo.n.u1Type == SVM_IOIO_WRITE ? 'w' : 'r'));
6554	AssertReturn(pCtx->dx == IoExitInfo.n.u16Port, VERR_SVM_IPE_2);
6555	static IEMMODE const s_aenmAddrMode[8] =
6556	{
6557	(IEMMODE)-1, IEMMODE_16BIT, IEMMODE_32BIT, (IEMMODE)-1, IEMMODE_64BIT, (IEMMODE)-1, (IEMMODE)-1, (IEMMODE)-1
6558	};
6559	IEMMODE enmAddrMode = s_aenmAddrMode[(IoExitInfo.u >> 7) & 0x7];
6560	if (enmAddrMode != (IEMMODE)-1)
6561	{
6562	uint64_t cbInstr = pVmcb->ctrl.u64ExitInfo2 - pCtx->rip;
6563	if (cbInstr <= 15 && cbInstr >= 1)
6564	{
6565	Assert(cbInstr >= 1U + IoExitInfo.n.u1Rep);
6566	if (IoExitInfo.n.u1Type == SVM_IOIO_WRITE)
6567	{
6568	/* Don't know exactly how to detect whether u3Seg is valid, currently
6569	only enabling it for Bulldozer and later with NRIP. OS/2 broke on
6570	2384 Opterons when only checking NRIP. */
6571	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
6572	if ( fSupportsNextRipSave
6573	&& pVM->cpum.ro.GuestFeatures.enmMicroarch >= kCpumMicroarch_AMD_15h_First)
6574	{
6575	AssertMsg(IoExitInfo.n.u3Seg == X86_SREG_DS \|\| cbInstr > 1U + IoExitInfo.n.u1Rep,
6576	("u32Seg=%d cbInstr=%d u1REP=%d", IoExitInfo.n.u3Seg, cbInstr, IoExitInfo.n.u1Rep));
6577	rcStrict = IEMExecStringIoWrite(pVCpu, cbValue, enmAddrMode, IoExitInfo.n.u1Rep, (uint8_t)cbInstr,
6578	IoExitInfo.n.u3Seg, true /fIoChecked/);
6579	}
6580	else if (cbInstr == 1U + IoExitInfo.n.u1Rep)
6581	rcStrict = IEMExecStringIoWrite(pVCpu, cbValue, enmAddrMode, IoExitInfo.n.u1Rep, (uint8_t)cbInstr,
6582	X86_SREG_DS, true /fIoChecked/);
6583	else
6584	rcStrict = IEMExecOne(pVCpu);
6585	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOStringWrite);
6586	}
6587	else
6588	{
6589	AssertMsg(IoExitInfo.n.u3Seg == X86_SREG_ES /=0/, ("%#x\n", IoExitInfo.n.u3Seg));
6590	rcStrict = IEMExecStringIoRead(pVCpu, cbValue, enmAddrMode, IoExitInfo.n.u1Rep, (uint8_t)cbInstr,
6591	true /fIoChecked/);
6592	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOStringRead);
6593	}
6594	}
6595	else
6596	{
6597	AssertMsgFailed(("rip=%RX64 nrip=%#RX64 cbInstr=%#RX64\n", pCtx->rip, pVmcb->ctrl.u64ExitInfo2, cbInstr));
6598	rcStrict = IEMExecOne(pVCpu);
6599	}
6600	}
6601	else
6602	{
6603	AssertMsgFailed(("IoExitInfo=%RX64\n", IoExitInfo.u));
6604	rcStrict = IEMExecOne(pVCpu);
6605	}
6606	fUpdateRipAlready = true;
6607	}
6608	else
6609	{
6610	/* IN/OUT - I/O instruction. */
6611	Assert(!IoExitInfo.n.u1Rep);
6612
6613	uint8_t const cbInstr = pVmcb->ctrl.u64ExitInfo2 - pCtx->rip;
6614	if (IoExitInfo.n.u1Type == SVM_IOIO_WRITE)
6615	{
6616	rcStrict = IOMIOPortWrite(pVM, pVCpu, IoExitInfo.n.u16Port, pCtx->eax & uAndVal, cbValue);
6617	if ( rcStrict == VINF_IOM_R3_IOPORT_WRITE
6618	&& !pCtx->eflags.Bits.u1TF)
6619	rcStrict = EMRZSetPendingIoPortWrite(pVCpu, IoExitInfo.n.u16Port, cbInstr, cbValue, pCtx->eax & uAndVal);
6620	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOWrite);
6621	}
6622	else
6623	{
6624	uint32_t u32Val = 0;
6625	rcStrict = IOMIOPortRead(pVM, pVCpu, IoExitInfo.n.u16Port, &u32Val, cbValue);
6626	if (IOM_SUCCESS(rcStrict))
6627	{
6628	/* Save result of I/O IN instr. in AL/AX/EAX. */
6629	/** @todo r=bird: 32-bit op size should clear high bits of rax! */
6630	pCtx->eax = (pCtx->eax & ~uAndVal) \| (u32Val & uAndVal);
6631	}
6632	else if ( rcStrict == VINF_IOM_R3_IOPORT_READ
6633	&& !pCtx->eflags.Bits.u1TF)
6634	rcStrict = EMRZSetPendingIoPortRead(pVCpu, IoExitInfo.n.u16Port, cbInstr, cbValue);
6635
6636	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIORead);
6637	}
6638	}
6639
6640	if (IOM_SUCCESS(rcStrict))
6641	{
6642	/* AMD-V saves the RIP of the instruction following the IO instruction in EXITINFO2. */
6643	if (!fUpdateRipAlready)
6644	pCtx->rip = pVmcb->ctrl.u64ExitInfo2;
6645
6646	/*
6647	* If any I/O breakpoints are armed, we need to check if one triggered
6648	* and take appropriate action.
6649	* Note that the I/O breakpoint type is undefined if CR4.DE is 0.
6650	*/
6651	/** @todo Optimize away the DBGFBpIsHwIoArmed call by having DBGF tell the
6652	* execution engines about whether hyper BPs and such are pending. */
6653	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_DR7);
6654	uint32_t const uDr7 = pCtx->dr[7];
6655	if (RT_UNLIKELY( ( (uDr7 & X86_DR7_ENABLED_MASK)
6656	&& X86_DR7_ANY_RW_IO(uDr7)
6657	&& (pCtx->cr4 & X86_CR4_DE))
6658	\|\| DBGFBpIsHwIoArmed(pVM)))
6659	{
6660	/* We're playing with the host CPU state here, make sure we don't preempt or longjmp. */
6661	VMMRZCallRing3Disable(pVCpu);
6662	HM_DISABLE_PREEMPT(pVCpu);
6663
6664	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxIoCheck);
6665	CPUMR0DebugStateMaybeSaveGuest(pVCpu, false /fDr6/);
6666
6667	VBOXSTRICTRC rcStrict2 = DBGFBpCheckIo(pVM, pVCpu, &pVCpu->cpum.GstCtx, IoExitInfo.n.u16Port, cbValue);
6668	if (rcStrict2 == VINF_EM_RAW_GUEST_TRAP)
6669	{
6670	/* Raise #DB. */
6671	pVmcb->guest.u64DR6 = pCtx->dr[6];
6672	pVmcb->guest.u64DR7 = pCtx->dr[7];
6673	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
6674	hmR0SvmSetPendingXcptDB(pVCpu);
6675	}
6676	/* rcStrict is VINF_SUCCESS, VINF_IOM_R3_IOPORT_COMMIT_WRITE, or in [VINF_EM_FIRST..VINF_EM_LAST],
6677	however we can ditch VINF_IOM_R3_IOPORT_COMMIT_WRITE as it has VMCPU_FF_IOM as backup. */
6678	else if ( rcStrict2 != VINF_SUCCESS
6679	&& (rcStrict == VINF_SUCCESS \|\| rcStrict2 < rcStrict))
6680	rcStrict = rcStrict2;
6681	AssertCompile(VINF_EM_LAST < VINF_IOM_R3_IOPORT_COMMIT_WRITE);
6682
6683	HM_RESTORE_PREEMPT();
6684	VMMRZCallRing3Enable(pVCpu);
6685	}
6686
6687	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6688	}
6689
6690	#ifdef VBOX_STRICT
6691	if ( rcStrict == VINF_IOM_R3_IOPORT_READ
6692	\|\| rcStrict == VINF_EM_PENDING_R3_IOPORT_READ)
6693	Assert(IoExitInfo.n.u1Type == SVM_IOIO_READ);
6694	else if ( rcStrict == VINF_IOM_R3_IOPORT_WRITE
6695	\|\| rcStrict == VINF_IOM_R3_IOPORT_COMMIT_WRITE
6696	\|\| rcStrict == VINF_EM_PENDING_R3_IOPORT_WRITE)
6697	Assert(IoExitInfo.n.u1Type == SVM_IOIO_WRITE);
6698	else
6699	{
6700	/** @todo r=bird: This is missing a bunch of VINF_EM_FIRST..VINF_EM_LAST
6701	* statuses, that the VMM device and some others may return. See
6702	* IOM_SUCCESS() for guidance. */
6703	AssertMsg( RT_FAILURE(rcStrict)
6704	\|\| rcStrict == VINF_SUCCESS
6705	\|\| rcStrict == VINF_EM_RAW_EMULATE_INSTR
6706	\|\| rcStrict == VINF_EM_DBG_BREAKPOINT
6707	\|\| rcStrict == VINF_EM_RAW_GUEST_TRAP
6708	\|\| rcStrict == VINF_EM_DBG_STEPPED
6709	\|\| rcStrict == VINF_EM_RAW_TO_R3
6710	\|\| rcStrict == VINF_TRPM_XCPT_DISPATCHED
6711	\|\| rcStrict == VINF_EM_TRIPLE_FAULT, ("%Rrc\n", VBOXSTRICTRC_VAL(rcStrict)));
6712	}
6713	#endif
6714	}
6715	else
6716	{
6717	/*
6718	* Frequent exit or something needing probing. Get state and call EMHistoryExec.
6719	*/
6720	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6721	STAM_COUNTER_INC(!IoExitInfo.n.u1Str
6722	? IoExitInfo.n.u1Type == SVM_IOIO_WRITE ? &pVCpu->hm.s.StatExitIOWrite : &pVCpu->hm.s.StatExitIORead
6723	: IoExitInfo.n.u1Type == SVM_IOIO_WRITE ? &pVCpu->hm.s.StatExitIOStringWrite : &pVCpu->hm.s.StatExitIOStringRead);
6724	Log4(("IOExit/%u: %04x:%08RX64: %s%s%s %#x LB %u -> EMHistoryExec\n",
6725	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip, IoExitInfo.n.u1Rep ? "REP " : "",
6726	IoExitInfo.n.u1Type == SVM_IOIO_WRITE ? "OUT" : "IN", IoExitInfo.n.u1Str ? "S" : "", IoExitInfo.n.u16Port, uIOWidth));
6727
6728	rcStrict = EMHistoryExec(pVCpu, pExitRec, 0);
6729	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
6730
6731	Log4(("IOExit/%u: %04x:%08RX64: EMHistoryExec -> %Rrc + %04x:%08RX64\n",
6732	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip,
6733	VBOXSTRICTRC_VAL(rcStrict), pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip));
6734	}
6735	return VBOXSTRICTRC_TODO(rcStrict);
6736	}
6737
6738
6739	/**
6740	* \#VMEXIT handler for Nested Page-faults (SVM_EXIT_NPF). Conditional \#VMEXIT.
6741	*/
6742	HMSVM_EXIT_DECL hmR0SvmExitNestedPF(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
6743	{
6744	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6745	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6746	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(pVCpu, pSvmTransient);
6747
6748	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
6749	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
6750	Assert(pVM->hm.s.fNestedPaging);
6751
6752	/* See AMD spec. 15.25.6 "Nested versus Guest Page Faults, Fault Ordering" for VMCB details for #NPF. */
6753	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
6754	RTGCPHYS GCPhysFaultAddr = pVmcb->ctrl.u64ExitInfo2;
6755	uint32_t u32ErrCode = pVmcb->ctrl.u64ExitInfo1; /* Note! High bits in EXITINFO1 may contain additional info and are
6756	thus intentionally not copied into u32ErrCode. */
6757
6758	Log4Func(("#NPF at CS:RIP=%04x:%#RX64 GCPhysFaultAddr=%RGp ErrCode=%#x \n", pCtx->cs.Sel, pCtx->rip, GCPhysFaultAddr,
6759	u32ErrCode));
6760
6761	/*
6762	* TPR patching for 32-bit guests, using the reserved bit in the page tables for MMIO regions.
6763	*/
6764	if ( pVM->hm.s.fTprPatchingAllowed
6765	&& (GCPhysFaultAddr & PAGE_OFFSET_MASK) == XAPIC_OFF_TPR
6766	&& ( !(u32ErrCode & X86_TRAP_PF_P) /* Not present */
6767	\|\| (u32ErrCode & (X86_TRAP_PF_P \| X86_TRAP_PF_RSVD)) == (X86_TRAP_PF_P \| X86_TRAP_PF_RSVD)) /* MMIO page. */
6768	&& !CPUMIsGuestInSvmNestedHwVirtMode(pCtx)
6769	&& !CPUMIsGuestInLongModeEx(pCtx)
6770	&& !CPUMGetGuestCPL(pVCpu)
6771	&& pVM->hm.s.cPatches < RT_ELEMENTS(pVM->hm.s.aPatches))
6772	{
6773	RTGCPHYS GCPhysApicBase = APICGetBaseMsrNoCheck(pVCpu);
6774	GCPhysApicBase &= PAGE_BASE_GC_MASK;
6775
6776	if (GCPhysFaultAddr == GCPhysApicBase + XAPIC_OFF_TPR)
6777	{
6778	/* Only attempt to patch the instruction once. */
6779	PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
6780	if (!pPatch)
6781	return VINF_EM_HM_PATCH_TPR_INSTR;
6782	}
6783	}
6784
6785	/*
6786	* Determine the nested paging mode.
6787	*/
6788	/** @todo r=bird: Gotta love this nested paging hacking we're still carrying with us... (Split PGM_TYPE_NESTED.) */
6789	PGMMODE const enmNestedPagingMode = PGMGetHostMode(pVM);
6790
6791	/*
6792	* MMIO optimization using the reserved (RSVD) bit in the guest page tables for MMIO pages.
6793	*/
6794	Assert((u32ErrCode & (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P)) != X86_TRAP_PF_RSVD);
6795	if ((u32ErrCode & (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P)) == (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P))
6796	{
6797	/*
6798	* If event delivery causes an MMIO #NPF, go back to instruction emulation as otherwise
6799	* injecting the original pending event would most likely cause the same MMIO #NPF.
6800	*/
6801	if (pVCpu->hm.s.Event.fPending)
6802	{
6803	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectInterpret);
6804	return VINF_EM_RAW_INJECT_TRPM_EVENT;
6805	}
6806
6807	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CS \| CPUMCTX_EXTRN_RIP);
6808	VBOXSTRICTRC rcStrict;
6809	PCEMEXITREC pExitRec = EMHistoryUpdateFlagsAndTypeAndPC(pVCpu,
6810	EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_MMIO),
6811	pVCpu->cpum.GstCtx.rip + pVCpu->cpum.GstCtx.cs.u64Base);
6812	if (!pExitRec)
6813	{
6814
6815	rcStrict = PGMR0Trap0eHandlerNPMisconfig(pVM, pVCpu, enmNestedPagingMode, CPUMCTX2CORE(pCtx), GCPhysFaultAddr,
6816	u32ErrCode);
6817
6818	/*
6819	* If we succeed, resume guest execution.
6820	*
6821	* If we fail in interpreting the instruction because we couldn't get the guest
6822	* physical address of the page containing the instruction via the guest's page
6823	* tables (we would invalidate the guest page in the host TLB), resume execution
6824	* which would cause a guest page fault to let the guest handle this weird case.
6825	*
6826	* See @bugref{6043}.
6827	*/
6828	if ( rcStrict == VINF_SUCCESS
6829	\|\| rcStrict == VERR_PAGE_TABLE_NOT_PRESENT
6830	\|\| rcStrict == VERR_PAGE_NOT_PRESENT)
6831	{
6832	/* Successfully handled MMIO operation. */
6833	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
6834	rcStrict = VINF_SUCCESS;
6835	}
6836	}
6837	else
6838	{
6839	/*
6840	* Frequent exit or something needing probing. Get state and call EMHistoryExec.
6841	*/
6842	Assert(pCtx == &pVCpu->cpum.GstCtx);
6843	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6844	Log4(("EptMisscfgExit/%u: %04x:%08RX64: %RGp -> EMHistoryExec\n",
6845	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip, GCPhysFaultAddr));
6846
6847	rcStrict = EMHistoryExec(pVCpu, pExitRec, 0);
6848	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
6849
6850	Log4(("EptMisscfgExit/%u: %04x:%08RX64: EMHistoryExec -> %Rrc + %04x:%08RX64\n",
6851	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip,
6852	VBOXSTRICTRC_VAL(rcStrict), pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip));
6853	}
6854	return VBOXSTRICTRC_TODO(rcStrict);
6855	}
6856
6857	/*
6858	* Nested page-fault.
6859	*/
6860	TRPMAssertXcptPF(pVCpu, GCPhysFaultAddr, u32ErrCode);
6861	int rc = PGMR0Trap0eHandlerNestedPaging(pVM, pVCpu, enmNestedPagingMode, u32ErrCode, CPUMCTX2CORE(pCtx), GCPhysFaultAddr);
6862	TRPMResetTrap(pVCpu);
6863
6864	Log4Func(("#NPF: PGMR0Trap0eHandlerNestedPaging returns %Rrc CS:RIP=%04x:%#RX64\n", rc, pCtx->cs.Sel, pCtx->rip));
6865
6866	/*
6867	* Same case as PGMR0Trap0eHandlerNPMisconfig(). See comment above, @bugref{6043}.
6868	*/
6869	if ( rc == VINF_SUCCESS
6870	\|\| rc == VERR_PAGE_TABLE_NOT_PRESENT
6871	\|\| rc == VERR_PAGE_NOT_PRESENT)
6872	{
6873	/* We've successfully synced our shadow page tables. */
6874	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPF);
6875	rc = VINF_SUCCESS;
6876	}
6877
6878	/*
6879	* If delivering an event causes an #NPF (and not MMIO), we shall resolve the fault and
6880	* re-inject the original event.
6881	*/
6882	if (pVCpu->hm.s.Event.fPending)
6883	{
6884	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectReflectNPF);
6885
6886	/*
6887	* If the #NPF handler requested emulation of the instruction, ignore it.
6888	* We need to re-inject the original event so as to not lose it.
6889	* Reproducible when booting ReactOS 0.4.12 with BTRFS (installed using BootCD,
6890	* LiveCD is broken for other reasons).
6891	*/
6892	if (rc == VINF_EM_RAW_EMULATE_INSTR)
6893	rc = VINF_EM_RAW_INJECT_TRPM_EVENT;
6894	}
6895
6896	return rc;
6897	}
6898
6899
6900	/**
6901	* \#VMEXIT handler for virtual interrupt (SVM_EXIT_VINTR). Conditional
6902	* \#VMEXIT.
6903	*/
6904	HMSVM_EXIT_DECL hmR0SvmExitVIntr(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
6905	{
6906	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6907	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(&pVCpu->cpum.GstCtx);
6908
6909	/* Indicate that we no longer need to #VMEXIT when the guest is ready to receive NMIs, it is now ready. */
6910	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
6911	hmR0SvmClearIntWindowExiting(pVCpu, pVmcb);
6912
6913	/* Deliver the pending interrupt via hmR0SvmEvaluatePendingEvent() and resume guest execution. */
6914	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIntWindow);
6915	return VINF_SUCCESS;
6916	}
6917
6918
6919	/**
6920	* \#VMEXIT handler for task switches (SVM_EXIT_TASK_SWITCH). Conditional
6921	* \#VMEXIT.
6922	*/
6923	HMSVM_EXIT_DECL hmR0SvmExitTaskSwitch(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
6924	{
6925	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6926	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(pVCpu, pSvmTransient);
6927
6928	#ifndef HMSVM_ALWAYS_TRAP_TASK_SWITCH
6929	Assert(!pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
6930	#endif
6931
6932	/* Check if this task-switch occurred while delivering an event through the guest IDT. */
6933	if (pVCpu->hm.s.Event.fPending) /* Can happen with exceptions/NMI. See @bugref{8411}. */
6934	{
6935	/*
6936	* AMD-V provides us with the exception which caused the TS; we collect
6937	* the information in the call to hmR0SvmCheckExitDueToEventDelivery().
6938	*/
6939	Log4Func(("TS occurred during event delivery\n"));
6940	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitTaskSwitch);
6941	return VINF_EM_RAW_INJECT_TRPM_EVENT;
6942	}
6943
6944	/** @todo Emulate task switch someday, currently just going back to ring-3 for
6945	* emulation. */
6946	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitTaskSwitch);
6947	return VERR_EM_INTERPRETER;
6948	}
6949
6950
6951	/**
6952	* \#VMEXIT handler for VMMCALL (SVM_EXIT_VMMCALL). Conditional \#VMEXIT.
6953	*/
6954	HMSVM_EXIT_DECL hmR0SvmExitVmmCall(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
6955	{
6956	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6957	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6958
6959	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
6960	if (pVM->hm.s.fTprPatchingAllowed)
6961	{
6962	int rc = hmEmulateSvmMovTpr(pVM, pVCpu);
6963	if (rc != VERR_NOT_FOUND)
6964	{
6965	Log4Func(("hmEmulateSvmMovTpr returns %Rrc\n", rc));
6966	return rc;
6967	}
6968	}
6969
6970	if (EMAreHypercallInstructionsEnabled(pVCpu))
6971	{
6972	unsigned cbInstr;
6973	if (hmR0SvmSupportsNextRipSave(pVCpu))
6974	{
6975	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
6976	cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
6977	}
6978	else
6979	{
6980	PDISCPUSTATE pDis = &pVCpu->hm.s.DisState;
6981	int rc = EMInterpretDisasCurrent(pVCpu->CTX_SUFF(pVM), pVCpu, pDis, &cbInstr);
6982	if ( rc == VINF_SUCCESS
6983	&& pDis->pCurInstr->uOpcode == OP_VMMCALL)
6984	Assert(cbInstr > 0);
6985	else
6986	cbInstr = 0;
6987	}
6988
6989	VBOXSTRICTRC rcStrict = GIMHypercall(pVCpu, &pVCpu->cpum.GstCtx);
6990	if (RT_SUCCESS(rcStrict))
6991	{
6992	/* Only update the RIP if we're continuing guest execution and not in the case
6993	of say VINF_GIM_R3_HYPERCALL. */
6994	if (rcStrict == VINF_SUCCESS)
6995	hmR0SvmAdvanceRip(pVCpu, cbInstr);
6996
6997	return VBOXSTRICTRC_VAL(rcStrict);
6998	}
6999	else
7000	Log4Func(("GIMHypercall returns %Rrc -> #UD\n", VBOXSTRICTRC_VAL(rcStrict)));
7001	}
7002
7003	hmR0SvmSetPendingXcptUD(pVCpu);
7004	return VINF_SUCCESS;
7005	}
7006
7007
7008	/**
7009	* \#VMEXIT handler for VMMCALL (SVM_EXIT_VMMCALL). Conditional \#VMEXIT.
7010	*/
7011	HMSVM_EXIT_DECL hmR0SvmExitPause(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7012	{
7013	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7014
7015	unsigned cbInstr;
7016	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
7017	if (fSupportsNextRipSave)
7018	{
7019	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7020	cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
7021	}
7022	else
7023	{
7024	PDISCPUSTATE pDis = &pVCpu->hm.s.DisState;
7025	int rc = EMInterpretDisasCurrent(pVCpu->CTX_SUFF(pVM), pVCpu, pDis, &cbInstr);
7026	if ( rc == VINF_SUCCESS
7027	&& pDis->pCurInstr->uOpcode == OP_PAUSE)
7028	Assert(cbInstr > 0);
7029	else
7030	cbInstr = 0;
7031	}
7032
7033	/** @todo The guest has likely hit a contended spinlock. We might want to
7034	* poke a schedule different guest VCPU. */
7035	hmR0SvmAdvanceRip(pVCpu, cbInstr);
7036	return VINF_EM_RAW_INTERRUPT;
7037	}
7038
7039
7040	/**
7041	* \#VMEXIT handler for FERR intercept (SVM_EXIT_FERR_FREEZE). Conditional
7042	* \#VMEXIT.
7043	*/
7044	HMSVM_EXIT_DECL hmR0SvmExitFerrFreeze(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7045	{
7046	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7047	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR0);
7048	Assert(!(pVCpu->cpum.GstCtx.cr0 & X86_CR0_NE));
7049
7050	Log4Func(("Raising IRQ 13 in response to #FERR\n"));
7051	return PDMIsaSetIrq(pVCpu->CTX_SUFF(pVM), 13 /* u8Irq /, 1 / u8Level /, 0 / uTagSrc */);
7052	}
7053
7054
7055	/**
7056	* \#VMEXIT handler for IRET (SVM_EXIT_IRET). Conditional \#VMEXIT.
7057	*/
7058	HMSVM_EXIT_DECL hmR0SvmExitIret(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7059	{
7060	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7061
7062	/* Indicate that we no longer need to #VMEXIT when the guest is ready to receive NMIs, it is now (almost) ready. */
7063	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7064	hmR0SvmClearCtrlIntercept(pVCpu, pVmcb, SVM_CTRL_INTERCEPT_IRET);
7065
7066	/* Emulate the IRET. We have to execute the IRET before an NMI, but must potentially
7067	* deliver a pending NMI right after. If the IRET faults, an NMI can come before the
7068	* handler executes. Yes, x86 is ugly.
7069	*/
7070	return VINF_EM_RAW_EMULATE_INSTR;
7071	}
7072
7073
7074	/**
7075	* \#VMEXIT handler for page-fault exceptions (SVM_EXIT_XCPT_14).
7076	* Conditional \#VMEXIT.
7077	*/
7078	HMSVM_EXIT_DECL hmR0SvmExitXcptPF(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7079	{
7080	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7081	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7082	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(pVCpu, pSvmTransient);
7083
7084	/* See AMD spec. 15.12.15 "#PF (Page Fault)". */
7085	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
7086	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
7087	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7088	uint32_t uErrCode = pVmcb->ctrl.u64ExitInfo1;
7089	uint64_t const uFaultAddress = pVmcb->ctrl.u64ExitInfo2;
7090
7091	#if defined(HMSVM_ALWAYS_TRAP_ALL_XCPTS) \|\| defined(HMSVM_ALWAYS_TRAP_PF)
7092	if (pVM->hm.s.fNestedPaging)
7093	{
7094	pVCpu->hm.s.Event.fPending = false; /* In case it's a contributory or vectoring #PF. */
7095	if ( !pSvmTransient->fVectoringDoublePF
7096	\|\| CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
7097	{
7098	/* A genuine guest #PF, reflect it to the guest. */
7099	hmR0SvmSetPendingXcptPF(pVCpu, uErrCode, uFaultAddress);
7100	Log4Func(("#PF: Guest page fault at %04X:%RGv FaultAddr=%RX64 ErrCode=%#x\n", pCtx->cs.Sel, (RTGCPTR)pCtx->rip,
7101	uFaultAddress, uErrCode));
7102	}
7103	else
7104	{
7105	/* A guest page-fault occurred during delivery of a page-fault. Inject #DF. */
7106	hmR0SvmSetPendingXcptDF(pVCpu);
7107	Log4Func(("Pending #DF due to vectoring #PF. NP\n"));
7108	}
7109	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestPF);
7110	return VINF_SUCCESS;
7111	}
7112	#endif
7113
7114	Assert(!pVM->hm.s.fNestedPaging);
7115
7116	/*
7117	* TPR patching shortcut for APIC TPR reads and writes; only applicable to 32-bit guests.
7118	*/
7119	if ( pVM->hm.s.fTprPatchingAllowed
7120	&& (uFaultAddress & 0xfff) == XAPIC_OFF_TPR
7121	&& !(uErrCode & X86_TRAP_PF_P) /* Not present. */
7122	&& !CPUMIsGuestInSvmNestedHwVirtMode(pCtx)
7123	&& !CPUMIsGuestInLongModeEx(pCtx)
7124	&& !CPUMGetGuestCPL(pVCpu)
7125	&& pVM->hm.s.cPatches < RT_ELEMENTS(pVM->hm.s.aPatches))
7126	{
7127	RTGCPHYS GCPhysApicBase;
7128	GCPhysApicBase = APICGetBaseMsrNoCheck(pVCpu);
7129	GCPhysApicBase &= PAGE_BASE_GC_MASK;
7130
7131	/* Check if the page at the fault-address is the APIC base. */
7132	RTGCPHYS GCPhysPage;
7133	int rc2 = PGMGstGetPage(pVCpu, (RTGCPTR)uFaultAddress, NULL /* pfFlags */, &GCPhysPage);
7134	if ( rc2 == VINF_SUCCESS
7135	&& GCPhysPage == GCPhysApicBase)
7136	{
7137	/* Only attempt to patch the instruction once. */
7138	PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
7139	if (!pPatch)
7140	return VINF_EM_HM_PATCH_TPR_INSTR;
7141	}
7142	}
7143
7144	Log4Func(("#PF: uFaultAddress=%#RX64 CS:RIP=%#04x:%#RX64 uErrCode %#RX32 cr3=%#RX64\n", uFaultAddress, pCtx->cs.Sel,
7145	pCtx->rip, uErrCode, pCtx->cr3));
7146
7147	/*
7148	* If it's a vectoring #PF, emulate injecting the original event injection as
7149	* PGMTrap0eHandler() is incapable of differentiating between instruction emulation and
7150	* event injection that caused a #PF. See @bugref{6607}.
7151	*/
7152	if (pSvmTransient->fVectoringPF)
7153	{
7154	Assert(pVCpu->hm.s.Event.fPending);
7155	return VINF_EM_RAW_INJECT_TRPM_EVENT;
7156	}
7157
7158	TRPMAssertXcptPF(pVCpu, uFaultAddress, uErrCode);
7159	int rc = PGMTrap0eHandler(pVCpu, uErrCode, CPUMCTX2CORE(pCtx), (RTGCPTR)uFaultAddress);
7160
7161	Log4Func(("#PF: rc=%Rrc\n", rc));
7162
7163	if (rc == VINF_SUCCESS)
7164	{
7165	/* Successfully synced shadow pages tables or emulated an MMIO instruction. */
7166	TRPMResetTrap(pVCpu);
7167	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPF);
7168	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
7169	return rc;
7170	}
7171
7172	if (rc == VINF_EM_RAW_GUEST_TRAP)
7173	{
7174	pVCpu->hm.s.Event.fPending = false; /* In case it's a contributory or vectoring #PF. */
7175
7176	/*
7177	* If a nested-guest delivers a #PF and that causes a #PF which is -not- a shadow #PF,
7178	* we should simply forward the #PF to the guest and is up to the nested-hypervisor to
7179	* determine whether it is a nested-shadow #PF or a #DF, see @bugref{7243#c121}.
7180	*/
7181	if ( !pSvmTransient->fVectoringDoublePF
7182	\|\| CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
7183	{
7184	/* It's a guest (or nested-guest) page fault and needs to be reflected. */
7185	uErrCode = TRPMGetErrorCode(pVCpu); /* The error code might have been changed. */
7186	TRPMResetTrap(pVCpu);
7187
7188	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
7189	/* If the nested-guest is intercepting #PFs, cause a #PF #VMEXIT. */
7190	if ( CPUMIsGuestInSvmNestedHwVirtMode(pCtx)
7191	&& CPUMIsGuestSvmXcptInterceptSet(pVCpu, pCtx, X86_XCPT_PF))
7192	return VBOXSTRICTRC_TODO(IEMExecSvmVmexit(pVCpu, SVM_EXIT_XCPT_PF, uErrCode, uFaultAddress));
7193	#endif
7194
7195	hmR0SvmSetPendingXcptPF(pVCpu, uErrCode, uFaultAddress);
7196	}
7197	else
7198	{
7199	/* A guest page-fault occurred during delivery of a page-fault. Inject #DF. */
7200	TRPMResetTrap(pVCpu);
7201	hmR0SvmSetPendingXcptDF(pVCpu);
7202	Log4Func(("#PF: Pending #DF due to vectoring #PF\n"));
7203	}
7204
7205	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestPF);
7206	return VINF_SUCCESS;
7207	}
7208
7209	TRPMResetTrap(pVCpu);
7210	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPFEM);
7211	return rc;
7212	}
7213
7214
7215	/**
7216	* \#VMEXIT handler for undefined opcode (SVM_EXIT_XCPT_6).
7217	* Conditional \#VMEXIT.
7218	*/
7219	HMSVM_EXIT_DECL hmR0SvmExitXcptUD(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7220	{
7221	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7222	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(&pVCpu->cpum.GstCtx);
7223	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestUD);
7224
7225	/* Paranoia; Ensure we cannot be called as a result of event delivery. */
7226	PSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
7227	Assert(!pVmcb->ctrl.ExitIntInfo.n.u1Valid); NOREF(pVmcb);
7228
7229	int rc = VERR_SVM_UNEXPECTED_XCPT_EXIT;
7230	if (pVCpu->hm.s.fGIMTrapXcptUD)
7231	{
7232	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7233	uint8_t cbInstr = 0;
7234	VBOXSTRICTRC rcStrict = GIMXcptUD(pVCpu, &pVCpu->cpum.GstCtx, NULL /* pDis */, &cbInstr);
7235	if (rcStrict == VINF_SUCCESS)
7236	{
7237	/* #UD #VMEXIT does not have valid NRIP information, manually advance RIP. See @bugref{7270#c170}. */
7238	hmR0SvmAdvanceRip(pVCpu, cbInstr);
7239	rc = VINF_SUCCESS;
7240	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
7241	}
7242	else if (rcStrict == VINF_GIM_HYPERCALL_CONTINUING)
7243	rc = VINF_SUCCESS;
7244	else if (rcStrict == VINF_GIM_R3_HYPERCALL)
7245	rc = VINF_GIM_R3_HYPERCALL;
7246	else
7247	Assert(RT_FAILURE(VBOXSTRICTRC_VAL(rcStrict)));
7248	}
7249
7250	/* If the GIM #UD exception handler didn't succeed for some reason or wasn't needed, raise #UD. */
7251	if (RT_FAILURE(rc))
7252	{
7253	hmR0SvmSetPendingXcptUD(pVCpu);
7254	rc = VINF_SUCCESS;
7255	}
7256
7257	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestUD);
7258	return rc;
7259	}
7260
7261
7262	/**
7263	* \#VMEXIT handler for math-fault exceptions (SVM_EXIT_XCPT_16).
7264	* Conditional \#VMEXIT.
7265	*/
7266	HMSVM_EXIT_DECL hmR0SvmExitXcptMF(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7267	{
7268	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7269	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7270	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestMF);
7271
7272	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
7273	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7274
7275	/* Paranoia; Ensure we cannot be called as a result of event delivery. */
7276	Assert(!pVmcb->ctrl.ExitIntInfo.n.u1Valid); NOREF(pVmcb);
7277
7278	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestMF);
7279
7280	if (!(pCtx->cr0 & X86_CR0_NE))
7281	{
7282	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
7283	PDISSTATE pDis = &pVCpu->hm.s.DisState;
7284	unsigned cbInstr;
7285	int rc = EMInterpretDisasCurrent(pVM, pVCpu, pDis, &cbInstr);
7286	if (RT_SUCCESS(rc))
7287	{
7288	/* Convert a #MF into a FERR -> IRQ 13. See @bugref{6117}. */
7289	rc = PDMIsaSetIrq(pVCpu->CTX_SUFF(pVM), 13 /* u8Irq /, 1 / u8Level /, 0 / uTagSrc */);
7290	if (RT_SUCCESS(rc))
7291	hmR0SvmAdvanceRip(pVCpu, cbInstr);
7292	}
7293	else
7294	Log4Func(("EMInterpretDisasCurrent returned %Rrc uOpCode=%#x\n", rc, pDis->pCurInstr->uOpcode));
7295	return rc;
7296	}
7297
7298	hmR0SvmSetPendingXcptMF(pVCpu);
7299	return VINF_SUCCESS;
7300	}
7301
7302
7303	/**
7304	* \#VMEXIT handler for debug exceptions (SVM_EXIT_XCPT_1). Conditional
7305	* \#VMEXIT.
7306	*/
7307	HMSVM_EXIT_DECL hmR0SvmExitXcptDB(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7308	{
7309	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7310	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7311	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(pVCpu, pSvmTransient);
7312	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestDB);
7313
7314	if (RT_UNLIKELY(pVCpu->hm.s.Event.fPending))
7315	{
7316	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectInterpret);
7317	return VINF_EM_RAW_INJECT_TRPM_EVENT;
7318	}
7319
7320	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestDB);
7321
7322	/*
7323	* This can be a fault-type #DB (instruction breakpoint) or a trap-type #DB (data
7324	* breakpoint). However, for both cases DR6 and DR7 are updated to what the exception
7325	* handler expects. See AMD spec. 15.12.2 "#DB (Debug)".
7326	*/
7327	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
7328	PSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
7329	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
7330	int rc = DBGFRZTrap01Handler(pVM, pVCpu, CPUMCTX2CORE(pCtx), pVmcb->guest.u64DR6, pVCpu->hm.s.fSingleInstruction);
7331	if (rc == VINF_EM_RAW_GUEST_TRAP)
7332	{
7333	Log5(("hmR0SvmExitXcptDB: DR6=%#RX64 -> guest trap\n", pVmcb->guest.u64DR6));
7334	if (CPUMIsHyperDebugStateActive(pVCpu))
7335	CPUMSetGuestDR6(pVCpu, CPUMGetGuestDR6(pVCpu) \| pVmcb->guest.u64DR6);
7336
7337	/* Reflect the exception back to the guest. */
7338	hmR0SvmSetPendingXcptDB(pVCpu);
7339	rc = VINF_SUCCESS;
7340	}
7341
7342	/*
7343	* Update DR6.
7344	*/
7345	if (CPUMIsHyperDebugStateActive(pVCpu))
7346	{
7347	Log5(("hmR0SvmExitXcptDB: DR6=%#RX64 -> %Rrc\n", pVmcb->guest.u64DR6, rc));
7348	pVmcb->guest.u64DR6 = X86_DR6_INIT_VAL;
7349	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
7350	}
7351	else
7352	{
7353	AssertMsg(rc == VINF_SUCCESS, ("rc=%Rrc\n", rc));
7354	Assert(!pVCpu->hm.s.fSingleInstruction && !DBGFIsStepping(pVCpu));
7355	}
7356
7357	return rc;
7358	}
7359
7360
7361	/**
7362	* \#VMEXIT handler for alignment check exceptions (SVM_EXIT_XCPT_17).
7363	* Conditional \#VMEXIT.
7364	*/
7365	HMSVM_EXIT_DECL hmR0SvmExitXcptAC(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7366	{
7367	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7368	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(pVCpu, pSvmTransient);
7369	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestAC);
7370
7371	SVMEVENT Event;
7372	Event.u = 0;
7373	Event.n.u1Valid = 1;
7374	Event.n.u3Type = SVM_EVENT_EXCEPTION;
7375	Event.n.u8Vector = X86_XCPT_AC;
7376	Event.n.u1ErrorCodeValid = 1;
7377	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
7378	return VINF_SUCCESS;
7379	}
7380
7381
7382	/**
7383	* \#VMEXIT handler for breakpoint exceptions (SVM_EXIT_XCPT_3).
7384	* Conditional \#VMEXIT.
7385	*/
7386	HMSVM_EXIT_DECL hmR0SvmExitXcptBP(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7387	{
7388	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7389	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7390	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(pVCpu, pSvmTransient);
7391	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestBP);
7392
7393	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
7394	int rc = DBGFRZTrap03Handler(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
7395	if (rc == VINF_EM_RAW_GUEST_TRAP)
7396	{
7397	SVMEVENT Event;
7398	Event.u = 0;
7399	Event.n.u1Valid = 1;
7400	Event.n.u3Type = SVM_EVENT_EXCEPTION;
7401	Event.n.u8Vector = X86_XCPT_BP;
7402	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
7403	rc = VINF_SUCCESS;
7404	}
7405
7406	Assert(rc == VINF_SUCCESS \|\| rc == VINF_EM_DBG_BREAKPOINT);
7407	return rc;
7408	}
7409
7410
7411	/**
7412	* Hacks its way around the lovely mesa driver's backdoor accesses.
7413	*
7414	* @sa hmR0VmxHandleMesaDrvGp
7415	*/
7416	static int hmR0SvmHandleMesaDrvGp(PVMCPUCC pVCpu, PCPUMCTX pCtx, PCSVMVMCB pVmcb)
7417	{
7418	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CS \| CPUMCTX_EXTRN_RIP \| CPUMCTX_EXTRN_RFLAGS \| CPUMCTX_EXTRN_GPRS_MASK);
7419	Log(("hmR0SvmHandleMesaDrvGp: at %04x:%08RX64 rcx=%RX64 rbx=%RX64\n",
7420	pVmcb->guest.CS.u16Sel, pVmcb->guest.u64RIP, pCtx->rcx, pCtx->rbx));
7421	RT_NOREF(pCtx, pVmcb);
7422
7423	/* For now we'll just skip the instruction. */
7424	hmR0SvmAdvanceRip(pVCpu, 1);
7425	return VINF_SUCCESS;
7426	}
7427
7428
7429	/**
7430	* Checks if the \#GP'ing instruction is the mesa driver doing it's lovely
7431	* backdoor logging w/o checking what it is running inside.
7432	*
7433	* This recognizes an "IN EAX,DX" instruction executed in flat ring-3, with the
7434	* backdoor port and magic numbers loaded in registers.
7435	*
7436	* @returns true if it is, false if it isn't.
7437	* @sa hmR0VmxIsMesaDrvGp
7438	*/
7439	DECLINLINE(bool) hmR0SvmIsMesaDrvGp(PVMCPUCC pVCpu, PCPUMCTX pCtx, PCSVMVMCB pVmcb)
7440	{
7441	/* Check magic and port. */
7442	Assert(!(pCtx->fExtrn & (CPUMCTX_EXTRN_RDX \| CPUMCTX_EXTRN_RCX)));
7443	/Log8(("hmR0SvmIsMesaDrvGp: rax=%RX64 rdx=%RX64\n", pCtx->fExtrn & CPUMCTX_EXTRN_RAX ? pVmcb->guest.u64RAX : pCtx->rax, pCtx->rdx));/
7444	if (pCtx->dx != UINT32_C(0x5658))
7445	return false;
7446	if ((pCtx->fExtrn & CPUMCTX_EXTRN_RAX ? pVmcb->guest.u64RAX : pCtx->rax) != UINT32_C(0x564d5868))
7447	return false;
7448
7449	/* Check that it is #GP(0). */
7450	if (pVmcb->ctrl.u64ExitInfo1 != 0)
7451	return false;
7452
7453	/* Flat ring-3 CS. */
7454	/Log8(("hmR0SvmIsMesaDrvGp: u8CPL=%d base=%RX64\n", pVmcb->guest.u8CPL, pCtx->fExtrn & CPUMCTX_EXTRN_CS ? pVmcb->guest.CS.u64Base : pCtx->cs.u64Base));/
7455	if (pVmcb->guest.u8CPL != 3)
7456	return false;
7457	if ((pCtx->fExtrn & CPUMCTX_EXTRN_CS ? pVmcb->guest.CS.u64Base : pCtx->cs.u64Base) != 0)
7458	return false;
7459
7460	/* 0xed: IN eAX,dx */
7461	if (pVmcb->ctrl.cbInstrFetched < 1) /* unlikely, it turns out. */
7462	{
7463	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CS \| CPUMCTX_EXTRN_RIP \| CPUMCTX_EXTRN_GPRS_MASK
7464	\| CPUMCTX_EXTRN_CR0 \| CPUMCTX_EXTRN_CR3 \| CPUMCTX_EXTRN_CR4 \| CPUMCTX_EXTRN_EFER);
7465	uint8_t abInstr[1];
7466	int rc = PGMPhysSimpleReadGCPtr(pVCpu, abInstr, pCtx->rip, sizeof(abInstr));
7467	/Log8(("hmR0SvmIsMesaDrvGp: PGMPhysSimpleReadGCPtr -> %Rrc %#x\n", rc, abInstr[0])); /
7468	if (RT_FAILURE(rc))
7469	return false;
7470	if (abInstr[0] != 0xed)
7471	return false;
7472	}
7473	else
7474	{
7475	/Log8(("hmR0SvmIsMesaDrvGp: %#x\n", pVmcb->ctrl.abInstr));/
7476	if (pVmcb->ctrl.abInstr[0] != 0xed)
7477	return false;
7478	}
7479	return true;
7480	}
7481
7482
7483	/**
7484	* \#VMEXIT handler for general protection faults (SVM_EXIT_XCPT_BP).
7485	* Conditional \#VMEXIT.
7486	*/
7487	HMSVM_EXIT_DECL hmR0SvmExitXcptGP(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7488	{
7489	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7490	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(pVCpu, pSvmTransient);
7491	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestGP);
7492
7493	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7494	Assert(pSvmTransient->u64ExitCode == pVmcb->ctrl.u64ExitCode);
7495
7496	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
7497	if ( !pVCpu->hm.s.fTrapXcptGpForLovelyMesaDrv
7498	\|\| !hmR0SvmIsMesaDrvGp(pVCpu, pCtx, pVmcb))
7499	{
7500	SVMEVENT Event;
7501	Event.u = 0;
7502	Event.n.u1Valid = 1;
7503	Event.n.u3Type = SVM_EVENT_EXCEPTION;
7504	Event.n.u8Vector = X86_XCPT_GP;
7505	Event.n.u1ErrorCodeValid = 1;
7506	Event.n.u32ErrorCode = (uint32_t)pVmcb->ctrl.u64ExitInfo1;
7507	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
7508	return VINF_SUCCESS;
7509	}
7510	return hmR0SvmHandleMesaDrvGp(pVCpu, pCtx, pVmcb);
7511	}
7512
7513
7514	#if defined(HMSVM_ALWAYS_TRAP_ALL_XCPTS) \|\| defined(VBOX_WITH_NESTED_HWVIRT_SVM)
7515	/**
7516	* \#VMEXIT handler for generic exceptions. Conditional \#VMEXIT.
7517	*/
7518	HMSVM_EXIT_DECL hmR0SvmExitXcptGeneric(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7519	{
7520	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7521	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(pVCpu, pSvmTransient);
7522
7523	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7524	uint8_t const uVector = pVmcb->ctrl.u64ExitCode - SVM_EXIT_XCPT_0;
7525	uint32_t const uErrCode = pVmcb->ctrl.u64ExitInfo1;
7526	Assert(pSvmTransient->u64ExitCode == pVmcb->ctrl.u64ExitCode);
7527	Assert(uVector <= X86_XCPT_LAST);
7528	Log4Func(("uVector=%#x uErrCode=%u\n", uVector, uErrCode));
7529
7530	SVMEVENT Event;
7531	Event.u = 0;
7532	Event.n.u1Valid = 1;
7533	Event.n.u3Type = SVM_EVENT_EXCEPTION;
7534	Event.n.u8Vector = uVector;
7535	switch (uVector)
7536	{
7537	/* Shouldn't be here for reflecting #PFs (among other things, the fault address isn't passed along). */
7538	case X86_XCPT_PF: AssertMsgFailed(("hmR0SvmExitXcptGeneric: Unexpected exception")); return VERR_SVM_IPE_5;
7539	case X86_XCPT_DF:
7540	case X86_XCPT_TS:
7541	case X86_XCPT_NP:
7542	case X86_XCPT_SS:
7543	case X86_XCPT_GP:
7544	case X86_XCPT_AC:
7545	{
7546	Event.n.u1ErrorCodeValid = 1;
7547	Event.n.u32ErrorCode = uErrCode;
7548	break;
7549	}
7550	}
7551
7552	#ifdef VBOX_WITH_STATISTICS
7553	switch (uVector)
7554	{
7555	case X86_XCPT_DE: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestDE); break;
7556	case X86_XCPT_DB: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestDB); break;
7557	case X86_XCPT_BP: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestBP); break;
7558	case X86_XCPT_OF: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestOF); break;
7559	case X86_XCPT_BR: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestBR); break;
7560	case X86_XCPT_UD: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestUD); break;
7561	case X86_XCPT_NM: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestOF); break;
7562	case X86_XCPT_DF: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestDF); break;
7563	case X86_XCPT_TS: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestTS); break;
7564	case X86_XCPT_NP: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestNP); break;
7565	case X86_XCPT_SS: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestSS); break;
7566	case X86_XCPT_GP: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestGP); break;
7567	case X86_XCPT_PF: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestPF); break;
7568	case X86_XCPT_MF: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestMF); break;
7569	case X86_XCPT_AC: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestAC); break;
7570	case X86_XCPT_XF: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestXF); break;
7571	default:
7572	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestXcpUnk);
7573	break;
7574	}
7575	#endif
7576
7577	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
7578	return VINF_SUCCESS;
7579	}
7580	#endif
7581
7582	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
7583	/**
7584	* \#VMEXIT handler for CLGI (SVM_EXIT_CLGI). Conditional \#VMEXIT.
7585	*/
7586	HMSVM_EXIT_DECL hmR0SvmExitClgi(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7587	{
7588	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7589
7590	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7591	Assert(pVmcb);
7592	Assert(!pVmcb->ctrl.IntCtrl.n.u1VGifEnable);
7593
7594	VBOXSTRICTRC rcStrict;
7595	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
7596	uint64_t const fImport = CPUMCTX_EXTRN_HWVIRT;
7597	if (fSupportsNextRipSave)
7598	{
7599	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK \| fImport);
7600	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
7601	rcStrict = IEMExecDecodedClgi(pVCpu, cbInstr);
7602	}
7603	else
7604	{
7605	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK \| fImport);
7606	rcStrict = IEMExecOne(pVCpu);
7607	}
7608
7609	if (rcStrict == VINF_SUCCESS)
7610	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_HWVIRT);
7611	else if (rcStrict == VINF_IEM_RAISED_XCPT)
7612	{
7613	rcStrict = VINF_SUCCESS;
7614	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
7615	}
7616	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
7617	return VBOXSTRICTRC_TODO(rcStrict);
7618	}
7619
7620
7621	/**
7622	* \#VMEXIT handler for STGI (SVM_EXIT_STGI). Conditional \#VMEXIT.
7623	*/
7624	HMSVM_EXIT_DECL hmR0SvmExitStgi(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7625	{
7626	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7627
7628	/*
7629	* When VGIF is not used we always intercept STGI instructions. When VGIF is used,
7630	* we only intercept STGI when events are pending for GIF to become 1.
7631	*/
7632	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7633	if (pVmcb->ctrl.IntCtrl.n.u1VGifEnable)
7634	hmR0SvmClearCtrlIntercept(pVCpu, pVmcb, SVM_CTRL_INTERCEPT_STGI);
7635
7636	VBOXSTRICTRC rcStrict;
7637	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
7638	uint64_t const fImport = CPUMCTX_EXTRN_HWVIRT;
7639	if (fSupportsNextRipSave)
7640	{
7641	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK \| fImport);
7642	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
7643	rcStrict = IEMExecDecodedStgi(pVCpu, cbInstr);
7644	}
7645	else
7646	{
7647	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK \| fImport);
7648	rcStrict = IEMExecOne(pVCpu);
7649	}
7650
7651	if (rcStrict == VINF_SUCCESS)
7652	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_HWVIRT);
7653	else if (rcStrict == VINF_IEM_RAISED_XCPT)
7654	{
7655	rcStrict = VINF_SUCCESS;
7656	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
7657	}
7658	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
7659	return VBOXSTRICTRC_TODO(rcStrict);
7660	}
7661
7662
7663	/**
7664	* \#VMEXIT handler for VMLOAD (SVM_EXIT_VMLOAD). Conditional \#VMEXIT.
7665	*/
7666	HMSVM_EXIT_DECL hmR0SvmExitVmload(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7667	{
7668	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7669
7670	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7671	Assert(pVmcb);
7672	Assert(!pVmcb->ctrl.LbrVirt.n.u1VirtVmsaveVmload);
7673
7674	VBOXSTRICTRC rcStrict;
7675	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
7676	uint64_t const fImport = CPUMCTX_EXTRN_FS \| CPUMCTX_EXTRN_GS \| CPUMCTX_EXTRN_KERNEL_GS_BASE
7677	\| CPUMCTX_EXTRN_TR \| CPUMCTX_EXTRN_LDTR \| CPUMCTX_EXTRN_SYSCALL_MSRS
7678	\| CPUMCTX_EXTRN_SYSENTER_MSRS;
7679	if (fSupportsNextRipSave)
7680	{
7681	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK \| fImport);
7682	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
7683	rcStrict = IEMExecDecodedVmload(pVCpu, cbInstr);
7684	}
7685	else
7686	{
7687	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK \| fImport);
7688	rcStrict = IEMExecOne(pVCpu);
7689	}
7690
7691	if (rcStrict == VINF_SUCCESS)
7692	{
7693	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_FS \| HM_CHANGED_GUEST_GS
7694	\| HM_CHANGED_GUEST_TR \| HM_CHANGED_GUEST_LDTR
7695	\| HM_CHANGED_GUEST_KERNEL_GS_BASE \| HM_CHANGED_GUEST_SYSCALL_MSRS
7696	\| HM_CHANGED_GUEST_SYSENTER_MSR_MASK);
7697	}
7698	else if (rcStrict == VINF_IEM_RAISED_XCPT)
7699	{
7700	rcStrict = VINF_SUCCESS;
7701	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
7702	}
7703	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
7704	return VBOXSTRICTRC_TODO(rcStrict);
7705	}
7706
7707
7708	/**
7709	* \#VMEXIT handler for VMSAVE (SVM_EXIT_VMSAVE). Conditional \#VMEXIT.
7710	*/
7711	HMSVM_EXIT_DECL hmR0SvmExitVmsave(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7712	{
7713	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7714
7715	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7716	Assert(!pVmcb->ctrl.LbrVirt.n.u1VirtVmsaveVmload);
7717
7718	VBOXSTRICTRC rcStrict;
7719	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
7720	if (fSupportsNextRipSave)
7721	{
7722	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK);
7723	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
7724	rcStrict = IEMExecDecodedVmsave(pVCpu, cbInstr);
7725	}
7726	else
7727	{
7728	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
7729	rcStrict = IEMExecOne(pVCpu);
7730	}
7731
7732	if (rcStrict == VINF_IEM_RAISED_XCPT)
7733	{
7734	rcStrict = VINF_SUCCESS;
7735	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
7736	}
7737	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
7738	return VBOXSTRICTRC_TODO(rcStrict);
7739	}
7740
7741
7742	/**
7743	* \#VMEXIT handler for INVLPGA (SVM_EXIT_INVLPGA). Conditional \#VMEXIT.
7744	*/
7745	HMSVM_EXIT_DECL hmR0SvmExitInvlpga(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7746	{
7747	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7748
7749	VBOXSTRICTRC rcStrict;
7750	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
7751	if (fSupportsNextRipSave)
7752	{
7753	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK);
7754	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7755	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
7756	rcStrict = IEMExecDecodedInvlpga(pVCpu, cbInstr);
7757	}
7758	else
7759	{
7760	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
7761	rcStrict = IEMExecOne(pVCpu);
7762	}
7763
7764	if (rcStrict == VINF_IEM_RAISED_XCPT)
7765	{
7766	rcStrict = VINF_SUCCESS;
7767	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
7768	}
7769	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
7770	return VBOXSTRICTRC_TODO(rcStrict);
7771	}
7772
7773
7774	/**
7775	* \#VMEXIT handler for STGI (SVM_EXIT_VMRUN). Conditional \#VMEXIT.
7776	*/
7777	HMSVM_EXIT_DECL hmR0SvmExitVmrun(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7778	{
7779	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7780	/* We shall import the entire state here, just in case we enter and continue execution of
7781	the nested-guest with hardware-assisted SVM in ring-0, we would be switching VMCBs and
7782	could lose lose part of CPU state. */
7783	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7784
7785	VBOXSTRICTRC rcStrict;
7786	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
7787	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatExitVmentry, z);
7788	if (fSupportsNextRipSave)
7789	{
7790	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7791	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
7792	rcStrict = IEMExecDecodedVmrun(pVCpu, cbInstr);
7793	}
7794	else
7795	{
7796	/* We use IEMExecOneBypassEx() here as it supresses attempt to continue emulating any
7797	instruction(s) when interrupt inhibition is set as part of emulating the VMRUN
7798	instruction itself, see @bugref{7243#c126} */
7799	rcStrict = IEMExecOneBypassEx(pVCpu, CPUMCTX2CORE(&pVCpu->cpum.GstCtx), NULL /* pcbWritten */);
7800	}
7801	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExitVmentry, z);
7802
7803	if (rcStrict == VINF_SUCCESS)
7804	{
7805	rcStrict = VINF_SVM_VMRUN;
7806	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_SVM_VMRUN_MASK);
7807	}
7808	else if (rcStrict == VINF_IEM_RAISED_XCPT)
7809	{
7810	rcStrict = VINF_SUCCESS;
7811	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
7812	}
7813	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
7814	return VBOXSTRICTRC_TODO(rcStrict);
7815	}
7816
7817
7818	/**
7819	* Nested-guest \#VMEXIT handler for debug exceptions (SVM_EXIT_XCPT_1).
7820	* Unconditional \#VMEXIT.
7821	*/
7822	HMSVM_EXIT_DECL hmR0SvmNestedExitXcptDB(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7823	{
7824	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7825	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(pVCpu, pSvmTransient);
7826
7827	if (pVCpu->hm.s.Event.fPending)
7828	{
7829	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectInterpret);
7830	return VINF_EM_RAW_INJECT_TRPM_EVENT;
7831	}
7832
7833	hmR0SvmSetPendingXcptDB(pVCpu);
7834	return VINF_SUCCESS;
7835	}
7836
7837
7838	/**
7839	* Nested-guest \#VMEXIT handler for breakpoint exceptions (SVM_EXIT_XCPT_3).
7840	* Conditional \#VMEXIT.
7841	*/
7842	HMSVM_EXIT_DECL hmR0SvmNestedExitXcptBP(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7843	{
7844	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7845	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(pVCpu, pSvmTransient);
7846
7847	SVMEVENT Event;
7848	Event.u = 0;
7849	Event.n.u1Valid = 1;
7850	Event.n.u3Type = SVM_EVENT_EXCEPTION;
7851	Event.n.u8Vector = X86_XCPT_BP;
7852	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
7853	return VINF_SUCCESS;
7854	}
7855	#endif /* VBOX_WITH_NESTED_HWVIRT_SVM */
7856
7857	/** @} */
7858

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

source: vbox/trunk/src/VBox/VMM/VMMR0/HMSVMR0.cpp@ 84948

Download in other formats: