HMSVMR0.cpp@ 72783

Last change on this file since 72783 was 72783, checked in by vboxsync, 6 years ago
HM: doxygen fixes
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 318.4 KB

Line
1	/* $Id: HMSVMR0.cpp 72783 2018-07-01 16:33:06Z vboxsync $ */
2	/** @file
3	* HM SVM (AMD-V) - Host Context Ring-0.
4	*/
5
6	/*
7	* Copyright (C) 2013-2017 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/gim.h>
33	#include <VBox/vmm/apic.h>
34	#include "HMInternal.h"
35	#include <VBox/vmm/vm.h>
36	#include "HMSVMR0.h"
37	#include "dtrace/VBoxVMM.h"
38
39	#ifdef DEBUG_ramshankar
40	# define HMSVM_SYNC_FULL_GUEST_STATE
41	# define HMSVM_SYNC_FULL_NESTED_GUEST_STATE
42	# define HMSVM_ALWAYS_TRAP_ALL_XCPTS
43	# define HMSVM_ALWAYS_TRAP_PF
44	# define HMSVM_ALWAYS_TRAP_TASK_SWITCH
45	#endif
46
47
48	/*********************************************************************************************************************************
49	* Defined Constants And Macros *
50	*********************************************************************************************************************************/
51	#ifdef VBOX_WITH_STATISTICS
52	# define HMSVM_EXITCODE_STAM_COUNTER_INC(u64ExitCode) do { \
53	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitAll); \
54	if ((u64ExitCode) == SVM_EXIT_NPF) \
55	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitReasonNpf); \
56	else \
57	STAM_COUNTER_INC(&pVCpu->hm.s.paStatExitReasonR0[(u64ExitCode) & MASK_EXITREASON_STAT]); \
58	} while (0)
59
60	# ifdef VBOX_WITH_NESTED_HWVIRT_SVM
61	# define HMSVM_NESTED_EXITCODE_STAM_COUNTER_INC(u64ExitCode) do { \
62	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitAll); \
63	if ((u64ExitCode) == SVM_EXIT_NPF) \
64	STAM_COUNTER_INC(&pVCpu->hm.s.StatNestedExitReasonNpf); \
65	else \
66	STAM_COUNTER_INC(&pVCpu->hm.s.paStatNestedExitReasonR0[(u64ExitCode) & MASK_EXITREASON_STAT]); \
67	} while (0)
68	# endif
69	#else
70	# define HMSVM_EXITCODE_STAM_COUNTER_INC(u64ExitCode) do { } while (0)
71	# ifdef VBOX_WITH_NESTED_HWVIRT_SVM
72	# define HMSVM_NESTED_EXITCODE_STAM_COUNTER_INC(u64ExitCode) do { } while (0)
73	# endif
74	#endif /* !VBOX_WITH_STATISTICS */
75
76
77	/** If we decide to use a function table approach this can be useful to
78	* switch to a "static DECLCALLBACK(int)". */
79	#define HMSVM_EXIT_DECL static int
80
81	/**
82	* Subset of the guest-CPU state that is kept by SVM R0 code while executing the
83	* guest using hardware-assisted SVM.
84	*
85	* This excludes state like TSC AUX, GPRs (other than RSP, RAX) which are always
86	* are swapped and restored across the world-switch and also registers like
87	* EFER, PAT MSR etc. which cannot be modified by the guest without causing a
88	* \#VMEXIT.
89	*/
90	#define HMSVM_CPUMCTX_EXTRN_ALL ( CPUMCTX_EXTRN_RIP \
91	\| CPUMCTX_EXTRN_RFLAGS \
92	\| CPUMCTX_EXTRN_RAX \
93	\| CPUMCTX_EXTRN_RSP \
94	\| CPUMCTX_EXTRN_SREG_MASK \
95	\| CPUMCTX_EXTRN_CR0 \
96	\| CPUMCTX_EXTRN_CR2 \
97	\| CPUMCTX_EXTRN_CR3 \
98	\| CPUMCTX_EXTRN_TABLE_MASK \
99	\| CPUMCTX_EXTRN_DR6 \
100	\| CPUMCTX_EXTRN_DR7 \
101	\| CPUMCTX_EXTRN_KERNEL_GS_BASE \
102	\| CPUMCTX_EXTRN_SYSCALL_MSRS \
103	\| CPUMCTX_EXTRN_SYSENTER_MSRS \
104	\| CPUMCTX_EXTRN_HWVIRT \
105	\| CPUMCTX_EXTRN_HM_SVM_MASK)
106
107	/**
108	* Subset of the guest-CPU state that is shared between the guest and host.
109	*/
110	#define HMSVM_CPUMCTX_SHARED_STATE CPUMCTX_EXTRN_DR_MASK
111
112	/** Macro for importing guest state from the VMCB back into CPUMCTX. */
113	#define HMSVM_CPUMCTX_IMPORT_STATE(a_pVCpu, a_fWhat) \
114	do { \
115	if ((a_pVCpu)->cpum.GstCtx.fExtrn & (a_fWhat)) \
116	hmR0SvmImportGuestState((a_pVCpu), (a_fWhat)); \
117	} while (0)
118
119	/** Assert that the required state bits are fetched. */
120	#define HMSVM_CPUMCTX_ASSERT(a_pVCpu, a_fExtrnMbz) AssertMsg(!((a_pVCpu)->cpum.GstCtx.fExtrn & (a_fExtrnMbz)), \
121	("fExtrn=%#RX64 fExtrnMbz=%#RX64\n", \
122	(a_pVCpu)->cpum.GstCtx.fExtrn, (a_fExtrnMbz)))
123
124	/** Macro for checking and returning from the using function for
125	* \#VMEXIT intercepts that maybe caused during delivering of another
126	* event in the guest. */
127	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
128	# define HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY() \
129	do \
130	{ \
131	int rc = hmR0SvmCheckExitDueToEventDelivery(pVCpu, pCtx, pSvmTransient); \
132	if (RT_LIKELY(rc == VINF_SUCCESS)) { /* continue #VMEXIT handling */ } \
133	else if ( rc == VINF_HM_DOUBLE_FAULT) { return VINF_SUCCESS; } \
134	else if ( rc == VINF_EM_RESET \
135	&& CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_SHUTDOWN)) \
136	{ \
137	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_SVM_VMEXIT_MASK); \
138	return VBOXSTRICTRC_TODO(IEMExecSvmVmexit(pVCpu, SVM_EXIT_SHUTDOWN, 0, 0)); \
139	} \
140	else \
141	return rc; \
142	} while (0)
143	#else
144	# define HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY() \
145	do \
146	{ \
147	int rc = hmR0SvmCheckExitDueToEventDelivery(pVCpu, pCtx, pSvmTransient); \
148	if (RT_LIKELY(rc == VINF_SUCCESS)) { /* continue #VMEXIT handling */ } \
149	else if ( rc == VINF_HM_DOUBLE_FAULT) { return VINF_SUCCESS; } \
150	else \
151	return rc; \
152	} while (0)
153	#endif
154
155	/** Macro which updates interrupt shadow for the current RIP. */
156	#define HMSVM_UPDATE_INTR_SHADOW(pVCpu, pCtx) \
157	do { \
158	/* Update interrupt shadow. */ \
159	if ( VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS) \
160	&& pCtx->rip != EMGetInhibitInterruptsPC(pVCpu)) \
161	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS); \
162	} while (0)
163
164	/** Macro for upgrading a @a a_rc to VINF_EM_DBG_STEPPED after emulating an
165	* instruction that exited. */
166	#define HMSVM_CHECK_SINGLE_STEP(a_pVCpu, a_rc) \
167	do { \
168	if ((a_pVCpu)->hm.s.fSingleInstruction && (a_rc) == VINF_SUCCESS) \
169	(a_rc) = VINF_EM_DBG_STEPPED; \
170	} while (0)
171
172	/** Assert that preemption is disabled or covered by thread-context hooks. */
173	#define HMSVM_ASSERT_PREEMPT_SAFE() Assert( VMMR0ThreadCtxHookIsEnabled(pVCpu) \
174	\|\| !RTThreadPreemptIsEnabled(NIL_RTTHREAD));
175
176	/** Assert that we haven't migrated CPUs when thread-context hooks are not
177	* used. */
178	#define HMSVM_ASSERT_CPU_SAFE() AssertMsg( VMMR0ThreadCtxHookIsEnabled(pVCpu) \
179	\|\| pVCpu->hm.s.idEnteredCpu == RTMpCpuId(), \
180	("Illegal migration! Entered on CPU %u Current %u\n", \
181	pVCpu->hm.s.idEnteredCpu, RTMpCpuId()));
182
183	/** Assert that we're not executing a nested-guest. */
184	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
185	# define HMSVM_ASSERT_NOT_IN_NESTED_GUEST(a_pCtx) Assert(!CPUMIsGuestInSvmNestedHwVirtMode((a_pCtx)))
186	#else
187	# define HMSVM_ASSERT_NOT_IN_NESTED_GUEST(a_pCtx) do { NOREF((a_pCtx)); } while (0)
188	#endif
189
190	/** Assert that we're executing a nested-guest. */
191	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
192	# define HMSVM_ASSERT_IN_NESTED_GUEST(a_pCtx) Assert(CPUMIsGuestInSvmNestedHwVirtMode((a_pCtx)))
193	#else
194	# define HMSVM_ASSERT_IN_NESTED_GUEST(a_pCtx) do { NOREF((a_pCtx)); } while (0)
195	#endif
196
197	/** Validate segment descriptor granularity bit. */
198	#ifdef VBOX_STRICT
199	# define HMSVM_ASSERT_SEG_GRANULARITY(a_pCtx, reg) \
200	AssertMsg( !(a_pCtx)->reg.Attr.n.u1Present \
201	\|\| ( (a_pCtx)->reg.Attr.n.u1Granularity \
202	? ((a_pCtx)->reg.u32Limit & 0xfff) == 0xfff \
203	: (a_pCtx)->reg.u32Limit <= UINT32_C(0xfffff)), \
204	("Invalid Segment Attributes Limit=%#RX32 Attr=%#RX32 Base=%#RX64\n", (a_pCtx)->reg.u32Limit, \
205	(a_pCtx)->reg.Attr.u, (a_pCtx)->reg.u64Base))
206	#else
207	# define HMSVM_ASSERT_SEG_GRANULARITY(a_pCtx, reg) do { } while (0)
208	#endif
209
210	/**
211	* Exception bitmap mask for all contributory exceptions.
212	*
213	* Page fault is deliberately excluded here as it's conditional as to whether
214	* it's contributory or benign. Page faults are handled separately.
215	*/
216	#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) \
217	\| RT_BIT(X86_XCPT_DE))
218
219	/**
220	* Mandatory/unconditional guest control intercepts.
221	*
222	* SMIs can and do happen in normal operation. We need not intercept them
223	* while executing the guest (or nested-guest).
224	*/
225	#define HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS ( SVM_CTRL_INTERCEPT_INTR \
226	\| SVM_CTRL_INTERCEPT_NMI \
227	\| SVM_CTRL_INTERCEPT_INIT \
228	\| SVM_CTRL_INTERCEPT_RDPMC \
229	\| SVM_CTRL_INTERCEPT_CPUID \
230	\| SVM_CTRL_INTERCEPT_RSM \
231	\| SVM_CTRL_INTERCEPT_HLT \
232	\| SVM_CTRL_INTERCEPT_IOIO_PROT \
233	\| SVM_CTRL_INTERCEPT_MSR_PROT \
234	\| SVM_CTRL_INTERCEPT_INVLPGA \
235	\| SVM_CTRL_INTERCEPT_SHUTDOWN \
236	\| SVM_CTRL_INTERCEPT_FERR_FREEZE \
237	\| SVM_CTRL_INTERCEPT_VMRUN \
238	\| SVM_CTRL_INTERCEPT_SKINIT \
239	\| SVM_CTRL_INTERCEPT_WBINVD \
240	\| SVM_CTRL_INTERCEPT_MONITOR \
241	\| SVM_CTRL_INTERCEPT_MWAIT \
242	\| SVM_CTRL_INTERCEPT_CR0_SEL_WRITE \
243	\| SVM_CTRL_INTERCEPT_XSETBV)
244
245	/** @name VMCB Clean Bits.
246	*
247	* These flags are used for VMCB-state caching. A set VMCB Clean bit indicates
248	* AMD-V doesn't need to reload the corresponding value(s) from the VMCB in
249	* memory.
250	*
251	* @{ */
252	/** All intercepts vectors, TSC offset, PAUSE filter counter. */
253	#define HMSVM_VMCB_CLEAN_INTERCEPTS RT_BIT(0)
254	/** I/O permission bitmap, MSR permission bitmap. */
255	#define HMSVM_VMCB_CLEAN_IOPM_MSRPM RT_BIT(1)
256	/** ASID. */
257	#define HMSVM_VMCB_CLEAN_ASID RT_BIT(2)
258	/** TRP: V_TPR, V_IRQ, V_INTR_PRIO, V_IGN_TPR, V_INTR_MASKING,
259	V_INTR_VECTOR. */
260	#define HMSVM_VMCB_CLEAN_INT_CTRL RT_BIT(3)
261	/** Nested Paging: Nested CR3 (nCR3), PAT. */
262	#define HMSVM_VMCB_CLEAN_NP RT_BIT(4)
263	/** Control registers (CR0, CR3, CR4, EFER). */
264	#define HMSVM_VMCB_CLEAN_CRX_EFER RT_BIT(5)
265	/** Debug registers (DR6, DR7). */
266	#define HMSVM_VMCB_CLEAN_DRX RT_BIT(6)
267	/** GDT, IDT limit and base. */
268	#define HMSVM_VMCB_CLEAN_DT RT_BIT(7)
269	/** Segment register: CS, SS, DS, ES limit and base. */
270	#define HMSVM_VMCB_CLEAN_SEG RT_BIT(8)
271	/** CR2.*/
272	#define HMSVM_VMCB_CLEAN_CR2 RT_BIT(9)
273	/** Last-branch record (DbgCtlMsr, br_from, br_to, lastint_from, lastint_to) */
274	#define HMSVM_VMCB_CLEAN_LBR RT_BIT(10)
275	/** AVIC (AVIC APIC_BAR; AVIC APIC_BACKING_PAGE, AVIC
276	PHYSICAL_TABLE and AVIC LOGICAL_TABLE Pointers). */
277	#define HMSVM_VMCB_CLEAN_AVIC RT_BIT(11)
278	/** Mask of all valid VMCB Clean bits. */
279	#define HMSVM_VMCB_CLEAN_ALL ( HMSVM_VMCB_CLEAN_INTERCEPTS \
280	\| HMSVM_VMCB_CLEAN_IOPM_MSRPM \
281	\| HMSVM_VMCB_CLEAN_ASID \
282	\| HMSVM_VMCB_CLEAN_INT_CTRL \
283	\| HMSVM_VMCB_CLEAN_NP \
284	\| HMSVM_VMCB_CLEAN_CRX_EFER \
285	\| HMSVM_VMCB_CLEAN_DRX \
286	\| HMSVM_VMCB_CLEAN_DT \
287	\| HMSVM_VMCB_CLEAN_SEG \
288	\| HMSVM_VMCB_CLEAN_CR2 \
289	\| HMSVM_VMCB_CLEAN_LBR \
290	\| HMSVM_VMCB_CLEAN_AVIC)
291	/** @} */
292
293	/** @name SVM transient.
294	*
295	* A state structure for holding miscellaneous information across AMD-V
296	* VMRUN/\#VMEXIT operation, restored after the transition.
297	*
298	* @{ */
299	typedef struct SVMTRANSIENT
300	{
301	/** The host's rflags/eflags. */
302	RTCCUINTREG fEFlags;
303	#if HC_ARCH_BITS == 32
304	uint32_t u32Alignment0;
305	#endif
306
307	/** The \#VMEXIT exit code (the EXITCODE field in the VMCB). */
308	uint64_t u64ExitCode;
309	/** The guest's TPR value used for TPR shadowing. */
310	uint8_t u8GuestTpr;
311	/** Alignment. */
312	uint8_t abAlignment0[7];
313
314	/** Pointer to the currently executing VMCB. */
315	PSVMVMCB pVmcb;
316	/** Whether we are currently executing a nested-guest. */
317	bool fIsNestedGuest;
318
319	/** Whether the guest debug state was active at the time of \#VMEXIT. */
320	bool fWasGuestDebugStateActive;
321	/** Whether the hyper debug state was active at the time of \#VMEXIT. */
322	bool fWasHyperDebugStateActive;
323	/** Whether the TSC offset mode needs to be updated. */
324	bool fUpdateTscOffsetting;
325	/** Whether the TSC_AUX MSR needs restoring on \#VMEXIT. */
326	bool fRestoreTscAuxMsr;
327	/** Whether the \#VMEXIT was caused by a page-fault during delivery of a
328	* contributary exception or a page-fault. */
329	bool fVectoringDoublePF;
330	/** Whether the \#VMEXIT was caused by a page-fault during delivery of an
331	* external interrupt or NMI. */
332	bool fVectoringPF;
333	} SVMTRANSIENT, *PSVMTRANSIENT;
334	AssertCompileMemberAlignment(SVMTRANSIENT, u64ExitCode, sizeof(uint64_t));
335	AssertCompileMemberAlignment(SVMTRANSIENT, pVmcb, sizeof(uint64_t));
336	/** @} */
337
338	/**
339	* MSRPM (MSR permission bitmap) read permissions (for guest RDMSR).
340	*/
341	typedef enum SVMMSREXITREAD
342	{
343	/** Reading this MSR causes a \#VMEXIT. */
344	SVMMSREXIT_INTERCEPT_READ = 0xb,
345	/** Reading this MSR does not cause a \#VMEXIT. */
346	SVMMSREXIT_PASSTHRU_READ
347	} SVMMSREXITREAD;
348
349	/**
350	* MSRPM (MSR permission bitmap) write permissions (for guest WRMSR).
351	*/
352	typedef enum SVMMSREXITWRITE
353	{
354	/** Writing to this MSR causes a \#VMEXIT. */
355	SVMMSREXIT_INTERCEPT_WRITE = 0xd,
356	/** Writing to this MSR does not cause a \#VMEXIT. */
357	SVMMSREXIT_PASSTHRU_WRITE
358	} SVMMSREXITWRITE;
359
360	/**
361	* SVM \#VMEXIT handler.
362	*
363	* @returns VBox status code.
364	* @param pVCpu The cross context virtual CPU structure.
365	* @param pCtx Pointer to the guest-CPU context.
366	* @param pSvmTransient Pointer to the SVM-transient structure.
367	*/
368	typedef int FNSVMEXITHANDLER(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
369
370
371	/*********************************************************************************************************************************
372	* Internal Functions *
373	*********************************************************************************************************************************/
374	static void hmR0SvmPendingEventToTrpmTrap(PVMCPU pVCpu);
375	static void hmR0SvmLeave(PVMCPU pVCpu, bool fImportState);
376
377
378	/** @name \#VMEXIT handlers.
379	* @{
380	*/
381	static FNSVMEXITHANDLER hmR0SvmExitIntr;
382	static FNSVMEXITHANDLER hmR0SvmExitWbinvd;
383	static FNSVMEXITHANDLER hmR0SvmExitInvd;
384	static FNSVMEXITHANDLER hmR0SvmExitCpuid;
385	static FNSVMEXITHANDLER hmR0SvmExitRdtsc;
386	static FNSVMEXITHANDLER hmR0SvmExitRdtscp;
387	static FNSVMEXITHANDLER hmR0SvmExitRdpmc;
388	static FNSVMEXITHANDLER hmR0SvmExitInvlpg;
389	static FNSVMEXITHANDLER hmR0SvmExitHlt;
390	static FNSVMEXITHANDLER hmR0SvmExitMonitor;
391	static FNSVMEXITHANDLER hmR0SvmExitMwait;
392	static FNSVMEXITHANDLER hmR0SvmExitShutdown;
393	static FNSVMEXITHANDLER hmR0SvmExitUnexpected;
394	static FNSVMEXITHANDLER hmR0SvmExitReadCRx;
395	static FNSVMEXITHANDLER hmR0SvmExitWriteCRx;
396	static FNSVMEXITHANDLER hmR0SvmExitMsr;
397	static FNSVMEXITHANDLER hmR0SvmExitReadDRx;
398	static FNSVMEXITHANDLER hmR0SvmExitWriteDRx;
399	static FNSVMEXITHANDLER hmR0SvmExitXsetbv;
400	static FNSVMEXITHANDLER hmR0SvmExitIOInstr;
401	static FNSVMEXITHANDLER hmR0SvmExitNestedPF;
402	static FNSVMEXITHANDLER hmR0SvmExitVIntr;
403	static FNSVMEXITHANDLER hmR0SvmExitTaskSwitch;
404	static FNSVMEXITHANDLER hmR0SvmExitVmmCall;
405	static FNSVMEXITHANDLER hmR0SvmExitPause;
406	static FNSVMEXITHANDLER hmR0SvmExitFerrFreeze;
407	static FNSVMEXITHANDLER hmR0SvmExitIret;
408	static FNSVMEXITHANDLER hmR0SvmExitXcptPF;
409	static FNSVMEXITHANDLER hmR0SvmExitXcptUD;
410	static FNSVMEXITHANDLER hmR0SvmExitXcptMF;
411	static FNSVMEXITHANDLER hmR0SvmExitXcptDB;
412	static FNSVMEXITHANDLER hmR0SvmExitXcptAC;
413	static FNSVMEXITHANDLER hmR0SvmExitXcptBP;
414	#if defined(HMSVM_ALWAYS_TRAP_ALL_XCPTS) \|\| defined(VBOX_WITH_NESTED_HWVIRT_SVM)
415	static FNSVMEXITHANDLER hmR0SvmExitXcptGeneric;
416	#endif
417	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
418	static FNSVMEXITHANDLER hmR0SvmExitClgi;
419	static FNSVMEXITHANDLER hmR0SvmExitStgi;
420	static FNSVMEXITHANDLER hmR0SvmExitVmload;
421	static FNSVMEXITHANDLER hmR0SvmExitVmsave;
422	static FNSVMEXITHANDLER hmR0SvmExitInvlpga;
423	static FNSVMEXITHANDLER hmR0SvmExitVmrun;
424	static FNSVMEXITHANDLER hmR0SvmNestedExitXcptDB;
425	static FNSVMEXITHANDLER hmR0SvmNestedExitXcptBP;
426	#endif
427	/** @} */
428
429	static int hmR0SvmHandleExit(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
430	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
431	static int hmR0SvmHandleExitNested(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
432	#endif
433
434
435	/*********************************************************************************************************************************
436	* Global Variables *
437	*********************************************************************************************************************************/
438	/** Ring-0 memory object for the IO bitmap. */
439	static RTR0MEMOBJ g_hMemObjIOBitmap = NIL_RTR0MEMOBJ;
440	/** Physical address of the IO bitmap. */
441	static RTHCPHYS g_HCPhysIOBitmap;
442	/** Pointer to the IO bitmap. */
443	static R0PTRTYPE(void *) g_pvIOBitmap;
444
445	#ifdef VBOX_STRICT
446	# define HMSVM_LOG_RBP_RSP RT_BIT_32(0)
447	# define HMSVM_LOG_CR_REGS RT_BIT_32(1)
448	# define HMSVM_LOG_CS RT_BIT_32(2)
449	# define HMSVM_LOG_SS RT_BIT_32(3)
450	# define HMSVM_LOG_FS RT_BIT_32(4)
451	# define HMSVM_LOG_GS RT_BIT_32(5)
452	# define HMSVM_LOG_LBR RT_BIT_32(6)
453	# define HMSVM_LOG_ALL ( HMSVM_LOG_RBP_RSP \
454	\| HMSVM_LOG_CR_REGS \
455	\| HMSVM_LOG_CS \
456	\| HMSVM_LOG_SS \
457	\| HMSVM_LOG_FS \
458	\| HMSVM_LOG_GS \
459	\| HMSVM_LOG_LBR)
460
461	/**
462	* Dumps virtual CPU state and additional info. to the logger for diagnostics.
463	*
464	* @param pVCpu The cross context virtual CPU structure.
465	* @param pVmcb Pointer to the VM control block.
466	* @param pszPrefix Log prefix.
467	* @param fFlags Log flags, see HMSVM_LOG_XXX.
468	* @param uVerbose The verbosity level, currently unused.
469	*/
470	static void hmR0SvmLogState(PVMCPU pVCpu, PCSVMVMCB pVmcb, const char *pszPrefix, uint32_t fFlags, uint8_t uVerbose)
471	{
472	RT_NOREF2(pVCpu, uVerbose);
473	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
474
475	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_CS \| CPUMCTX_EXTRN_RIP \| CPUMCTX_EXTRN_RFLAGS);
476	Log4(("%s: cs:rip=%04x:%RX64 efl=%#RX64\n", pszPrefix, pCtx->cs.Sel, pCtx->rip, pCtx->rflags.u));
477
478	if (fFlags & HMSVM_LOG_RBP_RSP)
479	{
480	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_RSP \| CPUMCTX_EXTRN_RBP);
481	Log4(("%s: rsp=%#RX64 rbp=%#RX64\n", pszPrefix, pCtx->rsp, pCtx->rbp));
482	}
483
484	if (fFlags & HMSVM_LOG_CR_REGS)
485	{
486	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_CR0 \| CPUMCTX_EXTRN_CR3 \| CPUMCTX_EXTRN_CR4);
487	Log4(("%s: cr0=%#RX64 cr3=%#RX64 cr4=%#RX64\n", pszPrefix, pCtx->cr0, pCtx->cr3, pCtx->cr4));
488	}
489
490	if (fFlags & HMSVM_LOG_CS)
491	{
492	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_CS);
493	Log4(("%s: cs={%04x base=%016RX64 limit=%08x flags=%08x}\n", pszPrefix, pCtx->cs.Sel, pCtx->cs.u64Base,
494	pCtx->cs.u32Limit, pCtx->cs.Attr.u));
495	}
496	if (fFlags & HMSVM_LOG_SS)
497	{
498	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_SS);
499	Log4(("%s: ss={%04x base=%016RX64 limit=%08x flags=%08x}\n", pszPrefix, pCtx->ss.Sel, pCtx->ss.u64Base,
500	pCtx->ss.u32Limit, pCtx->ss.Attr.u));
501	}
502	if (fFlags & HMSVM_LOG_FS)
503	{
504	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_FS);
505	Log4(("%s: fs={%04x base=%016RX64 limit=%08x flags=%08x}\n", pszPrefix, pCtx->fs.Sel, pCtx->fs.u64Base,
506	pCtx->fs.u32Limit, pCtx->fs.Attr.u));
507	}
508	if (fFlags & HMSVM_LOG_GS)
509	{
510	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_GS);
511	Log4(("%s: gs={%04x base=%016RX64 limit=%08x flags=%08x}\n", pszPrefix, pCtx->gs.Sel, pCtx->gs.u64Base,
512	pCtx->gs.u32Limit, pCtx->gs.Attr.u));
513	}
514
515	PCSVMVMCBSTATESAVE pVmcbGuest = &pVmcb->guest;
516	if (fFlags & HMSVM_LOG_LBR)
517	{
518	Log4(("%s: br_from=%#RX64 br_to=%#RX64 lastxcpt_from=%#RX64 lastxcpt_to=%#RX64\n", pszPrefix, pVmcbGuest->u64BR_FROM,
519	pVmcbGuest->u64BR_TO, pVmcbGuest->u64LASTEXCPFROM, pVmcbGuest->u64LASTEXCPTO));
520	}
521	NOREF(pVmcbGuest); NOREF(pCtx);
522	}
523	#endif /* VBOX_STRICT */
524
525
526	/**
527	* Sets up and activates AMD-V on the current CPU.
528	*
529	* @returns VBox status code.
530	* @param pCpu Pointer to the CPU info struct.
531	* @param pVM The cross context VM structure. Can be
532	* NULL after a resume!
533	* @param pvCpuPage Pointer to the global CPU page.
534	* @param HCPhysCpuPage Physical address of the global CPU page.
535	* @param fEnabledByHost Whether the host OS has already initialized AMD-V.
536	* @param pvArg Unused on AMD-V.
537	*/
538	VMMR0DECL(int) SVMR0EnableCpu(PHMGLOBALCPUINFO pCpu, PVM pVM, void *pvCpuPage, RTHCPHYS HCPhysCpuPage, bool fEnabledByHost,
539	void *pvArg)
540	{
541	Assert(!fEnabledByHost);
542	Assert(HCPhysCpuPage && HCPhysCpuPage != NIL_RTHCPHYS);
543	Assert(RT_ALIGN_T(HCPhysCpuPage, _4K, RTHCPHYS) == HCPhysCpuPage);
544	Assert(pvCpuPage); NOREF(pvCpuPage);
545	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
546
547	NOREF(pvArg);
548	NOREF(fEnabledByHost);
549
550	/* Paranoid: Disable interrupt as, in theory, interrupt handlers might mess with EFER. */
551	RTCCUINTREG const fEFlags = ASMIntDisableFlags();
552
553	/*
554	* We must turn on AMD-V and setup the host state physical address, as those MSRs are per CPU.
555	*/
556	uint64_t u64HostEfer = ASMRdMsr(MSR_K6_EFER);
557	if (u64HostEfer & MSR_K6_EFER_SVME)
558	{
559	/* If the VBOX_HWVIRTEX_IGNORE_SVM_IN_USE is active, then we blindly use AMD-V. */
560	if ( pVM
561	&& pVM->hm.s.svm.fIgnoreInUseError)
562	pCpu->fIgnoreAMDVInUseError = true;
563
564	if (!pCpu->fIgnoreAMDVInUseError)
565	{
566	ASMSetFlags(fEFlags);
567	return VERR_SVM_IN_USE;
568	}
569	}
570
571	/* Turn on AMD-V in the EFER MSR. */
572	ASMWrMsr(MSR_K6_EFER, u64HostEfer \| MSR_K6_EFER_SVME);
573
574	/* Write the physical page address where the CPU will store the host state while executing the VM. */
575	ASMWrMsr(MSR_K8_VM_HSAVE_PA, HCPhysCpuPage);
576
577	/* Restore interrupts. */
578	ASMSetFlags(fEFlags);
579
580	/*
581	* Theoretically, other hypervisors may have used ASIDs, ideally we should flush all
582	* non-zero ASIDs when enabling SVM. AMD doesn't have an SVM instruction to flush all
583	* ASIDs (flushing is done upon VMRUN). Therefore, flag that we need to flush the TLB
584	* entirely with before executing any guest code.
585	*/
586	pCpu->fFlushAsidBeforeUse = true;
587
588	/*
589	* Ensure each VCPU scheduled on this CPU gets a new ASID on resume. See @bugref{6255}.
590	*/
591	++pCpu->cTlbFlushes;
592
593	return VINF_SUCCESS;
594	}
595
596
597	/**
598	* Deactivates AMD-V on the current CPU.
599	*
600	* @returns VBox status code.
601	* @param pCpu Pointer to the CPU info struct.
602	* @param pvCpuPage Pointer to the global CPU page.
603	* @param HCPhysCpuPage Physical address of the global CPU page.
604	*/
605	VMMR0DECL(int) SVMR0DisableCpu(PHMGLOBALCPUINFO pCpu, void *pvCpuPage, RTHCPHYS HCPhysCpuPage)
606	{
607	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
608	AssertReturn( HCPhysCpuPage
609	&& HCPhysCpuPage != NIL_RTHCPHYS, VERR_INVALID_PARAMETER);
610	AssertReturn(pvCpuPage, VERR_INVALID_PARAMETER);
611	NOREF(pCpu);
612
613	/* Paranoid: Disable interrupts as, in theory, interrupt handlers might mess with EFER. */
614	RTCCUINTREG const fEFlags = ASMIntDisableFlags();
615
616	/* Turn off AMD-V in the EFER MSR. */
617	uint64_t u64HostEfer = ASMRdMsr(MSR_K6_EFER);
618	ASMWrMsr(MSR_K6_EFER, u64HostEfer & ~MSR_K6_EFER_SVME);
619
620	/* Invalidate host state physical address. */
621	ASMWrMsr(MSR_K8_VM_HSAVE_PA, 0);
622
623	/* Restore interrupts. */
624	ASMSetFlags(fEFlags);
625
626	return VINF_SUCCESS;
627	}
628
629
630	/**
631	* Does global AMD-V initialization (called during module initialization).
632	*
633	* @returns VBox status code.
634	*/
635	VMMR0DECL(int) SVMR0GlobalInit(void)
636	{
637	/*
638	* Allocate 12 KB (3 pages) for the IO bitmap. Since this is non-optional and we always
639	* intercept all IO accesses, it's done once globally here instead of per-VM.
640	*/
641	Assert(g_hMemObjIOBitmap == NIL_RTR0MEMOBJ);
642	int rc = RTR0MemObjAllocCont(&g_hMemObjIOBitmap, SVM_IOPM_PAGES << X86_PAGE_4K_SHIFT, false /* fExecutable */);
643	if (RT_FAILURE(rc))
644	return rc;
645
646	g_pvIOBitmap = RTR0MemObjAddress(g_hMemObjIOBitmap);
647	g_HCPhysIOBitmap = RTR0MemObjGetPagePhysAddr(g_hMemObjIOBitmap, 0 /* iPage */);
648
649	/* Set all bits to intercept all IO accesses. */
650	ASMMemFill32(g_pvIOBitmap, SVM_IOPM_PAGES << X86_PAGE_4K_SHIFT, UINT32_C(0xffffffff));
651
652	return VINF_SUCCESS;
653	}
654
655
656	/**
657	* Does global AMD-V termination (called during module termination).
658	*/
659	VMMR0DECL(void) SVMR0GlobalTerm(void)
660	{
661	if (g_hMemObjIOBitmap != NIL_RTR0MEMOBJ)
662	{
663	RTR0MemObjFree(g_hMemObjIOBitmap, true /* fFreeMappings */);
664	g_pvIOBitmap = NULL;
665	g_HCPhysIOBitmap = 0;
666	g_hMemObjIOBitmap = NIL_RTR0MEMOBJ;
667	}
668	}
669
670
671	/**
672	* Frees any allocated per-VCPU structures for a VM.
673	*
674	* @param pVM The cross context VM structure.
675	*/
676	DECLINLINE(void) hmR0SvmFreeStructs(PVM pVM)
677	{
678	for (uint32_t i = 0; i < pVM->cCpus; i++)
679	{
680	PVMCPU pVCpu = &pVM->aCpus[i];
681	AssertPtr(pVCpu);
682
683	if (pVCpu->hm.s.svm.hMemObjVmcbHost != NIL_RTR0MEMOBJ)
684	{
685	RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjVmcbHost, false);
686	pVCpu->hm.s.svm.HCPhysVmcbHost = 0;
687	pVCpu->hm.s.svm.hMemObjVmcbHost = NIL_RTR0MEMOBJ;
688	}
689
690	if (pVCpu->hm.s.svm.hMemObjVmcb != NIL_RTR0MEMOBJ)
691	{
692	RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjVmcb, false);
693	pVCpu->hm.s.svm.pVmcb = NULL;
694	pVCpu->hm.s.svm.HCPhysVmcb = 0;
695	pVCpu->hm.s.svm.hMemObjVmcb = NIL_RTR0MEMOBJ;
696	}
697
698	if (pVCpu->hm.s.svm.hMemObjMsrBitmap != NIL_RTR0MEMOBJ)
699	{
700	RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjMsrBitmap, false);
701	pVCpu->hm.s.svm.pvMsrBitmap = NULL;
702	pVCpu->hm.s.svm.HCPhysMsrBitmap = 0;
703	pVCpu->hm.s.svm.hMemObjMsrBitmap = NIL_RTR0MEMOBJ;
704	}
705	}
706	}
707
708
709	/**
710	* Does per-VM AMD-V initialization.
711	*
712	* @returns VBox status code.
713	* @param pVM The cross context VM structure.
714	*/
715	VMMR0DECL(int) SVMR0InitVM(PVM pVM)
716	{
717	int rc = VERR_INTERNAL_ERROR_5;
718
719	/*
720	* Check for an AMD CPU erratum which requires us to flush the TLB before every world-switch.
721	*/
722	uint32_t u32Family;
723	uint32_t u32Model;
724	uint32_t u32Stepping;
725	if (HMAmdIsSubjectToErratum170(&u32Family, &u32Model, &u32Stepping))
726	{
727	Log4Func(("AMD cpu with erratum 170 family %#x model %#x stepping %#x\n", u32Family, u32Model, u32Stepping));
728	pVM->hm.s.svm.fAlwaysFlushTLB = true;
729	}
730
731	/*
732	* Initialize the R0 memory objects up-front so we can properly cleanup on allocation failures.
733	*/
734	for (VMCPUID i = 0; i < pVM->cCpus; i++)
735	{
736	PVMCPU pVCpu = &pVM->aCpus[i];
737	pVCpu->hm.s.svm.hMemObjVmcbHost = NIL_RTR0MEMOBJ;
738	pVCpu->hm.s.svm.hMemObjVmcb = NIL_RTR0MEMOBJ;
739	pVCpu->hm.s.svm.hMemObjMsrBitmap = NIL_RTR0MEMOBJ;
740	}
741
742	for (VMCPUID i = 0; i < pVM->cCpus; i++)
743	{
744	PVMCPU pVCpu = &pVM->aCpus[i];
745
746	/*
747	* Allocate one page for the host-context VM control block (VMCB). This is used for additional host-state (such as
748	* FS, GS, Kernel GS Base, etc.) apart from the host-state save area specified in MSR_K8_VM_HSAVE_PA.
749	*/
750	rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjVmcbHost, SVM_VMCB_PAGES << PAGE_SHIFT, false /* fExecutable */);
751	if (RT_FAILURE(rc))
752	goto failure_cleanup;
753
754	void *pvVmcbHost = RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjVmcbHost);
755	pVCpu->hm.s.svm.HCPhysVmcbHost = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjVmcbHost, 0 /* iPage */);
756	Assert(pVCpu->hm.s.svm.HCPhysVmcbHost < _4G);
757	ASMMemZeroPage(pvVmcbHost);
758
759	/*
760	* Allocate one page for the guest-state VMCB.
761	*/
762	rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjVmcb, SVM_VMCB_PAGES << PAGE_SHIFT, false /* fExecutable */);
763	if (RT_FAILURE(rc))
764	goto failure_cleanup;
765
766	pVCpu->hm.s.svm.pVmcb = (PSVMVMCB)RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjVmcb);
767	pVCpu->hm.s.svm.HCPhysVmcb = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjVmcb, 0 /* iPage */);
768	Assert(pVCpu->hm.s.svm.HCPhysVmcb < _4G);
769	ASMMemZeroPage(pVCpu->hm.s.svm.pVmcb);
770
771	/*
772	* Allocate two pages (8 KB) for the MSR permission bitmap. There doesn't seem to be a way to convince
773	* SVM to not require one.
774	*/
775	rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjMsrBitmap, SVM_MSRPM_PAGES << X86_PAGE_4K_SHIFT,
776	false /* fExecutable */);
777	if (RT_FAILURE(rc))
778	goto failure_cleanup;
779
780	pVCpu->hm.s.svm.pvMsrBitmap = RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjMsrBitmap);
781	pVCpu->hm.s.svm.HCPhysMsrBitmap = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjMsrBitmap, 0 /* iPage */);
782	/* Set all bits to intercept all MSR accesses (changed later on). */
783	ASMMemFill32(pVCpu->hm.s.svm.pvMsrBitmap, SVM_MSRPM_PAGES << X86_PAGE_4K_SHIFT, UINT32_C(0xffffffff));
784	}
785
786	return VINF_SUCCESS;
787
788	failure_cleanup:
789	hmR0SvmFreeStructs(pVM);
790	return rc;
791	}
792
793
794	/**
795	* Does per-VM AMD-V termination.
796	*
797	* @returns VBox status code.
798	* @param pVM The cross context VM structure.
799	*/
800	VMMR0DECL(int) SVMR0TermVM(PVM pVM)
801	{
802	hmR0SvmFreeStructs(pVM);
803	return VINF_SUCCESS;
804	}
805
806
807	/**
808	* Returns whether the VMCB Clean Bits feature is supported.
809	*
810	* @return @c true if supported, @c false otherwise.
811	* @param pVCpu The cross context virtual CPU structure.
812	* @param pCtx Pointer to the guest-CPU context.
813	*/
814	DECLINLINE(bool) hmR0SvmSupportsVmcbCleanBits(PVMCPU pVCpu, PCCPUMCTX pCtx)
815	{
816	PVM pVM = pVCpu->CTX_SUFF(pVM);
817	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
818	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
819	{
820	return (pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_VMCB_CLEAN)
821	&& pVM->cpum.ro.GuestFeatures.fSvmVmcbClean;
822	}
823	#else
824	RT_NOREF(pCtx);
825	#endif
826	return RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_VMCB_CLEAN);
827	}
828
829
830	/**
831	* Returns whether the decode assists feature is supported.
832	*
833	* @return @c true if supported, @c false otherwise.
834	* @param pVCpu The cross context virtual CPU structure.
835	* @param pCtx Pointer to the guest-CPU context.
836	*/
837	DECLINLINE(bool) hmR0SvmSupportsDecodeAssists(PVMCPU pVCpu, PCPUMCTX pCtx)
838	{
839	PVM pVM = pVCpu->CTX_SUFF(pVM);
840	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
841	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
842	{
843	return (pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_DECODE_ASSISTS)
844	&& pVM->cpum.ro.GuestFeatures.fSvmDecodeAssists;
845	}
846	#else
847	RT_NOREF(pCtx);
848	#endif
849	return RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_DECODE_ASSISTS);
850	}
851
852
853	/**
854	* Returns whether the NRIP_SAVE feature is supported.
855	*
856	* @return @c true if supported, @c false otherwise.
857	* @param pVCpu The cross context virtual CPU structure.
858	* @param pCtx Pointer to the guest-CPU context.
859	*/
860	DECLINLINE(bool) hmR0SvmSupportsNextRipSave(PVMCPU pVCpu, PCPUMCTX pCtx)
861	{
862	PVM pVM = pVCpu->CTX_SUFF(pVM);
863	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
864	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
865	{
866	return (pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_NRIP_SAVE)
867	&& pVM->cpum.ro.GuestFeatures.fSvmNextRipSave;
868	}
869	#else
870	RT_NOREF(pCtx);
871	#endif
872	return RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_NRIP_SAVE);
873	}
874
875
876	/**
877	* Sets the permission bits for the specified MSR in the MSRPM bitmap.
878	*
879	* @param pCtx Pointer to the guest-CPU or nested-guest-CPU context.
880	* @param pbMsrBitmap Pointer to the MSR bitmap.
881	* @param idMsr The MSR for which the permissions are being set.
882	* @param enmRead MSR read permissions.
883	* @param enmWrite MSR write permissions.
884	*
885	* @remarks This function does -not- clear the VMCB clean bits for MSRPM. The
886	* caller needs to take care of this.
887	*/
888	static void hmR0SvmSetMsrPermission(PCCPUMCTX pCtx, uint8_t *pbMsrBitmap, uint32_t idMsr, SVMMSREXITREAD enmRead,
889	SVMMSREXITWRITE enmWrite)
890	{
891	bool const fInNestedGuestMode = CPUMIsGuestInSvmNestedHwVirtMode(pCtx);
892	uint16_t offMsrpm;
893	uint8_t uMsrpmBit;
894	int rc = HMSvmGetMsrpmOffsetAndBit(idMsr, &offMsrpm, &uMsrpmBit);
895	AssertRC(rc);
896
897	Assert(uMsrpmBit == 0 \|\| uMsrpmBit == 2 \|\| uMsrpmBit == 4 \|\| uMsrpmBit == 6);
898	Assert(offMsrpm < SVM_MSRPM_PAGES << X86_PAGE_4K_SHIFT);
899
900	pbMsrBitmap += offMsrpm;
901	if (enmRead == SVMMSREXIT_INTERCEPT_READ)
902	*pbMsrBitmap \|= RT_BIT(uMsrpmBit);
903	else
904	{
905	if (!fInNestedGuestMode)
906	*pbMsrBitmap &= ~RT_BIT(uMsrpmBit);
907	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
908	else
909	{
910	/* Only clear the bit if the nested-guest is also not intercepting the MSR read.*/
911	uint8_t const pbNstGstMsrBitmap = (uint8_t )pCtx->hwvirt.svm.CTX_SUFF(pvMsrBitmap);
912	pbNstGstMsrBitmap += offMsrpm;
913	if (!(*pbNstGstMsrBitmap & RT_BIT(uMsrpmBit)))
914	*pbMsrBitmap &= ~RT_BIT(uMsrpmBit);
915	else
916	Assert(*pbMsrBitmap & RT_BIT(uMsrpmBit));
917	}
918	#endif
919	}
920
921	if (enmWrite == SVMMSREXIT_INTERCEPT_WRITE)
922	*pbMsrBitmap \|= RT_BIT(uMsrpmBit + 1);
923	else
924	{
925	if (!fInNestedGuestMode)
926	*pbMsrBitmap &= ~RT_BIT(uMsrpmBit + 1);
927	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
928	else
929	{
930	/* Only clear the bit if the nested-guest is also not intercepting the MSR write.*/
931	uint8_t const pbNstGstMsrBitmap = (uint8_t )pCtx->hwvirt.svm.CTX_SUFF(pvMsrBitmap);
932	pbNstGstMsrBitmap += offMsrpm;
933	if (!(*pbNstGstMsrBitmap & RT_BIT(uMsrpmBit + 1)))
934	*pbMsrBitmap &= ~RT_BIT(uMsrpmBit + 1);
935	else
936	Assert(*pbMsrBitmap & RT_BIT(uMsrpmBit + 1));
937	}
938	#endif
939	}
940	}
941
942
943	/**
944	* Sets up AMD-V for the specified VM.
945	* This function is only called once per-VM during initalization.
946	*
947	* @returns VBox status code.
948	* @param pVM The cross context VM structure.
949	*/
950	VMMR0DECL(int) SVMR0SetupVM(PVM pVM)
951	{
952	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
953	AssertReturn(pVM, VERR_INVALID_PARAMETER);
954	Assert(pVM->hm.s.svm.fSupported);
955
956	bool const fPauseFilter = RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_PAUSE_FILTER);
957	bool const fPauseFilterThreshold = RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_PAUSE_FILTER_THRESHOLD);
958	bool const fUsePauseFilter = fPauseFilter && pVM->hm.s.svm.cPauseFilter;
959
960	bool const fLbrVirt = RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_LBR_VIRT);
961	bool const fUseLbrVirt = fLbrVirt; /** @todo CFGM, IEM implementation etc. */
962
963	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
964	bool const fVirtVmsaveVmload = RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_VIRT_VMSAVE_VMLOAD);
965	bool const fUseVirtVmsaveVmload = fVirtVmsaveVmload && pVM->hm.s.svm.fVirtVmsaveVmload && pVM->hm.s.fNestedPaging;
966
967	bool const fVGif = RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_VGIF);
968	bool const fUseVGif = fVGif && pVM->hm.s.svm.fVGif;
969	#endif
970
971	PVMCPU pVCpu = &pVM->aCpus[0];
972	PSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
973	AssertMsgReturn(pVmcb, ("Invalid pVmcb for vcpu[0]\n"), VERR_SVM_INVALID_PVMCB);
974	PSVMVMCBCTRL pVmcbCtrl = &pVmcb->ctrl;
975
976	/* Always trap #AC for reasons of security. */
977	pVmcbCtrl->u32InterceptXcpt \|= RT_BIT_32(X86_XCPT_AC);
978
979	/* Always trap #DB for reasons of security. */
980	pVmcbCtrl->u32InterceptXcpt \|= RT_BIT_32(X86_XCPT_DB);
981
982	/* Trap exceptions unconditionally (debug purposes). */
983	#ifdef HMSVM_ALWAYS_TRAP_PF
984	pVmcbCtrl->u32InterceptXcpt \|= RT_BIT(X86_XCPT_PF);
985	#endif
986	#ifdef HMSVM_ALWAYS_TRAP_ALL_XCPTS
987	/* If you add any exceptions here, make sure to update hmR0SvmHandleExit(). */
988	pVmcbCtrl->u32InterceptXcpt \|= 0
989	\| RT_BIT(X86_XCPT_BP)
990	\| RT_BIT(X86_XCPT_DE)
991	\| RT_BIT(X86_XCPT_NM)
992	\| RT_BIT(X86_XCPT_UD)
993	\| RT_BIT(X86_XCPT_NP)
994	\| RT_BIT(X86_XCPT_SS)
995	\| RT_BIT(X86_XCPT_GP)
996	\| RT_BIT(X86_XCPT_PF)
997	\| RT_BIT(X86_XCPT_MF)
998	;
999	#endif
1000
1001	/* Apply the exceptions intercepts needed by the GIM provider. */
1002	if (pVCpu->hm.s.fGIMTrapXcptUD)
1003	pVmcbCtrl->u32InterceptXcpt \|= RT_BIT(X86_XCPT_UD);
1004
1005	/* Set up unconditional intercepts and conditions. */
1006	pVmcbCtrl->u64InterceptCtrl = HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS
1007	\| SVM_CTRL_INTERCEPT_VMMCALL;
1008
1009	#ifdef HMSVM_ALWAYS_TRAP_TASK_SWITCH
1010	pVmcbCtrl->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_TASK_SWITCH;
1011	#endif
1012
1013	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1014	/* Virtualized VMSAVE/VMLOAD. */
1015	pVmcbCtrl->LbrVirt.n.u1VirtVmsaveVmload = fUseVirtVmsaveVmload;
1016	if (!fUseVirtVmsaveVmload)
1017	{
1018	pVmcbCtrl->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_VMSAVE
1019	\| SVM_CTRL_INTERCEPT_VMLOAD;
1020	}
1021
1022	/* Virtual GIF. */
1023	pVmcbCtrl->IntCtrl.n.u1VGifEnable = fUseVGif;
1024	if (!fUseVGif)
1025	{
1026	pVmcbCtrl->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_CLGI
1027	\| SVM_CTRL_INTERCEPT_STGI;
1028	}
1029	#endif
1030
1031	/* CR4 writes must always be intercepted for tracking PGM mode changes. */
1032	pVmcbCtrl->u16InterceptWrCRx = RT_BIT(4);
1033
1034	/* Intercept all DRx reads and writes by default. Changed later on. */
1035	pVmcbCtrl->u16InterceptRdDRx = 0xffff;
1036	pVmcbCtrl->u16InterceptWrDRx = 0xffff;
1037
1038	/* Virtualize masking of INTR interrupts. (reads/writes from/to CR8 go to the V_TPR register) */
1039	pVmcbCtrl->IntCtrl.n.u1VIntrMasking = 1;
1040
1041	/* Ignore the priority in the virtual TPR. This is necessary for delivering PIC style (ExtInt) interrupts
1042	and we currently deliver both PIC and APIC interrupts alike, see hmR0SvmEvaluatePendingEvent() */
1043	pVmcbCtrl->IntCtrl.n.u1IgnoreTPR = 1;
1044
1045	/* Set the IO permission bitmap physical addresses. */
1046	pVmcbCtrl->u64IOPMPhysAddr = g_HCPhysIOBitmap;
1047
1048	/* LBR virtualization. */
1049	pVmcbCtrl->LbrVirt.n.u1LbrVirt = fUseLbrVirt;
1050
1051	/* The host ASID MBZ, for the guest start with 1. */
1052	pVmcbCtrl->TLBCtrl.n.u32ASID = 1;
1053
1054	/* Setup Nested Paging. This doesn't change throughout the execution time of the VM. */
1055	pVmcbCtrl->NestedPagingCtrl.n.u1NestedPaging = pVM->hm.s.fNestedPaging;
1056
1057	/* Without Nested Paging, we need additionally intercepts. */
1058	if (!pVM->hm.s.fNestedPaging)
1059	{
1060	/* CR3 reads/writes must be intercepted; our shadow values differ from the guest values. */
1061	pVmcbCtrl->u16InterceptRdCRx \|= RT_BIT(3);
1062	pVmcbCtrl->u16InterceptWrCRx \|= RT_BIT(3);
1063
1064	/* Intercept INVLPG and task switches (may change CR3, EFLAGS, LDT). */
1065	pVmcbCtrl->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_INVLPG
1066	\| SVM_CTRL_INTERCEPT_TASK_SWITCH;
1067
1068	/* Page faults must be intercepted to implement shadow paging. */
1069	pVmcbCtrl->u32InterceptXcpt \|= RT_BIT(X86_XCPT_PF);
1070	}
1071
1072	/* Setup Pause Filter for guest pause-loop (spinlock) exiting. */
1073	if (fUsePauseFilter)
1074	{
1075	Assert(pVM->hm.s.svm.cPauseFilter > 0);
1076	pVmcbCtrl->u16PauseFilterCount = pVM->hm.s.svm.cPauseFilter;
1077	if (fPauseFilterThreshold)
1078	pVmcbCtrl->u16PauseFilterThreshold = pVM->hm.s.svm.cPauseFilterThresholdTicks;
1079	pVmcbCtrl->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_PAUSE;
1080	}
1081
1082	/*
1083	* Setup the MSR permission bitmap.
1084	* The following MSRs are saved/restored automatically during the world-switch.
1085	* Don't intercept guest read/write accesses to these MSRs.
1086	*/
1087	uint8_t pbMsrBitmap = (uint8_t )pVCpu->hm.s.svm.pvMsrBitmap;
1088	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1089	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1090	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_K8_CSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1091	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_K6_STAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1092	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_K8_SF_MASK, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1093	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_K8_FS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1094	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_K8_GS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1095	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_K8_KERNEL_GS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1096	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_IA32_SYSENTER_CS, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1097	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_IA32_SYSENTER_ESP, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1098	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_IA32_SYSENTER_EIP, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1099	pVmcbCtrl->u64MSRPMPhysAddr = pVCpu->hm.s.svm.HCPhysMsrBitmap;
1100
1101	/* Initially all VMCB clean bits MBZ indicating that everything should be loaded from the VMCB in memory. */
1102	Assert(pVmcbCtrl->u32VmcbCleanBits == 0);
1103
1104	for (VMCPUID i = 1; i < pVM->cCpus; i++)
1105	{
1106	PVMCPU pVCpuCur = &pVM->aCpus[i];
1107	PSVMVMCB pVmcbCur = pVM->aCpus[i].hm.s.svm.pVmcb;
1108	AssertMsgReturn(pVmcbCur, ("Invalid pVmcb for vcpu[%u]\n", i), VERR_SVM_INVALID_PVMCB);
1109	PSVMVMCBCTRL pVmcbCtrlCur = &pVmcbCur->ctrl;
1110
1111	/* Copy the VMCB control area. */
1112	memcpy(pVmcbCtrlCur, pVmcbCtrl, sizeof(*pVmcbCtrlCur));
1113
1114	/* Copy the MSR bitmap and setup the VCPU-specific host physical address. */
1115	uint8_t pbMsrBitmapCur = (uint8_t )pVCpuCur->hm.s.svm.pvMsrBitmap;
1116	memcpy(pbMsrBitmapCur, pbMsrBitmap, SVM_MSRPM_PAGES << X86_PAGE_4K_SHIFT);
1117	pVmcbCtrlCur->u64MSRPMPhysAddr = pVCpuCur->hm.s.svm.HCPhysMsrBitmap;
1118
1119	/* Initially all VMCB clean bits MBZ indicating that everything should be loaded from the VMCB in memory. */
1120	Assert(pVmcbCtrlCur->u32VmcbCleanBits == 0);
1121
1122	/* Verify our assumption that GIM providers trap #UD uniformly across VCPUs initially. */
1123	Assert(pVCpuCur->hm.s.fGIMTrapXcptUD == pVCpu->hm.s.fGIMTrapXcptUD);
1124	}
1125
1126	return VINF_SUCCESS;
1127	}
1128
1129
1130	/**
1131	* Gets a pointer to the currently active guest (or nested-guest) VMCB.
1132	*
1133	* @returns Pointer to the current context VMCB.
1134	* @param pVCpu The cross context virtual CPU structure.
1135	* @param pCtx Pointer to the guest-CPU context.
1136	*/
1137	DECLINLINE(PSVMVMCB) hmR0SvmGetCurrentVmcb(PVMCPU pVCpu, PCPUMCTX pCtx)
1138	{
1139	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1140	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
1141	return pCtx->hwvirt.svm.CTX_SUFF(pVmcb);
1142	#else
1143	RT_NOREF(pCtx);
1144	#endif
1145	return pVCpu->hm.s.svm.pVmcb;
1146	}
1147
1148
1149	/**
1150	* Gets a pointer to the nested-guest VMCB cache.
1151	*
1152	* @returns Pointer to the nested-guest VMCB cache.
1153	* @param pVCpu The cross context virtual CPU structure.
1154	*/
1155	DECLINLINE(PSVMNESTEDVMCBCACHE) hmR0SvmGetNestedVmcbCache(PVMCPU pVCpu)
1156	{
1157	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1158	Assert(pVCpu->hm.s.svm.NstGstVmcbCache.fCacheValid);
1159	return &pVCpu->hm.s.svm.NstGstVmcbCache;
1160	#else
1161	RT_NOREF(pVCpu);
1162	return NULL;
1163	#endif
1164	}
1165
1166
1167	/**
1168	* Invalidates a guest page by guest virtual address.
1169	*
1170	* @returns VBox status code.
1171	* @param pVM The cross context VM structure.
1172	* @param pVCpu The cross context virtual CPU structure.
1173	* @param GCVirt Guest virtual address of the page to invalidate.
1174	*/
1175	VMMR0DECL(int) SVMR0InvalidatePage(PVM pVM, PVMCPU pVCpu, RTGCPTR GCVirt)
1176	{
1177	AssertReturn(pVM, VERR_INVALID_PARAMETER);
1178	Assert(pVM->hm.s.svm.fSupported);
1179
1180	bool fFlushPending = pVM->hm.s.svm.fAlwaysFlushTLB \|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_TLB_FLUSH);
1181
1182	/* Skip it if a TLB flush is already pending. */
1183	if (!fFlushPending)
1184	{
1185	Log4Func(("%#RGv\n", GCVirt));
1186
1187	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1188	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
1189	AssertMsgReturn(pVmcb, ("Invalid pVmcb!\n"), VERR_SVM_INVALID_PVMCB);
1190
1191	#if HC_ARCH_BITS == 32
1192	/* If we get a flush in 64-bit guest mode, then force a full TLB flush. INVLPGA takes only 32-bit addresses. */
1193	if (CPUMIsGuestInLongMode(pVCpu))
1194	VMCPU_FF_SET(pVCpu, VMCPU_FF_TLB_FLUSH);
1195	else
1196	#endif
1197	{
1198	SVMR0InvlpgA(GCVirt, pVmcb->ctrl.TLBCtrl.n.u32ASID);
1199	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlbInvlpgVirt);
1200	}
1201	}
1202	return VINF_SUCCESS;
1203	}
1204
1205
1206	/**
1207	* Flushes the appropriate tagged-TLB entries.
1208	*
1209	* @param pVCpu The cross context virtual CPU structure.
1210	* @param pCtx Pointer to the guest-CPU or nested-guest-CPU context.
1211	* @param pVmcb Pointer to the VM control block.
1212	* @param pHostCpu Pointer to the HM host-CPU info.
1213	*/
1214	static void hmR0SvmFlushTaggedTlb(PVMCPU pVCpu, PCCPUMCTX pCtx, PSVMVMCB pVmcb, PHMGLOBALCPUINFO pHostCpu)
1215	{
1216	#ifndef VBOX_WITH_NESTED_HWVIRT_SVM
1217	RT_NOREF(pCtx);
1218	#endif
1219	PVM pVM = pVCpu->CTX_SUFF(pVM);
1220
1221	/*
1222	* Force a TLB flush for the first world switch if the current CPU differs from the one
1223	* we ran on last. This can happen both for start & resume due to long jumps back to
1224	* ring-3.
1225	*
1226	* We also force a TLB flush every time when executing a nested-guest VCPU as there is no
1227	* correlation between it and the physical CPU.
1228	*
1229	* If the TLB flush count changed, another VM (VCPU rather) has hit the ASID limit while
1230	* flushing the TLB, so we cannot reuse the ASIDs without flushing.
1231	*/
1232	bool fNewAsid = false;
1233	Assert(pHostCpu->idCpu != NIL_RTCPUID);
1234	if ( pVCpu->hm.s.idLastCpu != pHostCpu->idCpu
1235	\|\| pVCpu->hm.s.cTlbFlushes != pHostCpu->cTlbFlushes
1236	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1237	\|\| CPUMIsGuestInSvmNestedHwVirtMode(pCtx)
1238	#endif
1239	)
1240	{
1241	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlbWorldSwitch);
1242	pVCpu->hm.s.fForceTLBFlush = true;
1243	fNewAsid = true;
1244	}
1245
1246	/* Set TLB flush state as checked until we return from the world switch. */
1247	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, true);
1248
1249	/* Check for explicit TLB flushes. */
1250	if (VMCPU_FF_TEST_AND_CLEAR(pVCpu, VMCPU_FF_TLB_FLUSH))
1251	{
1252	pVCpu->hm.s.fForceTLBFlush = true;
1253	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlb);
1254	}
1255
1256	/*
1257	* If the AMD CPU erratum 170, We need to flush the entire TLB for each world switch. Sad.
1258	* This Host CPU requirement takes precedence.
1259	*/
1260	if (pVM->hm.s.svm.fAlwaysFlushTLB)
1261	{
1262	pHostCpu->uCurrentAsid = 1;
1263	pVCpu->hm.s.uCurrentAsid = 1;
1264	pVCpu->hm.s.cTlbFlushes = pHostCpu->cTlbFlushes;
1265	pVCpu->hm.s.idLastCpu = pHostCpu->idCpu;
1266	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
1267
1268	/* Clear the VMCB Clean Bit for NP while flushing the TLB. See @bugref{7152}. */
1269	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
1270	}
1271	else
1272	{
1273	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_NOTHING;
1274	if (pVCpu->hm.s.fForceTLBFlush)
1275	{
1276	/* Clear the VMCB Clean Bit for NP while flushing the TLB. See @bugref{7152}. */
1277	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
1278
1279	if (fNewAsid)
1280	{
1281	++pHostCpu->uCurrentAsid;
1282
1283	bool fHitASIDLimit = false;
1284	if (pHostCpu->uCurrentAsid >= pVM->hm.s.uMaxAsid)
1285	{
1286	pHostCpu->uCurrentAsid = 1; /* Wraparound at 1; host uses 0 */
1287	pHostCpu->cTlbFlushes++; /* All VCPUs that run on this host CPU must use a new ASID. */
1288	fHitASIDLimit = true;
1289	}
1290
1291	if ( fHitASIDLimit
1292	\|\| pHostCpu->fFlushAsidBeforeUse)
1293	{
1294	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
1295	pHostCpu->fFlushAsidBeforeUse = false;
1296	}
1297
1298	pVCpu->hm.s.uCurrentAsid = pHostCpu->uCurrentAsid;
1299	pVCpu->hm.s.idLastCpu = pHostCpu->idCpu;
1300	pVCpu->hm.s.cTlbFlushes = pHostCpu->cTlbFlushes;
1301	}
1302	else
1303	{
1304	if (pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_FLUSH_BY_ASID)
1305	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_SINGLE_CONTEXT;
1306	else
1307	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
1308	}
1309
1310	pVCpu->hm.s.fForceTLBFlush = false;
1311	}
1312	}
1313
1314	/* Update VMCB with the ASID. */
1315	if (pVmcb->ctrl.TLBCtrl.n.u32ASID != pVCpu->hm.s.uCurrentAsid)
1316	{
1317	pVmcb->ctrl.TLBCtrl.n.u32ASID = pVCpu->hm.s.uCurrentAsid;
1318	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_ASID;
1319	}
1320
1321	AssertMsg(pVCpu->hm.s.idLastCpu == pHostCpu->idCpu,
1322	("vcpu idLastCpu=%u hostcpu idCpu=%u\n", pVCpu->hm.s.idLastCpu, pHostCpu->idCpu));
1323	AssertMsg(pVCpu->hm.s.cTlbFlushes == pHostCpu->cTlbFlushes,
1324	("Flush count mismatch for cpu %u (%u vs %u)\n", pHostCpu->idCpu, pVCpu->hm.s.cTlbFlushes, pHostCpu->cTlbFlushes));
1325	AssertMsg(pHostCpu->uCurrentAsid >= 1 && pHostCpu->uCurrentAsid < pVM->hm.s.uMaxAsid,
1326	("cpu%d uCurrentAsid = %x\n", pHostCpu->idCpu, pHostCpu->uCurrentAsid));
1327	AssertMsg(pVCpu->hm.s.uCurrentAsid >= 1 && pVCpu->hm.s.uCurrentAsid < pVM->hm.s.uMaxAsid,
1328	("cpu%d VM uCurrentAsid = %x\n", pHostCpu->idCpu, pVCpu->hm.s.uCurrentAsid));
1329
1330	#ifdef VBOX_WITH_STATISTICS
1331	if (pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_NOTHING)
1332	STAM_COUNTER_INC(&pVCpu->hm.s.StatNoFlushTlbWorldSwitch);
1333	else if ( pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_SINGLE_CONTEXT
1334	\|\| pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_SINGLE_CONTEXT_RETAIN_GLOBALS)
1335	{
1336	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushAsid);
1337	}
1338	else
1339	{
1340	Assert(pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_ENTIRE);
1341	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushEntire);
1342	}
1343	#endif
1344	}
1345
1346
1347	/** @name 64-bit guest on 32-bit host OS helper functions.
1348	*
1349	* The host CPU is still 64-bit capable but the host OS is running in 32-bit
1350	* mode (code segment, paging). These wrappers/helpers perform the necessary
1351	* bits for the 32->64 switcher.
1352	*
1353	* @{ */
1354	#if HC_ARCH_BITS == 32 && defined(VBOX_ENABLE_64_BITS_GUESTS)
1355	/**
1356	* Prepares for and executes VMRUN (64-bit guests on a 32-bit host).
1357	*
1358	* @returns VBox status code.
1359	* @param HCPhysVmcbHost Physical address of host VMCB.
1360	* @param HCPhysVmcb Physical address of the VMCB.
1361	* @param pCtx Pointer to the guest-CPU context.
1362	* @param pVM The cross context VM structure.
1363	* @param pVCpu The cross context virtual CPU structure.
1364	*/
1365	DECLASM(int) SVMR0VMSwitcherRun64(RTHCPHYS HCPhysVmcbHost, RTHCPHYS HCPhysVmcb, PCPUMCTX pCtx, PVM pVM, PVMCPU pVCpu)
1366	{
1367	RT_NOREF2(pVM, pCtx);
1368	uint32_t aParam[8];
1369	aParam[0] = RT_LO_U32(HCPhysVmcbHost); /* Param 1: HCPhysVmcbHost - Lo. */
1370	aParam[1] = RT_HI_U32(HCPhysVmcbHost); /* Param 1: HCPhysVmcbHost - Hi. */
1371	aParam[2] = RT_LO_U32(HCPhysVmcb); /* Param 2: HCPhysVmcb - Lo. */
1372	aParam[3] = RT_HI_U32(HCPhysVmcb); /* Param 2: HCPhysVmcb - Hi. */
1373	aParam[4] = VM_RC_ADDR(pVM, pVM);
1374	aParam[5] = 0;
1375	aParam[6] = VM_RC_ADDR(pVM, pVCpu);
1376	aParam[7] = 0;
1377
1378	return SVMR0Execute64BitsHandler(pVCpu, HM64ON32OP_SVMRCVMRun64, RT_ELEMENTS(aParam), &aParam[0]);
1379	}
1380
1381
1382	/**
1383	* Executes the specified VMRUN handler in 64-bit mode.
1384	*
1385	* @returns VBox status code.
1386	* @param pVCpu The cross context virtual CPU structure.
1387	* @param enmOp The operation to perform.
1388	* @param cParams Number of parameters.
1389	* @param paParam Array of 32-bit parameters.
1390	*/
1391	VMMR0DECL(int) SVMR0Execute64BitsHandler(PVMCPU pVCpu, HM64ON32OP enmOp, uint32_t cParams, uint32_t *paParam)
1392	{
1393	PVM pVM = pVCpu->CTX_SUFF(pVM);
1394	AssertReturn(pVM->hm.s.pfnHost32ToGuest64R0, VERR_HM_NO_32_TO_64_SWITCHER);
1395	Assert(enmOp > HM64ON32OP_INVALID && enmOp < HM64ON32OP_END);
1396
1397	/* Disable interrupts. */
1398	RTHCUINTREG const fEFlags = ASMIntDisableFlags();
1399
1400	#ifdef VBOX_WITH_VMMR0_DISABLE_LAPIC_NMI
1401	RTCPUID idHostCpu = RTMpCpuId();
1402	CPUMR0SetLApic(pVCpu, idHostCpu);
1403	#endif
1404
1405	CPUMSetHyperESP(pVCpu, VMMGetStackRC(pVCpu));
1406	CPUMSetHyperEIP(pVCpu, enmOp);
1407	for (int i = (int)cParams - 1; i >= 0; i--)
1408	CPUMPushHyper(pVCpu, paParam[i]);
1409
1410	/* Call the switcher. */
1411	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatWorldSwitch3264, z);
1412	int rc = pVM->hm.s.pfnHost32ToGuest64R0(pVM, RT_OFFSETOF(VM, aCpus[pVCpu->idCpu].cpum) - RT_OFFSETOF(VM, cpum));
1413	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatWorldSwitch3264, z);
1414
1415	/* Restore interrupts. */
1416	ASMSetFlags(fEFlags);
1417	return rc;
1418	}
1419
1420	#endif /* HC_ARCH_BITS == 32 && defined(VBOX_ENABLE_64_BITS_GUESTS) */
1421	/** @} */
1422
1423
1424	/**
1425	* Sets an exception intercept in the specified VMCB.
1426	*
1427	* @param pVmcb Pointer to the VM control block.
1428	* @param uXcpt The exception (X86_XCPT_*).
1429	*/
1430	DECLINLINE(void) hmR0SvmSetXcptIntercept(PSVMVMCB pVmcb, uint8_t uXcpt)
1431	{
1432	if (!(pVmcb->ctrl.u32InterceptXcpt & RT_BIT(uXcpt)))
1433	{
1434	pVmcb->ctrl.u32InterceptXcpt \|= RT_BIT(uXcpt);
1435	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1436	}
1437	}
1438
1439
1440	/**
1441	* Clears an exception intercept in the specified VMCB.
1442	*
1443	* @param pVCpu The cross context virtual CPU structure.
1444	* @param pVmcb Pointer to the VM control block.
1445	* @param uXcpt The exception (X86_XCPT_*).
1446	*
1447	* @remarks This takes into account if we're executing a nested-guest and only
1448	* removes the exception intercept if both the guest -and- nested-guest
1449	* are not intercepting it.
1450	*/
1451	DECLINLINE(void) hmR0SvmClearXcptIntercept(PVMCPU pVCpu, PSVMVMCB pVmcb, uint8_t uXcpt)
1452	{
1453	Assert(uXcpt != X86_XCPT_DB);
1454	Assert(uXcpt != X86_XCPT_AC);
1455	#ifndef HMSVM_ALWAYS_TRAP_ALL_XCPTS
1456	if (pVmcb->ctrl.u32InterceptXcpt & RT_BIT(uXcpt))
1457	{
1458	bool fRemove = true;
1459	# ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1460	/* Only remove the intercept if the nested-guest is also not intercepting it! */
1461	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1462	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
1463	{
1464	PCSVMNESTEDVMCBCACHE pVmcbNstGstCache = hmR0SvmGetNestedVmcbCache(pVCpu);
1465	fRemove = !(pVmcbNstGstCache->u32InterceptXcpt & RT_BIT(uXcpt));
1466	}
1467	# else
1468	RT_NOREF(pVCpu);
1469	# endif
1470	if (fRemove)
1471	{
1472	pVmcb->ctrl.u32InterceptXcpt &= ~RT_BIT(uXcpt);
1473	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1474	}
1475	}
1476	#else
1477	RT_NOREF3(pVCpu, pVmcb, uXcpt);
1478	#endif
1479	}
1480
1481
1482	/**
1483	* Sets a control intercept in the specified VMCB.
1484	*
1485	* @param pVmcb Pointer to the VM control block.
1486	* @param fCtrlIntercept The control intercept (SVM_CTRL_INTERCEPT_*).
1487	*/
1488	DECLINLINE(void) hmR0SvmSetCtrlIntercept(PSVMVMCB pVmcb, uint64_t fCtrlIntercept)
1489	{
1490	if (!(pVmcb->ctrl.u64InterceptCtrl & fCtrlIntercept))
1491	{
1492	pVmcb->ctrl.u64InterceptCtrl \|= fCtrlIntercept;
1493	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1494	}
1495	}
1496
1497
1498	/**
1499	* Clears a control intercept in the specified VMCB.
1500	*
1501	* @returns @c true if the intercept is still set, @c false otherwise.
1502	* @param pVCpu The cross context virtual CPU structure.
1503	* @param pCtx Pointer to the guest-CPU context.
1504	* @param pVmcb Pointer to the VM control block.
1505	* @param fCtrlIntercept The control intercept (SVM_CTRL_INTERCEPT_*).
1506	*
1507	* @remarks This takes into account if we're executing a nested-guest and only
1508	* removes the control intercept if both the guest -and- nested-guest
1509	* are not intercepting it.
1510	*/
1511	DECLINLINE(bool) hmR0SvmClearCtrlIntercept(PVMCPU pVCpu, PCCPUMCTX pCtx, PSVMVMCB pVmcb, uint64_t fCtrlIntercept)
1512	{
1513	if (pVmcb->ctrl.u64InterceptCtrl & fCtrlIntercept)
1514	{
1515	bool fRemove = true;
1516	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1517	/* Only remove the control intercept if the nested-guest is also not intercepting it! */
1518	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
1519	{
1520	PCSVMNESTEDVMCBCACHE pVmcbNstGstCache = hmR0SvmGetNestedVmcbCache(pVCpu);
1521	fRemove = !(pVmcbNstGstCache->u64InterceptCtrl & fCtrlIntercept);
1522	}
1523	#else
1524	RT_NOREF2(pVCpu, pCtx);
1525	#endif
1526	if (fRemove)
1527	{
1528	pVmcb->ctrl.u64InterceptCtrl &= ~fCtrlIntercept;
1529	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1530	}
1531	}
1532
1533	return RT_BOOL(pVmcb->ctrl.u64InterceptCtrl & fCtrlIntercept);
1534	}
1535
1536
1537	/**
1538	* Exports the guest (or nested-guest) CR0 into the VMCB.
1539	*
1540	* @param pVCpu The cross context virtual CPU structure.
1541	* @param pVmcb Pointer to the VM control block.
1542	*
1543	* @remarks This assumes we always pre-load the guest FPU.
1544	* @remarks No-long-jump zone!!!
1545	*/
1546	static void hmR0SvmExportGuestCR0(PVMCPU pVCpu, PSVMVMCB pVmcb)
1547	{
1548	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1549
1550	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1551	uint64_t const uGuestCr0 = pCtx->cr0;
1552	uint64_t uShadowCr0 = uGuestCr0;
1553
1554	/* Always enable caching. */
1555	uShadowCr0 &= ~(X86_CR0_CD \| X86_CR0_NW);
1556
1557	/* When Nested Paging is not available use shadow page tables and intercept #PFs (latter done in SVMR0SetupVM()). */
1558	if (!pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging)
1559	{
1560	uShadowCr0 \|= X86_CR0_PG /* Use shadow page tables. */
1561	\| X86_CR0_WP; /* Guest CPL 0 writes to its read-only pages should cause a #PF #VMEXIT. */
1562	}
1563
1564	/*
1565	* Use the #MF style of legacy-FPU error reporting for now. Although AMD-V has MSRs that
1566	* lets us isolate the host from it, IEM/REM still needs work to emulate it properly,
1567	* see @bugref{7243#c103}.
1568	*/
1569	if (!(uGuestCr0 & X86_CR0_NE))
1570	{
1571	uShadowCr0 \|= X86_CR0_NE;
1572	hmR0SvmSetXcptIntercept(pVmcb, X86_XCPT_MF);
1573	}
1574	else
1575	hmR0SvmClearXcptIntercept(pVCpu, pVmcb, X86_XCPT_MF);
1576
1577	/*
1578	* If the shadow and guest CR0 are identical we can avoid intercepting CR0 reads.
1579	*
1580	* CR0 writes still needs interception as PGM requires tracking paging mode changes,
1581	* see @bugref{6944}.
1582	*
1583	* We also don't ever want to honor weird things like cache disable from the guest.
1584	* However, we can avoid intercepting changes to the TS & MP bits by clearing the CR0
1585	* write intercept below and keeping SVM_CTRL_INTERCEPT_CR0_SEL_WRITE instead.
1586	*/
1587	if (uShadowCr0 == uGuestCr0)
1588	{
1589	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
1590	{
1591	pVmcb->ctrl.u16InterceptRdCRx &= ~RT_BIT(0);
1592	pVmcb->ctrl.u16InterceptWrCRx &= ~RT_BIT(0);
1593	Assert(pVmcb->ctrl.u64InterceptCtrl & SVM_CTRL_INTERCEPT_CR0_SEL_WRITE);
1594	}
1595	else
1596	{
1597	/* If the nested-hypervisor intercepts CR0 reads/writes, we need to continue intercepting them. */
1598	PCSVMNESTEDVMCBCACHE pVmcbNstGstCache = hmR0SvmGetNestedVmcbCache(pVCpu);
1599	pVmcb->ctrl.u16InterceptRdCRx = (pVmcb->ctrl.u16InterceptRdCRx & ~RT_BIT(0))
1600	\| (pVmcbNstGstCache->u16InterceptRdCRx & RT_BIT(0));
1601	pVmcb->ctrl.u16InterceptWrCRx = (pVmcb->ctrl.u16InterceptWrCRx & ~RT_BIT(0))
1602	\| (pVmcbNstGstCache->u16InterceptWrCRx & RT_BIT(0));
1603	}
1604	}
1605	else
1606	{
1607	pVmcb->ctrl.u16InterceptRdCRx \|= RT_BIT(0);
1608	pVmcb->ctrl.u16InterceptWrCRx \|= RT_BIT(0);
1609	}
1610	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1611
1612	Assert(RT_HI_U32(uShadowCr0) == 0);
1613	if (pVmcb->guest.u64CR0 != uShadowCr0)
1614	{
1615	pVmcb->guest.u64CR0 = uShadowCr0;
1616	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1617	}
1618	}
1619
1620
1621	/**
1622	* Exports the guest (or nested-guest) CR3 into the VMCB.
1623	*
1624	* @param pVCpu The cross context virtual CPU structure.
1625	* @param pVmcb Pointer to the VM control block.
1626	*
1627	* @remarks No-long-jump zone!!!
1628	*/
1629	static void hmR0SvmExportGuestCR3(PVMCPU pVCpu, PSVMVMCB pVmcb)
1630	{
1631	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1632
1633	PVM pVM = pVCpu->CTX_SUFF(pVM);
1634	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1635	if (pVM->hm.s.fNestedPaging)
1636	{
1637	PGMMODE enmShwPagingMode;
1638	#if HC_ARCH_BITS == 32
1639	if (CPUMIsGuestInLongModeEx(pCtx))
1640	enmShwPagingMode = PGMMODE_AMD64_NX;
1641	else
1642	#endif
1643	enmShwPagingMode = PGMGetHostMode(pVM);
1644
1645	pVmcb->ctrl.u64NestedPagingCR3 = PGMGetNestedCR3(pVCpu, enmShwPagingMode);
1646	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
1647	pVmcb->guest.u64CR3 = pCtx->cr3;
1648	Assert(pVmcb->ctrl.u64NestedPagingCR3);
1649	}
1650	else
1651	pVmcb->guest.u64CR3 = PGMGetHyperCR3(pVCpu);
1652
1653	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1654	}
1655
1656
1657	/**
1658	* Exports the guest (or nested-guest) CR4 into the VMCB.
1659	*
1660	* @param pVCpu The cross context virtual CPU structure.
1661	* @param pVmcb Pointer to the VM control block.
1662	*
1663	* @remarks No-long-jump zone!!!
1664	*/
1665	static int hmR0SvmExportGuestCR4(PVMCPU pVCpu, PSVMVMCB pVmcb)
1666	{
1667	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1668
1669	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1670	uint64_t uShadowCr4 = pCtx->cr4;
1671	if (!pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging)
1672	{
1673	switch (pVCpu->hm.s.enmShadowMode)
1674	{
1675	case PGMMODE_REAL:
1676	case PGMMODE_PROTECTED: /* Protected mode, no paging. */
1677	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1678
1679	case PGMMODE_32_BIT: /* 32-bit paging. */
1680	uShadowCr4 &= ~X86_CR4_PAE;
1681	break;
1682
1683	case PGMMODE_PAE: /* PAE paging. */
1684	case PGMMODE_PAE_NX: /* PAE paging with NX enabled. */
1685	/** Must use PAE paging as we could use physical memory > 4 GB */
1686	uShadowCr4 \|= X86_CR4_PAE;
1687	break;
1688
1689	case PGMMODE_AMD64: /* 64-bit AMD paging (long mode). */
1690	case PGMMODE_AMD64_NX: /* 64-bit AMD paging (long mode) with NX enabled. */
1691	#ifdef VBOX_ENABLE_64_BITS_GUESTS
1692	break;
1693	#else
1694	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1695	#endif
1696
1697	default: /* shut up gcc */
1698	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1699	}
1700	}
1701
1702	/* Whether to save/load/restore XCR0 during world switch depends on CR4.OSXSAVE and host+guest XCR0. */
1703	pVCpu->hm.s.fLoadSaveGuestXcr0 = (pCtx->cr4 & X86_CR4_OSXSAVE) && pCtx->aXcr[0] != ASMGetXcr0();
1704
1705	/* Avoid intercepting CR4 reads if the guest and shadow CR4 values are identical. */
1706	if (uShadowCr4 == pCtx->cr4)
1707	{
1708	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
1709	pVmcb->ctrl.u16InterceptRdCRx &= ~RT_BIT(4);
1710	else
1711	{
1712	/* If the nested-hypervisor intercepts CR4 reads, we need to continue intercepting them. */
1713	PCSVMNESTEDVMCBCACHE pVmcbNstGstCache = hmR0SvmGetNestedVmcbCache(pVCpu);
1714	pVmcb->ctrl.u16InterceptRdCRx = (pVmcb->ctrl.u16InterceptRdCRx & ~RT_BIT(4))
1715	\| (pVmcbNstGstCache->u16InterceptRdCRx & RT_BIT(4));
1716	}
1717	}
1718	else
1719	pVmcb->ctrl.u16InterceptRdCRx \|= RT_BIT(4);
1720
1721	/* CR4 writes are always intercepted (both guest, nested-guest) for tracking PGM mode changes. */
1722	Assert(pVmcb->ctrl.u16InterceptWrCRx & RT_BIT(4));
1723
1724	/* Update VMCB with the shadow CR4 the appropriate VMCB clean bits. */
1725	Assert(RT_HI_U32(uShadowCr4) == 0);
1726	pVmcb->guest.u64CR4 = uShadowCr4;
1727	pVmcb->ctrl.u32VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_CRX_EFER \| HMSVM_VMCB_CLEAN_INTERCEPTS);
1728
1729	return VINF_SUCCESS;
1730	}
1731
1732
1733	/**
1734	* Exports the guest (or nested-guest) control registers into the VMCB.
1735	*
1736	* @returns VBox status code.
1737	* @param pVCpu The cross context virtual CPU structure.
1738	* @param pVmcb Pointer to the VM control block.
1739	*
1740	* @remarks No-long-jump zone!!!
1741	*/
1742	static int hmR0SvmExportGuestControlRegs(PVMCPU pVCpu, PSVMVMCB pVmcb)
1743	{
1744	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1745
1746	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_CR_MASK)
1747	{
1748	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_CR0)
1749	hmR0SvmExportGuestCR0(pVCpu, pVmcb);
1750
1751	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_CR2)
1752	{
1753	pVmcb->guest.u64CR2 = pVCpu->cpum.GstCtx.cr2;
1754	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CR2;
1755	}
1756
1757	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_CR3)
1758	hmR0SvmExportGuestCR3(pVCpu, pVmcb);
1759
1760	/* CR4 re-loading is ASSUMED to be done everytime we get in from ring-3! (XCR0) */
1761	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_CR4)
1762	{
1763	int rc = hmR0SvmExportGuestCR4(pVCpu, pVmcb);
1764	if (RT_FAILURE(rc))
1765	return rc;
1766	}
1767
1768	pVCpu->hm.s.fCtxChanged &= ~HM_CHANGED_GUEST_CR_MASK;
1769	}
1770	return VINF_SUCCESS;
1771	}
1772
1773
1774	/**
1775	* Exports the guest (or nested-guest) segment registers into the VMCB.
1776	*
1777	* @returns VBox status code.
1778	* @param pVCpu The cross context virtual CPU structure.
1779	* @param pVmcb Pointer to the VM control block.
1780	*
1781	* @remarks No-long-jump zone!!!
1782	*/
1783	static void hmR0SvmExportGuestSegmentRegs(PVMCPU pVCpu, PSVMVMCB pVmcb)
1784	{
1785	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1786	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1787
1788	/* Guest segment registers. */
1789	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SREG_MASK)
1790	{
1791	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_CS)
1792	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, CS, cs);
1793
1794	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SS)
1795	{
1796	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, SS, ss);
1797	pVmcb->guest.u8CPL = pCtx->ss.Attr.n.u2Dpl;
1798	}
1799
1800	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_DS)
1801	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, DS, ds);
1802
1803	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_ES)
1804	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, ES, es);
1805
1806	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_FS)
1807	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, FS, fs);
1808
1809	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_GS)
1810	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, GS, gs);
1811
1812	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_SEG;
1813	}
1814
1815	/* Guest TR. */
1816	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_TR)
1817	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, TR, tr);
1818
1819	/* Guest LDTR. */
1820	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_LDTR)
1821	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, LDTR, ldtr);
1822
1823	/* Guest GDTR. */
1824	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_GDTR)
1825	{
1826	pVmcb->guest.GDTR.u32Limit = pCtx->gdtr.cbGdt;
1827	pVmcb->guest.GDTR.u64Base = pCtx->gdtr.pGdt;
1828	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DT;
1829	}
1830
1831	/* Guest IDTR. */
1832	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_IDTR)
1833	{
1834	pVmcb->guest.IDTR.u32Limit = pCtx->idtr.cbIdt;
1835	pVmcb->guest.IDTR.u64Base = pCtx->idtr.pIdt;
1836	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DT;
1837	}
1838
1839	pVCpu->hm.s.fCtxChanged &= ~( HM_CHANGED_GUEST_SREG_MASK
1840	\| HM_CHANGED_GUEST_TABLE_MASK);
1841	}
1842
1843
1844	/**
1845	* Exports the guest (or nested-guest) MSRs into the VMCB.
1846	*
1847	* @param pVCpu The cross context virtual CPU structure.
1848	* @param pVmcb Pointer to the VM control block.
1849	*
1850	* @remarks No-long-jump zone!!!
1851	*/
1852	static void hmR0SvmExportGuestMsrs(PVMCPU pVCpu, PSVMVMCB pVmcb)
1853	{
1854	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1855	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1856
1857	/* Guest Sysenter MSRs. */
1858	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SYSENTER_MSR_MASK)
1859	{
1860	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SYSENTER_CS_MSR)
1861	pVmcb->guest.u64SysEnterCS = pCtx->SysEnter.cs;
1862
1863	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SYSENTER_EIP_MSR)
1864	pVmcb->guest.u64SysEnterEIP = pCtx->SysEnter.eip;
1865
1866	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SYSENTER_ESP_MSR)
1867	pVmcb->guest.u64SysEnterESP = pCtx->SysEnter.esp;
1868	}
1869
1870	/*
1871	* Guest EFER MSR.
1872	* AMD-V requires guest EFER.SVME to be set. Weird.
1873	* See AMD spec. 15.5.1 "Basic Operation" \| "Canonicalization and Consistency Checks".
1874	*/
1875	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_EFER_MSR)
1876	{
1877	pVmcb->guest.u64EFER = pCtx->msrEFER \| MSR_K6_EFER_SVME;
1878	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1879	}
1880
1881	/* If the guest isn't in 64-bit mode, clear MSR_K6_LME bit, otherwise SVM expects amd64 shadow paging. */
1882	if ( !CPUMIsGuestInLongModeEx(pCtx)
1883	&& (pCtx->msrEFER & MSR_K6_EFER_LME))
1884	{
1885	pVmcb->guest.u64EFER &= ~MSR_K6_EFER_LME;
1886	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1887	}
1888
1889	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SYSCALL_MSRS)
1890	{
1891	pVmcb->guest.u64STAR = pCtx->msrSTAR;
1892	pVmcb->guest.u64LSTAR = pCtx->msrLSTAR;
1893	pVmcb->guest.u64CSTAR = pCtx->msrCSTAR;
1894	pVmcb->guest.u64SFMASK = pCtx->msrSFMASK;
1895	}
1896
1897	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_KERNEL_GS_BASE)
1898	pVmcb->guest.u64KernelGSBase = pCtx->msrKERNELGSBASE;
1899
1900	pVCpu->hm.s.fCtxChanged &= ~( HM_CHANGED_GUEST_SYSENTER_MSR_MASK
1901	\| HM_CHANGED_GUEST_EFER_MSR
1902	\| HM_CHANGED_GUEST_SYSCALL_MSRS
1903	\| HM_CHANGED_GUEST_KERNEL_GS_BASE);
1904
1905	/*
1906	* Setup the PAT MSR (applicable for Nested Paging only).
1907	*
1908	* While guests can modify and see the modified values through the shadow values,
1909	* we shall not honor any guest modifications of this MSR to ensure caching is always
1910	* enabled similar to how we clear CR0.CD and NW bits.
1911	*
1912	* For nested-guests this needs to always be set as well, see @bugref{7243#c109}.
1913	*/
1914	pVmcb->guest.u64PAT = MSR_IA32_CR_PAT_INIT_VAL;
1915
1916	/* Enable the last branch record bit if LBR virtualization is enabled. */
1917	if (pVmcb->ctrl.LbrVirt.n.u1LbrVirt)
1918	pVmcb->guest.u64DBGCTL = MSR_IA32_DEBUGCTL_LBR;
1919	}
1920
1921
1922	/**
1923	* Exports the guest (or nested-guest) debug state into the VMCB and programs
1924	* the necessary intercepts accordingly.
1925	*
1926	* @param pVCpu The cross context virtual CPU structure.
1927	* @param pVmcb Pointer to the VM control block.
1928	*
1929	* @remarks No-long-jump zone!!!
1930	* @remarks Requires EFLAGS to be up-to-date in the VMCB!
1931	*/
1932	static void hmR0SvmExportSharedDebugState(PVMCPU pVCpu, PSVMVMCB pVmcb)
1933	{
1934	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1935
1936	/*
1937	* Anyone single stepping on the host side? If so, we'll have to use the
1938	* trap flag in the guest EFLAGS since AMD-V doesn't have a trap flag on
1939	* the VMM level like the VT-x implementations does.
1940	*/
1941	bool fInterceptMovDRx = false;
1942	bool const fStepping = pVCpu->hm.s.fSingleInstruction \|\| DBGFIsStepping(pVCpu);
1943	if (fStepping)
1944	{
1945	pVCpu->hm.s.fClearTrapFlag = true;
1946	pVmcb->guest.u64RFlags \|= X86_EFL_TF;
1947	fInterceptMovDRx = true; /* Need clean DR6, no guest mess. */
1948	}
1949
1950	if ( fStepping
1951	\|\| (CPUMGetHyperDR7(pVCpu) & X86_DR7_ENABLED_MASK))
1952	{
1953	/*
1954	* Use the combined guest and host DRx values found in the hypervisor
1955	* register set because the debugger has breakpoints active or someone
1956	* is single stepping on the host side.
1957	*
1958	* Note! DBGF expects a clean DR6 state before executing guest code.
1959	*/
1960	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS)
1961	if ( CPUMIsGuestInLongModeEx(pCtx)
1962	&& !CPUMIsHyperDebugStateActivePending(pVCpu))
1963	{
1964	CPUMR0LoadHyperDebugState(pVCpu, false /* include DR6 */);
1965	Assert(!CPUMIsGuestDebugStateActivePending(pVCpu));
1966	Assert(CPUMIsHyperDebugStateActivePending(pVCpu));
1967	}
1968	else
1969	#endif
1970	if (!CPUMIsHyperDebugStateActive(pVCpu))
1971	{
1972	CPUMR0LoadHyperDebugState(pVCpu, false /* include DR6 */);
1973	Assert(!CPUMIsGuestDebugStateActive(pVCpu));
1974	Assert(CPUMIsHyperDebugStateActive(pVCpu));
1975	}
1976
1977	/* Update DR6 & DR7. (The other DRx values are handled by CPUM one way or the other.) */
1978	if ( pVmcb->guest.u64DR6 != X86_DR6_INIT_VAL
1979	\|\| pVmcb->guest.u64DR7 != CPUMGetHyperDR7(pVCpu))
1980	{
1981	pVmcb->guest.u64DR7 = CPUMGetHyperDR7(pVCpu);
1982	pVmcb->guest.u64DR6 = X86_DR6_INIT_VAL;
1983	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
1984	}
1985
1986	/** @todo If we cared, we could optimize to allow the guest to read registers
1987	* with the same values. */
1988	fInterceptMovDRx = true;
1989	pVCpu->hm.s.fUsingHyperDR7 = true;
1990	Log5(("hmR0SvmExportSharedDebugState: Loaded hyper DRx\n"));
1991	}
1992	else
1993	{
1994	/*
1995	* Update DR6, DR7 with the guest values if necessary.
1996	*/
1997	if ( pVmcb->guest.u64DR7 != pCtx->dr[7]
1998	\|\| pVmcb->guest.u64DR6 != pCtx->dr[6])
1999	{
2000	pVmcb->guest.u64DR7 = pCtx->dr[7];
2001	pVmcb->guest.u64DR6 = pCtx->dr[6];
2002	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
2003	}
2004	pVCpu->hm.s.fUsingHyperDR7 = false;
2005
2006	/*
2007	* If the guest has enabled debug registers, we need to load them prior to
2008	* executing guest code so they'll trigger at the right time.
2009	*/
2010	if (pCtx->dr[7] & (X86_DR7_ENABLED_MASK \| X86_DR7_GD)) /** @todo Why GD? */
2011	{
2012	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS)
2013	if ( CPUMIsGuestInLongModeEx(pCtx)
2014	&& !CPUMIsGuestDebugStateActivePending(pVCpu))
2015	{
2016	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
2017	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxArmed);
2018	Assert(!CPUMIsHyperDebugStateActivePending(pVCpu));
2019	Assert(CPUMIsGuestDebugStateActivePending(pVCpu));
2020	}
2021	else
2022	#endif
2023	if (!CPUMIsGuestDebugStateActive(pVCpu))
2024	{
2025	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
2026	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxArmed);
2027	Assert(!CPUMIsHyperDebugStateActive(pVCpu));
2028	Assert(CPUMIsGuestDebugStateActive(pVCpu));
2029	}
2030	Log5(("hmR0SvmExportSharedDebugState: Loaded guest DRx\n"));
2031	}
2032	/*
2033	* If no debugging enabled, we'll lazy load DR0-3. We don't need to
2034	* intercept #DB as DR6 is updated in the VMCB.
2035	*
2036	* Note! If we cared and dared, we could skip intercepting \#DB here.
2037	* However, \#DB shouldn't be performance critical, so we'll play safe
2038	* and keep the code similar to the VT-x code and always intercept it.
2039	*/
2040	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS)
2041	else if ( !CPUMIsGuestDebugStateActivePending(pVCpu)
2042	&& !CPUMIsGuestDebugStateActive(pVCpu))
2043	#else
2044	else if (!CPUMIsGuestDebugStateActive(pVCpu))
2045	#endif
2046	{
2047	fInterceptMovDRx = true;
2048	}
2049	}
2050
2051	Assert(pVmcb->ctrl.u32InterceptXcpt & RT_BIT_32(X86_XCPT_DB));
2052	if (fInterceptMovDRx)
2053	{
2054	if ( pVmcb->ctrl.u16InterceptRdDRx != 0xffff
2055	\|\| pVmcb->ctrl.u16InterceptWrDRx != 0xffff)
2056	{
2057	pVmcb->ctrl.u16InterceptRdDRx = 0xffff;
2058	pVmcb->ctrl.u16InterceptWrDRx = 0xffff;
2059	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2060	}
2061	}
2062	else
2063	{
2064	if ( pVmcb->ctrl.u16InterceptRdDRx
2065	\|\| pVmcb->ctrl.u16InterceptWrDRx)
2066	{
2067	pVmcb->ctrl.u16InterceptRdDRx = 0;
2068	pVmcb->ctrl.u16InterceptWrDRx = 0;
2069	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2070	}
2071	}
2072	Log4Func(("DR6=%#RX64 DR7=%#RX64\n", pCtx->dr[6], pCtx->dr[7]));
2073	}
2074
2075	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
2076	/**
2077	* Exports the nested-guest hardware virtualization state into the nested-guest
2078	* VMCB.
2079	*
2080	* @param pVCpu The cross context virtual CPU structure.
2081	* @param pVmcbNstGst Pointer to the nested-guest VM control block.
2082	*
2083	* @remarks No-long-jump zone!!!
2084	*/
2085	static void hmR0SvmExportGuestHwvirtStateNested(PVMCPU pVCpu, PSVMVMCB pVmcbNstGst)
2086	{
2087	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2088
2089	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_HWVIRT)
2090	{
2091	/*
2092	* Ensure the nested-guest pause-filter counters don't exceed the outer guest values esp.
2093	* since SVM doesn't have a preemption timer.
2094	*
2095	* We do this here rather than in hmR0SvmSetupVmcbNested() as we may have been executing the
2096	* nested-guest in IEM incl. PAUSE instructions which would update the pause-filter counters
2097	* and may continue execution in SVM R0 without a nested-guest #VMEXIT in between.
2098	*/
2099	PVM pVM = pVCpu->CTX_SUFF(pVM);
2100	PSVMVMCBCTRL pVmcbNstGstCtrl = &pVmcbNstGst->ctrl;
2101	uint16_t const uGuestPauseFilterCount = pVM->hm.s.svm.cPauseFilter;
2102	uint16_t const uGuestPauseFilterThreshold = pVM->hm.s.svm.cPauseFilterThresholdTicks;
2103	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_PAUSE))
2104	{
2105	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2106	pVmcbNstGstCtrl->u16PauseFilterCount = RT_MIN(pCtx->hwvirt.svm.cPauseFilter, uGuestPauseFilterCount);
2107	pVmcbNstGstCtrl->u16PauseFilterThreshold = RT_MIN(pCtx->hwvirt.svm.cPauseFilterThreshold, uGuestPauseFilterThreshold);
2108	pVmcbNstGstCtrl->u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2109	}
2110	else
2111	{
2112	pVmcbNstGstCtrl->u16PauseFilterCount = uGuestPauseFilterCount;
2113	pVmcbNstGstCtrl->u16PauseFilterThreshold = uGuestPauseFilterThreshold;
2114	}
2115
2116	pVCpu->hm.s.fCtxChanged &= ~HM_CHANGED_GUEST_HWVIRT;
2117	}
2118	}
2119	#endif
2120
2121	/**
2122	* Exports the guest APIC TPR state into the VMCB.
2123	*
2124	* @returns VBox status code.
2125	* @param pVCpu The cross context virtual CPU structure.
2126	* @param pVmcb Pointer to the VM control block.
2127	*/
2128	static int hmR0SvmExportGuestApicTpr(PVMCPU pVCpu, PSVMVMCB pVmcb)
2129	{
2130	if (ASMAtomicUoReadU64(&pVCpu->hm.s.fCtxChanged) & HM_CHANGED_GUEST_APIC_TPR)
2131	{
2132	PVM pVM = pVCpu->CTX_SUFF(pVM);
2133	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2134	if ( PDMHasApic(pVM)
2135	&& APICIsEnabled(pVCpu))
2136	{
2137	bool fPendingIntr;
2138	uint8_t u8Tpr;
2139	int rc = APICGetTpr(pVCpu, &u8Tpr, &fPendingIntr, NULL /* pu8PendingIrq */);
2140	AssertRCReturn(rc, rc);
2141
2142	/* Assume that we need to trap all TPR accesses and thus need not check on
2143	every #VMEXIT if we should update the TPR. */
2144	Assert(pVmcb->ctrl.IntCtrl.n.u1VIntrMasking);
2145	pVCpu->hm.s.svm.fSyncVTpr = false;
2146
2147	if (!pVM->hm.s.fTPRPatchingActive)
2148	{
2149	/* Bits 3-0 of the VTPR field correspond to bits 7-4 of the TPR (which is the Task-Priority Class). */
2150	pVmcb->ctrl.IntCtrl.n.u8VTPR = (u8Tpr >> 4);
2151
2152	/* If there are interrupts pending, intercept CR8 writes to evaluate ASAP if we
2153	can deliver the interrupt to the guest. */
2154	if (fPendingIntr)
2155	pVmcb->ctrl.u16InterceptWrCRx \|= RT_BIT(8);
2156	else
2157	{
2158	pVmcb->ctrl.u16InterceptWrCRx &= ~RT_BIT(8);
2159	pVCpu->hm.s.svm.fSyncVTpr = true;
2160	}
2161
2162	pVmcb->ctrl.u32VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS \| HMSVM_VMCB_CLEAN_INT_CTRL);
2163	}
2164	else
2165	{
2166	/* 32-bit guests uses LSTAR MSR for patching guest code which touches the TPR. */
2167	pVmcb->guest.u64LSTAR = u8Tpr;
2168	uint8_t pbMsrBitmap = (uint8_t )pVCpu->hm.s.svm.pvMsrBitmap;
2169
2170	/* If there are interrupts pending, intercept LSTAR writes, otherwise don't intercept reads or writes. */
2171	if (fPendingIntr)
2172	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_INTERCEPT_WRITE);
2173	else
2174	{
2175	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
2176	pVCpu->hm.s.svm.fSyncVTpr = true;
2177	}
2178	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_IOPM_MSRPM;
2179	}
2180	}
2181	ASMAtomicUoAndU64(&pVCpu->hm.s.fCtxChanged, ~HM_CHANGED_GUEST_APIC_TPR);
2182	}
2183	return VINF_SUCCESS;
2184	}
2185
2186
2187	/**
2188	* Sets up the exception interrupts required for guest (or nested-guest)
2189	* execution in the VMCB.
2190	*
2191	* @param pVCpu The cross context virtual CPU structure.
2192	* @param pVmcb Pointer to the VM control block.
2193	*
2194	* @remarks No-long-jump zone!!!
2195	*/
2196	static void hmR0SvmExportGuestXcptIntercepts(PVMCPU pVCpu, PSVMVMCB pVmcb)
2197	{
2198	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2199
2200	/* If we modify intercepts from here, please check & adjust hmR0SvmMergeVmcbCtrlsNested() if required. */
2201	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_SVM_GUEST_XCPT_INTERCEPTS)
2202	{
2203	/* Trap #UD for GIM provider (e.g. for hypercalls). */
2204	if (pVCpu->hm.s.fGIMTrapXcptUD)
2205	hmR0SvmSetXcptIntercept(pVmcb, X86_XCPT_UD);
2206	else
2207	hmR0SvmClearXcptIntercept(pVCpu, pVmcb, X86_XCPT_UD);
2208
2209	/* Trap #BP for INT3 debug breakpoints set by the VM debugger. */
2210	if (pVCpu->CTX_SUFF(pVM)->dbgf.ro.cEnabledInt3Breakpoints)
2211	hmR0SvmSetXcptIntercept(pVmcb, X86_XCPT_BP);
2212	else
2213	hmR0SvmClearXcptIntercept(pVCpu, pVmcb, X86_XCPT_BP);
2214
2215	/* The remaining intercepts are handled elsewhere, e.g. in hmR0SvmExportGuestCR0(). */
2216	pVCpu->hm.s.fCtxChanged &= ~HM_CHANGED_SVM_GUEST_XCPT_INTERCEPTS;
2217	}
2218	}
2219
2220
2221	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
2222	/**
2223	* Merges guest and nested-guest intercepts for executing the nested-guest using
2224	* hardware-assisted SVM.
2225	*
2226	* This merges the guest and nested-guest intercepts in a way that if the outer
2227	* guest intercept is set we need to intercept it in the nested-guest as
2228	* well.
2229	*
2230	* @param pVCpu The cross context virtual CPU structure.
2231	* @param pVmcbNstGst Pointer to the nested-guest VM control block.
2232	* @param pCtx Pointer to the nested-guest-CPU context.
2233	*/
2234	static void hmR0SvmMergeVmcbCtrlsNested(PVMCPU pVCpu, PCCPUMCTX pCtx)
2235	{
2236	PVM pVM = pVCpu->CTX_SUFF(pVM);
2237	PCSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
2238	PSVMVMCB pVmcbNstGst = pCtx->hwvirt.svm.CTX_SUFF(pVmcb);
2239	PSVMVMCBCTRL pVmcbNstGstCtrl = &pVmcbNstGst->ctrl;
2240
2241	/* Merge the guest's CR intercepts into the nested-guest VMCB. */
2242	pVmcbNstGstCtrl->u16InterceptRdCRx \|= pVmcb->ctrl.u16InterceptRdCRx;
2243	pVmcbNstGstCtrl->u16InterceptWrCRx \|= pVmcb->ctrl.u16InterceptWrCRx;
2244
2245	/* Always intercept CR4 writes for tracking PGM mode changes. */
2246	pVmcbNstGstCtrl->u16InterceptWrCRx \|= RT_BIT(4);
2247
2248	/* Without nested paging, intercept CR3 reads and writes as we load shadow page tables. */
2249	if (!pVM->hm.s.fNestedPaging)
2250	{
2251	pVmcbNstGstCtrl->u16InterceptRdCRx \|= RT_BIT(3);
2252	pVmcbNstGstCtrl->u16InterceptWrCRx \|= RT_BIT(3);
2253	}
2254
2255	/** @todo Figure out debugging with nested-guests, till then just intercept
2256	* all DR[0-15] accesses. */
2257	pVmcbNstGstCtrl->u16InterceptRdDRx \|= 0xffff;
2258	pVmcbNstGstCtrl->u16InterceptWrDRx \|= 0xffff;
2259
2260	/*
2261	* Merge the guest's exception intercepts into the nested-guest VMCB.
2262	*
2263	* - \#UD: Exclude these as the outer guest's GIM hypercalls are not applicable
2264	* while executing the nested-guest.
2265	*
2266	* - \#BP: Exclude breakpoints set by the VM debugger for the outer guest. This can
2267	* be tweaked later depending on how we wish to implement breakpoints.
2268	*
2269	* Warning!! This ASSUMES we only intercept \#UD for hypercall purposes and \#BP
2270	* for VM debugger breakpoints, see hmR0SvmExportGuestXcptIntercepts().
2271	*/
2272	#ifndef HMSVM_ALWAYS_TRAP_ALL_XCPTS
2273	pVmcbNstGstCtrl->u32InterceptXcpt \|= (pVmcb->ctrl.u32InterceptXcpt & ~( RT_BIT(X86_XCPT_UD)
2274	\| RT_BIT(X86_XCPT_BP)));
2275	#else
2276	pVmcbNstGstCtrl->u32InterceptXcpt \|= pVmcb->ctrl.u32InterceptXcpt;
2277	#endif
2278
2279	/*
2280	* Adjust intercepts while executing the nested-guest that differ from the
2281	* outer guest intercepts.
2282	*
2283	* - VINTR: Exclude the outer guest intercept as we don't need to cause VINTR #VMEXITs
2284	* that belong to the nested-guest to the outer guest.
2285	*
2286	* - VMMCALL: Exclude the outer guest intercept as when it's also not intercepted by
2287	* the nested-guest, the physical CPU raises a \#UD exception as expected.
2288	*/
2289	pVmcbNstGstCtrl->u64InterceptCtrl \|= (pVmcb->ctrl.u64InterceptCtrl & ~( SVM_CTRL_INTERCEPT_VINTR
2290	\| SVM_CTRL_INTERCEPT_VMMCALL))
2291	\| HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS;
2292
2293	Assert( (pVmcbNstGstCtrl->u64InterceptCtrl & HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS)
2294	== HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS);
2295
2296	/* Finally, update the VMCB clean bits. */
2297	pVmcbNstGstCtrl->u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2298	}
2299	#endif
2300
2301
2302	/**
2303	* Selects the appropriate function to run guest code.
2304	*
2305	* @returns VBox status code.
2306	* @param pVCpu The cross context virtual CPU structure.
2307	*
2308	* @remarks No-long-jump zone!!!
2309	*/
2310	static int hmR0SvmSelectVMRunHandler(PVMCPU pVCpu)
2311	{
2312	if (CPUMIsGuestInLongMode(pVCpu))
2313	{
2314	#ifndef VBOX_ENABLE_64_BITS_GUESTS
2315	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
2316	#endif
2317	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.fAllow64BitGuests); /* Guaranteed by hmR3InitFinalizeR0(). */
2318	#if HC_ARCH_BITS == 32
2319	/* 32-bit host. We need to switch to 64-bit before running the 64-bit guest. */
2320	pVCpu->hm.s.svm.pfnVMRun = SVMR0VMSwitcherRun64;
2321	#else
2322	/* 64-bit host or hybrid host. */
2323	pVCpu->hm.s.svm.pfnVMRun = SVMR0VMRun64;
2324	#endif
2325	}
2326	else
2327	{
2328	/* Guest is not in long mode, use the 32-bit handler. */
2329	pVCpu->hm.s.svm.pfnVMRun = SVMR0VMRun;
2330	}
2331	return VINF_SUCCESS;
2332	}
2333
2334
2335	/**
2336	* Enters the AMD-V session.
2337	*
2338	* @returns VBox status code.
2339	* @param pVM The cross context VM structure.
2340	* @param pVCpu The cross context virtual CPU structure.
2341	* @param pCpu Pointer to the CPU info struct.
2342	*/
2343	VMMR0DECL(int) SVMR0Enter(PVM pVM, PVMCPU pVCpu, PHMGLOBALCPUINFO pCpu)
2344	{
2345	AssertPtr(pVM);
2346	AssertPtr(pVCpu);
2347	Assert(pVM->hm.s.svm.fSupported);
2348	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2349	NOREF(pVM); NOREF(pCpu);
2350
2351	LogFlowFunc(("pVM=%p pVCpu=%p\n", pVM, pVCpu));
2352	Assert((pVCpu->hm.s.fCtxChanged & (HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE))
2353	== (HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE));
2354
2355	pVCpu->hm.s.fLeaveDone = false;
2356	return VINF_SUCCESS;
2357	}
2358
2359
2360	/**
2361	* Thread-context callback for AMD-V.
2362	*
2363	* @param enmEvent The thread-context event.
2364	* @param pVCpu The cross context virtual CPU structure.
2365	* @param fGlobalInit Whether global VT-x/AMD-V init. is used.
2366	* @thread EMT(pVCpu)
2367	*/
2368	VMMR0DECL(void) SVMR0ThreadCtxCallback(RTTHREADCTXEVENT enmEvent, PVMCPU pVCpu, bool fGlobalInit)
2369	{
2370	NOREF(fGlobalInit);
2371
2372	switch (enmEvent)
2373	{
2374	case RTTHREADCTXEVENT_OUT:
2375	{
2376	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2377	Assert(VMMR0ThreadCtxHookIsEnabled(pVCpu));
2378	VMCPU_ASSERT_EMT(pVCpu);
2379
2380	/* No longjmps (log-flush, locks) in this fragile context. */
2381	VMMRZCallRing3Disable(pVCpu);
2382
2383	if (!pVCpu->hm.s.fLeaveDone)
2384	{
2385	hmR0SvmLeave(pVCpu, false /* fImportState */);
2386	pVCpu->hm.s.fLeaveDone = true;
2387	}
2388
2389	/* Leave HM context, takes care of local init (term). */
2390	int rc = HMR0LeaveCpu(pVCpu);
2391	AssertRC(rc); NOREF(rc);
2392
2393	/* Restore longjmp state. */
2394	VMMRZCallRing3Enable(pVCpu);
2395	STAM_REL_COUNTER_INC(&pVCpu->hm.s.StatSwitchPreempt);
2396	break;
2397	}
2398
2399	case RTTHREADCTXEVENT_IN:
2400	{
2401	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2402	Assert(VMMR0ThreadCtxHookIsEnabled(pVCpu));
2403	VMCPU_ASSERT_EMT(pVCpu);
2404
2405	/* No longjmps (log-flush, locks) in this fragile context. */
2406	VMMRZCallRing3Disable(pVCpu);
2407
2408	/*
2409	* Initialize the bare minimum state required for HM. This takes care of
2410	* initializing AMD-V if necessary (onlined CPUs, local init etc.)
2411	*/
2412	int rc = HMR0EnterCpu(pVCpu);
2413	AssertRC(rc); NOREF(rc);
2414	Assert((pVCpu->hm.s.fCtxChanged & (HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE))
2415	== (HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE));
2416
2417	pVCpu->hm.s.fLeaveDone = false;
2418
2419	/* Restore longjmp state. */
2420	VMMRZCallRing3Enable(pVCpu);
2421	break;
2422	}
2423
2424	default:
2425	break;
2426	}
2427	}
2428
2429
2430	/**
2431	* Saves the host state.
2432	*
2433	* @returns VBox status code.
2434	* @param pVCpu The cross context virtual CPU structure.
2435	*
2436	* @remarks No-long-jump zone!!!
2437	*/
2438	VMMR0DECL(int) SVMR0ExportHostState(PVMCPU pVCpu)
2439	{
2440	NOREF(pVCpu);
2441
2442	/* Nothing to do here. AMD-V does this for us automatically during the world-switch. */
2443	ASMAtomicUoAndU64(&pVCpu->hm.s.fCtxChanged, ~HM_CHANGED_HOST_CONTEXT);
2444	return VINF_SUCCESS;
2445	}
2446
2447
2448	/**
2449	* Exports the guest state from the guest-CPU context into the VMCB.
2450	*
2451	* The CPU state will be loaded from these fields on every successful VM-entry.
2452	* Also sets up the appropriate VMRUN function to execute guest code based on
2453	* the guest CPU mode.
2454	*
2455	* @returns VBox status code.
2456	* @param pVM The cross context VM structure.
2457	* @param pVCpu The cross context virtual CPU structure.
2458	* @param pCtx Pointer to the guest-CPU context.
2459	*
2460	* @remarks No-long-jump zone!!!
2461	*/
2462	static int hmR0SvmExportGuestState(PVMCPU pVCpu)
2463	{
2464	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatExportGuestState, x);
2465
2466	PSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
2467	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2468
2469	Assert(pVmcb);
2470	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(pCtx);
2471
2472	pVmcb->guest.u64RIP = pCtx->rip;
2473	pVmcb->guest.u64RSP = pCtx->rsp;
2474	pVmcb->guest.u64RFlags = pCtx->eflags.u32;
2475	pVmcb->guest.u64RAX = pCtx->rax;
2476	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
2477	if (pVmcb->ctrl.IntCtrl.n.u1VGifEnable)
2478	{
2479	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_VGIF);
2480	pVmcb->ctrl.IntCtrl.n.u1VGif = pCtx->hwvirt.fGif;
2481	}
2482	#endif
2483
2484	RTCCUINTREG const fEFlags = ASMIntDisableFlags();
2485
2486	int rc = hmR0SvmExportGuestControlRegs(pVCpu, pVmcb);
2487	AssertRCReturnStmt(rc, ASMSetFlags(fEFlags), rc);
2488
2489	hmR0SvmExportGuestSegmentRegs(pVCpu, pVmcb);
2490	hmR0SvmExportGuestMsrs(pVCpu, pVmcb);
2491	hmR0SvmExportGuestXcptIntercepts(pVCpu, pVmcb);
2492
2493	ASMSetFlags(fEFlags);
2494
2495	/* hmR0SvmExportGuestApicTpr() must be called -after- hmR0SvmExportGuestMsrs() as we
2496	otherwise we would overwrite the LSTAR MSR that we use for TPR patching. */
2497	hmR0SvmExportGuestApicTpr(pVCpu, pVmcb);
2498
2499	rc = hmR0SvmSelectVMRunHandler(pVCpu);
2500	AssertRCReturn(rc, rc);
2501
2502	/* Clear any bits that may be set but exported unconditionally or unused/reserved bits. */
2503	ASMAtomicUoAndU64(&pVCpu->hm.s.fCtxChanged, ~( HM_CHANGED_GUEST_RIP
2504	\| HM_CHANGED_GUEST_RFLAGS
2505	\| HM_CHANGED_GUEST_GPRS_MASK
2506	\| HM_CHANGED_GUEST_X87
2507	\| HM_CHANGED_GUEST_SSE_AVX
2508	\| HM_CHANGED_GUEST_OTHER_XSAVE
2509	\| HM_CHANGED_GUEST_XCRx
2510	\| HM_CHANGED_GUEST_TSC_AUX
2511	\| HM_CHANGED_GUEST_OTHER_MSRS
2512	\| HM_CHANGED_GUEST_HWVIRT
2513	\| (HM_CHANGED_KEEPER_STATE_MASK & ~HM_CHANGED_SVM_GUEST_XCPT_INTERCEPTS)));
2514
2515	#ifdef VBOX_STRICT
2516	/*
2517	* All of the guest-CPU state and SVM keeper bits should be exported here by now,
2518	* except for the host-context and/or shared host-guest context bits.
2519	*/
2520	uint64_t const fCtxChanged = ASMAtomicUoReadU64(&pVCpu->hm.s.fCtxChanged);
2521	AssertMsg(!(fCtxChanged & (HM_CHANGED_ALL_GUEST & ~HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE)),
2522	("fCtxChanged=%#RX64\n", fCtxChanged));
2523
2524	/*
2525	* If we need to log state that isn't always imported, we'll need to import them here.
2526	* See hmR0SvmPostRunGuest() for which part of the state is imported uncondtionally.
2527	*/
2528	hmR0SvmLogState(pVCpu, pVmcb, "hmR0SvmExportGuestState", 0 /* fFlags /, 0 / uVerbose */);
2529	#endif
2530
2531	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExportGuestState, x);
2532	return VINF_SUCCESS;
2533	}
2534
2535
2536	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
2537	/**
2538	* Merges the guest and nested-guest MSR permission bitmap.
2539	*
2540	* If the guest is intercepting an MSR we need to intercept it regardless of
2541	* whether the nested-guest is intercepting it or not.
2542	*
2543	* @param pHostCpu Pointer to the physical CPU HM info. struct.
2544	* @param pVCpu The cross context virtual CPU structure.
2545	* @param pCtx Pointer to the nested-guest-CPU context.
2546	*
2547	* @remarks No-long-jmp zone!!!
2548	*/
2549	DECLINLINE(void) hmR0SvmMergeMsrpmNested(PHMGLOBALCPUINFO pHostCpu, PVMCPU pVCpu, PCCPUMCTX pCtx)
2550	{
2551	uint64_t const pu64GstMsrpm = (uint64_t const )pVCpu->hm.s.svm.pvMsrBitmap;
2552	uint64_t const pu64NstGstMsrpm = (uint64_t const )pCtx->hwvirt.svm.CTX_SUFF(pvMsrBitmap);
2553	uint64_t pu64DstMsrpm = (uint64_t )pHostCpu->n.svm.pvNstGstMsrpm;
2554
2555	/* MSRPM bytes from offset 0x1800 are reserved, so we stop merging there. */
2556	uint32_t const offRsvdQwords = 0x1800 >> 3;
2557	for (uint32_t i = 0; i < offRsvdQwords; i++)
2558	pu64DstMsrpm[i] = pu64NstGstMsrpm[i] \| pu64GstMsrpm[i];
2559	}
2560
2561
2562	/**
2563	* Caches the nested-guest VMCB fields before we modify them for execution using
2564	* hardware-assisted SVM.
2565	*
2566	* @returns true if the VMCB was previously already cached, false otherwise.
2567	* @param pCtx Pointer to the guest-CPU context.
2568	*
2569	* @sa HMSvmNstGstVmExitNotify.
2570	*/
2571	static bool hmR0SvmCacheVmcbNested(PVMCPU pVCpu, PCCPUMCTX pCtx)
2572	{
2573	/*
2574	* Cache the nested-guest programmed VMCB fields if we have not cached it yet.
2575	* Otherwise we risk re-caching the values we may have modified, see @bugref{7243#c44}.
2576	*
2577	* Nested-paging CR3 is not saved back into the VMCB on #VMEXIT, hence no need to
2578	* cache and restore it, see AMD spec. 15.25.4 "Nested Paging and VMRUN/#VMEXIT".
2579	*/
2580	PSVMNESTEDVMCBCACHE pVmcbNstGstCache = &pVCpu->hm.s.svm.NstGstVmcbCache;
2581	bool const fWasCached = pVmcbNstGstCache->fCacheValid;
2582	if (!fWasCached)
2583	{
2584	PCSVMVMCB pVmcbNstGst = pCtx->hwvirt.svm.CTX_SUFF(pVmcb);
2585	PCSVMVMCBCTRL pVmcbNstGstCtrl = &pVmcbNstGst->ctrl;
2586	pVmcbNstGstCache->u16InterceptRdCRx = pVmcbNstGstCtrl->u16InterceptRdCRx;
2587	pVmcbNstGstCache->u16InterceptWrCRx = pVmcbNstGstCtrl->u16InterceptWrCRx;
2588	pVmcbNstGstCache->u16InterceptRdDRx = pVmcbNstGstCtrl->u16InterceptRdDRx;
2589	pVmcbNstGstCache->u16InterceptWrDRx = pVmcbNstGstCtrl->u16InterceptWrDRx;
2590	pVmcbNstGstCache->u16PauseFilterThreshold = pVmcbNstGstCtrl->u16PauseFilterThreshold;
2591	pVmcbNstGstCache->u16PauseFilterCount = pVmcbNstGstCtrl->u16PauseFilterCount;
2592	pVmcbNstGstCache->u32InterceptXcpt = pVmcbNstGstCtrl->u32InterceptXcpt;
2593	pVmcbNstGstCache->u64InterceptCtrl = pVmcbNstGstCtrl->u64InterceptCtrl;
2594	pVmcbNstGstCache->u64TSCOffset = pVmcbNstGstCtrl->u64TSCOffset;
2595	pVmcbNstGstCache->fVIntrMasking = pVmcbNstGstCtrl->IntCtrl.n.u1VIntrMasking;
2596	pVmcbNstGstCache->fNestedPaging = pVmcbNstGstCtrl->NestedPagingCtrl.n.u1NestedPaging;
2597	pVmcbNstGstCache->fLbrVirt = pVmcbNstGstCtrl->LbrVirt.n.u1LbrVirt;
2598	pVmcbNstGstCache->fCacheValid = true;
2599	Log4Func(("Cached VMCB fields\n"));
2600	}
2601
2602	return fWasCached;
2603	}
2604
2605
2606	/**
2607	* Sets up the nested-guest VMCB for execution using hardware-assisted SVM.
2608	*
2609	* This is done the first time we enter nested-guest execution using SVM R0
2610	* until the nested-guest \#VMEXIT (not to be confused with physical CPU
2611	* \#VMEXITs which may or may not cause a corresponding nested-guest \#VMEXIT).
2612	*
2613	* @param pVCpu The cross context virtual CPU structure.
2614	*/
2615	static void hmR0SvmSetupVmcbNested(PVMCPU pVCpu)
2616	{
2617	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2618	PSVMVMCB pVmcbNstGst = pCtx->hwvirt.svm.CTX_SUFF(pVmcb);
2619	PSVMVMCBCTRL pVmcbNstGstCtrl = &pVmcbNstGst->ctrl;
2620
2621	/*
2622	* First cache the nested-guest VMCB fields we may potentially modify.
2623	*/
2624	bool const fVmcbCached = hmR0SvmCacheVmcbNested(pVCpu, pCtx);
2625	if (!fVmcbCached)
2626	{
2627	/*
2628	* The IOPM of the nested-guest can be ignored because the the guest always
2629	* intercepts all IO port accesses. Thus, we'll swap to the guest IOPM rather
2630	* than the nested-guest IOPM and swap the field back on the #VMEXIT.
2631	*/
2632	pVmcbNstGstCtrl->u64IOPMPhysAddr = g_HCPhysIOBitmap;
2633
2634	/*
2635	* Use the same nested-paging as the outer guest. We can't dynamically switch off
2636	* nested-paging suddenly while executing a VM (see assertion at the end of
2637	* Trap0eHandler() in PGMAllBth.h).
2638	*/
2639	pVmcbNstGstCtrl->NestedPagingCtrl.n.u1NestedPaging = pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging;
2640
2641	/* Always enable V_INTR_MASKING as we do not want to allow access to the physical APIC TPR. */
2642	pVmcbNstGstCtrl->IntCtrl.n.u1VIntrMasking = 1;
2643
2644	/*
2645	* Turn off TPR syncing on #VMEXIT for nested-guests as CR8 intercepts are subject
2646	* to the nested-guest intercepts and we always run with V_INTR_MASKING.
2647	*/
2648	pVCpu->hm.s.svm.fSyncVTpr = false;
2649
2650	#ifdef DEBUG_ramshankar
2651	/* For debugging purposes - copy the LBR info. from outer guest VMCB. */
2652	pVmcbNstGstCtrl->LbrVirt.n.u1LbrVirt = pVmcb->ctrl.LbrVirt.n.u1LbrVirt;
2653	#endif
2654
2655	/*
2656	* If we don't expose Virtualized-VMSAVE/VMLOAD feature to the outer guest, we
2657	* need to intercept VMSAVE/VMLOAD instructions executed by the nested-guest.
2658	*/
2659	if (!pVCpu->CTX_SUFF(pVM)->cpum.ro.GuestFeatures.fSvmVirtVmsaveVmload)
2660	pVmcbNstGstCtrl->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_VMSAVE
2661	\| SVM_CTRL_INTERCEPT_VMLOAD;
2662
2663	/*
2664	* If we don't expose Virtual GIF feature to the outer guest, we need to intercept
2665	* CLGI/STGI instructions executed by the nested-guest.
2666	*/
2667	if (!pVCpu->CTX_SUFF(pVM)->cpum.ro.GuestFeatures.fSvmVGif)
2668	pVmcbNstGstCtrl->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_CLGI
2669	\| SVM_CTRL_INTERCEPT_STGI;
2670
2671	/* Merge the guest and nested-guest intercepts. */
2672	hmR0SvmMergeVmcbCtrlsNested(pVCpu, pCtx);
2673
2674	/* Update the VMCB clean bits. */
2675	pVmcbNstGstCtrl->u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2676	}
2677	else
2678	{
2679	Assert(!pVCpu->hm.s.svm.fSyncVTpr);
2680	Assert(pVmcbNstGstCtrl->u64IOPMPhysAddr == g_HCPhysIOBitmap);
2681	Assert(RT_BOOL(pVmcbNstGstCtrl->NestedPagingCtrl.n.u1NestedPaging) == pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
2682	}
2683	}
2684
2685
2686	/**
2687	* Exports the nested-guest state into the VMCB.
2688	*
2689	* We need to export the entire state as we could be continuing nested-guest
2690	* execution at any point (not just immediately after VMRUN) and thus the VMCB
2691	* can be out-of-sync with the nested-guest state if it was executed in IEM.
2692	*
2693	* @returns VBox status code.
2694	* @param pVCpu The cross context virtual CPU structure.
2695	* @param pCtx Pointer to the guest-CPU context.
2696	*
2697	* @remarks No-long-jump zone!!!
2698	*/
2699	static int hmR0SvmExportGuestStateNested(PVMCPU pVCpu)
2700	{
2701	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatExportGuestState, x);
2702
2703	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2704	PSVMVMCB pVmcbNstGst = pCtx->hwvirt.svm.CTX_SUFF(pVmcb);
2705	Assert(pVmcbNstGst);
2706
2707	hmR0SvmSetupVmcbNested(pVCpu);
2708
2709	pVmcbNstGst->guest.u64RIP = pCtx->rip;
2710	pVmcbNstGst->guest.u64RSP = pCtx->rsp;
2711	pVmcbNstGst->guest.u64RFlags = pCtx->eflags.u32;
2712	pVmcbNstGst->guest.u64RAX = pCtx->rax;
2713
2714	RTCCUINTREG const fEFlags = ASMIntDisableFlags();
2715
2716	int rc = hmR0SvmExportGuestControlRegs(pVCpu, pVmcbNstGst);
2717	AssertRCReturnStmt(rc, ASMSetFlags(fEFlags), rc);
2718
2719	hmR0SvmExportGuestSegmentRegs(pVCpu, pVmcbNstGst);
2720	hmR0SvmExportGuestMsrs(pVCpu, pVmcbNstGst);
2721	hmR0SvmExportGuestHwvirtStateNested(pVCpu, pVmcbNstGst);
2722
2723	ASMSetFlags(fEFlags);
2724
2725	/* Nested VGIF not supported yet. */
2726	Assert(!pVmcbNstGst->ctrl.IntCtrl.n.u1VGifEnable);
2727
2728	rc = hmR0SvmSelectVMRunHandler(pVCpu);
2729	AssertRCReturn(rc, rc);
2730
2731	/* Clear any bits that may be set but exported unconditionally or unused/reserved bits. */
2732	ASMAtomicUoAndU64(&pVCpu->hm.s.fCtxChanged, ~( HM_CHANGED_GUEST_RIP
2733	\| HM_CHANGED_GUEST_RFLAGS
2734	\| HM_CHANGED_GUEST_GPRS_MASK
2735	\| HM_CHANGED_GUEST_APIC_TPR
2736	\| HM_CHANGED_GUEST_X87
2737	\| HM_CHANGED_GUEST_SSE_AVX
2738	\| HM_CHANGED_GUEST_OTHER_XSAVE
2739	\| HM_CHANGED_GUEST_XCRx
2740	\| HM_CHANGED_GUEST_TSC_AUX
2741	\| HM_CHANGED_GUEST_OTHER_MSRS
2742	\| HM_CHANGED_SVM_GUEST_XCPT_INTERCEPTS
2743	\| (HM_CHANGED_KEEPER_STATE_MASK & ~HM_CHANGED_SVM_MASK)));
2744
2745	#ifdef VBOX_STRICT
2746	/*
2747	* All of the guest-CPU state and SVM keeper bits should be exported here by now, except
2748	* for the host-context and/or shared host-guest context bits.
2749	*/
2750	uint64_t const fCtxChanged = ASMAtomicUoReadU64(&pVCpu->hm.s.fCtxChanged);
2751	AssertMsg(!(fCtxChanged & (HM_CHANGED_ALL_GUEST & ~HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE)),
2752	("fCtxChanged=%#RX64\n", fCtxChanged));
2753
2754	/*
2755	* If we need to log state that isn't always imported, we'll need to import them here.
2756	* See hmR0SvmPostRunGuest() for which part of the state is imported uncondtionally.
2757	*/
2758	hmR0SvmLogState(pVCpu, pVmcbNstGst, "hmR0SvmExportGuestStateNested", 0 /* fFlags /, 0 / uVerbose */);
2759	#endif
2760
2761	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExportGuestState, x);
2762	return rc;
2763	}
2764	#endif /* VBOX_WITH_NESTED_HWVIRT_SVM */
2765
2766
2767	/**
2768	* Exports the state shared between the host and guest (or nested-guest) into
2769	* the VMCB.
2770	*
2771	* @param pVCpu The cross context virtual CPU structure.
2772	* @param pVmcb Pointer to the VM control block.
2773	*
2774	* @remarks No-long-jump zone!!!
2775	*/
2776	static void hmR0SvmExportSharedState(PVMCPU pVCpu, PSVMVMCB pVmcb)
2777	{
2778	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2779	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2780
2781	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_DR_MASK)
2782	{
2783	/** @todo Figure out stepping with nested-guest. */
2784	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2785	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
2786	hmR0SvmExportSharedDebugState(pVCpu, pVmcb);
2787	else
2788	{
2789	pVmcb->guest.u64DR6 = pCtx->dr[6];
2790	pVmcb->guest.u64DR7 = pCtx->dr[7];
2791	}
2792	}
2793
2794	pVCpu->hm.s.fCtxChanged &= ~HM_CHANGED_GUEST_DR_MASK;
2795	AssertMsg(!(pVCpu->hm.s.fCtxChanged & HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE),
2796	("fCtxChanged=%#RX64\n", pVCpu->hm.s.fCtxChanged));
2797	}
2798
2799
2800	/**
2801	* Worker for SVMR0ImportStateOnDemand.
2802	*
2803	* @param pVCpu The cross context virtual CPU structure.
2804	* @param pCtx Pointer to the guest-CPU or nested-guest-CPU context.
2805	* @param fWhat What to import, CPUMCTX_EXTRN_XXX.
2806	*/
2807	static void hmR0SvmImportGuestState(PVMCPU pVCpu, uint64_t fWhat)
2808	{
2809	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatImportGuestState, x);
2810
2811	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2812	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
2813	PCSVMVMCBSTATESAVE pVmcbGuest = &pVmcb->guest;
2814	PCSVMVMCBCTRL pVmcbCtrl = &pVmcb->ctrl;
2815
2816	Log4Func(("fExtrn=%#RX64 fWhat=%#RX64\n", pCtx->fExtrn, fWhat));
2817
2818	/*
2819	* We disable interrupts to make the updating of the state and in particular
2820	* the fExtrn modification atomic wrt to preemption hooks.
2821	*/
2822	RTCCUINTREG const fEFlags = ASMIntDisableFlags();
2823
2824	fWhat &= pCtx->fExtrn;
2825	if (fWhat)
2826	{
2827	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
2828	if (fWhat & CPUMCTX_EXTRN_HWVIRT)
2829	{
2830	if ( !CPUMIsGuestInSvmNestedHwVirtMode(pCtx)
2831	&& pVmcbCtrl->IntCtrl.n.u1VGifEnable)
2832	{
2833	/* We don't yet support passing VGIF feature to the guest. */
2834	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.svm.fVGif);
2835	pCtx->hwvirt.fGif = pVmcbCtrl->IntCtrl.n.u1VGif;
2836	}
2837	}
2838
2839	if (fWhat & CPUMCTX_EXTRN_HM_SVM_HWVIRT_VIRQ)
2840	{
2841	if ( !pVmcbCtrl->IntCtrl.n.u1VIrqPending
2842	&& VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_NESTED_GUEST))
2843	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INTERRUPT_NESTED_GUEST);
2844	}
2845	#endif
2846
2847	if (fWhat & CPUMCTX_EXTRN_HM_SVM_INT_SHADOW)
2848	{
2849	if (pVmcbCtrl->IntShadow.n.u1IntShadow)
2850	EMSetInhibitInterruptsPC(pVCpu, pVmcbGuest->u64RIP);
2851	else if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS))
2852	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS);
2853	}
2854
2855	if (fWhat & CPUMCTX_EXTRN_RIP)
2856	pCtx->rip = pVmcbGuest->u64RIP;
2857
2858	if (fWhat & CPUMCTX_EXTRN_RFLAGS)
2859	pCtx->eflags.u32 = pVmcbGuest->u64RFlags;
2860
2861	if (fWhat & CPUMCTX_EXTRN_RSP)
2862	pCtx->rsp = pVmcbGuest->u64RSP;
2863
2864	if (fWhat & CPUMCTX_EXTRN_RAX)
2865	pCtx->rax = pVmcbGuest->u64RAX;
2866
2867	if (fWhat & CPUMCTX_EXTRN_SREG_MASK)
2868	{
2869	if (fWhat & CPUMCTX_EXTRN_CS)
2870	{
2871	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, CS, cs);
2872	/* Correct the CS granularity bit. Haven't seen it being wrong in any other register (yet). */
2873	/** @todo SELM might need to be fixed as it too should not care about the
2874	* granularity bit. See @bugref{6785}. */
2875	if ( !pCtx->cs.Attr.n.u1Granularity
2876	&& pCtx->cs.Attr.n.u1Present
2877	&& pCtx->cs.u32Limit > UINT32_C(0xfffff))
2878	{
2879	Assert((pCtx->cs.u32Limit & 0xfff) == 0xfff);
2880	pCtx->cs.Attr.n.u1Granularity = 1;
2881	}
2882	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, cs);
2883	}
2884	if (fWhat & CPUMCTX_EXTRN_SS)
2885	{
2886	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, SS, ss);
2887	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, ss);
2888	/*
2889	* Sync the hidden SS DPL field. AMD CPUs have a separate CPL field in the
2890	* VMCB and uses that and thus it's possible that when the CPL changes during
2891	* guest execution that the SS DPL isn't updated by AMD-V. Observed on some
2892	* AMD Fusion CPUs with 64-bit guests.
2893	*
2894	* See AMD spec. 15.5.1 "Basic operation".
2895	*/
2896	Assert(!(pVmcbGuest->u8CPL & ~0x3));
2897	uint8_t const uCpl = pVmcbGuest->u8CPL;
2898	if (pCtx->ss.Attr.n.u2Dpl != uCpl)
2899	pCtx->ss.Attr.n.u2Dpl = uCpl & 0x3;
2900	}
2901	if (fWhat & CPUMCTX_EXTRN_DS)
2902	{
2903	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, DS, ds);
2904	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, ds);
2905	}
2906	if (fWhat & CPUMCTX_EXTRN_ES)
2907	{
2908	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, ES, es);
2909	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, es);
2910	}
2911	if (fWhat & CPUMCTX_EXTRN_FS)
2912	{
2913	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, FS, fs);
2914	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, fs);
2915	}
2916	if (fWhat & CPUMCTX_EXTRN_GS)
2917	{
2918	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, GS, gs);
2919	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, gs);
2920	}
2921	}
2922
2923	if (fWhat & CPUMCTX_EXTRN_TABLE_MASK)
2924	{
2925	if (fWhat & CPUMCTX_EXTRN_TR)
2926	{
2927	/*
2928	* Fixup TR attributes so it's compatible with Intel. Important when saved-states
2929	* are used between Intel and AMD, see @bugref{6208#c39}.
2930	* ASSUME that it's normally correct and that we're in 32-bit or 64-bit mode.
2931	*/
2932	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, TR, tr);
2933	if (pCtx->tr.Attr.n.u4Type != X86_SEL_TYPE_SYS_386_TSS_BUSY)
2934	{
2935	if ( pCtx->tr.Attr.n.u4Type == X86_SEL_TYPE_SYS_386_TSS_AVAIL
2936	\|\| CPUMIsGuestInLongModeEx(pCtx))
2937	pCtx->tr.Attr.n.u4Type = X86_SEL_TYPE_SYS_386_TSS_BUSY;
2938	else if (pCtx->tr.Attr.n.u4Type == X86_SEL_TYPE_SYS_286_TSS_AVAIL)
2939	pCtx->tr.Attr.n.u4Type = X86_SEL_TYPE_SYS_286_TSS_BUSY;
2940	}
2941	}
2942
2943	if (fWhat & CPUMCTX_EXTRN_LDTR)
2944	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, LDTR, ldtr);
2945
2946	if (fWhat & CPUMCTX_EXTRN_GDTR)
2947	{
2948	pCtx->gdtr.cbGdt = pVmcbGuest->GDTR.u32Limit;
2949	pCtx->gdtr.pGdt = pVmcbGuest->GDTR.u64Base;
2950	}
2951
2952	if (fWhat & CPUMCTX_EXTRN_IDTR)
2953	{
2954	pCtx->idtr.cbIdt = pVmcbGuest->IDTR.u32Limit;
2955	pCtx->idtr.pIdt = pVmcbGuest->IDTR.u64Base;
2956	}
2957	}
2958
2959	if (fWhat & CPUMCTX_EXTRN_SYSCALL_MSRS)
2960	{
2961	pCtx->msrSTAR = pVmcbGuest->u64STAR;
2962	pCtx->msrLSTAR = pVmcbGuest->u64LSTAR;
2963	pCtx->msrCSTAR = pVmcbGuest->u64CSTAR;
2964	pCtx->msrSFMASK = pVmcbGuest->u64SFMASK;
2965	}
2966
2967	if (fWhat & CPUMCTX_EXTRN_SYSENTER_MSRS)
2968	{
2969	pCtx->SysEnter.cs = pVmcbGuest->u64SysEnterCS;
2970	pCtx->SysEnter.eip = pVmcbGuest->u64SysEnterEIP;
2971	pCtx->SysEnter.esp = pVmcbGuest->u64SysEnterESP;
2972	}
2973
2974	if (fWhat & CPUMCTX_EXTRN_KERNEL_GS_BASE)
2975	pCtx->msrKERNELGSBASE = pVmcbGuest->u64KernelGSBase;
2976
2977	if (fWhat & CPUMCTX_EXTRN_DR_MASK)
2978	{
2979	if (fWhat & CPUMCTX_EXTRN_DR6)
2980	{
2981	if (!pVCpu->hm.s.fUsingHyperDR7)
2982	pCtx->dr[6] = pVmcbGuest->u64DR6;
2983	else
2984	CPUMSetHyperDR6(pVCpu, pVmcbGuest->u64DR6);
2985	}
2986
2987	if (fWhat & CPUMCTX_EXTRN_DR7)
2988	{
2989	if (!pVCpu->hm.s.fUsingHyperDR7)
2990	pCtx->dr[7] = pVmcbGuest->u64DR7;
2991	else
2992	Assert(pVmcbGuest->u64DR7 == CPUMGetHyperDR7(pVCpu));
2993	}
2994	}
2995
2996	if (fWhat & CPUMCTX_EXTRN_CR_MASK)
2997	{
2998	if (fWhat & CPUMCTX_EXTRN_CR0)
2999	{
3000	/* We intercept changes to all CR0 bits except maybe TS & MP bits. */
3001	uint64_t const uCr0 = (pCtx->cr0 & ~(X86_CR0_TS \| X86_CR0_MP))
3002	\| (pVmcbGuest->u64CR0 & (X86_CR0_TS \| X86_CR0_MP));
3003	VMMRZCallRing3Disable(pVCpu); /* Calls into PGM which has Log statements. */
3004	CPUMSetGuestCR0(pVCpu, uCr0);
3005	VMMRZCallRing3Enable(pVCpu);
3006	}
3007
3008	if (fWhat & CPUMCTX_EXTRN_CR2)
3009	pCtx->cr2 = pVmcbGuest->u64CR2;
3010
3011	if (fWhat & CPUMCTX_EXTRN_CR3)
3012	{
3013	if ( pVmcbCtrl->NestedPagingCtrl.n.u1NestedPaging
3014	&& pCtx->cr3 != pVmcbGuest->u64CR3)
3015	{
3016	CPUMSetGuestCR3(pVCpu, pVmcbGuest->u64CR3);
3017	VMCPU_FF_SET(pVCpu, VMCPU_FF_HM_UPDATE_CR3);
3018	}
3019	}
3020
3021	/* Changes to CR4 are always intercepted. */
3022	}
3023
3024	/* Update fExtrn. */
3025	pCtx->fExtrn &= ~fWhat;
3026
3027	/* If everything has been imported, clear the HM keeper bit. */
3028	if (!(pCtx->fExtrn & HMSVM_CPUMCTX_EXTRN_ALL))
3029	{
3030	pCtx->fExtrn &= ~CPUMCTX_EXTRN_KEEPER_HM;
3031	Assert(!pCtx->fExtrn);
3032	}
3033	}
3034	else
3035	Assert(!pCtx->fExtrn \|\| (pCtx->fExtrn & HMSVM_CPUMCTX_EXTRN_ALL));
3036
3037	ASMSetFlags(fEFlags);
3038
3039	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatImportGuestState, x);
3040
3041	/*
3042	* Honor any pending CR3 updates.
3043	*
3044	* Consider this scenario: #VMEXIT -> VMMRZCallRing3Enable() -> do stuff that causes a longjmp
3045	* -> hmR0SvmCallRing3Callback() -> VMMRZCallRing3Disable() -> hmR0SvmImportGuestState()
3046	* -> Sets VMCPU_FF_HM_UPDATE_CR3 pending -> return from the longjmp -> continue with #VMEXIT
3047	* handling -> hmR0SvmImportGuestState() and here we are.
3048	*
3049	* The reason for such complicated handling is because VM-exits that call into PGM expect
3050	* CR3 to be up-to-date and thus any CR3-saves -before- the VM-exit (longjmp) would've
3051	* postponed the CR3 update via the force-flag and cleared CR3 from fExtrn. Any SVM R0
3052	* VM-exit handler that requests CR3 to be saved will end up here and we call PGMUpdateCR3().
3053	*
3054	* The longjmp exit path can't check these CR3 force-flags and call code that takes a lock again,
3055	* and does not process force-flag like regular exits to ring-3 either, we cover for it here.
3056	*/
3057	if ( VMMRZCallRing3IsEnabled(pVCpu)
3058	&& VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_UPDATE_CR3))
3059	{
3060	Assert(pCtx->cr3 == pVmcbGuest->u64CR3);
3061	PGMUpdateCR3(pVCpu, pCtx->cr3);
3062	}
3063	}
3064
3065
3066	/**
3067	* Saves the guest (or nested-guest) state from the VMCB into the guest-CPU
3068	* context.
3069	*
3070	* Currently there is no residual state left in the CPU that is not updated in the
3071	* VMCB.
3072	*
3073	* @returns VBox status code.
3074	* @param pVCpu The cross context virtual CPU structure.
3075	* @param fWhat What to import, CPUMCTX_EXTRN_XXX.
3076	*/
3077	VMMR0DECL(int) SVMR0ImportStateOnDemand(PVMCPU pVCpu, uint64_t fWhat)
3078	{
3079	hmR0SvmImportGuestState(pVCpu, fWhat);
3080	return VINF_SUCCESS;
3081	}
3082
3083
3084	/**
3085	* Does the necessary state syncing before returning to ring-3 for any reason
3086	* (longjmp, preemption, voluntary exits to ring-3) from AMD-V.
3087	*
3088	* @param pVCpu The cross context virtual CPU structure.
3089	* @param fImportState Whether to import the guest state from the VMCB back
3090	* to the guest-CPU context.
3091	*
3092	* @remarks No-long-jmp zone!!!
3093	*/
3094	static void hmR0SvmLeave(PVMCPU pVCpu, bool fImportState)
3095	{
3096	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
3097	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
3098	Assert(VMMR0IsLogFlushDisabled(pVCpu));
3099
3100	/*
3101	* !!! IMPORTANT !!!
3102	* If you modify code here, make sure to check whether hmR0SvmCallRing3Callback() needs to be updated too.
3103	*/
3104
3105	/* Save the guest state if necessary. */
3106	if (fImportState)
3107	hmR0SvmImportGuestState(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
3108
3109	/* Restore host FPU state if necessary and resync on next R0 reentry. */
3110	CPUMR0FpuStateMaybeSaveGuestAndRestoreHost(pVCpu);
3111	Assert(!CPUMIsGuestFPUStateActive(pVCpu));
3112
3113	/*
3114	* Restore host debug registers if necessary and resync on next R0 reentry.
3115	*/
3116	#ifdef VBOX_STRICT
3117	if (CPUMIsHyperDebugStateActive(pVCpu))
3118	{
3119	PSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb; /** @todo nested-guest. */
3120	Assert(pVmcb->ctrl.u16InterceptRdDRx == 0xffff);
3121	Assert(pVmcb->ctrl.u16InterceptWrDRx == 0xffff);
3122	}
3123	#endif
3124	CPUMR0DebugStateMaybeSaveGuestAndRestoreHost(pVCpu, false /* save DR6 */);
3125	Assert(!CPUMIsHyperDebugStateActive(pVCpu));
3126	Assert(!CPUMIsGuestDebugStateActive(pVCpu));
3127
3128	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatEntry);
3129	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatImportGuestState);
3130	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatExportGuestState);
3131	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatExit1);
3132	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatExit2);
3133	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchLongJmpToR3);
3134
3135	VMCPU_CMPXCHG_STATE(pVCpu, VMCPUSTATE_STARTED_HM, VMCPUSTATE_STARTED_EXEC);
3136	}
3137
3138
3139	/**
3140	* Leaves the AMD-V session.
3141	*
3142	* Only used while returning to ring-3 either due to longjump or exits to
3143	* ring-3.
3144	*
3145	* @returns VBox status code.
3146	* @param pVCpu The cross context virtual CPU structure.
3147	*/
3148	static int hmR0SvmLeaveSession(PVMCPU pVCpu)
3149	{
3150	HM_DISABLE_PREEMPT();
3151	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
3152	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
3153
3154	/* When thread-context hooks are used, we can avoid doing the leave again if we had been preempted before
3155	and done this from the SVMR0ThreadCtxCallback(). */
3156	if (!pVCpu->hm.s.fLeaveDone)
3157	{
3158	hmR0SvmLeave(pVCpu, true /* fImportState */);
3159	pVCpu->hm.s.fLeaveDone = true;
3160	}
3161
3162	/*
3163	* !!! IMPORTANT !!!
3164	* If you modify code here, make sure to check whether hmR0SvmCallRing3Callback() needs to be updated too.
3165	*/
3166
3167	/** @todo eliminate the need for calling VMMR0ThreadCtxHookDisable here! */
3168	/* Deregister hook now that we've left HM context before re-enabling preemption. */
3169	VMMR0ThreadCtxHookDisable(pVCpu);
3170
3171	/* Leave HM context. This takes care of local init (term). */
3172	int rc = HMR0LeaveCpu(pVCpu);
3173
3174	HM_RESTORE_PREEMPT();
3175	return rc;
3176	}
3177
3178
3179	/**
3180	* Does the necessary state syncing before doing a longjmp to ring-3.
3181	*
3182	* @returns VBox status code.
3183	* @param pVCpu The cross context virtual CPU structure.
3184	*
3185	* @remarks No-long-jmp zone!!!
3186	*/
3187	static int hmR0SvmLongJmpToRing3(PVMCPU pVCpu)
3188	{
3189	return hmR0SvmLeaveSession(pVCpu);
3190	}
3191
3192
3193	/**
3194	* VMMRZCallRing3() callback wrapper which saves the guest state (or restores
3195	* any remaining host state) before we longjump to ring-3 and possibly get
3196	* preempted.
3197	*
3198	* @param pVCpu The cross context virtual CPU structure.
3199	* @param enmOperation The operation causing the ring-3 longjump.
3200	* @param pvUser The user argument (pointer to the possibly
3201	* out-of-date guest-CPU context).
3202	*/
3203	static DECLCALLBACK(int) hmR0SvmCallRing3Callback(PVMCPU pVCpu, VMMCALLRING3 enmOperation, void *pvUser)
3204	{
3205	RT_NOREF_PV(pvUser);
3206
3207	if (enmOperation == VMMCALLRING3_VM_R0_ASSERTION)
3208	{
3209	/*
3210	* !!! IMPORTANT !!!
3211	* If you modify code here, make sure to check whether hmR0SvmLeave() and hmR0SvmLeaveSession() needs
3212	* to be updated too. This is a stripped down version which gets out ASAP trying to not trigger any assertion.
3213	*/
3214	VMMRZCallRing3RemoveNotification(pVCpu);
3215	VMMRZCallRing3Disable(pVCpu);
3216	HM_DISABLE_PREEMPT();
3217
3218	/* Restore host FPU state if necessary and resync on next R0 reentry. */
3219	CPUMR0FpuStateMaybeSaveGuestAndRestoreHost(pVCpu);
3220
3221	/* Restore host debug registers if necessary and resync on next R0 reentry. */
3222	CPUMR0DebugStateMaybeSaveGuestAndRestoreHost(pVCpu, false /* save DR6 */);
3223
3224	/* Deregister the hook now that we've left HM context before re-enabling preemption. */
3225	/** @todo eliminate the need for calling VMMR0ThreadCtxHookDisable here! */
3226	VMMR0ThreadCtxHookDisable(pVCpu);
3227
3228	/* Leave HM context. This takes care of local init (term). */
3229	HMR0LeaveCpu(pVCpu);
3230
3231	HM_RESTORE_PREEMPT();
3232	return VINF_SUCCESS;
3233	}
3234
3235	Assert(pVCpu);
3236	Assert(pvUser);
3237	Assert(VMMRZCallRing3IsEnabled(pVCpu));
3238	HMSVM_ASSERT_PREEMPT_SAFE();
3239
3240	VMMRZCallRing3Disable(pVCpu);
3241	Assert(VMMR0IsLogFlushDisabled(pVCpu));
3242
3243	Log4Func(("Calling hmR0SvmLongJmpToRing3\n"));
3244	int rc = hmR0SvmLongJmpToRing3(pVCpu);
3245	AssertRCReturn(rc, rc);
3246
3247	VMMRZCallRing3Enable(pVCpu);
3248	return VINF_SUCCESS;
3249	}
3250
3251
3252	/**
3253	* Take necessary actions before going back to ring-3.
3254	*
3255	* An action requires us to go back to ring-3. This function does the necessary
3256	* steps before we can safely return to ring-3. This is not the same as longjmps
3257	* to ring-3, this is voluntary.
3258	*
3259	* @returns VBox status code.
3260	* @param pVM The cross context VM structure.
3261	* @param pVCpu The cross context virtual CPU structure.
3262	* @param pCtx Pointer to the guest-CPU context.
3263	* @param rcExit The reason for exiting to ring-3. Can be
3264	* VINF_VMM_UNKNOWN_RING3_CALL.
3265	*/
3266	static int hmR0SvmExitToRing3(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, int rcExit)
3267	{
3268	Assert(pVM);
3269	Assert(pVCpu);
3270	Assert(pCtx);
3271	HMSVM_ASSERT_PREEMPT_SAFE();
3272
3273	/* Please, no longjumps here (any logging shouldn't flush jump back to ring-3). NO LOGGING BEFORE THIS POINT! */
3274	VMMRZCallRing3Disable(pVCpu);
3275	Log4Func(("rcExit=%d LocalFF=%#RX32 GlobalFF=%#RX32\n", rcExit, pVCpu->fLocalForcedActions, pVM->fGlobalForcedActions));
3276
3277	/* We need to do this only while truly exiting the "inner loop" back to ring-3 and -not- for any longjmp to ring3. */
3278	if (pVCpu->hm.s.Event.fPending)
3279	{
3280	hmR0SvmPendingEventToTrpmTrap(pVCpu);
3281	Assert(!pVCpu->hm.s.Event.fPending);
3282	}
3283
3284	/* Sync. the necessary state for going back to ring-3. */
3285	hmR0SvmLeaveSession(pVCpu);
3286	STAM_COUNTER_DEC(&pVCpu->hm.s.StatSwitchLongJmpToR3);
3287
3288	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_TO_R3);
3289	CPUMSetChangedFlags(pVCpu, CPUM_CHANGED_SYSENTER_MSR
3290	\| CPUM_CHANGED_LDTR
3291	\| CPUM_CHANGED_GDTR
3292	\| CPUM_CHANGED_IDTR
3293	\| CPUM_CHANGED_TR
3294	\| CPUM_CHANGED_HIDDEN_SEL_REGS);
3295	if ( pVM->hm.s.fNestedPaging
3296	&& CPUMIsGuestPagingEnabledEx(pCtx))
3297	{
3298	CPUMSetChangedFlags(pVCpu, CPUM_CHANGED_GLOBAL_TLB_FLUSH);
3299	}
3300
3301	/* Update the exit-to-ring 3 reason. */
3302	pVCpu->hm.s.rcLastExitToR3 = rcExit;
3303
3304	/* On our way back from ring-3 reload the guest state if there is a possibility of it being changed. */
3305	if (rcExit != VINF_EM_RAW_INTERRUPT)
3306	{
3307	Assert(!(pCtx->fExtrn & HMSVM_CPUMCTX_EXTRN_ALL));
3308	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
3309	}
3310
3311	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchExitToR3);
3312
3313	/* We do -not- want any longjmp notifications after this! We must return to ring-3 ASAP. */
3314	VMMRZCallRing3RemoveNotification(pVCpu);
3315	VMMRZCallRing3Enable(pVCpu);
3316
3317	/*
3318	* If we're emulating an instruction, we shouldn't have any TRPM traps pending
3319	* and if we're injecting an event we should have a TRPM trap pending.
3320	*/
3321	AssertReturnStmt(rcExit != VINF_EM_RAW_INJECT_TRPM_EVENT \|\| TRPMHasTrap(pVCpu),
3322	pVCpu->hm.s.u32HMError = rcExit,
3323	VERR_SVM_IPE_5);
3324	AssertReturnStmt(rcExit != VINF_EM_RAW_EMULATE_INSTR \|\| !TRPMHasTrap(pVCpu),
3325	pVCpu->hm.s.u32HMError = rcExit,
3326	VERR_SVM_IPE_4);
3327
3328	return rcExit;
3329	}
3330
3331
3332	/**
3333	* Updates the use of TSC offsetting mode for the CPU and adjusts the necessary
3334	* intercepts.
3335	*
3336	* @param pVM The cross context VM structure.
3337	* @param pVCpu The cross context virtual CPU structure.
3338	* @param pCtx Pointer to the guest-CPU or nested-guest-CPU context.
3339	* @param pVmcb Pointer to the VM control block.
3340	*
3341	* @remarks No-long-jump zone!!!
3342	*/
3343	static void hmR0SvmUpdateTscOffsetting(PVM pVM, PVMCPU pVCpu, PCCPUMCTX pCtx, PSVMVMCB pVmcb)
3344	{
3345	/*
3346	* Avoid intercepting RDTSC/RDTSCP if we determined the host TSC (++) is stable
3347	* and in case of a nested-guest, if the nested-VMCB specifies it is not intercepting
3348	* RDTSC/RDTSCP as well.
3349	*/
3350	bool fParavirtTsc;
3351	uint64_t uTscOffset;
3352	bool const fCanUseRealTsc = TMCpuTickCanUseRealTSC(pVM, pVCpu, &uTscOffset, &fParavirtTsc);
3353
3354	bool fIntercept;
3355	if (fCanUseRealTsc)
3356	fIntercept = hmR0SvmClearCtrlIntercept(pVCpu, pCtx, pVmcb, SVM_CTRL_INTERCEPT_RDTSC \| SVM_CTRL_INTERCEPT_RDTSCP);
3357	else
3358	{
3359	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_RDTSC \| SVM_CTRL_INTERCEPT_RDTSCP);
3360	fIntercept = true;
3361	}
3362
3363	if (!fIntercept)
3364	{
3365	/* Apply the nested-guest VMCB's TSC offset over the guest TSC offset. */
3366	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
3367	uTscOffset = HMSvmNstGstApplyTscOffset(pVCpu, uTscOffset);
3368
3369	/* Update the TSC offset in the VMCB and the relevant clean bits. */
3370	pVmcb->ctrl.u64TSCOffset = uTscOffset;
3371	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
3372
3373	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscOffset);
3374	}
3375	else
3376	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscIntercept);
3377
3378	/* Currently neither Hyper-V nor KVM need to update their paravirt. TSC
3379	information before every VM-entry, hence we have nothing to do here at the moment. */
3380	if (fParavirtTsc)
3381	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscParavirt);
3382	}
3383
3384
3385	/**
3386	* Sets an event as a pending event to be injected into the guest.
3387	*
3388	* @param pVCpu The cross context virtual CPU structure.
3389	* @param pEvent Pointer to the SVM event.
3390	* @param GCPtrFaultAddress The fault-address (CR2) in case it's a
3391	* page-fault.
3392	*
3393	* @remarks Statistics counter assumes this is a guest event being reflected to
3394	* the guest i.e. 'StatInjectPendingReflect' is incremented always.
3395	*/
3396	DECLINLINE(void) hmR0SvmSetPendingEvent(PVMCPU pVCpu, PSVMEVENT pEvent, RTGCUINTPTR GCPtrFaultAddress)
3397	{
3398	Assert(!pVCpu->hm.s.Event.fPending);
3399	Assert(pEvent->n.u1Valid);
3400
3401	pVCpu->hm.s.Event.u64IntInfo = pEvent->u;
3402	pVCpu->hm.s.Event.fPending = true;
3403	pVCpu->hm.s.Event.GCPtrFaultAddress = GCPtrFaultAddress;
3404
3405	Log4Func(("u=%#RX64 u8Vector=%#x Type=%#x ErrorCodeValid=%RTbool ErrorCode=%#RX32\n", pEvent->u, pEvent->n.u8Vector,
3406	(uint8_t)pEvent->n.u3Type, !!pEvent->n.u1ErrorCodeValid, pEvent->n.u32ErrorCode));
3407	}
3408
3409
3410	/**
3411	* Sets an invalid-opcode (\#UD) exception as pending-for-injection into the VM.
3412	*
3413	* @param pVCpu The cross context virtual CPU structure.
3414	*/
3415	DECLINLINE(void) hmR0SvmSetPendingXcptUD(PVMCPU pVCpu)
3416	{
3417	SVMEVENT Event;
3418	Event.u = 0;
3419	Event.n.u1Valid = 1;
3420	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3421	Event.n.u8Vector = X86_XCPT_UD;
3422	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3423	}
3424
3425
3426	/**
3427	* Sets a debug (\#DB) exception as pending-for-injection into the VM.
3428	*
3429	* @param pVCpu The cross context virtual CPU structure.
3430	*/
3431	DECLINLINE(void) hmR0SvmSetPendingXcptDB(PVMCPU pVCpu)
3432	{
3433	SVMEVENT Event;
3434	Event.u = 0;
3435	Event.n.u1Valid = 1;
3436	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3437	Event.n.u8Vector = X86_XCPT_DB;
3438	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3439	}
3440
3441
3442	/**
3443	* Sets a page fault (\#PF) exception as pending-for-injection into the VM.
3444	*
3445	* @param pVCpu The cross context virtual CPU structure.
3446	* @param pCtx Pointer to the guest-CPU context.
3447	* @param u32ErrCode The error-code for the page-fault.
3448	* @param uFaultAddress The page fault address (CR2).
3449	*
3450	* @remarks This updates the guest CR2 with @a uFaultAddress!
3451	*/
3452	DECLINLINE(void) hmR0SvmSetPendingXcptPF(PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t u32ErrCode, RTGCUINTPTR uFaultAddress)
3453	{
3454	SVMEVENT Event;
3455	Event.u = 0;
3456	Event.n.u1Valid = 1;
3457	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3458	Event.n.u8Vector = X86_XCPT_PF;
3459	Event.n.u1ErrorCodeValid = 1;
3460	Event.n.u32ErrorCode = u32ErrCode;
3461
3462	/* Update CR2 of the guest. */
3463	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_CR2);
3464	if (pCtx->cr2 != uFaultAddress)
3465	{
3466	pCtx->cr2 = uFaultAddress;
3467	/* The VMCB clean bit for CR2 will be updated while re-loading the guest state. */
3468	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_CR2);
3469	}
3470
3471	hmR0SvmSetPendingEvent(pVCpu, &Event, uFaultAddress);
3472	}
3473
3474
3475	/**
3476	* Sets a math-fault (\#MF) exception as pending-for-injection into the VM.
3477	*
3478	* @param pVCpu The cross context virtual CPU structure.
3479	*/
3480	DECLINLINE(void) hmR0SvmSetPendingXcptMF(PVMCPU pVCpu)
3481	{
3482	SVMEVENT Event;
3483	Event.u = 0;
3484	Event.n.u1Valid = 1;
3485	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3486	Event.n.u8Vector = X86_XCPT_MF;
3487	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3488	}
3489
3490
3491	/**
3492	* Sets a double fault (\#DF) exception as pending-for-injection into the VM.
3493	*
3494	* @param pVCpu The cross context virtual CPU structure.
3495	*/
3496	DECLINLINE(void) hmR0SvmSetPendingXcptDF(PVMCPU pVCpu)
3497	{
3498	SVMEVENT Event;
3499	Event.u = 0;
3500	Event.n.u1Valid = 1;
3501	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3502	Event.n.u8Vector = X86_XCPT_DF;
3503	Event.n.u1ErrorCodeValid = 1;
3504	Event.n.u32ErrorCode = 0;
3505	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3506	}
3507
3508
3509	/**
3510	* Injects an event into the guest upon VMRUN by updating the relevant field
3511	* in the VMCB.
3512	*
3513	* @param pVCpu The cross context virtual CPU structure.
3514	* @param pVmcb Pointer to the guest VM control block.
3515	* @param pEvent Pointer to the event.
3516	*
3517	* @remarks No-long-jump zone!!!
3518	* @remarks Requires CR0!
3519	*/
3520	DECLINLINE(void) hmR0SvmInjectEventVmcb(PVMCPU pVCpu, PSVMVMCB pVmcb, PSVMEVENT pEvent)
3521	{
3522	Assert(!pVmcb->ctrl.EventInject.n.u1Valid);
3523	pVmcb->ctrl.EventInject.u = pEvent->u;
3524	STAM_COUNTER_INC(&pVCpu->hm.s.paStatInjectedIrqsR0[pEvent->n.u8Vector & MASK_INJECT_IRQ_STAT]);
3525	RT_NOREF(pVCpu);
3526
3527	Log4Func(("u=%#RX64 u8Vector=%#x Type=%#x ErrorCodeValid=%RTbool ErrorCode=%#RX32\n", pEvent->u, pEvent->n.u8Vector,
3528	(uint8_t)pEvent->n.u3Type, !!pEvent->n.u1ErrorCodeValid, pEvent->n.u32ErrorCode));
3529	}
3530
3531
3532
3533	/**
3534	* Converts any TRPM trap into a pending HM event. This is typically used when
3535	* entering from ring-3 (not longjmp returns).
3536	*
3537	* @param pVCpu The cross context virtual CPU structure.
3538	*/
3539	static void hmR0SvmTrpmTrapToPendingEvent(PVMCPU pVCpu)
3540	{
3541	Assert(TRPMHasTrap(pVCpu));
3542	Assert(!pVCpu->hm.s.Event.fPending);
3543
3544	uint8_t uVector;
3545	TRPMEVENT enmTrpmEvent;
3546	RTGCUINT uErrCode;
3547	RTGCUINTPTR GCPtrFaultAddress;
3548	uint8_t cbInstr;
3549
3550	int rc = TRPMQueryTrapAll(pVCpu, &uVector, &enmTrpmEvent, &uErrCode, &GCPtrFaultAddress, &cbInstr);
3551	AssertRC(rc);
3552
3553	SVMEVENT Event;
3554	Event.u = 0;
3555	Event.n.u1Valid = 1;
3556	Event.n.u8Vector = uVector;
3557
3558	/* Refer AMD spec. 15.20 "Event Injection" for the format. */
3559	if (enmTrpmEvent == TRPM_TRAP)
3560	{
3561	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3562	switch (uVector)
3563	{
3564	case X86_XCPT_NMI:
3565	{
3566	Event.n.u3Type = SVM_EVENT_NMI;
3567	break;
3568	}
3569
3570	case X86_XCPT_PF:
3571	case X86_XCPT_DF:
3572	case X86_XCPT_TS:
3573	case X86_XCPT_NP:
3574	case X86_XCPT_SS:
3575	case X86_XCPT_GP:
3576	case X86_XCPT_AC:
3577	{
3578	Event.n.u1ErrorCodeValid = 1;
3579	Event.n.u32ErrorCode = uErrCode;
3580	break;
3581	}
3582	}
3583	}
3584	else if (enmTrpmEvent == TRPM_HARDWARE_INT)
3585	Event.n.u3Type = SVM_EVENT_EXTERNAL_IRQ;
3586	else if (enmTrpmEvent == TRPM_SOFTWARE_INT)
3587	Event.n.u3Type = SVM_EVENT_SOFTWARE_INT;
3588	else
3589	AssertMsgFailed(("Invalid TRPM event type %d\n", enmTrpmEvent));
3590
3591	rc = TRPMResetTrap(pVCpu);
3592	AssertRC(rc);
3593
3594	Log4(("TRPM->HM event: u=%#RX64 u8Vector=%#x uErrorCodeValid=%RTbool uErrorCode=%#RX32\n", Event.u, Event.n.u8Vector,
3595	!!Event.n.u1ErrorCodeValid, Event.n.u32ErrorCode));
3596
3597	hmR0SvmSetPendingEvent(pVCpu, &Event, GCPtrFaultAddress);
3598	}
3599
3600
3601	/**
3602	* Converts any pending SVM event into a TRPM trap. Typically used when leaving
3603	* AMD-V to execute any instruction.
3604	*
3605	* @param pVCpu The cross context virtual CPU structure.
3606	*/
3607	static void hmR0SvmPendingEventToTrpmTrap(PVMCPU pVCpu)
3608	{
3609	Assert(pVCpu->hm.s.Event.fPending);
3610	Assert(TRPMQueryTrap(pVCpu, NULL /* pu8TrapNo /, NULL / pEnmType */) == VERR_TRPM_NO_ACTIVE_TRAP);
3611
3612	SVMEVENT Event;
3613	Event.u = pVCpu->hm.s.Event.u64IntInfo;
3614
3615	uint8_t uVector = Event.n.u8Vector;
3616	uint8_t uVectorType = Event.n.u3Type;
3617	TRPMEVENT enmTrapType = HMSvmEventToTrpmEventType(&Event);
3618
3619	Log4(("HM event->TRPM: uVector=%#x enmTrapType=%d\n", uVector, uVectorType));
3620
3621	int rc = TRPMAssertTrap(pVCpu, uVector, enmTrapType);
3622	AssertRC(rc);
3623
3624	if (Event.n.u1ErrorCodeValid)
3625	TRPMSetErrorCode(pVCpu, Event.n.u32ErrorCode);
3626
3627	if ( uVectorType == SVM_EVENT_EXCEPTION
3628	&& uVector == X86_XCPT_PF)
3629	{
3630	TRPMSetFaultAddress(pVCpu, pVCpu->hm.s.Event.GCPtrFaultAddress);
3631	Assert(pVCpu->hm.s.Event.GCPtrFaultAddress == CPUMGetGuestCR2(pVCpu));
3632	}
3633	else if (uVectorType == SVM_EVENT_SOFTWARE_INT)
3634	{
3635	AssertMsg( uVectorType == SVM_EVENT_SOFTWARE_INT
3636	\|\| (uVector == X86_XCPT_BP \|\| uVector == X86_XCPT_OF),
3637	("Invalid vector: uVector=%#x uVectorType=%#x\n", uVector, uVectorType));
3638	TRPMSetInstrLength(pVCpu, pVCpu->hm.s.Event.cbInstr);
3639	}
3640	pVCpu->hm.s.Event.fPending = false;
3641	}
3642
3643
3644	/**
3645	* Checks if the guest (or nested-guest) has an interrupt shadow active right
3646	* now.
3647	*
3648	* @returns @c true if the interrupt shadow is active, @c false otherwise.
3649	* @param pVCpu The cross context virtual CPU structure.
3650	* @param pCtx Pointer to the guest-CPU context.
3651	*
3652	* @remarks No-long-jump zone!!!
3653	* @remarks Has side-effects with VMCPU_FF_INHIBIT_INTERRUPTS force-flag.
3654	*/
3655	DECLINLINE(bool) hmR0SvmIsIntrShadowActive(PVMCPU pVCpu, PCCPUMCTX pCtx)
3656	{
3657	/*
3658	* Instructions like STI and MOV SS inhibit interrupts till the next instruction
3659	* completes. Check if we should inhibit interrupts or clear any existing
3660	* interrupt inhibition.
3661	*/
3662	if (VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS))
3663	{
3664	if (pCtx->rip != EMGetInhibitInterruptsPC(pVCpu))
3665	{
3666	/*
3667	* We can clear the inhibit force flag as even if we go back to the recompiler
3668	* without executing guest code in AMD-V, the flag's condition to be cleared is
3669	* met and thus the cleared state is correct.
3670	*/
3671	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS);
3672	return false;
3673	}
3674	return true;
3675	}
3676	return false;
3677	}
3678
3679
3680	/**
3681	* Sets the virtual interrupt intercept control in the VMCB.
3682	*
3683	* @param pVCpu The cross context virtual CPU structure.
3684	* @param pVmcb Pointer to the VM control block.
3685	* @param pCtx Pointer to the guest-CPU context.
3686	*/
3687	DECLINLINE(void) hmR0SvmSetIntWindowExiting(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
3688	{
3689	/*
3690	* When AVIC isn't supported, set up an interrupt window to cause a #VMEXIT when the guest
3691	* is ready to accept interrupts. At #VMEXIT, we then get the interrupt from the APIC
3692	* (updating ISR at the right time) and inject the interrupt.
3693	*
3694	* With AVIC is supported, we could make use of the asynchronously delivery without
3695	* #VMEXIT and we would be passing the AVIC page to SVM.
3696	*
3697	* In AMD-V, an interrupt window is achieved using a combination of V_IRQ (an interrupt
3698	* is pending), V_IGN_TPR (ignore TPR priorities) and the VINTR intercept all being set.
3699	*/
3700	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
3701	/*
3702	* Currently we don't overlay interupt windows and if there's any V_IRQ pending in the
3703	* nested-guest VMCB, we avoid setting up any interrupt window on behalf of the outer
3704	* guest.
3705	*/
3706	/** @todo Does this mean we end up prioritizing virtual interrupt
3707	* delivery/window over a physical interrupt (from the outer guest)
3708	* might be pending? */
3709	bool const fEnableIntWindow = !VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_NESTED_GUEST);
3710	if (!fEnableIntWindow)
3711	{
3712	Assert(CPUMIsGuestInSvmNestedHwVirtMode(pCtx)); RT_NOREF(pCtx);
3713	Log4(("Nested-guest V_IRQ already pending\n"));
3714	}
3715	#else
3716	RT_NOREF2(pVCpu, pCtx);
3717	bool const fEnableIntWindow = true;
3718	#endif
3719	if (fEnableIntWindow)
3720	{
3721	Assert(pVmcb->ctrl.IntCtrl.n.u1IgnoreTPR);
3722	pVmcb->ctrl.IntCtrl.n.u1VIrqPending = 1;
3723	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INT_CTRL;
3724	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_VINTR);
3725	Log4(("Set VINTR intercept\n"));
3726	}
3727	}
3728
3729
3730	/**
3731	* Clears the virtual interrupt intercept control in the VMCB as
3732	* we are figured the guest is unable process any interrupts
3733	* at this point of time.
3734	*
3735	* @param pVCpu The cross context virtual CPU structure.
3736	* @param pVmcb Pointer to the VM control block.
3737	* @param pCtx Pointer to the guest-CPU context.
3738	*/
3739	DECLINLINE(void) hmR0SvmClearIntWindowExiting(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
3740	{
3741	PSVMVMCBCTRL pVmcbCtrl = &pVmcb->ctrl;
3742	if ( pVmcbCtrl->IntCtrl.n.u1VIrqPending
3743	\|\| (pVmcbCtrl->u64InterceptCtrl & SVM_CTRL_INTERCEPT_VINTR))
3744	{
3745	pVmcbCtrl->IntCtrl.n.u1VIrqPending = 0;
3746	pVmcbCtrl->u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INT_CTRL;
3747	hmR0SvmClearCtrlIntercept(pVCpu, pCtx, pVmcb, SVM_CTRL_INTERCEPT_VINTR);
3748	Log4(("Cleared VINTR intercept\n"));
3749	}
3750	}
3751
3752	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
3753	/**
3754	* Evaluates the event to be delivered to the nested-guest and sets it as the
3755	* pending event.
3756	*
3757	* @returns VBox strict status code.
3758	* @param pVCpu The cross context virtual CPU structure.
3759	* @param pCtx Pointer to the guest-CPU context.
3760	*/
3761	static VBOXSTRICTRC hmR0SvmEvaluatePendingEventNested(PVMCPU pVCpu, PCPUMCTX pCtx)
3762	{
3763	HMSVM_ASSERT_IN_NESTED_GUEST(pCtx);
3764	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_HWVIRT
3765	\| CPUMCTX_EXTRN_RFLAGS
3766	\| CPUMCTX_EXTRN_HM_SVM_INT_SHADOW
3767	\| CPUMCTX_EXTRN_HM_SVM_HWVIRT_VIRQ);
3768
3769	Assert(!pVCpu->hm.s.Event.fPending);
3770	Assert(pCtx->hwvirt.fGif);
3771	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
3772	Assert(pVmcb);
3773
3774	bool const fVirtualGif = CPUMGetSvmNstGstVGif(pCtx);
3775	bool const fIntShadow = hmR0SvmIsIntrShadowActive(pVCpu, pCtx);
3776	bool const fBlockNmi = VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_BLOCK_NMIS);
3777
3778	Log4Func(("fVirtualGif=%RTbool fBlockNmi=%RTbool fIntShadow=%RTbool fIntPending=%RTbool fNmiPending=%RTbool\n",
3779	fVirtualGif, fBlockNmi, fIntShadow, VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC),
3780	VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_NMI)));
3781
3782	/** @todo SMI. SMIs take priority over NMIs. */
3783
3784	/*
3785	* Check if the guest can receive NMIs.
3786	* Nested NMIs are not allowed, see AMD spec. 8.1.4 "Masking External Interrupts".
3787	* NMIs take priority over maskable interrupts, see AMD spec. 8.5 "Priorities".
3788	*/
3789	if ( VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_NMI)
3790	&& !fBlockNmi)
3791	{
3792	if ( fVirtualGif
3793	&& !fIntShadow)
3794	{
3795	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_NMI))
3796	{
3797	Log4(("Intercepting NMI -> #VMEXIT\n"));
3798	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_SVM_VMEXIT_MASK);
3799	return IEMExecSvmVmexit(pVCpu, SVM_EXIT_NMI, 0, 0);
3800	}
3801
3802	Log4(("Setting NMI pending for injection\n"));
3803	SVMEVENT Event;
3804	Event.u = 0;
3805	Event.n.u1Valid = 1;
3806	Event.n.u8Vector = X86_XCPT_NMI;
3807	Event.n.u3Type = SVM_EVENT_NMI;
3808	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3809	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INTERRUPT_NMI);
3810	}
3811	else if (!fVirtualGif)
3812	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_STGI);
3813	else
3814	hmR0SvmSetIntWindowExiting(pVCpu, pVmcb, pCtx);
3815	}
3816	/*
3817	* Check if the nested-guest can receive external interrupts (generated by the guest's
3818	* PIC/APIC).
3819	*
3820	* External intercepts, NMI, SMI etc. from the physical CPU are -always- intercepted
3821	* when executing using hardware-assisted SVM, see HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS.
3822	*
3823	* External interrupts that are generated for the outer guest may be intercepted
3824	* depending on how the nested-guest VMCB was programmed by guest software.
3825	*
3826	* Physical interrupts always take priority over virtual interrupts,
3827	* see AMD spec. 15.21.4 "Injecting Virtual (INTR) Interrupts".
3828	*/
3829	else if ( VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC)
3830	&& !pVCpu->hm.s.fSingleInstruction)
3831	{
3832	if ( fVirtualGif
3833	&& !fIntShadow
3834	&& CPUMCanSvmNstGstTakePhysIntr(pVCpu, pCtx))
3835	{
3836	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_INTR))
3837	{
3838	Log4(("Intercepting INTR -> #VMEXIT\n"));
3839	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_SVM_VMEXIT_MASK);
3840	return IEMExecSvmVmexit(pVCpu, SVM_EXIT_INTR, 0, 0);
3841	}
3842
3843	uint8_t u8Interrupt;
3844	int rc = PDMGetInterrupt(pVCpu, &u8Interrupt);
3845	if (RT_SUCCESS(rc))
3846	{
3847	Log4(("Setting external interrupt %#x pending for injection\n", u8Interrupt));
3848	SVMEVENT Event;
3849	Event.u = 0;
3850	Event.n.u1Valid = 1;
3851	Event.n.u8Vector = u8Interrupt;
3852	Event.n.u3Type = SVM_EVENT_EXTERNAL_IRQ;
3853	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3854	}
3855	else if (rc == VERR_APIC_INTR_MASKED_BY_TPR)
3856	{
3857	/*
3858	* AMD-V has no TPR thresholding feature. TPR and the force-flag will be
3859	* updated eventually when the TPR is written by the guest.
3860	*/
3861	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchTprMaskedIrq);
3862	}
3863	else
3864	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchGuestIrq);
3865	}
3866	else if (!fVirtualGif)
3867	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_STGI);
3868	else
3869	hmR0SvmSetIntWindowExiting(pVCpu, pVmcb, pCtx);
3870	}
3871
3872	return VINF_SUCCESS;
3873	}
3874	#endif
3875
3876	/**
3877	* Evaluates the event to be delivered to the guest and sets it as the pending
3878	* event.
3879	*
3880	* @param pVCpu The cross context virtual CPU structure.
3881	* @param pCtx Pointer to the guest-CPU context.
3882	*/
3883	static void hmR0SvmEvaluatePendingEvent(PVMCPU pVCpu, PCPUMCTX pCtx)
3884	{
3885	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(pCtx);
3886	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_HWVIRT
3887	\| CPUMCTX_EXTRN_RFLAGS
3888	\| CPUMCTX_EXTRN_HM_SVM_INT_SHADOW);
3889
3890	Assert(!pVCpu->hm.s.Event.fPending);
3891	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
3892	Assert(pVmcb);
3893
3894	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
3895	bool const fGif = pCtx->hwvirt.fGif;
3896	#else
3897	bool const fGif = true;
3898	#endif
3899	bool const fIntShadow = hmR0SvmIsIntrShadowActive(pVCpu, pCtx);
3900	bool const fBlockInt = !(pCtx->eflags.u32 & X86_EFL_IF);
3901	bool const fBlockNmi = VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_BLOCK_NMIS);
3902
3903	Log4Func(("fGif=%RTbool fBlockNmi=%RTbool fBlockInt=%RTbool fIntShadow=%RTbool fIntPending=%RTbool NMI pending=%RTbool\n",
3904	fGif, fBlockNmi, fBlockInt, fIntShadow,
3905	VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC),
3906	VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_NMI)));
3907
3908	/** @todo SMI. SMIs take priority over NMIs. */
3909
3910	/*
3911	* Check if the guest can receive NMIs.
3912	* Nested NMIs are not allowed, see AMD spec. 8.1.4 "Masking External Interrupts".
3913	* NMIs take priority over maskable interrupts, see AMD spec. 8.5 "Priorities".
3914	*/
3915	if ( VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_NMI)
3916	&& !fBlockNmi)
3917	{
3918	if ( fGif
3919	&& !fIntShadow)
3920	{
3921	Log4(("Setting NMI pending for injection\n"));
3922	SVMEVENT Event;
3923	Event.u = 0;
3924	Event.n.u1Valid = 1;
3925	Event.n.u8Vector = X86_XCPT_NMI;
3926	Event.n.u3Type = SVM_EVENT_NMI;
3927	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3928	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INTERRUPT_NMI);
3929	}
3930	else if (!fGif)
3931	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_STGI);
3932	else
3933	hmR0SvmSetIntWindowExiting(pVCpu, pVmcb, pCtx);
3934	}
3935	/*
3936	* Check if the guest can receive external interrupts (PIC/APIC). Once PDMGetInterrupt()
3937	* returns a valid interrupt we -must- deliver the interrupt. We can no longer re-request
3938	* it from the APIC device.
3939	*/
3940	else if ( VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC)
3941	&& !pVCpu->hm.s.fSingleInstruction)
3942	{
3943	if ( fGif
3944	&& !fBlockInt
3945	&& !fIntShadow)
3946	{
3947	uint8_t u8Interrupt;
3948	int rc = PDMGetInterrupt(pVCpu, &u8Interrupt);
3949	if (RT_SUCCESS(rc))
3950	{
3951	Log4(("Setting external interrupt %#x pending for injection\n", u8Interrupt));
3952	SVMEVENT Event;
3953	Event.u = 0;
3954	Event.n.u1Valid = 1;
3955	Event.n.u8Vector = u8Interrupt;
3956	Event.n.u3Type = SVM_EVENT_EXTERNAL_IRQ;
3957	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3958	}
3959	else if (rc == VERR_APIC_INTR_MASKED_BY_TPR)
3960	{
3961	/*
3962	* AMD-V has no TPR thresholding feature. TPR and the force-flag will be
3963	* updated eventually when the TPR is written by the guest.
3964	*/
3965	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchTprMaskedIrq);
3966	}
3967	else
3968	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchGuestIrq);
3969	}
3970	else if (!fGif)
3971	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_STGI);
3972	else
3973	hmR0SvmSetIntWindowExiting(pVCpu, pVmcb, pCtx);
3974	}
3975	}
3976
3977
3978	/**
3979	* Injects any pending events into the guest (or nested-guest).
3980	*
3981	* @param pVCpu The cross context virtual CPU structure.
3982	* @param pCtx Pointer to the guest-CPU context.
3983	* @param pVmcb Pointer to the VM control block.
3984	*
3985	* @remarks Must only be called when we are guaranteed to enter
3986	* hardware-assisted SVM execution and not return to ring-3
3987	* prematurely.
3988	*/
3989	static void hmR0SvmInjectPendingEvent(PVMCPU pVCpu, PCCPUMCTX pCtx, PSVMVMCB pVmcb)
3990	{
3991	Assert(!TRPMHasTrap(pVCpu));
3992	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
3993
3994	bool const fIntShadow = hmR0SvmIsIntrShadowActive(pVCpu, pCtx);
3995	#ifdef VBOX_STRICT
3996	bool const fGif = pCtx->hwvirt.fGif;
3997	bool fAllowInt = fGif;
3998	if (fGif)
3999	{
4000	/*
4001	* For nested-guests we have no way to determine if we're injecting a physical or
4002	* virtual interrupt at this point. Hence the partial verification below.
4003	*/
4004	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
4005	fAllowInt = CPUMCanSvmNstGstTakePhysIntr(pVCpu, pCtx) \|\| CPUMCanSvmNstGstTakeVirtIntr(pVCpu, pCtx);
4006	else
4007	fAllowInt = RT_BOOL(pCtx->eflags.u32 & X86_EFL_IF);
4008	}
4009	#endif
4010
4011	if (pVCpu->hm.s.Event.fPending)
4012	{
4013	SVMEVENT Event;
4014	Event.u = pVCpu->hm.s.Event.u64IntInfo;
4015	Assert(Event.n.u1Valid);
4016
4017	/*
4018	* Validate event injection pre-conditions.
4019	*/
4020	if (Event.n.u3Type == SVM_EVENT_EXTERNAL_IRQ)
4021	{
4022	Assert(fAllowInt);
4023	Assert(!fIntShadow);
4024	}
4025	else if (Event.n.u3Type == SVM_EVENT_NMI)
4026	{
4027	Assert(fGif);
4028	Assert(!fIntShadow);
4029	}
4030
4031	/*
4032	* Before injecting an NMI we must set VMCPU_FF_BLOCK_NMIS to prevent nested NMIs. We
4033	* do this only when we are surely going to inject the NMI as otherwise if we return
4034	* to ring-3 prematurely we could leave NMIs blocked indefinitely upon re-entry into
4035	* SVM R0.
4036	*
4037	* With VT-x, this is handled by the Guest interruptibility information VMCS field
4038	* which will set the VMCS field after actually delivering the NMI which we read on
4039	* VM-exit to determine the state.
4040	*/
4041	if ( Event.n.u3Type == SVM_EVENT_NMI
4042	&& Event.n.u8Vector == X86_XCPT_NMI
4043	&& !VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_BLOCK_NMIS))
4044	{
4045	VMCPU_FF_SET(pVCpu, VMCPU_FF_BLOCK_NMIS);
4046	}
4047
4048	/*
4049	* Inject it (update VMCB for injection by the hardware).
4050	*/
4051	Log4(("Injecting pending HM event\n"));
4052	hmR0SvmInjectEventVmcb(pVCpu, pVmcb, &Event);
4053	pVCpu->hm.s.Event.fPending = false;
4054
4055	if (Event.n.u3Type == SVM_EVENT_EXTERNAL_IRQ)
4056	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectInterrupt);
4057	else
4058	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectXcpt);
4059	}
4060	else
4061	Assert(pVmcb->ctrl.EventInject.n.u1Valid == 0);
4062
4063	/*
4064	* We could have injected an NMI through IEM and continue guest execution using
4065	* hardware-assisted SVM. In which case, we would not have any events pending (above)
4066	* but we still need to intercept IRET in order to eventually clear NMI inhibition.
4067	*/
4068	if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_BLOCK_NMIS))
4069	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_IRET);
4070
4071	/*
4072	* Update the guest interrupt shadow in the guest (or nested-guest) VMCB.
4073	*
4074	* For nested-guests: We need to update it too for the scenario where IEM executes
4075	* the nested-guest but execution later continues here with an interrupt shadow active.
4076	*/
4077	pVmcb->ctrl.IntShadow.n.u1IntShadow = fIntShadow;
4078	}
4079
4080
4081	/**
4082	* Reports world-switch error and dumps some useful debug info.
4083	*
4084	* @param pVM The cross context VM structure.
4085	* @param pVCpu The cross context virtual CPU structure.
4086	* @param rcVMRun The return code from VMRUN (or
4087	* VERR_SVM_INVALID_GUEST_STATE for invalid
4088	* guest-state).
4089	* @param pCtx Pointer to the guest-CPU context.
4090	*/
4091	static void hmR0SvmReportWorldSwitchError(PVM pVM, PVMCPU pVCpu, int rcVMRun, PCPUMCTX pCtx)
4092	{
4093	NOREF(pCtx);
4094	HMSVM_ASSERT_PREEMPT_SAFE();
4095	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(pCtx);
4096	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
4097
4098	PCSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
4099	if (rcVMRun == VERR_SVM_INVALID_GUEST_STATE)
4100	{
4101	hmR0DumpRegs(pVM, pVCpu, pCtx); NOREF(pVM);
4102	/** @todo We probably don't need to dump this anymore or we can expand
4103	* hmR0DumpRegs()? */
4104	#ifdef VBOX_STRICT
4105	Log4(("ctrl.u32VmcbCleanBits %#RX32\n", pVmcb->ctrl.u32VmcbCleanBits));
4106	Log4(("ctrl.u16InterceptRdCRx %#x\n", pVmcb->ctrl.u16InterceptRdCRx));
4107	Log4(("ctrl.u16InterceptWrCRx %#x\n", pVmcb->ctrl.u16InterceptWrCRx));
4108	Log4(("ctrl.u16InterceptRdDRx %#x\n", pVmcb->ctrl.u16InterceptRdDRx));
4109	Log4(("ctrl.u16InterceptWrDRx %#x\n", pVmcb->ctrl.u16InterceptWrDRx));
4110	Log4(("ctrl.u32InterceptXcpt %#x\n", pVmcb->ctrl.u32InterceptXcpt));
4111	Log4(("ctrl.u64InterceptCtrl %#RX64\n", pVmcb->ctrl.u64InterceptCtrl));
4112	Log4(("ctrl.u64IOPMPhysAddr %#RX64\n", pVmcb->ctrl.u64IOPMPhysAddr));
4113	Log4(("ctrl.u64MSRPMPhysAddr %#RX64\n", pVmcb->ctrl.u64MSRPMPhysAddr));
4114	Log4(("ctrl.u64TSCOffset %#RX64\n", pVmcb->ctrl.u64TSCOffset));
4115
4116	Log4(("ctrl.TLBCtrl.u32ASID %#x\n", pVmcb->ctrl.TLBCtrl.n.u32ASID));
4117	Log4(("ctrl.TLBCtrl.u8TLBFlush %#x\n", pVmcb->ctrl.TLBCtrl.n.u8TLBFlush));
4118	Log4(("ctrl.TLBCtrl.u24Reserved %#x\n", pVmcb->ctrl.TLBCtrl.n.u24Reserved));
4119
4120	Log4(("ctrl.IntCtrl.u8VTPR %#x\n", pVmcb->ctrl.IntCtrl.n.u8VTPR));
4121	Log4(("ctrl.IntCtrl.u1VIrqPending %#x\n", pVmcb->ctrl.IntCtrl.n.u1VIrqPending));
4122	Log4(("ctrl.IntCtrl.u1VGif %#x\n", pVmcb->ctrl.IntCtrl.n.u1VGif));
4123	Log4(("ctrl.IntCtrl.u6Reserved0 %#x\n", pVmcb->ctrl.IntCtrl.n.u6Reserved));
4124	Log4(("ctrl.IntCtrl.u4VIntrPrio %#x\n", pVmcb->ctrl.IntCtrl.n.u4VIntrPrio));
4125	Log4(("ctrl.IntCtrl.u1IgnoreTPR %#x\n", pVmcb->ctrl.IntCtrl.n.u1IgnoreTPR));
4126	Log4(("ctrl.IntCtrl.u3Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u3Reserved));
4127	Log4(("ctrl.IntCtrl.u1VIntrMasking %#x\n", pVmcb->ctrl.IntCtrl.n.u1VIntrMasking));
4128	Log4(("ctrl.IntCtrl.u1VGifEnable %#x\n", pVmcb->ctrl.IntCtrl.n.u1VGifEnable));
4129	Log4(("ctrl.IntCtrl.u5Reserved1 %#x\n", pVmcb->ctrl.IntCtrl.n.u5Reserved));
4130	Log4(("ctrl.IntCtrl.u8VIntrVector %#x\n", pVmcb->ctrl.IntCtrl.n.u8VIntrVector));
4131	Log4(("ctrl.IntCtrl.u24Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u24Reserved));
4132
4133	Log4(("ctrl.IntShadow.u1IntShadow %#x\n", pVmcb->ctrl.IntShadow.n.u1IntShadow));
4134	Log4(("ctrl.IntShadow.u1GuestIntMask %#x\n", pVmcb->ctrl.IntShadow.n.u1GuestIntMask));
4135	Log4(("ctrl.u64ExitCode %#RX64\n", pVmcb->ctrl.u64ExitCode));
4136	Log4(("ctrl.u64ExitInfo1 %#RX64\n", pVmcb->ctrl.u64ExitInfo1));
4137	Log4(("ctrl.u64ExitInfo2 %#RX64\n", pVmcb->ctrl.u64ExitInfo2));
4138	Log4(("ctrl.ExitIntInfo.u8Vector %#x\n", pVmcb->ctrl.ExitIntInfo.n.u8Vector));
4139	Log4(("ctrl.ExitIntInfo.u3Type %#x\n", pVmcb->ctrl.ExitIntInfo.n.u3Type));
4140	Log4(("ctrl.ExitIntInfo.u1ErrorCodeValid %#x\n", pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid));
4141	Log4(("ctrl.ExitIntInfo.u19Reserved %#x\n", pVmcb->ctrl.ExitIntInfo.n.u19Reserved));
4142	Log4(("ctrl.ExitIntInfo.u1Valid %#x\n", pVmcb->ctrl.ExitIntInfo.n.u1Valid));
4143	Log4(("ctrl.ExitIntInfo.u32ErrorCode %#x\n", pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode));
4144	Log4(("ctrl.NestedPagingCtrl.u1NestedPaging %#x\n", pVmcb->ctrl.NestedPagingCtrl.n.u1NestedPaging));
4145	Log4(("ctrl.NestedPagingCtrl.u1Sev %#x\n", pVmcb->ctrl.NestedPagingCtrl.n.u1Sev));
4146	Log4(("ctrl.NestedPagingCtrl.u1SevEs %#x\n", pVmcb->ctrl.NestedPagingCtrl.n.u1SevEs));
4147	Log4(("ctrl.EventInject.u8Vector %#x\n", pVmcb->ctrl.EventInject.n.u8Vector));
4148	Log4(("ctrl.EventInject.u3Type %#x\n", pVmcb->ctrl.EventInject.n.u3Type));
4149	Log4(("ctrl.EventInject.u1ErrorCodeValid %#x\n", pVmcb->ctrl.EventInject.n.u1ErrorCodeValid));
4150	Log4(("ctrl.EventInject.u19Reserved %#x\n", pVmcb->ctrl.EventInject.n.u19Reserved));
4151	Log4(("ctrl.EventInject.u1Valid %#x\n", pVmcb->ctrl.EventInject.n.u1Valid));
4152	Log4(("ctrl.EventInject.u32ErrorCode %#x\n", pVmcb->ctrl.EventInject.n.u32ErrorCode));
4153
4154	Log4(("ctrl.u64NestedPagingCR3 %#RX64\n", pVmcb->ctrl.u64NestedPagingCR3));
4155
4156	Log4(("ctrl.LbrVirt.u1LbrVirt %#x\n", pVmcb->ctrl.LbrVirt.n.u1LbrVirt));
4157	Log4(("ctrl.LbrVirt.u1VirtVmsaveVmload %#x\n", pVmcb->ctrl.LbrVirt.n.u1VirtVmsaveVmload));
4158
4159	Log4(("guest.CS.u16Sel %RTsel\n", pVmcb->guest.CS.u16Sel));
4160	Log4(("guest.CS.u16Attr %#x\n", pVmcb->guest.CS.u16Attr));
4161	Log4(("guest.CS.u32Limit %#RX32\n", pVmcb->guest.CS.u32Limit));
4162	Log4(("guest.CS.u64Base %#RX64\n", pVmcb->guest.CS.u64Base));
4163	Log4(("guest.DS.u16Sel %#RTsel\n", pVmcb->guest.DS.u16Sel));
4164	Log4(("guest.DS.u16Attr %#x\n", pVmcb->guest.DS.u16Attr));
4165	Log4(("guest.DS.u32Limit %#RX32\n", pVmcb->guest.DS.u32Limit));
4166	Log4(("guest.DS.u64Base %#RX64\n", pVmcb->guest.DS.u64Base));
4167	Log4(("guest.ES.u16Sel %RTsel\n", pVmcb->guest.ES.u16Sel));
4168	Log4(("guest.ES.u16Attr %#x\n", pVmcb->guest.ES.u16Attr));
4169	Log4(("guest.ES.u32Limit %#RX32\n", pVmcb->guest.ES.u32Limit));
4170	Log4(("guest.ES.u64Base %#RX64\n", pVmcb->guest.ES.u64Base));
4171	Log4(("guest.FS.u16Sel %RTsel\n", pVmcb->guest.FS.u16Sel));
4172	Log4(("guest.FS.u16Attr %#x\n", pVmcb->guest.FS.u16Attr));
4173	Log4(("guest.FS.u32Limit %#RX32\n", pVmcb->guest.FS.u32Limit));
4174	Log4(("guest.FS.u64Base %#RX64\n", pVmcb->guest.FS.u64Base));
4175	Log4(("guest.GS.u16Sel %RTsel\n", pVmcb->guest.GS.u16Sel));
4176	Log4(("guest.GS.u16Attr %#x\n", pVmcb->guest.GS.u16Attr));
4177	Log4(("guest.GS.u32Limit %#RX32\n", pVmcb->guest.GS.u32Limit));
4178	Log4(("guest.GS.u64Base %#RX64\n", pVmcb->guest.GS.u64Base));
4179
4180	Log4(("guest.GDTR.u32Limit %#RX32\n", pVmcb->guest.GDTR.u32Limit));
4181	Log4(("guest.GDTR.u64Base %#RX64\n", pVmcb->guest.GDTR.u64Base));
4182
4183	Log4(("guest.LDTR.u16Sel %RTsel\n", pVmcb->guest.LDTR.u16Sel));
4184	Log4(("guest.LDTR.u16Attr %#x\n", pVmcb->guest.LDTR.u16Attr));
4185	Log4(("guest.LDTR.u32Limit %#RX32\n", pVmcb->guest.LDTR.u32Limit));
4186	Log4(("guest.LDTR.u64Base %#RX64\n", pVmcb->guest.LDTR.u64Base));
4187
4188	Log4(("guest.IDTR.u32Limit %#RX32\n", pVmcb->guest.IDTR.u32Limit));
4189	Log4(("guest.IDTR.u64Base %#RX64\n", pVmcb->guest.IDTR.u64Base));
4190
4191	Log4(("guest.TR.u16Sel %RTsel\n", pVmcb->guest.TR.u16Sel));
4192	Log4(("guest.TR.u16Attr %#x\n", pVmcb->guest.TR.u16Attr));
4193	Log4(("guest.TR.u32Limit %#RX32\n", pVmcb->guest.TR.u32Limit));
4194	Log4(("guest.TR.u64Base %#RX64\n", pVmcb->guest.TR.u64Base));
4195
4196	Log4(("guest.u8CPL %#x\n", pVmcb->guest.u8CPL));
4197	Log4(("guest.u64CR0 %#RX64\n", pVmcb->guest.u64CR0));
4198	Log4(("guest.u64CR2 %#RX64\n", pVmcb->guest.u64CR2));
4199	Log4(("guest.u64CR3 %#RX64\n", pVmcb->guest.u64CR3));
4200	Log4(("guest.u64CR4 %#RX64\n", pVmcb->guest.u64CR4));
4201	Log4(("guest.u64DR6 %#RX64\n", pVmcb->guest.u64DR6));
4202	Log4(("guest.u64DR7 %#RX64\n", pVmcb->guest.u64DR7));
4203
4204	Log4(("guest.u64RIP %#RX64\n", pVmcb->guest.u64RIP));
4205	Log4(("guest.u64RSP %#RX64\n", pVmcb->guest.u64RSP));
4206	Log4(("guest.u64RAX %#RX64\n", pVmcb->guest.u64RAX));
4207	Log4(("guest.u64RFlags %#RX64\n", pVmcb->guest.u64RFlags));
4208
4209	Log4(("guest.u64SysEnterCS %#RX64\n", pVmcb->guest.u64SysEnterCS));
4210	Log4(("guest.u64SysEnterEIP %#RX64\n", pVmcb->guest.u64SysEnterEIP));
4211	Log4(("guest.u64SysEnterESP %#RX64\n", pVmcb->guest.u64SysEnterESP));
4212
4213	Log4(("guest.u64EFER %#RX64\n", pVmcb->guest.u64EFER));
4214	Log4(("guest.u64STAR %#RX64\n", pVmcb->guest.u64STAR));
4215	Log4(("guest.u64LSTAR %#RX64\n", pVmcb->guest.u64LSTAR));
4216	Log4(("guest.u64CSTAR %#RX64\n", pVmcb->guest.u64CSTAR));
4217	Log4(("guest.u64SFMASK %#RX64\n", pVmcb->guest.u64SFMASK));
4218	Log4(("guest.u64KernelGSBase %#RX64\n", pVmcb->guest.u64KernelGSBase));
4219	Log4(("guest.u64PAT %#RX64\n", pVmcb->guest.u64PAT));
4220	Log4(("guest.u64DBGCTL %#RX64\n", pVmcb->guest.u64DBGCTL));
4221	Log4(("guest.u64BR_FROM %#RX64\n", pVmcb->guest.u64BR_FROM));
4222	Log4(("guest.u64BR_TO %#RX64\n", pVmcb->guest.u64BR_TO));
4223	Log4(("guest.u64LASTEXCPFROM %#RX64\n", pVmcb->guest.u64LASTEXCPFROM));
4224	Log4(("guest.u64LASTEXCPTO %#RX64\n", pVmcb->guest.u64LASTEXCPTO));
4225	#endif /* VBOX_STRICT */
4226	}
4227	else
4228	Log4Func(("rcVMRun=%d\n", rcVMRun));
4229
4230	NOREF(pVmcb);
4231	}
4232
4233
4234	/**
4235	* Check per-VM and per-VCPU force flag actions that require us to go back to
4236	* ring-3 for one reason or another.
4237	*
4238	* @returns VBox status code (information status code included).
4239	* @retval VINF_SUCCESS if we don't have any actions that require going back to
4240	* ring-3.
4241	* @retval VINF_PGM_SYNC_CR3 if we have pending PGM CR3 sync.
4242	* @retval VINF_EM_PENDING_REQUEST if we have pending requests (like hardware
4243	* interrupts)
4244	* @retval VINF_PGM_POOL_FLUSH_PENDING if PGM is doing a pool flush and requires
4245	* all EMTs to be in ring-3.
4246	* @retval VINF_EM_RAW_TO_R3 if there is pending DMA requests.
4247	* @retval VINF_EM_NO_MEMORY PGM is out of memory, we need to return
4248	* to the EM loop.
4249	*
4250	* @param pVM The cross context VM structure.
4251	* @param pVCpu The cross context virtual CPU structure.
4252	* @param pCtx Pointer to the guest-CPU context.
4253	*/
4254	static int hmR0SvmCheckForceFlags(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
4255	{
4256	Assert(VMMRZCallRing3IsEnabled(pVCpu));
4257	Assert(!VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_UPDATE_PAE_PDPES));
4258
4259	/* Could happen as a result of longjump. */
4260	if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_UPDATE_CR3))
4261	PGMUpdateCR3(pVCpu, CPUMGetGuestCR3(pVCpu));
4262
4263	/* Update pending interrupts into the APIC's IRR. */
4264	if (VMCPU_FF_TEST_AND_CLEAR(pVCpu, VMCPU_FF_UPDATE_APIC))
4265	APICUpdatePendingInterrupts(pVCpu);
4266
4267	if ( VM_FF_IS_PENDING(pVM, !pVCpu->hm.s.fSingleInstruction
4268	? VM_FF_HP_R0_PRE_HM_MASK : VM_FF_HP_R0_PRE_HM_STEP_MASK)
4269	\|\| VMCPU_FF_IS_PENDING(pVCpu, !pVCpu->hm.s.fSingleInstruction
4270	? VMCPU_FF_HP_R0_PRE_HM_MASK : VMCPU_FF_HP_R0_PRE_HM_STEP_MASK) )
4271	{
4272	/* Pending PGM C3 sync. */
4273	if (VMCPU_FF_IS_PENDING(pVCpu,VMCPU_FF_PGM_SYNC_CR3 \| VMCPU_FF_PGM_SYNC_CR3_NON_GLOBAL))
4274	{
4275	int rc = PGMSyncCR3(pVCpu, pCtx->cr0, pCtx->cr3, pCtx->cr4, VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_PGM_SYNC_CR3));
4276	if (rc != VINF_SUCCESS)
4277	{
4278	Log4Func(("PGMSyncCR3 forcing us back to ring-3. rc=%d\n", rc));
4279	return rc;
4280	}
4281	}
4282
4283	/* Pending HM-to-R3 operations (critsects, timers, EMT rendezvous etc.) */
4284	/* -XXX- what was that about single stepping? */
4285	if ( VM_FF_IS_PENDING(pVM, VM_FF_HM_TO_R3_MASK)
4286	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_TO_R3_MASK))
4287	{
4288	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHmToR3FF);
4289	int rc = RT_UNLIKELY(VM_FF_IS_PENDING(pVM, VM_FF_PGM_NO_MEMORY)) ? VINF_EM_NO_MEMORY : VINF_EM_RAW_TO_R3;
4290	Log4Func(("HM_TO_R3 forcing us back to ring-3. rc=%d\n", rc));
4291	return rc;
4292	}
4293
4294	/* Pending VM request packets, such as hardware interrupts. */
4295	if ( VM_FF_IS_PENDING(pVM, VM_FF_REQUEST)
4296	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_REQUEST))
4297	{
4298	Log4Func(("Pending VM request forcing us back to ring-3\n"));
4299	return VINF_EM_PENDING_REQUEST;
4300	}
4301
4302	/* Pending PGM pool flushes. */
4303	if (VM_FF_IS_PENDING(pVM, VM_FF_PGM_POOL_FLUSH_PENDING))
4304	{
4305	Log4Func(("PGM pool flush pending forcing us back to ring-3\n"));
4306	return VINF_PGM_POOL_FLUSH_PENDING;
4307	}
4308
4309	/* Pending DMA requests. */
4310	if (VM_FF_IS_PENDING(pVM, VM_FF_PDM_DMA))
4311	{
4312	Log4Func(("Pending DMA request forcing us back to ring-3\n"));
4313	return VINF_EM_RAW_TO_R3;
4314	}
4315	}
4316
4317	return VINF_SUCCESS;
4318	}
4319
4320
4321	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
4322	/**
4323	* Does the preparations before executing nested-guest code in AMD-V.
4324	*
4325	* @returns VBox status code (informational status codes included).
4326	* @retval VINF_SUCCESS if we can proceed with running the guest.
4327	* @retval VINF_* scheduling changes, we have to go back to ring-3.
4328	*
4329	* @param pVM The cross context VM structure.
4330	* @param pVCpu The cross context virtual CPU structure.
4331	* @param pCtx Pointer to the nested-guest-CPU context.
4332	* @param pSvmTransient Pointer to the SVM transient structure.
4333	*
4334	* @remarks Same caveats regarding longjumps as hmR0SvmPreRunGuest applies.
4335	* @sa hmR0SvmPreRunGuest.
4336	*/
4337	static int hmR0SvmPreRunGuestNested(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4338	{
4339	HMSVM_ASSERT_PREEMPT_SAFE();
4340	HMSVM_ASSERT_IN_NESTED_GUEST(pCtx);
4341
4342	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM_ONLY_IN_IEM
4343	Log2(("hmR0SvmPreRunGuest: Rescheduling to IEM due to nested-hwvirt or forced IEM exec -> VINF_EM_RESCHEDULE_REM\n"));
4344	return VINF_EM_RESCHEDULE_REM;
4345	#endif
4346
4347	/* Check force flag actions that might require us to go back to ring-3. */
4348	int rc = hmR0SvmCheckForceFlags(pVM, pVCpu, pCtx);
4349	if (rc != VINF_SUCCESS)
4350	return rc;
4351
4352	if (TRPMHasTrap(pVCpu))
4353	hmR0SvmTrpmTrapToPendingEvent(pVCpu);
4354	else if (!pVCpu->hm.s.Event.fPending)
4355	{
4356	VBOXSTRICTRC rcStrict = hmR0SvmEvaluatePendingEventNested(pVCpu, pCtx);
4357	if ( rcStrict != VINF_SUCCESS
4358	\|\| !CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
4359	return VBOXSTRICTRC_VAL(rcStrict);
4360	}
4361
4362	HMSVM_ASSERT_IN_NESTED_GUEST(pCtx);
4363
4364	/*
4365	* On the oldest AMD-V systems, we may not get enough information to reinject an NMI.
4366	* Just do it in software, see @bugref{8411}.
4367	* NB: If we could continue a task switch exit we wouldn't need to do this.
4368	*/
4369	if (RT_UNLIKELY( !pVM->hm.s.svm.u32Features
4370	&& pVCpu->hm.s.Event.fPending
4371	&& SVM_EVENT_GET_TYPE(pVCpu->hm.s.Event.u64IntInfo) == SVM_EVENT_NMI))
4372	{
4373	return VINF_EM_RAW_INJECT_TRPM_EVENT;
4374	}
4375
4376	#ifdef HMSVM_SYNC_FULL_NESTED_GUEST_STATE
4377	Assert(!(pCtx->fExtrn & HMSVM_CPUMCTX_EXTRN_ALL));
4378	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
4379	#endif
4380
4381	/*
4382	* Export the nested-guest state bits that are not shared with the host in any way as we
4383	* can longjmp or get preempted in the midst of exporting some of the state.
4384	*/
4385	rc = hmR0SvmExportGuestStateNested(pVCpu);
4386	AssertRCReturn(rc, rc);
4387	STAM_COUNTER_INC(&pVCpu->hm.s.StatExportFull);
4388
4389	/* Ensure we've cached (and hopefully modified) the VMCB for execution using hardware-assisted SVM. */
4390	Assert(pVCpu->hm.s.svm.NstGstVmcbCache.fCacheValid);
4391
4392	/*
4393	* No longjmps to ring-3 from this point on!!!
4394	*
4395	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional,
4396	* better than a kernel panic. This also disables flushing of the R0-logger instance.
4397	*/
4398	VMMRZCallRing3Disable(pVCpu);
4399
4400	/*
4401	* We disable interrupts so that we don't miss any interrupts that would flag preemption
4402	* (IPI/timers etc.) when thread-context hooks aren't used and we've been running with
4403	* preemption disabled for a while. Since this is purly to aid the
4404	* RTThreadPreemptIsPending() code, it doesn't matter that it may temporarily reenable and
4405	* disable interrupt on NT.
4406	*
4407	* We need to check for force-flags that could've possible been altered since we last
4408	* checked them (e.g. by PDMGetInterrupt() leaving the PDM critical section,
4409	* see @bugref{6398}).
4410	*
4411	* We also check a couple of other force-flags as a last opportunity to get the EMT back
4412	* to ring-3 before executing guest code.
4413	*/
4414	pSvmTransient->fEFlags = ASMIntDisableFlags();
4415	if ( VM_FF_IS_PENDING(pVM, VM_FF_EMT_RENDEZVOUS \| VM_FF_TM_VIRTUAL_SYNC)
4416	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_TO_R3_MASK))
4417	{
4418	ASMSetFlags(pSvmTransient->fEFlags);
4419	VMMRZCallRing3Enable(pVCpu);
4420	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHmToR3FF);
4421	return VINF_EM_RAW_TO_R3;
4422	}
4423	if (RTThreadPreemptIsPending(NIL_RTTHREAD))
4424	{
4425	ASMSetFlags(pSvmTransient->fEFlags);
4426	VMMRZCallRing3Enable(pVCpu);
4427	STAM_COUNTER_INC(&pVCpu->hm.s.StatPendingHostIrq);
4428	return VINF_EM_RAW_INTERRUPT;
4429	}
4430	return VINF_SUCCESS;
4431	}
4432	#endif
4433
4434
4435	/**
4436	* Does the preparations before executing guest code in AMD-V.
4437	*
4438	* This may cause longjmps to ring-3 and may even result in rescheduling to the
4439	* recompiler. We must be cautious what we do here regarding committing
4440	* guest-state information into the VMCB assuming we assuredly execute the guest
4441	* in AMD-V. If we fall back to the recompiler after updating the VMCB and
4442	* clearing the common-state (TRPM/forceflags), we must undo those changes so
4443	* that the recompiler can (and should) use them when it resumes guest
4444	* execution. Otherwise such operations must be done when we can no longer
4445	* exit to ring-3.
4446	*
4447	* @returns VBox status code (informational status codes included).
4448	* @retval VINF_SUCCESS if we can proceed with running the guest.
4449	* @retval VINF_* scheduling changes, we have to go back to ring-3.
4450	*
4451	* @param pVM The cross context VM structure.
4452	* @param pVCpu The cross context virtual CPU structure.
4453	* @param pCtx Pointer to the guest-CPU context.
4454	* @param pSvmTransient Pointer to the SVM transient structure.
4455	*/
4456	static int hmR0SvmPreRunGuest(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4457	{
4458	HMSVM_ASSERT_PREEMPT_SAFE();
4459	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(pCtx);
4460
4461	/* Check force flag actions that might require us to go back to ring-3. */
4462	int rc = hmR0SvmCheckForceFlags(pVM, pVCpu, pCtx);
4463	if (rc != VINF_SUCCESS)
4464	return rc;
4465
4466	if (TRPMHasTrap(pVCpu))
4467	hmR0SvmTrpmTrapToPendingEvent(pVCpu);
4468	else if (!pVCpu->hm.s.Event.fPending)
4469	hmR0SvmEvaluatePendingEvent(pVCpu, pCtx);
4470
4471	/*
4472	* On the oldest AMD-V systems, we may not get enough information to reinject an NMI.
4473	* Just do it in software, see @bugref{8411}.
4474	* NB: If we could continue a task switch exit we wouldn't need to do this.
4475	*/
4476	if (RT_UNLIKELY(pVCpu->hm.s.Event.fPending && (((pVCpu->hm.s.Event.u64IntInfo >> 8) & 7) == SVM_EVENT_NMI)))
4477	if (RT_UNLIKELY(!pVM->hm.s.svm.u32Features))
4478	return VINF_EM_RAW_INJECT_TRPM_EVENT;
4479
4480	#ifdef HMSVM_SYNC_FULL_GUEST_STATE
4481	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
4482	{
4483	Assert(!(pCtx->fExtrn & HMSVM_CPUMCTX_EXTRN_ALL));
4484	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
4485	}
4486	#endif
4487	#ifdef HMSVM_SYNC_FULL_NESTED_GUEST_STATE
4488	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
4489	{
4490	Assert(!(pCtx->fExtrn & HMSVM_CPUMCTX_EXTRN_ALL));
4491	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
4492	}
4493	#endif
4494
4495	/*
4496	* Export the guest state bits that are not shared with the host in any way as we can
4497	* longjmp or get preempted in the midst of exporting some of the state.
4498	*/
4499	rc = hmR0SvmExportGuestState(pVCpu);
4500	AssertRCReturn(rc, rc);
4501	STAM_COUNTER_INC(&pVCpu->hm.s.StatExportFull);
4502
4503	/*
4504	* If we're not intercepting TPR changes in the guest, save the guest TPR before the
4505	* world-switch so we can update it on the way back if the guest changed the TPR.
4506	*/
4507	if (pVCpu->hm.s.svm.fSyncVTpr)
4508	{
4509	PCSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
4510	if (pVM->hm.s.fTPRPatchingActive)
4511	pSvmTransient->u8GuestTpr = pVmcb->guest.u64LSTAR;
4512	else
4513	pSvmTransient->u8GuestTpr = pVmcb->ctrl.IntCtrl.n.u8VTPR;
4514	}
4515
4516	/*
4517	* No longjmps to ring-3 from this point on!!!
4518	*
4519	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional,
4520	* better than a kernel panic. This also disables flushing of the R0-logger instance.
4521	*/
4522	VMMRZCallRing3Disable(pVCpu);
4523
4524	/*
4525	* We disable interrupts so that we don't miss any interrupts that would flag preemption
4526	* (IPI/timers etc.) when thread-context hooks aren't used and we've been running with
4527	* preemption disabled for a while. Since this is purly to aid the
4528	* RTThreadPreemptIsPending() code, it doesn't matter that it may temporarily reenable and
4529	* disable interrupt on NT.
4530	*
4531	* We need to check for force-flags that could've possible been altered since we last
4532	* checked them (e.g. by PDMGetInterrupt() leaving the PDM critical section,
4533	* see @bugref{6398}).
4534	*
4535	* We also check a couple of other force-flags as a last opportunity to get the EMT back
4536	* to ring-3 before executing guest code.
4537	*/
4538	pSvmTransient->fEFlags = ASMIntDisableFlags();
4539	if ( VM_FF_IS_PENDING(pVM, VM_FF_EMT_RENDEZVOUS \| VM_FF_TM_VIRTUAL_SYNC)
4540	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_TO_R3_MASK))
4541	{
4542	ASMSetFlags(pSvmTransient->fEFlags);
4543	VMMRZCallRing3Enable(pVCpu);
4544	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHmToR3FF);
4545	return VINF_EM_RAW_TO_R3;
4546	}
4547	if (RTThreadPreemptIsPending(NIL_RTTHREAD))
4548	{
4549	ASMSetFlags(pSvmTransient->fEFlags);
4550	VMMRZCallRing3Enable(pVCpu);
4551	STAM_COUNTER_INC(&pVCpu->hm.s.StatPendingHostIrq);
4552	return VINF_EM_RAW_INTERRUPT;
4553	}
4554
4555	return VINF_SUCCESS;
4556	}
4557
4558
4559	/**
4560	* Prepares to run guest (or nested-guest) code in AMD-V and we've committed to
4561	* doing so.
4562	*
4563	* This means there is no backing out to ring-3 or anywhere else at this point.
4564	*
4565	* @param pVCpu The cross context virtual CPU structure.
4566	* @param pCtx Pointer to the guest-CPU context.
4567	* @param pSvmTransient Pointer to the SVM transient structure.
4568	*
4569	* @remarks Called with preemption disabled.
4570	* @remarks No-long-jump zone!!!
4571	*/
4572	static void hmR0SvmPreRunGuestCommitted(PVMCPU pVCpu, PCCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4573	{
4574	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
4575	Assert(VMMR0IsLogFlushDisabled(pVCpu));
4576	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
4577
4578	VMCPU_ASSERT_STATE(pVCpu, VMCPUSTATE_STARTED_HM);
4579	VMCPU_SET_STATE(pVCpu, VMCPUSTATE_STARTED_EXEC); /* Indicate the start of guest execution. */
4580
4581	PVM pVM = pVCpu->CTX_SUFF(pVM);
4582	PSVMVMCB pVmcb = pSvmTransient->pVmcb;
4583
4584	hmR0SvmInjectPendingEvent(pVCpu, pCtx, pVmcb);
4585
4586	if (!CPUMIsGuestFPUStateActive(pVCpu))
4587	{
4588	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatLoadGuestFpuState, x);
4589	CPUMR0LoadGuestFPU(pVM, pVCpu); /* (Ignore rc, no need to set HM_CHANGED_HOST_CONTEXT for SVM.) */
4590	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatLoadGuestFpuState, x);
4591	STAM_COUNTER_INC(&pVCpu->hm.s.StatLoadGuestFpu);
4592	}
4593
4594	/* Load the state shared between host and guest (FPU, debug). */
4595	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE)
4596	hmR0SvmExportSharedState(pVCpu, pVmcb);
4597
4598	pVCpu->hm.s.fCtxChanged &= ~HM_CHANGED_HOST_CONTEXT; /* Preemption might set this, nothing to do on AMD-V. */
4599	AssertMsg(!pVCpu->hm.s.fCtxChanged, ("fCtxChanged=%#RX64\n", pVCpu->hm.s.fCtxChanged));
4600
4601	PHMGLOBALCPUINFO pHostCpu = hmR0GetCurrentCpu();
4602	RTCPUID const idHostCpu = pHostCpu->idCpu;
4603	bool const fMigratedHostCpu = idHostCpu != pVCpu->hm.s.idLastCpu;
4604
4605	/* Setup TSC offsetting. */
4606	if ( pSvmTransient->fUpdateTscOffsetting
4607	\|\| fMigratedHostCpu)
4608	{
4609	hmR0SvmUpdateTscOffsetting(pVM, pVCpu, pCtx, pVmcb);
4610	pSvmTransient->fUpdateTscOffsetting = false;
4611	}
4612
4613	/* If we've migrating CPUs, mark the VMCB Clean bits as dirty. */
4614	if (fMigratedHostCpu)
4615	pVmcb->ctrl.u32VmcbCleanBits = 0;
4616
4617	/* Store status of the shared guest-host state at the time of VMRUN. */
4618	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS)
4619	if (CPUMIsGuestInLongModeEx(pCtx))
4620	{
4621	pSvmTransient->fWasGuestDebugStateActive = CPUMIsGuestDebugStateActivePending(pVCpu);
4622	pSvmTransient->fWasHyperDebugStateActive = CPUMIsHyperDebugStateActivePending(pVCpu);
4623	}
4624	else
4625	#endif
4626	{
4627	pSvmTransient->fWasGuestDebugStateActive = CPUMIsGuestDebugStateActive(pVCpu);
4628	pSvmTransient->fWasHyperDebugStateActive = CPUMIsHyperDebugStateActive(pVCpu);
4629	}
4630
4631	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
4632	uint8_t *pbMsrBitmap;
4633	if (!pSvmTransient->fIsNestedGuest)
4634	pbMsrBitmap = (uint8_t *)pVCpu->hm.s.svm.pvMsrBitmap;
4635	else
4636	{
4637	hmR0SvmMergeMsrpmNested(pHostCpu, pVCpu, pCtx);
4638
4639	/* Update the nested-guest VMCB with the newly merged MSRPM (clean bits updated below). */
4640	pVmcb->ctrl.u64MSRPMPhysAddr = pHostCpu->n.svm.HCPhysNstGstMsrpm;
4641	pbMsrBitmap = (uint8_t *)pHostCpu->n.svm.pvNstGstMsrpm;
4642	}
4643	#else
4644	uint8_t pbMsrBitmap = (uint8_t )pVCpu->hm.s.svm.pvMsrBitmap;
4645	#endif
4646
4647	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, true); /* Used for TLB flushing, set this across the world switch. */
4648	/* Flush the appropriate tagged-TLB entries. */
4649	hmR0SvmFlushTaggedTlb(pVCpu, pCtx, pVmcb, pHostCpu);
4650	Assert(pVCpu->hm.s.idLastCpu == idHostCpu);
4651
4652	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatEntry, &pVCpu->hm.s.StatInGC, x);
4653
4654	TMNotifyStartOfExecution(pVCpu); /* Finally, notify TM to resume its clocks as we're about
4655	to start executing. */
4656
4657	/*
4658	* Save the current Host TSC_AUX and write the guest TSC_AUX to the host, so that RDTSCPs
4659	* (that don't cause exits) reads the guest MSR, see @bugref{3324}.
4660	*
4661	* This should be done -after- any RDTSCPs for obtaining the host timestamp (TM, STAM etc).
4662	*/
4663	if ( (pVM->hm.s.cpuid.u32AMDFeatureEDX & X86_CPUID_EXT_FEATURE_EDX_RDTSCP)
4664	&& !(pVmcb->ctrl.u64InterceptCtrl & SVM_CTRL_INTERCEPT_RDTSCP))
4665	{
4666	uint64_t const uGuestTscAux = CPUMGetGuestTscAux(pVCpu);
4667	pVCpu->hm.s.u64HostTscAux = ASMRdMsr(MSR_K8_TSC_AUX);
4668	if (uGuestTscAux != pVCpu->hm.s.u64HostTscAux)
4669	ASMWrMsr(MSR_K8_TSC_AUX, uGuestTscAux);
4670	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_K8_TSC_AUX, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
4671	pSvmTransient->fRestoreTscAuxMsr = true;
4672	}
4673	else
4674	{
4675	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_K8_TSC_AUX, SVMMSREXIT_INTERCEPT_READ, SVMMSREXIT_INTERCEPT_WRITE);
4676	pSvmTransient->fRestoreTscAuxMsr = false;
4677	}
4678	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_IOPM_MSRPM;
4679
4680	/*
4681	* If VMCB Clean bits isn't supported by the CPU or exposed to the guest in the nested
4682	* virtualization case, mark all state-bits as dirty indicating to the CPU to re-load
4683	* from the VMCB.
4684	*/
4685	bool const fSupportsVmcbCleanBits = hmR0SvmSupportsVmcbCleanBits(pVCpu, pCtx);
4686	if (!fSupportsVmcbCleanBits)
4687	pVmcb->ctrl.u32VmcbCleanBits = 0;
4688	}
4689
4690
4691	/**
4692	* Wrapper for running the guest code in AMD-V.
4693	*
4694	* @returns VBox strict status code.
4695	* @param pVM The cross context VM structure.
4696	* @param pVCpu The cross context virtual CPU structure.
4697	* @param pCtx Pointer to the guest-CPU context.
4698	*
4699	* @remarks No-long-jump zone!!!
4700	*/
4701	DECLINLINE(int) hmR0SvmRunGuest(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
4702	{
4703	/* Mark that HM is the keeper of all guest-CPU registers now that we're going to execute guest code. */
4704	pCtx->fExtrn \|= HMSVM_CPUMCTX_EXTRN_ALL \| CPUMCTX_EXTRN_KEEPER_HM;
4705
4706	/*
4707	* 64-bit Windows uses XMM registers in the kernel as the Microsoft compiler expresses
4708	* floating-point operations using SSE instructions. Some XMM registers (XMM6-XMM15) are
4709	* callee-saved and thus the need for this XMM wrapper.
4710	*
4711	* Refer MSDN "Configuring Programs for 64-bit/x64 Software Conventions / Register Usage".
4712	*/
4713	#ifdef VBOX_WITH_KERNEL_USING_XMM
4714	return hmR0SVMRunWrapXMM(pVCpu->hm.s.svm.HCPhysVmcbHost, pVCpu->hm.s.svm.HCPhysVmcb, pCtx, pVM, pVCpu,
4715	pVCpu->hm.s.svm.pfnVMRun);
4716	#else
4717	return pVCpu->hm.s.svm.pfnVMRun(pVCpu->hm.s.svm.HCPhysVmcbHost, pVCpu->hm.s.svm.HCPhysVmcb, pCtx, pVM, pVCpu);
4718	#endif
4719	}
4720
4721
4722	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
4723	/**
4724	* Wrapper for running the nested-guest code in AMD-V.
4725	*
4726	* @returns VBox strict status code.
4727	* @param pVM The cross context VM structure.
4728	* @param pVCpu The cross context virtual CPU structure.
4729	* @param pCtx Pointer to the guest-CPU context.
4730	*
4731	* @remarks No-long-jump zone!!!
4732	*/
4733	DECLINLINE(int) hmR0SvmRunGuestNested(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
4734	{
4735	/* Mark that HM is the keeper of all guest-CPU registers now that we're going to execute guest code. */
4736	pCtx->fExtrn \|= HMSVM_CPUMCTX_EXTRN_ALL \| CPUMCTX_EXTRN_KEEPER_HM;
4737
4738	/*
4739	* 64-bit Windows uses XMM registers in the kernel as the Microsoft compiler expresses
4740	* floating-point operations using SSE instructions. Some XMM registers (XMM6-XMM15) are
4741	* callee-saved and thus the need for this XMM wrapper.
4742	*
4743	* Refer MSDN "Configuring Programs for 64-bit/x64 Software Conventions / Register Usage".
4744	*/
4745	#ifdef VBOX_WITH_KERNEL_USING_XMM
4746	return hmR0SVMRunWrapXMM(pVCpu->hm.s.svm.HCPhysVmcbHost, pCtx->hwvirt.svm.HCPhysVmcb, pCtx, pVM, pVCpu,
4747	pVCpu->hm.s.svm.pfnVMRun);
4748	#else
4749	return pVCpu->hm.s.svm.pfnVMRun(pVCpu->hm.s.svm.HCPhysVmcbHost, pCtx->hwvirt.svm.HCPhysVmcb, pCtx, pVM, pVCpu);
4750	#endif
4751	}
4752
4753
4754	/**
4755	* Undoes the TSC offset applied for an SVM nested-guest and returns the TSC
4756	* value for the guest.
4757	*
4758	* @returns The TSC offset after undoing any nested-guest TSC offset.
4759	* @param pVCpu The cross context virtual CPU structure of the calling EMT.
4760	* @param uTicks The nested-guest TSC.
4761	*
4762	* @note If you make any changes to this function, please check if
4763	* hmR0SvmNstGstUndoTscOffset() needs adjusting.
4764	*
4765	* @sa HMSvmNstGstApplyTscOffset().
4766	*/
4767	DECLINLINE(uint64_t) hmR0SvmNstGstUndoTscOffset(PVMCPU pVCpu, uint64_t uTicks)
4768	{
4769	PCSVMNESTEDVMCBCACHE pVmcbNstGstCache = &pVCpu->hm.s.svm.NstGstVmcbCache;
4770	Assert(pVmcbNstGstCache->fCacheValid);
4771	return uTicks - pVmcbNstGstCache->u64TSCOffset;
4772	}
4773	#endif
4774
4775	/**
4776	* Performs some essential restoration of state after running guest (or
4777	* nested-guest) code in AMD-V.
4778	*
4779	* @param pVCpu The cross context virtual CPU structure.
4780	* @param pCtx Pointer to the guest-CPU context. The data maybe
4781	* out-of-sync. Make sure to update the required fields
4782	* before using them.
4783	* @param pSvmTransient Pointer to the SVM transient structure.
4784	* @param rcVMRun Return code of VMRUN.
4785	*
4786	* @remarks Called with interrupts disabled.
4787	* @remarks No-long-jump zone!!! This function will however re-enable longjmps
4788	* unconditionally when it is safe to do so.
4789	*/
4790	static void hmR0SvmPostRunGuest(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient, int rcVMRun)
4791	{
4792	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
4793
4794	uint64_t const uHostTsc = ASMReadTSC(); /* Read the TSC as soon as possible. */
4795	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, false); /* See HMInvalidatePageOnAllVCpus(): used for TLB flushing. */
4796	ASMAtomicIncU32(&pVCpu->hm.s.cWorldSwitchExits); /* Initialized in vmR3CreateUVM(): used for EMT poking. */
4797
4798	PSVMVMCB pVmcb = pSvmTransient->pVmcb;
4799	PSVMVMCBCTRL pVmcbCtrl = &pVmcb->ctrl;
4800
4801	/* TSC read must be done early for maximum accuracy. */
4802	if (!(pVmcbCtrl->u64InterceptCtrl & SVM_CTRL_INTERCEPT_RDTSC))
4803	{
4804	if (!pSvmTransient->fIsNestedGuest)
4805	TMCpuTickSetLastSeen(pVCpu, uHostTsc + pVmcbCtrl->u64TSCOffset);
4806	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
4807	else
4808	{
4809	/* The nested-guest VMCB TSC offset shall eventually be restored on #VMEXIT via HMSvmNstGstVmExitNotify(). */
4810	uint64_t const uGstTsc = hmR0SvmNstGstUndoTscOffset(pVCpu, uHostTsc + pVmcbCtrl->u64TSCOffset);
4811	TMCpuTickSetLastSeen(pVCpu, uGstTsc);
4812	}
4813	#endif
4814	}
4815
4816	if (pSvmTransient->fRestoreTscAuxMsr)
4817	{
4818	uint64_t u64GuestTscAuxMsr = ASMRdMsr(MSR_K8_TSC_AUX);
4819	CPUMSetGuestTscAux(pVCpu, u64GuestTscAuxMsr);
4820	if (u64GuestTscAuxMsr != pVCpu->hm.s.u64HostTscAux)
4821	ASMWrMsr(MSR_K8_TSC_AUX, pVCpu->hm.s.u64HostTscAux);
4822	}
4823
4824	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatInGC, &pVCpu->hm.s.StatExit1, x);
4825	TMNotifyEndOfExecution(pVCpu); /* Notify TM that the guest is no longer running. */
4826	VMCPU_SET_STATE(pVCpu, VMCPUSTATE_STARTED_HM);
4827
4828	Assert(!(ASMGetFlags() & X86_EFL_IF));
4829	ASMSetFlags(pSvmTransient->fEFlags); /* Enable interrupts. */
4830	VMMRZCallRing3Enable(pVCpu); /* It is now safe to do longjmps to ring-3!!! */
4831
4832	/* If VMRUN failed, we can bail out early. This does -not- cover SVM_EXIT_INVALID. */
4833	if (RT_UNLIKELY(rcVMRun != VINF_SUCCESS))
4834	{
4835	Log4Func(("VMRUN failure: rcVMRun=%Rrc\n", rcVMRun));
4836	return;
4837	}
4838
4839	pSvmTransient->u64ExitCode = pVmcbCtrl->u64ExitCode; /* Save the #VMEXIT reason. */
4840	pVmcbCtrl->u32VmcbCleanBits = HMSVM_VMCB_CLEAN_ALL; /* Mark the VMCB-state cache as unmodified by VMM. */
4841	pSvmTransient->fVectoringDoublePF = false; /* Vectoring double page-fault needs to be determined later. */
4842	pSvmTransient->fVectoringPF = false; /* Vectoring page-fault needs to be determined later. */
4843
4844	#ifdef HMSVM_SYNC_FULL_GUEST_STATE
4845	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
4846	hmR0SvmImportGuestState(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
4847	#elif defined(HMSVM_SYNC_FULL_NESTED_GUEST_STATE)
4848	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
4849	hmR0SvmImportGuestState(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
4850	#else
4851	/*
4852	* Always import the following:
4853	*
4854	* - RIP for exit optimizations and evaluating event injection on re-entry.
4855	* - RFLAGS for evaluating event injection on VM re-entry and for exporting shared debug
4856	* state on preemption.
4857	* - Interrupt shadow, GIF for evaluating event injection on VM re-entry.
4858	* - CS for exit optimizations.
4859	* - RAX, RSP for simplifying assumptions on GPRs. All other GPRs are swapped by the
4860	* assembly switcher code.
4861	* - Shared state (only DR7 currently) for exporting shared debug state on preemption.
4862	*/
4863	hmR0SvmImportGuestState(pVCpu, CPUMCTX_EXTRN_RIP
4864	\| CPUMCTX_EXTRN_RFLAGS
4865	\| CPUMCTX_EXTRN_RAX
4866	\| CPUMCTX_EXTRN_RSP
4867	\| CPUMCTX_EXTRN_CS
4868	\| CPUMCTX_EXTRN_HWVIRT
4869	\| CPUMCTX_EXTRN_HM_SVM_INT_SHADOW
4870	\| CPUMCTX_EXTRN_HM_SVM_HWVIRT_VIRQ
4871	\| HMSVM_CPUMCTX_SHARED_STATE);
4872	#endif
4873
4874	#ifdef DEBUG_ramshankar
4875	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
4876	{
4877	hmR0SvmImportGuestState(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
4878	hmR0SvmLogState(pVCpu, pVmcb, pCtx, "hmR0SvmPostRunGuestNested", HMSVM_LOG_ALL & ~HMSVM_LOG_LBR, 0 /* uVerbose */);
4879	}
4880	#endif
4881
4882	if ( pSvmTransient->u64ExitCode != SVM_EXIT_INVALID
4883	&& pVCpu->hm.s.svm.fSyncVTpr)
4884	{
4885	Assert(!pSvmTransient->fIsNestedGuest);
4886	/* TPR patching (for 32-bit guests) uses LSTAR MSR for holding the TPR value, otherwise uses the VTPR. */
4887	if ( pVCpu->CTX_SUFF(pVM)->hm.s.fTPRPatchingActive
4888	&& (pVmcb->guest.u64LSTAR & 0xff) != pSvmTransient->u8GuestTpr)
4889	{
4890	int rc = APICSetTpr(pVCpu, pVmcb->guest.u64LSTAR & 0xff);
4891	AssertRC(rc);
4892	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
4893	}
4894	/* Sync TPR when we aren't intercepting CR8 writes. */
4895	else if (pSvmTransient->u8GuestTpr != pVmcbCtrl->IntCtrl.n.u8VTPR)
4896	{
4897	int rc = APICSetTpr(pVCpu, pVmcbCtrl->IntCtrl.n.u8VTPR << 4);
4898	AssertRC(rc);
4899	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
4900	}
4901	}
4902
4903	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_CS \| CPUMCTX_EXTRN_RIP);
4904	EMHistoryAddExit(pVCpu, EMEXIT_MAKE_FT(EMEXIT_F_KIND_SVM, pSvmTransient->u64ExitCode & EMEXIT_F_TYPE_MASK),
4905	pCtx->cs.u64Base + pCtx->rip, uHostTsc);
4906	}
4907
4908
4909	/**
4910	* Runs the guest code using AMD-V.
4911	*
4912	* @returns VBox status code.
4913	* @param pVM The cross context VM structure.
4914	* @param pVCpu The cross context virtual CPU structure.
4915	* @param pCtx Pointer to the guest-CPU context.
4916	* @param pcLoops Pointer to the number of executed loops.
4917	*/
4918	static int hmR0SvmRunGuestCodeNormal(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t *pcLoops)
4919	{
4920	uint32_t const cMaxResumeLoops = pVM->hm.s.cMaxResumeLoops;
4921	Assert(pcLoops);
4922	Assert(*pcLoops <= cMaxResumeLoops);
4923
4924	SVMTRANSIENT SvmTransient;
4925	RT_ZERO(SvmTransient);
4926	SvmTransient.fUpdateTscOffsetting = true;
4927	SvmTransient.pVmcb = pVCpu->hm.s.svm.pVmcb;
4928
4929	int rc = VERR_INTERNAL_ERROR_5;
4930	for (;;)
4931	{
4932	Assert(!HMR0SuspendPending());
4933	HMSVM_ASSERT_CPU_SAFE();
4934
4935	/* Preparatory work for running nested-guest code, this may force us to return to
4936	ring-3. This bugger disables interrupts on VINF_SUCCESS! */
4937	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatEntry, x);
4938	rc = hmR0SvmPreRunGuest(pVM, pVCpu, pCtx, &SvmTransient);
4939	if (rc != VINF_SUCCESS)
4940	break;
4941
4942	/*
4943	* No longjmps to ring-3 from this point on!!!
4944	*
4945	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional,
4946	* better than a kernel panic. This also disables flushing of the R0-logger instance.
4947	*/
4948	hmR0SvmPreRunGuestCommitted(pVCpu, pCtx, &SvmTransient);
4949	rc = hmR0SvmRunGuest(pVM, pVCpu, pCtx);
4950
4951	/* Restore any residual host-state and save any bits shared between host and guest
4952	into the guest-CPU state. Re-enables interrupts! */
4953	hmR0SvmPostRunGuest(pVCpu, pCtx, &SvmTransient, rc);
4954
4955	if (RT_UNLIKELY( rc != VINF_SUCCESS /* Check for VMRUN errors. */
4956	\|\| SvmTransient.u64ExitCode == SVM_EXIT_INVALID)) /* Check for invalid guest-state errors. */
4957	{
4958	if (rc == VINF_SUCCESS)
4959	rc = VERR_SVM_INVALID_GUEST_STATE;
4960	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit1, x);
4961	hmR0SvmReportWorldSwitchError(pVM, pVCpu, rc, pCtx);
4962	break;
4963	}
4964
4965	/* Handle the #VMEXIT. */
4966	HMSVM_EXITCODE_STAM_COUNTER_INC(SvmTransient.u64ExitCode);
4967	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatExit1, &pVCpu->hm.s.StatExit2, x);
4968	VBOXVMM_R0_HMSVM_VMEXIT(pVCpu, pCtx, SvmTransient.u64ExitCode, pVCpu->hm.s.svm.pVmcb);
4969	rc = hmR0SvmHandleExit(pVCpu, pCtx, &SvmTransient);
4970	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit2, x);
4971	if (rc != VINF_SUCCESS)
4972	break;
4973	if (++(*pcLoops) >= cMaxResumeLoops)
4974	{
4975	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchMaxResumeLoops);
4976	rc = VINF_EM_RAW_INTERRUPT;
4977	break;
4978	}
4979	}
4980
4981	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatEntry, x);
4982	return rc;
4983	}
4984
4985
4986	/**
4987	* Runs the guest code using AMD-V in single step mode.
4988	*
4989	* @returns VBox status code.
4990	* @param pVM The cross context VM structure.
4991	* @param pVCpu The cross context virtual CPU structure.
4992	* @param pCtx Pointer to the guest-CPU context.
4993	* @param pcLoops Pointer to the number of executed loops.
4994	*/
4995	static int hmR0SvmRunGuestCodeStep(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t *pcLoops)
4996	{
4997	uint32_t const cMaxResumeLoops = pVM->hm.s.cMaxResumeLoops;
4998	Assert(pcLoops);
4999	Assert(*pcLoops <= cMaxResumeLoops);
5000
5001	SVMTRANSIENT SvmTransient;
5002	RT_ZERO(SvmTransient);
5003	SvmTransient.fUpdateTscOffsetting = true;
5004	SvmTransient.pVmcb = pVCpu->hm.s.svm.pVmcb;
5005
5006	uint16_t uCsStart = pCtx->cs.Sel;
5007	uint64_t uRipStart = pCtx->rip;
5008
5009	int rc = VERR_INTERNAL_ERROR_5;
5010	for (;;)
5011	{
5012	Assert(!HMR0SuspendPending());
5013	AssertMsg(pVCpu->hm.s.idEnteredCpu == RTMpCpuId(),
5014	("Illegal migration! Entered on CPU %u Current %u cLoops=%u\n", (unsigned)pVCpu->hm.s.idEnteredCpu,
5015	(unsigned)RTMpCpuId(), *pcLoops));
5016
5017	/* Preparatory work for running nested-guest code, this may force us to return to
5018	ring-3. This bugger disables interrupts on VINF_SUCCESS! */
5019	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatEntry, x);
5020	rc = hmR0SvmPreRunGuest(pVM, pVCpu, pCtx, &SvmTransient);
5021	if (rc != VINF_SUCCESS)
5022	break;
5023
5024	/*
5025	* No longjmps to ring-3 from this point on!!!
5026	*
5027	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional,
5028	* better than a kernel panic. This also disables flushing of the R0-logger instance.
5029	*/
5030	VMMRZCallRing3Disable(pVCpu);
5031	VMMRZCallRing3RemoveNotification(pVCpu);
5032	hmR0SvmPreRunGuestCommitted(pVCpu, pCtx, &SvmTransient);
5033
5034	rc = hmR0SvmRunGuest(pVM, pVCpu, pCtx);
5035
5036	/* Restore any residual host-state and save any bits shared between host and guest
5037	into the guest-CPU state. Re-enables interrupts! */
5038	hmR0SvmPostRunGuest(pVCpu, pCtx, &SvmTransient, rc);
5039
5040	if (RT_UNLIKELY( rc != VINF_SUCCESS /* Check for VMRUN errors. */
5041	\|\| SvmTransient.u64ExitCode == SVM_EXIT_INVALID)) /* Check for invalid guest-state errors. */
5042	{
5043	if (rc == VINF_SUCCESS)
5044	rc = VERR_SVM_INVALID_GUEST_STATE;
5045	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit1, x);
5046	hmR0SvmReportWorldSwitchError(pVM, pVCpu, rc, pCtx);
5047	return rc;
5048	}
5049
5050	/* Handle the #VMEXIT. */
5051	HMSVM_EXITCODE_STAM_COUNTER_INC(SvmTransient.u64ExitCode);
5052	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatExit1, &pVCpu->hm.s.StatExit2, x);
5053	VBOXVMM_R0_HMSVM_VMEXIT(pVCpu, pCtx, SvmTransient.u64ExitCode, pVCpu->hm.s.svm.pVmcb);
5054	rc = hmR0SvmHandleExit(pVCpu, pCtx, &SvmTransient);
5055	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit2, x);
5056	if (rc != VINF_SUCCESS)
5057	break;
5058	if (++(*pcLoops) >= cMaxResumeLoops)
5059	{
5060	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchMaxResumeLoops);
5061	rc = VINF_EM_RAW_INTERRUPT;
5062	break;
5063	}
5064
5065	/*
5066	* Did the RIP change, if so, consider it a single step.
5067	* Otherwise, make sure one of the TFs gets set.
5068	*/
5069	if ( pCtx->rip != uRipStart
5070	\|\| pCtx->cs.Sel != uCsStart)
5071	{
5072	rc = VINF_EM_DBG_STEPPED;
5073	break;
5074	}
5075	pVCpu->hm.s.fCtxChanged \|= HM_CHANGED_GUEST_DR_MASK;
5076	}
5077
5078	/*
5079	* Clear the X86_EFL_TF if necessary.
5080	*/
5081	if (pVCpu->hm.s.fClearTrapFlag)
5082	{
5083	pVCpu->hm.s.fClearTrapFlag = false;
5084	pCtx->eflags.Bits.u1TF = 0;
5085	}
5086
5087	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatEntry, x);
5088	return rc;
5089	}
5090
5091	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
5092	/**
5093	* Runs the nested-guest code using AMD-V.
5094	*
5095	* @returns VBox status code.
5096	* @param pVM The cross context VM structure.
5097	* @param pVCpu The cross context virtual CPU structure.
5098	* @param pCtx Pointer to the guest-CPU context.
5099	* @param pcLoops Pointer to the number of executed loops. If we're switching
5100	* from the guest-code execution loop to this nested-guest
5101	* execution loop pass the remainder value, else pass 0.
5102	*/
5103	static int hmR0SvmRunGuestCodeNested(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t *pcLoops)
5104	{
5105	HMSVM_ASSERT_IN_NESTED_GUEST(pCtx);
5106	Assert(pcLoops);
5107	Assert(*pcLoops <= pVM->hm.s.cMaxResumeLoops);
5108
5109	SVMTRANSIENT SvmTransient;
5110	RT_ZERO(SvmTransient);
5111	SvmTransient.fUpdateTscOffsetting = true;
5112	SvmTransient.pVmcb = pCtx->hwvirt.svm.CTX_SUFF(pVmcb);
5113	SvmTransient.fIsNestedGuest = true;
5114
5115	int rc = VERR_INTERNAL_ERROR_4;
5116	for (;;)
5117	{
5118	Assert(!HMR0SuspendPending());
5119	HMSVM_ASSERT_CPU_SAFE();
5120
5121	/* Preparatory work for running nested-guest code, this may force us to return to
5122	ring-3. This bugger disables interrupts on VINF_SUCCESS! */
5123	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatEntry, x);
5124	rc = hmR0SvmPreRunGuestNested(pVM, pVCpu, pCtx, &SvmTransient);
5125	if ( rc != VINF_SUCCESS
5126	\|\| !CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
5127	{
5128	break;
5129	}
5130
5131	/*
5132	* No longjmps to ring-3 from this point on!!!
5133	*
5134	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional,
5135	* better than a kernel panic. This also disables flushing of the R0-logger instance.
5136	*/
5137	hmR0SvmPreRunGuestCommitted(pVCpu, pCtx, &SvmTransient);
5138
5139	rc = hmR0SvmRunGuestNested(pVM, pVCpu, pCtx);
5140
5141	/* Restore any residual host-state and save any bits shared between host and guest
5142	into the guest-CPU state. Re-enables interrupts! */
5143	hmR0SvmPostRunGuest(pVCpu, pCtx, &SvmTransient, rc);
5144
5145	if (RT_LIKELY( rc == VINF_SUCCESS
5146	&& SvmTransient.u64ExitCode != SVM_EXIT_INVALID))
5147	{ /* extremely likely */ }
5148	else
5149	{
5150	/* VMRUN failed, shouldn't really happen, Guru. */
5151	if (rc != VINF_SUCCESS)
5152	break;
5153
5154	/* Invalid nested-guest state. Cause a #VMEXIT but assert on strict builds. */
5155	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
5156	AssertMsgFailed(("Invalid nested-guest state. rc=%Rrc u64ExitCode=%#RX64\n", rc, SvmTransient.u64ExitCode));
5157	rc = VBOXSTRICTRC_TODO(IEMExecSvmVmexit(pVCpu, SVM_EXIT_INVALID, 0, 0));
5158	break;
5159	}
5160
5161	/* Handle the #VMEXIT. */
5162	HMSVM_NESTED_EXITCODE_STAM_COUNTER_INC(SvmTransient.u64ExitCode);
5163	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatExit1, &pVCpu->hm.s.StatExit2, x);
5164	VBOXVMM_R0_HMSVM_VMEXIT(pVCpu, pCtx, SvmTransient.u64ExitCode, pCtx->hwvirt.svm.CTX_SUFF(pVmcb));
5165	rc = hmR0SvmHandleExitNested(pVCpu, pCtx, &SvmTransient);
5166	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit2, x);
5167	if ( rc != VINF_SUCCESS
5168	\|\| !CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
5169	break;
5170	if (++(*pcLoops) >= pVM->hm.s.cMaxResumeLoops)
5171	{
5172	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchMaxResumeLoops);
5173	rc = VINF_EM_RAW_INTERRUPT;
5174	break;
5175	}
5176
5177	/** @todo handle single-stepping */
5178	}
5179
5180	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatEntry, x);
5181	return rc;
5182	}
5183	#endif
5184
5185
5186	/**
5187	* Runs the guest code using AMD-V.
5188	*
5189	* @returns Strict VBox status code.
5190	* @param pVM The cross context VM structure.
5191	* @param pVCpu The cross context virtual CPU structure.
5192	* @param pCtx Pointer to the guest-CPU context.
5193	*/
5194	VMMR0DECL(VBOXSTRICTRC) SVMR0RunGuestCode(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
5195	{
5196	Assert(VMMRZCallRing3IsEnabled(pVCpu));
5197	HMSVM_ASSERT_PREEMPT_SAFE();
5198	VMMRZCallRing3SetNotification(pVCpu, hmR0SvmCallRing3Callback, pCtx);
5199
5200	uint32_t cLoops = 0;
5201	int rc;
5202	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
5203	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
5204	#endif
5205	{
5206	if (!pVCpu->hm.s.fSingleInstruction)
5207	rc = hmR0SvmRunGuestCodeNormal(pVM, pVCpu, pCtx, &cLoops);
5208	else
5209	rc = hmR0SvmRunGuestCodeStep(pVM, pVCpu, pCtx, &cLoops);
5210	}
5211	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
5212	else
5213	{
5214	rc = VINF_SVM_VMRUN;
5215	}
5216
5217	/* Re-check the nested-guest condition here as we may be transitioning from the normal
5218	execution loop into the nested-guest, hence this is not placed in the 'else' part above. */
5219	if (rc == VINF_SVM_VMRUN)
5220	{
5221	rc = hmR0SvmRunGuestCodeNested(pVM, pVCpu, pCtx, &cLoops);
5222	if (rc == VINF_SVM_VMEXIT)
5223	rc = VINF_SUCCESS;
5224	}
5225	#endif
5226
5227	/* Fixup error codes. */
5228	if (rc == VERR_EM_INTERPRETER)
5229	rc = VINF_EM_RAW_EMULATE_INSTR;
5230	else if (rc == VINF_EM_RESET)
5231	rc = VINF_EM_TRIPLE_FAULT;
5232
5233	/* Prepare to return to ring-3. This will remove longjmp notifications. */
5234	rc = hmR0SvmExitToRing3(pVM, pVCpu, pCtx, rc);
5235	Assert(!VMMRZCallRing3IsNotificationSet(pVCpu));
5236	return rc;
5237	}
5238
5239
5240	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
5241	/**
5242	* Determines whether an IOIO intercept is active for the nested-guest or not.
5243	*
5244	* @param pvIoBitmap Pointer to the nested-guest IO bitmap.
5245	* @param pIoExitInfo Pointer to the SVMIOIOEXITINFO.
5246	*/
5247	static bool hmR0SvmIsIoInterceptActive(void *pvIoBitmap, PSVMIOIOEXITINFO pIoExitInfo)
5248	{
5249	const uint16_t u16Port = pIoExitInfo->n.u16Port;
5250	const SVMIOIOTYPE enmIoType = (SVMIOIOTYPE)pIoExitInfo->n.u1Type;
5251	const uint8_t cbReg = (pIoExitInfo->u >> SVM_IOIO_OP_SIZE_SHIFT) & 7;
5252	const uint8_t cAddrSizeBits = ((pIoExitInfo->u >> SVM_IOIO_ADDR_SIZE_SHIFT) & 7) << 4;
5253	const uint8_t iEffSeg = pIoExitInfo->n.u3Seg;
5254	const bool fRep = pIoExitInfo->n.u1Rep;
5255	const bool fStrIo = pIoExitInfo->n.u1Str;
5256
5257	return HMSvmIsIOInterceptActive(pvIoBitmap, u16Port, enmIoType, cbReg, cAddrSizeBits, iEffSeg, fRep, fStrIo,
5258	NULL /* pIoExitInfo */);
5259	}
5260
5261
5262	/**
5263	* Handles a nested-guest \#VMEXIT (for all EXITCODE values except
5264	* SVM_EXIT_INVALID).
5265	*
5266	* @returns VBox status code (informational status codes included).
5267	* @param pVCpu The cross context virtual CPU structure.
5268	* @param pCtx Pointer to the guest-CPU context.
5269	* @param pSvmTransient Pointer to the SVM transient structure.
5270	*/
5271	static int hmR0SvmHandleExitNested(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5272	{
5273	HMSVM_ASSERT_IN_NESTED_GUEST(pCtx);
5274	Assert(pSvmTransient->u64ExitCode != SVM_EXIT_INVALID);
5275	Assert(pSvmTransient->u64ExitCode <= SVM_EXIT_MAX);
5276
5277	/** @todo Figure out why using IEM_CPUMCTX_EXTRN_SVM_VMEXIT_MASK instead of
5278	* HMSVM_CPUMCTX_EXTRN_ALL breaks nested guests (XP Pro, DSL etc.), see
5279	* also HMSvmNstGstVmExitNotify(). */
5280	#define NST_GST_VMEXIT_CALL_RET(a_pVCpu, a_pCtx, a_uExitCode, a_uExitInfo1, a_uExitInfo2) \
5281	do { \
5282	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL); \
5283	return VBOXSTRICTRC_TODO(IEMExecSvmVmexit((a_pVCpu), (a_uExitCode), (a_uExitInfo1), (a_uExitInfo2))); \
5284	} while (0)
5285
5286	/*
5287	* For all the #VMEXITs here we primarily figure out if the #VMEXIT is expected by the
5288	* nested-guest. If it isn't, it should be handled by the (outer) guest.
5289	*/
5290	PSVMVMCB pVmcbNstGst = pCtx->hwvirt.svm.CTX_SUFF(pVmcb);
5291	PSVMVMCBCTRL pVmcbNstGstCtrl = &pVmcbNstGst->ctrl;
5292	uint64_t const uExitCode = pVmcbNstGstCtrl->u64ExitCode;
5293	uint64_t const uExitInfo1 = pVmcbNstGstCtrl->u64ExitInfo1;
5294	uint64_t const uExitInfo2 = pVmcbNstGstCtrl->u64ExitInfo2;
5295
5296	Assert(uExitCode == pVmcbNstGstCtrl->u64ExitCode);
5297	switch (uExitCode)
5298	{
5299	case SVM_EXIT_CPUID:
5300	{
5301	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_CPUID))
5302	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5303	return hmR0SvmExitCpuid(pVCpu, pCtx, pSvmTransient);
5304	}
5305
5306	case SVM_EXIT_RDTSC:
5307	{
5308	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_RDTSC))
5309	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5310	return hmR0SvmExitRdtsc(pVCpu, pCtx, pSvmTransient);
5311	}
5312
5313	case SVM_EXIT_RDTSCP:
5314	{
5315	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_RDTSCP))
5316	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5317	return hmR0SvmExitRdtscp(pVCpu, pCtx, pSvmTransient);
5318	}
5319
5320	case SVM_EXIT_MONITOR:
5321	{
5322	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_MONITOR))
5323	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5324	return hmR0SvmExitMonitor(pVCpu, pCtx, pSvmTransient);
5325	}
5326
5327	case SVM_EXIT_MWAIT:
5328	{
5329	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_MWAIT))
5330	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5331	return hmR0SvmExitMwait(pVCpu, pCtx, pSvmTransient);
5332	}
5333
5334	case SVM_EXIT_HLT:
5335	{
5336	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_HLT))
5337	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5338	return hmR0SvmExitHlt(pVCpu, pCtx, pSvmTransient);
5339	}
5340
5341	case SVM_EXIT_MSR:
5342	{
5343	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_MSR_PROT))
5344	{
5345	uint32_t const idMsr = pCtx->ecx;
5346	uint16_t offMsrpm;
5347	uint8_t uMsrpmBit;
5348	int rc = HMSvmGetMsrpmOffsetAndBit(idMsr, &offMsrpm, &uMsrpmBit);
5349	if (RT_SUCCESS(rc))
5350	{
5351	Assert(uMsrpmBit == 0 \|\| uMsrpmBit == 2 \|\| uMsrpmBit == 4 \|\| uMsrpmBit == 6);
5352	Assert(offMsrpm < SVM_MSRPM_PAGES << X86_PAGE_4K_SHIFT);
5353
5354	uint8_t const pbMsrBitmap = (uint8_t const )pCtx->hwvirt.svm.CTX_SUFF(pvMsrBitmap);
5355	pbMsrBitmap += offMsrpm;
5356	bool const fInterceptRead = RT_BOOL(*pbMsrBitmap & RT_BIT(uMsrpmBit));
5357	bool const fInterceptWrite = RT_BOOL(*pbMsrBitmap & RT_BIT(uMsrpmBit + 1));
5358
5359	if ( (fInterceptWrite && pVmcbNstGstCtrl->u64ExitInfo1 == SVM_EXIT1_MSR_WRITE)
5360	\|\| (fInterceptRead && pVmcbNstGstCtrl->u64ExitInfo1 == SVM_EXIT1_MSR_READ))
5361	{
5362	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5363	}
5364	}
5365	else
5366	{
5367	/*
5368	* MSRs not covered by the MSRPM automatically cause an #VMEXIT.
5369	* See AMD-V spec. "15.11 MSR Intercepts".
5370	*/
5371	Assert(rc == VERR_OUT_OF_RANGE);
5372	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5373	}
5374	}
5375	return hmR0SvmExitMsr(pVCpu, pCtx, pSvmTransient);
5376	}
5377
5378	case SVM_EXIT_IOIO:
5379	{
5380	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_IOIO_PROT))
5381	{
5382	void *pvIoBitmap = pCtx->hwvirt.svm.CTX_SUFF(pvIoBitmap);
5383	SVMIOIOEXITINFO IoExitInfo;
5384	IoExitInfo.u = pVmcbNstGst->ctrl.u64ExitInfo1;
5385	bool const fIntercept = hmR0SvmIsIoInterceptActive(pvIoBitmap, &IoExitInfo);
5386	if (fIntercept)
5387	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5388	}
5389	return hmR0SvmExitIOInstr(pVCpu, pCtx, pSvmTransient);
5390	}
5391
5392	case SVM_EXIT_XCPT_PF:
5393	{
5394	PVM pVM = pVCpu->CTX_SUFF(pVM);
5395	if (pVM->hm.s.fNestedPaging)
5396	{
5397	uint32_t const u32ErrCode = pVmcbNstGstCtrl->u64ExitInfo1;
5398	uint64_t const uFaultAddress = pVmcbNstGstCtrl->u64ExitInfo2;
5399
5400	/* If the nested-guest is intercepting #PFs, cause a #PF #VMEXIT. */
5401	if (HMIsGuestSvmXcptInterceptSet(pVCpu, X86_XCPT_PF))
5402	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, u32ErrCode, uFaultAddress);
5403
5404	/* If the nested-guest is not intercepting #PFs, forward the #PF to the guest. */
5405	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR2);
5406	hmR0SvmSetPendingXcptPF(pVCpu, pCtx, u32ErrCode, uFaultAddress);
5407	return VINF_SUCCESS;
5408	}
5409	return hmR0SvmExitXcptPF(pVCpu, pCtx,pSvmTransient);
5410	}
5411
5412	case SVM_EXIT_XCPT_UD:
5413	{
5414	if (HMIsGuestSvmXcptInterceptSet(pVCpu, X86_XCPT_UD))
5415	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5416	hmR0SvmSetPendingXcptUD(pVCpu);
5417	return VINF_SUCCESS;
5418	}
5419
5420	case SVM_EXIT_XCPT_MF:
5421	{
5422	if (HMIsGuestSvmXcptInterceptSet(pVCpu, X86_XCPT_MF))
5423	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5424	return hmR0SvmExitXcptMF(pVCpu, pCtx, pSvmTransient);
5425	}
5426
5427	case SVM_EXIT_XCPT_DB:
5428	{
5429	if (HMIsGuestSvmXcptInterceptSet(pVCpu, X86_XCPT_DB))
5430	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5431	return hmR0SvmNestedExitXcptDB(pVCpu, pCtx, pSvmTransient);
5432	}
5433
5434	case SVM_EXIT_XCPT_AC:
5435	{
5436	if (HMIsGuestSvmXcptInterceptSet(pVCpu, X86_XCPT_AC))
5437	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5438	return hmR0SvmExitXcptAC(pVCpu, pCtx, pSvmTransient);
5439	}
5440
5441	case SVM_EXIT_XCPT_BP:
5442	{
5443	if (HMIsGuestSvmXcptInterceptSet(pVCpu, X86_XCPT_BP))
5444	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5445	return hmR0SvmNestedExitXcptBP(pVCpu, pCtx, pSvmTransient);
5446	}
5447
5448	case SVM_EXIT_READ_CR0:
5449	case SVM_EXIT_READ_CR3:
5450	case SVM_EXIT_READ_CR4:
5451	{
5452	uint8_t const uCr = uExitCode - SVM_EXIT_READ_CR0;
5453	if (HMIsGuestSvmReadCRxInterceptSet(pVCpu, uCr))
5454	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5455	return hmR0SvmExitReadCRx(pVCpu, pCtx, pSvmTransient);
5456	}
5457
5458	case SVM_EXIT_CR0_SEL_WRITE:
5459	{
5460	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_CR0_SEL_WRITE))
5461	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5462	return hmR0SvmExitWriteCRx(pVCpu, pCtx, pSvmTransient);
5463	}
5464
5465	case SVM_EXIT_WRITE_CR0:
5466	case SVM_EXIT_WRITE_CR3:
5467	case SVM_EXIT_WRITE_CR4:
5468	case SVM_EXIT_WRITE_CR8: /* CR8 writes would go to the V_TPR rather than here, since we run with V_INTR_MASKING. */
5469	{
5470	uint8_t const uCr = uExitCode - SVM_EXIT_WRITE_CR0;
5471	Log4Func(("Write CR%u: uExitInfo1=%#RX64 uExitInfo2=%#RX64\n", uCr, uExitInfo1, uExitInfo2));
5472
5473	if (HMIsGuestSvmWriteCRxInterceptSet(pVCpu, uCr))
5474	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5475	return hmR0SvmExitWriteCRx(pVCpu, pCtx, pSvmTransient);
5476	}
5477
5478	case SVM_EXIT_PAUSE:
5479	{
5480	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_PAUSE))
5481	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5482	return hmR0SvmExitPause(pVCpu, pCtx, pSvmTransient);
5483	}
5484
5485	case SVM_EXIT_VINTR:
5486	{
5487	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_VINTR))
5488	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5489	return hmR0SvmExitUnexpected(pVCpu, pCtx, pSvmTransient);
5490	}
5491
5492	case SVM_EXIT_INTR:
5493	case SVM_EXIT_NMI:
5494	case SVM_EXIT_SMI:
5495	case SVM_EXIT_XCPT_NMI: /* Should not occur, SVM_EXIT_NMI is used instead. */
5496	{
5497	/*
5498	* We shouldn't direct physical interrupts, NMIs, SMIs to the nested-guest.
5499	*
5500	* Although we don't intercept SMIs, the nested-guest might. Therefore, we might
5501	* get an SMI #VMEXIT here so simply ignore rather than causing a corresponding
5502	* nested-guest #VMEXIT.
5503	*/
5504	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_SVM_VMEXIT_MASK);
5505	return hmR0SvmExitIntr(pVCpu, pCtx, pSvmTransient);
5506	}
5507
5508	case SVM_EXIT_FERR_FREEZE:
5509	{
5510	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_FERR_FREEZE))
5511	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5512	return hmR0SvmExitFerrFreeze(pVCpu, pCtx, pSvmTransient);
5513	}
5514
5515	case SVM_EXIT_INVLPG:
5516	{
5517	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_INVLPG))
5518	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5519	return hmR0SvmExitInvlpg(pVCpu, pCtx, pSvmTransient);
5520	}
5521
5522	case SVM_EXIT_WBINVD:
5523	{
5524	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_WBINVD))
5525	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5526	return hmR0SvmExitWbinvd(pVCpu, pCtx, pSvmTransient);
5527	}
5528
5529	case SVM_EXIT_INVD:
5530	{
5531	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_INVD))
5532	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5533	return hmR0SvmExitInvd(pVCpu, pCtx, pSvmTransient);
5534	}
5535
5536	case SVM_EXIT_RDPMC:
5537	{
5538	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_RDPMC))
5539	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5540	return hmR0SvmExitRdpmc(pVCpu, pCtx, pSvmTransient);
5541	}
5542
5543	default:
5544	{
5545	switch (uExitCode)
5546	{
5547	case SVM_EXIT_READ_DR0: case SVM_EXIT_READ_DR1: case SVM_EXIT_READ_DR2: case SVM_EXIT_READ_DR3:
5548	case SVM_EXIT_READ_DR6: case SVM_EXIT_READ_DR7: case SVM_EXIT_READ_DR8: case SVM_EXIT_READ_DR9:
5549	case SVM_EXIT_READ_DR10: case SVM_EXIT_READ_DR11: case SVM_EXIT_READ_DR12: case SVM_EXIT_READ_DR13:
5550	case SVM_EXIT_READ_DR14: case SVM_EXIT_READ_DR15:
5551	{
5552	uint8_t const uDr = uExitCode - SVM_EXIT_READ_DR0;
5553	if (HMIsGuestSvmReadDRxInterceptSet(pVCpu, uDr))
5554	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5555	return hmR0SvmExitReadDRx(pVCpu, pCtx, pSvmTransient);
5556	}
5557
5558	case SVM_EXIT_WRITE_DR0: case SVM_EXIT_WRITE_DR1: case SVM_EXIT_WRITE_DR2: case SVM_EXIT_WRITE_DR3:
5559	case SVM_EXIT_WRITE_DR6: case SVM_EXIT_WRITE_DR7: case SVM_EXIT_WRITE_DR8: case SVM_EXIT_WRITE_DR9:
5560	case SVM_EXIT_WRITE_DR10: case SVM_EXIT_WRITE_DR11: case SVM_EXIT_WRITE_DR12: case SVM_EXIT_WRITE_DR13:
5561	case SVM_EXIT_WRITE_DR14: case SVM_EXIT_WRITE_DR15:
5562	{
5563	uint8_t const uDr = uExitCode - SVM_EXIT_WRITE_DR0;
5564	if (HMIsGuestSvmWriteDRxInterceptSet(pVCpu, uDr))
5565	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5566	return hmR0SvmExitWriteDRx(pVCpu, pCtx, pSvmTransient);
5567	}
5568
5569	case SVM_EXIT_XCPT_DE:
5570	/* SVM_EXIT_XCPT_DB: / / Handled above. */
5571	/* SVM_EXIT_XCPT_NMI: / / Handled above. */
5572	/* SVM_EXIT_XCPT_BP: / / Handled above. */
5573	case SVM_EXIT_XCPT_OF:
5574	case SVM_EXIT_XCPT_BR:
5575	/* SVM_EXIT_XCPT_UD: / / Handled above. */
5576	case SVM_EXIT_XCPT_NM:
5577	case SVM_EXIT_XCPT_DF:
5578	case SVM_EXIT_XCPT_CO_SEG_OVERRUN:
5579	case SVM_EXIT_XCPT_TS:
5580	case SVM_EXIT_XCPT_NP:
5581	case SVM_EXIT_XCPT_SS:
5582	case SVM_EXIT_XCPT_GP:
5583	/* SVM_EXIT_XCPT_PF: / / Handled above. */
5584	case SVM_EXIT_XCPT_15: /* Reserved. */
5585	/* SVM_EXIT_XCPT_MF: / / Handled above. */
5586	/* SVM_EXIT_XCPT_AC: / / Handled above. */
5587	case SVM_EXIT_XCPT_MC:
5588	case SVM_EXIT_XCPT_XF:
5589	case SVM_EXIT_XCPT_20: case SVM_EXIT_XCPT_21: case SVM_EXIT_XCPT_22: case SVM_EXIT_XCPT_23:
5590	case SVM_EXIT_XCPT_24: case SVM_EXIT_XCPT_25: case SVM_EXIT_XCPT_26: case SVM_EXIT_XCPT_27:
5591	case SVM_EXIT_XCPT_28: case SVM_EXIT_XCPT_29: case SVM_EXIT_XCPT_30: case SVM_EXIT_XCPT_31:
5592	{
5593	uint8_t const uVector = uExitCode - SVM_EXIT_XCPT_0;
5594	if (HMIsGuestSvmXcptInterceptSet(pVCpu, uVector))
5595	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5596	return hmR0SvmExitXcptGeneric(pVCpu, pCtx, pSvmTransient);
5597	}
5598
5599	case SVM_EXIT_XSETBV:
5600	{
5601	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_XSETBV))
5602	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5603	return hmR0SvmExitXsetbv(pVCpu, pCtx, pSvmTransient);
5604	}
5605
5606	case SVM_EXIT_TASK_SWITCH:
5607	{
5608	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_TASK_SWITCH))
5609	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5610	return hmR0SvmExitTaskSwitch(pVCpu, pCtx, pSvmTransient);
5611	}
5612
5613	case SVM_EXIT_IRET:
5614	{
5615	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_IRET))
5616	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5617	return hmR0SvmExitIret(pVCpu, pCtx, pSvmTransient);
5618	}
5619
5620	case SVM_EXIT_SHUTDOWN:
5621	{
5622	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_SHUTDOWN))
5623	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5624	return hmR0SvmExitShutdown(pVCpu, pCtx, pSvmTransient);
5625	}
5626
5627	case SVM_EXIT_VMMCALL:
5628	{
5629	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_VMMCALL))
5630	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5631	return hmR0SvmExitVmmCall(pVCpu, pCtx, pSvmTransient);
5632	}
5633
5634	case SVM_EXIT_CLGI:
5635	{
5636	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_CLGI))
5637	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5638	return hmR0SvmExitClgi(pVCpu, pCtx, pSvmTransient);
5639	}
5640
5641	case SVM_EXIT_STGI:
5642	{
5643	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_STGI))
5644	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5645	return hmR0SvmExitStgi(pVCpu, pCtx, pSvmTransient);
5646	}
5647
5648	case SVM_EXIT_VMLOAD:
5649	{
5650	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_VMLOAD))
5651	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5652	return hmR0SvmExitVmload(pVCpu, pCtx, pSvmTransient);
5653	}
5654
5655	case SVM_EXIT_VMSAVE:
5656	{
5657	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_VMSAVE))
5658	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5659	return hmR0SvmExitVmsave(pVCpu, pCtx, pSvmTransient);
5660	}
5661
5662	case SVM_EXIT_INVLPGA:
5663	{
5664	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_INVLPGA))
5665	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5666	return hmR0SvmExitInvlpga(pVCpu, pCtx, pSvmTransient);
5667	}
5668
5669	case SVM_EXIT_VMRUN:
5670	{
5671	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_VMRUN))
5672	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5673	return hmR0SvmExitVmrun(pVCpu, pCtx, pSvmTransient);
5674	}
5675
5676	case SVM_EXIT_RSM:
5677	{
5678	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_RSM))
5679	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5680	hmR0SvmSetPendingXcptUD(pVCpu);
5681	return VINF_SUCCESS;
5682	}
5683
5684	case SVM_EXIT_SKINIT:
5685	{
5686	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_SKINIT))
5687	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5688	hmR0SvmSetPendingXcptUD(pVCpu);
5689	return VINF_SUCCESS;
5690	}
5691
5692	case SVM_EXIT_NPF:
5693	{
5694	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
5695	return hmR0SvmExitNestedPF(pVCpu, pCtx, pSvmTransient);
5696	}
5697
5698	case SVM_EXIT_INIT: /* We shouldn't get INIT signals while executing a nested-guest. */
5699	return hmR0SvmExitUnexpected(pVCpu, pCtx, pSvmTransient);
5700
5701	default:
5702	{
5703	AssertMsgFailed(("hmR0SvmHandleExitNested: Unknown exit code %#x\n", pSvmTransient->u64ExitCode));
5704	pVCpu->hm.s.u32HMError = pSvmTransient->u64ExitCode;
5705	return VERR_SVM_UNKNOWN_EXIT;
5706	}
5707	}
5708	}
5709	}
5710	/* not reached */
5711
5712	#undef NST_GST_VMEXIT_CALL_RET
5713	}
5714	#endif
5715
5716
5717	/**
5718	* Handles a guest \#VMEXIT (for all EXITCODE values except SVM_EXIT_INVALID).
5719	*
5720	* @returns VBox status code (informational status codes included).
5721	* @param pVCpu The cross context virtual CPU structure.
5722	* @param pCtx Pointer to the guest-CPU context.
5723	* @param pSvmTransient Pointer to the SVM transient structure.
5724	*/
5725	static int hmR0SvmHandleExit(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5726	{
5727	Assert(pSvmTransient->u64ExitCode != SVM_EXIT_INVALID);
5728	Assert(pSvmTransient->u64ExitCode <= SVM_EXIT_MAX);
5729
5730	#ifdef DEBUG_ramshankar
5731	# define VMEXIT_CALL_RET(a_fDbg, a_CallExpr) \
5732	do { \
5733	if ((a_fDbg) == 1) \
5734	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL); \
5735	int rc = a_CallExpr; \
5736	if ((a_fDbg) == 1) \
5737	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST); \
5738	return rc; \
5739	} while (0)
5740	#else
5741	# define VMEXIT_CALL_RET(a_fDbg, a_CallExpr) return a_CallExpr
5742	#endif
5743
5744	/*
5745	* The ordering of the case labels is based on most-frequently-occurring #VMEXITs
5746	* for most guests under normal workloads (for some definition of "normal").
5747	*/
5748	uint64_t const uExitCode = pSvmTransient->u64ExitCode;
5749	switch (uExitCode)
5750	{
5751	case SVM_EXIT_NPF: VMEXIT_CALL_RET(0, hmR0SvmExitNestedPF(pVCpu, pCtx, pSvmTransient));
5752	case SVM_EXIT_IOIO: VMEXIT_CALL_RET(0, hmR0SvmExitIOInstr(pVCpu, pCtx, pSvmTransient));
5753	case SVM_EXIT_RDTSC: VMEXIT_CALL_RET(0, hmR0SvmExitRdtsc(pVCpu, pCtx, pSvmTransient));
5754	case SVM_EXIT_RDTSCP: VMEXIT_CALL_RET(0, hmR0SvmExitRdtscp(pVCpu, pCtx, pSvmTransient));
5755	case SVM_EXIT_CPUID: VMEXIT_CALL_RET(0, hmR0SvmExitCpuid(pVCpu, pCtx, pSvmTransient));
5756	case SVM_EXIT_XCPT_PF: VMEXIT_CALL_RET(0, hmR0SvmExitXcptPF(pVCpu, pCtx, pSvmTransient));
5757	case SVM_EXIT_MSR: VMEXIT_CALL_RET(0, hmR0SvmExitMsr(pVCpu, pCtx, pSvmTransient));
5758	case SVM_EXIT_MONITOR: VMEXIT_CALL_RET(0, hmR0SvmExitMonitor(pVCpu, pCtx, pSvmTransient));
5759	case SVM_EXIT_MWAIT: VMEXIT_CALL_RET(0, hmR0SvmExitMwait(pVCpu, pCtx, pSvmTransient));
5760	case SVM_EXIT_HLT: VMEXIT_CALL_RET(0, hmR0SvmExitHlt(pVCpu, pCtx, pSvmTransient));
5761
5762	case SVM_EXIT_XCPT_NMI: /* Should not occur, SVM_EXIT_NMI is used instead. */
5763	case SVM_EXIT_INTR:
5764	case SVM_EXIT_NMI: VMEXIT_CALL_RET(0, hmR0SvmExitIntr(pVCpu, pCtx, pSvmTransient));
5765
5766	case SVM_EXIT_READ_CR0:
5767	case SVM_EXIT_READ_CR3:
5768	case SVM_EXIT_READ_CR4: VMEXIT_CALL_RET(0, hmR0SvmExitReadCRx(pVCpu, pCtx, pSvmTransient));
5769
5770	case SVM_EXIT_CR0_SEL_WRITE:
5771	case SVM_EXIT_WRITE_CR0:
5772	case SVM_EXIT_WRITE_CR3:
5773	case SVM_EXIT_WRITE_CR4:
5774	case SVM_EXIT_WRITE_CR8: VMEXIT_CALL_RET(0, hmR0SvmExitWriteCRx(pVCpu, pCtx, pSvmTransient));
5775
5776	case SVM_EXIT_VINTR: VMEXIT_CALL_RET(0, hmR0SvmExitVIntr(pVCpu, pCtx, pSvmTransient));
5777	case SVM_EXIT_PAUSE: VMEXIT_CALL_RET(0, hmR0SvmExitPause(pVCpu, pCtx, pSvmTransient));
5778	case SVM_EXIT_VMMCALL: VMEXIT_CALL_RET(0, hmR0SvmExitVmmCall(pVCpu, pCtx, pSvmTransient));
5779	case SVM_EXIT_INVLPG: VMEXIT_CALL_RET(0, hmR0SvmExitInvlpg(pVCpu, pCtx, pSvmTransient));
5780	case SVM_EXIT_WBINVD: VMEXIT_CALL_RET(0, hmR0SvmExitWbinvd(pVCpu, pCtx, pSvmTransient));
5781	case SVM_EXIT_INVD: VMEXIT_CALL_RET(0, hmR0SvmExitInvd(pVCpu, pCtx, pSvmTransient));
5782	case SVM_EXIT_RDPMC: VMEXIT_CALL_RET(0, hmR0SvmExitRdpmc(pVCpu, pCtx, pSvmTransient));
5783	case SVM_EXIT_IRET: VMEXIT_CALL_RET(0, hmR0SvmExitIret(pVCpu, pCtx, pSvmTransient));
5784	case SVM_EXIT_XCPT_UD: VMEXIT_CALL_RET(0, hmR0SvmExitXcptUD(pVCpu, pCtx, pSvmTransient));
5785	case SVM_EXIT_XCPT_MF: VMEXIT_CALL_RET(0, hmR0SvmExitXcptMF(pVCpu, pCtx, pSvmTransient));
5786	case SVM_EXIT_XCPT_DB: VMEXIT_CALL_RET(0, hmR0SvmExitXcptDB(pVCpu, pCtx, pSvmTransient));
5787	case SVM_EXIT_XCPT_AC: VMEXIT_CALL_RET(0, hmR0SvmExitXcptAC(pVCpu, pCtx, pSvmTransient));
5788	case SVM_EXIT_XCPT_BP: VMEXIT_CALL_RET(0, hmR0SvmExitXcptBP(pVCpu, pCtx, pSvmTransient));
5789	case SVM_EXIT_XSETBV: VMEXIT_CALL_RET(0, hmR0SvmExitXsetbv(pVCpu, pCtx, pSvmTransient));
5790	case SVM_EXIT_FERR_FREEZE: VMEXIT_CALL_RET(0, hmR0SvmExitFerrFreeze(pVCpu, pCtx, pSvmTransient));
5791
5792	default:
5793	{
5794	switch (pSvmTransient->u64ExitCode)
5795	{
5796	case SVM_EXIT_READ_DR0: case SVM_EXIT_READ_DR1: case SVM_EXIT_READ_DR2: case SVM_EXIT_READ_DR3:
5797	case SVM_EXIT_READ_DR6: case SVM_EXIT_READ_DR7: case SVM_EXIT_READ_DR8: case SVM_EXIT_READ_DR9:
5798	case SVM_EXIT_READ_DR10: case SVM_EXIT_READ_DR11: case SVM_EXIT_READ_DR12: case SVM_EXIT_READ_DR13:
5799	case SVM_EXIT_READ_DR14: case SVM_EXIT_READ_DR15:
5800	VMEXIT_CALL_RET(0, hmR0SvmExitReadDRx(pVCpu, pCtx, pSvmTransient));
5801
5802	case SVM_EXIT_WRITE_DR0: case SVM_EXIT_WRITE_DR1: case SVM_EXIT_WRITE_DR2: case SVM_EXIT_WRITE_DR3:
5803	case SVM_EXIT_WRITE_DR6: case SVM_EXIT_WRITE_DR7: case SVM_EXIT_WRITE_DR8: case SVM_EXIT_WRITE_DR9:
5804	case SVM_EXIT_WRITE_DR10: case SVM_EXIT_WRITE_DR11: case SVM_EXIT_WRITE_DR12: case SVM_EXIT_WRITE_DR13:
5805	case SVM_EXIT_WRITE_DR14: case SVM_EXIT_WRITE_DR15:
5806	VMEXIT_CALL_RET(0, hmR0SvmExitWriteDRx(pVCpu, pCtx, pSvmTransient));
5807
5808	case SVM_EXIT_TASK_SWITCH: VMEXIT_CALL_RET(0, hmR0SvmExitTaskSwitch(pVCpu, pCtx, pSvmTransient));
5809	case SVM_EXIT_SHUTDOWN: VMEXIT_CALL_RET(0, hmR0SvmExitShutdown(pVCpu, pCtx, pSvmTransient));
5810
5811	case SVM_EXIT_SMI:
5812	case SVM_EXIT_INIT:
5813	{
5814	/*
5815	* We don't intercept SMIs. As for INIT signals, it really shouldn't ever happen here.
5816	* If it ever does, we want to know about it so log the exit code and bail.
5817	*/
5818	VMEXIT_CALL_RET(0, hmR0SvmExitUnexpected(pVCpu, pCtx, pSvmTransient));
5819	}
5820
5821	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
5822	case SVM_EXIT_CLGI: VMEXIT_CALL_RET(0, hmR0SvmExitClgi(pVCpu, pCtx, pSvmTransient));
5823	case SVM_EXIT_STGI: VMEXIT_CALL_RET(0, hmR0SvmExitStgi(pVCpu, pCtx, pSvmTransient));
5824	case SVM_EXIT_VMLOAD: VMEXIT_CALL_RET(0, hmR0SvmExitVmload(pVCpu, pCtx, pSvmTransient));
5825	case SVM_EXIT_VMSAVE: VMEXIT_CALL_RET(0, hmR0SvmExitVmsave(pVCpu, pCtx, pSvmTransient));
5826	case SVM_EXIT_INVLPGA: VMEXIT_CALL_RET(0, hmR0SvmExitInvlpga(pVCpu, pCtx, pSvmTransient));
5827	case SVM_EXIT_VMRUN: VMEXIT_CALL_RET(0, hmR0SvmExitVmrun(pVCpu, pCtx, pSvmTransient));
5828	#else
5829	case SVM_EXIT_CLGI:
5830	case SVM_EXIT_STGI:
5831	case SVM_EXIT_VMLOAD:
5832	case SVM_EXIT_VMSAVE:
5833	case SVM_EXIT_INVLPGA:
5834	case SVM_EXIT_VMRUN:
5835	#endif
5836	case SVM_EXIT_RSM:
5837	case SVM_EXIT_SKINIT:
5838	{
5839	hmR0SvmSetPendingXcptUD(pVCpu);
5840	return VINF_SUCCESS;
5841	}
5842
5843	#ifdef HMSVM_ALWAYS_TRAP_ALL_XCPTS
5844	case SVM_EXIT_XCPT_DE:
5845	/* SVM_EXIT_XCPT_DB: / / Handled above. */
5846	/* SVM_EXIT_XCPT_NMI: / / Handled above. */
5847	/* SVM_EXIT_XCPT_BP: / / Handled above. */
5848	case SVM_EXIT_XCPT_OF:
5849	case SVM_EXIT_XCPT_BR:
5850	/* SVM_EXIT_XCPT_UD: / / Handled above. */
5851	case SVM_EXIT_XCPT_NM:
5852	case SVM_EXIT_XCPT_DF:
5853	case SVM_EXIT_XCPT_CO_SEG_OVERRUN:
5854	case SVM_EXIT_XCPT_TS:
5855	case SVM_EXIT_XCPT_NP:
5856	case SVM_EXIT_XCPT_SS:
5857	case SVM_EXIT_XCPT_GP:
5858	/* SVM_EXIT_XCPT_PF: */
5859	case SVM_EXIT_XCPT_15: /* Reserved. */
5860	/* SVM_EXIT_XCPT_MF: / / Handled above. */
5861	/* SVM_EXIT_XCPT_AC: / / Handled above. */
5862	case SVM_EXIT_XCPT_MC:
5863	case SVM_EXIT_XCPT_XF:
5864	case SVM_EXIT_XCPT_20: case SVM_EXIT_XCPT_21: case SVM_EXIT_XCPT_22: case SVM_EXIT_XCPT_23:
5865	case SVM_EXIT_XCPT_24: case SVM_EXIT_XCPT_25: case SVM_EXIT_XCPT_26: case SVM_EXIT_XCPT_27:
5866	case SVM_EXIT_XCPT_28: case SVM_EXIT_XCPT_29: case SVM_EXIT_XCPT_30: case SVM_EXIT_XCPT_31:
5867	VMEXIT_CALL_RET(0, hmR0SvmExitXcptGeneric(pVCpu, pCtx, pSvmTransient));
5868	#endif /* HMSVM_ALWAYS_TRAP_ALL_XCPTS */
5869
5870	default:
5871	{
5872	AssertMsgFailed(("hmR0SvmHandleExit: Unknown exit code %#RX64\n", uExitCode));
5873	pVCpu->hm.s.u32HMError = uExitCode;
5874	return VERR_SVM_UNKNOWN_EXIT;
5875	}
5876	}
5877	}
5878	}
5879	/* not reached */
5880	#undef VMEXIT_CALL_RET
5881	}
5882
5883
5884	#ifdef DEBUG
5885	/* Is there some generic IPRT define for this that are not in Runtime/internal/\* ?? */
5886	# define HMSVM_ASSERT_PREEMPT_CPUID_VAR() \
5887	RTCPUID const idAssertCpu = RTThreadPreemptIsEnabled(NIL_RTTHREAD) ? NIL_RTCPUID : RTMpCpuId()
5888
5889	# define HMSVM_ASSERT_PREEMPT_CPUID() \
5890	do \
5891	{ \
5892	RTCPUID const idAssertCpuNow = RTThreadPreemptIsEnabled(NIL_RTTHREAD) ? NIL_RTCPUID : RTMpCpuId(); \
5893	AssertMsg(idAssertCpu == idAssertCpuNow, ("SVM %#x, %#x\n", idAssertCpu, idAssertCpuNow)); \
5894	} while (0)
5895
5896	# define HMSVM_VALIDATE_EXIT_HANDLER_PARAMS() \
5897	do { \
5898	AssertPtr(pVCpu); \
5899	AssertPtr(pCtx); \
5900	AssertPtr(pSvmTransient); \
5901	Assert(ASMIntAreEnabled()); \
5902	HMSVM_ASSERT_PREEMPT_SAFE(); \
5903	HMSVM_ASSERT_PREEMPT_CPUID_VAR(); \
5904	Log4Func(("vcpu[%u] -v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-\n", (uint32_t)pVCpu->idCpu)); \
5905	HMSVM_ASSERT_PREEMPT_SAFE(); \
5906	if (VMMR0IsLogFlushDisabled(pVCpu)) \
5907	HMSVM_ASSERT_PREEMPT_CPUID(); \
5908	} while (0)
5909	#else /* Release builds */
5910	# define HMSVM_VALIDATE_EXIT_HANDLER_PARAMS() do { NOREF(pVCpu); NOREF(pCtx); NOREF(pSvmTransient); } while (0)
5911	#endif
5912
5913
5914	/**
5915	* Worker for hmR0SvmInterpretInvlpg().
5916	*
5917	* @return VBox status code.
5918	* @param pVCpu The cross context virtual CPU structure.
5919	* @param pCpu Pointer to the disassembler state.
5920	* @param pCtx The guest CPU context.
5921	*/
5922	static int hmR0SvmInterpretInvlPgEx(PVMCPU pVCpu, PDISCPUSTATE pCpu, PCPUMCTX pCtx)
5923	{
5924	DISQPVPARAMVAL Param1;
5925	RTGCPTR GCPtrPage;
5926
5927	int rc = DISQueryParamVal(CPUMCTX2CORE(pCtx), pCpu, &pCpu->Param1, &Param1, DISQPVWHICH_SRC);
5928	if (RT_FAILURE(rc))
5929	return VERR_EM_INTERPRETER;
5930
5931	if ( Param1.type == DISQPV_TYPE_IMMEDIATE
5932	\|\| Param1.type == DISQPV_TYPE_ADDRESS)
5933	{
5934	if (!(Param1.flags & (DISQPV_FLAG_32 \| DISQPV_FLAG_64)))
5935	return VERR_EM_INTERPRETER;
5936
5937	GCPtrPage = Param1.val.val64;
5938	VBOXSTRICTRC rc2 = EMInterpretInvlpg(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx), GCPtrPage);
5939	rc = VBOXSTRICTRC_VAL(rc2);
5940	}
5941	else
5942	{
5943	Log4Func(("Invalid parameter type %#x\n", Param1.type));
5944	rc = VERR_EM_INTERPRETER;
5945	}
5946
5947	return rc;
5948	}
5949
5950
5951	/**
5952	* Interprets INVLPG.
5953	*
5954	* @returns VBox status code.
5955	* @retval VINF_* Scheduling instructions.
5956	* @retval VERR_EM_INTERPRETER Something we can't cope with.
5957	* @retval VERR_* Fatal errors.
5958	*
5959	* @param pVM The cross context VM structure.
5960	* @param pVCpu The cross context virtual CPU structure.
5961	* @param pCtx The guest CPU context.
5962	*
5963	* @remarks Updates the RIP if the instruction was executed successfully.
5964	*/
5965	static int hmR0SvmInterpretInvlpg(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
5966	{
5967	/* Only allow 32 & 64 bit code. */
5968	if (CPUMGetGuestCodeBits(pVCpu) != 16)
5969	{
5970	PDISSTATE pDis = &pVCpu->hm.s.DisState;
5971	int rc = EMInterpretDisasCurrent(pVM, pVCpu, pDis, NULL /* pcbInstr */);
5972	if ( RT_SUCCESS(rc)
5973	&& pDis->pCurInstr->uOpcode == OP_INVLPG)
5974	{
5975	rc = hmR0SvmInterpretInvlPgEx(pVCpu, pDis, pCtx);
5976	if (RT_SUCCESS(rc))
5977	pCtx->rip += pDis->cbInstr;
5978	return rc;
5979	}
5980	else
5981	Log4Func(("EMInterpretDisasCurrent failed! rc=%Rrc uOpCode=%#x\n", rc, pDis->pCurInstr->uOpcode));
5982	}
5983	return VERR_EM_INTERPRETER;
5984	}
5985
5986
5987	/**
5988	* Gets the IEM exception flags for the specified SVM event.
5989	*
5990	* @returns The IEM exception flags.
5991	* @param pEvent Pointer to the SVM event.
5992	*
5993	* @remarks This function currently only constructs flags required for
5994	* IEMEvaluateRecursiveXcpt and not the complete flags (e.g. error-code
5995	* and CR2 aspects of an exception are not included).
5996	*/
5997	static uint32_t hmR0SvmGetIemXcptFlags(PCSVMEVENT pEvent)
5998	{
5999	uint8_t const uEventType = pEvent->n.u3Type;
6000	uint32_t fIemXcptFlags;
6001	switch (uEventType)
6002	{
6003	case SVM_EVENT_EXCEPTION:
6004	/*
6005	* Only INT3 and INTO instructions can raise #BP and #OF exceptions.
6006	* See AMD spec. Table 8-1. "Interrupt Vector Source and Cause".
6007	*/
6008	if (pEvent->n.u8Vector == X86_XCPT_BP)
6009	{
6010	fIemXcptFlags = IEM_XCPT_FLAGS_T_SOFT_INT \| IEM_XCPT_FLAGS_BP_INSTR;
6011	break;
6012	}
6013	if (pEvent->n.u8Vector == X86_XCPT_OF)
6014	{
6015	fIemXcptFlags = IEM_XCPT_FLAGS_T_SOFT_INT \| IEM_XCPT_FLAGS_OF_INSTR;
6016	break;
6017	}
6018	/** @todo How do we distinguish ICEBP \#DB from the regular one? */
6019	RT_FALL_THRU();
6020	case SVM_EVENT_NMI:
6021	fIemXcptFlags = IEM_XCPT_FLAGS_T_CPU_XCPT;
6022	break;
6023
6024	case SVM_EVENT_EXTERNAL_IRQ:
6025	fIemXcptFlags = IEM_XCPT_FLAGS_T_EXT_INT;
6026	break;
6027
6028	case SVM_EVENT_SOFTWARE_INT:
6029	fIemXcptFlags = IEM_XCPT_FLAGS_T_SOFT_INT;
6030	break;
6031
6032	default:
6033	fIemXcptFlags = 0;
6034	AssertMsgFailed(("Unexpected event type! uEventType=%#x uVector=%#x", uEventType, pEvent->n.u8Vector));
6035	break;
6036	}
6037	return fIemXcptFlags;
6038	}
6039
6040
6041	/**
6042	* Handle a condition that occurred while delivering an event through the guest
6043	* IDT.
6044	*
6045	* @returns VBox status code (informational error codes included).
6046	* @retval VINF_SUCCESS if we should continue handling the \#VMEXIT.
6047	* @retval VINF_HM_DOUBLE_FAULT if a \#DF condition was detected and we ought to
6048	* continue execution of the guest which will delivery the \#DF.
6049	* @retval VINF_EM_RESET if we detected a triple-fault condition.
6050	* @retval VERR_EM_GUEST_CPU_HANG if we detected a guest CPU hang.
6051	*
6052	* @param pVCpu The cross context virtual CPU structure.
6053	* @param pCtx Pointer to the guest-CPU context.
6054	* @param pSvmTransient Pointer to the SVM transient structure.
6055	*
6056	* @remarks No-long-jump zone!!!
6057	*/
6058	static int hmR0SvmCheckExitDueToEventDelivery(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6059	{
6060	int rc = VINF_SUCCESS;
6061	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
6062	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR2);
6063
6064	Log4(("EXITINTINFO: Pending vectoring event %#RX64 Valid=%RTbool ErrValid=%RTbool Err=%#RX32 Type=%u Vector=%u\n",
6065	pVmcb->ctrl.ExitIntInfo.u, !!pVmcb->ctrl.ExitIntInfo.n.u1Valid, !!pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid,
6066	pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode, pVmcb->ctrl.ExitIntInfo.n.u3Type, pVmcb->ctrl.ExitIntInfo.n.u8Vector));
6067
6068	/*
6069	* The EXITINTINFO (if valid) contains the prior exception (IDT vector) that was trying to
6070	* be delivered to the guest which caused a #VMEXIT which was intercepted (Exit vector).
6071	*
6072	* See AMD spec. 15.7.3 "EXITINFO Pseudo-Code".
6073	*/
6074	if (pVmcb->ctrl.ExitIntInfo.n.u1Valid)
6075	{
6076	IEMXCPTRAISE enmRaise;
6077	IEMXCPTRAISEINFO fRaiseInfo;
6078	bool const fExitIsHwXcpt = pSvmTransient->u64ExitCode - SVM_EXIT_XCPT_0 <= SVM_EXIT_XCPT_31;
6079	uint8_t const uIdtVector = pVmcb->ctrl.ExitIntInfo.n.u8Vector;
6080	if (fExitIsHwXcpt)
6081	{
6082	uint8_t const uExitVector = pSvmTransient->u64ExitCode - SVM_EXIT_XCPT_0;
6083	uint32_t const fIdtVectorFlags = hmR0SvmGetIemXcptFlags(&pVmcb->ctrl.ExitIntInfo);
6084	uint32_t const fExitVectorFlags = IEM_XCPT_FLAGS_T_CPU_XCPT;
6085	enmRaise = IEMEvaluateRecursiveXcpt(pVCpu, fIdtVectorFlags, uIdtVector, fExitVectorFlags, uExitVector, &fRaiseInfo);
6086	}
6087	else
6088	{
6089	/*
6090	* If delivery of an event caused a #VMEXIT that is not an exception (e.g. #NPF)
6091	* then we end up here.
6092	*
6093	* If the event was:
6094	* - a software interrupt, we can re-execute the instruction which will
6095	* regenerate the event.
6096	* - an NMI, we need to clear NMI blocking and re-inject the NMI.
6097	* - a hardware exception or external interrupt, we re-inject it.
6098	*/
6099	fRaiseInfo = IEMXCPTRAISEINFO_NONE;
6100	if (pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_SOFTWARE_INT)
6101	enmRaise = IEMXCPTRAISE_REEXEC_INSTR;
6102	else
6103	enmRaise = IEMXCPTRAISE_PREV_EVENT;
6104	}
6105
6106	switch (enmRaise)
6107	{
6108	case IEMXCPTRAISE_CURRENT_XCPT:
6109	case IEMXCPTRAISE_PREV_EVENT:
6110	{
6111	/* For software interrupts, we shall re-execute the instruction. */
6112	if (!(fRaiseInfo & IEMXCPTRAISEINFO_SOFT_INT_XCPT))
6113	{
6114	RTGCUINTPTR GCPtrFaultAddress = 0;
6115
6116	/* If we are re-injecting an NMI, clear NMI blocking. */
6117	if (pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_NMI)
6118	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_BLOCK_NMIS);
6119
6120	/* Determine a vectoring #PF condition, see comment in hmR0SvmExitXcptPF(). */
6121	if (fRaiseInfo & (IEMXCPTRAISEINFO_EXT_INT_PF \| IEMXCPTRAISEINFO_NMI_PF))
6122	{
6123	pSvmTransient->fVectoringPF = true;
6124	Log4Func(("IDT: Pending vectoring #PF due to delivery of Ext-Int/NMI. uCR2=%#RX64\n", pCtx->cr2));
6125	}
6126	else if ( pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_EXCEPTION
6127	&& uIdtVector == X86_XCPT_PF)
6128	{
6129	/*
6130	* If the previous exception was a #PF, we need to recover the CR2 value.
6131	* This can't happen with shadow paging.
6132	*/
6133	GCPtrFaultAddress = pCtx->cr2;
6134	}
6135
6136	/*
6137	* Without nested paging, when uExitVector is #PF, CR2 value will be updated from the VMCB's
6138	* exit info. fields, if it's a guest #PF, see hmR0SvmExitXcptPF().
6139	*/
6140	Assert(pVmcb->ctrl.ExitIntInfo.n.u3Type != SVM_EVENT_SOFTWARE_INT);
6141	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectPendingReflect);
6142	hmR0SvmSetPendingEvent(pVCpu, &pVmcb->ctrl.ExitIntInfo, GCPtrFaultAddress);
6143
6144	Log4Func(("IDT: Pending vectoring event %#RX64 ErrValid=%RTbool Err=%#RX32 GCPtrFaultAddress=%#RX64\n",
6145	pVmcb->ctrl.ExitIntInfo.u, RT_BOOL(pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid),
6146	pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode, GCPtrFaultAddress));
6147	}
6148	break;
6149	}
6150
6151	case IEMXCPTRAISE_REEXEC_INSTR:
6152	{
6153	Assert(rc == VINF_SUCCESS);
6154	break;
6155	}
6156
6157	case IEMXCPTRAISE_DOUBLE_FAULT:
6158	{
6159	/*
6160	* Determing a vectoring double #PF condition. Used later, when PGM evaluates
6161	* the second #PF as a guest #PF (and not a shadow #PF) and needs to be
6162	* converted into a #DF.
6163	*/
6164	if (fRaiseInfo & IEMXCPTRAISEINFO_PF_PF)
6165	{
6166	Log4Func(("IDT: Pending vectoring double #PF uCR2=%#RX64\n", pCtx->cr2));
6167	pSvmTransient->fVectoringDoublePF = true;
6168	Assert(rc == VINF_SUCCESS);
6169	}
6170	else
6171	{
6172	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectPendingReflect);
6173	hmR0SvmSetPendingXcptDF(pVCpu);
6174	rc = VINF_HM_DOUBLE_FAULT;
6175	}
6176	break;
6177	}
6178
6179	case IEMXCPTRAISE_TRIPLE_FAULT:
6180	{
6181	rc = VINF_EM_RESET;
6182	break;
6183	}
6184
6185	case IEMXCPTRAISE_CPU_HANG:
6186	{
6187	rc = VERR_EM_GUEST_CPU_HANG;
6188	break;
6189	}
6190
6191	default:
6192	{
6193	AssertMsgFailed(("hmR0SvmExitCpuid: EMInterpretCpuId failed with %Rrc\n", rc));
6194	rc = VERR_SVM_IPE_2;
6195	break;
6196	}
6197	}
6198	}
6199	Assert(rc == VINF_SUCCESS \|\| rc == VINF_HM_DOUBLE_FAULT \|\| rc == VINF_EM_RESET \|\| rc == VERR_EM_GUEST_CPU_HANG);
6200	NOREF(pCtx);
6201	return rc;
6202	}
6203
6204
6205	/**
6206	* Advances the guest RIP making use of the CPU's NRIP_SAVE feature if
6207	* supported, otherwise advances the RIP by the number of bytes specified in
6208	* @a cb.
6209	*
6210	* @param pVCpu The cross context virtual CPU structure.
6211	* @param pCtx Pointer to the guest-CPU context.
6212	* @param cb RIP increment value in bytes.
6213	*
6214	* @remarks Use this function only from \#VMEXIT's where the NRIP value is valid
6215	* when NRIP_SAVE is supported by the CPU, otherwise use
6216	* hmR0SvmAdvanceRipDumb!
6217	*/
6218	DECLINLINE(void) hmR0SvmAdvanceRipHwAssist(PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t cb)
6219	{
6220	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu, pCtx);
6221	if (fSupportsNextRipSave)
6222	{
6223	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
6224	Assert(pVmcb);
6225	Assert(pVmcb->ctrl.u64NextRIP);
6226	Assert(!(pCtx->fExtrn & CPUMCTX_EXTRN_RIP));
6227	AssertRelease(pVmcb->ctrl.u64NextRIP - pCtx->rip == cb); /* temporary, remove later */
6228	pCtx->rip = pVmcb->ctrl.u64NextRIP;
6229	}
6230	else
6231	pCtx->rip += cb;
6232
6233	HMSVM_UPDATE_INTR_SHADOW(pVCpu, pCtx);
6234	}
6235
6236
6237	/**
6238	* Gets the length of the current instruction if the CPU supports the NRIP_SAVE
6239	* feature. Otherwise, returns the value in @a cbLikely.
6240	*
6241	* @param pVCpu The cross context virtual CPU structure.
6242	* @param pCtx Pointer to the guest-CPU context.
6243	* @param cbLikely The likely instruction length.
6244	*/
6245	DECLINLINE(uint8_t) hmR0SvmGetInstrLengthHwAssist(PVMCPU pVCpu, PCPUMCTX pCtx, uint8_t cbLikely)
6246	{
6247	Assert(cbLikely <= 15); /* See Intel spec. 2.3.11 "AVX Instruction Length" */
6248	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu, pCtx);
6249	if (fSupportsNextRipSave)
6250	{
6251	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
6252	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pCtx->rip;
6253	Assert(cbInstr == cbLikely);
6254	return cbInstr;
6255	}
6256	return cbLikely;
6257	}
6258
6259
6260	/**
6261	* Advances the guest RIP by the number of bytes specified in @a cb. This does
6262	* not make use of any hardware features to determine the instruction length.
6263	*
6264	* @param pVCpu The cross context virtual CPU structure.
6265	* @param pCtx Pointer to the guest-CPU context.
6266	* @param cb RIP increment value in bytes.
6267	*/
6268	DECLINLINE(void) hmR0SvmAdvanceRipDumb(PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t cb)
6269	{
6270	pCtx->rip += cb;
6271	HMSVM_UPDATE_INTR_SHADOW(pVCpu, pCtx);
6272	}
6273	#undef HMSVM_UPDATE_INTR_SHADOW
6274
6275
6276	/* -=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= */
6277	/* -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- #VMEXIT handlers -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- */
6278	/* -=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= */
6279
6280	/** @name \#VMEXIT handlers.
6281	* @{
6282	*/
6283
6284	/**
6285	* \#VMEXIT handler for external interrupts, NMIs, FPU assertion freeze and INIT
6286	* signals (SVM_EXIT_INTR, SVM_EXIT_NMI, SVM_EXIT_FERR_FREEZE, SVM_EXIT_INIT).
6287	*/
6288	HMSVM_EXIT_DECL hmR0SvmExitIntr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6289	{
6290	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6291
6292	if (pSvmTransient->u64ExitCode == SVM_EXIT_NMI)
6293	STAM_REL_COUNTER_INC(&pVCpu->hm.s.StatExitHostNmiInGC);
6294	else if (pSvmTransient->u64ExitCode == SVM_EXIT_INTR)
6295	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitExtInt);
6296
6297	/*
6298	* AMD-V has no preemption timer and the generic periodic preemption timer has no way to
6299	* signal -before- the timer fires if the current interrupt is our own timer or a some
6300	* other host interrupt. We also cannot examine what interrupt it is until the host
6301	* actually take the interrupt.
6302	*
6303	* Going back to executing guest code here unconditionally causes random scheduling
6304	* problems (observed on an AMD Phenom 9850 Quad-Core on Windows 64-bit host).
6305	*/
6306	return VINF_EM_RAW_INTERRUPT;
6307	}
6308
6309
6310	/**
6311	* \#VMEXIT handler for WBINVD (SVM_EXIT_WBINVD). Conditional \#VMEXIT.
6312	*/
6313	HMSVM_EXIT_DECL hmR0SvmExitWbinvd(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6314	{
6315	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6316
6317	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 2);
6318	int rc = VINF_SUCCESS;
6319	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6320	return rc;
6321	}
6322
6323
6324	/**
6325	* \#VMEXIT handler for INVD (SVM_EXIT_INVD). Unconditional \#VMEXIT.
6326	*/
6327	HMSVM_EXIT_DECL hmR0SvmExitInvd(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6328	{
6329	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6330
6331	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 2);
6332	int rc = VINF_SUCCESS;
6333	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6334	return rc;
6335	}
6336
6337
6338	/**
6339	* \#VMEXIT handler for INVD (SVM_EXIT_CPUID). Conditional \#VMEXIT.
6340	*/
6341	HMSVM_EXIT_DECL hmR0SvmExitCpuid(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6342	{
6343	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6344
6345	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_RIP
6346	\| CPUMCTX_EXTRN_CS);
6347	VBOXSTRICTRC rcStrict;
6348	PCEMEXITREC pExitRec = EMHistoryUpdateFlagsAndTypeAndPC(pVCpu,
6349	EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_CPUID),
6350	pVCpu->cpum.GstCtx.rip + pVCpu->cpum.GstCtx.cs.u64Base);
6351	if (!pExitRec)
6352	{
6353	PVM pVM = pVCpu->CTX_SUFF(pVM);
6354	rcStrict = EMInterpretCpuId(pVM, pVCpu, CPUMCTX2CORE(pCtx));
6355	if (RT_LIKELY(rcStrict == VINF_SUCCESS))
6356	{
6357	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 2);
6358	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6359	}
6360	else
6361	{
6362	AssertMsgFailed(("hmR0SvmExitCpuid: EMInterpretCpuId failed with %Rrc\n", VBOXSTRICTRC_VAL(rcStrict)));
6363	rcStrict = VERR_EM_INTERPRETER;
6364	}
6365	}
6366	else
6367	{
6368	/*
6369	* Frequent exit or something needing probing. Get state and call EMHistoryExec.
6370	*/
6371	Assert(pCtx == &pVCpu->cpum.GstCtx);
6372	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6373
6374	Log4(("CpuIdExit/%u: %04x:%08RX64: %#x/%#x -> EMHistoryExec\n",
6375	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip, pVCpu->cpum.GstCtx.eax, pVCpu->cpum.GstCtx.ecx));
6376
6377	rcStrict = EMHistoryExec(pVCpu, pExitRec, 0);
6378
6379	Log4(("CpuIdExit/%u: %04x:%08RX64: EMHistoryExec -> %Rrc + %04x:%08RX64\n",
6380	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip,
6381	VBOXSTRICTRC_VAL(rcStrict), pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip));
6382	}
6383	return VBOXSTRICTRC_TODO(rcStrict);
6384	}
6385
6386
6387	/**
6388	* \#VMEXIT handler for RDTSC (SVM_EXIT_RDTSC). Conditional \#VMEXIT.
6389	*/
6390	HMSVM_EXIT_DECL hmR0SvmExitRdtsc(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6391	{
6392	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6393	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6394	VBOXSTRICTRC rcStrict = IEMExecDecodedRdtsc(pVCpu, hmR0SvmGetInstrLengthHwAssist(pVCpu, pCtx, 2));
6395	if (rcStrict == VINF_SUCCESS)
6396	pSvmTransient->fUpdateTscOffsetting = true;
6397	else if (rcStrict == VINF_IEM_RAISED_XCPT)
6398	rcStrict = VINF_SUCCESS;
6399	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6400	return VBOXSTRICTRC_TODO(rcStrict);
6401	}
6402
6403
6404	/**
6405	* \#VMEXIT handler for RDTSCP (SVM_EXIT_RDTSCP). Conditional \#VMEXIT.
6406	*/
6407	HMSVM_EXIT_DECL hmR0SvmExitRdtscp(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6408	{
6409	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6410	VBOXSTRICTRC rcStrict = IEMExecDecodedRdtscp(pVCpu, hmR0SvmGetInstrLengthHwAssist(pVCpu, pCtx, 3));
6411	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6412	if (rcStrict == VINF_SUCCESS)
6413	pSvmTransient->fUpdateTscOffsetting = true;
6414	else if (rcStrict == VINF_IEM_RAISED_XCPT)
6415	rcStrict = VINF_SUCCESS;
6416	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6417	return VBOXSTRICTRC_TODO(rcStrict);
6418	}
6419
6420
6421	/**
6422	* \#VMEXIT handler for RDPMC (SVM_EXIT_RDPMC). Conditional \#VMEXIT.
6423	*/
6424	HMSVM_EXIT_DECL hmR0SvmExitRdpmc(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6425	{
6426	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6427	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR0
6428	\| CPUMCTX_EXTRN_CR4
6429	\| CPUMCTX_EXTRN_SS);
6430
6431	int rc = EMInterpretRdpmc(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
6432	if (RT_LIKELY(rc == VINF_SUCCESS))
6433	{
6434	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 2);
6435	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6436	}
6437	else
6438	{
6439	AssertMsgFailed(("hmR0SvmExitRdpmc: EMInterpretRdpmc failed with %Rrc\n", rc));
6440	rc = VERR_EM_INTERPRETER;
6441	}
6442	return rc;
6443	}
6444
6445
6446	/**
6447	* \#VMEXIT handler for INVLPG (SVM_EXIT_INVLPG). Conditional \#VMEXIT.
6448	*/
6449	HMSVM_EXIT_DECL hmR0SvmExitInvlpg(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6450	{
6451	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6452	PVM pVM = pVCpu->CTX_SUFF(pVM);
6453	Assert(!pVM->hm.s.fNestedPaging);
6454
6455	bool const fSupportsDecodeAssists = hmR0SvmSupportsDecodeAssists(pVCpu, pCtx);
6456	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu, pCtx);
6457	if ( fSupportsDecodeAssists
6458	&& fSupportsNextRipSave)
6459	{
6460	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6461	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
6462	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pCtx->rip;
6463	RTGCPTR const GCPtrPage = pVmcb->ctrl.u64ExitInfo1;
6464	VBOXSTRICTRC rcStrict = IEMExecDecodedInvlpg(pVCpu, cbInstr, GCPtrPage);
6465	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6466	return VBOXSTRICTRC_VAL(rcStrict);
6467	}
6468
6469	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6470	int rc = hmR0SvmInterpretInvlpg(pVM, pVCpu, pCtx); /* Updates RIP if successful. */
6471	Assert(rc == VINF_SUCCESS \|\| rc == VERR_EM_INTERPRETER);
6472	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6473	return rc;
6474	}
6475
6476
6477	/**
6478	* \#VMEXIT handler for HLT (SVM_EXIT_HLT). Conditional \#VMEXIT.
6479	*/
6480	HMSVM_EXIT_DECL hmR0SvmExitHlt(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6481	{
6482	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6483
6484	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 1);
6485	int rc = EMShouldContinueAfterHalt(pVCpu, pCtx) ? VINF_SUCCESS : VINF_EM_HALT;
6486	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6487
6488	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitHlt);
6489	if (rc != VINF_SUCCESS)
6490	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHltToR3);
6491	return rc;
6492	}
6493
6494
6495	/**
6496	* \#VMEXIT handler for MONITOR (SVM_EXIT_MONITOR). Conditional \#VMEXIT.
6497	*/
6498	HMSVM_EXIT_DECL hmR0SvmExitMonitor(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6499	{
6500	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6501	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR0
6502	\| CPUMCTX_EXTRN_SS);
6503
6504	int rc = EMInterpretMonitor(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
6505	if (RT_LIKELY(rc == VINF_SUCCESS))
6506	{
6507	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 3);
6508	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6509	}
6510	else
6511	{
6512	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMonitor: EMInterpretMonitor failed with %Rrc\n", rc));
6513	rc = VERR_EM_INTERPRETER;
6514	}
6515	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitMonitor);
6516	return rc;
6517	}
6518
6519
6520	/**
6521	* \#VMEXIT handler for MWAIT (SVM_EXIT_MWAIT). Conditional \#VMEXIT.
6522	*/
6523	HMSVM_EXIT_DECL hmR0SvmExitMwait(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6524	{
6525	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6526	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR0
6527	\| CPUMCTX_EXTRN_SS);
6528
6529	VBOXSTRICTRC rc2 = EMInterpretMWait(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
6530	int rc = VBOXSTRICTRC_VAL(rc2);
6531	if ( rc == VINF_EM_HALT
6532	\|\| rc == VINF_SUCCESS)
6533	{
6534	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 3);
6535
6536	if ( rc == VINF_EM_HALT
6537	&& EMMonitorWaitShouldContinue(pVCpu, pCtx))
6538	{
6539	rc = VINF_SUCCESS;
6540	}
6541	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6542	}
6543	else
6544	{
6545	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMwait: EMInterpretMWait failed with %Rrc\n", rc));
6546	rc = VERR_EM_INTERPRETER;
6547	}
6548	AssertMsg(rc == VINF_SUCCESS \|\| rc == VINF_EM_HALT \|\| rc == VERR_EM_INTERPRETER,
6549	("hmR0SvmExitMwait: EMInterpretMWait failed rc=%Rrc\n", rc));
6550	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitMwait);
6551	return rc;
6552	}
6553
6554
6555	/**
6556	* \#VMEXIT handler for shutdown (triple-fault) (SVM_EXIT_SHUTDOWN). Conditional
6557	* \#VMEXIT.
6558	*/
6559	HMSVM_EXIT_DECL hmR0SvmExitShutdown(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6560	{
6561	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6562	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6563	return VINF_EM_RESET;
6564	}
6565
6566
6567	/**
6568	* \#VMEXIT handler for unexpected exits. Conditional \#VMEXIT.
6569	*/
6570	HMSVM_EXIT_DECL hmR0SvmExitUnexpected(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6571	{
6572	RT_NOREF(pCtx);
6573	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
6574	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6575	AssertMsgFailed(("hmR0SvmExitUnexpected: ExitCode=%#RX64 uExitInfo1=%#RX64 uExitInfo2=%#RX64\n", pSvmTransient->u64ExitCode,
6576	pVmcb->ctrl.u64ExitInfo1, pVmcb->ctrl.u64ExitInfo2));
6577	RT_NOREF(pVmcb);
6578	pVCpu->hm.s.u32HMError = (uint32_t)pSvmTransient->u64ExitCode;
6579	return VERR_SVM_UNEXPECTED_EXIT;
6580	}
6581
6582
6583	/**
6584	* \#VMEXIT handler for CRx reads (SVM_EXIT_READ_CR*). Conditional \#VMEXIT.
6585	*/
6586	HMSVM_EXIT_DECL hmR0SvmExitReadCRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6587	{
6588	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6589
6590	Log4Func(("CS:RIP=%04x:%#RX64\n", pCtx->cs.Sel, pCtx->rip));
6591	#ifdef VBOX_WITH_STATISTICS
6592	switch (pSvmTransient->u64ExitCode)
6593	{
6594	case SVM_EXIT_READ_CR0: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR0Read); break;
6595	case SVM_EXIT_READ_CR2: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR2Read); break;
6596	case SVM_EXIT_READ_CR3: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR3Read); break;
6597	case SVM_EXIT_READ_CR4: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR4Read); break;
6598	case SVM_EXIT_READ_CR8: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR8Read); break;
6599	}
6600	#endif
6601
6602	bool const fSupportsDecodeAssists = hmR0SvmSupportsDecodeAssists(pVCpu, pCtx);
6603	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu, pCtx);
6604	if ( fSupportsDecodeAssists
6605	&& fSupportsNextRipSave)
6606	{
6607	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
6608	bool const fMovCRx = RT_BOOL(pVmcb->ctrl.u64ExitInfo1 & SVM_EXIT1_MOV_CRX_MASK);
6609	if (fMovCRx)
6610	{
6611	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6612	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pCtx->rip;
6613	uint8_t const iCrReg = pSvmTransient->u64ExitCode - SVM_EXIT_READ_CR0;
6614	uint8_t const iGReg = pVmcb->ctrl.u64ExitInfo1 & SVM_EXIT1_MOV_CRX_GPR_NUMBER;
6615	VBOXSTRICTRC rcStrict = IEMExecDecodedMovCRxRead(pVCpu, cbInstr, iGReg, iCrReg);
6616	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6617	return VBOXSTRICTRC_VAL(rcStrict);
6618	}
6619	/* else: SMSW instruction, fall back below to IEM for this. */
6620	}
6621
6622	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6623	VBOXSTRICTRC rc2 = EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */);
6624	int rc = VBOXSTRICTRC_VAL(rc2);
6625	AssertMsg(rc == VINF_SUCCESS \|\| rc == VERR_EM_INTERPRETER \|\| rc == VINF_PGM_CHANGE_MODE \|\| rc == VINF_PGM_SYNC_CR3,
6626	("hmR0SvmExitReadCRx: EMInterpretInstruction failed rc=%Rrc\n", rc));
6627	Assert((pSvmTransient->u64ExitCode - SVM_EXIT_READ_CR0) <= 15);
6628	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6629	return rc;
6630	}
6631
6632
6633	/**
6634	* \#VMEXIT handler for CRx writes (SVM_EXIT_WRITE_CR*). Conditional \#VMEXIT.
6635	*/
6636	HMSVM_EXIT_DECL hmR0SvmExitWriteCRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6637	{
6638	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6639
6640	uint64_t const uExitCode = pSvmTransient->u64ExitCode;
6641	uint8_t const iCrReg = uExitCode == SVM_EXIT_CR0_SEL_WRITE ? 0 : (pSvmTransient->u64ExitCode - SVM_EXIT_WRITE_CR0);
6642	Assert(iCrReg <= 15);
6643
6644	VBOXSTRICTRC rcStrict = VERR_SVM_IPE_5;
6645	bool fDecodedInstr = false;
6646	bool const fSupportsDecodeAssists = hmR0SvmSupportsDecodeAssists(pVCpu, pCtx);
6647	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu, pCtx);
6648	if ( fSupportsDecodeAssists
6649	&& fSupportsNextRipSave)
6650	{
6651	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
6652	bool const fMovCRx = RT_BOOL(pVmcb->ctrl.u64ExitInfo1 & SVM_EXIT1_MOV_CRX_MASK);
6653	if (fMovCRx)
6654	{
6655	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6656	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pCtx->rip;
6657	uint8_t const iGReg = pVmcb->ctrl.u64ExitInfo1 & SVM_EXIT1_MOV_CRX_GPR_NUMBER;
6658	Log4Func(("Mov CR%u w/ iGReg=%#x\n", iCrReg, iGReg));
6659	rcStrict = IEMExecDecodedMovCRxWrite(pVCpu, cbInstr, iCrReg, iGReg);
6660	fDecodedInstr = true;
6661	}
6662	/* else: LMSW or CLTS instruction, fall back below to IEM for this. */
6663	}
6664
6665	if (!fDecodedInstr)
6666	{
6667	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6668	Log4Func(("iCrReg=%#x\n", iCrReg));
6669	rcStrict = IEMExecOneBypassEx(pVCpu, CPUMCTX2CORE(pCtx), NULL);
6670	if (RT_UNLIKELY( rcStrict == VERR_IEM_ASPECT_NOT_IMPLEMENTED
6671	\|\| rcStrict == VERR_IEM_INSTR_NOT_IMPLEMENTED))
6672	rcStrict = VERR_EM_INTERPRETER;
6673	}
6674
6675	if (rcStrict == VINF_SUCCESS)
6676	{
6677	switch (iCrReg)
6678	{
6679	case 0:
6680	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_CR0);
6681	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR0Write);
6682	break;
6683
6684	case 2:
6685	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_CR2);
6686	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR2Write);
6687	break;
6688
6689	case 3:
6690	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_CR3);
6691	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR3Write);
6692	break;
6693
6694	case 4:
6695	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_CR4);
6696	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR4Write);
6697	break;
6698
6699	case 8:
6700	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
6701	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR8Write);
6702	break;
6703
6704	default:
6705	{
6706	AssertMsgFailed(("hmR0SvmExitWriteCRx: Invalid/Unexpected Write-CRx exit. u64ExitCode=%#RX64 %#x\n",
6707	pSvmTransient->u64ExitCode, iCrReg));
6708	break;
6709	}
6710	}
6711	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6712	}
6713	else
6714	Assert(rcStrict == VERR_EM_INTERPRETER \|\| rcStrict == VINF_PGM_CHANGE_MODE \|\| rcStrict == VINF_PGM_SYNC_CR3);
6715	return VBOXSTRICTRC_TODO(rcStrict);
6716	}
6717
6718
6719	/**
6720	* \#VMEXIT helper for read MSRs, see hmR0SvmExitMsr.
6721	*
6722	* @returns VBox status code.
6723	* @param pVCpu The cross context virtual CPU structure.
6724	* @param pVmcb Pointer to the VM control block.
6725	*/
6726	static int hmR0SvmExitReadMsr(PVMCPU pVCpu, PSVMVMCB pVmcb)
6727	{
6728	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
6729	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR0
6730	\| CPUMCTX_EXTRN_RFLAGS
6731	\| CPUMCTX_EXTRN_SS
6732	\| CPUMCTX_EXTRN_ALL_MSRS);
6733
6734	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdmsr);
6735	Log4Func(("idMsr=%#RX32\n", pCtx->ecx));
6736
6737	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu, pCtx);
6738	if (fSupportsNextRipSave)
6739	{
6740	int rc = EMInterpretRdmsr(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
6741	if (RT_LIKELY(rc == VINF_SUCCESS))
6742	{
6743	pCtx->rip = pVmcb->ctrl.u64NextRIP;
6744	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6745	return VINF_SUCCESS;
6746	}
6747
6748	AssertMsg( rc == VERR_EM_INTERPRETER
6749	\|\| rc == VINF_CPUM_R3_MSR_READ, ("EMInterpretRdmsr failed rc=%Rrc\n", rc));
6750	return rc;
6751	}
6752
6753	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6754	int rc = VBOXSTRICTRC_TODO(EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0));
6755	if (RT_UNLIKELY(rc != VINF_SUCCESS))
6756	{
6757	AssertMsg( rc == VERR_EM_INTERPRETER
6758	\|\| rc == VINF_CPUM_R3_MSR_READ, ("EMInterpretInstruction failed rc=%Rrc\n", rc));
6759	}
6760	/* RIP updated by EMInterpretInstruction(). */
6761	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6762	return rc;
6763	}
6764
6765
6766	/**
6767	* \#VMEXIT helper for write MSRs, see hmR0SvmExitMsr.
6768	*
6769	* @returns VBox status code.
6770	* @param pVCpu The cross context virtual CPU structure.
6771	* @param pVmcb Pointer to the VM control block.
6772	* @param pSvmTransient Pointer to the SVM-transient structure.
6773	*/
6774	static int hmR0SvmExitWriteMsr(PVMCPU pVCpu, PSVMVMCB pVmcb, PSVMTRANSIENT pSvmTransient)
6775	{
6776	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
6777	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR0
6778	\| CPUMCTX_EXTRN_RFLAGS
6779	\| CPUMCTX_EXTRN_SS
6780	\| CPUMCTX_EXTRN_ALL_MSRS);
6781
6782	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitWrmsr);
6783	Log4Func(("idMsr=%#RX32\n", pCtx->ecx));
6784
6785	/*
6786	* Handle TPR patching MSR writes.
6787	* We utilitize the LSTAR MSR for patching.
6788	*/
6789	if ( pVCpu->CTX_SUFF(pVM)->hm.s.fTPRPatchingActive
6790	&& pCtx->ecx == MSR_K8_LSTAR)
6791	{
6792	if ((pCtx->eax & 0xff) != pSvmTransient->u8GuestTpr)
6793	{
6794	/* Our patch code uses LSTAR for TPR caching for 32-bit guests. */
6795	int rc2 = APICSetTpr(pVCpu, pCtx->eax & 0xff);
6796	AssertRC(rc2);
6797	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
6798	}
6799
6800	int rc = VINF_SUCCESS;
6801	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 2);
6802	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6803	return rc;
6804	}
6805
6806	/*
6807	* Handle regular MSR writes.
6808	*/
6809	int rc;
6810	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu, pCtx);
6811	if (fSupportsNextRipSave)
6812	{
6813	rc = EMInterpretWrmsr(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
6814	if (RT_LIKELY(rc == VINF_SUCCESS))
6815	{
6816	pCtx->rip = pVmcb->ctrl.u64NextRIP;
6817	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6818	}
6819	else
6820	AssertMsg( rc == VERR_EM_INTERPRETER
6821	\|\| rc == VINF_CPUM_R3_MSR_WRITE, ("hmR0SvmExitMsr: EMInterpretWrmsr failed rc=%Rrc\n", rc));
6822	}
6823	else
6824	{
6825	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6826	rc = VBOXSTRICTRC_TODO(EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */));
6827	if (RT_LIKELY(rc == VINF_SUCCESS))
6828	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc); /* RIP updated by EMInterpretInstruction(). */
6829	else
6830	AssertMsg( rc == VERR_EM_INTERPRETER
6831	\|\| rc == VINF_CPUM_R3_MSR_WRITE, ("hmR0SvmExitMsr: WrMsr. EMInterpretInstruction failed rc=%Rrc\n", rc));
6832	}
6833
6834	if (rc == VINF_SUCCESS)
6835	{
6836	/* If this is an X2APIC WRMSR access, update the APIC TPR state. */
6837	if ( pCtx->ecx >= MSR_IA32_X2APIC_START
6838	&& pCtx->ecx <= MSR_IA32_X2APIC_END)
6839	{
6840	/*
6841	* We've already saved the APIC related guest-state (TPR) in hmR0SvmPostRunGuest().
6842	* When full APIC register virtualization is implemented we'll have to make sure
6843	* APIC state is saved from the VMCB before EMInterpretWrmsr() changes it.
6844	*/
6845	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
6846	}
6847	else
6848	{
6849	switch (pCtx->ecx)
6850	{
6851	case MSR_IA32_TSC: pSvmTransient->fUpdateTscOffsetting = true; break;
6852	case MSR_K6_EFER: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_EFER_MSR); break;
6853	case MSR_K8_FS_BASE: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_FS); break;
6854	case MSR_K8_GS_BASE: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_GS); break;
6855	case MSR_IA32_SYSENTER_CS: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_SYSENTER_CS_MSR); break;
6856	case MSR_IA32_SYSENTER_EIP: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_SYSENTER_EIP_MSR); break;
6857	case MSR_IA32_SYSENTER_ESP: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_SYSENTER_ESP_MSR); break;
6858	}
6859	}
6860	}
6861
6862	/* RIP has been updated by above after EMInterpretWrmsr() or by EMInterpretInstruction(). */
6863	return rc;
6864	}
6865
6866
6867	/**
6868	* \#VMEXIT handler for MSR read and writes (SVM_EXIT_MSR). Conditional
6869	* \#VMEXIT.
6870	*/
6871	HMSVM_EXIT_DECL hmR0SvmExitMsr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6872	{
6873	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6874
6875	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
6876	if (pVmcb->ctrl.u64ExitInfo1 == SVM_EXIT1_MSR_READ)
6877	return hmR0SvmExitReadMsr(pVCpu, pVmcb);
6878
6879	Assert(pVmcb->ctrl.u64ExitInfo1 == SVM_EXIT1_MSR_WRITE);
6880	return hmR0SvmExitWriteMsr(pVCpu, pVmcb, pSvmTransient);
6881	}
6882
6883
6884	/**
6885	* \#VMEXIT handler for DRx read (SVM_EXIT_READ_DRx). Conditional \#VMEXIT.
6886	*/
6887	HMSVM_EXIT_DECL hmR0SvmExitReadDRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6888	{
6889	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6890	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6891
6892	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitDRxRead);
6893
6894	/** @todo Stepping with nested-guest. */
6895	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
6896	{
6897	/* We should -not- get this #VMEXIT if the guest's debug registers were active. */
6898	if (pSvmTransient->fWasGuestDebugStateActive)
6899	{
6900	AssertMsgFailed(("hmR0SvmExitReadDRx: Unexpected exit %#RX32\n", (uint32_t)pSvmTransient->u64ExitCode));
6901	pVCpu->hm.s.u32HMError = (uint32_t)pSvmTransient->u64ExitCode;
6902	return VERR_SVM_UNEXPECTED_EXIT;
6903	}
6904
6905	/*
6906	* Lazy DR0-3 loading.
6907	*/
6908	if (!pSvmTransient->fWasHyperDebugStateActive)
6909	{
6910	Assert(!DBGFIsStepping(pVCpu)); Assert(!pVCpu->hm.s.fSingleInstruction);
6911	Log5(("hmR0SvmExitReadDRx: Lazy loading guest debug registers\n"));
6912
6913	/* Don't intercept DRx read and writes. */
6914	PSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
6915	pVmcb->ctrl.u16InterceptRdDRx = 0;
6916	pVmcb->ctrl.u16InterceptWrDRx = 0;
6917	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
6918
6919	/* We're playing with the host CPU state here, make sure we don't preempt or longjmp. */
6920	VMMRZCallRing3Disable(pVCpu);
6921	HM_DISABLE_PREEMPT();
6922
6923	/* Save the host & load the guest debug state, restart execution of the MOV DRx instruction. */
6924	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
6925	Assert(CPUMIsGuestDebugStateActive(pVCpu) \|\| HC_ARCH_BITS == 32);
6926
6927	HM_RESTORE_PREEMPT();
6928	VMMRZCallRing3Enable(pVCpu);
6929
6930	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxContextSwitch);
6931	return VINF_SUCCESS;
6932	}
6933	}
6934
6935	/*
6936	* Interpret the read/writing of DRx.
6937	*/
6938	/** @todo Decode assist. */
6939	VBOXSTRICTRC rc = EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */);
6940	Log5(("hmR0SvmExitReadDRx: Emulated DRx access: rc=%Rrc\n", VBOXSTRICTRC_VAL(rc)));
6941	if (RT_LIKELY(rc == VINF_SUCCESS))
6942	{
6943	/* Not necessary for read accesses but whatever doesn't hurt for now, will be fixed with decode assist. */
6944	/** @todo CPUM should set this flag! */
6945	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_DR_MASK);
6946	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6947	}
6948	else
6949	Assert(rc == VERR_EM_INTERPRETER);
6950	return VBOXSTRICTRC_TODO(rc);
6951	}
6952
6953
6954	/**
6955	* \#VMEXIT handler for DRx write (SVM_EXIT_WRITE_DRx). Conditional \#VMEXIT.
6956	*/
6957	HMSVM_EXIT_DECL hmR0SvmExitWriteDRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6958	{
6959	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6960	/* For now it's the same since we interpret the instruction anyway. Will change when using of Decode Assist is implemented. */
6961	int rc = hmR0SvmExitReadDRx(pVCpu, pCtx, pSvmTransient);
6962	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitDRxWrite);
6963	STAM_COUNTER_DEC(&pVCpu->hm.s.StatExitDRxRead);
6964	return rc;
6965	}
6966
6967
6968	/**
6969	* \#VMEXIT handler for XCRx write (SVM_EXIT_XSETBV). Conditional \#VMEXIT.
6970	*/
6971	HMSVM_EXIT_DECL hmR0SvmExitXsetbv(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6972	{
6973	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6974	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6975
6976	/** @todo decode assists... */
6977	VBOXSTRICTRC rcStrict = IEMExecOne(pVCpu);
6978	if (rcStrict == VINF_IEM_RAISED_XCPT)
6979	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_XCPT_RAISED_MASK);
6980
6981	pVCpu->hm.s.fLoadSaveGuestXcr0 = (pCtx->cr4 & X86_CR4_OSXSAVE) && pCtx->aXcr[0] != ASMGetXcr0();
6982	Log4Func(("New XCR0=%#RX64 fLoadSaveGuestXcr0=%d (cr4=%RX64) rcStrict=%Rrc\n", pCtx->aXcr[0], pVCpu->hm.s.fLoadSaveGuestXcr0,
6983	pCtx->cr4, VBOXSTRICTRC_VAL(rcStrict)));
6984
6985	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6986	return VBOXSTRICTRC_TODO(rcStrict);
6987	}
6988
6989
6990	/**
6991	* \#VMEXIT handler for I/O instructions (SVM_EXIT_IOIO). Conditional \#VMEXIT.
6992	*/
6993	HMSVM_EXIT_DECL hmR0SvmExitIOInstr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6994	{
6995	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6996	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK
6997	\| CPUMCTX_EXTRN_SREG_MASK);
6998
6999	/* I/O operation lookup arrays. */
7000	static uint32_t const s_aIOSize[8] = { 0, 1, 2, 0, 4, 0, 0, 0 }; /* Size of the I/O accesses in bytes. */
7001	static uint32_t const s_aIOOpAnd[8] = { 0, 0xff, 0xffff, 0, 0xffffffff, 0, 0, 0 }; /* AND masks for saving
7002	the result (in AL/AX/EAX). */
7003	Log4Func(("CS:RIP=%04x:%#RX64\n", pCtx->cs.Sel, pCtx->rip));
7004
7005	PVM pVM = pVCpu->CTX_SUFF(pVM);
7006	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7007
7008	/* Refer AMD spec. 15.10.2 "IN and OUT Behaviour" and Figure 15-2. "EXITINFO1 for IOIO Intercept" for the format. */
7009	SVMIOIOEXITINFO IoExitInfo;
7010	IoExitInfo.u = (uint32_t)pVmcb->ctrl.u64ExitInfo1;
7011	uint32_t uIOWidth = (IoExitInfo.u >> 4) & 0x7;
7012	uint32_t cbValue = s_aIOSize[uIOWidth];
7013	uint32_t uAndVal = s_aIOOpAnd[uIOWidth];
7014
7015	if (RT_UNLIKELY(!cbValue))
7016	{
7017	AssertMsgFailed(("hmR0SvmExitIOInstr: Invalid IO operation. uIOWidth=%u\n", uIOWidth));
7018	return VERR_EM_INTERPRETER;
7019	}
7020
7021	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_RIP
7022	\| CPUMCTX_EXTRN_CS
7023	\| CPUMCTX_EXTRN_RFLAGS);
7024	VBOXSTRICTRC rcStrict;
7025	PCEMEXITREC pExitRec = NULL;
7026	if ( !pVCpu->hm.s.fSingleInstruction
7027	&& !pVCpu->cpum.GstCtx.eflags.Bits.u1TF)
7028	pExitRec = EMHistoryUpdateFlagsAndTypeAndPC(pVCpu,
7029	!IoExitInfo.n.u1Str
7030	? IoExitInfo.n.u1Type == SVM_IOIO_READ
7031	? EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_IO_PORT_READ)
7032	: EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_IO_PORT_WRITE)
7033	: IoExitInfo.n.u1Type == SVM_IOIO_READ
7034	? EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_IO_PORT_STR_READ)
7035	: EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_IO_PORT_STR_WRITE),
7036	pVCpu->cpum.GstCtx.rip + pVCpu->cpum.GstCtx.cs.u64Base);
7037	if (!pExitRec)
7038	{
7039	bool fUpdateRipAlready = false;
7040	if (IoExitInfo.n.u1Str)
7041	{
7042	/* INS/OUTS - I/O String instruction. */
7043	/** @todo Huh? why can't we use the segment prefix information given by AMD-V
7044	* in EXITINFO1? Investigate once this thing is up and running. */
7045	Log4Func(("CS:RIP=%04x:%08RX64 %#06x/%u %c str\n", pCtx->cs.Sel, pCtx->rip, IoExitInfo.n.u16Port, cbValue,
7046	IoExitInfo.n.u1Type == SVM_IOIO_WRITE ? 'w' : 'r'));
7047	AssertReturn(pCtx->dx == IoExitInfo.n.u16Port, VERR_SVM_IPE_2);
7048	static IEMMODE const s_aenmAddrMode[8] =
7049	{
7050	(IEMMODE)-1, IEMMODE_16BIT, IEMMODE_32BIT, (IEMMODE)-1, IEMMODE_64BIT, (IEMMODE)-1, (IEMMODE)-1, (IEMMODE)-1
7051	};
7052	IEMMODE enmAddrMode = s_aenmAddrMode[(IoExitInfo.u >> 7) & 0x7];
7053	if (enmAddrMode != (IEMMODE)-1)
7054	{
7055	uint64_t cbInstr = pVmcb->ctrl.u64ExitInfo2 - pCtx->rip;
7056	if (cbInstr <= 15 && cbInstr >= 1)
7057	{
7058	Assert(cbInstr >= 1U + IoExitInfo.n.u1Rep);
7059	if (IoExitInfo.n.u1Type == SVM_IOIO_WRITE)
7060	{
7061	/* Don't know exactly how to detect whether u3Seg is valid, currently
7062	only enabling it for Bulldozer and later with NRIP. OS/2 broke on
7063	2384 Opterons when only checking NRIP. */
7064	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu, pCtx);
7065	if ( fSupportsNextRipSave
7066	&& pVM->cpum.ro.GuestFeatures.enmMicroarch >= kCpumMicroarch_AMD_15h_First)
7067	{
7068	AssertMsg(IoExitInfo.n.u3Seg == X86_SREG_DS \|\| cbInstr > 1U + IoExitInfo.n.u1Rep,
7069	("u32Seg=%d cbInstr=%d u1REP=%d", IoExitInfo.n.u3Seg, cbInstr, IoExitInfo.n.u1Rep));
7070	rcStrict = IEMExecStringIoWrite(pVCpu, cbValue, enmAddrMode, IoExitInfo.n.u1Rep, (uint8_t)cbInstr,
7071	IoExitInfo.n.u3Seg, true /fIoChecked/);
7072	}
7073	else if (cbInstr == 1U + IoExitInfo.n.u1Rep)
7074	rcStrict = IEMExecStringIoWrite(pVCpu, cbValue, enmAddrMode, IoExitInfo.n.u1Rep, (uint8_t)cbInstr,
7075	X86_SREG_DS, true /fIoChecked/);
7076	else
7077	rcStrict = IEMExecOne(pVCpu);
7078	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOStringWrite);
7079	}
7080	else
7081	{
7082	AssertMsg(IoExitInfo.n.u3Seg == X86_SREG_ES /=0/, ("%#x\n", IoExitInfo.n.u3Seg));
7083	rcStrict = IEMExecStringIoRead(pVCpu, cbValue, enmAddrMode, IoExitInfo.n.u1Rep, (uint8_t)cbInstr,
7084	true /fIoChecked/);
7085	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOStringRead);
7086	}
7087	}
7088	else
7089	{
7090	AssertMsgFailed(("rip=%RX64 nrip=%#RX64 cbInstr=%#RX64\n", pCtx->rip, pVmcb->ctrl.u64ExitInfo2, cbInstr));
7091	rcStrict = IEMExecOne(pVCpu);
7092	}
7093	}
7094	else
7095	{
7096	AssertMsgFailed(("IoExitInfo=%RX64\n", IoExitInfo.u));
7097	rcStrict = IEMExecOne(pVCpu);
7098	}
7099	fUpdateRipAlready = true;
7100	}
7101	else
7102	{
7103	/* IN/OUT - I/O instruction. */
7104	Assert(!IoExitInfo.n.u1Rep);
7105
7106	if (IoExitInfo.n.u1Type == SVM_IOIO_WRITE)
7107	{
7108	rcStrict = IOMIOPortWrite(pVM, pVCpu, IoExitInfo.n.u16Port, pCtx->eax & uAndVal, cbValue);
7109	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOWrite);
7110	}
7111	else
7112	{
7113	uint32_t u32Val = 0;
7114	rcStrict = IOMIOPortRead(pVM, pVCpu, IoExitInfo.n.u16Port, &u32Val, cbValue);
7115	if (IOM_SUCCESS(rcStrict))
7116	{
7117	/* Save result of I/O IN instr. in AL/AX/EAX. */
7118	/** @todo r=bird: 32-bit op size should clear high bits of rax! */
7119	pCtx->eax = (pCtx->eax & ~uAndVal) \| (u32Val & uAndVal);
7120	}
7121	else if (rcStrict == VINF_IOM_R3_IOPORT_READ)
7122	HMR0SavePendingIOPortRead(pVCpu, pCtx->rip, pVmcb->ctrl.u64ExitInfo2, IoExitInfo.n.u16Port, uAndVal, cbValue);
7123
7124	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIORead);
7125	}
7126	}
7127
7128	if (IOM_SUCCESS(rcStrict))
7129	{
7130	/* AMD-V saves the RIP of the instruction following the IO instruction in EXITINFO2. */
7131	if (!fUpdateRipAlready)
7132	pCtx->rip = pVmcb->ctrl.u64ExitInfo2;
7133
7134	/*
7135	* If any I/O breakpoints are armed, we need to check if one triggered
7136	* and take appropriate action.
7137	* Note that the I/O breakpoint type is undefined if CR4.DE is 0.
7138	*/
7139	/** @todo Optimize away the DBGFBpIsHwIoArmed call by having DBGF tell the
7140	* execution engines about whether hyper BPs and such are pending. */
7141	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_DR7);
7142	uint32_t const uDr7 = pCtx->dr[7];
7143	if (RT_UNLIKELY( ( (uDr7 & X86_DR7_ENABLED_MASK)
7144	&& X86_DR7_ANY_RW_IO(uDr7)
7145	&& (pCtx->cr4 & X86_CR4_DE))
7146	\|\| DBGFBpIsHwIoArmed(pVM)))
7147	{
7148	/* We're playing with the host CPU state here, make sure we don't preempt or longjmp. */
7149	VMMRZCallRing3Disable(pVCpu);
7150	HM_DISABLE_PREEMPT();
7151
7152	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxIoCheck);
7153	CPUMR0DebugStateMaybeSaveGuest(pVCpu, false /fDr6/);
7154
7155	VBOXSTRICTRC rcStrict2 = DBGFBpCheckIo(pVM, pVCpu, pCtx, IoExitInfo.n.u16Port, cbValue);
7156	if (rcStrict2 == VINF_EM_RAW_GUEST_TRAP)
7157	{
7158	/* Raise #DB. */
7159	pVmcb->guest.u64DR6 = pCtx->dr[6];
7160	pVmcb->guest.u64DR7 = pCtx->dr[7];
7161	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
7162	hmR0SvmSetPendingXcptDB(pVCpu);
7163	}
7164	/* rcStrict is VINF_SUCCESS, VINF_IOM_R3_IOPORT_COMMIT_WRITE, or in [VINF_EM_FIRST..VINF_EM_LAST],
7165	however we can ditch VINF_IOM_R3_IOPORT_COMMIT_WRITE as it has VMCPU_FF_IOM as backup. */
7166	else if ( rcStrict2 != VINF_SUCCESS
7167	&& (rcStrict == VINF_SUCCESS \|\| rcStrict2 < rcStrict))
7168	rcStrict = rcStrict2;
7169	AssertCompile(VINF_EM_LAST < VINF_IOM_R3_IOPORT_COMMIT_WRITE);
7170
7171	HM_RESTORE_PREEMPT();
7172	VMMRZCallRing3Enable(pVCpu);
7173	}
7174
7175	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
7176	}
7177
7178	#ifdef VBOX_STRICT
7179	if (rcStrict == VINF_IOM_R3_IOPORT_READ)
7180	Assert(IoExitInfo.n.u1Type == SVM_IOIO_READ);
7181	else if (rcStrict == VINF_IOM_R3_IOPORT_WRITE \|\| rcStrict == VINF_IOM_R3_IOPORT_COMMIT_WRITE)
7182	Assert(IoExitInfo.n.u1Type == SVM_IOIO_WRITE);
7183	else
7184	{
7185	/** @todo r=bird: This is missing a bunch of VINF_EM_FIRST..VINF_EM_LAST
7186	* statuses, that the VMM device and some others may return. See
7187	* IOM_SUCCESS() for guidance. */
7188	AssertMsg( RT_FAILURE(rcStrict)
7189	\|\| rcStrict == VINF_SUCCESS
7190	\|\| rcStrict == VINF_EM_RAW_EMULATE_INSTR
7191	\|\| rcStrict == VINF_EM_DBG_BREAKPOINT
7192	\|\| rcStrict == VINF_EM_RAW_GUEST_TRAP
7193	\|\| rcStrict == VINF_EM_RAW_TO_R3
7194	\|\| rcStrict == VINF_TRPM_XCPT_DISPATCHED
7195	\|\| rcStrict == VINF_EM_TRIPLE_FAULT, ("%Rrc\n", VBOXSTRICTRC_VAL(rcStrict)));
7196	}
7197	#endif
7198	}
7199	else
7200	{
7201	/*
7202	* Frequent exit or something needing probing. Get state and call EMHistoryExec.
7203	*/
7204	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7205	STAM_COUNTER_INC(!IoExitInfo.n.u1Str
7206	? IoExitInfo.n.u1Type == SVM_IOIO_WRITE ? &pVCpu->hm.s.StatExitIOWrite : &pVCpu->hm.s.StatExitIORead
7207	: IoExitInfo.n.u1Type == SVM_IOIO_WRITE ? &pVCpu->hm.s.StatExitIOStringWrite : &pVCpu->hm.s.StatExitIOStringRead);
7208	Log4(("IOExit/%u: %04x:%08RX64: %s%s%s %#x LB %u -> EMHistoryExec\n",
7209	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip, IoExitInfo.n.u1Rep ? "REP " : "",
7210	IoExitInfo.n.u1Type == SVM_IOIO_WRITE ? "OUT" : "IN", IoExitInfo.n.u1Str ? "S" : "", IoExitInfo.n.u16Port, uIOWidth));
7211
7212	rcStrict = EMHistoryExec(pVCpu, pExitRec, 0);
7213	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
7214
7215	Log4(("IOExit/%u: %04x:%08RX64: EMHistoryExec -> %Rrc + %04x:%08RX64\n",
7216	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip,
7217	VBOXSTRICTRC_VAL(rcStrict), pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip));
7218	}
7219	return VBOXSTRICTRC_TODO(rcStrict);
7220	}
7221
7222
7223	/**
7224	* \#VMEXIT handler for Nested Page-faults (SVM_EXIT_NPF). Conditional \#VMEXIT.
7225	*/
7226	HMSVM_EXIT_DECL hmR0SvmExitNestedPF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7227	{
7228	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7229	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7230	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
7231
7232	PVM pVM = pVCpu->CTX_SUFF(pVM);
7233	Assert(pVM->hm.s.fNestedPaging);
7234
7235	/* See AMD spec. 15.25.6 "Nested versus Guest Page Faults, Fault Ordering" for VMCB details for #NPF. */
7236	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7237	RTGCPHYS GCPhysFaultAddr = pVmcb->ctrl.u64ExitInfo2;
7238	uint32_t u32ErrCode = pVmcb->ctrl.u64ExitInfo1; /* Note! High bits in EXITINFO1 may contain additional info and are
7239	thus intentionally not copied into u32ErrCode. */
7240
7241	Log4Func(("#NPF at CS:RIP=%04x:%#RX64 GCPhysFaultAddr=%RGp ErrCode=%#x \n", pCtx->cs.Sel, pCtx->rip, GCPhysFaultAddr,
7242	u32ErrCode));
7243
7244	/*
7245	* TPR patching for 32-bit guests, using the reserved bit in the page tables for MMIO regions.
7246	*/
7247	if ( pVM->hm.s.fTprPatchingAllowed
7248	&& !CPUMIsGuestInSvmNestedHwVirtMode(pCtx)
7249	&& (GCPhysFaultAddr & PAGE_OFFSET_MASK) == XAPIC_OFF_TPR
7250	&& ( !(u32ErrCode & X86_TRAP_PF_P) /* Not present */
7251	\|\| (u32ErrCode & (X86_TRAP_PF_P \| X86_TRAP_PF_RSVD)) == (X86_TRAP_PF_P \| X86_TRAP_PF_RSVD)) /* MMIO page. */
7252	&& !CPUMIsGuestInLongModeEx(pCtx)
7253	&& !CPUMGetGuestCPL(pVCpu)
7254	&& pVM->hm.s.cPatches < RT_ELEMENTS(pVM->hm.s.aPatches))
7255	{
7256	RTGCPHYS GCPhysApicBase = APICGetBaseMsrNoCheck(pVCpu);
7257	GCPhysApicBase &= PAGE_BASE_GC_MASK;
7258
7259	if (GCPhysFaultAddr == GCPhysApicBase + XAPIC_OFF_TPR)
7260	{
7261	/* Only attempt to patch the instruction once. */
7262	PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
7263	if (!pPatch)
7264	return VINF_EM_HM_PATCH_TPR_INSTR;
7265	}
7266	}
7267
7268	/*
7269	* Determine the nested paging mode.
7270	*/
7271	PGMMODE enmNestedPagingMode;
7272	#if HC_ARCH_BITS == 32
7273	if (CPUMIsGuestInLongModeEx(pCtx))
7274	enmNestedPagingMode = PGMMODE_AMD64_NX;
7275	else
7276	#endif
7277	enmNestedPagingMode = PGMGetHostMode(pVM);
7278
7279	/*
7280	* MMIO optimization using the reserved (RSVD) bit in the guest page tables for MMIO pages.
7281	*/
7282	Assert((u32ErrCode & (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P)) != X86_TRAP_PF_RSVD);
7283	if ((u32ErrCode & (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P)) == (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P))
7284	{
7285	/*
7286	* If event delivery causes an MMIO #NPF, go back to instruction emulation as otherwise
7287	* injecting the original pending event would most likely cause the same MMIO #NPF.
7288	*/
7289	if (pVCpu->hm.s.Event.fPending)
7290	return VINF_EM_RAW_INJECT_TRPM_EVENT;
7291
7292	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_RIP
7293	\| CPUMCTX_EXTRN_CS);
7294	VBOXSTRICTRC rcStrict;
7295	PCEMEXITREC pExitRec = EMHistoryUpdateFlagsAndTypeAndPC(pVCpu,
7296	EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_MMIO),
7297	pVCpu->cpum.GstCtx.rip + pVCpu->cpum.GstCtx.cs.u64Base);
7298	if (!pExitRec)
7299	{
7300
7301	rcStrict = PGMR0Trap0eHandlerNPMisconfig(pVM, pVCpu, enmNestedPagingMode, CPUMCTX2CORE(pCtx), GCPhysFaultAddr,
7302	u32ErrCode);
7303
7304	/*
7305	* If we succeed, resume guest execution.
7306	*
7307	* If we fail in interpreting the instruction because we couldn't get the guest
7308	* physical address of the page containing the instruction via the guest's page
7309	* tables (we would invalidate the guest page in the host TLB), resume execution
7310	* which would cause a guest page fault to let the guest handle this weird case.
7311	*
7312	* See @bugref{6043}.
7313	*/
7314	if ( rcStrict == VINF_SUCCESS
7315	\|\| rcStrict == VERR_PAGE_TABLE_NOT_PRESENT
7316	\|\| rcStrict == VERR_PAGE_NOT_PRESENT)
7317	{
7318	/* Successfully handled MMIO operation. */
7319	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
7320	rcStrict = VINF_SUCCESS;
7321	}
7322	}
7323	else
7324	{
7325	/*
7326	* Frequent exit or something needing probing. Get state and call EMHistoryExec.
7327	*/
7328	Assert(pCtx == &pVCpu->cpum.GstCtx);
7329	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7330	Log4(("EptMisscfgExit/%u: %04x:%08RX64: %RGp -> EMHistoryExec\n",
7331	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip, GCPhysFaultAddr));
7332
7333	rcStrict = EMHistoryExec(pVCpu, pExitRec, 0);
7334	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
7335
7336	Log4(("EptMisscfgExit/%u: %04x:%08RX64: EMHistoryExec -> %Rrc + %04x:%08RX64\n",
7337	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip,
7338	VBOXSTRICTRC_VAL(rcStrict), pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip));
7339	}
7340	return VBOXSTRICTRC_TODO(rcStrict);
7341	}
7342
7343	TRPMAssertXcptPF(pVCpu, GCPhysFaultAddr, u32ErrCode);
7344	int rc = PGMR0Trap0eHandlerNestedPaging(pVM, pVCpu, enmNestedPagingMode, u32ErrCode, CPUMCTX2CORE(pCtx), GCPhysFaultAddr);
7345	TRPMResetTrap(pVCpu);
7346
7347	Log4Func(("#NPF: PGMR0Trap0eHandlerNestedPaging returns %Rrc CS:RIP=%04x:%#RX64\n", rc, pCtx->cs.Sel, pCtx->rip));
7348
7349	/*
7350	* Same case as PGMR0Trap0eHandlerNPMisconfig(). See comment above, @bugref{6043}.
7351	*/
7352	if ( rc == VINF_SUCCESS
7353	\|\| rc == VERR_PAGE_TABLE_NOT_PRESENT
7354	\|\| rc == VERR_PAGE_NOT_PRESENT)
7355	{
7356	/* We've successfully synced our shadow page tables. */
7357	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPF);
7358	rc = VINF_SUCCESS;
7359	}
7360
7361	return rc;
7362	}
7363
7364
7365	/**
7366	* \#VMEXIT handler for virtual interrupt (SVM_EXIT_VINTR). Conditional
7367	* \#VMEXIT.
7368	*/
7369	HMSVM_EXIT_DECL hmR0SvmExitVIntr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7370	{
7371	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7372	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(pCtx);
7373
7374	/* Indicate that we no longer need to #VMEXIT when the guest is ready to receive NMIs, it is now ready. */
7375	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7376	hmR0SvmClearIntWindowExiting(pVCpu, pVmcb, pCtx);
7377
7378	/* Deliver the pending interrupt via hmR0SvmEvaluatePendingEvent() and resume guest execution. */
7379	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIntWindow);
7380	return VINF_SUCCESS;
7381	}
7382
7383
7384	/**
7385	* \#VMEXIT handler for task switches (SVM_EXIT_TASK_SWITCH). Conditional
7386	* \#VMEXIT.
7387	*/
7388	HMSVM_EXIT_DECL hmR0SvmExitTaskSwitch(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7389	{
7390	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7391	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
7392
7393	#ifndef HMSVM_ALWAYS_TRAP_TASK_SWITCH
7394	Assert(!pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
7395	#endif
7396
7397	/* Check if this task-switch occurred while delivering an event through the guest IDT. */
7398	if (pVCpu->hm.s.Event.fPending) /* Can happen with exceptions/NMI. See @bugref{8411}. */
7399	{
7400	/*
7401	* AMD-V provides us with the exception which caused the TS; we collect
7402	* the information in the call to hmR0SvmCheckExitDueToEventDelivery().
7403	*/
7404	Log4Func(("TS occurred during event delivery\n"));
7405	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitTaskSwitch);
7406	return VINF_EM_RAW_INJECT_TRPM_EVENT;
7407	}
7408
7409	/** @todo Emulate task switch someday, currently just going back to ring-3 for
7410	* emulation. */
7411	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitTaskSwitch);
7412	return VERR_EM_INTERPRETER;
7413	}
7414
7415
7416	/**
7417	* \#VMEXIT handler for VMMCALL (SVM_EXIT_VMMCALL). Conditional \#VMEXIT.
7418	*/
7419	HMSVM_EXIT_DECL hmR0SvmExitVmmCall(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7420	{
7421	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7422	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7423
7424	if (pVCpu->CTX_SUFF(pVM)->hm.s.fTprPatchingAllowed)
7425	{
7426	int rc = hmSvmEmulateMovTpr(pVCpu, pCtx);
7427	if (rc != VERR_NOT_FOUND)
7428	{
7429	Log4Func(("hmSvmEmulateMovTpr returns %Rrc\n", rc));
7430	return rc;
7431	}
7432	}
7433
7434	if (EMAreHypercallInstructionsEnabled(pVCpu))
7435	{
7436	VBOXSTRICTRC rcStrict = GIMHypercall(pVCpu, pCtx);
7437	if (RT_SUCCESS(rcStrict))
7438	{
7439	/* Only update the RIP if we're continuing guest execution and not in the case
7440	of say VINF_GIM_R3_HYPERCALL. */
7441	if (rcStrict == VINF_SUCCESS)
7442	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 3 /* cbInstr */);
7443
7444	return VBOXSTRICTRC_VAL(rcStrict);
7445	}
7446	else
7447	Log4Func(("GIMHypercall returns %Rrc -> #UD\n", VBOXSTRICTRC_VAL(rcStrict)));
7448	}
7449
7450	hmR0SvmSetPendingXcptUD(pVCpu);
7451	return VINF_SUCCESS;
7452	}
7453
7454
7455	/**
7456	* \#VMEXIT handler for VMMCALL (SVM_EXIT_VMMCALL). Conditional \#VMEXIT.
7457	*/
7458	HMSVM_EXIT_DECL hmR0SvmExitPause(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7459	{
7460	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7461	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 2);
7462	/** @todo The guest has likely hit a contended spinlock. We might want to
7463	* poke a schedule different guest VCPU. */
7464	return VINF_EM_RAW_INTERRUPT;
7465	}
7466
7467
7468	/**
7469	* \#VMEXIT handler for FERR intercept (SVM_EXIT_FERR_FREEZE). Conditional
7470	* \#VMEXIT.
7471	*/
7472	HMSVM_EXIT_DECL hmR0SvmExitFerrFreeze(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7473	{
7474	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7475	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR0);
7476	Assert(!(pCtx->cr0 & X86_CR0_NE));
7477
7478	Log4Func(("Raising IRQ 13 in response to #FERR\n"));
7479	return PDMIsaSetIrq(pVCpu->CTX_SUFF(pVM), 13 /* u8Irq /, 1 / u8Level /, 0 / uTagSrc */);
7480	}
7481
7482
7483	/**
7484	* \#VMEXIT handler for IRET (SVM_EXIT_IRET). Conditional \#VMEXIT.
7485	*/
7486	HMSVM_EXIT_DECL hmR0SvmExitIret(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7487	{
7488	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7489
7490	/* Clear NMI blocking. */
7491	if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_BLOCK_NMIS))
7492	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_BLOCK_NMIS);
7493
7494	/* Indicate that we no longer need to #VMEXIT when the guest is ready to receive NMIs, it is now ready. */
7495	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7496	hmR0SvmClearCtrlIntercept(pVCpu, pCtx, pVmcb, SVM_CTRL_INTERCEPT_IRET);
7497
7498	/* Deliver the pending NMI via hmR0SvmEvaluatePendingEvent() and resume guest execution. */
7499	return VINF_SUCCESS;
7500	}
7501
7502
7503	/**
7504	* \#VMEXIT handler for page-fault exceptions (SVM_EXIT_XCPT_14).
7505	* Conditional \#VMEXIT.
7506	*/
7507	HMSVM_EXIT_DECL hmR0SvmExitXcptPF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7508	{
7509	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7510	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7511	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
7512
7513	/* See AMD spec. 15.12.15 "#PF (Page Fault)". */
7514	PVM pVM = pVCpu->CTX_SUFF(pVM);
7515	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7516	uint32_t uErrCode = pVmcb->ctrl.u64ExitInfo1;
7517	uint64_t const uFaultAddress = pVmcb->ctrl.u64ExitInfo2;
7518
7519	#if defined(HMSVM_ALWAYS_TRAP_ALL_XCPTS) \|\| defined(HMSVM_ALWAYS_TRAP_PF)
7520	if (pVM->hm.s.fNestedPaging)
7521	{
7522	pVCpu->hm.s.Event.fPending = false; /* In case it's a contributory or vectoring #PF. */
7523	if ( !pSvmTransient->fVectoringDoublePF
7524	\|\| CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
7525	{
7526	/* A genuine guest #PF, reflect it to the guest. */
7527	hmR0SvmSetPendingXcptPF(pVCpu, pCtx, uErrCode, uFaultAddress);
7528	Log4Func(("#PF: Guest page fault at %04X:%RGv FaultAddr=%RX64 ErrCode=%#x\n", pCtx->cs.Sel, (RTGCPTR)pCtx->rip,
7529	uFaultAddress, uErrCode));
7530	}
7531	else
7532	{
7533	/* A guest page-fault occurred during delivery of a page-fault. Inject #DF. */
7534	hmR0SvmSetPendingXcptDF(pVCpu);
7535	Log4Func(("Pending #DF due to vectoring #PF. NP\n"));
7536	}
7537	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestPF);
7538	return VINF_SUCCESS;
7539	}
7540	#endif
7541
7542	Assert(!pVM->hm.s.fNestedPaging);
7543
7544	/*
7545	* TPR patching shortcut for APIC TPR reads and writes; only applicable to 32-bit guests.
7546	*/
7547	if ( pVM->hm.s.fTprPatchingAllowed
7548	&& (uFaultAddress & 0xfff) == XAPIC_OFF_TPR
7549	&& !(uErrCode & X86_TRAP_PF_P) /* Not present. */
7550	&& !CPUMIsGuestInSvmNestedHwVirtMode(pCtx)
7551	&& !CPUMIsGuestInLongModeEx(pCtx)
7552	&& !CPUMGetGuestCPL(pVCpu)
7553	&& pVM->hm.s.cPatches < RT_ELEMENTS(pVM->hm.s.aPatches))
7554	{
7555	RTGCPHYS GCPhysApicBase;
7556	GCPhysApicBase = APICGetBaseMsrNoCheck(pVCpu);
7557	GCPhysApicBase &= PAGE_BASE_GC_MASK;
7558
7559	/* Check if the page at the fault-address is the APIC base. */
7560	RTGCPHYS GCPhysPage;
7561	int rc2 = PGMGstGetPage(pVCpu, (RTGCPTR)uFaultAddress, NULL /* pfFlags */, &GCPhysPage);
7562	if ( rc2 == VINF_SUCCESS
7563	&& GCPhysPage == GCPhysApicBase)
7564	{
7565	/* Only attempt to patch the instruction once. */
7566	PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
7567	if (!pPatch)
7568	return VINF_EM_HM_PATCH_TPR_INSTR;
7569	}
7570	}
7571
7572	Log4Func(("#PF: uFaultAddress=%#RX64 CS:RIP=%#04x:%#RX64 uErrCode %#RX32 cr3=%#RX64\n", uFaultAddress, pCtx->cs.Sel,
7573	pCtx->rip, uErrCode, pCtx->cr3));
7574
7575	/*
7576	* If it's a vectoring #PF, emulate injecting the original event injection as
7577	* PGMTrap0eHandler() is incapable of differentiating between instruction emulation and
7578	* event injection that caused a #PF. See @bugref{6607}.
7579	*/
7580	if (pSvmTransient->fVectoringPF)
7581	{
7582	Assert(pVCpu->hm.s.Event.fPending);
7583	return VINF_EM_RAW_INJECT_TRPM_EVENT;
7584	}
7585
7586	TRPMAssertXcptPF(pVCpu, uFaultAddress, uErrCode);
7587	int rc = PGMTrap0eHandler(pVCpu, uErrCode, CPUMCTX2CORE(pCtx), (RTGCPTR)uFaultAddress);
7588
7589	Log4Func(("#PF: rc=%Rrc\n", rc));
7590
7591	if (rc == VINF_SUCCESS)
7592	{
7593	/* Successfully synced shadow pages tables or emulated an MMIO instruction. */
7594	TRPMResetTrap(pVCpu);
7595	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPF);
7596	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
7597	return rc;
7598	}
7599
7600	if (rc == VINF_EM_RAW_GUEST_TRAP)
7601	{
7602	pVCpu->hm.s.Event.fPending = false; /* In case it's a contributory or vectoring #PF. */
7603
7604	/*
7605	* If a nested-guest delivers a #PF and that causes a #PF which is -not- a shadow #PF,
7606	* we should simply forward the #PF to the guest and is up to the nested-hypervisor to
7607	* determine whether it is a nested-shadow #PF or a #DF, see @bugref{7243#c121}.
7608	*/
7609	if ( !pSvmTransient->fVectoringDoublePF
7610	\|\| CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
7611	{
7612	/* It's a guest (or nested-guest) page fault and needs to be reflected. */
7613	uErrCode = TRPMGetErrorCode(pVCpu); /* The error code might have been changed. */
7614	TRPMResetTrap(pVCpu);
7615
7616	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
7617	/* If the nested-guest is intercepting #PFs, cause a #PF #VMEXIT. */
7618	if ( CPUMIsGuestInSvmNestedHwVirtMode(pCtx)
7619	&& HMIsGuestSvmXcptInterceptSet(pVCpu, X86_XCPT_PF))
7620	return VBOXSTRICTRC_TODO(IEMExecSvmVmexit(pVCpu, SVM_EXIT_XCPT_PF, uErrCode, uFaultAddress));
7621	#endif
7622
7623	hmR0SvmSetPendingXcptPF(pVCpu, pCtx, uErrCode, uFaultAddress);
7624	}
7625	else
7626	{
7627	/* A guest page-fault occurred during delivery of a page-fault. Inject #DF. */
7628	TRPMResetTrap(pVCpu);
7629	hmR0SvmSetPendingXcptDF(pVCpu);
7630	Log4Func(("#PF: Pending #DF due to vectoring #PF\n"));
7631	}
7632
7633	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestPF);
7634	return VINF_SUCCESS;
7635	}
7636
7637	TRPMResetTrap(pVCpu);
7638	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPFEM);
7639	return rc;
7640	}
7641
7642
7643	/**
7644	* \#VMEXIT handler for undefined opcode (SVM_EXIT_XCPT_6).
7645	* Conditional \#VMEXIT.
7646	*/
7647	HMSVM_EXIT_DECL hmR0SvmExitXcptUD(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7648	{
7649	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7650	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(pCtx);
7651
7652	/* Paranoia; Ensure we cannot be called as a result of event delivery. */
7653	PSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
7654	Assert(!pVmcb->ctrl.ExitIntInfo.n.u1Valid); NOREF(pVmcb);
7655
7656	int rc = VERR_SVM_UNEXPECTED_XCPT_EXIT;
7657	if (pVCpu->hm.s.fGIMTrapXcptUD)
7658	{
7659	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7660	uint8_t cbInstr = 0;
7661	VBOXSTRICTRC rcStrict = GIMXcptUD(pVCpu, pCtx, NULL /* pDis */, &cbInstr);
7662	if (rcStrict == VINF_SUCCESS)
7663	{
7664	/* #UD #VMEXIT does not have valid NRIP information, manually advance RIP. See @bugref{7270#c170}. */
7665	hmR0SvmAdvanceRipDumb(pVCpu, pCtx, cbInstr);
7666	rc = VINF_SUCCESS;
7667	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
7668	}
7669	else if (rcStrict == VINF_GIM_HYPERCALL_CONTINUING)
7670	rc = VINF_SUCCESS;
7671	else if (rcStrict == VINF_GIM_R3_HYPERCALL)
7672	rc = VINF_GIM_R3_HYPERCALL;
7673	else
7674	Assert(RT_FAILURE(VBOXSTRICTRC_VAL(rcStrict)));
7675	}
7676
7677	/* If the GIM #UD exception handler didn't succeed for some reason or wasn't needed, raise #UD. */
7678	if (RT_FAILURE(rc))
7679	{
7680	hmR0SvmSetPendingXcptUD(pVCpu);
7681	rc = VINF_SUCCESS;
7682	}
7683
7684	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestUD);
7685	return rc;
7686	}
7687
7688
7689	/**
7690	* \#VMEXIT handler for math-fault exceptions (SVM_EXIT_XCPT_16).
7691	* Conditional \#VMEXIT.
7692	*/
7693	HMSVM_EXIT_DECL hmR0SvmExitXcptMF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7694	{
7695	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7696	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7697
7698	/* Paranoia; Ensure we cannot be called as a result of event delivery. */
7699	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7700	Assert(!pVmcb->ctrl.ExitIntInfo.n.u1Valid); NOREF(pVmcb);
7701
7702	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestMF);
7703
7704	if (!(pCtx->cr0 & X86_CR0_NE))
7705	{
7706	PVM pVM = pVCpu->CTX_SUFF(pVM);
7707	PDISSTATE pDis = &pVCpu->hm.s.DisState;
7708	unsigned cbOp;
7709	int rc = EMInterpretDisasCurrent(pVM, pVCpu, pDis, &cbOp);
7710	if (RT_SUCCESS(rc))
7711	{
7712	/* Convert a #MF into a FERR -> IRQ 13. See @bugref{6117}. */
7713	rc = PDMIsaSetIrq(pVCpu->CTX_SUFF(pVM), 13 /* u8Irq /, 1 / u8Level /, 0 / uTagSrc */);
7714	if (RT_SUCCESS(rc))
7715	pCtx->rip += cbOp;
7716	}
7717	else
7718	Log4Func(("EMInterpretDisasCurrent returned %Rrc uOpCode=%#x\n", rc, pDis->pCurInstr->uOpcode));
7719	return rc;
7720	}
7721
7722	hmR0SvmSetPendingXcptMF(pVCpu);
7723	return VINF_SUCCESS;
7724	}
7725
7726
7727	/**
7728	* \#VMEXIT handler for debug exceptions (SVM_EXIT_XCPT_1). Conditional
7729	* \#VMEXIT.
7730	*/
7731	HMSVM_EXIT_DECL hmR0SvmExitXcptDB(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7732	{
7733	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7734	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7735	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
7736
7737	if (RT_UNLIKELY(pVCpu->hm.s.Event.fPending))
7738	{
7739	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectPendingInterpret);
7740	return VINF_EM_RAW_INJECT_TRPM_EVENT;
7741	}
7742
7743	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestDB);
7744
7745	/*
7746	* This can be a fault-type #DB (instruction breakpoint) or a trap-type #DB (data
7747	* breakpoint). However, for both cases DR6 and DR7 are updated to what the exception
7748	* handler expects. See AMD spec. 15.12.2 "#DB (Debug)".
7749	*/
7750	PVM pVM = pVCpu->CTX_SUFF(pVM);
7751	PSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
7752	int rc = DBGFRZTrap01Handler(pVM, pVCpu, CPUMCTX2CORE(pCtx), pVmcb->guest.u64DR6, pVCpu->hm.s.fSingleInstruction);
7753	if (rc == VINF_EM_RAW_GUEST_TRAP)
7754	{
7755	Log5(("hmR0SvmExitXcptDB: DR6=%#RX64 -> guest trap\n", pVmcb->guest.u64DR6));
7756	if (CPUMIsHyperDebugStateActive(pVCpu))
7757	CPUMSetGuestDR6(pVCpu, CPUMGetGuestDR6(pVCpu) \| pVmcb->guest.u64DR6);
7758
7759	/* Reflect the exception back to the guest. */
7760	hmR0SvmSetPendingXcptDB(pVCpu);
7761	rc = VINF_SUCCESS;
7762	}
7763
7764	/*
7765	* Update DR6.
7766	*/
7767	if (CPUMIsHyperDebugStateActive(pVCpu))
7768	{
7769	Log5(("hmR0SvmExitXcptDB: DR6=%#RX64 -> %Rrc\n", pVmcb->guest.u64DR6, rc));
7770	pVmcb->guest.u64DR6 = X86_DR6_INIT_VAL;
7771	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
7772	}
7773	else
7774	{
7775	AssertMsg(rc == VINF_SUCCESS, ("rc=%Rrc\n", rc));
7776	Assert(!pVCpu->hm.s.fSingleInstruction && !DBGFIsStepping(pVCpu));
7777	}
7778
7779	return rc;
7780	}
7781
7782
7783	/**
7784	* \#VMEXIT handler for alignment check exceptions (SVM_EXIT_XCPT_17).
7785	* Conditional \#VMEXIT.
7786	*/
7787	HMSVM_EXIT_DECL hmR0SvmExitXcptAC(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7788	{
7789	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7790	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
7791
7792	SVMEVENT Event;
7793	Event.u = 0;
7794	Event.n.u1Valid = 1;
7795	Event.n.u3Type = SVM_EVENT_EXCEPTION;
7796	Event.n.u8Vector = X86_XCPT_AC;
7797	Event.n.u1ErrorCodeValid = 1;
7798	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
7799	return VINF_SUCCESS;
7800	}
7801
7802
7803	/**
7804	* \#VMEXIT handler for breakpoint exceptions (SVM_EXIT_XCPT_3).
7805	* Conditional \#VMEXIT.
7806	*/
7807	HMSVM_EXIT_DECL hmR0SvmExitXcptBP(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7808	{
7809	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7810	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7811	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
7812
7813	int rc = DBGFRZTrap03Handler(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
7814	if (rc == VINF_EM_RAW_GUEST_TRAP)
7815	{
7816	SVMEVENT Event;
7817	Event.u = 0;
7818	Event.n.u1Valid = 1;
7819	Event.n.u3Type = SVM_EVENT_EXCEPTION;
7820	Event.n.u8Vector = X86_XCPT_BP;
7821	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
7822	}
7823
7824	Assert(rc == VINF_SUCCESS \|\| rc == VINF_EM_RAW_GUEST_TRAP \|\| rc == VINF_EM_DBG_BREAKPOINT);
7825	return rc;
7826	}
7827
7828
7829	#if defined(HMSVM_ALWAYS_TRAP_ALL_XCPTS) \|\| defined(VBOX_WITH_NESTED_HWVIRT_SVM)
7830	/**
7831	* \#VMEXIT handler for generic exceptions. Conditional \#VMEXIT.
7832	*/
7833	HMSVM_EXIT_DECL hmR0SvmExitXcptGeneric(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7834	{
7835	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7836	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
7837
7838	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7839	uint8_t const uVector = pVmcb->ctrl.u64ExitCode - SVM_EXIT_XCPT_0;
7840	uint32_t const uErrCode = pVmcb->ctrl.u64ExitInfo1;
7841	Assert(pSvmTransient->u64ExitCode == pVmcb->ctrl.u64ExitCode);
7842	Assert(uVector <= X86_XCPT_LAST);
7843	Log4Func(("uVector=%#x uErrCode=%u\n", uVector, uErrCode));
7844
7845	SVMEVENT Event;
7846	Event.u = 0;
7847	Event.n.u1Valid = 1;
7848	Event.n.u3Type = SVM_EVENT_EXCEPTION;
7849	Event.n.u8Vector = uVector;
7850	switch (uVector)
7851	{
7852	/* Shouldn't be here for reflecting #PFs (among other things, the fault address isn't passed along). */
7853	case X86_XCPT_PF: AssertMsgFailed(("hmR0SvmExitXcptGeneric: Unexpected exception")); return VERR_SVM_IPE_5;
7854	case X86_XCPT_DF:
7855	case X86_XCPT_TS:
7856	case X86_XCPT_NP:
7857	case X86_XCPT_SS:
7858	case X86_XCPT_GP:
7859	case X86_XCPT_AC:
7860	{
7861	Event.n.u1ErrorCodeValid = 1;
7862	Event.n.u32ErrorCode = uErrCode;
7863	break;
7864	}
7865	}
7866
7867	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
7868	return VINF_SUCCESS;
7869	}
7870	#endif
7871
7872	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
7873	/**
7874	* \#VMEXIT handler for CLGI (SVM_EXIT_CLGI). Conditional \#VMEXIT.
7875	*/
7876	HMSVM_EXIT_DECL hmR0SvmExitClgi(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7877	{
7878	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7879	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK
7880	\| CPUMCTX_EXTRN_HWVIRT);
7881
7882	#ifdef VBOX_STRICT
7883	PCSVMVMCB pVmcbTmp = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7884	Assert(pVmcbTmp);
7885	Assert(!pVmcbTmp->ctrl.IntCtrl.n.u1VGifEnable);
7886	RT_NOREF(pVmcbTmp);
7887	#endif
7888
7889	uint8_t const cbInstr = hmR0SvmGetInstrLengthHwAssist(pVCpu, pCtx, 3);
7890	VBOXSTRICTRC rcStrict = IEMExecDecodedClgi(pVCpu, cbInstr);
7891	if (rcStrict == VINF_SUCCESS)
7892	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_HWVIRT);
7893	return VBOXSTRICTRC_VAL(rcStrict);
7894	}
7895
7896
7897	/**
7898	* \#VMEXIT handler for STGI (SVM_EXIT_STGI). Conditional \#VMEXIT.
7899	*/
7900	HMSVM_EXIT_DECL hmR0SvmExitStgi(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7901	{
7902	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7903	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK
7904	\| CPUMCTX_EXTRN_HWVIRT);
7905
7906	/*
7907	* When VGIF is not used we always intercept STGI instructions. When VGIF is used,
7908	* we only intercept STGI when events are pending for GIF to become 1.
7909	*/
7910	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7911	if (pVmcb->ctrl.IntCtrl.n.u1VGifEnable)
7912	hmR0SvmClearCtrlIntercept(pVCpu, pCtx, pVmcb, SVM_CTRL_INTERCEPT_STGI);
7913
7914	uint8_t const cbInstr = hmR0SvmGetInstrLengthHwAssist(pVCpu, pCtx, 3);
7915	VBOXSTRICTRC rcStrict = IEMExecDecodedStgi(pVCpu, cbInstr);
7916	if (rcStrict == VINF_SUCCESS)
7917	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_HWVIRT);
7918	return VBOXSTRICTRC_VAL(rcStrict);
7919	}
7920
7921
7922	/**
7923	* \#VMEXIT handler for VMLOAD (SVM_EXIT_VMLOAD). Conditional \#VMEXIT.
7924	*/
7925	HMSVM_EXIT_DECL hmR0SvmExitVmload(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7926	{
7927	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7928	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK
7929	\| CPUMCTX_EXTRN_FS
7930	\| CPUMCTX_EXTRN_GS
7931	\| CPUMCTX_EXTRN_TR
7932	\| CPUMCTX_EXTRN_LDTR
7933	\| CPUMCTX_EXTRN_KERNEL_GS_BASE
7934	\| CPUMCTX_EXTRN_SYSCALL_MSRS
7935	\| CPUMCTX_EXTRN_SYSENTER_MSRS);
7936
7937	#ifdef VBOX_STRICT
7938	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7939	Assert(pVmcb);
7940	Assert(!pVmcb->ctrl.LbrVirt.n.u1VirtVmsaveVmload);
7941	RT_NOREF(pVmcb);
7942	#endif
7943
7944	uint8_t const cbInstr = hmR0SvmGetInstrLengthHwAssist(pVCpu, pCtx, 3);
7945	VBOXSTRICTRC rcStrict = IEMExecDecodedVmload(pVCpu, cbInstr);
7946	if (rcStrict == VINF_SUCCESS)
7947	{
7948	/* We skip flagging changes made to LSTAR, STAR, SFMASK and other MSRs as they are always re-loaded. */
7949	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_FS
7950	\| HM_CHANGED_GUEST_GS
7951	\| HM_CHANGED_GUEST_TR
7952	\| HM_CHANGED_GUEST_LDTR
7953	\| HM_CHANGED_GUEST_SYSENTER_CS_MSR
7954	\| HM_CHANGED_GUEST_SYSENTER_EIP_MSR
7955	\| HM_CHANGED_GUEST_SYSENTER_ESP_MSR);
7956	}
7957	return VBOXSTRICTRC_VAL(rcStrict);
7958	}
7959
7960
7961	/**
7962	* \#VMEXIT handler for VMSAVE (SVM_EXIT_VMSAVE). Conditional \#VMEXIT.
7963	*/
7964	HMSVM_EXIT_DECL hmR0SvmExitVmsave(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7965	{
7966	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7967	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
7968
7969	#ifdef VBOX_STRICT
7970	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7971	Assert(pVmcb);
7972	Assert(!pVmcb->ctrl.LbrVirt.n.u1VirtVmsaveVmload);
7973	RT_NOREF(pVmcb);
7974	#endif
7975
7976	uint8_t const cbInstr = hmR0SvmGetInstrLengthHwAssist(pVCpu, pCtx, 3);
7977	VBOXSTRICTRC rcStrict = IEMExecDecodedVmsave(pVCpu, cbInstr);
7978	return VBOXSTRICTRC_VAL(rcStrict);
7979	}
7980
7981
7982	/**
7983	* \#VMEXIT handler for INVLPGA (SVM_EXIT_INVLPGA). Conditional \#VMEXIT.
7984	*/
7985	HMSVM_EXIT_DECL hmR0SvmExitInvlpga(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7986	{
7987	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7988	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
7989
7990	uint8_t const cbInstr = hmR0SvmGetInstrLengthHwAssist(pVCpu, pCtx, 3);
7991	VBOXSTRICTRC rcStrict = IEMExecDecodedInvlpga(pVCpu, cbInstr);
7992	return VBOXSTRICTRC_VAL(rcStrict);
7993	}
7994
7995
7996	/**
7997	* \#VMEXIT handler for STGI (SVM_EXIT_VMRUN). Conditional \#VMEXIT.
7998	*/
7999	HMSVM_EXIT_DECL hmR0SvmExitVmrun(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
8000	{
8001	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
8002	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK
8003	\| IEM_CPUMCTX_EXTRN_SVM_VMRUN_MASK);
8004	VBOXSTRICTRC rcStrict;
8005	uint8_t const cbInstr = hmR0SvmGetInstrLengthHwAssist(pVCpu, pCtx, 3);
8006	rcStrict = IEMExecDecodedVmrun(pVCpu, cbInstr);
8007	Log4Func(("IEMExecDecodedVmrun returns %Rrc\n", VBOXSTRICTRC_VAL(rcStrict)));
8008	if (rcStrict == VINF_SUCCESS)
8009	{
8010	rcStrict = VINF_SVM_VMRUN;
8011	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_SVM_VMRUN_MASK);
8012	}
8013	return VBOXSTRICTRC_VAL(rcStrict);
8014	}
8015
8016
8017	/**
8018	* Nested-guest \#VMEXIT handler for debug exceptions (SVM_EXIT_XCPT_1).
8019	* Unconditional \#VMEXIT.
8020	*/
8021	HMSVM_EXIT_DECL hmR0SvmNestedExitXcptDB(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
8022	{
8023	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
8024	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
8025
8026	if (pVCpu->hm.s.Event.fPending)
8027	{
8028	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectPendingInterpret);
8029	return VINF_EM_RAW_INJECT_TRPM_EVENT;
8030	}
8031
8032	hmR0SvmSetPendingXcptDB(pVCpu);
8033	return VINF_SUCCESS;
8034	}
8035
8036
8037	/**
8038	* Nested-guest \#VMEXIT handler for breakpoint exceptions (SVM_EXIT_XCPT_3).
8039	* Conditional \#VMEXIT.
8040	*/
8041	HMSVM_EXIT_DECL hmR0SvmNestedExitXcptBP(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
8042	{
8043	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
8044	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
8045
8046	SVMEVENT Event;
8047	Event.u = 0;
8048	Event.n.u1Valid = 1;
8049	Event.n.u3Type = SVM_EVENT_EXCEPTION;
8050	Event.n.u8Vector = X86_XCPT_BP;
8051	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
8052	return VINF_SUCCESS;
8053	}
8054
8055	#endif /* VBOX_WITH_NESTED_HWVIRT_SVM */
8056
8057
8058	/** @} */
8059

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

source: vbox/trunk/src/VBox/VMM/VMMR0/HMSVMR0.cpp@ 72783

Download in other formats: