HMSVMR0.cpp@ 72747

Last change on this file since 72747 was 72744, checked in by vboxsync, 6 years ago
VMM: Extend HM changed flags. bugref:9193 [HM, CPUM]
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 72744 2018-06-29 07:36:19Z 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);
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 = CPUMQueryGuestCtxPtr(pVCpu);
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	uint32_t aParam[8];
1368	aParam[0] = RT_LO_U32(HCPhysVmcbHost); /* Param 1: HCPhysVmcbHost - Lo. */
1369	aParam[1] = RT_HI_U32(HCPhysVmcbHost); /* Param 1: HCPhysVmcbHost - Hi. */
1370	aParam[2] = RT_LO_U32(HCPhysVmcb); /* Param 2: HCPhysVmcb - Lo. */
1371	aParam[3] = RT_HI_U32(HCPhysVmcb); /* Param 2: HCPhysVmcb - Hi. */
1372	aParam[4] = VM_RC_ADDR(pVM, pVM);
1373	aParam[5] = 0;
1374	aParam[6] = VM_RC_ADDR(pVM, pVCpu);
1375	aParam[7] = 0;
1376
1377	return SVMR0Execute64BitsHandler(pVCpu, HM64ON32OP_SVMRCVMRun64, RT_ELEMENTS(aParam), &aParam[0]);
1378	}
1379
1380
1381	/**
1382	* Executes the specified VMRUN handler in 64-bit mode.
1383	*
1384	* @returns VBox status code.
1385	* @param pVCpu The cross context virtual CPU structure.
1386	* @param enmOp The operation to perform.
1387	* @param cParams Number of parameters.
1388	* @param paParam Array of 32-bit parameters.
1389	*/
1390	VMMR0DECL(int) SVMR0Execute64BitsHandler(PVMCPU pVCpu, HM64ON32OP enmOp, uint32_t cParams, uint32_t *paParam)
1391	{
1392	PVM pVM = pVCpu->CTX_SUFF(pVM);
1393	AssertReturn(pVM->hm.s.pfnHost32ToGuest64R0, VERR_HM_NO_32_TO_64_SWITCHER);
1394	Assert(enmOp > HM64ON32OP_INVALID && enmOp < HM64ON32OP_END);
1395
1396	/* Disable interrupts. */
1397	RTHCUINTREG const fEFlags = ASMIntDisableFlags();
1398
1399	#ifdef VBOX_WITH_VMMR0_DISABLE_LAPIC_NMI
1400	RTCPUID idHostCpu = RTMpCpuId();
1401	CPUMR0SetLApic(pVCpu, idHostCpu);
1402	#endif
1403
1404	CPUMSetHyperESP(pVCpu, VMMGetStackRC(pVCpu));
1405	CPUMSetHyperEIP(pVCpu, enmOp);
1406	for (int i = (int)cParams - 1; i >= 0; i--)
1407	CPUMPushHyper(pVCpu, paParam[i]);
1408
1409	/* Call the switcher. */
1410	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatWorldSwitch3264, z);
1411	int rc = pVM->hm.s.pfnHost32ToGuest64R0(pVM, RT_OFFSETOF(VM, aCpus[pVCpu->idCpu].cpum) - RT_OFFSETOF(VM, cpum));
1412	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatWorldSwitch3264, z);
1413
1414	/* Restore interrupts. */
1415	ASMSetFlags(fEFlags);
1416	return rc;
1417	}
1418
1419	#endif /* HC_ARCH_BITS == 32 && defined(VBOX_ENABLE_64_BITS_GUESTS) */
1420	/** @} */
1421
1422
1423	/**
1424	* Sets an exception intercept in the specified VMCB.
1425	*
1426	* @param pVmcb Pointer to the VM control block.
1427	* @param uXcpt The exception (X86_XCPT_*).
1428	*/
1429	DECLINLINE(void) hmR0SvmSetXcptIntercept(PSVMVMCB pVmcb, uint8_t uXcpt)
1430	{
1431	if (!(pVmcb->ctrl.u32InterceptXcpt & RT_BIT(uXcpt)))
1432	{
1433	pVmcb->ctrl.u32InterceptXcpt \|= RT_BIT(uXcpt);
1434	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1435	}
1436	}
1437
1438
1439	/**
1440	* Clears an exception intercept in the specified VMCB.
1441	*
1442	* @param pVCpu The cross context virtual CPU structure.
1443	* @param pVmcb Pointer to the VM control block.
1444	* @param uXcpt The exception (X86_XCPT_*).
1445	*
1446	* @remarks This takes into account if we're executing a nested-guest and only
1447	* removes the exception intercept if both the guest -and- nested-guest
1448	* are not intercepting it.
1449	*/
1450	DECLINLINE(void) hmR0SvmClearXcptIntercept(PVMCPU pVCpu, PSVMVMCB pVmcb, uint8_t uXcpt)
1451	{
1452	Assert(uXcpt != X86_XCPT_DB);
1453	Assert(uXcpt != X86_XCPT_AC);
1454	#ifndef HMSVM_ALWAYS_TRAP_ALL_XCPTS
1455	if (pVmcb->ctrl.u32InterceptXcpt & RT_BIT(uXcpt))
1456	{
1457	bool fRemove = true;
1458	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1459	/* Only remove the intercept if the nested-guest is also not intercepting it! */
1460	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1461	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
1462	{
1463	PCSVMNESTEDVMCBCACHE pVmcbNstGstCache = hmR0SvmGetNestedVmcbCache(pVCpu);
1464	fRemove = !(pVmcbNstGstCache->u32InterceptXcpt & RT_BIT(uXcpt));
1465	}
1466	#endif
1467	if (fRemove)
1468	{
1469	pVmcb->ctrl.u32InterceptXcpt &= ~RT_BIT(uXcpt);
1470	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1471	}
1472	}
1473	#else
1474	RT_NOREF3(pVCpu, pVmcb, uXcpt);
1475	#endif
1476	}
1477
1478
1479	/**
1480	* Sets a control intercept in the specified VMCB.
1481	*
1482	* @param pVmcb Pointer to the VM control block.
1483	* @param fCtrlIntercept The control intercept (SVM_CTRL_INTERCEPT_*).
1484	*/
1485	DECLINLINE(void) hmR0SvmSetCtrlIntercept(PSVMVMCB pVmcb, uint64_t fCtrlIntercept)
1486	{
1487	if (!(pVmcb->ctrl.u64InterceptCtrl & fCtrlIntercept))
1488	{
1489	pVmcb->ctrl.u64InterceptCtrl \|= fCtrlIntercept;
1490	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1491	}
1492	}
1493
1494
1495	/**
1496	* Clears a control intercept in the specified VMCB.
1497	*
1498	* @returns @c true if the intercept is still set, @c false otherwise.
1499	* @param pVCpu The cross context virtual CPU structure.
1500	* @param pCtx Pointer to the guest-CPU context.
1501	* @param pVmcb Pointer to the VM control block.
1502	* @param fCtrlIntercept The control intercept (SVM_CTRL_INTERCEPT_*).
1503	*
1504	* @remarks This takes into account if we're executing a nested-guest and only
1505	* removes the control intercept if both the guest -and- nested-guest
1506	* are not intercepting it.
1507	*/
1508	DECLINLINE(bool) hmR0SvmClearCtrlIntercept(PVMCPU pVCpu, PCCPUMCTX pCtx, PSVMVMCB pVmcb, uint64_t fCtrlIntercept)
1509	{
1510	if (pVmcb->ctrl.u64InterceptCtrl & fCtrlIntercept)
1511	{
1512	bool fRemove = true;
1513	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1514	/* Only remove the control intercept if the nested-guest is also not intercepting it! */
1515	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
1516	{
1517	PCSVMNESTEDVMCBCACHE pVmcbNstGstCache = hmR0SvmGetNestedVmcbCache(pVCpu);
1518	fRemove = !(pVmcbNstGstCache->u64InterceptCtrl & fCtrlIntercept);
1519	}
1520	#else
1521	RT_NOREF2(pVCpu, pCtx);
1522	#endif
1523	if (fRemove)
1524	{
1525	pVmcb->ctrl.u64InterceptCtrl &= ~fCtrlIntercept;
1526	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1527	}
1528	}
1529
1530	return RT_BOOL(pVmcb->ctrl.u64InterceptCtrl & fCtrlIntercept);
1531	}
1532
1533
1534	/**
1535	* Exports the guest (or nested-guest) CR0 into the VMCB.
1536	*
1537	* @param pVCpu The cross context virtual CPU structure.
1538	* @param pVmcb Pointer to the VM control block.
1539	*
1540	* @remarks This assumes we always pre-load the guest FPU.
1541	* @remarks No-long-jump zone!!!
1542	*/
1543	static void hmR0SvmExportGuestCR0(PVMCPU pVCpu, PSVMVMCB pVmcb)
1544	{
1545	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1546
1547	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1548	uint64_t const uGuestCr0 = pCtx->cr0;
1549	uint64_t uShadowCr0 = uGuestCr0;
1550
1551	/* Always enable caching. */
1552	uShadowCr0 &= ~(X86_CR0_CD \| X86_CR0_NW);
1553
1554	/* When Nested Paging is not available use shadow page tables and intercept #PFs (latter done in SVMR0SetupVM()). */
1555	if (!pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging)
1556	{
1557	uShadowCr0 \|= X86_CR0_PG /* Use shadow page tables. */
1558	\| X86_CR0_WP; /* Guest CPL 0 writes to its read-only pages should cause a #PF #VMEXIT. */
1559	}
1560
1561	/*
1562	* Use the #MF style of legacy-FPU error reporting for now. Although AMD-V has MSRs that
1563	* lets us isolate the host from it, IEM/REM still needs work to emulate it properly,
1564	* see @bugref{7243#c103}.
1565	*/
1566	if (!(uGuestCr0 & X86_CR0_NE))
1567	{
1568	uShadowCr0 \|= X86_CR0_NE;
1569	hmR0SvmSetXcptIntercept(pVmcb, X86_XCPT_MF);
1570	}
1571	else
1572	hmR0SvmClearXcptIntercept(pVCpu, pVmcb, X86_XCPT_MF);
1573
1574	/*
1575	* If the shadow and guest CR0 are identical we can avoid intercepting CR0 reads.
1576	*
1577	* CR0 writes still needs interception as PGM requires tracking paging mode changes,
1578	* see @bugref{6944}.
1579	*
1580	* We also don't ever want to honor weird things like cache disable from the guest.
1581	* However, we can avoid intercepting changes to the TS & MP bits by clearing the CR0
1582	* write intercept below and keeping SVM_CTRL_INTERCEPT_CR0_SEL_WRITE instead.
1583	*/
1584	if (uShadowCr0 == uGuestCr0)
1585	{
1586	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
1587	{
1588	pVmcb->ctrl.u16InterceptRdCRx &= ~RT_BIT(0);
1589	pVmcb->ctrl.u16InterceptWrCRx &= ~RT_BIT(0);
1590	Assert(pVmcb->ctrl.u64InterceptCtrl & SVM_CTRL_INTERCEPT_CR0_SEL_WRITE);
1591	}
1592	else
1593	{
1594	/* If the nested-hypervisor intercepts CR0 reads/writes, we need to continue intercepting them. */
1595	PCSVMNESTEDVMCBCACHE pVmcbNstGstCache = hmR0SvmGetNestedVmcbCache(pVCpu);
1596	pVmcb->ctrl.u16InterceptRdCRx = (pVmcb->ctrl.u16InterceptRdCRx & ~RT_BIT(0))
1597	\| (pVmcbNstGstCache->u16InterceptRdCRx & RT_BIT(0));
1598	pVmcb->ctrl.u16InterceptWrCRx = (pVmcb->ctrl.u16InterceptWrCRx & ~RT_BIT(0))
1599	\| (pVmcbNstGstCache->u16InterceptWrCRx & RT_BIT(0));
1600	}
1601	}
1602	else
1603	{
1604	pVmcb->ctrl.u16InterceptRdCRx \|= RT_BIT(0);
1605	pVmcb->ctrl.u16InterceptWrCRx \|= RT_BIT(0);
1606	}
1607	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1608
1609	Assert(RT_HI_U32(uShadowCr0) == 0);
1610	if (pVmcb->guest.u64CR0 != uShadowCr0)
1611	{
1612	pVmcb->guest.u64CR0 = uShadowCr0;
1613	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1614	}
1615	}
1616
1617
1618	/**
1619	* Exports the guest (or nested-guest) CR3 into the VMCB.
1620	*
1621	* @param pVCpu The cross context virtual CPU structure.
1622	* @param pVmcb Pointer to the VM control block.
1623	*
1624	* @remarks No-long-jump zone!!!
1625	*/
1626	static void hmR0SvmExportGuestCR3(PVMCPU pVCpu, PSVMVMCB pVmcb)
1627	{
1628	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1629
1630	PVM pVM = pVCpu->CTX_SUFF(pVM);
1631	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1632	if (pVM->hm.s.fNestedPaging)
1633	{
1634	PGMMODE enmShwPagingMode;
1635	#if HC_ARCH_BITS == 32
1636	if (CPUMIsGuestInLongModeEx(pCtx))
1637	enmShwPagingMode = PGMMODE_AMD64_NX;
1638	else
1639	#endif
1640	enmShwPagingMode = PGMGetHostMode(pVM);
1641
1642	pVmcb->ctrl.u64NestedPagingCR3 = PGMGetNestedCR3(pVCpu, enmShwPagingMode);
1643	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
1644	pVmcb->guest.u64CR3 = pCtx->cr3;
1645	Assert(pVmcb->ctrl.u64NestedPagingCR3);
1646	}
1647	else
1648	pVmcb->guest.u64CR3 = PGMGetHyperCR3(pVCpu);
1649
1650	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1651	}
1652
1653
1654	/**
1655	* Exports the guest (or nested-guest) CR4 into the VMCB.
1656	*
1657	* @param pVCpu The cross context virtual CPU structure.
1658	* @param pVmcb Pointer to the VM control block.
1659	*
1660	* @remarks No-long-jump zone!!!
1661	*/
1662	static int hmR0SvmExportGuestCR4(PVMCPU pVCpu, PSVMVMCB pVmcb)
1663	{
1664	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1665
1666	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1667	uint64_t uShadowCr4 = pCtx->cr4;
1668	if (!pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging)
1669	{
1670	switch (pVCpu->hm.s.enmShadowMode)
1671	{
1672	case PGMMODE_REAL:
1673	case PGMMODE_PROTECTED: /* Protected mode, no paging. */
1674	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1675
1676	case PGMMODE_32_BIT: /* 32-bit paging. */
1677	uShadowCr4 &= ~X86_CR4_PAE;
1678	break;
1679
1680	case PGMMODE_PAE: /* PAE paging. */
1681	case PGMMODE_PAE_NX: /* PAE paging with NX enabled. */
1682	/** Must use PAE paging as we could use physical memory > 4 GB */
1683	uShadowCr4 \|= X86_CR4_PAE;
1684	break;
1685
1686	case PGMMODE_AMD64: /* 64-bit AMD paging (long mode). */
1687	case PGMMODE_AMD64_NX: /* 64-bit AMD paging (long mode) with NX enabled. */
1688	#ifdef VBOX_ENABLE_64_BITS_GUESTS
1689	break;
1690	#else
1691	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1692	#endif
1693
1694	default: /* shut up gcc */
1695	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1696	}
1697	}
1698
1699	/* Whether to save/load/restore XCR0 during world switch depends on CR4.OSXSAVE and host+guest XCR0. */
1700	pVCpu->hm.s.fLoadSaveGuestXcr0 = (pCtx->cr4 & X86_CR4_OSXSAVE) && pCtx->aXcr[0] != ASMGetXcr0();
1701
1702	/* Avoid intercepting CR4 reads if the guest and shadow CR4 values are identical. */
1703	if (uShadowCr4 == pCtx->cr4)
1704	{
1705	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
1706	pVmcb->ctrl.u16InterceptRdCRx &= ~RT_BIT(4);
1707	else
1708	{
1709	/* If the nested-hypervisor intercepts CR4 reads, we need to continue intercepting them. */
1710	PCSVMNESTEDVMCBCACHE pVmcbNstGstCache = hmR0SvmGetNestedVmcbCache(pVCpu);
1711	pVmcb->ctrl.u16InterceptRdCRx = (pVmcb->ctrl.u16InterceptRdCRx & ~RT_BIT(4))
1712	\| (pVmcbNstGstCache->u16InterceptRdCRx & RT_BIT(4));
1713	}
1714	}
1715	else
1716	pVmcb->ctrl.u16InterceptRdCRx \|= RT_BIT(4);
1717
1718	/* CR4 writes are always intercepted (both guest, nested-guest) for tracking PGM mode changes. */
1719	Assert(pVmcb->ctrl.u16InterceptWrCRx & RT_BIT(4));
1720
1721	/* Update VMCB with the shadow CR4 the appropriate VMCB clean bits. */
1722	Assert(RT_HI_U32(uShadowCr4) == 0);
1723	pVmcb->guest.u64CR4 = uShadowCr4;
1724	pVmcb->ctrl.u32VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_CRX_EFER \| HMSVM_VMCB_CLEAN_INTERCEPTS);
1725
1726	return VINF_SUCCESS;
1727	}
1728
1729
1730	/**
1731	* Exports the guest (or nested-guest) control registers into the VMCB.
1732	*
1733	* @returns VBox status code.
1734	* @param pVCpu The cross context virtual CPU structure.
1735	* @param pVmcb Pointer to the VM control block.
1736	*
1737	* @remarks No-long-jump zone!!!
1738	*/
1739	static int hmR0SvmExportGuestControlRegs(PVMCPU pVCpu, PSVMVMCB pVmcb)
1740	{
1741	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1742
1743	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_CR_MASK)
1744	{
1745	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_CR0)
1746	hmR0SvmExportGuestCR0(pVCpu, pVmcb);
1747
1748	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_CR2)
1749	{
1750	pVmcb->guest.u64CR2 = pVCpu->cpum.GstCtx.cr2;
1751	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CR2;
1752	}
1753
1754	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_CR3)
1755	hmR0SvmExportGuestCR3(pVCpu, pVmcb);
1756
1757	/* CR4 re-loading is ASSUMED to be done everytime we get in from ring-3! (XCR0) */
1758	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_CR4)
1759	{
1760	int rc = hmR0SvmExportGuestCR4(pVCpu, pVmcb);
1761	if (RT_FAILURE(rc))
1762	return rc;
1763	}
1764
1765	pVCpu->hm.s.fCtxChanged &= ~HM_CHANGED_GUEST_CR_MASK;
1766	}
1767	return VINF_SUCCESS;
1768	}
1769
1770
1771	/**
1772	* Exports the guest (or nested-guest) segment registers into the VMCB.
1773	*
1774	* @returns VBox status code.
1775	* @param pVCpu The cross context virtual CPU structure.
1776	* @param pVmcb Pointer to the VM control block.
1777	*
1778	* @remarks No-long-jump zone!!!
1779	*/
1780	static void hmR0SvmExportGuestSegmentRegs(PVMCPU pVCpu, PSVMVMCB pVmcb)
1781	{
1782	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1783	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1784
1785	/* Guest segment registers. */
1786	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SREG_MASK)
1787	{
1788	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_CS)
1789	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, CS, cs);
1790
1791	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SS)
1792	{
1793	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, SS, ss);
1794	pVmcb->guest.u8CPL = pCtx->ss.Attr.n.u2Dpl;
1795	}
1796
1797	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_DS)
1798	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, DS, ds);
1799
1800	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_ES)
1801	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, ES, es);
1802
1803	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_FS)
1804	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, FS, fs);
1805
1806	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_GS)
1807	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, GS, gs);
1808
1809	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_SEG;
1810	}
1811
1812	/* Guest TR. */
1813	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_TR)
1814	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, TR, tr);
1815
1816	/* Guest LDTR. */
1817	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_LDTR)
1818	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, LDTR, ldtr);
1819
1820	/* Guest GDTR. */
1821	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_GDTR)
1822	{
1823	pVmcb->guest.GDTR.u32Limit = pCtx->gdtr.cbGdt;
1824	pVmcb->guest.GDTR.u64Base = pCtx->gdtr.pGdt;
1825	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DT;
1826	}
1827
1828	/* Guest IDTR. */
1829	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_IDTR)
1830	{
1831	pVmcb->guest.IDTR.u32Limit = pCtx->idtr.cbIdt;
1832	pVmcb->guest.IDTR.u64Base = pCtx->idtr.pIdt;
1833	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DT;
1834	}
1835
1836	pVCpu->hm.s.fCtxChanged &= ~( HM_CHANGED_GUEST_SREG_MASK
1837	\| HM_CHANGED_GUEST_TABLE_MASK);
1838	}
1839
1840
1841	/**
1842	* Exports the guest (or nested-guest) MSRs into the VMCB.
1843	*
1844	* @param pVCpu The cross context virtual CPU structure.
1845	* @param pVmcb Pointer to the VM control block.
1846	* @param pCtx Pointer to the guest-CPU context.
1847	*
1848	* @remarks No-long-jump zone!!!
1849	*/
1850	static void hmR0SvmExportGuestMsrs(PVMCPU pVCpu, PSVMVMCB pVmcb)
1851	{
1852	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1853	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1854
1855	/* Guest Sysenter MSRs. */
1856	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SYSENTER_MSR_MASK)
1857	{
1858	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SYSENTER_CS_MSR)
1859	pVmcb->guest.u64SysEnterCS = pCtx->SysEnter.cs;
1860
1861	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SYSENTER_EIP_MSR)
1862	pVmcb->guest.u64SysEnterEIP = pCtx->SysEnter.eip;
1863
1864	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SYSENTER_ESP_MSR)
1865	pVmcb->guest.u64SysEnterESP = pCtx->SysEnter.esp;
1866	}
1867
1868	/*
1869	* Guest EFER MSR.
1870	* AMD-V requires guest EFER.SVME to be set. Weird.
1871	* See AMD spec. 15.5.1 "Basic Operation" \| "Canonicalization and Consistency Checks".
1872	*/
1873	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_EFER_MSR)
1874	{
1875	pVmcb->guest.u64EFER = pCtx->msrEFER \| MSR_K6_EFER_SVME;
1876	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1877	}
1878
1879	/* If the guest isn't in 64-bit mode, clear MSR_K6_LME bit, otherwise SVM expects amd64 shadow paging. */
1880	if ( !CPUMIsGuestInLongModeEx(pCtx)
1881	&& (pCtx->msrEFER & MSR_K6_EFER_LME))
1882	{
1883	pVmcb->guest.u64EFER &= ~MSR_K6_EFER_LME;
1884	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1885	}
1886
1887	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SYSCALL_MSRS)
1888	{
1889	pVmcb->guest.u64STAR = pCtx->msrSTAR;
1890	pVmcb->guest.u64LSTAR = pCtx->msrLSTAR;
1891	pVmcb->guest.u64CSTAR = pCtx->msrCSTAR;
1892	pVmcb->guest.u64SFMASK = pCtx->msrSFMASK;
1893	}
1894
1895	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_KERNEL_GS_BASE)
1896	pVmcb->guest.u64KernelGSBase = pCtx->msrKERNELGSBASE;
1897
1898	pVCpu->hm.s.fCtxChanged &= ~( HM_CHANGED_GUEST_SYSENTER_MSR_MASK
1899	\| HM_CHANGED_GUEST_EFER_MSR
1900	\| HM_CHANGED_GUEST_SYSCALL_MSRS
1901	\| HM_CHANGED_GUEST_KERNEL_GS_BASE);
1902
1903	/*
1904	* Setup the PAT MSR (applicable for Nested Paging only).
1905	*
1906	* While guests can modify and see the modified values through the shadow values,
1907	* we shall not honor any guest modifications of this MSR to ensure caching is always
1908	* enabled similar to how we clear CR0.CD and NW bits.
1909	*
1910	* For nested-guests this needs to always be set as well, see @bugref{7243#c109}.
1911	*/
1912	pVmcb->guest.u64PAT = MSR_IA32_CR_PAT_INIT_VAL;
1913
1914	/* Enable the last branch record bit if LBR virtualization is enabled. */
1915	if (pVmcb->ctrl.LbrVirt.n.u1LbrVirt)
1916	pVmcb->guest.u64DBGCTL = MSR_IA32_DEBUGCTL_LBR;
1917	}
1918
1919
1920	/**
1921	* Exports the guest (or nested-guest) debug state into the VMCB and programs
1922	* the necessary intercepts accordingly.
1923	*
1924	* @param pVCpu The cross context virtual CPU structure.
1925	* @param pVmcb Pointer to the VM control block.
1926	*
1927	* @remarks No-long-jump zone!!!
1928	* @remarks Requires EFLAGS to be up-to-date in the VMCB!
1929	*/
1930	static void hmR0SvmExportSharedDebugState(PVMCPU pVCpu, PSVMVMCB pVmcb)
1931	{
1932	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1933
1934	/*
1935	* Anyone single stepping on the host side? If so, we'll have to use the
1936	* trap flag in the guest EFLAGS since AMD-V doesn't have a trap flag on
1937	* the VMM level like the VT-x implementations does.
1938	*/
1939	bool fInterceptMovDRx = false;
1940	bool const fStepping = pVCpu->hm.s.fSingleInstruction \|\| DBGFIsStepping(pVCpu);
1941	if (fStepping)
1942	{
1943	pVCpu->hm.s.fClearTrapFlag = true;
1944	pVmcb->guest.u64RFlags \|= X86_EFL_TF;
1945	fInterceptMovDRx = true; /* Need clean DR6, no guest mess. */
1946	}
1947
1948	if ( fStepping
1949	\|\| (CPUMGetHyperDR7(pVCpu) & X86_DR7_ENABLED_MASK))
1950	{
1951	/*
1952	* Use the combined guest and host DRx values found in the hypervisor
1953	* register set because the debugger has breakpoints active or someone
1954	* is single stepping on the host side.
1955	*
1956	* Note! DBGF expects a clean DR6 state before executing guest code.
1957	*/
1958	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS)
1959	if ( CPUMIsGuestInLongModeEx(pCtx)
1960	&& !CPUMIsHyperDebugStateActivePending(pVCpu))
1961	{
1962	CPUMR0LoadHyperDebugState(pVCpu, false /* include DR6 */);
1963	Assert(!CPUMIsGuestDebugStateActivePending(pVCpu));
1964	Assert(CPUMIsHyperDebugStateActivePending(pVCpu));
1965	}
1966	else
1967	#endif
1968	if (!CPUMIsHyperDebugStateActive(pVCpu))
1969	{
1970	CPUMR0LoadHyperDebugState(pVCpu, false /* include DR6 */);
1971	Assert(!CPUMIsGuestDebugStateActive(pVCpu));
1972	Assert(CPUMIsHyperDebugStateActive(pVCpu));
1973	}
1974
1975	/* Update DR6 & DR7. (The other DRx values are handled by CPUM one way or the other.) */
1976	if ( pVmcb->guest.u64DR6 != X86_DR6_INIT_VAL
1977	\|\| pVmcb->guest.u64DR7 != CPUMGetHyperDR7(pVCpu))
1978	{
1979	pVmcb->guest.u64DR7 = CPUMGetHyperDR7(pVCpu);
1980	pVmcb->guest.u64DR6 = X86_DR6_INIT_VAL;
1981	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
1982	}
1983
1984	/** @todo If we cared, we could optimize to allow the guest to read registers
1985	* with the same values. */
1986	fInterceptMovDRx = true;
1987	pVCpu->hm.s.fUsingHyperDR7 = true;
1988	Log5(("hmR0SvmExportSharedDebugState: Loaded hyper DRx\n"));
1989	}
1990	else
1991	{
1992	/*
1993	* Update DR6, DR7 with the guest values if necessary.
1994	*/
1995	if ( pVmcb->guest.u64DR7 != pCtx->dr[7]
1996	\|\| pVmcb->guest.u64DR6 != pCtx->dr[6])
1997	{
1998	pVmcb->guest.u64DR7 = pCtx->dr[7];
1999	pVmcb->guest.u64DR6 = pCtx->dr[6];
2000	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
2001	}
2002	pVCpu->hm.s.fUsingHyperDR7 = false;
2003
2004	/*
2005	* If the guest has enabled debug registers, we need to load them prior to
2006	* executing guest code so they'll trigger at the right time.
2007	*/
2008	if (pCtx->dr[7] & (X86_DR7_ENABLED_MASK \| X86_DR7_GD)) /** @todo Why GD? */
2009	{
2010	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS)
2011	if ( CPUMIsGuestInLongModeEx(pCtx)
2012	&& !CPUMIsGuestDebugStateActivePending(pVCpu))
2013	{
2014	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
2015	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxArmed);
2016	Assert(!CPUMIsHyperDebugStateActivePending(pVCpu));
2017	Assert(CPUMIsGuestDebugStateActivePending(pVCpu));
2018	}
2019	else
2020	#endif
2021	if (!CPUMIsGuestDebugStateActive(pVCpu))
2022	{
2023	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
2024	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxArmed);
2025	Assert(!CPUMIsHyperDebugStateActive(pVCpu));
2026	Assert(CPUMIsGuestDebugStateActive(pVCpu));
2027	}
2028	Log5(("hmR0SvmExportSharedDebugState: Loaded guest DRx\n"));
2029	}
2030	/*
2031	* If no debugging enabled, we'll lazy load DR0-3. We don't need to
2032	* intercept #DB as DR6 is updated in the VMCB.
2033	*
2034	* Note! If we cared and dared, we could skip intercepting \#DB here.
2035	* However, \#DB shouldn't be performance critical, so we'll play safe
2036	* and keep the code similar to the VT-x code and always intercept it.
2037	*/
2038	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS)
2039	else if ( !CPUMIsGuestDebugStateActivePending(pVCpu)
2040	&& !CPUMIsGuestDebugStateActive(pVCpu))
2041	#else
2042	else if (!CPUMIsGuestDebugStateActive(pVCpu))
2043	#endif
2044	{
2045	fInterceptMovDRx = true;
2046	}
2047	}
2048
2049	Assert(pVmcb->ctrl.u32InterceptXcpt & RT_BIT_32(X86_XCPT_DB));
2050	if (fInterceptMovDRx)
2051	{
2052	if ( pVmcb->ctrl.u16InterceptRdDRx != 0xffff
2053	\|\| pVmcb->ctrl.u16InterceptWrDRx != 0xffff)
2054	{
2055	pVmcb->ctrl.u16InterceptRdDRx = 0xffff;
2056	pVmcb->ctrl.u16InterceptWrDRx = 0xffff;
2057	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2058	}
2059	}
2060	else
2061	{
2062	if ( pVmcb->ctrl.u16InterceptRdDRx
2063	\|\| pVmcb->ctrl.u16InterceptWrDRx)
2064	{
2065	pVmcb->ctrl.u16InterceptRdDRx = 0;
2066	pVmcb->ctrl.u16InterceptWrDRx = 0;
2067	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2068	}
2069	}
2070	Log4Func(("DR6=%#RX64 DR7=%#RX64\n", pCtx->dr[6], pCtx->dr[7]));
2071	}
2072
2073	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
2074	/**
2075	* Exports the nested-guest hardware virtualization state into the nested-guest
2076	* VMCB.
2077	*
2078	* @param pVCpu The cross context virtual CPU structure.
2079	* @param pVmcbNstGst Pointer to the nested-guest VM control block.
2080	*
2081	* @remarks No-long-jump zone!!!
2082	*/
2083	static void hmR0SvmExportGuestHwvirtStateNested(PVMCPU pVCpu, PSVMVMCB pVmcbNstGst)
2084	{
2085	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2086
2087	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_HWVIRT)
2088	{
2089	/*
2090	* Ensure the nested-guest pause-filter counters don't exceed the outer guest values esp.
2091	* since SVM doesn't have a preemption timer.
2092	*
2093	* We do this here rather than in hmR0SvmSetupVmcbNested() as we may have been executing the
2094	* nested-guest in IEM incl. PAUSE instructions which would update the pause-filter counters
2095	* and may continue execution in SVM R0 without a nested-guest #VMEXIT in between.
2096	*/
2097	PVM pVM = pVCpu->CTX_SUFF(pVM);
2098	PSVMVMCBCTRL pVmcbNstGstCtrl = &pVmcbNstGst->ctrl;
2099	uint16_t const uGuestPauseFilterCount = pVM->hm.s.svm.cPauseFilter;
2100	uint16_t const uGuestPauseFilterThreshold = pVM->hm.s.svm.cPauseFilterThresholdTicks;
2101	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_PAUSE))
2102	{
2103	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2104	pVmcbNstGstCtrl->u16PauseFilterCount = RT_MIN(pCtx->hwvirt.svm.cPauseFilter, uGuestPauseFilterCount);
2105	pVmcbNstGstCtrl->u16PauseFilterThreshold = RT_MIN(pCtx->hwvirt.svm.cPauseFilterThreshold, uGuestPauseFilterThreshold);
2106	pVmcbNstGstCtrl->u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2107	}
2108	else
2109	{
2110	pVmcbNstGstCtrl->u16PauseFilterCount = uGuestPauseFilterCount;
2111	pVmcbNstGstCtrl->u16PauseFilterThreshold = uGuestPauseFilterThreshold;
2112	}
2113
2114	pVCpu->hm.s.fCtxChanged &= ~HM_CHANGED_GUEST_HWVIRT;
2115	}
2116	}
2117	#endif
2118
2119	/**
2120	* Exports the guest APIC TPR state into the VMCB.
2121	*
2122	* @returns VBox status code.
2123	* @param pVCpu The cross context virtual CPU structure.
2124	* @param pVmcb Pointer to the VM control block.
2125	*/
2126	static int hmR0SvmExportGuestApicTpr(PVMCPU pVCpu, PSVMVMCB pVmcb)
2127	{
2128	if (ASMAtomicUoReadU64(&pVCpu->hm.s.fCtxChanged) & HM_CHANGED_GUEST_APIC_TPR)
2129	{
2130	PVM pVM = pVCpu->CTX_SUFF(pVM);
2131	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2132	if ( PDMHasApic(pVM)
2133	&& APICIsEnabled(pVCpu))
2134	{
2135	bool fPendingIntr;
2136	uint8_t u8Tpr;
2137	int rc = APICGetTpr(pVCpu, &u8Tpr, &fPendingIntr, NULL /* pu8PendingIrq */);
2138	AssertRCReturn(rc, rc);
2139
2140	/* Assume that we need to trap all TPR accesses and thus need not check on
2141	every #VMEXIT if we should update the TPR. */
2142	Assert(pVmcb->ctrl.IntCtrl.n.u1VIntrMasking);
2143	pVCpu->hm.s.svm.fSyncVTpr = false;
2144
2145	if (!pVM->hm.s.fTPRPatchingActive)
2146	{
2147	/* Bits 3-0 of the VTPR field correspond to bits 7-4 of the TPR (which is the Task-Priority Class). */
2148	pVmcb->ctrl.IntCtrl.n.u8VTPR = (u8Tpr >> 4);
2149
2150	/* If there are interrupts pending, intercept CR8 writes to evaluate ASAP if we
2151	can deliver the interrupt to the guest. */
2152	if (fPendingIntr)
2153	pVmcb->ctrl.u16InterceptWrCRx \|= RT_BIT(8);
2154	else
2155	{
2156	pVmcb->ctrl.u16InterceptWrCRx &= ~RT_BIT(8);
2157	pVCpu->hm.s.svm.fSyncVTpr = true;
2158	}
2159
2160	pVmcb->ctrl.u32VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS \| HMSVM_VMCB_CLEAN_INT_CTRL);
2161	}
2162	else
2163	{
2164	/* 32-bit guests uses LSTAR MSR for patching guest code which touches the TPR. */
2165	pVmcb->guest.u64LSTAR = u8Tpr;
2166	uint8_t pbMsrBitmap = (uint8_t )pVCpu->hm.s.svm.pvMsrBitmap;
2167
2168	/* If there are interrupts pending, intercept LSTAR writes, otherwise don't intercept reads or writes. */
2169	if (fPendingIntr)
2170	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_INTERCEPT_WRITE);
2171	else
2172	{
2173	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
2174	pVCpu->hm.s.svm.fSyncVTpr = true;
2175	}
2176	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_IOPM_MSRPM;
2177	}
2178	}
2179	ASMAtomicUoAndU64(&pVCpu->hm.s.fCtxChanged, ~HM_CHANGED_GUEST_APIC_TPR);
2180	}
2181	return VINF_SUCCESS;
2182	}
2183
2184
2185	/**
2186	* Sets up the exception interrupts required for guest (or nested-guest)
2187	* execution in the VMCB.
2188	*
2189	* @param pVCpu The cross context virtual CPU structure.
2190	* @param pVmcb Pointer to the VM control block.
2191	*
2192	* @remarks No-long-jump zone!!!
2193	*/
2194	static void hmR0SvmExportGuestXcptIntercepts(PVMCPU pVCpu, PSVMVMCB pVmcb)
2195	{
2196	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2197
2198	/* If we modify intercepts from here, please check & adjust hmR0SvmMergeVmcbCtrlsNested() if required. */
2199	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_SVM_GUEST_XCPT_INTERCEPTS)
2200	{
2201	/* Trap #UD for GIM provider (e.g. for hypercalls). */
2202	if (pVCpu->hm.s.fGIMTrapXcptUD)
2203	hmR0SvmSetXcptIntercept(pVmcb, X86_XCPT_UD);
2204	else
2205	hmR0SvmClearXcptIntercept(pVCpu, pVmcb, X86_XCPT_UD);
2206
2207	/* Trap #BP for INT3 debug breakpoints set by the VM debugger. */
2208	if (pVCpu->CTX_SUFF(pVM)->dbgf.ro.cEnabledInt3Breakpoints)
2209	hmR0SvmSetXcptIntercept(pVmcb, X86_XCPT_BP);
2210	else
2211	hmR0SvmClearXcptIntercept(pVCpu, pVmcb, X86_XCPT_BP);
2212
2213	/* The remaining intercepts are handled elsewhere, e.g. in hmR0SvmExportGuestCR0(). */
2214	pVCpu->hm.s.fCtxChanged &= ~HM_CHANGED_SVM_GUEST_XCPT_INTERCEPTS;
2215	}
2216	}
2217
2218
2219	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
2220	/**
2221	* Merges guest and nested-guest intercepts for executing the nested-guest using
2222	* hardware-assisted SVM.
2223	*
2224	* This merges the guest and nested-guest intercepts in a way that if the outer
2225	* guest intercept is set we need to intercept it in the nested-guest as
2226	* well.
2227	*
2228	* @param pVCpu The cross context virtual CPU structure.
2229	* @param pVmcbNstGst Pointer to the nested-guest VM control block.
2230	* @param pCtx Pointer to the nested-guest-CPU context.
2231	*/
2232	static void hmR0SvmMergeVmcbCtrlsNested(PVMCPU pVCpu, PCCPUMCTX pCtx)
2233	{
2234	PVM pVM = pVCpu->CTX_SUFF(pVM);
2235	PCSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
2236	PSVMVMCB pVmcbNstGst = pCtx->hwvirt.svm.CTX_SUFF(pVmcb);
2237	PSVMVMCBCTRL pVmcbNstGstCtrl = &pVmcbNstGst->ctrl;
2238
2239	/* Merge the guest's CR intercepts into the nested-guest VMCB. */
2240	pVmcbNstGstCtrl->u16InterceptRdCRx \|= pVmcb->ctrl.u16InterceptRdCRx;
2241	pVmcbNstGstCtrl->u16InterceptWrCRx \|= pVmcb->ctrl.u16InterceptWrCRx;
2242
2243	/* Always intercept CR4 writes for tracking PGM mode changes. */
2244	pVmcbNstGstCtrl->u16InterceptWrCRx \|= RT_BIT(4);
2245
2246	/* Without nested paging, intercept CR3 reads and writes as we load shadow page tables. */
2247	if (!pVM->hm.s.fNestedPaging)
2248	{
2249	pVmcbNstGstCtrl->u16InterceptRdCRx \|= RT_BIT(3);
2250	pVmcbNstGstCtrl->u16InterceptWrCRx \|= RT_BIT(3);
2251	}
2252
2253	/** @todo Figure out debugging with nested-guests, till then just intercept
2254	* all DR[0-15] accesses. */
2255	pVmcbNstGstCtrl->u16InterceptRdDRx \|= 0xffff;
2256	pVmcbNstGstCtrl->u16InterceptWrDRx \|= 0xffff;
2257
2258	/*
2259	* Merge the guest's exception intercepts into the nested-guest VMCB.
2260	*
2261	* - \#UD: Exclude these as the outer guest's GIM hypercalls are not applicable
2262	* while executing the nested-guest.
2263	*
2264	* - \#BP: Exclude breakpoints set by the VM debugger for the outer guest. This can
2265	* be tweaked later depending on how we wish to implement breakpoints.
2266	*
2267	* Warning!! This ASSUMES we only intercept \#UD for hypercall purposes and \#BP
2268	* for VM debugger breakpoints, see hmR0SvmExportGuestXcptIntercepts().
2269	*/
2270	#ifndef HMSVM_ALWAYS_TRAP_ALL_XCPTS
2271	pVmcbNstGstCtrl->u32InterceptXcpt \|= (pVmcb->ctrl.u32InterceptXcpt & ~( RT_BIT(X86_XCPT_UD)
2272	\| RT_BIT(X86_XCPT_BP)));
2273	#else
2274	pVmcbNstGstCtrl->u32InterceptXcpt \|= pVmcb->ctrl.u32InterceptXcpt;
2275	#endif
2276
2277	/*
2278	* Adjust intercepts while executing the nested-guest that differ from the
2279	* outer guest intercepts.
2280	*
2281	* - VINTR: Exclude the outer guest intercept as we don't need to cause VINTR #VMEXITs
2282	* that belong to the nested-guest to the outer guest.
2283	*
2284	* - VMMCALL: Exclude the outer guest intercept as when it's also not intercepted by
2285	* the nested-guest, the physical CPU raises a \#UD exception as expected.
2286	*/
2287	pVmcbNstGstCtrl->u64InterceptCtrl \|= (pVmcb->ctrl.u64InterceptCtrl & ~( SVM_CTRL_INTERCEPT_VINTR
2288	\| SVM_CTRL_INTERCEPT_VMMCALL))
2289	\| HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS;
2290
2291	Assert( (pVmcbNstGstCtrl->u64InterceptCtrl & HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS)
2292	== HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS);
2293
2294	/* Finally, update the VMCB clean bits. */
2295	pVmcbNstGstCtrl->u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2296	}
2297	#endif
2298
2299
2300	/**
2301	* Selects the appropriate function to run guest code.
2302	*
2303	* @returns VBox status code.
2304	* @param pVCpu The cross context virtual CPU structure.
2305	*
2306	* @remarks No-long-jump zone!!!
2307	*/
2308	static int hmR0SvmSelectVMRunHandler(PVMCPU pVCpu)
2309	{
2310	if (CPUMIsGuestInLongMode(pVCpu))
2311	{
2312	#ifndef VBOX_ENABLE_64_BITS_GUESTS
2313	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
2314	#endif
2315	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.fAllow64BitGuests); /* Guaranteed by hmR3InitFinalizeR0(). */
2316	#if HC_ARCH_BITS == 32
2317	/* 32-bit host. We need to switch to 64-bit before running the 64-bit guest. */
2318	pVCpu->hm.s.svm.pfnVMRun = SVMR0VMSwitcherRun64;
2319	#else
2320	/* 64-bit host or hybrid host. */
2321	pVCpu->hm.s.svm.pfnVMRun = SVMR0VMRun64;
2322	#endif
2323	}
2324	else
2325	{
2326	/* Guest is not in long mode, use the 32-bit handler. */
2327	pVCpu->hm.s.svm.pfnVMRun = SVMR0VMRun;
2328	}
2329	return VINF_SUCCESS;
2330	}
2331
2332
2333	/**
2334	* Enters the AMD-V session.
2335	*
2336	* @returns VBox status code.
2337	* @param pVM The cross context VM structure.
2338	* @param pVCpu The cross context virtual CPU structure.
2339	* @param pCpu Pointer to the CPU info struct.
2340	*/
2341	VMMR0DECL(int) SVMR0Enter(PVM pVM, PVMCPU pVCpu, PHMGLOBALCPUINFO pCpu)
2342	{
2343	AssertPtr(pVM);
2344	AssertPtr(pVCpu);
2345	Assert(pVM->hm.s.svm.fSupported);
2346	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2347	NOREF(pVM); NOREF(pCpu);
2348
2349	LogFlowFunc(("pVM=%p pVCpu=%p\n", pVM, pVCpu));
2350	Assert((pVCpu->hm.s.fCtxChanged & (HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE))
2351	== (HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE));
2352
2353	pVCpu->hm.s.fLeaveDone = false;
2354	return VINF_SUCCESS;
2355	}
2356
2357
2358	/**
2359	* Thread-context callback for AMD-V.
2360	*
2361	* @param enmEvent The thread-context event.
2362	* @param pVCpu The cross context virtual CPU structure.
2363	* @param fGlobalInit Whether global VT-x/AMD-V init. is used.
2364	* @thread EMT(pVCpu)
2365	*/
2366	VMMR0DECL(void) SVMR0ThreadCtxCallback(RTTHREADCTXEVENT enmEvent, PVMCPU pVCpu, bool fGlobalInit)
2367	{
2368	NOREF(fGlobalInit);
2369
2370	switch (enmEvent)
2371	{
2372	case RTTHREADCTXEVENT_OUT:
2373	{
2374	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2375	Assert(VMMR0ThreadCtxHookIsEnabled(pVCpu));
2376	VMCPU_ASSERT_EMT(pVCpu);
2377
2378	/* No longjmps (log-flush, locks) in this fragile context. */
2379	VMMRZCallRing3Disable(pVCpu);
2380
2381	if (!pVCpu->hm.s.fLeaveDone)
2382	{
2383	hmR0SvmLeave(pVCpu, false /* fImportState */);
2384	pVCpu->hm.s.fLeaveDone = true;
2385	}
2386
2387	/* Leave HM context, takes care of local init (term). */
2388	int rc = HMR0LeaveCpu(pVCpu);
2389	AssertRC(rc); NOREF(rc);
2390
2391	/* Restore longjmp state. */
2392	VMMRZCallRing3Enable(pVCpu);
2393	STAM_REL_COUNTER_INC(&pVCpu->hm.s.StatSwitchPreempt);
2394	break;
2395	}
2396
2397	case RTTHREADCTXEVENT_IN:
2398	{
2399	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2400	Assert(VMMR0ThreadCtxHookIsEnabled(pVCpu));
2401	VMCPU_ASSERT_EMT(pVCpu);
2402
2403	/* No longjmps (log-flush, locks) in this fragile context. */
2404	VMMRZCallRing3Disable(pVCpu);
2405
2406	/*
2407	* Initialize the bare minimum state required for HM. This takes care of
2408	* initializing AMD-V if necessary (onlined CPUs, local init etc.)
2409	*/
2410	int rc = HMR0EnterCpu(pVCpu);
2411	AssertRC(rc); NOREF(rc);
2412	Assert((pVCpu->hm.s.fCtxChanged & (HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE))
2413	== (HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE));
2414
2415	pVCpu->hm.s.fLeaveDone = false;
2416
2417	/* Restore longjmp state. */
2418	VMMRZCallRing3Enable(pVCpu);
2419	break;
2420	}
2421
2422	default:
2423	break;
2424	}
2425	}
2426
2427
2428	/**
2429	* Saves the host state.
2430	*
2431	* @returns VBox status code.
2432	* @param pVM The cross context VM structure.
2433	* @param pVCpu The cross context virtual CPU structure.
2434	*
2435	* @remarks No-long-jump zone!!!
2436	*/
2437	VMMR0DECL(int) SVMR0ExportHostState(PVMCPU pVCpu)
2438	{
2439	NOREF(pVCpu);
2440
2441	/* Nothing to do here. AMD-V does this for us automatically during the world-switch. */
2442	ASMAtomicUoAndU64(&pVCpu->hm.s.fCtxChanged, ~HM_CHANGED_HOST_CONTEXT);
2443	return VINF_SUCCESS;
2444	}
2445
2446
2447	/**
2448	* Exports the guest state from the guest-CPU context into the VMCB.
2449	*
2450	* The CPU state will be loaded from these fields on every successful VM-entry.
2451	* Also sets up the appropriate VMRUN function to execute guest code based on
2452	* the guest CPU mode.
2453	*
2454	* @returns VBox status code.
2455	* @param pVM The cross context VM structure.
2456	* @param pVCpu The cross context virtual CPU structure.
2457	* @param pCtx Pointer to the guest-CPU context.
2458	*
2459	* @remarks No-long-jump zone!!!
2460	*/
2461	static int hmR0SvmExportGuestState(PVMCPU pVCpu)
2462	{
2463	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatExportGuestState, x);
2464
2465	PSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
2466	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2467
2468	Assert(pVmcb);
2469	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(pCtx);
2470
2471	pVmcb->guest.u64RIP = pCtx->rip;
2472	pVmcb->guest.u64RSP = pCtx->rsp;
2473	pVmcb->guest.u64RFlags = pCtx->eflags.u32;
2474	pVmcb->guest.u64RAX = pCtx->rax;
2475	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
2476	if (pVmcb->ctrl.IntCtrl.n.u1VGifEnable)
2477	{
2478	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_VGIF);
2479	pVmcb->ctrl.IntCtrl.n.u1VGif = pCtx->hwvirt.fGif;
2480	}
2481	#endif
2482
2483	RTCCUINTREG const fEFlags = ASMIntDisableFlags();
2484
2485	int rc = hmR0SvmExportGuestControlRegs(pVCpu, pVmcb);
2486	AssertRCReturnStmt(rc, ASMSetFlags(fEFlags), rc);
2487
2488	hmR0SvmExportGuestSegmentRegs(pVCpu, pVmcb);
2489	hmR0SvmExportGuestMsrs(pVCpu, pVmcb);
2490	hmR0SvmExportGuestXcptIntercepts(pVCpu, pVmcb);
2491
2492	ASMSetFlags(fEFlags);
2493
2494	/* hmR0SvmExportGuestApicTpr() must be called -after- hmR0SvmExportGuestMsrs() as we
2495	otherwise we would overwrite the LSTAR MSR that we use for TPR patching. */
2496	hmR0SvmExportGuestApicTpr(pVCpu, pVmcb);
2497
2498	rc = hmR0SvmSelectVMRunHandler(pVCpu);
2499	AssertRCReturn(rc, rc);
2500
2501	/* Clear any bits that may be set but exported unconditionally or unused/reserved bits. */
2502	ASMAtomicUoAndU64(&pVCpu->hm.s.fCtxChanged, ~( HM_CHANGED_GUEST_RIP
2503	\| HM_CHANGED_GUEST_RFLAGS
2504	\| HM_CHANGED_GUEST_GPRS_MASK
2505	\| HM_CHANGED_GUEST_X87
2506	\| HM_CHANGED_GUEST_SSE_AVX
2507	\| HM_CHANGED_GUEST_OTHER_XSAVE
2508	\| HM_CHANGED_GUEST_XCRx
2509	\| HM_CHANGED_GUEST_TSC_AUX
2510	\| HM_CHANGED_GUEST_OTHER_MSRS
2511	\| HM_CHANGED_GUEST_HWVIRT
2512	\| (HM_CHANGED_KEEPER_STATE_MASK & ~HM_CHANGED_SVM_GUEST_XCPT_INTERCEPTS)));
2513
2514	#ifdef VBOX_STRICT
2515	/*
2516	* All of the guest-CPU state and SVM keeper bits should be exported here by now,
2517	* except for the host-context and/or shared host-guest context bits.
2518	*/
2519	uint64_t const fCtxChanged = ASMAtomicUoReadU64(&pVCpu->hm.s.fCtxChanged);
2520	AssertMsg(!(fCtxChanged & (HM_CHANGED_ALL_GUEST & ~HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE)),
2521	("fCtxChanged=%#RX64\n", fCtxChanged));
2522
2523	/*
2524	* If we need to log state that isn't always imported, we'll need to import them here.
2525	* See hmR0SvmPostRunGuest() for which part of the state is imported uncondtionally.
2526	*/
2527	hmR0SvmLogState(pVCpu, pVmcb, "hmR0SvmExportGuestState", 0 /* fFlags /, 0 / uVerbose */);
2528	#endif
2529
2530	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExportGuestState, x);
2531	return VINF_SUCCESS;
2532	}
2533
2534
2535	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
2536	/**
2537	* Merges the guest and nested-guest MSR permission bitmap.
2538	*
2539	* If the guest is intercepting an MSR we need to intercept it regardless of
2540	* whether the nested-guest is intercepting it or not.
2541	*
2542	* @param pHostCpu Pointer to the physical CPU HM info. struct.
2543	* @param pVCpu The cross context virtual CPU structure.
2544	* @param pCtx Pointer to the nested-guest-CPU context.
2545	*/
2546	static void hmR0SvmMergeMsrpmNested(PHMGLOBALCPUINFO pHostCpu, PVMCPU pVCpu, PCCPUMCTX pCtx)
2547	{
2548	uint64_t const pu64GstMsrpm = (uint64_t const )pVCpu->hm.s.svm.pvMsrBitmap;
2549	uint64_t const pu64NstGstMsrpm = (uint64_t const )pCtx->hwvirt.svm.CTX_SUFF(pvMsrBitmap);
2550	uint64_t pu64DstMsrpm = (uint64_t )pHostCpu->n.svm.pvNstGstMsrpm;
2551
2552	/* MSRPM bytes from offset 0x1800 are reserved, so we stop merging there. */
2553	uint32_t const offRsvdQwords = 0x1800 >> 3;
2554	for (uint32_t i = 0; i < offRsvdQwords; i++)
2555	pu64DstMsrpm[i] = pu64NstGstMsrpm[i] \| pu64GstMsrpm[i];
2556	}
2557
2558
2559	/**
2560	* Caches the nested-guest VMCB fields before we modify them for execution using
2561	* hardware-assisted SVM.
2562	*
2563	* @returns true if the VMCB was previously already cached, false otherwise.
2564	* @param pCtx Pointer to the guest-CPU context.
2565	*
2566	* @sa HMSvmNstGstVmExitNotify.
2567	*/
2568	static bool hmR0SvmCacheVmcbNested(PVMCPU pVCpu, PCCPUMCTX pCtx)
2569	{
2570	/*
2571	* Cache the nested-guest programmed VMCB fields if we have not cached it yet.
2572	* Otherwise we risk re-caching the values we may have modified, see @bugref{7243#c44}.
2573	*
2574	* Nested-paging CR3 is not saved back into the VMCB on #VMEXIT, hence no need to
2575	* cache and restore it, see AMD spec. 15.25.4 "Nested Paging and VMRUN/#VMEXIT".
2576	*/
2577	PSVMNESTEDVMCBCACHE pVmcbNstGstCache = &pVCpu->hm.s.svm.NstGstVmcbCache;
2578	bool const fWasCached = pVmcbNstGstCache->fCacheValid;
2579	if (!fWasCached)
2580	{
2581	PCSVMVMCB pVmcbNstGst = pCtx->hwvirt.svm.CTX_SUFF(pVmcb);
2582	PCSVMVMCBCTRL pVmcbNstGstCtrl = &pVmcbNstGst->ctrl;
2583	pVmcbNstGstCache->u16InterceptRdCRx = pVmcbNstGstCtrl->u16InterceptRdCRx;
2584	pVmcbNstGstCache->u16InterceptWrCRx = pVmcbNstGstCtrl->u16InterceptWrCRx;
2585	pVmcbNstGstCache->u16InterceptRdDRx = pVmcbNstGstCtrl->u16InterceptRdDRx;
2586	pVmcbNstGstCache->u16InterceptWrDRx = pVmcbNstGstCtrl->u16InterceptWrDRx;
2587	pVmcbNstGstCache->u16PauseFilterThreshold = pVmcbNstGstCtrl->u16PauseFilterThreshold;
2588	pVmcbNstGstCache->u16PauseFilterCount = pVmcbNstGstCtrl->u16PauseFilterCount;
2589	pVmcbNstGstCache->u32InterceptXcpt = pVmcbNstGstCtrl->u32InterceptXcpt;
2590	pVmcbNstGstCache->u64InterceptCtrl = pVmcbNstGstCtrl->u64InterceptCtrl;
2591	pVmcbNstGstCache->u64TSCOffset = pVmcbNstGstCtrl->u64TSCOffset;
2592	pVmcbNstGstCache->fVIntrMasking = pVmcbNstGstCtrl->IntCtrl.n.u1VIntrMasking;
2593	pVmcbNstGstCache->fNestedPaging = pVmcbNstGstCtrl->NestedPagingCtrl.n.u1NestedPaging;
2594	pVmcbNstGstCache->fLbrVirt = pVmcbNstGstCtrl->LbrVirt.n.u1LbrVirt;
2595	pVmcbNstGstCache->fCacheValid = true;
2596	Log4Func(("Cached VMCB fields\n"));
2597	}
2598
2599	return fWasCached;
2600	}
2601
2602
2603	/**
2604	* Sets up the nested-guest VMCB for execution using hardware-assisted SVM.
2605	*
2606	* This is done the first time we enter nested-guest execution using SVM R0
2607	* until the nested-guest \#VMEXIT (not to be confused with physical CPU
2608	* \#VMEXITs which may or may not cause a corresponding nested-guest \#VMEXIT).
2609	*
2610	* @param pVCpu The cross context virtual CPU structure.
2611	*/
2612	static void hmR0SvmSetupVmcbNested(PVMCPU pVCpu)
2613	{
2614	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2615	PSVMVMCB pVmcbNstGst = pCtx->hwvirt.svm.CTX_SUFF(pVmcb);
2616	PSVMVMCBCTRL pVmcbNstGstCtrl = &pVmcbNstGst->ctrl;
2617
2618	/*
2619	* First cache the nested-guest VMCB fields we may potentially modify.
2620	*/
2621	bool const fVmcbCached = hmR0SvmCacheVmcbNested(pVCpu, pCtx);
2622	if (!fVmcbCached)
2623	{
2624	/*
2625	* The IOPM of the nested-guest can be ignored because the the guest always
2626	* intercepts all IO port accesses. Thus, we'll swap to the guest IOPM rather
2627	* than the nested-guest IOPM and swap the field back on the #VMEXIT.
2628	*/
2629	pVmcbNstGstCtrl->u64IOPMPhysAddr = g_HCPhysIOBitmap;
2630
2631	/*
2632	* Use the same nested-paging as the outer guest. We can't dynamically switch off
2633	* nested-paging suddenly while executing a VM (see assertion at the end of
2634	* Trap0eHandler() in PGMAllBth.h).
2635	*/
2636	pVmcbNstGstCtrl->NestedPagingCtrl.n.u1NestedPaging = pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging;
2637
2638	/* Always enable V_INTR_MASKING as we do not want to allow access to the physical APIC TPR. */
2639	pVmcbNstGstCtrl->IntCtrl.n.u1VIntrMasking = 1;
2640
2641	/*
2642	* Turn off TPR syncing on #VMEXIT for nested-guests as CR8 intercepts are subject
2643	* to the nested-guest intercepts and we always run with V_INTR_MASKING.
2644	*/
2645	pVCpu->hm.s.svm.fSyncVTpr = false;
2646
2647	#ifdef DEBUG_ramshankar
2648	/* For debugging purposes - copy the LBR info. from outer guest VMCB. */
2649	pVmcbNstGstCtrl->LbrVirt.n.u1LbrVirt = pVmcb->ctrl.LbrVirt.n.u1LbrVirt;
2650	#endif
2651
2652	/*
2653	* If we don't expose Virtualized-VMSAVE/VMLOAD feature to the outer guest, we
2654	* need to intercept VMSAVE/VMLOAD instructions executed by the nested-guest.
2655	*/
2656	if (!pVCpu->CTX_SUFF(pVM)->cpum.ro.GuestFeatures.fSvmVirtVmsaveVmload)
2657	pVmcbNstGstCtrl->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_VMSAVE
2658	\| SVM_CTRL_INTERCEPT_VMLOAD;
2659
2660	/*
2661	* If we don't expose Virtual GIF feature to the outer guest, we need to intercept
2662	* CLGI/STGI instructions executed by the nested-guest.
2663	*/
2664	if (!pVCpu->CTX_SUFF(pVM)->cpum.ro.GuestFeatures.fSvmVGif)
2665	pVmcbNstGstCtrl->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_CLGI
2666	\| SVM_CTRL_INTERCEPT_STGI;
2667
2668	/* Merge the guest and nested-guest intercepts. */
2669	hmR0SvmMergeVmcbCtrlsNested(pVCpu, pCtx);
2670
2671	/* Update the VMCB clean bits. */
2672	pVmcbNstGstCtrl->u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2673	}
2674	else
2675	{
2676	Assert(!pVCpu->hm.s.svm.fSyncVTpr);
2677	Assert(pVmcbNstGstCtrl->u64IOPMPhysAddr == g_HCPhysIOBitmap);
2678	Assert(RT_BOOL(pVmcbNstGstCtrl->NestedPagingCtrl.n.u1NestedPaging) == pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
2679	}
2680	}
2681
2682
2683	/**
2684	* Exports the nested-guest state into the VMCB.
2685	*
2686	* We need to export the entire state as we could be continuing nested-guest
2687	* execution at any point (not just immediately after VMRUN) and thus the VMCB
2688	* can be out-of-sync with the nested-guest state if it was executed in IEM.
2689	*
2690	* @returns VBox status code.
2691	* @param pVCpu The cross context virtual CPU structure.
2692	* @param pCtx Pointer to the guest-CPU context.
2693	*
2694	* @remarks No-long-jump zone!!!
2695	*/
2696	static int hmR0SvmExportGuestStateNested(PVMCPU pVCpu)
2697	{
2698	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatExportGuestState, x);
2699
2700	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2701	PSVMVMCB pVmcbNstGst = pCtx->hwvirt.svm.CTX_SUFF(pVmcb);
2702	Assert(pVmcbNstGst);
2703
2704	hmR0SvmSetupVmcbNested(pVCpu);
2705
2706	pVmcbNstGst->guest.u64RIP = pCtx->rip;
2707	pVmcbNstGst->guest.u64RSP = pCtx->rsp;
2708	pVmcbNstGst->guest.u64RFlags = pCtx->eflags.u32;
2709	pVmcbNstGst->guest.u64RAX = pCtx->rax;
2710
2711	RTCCUINTREG const fEFlags = ASMIntDisableFlags();
2712
2713	int rc = hmR0SvmExportGuestControlRegs(pVCpu, pVmcbNstGst);
2714	AssertRCReturnStmt(rc, ASMSetFlags(fEFlags), rc);
2715
2716	hmR0SvmExportGuestSegmentRegs(pVCpu, pVmcbNstGst);
2717	hmR0SvmExportGuestMsrs(pVCpu, pVmcbNstGst);
2718	hmR0SvmExportGuestHwvirtStateNested(pVCpu, pVmcbNstGst);
2719
2720	ASMSetFlags(fEFlags);
2721
2722	/* Nested VGIF not supported yet. */
2723	Assert(!pVmcbNstGst->ctrl.IntCtrl.n.u1VGifEnable);
2724
2725	rc = hmR0SvmSelectVMRunHandler(pVCpu);
2726	AssertRCReturn(rc, rc);
2727
2728	/* Clear any bits that may be set but exported unconditionally or unused/reserved bits. */
2729	ASMAtomicUoAndU64(&pVCpu->hm.s.fCtxChanged, ~( HM_CHANGED_GUEST_RIP
2730	\| HM_CHANGED_GUEST_RFLAGS
2731	\| HM_CHANGED_GUEST_GPRS_MASK
2732	\| HM_CHANGED_GUEST_APIC_TPR
2733	\| HM_CHANGED_GUEST_X87
2734	\| HM_CHANGED_GUEST_SSE_AVX
2735	\| HM_CHANGED_GUEST_OTHER_XSAVE
2736	\| HM_CHANGED_GUEST_XCRx
2737	\| HM_CHANGED_GUEST_TSC_AUX
2738	\| HM_CHANGED_GUEST_OTHER_MSRS
2739	\| HM_CHANGED_SVM_GUEST_XCPT_INTERCEPTS
2740	\| (HM_CHANGED_KEEPER_STATE_MASK & ~HM_CHANGED_SVM_MASK)));
2741
2742	#ifdef VBOX_STRICT
2743	/*
2744	* All of the guest-CPU state and SVM keeper bits should be exported here by now, except
2745	* for the host-context and/or shared host-guest context bits.
2746	*/
2747	uint64_t const fCtxChanged = ASMAtomicUoReadU64(&pVCpu->hm.s.fCtxChanged);
2748	AssertMsg(!(fCtxChanged & (HM_CHANGED_ALL_GUEST & ~HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE)),
2749	("fCtxChanged=%#RX64\n", fCtxChanged));
2750
2751	/*
2752	* If we need to log state that isn't always imported, we'll need to import them here.
2753	* See hmR0SvmPostRunGuest() for which part of the state is imported uncondtionally.
2754	*/
2755	hmR0SvmLogState(pVCpu, pVmcbNstGst, "hmR0SvmExportGuestStateNested", 0 /* fFlags /, 0 / uVerbose */);
2756	#endif
2757
2758	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExportGuestState, x);
2759	return rc;
2760	}
2761	#endif /* VBOX_WITH_NESTED_HWVIRT_SVM */
2762
2763
2764	/**
2765	* Exports the state shared between the host and guest (or nested-guest) into
2766	* the VMCB.
2767	*
2768	* @param pVCpu The cross context virtual CPU structure.
2769	* @param pVmcb Pointer to the VM control block.
2770	*
2771	* @remarks No-long-jump zone!!!
2772	*/
2773	static void hmR0SvmExportSharedState(PVMCPU pVCpu, PSVMVMCB pVmcb)
2774	{
2775	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2776	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2777
2778	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_DR_MASK)
2779	{
2780	/** @todo Figure out stepping with nested-guest. */
2781	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2782	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
2783	hmR0SvmExportSharedDebugState(pVCpu, pVmcb);
2784	else
2785	{
2786	pVmcb->guest.u64DR6 = pCtx->dr[6];
2787	pVmcb->guest.u64DR7 = pCtx->dr[7];
2788	}
2789	}
2790
2791	pVCpu->hm.s.fCtxChanged &= ~HM_CHANGED_GUEST_DR_MASK;
2792	AssertMsg(!(pVCpu->hm.s.fCtxChanged & HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE),
2793	("fCtxChanged=%#RX64\n", pVCpu->hm.s.fCtxChanged));
2794	}
2795
2796
2797	/**
2798	* Worker for SVMR0ImportStateOnDemand.
2799	*
2800	* @param pVCpu The cross context virtual CPU structure.
2801	* @param pCtx Pointer to the guest-CPU or nested-guest-CPU context.
2802	* @param fWhat What to import, CPUMCTX_EXTRN_XXX.
2803	*/
2804	static void hmR0SvmImportGuestState(PVMCPU pVCpu, uint64_t fWhat)
2805	{
2806	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatImportGuestState, x);
2807
2808	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2809	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
2810	PCSVMVMCBSTATESAVE pVmcbGuest = &pVmcb->guest;
2811	PCSVMVMCBCTRL pVmcbCtrl = &pVmcb->ctrl;
2812
2813	Log4Func(("fExtrn=%#RX64 fWhat=%#RX64\n", pCtx->fExtrn, fWhat));
2814
2815	/*
2816	* We disable interrupts to make the updating of the state and in particular
2817	* the fExtrn modification atomic wrt to preemption hooks.
2818	*/
2819	RTCCUINTREG const fEFlags = ASMIntDisableFlags();
2820
2821	fWhat &= pCtx->fExtrn;
2822	if (fWhat & pCtx->fExtrn)
2823	{
2824	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
2825	if (fWhat & CPUMCTX_EXTRN_HWVIRT)
2826	{
2827	if ( !CPUMIsGuestInSvmNestedHwVirtMode(pCtx)
2828	&& pVmcbCtrl->IntCtrl.n.u1VGifEnable)
2829	{
2830	/* We don't yet support passing VGIF feature to the guest. */
2831	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.svm.fVGif);
2832	pCtx->hwvirt.fGif = pVmcbCtrl->IntCtrl.n.u1VGif;
2833	}
2834	}
2835
2836	if (fWhat & CPUMCTX_EXTRN_HM_SVM_HWVIRT_VIRQ)
2837	{
2838	if ( !pVmcbCtrl->IntCtrl.n.u1VIrqPending
2839	&& VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_NESTED_GUEST))
2840	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INTERRUPT_NESTED_GUEST);
2841	}
2842	#endif
2843
2844	if (fWhat & CPUMCTX_EXTRN_HM_SVM_INT_SHADOW)
2845	{
2846	if (pVmcbCtrl->IntShadow.n.u1IntShadow)
2847	EMSetInhibitInterruptsPC(pVCpu, pVmcbGuest->u64RIP);
2848	else if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS))
2849	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS);
2850	}
2851
2852	if (fWhat & CPUMCTX_EXTRN_RIP)
2853	pCtx->rip = pVmcbGuest->u64RIP;
2854
2855	if (fWhat & CPUMCTX_EXTRN_RFLAGS)
2856	pCtx->eflags.u32 = pVmcbGuest->u64RFlags;
2857
2858	if (fWhat & CPUMCTX_EXTRN_RSP)
2859	pCtx->rsp = pVmcbGuest->u64RSP;
2860
2861	if (fWhat & CPUMCTX_EXTRN_RAX)
2862	pCtx->rax = pVmcbGuest->u64RAX;
2863
2864	if (fWhat & CPUMCTX_EXTRN_SREG_MASK)
2865	{
2866	if (fWhat & CPUMCTX_EXTRN_CS)
2867	{
2868	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, CS, cs);
2869	/* Correct the CS granularity bit. Haven't seen it being wrong in any other register (yet). */
2870	/** @todo SELM might need to be fixed as it too should not care about the
2871	* granularity bit. See @bugref{6785}. */
2872	if ( !pCtx->cs.Attr.n.u1Granularity
2873	&& pCtx->cs.Attr.n.u1Present
2874	&& pCtx->cs.u32Limit > UINT32_C(0xfffff))
2875	{
2876	Assert((pCtx->cs.u32Limit & 0xfff) == 0xfff);
2877	pCtx->cs.Attr.n.u1Granularity = 1;
2878	}
2879	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, cs);
2880	}
2881	if (fWhat & CPUMCTX_EXTRN_SS)
2882	{
2883	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, SS, ss);
2884	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, ss);
2885	/*
2886	* Sync the hidden SS DPL field. AMD CPUs have a separate CPL field in the
2887	* VMCB and uses that and thus it's possible that when the CPL changes during
2888	* guest execution that the SS DPL isn't updated by AMD-V. Observed on some
2889	* AMD Fusion CPUs with 64-bit guests.
2890	*
2891	* See AMD spec. 15.5.1 "Basic operation".
2892	*/
2893	Assert(!(pVmcbGuest->u8CPL & ~0x3));
2894	uint8_t const uCpl = pVmcbGuest->u8CPL;
2895	if (pCtx->ss.Attr.n.u2Dpl != uCpl)
2896	pCtx->ss.Attr.n.u2Dpl = uCpl & 0x3;
2897	}
2898	if (fWhat & CPUMCTX_EXTRN_DS)
2899	{
2900	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, DS, ds);
2901	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, ds);
2902	}
2903	if (fWhat & CPUMCTX_EXTRN_ES)
2904	{
2905	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, ES, es);
2906	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, es);
2907	}
2908	if (fWhat & CPUMCTX_EXTRN_FS)
2909	{
2910	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, FS, fs);
2911	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, fs);
2912	}
2913	if (fWhat & CPUMCTX_EXTRN_GS)
2914	{
2915	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, GS, gs);
2916	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, gs);
2917	}
2918	}
2919
2920	if (fWhat & CPUMCTX_EXTRN_TABLE_MASK)
2921	{
2922	if (fWhat & CPUMCTX_EXTRN_TR)
2923	{
2924	/*
2925	* Fixup TR attributes so it's compatible with Intel. Important when saved-states
2926	* are used between Intel and AMD, see @bugref{6208#c39}.
2927	* ASSUME that it's normally correct and that we're in 32-bit or 64-bit mode.
2928	*/
2929	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, TR, tr);
2930	if (pCtx->tr.Attr.n.u4Type != X86_SEL_TYPE_SYS_386_TSS_BUSY)
2931	{
2932	if ( pCtx->tr.Attr.n.u4Type == X86_SEL_TYPE_SYS_386_TSS_AVAIL
2933	\|\| CPUMIsGuestInLongModeEx(pCtx))
2934	pCtx->tr.Attr.n.u4Type = X86_SEL_TYPE_SYS_386_TSS_BUSY;
2935	else if (pCtx->tr.Attr.n.u4Type == X86_SEL_TYPE_SYS_286_TSS_AVAIL)
2936	pCtx->tr.Attr.n.u4Type = X86_SEL_TYPE_SYS_286_TSS_BUSY;
2937	}
2938	}
2939
2940	if (fWhat & CPUMCTX_EXTRN_LDTR)
2941	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, LDTR, ldtr);
2942
2943	if (fWhat & CPUMCTX_EXTRN_GDTR)
2944	{
2945	pCtx->gdtr.cbGdt = pVmcbGuest->GDTR.u32Limit;
2946	pCtx->gdtr.pGdt = pVmcbGuest->GDTR.u64Base;
2947	}
2948
2949	if (fWhat & CPUMCTX_EXTRN_IDTR)
2950	{
2951	pCtx->idtr.cbIdt = pVmcbGuest->IDTR.u32Limit;
2952	pCtx->idtr.pIdt = pVmcbGuest->IDTR.u64Base;
2953	}
2954	}
2955
2956	if (fWhat & CPUMCTX_EXTRN_SYSCALL_MSRS)
2957	{
2958	pCtx->msrSTAR = pVmcbGuest->u64STAR;
2959	pCtx->msrLSTAR = pVmcbGuest->u64LSTAR;
2960	pCtx->msrCSTAR = pVmcbGuest->u64CSTAR;
2961	pCtx->msrSFMASK = pVmcbGuest->u64SFMASK;
2962	}
2963
2964	if (fWhat & CPUMCTX_EXTRN_SYSENTER_MSRS)
2965	{
2966	pCtx->SysEnter.cs = pVmcbGuest->u64SysEnterCS;
2967	pCtx->SysEnter.eip = pVmcbGuest->u64SysEnterEIP;
2968	pCtx->SysEnter.esp = pVmcbGuest->u64SysEnterESP;
2969	}
2970
2971	if (fWhat & CPUMCTX_EXTRN_KERNEL_GS_BASE)
2972	pCtx->msrKERNELGSBASE = pVmcbGuest->u64KernelGSBase;
2973
2974	if (fWhat & CPUMCTX_EXTRN_DR_MASK)
2975	{
2976	if (fWhat & CPUMCTX_EXTRN_DR6)
2977	{
2978	if (!pVCpu->hm.s.fUsingHyperDR7)
2979	pCtx->dr[6] = pVmcbGuest->u64DR6;
2980	else
2981	CPUMSetHyperDR6(pVCpu, pVmcbGuest->u64DR6);
2982	}
2983
2984	if (fWhat & CPUMCTX_EXTRN_DR7)
2985	{
2986	if (!pVCpu->hm.s.fUsingHyperDR7)
2987	pCtx->dr[7] = pVmcbGuest->u64DR7;
2988	else
2989	Assert(pVmcbGuest->u64DR7 == CPUMGetHyperDR7(pVCpu));
2990	}
2991	}
2992
2993	if (fWhat & CPUMCTX_EXTRN_CR_MASK)
2994	{
2995	if (fWhat & CPUMCTX_EXTRN_CR0)
2996	{
2997	/* We intercept changes to all CR0 bits except maybe TS & MP bits. */
2998	uint64_t const uCr0 = (pCtx->cr0 & ~(X86_CR0_TS \| X86_CR0_MP))
2999	\| (pVmcbGuest->u64CR0 & (X86_CR0_TS \| X86_CR0_MP));
3000	VMMRZCallRing3Disable(pVCpu); /* Calls into PGM which has Log statements. */
3001	CPUMSetGuestCR0(pVCpu, uCr0);
3002	VMMRZCallRing3Enable(pVCpu);
3003	}
3004
3005	if (fWhat & CPUMCTX_EXTRN_CR2)
3006	pCtx->cr2 = pVmcbGuest->u64CR2;
3007
3008	if (fWhat & CPUMCTX_EXTRN_CR3)
3009	{
3010	if ( pVmcbCtrl->NestedPagingCtrl.n.u1NestedPaging
3011	&& pCtx->cr3 != pVmcbGuest->u64CR3)
3012	{
3013	CPUMSetGuestCR3(pVCpu, pVmcbGuest->u64CR3);
3014	VMCPU_FF_SET(pVCpu, VMCPU_FF_HM_UPDATE_CR3);
3015	}
3016	}
3017
3018	/* Changes to CR4 are always intercepted. */
3019	}
3020
3021	/* Update fExtrn. */
3022	pCtx->fExtrn &= ~fWhat;
3023
3024	/* If everything has been imported, clear the HM keeper bit. */
3025	if (!(pCtx->fExtrn & HMSVM_CPUMCTX_EXTRN_ALL))
3026	{
3027	pCtx->fExtrn &= ~CPUMCTX_EXTRN_KEEPER_HM;
3028	Assert(!pCtx->fExtrn);
3029	}
3030	}
3031	else
3032	Assert(!pCtx->fExtrn \|\| (pCtx->fExtrn & HMSVM_CPUMCTX_EXTRN_ALL));
3033
3034	ASMSetFlags(fEFlags);
3035
3036	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatImportGuestState, x);
3037
3038	/*
3039	* Honor any pending CR3 updates.
3040	*
3041	* Consider this scenario: #VMEXIT -> VMMRZCallRing3Enable() -> do stuff that causes a longjmp
3042	* -> hmR0SvmCallRing3Callback() -> VMMRZCallRing3Disable() -> hmR0SvmImportGuestState()
3043	* -> Sets VMCPU_FF_HM_UPDATE_CR3 pending -> return from the longjmp -> continue with #VMEXIT
3044	* handling -> hmR0SvmImportGuestState() and here we are.
3045	*
3046	* The reason for such complicated handling is because VM-exits that call into PGM expect
3047	* CR3 to be up-to-date and thus any CR3-saves -before- the VM-exit (longjmp) would've
3048	* postponed the CR3 update via the force-flag and cleared CR3 from fExtrn. Any SVM R0
3049	* VM-exit handler that requests CR3 to be saved will end up here and we call PGMUpdateCR3().
3050	*
3051	* The longjmp exit path can't check these CR3 force-flags and call code that takes a lock again,
3052	* and does not process force-flag like regular exits to ring-3 either, we cover for it here.
3053	*/
3054	if ( VMMRZCallRing3IsEnabled(pVCpu)
3055	&& VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_UPDATE_CR3))
3056	{
3057	Assert(pCtx->cr3 == pVmcbGuest->u64CR3);
3058	PGMUpdateCR3(pVCpu, pCtx->cr3);
3059	}
3060	}
3061
3062
3063	/**
3064	* Saves the guest (or nested-guest) state from the VMCB into the guest-CPU
3065	* context.
3066	*
3067	* Currently there is no residual state left in the CPU that is not updated in the
3068	* VMCB.
3069	*
3070	* @returns VBox status code.
3071	* @param pVCpu The cross context virtual CPU structure.
3072	* @param fWhat What to import, CPUMCTX_EXTRN_XXX.
3073	*/
3074	VMMR0DECL(int) SVMR0ImportStateOnDemand(PVMCPU pVCpu, uint64_t fWhat)
3075	{
3076	hmR0SvmImportGuestState(pVCpu, fWhat);
3077	return VINF_SUCCESS;
3078	}
3079
3080
3081	/**
3082	* Does the necessary state syncing before returning to ring-3 for any reason
3083	* (longjmp, preemption, voluntary exits to ring-3) from AMD-V.
3084	*
3085	* @param pVCpu The cross context virtual CPU structure.
3086	* @param fImportState Whether to import the guest state from the VMCB back
3087	* to the guest-CPU context.
3088	*
3089	* @remarks No-long-jmp zone!!!
3090	*/
3091	static void hmR0SvmLeave(PVMCPU pVCpu, bool fImportState)
3092	{
3093	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
3094	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
3095	Assert(VMMR0IsLogFlushDisabled(pVCpu));
3096
3097	/*
3098	* !!! IMPORTANT !!!
3099	* If you modify code here, make sure to check whether hmR0SvmCallRing3Callback() needs to be updated too.
3100	*/
3101
3102	/* Save the guest state if necessary. */
3103	if (fImportState)
3104	hmR0SvmImportGuestState(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
3105
3106	/* Restore host FPU state if necessary and resync on next R0 reentry. */
3107	CPUMR0FpuStateMaybeSaveGuestAndRestoreHost(pVCpu);
3108	Assert(!CPUMIsGuestFPUStateActive(pVCpu));
3109
3110	/*
3111	* Restore host debug registers if necessary and resync on next R0 reentry.
3112	*/
3113	#ifdef VBOX_STRICT
3114	if (CPUMIsHyperDebugStateActive(pVCpu))
3115	{
3116	PSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb; /** @todo nested-guest. */
3117	Assert(pVmcb->ctrl.u16InterceptRdDRx == 0xffff);
3118	Assert(pVmcb->ctrl.u16InterceptWrDRx == 0xffff);
3119	}
3120	#endif
3121	CPUMR0DebugStateMaybeSaveGuestAndRestoreHost(pVCpu, false /* save DR6 */);
3122	Assert(!CPUMIsHyperDebugStateActive(pVCpu));
3123	Assert(!CPUMIsGuestDebugStateActive(pVCpu));
3124
3125	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatEntry);
3126	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatImportGuestState);
3127	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatExportGuestState);
3128	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatExit1);
3129	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatExit2);
3130	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchLongJmpToR3);
3131
3132	VMCPU_CMPXCHG_STATE(pVCpu, VMCPUSTATE_STARTED_HM, VMCPUSTATE_STARTED_EXEC);
3133	}
3134
3135
3136	/**
3137	* Leaves the AMD-V session.
3138	*
3139	* Only used while returning to ring-3 either due to longjump or exits to
3140	* ring-3.
3141	*
3142	* @returns VBox status code.
3143	* @param pVCpu The cross context virtual CPU structure.
3144	*/
3145	static int hmR0SvmLeaveSession(PVMCPU pVCpu)
3146	{
3147	HM_DISABLE_PREEMPT();
3148	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
3149	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
3150
3151	/* When thread-context hooks are used, we can avoid doing the leave again if we had been preempted before
3152	and done this from the SVMR0ThreadCtxCallback(). */
3153	if (!pVCpu->hm.s.fLeaveDone)
3154	{
3155	hmR0SvmLeave(pVCpu, true /* fImportState */);
3156	pVCpu->hm.s.fLeaveDone = true;
3157	}
3158
3159	/*
3160	* !!! IMPORTANT !!!
3161	* If you modify code here, make sure to check whether hmR0SvmCallRing3Callback() needs to be updated too.
3162	*/
3163
3164	/** @todo eliminate the need for calling VMMR0ThreadCtxHookDisable here! */
3165	/* Deregister hook now that we've left HM context before re-enabling preemption. */
3166	VMMR0ThreadCtxHookDisable(pVCpu);
3167
3168	/* Leave HM context. This takes care of local init (term). */
3169	int rc = HMR0LeaveCpu(pVCpu);
3170
3171	HM_RESTORE_PREEMPT();
3172	return rc;
3173	}
3174
3175
3176	/**
3177	* Does the necessary state syncing before doing a longjmp to ring-3.
3178	*
3179	* @returns VBox status code.
3180	* @param pVCpu The cross context virtual CPU structure.
3181	*
3182	* @remarks No-long-jmp zone!!!
3183	*/
3184	static int hmR0SvmLongJmpToRing3(PVMCPU pVCpu)
3185	{
3186	return hmR0SvmLeaveSession(pVCpu);
3187	}
3188
3189
3190	/**
3191	* VMMRZCallRing3() callback wrapper which saves the guest state (or restores
3192	* any remaining host state) before we longjump to ring-3 and possibly get
3193	* preempted.
3194	*
3195	* @param pVCpu The cross context virtual CPU structure.
3196	* @param enmOperation The operation causing the ring-3 longjump.
3197	* @param pvUser The user argument (pointer to the possibly
3198	* out-of-date guest-CPU context).
3199	*/
3200	static DECLCALLBACK(int) hmR0SvmCallRing3Callback(PVMCPU pVCpu, VMMCALLRING3 enmOperation, void *pvUser)
3201	{
3202	RT_NOREF_PV(pvUser);
3203
3204	if (enmOperation == VMMCALLRING3_VM_R0_ASSERTION)
3205	{
3206	/*
3207	* !!! IMPORTANT !!!
3208	* If you modify code here, make sure to check whether hmR0SvmLeave() and hmR0SvmLeaveSession() needs
3209	* to be updated too. This is a stripped down version which gets out ASAP trying to not trigger any assertion.
3210	*/
3211	VMMRZCallRing3RemoveNotification(pVCpu);
3212	VMMRZCallRing3Disable(pVCpu);
3213	HM_DISABLE_PREEMPT();
3214
3215	/* Restore host FPU state if necessary and resync on next R0 reentry. */
3216	CPUMR0FpuStateMaybeSaveGuestAndRestoreHost(pVCpu);
3217
3218	/* Restore host debug registers if necessary and resync on next R0 reentry. */
3219	CPUMR0DebugStateMaybeSaveGuestAndRestoreHost(pVCpu, false /* save DR6 */);
3220
3221	/* Deregister the hook now that we've left HM context before re-enabling preemption. */
3222	/** @todo eliminate the need for calling VMMR0ThreadCtxHookDisable here! */
3223	VMMR0ThreadCtxHookDisable(pVCpu);
3224
3225	/* Leave HM context. This takes care of local init (term). */
3226	HMR0LeaveCpu(pVCpu);
3227
3228	HM_RESTORE_PREEMPT();
3229	return VINF_SUCCESS;
3230	}
3231
3232	Assert(pVCpu);
3233	Assert(pvUser);
3234	Assert(VMMRZCallRing3IsEnabled(pVCpu));
3235	HMSVM_ASSERT_PREEMPT_SAFE();
3236
3237	VMMRZCallRing3Disable(pVCpu);
3238	Assert(VMMR0IsLogFlushDisabled(pVCpu));
3239
3240	Log4Func(("Calling hmR0SvmLongJmpToRing3\n"));
3241	int rc = hmR0SvmLongJmpToRing3(pVCpu);
3242	AssertRCReturn(rc, rc);
3243
3244	VMMRZCallRing3Enable(pVCpu);
3245	return VINF_SUCCESS;
3246	}
3247
3248
3249	/**
3250	* Take necessary actions before going back to ring-3.
3251	*
3252	* An action requires us to go back to ring-3. This function does the necessary
3253	* steps before we can safely return to ring-3. This is not the same as longjmps
3254	* to ring-3, this is voluntary.
3255	*
3256	* @returns VBox status code.
3257	* @param pVM The cross context VM structure.
3258	* @param pVCpu The cross context virtual CPU structure.
3259	* @param pCtx Pointer to the guest-CPU context.
3260	* @param rcExit The reason for exiting to ring-3. Can be
3261	* VINF_VMM_UNKNOWN_RING3_CALL.
3262	*/
3263	static int hmR0SvmExitToRing3(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, int rcExit)
3264	{
3265	Assert(pVM);
3266	Assert(pVCpu);
3267	Assert(pCtx);
3268	HMSVM_ASSERT_PREEMPT_SAFE();
3269
3270	/* Please, no longjumps here (any logging shouldn't flush jump back to ring-3). NO LOGGING BEFORE THIS POINT! */
3271	VMMRZCallRing3Disable(pVCpu);
3272	Log4Func(("rcExit=%d LocalFF=%#RX32 GlobalFF=%#RX32\n", rcExit, pVCpu->fLocalForcedActions, pVM->fGlobalForcedActions));
3273
3274	/* We need to do this only while truly exiting the "inner loop" back to ring-3 and -not- for any longjmp to ring3. */
3275	if (pVCpu->hm.s.Event.fPending)
3276	{
3277	hmR0SvmPendingEventToTrpmTrap(pVCpu);
3278	Assert(!pVCpu->hm.s.Event.fPending);
3279	}
3280
3281	/* Sync. the necessary state for going back to ring-3. */
3282	hmR0SvmLeaveSession(pVCpu);
3283	STAM_COUNTER_DEC(&pVCpu->hm.s.StatSwitchLongJmpToR3);
3284
3285	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_TO_R3);
3286	CPUMSetChangedFlags(pVCpu, CPUM_CHANGED_SYSENTER_MSR
3287	\| CPUM_CHANGED_LDTR
3288	\| CPUM_CHANGED_GDTR
3289	\| CPUM_CHANGED_IDTR
3290	\| CPUM_CHANGED_TR
3291	\| CPUM_CHANGED_HIDDEN_SEL_REGS);
3292	if ( pVM->hm.s.fNestedPaging
3293	&& CPUMIsGuestPagingEnabledEx(pCtx))
3294	{
3295	CPUMSetChangedFlags(pVCpu, CPUM_CHANGED_GLOBAL_TLB_FLUSH);
3296	}
3297
3298	/* Update the exit-to-ring 3 reason. */
3299	pVCpu->hm.s.rcLastExitToR3 = rcExit;
3300
3301	/* On our way back from ring-3 reload the guest state if there is a possibility of it being changed. */
3302	if (rcExit != VINF_EM_RAW_INTERRUPT)
3303	{
3304	Assert(!(pCtx->fExtrn & HMSVM_CPUMCTX_EXTRN_ALL));
3305	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
3306	}
3307
3308	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchExitToR3);
3309
3310	/* We do -not- want any longjmp notifications after this! We must return to ring-3 ASAP. */
3311	VMMRZCallRing3RemoveNotification(pVCpu);
3312	VMMRZCallRing3Enable(pVCpu);
3313
3314	/*
3315	* If we're emulating an instruction, we shouldn't have any TRPM traps pending
3316	* and if we're injecting an event we should have a TRPM trap pending.
3317	*/
3318	AssertReturnStmt(rcExit != VINF_EM_RAW_INJECT_TRPM_EVENT \|\| TRPMHasTrap(pVCpu),
3319	pVCpu->hm.s.u32HMError = rcExit,
3320	VERR_SVM_IPE_5);
3321	AssertReturnStmt(rcExit != VINF_EM_RAW_EMULATE_INSTR \|\| !TRPMHasTrap(pVCpu),
3322	pVCpu->hm.s.u32HMError = rcExit,
3323	VERR_SVM_IPE_4);
3324
3325	return rcExit;
3326	}
3327
3328
3329	/**
3330	* Updates the use of TSC offsetting mode for the CPU and adjusts the necessary
3331	* intercepts.
3332	*
3333	* @param pVM The cross context VM structure.
3334	* @param pVCpu The cross context virtual CPU structure.
3335	* @param pCtx Pointer to the guest-CPU or nested-guest-CPU context.
3336	* @param pVmcb Pointer to the VM control block.
3337	*
3338	* @remarks No-long-jump zone!!!
3339	*/
3340	static void hmR0SvmUpdateTscOffsetting(PVM pVM, PVMCPU pVCpu, PCCPUMCTX pCtx, PSVMVMCB pVmcb)
3341	{
3342	/*
3343	* Avoid intercepting RDTSC/RDTSCP if we determined the host TSC (++) is stable
3344	* and in case of a nested-guest, if the nested-VMCB specifies it is not intercepting
3345	* RDTSC/RDTSCP as well.
3346	*/
3347	bool fParavirtTsc;
3348	uint64_t uTscOffset;
3349	bool const fCanUseRealTsc = TMCpuTickCanUseRealTSC(pVM, pVCpu, &uTscOffset, &fParavirtTsc);
3350
3351	bool fIntercept;
3352	if (fCanUseRealTsc)
3353	fIntercept = hmR0SvmClearCtrlIntercept(pVCpu, pCtx, pVmcb, SVM_CTRL_INTERCEPT_RDTSC \| SVM_CTRL_INTERCEPT_RDTSCP);
3354	else
3355	{
3356	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_RDTSC \| SVM_CTRL_INTERCEPT_RDTSCP);
3357	fIntercept = true;
3358	}
3359
3360	if (!fIntercept)
3361	{
3362	/* Apply the nested-guest VMCB's TSC offset over the guest TSC offset. */
3363	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
3364	uTscOffset = HMSvmNstGstApplyTscOffset(pVCpu, uTscOffset);
3365
3366	/* Update the TSC offset in the VMCB and the relevant clean bits. */
3367	pVmcb->ctrl.u64TSCOffset = uTscOffset;
3368	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
3369
3370	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscOffset);
3371	}
3372	else
3373	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscIntercept);
3374
3375	/* Currently neither Hyper-V nor KVM need to update their paravirt. TSC
3376	information before every VM-entry, hence we have nothing to do here at the moment. */
3377	if (fParavirtTsc)
3378	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscParavirt);
3379	}
3380
3381
3382	/**
3383	* Sets an event as a pending event to be injected into the guest.
3384	*
3385	* @param pVCpu The cross context virtual CPU structure.
3386	* @param pEvent Pointer to the SVM event.
3387	* @param GCPtrFaultAddress The fault-address (CR2) in case it's a
3388	* page-fault.
3389	*
3390	* @remarks Statistics counter assumes this is a guest event being reflected to
3391	* the guest i.e. 'StatInjectPendingReflect' is incremented always.
3392	*/
3393	DECLINLINE(void) hmR0SvmSetPendingEvent(PVMCPU pVCpu, PSVMEVENT pEvent, RTGCUINTPTR GCPtrFaultAddress)
3394	{
3395	Assert(!pVCpu->hm.s.Event.fPending);
3396	Assert(pEvent->n.u1Valid);
3397
3398	pVCpu->hm.s.Event.u64IntInfo = pEvent->u;
3399	pVCpu->hm.s.Event.fPending = true;
3400	pVCpu->hm.s.Event.GCPtrFaultAddress = GCPtrFaultAddress;
3401
3402	Log4Func(("u=%#RX64 u8Vector=%#x Type=%#x ErrorCodeValid=%RTbool ErrorCode=%#RX32\n", pEvent->u, pEvent->n.u8Vector,
3403	(uint8_t)pEvent->n.u3Type, !!pEvent->n.u1ErrorCodeValid, pEvent->n.u32ErrorCode));
3404	}
3405
3406
3407	/**
3408	* Sets an invalid-opcode (\#UD) exception as pending-for-injection into the VM.
3409	*
3410	* @param pVCpu The cross context virtual CPU structure.
3411	*/
3412	DECLINLINE(void) hmR0SvmSetPendingXcptUD(PVMCPU pVCpu)
3413	{
3414	SVMEVENT Event;
3415	Event.u = 0;
3416	Event.n.u1Valid = 1;
3417	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3418	Event.n.u8Vector = X86_XCPT_UD;
3419	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3420	}
3421
3422
3423	/**
3424	* Sets a debug (\#DB) exception as pending-for-injection into the VM.
3425	*
3426	* @param pVCpu The cross context virtual CPU structure.
3427	*/
3428	DECLINLINE(void) hmR0SvmSetPendingXcptDB(PVMCPU pVCpu)
3429	{
3430	SVMEVENT Event;
3431	Event.u = 0;
3432	Event.n.u1Valid = 1;
3433	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3434	Event.n.u8Vector = X86_XCPT_DB;
3435	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3436	}
3437
3438
3439	/**
3440	* Sets a page fault (\#PF) exception as pending-for-injection into the VM.
3441	*
3442	* @param pVCpu The cross context virtual CPU structure.
3443	* @param pCtx Pointer to the guest-CPU context.
3444	* @param u32ErrCode The error-code for the page-fault.
3445	* @param uFaultAddress The page fault address (CR2).
3446	*
3447	* @remarks This updates the guest CR2 with @a uFaultAddress!
3448	*/
3449	DECLINLINE(void) hmR0SvmSetPendingXcptPF(PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t u32ErrCode, RTGCUINTPTR uFaultAddress)
3450	{
3451	SVMEVENT Event;
3452	Event.u = 0;
3453	Event.n.u1Valid = 1;
3454	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3455	Event.n.u8Vector = X86_XCPT_PF;
3456	Event.n.u1ErrorCodeValid = 1;
3457	Event.n.u32ErrorCode = u32ErrCode;
3458
3459	/* Update CR2 of the guest. */
3460	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_CR2);
3461	if (pCtx->cr2 != uFaultAddress)
3462	{
3463	pCtx->cr2 = uFaultAddress;
3464	/* The VMCB clean bit for CR2 will be updated while re-loading the guest state. */
3465	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_CR2);
3466	}
3467
3468	hmR0SvmSetPendingEvent(pVCpu, &Event, uFaultAddress);
3469	}
3470
3471
3472	/**
3473	* Sets a math-fault (\#MF) exception as pending-for-injection into the VM.
3474	*
3475	* @param pVCpu The cross context virtual CPU structure.
3476	*/
3477	DECLINLINE(void) hmR0SvmSetPendingXcptMF(PVMCPU pVCpu)
3478	{
3479	SVMEVENT Event;
3480	Event.u = 0;
3481	Event.n.u1Valid = 1;
3482	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3483	Event.n.u8Vector = X86_XCPT_MF;
3484	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3485	}
3486
3487
3488	/**
3489	* Sets a double fault (\#DF) exception as pending-for-injection into the VM.
3490	*
3491	* @param pVCpu The cross context virtual CPU structure.
3492	*/
3493	DECLINLINE(void) hmR0SvmSetPendingXcptDF(PVMCPU pVCpu)
3494	{
3495	SVMEVENT Event;
3496	Event.u = 0;
3497	Event.n.u1Valid = 1;
3498	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3499	Event.n.u8Vector = X86_XCPT_DF;
3500	Event.n.u1ErrorCodeValid = 1;
3501	Event.n.u32ErrorCode = 0;
3502	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3503	}
3504
3505
3506	/**
3507	* Injects an event into the guest upon VMRUN by updating the relevant field
3508	* in the VMCB.
3509	*
3510	* @param pVCpu The cross context virtual CPU structure.
3511	* @param pVmcb Pointer to the guest VM control block.
3512	* @param pEvent Pointer to the event.
3513	*
3514	* @remarks No-long-jump zone!!!
3515	* @remarks Requires CR0!
3516	*/
3517	DECLINLINE(void) hmR0SvmInjectEventVmcb(PVMCPU pVCpu, PSVMVMCB pVmcb, PSVMEVENT pEvent)
3518	{
3519	Assert(!pVmcb->ctrl.EventInject.n.u1Valid);
3520	pVmcb->ctrl.EventInject.u = pEvent->u;
3521	STAM_COUNTER_INC(&pVCpu->hm.s.paStatInjectedIrqsR0[pEvent->n.u8Vector & MASK_INJECT_IRQ_STAT]);
3522	RT_NOREF(pVCpu);
3523
3524	Log4Func(("u=%#RX64 u8Vector=%#x Type=%#x ErrorCodeValid=%RTbool ErrorCode=%#RX32\n", pEvent->u, pEvent->n.u8Vector,
3525	(uint8_t)pEvent->n.u3Type, !!pEvent->n.u1ErrorCodeValid, pEvent->n.u32ErrorCode));
3526	}
3527
3528
3529
3530	/**
3531	* Converts any TRPM trap into a pending HM event. This is typically used when
3532	* entering from ring-3 (not longjmp returns).
3533	*
3534	* @param pVCpu The cross context virtual CPU structure.
3535	*/
3536	static void hmR0SvmTrpmTrapToPendingEvent(PVMCPU pVCpu)
3537	{
3538	Assert(TRPMHasTrap(pVCpu));
3539	Assert(!pVCpu->hm.s.Event.fPending);
3540
3541	uint8_t uVector;
3542	TRPMEVENT enmTrpmEvent;
3543	RTGCUINT uErrCode;
3544	RTGCUINTPTR GCPtrFaultAddress;
3545	uint8_t cbInstr;
3546
3547	int rc = TRPMQueryTrapAll(pVCpu, &uVector, &enmTrpmEvent, &uErrCode, &GCPtrFaultAddress, &cbInstr);
3548	AssertRC(rc);
3549
3550	SVMEVENT Event;
3551	Event.u = 0;
3552	Event.n.u1Valid = 1;
3553	Event.n.u8Vector = uVector;
3554
3555	/* Refer AMD spec. 15.20 "Event Injection" for the format. */
3556	if (enmTrpmEvent == TRPM_TRAP)
3557	{
3558	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3559	switch (uVector)
3560	{
3561	case X86_XCPT_NMI:
3562	{
3563	Event.n.u3Type = SVM_EVENT_NMI;
3564	break;
3565	}
3566
3567	case X86_XCPT_PF:
3568	case X86_XCPT_DF:
3569	case X86_XCPT_TS:
3570	case X86_XCPT_NP:
3571	case X86_XCPT_SS:
3572	case X86_XCPT_GP:
3573	case X86_XCPT_AC:
3574	{
3575	Event.n.u1ErrorCodeValid = 1;
3576	Event.n.u32ErrorCode = uErrCode;
3577	break;
3578	}
3579	}
3580	}
3581	else if (enmTrpmEvent == TRPM_HARDWARE_INT)
3582	Event.n.u3Type = SVM_EVENT_EXTERNAL_IRQ;
3583	else if (enmTrpmEvent == TRPM_SOFTWARE_INT)
3584	Event.n.u3Type = SVM_EVENT_SOFTWARE_INT;
3585	else
3586	AssertMsgFailed(("Invalid TRPM event type %d\n", enmTrpmEvent));
3587
3588	rc = TRPMResetTrap(pVCpu);
3589	AssertRC(rc);
3590
3591	Log4(("TRPM->HM event: u=%#RX64 u8Vector=%#x uErrorCodeValid=%RTbool uErrorCode=%#RX32\n", Event.u, Event.n.u8Vector,
3592	!!Event.n.u1ErrorCodeValid, Event.n.u32ErrorCode));
3593
3594	hmR0SvmSetPendingEvent(pVCpu, &Event, GCPtrFaultAddress);
3595	}
3596
3597
3598	/**
3599	* Converts any pending SVM event into a TRPM trap. Typically used when leaving
3600	* AMD-V to execute any instruction.
3601	*
3602	* @param pVCpu The cross context virtual CPU structure.
3603	*/
3604	static void hmR0SvmPendingEventToTrpmTrap(PVMCPU pVCpu)
3605	{
3606	Assert(pVCpu->hm.s.Event.fPending);
3607	Assert(TRPMQueryTrap(pVCpu, NULL /* pu8TrapNo /, NULL / pEnmType */) == VERR_TRPM_NO_ACTIVE_TRAP);
3608
3609	SVMEVENT Event;
3610	Event.u = pVCpu->hm.s.Event.u64IntInfo;
3611
3612	uint8_t uVector = Event.n.u8Vector;
3613	uint8_t uVectorType = Event.n.u3Type;
3614	TRPMEVENT enmTrapType = HMSvmEventToTrpmEventType(&Event);
3615
3616	Log4(("HM event->TRPM: uVector=%#x enmTrapType=%d\n", uVector, uVectorType));
3617
3618	int rc = TRPMAssertTrap(pVCpu, uVector, enmTrapType);
3619	AssertRC(rc);
3620
3621	if (Event.n.u1ErrorCodeValid)
3622	TRPMSetErrorCode(pVCpu, Event.n.u32ErrorCode);
3623
3624	if ( uVectorType == SVM_EVENT_EXCEPTION
3625	&& uVector == X86_XCPT_PF)
3626	{
3627	TRPMSetFaultAddress(pVCpu, pVCpu->hm.s.Event.GCPtrFaultAddress);
3628	Assert(pVCpu->hm.s.Event.GCPtrFaultAddress == CPUMGetGuestCR2(pVCpu));
3629	}
3630	else if (uVectorType == SVM_EVENT_SOFTWARE_INT)
3631	{
3632	AssertMsg( uVectorType == SVM_EVENT_SOFTWARE_INT
3633	\|\| (uVector == X86_XCPT_BP \|\| uVector == X86_XCPT_OF),
3634	("Invalid vector: uVector=%#x uVectorType=%#x\n", uVector, uVectorType));
3635	TRPMSetInstrLength(pVCpu, pVCpu->hm.s.Event.cbInstr);
3636	}
3637	pVCpu->hm.s.Event.fPending = false;
3638	}
3639
3640
3641	/**
3642	* Checks if the guest (or nested-guest) has an interrupt shadow active right
3643	* now.
3644	*
3645	* @returns @c true if the interrupt shadow is active, @c false otherwise.
3646	* @param pVCpu The cross context virtual CPU structure.
3647	* @param pCtx Pointer to the guest-CPU context.
3648	*
3649	* @remarks No-long-jump zone!!!
3650	* @remarks Has side-effects with VMCPU_FF_INHIBIT_INTERRUPTS force-flag.
3651	*/
3652	DECLINLINE(bool) hmR0SvmIsIntrShadowActive(PVMCPU pVCpu, PCCPUMCTX pCtx)
3653	{
3654	/*
3655	* Instructions like STI and MOV SS inhibit interrupts till the next instruction
3656	* completes. Check if we should inhibit interrupts or clear any existing
3657	* interrupt inhibition.
3658	*/
3659	if (VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS))
3660	{
3661	if (pCtx->rip != EMGetInhibitInterruptsPC(pVCpu))
3662	{
3663	/*
3664	* We can clear the inhibit force flag as even if we go back to the recompiler
3665	* without executing guest code in AMD-V, the flag's condition to be cleared is
3666	* met and thus the cleared state is correct.
3667	*/
3668	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS);
3669	return false;
3670	}
3671	return true;
3672	}
3673	return false;
3674	}
3675
3676
3677	/**
3678	* Sets the virtual interrupt intercept control in the VMCB.
3679	*
3680	* @param pVCpu The cross context virtual CPU structure.
3681	* @param pVmcb Pointer to the VM control block.
3682	* @param pCtx Pointer to the guest-CPU context.
3683	*/
3684	DECLINLINE(void) hmR0SvmSetIntWindowExiting(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
3685	{
3686	/*
3687	* When AVIC isn't supported, set up an interrupt window to cause a #VMEXIT when the guest
3688	* is ready to accept interrupts. At #VMEXIT, we then get the interrupt from the APIC
3689	* (updating ISR at the right time) and inject the interrupt.
3690	*
3691	* With AVIC is supported, we could make use of the asynchronously delivery without
3692	* #VMEXIT and we would be passing the AVIC page to SVM.
3693	*
3694	* In AMD-V, an interrupt window is achieved using a combination of V_IRQ (an interrupt
3695	* is pending), V_IGN_TPR (ignore TPR priorities) and the VINTR intercept all being set.
3696	*/
3697	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
3698	/*
3699	* Currently we don't overlay interupt windows and if there's any V_IRQ pending in the
3700	* nested-guest VMCB, we avoid setting up any interrupt window on behalf of the outer
3701	* guest.
3702	*/
3703	/** @todo Does this mean we end up prioritizing virtual interrupt
3704	* delivery/window over a physical interrupt (from the outer guest)
3705	* might be pending? */
3706	bool const fEnableIntWindow = !VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_NESTED_GUEST);
3707	if (!fEnableIntWindow)
3708	{
3709	Assert(CPUMIsGuestInSvmNestedHwVirtMode(pCtx)); RT_NOREF(pCtx);
3710	Log4(("Nested-guest V_IRQ already pending\n"));
3711	}
3712	#else
3713	RT_NOREF2(pVCpu, pCtx);
3714	bool const fEnableIntWindow = true;
3715	#endif
3716	if (fEnableIntWindow)
3717	{
3718	Assert(pVmcb->ctrl.IntCtrl.n.u1IgnoreTPR);
3719	pVmcb->ctrl.IntCtrl.n.u1VIrqPending = 1;
3720	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INT_CTRL;
3721	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_VINTR);
3722	Log4(("Set VINTR intercept\n"));
3723	}
3724	}
3725
3726
3727	/**
3728	* Clears the virtual interrupt intercept control in the VMCB as
3729	* we are figured the guest is unable process any interrupts
3730	* at this point of time.
3731	*
3732	* @param pVCpu The cross context virtual CPU structure.
3733	* @param pVmcb Pointer to the VM control block.
3734	* @param pCtx Pointer to the guest-CPU context.
3735	*/
3736	DECLINLINE(void) hmR0SvmClearIntWindowExiting(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
3737	{
3738	PSVMVMCBCTRL pVmcbCtrl = &pVmcb->ctrl;
3739	if ( pVmcbCtrl->IntCtrl.n.u1VIrqPending
3740	\|\| (pVmcbCtrl->u64InterceptCtrl & SVM_CTRL_INTERCEPT_VINTR))
3741	{
3742	pVmcbCtrl->IntCtrl.n.u1VIrqPending = 0;
3743	pVmcbCtrl->u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INT_CTRL;
3744	hmR0SvmClearCtrlIntercept(pVCpu, pCtx, pVmcb, SVM_CTRL_INTERCEPT_VINTR);
3745	Log4(("Cleared VINTR intercept\n"));
3746	}
3747	}
3748
3749	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
3750	/**
3751	* Evaluates the event to be delivered to the nested-guest and sets it as the
3752	* pending event.
3753	*
3754	* @returns VBox strict status code.
3755	* @param pVCpu The cross context virtual CPU structure.
3756	* @param pCtx Pointer to the guest-CPU context.
3757	*/
3758	static VBOXSTRICTRC hmR0SvmEvaluatePendingEventNested(PVMCPU pVCpu, PCPUMCTX pCtx)
3759	{
3760	HMSVM_ASSERT_IN_NESTED_GUEST(pCtx);
3761	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_HWVIRT
3762	\| CPUMCTX_EXTRN_RFLAGS
3763	\| CPUMCTX_EXTRN_HM_SVM_INT_SHADOW
3764	\| CPUMCTX_EXTRN_HM_SVM_HWVIRT_VIRQ);
3765
3766	Assert(!pVCpu->hm.s.Event.fPending);
3767	Assert(pCtx->hwvirt.fGif);
3768	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
3769	Assert(pVmcb);
3770
3771	bool const fVirtualGif = CPUMGetSvmNstGstVGif(pCtx);
3772	bool const fIntShadow = hmR0SvmIsIntrShadowActive(pVCpu, pCtx);
3773	bool const fBlockNmi = VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_BLOCK_NMIS);
3774
3775	Log4Func(("fVirtualGif=%RTbool fBlockNmi=%RTbool fIntShadow=%RTbool fIntPending=%RTbool fNmiPending=%RTbool\n",
3776	fVirtualGif, fBlockNmi, fIntShadow, VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC),
3777	VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_NMI)));
3778
3779	/** @todo SMI. SMIs take priority over NMIs. */
3780
3781	/*
3782	* Check if the guest can receive NMIs.
3783	* Nested NMIs are not allowed, see AMD spec. 8.1.4 "Masking External Interrupts".
3784	* NMIs take priority over maskable interrupts, see AMD spec. 8.5 "Priorities".
3785	*/
3786	if ( VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_NMI)
3787	&& !fBlockNmi)
3788	{
3789	if ( fVirtualGif
3790	&& !fIntShadow)
3791	{
3792	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_NMI))
3793	{
3794	Log4(("Intercepting NMI -> #VMEXIT\n"));
3795	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_SVM_VMEXIT_MASK);
3796	return IEMExecSvmVmexit(pVCpu, SVM_EXIT_NMI, 0, 0);
3797	}
3798
3799	Log4(("Setting NMI pending for injection\n"));
3800	SVMEVENT Event;
3801	Event.u = 0;
3802	Event.n.u1Valid = 1;
3803	Event.n.u8Vector = X86_XCPT_NMI;
3804	Event.n.u3Type = SVM_EVENT_NMI;
3805	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3806	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INTERRUPT_NMI);
3807	}
3808	else if (!fVirtualGif)
3809	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_STGI);
3810	else
3811	hmR0SvmSetIntWindowExiting(pVCpu, pVmcb, pCtx);
3812	}
3813	/*
3814	* Check if the nested-guest can receive external interrupts (generated by the guest's
3815	* PIC/APIC).
3816	*
3817	* External intercepts, NMI, SMI etc. from the physical CPU are -always- intercepted
3818	* when executing using hardware-assisted SVM, see HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS.
3819	*
3820	* External interrupts that are generated for the outer guest may be intercepted
3821	* depending on how the nested-guest VMCB was programmed by guest software.
3822	*
3823	* Physical interrupts always take priority over virtual interrupts,
3824	* see AMD spec. 15.21.4 "Injecting Virtual (INTR) Interrupts".
3825	*/
3826	else if ( VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC)
3827	&& !pVCpu->hm.s.fSingleInstruction)
3828	{
3829	if ( fVirtualGif
3830	&& !fIntShadow
3831	&& CPUMCanSvmNstGstTakePhysIntr(pVCpu, pCtx))
3832	{
3833	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_INTR))
3834	{
3835	Log4(("Intercepting INTR -> #VMEXIT\n"));
3836	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_SVM_VMEXIT_MASK);
3837	return IEMExecSvmVmexit(pVCpu, SVM_EXIT_INTR, 0, 0);
3838	}
3839
3840	uint8_t u8Interrupt;
3841	int rc = PDMGetInterrupt(pVCpu, &u8Interrupt);
3842	if (RT_SUCCESS(rc))
3843	{
3844	Log4(("Setting external interrupt %#x pending for injection\n", u8Interrupt));
3845	SVMEVENT Event;
3846	Event.u = 0;
3847	Event.n.u1Valid = 1;
3848	Event.n.u8Vector = u8Interrupt;
3849	Event.n.u3Type = SVM_EVENT_EXTERNAL_IRQ;
3850	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3851	}
3852	else if (rc == VERR_APIC_INTR_MASKED_BY_TPR)
3853	{
3854	/*
3855	* AMD-V has no TPR thresholding feature. TPR and the force-flag will be
3856	* updated eventually when the TPR is written by the guest.
3857	*/
3858	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchTprMaskedIrq);
3859	}
3860	else
3861	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchGuestIrq);
3862	}
3863	else if (!fVirtualGif)
3864	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_STGI);
3865	else
3866	hmR0SvmSetIntWindowExiting(pVCpu, pVmcb, pCtx);
3867	}
3868
3869	return VINF_SUCCESS;
3870	}
3871	#endif
3872
3873	/**
3874	* Evaluates the event to be delivered to the guest and sets it as the pending
3875	* event.
3876	*
3877	* @param pVCpu The cross context virtual CPU structure.
3878	* @param pCtx Pointer to the guest-CPU context.
3879	*/
3880	static void hmR0SvmEvaluatePendingEvent(PVMCPU pVCpu, PCPUMCTX pCtx)
3881	{
3882	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(pCtx);
3883	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_HWVIRT
3884	\| CPUMCTX_EXTRN_RFLAGS
3885	\| CPUMCTX_EXTRN_HM_SVM_INT_SHADOW);
3886
3887	Assert(!pVCpu->hm.s.Event.fPending);
3888	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
3889	Assert(pVmcb);
3890
3891	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
3892	bool const fGif = pCtx->hwvirt.fGif;
3893	#else
3894	bool const fGif = true;
3895	#endif
3896	bool const fIntShadow = hmR0SvmIsIntrShadowActive(pVCpu, pCtx);
3897	bool const fBlockInt = !(pCtx->eflags.u32 & X86_EFL_IF);
3898	bool const fBlockNmi = VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_BLOCK_NMIS);
3899
3900	Log4Func(("fGif=%RTbool fBlockNmi=%RTbool fBlockInt=%RTbool fIntShadow=%RTbool fIntPending=%RTbool NMI pending=%RTbool\n",
3901	fGif, fBlockNmi, fBlockInt, fIntShadow,
3902	VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC),
3903	VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_NMI)));
3904
3905	/** @todo SMI. SMIs take priority over NMIs. */
3906
3907	/*
3908	* Check if the guest can receive NMIs.
3909	* Nested NMIs are not allowed, see AMD spec. 8.1.4 "Masking External Interrupts".
3910	* NMIs take priority over maskable interrupts, see AMD spec. 8.5 "Priorities".
3911	*/
3912	if ( VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_NMI)
3913	&& !fBlockNmi)
3914	{
3915	if ( fGif
3916	&& !fIntShadow)
3917	{
3918	Log4(("Setting NMI pending for injection\n"));
3919	SVMEVENT Event;
3920	Event.u = 0;
3921	Event.n.u1Valid = 1;
3922	Event.n.u8Vector = X86_XCPT_NMI;
3923	Event.n.u3Type = SVM_EVENT_NMI;
3924	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3925	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INTERRUPT_NMI);
3926	}
3927	else if (!fGif)
3928	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_STGI);
3929	else
3930	hmR0SvmSetIntWindowExiting(pVCpu, pVmcb, pCtx);
3931	}
3932	/*
3933	* Check if the guest can receive external interrupts (PIC/APIC). Once PDMGetInterrupt()
3934	* returns a valid interrupt we -must- deliver the interrupt. We can no longer re-request
3935	* it from the APIC device.
3936	*/
3937	else if ( VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC)
3938	&& !pVCpu->hm.s.fSingleInstruction)
3939	{
3940	if ( fGif
3941	&& !fBlockInt
3942	&& !fIntShadow)
3943	{
3944	uint8_t u8Interrupt;
3945	int rc = PDMGetInterrupt(pVCpu, &u8Interrupt);
3946	if (RT_SUCCESS(rc))
3947	{
3948	Log4(("Setting external interrupt %#x pending for injection\n", u8Interrupt));
3949	SVMEVENT Event;
3950	Event.u = 0;
3951	Event.n.u1Valid = 1;
3952	Event.n.u8Vector = u8Interrupt;
3953	Event.n.u3Type = SVM_EVENT_EXTERNAL_IRQ;
3954	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3955	}
3956	else if (rc == VERR_APIC_INTR_MASKED_BY_TPR)
3957	{
3958	/*
3959	* AMD-V has no TPR thresholding feature. TPR and the force-flag will be
3960	* updated eventually when the TPR is written by the guest.
3961	*/
3962	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchTprMaskedIrq);
3963	}
3964	else
3965	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchGuestIrq);
3966	}
3967	else if (!fGif)
3968	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_STGI);
3969	else
3970	hmR0SvmSetIntWindowExiting(pVCpu, pVmcb, pCtx);
3971	}
3972	}
3973
3974
3975	/**
3976	* Injects any pending events into the guest (or nested-guest).
3977	*
3978	* @param pVCpu The cross context virtual CPU structure.
3979	* @param pCtx Pointer to the guest-CPU context.
3980	* @param pVmcb Pointer to the VM control block.
3981	*
3982	* @remarks Must only be called when we are guaranteed to enter
3983	* hardware-assisted SVM execution and not return to ring-3
3984	* prematurely.
3985	*/
3986	static void hmR0SvmInjectPendingEvent(PVMCPU pVCpu, PCCPUMCTX pCtx, PSVMVMCB pVmcb)
3987	{
3988	Assert(!TRPMHasTrap(pVCpu));
3989	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
3990
3991	bool const fIntShadow = hmR0SvmIsIntrShadowActive(pVCpu, pCtx);
3992	#ifdef VBOX_STRICT
3993	bool const fGif = pCtx->hwvirt.fGif;
3994	bool fAllowInt = fGif;
3995	if (fGif)
3996	{
3997	/*
3998	* For nested-guests we have no way to determine if we're injecting a physical or
3999	* virtual interrupt at this point. Hence the partial verification below.
4000	*/
4001	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
4002	fAllowInt = CPUMCanSvmNstGstTakePhysIntr(pVCpu, pCtx) \|\| CPUMCanSvmNstGstTakeVirtIntr(pVCpu, pCtx);
4003	else
4004	fAllowInt = RT_BOOL(pCtx->eflags.u32 & X86_EFL_IF);
4005	}
4006	#endif
4007
4008	if (pVCpu->hm.s.Event.fPending)
4009	{
4010	SVMEVENT Event;
4011	Event.u = pVCpu->hm.s.Event.u64IntInfo;
4012	Assert(Event.n.u1Valid);
4013
4014	/*
4015	* Validate event injection pre-conditions.
4016	*/
4017	if (Event.n.u3Type == SVM_EVENT_EXTERNAL_IRQ)
4018	{
4019	Assert(fAllowInt);
4020	Assert(!fIntShadow);
4021	}
4022	else if (Event.n.u3Type == SVM_EVENT_NMI)
4023	{
4024	Assert(fGif);
4025	Assert(!fIntShadow);
4026	}
4027
4028	/*
4029	* Before injecting an NMI we must set VMCPU_FF_BLOCK_NMIS to prevent nested NMIs. We
4030	* do this only when we are surely going to inject the NMI as otherwise if we return
4031	* to ring-3 prematurely we could leave NMIs blocked indefinitely upon re-entry into
4032	* SVM R0.
4033	*
4034	* With VT-x, this is handled by the Guest interruptibility information VMCS field
4035	* which will set the VMCS field after actually delivering the NMI which we read on
4036	* VM-exit to determine the state.
4037	*/
4038	if ( Event.n.u3Type == SVM_EVENT_NMI
4039	&& Event.n.u8Vector == X86_XCPT_NMI
4040	&& !VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_BLOCK_NMIS))
4041	{
4042	VMCPU_FF_SET(pVCpu, VMCPU_FF_BLOCK_NMIS);
4043	}
4044
4045	/*
4046	* Inject it (update VMCB for injection by the hardware).
4047	*/
4048	Log4(("Injecting pending HM event\n"));
4049	hmR0SvmInjectEventVmcb(pVCpu, pVmcb, &Event);
4050	pVCpu->hm.s.Event.fPending = false;
4051
4052	if (Event.n.u3Type == SVM_EVENT_EXTERNAL_IRQ)
4053	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectInterrupt);
4054	else
4055	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectXcpt);
4056	}
4057	else
4058	Assert(pVmcb->ctrl.EventInject.n.u1Valid == 0);
4059
4060	/*
4061	* We could have injected an NMI through IEM and continue guest execution using
4062	* hardware-assisted SVM. In which case, we would not have any events pending (above)
4063	* but we still need to intercept IRET in order to eventually clear NMI inhibition.
4064	*/
4065	if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_BLOCK_NMIS))
4066	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_IRET);
4067
4068	/*
4069	* Update the guest interrupt shadow in the guest (or nested-guest) VMCB.
4070	*
4071	* For nested-guests: We need to update it too for the scenario where IEM executes
4072	* the nested-guest but execution later continues here with an interrupt shadow active.
4073	*/
4074	pVmcb->ctrl.IntShadow.n.u1IntShadow = fIntShadow;
4075	}
4076
4077
4078	/**
4079	* Reports world-switch error and dumps some useful debug info.
4080	*
4081	* @param pVM The cross context VM structure.
4082	* @param pVCpu The cross context virtual CPU structure.
4083	* @param rcVMRun The return code from VMRUN (or
4084	* VERR_SVM_INVALID_GUEST_STATE for invalid
4085	* guest-state).
4086	* @param pCtx Pointer to the guest-CPU context.
4087	*/
4088	static void hmR0SvmReportWorldSwitchError(PVM pVM, PVMCPU pVCpu, int rcVMRun, PCPUMCTX pCtx)
4089	{
4090	NOREF(pCtx);
4091	HMSVM_ASSERT_PREEMPT_SAFE();
4092	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(pCtx);
4093	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
4094
4095	PCSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
4096	if (rcVMRun == VERR_SVM_INVALID_GUEST_STATE)
4097	{
4098	hmR0DumpRegs(pVM, pVCpu, pCtx); NOREF(pVM);
4099	/** @todo We probably don't need to dump this anymore or we can expand
4100	* hmR0DumpRegs()? */
4101	#ifdef VBOX_STRICT
4102	Log4(("ctrl.u32VmcbCleanBits %#RX32\n", pVmcb->ctrl.u32VmcbCleanBits));
4103	Log4(("ctrl.u16InterceptRdCRx %#x\n", pVmcb->ctrl.u16InterceptRdCRx));
4104	Log4(("ctrl.u16InterceptWrCRx %#x\n", pVmcb->ctrl.u16InterceptWrCRx));
4105	Log4(("ctrl.u16InterceptRdDRx %#x\n", pVmcb->ctrl.u16InterceptRdDRx));
4106	Log4(("ctrl.u16InterceptWrDRx %#x\n", pVmcb->ctrl.u16InterceptWrDRx));
4107	Log4(("ctrl.u32InterceptXcpt %#x\n", pVmcb->ctrl.u32InterceptXcpt));
4108	Log4(("ctrl.u64InterceptCtrl %#RX64\n", pVmcb->ctrl.u64InterceptCtrl));
4109	Log4(("ctrl.u64IOPMPhysAddr %#RX64\n", pVmcb->ctrl.u64IOPMPhysAddr));
4110	Log4(("ctrl.u64MSRPMPhysAddr %#RX64\n", pVmcb->ctrl.u64MSRPMPhysAddr));
4111	Log4(("ctrl.u64TSCOffset %#RX64\n", pVmcb->ctrl.u64TSCOffset));
4112
4113	Log4(("ctrl.TLBCtrl.u32ASID %#x\n", pVmcb->ctrl.TLBCtrl.n.u32ASID));
4114	Log4(("ctrl.TLBCtrl.u8TLBFlush %#x\n", pVmcb->ctrl.TLBCtrl.n.u8TLBFlush));
4115	Log4(("ctrl.TLBCtrl.u24Reserved %#x\n", pVmcb->ctrl.TLBCtrl.n.u24Reserved));
4116
4117	Log4(("ctrl.IntCtrl.u8VTPR %#x\n", pVmcb->ctrl.IntCtrl.n.u8VTPR));
4118	Log4(("ctrl.IntCtrl.u1VIrqPending %#x\n", pVmcb->ctrl.IntCtrl.n.u1VIrqPending));
4119	Log4(("ctrl.IntCtrl.u1VGif %#x\n", pVmcb->ctrl.IntCtrl.n.u1VGif));
4120	Log4(("ctrl.IntCtrl.u6Reserved0 %#x\n", pVmcb->ctrl.IntCtrl.n.u6Reserved));
4121	Log4(("ctrl.IntCtrl.u4VIntrPrio %#x\n", pVmcb->ctrl.IntCtrl.n.u4VIntrPrio));
4122	Log4(("ctrl.IntCtrl.u1IgnoreTPR %#x\n", pVmcb->ctrl.IntCtrl.n.u1IgnoreTPR));
4123	Log4(("ctrl.IntCtrl.u3Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u3Reserved));
4124	Log4(("ctrl.IntCtrl.u1VIntrMasking %#x\n", pVmcb->ctrl.IntCtrl.n.u1VIntrMasking));
4125	Log4(("ctrl.IntCtrl.u1VGifEnable %#x\n", pVmcb->ctrl.IntCtrl.n.u1VGifEnable));
4126	Log4(("ctrl.IntCtrl.u5Reserved1 %#x\n", pVmcb->ctrl.IntCtrl.n.u5Reserved));
4127	Log4(("ctrl.IntCtrl.u8VIntrVector %#x\n", pVmcb->ctrl.IntCtrl.n.u8VIntrVector));
4128	Log4(("ctrl.IntCtrl.u24Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u24Reserved));
4129
4130	Log4(("ctrl.IntShadow.u1IntShadow %#x\n", pVmcb->ctrl.IntShadow.n.u1IntShadow));
4131	Log4(("ctrl.IntShadow.u1GuestIntMask %#x\n", pVmcb->ctrl.IntShadow.n.u1GuestIntMask));
4132	Log4(("ctrl.u64ExitCode %#RX64\n", pVmcb->ctrl.u64ExitCode));
4133	Log4(("ctrl.u64ExitInfo1 %#RX64\n", pVmcb->ctrl.u64ExitInfo1));
4134	Log4(("ctrl.u64ExitInfo2 %#RX64\n", pVmcb->ctrl.u64ExitInfo2));
4135	Log4(("ctrl.ExitIntInfo.u8Vector %#x\n", pVmcb->ctrl.ExitIntInfo.n.u8Vector));
4136	Log4(("ctrl.ExitIntInfo.u3Type %#x\n", pVmcb->ctrl.ExitIntInfo.n.u3Type));
4137	Log4(("ctrl.ExitIntInfo.u1ErrorCodeValid %#x\n", pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid));
4138	Log4(("ctrl.ExitIntInfo.u19Reserved %#x\n", pVmcb->ctrl.ExitIntInfo.n.u19Reserved));
4139	Log4(("ctrl.ExitIntInfo.u1Valid %#x\n", pVmcb->ctrl.ExitIntInfo.n.u1Valid));
4140	Log4(("ctrl.ExitIntInfo.u32ErrorCode %#x\n", pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode));
4141	Log4(("ctrl.NestedPagingCtrl.u1NestedPaging %#x\n", pVmcb->ctrl.NestedPagingCtrl.n.u1NestedPaging));
4142	Log4(("ctrl.NestedPagingCtrl.u1Sev %#x\n", pVmcb->ctrl.NestedPagingCtrl.n.u1Sev));
4143	Log4(("ctrl.NestedPagingCtrl.u1SevEs %#x\n", pVmcb->ctrl.NestedPagingCtrl.n.u1SevEs));
4144	Log4(("ctrl.EventInject.u8Vector %#x\n", pVmcb->ctrl.EventInject.n.u8Vector));
4145	Log4(("ctrl.EventInject.u3Type %#x\n", pVmcb->ctrl.EventInject.n.u3Type));
4146	Log4(("ctrl.EventInject.u1ErrorCodeValid %#x\n", pVmcb->ctrl.EventInject.n.u1ErrorCodeValid));
4147	Log4(("ctrl.EventInject.u19Reserved %#x\n", pVmcb->ctrl.EventInject.n.u19Reserved));
4148	Log4(("ctrl.EventInject.u1Valid %#x\n", pVmcb->ctrl.EventInject.n.u1Valid));
4149	Log4(("ctrl.EventInject.u32ErrorCode %#x\n", pVmcb->ctrl.EventInject.n.u32ErrorCode));
4150
4151	Log4(("ctrl.u64NestedPagingCR3 %#RX64\n", pVmcb->ctrl.u64NestedPagingCR3));
4152
4153	Log4(("ctrl.LbrVirt.u1LbrVirt %#x\n", pVmcb->ctrl.LbrVirt.n.u1LbrVirt));
4154	Log4(("ctrl.LbrVirt.u1VirtVmsaveVmload %#x\n", pVmcb->ctrl.LbrVirt.n.u1VirtVmsaveVmload));
4155
4156	Log4(("guest.CS.u16Sel %RTsel\n", pVmcb->guest.CS.u16Sel));
4157	Log4(("guest.CS.u16Attr %#x\n", pVmcb->guest.CS.u16Attr));
4158	Log4(("guest.CS.u32Limit %#RX32\n", pVmcb->guest.CS.u32Limit));
4159	Log4(("guest.CS.u64Base %#RX64\n", pVmcb->guest.CS.u64Base));
4160	Log4(("guest.DS.u16Sel %#RTsel\n", pVmcb->guest.DS.u16Sel));
4161	Log4(("guest.DS.u16Attr %#x\n", pVmcb->guest.DS.u16Attr));
4162	Log4(("guest.DS.u32Limit %#RX32\n", pVmcb->guest.DS.u32Limit));
4163	Log4(("guest.DS.u64Base %#RX64\n", pVmcb->guest.DS.u64Base));
4164	Log4(("guest.ES.u16Sel %RTsel\n", pVmcb->guest.ES.u16Sel));
4165	Log4(("guest.ES.u16Attr %#x\n", pVmcb->guest.ES.u16Attr));
4166	Log4(("guest.ES.u32Limit %#RX32\n", pVmcb->guest.ES.u32Limit));
4167	Log4(("guest.ES.u64Base %#RX64\n", pVmcb->guest.ES.u64Base));
4168	Log4(("guest.FS.u16Sel %RTsel\n", pVmcb->guest.FS.u16Sel));
4169	Log4(("guest.FS.u16Attr %#x\n", pVmcb->guest.FS.u16Attr));
4170	Log4(("guest.FS.u32Limit %#RX32\n", pVmcb->guest.FS.u32Limit));
4171	Log4(("guest.FS.u64Base %#RX64\n", pVmcb->guest.FS.u64Base));
4172	Log4(("guest.GS.u16Sel %RTsel\n", pVmcb->guest.GS.u16Sel));
4173	Log4(("guest.GS.u16Attr %#x\n", pVmcb->guest.GS.u16Attr));
4174	Log4(("guest.GS.u32Limit %#RX32\n", pVmcb->guest.GS.u32Limit));
4175	Log4(("guest.GS.u64Base %#RX64\n", pVmcb->guest.GS.u64Base));
4176
4177	Log4(("guest.GDTR.u32Limit %#RX32\n", pVmcb->guest.GDTR.u32Limit));
4178	Log4(("guest.GDTR.u64Base %#RX64\n", pVmcb->guest.GDTR.u64Base));
4179
4180	Log4(("guest.LDTR.u16Sel %RTsel\n", pVmcb->guest.LDTR.u16Sel));
4181	Log4(("guest.LDTR.u16Attr %#x\n", pVmcb->guest.LDTR.u16Attr));
4182	Log4(("guest.LDTR.u32Limit %#RX32\n", pVmcb->guest.LDTR.u32Limit));
4183	Log4(("guest.LDTR.u64Base %#RX64\n", pVmcb->guest.LDTR.u64Base));
4184
4185	Log4(("guest.IDTR.u32Limit %#RX32\n", pVmcb->guest.IDTR.u32Limit));
4186	Log4(("guest.IDTR.u64Base %#RX64\n", pVmcb->guest.IDTR.u64Base));
4187
4188	Log4(("guest.TR.u16Sel %RTsel\n", pVmcb->guest.TR.u16Sel));
4189	Log4(("guest.TR.u16Attr %#x\n", pVmcb->guest.TR.u16Attr));
4190	Log4(("guest.TR.u32Limit %#RX32\n", pVmcb->guest.TR.u32Limit));
4191	Log4(("guest.TR.u64Base %#RX64\n", pVmcb->guest.TR.u64Base));
4192
4193	Log4(("guest.u8CPL %#x\n", pVmcb->guest.u8CPL));
4194	Log4(("guest.u64CR0 %#RX64\n", pVmcb->guest.u64CR0));
4195	Log4(("guest.u64CR2 %#RX64\n", pVmcb->guest.u64CR2));
4196	Log4(("guest.u64CR3 %#RX64\n", pVmcb->guest.u64CR3));
4197	Log4(("guest.u64CR4 %#RX64\n", pVmcb->guest.u64CR4));
4198	Log4(("guest.u64DR6 %#RX64\n", pVmcb->guest.u64DR6));
4199	Log4(("guest.u64DR7 %#RX64\n", pVmcb->guest.u64DR7));
4200
4201	Log4(("guest.u64RIP %#RX64\n", pVmcb->guest.u64RIP));
4202	Log4(("guest.u64RSP %#RX64\n", pVmcb->guest.u64RSP));
4203	Log4(("guest.u64RAX %#RX64\n", pVmcb->guest.u64RAX));
4204	Log4(("guest.u64RFlags %#RX64\n", pVmcb->guest.u64RFlags));
4205
4206	Log4(("guest.u64SysEnterCS %#RX64\n", pVmcb->guest.u64SysEnterCS));
4207	Log4(("guest.u64SysEnterEIP %#RX64\n", pVmcb->guest.u64SysEnterEIP));
4208	Log4(("guest.u64SysEnterESP %#RX64\n", pVmcb->guest.u64SysEnterESP));
4209
4210	Log4(("guest.u64EFER %#RX64\n", pVmcb->guest.u64EFER));
4211	Log4(("guest.u64STAR %#RX64\n", pVmcb->guest.u64STAR));
4212	Log4(("guest.u64LSTAR %#RX64\n", pVmcb->guest.u64LSTAR));
4213	Log4(("guest.u64CSTAR %#RX64\n", pVmcb->guest.u64CSTAR));
4214	Log4(("guest.u64SFMASK %#RX64\n", pVmcb->guest.u64SFMASK));
4215	Log4(("guest.u64KernelGSBase %#RX64\n", pVmcb->guest.u64KernelGSBase));
4216	Log4(("guest.u64PAT %#RX64\n", pVmcb->guest.u64PAT));
4217	Log4(("guest.u64DBGCTL %#RX64\n", pVmcb->guest.u64DBGCTL));
4218	Log4(("guest.u64BR_FROM %#RX64\n", pVmcb->guest.u64BR_FROM));
4219	Log4(("guest.u64BR_TO %#RX64\n", pVmcb->guest.u64BR_TO));
4220	Log4(("guest.u64LASTEXCPFROM %#RX64\n", pVmcb->guest.u64LASTEXCPFROM));
4221	Log4(("guest.u64LASTEXCPTO %#RX64\n", pVmcb->guest.u64LASTEXCPTO));
4222	#endif /* VBOX_STRICT */
4223	}
4224	else
4225	Log4Func(("rcVMRun=%d\n", rcVMRun));
4226
4227	NOREF(pVmcb);
4228	}
4229
4230
4231	/**
4232	* Check per-VM and per-VCPU force flag actions that require us to go back to
4233	* ring-3 for one reason or another.
4234	*
4235	* @returns VBox status code (information status code included).
4236	* @retval VINF_SUCCESS if we don't have any actions that require going back to
4237	* ring-3.
4238	* @retval VINF_PGM_SYNC_CR3 if we have pending PGM CR3 sync.
4239	* @retval VINF_EM_PENDING_REQUEST if we have pending requests (like hardware
4240	* interrupts)
4241	* @retval VINF_PGM_POOL_FLUSH_PENDING if PGM is doing a pool flush and requires
4242	* all EMTs to be in ring-3.
4243	* @retval VINF_EM_RAW_TO_R3 if there is pending DMA requests.
4244	* @retval VINF_EM_NO_MEMORY PGM is out of memory, we need to return
4245	* to the EM loop.
4246	*
4247	* @param pVM The cross context VM structure.
4248	* @param pVCpu The cross context virtual CPU structure.
4249	* @param pCtx Pointer to the guest-CPU context.
4250	*/
4251	static int hmR0SvmCheckForceFlags(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
4252	{
4253	Assert(VMMRZCallRing3IsEnabled(pVCpu));
4254	Assert(!VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_UPDATE_PAE_PDPES));
4255
4256	/* Could happen as a result of longjump. */
4257	if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_UPDATE_CR3))
4258	PGMUpdateCR3(pVCpu, CPUMGetGuestCR3(pVCpu));
4259
4260	/* Update pending interrupts into the APIC's IRR. */
4261	if (VMCPU_FF_TEST_AND_CLEAR(pVCpu, VMCPU_FF_UPDATE_APIC))
4262	APICUpdatePendingInterrupts(pVCpu);
4263
4264	if ( VM_FF_IS_PENDING(pVM, !pVCpu->hm.s.fSingleInstruction
4265	? VM_FF_HP_R0_PRE_HM_MASK : VM_FF_HP_R0_PRE_HM_STEP_MASK)
4266	\|\| VMCPU_FF_IS_PENDING(pVCpu, !pVCpu->hm.s.fSingleInstruction
4267	? VMCPU_FF_HP_R0_PRE_HM_MASK : VMCPU_FF_HP_R0_PRE_HM_STEP_MASK) )
4268	{
4269	/* Pending PGM C3 sync. */
4270	if (VMCPU_FF_IS_PENDING(pVCpu,VMCPU_FF_PGM_SYNC_CR3 \| VMCPU_FF_PGM_SYNC_CR3_NON_GLOBAL))
4271	{
4272	int rc = PGMSyncCR3(pVCpu, pCtx->cr0, pCtx->cr3, pCtx->cr4, VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_PGM_SYNC_CR3));
4273	if (rc != VINF_SUCCESS)
4274	{
4275	Log4Func(("PGMSyncCR3 forcing us back to ring-3. rc=%d\n", rc));
4276	return rc;
4277	}
4278	}
4279
4280	/* Pending HM-to-R3 operations (critsects, timers, EMT rendezvous etc.) */
4281	/* -XXX- what was that about single stepping? */
4282	if ( VM_FF_IS_PENDING(pVM, VM_FF_HM_TO_R3_MASK)
4283	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_TO_R3_MASK))
4284	{
4285	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHmToR3FF);
4286	int rc = RT_UNLIKELY(VM_FF_IS_PENDING(pVM, VM_FF_PGM_NO_MEMORY)) ? VINF_EM_NO_MEMORY : VINF_EM_RAW_TO_R3;
4287	Log4Func(("HM_TO_R3 forcing us back to ring-3. rc=%d\n", rc));
4288	return rc;
4289	}
4290
4291	/* Pending VM request packets, such as hardware interrupts. */
4292	if ( VM_FF_IS_PENDING(pVM, VM_FF_REQUEST)
4293	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_REQUEST))
4294	{
4295	Log4Func(("Pending VM request forcing us back to ring-3\n"));
4296	return VINF_EM_PENDING_REQUEST;
4297	}
4298
4299	/* Pending PGM pool flushes. */
4300	if (VM_FF_IS_PENDING(pVM, VM_FF_PGM_POOL_FLUSH_PENDING))
4301	{
4302	Log4Func(("PGM pool flush pending forcing us back to ring-3\n"));
4303	return VINF_PGM_POOL_FLUSH_PENDING;
4304	}
4305
4306	/* Pending DMA requests. */
4307	if (VM_FF_IS_PENDING(pVM, VM_FF_PDM_DMA))
4308	{
4309	Log4Func(("Pending DMA request forcing us back to ring-3\n"));
4310	return VINF_EM_RAW_TO_R3;
4311	}
4312	}
4313
4314	return VINF_SUCCESS;
4315	}
4316
4317
4318	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
4319	/**
4320	* Does the preparations before executing nested-guest code in AMD-V.
4321	*
4322	* @returns VBox status code (informational status codes included).
4323	* @retval VINF_SUCCESS if we can proceed with running the guest.
4324	* @retval VINF_* scheduling changes, we have to go back to ring-3.
4325	*
4326	* @param pVM The cross context VM structure.
4327	* @param pVCpu The cross context virtual CPU structure.
4328	* @param pCtx Pointer to the nested-guest-CPU context.
4329	* @param pSvmTransient Pointer to the SVM transient structure.
4330	*
4331	* @remarks Same caveats regarding longjumps as hmR0SvmPreRunGuest applies.
4332	* @sa hmR0SvmPreRunGuest.
4333	*/
4334	static int hmR0SvmPreRunGuestNested(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4335	{
4336	HMSVM_ASSERT_PREEMPT_SAFE();
4337	HMSVM_ASSERT_IN_NESTED_GUEST(pCtx);
4338
4339	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM_ONLY_IN_IEM
4340	Log2(("hmR0SvmPreRunGuest: Rescheduling to IEM due to nested-hwvirt or forced IEM exec -> VINF_EM_RESCHEDULE_REM\n"));
4341	return VINF_EM_RESCHEDULE_REM;
4342	#endif
4343
4344	/* Check force flag actions that might require us to go back to ring-3. */
4345	int rc = hmR0SvmCheckForceFlags(pVM, pVCpu, pCtx);
4346	if (rc != VINF_SUCCESS)
4347	return rc;
4348
4349	if (TRPMHasTrap(pVCpu))
4350	hmR0SvmTrpmTrapToPendingEvent(pVCpu);
4351	else if (!pVCpu->hm.s.Event.fPending)
4352	{
4353	VBOXSTRICTRC rcStrict = hmR0SvmEvaluatePendingEventNested(pVCpu, pCtx);
4354	if ( rcStrict != VINF_SUCCESS
4355	\|\| !CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
4356	return VBOXSTRICTRC_VAL(rcStrict);
4357	}
4358
4359	HMSVM_ASSERT_IN_NESTED_GUEST(pCtx);
4360
4361	/*
4362	* On the oldest AMD-V systems, we may not get enough information to reinject an NMI.
4363	* Just do it in software, see @bugref{8411}.
4364	* NB: If we could continue a task switch exit we wouldn't need to do this.
4365	*/
4366	if (RT_UNLIKELY( !pVM->hm.s.svm.u32Features
4367	&& pVCpu->hm.s.Event.fPending
4368	&& SVM_EVENT_GET_TYPE(pVCpu->hm.s.Event.u64IntInfo) == SVM_EVENT_NMI))
4369	{
4370	return VINF_EM_RAW_INJECT_TRPM_EVENT;
4371	}
4372
4373	#ifdef HMSVM_SYNC_FULL_NESTED_GUEST_STATE
4374	Assert(!(pCtx->fExtrn & HMSVM_CPUMCTX_EXTRN_ALL));
4375	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
4376	#endif
4377
4378	/*
4379	* Export the nested-guest state bits that are not shared with the host in any way as we
4380	* can longjmp or get preempted in the midst of exporting some of the state.
4381	*/
4382	rc = hmR0SvmExportGuestStateNested(pVCpu);
4383	AssertRCReturn(rc, rc);
4384	STAM_COUNTER_INC(&pVCpu->hm.s.StatExportFull);
4385
4386	/* Ensure we've cached (and hopefully modified) the VMCB for execution using hardware-assisted SVM. */
4387	Assert(pVCpu->hm.s.svm.NstGstVmcbCache.fCacheValid);
4388
4389	/*
4390	* No longjmps to ring-3 from this point on!!!
4391	*
4392	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional,
4393	* better than a kernel panic. This also disables flushing of the R0-logger instance.
4394	*/
4395	VMMRZCallRing3Disable(pVCpu);
4396
4397	/*
4398	* We disable interrupts so that we don't miss any interrupts that would flag preemption
4399	* (IPI/timers etc.) when thread-context hooks aren't used and we've been running with
4400	* preemption disabled for a while. Since this is purly to aid the
4401	* RTThreadPreemptIsPending() code, it doesn't matter that it may temporarily reenable and
4402	* disable interrupt on NT.
4403	*
4404	* We need to check for force-flags that could've possible been altered since we last
4405	* checked them (e.g. by PDMGetInterrupt() leaving the PDM critical section,
4406	* see @bugref{6398}).
4407	*
4408	* We also check a couple of other force-flags as a last opportunity to get the EMT back
4409	* to ring-3 before executing guest code.
4410	*/
4411	pSvmTransient->fEFlags = ASMIntDisableFlags();
4412	if ( VM_FF_IS_PENDING(pVM, VM_FF_EMT_RENDEZVOUS \| VM_FF_TM_VIRTUAL_SYNC)
4413	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_TO_R3_MASK))
4414	{
4415	ASMSetFlags(pSvmTransient->fEFlags);
4416	VMMRZCallRing3Enable(pVCpu);
4417	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHmToR3FF);
4418	return VINF_EM_RAW_TO_R3;
4419	}
4420	if (RTThreadPreemptIsPending(NIL_RTTHREAD))
4421	{
4422	ASMSetFlags(pSvmTransient->fEFlags);
4423	VMMRZCallRing3Enable(pVCpu);
4424	STAM_COUNTER_INC(&pVCpu->hm.s.StatPendingHostIrq);
4425	return VINF_EM_RAW_INTERRUPT;
4426	}
4427	return VINF_SUCCESS;
4428	}
4429	#endif
4430
4431
4432	/**
4433	* Does the preparations before executing guest code in AMD-V.
4434	*
4435	* This may cause longjmps to ring-3 and may even result in rescheduling to the
4436	* recompiler. We must be cautious what we do here regarding committing
4437	* guest-state information into the VMCB assuming we assuredly execute the guest
4438	* in AMD-V. If we fall back to the recompiler after updating the VMCB and
4439	* clearing the common-state (TRPM/forceflags), we must undo those changes so
4440	* that the recompiler can (and should) use them when it resumes guest
4441	* execution. Otherwise such operations must be done when we can no longer
4442	* exit to ring-3.
4443	*
4444	* @returns VBox status code (informational status codes included).
4445	* @retval VINF_SUCCESS if we can proceed with running the guest.
4446	* @retval VINF_* scheduling changes, we have to go back to ring-3.
4447	*
4448	* @param pVM The cross context VM structure.
4449	* @param pVCpu The cross context virtual CPU structure.
4450	* @param pCtx Pointer to the guest-CPU context.
4451	* @param pSvmTransient Pointer to the SVM transient structure.
4452	*/
4453	static int hmR0SvmPreRunGuest(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4454	{
4455	HMSVM_ASSERT_PREEMPT_SAFE();
4456	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(pCtx);
4457
4458	/* Check force flag actions that might require us to go back to ring-3. */
4459	int rc = hmR0SvmCheckForceFlags(pVM, pVCpu, pCtx);
4460	if (rc != VINF_SUCCESS)
4461	return rc;
4462
4463	if (TRPMHasTrap(pVCpu))
4464	hmR0SvmTrpmTrapToPendingEvent(pVCpu);
4465	else if (!pVCpu->hm.s.Event.fPending)
4466	hmR0SvmEvaluatePendingEvent(pVCpu, pCtx);
4467
4468	/*
4469	* On the oldest AMD-V systems, we may not get enough information to reinject an NMI.
4470	* Just do it in software, see @bugref{8411}.
4471	* NB: If we could continue a task switch exit we wouldn't need to do this.
4472	*/
4473	if (RT_UNLIKELY(pVCpu->hm.s.Event.fPending && (((pVCpu->hm.s.Event.u64IntInfo >> 8) & 7) == SVM_EVENT_NMI)))
4474	if (RT_UNLIKELY(!pVM->hm.s.svm.u32Features))
4475	return VINF_EM_RAW_INJECT_TRPM_EVENT;
4476
4477	#ifdef HMSVM_SYNC_FULL_GUEST_STATE
4478	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
4479	{
4480	Assert(!(pCtx->fExtrn & HMSVM_CPUMCTX_EXTRN_ALL));
4481	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
4482	}
4483	#endif
4484	#ifdef HMSVM_SYNC_FULL_NESTED_GUEST_STATE
4485	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
4486	{
4487	Assert(!(pCtx->fExtrn & HMSVM_CPUMCTX_EXTRN_ALL));
4488	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
4489	}
4490	#endif
4491
4492	/*
4493	* Export the guest state bits that are not shared with the host in any way as we can
4494	* longjmp or get preempted in the midst of exporting some of the state.
4495	*/
4496	rc = hmR0SvmExportGuestState(pVCpu);
4497	AssertRCReturn(rc, rc);
4498	STAM_COUNTER_INC(&pVCpu->hm.s.StatExportFull);
4499
4500	/*
4501	* If we're not intercepting TPR changes in the guest, save the guest TPR before the
4502	* world-switch so we can update it on the way back if the guest changed the TPR.
4503	*/
4504	if (pVCpu->hm.s.svm.fSyncVTpr)
4505	{
4506	PCSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
4507	if (pVM->hm.s.fTPRPatchingActive)
4508	pSvmTransient->u8GuestTpr = pVmcb->guest.u64LSTAR;
4509	else
4510	pSvmTransient->u8GuestTpr = pVmcb->ctrl.IntCtrl.n.u8VTPR;
4511	}
4512
4513	/*
4514	* No longjmps to ring-3 from this point on!!!
4515	*
4516	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional,
4517	* better than a kernel panic. This also disables flushing of the R0-logger instance.
4518	*/
4519	VMMRZCallRing3Disable(pVCpu);
4520
4521	/*
4522	* We disable interrupts so that we don't miss any interrupts that would flag preemption
4523	* (IPI/timers etc.) when thread-context hooks aren't used and we've been running with
4524	* preemption disabled for a while. Since this is purly to aid the
4525	* RTThreadPreemptIsPending() code, it doesn't matter that it may temporarily reenable and
4526	* disable interrupt on NT.
4527	*
4528	* We need to check for force-flags that could've possible been altered since we last
4529	* checked them (e.g. by PDMGetInterrupt() leaving the PDM critical section,
4530	* see @bugref{6398}).
4531	*
4532	* We also check a couple of other force-flags as a last opportunity to get the EMT back
4533	* to ring-3 before executing guest code.
4534	*/
4535	pSvmTransient->fEFlags = ASMIntDisableFlags();
4536	if ( VM_FF_IS_PENDING(pVM, VM_FF_EMT_RENDEZVOUS \| VM_FF_TM_VIRTUAL_SYNC)
4537	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_TO_R3_MASK))
4538	{
4539	ASMSetFlags(pSvmTransient->fEFlags);
4540	VMMRZCallRing3Enable(pVCpu);
4541	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHmToR3FF);
4542	return VINF_EM_RAW_TO_R3;
4543	}
4544	if (RTThreadPreemptIsPending(NIL_RTTHREAD))
4545	{
4546	ASMSetFlags(pSvmTransient->fEFlags);
4547	VMMRZCallRing3Enable(pVCpu);
4548	STAM_COUNTER_INC(&pVCpu->hm.s.StatPendingHostIrq);
4549	return VINF_EM_RAW_INTERRUPT;
4550	}
4551
4552	return VINF_SUCCESS;
4553	}
4554
4555
4556	/**
4557	* Prepares to run guest (or nested-guest) code in AMD-V and we've committed to
4558	* doing so.
4559	*
4560	* This means there is no backing out to ring-3 or anywhere else at this point.
4561	*
4562	* @param pVCpu The cross context virtual CPU structure.
4563	* @param pCtx Pointer to the guest-CPU context.
4564	* @param pSvmTransient Pointer to the SVM transient structure.
4565	*
4566	* @remarks Called with preemption disabled.
4567	* @remarks No-long-jump zone!!!
4568	*/
4569	static void hmR0SvmPreRunGuestCommitted(PVMCPU pVCpu, PCCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4570	{
4571	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
4572	Assert(VMMR0IsLogFlushDisabled(pVCpu));
4573	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
4574
4575	VMCPU_ASSERT_STATE(pVCpu, VMCPUSTATE_STARTED_HM);
4576	VMCPU_SET_STATE(pVCpu, VMCPUSTATE_STARTED_EXEC); /* Indicate the start of guest execution. */
4577
4578	PVM pVM = pVCpu->CTX_SUFF(pVM);
4579	PSVMVMCB pVmcb = pSvmTransient->pVmcb;
4580
4581	hmR0SvmInjectPendingEvent(pVCpu, pCtx, pVmcb);
4582
4583	if (!CPUMIsGuestFPUStateActive(pVCpu))
4584	{
4585	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatLoadGuestFpuState, x);
4586	CPUMR0LoadGuestFPU(pVM, pVCpu); /* (Ignore rc, no need to set HM_CHANGED_HOST_CONTEXT for SVM.) */
4587	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatLoadGuestFpuState, x);
4588	STAM_COUNTER_INC(&pVCpu->hm.s.StatLoadGuestFpu);
4589	}
4590
4591	/* Load the state shared between host and guest (FPU, debug). */
4592	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE)
4593	hmR0SvmExportSharedState(pVCpu, pVmcb);
4594
4595	pVCpu->hm.s.fCtxChanged &= ~HM_CHANGED_HOST_CONTEXT; /* Preemption might set this, nothing to do on AMD-V. */
4596	AssertMsg(!pVCpu->hm.s.fCtxChanged, ("fCtxChanged=%#RX64\n", pVCpu->hm.s.fCtxChanged));
4597
4598	PHMGLOBALCPUINFO pHostCpu = hmR0GetCurrentCpu();
4599	RTCPUID const idHostCpu = pHostCpu->idCpu;
4600	bool const fMigratedHostCpu = idHostCpu != pVCpu->hm.s.idLastCpu;
4601
4602	/* Setup TSC offsetting. */
4603	if ( pSvmTransient->fUpdateTscOffsetting
4604	\|\| fMigratedHostCpu)
4605	{
4606	hmR0SvmUpdateTscOffsetting(pVM, pVCpu, pCtx, pVmcb);
4607	pSvmTransient->fUpdateTscOffsetting = false;
4608	}
4609
4610	/* If we've migrating CPUs, mark the VMCB Clean bits as dirty. */
4611	if (fMigratedHostCpu)
4612	pVmcb->ctrl.u32VmcbCleanBits = 0;
4613
4614	/* Store status of the shared guest-host state at the time of VMRUN. */
4615	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS)
4616	if (CPUMIsGuestInLongModeEx(pCtx))
4617	{
4618	pSvmTransient->fWasGuestDebugStateActive = CPUMIsGuestDebugStateActivePending(pVCpu);
4619	pSvmTransient->fWasHyperDebugStateActive = CPUMIsHyperDebugStateActivePending(pVCpu);
4620	}
4621	else
4622	#endif
4623	{
4624	pSvmTransient->fWasGuestDebugStateActive = CPUMIsGuestDebugStateActive(pVCpu);
4625	pSvmTransient->fWasHyperDebugStateActive = CPUMIsHyperDebugStateActive(pVCpu);
4626	}
4627
4628	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
4629	uint8_t *pbMsrBitmap;
4630	if (!pSvmTransient->fIsNestedGuest)
4631	pbMsrBitmap = (uint8_t *)pVCpu->hm.s.svm.pvMsrBitmap;
4632	else
4633	{
4634	hmR0SvmMergeMsrpmNested(pHostCpu, pVCpu, pCtx);
4635
4636	/* Update the nested-guest VMCB with the newly merged MSRPM (clean bits updated below). */
4637	pVmcb->ctrl.u64MSRPMPhysAddr = pHostCpu->n.svm.HCPhysNstGstMsrpm;
4638	pbMsrBitmap = (uint8_t *)pHostCpu->n.svm.pvNstGstMsrpm;
4639	}
4640	#else
4641	uint8_t pbMsrBitmap = (uint8_t )pVCpu->hm.s.svm.pvMsrBitmap;
4642	#endif
4643
4644	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, true); /* Used for TLB flushing, set this across the world switch. */
4645	/* Flush the appropriate tagged-TLB entries. */
4646	hmR0SvmFlushTaggedTlb(pVCpu, pCtx, pVmcb, pHostCpu);
4647	Assert(pVCpu->hm.s.idLastCpu == idHostCpu);
4648
4649	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatEntry, &pVCpu->hm.s.StatInGC, x);
4650
4651	TMNotifyStartOfExecution(pVCpu); /* Finally, notify TM to resume its clocks as we're about
4652	to start executing. */
4653
4654	/*
4655	* Save the current Host TSC_AUX and write the guest TSC_AUX to the host, so that RDTSCPs
4656	* (that don't cause exits) reads the guest MSR, see @bugref{3324}.
4657	*
4658	* This should be done -after- any RDTSCPs for obtaining the host timestamp (TM, STAM etc).
4659	*/
4660	if ( (pVM->hm.s.cpuid.u32AMDFeatureEDX & X86_CPUID_EXT_FEATURE_EDX_RDTSCP)
4661	&& !(pVmcb->ctrl.u64InterceptCtrl & SVM_CTRL_INTERCEPT_RDTSCP))
4662	{
4663	uint64_t const uGuestTscAux = CPUMGetGuestTscAux(pVCpu);
4664	pVCpu->hm.s.u64HostTscAux = ASMRdMsr(MSR_K8_TSC_AUX);
4665	if (uGuestTscAux != pVCpu->hm.s.u64HostTscAux)
4666	ASMWrMsr(MSR_K8_TSC_AUX, uGuestTscAux);
4667	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_K8_TSC_AUX, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
4668	pSvmTransient->fRestoreTscAuxMsr = true;
4669	}
4670	else
4671	{
4672	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_K8_TSC_AUX, SVMMSREXIT_INTERCEPT_READ, SVMMSREXIT_INTERCEPT_WRITE);
4673	pSvmTransient->fRestoreTscAuxMsr = false;
4674	}
4675	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_IOPM_MSRPM;
4676
4677	/*
4678	* If VMCB Clean bits isn't supported by the CPU or exposed to the guest in the nested
4679	* virtualization case, mark all state-bits as dirty indicating to the CPU to re-load
4680	* from the VMCB.
4681	*/
4682	bool const fSupportsVmcbCleanBits = hmR0SvmSupportsVmcbCleanBits(pVCpu, pCtx);
4683	if (!fSupportsVmcbCleanBits)
4684	pVmcb->ctrl.u32VmcbCleanBits = 0;
4685	}
4686
4687
4688	/**
4689	* Wrapper for running the guest code in AMD-V.
4690	*
4691	* @returns VBox strict status code.
4692	* @param pVM The cross context VM structure.
4693	* @param pVCpu The cross context virtual CPU structure.
4694	* @param pCtx Pointer to the guest-CPU context.
4695	*
4696	* @remarks No-long-jump zone!!!
4697	*/
4698	DECLINLINE(int) hmR0SvmRunGuest(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
4699	{
4700	/* Mark that HM is the keeper of all guest-CPU registers now that we're going to execute guest code. */
4701	pCtx->fExtrn \|= HMSVM_CPUMCTX_EXTRN_ALL \| CPUMCTX_EXTRN_KEEPER_HM;
4702
4703	/*
4704	* 64-bit Windows uses XMM registers in the kernel as the Microsoft compiler expresses
4705	* floating-point operations using SSE instructions. Some XMM registers (XMM6-XMM15) are
4706	* callee-saved and thus the need for this XMM wrapper.
4707	*
4708	* Refer MSDN "Configuring Programs for 64-bit/x64 Software Conventions / Register Usage".
4709	*/
4710	#ifdef VBOX_WITH_KERNEL_USING_XMM
4711	return hmR0SVMRunWrapXMM(pVCpu->hm.s.svm.HCPhysVmcbHost, pVCpu->hm.s.svm.HCPhysVmcb, pCtx, pVM, pVCpu,
4712	pVCpu->hm.s.svm.pfnVMRun);
4713	#else
4714	return pVCpu->hm.s.svm.pfnVMRun(pVCpu->hm.s.svm.HCPhysVmcbHost, pVCpu->hm.s.svm.HCPhysVmcb, pCtx, pVM, pVCpu);
4715	#endif
4716	}
4717
4718
4719	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
4720	/**
4721	* Wrapper for running the nested-guest code in AMD-V.
4722	*
4723	* @returns VBox strict status code.
4724	* @param pVM The cross context VM structure.
4725	* @param pVCpu The cross context virtual CPU structure.
4726	* @param pCtx Pointer to the guest-CPU context.
4727	*
4728	* @remarks No-long-jump zone!!!
4729	*/
4730	DECLINLINE(int) hmR0SvmRunGuestNested(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
4731	{
4732	/* Mark that HM is the keeper of all guest-CPU registers now that we're going to execute guest code. */
4733	pCtx->fExtrn \|= HMSVM_CPUMCTX_EXTRN_ALL \| CPUMCTX_EXTRN_KEEPER_HM;
4734
4735	/*
4736	* 64-bit Windows uses XMM registers in the kernel as the Microsoft compiler expresses
4737	* floating-point operations using SSE instructions. Some XMM registers (XMM6-XMM15) are
4738	* callee-saved and thus the need for this XMM wrapper.
4739	*
4740	* Refer MSDN "Configuring Programs for 64-bit/x64 Software Conventions / Register Usage".
4741	*/
4742	#ifdef VBOX_WITH_KERNEL_USING_XMM
4743	return hmR0SVMRunWrapXMM(pVCpu->hm.s.svm.HCPhysVmcbHost, pCtx->hwvirt.svm.HCPhysVmcb, pCtx, pVM, pVCpu,
4744	pVCpu->hm.s.svm.pfnVMRun);
4745	#else
4746	return pVCpu->hm.s.svm.pfnVMRun(pVCpu->hm.s.svm.HCPhysVmcbHost, pCtx->hwvirt.svm.HCPhysVmcb, pCtx, pVM, pVCpu);
4747	#endif
4748	}
4749
4750
4751	/**
4752	* Undoes the TSC offset applied for an SVM nested-guest and returns the TSC
4753	* value for the guest.
4754	*
4755	* @returns The TSC offset after undoing any nested-guest TSC offset.
4756	* @param pVCpu The cross context virtual CPU structure of the calling EMT.
4757	* @param uTicks The nested-guest TSC.
4758	*
4759	* @note If you make any changes to this function, please check if
4760	* hmR0SvmNstGstUndoTscOffset() needs adjusting.
4761	*
4762	* @sa HMSvmNstGstApplyTscOffset().
4763	*/
4764	DECLINLINE(uint64_t) hmR0SvmNstGstUndoTscOffset(PVMCPU pVCpu, uint64_t uTicks)
4765	{
4766	PCSVMNESTEDVMCBCACHE pVmcbNstGstCache = &pVCpu->hm.s.svm.NstGstVmcbCache;
4767	Assert(pVmcbNstGstCache->fCacheValid);
4768	return uTicks - pVmcbNstGstCache->u64TSCOffset;
4769	}
4770	#endif
4771
4772	/**
4773	* Performs some essential restoration of state after running guest (or
4774	* nested-guest) code in AMD-V.
4775	*
4776	* @param pVCpu The cross context virtual CPU structure.
4777	* @param pCtx Pointer to the guest-CPU context. The data maybe
4778	* out-of-sync. Make sure to update the required fields
4779	* before using them.
4780	* @param pSvmTransient Pointer to the SVM transient structure.
4781	* @param rcVMRun Return code of VMRUN.
4782	*
4783	* @remarks Called with interrupts disabled.
4784	* @remarks No-long-jump zone!!! This function will however re-enable longjmps
4785	* unconditionally when it is safe to do so.
4786	*/
4787	static void hmR0SvmPostRunGuest(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient, int rcVMRun)
4788	{
4789	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
4790
4791	uint64_t const uHostTsc = ASMReadTSC(); /* Read the TSC as soon as possible. */
4792	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, false); /* See HMInvalidatePageOnAllVCpus(): used for TLB flushing. */
4793	ASMAtomicIncU32(&pVCpu->hm.s.cWorldSwitchExits); /* Initialized in vmR3CreateUVM(): used for EMT poking. */
4794
4795	PSVMVMCB pVmcb = pSvmTransient->pVmcb;
4796	PSVMVMCBCTRL pVmcbCtrl = &pVmcb->ctrl;
4797
4798	/* TSC read must be done early for maximum accuracy. */
4799	if (!(pVmcbCtrl->u64InterceptCtrl & SVM_CTRL_INTERCEPT_RDTSC))
4800	{
4801	if (!pSvmTransient->fIsNestedGuest)
4802	TMCpuTickSetLastSeen(pVCpu, uHostTsc + pVmcbCtrl->u64TSCOffset);
4803	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
4804	else
4805	{
4806	/* The nested-guest VMCB TSC offset shall eventually be restored on #VMEXIT via HMSvmNstGstVmExitNotify(). */
4807	uint64_t const uGstTsc = hmR0SvmNstGstUndoTscOffset(pVCpu, uHostTsc + pVmcbCtrl->u64TSCOffset);
4808	TMCpuTickSetLastSeen(pVCpu, uGstTsc);
4809	}
4810	#endif
4811	}
4812
4813	if (pSvmTransient->fRestoreTscAuxMsr)
4814	{
4815	uint64_t u64GuestTscAuxMsr = ASMRdMsr(MSR_K8_TSC_AUX);
4816	CPUMSetGuestTscAux(pVCpu, u64GuestTscAuxMsr);
4817	if (u64GuestTscAuxMsr != pVCpu->hm.s.u64HostTscAux)
4818	ASMWrMsr(MSR_K8_TSC_AUX, pVCpu->hm.s.u64HostTscAux);
4819	}
4820
4821	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatInGC, &pVCpu->hm.s.StatExit1, x);
4822	TMNotifyEndOfExecution(pVCpu); /* Notify TM that the guest is no longer running. */
4823	VMCPU_SET_STATE(pVCpu, VMCPUSTATE_STARTED_HM);
4824
4825	Assert(!(ASMGetFlags() & X86_EFL_IF));
4826	ASMSetFlags(pSvmTransient->fEFlags); /* Enable interrupts. */
4827	VMMRZCallRing3Enable(pVCpu); /* It is now safe to do longjmps to ring-3!!! */
4828
4829	/* If VMRUN failed, we can bail out early. This does -not- cover SVM_EXIT_INVALID. */
4830	if (RT_UNLIKELY(rcVMRun != VINF_SUCCESS))
4831	{
4832	Log4Func(("VMRUN failure: rcVMRun=%Rrc\n", rcVMRun));
4833	return;
4834	}
4835
4836	pSvmTransient->u64ExitCode = pVmcbCtrl->u64ExitCode; /* Save the #VMEXIT reason. */
4837	pVmcbCtrl->u32VmcbCleanBits = HMSVM_VMCB_CLEAN_ALL; /* Mark the VMCB-state cache as unmodified by VMM. */
4838	pSvmTransient->fVectoringDoublePF = false; /* Vectoring double page-fault needs to be determined later. */
4839	pSvmTransient->fVectoringPF = false; /* Vectoring page-fault needs to be determined later. */
4840
4841	#ifdef HMSVM_SYNC_FULL_GUEST_STATE
4842	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
4843	hmR0SvmImportGuestState(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
4844	#elif defined(HMSVM_SYNC_FULL_NESTED_GUEST_STATE)
4845	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
4846	hmR0SvmImportGuestState(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
4847	#else
4848	/*
4849	* Always import the following:
4850	*
4851	* - RIP for exit optimizations and evaluating event injection on re-entry.
4852	* - RFLAGS for evaluating event injection on VM re-entry and for exporting shared debug
4853	* state on preemption.
4854	* - Interrupt shadow, GIF for evaluating event injection on VM re-entry.
4855	* - CS for exit optimizations.
4856	* - RAX, RSP for simplifying assumptions on GPRs. All other GPRs are swapped by the
4857	* assembly switcher code.
4858	* - Shared state (only DR7 currently) for exporting shared debug state on preemption.
4859	*/
4860	hmR0SvmImportGuestState(pVCpu, CPUMCTX_EXTRN_RIP
4861	\| CPUMCTX_EXTRN_RFLAGS
4862	\| CPUMCTX_EXTRN_RAX
4863	\| CPUMCTX_EXTRN_RSP
4864	\| CPUMCTX_EXTRN_CS
4865	\| CPUMCTX_EXTRN_HWVIRT
4866	\| CPUMCTX_EXTRN_HM_SVM_INT_SHADOW
4867	\| CPUMCTX_EXTRN_HM_SVM_HWVIRT_VIRQ
4868	\| HMSVM_CPUMCTX_SHARED_STATE);
4869	#endif
4870
4871	#ifdef DEBUG_ramshankar
4872	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
4873	{
4874	hmR0SvmImportGuestState(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
4875	hmR0SvmLogState(pVCpu, pVmcb, pCtx, "hmR0SvmPostRunGuestNested", HMSVM_LOG_ALL & ~HMSVM_LOG_LBR, 0 /* uVerbose */);
4876	}
4877	#endif
4878
4879	if ( pSvmTransient->u64ExitCode != SVM_EXIT_INVALID
4880	&& pVCpu->hm.s.svm.fSyncVTpr)
4881	{
4882	Assert(!pSvmTransient->fIsNestedGuest);
4883	/* TPR patching (for 32-bit guests) uses LSTAR MSR for holding the TPR value, otherwise uses the VTPR. */
4884	if ( pVCpu->CTX_SUFF(pVM)->hm.s.fTPRPatchingActive
4885	&& (pVmcb->guest.u64LSTAR & 0xff) != pSvmTransient->u8GuestTpr)
4886	{
4887	int rc = APICSetTpr(pVCpu, pVmcb->guest.u64LSTAR & 0xff);
4888	AssertRC(rc);
4889	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
4890	}
4891	/* Sync TPR when we aren't intercepting CR8 writes. */
4892	else if (pSvmTransient->u8GuestTpr != pVmcbCtrl->IntCtrl.n.u8VTPR)
4893	{
4894	int rc = APICSetTpr(pVCpu, pVmcbCtrl->IntCtrl.n.u8VTPR << 4);
4895	AssertRC(rc);
4896	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
4897	}
4898	}
4899
4900	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_CS \| CPUMCTX_EXTRN_RIP);
4901	EMHistoryAddExit(pVCpu, EMEXIT_MAKE_FT(EMEXIT_F_KIND_SVM, pSvmTransient->u64ExitCode & EMEXIT_F_TYPE_MASK),
4902	pCtx->cs.u64Base + pCtx->rip, uHostTsc);
4903	}
4904
4905
4906	/**
4907	* Runs the guest code using AMD-V.
4908	*
4909	* @returns VBox status code.
4910	* @param pVM The cross context VM structure.
4911	* @param pVCpu The cross context virtual CPU structure.
4912	* @param pCtx Pointer to the guest-CPU context.
4913	* @param pcLoops Pointer to the number of executed loops.
4914	*/
4915	static int hmR0SvmRunGuestCodeNormal(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t *pcLoops)
4916	{
4917	uint32_t const cMaxResumeLoops = pVM->hm.s.cMaxResumeLoops;
4918	Assert(pcLoops);
4919	Assert(*pcLoops <= cMaxResumeLoops);
4920
4921	SVMTRANSIENT SvmTransient;
4922	RT_ZERO(SvmTransient);
4923	SvmTransient.fUpdateTscOffsetting = true;
4924	SvmTransient.pVmcb = pVCpu->hm.s.svm.pVmcb;
4925
4926	int rc = VERR_INTERNAL_ERROR_5;
4927	for (;;)
4928	{
4929	Assert(!HMR0SuspendPending());
4930	HMSVM_ASSERT_CPU_SAFE();
4931
4932	/* Preparatory work for running nested-guest code, this may force us to return to
4933	ring-3. This bugger disables interrupts on VINF_SUCCESS! */
4934	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatEntry, x);
4935	rc = hmR0SvmPreRunGuest(pVM, pVCpu, pCtx, &SvmTransient);
4936	if (rc != VINF_SUCCESS)
4937	break;
4938
4939	/*
4940	* No longjmps to ring-3 from this point on!!!
4941	*
4942	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional,
4943	* better than a kernel panic. This also disables flushing of the R0-logger instance.
4944	*/
4945	hmR0SvmPreRunGuestCommitted(pVCpu, pCtx, &SvmTransient);
4946	rc = hmR0SvmRunGuest(pVM, pVCpu, pCtx);
4947
4948	/* Restore any residual host-state and save any bits shared between host and guest
4949	into the guest-CPU state. Re-enables interrupts! */
4950	hmR0SvmPostRunGuest(pVCpu, pCtx, &SvmTransient, rc);
4951
4952	if (RT_UNLIKELY( rc != VINF_SUCCESS /* Check for VMRUN errors. */
4953	\|\| SvmTransient.u64ExitCode == SVM_EXIT_INVALID)) /* Check for invalid guest-state errors. */
4954	{
4955	if (rc == VINF_SUCCESS)
4956	rc = VERR_SVM_INVALID_GUEST_STATE;
4957	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit1, x);
4958	hmR0SvmReportWorldSwitchError(pVM, pVCpu, rc, pCtx);
4959	break;
4960	}
4961
4962	/* Handle the #VMEXIT. */
4963	HMSVM_EXITCODE_STAM_COUNTER_INC(SvmTransient.u64ExitCode);
4964	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatExit1, &pVCpu->hm.s.StatExit2, x);
4965	VBOXVMM_R0_HMSVM_VMEXIT(pVCpu, pCtx, SvmTransient.u64ExitCode, pVCpu->hm.s.svm.pVmcb);
4966	rc = hmR0SvmHandleExit(pVCpu, pCtx, &SvmTransient);
4967	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit2, x);
4968	if (rc != VINF_SUCCESS)
4969	break;
4970	if (++(*pcLoops) >= cMaxResumeLoops)
4971	{
4972	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchMaxResumeLoops);
4973	rc = VINF_EM_RAW_INTERRUPT;
4974	break;
4975	}
4976	}
4977
4978	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatEntry, x);
4979	return rc;
4980	}
4981
4982
4983	/**
4984	* Runs the guest code using AMD-V in single step mode.
4985	*
4986	* @returns VBox status code.
4987	* @param pVM The cross context VM structure.
4988	* @param pVCpu The cross context virtual CPU structure.
4989	* @param pCtx Pointer to the guest-CPU context.
4990	* @param pcLoops Pointer to the number of executed loops.
4991	*/
4992	static int hmR0SvmRunGuestCodeStep(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t *pcLoops)
4993	{
4994	uint32_t const cMaxResumeLoops = pVM->hm.s.cMaxResumeLoops;
4995	Assert(pcLoops);
4996	Assert(*pcLoops <= cMaxResumeLoops);
4997
4998	SVMTRANSIENT SvmTransient;
4999	RT_ZERO(SvmTransient);
5000	SvmTransient.fUpdateTscOffsetting = true;
5001	SvmTransient.pVmcb = pVCpu->hm.s.svm.pVmcb;
5002
5003	uint16_t uCsStart = pCtx->cs.Sel;
5004	uint64_t uRipStart = pCtx->rip;
5005
5006	int rc = VERR_INTERNAL_ERROR_5;
5007	for (;;)
5008	{
5009	Assert(!HMR0SuspendPending());
5010	AssertMsg(pVCpu->hm.s.idEnteredCpu == RTMpCpuId(),
5011	("Illegal migration! Entered on CPU %u Current %u cLoops=%u\n", (unsigned)pVCpu->hm.s.idEnteredCpu,
5012	(unsigned)RTMpCpuId(), *pcLoops));
5013
5014	/* Preparatory work for running nested-guest code, this may force us to return to
5015	ring-3. This bugger disables interrupts on VINF_SUCCESS! */
5016	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatEntry, x);
5017	rc = hmR0SvmPreRunGuest(pVM, pVCpu, pCtx, &SvmTransient);
5018	if (rc != VINF_SUCCESS)
5019	break;
5020
5021	/*
5022	* No longjmps to ring-3 from this point on!!!
5023	*
5024	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional,
5025	* better than a kernel panic. This also disables flushing of the R0-logger instance.
5026	*/
5027	VMMRZCallRing3Disable(pVCpu);
5028	VMMRZCallRing3RemoveNotification(pVCpu);
5029	hmR0SvmPreRunGuestCommitted(pVCpu, pCtx, &SvmTransient);
5030
5031	rc = hmR0SvmRunGuest(pVM, pVCpu, pCtx);
5032
5033	/* Restore any residual host-state and save any bits shared between host and guest
5034	into the guest-CPU state. Re-enables interrupts! */
5035	hmR0SvmPostRunGuest(pVCpu, pCtx, &SvmTransient, rc);
5036
5037	if (RT_UNLIKELY( rc != VINF_SUCCESS /* Check for VMRUN errors. */
5038	\|\| SvmTransient.u64ExitCode == SVM_EXIT_INVALID)) /* Check for invalid guest-state errors. */
5039	{
5040	if (rc == VINF_SUCCESS)
5041	rc = VERR_SVM_INVALID_GUEST_STATE;
5042	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit1, x);
5043	hmR0SvmReportWorldSwitchError(pVM, pVCpu, rc, pCtx);
5044	return rc;
5045	}
5046
5047	/* Handle the #VMEXIT. */
5048	HMSVM_EXITCODE_STAM_COUNTER_INC(SvmTransient.u64ExitCode);
5049	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatExit1, &pVCpu->hm.s.StatExit2, x);
5050	VBOXVMM_R0_HMSVM_VMEXIT(pVCpu, pCtx, SvmTransient.u64ExitCode, pVCpu->hm.s.svm.pVmcb);
5051	rc = hmR0SvmHandleExit(pVCpu, pCtx, &SvmTransient);
5052	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit2, x);
5053	if (rc != VINF_SUCCESS)
5054	break;
5055	if (++(*pcLoops) >= cMaxResumeLoops)
5056	{
5057	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchMaxResumeLoops);
5058	rc = VINF_EM_RAW_INTERRUPT;
5059	break;
5060	}
5061
5062	/*
5063	* Did the RIP change, if so, consider it a single step.
5064	* Otherwise, make sure one of the TFs gets set.
5065	*/
5066	if ( pCtx->rip != uRipStart
5067	\|\| pCtx->cs.Sel != uCsStart)
5068	{
5069	rc = VINF_EM_DBG_STEPPED;
5070	break;
5071	}
5072	pVCpu->hm.s.fCtxChanged \|= HM_CHANGED_GUEST_DR_MASK;
5073	}
5074
5075	/*
5076	* Clear the X86_EFL_TF if necessary.
5077	*/
5078	if (pVCpu->hm.s.fClearTrapFlag)
5079	{
5080	pVCpu->hm.s.fClearTrapFlag = false;
5081	pCtx->eflags.Bits.u1TF = 0;
5082	}
5083
5084	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatEntry, x);
5085	return rc;
5086	}
5087
5088	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
5089	/**
5090	* Runs the nested-guest code using AMD-V.
5091	*
5092	* @returns VBox status code.
5093	* @param pVM The cross context VM structure.
5094	* @param pVCpu The cross context virtual CPU structure.
5095	* @param pCtx Pointer to the guest-CPU context.
5096	* @param pcLoops Pointer to the number of executed loops. If we're switching
5097	* from the guest-code execution loop to this nested-guest
5098	* execution loop pass the remainder value, else pass 0.
5099	*/
5100	static int hmR0SvmRunGuestCodeNested(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t *pcLoops)
5101	{
5102	HMSVM_ASSERT_IN_NESTED_GUEST(pCtx);
5103	Assert(pcLoops);
5104	Assert(*pcLoops <= pVM->hm.s.cMaxResumeLoops);
5105
5106	SVMTRANSIENT SvmTransient;
5107	RT_ZERO(SvmTransient);
5108	SvmTransient.fUpdateTscOffsetting = true;
5109	SvmTransient.pVmcb = pCtx->hwvirt.svm.CTX_SUFF(pVmcb);
5110	SvmTransient.fIsNestedGuest = true;
5111
5112	int rc = VERR_INTERNAL_ERROR_4;
5113	for (;;)
5114	{
5115	Assert(!HMR0SuspendPending());
5116	HMSVM_ASSERT_CPU_SAFE();
5117
5118	/* Preparatory work for running nested-guest code, this may force us to return to
5119	ring-3. This bugger disables interrupts on VINF_SUCCESS! */
5120	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatEntry, x);
5121	rc = hmR0SvmPreRunGuestNested(pVM, pVCpu, pCtx, &SvmTransient);
5122	if ( rc != VINF_SUCCESS
5123	\|\| !CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
5124	{
5125	break;
5126	}
5127
5128	/*
5129	* No longjmps to ring-3 from this point on!!!
5130	*
5131	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional,
5132	* better than a kernel panic. This also disables flushing of the R0-logger instance.
5133	*/
5134	hmR0SvmPreRunGuestCommitted(pVCpu, pCtx, &SvmTransient);
5135
5136	rc = hmR0SvmRunGuestNested(pVM, pVCpu, pCtx);
5137
5138	/* Restore any residual host-state and save any bits shared between host and guest
5139	into the guest-CPU state. Re-enables interrupts! */
5140	hmR0SvmPostRunGuest(pVCpu, pCtx, &SvmTransient, rc);
5141
5142	if (RT_LIKELY( rc == VINF_SUCCESS
5143	&& SvmTransient.u64ExitCode != SVM_EXIT_INVALID))
5144	{ /* extremely likely */ }
5145	else
5146	{
5147	/* VMRUN failed, shouldn't really happen, Guru. */
5148	if (rc != VINF_SUCCESS)
5149	break;
5150
5151	/* Invalid nested-guest state. Cause a #VMEXIT but assert on strict builds. */
5152	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
5153	AssertMsgFailed(("Invalid nested-guest state. rc=%Rrc u64ExitCode=%#RX64\n", rc, SvmTransient.u64ExitCode));
5154	rc = VBOXSTRICTRC_TODO(IEMExecSvmVmexit(pVCpu, SVM_EXIT_INVALID, 0, 0));
5155	break;
5156	}
5157
5158	/* Handle the #VMEXIT. */
5159	HMSVM_NESTED_EXITCODE_STAM_COUNTER_INC(SvmTransient.u64ExitCode);
5160	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatExit1, &pVCpu->hm.s.StatExit2, x);
5161	VBOXVMM_R0_HMSVM_VMEXIT(pVCpu, pCtx, SvmTransient.u64ExitCode, pCtx->hwvirt.svm.CTX_SUFF(pVmcb));
5162	rc = hmR0SvmHandleExitNested(pVCpu, pCtx, &SvmTransient);
5163	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit2, x);
5164	if ( rc != VINF_SUCCESS
5165	\|\| !CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
5166	break;
5167	if (++(*pcLoops) >= pVM->hm.s.cMaxResumeLoops)
5168	{
5169	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchMaxResumeLoops);
5170	rc = VINF_EM_RAW_INTERRUPT;
5171	break;
5172	}
5173
5174	/** @todo handle single-stepping */
5175	}
5176
5177	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatEntry, x);
5178	return rc;
5179	}
5180	#endif
5181
5182
5183	/**
5184	* Runs the guest code using AMD-V.
5185	*
5186	* @returns Strict VBox status code.
5187	* @param pVM The cross context VM structure.
5188	* @param pVCpu The cross context virtual CPU structure.
5189	* @param pCtx Pointer to the guest-CPU context.
5190	*/
5191	VMMR0DECL(VBOXSTRICTRC) SVMR0RunGuestCode(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
5192	{
5193	Assert(VMMRZCallRing3IsEnabled(pVCpu));
5194	HMSVM_ASSERT_PREEMPT_SAFE();
5195	VMMRZCallRing3SetNotification(pVCpu, hmR0SvmCallRing3Callback, pCtx);
5196
5197	uint32_t cLoops = 0;
5198	int rc;
5199	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
5200	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
5201	#endif
5202	{
5203	if (!pVCpu->hm.s.fSingleInstruction)
5204	rc = hmR0SvmRunGuestCodeNormal(pVM, pVCpu, pCtx, &cLoops);
5205	else
5206	rc = hmR0SvmRunGuestCodeStep(pVM, pVCpu, pCtx, &cLoops);
5207	}
5208	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
5209	else
5210	{
5211	rc = VINF_SVM_VMRUN;
5212	}
5213
5214	/* Re-check the nested-guest condition here as we may be transitioning from the normal
5215	execution loop into the nested-guest, hence this is not placed in the 'else' part above. */
5216	if (rc == VINF_SVM_VMRUN)
5217	{
5218	rc = hmR0SvmRunGuestCodeNested(pVM, pVCpu, pCtx, &cLoops);
5219	if (rc == VINF_SVM_VMEXIT)
5220	rc = VINF_SUCCESS;
5221	}
5222	#endif
5223
5224	/* Fixup error codes. */
5225	if (rc == VERR_EM_INTERPRETER)
5226	rc = VINF_EM_RAW_EMULATE_INSTR;
5227	else if (rc == VINF_EM_RESET)
5228	rc = VINF_EM_TRIPLE_FAULT;
5229
5230	/* Prepare to return to ring-3. This will remove longjmp notifications. */
5231	rc = hmR0SvmExitToRing3(pVM, pVCpu, pCtx, rc);
5232	Assert(!VMMRZCallRing3IsNotificationSet(pVCpu));
5233	return rc;
5234	}
5235
5236
5237	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
5238	/**
5239	* Determines whether an IOIO intercept is active for the nested-guest or not.
5240	*
5241	* @param pvIoBitmap Pointer to the nested-guest IO bitmap.
5242	* @param pIoExitInfo Pointer to the SVMIOIOEXITINFO.
5243	*/
5244	static bool hmR0SvmIsIoInterceptActive(void *pvIoBitmap, PSVMIOIOEXITINFO pIoExitInfo)
5245	{
5246	const uint16_t u16Port = pIoExitInfo->n.u16Port;
5247	const SVMIOIOTYPE enmIoType = (SVMIOIOTYPE)pIoExitInfo->n.u1Type;
5248	const uint8_t cbReg = (pIoExitInfo->u >> SVM_IOIO_OP_SIZE_SHIFT) & 7;
5249	const uint8_t cAddrSizeBits = ((pIoExitInfo->u >> SVM_IOIO_ADDR_SIZE_SHIFT) & 7) << 4;
5250	const uint8_t iEffSeg = pIoExitInfo->n.u3Seg;
5251	const bool fRep = pIoExitInfo->n.u1Rep;
5252	const bool fStrIo = pIoExitInfo->n.u1Str;
5253
5254	return HMSvmIsIOInterceptActive(pvIoBitmap, u16Port, enmIoType, cbReg, cAddrSizeBits, iEffSeg, fRep, fStrIo,
5255	NULL /* pIoExitInfo */);
5256	}
5257
5258
5259	/**
5260	* Handles a nested-guest \#VMEXIT (for all EXITCODE values except
5261	* SVM_EXIT_INVALID).
5262	*
5263	* @returns VBox status code (informational status codes included).
5264	* @param pVCpu The cross context virtual CPU structure.
5265	* @param pCtx Pointer to the guest-CPU context.
5266	* @param pSvmTransient Pointer to the SVM transient structure.
5267	*/
5268	static int hmR0SvmHandleExitNested(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5269	{
5270	HMSVM_ASSERT_IN_NESTED_GUEST(pCtx);
5271	Assert(pSvmTransient->u64ExitCode != SVM_EXIT_INVALID);
5272	Assert(pSvmTransient->u64ExitCode <= SVM_EXIT_MAX);
5273
5274	/** @todo Figure out why using IEM_CPUMCTX_EXTRN_SVM_VMEXIT_MASK instead of
5275	* HMSVM_CPUMCTX_EXTRN_ALL breaks nested guests (XP Pro, DSL etc.), see
5276	* also HMSvmNstGstVmExitNotify(). */
5277	#define NST_GST_VMEXIT_CALL_RET(a_pVCpu, a_pCtx, a_uExitCode, a_uExitInfo1, a_uExitInfo2) \
5278	do { \
5279	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL); \
5280	return VBOXSTRICTRC_TODO(IEMExecSvmVmexit((a_pVCpu), (a_uExitCode), (a_uExitInfo1), (a_uExitInfo2))); \
5281	} while (0)
5282
5283	/*
5284	* For all the #VMEXITs here we primarily figure out if the #VMEXIT is expected by the
5285	* nested-guest. If it isn't, it should be handled by the (outer) guest.
5286	*/
5287	PSVMVMCB pVmcbNstGst = pCtx->hwvirt.svm.CTX_SUFF(pVmcb);
5288	PSVMVMCBCTRL pVmcbNstGstCtrl = &pVmcbNstGst->ctrl;
5289	uint64_t const uExitCode = pVmcbNstGstCtrl->u64ExitCode;
5290	uint64_t const uExitInfo1 = pVmcbNstGstCtrl->u64ExitInfo1;
5291	uint64_t const uExitInfo2 = pVmcbNstGstCtrl->u64ExitInfo2;
5292
5293	Assert(uExitCode == pVmcbNstGstCtrl->u64ExitCode);
5294	switch (uExitCode)
5295	{
5296	case SVM_EXIT_CPUID:
5297	{
5298	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_CPUID))
5299	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5300	return hmR0SvmExitCpuid(pVCpu, pCtx, pSvmTransient);
5301	}
5302
5303	case SVM_EXIT_RDTSC:
5304	{
5305	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_RDTSC))
5306	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5307	return hmR0SvmExitRdtsc(pVCpu, pCtx, pSvmTransient);
5308	}
5309
5310	case SVM_EXIT_RDTSCP:
5311	{
5312	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_RDTSCP))
5313	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5314	return hmR0SvmExitRdtscp(pVCpu, pCtx, pSvmTransient);
5315	}
5316
5317	case SVM_EXIT_MONITOR:
5318	{
5319	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_MONITOR))
5320	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5321	return hmR0SvmExitMonitor(pVCpu, pCtx, pSvmTransient);
5322	}
5323
5324	case SVM_EXIT_MWAIT:
5325	{
5326	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_MWAIT))
5327	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5328	return hmR0SvmExitMwait(pVCpu, pCtx, pSvmTransient);
5329	}
5330
5331	case SVM_EXIT_HLT:
5332	{
5333	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_HLT))
5334	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5335	return hmR0SvmExitHlt(pVCpu, pCtx, pSvmTransient);
5336	}
5337
5338	case SVM_EXIT_MSR:
5339	{
5340	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_MSR_PROT))
5341	{
5342	uint32_t const idMsr = pCtx->ecx;
5343	uint16_t offMsrpm;
5344	uint8_t uMsrpmBit;
5345	int rc = HMSvmGetMsrpmOffsetAndBit(idMsr, &offMsrpm, &uMsrpmBit);
5346	if (RT_SUCCESS(rc))
5347	{
5348	Assert(uMsrpmBit == 0 \|\| uMsrpmBit == 2 \|\| uMsrpmBit == 4 \|\| uMsrpmBit == 6);
5349	Assert(offMsrpm < SVM_MSRPM_PAGES << X86_PAGE_4K_SHIFT);
5350
5351	uint8_t const pbMsrBitmap = (uint8_t const )pCtx->hwvirt.svm.CTX_SUFF(pvMsrBitmap);
5352	pbMsrBitmap += offMsrpm;
5353	bool const fInterceptRead = RT_BOOL(*pbMsrBitmap & RT_BIT(uMsrpmBit));
5354	bool const fInterceptWrite = RT_BOOL(*pbMsrBitmap & RT_BIT(uMsrpmBit + 1));
5355
5356	if ( (fInterceptWrite && pVmcbNstGstCtrl->u64ExitInfo1 == SVM_EXIT1_MSR_WRITE)
5357	\|\| (fInterceptRead && pVmcbNstGstCtrl->u64ExitInfo1 == SVM_EXIT1_MSR_READ))
5358	{
5359	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5360	}
5361	}
5362	else
5363	{
5364	/*
5365	* MSRs not covered by the MSRPM automatically cause an #VMEXIT.
5366	* See AMD-V spec. "15.11 MSR Intercepts".
5367	*/
5368	Assert(rc == VERR_OUT_OF_RANGE);
5369	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5370	}
5371	}
5372	return hmR0SvmExitMsr(pVCpu, pCtx, pSvmTransient);
5373	}
5374
5375	case SVM_EXIT_IOIO:
5376	{
5377	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_IOIO_PROT))
5378	{
5379	void *pvIoBitmap = pCtx->hwvirt.svm.CTX_SUFF(pvIoBitmap);
5380	SVMIOIOEXITINFO IoExitInfo;
5381	IoExitInfo.u = pVmcbNstGst->ctrl.u64ExitInfo1;
5382	bool const fIntercept = hmR0SvmIsIoInterceptActive(pvIoBitmap, &IoExitInfo);
5383	if (fIntercept)
5384	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5385	}
5386	return hmR0SvmExitIOInstr(pVCpu, pCtx, pSvmTransient);
5387	}
5388
5389	case SVM_EXIT_XCPT_PF:
5390	{
5391	PVM pVM = pVCpu->CTX_SUFF(pVM);
5392	if (pVM->hm.s.fNestedPaging)
5393	{
5394	uint32_t const u32ErrCode = pVmcbNstGstCtrl->u64ExitInfo1;
5395	uint64_t const uFaultAddress = pVmcbNstGstCtrl->u64ExitInfo2;
5396
5397	/* If the nested-guest is intercepting #PFs, cause a #PF #VMEXIT. */
5398	if (HMIsGuestSvmXcptInterceptSet(pVCpu, X86_XCPT_PF))
5399	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, u32ErrCode, uFaultAddress);
5400
5401	/* If the nested-guest is not intercepting #PFs, forward the #PF to the guest. */
5402	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR2);
5403	hmR0SvmSetPendingXcptPF(pVCpu, pCtx, u32ErrCode, uFaultAddress);
5404	return VINF_SUCCESS;
5405	}
5406	return hmR0SvmExitXcptPF(pVCpu, pCtx,pSvmTransient);
5407	}
5408
5409	case SVM_EXIT_XCPT_UD:
5410	{
5411	if (HMIsGuestSvmXcptInterceptSet(pVCpu, X86_XCPT_UD))
5412	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5413	hmR0SvmSetPendingXcptUD(pVCpu);
5414	return VINF_SUCCESS;
5415	}
5416
5417	case SVM_EXIT_XCPT_MF:
5418	{
5419	if (HMIsGuestSvmXcptInterceptSet(pVCpu, X86_XCPT_MF))
5420	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5421	return hmR0SvmExitXcptMF(pVCpu, pCtx, pSvmTransient);
5422	}
5423
5424	case SVM_EXIT_XCPT_DB:
5425	{
5426	if (HMIsGuestSvmXcptInterceptSet(pVCpu, X86_XCPT_DB))
5427	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5428	return hmR0SvmNestedExitXcptDB(pVCpu, pCtx, pSvmTransient);
5429	}
5430
5431	case SVM_EXIT_XCPT_AC:
5432	{
5433	if (HMIsGuestSvmXcptInterceptSet(pVCpu, X86_XCPT_AC))
5434	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5435	return hmR0SvmExitXcptAC(pVCpu, pCtx, pSvmTransient);
5436	}
5437
5438	case SVM_EXIT_XCPT_BP:
5439	{
5440	if (HMIsGuestSvmXcptInterceptSet(pVCpu, X86_XCPT_BP))
5441	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5442	return hmR0SvmNestedExitXcptBP(pVCpu, pCtx, pSvmTransient);
5443	}
5444
5445	case SVM_EXIT_READ_CR0:
5446	case SVM_EXIT_READ_CR3:
5447	case SVM_EXIT_READ_CR4:
5448	{
5449	uint8_t const uCr = uExitCode - SVM_EXIT_READ_CR0;
5450	if (HMIsGuestSvmReadCRxInterceptSet(pVCpu, uCr))
5451	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5452	return hmR0SvmExitReadCRx(pVCpu, pCtx, pSvmTransient);
5453	}
5454
5455	case SVM_EXIT_CR0_SEL_WRITE:
5456	{
5457	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_CR0_SEL_WRITE))
5458	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5459	return hmR0SvmExitWriteCRx(pVCpu, pCtx, pSvmTransient);
5460	}
5461
5462	case SVM_EXIT_WRITE_CR0:
5463	case SVM_EXIT_WRITE_CR3:
5464	case SVM_EXIT_WRITE_CR4:
5465	case SVM_EXIT_WRITE_CR8: /* CR8 writes would go to the V_TPR rather than here, since we run with V_INTR_MASKING. */
5466	{
5467	uint8_t const uCr = uExitCode - SVM_EXIT_WRITE_CR0;
5468	Log4Func(("Write CR%u: uExitInfo1=%#RX64 uExitInfo2=%#RX64\n", uCr, uExitInfo1, uExitInfo2));
5469
5470	if (HMIsGuestSvmWriteCRxInterceptSet(pVCpu, uCr))
5471	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5472	return hmR0SvmExitWriteCRx(pVCpu, pCtx, pSvmTransient);
5473	}
5474
5475	case SVM_EXIT_PAUSE:
5476	{
5477	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_PAUSE))
5478	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5479	return hmR0SvmExitPause(pVCpu, pCtx, pSvmTransient);
5480	}
5481
5482	case SVM_EXIT_VINTR:
5483	{
5484	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_VINTR))
5485	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5486	return hmR0SvmExitUnexpected(pVCpu, pCtx, pSvmTransient);
5487	}
5488
5489	case SVM_EXIT_INTR:
5490	case SVM_EXIT_NMI:
5491	case SVM_EXIT_SMI:
5492	case SVM_EXIT_XCPT_NMI: /* Should not occur, SVM_EXIT_NMI is used instead. */
5493	{
5494	/*
5495	* We shouldn't direct physical interrupts, NMIs, SMIs to the nested-guest.
5496	*
5497	* Although we don't intercept SMIs, the nested-guest might. Therefore, we might
5498	* get an SMI #VMEXIT here so simply ignore rather than causing a corresponding
5499	* nested-guest #VMEXIT.
5500	*/
5501	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_SVM_VMEXIT_MASK);
5502	return hmR0SvmExitIntr(pVCpu, pCtx, pSvmTransient);
5503	}
5504
5505	case SVM_EXIT_FERR_FREEZE:
5506	{
5507	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_FERR_FREEZE))
5508	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5509	return hmR0SvmExitFerrFreeze(pVCpu, pCtx, pSvmTransient);
5510	}
5511
5512	case SVM_EXIT_INVLPG:
5513	{
5514	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_INVLPG))
5515	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5516	return hmR0SvmExitInvlpg(pVCpu, pCtx, pSvmTransient);
5517	}
5518
5519	case SVM_EXIT_WBINVD:
5520	{
5521	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_WBINVD))
5522	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5523	return hmR0SvmExitWbinvd(pVCpu, pCtx, pSvmTransient);
5524	}
5525
5526	case SVM_EXIT_INVD:
5527	{
5528	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_INVD))
5529	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5530	return hmR0SvmExitInvd(pVCpu, pCtx, pSvmTransient);
5531	}
5532
5533	case SVM_EXIT_RDPMC:
5534	{
5535	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_RDPMC))
5536	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5537	return hmR0SvmExitRdpmc(pVCpu, pCtx, pSvmTransient);
5538	}
5539
5540	default:
5541	{
5542	switch (uExitCode)
5543	{
5544	case SVM_EXIT_READ_DR0: case SVM_EXIT_READ_DR1: case SVM_EXIT_READ_DR2: case SVM_EXIT_READ_DR3:
5545	case SVM_EXIT_READ_DR6: case SVM_EXIT_READ_DR7: case SVM_EXIT_READ_DR8: case SVM_EXIT_READ_DR9:
5546	case SVM_EXIT_READ_DR10: case SVM_EXIT_READ_DR11: case SVM_EXIT_READ_DR12: case SVM_EXIT_READ_DR13:
5547	case SVM_EXIT_READ_DR14: case SVM_EXIT_READ_DR15:
5548	{
5549	uint8_t const uDr = uExitCode - SVM_EXIT_READ_DR0;
5550	if (HMIsGuestSvmReadDRxInterceptSet(pVCpu, uDr))
5551	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5552	return hmR0SvmExitReadDRx(pVCpu, pCtx, pSvmTransient);
5553	}
5554
5555	case SVM_EXIT_WRITE_DR0: case SVM_EXIT_WRITE_DR1: case SVM_EXIT_WRITE_DR2: case SVM_EXIT_WRITE_DR3:
5556	case SVM_EXIT_WRITE_DR6: case SVM_EXIT_WRITE_DR7: case SVM_EXIT_WRITE_DR8: case SVM_EXIT_WRITE_DR9:
5557	case SVM_EXIT_WRITE_DR10: case SVM_EXIT_WRITE_DR11: case SVM_EXIT_WRITE_DR12: case SVM_EXIT_WRITE_DR13:
5558	case SVM_EXIT_WRITE_DR14: case SVM_EXIT_WRITE_DR15:
5559	{
5560	uint8_t const uDr = uExitCode - SVM_EXIT_WRITE_DR0;
5561	if (HMIsGuestSvmWriteDRxInterceptSet(pVCpu, uDr))
5562	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5563	return hmR0SvmExitWriteDRx(pVCpu, pCtx, pSvmTransient);
5564	}
5565
5566	case SVM_EXIT_XCPT_DE:
5567	/* SVM_EXIT_XCPT_DB: / / Handled above. */
5568	/* SVM_EXIT_XCPT_NMI: / / Handled above. */
5569	/* SVM_EXIT_XCPT_BP: / / Handled above. */
5570	case SVM_EXIT_XCPT_OF:
5571	case SVM_EXIT_XCPT_BR:
5572	/* SVM_EXIT_XCPT_UD: / / Handled above. */
5573	case SVM_EXIT_XCPT_NM:
5574	case SVM_EXIT_XCPT_DF:
5575	case SVM_EXIT_XCPT_CO_SEG_OVERRUN:
5576	case SVM_EXIT_XCPT_TS:
5577	case SVM_EXIT_XCPT_NP:
5578	case SVM_EXIT_XCPT_SS:
5579	case SVM_EXIT_XCPT_GP:
5580	/* SVM_EXIT_XCPT_PF: / / Handled above. */
5581	case SVM_EXIT_XCPT_15: /* Reserved. */
5582	/* SVM_EXIT_XCPT_MF: / / Handled above. */
5583	/* SVM_EXIT_XCPT_AC: / / Handled above. */
5584	case SVM_EXIT_XCPT_MC:
5585	case SVM_EXIT_XCPT_XF:
5586	case SVM_EXIT_XCPT_20: case SVM_EXIT_XCPT_21: case SVM_EXIT_XCPT_22: case SVM_EXIT_XCPT_23:
5587	case SVM_EXIT_XCPT_24: case SVM_EXIT_XCPT_25: case SVM_EXIT_XCPT_26: case SVM_EXIT_XCPT_27:
5588	case SVM_EXIT_XCPT_28: case SVM_EXIT_XCPT_29: case SVM_EXIT_XCPT_30: case SVM_EXIT_XCPT_31:
5589	{
5590	uint8_t const uVector = uExitCode - SVM_EXIT_XCPT_0;
5591	if (HMIsGuestSvmXcptInterceptSet(pVCpu, uVector))
5592	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5593	return hmR0SvmExitXcptGeneric(pVCpu, pCtx, pSvmTransient);
5594	}
5595
5596	case SVM_EXIT_XSETBV:
5597	{
5598	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_XSETBV))
5599	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5600	return hmR0SvmExitXsetbv(pVCpu, pCtx, pSvmTransient);
5601	}
5602
5603	case SVM_EXIT_TASK_SWITCH:
5604	{
5605	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_TASK_SWITCH))
5606	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5607	return hmR0SvmExitTaskSwitch(pVCpu, pCtx, pSvmTransient);
5608	}
5609
5610	case SVM_EXIT_IRET:
5611	{
5612	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_IRET))
5613	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5614	return hmR0SvmExitIret(pVCpu, pCtx, pSvmTransient);
5615	}
5616
5617	case SVM_EXIT_SHUTDOWN:
5618	{
5619	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_SHUTDOWN))
5620	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5621	return hmR0SvmExitShutdown(pVCpu, pCtx, pSvmTransient);
5622	}
5623
5624	case SVM_EXIT_VMMCALL:
5625	{
5626	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_VMMCALL))
5627	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5628	return hmR0SvmExitVmmCall(pVCpu, pCtx, pSvmTransient);
5629	}
5630
5631	case SVM_EXIT_CLGI:
5632	{
5633	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_CLGI))
5634	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5635	return hmR0SvmExitClgi(pVCpu, pCtx, pSvmTransient);
5636	}
5637
5638	case SVM_EXIT_STGI:
5639	{
5640	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_STGI))
5641	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5642	return hmR0SvmExitStgi(pVCpu, pCtx, pSvmTransient);
5643	}
5644
5645	case SVM_EXIT_VMLOAD:
5646	{
5647	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_VMLOAD))
5648	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5649	return hmR0SvmExitVmload(pVCpu, pCtx, pSvmTransient);
5650	}
5651
5652	case SVM_EXIT_VMSAVE:
5653	{
5654	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_VMSAVE))
5655	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5656	return hmR0SvmExitVmsave(pVCpu, pCtx, pSvmTransient);
5657	}
5658
5659	case SVM_EXIT_INVLPGA:
5660	{
5661	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_INVLPGA))
5662	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5663	return hmR0SvmExitInvlpga(pVCpu, pCtx, pSvmTransient);
5664	}
5665
5666	case SVM_EXIT_VMRUN:
5667	{
5668	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_VMRUN))
5669	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5670	return hmR0SvmExitVmrun(pVCpu, pCtx, pSvmTransient);
5671	}
5672
5673	case SVM_EXIT_RSM:
5674	{
5675	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_RSM))
5676	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5677	hmR0SvmSetPendingXcptUD(pVCpu);
5678	return VINF_SUCCESS;
5679	}
5680
5681	case SVM_EXIT_SKINIT:
5682	{
5683	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_SKINIT))
5684	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5685	hmR0SvmSetPendingXcptUD(pVCpu);
5686	return VINF_SUCCESS;
5687	}
5688
5689	case SVM_EXIT_NPF:
5690	{
5691	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
5692	return hmR0SvmExitNestedPF(pVCpu, pCtx, pSvmTransient);
5693	}
5694
5695	case SVM_EXIT_INIT: /* We shouldn't get INIT signals while executing a nested-guest. */
5696	return hmR0SvmExitUnexpected(pVCpu, pCtx, pSvmTransient);
5697
5698	default:
5699	{
5700	AssertMsgFailed(("hmR0SvmHandleExitNested: Unknown exit code %#x\n", pSvmTransient->u64ExitCode));
5701	pVCpu->hm.s.u32HMError = pSvmTransient->u64ExitCode;
5702	return VERR_SVM_UNKNOWN_EXIT;
5703	}
5704	}
5705	}
5706	}
5707	/* not reached */
5708
5709	#undef NST_GST_VMEXIT_CALL_RET
5710	}
5711	#endif
5712
5713
5714	/**
5715	* Handles a guest \#VMEXIT (for all EXITCODE values except SVM_EXIT_INVALID).
5716	*
5717	* @returns VBox status code (informational status codes included).
5718	* @param pVCpu The cross context virtual CPU structure.
5719	* @param pCtx Pointer to the guest-CPU context.
5720	* @param pSvmTransient Pointer to the SVM transient structure.
5721	*/
5722	static int hmR0SvmHandleExit(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5723	{
5724	Assert(pSvmTransient->u64ExitCode != SVM_EXIT_INVALID);
5725	Assert(pSvmTransient->u64ExitCode <= SVM_EXIT_MAX);
5726
5727	#ifdef DEBUG_ramshankar
5728	# define VMEXIT_CALL_RET(a_fDbg, a_CallExpr) \
5729	do { \
5730	if ((a_fDbg) == 1) \
5731	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL); \
5732	int rc = a_CallExpr; \
5733	if ((a_fDbg) == 1) \
5734	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST); \
5735	return rc; \
5736	} while (0)
5737	#else
5738	# define VMEXIT_CALL_RET(a_fDbg, a_CallExpr) return a_CallExpr
5739	#endif
5740
5741	/*
5742	* The ordering of the case labels is based on most-frequently-occurring #VMEXITs
5743	* for most guests under normal workloads (for some definition of "normal").
5744	*/
5745	uint64_t const uExitCode = pSvmTransient->u64ExitCode;
5746	switch (uExitCode)
5747	{
5748	case SVM_EXIT_NPF: VMEXIT_CALL_RET(0, hmR0SvmExitNestedPF(pVCpu, pCtx, pSvmTransient));
5749	case SVM_EXIT_IOIO: VMEXIT_CALL_RET(0, hmR0SvmExitIOInstr(pVCpu, pCtx, pSvmTransient));
5750	case SVM_EXIT_RDTSC: VMEXIT_CALL_RET(0, hmR0SvmExitRdtsc(pVCpu, pCtx, pSvmTransient));
5751	case SVM_EXIT_RDTSCP: VMEXIT_CALL_RET(0, hmR0SvmExitRdtscp(pVCpu, pCtx, pSvmTransient));
5752	case SVM_EXIT_CPUID: VMEXIT_CALL_RET(0, hmR0SvmExitCpuid(pVCpu, pCtx, pSvmTransient));
5753	case SVM_EXIT_XCPT_PF: VMEXIT_CALL_RET(0, hmR0SvmExitXcptPF(pVCpu, pCtx, pSvmTransient));
5754	case SVM_EXIT_MSR: VMEXIT_CALL_RET(0, hmR0SvmExitMsr(pVCpu, pCtx, pSvmTransient));
5755	case SVM_EXIT_MONITOR: VMEXIT_CALL_RET(0, hmR0SvmExitMonitor(pVCpu, pCtx, pSvmTransient));
5756	case SVM_EXIT_MWAIT: VMEXIT_CALL_RET(0, hmR0SvmExitMwait(pVCpu, pCtx, pSvmTransient));
5757	case SVM_EXIT_HLT: VMEXIT_CALL_RET(0, hmR0SvmExitHlt(pVCpu, pCtx, pSvmTransient));
5758
5759	case SVM_EXIT_XCPT_NMI: /* Should not occur, SVM_EXIT_NMI is used instead. */
5760	case SVM_EXIT_INTR:
5761	case SVM_EXIT_NMI: VMEXIT_CALL_RET(0, hmR0SvmExitIntr(pVCpu, pCtx, pSvmTransient));
5762
5763	case SVM_EXIT_READ_CR0:
5764	case SVM_EXIT_READ_CR3:
5765	case SVM_EXIT_READ_CR4: VMEXIT_CALL_RET(0, hmR0SvmExitReadCRx(pVCpu, pCtx, pSvmTransient));
5766
5767	case SVM_EXIT_CR0_SEL_WRITE:
5768	case SVM_EXIT_WRITE_CR0:
5769	case SVM_EXIT_WRITE_CR3:
5770	case SVM_EXIT_WRITE_CR4:
5771	case SVM_EXIT_WRITE_CR8: VMEXIT_CALL_RET(0, hmR0SvmExitWriteCRx(pVCpu, pCtx, pSvmTransient));
5772
5773	case SVM_EXIT_VINTR: VMEXIT_CALL_RET(0, hmR0SvmExitVIntr(pVCpu, pCtx, pSvmTransient));
5774	case SVM_EXIT_PAUSE: VMEXIT_CALL_RET(0, hmR0SvmExitPause(pVCpu, pCtx, pSvmTransient));
5775	case SVM_EXIT_VMMCALL: VMEXIT_CALL_RET(0, hmR0SvmExitVmmCall(pVCpu, pCtx, pSvmTransient));
5776	case SVM_EXIT_INVLPG: VMEXIT_CALL_RET(0, hmR0SvmExitInvlpg(pVCpu, pCtx, pSvmTransient));
5777	case SVM_EXIT_WBINVD: VMEXIT_CALL_RET(0, hmR0SvmExitWbinvd(pVCpu, pCtx, pSvmTransient));
5778	case SVM_EXIT_INVD: VMEXIT_CALL_RET(0, hmR0SvmExitInvd(pVCpu, pCtx, pSvmTransient));
5779	case SVM_EXIT_RDPMC: VMEXIT_CALL_RET(0, hmR0SvmExitRdpmc(pVCpu, pCtx, pSvmTransient));
5780	case SVM_EXIT_IRET: VMEXIT_CALL_RET(0, hmR0SvmExitIret(pVCpu, pCtx, pSvmTransient));
5781	case SVM_EXIT_XCPT_UD: VMEXIT_CALL_RET(0, hmR0SvmExitXcptUD(pVCpu, pCtx, pSvmTransient));
5782	case SVM_EXIT_XCPT_MF: VMEXIT_CALL_RET(0, hmR0SvmExitXcptMF(pVCpu, pCtx, pSvmTransient));
5783	case SVM_EXIT_XCPT_DB: VMEXIT_CALL_RET(0, hmR0SvmExitXcptDB(pVCpu, pCtx, pSvmTransient));
5784	case SVM_EXIT_XCPT_AC: VMEXIT_CALL_RET(0, hmR0SvmExitXcptAC(pVCpu, pCtx, pSvmTransient));
5785	case SVM_EXIT_XCPT_BP: VMEXIT_CALL_RET(0, hmR0SvmExitXcptBP(pVCpu, pCtx, pSvmTransient));
5786	case SVM_EXIT_XSETBV: VMEXIT_CALL_RET(0, hmR0SvmExitXsetbv(pVCpu, pCtx, pSvmTransient));
5787	case SVM_EXIT_FERR_FREEZE: VMEXIT_CALL_RET(0, hmR0SvmExitFerrFreeze(pVCpu, pCtx, pSvmTransient));
5788
5789	default:
5790	{
5791	switch (pSvmTransient->u64ExitCode)
5792	{
5793	case SVM_EXIT_READ_DR0: case SVM_EXIT_READ_DR1: case SVM_EXIT_READ_DR2: case SVM_EXIT_READ_DR3:
5794	case SVM_EXIT_READ_DR6: case SVM_EXIT_READ_DR7: case SVM_EXIT_READ_DR8: case SVM_EXIT_READ_DR9:
5795	case SVM_EXIT_READ_DR10: case SVM_EXIT_READ_DR11: case SVM_EXIT_READ_DR12: case SVM_EXIT_READ_DR13:
5796	case SVM_EXIT_READ_DR14: case SVM_EXIT_READ_DR15:
5797	VMEXIT_CALL_RET(0, hmR0SvmExitReadDRx(pVCpu, pCtx, pSvmTransient));
5798
5799	case SVM_EXIT_WRITE_DR0: case SVM_EXIT_WRITE_DR1: case SVM_EXIT_WRITE_DR2: case SVM_EXIT_WRITE_DR3:
5800	case SVM_EXIT_WRITE_DR6: case SVM_EXIT_WRITE_DR7: case SVM_EXIT_WRITE_DR8: case SVM_EXIT_WRITE_DR9:
5801	case SVM_EXIT_WRITE_DR10: case SVM_EXIT_WRITE_DR11: case SVM_EXIT_WRITE_DR12: case SVM_EXIT_WRITE_DR13:
5802	case SVM_EXIT_WRITE_DR14: case SVM_EXIT_WRITE_DR15:
5803	VMEXIT_CALL_RET(0, hmR0SvmExitWriteDRx(pVCpu, pCtx, pSvmTransient));
5804
5805	case SVM_EXIT_TASK_SWITCH: VMEXIT_CALL_RET(0, hmR0SvmExitTaskSwitch(pVCpu, pCtx, pSvmTransient));
5806	case SVM_EXIT_SHUTDOWN: VMEXIT_CALL_RET(0, hmR0SvmExitShutdown(pVCpu, pCtx, pSvmTransient));
5807
5808	case SVM_EXIT_SMI:
5809	case SVM_EXIT_INIT:
5810	{
5811	/*
5812	* We don't intercept SMIs. As for INIT signals, it really shouldn't ever happen here.
5813	* If it ever does, we want to know about it so log the exit code and bail.
5814	*/
5815	VMEXIT_CALL_RET(0, hmR0SvmExitUnexpected(pVCpu, pCtx, pSvmTransient));
5816	}
5817
5818	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
5819	case SVM_EXIT_CLGI: VMEXIT_CALL_RET(0, hmR0SvmExitClgi(pVCpu, pCtx, pSvmTransient));
5820	case SVM_EXIT_STGI: VMEXIT_CALL_RET(0, hmR0SvmExitStgi(pVCpu, pCtx, pSvmTransient));
5821	case SVM_EXIT_VMLOAD: VMEXIT_CALL_RET(0, hmR0SvmExitVmload(pVCpu, pCtx, pSvmTransient));
5822	case SVM_EXIT_VMSAVE: VMEXIT_CALL_RET(0, hmR0SvmExitVmsave(pVCpu, pCtx, pSvmTransient));
5823	case SVM_EXIT_INVLPGA: VMEXIT_CALL_RET(0, hmR0SvmExitInvlpga(pVCpu, pCtx, pSvmTransient));
5824	case SVM_EXIT_VMRUN: VMEXIT_CALL_RET(0, hmR0SvmExitVmrun(pVCpu, pCtx, pSvmTransient));
5825	#else
5826	case SVM_EXIT_CLGI:
5827	case SVM_EXIT_STGI:
5828	case SVM_EXIT_VMLOAD:
5829	case SVM_EXIT_VMSAVE:
5830	case SVM_EXIT_INVLPGA:
5831	case SVM_EXIT_VMRUN:
5832	#endif
5833	case SVM_EXIT_RSM:
5834	case SVM_EXIT_SKINIT:
5835	{
5836	hmR0SvmSetPendingXcptUD(pVCpu);
5837	return VINF_SUCCESS;
5838	}
5839
5840	#ifdef HMSVM_ALWAYS_TRAP_ALL_XCPTS
5841	case SVM_EXIT_XCPT_DE:
5842	/* SVM_EXIT_XCPT_DB: / / Handled above. */
5843	/* SVM_EXIT_XCPT_NMI: / / Handled above. */
5844	/* SVM_EXIT_XCPT_BP: / / Handled above. */
5845	case SVM_EXIT_XCPT_OF:
5846	case SVM_EXIT_XCPT_BR:
5847	/* SVM_EXIT_XCPT_UD: / / Handled above. */
5848	case SVM_EXIT_XCPT_NM:
5849	case SVM_EXIT_XCPT_DF:
5850	case SVM_EXIT_XCPT_CO_SEG_OVERRUN:
5851	case SVM_EXIT_XCPT_TS:
5852	case SVM_EXIT_XCPT_NP:
5853	case SVM_EXIT_XCPT_SS:
5854	case SVM_EXIT_XCPT_GP:
5855	/* SVM_EXIT_XCPT_PF: */
5856	case SVM_EXIT_XCPT_15: /* Reserved. */
5857	/* SVM_EXIT_XCPT_MF: / / Handled above. */
5858	/* SVM_EXIT_XCPT_AC: / / Handled above. */
5859	case SVM_EXIT_XCPT_MC:
5860	case SVM_EXIT_XCPT_XF:
5861	case SVM_EXIT_XCPT_20: case SVM_EXIT_XCPT_21: case SVM_EXIT_XCPT_22: case SVM_EXIT_XCPT_23:
5862	case SVM_EXIT_XCPT_24: case SVM_EXIT_XCPT_25: case SVM_EXIT_XCPT_26: case SVM_EXIT_XCPT_27:
5863	case SVM_EXIT_XCPT_28: case SVM_EXIT_XCPT_29: case SVM_EXIT_XCPT_30: case SVM_EXIT_XCPT_31:
5864	VMEXIT_CALL_RET(0, hmR0SvmExitXcptGeneric(pVCpu, pCtx, pSvmTransient));
5865	#endif /* HMSVM_ALWAYS_TRAP_ALL_XCPTS */
5866
5867	default:
5868	{
5869	AssertMsgFailed(("hmR0SvmHandleExit: Unknown exit code %#RX64\n", uExitCode));
5870	pVCpu->hm.s.u32HMError = uExitCode;
5871	return VERR_SVM_UNKNOWN_EXIT;
5872	}
5873	}
5874	}
5875	}
5876	/* not reached */
5877	#undef VMEXIT_CALL_RET
5878	}
5879
5880
5881	#ifdef DEBUG
5882	/* Is there some generic IPRT define for this that are not in Runtime/internal/\* ?? */
5883	# define HMSVM_ASSERT_PREEMPT_CPUID_VAR() \
5884	RTCPUID const idAssertCpu = RTThreadPreemptIsEnabled(NIL_RTTHREAD) ? NIL_RTCPUID : RTMpCpuId()
5885
5886	# define HMSVM_ASSERT_PREEMPT_CPUID() \
5887	do \
5888	{ \
5889	RTCPUID const idAssertCpuNow = RTThreadPreemptIsEnabled(NIL_RTTHREAD) ? NIL_RTCPUID : RTMpCpuId(); \
5890	AssertMsg(idAssertCpu == idAssertCpuNow, ("SVM %#x, %#x\n", idAssertCpu, idAssertCpuNow)); \
5891	} while (0)
5892
5893	# define HMSVM_VALIDATE_EXIT_HANDLER_PARAMS() \
5894	do { \
5895	AssertPtr(pVCpu); \
5896	AssertPtr(pCtx); \
5897	AssertPtr(pSvmTransient); \
5898	Assert(ASMIntAreEnabled()); \
5899	HMSVM_ASSERT_PREEMPT_SAFE(); \
5900	HMSVM_ASSERT_PREEMPT_CPUID_VAR(); \
5901	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)); \
5902	HMSVM_ASSERT_PREEMPT_SAFE(); \
5903	if (VMMR0IsLogFlushDisabled(pVCpu)) \
5904	HMSVM_ASSERT_PREEMPT_CPUID(); \
5905	} while (0)
5906	#else /* Release builds */
5907	# define HMSVM_VALIDATE_EXIT_HANDLER_PARAMS() do { NOREF(pVCpu); NOREF(pCtx); NOREF(pSvmTransient); } while (0)
5908	#endif
5909
5910
5911	/**
5912	* Worker for hmR0SvmInterpretInvlpg().
5913	*
5914	* @return VBox status code.
5915	* @param pVCpu The cross context virtual CPU structure.
5916	* @param pCpu Pointer to the disassembler state.
5917	* @param pCtx The guest CPU context.
5918	*/
5919	static int hmR0SvmInterpretInvlPgEx(PVMCPU pVCpu, PDISCPUSTATE pCpu, PCPUMCTX pCtx)
5920	{
5921	DISQPVPARAMVAL Param1;
5922	RTGCPTR GCPtrPage;
5923
5924	int rc = DISQueryParamVal(CPUMCTX2CORE(pCtx), pCpu, &pCpu->Param1, &Param1, DISQPVWHICH_SRC);
5925	if (RT_FAILURE(rc))
5926	return VERR_EM_INTERPRETER;
5927
5928	if ( Param1.type == DISQPV_TYPE_IMMEDIATE
5929	\|\| Param1.type == DISQPV_TYPE_ADDRESS)
5930	{
5931	if (!(Param1.flags & (DISQPV_FLAG_32 \| DISQPV_FLAG_64)))
5932	return VERR_EM_INTERPRETER;
5933
5934	GCPtrPage = Param1.val.val64;
5935	VBOXSTRICTRC rc2 = EMInterpretInvlpg(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx), GCPtrPage);
5936	rc = VBOXSTRICTRC_VAL(rc2);
5937	}
5938	else
5939	{
5940	Log4Func(("Invalid parameter type %#x\n", Param1.type));
5941	rc = VERR_EM_INTERPRETER;
5942	}
5943
5944	return rc;
5945	}
5946
5947
5948	/**
5949	* Interprets INVLPG.
5950	*
5951	* @returns VBox status code.
5952	* @retval VINF_* Scheduling instructions.
5953	* @retval VERR_EM_INTERPRETER Something we can't cope with.
5954	* @retval VERR_* Fatal errors.
5955	*
5956	* @param pVM The cross context VM structure.
5957	* @param pVCpu The cross context virtual CPU structure.
5958	* @param pCtx The guest CPU context.
5959	*
5960	* @remarks Updates the RIP if the instruction was executed successfully.
5961	*/
5962	static int hmR0SvmInterpretInvlpg(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
5963	{
5964	/* Only allow 32 & 64 bit code. */
5965	if (CPUMGetGuestCodeBits(pVCpu) != 16)
5966	{
5967	PDISSTATE pDis = &pVCpu->hm.s.DisState;
5968	int rc = EMInterpretDisasCurrent(pVM, pVCpu, pDis, NULL /* pcbInstr */);
5969	if ( RT_SUCCESS(rc)
5970	&& pDis->pCurInstr->uOpcode == OP_INVLPG)
5971	{
5972	rc = hmR0SvmInterpretInvlPgEx(pVCpu, pDis, pCtx);
5973	if (RT_SUCCESS(rc))
5974	pCtx->rip += pDis->cbInstr;
5975	return rc;
5976	}
5977	else
5978	Log4Func(("EMInterpretDisasCurrent failed! rc=%Rrc uOpCode=%#x\n", rc, pDis->pCurInstr->uOpcode));
5979	}
5980	return VERR_EM_INTERPRETER;
5981	}
5982
5983
5984	/**
5985	* Gets the IEM exception flags for the specified SVM event.
5986	*
5987	* @returns The IEM exception flags.
5988	* @param pEvent Pointer to the SVM event.
5989	*
5990	* @remarks This function currently only constructs flags required for
5991	* IEMEvaluateRecursiveXcpt and not the complete flags (e.g. error-code
5992	* and CR2 aspects of an exception are not included).
5993	*/
5994	static uint32_t hmR0SvmGetIemXcptFlags(PCSVMEVENT pEvent)
5995	{
5996	uint8_t const uEventType = pEvent->n.u3Type;
5997	uint32_t fIemXcptFlags;
5998	switch (uEventType)
5999	{
6000	case SVM_EVENT_EXCEPTION:
6001	/*
6002	* Only INT3 and INTO instructions can raise #BP and #OF exceptions.
6003	* See AMD spec. Table 8-1. "Interrupt Vector Source and Cause".
6004	*/
6005	if (pEvent->n.u8Vector == X86_XCPT_BP)
6006	{
6007	fIemXcptFlags = IEM_XCPT_FLAGS_T_SOFT_INT \| IEM_XCPT_FLAGS_BP_INSTR;
6008	break;
6009	}
6010	if (pEvent->n.u8Vector == X86_XCPT_OF)
6011	{
6012	fIemXcptFlags = IEM_XCPT_FLAGS_T_SOFT_INT \| IEM_XCPT_FLAGS_OF_INSTR;
6013	break;
6014	}
6015	/** @todo How do we distinguish ICEBP \#DB from the regular one? */
6016	RT_FALL_THRU();
6017	case SVM_EVENT_NMI:
6018	fIemXcptFlags = IEM_XCPT_FLAGS_T_CPU_XCPT;
6019	break;
6020
6021	case SVM_EVENT_EXTERNAL_IRQ:
6022	fIemXcptFlags = IEM_XCPT_FLAGS_T_EXT_INT;
6023	break;
6024
6025	case SVM_EVENT_SOFTWARE_INT:
6026	fIemXcptFlags = IEM_XCPT_FLAGS_T_SOFT_INT;
6027	break;
6028
6029	default:
6030	fIemXcptFlags = 0;
6031	AssertMsgFailed(("Unexpected event type! uEventType=%#x uVector=%#x", uEventType, pEvent->n.u8Vector));
6032	break;
6033	}
6034	return fIemXcptFlags;
6035	}
6036
6037
6038	/**
6039	* Handle a condition that occurred while delivering an event through the guest
6040	* IDT.
6041	*
6042	* @returns VBox status code (informational error codes included).
6043	* @retval VINF_SUCCESS if we should continue handling the \#VMEXIT.
6044	* @retval VINF_HM_DOUBLE_FAULT if a \#DF condition was detected and we ought to
6045	* continue execution of the guest which will delivery the \#DF.
6046	* @retval VINF_EM_RESET if we detected a triple-fault condition.
6047	* @retval VERR_EM_GUEST_CPU_HANG if we detected a guest CPU hang.
6048	*
6049	* @param pVCpu The cross context virtual CPU structure.
6050	* @param pCtx Pointer to the guest-CPU context.
6051	* @param pSvmTransient Pointer to the SVM transient structure.
6052	*
6053	* @remarks No-long-jump zone!!!
6054	*/
6055	static int hmR0SvmCheckExitDueToEventDelivery(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6056	{
6057	int rc = VINF_SUCCESS;
6058	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
6059	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR2);
6060
6061	Log4(("EXITINTINFO: Pending vectoring event %#RX64 Valid=%RTbool ErrValid=%RTbool Err=%#RX32 Type=%u Vector=%u\n",
6062	pVmcb->ctrl.ExitIntInfo.u, !!pVmcb->ctrl.ExitIntInfo.n.u1Valid, !!pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid,
6063	pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode, pVmcb->ctrl.ExitIntInfo.n.u3Type, pVmcb->ctrl.ExitIntInfo.n.u8Vector));
6064
6065	/*
6066	* The EXITINTINFO (if valid) contains the prior exception (IDT vector) that was trying to
6067	* be delivered to the guest which caused a #VMEXIT which was intercepted (Exit vector).
6068	*
6069	* See AMD spec. 15.7.3 "EXITINFO Pseudo-Code".
6070	*/
6071	if (pVmcb->ctrl.ExitIntInfo.n.u1Valid)
6072	{
6073	IEMXCPTRAISE enmRaise;
6074	IEMXCPTRAISEINFO fRaiseInfo;
6075	bool const fExitIsHwXcpt = pSvmTransient->u64ExitCode - SVM_EXIT_XCPT_0 <= SVM_EXIT_XCPT_31;
6076	uint8_t const uIdtVector = pVmcb->ctrl.ExitIntInfo.n.u8Vector;
6077	if (fExitIsHwXcpt)
6078	{
6079	uint8_t const uExitVector = pSvmTransient->u64ExitCode - SVM_EXIT_XCPT_0;
6080	uint32_t const fIdtVectorFlags = hmR0SvmGetIemXcptFlags(&pVmcb->ctrl.ExitIntInfo);
6081	uint32_t const fExitVectorFlags = IEM_XCPT_FLAGS_T_CPU_XCPT;
6082	enmRaise = IEMEvaluateRecursiveXcpt(pVCpu, fIdtVectorFlags, uIdtVector, fExitVectorFlags, uExitVector, &fRaiseInfo);
6083	}
6084	else
6085	{
6086	/*
6087	* If delivery of an event caused a #VMEXIT that is not an exception (e.g. #NPF)
6088	* then we end up here.
6089	*
6090	* If the event was:
6091	* - a software interrupt, we can re-execute the instruction which will
6092	* regenerate the event.
6093	* - an NMI, we need to clear NMI blocking and re-inject the NMI.
6094	* - a hardware exception or external interrupt, we re-inject it.
6095	*/
6096	fRaiseInfo = IEMXCPTRAISEINFO_NONE;
6097	if (pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_SOFTWARE_INT)
6098	enmRaise = IEMXCPTRAISE_REEXEC_INSTR;
6099	else
6100	enmRaise = IEMXCPTRAISE_PREV_EVENT;
6101	}
6102
6103	switch (enmRaise)
6104	{
6105	case IEMXCPTRAISE_CURRENT_XCPT:
6106	case IEMXCPTRAISE_PREV_EVENT:
6107	{
6108	/* For software interrupts, we shall re-execute the instruction. */
6109	if (!(fRaiseInfo & IEMXCPTRAISEINFO_SOFT_INT_XCPT))
6110	{
6111	RTGCUINTPTR GCPtrFaultAddress = 0;
6112
6113	/* If we are re-injecting an NMI, clear NMI blocking. */
6114	if (pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_NMI)
6115	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_BLOCK_NMIS);
6116
6117	/* Determine a vectoring #PF condition, see comment in hmR0SvmExitXcptPF(). */
6118	if (fRaiseInfo & (IEMXCPTRAISEINFO_EXT_INT_PF \| IEMXCPTRAISEINFO_NMI_PF))
6119	{
6120	pSvmTransient->fVectoringPF = true;
6121	Log4Func(("IDT: Pending vectoring #PF due to delivery of Ext-Int/NMI. uCR2=%#RX64\n", pCtx->cr2));
6122	}
6123	else if ( pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_EXCEPTION
6124	&& uIdtVector == X86_XCPT_PF)
6125	{
6126	/*
6127	* If the previous exception was a #PF, we need to recover the CR2 value.
6128	* This can't happen with shadow paging.
6129	*/
6130	GCPtrFaultAddress = pCtx->cr2;
6131	}
6132
6133	/*
6134	* Without nested paging, when uExitVector is #PF, CR2 value will be updated from the VMCB's
6135	* exit info. fields, if it's a guest #PF, see hmR0SvmExitXcptPF().
6136	*/
6137	Assert(pVmcb->ctrl.ExitIntInfo.n.u3Type != SVM_EVENT_SOFTWARE_INT);
6138	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectPendingReflect);
6139	hmR0SvmSetPendingEvent(pVCpu, &pVmcb->ctrl.ExitIntInfo, GCPtrFaultAddress);
6140
6141	Log4Func(("IDT: Pending vectoring event %#RX64 ErrValid=%RTbool Err=%#RX32 GCPtrFaultAddress=%#RX64\n",
6142	pVmcb->ctrl.ExitIntInfo.u, RT_BOOL(pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid),
6143	pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode, GCPtrFaultAddress));
6144	}
6145	break;
6146	}
6147
6148	case IEMXCPTRAISE_REEXEC_INSTR:
6149	{
6150	Assert(rc == VINF_SUCCESS);
6151	break;
6152	}
6153
6154	case IEMXCPTRAISE_DOUBLE_FAULT:
6155	{
6156	/*
6157	* Determing a vectoring double #PF condition. Used later, when PGM evaluates
6158	* the second #PF as a guest #PF (and not a shadow #PF) and needs to be
6159	* converted into a #DF.
6160	*/
6161	if (fRaiseInfo & IEMXCPTRAISEINFO_PF_PF)
6162	{
6163	Log4Func(("IDT: Pending vectoring double #PF uCR2=%#RX64\n", pCtx->cr2));
6164	pSvmTransient->fVectoringDoublePF = true;
6165	Assert(rc == VINF_SUCCESS);
6166	}
6167	else
6168	{
6169	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectPendingReflect);
6170	hmR0SvmSetPendingXcptDF(pVCpu);
6171	rc = VINF_HM_DOUBLE_FAULT;
6172	}
6173	break;
6174	}
6175
6176	case IEMXCPTRAISE_TRIPLE_FAULT:
6177	{
6178	rc = VINF_EM_RESET;
6179	break;
6180	}
6181
6182	case IEMXCPTRAISE_CPU_HANG:
6183	{
6184	rc = VERR_EM_GUEST_CPU_HANG;
6185	break;
6186	}
6187
6188	default:
6189	{
6190	AssertMsgFailed(("hmR0SvmExitCpuid: EMInterpretCpuId failed with %Rrc\n", rc));
6191	rc = VERR_SVM_IPE_2;
6192	break;
6193	}
6194	}
6195	}
6196	Assert(rc == VINF_SUCCESS \|\| rc == VINF_HM_DOUBLE_FAULT \|\| rc == VINF_EM_RESET \|\| rc == VERR_EM_GUEST_CPU_HANG);
6197	NOREF(pCtx);
6198	return rc;
6199	}
6200
6201
6202	/**
6203	* Advances the guest RIP making use of the CPU's NRIP_SAVE feature if
6204	* supported, otherwise advances the RIP by the number of bytes specified in
6205	* @a cb.
6206	*
6207	* @param pVCpu The cross context virtual CPU structure.
6208	* @param pCtx Pointer to the guest-CPU context.
6209	* @param cb RIP increment value in bytes.
6210	*
6211	* @remarks Use this function only from \#VMEXIT's where the NRIP value is valid
6212	* when NRIP_SAVE is supported by the CPU, otherwise use
6213	* hmR0SvmAdvanceRipDumb!
6214	*/
6215	DECLINLINE(void) hmR0SvmAdvanceRipHwAssist(PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t cb)
6216	{
6217	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu, pCtx);
6218	if (fSupportsNextRipSave)
6219	{
6220	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
6221	Assert(pVmcb);
6222	Assert(pVmcb->ctrl.u64NextRIP);
6223	Assert(!(pCtx->fExtrn & CPUMCTX_EXTRN_RIP));
6224	AssertRelease(pVmcb->ctrl.u64NextRIP - pCtx->rip == cb); /* temporary, remove later */
6225	pCtx->rip = pVmcb->ctrl.u64NextRIP;
6226	}
6227	else
6228	pCtx->rip += cb;
6229
6230	HMSVM_UPDATE_INTR_SHADOW(pVCpu, pCtx);
6231	}
6232
6233
6234	/**
6235	* Gets the length of the current instruction if the CPU supports the NRIP_SAVE
6236	* feature. Otherwise, returns the value in @a cbLikely.
6237	*
6238	* @param pVCpu The cross context virtual CPU structure.
6239	* @param pCtx Pointer to the guest-CPU context.
6240	* @param cbLikely The likely instruction length.
6241	*/
6242	DECLINLINE(uint8_t) hmR0SvmGetInstrLengthHwAssist(PVMCPU pVCpu, PCPUMCTX pCtx, uint8_t cbLikely)
6243	{
6244	Assert(cbLikely <= 15); /* See Intel spec. 2.3.11 "AVX Instruction Length" */
6245	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu, pCtx);
6246	if (fSupportsNextRipSave)
6247	{
6248	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
6249	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pCtx->rip;
6250	Assert(cbInstr == cbLikely);
6251	return cbInstr;
6252	}
6253	return cbLikely;
6254	}
6255
6256
6257	/**
6258	* Advances the guest RIP by the number of bytes specified in @a cb. This does
6259	* not make use of any hardware features to determine the instruction length.
6260	*
6261	* @param pVCpu The cross context virtual CPU structure.
6262	* @param pCtx Pointer to the guest-CPU context.
6263	* @param cb RIP increment value in bytes.
6264	*/
6265	DECLINLINE(void) hmR0SvmAdvanceRipDumb(PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t cb)
6266	{
6267	pCtx->rip += cb;
6268	HMSVM_UPDATE_INTR_SHADOW(pVCpu, pCtx);
6269	}
6270	#undef HMSVM_UPDATE_INTR_SHADOW
6271
6272
6273	/* -=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= */
6274	/* -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- #VMEXIT handlers -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- */
6275	/* -=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= */
6276
6277	/** @name \#VMEXIT handlers.
6278	* @{
6279	*/
6280
6281	/**
6282	* \#VMEXIT handler for external interrupts, NMIs, FPU assertion freeze and INIT
6283	* signals (SVM_EXIT_INTR, SVM_EXIT_NMI, SVM_EXIT_FERR_FREEZE, SVM_EXIT_INIT).
6284	*/
6285	HMSVM_EXIT_DECL hmR0SvmExitIntr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6286	{
6287	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6288
6289	if (pSvmTransient->u64ExitCode == SVM_EXIT_NMI)
6290	STAM_REL_COUNTER_INC(&pVCpu->hm.s.StatExitHostNmiInGC);
6291	else if (pSvmTransient->u64ExitCode == SVM_EXIT_INTR)
6292	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitExtInt);
6293
6294	/*
6295	* AMD-V has no preemption timer and the generic periodic preemption timer has no way to
6296	* signal -before- the timer fires if the current interrupt is our own timer or a some
6297	* other host interrupt. We also cannot examine what interrupt it is until the host
6298	* actually take the interrupt.
6299	*
6300	* Going back to executing guest code here unconditionally causes random scheduling
6301	* problems (observed on an AMD Phenom 9850 Quad-Core on Windows 64-bit host).
6302	*/
6303	return VINF_EM_RAW_INTERRUPT;
6304	}
6305
6306
6307	/**
6308	* \#VMEXIT handler for WBINVD (SVM_EXIT_WBINVD). Conditional \#VMEXIT.
6309	*/
6310	HMSVM_EXIT_DECL hmR0SvmExitWbinvd(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6311	{
6312	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6313
6314	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 2);
6315	int rc = VINF_SUCCESS;
6316	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6317	return rc;
6318	}
6319
6320
6321	/**
6322	* \#VMEXIT handler for INVD (SVM_EXIT_INVD). Unconditional \#VMEXIT.
6323	*/
6324	HMSVM_EXIT_DECL hmR0SvmExitInvd(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6325	{
6326	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6327
6328	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 2);
6329	int rc = VINF_SUCCESS;
6330	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6331	return rc;
6332	}
6333
6334
6335	/**
6336	* \#VMEXIT handler for INVD (SVM_EXIT_CPUID). Conditional \#VMEXIT.
6337	*/
6338	HMSVM_EXIT_DECL hmR0SvmExitCpuid(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6339	{
6340	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6341
6342	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_RIP
6343	\| CPUMCTX_EXTRN_CS);
6344	VBOXSTRICTRC rcStrict;
6345	PCEMEXITREC pExitRec = EMHistoryUpdateFlagsAndTypeAndPC(pVCpu,
6346	EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_CPUID),
6347	pVCpu->cpum.GstCtx.rip + pVCpu->cpum.GstCtx.cs.u64Base);
6348	if (!pExitRec)
6349	{
6350	PVM pVM = pVCpu->CTX_SUFF(pVM);
6351	rcStrict = EMInterpretCpuId(pVM, pVCpu, CPUMCTX2CORE(pCtx));
6352	if (RT_LIKELY(rcStrict == VINF_SUCCESS))
6353	{
6354	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 2);
6355	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6356	}
6357	else
6358	{
6359	AssertMsgFailed(("hmR0SvmExitCpuid: EMInterpretCpuId failed with %Rrc\n", VBOXSTRICTRC_VAL(rcStrict)));
6360	rcStrict = VERR_EM_INTERPRETER;
6361	}
6362	}
6363	else
6364	{
6365	/*
6366	* Frequent exit or something needing probing. Get state and call EMHistoryExec.
6367	*/
6368	Assert(pCtx == &pVCpu->cpum.GstCtx);
6369	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6370
6371	Log4(("CpuIdExit/%u: %04x:%08RX64: %#x/%#x -> EMHistoryExec\n",
6372	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip, pVCpu->cpum.GstCtx.eax, pVCpu->cpum.GstCtx.ecx));
6373
6374	rcStrict = EMHistoryExec(pVCpu, pExitRec, 0);
6375
6376	Log4(("CpuIdExit/%u: %04x:%08RX64: EMHistoryExec -> %Rrc + %04x:%08RX64\n",
6377	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip,
6378	VBOXSTRICTRC_VAL(rcStrict), pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip));
6379	}
6380	return VBOXSTRICTRC_TODO(rcStrict);
6381	}
6382
6383
6384	/**
6385	* \#VMEXIT handler for RDTSC (SVM_EXIT_RDTSC). Conditional \#VMEXIT.
6386	*/
6387	HMSVM_EXIT_DECL hmR0SvmExitRdtsc(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6388	{
6389	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6390	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6391	VBOXSTRICTRC rcStrict = IEMExecDecodedRdtsc(pVCpu, hmR0SvmGetInstrLengthHwAssist(pVCpu, pCtx, 2));
6392	if (rcStrict == VINF_SUCCESS)
6393	pSvmTransient->fUpdateTscOffsetting = true;
6394	else if (rcStrict == VINF_IEM_RAISED_XCPT)
6395	rcStrict = VINF_SUCCESS;
6396	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6397	return VBOXSTRICTRC_TODO(rcStrict);
6398	}
6399
6400
6401	/**
6402	* \#VMEXIT handler for RDTSCP (SVM_EXIT_RDTSCP). Conditional \#VMEXIT.
6403	*/
6404	HMSVM_EXIT_DECL hmR0SvmExitRdtscp(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6405	{
6406	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6407	VBOXSTRICTRC rcStrict = IEMExecDecodedRdtscp(pVCpu, hmR0SvmGetInstrLengthHwAssist(pVCpu, pCtx, 3));
6408	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6409	if (rcStrict == VINF_SUCCESS)
6410	pSvmTransient->fUpdateTscOffsetting = true;
6411	else if (rcStrict == VINF_IEM_RAISED_XCPT)
6412	rcStrict = VINF_SUCCESS;
6413	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6414	return VBOXSTRICTRC_TODO(rcStrict);
6415	}
6416
6417
6418	/**
6419	* \#VMEXIT handler for RDPMC (SVM_EXIT_RDPMC). Conditional \#VMEXIT.
6420	*/
6421	HMSVM_EXIT_DECL hmR0SvmExitRdpmc(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6422	{
6423	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6424	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR0
6425	\| CPUMCTX_EXTRN_CR4
6426	\| CPUMCTX_EXTRN_SS);
6427
6428	int rc = EMInterpretRdpmc(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
6429	if (RT_LIKELY(rc == VINF_SUCCESS))
6430	{
6431	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 2);
6432	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6433	}
6434	else
6435	{
6436	AssertMsgFailed(("hmR0SvmExitRdpmc: EMInterpretRdpmc failed with %Rrc\n", rc));
6437	rc = VERR_EM_INTERPRETER;
6438	}
6439	return rc;
6440	}
6441
6442
6443	/**
6444	* \#VMEXIT handler for INVLPG (SVM_EXIT_INVLPG). Conditional \#VMEXIT.
6445	*/
6446	HMSVM_EXIT_DECL hmR0SvmExitInvlpg(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6447	{
6448	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6449	PVM pVM = pVCpu->CTX_SUFF(pVM);
6450	Assert(!pVM->hm.s.fNestedPaging);
6451
6452	bool const fSupportsDecodeAssists = hmR0SvmSupportsDecodeAssists(pVCpu, pCtx);
6453	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu, pCtx);
6454	if ( fSupportsDecodeAssists
6455	&& fSupportsNextRipSave)
6456	{
6457	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6458	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
6459	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pCtx->rip;
6460	RTGCPTR const GCPtrPage = pVmcb->ctrl.u64ExitInfo1;
6461	VBOXSTRICTRC rcStrict = IEMExecDecodedInvlpg(pVCpu, cbInstr, GCPtrPage);
6462	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6463	return VBOXSTRICTRC_VAL(rcStrict);
6464	}
6465
6466	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6467	int rc = hmR0SvmInterpretInvlpg(pVM, pVCpu, pCtx); /* Updates RIP if successful. */
6468	Assert(rc == VINF_SUCCESS \|\| rc == VERR_EM_INTERPRETER);
6469	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6470	return rc;
6471	}
6472
6473
6474	/**
6475	* \#VMEXIT handler for HLT (SVM_EXIT_HLT). Conditional \#VMEXIT.
6476	*/
6477	HMSVM_EXIT_DECL hmR0SvmExitHlt(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6478	{
6479	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6480
6481	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 1);
6482	int rc = EMShouldContinueAfterHalt(pVCpu, pCtx) ? VINF_SUCCESS : VINF_EM_HALT;
6483	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6484
6485	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitHlt);
6486	if (rc != VINF_SUCCESS)
6487	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHltToR3);
6488	return rc;
6489	}
6490
6491
6492	/**
6493	* \#VMEXIT handler for MONITOR (SVM_EXIT_MONITOR). Conditional \#VMEXIT.
6494	*/
6495	HMSVM_EXIT_DECL hmR0SvmExitMonitor(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6496	{
6497	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6498	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR0
6499	\| CPUMCTX_EXTRN_SS);
6500
6501	int rc = EMInterpretMonitor(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
6502	if (RT_LIKELY(rc == VINF_SUCCESS))
6503	{
6504	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 3);
6505	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6506	}
6507	else
6508	{
6509	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMonitor: EMInterpretMonitor failed with %Rrc\n", rc));
6510	rc = VERR_EM_INTERPRETER;
6511	}
6512	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitMonitor);
6513	return rc;
6514	}
6515
6516
6517	/**
6518	* \#VMEXIT handler for MWAIT (SVM_EXIT_MWAIT). Conditional \#VMEXIT.
6519	*/
6520	HMSVM_EXIT_DECL hmR0SvmExitMwait(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6521	{
6522	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6523	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR0
6524	\| CPUMCTX_EXTRN_SS);
6525
6526	VBOXSTRICTRC rc2 = EMInterpretMWait(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
6527	int rc = VBOXSTRICTRC_VAL(rc2);
6528	if ( rc == VINF_EM_HALT
6529	\|\| rc == VINF_SUCCESS)
6530	{
6531	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 3);
6532
6533	if ( rc == VINF_EM_HALT
6534	&& EMMonitorWaitShouldContinue(pVCpu, pCtx))
6535	{
6536	rc = VINF_SUCCESS;
6537	}
6538	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6539	}
6540	else
6541	{
6542	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMwait: EMInterpretMWait failed with %Rrc\n", rc));
6543	rc = VERR_EM_INTERPRETER;
6544	}
6545	AssertMsg(rc == VINF_SUCCESS \|\| rc == VINF_EM_HALT \|\| rc == VERR_EM_INTERPRETER,
6546	("hmR0SvmExitMwait: EMInterpretMWait failed rc=%Rrc\n", rc));
6547	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitMwait);
6548	return rc;
6549	}
6550
6551
6552	/**
6553	* \#VMEXIT handler for shutdown (triple-fault) (SVM_EXIT_SHUTDOWN). Conditional
6554	* \#VMEXIT.
6555	*/
6556	HMSVM_EXIT_DECL hmR0SvmExitShutdown(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6557	{
6558	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6559	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6560	return VINF_EM_RESET;
6561	}
6562
6563
6564	/**
6565	* \#VMEXIT handler for unexpected exits. Conditional \#VMEXIT.
6566	*/
6567	HMSVM_EXIT_DECL hmR0SvmExitUnexpected(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6568	{
6569	RT_NOREF(pCtx);
6570	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
6571	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6572	AssertMsgFailed(("hmR0SvmExitUnexpected: ExitCode=%#RX64 uExitInfo1=%#RX64 uExitInfo2=%#RX64\n", pSvmTransient->u64ExitCode,
6573	pVmcb->ctrl.u64ExitInfo1, pVmcb->ctrl.u64ExitInfo2));
6574	RT_NOREF(pVmcb);
6575	pVCpu->hm.s.u32HMError = (uint32_t)pSvmTransient->u64ExitCode;
6576	return VERR_SVM_UNEXPECTED_EXIT;
6577	}
6578
6579
6580	/**
6581	* \#VMEXIT handler for CRx reads (SVM_EXIT_READ_CR*). Conditional \#VMEXIT.
6582	*/
6583	HMSVM_EXIT_DECL hmR0SvmExitReadCRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6584	{
6585	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6586
6587	Log4Func(("CS:RIP=%04x:%#RX64\n", pCtx->cs.Sel, pCtx->rip));
6588	#ifdef VBOX_WITH_STATISTICS
6589	switch (pSvmTransient->u64ExitCode)
6590	{
6591	case SVM_EXIT_READ_CR0: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR0Read); break;
6592	case SVM_EXIT_READ_CR2: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR2Read); break;
6593	case SVM_EXIT_READ_CR3: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR3Read); break;
6594	case SVM_EXIT_READ_CR4: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR4Read); break;
6595	case SVM_EXIT_READ_CR8: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR8Read); break;
6596	}
6597	#endif
6598
6599	bool const fSupportsDecodeAssists = hmR0SvmSupportsDecodeAssists(pVCpu, pCtx);
6600	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu, pCtx);
6601	if ( fSupportsDecodeAssists
6602	&& fSupportsNextRipSave)
6603	{
6604	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
6605	bool const fMovCRx = RT_BOOL(pVmcb->ctrl.u64ExitInfo1 & SVM_EXIT1_MOV_CRX_MASK);
6606	if (fMovCRx)
6607	{
6608	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6609	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pCtx->rip;
6610	uint8_t const iCrReg = pSvmTransient->u64ExitCode - SVM_EXIT_READ_CR0;
6611	uint8_t const iGReg = pVmcb->ctrl.u64ExitInfo1 & SVM_EXIT1_MOV_CRX_GPR_NUMBER;
6612	VBOXSTRICTRC rcStrict = IEMExecDecodedMovCRxRead(pVCpu, cbInstr, iGReg, iCrReg);
6613	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6614	return VBOXSTRICTRC_VAL(rcStrict);
6615	}
6616	/* else: SMSW instruction, fall back below to IEM for this. */
6617	}
6618
6619	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6620	VBOXSTRICTRC rc2 = EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */);
6621	int rc = VBOXSTRICTRC_VAL(rc2);
6622	AssertMsg(rc == VINF_SUCCESS \|\| rc == VERR_EM_INTERPRETER \|\| rc == VINF_PGM_CHANGE_MODE \|\| rc == VINF_PGM_SYNC_CR3,
6623	("hmR0SvmExitReadCRx: EMInterpretInstruction failed rc=%Rrc\n", rc));
6624	Assert((pSvmTransient->u64ExitCode - SVM_EXIT_READ_CR0) <= 15);
6625	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6626	return rc;
6627	}
6628
6629
6630	/**
6631	* \#VMEXIT handler for CRx writes (SVM_EXIT_WRITE_CR*). Conditional \#VMEXIT.
6632	*/
6633	HMSVM_EXIT_DECL hmR0SvmExitWriteCRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6634	{
6635	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6636
6637	uint64_t const uExitCode = pSvmTransient->u64ExitCode;
6638	uint8_t const iCrReg = uExitCode == SVM_EXIT_CR0_SEL_WRITE ? 0 : (pSvmTransient->u64ExitCode - SVM_EXIT_WRITE_CR0);
6639	Assert(iCrReg <= 15);
6640
6641	VBOXSTRICTRC rcStrict = VERR_SVM_IPE_5;
6642	bool fDecodedInstr = false;
6643	bool const fSupportsDecodeAssists = hmR0SvmSupportsDecodeAssists(pVCpu, pCtx);
6644	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu, pCtx);
6645	if ( fSupportsDecodeAssists
6646	&& fSupportsNextRipSave)
6647	{
6648	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
6649	bool const fMovCRx = RT_BOOL(pVmcb->ctrl.u64ExitInfo1 & SVM_EXIT1_MOV_CRX_MASK);
6650	if (fMovCRx)
6651	{
6652	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6653	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pCtx->rip;
6654	uint8_t const iGReg = pVmcb->ctrl.u64ExitInfo1 & SVM_EXIT1_MOV_CRX_GPR_NUMBER;
6655	Log4Func(("Mov CR%u w/ iGReg=%#x\n", iCrReg, iGReg));
6656	rcStrict = IEMExecDecodedMovCRxWrite(pVCpu, cbInstr, iCrReg, iGReg);
6657	fDecodedInstr = true;
6658	}
6659	/* else: LMSW or CLTS instruction, fall back below to IEM for this. */
6660	}
6661
6662	if (!fDecodedInstr)
6663	{
6664	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6665	Log4Func(("iCrReg=%#x\n", iCrReg));
6666	rcStrict = IEMExecOneBypassEx(pVCpu, CPUMCTX2CORE(pCtx), NULL);
6667	if (RT_UNLIKELY( rcStrict == VERR_IEM_ASPECT_NOT_IMPLEMENTED
6668	\|\| rcStrict == VERR_IEM_INSTR_NOT_IMPLEMENTED))
6669	rcStrict = VERR_EM_INTERPRETER;
6670	}
6671
6672	if (rcStrict == VINF_SUCCESS)
6673	{
6674	switch (iCrReg)
6675	{
6676	case 0:
6677	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_CR0);
6678	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR0Write);
6679	break;
6680
6681	case 2:
6682	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_CR2);
6683	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR2Write);
6684	break;
6685
6686	case 3:
6687	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_CR3);
6688	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR3Write);
6689	break;
6690
6691	case 4:
6692	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_CR4);
6693	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR4Write);
6694	break;
6695
6696	case 8:
6697	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
6698	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR8Write);
6699	break;
6700
6701	default:
6702	{
6703	AssertMsgFailed(("hmR0SvmExitWriteCRx: Invalid/Unexpected Write-CRx exit. u64ExitCode=%#RX64 %#x\n",
6704	pSvmTransient->u64ExitCode, iCrReg));
6705	break;
6706	}
6707	}
6708	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6709	}
6710	else
6711	Assert(rcStrict == VERR_EM_INTERPRETER \|\| rcStrict == VINF_PGM_CHANGE_MODE \|\| rcStrict == VINF_PGM_SYNC_CR3);
6712	return VBOXSTRICTRC_TODO(rcStrict);
6713	}
6714
6715
6716	/**
6717	* \#VMEXIT helper for read MSRs, see hmR0SvmExitMsr.
6718	*
6719	* @returns VBox status code.
6720	* @param pVCpu The cross context virtual CPU structure.
6721	* @param pVmcb Pointer to the VM control block.
6722	*/
6723	static int hmR0SvmExitReadMsr(PVMCPU pVCpu, PSVMVMCB pVmcb)
6724	{
6725	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
6726	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR0
6727	\| CPUMCTX_EXTRN_RFLAGS
6728	\| CPUMCTX_EXTRN_SS
6729	\| CPUMCTX_EXTRN_ALL_MSRS);
6730
6731	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdmsr);
6732	Log4Func(("idMsr=%#RX32\n", pCtx->ecx));
6733
6734	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu, pCtx);
6735	if (fSupportsNextRipSave)
6736	{
6737	int rc = EMInterpretRdmsr(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
6738	if (RT_LIKELY(rc == VINF_SUCCESS))
6739	{
6740	pCtx->rip = pVmcb->ctrl.u64NextRIP;
6741	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6742	return VINF_SUCCESS;
6743	}
6744
6745	AssertMsg( rc == VERR_EM_INTERPRETER
6746	\|\| rc == VINF_CPUM_R3_MSR_READ, ("EMInterpretRdmsr failed rc=%Rrc\n", rc));
6747	return rc;
6748	}
6749
6750	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6751	int rc = VBOXSTRICTRC_TODO(EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0));
6752	if (RT_UNLIKELY(rc != VINF_SUCCESS))
6753	{
6754	AssertMsg( rc == VERR_EM_INTERPRETER
6755	\|\| rc == VINF_CPUM_R3_MSR_READ, ("EMInterpretInstruction failed rc=%Rrc\n", rc));
6756	}
6757	/* RIP updated by EMInterpretInstruction(). */
6758	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6759	return rc;
6760	}
6761
6762
6763	/**
6764	* \#VMEXIT helper for write MSRs, see hmR0SvmExitMsr.
6765	*
6766	* @returns VBox status code.
6767	* @param pVCpu The cross context virtual CPU structure.
6768	* @param pVmcb Pointer to the VM control block.
6769	* @param pSvmTransient Pointer to the SVM-transient structure.
6770	*/
6771	static int hmR0SvmExitWriteMsr(PVMCPU pVCpu, PSVMVMCB pVmcb, PSVMTRANSIENT pSvmTransient)
6772	{
6773	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
6774	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR0
6775	\| CPUMCTX_EXTRN_RFLAGS
6776	\| CPUMCTX_EXTRN_SS
6777	\| CPUMCTX_EXTRN_ALL_MSRS);
6778
6779	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitWrmsr);
6780	Log4Func(("idMsr=%#RX32\n", pCtx->ecx));
6781
6782	/*
6783	* Handle TPR patching MSR writes.
6784	* We utilitize the LSTAR MSR for patching.
6785	*/
6786	if ( pVCpu->CTX_SUFF(pVM)->hm.s.fTPRPatchingActive
6787	&& pCtx->ecx == MSR_K8_LSTAR)
6788	{
6789	if ((pCtx->eax & 0xff) != pSvmTransient->u8GuestTpr)
6790	{
6791	/* Our patch code uses LSTAR for TPR caching for 32-bit guests. */
6792	int rc2 = APICSetTpr(pVCpu, pCtx->eax & 0xff);
6793	AssertRC(rc2);
6794	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
6795	}
6796
6797	int rc = VINF_SUCCESS;
6798	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 2);
6799	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6800	return rc;
6801	}
6802
6803	/*
6804	* Handle regular MSR writes.
6805	*/
6806	int rc;
6807	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu, pCtx);
6808	if (fSupportsNextRipSave)
6809	{
6810	rc = EMInterpretWrmsr(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
6811	if (RT_LIKELY(rc == VINF_SUCCESS))
6812	{
6813	pCtx->rip = pVmcb->ctrl.u64NextRIP;
6814	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6815	}
6816	else
6817	AssertMsg( rc == VERR_EM_INTERPRETER
6818	\|\| rc == VINF_CPUM_R3_MSR_WRITE, ("hmR0SvmExitMsr: EMInterpretWrmsr failed rc=%Rrc\n", rc));
6819	}
6820	else
6821	{
6822	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6823	rc = VBOXSTRICTRC_TODO(EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */));
6824	if (RT_LIKELY(rc == VINF_SUCCESS))
6825	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc); /* RIP updated by EMInterpretInstruction(). */
6826	else
6827	AssertMsg( rc == VERR_EM_INTERPRETER
6828	\|\| rc == VINF_CPUM_R3_MSR_WRITE, ("hmR0SvmExitMsr: WrMsr. EMInterpretInstruction failed rc=%Rrc\n", rc));
6829	}
6830
6831	if (rc == VINF_SUCCESS)
6832	{
6833	/* If this is an X2APIC WRMSR access, update the APIC TPR state. */
6834	if ( pCtx->ecx >= MSR_IA32_X2APIC_START
6835	&& pCtx->ecx <= MSR_IA32_X2APIC_END)
6836	{
6837	/*
6838	* We've already saved the APIC related guest-state (TPR) in hmR0SvmPostRunGuest().
6839	* When full APIC register virtualization is implemented we'll have to make sure
6840	* APIC state is saved from the VMCB before EMInterpretWrmsr() changes it.
6841	*/
6842	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
6843	}
6844	else
6845	{
6846	switch (pCtx->ecx)
6847	{
6848	case MSR_IA32_TSC: pSvmTransient->fUpdateTscOffsetting = true; break;
6849	case MSR_K6_EFER: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_EFER_MSR); break;
6850	case MSR_K8_FS_BASE: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_FS); break;
6851	case MSR_K8_GS_BASE: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_GS); break;
6852	case MSR_IA32_SYSENTER_CS: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_SYSENTER_CS_MSR); break;
6853	case MSR_IA32_SYSENTER_EIP: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_SYSENTER_EIP_MSR); break;
6854	case MSR_IA32_SYSENTER_ESP: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_SYSENTER_ESP_MSR); break;
6855	}
6856	}
6857	}
6858
6859	/* RIP has been updated by above after EMInterpretWrmsr() or by EMInterpretInstruction(). */
6860	return rc;
6861	}
6862
6863
6864	/**
6865	* \#VMEXIT handler for MSR read and writes (SVM_EXIT_MSR). Conditional
6866	* \#VMEXIT.
6867	*/
6868	HMSVM_EXIT_DECL hmR0SvmExitMsr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6869	{
6870	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6871
6872	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
6873	if (pVmcb->ctrl.u64ExitInfo1 == SVM_EXIT1_MSR_READ)
6874	return hmR0SvmExitReadMsr(pVCpu, pVmcb);
6875
6876	Assert(pVmcb->ctrl.u64ExitInfo1 == SVM_EXIT1_MSR_WRITE);
6877	return hmR0SvmExitWriteMsr(pVCpu, pVmcb, pSvmTransient);
6878	}
6879
6880
6881	/**
6882	* \#VMEXIT handler for DRx read (SVM_EXIT_READ_DRx). Conditional \#VMEXIT.
6883	*/
6884	HMSVM_EXIT_DECL hmR0SvmExitReadDRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6885	{
6886	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6887	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6888
6889	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitDRxRead);
6890
6891	/** @todo Stepping with nested-guest. */
6892	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
6893	{
6894	/* We should -not- get this #VMEXIT if the guest's debug registers were active. */
6895	if (pSvmTransient->fWasGuestDebugStateActive)
6896	{
6897	AssertMsgFailed(("hmR0SvmExitReadDRx: Unexpected exit %#RX32\n", (uint32_t)pSvmTransient->u64ExitCode));
6898	pVCpu->hm.s.u32HMError = (uint32_t)pSvmTransient->u64ExitCode;
6899	return VERR_SVM_UNEXPECTED_EXIT;
6900	}
6901
6902	/*
6903	* Lazy DR0-3 loading.
6904	*/
6905	if (!pSvmTransient->fWasHyperDebugStateActive)
6906	{
6907	Assert(!DBGFIsStepping(pVCpu)); Assert(!pVCpu->hm.s.fSingleInstruction);
6908	Log5(("hmR0SvmExitReadDRx: Lazy loading guest debug registers\n"));
6909
6910	/* Don't intercept DRx read and writes. */
6911	PSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
6912	pVmcb->ctrl.u16InterceptRdDRx = 0;
6913	pVmcb->ctrl.u16InterceptWrDRx = 0;
6914	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
6915
6916	/* We're playing with the host CPU state here, make sure we don't preempt or longjmp. */
6917	VMMRZCallRing3Disable(pVCpu);
6918	HM_DISABLE_PREEMPT();
6919
6920	/* Save the host & load the guest debug state, restart execution of the MOV DRx instruction. */
6921	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
6922	Assert(CPUMIsGuestDebugStateActive(pVCpu) \|\| HC_ARCH_BITS == 32);
6923
6924	HM_RESTORE_PREEMPT();
6925	VMMRZCallRing3Enable(pVCpu);
6926
6927	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxContextSwitch);
6928	return VINF_SUCCESS;
6929	}
6930	}
6931
6932	/*
6933	* Interpret the read/writing of DRx.
6934	*/
6935	/** @todo Decode assist. */
6936	VBOXSTRICTRC rc = EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */);
6937	Log5(("hmR0SvmExitReadDRx: Emulated DRx access: rc=%Rrc\n", VBOXSTRICTRC_VAL(rc)));
6938	if (RT_LIKELY(rc == VINF_SUCCESS))
6939	{
6940	/* Not necessary for read accesses but whatever doesn't hurt for now, will be fixed with decode assist. */
6941	/** @todo CPUM should set this flag! */
6942	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_DR_MASK);
6943	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6944	}
6945	else
6946	Assert(rc == VERR_EM_INTERPRETER);
6947	return VBOXSTRICTRC_TODO(rc);
6948	}
6949
6950
6951	/**
6952	* \#VMEXIT handler for DRx write (SVM_EXIT_WRITE_DRx). Conditional \#VMEXIT.
6953	*/
6954	HMSVM_EXIT_DECL hmR0SvmExitWriteDRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6955	{
6956	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6957	/* For now it's the same since we interpret the instruction anyway. Will change when using of Decode Assist is implemented. */
6958	int rc = hmR0SvmExitReadDRx(pVCpu, pCtx, pSvmTransient);
6959	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitDRxWrite);
6960	STAM_COUNTER_DEC(&pVCpu->hm.s.StatExitDRxRead);
6961	return rc;
6962	}
6963
6964
6965	/**
6966	* \#VMEXIT handler for XCRx write (SVM_EXIT_XSETBV). Conditional \#VMEXIT.
6967	*/
6968	HMSVM_EXIT_DECL hmR0SvmExitXsetbv(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6969	{
6970	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6971	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6972
6973	/** @todo decode assists... */
6974	VBOXSTRICTRC rcStrict = IEMExecOne(pVCpu);
6975	if (rcStrict == VINF_IEM_RAISED_XCPT)
6976	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_XCPT_RAISED_MASK);
6977
6978	pVCpu->hm.s.fLoadSaveGuestXcr0 = (pCtx->cr4 & X86_CR4_OSXSAVE) && pCtx->aXcr[0] != ASMGetXcr0();
6979	Log4Func(("New XCR0=%#RX64 fLoadSaveGuestXcr0=%d (cr4=%RX64) rcStrict=%Rrc\n", pCtx->aXcr[0], pVCpu->hm.s.fLoadSaveGuestXcr0,
6980	pCtx->cr4, VBOXSTRICTRC_VAL(rcStrict)));
6981
6982	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6983	return VBOXSTRICTRC_TODO(rcStrict);
6984	}
6985
6986
6987	/**
6988	* \#VMEXIT handler for I/O instructions (SVM_EXIT_IOIO). Conditional \#VMEXIT.
6989	*/
6990	HMSVM_EXIT_DECL hmR0SvmExitIOInstr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6991	{
6992	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6993	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK
6994	\| CPUMCTX_EXTRN_SREG_MASK);
6995
6996	/* I/O operation lookup arrays. */
6997	static uint32_t const s_aIOSize[8] = { 0, 1, 2, 0, 4, 0, 0, 0 }; /* Size of the I/O accesses in bytes. */
6998	static uint32_t const s_aIOOpAnd[8] = { 0, 0xff, 0xffff, 0, 0xffffffff, 0, 0, 0 }; /* AND masks for saving
6999	the result (in AL/AX/EAX). */
7000	Log4Func(("CS:RIP=%04x:%#RX64\n", pCtx->cs.Sel, pCtx->rip));
7001
7002	PVM pVM = pVCpu->CTX_SUFF(pVM);
7003	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7004
7005	/* Refer AMD spec. 15.10.2 "IN and OUT Behaviour" and Figure 15-2. "EXITINFO1 for IOIO Intercept" for the format. */
7006	SVMIOIOEXITINFO IoExitInfo;
7007	IoExitInfo.u = (uint32_t)pVmcb->ctrl.u64ExitInfo1;
7008	uint32_t uIOWidth = (IoExitInfo.u >> 4) & 0x7;
7009	uint32_t cbValue = s_aIOSize[uIOWidth];
7010	uint32_t uAndVal = s_aIOOpAnd[uIOWidth];
7011
7012	if (RT_UNLIKELY(!cbValue))
7013	{
7014	AssertMsgFailed(("hmR0SvmExitIOInstr: Invalid IO operation. uIOWidth=%u\n", uIOWidth));
7015	return VERR_EM_INTERPRETER;
7016	}
7017
7018	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_RIP
7019	\| CPUMCTX_EXTRN_CS
7020	\| CPUMCTX_EXTRN_RFLAGS);
7021	VBOXSTRICTRC rcStrict;
7022	PCEMEXITREC pExitRec = NULL;
7023	if ( !pVCpu->hm.s.fSingleInstruction
7024	&& !pVCpu->cpum.GstCtx.eflags.Bits.u1TF)
7025	pExitRec = EMHistoryUpdateFlagsAndTypeAndPC(pVCpu,
7026	!IoExitInfo.n.u1Str
7027	? IoExitInfo.n.u1Type == SVM_IOIO_READ
7028	? EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_IO_PORT_READ)
7029	: EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_IO_PORT_WRITE)
7030	: IoExitInfo.n.u1Type == SVM_IOIO_READ
7031	? EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_IO_PORT_STR_READ)
7032	: EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_IO_PORT_STR_WRITE),
7033	pVCpu->cpum.GstCtx.rip + pVCpu->cpum.GstCtx.cs.u64Base);
7034	if (!pExitRec)
7035	{
7036	bool fUpdateRipAlready = false;
7037	if (IoExitInfo.n.u1Str)
7038	{
7039	/* INS/OUTS - I/O String instruction. */
7040	/** @todo Huh? why can't we use the segment prefix information given by AMD-V
7041	* in EXITINFO1? Investigate once this thing is up and running. */
7042	Log4Func(("CS:RIP=%04x:%08RX64 %#06x/%u %c str\n", pCtx->cs.Sel, pCtx->rip, IoExitInfo.n.u16Port, cbValue,
7043	IoExitInfo.n.u1Type == SVM_IOIO_WRITE ? 'w' : 'r'));
7044	AssertReturn(pCtx->dx == IoExitInfo.n.u16Port, VERR_SVM_IPE_2);
7045	static IEMMODE const s_aenmAddrMode[8] =
7046	{
7047	(IEMMODE)-1, IEMMODE_16BIT, IEMMODE_32BIT, (IEMMODE)-1, IEMMODE_64BIT, (IEMMODE)-1, (IEMMODE)-1, (IEMMODE)-1
7048	};
7049	IEMMODE enmAddrMode = s_aenmAddrMode[(IoExitInfo.u >> 7) & 0x7];
7050	if (enmAddrMode != (IEMMODE)-1)
7051	{
7052	uint64_t cbInstr = pVmcb->ctrl.u64ExitInfo2 - pCtx->rip;
7053	if (cbInstr <= 15 && cbInstr >= 1)
7054	{
7055	Assert(cbInstr >= 1U + IoExitInfo.n.u1Rep);
7056	if (IoExitInfo.n.u1Type == SVM_IOIO_WRITE)
7057	{
7058	/* Don't know exactly how to detect whether u3Seg is valid, currently
7059	only enabling it for Bulldozer and later with NRIP. OS/2 broke on
7060	2384 Opterons when only checking NRIP. */
7061	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu, pCtx);
7062	if ( fSupportsNextRipSave
7063	&& pVM->cpum.ro.GuestFeatures.enmMicroarch >= kCpumMicroarch_AMD_15h_First)
7064	{
7065	AssertMsg(IoExitInfo.n.u3Seg == X86_SREG_DS \|\| cbInstr > 1U + IoExitInfo.n.u1Rep,
7066	("u32Seg=%d cbInstr=%d u1REP=%d", IoExitInfo.n.u3Seg, cbInstr, IoExitInfo.n.u1Rep));
7067	rcStrict = IEMExecStringIoWrite(pVCpu, cbValue, enmAddrMode, IoExitInfo.n.u1Rep, (uint8_t)cbInstr,
7068	IoExitInfo.n.u3Seg, true /fIoChecked/);
7069	}
7070	else if (cbInstr == 1U + IoExitInfo.n.u1Rep)
7071	rcStrict = IEMExecStringIoWrite(pVCpu, cbValue, enmAddrMode, IoExitInfo.n.u1Rep, (uint8_t)cbInstr,
7072	X86_SREG_DS, true /fIoChecked/);
7073	else
7074	rcStrict = IEMExecOne(pVCpu);
7075	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOStringWrite);
7076	}
7077	else
7078	{
7079	AssertMsg(IoExitInfo.n.u3Seg == X86_SREG_ES /=0/, ("%#x\n", IoExitInfo.n.u3Seg));
7080	rcStrict = IEMExecStringIoRead(pVCpu, cbValue, enmAddrMode, IoExitInfo.n.u1Rep, (uint8_t)cbInstr,
7081	true /fIoChecked/);
7082	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOStringRead);
7083	}
7084	}
7085	else
7086	{
7087	AssertMsgFailed(("rip=%RX64 nrip=%#RX64 cbInstr=%#RX64\n", pCtx->rip, pVmcb->ctrl.u64ExitInfo2, cbInstr));
7088	rcStrict = IEMExecOne(pVCpu);
7089	}
7090	}
7091	else
7092	{
7093	AssertMsgFailed(("IoExitInfo=%RX64\n", IoExitInfo.u));
7094	rcStrict = IEMExecOne(pVCpu);
7095	}
7096	fUpdateRipAlready = true;
7097	}
7098	else
7099	{
7100	/* IN/OUT - I/O instruction. */
7101	Assert(!IoExitInfo.n.u1Rep);
7102
7103	if (IoExitInfo.n.u1Type == SVM_IOIO_WRITE)
7104	{
7105	rcStrict = IOMIOPortWrite(pVM, pVCpu, IoExitInfo.n.u16Port, pCtx->eax & uAndVal, cbValue);
7106	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOWrite);
7107	}
7108	else
7109	{
7110	uint32_t u32Val = 0;
7111	rcStrict = IOMIOPortRead(pVM, pVCpu, IoExitInfo.n.u16Port, &u32Val, cbValue);
7112	if (IOM_SUCCESS(rcStrict))
7113	{
7114	/* Save result of I/O IN instr. in AL/AX/EAX. */
7115	/** @todo r=bird: 32-bit op size should clear high bits of rax! */
7116	pCtx->eax = (pCtx->eax & ~uAndVal) \| (u32Val & uAndVal);
7117	}
7118	else if (rcStrict == VINF_IOM_R3_IOPORT_READ)
7119	HMR0SavePendingIOPortRead(pVCpu, pCtx->rip, pVmcb->ctrl.u64ExitInfo2, IoExitInfo.n.u16Port, uAndVal, cbValue);
7120
7121	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIORead);
7122	}
7123	}
7124
7125	if (IOM_SUCCESS(rcStrict))
7126	{
7127	/* AMD-V saves the RIP of the instruction following the IO instruction in EXITINFO2. */
7128	if (!fUpdateRipAlready)
7129	pCtx->rip = pVmcb->ctrl.u64ExitInfo2;
7130
7131	/*
7132	* If any I/O breakpoints are armed, we need to check if one triggered
7133	* and take appropriate action.
7134	* Note that the I/O breakpoint type is undefined if CR4.DE is 0.
7135	*/
7136	/** @todo Optimize away the DBGFBpIsHwIoArmed call by having DBGF tell the
7137	* execution engines about whether hyper BPs and such are pending. */
7138	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_DR7);
7139	uint32_t const uDr7 = pCtx->dr[7];
7140	if (RT_UNLIKELY( ( (uDr7 & X86_DR7_ENABLED_MASK)
7141	&& X86_DR7_ANY_RW_IO(uDr7)
7142	&& (pCtx->cr4 & X86_CR4_DE))
7143	\|\| DBGFBpIsHwIoArmed(pVM)))
7144	{
7145	/* We're playing with the host CPU state here, make sure we don't preempt or longjmp. */
7146	VMMRZCallRing3Disable(pVCpu);
7147	HM_DISABLE_PREEMPT();
7148
7149	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxIoCheck);
7150	CPUMR0DebugStateMaybeSaveGuest(pVCpu, false /fDr6/);
7151
7152	VBOXSTRICTRC rcStrict2 = DBGFBpCheckIo(pVM, pVCpu, pCtx, IoExitInfo.n.u16Port, cbValue);
7153	if (rcStrict2 == VINF_EM_RAW_GUEST_TRAP)
7154	{
7155	/* Raise #DB. */
7156	pVmcb->guest.u64DR6 = pCtx->dr[6];
7157	pVmcb->guest.u64DR7 = pCtx->dr[7];
7158	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
7159	hmR0SvmSetPendingXcptDB(pVCpu);
7160	}
7161	/* rcStrict is VINF_SUCCESS, VINF_IOM_R3_IOPORT_COMMIT_WRITE, or in [VINF_EM_FIRST..VINF_EM_LAST],
7162	however we can ditch VINF_IOM_R3_IOPORT_COMMIT_WRITE as it has VMCPU_FF_IOM as backup. */
7163	else if ( rcStrict2 != VINF_SUCCESS
7164	&& (rcStrict == VINF_SUCCESS \|\| rcStrict2 < rcStrict))
7165	rcStrict = rcStrict2;
7166	AssertCompile(VINF_EM_LAST < VINF_IOM_R3_IOPORT_COMMIT_WRITE);
7167
7168	HM_RESTORE_PREEMPT();
7169	VMMRZCallRing3Enable(pVCpu);
7170	}
7171
7172	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
7173	}
7174
7175	#ifdef VBOX_STRICT
7176	if (rcStrict == VINF_IOM_R3_IOPORT_READ)
7177	Assert(IoExitInfo.n.u1Type == SVM_IOIO_READ);
7178	else if (rcStrict == VINF_IOM_R3_IOPORT_WRITE \|\| rcStrict == VINF_IOM_R3_IOPORT_COMMIT_WRITE)
7179	Assert(IoExitInfo.n.u1Type == SVM_IOIO_WRITE);
7180	else
7181	{
7182	/** @todo r=bird: This is missing a bunch of VINF_EM_FIRST..VINF_EM_LAST
7183	* statuses, that the VMM device and some others may return. See
7184	* IOM_SUCCESS() for guidance. */
7185	AssertMsg( RT_FAILURE(rcStrict)
7186	\|\| rcStrict == VINF_SUCCESS
7187	\|\| rcStrict == VINF_EM_RAW_EMULATE_INSTR
7188	\|\| rcStrict == VINF_EM_DBG_BREAKPOINT
7189	\|\| rcStrict == VINF_EM_RAW_GUEST_TRAP
7190	\|\| rcStrict == VINF_EM_RAW_TO_R3
7191	\|\| rcStrict == VINF_TRPM_XCPT_DISPATCHED
7192	\|\| rcStrict == VINF_EM_TRIPLE_FAULT, ("%Rrc\n", VBOXSTRICTRC_VAL(rcStrict)));
7193	}
7194	#endif
7195	}
7196	else
7197	{
7198	/*
7199	* Frequent exit or something needing probing. Get state and call EMHistoryExec.
7200	*/
7201	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7202	STAM_COUNTER_INC(!IoExitInfo.n.u1Str
7203	? IoExitInfo.n.u1Type == SVM_IOIO_WRITE ? &pVCpu->hm.s.StatExitIOWrite : &pVCpu->hm.s.StatExitIORead
7204	: IoExitInfo.n.u1Type == SVM_IOIO_WRITE ? &pVCpu->hm.s.StatExitIOStringWrite : &pVCpu->hm.s.StatExitIOStringRead);
7205	Log4(("IOExit/%u: %04x:%08RX64: %s%s%s %#x LB %u -> EMHistoryExec\n",
7206	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip, IoExitInfo.n.u1Rep ? "REP " : "",
7207	IoExitInfo.n.u1Type == SVM_IOIO_WRITE ? "OUT" : "IN", IoExitInfo.n.u1Str ? "S" : "", IoExitInfo.n.u16Port, uIOWidth));
7208
7209	rcStrict = EMHistoryExec(pVCpu, pExitRec, 0);
7210	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
7211
7212	Log4(("IOExit/%u: %04x:%08RX64: EMHistoryExec -> %Rrc + %04x:%08RX64\n",
7213	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip,
7214	VBOXSTRICTRC_VAL(rcStrict), pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip));
7215	}
7216	return VBOXSTRICTRC_TODO(rcStrict);
7217	}
7218
7219
7220	/**
7221	* \#VMEXIT handler for Nested Page-faults (SVM_EXIT_NPF). Conditional \#VMEXIT.
7222	*/
7223	HMSVM_EXIT_DECL hmR0SvmExitNestedPF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7224	{
7225	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7226	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7227	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
7228
7229	PVM pVM = pVCpu->CTX_SUFF(pVM);
7230	Assert(pVM->hm.s.fNestedPaging);
7231
7232	/* See AMD spec. 15.25.6 "Nested versus Guest Page Faults, Fault Ordering" for VMCB details for #NPF. */
7233	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7234	RTGCPHYS GCPhysFaultAddr = pVmcb->ctrl.u64ExitInfo2;
7235	uint32_t u32ErrCode = pVmcb->ctrl.u64ExitInfo1; /* Note! High bits in EXITINFO1 may contain additional info and are
7236	thus intentionally not copied into u32ErrCode. */
7237
7238	Log4Func(("#NPF at CS:RIP=%04x:%#RX64 GCPhysFaultAddr=%RGp ErrCode=%#x \n", pCtx->cs.Sel, pCtx->rip, GCPhysFaultAddr,
7239	u32ErrCode));
7240
7241	/*
7242	* TPR patching for 32-bit guests, using the reserved bit in the page tables for MMIO regions.
7243	*/
7244	if ( pVM->hm.s.fTprPatchingAllowed
7245	&& !CPUMIsGuestInSvmNestedHwVirtMode(pCtx)
7246	&& (GCPhysFaultAddr & PAGE_OFFSET_MASK) == XAPIC_OFF_TPR
7247	&& ( !(u32ErrCode & X86_TRAP_PF_P) /* Not present */
7248	\|\| (u32ErrCode & (X86_TRAP_PF_P \| X86_TRAP_PF_RSVD)) == (X86_TRAP_PF_P \| X86_TRAP_PF_RSVD)) /* MMIO page. */
7249	&& !CPUMIsGuestInLongModeEx(pCtx)
7250	&& !CPUMGetGuestCPL(pVCpu)
7251	&& pVM->hm.s.cPatches < RT_ELEMENTS(pVM->hm.s.aPatches))
7252	{
7253	RTGCPHYS GCPhysApicBase = APICGetBaseMsrNoCheck(pVCpu);
7254	GCPhysApicBase &= PAGE_BASE_GC_MASK;
7255
7256	if (GCPhysFaultAddr == GCPhysApicBase + XAPIC_OFF_TPR)
7257	{
7258	/* Only attempt to patch the instruction once. */
7259	PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
7260	if (!pPatch)
7261	return VINF_EM_HM_PATCH_TPR_INSTR;
7262	}
7263	}
7264
7265	/*
7266	* Determine the nested paging mode.
7267	*/
7268	PGMMODE enmNestedPagingMode;
7269	#if HC_ARCH_BITS == 32
7270	if (CPUMIsGuestInLongModeEx(pCtx))
7271	enmNestedPagingMode = PGMMODE_AMD64_NX;
7272	else
7273	#endif
7274	enmNestedPagingMode = PGMGetHostMode(pVM);
7275
7276	/*
7277	* MMIO optimization using the reserved (RSVD) bit in the guest page tables for MMIO pages.
7278	*/
7279	Assert((u32ErrCode & (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P)) != X86_TRAP_PF_RSVD);
7280	if ((u32ErrCode & (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P)) == (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P))
7281	{
7282	/*
7283	* If event delivery causes an MMIO #NPF, go back to instruction emulation as otherwise
7284	* injecting the original pending event would most likely cause the same MMIO #NPF.
7285	*/
7286	if (pVCpu->hm.s.Event.fPending)
7287	return VINF_EM_RAW_INJECT_TRPM_EVENT;
7288
7289	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_RIP
7290	\| CPUMCTX_EXTRN_CS);
7291	VBOXSTRICTRC rcStrict;
7292	PCEMEXITREC pExitRec = EMHistoryUpdateFlagsAndTypeAndPC(pVCpu,
7293	EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_MMIO),
7294	pVCpu->cpum.GstCtx.rip + pVCpu->cpum.GstCtx.cs.u64Base);
7295	if (!pExitRec)
7296	{
7297
7298	rcStrict = PGMR0Trap0eHandlerNPMisconfig(pVM, pVCpu, enmNestedPagingMode, CPUMCTX2CORE(pCtx), GCPhysFaultAddr,
7299	u32ErrCode);
7300
7301	/*
7302	* If we succeed, resume guest execution.
7303	*
7304	* If we fail in interpreting the instruction because we couldn't get the guest
7305	* physical address of the page containing the instruction via the guest's page
7306	* tables (we would invalidate the guest page in the host TLB), resume execution
7307	* which would cause a guest page fault to let the guest handle this weird case.
7308	*
7309	* See @bugref{6043}.
7310	*/
7311	if ( rcStrict == VINF_SUCCESS
7312	\|\| rcStrict == VERR_PAGE_TABLE_NOT_PRESENT
7313	\|\| rcStrict == VERR_PAGE_NOT_PRESENT)
7314	{
7315	/* Successfully handled MMIO operation. */
7316	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
7317	rcStrict = VINF_SUCCESS;
7318	}
7319	}
7320	else
7321	{
7322	/*
7323	* Frequent exit or something needing probing. Get state and call EMHistoryExec.
7324	*/
7325	Assert(pCtx == &pVCpu->cpum.GstCtx);
7326	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7327	Log4(("EptMisscfgExit/%u: %04x:%08RX64: %RGp -> EMHistoryExec\n",
7328	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip, GCPhysFaultAddr));
7329
7330	rcStrict = EMHistoryExec(pVCpu, pExitRec, 0);
7331	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
7332
7333	Log4(("EptMisscfgExit/%u: %04x:%08RX64: EMHistoryExec -> %Rrc + %04x:%08RX64\n",
7334	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip,
7335	VBOXSTRICTRC_VAL(rcStrict), pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip));
7336	}
7337	return VBOXSTRICTRC_TODO(rcStrict);
7338	}
7339
7340	TRPMAssertXcptPF(pVCpu, GCPhysFaultAddr, u32ErrCode);
7341	int rc = PGMR0Trap0eHandlerNestedPaging(pVM, pVCpu, enmNestedPagingMode, u32ErrCode, CPUMCTX2CORE(pCtx), GCPhysFaultAddr);
7342	TRPMResetTrap(pVCpu);
7343
7344	Log4Func(("#NPF: PGMR0Trap0eHandlerNestedPaging returns %Rrc CS:RIP=%04x:%#RX64\n", rc, pCtx->cs.Sel, pCtx->rip));
7345
7346	/*
7347	* Same case as PGMR0Trap0eHandlerNPMisconfig(). See comment above, @bugref{6043}.
7348	*/
7349	if ( rc == VINF_SUCCESS
7350	\|\| rc == VERR_PAGE_TABLE_NOT_PRESENT
7351	\|\| rc == VERR_PAGE_NOT_PRESENT)
7352	{
7353	/* We've successfully synced our shadow page tables. */
7354	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPF);
7355	rc = VINF_SUCCESS;
7356	}
7357
7358	return rc;
7359	}
7360
7361
7362	/**
7363	* \#VMEXIT handler for virtual interrupt (SVM_EXIT_VINTR). Conditional
7364	* \#VMEXIT.
7365	*/
7366	HMSVM_EXIT_DECL hmR0SvmExitVIntr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7367	{
7368	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7369	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(pCtx);
7370
7371	/* Indicate that we no longer need to #VMEXIT when the guest is ready to receive NMIs, it is now ready. */
7372	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7373	hmR0SvmClearIntWindowExiting(pVCpu, pVmcb, pCtx);
7374
7375	/* Deliver the pending interrupt via hmR0SvmEvaluatePendingEvent() and resume guest execution. */
7376	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIntWindow);
7377	return VINF_SUCCESS;
7378	}
7379
7380
7381	/**
7382	* \#VMEXIT handler for task switches (SVM_EXIT_TASK_SWITCH). Conditional
7383	* \#VMEXIT.
7384	*/
7385	HMSVM_EXIT_DECL hmR0SvmExitTaskSwitch(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7386	{
7387	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7388	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
7389
7390	#ifndef HMSVM_ALWAYS_TRAP_TASK_SWITCH
7391	Assert(!pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
7392	#endif
7393
7394	/* Check if this task-switch occurred while delivering an event through the guest IDT. */
7395	if (pVCpu->hm.s.Event.fPending) /* Can happen with exceptions/NMI. See @bugref{8411}. */
7396	{
7397	/*
7398	* AMD-V provides us with the exception which caused the TS; we collect
7399	* the information in the call to hmR0SvmCheckExitDueToEventDelivery().
7400	*/
7401	Log4Func(("TS occurred during event delivery\n"));
7402	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitTaskSwitch);
7403	return VINF_EM_RAW_INJECT_TRPM_EVENT;
7404	}
7405
7406	/** @todo Emulate task switch someday, currently just going back to ring-3 for
7407	* emulation. */
7408	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitTaskSwitch);
7409	return VERR_EM_INTERPRETER;
7410	}
7411
7412
7413	/**
7414	* \#VMEXIT handler for VMMCALL (SVM_EXIT_VMMCALL). Conditional \#VMEXIT.
7415	*/
7416	HMSVM_EXIT_DECL hmR0SvmExitVmmCall(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7417	{
7418	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7419	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7420
7421	if (pVCpu->CTX_SUFF(pVM)->hm.s.fTprPatchingAllowed)
7422	{
7423	int rc = hmSvmEmulateMovTpr(pVCpu, pCtx);
7424	if (rc != VERR_NOT_FOUND)
7425	{
7426	Log4Func(("hmSvmEmulateMovTpr returns %Rrc\n", rc));
7427	return rc;
7428	}
7429	}
7430
7431	if (EMAreHypercallInstructionsEnabled(pVCpu))
7432	{
7433	VBOXSTRICTRC rcStrict = GIMHypercall(pVCpu, pCtx);
7434	if (RT_SUCCESS(rcStrict))
7435	{
7436	/* Only update the RIP if we're continuing guest execution and not in the case
7437	of say VINF_GIM_R3_HYPERCALL. */
7438	if (rcStrict == VINF_SUCCESS)
7439	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 3 /* cbInstr */);
7440
7441	return VBOXSTRICTRC_VAL(rcStrict);
7442	}
7443	else
7444	Log4Func(("GIMHypercall returns %Rrc -> #UD\n", VBOXSTRICTRC_VAL(rcStrict)));
7445	}
7446
7447	hmR0SvmSetPendingXcptUD(pVCpu);
7448	return VINF_SUCCESS;
7449	}
7450
7451
7452	/**
7453	* \#VMEXIT handler for VMMCALL (SVM_EXIT_VMMCALL). Conditional \#VMEXIT.
7454	*/
7455	HMSVM_EXIT_DECL hmR0SvmExitPause(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7456	{
7457	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7458	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 2);
7459	/** @todo The guest has likely hit a contended spinlock. We might want to
7460	* poke a schedule different guest VCPU. */
7461	return VINF_EM_RAW_INTERRUPT;
7462	}
7463
7464
7465	/**
7466	* \#VMEXIT handler for FERR intercept (SVM_EXIT_FERR_FREEZE). Conditional
7467	* \#VMEXIT.
7468	*/
7469	HMSVM_EXIT_DECL hmR0SvmExitFerrFreeze(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7470	{
7471	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7472	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR0);
7473	Assert(!(pCtx->cr0 & X86_CR0_NE));
7474
7475	Log4Func(("Raising IRQ 13 in response to #FERR\n"));
7476	return PDMIsaSetIrq(pVCpu->CTX_SUFF(pVM), 13 /* u8Irq /, 1 / u8Level /, 0 / uTagSrc */);
7477	}
7478
7479
7480	/**
7481	* \#VMEXIT handler for IRET (SVM_EXIT_IRET). Conditional \#VMEXIT.
7482	*/
7483	HMSVM_EXIT_DECL hmR0SvmExitIret(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7484	{
7485	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7486
7487	/* Clear NMI blocking. */
7488	if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_BLOCK_NMIS))
7489	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_BLOCK_NMIS);
7490
7491	/* Indicate that we no longer need to #VMEXIT when the guest is ready to receive NMIs, it is now ready. */
7492	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7493	hmR0SvmClearCtrlIntercept(pVCpu, pCtx, pVmcb, SVM_CTRL_INTERCEPT_IRET);
7494
7495	/* Deliver the pending NMI via hmR0SvmEvaluatePendingEvent() and resume guest execution. */
7496	return VINF_SUCCESS;
7497	}
7498
7499
7500	/**
7501	* \#VMEXIT handler for page-fault exceptions (SVM_EXIT_XCPT_14).
7502	* Conditional \#VMEXIT.
7503	*/
7504	HMSVM_EXIT_DECL hmR0SvmExitXcptPF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7505	{
7506	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7507	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7508	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
7509
7510	/* See AMD spec. 15.12.15 "#PF (Page Fault)". */
7511	PVM pVM = pVCpu->CTX_SUFF(pVM);
7512	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7513	uint32_t uErrCode = pVmcb->ctrl.u64ExitInfo1;
7514	uint64_t const uFaultAddress = pVmcb->ctrl.u64ExitInfo2;
7515
7516	#if defined(HMSVM_ALWAYS_TRAP_ALL_XCPTS) \|\| defined(HMSVM_ALWAYS_TRAP_PF)
7517	if (pVM->hm.s.fNestedPaging)
7518	{
7519	pVCpu->hm.s.Event.fPending = false; /* In case it's a contributory or vectoring #PF. */
7520	if ( !pSvmTransient->fVectoringDoublePF
7521	\|\| CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
7522	{
7523	/* A genuine guest #PF, reflect it to the guest. */
7524	hmR0SvmSetPendingXcptPF(pVCpu, pCtx, uErrCode, uFaultAddress);
7525	Log4Func(("#PF: Guest page fault at %04X:%RGv FaultAddr=%RX64 ErrCode=%#x\n", pCtx->cs.Sel, (RTGCPTR)pCtx->rip,
7526	uFaultAddress, uErrCode));
7527	}
7528	else
7529	{
7530	/* A guest page-fault occurred during delivery of a page-fault. Inject #DF. */
7531	hmR0SvmSetPendingXcptDF(pVCpu);
7532	Log4Func(("Pending #DF due to vectoring #PF. NP\n"));
7533	}
7534	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestPF);
7535	return VINF_SUCCESS;
7536	}
7537	#endif
7538
7539	Assert(!pVM->hm.s.fNestedPaging);
7540
7541	/*
7542	* TPR patching shortcut for APIC TPR reads and writes; only applicable to 32-bit guests.
7543	*/
7544	if ( pVM->hm.s.fTprPatchingAllowed
7545	&& (uFaultAddress & 0xfff) == XAPIC_OFF_TPR
7546	&& !(uErrCode & X86_TRAP_PF_P) /* Not present. */
7547	&& !CPUMIsGuestInSvmNestedHwVirtMode(pCtx)
7548	&& !CPUMIsGuestInLongModeEx(pCtx)
7549	&& !CPUMGetGuestCPL(pVCpu)
7550	&& pVM->hm.s.cPatches < RT_ELEMENTS(pVM->hm.s.aPatches))
7551	{
7552	RTGCPHYS GCPhysApicBase;
7553	GCPhysApicBase = APICGetBaseMsrNoCheck(pVCpu);
7554	GCPhysApicBase &= PAGE_BASE_GC_MASK;
7555
7556	/* Check if the page at the fault-address is the APIC base. */
7557	RTGCPHYS GCPhysPage;
7558	int rc2 = PGMGstGetPage(pVCpu, (RTGCPTR)uFaultAddress, NULL /* pfFlags */, &GCPhysPage);
7559	if ( rc2 == VINF_SUCCESS
7560	&& GCPhysPage == GCPhysApicBase)
7561	{
7562	/* Only attempt to patch the instruction once. */
7563	PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
7564	if (!pPatch)
7565	return VINF_EM_HM_PATCH_TPR_INSTR;
7566	}
7567	}
7568
7569	Log4Func(("#PF: uFaultAddress=%#RX64 CS:RIP=%#04x:%#RX64 uErrCode %#RX32 cr3=%#RX64\n", uFaultAddress, pCtx->cs.Sel,
7570	pCtx->rip, uErrCode, pCtx->cr3));
7571
7572	/*
7573	* If it's a vectoring #PF, emulate injecting the original event injection as
7574	* PGMTrap0eHandler() is incapable of differentiating between instruction emulation and
7575	* event injection that caused a #PF. See @bugref{6607}.
7576	*/
7577	if (pSvmTransient->fVectoringPF)
7578	{
7579	Assert(pVCpu->hm.s.Event.fPending);
7580	return VINF_EM_RAW_INJECT_TRPM_EVENT;
7581	}
7582
7583	TRPMAssertXcptPF(pVCpu, uFaultAddress, uErrCode);
7584	int rc = PGMTrap0eHandler(pVCpu, uErrCode, CPUMCTX2CORE(pCtx), (RTGCPTR)uFaultAddress);
7585
7586	Log4Func(("#PF: rc=%Rrc\n", rc));
7587
7588	if (rc == VINF_SUCCESS)
7589	{
7590	/* Successfully synced shadow pages tables or emulated an MMIO instruction. */
7591	TRPMResetTrap(pVCpu);
7592	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPF);
7593	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
7594	return rc;
7595	}
7596
7597	if (rc == VINF_EM_RAW_GUEST_TRAP)
7598	{
7599	pVCpu->hm.s.Event.fPending = false; /* In case it's a contributory or vectoring #PF. */
7600
7601	/*
7602	* If a nested-guest delivers a #PF and that causes a #PF which is -not- a shadow #PF,
7603	* we should simply forward the #PF to the guest and is up to the nested-hypervisor to
7604	* determine whether it is a nested-shadow #PF or a #DF, see @bugref{7243#c121}.
7605	*/
7606	if ( !pSvmTransient->fVectoringDoublePF
7607	\|\| CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
7608	{
7609	/* It's a guest (or nested-guest) page fault and needs to be reflected. */
7610	uErrCode = TRPMGetErrorCode(pVCpu); /* The error code might have been changed. */
7611	TRPMResetTrap(pVCpu);
7612
7613	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
7614	/* If the nested-guest is intercepting #PFs, cause a #PF #VMEXIT. */
7615	if ( CPUMIsGuestInSvmNestedHwVirtMode(pCtx)
7616	&& HMIsGuestSvmXcptInterceptSet(pVCpu, X86_XCPT_PF))
7617	return VBOXSTRICTRC_TODO(IEMExecSvmVmexit(pVCpu, SVM_EXIT_XCPT_PF, uErrCode, uFaultAddress));
7618	#endif
7619
7620	hmR0SvmSetPendingXcptPF(pVCpu, pCtx, uErrCode, uFaultAddress);
7621	}
7622	else
7623	{
7624	/* A guest page-fault occurred during delivery of a page-fault. Inject #DF. */
7625	TRPMResetTrap(pVCpu);
7626	hmR0SvmSetPendingXcptDF(pVCpu);
7627	Log4Func(("#PF: Pending #DF due to vectoring #PF\n"));
7628	}
7629
7630	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestPF);
7631	return VINF_SUCCESS;
7632	}
7633
7634	TRPMResetTrap(pVCpu);
7635	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPFEM);
7636	return rc;
7637	}
7638
7639
7640	/**
7641	* \#VMEXIT handler for undefined opcode (SVM_EXIT_XCPT_6).
7642	* Conditional \#VMEXIT.
7643	*/
7644	HMSVM_EXIT_DECL hmR0SvmExitXcptUD(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7645	{
7646	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7647	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(pCtx);
7648
7649	/* Paranoia; Ensure we cannot be called as a result of event delivery. */
7650	PSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
7651	Assert(!pVmcb->ctrl.ExitIntInfo.n.u1Valid); NOREF(pVmcb);
7652
7653	int rc = VERR_SVM_UNEXPECTED_XCPT_EXIT;
7654	if (pVCpu->hm.s.fGIMTrapXcptUD)
7655	{
7656	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7657	uint8_t cbInstr = 0;
7658	VBOXSTRICTRC rcStrict = GIMXcptUD(pVCpu, pCtx, NULL /* pDis */, &cbInstr);
7659	if (rcStrict == VINF_SUCCESS)
7660	{
7661	/* #UD #VMEXIT does not have valid NRIP information, manually advance RIP. See @bugref{7270#c170}. */
7662	hmR0SvmAdvanceRipDumb(pVCpu, pCtx, cbInstr);
7663	rc = VINF_SUCCESS;
7664	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
7665	}
7666	else if (rcStrict == VINF_GIM_HYPERCALL_CONTINUING)
7667	rc = VINF_SUCCESS;
7668	else if (rcStrict == VINF_GIM_R3_HYPERCALL)
7669	rc = VINF_GIM_R3_HYPERCALL;
7670	else
7671	Assert(RT_FAILURE(VBOXSTRICTRC_VAL(rcStrict)));
7672	}
7673
7674	/* If the GIM #UD exception handler didn't succeed for some reason or wasn't needed, raise #UD. */
7675	if (RT_FAILURE(rc))
7676	{
7677	hmR0SvmSetPendingXcptUD(pVCpu);
7678	rc = VINF_SUCCESS;
7679	}
7680
7681	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestUD);
7682	return rc;
7683	}
7684
7685
7686	/**
7687	* \#VMEXIT handler for math-fault exceptions (SVM_EXIT_XCPT_16).
7688	* Conditional \#VMEXIT.
7689	*/
7690	HMSVM_EXIT_DECL hmR0SvmExitXcptMF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7691	{
7692	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7693	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7694
7695	/* Paranoia; Ensure we cannot be called as a result of event delivery. */
7696	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7697	Assert(!pVmcb->ctrl.ExitIntInfo.n.u1Valid); NOREF(pVmcb);
7698
7699	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestMF);
7700
7701	if (!(pCtx->cr0 & X86_CR0_NE))
7702	{
7703	PVM pVM = pVCpu->CTX_SUFF(pVM);
7704	PDISSTATE pDis = &pVCpu->hm.s.DisState;
7705	unsigned cbOp;
7706	int rc = EMInterpretDisasCurrent(pVM, pVCpu, pDis, &cbOp);
7707	if (RT_SUCCESS(rc))
7708	{
7709	/* Convert a #MF into a FERR -> IRQ 13. See @bugref{6117}. */
7710	rc = PDMIsaSetIrq(pVCpu->CTX_SUFF(pVM), 13 /* u8Irq /, 1 / u8Level /, 0 / uTagSrc */);
7711	if (RT_SUCCESS(rc))
7712	pCtx->rip += cbOp;
7713	}
7714	else
7715	Log4Func(("EMInterpretDisasCurrent returned %Rrc uOpCode=%#x\n", rc, pDis->pCurInstr->uOpcode));
7716	return rc;
7717	}
7718
7719	hmR0SvmSetPendingXcptMF(pVCpu);
7720	return VINF_SUCCESS;
7721	}
7722
7723
7724	/**
7725	* \#VMEXIT handler for debug exceptions (SVM_EXIT_XCPT_1). Conditional
7726	* \#VMEXIT.
7727	*/
7728	HMSVM_EXIT_DECL hmR0SvmExitXcptDB(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7729	{
7730	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7731	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7732	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
7733
7734	if (RT_UNLIKELY(pVCpu->hm.s.Event.fPending))
7735	{
7736	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectPendingInterpret);
7737	return VINF_EM_RAW_INJECT_TRPM_EVENT;
7738	}
7739
7740	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestDB);
7741
7742	/*
7743	* This can be a fault-type #DB (instruction breakpoint) or a trap-type #DB (data
7744	* breakpoint). However, for both cases DR6 and DR7 are updated to what the exception
7745	* handler expects. See AMD spec. 15.12.2 "#DB (Debug)".
7746	*/
7747	PVM pVM = pVCpu->CTX_SUFF(pVM);
7748	PSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
7749	int rc = DBGFRZTrap01Handler(pVM, pVCpu, CPUMCTX2CORE(pCtx), pVmcb->guest.u64DR6, pVCpu->hm.s.fSingleInstruction);
7750	if (rc == VINF_EM_RAW_GUEST_TRAP)
7751	{
7752	Log5(("hmR0SvmExitXcptDB: DR6=%#RX64 -> guest trap\n", pVmcb->guest.u64DR6));
7753	if (CPUMIsHyperDebugStateActive(pVCpu))
7754	CPUMSetGuestDR6(pVCpu, CPUMGetGuestDR6(pVCpu) \| pVmcb->guest.u64DR6);
7755
7756	/* Reflect the exception back to the guest. */
7757	hmR0SvmSetPendingXcptDB(pVCpu);
7758	rc = VINF_SUCCESS;
7759	}
7760
7761	/*
7762	* Update DR6.
7763	*/
7764	if (CPUMIsHyperDebugStateActive(pVCpu))
7765	{
7766	Log5(("hmR0SvmExitXcptDB: DR6=%#RX64 -> %Rrc\n", pVmcb->guest.u64DR6, rc));
7767	pVmcb->guest.u64DR6 = X86_DR6_INIT_VAL;
7768	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
7769	}
7770	else
7771	{
7772	AssertMsg(rc == VINF_SUCCESS, ("rc=%Rrc\n", rc));
7773	Assert(!pVCpu->hm.s.fSingleInstruction && !DBGFIsStepping(pVCpu));
7774	}
7775
7776	return rc;
7777	}
7778
7779
7780	/**
7781	* \#VMEXIT handler for alignment check exceptions (SVM_EXIT_XCPT_17).
7782	* Conditional \#VMEXIT.
7783	*/
7784	HMSVM_EXIT_DECL hmR0SvmExitXcptAC(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7785	{
7786	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7787	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
7788
7789	SVMEVENT Event;
7790	Event.u = 0;
7791	Event.n.u1Valid = 1;
7792	Event.n.u3Type = SVM_EVENT_EXCEPTION;
7793	Event.n.u8Vector = X86_XCPT_AC;
7794	Event.n.u1ErrorCodeValid = 1;
7795	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
7796	return VINF_SUCCESS;
7797	}
7798
7799
7800	/**
7801	* \#VMEXIT handler for breakpoint exceptions (SVM_EXIT_XCPT_3).
7802	* Conditional \#VMEXIT.
7803	*/
7804	HMSVM_EXIT_DECL hmR0SvmExitXcptBP(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7805	{
7806	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7807	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7808	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
7809
7810	int rc = DBGFRZTrap03Handler(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
7811	if (rc == VINF_EM_RAW_GUEST_TRAP)
7812	{
7813	SVMEVENT Event;
7814	Event.u = 0;
7815	Event.n.u1Valid = 1;
7816	Event.n.u3Type = SVM_EVENT_EXCEPTION;
7817	Event.n.u8Vector = X86_XCPT_BP;
7818	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
7819	}
7820
7821	Assert(rc == VINF_SUCCESS \|\| rc == VINF_EM_RAW_GUEST_TRAP \|\| rc == VINF_EM_DBG_BREAKPOINT);
7822	return rc;
7823	}
7824
7825
7826	#if defined(HMSVM_ALWAYS_TRAP_ALL_XCPTS) \|\| defined(VBOX_WITH_NESTED_HWVIRT_SVM)
7827	/**
7828	* \#VMEXIT handler for generic exceptions. Conditional \#VMEXIT.
7829	*/
7830	HMSVM_EXIT_DECL hmR0SvmExitXcptGeneric(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7831	{
7832	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7833	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
7834
7835	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7836	uint8_t const uVector = pVmcb->ctrl.u64ExitCode - SVM_EXIT_XCPT_0;
7837	uint32_t const uErrCode = pVmcb->ctrl.u64ExitInfo1;
7838	Assert(pSvmTransient->u64ExitCode == pVmcb->ctrl.u64ExitCode);
7839	Assert(uVector <= X86_XCPT_LAST);
7840	Log4Func(("uVector=%#x uErrCode=%u\n", uVector, uErrCode));
7841
7842	SVMEVENT Event;
7843	Event.u = 0;
7844	Event.n.u1Valid = 1;
7845	Event.n.u3Type = SVM_EVENT_EXCEPTION;
7846	Event.n.u8Vector = uVector;
7847	switch (uVector)
7848	{
7849	/* Shouldn't be here for reflecting #PFs (among other things, the fault address isn't passed along). */
7850	case X86_XCPT_PF: AssertMsgFailed(("hmR0SvmExitXcptGeneric: Unexpected exception")); return VERR_SVM_IPE_5;
7851	case X86_XCPT_DF:
7852	case X86_XCPT_TS:
7853	case X86_XCPT_NP:
7854	case X86_XCPT_SS:
7855	case X86_XCPT_GP:
7856	case X86_XCPT_AC:
7857	{
7858	Event.n.u1ErrorCodeValid = 1;
7859	Event.n.u32ErrorCode = uErrCode;
7860	break;
7861	}
7862	}
7863
7864	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
7865	return VINF_SUCCESS;
7866	}
7867	#endif
7868
7869	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
7870	/**
7871	* \#VMEXIT handler for CLGI (SVM_EXIT_CLGI). Conditional \#VMEXIT.
7872	*/
7873	HMSVM_EXIT_DECL hmR0SvmExitClgi(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7874	{
7875	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7876	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK
7877	\| CPUMCTX_EXTRN_HWVIRT);
7878
7879	#ifdef VBOX_STRICT
7880	PCSVMVMCB pVmcbTmp = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7881	Assert(pVmcbTmp);
7882	Assert(!pVmcbTmp->ctrl.IntCtrl.n.u1VGifEnable);
7883	RT_NOREF(pVmcbTmp);
7884	#endif
7885
7886	uint8_t const cbInstr = hmR0SvmGetInstrLengthHwAssist(pVCpu, pCtx, 3);
7887	VBOXSTRICTRC rcStrict = IEMExecDecodedClgi(pVCpu, cbInstr);
7888	if (rcStrict == VINF_SUCCESS)
7889	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_HWVIRT);
7890	return VBOXSTRICTRC_VAL(rcStrict);
7891	}
7892
7893
7894	/**
7895	* \#VMEXIT handler for STGI (SVM_EXIT_STGI). Conditional \#VMEXIT.
7896	*/
7897	HMSVM_EXIT_DECL hmR0SvmExitStgi(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7898	{
7899	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7900	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK
7901	\| CPUMCTX_EXTRN_HWVIRT);
7902
7903	/*
7904	* When VGIF is not used we always intercept STGI instructions. When VGIF is used,
7905	* we only intercept STGI when events are pending for GIF to become 1.
7906	*/
7907	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7908	if (pVmcb->ctrl.IntCtrl.n.u1VGifEnable)
7909	hmR0SvmClearCtrlIntercept(pVCpu, pCtx, pVmcb, SVM_CTRL_INTERCEPT_STGI);
7910
7911	uint8_t const cbInstr = hmR0SvmGetInstrLengthHwAssist(pVCpu, pCtx, 3);
7912	VBOXSTRICTRC rcStrict = IEMExecDecodedStgi(pVCpu, cbInstr);
7913	if (rcStrict == VINF_SUCCESS)
7914	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_HWVIRT);
7915	return VBOXSTRICTRC_VAL(rcStrict);
7916	}
7917
7918
7919	/**
7920	* \#VMEXIT handler for VMLOAD (SVM_EXIT_VMLOAD). Conditional \#VMEXIT.
7921	*/
7922	HMSVM_EXIT_DECL hmR0SvmExitVmload(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7923	{
7924	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7925	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK
7926	\| CPUMCTX_EXTRN_FS
7927	\| CPUMCTX_EXTRN_GS
7928	\| CPUMCTX_EXTRN_TR
7929	\| CPUMCTX_EXTRN_LDTR
7930	\| CPUMCTX_EXTRN_KERNEL_GS_BASE
7931	\| CPUMCTX_EXTRN_SYSCALL_MSRS
7932	\| CPUMCTX_EXTRN_SYSENTER_MSRS);
7933
7934	#ifdef VBOX_STRICT
7935	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7936	Assert(pVmcb);
7937	Assert(!pVmcb->ctrl.LbrVirt.n.u1VirtVmsaveVmload);
7938	RT_NOREF(pVmcb);
7939	#endif
7940
7941	uint8_t const cbInstr = hmR0SvmGetInstrLengthHwAssist(pVCpu, pCtx, 3);
7942	VBOXSTRICTRC rcStrict = IEMExecDecodedVmload(pVCpu, cbInstr);
7943	if (rcStrict == VINF_SUCCESS)
7944	{
7945	/* We skip flagging changes made to LSTAR, STAR, SFMASK and other MSRs as they are always re-loaded. */
7946	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_FS
7947	\| HM_CHANGED_GUEST_GS
7948	\| HM_CHANGED_GUEST_TR
7949	\| HM_CHANGED_GUEST_LDTR
7950	\| HM_CHANGED_GUEST_SYSENTER_CS_MSR
7951	\| HM_CHANGED_GUEST_SYSENTER_EIP_MSR
7952	\| HM_CHANGED_GUEST_SYSENTER_ESP_MSR);
7953	}
7954	return VBOXSTRICTRC_VAL(rcStrict);
7955	}
7956
7957
7958	/**
7959	* \#VMEXIT handler for VMSAVE (SVM_EXIT_VMSAVE). Conditional \#VMEXIT.
7960	*/
7961	HMSVM_EXIT_DECL hmR0SvmExitVmsave(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7962	{
7963	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7964	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
7965
7966	#ifdef VBOX_STRICT
7967	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7968	Assert(pVmcb);
7969	Assert(!pVmcb->ctrl.LbrVirt.n.u1VirtVmsaveVmload);
7970	RT_NOREF(pVmcb);
7971	#endif
7972
7973	uint8_t const cbInstr = hmR0SvmGetInstrLengthHwAssist(pVCpu, pCtx, 3);
7974	VBOXSTRICTRC rcStrict = IEMExecDecodedVmsave(pVCpu, cbInstr);
7975	return VBOXSTRICTRC_VAL(rcStrict);
7976	}
7977
7978
7979	/**
7980	* \#VMEXIT handler for INVLPGA (SVM_EXIT_INVLPGA). Conditional \#VMEXIT.
7981	*/
7982	HMSVM_EXIT_DECL hmR0SvmExitInvlpga(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7983	{
7984	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7985	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
7986
7987	uint8_t const cbInstr = hmR0SvmGetInstrLengthHwAssist(pVCpu, pCtx, 3);
7988	VBOXSTRICTRC rcStrict = IEMExecDecodedInvlpga(pVCpu, cbInstr);
7989	return VBOXSTRICTRC_VAL(rcStrict);
7990	}
7991
7992
7993	/**
7994	* \#VMEXIT handler for STGI (SVM_EXIT_VMRUN). Conditional \#VMEXIT.
7995	*/
7996	HMSVM_EXIT_DECL hmR0SvmExitVmrun(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7997	{
7998	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7999	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK
8000	\| IEM_CPUMCTX_EXTRN_SVM_VMRUN_MASK);
8001	VBOXSTRICTRC rcStrict;
8002	uint8_t const cbInstr = hmR0SvmGetInstrLengthHwAssist(pVCpu, pCtx, 3);
8003	rcStrict = IEMExecDecodedVmrun(pVCpu, cbInstr);
8004	Log4Func(("IEMExecDecodedVmrun returns %Rrc\n", VBOXSTRICTRC_VAL(rcStrict)));
8005	if (rcStrict == VINF_SUCCESS)
8006	{
8007	rcStrict = VINF_SVM_VMRUN;
8008	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_SVM_VMRUN_MASK);
8009	}
8010	return VBOXSTRICTRC_VAL(rcStrict);
8011	}
8012
8013
8014	/**
8015	* Nested-guest \#VMEXIT handler for debug exceptions (SVM_EXIT_XCPT_1).
8016	* Unconditional \#VMEXIT.
8017	*/
8018	HMSVM_EXIT_DECL hmR0SvmNestedExitXcptDB(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
8019	{
8020	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
8021	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
8022
8023	if (pVCpu->hm.s.Event.fPending)
8024	{
8025	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectPendingInterpret);
8026	return VINF_EM_RAW_INJECT_TRPM_EVENT;
8027	}
8028
8029	hmR0SvmSetPendingXcptDB(pVCpu);
8030	return VINF_SUCCESS;
8031	}
8032
8033
8034	/**
8035	* Nested-guest \#VMEXIT handler for breakpoint exceptions (SVM_EXIT_XCPT_3).
8036	* Conditional \#VMEXIT.
8037	*/
8038	HMSVM_EXIT_DECL hmR0SvmNestedExitXcptBP(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
8039	{
8040	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
8041	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
8042
8043	SVMEVENT Event;
8044	Event.u = 0;
8045	Event.n.u1Valid = 1;
8046	Event.n.u3Type = SVM_EVENT_EXCEPTION;
8047	Event.n.u8Vector = X86_XCPT_BP;
8048	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
8049	return VINF_SUCCESS;
8050	}
8051
8052	#endif /* VBOX_WITH_NESTED_HWVIRT_SVM */
8053
8054
8055	/** @} */
8056

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

source: vbox/trunk/src/VBox/VMM/VMMR0/HMSVMR0.cpp@ 72747

Download in other formats: