HMSVMR0.cpp@ 72881

Last change on this file since 72881 was 72881, checked in by vboxsync, 6 years ago
EM,HM,IEM: Replaced EMInterpretRdmsr with IEMExecDecodedRdmsr.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 316.6 KB

Line
1	/* $Id: HMSVMR0.cpp 72881 2018-07-04 15:19:39Z 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	/** If we decide to use a function table approach this can be useful to
77	* switch to a "static DECLCALLBACK(int)". */
78	#define HMSVM_EXIT_DECL static int
79
80	/**
81	* Subset of the guest-CPU state that is kept by SVM R0 code while executing the
82	* guest using hardware-assisted SVM.
83	*
84	* This excludes state like TSC AUX, GPRs (other than RSP, RAX) which are always
85	* are swapped and restored across the world-switch and also registers like
86	* EFER, PAT MSR etc. which cannot be modified by the guest without causing a
87	* \#VMEXIT.
88	*/
89	#define HMSVM_CPUMCTX_EXTRN_ALL ( CPUMCTX_EXTRN_RIP \
90	\| CPUMCTX_EXTRN_RFLAGS \
91	\| CPUMCTX_EXTRN_RAX \
92	\| CPUMCTX_EXTRN_RSP \
93	\| CPUMCTX_EXTRN_SREG_MASK \
94	\| CPUMCTX_EXTRN_CR0 \
95	\| CPUMCTX_EXTRN_CR2 \
96	\| CPUMCTX_EXTRN_CR3 \
97	\| CPUMCTX_EXTRN_TABLE_MASK \
98	\| CPUMCTX_EXTRN_DR6 \
99	\| CPUMCTX_EXTRN_DR7 \
100	\| CPUMCTX_EXTRN_KERNEL_GS_BASE \
101	\| CPUMCTX_EXTRN_SYSCALL_MSRS \
102	\| CPUMCTX_EXTRN_SYSENTER_MSRS \
103	\| CPUMCTX_EXTRN_HWVIRT \
104	\| CPUMCTX_EXTRN_HM_SVM_MASK)
105
106	/**
107	* Subset of the guest-CPU state that is shared between the guest and host.
108	*/
109	#define HMSVM_CPUMCTX_SHARED_STATE CPUMCTX_EXTRN_DR_MASK
110
111	/** Macro for importing guest state from the VMCB back into CPUMCTX. */
112	#define HMSVM_CPUMCTX_IMPORT_STATE(a_pVCpu, a_fWhat) \
113	do { \
114	if ((a_pVCpu)->cpum.GstCtx.fExtrn & (a_fWhat)) \
115	hmR0SvmImportGuestState((a_pVCpu), (a_fWhat)); \
116	} while (0)
117
118	/** Assert that the required state bits are fetched. */
119	#define HMSVM_CPUMCTX_ASSERT(a_pVCpu, a_fExtrnMbz) AssertMsg(!((a_pVCpu)->cpum.GstCtx.fExtrn & (a_fExtrnMbz)), \
120	("fExtrn=%#RX64 fExtrnMbz=%#RX64\n", \
121	(a_pVCpu)->cpum.GstCtx.fExtrn, (a_fExtrnMbz)))
122
123	/** Macro for checking and returning from the using function for
124	* \#VMEXIT intercepts that maybe caused during delivering of another
125	* event in the guest. */
126	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
127	# define HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY() \
128	do \
129	{ \
130	int rc = hmR0SvmCheckExitDueToEventDelivery(pVCpu, pCtx, pSvmTransient); \
131	if (RT_LIKELY(rc == VINF_SUCCESS)) { /* continue #VMEXIT handling */ } \
132	else if ( rc == VINF_HM_DOUBLE_FAULT) { return VINF_SUCCESS; } \
133	else if ( rc == VINF_EM_RESET \
134	&& CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_SHUTDOWN)) \
135	{ \
136	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_SVM_VMEXIT_MASK); \
137	return VBOXSTRICTRC_TODO(IEMExecSvmVmexit(pVCpu, SVM_EXIT_SHUTDOWN, 0, 0)); \
138	} \
139	else \
140	return rc; \
141	} while (0)
142	#else
143	# define HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY() \
144	do \
145	{ \
146	int rc = hmR0SvmCheckExitDueToEventDelivery(pVCpu, pCtx, pSvmTransient); \
147	if (RT_LIKELY(rc == VINF_SUCCESS)) { /* continue #VMEXIT handling */ } \
148	else if ( rc == VINF_HM_DOUBLE_FAULT) { return VINF_SUCCESS; } \
149	else \
150	return rc; \
151	} while (0)
152	#endif
153
154	/** Macro which updates interrupt shadow for the current RIP. */
155	#define HMSVM_UPDATE_INTR_SHADOW(pVCpu, pCtx) \
156	do { \
157	/* Update interrupt shadow. */ \
158	if ( VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS) \
159	&& pCtx->rip != EMGetInhibitInterruptsPC(pVCpu)) \
160	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS); \
161	} while (0)
162
163	/** Macro for upgrading a @a a_rc to VINF_EM_DBG_STEPPED after emulating an
164	* instruction that exited. */
165	#define HMSVM_CHECK_SINGLE_STEP(a_pVCpu, a_rc) \
166	do { \
167	if ((a_pVCpu)->hm.s.fSingleInstruction && (a_rc) == VINF_SUCCESS) \
168	(a_rc) = VINF_EM_DBG_STEPPED; \
169	} while (0)
170
171	/** Assert that preemption is disabled or covered by thread-context hooks. */
172	#define HMSVM_ASSERT_PREEMPT_SAFE() Assert( VMMR0ThreadCtxHookIsEnabled(pVCpu) \
173	\|\| !RTThreadPreemptIsEnabled(NIL_RTTHREAD));
174
175	/** Assert that we haven't migrated CPUs when thread-context hooks are not
176	* used. */
177	#define HMSVM_ASSERT_CPU_SAFE() AssertMsg( VMMR0ThreadCtxHookIsEnabled(pVCpu) \
178	\|\| pVCpu->hm.s.idEnteredCpu == RTMpCpuId(), \
179	("Illegal migration! Entered on CPU %u Current %u\n", \
180	pVCpu->hm.s.idEnteredCpu, RTMpCpuId()));
181
182	/** Assert that we're not executing a nested-guest. */
183	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
184	# define HMSVM_ASSERT_NOT_IN_NESTED_GUEST(a_pCtx) Assert(!CPUMIsGuestInSvmNestedHwVirtMode((a_pCtx)))
185	#else
186	# define HMSVM_ASSERT_NOT_IN_NESTED_GUEST(a_pCtx) do { NOREF((a_pCtx)); } while (0)
187	#endif
188
189	/** Assert that we're executing a nested-guest. */
190	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
191	# define HMSVM_ASSERT_IN_NESTED_GUEST(a_pCtx) Assert(CPUMIsGuestInSvmNestedHwVirtMode((a_pCtx)))
192	#else
193	# define HMSVM_ASSERT_IN_NESTED_GUEST(a_pCtx) do { NOREF((a_pCtx)); } while (0)
194	#endif
195
196	/** Validate segment descriptor granularity bit. */
197	#ifdef VBOX_STRICT
198	# define HMSVM_ASSERT_SEG_GRANULARITY(a_pCtx, reg) \
199	AssertMsg( !(a_pCtx)->reg.Attr.n.u1Present \
200	\|\| ( (a_pCtx)->reg.Attr.n.u1Granularity \
201	? ((a_pCtx)->reg.u32Limit & 0xfff) == 0xfff \
202	: (a_pCtx)->reg.u32Limit <= UINT32_C(0xfffff)), \
203	("Invalid Segment Attributes Limit=%#RX32 Attr=%#RX32 Base=%#RX64\n", (a_pCtx)->reg.u32Limit, \
204	(a_pCtx)->reg.Attr.u, (a_pCtx)->reg.u64Base))
205	#else
206	# define HMSVM_ASSERT_SEG_GRANULARITY(a_pCtx, reg) do { } while (0)
207	#endif
208
209	/**
210	* Exception bitmap mask for all contributory exceptions.
211	*
212	* Page fault is deliberately excluded here as it's conditional as to whether
213	* it's contributory or benign. Page faults are handled separately.
214	*/
215	#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) \
216	\| RT_BIT(X86_XCPT_DE))
217
218	/**
219	* Mandatory/unconditional guest control intercepts.
220	*
221	* SMIs can and do happen in normal operation. We need not intercept them
222	* while executing the guest (or nested-guest).
223	*/
224	#define HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS ( SVM_CTRL_INTERCEPT_INTR \
225	\| SVM_CTRL_INTERCEPT_NMI \
226	\| SVM_CTRL_INTERCEPT_INIT \
227	\| SVM_CTRL_INTERCEPT_RDPMC \
228	\| SVM_CTRL_INTERCEPT_CPUID \
229	\| SVM_CTRL_INTERCEPT_RSM \
230	\| SVM_CTRL_INTERCEPT_HLT \
231	\| SVM_CTRL_INTERCEPT_IOIO_PROT \
232	\| SVM_CTRL_INTERCEPT_MSR_PROT \
233	\| SVM_CTRL_INTERCEPT_INVLPGA \
234	\| SVM_CTRL_INTERCEPT_SHUTDOWN \
235	\| SVM_CTRL_INTERCEPT_FERR_FREEZE \
236	\| SVM_CTRL_INTERCEPT_VMRUN \
237	\| SVM_CTRL_INTERCEPT_SKINIT \
238	\| SVM_CTRL_INTERCEPT_WBINVD \
239	\| SVM_CTRL_INTERCEPT_MONITOR \
240	\| SVM_CTRL_INTERCEPT_MWAIT \
241	\| SVM_CTRL_INTERCEPT_CR0_SEL_WRITE \
242	\| SVM_CTRL_INTERCEPT_XSETBV)
243
244	/** @name VMCB Clean Bits.
245	*
246	* These flags are used for VMCB-state caching. A set VMCB Clean bit indicates
247	* AMD-V doesn't need to reload the corresponding value(s) from the VMCB in
248	* memory.
249	*
250	* @{ */
251	/** All intercepts vectors, TSC offset, PAUSE filter counter. */
252	#define HMSVM_VMCB_CLEAN_INTERCEPTS RT_BIT(0)
253	/** I/O permission bitmap, MSR permission bitmap. */
254	#define HMSVM_VMCB_CLEAN_IOPM_MSRPM RT_BIT(1)
255	/** ASID. */
256	#define HMSVM_VMCB_CLEAN_ASID RT_BIT(2)
257	/** TRP: V_TPR, V_IRQ, V_INTR_PRIO, V_IGN_TPR, V_INTR_MASKING,
258	V_INTR_VECTOR. */
259	#define HMSVM_VMCB_CLEAN_INT_CTRL RT_BIT(3)
260	/** Nested Paging: Nested CR3 (nCR3), PAT. */
261	#define HMSVM_VMCB_CLEAN_NP RT_BIT(4)
262	/** Control registers (CR0, CR3, CR4, EFER). */
263	#define HMSVM_VMCB_CLEAN_CRX_EFER RT_BIT(5)
264	/** Debug registers (DR6, DR7). */
265	#define HMSVM_VMCB_CLEAN_DRX RT_BIT(6)
266	/** GDT, IDT limit and base. */
267	#define HMSVM_VMCB_CLEAN_DT RT_BIT(7)
268	/** Segment register: CS, SS, DS, ES limit and base. */
269	#define HMSVM_VMCB_CLEAN_SEG RT_BIT(8)
270	/** CR2.*/
271	#define HMSVM_VMCB_CLEAN_CR2 RT_BIT(9)
272	/** Last-branch record (DbgCtlMsr, br_from, br_to, lastint_from, lastint_to) */
273	#define HMSVM_VMCB_CLEAN_LBR RT_BIT(10)
274	/** AVIC (AVIC APIC_BAR; AVIC APIC_BACKING_PAGE, AVIC
275	PHYSICAL_TABLE and AVIC LOGICAL_TABLE Pointers). */
276	#define HMSVM_VMCB_CLEAN_AVIC RT_BIT(11)
277	/** Mask of all valid VMCB Clean bits. */
278	#define HMSVM_VMCB_CLEAN_ALL ( HMSVM_VMCB_CLEAN_INTERCEPTS \
279	\| HMSVM_VMCB_CLEAN_IOPM_MSRPM \
280	\| HMSVM_VMCB_CLEAN_ASID \
281	\| HMSVM_VMCB_CLEAN_INT_CTRL \
282	\| HMSVM_VMCB_CLEAN_NP \
283	\| HMSVM_VMCB_CLEAN_CRX_EFER \
284	\| HMSVM_VMCB_CLEAN_DRX \
285	\| HMSVM_VMCB_CLEAN_DT \
286	\| HMSVM_VMCB_CLEAN_SEG \
287	\| HMSVM_VMCB_CLEAN_CR2 \
288	\| HMSVM_VMCB_CLEAN_LBR \
289	\| HMSVM_VMCB_CLEAN_AVIC)
290	/** @} */
291
292	/** @name SVM transient.
293	*
294	* A state structure for holding miscellaneous information across AMD-V
295	* VMRUN/\#VMEXIT operation, restored after the transition.
296	*
297	* @{ */
298	typedef struct SVMTRANSIENT
299	{
300	/** The host's rflags/eflags. */
301	RTCCUINTREG fEFlags;
302	#if HC_ARCH_BITS == 32
303	uint32_t u32Alignment0;
304	#endif
305
306	/** The \#VMEXIT exit code (the EXITCODE field in the VMCB). */
307	uint64_t u64ExitCode;
308	/** The guest's TPR value used for TPR shadowing. */
309	uint8_t u8GuestTpr;
310	/** Alignment. */
311	uint8_t abAlignment0[7];
312
313	/** Pointer to the currently executing VMCB. */
314	PSVMVMCB pVmcb;
315	/** Whether we are currently executing a nested-guest. */
316	bool fIsNestedGuest;
317
318	/** Whether the guest debug state was active at the time of \#VMEXIT. */
319	bool fWasGuestDebugStateActive;
320	/** Whether the hyper debug state was active at the time of \#VMEXIT. */
321	bool fWasHyperDebugStateActive;
322	/** Whether the TSC offset mode needs to be updated. */
323	bool fUpdateTscOffsetting;
324	/** Whether the TSC_AUX MSR needs restoring on \#VMEXIT. */
325	bool fRestoreTscAuxMsr;
326	/** Whether the \#VMEXIT was caused by a page-fault during delivery of a
327	* contributary exception or a page-fault. */
328	bool fVectoringDoublePF;
329	/** Whether the \#VMEXIT was caused by a page-fault during delivery of an
330	* external interrupt or NMI. */
331	bool fVectoringPF;
332	} SVMTRANSIENT, *PSVMTRANSIENT;
333	AssertCompileMemberAlignment(SVMTRANSIENT, u64ExitCode, sizeof(uint64_t));
334	AssertCompileMemberAlignment(SVMTRANSIENT, pVmcb, sizeof(uint64_t));
335	/** @} */
336
337	/**
338	* MSRPM (MSR permission bitmap) read permissions (for guest RDMSR).
339	*/
340	typedef enum SVMMSREXITREAD
341	{
342	/** Reading this MSR causes a \#VMEXIT. */
343	SVMMSREXIT_INTERCEPT_READ = 0xb,
344	/** Reading this MSR does not cause a \#VMEXIT. */
345	SVMMSREXIT_PASSTHRU_READ
346	} SVMMSREXITREAD;
347
348	/**
349	* MSRPM (MSR permission bitmap) write permissions (for guest WRMSR).
350	*/
351	typedef enum SVMMSREXITWRITE
352	{
353	/** Writing to this MSR causes a \#VMEXIT. */
354	SVMMSREXIT_INTERCEPT_WRITE = 0xd,
355	/** Writing to this MSR does not cause a \#VMEXIT. */
356	SVMMSREXIT_PASSTHRU_WRITE
357	} SVMMSREXITWRITE;
358
359	/**
360	* SVM \#VMEXIT handler.
361	*
362	* @returns VBox status code.
363	* @param pVCpu The cross context virtual CPU structure.
364	* @param pCtx Pointer to the guest-CPU context.
365	* @param pSvmTransient Pointer to the SVM-transient structure.
366	*/
367	typedef int FNSVMEXITHANDLER(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
368
369
370	/*********************************************************************************************************************************
371	* Internal Functions *
372	*********************************************************************************************************************************/
373	static void hmR0SvmPendingEventToTrpmTrap(PVMCPU pVCpu);
374	static void hmR0SvmLeave(PVMCPU pVCpu, bool fImportState);
375
376
377	/** @name \#VMEXIT handlers.
378	* @{
379	*/
380	static FNSVMEXITHANDLER hmR0SvmExitIntr;
381	static FNSVMEXITHANDLER hmR0SvmExitWbinvd;
382	static FNSVMEXITHANDLER hmR0SvmExitInvd;
383	static FNSVMEXITHANDLER hmR0SvmExitCpuid;
384	static FNSVMEXITHANDLER hmR0SvmExitRdtsc;
385	static FNSVMEXITHANDLER hmR0SvmExitRdtscp;
386	static FNSVMEXITHANDLER hmR0SvmExitRdpmc;
387	static FNSVMEXITHANDLER hmR0SvmExitInvlpg;
388	static FNSVMEXITHANDLER hmR0SvmExitHlt;
389	static FNSVMEXITHANDLER hmR0SvmExitMonitor;
390	static FNSVMEXITHANDLER hmR0SvmExitMwait;
391	static FNSVMEXITHANDLER hmR0SvmExitShutdown;
392	static FNSVMEXITHANDLER hmR0SvmExitUnexpected;
393	static FNSVMEXITHANDLER hmR0SvmExitReadCRx;
394	static FNSVMEXITHANDLER hmR0SvmExitWriteCRx;
395	static FNSVMEXITHANDLER hmR0SvmExitMsr;
396	static FNSVMEXITHANDLER hmR0SvmExitReadDRx;
397	static FNSVMEXITHANDLER hmR0SvmExitWriteDRx;
398	static FNSVMEXITHANDLER hmR0SvmExitXsetbv;
399	static FNSVMEXITHANDLER hmR0SvmExitIOInstr;
400	static FNSVMEXITHANDLER hmR0SvmExitNestedPF;
401	static FNSVMEXITHANDLER hmR0SvmExitVIntr;
402	static FNSVMEXITHANDLER hmR0SvmExitTaskSwitch;
403	static FNSVMEXITHANDLER hmR0SvmExitVmmCall;
404	static FNSVMEXITHANDLER hmR0SvmExitPause;
405	static FNSVMEXITHANDLER hmR0SvmExitFerrFreeze;
406	static FNSVMEXITHANDLER hmR0SvmExitIret;
407	static FNSVMEXITHANDLER hmR0SvmExitXcptPF;
408	static FNSVMEXITHANDLER hmR0SvmExitXcptUD;
409	static FNSVMEXITHANDLER hmR0SvmExitXcptMF;
410	static FNSVMEXITHANDLER hmR0SvmExitXcptDB;
411	static FNSVMEXITHANDLER hmR0SvmExitXcptAC;
412	static FNSVMEXITHANDLER hmR0SvmExitXcptBP;
413	#if defined(HMSVM_ALWAYS_TRAP_ALL_XCPTS) \|\| defined(VBOX_WITH_NESTED_HWVIRT_SVM)
414	static FNSVMEXITHANDLER hmR0SvmExitXcptGeneric;
415	#endif
416	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
417	static FNSVMEXITHANDLER hmR0SvmExitClgi;
418	static FNSVMEXITHANDLER hmR0SvmExitStgi;
419	static FNSVMEXITHANDLER hmR0SvmExitVmload;
420	static FNSVMEXITHANDLER hmR0SvmExitVmsave;
421	static FNSVMEXITHANDLER hmR0SvmExitInvlpga;
422	static FNSVMEXITHANDLER hmR0SvmExitVmrun;
423	static FNSVMEXITHANDLER hmR0SvmNestedExitXcptDB;
424	static FNSVMEXITHANDLER hmR0SvmNestedExitXcptBP;
425	#endif
426	/** @} */
427
428	static int hmR0SvmHandleExit(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
429	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
430	static int hmR0SvmHandleExitNested(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
431	#endif
432
433
434	/*********************************************************************************************************************************
435	* Global Variables *
436	*********************************************************************************************************************************/
437	/** Ring-0 memory object for the IO bitmap. */
438	static RTR0MEMOBJ g_hMemObjIOBitmap = NIL_RTR0MEMOBJ;
439	/** Physical address of the IO bitmap. */
440	static RTHCPHYS g_HCPhysIOBitmap;
441	/** Pointer to the IO bitmap. */
442	static R0PTRTYPE(void *) g_pvIOBitmap;
443
444	#ifdef VBOX_STRICT
445	# define HMSVM_LOG_RBP_RSP RT_BIT_32(0)
446	# define HMSVM_LOG_CR_REGS RT_BIT_32(1)
447	# define HMSVM_LOG_CS RT_BIT_32(2)
448	# define HMSVM_LOG_SS RT_BIT_32(3)
449	# define HMSVM_LOG_FS RT_BIT_32(4)
450	# define HMSVM_LOG_GS RT_BIT_32(5)
451	# define HMSVM_LOG_LBR RT_BIT_32(6)
452	# define HMSVM_LOG_ALL ( HMSVM_LOG_RBP_RSP \
453	\| HMSVM_LOG_CR_REGS \
454	\| HMSVM_LOG_CS \
455	\| HMSVM_LOG_SS \
456	\| HMSVM_LOG_FS \
457	\| HMSVM_LOG_GS \
458	\| HMSVM_LOG_LBR)
459
460	/**
461	* Dumps virtual CPU state and additional info. to the logger for diagnostics.
462	*
463	* @param pVCpu The cross context virtual CPU structure.
464	* @param pVmcb Pointer to the VM control block.
465	* @param pszPrefix Log prefix.
466	* @param fFlags Log flags, see HMSVM_LOG_XXX.
467	* @param uVerbose The verbosity level, currently unused.
468	*/
469	static void hmR0SvmLogState(PVMCPU pVCpu, PCSVMVMCB pVmcb, const char *pszPrefix, uint32_t fFlags, uint8_t uVerbose)
470	{
471	RT_NOREF2(pVCpu, uVerbose);
472	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
473
474	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_CS \| CPUMCTX_EXTRN_RIP \| CPUMCTX_EXTRN_RFLAGS);
475	Log4(("%s: cs:rip=%04x:%RX64 efl=%#RX64\n", pszPrefix, pCtx->cs.Sel, pCtx->rip, pCtx->rflags.u));
476
477	if (fFlags & HMSVM_LOG_RBP_RSP)
478	{
479	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_RSP \| CPUMCTX_EXTRN_RBP);
480	Log4(("%s: rsp=%#RX64 rbp=%#RX64\n", pszPrefix, pCtx->rsp, pCtx->rbp));
481	}
482
483	if (fFlags & HMSVM_LOG_CR_REGS)
484	{
485	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_CR0 \| CPUMCTX_EXTRN_CR3 \| CPUMCTX_EXTRN_CR4);
486	Log4(("%s: cr0=%#RX64 cr3=%#RX64 cr4=%#RX64\n", pszPrefix, pCtx->cr0, pCtx->cr3, pCtx->cr4));
487	}
488
489	if (fFlags & HMSVM_LOG_CS)
490	{
491	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_CS);
492	Log4(("%s: cs={%04x base=%016RX64 limit=%08x flags=%08x}\n", pszPrefix, pCtx->cs.Sel, pCtx->cs.u64Base,
493	pCtx->cs.u32Limit, pCtx->cs.Attr.u));
494	}
495	if (fFlags & HMSVM_LOG_SS)
496	{
497	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_SS);
498	Log4(("%s: ss={%04x base=%016RX64 limit=%08x flags=%08x}\n", pszPrefix, pCtx->ss.Sel, pCtx->ss.u64Base,
499	pCtx->ss.u32Limit, pCtx->ss.Attr.u));
500	}
501	if (fFlags & HMSVM_LOG_FS)
502	{
503	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_FS);
504	Log4(("%s: fs={%04x base=%016RX64 limit=%08x flags=%08x}\n", pszPrefix, pCtx->fs.Sel, pCtx->fs.u64Base,
505	pCtx->fs.u32Limit, pCtx->fs.Attr.u));
506	}
507	if (fFlags & HMSVM_LOG_GS)
508	{
509	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_GS);
510	Log4(("%s: gs={%04x base=%016RX64 limit=%08x flags=%08x}\n", pszPrefix, pCtx->gs.Sel, pCtx->gs.u64Base,
511	pCtx->gs.u32Limit, pCtx->gs.Attr.u));
512	}
513
514	PCSVMVMCBSTATESAVE pVmcbGuest = &pVmcb->guest;
515	if (fFlags & HMSVM_LOG_LBR)
516	{
517	Log4(("%s: br_from=%#RX64 br_to=%#RX64 lastxcpt_from=%#RX64 lastxcpt_to=%#RX64\n", pszPrefix, pVmcbGuest->u64BR_FROM,
518	pVmcbGuest->u64BR_TO, pVmcbGuest->u64LASTEXCPFROM, pVmcbGuest->u64LASTEXCPTO));
519	}
520	NOREF(pVmcbGuest); NOREF(pCtx);
521	}
522	#endif /* VBOX_STRICT */
523
524
525	/**
526	* Sets up and activates AMD-V on the current CPU.
527	*
528	* @returns VBox status code.
529	* @param pHostCpu Pointer to the CPU info struct.
530	* @param pVM The cross context VM structure. Can be
531	* NULL after a resume!
532	* @param pvCpuPage Pointer to the global CPU page.
533	* @param HCPhysCpuPage Physical address of the global CPU page.
534	* @param fEnabledByHost Whether the host OS has already initialized AMD-V.
535	* @param pvArg Unused on AMD-V.
536	*/
537	VMMR0DECL(int) SVMR0EnableCpu(PHMGLOBALCPUINFO pHostCpu, PVM pVM, void *pvCpuPage, RTHCPHYS HCPhysCpuPage, bool fEnabledByHost,
538	void *pvArg)
539	{
540	Assert(!fEnabledByHost);
541	Assert(HCPhysCpuPage && HCPhysCpuPage != NIL_RTHCPHYS);
542	Assert(RT_ALIGN_T(HCPhysCpuPage, _4K, RTHCPHYS) == HCPhysCpuPage);
543	Assert(pvCpuPage); NOREF(pvCpuPage);
544	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
545
546	NOREF(pvArg);
547	NOREF(fEnabledByHost);
548
549	/* Paranoid: Disable interrupt as, in theory, interrupt handlers might mess with EFER. */
550	RTCCUINTREG const fEFlags = ASMIntDisableFlags();
551
552	/*
553	* We must turn on AMD-V and setup the host state physical address, as those MSRs are per CPU.
554	*/
555	uint64_t u64HostEfer = ASMRdMsr(MSR_K6_EFER);
556	if (u64HostEfer & MSR_K6_EFER_SVME)
557	{
558	/* If the VBOX_HWVIRTEX_IGNORE_SVM_IN_USE is active, then we blindly use AMD-V. */
559	if ( pVM
560	&& pVM->hm.s.svm.fIgnoreInUseError)
561	pHostCpu->fIgnoreAMDVInUseError = true;
562
563	if (!pHostCpu->fIgnoreAMDVInUseError)
564	{
565	ASMSetFlags(fEFlags);
566	return VERR_SVM_IN_USE;
567	}
568	}
569
570	/* Turn on AMD-V in the EFER MSR. */
571	ASMWrMsr(MSR_K6_EFER, u64HostEfer \| MSR_K6_EFER_SVME);
572
573	/* Write the physical page address where the CPU will store the host state while executing the VM. */
574	ASMWrMsr(MSR_K8_VM_HSAVE_PA, HCPhysCpuPage);
575
576	/* Restore interrupts. */
577	ASMSetFlags(fEFlags);
578
579	/*
580	* Theoretically, other hypervisors may have used ASIDs, ideally we should flush all
581	* non-zero ASIDs when enabling SVM. AMD doesn't have an SVM instruction to flush all
582	* ASIDs (flushing is done upon VMRUN). Therefore, flag that we need to flush the TLB
583	* entirely with before executing any guest code.
584	*/
585	pHostCpu->fFlushAsidBeforeUse = true;
586
587	/*
588	* Ensure each VCPU scheduled on this CPU gets a new ASID on resume. See @bugref{6255}.
589	*/
590	++pHostCpu->cTlbFlushes;
591
592	return VINF_SUCCESS;
593	}
594
595
596	/**
597	* Deactivates AMD-V on the current CPU.
598	*
599	* @returns VBox status code.
600	* @param pHostCpu Pointer to the CPU info struct.
601	* @param pvCpuPage Pointer to the global CPU page.
602	* @param HCPhysCpuPage Physical address of the global CPU page.
603	*/
604	VMMR0DECL(int) SVMR0DisableCpu(PHMGLOBALCPUINFO pHostCpu, void *pvCpuPage, RTHCPHYS HCPhysCpuPage)
605	{
606	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
607	AssertReturn( HCPhysCpuPage
608	&& HCPhysCpuPage != NIL_RTHCPHYS, VERR_INVALID_PARAMETER);
609	AssertReturn(pvCpuPage, VERR_INVALID_PARAMETER);
610	RT_NOREF(pHostCpu);
611
612	/* Paranoid: Disable interrupts as, in theory, interrupt handlers might mess with EFER. */
613	RTCCUINTREG const fEFlags = ASMIntDisableFlags();
614
615	/* Turn off AMD-V in the EFER MSR. */
616	uint64_t u64HostEfer = ASMRdMsr(MSR_K6_EFER);
617	ASMWrMsr(MSR_K6_EFER, u64HostEfer & ~MSR_K6_EFER_SVME);
618
619	/* Invalidate host state physical address. */
620	ASMWrMsr(MSR_K8_VM_HSAVE_PA, 0);
621
622	/* Restore interrupts. */
623	ASMSetFlags(fEFlags);
624
625	return VINF_SUCCESS;
626	}
627
628
629	/**
630	* Does global AMD-V initialization (called during module initialization).
631	*
632	* @returns VBox status code.
633	*/
634	VMMR0DECL(int) SVMR0GlobalInit(void)
635	{
636	/*
637	* Allocate 12 KB (3 pages) for the IO bitmap. Since this is non-optional and we always
638	* intercept all IO accesses, it's done once globally here instead of per-VM.
639	*/
640	Assert(g_hMemObjIOBitmap == NIL_RTR0MEMOBJ);
641	int rc = RTR0MemObjAllocCont(&g_hMemObjIOBitmap, SVM_IOPM_PAGES << X86_PAGE_4K_SHIFT, false /* fExecutable */);
642	if (RT_FAILURE(rc))
643	return rc;
644
645	g_pvIOBitmap = RTR0MemObjAddress(g_hMemObjIOBitmap);
646	g_HCPhysIOBitmap = RTR0MemObjGetPagePhysAddr(g_hMemObjIOBitmap, 0 /* iPage */);
647
648	/* Set all bits to intercept all IO accesses. */
649	ASMMemFill32(g_pvIOBitmap, SVM_IOPM_PAGES << X86_PAGE_4K_SHIFT, UINT32_C(0xffffffff));
650
651	return VINF_SUCCESS;
652	}
653
654
655	/**
656	* Does global AMD-V termination (called during module termination).
657	*/
658	VMMR0DECL(void) SVMR0GlobalTerm(void)
659	{
660	if (g_hMemObjIOBitmap != NIL_RTR0MEMOBJ)
661	{
662	RTR0MemObjFree(g_hMemObjIOBitmap, true /* fFreeMappings */);
663	g_pvIOBitmap = NULL;
664	g_HCPhysIOBitmap = 0;
665	g_hMemObjIOBitmap = NIL_RTR0MEMOBJ;
666	}
667	}
668
669
670	/**
671	* Frees any allocated per-VCPU structures for a VM.
672	*
673	* @param pVM The cross context VM structure.
674	*/
675	DECLINLINE(void) hmR0SvmFreeStructs(PVM pVM)
676	{
677	for (uint32_t i = 0; i < pVM->cCpus; i++)
678	{
679	PVMCPU pVCpu = &pVM->aCpus[i];
680	AssertPtr(pVCpu);
681
682	if (pVCpu->hm.s.svm.hMemObjVmcbHost != NIL_RTR0MEMOBJ)
683	{
684	RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjVmcbHost, false);
685	pVCpu->hm.s.svm.HCPhysVmcbHost = 0;
686	pVCpu->hm.s.svm.hMemObjVmcbHost = NIL_RTR0MEMOBJ;
687	}
688
689	if (pVCpu->hm.s.svm.hMemObjVmcb != NIL_RTR0MEMOBJ)
690	{
691	RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjVmcb, false);
692	pVCpu->hm.s.svm.pVmcb = NULL;
693	pVCpu->hm.s.svm.HCPhysVmcb = 0;
694	pVCpu->hm.s.svm.hMemObjVmcb = NIL_RTR0MEMOBJ;
695	}
696
697	if (pVCpu->hm.s.svm.hMemObjMsrBitmap != NIL_RTR0MEMOBJ)
698	{
699	RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjMsrBitmap, false);
700	pVCpu->hm.s.svm.pvMsrBitmap = NULL;
701	pVCpu->hm.s.svm.HCPhysMsrBitmap = 0;
702	pVCpu->hm.s.svm.hMemObjMsrBitmap = NIL_RTR0MEMOBJ;
703	}
704	}
705	}
706
707
708	/**
709	* Does per-VM AMD-V initialization.
710	*
711	* @returns VBox status code.
712	* @param pVM The cross context VM structure.
713	*/
714	VMMR0DECL(int) SVMR0InitVM(PVM pVM)
715	{
716	int rc = VERR_INTERNAL_ERROR_5;
717
718	/*
719	* Check for an AMD CPU erratum which requires us to flush the TLB before every world-switch.
720	*/
721	uint32_t u32Family;
722	uint32_t u32Model;
723	uint32_t u32Stepping;
724	if (HMAmdIsSubjectToErratum170(&u32Family, &u32Model, &u32Stepping))
725	{
726	Log4Func(("AMD cpu with erratum 170 family %#x model %#x stepping %#x\n", u32Family, u32Model, u32Stepping));
727	pVM->hm.s.svm.fAlwaysFlushTLB = true;
728	}
729
730	/*
731	* Initialize the R0 memory objects up-front so we can properly cleanup on allocation failures.
732	*/
733	for (VMCPUID i = 0; i < pVM->cCpus; i++)
734	{
735	PVMCPU pVCpu = &pVM->aCpus[i];
736	pVCpu->hm.s.svm.hMemObjVmcbHost = NIL_RTR0MEMOBJ;
737	pVCpu->hm.s.svm.hMemObjVmcb = NIL_RTR0MEMOBJ;
738	pVCpu->hm.s.svm.hMemObjMsrBitmap = NIL_RTR0MEMOBJ;
739	}
740
741	for (VMCPUID i = 0; i < pVM->cCpus; i++)
742	{
743	PVMCPU pVCpu = &pVM->aCpus[i];
744
745	/*
746	* Allocate one page for the host-context VM control block (VMCB). This is used for additional host-state (such as
747	* FS, GS, Kernel GS Base, etc.) apart from the host-state save area specified in MSR_K8_VM_HSAVE_PA.
748	*/
749	rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjVmcbHost, SVM_VMCB_PAGES << PAGE_SHIFT, false /* fExecutable */);
750	if (RT_FAILURE(rc))
751	goto failure_cleanup;
752
753	void *pvVmcbHost = RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjVmcbHost);
754	pVCpu->hm.s.svm.HCPhysVmcbHost = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjVmcbHost, 0 /* iPage */);
755	Assert(pVCpu->hm.s.svm.HCPhysVmcbHost < _4G);
756	ASMMemZeroPage(pvVmcbHost);
757
758	/*
759	* Allocate one page for the guest-state VMCB.
760	*/
761	rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjVmcb, SVM_VMCB_PAGES << PAGE_SHIFT, false /* fExecutable */);
762	if (RT_FAILURE(rc))
763	goto failure_cleanup;
764
765	pVCpu->hm.s.svm.pVmcb = (PSVMVMCB)RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjVmcb);
766	pVCpu->hm.s.svm.HCPhysVmcb = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjVmcb, 0 /* iPage */);
767	Assert(pVCpu->hm.s.svm.HCPhysVmcb < _4G);
768	ASMMemZeroPage(pVCpu->hm.s.svm.pVmcb);
769
770	/*
771	* Allocate two pages (8 KB) for the MSR permission bitmap. There doesn't seem to be a way to convince
772	* SVM to not require one.
773	*/
774	rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjMsrBitmap, SVM_MSRPM_PAGES << X86_PAGE_4K_SHIFT,
775	false /* fExecutable */);
776	if (RT_FAILURE(rc))
777	goto failure_cleanup;
778
779	pVCpu->hm.s.svm.pvMsrBitmap = RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjMsrBitmap);
780	pVCpu->hm.s.svm.HCPhysMsrBitmap = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjMsrBitmap, 0 /* iPage */);
781	/* Set all bits to intercept all MSR accesses (changed later on). */
782	ASMMemFill32(pVCpu->hm.s.svm.pvMsrBitmap, SVM_MSRPM_PAGES << X86_PAGE_4K_SHIFT, UINT32_C(0xffffffff));
783	}
784
785	return VINF_SUCCESS;
786
787	failure_cleanup:
788	hmR0SvmFreeStructs(pVM);
789	return rc;
790	}
791
792
793	/**
794	* Does per-VM AMD-V termination.
795	*
796	* @returns VBox status code.
797	* @param pVM The cross context VM structure.
798	*/
799	VMMR0DECL(int) SVMR0TermVM(PVM pVM)
800	{
801	hmR0SvmFreeStructs(pVM);
802	return VINF_SUCCESS;
803	}
804
805
806	/**
807	* Returns whether the VMCB Clean Bits feature is supported.
808	*
809	* @return @c true if supported, @c false otherwise.
810	* @param pVCpu The cross context virtual CPU structure.
811	* @param pCtx Pointer to the guest-CPU context.
812	*/
813	DECLINLINE(bool) hmR0SvmSupportsVmcbCleanBits(PVMCPU pVCpu, PCCPUMCTX pCtx)
814	{
815	PVM pVM = pVCpu->CTX_SUFF(pVM);
816	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
817	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
818	{
819	return (pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_VMCB_CLEAN)
820	&& pVM->cpum.ro.GuestFeatures.fSvmVmcbClean;
821	}
822	#else
823	RT_NOREF(pCtx);
824	#endif
825	return RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_VMCB_CLEAN);
826	}
827
828
829	/**
830	* Returns whether the decode assists feature is supported.
831	*
832	* @return @c true if supported, @c false otherwise.
833	* @param pVCpu The cross context virtual CPU structure.
834	* @param pCtx Pointer to the guest-CPU context.
835	*/
836	DECLINLINE(bool) hmR0SvmSupportsDecodeAssists(PVMCPU pVCpu, PCPUMCTX pCtx)
837	{
838	PVM pVM = pVCpu->CTX_SUFF(pVM);
839	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
840	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
841	{
842	return (pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_DECODE_ASSISTS)
843	&& pVM->cpum.ro.GuestFeatures.fSvmDecodeAssists;
844	}
845	#else
846	RT_NOREF(pCtx);
847	#endif
848	return RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_DECODE_ASSISTS);
849	}
850
851
852	/**
853	* Returns whether the NRIP_SAVE feature is supported.
854	*
855	* @return @c true if supported, @c false otherwise.
856	* @param pVCpu The cross context virtual CPU structure.
857	* @param pCtx Pointer to the guest-CPU context.
858	*/
859	DECLINLINE(bool) hmR0SvmSupportsNextRipSave(PVMCPU pVCpu, PCPUMCTX pCtx)
860	{
861	PVM pVM = pVCpu->CTX_SUFF(pVM);
862	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
863	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
864	{
865	return (pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_NRIP_SAVE)
866	&& pVM->cpum.ro.GuestFeatures.fSvmNextRipSave;
867	}
868	#else
869	RT_NOREF(pCtx);
870	#endif
871	return RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_NRIP_SAVE);
872	}
873
874
875	/**
876	* Sets the permission bits for the specified MSR in the MSRPM bitmap.
877	*
878	* @param pCtx Pointer to the guest-CPU or nested-guest-CPU context.
879	* @param pbMsrBitmap Pointer to the MSR bitmap.
880	* @param idMsr The MSR for which the permissions are being set.
881	* @param enmRead MSR read permissions.
882	* @param enmWrite MSR write permissions.
883	*
884	* @remarks This function does -not- clear the VMCB clean bits for MSRPM. The
885	* caller needs to take care of this.
886	*/
887	static void hmR0SvmSetMsrPermission(PCCPUMCTX pCtx, uint8_t *pbMsrBitmap, uint32_t idMsr, SVMMSREXITREAD enmRead,
888	SVMMSREXITWRITE enmWrite)
889	{
890	bool const fInNestedGuestMode = CPUMIsGuestInSvmNestedHwVirtMode(pCtx);
891	uint16_t offMsrpm;
892	uint8_t uMsrpmBit;
893	int rc = HMSvmGetMsrpmOffsetAndBit(idMsr, &offMsrpm, &uMsrpmBit);
894	AssertRC(rc);
895
896	Assert(uMsrpmBit == 0 \|\| uMsrpmBit == 2 \|\| uMsrpmBit == 4 \|\| uMsrpmBit == 6);
897	Assert(offMsrpm < SVM_MSRPM_PAGES << X86_PAGE_4K_SHIFT);
898
899	pbMsrBitmap += offMsrpm;
900	if (enmRead == SVMMSREXIT_INTERCEPT_READ)
901	*pbMsrBitmap \|= RT_BIT(uMsrpmBit);
902	else
903	{
904	if (!fInNestedGuestMode)
905	*pbMsrBitmap &= ~RT_BIT(uMsrpmBit);
906	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
907	else
908	{
909	/* Only clear the bit if the nested-guest is also not intercepting the MSR read.*/
910	uint8_t const pbNstGstMsrBitmap = (uint8_t )pCtx->hwvirt.svm.CTX_SUFF(pvMsrBitmap);
911	pbNstGstMsrBitmap += offMsrpm;
912	if (!(*pbNstGstMsrBitmap & RT_BIT(uMsrpmBit)))
913	*pbMsrBitmap &= ~RT_BIT(uMsrpmBit);
914	else
915	Assert(*pbMsrBitmap & RT_BIT(uMsrpmBit));
916	}
917	#endif
918	}
919
920	if (enmWrite == SVMMSREXIT_INTERCEPT_WRITE)
921	*pbMsrBitmap \|= RT_BIT(uMsrpmBit + 1);
922	else
923	{
924	if (!fInNestedGuestMode)
925	*pbMsrBitmap &= ~RT_BIT(uMsrpmBit + 1);
926	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
927	else
928	{
929	/* Only clear the bit if the nested-guest is also not intercepting the MSR write.*/
930	uint8_t const pbNstGstMsrBitmap = (uint8_t )pCtx->hwvirt.svm.CTX_SUFF(pvMsrBitmap);
931	pbNstGstMsrBitmap += offMsrpm;
932	if (!(*pbNstGstMsrBitmap & RT_BIT(uMsrpmBit + 1)))
933	*pbMsrBitmap &= ~RT_BIT(uMsrpmBit + 1);
934	else
935	Assert(*pbMsrBitmap & RT_BIT(uMsrpmBit + 1));
936	}
937	#endif
938	}
939	}
940
941
942	/**
943	* Sets up AMD-V for the specified VM.
944	* This function is only called once per-VM during initalization.
945	*
946	* @returns VBox status code.
947	* @param pVM The cross context VM structure.
948	*/
949	VMMR0DECL(int) SVMR0SetupVM(PVM pVM)
950	{
951	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
952	AssertReturn(pVM, VERR_INVALID_PARAMETER);
953	Assert(pVM->hm.s.svm.fSupported);
954
955	bool const fPauseFilter = RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_PAUSE_FILTER);
956	bool const fPauseFilterThreshold = RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_PAUSE_FILTER_THRESHOLD);
957	bool const fUsePauseFilter = fPauseFilter && pVM->hm.s.svm.cPauseFilter;
958
959	bool const fLbrVirt = RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_LBR_VIRT);
960	bool const fUseLbrVirt = fLbrVirt; /** @todo CFGM, IEM implementation etc. */
961
962	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
963	bool const fVirtVmsaveVmload = RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_VIRT_VMSAVE_VMLOAD);
964	bool const fUseVirtVmsaveVmload = fVirtVmsaveVmload && pVM->hm.s.svm.fVirtVmsaveVmload && pVM->hm.s.fNestedPaging;
965
966	bool const fVGif = RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_VGIF);
967	bool const fUseVGif = fVGif && pVM->hm.s.svm.fVGif;
968	#endif
969
970	PVMCPU pVCpu = &pVM->aCpus[0];
971	PSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
972	AssertMsgReturn(pVmcb, ("Invalid pVmcb for vcpu[0]\n"), VERR_SVM_INVALID_PVMCB);
973	PSVMVMCBCTRL pVmcbCtrl = &pVmcb->ctrl;
974
975	/* Always trap #AC for reasons of security. */
976	pVmcbCtrl->u32InterceptXcpt \|= RT_BIT_32(X86_XCPT_AC);
977
978	/* Always trap #DB for reasons of security. */
979	pVmcbCtrl->u32InterceptXcpt \|= RT_BIT_32(X86_XCPT_DB);
980
981	/* Trap exceptions unconditionally (debug purposes). */
982	#ifdef HMSVM_ALWAYS_TRAP_PF
983	pVmcbCtrl->u32InterceptXcpt \|= RT_BIT(X86_XCPT_PF);
984	#endif
985	#ifdef HMSVM_ALWAYS_TRAP_ALL_XCPTS
986	/* If you add any exceptions here, make sure to update hmR0SvmHandleExit(). */
987	pVmcbCtrl->u32InterceptXcpt \|= 0
988	\| RT_BIT(X86_XCPT_BP)
989	\| RT_BIT(X86_XCPT_DE)
990	\| RT_BIT(X86_XCPT_NM)
991	\| RT_BIT(X86_XCPT_UD)
992	\| RT_BIT(X86_XCPT_NP)
993	\| RT_BIT(X86_XCPT_SS)
994	\| RT_BIT(X86_XCPT_GP)
995	\| RT_BIT(X86_XCPT_PF)
996	\| RT_BIT(X86_XCPT_MF)
997	;
998	#endif
999
1000	/* Apply the exceptions intercepts needed by the GIM provider. */
1001	if (pVCpu->hm.s.fGIMTrapXcptUD)
1002	pVmcbCtrl->u32InterceptXcpt \|= RT_BIT(X86_XCPT_UD);
1003
1004	/* Set up unconditional intercepts and conditions. */
1005	pVmcbCtrl->u64InterceptCtrl = HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS
1006	\| SVM_CTRL_INTERCEPT_VMMCALL;
1007
1008	#ifdef HMSVM_ALWAYS_TRAP_TASK_SWITCH
1009	pVmcbCtrl->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_TASK_SWITCH;
1010	#endif
1011
1012	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1013	/* Virtualized VMSAVE/VMLOAD. */
1014	pVmcbCtrl->LbrVirt.n.u1VirtVmsaveVmload = fUseVirtVmsaveVmload;
1015	if (!fUseVirtVmsaveVmload)
1016	{
1017	pVmcbCtrl->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_VMSAVE
1018	\| SVM_CTRL_INTERCEPT_VMLOAD;
1019	}
1020
1021	/* Virtual GIF. */
1022	pVmcbCtrl->IntCtrl.n.u1VGifEnable = fUseVGif;
1023	if (!fUseVGif)
1024	{
1025	pVmcbCtrl->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_CLGI
1026	\| SVM_CTRL_INTERCEPT_STGI;
1027	}
1028	#endif
1029
1030	/* CR4 writes must always be intercepted for tracking PGM mode changes. */
1031	pVmcbCtrl->u16InterceptWrCRx = RT_BIT(4);
1032
1033	/* Intercept all DRx reads and writes by default. Changed later on. */
1034	pVmcbCtrl->u16InterceptRdDRx = 0xffff;
1035	pVmcbCtrl->u16InterceptWrDRx = 0xffff;
1036
1037	/* Virtualize masking of INTR interrupts. (reads/writes from/to CR8 go to the V_TPR register) */
1038	pVmcbCtrl->IntCtrl.n.u1VIntrMasking = 1;
1039
1040	/* Ignore the priority in the virtual TPR. This is necessary for delivering PIC style (ExtInt) interrupts
1041	and we currently deliver both PIC and APIC interrupts alike, see hmR0SvmEvaluatePendingEvent() */
1042	pVmcbCtrl->IntCtrl.n.u1IgnoreTPR = 1;
1043
1044	/* Set the IO permission bitmap physical addresses. */
1045	pVmcbCtrl->u64IOPMPhysAddr = g_HCPhysIOBitmap;
1046
1047	/* LBR virtualization. */
1048	pVmcbCtrl->LbrVirt.n.u1LbrVirt = fUseLbrVirt;
1049
1050	/* The host ASID MBZ, for the guest start with 1. */
1051	pVmcbCtrl->TLBCtrl.n.u32ASID = 1;
1052
1053	/* Setup Nested Paging. This doesn't change throughout the execution time of the VM. */
1054	pVmcbCtrl->NestedPagingCtrl.n.u1NestedPaging = pVM->hm.s.fNestedPaging;
1055
1056	/* Without Nested Paging, we need additionally intercepts. */
1057	if (!pVM->hm.s.fNestedPaging)
1058	{
1059	/* CR3 reads/writes must be intercepted; our shadow values differ from the guest values. */
1060	pVmcbCtrl->u16InterceptRdCRx \|= RT_BIT(3);
1061	pVmcbCtrl->u16InterceptWrCRx \|= RT_BIT(3);
1062
1063	/* Intercept INVLPG and task switches (may change CR3, EFLAGS, LDT). */
1064	pVmcbCtrl->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_INVLPG
1065	\| SVM_CTRL_INTERCEPT_TASK_SWITCH;
1066
1067	/* Page faults must be intercepted to implement shadow paging. */
1068	pVmcbCtrl->u32InterceptXcpt \|= RT_BIT(X86_XCPT_PF);
1069	}
1070
1071	/* Setup Pause Filter for guest pause-loop (spinlock) exiting. */
1072	if (fUsePauseFilter)
1073	{
1074	Assert(pVM->hm.s.svm.cPauseFilter > 0);
1075	pVmcbCtrl->u16PauseFilterCount = pVM->hm.s.svm.cPauseFilter;
1076	if (fPauseFilterThreshold)
1077	pVmcbCtrl->u16PauseFilterThreshold = pVM->hm.s.svm.cPauseFilterThresholdTicks;
1078	pVmcbCtrl->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_PAUSE;
1079	}
1080
1081	/*
1082	* Setup the MSR permission bitmap.
1083	* The following MSRs are saved/restored automatically during the world-switch.
1084	* Don't intercept guest read/write accesses to these MSRs.
1085	*/
1086	uint8_t pbMsrBitmap = (uint8_t )pVCpu->hm.s.svm.pvMsrBitmap;
1087	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1088	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1089	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_K8_CSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1090	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_K6_STAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1091	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_K8_SF_MASK, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1092	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_K8_FS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1093	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_K8_GS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1094	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_K8_KERNEL_GS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1095	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_IA32_SYSENTER_CS, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1096	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_IA32_SYSENTER_ESP, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1097	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_IA32_SYSENTER_EIP, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1098	pVmcbCtrl->u64MSRPMPhysAddr = pVCpu->hm.s.svm.HCPhysMsrBitmap;
1099
1100	/* Initially all VMCB clean bits MBZ indicating that everything should be loaded from the VMCB in memory. */
1101	Assert(pVmcbCtrl->u32VmcbCleanBits == 0);
1102
1103	for (VMCPUID i = 1; i < pVM->cCpus; i++)
1104	{
1105	PVMCPU pVCpuCur = &pVM->aCpus[i];
1106	PSVMVMCB pVmcbCur = pVM->aCpus[i].hm.s.svm.pVmcb;
1107	AssertMsgReturn(pVmcbCur, ("Invalid pVmcb for vcpu[%u]\n", i), VERR_SVM_INVALID_PVMCB);
1108	PSVMVMCBCTRL pVmcbCtrlCur = &pVmcbCur->ctrl;
1109
1110	/* Copy the VMCB control area. */
1111	memcpy(pVmcbCtrlCur, pVmcbCtrl, sizeof(*pVmcbCtrlCur));
1112
1113	/* Copy the MSR bitmap and setup the VCPU-specific host physical address. */
1114	uint8_t pbMsrBitmapCur = (uint8_t )pVCpuCur->hm.s.svm.pvMsrBitmap;
1115	memcpy(pbMsrBitmapCur, pbMsrBitmap, SVM_MSRPM_PAGES << X86_PAGE_4K_SHIFT);
1116	pVmcbCtrlCur->u64MSRPMPhysAddr = pVCpuCur->hm.s.svm.HCPhysMsrBitmap;
1117
1118	/* Initially all VMCB clean bits MBZ indicating that everything should be loaded from the VMCB in memory. */
1119	Assert(pVmcbCtrlCur->u32VmcbCleanBits == 0);
1120
1121	/* Verify our assumption that GIM providers trap #UD uniformly across VCPUs initially. */
1122	Assert(pVCpuCur->hm.s.fGIMTrapXcptUD == pVCpu->hm.s.fGIMTrapXcptUD);
1123	}
1124
1125	return VINF_SUCCESS;
1126	}
1127
1128
1129	/**
1130	* Gets a pointer to the currently active guest (or nested-guest) VMCB.
1131	*
1132	* @returns Pointer to the current context VMCB.
1133	* @param pVCpu The cross context virtual CPU structure.
1134	* @param pCtx Pointer to the guest-CPU context.
1135	*/
1136	DECLINLINE(PSVMVMCB) hmR0SvmGetCurrentVmcb(PVMCPU pVCpu, PCPUMCTX pCtx)
1137	{
1138	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1139	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
1140	return pCtx->hwvirt.svm.CTX_SUFF(pVmcb);
1141	#else
1142	RT_NOREF(pCtx);
1143	#endif
1144	return pVCpu->hm.s.svm.pVmcb;
1145	}
1146
1147
1148	/**
1149	* Gets a pointer to the nested-guest VMCB cache.
1150	*
1151	* @returns Pointer to the nested-guest VMCB cache.
1152	* @param pVCpu The cross context virtual CPU structure.
1153	*/
1154	DECLINLINE(PSVMNESTEDVMCBCACHE) hmR0SvmGetNestedVmcbCache(PVMCPU pVCpu)
1155	{
1156	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1157	Assert(pVCpu->hm.s.svm.NstGstVmcbCache.fCacheValid);
1158	return &pVCpu->hm.s.svm.NstGstVmcbCache;
1159	#else
1160	RT_NOREF(pVCpu);
1161	return NULL;
1162	#endif
1163	}
1164
1165
1166	/**
1167	* Invalidates a guest page by guest virtual address.
1168	*
1169	* @returns VBox status code.
1170	* @param pVCpu The cross context virtual CPU structure.
1171	* @param GCVirt Guest virtual address of the page to invalidate.
1172	*/
1173	VMMR0DECL(int) SVMR0InvalidatePage(PVMCPU pVCpu, RTGCPTR GCVirt)
1174	{
1175	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.svm.fSupported);
1176
1177	bool const fFlushPending = pVCpu->CTX_SUFF(pVM)->hm.s.svm.fAlwaysFlushTLB \|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_TLB_FLUSH);
1178
1179	/* Skip it if a TLB flush is already pending. */
1180	if (!fFlushPending)
1181	{
1182	Log4Func(("%#RGv\n", GCVirt));
1183
1184	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1185	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
1186	AssertMsgReturn(pVmcb, ("Invalid pVmcb!\n"), VERR_SVM_INVALID_PVMCB);
1187
1188	#if HC_ARCH_BITS == 32
1189	/* If we get a flush in 64-bit guest mode, then force a full TLB flush. INVLPGA takes only 32-bit addresses. */
1190	if (CPUMIsGuestInLongMode(pVCpu))
1191	VMCPU_FF_SET(pVCpu, VMCPU_FF_TLB_FLUSH);
1192	else
1193	#endif
1194	{
1195	SVMR0InvlpgA(GCVirt, pVmcb->ctrl.TLBCtrl.n.u32ASID);
1196	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlbInvlpgVirt);
1197	}
1198	}
1199	return VINF_SUCCESS;
1200	}
1201
1202
1203	/**
1204	* Flushes the appropriate tagged-TLB entries.
1205	*
1206	* @param pVCpu The cross context virtual CPU structure.
1207	* @param pCtx Pointer to the guest-CPU or nested-guest-CPU context.
1208	* @param pVmcb Pointer to the VM control block.
1209	* @param pHostCpu Pointer to the HM host-CPU info.
1210	*/
1211	static void hmR0SvmFlushTaggedTlb(PVMCPU pVCpu, PCCPUMCTX pCtx, PSVMVMCB pVmcb, PHMGLOBALCPUINFO pHostCpu)
1212	{
1213	#ifndef VBOX_WITH_NESTED_HWVIRT_SVM
1214	RT_NOREF(pCtx);
1215	#endif
1216	PVM pVM = pVCpu->CTX_SUFF(pVM);
1217
1218	/*
1219	* Force a TLB flush for the first world switch if the current CPU differs from the one
1220	* we ran on last. This can happen both for start & resume due to long jumps back to
1221	* ring-3.
1222	*
1223	* We also force a TLB flush every time when executing a nested-guest VCPU as there is no
1224	* correlation between it and the physical CPU.
1225	*
1226	* If the TLB flush count changed, another VM (VCPU rather) has hit the ASID limit while
1227	* flushing the TLB, so we cannot reuse the ASIDs without flushing.
1228	*/
1229	bool fNewAsid = false;
1230	Assert(pHostCpu->idCpu != NIL_RTCPUID);
1231	if ( pVCpu->hm.s.idLastCpu != pHostCpu->idCpu
1232	\|\| pVCpu->hm.s.cTlbFlushes != pHostCpu->cTlbFlushes
1233	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1234	\|\| CPUMIsGuestInSvmNestedHwVirtMode(pCtx)
1235	#endif
1236	)
1237	{
1238	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlbWorldSwitch);
1239	pVCpu->hm.s.fForceTLBFlush = true;
1240	fNewAsid = true;
1241	}
1242
1243	/* Set TLB flush state as checked until we return from the world switch. */
1244	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, true);
1245
1246	/* Check for explicit TLB flushes. */
1247	if (VMCPU_FF_TEST_AND_CLEAR(pVCpu, VMCPU_FF_TLB_FLUSH))
1248	{
1249	pVCpu->hm.s.fForceTLBFlush = true;
1250	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlb);
1251	}
1252
1253	/*
1254	* If the AMD CPU erratum 170, We need to flush the entire TLB for each world switch. Sad.
1255	* This Host CPU requirement takes precedence.
1256	*/
1257	if (pVM->hm.s.svm.fAlwaysFlushTLB)
1258	{
1259	pHostCpu->uCurrentAsid = 1;
1260	pVCpu->hm.s.uCurrentAsid = 1;
1261	pVCpu->hm.s.cTlbFlushes = pHostCpu->cTlbFlushes;
1262	pVCpu->hm.s.idLastCpu = pHostCpu->idCpu;
1263	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
1264
1265	/* Clear the VMCB Clean Bit for NP while flushing the TLB. See @bugref{7152}. */
1266	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
1267	}
1268	else
1269	{
1270	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_NOTHING;
1271	if (pVCpu->hm.s.fForceTLBFlush)
1272	{
1273	/* Clear the VMCB Clean Bit for NP while flushing the TLB. See @bugref{7152}. */
1274	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
1275
1276	if (fNewAsid)
1277	{
1278	++pHostCpu->uCurrentAsid;
1279
1280	bool fHitASIDLimit = false;
1281	if (pHostCpu->uCurrentAsid >= pVM->hm.s.uMaxAsid)
1282	{
1283	pHostCpu->uCurrentAsid = 1; /* Wraparound at 1; host uses 0 */
1284	pHostCpu->cTlbFlushes++; /* All VCPUs that run on this host CPU must use a new ASID. */
1285	fHitASIDLimit = true;
1286	}
1287
1288	if ( fHitASIDLimit
1289	\|\| pHostCpu->fFlushAsidBeforeUse)
1290	{
1291	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
1292	pHostCpu->fFlushAsidBeforeUse = false;
1293	}
1294
1295	pVCpu->hm.s.uCurrentAsid = pHostCpu->uCurrentAsid;
1296	pVCpu->hm.s.idLastCpu = pHostCpu->idCpu;
1297	pVCpu->hm.s.cTlbFlushes = pHostCpu->cTlbFlushes;
1298	}
1299	else
1300	{
1301	if (pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_FLUSH_BY_ASID)
1302	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_SINGLE_CONTEXT;
1303	else
1304	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
1305	}
1306
1307	pVCpu->hm.s.fForceTLBFlush = false;
1308	}
1309	}
1310
1311	/* Update VMCB with the ASID. */
1312	if (pVmcb->ctrl.TLBCtrl.n.u32ASID != pVCpu->hm.s.uCurrentAsid)
1313	{
1314	pVmcb->ctrl.TLBCtrl.n.u32ASID = pVCpu->hm.s.uCurrentAsid;
1315	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_ASID;
1316	}
1317
1318	AssertMsg(pVCpu->hm.s.idLastCpu == pHostCpu->idCpu,
1319	("vcpu idLastCpu=%u hostcpu idCpu=%u\n", pVCpu->hm.s.idLastCpu, pHostCpu->idCpu));
1320	AssertMsg(pVCpu->hm.s.cTlbFlushes == pHostCpu->cTlbFlushes,
1321	("Flush count mismatch for cpu %u (%u vs %u)\n", pHostCpu->idCpu, pVCpu->hm.s.cTlbFlushes, pHostCpu->cTlbFlushes));
1322	AssertMsg(pHostCpu->uCurrentAsid >= 1 && pHostCpu->uCurrentAsid < pVM->hm.s.uMaxAsid,
1323	("cpu%d uCurrentAsid = %x\n", pHostCpu->idCpu, pHostCpu->uCurrentAsid));
1324	AssertMsg(pVCpu->hm.s.uCurrentAsid >= 1 && pVCpu->hm.s.uCurrentAsid < pVM->hm.s.uMaxAsid,
1325	("cpu%d VM uCurrentAsid = %x\n", pHostCpu->idCpu, pVCpu->hm.s.uCurrentAsid));
1326
1327	#ifdef VBOX_WITH_STATISTICS
1328	if (pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_NOTHING)
1329	STAM_COUNTER_INC(&pVCpu->hm.s.StatNoFlushTlbWorldSwitch);
1330	else if ( pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_SINGLE_CONTEXT
1331	\|\| pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_SINGLE_CONTEXT_RETAIN_GLOBALS)
1332	{
1333	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushAsid);
1334	}
1335	else
1336	{
1337	Assert(pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_ENTIRE);
1338	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushEntire);
1339	}
1340	#endif
1341	}
1342
1343
1344	/** @name 64-bit guest on 32-bit host OS helper functions.
1345	*
1346	* The host CPU is still 64-bit capable but the host OS is running in 32-bit
1347	* mode (code segment, paging). These wrappers/helpers perform the necessary
1348	* bits for the 32->64 switcher.
1349	*
1350	* @{ */
1351	#if HC_ARCH_BITS == 32 && defined(VBOX_ENABLE_64_BITS_GUESTS)
1352	/**
1353	* Prepares for and executes VMRUN (64-bit guests on a 32-bit host).
1354	*
1355	* @returns VBox status code.
1356	* @param HCPhysVmcbHost Physical address of host VMCB.
1357	* @param HCPhysVmcb Physical address of the VMCB.
1358	* @param pCtx Pointer to the guest-CPU context.
1359	* @param pVM The cross context VM structure.
1360	* @param pVCpu The cross context virtual CPU structure.
1361	*/
1362	DECLASM(int) SVMR0VMSwitcherRun64(RTHCPHYS HCPhysVmcbHost, RTHCPHYS HCPhysVmcb, PCPUMCTX pCtx, PVM pVM, PVMCPU pVCpu)
1363	{
1364	RT_NOREF2(pVM, pCtx);
1365	uint32_t aParam[8];
1366	aParam[0] = RT_LO_U32(HCPhysVmcbHost); /* Param 1: HCPhysVmcbHost - Lo. */
1367	aParam[1] = RT_HI_U32(HCPhysVmcbHost); /* Param 1: HCPhysVmcbHost - Hi. */
1368	aParam[2] = RT_LO_U32(HCPhysVmcb); /* Param 2: HCPhysVmcb - Lo. */
1369	aParam[3] = RT_HI_U32(HCPhysVmcb); /* Param 2: HCPhysVmcb - Hi. */
1370	aParam[4] = VM_RC_ADDR(pVM, pVM);
1371	aParam[5] = 0;
1372	aParam[6] = VM_RC_ADDR(pVM, pVCpu);
1373	aParam[7] = 0;
1374
1375	return SVMR0Execute64BitsHandler(pVCpu, HM64ON32OP_SVMRCVMRun64, RT_ELEMENTS(aParam), &aParam[0]);
1376	}
1377
1378
1379	/**
1380	* Executes the specified VMRUN handler in 64-bit mode.
1381	*
1382	* @returns VBox status code.
1383	* @param pVCpu The cross context virtual CPU structure.
1384	* @param enmOp The operation to perform.
1385	* @param cParams Number of parameters.
1386	* @param paParam Array of 32-bit parameters.
1387	*/
1388	VMMR0DECL(int) SVMR0Execute64BitsHandler(PVMCPU pVCpu, HM64ON32OP enmOp, uint32_t cParams, uint32_t *paParam)
1389	{
1390	PVM pVM = pVCpu->CTX_SUFF(pVM);
1391	AssertReturn(pVM->hm.s.pfnHost32ToGuest64R0, VERR_HM_NO_32_TO_64_SWITCHER);
1392	Assert(enmOp > HM64ON32OP_INVALID && enmOp < HM64ON32OP_END);
1393
1394	/* Disable interrupts. */
1395	RTHCUINTREG const fEFlags = ASMIntDisableFlags();
1396
1397	#ifdef VBOX_WITH_VMMR0_DISABLE_LAPIC_NMI
1398	RTCPUID idHostCpu = RTMpCpuId();
1399	CPUMR0SetLApic(pVCpu, idHostCpu);
1400	#endif
1401
1402	CPUMSetHyperESP(pVCpu, VMMGetStackRC(pVCpu));
1403	CPUMSetHyperEIP(pVCpu, enmOp);
1404	for (int i = (int)cParams - 1; i >= 0; i--)
1405	CPUMPushHyper(pVCpu, paParam[i]);
1406
1407	/* Call the switcher. */
1408	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatWorldSwitch3264, z);
1409	int rc = pVM->hm.s.pfnHost32ToGuest64R0(pVM, RT_OFFSETOF(VM, aCpus[pVCpu->idCpu].cpum) - RT_OFFSETOF(VM, cpum));
1410	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatWorldSwitch3264, z);
1411
1412	/* Restore interrupts. */
1413	ASMSetFlags(fEFlags);
1414	return rc;
1415	}
1416
1417	#endif /* HC_ARCH_BITS == 32 && defined(VBOX_ENABLE_64_BITS_GUESTS) */
1418	/** @} */
1419
1420
1421	/**
1422	* Sets an exception intercept in the specified VMCB.
1423	*
1424	* @param pVmcb Pointer to the VM control block.
1425	* @param uXcpt The exception (X86_XCPT_*).
1426	*/
1427	DECLINLINE(void) hmR0SvmSetXcptIntercept(PSVMVMCB pVmcb, uint8_t uXcpt)
1428	{
1429	if (!(pVmcb->ctrl.u32InterceptXcpt & RT_BIT(uXcpt)))
1430	{
1431	pVmcb->ctrl.u32InterceptXcpt \|= RT_BIT(uXcpt);
1432	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1433	}
1434	}
1435
1436
1437	/**
1438	* Clears an exception intercept in the specified VMCB.
1439	*
1440	* @param pVCpu The cross context virtual CPU structure.
1441	* @param pVmcb Pointer to the VM control block.
1442	* @param uXcpt The exception (X86_XCPT_*).
1443	*
1444	* @remarks This takes into account if we're executing a nested-guest and only
1445	* removes the exception intercept if both the guest -and- nested-guest
1446	* are not intercepting it.
1447	*/
1448	DECLINLINE(void) hmR0SvmClearXcptIntercept(PVMCPU pVCpu, PSVMVMCB pVmcb, uint8_t uXcpt)
1449	{
1450	Assert(uXcpt != X86_XCPT_DB);
1451	Assert(uXcpt != X86_XCPT_AC);
1452	#ifndef HMSVM_ALWAYS_TRAP_ALL_XCPTS
1453	if (pVmcb->ctrl.u32InterceptXcpt & RT_BIT(uXcpt))
1454	{
1455	bool fRemove = true;
1456	# ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1457	/* Only remove the intercept if the nested-guest is also not intercepting it! */
1458	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1459	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
1460	{
1461	PCSVMNESTEDVMCBCACHE pVmcbNstGstCache = hmR0SvmGetNestedVmcbCache(pVCpu);
1462	fRemove = !(pVmcbNstGstCache->u32InterceptXcpt & RT_BIT(uXcpt));
1463	}
1464	# else
1465	RT_NOREF(pVCpu);
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));
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));
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	*
1847	* @remarks No-long-jump zone!!!
1848	*/
1849	static void hmR0SvmExportGuestMsrs(PVMCPU pVCpu, PSVMVMCB pVmcb)
1850	{
1851	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1852	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1853
1854	/* Guest Sysenter MSRs. */
1855	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SYSENTER_MSR_MASK)
1856	{
1857	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SYSENTER_CS_MSR)
1858	pVmcb->guest.u64SysEnterCS = pCtx->SysEnter.cs;
1859
1860	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SYSENTER_EIP_MSR)
1861	pVmcb->guest.u64SysEnterEIP = pCtx->SysEnter.eip;
1862
1863	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SYSENTER_ESP_MSR)
1864	pVmcb->guest.u64SysEnterESP = pCtx->SysEnter.esp;
1865	}
1866
1867	/*
1868	* Guest EFER MSR.
1869	* AMD-V requires guest EFER.SVME to be set. Weird.
1870	* See AMD spec. 15.5.1 "Basic Operation" \| "Canonicalization and Consistency Checks".
1871	*/
1872	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_EFER_MSR)
1873	{
1874	pVmcb->guest.u64EFER = pCtx->msrEFER \| MSR_K6_EFER_SVME;
1875	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1876	}
1877
1878	/* If the guest isn't in 64-bit mode, clear MSR_K6_LME bit, otherwise SVM expects amd64 shadow paging. */
1879	if ( !CPUMIsGuestInLongModeEx(pCtx)
1880	&& (pCtx->msrEFER & MSR_K6_EFER_LME))
1881	{
1882	pVmcb->guest.u64EFER &= ~MSR_K6_EFER_LME;
1883	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1884	}
1885
1886	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SYSCALL_MSRS)
1887	{
1888	pVmcb->guest.u64STAR = pCtx->msrSTAR;
1889	pVmcb->guest.u64LSTAR = pCtx->msrLSTAR;
1890	pVmcb->guest.u64CSTAR = pCtx->msrCSTAR;
1891	pVmcb->guest.u64SFMASK = pCtx->msrSFMASK;
1892	}
1893
1894	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_KERNEL_GS_BASE)
1895	pVmcb->guest.u64KernelGSBase = pCtx->msrKERNELGSBASE;
1896
1897	pVCpu->hm.s.fCtxChanged &= ~( HM_CHANGED_GUEST_SYSENTER_MSR_MASK
1898	\| HM_CHANGED_GUEST_EFER_MSR
1899	\| HM_CHANGED_GUEST_SYSCALL_MSRS
1900	\| HM_CHANGED_GUEST_KERNEL_GS_BASE);
1901
1902	/*
1903	* Setup the PAT MSR (applicable for Nested Paging only).
1904	*
1905	* While guests can modify and see the modified values through the shadow values,
1906	* we shall not honor any guest modifications of this MSR to ensure caching is always
1907	* enabled similar to how we clear CR0.CD and NW bits.
1908	*
1909	* For nested-guests this needs to always be set as well, see @bugref{7243#c109}.
1910	*/
1911	pVmcb->guest.u64PAT = MSR_IA32_CR_PAT_INIT_VAL;
1912
1913	/* Enable the last branch record bit if LBR virtualization is enabled. */
1914	if (pVmcb->ctrl.LbrVirt.n.u1LbrVirt)
1915	pVmcb->guest.u64DBGCTL = MSR_IA32_DEBUGCTL_LBR;
1916	}
1917
1918
1919	/**
1920	* Exports the guest (or nested-guest) debug state into the VMCB and programs
1921	* the necessary intercepts accordingly.
1922	*
1923	* @param pVCpu The cross context virtual CPU structure.
1924	* @param pVmcb Pointer to the VM control block.
1925	*
1926	* @remarks No-long-jump zone!!!
1927	* @remarks Requires EFLAGS to be up-to-date in the VMCB!
1928	*/
1929	static void hmR0SvmExportSharedDebugState(PVMCPU pVCpu, PSVMVMCB pVmcb)
1930	{
1931	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1932
1933	/*
1934	* Anyone single stepping on the host side? If so, we'll have to use the
1935	* trap flag in the guest EFLAGS since AMD-V doesn't have a trap flag on
1936	* the VMM level like the VT-x implementations does.
1937	*/
1938	bool fInterceptMovDRx = false;
1939	bool const fStepping = pVCpu->hm.s.fSingleInstruction \|\| DBGFIsStepping(pVCpu);
1940	if (fStepping)
1941	{
1942	pVCpu->hm.s.fClearTrapFlag = true;
1943	pVmcb->guest.u64RFlags \|= X86_EFL_TF;
1944	fInterceptMovDRx = true; /* Need clean DR6, no guest mess. */
1945	}
1946
1947	if ( fStepping
1948	\|\| (CPUMGetHyperDR7(pVCpu) & X86_DR7_ENABLED_MASK))
1949	{
1950	/*
1951	* Use the combined guest and host DRx values found in the hypervisor
1952	* register set because the debugger has breakpoints active or someone
1953	* is single stepping on the host side.
1954	*
1955	* Note! DBGF expects a clean DR6 state before executing guest code.
1956	*/
1957	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS)
1958	if ( CPUMIsGuestInLongModeEx(pCtx)
1959	&& !CPUMIsHyperDebugStateActivePending(pVCpu))
1960	{
1961	CPUMR0LoadHyperDebugState(pVCpu, false /* include DR6 */);
1962	Assert(!CPUMIsGuestDebugStateActivePending(pVCpu));
1963	Assert(CPUMIsHyperDebugStateActivePending(pVCpu));
1964	}
1965	else
1966	#endif
1967	if (!CPUMIsHyperDebugStateActive(pVCpu))
1968	{
1969	CPUMR0LoadHyperDebugState(pVCpu, false /* include DR6 */);
1970	Assert(!CPUMIsGuestDebugStateActive(pVCpu));
1971	Assert(CPUMIsHyperDebugStateActive(pVCpu));
1972	}
1973
1974	/* Update DR6 & DR7. (The other DRx values are handled by CPUM one way or the other.) */
1975	if ( pVmcb->guest.u64DR6 != X86_DR6_INIT_VAL
1976	\|\| pVmcb->guest.u64DR7 != CPUMGetHyperDR7(pVCpu))
1977	{
1978	pVmcb->guest.u64DR7 = CPUMGetHyperDR7(pVCpu);
1979	pVmcb->guest.u64DR6 = X86_DR6_INIT_VAL;
1980	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
1981	}
1982
1983	/** @todo If we cared, we could optimize to allow the guest to read registers
1984	* with the same values. */
1985	fInterceptMovDRx = true;
1986	pVCpu->hm.s.fUsingHyperDR7 = true;
1987	Log5(("hmR0SvmExportSharedDebugState: Loaded hyper DRx\n"));
1988	}
1989	else
1990	{
1991	/*
1992	* Update DR6, DR7 with the guest values if necessary.
1993	*/
1994	if ( pVmcb->guest.u64DR7 != pCtx->dr[7]
1995	\|\| pVmcb->guest.u64DR6 != pCtx->dr[6])
1996	{
1997	pVmcb->guest.u64DR7 = pCtx->dr[7];
1998	pVmcb->guest.u64DR6 = pCtx->dr[6];
1999	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
2000	}
2001	pVCpu->hm.s.fUsingHyperDR7 = false;
2002
2003	/*
2004	* If the guest has enabled debug registers, we need to load them prior to
2005	* executing guest code so they'll trigger at the right time.
2006	*/
2007	if (pCtx->dr[7] & (X86_DR7_ENABLED_MASK \| X86_DR7_GD)) /** @todo Why GD? */
2008	{
2009	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS)
2010	if ( CPUMIsGuestInLongModeEx(pCtx)
2011	&& !CPUMIsGuestDebugStateActivePending(pVCpu))
2012	{
2013	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
2014	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxArmed);
2015	Assert(!CPUMIsHyperDebugStateActivePending(pVCpu));
2016	Assert(CPUMIsGuestDebugStateActivePending(pVCpu));
2017	}
2018	else
2019	#endif
2020	if (!CPUMIsGuestDebugStateActive(pVCpu))
2021	{
2022	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
2023	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxArmed);
2024	Assert(!CPUMIsHyperDebugStateActive(pVCpu));
2025	Assert(CPUMIsGuestDebugStateActive(pVCpu));
2026	}
2027	Log5(("hmR0SvmExportSharedDebugState: Loaded guest DRx\n"));
2028	}
2029	/*
2030	* If no debugging enabled, we'll lazy load DR0-3. We don't need to
2031	* intercept #DB as DR6 is updated in the VMCB.
2032	*
2033	* Note! If we cared and dared, we could skip intercepting \#DB here.
2034	* However, \#DB shouldn't be performance critical, so we'll play safe
2035	* and keep the code similar to the VT-x code and always intercept it.
2036	*/
2037	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS)
2038	else if ( !CPUMIsGuestDebugStateActivePending(pVCpu)
2039	&& !CPUMIsGuestDebugStateActive(pVCpu))
2040	#else
2041	else if (!CPUMIsGuestDebugStateActive(pVCpu))
2042	#endif
2043	{
2044	fInterceptMovDRx = true;
2045	}
2046	}
2047
2048	Assert(pVmcb->ctrl.u32InterceptXcpt & RT_BIT_32(X86_XCPT_DB));
2049	if (fInterceptMovDRx)
2050	{
2051	if ( pVmcb->ctrl.u16InterceptRdDRx != 0xffff
2052	\|\| pVmcb->ctrl.u16InterceptWrDRx != 0xffff)
2053	{
2054	pVmcb->ctrl.u16InterceptRdDRx = 0xffff;
2055	pVmcb->ctrl.u16InterceptWrDRx = 0xffff;
2056	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2057	}
2058	}
2059	else
2060	{
2061	if ( pVmcb->ctrl.u16InterceptRdDRx
2062	\|\| pVmcb->ctrl.u16InterceptWrDRx)
2063	{
2064	pVmcb->ctrl.u16InterceptRdDRx = 0;
2065	pVmcb->ctrl.u16InterceptWrDRx = 0;
2066	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2067	}
2068	}
2069	Log4Func(("DR6=%#RX64 DR7=%#RX64\n", pCtx->dr[6], pCtx->dr[7]));
2070	}
2071
2072	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
2073	/**
2074	* Exports the nested-guest hardware virtualization state into the nested-guest
2075	* VMCB.
2076	*
2077	* @param pVCpu The cross context virtual CPU structure.
2078	* @param pVmcbNstGst Pointer to the nested-guest VM control block.
2079	*
2080	* @remarks No-long-jump zone!!!
2081	*/
2082	static void hmR0SvmExportGuestHwvirtStateNested(PVMCPU pVCpu, PSVMVMCB pVmcbNstGst)
2083	{
2084	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2085
2086	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_HWVIRT)
2087	{
2088	/*
2089	* Ensure the nested-guest pause-filter counters don't exceed the outer guest values esp.
2090	* since SVM doesn't have a preemption timer.
2091	*
2092	* We do this here rather than in hmR0SvmSetupVmcbNested() as we may have been executing the
2093	* nested-guest in IEM incl. PAUSE instructions which would update the pause-filter counters
2094	* and may continue execution in SVM R0 without a nested-guest #VMEXIT in between.
2095	*/
2096	PVM pVM = pVCpu->CTX_SUFF(pVM);
2097	PSVMVMCBCTRL pVmcbNstGstCtrl = &pVmcbNstGst->ctrl;
2098	uint16_t const uGuestPauseFilterCount = pVM->hm.s.svm.cPauseFilter;
2099	uint16_t const uGuestPauseFilterThreshold = pVM->hm.s.svm.cPauseFilterThresholdTicks;
2100	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_PAUSE))
2101	{
2102	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2103	pVmcbNstGstCtrl->u16PauseFilterCount = RT_MIN(pCtx->hwvirt.svm.cPauseFilter, uGuestPauseFilterCount);
2104	pVmcbNstGstCtrl->u16PauseFilterThreshold = RT_MIN(pCtx->hwvirt.svm.cPauseFilterThreshold, uGuestPauseFilterThreshold);
2105	pVmcbNstGstCtrl->u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2106	}
2107	else
2108	{
2109	pVmcbNstGstCtrl->u16PauseFilterCount = uGuestPauseFilterCount;
2110	pVmcbNstGstCtrl->u16PauseFilterThreshold = uGuestPauseFilterThreshold;
2111	}
2112
2113	pVCpu->hm.s.fCtxChanged &= ~HM_CHANGED_GUEST_HWVIRT;
2114	}
2115	}
2116	#endif
2117
2118	/**
2119	* Exports the guest APIC TPR state into the VMCB.
2120	*
2121	* @returns VBox status code.
2122	* @param pVCpu The cross context virtual CPU structure.
2123	* @param pVmcb Pointer to the VM control block.
2124	*/
2125	static int hmR0SvmExportGuestApicTpr(PVMCPU pVCpu, PSVMVMCB pVmcb)
2126	{
2127	if (ASMAtomicUoReadU64(&pVCpu->hm.s.fCtxChanged) & HM_CHANGED_GUEST_APIC_TPR)
2128	{
2129	PVM pVM = pVCpu->CTX_SUFF(pVM);
2130	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2131	if ( PDMHasApic(pVM)
2132	&& APICIsEnabled(pVCpu))
2133	{
2134	bool fPendingIntr;
2135	uint8_t u8Tpr;
2136	int rc = APICGetTpr(pVCpu, &u8Tpr, &fPendingIntr, NULL /* pu8PendingIrq */);
2137	AssertRCReturn(rc, rc);
2138
2139	/* Assume that we need to trap all TPR accesses and thus need not check on
2140	every #VMEXIT if we should update the TPR. */
2141	Assert(pVmcb->ctrl.IntCtrl.n.u1VIntrMasking);
2142	pVCpu->hm.s.svm.fSyncVTpr = false;
2143
2144	if (!pVM->hm.s.fTPRPatchingActive)
2145	{
2146	/* Bits 3-0 of the VTPR field correspond to bits 7-4 of the TPR (which is the Task-Priority Class). */
2147	pVmcb->ctrl.IntCtrl.n.u8VTPR = (u8Tpr >> 4);
2148
2149	/* If there are interrupts pending, intercept CR8 writes to evaluate ASAP if we
2150	can deliver the interrupt to the guest. */
2151	if (fPendingIntr)
2152	pVmcb->ctrl.u16InterceptWrCRx \|= RT_BIT(8);
2153	else
2154	{
2155	pVmcb->ctrl.u16InterceptWrCRx &= ~RT_BIT(8);
2156	pVCpu->hm.s.svm.fSyncVTpr = true;
2157	}
2158
2159	pVmcb->ctrl.u32VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS \| HMSVM_VMCB_CLEAN_INT_CTRL);
2160	}
2161	else
2162	{
2163	/* 32-bit guests uses LSTAR MSR for patching guest code which touches the TPR. */
2164	pVmcb->guest.u64LSTAR = u8Tpr;
2165	uint8_t pbMsrBitmap = (uint8_t )pVCpu->hm.s.svm.pvMsrBitmap;
2166
2167	/* If there are interrupts pending, intercept LSTAR writes, otherwise don't intercept reads or writes. */
2168	if (fPendingIntr)
2169	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_INTERCEPT_WRITE);
2170	else
2171	{
2172	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
2173	pVCpu->hm.s.svm.fSyncVTpr = true;
2174	}
2175	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_IOPM_MSRPM;
2176	}
2177	}
2178	ASMAtomicUoAndU64(&pVCpu->hm.s.fCtxChanged, ~HM_CHANGED_GUEST_APIC_TPR);
2179	}
2180	return VINF_SUCCESS;
2181	}
2182
2183
2184	/**
2185	* Sets up the exception interrupts required for guest (or nested-guest)
2186	* execution in the VMCB.
2187	*
2188	* @param pVCpu The cross context virtual CPU structure.
2189	* @param pVmcb Pointer to the VM control block.
2190	*
2191	* @remarks No-long-jump zone!!!
2192	*/
2193	static void hmR0SvmExportGuestXcptIntercepts(PVMCPU pVCpu, PSVMVMCB pVmcb)
2194	{
2195	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2196
2197	/* If we modify intercepts from here, please check & adjust hmR0SvmMergeVmcbCtrlsNested() if required. */
2198	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_SVM_GUEST_XCPT_INTERCEPTS)
2199	{
2200	/* Trap #UD for GIM provider (e.g. for hypercalls). */
2201	if (pVCpu->hm.s.fGIMTrapXcptUD)
2202	hmR0SvmSetXcptIntercept(pVmcb, X86_XCPT_UD);
2203	else
2204	hmR0SvmClearXcptIntercept(pVCpu, pVmcb, X86_XCPT_UD);
2205
2206	/* Trap #BP for INT3 debug breakpoints set by the VM debugger. */
2207	if (pVCpu->CTX_SUFF(pVM)->dbgf.ro.cEnabledInt3Breakpoints)
2208	hmR0SvmSetXcptIntercept(pVmcb, X86_XCPT_BP);
2209	else
2210	hmR0SvmClearXcptIntercept(pVCpu, pVmcb, X86_XCPT_BP);
2211
2212	/* The remaining intercepts are handled elsewhere, e.g. in hmR0SvmExportGuestCR0(). */
2213	pVCpu->hm.s.fCtxChanged &= ~HM_CHANGED_SVM_GUEST_XCPT_INTERCEPTS;
2214	}
2215	}
2216
2217
2218	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
2219	/**
2220	* Merges guest and nested-guest intercepts for executing the nested-guest using
2221	* hardware-assisted SVM.
2222	*
2223	* This merges the guest and nested-guest intercepts in a way that if the outer
2224	* guest intercept is set we need to intercept it in the nested-guest as
2225	* well.
2226	*
2227	* @param pVCpu The cross context virtual CPU structure.
2228	* @param pVmcbNstGst Pointer to the nested-guest VM control block.
2229	* @param pCtx Pointer to the nested-guest-CPU context.
2230	*/
2231	static void hmR0SvmMergeVmcbCtrlsNested(PVMCPU pVCpu, PCCPUMCTX pCtx)
2232	{
2233	PVM pVM = pVCpu->CTX_SUFF(pVM);
2234	PCSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
2235	PSVMVMCB pVmcbNstGst = pCtx->hwvirt.svm.CTX_SUFF(pVmcb);
2236	PSVMVMCBCTRL pVmcbNstGstCtrl = &pVmcbNstGst->ctrl;
2237
2238	/* Merge the guest's CR intercepts into the nested-guest VMCB. */
2239	pVmcbNstGstCtrl->u16InterceptRdCRx \|= pVmcb->ctrl.u16InterceptRdCRx;
2240	pVmcbNstGstCtrl->u16InterceptWrCRx \|= pVmcb->ctrl.u16InterceptWrCRx;
2241
2242	/* Always intercept CR4 writes for tracking PGM mode changes. */
2243	pVmcbNstGstCtrl->u16InterceptWrCRx \|= RT_BIT(4);
2244
2245	/* Without nested paging, intercept CR3 reads and writes as we load shadow page tables. */
2246	if (!pVM->hm.s.fNestedPaging)
2247	{
2248	pVmcbNstGstCtrl->u16InterceptRdCRx \|= RT_BIT(3);
2249	pVmcbNstGstCtrl->u16InterceptWrCRx \|= RT_BIT(3);
2250	}
2251
2252	/** @todo Figure out debugging with nested-guests, till then just intercept
2253	* all DR[0-15] accesses. */
2254	pVmcbNstGstCtrl->u16InterceptRdDRx \|= 0xffff;
2255	pVmcbNstGstCtrl->u16InterceptWrDRx \|= 0xffff;
2256
2257	/*
2258	* Merge the guest's exception intercepts into the nested-guest VMCB.
2259	*
2260	* - \#UD: Exclude these as the outer guest's GIM hypercalls are not applicable
2261	* while executing the nested-guest.
2262	*
2263	* - \#BP: Exclude breakpoints set by the VM debugger for the outer guest. This can
2264	* be tweaked later depending on how we wish to implement breakpoints.
2265	*
2266	* Warning!! This ASSUMES we only intercept \#UD for hypercall purposes and \#BP
2267	* for VM debugger breakpoints, see hmR0SvmExportGuestXcptIntercepts().
2268	*/
2269	#ifndef HMSVM_ALWAYS_TRAP_ALL_XCPTS
2270	pVmcbNstGstCtrl->u32InterceptXcpt \|= (pVmcb->ctrl.u32InterceptXcpt & ~( RT_BIT(X86_XCPT_UD)
2271	\| RT_BIT(X86_XCPT_BP)));
2272	#else
2273	pVmcbNstGstCtrl->u32InterceptXcpt \|= pVmcb->ctrl.u32InterceptXcpt;
2274	#endif
2275
2276	/*
2277	* Adjust intercepts while executing the nested-guest that differ from the
2278	* outer guest intercepts.
2279	*
2280	* - VINTR: Exclude the outer guest intercept as we don't need to cause VINTR #VMEXITs
2281	* that belong to the nested-guest to the outer guest.
2282	*
2283	* - VMMCALL: Exclude the outer guest intercept as when it's also not intercepted by
2284	* the nested-guest, the physical CPU raises a \#UD exception as expected.
2285	*/
2286	pVmcbNstGstCtrl->u64InterceptCtrl \|= (pVmcb->ctrl.u64InterceptCtrl & ~( SVM_CTRL_INTERCEPT_VINTR
2287	\| SVM_CTRL_INTERCEPT_VMMCALL))
2288	\| HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS;
2289
2290	Assert( (pVmcbNstGstCtrl->u64InterceptCtrl & HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS)
2291	== HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS);
2292
2293	/* Finally, update the VMCB clean bits. */
2294	pVmcbNstGstCtrl->u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2295	}
2296	#endif
2297
2298
2299	/**
2300	* Selects the appropriate function to run guest code.
2301	*
2302	* @returns VBox status code.
2303	* @param pVCpu The cross context virtual CPU structure.
2304	*
2305	* @remarks No-long-jump zone!!!
2306	*/
2307	static int hmR0SvmSelectVMRunHandler(PVMCPU pVCpu)
2308	{
2309	if (CPUMIsGuestInLongMode(pVCpu))
2310	{
2311	#ifndef VBOX_ENABLE_64_BITS_GUESTS
2312	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
2313	#endif
2314	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.fAllow64BitGuests); /* Guaranteed by hmR3InitFinalizeR0(). */
2315	#if HC_ARCH_BITS == 32
2316	/* 32-bit host. We need to switch to 64-bit before running the 64-bit guest. */
2317	pVCpu->hm.s.svm.pfnVMRun = SVMR0VMSwitcherRun64;
2318	#else
2319	/* 64-bit host or hybrid host. */
2320	pVCpu->hm.s.svm.pfnVMRun = SVMR0VMRun64;
2321	#endif
2322	}
2323	else
2324	{
2325	/* Guest is not in long mode, use the 32-bit handler. */
2326	pVCpu->hm.s.svm.pfnVMRun = SVMR0VMRun;
2327	}
2328	return VINF_SUCCESS;
2329	}
2330
2331
2332	/**
2333	* Enters the AMD-V session.
2334	*
2335	* @returns VBox status code.
2336	* @param pVCpu The cross context virtual CPU structure.
2337	* @param pHostCpu Pointer to the CPU info struct.
2338	*/
2339	VMMR0DECL(int) SVMR0Enter(PVMCPU pVCpu, PHMGLOBALCPUINFO pHostCpu)
2340	{
2341	AssertPtr(pVCpu);
2342	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.svm.fSupported);
2343	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2344	RT_NOREF(pHostCpu);
2345
2346	LogFlowFunc(("pVCpu=%p\n", pVCpu));
2347	Assert((pVCpu->hm.s.fCtxChanged & (HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE))
2348	== (HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE));
2349
2350	pVCpu->hm.s.fLeaveDone = false;
2351	return VINF_SUCCESS;
2352	}
2353
2354
2355	/**
2356	* Thread-context callback for AMD-V.
2357	*
2358	* @param enmEvent The thread-context event.
2359	* @param pVCpu The cross context virtual CPU structure.
2360	* @param fGlobalInit Whether global VT-x/AMD-V init. is used.
2361	* @thread EMT(pVCpu)
2362	*/
2363	VMMR0DECL(void) SVMR0ThreadCtxCallback(RTTHREADCTXEVENT enmEvent, PVMCPU pVCpu, bool fGlobalInit)
2364	{
2365	NOREF(fGlobalInit);
2366
2367	switch (enmEvent)
2368	{
2369	case RTTHREADCTXEVENT_OUT:
2370	{
2371	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2372	Assert(VMMR0ThreadCtxHookIsEnabled(pVCpu));
2373	VMCPU_ASSERT_EMT(pVCpu);
2374
2375	/* No longjmps (log-flush, locks) in this fragile context. */
2376	VMMRZCallRing3Disable(pVCpu);
2377
2378	if (!pVCpu->hm.s.fLeaveDone)
2379	{
2380	hmR0SvmLeave(pVCpu, false /* fImportState */);
2381	pVCpu->hm.s.fLeaveDone = true;
2382	}
2383
2384	/* Leave HM context, takes care of local init (term). */
2385	int rc = HMR0LeaveCpu(pVCpu);
2386	AssertRC(rc); NOREF(rc);
2387
2388	/* Restore longjmp state. */
2389	VMMRZCallRing3Enable(pVCpu);
2390	STAM_REL_COUNTER_INC(&pVCpu->hm.s.StatSwitchPreempt);
2391	break;
2392	}
2393
2394	case RTTHREADCTXEVENT_IN:
2395	{
2396	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2397	Assert(VMMR0ThreadCtxHookIsEnabled(pVCpu));
2398	VMCPU_ASSERT_EMT(pVCpu);
2399
2400	/* No longjmps (log-flush, locks) in this fragile context. */
2401	VMMRZCallRing3Disable(pVCpu);
2402
2403	/*
2404	* Initialize the bare minimum state required for HM. This takes care of
2405	* initializing AMD-V if necessary (onlined CPUs, local init etc.)
2406	*/
2407	int rc = hmR0EnterCpu(pVCpu);
2408	AssertRC(rc); NOREF(rc);
2409	Assert((pVCpu->hm.s.fCtxChanged & (HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE))
2410	== (HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE));
2411
2412	pVCpu->hm.s.fLeaveDone = false;
2413
2414	/* Restore longjmp state. */
2415	VMMRZCallRing3Enable(pVCpu);
2416	break;
2417	}
2418
2419	default:
2420	break;
2421	}
2422	}
2423
2424
2425	/**
2426	* Saves the host state.
2427	*
2428	* @returns VBox status code.
2429	* @param pVCpu The cross context virtual CPU structure.
2430	*
2431	* @remarks No-long-jump zone!!!
2432	*/
2433	VMMR0DECL(int) SVMR0ExportHostState(PVMCPU pVCpu)
2434	{
2435	NOREF(pVCpu);
2436
2437	/* Nothing to do here. AMD-V does this for us automatically during the world-switch. */
2438	ASMAtomicUoAndU64(&pVCpu->hm.s.fCtxChanged, ~HM_CHANGED_HOST_CONTEXT);
2439	return VINF_SUCCESS;
2440	}
2441
2442
2443	/**
2444	* Exports the guest state from the guest-CPU context into the VMCB.
2445	*
2446	* The CPU state will be loaded from these fields on every successful VM-entry.
2447	* Also sets up the appropriate VMRUN function to execute guest code based on
2448	* the guest CPU mode.
2449	*
2450	* @returns VBox status code.
2451	* @param pVCpu The cross context virtual CPU structure.
2452	*
2453	* @remarks No-long-jump zone!!!
2454	*/
2455	static int hmR0SvmExportGuestState(PVMCPU pVCpu)
2456	{
2457	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatExportGuestState, x);
2458
2459	PSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
2460	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2461
2462	Assert(pVmcb);
2463	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(pCtx);
2464
2465	pVmcb->guest.u64RIP = pCtx->rip;
2466	pVmcb->guest.u64RSP = pCtx->rsp;
2467	pVmcb->guest.u64RFlags = pCtx->eflags.u32;
2468	pVmcb->guest.u64RAX = pCtx->rax;
2469	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
2470	if (pVmcb->ctrl.IntCtrl.n.u1VGifEnable)
2471	{
2472	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_VGIF);
2473	pVmcb->ctrl.IntCtrl.n.u1VGif = pCtx->hwvirt.fGif;
2474	}
2475	#endif
2476
2477	RTCCUINTREG const fEFlags = ASMIntDisableFlags();
2478
2479	int rc = hmR0SvmExportGuestControlRegs(pVCpu, pVmcb);
2480	AssertRCReturnStmt(rc, ASMSetFlags(fEFlags), rc);
2481
2482	hmR0SvmExportGuestSegmentRegs(pVCpu, pVmcb);
2483	hmR0SvmExportGuestMsrs(pVCpu, pVmcb);
2484	hmR0SvmExportGuestXcptIntercepts(pVCpu, pVmcb);
2485
2486	ASMSetFlags(fEFlags);
2487
2488	/* hmR0SvmExportGuestApicTpr() must be called -after- hmR0SvmExportGuestMsrs() as we
2489	otherwise we would overwrite the LSTAR MSR that we use for TPR patching. */
2490	hmR0SvmExportGuestApicTpr(pVCpu, pVmcb);
2491
2492	rc = hmR0SvmSelectVMRunHandler(pVCpu);
2493	AssertRCReturn(rc, rc);
2494
2495	/* Clear any bits that may be set but exported unconditionally or unused/reserved bits. */
2496	ASMAtomicUoAndU64(&pVCpu->hm.s.fCtxChanged, ~( HM_CHANGED_GUEST_RIP
2497	\| HM_CHANGED_GUEST_RFLAGS
2498	\| HM_CHANGED_GUEST_GPRS_MASK
2499	\| HM_CHANGED_GUEST_X87
2500	\| HM_CHANGED_GUEST_SSE_AVX
2501	\| HM_CHANGED_GUEST_OTHER_XSAVE
2502	\| HM_CHANGED_GUEST_XCRx
2503	\| HM_CHANGED_GUEST_TSC_AUX
2504	\| HM_CHANGED_GUEST_OTHER_MSRS
2505	\| HM_CHANGED_GUEST_HWVIRT
2506	\| (HM_CHANGED_KEEPER_STATE_MASK & ~HM_CHANGED_SVM_GUEST_XCPT_INTERCEPTS)));
2507
2508	#ifdef VBOX_STRICT
2509	/*
2510	* All of the guest-CPU state and SVM keeper bits should be exported here by now,
2511	* except for the host-context and/or shared host-guest context bits.
2512	*/
2513	uint64_t const fCtxChanged = ASMAtomicUoReadU64(&pVCpu->hm.s.fCtxChanged);
2514	RT_UNTRUSTED_NONVOLATILE_COPY_FENCE();
2515	AssertMsg(!(fCtxChanged & (HM_CHANGED_ALL_GUEST & ~HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE)),
2516	("fCtxChanged=%#RX64\n", fCtxChanged));
2517
2518	/*
2519	* If we need to log state that isn't always imported, we'll need to import them here.
2520	* See hmR0SvmPostRunGuest() for which part of the state is imported uncondtionally.
2521	*/
2522	hmR0SvmLogState(pVCpu, pVmcb, "hmR0SvmExportGuestState", 0 /* fFlags /, 0 / uVerbose */);
2523	#endif
2524
2525	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExportGuestState, x);
2526	return VINF_SUCCESS;
2527	}
2528
2529
2530	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
2531	/**
2532	* Merges the guest and nested-guest MSR permission bitmap.
2533	*
2534	* If the guest is intercepting an MSR we need to intercept it regardless of
2535	* whether the nested-guest is intercepting it or not.
2536	*
2537	* @param pHostCpu Pointer to the physical CPU HM info. struct.
2538	* @param pVCpu The cross context virtual CPU structure.
2539	* @param pCtx Pointer to the nested-guest-CPU context.
2540	*
2541	* @remarks No-long-jmp zone!!!
2542	*/
2543	DECLINLINE(void) hmR0SvmMergeMsrpmNested(PHMGLOBALCPUINFO pHostCpu, PVMCPU pVCpu, PCCPUMCTX pCtx)
2544	{
2545	uint64_t const pu64GstMsrpm = (uint64_t const )pVCpu->hm.s.svm.pvMsrBitmap;
2546	uint64_t const pu64NstGstMsrpm = (uint64_t const )pCtx->hwvirt.svm.CTX_SUFF(pvMsrBitmap);
2547	uint64_t pu64DstMsrpm = (uint64_t )pHostCpu->n.svm.pvNstGstMsrpm;
2548
2549	/* MSRPM bytes from offset 0x1800 are reserved, so we stop merging there. */
2550	uint32_t const offRsvdQwords = 0x1800 >> 3;
2551	for (uint32_t i = 0; i < offRsvdQwords; i++)
2552	pu64DstMsrpm[i] = pu64NstGstMsrpm[i] \| pu64GstMsrpm[i];
2553	}
2554
2555
2556	/**
2557	* Caches the nested-guest VMCB fields before we modify them for execution using
2558	* hardware-assisted SVM.
2559	*
2560	* @returns true if the VMCB was previously already cached, false otherwise.
2561	* @param pCtx Pointer to the guest-CPU context.
2562	*
2563	* @sa HMSvmNstGstVmExitNotify.
2564	*/
2565	static bool hmR0SvmCacheVmcbNested(PVMCPU pVCpu, PCCPUMCTX pCtx)
2566	{
2567	/*
2568	* Cache the nested-guest programmed VMCB fields if we have not cached it yet.
2569	* Otherwise we risk re-caching the values we may have modified, see @bugref{7243#c44}.
2570	*
2571	* Nested-paging CR3 is not saved back into the VMCB on #VMEXIT, hence no need to
2572	* cache and restore it, see AMD spec. 15.25.4 "Nested Paging and VMRUN/#VMEXIT".
2573	*/
2574	PSVMNESTEDVMCBCACHE pVmcbNstGstCache = &pVCpu->hm.s.svm.NstGstVmcbCache;
2575	bool const fWasCached = pVmcbNstGstCache->fCacheValid;
2576	if (!fWasCached)
2577	{
2578	PCSVMVMCB pVmcbNstGst = pCtx->hwvirt.svm.CTX_SUFF(pVmcb);
2579	PCSVMVMCBCTRL pVmcbNstGstCtrl = &pVmcbNstGst->ctrl;
2580	pVmcbNstGstCache->u16InterceptRdCRx = pVmcbNstGstCtrl->u16InterceptRdCRx;
2581	pVmcbNstGstCache->u16InterceptWrCRx = pVmcbNstGstCtrl->u16InterceptWrCRx;
2582	pVmcbNstGstCache->u16InterceptRdDRx = pVmcbNstGstCtrl->u16InterceptRdDRx;
2583	pVmcbNstGstCache->u16InterceptWrDRx = pVmcbNstGstCtrl->u16InterceptWrDRx;
2584	pVmcbNstGstCache->u16PauseFilterThreshold = pVmcbNstGstCtrl->u16PauseFilterThreshold;
2585	pVmcbNstGstCache->u16PauseFilterCount = pVmcbNstGstCtrl->u16PauseFilterCount;
2586	pVmcbNstGstCache->u32InterceptXcpt = pVmcbNstGstCtrl->u32InterceptXcpt;
2587	pVmcbNstGstCache->u64InterceptCtrl = pVmcbNstGstCtrl->u64InterceptCtrl;
2588	pVmcbNstGstCache->u64TSCOffset = pVmcbNstGstCtrl->u64TSCOffset;
2589	pVmcbNstGstCache->fVIntrMasking = pVmcbNstGstCtrl->IntCtrl.n.u1VIntrMasking;
2590	pVmcbNstGstCache->fNestedPaging = pVmcbNstGstCtrl->NestedPagingCtrl.n.u1NestedPaging;
2591	pVmcbNstGstCache->fLbrVirt = pVmcbNstGstCtrl->LbrVirt.n.u1LbrVirt;
2592	pVmcbNstGstCache->fCacheValid = true;
2593	Log4Func(("Cached VMCB fields\n"));
2594	}
2595
2596	return fWasCached;
2597	}
2598
2599
2600	/**
2601	* Sets up the nested-guest VMCB for execution using hardware-assisted SVM.
2602	*
2603	* This is done the first time we enter nested-guest execution using SVM R0
2604	* until the nested-guest \#VMEXIT (not to be confused with physical CPU
2605	* \#VMEXITs which may or may not cause a corresponding nested-guest \#VMEXIT).
2606	*
2607	* @param pVCpu The cross context virtual CPU structure.
2608	*/
2609	static void hmR0SvmSetupVmcbNested(PVMCPU pVCpu)
2610	{
2611	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2612	PSVMVMCB pVmcbNstGst = pCtx->hwvirt.svm.CTX_SUFF(pVmcb);
2613	PSVMVMCBCTRL pVmcbNstGstCtrl = &pVmcbNstGst->ctrl;
2614
2615	/*
2616	* First cache the nested-guest VMCB fields we may potentially modify.
2617	*/
2618	bool const fVmcbCached = hmR0SvmCacheVmcbNested(pVCpu, pCtx);
2619	if (!fVmcbCached)
2620	{
2621	/*
2622	* The IOPM of the nested-guest can be ignored because the the guest always
2623	* intercepts all IO port accesses. Thus, we'll swap to the guest IOPM rather
2624	* than the nested-guest IOPM and swap the field back on the #VMEXIT.
2625	*/
2626	pVmcbNstGstCtrl->u64IOPMPhysAddr = g_HCPhysIOBitmap;
2627
2628	/*
2629	* Use the same nested-paging as the outer guest. We can't dynamically switch off
2630	* nested-paging suddenly while executing a VM (see assertion at the end of
2631	* Trap0eHandler() in PGMAllBth.h).
2632	*/
2633	pVmcbNstGstCtrl->NestedPagingCtrl.n.u1NestedPaging = pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging;
2634
2635	/* Always enable V_INTR_MASKING as we do not want to allow access to the physical APIC TPR. */
2636	pVmcbNstGstCtrl->IntCtrl.n.u1VIntrMasking = 1;
2637
2638	/*
2639	* Turn off TPR syncing on #VMEXIT for nested-guests as CR8 intercepts are subject
2640	* to the nested-guest intercepts and we always run with V_INTR_MASKING.
2641	*/
2642	pVCpu->hm.s.svm.fSyncVTpr = false;
2643
2644	#ifdef DEBUG_ramshankar
2645	/* For debugging purposes - copy the LBR info. from outer guest VMCB. */
2646	pVmcbNstGstCtrl->LbrVirt.n.u1LbrVirt = pVmcb->ctrl.LbrVirt.n.u1LbrVirt;
2647	#endif
2648
2649	/*
2650	* If we don't expose Virtualized-VMSAVE/VMLOAD feature to the outer guest, we
2651	* need to intercept VMSAVE/VMLOAD instructions executed by the nested-guest.
2652	*/
2653	if (!pVCpu->CTX_SUFF(pVM)->cpum.ro.GuestFeatures.fSvmVirtVmsaveVmload)
2654	pVmcbNstGstCtrl->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_VMSAVE
2655	\| SVM_CTRL_INTERCEPT_VMLOAD;
2656
2657	/*
2658	* If we don't expose Virtual GIF feature to the outer guest, we need to intercept
2659	* CLGI/STGI instructions executed by the nested-guest.
2660	*/
2661	if (!pVCpu->CTX_SUFF(pVM)->cpum.ro.GuestFeatures.fSvmVGif)
2662	pVmcbNstGstCtrl->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_CLGI
2663	\| SVM_CTRL_INTERCEPT_STGI;
2664
2665	/* Merge the guest and nested-guest intercepts. */
2666	hmR0SvmMergeVmcbCtrlsNested(pVCpu, pCtx);
2667
2668	/* Update the VMCB clean bits. */
2669	pVmcbNstGstCtrl->u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2670	}
2671	else
2672	{
2673	Assert(!pVCpu->hm.s.svm.fSyncVTpr);
2674	Assert(pVmcbNstGstCtrl->u64IOPMPhysAddr == g_HCPhysIOBitmap);
2675	Assert(RT_BOOL(pVmcbNstGstCtrl->NestedPagingCtrl.n.u1NestedPaging) == pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
2676	}
2677	}
2678
2679
2680	/**
2681	* Exports the nested-guest state into the VMCB.
2682	*
2683	* We need to export the entire state as we could be continuing nested-guest
2684	* execution at any point (not just immediately after VMRUN) and thus the VMCB
2685	* can be out-of-sync with the nested-guest state if it was executed in IEM.
2686	*
2687	* @returns VBox status code.
2688	* @param pVCpu The cross context virtual CPU structure.
2689	* @param pCtx Pointer to the guest-CPU context.
2690	*
2691	* @remarks No-long-jump zone!!!
2692	*/
2693	static int hmR0SvmExportGuestStateNested(PVMCPU pVCpu)
2694	{
2695	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatExportGuestState, x);
2696
2697	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2698	PSVMVMCB pVmcbNstGst = pCtx->hwvirt.svm.CTX_SUFF(pVmcb);
2699	Assert(pVmcbNstGst);
2700
2701	hmR0SvmSetupVmcbNested(pVCpu);
2702
2703	pVmcbNstGst->guest.u64RIP = pCtx->rip;
2704	pVmcbNstGst->guest.u64RSP = pCtx->rsp;
2705	pVmcbNstGst->guest.u64RFlags = pCtx->eflags.u32;
2706	pVmcbNstGst->guest.u64RAX = pCtx->rax;
2707
2708	RTCCUINTREG const fEFlags = ASMIntDisableFlags();
2709
2710	int rc = hmR0SvmExportGuestControlRegs(pVCpu, pVmcbNstGst);
2711	AssertRCReturnStmt(rc, ASMSetFlags(fEFlags), rc);
2712
2713	hmR0SvmExportGuestSegmentRegs(pVCpu, pVmcbNstGst);
2714	hmR0SvmExportGuestMsrs(pVCpu, pVmcbNstGst);
2715	hmR0SvmExportGuestHwvirtStateNested(pVCpu, pVmcbNstGst);
2716
2717	ASMSetFlags(fEFlags);
2718
2719	/* Nested VGIF not supported yet. */
2720	Assert(!pVmcbNstGst->ctrl.IntCtrl.n.u1VGifEnable);
2721
2722	rc = hmR0SvmSelectVMRunHandler(pVCpu);
2723	AssertRCReturn(rc, rc);
2724
2725	/* Clear any bits that may be set but exported unconditionally or unused/reserved bits. */
2726	ASMAtomicUoAndU64(&pVCpu->hm.s.fCtxChanged, ~( HM_CHANGED_GUEST_RIP
2727	\| HM_CHANGED_GUEST_RFLAGS
2728	\| HM_CHANGED_GUEST_GPRS_MASK
2729	\| HM_CHANGED_GUEST_APIC_TPR
2730	\| HM_CHANGED_GUEST_X87
2731	\| HM_CHANGED_GUEST_SSE_AVX
2732	\| HM_CHANGED_GUEST_OTHER_XSAVE
2733	\| HM_CHANGED_GUEST_XCRx
2734	\| HM_CHANGED_GUEST_TSC_AUX
2735	\| HM_CHANGED_GUEST_OTHER_MSRS
2736	\| HM_CHANGED_SVM_GUEST_XCPT_INTERCEPTS
2737	\| (HM_CHANGED_KEEPER_STATE_MASK & ~HM_CHANGED_SVM_MASK)));
2738
2739	#ifdef VBOX_STRICT
2740	/*
2741	* All of the guest-CPU state and SVM keeper bits should be exported here by now, except
2742	* for the host-context and/or shared host-guest context bits.
2743	*/
2744	uint64_t const fCtxChanged = ASMAtomicUoReadU64(&pVCpu->hm.s.fCtxChanged);
2745	RT_UNTRUSTED_NONVOLATILE_COPY_FENCE();
2746	AssertMsg(!(fCtxChanged & (HM_CHANGED_ALL_GUEST & ~HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE)),
2747	("fCtxChanged=%#RX64\n", fCtxChanged));
2748
2749	/*
2750	* If we need to log state that isn't always imported, we'll need to import them here.
2751	* See hmR0SvmPostRunGuest() for which part of the state is imported uncondtionally.
2752	*/
2753	hmR0SvmLogState(pVCpu, pVmcbNstGst, "hmR0SvmExportGuestStateNested", 0 /* fFlags /, 0 / uVerbose */);
2754	#endif
2755
2756	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExportGuestState, x);
2757	return rc;
2758	}
2759	#endif /* VBOX_WITH_NESTED_HWVIRT_SVM */
2760
2761
2762	/**
2763	* Exports the state shared between the host and guest (or nested-guest) into
2764	* the VMCB.
2765	*
2766	* @param pVCpu The cross context virtual CPU structure.
2767	* @param pVmcb Pointer to the VM control block.
2768	*
2769	* @remarks No-long-jump zone!!!
2770	*/
2771	static void hmR0SvmExportSharedState(PVMCPU pVCpu, PSVMVMCB pVmcb)
2772	{
2773	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2774	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2775
2776	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_DR_MASK)
2777	{
2778	/** @todo Figure out stepping with nested-guest. */
2779	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2780	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
2781	hmR0SvmExportSharedDebugState(pVCpu, pVmcb);
2782	else
2783	{
2784	pVmcb->guest.u64DR6 = pCtx->dr[6];
2785	pVmcb->guest.u64DR7 = pCtx->dr[7];
2786	}
2787	}
2788
2789	pVCpu->hm.s.fCtxChanged &= ~HM_CHANGED_GUEST_DR_MASK;
2790	AssertMsg(!(pVCpu->hm.s.fCtxChanged & HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE),
2791	("fCtxChanged=%#RX64\n", pVCpu->hm.s.fCtxChanged));
2792	}
2793
2794
2795	/**
2796	* Worker for SVMR0ImportStateOnDemand.
2797	*
2798	* @param pVCpu The cross context virtual CPU structure.
2799	* @param fWhat What to import, CPUMCTX_EXTRN_XXX.
2800	*/
2801	static void hmR0SvmImportGuestState(PVMCPU pVCpu, uint64_t fWhat)
2802	{
2803	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatImportGuestState, x);
2804
2805	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2806	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
2807	PCSVMVMCBSTATESAVE pVmcbGuest = &pVmcb->guest;
2808	PCSVMVMCBCTRL pVmcbCtrl = &pVmcb->ctrl;
2809
2810	Log4Func(("fExtrn=%#RX64 fWhat=%#RX64\n", pCtx->fExtrn, fWhat));
2811
2812	/*
2813	* We disable interrupts to make the updating of the state and in particular
2814	* the fExtrn modification atomic wrt to preemption hooks.
2815	*/
2816	RTCCUINTREG const fEFlags = ASMIntDisableFlags();
2817
2818	fWhat &= pCtx->fExtrn;
2819	if (fWhat)
2820	{
2821	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
2822	if (fWhat & CPUMCTX_EXTRN_HWVIRT)
2823	{
2824	if ( !CPUMIsGuestInSvmNestedHwVirtMode(pCtx)
2825	&& pVmcbCtrl->IntCtrl.n.u1VGifEnable)
2826	{
2827	/* We don't yet support passing VGIF feature to the guest. */
2828	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.svm.fVGif);
2829	pCtx->hwvirt.fGif = pVmcbCtrl->IntCtrl.n.u1VGif;
2830	}
2831	}
2832
2833	if (fWhat & CPUMCTX_EXTRN_HM_SVM_HWVIRT_VIRQ)
2834	{
2835	if ( !pVmcbCtrl->IntCtrl.n.u1VIrqPending
2836	&& VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_NESTED_GUEST))
2837	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INTERRUPT_NESTED_GUEST);
2838	}
2839	#endif
2840
2841	if (fWhat & CPUMCTX_EXTRN_HM_SVM_INT_SHADOW)
2842	{
2843	if (pVmcbCtrl->IntShadow.n.u1IntShadow)
2844	EMSetInhibitInterruptsPC(pVCpu, pVmcbGuest->u64RIP);
2845	else if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS))
2846	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS);
2847	}
2848
2849	if (fWhat & CPUMCTX_EXTRN_RIP)
2850	pCtx->rip = pVmcbGuest->u64RIP;
2851
2852	if (fWhat & CPUMCTX_EXTRN_RFLAGS)
2853	pCtx->eflags.u32 = pVmcbGuest->u64RFlags;
2854
2855	if (fWhat & CPUMCTX_EXTRN_RSP)
2856	pCtx->rsp = pVmcbGuest->u64RSP;
2857
2858	if (fWhat & CPUMCTX_EXTRN_RAX)
2859	pCtx->rax = pVmcbGuest->u64RAX;
2860
2861	if (fWhat & CPUMCTX_EXTRN_SREG_MASK)
2862	{
2863	if (fWhat & CPUMCTX_EXTRN_CS)
2864	{
2865	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, CS, cs);
2866	/* Correct the CS granularity bit. Haven't seen it being wrong in any other register (yet). */
2867	/** @todo SELM might need to be fixed as it too should not care about the
2868	* granularity bit. See @bugref{6785}. */
2869	if ( !pCtx->cs.Attr.n.u1Granularity
2870	&& pCtx->cs.Attr.n.u1Present
2871	&& pCtx->cs.u32Limit > UINT32_C(0xfffff))
2872	{
2873	Assert((pCtx->cs.u32Limit & 0xfff) == 0xfff);
2874	pCtx->cs.Attr.n.u1Granularity = 1;
2875	}
2876	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, cs);
2877	}
2878	if (fWhat & CPUMCTX_EXTRN_SS)
2879	{
2880	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, SS, ss);
2881	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, ss);
2882	/*
2883	* Sync the hidden SS DPL field. AMD CPUs have a separate CPL field in the
2884	* VMCB and uses that and thus it's possible that when the CPL changes during
2885	* guest execution that the SS DPL isn't updated by AMD-V. Observed on some
2886	* AMD Fusion CPUs with 64-bit guests.
2887	*
2888	* See AMD spec. 15.5.1 "Basic operation".
2889	*/
2890	Assert(!(pVmcbGuest->u8CPL & ~0x3));
2891	uint8_t const uCpl = pVmcbGuest->u8CPL;
2892	if (pCtx->ss.Attr.n.u2Dpl != uCpl)
2893	pCtx->ss.Attr.n.u2Dpl = uCpl & 0x3;
2894	}
2895	if (fWhat & CPUMCTX_EXTRN_DS)
2896	{
2897	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, DS, ds);
2898	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, ds);
2899	}
2900	if (fWhat & CPUMCTX_EXTRN_ES)
2901	{
2902	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, ES, es);
2903	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, es);
2904	}
2905	if (fWhat & CPUMCTX_EXTRN_FS)
2906	{
2907	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, FS, fs);
2908	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, fs);
2909	}
2910	if (fWhat & CPUMCTX_EXTRN_GS)
2911	{
2912	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, GS, gs);
2913	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, gs);
2914	}
2915	}
2916
2917	if (fWhat & CPUMCTX_EXTRN_TABLE_MASK)
2918	{
2919	if (fWhat & CPUMCTX_EXTRN_TR)
2920	{
2921	/*
2922	* Fixup TR attributes so it's compatible with Intel. Important when saved-states
2923	* are used between Intel and AMD, see @bugref{6208#c39}.
2924	* ASSUME that it's normally correct and that we're in 32-bit or 64-bit mode.
2925	*/
2926	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, TR, tr);
2927	if (pCtx->tr.Attr.n.u4Type != X86_SEL_TYPE_SYS_386_TSS_BUSY)
2928	{
2929	if ( pCtx->tr.Attr.n.u4Type == X86_SEL_TYPE_SYS_386_TSS_AVAIL
2930	\|\| CPUMIsGuestInLongModeEx(pCtx))
2931	pCtx->tr.Attr.n.u4Type = X86_SEL_TYPE_SYS_386_TSS_BUSY;
2932	else if (pCtx->tr.Attr.n.u4Type == X86_SEL_TYPE_SYS_286_TSS_AVAIL)
2933	pCtx->tr.Attr.n.u4Type = X86_SEL_TYPE_SYS_286_TSS_BUSY;
2934	}
2935	}
2936
2937	if (fWhat & CPUMCTX_EXTRN_LDTR)
2938	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, LDTR, ldtr);
2939
2940	if (fWhat & CPUMCTX_EXTRN_GDTR)
2941	{
2942	pCtx->gdtr.cbGdt = pVmcbGuest->GDTR.u32Limit;
2943	pCtx->gdtr.pGdt = pVmcbGuest->GDTR.u64Base;
2944	}
2945
2946	if (fWhat & CPUMCTX_EXTRN_IDTR)
2947	{
2948	pCtx->idtr.cbIdt = pVmcbGuest->IDTR.u32Limit;
2949	pCtx->idtr.pIdt = pVmcbGuest->IDTR.u64Base;
2950	}
2951	}
2952
2953	if (fWhat & CPUMCTX_EXTRN_SYSCALL_MSRS)
2954	{
2955	pCtx->msrSTAR = pVmcbGuest->u64STAR;
2956	pCtx->msrLSTAR = pVmcbGuest->u64LSTAR;
2957	pCtx->msrCSTAR = pVmcbGuest->u64CSTAR;
2958	pCtx->msrSFMASK = pVmcbGuest->u64SFMASK;
2959	}
2960
2961	if (fWhat & CPUMCTX_EXTRN_SYSENTER_MSRS)
2962	{
2963	pCtx->SysEnter.cs = pVmcbGuest->u64SysEnterCS;
2964	pCtx->SysEnter.eip = pVmcbGuest->u64SysEnterEIP;
2965	pCtx->SysEnter.esp = pVmcbGuest->u64SysEnterESP;
2966	}
2967
2968	if (fWhat & CPUMCTX_EXTRN_KERNEL_GS_BASE)
2969	pCtx->msrKERNELGSBASE = pVmcbGuest->u64KernelGSBase;
2970
2971	if (fWhat & CPUMCTX_EXTRN_DR_MASK)
2972	{
2973	if (fWhat & CPUMCTX_EXTRN_DR6)
2974	{
2975	if (!pVCpu->hm.s.fUsingHyperDR7)
2976	pCtx->dr[6] = pVmcbGuest->u64DR6;
2977	else
2978	CPUMSetHyperDR6(pVCpu, pVmcbGuest->u64DR6);
2979	}
2980
2981	if (fWhat & CPUMCTX_EXTRN_DR7)
2982	{
2983	if (!pVCpu->hm.s.fUsingHyperDR7)
2984	pCtx->dr[7] = pVmcbGuest->u64DR7;
2985	else
2986	Assert(pVmcbGuest->u64DR7 == CPUMGetHyperDR7(pVCpu));
2987	}
2988	}
2989
2990	if (fWhat & CPUMCTX_EXTRN_CR_MASK)
2991	{
2992	if (fWhat & CPUMCTX_EXTRN_CR0)
2993	{
2994	/* We intercept changes to all CR0 bits except maybe TS & MP bits. */
2995	uint64_t const uCr0 = (pCtx->cr0 & ~(X86_CR0_TS \| X86_CR0_MP))
2996	\| (pVmcbGuest->u64CR0 & (X86_CR0_TS \| X86_CR0_MP));
2997	VMMRZCallRing3Disable(pVCpu); /* Calls into PGM which has Log statements. */
2998	CPUMSetGuestCR0(pVCpu, uCr0);
2999	VMMRZCallRing3Enable(pVCpu);
3000	}
3001
3002	if (fWhat & CPUMCTX_EXTRN_CR2)
3003	pCtx->cr2 = pVmcbGuest->u64CR2;
3004
3005	if (fWhat & CPUMCTX_EXTRN_CR3)
3006	{
3007	if ( pVmcbCtrl->NestedPagingCtrl.n.u1NestedPaging
3008	&& pCtx->cr3 != pVmcbGuest->u64CR3)
3009	{
3010	CPUMSetGuestCR3(pVCpu, pVmcbGuest->u64CR3);
3011	VMCPU_FF_SET(pVCpu, VMCPU_FF_HM_UPDATE_CR3);
3012	}
3013	}
3014
3015	/* Changes to CR4 are always intercepted. */
3016	}
3017
3018	/* Update fExtrn. */
3019	pCtx->fExtrn &= ~fWhat;
3020
3021	/* If everything has been imported, clear the HM keeper bit. */
3022	if (!(pCtx->fExtrn & HMSVM_CPUMCTX_EXTRN_ALL))
3023	{
3024	pCtx->fExtrn &= ~CPUMCTX_EXTRN_KEEPER_HM;
3025	Assert(!pCtx->fExtrn);
3026	}
3027	}
3028	else
3029	Assert(!pCtx->fExtrn \|\| (pCtx->fExtrn & HMSVM_CPUMCTX_EXTRN_ALL));
3030
3031	ASMSetFlags(fEFlags);
3032
3033	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatImportGuestState, x);
3034
3035	/*
3036	* Honor any pending CR3 updates.
3037	*
3038	* Consider this scenario: #VMEXIT -> VMMRZCallRing3Enable() -> do stuff that causes a longjmp
3039	* -> hmR0SvmCallRing3Callback() -> VMMRZCallRing3Disable() -> hmR0SvmImportGuestState()
3040	* -> Sets VMCPU_FF_HM_UPDATE_CR3 pending -> return from the longjmp -> continue with #VMEXIT
3041	* handling -> hmR0SvmImportGuestState() and here we are.
3042	*
3043	* The reason for such complicated handling is because VM-exits that call into PGM expect
3044	* CR3 to be up-to-date and thus any CR3-saves -before- the VM-exit (longjmp) would've
3045	* postponed the CR3 update via the force-flag and cleared CR3 from fExtrn. Any SVM R0
3046	* VM-exit handler that requests CR3 to be saved will end up here and we call PGMUpdateCR3().
3047	*
3048	* The longjmp exit path can't check these CR3 force-flags and call code that takes a lock again,
3049	* and does not process force-flag like regular exits to ring-3 either, we cover for it here.
3050	*/
3051	if ( VMMRZCallRing3IsEnabled(pVCpu)
3052	&& VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_UPDATE_CR3))
3053	{
3054	Assert(pCtx->cr3 == pVmcbGuest->u64CR3);
3055	PGMUpdateCR3(pVCpu, pCtx->cr3);
3056	}
3057	}
3058
3059
3060	/**
3061	* Saves the guest (or nested-guest) state from the VMCB into the guest-CPU
3062	* context.
3063	*
3064	* Currently there is no residual state left in the CPU that is not updated in the
3065	* VMCB.
3066	*
3067	* @returns VBox status code.
3068	* @param pVCpu The cross context virtual CPU structure.
3069	* @param fWhat What to import, CPUMCTX_EXTRN_XXX.
3070	*/
3071	VMMR0DECL(int) SVMR0ImportStateOnDemand(PVMCPU pVCpu, uint64_t fWhat)
3072	{
3073	hmR0SvmImportGuestState(pVCpu, fWhat);
3074	return VINF_SUCCESS;
3075	}
3076
3077
3078	/**
3079	* Does the necessary state syncing before returning to ring-3 for any reason
3080	* (longjmp, preemption, voluntary exits to ring-3) from AMD-V.
3081	*
3082	* @param pVCpu The cross context virtual CPU structure.
3083	* @param fImportState Whether to import the guest state from the VMCB back
3084	* to the guest-CPU context.
3085	*
3086	* @remarks No-long-jmp zone!!!
3087	*/
3088	static void hmR0SvmLeave(PVMCPU pVCpu, bool fImportState)
3089	{
3090	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
3091	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
3092	Assert(VMMR0IsLogFlushDisabled(pVCpu));
3093
3094	/*
3095	* !!! IMPORTANT !!!
3096	* If you modify code here, make sure to check whether hmR0SvmCallRing3Callback() needs to be updated too.
3097	*/
3098
3099	/* Save the guest state if necessary. */
3100	if (fImportState)
3101	hmR0SvmImportGuestState(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
3102
3103	/* Restore host FPU state if necessary and resync on next R0 reentry. */
3104	CPUMR0FpuStateMaybeSaveGuestAndRestoreHost(pVCpu);
3105	Assert(!CPUMIsGuestFPUStateActive(pVCpu));
3106
3107	/*
3108	* Restore host debug registers if necessary and resync on next R0 reentry.
3109	*/
3110	#ifdef VBOX_STRICT
3111	if (CPUMIsHyperDebugStateActive(pVCpu))
3112	{
3113	PSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb; /** @todo nested-guest. */
3114	Assert(pVmcb->ctrl.u16InterceptRdDRx == 0xffff);
3115	Assert(pVmcb->ctrl.u16InterceptWrDRx == 0xffff);
3116	}
3117	#endif
3118	CPUMR0DebugStateMaybeSaveGuestAndRestoreHost(pVCpu, false /* save DR6 */);
3119	Assert(!CPUMIsHyperDebugStateActive(pVCpu));
3120	Assert(!CPUMIsGuestDebugStateActive(pVCpu));
3121
3122	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatEntry);
3123	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatImportGuestState);
3124	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatExportGuestState);
3125	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatPreExit);
3126	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatExitHandling);
3127	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchLongJmpToR3);
3128
3129	VMCPU_CMPXCHG_STATE(pVCpu, VMCPUSTATE_STARTED_HM, VMCPUSTATE_STARTED_EXEC);
3130	}
3131
3132
3133	/**
3134	* Leaves the AMD-V session.
3135	*
3136	* Only used while returning to ring-3 either due to longjump or exits to
3137	* ring-3.
3138	*
3139	* @returns VBox status code.
3140	* @param pVCpu The cross context virtual CPU structure.
3141	*/
3142	static int hmR0SvmLeaveSession(PVMCPU pVCpu)
3143	{
3144	HM_DISABLE_PREEMPT();
3145	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
3146	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
3147
3148	/* When thread-context hooks are used, we can avoid doing the leave again if we had been preempted before
3149	and done this from the SVMR0ThreadCtxCallback(). */
3150	if (!pVCpu->hm.s.fLeaveDone)
3151	{
3152	hmR0SvmLeave(pVCpu, true /* fImportState */);
3153	pVCpu->hm.s.fLeaveDone = true;
3154	}
3155
3156	/*
3157	* !!! IMPORTANT !!!
3158	* If you modify code here, make sure to check whether hmR0SvmCallRing3Callback() needs to be updated too.
3159	*/
3160
3161	/** @todo eliminate the need for calling VMMR0ThreadCtxHookDisable here! */
3162	/* Deregister hook now that we've left HM context before re-enabling preemption. */
3163	VMMR0ThreadCtxHookDisable(pVCpu);
3164
3165	/* Leave HM context. This takes care of local init (term). */
3166	int rc = HMR0LeaveCpu(pVCpu);
3167
3168	HM_RESTORE_PREEMPT();
3169	return rc;
3170	}
3171
3172
3173	/**
3174	* Does the necessary state syncing before doing a longjmp to ring-3.
3175	*
3176	* @returns VBox status code.
3177	* @param pVCpu The cross context virtual CPU structure.
3178	*
3179	* @remarks No-long-jmp zone!!!
3180	*/
3181	static int hmR0SvmLongJmpToRing3(PVMCPU pVCpu)
3182	{
3183	return hmR0SvmLeaveSession(pVCpu);
3184	}
3185
3186
3187	/**
3188	* VMMRZCallRing3() callback wrapper which saves the guest state (or restores
3189	* any remaining host state) before we longjump to ring-3 and possibly get
3190	* preempted.
3191	*
3192	* @param pVCpu The cross context virtual CPU structure.
3193	* @param enmOperation The operation causing the ring-3 longjump.
3194	* @param pvUser The user argument (pointer to the possibly
3195	* out-of-date guest-CPU context).
3196	*/
3197	static DECLCALLBACK(int) hmR0SvmCallRing3Callback(PVMCPU pVCpu, VMMCALLRING3 enmOperation, void *pvUser)
3198	{
3199	RT_NOREF_PV(pvUser);
3200
3201	if (enmOperation == VMMCALLRING3_VM_R0_ASSERTION)
3202	{
3203	/*
3204	* !!! IMPORTANT !!!
3205	* If you modify code here, make sure to check whether hmR0SvmLeave() and hmR0SvmLeaveSession() needs
3206	* to be updated too. This is a stripped down version which gets out ASAP trying to not trigger any assertion.
3207	*/
3208	VMMRZCallRing3RemoveNotification(pVCpu);
3209	VMMRZCallRing3Disable(pVCpu);
3210	HM_DISABLE_PREEMPT();
3211
3212	/* Restore host FPU state if necessary and resync on next R0 reentry. */
3213	CPUMR0FpuStateMaybeSaveGuestAndRestoreHost(pVCpu);
3214
3215	/* Restore host debug registers if necessary and resync on next R0 reentry. */
3216	CPUMR0DebugStateMaybeSaveGuestAndRestoreHost(pVCpu, false /* save DR6 */);
3217
3218	/* Deregister the hook now that we've left HM context before re-enabling preemption. */
3219	/** @todo eliminate the need for calling VMMR0ThreadCtxHookDisable here! */
3220	VMMR0ThreadCtxHookDisable(pVCpu);
3221
3222	/* Leave HM context. This takes care of local init (term). */
3223	HMR0LeaveCpu(pVCpu);
3224
3225	HM_RESTORE_PREEMPT();
3226	return VINF_SUCCESS;
3227	}
3228
3229	Assert(pVCpu);
3230	Assert(pvUser);
3231	Assert(VMMRZCallRing3IsEnabled(pVCpu));
3232	HMSVM_ASSERT_PREEMPT_SAFE();
3233
3234	VMMRZCallRing3Disable(pVCpu);
3235	Assert(VMMR0IsLogFlushDisabled(pVCpu));
3236
3237	Log4Func(("Calling hmR0SvmLongJmpToRing3\n"));
3238	int rc = hmR0SvmLongJmpToRing3(pVCpu);
3239	AssertRCReturn(rc, rc);
3240
3241	VMMRZCallRing3Enable(pVCpu);
3242	return VINF_SUCCESS;
3243	}
3244
3245
3246	/**
3247	* Take necessary actions before going back to ring-3.
3248	*
3249	* An action requires us to go back to ring-3. This function does the necessary
3250	* steps before we can safely return to ring-3. This is not the same as longjmps
3251	* to ring-3, this is voluntary.
3252	*
3253	* @returns VBox status code.
3254	* @param pVCpu The cross context virtual CPU structure.
3255	* @param pCtx Pointer to the guest-CPU context.
3256	* @param rcExit The reason for exiting to ring-3. Can be
3257	* VINF_VMM_UNKNOWN_RING3_CALL.
3258	*/
3259	static int hmR0SvmExitToRing3(PVMCPU pVCpu, PCPUMCTX pCtx, int rcExit)
3260	{
3261	Assert(pVCpu);
3262	Assert(pCtx);
3263	HMSVM_ASSERT_PREEMPT_SAFE();
3264
3265	/* Please, no longjumps here (any logging shouldn't flush jump back to ring-3). NO LOGGING BEFORE THIS POINT! */
3266	VMMRZCallRing3Disable(pVCpu);
3267	Log4Func(("rcExit=%d LocalFF=%#RX32 GlobalFF=%#RX32\n", rcExit, pVCpu->fLocalForcedActions,
3268	pVCpu->CTX_SUFF(pVM)->fGlobalForcedActions));
3269
3270	/* We need to do this only while truly exiting the "inner loop" back to ring-3 and -not- for any longjmp to ring3. */
3271	if (pVCpu->hm.s.Event.fPending)
3272	{
3273	hmR0SvmPendingEventToTrpmTrap(pVCpu);
3274	Assert(!pVCpu->hm.s.Event.fPending);
3275	}
3276
3277	/* Sync. the necessary state for going back to ring-3. */
3278	hmR0SvmLeaveSession(pVCpu);
3279	STAM_COUNTER_DEC(&pVCpu->hm.s.StatSwitchLongJmpToR3);
3280
3281	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_TO_R3);
3282	CPUMSetChangedFlags(pVCpu, CPUM_CHANGED_SYSENTER_MSR
3283	\| CPUM_CHANGED_LDTR
3284	\| CPUM_CHANGED_GDTR
3285	\| CPUM_CHANGED_IDTR
3286	\| CPUM_CHANGED_TR
3287	\| CPUM_CHANGED_HIDDEN_SEL_REGS);
3288	if ( pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging
3289	&& CPUMIsGuestPagingEnabledEx(pCtx))
3290	{
3291	CPUMSetChangedFlags(pVCpu, CPUM_CHANGED_GLOBAL_TLB_FLUSH);
3292	}
3293
3294	/* Update the exit-to-ring 3 reason. */
3295	pVCpu->hm.s.rcLastExitToR3 = rcExit;
3296
3297	/* On our way back from ring-3 reload the guest state if there is a possibility of it being changed. */
3298	if (rcExit != VINF_EM_RAW_INTERRUPT)
3299	{
3300	Assert(!(pCtx->fExtrn & HMSVM_CPUMCTX_EXTRN_ALL));
3301	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
3302	}
3303
3304	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchExitToR3);
3305
3306	/* We do -not- want any longjmp notifications after this! We must return to ring-3 ASAP. */
3307	VMMRZCallRing3RemoveNotification(pVCpu);
3308	VMMRZCallRing3Enable(pVCpu);
3309
3310	/*
3311	* If we're emulating an instruction, we shouldn't have any TRPM traps pending
3312	* and if we're injecting an event we should have a TRPM trap pending.
3313	*/
3314	AssertReturnStmt(rcExit != VINF_EM_RAW_INJECT_TRPM_EVENT \|\| TRPMHasTrap(pVCpu),
3315	pVCpu->hm.s.u32HMError = rcExit,
3316	VERR_SVM_IPE_5);
3317	AssertReturnStmt(rcExit != VINF_EM_RAW_EMULATE_INSTR \|\| !TRPMHasTrap(pVCpu),
3318	pVCpu->hm.s.u32HMError = rcExit,
3319	VERR_SVM_IPE_4);
3320
3321	return rcExit;
3322	}
3323
3324
3325	/**
3326	* Updates the use of TSC offsetting mode for the CPU and adjusts the necessary
3327	* intercepts.
3328	*
3329	* @param pVCpu The cross context virtual CPU structure.
3330	* @param pCtx Pointer to the guest-CPU or nested-guest-CPU context.
3331	* @param pVmcb Pointer to the VM control block.
3332	*
3333	* @remarks No-long-jump zone!!!
3334	*/
3335	static void hmR0SvmUpdateTscOffsetting(PVMCPU pVCpu, PCCPUMCTX pCtx, PSVMVMCB pVmcb)
3336	{
3337	/*
3338	* Avoid intercepting RDTSC/RDTSCP if we determined the host TSC (++) is stable
3339	* and in case of a nested-guest, if the nested-VMCB specifies it is not intercepting
3340	* RDTSC/RDTSCP as well.
3341	*/
3342	bool fParavirtTsc;
3343	uint64_t uTscOffset;
3344	bool const fCanUseRealTsc = TMCpuTickCanUseRealTSC(pVCpu->CTX_SUFF(pVM), pVCpu, &uTscOffset, &fParavirtTsc);
3345
3346	bool fIntercept;
3347	if (fCanUseRealTsc)
3348	fIntercept = hmR0SvmClearCtrlIntercept(pVCpu, pCtx, pVmcb, SVM_CTRL_INTERCEPT_RDTSC \| SVM_CTRL_INTERCEPT_RDTSCP);
3349	else
3350	{
3351	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_RDTSC \| SVM_CTRL_INTERCEPT_RDTSCP);
3352	fIntercept = true;
3353	}
3354
3355	if (!fIntercept)
3356	{
3357	/* Apply the nested-guest VMCB's TSC offset over the guest TSC offset. */
3358	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
3359	uTscOffset = HMSvmNstGstApplyTscOffset(pVCpu, uTscOffset);
3360
3361	/* Update the TSC offset in the VMCB and the relevant clean bits. */
3362	pVmcb->ctrl.u64TSCOffset = uTscOffset;
3363	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
3364
3365	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscOffset);
3366	}
3367	else
3368	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscIntercept);
3369
3370	/* Currently neither Hyper-V nor KVM need to update their paravirt. TSC
3371	information before every VM-entry, hence we have nothing to do here at the moment. */
3372	if (fParavirtTsc)
3373	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscParavirt);
3374	}
3375
3376
3377	/**
3378	* Sets an event as a pending event to be injected into the guest.
3379	*
3380	* @param pVCpu The cross context virtual CPU structure.
3381	* @param pEvent Pointer to the SVM event.
3382	* @param GCPtrFaultAddress The fault-address (CR2) in case it's a
3383	* page-fault.
3384	*
3385	* @remarks Statistics counter assumes this is a guest event being reflected to
3386	* the guest i.e. 'StatInjectPendingReflect' is incremented always.
3387	*/
3388	DECLINLINE(void) hmR0SvmSetPendingEvent(PVMCPU pVCpu, PSVMEVENT pEvent, RTGCUINTPTR GCPtrFaultAddress)
3389	{
3390	Assert(!pVCpu->hm.s.Event.fPending);
3391	Assert(pEvent->n.u1Valid);
3392
3393	pVCpu->hm.s.Event.u64IntInfo = pEvent->u;
3394	pVCpu->hm.s.Event.fPending = true;
3395	pVCpu->hm.s.Event.GCPtrFaultAddress = GCPtrFaultAddress;
3396
3397	Log4Func(("u=%#RX64 u8Vector=%#x Type=%#x ErrorCodeValid=%RTbool ErrorCode=%#RX32\n", pEvent->u, pEvent->n.u8Vector,
3398	(uint8_t)pEvent->n.u3Type, !!pEvent->n.u1ErrorCodeValid, pEvent->n.u32ErrorCode));
3399	}
3400
3401
3402	/**
3403	* Sets an invalid-opcode (\#UD) exception as pending-for-injection into the VM.
3404	*
3405	* @param pVCpu The cross context virtual CPU structure.
3406	*/
3407	DECLINLINE(void) hmR0SvmSetPendingXcptUD(PVMCPU pVCpu)
3408	{
3409	SVMEVENT Event;
3410	Event.u = 0;
3411	Event.n.u1Valid = 1;
3412	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3413	Event.n.u8Vector = X86_XCPT_UD;
3414	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3415	}
3416
3417
3418	/**
3419	* Sets a debug (\#DB) exception as pending-for-injection into the VM.
3420	*
3421	* @param pVCpu The cross context virtual CPU structure.
3422	*/
3423	DECLINLINE(void) hmR0SvmSetPendingXcptDB(PVMCPU pVCpu)
3424	{
3425	SVMEVENT Event;
3426	Event.u = 0;
3427	Event.n.u1Valid = 1;
3428	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3429	Event.n.u8Vector = X86_XCPT_DB;
3430	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3431	}
3432
3433
3434	/**
3435	* Sets a page fault (\#PF) exception as pending-for-injection into the VM.
3436	*
3437	* @param pVCpu The cross context virtual CPU structure.
3438	* @param pCtx Pointer to the guest-CPU context.
3439	* @param u32ErrCode The error-code for the page-fault.
3440	* @param uFaultAddress The page fault address (CR2).
3441	*
3442	* @remarks This updates the guest CR2 with @a uFaultAddress!
3443	*/
3444	DECLINLINE(void) hmR0SvmSetPendingXcptPF(PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t u32ErrCode, RTGCUINTPTR uFaultAddress)
3445	{
3446	SVMEVENT Event;
3447	Event.u = 0;
3448	Event.n.u1Valid = 1;
3449	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3450	Event.n.u8Vector = X86_XCPT_PF;
3451	Event.n.u1ErrorCodeValid = 1;
3452	Event.n.u32ErrorCode = u32ErrCode;
3453
3454	/* Update CR2 of the guest. */
3455	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_CR2);
3456	if (pCtx->cr2 != uFaultAddress)
3457	{
3458	pCtx->cr2 = uFaultAddress;
3459	/* The VMCB clean bit for CR2 will be updated while re-loading the guest state. */
3460	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_CR2);
3461	}
3462
3463	hmR0SvmSetPendingEvent(pVCpu, &Event, uFaultAddress);
3464	}
3465
3466
3467	/**
3468	* Sets a math-fault (\#MF) exception as pending-for-injection into the VM.
3469	*
3470	* @param pVCpu The cross context virtual CPU structure.
3471	*/
3472	DECLINLINE(void) hmR0SvmSetPendingXcptMF(PVMCPU pVCpu)
3473	{
3474	SVMEVENT Event;
3475	Event.u = 0;
3476	Event.n.u1Valid = 1;
3477	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3478	Event.n.u8Vector = X86_XCPT_MF;
3479	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3480	}
3481
3482
3483	/**
3484	* Sets a double fault (\#DF) exception as pending-for-injection into the VM.
3485	*
3486	* @param pVCpu The cross context virtual CPU structure.
3487	*/
3488	DECLINLINE(void) hmR0SvmSetPendingXcptDF(PVMCPU pVCpu)
3489	{
3490	SVMEVENT Event;
3491	Event.u = 0;
3492	Event.n.u1Valid = 1;
3493	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3494	Event.n.u8Vector = X86_XCPT_DF;
3495	Event.n.u1ErrorCodeValid = 1;
3496	Event.n.u32ErrorCode = 0;
3497	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3498	}
3499
3500
3501	/**
3502	* Injects an event into the guest upon VMRUN by updating the relevant field
3503	* in the VMCB.
3504	*
3505	* @param pVCpu The cross context virtual CPU structure.
3506	* @param pVmcb Pointer to the guest VM control block.
3507	* @param pEvent Pointer to the event.
3508	*
3509	* @remarks No-long-jump zone!!!
3510	* @remarks Requires CR0!
3511	*/
3512	DECLINLINE(void) hmR0SvmInjectEventVmcb(PVMCPU pVCpu, PSVMVMCB pVmcb, PSVMEVENT pEvent)
3513	{
3514	Assert(!pVmcb->ctrl.EventInject.n.u1Valid);
3515	pVmcb->ctrl.EventInject.u = pEvent->u;
3516	STAM_COUNTER_INC(&pVCpu->hm.s.paStatInjectedIrqsR0[pEvent->n.u8Vector & MASK_INJECT_IRQ_STAT]);
3517	RT_NOREF(pVCpu);
3518
3519	Log4Func(("u=%#RX64 u8Vector=%#x Type=%#x ErrorCodeValid=%RTbool ErrorCode=%#RX32\n", pEvent->u, pEvent->n.u8Vector,
3520	(uint8_t)pEvent->n.u3Type, !!pEvent->n.u1ErrorCodeValid, pEvent->n.u32ErrorCode));
3521	}
3522
3523
3524
3525	/**
3526	* Converts any TRPM trap into a pending HM event. This is typically used when
3527	* entering from ring-3 (not longjmp returns).
3528	*
3529	* @param pVCpu The cross context virtual CPU structure.
3530	*/
3531	static void hmR0SvmTrpmTrapToPendingEvent(PVMCPU pVCpu)
3532	{
3533	Assert(TRPMHasTrap(pVCpu));
3534	Assert(!pVCpu->hm.s.Event.fPending);
3535
3536	uint8_t uVector;
3537	TRPMEVENT enmTrpmEvent;
3538	RTGCUINT uErrCode;
3539	RTGCUINTPTR GCPtrFaultAddress;
3540	uint8_t cbInstr;
3541
3542	int rc = TRPMQueryTrapAll(pVCpu, &uVector, &enmTrpmEvent, &uErrCode, &GCPtrFaultAddress, &cbInstr);
3543	AssertRC(rc);
3544
3545	SVMEVENT Event;
3546	Event.u = 0;
3547	Event.n.u1Valid = 1;
3548	Event.n.u8Vector = uVector;
3549
3550	/* Refer AMD spec. 15.20 "Event Injection" for the format. */
3551	if (enmTrpmEvent == TRPM_TRAP)
3552	{
3553	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3554	switch (uVector)
3555	{
3556	case X86_XCPT_NMI:
3557	{
3558	Event.n.u3Type = SVM_EVENT_NMI;
3559	break;
3560	}
3561
3562	case X86_XCPT_PF:
3563	case X86_XCPT_DF:
3564	case X86_XCPT_TS:
3565	case X86_XCPT_NP:
3566	case X86_XCPT_SS:
3567	case X86_XCPT_GP:
3568	case X86_XCPT_AC:
3569	{
3570	Event.n.u1ErrorCodeValid = 1;
3571	Event.n.u32ErrorCode = uErrCode;
3572	break;
3573	}
3574	}
3575	}
3576	else if (enmTrpmEvent == TRPM_HARDWARE_INT)
3577	Event.n.u3Type = SVM_EVENT_EXTERNAL_IRQ;
3578	else if (enmTrpmEvent == TRPM_SOFTWARE_INT)
3579	Event.n.u3Type = SVM_EVENT_SOFTWARE_INT;
3580	else
3581	AssertMsgFailed(("Invalid TRPM event type %d\n", enmTrpmEvent));
3582
3583	rc = TRPMResetTrap(pVCpu);
3584	AssertRC(rc);
3585
3586	Log4(("TRPM->HM event: u=%#RX64 u8Vector=%#x uErrorCodeValid=%RTbool uErrorCode=%#RX32\n", Event.u, Event.n.u8Vector,
3587	!!Event.n.u1ErrorCodeValid, Event.n.u32ErrorCode));
3588
3589	hmR0SvmSetPendingEvent(pVCpu, &Event, GCPtrFaultAddress);
3590	}
3591
3592
3593	/**
3594	* Converts any pending SVM event into a TRPM trap. Typically used when leaving
3595	* AMD-V to execute any instruction.
3596	*
3597	* @param pVCpu The cross context virtual CPU structure.
3598	*/
3599	static void hmR0SvmPendingEventToTrpmTrap(PVMCPU pVCpu)
3600	{
3601	Assert(pVCpu->hm.s.Event.fPending);
3602	Assert(TRPMQueryTrap(pVCpu, NULL /* pu8TrapNo /, NULL / pEnmType */) == VERR_TRPM_NO_ACTIVE_TRAP);
3603
3604	SVMEVENT Event;
3605	Event.u = pVCpu->hm.s.Event.u64IntInfo;
3606
3607	uint8_t uVector = Event.n.u8Vector;
3608	uint8_t uVectorType = Event.n.u3Type;
3609	TRPMEVENT enmTrapType = HMSvmEventToTrpmEventType(&Event);
3610
3611	Log4(("HM event->TRPM: uVector=%#x enmTrapType=%d\n", uVector, uVectorType));
3612
3613	int rc = TRPMAssertTrap(pVCpu, uVector, enmTrapType);
3614	AssertRC(rc);
3615
3616	if (Event.n.u1ErrorCodeValid)
3617	TRPMSetErrorCode(pVCpu, Event.n.u32ErrorCode);
3618
3619	if ( uVectorType == SVM_EVENT_EXCEPTION
3620	&& uVector == X86_XCPT_PF)
3621	{
3622	TRPMSetFaultAddress(pVCpu, pVCpu->hm.s.Event.GCPtrFaultAddress);
3623	Assert(pVCpu->hm.s.Event.GCPtrFaultAddress == CPUMGetGuestCR2(pVCpu));
3624	}
3625	else if (uVectorType == SVM_EVENT_SOFTWARE_INT)
3626	{
3627	AssertMsg( uVectorType == SVM_EVENT_SOFTWARE_INT
3628	\|\| (uVector == X86_XCPT_BP \|\| uVector == X86_XCPT_OF),
3629	("Invalid vector: uVector=%#x uVectorType=%#x\n", uVector, uVectorType));
3630	TRPMSetInstrLength(pVCpu, pVCpu->hm.s.Event.cbInstr);
3631	}
3632	pVCpu->hm.s.Event.fPending = false;
3633	}
3634
3635
3636	/**
3637	* Checks if the guest (or nested-guest) has an interrupt shadow active right
3638	* now.
3639	*
3640	* @returns @c true if the interrupt shadow is active, @c false otherwise.
3641	* @param pVCpu The cross context virtual CPU structure.
3642	* @param pCtx Pointer to the guest-CPU context.
3643	*
3644	* @remarks No-long-jump zone!!!
3645	* @remarks Has side-effects with VMCPU_FF_INHIBIT_INTERRUPTS force-flag.
3646	*/
3647	DECLINLINE(bool) hmR0SvmIsIntrShadowActive(PVMCPU pVCpu, PCCPUMCTX pCtx)
3648	{
3649	/*
3650	* Instructions like STI and MOV SS inhibit interrupts till the next instruction
3651	* completes. Check if we should inhibit interrupts or clear any existing
3652	* interrupt inhibition.
3653	*/
3654	if (VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS))
3655	{
3656	if (pCtx->rip != EMGetInhibitInterruptsPC(pVCpu))
3657	{
3658	/*
3659	* We can clear the inhibit force flag as even if we go back to the recompiler
3660	* without executing guest code in AMD-V, the flag's condition to be cleared is
3661	* met and thus the cleared state is correct.
3662	*/
3663	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS);
3664	return false;
3665	}
3666	return true;
3667	}
3668	return false;
3669	}
3670
3671
3672	/**
3673	* Sets the virtual interrupt intercept control in the VMCB.
3674	*
3675	* @param pVCpu The cross context virtual CPU structure.
3676	* @param pVmcb Pointer to the VM control block.
3677	* @param pCtx Pointer to the guest-CPU context.
3678	*/
3679	DECLINLINE(void) hmR0SvmSetIntWindowExiting(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
3680	{
3681	/*
3682	* When AVIC isn't supported, set up an interrupt window to cause a #VMEXIT when the guest
3683	* is ready to accept interrupts. At #VMEXIT, we then get the interrupt from the APIC
3684	* (updating ISR at the right time) and inject the interrupt.
3685	*
3686	* With AVIC is supported, we could make use of the asynchronously delivery without
3687	* #VMEXIT and we would be passing the AVIC page to SVM.
3688	*
3689	* In AMD-V, an interrupt window is achieved using a combination of V_IRQ (an interrupt
3690	* is pending), V_IGN_TPR (ignore TPR priorities) and the VINTR intercept all being set.
3691	*/
3692	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
3693	/*
3694	* Currently we don't overlay interupt windows and if there's any V_IRQ pending in the
3695	* nested-guest VMCB, we avoid setting up any interrupt window on behalf of the outer
3696	* guest.
3697	*/
3698	/** @todo Does this mean we end up prioritizing virtual interrupt
3699	* delivery/window over a physical interrupt (from the outer guest)
3700	* might be pending? */
3701	bool const fEnableIntWindow = !VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_NESTED_GUEST);
3702	if (!fEnableIntWindow)
3703	{
3704	Assert(CPUMIsGuestInSvmNestedHwVirtMode(pCtx)); RT_NOREF(pCtx);
3705	Log4(("Nested-guest V_IRQ already pending\n"));
3706	}
3707	#else
3708	RT_NOREF2(pVCpu, pCtx);
3709	bool const fEnableIntWindow = true;
3710	#endif
3711	if (fEnableIntWindow)
3712	{
3713	Assert(pVmcb->ctrl.IntCtrl.n.u1IgnoreTPR);
3714	pVmcb->ctrl.IntCtrl.n.u1VIrqPending = 1;
3715	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INT_CTRL;
3716	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_VINTR);
3717	Log4(("Set VINTR intercept\n"));
3718	}
3719	}
3720
3721
3722	/**
3723	* Clears the virtual interrupt intercept control in the VMCB as
3724	* we are figured the guest is unable process any interrupts
3725	* at this point of time.
3726	*
3727	* @param pVCpu The cross context virtual CPU structure.
3728	* @param pVmcb Pointer to the VM control block.
3729	* @param pCtx Pointer to the guest-CPU context.
3730	*/
3731	DECLINLINE(void) hmR0SvmClearIntWindowExiting(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
3732	{
3733	PSVMVMCBCTRL pVmcbCtrl = &pVmcb->ctrl;
3734	if ( pVmcbCtrl->IntCtrl.n.u1VIrqPending
3735	\|\| (pVmcbCtrl->u64InterceptCtrl & SVM_CTRL_INTERCEPT_VINTR))
3736	{
3737	pVmcbCtrl->IntCtrl.n.u1VIrqPending = 0;
3738	pVmcbCtrl->u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INT_CTRL;
3739	hmR0SvmClearCtrlIntercept(pVCpu, pCtx, pVmcb, SVM_CTRL_INTERCEPT_VINTR);
3740	Log4(("Cleared VINTR intercept\n"));
3741	}
3742	}
3743
3744	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
3745	/**
3746	* Evaluates the event to be delivered to the nested-guest and sets it as the
3747	* pending event.
3748	*
3749	* @returns VBox strict status code.
3750	* @param pVCpu The cross context virtual CPU structure.
3751	* @param pCtx Pointer to the guest-CPU context.
3752	*/
3753	static VBOXSTRICTRC hmR0SvmEvaluatePendingEventNested(PVMCPU pVCpu, PCPUMCTX pCtx)
3754	{
3755	HMSVM_ASSERT_IN_NESTED_GUEST(pCtx);
3756	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_HWVIRT
3757	\| CPUMCTX_EXTRN_RFLAGS
3758	\| CPUMCTX_EXTRN_HM_SVM_INT_SHADOW
3759	\| CPUMCTX_EXTRN_HM_SVM_HWVIRT_VIRQ);
3760
3761	Assert(!pVCpu->hm.s.Event.fPending);
3762	Assert(pCtx->hwvirt.fGif);
3763	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
3764	Assert(pVmcb);
3765
3766	bool const fVirtualGif = CPUMGetSvmNstGstVGif(pCtx);
3767	bool const fIntShadow = hmR0SvmIsIntrShadowActive(pVCpu, pCtx);
3768	bool const fBlockNmi = VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_BLOCK_NMIS);
3769
3770	Log4Func(("fVirtualGif=%RTbool fBlockNmi=%RTbool fIntShadow=%RTbool fIntPending=%RTbool fNmiPending=%RTbool\n",
3771	fVirtualGif, fBlockNmi, fIntShadow, VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC),
3772	VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_NMI)));
3773
3774	/** @todo SMI. SMIs take priority over NMIs. */
3775
3776	/*
3777	* Check if the guest can receive NMIs.
3778	* Nested NMIs are not allowed, see AMD spec. 8.1.4 "Masking External Interrupts".
3779	* NMIs take priority over maskable interrupts, see AMD spec. 8.5 "Priorities".
3780	*/
3781	if ( VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_NMI)
3782	&& !fBlockNmi)
3783	{
3784	if ( fVirtualGif
3785	&& !fIntShadow)
3786	{
3787	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_NMI))
3788	{
3789	Log4(("Intercepting NMI -> #VMEXIT\n"));
3790	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_SVM_VMEXIT_MASK);
3791	return IEMExecSvmVmexit(pVCpu, SVM_EXIT_NMI, 0, 0);
3792	}
3793
3794	Log4(("Setting NMI pending for injection\n"));
3795	SVMEVENT Event;
3796	Event.u = 0;
3797	Event.n.u1Valid = 1;
3798	Event.n.u8Vector = X86_XCPT_NMI;
3799	Event.n.u3Type = SVM_EVENT_NMI;
3800	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3801	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INTERRUPT_NMI);
3802	}
3803	else if (!fVirtualGif)
3804	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_STGI);
3805	else
3806	hmR0SvmSetIntWindowExiting(pVCpu, pVmcb, pCtx);
3807	}
3808	/*
3809	* Check if the nested-guest can receive external interrupts (generated by the guest's
3810	* PIC/APIC).
3811	*
3812	* External intercepts, NMI, SMI etc. from the physical CPU are -always- intercepted
3813	* when executing using hardware-assisted SVM, see HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS.
3814	*
3815	* External interrupts that are generated for the outer guest may be intercepted
3816	* depending on how the nested-guest VMCB was programmed by guest software.
3817	*
3818	* Physical interrupts always take priority over virtual interrupts,
3819	* see AMD spec. 15.21.4 "Injecting Virtual (INTR) Interrupts".
3820	*/
3821	else if ( VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC)
3822	&& !pVCpu->hm.s.fSingleInstruction)
3823	{
3824	if ( fVirtualGif
3825	&& !fIntShadow
3826	&& CPUMCanSvmNstGstTakePhysIntr(pVCpu, pCtx))
3827	{
3828	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_INTR))
3829	{
3830	Log4(("Intercepting INTR -> #VMEXIT\n"));
3831	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_SVM_VMEXIT_MASK);
3832	return IEMExecSvmVmexit(pVCpu, SVM_EXIT_INTR, 0, 0);
3833	}
3834
3835	uint8_t u8Interrupt;
3836	int rc = PDMGetInterrupt(pVCpu, &u8Interrupt);
3837	if (RT_SUCCESS(rc))
3838	{
3839	Log4(("Setting external interrupt %#x pending for injection\n", u8Interrupt));
3840	SVMEVENT Event;
3841	Event.u = 0;
3842	Event.n.u1Valid = 1;
3843	Event.n.u8Vector = u8Interrupt;
3844	Event.n.u3Type = SVM_EVENT_EXTERNAL_IRQ;
3845	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3846	}
3847	else if (rc == VERR_APIC_INTR_MASKED_BY_TPR)
3848	{
3849	/*
3850	* AMD-V has no TPR thresholding feature. TPR and the force-flag will be
3851	* updated eventually when the TPR is written by the guest.
3852	*/
3853	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchTprMaskedIrq);
3854	}
3855	else
3856	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchGuestIrq);
3857	}
3858	else if (!fVirtualGif)
3859	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_STGI);
3860	else
3861	hmR0SvmSetIntWindowExiting(pVCpu, pVmcb, pCtx);
3862	}
3863
3864	return VINF_SUCCESS;
3865	}
3866	#endif
3867
3868	/**
3869	* Evaluates the event to be delivered to the guest and sets it as the pending
3870	* event.
3871	*
3872	* @param pVCpu The cross context virtual CPU structure.
3873	* @param pCtx Pointer to the guest-CPU context.
3874	*/
3875	static void hmR0SvmEvaluatePendingEvent(PVMCPU pVCpu, PCPUMCTX pCtx)
3876	{
3877	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(pCtx);
3878	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_HWVIRT
3879	\| CPUMCTX_EXTRN_RFLAGS
3880	\| CPUMCTX_EXTRN_HM_SVM_INT_SHADOW);
3881
3882	Assert(!pVCpu->hm.s.Event.fPending);
3883	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
3884	Assert(pVmcb);
3885
3886	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
3887	bool const fGif = pCtx->hwvirt.fGif;
3888	#else
3889	bool const fGif = true;
3890	#endif
3891	bool const fIntShadow = hmR0SvmIsIntrShadowActive(pVCpu, pCtx);
3892	bool const fBlockInt = !(pCtx->eflags.u32 & X86_EFL_IF);
3893	bool const fBlockNmi = VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_BLOCK_NMIS);
3894
3895	Log4Func(("fGif=%RTbool fBlockNmi=%RTbool fBlockInt=%RTbool fIntShadow=%RTbool fIntPending=%RTbool NMI pending=%RTbool\n",
3896	fGif, fBlockNmi, fBlockInt, fIntShadow,
3897	VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC),
3898	VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_NMI)));
3899
3900	/** @todo SMI. SMIs take priority over NMIs. */
3901
3902	/*
3903	* Check if the guest can receive NMIs.
3904	* Nested NMIs are not allowed, see AMD spec. 8.1.4 "Masking External Interrupts".
3905	* NMIs take priority over maskable interrupts, see AMD spec. 8.5 "Priorities".
3906	*/
3907	if ( VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_NMI)
3908	&& !fBlockNmi)
3909	{
3910	if ( fGif
3911	&& !fIntShadow)
3912	{
3913	Log4(("Setting NMI pending for injection\n"));
3914	SVMEVENT Event;
3915	Event.u = 0;
3916	Event.n.u1Valid = 1;
3917	Event.n.u8Vector = X86_XCPT_NMI;
3918	Event.n.u3Type = SVM_EVENT_NMI;
3919	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3920	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INTERRUPT_NMI);
3921	}
3922	else if (!fGif)
3923	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_STGI);
3924	else
3925	hmR0SvmSetIntWindowExiting(pVCpu, pVmcb, pCtx);
3926	}
3927	/*
3928	* Check if the guest can receive external interrupts (PIC/APIC). Once PDMGetInterrupt()
3929	* returns a valid interrupt we -must- deliver the interrupt. We can no longer re-request
3930	* it from the APIC device.
3931	*/
3932	else if ( VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC)
3933	&& !pVCpu->hm.s.fSingleInstruction)
3934	{
3935	if ( fGif
3936	&& !fBlockInt
3937	&& !fIntShadow)
3938	{
3939	uint8_t u8Interrupt;
3940	int rc = PDMGetInterrupt(pVCpu, &u8Interrupt);
3941	if (RT_SUCCESS(rc))
3942	{
3943	Log4(("Setting external interrupt %#x pending for injection\n", u8Interrupt));
3944	SVMEVENT Event;
3945	Event.u = 0;
3946	Event.n.u1Valid = 1;
3947	Event.n.u8Vector = u8Interrupt;
3948	Event.n.u3Type = SVM_EVENT_EXTERNAL_IRQ;
3949	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3950	}
3951	else if (rc == VERR_APIC_INTR_MASKED_BY_TPR)
3952	{
3953	/*
3954	* AMD-V has no TPR thresholding feature. TPR and the force-flag will be
3955	* updated eventually when the TPR is written by the guest.
3956	*/
3957	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchTprMaskedIrq);
3958	}
3959	else
3960	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchGuestIrq);
3961	}
3962	else if (!fGif)
3963	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_STGI);
3964	else
3965	hmR0SvmSetIntWindowExiting(pVCpu, pVmcb, pCtx);
3966	}
3967	}
3968
3969
3970	/**
3971	* Injects any pending events into the guest (or nested-guest).
3972	*
3973	* @param pVCpu The cross context virtual CPU structure.
3974	* @param pCtx Pointer to the guest-CPU context.
3975	* @param pVmcb Pointer to the VM control block.
3976	*
3977	* @remarks Must only be called when we are guaranteed to enter
3978	* hardware-assisted SVM execution and not return to ring-3
3979	* prematurely.
3980	*/
3981	static void hmR0SvmInjectPendingEvent(PVMCPU pVCpu, PCCPUMCTX pCtx, PSVMVMCB pVmcb)
3982	{
3983	Assert(!TRPMHasTrap(pVCpu));
3984	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
3985
3986	bool const fIntShadow = hmR0SvmIsIntrShadowActive(pVCpu, pCtx);
3987	#ifdef VBOX_STRICT
3988	bool const fGif = pCtx->hwvirt.fGif;
3989	bool fAllowInt = fGif;
3990	if (fGif)
3991	{
3992	/*
3993	* For nested-guests we have no way to determine if we're injecting a physical or
3994	* virtual interrupt at this point. Hence the partial verification below.
3995	*/
3996	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
3997	fAllowInt = CPUMCanSvmNstGstTakePhysIntr(pVCpu, pCtx) \|\| CPUMCanSvmNstGstTakeVirtIntr(pVCpu, pCtx);
3998	else
3999	fAllowInt = RT_BOOL(pCtx->eflags.u32 & X86_EFL_IF);
4000	}
4001	#endif
4002
4003	if (pVCpu->hm.s.Event.fPending)
4004	{
4005	SVMEVENT Event;
4006	Event.u = pVCpu->hm.s.Event.u64IntInfo;
4007	Assert(Event.n.u1Valid);
4008
4009	/*
4010	* Validate event injection pre-conditions.
4011	*/
4012	if (Event.n.u3Type == SVM_EVENT_EXTERNAL_IRQ)
4013	{
4014	Assert(fAllowInt);
4015	Assert(!fIntShadow);
4016	}
4017	else if (Event.n.u3Type == SVM_EVENT_NMI)
4018	{
4019	Assert(fGif);
4020	Assert(!fIntShadow);
4021	}
4022
4023	/*
4024	* Before injecting an NMI we must set VMCPU_FF_BLOCK_NMIS to prevent nested NMIs. We
4025	* do this only when we are surely going to inject the NMI as otherwise if we return
4026	* to ring-3 prematurely we could leave NMIs blocked indefinitely upon re-entry into
4027	* SVM R0.
4028	*
4029	* With VT-x, this is handled by the Guest interruptibility information VMCS field
4030	* which will set the VMCS field after actually delivering the NMI which we read on
4031	* VM-exit to determine the state.
4032	*/
4033	if ( Event.n.u3Type == SVM_EVENT_NMI
4034	&& Event.n.u8Vector == X86_XCPT_NMI
4035	&& !VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_BLOCK_NMIS))
4036	{
4037	VMCPU_FF_SET(pVCpu, VMCPU_FF_BLOCK_NMIS);
4038	}
4039
4040	/*
4041	* Inject it (update VMCB for injection by the hardware).
4042	*/
4043	Log4(("Injecting pending HM event\n"));
4044	hmR0SvmInjectEventVmcb(pVCpu, pVmcb, &Event);
4045	pVCpu->hm.s.Event.fPending = false;
4046
4047	if (Event.n.u3Type == SVM_EVENT_EXTERNAL_IRQ)
4048	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectInterrupt);
4049	else
4050	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectXcpt);
4051	}
4052	else
4053	Assert(pVmcb->ctrl.EventInject.n.u1Valid == 0);
4054
4055	/*
4056	* We could have injected an NMI through IEM and continue guest execution using
4057	* hardware-assisted SVM. In which case, we would not have any events pending (above)
4058	* but we still need to intercept IRET in order to eventually clear NMI inhibition.
4059	*/
4060	if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_BLOCK_NMIS))
4061	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_IRET);
4062
4063	/*
4064	* Update the guest interrupt shadow in the guest (or nested-guest) VMCB.
4065	*
4066	* For nested-guests: We need to update it too for the scenario where IEM executes
4067	* the nested-guest but execution later continues here with an interrupt shadow active.
4068	*/
4069	pVmcb->ctrl.IntShadow.n.u1IntShadow = fIntShadow;
4070	}
4071
4072
4073	/**
4074	* Reports world-switch error and dumps some useful debug info.
4075	*
4076	* @param pVCpu The cross context virtual CPU structure.
4077	* @param rcVMRun The return code from VMRUN (or
4078	* VERR_SVM_INVALID_GUEST_STATE for invalid
4079	* guest-state).
4080	* @param pCtx Pointer to the guest-CPU context.
4081	*/
4082	static void hmR0SvmReportWorldSwitchError(PVMCPU pVCpu, int rcVMRun, PCPUMCTX pCtx)
4083	{
4084	HMSVM_ASSERT_PREEMPT_SAFE();
4085	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(pCtx);
4086	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
4087
4088	if (rcVMRun == VERR_SVM_INVALID_GUEST_STATE)
4089	{
4090	#ifdef VBOX_STRICT
4091	hmR0DumpRegs(pVCpu, pCtx);
4092	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
4093	Log4(("ctrl.u32VmcbCleanBits %#RX32\n", pVmcb->ctrl.u32VmcbCleanBits));
4094	Log4(("ctrl.u16InterceptRdCRx %#x\n", pVmcb->ctrl.u16InterceptRdCRx));
4095	Log4(("ctrl.u16InterceptWrCRx %#x\n", pVmcb->ctrl.u16InterceptWrCRx));
4096	Log4(("ctrl.u16InterceptRdDRx %#x\n", pVmcb->ctrl.u16InterceptRdDRx));
4097	Log4(("ctrl.u16InterceptWrDRx %#x\n", pVmcb->ctrl.u16InterceptWrDRx));
4098	Log4(("ctrl.u32InterceptXcpt %#x\n", pVmcb->ctrl.u32InterceptXcpt));
4099	Log4(("ctrl.u64InterceptCtrl %#RX64\n", pVmcb->ctrl.u64InterceptCtrl));
4100	Log4(("ctrl.u64IOPMPhysAddr %#RX64\n", pVmcb->ctrl.u64IOPMPhysAddr));
4101	Log4(("ctrl.u64MSRPMPhysAddr %#RX64\n", pVmcb->ctrl.u64MSRPMPhysAddr));
4102	Log4(("ctrl.u64TSCOffset %#RX64\n", pVmcb->ctrl.u64TSCOffset));
4103
4104	Log4(("ctrl.TLBCtrl.u32ASID %#x\n", pVmcb->ctrl.TLBCtrl.n.u32ASID));
4105	Log4(("ctrl.TLBCtrl.u8TLBFlush %#x\n", pVmcb->ctrl.TLBCtrl.n.u8TLBFlush));
4106	Log4(("ctrl.TLBCtrl.u24Reserved %#x\n", pVmcb->ctrl.TLBCtrl.n.u24Reserved));
4107
4108	Log4(("ctrl.IntCtrl.u8VTPR %#x\n", pVmcb->ctrl.IntCtrl.n.u8VTPR));
4109	Log4(("ctrl.IntCtrl.u1VIrqPending %#x\n", pVmcb->ctrl.IntCtrl.n.u1VIrqPending));
4110	Log4(("ctrl.IntCtrl.u1VGif %#x\n", pVmcb->ctrl.IntCtrl.n.u1VGif));
4111	Log4(("ctrl.IntCtrl.u6Reserved0 %#x\n", pVmcb->ctrl.IntCtrl.n.u6Reserved));
4112	Log4(("ctrl.IntCtrl.u4VIntrPrio %#x\n", pVmcb->ctrl.IntCtrl.n.u4VIntrPrio));
4113	Log4(("ctrl.IntCtrl.u1IgnoreTPR %#x\n", pVmcb->ctrl.IntCtrl.n.u1IgnoreTPR));
4114	Log4(("ctrl.IntCtrl.u3Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u3Reserved));
4115	Log4(("ctrl.IntCtrl.u1VIntrMasking %#x\n", pVmcb->ctrl.IntCtrl.n.u1VIntrMasking));
4116	Log4(("ctrl.IntCtrl.u1VGifEnable %#x\n", pVmcb->ctrl.IntCtrl.n.u1VGifEnable));
4117	Log4(("ctrl.IntCtrl.u5Reserved1 %#x\n", pVmcb->ctrl.IntCtrl.n.u5Reserved));
4118	Log4(("ctrl.IntCtrl.u8VIntrVector %#x\n", pVmcb->ctrl.IntCtrl.n.u8VIntrVector));
4119	Log4(("ctrl.IntCtrl.u24Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u24Reserved));
4120
4121	Log4(("ctrl.IntShadow.u1IntShadow %#x\n", pVmcb->ctrl.IntShadow.n.u1IntShadow));
4122	Log4(("ctrl.IntShadow.u1GuestIntMask %#x\n", pVmcb->ctrl.IntShadow.n.u1GuestIntMask));
4123	Log4(("ctrl.u64ExitCode %#RX64\n", pVmcb->ctrl.u64ExitCode));
4124	Log4(("ctrl.u64ExitInfo1 %#RX64\n", pVmcb->ctrl.u64ExitInfo1));
4125	Log4(("ctrl.u64ExitInfo2 %#RX64\n", pVmcb->ctrl.u64ExitInfo2));
4126	Log4(("ctrl.ExitIntInfo.u8Vector %#x\n", pVmcb->ctrl.ExitIntInfo.n.u8Vector));
4127	Log4(("ctrl.ExitIntInfo.u3Type %#x\n", pVmcb->ctrl.ExitIntInfo.n.u3Type));
4128	Log4(("ctrl.ExitIntInfo.u1ErrorCodeValid %#x\n", pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid));
4129	Log4(("ctrl.ExitIntInfo.u19Reserved %#x\n", pVmcb->ctrl.ExitIntInfo.n.u19Reserved));
4130	Log4(("ctrl.ExitIntInfo.u1Valid %#x\n", pVmcb->ctrl.ExitIntInfo.n.u1Valid));
4131	Log4(("ctrl.ExitIntInfo.u32ErrorCode %#x\n", pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode));
4132	Log4(("ctrl.NestedPagingCtrl.u1NestedPaging %#x\n", pVmcb->ctrl.NestedPagingCtrl.n.u1NestedPaging));
4133	Log4(("ctrl.NestedPagingCtrl.u1Sev %#x\n", pVmcb->ctrl.NestedPagingCtrl.n.u1Sev));
4134	Log4(("ctrl.NestedPagingCtrl.u1SevEs %#x\n", pVmcb->ctrl.NestedPagingCtrl.n.u1SevEs));
4135	Log4(("ctrl.EventInject.u8Vector %#x\n", pVmcb->ctrl.EventInject.n.u8Vector));
4136	Log4(("ctrl.EventInject.u3Type %#x\n", pVmcb->ctrl.EventInject.n.u3Type));
4137	Log4(("ctrl.EventInject.u1ErrorCodeValid %#x\n", pVmcb->ctrl.EventInject.n.u1ErrorCodeValid));
4138	Log4(("ctrl.EventInject.u19Reserved %#x\n", pVmcb->ctrl.EventInject.n.u19Reserved));
4139	Log4(("ctrl.EventInject.u1Valid %#x\n", pVmcb->ctrl.EventInject.n.u1Valid));
4140	Log4(("ctrl.EventInject.u32ErrorCode %#x\n", pVmcb->ctrl.EventInject.n.u32ErrorCode));
4141
4142	Log4(("ctrl.u64NestedPagingCR3 %#RX64\n", pVmcb->ctrl.u64NestedPagingCR3));
4143
4144	Log4(("ctrl.LbrVirt.u1LbrVirt %#x\n", pVmcb->ctrl.LbrVirt.n.u1LbrVirt));
4145	Log4(("ctrl.LbrVirt.u1VirtVmsaveVmload %#x\n", pVmcb->ctrl.LbrVirt.n.u1VirtVmsaveVmload));
4146
4147	Log4(("guest.CS.u16Sel %RTsel\n", pVmcb->guest.CS.u16Sel));
4148	Log4(("guest.CS.u16Attr %#x\n", pVmcb->guest.CS.u16Attr));
4149	Log4(("guest.CS.u32Limit %#RX32\n", pVmcb->guest.CS.u32Limit));
4150	Log4(("guest.CS.u64Base %#RX64\n", pVmcb->guest.CS.u64Base));
4151	Log4(("guest.DS.u16Sel %#RTsel\n", pVmcb->guest.DS.u16Sel));
4152	Log4(("guest.DS.u16Attr %#x\n", pVmcb->guest.DS.u16Attr));
4153	Log4(("guest.DS.u32Limit %#RX32\n", pVmcb->guest.DS.u32Limit));
4154	Log4(("guest.DS.u64Base %#RX64\n", pVmcb->guest.DS.u64Base));
4155	Log4(("guest.ES.u16Sel %RTsel\n", pVmcb->guest.ES.u16Sel));
4156	Log4(("guest.ES.u16Attr %#x\n", pVmcb->guest.ES.u16Attr));
4157	Log4(("guest.ES.u32Limit %#RX32\n", pVmcb->guest.ES.u32Limit));
4158	Log4(("guest.ES.u64Base %#RX64\n", pVmcb->guest.ES.u64Base));
4159	Log4(("guest.FS.u16Sel %RTsel\n", pVmcb->guest.FS.u16Sel));
4160	Log4(("guest.FS.u16Attr %#x\n", pVmcb->guest.FS.u16Attr));
4161	Log4(("guest.FS.u32Limit %#RX32\n", pVmcb->guest.FS.u32Limit));
4162	Log4(("guest.FS.u64Base %#RX64\n", pVmcb->guest.FS.u64Base));
4163	Log4(("guest.GS.u16Sel %RTsel\n", pVmcb->guest.GS.u16Sel));
4164	Log4(("guest.GS.u16Attr %#x\n", pVmcb->guest.GS.u16Attr));
4165	Log4(("guest.GS.u32Limit %#RX32\n", pVmcb->guest.GS.u32Limit));
4166	Log4(("guest.GS.u64Base %#RX64\n", pVmcb->guest.GS.u64Base));
4167
4168	Log4(("guest.GDTR.u32Limit %#RX32\n", pVmcb->guest.GDTR.u32Limit));
4169	Log4(("guest.GDTR.u64Base %#RX64\n", pVmcb->guest.GDTR.u64Base));
4170
4171	Log4(("guest.LDTR.u16Sel %RTsel\n", pVmcb->guest.LDTR.u16Sel));
4172	Log4(("guest.LDTR.u16Attr %#x\n", pVmcb->guest.LDTR.u16Attr));
4173	Log4(("guest.LDTR.u32Limit %#RX32\n", pVmcb->guest.LDTR.u32Limit));
4174	Log4(("guest.LDTR.u64Base %#RX64\n", pVmcb->guest.LDTR.u64Base));
4175
4176	Log4(("guest.IDTR.u32Limit %#RX32\n", pVmcb->guest.IDTR.u32Limit));
4177	Log4(("guest.IDTR.u64Base %#RX64\n", pVmcb->guest.IDTR.u64Base));
4178
4179	Log4(("guest.TR.u16Sel %RTsel\n", pVmcb->guest.TR.u16Sel));
4180	Log4(("guest.TR.u16Attr %#x\n", pVmcb->guest.TR.u16Attr));
4181	Log4(("guest.TR.u32Limit %#RX32\n", pVmcb->guest.TR.u32Limit));
4182	Log4(("guest.TR.u64Base %#RX64\n", pVmcb->guest.TR.u64Base));
4183
4184	Log4(("guest.u8CPL %#x\n", pVmcb->guest.u8CPL));
4185	Log4(("guest.u64CR0 %#RX64\n", pVmcb->guest.u64CR0));
4186	Log4(("guest.u64CR2 %#RX64\n", pVmcb->guest.u64CR2));
4187	Log4(("guest.u64CR3 %#RX64\n", pVmcb->guest.u64CR3));
4188	Log4(("guest.u64CR4 %#RX64\n", pVmcb->guest.u64CR4));
4189	Log4(("guest.u64DR6 %#RX64\n", pVmcb->guest.u64DR6));
4190	Log4(("guest.u64DR7 %#RX64\n", pVmcb->guest.u64DR7));
4191
4192	Log4(("guest.u64RIP %#RX64\n", pVmcb->guest.u64RIP));
4193	Log4(("guest.u64RSP %#RX64\n", pVmcb->guest.u64RSP));
4194	Log4(("guest.u64RAX %#RX64\n", pVmcb->guest.u64RAX));
4195	Log4(("guest.u64RFlags %#RX64\n", pVmcb->guest.u64RFlags));
4196
4197	Log4(("guest.u64SysEnterCS %#RX64\n", pVmcb->guest.u64SysEnterCS));
4198	Log4(("guest.u64SysEnterEIP %#RX64\n", pVmcb->guest.u64SysEnterEIP));
4199	Log4(("guest.u64SysEnterESP %#RX64\n", pVmcb->guest.u64SysEnterESP));
4200
4201	Log4(("guest.u64EFER %#RX64\n", pVmcb->guest.u64EFER));
4202	Log4(("guest.u64STAR %#RX64\n", pVmcb->guest.u64STAR));
4203	Log4(("guest.u64LSTAR %#RX64\n", pVmcb->guest.u64LSTAR));
4204	Log4(("guest.u64CSTAR %#RX64\n", pVmcb->guest.u64CSTAR));
4205	Log4(("guest.u64SFMASK %#RX64\n", pVmcb->guest.u64SFMASK));
4206	Log4(("guest.u64KernelGSBase %#RX64\n", pVmcb->guest.u64KernelGSBase));
4207	Log4(("guest.u64PAT %#RX64\n", pVmcb->guest.u64PAT));
4208	Log4(("guest.u64DBGCTL %#RX64\n", pVmcb->guest.u64DBGCTL));
4209	Log4(("guest.u64BR_FROM %#RX64\n", pVmcb->guest.u64BR_FROM));
4210	Log4(("guest.u64BR_TO %#RX64\n", pVmcb->guest.u64BR_TO));
4211	Log4(("guest.u64LASTEXCPFROM %#RX64\n", pVmcb->guest.u64LASTEXCPFROM));
4212	Log4(("guest.u64LASTEXCPTO %#RX64\n", pVmcb->guest.u64LASTEXCPTO));
4213
4214	NOREF(pVmcb);
4215	#endif /* VBOX_STRICT */
4216	}
4217	else
4218	Log4Func(("rcVMRun=%d\n", rcVMRun));
4219
4220	NOREF(pCtx);
4221	}
4222
4223
4224	/**
4225	* Check per-VM and per-VCPU force flag actions that require us to go back to
4226	* ring-3 for one reason or another.
4227	*
4228	* @returns VBox status code (information status code included).
4229	* @retval VINF_SUCCESS if we don't have any actions that require going back to
4230	* ring-3.
4231	* @retval VINF_PGM_SYNC_CR3 if we have pending PGM CR3 sync.
4232	* @retval VINF_EM_PENDING_REQUEST if we have pending requests (like hardware
4233	* interrupts)
4234	* @retval VINF_PGM_POOL_FLUSH_PENDING if PGM is doing a pool flush and requires
4235	* all EMTs to be in ring-3.
4236	* @retval VINF_EM_RAW_TO_R3 if there is pending DMA requests.
4237	* @retval VINF_EM_NO_MEMORY PGM is out of memory, we need to return
4238	* to the EM loop.
4239	*
4240	* @param pVCpu The cross context virtual CPU structure.
4241	* @param pCtx Pointer to the guest-CPU context.
4242	*/
4243	static int hmR0SvmCheckForceFlags(PVMCPU pVCpu, PCPUMCTX pCtx)
4244	{
4245	Assert(VMMRZCallRing3IsEnabled(pVCpu));
4246	Assert(!VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_UPDATE_PAE_PDPES));
4247
4248	/* Could happen as a result of longjump. */
4249	if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_UPDATE_CR3))
4250	PGMUpdateCR3(pVCpu, CPUMGetGuestCR3(pVCpu));
4251
4252	/* Update pending interrupts into the APIC's IRR. */
4253	if (VMCPU_FF_TEST_AND_CLEAR(pVCpu, VMCPU_FF_UPDATE_APIC))
4254	APICUpdatePendingInterrupts(pVCpu);
4255
4256	PVM pVM = pVCpu->CTX_SUFF(pVM);
4257	if ( VM_FF_IS_PENDING(pVM, !pVCpu->hm.s.fSingleInstruction
4258	? VM_FF_HP_R0_PRE_HM_MASK : VM_FF_HP_R0_PRE_HM_STEP_MASK)
4259	\|\| VMCPU_FF_IS_PENDING(pVCpu, !pVCpu->hm.s.fSingleInstruction
4260	? VMCPU_FF_HP_R0_PRE_HM_MASK : VMCPU_FF_HP_R0_PRE_HM_STEP_MASK) )
4261	{
4262	/* Pending PGM C3 sync. */
4263	if (VMCPU_FF_IS_PENDING(pVCpu,VMCPU_FF_PGM_SYNC_CR3 \| VMCPU_FF_PGM_SYNC_CR3_NON_GLOBAL))
4264	{
4265	int rc = PGMSyncCR3(pVCpu, pCtx->cr0, pCtx->cr3, pCtx->cr4, VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_PGM_SYNC_CR3));
4266	if (rc != VINF_SUCCESS)
4267	{
4268	Log4Func(("PGMSyncCR3 forcing us back to ring-3. rc=%d\n", rc));
4269	return rc;
4270	}
4271	}
4272
4273	/* Pending HM-to-R3 operations (critsects, timers, EMT rendezvous etc.) */
4274	/* -XXX- what was that about single stepping? */
4275	if ( VM_FF_IS_PENDING(pVM, VM_FF_HM_TO_R3_MASK)
4276	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_TO_R3_MASK))
4277	{
4278	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHmToR3FF);
4279	int rc = RT_UNLIKELY(VM_FF_IS_PENDING(pVM, VM_FF_PGM_NO_MEMORY)) ? VINF_EM_NO_MEMORY : VINF_EM_RAW_TO_R3;
4280	Log4Func(("HM_TO_R3 forcing us back to ring-3. rc=%d\n", rc));
4281	return rc;
4282	}
4283
4284	/* Pending VM request packets, such as hardware interrupts. */
4285	if ( VM_FF_IS_PENDING(pVM, VM_FF_REQUEST)
4286	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_REQUEST))
4287	{
4288	Log4Func(("Pending VM request forcing us back to ring-3\n"));
4289	return VINF_EM_PENDING_REQUEST;
4290	}
4291
4292	/* Pending PGM pool flushes. */
4293	if (VM_FF_IS_PENDING(pVM, VM_FF_PGM_POOL_FLUSH_PENDING))
4294	{
4295	Log4Func(("PGM pool flush pending forcing us back to ring-3\n"));
4296	return VINF_PGM_POOL_FLUSH_PENDING;
4297	}
4298
4299	/* Pending DMA requests. */
4300	if (VM_FF_IS_PENDING(pVM, VM_FF_PDM_DMA))
4301	{
4302	Log4Func(("Pending DMA request forcing us back to ring-3\n"));
4303	return VINF_EM_RAW_TO_R3;
4304	}
4305	}
4306
4307	return VINF_SUCCESS;
4308	}
4309
4310
4311	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
4312	/**
4313	* Does the preparations before executing nested-guest code in AMD-V.
4314	*
4315	* @returns VBox status code (informational status codes included).
4316	* @retval VINF_SUCCESS if we can proceed with running the guest.
4317	* @retval VINF_* scheduling changes, we have to go back to ring-3.
4318	*
4319	* @param pVCpu The cross context virtual CPU structure.
4320	* @param pCtx Pointer to the nested-guest-CPU context.
4321	* @param pSvmTransient Pointer to the SVM transient structure.
4322	*
4323	* @remarks Same caveats regarding longjumps as hmR0SvmPreRunGuest applies.
4324	* @sa hmR0SvmPreRunGuest.
4325	*/
4326	static int hmR0SvmPreRunGuestNested(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4327	{
4328	HMSVM_ASSERT_PREEMPT_SAFE();
4329	HMSVM_ASSERT_IN_NESTED_GUEST(pCtx);
4330
4331	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM_ONLY_IN_IEM
4332	Log2(("hmR0SvmPreRunGuest: Rescheduling to IEM due to nested-hwvirt or forced IEM exec -> VINF_EM_RESCHEDULE_REM\n"));
4333	return VINF_EM_RESCHEDULE_REM;
4334	#endif
4335
4336	/* Check force flag actions that might require us to go back to ring-3. */
4337	int rc = hmR0SvmCheckForceFlags(pVCpu, pCtx);
4338	if (rc != VINF_SUCCESS)
4339	return rc;
4340
4341	if (TRPMHasTrap(pVCpu))
4342	hmR0SvmTrpmTrapToPendingEvent(pVCpu);
4343	else if (!pVCpu->hm.s.Event.fPending)
4344	{
4345	VBOXSTRICTRC rcStrict = hmR0SvmEvaluatePendingEventNested(pVCpu, pCtx);
4346	if ( rcStrict != VINF_SUCCESS
4347	\|\| !CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
4348	return VBOXSTRICTRC_VAL(rcStrict);
4349	}
4350
4351	HMSVM_ASSERT_IN_NESTED_GUEST(pCtx);
4352
4353	/*
4354	* On the oldest AMD-V systems, we may not get enough information to reinject an NMI.
4355	* Just do it in software, see @bugref{8411}.
4356	* NB: If we could continue a task switch exit we wouldn't need to do this.
4357	*/
4358	PVM pVM = pVCpu->CTX_SUFF(pVM);
4359	if (RT_UNLIKELY( !pVM->hm.s.svm.u32Features
4360	&& pVCpu->hm.s.Event.fPending
4361	&& SVM_EVENT_GET_TYPE(pVCpu->hm.s.Event.u64IntInfo) == SVM_EVENT_NMI))
4362	{
4363	return VINF_EM_RAW_INJECT_TRPM_EVENT;
4364	}
4365
4366	#ifdef HMSVM_SYNC_FULL_NESTED_GUEST_STATE
4367	Assert(!(pCtx->fExtrn & HMSVM_CPUMCTX_EXTRN_ALL));
4368	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
4369	#endif
4370
4371	/*
4372	* Export the nested-guest state bits that are not shared with the host in any way as we
4373	* can longjmp or get preempted in the midst of exporting some of the state.
4374	*/
4375	rc = hmR0SvmExportGuestStateNested(pVCpu);
4376	AssertRCReturn(rc, rc);
4377	STAM_COUNTER_INC(&pVCpu->hm.s.StatExportFull);
4378
4379	/* Ensure we've cached (and hopefully modified) the VMCB for execution using hardware-assisted SVM. */
4380	Assert(pVCpu->hm.s.svm.NstGstVmcbCache.fCacheValid);
4381
4382	/*
4383	* No longjmps to ring-3 from this point on!!!
4384	*
4385	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional,
4386	* better than a kernel panic. This also disables flushing of the R0-logger instance.
4387	*/
4388	VMMRZCallRing3Disable(pVCpu);
4389
4390	/*
4391	* We disable interrupts so that we don't miss any interrupts that would flag preemption
4392	* (IPI/timers etc.) when thread-context hooks aren't used and we've been running with
4393	* preemption disabled for a while. Since this is purly to aid the
4394	* RTThreadPreemptIsPending() code, it doesn't matter that it may temporarily reenable and
4395	* disable interrupt on NT.
4396	*
4397	* We need to check for force-flags that could've possible been altered since we last
4398	* checked them (e.g. by PDMGetInterrupt() leaving the PDM critical section,
4399	* see @bugref{6398}).
4400	*
4401	* We also check a couple of other force-flags as a last opportunity to get the EMT back
4402	* to ring-3 before executing guest code.
4403	*/
4404	pSvmTransient->fEFlags = ASMIntDisableFlags();
4405	if ( VM_FF_IS_PENDING(pVM, VM_FF_EMT_RENDEZVOUS \| VM_FF_TM_VIRTUAL_SYNC)
4406	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_TO_R3_MASK))
4407	{
4408	ASMSetFlags(pSvmTransient->fEFlags);
4409	VMMRZCallRing3Enable(pVCpu);
4410	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHmToR3FF);
4411	return VINF_EM_RAW_TO_R3;
4412	}
4413	if (RTThreadPreemptIsPending(NIL_RTTHREAD))
4414	{
4415	ASMSetFlags(pSvmTransient->fEFlags);
4416	VMMRZCallRing3Enable(pVCpu);
4417	STAM_COUNTER_INC(&pVCpu->hm.s.StatPendingHostIrq);
4418	return VINF_EM_RAW_INTERRUPT;
4419	}
4420	return VINF_SUCCESS;
4421	}
4422	#endif
4423
4424
4425	/**
4426	* Does the preparations before executing guest code in AMD-V.
4427	*
4428	* This may cause longjmps to ring-3 and may even result in rescheduling to the
4429	* recompiler. We must be cautious what we do here regarding committing
4430	* guest-state information into the VMCB assuming we assuredly execute the guest
4431	* in AMD-V. If we fall back to the recompiler after updating the VMCB and
4432	* clearing the common-state (TRPM/forceflags), we must undo those changes so
4433	* that the recompiler can (and should) use them when it resumes guest
4434	* execution. Otherwise such operations must be done when we can no longer
4435	* exit to ring-3.
4436	*
4437	* @returns VBox status code (informational status codes included).
4438	* @retval VINF_SUCCESS if we can proceed with running the guest.
4439	* @retval VINF_* scheduling changes, we have to go back to ring-3.
4440	*
4441	* @param pVCpu The cross context virtual CPU structure.
4442	* @param pCtx Pointer to the guest-CPU context.
4443	* @param pSvmTransient Pointer to the SVM transient structure.
4444	*/
4445	static int hmR0SvmPreRunGuest(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4446	{
4447	HMSVM_ASSERT_PREEMPT_SAFE();
4448	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(pCtx);
4449
4450	/* Check force flag actions that might require us to go back to ring-3. */
4451	int rc = hmR0SvmCheckForceFlags(pVCpu, pCtx);
4452	if (rc != VINF_SUCCESS)
4453	return rc;
4454
4455	if (TRPMHasTrap(pVCpu))
4456	hmR0SvmTrpmTrapToPendingEvent(pVCpu);
4457	else if (!pVCpu->hm.s.Event.fPending)
4458	hmR0SvmEvaluatePendingEvent(pVCpu, pCtx);
4459
4460	/*
4461	* On the oldest AMD-V systems, we may not get enough information to reinject an NMI.
4462	* Just do it in software, see @bugref{8411}.
4463	* NB: If we could continue a task switch exit we wouldn't need to do this.
4464	*/
4465	PVM pVM = pVCpu->CTX_SUFF(pVM);
4466	if (RT_UNLIKELY(pVCpu->hm.s.Event.fPending && (((pVCpu->hm.s.Event.u64IntInfo >> 8) & 7) == SVM_EVENT_NMI)))
4467	if (RT_UNLIKELY(!pVM->hm.s.svm.u32Features))
4468	return VINF_EM_RAW_INJECT_TRPM_EVENT;
4469
4470	#ifdef HMSVM_SYNC_FULL_GUEST_STATE
4471	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
4472	{
4473	Assert(!(pCtx->fExtrn & HMSVM_CPUMCTX_EXTRN_ALL));
4474	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
4475	}
4476	#endif
4477	#ifdef HMSVM_SYNC_FULL_NESTED_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
4485	/*
4486	* Export the guest state bits that are not shared with the host in any way as we can
4487	* longjmp or get preempted in the midst of exporting some of the state.
4488	*/
4489	rc = hmR0SvmExportGuestState(pVCpu);
4490	AssertRCReturn(rc, rc);
4491	STAM_COUNTER_INC(&pVCpu->hm.s.StatExportFull);
4492
4493	/*
4494	* If we're not intercepting TPR changes in the guest, save the guest TPR before the
4495	* world-switch so we can update it on the way back if the guest changed the TPR.
4496	*/
4497	if (pVCpu->hm.s.svm.fSyncVTpr)
4498	{
4499	PCSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
4500	if (pVM->hm.s.fTPRPatchingActive)
4501	pSvmTransient->u8GuestTpr = pVmcb->guest.u64LSTAR;
4502	else
4503	pSvmTransient->u8GuestTpr = pVmcb->ctrl.IntCtrl.n.u8VTPR;
4504	}
4505
4506	/*
4507	* No longjmps to ring-3 from this point on!!!
4508	*
4509	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional,
4510	* better than a kernel panic. This also disables flushing of the R0-logger instance.
4511	*/
4512	VMMRZCallRing3Disable(pVCpu);
4513
4514	/*
4515	* We disable interrupts so that we don't miss any interrupts that would flag preemption
4516	* (IPI/timers etc.) when thread-context hooks aren't used and we've been running with
4517	* preemption disabled for a while. Since this is purly to aid the
4518	* RTThreadPreemptIsPending() code, it doesn't matter that it may temporarily reenable and
4519	* disable interrupt on NT.
4520	*
4521	* We need to check for force-flags that could've possible been altered since we last
4522	* checked them (e.g. by PDMGetInterrupt() leaving the PDM critical section,
4523	* see @bugref{6398}).
4524	*
4525	* We also check a couple of other force-flags as a last opportunity to get the EMT back
4526	* to ring-3 before executing guest code.
4527	*/
4528	pSvmTransient->fEFlags = ASMIntDisableFlags();
4529	if ( VM_FF_IS_PENDING(pVM, VM_FF_EMT_RENDEZVOUS \| VM_FF_TM_VIRTUAL_SYNC)
4530	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_TO_R3_MASK))
4531	{
4532	ASMSetFlags(pSvmTransient->fEFlags);
4533	VMMRZCallRing3Enable(pVCpu);
4534	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHmToR3FF);
4535	return VINF_EM_RAW_TO_R3;
4536	}
4537	if (RTThreadPreemptIsPending(NIL_RTTHREAD))
4538	{
4539	ASMSetFlags(pSvmTransient->fEFlags);
4540	VMMRZCallRing3Enable(pVCpu);
4541	STAM_COUNTER_INC(&pVCpu->hm.s.StatPendingHostIrq);
4542	return VINF_EM_RAW_INTERRUPT;
4543	}
4544
4545	return VINF_SUCCESS;
4546	}
4547
4548
4549	/**
4550	* Prepares to run guest (or nested-guest) code in AMD-V and we've committed to
4551	* doing so.
4552	*
4553	* This means there is no backing out to ring-3 or anywhere else at this point.
4554	*
4555	* @param pVCpu The cross context virtual CPU structure.
4556	* @param pCtx Pointer to the guest-CPU context.
4557	* @param pSvmTransient Pointer to the SVM transient structure.
4558	*
4559	* @remarks Called with preemption disabled.
4560	* @remarks No-long-jump zone!!!
4561	*/
4562	static void hmR0SvmPreRunGuestCommitted(PVMCPU pVCpu, PCCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4563	{
4564	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
4565	Assert(VMMR0IsLogFlushDisabled(pVCpu));
4566	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
4567
4568	VMCPU_ASSERT_STATE(pVCpu, VMCPUSTATE_STARTED_HM);
4569	VMCPU_SET_STATE(pVCpu, VMCPUSTATE_STARTED_EXEC); /* Indicate the start of guest execution. */
4570
4571	PVM pVM = pVCpu->CTX_SUFF(pVM);
4572	PSVMVMCB pVmcb = pSvmTransient->pVmcb;
4573
4574	hmR0SvmInjectPendingEvent(pVCpu, pCtx, pVmcb);
4575
4576	if (!CPUMIsGuestFPUStateActive(pVCpu))
4577	{
4578	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatLoadGuestFpuState, x);
4579	CPUMR0LoadGuestFPU(pVM, pVCpu); /* (Ignore rc, no need to set HM_CHANGED_HOST_CONTEXT for SVM.) */
4580	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatLoadGuestFpuState, x);
4581	STAM_COUNTER_INC(&pVCpu->hm.s.StatLoadGuestFpu);
4582	}
4583
4584	/* Load the state shared between host and guest (FPU, debug). */
4585	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE)
4586	hmR0SvmExportSharedState(pVCpu, pVmcb);
4587
4588	pVCpu->hm.s.fCtxChanged &= ~HM_CHANGED_HOST_CONTEXT; /* Preemption might set this, nothing to do on AMD-V. */
4589	AssertMsg(!pVCpu->hm.s.fCtxChanged, ("fCtxChanged=%#RX64\n", pVCpu->hm.s.fCtxChanged));
4590
4591	PHMGLOBALCPUINFO pHostCpu = hmR0GetCurrentCpu();
4592	RTCPUID const idHostCpu = pHostCpu->idCpu;
4593	bool const fMigratedHostCpu = idHostCpu != pVCpu->hm.s.idLastCpu;
4594
4595	/* Setup TSC offsetting. */
4596	if ( pSvmTransient->fUpdateTscOffsetting
4597	\|\| fMigratedHostCpu)
4598	{
4599	hmR0SvmUpdateTscOffsetting(pVCpu, pCtx, pVmcb);
4600	pSvmTransient->fUpdateTscOffsetting = false;
4601	}
4602
4603	/* If we've migrating CPUs, mark the VMCB Clean bits as dirty. */
4604	if (fMigratedHostCpu)
4605	pVmcb->ctrl.u32VmcbCleanBits = 0;
4606
4607	/* Store status of the shared guest-host state at the time of VMRUN. */
4608	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS)
4609	if (CPUMIsGuestInLongModeEx(pCtx))
4610	{
4611	pSvmTransient->fWasGuestDebugStateActive = CPUMIsGuestDebugStateActivePending(pVCpu);
4612	pSvmTransient->fWasHyperDebugStateActive = CPUMIsHyperDebugStateActivePending(pVCpu);
4613	}
4614	else
4615	#endif
4616	{
4617	pSvmTransient->fWasGuestDebugStateActive = CPUMIsGuestDebugStateActive(pVCpu);
4618	pSvmTransient->fWasHyperDebugStateActive = CPUMIsHyperDebugStateActive(pVCpu);
4619	}
4620
4621	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
4622	uint8_t *pbMsrBitmap;
4623	if (!pSvmTransient->fIsNestedGuest)
4624	pbMsrBitmap = (uint8_t *)pVCpu->hm.s.svm.pvMsrBitmap;
4625	else
4626	{
4627	hmR0SvmMergeMsrpmNested(pHostCpu, pVCpu, pCtx);
4628
4629	/* Update the nested-guest VMCB with the newly merged MSRPM (clean bits updated below). */
4630	pVmcb->ctrl.u64MSRPMPhysAddr = pHostCpu->n.svm.HCPhysNstGstMsrpm;
4631	pbMsrBitmap = (uint8_t *)pHostCpu->n.svm.pvNstGstMsrpm;
4632	}
4633	#else
4634	uint8_t pbMsrBitmap = (uint8_t )pVCpu->hm.s.svm.pvMsrBitmap;
4635	#endif
4636
4637	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, true); /* Used for TLB flushing, set this across the world switch. */
4638	/* Flush the appropriate tagged-TLB entries. */
4639	hmR0SvmFlushTaggedTlb(pVCpu, pCtx, pVmcb, pHostCpu);
4640	Assert(pVCpu->hm.s.idLastCpu == idHostCpu);
4641
4642	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatEntry, &pVCpu->hm.s.StatInGC, x);
4643
4644	TMNotifyStartOfExecution(pVCpu); /* Finally, notify TM to resume its clocks as we're about
4645	to start executing. */
4646
4647	/*
4648	* Save the current Host TSC_AUX and write the guest TSC_AUX to the host, so that RDTSCPs
4649	* (that don't cause exits) reads the guest MSR, see @bugref{3324}.
4650	*
4651	* This should be done -after- any RDTSCPs for obtaining the host timestamp (TM, STAM etc).
4652	*/
4653	if ( (pVM->hm.s.cpuid.u32AMDFeatureEDX & X86_CPUID_EXT_FEATURE_EDX_RDTSCP)
4654	&& !(pVmcb->ctrl.u64InterceptCtrl & SVM_CTRL_INTERCEPT_RDTSCP))
4655	{
4656	uint64_t const uGuestTscAux = CPUMGetGuestTscAux(pVCpu);
4657	pVCpu->hm.s.u64HostTscAux = ASMRdMsr(MSR_K8_TSC_AUX);
4658	if (uGuestTscAux != pVCpu->hm.s.u64HostTscAux)
4659	ASMWrMsr(MSR_K8_TSC_AUX, uGuestTscAux);
4660	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_K8_TSC_AUX, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
4661	pSvmTransient->fRestoreTscAuxMsr = true;
4662	}
4663	else
4664	{
4665	hmR0SvmSetMsrPermission(pCtx, pbMsrBitmap, MSR_K8_TSC_AUX, SVMMSREXIT_INTERCEPT_READ, SVMMSREXIT_INTERCEPT_WRITE);
4666	pSvmTransient->fRestoreTscAuxMsr = false;
4667	}
4668	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_IOPM_MSRPM;
4669
4670	/*
4671	* If VMCB Clean bits isn't supported by the CPU or exposed to the guest in the nested
4672	* virtualization case, mark all state-bits as dirty indicating to the CPU to re-load
4673	* from the VMCB.
4674	*/
4675	bool const fSupportsVmcbCleanBits = hmR0SvmSupportsVmcbCleanBits(pVCpu, pCtx);
4676	if (!fSupportsVmcbCleanBits)
4677	pVmcb->ctrl.u32VmcbCleanBits = 0;
4678	}
4679
4680
4681	/**
4682	* Wrapper for running the guest (or nested-guest) code in AMD-V.
4683	*
4684	* @returns VBox strict status code.
4685	* @param pVCpu The cross context virtual CPU structure.
4686	* @param pCtx Pointer to the guest-CPU context.
4687	* @param HCPhysVmcb The host physical address of the VMCB.
4688	*
4689	* @remarks No-long-jump zone!!!
4690	*/
4691	DECLINLINE(int) hmR0SvmRunGuest(PVMCPU pVCpu, PCPUMCTX pCtx, RTHCPHYS HCPhysVmcb)
4692	{
4693	/* Mark that HM is the keeper of all guest-CPU registers now that we're going to execute guest code. */
4694	pCtx->fExtrn \|= HMSVM_CPUMCTX_EXTRN_ALL \| CPUMCTX_EXTRN_KEEPER_HM;
4695
4696	/*
4697	* 64-bit Windows uses XMM registers in the kernel as the Microsoft compiler expresses
4698	* floating-point operations using SSE instructions. Some XMM registers (XMM6-XMM15) are
4699	* callee-saved and thus the need for this XMM wrapper.
4700	*
4701	* Refer MSDN "Configuring Programs for 64-bit/x64 Software Conventions / Register Usage".
4702	*/
4703	PVM pVM = pVCpu->CTX_SUFF(pVM);
4704	#ifdef VBOX_WITH_KERNEL_USING_XMM
4705	return hmR0SVMRunWrapXMM(pVCpu->hm.s.svm.HCPhysVmcbHost, HCPhysVmcb, pCtx, pVM, pVCpu, pVCpu->hm.s.svm.pfnVMRun);
4706	#else
4707	return pVCpu->hm.s.svm.pfnVMRun(pVCpu->hm.s.svm.HCPhysVmcbHost, HCPhysVmcb, pCtx, pVM, pVCpu);
4708	#endif
4709	}
4710
4711
4712	/**
4713	* Undoes the TSC offset applied for an SVM nested-guest and returns the TSC
4714	* value for the guest.
4715	*
4716	* @returns The TSC offset after undoing any nested-guest TSC offset.
4717	* @param pVCpu The cross context virtual CPU structure of the calling EMT.
4718	* @param uTicks The nested-guest TSC.
4719	*
4720	* @note If you make any changes to this function, please check if
4721	* hmR0SvmNstGstUndoTscOffset() needs adjusting.
4722	*
4723	* @sa HMSvmNstGstApplyTscOffset().
4724	*/
4725	DECLINLINE(uint64_t) hmR0SvmNstGstUndoTscOffset(PVMCPU pVCpu, uint64_t uTicks)
4726	{
4727	PCSVMNESTEDVMCBCACHE pVmcbNstGstCache = &pVCpu->hm.s.svm.NstGstVmcbCache;
4728	Assert(pVmcbNstGstCache->fCacheValid);
4729	return uTicks - pVmcbNstGstCache->u64TSCOffset;
4730	}
4731
4732
4733	/**
4734	* Performs some essential restoration of state after running guest (or
4735	* nested-guest) code in AMD-V.
4736	*
4737	* @param pVCpu The cross context virtual CPU structure.
4738	* @param pCtx Pointer to the guest-CPU context. The data maybe
4739	* out-of-sync. Make sure to update the required fields
4740	* before using them.
4741	* @param pSvmTransient Pointer to the SVM transient structure.
4742	* @param rcVMRun Return code of VMRUN.
4743	*
4744	* @remarks Called with interrupts disabled.
4745	* @remarks No-long-jump zone!!! This function will however re-enable longjmps
4746	* unconditionally when it is safe to do so.
4747	*/
4748	static void hmR0SvmPostRunGuest(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient, int rcVMRun)
4749	{
4750	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
4751
4752	uint64_t const uHostTsc = ASMReadTSC(); /* Read the TSC as soon as possible. */
4753	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, false); /* See HMInvalidatePageOnAllVCpus(): used for TLB flushing. */
4754	ASMAtomicIncU32(&pVCpu->hm.s.cWorldSwitchExits); /* Initialized in vmR3CreateUVM(): used for EMT poking. */
4755
4756	PSVMVMCB pVmcb = pSvmTransient->pVmcb;
4757	PSVMVMCBCTRL pVmcbCtrl = &pVmcb->ctrl;
4758
4759	/* TSC read must be done early for maximum accuracy. */
4760	if (!(pVmcbCtrl->u64InterceptCtrl & SVM_CTRL_INTERCEPT_RDTSC))
4761	{
4762	if (!pSvmTransient->fIsNestedGuest)
4763	TMCpuTickSetLastSeen(pVCpu, uHostTsc + pVmcbCtrl->u64TSCOffset);
4764	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
4765	else
4766	{
4767	/* The nested-guest VMCB TSC offset shall eventually be restored on #VMEXIT via HMSvmNstGstVmExitNotify(). */
4768	uint64_t const uGstTsc = hmR0SvmNstGstUndoTscOffset(pVCpu, uHostTsc + pVmcbCtrl->u64TSCOffset);
4769	TMCpuTickSetLastSeen(pVCpu, uGstTsc);
4770	}
4771	#endif
4772	}
4773
4774	if (pSvmTransient->fRestoreTscAuxMsr)
4775	{
4776	uint64_t u64GuestTscAuxMsr = ASMRdMsr(MSR_K8_TSC_AUX);
4777	CPUMSetGuestTscAux(pVCpu, u64GuestTscAuxMsr);
4778	if (u64GuestTscAuxMsr != pVCpu->hm.s.u64HostTscAux)
4779	ASMWrMsr(MSR_K8_TSC_AUX, pVCpu->hm.s.u64HostTscAux);
4780	}
4781
4782	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatInGC, &pVCpu->hm.s.StatPreExit, x);
4783	TMNotifyEndOfExecution(pVCpu); /* Notify TM that the guest is no longer running. */
4784	VMCPU_SET_STATE(pVCpu, VMCPUSTATE_STARTED_HM);
4785
4786	Assert(!(ASMGetFlags() & X86_EFL_IF));
4787	ASMSetFlags(pSvmTransient->fEFlags); /* Enable interrupts. */
4788	VMMRZCallRing3Enable(pVCpu); /* It is now safe to do longjmps to ring-3!!! */
4789
4790	/* If VMRUN failed, we can bail out early. This does -not- cover SVM_EXIT_INVALID. */
4791	if (RT_UNLIKELY(rcVMRun != VINF_SUCCESS))
4792	{
4793	Log4Func(("VMRUN failure: rcVMRun=%Rrc\n", rcVMRun));
4794	return;
4795	}
4796
4797	pSvmTransient->u64ExitCode = pVmcbCtrl->u64ExitCode; /* Save the #VMEXIT reason. */
4798	pVmcbCtrl->u32VmcbCleanBits = HMSVM_VMCB_CLEAN_ALL; /* Mark the VMCB-state cache as unmodified by VMM. */
4799	pSvmTransient->fVectoringDoublePF = false; /* Vectoring double page-fault needs to be determined later. */
4800	pSvmTransient->fVectoringPF = false; /* Vectoring page-fault needs to be determined later. */
4801
4802	#ifdef HMSVM_SYNC_FULL_GUEST_STATE
4803	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
4804	hmR0SvmImportGuestState(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
4805	#elif defined(HMSVM_SYNC_FULL_NESTED_GUEST_STATE)
4806	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
4807	hmR0SvmImportGuestState(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
4808	#else
4809	/*
4810	* Always import the following:
4811	*
4812	* - RIP for exit optimizations and evaluating event injection on re-entry.
4813	* - RFLAGS for evaluating event injection on VM re-entry and for exporting shared debug
4814	* state on preemption.
4815	* - Interrupt shadow, GIF for evaluating event injection on VM re-entry.
4816	* - CS for exit optimizations.
4817	* - RAX, RSP for simplifying assumptions on GPRs. All other GPRs are swapped by the
4818	* assembly switcher code.
4819	* - Shared state (only DR7 currently) for exporting shared debug state on preemption.
4820	*/
4821	hmR0SvmImportGuestState(pVCpu, CPUMCTX_EXTRN_RIP
4822	\| CPUMCTX_EXTRN_RFLAGS
4823	\| CPUMCTX_EXTRN_RAX
4824	\| CPUMCTX_EXTRN_RSP
4825	\| CPUMCTX_EXTRN_CS
4826	\| CPUMCTX_EXTRN_HWVIRT
4827	\| CPUMCTX_EXTRN_HM_SVM_INT_SHADOW
4828	\| CPUMCTX_EXTRN_HM_SVM_HWVIRT_VIRQ
4829	\| HMSVM_CPUMCTX_SHARED_STATE);
4830	#endif
4831
4832	#ifdef DEBUG_ramshankar
4833	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
4834	{
4835	hmR0SvmImportGuestState(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
4836	hmR0SvmLogState(pVCpu, pVmcb, pCtx, "hmR0SvmPostRunGuestNested", HMSVM_LOG_ALL & ~HMSVM_LOG_LBR, 0 /* uVerbose */);
4837	}
4838	#endif
4839
4840	if ( pSvmTransient->u64ExitCode != SVM_EXIT_INVALID
4841	&& pVCpu->hm.s.svm.fSyncVTpr)
4842	{
4843	Assert(!pSvmTransient->fIsNestedGuest);
4844	/* TPR patching (for 32-bit guests) uses LSTAR MSR for holding the TPR value, otherwise uses the VTPR. */
4845	if ( pVCpu->CTX_SUFF(pVM)->hm.s.fTPRPatchingActive
4846	&& (pVmcb->guest.u64LSTAR & 0xff) != pSvmTransient->u8GuestTpr)
4847	{
4848	int rc = APICSetTpr(pVCpu, pVmcb->guest.u64LSTAR & 0xff);
4849	AssertRC(rc);
4850	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
4851	}
4852	/* Sync TPR when we aren't intercepting CR8 writes. */
4853	else if (pSvmTransient->u8GuestTpr != pVmcbCtrl->IntCtrl.n.u8VTPR)
4854	{
4855	int rc = APICSetTpr(pVCpu, pVmcbCtrl->IntCtrl.n.u8VTPR << 4);
4856	AssertRC(rc);
4857	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
4858	}
4859	}
4860
4861	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_CS \| CPUMCTX_EXTRN_RIP);
4862	EMHistoryAddExit(pVCpu, EMEXIT_MAKE_FT(EMEXIT_F_KIND_SVM, pSvmTransient->u64ExitCode & EMEXIT_F_TYPE_MASK),
4863	pCtx->cs.u64Base + pCtx->rip, uHostTsc);
4864	}
4865
4866
4867	/**
4868	* Runs the guest code using AMD-V.
4869	*
4870	* @returns VBox status code.
4871	* @param pVCpu The cross context virtual CPU structure.
4872	* @param pCtx Pointer to the guest-CPU context.
4873	* @param pcLoops Pointer to the number of executed loops.
4874	*/
4875	static int hmR0SvmRunGuestCodeNormal(PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t *pcLoops)
4876	{
4877	uint32_t const cMaxResumeLoops = pVCpu->CTX_SUFF(pVM)->hm.s.cMaxResumeLoops;
4878	Assert(pcLoops);
4879	Assert(*pcLoops <= cMaxResumeLoops);
4880
4881	SVMTRANSIENT SvmTransient;
4882	RT_ZERO(SvmTransient);
4883	SvmTransient.fUpdateTscOffsetting = true;
4884	SvmTransient.pVmcb = pVCpu->hm.s.svm.pVmcb;
4885
4886	int rc = VERR_INTERNAL_ERROR_5;
4887	for (;;)
4888	{
4889	Assert(!HMR0SuspendPending());
4890	HMSVM_ASSERT_CPU_SAFE();
4891
4892	/* Preparatory work for running nested-guest code, this may force us to return to
4893	ring-3. This bugger disables interrupts on VINF_SUCCESS! */
4894	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatEntry, x);
4895	rc = hmR0SvmPreRunGuest(pVCpu, pCtx, &SvmTransient);
4896	if (rc != VINF_SUCCESS)
4897	break;
4898
4899	/*
4900	* No longjmps to ring-3 from this point on!!!
4901	*
4902	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional,
4903	* better than a kernel panic. This also disables flushing of the R0-logger instance.
4904	*/
4905	hmR0SvmPreRunGuestCommitted(pVCpu, pCtx, &SvmTransient);
4906	rc = hmR0SvmRunGuest(pVCpu, pCtx, pVCpu->hm.s.svm.HCPhysVmcb);
4907
4908	/* Restore any residual host-state and save any bits shared between host and guest
4909	into the guest-CPU state. Re-enables interrupts! */
4910	hmR0SvmPostRunGuest(pVCpu, pCtx, &SvmTransient, rc);
4911
4912	if (RT_UNLIKELY( rc != VINF_SUCCESS /* Check for VMRUN errors. */
4913	\|\| SvmTransient.u64ExitCode == SVM_EXIT_INVALID)) /* Check for invalid guest-state errors. */
4914	{
4915	if (rc == VINF_SUCCESS)
4916	rc = VERR_SVM_INVALID_GUEST_STATE;
4917	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatPreExit, x);
4918	hmR0SvmReportWorldSwitchError(pVCpu, rc, pCtx);
4919	break;
4920	}
4921
4922	/* Handle the #VMEXIT. */
4923	HMSVM_EXITCODE_STAM_COUNTER_INC(SvmTransient.u64ExitCode);
4924	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatPreExit, &pVCpu->hm.s.StatExitHandling, x);
4925	VBOXVMM_R0_HMSVM_VMEXIT(pVCpu, pCtx, SvmTransient.u64ExitCode, pVCpu->hm.s.svm.pVmcb);
4926	rc = hmR0SvmHandleExit(pVCpu, pCtx, &SvmTransient);
4927	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExitHandling, x);
4928	if (rc != VINF_SUCCESS)
4929	break;
4930	if (++(*pcLoops) >= cMaxResumeLoops)
4931	{
4932	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchMaxResumeLoops);
4933	rc = VINF_EM_RAW_INTERRUPT;
4934	break;
4935	}
4936	}
4937
4938	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatEntry, x);
4939	return rc;
4940	}
4941
4942
4943	/**
4944	* Runs the guest code using AMD-V in single step mode.
4945	*
4946	* @returns VBox status code.
4947	* @param pVCpu The cross context virtual CPU structure.
4948	* @param pCtx Pointer to the guest-CPU context.
4949	* @param pcLoops Pointer to the number of executed loops.
4950	*/
4951	static int hmR0SvmRunGuestCodeStep(PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t *pcLoops)
4952	{
4953	uint32_t const cMaxResumeLoops = pVCpu->CTX_SUFF(pVM)->hm.s.cMaxResumeLoops;
4954	Assert(pcLoops);
4955	Assert(*pcLoops <= cMaxResumeLoops);
4956
4957	SVMTRANSIENT SvmTransient;
4958	RT_ZERO(SvmTransient);
4959	SvmTransient.fUpdateTscOffsetting = true;
4960	SvmTransient.pVmcb = pVCpu->hm.s.svm.pVmcb;
4961
4962	uint16_t uCsStart = pCtx->cs.Sel;
4963	uint64_t uRipStart = pCtx->rip;
4964
4965	int rc = VERR_INTERNAL_ERROR_5;
4966	for (;;)
4967	{
4968	Assert(!HMR0SuspendPending());
4969	AssertMsg(pVCpu->hm.s.idEnteredCpu == RTMpCpuId(),
4970	("Illegal migration! Entered on CPU %u Current %u cLoops=%u\n", (unsigned)pVCpu->hm.s.idEnteredCpu,
4971	(unsigned)RTMpCpuId(), *pcLoops));
4972
4973	/* Preparatory work for running nested-guest code, this may force us to return to
4974	ring-3. This bugger disables interrupts on VINF_SUCCESS! */
4975	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatEntry, x);
4976	rc = hmR0SvmPreRunGuest(pVCpu, pCtx, &SvmTransient);
4977	if (rc != VINF_SUCCESS)
4978	break;
4979
4980	/*
4981	* No longjmps to ring-3 from this point on!!!
4982	*
4983	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional,
4984	* better than a kernel panic. This also disables flushing of the R0-logger instance.
4985	*/
4986	VMMRZCallRing3Disable(pVCpu);
4987	VMMRZCallRing3RemoveNotification(pVCpu);
4988	hmR0SvmPreRunGuestCommitted(pVCpu, pCtx, &SvmTransient);
4989
4990	rc = hmR0SvmRunGuest(pVCpu, pCtx, pVCpu->hm.s.svm.HCPhysVmcb);
4991
4992	/* Restore any residual host-state and save any bits shared between host and guest
4993	into the guest-CPU state. Re-enables interrupts! */
4994	hmR0SvmPostRunGuest(pVCpu, pCtx, &SvmTransient, rc);
4995
4996	if (RT_UNLIKELY( rc != VINF_SUCCESS /* Check for VMRUN errors. */
4997	\|\| SvmTransient.u64ExitCode == SVM_EXIT_INVALID)) /* Check for invalid guest-state errors. */
4998	{
4999	if (rc == VINF_SUCCESS)
5000	rc = VERR_SVM_INVALID_GUEST_STATE;
5001	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatPreExit, x);
5002	hmR0SvmReportWorldSwitchError(pVCpu, rc, pCtx);
5003	return rc;
5004	}
5005
5006	/* Handle the #VMEXIT. */
5007	HMSVM_EXITCODE_STAM_COUNTER_INC(SvmTransient.u64ExitCode);
5008	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatPreExit, &pVCpu->hm.s.StatExitHandling, x);
5009	VBOXVMM_R0_HMSVM_VMEXIT(pVCpu, pCtx, SvmTransient.u64ExitCode, pVCpu->hm.s.svm.pVmcb);
5010	rc = hmR0SvmHandleExit(pVCpu, pCtx, &SvmTransient);
5011	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExitHandling, x);
5012	if (rc != VINF_SUCCESS)
5013	break;
5014	if (++(*pcLoops) >= cMaxResumeLoops)
5015	{
5016	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchMaxResumeLoops);
5017	rc = VINF_EM_RAW_INTERRUPT;
5018	break;
5019	}
5020
5021	/*
5022	* Did the RIP change, if so, consider it a single step.
5023	* Otherwise, make sure one of the TFs gets set.
5024	*/
5025	if ( pCtx->rip != uRipStart
5026	\|\| pCtx->cs.Sel != uCsStart)
5027	{
5028	rc = VINF_EM_DBG_STEPPED;
5029	break;
5030	}
5031	pVCpu->hm.s.fCtxChanged \|= HM_CHANGED_GUEST_DR_MASK;
5032	}
5033
5034	/*
5035	* Clear the X86_EFL_TF if necessary.
5036	*/
5037	if (pVCpu->hm.s.fClearTrapFlag)
5038	{
5039	pVCpu->hm.s.fClearTrapFlag = false;
5040	pCtx->eflags.Bits.u1TF = 0;
5041	}
5042
5043	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatEntry, x);
5044	return rc;
5045	}
5046
5047	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
5048	/**
5049	* Runs the nested-guest code using AMD-V.
5050	*
5051	* @returns VBox status code.
5052	* @param pVCpu The cross context virtual CPU structure.
5053	* @param pCtx Pointer to the guest-CPU context.
5054	* @param pcLoops Pointer to the number of executed loops. If we're switching
5055	* from the guest-code execution loop to this nested-guest
5056	* execution loop pass the remainder value, else pass 0.
5057	*/
5058	static int hmR0SvmRunGuestCodeNested(PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t *pcLoops)
5059	{
5060	HMSVM_ASSERT_IN_NESTED_GUEST(pCtx);
5061	Assert(pcLoops);
5062	Assert(*pcLoops <= pVCpu->CTX_SUFF(pVM)->hm.s.cMaxResumeLoops);
5063
5064	SVMTRANSIENT SvmTransient;
5065	RT_ZERO(SvmTransient);
5066	SvmTransient.fUpdateTscOffsetting = true;
5067	SvmTransient.pVmcb = pCtx->hwvirt.svm.CTX_SUFF(pVmcb);
5068	SvmTransient.fIsNestedGuest = true;
5069
5070	int rc = VERR_INTERNAL_ERROR_4;
5071	for (;;)
5072	{
5073	Assert(!HMR0SuspendPending());
5074	HMSVM_ASSERT_CPU_SAFE();
5075
5076	/* Preparatory work for running nested-guest code, this may force us to return to
5077	ring-3. This bugger disables interrupts on VINF_SUCCESS! */
5078	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatEntry, x);
5079	rc = hmR0SvmPreRunGuestNested(pVCpu, pCtx, &SvmTransient);
5080	if ( rc != VINF_SUCCESS
5081	\|\| !CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
5082	{
5083	break;
5084	}
5085
5086	/*
5087	* No longjmps to ring-3 from this point on!!!
5088	*
5089	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional,
5090	* better than a kernel panic. This also disables flushing of the R0-logger instance.
5091	*/
5092	hmR0SvmPreRunGuestCommitted(pVCpu, pCtx, &SvmTransient);
5093
5094	rc = hmR0SvmRunGuest(pVCpu, pCtx, pCtx->hwvirt.svm.HCPhysVmcb);
5095
5096	/* Restore any residual host-state and save any bits shared between host and guest
5097	into the guest-CPU state. Re-enables interrupts! */
5098	hmR0SvmPostRunGuest(pVCpu, pCtx, &SvmTransient, rc);
5099
5100	if (RT_LIKELY( rc == VINF_SUCCESS
5101	&& SvmTransient.u64ExitCode != SVM_EXIT_INVALID))
5102	{ /* extremely likely */ }
5103	else
5104	{
5105	/* VMRUN failed, shouldn't really happen, Guru. */
5106	if (rc != VINF_SUCCESS)
5107	break;
5108
5109	/* Invalid nested-guest state. Cause a #VMEXIT but assert on strict builds. */
5110	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
5111	AssertMsgFailed(("Invalid nested-guest state. rc=%Rrc u64ExitCode=%#RX64\n", rc, SvmTransient.u64ExitCode));
5112	rc = VBOXSTRICTRC_TODO(IEMExecSvmVmexit(pVCpu, SVM_EXIT_INVALID, 0, 0));
5113	break;
5114	}
5115
5116	/* Handle the #VMEXIT. */
5117	HMSVM_NESTED_EXITCODE_STAM_COUNTER_INC(SvmTransient.u64ExitCode);
5118	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatPreExit, &pVCpu->hm.s.StatExitHandling, x);
5119	VBOXVMM_R0_HMSVM_VMEXIT(pVCpu, pCtx, SvmTransient.u64ExitCode, pCtx->hwvirt.svm.CTX_SUFF(pVmcb));
5120	rc = hmR0SvmHandleExitNested(pVCpu, pCtx, &SvmTransient);
5121	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExitHandling, x);
5122	if ( rc != VINF_SUCCESS
5123	\|\| !CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
5124	break;
5125	if (++(*pcLoops) >= pVCpu->CTX_SUFF(pVM)->hm.s.cMaxResumeLoops)
5126	{
5127	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchMaxResumeLoops);
5128	rc = VINF_EM_RAW_INTERRUPT;
5129	break;
5130	}
5131
5132	/** @todo handle single-stepping */
5133	}
5134
5135	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatEntry, x);
5136	return rc;
5137	}
5138	#endif
5139
5140
5141	/**
5142	* Runs the guest code using AMD-V.
5143	*
5144	* @returns Strict VBox status code.
5145	* @param pVCpu The cross context virtual CPU structure.
5146	* @param pCtx Pointer to the guest-CPU context.
5147	*/
5148	VMMR0DECL(VBOXSTRICTRC) SVMR0RunGuestCode(PVMCPU pVCpu, PCPUMCTX pCtx)
5149	{
5150	Assert(VMMRZCallRing3IsEnabled(pVCpu));
5151	HMSVM_ASSERT_PREEMPT_SAFE();
5152	VMMRZCallRing3SetNotification(pVCpu, hmR0SvmCallRing3Callback, pCtx);
5153
5154	uint32_t cLoops = 0;
5155	int rc;
5156	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
5157	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
5158	#endif
5159	{
5160	if (!pVCpu->hm.s.fSingleInstruction)
5161	rc = hmR0SvmRunGuestCodeNormal(pVCpu, pCtx, &cLoops);
5162	else
5163	rc = hmR0SvmRunGuestCodeStep(pVCpu, pCtx, &cLoops);
5164	}
5165	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
5166	else
5167	{
5168	rc = VINF_SVM_VMRUN;
5169	}
5170
5171	/* Re-check the nested-guest condition here as we may be transitioning from the normal
5172	execution loop into the nested-guest, hence this is not placed in the 'else' part above. */
5173	if (rc == VINF_SVM_VMRUN)
5174	{
5175	rc = hmR0SvmRunGuestCodeNested(pVCpu, pCtx, &cLoops);
5176	if (rc == VINF_SVM_VMEXIT)
5177	rc = VINF_SUCCESS;
5178	}
5179	#endif
5180
5181	/* Fixup error codes. */
5182	if (rc == VERR_EM_INTERPRETER)
5183	rc = VINF_EM_RAW_EMULATE_INSTR;
5184	else if (rc == VINF_EM_RESET)
5185	rc = VINF_EM_TRIPLE_FAULT;
5186
5187	/* Prepare to return to ring-3. This will remove longjmp notifications. */
5188	rc = hmR0SvmExitToRing3(pVCpu, pCtx, rc);
5189	Assert(!VMMRZCallRing3IsNotificationSet(pVCpu));
5190	return rc;
5191	}
5192
5193
5194	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
5195	/**
5196	* Determines whether an IOIO intercept is active for the nested-guest or not.
5197	*
5198	* @param pvIoBitmap Pointer to the nested-guest IO bitmap.
5199	* @param pIoExitInfo Pointer to the SVMIOIOEXITINFO.
5200	*/
5201	static bool hmR0SvmIsIoInterceptActive(void *pvIoBitmap, PSVMIOIOEXITINFO pIoExitInfo)
5202	{
5203	const uint16_t u16Port = pIoExitInfo->n.u16Port;
5204	const SVMIOIOTYPE enmIoType = (SVMIOIOTYPE)pIoExitInfo->n.u1Type;
5205	const uint8_t cbReg = (pIoExitInfo->u >> SVM_IOIO_OP_SIZE_SHIFT) & 7;
5206	const uint8_t cAddrSizeBits = ((pIoExitInfo->u >> SVM_IOIO_ADDR_SIZE_SHIFT) & 7) << 4;
5207	const uint8_t iEffSeg = pIoExitInfo->n.u3Seg;
5208	const bool fRep = pIoExitInfo->n.u1Rep;
5209	const bool fStrIo = pIoExitInfo->n.u1Str;
5210
5211	return HMSvmIsIOInterceptActive(pvIoBitmap, u16Port, enmIoType, cbReg, cAddrSizeBits, iEffSeg, fRep, fStrIo,
5212	NULL /* pIoExitInfo */);
5213	}
5214
5215
5216	/**
5217	* Handles a nested-guest \#VMEXIT (for all EXITCODE values except
5218	* SVM_EXIT_INVALID).
5219	*
5220	* @returns VBox status code (informational status codes included).
5221	* @param pVCpu The cross context virtual CPU structure.
5222	* @param pCtx Pointer to the guest-CPU context.
5223	* @param pSvmTransient Pointer to the SVM transient structure.
5224	*/
5225	static int hmR0SvmHandleExitNested(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5226	{
5227	HMSVM_ASSERT_IN_NESTED_GUEST(pCtx);
5228	Assert(pSvmTransient->u64ExitCode != SVM_EXIT_INVALID);
5229	Assert(pSvmTransient->u64ExitCode <= SVM_EXIT_MAX);
5230
5231	/** @todo Figure out why using IEM_CPUMCTX_EXTRN_SVM_VMEXIT_MASK instead of
5232	* HMSVM_CPUMCTX_EXTRN_ALL breaks nested guests (XP Pro, DSL etc.), see
5233	* also HMSvmNstGstVmExitNotify(). */
5234	#define NST_GST_VMEXIT_CALL_RET(a_pVCpu, a_pCtx, a_uExitCode, a_uExitInfo1, a_uExitInfo2) \
5235	do { \
5236	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL); \
5237	return VBOXSTRICTRC_TODO(IEMExecSvmVmexit((a_pVCpu), (a_uExitCode), (a_uExitInfo1), (a_uExitInfo2))); \
5238	} while (0)
5239
5240	/*
5241	* For all the #VMEXITs here we primarily figure out if the #VMEXIT is expected by the
5242	* nested-guest. If it isn't, it should be handled by the (outer) guest.
5243	*/
5244	PSVMVMCB pVmcbNstGst = pCtx->hwvirt.svm.CTX_SUFF(pVmcb);
5245	PSVMVMCBCTRL pVmcbNstGstCtrl = &pVmcbNstGst->ctrl;
5246	uint64_t const uExitCode = pVmcbNstGstCtrl->u64ExitCode;
5247	uint64_t const uExitInfo1 = pVmcbNstGstCtrl->u64ExitInfo1;
5248	uint64_t const uExitInfo2 = pVmcbNstGstCtrl->u64ExitInfo2;
5249
5250	Assert(uExitCode == pVmcbNstGstCtrl->u64ExitCode);
5251	switch (uExitCode)
5252	{
5253	case SVM_EXIT_CPUID:
5254	{
5255	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_CPUID))
5256	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5257	return hmR0SvmExitCpuid(pVCpu, pCtx, pSvmTransient);
5258	}
5259
5260	case SVM_EXIT_RDTSC:
5261	{
5262	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_RDTSC))
5263	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5264	return hmR0SvmExitRdtsc(pVCpu, pCtx, pSvmTransient);
5265	}
5266
5267	case SVM_EXIT_RDTSCP:
5268	{
5269	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_RDTSCP))
5270	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5271	return hmR0SvmExitRdtscp(pVCpu, pCtx, pSvmTransient);
5272	}
5273
5274	case SVM_EXIT_MONITOR:
5275	{
5276	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_MONITOR))
5277	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5278	return hmR0SvmExitMonitor(pVCpu, pCtx, pSvmTransient);
5279	}
5280
5281	case SVM_EXIT_MWAIT:
5282	{
5283	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_MWAIT))
5284	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5285	return hmR0SvmExitMwait(pVCpu, pCtx, pSvmTransient);
5286	}
5287
5288	case SVM_EXIT_HLT:
5289	{
5290	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_HLT))
5291	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5292	return hmR0SvmExitHlt(pVCpu, pCtx, pSvmTransient);
5293	}
5294
5295	case SVM_EXIT_MSR:
5296	{
5297	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_MSR_PROT))
5298	{
5299	uint32_t const idMsr = pCtx->ecx;
5300	uint16_t offMsrpm;
5301	uint8_t uMsrpmBit;
5302	int rc = HMSvmGetMsrpmOffsetAndBit(idMsr, &offMsrpm, &uMsrpmBit);
5303	if (RT_SUCCESS(rc))
5304	{
5305	Assert(uMsrpmBit == 0 \|\| uMsrpmBit == 2 \|\| uMsrpmBit == 4 \|\| uMsrpmBit == 6);
5306	Assert(offMsrpm < SVM_MSRPM_PAGES << X86_PAGE_4K_SHIFT);
5307
5308	uint8_t const pbMsrBitmap = (uint8_t const )pCtx->hwvirt.svm.CTX_SUFF(pvMsrBitmap);
5309	pbMsrBitmap += offMsrpm;
5310	bool const fInterceptRead = RT_BOOL(*pbMsrBitmap & RT_BIT(uMsrpmBit));
5311	bool const fInterceptWrite = RT_BOOL(*pbMsrBitmap & RT_BIT(uMsrpmBit + 1));
5312
5313	if ( (fInterceptWrite && pVmcbNstGstCtrl->u64ExitInfo1 == SVM_EXIT1_MSR_WRITE)
5314	\|\| (fInterceptRead && pVmcbNstGstCtrl->u64ExitInfo1 == SVM_EXIT1_MSR_READ))
5315	{
5316	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5317	}
5318	}
5319	else
5320	{
5321	/*
5322	* MSRs not covered by the MSRPM automatically cause an #VMEXIT.
5323	* See AMD-V spec. "15.11 MSR Intercepts".
5324	*/
5325	Assert(rc == VERR_OUT_OF_RANGE);
5326	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5327	}
5328	}
5329	return hmR0SvmExitMsr(pVCpu, pCtx, pSvmTransient);
5330	}
5331
5332	case SVM_EXIT_IOIO:
5333	{
5334	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_IOIO_PROT))
5335	{
5336	void *pvIoBitmap = pCtx->hwvirt.svm.CTX_SUFF(pvIoBitmap);
5337	SVMIOIOEXITINFO IoExitInfo;
5338	IoExitInfo.u = pVmcbNstGst->ctrl.u64ExitInfo1;
5339	bool const fIntercept = hmR0SvmIsIoInterceptActive(pvIoBitmap, &IoExitInfo);
5340	if (fIntercept)
5341	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5342	}
5343	return hmR0SvmExitIOInstr(pVCpu, pCtx, pSvmTransient);
5344	}
5345
5346	case SVM_EXIT_XCPT_PF:
5347	{
5348	PVM pVM = pVCpu->CTX_SUFF(pVM);
5349	if (pVM->hm.s.fNestedPaging)
5350	{
5351	uint32_t const u32ErrCode = pVmcbNstGstCtrl->u64ExitInfo1;
5352	uint64_t const uFaultAddress = pVmcbNstGstCtrl->u64ExitInfo2;
5353
5354	/* If the nested-guest is intercepting #PFs, cause a #PF #VMEXIT. */
5355	if (HMIsGuestSvmXcptInterceptSet(pVCpu, X86_XCPT_PF))
5356	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, u32ErrCode, uFaultAddress);
5357
5358	/* If the nested-guest is not intercepting #PFs, forward the #PF to the guest. */
5359	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR2);
5360	hmR0SvmSetPendingXcptPF(pVCpu, pCtx, u32ErrCode, uFaultAddress);
5361	return VINF_SUCCESS;
5362	}
5363	return hmR0SvmExitXcptPF(pVCpu, pCtx,pSvmTransient);
5364	}
5365
5366	case SVM_EXIT_XCPT_UD:
5367	{
5368	if (HMIsGuestSvmXcptInterceptSet(pVCpu, X86_XCPT_UD))
5369	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5370	hmR0SvmSetPendingXcptUD(pVCpu);
5371	return VINF_SUCCESS;
5372	}
5373
5374	case SVM_EXIT_XCPT_MF:
5375	{
5376	if (HMIsGuestSvmXcptInterceptSet(pVCpu, X86_XCPT_MF))
5377	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5378	return hmR0SvmExitXcptMF(pVCpu, pCtx, pSvmTransient);
5379	}
5380
5381	case SVM_EXIT_XCPT_DB:
5382	{
5383	if (HMIsGuestSvmXcptInterceptSet(pVCpu, X86_XCPT_DB))
5384	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5385	return hmR0SvmNestedExitXcptDB(pVCpu, pCtx, pSvmTransient);
5386	}
5387
5388	case SVM_EXIT_XCPT_AC:
5389	{
5390	if (HMIsGuestSvmXcptInterceptSet(pVCpu, X86_XCPT_AC))
5391	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5392	return hmR0SvmExitXcptAC(pVCpu, pCtx, pSvmTransient);
5393	}
5394
5395	case SVM_EXIT_XCPT_BP:
5396	{
5397	if (HMIsGuestSvmXcptInterceptSet(pVCpu, X86_XCPT_BP))
5398	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5399	return hmR0SvmNestedExitXcptBP(pVCpu, pCtx, pSvmTransient);
5400	}
5401
5402	case SVM_EXIT_READ_CR0:
5403	case SVM_EXIT_READ_CR3:
5404	case SVM_EXIT_READ_CR4:
5405	{
5406	uint8_t const uCr = uExitCode - SVM_EXIT_READ_CR0;
5407	if (HMIsGuestSvmReadCRxInterceptSet(pVCpu, uCr))
5408	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5409	return hmR0SvmExitReadCRx(pVCpu, pCtx, pSvmTransient);
5410	}
5411
5412	case SVM_EXIT_CR0_SEL_WRITE:
5413	{
5414	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_CR0_SEL_WRITE))
5415	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5416	return hmR0SvmExitWriteCRx(pVCpu, pCtx, pSvmTransient);
5417	}
5418
5419	case SVM_EXIT_WRITE_CR0:
5420	case SVM_EXIT_WRITE_CR3:
5421	case SVM_EXIT_WRITE_CR4:
5422	case SVM_EXIT_WRITE_CR8: /* CR8 writes would go to the V_TPR rather than here, since we run with V_INTR_MASKING. */
5423	{
5424	uint8_t const uCr = uExitCode - SVM_EXIT_WRITE_CR0;
5425	Log4Func(("Write CR%u: uExitInfo1=%#RX64 uExitInfo2=%#RX64\n", uCr, uExitInfo1, uExitInfo2));
5426
5427	if (HMIsGuestSvmWriteCRxInterceptSet(pVCpu, uCr))
5428	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5429	return hmR0SvmExitWriteCRx(pVCpu, pCtx, pSvmTransient);
5430	}
5431
5432	case SVM_EXIT_PAUSE:
5433	{
5434	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_PAUSE))
5435	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5436	return hmR0SvmExitPause(pVCpu, pCtx, pSvmTransient);
5437	}
5438
5439	case SVM_EXIT_VINTR:
5440	{
5441	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_VINTR))
5442	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5443	return hmR0SvmExitUnexpected(pVCpu, pCtx, pSvmTransient);
5444	}
5445
5446	case SVM_EXIT_INTR:
5447	case SVM_EXIT_NMI:
5448	case SVM_EXIT_SMI:
5449	case SVM_EXIT_XCPT_NMI: /* Should not occur, SVM_EXIT_NMI is used instead. */
5450	{
5451	/*
5452	* We shouldn't direct physical interrupts, NMIs, SMIs to the nested-guest.
5453	*
5454	* Although we don't intercept SMIs, the nested-guest might. Therefore, we might
5455	* get an SMI #VMEXIT here so simply ignore rather than causing a corresponding
5456	* nested-guest #VMEXIT.
5457	*/
5458	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_SVM_VMEXIT_MASK);
5459	return hmR0SvmExitIntr(pVCpu, pCtx, pSvmTransient);
5460	}
5461
5462	case SVM_EXIT_FERR_FREEZE:
5463	{
5464	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_FERR_FREEZE))
5465	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5466	return hmR0SvmExitFerrFreeze(pVCpu, pCtx, pSvmTransient);
5467	}
5468
5469	case SVM_EXIT_INVLPG:
5470	{
5471	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_INVLPG))
5472	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5473	return hmR0SvmExitInvlpg(pVCpu, pCtx, pSvmTransient);
5474	}
5475
5476	case SVM_EXIT_WBINVD:
5477	{
5478	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_WBINVD))
5479	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5480	return hmR0SvmExitWbinvd(pVCpu, pCtx, pSvmTransient);
5481	}
5482
5483	case SVM_EXIT_INVD:
5484	{
5485	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_INVD))
5486	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5487	return hmR0SvmExitInvd(pVCpu, pCtx, pSvmTransient);
5488	}
5489
5490	case SVM_EXIT_RDPMC:
5491	{
5492	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_RDPMC))
5493	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5494	return hmR0SvmExitRdpmc(pVCpu, pCtx, pSvmTransient);
5495	}
5496
5497	default:
5498	{
5499	switch (uExitCode)
5500	{
5501	case SVM_EXIT_READ_DR0: case SVM_EXIT_READ_DR1: case SVM_EXIT_READ_DR2: case SVM_EXIT_READ_DR3:
5502	case SVM_EXIT_READ_DR6: case SVM_EXIT_READ_DR7: case SVM_EXIT_READ_DR8: case SVM_EXIT_READ_DR9:
5503	case SVM_EXIT_READ_DR10: case SVM_EXIT_READ_DR11: case SVM_EXIT_READ_DR12: case SVM_EXIT_READ_DR13:
5504	case SVM_EXIT_READ_DR14: case SVM_EXIT_READ_DR15:
5505	{
5506	uint8_t const uDr = uExitCode - SVM_EXIT_READ_DR0;
5507	if (HMIsGuestSvmReadDRxInterceptSet(pVCpu, uDr))
5508	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5509	return hmR0SvmExitReadDRx(pVCpu, pCtx, pSvmTransient);
5510	}
5511
5512	case SVM_EXIT_WRITE_DR0: case SVM_EXIT_WRITE_DR1: case SVM_EXIT_WRITE_DR2: case SVM_EXIT_WRITE_DR3:
5513	case SVM_EXIT_WRITE_DR6: case SVM_EXIT_WRITE_DR7: case SVM_EXIT_WRITE_DR8: case SVM_EXIT_WRITE_DR9:
5514	case SVM_EXIT_WRITE_DR10: case SVM_EXIT_WRITE_DR11: case SVM_EXIT_WRITE_DR12: case SVM_EXIT_WRITE_DR13:
5515	case SVM_EXIT_WRITE_DR14: case SVM_EXIT_WRITE_DR15:
5516	{
5517	uint8_t const uDr = uExitCode - SVM_EXIT_WRITE_DR0;
5518	if (HMIsGuestSvmWriteDRxInterceptSet(pVCpu, uDr))
5519	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5520	return hmR0SvmExitWriteDRx(pVCpu, pCtx, pSvmTransient);
5521	}
5522
5523	case SVM_EXIT_XCPT_DE:
5524	/* SVM_EXIT_XCPT_DB: / / Handled above. */
5525	/* SVM_EXIT_XCPT_NMI: / / Handled above. */
5526	/* SVM_EXIT_XCPT_BP: / / Handled above. */
5527	case SVM_EXIT_XCPT_OF:
5528	case SVM_EXIT_XCPT_BR:
5529	/* SVM_EXIT_XCPT_UD: / / Handled above. */
5530	case SVM_EXIT_XCPT_NM:
5531	case SVM_EXIT_XCPT_DF:
5532	case SVM_EXIT_XCPT_CO_SEG_OVERRUN:
5533	case SVM_EXIT_XCPT_TS:
5534	case SVM_EXIT_XCPT_NP:
5535	case SVM_EXIT_XCPT_SS:
5536	case SVM_EXIT_XCPT_GP:
5537	/* SVM_EXIT_XCPT_PF: / / Handled above. */
5538	case SVM_EXIT_XCPT_15: /* Reserved. */
5539	/* SVM_EXIT_XCPT_MF: / / Handled above. */
5540	/* SVM_EXIT_XCPT_AC: / / Handled above. */
5541	case SVM_EXIT_XCPT_MC:
5542	case SVM_EXIT_XCPT_XF:
5543	case SVM_EXIT_XCPT_20: case SVM_EXIT_XCPT_21: case SVM_EXIT_XCPT_22: case SVM_EXIT_XCPT_23:
5544	case SVM_EXIT_XCPT_24: case SVM_EXIT_XCPT_25: case SVM_EXIT_XCPT_26: case SVM_EXIT_XCPT_27:
5545	case SVM_EXIT_XCPT_28: case SVM_EXIT_XCPT_29: case SVM_EXIT_XCPT_30: case SVM_EXIT_XCPT_31:
5546	{
5547	uint8_t const uVector = uExitCode - SVM_EXIT_XCPT_0;
5548	if (HMIsGuestSvmXcptInterceptSet(pVCpu, uVector))
5549	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5550	return hmR0SvmExitXcptGeneric(pVCpu, pCtx, pSvmTransient);
5551	}
5552
5553	case SVM_EXIT_XSETBV:
5554	{
5555	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_XSETBV))
5556	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5557	return hmR0SvmExitXsetbv(pVCpu, pCtx, pSvmTransient);
5558	}
5559
5560	case SVM_EXIT_TASK_SWITCH:
5561	{
5562	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_TASK_SWITCH))
5563	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5564	return hmR0SvmExitTaskSwitch(pVCpu, pCtx, pSvmTransient);
5565	}
5566
5567	case SVM_EXIT_IRET:
5568	{
5569	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_IRET))
5570	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5571	return hmR0SvmExitIret(pVCpu, pCtx, pSvmTransient);
5572	}
5573
5574	case SVM_EXIT_SHUTDOWN:
5575	{
5576	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_SHUTDOWN))
5577	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5578	return hmR0SvmExitShutdown(pVCpu, pCtx, pSvmTransient);
5579	}
5580
5581	case SVM_EXIT_VMMCALL:
5582	{
5583	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_VMMCALL))
5584	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5585	return hmR0SvmExitVmmCall(pVCpu, pCtx, pSvmTransient);
5586	}
5587
5588	case SVM_EXIT_CLGI:
5589	{
5590	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_CLGI))
5591	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5592	return hmR0SvmExitClgi(pVCpu, pCtx, pSvmTransient);
5593	}
5594
5595	case SVM_EXIT_STGI:
5596	{
5597	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_STGI))
5598	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5599	return hmR0SvmExitStgi(pVCpu, pCtx, pSvmTransient);
5600	}
5601
5602	case SVM_EXIT_VMLOAD:
5603	{
5604	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_VMLOAD))
5605	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5606	return hmR0SvmExitVmload(pVCpu, pCtx, pSvmTransient);
5607	}
5608
5609	case SVM_EXIT_VMSAVE:
5610	{
5611	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_VMSAVE))
5612	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5613	return hmR0SvmExitVmsave(pVCpu, pCtx, pSvmTransient);
5614	}
5615
5616	case SVM_EXIT_INVLPGA:
5617	{
5618	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_INVLPGA))
5619	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5620	return hmR0SvmExitInvlpga(pVCpu, pCtx, pSvmTransient);
5621	}
5622
5623	case SVM_EXIT_VMRUN:
5624	{
5625	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_VMRUN))
5626	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5627	return hmR0SvmExitVmrun(pVCpu, pCtx, pSvmTransient);
5628	}
5629
5630	case SVM_EXIT_RSM:
5631	{
5632	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_RSM))
5633	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5634	hmR0SvmSetPendingXcptUD(pVCpu);
5635	return VINF_SUCCESS;
5636	}
5637
5638	case SVM_EXIT_SKINIT:
5639	{
5640	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_SKINIT))
5641	NST_GST_VMEXIT_CALL_RET(pVCpu, pCtx, uExitCode, uExitInfo1, uExitInfo2);
5642	hmR0SvmSetPendingXcptUD(pVCpu);
5643	return VINF_SUCCESS;
5644	}
5645
5646	case SVM_EXIT_NPF:
5647	{
5648	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
5649	return hmR0SvmExitNestedPF(pVCpu, pCtx, pSvmTransient);
5650	}
5651
5652	case SVM_EXIT_INIT: /* We shouldn't get INIT signals while executing a nested-guest. */
5653	return hmR0SvmExitUnexpected(pVCpu, pCtx, pSvmTransient);
5654
5655	default:
5656	{
5657	AssertMsgFailed(("hmR0SvmHandleExitNested: Unknown exit code %#x\n", pSvmTransient->u64ExitCode));
5658	pVCpu->hm.s.u32HMError = pSvmTransient->u64ExitCode;
5659	return VERR_SVM_UNKNOWN_EXIT;
5660	}
5661	}
5662	}
5663	}
5664	/* not reached */
5665
5666	#undef NST_GST_VMEXIT_CALL_RET
5667	}
5668	#endif
5669
5670
5671	/**
5672	* Handles a guest \#VMEXIT (for all EXITCODE values except SVM_EXIT_INVALID).
5673	*
5674	* @returns VBox status code (informational status codes included).
5675	* @param pVCpu The cross context virtual CPU structure.
5676	* @param pCtx Pointer to the guest-CPU context.
5677	* @param pSvmTransient Pointer to the SVM transient structure.
5678	*/
5679	static int hmR0SvmHandleExit(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5680	{
5681	Assert(pSvmTransient->u64ExitCode != SVM_EXIT_INVALID);
5682	Assert(pSvmTransient->u64ExitCode <= SVM_EXIT_MAX);
5683
5684	#ifdef DEBUG_ramshankar
5685	# define VMEXIT_CALL_RET(a_fDbg, a_CallExpr) \
5686	do { \
5687	if ((a_fDbg) == 1) \
5688	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL); \
5689	int rc = a_CallExpr; \
5690	if ((a_fDbg) == 1) \
5691	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST); \
5692	return rc; \
5693	} while (0)
5694	#else
5695	# define VMEXIT_CALL_RET(a_fDbg, a_CallExpr) return a_CallExpr
5696	#endif
5697
5698	/*
5699	* The ordering of the case labels is based on most-frequently-occurring #VMEXITs
5700	* for most guests under normal workloads (for some definition of "normal").
5701	*/
5702	uint64_t const uExitCode = pSvmTransient->u64ExitCode;
5703	switch (uExitCode)
5704	{
5705	case SVM_EXIT_NPF: VMEXIT_CALL_RET(0, hmR0SvmExitNestedPF(pVCpu, pCtx, pSvmTransient));
5706	case SVM_EXIT_IOIO: VMEXIT_CALL_RET(0, hmR0SvmExitIOInstr(pVCpu, pCtx, pSvmTransient));
5707	case SVM_EXIT_RDTSC: VMEXIT_CALL_RET(0, hmR0SvmExitRdtsc(pVCpu, pCtx, pSvmTransient));
5708	case SVM_EXIT_RDTSCP: VMEXIT_CALL_RET(0, hmR0SvmExitRdtscp(pVCpu, pCtx, pSvmTransient));
5709	case SVM_EXIT_CPUID: VMEXIT_CALL_RET(0, hmR0SvmExitCpuid(pVCpu, pCtx, pSvmTransient));
5710	case SVM_EXIT_XCPT_PF: VMEXIT_CALL_RET(0, hmR0SvmExitXcptPF(pVCpu, pCtx, pSvmTransient));
5711	case SVM_EXIT_MSR: VMEXIT_CALL_RET(0, hmR0SvmExitMsr(pVCpu, pCtx, pSvmTransient));
5712	case SVM_EXIT_MONITOR: VMEXIT_CALL_RET(0, hmR0SvmExitMonitor(pVCpu, pCtx, pSvmTransient));
5713	case SVM_EXIT_MWAIT: VMEXIT_CALL_RET(0, hmR0SvmExitMwait(pVCpu, pCtx, pSvmTransient));
5714	case SVM_EXIT_HLT: VMEXIT_CALL_RET(0, hmR0SvmExitHlt(pVCpu, pCtx, pSvmTransient));
5715
5716	case SVM_EXIT_XCPT_NMI: /* Should not occur, SVM_EXIT_NMI is used instead. */
5717	case SVM_EXIT_INTR:
5718	case SVM_EXIT_NMI: VMEXIT_CALL_RET(0, hmR0SvmExitIntr(pVCpu, pCtx, pSvmTransient));
5719
5720	case SVM_EXIT_READ_CR0:
5721	case SVM_EXIT_READ_CR3:
5722	case SVM_EXIT_READ_CR4: VMEXIT_CALL_RET(0, hmR0SvmExitReadCRx(pVCpu, pCtx, pSvmTransient));
5723
5724	case SVM_EXIT_CR0_SEL_WRITE:
5725	case SVM_EXIT_WRITE_CR0:
5726	case SVM_EXIT_WRITE_CR3:
5727	case SVM_EXIT_WRITE_CR4:
5728	case SVM_EXIT_WRITE_CR8: VMEXIT_CALL_RET(0, hmR0SvmExitWriteCRx(pVCpu, pCtx, pSvmTransient));
5729
5730	case SVM_EXIT_VINTR: VMEXIT_CALL_RET(0, hmR0SvmExitVIntr(pVCpu, pCtx, pSvmTransient));
5731	case SVM_EXIT_PAUSE: VMEXIT_CALL_RET(0, hmR0SvmExitPause(pVCpu, pCtx, pSvmTransient));
5732	case SVM_EXIT_VMMCALL: VMEXIT_CALL_RET(0, hmR0SvmExitVmmCall(pVCpu, pCtx, pSvmTransient));
5733	case SVM_EXIT_INVLPG: VMEXIT_CALL_RET(0, hmR0SvmExitInvlpg(pVCpu, pCtx, pSvmTransient));
5734	case SVM_EXIT_WBINVD: VMEXIT_CALL_RET(0, hmR0SvmExitWbinvd(pVCpu, pCtx, pSvmTransient));
5735	case SVM_EXIT_INVD: VMEXIT_CALL_RET(0, hmR0SvmExitInvd(pVCpu, pCtx, pSvmTransient));
5736	case SVM_EXIT_RDPMC: VMEXIT_CALL_RET(0, hmR0SvmExitRdpmc(pVCpu, pCtx, pSvmTransient));
5737	case SVM_EXIT_IRET: VMEXIT_CALL_RET(0, hmR0SvmExitIret(pVCpu, pCtx, pSvmTransient));
5738	case SVM_EXIT_XCPT_UD: VMEXIT_CALL_RET(0, hmR0SvmExitXcptUD(pVCpu, pCtx, pSvmTransient));
5739	case SVM_EXIT_XCPT_MF: VMEXIT_CALL_RET(0, hmR0SvmExitXcptMF(pVCpu, pCtx, pSvmTransient));
5740	case SVM_EXIT_XCPT_DB: VMEXIT_CALL_RET(0, hmR0SvmExitXcptDB(pVCpu, pCtx, pSvmTransient));
5741	case SVM_EXIT_XCPT_AC: VMEXIT_CALL_RET(0, hmR0SvmExitXcptAC(pVCpu, pCtx, pSvmTransient));
5742	case SVM_EXIT_XCPT_BP: VMEXIT_CALL_RET(0, hmR0SvmExitXcptBP(pVCpu, pCtx, pSvmTransient));
5743	case SVM_EXIT_XSETBV: VMEXIT_CALL_RET(0, hmR0SvmExitXsetbv(pVCpu, pCtx, pSvmTransient));
5744	case SVM_EXIT_FERR_FREEZE: VMEXIT_CALL_RET(0, hmR0SvmExitFerrFreeze(pVCpu, pCtx, pSvmTransient));
5745
5746	default:
5747	{
5748	switch (pSvmTransient->u64ExitCode)
5749	{
5750	case SVM_EXIT_READ_DR0: case SVM_EXIT_READ_DR1: case SVM_EXIT_READ_DR2: case SVM_EXIT_READ_DR3:
5751	case SVM_EXIT_READ_DR6: case SVM_EXIT_READ_DR7: case SVM_EXIT_READ_DR8: case SVM_EXIT_READ_DR9:
5752	case SVM_EXIT_READ_DR10: case SVM_EXIT_READ_DR11: case SVM_EXIT_READ_DR12: case SVM_EXIT_READ_DR13:
5753	case SVM_EXIT_READ_DR14: case SVM_EXIT_READ_DR15:
5754	VMEXIT_CALL_RET(0, hmR0SvmExitReadDRx(pVCpu, pCtx, pSvmTransient));
5755
5756	case SVM_EXIT_WRITE_DR0: case SVM_EXIT_WRITE_DR1: case SVM_EXIT_WRITE_DR2: case SVM_EXIT_WRITE_DR3:
5757	case SVM_EXIT_WRITE_DR6: case SVM_EXIT_WRITE_DR7: case SVM_EXIT_WRITE_DR8: case SVM_EXIT_WRITE_DR9:
5758	case SVM_EXIT_WRITE_DR10: case SVM_EXIT_WRITE_DR11: case SVM_EXIT_WRITE_DR12: case SVM_EXIT_WRITE_DR13:
5759	case SVM_EXIT_WRITE_DR14: case SVM_EXIT_WRITE_DR15:
5760	VMEXIT_CALL_RET(0, hmR0SvmExitWriteDRx(pVCpu, pCtx, pSvmTransient));
5761
5762	case SVM_EXIT_TASK_SWITCH: VMEXIT_CALL_RET(0, hmR0SvmExitTaskSwitch(pVCpu, pCtx, pSvmTransient));
5763	case SVM_EXIT_SHUTDOWN: VMEXIT_CALL_RET(0, hmR0SvmExitShutdown(pVCpu, pCtx, pSvmTransient));
5764
5765	case SVM_EXIT_SMI:
5766	case SVM_EXIT_INIT:
5767	{
5768	/*
5769	* We don't intercept SMIs. As for INIT signals, it really shouldn't ever happen here.
5770	* If it ever does, we want to know about it so log the exit code and bail.
5771	*/
5772	VMEXIT_CALL_RET(0, hmR0SvmExitUnexpected(pVCpu, pCtx, pSvmTransient));
5773	}
5774
5775	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
5776	case SVM_EXIT_CLGI: VMEXIT_CALL_RET(0, hmR0SvmExitClgi(pVCpu, pCtx, pSvmTransient));
5777	case SVM_EXIT_STGI: VMEXIT_CALL_RET(0, hmR0SvmExitStgi(pVCpu, pCtx, pSvmTransient));
5778	case SVM_EXIT_VMLOAD: VMEXIT_CALL_RET(0, hmR0SvmExitVmload(pVCpu, pCtx, pSvmTransient));
5779	case SVM_EXIT_VMSAVE: VMEXIT_CALL_RET(0, hmR0SvmExitVmsave(pVCpu, pCtx, pSvmTransient));
5780	case SVM_EXIT_INVLPGA: VMEXIT_CALL_RET(0, hmR0SvmExitInvlpga(pVCpu, pCtx, pSvmTransient));
5781	case SVM_EXIT_VMRUN: VMEXIT_CALL_RET(0, hmR0SvmExitVmrun(pVCpu, pCtx, pSvmTransient));
5782	#else
5783	case SVM_EXIT_CLGI:
5784	case SVM_EXIT_STGI:
5785	case SVM_EXIT_VMLOAD:
5786	case SVM_EXIT_VMSAVE:
5787	case SVM_EXIT_INVLPGA:
5788	case SVM_EXIT_VMRUN:
5789	#endif
5790	case SVM_EXIT_RSM:
5791	case SVM_EXIT_SKINIT:
5792	{
5793	hmR0SvmSetPendingXcptUD(pVCpu);
5794	return VINF_SUCCESS;
5795	}
5796
5797	#ifdef HMSVM_ALWAYS_TRAP_ALL_XCPTS
5798	case SVM_EXIT_XCPT_DE:
5799	/* SVM_EXIT_XCPT_DB: / / Handled above. */
5800	/* SVM_EXIT_XCPT_NMI: / / Handled above. */
5801	/* SVM_EXIT_XCPT_BP: / / Handled above. */
5802	case SVM_EXIT_XCPT_OF:
5803	case SVM_EXIT_XCPT_BR:
5804	/* SVM_EXIT_XCPT_UD: / / Handled above. */
5805	case SVM_EXIT_XCPT_NM:
5806	case SVM_EXIT_XCPT_DF:
5807	case SVM_EXIT_XCPT_CO_SEG_OVERRUN:
5808	case SVM_EXIT_XCPT_TS:
5809	case SVM_EXIT_XCPT_NP:
5810	case SVM_EXIT_XCPT_SS:
5811	case SVM_EXIT_XCPT_GP:
5812	/* SVM_EXIT_XCPT_PF: */
5813	case SVM_EXIT_XCPT_15: /* Reserved. */
5814	/* SVM_EXIT_XCPT_MF: / / Handled above. */
5815	/* SVM_EXIT_XCPT_AC: / / Handled above. */
5816	case SVM_EXIT_XCPT_MC:
5817	case SVM_EXIT_XCPT_XF:
5818	case SVM_EXIT_XCPT_20: case SVM_EXIT_XCPT_21: case SVM_EXIT_XCPT_22: case SVM_EXIT_XCPT_23:
5819	case SVM_EXIT_XCPT_24: case SVM_EXIT_XCPT_25: case SVM_EXIT_XCPT_26: case SVM_EXIT_XCPT_27:
5820	case SVM_EXIT_XCPT_28: case SVM_EXIT_XCPT_29: case SVM_EXIT_XCPT_30: case SVM_EXIT_XCPT_31:
5821	VMEXIT_CALL_RET(0, hmR0SvmExitXcptGeneric(pVCpu, pCtx, pSvmTransient));
5822	#endif /* HMSVM_ALWAYS_TRAP_ALL_XCPTS */
5823
5824	default:
5825	{
5826	AssertMsgFailed(("hmR0SvmHandleExit: Unknown exit code %#RX64\n", uExitCode));
5827	pVCpu->hm.s.u32HMError = uExitCode;
5828	return VERR_SVM_UNKNOWN_EXIT;
5829	}
5830	}
5831	}
5832	}
5833	/* not reached */
5834	#undef VMEXIT_CALL_RET
5835	}
5836
5837
5838	#ifdef DEBUG
5839	/* Is there some generic IPRT define for this that are not in Runtime/internal/\* ?? */
5840	# define HMSVM_ASSERT_PREEMPT_CPUID_VAR() \
5841	RTCPUID const idAssertCpu = RTThreadPreemptIsEnabled(NIL_RTTHREAD) ? NIL_RTCPUID : RTMpCpuId()
5842
5843	# define HMSVM_ASSERT_PREEMPT_CPUID() \
5844	do \
5845	{ \
5846	RTCPUID const idAssertCpuNow = RTThreadPreemptIsEnabled(NIL_RTTHREAD) ? NIL_RTCPUID : RTMpCpuId(); \
5847	AssertMsg(idAssertCpu == idAssertCpuNow, ("SVM %#x, %#x\n", idAssertCpu, idAssertCpuNow)); \
5848	} while (0)
5849
5850	# define HMSVM_VALIDATE_EXIT_HANDLER_PARAMS() \
5851	do { \
5852	AssertPtr(pVCpu); \
5853	AssertPtr(pCtx); \
5854	AssertPtr(pSvmTransient); \
5855	Assert(ASMIntAreEnabled()); \
5856	HMSVM_ASSERT_PREEMPT_SAFE(); \
5857	HMSVM_ASSERT_PREEMPT_CPUID_VAR(); \
5858	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)); \
5859	HMSVM_ASSERT_PREEMPT_SAFE(); \
5860	if (VMMR0IsLogFlushDisabled(pVCpu)) \
5861	HMSVM_ASSERT_PREEMPT_CPUID(); \
5862	} while (0)
5863	#else /* Release builds */
5864	# define HMSVM_VALIDATE_EXIT_HANDLER_PARAMS() do { NOREF(pVCpu); NOREF(pCtx); NOREF(pSvmTransient); } while (0)
5865	#endif
5866
5867
5868	/**
5869	* Worker for hmR0SvmInterpretInvlpg().
5870	*
5871	* @return VBox status code.
5872	* @param pVCpu The cross context virtual CPU structure.
5873	* @param pCpu Pointer to the disassembler state.
5874	* @param pCtx The guest CPU context.
5875	*/
5876	static int hmR0SvmInterpretInvlPgEx(PVMCPU pVCpu, PDISCPUSTATE pCpu, PCPUMCTX pCtx)
5877	{
5878	DISQPVPARAMVAL Param1;
5879	RTGCPTR GCPtrPage;
5880
5881	int rc = DISQueryParamVal(CPUMCTX2CORE(pCtx), pCpu, &pCpu->Param1, &Param1, DISQPVWHICH_SRC);
5882	if (RT_FAILURE(rc))
5883	return VERR_EM_INTERPRETER;
5884
5885	if ( Param1.type == DISQPV_TYPE_IMMEDIATE
5886	\|\| Param1.type == DISQPV_TYPE_ADDRESS)
5887	{
5888	if (!(Param1.flags & (DISQPV_FLAG_32 \| DISQPV_FLAG_64)))
5889	return VERR_EM_INTERPRETER;
5890
5891	GCPtrPage = Param1.val.val64;
5892	VBOXSTRICTRC rc2 = EMInterpretInvlpg(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx), GCPtrPage);
5893	rc = VBOXSTRICTRC_VAL(rc2);
5894	}
5895	else
5896	{
5897	Log4Func(("Invalid parameter type %#x\n", Param1.type));
5898	rc = VERR_EM_INTERPRETER;
5899	}
5900
5901	return rc;
5902	}
5903
5904
5905	/**
5906	* Interprets INVLPG.
5907	*
5908	* @returns VBox status code.
5909	* @retval VINF_* Scheduling instructions.
5910	* @retval VERR_EM_INTERPRETER Something we can't cope with.
5911	* @retval VERR_* Fatal errors.
5912	*
5913	* @param pVCpu The cross context virtual CPU structure.
5914	* @param pCtx The guest CPU context.
5915	*
5916	* @remarks Updates the RIP if the instruction was executed successfully.
5917	*/
5918	static int hmR0SvmInterpretInvlpg(PVMCPU pVCpu, PCPUMCTX pCtx)
5919	{
5920	/* Only allow 32 & 64 bit code. */
5921	if (CPUMGetGuestCodeBits(pVCpu) != 16)
5922	{
5923	PDISSTATE pDis = &pVCpu->hm.s.DisState;
5924	int rc = EMInterpretDisasCurrent(pVCpu->CTX_SUFF(pVM), pVCpu, pDis, NULL /* pcbInstr */);
5925	if ( RT_SUCCESS(rc)
5926	&& pDis->pCurInstr->uOpcode == OP_INVLPG)
5927	{
5928	rc = hmR0SvmInterpretInvlPgEx(pVCpu, pDis, pCtx);
5929	if (RT_SUCCESS(rc))
5930	pCtx->rip += pDis->cbInstr;
5931	return rc;
5932	}
5933	else
5934	Log4Func(("EMInterpretDisasCurrent failed! rc=%Rrc uOpCode=%#x\n", rc, pDis->pCurInstr->uOpcode));
5935	}
5936	return VERR_EM_INTERPRETER;
5937	}
5938
5939
5940	/**
5941	* Gets the IEM exception flags for the specified SVM event.
5942	*
5943	* @returns The IEM exception flags.
5944	* @param pEvent Pointer to the SVM event.
5945	*
5946	* @remarks This function currently only constructs flags required for
5947	* IEMEvaluateRecursiveXcpt and not the complete flags (e.g. error-code
5948	* and CR2 aspects of an exception are not included).
5949	*/
5950	static uint32_t hmR0SvmGetIemXcptFlags(PCSVMEVENT pEvent)
5951	{
5952	uint8_t const uEventType = pEvent->n.u3Type;
5953	uint32_t fIemXcptFlags;
5954	switch (uEventType)
5955	{
5956	case SVM_EVENT_EXCEPTION:
5957	/*
5958	* Only INT3 and INTO instructions can raise #BP and #OF exceptions.
5959	* See AMD spec. Table 8-1. "Interrupt Vector Source and Cause".
5960	*/
5961	if (pEvent->n.u8Vector == X86_XCPT_BP)
5962	{
5963	fIemXcptFlags = IEM_XCPT_FLAGS_T_SOFT_INT \| IEM_XCPT_FLAGS_BP_INSTR;
5964	break;
5965	}
5966	if (pEvent->n.u8Vector == X86_XCPT_OF)
5967	{
5968	fIemXcptFlags = IEM_XCPT_FLAGS_T_SOFT_INT \| IEM_XCPT_FLAGS_OF_INSTR;
5969	break;
5970	}
5971	/** @todo How do we distinguish ICEBP \#DB from the regular one? */
5972	RT_FALL_THRU();
5973	case SVM_EVENT_NMI:
5974	fIemXcptFlags = IEM_XCPT_FLAGS_T_CPU_XCPT;
5975	break;
5976
5977	case SVM_EVENT_EXTERNAL_IRQ:
5978	fIemXcptFlags = IEM_XCPT_FLAGS_T_EXT_INT;
5979	break;
5980
5981	case SVM_EVENT_SOFTWARE_INT:
5982	fIemXcptFlags = IEM_XCPT_FLAGS_T_SOFT_INT;
5983	break;
5984
5985	default:
5986	fIemXcptFlags = 0;
5987	AssertMsgFailed(("Unexpected event type! uEventType=%#x uVector=%#x", uEventType, pEvent->n.u8Vector));
5988	break;
5989	}
5990	return fIemXcptFlags;
5991	}
5992
5993
5994	/**
5995	* Handle a condition that occurred while delivering an event through the guest
5996	* IDT.
5997	*
5998	* @returns VBox status code (informational error codes included).
5999	* @retval VINF_SUCCESS if we should continue handling the \#VMEXIT.
6000	* @retval VINF_HM_DOUBLE_FAULT if a \#DF condition was detected and we ought to
6001	* continue execution of the guest which will delivery the \#DF.
6002	* @retval VINF_EM_RESET if we detected a triple-fault condition.
6003	* @retval VERR_EM_GUEST_CPU_HANG if we detected a guest CPU hang.
6004	*
6005	* @param pVCpu The cross context virtual CPU structure.
6006	* @param pCtx Pointer to the guest-CPU context.
6007	* @param pSvmTransient Pointer to the SVM transient structure.
6008	*
6009	* @remarks No-long-jump zone!!!
6010	*/
6011	static int hmR0SvmCheckExitDueToEventDelivery(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6012	{
6013	int rc = VINF_SUCCESS;
6014	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
6015	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR2);
6016
6017	Log4(("EXITINTINFO: Pending vectoring event %#RX64 Valid=%RTbool ErrValid=%RTbool Err=%#RX32 Type=%u Vector=%u\n",
6018	pVmcb->ctrl.ExitIntInfo.u, !!pVmcb->ctrl.ExitIntInfo.n.u1Valid, !!pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid,
6019	pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode, pVmcb->ctrl.ExitIntInfo.n.u3Type, pVmcb->ctrl.ExitIntInfo.n.u8Vector));
6020
6021	/*
6022	* The EXITINTINFO (if valid) contains the prior exception (IDT vector) that was trying to
6023	* be delivered to the guest which caused a #VMEXIT which was intercepted (Exit vector).
6024	*
6025	* See AMD spec. 15.7.3 "EXITINFO Pseudo-Code".
6026	*/
6027	if (pVmcb->ctrl.ExitIntInfo.n.u1Valid)
6028	{
6029	IEMXCPTRAISE enmRaise;
6030	IEMXCPTRAISEINFO fRaiseInfo;
6031	bool const fExitIsHwXcpt = pSvmTransient->u64ExitCode - SVM_EXIT_XCPT_0 <= SVM_EXIT_XCPT_31;
6032	uint8_t const uIdtVector = pVmcb->ctrl.ExitIntInfo.n.u8Vector;
6033	if (fExitIsHwXcpt)
6034	{
6035	uint8_t const uExitVector = pSvmTransient->u64ExitCode - SVM_EXIT_XCPT_0;
6036	uint32_t const fIdtVectorFlags = hmR0SvmGetIemXcptFlags(&pVmcb->ctrl.ExitIntInfo);
6037	uint32_t const fExitVectorFlags = IEM_XCPT_FLAGS_T_CPU_XCPT;
6038	enmRaise = IEMEvaluateRecursiveXcpt(pVCpu, fIdtVectorFlags, uIdtVector, fExitVectorFlags, uExitVector, &fRaiseInfo);
6039	}
6040	else
6041	{
6042	/*
6043	* If delivery of an event caused a #VMEXIT that is not an exception (e.g. #NPF)
6044	* then we end up here.
6045	*
6046	* If the event was:
6047	* - a software interrupt, we can re-execute the instruction which will
6048	* regenerate the event.
6049	* - an NMI, we need to clear NMI blocking and re-inject the NMI.
6050	* - a hardware exception or external interrupt, we re-inject it.
6051	*/
6052	fRaiseInfo = IEMXCPTRAISEINFO_NONE;
6053	if (pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_SOFTWARE_INT)
6054	enmRaise = IEMXCPTRAISE_REEXEC_INSTR;
6055	else
6056	enmRaise = IEMXCPTRAISE_PREV_EVENT;
6057	}
6058
6059	switch (enmRaise)
6060	{
6061	case IEMXCPTRAISE_CURRENT_XCPT:
6062	case IEMXCPTRAISE_PREV_EVENT:
6063	{
6064	/* For software interrupts, we shall re-execute the instruction. */
6065	if (!(fRaiseInfo & IEMXCPTRAISEINFO_SOFT_INT_XCPT))
6066	{
6067	RTGCUINTPTR GCPtrFaultAddress = 0;
6068
6069	/* If we are re-injecting an NMI, clear NMI blocking. */
6070	if (pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_NMI)
6071	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_BLOCK_NMIS);
6072
6073	/* Determine a vectoring #PF condition, see comment in hmR0SvmExitXcptPF(). */
6074	if (fRaiseInfo & (IEMXCPTRAISEINFO_EXT_INT_PF \| IEMXCPTRAISEINFO_NMI_PF))
6075	{
6076	pSvmTransient->fVectoringPF = true;
6077	Log4Func(("IDT: Pending vectoring #PF due to delivery of Ext-Int/NMI. uCR2=%#RX64\n", pCtx->cr2));
6078	}
6079	else if ( pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_EXCEPTION
6080	&& uIdtVector == X86_XCPT_PF)
6081	{
6082	/*
6083	* If the previous exception was a #PF, we need to recover the CR2 value.
6084	* This can't happen with shadow paging.
6085	*/
6086	GCPtrFaultAddress = pCtx->cr2;
6087	}
6088
6089	/*
6090	* Without nested paging, when uExitVector is #PF, CR2 value will be updated from the VMCB's
6091	* exit info. fields, if it's a guest #PF, see hmR0SvmExitXcptPF().
6092	*/
6093	Assert(pVmcb->ctrl.ExitIntInfo.n.u3Type != SVM_EVENT_SOFTWARE_INT);
6094	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectPendingReflect);
6095	hmR0SvmSetPendingEvent(pVCpu, &pVmcb->ctrl.ExitIntInfo, GCPtrFaultAddress);
6096
6097	Log4Func(("IDT: Pending vectoring event %#RX64 ErrValid=%RTbool Err=%#RX32 GCPtrFaultAddress=%#RX64\n",
6098	pVmcb->ctrl.ExitIntInfo.u, RT_BOOL(pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid),
6099	pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode, GCPtrFaultAddress));
6100	}
6101	break;
6102	}
6103
6104	case IEMXCPTRAISE_REEXEC_INSTR:
6105	{
6106	Assert(rc == VINF_SUCCESS);
6107	break;
6108	}
6109
6110	case IEMXCPTRAISE_DOUBLE_FAULT:
6111	{
6112	/*
6113	* Determing a vectoring double #PF condition. Used later, when PGM evaluates
6114	* the second #PF as a guest #PF (and not a shadow #PF) and needs to be
6115	* converted into a #DF.
6116	*/
6117	if (fRaiseInfo & IEMXCPTRAISEINFO_PF_PF)
6118	{
6119	Log4Func(("IDT: Pending vectoring double #PF uCR2=%#RX64\n", pCtx->cr2));
6120	pSvmTransient->fVectoringDoublePF = true;
6121	Assert(rc == VINF_SUCCESS);
6122	}
6123	else
6124	{
6125	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectPendingReflect);
6126	hmR0SvmSetPendingXcptDF(pVCpu);
6127	rc = VINF_HM_DOUBLE_FAULT;
6128	}
6129	break;
6130	}
6131
6132	case IEMXCPTRAISE_TRIPLE_FAULT:
6133	{
6134	rc = VINF_EM_RESET;
6135	break;
6136	}
6137
6138	case IEMXCPTRAISE_CPU_HANG:
6139	{
6140	rc = VERR_EM_GUEST_CPU_HANG;
6141	break;
6142	}
6143
6144	default:
6145	{
6146	AssertMsgFailed(("hmR0SvmExitCpuid: EMInterpretCpuId failed with %Rrc\n", rc));
6147	rc = VERR_SVM_IPE_2;
6148	break;
6149	}
6150	}
6151	}
6152	Assert(rc == VINF_SUCCESS \|\| rc == VINF_HM_DOUBLE_FAULT \|\| rc == VINF_EM_RESET \|\| rc == VERR_EM_GUEST_CPU_HANG);
6153	NOREF(pCtx);
6154	return rc;
6155	}
6156
6157
6158	/**
6159	* Advances the guest RIP making use of the CPU's NRIP_SAVE feature if
6160	* supported, otherwise advances the RIP by the number of bytes specified in
6161	* @a cb.
6162	*
6163	* @param pVCpu The cross context virtual CPU structure.
6164	* @param pCtx Pointer to the guest-CPU context.
6165	* @param cb RIP increment value in bytes.
6166	*
6167	* @remarks Use this function only from \#VMEXIT's where the NRIP value is valid
6168	* when NRIP_SAVE is supported by the CPU, otherwise use
6169	* hmR0SvmAdvanceRipDumb!
6170	*/
6171	DECLINLINE(void) hmR0SvmAdvanceRipHwAssist(PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t cb)
6172	{
6173	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu, pCtx);
6174	if (fSupportsNextRipSave)
6175	{
6176	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
6177	Assert(pVmcb);
6178	Assert(pVmcb->ctrl.u64NextRIP);
6179	Assert(!(pCtx->fExtrn & CPUMCTX_EXTRN_RIP));
6180	AssertRelease(pVmcb->ctrl.u64NextRIP - pCtx->rip == cb); /* temporary, remove later */
6181	pCtx->rip = pVmcb->ctrl.u64NextRIP;
6182	}
6183	else
6184	pCtx->rip += cb;
6185
6186	HMSVM_UPDATE_INTR_SHADOW(pVCpu, pCtx);
6187	}
6188
6189
6190	/**
6191	* Gets the length of the current instruction if the CPU supports the NRIP_SAVE
6192	* feature. Otherwise, returns the value in @a cbLikely.
6193	*
6194	* @param pVCpu The cross context virtual CPU structure.
6195	* @param pCtx Pointer to the guest-CPU context.
6196	* @param cbLikely The likely instruction length.
6197	*/
6198	DECLINLINE(uint8_t) hmR0SvmGetInstrLengthHwAssist(PVMCPU pVCpu, PCPUMCTX pCtx, uint8_t cbLikely)
6199	{
6200	Assert(cbLikely <= 15); /* See Intel spec. 2.3.11 "AVX Instruction Length" */
6201	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu, pCtx);
6202	if (fSupportsNextRipSave)
6203	{
6204	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
6205	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pCtx->rip;
6206	Assert(cbInstr == cbLikely);
6207	return cbInstr;
6208	}
6209	return cbLikely;
6210	}
6211
6212
6213	/**
6214	* Advances the guest RIP by the number of bytes specified in @a cb. This does
6215	* not make use of any hardware features to determine the instruction length.
6216	*
6217	* @param pVCpu The cross context virtual CPU structure.
6218	* @param pCtx Pointer to the guest-CPU context.
6219	* @param cb RIP increment value in bytes.
6220	*/
6221	DECLINLINE(void) hmR0SvmAdvanceRipDumb(PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t cb)
6222	{
6223	pCtx->rip += cb;
6224	HMSVM_UPDATE_INTR_SHADOW(pVCpu, pCtx);
6225	}
6226	#undef HMSVM_UPDATE_INTR_SHADOW
6227
6228
6229	/* -=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= */
6230	/* -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- #VMEXIT handlers -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- */
6231	/* -=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= */
6232
6233	/** @name \#VMEXIT handlers.
6234	* @{
6235	*/
6236
6237	/**
6238	* \#VMEXIT handler for external interrupts, NMIs, FPU assertion freeze and INIT
6239	* signals (SVM_EXIT_INTR, SVM_EXIT_NMI, SVM_EXIT_FERR_FREEZE, SVM_EXIT_INIT).
6240	*/
6241	HMSVM_EXIT_DECL hmR0SvmExitIntr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6242	{
6243	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6244
6245	if (pSvmTransient->u64ExitCode == SVM_EXIT_NMI)
6246	STAM_REL_COUNTER_INC(&pVCpu->hm.s.StatExitHostNmiInGC);
6247	else if (pSvmTransient->u64ExitCode == SVM_EXIT_INTR)
6248	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitExtInt);
6249
6250	/*
6251	* AMD-V has no preemption timer and the generic periodic preemption timer has no way to
6252	* signal -before- the timer fires if the current interrupt is our own timer or a some
6253	* other host interrupt. We also cannot examine what interrupt it is until the host
6254	* actually take the interrupt.
6255	*
6256	* Going back to executing guest code here unconditionally causes random scheduling
6257	* problems (observed on an AMD Phenom 9850 Quad-Core on Windows 64-bit host).
6258	*/
6259	return VINF_EM_RAW_INTERRUPT;
6260	}
6261
6262
6263	/**
6264	* \#VMEXIT handler for WBINVD (SVM_EXIT_WBINVD). Conditional \#VMEXIT.
6265	*/
6266	HMSVM_EXIT_DECL hmR0SvmExitWbinvd(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6267	{
6268	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6269
6270	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 2);
6271	int rc = VINF_SUCCESS;
6272	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6273	return rc;
6274	}
6275
6276
6277	/**
6278	* \#VMEXIT handler for INVD (SVM_EXIT_INVD). Unconditional \#VMEXIT.
6279	*/
6280	HMSVM_EXIT_DECL hmR0SvmExitInvd(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6281	{
6282	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6283
6284	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 2);
6285	int rc = VINF_SUCCESS;
6286	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6287	return rc;
6288	}
6289
6290
6291	/**
6292	* \#VMEXIT handler for INVD (SVM_EXIT_CPUID). Conditional \#VMEXIT.
6293	*/
6294	HMSVM_EXIT_DECL hmR0SvmExitCpuid(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6295	{
6296	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6297
6298	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_RIP
6299	\| CPUMCTX_EXTRN_CS);
6300	VBOXSTRICTRC rcStrict;
6301	PCEMEXITREC pExitRec = EMHistoryUpdateFlagsAndTypeAndPC(pVCpu,
6302	EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_CPUID),
6303	pVCpu->cpum.GstCtx.rip + pVCpu->cpum.GstCtx.cs.u64Base);
6304	if (!pExitRec)
6305	{
6306	PVM pVM = pVCpu->CTX_SUFF(pVM);
6307	rcStrict = EMInterpretCpuId(pVM, pVCpu, CPUMCTX2CORE(pCtx));
6308	if (RT_LIKELY(rcStrict == VINF_SUCCESS))
6309	{
6310	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 2);
6311	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6312	}
6313	else
6314	{
6315	AssertMsgFailed(("hmR0SvmExitCpuid: EMInterpretCpuId failed with %Rrc\n", VBOXSTRICTRC_VAL(rcStrict)));
6316	rcStrict = VERR_EM_INTERPRETER;
6317	}
6318	}
6319	else
6320	{
6321	/*
6322	* Frequent exit or something needing probing. Get state and call EMHistoryExec.
6323	*/
6324	Assert(pCtx == &pVCpu->cpum.GstCtx);
6325	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6326
6327	Log4(("CpuIdExit/%u: %04x:%08RX64: %#x/%#x -> EMHistoryExec\n",
6328	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip, pVCpu->cpum.GstCtx.eax, pVCpu->cpum.GstCtx.ecx));
6329
6330	rcStrict = EMHistoryExec(pVCpu, pExitRec, 0);
6331
6332	Log4(("CpuIdExit/%u: %04x:%08RX64: EMHistoryExec -> %Rrc + %04x:%08RX64\n",
6333	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip,
6334	VBOXSTRICTRC_VAL(rcStrict), pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip));
6335	}
6336	return VBOXSTRICTRC_TODO(rcStrict);
6337	}
6338
6339
6340	/**
6341	* \#VMEXIT handler for RDTSC (SVM_EXIT_RDTSC). Conditional \#VMEXIT.
6342	*/
6343	HMSVM_EXIT_DECL hmR0SvmExitRdtsc(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6344	{
6345	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6346	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6347	VBOXSTRICTRC rcStrict = IEMExecDecodedRdtsc(pVCpu, hmR0SvmGetInstrLengthHwAssist(pVCpu, pCtx, 2));
6348	if (rcStrict == VINF_SUCCESS)
6349	pSvmTransient->fUpdateTscOffsetting = true;
6350	else if (rcStrict == VINF_IEM_RAISED_XCPT)
6351	rcStrict = VINF_SUCCESS;
6352	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6353	return VBOXSTRICTRC_TODO(rcStrict);
6354	}
6355
6356
6357	/**
6358	* \#VMEXIT handler for RDTSCP (SVM_EXIT_RDTSCP). Conditional \#VMEXIT.
6359	*/
6360	HMSVM_EXIT_DECL hmR0SvmExitRdtscp(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6361	{
6362	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6363	VBOXSTRICTRC rcStrict = IEMExecDecodedRdtscp(pVCpu, hmR0SvmGetInstrLengthHwAssist(pVCpu, pCtx, 3));
6364	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6365	if (rcStrict == VINF_SUCCESS)
6366	pSvmTransient->fUpdateTscOffsetting = true;
6367	else if (rcStrict == VINF_IEM_RAISED_XCPT)
6368	rcStrict = VINF_SUCCESS;
6369	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6370	return VBOXSTRICTRC_TODO(rcStrict);
6371	}
6372
6373
6374	/**
6375	* \#VMEXIT handler for RDPMC (SVM_EXIT_RDPMC). Conditional \#VMEXIT.
6376	*/
6377	HMSVM_EXIT_DECL hmR0SvmExitRdpmc(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6378	{
6379	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6380	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR0
6381	\| CPUMCTX_EXTRN_CR4
6382	\| CPUMCTX_EXTRN_SS);
6383
6384	int rc = EMInterpretRdpmc(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
6385	if (RT_LIKELY(rc == VINF_SUCCESS))
6386	{
6387	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 2);
6388	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6389	}
6390	else
6391	{
6392	AssertMsgFailed(("hmR0SvmExitRdpmc: EMInterpretRdpmc failed with %Rrc\n", rc));
6393	rc = VERR_EM_INTERPRETER;
6394	}
6395	return rc;
6396	}
6397
6398
6399	/**
6400	* \#VMEXIT handler for INVLPG (SVM_EXIT_INVLPG). Conditional \#VMEXIT.
6401	*/
6402	HMSVM_EXIT_DECL hmR0SvmExitInvlpg(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6403	{
6404	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6405	Assert(!pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
6406
6407	bool const fSupportsDecodeAssists = hmR0SvmSupportsDecodeAssists(pVCpu, pCtx);
6408	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu, pCtx);
6409	if ( fSupportsDecodeAssists
6410	&& fSupportsNextRipSave)
6411	{
6412	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6413	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
6414	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pCtx->rip;
6415	RTGCPTR const GCPtrPage = pVmcb->ctrl.u64ExitInfo1;
6416	VBOXSTRICTRC rcStrict = IEMExecDecodedInvlpg(pVCpu, cbInstr, GCPtrPage);
6417	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6418	return VBOXSTRICTRC_VAL(rcStrict);
6419	}
6420
6421	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6422	int rc = hmR0SvmInterpretInvlpg(pVCpu, pCtx); /* Updates RIP if successful. */
6423	Assert(rc == VINF_SUCCESS \|\| rc == VERR_EM_INTERPRETER);
6424	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6425	return rc;
6426	}
6427
6428
6429	/**
6430	* \#VMEXIT handler for HLT (SVM_EXIT_HLT). Conditional \#VMEXIT.
6431	*/
6432	HMSVM_EXIT_DECL hmR0SvmExitHlt(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6433	{
6434	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6435
6436	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 1);
6437	int rc = EMShouldContinueAfterHalt(pVCpu, pCtx) ? VINF_SUCCESS : VINF_EM_HALT;
6438	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6439
6440	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitHlt);
6441	if (rc != VINF_SUCCESS)
6442	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHltToR3);
6443	return rc;
6444	}
6445
6446
6447	/**
6448	* \#VMEXIT handler for MONITOR (SVM_EXIT_MONITOR). Conditional \#VMEXIT.
6449	*/
6450	HMSVM_EXIT_DECL hmR0SvmExitMonitor(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6451	{
6452	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6453	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR0
6454	\| CPUMCTX_EXTRN_SS);
6455
6456	int rc = EMInterpretMonitor(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
6457	if (RT_LIKELY(rc == VINF_SUCCESS))
6458	{
6459	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 3);
6460	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6461	}
6462	else
6463	{
6464	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMonitor: EMInterpretMonitor failed with %Rrc\n", rc));
6465	rc = VERR_EM_INTERPRETER;
6466	}
6467	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitMonitor);
6468	return rc;
6469	}
6470
6471
6472	/**
6473	* \#VMEXIT handler for MWAIT (SVM_EXIT_MWAIT). Conditional \#VMEXIT.
6474	*/
6475	HMSVM_EXIT_DECL hmR0SvmExitMwait(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6476	{
6477	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6478	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR0
6479	\| CPUMCTX_EXTRN_SS);
6480
6481	VBOXSTRICTRC rc2 = EMInterpretMWait(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
6482	int rc = VBOXSTRICTRC_VAL(rc2);
6483	if ( rc == VINF_EM_HALT
6484	\|\| rc == VINF_SUCCESS)
6485	{
6486	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 3);
6487
6488	if ( rc == VINF_EM_HALT
6489	&& EMMonitorWaitShouldContinue(pVCpu, pCtx))
6490	{
6491	rc = VINF_SUCCESS;
6492	}
6493	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6494	}
6495	else
6496	{
6497	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMwait: EMInterpretMWait failed with %Rrc\n", rc));
6498	rc = VERR_EM_INTERPRETER;
6499	}
6500	AssertMsg(rc == VINF_SUCCESS \|\| rc == VINF_EM_HALT \|\| rc == VERR_EM_INTERPRETER,
6501	("hmR0SvmExitMwait: EMInterpretMWait failed rc=%Rrc\n", rc));
6502	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitMwait);
6503	return rc;
6504	}
6505
6506
6507	/**
6508	* \#VMEXIT handler for shutdown (triple-fault) (SVM_EXIT_SHUTDOWN). Conditional
6509	* \#VMEXIT.
6510	*/
6511	HMSVM_EXIT_DECL hmR0SvmExitShutdown(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6512	{
6513	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6514	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6515	return VINF_EM_RESET;
6516	}
6517
6518
6519	/**
6520	* \#VMEXIT handler for unexpected exits. Conditional \#VMEXIT.
6521	*/
6522	HMSVM_EXIT_DECL hmR0SvmExitUnexpected(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6523	{
6524	RT_NOREF(pCtx);
6525	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
6526	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6527	AssertMsgFailed(("hmR0SvmExitUnexpected: ExitCode=%#RX64 uExitInfo1=%#RX64 uExitInfo2=%#RX64\n", pSvmTransient->u64ExitCode,
6528	pVmcb->ctrl.u64ExitInfo1, pVmcb->ctrl.u64ExitInfo2));
6529	RT_NOREF(pVmcb);
6530	pVCpu->hm.s.u32HMError = (uint32_t)pSvmTransient->u64ExitCode;
6531	return VERR_SVM_UNEXPECTED_EXIT;
6532	}
6533
6534
6535	/**
6536	* \#VMEXIT handler for CRx reads (SVM_EXIT_READ_CR*). Conditional \#VMEXIT.
6537	*/
6538	HMSVM_EXIT_DECL hmR0SvmExitReadCRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6539	{
6540	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6541
6542	Log4Func(("CS:RIP=%04x:%#RX64\n", pCtx->cs.Sel, pCtx->rip));
6543	#ifdef VBOX_WITH_STATISTICS
6544	switch (pSvmTransient->u64ExitCode)
6545	{
6546	case SVM_EXIT_READ_CR0: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR0Read); break;
6547	case SVM_EXIT_READ_CR2: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR2Read); break;
6548	case SVM_EXIT_READ_CR3: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR3Read); break;
6549	case SVM_EXIT_READ_CR4: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR4Read); break;
6550	case SVM_EXIT_READ_CR8: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR8Read); break;
6551	}
6552	#endif
6553
6554	bool const fSupportsDecodeAssists = hmR0SvmSupportsDecodeAssists(pVCpu, pCtx);
6555	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu, pCtx);
6556	if ( fSupportsDecodeAssists
6557	&& fSupportsNextRipSave)
6558	{
6559	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
6560	bool const fMovCRx = RT_BOOL(pVmcb->ctrl.u64ExitInfo1 & SVM_EXIT1_MOV_CRX_MASK);
6561	if (fMovCRx)
6562	{
6563	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6564	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pCtx->rip;
6565	uint8_t const iCrReg = pSvmTransient->u64ExitCode - SVM_EXIT_READ_CR0;
6566	uint8_t const iGReg = pVmcb->ctrl.u64ExitInfo1 & SVM_EXIT1_MOV_CRX_GPR_NUMBER;
6567	VBOXSTRICTRC rcStrict = IEMExecDecodedMovCRxRead(pVCpu, cbInstr, iGReg, iCrReg);
6568	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6569	return VBOXSTRICTRC_VAL(rcStrict);
6570	}
6571	/* else: SMSW instruction, fall back below to IEM for this. */
6572	}
6573
6574	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6575	VBOXSTRICTRC rc2 = EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */);
6576	int rc = VBOXSTRICTRC_VAL(rc2);
6577	AssertMsg(rc == VINF_SUCCESS \|\| rc == VERR_EM_INTERPRETER \|\| rc == VINF_PGM_CHANGE_MODE \|\| rc == VINF_PGM_SYNC_CR3,
6578	("hmR0SvmExitReadCRx: EMInterpretInstruction failed rc=%Rrc\n", rc));
6579	Assert((pSvmTransient->u64ExitCode - SVM_EXIT_READ_CR0) <= 15);
6580	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6581	return rc;
6582	}
6583
6584
6585	/**
6586	* \#VMEXIT handler for CRx writes (SVM_EXIT_WRITE_CR*). Conditional \#VMEXIT.
6587	*/
6588	HMSVM_EXIT_DECL hmR0SvmExitWriteCRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6589	{
6590	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6591
6592	uint64_t const uExitCode = pSvmTransient->u64ExitCode;
6593	uint8_t const iCrReg = uExitCode == SVM_EXIT_CR0_SEL_WRITE ? 0 : (pSvmTransient->u64ExitCode - SVM_EXIT_WRITE_CR0);
6594	Assert(iCrReg <= 15);
6595
6596	VBOXSTRICTRC rcStrict = VERR_SVM_IPE_5;
6597	bool fDecodedInstr = false;
6598	bool const fSupportsDecodeAssists = hmR0SvmSupportsDecodeAssists(pVCpu, pCtx);
6599	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu, pCtx);
6600	if ( fSupportsDecodeAssists
6601	&& fSupportsNextRipSave)
6602	{
6603	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
6604	bool const fMovCRx = RT_BOOL(pVmcb->ctrl.u64ExitInfo1 & SVM_EXIT1_MOV_CRX_MASK);
6605	if (fMovCRx)
6606	{
6607	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6608	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pCtx->rip;
6609	uint8_t const iGReg = pVmcb->ctrl.u64ExitInfo1 & SVM_EXIT1_MOV_CRX_GPR_NUMBER;
6610	Log4Func(("Mov CR%u w/ iGReg=%#x\n", iCrReg, iGReg));
6611	rcStrict = IEMExecDecodedMovCRxWrite(pVCpu, cbInstr, iCrReg, iGReg);
6612	fDecodedInstr = true;
6613	}
6614	/* else: LMSW or CLTS instruction, fall back below to IEM for this. */
6615	}
6616
6617	if (!fDecodedInstr)
6618	{
6619	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6620	Log4Func(("iCrReg=%#x\n", iCrReg));
6621	rcStrict = IEMExecOneBypassEx(pVCpu, CPUMCTX2CORE(pCtx), NULL);
6622	if (RT_UNLIKELY( rcStrict == VERR_IEM_ASPECT_NOT_IMPLEMENTED
6623	\|\| rcStrict == VERR_IEM_INSTR_NOT_IMPLEMENTED))
6624	rcStrict = VERR_EM_INTERPRETER;
6625	}
6626
6627	if (rcStrict == VINF_SUCCESS)
6628	{
6629	switch (iCrReg)
6630	{
6631	case 0:
6632	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_CR0);
6633	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR0Write);
6634	break;
6635
6636	case 2:
6637	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_CR2);
6638	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR2Write);
6639	break;
6640
6641	case 3:
6642	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_CR3);
6643	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR3Write);
6644	break;
6645
6646	case 4:
6647	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_CR4);
6648	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR4Write);
6649	break;
6650
6651	case 8:
6652	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
6653	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR8Write);
6654	break;
6655
6656	default:
6657	{
6658	AssertMsgFailed(("hmR0SvmExitWriteCRx: Invalid/Unexpected Write-CRx exit. u64ExitCode=%#RX64 %#x\n",
6659	pSvmTransient->u64ExitCode, iCrReg));
6660	break;
6661	}
6662	}
6663	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6664	}
6665	else
6666	Assert(rcStrict == VERR_EM_INTERPRETER \|\| rcStrict == VINF_PGM_CHANGE_MODE \|\| rcStrict == VINF_PGM_SYNC_CR3);
6667	return VBOXSTRICTRC_TODO(rcStrict);
6668	}
6669
6670
6671	/**
6672	* \#VMEXIT helper for read MSRs, see hmR0SvmExitMsr.
6673	*
6674	* @returns Strict VBox status code.
6675	* @param pVCpu The cross context virtual CPU structure.
6676	* @param pVmcb Pointer to the VM control block.
6677	*/
6678	static VBOXSTRICTRC hmR0SvmExitReadMsr(PVMCPU pVCpu, PSVMVMCB pVmcb)
6679	{
6680	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
6681	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdmsr);
6682	Log4Func(("idMsr=%#RX32\n", pCtx->ecx));
6683
6684	VBOXSTRICTRC rcStrict;
6685	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu, pCtx);
6686	if (fSupportsNextRipSave)
6687	{
6688	/** @todo Optimize this: Only retrieve the MSR bits we need here. CPUMAllMsrs.cpp
6689	* can ask for what it needs instead of using CPUMCTX_EXTRN_ALL_MSRS. */
6690	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK \| CPUMCTX_EXTRN_ALL_MSRS);
6691	rcStrict = IEMExecDecodedRdmsr(pVCpu, pVmcb->ctrl.u64NextRIP - pCtx->rip);
6692	if (RT_LIKELY(rcStrict == VINF_SUCCESS))
6693	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict); /* RIP updated by IEMExecDecodedRdmsr(). */
6694	else
6695	AssertMsg( rcStrict == VINF_IEM_RAISED_XCPT
6696	\|\| rcStrict == VINF_CPUM_R3_MSR_WRITE,
6697	("Unexpected IEMExecDecodedWrmsr status: %Rrc\n", VBOXSTRICTRC_VAL(rcStrict)));
6698	}
6699	else
6700	{
6701	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK \| CPUMCTX_EXTRN_ALL_MSRS);
6702	rcStrict = IEMExecOne(pVCpu);
6703	if (RT_LIKELY(rcStrict == VINF_SUCCESS))
6704	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict); /* RIP updated by IEMExecOne(). */
6705	else
6706	AssertMsg( rcStrict == VINF_IEM_RAISED_XCPT
6707	\|\| rcStrict == VINF_CPUM_R3_MSR_READ, ("Unexpected IEMExecOne status: %Rrc\n", VBOXSTRICTRC_VAL(rcStrict)));
6708	}
6709	return rcStrict;
6710	}
6711
6712
6713	/**
6714	* \#VMEXIT helper for write MSRs, see hmR0SvmExitMsr.
6715	*
6716	* @returns Strict VBox status code.
6717	* @param pVCpu The cross context virtual CPU structure.
6718	* @param pVmcb Pointer to the VM control block.
6719	* @param pSvmTransient Pointer to the SVM-transient structure.
6720	*/
6721	static VBOXSTRICTRC hmR0SvmExitWriteMsr(PVMCPU pVCpu, PSVMVMCB pVmcb, PSVMTRANSIENT pSvmTransient)
6722	{
6723	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
6724	uint32_t const idMsr = pCtx->ecx;
6725	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitWrmsr);
6726	Log4Func(("idMsr=%#RX32\n", idMsr));
6727
6728	/*
6729	* Handle TPR patching MSR writes.
6730	* We utilitize the LSTAR MSR for patching.
6731	*/
6732	if ( pVCpu->CTX_SUFF(pVM)->hm.s.fTPRPatchingActive
6733	&& idMsr == MSR_K8_LSTAR)
6734	{
6735	if ((pCtx->eax & 0xff) != pSvmTransient->u8GuestTpr)
6736	{
6737	/* Our patch code uses LSTAR for TPR caching for 32-bit guests. */
6738	int rc2 = APICSetTpr(pVCpu, pCtx->eax & 0xff);
6739	AssertRC(rc2);
6740	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
6741	}
6742
6743	int rc = VINF_SUCCESS;
6744	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 2);
6745	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6746	return rc;
6747	}
6748
6749	/*
6750	* Handle regular MSR writes.
6751	*/
6752	VBOXSTRICTRC rcStrict;
6753	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu, pCtx);
6754	if (fSupportsNextRipSave)
6755	{
6756	/** @todo Optimize this: We don't need to get much of the MSR state here
6757	* since we're only updating. CPUMAllMsrs.cpp can ask for what it needs and
6758	* clear the applicable extern flags. */
6759	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK \| CPUMCTX_EXTRN_ALL_MSRS);
6760	rcStrict = IEMExecDecodedWrmsr(pVCpu, pVmcb->ctrl.u64NextRIP - pCtx->rip);
6761	if (RT_LIKELY(rcStrict == VINF_SUCCESS))
6762	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict); /* RIP updated by IEMExecDecodedWrmsr(). */
6763	else
6764	AssertMsg( rcStrict == VINF_IEM_RAISED_XCPT
6765	\|\| rcStrict == VINF_CPUM_R3_MSR_WRITE,
6766	("Unexpected IEMExecDecodedWrmsr status: %Rrc\n", VBOXSTRICTRC_VAL(rcStrict)));
6767	}
6768	else
6769	{
6770	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK \| CPUMCTX_EXTRN_ALL_MSRS);
6771	rcStrict = IEMExecOne(pVCpu);
6772	if (RT_LIKELY(rcStrict == VINF_SUCCESS))
6773	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict); /* RIP updated by IEMExecOne(). */
6774	else
6775	AssertMsg( rcStrict == VINF_IEM_RAISED_XCPT
6776	\|\| rcStrict == VINF_CPUM_R3_MSR_WRITE, ("Unexpected IEMExecOne status: %Rrc\n", VBOXSTRICTRC_VAL(rcStrict)));
6777	}
6778
6779	if (rcStrict == VINF_SUCCESS)
6780	{
6781	/* If this is an X2APIC WRMSR access, update the APIC TPR state. */
6782	if ( idMsr >= MSR_IA32_X2APIC_START
6783	&& idMsr <= MSR_IA32_X2APIC_END)
6784	{
6785	/*
6786	* We've already saved the APIC related guest-state (TPR) in hmR0SvmPostRunGuest().
6787	* When full APIC register virtualization is implemented we'll have to make sure
6788	* APIC state is saved from the VMCB before IEM changes it.
6789	*/
6790	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
6791	}
6792	else
6793	{
6794	switch (idMsr)
6795	{
6796	case MSR_IA32_TSC: pSvmTransient->fUpdateTscOffsetting = true; break;
6797	case MSR_K6_EFER: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_EFER_MSR); break;
6798	case MSR_K8_FS_BASE: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_FS); break;
6799	case MSR_K8_GS_BASE: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_GS); break;
6800	case MSR_IA32_SYSENTER_CS: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_SYSENTER_CS_MSR); break;
6801	case MSR_IA32_SYSENTER_EIP: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_SYSENTER_EIP_MSR); break;
6802	case MSR_IA32_SYSENTER_ESP: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_SYSENTER_ESP_MSR); break;
6803	}
6804	}
6805	}
6806
6807	return rcStrict;
6808	}
6809
6810
6811	/**
6812	* \#VMEXIT handler for MSR read and writes (SVM_EXIT_MSR). Conditional
6813	* \#VMEXIT.
6814	*/
6815	HMSVM_EXIT_DECL hmR0SvmExitMsr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6816	{
6817	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6818
6819	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
6820	if (pVmcb->ctrl.u64ExitInfo1 == SVM_EXIT1_MSR_READ)
6821	return VBOXSTRICTRC_TODO(hmR0SvmExitReadMsr(pVCpu, pVmcb));
6822
6823	Assert(pVmcb->ctrl.u64ExitInfo1 == SVM_EXIT1_MSR_WRITE);
6824	return VBOXSTRICTRC_TODO(hmR0SvmExitWriteMsr(pVCpu, pVmcb, pSvmTransient));
6825	}
6826
6827
6828	/**
6829	* \#VMEXIT handler for DRx read (SVM_EXIT_READ_DRx). Conditional \#VMEXIT.
6830	*/
6831	HMSVM_EXIT_DECL hmR0SvmExitReadDRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6832	{
6833	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6834	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6835
6836	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitDRxRead);
6837
6838	/** @todo Stepping with nested-guest. */
6839	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
6840	{
6841	/* We should -not- get this #VMEXIT if the guest's debug registers were active. */
6842	if (pSvmTransient->fWasGuestDebugStateActive)
6843	{
6844	AssertMsgFailed(("hmR0SvmExitReadDRx: Unexpected exit %#RX32\n", (uint32_t)pSvmTransient->u64ExitCode));
6845	pVCpu->hm.s.u32HMError = (uint32_t)pSvmTransient->u64ExitCode;
6846	return VERR_SVM_UNEXPECTED_EXIT;
6847	}
6848
6849	/*
6850	* Lazy DR0-3 loading.
6851	*/
6852	if (!pSvmTransient->fWasHyperDebugStateActive)
6853	{
6854	Assert(!DBGFIsStepping(pVCpu)); Assert(!pVCpu->hm.s.fSingleInstruction);
6855	Log5(("hmR0SvmExitReadDRx: Lazy loading guest debug registers\n"));
6856
6857	/* Don't intercept DRx read and writes. */
6858	PSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
6859	pVmcb->ctrl.u16InterceptRdDRx = 0;
6860	pVmcb->ctrl.u16InterceptWrDRx = 0;
6861	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
6862
6863	/* We're playing with the host CPU state here, make sure we don't preempt or longjmp. */
6864	VMMRZCallRing3Disable(pVCpu);
6865	HM_DISABLE_PREEMPT();
6866
6867	/* Save the host & load the guest debug state, restart execution of the MOV DRx instruction. */
6868	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
6869	Assert(CPUMIsGuestDebugStateActive(pVCpu) \|\| HC_ARCH_BITS == 32);
6870
6871	HM_RESTORE_PREEMPT();
6872	VMMRZCallRing3Enable(pVCpu);
6873
6874	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxContextSwitch);
6875	return VINF_SUCCESS;
6876	}
6877	}
6878
6879	/*
6880	* Interpret the read/writing of DRx.
6881	*/
6882	/** @todo Decode assist. */
6883	VBOXSTRICTRC rc = EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */);
6884	Log5(("hmR0SvmExitReadDRx: Emulated DRx access: rc=%Rrc\n", VBOXSTRICTRC_VAL(rc)));
6885	if (RT_LIKELY(rc == VINF_SUCCESS))
6886	{
6887	/* Not necessary for read accesses but whatever doesn't hurt for now, will be fixed with decode assist. */
6888	/** @todo CPUM should set this flag! */
6889	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_DR_MASK);
6890	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6891	}
6892	else
6893	Assert(rc == VERR_EM_INTERPRETER);
6894	return VBOXSTRICTRC_TODO(rc);
6895	}
6896
6897
6898	/**
6899	* \#VMEXIT handler for DRx write (SVM_EXIT_WRITE_DRx). Conditional \#VMEXIT.
6900	*/
6901	HMSVM_EXIT_DECL hmR0SvmExitWriteDRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6902	{
6903	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6904	/* For now it's the same since we interpret the instruction anyway. Will change when using of Decode Assist is implemented. */
6905	int rc = hmR0SvmExitReadDRx(pVCpu, pCtx, pSvmTransient);
6906	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitDRxWrite);
6907	STAM_COUNTER_DEC(&pVCpu->hm.s.StatExitDRxRead);
6908	return rc;
6909	}
6910
6911
6912	/**
6913	* \#VMEXIT handler for XCRx write (SVM_EXIT_XSETBV). Conditional \#VMEXIT.
6914	*/
6915	HMSVM_EXIT_DECL hmR0SvmExitXsetbv(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6916	{
6917	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6918	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6919
6920	/** @todo decode assists... */
6921	VBOXSTRICTRC rcStrict = IEMExecOne(pVCpu);
6922	if (rcStrict == VINF_IEM_RAISED_XCPT)
6923	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_XCPT_RAISED_MASK);
6924
6925	pVCpu->hm.s.fLoadSaveGuestXcr0 = (pCtx->cr4 & X86_CR4_OSXSAVE) && pCtx->aXcr[0] != ASMGetXcr0();
6926	Log4Func(("New XCR0=%#RX64 fLoadSaveGuestXcr0=%d (cr4=%RX64) rcStrict=%Rrc\n", pCtx->aXcr[0], pVCpu->hm.s.fLoadSaveGuestXcr0,
6927	pCtx->cr4, VBOXSTRICTRC_VAL(rcStrict)));
6928
6929	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6930	return VBOXSTRICTRC_TODO(rcStrict);
6931	}
6932
6933
6934	/**
6935	* \#VMEXIT handler for I/O instructions (SVM_EXIT_IOIO). Conditional \#VMEXIT.
6936	*/
6937	HMSVM_EXIT_DECL hmR0SvmExitIOInstr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
6938	{
6939	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
6940	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK
6941	\| CPUMCTX_EXTRN_SREG_MASK);
6942
6943	/* I/O operation lookup arrays. */
6944	static uint32_t const s_aIOSize[8] = { 0, 1, 2, 0, 4, 0, 0, 0 }; /* Size of the I/O accesses in bytes. */
6945	static uint32_t const s_aIOOpAnd[8] = { 0, 0xff, 0xffff, 0, 0xffffffff, 0, 0, 0 }; /* AND masks for saving
6946	the result (in AL/AX/EAX). */
6947	Log4Func(("CS:RIP=%04x:%#RX64\n", pCtx->cs.Sel, pCtx->rip));
6948
6949	PVM pVM = pVCpu->CTX_SUFF(pVM);
6950	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
6951
6952	/* Refer AMD spec. 15.10.2 "IN and OUT Behaviour" and Figure 15-2. "EXITINFO1 for IOIO Intercept" for the format. */
6953	SVMIOIOEXITINFO IoExitInfo;
6954	IoExitInfo.u = (uint32_t)pVmcb->ctrl.u64ExitInfo1;
6955	uint32_t uIOWidth = (IoExitInfo.u >> 4) & 0x7;
6956	uint32_t cbValue = s_aIOSize[uIOWidth];
6957	uint32_t uAndVal = s_aIOOpAnd[uIOWidth];
6958
6959	if (RT_UNLIKELY(!cbValue))
6960	{
6961	AssertMsgFailed(("hmR0SvmExitIOInstr: Invalid IO operation. uIOWidth=%u\n", uIOWidth));
6962	return VERR_EM_INTERPRETER;
6963	}
6964
6965	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_RIP
6966	\| CPUMCTX_EXTRN_CS
6967	\| CPUMCTX_EXTRN_RFLAGS);
6968	VBOXSTRICTRC rcStrict;
6969	PCEMEXITREC pExitRec = NULL;
6970	if ( !pVCpu->hm.s.fSingleInstruction
6971	&& !pVCpu->cpum.GstCtx.eflags.Bits.u1TF)
6972	pExitRec = EMHistoryUpdateFlagsAndTypeAndPC(pVCpu,
6973	!IoExitInfo.n.u1Str
6974	? IoExitInfo.n.u1Type == SVM_IOIO_READ
6975	? EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_IO_PORT_READ)
6976	: EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_IO_PORT_WRITE)
6977	: IoExitInfo.n.u1Type == SVM_IOIO_READ
6978	? EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_IO_PORT_STR_READ)
6979	: EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_IO_PORT_STR_WRITE),
6980	pVCpu->cpum.GstCtx.rip + pVCpu->cpum.GstCtx.cs.u64Base);
6981	if (!pExitRec)
6982	{
6983	bool fUpdateRipAlready = false;
6984	if (IoExitInfo.n.u1Str)
6985	{
6986	/* INS/OUTS - I/O String instruction. */
6987	/** @todo Huh? why can't we use the segment prefix information given by AMD-V
6988	* in EXITINFO1? Investigate once this thing is up and running. */
6989	Log4Func(("CS:RIP=%04x:%08RX64 %#06x/%u %c str\n", pCtx->cs.Sel, pCtx->rip, IoExitInfo.n.u16Port, cbValue,
6990	IoExitInfo.n.u1Type == SVM_IOIO_WRITE ? 'w' : 'r'));
6991	AssertReturn(pCtx->dx == IoExitInfo.n.u16Port, VERR_SVM_IPE_2);
6992	static IEMMODE const s_aenmAddrMode[8] =
6993	{
6994	(IEMMODE)-1, IEMMODE_16BIT, IEMMODE_32BIT, (IEMMODE)-1, IEMMODE_64BIT, (IEMMODE)-1, (IEMMODE)-1, (IEMMODE)-1
6995	};
6996	IEMMODE enmAddrMode = s_aenmAddrMode[(IoExitInfo.u >> 7) & 0x7];
6997	if (enmAddrMode != (IEMMODE)-1)
6998	{
6999	uint64_t cbInstr = pVmcb->ctrl.u64ExitInfo2 - pCtx->rip;
7000	if (cbInstr <= 15 && cbInstr >= 1)
7001	{
7002	Assert(cbInstr >= 1U + IoExitInfo.n.u1Rep);
7003	if (IoExitInfo.n.u1Type == SVM_IOIO_WRITE)
7004	{
7005	/* Don't know exactly how to detect whether u3Seg is valid, currently
7006	only enabling it for Bulldozer and later with NRIP. OS/2 broke on
7007	2384 Opterons when only checking NRIP. */
7008	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu, pCtx);
7009	if ( fSupportsNextRipSave
7010	&& pVM->cpum.ro.GuestFeatures.enmMicroarch >= kCpumMicroarch_AMD_15h_First)
7011	{
7012	AssertMsg(IoExitInfo.n.u3Seg == X86_SREG_DS \|\| cbInstr > 1U + IoExitInfo.n.u1Rep,
7013	("u32Seg=%d cbInstr=%d u1REP=%d", IoExitInfo.n.u3Seg, cbInstr, IoExitInfo.n.u1Rep));
7014	rcStrict = IEMExecStringIoWrite(pVCpu, cbValue, enmAddrMode, IoExitInfo.n.u1Rep, (uint8_t)cbInstr,
7015	IoExitInfo.n.u3Seg, true /fIoChecked/);
7016	}
7017	else if (cbInstr == 1U + IoExitInfo.n.u1Rep)
7018	rcStrict = IEMExecStringIoWrite(pVCpu, cbValue, enmAddrMode, IoExitInfo.n.u1Rep, (uint8_t)cbInstr,
7019	X86_SREG_DS, true /fIoChecked/);
7020	else
7021	rcStrict = IEMExecOne(pVCpu);
7022	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOStringWrite);
7023	}
7024	else
7025	{
7026	AssertMsg(IoExitInfo.n.u3Seg == X86_SREG_ES /=0/, ("%#x\n", IoExitInfo.n.u3Seg));
7027	rcStrict = IEMExecStringIoRead(pVCpu, cbValue, enmAddrMode, IoExitInfo.n.u1Rep, (uint8_t)cbInstr,
7028	true /fIoChecked/);
7029	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOStringRead);
7030	}
7031	}
7032	else
7033	{
7034	AssertMsgFailed(("rip=%RX64 nrip=%#RX64 cbInstr=%#RX64\n", pCtx->rip, pVmcb->ctrl.u64ExitInfo2, cbInstr));
7035	rcStrict = IEMExecOne(pVCpu);
7036	}
7037	}
7038	else
7039	{
7040	AssertMsgFailed(("IoExitInfo=%RX64\n", IoExitInfo.u));
7041	rcStrict = IEMExecOne(pVCpu);
7042	}
7043	fUpdateRipAlready = true;
7044	}
7045	else
7046	{
7047	/* IN/OUT - I/O instruction. */
7048	Assert(!IoExitInfo.n.u1Rep);
7049
7050	if (IoExitInfo.n.u1Type == SVM_IOIO_WRITE)
7051	{
7052	rcStrict = IOMIOPortWrite(pVM, pVCpu, IoExitInfo.n.u16Port, pCtx->eax & uAndVal, cbValue);
7053	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOWrite);
7054	}
7055	else
7056	{
7057	uint32_t u32Val = 0;
7058	rcStrict = IOMIOPortRead(pVM, pVCpu, IoExitInfo.n.u16Port, &u32Val, cbValue);
7059	if (IOM_SUCCESS(rcStrict))
7060	{
7061	/* Save result of I/O IN instr. in AL/AX/EAX. */
7062	/** @todo r=bird: 32-bit op size should clear high bits of rax! */
7063	pCtx->eax = (pCtx->eax & ~uAndVal) \| (u32Val & uAndVal);
7064	}
7065	else if (rcStrict == VINF_IOM_R3_IOPORT_READ)
7066	HMR0SavePendingIOPortRead(pVCpu, pCtx->rip, pVmcb->ctrl.u64ExitInfo2, IoExitInfo.n.u16Port, uAndVal, cbValue);
7067
7068	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIORead);
7069	}
7070	}
7071
7072	if (IOM_SUCCESS(rcStrict))
7073	{
7074	/* AMD-V saves the RIP of the instruction following the IO instruction in EXITINFO2. */
7075	if (!fUpdateRipAlready)
7076	pCtx->rip = pVmcb->ctrl.u64ExitInfo2;
7077
7078	/*
7079	* If any I/O breakpoints are armed, we need to check if one triggered
7080	* and take appropriate action.
7081	* Note that the I/O breakpoint type is undefined if CR4.DE is 0.
7082	*/
7083	/** @todo Optimize away the DBGFBpIsHwIoArmed call by having DBGF tell the
7084	* execution engines about whether hyper BPs and such are pending. */
7085	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_DR7);
7086	uint32_t const uDr7 = pCtx->dr[7];
7087	if (RT_UNLIKELY( ( (uDr7 & X86_DR7_ENABLED_MASK)
7088	&& X86_DR7_ANY_RW_IO(uDr7)
7089	&& (pCtx->cr4 & X86_CR4_DE))
7090	\|\| DBGFBpIsHwIoArmed(pVM)))
7091	{
7092	/* We're playing with the host CPU state here, make sure we don't preempt or longjmp. */
7093	VMMRZCallRing3Disable(pVCpu);
7094	HM_DISABLE_PREEMPT();
7095
7096	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxIoCheck);
7097	CPUMR0DebugStateMaybeSaveGuest(pVCpu, false /fDr6/);
7098
7099	VBOXSTRICTRC rcStrict2 = DBGFBpCheckIo(pVM, pVCpu, pCtx, IoExitInfo.n.u16Port, cbValue);
7100	if (rcStrict2 == VINF_EM_RAW_GUEST_TRAP)
7101	{
7102	/* Raise #DB. */
7103	pVmcb->guest.u64DR6 = pCtx->dr[6];
7104	pVmcb->guest.u64DR7 = pCtx->dr[7];
7105	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
7106	hmR0SvmSetPendingXcptDB(pVCpu);
7107	}
7108	/* rcStrict is VINF_SUCCESS, VINF_IOM_R3_IOPORT_COMMIT_WRITE, or in [VINF_EM_FIRST..VINF_EM_LAST],
7109	however we can ditch VINF_IOM_R3_IOPORT_COMMIT_WRITE as it has VMCPU_FF_IOM as backup. */
7110	else if ( rcStrict2 != VINF_SUCCESS
7111	&& (rcStrict == VINF_SUCCESS \|\| rcStrict2 < rcStrict))
7112	rcStrict = rcStrict2;
7113	AssertCompile(VINF_EM_LAST < VINF_IOM_R3_IOPORT_COMMIT_WRITE);
7114
7115	HM_RESTORE_PREEMPT();
7116	VMMRZCallRing3Enable(pVCpu);
7117	}
7118
7119	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
7120	}
7121
7122	#ifdef VBOX_STRICT
7123	if (rcStrict == VINF_IOM_R3_IOPORT_READ)
7124	Assert(IoExitInfo.n.u1Type == SVM_IOIO_READ);
7125	else if (rcStrict == VINF_IOM_R3_IOPORT_WRITE \|\| rcStrict == VINF_IOM_R3_IOPORT_COMMIT_WRITE)
7126	Assert(IoExitInfo.n.u1Type == SVM_IOIO_WRITE);
7127	else
7128	{
7129	/** @todo r=bird: This is missing a bunch of VINF_EM_FIRST..VINF_EM_LAST
7130	* statuses, that the VMM device and some others may return. See
7131	* IOM_SUCCESS() for guidance. */
7132	AssertMsg( RT_FAILURE(rcStrict)
7133	\|\| rcStrict == VINF_SUCCESS
7134	\|\| rcStrict == VINF_EM_RAW_EMULATE_INSTR
7135	\|\| rcStrict == VINF_EM_DBG_BREAKPOINT
7136	\|\| rcStrict == VINF_EM_RAW_GUEST_TRAP
7137	\|\| rcStrict == VINF_EM_RAW_TO_R3
7138	\|\| rcStrict == VINF_TRPM_XCPT_DISPATCHED
7139	\|\| rcStrict == VINF_EM_TRIPLE_FAULT, ("%Rrc\n", VBOXSTRICTRC_VAL(rcStrict)));
7140	}
7141	#endif
7142	}
7143	else
7144	{
7145	/*
7146	* Frequent exit or something needing probing. Get state and call EMHistoryExec.
7147	*/
7148	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7149	STAM_COUNTER_INC(!IoExitInfo.n.u1Str
7150	? IoExitInfo.n.u1Type == SVM_IOIO_WRITE ? &pVCpu->hm.s.StatExitIOWrite : &pVCpu->hm.s.StatExitIORead
7151	: IoExitInfo.n.u1Type == SVM_IOIO_WRITE ? &pVCpu->hm.s.StatExitIOStringWrite : &pVCpu->hm.s.StatExitIOStringRead);
7152	Log4(("IOExit/%u: %04x:%08RX64: %s%s%s %#x LB %u -> EMHistoryExec\n",
7153	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip, IoExitInfo.n.u1Rep ? "REP " : "",
7154	IoExitInfo.n.u1Type == SVM_IOIO_WRITE ? "OUT" : "IN", IoExitInfo.n.u1Str ? "S" : "", IoExitInfo.n.u16Port, uIOWidth));
7155
7156	rcStrict = EMHistoryExec(pVCpu, pExitRec, 0);
7157	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
7158
7159	Log4(("IOExit/%u: %04x:%08RX64: EMHistoryExec -> %Rrc + %04x:%08RX64\n",
7160	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip,
7161	VBOXSTRICTRC_VAL(rcStrict), pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip));
7162	}
7163	return VBOXSTRICTRC_TODO(rcStrict);
7164	}
7165
7166
7167	/**
7168	* \#VMEXIT handler for Nested Page-faults (SVM_EXIT_NPF). Conditional \#VMEXIT.
7169	*/
7170	HMSVM_EXIT_DECL hmR0SvmExitNestedPF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7171	{
7172	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7173	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7174	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
7175
7176	PVM pVM = pVCpu->CTX_SUFF(pVM);
7177	Assert(pVM->hm.s.fNestedPaging);
7178
7179	/* See AMD spec. 15.25.6 "Nested versus Guest Page Faults, Fault Ordering" for VMCB details for #NPF. */
7180	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7181	RTGCPHYS GCPhysFaultAddr = pVmcb->ctrl.u64ExitInfo2;
7182	uint32_t u32ErrCode = pVmcb->ctrl.u64ExitInfo1; /* Note! High bits in EXITINFO1 may contain additional info and are
7183	thus intentionally not copied into u32ErrCode. */
7184
7185	Log4Func(("#NPF at CS:RIP=%04x:%#RX64 GCPhysFaultAddr=%RGp ErrCode=%#x \n", pCtx->cs.Sel, pCtx->rip, GCPhysFaultAddr,
7186	u32ErrCode));
7187
7188	/*
7189	* TPR patching for 32-bit guests, using the reserved bit in the page tables for MMIO regions.
7190	*/
7191	if ( pVM->hm.s.fTprPatchingAllowed
7192	&& (GCPhysFaultAddr & PAGE_OFFSET_MASK) == XAPIC_OFF_TPR
7193	&& ( !(u32ErrCode & X86_TRAP_PF_P) /* Not present */
7194	\|\| (u32ErrCode & (X86_TRAP_PF_P \| X86_TRAP_PF_RSVD)) == (X86_TRAP_PF_P \| X86_TRAP_PF_RSVD)) /* MMIO page. */
7195	&& !CPUMIsGuestInSvmNestedHwVirtMode(pCtx)
7196	&& !CPUMIsGuestInLongModeEx(pCtx)
7197	&& !CPUMGetGuestCPL(pVCpu)
7198	&& pVM->hm.s.cPatches < RT_ELEMENTS(pVM->hm.s.aPatches))
7199	{
7200	RTGCPHYS GCPhysApicBase = APICGetBaseMsrNoCheck(pVCpu);
7201	GCPhysApicBase &= PAGE_BASE_GC_MASK;
7202
7203	if (GCPhysFaultAddr == GCPhysApicBase + XAPIC_OFF_TPR)
7204	{
7205	/* Only attempt to patch the instruction once. */
7206	PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
7207	if (!pPatch)
7208	return VINF_EM_HM_PATCH_TPR_INSTR;
7209	}
7210	}
7211
7212	/*
7213	* Determine the nested paging mode.
7214	*/
7215	PGMMODE enmNestedPagingMode;
7216	#if HC_ARCH_BITS == 32
7217	if (CPUMIsGuestInLongModeEx(pCtx))
7218	enmNestedPagingMode = PGMMODE_AMD64_NX;
7219	else
7220	#endif
7221	enmNestedPagingMode = PGMGetHostMode(pVM);
7222
7223	/*
7224	* MMIO optimization using the reserved (RSVD) bit in the guest page tables for MMIO pages.
7225	*/
7226	Assert((u32ErrCode & (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P)) != X86_TRAP_PF_RSVD);
7227	if ((u32ErrCode & (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P)) == (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P))
7228	{
7229	/*
7230	* If event delivery causes an MMIO #NPF, go back to instruction emulation as otherwise
7231	* injecting the original pending event would most likely cause the same MMIO #NPF.
7232	*/
7233	if (pVCpu->hm.s.Event.fPending)
7234	return VINF_EM_RAW_INJECT_TRPM_EVENT;
7235
7236	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_RIP
7237	\| CPUMCTX_EXTRN_CS);
7238	VBOXSTRICTRC rcStrict;
7239	PCEMEXITREC pExitRec = EMHistoryUpdateFlagsAndTypeAndPC(pVCpu,
7240	EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_MMIO),
7241	pVCpu->cpum.GstCtx.rip + pVCpu->cpum.GstCtx.cs.u64Base);
7242	if (!pExitRec)
7243	{
7244
7245	rcStrict = PGMR0Trap0eHandlerNPMisconfig(pVM, pVCpu, enmNestedPagingMode, CPUMCTX2CORE(pCtx), GCPhysFaultAddr,
7246	u32ErrCode);
7247
7248	/*
7249	* If we succeed, resume guest execution.
7250	*
7251	* If we fail in interpreting the instruction because we couldn't get the guest
7252	* physical address of the page containing the instruction via the guest's page
7253	* tables (we would invalidate the guest page in the host TLB), resume execution
7254	* which would cause a guest page fault to let the guest handle this weird case.
7255	*
7256	* See @bugref{6043}.
7257	*/
7258	if ( rcStrict == VINF_SUCCESS
7259	\|\| rcStrict == VERR_PAGE_TABLE_NOT_PRESENT
7260	\|\| rcStrict == VERR_PAGE_NOT_PRESENT)
7261	{
7262	/* Successfully handled MMIO operation. */
7263	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
7264	rcStrict = VINF_SUCCESS;
7265	}
7266	}
7267	else
7268	{
7269	/*
7270	* Frequent exit or something needing probing. Get state and call EMHistoryExec.
7271	*/
7272	Assert(pCtx == &pVCpu->cpum.GstCtx);
7273	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7274	Log4(("EptMisscfgExit/%u: %04x:%08RX64: %RGp -> EMHistoryExec\n",
7275	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip, GCPhysFaultAddr));
7276
7277	rcStrict = EMHistoryExec(pVCpu, pExitRec, 0);
7278	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
7279
7280	Log4(("EptMisscfgExit/%u: %04x:%08RX64: EMHistoryExec -> %Rrc + %04x:%08RX64\n",
7281	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip,
7282	VBOXSTRICTRC_VAL(rcStrict), pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip));
7283	}
7284	return VBOXSTRICTRC_TODO(rcStrict);
7285	}
7286
7287	TRPMAssertXcptPF(pVCpu, GCPhysFaultAddr, u32ErrCode);
7288	int rc = PGMR0Trap0eHandlerNestedPaging(pVM, pVCpu, enmNestedPagingMode, u32ErrCode, CPUMCTX2CORE(pCtx), GCPhysFaultAddr);
7289	TRPMResetTrap(pVCpu);
7290
7291	Log4Func(("#NPF: PGMR0Trap0eHandlerNestedPaging returns %Rrc CS:RIP=%04x:%#RX64\n", rc, pCtx->cs.Sel, pCtx->rip));
7292
7293	/*
7294	* Same case as PGMR0Trap0eHandlerNPMisconfig(). See comment above, @bugref{6043}.
7295	*/
7296	if ( rc == VINF_SUCCESS
7297	\|\| rc == VERR_PAGE_TABLE_NOT_PRESENT
7298	\|\| rc == VERR_PAGE_NOT_PRESENT)
7299	{
7300	/* We've successfully synced our shadow page tables. */
7301	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPF);
7302	rc = VINF_SUCCESS;
7303	}
7304
7305	return rc;
7306	}
7307
7308
7309	/**
7310	* \#VMEXIT handler for virtual interrupt (SVM_EXIT_VINTR). Conditional
7311	* \#VMEXIT.
7312	*/
7313	HMSVM_EXIT_DECL hmR0SvmExitVIntr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7314	{
7315	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7316	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(pCtx);
7317
7318	/* Indicate that we no longer need to #VMEXIT when the guest is ready to receive NMIs, it is now ready. */
7319	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7320	hmR0SvmClearIntWindowExiting(pVCpu, pVmcb, pCtx);
7321
7322	/* Deliver the pending interrupt via hmR0SvmEvaluatePendingEvent() and resume guest execution. */
7323	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIntWindow);
7324	return VINF_SUCCESS;
7325	}
7326
7327
7328	/**
7329	* \#VMEXIT handler for task switches (SVM_EXIT_TASK_SWITCH). Conditional
7330	* \#VMEXIT.
7331	*/
7332	HMSVM_EXIT_DECL hmR0SvmExitTaskSwitch(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7333	{
7334	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7335	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
7336
7337	#ifndef HMSVM_ALWAYS_TRAP_TASK_SWITCH
7338	Assert(!pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
7339	#endif
7340
7341	/* Check if this task-switch occurred while delivering an event through the guest IDT. */
7342	if (pVCpu->hm.s.Event.fPending) /* Can happen with exceptions/NMI. See @bugref{8411}. */
7343	{
7344	/*
7345	* AMD-V provides us with the exception which caused the TS; we collect
7346	* the information in the call to hmR0SvmCheckExitDueToEventDelivery().
7347	*/
7348	Log4Func(("TS occurred during event delivery\n"));
7349	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitTaskSwitch);
7350	return VINF_EM_RAW_INJECT_TRPM_EVENT;
7351	}
7352
7353	/** @todo Emulate task switch someday, currently just going back to ring-3 for
7354	* emulation. */
7355	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitTaskSwitch);
7356	return VERR_EM_INTERPRETER;
7357	}
7358
7359
7360	/**
7361	* \#VMEXIT handler for VMMCALL (SVM_EXIT_VMMCALL). Conditional \#VMEXIT.
7362	*/
7363	HMSVM_EXIT_DECL hmR0SvmExitVmmCall(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7364	{
7365	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7366	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7367
7368	if (pVCpu->CTX_SUFF(pVM)->hm.s.fTprPatchingAllowed)
7369	{
7370	int rc = hmSvmEmulateMovTpr(pVCpu, pCtx);
7371	if (rc != VERR_NOT_FOUND)
7372	{
7373	Log4Func(("hmSvmEmulateMovTpr returns %Rrc\n", rc));
7374	return rc;
7375	}
7376	}
7377
7378	if (EMAreHypercallInstructionsEnabled(pVCpu))
7379	{
7380	VBOXSTRICTRC rcStrict = GIMHypercall(pVCpu, pCtx);
7381	if (RT_SUCCESS(rcStrict))
7382	{
7383	/* Only update the RIP if we're continuing guest execution and not in the case
7384	of say VINF_GIM_R3_HYPERCALL. */
7385	if (rcStrict == VINF_SUCCESS)
7386	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 3 /* cbInstr */);
7387
7388	return VBOXSTRICTRC_VAL(rcStrict);
7389	}
7390	else
7391	Log4Func(("GIMHypercall returns %Rrc -> #UD\n", VBOXSTRICTRC_VAL(rcStrict)));
7392	}
7393
7394	hmR0SvmSetPendingXcptUD(pVCpu);
7395	return VINF_SUCCESS;
7396	}
7397
7398
7399	/**
7400	* \#VMEXIT handler for VMMCALL (SVM_EXIT_VMMCALL). Conditional \#VMEXIT.
7401	*/
7402	HMSVM_EXIT_DECL hmR0SvmExitPause(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7403	{
7404	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7405	hmR0SvmAdvanceRipHwAssist(pVCpu, pCtx, 2);
7406	/** @todo The guest has likely hit a contended spinlock. We might want to
7407	* poke a schedule different guest VCPU. */
7408	return VINF_EM_RAW_INTERRUPT;
7409	}
7410
7411
7412	/**
7413	* \#VMEXIT handler for FERR intercept (SVM_EXIT_FERR_FREEZE). Conditional
7414	* \#VMEXIT.
7415	*/
7416	HMSVM_EXIT_DECL hmR0SvmExitFerrFreeze(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7417	{
7418	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7419	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR0);
7420	Assert(!(pCtx->cr0 & X86_CR0_NE));
7421
7422	Log4Func(("Raising IRQ 13 in response to #FERR\n"));
7423	return PDMIsaSetIrq(pVCpu->CTX_SUFF(pVM), 13 /* u8Irq /, 1 / u8Level /, 0 / uTagSrc */);
7424	}
7425
7426
7427	/**
7428	* \#VMEXIT handler for IRET (SVM_EXIT_IRET). Conditional \#VMEXIT.
7429	*/
7430	HMSVM_EXIT_DECL hmR0SvmExitIret(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7431	{
7432	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7433
7434	/* Clear NMI blocking. */
7435	if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_BLOCK_NMIS))
7436	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_BLOCK_NMIS);
7437
7438	/* Indicate that we no longer need to #VMEXIT when the guest is ready to receive NMIs, it is now ready. */
7439	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7440	hmR0SvmClearCtrlIntercept(pVCpu, pCtx, pVmcb, SVM_CTRL_INTERCEPT_IRET);
7441
7442	/* Deliver the pending NMI via hmR0SvmEvaluatePendingEvent() and resume guest execution. */
7443	return VINF_SUCCESS;
7444	}
7445
7446
7447	/**
7448	* \#VMEXIT handler for page-fault exceptions (SVM_EXIT_XCPT_14).
7449	* Conditional \#VMEXIT.
7450	*/
7451	HMSVM_EXIT_DECL hmR0SvmExitXcptPF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7452	{
7453	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7454	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7455	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
7456
7457	/* See AMD spec. 15.12.15 "#PF (Page Fault)". */
7458	PVM pVM = pVCpu->CTX_SUFF(pVM);
7459	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7460	uint32_t uErrCode = pVmcb->ctrl.u64ExitInfo1;
7461	uint64_t const uFaultAddress = pVmcb->ctrl.u64ExitInfo2;
7462
7463	#if defined(HMSVM_ALWAYS_TRAP_ALL_XCPTS) \|\| defined(HMSVM_ALWAYS_TRAP_PF)
7464	if (pVM->hm.s.fNestedPaging)
7465	{
7466	pVCpu->hm.s.Event.fPending = false; /* In case it's a contributory or vectoring #PF. */
7467	if ( !pSvmTransient->fVectoringDoublePF
7468	\|\| CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
7469	{
7470	/* A genuine guest #PF, reflect it to the guest. */
7471	hmR0SvmSetPendingXcptPF(pVCpu, pCtx, uErrCode, uFaultAddress);
7472	Log4Func(("#PF: Guest page fault at %04X:%RGv FaultAddr=%RX64 ErrCode=%#x\n", pCtx->cs.Sel, (RTGCPTR)pCtx->rip,
7473	uFaultAddress, uErrCode));
7474	}
7475	else
7476	{
7477	/* A guest page-fault occurred during delivery of a page-fault. Inject #DF. */
7478	hmR0SvmSetPendingXcptDF(pVCpu);
7479	Log4Func(("Pending #DF due to vectoring #PF. NP\n"));
7480	}
7481	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestPF);
7482	return VINF_SUCCESS;
7483	}
7484	#endif
7485
7486	Assert(!pVM->hm.s.fNestedPaging);
7487
7488	/*
7489	* TPR patching shortcut for APIC TPR reads and writes; only applicable to 32-bit guests.
7490	*/
7491	if ( pVM->hm.s.fTprPatchingAllowed
7492	&& (uFaultAddress & 0xfff) == XAPIC_OFF_TPR
7493	&& !(uErrCode & X86_TRAP_PF_P) /* Not present. */
7494	&& !CPUMIsGuestInSvmNestedHwVirtMode(pCtx)
7495	&& !CPUMIsGuestInLongModeEx(pCtx)
7496	&& !CPUMGetGuestCPL(pVCpu)
7497	&& pVM->hm.s.cPatches < RT_ELEMENTS(pVM->hm.s.aPatches))
7498	{
7499	RTGCPHYS GCPhysApicBase;
7500	GCPhysApicBase = APICGetBaseMsrNoCheck(pVCpu);
7501	GCPhysApicBase &= PAGE_BASE_GC_MASK;
7502
7503	/* Check if the page at the fault-address is the APIC base. */
7504	RTGCPHYS GCPhysPage;
7505	int rc2 = PGMGstGetPage(pVCpu, (RTGCPTR)uFaultAddress, NULL /* pfFlags */, &GCPhysPage);
7506	if ( rc2 == VINF_SUCCESS
7507	&& GCPhysPage == GCPhysApicBase)
7508	{
7509	/* Only attempt to patch the instruction once. */
7510	PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
7511	if (!pPatch)
7512	return VINF_EM_HM_PATCH_TPR_INSTR;
7513	}
7514	}
7515
7516	Log4Func(("#PF: uFaultAddress=%#RX64 CS:RIP=%#04x:%#RX64 uErrCode %#RX32 cr3=%#RX64\n", uFaultAddress, pCtx->cs.Sel,
7517	pCtx->rip, uErrCode, pCtx->cr3));
7518
7519	/*
7520	* If it's a vectoring #PF, emulate injecting the original event injection as
7521	* PGMTrap0eHandler() is incapable of differentiating between instruction emulation and
7522	* event injection that caused a #PF. See @bugref{6607}.
7523	*/
7524	if (pSvmTransient->fVectoringPF)
7525	{
7526	Assert(pVCpu->hm.s.Event.fPending);
7527	return VINF_EM_RAW_INJECT_TRPM_EVENT;
7528	}
7529
7530	TRPMAssertXcptPF(pVCpu, uFaultAddress, uErrCode);
7531	int rc = PGMTrap0eHandler(pVCpu, uErrCode, CPUMCTX2CORE(pCtx), (RTGCPTR)uFaultAddress);
7532
7533	Log4Func(("#PF: rc=%Rrc\n", rc));
7534
7535	if (rc == VINF_SUCCESS)
7536	{
7537	/* Successfully synced shadow pages tables or emulated an MMIO instruction. */
7538	TRPMResetTrap(pVCpu);
7539	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPF);
7540	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
7541	return rc;
7542	}
7543
7544	if (rc == VINF_EM_RAW_GUEST_TRAP)
7545	{
7546	pVCpu->hm.s.Event.fPending = false; /* In case it's a contributory or vectoring #PF. */
7547
7548	/*
7549	* If a nested-guest delivers a #PF and that causes a #PF which is -not- a shadow #PF,
7550	* we should simply forward the #PF to the guest and is up to the nested-hypervisor to
7551	* determine whether it is a nested-shadow #PF or a #DF, see @bugref{7243#c121}.
7552	*/
7553	if ( !pSvmTransient->fVectoringDoublePF
7554	\|\| CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
7555	{
7556	/* It's a guest (or nested-guest) page fault and needs to be reflected. */
7557	uErrCode = TRPMGetErrorCode(pVCpu); /* The error code might have been changed. */
7558	TRPMResetTrap(pVCpu);
7559
7560	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
7561	/* If the nested-guest is intercepting #PFs, cause a #PF #VMEXIT. */
7562	if ( CPUMIsGuestInSvmNestedHwVirtMode(pCtx)
7563	&& HMIsGuestSvmXcptInterceptSet(pVCpu, X86_XCPT_PF))
7564	return VBOXSTRICTRC_TODO(IEMExecSvmVmexit(pVCpu, SVM_EXIT_XCPT_PF, uErrCode, uFaultAddress));
7565	#endif
7566
7567	hmR0SvmSetPendingXcptPF(pVCpu, pCtx, uErrCode, uFaultAddress);
7568	}
7569	else
7570	{
7571	/* A guest page-fault occurred during delivery of a page-fault. Inject #DF. */
7572	TRPMResetTrap(pVCpu);
7573	hmR0SvmSetPendingXcptDF(pVCpu);
7574	Log4Func(("#PF: Pending #DF due to vectoring #PF\n"));
7575	}
7576
7577	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestPF);
7578	return VINF_SUCCESS;
7579	}
7580
7581	TRPMResetTrap(pVCpu);
7582	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPFEM);
7583	return rc;
7584	}
7585
7586
7587	/**
7588	* \#VMEXIT handler for undefined opcode (SVM_EXIT_XCPT_6).
7589	* Conditional \#VMEXIT.
7590	*/
7591	HMSVM_EXIT_DECL hmR0SvmExitXcptUD(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7592	{
7593	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7594	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(pCtx);
7595
7596	/* Paranoia; Ensure we cannot be called as a result of event delivery. */
7597	PSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
7598	Assert(!pVmcb->ctrl.ExitIntInfo.n.u1Valid); NOREF(pVmcb);
7599
7600	int rc = VERR_SVM_UNEXPECTED_XCPT_EXIT;
7601	if (pVCpu->hm.s.fGIMTrapXcptUD)
7602	{
7603	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7604	uint8_t cbInstr = 0;
7605	VBOXSTRICTRC rcStrict = GIMXcptUD(pVCpu, pCtx, NULL /* pDis */, &cbInstr);
7606	if (rcStrict == VINF_SUCCESS)
7607	{
7608	/* #UD #VMEXIT does not have valid NRIP information, manually advance RIP. See @bugref{7270#c170}. */
7609	hmR0SvmAdvanceRipDumb(pVCpu, pCtx, cbInstr);
7610	rc = VINF_SUCCESS;
7611	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
7612	}
7613	else if (rcStrict == VINF_GIM_HYPERCALL_CONTINUING)
7614	rc = VINF_SUCCESS;
7615	else if (rcStrict == VINF_GIM_R3_HYPERCALL)
7616	rc = VINF_GIM_R3_HYPERCALL;
7617	else
7618	Assert(RT_FAILURE(VBOXSTRICTRC_VAL(rcStrict)));
7619	}
7620
7621	/* If the GIM #UD exception handler didn't succeed for some reason or wasn't needed, raise #UD. */
7622	if (RT_FAILURE(rc))
7623	{
7624	hmR0SvmSetPendingXcptUD(pVCpu);
7625	rc = VINF_SUCCESS;
7626	}
7627
7628	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestUD);
7629	return rc;
7630	}
7631
7632
7633	/**
7634	* \#VMEXIT handler for math-fault exceptions (SVM_EXIT_XCPT_16).
7635	* Conditional \#VMEXIT.
7636	*/
7637	HMSVM_EXIT_DECL hmR0SvmExitXcptMF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7638	{
7639	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7640	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7641
7642	/* Paranoia; Ensure we cannot be called as a result of event delivery. */
7643	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7644	Assert(!pVmcb->ctrl.ExitIntInfo.n.u1Valid); NOREF(pVmcb);
7645
7646	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestMF);
7647
7648	if (!(pCtx->cr0 & X86_CR0_NE))
7649	{
7650	PVM pVM = pVCpu->CTX_SUFF(pVM);
7651	PDISSTATE pDis = &pVCpu->hm.s.DisState;
7652	unsigned cbOp;
7653	int rc = EMInterpretDisasCurrent(pVM, pVCpu, pDis, &cbOp);
7654	if (RT_SUCCESS(rc))
7655	{
7656	/* Convert a #MF into a FERR -> IRQ 13. See @bugref{6117}. */
7657	rc = PDMIsaSetIrq(pVCpu->CTX_SUFF(pVM), 13 /* u8Irq /, 1 / u8Level /, 0 / uTagSrc */);
7658	if (RT_SUCCESS(rc))
7659	pCtx->rip += cbOp;
7660	}
7661	else
7662	Log4Func(("EMInterpretDisasCurrent returned %Rrc uOpCode=%#x\n", rc, pDis->pCurInstr->uOpcode));
7663	return rc;
7664	}
7665
7666	hmR0SvmSetPendingXcptMF(pVCpu);
7667	return VINF_SUCCESS;
7668	}
7669
7670
7671	/**
7672	* \#VMEXIT handler for debug exceptions (SVM_EXIT_XCPT_1). Conditional
7673	* \#VMEXIT.
7674	*/
7675	HMSVM_EXIT_DECL hmR0SvmExitXcptDB(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7676	{
7677	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7678	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7679	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
7680
7681	if (RT_UNLIKELY(pVCpu->hm.s.Event.fPending))
7682	{
7683	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectPendingInterpret);
7684	return VINF_EM_RAW_INJECT_TRPM_EVENT;
7685	}
7686
7687	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestDB);
7688
7689	/*
7690	* This can be a fault-type #DB (instruction breakpoint) or a trap-type #DB (data
7691	* breakpoint). However, for both cases DR6 and DR7 are updated to what the exception
7692	* handler expects. See AMD spec. 15.12.2 "#DB (Debug)".
7693	*/
7694	PVM pVM = pVCpu->CTX_SUFF(pVM);
7695	PSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
7696	int rc = DBGFRZTrap01Handler(pVM, pVCpu, CPUMCTX2CORE(pCtx), pVmcb->guest.u64DR6, pVCpu->hm.s.fSingleInstruction);
7697	if (rc == VINF_EM_RAW_GUEST_TRAP)
7698	{
7699	Log5(("hmR0SvmExitXcptDB: DR6=%#RX64 -> guest trap\n", pVmcb->guest.u64DR6));
7700	if (CPUMIsHyperDebugStateActive(pVCpu))
7701	CPUMSetGuestDR6(pVCpu, CPUMGetGuestDR6(pVCpu) \| pVmcb->guest.u64DR6);
7702
7703	/* Reflect the exception back to the guest. */
7704	hmR0SvmSetPendingXcptDB(pVCpu);
7705	rc = VINF_SUCCESS;
7706	}
7707
7708	/*
7709	* Update DR6.
7710	*/
7711	if (CPUMIsHyperDebugStateActive(pVCpu))
7712	{
7713	Log5(("hmR0SvmExitXcptDB: DR6=%#RX64 -> %Rrc\n", pVmcb->guest.u64DR6, rc));
7714	pVmcb->guest.u64DR6 = X86_DR6_INIT_VAL;
7715	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
7716	}
7717	else
7718	{
7719	AssertMsg(rc == VINF_SUCCESS, ("rc=%Rrc\n", rc));
7720	Assert(!pVCpu->hm.s.fSingleInstruction && !DBGFIsStepping(pVCpu));
7721	}
7722
7723	return rc;
7724	}
7725
7726
7727	/**
7728	* \#VMEXIT handler for alignment check exceptions (SVM_EXIT_XCPT_17).
7729	* Conditional \#VMEXIT.
7730	*/
7731	HMSVM_EXIT_DECL hmR0SvmExitXcptAC(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7732	{
7733	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7734	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
7735
7736	SVMEVENT Event;
7737	Event.u = 0;
7738	Event.n.u1Valid = 1;
7739	Event.n.u3Type = SVM_EVENT_EXCEPTION;
7740	Event.n.u8Vector = X86_XCPT_AC;
7741	Event.n.u1ErrorCodeValid = 1;
7742	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
7743	return VINF_SUCCESS;
7744	}
7745
7746
7747	/**
7748	* \#VMEXIT handler for breakpoint exceptions (SVM_EXIT_XCPT_3).
7749	* Conditional \#VMEXIT.
7750	*/
7751	HMSVM_EXIT_DECL hmR0SvmExitXcptBP(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7752	{
7753	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7754	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7755	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
7756
7757	int rc = DBGFRZTrap03Handler(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
7758	if (rc == VINF_EM_RAW_GUEST_TRAP)
7759	{
7760	SVMEVENT Event;
7761	Event.u = 0;
7762	Event.n.u1Valid = 1;
7763	Event.n.u3Type = SVM_EVENT_EXCEPTION;
7764	Event.n.u8Vector = X86_XCPT_BP;
7765	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
7766	}
7767
7768	Assert(rc == VINF_SUCCESS \|\| rc == VINF_EM_RAW_GUEST_TRAP \|\| rc == VINF_EM_DBG_BREAKPOINT);
7769	return rc;
7770	}
7771
7772
7773	#if defined(HMSVM_ALWAYS_TRAP_ALL_XCPTS) \|\| defined(VBOX_WITH_NESTED_HWVIRT_SVM)
7774	/**
7775	* \#VMEXIT handler for generic exceptions. Conditional \#VMEXIT.
7776	*/
7777	HMSVM_EXIT_DECL hmR0SvmExitXcptGeneric(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7778	{
7779	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7780	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
7781
7782	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7783	uint8_t const uVector = pVmcb->ctrl.u64ExitCode - SVM_EXIT_XCPT_0;
7784	uint32_t const uErrCode = pVmcb->ctrl.u64ExitInfo1;
7785	Assert(pSvmTransient->u64ExitCode == pVmcb->ctrl.u64ExitCode);
7786	Assert(uVector <= X86_XCPT_LAST);
7787	Log4Func(("uVector=%#x uErrCode=%u\n", uVector, uErrCode));
7788
7789	SVMEVENT Event;
7790	Event.u = 0;
7791	Event.n.u1Valid = 1;
7792	Event.n.u3Type = SVM_EVENT_EXCEPTION;
7793	Event.n.u8Vector = uVector;
7794	switch (uVector)
7795	{
7796	/* Shouldn't be here for reflecting #PFs (among other things, the fault address isn't passed along). */
7797	case X86_XCPT_PF: AssertMsgFailed(("hmR0SvmExitXcptGeneric: Unexpected exception")); return VERR_SVM_IPE_5;
7798	case X86_XCPT_DF:
7799	case X86_XCPT_TS:
7800	case X86_XCPT_NP:
7801	case X86_XCPT_SS:
7802	case X86_XCPT_GP:
7803	case X86_XCPT_AC:
7804	{
7805	Event.n.u1ErrorCodeValid = 1;
7806	Event.n.u32ErrorCode = uErrCode;
7807	break;
7808	}
7809	}
7810
7811	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
7812	return VINF_SUCCESS;
7813	}
7814	#endif
7815
7816	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
7817	/**
7818	* \#VMEXIT handler for CLGI (SVM_EXIT_CLGI). Conditional \#VMEXIT.
7819	*/
7820	HMSVM_EXIT_DECL hmR0SvmExitClgi(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7821	{
7822	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7823	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK
7824	\| CPUMCTX_EXTRN_HWVIRT);
7825
7826	#ifdef VBOX_STRICT
7827	PCSVMVMCB pVmcbTmp = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7828	Assert(pVmcbTmp);
7829	Assert(!pVmcbTmp->ctrl.IntCtrl.n.u1VGifEnable);
7830	RT_NOREF(pVmcbTmp);
7831	#endif
7832
7833	uint8_t const cbInstr = hmR0SvmGetInstrLengthHwAssist(pVCpu, pCtx, 3);
7834	VBOXSTRICTRC rcStrict = IEMExecDecodedClgi(pVCpu, cbInstr);
7835	if (rcStrict == VINF_SUCCESS)
7836	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_HWVIRT);
7837	return VBOXSTRICTRC_VAL(rcStrict);
7838	}
7839
7840
7841	/**
7842	* \#VMEXIT handler for STGI (SVM_EXIT_STGI). Conditional \#VMEXIT.
7843	*/
7844	HMSVM_EXIT_DECL hmR0SvmExitStgi(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7845	{
7846	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7847	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK
7848	\| CPUMCTX_EXTRN_HWVIRT);
7849
7850	/*
7851	* When VGIF is not used we always intercept STGI instructions. When VGIF is used,
7852	* we only intercept STGI when events are pending for GIF to become 1.
7853	*/
7854	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7855	if (pVmcb->ctrl.IntCtrl.n.u1VGifEnable)
7856	hmR0SvmClearCtrlIntercept(pVCpu, pCtx, pVmcb, SVM_CTRL_INTERCEPT_STGI);
7857
7858	uint8_t const cbInstr = hmR0SvmGetInstrLengthHwAssist(pVCpu, pCtx, 3);
7859	VBOXSTRICTRC rcStrict = IEMExecDecodedStgi(pVCpu, cbInstr);
7860	if (rcStrict == VINF_SUCCESS)
7861	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_HWVIRT);
7862	return VBOXSTRICTRC_VAL(rcStrict);
7863	}
7864
7865
7866	/**
7867	* \#VMEXIT handler for VMLOAD (SVM_EXIT_VMLOAD). Conditional \#VMEXIT.
7868	*/
7869	HMSVM_EXIT_DECL hmR0SvmExitVmload(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7870	{
7871	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7872	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK
7873	\| CPUMCTX_EXTRN_FS
7874	\| CPUMCTX_EXTRN_GS
7875	\| CPUMCTX_EXTRN_TR
7876	\| CPUMCTX_EXTRN_LDTR
7877	\| CPUMCTX_EXTRN_KERNEL_GS_BASE
7878	\| CPUMCTX_EXTRN_SYSCALL_MSRS
7879	\| CPUMCTX_EXTRN_SYSENTER_MSRS);
7880
7881	#ifdef VBOX_STRICT
7882	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7883	Assert(pVmcb);
7884	Assert(!pVmcb->ctrl.LbrVirt.n.u1VirtVmsaveVmload);
7885	RT_NOREF(pVmcb);
7886	#endif
7887
7888	uint8_t const cbInstr = hmR0SvmGetInstrLengthHwAssist(pVCpu, pCtx, 3);
7889	VBOXSTRICTRC rcStrict = IEMExecDecodedVmload(pVCpu, cbInstr);
7890	if (rcStrict == VINF_SUCCESS)
7891	{
7892	/* We skip flagging changes made to LSTAR, STAR, SFMASK and other MSRs as they are always re-loaded. */
7893	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_FS
7894	\| HM_CHANGED_GUEST_GS
7895	\| HM_CHANGED_GUEST_TR
7896	\| HM_CHANGED_GUEST_LDTR
7897	\| HM_CHANGED_GUEST_SYSENTER_CS_MSR
7898	\| HM_CHANGED_GUEST_SYSENTER_EIP_MSR
7899	\| HM_CHANGED_GUEST_SYSENTER_ESP_MSR);
7900	}
7901	return VBOXSTRICTRC_VAL(rcStrict);
7902	}
7903
7904
7905	/**
7906	* \#VMEXIT handler for VMSAVE (SVM_EXIT_VMSAVE). Conditional \#VMEXIT.
7907	*/
7908	HMSVM_EXIT_DECL hmR0SvmExitVmsave(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7909	{
7910	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7911	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
7912
7913	#ifdef VBOX_STRICT
7914	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu, pCtx);
7915	Assert(pVmcb);
7916	Assert(!pVmcb->ctrl.LbrVirt.n.u1VirtVmsaveVmload);
7917	RT_NOREF(pVmcb);
7918	#endif
7919
7920	uint8_t const cbInstr = hmR0SvmGetInstrLengthHwAssist(pVCpu, pCtx, 3);
7921	VBOXSTRICTRC rcStrict = IEMExecDecodedVmsave(pVCpu, cbInstr);
7922	return VBOXSTRICTRC_VAL(rcStrict);
7923	}
7924
7925
7926	/**
7927	* \#VMEXIT handler for INVLPGA (SVM_EXIT_INVLPGA). Conditional \#VMEXIT.
7928	*/
7929	HMSVM_EXIT_DECL hmR0SvmExitInvlpga(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7930	{
7931	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7932	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
7933
7934	uint8_t const cbInstr = hmR0SvmGetInstrLengthHwAssist(pVCpu, pCtx, 3);
7935	VBOXSTRICTRC rcStrict = IEMExecDecodedInvlpga(pVCpu, cbInstr);
7936	return VBOXSTRICTRC_VAL(rcStrict);
7937	}
7938
7939
7940	/**
7941	* \#VMEXIT handler for STGI (SVM_EXIT_VMRUN). Conditional \#VMEXIT.
7942	*/
7943	HMSVM_EXIT_DECL hmR0SvmExitVmrun(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7944	{
7945	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7946	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK
7947	\| IEM_CPUMCTX_EXTRN_SVM_VMRUN_MASK);
7948	VBOXSTRICTRC rcStrict;
7949	uint8_t const cbInstr = hmR0SvmGetInstrLengthHwAssist(pVCpu, pCtx, 3);
7950	rcStrict = IEMExecDecodedVmrun(pVCpu, cbInstr);
7951	Log4Func(("IEMExecDecodedVmrun returns %Rrc\n", VBOXSTRICTRC_VAL(rcStrict)));
7952	if (rcStrict == VINF_SUCCESS)
7953	{
7954	rcStrict = VINF_SVM_VMRUN;
7955	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_SVM_VMRUN_MASK);
7956	}
7957	return VBOXSTRICTRC_VAL(rcStrict);
7958	}
7959
7960
7961	/**
7962	* Nested-guest \#VMEXIT handler for debug exceptions (SVM_EXIT_XCPT_1).
7963	* Unconditional \#VMEXIT.
7964	*/
7965	HMSVM_EXIT_DECL hmR0SvmNestedExitXcptDB(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7966	{
7967	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7968	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
7969
7970	if (pVCpu->hm.s.Event.fPending)
7971	{
7972	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectPendingInterpret);
7973	return VINF_EM_RAW_INJECT_TRPM_EVENT;
7974	}
7975
7976	hmR0SvmSetPendingXcptDB(pVCpu);
7977	return VINF_SUCCESS;
7978	}
7979
7980
7981	/**
7982	* Nested-guest \#VMEXIT handler for breakpoint exceptions (SVM_EXIT_XCPT_3).
7983	* Conditional \#VMEXIT.
7984	*/
7985	HMSVM_EXIT_DECL hmR0SvmNestedExitXcptBP(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
7986	{
7987	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
7988	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
7989
7990	SVMEVENT Event;
7991	Event.u = 0;
7992	Event.n.u1Valid = 1;
7993	Event.n.u3Type = SVM_EVENT_EXCEPTION;
7994	Event.n.u8Vector = X86_XCPT_BP;
7995	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
7996	return VINF_SUCCESS;
7997	}
7998	#endif /* VBOX_WITH_NESTED_HWVIRT_SVM */
7999
8000	/** @} */
8001

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

source: vbox/trunk/src/VBox/VMM/VMMR0/HMSVMR0.cpp@ 72881

Download in other formats: