HMSVMR0.cpp@ 92583

Last change on this file since 92583 was 92583, checked in by vboxsync, 3 years ago
VMM: Nested VMX: bugref:10092 Renamed fPdpesMapped as it's rather misleading. More importantly CR3 is mapped and in case of PAE paging, the PAE PDPTEs have been mapped.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 319.2 KB

Line
1	/* $Id: HMSVMR0.cpp 92583 2021-11-24 09:13:14Z vboxsync $ */
2	/** @file
3	* HM SVM (AMD-V) - Host Context Ring-0.
4	*/
5
6	/*
7	* Copyright (C) 2013-2020 Oracle Corporation
8	*
9	* This file is part of VirtualBox Open Source Edition (OSE), as
10	* available from http://www.virtualbox.org. This file is free software;
11	* you can redistribute it and/or modify it under the terms of the GNU
12	* General Public License (GPL) as published by the Free Software
13	* Foundation, in version 2 as it comes in the "COPYING" file of the
14	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16	*/
17
18
19	/*********************************************************************************************************************************
20	* Header Files *
21	*********************************************************************************************************************************/
22	#define LOG_GROUP LOG_GROUP_HM
23	#define VMCPU_INCL_CPUM_GST_CTX
24	#include <iprt/asm-amd64-x86.h>
25	#include <iprt/thread.h>
26
27	#include <VBox/vmm/pdmapi.h>
28	#include <VBox/vmm/dbgf.h>
29	#include <VBox/vmm/iem.h>
30	#include <VBox/vmm/iom.h>
31	#include <VBox/vmm/tm.h>
32	#include <VBox/vmm/em.h>
33	#include <VBox/vmm/gim.h>
34	#include <VBox/vmm/apic.h>
35	#include "HMInternal.h"
36	#include <VBox/vmm/vmcc.h>
37	#include <VBox/err.h>
38	#include "HMSVMR0.h"
39	#include "dtrace/VBoxVMM.h"
40
41	#ifdef DEBUG_ramshankar
42	# define HMSVM_SYNC_FULL_GUEST_STATE
43	# define HMSVM_ALWAYS_TRAP_ALL_XCPTS
44	# define HMSVM_ALWAYS_TRAP_PF
45	# define HMSVM_ALWAYS_TRAP_TASK_SWITCH
46	#endif
47
48
49	/*********************************************************************************************************************************
50	* Defined Constants And Macros *
51	*********************************************************************************************************************************/
52	#ifdef VBOX_WITH_STATISTICS
53	# define HMSVM_EXITCODE_STAM_COUNTER_INC(u64ExitCode) do { \
54	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitAll); \
55	if ((u64ExitCode) == SVM_EXIT_NPF) \
56	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitReasonNpf); \
57	else \
58	STAM_COUNTER_INC(&pVCpu->hm.s.aStatExitReason[(u64ExitCode) & MASK_EXITREASON_STAT]); \
59	} while (0)
60
61	# ifdef VBOX_WITH_NESTED_HWVIRT_SVM
62	# define HMSVM_NESTED_EXITCODE_STAM_COUNTER_INC(u64ExitCode) do { \
63	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitAll); \
64	STAM_COUNTER_INC(&pVCpu->hm.s.StatNestedExitAll); \
65	if ((u64ExitCode) == SVM_EXIT_NPF) \
66	STAM_COUNTER_INC(&pVCpu->hm.s.StatNestedExitReasonNpf); \
67	else \
68	STAM_COUNTER_INC(&pVCpu->hm.s.aStatNestedExitReason[(u64ExitCode) & MASK_EXITREASON_STAT]); \
69	} while (0)
70	# endif
71	#else
72	# define HMSVM_EXITCODE_STAM_COUNTER_INC(u64ExitCode) do { } while (0)
73	# ifdef VBOX_WITH_NESTED_HWVIRT_SVM
74	# define HMSVM_NESTED_EXITCODE_STAM_COUNTER_INC(u64ExitCode) do { } while (0)
75	# endif
76	#endif /* !VBOX_WITH_STATISTICS */
77
78	/** If we decide to use a function table approach this can be useful to
79	* switch to a "static DECLCALLBACK(int)". */
80	#define HMSVM_EXIT_DECL static VBOXSTRICTRC
81
82	/**
83	* Subset of the guest-CPU state that is kept by SVM R0 code while executing the
84	* guest using hardware-assisted SVM.
85	*
86	* This excludes state like TSC AUX, GPRs (other than RSP, RAX) which are always
87	* are swapped and restored across the world-switch and also registers like
88	* EFER, PAT MSR etc. which cannot be modified by the guest without causing a
89	* \#VMEXIT.
90	*/
91	#define HMSVM_CPUMCTX_EXTRN_ALL ( CPUMCTX_EXTRN_RIP \
92	\| CPUMCTX_EXTRN_RFLAGS \
93	\| CPUMCTX_EXTRN_RAX \
94	\| CPUMCTX_EXTRN_RSP \
95	\| CPUMCTX_EXTRN_SREG_MASK \
96	\| CPUMCTX_EXTRN_CR0 \
97	\| CPUMCTX_EXTRN_CR2 \
98	\| CPUMCTX_EXTRN_CR3 \
99	\| CPUMCTX_EXTRN_TABLE_MASK \
100	\| CPUMCTX_EXTRN_DR6 \
101	\| CPUMCTX_EXTRN_DR7 \
102	\| CPUMCTX_EXTRN_KERNEL_GS_BASE \
103	\| CPUMCTX_EXTRN_SYSCALL_MSRS \
104	\| CPUMCTX_EXTRN_SYSENTER_MSRS \
105	\| CPUMCTX_EXTRN_HWVIRT \
106	\| CPUMCTX_EXTRN_INHIBIT_INT \
107	\| CPUMCTX_EXTRN_HM_SVM_MASK)
108
109	/**
110	* Subset of the guest-CPU state that is shared between the guest and host.
111	*/
112	#define HMSVM_CPUMCTX_SHARED_STATE CPUMCTX_EXTRN_DR_MASK
113
114	/** Macro for importing guest state from the VMCB back into CPUMCTX. */
115	#define HMSVM_CPUMCTX_IMPORT_STATE(a_pVCpu, a_fWhat) \
116	do { \
117	if ((a_pVCpu)->cpum.GstCtx.fExtrn & (a_fWhat)) \
118	hmR0SvmImportGuestState((a_pVCpu), (a_fWhat)); \
119	} while (0)
120
121	/** Assert that the required state bits are fetched. */
122	#define HMSVM_CPUMCTX_ASSERT(a_pVCpu, a_fExtrnMbz) AssertMsg(!((a_pVCpu)->cpum.GstCtx.fExtrn & (a_fExtrnMbz)), \
123	("fExtrn=%#RX64 fExtrnMbz=%#RX64\n", \
124	(a_pVCpu)->cpum.GstCtx.fExtrn, (a_fExtrnMbz)))
125
126	/** Assert that preemption is disabled or covered by thread-context hooks. */
127	#define HMSVM_ASSERT_PREEMPT_SAFE(a_pVCpu) Assert( VMMR0ThreadCtxHookIsEnabled((a_pVCpu)) \
128	\|\| !RTThreadPreemptIsEnabled(NIL_RTTHREAD));
129
130	/** Assert that we haven't migrated CPUs when thread-context hooks are not
131	* used. */
132	#define HMSVM_ASSERT_CPU_SAFE(a_pVCpu) AssertMsg( VMMR0ThreadCtxHookIsEnabled((a_pVCpu)) \
133	\|\| (a_pVCpu)->hmr0.s.idEnteredCpu == RTMpCpuId(), \
134	("Illegal migration! Entered on CPU %u Current %u\n", \
135	(a_pVCpu)->hmr0.s.idEnteredCpu, RTMpCpuId()));
136
137	/** Assert that we're not executing a nested-guest. */
138	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
139	# define HMSVM_ASSERT_NOT_IN_NESTED_GUEST(a_pCtx) Assert(!CPUMIsGuestInSvmNestedHwVirtMode((a_pCtx)))
140	#else
141	# define HMSVM_ASSERT_NOT_IN_NESTED_GUEST(a_pCtx) do { NOREF((a_pCtx)); } while (0)
142	#endif
143
144	/** Assert that we're executing a nested-guest. */
145	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
146	# define HMSVM_ASSERT_IN_NESTED_GUEST(a_pCtx) Assert(CPUMIsGuestInSvmNestedHwVirtMode((a_pCtx)))
147	#else
148	# define HMSVM_ASSERT_IN_NESTED_GUEST(a_pCtx) do { NOREF((a_pCtx)); } while (0)
149	#endif
150
151	/** Macro for checking and returning from the using function for
152	* \#VMEXIT intercepts that maybe caused during delivering of another
153	* event in the guest. */
154	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
155	# define HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(a_pVCpu, a_pSvmTransient) \
156	do \
157	{ \
158	int rc = hmR0SvmCheckExitDueToEventDelivery((a_pVCpu), (a_pSvmTransient)); \
159	if (RT_LIKELY(rc == VINF_SUCCESS)) { /* continue #VMEXIT handling */ } \
160	else if ( rc == VINF_HM_DOUBLE_FAULT) { return VINF_SUCCESS; } \
161	else if ( rc == VINF_EM_RESET \
162	&& CPUMIsGuestSvmCtrlInterceptSet((a_pVCpu), &(a_pVCpu)->cpum.GstCtx, SVM_CTRL_INTERCEPT_SHUTDOWN)) \
163	{ \
164	HMSVM_CPUMCTX_IMPORT_STATE((a_pVCpu), HMSVM_CPUMCTX_EXTRN_ALL); \
165	return IEMExecSvmVmexit((a_pVCpu), SVM_EXIT_SHUTDOWN, 0, 0); \
166	} \
167	else \
168	return rc; \
169	} while (0)
170	#else
171	# define HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(a_pVCpu, a_pSvmTransient) \
172	do \
173	{ \
174	int rc = hmR0SvmCheckExitDueToEventDelivery((a_pVCpu), (a_pSvmTransient)); \
175	if (RT_LIKELY(rc == VINF_SUCCESS)) { /* continue #VMEXIT handling */ } \
176	else if ( rc == VINF_HM_DOUBLE_FAULT) { return VINF_SUCCESS; } \
177	else \
178	return rc; \
179	} while (0)
180	#endif
181
182	/** Macro for upgrading a @a a_rc to VINF_EM_DBG_STEPPED after emulating an
183	* instruction that exited. */
184	#define HMSVM_CHECK_SINGLE_STEP(a_pVCpu, a_rc) \
185	do { \
186	if ((a_pVCpu)->hm.s.fSingleInstruction && (a_rc) == VINF_SUCCESS) \
187	(a_rc) = VINF_EM_DBG_STEPPED; \
188	} while (0)
189
190	/** Validate segment descriptor granularity bit. */
191	#ifdef VBOX_STRICT
192	# define HMSVM_ASSERT_SEG_GRANULARITY(a_pCtx, reg) \
193	AssertMsg( !(a_pCtx)->reg.Attr.n.u1Present \
194	\|\| ( (a_pCtx)->reg.Attr.n.u1Granularity \
195	? ((a_pCtx)->reg.u32Limit & 0xfff) == 0xfff \
196	: (a_pCtx)->reg.u32Limit <= UINT32_C(0xfffff)), \
197	("Invalid Segment Attributes Limit=%#RX32 Attr=%#RX32 Base=%#RX64\n", (a_pCtx)->reg.u32Limit, \
198	(a_pCtx)->reg.Attr.u, (a_pCtx)->reg.u64Base))
199	#else
200	# define HMSVM_ASSERT_SEG_GRANULARITY(a_pCtx, reg) do { } while (0)
201	#endif
202
203	/**
204	* Exception bitmap mask for all contributory exceptions.
205	*
206	* Page fault is deliberately excluded here as it's conditional as to whether
207	* it's contributory or benign. Page faults are handled separately.
208	*/
209	#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) \
210	\| RT_BIT(X86_XCPT_DE))
211
212	/**
213	* Mandatory/unconditional guest control intercepts.
214	*
215	* SMIs can and do happen in normal operation. We need not intercept them
216	* while executing the guest (or nested-guest).
217	*/
218	#define HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS ( SVM_CTRL_INTERCEPT_INTR \
219	\| SVM_CTRL_INTERCEPT_NMI \
220	\| SVM_CTRL_INTERCEPT_INIT \
221	\| SVM_CTRL_INTERCEPT_RDPMC \
222	\| SVM_CTRL_INTERCEPT_CPUID \
223	\| SVM_CTRL_INTERCEPT_RSM \
224	\| SVM_CTRL_INTERCEPT_HLT \
225	\| SVM_CTRL_INTERCEPT_IOIO_PROT \
226	\| SVM_CTRL_INTERCEPT_MSR_PROT \
227	\| SVM_CTRL_INTERCEPT_INVLPGA \
228	\| SVM_CTRL_INTERCEPT_SHUTDOWN \
229	\| SVM_CTRL_INTERCEPT_FERR_FREEZE \
230	\| SVM_CTRL_INTERCEPT_VMRUN \
231	\| SVM_CTRL_INTERCEPT_SKINIT \
232	\| SVM_CTRL_INTERCEPT_WBINVD \
233	\| SVM_CTRL_INTERCEPT_MONITOR \
234	\| SVM_CTRL_INTERCEPT_MWAIT \
235	\| SVM_CTRL_INTERCEPT_CR0_SEL_WRITE \
236	\| SVM_CTRL_INTERCEPT_XSETBV)
237
238	/** @name VMCB Clean Bits.
239	*
240	* These flags are used for VMCB-state caching. A set VMCB Clean bit indicates
241	* AMD-V doesn't need to reload the corresponding value(s) from the VMCB in
242	* memory.
243	*
244	* @{ */
245	/** All intercepts vectors, TSC offset, PAUSE filter counter. */
246	#define HMSVM_VMCB_CLEAN_INTERCEPTS RT_BIT(0)
247	/** I/O permission bitmap, MSR permission bitmap. */
248	#define HMSVM_VMCB_CLEAN_IOPM_MSRPM RT_BIT(1)
249	/** ASID. */
250	#define HMSVM_VMCB_CLEAN_ASID RT_BIT(2)
251	/** TRP: V_TPR, V_IRQ, V_INTR_PRIO, V_IGN_TPR, V_INTR_MASKING,
252	V_INTR_VECTOR. */
253	#define HMSVM_VMCB_CLEAN_INT_CTRL RT_BIT(3)
254	/** Nested Paging: Nested CR3 (nCR3), PAT. */
255	#define HMSVM_VMCB_CLEAN_NP RT_BIT(4)
256	/** Control registers (CR0, CR3, CR4, EFER). */
257	#define HMSVM_VMCB_CLEAN_CRX_EFER RT_BIT(5)
258	/** Debug registers (DR6, DR7). */
259	#define HMSVM_VMCB_CLEAN_DRX RT_BIT(6)
260	/** GDT, IDT limit and base. */
261	#define HMSVM_VMCB_CLEAN_DT RT_BIT(7)
262	/** Segment register: CS, SS, DS, ES limit and base. */
263	#define HMSVM_VMCB_CLEAN_SEG RT_BIT(8)
264	/** CR2.*/
265	#define HMSVM_VMCB_CLEAN_CR2 RT_BIT(9)
266	/** Last-branch record (DbgCtlMsr, br_from, br_to, lastint_from, lastint_to) */
267	#define HMSVM_VMCB_CLEAN_LBR RT_BIT(10)
268	/** AVIC (AVIC APIC_BAR; AVIC APIC_BACKING_PAGE, AVIC
269	PHYSICAL_TABLE and AVIC LOGICAL_TABLE Pointers). */
270	#define HMSVM_VMCB_CLEAN_AVIC RT_BIT(11)
271	/** Mask of all valid VMCB Clean bits. */
272	#define HMSVM_VMCB_CLEAN_ALL ( HMSVM_VMCB_CLEAN_INTERCEPTS \
273	\| HMSVM_VMCB_CLEAN_IOPM_MSRPM \
274	\| HMSVM_VMCB_CLEAN_ASID \
275	\| HMSVM_VMCB_CLEAN_INT_CTRL \
276	\| HMSVM_VMCB_CLEAN_NP \
277	\| HMSVM_VMCB_CLEAN_CRX_EFER \
278	\| HMSVM_VMCB_CLEAN_DRX \
279	\| HMSVM_VMCB_CLEAN_DT \
280	\| HMSVM_VMCB_CLEAN_SEG \
281	\| HMSVM_VMCB_CLEAN_CR2 \
282	\| HMSVM_VMCB_CLEAN_LBR \
283	\| HMSVM_VMCB_CLEAN_AVIC)
284	/** @} */
285
286	/** @name SVM transient.
287	*
288	* A state structure for holding miscellaneous information across AMD-V
289	* VMRUN/\#VMEXIT operation, restored after the transition.
290	*
291	* @{ */
292	typedef struct SVMTRANSIENT
293	{
294	/** The host's rflags/eflags. */
295	RTCCUINTREG fEFlags;
296	/** The \#VMEXIT exit code (the EXITCODE field in the VMCB). */
297	uint64_t u64ExitCode;
298
299	/** The guest's TPR value used for TPR shadowing. */
300	uint8_t u8GuestTpr;
301	/** Alignment. */
302	uint8_t abAlignment0[7];
303
304	/** Pointer to the currently executing VMCB. */
305	PSVMVMCB pVmcb;
306
307	/** Whether we are currently executing a nested-guest. */
308	bool fIsNestedGuest;
309	/** Whether the guest debug state was active at the time of \#VMEXIT. */
310	bool fWasGuestDebugStateActive;
311	/** Whether the hyper debug state was active at the time of \#VMEXIT. */
312	bool fWasHyperDebugStateActive;
313	/** Whether the TSC offset mode needs to be updated. */
314	bool fUpdateTscOffsetting;
315	/** Whether the TSC_AUX MSR needs restoring on \#VMEXIT. */
316	bool fRestoreTscAuxMsr;
317	/** Whether the \#VMEXIT was caused by a page-fault during delivery of a
318	* contributary exception or a page-fault. */
319	bool fVectoringDoublePF;
320	/** Whether the \#VMEXIT was caused by a page-fault during delivery of an
321	* external interrupt or NMI. */
322	bool fVectoringPF;
323	/** Padding. */
324	bool afPadding0;
325	} SVMTRANSIENT;
326	/** Pointer to SVM transient state. */
327	typedef SVMTRANSIENT *PSVMTRANSIENT;
328	/** Pointer to a const SVM transient state. */
329	typedef const SVMTRANSIENT *PCSVMTRANSIENT;
330
331	AssertCompileSizeAlignment(SVMTRANSIENT, sizeof(uint64_t));
332	AssertCompileMemberAlignment(SVMTRANSIENT, u64ExitCode, sizeof(uint64_t));
333	AssertCompileMemberAlignment(SVMTRANSIENT, pVmcb, sizeof(uint64_t));
334	/** @} */
335
336	/**
337	* MSRPM (MSR permission bitmap) read permissions (for guest RDMSR).
338	*/
339	typedef enum SVMMSREXITREAD
340	{
341	/** Reading this MSR causes a \#VMEXIT. */
342	SVMMSREXIT_INTERCEPT_READ = 0xb,
343	/** Reading this MSR does not cause a \#VMEXIT. */
344	SVMMSREXIT_PASSTHRU_READ
345	} SVMMSREXITREAD;
346
347	/**
348	* MSRPM (MSR permission bitmap) write permissions (for guest WRMSR).
349	*/
350	typedef enum SVMMSREXITWRITE
351	{
352	/** Writing to this MSR causes a \#VMEXIT. */
353	SVMMSREXIT_INTERCEPT_WRITE = 0xd,
354	/** Writing to this MSR does not cause a \#VMEXIT. */
355	SVMMSREXIT_PASSTHRU_WRITE
356	} SVMMSREXITWRITE;
357
358	/**
359	* SVM \#VMEXIT handler.
360	*
361	* @returns Strict VBox status code.
362	* @param pVCpu The cross context virtual CPU structure.
363	* @param pSvmTransient Pointer to the SVM-transient structure.
364	*/
365	typedef VBOXSTRICTRC FNSVMEXITHANDLER(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient);
366
367
368	/*********************************************************************************************************************************
369	* Internal Functions *
370	*********************************************************************************************************************************/
371	static void hmR0SvmPendingEventToTrpmTrap(PVMCPUCC pVCpu);
372	static void hmR0SvmLeave(PVMCPUCC pVCpu, bool fImportState);
373
374
375	/** @name \#VMEXIT handlers.
376	* @{
377	*/
378	static FNSVMEXITHANDLER hmR0SvmExitIntr;
379	static FNSVMEXITHANDLER hmR0SvmExitWbinvd;
380	static FNSVMEXITHANDLER hmR0SvmExitInvd;
381	static FNSVMEXITHANDLER hmR0SvmExitCpuid;
382	static FNSVMEXITHANDLER hmR0SvmExitRdtsc;
383	static FNSVMEXITHANDLER hmR0SvmExitRdtscp;
384	static FNSVMEXITHANDLER hmR0SvmExitRdpmc;
385	static FNSVMEXITHANDLER hmR0SvmExitInvlpg;
386	static FNSVMEXITHANDLER hmR0SvmExitHlt;
387	static FNSVMEXITHANDLER hmR0SvmExitMonitor;
388	static FNSVMEXITHANDLER hmR0SvmExitMwait;
389	static FNSVMEXITHANDLER hmR0SvmExitShutdown;
390	static FNSVMEXITHANDLER hmR0SvmExitUnexpected;
391	static FNSVMEXITHANDLER hmR0SvmExitReadCRx;
392	static FNSVMEXITHANDLER hmR0SvmExitWriteCRx;
393	static FNSVMEXITHANDLER hmR0SvmExitMsr;
394	static FNSVMEXITHANDLER hmR0SvmExitReadDRx;
395	static FNSVMEXITHANDLER hmR0SvmExitWriteDRx;
396	static FNSVMEXITHANDLER hmR0SvmExitXsetbv;
397	static FNSVMEXITHANDLER hmR0SvmExitIOInstr;
398	static FNSVMEXITHANDLER hmR0SvmExitNestedPF;
399	static FNSVMEXITHANDLER hmR0SvmExitVIntr;
400	static FNSVMEXITHANDLER hmR0SvmExitTaskSwitch;
401	static FNSVMEXITHANDLER hmR0SvmExitVmmCall;
402	static FNSVMEXITHANDLER hmR0SvmExitPause;
403	static FNSVMEXITHANDLER hmR0SvmExitFerrFreeze;
404	static FNSVMEXITHANDLER hmR0SvmExitIret;
405	static FNSVMEXITHANDLER hmR0SvmExitXcptPF;
406	static FNSVMEXITHANDLER hmR0SvmExitXcptUD;
407	static FNSVMEXITHANDLER hmR0SvmExitXcptMF;
408	static FNSVMEXITHANDLER hmR0SvmExitXcptDB;
409	static FNSVMEXITHANDLER hmR0SvmExitXcptAC;
410	static FNSVMEXITHANDLER hmR0SvmExitXcptBP;
411	static FNSVMEXITHANDLER hmR0SvmExitXcptGP;
412	#if defined(HMSVM_ALWAYS_TRAP_ALL_XCPTS) \|\| defined(VBOX_WITH_NESTED_HWVIRT_SVM)
413	static FNSVMEXITHANDLER hmR0SvmExitXcptGeneric;
414	#endif
415	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
416	static FNSVMEXITHANDLER hmR0SvmExitClgi;
417	static FNSVMEXITHANDLER hmR0SvmExitStgi;
418	static FNSVMEXITHANDLER hmR0SvmExitVmload;
419	static FNSVMEXITHANDLER hmR0SvmExitVmsave;
420	static FNSVMEXITHANDLER hmR0SvmExitInvlpga;
421	static FNSVMEXITHANDLER hmR0SvmExitVmrun;
422	static FNSVMEXITHANDLER hmR0SvmNestedExitXcptDB;
423	static FNSVMEXITHANDLER hmR0SvmNestedExitXcptBP;
424	#endif
425	/** @} */
426
427	static VBOXSTRICTRC hmR0SvmHandleExit(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient);
428	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
429	static VBOXSTRICTRC hmR0SvmHandleExitNested(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient);
430	#endif
431
432
433	/*********************************************************************************************************************************
434	* Global Variables *
435	*********************************************************************************************************************************/
436	/** Ring-0 memory object for the IO bitmap. */
437	static RTR0MEMOBJ g_hMemObjIOBitmap = NIL_RTR0MEMOBJ;
438	/** Physical address of the IO bitmap. */
439	static RTHCPHYS g_HCPhysIOBitmap;
440	/** Pointer to the IO bitmap. */
441	static R0PTRTYPE(void *) g_pvIOBitmap;
442
443	#ifdef VBOX_STRICT
444	# define HMSVM_LOG_RBP_RSP RT_BIT_32(0)
445	# define HMSVM_LOG_CR_REGS RT_BIT_32(1)
446	# define HMSVM_LOG_CS RT_BIT_32(2)
447	# define HMSVM_LOG_SS RT_BIT_32(3)
448	# define HMSVM_LOG_FS RT_BIT_32(4)
449	# define HMSVM_LOG_GS RT_BIT_32(5)
450	# define HMSVM_LOG_LBR RT_BIT_32(6)
451	# define HMSVM_LOG_ALL ( HMSVM_LOG_RBP_RSP \
452	\| HMSVM_LOG_CR_REGS \
453	\| HMSVM_LOG_CS \
454	\| HMSVM_LOG_SS \
455	\| HMSVM_LOG_FS \
456	\| HMSVM_LOG_GS \
457	\| HMSVM_LOG_LBR)
458
459	/**
460	* Dumps virtual CPU state and additional info. to the logger for diagnostics.
461	*
462	* @param pVCpu The cross context virtual CPU structure.
463	* @param pVmcb Pointer to the VM control block.
464	* @param pszPrefix Log prefix.
465	* @param fFlags Log flags, see HMSVM_LOG_XXX.
466	* @param uVerbose The verbosity level, currently unused.
467	*/
468	static void hmR0SvmLogState(PVMCPUCC pVCpu, PCSVMVMCB pVmcb, const char *pszPrefix, uint32_t fFlags, uint8_t uVerbose)
469	{
470	RT_NOREF2(pVCpu, uVerbose);
471	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
472
473	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_CS \| CPUMCTX_EXTRN_RIP \| CPUMCTX_EXTRN_RFLAGS);
474	Log4(("%s: cs:rip=%04x:%RX64 efl=%#RX64\n", pszPrefix, pCtx->cs.Sel, pCtx->rip, pCtx->rflags.u));
475
476	if (fFlags & HMSVM_LOG_RBP_RSP)
477	{
478	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_RSP \| CPUMCTX_EXTRN_RBP);
479	Log4(("%s: rsp=%#RX64 rbp=%#RX64\n", pszPrefix, pCtx->rsp, pCtx->rbp));
480	}
481
482	if (fFlags & HMSVM_LOG_CR_REGS)
483	{
484	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_CR0 \| CPUMCTX_EXTRN_CR3 \| CPUMCTX_EXTRN_CR4);
485	Log4(("%s: cr0=%#RX64 cr3=%#RX64 cr4=%#RX64\n", pszPrefix, pCtx->cr0, pCtx->cr3, pCtx->cr4));
486	}
487
488	if (fFlags & HMSVM_LOG_CS)
489	{
490	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_CS);
491	Log4(("%s: cs={%04x base=%016RX64 limit=%08x flags=%08x}\n", pszPrefix, pCtx->cs.Sel, pCtx->cs.u64Base,
492	pCtx->cs.u32Limit, pCtx->cs.Attr.u));
493	}
494	if (fFlags & HMSVM_LOG_SS)
495	{
496	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_SS);
497	Log4(("%s: ss={%04x base=%016RX64 limit=%08x flags=%08x}\n", pszPrefix, pCtx->ss.Sel, pCtx->ss.u64Base,
498	pCtx->ss.u32Limit, pCtx->ss.Attr.u));
499	}
500	if (fFlags & HMSVM_LOG_FS)
501	{
502	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_FS);
503	Log4(("%s: fs={%04x base=%016RX64 limit=%08x flags=%08x}\n", pszPrefix, pCtx->fs.Sel, pCtx->fs.u64Base,
504	pCtx->fs.u32Limit, pCtx->fs.Attr.u));
505	}
506	if (fFlags & HMSVM_LOG_GS)
507	{
508	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_GS);
509	Log4(("%s: gs={%04x base=%016RX64 limit=%08x flags=%08x}\n", pszPrefix, pCtx->gs.Sel, pCtx->gs.u64Base,
510	pCtx->gs.u32Limit, pCtx->gs.Attr.u));
511	}
512
513	PCSVMVMCBSTATESAVE pVmcbGuest = &pVmcb->guest;
514	if (fFlags & HMSVM_LOG_LBR)
515	{
516	Log4(("%s: br_from=%#RX64 br_to=%#RX64 lastxcpt_from=%#RX64 lastxcpt_to=%#RX64\n", pszPrefix, pVmcbGuest->u64BR_FROM,
517	pVmcbGuest->u64BR_TO, pVmcbGuest->u64LASTEXCPFROM, pVmcbGuest->u64LASTEXCPTO));
518	}
519	NOREF(pszPrefix); NOREF(pVmcbGuest); NOREF(pCtx);
520	}
521	#endif /* VBOX_STRICT */
522
523
524	/**
525	* Sets up and activates AMD-V on the current CPU.
526	*
527	* @returns VBox status code.
528	* @param pHostCpu The HM physical-CPU structure.
529	* @param pVM The cross context VM structure. Can be
530	* NULL after a resume!
531	* @param pvCpuPage Pointer to the global CPU page.
532	* @param HCPhysCpuPage Physical address of the global CPU page.
533	* @param fEnabledByHost Whether the host OS has already initialized AMD-V.
534	* @param pHwvirtMsrs Pointer to the hardware-virtualization MSRs (currently
535	* unused).
536	*/
537	VMMR0DECL(int) SVMR0EnableCpu(PHMPHYSCPU pHostCpu, PVMCC pVM, void *pvCpuPage, RTHCPHYS HCPhysCpuPage, bool fEnabledByHost,
538	PCSUPHWVIRTMSRS pHwvirtMsrs)
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	RT_NOREF2(fEnabledByHost, pHwvirtMsrs);
547
548	/* Paranoid: Disable interrupt as, in theory, interrupt handlers might mess with EFER. */
549	RTCCUINTREG const fEFlags = ASMIntDisableFlags();
550
551	/*
552	* We must turn on AMD-V and setup the host state physical address, as those MSRs are per CPU.
553	*/
554	uint64_t u64HostEfer = ASMRdMsr(MSR_K6_EFER);
555	if (u64HostEfer & MSR_K6_EFER_SVME)
556	{
557	/* If the VBOX_HWVIRTEX_IGNORE_SVM_IN_USE is active, then we blindly use AMD-V. */
558	if ( pVM
559	&& pVM->hm.s.svm.fIgnoreInUseError)
560	pHostCpu->fIgnoreAMDVInUseError = true;
561
562	if (!pHostCpu->fIgnoreAMDVInUseError)
563	{
564	ASMSetFlags(fEFlags);
565	return VERR_SVM_IN_USE;
566	}
567	}
568
569	/* Turn on AMD-V in the EFER MSR. */
570	ASMWrMsr(MSR_K6_EFER, u64HostEfer \| MSR_K6_EFER_SVME);
571
572	/* Write the physical page address where the CPU will store the host state while executing the VM. */
573	ASMWrMsr(MSR_K8_VM_HSAVE_PA, HCPhysCpuPage);
574
575	/* Restore interrupts. */
576	ASMSetFlags(fEFlags);
577
578	/*
579	* Theoretically, other hypervisors may have used ASIDs, ideally we should flush all
580	* non-zero ASIDs when enabling SVM. AMD doesn't have an SVM instruction to flush all
581	* ASIDs (flushing is done upon VMRUN). Therefore, flag that we need to flush the TLB
582	* entirely with before executing any guest code.
583	*/
584	pHostCpu->fFlushAsidBeforeUse = true;
585
586	/*
587	* Ensure each VCPU scheduled on this CPU gets a new ASID on resume. See @bugref{6255}.
588	*/
589	++pHostCpu->cTlbFlushes;
590
591	return VINF_SUCCESS;
592	}
593
594
595	/**
596	* Deactivates AMD-V on the current CPU.
597	*
598	* @returns VBox status code.
599	* @param pHostCpu The HM physical-CPU structure.
600	* @param pvCpuPage Pointer to the global CPU page.
601	* @param HCPhysCpuPage Physical address of the global CPU page.
602	*/
603	VMMR0DECL(int) SVMR0DisableCpu(PHMPHYSCPU pHostCpu, void *pvCpuPage, RTHCPHYS HCPhysCpuPage)
604	{
605	RT_NOREF1(pHostCpu);
606	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
607	AssertReturn( HCPhysCpuPage
608	&& HCPhysCpuPage != NIL_RTHCPHYS, VERR_INVALID_PARAMETER);
609	AssertReturn(pvCpuPage, VERR_INVALID_PARAMETER);
610
611	/* Paranoid: Disable interrupts as, in theory, interrupt handlers might mess with EFER. */
612	RTCCUINTREG const fEFlags = ASMIntDisableFlags();
613
614	/* Turn off AMD-V in the EFER MSR. */
615	uint64_t u64HostEfer = ASMRdMsr(MSR_K6_EFER);
616	ASMWrMsr(MSR_K6_EFER, u64HostEfer & ~MSR_K6_EFER_SVME);
617
618	/* Invalidate host state physical address. */
619	ASMWrMsr(MSR_K8_VM_HSAVE_PA, 0);
620
621	/* Restore interrupts. */
622	ASMSetFlags(fEFlags);
623
624	return VINF_SUCCESS;
625	}
626
627
628	/**
629	* Does global AMD-V initialization (called during module initialization).
630	*
631	* @returns VBox status code.
632	*/
633	VMMR0DECL(int) SVMR0GlobalInit(void)
634	{
635	/*
636	* Allocate 12 KB (3 pages) for the IO bitmap. Since this is non-optional and we always
637	* intercept all IO accesses, it's done once globally here instead of per-VM.
638	*/
639	Assert(g_hMemObjIOBitmap == NIL_RTR0MEMOBJ);
640	int rc = RTR0MemObjAllocCont(&g_hMemObjIOBitmap, SVM_IOPM_PAGES << X86_PAGE_4K_SHIFT, false /* fExecutable */);
641	if (RT_FAILURE(rc))
642	return rc;
643
644	g_pvIOBitmap = RTR0MemObjAddress(g_hMemObjIOBitmap);
645	g_HCPhysIOBitmap = RTR0MemObjGetPagePhysAddr(g_hMemObjIOBitmap, 0 /* iPage */);
646
647	/* Set all bits to intercept all IO accesses. */
648	ASMMemFill32(g_pvIOBitmap, SVM_IOPM_PAGES << X86_PAGE_4K_SHIFT, UINT32_C(0xffffffff));
649
650	return VINF_SUCCESS;
651	}
652
653
654	/**
655	* Does global AMD-V termination (called during module termination).
656	*/
657	VMMR0DECL(void) SVMR0GlobalTerm(void)
658	{
659	if (g_hMemObjIOBitmap != NIL_RTR0MEMOBJ)
660	{
661	RTR0MemObjFree(g_hMemObjIOBitmap, true /* fFreeMappings */);
662	g_pvIOBitmap = NULL;
663	g_HCPhysIOBitmap = 0;
664	g_hMemObjIOBitmap = NIL_RTR0MEMOBJ;
665	}
666	}
667
668
669	/**
670	* Frees any allocated per-VCPU structures for a VM.
671	*
672	* @param pVM The cross context VM structure.
673	*/
674	DECLINLINE(void) hmR0SvmFreeStructs(PVMCC pVM)
675	{
676	for (VMCPUID idCpu = 0; idCpu < pVM->cCpus; idCpu++)
677	{
678	PVMCPUCC pVCpu = VMCC_GET_CPU(pVM, idCpu);
679	AssertPtr(pVCpu);
680
681	if (pVCpu->hmr0.s.svm.hMemObjVmcbHost != NIL_RTR0MEMOBJ)
682	{
683	RTR0MemObjFree(pVCpu->hmr0.s.svm.hMemObjVmcbHost, false);
684	pVCpu->hmr0.s.svm.HCPhysVmcbHost = 0;
685	pVCpu->hmr0.s.svm.hMemObjVmcbHost = NIL_RTR0MEMOBJ;
686	}
687
688	if (pVCpu->hmr0.s.svm.hMemObjVmcb != NIL_RTR0MEMOBJ)
689	{
690	RTR0MemObjFree(pVCpu->hmr0.s.svm.hMemObjVmcb, false);
691	pVCpu->hmr0.s.svm.pVmcb = NULL;
692	pVCpu->hmr0.s.svm.HCPhysVmcb = 0;
693	pVCpu->hmr0.s.svm.hMemObjVmcb = NIL_RTR0MEMOBJ;
694	}
695
696	if (pVCpu->hmr0.s.svm.hMemObjMsrBitmap != NIL_RTR0MEMOBJ)
697	{
698	RTR0MemObjFree(pVCpu->hmr0.s.svm.hMemObjMsrBitmap, false);
699	pVCpu->hmr0.s.svm.pvMsrBitmap = NULL;
700	pVCpu->hmr0.s.svm.HCPhysMsrBitmap = 0;
701	pVCpu->hmr0.s.svm.hMemObjMsrBitmap = NIL_RTR0MEMOBJ;
702	}
703	}
704	}
705
706
707	/**
708	* Sets pfnVMRun to the best suited variant.
709	*
710	* This must be called whenever anything changes relative to the SVMR0VMRun
711	* variant selection:
712	* - pVCpu->hm.s.fLoadSaveGuestXcr0
713	* - CPUMCTX_WSF_IBPB_ENTRY in pVCpu->cpum.GstCtx.fWorldSwitcher
714	* - CPUMCTX_WSF_IBPB_EXIT in pVCpu->cpum.GstCtx.fWorldSwitcher
715	* - Perhaps: CPUMIsGuestFPUStateActive() (windows only)
716	* - Perhaps: CPUMCTX.fXStateMask (windows only)
717	*
718	* We currently ASSUME that neither CPUMCTX_WSF_IBPB_ENTRY nor
719	* CPUMCTX_WSF_IBPB_EXIT cannot be changed at runtime.
720	*/
721	static void hmR0SvmUpdateVmRunFunction(PVMCPUCC pVCpu)
722	{
723	static const struct CLANGWORKAROUND { PFNHMSVMVMRUN pfn; } s_aHmR0SvmVmRunFunctions[] =
724	{
725	{ hmR0SvmVmRun_SansXcr0_SansIbpbEntry_SansIbpbExit },
726	{ hmR0SvmVmRun_WithXcr0_SansIbpbEntry_SansIbpbExit },
727	{ hmR0SvmVmRun_SansXcr0_WithIbpbEntry_SansIbpbExit },
728	{ hmR0SvmVmRun_WithXcr0_WithIbpbEntry_SansIbpbExit },
729	{ hmR0SvmVmRun_SansXcr0_SansIbpbEntry_WithIbpbExit },
730	{ hmR0SvmVmRun_WithXcr0_SansIbpbEntry_WithIbpbExit },
731	{ hmR0SvmVmRun_SansXcr0_WithIbpbEntry_WithIbpbExit },
732	{ hmR0SvmVmRun_WithXcr0_WithIbpbEntry_WithIbpbExit },
733	};
734	uintptr_t const idx = (pVCpu->hmr0.s.fLoadSaveGuestXcr0 ? 1 : 0)
735	\| (pVCpu->hmr0.s.fWorldSwitcher & HM_WSF_IBPB_ENTRY ? 2 : 0)
736	\| (pVCpu->hmr0.s.fWorldSwitcher & HM_WSF_IBPB_EXIT ? 4 : 0);
737	PFNHMSVMVMRUN const pfnVMRun = s_aHmR0SvmVmRunFunctions[idx].pfn;
738	if (pVCpu->hmr0.s.svm.pfnVMRun != pfnVMRun)
739	pVCpu->hmr0.s.svm.pfnVMRun = pfnVMRun;
740	}
741
742
743	/**
744	* Selector FNHMSVMVMRUN implementation.
745	*/
746	static DECLCALLBACK(int) hmR0SvmVMRunSelector(PVMCC pVM, PVMCPUCC pVCpu, RTHCPHYS HCPhysVMCB)
747	{
748	hmR0SvmUpdateVmRunFunction(pVCpu);
749	return pVCpu->hmr0.s.svm.pfnVMRun(pVM, pVCpu, HCPhysVMCB);
750	}
751
752
753	/**
754	* Does per-VM AMD-V initialization.
755	*
756	* @returns VBox status code.
757	* @param pVM The cross context VM structure.
758	*/
759	VMMR0DECL(int) SVMR0InitVM(PVMCC pVM)
760	{
761	int rc = VERR_INTERNAL_ERROR_5;
762
763	/*
764	* Check for an AMD CPU erratum which requires us to flush the TLB before every world-switch.
765	*/
766	uint32_t u32Family;
767	uint32_t u32Model;
768	uint32_t u32Stepping;
769	if (HMIsSubjectToSvmErratum170(&u32Family, &u32Model, &u32Stepping))
770	{
771	Log4Func(("AMD cpu with erratum 170 family %#x model %#x stepping %#x\n", u32Family, u32Model, u32Stepping));
772	pVM->hmr0.s.svm.fAlwaysFlushTLB = true;
773	}
774
775	/*
776	* Initialize the R0 memory objects up-front so we can properly cleanup on allocation failures.
777	*/
778	for (VMCPUID idCpu = 0; idCpu < pVM->cCpus; idCpu++)
779	{
780	PVMCPUCC pVCpu = VMCC_GET_CPU(pVM, idCpu);
781	pVCpu->hmr0.s.svm.hMemObjVmcbHost = NIL_RTR0MEMOBJ;
782	pVCpu->hmr0.s.svm.hMemObjVmcb = NIL_RTR0MEMOBJ;
783	pVCpu->hmr0.s.svm.hMemObjMsrBitmap = NIL_RTR0MEMOBJ;
784	}
785
786	for (VMCPUID idCpu = 0; idCpu < pVM->cCpus; idCpu++)
787	{
788	PVMCPUCC pVCpu = VMCC_GET_CPU(pVM, idCpu);
789
790	/*
791	* Initialize the hardware-assisted SVM guest-execution handler.
792	* We now use a single handler for both 32-bit and 64-bit guests, see @bugref{6208#c73}.
793	*/
794	pVCpu->hmr0.s.svm.pfnVMRun = hmR0SvmVMRunSelector;
795
796	/*
797	* Allocate one page for the host-context VM control block (VMCB). This is used for additional host-state (such as
798	* FS, GS, Kernel GS Base, etc.) apart from the host-state save area specified in MSR_K8_VM_HSAVE_PA.
799	*/
800	/** @todo Does this need to be below 4G? */
801	rc = RTR0MemObjAllocCont(&pVCpu->hmr0.s.svm.hMemObjVmcbHost, SVM_VMCB_PAGES << PAGE_SHIFT, false /* fExecutable */);
802	if (RT_FAILURE(rc))
803	goto failure_cleanup;
804
805	void *pvVmcbHost = RTR0MemObjAddress(pVCpu->hmr0.s.svm.hMemObjVmcbHost);
806	pVCpu->hmr0.s.svm.HCPhysVmcbHost = RTR0MemObjGetPagePhysAddr(pVCpu->hmr0.s.svm.hMemObjVmcbHost, 0 /* iPage */);
807	Assert(pVCpu->hmr0.s.svm.HCPhysVmcbHost < _4G);
808	ASMMemZeroPage(pvVmcbHost);
809
810	/*
811	* Allocate one page for the guest-state VMCB.
812	*/
813	/** @todo Does this need to be below 4G? */
814	rc = RTR0MemObjAllocCont(&pVCpu->hmr0.s.svm.hMemObjVmcb, SVM_VMCB_PAGES << PAGE_SHIFT, false /* fExecutable */);
815	if (RT_FAILURE(rc))
816	goto failure_cleanup;
817
818	pVCpu->hmr0.s.svm.pVmcb = (PSVMVMCB)RTR0MemObjAddress(pVCpu->hmr0.s.svm.hMemObjVmcb);
819	pVCpu->hmr0.s.svm.HCPhysVmcb = RTR0MemObjGetPagePhysAddr(pVCpu->hmr0.s.svm.hMemObjVmcb, 0 /* iPage */);
820	Assert(pVCpu->hmr0.s.svm.HCPhysVmcb < _4G);
821	ASMMemZeroPage(pVCpu->hmr0.s.svm.pVmcb);
822
823	/*
824	* Allocate two pages (8 KB) for the MSR permission bitmap. There doesn't seem to be a way to convince
825	* SVM to not require one.
826	*/
827	/** @todo Does this need to be below 4G? */
828	rc = RTR0MemObjAllocCont(&pVCpu->hmr0.s.svm.hMemObjMsrBitmap, SVM_MSRPM_PAGES << X86_PAGE_4K_SHIFT,
829	false /* fExecutable */);
830	if (RT_FAILURE(rc))
831	goto failure_cleanup;
832
833	pVCpu->hmr0.s.svm.pvMsrBitmap = RTR0MemObjAddress(pVCpu->hmr0.s.svm.hMemObjMsrBitmap);
834	pVCpu->hmr0.s.svm.HCPhysMsrBitmap = RTR0MemObjGetPagePhysAddr(pVCpu->hmr0.s.svm.hMemObjMsrBitmap, 0 /* iPage */);
835	/* Set all bits to intercept all MSR accesses (changed later on). */
836	ASMMemFill32(pVCpu->hmr0.s.svm.pvMsrBitmap, SVM_MSRPM_PAGES << X86_PAGE_4K_SHIFT, UINT32_C(0xffffffff));
837	}
838
839	return VINF_SUCCESS;
840
841	failure_cleanup:
842	hmR0SvmFreeStructs(pVM);
843	return rc;
844	}
845
846
847	/**
848	* Does per-VM AMD-V termination.
849	*
850	* @returns VBox status code.
851	* @param pVM The cross context VM structure.
852	*/
853	VMMR0DECL(int) SVMR0TermVM(PVMCC pVM)
854	{
855	hmR0SvmFreeStructs(pVM);
856	return VINF_SUCCESS;
857	}
858
859
860	/**
861	* Returns whether the VMCB Clean Bits feature is supported.
862	*
863	* @returns @c true if supported, @c false otherwise.
864	* @param pVCpu The cross context virtual CPU structure.
865	* @param fIsNestedGuest Whether we are currently executing the nested-guest.
866	*/
867	DECL_FORCE_INLINE(bool) hmR0SvmSupportsVmcbCleanBits(PVMCPUCC pVCpu, bool fIsNestedGuest)
868	{
869	PCVMCC pVM = pVCpu->CTX_SUFF(pVM);
870	bool const fHostVmcbCleanBits = RT_BOOL(g_fHmSvmFeatures & X86_CPUID_SVM_FEATURE_EDX_VMCB_CLEAN);
871	if (!fIsNestedGuest)
872	return fHostVmcbCleanBits;
873	return fHostVmcbCleanBits && pVM->cpum.ro.GuestFeatures.fSvmVmcbClean;
874	}
875
876
877	/**
878	* Returns whether the decode assists feature is supported.
879	*
880	* @returns @c true if supported, @c false otherwise.
881	* @param pVCpu The cross context virtual CPU structure.
882	*/
883	DECLINLINE(bool) hmR0SvmSupportsDecodeAssists(PVMCPUCC pVCpu)
884	{
885	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
886	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
887	if (CPUMIsGuestInSvmNestedHwVirtMode(&pVCpu->cpum.GstCtx))
888	return (g_fHmSvmFeatures & X86_CPUID_SVM_FEATURE_EDX_DECODE_ASSISTS)
889	&& pVM->cpum.ro.GuestFeatures.fSvmDecodeAssists;
890	#endif
891	return RT_BOOL(g_fHmSvmFeatures & X86_CPUID_SVM_FEATURE_EDX_DECODE_ASSISTS);
892	}
893
894
895	/**
896	* Returns whether the NRIP_SAVE feature is supported.
897	*
898	* @returns @c true if supported, @c false otherwise.
899	* @param pVCpu The cross context virtual CPU structure.
900	*/
901	DECLINLINE(bool) hmR0SvmSupportsNextRipSave(PVMCPUCC pVCpu)
902	{
903	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
904	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
905	if (CPUMIsGuestInSvmNestedHwVirtMode(&pVCpu->cpum.GstCtx))
906	return (g_fHmSvmFeatures & X86_CPUID_SVM_FEATURE_EDX_NRIP_SAVE)
907	&& pVM->cpum.ro.GuestFeatures.fSvmNextRipSave;
908	#endif
909	return RT_BOOL(g_fHmSvmFeatures & X86_CPUID_SVM_FEATURE_EDX_NRIP_SAVE);
910	}
911
912
913	/**
914	* Sets the permission bits for the specified MSR in the MSRPM bitmap.
915	*
916	* @param pVCpu The cross context virtual CPU structure.
917	* @param pbMsrBitmap Pointer to the MSR bitmap.
918	* @param idMsr The MSR for which the permissions are being set.
919	* @param enmRead MSR read permissions.
920	* @param enmWrite MSR write permissions.
921	*
922	* @remarks This function does -not- clear the VMCB clean bits for MSRPM. The
923	* caller needs to take care of this.
924	*/
925	static void hmR0SvmSetMsrPermission(PVMCPUCC pVCpu, uint8_t *pbMsrBitmap, uint32_t idMsr, SVMMSREXITREAD enmRead,
926	SVMMSREXITWRITE enmWrite)
927	{
928	bool const fInNestedGuestMode = CPUMIsGuestInSvmNestedHwVirtMode(&pVCpu->cpum.GstCtx);
929	uint16_t offMsrpm;
930	uint8_t uMsrpmBit;
931	int rc = CPUMGetSvmMsrpmOffsetAndBit(idMsr, &offMsrpm, &uMsrpmBit);
932	AssertRC(rc);
933
934	Assert(uMsrpmBit == 0 \|\| uMsrpmBit == 2 \|\| uMsrpmBit == 4 \|\| uMsrpmBit == 6);
935	Assert(offMsrpm < SVM_MSRPM_PAGES << X86_PAGE_4K_SHIFT);
936
937	pbMsrBitmap += offMsrpm;
938	if (enmRead == SVMMSREXIT_INTERCEPT_READ)
939	*pbMsrBitmap \|= RT_BIT(uMsrpmBit);
940	else
941	{
942	if (!fInNestedGuestMode)
943	*pbMsrBitmap &= ~RT_BIT(uMsrpmBit);
944	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
945	else
946	{
947	/* Only clear the bit if the nested-guest is also not intercepting the MSR read.*/
948	if (!(pVCpu->cpum.GstCtx.hwvirt.svm.abMsrBitmap[offMsrpm] & RT_BIT(uMsrpmBit)))
949	*pbMsrBitmap &= ~RT_BIT(uMsrpmBit);
950	else
951	Assert(*pbMsrBitmap & RT_BIT(uMsrpmBit));
952	}
953	#endif
954	}
955
956	if (enmWrite == SVMMSREXIT_INTERCEPT_WRITE)
957	*pbMsrBitmap \|= RT_BIT(uMsrpmBit + 1);
958	else
959	{
960	if (!fInNestedGuestMode)
961	*pbMsrBitmap &= ~RT_BIT(uMsrpmBit + 1);
962	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
963	else
964	{
965	/* Only clear the bit if the nested-guest is also not intercepting the MSR write.*/
966	if (!(pVCpu->cpum.GstCtx.hwvirt.svm.abMsrBitmap[offMsrpm] & RT_BIT(uMsrpmBit + 1)))
967	*pbMsrBitmap &= ~RT_BIT(uMsrpmBit + 1);
968	else
969	Assert(*pbMsrBitmap & RT_BIT(uMsrpmBit + 1));
970	}
971	#endif
972	}
973	}
974
975
976	/**
977	* Sets up AMD-V for the specified VM.
978	* This function is only called once per-VM during initalization.
979	*
980	* @returns VBox status code.
981	* @param pVM The cross context VM structure.
982	*/
983	VMMR0DECL(int) SVMR0SetupVM(PVMCC pVM)
984	{
985	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
986	AssertReturn(pVM, VERR_INVALID_PARAMETER);
987
988	/*
989	* Validate and copy over some parameters.
990	*/
991	AssertReturn(pVM->hm.s.svm.fSupported, VERR_INCOMPATIBLE_CONFIG);
992	bool const fNestedPaging = pVM->hm.s.fNestedPagingCfg;
993	AssertReturn(!fNestedPaging \|\| (g_fHmSvmFeatures & X86_CPUID_SVM_FEATURE_EDX_NESTED_PAGING), VERR_INCOMPATIBLE_CONFIG);
994	pVM->hmr0.s.fNestedPaging = fNestedPaging;
995	pVM->hmr0.s.fAllow64BitGuests = pVM->hm.s.fAllow64BitGuestsCfg;
996
997	/*
998	* Determin some configuration parameters.
999	*/
1000	bool const fPauseFilter = RT_BOOL(g_fHmSvmFeatures & X86_CPUID_SVM_FEATURE_EDX_PAUSE_FILTER);
1001	bool const fPauseFilterThreshold = RT_BOOL(g_fHmSvmFeatures & X86_CPUID_SVM_FEATURE_EDX_PAUSE_FILTER_THRESHOLD);
1002	bool const fUsePauseFilter = fPauseFilter && pVM->hm.s.svm.cPauseFilter;
1003
1004	bool const fLbrVirt = RT_BOOL(g_fHmSvmFeatures & X86_CPUID_SVM_FEATURE_EDX_LBR_VIRT);
1005	bool const fUseLbrVirt = fLbrVirt && pVM->hm.s.svm.fLbrVirt; /** @todo IEM implementation etc. */
1006
1007	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1008	bool const fVirtVmsaveVmload = RT_BOOL(g_fHmSvmFeatures & X86_CPUID_SVM_FEATURE_EDX_VIRT_VMSAVE_VMLOAD);
1009	bool const fUseVirtVmsaveVmload = fVirtVmsaveVmload && pVM->hm.s.svm.fVirtVmsaveVmload && fNestedPaging;
1010
1011	bool const fVGif = RT_BOOL(g_fHmSvmFeatures & X86_CPUID_SVM_FEATURE_EDX_VGIF);
1012	bool const fUseVGif = fVGif && pVM->hm.s.svm.fVGif;
1013	#endif
1014
1015	PVMCPUCC pVCpu0 = VMCC_GET_CPU_0(pVM);
1016	PSVMVMCB pVmcb0 = pVCpu0->hmr0.s.svm.pVmcb;
1017	AssertMsgReturn(RT_VALID_PTR(pVmcb0), ("Invalid pVmcb (%p) for vcpu[0]\n", pVmcb0), VERR_SVM_INVALID_PVMCB);
1018	PSVMVMCBCTRL pVmcbCtrl0 = &pVmcb0->ctrl;
1019
1020	/* Always trap #AC for reasons of security. */
1021	pVmcbCtrl0->u32InterceptXcpt \|= RT_BIT_32(X86_XCPT_AC);
1022
1023	/* Always trap #DB for reasons of security. */
1024	pVmcbCtrl0->u32InterceptXcpt \|= RT_BIT_32(X86_XCPT_DB);
1025
1026	/* Trap exceptions unconditionally (debug purposes). */
1027	#ifdef HMSVM_ALWAYS_TRAP_PF
1028	pVmcbCtrl0->u32InterceptXcpt \|= RT_BIT_32(X86_XCPT_PF);
1029	#endif
1030	#ifdef HMSVM_ALWAYS_TRAP_ALL_XCPTS
1031	/* If you add any exceptions here, make sure to update hmR0SvmHandleExit(). */
1032	pVmcbCtrl0->u32InterceptXcpt \|= RT_BIT_32(X86_XCPT_BP)
1033	\| RT_BIT_32(X86_XCPT_DE)
1034	\| RT_BIT_32(X86_XCPT_NM)
1035	\| RT_BIT_32(X86_XCPT_UD)
1036	\| RT_BIT_32(X86_XCPT_NP)
1037	\| RT_BIT_32(X86_XCPT_SS)
1038	\| RT_BIT_32(X86_XCPT_GP)
1039	\| RT_BIT_32(X86_XCPT_PF)
1040	\| RT_BIT_32(X86_XCPT_MF)
1041	;
1042	#endif
1043
1044	/* Apply the exceptions intercepts needed by the GIM provider. */
1045	if (pVCpu0->hm.s.fGIMTrapXcptUD \|\| pVCpu0->hm.s.svm.fEmulateLongModeSysEnterExit)
1046	pVmcbCtrl0->u32InterceptXcpt \|= RT_BIT(X86_XCPT_UD);
1047
1048	/* The mesa 3d driver hack needs #GP. */
1049	if (pVCpu0->hm.s.fTrapXcptGpForLovelyMesaDrv)
1050	pVmcbCtrl0->u32InterceptXcpt \|= RT_BIT(X86_XCPT_GP);
1051
1052	/* Set up unconditional intercepts and conditions. */
1053	pVmcbCtrl0->u64InterceptCtrl = HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS
1054	\| SVM_CTRL_INTERCEPT_VMMCALL
1055	\| SVM_CTRL_INTERCEPT_VMSAVE
1056	\| SVM_CTRL_INTERCEPT_VMLOAD
1057	\| SVM_CTRL_INTERCEPT_CLGI
1058	\| SVM_CTRL_INTERCEPT_STGI;
1059
1060	#ifdef HMSVM_ALWAYS_TRAP_TASK_SWITCH
1061	pVmcbCtrl0->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_TASK_SWITCH;
1062	#endif
1063
1064	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1065	if (pVCpu0->CTX_SUFF(pVM)->cpum.ro.GuestFeatures.fSvm)
1066	{
1067	/* Virtualized VMSAVE/VMLOAD. */
1068	if (fUseVirtVmsaveVmload)
1069	{
1070	pVmcbCtrl0->LbrVirt.n.u1VirtVmsaveVmload = 1;
1071	pVmcbCtrl0->u64InterceptCtrl &= ~( SVM_CTRL_INTERCEPT_VMSAVE
1072	\| SVM_CTRL_INTERCEPT_VMLOAD);
1073	}
1074	else
1075	Assert(!pVmcbCtrl0->LbrVirt.n.u1VirtVmsaveVmload);
1076
1077	/* Virtual GIF. */
1078	if (fUseVGif)
1079	{
1080	pVmcbCtrl0->IntCtrl.n.u1VGifEnable = 1;
1081	pVmcbCtrl0->u64InterceptCtrl &= ~( SVM_CTRL_INTERCEPT_CLGI
1082	\| SVM_CTRL_INTERCEPT_STGI);
1083	}
1084	else
1085	Assert(!pVmcbCtrl0->IntCtrl.n.u1VGifEnable);
1086	}
1087	else
1088	#endif
1089	{
1090	Assert(!pVCpu0->CTX_SUFF(pVM)->cpum.ro.GuestFeatures.fSvm);
1091	Assert(!pVmcbCtrl0->LbrVirt.n.u1VirtVmsaveVmload);
1092	Assert(!pVmcbCtrl0->IntCtrl.n.u1VGifEnable);
1093	}
1094
1095	/* CR4 writes must always be intercepted for tracking PGM mode changes. */
1096	pVmcbCtrl0->u16InterceptWrCRx = RT_BIT(4);
1097
1098	/* Intercept all DRx reads and writes by default. Changed later on. */
1099	pVmcbCtrl0->u16InterceptRdDRx = 0xffff;
1100	pVmcbCtrl0->u16InterceptWrDRx = 0xffff;
1101
1102	/* Virtualize masking of INTR interrupts. (reads/writes from/to CR8 go to the V_TPR register) */
1103	pVmcbCtrl0->IntCtrl.n.u1VIntrMasking = 1;
1104
1105	/* Ignore the priority in the virtual TPR. This is necessary for delivering PIC style (ExtInt) interrupts
1106	and we currently deliver both PIC and APIC interrupts alike, see hmR0SvmEvaluatePendingEvent() */
1107	pVmcbCtrl0->IntCtrl.n.u1IgnoreTPR = 1;
1108
1109	/* Set the IO permission bitmap physical addresses. */
1110	pVmcbCtrl0->u64IOPMPhysAddr = g_HCPhysIOBitmap;
1111
1112	/* LBR virtualization. */
1113	pVmcbCtrl0->LbrVirt.n.u1LbrVirt = fUseLbrVirt;
1114
1115	/* The host ASID MBZ, for the guest start with 1. */
1116	pVmcbCtrl0->TLBCtrl.n.u32ASID = 1;
1117
1118	/* Setup Nested Paging. This doesn't change throughout the execution time of the VM. */
1119	pVmcbCtrl0->NestedPagingCtrl.n.u1NestedPaging = fNestedPaging;
1120
1121	/* Without Nested Paging, we need additionally intercepts. */
1122	if (!fNestedPaging)
1123	{
1124	/* CR3 reads/writes must be intercepted; our shadow values differ from the guest values. */
1125	pVmcbCtrl0->u16InterceptRdCRx \|= RT_BIT(3);
1126	pVmcbCtrl0->u16InterceptWrCRx \|= RT_BIT(3);
1127
1128	/* Intercept INVLPG and task switches (may change CR3, EFLAGS, LDT). */
1129	pVmcbCtrl0->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_INVLPG
1130	\| SVM_CTRL_INTERCEPT_TASK_SWITCH;
1131
1132	/* Page faults must be intercepted to implement shadow paging. */
1133	pVmcbCtrl0->u32InterceptXcpt \|= RT_BIT(X86_XCPT_PF);
1134	}
1135
1136	/* Setup Pause Filter for guest pause-loop (spinlock) exiting. */
1137	if (fUsePauseFilter)
1138	{
1139	Assert(pVM->hm.s.svm.cPauseFilter > 0);
1140	pVmcbCtrl0->u16PauseFilterCount = pVM->hm.s.svm.cPauseFilter;
1141	if (fPauseFilterThreshold)
1142	pVmcbCtrl0->u16PauseFilterThreshold = pVM->hm.s.svm.cPauseFilterThresholdTicks;
1143	pVmcbCtrl0->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_PAUSE;
1144	}
1145
1146	/*
1147	* Setup the MSR permission bitmap.
1148	* The following MSRs are saved/restored automatically during the world-switch.
1149	* Don't intercept guest read/write accesses to these MSRs.
1150	*/
1151	uint8_t pbMsrBitmap0 = (uint8_t )pVCpu0->hmr0.s.svm.pvMsrBitmap;
1152	hmR0SvmSetMsrPermission(pVCpu0, pbMsrBitmap0, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1153	hmR0SvmSetMsrPermission(pVCpu0, pbMsrBitmap0, MSR_K8_CSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1154	hmR0SvmSetMsrPermission(pVCpu0, pbMsrBitmap0, MSR_K6_STAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1155	hmR0SvmSetMsrPermission(pVCpu0, pbMsrBitmap0, MSR_K8_SF_MASK, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1156	hmR0SvmSetMsrPermission(pVCpu0, pbMsrBitmap0, MSR_K8_FS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1157	hmR0SvmSetMsrPermission(pVCpu0, pbMsrBitmap0, MSR_K8_GS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1158	hmR0SvmSetMsrPermission(pVCpu0, pbMsrBitmap0, MSR_K8_KERNEL_GS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1159	if (!pVCpu0->hm.s.svm.fEmulateLongModeSysEnterExit)
1160	{
1161	hmR0SvmSetMsrPermission(pVCpu0, pbMsrBitmap0, MSR_IA32_SYSENTER_CS, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1162	hmR0SvmSetMsrPermission(pVCpu0, pbMsrBitmap0, MSR_IA32_SYSENTER_ESP, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1163	hmR0SvmSetMsrPermission(pVCpu0, pbMsrBitmap0, MSR_IA32_SYSENTER_EIP, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1164	}
1165	else
1166	{
1167	hmR0SvmSetMsrPermission(pVCpu0, pbMsrBitmap0, MSR_IA32_SYSENTER_CS, SVMMSREXIT_INTERCEPT_READ, SVMMSREXIT_INTERCEPT_WRITE);
1168	hmR0SvmSetMsrPermission(pVCpu0, pbMsrBitmap0, MSR_IA32_SYSENTER_ESP, SVMMSREXIT_INTERCEPT_READ, SVMMSREXIT_INTERCEPT_WRITE);
1169	hmR0SvmSetMsrPermission(pVCpu0, pbMsrBitmap0, MSR_IA32_SYSENTER_EIP, SVMMSREXIT_INTERCEPT_READ, SVMMSREXIT_INTERCEPT_WRITE);
1170	}
1171	pVmcbCtrl0->u64MSRPMPhysAddr = pVCpu0->hmr0.s.svm.HCPhysMsrBitmap;
1172
1173	/* Initially all VMCB clean bits MBZ indicating that everything should be loaded from the VMCB in memory. */
1174	Assert(pVmcbCtrl0->u32VmcbCleanBits == 0);
1175
1176	for (VMCPUID idCpu = 1; idCpu < pVM->cCpus; idCpu++)
1177	{
1178	PVMCPUCC pVCpuCur = VMCC_GET_CPU(pVM, idCpu);
1179	PSVMVMCB pVmcbCur = pVCpuCur->hmr0.s.svm.pVmcb;
1180	AssertMsgReturn(RT_VALID_PTR(pVmcbCur), ("Invalid pVmcb (%p) for vcpu[%u]\n", pVmcbCur, idCpu), VERR_SVM_INVALID_PVMCB);
1181	PSVMVMCBCTRL pVmcbCtrlCur = &pVmcbCur->ctrl;
1182
1183	/* Copy the VMCB control area. */
1184	memcpy(pVmcbCtrlCur, pVmcbCtrl0, sizeof(*pVmcbCtrlCur));
1185
1186	/* Copy the MSR bitmap and setup the VCPU-specific host physical address. */
1187	uint8_t pbMsrBitmapCur = (uint8_t )pVCpuCur->hmr0.s.svm.pvMsrBitmap;
1188	memcpy(pbMsrBitmapCur, pbMsrBitmap0, SVM_MSRPM_PAGES << X86_PAGE_4K_SHIFT);
1189	pVmcbCtrlCur->u64MSRPMPhysAddr = pVCpuCur->hmr0.s.svm.HCPhysMsrBitmap;
1190
1191	/* Initially all VMCB clean bits MBZ indicating that everything should be loaded from the VMCB in memory. */
1192	Assert(pVmcbCtrlCur->u32VmcbCleanBits == 0);
1193
1194	/* Verify our assumption that GIM providers trap #UD uniformly across VCPUs initially. */
1195	Assert(pVCpuCur->hm.s.fGIMTrapXcptUD == pVCpu0->hm.s.fGIMTrapXcptUD);
1196	}
1197
1198	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1199	LogRel(("HM: fUsePauseFilter=%RTbool fUseLbrVirt=%RTbool fUseVGif=%RTbool fUseVirtVmsaveVmload=%RTbool\n", fUsePauseFilter,
1200	fUseLbrVirt, fUseVGif, fUseVirtVmsaveVmload));
1201	#else
1202	LogRel(("HM: fUsePauseFilter=%RTbool fUseLbrVirt=%RTbool\n", fUsePauseFilter, fUseLbrVirt));
1203	#endif
1204	return VINF_SUCCESS;
1205	}
1206
1207
1208	/**
1209	* Gets a pointer to the currently active guest (or nested-guest) VMCB.
1210	*
1211	* @returns Pointer to the current context VMCB.
1212	* @param pVCpu The cross context virtual CPU structure.
1213	*/
1214	DECLINLINE(PSVMVMCB) hmR0SvmGetCurrentVmcb(PVMCPUCC pVCpu)
1215	{
1216	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1217	if (CPUMIsGuestInSvmNestedHwVirtMode(&pVCpu->cpum.GstCtx))
1218	return &pVCpu->cpum.GstCtx.hwvirt.svm.Vmcb;
1219	#endif
1220	return pVCpu->hmr0.s.svm.pVmcb;
1221	}
1222
1223
1224	/**
1225	* Gets a pointer to the nested-guest VMCB cache.
1226	*
1227	* @returns Pointer to the nested-guest VMCB cache.
1228	* @param pVCpu The cross context virtual CPU structure.
1229	*/
1230	DECLINLINE(PSVMNESTEDVMCBCACHE) hmR0SvmGetNestedVmcbCache(PVMCPUCC pVCpu)
1231	{
1232	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1233	Assert(pVCpu->hm.s.svm.NstGstVmcbCache.fCacheValid);
1234	return &pVCpu->hm.s.svm.NstGstVmcbCache;
1235	#else
1236	RT_NOREF(pVCpu);
1237	return NULL;
1238	#endif
1239	}
1240
1241
1242	/**
1243	* Invalidates a guest page by guest virtual address.
1244	*
1245	* @returns VBox status code.
1246	* @param pVCpu The cross context virtual CPU structure.
1247	* @param GCVirt Guest virtual address of the page to invalidate.
1248	*/
1249	VMMR0DECL(int) SVMR0InvalidatePage(PVMCPUCC pVCpu, RTGCPTR GCVirt)
1250	{
1251	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.svm.fSupported);
1252
1253	bool const fFlushPending = VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_TLB_FLUSH) \|\| pVCpu->CTX_SUFF(pVM)->hmr0.s.svm.fAlwaysFlushTLB;
1254
1255	/* Skip it if a TLB flush is already pending. */
1256	if (!fFlushPending)
1257	{
1258	Log4Func(("%#RGv\n", GCVirt));
1259
1260	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
1261	AssertMsgReturn(pVmcb, ("Invalid pVmcb!\n"), VERR_SVM_INVALID_PVMCB);
1262
1263	SVMR0InvlpgA(GCVirt, pVmcb->ctrl.TLBCtrl.n.u32ASID);
1264	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlbInvlpgVirt);
1265	}
1266	return VINF_SUCCESS;
1267	}
1268
1269
1270	/**
1271	* Flushes the appropriate tagged-TLB entries.
1272	*
1273	* @param pHostCpu The HM physical-CPU structure.
1274	* @param pVCpu The cross context virtual CPU structure.
1275	* @param pVmcb Pointer to the VM control block.
1276	*/
1277	static void hmR0SvmFlushTaggedTlb(PHMPHYSCPU pHostCpu, PVMCPUCC pVCpu, PSVMVMCB pVmcb)
1278	{
1279	/*
1280	* Force a TLB flush for the first world switch if the current CPU differs from the one
1281	* we ran on last. This can happen both for start & resume due to long jumps back to
1282	* ring-3.
1283	*
1284	* We also force a TLB flush every time when executing a nested-guest VCPU as there is no
1285	* correlation between it and the physical CPU.
1286	*
1287	* If the TLB flush count changed, another VM (VCPU rather) has hit the ASID limit while
1288	* flushing the TLB, so we cannot reuse the ASIDs without flushing.
1289	*/
1290	bool fNewAsid = false;
1291	Assert(pHostCpu->idCpu != NIL_RTCPUID);
1292	if ( pVCpu->hmr0.s.idLastCpu != pHostCpu->idCpu
1293	\|\| pVCpu->hmr0.s.cTlbFlushes != pHostCpu->cTlbFlushes
1294	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1295	\|\| CPUMIsGuestInSvmNestedHwVirtMode(&pVCpu->cpum.GstCtx)
1296	#endif
1297	)
1298	{
1299	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlbWorldSwitch);
1300	pVCpu->hmr0.s.fForceTLBFlush = true;
1301	fNewAsid = true;
1302	}
1303
1304	/* Set TLB flush state as checked until we return from the world switch. */
1305	ASMAtomicUoWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, true);
1306
1307	/* Check for explicit TLB flushes. */
1308	if (VMCPU_FF_TEST_AND_CLEAR(pVCpu, VMCPU_FF_TLB_FLUSH))
1309	{
1310	pVCpu->hmr0.s.fForceTLBFlush = true;
1311	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlb);
1312	}
1313
1314	/*
1315	* If the AMD CPU erratum 170, We need to flush the entire TLB for each world switch. Sad.
1316	* This Host CPU requirement takes precedence.
1317	*/
1318	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
1319	if (pVM->hmr0.s.svm.fAlwaysFlushTLB)
1320	{
1321	pHostCpu->uCurrentAsid = 1;
1322	pVCpu->hmr0.s.uCurrentAsid = 1;
1323	pVCpu->hmr0.s.cTlbFlushes = pHostCpu->cTlbFlushes;
1324	pVCpu->hmr0.s.idLastCpu = pHostCpu->idCpu;
1325	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
1326
1327	/* Clear the VMCB Clean Bit for NP while flushing the TLB. See @bugref{7152}. */
1328	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
1329	}
1330	else
1331	{
1332	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_NOTHING;
1333	if (pVCpu->hmr0.s.fForceTLBFlush)
1334	{
1335	/* Clear the VMCB Clean Bit for NP while flushing the TLB. See @bugref{7152}. */
1336	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
1337
1338	if (fNewAsid)
1339	{
1340	++pHostCpu->uCurrentAsid;
1341
1342	bool fHitASIDLimit = false;
1343	if (pHostCpu->uCurrentAsid >= g_uHmMaxAsid)
1344	{
1345	pHostCpu->uCurrentAsid = 1; /* Wraparound at 1; host uses 0 */
1346	pHostCpu->cTlbFlushes++; /* All VCPUs that run on this host CPU must use a new ASID. */
1347	fHitASIDLimit = true;
1348	}
1349
1350	if ( fHitASIDLimit
1351	\|\| pHostCpu->fFlushAsidBeforeUse)
1352	{
1353	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
1354	pHostCpu->fFlushAsidBeforeUse = false;
1355	}
1356
1357	pVCpu->hmr0.s.uCurrentAsid = pHostCpu->uCurrentAsid;
1358	pVCpu->hmr0.s.idLastCpu = pHostCpu->idCpu;
1359	pVCpu->hmr0.s.cTlbFlushes = pHostCpu->cTlbFlushes;
1360	}
1361	else
1362	{
1363	if (g_fHmSvmFeatures & X86_CPUID_SVM_FEATURE_EDX_FLUSH_BY_ASID)
1364	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_SINGLE_CONTEXT;
1365	else
1366	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
1367	}
1368
1369	pVCpu->hmr0.s.fForceTLBFlush = false;
1370	}
1371	}
1372
1373	/* Update VMCB with the ASID. */
1374	if (pVmcb->ctrl.TLBCtrl.n.u32ASID != pVCpu->hmr0.s.uCurrentAsid)
1375	{
1376	pVmcb->ctrl.TLBCtrl.n.u32ASID = pVCpu->hmr0.s.uCurrentAsid;
1377	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_ASID;
1378	}
1379
1380	AssertMsg(pVCpu->hmr0.s.idLastCpu == pHostCpu->idCpu,
1381	("vcpu idLastCpu=%u hostcpu idCpu=%u\n", pVCpu->hmr0.s.idLastCpu, pHostCpu->idCpu));
1382	AssertMsg(pVCpu->hmr0.s.cTlbFlushes == pHostCpu->cTlbFlushes,
1383	("Flush count mismatch for cpu %u (%u vs %u)\n", pHostCpu->idCpu, pVCpu->hmr0.s.cTlbFlushes, pHostCpu->cTlbFlushes));
1384	AssertMsg(pHostCpu->uCurrentAsid >= 1 && pHostCpu->uCurrentAsid < g_uHmMaxAsid,
1385	("cpu%d uCurrentAsid = %x\n", pHostCpu->idCpu, pHostCpu->uCurrentAsid));
1386	AssertMsg(pVCpu->hmr0.s.uCurrentAsid >= 1 && pVCpu->hmr0.s.uCurrentAsid < g_uHmMaxAsid,
1387	("cpu%d VM uCurrentAsid = %x\n", pHostCpu->idCpu, pVCpu->hmr0.s.uCurrentAsid));
1388
1389	#ifdef VBOX_WITH_STATISTICS
1390	if (pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_NOTHING)
1391	STAM_COUNTER_INC(&pVCpu->hm.s.StatNoFlushTlbWorldSwitch);
1392	else if ( pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_SINGLE_CONTEXT
1393	\|\| pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_SINGLE_CONTEXT_RETAIN_GLOBALS)
1394	{
1395	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushAsid);
1396	}
1397	else
1398	{
1399	Assert(pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_ENTIRE);
1400	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushEntire);
1401	}
1402	#endif
1403	}
1404
1405
1406	/**
1407	* Sets an exception intercept in the specified VMCB.
1408	*
1409	* @param pVmcb Pointer to the VM control block.
1410	* @param uXcpt The exception (X86_XCPT_*).
1411	*/
1412	DECLINLINE(void) hmR0SvmSetXcptIntercept(PSVMVMCB pVmcb, uint8_t uXcpt)
1413	{
1414	if (!(pVmcb->ctrl.u32InterceptXcpt & RT_BIT(uXcpt)))
1415	{
1416	pVmcb->ctrl.u32InterceptXcpt \|= RT_BIT(uXcpt);
1417	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1418	}
1419	}
1420
1421
1422	/**
1423	* Clears an exception intercept in the specified VMCB.
1424	*
1425	* @param pVCpu The cross context virtual CPU structure.
1426	* @param pVmcb Pointer to the VM control block.
1427	* @param uXcpt The exception (X86_XCPT_*).
1428	*
1429	* @remarks This takes into account if we're executing a nested-guest and only
1430	* removes the exception intercept if both the guest -and- nested-guest
1431	* are not intercepting it.
1432	*/
1433	DECLINLINE(void) hmR0SvmClearXcptIntercept(PVMCPUCC pVCpu, PSVMVMCB pVmcb, uint8_t uXcpt)
1434	{
1435	Assert(uXcpt != X86_XCPT_DB);
1436	Assert(uXcpt != X86_XCPT_AC);
1437	Assert(uXcpt != X86_XCPT_GP);
1438	#ifndef HMSVM_ALWAYS_TRAP_ALL_XCPTS
1439	if (pVmcb->ctrl.u32InterceptXcpt & RT_BIT(uXcpt))
1440	{
1441	bool fRemove = true;
1442	# ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1443	/* Only remove the intercept if the nested-guest is also not intercepting it! */
1444	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1445	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
1446	{
1447	PCSVMNESTEDVMCBCACHE pVmcbNstGstCache = hmR0SvmGetNestedVmcbCache(pVCpu);
1448	fRemove = !(pVmcbNstGstCache->u32InterceptXcpt & RT_BIT(uXcpt));
1449	}
1450	# else
1451	RT_NOREF(pVCpu);
1452	# endif
1453	if (fRemove)
1454	{
1455	pVmcb->ctrl.u32InterceptXcpt &= ~RT_BIT(uXcpt);
1456	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1457	}
1458	}
1459	#else
1460	RT_NOREF3(pVCpu, pVmcb, uXcpt);
1461	#endif
1462	}
1463
1464
1465	/**
1466	* Sets a control intercept in the specified VMCB.
1467	*
1468	* @param pVmcb Pointer to the VM control block.
1469	* @param fCtrlIntercept The control intercept (SVM_CTRL_INTERCEPT_*).
1470	*/
1471	DECLINLINE(void) hmR0SvmSetCtrlIntercept(PSVMVMCB pVmcb, uint64_t fCtrlIntercept)
1472	{
1473	if (!(pVmcb->ctrl.u64InterceptCtrl & fCtrlIntercept))
1474	{
1475	pVmcb->ctrl.u64InterceptCtrl \|= fCtrlIntercept;
1476	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1477	}
1478	}
1479
1480
1481	/**
1482	* Clears a control intercept in the specified VMCB.
1483	*
1484	* @returns @c true if the intercept is still set, @c false otherwise.
1485	* @param pVCpu The cross context virtual CPU structure.
1486	* @param pVmcb Pointer to the VM control block.
1487	* @param fCtrlIntercept The control intercept (SVM_CTRL_INTERCEPT_*).
1488	*
1489	* @remarks This takes into account if we're executing a nested-guest and only
1490	* removes the control intercept if both the guest -and- nested-guest
1491	* are not intercepting it.
1492	*/
1493	static bool hmR0SvmClearCtrlIntercept(PVMCPUCC pVCpu, PSVMVMCB pVmcb, uint64_t fCtrlIntercept)
1494	{
1495	if (pVmcb->ctrl.u64InterceptCtrl & fCtrlIntercept)
1496	{
1497	bool fRemove = true;
1498	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1499	/* Only remove the control intercept if the nested-guest is also not intercepting it! */
1500	if (CPUMIsGuestInSvmNestedHwVirtMode(&pVCpu->cpum.GstCtx))
1501	{
1502	PCSVMNESTEDVMCBCACHE pVmcbNstGstCache = hmR0SvmGetNestedVmcbCache(pVCpu);
1503	fRemove = !(pVmcbNstGstCache->u64InterceptCtrl & fCtrlIntercept);
1504	}
1505	#else
1506	RT_NOREF(pVCpu);
1507	#endif
1508	if (fRemove)
1509	{
1510	pVmcb->ctrl.u64InterceptCtrl &= ~fCtrlIntercept;
1511	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1512	}
1513	}
1514
1515	return RT_BOOL(pVmcb->ctrl.u64InterceptCtrl & fCtrlIntercept);
1516	}
1517
1518
1519	/**
1520	* Exports the guest (or nested-guest) CR0 into the VMCB.
1521	*
1522	* @param pVCpu The cross context virtual CPU structure.
1523	* @param pVmcb Pointer to the VM control block.
1524	*
1525	* @remarks This assumes we always pre-load the guest FPU.
1526	* @remarks No-long-jump zone!!!
1527	*/
1528	static void hmR0SvmExportGuestCR0(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
1529	{
1530	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1531
1532	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1533	uint64_t const uGuestCr0 = pCtx->cr0;
1534	uint64_t uShadowCr0 = uGuestCr0;
1535
1536	/* Always enable caching. */
1537	uShadowCr0 &= ~(X86_CR0_CD \| X86_CR0_NW);
1538
1539	/* When Nested Paging is not available use shadow page tables and intercept #PFs (latter done in SVMR0SetupVM()). */
1540	if (!pVCpu->CTX_SUFF(pVM)->hmr0.s.fNestedPaging)
1541	{
1542	uShadowCr0 \|= X86_CR0_PG /* Use shadow page tables. */
1543	\| X86_CR0_WP; /* Guest CPL 0 writes to its read-only pages should cause a #PF #VMEXIT. */
1544	}
1545
1546	/*
1547	* Use the #MF style of legacy-FPU error reporting for now. Although AMD-V has MSRs that
1548	* lets us isolate the host from it, IEM/REM still needs work to emulate it properly,
1549	* see @bugref{7243#c103}.
1550	*/
1551	if (!(uGuestCr0 & X86_CR0_NE))
1552	{
1553	uShadowCr0 \|= X86_CR0_NE;
1554	hmR0SvmSetXcptIntercept(pVmcb, X86_XCPT_MF);
1555	}
1556	else
1557	hmR0SvmClearXcptIntercept(pVCpu, pVmcb, X86_XCPT_MF);
1558
1559	/*
1560	* If the shadow and guest CR0 are identical we can avoid intercepting CR0 reads.
1561	*
1562	* CR0 writes still needs interception as PGM requires tracking paging mode changes,
1563	* see @bugref{6944}.
1564	*
1565	* We also don't ever want to honor weird things like cache disable from the guest.
1566	* However, we can avoid intercepting changes to the TS & MP bits by clearing the CR0
1567	* write intercept below and keeping SVM_CTRL_INTERCEPT_CR0_SEL_WRITE instead.
1568	*/
1569	if (uShadowCr0 == uGuestCr0)
1570	{
1571	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
1572	{
1573	pVmcb->ctrl.u16InterceptRdCRx &= ~RT_BIT(0);
1574	pVmcb->ctrl.u16InterceptWrCRx &= ~RT_BIT(0);
1575	Assert(pVmcb->ctrl.u64InterceptCtrl & SVM_CTRL_INTERCEPT_CR0_SEL_WRITE);
1576	}
1577	else
1578	{
1579	/* If the nested-hypervisor intercepts CR0 reads/writes, we need to continue intercepting them. */
1580	PCSVMNESTEDVMCBCACHE pVmcbNstGstCache = hmR0SvmGetNestedVmcbCache(pVCpu);
1581	pVmcb->ctrl.u16InterceptRdCRx = (pVmcb->ctrl.u16InterceptRdCRx & ~RT_BIT(0))
1582	\| (pVmcbNstGstCache->u16InterceptRdCRx & RT_BIT(0));
1583	pVmcb->ctrl.u16InterceptWrCRx = (pVmcb->ctrl.u16InterceptWrCRx & ~RT_BIT(0))
1584	\| (pVmcbNstGstCache->u16InterceptWrCRx & RT_BIT(0));
1585	}
1586	}
1587	else
1588	{
1589	pVmcb->ctrl.u16InterceptRdCRx \|= RT_BIT(0);
1590	pVmcb->ctrl.u16InterceptWrCRx \|= RT_BIT(0);
1591	}
1592	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1593
1594	Assert(!RT_HI_U32(uShadowCr0));
1595	if (pVmcb->guest.u64CR0 != uShadowCr0)
1596	{
1597	pVmcb->guest.u64CR0 = uShadowCr0;
1598	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1599	}
1600	}
1601
1602
1603	/**
1604	* Exports the guest (or nested-guest) CR3 into the VMCB.
1605	*
1606	* @param pVCpu The cross context virtual CPU structure.
1607	* @param pVmcb Pointer to the VM control block.
1608	*
1609	* @remarks No-long-jump zone!!!
1610	*/
1611	static void hmR0SvmExportGuestCR3(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
1612	{
1613	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1614
1615	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
1616	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1617	if (pVM->hmr0.s.fNestedPaging)
1618	{
1619	pVmcb->ctrl.u64NestedPagingCR3 = PGMGetHyperCR3(pVCpu);
1620	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
1621	pVmcb->guest.u64CR3 = pCtx->cr3;
1622	Assert(pVmcb->ctrl.u64NestedPagingCR3);
1623	}
1624	else
1625	pVmcb->guest.u64CR3 = PGMGetHyperCR3(pVCpu);
1626
1627	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1628	}
1629
1630
1631	/**
1632	* Exports the guest (or nested-guest) CR4 into the VMCB.
1633	*
1634	* @param pVCpu The cross context virtual CPU structure.
1635	* @param pVmcb Pointer to the VM control block.
1636	*
1637	* @remarks No-long-jump zone!!!
1638	*/
1639	static int hmR0SvmExportGuestCR4(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
1640	{
1641	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1642
1643	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1644	uint64_t uShadowCr4 = pCtx->cr4;
1645	if (!pVCpu->CTX_SUFF(pVM)->hmr0.s.fNestedPaging)
1646	{
1647	switch (pVCpu->hm.s.enmShadowMode)
1648	{
1649	case PGMMODE_REAL:
1650	case PGMMODE_PROTECTED: /* Protected mode, no paging. */
1651	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1652
1653	case PGMMODE_32_BIT: /* 32-bit paging. */
1654	uShadowCr4 &= ~X86_CR4_PAE;
1655	break;
1656
1657	case PGMMODE_PAE: /* PAE paging. */
1658	case PGMMODE_PAE_NX: /* PAE paging with NX enabled. */
1659	/** Must use PAE paging as we could use physical memory > 4 GB */
1660	uShadowCr4 \|= X86_CR4_PAE;
1661	break;
1662
1663	case PGMMODE_AMD64: /* 64-bit AMD paging (long mode). */
1664	case PGMMODE_AMD64_NX: /* 64-bit AMD paging (long mode) with NX enabled. */
1665	#ifdef VBOX_WITH_64_BITS_GUESTS
1666	break;
1667	#else
1668	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1669	#endif
1670
1671	default: /* shut up gcc */
1672	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1673	}
1674	}
1675
1676	/* Whether to save/load/restore XCR0 during world switch depends on CR4.OSXSAVE and host+guest XCR0. */
1677	bool const fLoadSaveGuestXcr0 = (pCtx->cr4 & X86_CR4_OSXSAVE) && pCtx->aXcr[0] != ASMGetXcr0();
1678	if (fLoadSaveGuestXcr0 != pVCpu->hmr0.s.fLoadSaveGuestXcr0)
1679	{
1680	pVCpu->hmr0.s.fLoadSaveGuestXcr0 = fLoadSaveGuestXcr0;
1681	hmR0SvmUpdateVmRunFunction(pVCpu);
1682	}
1683
1684	/* Avoid intercepting CR4 reads if the guest and shadow CR4 values are identical. */
1685	if (uShadowCr4 == pCtx->cr4)
1686	{
1687	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
1688	pVmcb->ctrl.u16InterceptRdCRx &= ~RT_BIT(4);
1689	else
1690	{
1691	/* If the nested-hypervisor intercepts CR4 reads, we need to continue intercepting them. */
1692	PCSVMNESTEDVMCBCACHE pVmcbNstGstCache = hmR0SvmGetNestedVmcbCache(pVCpu);
1693	pVmcb->ctrl.u16InterceptRdCRx = (pVmcb->ctrl.u16InterceptRdCRx & ~RT_BIT(4))
1694	\| (pVmcbNstGstCache->u16InterceptRdCRx & RT_BIT(4));
1695	}
1696	}
1697	else
1698	pVmcb->ctrl.u16InterceptRdCRx \|= RT_BIT(4);
1699
1700	/* CR4 writes are always intercepted (both guest, nested-guest) for tracking PGM mode changes. */
1701	Assert(pVmcb->ctrl.u16InterceptWrCRx & RT_BIT(4));
1702
1703	/* Update VMCB with the shadow CR4 the appropriate VMCB clean bits. */
1704	Assert(!RT_HI_U32(uShadowCr4));
1705	pVmcb->guest.u64CR4 = uShadowCr4;
1706	pVmcb->ctrl.u32VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_CRX_EFER \| HMSVM_VMCB_CLEAN_INTERCEPTS);
1707
1708	return VINF_SUCCESS;
1709	}
1710
1711
1712	/**
1713	* Exports the guest (or nested-guest) control registers into the VMCB.
1714	*
1715	* @returns VBox status code.
1716	* @param pVCpu The cross context virtual CPU structure.
1717	* @param pVmcb Pointer to the VM control block.
1718	*
1719	* @remarks No-long-jump zone!!!
1720	*/
1721	static int hmR0SvmExportGuestControlRegs(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
1722	{
1723	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1724
1725	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_CR_MASK)
1726	{
1727	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_CR0)
1728	hmR0SvmExportGuestCR0(pVCpu, pVmcb);
1729
1730	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_CR2)
1731	{
1732	pVmcb->guest.u64CR2 = pVCpu->cpum.GstCtx.cr2;
1733	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CR2;
1734	}
1735
1736	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_CR3)
1737	hmR0SvmExportGuestCR3(pVCpu, pVmcb);
1738
1739	/* CR4 re-loading is ASSUMED to be done everytime we get in from ring-3! (XCR0) */
1740	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_CR4)
1741	{
1742	int rc = hmR0SvmExportGuestCR4(pVCpu, pVmcb);
1743	if (RT_FAILURE(rc))
1744	return rc;
1745	}
1746
1747	pVCpu->hm.s.fCtxChanged &= ~HM_CHANGED_GUEST_CR_MASK;
1748	}
1749	return VINF_SUCCESS;
1750	}
1751
1752
1753	/**
1754	* Exports the guest (or nested-guest) segment registers into the VMCB.
1755	*
1756	* @returns VBox status code.
1757	* @param pVCpu The cross context virtual CPU structure.
1758	* @param pVmcb Pointer to the VM control block.
1759	*
1760	* @remarks No-long-jump zone!!!
1761	*/
1762	static void hmR0SvmExportGuestSegmentRegs(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
1763	{
1764	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1765	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1766
1767	/* Guest segment registers. */
1768	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SREG_MASK)
1769	{
1770	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_CS)
1771	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, CS, cs);
1772
1773	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SS)
1774	{
1775	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, SS, ss);
1776	pVmcb->guest.u8CPL = pCtx->ss.Attr.n.u2Dpl;
1777	}
1778
1779	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_DS)
1780	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, DS, ds);
1781
1782	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_ES)
1783	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, ES, es);
1784
1785	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_FS)
1786	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, FS, fs);
1787
1788	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_GS)
1789	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, GS, gs);
1790
1791	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_SEG;
1792	}
1793
1794	/* Guest TR. */
1795	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_TR)
1796	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, TR, tr);
1797
1798	/* Guest LDTR. */
1799	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_LDTR)
1800	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, LDTR, ldtr);
1801
1802	/* Guest GDTR. */
1803	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_GDTR)
1804	{
1805	pVmcb->guest.GDTR.u32Limit = pCtx->gdtr.cbGdt;
1806	pVmcb->guest.GDTR.u64Base = pCtx->gdtr.pGdt;
1807	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DT;
1808	}
1809
1810	/* Guest IDTR. */
1811	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_IDTR)
1812	{
1813	pVmcb->guest.IDTR.u32Limit = pCtx->idtr.cbIdt;
1814	pVmcb->guest.IDTR.u64Base = pCtx->idtr.pIdt;
1815	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DT;
1816	}
1817
1818	pVCpu->hm.s.fCtxChanged &= ~( HM_CHANGED_GUEST_SREG_MASK
1819	\| HM_CHANGED_GUEST_TABLE_MASK);
1820	}
1821
1822
1823	/**
1824	* Exports the guest (or nested-guest) MSRs into the VMCB.
1825	*
1826	* @param pVCpu The cross context virtual CPU structure.
1827	* @param pVmcb Pointer to the VM control block.
1828	*
1829	* @remarks No-long-jump zone!!!
1830	*/
1831	static void hmR0SvmExportGuestMsrs(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
1832	{
1833	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1834	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1835
1836	/* Guest Sysenter MSRs. */
1837	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SYSENTER_MSR_MASK)
1838	{
1839	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SYSENTER_CS_MSR)
1840	pVmcb->guest.u64SysEnterCS = pCtx->SysEnter.cs;
1841
1842	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SYSENTER_EIP_MSR)
1843	pVmcb->guest.u64SysEnterEIP = pCtx->SysEnter.eip;
1844
1845	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SYSENTER_ESP_MSR)
1846	pVmcb->guest.u64SysEnterESP = pCtx->SysEnter.esp;
1847	}
1848
1849	/*
1850	* Guest EFER MSR.
1851	* AMD-V requires guest EFER.SVME to be set. Weird.
1852	* See AMD spec. 15.5.1 "Basic Operation" \| "Canonicalization and Consistency Checks".
1853	*/
1854	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_EFER_MSR)
1855	{
1856	pVmcb->guest.u64EFER = pCtx->msrEFER \| MSR_K6_EFER_SVME;
1857	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1858	}
1859
1860	/* If the guest isn't in 64-bit mode, clear MSR_K6_LME bit, otherwise SVM expects amd64 shadow paging. */
1861	if ( !CPUMIsGuestInLongModeEx(pCtx)
1862	&& (pCtx->msrEFER & MSR_K6_EFER_LME))
1863	{
1864	pVmcb->guest.u64EFER &= ~MSR_K6_EFER_LME;
1865	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1866	}
1867
1868	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SYSCALL_MSRS)
1869	{
1870	pVmcb->guest.u64STAR = pCtx->msrSTAR;
1871	pVmcb->guest.u64LSTAR = pCtx->msrLSTAR;
1872	pVmcb->guest.u64CSTAR = pCtx->msrCSTAR;
1873	pVmcb->guest.u64SFMASK = pCtx->msrSFMASK;
1874	}
1875
1876	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_KERNEL_GS_BASE)
1877	pVmcb->guest.u64KernelGSBase = pCtx->msrKERNELGSBASE;
1878
1879	pVCpu->hm.s.fCtxChanged &= ~( HM_CHANGED_GUEST_SYSENTER_MSR_MASK
1880	\| HM_CHANGED_GUEST_EFER_MSR
1881	\| HM_CHANGED_GUEST_SYSCALL_MSRS
1882	\| HM_CHANGED_GUEST_KERNEL_GS_BASE);
1883
1884	/*
1885	* Setup the PAT MSR (applicable for Nested Paging only).
1886	*
1887	* The default value should be MSR_IA32_CR_PAT_INIT_VAL, but we treat all guest memory
1888	* as WB, so choose type 6 for all PAT slots, see @bugref{9634}.
1889	*
1890	* While guests can modify and see the modified values through the shadow values,
1891	* we shall not honor any guest modifications of this MSR to ensure caching is always
1892	* enabled similar to how we clear CR0.CD and NW bits.
1893	*
1894	* For nested-guests this needs to always be set as well, see @bugref{7243#c109}.
1895	*/
1896	pVmcb->guest.u64PAT = UINT64_C(0x0006060606060606);
1897
1898	/* Enable the last branch record bit if LBR virtualization is enabled. */
1899	if (pVmcb->ctrl.LbrVirt.n.u1LbrVirt)
1900	pVmcb->guest.u64DBGCTL = MSR_IA32_DEBUGCTL_LBR;
1901	}
1902
1903
1904	/**
1905	* Exports the guest (or nested-guest) debug state into the VMCB and programs
1906	* the necessary intercepts accordingly.
1907	*
1908	* @param pVCpu The cross context virtual CPU structure.
1909	* @param pVmcb Pointer to the VM control block.
1910	*
1911	* @remarks No-long-jump zone!!!
1912	* @remarks Requires EFLAGS to be up-to-date in the VMCB!
1913	*/
1914	static void hmR0SvmExportSharedDebugState(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
1915	{
1916	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1917
1918	/** @todo Figure out stepping with nested-guest. */
1919	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
1920	{
1921	/*
1922	* We don't want to always intercept DRx read/writes for nested-guests as it causes
1923	* problems when the nested hypervisor isn't intercepting them, see @bugref{10080}.
1924	* Instead, they are strictly only requested when the nested hypervisor intercepts
1925	* them -- handled while merging VMCB controls.
1926	*
1927	* If neither the outer nor the nested-hypervisor is intercepting DRx read/writes,
1928	* then the nested-guest debug state should be actively loaded on the host so that
1929	* nested-guest reads/writes its own debug registers without causing VM-exits.
1930	*/
1931	if ( ( pVmcb->ctrl.u16InterceptRdDRx != 0xffff
1932	\|\| pVmcb->ctrl.u16InterceptWrDRx != 0xffff)
1933	&& !CPUMIsGuestDebugStateActive(pVCpu))
1934	{
1935	CPUMR0LoadGuestDebugState(pVCpu, true /* include DR6 */);
1936	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxArmed);
1937	Assert(!CPUMIsHyperDebugStateActive(pVCpu));
1938	Assert(CPUMIsGuestDebugStateActive(pVCpu));
1939	}
1940
1941	pVmcb->guest.u64DR6 = pCtx->dr[6];
1942	pVmcb->guest.u64DR7 = pCtx->dr[7];
1943	return;
1944	}
1945
1946	/*
1947	* Anyone single stepping on the host side? If so, we'll have to use the
1948	* trap flag in the guest EFLAGS since AMD-V doesn't have a trap flag on
1949	* the VMM level like the VT-x implementations does.
1950	*/
1951	bool fInterceptMovDRx = false;
1952	bool const fStepping = pVCpu->hm.s.fSingleInstruction \|\| DBGFIsStepping(pVCpu);
1953	if (fStepping)
1954	{
1955	pVCpu->hmr0.s.fClearTrapFlag = true;
1956	pVmcb->guest.u64RFlags \|= X86_EFL_TF;
1957	fInterceptMovDRx = true; /* Need clean DR6, no guest mess. */
1958	}
1959
1960	if ( fStepping
1961	\|\| (CPUMGetHyperDR7(pVCpu) & X86_DR7_ENABLED_MASK))
1962	{
1963	/*
1964	* Use the combined guest and host DRx values found in the hypervisor
1965	* register set because the debugger has breakpoints active or someone
1966	* is single stepping on the host side.
1967	*
1968	* Note! DBGF expects a clean DR6 state before executing guest code.
1969	*/
1970	if (!CPUMIsHyperDebugStateActive(pVCpu))
1971	{
1972	CPUMR0LoadHyperDebugState(pVCpu, false /* include DR6 */);
1973	Assert(!CPUMIsGuestDebugStateActive(pVCpu));
1974	Assert(CPUMIsHyperDebugStateActive(pVCpu));
1975	}
1976
1977	/* Update DR6 & DR7. (The other DRx values are handled by CPUM one way or the other.) */
1978	if ( pVmcb->guest.u64DR6 != X86_DR6_INIT_VAL
1979	\|\| pVmcb->guest.u64DR7 != CPUMGetHyperDR7(pVCpu))
1980	{
1981	pVmcb->guest.u64DR7 = CPUMGetHyperDR7(pVCpu);
1982	pVmcb->guest.u64DR6 = X86_DR6_INIT_VAL;
1983	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
1984	}
1985
1986	/** @todo If we cared, we could optimize to allow the guest to read registers
1987	* with the same values. */
1988	fInterceptMovDRx = true;
1989	pVCpu->hmr0.s.fUsingHyperDR7 = true;
1990	Log5(("hmR0SvmExportSharedDebugState: Loaded hyper DRx\n"));
1991	}
1992	else
1993	{
1994	/*
1995	* Update DR6, DR7 with the guest values if necessary.
1996	*/
1997	if ( pVmcb->guest.u64DR7 != pCtx->dr[7]
1998	\|\| pVmcb->guest.u64DR6 != pCtx->dr[6])
1999	{
2000	pVmcb->guest.u64DR7 = pCtx->dr[7];
2001	pVmcb->guest.u64DR6 = pCtx->dr[6];
2002	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
2003	}
2004	pVCpu->hmr0.s.fUsingHyperDR7 = false;
2005
2006	/*
2007	* If the guest has enabled debug registers, we need to load them prior to
2008	* executing guest code so they'll trigger at the right time.
2009	*/
2010	if (pCtx->dr[7] & (X86_DR7_ENABLED_MASK \| X86_DR7_GD)) /** @todo Why GD? */
2011	{
2012	if (!CPUMIsGuestDebugStateActive(pVCpu))
2013	{
2014	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
2015	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxArmed);
2016	Assert(!CPUMIsHyperDebugStateActive(pVCpu));
2017	Assert(CPUMIsGuestDebugStateActive(pVCpu));
2018	}
2019	Log5(("hmR0SvmExportSharedDebugState: Loaded guest DRx\n"));
2020	}
2021	/*
2022	* If no debugging enabled, we'll lazy load DR0-3. We don't need to
2023	* intercept #DB as DR6 is updated in the VMCB.
2024	*
2025	* Note! If we cared and dared, we could skip intercepting \#DB here.
2026	* However, \#DB shouldn't be performance critical, so we'll play safe
2027	* and keep the code similar to the VT-x code and always intercept it.
2028	*/
2029	else if (!CPUMIsGuestDebugStateActive(pVCpu))
2030	fInterceptMovDRx = true;
2031	}
2032
2033	Assert(pVmcb->ctrl.u32InterceptXcpt & RT_BIT_32(X86_XCPT_DB));
2034	if (fInterceptMovDRx)
2035	{
2036	if ( pVmcb->ctrl.u16InterceptRdDRx != 0xffff
2037	\|\| pVmcb->ctrl.u16InterceptWrDRx != 0xffff)
2038	{
2039	pVmcb->ctrl.u16InterceptRdDRx = 0xffff;
2040	pVmcb->ctrl.u16InterceptWrDRx = 0xffff;
2041	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2042	}
2043	}
2044	else
2045	{
2046	if ( pVmcb->ctrl.u16InterceptRdDRx
2047	\|\| pVmcb->ctrl.u16InterceptWrDRx)
2048	{
2049	pVmcb->ctrl.u16InterceptRdDRx = 0;
2050	pVmcb->ctrl.u16InterceptWrDRx = 0;
2051	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2052	}
2053	}
2054	Log4Func(("DR6=%#RX64 DR7=%#RX64\n", pCtx->dr[6], pCtx->dr[7]));
2055	}
2056
2057	/**
2058	* Exports the hardware virtualization state into the nested-guest
2059	* VMCB.
2060	*
2061	* @param pVCpu The cross context virtual CPU structure.
2062	* @param pVmcb Pointer to the VM control block.
2063	*
2064	* @remarks No-long-jump zone!!!
2065	*/
2066	static void hmR0SvmExportGuestHwvirtState(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
2067	{
2068	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2069
2070	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_HWVIRT)
2071	{
2072	if (pVmcb->ctrl.IntCtrl.n.u1VGifEnable)
2073	{
2074	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2075	PCVMCC pVM = pVCpu->CTX_SUFF(pVM);
2076
2077	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(pCtx); /* Nested VGIF is not supported yet. */
2078	Assert(g_fHmSvmFeatures & X86_CPUID_SVM_FEATURE_EDX_VGIF); /* Physical hardware supports VGIF. */
2079	Assert(HMIsSvmVGifActive(pVM)); /* Outer VM has enabled VGIF. */
2080	NOREF(pVM);
2081
2082	pVmcb->ctrl.IntCtrl.n.u1VGif = CPUMGetGuestGif(pCtx);
2083	}
2084
2085	/*
2086	* Ensure the nested-guest pause-filter counters don't exceed the outer guest values esp.
2087	* since SVM doesn't have a preemption timer.
2088	*
2089	* We do this here rather than in hmR0SvmSetupVmcbNested() as we may have been executing the
2090	* nested-guest in IEM incl. PAUSE instructions which would update the pause-filter counters
2091	* and may continue execution in SVM R0 without a nested-guest #VMEXIT in between.
2092	*/
2093	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
2094	PSVMVMCBCTRL pVmcbCtrl = &pVmcb->ctrl;
2095	uint16_t const uGuestPauseFilterCount = pVM->hm.s.svm.cPauseFilter;
2096	uint16_t const uGuestPauseFilterThreshold = pVM->hm.s.svm.cPauseFilterThresholdTicks;
2097	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, &pVCpu->cpum.GstCtx, SVM_CTRL_INTERCEPT_PAUSE))
2098	{
2099	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2100	pVmcbCtrl->u16PauseFilterCount = RT_MIN(pCtx->hwvirt.svm.cPauseFilter, uGuestPauseFilterCount);
2101	pVmcbCtrl->u16PauseFilterThreshold = RT_MIN(pCtx->hwvirt.svm.cPauseFilterThreshold, uGuestPauseFilterThreshold);
2102	}
2103	else
2104	{
2105	/** @todo r=ramshankar: We can turn these assignments into assertions. */
2106	pVmcbCtrl->u16PauseFilterCount = uGuestPauseFilterCount;
2107	pVmcbCtrl->u16PauseFilterThreshold = uGuestPauseFilterThreshold;
2108	}
2109	pVmcbCtrl->u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2110
2111	pVCpu->hm.s.fCtxChanged &= ~HM_CHANGED_GUEST_HWVIRT;
2112	}
2113	}
2114
2115
2116	/**
2117	* Exports the guest APIC TPR state into the VMCB.
2118	*
2119	* @returns VBox status code.
2120	* @param pVCpu The cross context virtual CPU structure.
2121	* @param pVmcb Pointer to the VM control block.
2122	*/
2123	static int hmR0SvmExportGuestApicTpr(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
2124	{
2125	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(&pVCpu->cpum.GstCtx);
2126
2127	if (ASMAtomicUoReadU64(&pVCpu->hm.s.fCtxChanged) & HM_CHANGED_GUEST_APIC_TPR)
2128	{
2129	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
2130	if ( PDMHasApic(pVM)
2131	&& APICIsEnabled(pVCpu))
2132	{
2133	bool fPendingIntr;
2134	uint8_t u8Tpr;
2135	int rc = APICGetTpr(pVCpu, &u8Tpr, &fPendingIntr, NULL /* pu8PendingIrq */);
2136	AssertRCReturn(rc, rc);
2137
2138	/* Assume that we need to trap all TPR accesses and thus need not check on
2139	every #VMEXIT if we should update the TPR. */
2140	Assert(pVmcb->ctrl.IntCtrl.n.u1VIntrMasking);
2141	pVCpu->hmr0.s.svm.fSyncVTpr = false;
2142
2143	if (!pVM->hm.s.fTprPatchingActive)
2144	{
2145	/* Bits 3-0 of the VTPR field correspond to bits 7-4 of the TPR (which is the Task-Priority Class). */
2146	pVmcb->ctrl.IntCtrl.n.u8VTPR = (u8Tpr >> 4);
2147
2148	/* If there are interrupts pending, intercept CR8 writes to evaluate ASAP if we
2149	can deliver the interrupt to the guest. */
2150	if (fPendingIntr)
2151	pVmcb->ctrl.u16InterceptWrCRx \|= RT_BIT(8);
2152	else
2153	{
2154	pVmcb->ctrl.u16InterceptWrCRx &= ~RT_BIT(8);
2155	pVCpu->hmr0.s.svm.fSyncVTpr = true;
2156	}
2157
2158	pVmcb->ctrl.u32VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS \| HMSVM_VMCB_CLEAN_INT_CTRL);
2159	}
2160	else
2161	{
2162	/* 32-bit guests uses LSTAR MSR for patching guest code which touches the TPR. */
2163	pVmcb->guest.u64LSTAR = u8Tpr;
2164	uint8_t pbMsrBitmap = (uint8_t )pVCpu->hmr0.s.svm.pvMsrBitmap;
2165
2166	/* If there are interrupts pending, intercept LSTAR writes, otherwise don't intercept reads or writes. */
2167	if (fPendingIntr)
2168	hmR0SvmSetMsrPermission(pVCpu, pbMsrBitmap, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_INTERCEPT_WRITE);
2169	else
2170	{
2171	hmR0SvmSetMsrPermission(pVCpu, pbMsrBitmap, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
2172	pVCpu->hmr0.s.svm.fSyncVTpr = true;
2173	}
2174	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_IOPM_MSRPM;
2175	}
2176	}
2177	ASMAtomicUoAndU64(&pVCpu->hm.s.fCtxChanged, ~HM_CHANGED_GUEST_APIC_TPR);
2178	}
2179	return VINF_SUCCESS;
2180	}
2181
2182
2183	/**
2184	* Sets up the exception interrupts required for guest execution in the VMCB.
2185	*
2186	* @param pVCpu The cross context virtual CPU structure.
2187	* @param pVmcb Pointer to the VM control block.
2188	*
2189	* @remarks No-long-jump zone!!!
2190	*/
2191	static void hmR0SvmExportGuestXcptIntercepts(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
2192	{
2193	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(&pVCpu->cpum.GstCtx);
2194
2195	/* If we modify intercepts from here, please check & adjust hmR0SvmMergeVmcbCtrlsNested() if required. */
2196	if (ASMAtomicUoReadU64(&pVCpu->hm.s.fCtxChanged) & HM_CHANGED_SVM_XCPT_INTERCEPTS)
2197	{
2198	/* Trap #UD for GIM provider (e.g. for hypercalls). */
2199	if (pVCpu->hm.s.fGIMTrapXcptUD \|\| pVCpu->hm.s.svm.fEmulateLongModeSysEnterExit)
2200	hmR0SvmSetXcptIntercept(pVmcb, X86_XCPT_UD);
2201	else
2202	hmR0SvmClearXcptIntercept(pVCpu, pVmcb, X86_XCPT_UD);
2203
2204	/* Trap #BP for INT3 debug breakpoints set by the VM debugger. */
2205	if (pVCpu->CTX_SUFF(pVM)->dbgf.ro.cEnabledInt3Breakpoints)
2206	hmR0SvmSetXcptIntercept(pVmcb, X86_XCPT_BP);
2207	else
2208	hmR0SvmClearXcptIntercept(pVCpu, pVmcb, X86_XCPT_BP);
2209
2210	/* The remaining intercepts are handled elsewhere, e.g. in hmR0SvmExportGuestCR0(). */
2211	ASMAtomicUoAndU64(&pVCpu->hm.s.fCtxChanged, ~HM_CHANGED_SVM_XCPT_INTERCEPTS);
2212	}
2213	}
2214
2215
2216	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
2217	/**
2218	* Merges guest and nested-guest intercepts for executing the nested-guest using
2219	* hardware-assisted SVM.
2220	*
2221	* This merges the guest and nested-guest intercepts in a way that if the outer
2222	* guest intercept is set we need to intercept it in the nested-guest as
2223	* well.
2224	*
2225	* @param pVCpu The cross context virtual CPU structure.
2226	* @param pVmcbNstGst Pointer to the nested-guest VM control block.
2227	*/
2228	static void hmR0SvmMergeVmcbCtrlsNested(PVMCPUCC pVCpu)
2229	{
2230	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
2231	PCSVMVMCB pVmcb = pVCpu->hmr0.s.svm.pVmcb;
2232	PSVMVMCB pVmcbNstGst = &pVCpu->cpum.GstCtx.hwvirt.svm.Vmcb;
2233	PSVMVMCBCTRL pVmcbNstGstCtrl = &pVmcbNstGst->ctrl;
2234
2235	/* Merge the guest's CR intercepts into the nested-guest VMCB. */
2236	pVmcbNstGstCtrl->u16InterceptRdCRx \|= pVmcb->ctrl.u16InterceptRdCRx;
2237	pVmcbNstGstCtrl->u16InterceptWrCRx \|= pVmcb->ctrl.u16InterceptWrCRx;
2238
2239	/* Always intercept CR4 writes for tracking PGM mode changes. */
2240	pVmcbNstGstCtrl->u16InterceptWrCRx \|= RT_BIT(4);
2241
2242	/* Without nested paging, intercept CR3 reads and writes as we load shadow page tables. */
2243	if (!pVM->hmr0.s.fNestedPaging)
2244	{
2245	pVmcbNstGstCtrl->u16InterceptRdCRx \|= RT_BIT(3);
2246	pVmcbNstGstCtrl->u16InterceptWrCRx \|= RT_BIT(3);
2247	}
2248
2249	/* Merge the guest's DR intercepts into the nested-guest VMCB. */
2250	pVmcbNstGstCtrl->u16InterceptRdDRx \|= pVmcb->ctrl.u16InterceptRdDRx;
2251	pVmcbNstGstCtrl->u16InterceptWrDRx \|= pVmcb->ctrl.u16InterceptWrDRx;
2252
2253	/*
2254	* Merge the guest's exception intercepts into the nested-guest VMCB.
2255	*
2256	* - #UD: Exclude these as the outer guest's GIM hypercalls are not applicable
2257	* while executing the nested-guest.
2258	*
2259	* - #BP: Exclude breakpoints set by the VM debugger for the outer guest. This can
2260	* be tweaked later depending on how we wish to implement breakpoints.
2261	*
2262	* - #GP: Exclude these as it's the inner VMMs problem to get vmsvga 3d drivers
2263	* loaded into their guests, not ours.
2264	*
2265	* Warning!! This ASSUMES we only intercept \#UD for hypercall purposes and \#BP
2266	* for VM debugger breakpoints, see hmR0SvmExportGuestXcptIntercepts().
2267	*/
2268	#ifndef HMSVM_ALWAYS_TRAP_ALL_XCPTS
2269	pVmcbNstGstCtrl->u32InterceptXcpt \|= pVmcb->ctrl.u32InterceptXcpt
2270	& ~( RT_BIT(X86_XCPT_UD)
2271	\| RT_BIT(X86_XCPT_BP)
2272	\| (pVCpu->hm.s.fTrapXcptGpForLovelyMesaDrv ? RT_BIT(X86_XCPT_GP) : 0));
2273	#else
2274	pVmcbNstGstCtrl->u32InterceptXcpt \|= pVmcb->ctrl.u32InterceptXcpt;
2275	#endif
2276
2277	/*
2278	* Adjust intercepts while executing the nested-guest that differ from the
2279	* outer guest intercepts.
2280	*
2281	* - VINTR: Exclude the outer guest intercept as we don't need to cause VINTR #VMEXITs
2282	* that belong to the nested-guest to the outer guest.
2283	*
2284	* - VMMCALL: Exclude the outer guest intercept as when it's also not intercepted by
2285	* the nested-guest, the physical CPU raises a \#UD exception as expected.
2286	*/
2287	pVmcbNstGstCtrl->u64InterceptCtrl \|= (pVmcb->ctrl.u64InterceptCtrl & ~( SVM_CTRL_INTERCEPT_VINTR
2288	\| SVM_CTRL_INTERCEPT_VMMCALL))
2289	\| HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS;
2290
2291	Assert( (pVmcbNstGstCtrl->u64InterceptCtrl & HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS)
2292	== HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS);
2293
2294	/* Finally, update the VMCB clean bits. */
2295	pVmcbNstGstCtrl->u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2296	}
2297	#endif
2298
2299
2300	/**
2301	* Enters the AMD-V session.
2302	*
2303	* @returns VBox status code.
2304	* @param pVCpu The cross context virtual CPU structure.
2305	*/
2306	VMMR0DECL(int) SVMR0Enter(PVMCPUCC pVCpu)
2307	{
2308	AssertPtr(pVCpu);
2309	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.svm.fSupported);
2310	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2311
2312	LogFlowFunc(("pVCpu=%p\n", pVCpu));
2313	Assert((pVCpu->hm.s.fCtxChanged & (HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE))
2314	== (HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE));
2315
2316	pVCpu->hmr0.s.fLeaveDone = false;
2317	return VINF_SUCCESS;
2318	}
2319
2320
2321	/**
2322	* Thread-context callback for AMD-V.
2323	*
2324	* This is used together with RTThreadCtxHookCreate() on platforms which
2325	* supports it, and directly from VMMR0EmtPrepareForBlocking() and
2326	* VMMR0EmtResumeAfterBlocking() on platforms which don't.
2327	*
2328	* @param enmEvent The thread-context event.
2329	* @param pVCpu The cross context virtual CPU structure.
2330	* @param fGlobalInit Whether global VT-x/AMD-V init. is used.
2331	* @thread EMT(pVCpu)
2332	*/
2333	VMMR0DECL(void) SVMR0ThreadCtxCallback(RTTHREADCTXEVENT enmEvent, PVMCPUCC pVCpu, bool fGlobalInit)
2334	{
2335	NOREF(fGlobalInit);
2336
2337	switch (enmEvent)
2338	{
2339	case RTTHREADCTXEVENT_OUT:
2340	{
2341	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2342	VMCPU_ASSERT_EMT(pVCpu);
2343
2344	/* No longjmps (log-flush, locks) in this fragile context. */
2345	VMMRZCallRing3Disable(pVCpu);
2346
2347	if (!pVCpu->hmr0.s.fLeaveDone)
2348	{
2349	hmR0SvmLeave(pVCpu, false /* fImportState */);
2350	pVCpu->hmr0.s.fLeaveDone = true;
2351	}
2352
2353	/* Leave HM context, takes care of local init (term). */
2354	int rc = HMR0LeaveCpu(pVCpu);
2355	AssertRC(rc); NOREF(rc);
2356
2357	/* Restore longjmp state. */
2358	VMMRZCallRing3Enable(pVCpu);
2359	STAM_REL_COUNTER_INC(&pVCpu->hm.s.StatSwitchPreempt);
2360	break;
2361	}
2362
2363	case RTTHREADCTXEVENT_IN:
2364	{
2365	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2366	VMCPU_ASSERT_EMT(pVCpu);
2367
2368	/* No longjmps (log-flush, locks) in this fragile context. */
2369	VMMRZCallRing3Disable(pVCpu);
2370
2371	/*
2372	* Initialize the bare minimum state required for HM. This takes care of
2373	* initializing AMD-V if necessary (onlined CPUs, local init etc.)
2374	*/
2375	int rc = hmR0EnterCpu(pVCpu);
2376	AssertRC(rc); NOREF(rc);
2377	Assert( (pVCpu->hm.s.fCtxChanged & (HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE))
2378	== (HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE));
2379
2380	pVCpu->hmr0.s.fLeaveDone = false;
2381
2382	/* Restore longjmp state. */
2383	VMMRZCallRing3Enable(pVCpu);
2384	break;
2385	}
2386
2387	default:
2388	break;
2389	}
2390	}
2391
2392
2393	/**
2394	* Saves the host state.
2395	*
2396	* @returns VBox status code.
2397	* @param pVCpu The cross context virtual CPU structure.
2398	*
2399	* @remarks No-long-jump zone!!!
2400	*/
2401	VMMR0DECL(int) SVMR0ExportHostState(PVMCPUCC pVCpu)
2402	{
2403	NOREF(pVCpu);
2404
2405	/* Nothing to do here. AMD-V does this for us automatically during the world-switch. */
2406	ASMAtomicUoAndU64(&pVCpu->hm.s.fCtxChanged, ~HM_CHANGED_HOST_CONTEXT);
2407	return VINF_SUCCESS;
2408	}
2409
2410
2411	/**
2412	* Exports the guest or nested-guest state from the virtual-CPU context into the
2413	* VMCB.
2414	*
2415	* Also sets up the appropriate VMRUN function to execute guest or nested-guest
2416	* code based on the virtual-CPU mode.
2417	*
2418	* @returns VBox status code.
2419	* @param pVCpu The cross context virtual CPU structure.
2420	* @param pSvmTransient Pointer to the SVM-transient structure.
2421	*
2422	* @remarks No-long-jump zone!!!
2423	*/
2424	static int hmR0SvmExportGuestState(PVMCPUCC pVCpu, PCSVMTRANSIENT pSvmTransient)
2425	{
2426	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatExportGuestState, x);
2427
2428	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
2429	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2430	Assert(pVmcb);
2431
2432	pVmcb->guest.u64RIP = pCtx->rip;
2433	pVmcb->guest.u64RSP = pCtx->rsp;
2434	pVmcb->guest.u64RFlags = pCtx->eflags.u32;
2435	pVmcb->guest.u64RAX = pCtx->rax;
2436
2437	bool const fIsNestedGuest = pSvmTransient->fIsNestedGuest;
2438	RTCCUINTREG const fEFlags = ASMIntDisableFlags();
2439
2440	int rc = hmR0SvmExportGuestControlRegs(pVCpu, pVmcb);
2441	AssertRCReturnStmt(rc, ASMSetFlags(fEFlags), rc);
2442	hmR0SvmExportGuestSegmentRegs(pVCpu, pVmcb);
2443	hmR0SvmExportGuestMsrs(pVCpu, pVmcb);
2444	hmR0SvmExportGuestHwvirtState(pVCpu, pVmcb);
2445
2446	ASMSetFlags(fEFlags);
2447
2448	if (!fIsNestedGuest)
2449	{
2450	/* hmR0SvmExportGuestApicTpr() must be called -after- hmR0SvmExportGuestMsrs() as we
2451	otherwise we would overwrite the LSTAR MSR that we use for TPR patching. */
2452	hmR0SvmExportGuestApicTpr(pVCpu, pVmcb);
2453	hmR0SvmExportGuestXcptIntercepts(pVCpu, pVmcb);
2454	}
2455
2456	/* Clear any bits that may be set but exported unconditionally or unused/reserved bits. */
2457	uint64_t fUnusedMask = HM_CHANGED_GUEST_RIP
2458	\| HM_CHANGED_GUEST_RFLAGS
2459	\| HM_CHANGED_GUEST_GPRS_MASK
2460	\| HM_CHANGED_GUEST_X87
2461	\| HM_CHANGED_GUEST_SSE_AVX
2462	\| HM_CHANGED_GUEST_OTHER_XSAVE
2463	\| HM_CHANGED_GUEST_XCRx
2464	\| HM_CHANGED_GUEST_TSC_AUX
2465	\| HM_CHANGED_GUEST_OTHER_MSRS;
2466	if (fIsNestedGuest)
2467	fUnusedMask \|= HM_CHANGED_SVM_XCPT_INTERCEPTS
2468	\| HM_CHANGED_GUEST_APIC_TPR;
2469
2470	ASMAtomicUoAndU64(&pVCpu->hm.s.fCtxChanged, ~( fUnusedMask
2471	\| (HM_CHANGED_KEEPER_STATE_MASK & ~HM_CHANGED_SVM_MASK)));
2472
2473	#ifdef VBOX_STRICT
2474	/*
2475	* All of the guest-CPU state and SVM keeper bits should be exported here by now,
2476	* except for the host-context and/or shared host-guest context bits.
2477	*/
2478	uint64_t const fCtxChanged = ASMAtomicUoReadU64(&pVCpu->hm.s.fCtxChanged);
2479	AssertMsg(!(fCtxChanged & (HM_CHANGED_ALL_GUEST & ~HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE)),
2480	("fCtxChanged=%#RX64\n", fCtxChanged));
2481
2482	/*
2483	* If we need to log state that isn't always imported, we'll need to import them here.
2484	* See hmR0SvmPostRunGuest() for which part of the state is imported uncondtionally.
2485	*/
2486	hmR0SvmLogState(pVCpu, pVmcb, "hmR0SvmExportGuestState", 0 /* fFlags /, 0 / uVerbose */);
2487	#endif
2488
2489	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExportGuestState, x);
2490	return VINF_SUCCESS;
2491	}
2492
2493	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
2494
2495	/**
2496	* Merges the guest and nested-guest MSR permission bitmap.
2497	*
2498	* If the guest is intercepting an MSR we need to intercept it regardless of
2499	* whether the nested-guest is intercepting it or not.
2500	*
2501	* @param pHostCpu The HM physical-CPU structure.
2502	* @param pVCpu The cross context virtual CPU structure.
2503	*
2504	* @remarks No-long-jmp zone!!!
2505	*/
2506	DECLINLINE(void) hmR0SvmMergeMsrpmNested(PHMPHYSCPU pHostCpu, PVMCPUCC pVCpu)
2507	{
2508	uint64_t const pu64GstMsrpm = (uint64_t const )pVCpu->hmr0.s.svm.pvMsrBitmap;
2509	uint64_t const pu64NstGstMsrpm = (uint64_t const )&pVCpu->cpum.GstCtx.hwvirt.svm.abMsrBitmap[0];
2510	uint64_t pu64DstMsrpm = (uint64_t )pHostCpu->n.svm.pvNstGstMsrpm;
2511
2512	/* MSRPM bytes from offset 0x1800 are reserved, so we stop merging there. */
2513	uint32_t const offRsvdQwords = 0x1800 >> 3;
2514	for (uint32_t i = 0; i < offRsvdQwords; i++)
2515	pu64DstMsrpm[i] = pu64NstGstMsrpm[i] \| pu64GstMsrpm[i];
2516	}
2517
2518
2519	/**
2520	* Caches the nested-guest VMCB fields before we modify them for execution using
2521	* hardware-assisted SVM.
2522	*
2523	* @returns true if the VMCB was previously already cached, false otherwise.
2524	* @param pVCpu The cross context virtual CPU structure.
2525	*
2526	* @sa HMNotifySvmNstGstVmexit.
2527	*/
2528	static bool hmR0SvmCacheVmcbNested(PVMCPUCC pVCpu)
2529	{
2530	/*
2531	* Cache the nested-guest programmed VMCB fields if we have not cached it yet.
2532	* Otherwise we risk re-caching the values we may have modified, see @bugref{7243#c44}.
2533	*
2534	* Nested-paging CR3 is not saved back into the VMCB on #VMEXIT, hence no need to
2535	* cache and restore it, see AMD spec. 15.25.4 "Nested Paging and VMRUN/#VMEXIT".
2536	*/
2537	PSVMNESTEDVMCBCACHE pVmcbNstGstCache = &pVCpu->hm.s.svm.NstGstVmcbCache;
2538	bool const fWasCached = pVmcbNstGstCache->fCacheValid;
2539	if (!fWasCached)
2540	{
2541	PCSVMVMCB pVmcbNstGst = &pVCpu->cpum.GstCtx.hwvirt.svm.Vmcb;
2542	PCSVMVMCBCTRL pVmcbNstGstCtrl = &pVmcbNstGst->ctrl;
2543	pVmcbNstGstCache->u16InterceptRdCRx = pVmcbNstGstCtrl->u16InterceptRdCRx;
2544	pVmcbNstGstCache->u16InterceptWrCRx = pVmcbNstGstCtrl->u16InterceptWrCRx;
2545	pVmcbNstGstCache->u16InterceptRdDRx = pVmcbNstGstCtrl->u16InterceptRdDRx;
2546	pVmcbNstGstCache->u16InterceptWrDRx = pVmcbNstGstCtrl->u16InterceptWrDRx;
2547	pVmcbNstGstCache->u16PauseFilterThreshold = pVmcbNstGstCtrl->u16PauseFilterThreshold;
2548	pVmcbNstGstCache->u16PauseFilterCount = pVmcbNstGstCtrl->u16PauseFilterCount;
2549	pVmcbNstGstCache->u32InterceptXcpt = pVmcbNstGstCtrl->u32InterceptXcpt;
2550	pVmcbNstGstCache->u64InterceptCtrl = pVmcbNstGstCtrl->u64InterceptCtrl;
2551	pVmcbNstGstCache->u64TSCOffset = pVmcbNstGstCtrl->u64TSCOffset;
2552	pVmcbNstGstCache->fVIntrMasking = pVmcbNstGstCtrl->IntCtrl.n.u1VIntrMasking;
2553	pVmcbNstGstCache->fNestedPaging = pVmcbNstGstCtrl->NestedPagingCtrl.n.u1NestedPaging;
2554	pVmcbNstGstCache->fLbrVirt = pVmcbNstGstCtrl->LbrVirt.n.u1LbrVirt;
2555	pVmcbNstGstCache->fCacheValid = true;
2556	Log4Func(("Cached VMCB fields\n"));
2557	}
2558
2559	return fWasCached;
2560	}
2561
2562
2563	/**
2564	* Sets up the nested-guest VMCB for execution using hardware-assisted SVM.
2565	*
2566	* This is done the first time we enter nested-guest execution using SVM R0
2567	* until the nested-guest \#VMEXIT (not to be confused with physical CPU
2568	* \#VMEXITs which may or may not cause a corresponding nested-guest \#VMEXIT).
2569	*
2570	* @param pVCpu The cross context virtual CPU structure.
2571	*/
2572	static void hmR0SvmSetupVmcbNested(PVMCPUCC pVCpu)
2573	{
2574	PSVMVMCB pVmcbNstGst = &pVCpu->cpum.GstCtx.hwvirt.svm.Vmcb;
2575	PSVMVMCBCTRL pVmcbNstGstCtrl = &pVmcbNstGst->ctrl;
2576
2577	HMSVM_ASSERT_IN_NESTED_GUEST(&pVCpu->cpum.GstCtx);
2578
2579	/*
2580	* First cache the nested-guest VMCB fields we may potentially modify.
2581	*/
2582	bool const fVmcbCached = hmR0SvmCacheVmcbNested(pVCpu);
2583	if (!fVmcbCached)
2584	{
2585	/*
2586	* The IOPM of the nested-guest can be ignored because the the guest always
2587	* intercepts all IO port accesses. Thus, we'll swap to the guest IOPM rather
2588	* than the nested-guest IOPM and swap the field back on the #VMEXIT.
2589	*/
2590	pVmcbNstGstCtrl->u64IOPMPhysAddr = g_HCPhysIOBitmap;
2591
2592	/*
2593	* Use the same nested-paging as the outer guest. We can't dynamically switch off
2594	* nested-paging suddenly while executing a VM (see assertion at the end of
2595	* Trap0eHandler() in PGMAllBth.h).
2596	*/
2597	pVmcbNstGstCtrl->NestedPagingCtrl.n.u1NestedPaging = pVCpu->CTX_SUFF(pVM)->hmr0.s.fNestedPaging;
2598
2599	/* Always enable V_INTR_MASKING as we do not want to allow access to the physical APIC TPR. */
2600	pVmcbNstGstCtrl->IntCtrl.n.u1VIntrMasking = 1;
2601
2602	/*
2603	* Turn off TPR syncing on #VMEXIT for nested-guests as CR8 intercepts are subject
2604	* to the nested-guest intercepts and we always run with V_INTR_MASKING.
2605	*/
2606	pVCpu->hmr0.s.svm.fSyncVTpr = false;
2607
2608	# ifdef DEBUG_ramshankar
2609	/* For debugging purposes - copy the LBR info. from outer guest VMCB. */
2610	pVmcbNstGstCtrl->LbrVirt.n.u1LbrVirt = pVmcb->ctrl.LbrVirt.n.u1LbrVirt;
2611	# endif
2612
2613	/*
2614	* If we don't expose Virtualized-VMSAVE/VMLOAD feature to the outer guest, we
2615	* need to intercept VMSAVE/VMLOAD instructions executed by the nested-guest.
2616	*/
2617	if (!pVCpu->CTX_SUFF(pVM)->cpum.ro.GuestFeatures.fSvmVirtVmsaveVmload)
2618	pVmcbNstGstCtrl->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_VMSAVE
2619	\| SVM_CTRL_INTERCEPT_VMLOAD;
2620
2621	/*
2622	* If we don't expose Virtual GIF feature to the outer guest, we need to intercept
2623	* CLGI/STGI instructions executed by the nested-guest.
2624	*/
2625	if (!pVCpu->CTX_SUFF(pVM)->cpum.ro.GuestFeatures.fSvmVGif)
2626	pVmcbNstGstCtrl->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_CLGI
2627	\| SVM_CTRL_INTERCEPT_STGI;
2628
2629	/* Merge the guest and nested-guest intercepts. */
2630	hmR0SvmMergeVmcbCtrlsNested(pVCpu);
2631
2632	/* Update the VMCB clean bits. */
2633	pVmcbNstGstCtrl->u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2634	}
2635	else
2636	{
2637	Assert(!pVCpu->hmr0.s.svm.fSyncVTpr);
2638	Assert(pVmcbNstGstCtrl->u64IOPMPhysAddr == g_HCPhysIOBitmap);
2639	Assert(RT_BOOL(pVmcbNstGstCtrl->NestedPagingCtrl.n.u1NestedPaging) == pVCpu->CTX_SUFF(pVM)->hmr0.s.fNestedPaging);
2640	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPagingCfg == pVCpu->CTX_SUFF(pVM)->hmr0.s.fNestedPaging);
2641	}
2642	}
2643
2644	#endif /* VBOX_WITH_NESTED_HWVIRT_SVM */
2645
2646	/**
2647	* Exports the state shared between the host and guest (or nested-guest) into
2648	* the VMCB.
2649	*
2650	* @param pVCpu The cross context virtual CPU structure.
2651	* @param pVmcb Pointer to the VM control block.
2652	*
2653	* @remarks No-long-jump zone!!!
2654	*/
2655	static void hmR0SvmExportSharedState(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
2656	{
2657	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2658	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2659
2660	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_DR_MASK)
2661	hmR0SvmExportSharedDebugState(pVCpu, pVmcb);
2662
2663	pVCpu->hm.s.fCtxChanged &= ~HM_CHANGED_GUEST_DR_MASK;
2664	AssertMsg(!(pVCpu->hm.s.fCtxChanged & HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE),
2665	("fCtxChanged=%#RX64\n", pVCpu->hm.s.fCtxChanged));
2666	}
2667
2668
2669	/**
2670	* Worker for SVMR0ImportStateOnDemand.
2671	*
2672	* @param pVCpu The cross context virtual CPU structure.
2673	* @param fWhat What to import, CPUMCTX_EXTRN_XXX.
2674	*/
2675	static void hmR0SvmImportGuestState(PVMCPUCC pVCpu, uint64_t fWhat)
2676	{
2677	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatImportGuestState, x);
2678
2679	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2680	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
2681	PCSVMVMCBSTATESAVE pVmcbGuest = &pVmcb->guest;
2682	PCSVMVMCBCTRL pVmcbCtrl = &pVmcb->ctrl;
2683
2684	/*
2685	* We disable interrupts to make the updating of the state and in particular
2686	* the fExtrn modification atomic wrt to preemption hooks.
2687	*/
2688	RTCCUINTREG const fEFlags = ASMIntDisableFlags();
2689
2690	fWhat &= pCtx->fExtrn;
2691	if (fWhat)
2692	{
2693	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
2694	if (fWhat & CPUMCTX_EXTRN_HWVIRT)
2695	{
2696	if (pVmcbCtrl->IntCtrl.n.u1VGifEnable)
2697	{
2698	Assert(!CPUMIsGuestInSvmNestedHwVirtMode(pCtx)); /* We don't yet support passing VGIF feature to the guest. */
2699	Assert(HMIsSvmVGifActive(pVCpu->CTX_SUFF(pVM))); /* VM has configured it. */
2700	CPUMSetGuestGif(pCtx, pVmcbCtrl->IntCtrl.n.u1VGif);
2701	}
2702	}
2703
2704	if (fWhat & CPUMCTX_EXTRN_HM_SVM_HWVIRT_VIRQ)
2705	{
2706	if ( !pVmcbCtrl->IntCtrl.n.u1VIrqPending
2707	&& VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_INTERRUPT_NESTED_GUEST))
2708	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INTERRUPT_NESTED_GUEST);
2709	}
2710	#endif
2711
2712	if (fWhat & CPUMCTX_EXTRN_INHIBIT_INT)
2713	{
2714	if (pVmcbCtrl->IntShadow.n.u1IntShadow)
2715	EMSetInhibitInterruptsPC(pVCpu, pVmcbGuest->u64RIP);
2716	else if (VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS))
2717	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS);
2718	}
2719
2720	if (fWhat & CPUMCTX_EXTRN_RIP)
2721	pCtx->rip = pVmcbGuest->u64RIP;
2722
2723	if (fWhat & CPUMCTX_EXTRN_RFLAGS)
2724	pCtx->eflags.u32 = pVmcbGuest->u64RFlags;
2725
2726	if (fWhat & CPUMCTX_EXTRN_RSP)
2727	pCtx->rsp = pVmcbGuest->u64RSP;
2728
2729	if (fWhat & CPUMCTX_EXTRN_RAX)
2730	pCtx->rax = pVmcbGuest->u64RAX;
2731
2732	if (fWhat & CPUMCTX_EXTRN_SREG_MASK)
2733	{
2734	if (fWhat & CPUMCTX_EXTRN_CS)
2735	{
2736	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, CS, cs);
2737	/* Correct the CS granularity bit. Haven't seen it being wrong in any other register (yet). */
2738	/** @todo SELM might need to be fixed as it too should not care about the
2739	* granularity bit. See @bugref{6785}. */
2740	if ( !pCtx->cs.Attr.n.u1Granularity
2741	&& pCtx->cs.Attr.n.u1Present
2742	&& pCtx->cs.u32Limit > UINT32_C(0xfffff))
2743	{
2744	Assert((pCtx->cs.u32Limit & 0xfff) == 0xfff);
2745	pCtx->cs.Attr.n.u1Granularity = 1;
2746	}
2747	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, cs);
2748	}
2749	if (fWhat & CPUMCTX_EXTRN_SS)
2750	{
2751	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, SS, ss);
2752	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, ss);
2753	/*
2754	* Sync the hidden SS DPL field. AMD CPUs have a separate CPL field in the
2755	* VMCB and uses that and thus it's possible that when the CPL changes during
2756	* guest execution that the SS DPL isn't updated by AMD-V. Observed on some
2757	* AMD Fusion CPUs with 64-bit guests.
2758	*
2759	* See AMD spec. 15.5.1 "Basic operation".
2760	*/
2761	Assert(!(pVmcbGuest->u8CPL & ~0x3));
2762	uint8_t const uCpl = pVmcbGuest->u8CPL;
2763	if (pCtx->ss.Attr.n.u2Dpl != uCpl)
2764	pCtx->ss.Attr.n.u2Dpl = uCpl & 0x3;
2765	}
2766	if (fWhat & CPUMCTX_EXTRN_DS)
2767	{
2768	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, DS, ds);
2769	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, ds);
2770	}
2771	if (fWhat & CPUMCTX_EXTRN_ES)
2772	{
2773	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, ES, es);
2774	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, es);
2775	}
2776	if (fWhat & CPUMCTX_EXTRN_FS)
2777	{
2778	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, FS, fs);
2779	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, fs);
2780	}
2781	if (fWhat & CPUMCTX_EXTRN_GS)
2782	{
2783	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, GS, gs);
2784	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, gs);
2785	}
2786	}
2787
2788	if (fWhat & CPUMCTX_EXTRN_TABLE_MASK)
2789	{
2790	if (fWhat & CPUMCTX_EXTRN_TR)
2791	{
2792	/*
2793	* Fixup TR attributes so it's compatible with Intel. Important when saved-states
2794	* are used between Intel and AMD, see @bugref{6208#c39}.
2795	* ASSUME that it's normally correct and that we're in 32-bit or 64-bit mode.
2796	*/
2797	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, TR, tr);
2798	if (pCtx->tr.Attr.n.u4Type != X86_SEL_TYPE_SYS_386_TSS_BUSY)
2799	{
2800	if ( pCtx->tr.Attr.n.u4Type == X86_SEL_TYPE_SYS_386_TSS_AVAIL
2801	\|\| CPUMIsGuestInLongModeEx(pCtx))
2802	pCtx->tr.Attr.n.u4Type = X86_SEL_TYPE_SYS_386_TSS_BUSY;
2803	else if (pCtx->tr.Attr.n.u4Type == X86_SEL_TYPE_SYS_286_TSS_AVAIL)
2804	pCtx->tr.Attr.n.u4Type = X86_SEL_TYPE_SYS_286_TSS_BUSY;
2805	}
2806	}
2807
2808	if (fWhat & CPUMCTX_EXTRN_LDTR)
2809	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, LDTR, ldtr);
2810
2811	if (fWhat & CPUMCTX_EXTRN_GDTR)
2812	{
2813	pCtx->gdtr.cbGdt = pVmcbGuest->GDTR.u32Limit;
2814	pCtx->gdtr.pGdt = pVmcbGuest->GDTR.u64Base;
2815	}
2816
2817	if (fWhat & CPUMCTX_EXTRN_IDTR)
2818	{
2819	pCtx->idtr.cbIdt = pVmcbGuest->IDTR.u32Limit;
2820	pCtx->idtr.pIdt = pVmcbGuest->IDTR.u64Base;
2821	}
2822	}
2823
2824	if (fWhat & CPUMCTX_EXTRN_SYSCALL_MSRS)
2825	{
2826	pCtx->msrSTAR = pVmcbGuest->u64STAR;
2827	pCtx->msrLSTAR = pVmcbGuest->u64LSTAR;
2828	pCtx->msrCSTAR = pVmcbGuest->u64CSTAR;
2829	pCtx->msrSFMASK = pVmcbGuest->u64SFMASK;
2830	}
2831
2832	if ( (fWhat & CPUMCTX_EXTRN_SYSENTER_MSRS)
2833	&& !pVCpu->hm.s.svm.fEmulateLongModeSysEnterExit /* Intercepted. AMD-V would clear the high 32 bits of EIP & ESP. */)
2834	{
2835	pCtx->SysEnter.cs = pVmcbGuest->u64SysEnterCS;
2836	pCtx->SysEnter.eip = pVmcbGuest->u64SysEnterEIP;
2837	pCtx->SysEnter.esp = pVmcbGuest->u64SysEnterESP;
2838	}
2839
2840	if (fWhat & CPUMCTX_EXTRN_KERNEL_GS_BASE)
2841	pCtx->msrKERNELGSBASE = pVmcbGuest->u64KernelGSBase;
2842
2843	if (fWhat & CPUMCTX_EXTRN_DR_MASK)
2844	{
2845	if (fWhat & CPUMCTX_EXTRN_DR6)
2846	{
2847	if (!pVCpu->hmr0.s.fUsingHyperDR7)
2848	pCtx->dr[6] = pVmcbGuest->u64DR6;
2849	else
2850	CPUMSetHyperDR6(pVCpu, pVmcbGuest->u64DR6);
2851	}
2852
2853	if (fWhat & CPUMCTX_EXTRN_DR7)
2854	{
2855	if (!pVCpu->hmr0.s.fUsingHyperDR7)
2856	pCtx->dr[7] = pVmcbGuest->u64DR7;
2857	else
2858	Assert(pVmcbGuest->u64DR7 == CPUMGetHyperDR7(pVCpu));
2859	}
2860	}
2861
2862	if (fWhat & CPUMCTX_EXTRN_CR_MASK)
2863	{
2864	if (fWhat & CPUMCTX_EXTRN_CR0)
2865	{
2866	/* We intercept changes to all CR0 bits except maybe TS & MP bits. */
2867	uint64_t const uCr0 = (pCtx->cr0 & ~(X86_CR0_TS \| X86_CR0_MP))
2868	\| (pVmcbGuest->u64CR0 & (X86_CR0_TS \| X86_CR0_MP));
2869	VMMRZCallRing3Disable(pVCpu); /* Calls into PGM which has Log statements. */
2870	CPUMSetGuestCR0(pVCpu, uCr0);
2871	VMMRZCallRing3Enable(pVCpu);
2872	}
2873
2874	if (fWhat & CPUMCTX_EXTRN_CR2)
2875	pCtx->cr2 = pVmcbGuest->u64CR2;
2876
2877	if (fWhat & CPUMCTX_EXTRN_CR3)
2878	{
2879	if ( pVmcbCtrl->NestedPagingCtrl.n.u1NestedPaging
2880	&& pCtx->cr3 != pVmcbGuest->u64CR3)
2881	{
2882	CPUMSetGuestCR3(pVCpu, pVmcbGuest->u64CR3);
2883	VMCPU_FF_SET(pVCpu, VMCPU_FF_HM_UPDATE_CR3);
2884	}
2885	}
2886
2887	/* Changes to CR4 are always intercepted. */
2888	}
2889
2890	/* Update fExtrn. */
2891	pCtx->fExtrn &= ~fWhat;
2892
2893	/* If everything has been imported, clear the HM keeper bit. */
2894	if (!(pCtx->fExtrn & HMSVM_CPUMCTX_EXTRN_ALL))
2895	{
2896	pCtx->fExtrn &= ~CPUMCTX_EXTRN_KEEPER_HM;
2897	Assert(!pCtx->fExtrn);
2898	}
2899	}
2900	else
2901	Assert(!pCtx->fExtrn \|\| (pCtx->fExtrn & HMSVM_CPUMCTX_EXTRN_ALL));
2902
2903	ASMSetFlags(fEFlags);
2904
2905	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatImportGuestState, x);
2906
2907	/*
2908	* Honor any pending CR3 updates.
2909	*
2910	* Consider this scenario: #VMEXIT -> VMMRZCallRing3Enable() -> do stuff that causes a longjmp
2911	* -> SVMR0CallRing3Callback() -> VMMRZCallRing3Disable() -> hmR0SvmImportGuestState()
2912	* -> Sets VMCPU_FF_HM_UPDATE_CR3 pending -> return from the longjmp -> continue with #VMEXIT
2913	* handling -> hmR0SvmImportGuestState() and here we are.
2914	*
2915	* The reason for such complicated handling is because VM-exits that call into PGM expect
2916	* CR3 to be up-to-date and thus any CR3-saves -before- the VM-exit (longjmp) would've
2917	* postponed the CR3 update via the force-flag and cleared CR3 from fExtrn. Any SVM R0
2918	* VM-exit handler that requests CR3 to be saved will end up here and we call PGMUpdateCR3().
2919	*
2920	* The longjmp exit path can't check these CR3 force-flags and call code that takes a lock again,
2921	* and does not process force-flag like regular exits to ring-3 either, we cover for it here.
2922	*/
2923	if ( VMMRZCallRing3IsEnabled(pVCpu)
2924	&& VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_HM_UPDATE_CR3))
2925	{
2926	AssertMsg(pCtx->cr3 == pVmcbGuest->u64CR3, ("cr3=%#RX64 vmcb_cr3=%#RX64\n", pCtx->cr3, pVmcbGuest->u64CR3));
2927	PGMUpdateCR3(pVCpu, pCtx->cr3, false /* fCr3Mapped */);
2928	}
2929	}
2930
2931
2932	/**
2933	* Saves the guest (or nested-guest) state from the VMCB into the guest-CPU
2934	* context.
2935	*
2936	* Currently there is no residual state left in the CPU that is not updated in the
2937	* VMCB.
2938	*
2939	* @returns VBox status code.
2940	* @param pVCpu The cross context virtual CPU structure.
2941	* @param fWhat What to import, CPUMCTX_EXTRN_XXX.
2942	*/
2943	VMMR0DECL(int) SVMR0ImportStateOnDemand(PVMCPUCC pVCpu, uint64_t fWhat)
2944	{
2945	hmR0SvmImportGuestState(pVCpu, fWhat);
2946	return VINF_SUCCESS;
2947	}
2948
2949
2950	/**
2951	* Does the necessary state syncing before returning to ring-3 for any reason
2952	* (longjmp, preemption, voluntary exits to ring-3) from AMD-V.
2953	*
2954	* @param pVCpu The cross context virtual CPU structure.
2955	* @param fImportState Whether to import the guest state from the VMCB back
2956	* to the guest-CPU context.
2957	*
2958	* @remarks No-long-jmp zone!!!
2959	*/
2960	static void hmR0SvmLeave(PVMCPUCC pVCpu, bool fImportState)
2961	{
2962	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2963	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2964
2965	/*
2966	* !!! IMPORTANT !!!
2967	* If you modify code here, make sure to check whether SVMR0CallRing3Callback() needs to be updated too.
2968	*/
2969
2970	/* Save the guest state if necessary. */
2971	if (fImportState)
2972	hmR0SvmImportGuestState(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
2973
2974	/* Restore host FPU state if necessary and resync on next R0 reentry. */
2975	CPUMR0FpuStateMaybeSaveGuestAndRestoreHost(pVCpu);
2976	Assert(!CPUMIsGuestFPUStateActive(pVCpu));
2977
2978	/*
2979	* Restore host debug registers if necessary and resync on next R0 reentry.
2980	*/
2981	#ifdef VBOX_STRICT
2982	if (CPUMIsHyperDebugStateActive(pVCpu))
2983	{
2984	PSVMVMCB pVmcb = pVCpu->hmr0.s.svm.pVmcb; /** @todo nested-guest. */
2985	Assert(pVmcb->ctrl.u16InterceptRdDRx == 0xffff);
2986	Assert(pVmcb->ctrl.u16InterceptWrDRx == 0xffff);
2987	}
2988	#endif
2989	CPUMR0DebugStateMaybeSaveGuestAndRestoreHost(pVCpu, false /* save DR6 */);
2990	Assert(!CPUMIsHyperDebugStateActive(pVCpu));
2991	Assert(!CPUMIsGuestDebugStateActive(pVCpu));
2992
2993	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatEntry);
2994	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatImportGuestState);
2995	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatExportGuestState);
2996	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatPreExit);
2997	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatExitHandling);
2998	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatExitVmentry);
2999	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchLongJmpToR3);
3000
3001	VMCPU_CMPXCHG_STATE(pVCpu, VMCPUSTATE_STARTED_HM, VMCPUSTATE_STARTED_EXEC);
3002	}
3003
3004
3005	/**
3006	* Leaves the AMD-V session.
3007	*
3008	* Only used while returning to ring-3 either due to longjump or exits to
3009	* ring-3.
3010	*
3011	* @returns VBox status code.
3012	* @param pVCpu The cross context virtual CPU structure.
3013	*/
3014	static int hmR0SvmLeaveSession(PVMCPUCC pVCpu)
3015	{
3016	HM_DISABLE_PREEMPT(pVCpu);
3017	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
3018	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
3019
3020	/* When thread-context hooks are used, we can avoid doing the leave again if we had been preempted before
3021	and done this from the SVMR0ThreadCtxCallback(). */
3022	if (!pVCpu->hmr0.s.fLeaveDone)
3023	{
3024	hmR0SvmLeave(pVCpu, true /* fImportState */);
3025	pVCpu->hmr0.s.fLeaveDone = true;
3026	}
3027
3028	/*
3029	* !!! IMPORTANT !!!
3030	* If you modify code here, make sure to check whether SVMR0CallRing3Callback() needs to be updated too.
3031	*/
3032
3033	/** @todo eliminate the need for calling VMMR0ThreadCtxHookDisable here! */
3034	/* Deregister hook now that we've left HM context before re-enabling preemption. */
3035	VMMR0ThreadCtxHookDisable(pVCpu);
3036
3037	/* Leave HM context. This takes care of local init (term). */
3038	int rc = HMR0LeaveCpu(pVCpu);
3039
3040	HM_RESTORE_PREEMPT();
3041	return rc;
3042	}
3043
3044
3045	/**
3046	* VMMRZCallRing3() callback wrapper which saves the guest state (or restores
3047	* any remaining host state) before we go back to ring-3 due to an assertion.
3048	*
3049	* @param pVCpu The cross context virtual CPU structure.
3050	*/
3051	VMMR0DECL(int) SVMR0AssertionCallback(PVMCPUCC pVCpu)
3052	{
3053	/*
3054	* !!! IMPORTANT !!!
3055	* If you modify code here, make sure to check whether hmR0SvmLeave() and hmR0SvmLeaveSession() needs
3056	* to be updated too. This is a stripped down version which gets out ASAP trying to not trigger any assertion.
3057	*/
3058	VMMR0AssertionRemoveNotification(pVCpu);
3059	VMMRZCallRing3Disable(pVCpu);
3060	HM_DISABLE_PREEMPT(pVCpu);
3061
3062	/* Import the entire guest state. */
3063	hmR0SvmImportGuestState(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
3064
3065	/* Restore host FPU state if necessary and resync on next R0 reentry. */
3066	CPUMR0FpuStateMaybeSaveGuestAndRestoreHost(pVCpu);
3067
3068	/* Restore host debug registers if necessary and resync on next R0 reentry. */
3069	CPUMR0DebugStateMaybeSaveGuestAndRestoreHost(pVCpu, false /* save DR6 */);
3070
3071	/* Deregister the hook now that we've left HM context before re-enabling preemption. */
3072	/** @todo eliminate the need for calling VMMR0ThreadCtxHookDisable here! */
3073	VMMR0ThreadCtxHookDisable(pVCpu);
3074
3075	/* Leave HM context. This takes care of local init (term). */
3076	HMR0LeaveCpu(pVCpu);
3077
3078	HM_RESTORE_PREEMPT();
3079	return VINF_SUCCESS;
3080	}
3081
3082
3083	/**
3084	* Take necessary actions before going back to ring-3.
3085	*
3086	* An action requires us to go back to ring-3. This function does the necessary
3087	* steps before we can safely return to ring-3. This is not the same as longjmps
3088	* to ring-3, this is voluntary.
3089	*
3090	* @returns Strict VBox status code.
3091	* @param pVCpu The cross context virtual CPU structure.
3092	* @param rcExit The reason for exiting to ring-3. Can be
3093	* VINF_VMM_UNKNOWN_RING3_CALL.
3094	*/
3095	static VBOXSTRICTRC hmR0SvmExitToRing3(PVMCPUCC pVCpu, VBOXSTRICTRC rcExit)
3096	{
3097	Assert(pVCpu);
3098	HMSVM_ASSERT_PREEMPT_SAFE(pVCpu);
3099
3100	/* Please, no longjumps here (any logging shouldn't flush jump back to ring-3). NO LOGGING BEFORE THIS POINT! */
3101	VMMRZCallRing3Disable(pVCpu);
3102	Log4Func(("rcExit=%d LocalFF=%#RX64 GlobalFF=%#RX32\n", VBOXSTRICTRC_VAL(rcExit), (uint64_t)pVCpu->fLocalForcedActions,
3103	pVCpu->CTX_SUFF(pVM)->fGlobalForcedActions));
3104
3105	/* We need to do this only while truly exiting the "inner loop" back to ring-3 and -not- for any longjmp to ring3. */
3106	if (pVCpu->hm.s.Event.fPending)
3107	{
3108	hmR0SvmPendingEventToTrpmTrap(pVCpu);
3109	Assert(!pVCpu->hm.s.Event.fPending);
3110	}
3111
3112	/* Sync. the necessary state for going back to ring-3. */
3113	hmR0SvmLeaveSession(pVCpu);
3114	STAM_COUNTER_DEC(&pVCpu->hm.s.StatSwitchLongJmpToR3);
3115
3116	/* Thread-context hooks are unregistered at this point!!! */
3117	/* Ring-3 callback notifications are unregistered at this point!!! */
3118
3119	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_TO_R3);
3120	CPUMSetChangedFlags(pVCpu, CPUM_CHANGED_SYSENTER_MSR
3121	\| CPUM_CHANGED_LDTR
3122	\| CPUM_CHANGED_GDTR
3123	\| CPUM_CHANGED_IDTR
3124	\| CPUM_CHANGED_TR
3125	\| CPUM_CHANGED_HIDDEN_SEL_REGS);
3126	if ( pVCpu->CTX_SUFF(pVM)->hmr0.s.fNestedPaging
3127	&& CPUMIsGuestPagingEnabledEx(&pVCpu->cpum.GstCtx))
3128	{
3129	CPUMSetChangedFlags(pVCpu, CPUM_CHANGED_GLOBAL_TLB_FLUSH);
3130	}
3131
3132	/* Update the exit-to-ring 3 reason. */
3133	pVCpu->hm.s.rcLastExitToR3 = VBOXSTRICTRC_VAL(rcExit);
3134
3135	/* On our way back from ring-3, reload the guest-CPU state if it may change while in ring-3. */
3136	if ( rcExit != VINF_EM_RAW_INTERRUPT
3137	\|\| CPUMIsGuestInSvmNestedHwVirtMode(&pVCpu->cpum.GstCtx))
3138	{
3139	Assert(!(pVCpu->cpum.GstCtx.fExtrn & HMSVM_CPUMCTX_EXTRN_ALL));
3140	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
3141	}
3142
3143	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchExitToR3);
3144	VMMRZCallRing3Enable(pVCpu);
3145
3146	/*
3147	* If we're emulating an instruction, we shouldn't have any TRPM traps pending
3148	* and if we're injecting an event we should have a TRPM trap pending.
3149	*/
3150	AssertReturnStmt(rcExit != VINF_EM_RAW_INJECT_TRPM_EVENT \|\| TRPMHasTrap(pVCpu),
3151	pVCpu->hm.s.u32HMError = VBOXSTRICTRC_VAL(rcExit),
3152	VERR_SVM_IPE_5);
3153	AssertReturnStmt(rcExit != VINF_EM_RAW_EMULATE_INSTR \|\| !TRPMHasTrap(pVCpu),
3154	pVCpu->hm.s.u32HMError = VBOXSTRICTRC_VAL(rcExit),
3155	VERR_SVM_IPE_4);
3156
3157	return rcExit;
3158	}
3159
3160
3161	/**
3162	* Updates the use of TSC offsetting mode for the CPU and adjusts the necessary
3163	* intercepts.
3164	*
3165	* @param pVCpu The cross context virtual CPU structure.
3166	* @param pVmcb Pointer to the VM control block.
3167	*
3168	* @remarks No-long-jump zone!!!
3169	*/
3170	static void hmR0SvmUpdateTscOffsetting(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
3171	{
3172	/*
3173	* Avoid intercepting RDTSC/RDTSCP if we determined the host TSC (++) is stable
3174	* and in case of a nested-guest, if the nested-VMCB specifies it is not intercepting
3175	* RDTSC/RDTSCP as well.
3176	*/
3177	bool fParavirtTsc;
3178	uint64_t uTscOffset;
3179	bool const fCanUseRealTsc = TMCpuTickCanUseRealTSC(pVCpu->CTX_SUFF(pVM), pVCpu, &uTscOffset, &fParavirtTsc);
3180
3181	bool fIntercept;
3182	if (fCanUseRealTsc)
3183	fIntercept = hmR0SvmClearCtrlIntercept(pVCpu, pVmcb, SVM_CTRL_INTERCEPT_RDTSC \| SVM_CTRL_INTERCEPT_RDTSCP);
3184	else
3185	{
3186	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_RDTSC \| SVM_CTRL_INTERCEPT_RDTSCP);
3187	fIntercept = true;
3188	}
3189
3190	if (!fIntercept)
3191	{
3192	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
3193	/* Apply the nested-guest VMCB's TSC offset over the guest TSC offset. */
3194	if (CPUMIsGuestInSvmNestedHwVirtMode(&pVCpu->cpum.GstCtx))
3195	uTscOffset = CPUMApplyNestedGuestTscOffset(pVCpu, uTscOffset);
3196	#endif
3197
3198	/* Update the TSC offset in the VMCB and the relevant clean bits. */
3199	pVmcb->ctrl.u64TSCOffset = uTscOffset;
3200	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
3201	}
3202
3203	/* Currently neither Hyper-V nor KVM need to update their paravirt. TSC
3204	information before every VM-entry, hence we have nothing to do here at the moment. */
3205	if (fParavirtTsc)
3206	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscParavirt);
3207	}
3208
3209
3210	/**
3211	* Sets an event as a pending event to be injected into the guest.
3212	*
3213	* @param pVCpu The cross context virtual CPU structure.
3214	* @param pEvent Pointer to the SVM event.
3215	* @param GCPtrFaultAddress The fault-address (CR2) in case it's a
3216	* page-fault.
3217	*
3218	* @remarks Statistics counter assumes this is a guest event being reflected to
3219	* the guest i.e. 'StatInjectPendingReflect' is incremented always.
3220	*/
3221	DECLINLINE(void) hmR0SvmSetPendingEvent(PVMCPUCC pVCpu, PSVMEVENT pEvent, RTGCUINTPTR GCPtrFaultAddress)
3222	{
3223	Assert(!pVCpu->hm.s.Event.fPending);
3224	Assert(pEvent->n.u1Valid);
3225
3226	pVCpu->hm.s.Event.u64IntInfo = pEvent->u;
3227	pVCpu->hm.s.Event.fPending = true;
3228	pVCpu->hm.s.Event.GCPtrFaultAddress = GCPtrFaultAddress;
3229
3230	Log4Func(("u=%#RX64 u8Vector=%#x Type=%#x ErrorCodeValid=%RTbool ErrorCode=%#RX32\n", pEvent->u, pEvent->n.u8Vector,
3231	(uint8_t)pEvent->n.u3Type, !!pEvent->n.u1ErrorCodeValid, pEvent->n.u32ErrorCode));
3232	}
3233
3234
3235	/**
3236	* Sets an invalid-opcode (\#UD) exception as pending-for-injection into the VM.
3237	*
3238	* @param pVCpu The cross context virtual CPU structure.
3239	*/
3240	DECLINLINE(void) hmR0SvmSetPendingXcptUD(PVMCPUCC pVCpu)
3241	{
3242	SVMEVENT Event;
3243	Event.u = 0;
3244	Event.n.u1Valid = 1;
3245	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3246	Event.n.u8Vector = X86_XCPT_UD;
3247	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3248	}
3249
3250
3251	/**
3252	* Sets a debug (\#DB) exception as pending-for-injection into the VM.
3253	*
3254	* @param pVCpu The cross context virtual CPU structure.
3255	*/
3256	DECLINLINE(void) hmR0SvmSetPendingXcptDB(PVMCPUCC pVCpu)
3257	{
3258	SVMEVENT Event;
3259	Event.u = 0;
3260	Event.n.u1Valid = 1;
3261	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3262	Event.n.u8Vector = X86_XCPT_DB;
3263	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3264	}
3265
3266
3267	/**
3268	* Sets a page fault (\#PF) exception as pending-for-injection into the VM.
3269	*
3270	* @param pVCpu The cross context virtual CPU structure.
3271	* @param u32ErrCode The error-code for the page-fault.
3272	* @param uFaultAddress The page fault address (CR2).
3273	*
3274	* @remarks This updates the guest CR2 with @a uFaultAddress!
3275	*/
3276	DECLINLINE(void) hmR0SvmSetPendingXcptPF(PVMCPUCC pVCpu, uint32_t u32ErrCode, RTGCUINTPTR uFaultAddress)
3277	{
3278	SVMEVENT Event;
3279	Event.u = 0;
3280	Event.n.u1Valid = 1;
3281	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3282	Event.n.u8Vector = X86_XCPT_PF;
3283	Event.n.u1ErrorCodeValid = 1;
3284	Event.n.u32ErrorCode = u32ErrCode;
3285
3286	/* Update CR2 of the guest. */
3287	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_CR2);
3288	if (pVCpu->cpum.GstCtx.cr2 != uFaultAddress)
3289	{
3290	pVCpu->cpum.GstCtx.cr2 = uFaultAddress;
3291	/* The VMCB clean bit for CR2 will be updated while re-loading the guest state. */
3292	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_CR2);
3293	}
3294
3295	hmR0SvmSetPendingEvent(pVCpu, &Event, uFaultAddress);
3296	}
3297
3298
3299	/**
3300	* Sets a math-fault (\#MF) exception as pending-for-injection into the VM.
3301	*
3302	* @param pVCpu The cross context virtual CPU structure.
3303	*/
3304	DECLINLINE(void) hmR0SvmSetPendingXcptMF(PVMCPUCC pVCpu)
3305	{
3306	SVMEVENT Event;
3307	Event.u = 0;
3308	Event.n.u1Valid = 1;
3309	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3310	Event.n.u8Vector = X86_XCPT_MF;
3311	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3312	}
3313
3314
3315	/**
3316	* Sets a double fault (\#DF) exception as pending-for-injection into the VM.
3317	*
3318	* @param pVCpu The cross context virtual CPU structure.
3319	*/
3320	DECLINLINE(void) hmR0SvmSetPendingXcptDF(PVMCPUCC pVCpu)
3321	{
3322	SVMEVENT Event;
3323	Event.u = 0;
3324	Event.n.u1Valid = 1;
3325	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3326	Event.n.u8Vector = X86_XCPT_DF;
3327	Event.n.u1ErrorCodeValid = 1;
3328	Event.n.u32ErrorCode = 0;
3329	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3330	}
3331
3332
3333	/**
3334	* Injects an event into the guest upon VMRUN by updating the relevant field
3335	* in the VMCB.
3336	*
3337	* @param pVCpu The cross context virtual CPU structure.
3338	* @param pVmcb Pointer to the guest VM control block.
3339	* @param pEvent Pointer to the event.
3340	*
3341	* @remarks No-long-jump zone!!!
3342	* @remarks Requires CR0!
3343	*/
3344	DECLINLINE(void) hmR0SvmInjectEventVmcb(PVMCPUCC pVCpu, PSVMVMCB pVmcb, PSVMEVENT pEvent)
3345	{
3346	Assert(!pVmcb->ctrl.EventInject.n.u1Valid);
3347	pVmcb->ctrl.EventInject.u = pEvent->u;
3348	if ( pVmcb->ctrl.EventInject.n.u3Type == SVM_EVENT_EXCEPTION
3349	\|\| pVmcb->ctrl.EventInject.n.u3Type == SVM_EVENT_NMI)
3350	{
3351	Assert(pEvent->n.u8Vector <= X86_XCPT_LAST);
3352	STAM_COUNTER_INC(&pVCpu->hm.s.aStatInjectedXcpts[pEvent->n.u8Vector]);
3353	}
3354	else
3355	STAM_COUNTER_INC(&pVCpu->hm.s.aStatInjectedIrqs[pEvent->n.u8Vector & MASK_INJECT_IRQ_STAT]);
3356	RT_NOREF(pVCpu);
3357
3358	Log4Func(("u=%#RX64 u8Vector=%#x Type=%#x ErrorCodeValid=%RTbool ErrorCode=%#RX32\n", pEvent->u, pEvent->n.u8Vector,
3359	(uint8_t)pEvent->n.u3Type, !!pEvent->n.u1ErrorCodeValid, pEvent->n.u32ErrorCode));
3360	}
3361
3362
3363
3364	/**
3365	* Converts any TRPM trap into a pending HM event. This is typically used when
3366	* entering from ring-3 (not longjmp returns).
3367	*
3368	* @param pVCpu The cross context virtual CPU structure.
3369	*/
3370	static void hmR0SvmTrpmTrapToPendingEvent(PVMCPUCC pVCpu)
3371	{
3372	Assert(TRPMHasTrap(pVCpu));
3373	Assert(!pVCpu->hm.s.Event.fPending);
3374
3375	uint8_t uVector;
3376	TRPMEVENT enmTrpmEvent;
3377	uint32_t uErrCode;
3378	RTGCUINTPTR GCPtrFaultAddress;
3379	uint8_t cbInstr;
3380
3381	int rc = TRPMQueryTrapAll(pVCpu, &uVector, &enmTrpmEvent, &uErrCode, &GCPtrFaultAddress, &cbInstr, NULL /* pfIcebp */);
3382	AssertRC(rc);
3383
3384	SVMEVENT Event;
3385	Event.u = 0;
3386	Event.n.u1Valid = 1;
3387	Event.n.u8Vector = uVector;
3388
3389	/* Refer AMD spec. 15.20 "Event Injection" for the format. */
3390	if (enmTrpmEvent == TRPM_TRAP)
3391	{
3392	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3393	switch (uVector)
3394	{
3395	case X86_XCPT_NMI:
3396	{
3397	Event.n.u3Type = SVM_EVENT_NMI;
3398	break;
3399	}
3400
3401	case X86_XCPT_BP:
3402	case X86_XCPT_OF:
3403	AssertMsgFailed(("Invalid TRPM vector %d for event type %d\n", uVector, enmTrpmEvent));
3404	RT_FALL_THRU();
3405
3406	case X86_XCPT_PF:
3407	case X86_XCPT_DF:
3408	case X86_XCPT_TS:
3409	case X86_XCPT_NP:
3410	case X86_XCPT_SS:
3411	case X86_XCPT_GP:
3412	case X86_XCPT_AC:
3413	{
3414	Event.n.u1ErrorCodeValid = 1;
3415	Event.n.u32ErrorCode = uErrCode;
3416	break;
3417	}
3418	}
3419	}
3420	else if (enmTrpmEvent == TRPM_HARDWARE_INT)
3421	Event.n.u3Type = SVM_EVENT_EXTERNAL_IRQ;
3422	else if (enmTrpmEvent == TRPM_SOFTWARE_INT)
3423	Event.n.u3Type = SVM_EVENT_SOFTWARE_INT;
3424	else
3425	AssertMsgFailed(("Invalid TRPM event type %d\n", enmTrpmEvent));
3426
3427	rc = TRPMResetTrap(pVCpu);
3428	AssertRC(rc);
3429
3430	Log4(("TRPM->HM event: u=%#RX64 u8Vector=%#x uErrorCodeValid=%RTbool uErrorCode=%#RX32\n", Event.u, Event.n.u8Vector,
3431	!!Event.n.u1ErrorCodeValid, Event.n.u32ErrorCode));
3432
3433	hmR0SvmSetPendingEvent(pVCpu, &Event, GCPtrFaultAddress);
3434	}
3435
3436
3437	/**
3438	* Converts any pending SVM event into a TRPM trap. Typically used when leaving
3439	* AMD-V to execute any instruction.
3440	*
3441	* @param pVCpu The cross context virtual CPU structure.
3442	*/
3443	static void hmR0SvmPendingEventToTrpmTrap(PVMCPUCC pVCpu)
3444	{
3445	Assert(pVCpu->hm.s.Event.fPending);
3446	Assert(TRPMQueryTrap(pVCpu, NULL /* pu8TrapNo /, NULL / pEnmType */) == VERR_TRPM_NO_ACTIVE_TRAP);
3447
3448	SVMEVENT Event;
3449	Event.u = pVCpu->hm.s.Event.u64IntInfo;
3450
3451	uint8_t uVector = Event.n.u8Vector;
3452	TRPMEVENT enmTrapType = HMSvmEventToTrpmEventType(&Event, uVector);
3453
3454	Log4(("HM event->TRPM: uVector=%#x enmTrapType=%d\n", uVector, Event.n.u3Type));
3455
3456	int rc = TRPMAssertTrap(pVCpu, uVector, enmTrapType);
3457	AssertRC(rc);
3458
3459	if (Event.n.u1ErrorCodeValid)
3460	TRPMSetErrorCode(pVCpu, Event.n.u32ErrorCode);
3461
3462	if ( enmTrapType == TRPM_TRAP
3463	&& uVector == X86_XCPT_PF)
3464	{
3465	TRPMSetFaultAddress(pVCpu, pVCpu->hm.s.Event.GCPtrFaultAddress);
3466	Assert(pVCpu->hm.s.Event.GCPtrFaultAddress == CPUMGetGuestCR2(pVCpu));
3467	}
3468	else if (enmTrapType == TRPM_SOFTWARE_INT)
3469	TRPMSetInstrLength(pVCpu, pVCpu->hm.s.Event.cbInstr);
3470	pVCpu->hm.s.Event.fPending = false;
3471	}
3472
3473
3474	/**
3475	* Checks if the guest (or nested-guest) has an interrupt shadow active right
3476	* now.
3477	*
3478	* @returns @c true if the interrupt shadow is active, @c false otherwise.
3479	* @param pVCpu The cross context virtual CPU structure.
3480	*
3481	* @remarks No-long-jump zone!!!
3482	* @remarks Has side-effects with VMCPU_FF_INHIBIT_INTERRUPTS force-flag.
3483	*/
3484	static bool hmR0SvmIsIntrShadowActive(PVMCPUCC pVCpu)
3485	{
3486	/*
3487	* Instructions like STI and MOV SS inhibit interrupts till the next instruction
3488	* completes. Check if we should inhibit interrupts or clear any existing
3489	* interrupt inhibition.
3490	*/
3491	if (VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS))
3492	{
3493	if (pVCpu->cpum.GstCtx.rip != EMGetInhibitInterruptsPC(pVCpu))
3494	{
3495	/*
3496	* We can clear the inhibit force flag as even if we go back to the recompiler
3497	* without executing guest code in AMD-V, the flag's condition to be cleared is
3498	* met and thus the cleared state is correct.
3499	*/
3500	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS);
3501	return false;
3502	}
3503	return true;
3504	}
3505	return false;
3506	}
3507
3508
3509	/**
3510	* Sets the virtual interrupt intercept control in the VMCB.
3511	*
3512	* @param pVCpu The cross context virtual CPU structure.
3513	* @param pVmcb Pointer to the VM control block.
3514	*/
3515	static void hmR0SvmSetIntWindowExiting(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
3516	{
3517	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(&pVCpu->cpum.GstCtx); NOREF(pVCpu);
3518
3519	/*
3520	* When AVIC isn't supported, set up an interrupt window to cause a #VMEXIT when the guest
3521	* is ready to accept interrupts. At #VMEXIT, we then get the interrupt from the APIC
3522	* (updating ISR at the right time) and inject the interrupt.
3523	*
3524	* With AVIC is supported, we could make use of the asynchronously delivery without
3525	* #VMEXIT and we would be passing the AVIC page to SVM.
3526	*
3527	* In AMD-V, an interrupt window is achieved using a combination of V_IRQ (an interrupt
3528	* is pending), V_IGN_TPR (ignore TPR priorities) and the VINTR intercept all being set.
3529	*/
3530	Assert(pVmcb->ctrl.IntCtrl.n.u1IgnoreTPR);
3531	pVmcb->ctrl.IntCtrl.n.u1VIrqPending = 1;
3532	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INT_CTRL;
3533	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_VINTR);
3534	Log4(("Set VINTR intercept\n"));
3535	}
3536
3537
3538	/**
3539	* Clears the virtual interrupt intercept control in the VMCB as
3540	* we are figured the guest is unable process any interrupts
3541	* at this point of time.
3542	*
3543	* @param pVCpu The cross context virtual CPU structure.
3544	* @param pVmcb Pointer to the VM control block.
3545	*/
3546	static void hmR0SvmClearIntWindowExiting(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
3547	{
3548	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(&pVCpu->cpum.GstCtx); NOREF(pVCpu);
3549
3550	PSVMVMCBCTRL pVmcbCtrl = &pVmcb->ctrl;
3551	if ( pVmcbCtrl->IntCtrl.n.u1VIrqPending
3552	\|\| (pVmcbCtrl->u64InterceptCtrl & SVM_CTRL_INTERCEPT_VINTR))
3553	{
3554	pVmcbCtrl->IntCtrl.n.u1VIrqPending = 0;
3555	pVmcbCtrl->u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INT_CTRL;
3556	hmR0SvmClearCtrlIntercept(pVCpu, pVmcb, SVM_CTRL_INTERCEPT_VINTR);
3557	Log4(("Cleared VINTR intercept\n"));
3558	}
3559	}
3560
3561
3562	/**
3563	* Evaluates the event to be delivered to the guest and sets it as the pending
3564	* event.
3565	*
3566	* @returns Strict VBox status code.
3567	* @param pVCpu The cross context virtual CPU structure.
3568	* @param pSvmTransient Pointer to the SVM transient structure.
3569	*/
3570	static VBOXSTRICTRC hmR0SvmEvaluatePendingEvent(PVMCPUCC pVCpu, PCSVMTRANSIENT pSvmTransient)
3571	{
3572	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
3573	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_HWVIRT
3574	\| CPUMCTX_EXTRN_RFLAGS
3575	\| CPUMCTX_EXTRN_INHIBIT_INT
3576	\| CPUMCTX_EXTRN_HM_SVM_HWVIRT_VIRQ);
3577
3578	Assert(!pVCpu->hm.s.Event.fPending);
3579	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
3580	Assert(pVmcb);
3581
3582	bool const fGif = CPUMGetGuestGif(pCtx);
3583	bool const fIntShadow = hmR0SvmIsIntrShadowActive(pVCpu);
3584	bool const fBlockNmi = VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_BLOCK_NMIS);
3585
3586	Log4Func(("fGif=%RTbool fBlockNmi=%RTbool fIntShadow=%RTbool fIntPending=%RTbool fNmiPending=%RTbool\n",
3587	fGif, fBlockNmi, fIntShadow, VMCPU_FF_IS_ANY_SET(pVCpu, VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC),
3588	VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_INTERRUPT_NMI)));
3589
3590	/** @todo SMI. SMIs take priority over NMIs. */
3591
3592	/*
3593	* Check if the guest or nested-guest can receive NMIs.
3594	* Nested NMIs are not allowed, see AMD spec. 8.1.4 "Masking External Interrupts".
3595	* NMIs take priority over maskable interrupts, see AMD spec. 8.5 "Priorities".
3596	*/
3597	if ( VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_INTERRUPT_NMI)
3598	&& !fBlockNmi)
3599	{
3600	if ( fGif
3601	&& !fIntShadow)
3602	{
3603	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
3604	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_NMI))
3605	{
3606	Log4(("Intercepting NMI -> #VMEXIT\n"));
3607	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
3608	return IEMExecSvmVmexit(pVCpu, SVM_EXIT_NMI, 0, 0);
3609	}
3610	#endif
3611	Log4(("Setting NMI pending for injection\n"));
3612	SVMEVENT Event;
3613	Event.u = 0;
3614	Event.n.u1Valid = 1;
3615	Event.n.u8Vector = X86_XCPT_NMI;
3616	Event.n.u3Type = SVM_EVENT_NMI;
3617	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3618	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INTERRUPT_NMI);
3619	}
3620	else if (!fGif)
3621	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_STGI);
3622	else if (!pSvmTransient->fIsNestedGuest)
3623	hmR0SvmSetIntWindowExiting(pVCpu, pVmcb);
3624	/* else: for nested-guests, interrupt-window exiting will be picked up when merging VMCB controls. */
3625	}
3626	/*
3627	* Check if the guest can receive external interrupts (PIC/APIC). Once PDMGetInterrupt()
3628	* returns a valid interrupt we -must- deliver the interrupt. We can no longer re-request
3629	* it from the APIC device.
3630	*
3631	* For nested-guests, physical interrupts always take priority over virtual interrupts.
3632	* We don't need to inject nested-guest virtual interrupts here, we can let the hardware
3633	* do that work when we execute nested-guest code esp. since all the required information
3634	* is in the VMCB, unlike physical interrupts where we need to fetch the interrupt from
3635	* the virtual interrupt controller.
3636	*
3637	* See AMD spec. 15.21.4 "Injecting Virtual (INTR) Interrupts".
3638	*/
3639	else if ( VMCPU_FF_IS_ANY_SET(pVCpu, VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC)
3640	&& !pVCpu->hm.s.fSingleInstruction)
3641	{
3642	bool const fBlockInt = !pSvmTransient->fIsNestedGuest ? !(pCtx->eflags.u32 & X86_EFL_IF)
3643	: CPUMIsGuestSvmPhysIntrEnabled(pVCpu, pCtx);
3644	if ( fGif
3645	&& !fBlockInt
3646	&& !fIntShadow)
3647	{
3648	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
3649	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_INTR))
3650	{
3651	Log4(("Intercepting INTR -> #VMEXIT\n"));
3652	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
3653	return IEMExecSvmVmexit(pVCpu, SVM_EXIT_INTR, 0, 0);
3654	}
3655	#endif
3656	uint8_t u8Interrupt;
3657	int rc = PDMGetInterrupt(pVCpu, &u8Interrupt);
3658	if (RT_SUCCESS(rc))
3659	{
3660	Log4(("Setting external interrupt %#x pending for injection\n", u8Interrupt));
3661	SVMEVENT Event;
3662	Event.u = 0;
3663	Event.n.u1Valid = 1;
3664	Event.n.u8Vector = u8Interrupt;
3665	Event.n.u3Type = SVM_EVENT_EXTERNAL_IRQ;
3666	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3667	}
3668	else if (rc == VERR_APIC_INTR_MASKED_BY_TPR)
3669	{
3670	/*
3671	* AMD-V has no TPR thresholding feature. TPR and the force-flag will be
3672	* updated eventually when the TPR is written by the guest.
3673	*/
3674	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchTprMaskedIrq);
3675	}
3676	else
3677	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchGuestIrq);
3678	}
3679	else if (!fGif)
3680	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_STGI);
3681	else if (!pSvmTransient->fIsNestedGuest)
3682	hmR0SvmSetIntWindowExiting(pVCpu, pVmcb);
3683	/* else: for nested-guests, interrupt-window exiting will be picked up when merging VMCB controls. */
3684	}
3685
3686	return VINF_SUCCESS;
3687	}
3688
3689
3690	/**
3691	* Injects any pending events into the guest (or nested-guest).
3692	*
3693	* @param pVCpu The cross context virtual CPU structure.
3694	* @param pVmcb Pointer to the VM control block.
3695	*
3696	* @remarks Must only be called when we are guaranteed to enter
3697	* hardware-assisted SVM execution and not return to ring-3
3698	* prematurely.
3699	*/
3700	static void hmR0SvmInjectPendingEvent(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
3701	{
3702	Assert(!TRPMHasTrap(pVCpu));
3703	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
3704
3705	bool const fIntShadow = hmR0SvmIsIntrShadowActive(pVCpu);
3706	#ifdef VBOX_STRICT
3707	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
3708	bool const fGif = CPUMGetGuestGif(pCtx);
3709	bool fAllowInt = fGif;
3710	if (fGif)
3711	{
3712	/*
3713	* For nested-guests we have no way to determine if we're injecting a physical or
3714	* virtual interrupt at this point. Hence the partial verification below.
3715	*/
3716	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
3717	fAllowInt = CPUMIsGuestSvmPhysIntrEnabled(pVCpu, pCtx) \|\| CPUMIsGuestSvmVirtIntrEnabled(pVCpu, pCtx);
3718	else
3719	fAllowInt = RT_BOOL(pCtx->eflags.u32 & X86_EFL_IF);
3720	}
3721	#endif
3722
3723	if (pVCpu->hm.s.Event.fPending)
3724	{
3725	SVMEVENT Event;
3726	Event.u = pVCpu->hm.s.Event.u64IntInfo;
3727	Assert(Event.n.u1Valid);
3728
3729	/*
3730	* Validate event injection pre-conditions.
3731	*/
3732	if (Event.n.u3Type == SVM_EVENT_EXTERNAL_IRQ)
3733	{
3734	Assert(fAllowInt);
3735	Assert(!fIntShadow);
3736	}
3737	else if (Event.n.u3Type == SVM_EVENT_NMI)
3738	{
3739	Assert(fGif);
3740	Assert(!fIntShadow);
3741	}
3742
3743	/*
3744	* Before injecting an NMI we must set VMCPU_FF_BLOCK_NMIS to prevent nested NMIs. We
3745	* do this only when we are surely going to inject the NMI as otherwise if we return
3746	* to ring-3 prematurely we could leave NMIs blocked indefinitely upon re-entry into
3747	* SVM R0.
3748	*
3749	* With VT-x, this is handled by the Guest interruptibility information VMCS field
3750	* which will set the VMCS field after actually delivering the NMI which we read on
3751	* VM-exit to determine the state.
3752	*/
3753	if ( Event.n.u3Type == SVM_EVENT_NMI
3754	&& Event.n.u8Vector == X86_XCPT_NMI
3755	&& !VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_BLOCK_NMIS))
3756	{
3757	VMCPU_FF_SET(pVCpu, VMCPU_FF_BLOCK_NMIS);
3758	}
3759
3760	/*
3761	* Inject it (update VMCB for injection by the hardware).
3762	*/
3763	Log4(("Injecting pending HM event\n"));
3764	hmR0SvmInjectEventVmcb(pVCpu, pVmcb, &Event);
3765	pVCpu->hm.s.Event.fPending = false;
3766
3767	if (Event.n.u3Type == SVM_EVENT_EXTERNAL_IRQ)
3768	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectInterrupt);
3769	else
3770	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectXcpt);
3771	}
3772	else
3773	Assert(pVmcb->ctrl.EventInject.n.u1Valid == 0);
3774
3775	/*
3776	* We could have injected an NMI through IEM and continue guest execution using
3777	* hardware-assisted SVM. In which case, we would not have any events pending (above)
3778	* but we still need to intercept IRET in order to eventually clear NMI inhibition.
3779	*/
3780	if (VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_BLOCK_NMIS))
3781	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_IRET);
3782
3783	/*
3784	* Update the guest interrupt shadow in the guest (or nested-guest) VMCB.
3785	*
3786	* For nested-guests: We need to update it too for the scenario where IEM executes
3787	* the nested-guest but execution later continues here with an interrupt shadow active.
3788	*/
3789	pVmcb->ctrl.IntShadow.n.u1IntShadow = fIntShadow;
3790	}
3791
3792
3793	/**
3794	* Reports world-switch error and dumps some useful debug info.
3795	*
3796	* @param pVCpu The cross context virtual CPU structure.
3797	* @param rcVMRun The return code from VMRUN (or
3798	* VERR_SVM_INVALID_GUEST_STATE for invalid
3799	* guest-state).
3800	*/
3801	static void hmR0SvmReportWorldSwitchError(PVMCPUCC pVCpu, int rcVMRun)
3802	{
3803	HMSVM_ASSERT_PREEMPT_SAFE(pVCpu);
3804	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(&pVCpu->cpum.GstCtx);
3805	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
3806
3807	if (rcVMRun == VERR_SVM_INVALID_GUEST_STATE)
3808	{
3809	#ifdef VBOX_STRICT
3810	hmR0DumpRegs(pVCpu, HM_DUMP_REG_FLAGS_ALL);
3811	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
3812	Log4(("ctrl.u32VmcbCleanBits %#RX32\n", pVmcb->ctrl.u32VmcbCleanBits));
3813	Log4(("ctrl.u16InterceptRdCRx %#x\n", pVmcb->ctrl.u16InterceptRdCRx));
3814	Log4(("ctrl.u16InterceptWrCRx %#x\n", pVmcb->ctrl.u16InterceptWrCRx));
3815	Log4(("ctrl.u16InterceptRdDRx %#x\n", pVmcb->ctrl.u16InterceptRdDRx));
3816	Log4(("ctrl.u16InterceptWrDRx %#x\n", pVmcb->ctrl.u16InterceptWrDRx));
3817	Log4(("ctrl.u32InterceptXcpt %#x\n", pVmcb->ctrl.u32InterceptXcpt));
3818	Log4(("ctrl.u64InterceptCtrl %#RX64\n", pVmcb->ctrl.u64InterceptCtrl));
3819	Log4(("ctrl.u64IOPMPhysAddr %#RX64\n", pVmcb->ctrl.u64IOPMPhysAddr));
3820	Log4(("ctrl.u64MSRPMPhysAddr %#RX64\n", pVmcb->ctrl.u64MSRPMPhysAddr));
3821	Log4(("ctrl.u64TSCOffset %#RX64\n", pVmcb->ctrl.u64TSCOffset));
3822
3823	Log4(("ctrl.TLBCtrl.u32ASID %#x\n", pVmcb->ctrl.TLBCtrl.n.u32ASID));
3824	Log4(("ctrl.TLBCtrl.u8TLBFlush %#x\n", pVmcb->ctrl.TLBCtrl.n.u8TLBFlush));
3825	Log4(("ctrl.TLBCtrl.u24Reserved %#x\n", pVmcb->ctrl.TLBCtrl.n.u24Reserved));
3826
3827	Log4(("ctrl.IntCtrl.u8VTPR %#x\n", pVmcb->ctrl.IntCtrl.n.u8VTPR));
3828	Log4(("ctrl.IntCtrl.u1VIrqPending %#x\n", pVmcb->ctrl.IntCtrl.n.u1VIrqPending));
3829	Log4(("ctrl.IntCtrl.u1VGif %#x\n", pVmcb->ctrl.IntCtrl.n.u1VGif));
3830	Log4(("ctrl.IntCtrl.u6Reserved0 %#x\n", pVmcb->ctrl.IntCtrl.n.u6Reserved));
3831	Log4(("ctrl.IntCtrl.u4VIntrPrio %#x\n", pVmcb->ctrl.IntCtrl.n.u4VIntrPrio));
3832	Log4(("ctrl.IntCtrl.u1IgnoreTPR %#x\n", pVmcb->ctrl.IntCtrl.n.u1IgnoreTPR));
3833	Log4(("ctrl.IntCtrl.u3Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u3Reserved));
3834	Log4(("ctrl.IntCtrl.u1VIntrMasking %#x\n", pVmcb->ctrl.IntCtrl.n.u1VIntrMasking));
3835	Log4(("ctrl.IntCtrl.u1VGifEnable %#x\n", pVmcb->ctrl.IntCtrl.n.u1VGifEnable));
3836	Log4(("ctrl.IntCtrl.u5Reserved1 %#x\n", pVmcb->ctrl.IntCtrl.n.u5Reserved));
3837	Log4(("ctrl.IntCtrl.u8VIntrVector %#x\n", pVmcb->ctrl.IntCtrl.n.u8VIntrVector));
3838	Log4(("ctrl.IntCtrl.u24Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u24Reserved));
3839
3840	Log4(("ctrl.IntShadow.u1IntShadow %#x\n", pVmcb->ctrl.IntShadow.n.u1IntShadow));
3841	Log4(("ctrl.IntShadow.u1GuestIntMask %#x\n", pVmcb->ctrl.IntShadow.n.u1GuestIntMask));
3842	Log4(("ctrl.u64ExitCode %#RX64\n", pVmcb->ctrl.u64ExitCode));
3843	Log4(("ctrl.u64ExitInfo1 %#RX64\n", pVmcb->ctrl.u64ExitInfo1));
3844	Log4(("ctrl.u64ExitInfo2 %#RX64\n", pVmcb->ctrl.u64ExitInfo2));
3845	Log4(("ctrl.ExitIntInfo.u8Vector %#x\n", pVmcb->ctrl.ExitIntInfo.n.u8Vector));
3846	Log4(("ctrl.ExitIntInfo.u3Type %#x\n", pVmcb->ctrl.ExitIntInfo.n.u3Type));
3847	Log4(("ctrl.ExitIntInfo.u1ErrorCodeValid %#x\n", pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid));
3848	Log4(("ctrl.ExitIntInfo.u19Reserved %#x\n", pVmcb->ctrl.ExitIntInfo.n.u19Reserved));
3849	Log4(("ctrl.ExitIntInfo.u1Valid %#x\n", pVmcb->ctrl.ExitIntInfo.n.u1Valid));
3850	Log4(("ctrl.ExitIntInfo.u32ErrorCode %#x\n", pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode));
3851	Log4(("ctrl.NestedPagingCtrl.u1NestedPaging %#x\n", pVmcb->ctrl.NestedPagingCtrl.n.u1NestedPaging));
3852	Log4(("ctrl.NestedPagingCtrl.u1Sev %#x\n", pVmcb->ctrl.NestedPagingCtrl.n.u1Sev));
3853	Log4(("ctrl.NestedPagingCtrl.u1SevEs %#x\n", pVmcb->ctrl.NestedPagingCtrl.n.u1SevEs));
3854	Log4(("ctrl.EventInject.u8Vector %#x\n", pVmcb->ctrl.EventInject.n.u8Vector));
3855	Log4(("ctrl.EventInject.u3Type %#x\n", pVmcb->ctrl.EventInject.n.u3Type));
3856	Log4(("ctrl.EventInject.u1ErrorCodeValid %#x\n", pVmcb->ctrl.EventInject.n.u1ErrorCodeValid));
3857	Log4(("ctrl.EventInject.u19Reserved %#x\n", pVmcb->ctrl.EventInject.n.u19Reserved));
3858	Log4(("ctrl.EventInject.u1Valid %#x\n", pVmcb->ctrl.EventInject.n.u1Valid));
3859	Log4(("ctrl.EventInject.u32ErrorCode %#x\n", pVmcb->ctrl.EventInject.n.u32ErrorCode));
3860
3861	Log4(("ctrl.u64NestedPagingCR3 %#RX64\n", pVmcb->ctrl.u64NestedPagingCR3));
3862
3863	Log4(("ctrl.LbrVirt.u1LbrVirt %#x\n", pVmcb->ctrl.LbrVirt.n.u1LbrVirt));
3864	Log4(("ctrl.LbrVirt.u1VirtVmsaveVmload %#x\n", pVmcb->ctrl.LbrVirt.n.u1VirtVmsaveVmload));
3865
3866	Log4(("guest.CS.u16Sel %RTsel\n", pVmcb->guest.CS.u16Sel));
3867	Log4(("guest.CS.u16Attr %#x\n", pVmcb->guest.CS.u16Attr));
3868	Log4(("guest.CS.u32Limit %#RX32\n", pVmcb->guest.CS.u32Limit));
3869	Log4(("guest.CS.u64Base %#RX64\n", pVmcb->guest.CS.u64Base));
3870	Log4(("guest.DS.u16Sel %#RTsel\n", pVmcb->guest.DS.u16Sel));
3871	Log4(("guest.DS.u16Attr %#x\n", pVmcb->guest.DS.u16Attr));
3872	Log4(("guest.DS.u32Limit %#RX32\n", pVmcb->guest.DS.u32Limit));
3873	Log4(("guest.DS.u64Base %#RX64\n", pVmcb->guest.DS.u64Base));
3874	Log4(("guest.ES.u16Sel %RTsel\n", pVmcb->guest.ES.u16Sel));
3875	Log4(("guest.ES.u16Attr %#x\n", pVmcb->guest.ES.u16Attr));
3876	Log4(("guest.ES.u32Limit %#RX32\n", pVmcb->guest.ES.u32Limit));
3877	Log4(("guest.ES.u64Base %#RX64\n", pVmcb->guest.ES.u64Base));
3878	Log4(("guest.FS.u16Sel %RTsel\n", pVmcb->guest.FS.u16Sel));
3879	Log4(("guest.FS.u16Attr %#x\n", pVmcb->guest.FS.u16Attr));
3880	Log4(("guest.FS.u32Limit %#RX32\n", pVmcb->guest.FS.u32Limit));
3881	Log4(("guest.FS.u64Base %#RX64\n", pVmcb->guest.FS.u64Base));
3882	Log4(("guest.GS.u16Sel %RTsel\n", pVmcb->guest.GS.u16Sel));
3883	Log4(("guest.GS.u16Attr %#x\n", pVmcb->guest.GS.u16Attr));
3884	Log4(("guest.GS.u32Limit %#RX32\n", pVmcb->guest.GS.u32Limit));
3885	Log4(("guest.GS.u64Base %#RX64\n", pVmcb->guest.GS.u64Base));
3886
3887	Log4(("guest.GDTR.u32Limit %#RX32\n", pVmcb->guest.GDTR.u32Limit));
3888	Log4(("guest.GDTR.u64Base %#RX64\n", pVmcb->guest.GDTR.u64Base));
3889
3890	Log4(("guest.LDTR.u16Sel %RTsel\n", pVmcb->guest.LDTR.u16Sel));
3891	Log4(("guest.LDTR.u16Attr %#x\n", pVmcb->guest.LDTR.u16Attr));
3892	Log4(("guest.LDTR.u32Limit %#RX32\n", pVmcb->guest.LDTR.u32Limit));
3893	Log4(("guest.LDTR.u64Base %#RX64\n", pVmcb->guest.LDTR.u64Base));
3894
3895	Log4(("guest.IDTR.u32Limit %#RX32\n", pVmcb->guest.IDTR.u32Limit));
3896	Log4(("guest.IDTR.u64Base %#RX64\n", pVmcb->guest.IDTR.u64Base));
3897
3898	Log4(("guest.TR.u16Sel %RTsel\n", pVmcb->guest.TR.u16Sel));
3899	Log4(("guest.TR.u16Attr %#x\n", pVmcb->guest.TR.u16Attr));
3900	Log4(("guest.TR.u32Limit %#RX32\n", pVmcb->guest.TR.u32Limit));
3901	Log4(("guest.TR.u64Base %#RX64\n", pVmcb->guest.TR.u64Base));
3902
3903	Log4(("guest.u8CPL %#x\n", pVmcb->guest.u8CPL));
3904	Log4(("guest.u64CR0 %#RX64\n", pVmcb->guest.u64CR0));
3905	Log4(("guest.u64CR2 %#RX64\n", pVmcb->guest.u64CR2));
3906	Log4(("guest.u64CR3 %#RX64\n", pVmcb->guest.u64CR3));
3907	Log4(("guest.u64CR4 %#RX64\n", pVmcb->guest.u64CR4));
3908	Log4(("guest.u64DR6 %#RX64\n", pVmcb->guest.u64DR6));
3909	Log4(("guest.u64DR7 %#RX64\n", pVmcb->guest.u64DR7));
3910
3911	Log4(("guest.u64RIP %#RX64\n", pVmcb->guest.u64RIP));
3912	Log4(("guest.u64RSP %#RX64\n", pVmcb->guest.u64RSP));
3913	Log4(("guest.u64RAX %#RX64\n", pVmcb->guest.u64RAX));
3914	Log4(("guest.u64RFlags %#RX64\n", pVmcb->guest.u64RFlags));
3915
3916	Log4(("guest.u64SysEnterCS %#RX64\n", pVmcb->guest.u64SysEnterCS));
3917	Log4(("guest.u64SysEnterEIP %#RX64\n", pVmcb->guest.u64SysEnterEIP));
3918	Log4(("guest.u64SysEnterESP %#RX64\n", pVmcb->guest.u64SysEnterESP));
3919
3920	Log4(("guest.u64EFER %#RX64\n", pVmcb->guest.u64EFER));
3921	Log4(("guest.u64STAR %#RX64\n", pVmcb->guest.u64STAR));
3922	Log4(("guest.u64LSTAR %#RX64\n", pVmcb->guest.u64LSTAR));
3923	Log4(("guest.u64CSTAR %#RX64\n", pVmcb->guest.u64CSTAR));
3924	Log4(("guest.u64SFMASK %#RX64\n", pVmcb->guest.u64SFMASK));
3925	Log4(("guest.u64KernelGSBase %#RX64\n", pVmcb->guest.u64KernelGSBase));
3926	Log4(("guest.u64PAT %#RX64\n", pVmcb->guest.u64PAT));
3927	Log4(("guest.u64DBGCTL %#RX64\n", pVmcb->guest.u64DBGCTL));
3928	Log4(("guest.u64BR_FROM %#RX64\n", pVmcb->guest.u64BR_FROM));
3929	Log4(("guest.u64BR_TO %#RX64\n", pVmcb->guest.u64BR_TO));
3930	Log4(("guest.u64LASTEXCPFROM %#RX64\n", pVmcb->guest.u64LASTEXCPFROM));
3931	Log4(("guest.u64LASTEXCPTO %#RX64\n", pVmcb->guest.u64LASTEXCPTO));
3932
3933	NOREF(pVmcb);
3934	#endif /* VBOX_STRICT */
3935	}
3936	else
3937	Log4Func(("rcVMRun=%d\n", rcVMRun));
3938	}
3939
3940
3941	/**
3942	* Check per-VM and per-VCPU force flag actions that require us to go back to
3943	* ring-3 for one reason or another.
3944	*
3945	* @returns Strict VBox status code (information status code included).
3946	* @retval VINF_SUCCESS if we don't have any actions that require going back to
3947	* ring-3.
3948	* @retval VINF_PGM_SYNC_CR3 if we have pending PGM CR3 sync.
3949	* @retval VINF_EM_PENDING_REQUEST if we have pending requests (like hardware
3950	* interrupts)
3951	* @retval VINF_PGM_POOL_FLUSH_PENDING if PGM is doing a pool flush and requires
3952	* all EMTs to be in ring-3.
3953	* @retval VINF_EM_RAW_TO_R3 if there is pending DMA requests.
3954	* @retval VINF_EM_NO_MEMORY PGM is out of memory, we need to return
3955	* to the EM loop.
3956	*
3957	* @param pVCpu The cross context virtual CPU structure.
3958	*/
3959	static VBOXSTRICTRC hmR0SvmCheckForceFlags(PVMCPUCC pVCpu)
3960	{
3961	Assert(VMMRZCallRing3IsEnabled(pVCpu));
3962
3963	/* Could happen as a result of longjump. */
3964	if (VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_HM_UPDATE_CR3))
3965	PGMUpdateCR3(pVCpu, CPUMGetGuestCR3(pVCpu), false /* fPdpesMapped */);
3966
3967	/* Update pending interrupts into the APIC's IRR. */
3968	if (VMCPU_FF_TEST_AND_CLEAR(pVCpu, VMCPU_FF_UPDATE_APIC))
3969	APICUpdatePendingInterrupts(pVCpu);
3970
3971	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
3972	if ( VM_FF_IS_ANY_SET(pVM, !pVCpu->hm.s.fSingleInstruction
3973	? VM_FF_HP_R0_PRE_HM_MASK : VM_FF_HP_R0_PRE_HM_STEP_MASK)
3974	\|\| VMCPU_FF_IS_ANY_SET(pVCpu, !pVCpu->hm.s.fSingleInstruction
3975	? VMCPU_FF_HP_R0_PRE_HM_MASK : VMCPU_FF_HP_R0_PRE_HM_STEP_MASK) )
3976	{
3977	/* Pending PGM C3 sync. */
3978	if (VMCPU_FF_IS_ANY_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3 \| VMCPU_FF_PGM_SYNC_CR3_NON_GLOBAL))
3979	{
3980	int rc = PGMSyncCR3(pVCpu, pVCpu->cpum.GstCtx.cr0, pVCpu->cpum.GstCtx.cr3, pVCpu->cpum.GstCtx.cr4,
3981	VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3));
3982	if (rc != VINF_SUCCESS)
3983	{
3984	Log4Func(("PGMSyncCR3 forcing us back to ring-3. rc=%d\n", rc));
3985	return rc;
3986	}
3987	}
3988
3989	/* Pending HM-to-R3 operations (critsects, timers, EMT rendezvous etc.) */
3990	/* -XXX- what was that about single stepping? */
3991	if ( VM_FF_IS_ANY_SET(pVM, VM_FF_HM_TO_R3_MASK)
3992	\|\| VMCPU_FF_IS_ANY_SET(pVCpu, VMCPU_FF_HM_TO_R3_MASK))
3993	{
3994	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHmToR3FF);
3995	int rc = RT_LIKELY(!VM_FF_IS_SET(pVM, VM_FF_PGM_NO_MEMORY)) ? VINF_EM_RAW_TO_R3 : VINF_EM_NO_MEMORY;
3996	Log4Func(("HM_TO_R3 forcing us back to ring-3. rc=%d\n", rc));
3997	return rc;
3998	}
3999
4000	/* Pending VM request packets, such as hardware interrupts. */
4001	if ( VM_FF_IS_SET(pVM, VM_FF_REQUEST)
4002	\|\| VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_REQUEST))
4003	{
4004	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchVmReq);
4005	Log4Func(("Pending VM request forcing us back to ring-3\n"));
4006	return VINF_EM_PENDING_REQUEST;
4007	}
4008
4009	/* Pending PGM pool flushes. */
4010	if (VM_FF_IS_SET(pVM, VM_FF_PGM_POOL_FLUSH_PENDING))
4011	{
4012	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchPgmPoolFlush);
4013	Log4Func(("PGM pool flush pending forcing us back to ring-3\n"));
4014	return VINF_PGM_POOL_FLUSH_PENDING;
4015	}
4016
4017	/* Pending DMA requests. */
4018	if (VM_FF_IS_SET(pVM, VM_FF_PDM_DMA))
4019	{
4020	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchDma);
4021	Log4Func(("Pending DMA request forcing us back to ring-3\n"));
4022	return VINF_EM_RAW_TO_R3;
4023	}
4024	}
4025
4026	return VINF_SUCCESS;
4027	}
4028
4029
4030	/**
4031	* Does the preparations before executing guest code in AMD-V.
4032	*
4033	* This may cause longjmps to ring-3 and may even result in rescheduling to the
4034	* recompiler. We must be cautious what we do here regarding committing
4035	* guest-state information into the VMCB assuming we assuredly execute the guest
4036	* in AMD-V. If we fall back to the recompiler after updating the VMCB and
4037	* clearing the common-state (TRPM/forceflags), we must undo those changes so
4038	* that the recompiler can (and should) use them when it resumes guest
4039	* execution. Otherwise such operations must be done when we can no longer
4040	* exit to ring-3.
4041	*
4042	* @returns Strict VBox status code (informational status codes included).
4043	* @retval VINF_SUCCESS if we can proceed with running the guest.
4044	* @retval VINF_* scheduling changes, we have to go back to ring-3.
4045	*
4046	* @param pVCpu The cross context virtual CPU structure.
4047	* @param pSvmTransient Pointer to the SVM transient structure.
4048	*/
4049	static VBOXSTRICTRC hmR0SvmPreRunGuest(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
4050	{
4051	HMSVM_ASSERT_PREEMPT_SAFE(pVCpu);
4052
4053	#ifdef VBOX_WITH_NESTED_HWVIRT_ONLY_IN_IEM
4054	if (pSvmTransient->fIsNestedGuest)
4055	{
4056	Log2(("hmR0SvmPreRunGuest: Rescheduling to IEM due to nested-hwvirt or forced IEM exec -> VINF_EM_RESCHEDULE_REM\n"));
4057	return VINF_EM_RESCHEDULE_REM;
4058	}
4059	#endif
4060
4061	/* Check force flag actions that might require us to go back to ring-3. */
4062	VBOXSTRICTRC rc = hmR0SvmCheckForceFlags(pVCpu);
4063	if (rc != VINF_SUCCESS)
4064	return rc;
4065
4066	if (TRPMHasTrap(pVCpu))
4067	hmR0SvmTrpmTrapToPendingEvent(pVCpu);
4068	else if (!pVCpu->hm.s.Event.fPending)
4069	{
4070	rc = hmR0SvmEvaluatePendingEvent(pVCpu, pSvmTransient);
4071	if ( rc != VINF_SUCCESS
4072	\|\| pSvmTransient->fIsNestedGuest != CPUMIsGuestInSvmNestedHwVirtMode(&pVCpu->cpum.GstCtx))
4073	{
4074	/* If a nested-guest VM-exit occurred, bail. */
4075	if (pSvmTransient->fIsNestedGuest)
4076	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchNstGstVmexit);
4077	return rc;
4078	}
4079	}
4080
4081	/*
4082	* On the oldest AMD-V systems, we may not get enough information to reinject an NMI.
4083	* Just do it in software, see @bugref{8411}.
4084	* NB: If we could continue a task switch exit we wouldn't need to do this.
4085	*/
4086	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
4087	if (RT_UNLIKELY( !g_fHmSvmFeatures
4088	&& pVCpu->hm.s.Event.fPending
4089	&& SVM_EVENT_GET_TYPE(pVCpu->hm.s.Event.u64IntInfo) == SVM_EVENT_NMI))
4090	return VINF_EM_RAW_INJECT_TRPM_EVENT;
4091
4092	#ifdef HMSVM_SYNC_FULL_GUEST_STATE
4093	Assert(!(pVCpu->cpum.GstCtx.fExtrn & HMSVM_CPUMCTX_EXTRN_ALL));
4094	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
4095	#endif
4096
4097	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
4098	/*
4099	* Set up the nested-guest VMCB for execution using hardware-assisted SVM.
4100	*/
4101	if (pSvmTransient->fIsNestedGuest)
4102	hmR0SvmSetupVmcbNested(pVCpu);
4103	#endif
4104
4105	/*
4106	* Export the guest state bits that are not shared with the host in any way as we can
4107	* longjmp or get preempted in the midst of exporting some of the state.
4108	*/
4109	rc = hmR0SvmExportGuestState(pVCpu, pSvmTransient);
4110	AssertRCReturn(rc, rc);
4111	STAM_COUNTER_INC(&pVCpu->hm.s.StatExportFull);
4112
4113	/* Ensure we've cached (and hopefully modified) the nested-guest VMCB for execution using hardware-assisted SVM. */
4114	Assert(!pSvmTransient->fIsNestedGuest \|\| pVCpu->hm.s.svm.NstGstVmcbCache.fCacheValid);
4115
4116	/*
4117	* If we're not intercepting TPR changes in the guest, save the guest TPR before the
4118	* world-switch so we can update it on the way back if the guest changed the TPR.
4119	*/
4120	if (pVCpu->hmr0.s.svm.fSyncVTpr)
4121	{
4122	Assert(!pSvmTransient->fIsNestedGuest);
4123	PCSVMVMCB pVmcb = pVCpu->hmr0.s.svm.pVmcb;
4124	if (pVM->hm.s.fTprPatchingActive)
4125	pSvmTransient->u8GuestTpr = pVmcb->guest.u64LSTAR;
4126	else
4127	pSvmTransient->u8GuestTpr = pVmcb->ctrl.IntCtrl.n.u8VTPR;
4128	}
4129
4130	/*
4131	* No longjmps to ring-3 from this point on!!!
4132	*
4133	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional,
4134	* better than a kernel panic. This also disables flushing of the R0-logger instance.
4135	*/
4136	VMMRZCallRing3Disable(pVCpu);
4137
4138	/*
4139	* We disable interrupts so that we don't miss any interrupts that would flag preemption
4140	* (IPI/timers etc.) when thread-context hooks aren't used and we've been running with
4141	* preemption disabled for a while. Since this is purly to aid the
4142	* RTThreadPreemptIsPending() code, it doesn't matter that it may temporarily reenable and
4143	* disable interrupt on NT.
4144	*
4145	* We need to check for force-flags that could've possible been altered since we last
4146	* checked them (e.g. by PDMGetInterrupt() leaving the PDM critical section,
4147	* see @bugref{6398}).
4148	*
4149	* We also check a couple of other force-flags as a last opportunity to get the EMT back
4150	* to ring-3 before executing guest code.
4151	*/
4152	pSvmTransient->fEFlags = ASMIntDisableFlags();
4153	if ( VM_FF_IS_ANY_SET(pVM, VM_FF_EMT_RENDEZVOUS \| VM_FF_TM_VIRTUAL_SYNC)
4154	\|\| VMCPU_FF_IS_ANY_SET(pVCpu, VMCPU_FF_HM_TO_R3_MASK))
4155	{
4156	ASMSetFlags(pSvmTransient->fEFlags);
4157	VMMRZCallRing3Enable(pVCpu);
4158	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHmToR3FF);
4159	return VINF_EM_RAW_TO_R3;
4160	}
4161	if (RTThreadPreemptIsPending(NIL_RTTHREAD))
4162	{
4163	ASMSetFlags(pSvmTransient->fEFlags);
4164	VMMRZCallRing3Enable(pVCpu);
4165	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchPendingHostIrq);
4166	return VINF_EM_RAW_INTERRUPT;
4167	}
4168
4169	return VINF_SUCCESS;
4170	}
4171
4172
4173	/**
4174	* Prepares to run guest (or nested-guest) code in AMD-V and we've committed to
4175	* doing so.
4176	*
4177	* This means there is no backing out to ring-3 or anywhere else at this point.
4178	*
4179	* @param pVCpu The cross context virtual CPU structure.
4180	* @param pSvmTransient Pointer to the SVM transient structure.
4181	*
4182	* @remarks Called with preemption disabled.
4183	* @remarks No-long-jump zone!!!
4184	*/
4185	static void hmR0SvmPreRunGuestCommitted(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
4186	{
4187	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
4188	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
4189
4190	VMCPU_ASSERT_STATE(pVCpu, VMCPUSTATE_STARTED_HM);
4191	VMCPU_SET_STATE(pVCpu, VMCPUSTATE_STARTED_EXEC); /* Indicate the start of guest execution. */
4192
4193	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
4194	PSVMVMCB pVmcb = pSvmTransient->pVmcb;
4195
4196	hmR0SvmInjectPendingEvent(pVCpu, pVmcb);
4197
4198	if (!CPUMIsGuestFPUStateActive(pVCpu))
4199	{
4200	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatLoadGuestFpuState, x);
4201	CPUMR0LoadGuestFPU(pVM, pVCpu); /* (Ignore rc, no need to set HM_CHANGED_HOST_CONTEXT for SVM.) */
4202	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatLoadGuestFpuState, x);
4203	STAM_COUNTER_INC(&pVCpu->hm.s.StatLoadGuestFpu);
4204	}
4205
4206	/* Load the state shared between host and guest (FPU, debug). */
4207	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE)
4208	hmR0SvmExportSharedState(pVCpu, pVmcb);
4209
4210	pVCpu->hm.s.fCtxChanged &= ~HM_CHANGED_HOST_CONTEXT; /* Preemption might set this, nothing to do on AMD-V. */
4211	AssertMsg(!pVCpu->hm.s.fCtxChanged, ("fCtxChanged=%#RX64\n", pVCpu->hm.s.fCtxChanged));
4212
4213	PHMPHYSCPU pHostCpu = hmR0GetCurrentCpu();
4214	RTCPUID const idHostCpu = pHostCpu->idCpu;
4215	bool const fMigratedHostCpu = idHostCpu != pVCpu->hmr0.s.idLastCpu;
4216
4217	/* Setup TSC offsetting. */
4218	if ( pSvmTransient->fUpdateTscOffsetting
4219	\|\| fMigratedHostCpu)
4220	{
4221	hmR0SvmUpdateTscOffsetting(pVCpu, pVmcb);
4222	pSvmTransient->fUpdateTscOffsetting = false;
4223	}
4224
4225	/* Record statistics of how often we use TSC offsetting as opposed to intercepting RDTSC/P. */
4226	if (!(pVmcb->ctrl.u64InterceptCtrl & (SVM_CTRL_INTERCEPT_RDTSC \| SVM_CTRL_INTERCEPT_RDTSCP)))
4227	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscOffset);
4228	else
4229	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscIntercept);
4230
4231	/* If we've migrating CPUs, mark the VMCB Clean bits as dirty. */
4232	if (fMigratedHostCpu)
4233	pVmcb->ctrl.u32VmcbCleanBits = 0;
4234
4235	/* Store status of the shared guest-host state at the time of VMRUN. */
4236	pSvmTransient->fWasGuestDebugStateActive = CPUMIsGuestDebugStateActive(pVCpu);
4237	pSvmTransient->fWasHyperDebugStateActive = CPUMIsHyperDebugStateActive(pVCpu);
4238
4239	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
4240	uint8_t *pbMsrBitmap;
4241	if (!pSvmTransient->fIsNestedGuest)
4242	pbMsrBitmap = (uint8_t *)pVCpu->hmr0.s.svm.pvMsrBitmap;
4243	else
4244	{
4245	/** @todo We could perhaps optimize this by monitoring if the guest modifies its
4246	* MSRPM and only perform this if it changed also use EVEX.POR when it
4247	* does. */
4248	hmR0SvmMergeMsrpmNested(pHostCpu, pVCpu);
4249
4250	/* Update the nested-guest VMCB with the newly merged MSRPM (clean bits updated below). */
4251	pVmcb->ctrl.u64MSRPMPhysAddr = pHostCpu->n.svm.HCPhysNstGstMsrpm;
4252	pbMsrBitmap = (uint8_t *)pHostCpu->n.svm.pvNstGstMsrpm;
4253	}
4254	#else
4255	uint8_t pbMsrBitmap = (uint8_t )pVCpu->hm.s.svm.pvMsrBitmap;
4256	#endif
4257
4258	ASMAtomicUoWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, true); /* Used for TLB flushing, set this across the world switch. */
4259	/* Flush the appropriate tagged-TLB entries. */
4260	hmR0SvmFlushTaggedTlb(pHostCpu, pVCpu, pVmcb);
4261	Assert(pVCpu->hmr0.s.idLastCpu == idHostCpu);
4262
4263	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatEntry, &pVCpu->hm.s.StatInGC, x);
4264
4265	TMNotifyStartOfExecution(pVM, pVCpu); /* Finally, notify TM to resume its clocks as we're about
4266	to start executing. */
4267
4268	/*
4269	* Save the current Host TSC_AUX and write the guest TSC_AUX to the host, so that RDTSCPs
4270	* (that don't cause exits) reads the guest MSR, see @bugref{3324}.
4271	*
4272	* This should be done -after- any RDTSCPs for obtaining the host timestamp (TM, STAM etc).
4273	*/
4274	if ( pVM->cpum.ro.HostFeatures.fRdTscP
4275	&& !(pVmcb->ctrl.u64InterceptCtrl & SVM_CTRL_INTERCEPT_RDTSCP))
4276	{
4277	uint64_t const uGuestTscAux = CPUMGetGuestTscAux(pVCpu);
4278	pVCpu->hmr0.s.svm.u64HostTscAux = ASMRdMsr(MSR_K8_TSC_AUX);
4279	if (uGuestTscAux != pVCpu->hmr0.s.svm.u64HostTscAux)
4280	ASMWrMsr(MSR_K8_TSC_AUX, uGuestTscAux);
4281	hmR0SvmSetMsrPermission(pVCpu, pbMsrBitmap, MSR_K8_TSC_AUX, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
4282	pSvmTransient->fRestoreTscAuxMsr = true;
4283	}
4284	else
4285	{
4286	hmR0SvmSetMsrPermission(pVCpu, pbMsrBitmap, MSR_K8_TSC_AUX, SVMMSREXIT_INTERCEPT_READ, SVMMSREXIT_INTERCEPT_WRITE);
4287	pSvmTransient->fRestoreTscAuxMsr = false;
4288	}
4289	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_IOPM_MSRPM;
4290
4291	/*
4292	* If VMCB Clean bits isn't supported by the CPU or exposed to the guest in the nested
4293	* virtualization case, mark all state-bits as dirty indicating to the CPU to re-load
4294	* from the VMCB.
4295	*/
4296	bool const fSupportsVmcbCleanBits = hmR0SvmSupportsVmcbCleanBits(pVCpu, pSvmTransient->fIsNestedGuest);
4297	if (!fSupportsVmcbCleanBits)
4298	pVmcb->ctrl.u32VmcbCleanBits = 0;
4299	}
4300
4301
4302	/**
4303	* Wrapper for running the guest (or nested-guest) code in AMD-V.
4304	*
4305	* @returns VBox strict status code.
4306	* @param pVCpu The cross context virtual CPU structure.
4307	* @param HCPhysVmcb The host physical address of the VMCB.
4308	*
4309	* @remarks No-long-jump zone!!!
4310	*/
4311	DECLINLINE(int) hmR0SvmRunGuest(PVMCPUCC pVCpu, RTHCPHYS HCPhysVmcb)
4312	{
4313	/* Mark that HM is the keeper of all guest-CPU registers now that we're going to execute guest code. */
4314	pVCpu->cpum.GstCtx.fExtrn \|= HMSVM_CPUMCTX_EXTRN_ALL \| CPUMCTX_EXTRN_KEEPER_HM;
4315	return pVCpu->hmr0.s.svm.pfnVMRun(pVCpu->CTX_SUFF(pVM), pVCpu, HCPhysVmcb);
4316	}
4317
4318
4319	/**
4320	* Performs some essential restoration of state after running guest (or
4321	* nested-guest) code in AMD-V.
4322	*
4323	* @param pVCpu The cross context virtual CPU structure.
4324	* @param pSvmTransient Pointer to the SVM transient structure.
4325	* @param rcVMRun Return code of VMRUN.
4326	*
4327	* @remarks Called with interrupts disabled.
4328	* @remarks No-long-jump zone!!! This function will however re-enable longjmps
4329	* unconditionally when it is safe to do so.
4330	*/
4331	static void hmR0SvmPostRunGuest(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient, VBOXSTRICTRC rcVMRun)
4332	{
4333	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
4334
4335	ASMAtomicUoWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, false); /* See HMInvalidatePageOnAllVCpus(): used for TLB flushing. */
4336	ASMAtomicIncU32(&pVCpu->hmr0.s.cWorldSwitchExits); /* Initialized in vmR3CreateUVM(): used for EMT poking. */
4337
4338	PSVMVMCB pVmcb = pSvmTransient->pVmcb;
4339	PSVMVMCBCTRL pVmcbCtrl = &pVmcb->ctrl;
4340
4341	/* TSC read must be done early for maximum accuracy. */
4342	if (!(pVmcbCtrl->u64InterceptCtrl & SVM_CTRL_INTERCEPT_RDTSC))
4343	{
4344	if (!pSvmTransient->fIsNestedGuest)
4345	TMCpuTickSetLastSeen(pVCpu, pVCpu->hmr0.s.uTscExit + pVmcbCtrl->u64TSCOffset);
4346	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
4347	else
4348	{
4349	/* The nested-guest VMCB TSC offset shall eventually be restored on #VMEXIT via HMNotifySvmNstGstVmexit(). */
4350	uint64_t const uGstTsc = CPUMRemoveNestedGuestTscOffset(pVCpu, pVCpu->hmr0.s.uTscExit + pVmcbCtrl->u64TSCOffset);
4351	TMCpuTickSetLastSeen(pVCpu, uGstTsc);
4352	}
4353	#endif
4354	}
4355
4356	if (pSvmTransient->fRestoreTscAuxMsr)
4357	{
4358	uint64_t u64GuestTscAuxMsr = ASMRdMsr(MSR_K8_TSC_AUX);
4359	CPUMSetGuestTscAux(pVCpu, u64GuestTscAuxMsr);
4360	if (u64GuestTscAuxMsr != pVCpu->hmr0.s.svm.u64HostTscAux)
4361	ASMWrMsr(MSR_K8_TSC_AUX, pVCpu->hmr0.s.svm.u64HostTscAux);
4362	}
4363
4364	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatInGC, &pVCpu->hm.s.StatPreExit, x);
4365	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
4366	TMNotifyEndOfExecution(pVM, pVCpu, pVCpu->hmr0.s.uTscExit); /* Notify TM that the guest is no longer running. */
4367	VMCPU_SET_STATE(pVCpu, VMCPUSTATE_STARTED_HM);
4368
4369	Assert(!(ASMGetFlags() & X86_EFL_IF));
4370	ASMSetFlags(pSvmTransient->fEFlags); /* Enable interrupts. */
4371	VMMRZCallRing3Enable(pVCpu); /* It is now safe to do longjmps to ring-3!!! */
4372
4373	/* If VMRUN failed, we can bail out early. This does -not- cover SVM_EXIT_INVALID. */
4374	if (RT_UNLIKELY(rcVMRun != VINF_SUCCESS))
4375	{
4376	Log4Func(("VMRUN failure: rcVMRun=%Rrc\n", VBOXSTRICTRC_VAL(rcVMRun)));
4377	return;
4378	}
4379
4380	pSvmTransient->u64ExitCode = pVmcbCtrl->u64ExitCode; /* Save the #VMEXIT reason. */
4381	pSvmTransient->fVectoringDoublePF = false; /* Vectoring double page-fault needs to be determined later. */
4382	pSvmTransient->fVectoringPF = false; /* Vectoring page-fault needs to be determined later. */
4383	pVmcbCtrl->u32VmcbCleanBits = HMSVM_VMCB_CLEAN_ALL; /* Mark the VMCB-state cache as unmodified by VMM. */
4384
4385	#ifdef HMSVM_SYNC_FULL_GUEST_STATE
4386	hmR0SvmImportGuestState(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
4387	Assert(!(pVCpu->cpum.GstCtx.fExtrn & HMSVM_CPUMCTX_EXTRN_ALL));
4388	#else
4389	/*
4390	* Always import the following:
4391	*
4392	* - RIP for exit optimizations and evaluating event injection on re-entry.
4393	* - RFLAGS for evaluating event injection on VM re-entry and for exporting shared debug
4394	* state on preemption.
4395	* - Interrupt shadow, GIF for evaluating event injection on VM re-entry.
4396	* - CS for exit optimizations.
4397	* - RAX, RSP for simplifying assumptions on GPRs. All other GPRs are swapped by the
4398	* assembly switcher code.
4399	* - Shared state (only DR7 currently) for exporting shared debug state on preemption.
4400	*/
4401	hmR0SvmImportGuestState(pVCpu, CPUMCTX_EXTRN_RIP
4402	\| CPUMCTX_EXTRN_RFLAGS
4403	\| CPUMCTX_EXTRN_RAX
4404	\| CPUMCTX_EXTRN_RSP
4405	\| CPUMCTX_EXTRN_CS
4406	\| CPUMCTX_EXTRN_HWVIRT
4407	\| CPUMCTX_EXTRN_INHIBIT_INT
4408	\| CPUMCTX_EXTRN_HM_SVM_HWVIRT_VIRQ
4409	\| HMSVM_CPUMCTX_SHARED_STATE);
4410	#endif
4411
4412	if ( pSvmTransient->u64ExitCode != SVM_EXIT_INVALID
4413	&& pVCpu->hmr0.s.svm.fSyncVTpr)
4414	{
4415	Assert(!pSvmTransient->fIsNestedGuest);
4416	/* TPR patching (for 32-bit guests) uses LSTAR MSR for holding the TPR value, otherwise uses the VTPR. */
4417	if ( pVM->hm.s.fTprPatchingActive
4418	&& (pVmcb->guest.u64LSTAR & 0xff) != pSvmTransient->u8GuestTpr)
4419	{
4420	int rc = APICSetTpr(pVCpu, pVmcb->guest.u64LSTAR & 0xff);
4421	AssertRC(rc);
4422	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
4423	}
4424	/* Sync TPR when we aren't intercepting CR8 writes. */
4425	else if (pSvmTransient->u8GuestTpr != pVmcbCtrl->IntCtrl.n.u8VTPR)
4426	{
4427	int rc = APICSetTpr(pVCpu, pVmcbCtrl->IntCtrl.n.u8VTPR << 4);
4428	AssertRC(rc);
4429	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
4430	}
4431	}
4432
4433	#ifdef DEBUG_ramshankar
4434	if (CPUMIsGuestInSvmNestedHwVirtMode(&pVCpu->cpum.GstCtx))
4435	{
4436	hmR0SvmImportGuestState(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
4437	hmR0SvmLogState(pVCpu, pVmcb, pVCpu->cpum.GstCtx, "hmR0SvmPostRunGuestNested", HMSVM_LOG_ALL & ~HMSVM_LOG_LBR,
4438	0 /* uVerbose */);
4439	}
4440	#endif
4441
4442	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_CS \| CPUMCTX_EXTRN_RIP);
4443	EMHistoryAddExit(pVCpu, EMEXIT_MAKE_FT(EMEXIT_F_KIND_SVM, pSvmTransient->u64ExitCode & EMEXIT_F_TYPE_MASK),
4444	pVCpu->cpum.GstCtx.cs.u64Base + pVCpu->cpum.GstCtx.rip, pVCpu->hmr0.s.uTscExit);
4445	}
4446
4447
4448	/**
4449	* Runs the guest code using AMD-V.
4450	*
4451	* @returns Strict VBox status code.
4452	* @param pVCpu The cross context virtual CPU structure.
4453	* @param pcLoops Pointer to the number of executed loops.
4454	*/
4455	static VBOXSTRICTRC hmR0SvmRunGuestCodeNormal(PVMCPUCC pVCpu, uint32_t *pcLoops)
4456	{
4457	uint32_t const cMaxResumeLoops = pVCpu->CTX_SUFF(pVM)->hmr0.s.cMaxResumeLoops;
4458	Assert(pcLoops);
4459	Assert(*pcLoops <= cMaxResumeLoops);
4460
4461	SVMTRANSIENT SvmTransient;
4462	RT_ZERO(SvmTransient);
4463	SvmTransient.fUpdateTscOffsetting = true;
4464	SvmTransient.pVmcb = pVCpu->hmr0.s.svm.pVmcb;
4465
4466	VBOXSTRICTRC rc = VERR_INTERNAL_ERROR_5;
4467	for (;;)
4468	{
4469	Assert(!HMR0SuspendPending());
4470	HMSVM_ASSERT_CPU_SAFE(pVCpu);
4471
4472	/* Preparatory work for running nested-guest code, this may force us to return to
4473	ring-3. This bugger disables interrupts on VINF_SUCCESS! */
4474	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatEntry, x);
4475	rc = hmR0SvmPreRunGuest(pVCpu, &SvmTransient);
4476	if (rc != VINF_SUCCESS)
4477	break;
4478
4479	/*
4480	* No longjmps to ring-3 from this point on!!!
4481	*
4482	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional,
4483	* better than a kernel panic. This also disables flushing of the R0-logger instance.
4484	*/
4485	hmR0SvmPreRunGuestCommitted(pVCpu, &SvmTransient);
4486	rc = hmR0SvmRunGuest(pVCpu, pVCpu->hmr0.s.svm.HCPhysVmcb);
4487
4488	/* Restore any residual host-state and save any bits shared between host and guest
4489	into the guest-CPU state. Re-enables interrupts! */
4490	hmR0SvmPostRunGuest(pVCpu, &SvmTransient, rc);
4491
4492	if (RT_UNLIKELY( rc != VINF_SUCCESS /* Check for VMRUN errors. */
4493	\|\| SvmTransient.u64ExitCode == SVM_EXIT_INVALID)) /* Check for invalid guest-state errors. */
4494	{
4495	if (rc == VINF_SUCCESS)
4496	rc = VERR_SVM_INVALID_GUEST_STATE;
4497	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatPreExit, x);
4498	hmR0SvmReportWorldSwitchError(pVCpu, VBOXSTRICTRC_VAL(rc));
4499	break;
4500	}
4501
4502	/* Handle the #VMEXIT. */
4503	HMSVM_EXITCODE_STAM_COUNTER_INC(SvmTransient.u64ExitCode);
4504	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatPreExit, &pVCpu->hm.s.StatExitHandling, x);
4505	VBOXVMM_R0_HMSVM_VMEXIT(pVCpu, &pVCpu->cpum.GstCtx, SvmTransient.u64ExitCode, pVCpu->hmr0.s.svm.pVmcb);
4506	rc = hmR0SvmHandleExit(pVCpu, &SvmTransient);
4507	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExitHandling, x);
4508	if (rc != VINF_SUCCESS)
4509	break;
4510	if (++(*pcLoops) >= cMaxResumeLoops)
4511	{
4512	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchMaxResumeLoops);
4513	rc = VINF_EM_RAW_INTERRUPT;
4514	break;
4515	}
4516	}
4517
4518	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatEntry, x);
4519	return rc;
4520	}
4521
4522
4523	/**
4524	* Runs the guest code using AMD-V in single step mode.
4525	*
4526	* @returns Strict VBox status code.
4527	* @param pVCpu The cross context virtual CPU structure.
4528	* @param pcLoops Pointer to the number of executed loops.
4529	*/
4530	static VBOXSTRICTRC hmR0SvmRunGuestCodeStep(PVMCPUCC pVCpu, uint32_t *pcLoops)
4531	{
4532	uint32_t const cMaxResumeLoops = pVCpu->CTX_SUFF(pVM)->hmr0.s.cMaxResumeLoops;
4533	Assert(pcLoops);
4534	Assert(*pcLoops <= cMaxResumeLoops);
4535
4536	SVMTRANSIENT SvmTransient;
4537	RT_ZERO(SvmTransient);
4538	SvmTransient.fUpdateTscOffsetting = true;
4539	SvmTransient.pVmcb = pVCpu->hmr0.s.svm.pVmcb;
4540
4541	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
4542	uint16_t const uCsStart = pCtx->cs.Sel;
4543	uint64_t const uRipStart = pCtx->rip;
4544
4545	VBOXSTRICTRC rc = VERR_INTERNAL_ERROR_5;
4546	for (;;)
4547	{
4548	Assert(!HMR0SuspendPending());
4549	AssertMsg(pVCpu->hmr0.s.idEnteredCpu == RTMpCpuId(),
4550	("Illegal migration! Entered on CPU %u Current %u cLoops=%u\n", (unsigned)pVCpu->hmr0.s.idEnteredCpu,
4551	(unsigned)RTMpCpuId(), *pcLoops));
4552
4553	/* Preparatory work for running nested-guest code, this may force us to return to
4554	ring-3. This bugger disables interrupts on VINF_SUCCESS! */
4555	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatEntry, x);
4556	rc = hmR0SvmPreRunGuest(pVCpu, &SvmTransient);
4557	if (rc != VINF_SUCCESS)
4558	break;
4559
4560	/*
4561	* No longjmps to ring-3 from this point on!!!
4562	*
4563	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional,
4564	* better than a kernel panic. This also disables flushing of the R0-logger instance.
4565	*/
4566	hmR0SvmPreRunGuestCommitted(pVCpu, &SvmTransient);
4567
4568	rc = hmR0SvmRunGuest(pVCpu, pVCpu->hmr0.s.svm.HCPhysVmcb);
4569
4570	/* Restore any residual host-state and save any bits shared between host and guest
4571	into the guest-CPU state. Re-enables interrupts! */
4572	hmR0SvmPostRunGuest(pVCpu, &SvmTransient, rc);
4573
4574	if (RT_UNLIKELY( rc != VINF_SUCCESS /* Check for VMRUN errors. */
4575	\|\| SvmTransient.u64ExitCode == SVM_EXIT_INVALID)) /* Check for invalid guest-state errors. */
4576	{
4577	if (rc == VINF_SUCCESS)
4578	rc = VERR_SVM_INVALID_GUEST_STATE;
4579	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatPreExit, x);
4580	hmR0SvmReportWorldSwitchError(pVCpu, VBOXSTRICTRC_VAL(rc));
4581	return rc;
4582	}
4583
4584	/* Handle the #VMEXIT. */
4585	HMSVM_EXITCODE_STAM_COUNTER_INC(SvmTransient.u64ExitCode);
4586	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatPreExit, &pVCpu->hm.s.StatExitHandling, x);
4587	VBOXVMM_R0_HMSVM_VMEXIT(pVCpu, pCtx, SvmTransient.u64ExitCode, pVCpu->hmr0.s.svm.pVmcb);
4588	rc = hmR0SvmHandleExit(pVCpu, &SvmTransient);
4589	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExitHandling, x);
4590	if (rc != VINF_SUCCESS)
4591	break;
4592	if (++(*pcLoops) >= cMaxResumeLoops)
4593	{
4594	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchMaxResumeLoops);
4595	rc = VINF_EM_RAW_INTERRUPT;
4596	break;
4597	}
4598
4599	/*
4600	* Did the RIP change, if so, consider it a single step.
4601	* Otherwise, make sure one of the TFs gets set.
4602	*/
4603	if ( pCtx->rip != uRipStart
4604	\|\| pCtx->cs.Sel != uCsStart)
4605	{
4606	rc = VINF_EM_DBG_STEPPED;
4607	break;
4608	}
4609	pVCpu->hm.s.fCtxChanged \|= HM_CHANGED_GUEST_DR_MASK;
4610	}
4611
4612	/*
4613	* Clear the X86_EFL_TF if necessary.
4614	*/
4615	if (pVCpu->hmr0.s.fClearTrapFlag)
4616	{
4617	pVCpu->hmr0.s.fClearTrapFlag = false;
4618	pCtx->eflags.Bits.u1TF = 0;
4619	}
4620
4621	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatEntry, x);
4622	return rc;
4623	}
4624
4625	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
4626	/**
4627	* Runs the nested-guest code using AMD-V.
4628	*
4629	* @returns Strict VBox status code.
4630	* @param pVCpu The cross context virtual CPU structure.
4631	* @param pcLoops Pointer to the number of executed loops. If we're switching
4632	* from the guest-code execution loop to this nested-guest
4633	* execution loop pass the remainder value, else pass 0.
4634	*/
4635	static VBOXSTRICTRC hmR0SvmRunGuestCodeNested(PVMCPUCC pVCpu, uint32_t *pcLoops)
4636	{
4637	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
4638	HMSVM_ASSERT_IN_NESTED_GUEST(pCtx);
4639	Assert(pcLoops);
4640	Assert(*pcLoops <= pVCpu->CTX_SUFF(pVM)->hmr0.s.cMaxResumeLoops);
4641	/** @todo r=bird: Sharing this with ring-3 isn't safe in the long run, I fear... */
4642	RTHCPHYS const HCPhysVmcb = GVMMR0ConvertGVMPtr2HCPhys(pVCpu->pGVM, &pCtx->hwvirt.svm.Vmcb);
4643
4644	SVMTRANSIENT SvmTransient;
4645	RT_ZERO(SvmTransient);
4646	SvmTransient.fUpdateTscOffsetting = true;
4647	SvmTransient.pVmcb = &pCtx->hwvirt.svm.Vmcb;
4648	SvmTransient.fIsNestedGuest = true;
4649
4650	VBOXSTRICTRC rc = VERR_INTERNAL_ERROR_4;
4651	for (;;)
4652	{
4653	Assert(!HMR0SuspendPending());
4654	HMSVM_ASSERT_CPU_SAFE(pVCpu);
4655
4656	/* Preparatory work for running nested-guest code, this may force us to return to
4657	ring-3. This bugger disables interrupts on VINF_SUCCESS! */
4658	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatEntry, x);
4659	rc = hmR0SvmPreRunGuest(pVCpu, &SvmTransient);
4660	if ( rc != VINF_SUCCESS
4661	\|\| !CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
4662	break;
4663
4664	/*
4665	* No longjmps to ring-3 from this point on!!!
4666	*
4667	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional,
4668	* better than a kernel panic. This also disables flushing of the R0-logger instance.
4669	*/
4670	hmR0SvmPreRunGuestCommitted(pVCpu, &SvmTransient);
4671
4672	rc = hmR0SvmRunGuest(pVCpu, HCPhysVmcb);
4673
4674	/* Restore any residual host-state and save any bits shared between host and guest
4675	into the guest-CPU state. Re-enables interrupts! */
4676	hmR0SvmPostRunGuest(pVCpu, &SvmTransient, rc);
4677
4678	if (RT_LIKELY( rc == VINF_SUCCESS
4679	&& SvmTransient.u64ExitCode != SVM_EXIT_INVALID))
4680	{ /* extremely likely */ }
4681	else
4682	{
4683	/* VMRUN failed, shouldn't really happen, Guru. */
4684	if (rc != VINF_SUCCESS)
4685	break;
4686
4687	/* Invalid nested-guest state. Cause a #VMEXIT but assert on strict builds. */
4688	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
4689	AssertMsgFailed(("Invalid nested-guest state. rc=%Rrc u64ExitCode=%#RX64\n", rc, SvmTransient.u64ExitCode));
4690	rc = IEMExecSvmVmexit(pVCpu, SVM_EXIT_INVALID, 0, 0);
4691	break;
4692	}
4693
4694	/* Handle the #VMEXIT. */
4695	HMSVM_NESTED_EXITCODE_STAM_COUNTER_INC(SvmTransient.u64ExitCode);
4696	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatPreExit, &pVCpu->hm.s.StatExitHandling, x);
4697	VBOXVMM_R0_HMSVM_VMEXIT(pVCpu, pCtx, SvmTransient.u64ExitCode, &pCtx->hwvirt.svm.Vmcb);
4698	rc = hmR0SvmHandleExitNested(pVCpu, &SvmTransient);
4699	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExitHandling, x);
4700	if (rc == VINF_SUCCESS)
4701	{
4702	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
4703	{
4704	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchNstGstVmexit);
4705	rc = VINF_SVM_VMEXIT;
4706	}
4707	else
4708	{
4709	if (++(*pcLoops) <= pVCpu->CTX_SUFF(pVM)->hmr0.s.cMaxResumeLoops)
4710	continue;
4711	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchMaxResumeLoops);
4712	rc = VINF_EM_RAW_INTERRUPT;
4713	}
4714	}
4715	else
4716	Assert(rc != VINF_SVM_VMEXIT);
4717	break;
4718	/** @todo NSTSVM: handle single-stepping. */
4719	}
4720
4721	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatEntry, x);
4722	return rc;
4723	}
4724	#endif /* VBOX_WITH_NESTED_HWVIRT_SVM */
4725
4726
4727	/**
4728	* Runs the guest code using AMD-V.
4729	*
4730	* @returns Strict VBox status code.
4731	* @param pVCpu The cross context virtual CPU structure.
4732	*/
4733	VMMR0DECL(VBOXSTRICTRC) SVMR0RunGuestCode(PVMCPUCC pVCpu)
4734	{
4735	AssertPtr(pVCpu);
4736	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
4737	Assert(VMMRZCallRing3IsEnabled(pVCpu));
4738	Assert(!ASMAtomicUoReadU64(&pCtx->fExtrn));
4739	HMSVM_ASSERT_PREEMPT_SAFE(pVCpu);
4740
4741	uint32_t cLoops = 0;
4742	VBOXSTRICTRC rc;
4743	for (;;)
4744	{
4745	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
4746	bool const fInNestedGuestMode = CPUMIsGuestInSvmNestedHwVirtMode(pCtx);
4747	#else
4748	NOREF(pCtx);
4749	bool const fInNestedGuestMode = false;
4750	#endif
4751	if (!fInNestedGuestMode)
4752	{
4753	if (!pVCpu->hm.s.fSingleInstruction)
4754	rc = hmR0SvmRunGuestCodeNormal(pVCpu, &cLoops);
4755	else
4756	rc = hmR0SvmRunGuestCodeStep(pVCpu, &cLoops);
4757	}
4758	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
4759	else
4760	rc = hmR0SvmRunGuestCodeNested(pVCpu, &cLoops);
4761
4762	if (rc == VINF_SVM_VMRUN)
4763	{
4764	Assert(CPUMIsGuestInSvmNestedHwVirtMode(pCtx));
4765	continue;
4766	}
4767	if (rc == VINF_SVM_VMEXIT)
4768	{
4769	Assert(!CPUMIsGuestInSvmNestedHwVirtMode(pCtx));
4770	continue;
4771	}
4772	#endif
4773	break;
4774	}
4775
4776	/* Fixup error codes. */
4777	if (rc == VERR_EM_INTERPRETER)
4778	rc = VINF_EM_RAW_EMULATE_INSTR;
4779	else if (rc == VINF_EM_RESET)
4780	rc = VINF_EM_TRIPLE_FAULT;
4781
4782	/* Prepare to return to ring-3. This will remove longjmp notifications. */
4783	rc = hmR0SvmExitToRing3(pVCpu, rc);
4784	Assert(!ASMAtomicUoReadU64(&pCtx->fExtrn));
4785	Assert(!VMMR0AssertionIsNotificationSet(pVCpu));
4786	return rc;
4787	}
4788
4789
4790	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
4791	/**
4792	* Determines whether the given I/O access should cause a nested-guest \#VMEXIT.
4793	*
4794	* @param pvIoBitmap Pointer to the nested-guest IO bitmap.
4795	* @param pIoExitInfo Pointer to the SVMIOIOEXITINFO.
4796	*/
4797	static bool hmR0SvmIsIoInterceptSet(void *pvIoBitmap, PSVMIOIOEXITINFO pIoExitInfo)
4798	{
4799	const uint16_t u16Port = pIoExitInfo->n.u16Port;
4800	const SVMIOIOTYPE enmIoType = (SVMIOIOTYPE)pIoExitInfo->n.u1Type;
4801	const uint8_t cbReg = (pIoExitInfo->u >> SVM_IOIO_OP_SIZE_SHIFT) & 7;
4802	const uint8_t cAddrSizeBits = ((pIoExitInfo->u >> SVM_IOIO_ADDR_SIZE_SHIFT) & 7) << 4;
4803	const uint8_t iEffSeg = pIoExitInfo->n.u3Seg;
4804	const bool fRep = pIoExitInfo->n.u1Rep;
4805	const bool fStrIo = pIoExitInfo->n.u1Str;
4806
4807	return CPUMIsSvmIoInterceptSet(pvIoBitmap, u16Port, enmIoType, cbReg, cAddrSizeBits, iEffSeg, fRep, fStrIo,
4808	NULL /* pIoExitInfo */);
4809	}
4810
4811
4812	/**
4813	* Handles a nested-guest \#VMEXIT (for all EXITCODE values except
4814	* SVM_EXIT_INVALID).
4815	*
4816	* @returns VBox status code (informational status codes included).
4817	* @param pVCpu The cross context virtual CPU structure.
4818	* @param pSvmTransient Pointer to the SVM transient structure.
4819	*/
4820	static VBOXSTRICTRC hmR0SvmHandleExitNested(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
4821	{
4822	HMSVM_ASSERT_IN_NESTED_GUEST(&pVCpu->cpum.GstCtx);
4823	Assert(pSvmTransient->u64ExitCode != SVM_EXIT_INVALID);
4824	Assert(pSvmTransient->u64ExitCode <= SVM_EXIT_MAX);
4825
4826	/*
4827	* We import the complete state here because we use separate VMCBs for the guest and the
4828	* nested-guest, and the guest's VMCB is used after the #VMEXIT. We can only save/restore
4829	* the #VMEXIT specific state if we used the same VMCB for both guest and nested-guest.
4830	*/
4831	#define NST_GST_VMEXIT_CALL_RET(a_pVCpu, a_uExitCode, a_uExitInfo1, a_uExitInfo2) \
4832	do { \
4833	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL); \
4834	return IEMExecSvmVmexit((a_pVCpu), (a_uExitCode), (a_uExitInfo1), (a_uExitInfo2)); \
4835	} while (0)
4836
4837	/*
4838	* For all the #VMEXITs here we primarily figure out if the #VMEXIT is expected by the
4839	* nested-guest. If it isn't, it should be handled by the (outer) guest.
4840	*/
4841	PSVMVMCB pVmcbNstGst = &pVCpu->cpum.GstCtx.hwvirt.svm.Vmcb;
4842	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
4843	PSVMVMCBCTRL pVmcbNstGstCtrl = &pVmcbNstGst->ctrl;
4844	uint64_t const uExitCode = pVmcbNstGstCtrl->u64ExitCode;
4845	uint64_t const uExitInfo1 = pVmcbNstGstCtrl->u64ExitInfo1;
4846	uint64_t const uExitInfo2 = pVmcbNstGstCtrl->u64ExitInfo2;
4847
4848	Assert(uExitCode == pVmcbNstGstCtrl->u64ExitCode);
4849	switch (uExitCode)
4850	{
4851	case SVM_EXIT_CPUID:
4852	{
4853	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_CPUID))
4854	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4855	return hmR0SvmExitCpuid(pVCpu, pSvmTransient);
4856	}
4857
4858	case SVM_EXIT_RDTSC:
4859	{
4860	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_RDTSC))
4861	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4862	return hmR0SvmExitRdtsc(pVCpu, pSvmTransient);
4863	}
4864
4865	case SVM_EXIT_RDTSCP:
4866	{
4867	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_RDTSCP))
4868	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4869	return hmR0SvmExitRdtscp(pVCpu, pSvmTransient);
4870	}
4871
4872	case SVM_EXIT_MONITOR:
4873	{
4874	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_MONITOR))
4875	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4876	return hmR0SvmExitMonitor(pVCpu, pSvmTransient);
4877	}
4878
4879	case SVM_EXIT_MWAIT:
4880	{
4881	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_MWAIT))
4882	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4883	return hmR0SvmExitMwait(pVCpu, pSvmTransient);
4884	}
4885
4886	case SVM_EXIT_HLT:
4887	{
4888	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_HLT))
4889	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4890	return hmR0SvmExitHlt(pVCpu, pSvmTransient);
4891	}
4892
4893	case SVM_EXIT_MSR:
4894	{
4895	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_MSR_PROT))
4896	{
4897	uint32_t const idMsr = pVCpu->cpum.GstCtx.ecx;
4898	uint16_t offMsrpm;
4899	uint8_t uMsrpmBit;
4900	int rc = CPUMGetSvmMsrpmOffsetAndBit(idMsr, &offMsrpm, &uMsrpmBit);
4901	if (RT_SUCCESS(rc))
4902	{
4903	Assert(uMsrpmBit == 0 \|\| uMsrpmBit == 2 \|\| uMsrpmBit == 4 \|\| uMsrpmBit == 6);
4904	Assert(offMsrpm < SVM_MSRPM_PAGES << X86_PAGE_4K_SHIFT);
4905
4906	uint8_t const * const pbMsrBitmap = &pVCpu->cpum.GstCtx.hwvirt.svm.abMsrBitmap[offMsrpm];
4907	bool const fInterceptRead = RT_BOOL(*pbMsrBitmap & RT_BIT(uMsrpmBit));
4908	bool const fInterceptWrite = RT_BOOL(*pbMsrBitmap & RT_BIT(uMsrpmBit + 1));
4909
4910	if ( (fInterceptWrite && pVmcbNstGstCtrl->u64ExitInfo1 == SVM_EXIT1_MSR_WRITE)
4911	\|\| (fInterceptRead && pVmcbNstGstCtrl->u64ExitInfo1 == SVM_EXIT1_MSR_READ))
4912	{
4913	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4914	}
4915	}
4916	else
4917	{
4918	/*
4919	* MSRs not covered by the MSRPM automatically cause an #VMEXIT.
4920	* See AMD-V spec. "15.11 MSR Intercepts".
4921	*/
4922	Assert(rc == VERR_OUT_OF_RANGE);
4923	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4924	}
4925	}
4926	return hmR0SvmExitMsr(pVCpu, pSvmTransient);
4927	}
4928
4929	case SVM_EXIT_IOIO:
4930	{
4931	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_IOIO_PROT))
4932	{
4933	SVMIOIOEXITINFO IoExitInfo;
4934	IoExitInfo.u = pVmcbNstGst->ctrl.u64ExitInfo1;
4935	bool const fIntercept = hmR0SvmIsIoInterceptSet(pVCpu->cpum.GstCtx.hwvirt.svm.abIoBitmap, &IoExitInfo);
4936	if (fIntercept)
4937	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4938	}
4939	return hmR0SvmExitIOInstr(pVCpu, pSvmTransient);
4940	}
4941
4942	case SVM_EXIT_XCPT_PF:
4943	{
4944	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
4945	if (pVM->hmr0.s.fNestedPaging)
4946	{
4947	uint32_t const u32ErrCode = pVmcbNstGstCtrl->u64ExitInfo1;
4948	uint64_t const uFaultAddress = pVmcbNstGstCtrl->u64ExitInfo2;
4949
4950	/* If the nested-guest is intercepting #PFs, cause a #PF #VMEXIT. */
4951	if (CPUMIsGuestSvmXcptInterceptSet(pVCpu, pCtx, X86_XCPT_PF))
4952	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, u32ErrCode, uFaultAddress);
4953
4954	/* If the nested-guest is not intercepting #PFs, forward the #PF to the guest. */
4955	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR2);
4956	hmR0SvmSetPendingXcptPF(pVCpu, u32ErrCode, uFaultAddress);
4957	return VINF_SUCCESS;
4958	}
4959	return hmR0SvmExitXcptPF(pVCpu, pSvmTransient);
4960	}
4961
4962	case SVM_EXIT_XCPT_UD:
4963	{
4964	if (CPUMIsGuestSvmXcptInterceptSet(pVCpu, pCtx, X86_XCPT_UD))
4965	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4966	hmR0SvmSetPendingXcptUD(pVCpu);
4967	return VINF_SUCCESS;
4968	}
4969
4970	case SVM_EXIT_XCPT_MF:
4971	{
4972	if (CPUMIsGuestSvmXcptInterceptSet(pVCpu, pCtx, X86_XCPT_MF))
4973	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4974	return hmR0SvmExitXcptMF(pVCpu, pSvmTransient);
4975	}
4976
4977	case SVM_EXIT_XCPT_DB:
4978	{
4979	if (CPUMIsGuestSvmXcptInterceptSet(pVCpu, pCtx, X86_XCPT_DB))
4980	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4981	return hmR0SvmNestedExitXcptDB(pVCpu, pSvmTransient);
4982	}
4983
4984	case SVM_EXIT_XCPT_AC:
4985	{
4986	if (CPUMIsGuestSvmXcptInterceptSet(pVCpu, pCtx, X86_XCPT_AC))
4987	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4988	return hmR0SvmExitXcptAC(pVCpu, pSvmTransient);
4989	}
4990
4991	case SVM_EXIT_XCPT_BP:
4992	{
4993	if (CPUMIsGuestSvmXcptInterceptSet(pVCpu, pCtx, X86_XCPT_BP))
4994	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
4995	return hmR0SvmNestedExitXcptBP(pVCpu, pSvmTransient);
4996	}
4997
4998	case SVM_EXIT_READ_CR0:
4999	case SVM_EXIT_READ_CR3:
5000	case SVM_EXIT_READ_CR4:
5001	{
5002	uint8_t const uCr = uExitCode - SVM_EXIT_READ_CR0;
5003	if (CPUMIsGuestSvmReadCRxInterceptSet(pVCpu, pCtx, uCr))
5004	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5005	return hmR0SvmExitReadCRx(pVCpu, pSvmTransient);
5006	}
5007
5008	case SVM_EXIT_CR0_SEL_WRITE:
5009	{
5010	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_CR0_SEL_WRITE))
5011	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5012	return hmR0SvmExitWriteCRx(pVCpu, pSvmTransient);
5013	}
5014
5015	case SVM_EXIT_WRITE_CR0:
5016	case SVM_EXIT_WRITE_CR3:
5017	case SVM_EXIT_WRITE_CR4:
5018	case SVM_EXIT_WRITE_CR8: /* CR8 writes would go to the V_TPR rather than here, since we run with V_INTR_MASKING. */
5019	{
5020	uint8_t const uCr = uExitCode - SVM_EXIT_WRITE_CR0;
5021	Log4Func(("Write CR%u: uExitInfo1=%#RX64 uExitInfo2=%#RX64\n", uCr, uExitInfo1, uExitInfo2));
5022
5023	if (CPUMIsGuestSvmWriteCRxInterceptSet(pVCpu, pCtx, uCr))
5024	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5025	return hmR0SvmExitWriteCRx(pVCpu, pSvmTransient);
5026	}
5027
5028	case SVM_EXIT_PAUSE:
5029	{
5030	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_PAUSE))
5031	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5032	return hmR0SvmExitPause(pVCpu, pSvmTransient);
5033	}
5034
5035	case SVM_EXIT_VINTR:
5036	{
5037	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_VINTR))
5038	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5039	return hmR0SvmExitUnexpected(pVCpu, pSvmTransient);
5040	}
5041
5042	case SVM_EXIT_INTR:
5043	case SVM_EXIT_NMI:
5044	case SVM_EXIT_SMI:
5045	case SVM_EXIT_XCPT_NMI: /* Should not occur, SVM_EXIT_NMI is used instead. */
5046	{
5047	/*
5048	* We shouldn't direct physical interrupts, NMIs, SMIs to the nested-guest.
5049	*
5050	* Although we don't intercept SMIs, the nested-guest might. Therefore, we might
5051	* get an SMI #VMEXIT here so simply ignore rather than causing a corresponding
5052	* nested-guest #VMEXIT.
5053	*
5054	* We shall import the complete state here as we may cause #VMEXITs from ring-3
5055	* while trying to inject interrupts, see comment at the top of this function.
5056	*/
5057	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_ALL);
5058	return hmR0SvmExitIntr(pVCpu, pSvmTransient);
5059	}
5060
5061	case SVM_EXIT_FERR_FREEZE:
5062	{
5063	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_FERR_FREEZE))
5064	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5065	return hmR0SvmExitFerrFreeze(pVCpu, pSvmTransient);
5066	}
5067
5068	case SVM_EXIT_INVLPG:
5069	{
5070	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_INVLPG))
5071	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5072	return hmR0SvmExitInvlpg(pVCpu, pSvmTransient);
5073	}
5074
5075	case SVM_EXIT_WBINVD:
5076	{
5077	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_WBINVD))
5078	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5079	return hmR0SvmExitWbinvd(pVCpu, pSvmTransient);
5080	}
5081
5082	case SVM_EXIT_INVD:
5083	{
5084	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_INVD))
5085	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5086	return hmR0SvmExitInvd(pVCpu, pSvmTransient);
5087	}
5088
5089	case SVM_EXIT_RDPMC:
5090	{
5091	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_RDPMC))
5092	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5093	return hmR0SvmExitRdpmc(pVCpu, pSvmTransient);
5094	}
5095
5096	default:
5097	{
5098	switch (uExitCode)
5099	{
5100	case SVM_EXIT_READ_DR0: case SVM_EXIT_READ_DR1: case SVM_EXIT_READ_DR2: case SVM_EXIT_READ_DR3:
5101	case SVM_EXIT_READ_DR6: case SVM_EXIT_READ_DR7: case SVM_EXIT_READ_DR8: case SVM_EXIT_READ_DR9:
5102	case SVM_EXIT_READ_DR10: case SVM_EXIT_READ_DR11: case SVM_EXIT_READ_DR12: case SVM_EXIT_READ_DR13:
5103	case SVM_EXIT_READ_DR14: case SVM_EXIT_READ_DR15:
5104	{
5105	uint8_t const uDr = uExitCode - SVM_EXIT_READ_DR0;
5106	if (CPUMIsGuestSvmReadDRxInterceptSet(pVCpu, pCtx, uDr))
5107	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5108	return hmR0SvmExitReadDRx(pVCpu, pSvmTransient);
5109	}
5110
5111	case SVM_EXIT_WRITE_DR0: case SVM_EXIT_WRITE_DR1: case SVM_EXIT_WRITE_DR2: case SVM_EXIT_WRITE_DR3:
5112	case SVM_EXIT_WRITE_DR6: case SVM_EXIT_WRITE_DR7: case SVM_EXIT_WRITE_DR8: case SVM_EXIT_WRITE_DR9:
5113	case SVM_EXIT_WRITE_DR10: case SVM_EXIT_WRITE_DR11: case SVM_EXIT_WRITE_DR12: case SVM_EXIT_WRITE_DR13:
5114	case SVM_EXIT_WRITE_DR14: case SVM_EXIT_WRITE_DR15:
5115	{
5116	uint8_t const uDr = uExitCode - SVM_EXIT_WRITE_DR0;
5117	if (CPUMIsGuestSvmWriteDRxInterceptSet(pVCpu, pCtx, uDr))
5118	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5119	return hmR0SvmExitWriteDRx(pVCpu, pSvmTransient);
5120	}
5121
5122	case SVM_EXIT_XCPT_DE:
5123	/* SVM_EXIT_XCPT_DB: / / Handled above. */
5124	/* SVM_EXIT_XCPT_NMI: / / Handled above. */
5125	/* SVM_EXIT_XCPT_BP: / / Handled above. */
5126	case SVM_EXIT_XCPT_OF:
5127	case SVM_EXIT_XCPT_BR:
5128	/* SVM_EXIT_XCPT_UD: / / Handled above. */
5129	case SVM_EXIT_XCPT_NM:
5130	case SVM_EXIT_XCPT_DF:
5131	case SVM_EXIT_XCPT_CO_SEG_OVERRUN:
5132	case SVM_EXIT_XCPT_TS:
5133	case SVM_EXIT_XCPT_NP:
5134	case SVM_EXIT_XCPT_SS:
5135	case SVM_EXIT_XCPT_GP:
5136	/* SVM_EXIT_XCPT_PF: / / Handled above. */
5137	case SVM_EXIT_XCPT_15: /* Reserved. */
5138	/* SVM_EXIT_XCPT_MF: / / Handled above. */
5139	/* SVM_EXIT_XCPT_AC: / / Handled above. */
5140	case SVM_EXIT_XCPT_MC:
5141	case SVM_EXIT_XCPT_XF:
5142	case SVM_EXIT_XCPT_20: case SVM_EXIT_XCPT_21: case SVM_EXIT_XCPT_22: case SVM_EXIT_XCPT_23:
5143	case SVM_EXIT_XCPT_24: case SVM_EXIT_XCPT_25: case SVM_EXIT_XCPT_26: case SVM_EXIT_XCPT_27:
5144	case SVM_EXIT_XCPT_28: case SVM_EXIT_XCPT_29: case SVM_EXIT_XCPT_30: case SVM_EXIT_XCPT_31:
5145	{
5146	uint8_t const uVector = uExitCode - SVM_EXIT_XCPT_0;
5147	if (CPUMIsGuestSvmXcptInterceptSet(pVCpu, pCtx, uVector))
5148	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5149	return hmR0SvmExitXcptGeneric(pVCpu, pSvmTransient);
5150	}
5151
5152	case SVM_EXIT_XSETBV:
5153	{
5154	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_XSETBV))
5155	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5156	return hmR0SvmExitXsetbv(pVCpu, pSvmTransient);
5157	}
5158
5159	case SVM_EXIT_TASK_SWITCH:
5160	{
5161	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_TASK_SWITCH))
5162	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5163	return hmR0SvmExitTaskSwitch(pVCpu, pSvmTransient);
5164	}
5165
5166	case SVM_EXIT_IRET:
5167	{
5168	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_IRET))
5169	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5170	return hmR0SvmExitIret(pVCpu, pSvmTransient);
5171	}
5172
5173	case SVM_EXIT_SHUTDOWN:
5174	{
5175	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_SHUTDOWN))
5176	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5177	return hmR0SvmExitShutdown(pVCpu, pSvmTransient);
5178	}
5179
5180	case SVM_EXIT_VMMCALL:
5181	{
5182	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_VMMCALL))
5183	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5184	return hmR0SvmExitVmmCall(pVCpu, pSvmTransient);
5185	}
5186
5187	case SVM_EXIT_CLGI:
5188	{
5189	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_CLGI))
5190	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5191	return hmR0SvmExitClgi(pVCpu, pSvmTransient);
5192	}
5193
5194	case SVM_EXIT_STGI:
5195	{
5196	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_STGI))
5197	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5198	return hmR0SvmExitStgi(pVCpu, pSvmTransient);
5199	}
5200
5201	case SVM_EXIT_VMLOAD:
5202	{
5203	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_VMLOAD))
5204	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5205	return hmR0SvmExitVmload(pVCpu, pSvmTransient);
5206	}
5207
5208	case SVM_EXIT_VMSAVE:
5209	{
5210	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_VMSAVE))
5211	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5212	return hmR0SvmExitVmsave(pVCpu, pSvmTransient);
5213	}
5214
5215	case SVM_EXIT_INVLPGA:
5216	{
5217	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_INVLPGA))
5218	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5219	return hmR0SvmExitInvlpga(pVCpu, pSvmTransient);
5220	}
5221
5222	case SVM_EXIT_VMRUN:
5223	{
5224	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_VMRUN))
5225	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5226	return hmR0SvmExitVmrun(pVCpu, pSvmTransient);
5227	}
5228
5229	case SVM_EXIT_RSM:
5230	{
5231	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_RSM))
5232	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5233	hmR0SvmSetPendingXcptUD(pVCpu);
5234	return VINF_SUCCESS;
5235	}
5236
5237	case SVM_EXIT_SKINIT:
5238	{
5239	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_SKINIT))
5240	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5241	hmR0SvmSetPendingXcptUD(pVCpu);
5242	return VINF_SUCCESS;
5243	}
5244
5245	case SVM_EXIT_NPF:
5246	{
5247	Assert(pVCpu->CTX_SUFF(pVM)->hmr0.s.fNestedPaging);
5248	return hmR0SvmExitNestedPF(pVCpu, pSvmTransient);
5249	}
5250
5251	case SVM_EXIT_INIT: /* We shouldn't get INIT signals while executing a nested-guest. */
5252	return hmR0SvmExitUnexpected(pVCpu, pSvmTransient);
5253
5254	default:
5255	{
5256	AssertMsgFailed(("hmR0SvmHandleExitNested: Unknown exit code %#x\n", pSvmTransient->u64ExitCode));
5257	pVCpu->hm.s.u32HMError = pSvmTransient->u64ExitCode;
5258	return VERR_SVM_UNKNOWN_EXIT;
5259	}
5260	}
5261	}
5262	}
5263	/* not reached */
5264
5265	#undef NST_GST_VMEXIT_CALL_RET
5266	}
5267	#endif
5268
5269
5270	/**
5271	* Handles a guest \#VMEXIT (for all EXITCODE values except SVM_EXIT_INVALID).
5272	*
5273	* @returns Strict VBox status code (informational status codes included).
5274	* @param pVCpu The cross context virtual CPU structure.
5275	* @param pSvmTransient Pointer to the SVM transient structure.
5276	*/
5277	static VBOXSTRICTRC hmR0SvmHandleExit(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
5278	{
5279	Assert(pSvmTransient->u64ExitCode != SVM_EXIT_INVALID);
5280	Assert(pSvmTransient->u64ExitCode <= SVM_EXIT_MAX);
5281
5282	#ifdef DEBUG_ramshankar
5283	# define VMEXIT_CALL_RET(a_fDbg, a_CallExpr) \
5284	do { \
5285	if ((a_fDbg) == 1) \
5286	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL); \
5287	int rc = a_CallExpr; \
5288	if ((a_fDbg) == 1) \
5289	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST); \
5290	return rc; \
5291	} while (0)
5292	#else
5293	# define VMEXIT_CALL_RET(a_fDbg, a_CallExpr) return a_CallExpr
5294	#endif
5295
5296	/*
5297	* The ordering of the case labels is based on most-frequently-occurring #VMEXITs
5298	* for most guests under normal workloads (for some definition of "normal").
5299	*/
5300	uint64_t const uExitCode = pSvmTransient->u64ExitCode;
5301	switch (uExitCode)
5302	{
5303	case SVM_EXIT_NPF: VMEXIT_CALL_RET(0, hmR0SvmExitNestedPF(pVCpu, pSvmTransient));
5304	case SVM_EXIT_IOIO: VMEXIT_CALL_RET(0, hmR0SvmExitIOInstr(pVCpu, pSvmTransient));
5305	case SVM_EXIT_RDTSC: VMEXIT_CALL_RET(0, hmR0SvmExitRdtsc(pVCpu, pSvmTransient));
5306	case SVM_EXIT_RDTSCP: VMEXIT_CALL_RET(0, hmR0SvmExitRdtscp(pVCpu, pSvmTransient));
5307	case SVM_EXIT_CPUID: VMEXIT_CALL_RET(0, hmR0SvmExitCpuid(pVCpu, pSvmTransient));
5308	case SVM_EXIT_XCPT_PF: VMEXIT_CALL_RET(0, hmR0SvmExitXcptPF(pVCpu, pSvmTransient));
5309	case SVM_EXIT_MSR: VMEXIT_CALL_RET(0, hmR0SvmExitMsr(pVCpu, pSvmTransient));
5310	case SVM_EXIT_MONITOR: VMEXIT_CALL_RET(0, hmR0SvmExitMonitor(pVCpu, pSvmTransient));
5311	case SVM_EXIT_MWAIT: VMEXIT_CALL_RET(0, hmR0SvmExitMwait(pVCpu, pSvmTransient));
5312	case SVM_EXIT_HLT: VMEXIT_CALL_RET(0, hmR0SvmExitHlt(pVCpu, pSvmTransient));
5313
5314	case SVM_EXIT_XCPT_NMI: /* Should not occur, SVM_EXIT_NMI is used instead. */
5315	case SVM_EXIT_INTR:
5316	case SVM_EXIT_NMI: VMEXIT_CALL_RET(0, hmR0SvmExitIntr(pVCpu, pSvmTransient));
5317
5318	case SVM_EXIT_READ_CR0:
5319	case SVM_EXIT_READ_CR3:
5320	case SVM_EXIT_READ_CR4: VMEXIT_CALL_RET(0, hmR0SvmExitReadCRx(pVCpu, pSvmTransient));
5321
5322	case SVM_EXIT_CR0_SEL_WRITE:
5323	case SVM_EXIT_WRITE_CR0:
5324	case SVM_EXIT_WRITE_CR3:
5325	case SVM_EXIT_WRITE_CR4:
5326	case SVM_EXIT_WRITE_CR8: VMEXIT_CALL_RET(0, hmR0SvmExitWriteCRx(pVCpu, pSvmTransient));
5327
5328	case SVM_EXIT_VINTR: VMEXIT_CALL_RET(0, hmR0SvmExitVIntr(pVCpu, pSvmTransient));
5329	case SVM_EXIT_PAUSE: VMEXIT_CALL_RET(0, hmR0SvmExitPause(pVCpu, pSvmTransient));
5330	case SVM_EXIT_VMMCALL: VMEXIT_CALL_RET(0, hmR0SvmExitVmmCall(pVCpu, pSvmTransient));
5331	case SVM_EXIT_INVLPG: VMEXIT_CALL_RET(0, hmR0SvmExitInvlpg(pVCpu, pSvmTransient));
5332	case SVM_EXIT_WBINVD: VMEXIT_CALL_RET(0, hmR0SvmExitWbinvd(pVCpu, pSvmTransient));
5333	case SVM_EXIT_INVD: VMEXIT_CALL_RET(0, hmR0SvmExitInvd(pVCpu, pSvmTransient));
5334	case SVM_EXIT_RDPMC: VMEXIT_CALL_RET(0, hmR0SvmExitRdpmc(pVCpu, pSvmTransient));
5335	case SVM_EXIT_IRET: VMEXIT_CALL_RET(0, hmR0SvmExitIret(pVCpu, pSvmTransient));
5336	case SVM_EXIT_XCPT_UD: VMEXIT_CALL_RET(0, hmR0SvmExitXcptUD(pVCpu, pSvmTransient));
5337	case SVM_EXIT_XCPT_MF: VMEXIT_CALL_RET(0, hmR0SvmExitXcptMF(pVCpu, pSvmTransient));
5338	case SVM_EXIT_XCPT_DB: VMEXIT_CALL_RET(0, hmR0SvmExitXcptDB(pVCpu, pSvmTransient));
5339	case SVM_EXIT_XCPT_AC: VMEXIT_CALL_RET(0, hmR0SvmExitXcptAC(pVCpu, pSvmTransient));
5340	case SVM_EXIT_XCPT_BP: VMEXIT_CALL_RET(0, hmR0SvmExitXcptBP(pVCpu, pSvmTransient));
5341	case SVM_EXIT_XCPT_GP: VMEXIT_CALL_RET(0, hmR0SvmExitXcptGP(pVCpu, pSvmTransient));
5342	case SVM_EXIT_XSETBV: VMEXIT_CALL_RET(0, hmR0SvmExitXsetbv(pVCpu, pSvmTransient));
5343	case SVM_EXIT_FERR_FREEZE: VMEXIT_CALL_RET(0, hmR0SvmExitFerrFreeze(pVCpu, pSvmTransient));
5344
5345	default:
5346	{
5347	switch (pSvmTransient->u64ExitCode)
5348	{
5349	case SVM_EXIT_READ_DR0: case SVM_EXIT_READ_DR1: case SVM_EXIT_READ_DR2: case SVM_EXIT_READ_DR3:
5350	case SVM_EXIT_READ_DR6: case SVM_EXIT_READ_DR7: case SVM_EXIT_READ_DR8: case SVM_EXIT_READ_DR9:
5351	case SVM_EXIT_READ_DR10: case SVM_EXIT_READ_DR11: case SVM_EXIT_READ_DR12: case SVM_EXIT_READ_DR13:
5352	case SVM_EXIT_READ_DR14: case SVM_EXIT_READ_DR15:
5353	VMEXIT_CALL_RET(0, hmR0SvmExitReadDRx(pVCpu, pSvmTransient));
5354
5355	case SVM_EXIT_WRITE_DR0: case SVM_EXIT_WRITE_DR1: case SVM_EXIT_WRITE_DR2: case SVM_EXIT_WRITE_DR3:
5356	case SVM_EXIT_WRITE_DR6: case SVM_EXIT_WRITE_DR7: case SVM_EXIT_WRITE_DR8: case SVM_EXIT_WRITE_DR9:
5357	case SVM_EXIT_WRITE_DR10: case SVM_EXIT_WRITE_DR11: case SVM_EXIT_WRITE_DR12: case SVM_EXIT_WRITE_DR13:
5358	case SVM_EXIT_WRITE_DR14: case SVM_EXIT_WRITE_DR15:
5359	VMEXIT_CALL_RET(0, hmR0SvmExitWriteDRx(pVCpu, pSvmTransient));
5360
5361	case SVM_EXIT_TASK_SWITCH: VMEXIT_CALL_RET(0, hmR0SvmExitTaskSwitch(pVCpu, pSvmTransient));
5362	case SVM_EXIT_SHUTDOWN: VMEXIT_CALL_RET(0, hmR0SvmExitShutdown(pVCpu, pSvmTransient));
5363
5364	case SVM_EXIT_SMI:
5365	case SVM_EXIT_INIT:
5366	{
5367	/*
5368	* We don't intercept SMIs. As for INIT signals, it really shouldn't ever happen here.
5369	* If it ever does, we want to know about it so log the exit code and bail.
5370	*/
5371	VMEXIT_CALL_RET(0, hmR0SvmExitUnexpected(pVCpu, pSvmTransient));
5372	}
5373
5374	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
5375	case SVM_EXIT_CLGI: VMEXIT_CALL_RET(0, hmR0SvmExitClgi(pVCpu, pSvmTransient));
5376	case SVM_EXIT_STGI: VMEXIT_CALL_RET(0, hmR0SvmExitStgi(pVCpu, pSvmTransient));
5377	case SVM_EXIT_VMLOAD: VMEXIT_CALL_RET(0, hmR0SvmExitVmload(pVCpu, pSvmTransient));
5378	case SVM_EXIT_VMSAVE: VMEXIT_CALL_RET(0, hmR0SvmExitVmsave(pVCpu, pSvmTransient));
5379	case SVM_EXIT_INVLPGA: VMEXIT_CALL_RET(0, hmR0SvmExitInvlpga(pVCpu, pSvmTransient));
5380	case SVM_EXIT_VMRUN: VMEXIT_CALL_RET(0, hmR0SvmExitVmrun(pVCpu, pSvmTransient));
5381	#else
5382	case SVM_EXIT_CLGI:
5383	case SVM_EXIT_STGI:
5384	case SVM_EXIT_VMLOAD:
5385	case SVM_EXIT_VMSAVE:
5386	case SVM_EXIT_INVLPGA:
5387	case SVM_EXIT_VMRUN:
5388	#endif
5389	case SVM_EXIT_RSM:
5390	case SVM_EXIT_SKINIT:
5391	{
5392	hmR0SvmSetPendingXcptUD(pVCpu);
5393	return VINF_SUCCESS;
5394	}
5395
5396	#ifdef HMSVM_ALWAYS_TRAP_ALL_XCPTS
5397	case SVM_EXIT_XCPT_DE:
5398	/* SVM_EXIT_XCPT_DB: / / Handled above. */
5399	/* SVM_EXIT_XCPT_NMI: / / Handled above. */
5400	/* SVM_EXIT_XCPT_BP: / / Handled above. */
5401	case SVM_EXIT_XCPT_OF:
5402	case SVM_EXIT_XCPT_BR:
5403	/* SVM_EXIT_XCPT_UD: / / Handled above. */
5404	case SVM_EXIT_XCPT_NM:
5405	case SVM_EXIT_XCPT_DF:
5406	case SVM_EXIT_XCPT_CO_SEG_OVERRUN:
5407	case SVM_EXIT_XCPT_TS:
5408	case SVM_EXIT_XCPT_NP:
5409	case SVM_EXIT_XCPT_SS:
5410	/* SVM_EXIT_XCPT_GP: / / Handled above. */
5411	/* SVM_EXIT_XCPT_PF: */
5412	case SVM_EXIT_XCPT_15: /* Reserved. */
5413	/* SVM_EXIT_XCPT_MF: / / Handled above. */
5414	/* SVM_EXIT_XCPT_AC: / / Handled above. */
5415	case SVM_EXIT_XCPT_MC:
5416	case SVM_EXIT_XCPT_XF:
5417	case SVM_EXIT_XCPT_20: case SVM_EXIT_XCPT_21: case SVM_EXIT_XCPT_22: case SVM_EXIT_XCPT_23:
5418	case SVM_EXIT_XCPT_24: case SVM_EXIT_XCPT_25: case SVM_EXIT_XCPT_26: case SVM_EXIT_XCPT_27:
5419	case SVM_EXIT_XCPT_28: case SVM_EXIT_XCPT_29: case SVM_EXIT_XCPT_30: case SVM_EXIT_XCPT_31:
5420	VMEXIT_CALL_RET(0, hmR0SvmExitXcptGeneric(pVCpu, pSvmTransient));
5421	#endif /* HMSVM_ALWAYS_TRAP_ALL_XCPTS */
5422
5423	default:
5424	{
5425	AssertMsgFailed(("hmR0SvmHandleExit: Unknown exit code %#RX64\n", uExitCode));
5426	pVCpu->hm.s.u32HMError = uExitCode;
5427	return VERR_SVM_UNKNOWN_EXIT;
5428	}
5429	}
5430	}
5431	}
5432	/* not reached */
5433	#undef VMEXIT_CALL_RET
5434	}
5435
5436
5437	#ifdef VBOX_STRICT
5438	/* Is there some generic IPRT define for this that are not in Runtime/internal/\* ?? */
5439	# define HMSVM_ASSERT_PREEMPT_CPUID_VAR() \
5440	RTCPUID const idAssertCpu = RTThreadPreemptIsEnabled(NIL_RTTHREAD) ? NIL_RTCPUID : RTMpCpuId()
5441
5442	# define HMSVM_ASSERT_PREEMPT_CPUID() \
5443	do \
5444	{ \
5445	RTCPUID const idAssertCpuNow = RTThreadPreemptIsEnabled(NIL_RTTHREAD) ? NIL_RTCPUID : RTMpCpuId(); \
5446	AssertMsg(idAssertCpu == idAssertCpuNow, ("SVM %#x, %#x\n", idAssertCpu, idAssertCpuNow)); \
5447	} while (0)
5448
5449	# define HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(a_pVCpu, a_pSvmTransient) \
5450	do { \
5451	AssertPtr((a_pVCpu)); \
5452	AssertPtr((a_pSvmTransient)); \
5453	Assert(ASMIntAreEnabled()); \
5454	HMSVM_ASSERT_PREEMPT_SAFE((a_pVCpu)); \
5455	HMSVM_ASSERT_PREEMPT_CPUID_VAR(); \
5456	Log4Func(("vcpu[%u] -v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-\n", (a_pVCpu)->idCpu)); \
5457	HMSVM_ASSERT_PREEMPT_SAFE((a_pVCpu)); \
5458	if (!VMMRZCallRing3IsEnabled((a_pVCpu))) \
5459	HMSVM_ASSERT_PREEMPT_CPUID(); \
5460	} while (0)
5461	#else
5462	# define HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(a_pVCpu, a_pSvmTransient) \
5463	do { \
5464	RT_NOREF2(a_pVCpu, a_pSvmTransient); \
5465	} while (0)
5466	#endif
5467
5468
5469	/**
5470	* Gets the IEM exception flags for the specified SVM event.
5471	*
5472	* @returns The IEM exception flags.
5473	* @param pEvent Pointer to the SVM event.
5474	*
5475	* @remarks This function currently only constructs flags required for
5476	* IEMEvaluateRecursiveXcpt and not the complete flags (e.g. error-code
5477	* and CR2 aspects of an exception are not included).
5478	*/
5479	static uint32_t hmR0SvmGetIemXcptFlags(PCSVMEVENT pEvent)
5480	{
5481	uint8_t const uEventType = pEvent->n.u3Type;
5482	uint32_t fIemXcptFlags;
5483	switch (uEventType)
5484	{
5485	case SVM_EVENT_EXCEPTION:
5486	/*
5487	* Only INT3 and INTO instructions can raise #BP and #OF exceptions.
5488	* See AMD spec. Table 8-1. "Interrupt Vector Source and Cause".
5489	*/
5490	if (pEvent->n.u8Vector == X86_XCPT_BP)
5491	{
5492	fIemXcptFlags = IEM_XCPT_FLAGS_T_SOFT_INT \| IEM_XCPT_FLAGS_BP_INSTR;
5493	break;
5494	}
5495	if (pEvent->n.u8Vector == X86_XCPT_OF)
5496	{
5497	fIemXcptFlags = IEM_XCPT_FLAGS_T_SOFT_INT \| IEM_XCPT_FLAGS_OF_INSTR;
5498	break;
5499	}
5500	/** @todo How do we distinguish ICEBP \#DB from the regular one? */
5501	RT_FALL_THRU();
5502	case SVM_EVENT_NMI:
5503	fIemXcptFlags = IEM_XCPT_FLAGS_T_CPU_XCPT;
5504	break;
5505
5506	case SVM_EVENT_EXTERNAL_IRQ:
5507	fIemXcptFlags = IEM_XCPT_FLAGS_T_EXT_INT;
5508	break;
5509
5510	case SVM_EVENT_SOFTWARE_INT:
5511	fIemXcptFlags = IEM_XCPT_FLAGS_T_SOFT_INT;
5512	break;
5513
5514	default:
5515	fIemXcptFlags = 0;
5516	AssertMsgFailed(("Unexpected event type! uEventType=%#x uVector=%#x", uEventType, pEvent->n.u8Vector));
5517	break;
5518	}
5519	return fIemXcptFlags;
5520	}
5521
5522
5523	/**
5524	* Handle a condition that occurred while delivering an event through the guest
5525	* IDT.
5526	*
5527	* @returns VBox status code (informational error codes included).
5528	* @retval VINF_SUCCESS if we should continue handling the \#VMEXIT.
5529	* @retval VINF_HM_DOUBLE_FAULT if a \#DF condition was detected and we ought to
5530	* continue execution of the guest which will delivery the \#DF.
5531	* @retval VINF_EM_RESET if we detected a triple-fault condition.
5532	* @retval VERR_EM_GUEST_CPU_HANG if we detected a guest CPU hang.
5533	*
5534	* @param pVCpu The cross context virtual CPU structure.
5535	* @param pSvmTransient Pointer to the SVM transient structure.
5536	*
5537	* @remarks No-long-jump zone!!!
5538	*/
5539	static int hmR0SvmCheckExitDueToEventDelivery(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
5540	{
5541	int rc = VINF_SUCCESS;
5542	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
5543	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR2);
5544
5545	Log4(("EXITINTINFO: Pending vectoring event %#RX64 Valid=%RTbool ErrValid=%RTbool Err=%#RX32 Type=%u Vector=%u\n",
5546	pVmcb->ctrl.ExitIntInfo.u, !!pVmcb->ctrl.ExitIntInfo.n.u1Valid, !!pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid,
5547	pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode, pVmcb->ctrl.ExitIntInfo.n.u3Type, pVmcb->ctrl.ExitIntInfo.n.u8Vector));
5548
5549	/*
5550	* The EXITINTINFO (if valid) contains the prior exception (IDT vector) that was trying to
5551	* be delivered to the guest which caused a #VMEXIT which was intercepted (Exit vector).
5552	*
5553	* See AMD spec. 15.7.3 "EXITINFO Pseudo-Code".
5554	*/
5555	if (pVmcb->ctrl.ExitIntInfo.n.u1Valid)
5556	{
5557	IEMXCPTRAISE enmRaise;
5558	IEMXCPTRAISEINFO fRaiseInfo;
5559	bool const fExitIsHwXcpt = pSvmTransient->u64ExitCode - SVM_EXIT_XCPT_0 <= SVM_EXIT_XCPT_31;
5560	uint8_t const uIdtVector = pVmcb->ctrl.ExitIntInfo.n.u8Vector;
5561	if (fExitIsHwXcpt)
5562	{
5563	uint8_t const uExitVector = pSvmTransient->u64ExitCode - SVM_EXIT_XCPT_0;
5564	uint32_t const fIdtVectorFlags = hmR0SvmGetIemXcptFlags(&pVmcb->ctrl.ExitIntInfo);
5565	uint32_t const fExitVectorFlags = IEM_XCPT_FLAGS_T_CPU_XCPT;
5566	enmRaise = IEMEvaluateRecursiveXcpt(pVCpu, fIdtVectorFlags, uIdtVector, fExitVectorFlags, uExitVector, &fRaiseInfo);
5567	}
5568	else
5569	{
5570	/*
5571	* If delivery of an event caused a #VMEXIT that is not an exception (e.g. #NPF)
5572	* then we end up here.
5573	*
5574	* If the event was:
5575	* - a software interrupt, we can re-execute the instruction which will
5576	* regenerate the event.
5577	* - an NMI, we need to clear NMI blocking and re-inject the NMI.
5578	* - a hardware exception or external interrupt, we re-inject it.
5579	*/
5580	fRaiseInfo = IEMXCPTRAISEINFO_NONE;
5581	if (pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_SOFTWARE_INT)
5582	enmRaise = IEMXCPTRAISE_REEXEC_INSTR;
5583	else
5584	enmRaise = IEMXCPTRAISE_PREV_EVENT;
5585	}
5586
5587	switch (enmRaise)
5588	{
5589	case IEMXCPTRAISE_CURRENT_XCPT:
5590	case IEMXCPTRAISE_PREV_EVENT:
5591	{
5592	/* For software interrupts, we shall re-execute the instruction. */
5593	if (!(fRaiseInfo & IEMXCPTRAISEINFO_SOFT_INT_XCPT))
5594	{
5595	RTGCUINTPTR GCPtrFaultAddress = 0;
5596
5597	/* If we are re-injecting an NMI, clear NMI blocking. */
5598	if (pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_NMI)
5599	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_BLOCK_NMIS);
5600
5601	/* Determine a vectoring #PF condition, see comment in hmR0SvmExitXcptPF(). */
5602	if (fRaiseInfo & (IEMXCPTRAISEINFO_EXT_INT_PF \| IEMXCPTRAISEINFO_NMI_PF))
5603	{
5604	pSvmTransient->fVectoringPF = true;
5605	Log4Func(("IDT: Pending vectoring #PF due to delivery of Ext-Int/NMI. uCR2=%#RX64\n",
5606	pVCpu->cpum.GstCtx.cr2));
5607	}
5608	else if ( pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_EXCEPTION
5609	&& uIdtVector == X86_XCPT_PF)
5610	{
5611	/*
5612	* If the previous exception was a #PF, we need to recover the CR2 value.
5613	* This can't happen with shadow paging.
5614	*/
5615	GCPtrFaultAddress = pVCpu->cpum.GstCtx.cr2;
5616	}
5617
5618	/*
5619	* Without nested paging, when uExitVector is #PF, CR2 value will be updated from the VMCB's
5620	* exit info. fields, if it's a guest #PF, see hmR0SvmExitXcptPF().
5621	*/
5622	Assert(pVmcb->ctrl.ExitIntInfo.n.u3Type != SVM_EVENT_SOFTWARE_INT);
5623	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectReflect);
5624	hmR0SvmSetPendingEvent(pVCpu, &pVmcb->ctrl.ExitIntInfo, GCPtrFaultAddress);
5625
5626	Log4Func(("IDT: Pending vectoring event %#RX64 ErrValid=%RTbool Err=%#RX32 GCPtrFaultAddress=%#RX64\n",
5627	pVmcb->ctrl.ExitIntInfo.u, RT_BOOL(pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid),
5628	pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode, GCPtrFaultAddress));
5629	}
5630	break;
5631	}
5632
5633	case IEMXCPTRAISE_REEXEC_INSTR:
5634	{
5635	Assert(rc == VINF_SUCCESS);
5636	break;
5637	}
5638
5639	case IEMXCPTRAISE_DOUBLE_FAULT:
5640	{
5641	/*
5642	* Determing a vectoring double #PF condition. Used later, when PGM evaluates
5643	* the second #PF as a guest #PF (and not a shadow #PF) and needs to be
5644	* converted into a #DF.
5645	*/
5646	if (fRaiseInfo & IEMXCPTRAISEINFO_PF_PF)
5647	{
5648	Log4Func(("IDT: Pending vectoring double #PF uCR2=%#RX64\n", pVCpu->cpum.GstCtx.cr2));
5649	pSvmTransient->fVectoringDoublePF = true;
5650	Assert(rc == VINF_SUCCESS);
5651	}
5652	else
5653	{
5654	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectConvertDF);
5655	hmR0SvmSetPendingXcptDF(pVCpu);
5656	rc = VINF_HM_DOUBLE_FAULT;
5657	}
5658	break;
5659	}
5660
5661	case IEMXCPTRAISE_TRIPLE_FAULT:
5662	{
5663	rc = VINF_EM_RESET;
5664	break;
5665	}
5666
5667	case IEMXCPTRAISE_CPU_HANG:
5668	{
5669	rc = VERR_EM_GUEST_CPU_HANG;
5670	break;
5671	}
5672
5673	default:
5674	AssertMsgFailedBreakStmt(("Bogus enmRaise value: %d (%#x)\n", enmRaise, enmRaise), rc = VERR_SVM_IPE_2);
5675	}
5676	}
5677	Assert(rc == VINF_SUCCESS \|\| rc == VINF_HM_DOUBLE_FAULT \|\| rc == VINF_EM_RESET \|\| rc == VERR_EM_GUEST_CPU_HANG);
5678	return rc;
5679	}
5680
5681
5682	/**
5683	* Advances the guest RIP by the number of bytes specified in @a cb.
5684	*
5685	* @param pVCpu The cross context virtual CPU structure.
5686	* @param cb RIP increment value in bytes.
5687	*/
5688	DECLINLINE(void) hmR0SvmAdvanceRip(PVMCPUCC pVCpu, uint32_t cb)
5689	{
5690	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
5691	pCtx->rip += cb;
5692
5693	/* Update interrupt shadow. */
5694	if ( VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS)
5695	&& pCtx->rip != EMGetInhibitInterruptsPC(pVCpu))
5696	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS);
5697	}
5698
5699
5700	/* -=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= */
5701	/* -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- #VMEXIT handlers -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- */
5702	/* -=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= */
5703
5704	/** @name \#VMEXIT handlers.
5705	* @{
5706	*/
5707
5708	/**
5709	* \#VMEXIT handler for external interrupts, NMIs, FPU assertion freeze and INIT
5710	* signals (SVM_EXIT_INTR, SVM_EXIT_NMI, SVM_EXIT_FERR_FREEZE, SVM_EXIT_INIT).
5711	*/
5712	HMSVM_EXIT_DECL hmR0SvmExitIntr(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
5713	{
5714	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
5715
5716	if (pSvmTransient->u64ExitCode == SVM_EXIT_NMI)
5717	STAM_REL_COUNTER_INC(&pVCpu->hm.s.StatExitHostNmiInGC);
5718	else if (pSvmTransient->u64ExitCode == SVM_EXIT_INTR)
5719	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitExtInt);
5720
5721	/*
5722	* AMD-V has no preemption timer and the generic periodic preemption timer has no way to
5723	* signal -before- the timer fires if the current interrupt is our own timer or a some
5724	* other host interrupt. We also cannot examine what interrupt it is until the host
5725	* actually take the interrupt.
5726	*
5727	* Going back to executing guest code here unconditionally causes random scheduling
5728	* problems (observed on an AMD Phenom 9850 Quad-Core on Windows 64-bit host).
5729	*/
5730	return VINF_EM_RAW_INTERRUPT;
5731	}
5732
5733
5734	/**
5735	* \#VMEXIT handler for WBINVD (SVM_EXIT_WBINVD). Conditional \#VMEXIT.
5736	*/
5737	HMSVM_EXIT_DECL hmR0SvmExitWbinvd(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
5738	{
5739	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
5740
5741	VBOXSTRICTRC rcStrict;
5742	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
5743	if (fSupportsNextRipSave)
5744	{
5745	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK);
5746	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
5747	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
5748	rcStrict = IEMExecDecodedWbinvd(pVCpu, cbInstr);
5749	}
5750	else
5751	{
5752	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
5753	rcStrict = IEMExecOne(pVCpu);
5754	}
5755
5756	if (rcStrict == VINF_IEM_RAISED_XCPT)
5757	{
5758	rcStrict = VINF_SUCCESS;
5759	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
5760	}
5761	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
5762	return rcStrict;
5763	}
5764
5765
5766	/**
5767	* \#VMEXIT handler for INVD (SVM_EXIT_INVD). Unconditional \#VMEXIT.
5768	*/
5769	HMSVM_EXIT_DECL hmR0SvmExitInvd(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
5770	{
5771	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
5772
5773	VBOXSTRICTRC rcStrict;
5774	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
5775	if (fSupportsNextRipSave)
5776	{
5777	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK);
5778	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
5779	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
5780	rcStrict = IEMExecDecodedInvd(pVCpu, cbInstr);
5781	}
5782	else
5783	{
5784	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
5785	rcStrict = IEMExecOne(pVCpu);
5786	}
5787
5788	if (rcStrict == VINF_IEM_RAISED_XCPT)
5789	{
5790	rcStrict = VINF_SUCCESS;
5791	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
5792	}
5793	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
5794	return rcStrict;
5795	}
5796
5797
5798	/**
5799	* \#VMEXIT handler for INVD (SVM_EXIT_CPUID). Conditional \#VMEXIT.
5800	*/
5801	HMSVM_EXIT_DECL hmR0SvmExitCpuid(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
5802	{
5803	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
5804
5805	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK \| CPUMCTX_EXTRN_RAX \| CPUMCTX_EXTRN_RCX);
5806	VBOXSTRICTRC rcStrict;
5807	PCEMEXITREC pExitRec = EMHistoryUpdateFlagsAndTypeAndPC(pVCpu,
5808	EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_CPUID),
5809	pVCpu->cpum.GstCtx.rip + pVCpu->cpum.GstCtx.cs.u64Base);
5810	if (!pExitRec)
5811	{
5812	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
5813	if (fSupportsNextRipSave)
5814	{
5815	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
5816	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
5817	rcStrict = IEMExecDecodedCpuid(pVCpu, cbInstr);
5818	}
5819	else
5820	{
5821	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
5822	rcStrict = IEMExecOne(pVCpu);
5823	}
5824
5825	if (rcStrict == VINF_IEM_RAISED_XCPT)
5826	{
5827	rcStrict = VINF_SUCCESS;
5828	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
5829	}
5830	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
5831	}
5832	else
5833	{
5834	/*
5835	* Frequent exit or something needing probing. Get state and call EMHistoryExec.
5836	*/
5837	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
5838
5839	Log4(("CpuIdExit/%u: %04x:%08RX64: %#x/%#x -> EMHistoryExec\n",
5840	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip, pVCpu->cpum.GstCtx.eax, pVCpu->cpum.GstCtx.ecx));
5841
5842	rcStrict = EMHistoryExec(pVCpu, pExitRec, 0);
5843
5844	Log4(("CpuIdExit/%u: %04x:%08RX64: EMHistoryExec -> %Rrc + %04x:%08RX64\n",
5845	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip,
5846	VBOXSTRICTRC_VAL(rcStrict), pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip));
5847	}
5848	return rcStrict;
5849	}
5850
5851
5852	/**
5853	* \#VMEXIT handler for RDTSC (SVM_EXIT_RDTSC). Conditional \#VMEXIT.
5854	*/
5855	HMSVM_EXIT_DECL hmR0SvmExitRdtsc(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
5856	{
5857	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
5858
5859	VBOXSTRICTRC rcStrict;
5860	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
5861	if (fSupportsNextRipSave)
5862	{
5863	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK \| CPUMCTX_EXTRN_CR4);
5864	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
5865	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
5866	rcStrict = IEMExecDecodedRdtsc(pVCpu, cbInstr);
5867	}
5868	else
5869	{
5870	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
5871	rcStrict = IEMExecOne(pVCpu);
5872	}
5873
5874	if (rcStrict == VINF_SUCCESS)
5875	pSvmTransient->fUpdateTscOffsetting = true;
5876	else if (rcStrict == VINF_IEM_RAISED_XCPT)
5877	{
5878	rcStrict = VINF_SUCCESS;
5879	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
5880	}
5881	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
5882	return rcStrict;
5883	}
5884
5885
5886	/**
5887	* \#VMEXIT handler for RDTSCP (SVM_EXIT_RDTSCP). Conditional \#VMEXIT.
5888	*/
5889	HMSVM_EXIT_DECL hmR0SvmExitRdtscp(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
5890	{
5891	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
5892
5893	VBOXSTRICTRC rcStrict;
5894	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
5895	if (fSupportsNextRipSave)
5896	{
5897	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK \| CPUMCTX_EXTRN_CR4 \| CPUMCTX_EXTRN_TSC_AUX);
5898	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
5899	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
5900	rcStrict = IEMExecDecodedRdtscp(pVCpu, cbInstr);
5901	}
5902	else
5903	{
5904	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
5905	rcStrict = IEMExecOne(pVCpu);
5906	}
5907
5908	if (rcStrict == VINF_SUCCESS)
5909	pSvmTransient->fUpdateTscOffsetting = true;
5910	else if (rcStrict == VINF_IEM_RAISED_XCPT)
5911	{
5912	rcStrict = VINF_SUCCESS;
5913	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
5914	}
5915	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
5916	return rcStrict;
5917	}
5918
5919
5920	/**
5921	* \#VMEXIT handler for RDPMC (SVM_EXIT_RDPMC). Conditional \#VMEXIT.
5922	*/
5923	HMSVM_EXIT_DECL hmR0SvmExitRdpmc(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
5924	{
5925	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
5926
5927	VBOXSTRICTRC rcStrict;
5928	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
5929	if (fSupportsNextRipSave)
5930	{
5931	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK \| CPUMCTX_EXTRN_CR4);
5932	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
5933	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
5934	rcStrict = IEMExecDecodedRdpmc(pVCpu, cbInstr);
5935	}
5936	else
5937	{
5938	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
5939	rcStrict = IEMExecOne(pVCpu);
5940	}
5941
5942	if (rcStrict == VINF_IEM_RAISED_XCPT)
5943	{
5944	rcStrict = VINF_SUCCESS;
5945	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
5946	}
5947	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
5948	return rcStrict;
5949	}
5950
5951
5952	/**
5953	* \#VMEXIT handler for INVLPG (SVM_EXIT_INVLPG). Conditional \#VMEXIT.
5954	*/
5955	HMSVM_EXIT_DECL hmR0SvmExitInvlpg(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
5956	{
5957	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
5958	Assert(!pVCpu->CTX_SUFF(pVM)->hmr0.s.fNestedPaging);
5959
5960	VBOXSTRICTRC rcStrict;
5961	bool const fSupportsDecodeAssists = hmR0SvmSupportsDecodeAssists(pVCpu);
5962	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
5963	if ( fSupportsDecodeAssists
5964	&& fSupportsNextRipSave)
5965	{
5966	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_MEM_MASK);
5967	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
5968	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
5969	RTGCPTR const GCPtrPage = pVmcb->ctrl.u64ExitInfo1;
5970	rcStrict = IEMExecDecodedInvlpg(pVCpu, cbInstr, GCPtrPage);
5971	}
5972	else
5973	{
5974	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
5975	rcStrict = IEMExecOne(pVCpu);
5976	}
5977
5978	if (rcStrict == VINF_IEM_RAISED_XCPT)
5979	{
5980	rcStrict = VINF_SUCCESS;
5981	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
5982	}
5983	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
5984	return VBOXSTRICTRC_VAL(rcStrict);
5985	}
5986
5987
5988	/**
5989	* \#VMEXIT handler for HLT (SVM_EXIT_HLT). Conditional \#VMEXIT.
5990	*/
5991	HMSVM_EXIT_DECL hmR0SvmExitHlt(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
5992	{
5993	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
5994
5995	VBOXSTRICTRC rcStrict;
5996	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
5997	if (fSupportsNextRipSave)
5998	{
5999	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK);
6000	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
6001	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
6002	rcStrict = IEMExecDecodedHlt(pVCpu, cbInstr);
6003	}
6004	else
6005	{
6006	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6007	rcStrict = IEMExecOne(pVCpu);
6008	}
6009
6010	if ( rcStrict == VINF_EM_HALT
6011	\|\| rcStrict == VINF_SUCCESS)
6012	rcStrict = EMShouldContinueAfterHalt(pVCpu, &pVCpu->cpum.GstCtx) ? VINF_SUCCESS : VINF_EM_HALT;
6013	else if (rcStrict == VINF_IEM_RAISED_XCPT)
6014	{
6015	rcStrict = VINF_SUCCESS;
6016	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
6017	}
6018	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6019	if (rcStrict != VINF_SUCCESS)
6020	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHltToR3);
6021	return VBOXSTRICTRC_VAL(rcStrict);;
6022	}
6023
6024
6025	/**
6026	* \#VMEXIT handler for MONITOR (SVM_EXIT_MONITOR). Conditional \#VMEXIT.
6027	*/
6028	HMSVM_EXIT_DECL hmR0SvmExitMonitor(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
6029	{
6030	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6031
6032	/*
6033	* If the instruction length is supplied by the CPU is 3 bytes, we can be certain that no
6034	* segment override prefix is present (and thus use the default segment DS). Otherwise, a
6035	* segment override prefix or other prefixes might be used, in which case we fallback to
6036	* IEMExecOne() to figure out.
6037	*/
6038	VBOXSTRICTRC rcStrict;
6039	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
6040	uint8_t const cbInstr = hmR0SvmSupportsNextRipSave(pVCpu) ? pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip : 0;
6041	if (cbInstr)
6042	{
6043	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_MEM_MASK \| CPUMCTX_EXTRN_DS);
6044	rcStrict = IEMExecDecodedMonitor(pVCpu, cbInstr);
6045	}
6046	else
6047	{
6048	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6049	rcStrict = IEMExecOne(pVCpu);
6050	}
6051
6052	if (rcStrict == VINF_IEM_RAISED_XCPT)
6053	{
6054	rcStrict = VINF_SUCCESS;
6055	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
6056	}
6057	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6058	return rcStrict;
6059	}
6060
6061
6062	/**
6063	* \#VMEXIT handler for MWAIT (SVM_EXIT_MWAIT). Conditional \#VMEXIT.
6064	*/
6065	HMSVM_EXIT_DECL hmR0SvmExitMwait(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
6066	{
6067	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6068
6069	VBOXSTRICTRC rcStrict;
6070	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
6071	if (fSupportsNextRipSave)
6072	{
6073	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK);
6074	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
6075	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
6076	rcStrict = IEMExecDecodedMwait(pVCpu, cbInstr);
6077	}
6078	else
6079	{
6080	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6081	rcStrict = IEMExecOne(pVCpu);
6082	}
6083
6084	if ( rcStrict == VINF_EM_HALT
6085	&& EMMonitorWaitShouldContinue(pVCpu, &pVCpu->cpum.GstCtx))
6086	rcStrict = VINF_SUCCESS;
6087	else if (rcStrict == VINF_IEM_RAISED_XCPT)
6088	{
6089	rcStrict = VINF_SUCCESS;
6090	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
6091	}
6092	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6093	return rcStrict;
6094	}
6095
6096
6097	/**
6098	* \#VMEXIT handler for shutdown (triple-fault) (SVM_EXIT_SHUTDOWN). Conditional
6099	* \#VMEXIT.
6100	*/
6101	HMSVM_EXIT_DECL hmR0SvmExitShutdown(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
6102	{
6103	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6104	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6105	return VINF_EM_RESET;
6106	}
6107
6108
6109	/**
6110	* \#VMEXIT handler for unexpected exits. Conditional \#VMEXIT.
6111	*/
6112	HMSVM_EXIT_DECL hmR0SvmExitUnexpected(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
6113	{
6114	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
6115	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6116	AssertMsgFailed(("hmR0SvmExitUnexpected: ExitCode=%#RX64 uExitInfo1=%#RX64 uExitInfo2=%#RX64\n", pSvmTransient->u64ExitCode,
6117	pVmcb->ctrl.u64ExitInfo1, pVmcb->ctrl.u64ExitInfo2));
6118	RT_NOREF(pVmcb);
6119	pVCpu->hm.s.u32HMError = (uint32_t)pSvmTransient->u64ExitCode;
6120	return VERR_SVM_UNEXPECTED_EXIT;
6121	}
6122
6123
6124	/**
6125	* \#VMEXIT handler for CRx reads (SVM_EXIT_READ_CR*). Conditional \#VMEXIT.
6126	*/
6127	HMSVM_EXIT_DECL hmR0SvmExitReadCRx(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
6128	{
6129	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6130
6131	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
6132	Log4Func(("CS:RIP=%04x:%#RX64\n", pCtx->cs.Sel, pCtx->rip));
6133	#ifdef VBOX_WITH_STATISTICS
6134	switch (pSvmTransient->u64ExitCode)
6135	{
6136	case SVM_EXIT_READ_CR0: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR0Read); break;
6137	case SVM_EXIT_READ_CR2: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR2Read); break;
6138	case SVM_EXIT_READ_CR3: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR3Read); break;
6139	case SVM_EXIT_READ_CR4: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR4Read); break;
6140	case SVM_EXIT_READ_CR8: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR8Read); break;
6141	}
6142	#endif
6143
6144	bool const fSupportsDecodeAssists = hmR0SvmSupportsDecodeAssists(pVCpu);
6145	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
6146	if ( fSupportsDecodeAssists
6147	&& fSupportsNextRipSave)
6148	{
6149	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
6150	bool const fMovCRx = RT_BOOL(pVmcb->ctrl.u64ExitInfo1 & SVM_EXIT1_MOV_CRX_MASK);
6151	if (fMovCRx)
6152	{
6153	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK \| CPUMCTX_EXTRN_CR_MASK
6154	\| CPUMCTX_EXTRN_APIC_TPR);
6155	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pCtx->rip;
6156	uint8_t const iCrReg = pSvmTransient->u64ExitCode - SVM_EXIT_READ_CR0;
6157	uint8_t const iGReg = pVmcb->ctrl.u64ExitInfo1 & SVM_EXIT1_MOV_CRX_GPR_NUMBER;
6158	VBOXSTRICTRC rcStrict = IEMExecDecodedMovCRxRead(pVCpu, cbInstr, iGReg, iCrReg);
6159	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6160	return VBOXSTRICTRC_VAL(rcStrict);
6161	}
6162	/* else: SMSW instruction, fall back below to IEM for this. */
6163	}
6164
6165	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6166	VBOXSTRICTRC rcStrict = IEMExecOne(pVCpu);
6167	AssertMsg( rcStrict == VINF_SUCCESS
6168	\|\| rcStrict == VINF_PGM_SYNC_CR3
6169	\|\| rcStrict == VINF_IEM_RAISED_XCPT,
6170	("hmR0SvmExitReadCRx: IEMExecOne failed rc=%Rrc\n", VBOXSTRICTRC_VAL(rcStrict)));
6171	Assert((pSvmTransient->u64ExitCode - SVM_EXIT_READ_CR0) <= 15);
6172	if (rcStrict == VINF_IEM_RAISED_XCPT)
6173	{
6174	rcStrict = VINF_SUCCESS;
6175	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
6176	}
6177	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6178	return rcStrict;
6179	}
6180
6181
6182	/**
6183	* \#VMEXIT handler for CRx writes (SVM_EXIT_WRITE_CR*). Conditional \#VMEXIT.
6184	*/
6185	HMSVM_EXIT_DECL hmR0SvmExitWriteCRx(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
6186	{
6187	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6188
6189	uint64_t const uExitCode = pSvmTransient->u64ExitCode;
6190	uint8_t const iCrReg = uExitCode == SVM_EXIT_CR0_SEL_WRITE ? 0 : (pSvmTransient->u64ExitCode - SVM_EXIT_WRITE_CR0);
6191	Assert(iCrReg <= 15);
6192
6193	VBOXSTRICTRC rcStrict = VERR_SVM_IPE_5;
6194	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
6195	bool fDecodedInstr = false;
6196	bool const fSupportsDecodeAssists = hmR0SvmSupportsDecodeAssists(pVCpu);
6197	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
6198	if ( fSupportsDecodeAssists
6199	&& fSupportsNextRipSave)
6200	{
6201	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
6202	bool const fMovCRx = RT_BOOL(pVmcb->ctrl.u64ExitInfo1 & SVM_EXIT1_MOV_CRX_MASK);
6203	if (fMovCRx)
6204	{
6205	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_MEM_MASK \| CPUMCTX_EXTRN_CR3 \| CPUMCTX_EXTRN_CR4
6206	\| CPUMCTX_EXTRN_APIC_TPR);
6207	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pCtx->rip;
6208	uint8_t const iGReg = pVmcb->ctrl.u64ExitInfo1 & SVM_EXIT1_MOV_CRX_GPR_NUMBER;
6209	Log4Func(("Mov CR%u w/ iGReg=%#x\n", iCrReg, iGReg));
6210	rcStrict = IEMExecDecodedMovCRxWrite(pVCpu, cbInstr, iCrReg, iGReg);
6211	fDecodedInstr = true;
6212	}
6213	/* else: LMSW or CLTS instruction, fall back below to IEM for this. */
6214	}
6215
6216	if (!fDecodedInstr)
6217	{
6218	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6219	Log4Func(("iCrReg=%#x\n", iCrReg));
6220	rcStrict = IEMExecOne(pVCpu);
6221	if (RT_UNLIKELY( rcStrict == VERR_IEM_ASPECT_NOT_IMPLEMENTED
6222	\|\| rcStrict == VERR_IEM_INSTR_NOT_IMPLEMENTED))
6223	rcStrict = VERR_EM_INTERPRETER;
6224	}
6225
6226	if (rcStrict == VINF_SUCCESS)
6227	{
6228	switch (iCrReg)
6229	{
6230	case 0:
6231	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_CR0);
6232	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR0Write);
6233	break;
6234
6235	case 2:
6236	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_CR2);
6237	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR2Write);
6238	break;
6239
6240	case 3:
6241	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_CR3);
6242	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR3Write);
6243	break;
6244
6245	case 4:
6246	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_CR4);
6247	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR4Write);
6248	break;
6249
6250	case 8:
6251	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
6252	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR8Write);
6253	break;
6254
6255	default:
6256	{
6257	AssertMsgFailed(("hmR0SvmExitWriteCRx: Invalid/Unexpected Write-CRx exit. u64ExitCode=%#RX64 %#x\n",
6258	pSvmTransient->u64ExitCode, iCrReg));
6259	break;
6260	}
6261	}
6262	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6263	}
6264	else if (rcStrict == VINF_IEM_RAISED_XCPT)
6265	{
6266	rcStrict = VINF_SUCCESS;
6267	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
6268	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6269	}
6270	else
6271	Assert(rcStrict == VERR_EM_INTERPRETER \|\| rcStrict == VINF_PGM_SYNC_CR3);
6272	return rcStrict;
6273	}
6274
6275
6276	/**
6277	* \#VMEXIT helper for read MSRs, see hmR0SvmExitMsr.
6278	*
6279	* @returns Strict VBox status code.
6280	* @param pVCpu The cross context virtual CPU structure.
6281	* @param pVmcb Pointer to the VM control block.
6282	*/
6283	static VBOXSTRICTRC hmR0SvmExitReadMsr(PVMCPUCC pVCpu, PSVMVMCB pVmcb)
6284	{
6285	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdmsr);
6286	Log4Func(("idMsr=%#RX32\n", pVCpu->cpum.GstCtx.ecx));
6287
6288	VBOXSTRICTRC rcStrict;
6289	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
6290	if (fSupportsNextRipSave)
6291	{
6292	/** @todo Optimize this: Only retrieve the MSR bits we need here. CPUMAllMsrs.cpp
6293	* can ask for what it needs instead of using CPUMCTX_EXTRN_ALL_MSRS. */
6294	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK \| CPUMCTX_EXTRN_ALL_MSRS);
6295	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
6296	rcStrict = IEMExecDecodedRdmsr(pVCpu, cbInstr);
6297	}
6298	else
6299	{
6300	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK \| CPUMCTX_EXTRN_ALL_MSRS);
6301	rcStrict = IEMExecOne(pVCpu);
6302	}
6303
6304	AssertMsg( rcStrict == VINF_SUCCESS
6305	\|\| rcStrict == VINF_IEM_RAISED_XCPT
6306	\|\| rcStrict == VINF_CPUM_R3_MSR_READ,
6307	("hmR0SvmExitReadMsr: Unexpected status %Rrc\n", VBOXSTRICTRC_VAL(rcStrict)));
6308
6309	if (rcStrict == VINF_IEM_RAISED_XCPT)
6310	{
6311	rcStrict = VINF_SUCCESS;
6312	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
6313	}
6314	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6315	return rcStrict;
6316	}
6317
6318
6319	/**
6320	* \#VMEXIT helper for write MSRs, see hmR0SvmExitMsr.
6321	*
6322	* @returns Strict VBox status code.
6323	* @param pVCpu The cross context virtual CPU structure.
6324	* @param pVmcb Pointer to the VM control block.
6325	* @param pSvmTransient Pointer to the SVM-transient structure.
6326	*/
6327	static VBOXSTRICTRC hmR0SvmExitWriteMsr(PVMCPUCC pVCpu, PSVMVMCB pVmcb, PSVMTRANSIENT pSvmTransient)
6328	{
6329	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
6330	uint32_t const idMsr = pCtx->ecx;
6331	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitWrmsr);
6332	Log4Func(("idMsr=%#RX32\n", idMsr));
6333
6334	/*
6335	* Handle TPR patching MSR writes.
6336	* We utilitize the LSTAR MSR for patching.
6337	*/
6338	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
6339	if ( idMsr == MSR_K8_LSTAR
6340	&& pVCpu->CTX_SUFF(pVM)->hm.s.fTprPatchingActive)
6341	{
6342	unsigned cbInstr;
6343	if (fSupportsNextRipSave)
6344	cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
6345	else
6346	{
6347	PDISCPUSTATE pDis = &pVCpu->hmr0.s.svm.DisState;
6348	int rc = EMInterpretDisasCurrent(pVCpu->CTX_SUFF(pVM), pVCpu, pDis, &cbInstr);
6349	if ( rc == VINF_SUCCESS
6350	&& pDis->pCurInstr->uOpcode == OP_WRMSR)
6351	Assert(cbInstr > 0);
6352	else
6353	cbInstr = 0;
6354	}
6355
6356	/* Our patch code uses LSTAR for TPR caching for 32-bit guests. */
6357	if ((pCtx->eax & 0xff) != pSvmTransient->u8GuestTpr)
6358	{
6359	int rc = APICSetTpr(pVCpu, pCtx->eax & 0xff);
6360	AssertRCReturn(rc, rc);
6361	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
6362	}
6363
6364	int rc = VINF_SUCCESS;
6365	hmR0SvmAdvanceRip(pVCpu, cbInstr);
6366	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6367	return rc;
6368	}
6369
6370	/*
6371	* Handle regular MSR writes.
6372	*/
6373	VBOXSTRICTRC rcStrict;
6374	if (fSupportsNextRipSave)
6375	{
6376	/** @todo Optimize this: We don't need to get much of the MSR state here
6377	* since we're only updating. CPUMAllMsrs.cpp can ask for what it needs and
6378	* clear the applicable extern flags. */
6379	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK \| CPUMCTX_EXTRN_ALL_MSRS);
6380	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
6381	rcStrict = IEMExecDecodedWrmsr(pVCpu, cbInstr);
6382	}
6383	else
6384	{
6385	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK \| CPUMCTX_EXTRN_ALL_MSRS);
6386	rcStrict = IEMExecOne(pVCpu);
6387	}
6388
6389	AssertMsg( rcStrict == VINF_SUCCESS
6390	\|\| rcStrict == VINF_IEM_RAISED_XCPT
6391	\|\| rcStrict == VINF_CPUM_R3_MSR_WRITE,
6392	("hmR0SvmExitWriteMsr: Unexpected status %Rrc\n", VBOXSTRICTRC_VAL(rcStrict)));
6393
6394	if (rcStrict == VINF_SUCCESS)
6395	{
6396	/* If this is an X2APIC WRMSR access, update the APIC TPR state. */
6397	if ( idMsr >= MSR_IA32_X2APIC_START
6398	&& idMsr <= MSR_IA32_X2APIC_END)
6399	{
6400	/*
6401	* We've already saved the APIC related guest-state (TPR) in hmR0SvmPostRunGuest().
6402	* When full APIC register virtualization is implemented we'll have to make sure
6403	* APIC state is saved from the VMCB before IEM changes it.
6404	*/
6405	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
6406	}
6407	else
6408	{
6409	switch (idMsr)
6410	{
6411	case MSR_IA32_TSC: pSvmTransient->fUpdateTscOffsetting = true; break;
6412	case MSR_K6_EFER: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_EFER_MSR); break;
6413	case MSR_K8_FS_BASE: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_FS); break;
6414	case MSR_K8_GS_BASE: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_GS); break;
6415	case MSR_IA32_SYSENTER_CS: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_SYSENTER_CS_MSR); break;
6416	case MSR_IA32_SYSENTER_EIP: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_SYSENTER_EIP_MSR); break;
6417	case MSR_IA32_SYSENTER_ESP: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_SYSENTER_ESP_MSR); break;
6418	}
6419	}
6420	}
6421	else if (rcStrict == VINF_IEM_RAISED_XCPT)
6422	{
6423	rcStrict = VINF_SUCCESS;
6424	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
6425	}
6426	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6427	return rcStrict;
6428	}
6429
6430
6431	/**
6432	* \#VMEXIT handler for MSR read and writes (SVM_EXIT_MSR). Conditional
6433	* \#VMEXIT.
6434	*/
6435	HMSVM_EXIT_DECL hmR0SvmExitMsr(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
6436	{
6437	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6438
6439	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
6440	if (pVmcb->ctrl.u64ExitInfo1 == SVM_EXIT1_MSR_READ)
6441	return hmR0SvmExitReadMsr(pVCpu, pVmcb);
6442
6443	Assert(pVmcb->ctrl.u64ExitInfo1 == SVM_EXIT1_MSR_WRITE);
6444	return hmR0SvmExitWriteMsr(pVCpu, pVmcb, pSvmTransient);
6445	}
6446
6447
6448	/**
6449	* \#VMEXIT handler for DRx read (SVM_EXIT_READ_DRx). Conditional \#VMEXIT.
6450	*/
6451	HMSVM_EXIT_DECL hmR0SvmExitReadDRx(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
6452	{
6453	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6454	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6455
6456	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitDRxRead);
6457
6458	/** @todo Stepping with nested-guest. */
6459	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
6460	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
6461	{
6462	/* We should -not- get this #VMEXIT if the guest's debug registers were active. */
6463	if (pSvmTransient->fWasGuestDebugStateActive)
6464	{
6465	AssertMsgFailed(("hmR0SvmExitReadDRx: Unexpected exit %#RX32\n", (uint32_t)pSvmTransient->u64ExitCode));
6466	pVCpu->hm.s.u32HMError = (uint32_t)pSvmTransient->u64ExitCode;
6467	return VERR_SVM_UNEXPECTED_EXIT;
6468	}
6469
6470	/*
6471	* Lazy DR0-3 loading.
6472	*/
6473	if (!pSvmTransient->fWasHyperDebugStateActive)
6474	{
6475	Assert(!DBGFIsStepping(pVCpu)); Assert(!pVCpu->hm.s.fSingleInstruction);
6476	Log5(("hmR0SvmExitReadDRx: Lazy loading guest debug registers\n"));
6477
6478	/* Don't intercept DRx read and writes. */
6479	PSVMVMCB pVmcb = pVCpu->hmr0.s.svm.pVmcb;
6480	pVmcb->ctrl.u16InterceptRdDRx = 0;
6481	pVmcb->ctrl.u16InterceptWrDRx = 0;
6482	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
6483
6484	/* We're playing with the host CPU state here, make sure we don't preempt or longjmp. */
6485	VMMRZCallRing3Disable(pVCpu);
6486	HM_DISABLE_PREEMPT(pVCpu);
6487
6488	/* Save the host & load the guest debug state, restart execution of the MOV DRx instruction. */
6489	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
6490	Assert(CPUMIsGuestDebugStateActive(pVCpu));
6491
6492	HM_RESTORE_PREEMPT();
6493	VMMRZCallRing3Enable(pVCpu);
6494
6495	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxContextSwitch);
6496	return VINF_SUCCESS;
6497	}
6498	}
6499
6500	/*
6501	* Interpret the read/writing of DRx.
6502	*/
6503	/** @todo Decode assist. */
6504	VBOXSTRICTRC rc = EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */);
6505	Log5(("hmR0SvmExitReadDRx: Emulated DRx access: rc=%Rrc\n", VBOXSTRICTRC_VAL(rc)));
6506	if (RT_LIKELY(rc == VINF_SUCCESS))
6507	{
6508	/* Not necessary for read accesses but whatever doesn't hurt for now, will be fixed with decode assist. */
6509	/** @todo CPUM should set this flag! */
6510	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_DR_MASK);
6511	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6512	}
6513	else
6514	Assert(rc == VERR_EM_INTERPRETER);
6515	return rc;
6516	}
6517
6518
6519	/**
6520	* \#VMEXIT handler for DRx write (SVM_EXIT_WRITE_DRx). Conditional \#VMEXIT.
6521	*/
6522	HMSVM_EXIT_DECL hmR0SvmExitWriteDRx(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
6523	{
6524	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6525	/* For now it's the same since we interpret the instruction anyway. Will change when using of Decode Assist is implemented. */
6526	VBOXSTRICTRC rc = hmR0SvmExitReadDRx(pVCpu, pSvmTransient);
6527	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitDRxWrite);
6528	STAM_COUNTER_DEC(&pVCpu->hm.s.StatExitDRxRead);
6529	return rc;
6530	}
6531
6532
6533	/**
6534	* \#VMEXIT handler for XCRx write (SVM_EXIT_XSETBV). Conditional \#VMEXIT.
6535	*/
6536	HMSVM_EXIT_DECL hmR0SvmExitXsetbv(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
6537	{
6538	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6539	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6540
6541	/** @todo decode assists... */
6542	VBOXSTRICTRC rcStrict = IEMExecOne(pVCpu);
6543	if (RT_LIKELY(rcStrict == VINF_SUCCESS))
6544	{
6545	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
6546	bool const fLoadSaveGuestXcr0 = (pCtx->cr4 & X86_CR4_OSXSAVE) && pCtx->aXcr[0] != ASMGetXcr0();
6547	Log4Func(("New XCR0=%#RX64 fLoadSaveGuestXcr0=%RTbool (cr4=%#RX64)\n", pCtx->aXcr[0], fLoadSaveGuestXcr0, pCtx->cr4));
6548	if (fLoadSaveGuestXcr0 != pVCpu->hmr0.s.fLoadSaveGuestXcr0)
6549	{
6550	pVCpu->hmr0.s.fLoadSaveGuestXcr0 = fLoadSaveGuestXcr0;
6551	hmR0SvmUpdateVmRunFunction(pVCpu);
6552	}
6553	}
6554	else if (rcStrict == VINF_IEM_RAISED_XCPT)
6555	{
6556	rcStrict = VINF_SUCCESS;
6557	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
6558	}
6559	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6560	return rcStrict;
6561	}
6562
6563
6564	/**
6565	* \#VMEXIT handler for I/O instructions (SVM_EXIT_IOIO). Conditional \#VMEXIT.
6566	*/
6567	HMSVM_EXIT_DECL hmR0SvmExitIOInstr(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
6568	{
6569	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6570	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK \| CPUMCTX_EXTRN_SREG_MASK);
6571
6572	/* I/O operation lookup arrays. */
6573	static uint32_t const s_aIOSize[8] = { 0, 1, 2, 0, 4, 0, 0, 0 }; /* Size of the I/O accesses in bytes. */
6574	static uint32_t const s_aIOOpAnd[8] = { 0, 0xff, 0xffff, 0, 0xffffffff, 0, 0, 0 }; /* AND masks for saving
6575	the result (in AL/AX/EAX). */
6576	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
6577	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
6578	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
6579
6580	Log4Func(("CS:RIP=%04x:%#RX64\n", pCtx->cs.Sel, pCtx->rip));
6581
6582	/* Refer AMD spec. 15.10.2 "IN and OUT Behaviour" and Figure 15-2. "EXITINFO1 for IOIO Intercept" for the format. */
6583	SVMIOIOEXITINFO IoExitInfo;
6584	IoExitInfo.u = (uint32_t)pVmcb->ctrl.u64ExitInfo1;
6585	uint32_t uIOWidth = (IoExitInfo.u >> 4) & 0x7;
6586	uint32_t cbValue = s_aIOSize[uIOWidth];
6587	uint32_t uAndVal = s_aIOOpAnd[uIOWidth];
6588
6589	if (RT_UNLIKELY(!cbValue))
6590	{
6591	AssertMsgFailed(("hmR0SvmExitIOInstr: Invalid IO operation. uIOWidth=%u\n", uIOWidth));
6592	return VERR_EM_INTERPRETER;
6593	}
6594
6595	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CS \| CPUMCTX_EXTRN_RIP \| CPUMCTX_EXTRN_RFLAGS);
6596	VBOXSTRICTRC rcStrict;
6597	PCEMEXITREC pExitRec = NULL;
6598	if ( !pVCpu->hm.s.fSingleInstruction
6599	&& !pVCpu->cpum.GstCtx.eflags.Bits.u1TF)
6600	pExitRec = EMHistoryUpdateFlagsAndTypeAndPC(pVCpu,
6601	!IoExitInfo.n.u1Str
6602	? IoExitInfo.n.u1Type == SVM_IOIO_READ
6603	? EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_IO_PORT_READ)
6604	: EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_IO_PORT_WRITE)
6605	: IoExitInfo.n.u1Type == SVM_IOIO_READ
6606	? EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_IO_PORT_STR_READ)
6607	: EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_IO_PORT_STR_WRITE),
6608	pVCpu->cpum.GstCtx.rip + pVCpu->cpum.GstCtx.cs.u64Base);
6609	if (!pExitRec)
6610	{
6611	bool fUpdateRipAlready = false;
6612	if (IoExitInfo.n.u1Str)
6613	{
6614	/* INS/OUTS - I/O String instruction. */
6615	/** @todo Huh? why can't we use the segment prefix information given by AMD-V
6616	* in EXITINFO1? Investigate once this thing is up and running. */
6617	Log4Func(("CS:RIP=%04x:%08RX64 %#06x/%u %c str\n", pCtx->cs.Sel, pCtx->rip, IoExitInfo.n.u16Port, cbValue,
6618	IoExitInfo.n.u1Type == SVM_IOIO_WRITE ? 'w' : 'r'));
6619	AssertReturn(pCtx->dx == IoExitInfo.n.u16Port, VERR_SVM_IPE_2);
6620	static IEMMODE const s_aenmAddrMode[8] =
6621	{
6622	(IEMMODE)-1, IEMMODE_16BIT, IEMMODE_32BIT, (IEMMODE)-1, IEMMODE_64BIT, (IEMMODE)-1, (IEMMODE)-1, (IEMMODE)-1
6623	};
6624	IEMMODE enmAddrMode = s_aenmAddrMode[(IoExitInfo.u >> 7) & 0x7];
6625	if (enmAddrMode != (IEMMODE)-1)
6626	{
6627	uint64_t cbInstr = pVmcb->ctrl.u64ExitInfo2 - pCtx->rip;
6628	if (cbInstr <= 15 && cbInstr >= 1)
6629	{
6630	Assert(cbInstr >= 1U + IoExitInfo.n.u1Rep);
6631	if (IoExitInfo.n.u1Type == SVM_IOIO_WRITE)
6632	{
6633	/* Don't know exactly how to detect whether u3Seg is valid, currently
6634	only enabling it for Bulldozer and later with NRIP. OS/2 broke on
6635	2384 Opterons when only checking NRIP. */
6636	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
6637	if ( fSupportsNextRipSave
6638	&& pVM->cpum.ro.GuestFeatures.enmMicroarch >= kCpumMicroarch_AMD_15h_First)
6639	{
6640	AssertMsg(IoExitInfo.n.u3Seg == X86_SREG_DS \|\| cbInstr > 1U + IoExitInfo.n.u1Rep,
6641	("u32Seg=%d cbInstr=%d u1REP=%d", IoExitInfo.n.u3Seg, cbInstr, IoExitInfo.n.u1Rep));
6642	rcStrict = IEMExecStringIoWrite(pVCpu, cbValue, enmAddrMode, IoExitInfo.n.u1Rep, (uint8_t)cbInstr,
6643	IoExitInfo.n.u3Seg, true /fIoChecked/);
6644	}
6645	else if (cbInstr == 1U + IoExitInfo.n.u1Rep)
6646	rcStrict = IEMExecStringIoWrite(pVCpu, cbValue, enmAddrMode, IoExitInfo.n.u1Rep, (uint8_t)cbInstr,
6647	X86_SREG_DS, true /fIoChecked/);
6648	else
6649	rcStrict = IEMExecOne(pVCpu);
6650	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOStringWrite);
6651	}
6652	else
6653	{
6654	AssertMsg(IoExitInfo.n.u3Seg == X86_SREG_ES /=0/, ("%#x\n", IoExitInfo.n.u3Seg));
6655	rcStrict = IEMExecStringIoRead(pVCpu, cbValue, enmAddrMode, IoExitInfo.n.u1Rep, (uint8_t)cbInstr,
6656	true /fIoChecked/);
6657	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOStringRead);
6658	}
6659	}
6660	else
6661	{
6662	AssertMsgFailed(("rip=%RX64 nrip=%#RX64 cbInstr=%#RX64\n", pCtx->rip, pVmcb->ctrl.u64ExitInfo2, cbInstr));
6663	rcStrict = IEMExecOne(pVCpu);
6664	}
6665	}
6666	else
6667	{
6668	AssertMsgFailed(("IoExitInfo=%RX64\n", IoExitInfo.u));
6669	rcStrict = IEMExecOne(pVCpu);
6670	}
6671	fUpdateRipAlready = true;
6672	}
6673	else
6674	{
6675	/* IN/OUT - I/O instruction. */
6676	Assert(!IoExitInfo.n.u1Rep);
6677
6678	uint8_t const cbInstr = pVmcb->ctrl.u64ExitInfo2 - pCtx->rip;
6679	if (IoExitInfo.n.u1Type == SVM_IOIO_WRITE)
6680	{
6681	rcStrict = IOMIOPortWrite(pVM, pVCpu, IoExitInfo.n.u16Port, pCtx->eax & uAndVal, cbValue);
6682	if ( rcStrict == VINF_IOM_R3_IOPORT_WRITE
6683	&& !pCtx->eflags.Bits.u1TF)
6684	rcStrict = EMRZSetPendingIoPortWrite(pVCpu, IoExitInfo.n.u16Port, cbInstr, cbValue, pCtx->eax & uAndVal);
6685	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOWrite);
6686	}
6687	else
6688	{
6689	uint32_t u32Val = 0;
6690	rcStrict = IOMIOPortRead(pVM, pVCpu, IoExitInfo.n.u16Port, &u32Val, cbValue);
6691	if (IOM_SUCCESS(rcStrict))
6692	{
6693	/* Save result of I/O IN instr. in AL/AX/EAX. */
6694	/** @todo r=bird: 32-bit op size should clear high bits of rax! */
6695	pCtx->eax = (pCtx->eax & ~uAndVal) \| (u32Val & uAndVal);
6696	}
6697	else if ( rcStrict == VINF_IOM_R3_IOPORT_READ
6698	&& !pCtx->eflags.Bits.u1TF)
6699	rcStrict = EMRZSetPendingIoPortRead(pVCpu, IoExitInfo.n.u16Port, cbInstr, cbValue);
6700
6701	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIORead);
6702	}
6703	}
6704
6705	if (IOM_SUCCESS(rcStrict))
6706	{
6707	/* AMD-V saves the RIP of the instruction following the IO instruction in EXITINFO2. */
6708	if (!fUpdateRipAlready)
6709	pCtx->rip = pVmcb->ctrl.u64ExitInfo2;
6710
6711	/*
6712	* If any I/O breakpoints are armed, we need to check if one triggered
6713	* and take appropriate action.
6714	* Note that the I/O breakpoint type is undefined if CR4.DE is 0.
6715	*/
6716	/** @todo Optimize away the DBGFBpIsHwIoArmed call by having DBGF tell the
6717	* execution engines about whether hyper BPs and such are pending. */
6718	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_DR7);
6719	uint32_t const uDr7 = pCtx->dr[7];
6720	if (RT_UNLIKELY( ( (uDr7 & X86_DR7_ENABLED_MASK)
6721	&& X86_DR7_ANY_RW_IO(uDr7)
6722	&& (pCtx->cr4 & X86_CR4_DE))
6723	\|\| DBGFBpIsHwIoArmed(pVM)))
6724	{
6725	/* We're playing with the host CPU state here, make sure we don't preempt or longjmp. */
6726	VMMRZCallRing3Disable(pVCpu);
6727	HM_DISABLE_PREEMPT(pVCpu);
6728
6729	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxIoCheck);
6730	CPUMR0DebugStateMaybeSaveGuest(pVCpu, false /fDr6/);
6731
6732	VBOXSTRICTRC rcStrict2 = DBGFBpCheckIo(pVM, pVCpu, &pVCpu->cpum.GstCtx, IoExitInfo.n.u16Port, cbValue);
6733	if (rcStrict2 == VINF_EM_RAW_GUEST_TRAP)
6734	{
6735	/* Raise #DB. */
6736	pVmcb->guest.u64DR6 = pCtx->dr[6];
6737	pVmcb->guest.u64DR7 = pCtx->dr[7];
6738	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
6739	hmR0SvmSetPendingXcptDB(pVCpu);
6740	}
6741	/* rcStrict is VINF_SUCCESS, VINF_IOM_R3_IOPORT_COMMIT_WRITE, or in [VINF_EM_FIRST..VINF_EM_LAST],
6742	however we can ditch VINF_IOM_R3_IOPORT_COMMIT_WRITE as it has VMCPU_FF_IOM as backup. */
6743	else if ( rcStrict2 != VINF_SUCCESS
6744	&& (rcStrict == VINF_SUCCESS \|\| rcStrict2 < rcStrict))
6745	rcStrict = rcStrict2;
6746	AssertCompile(VINF_EM_LAST < VINF_IOM_R3_IOPORT_COMMIT_WRITE);
6747
6748	HM_RESTORE_PREEMPT();
6749	VMMRZCallRing3Enable(pVCpu);
6750	}
6751
6752	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6753	}
6754
6755	#ifdef VBOX_STRICT
6756	if ( rcStrict == VINF_IOM_R3_IOPORT_READ
6757	\|\| rcStrict == VINF_EM_PENDING_R3_IOPORT_READ)
6758	Assert(IoExitInfo.n.u1Type == SVM_IOIO_READ);
6759	else if ( rcStrict == VINF_IOM_R3_IOPORT_WRITE
6760	\|\| rcStrict == VINF_IOM_R3_IOPORT_COMMIT_WRITE
6761	\|\| rcStrict == VINF_EM_PENDING_R3_IOPORT_WRITE)
6762	Assert(IoExitInfo.n.u1Type == SVM_IOIO_WRITE);
6763	else
6764	{
6765	/** @todo r=bird: This is missing a bunch of VINF_EM_FIRST..VINF_EM_LAST
6766	* statuses, that the VMM device and some others may return. See
6767	* IOM_SUCCESS() for guidance. */
6768	AssertMsg( RT_FAILURE(rcStrict)
6769	\|\| rcStrict == VINF_SUCCESS
6770	\|\| rcStrict == VINF_EM_RAW_EMULATE_INSTR
6771	\|\| rcStrict == VINF_EM_DBG_BREAKPOINT
6772	\|\| rcStrict == VINF_EM_RAW_GUEST_TRAP
6773	\|\| rcStrict == VINF_EM_DBG_STEPPED
6774	\|\| rcStrict == VINF_EM_RAW_TO_R3
6775	\|\| rcStrict == VINF_EM_TRIPLE_FAULT, ("%Rrc\n", VBOXSTRICTRC_VAL(rcStrict)));
6776	}
6777	#endif
6778	}
6779	else
6780	{
6781	/*
6782	* Frequent exit or something needing probing. Get state and call EMHistoryExec.
6783	*/
6784	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6785	STAM_COUNTER_INC(!IoExitInfo.n.u1Str
6786	? IoExitInfo.n.u1Type == SVM_IOIO_WRITE ? &pVCpu->hm.s.StatExitIOWrite : &pVCpu->hm.s.StatExitIORead
6787	: IoExitInfo.n.u1Type == SVM_IOIO_WRITE ? &pVCpu->hm.s.StatExitIOStringWrite : &pVCpu->hm.s.StatExitIOStringRead);
6788	Log4(("IOExit/%u: %04x:%08RX64: %s%s%s %#x LB %u -> EMHistoryExec\n",
6789	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip, IoExitInfo.n.u1Rep ? "REP " : "",
6790	IoExitInfo.n.u1Type == SVM_IOIO_WRITE ? "OUT" : "IN", IoExitInfo.n.u1Str ? "S" : "", IoExitInfo.n.u16Port, uIOWidth));
6791
6792	rcStrict = EMHistoryExec(pVCpu, pExitRec, 0);
6793	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
6794
6795	Log4(("IOExit/%u: %04x:%08RX64: EMHistoryExec -> %Rrc + %04x:%08RX64\n",
6796	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip,
6797	VBOXSTRICTRC_VAL(rcStrict), pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip));
6798	}
6799	return rcStrict;
6800	}
6801
6802
6803	/**
6804	* \#VMEXIT handler for Nested Page-faults (SVM_EXIT_NPF). Conditional \#VMEXIT.
6805	*/
6806	HMSVM_EXIT_DECL hmR0SvmExitNestedPF(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
6807	{
6808	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6809	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6810	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(pVCpu, pSvmTransient);
6811
6812	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
6813	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
6814	Assert(pVM->hmr0.s.fNestedPaging);
6815
6816	/* See AMD spec. 15.25.6 "Nested versus Guest Page Faults, Fault Ordering" for VMCB details for #NPF. */
6817	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
6818	RTGCPHYS GCPhysFaultAddr = pVmcb->ctrl.u64ExitInfo2;
6819	uint32_t u32ErrCode = pVmcb->ctrl.u64ExitInfo1; /* Note! High bits in EXITINFO1 may contain additional info and are
6820	thus intentionally not copied into u32ErrCode. */
6821
6822	Log4Func(("#NPF at CS:RIP=%04x:%#RX64 GCPhysFaultAddr=%RGp ErrCode=%#x cbInstrFetched=%u %.15Rhxs\n", pCtx->cs.Sel, pCtx->rip, GCPhysFaultAddr,
6823	u32ErrCode, pVmcb->ctrl.cbInstrFetched, pVmcb->ctrl.abInstr));
6824
6825	/*
6826	* TPR patching for 32-bit guests, using the reserved bit in the page tables for MMIO regions.
6827	*/
6828	if ( pVM->hm.s.fTprPatchingAllowed
6829	&& (GCPhysFaultAddr & PAGE_OFFSET_MASK) == XAPIC_OFF_TPR
6830	&& ( !(u32ErrCode & X86_TRAP_PF_P) /* Not present */
6831	\|\| (u32ErrCode & (X86_TRAP_PF_P \| X86_TRAP_PF_RSVD)) == (X86_TRAP_PF_P \| X86_TRAP_PF_RSVD)) /* MMIO page. */
6832	&& !CPUMIsGuestInSvmNestedHwVirtMode(pCtx)
6833	&& !CPUMIsGuestInLongModeEx(pCtx)
6834	&& !CPUMGetGuestCPL(pVCpu)
6835	&& pVM->hm.s.cPatches < RT_ELEMENTS(pVM->hm.s.aPatches))
6836	{
6837	RTGCPHYS GCPhysApicBase = APICGetBaseMsrNoCheck(pVCpu);
6838	GCPhysApicBase &= PAGE_BASE_GC_MASK;
6839
6840	if (GCPhysFaultAddr == GCPhysApicBase + XAPIC_OFF_TPR)
6841	{
6842	/* Only attempt to patch the instruction once. */
6843	PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
6844	if (!pPatch)
6845	return VINF_EM_HM_PATCH_TPR_INSTR;
6846	}
6847	}
6848
6849	/*
6850	* Determine the nested paging mode.
6851	*/
6852	/** @todo r=bird: Gotta love this nested paging hacking we're still carrying with us... (Split PGM_TYPE_NESTED.) */
6853	PGMMODE const enmNestedPagingMode = PGMGetHostMode(pVM);
6854
6855	/*
6856	* MMIO optimization using the reserved (RSVD) bit in the guest page tables for MMIO pages.
6857	*/
6858	Assert((u32ErrCode & (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P)) != X86_TRAP_PF_RSVD);
6859	if ((u32ErrCode & (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P)) == (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P))
6860	{
6861	/*
6862	* If event delivery causes an MMIO #NPF, go back to instruction emulation as otherwise
6863	* injecting the original pending event would most likely cause the same MMIO #NPF.
6864	*/
6865	if (pVCpu->hm.s.Event.fPending)
6866	{
6867	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectInterpret);
6868	return VINF_EM_RAW_INJECT_TRPM_EVENT;
6869	}
6870
6871	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CS \| CPUMCTX_EXTRN_RIP);
6872	VBOXSTRICTRC rcStrict;
6873	PCEMEXITREC pExitRec = EMHistoryUpdateFlagsAndTypeAndPC(pVCpu,
6874	EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_MMIO),
6875	pVCpu->cpum.GstCtx.rip + pVCpu->cpum.GstCtx.cs.u64Base);
6876	if (!pExitRec)
6877	{
6878
6879	rcStrict = PGMR0Trap0eHandlerNPMisconfig(pVM, pVCpu, enmNestedPagingMode, CPUMCTX2CORE(pCtx), GCPhysFaultAddr,
6880	u32ErrCode);
6881
6882	/*
6883	* If we succeed, resume guest execution.
6884	*
6885	* If we fail in interpreting the instruction because we couldn't get the guest
6886	* physical address of the page containing the instruction via the guest's page
6887	* tables (we would invalidate the guest page in the host TLB), resume execution
6888	* which would cause a guest page fault to let the guest handle this weird case.
6889	*
6890	* See @bugref{6043}.
6891	*/
6892	if ( rcStrict == VINF_SUCCESS
6893	\|\| rcStrict == VERR_PAGE_TABLE_NOT_PRESENT
6894	\|\| rcStrict == VERR_PAGE_NOT_PRESENT)
6895	{
6896	/* Successfully handled MMIO operation. */
6897	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
6898	rcStrict = VINF_SUCCESS;
6899	}
6900	}
6901	else
6902	{
6903	/*
6904	* Frequent exit or something needing probing. Get state and call EMHistoryExec.
6905	*/
6906	Assert(pCtx == &pVCpu->cpum.GstCtx);
6907	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6908	Log4(("EptMisscfgExit/%u: %04x:%08RX64: %RGp -> EMHistoryExec\n",
6909	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip, GCPhysFaultAddr));
6910
6911	rcStrict = EMHistoryExec(pVCpu, pExitRec, 0);
6912	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
6913
6914	Log4(("EptMisscfgExit/%u: %04x:%08RX64: EMHistoryExec -> %Rrc + %04x:%08RX64\n",
6915	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip,
6916	VBOXSTRICTRC_VAL(rcStrict), pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip));
6917	}
6918	return rcStrict;
6919	}
6920
6921	/*
6922	* Nested page-fault.
6923	*/
6924	TRPMAssertXcptPF(pVCpu, GCPhysFaultAddr, u32ErrCode);
6925	int rc = PGMR0Trap0eHandlerNestedPaging(pVM, pVCpu, enmNestedPagingMode, u32ErrCode, CPUMCTX2CORE(pCtx), GCPhysFaultAddr);
6926	TRPMResetTrap(pVCpu);
6927
6928	Log4Func(("#NPF: PGMR0Trap0eHandlerNestedPaging returns %Rrc CS:RIP=%04x:%#RX64\n", rc, pCtx->cs.Sel, pCtx->rip));
6929
6930	/*
6931	* Same case as PGMR0Trap0eHandlerNPMisconfig(). See comment above, @bugref{6043}.
6932	*/
6933	if ( rc == VINF_SUCCESS
6934	\|\| rc == VERR_PAGE_TABLE_NOT_PRESENT
6935	\|\| rc == VERR_PAGE_NOT_PRESENT)
6936	{
6937	/* We've successfully synced our shadow page tables. */
6938	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPF);
6939	rc = VINF_SUCCESS;
6940	}
6941
6942	/*
6943	* If delivering an event causes an #NPF (and not MMIO), we shall resolve the fault and
6944	* re-inject the original event.
6945	*/
6946	if (pVCpu->hm.s.Event.fPending)
6947	{
6948	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectReflectNPF);
6949
6950	/*
6951	* If the #NPF handler requested emulation of the instruction, ignore it.
6952	* We need to re-inject the original event so as to not lose it.
6953	* Reproducible when booting ReactOS 0.4.12 with BTRFS (installed using BootCD,
6954	* LiveCD is broken for other reasons).
6955	*/
6956	if (rc == VINF_EM_RAW_EMULATE_INSTR)
6957	rc = VINF_EM_RAW_INJECT_TRPM_EVENT;
6958	}
6959
6960	return rc;
6961	}
6962
6963
6964	/**
6965	* \#VMEXIT handler for virtual interrupt (SVM_EXIT_VINTR). Conditional
6966	* \#VMEXIT.
6967	*/
6968	HMSVM_EXIT_DECL hmR0SvmExitVIntr(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
6969	{
6970	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6971	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(&pVCpu->cpum.GstCtx);
6972
6973	/* Indicate that we no longer need to #VMEXIT when the guest is ready to receive NMIs, it is now ready. */
6974	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
6975	hmR0SvmClearIntWindowExiting(pVCpu, pVmcb);
6976
6977	/* Deliver the pending interrupt via hmR0SvmEvaluatePendingEvent() and resume guest execution. */
6978	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIntWindow);
6979	return VINF_SUCCESS;
6980	}
6981
6982
6983	/**
6984	* \#VMEXIT handler for task switches (SVM_EXIT_TASK_SWITCH). Conditional
6985	* \#VMEXIT.
6986	*/
6987	HMSVM_EXIT_DECL hmR0SvmExitTaskSwitch(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
6988	{
6989	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6990	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(pVCpu, pSvmTransient);
6991
6992	#ifndef HMSVM_ALWAYS_TRAP_TASK_SWITCH
6993	Assert(!pVCpu->CTX_SUFF(pVM)->hmr0.s.fNestedPaging);
6994	#endif
6995
6996	/* Check if this task-switch occurred while delivering an event through the guest IDT. */
6997	if (pVCpu->hm.s.Event.fPending) /* Can happen with exceptions/NMI. See @bugref{8411}. */
6998	{
6999	/*
7000	* AMD-V provides us with the exception which caused the TS; we collect
7001	* the information in the call to hmR0SvmCheckExitDueToEventDelivery().
7002	*/
7003	Log4Func(("TS occurred during event delivery\n"));
7004	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitTaskSwitch);
7005	return VINF_EM_RAW_INJECT_TRPM_EVENT;
7006	}
7007
7008	/** @todo Emulate task switch someday, currently just going back to ring-3 for
7009	* emulation. */
7010	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitTaskSwitch);
7011	return VERR_EM_INTERPRETER;
7012	}
7013
7014
7015	/**
7016	* \#VMEXIT handler for VMMCALL (SVM_EXIT_VMMCALL). Conditional \#VMEXIT.
7017	*/
7018	HMSVM_EXIT_DECL hmR0SvmExitVmmCall(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7019	{
7020	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7021	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7022
7023	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
7024	if (pVM->hm.s.fTprPatchingAllowed)
7025	{
7026	int rc = hmEmulateSvmMovTpr(pVM, pVCpu);
7027	if (rc != VERR_NOT_FOUND)
7028	{
7029	Log4Func(("hmEmulateSvmMovTpr returns %Rrc\n", rc));
7030	return rc;
7031	}
7032	}
7033
7034	if (EMAreHypercallInstructionsEnabled(pVCpu))
7035	{
7036	unsigned cbInstr;
7037	if (hmR0SvmSupportsNextRipSave(pVCpu))
7038	{
7039	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7040	cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
7041	}
7042	else
7043	{
7044	PDISCPUSTATE pDis = &pVCpu->hmr0.s.svm.DisState;
7045	int rc = EMInterpretDisasCurrent(pVCpu->CTX_SUFF(pVM), pVCpu, pDis, &cbInstr);
7046	if ( rc == VINF_SUCCESS
7047	&& pDis->pCurInstr->uOpcode == OP_VMMCALL)
7048	Assert(cbInstr > 0);
7049	else
7050	cbInstr = 0;
7051	}
7052
7053	VBOXSTRICTRC rcStrict = GIMHypercall(pVCpu, &pVCpu->cpum.GstCtx);
7054	if (RT_SUCCESS(rcStrict))
7055	{
7056	/* Only update the RIP if we're continuing guest execution and not in the case
7057	of say VINF_GIM_R3_HYPERCALL. */
7058	if (rcStrict == VINF_SUCCESS)
7059	hmR0SvmAdvanceRip(pVCpu, cbInstr);
7060
7061	return VBOXSTRICTRC_VAL(rcStrict);
7062	}
7063	else
7064	Log4Func(("GIMHypercall returns %Rrc -> #UD\n", VBOXSTRICTRC_VAL(rcStrict)));
7065	}
7066
7067	hmR0SvmSetPendingXcptUD(pVCpu);
7068	return VINF_SUCCESS;
7069	}
7070
7071
7072	/**
7073	* \#VMEXIT handler for VMMCALL (SVM_EXIT_VMMCALL). Conditional \#VMEXIT.
7074	*/
7075	HMSVM_EXIT_DECL hmR0SvmExitPause(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7076	{
7077	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7078
7079	unsigned cbInstr;
7080	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
7081	if (fSupportsNextRipSave)
7082	{
7083	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7084	cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
7085	}
7086	else
7087	{
7088	PDISCPUSTATE pDis = &pVCpu->hmr0.s.svm.DisState;
7089	int rc = EMInterpretDisasCurrent(pVCpu->CTX_SUFF(pVM), pVCpu, pDis, &cbInstr);
7090	if ( rc == VINF_SUCCESS
7091	&& pDis->pCurInstr->uOpcode == OP_PAUSE)
7092	Assert(cbInstr > 0);
7093	else
7094	cbInstr = 0;
7095	}
7096
7097	/** @todo The guest has likely hit a contended spinlock. We might want to
7098	* poke a schedule different guest VCPU. */
7099	hmR0SvmAdvanceRip(pVCpu, cbInstr);
7100	return VINF_EM_RAW_INTERRUPT;
7101	}
7102
7103
7104	/**
7105	* \#VMEXIT handler for FERR intercept (SVM_EXIT_FERR_FREEZE). Conditional
7106	* \#VMEXIT.
7107	*/
7108	HMSVM_EXIT_DECL hmR0SvmExitFerrFreeze(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7109	{
7110	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7111	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR0);
7112	Assert(!(pVCpu->cpum.GstCtx.cr0 & X86_CR0_NE));
7113
7114	Log4Func(("Raising IRQ 13 in response to #FERR\n"));
7115	return PDMIsaSetIrq(pVCpu->CTX_SUFF(pVM), 13 /* u8Irq /, 1 / u8Level /, 0 / uTagSrc */);
7116	}
7117
7118
7119	/**
7120	* \#VMEXIT handler for IRET (SVM_EXIT_IRET). Conditional \#VMEXIT.
7121	*/
7122	HMSVM_EXIT_DECL hmR0SvmExitIret(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7123	{
7124	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7125
7126	/* Indicate that we no longer need to #VMEXIT when the guest is ready to receive NMIs, it is now (almost) ready. */
7127	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7128	hmR0SvmClearCtrlIntercept(pVCpu, pVmcb, SVM_CTRL_INTERCEPT_IRET);
7129
7130	/* Emulate the IRET. We have to execute the IRET before an NMI, but must potentially
7131	* deliver a pending NMI right after. If the IRET faults, an NMI can come before the
7132	* handler executes. Yes, x86 is ugly.
7133	*/
7134	return VINF_EM_RAW_EMULATE_INSTR;
7135	}
7136
7137
7138	/**
7139	* \#VMEXIT handler for page-fault exceptions (SVM_EXIT_XCPT_14).
7140	* Conditional \#VMEXIT.
7141	*/
7142	HMSVM_EXIT_DECL hmR0SvmExitXcptPF(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7143	{
7144	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7145	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7146	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(pVCpu, pSvmTransient);
7147
7148	/* See AMD spec. 15.12.15 "#PF (Page Fault)". */
7149	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
7150	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
7151	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7152	uint32_t uErrCode = pVmcb->ctrl.u64ExitInfo1;
7153	uint64_t const uFaultAddress = pVmcb->ctrl.u64ExitInfo2;
7154
7155	#if defined(HMSVM_ALWAYS_TRAP_ALL_XCPTS) \|\| defined(HMSVM_ALWAYS_TRAP_PF)
7156	if (pVM->hmr0.s.fNestedPaging)
7157	{
7158	pVCpu->hm.s.Event.fPending = false; /* In case it's a contributory or vectoring #PF. */
7159	if ( !pSvmTransient->fVectoringDoublePF
7160	\|\| CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
7161	{
7162	/* A genuine guest #PF, reflect it to the guest. */
7163	hmR0SvmSetPendingXcptPF(pVCpu, uErrCode, uFaultAddress);
7164	Log4Func(("#PF: Guest page fault at %04X:%RGv FaultAddr=%RX64 ErrCode=%#x\n", pCtx->cs.Sel, (RTGCPTR)pCtx->rip,
7165	uFaultAddress, uErrCode));
7166	}
7167	else
7168	{
7169	/* A guest page-fault occurred during delivery of a page-fault. Inject #DF. */
7170	hmR0SvmSetPendingXcptDF(pVCpu);
7171	Log4Func(("Pending #DF due to vectoring #PF. NP\n"));
7172	}
7173	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestPF);
7174	return VINF_SUCCESS;
7175	}
7176	#endif
7177
7178	Assert(!pVM->hmr0.s.fNestedPaging);
7179
7180	/*
7181	* TPR patching shortcut for APIC TPR reads and writes; only applicable to 32-bit guests.
7182	*/
7183	if ( pVM->hm.s.fTprPatchingAllowed
7184	&& (uFaultAddress & 0xfff) == XAPIC_OFF_TPR
7185	&& !(uErrCode & X86_TRAP_PF_P) /* Not present. */
7186	&& !CPUMIsGuestInSvmNestedHwVirtMode(pCtx)
7187	&& !CPUMIsGuestInLongModeEx(pCtx)
7188	&& !CPUMGetGuestCPL(pVCpu)
7189	&& pVM->hm.s.cPatches < RT_ELEMENTS(pVM->hm.s.aPatches))
7190	{
7191	RTGCPHYS GCPhysApicBase;
7192	GCPhysApicBase = APICGetBaseMsrNoCheck(pVCpu);
7193	GCPhysApicBase &= PAGE_BASE_GC_MASK;
7194
7195	/* Check if the page at the fault-address is the APIC base. */
7196	PGMPTWALK Walk;
7197	int rc2 = PGMGstGetPage(pVCpu, (RTGCPTR)uFaultAddress, &Walk);
7198	if ( rc2 == VINF_SUCCESS
7199	&& Walk.GCPhys == GCPhysApicBase)
7200	{
7201	/* Only attempt to patch the instruction once. */
7202	PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
7203	if (!pPatch)
7204	return VINF_EM_HM_PATCH_TPR_INSTR;
7205	}
7206	}
7207
7208	Log4Func(("#PF: uFaultAddress=%#RX64 CS:RIP=%#04x:%#RX64 uErrCode %#RX32 cr3=%#RX64\n", uFaultAddress, pCtx->cs.Sel,
7209	pCtx->rip, uErrCode, pCtx->cr3));
7210
7211	/*
7212	* If it's a vectoring #PF, emulate injecting the original event injection as
7213	* PGMTrap0eHandler() is incapable of differentiating between instruction emulation and
7214	* event injection that caused a #PF. See @bugref{6607}.
7215	*/
7216	if (pSvmTransient->fVectoringPF)
7217	{
7218	Assert(pVCpu->hm.s.Event.fPending);
7219	return VINF_EM_RAW_INJECT_TRPM_EVENT;
7220	}
7221
7222	TRPMAssertXcptPF(pVCpu, uFaultAddress, uErrCode);
7223	int rc = PGMTrap0eHandler(pVCpu, uErrCode, CPUMCTX2CORE(pCtx), (RTGCPTR)uFaultAddress);
7224
7225	Log4Func(("#PF: rc=%Rrc\n", rc));
7226
7227	if (rc == VINF_SUCCESS)
7228	{
7229	/* Successfully synced shadow pages tables or emulated an MMIO instruction. */
7230	TRPMResetTrap(pVCpu);
7231	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPF);
7232	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
7233	return rc;
7234	}
7235
7236	if (rc == VINF_EM_RAW_GUEST_TRAP)
7237	{
7238	pVCpu->hm.s.Event.fPending = false; /* In case it's a contributory or vectoring #PF. */
7239
7240	/*
7241	* If a nested-guest delivers a #PF and that causes a #PF which is -not- a shadow #PF,
7242	* we should simply forward the #PF to the guest and is up to the nested-hypervisor to
7243	* determine whether it is a nested-shadow #PF or a #DF, see @bugref{7243#c121}.
7244	*/
7245	if ( !pSvmTransient->fVectoringDoublePF
7246	\|\| CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
7247	{
7248	/* It's a guest (or nested-guest) page fault and needs to be reflected. */
7249	uErrCode = TRPMGetErrorCode(pVCpu); /* The error code might have been changed. */
7250	TRPMResetTrap(pVCpu);
7251
7252	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
7253	/* If the nested-guest is intercepting #PFs, cause a #PF #VMEXIT. */
7254	if ( CPUMIsGuestInSvmNestedHwVirtMode(pCtx)
7255	&& CPUMIsGuestSvmXcptInterceptSet(pVCpu, pCtx, X86_XCPT_PF))
7256	return IEMExecSvmVmexit(pVCpu, SVM_EXIT_XCPT_PF, uErrCode, uFaultAddress);
7257	#endif
7258
7259	hmR0SvmSetPendingXcptPF(pVCpu, uErrCode, uFaultAddress);
7260	}
7261	else
7262	{
7263	/* A guest page-fault occurred during delivery of a page-fault. Inject #DF. */
7264	TRPMResetTrap(pVCpu);
7265	hmR0SvmSetPendingXcptDF(pVCpu);
7266	Log4Func(("#PF: Pending #DF due to vectoring #PF\n"));
7267	}
7268
7269	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestPF);
7270	return VINF_SUCCESS;
7271	}
7272
7273	TRPMResetTrap(pVCpu);
7274	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPFEM);
7275	return rc;
7276	}
7277
7278
7279	/**
7280	* \#VMEXIT handler for undefined opcode (SVM_EXIT_XCPT_6).
7281	* Conditional \#VMEXIT.
7282	*/
7283	HMSVM_EXIT_DECL hmR0SvmExitXcptUD(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7284	{
7285	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7286	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(&pVCpu->cpum.GstCtx);
7287	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestUD);
7288
7289	/* Paranoia; Ensure we cannot be called as a result of event delivery. */
7290	PSVMVMCB pVmcb = pVCpu->hmr0.s.svm.pVmcb;
7291	Assert(!pVmcb->ctrl.ExitIntInfo.n.u1Valid); NOREF(pVmcb);
7292
7293	/** @todo if we accumulate more optional stuff here, we ought to combine the
7294	* reading of opcode bytes to avoid doing more than once. */
7295
7296	VBOXSTRICTRC rcStrict = VERR_SVM_UNEXPECTED_XCPT_EXIT;
7297	if (pVCpu->hm.s.fGIMTrapXcptUD)
7298	{
7299	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7300	uint8_t cbInstr = 0;
7301	rcStrict = GIMXcptUD(pVCpu, &pVCpu->cpum.GstCtx, NULL /* pDis */, &cbInstr);
7302	if (rcStrict == VINF_SUCCESS)
7303	{
7304	/* #UD #VMEXIT does not have valid NRIP information, manually advance RIP. See @bugref{7270#c170}. */
7305	hmR0SvmAdvanceRip(pVCpu, cbInstr);
7306	rcStrict = VINF_SUCCESS;
7307	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
7308	}
7309	else if (rcStrict == VINF_GIM_HYPERCALL_CONTINUING)
7310	rcStrict = VINF_SUCCESS;
7311	else if (rcStrict == VINF_GIM_R3_HYPERCALL)
7312	rcStrict = VINF_GIM_R3_HYPERCALL;
7313	else
7314	{
7315	Assert(RT_FAILURE(VBOXSTRICTRC_VAL(rcStrict)));
7316	rcStrict = VERR_SVM_UNEXPECTED_XCPT_EXIT;
7317	}
7318	}
7319
7320	if (pVCpu->hm.s.svm.fEmulateLongModeSysEnterExit)
7321	{
7322	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CS \| CPUMCTX_EXTRN_SS \| CPUMCTX_EXTRN_RIP \| CPUMCTX_EXTRN_RFLAGS
7323	\| CPUMCTX_EXTRN_CR0 \| CPUMCTX_EXTRN_CR3 \| CPUMCTX_EXTRN_CR4 \| CPUMCTX_EXTRN_EFER);
7324	if (CPUMIsGuestInLongModeEx(&pVCpu->cpum.GstCtx))
7325	{
7326	/* Ideally, IEM should just handle all these special #UD situations, but
7327	we don't quite trust things to behave optimially when doing that. So,
7328	for now we'll restrict ourselves to a handful of possible sysenter and
7329	sysexit encodings that we filter right here. */
7330	uint8_t abInstr[SVM_CTRL_GUEST_INSTR_BYTES_MAX];
7331	uint8_t cbInstr = pVmcb->ctrl.cbInstrFetched;
7332	uint32_t const uCpl = CPUMGetGuestCPL(pVCpu);
7333	uint8_t const cbMin = uCpl != 0 ? 2 : 1 + 2;
7334	RTGCPTR const GCPtrInstr = pVCpu->cpum.GstCtx.rip + pVCpu->cpum.GstCtx.cs.u64Base;
7335	if (cbInstr < cbMin \|\| cbInstr > SVM_CTRL_GUEST_INSTR_BYTES_MAX)
7336	{
7337	cbInstr = cbMin;
7338	int rc2 = PGMPhysSimpleReadGCPtr(pVCpu, abInstr, GCPtrInstr, cbInstr);
7339	AssertRCStmt(rc2, cbInstr = 0);
7340	}
7341	else
7342	memcpy(abInstr, pVmcb->ctrl.abInstr, cbInstr); /* unlikely */
7343	if ( cbInstr == 0 /* read error */
7344	\|\| (cbInstr >= 2 && abInstr[0] == 0x0f && abInstr[1] == 0x34) /* sysenter */
7345	\|\| ( uCpl == 0
7346	&& ( ( cbInstr >= 2 && abInstr[0] == 0x0f && abInstr[1] == 0x35) /* sysexit */
7347	\|\| ( cbInstr >= 3 && abInstr[1] == 0x0f && abInstr[2] == 0x35 /* rex.w sysexit */
7348	&& (abInstr[0] & (X86_OP_REX_W \| 0xf0)) == X86_OP_REX_W))))
7349	{
7350	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK
7351	\| CPUMCTX_EXTRN_SREG_MASK /* without ES+DS+GS the app will #GP later - go figure */);
7352	Log6(("hmR0SvmExitXcptUD: sysenter/sysexit: %.*Rhxs at %#llx CPL=%u\n", cbInstr, abInstr, GCPtrInstr, uCpl));
7353	rcStrict = IEMExecOneWithPrefetchedByPC(pVCpu, CPUMCTX2CORE(&pVCpu->cpum.GstCtx), GCPtrInstr, abInstr, cbInstr);
7354	Log6(("hmR0SvmExitXcptUD: sysenter/sysexit: rcStrict=%Rrc %04x:%08RX64 %08RX64 %04x:%08RX64\n",
7355	VBOXSTRICTRC_VAL(rcStrict), pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip, pVCpu->cpum.GstCtx.rflags.u,
7356	pVCpu->cpum.GstCtx.ss.Sel, pVCpu->cpum.GstCtx.rsp));
7357	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestUD);
7358	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK); /** @todo Lazy bird. */
7359	if (rcStrict == VINF_IEM_RAISED_XCPT)
7360	rcStrict = VINF_SUCCESS;
7361	return rcStrict;
7362	}
7363	Log6(("hmR0SvmExitXcptUD: not sysenter/sysexit: %.*Rhxs at %#llx CPL=%u\n", cbInstr, abInstr, GCPtrInstr, uCpl));
7364	}
7365	else
7366	Log6(("hmR0SvmExitXcptUD: not in long mode at %04x:%llx\n", pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip));
7367	}
7368
7369	/* If the GIM #UD exception handler didn't succeed for some reason or wasn't needed, raise #UD. */
7370	if (RT_FAILURE(rcStrict))
7371	{
7372	hmR0SvmSetPendingXcptUD(pVCpu);
7373	rcStrict = VINF_SUCCESS;
7374	}
7375
7376	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestUD);
7377	return rcStrict;
7378	}
7379
7380
7381	/**
7382	* \#VMEXIT handler for math-fault exceptions (SVM_EXIT_XCPT_16).
7383	* Conditional \#VMEXIT.
7384	*/
7385	HMSVM_EXIT_DECL hmR0SvmExitXcptMF(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7386	{
7387	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7388	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7389	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestMF);
7390
7391	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
7392	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7393
7394	/* Paranoia; Ensure we cannot be called as a result of event delivery. */
7395	Assert(!pVmcb->ctrl.ExitIntInfo.n.u1Valid); NOREF(pVmcb);
7396
7397	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestMF);
7398
7399	if (!(pCtx->cr0 & X86_CR0_NE))
7400	{
7401	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
7402	PDISSTATE pDis = &pVCpu->hmr0.s.svm.DisState;
7403	unsigned cbInstr;
7404	int rc = EMInterpretDisasCurrent(pVM, pVCpu, pDis, &cbInstr);
7405	if (RT_SUCCESS(rc))
7406	{
7407	/* Convert a #MF into a FERR -> IRQ 13. See @bugref{6117}. */
7408	rc = PDMIsaSetIrq(pVCpu->CTX_SUFF(pVM), 13 /* u8Irq /, 1 / u8Level /, 0 / uTagSrc */);
7409	if (RT_SUCCESS(rc))
7410	hmR0SvmAdvanceRip(pVCpu, cbInstr);
7411	}
7412	else
7413	Log4Func(("EMInterpretDisasCurrent returned %Rrc uOpCode=%#x\n", rc, pDis->pCurInstr->uOpcode));
7414	return rc;
7415	}
7416
7417	hmR0SvmSetPendingXcptMF(pVCpu);
7418	return VINF_SUCCESS;
7419	}
7420
7421
7422	/**
7423	* \#VMEXIT handler for debug exceptions (SVM_EXIT_XCPT_1). Conditional
7424	* \#VMEXIT.
7425	*/
7426	HMSVM_EXIT_DECL hmR0SvmExitXcptDB(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7427	{
7428	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7429	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7430	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(pVCpu, pSvmTransient);
7431	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestDB);
7432
7433	if (RT_UNLIKELY(pVCpu->hm.s.Event.fPending))
7434	{
7435	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectInterpret);
7436	return VINF_EM_RAW_INJECT_TRPM_EVENT;
7437	}
7438
7439	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestDB);
7440
7441	/*
7442	* This can be a fault-type #DB (instruction breakpoint) or a trap-type #DB (data
7443	* breakpoint). However, for both cases DR6 and DR7 are updated to what the exception
7444	* handler expects. See AMD spec. 15.12.2 "#DB (Debug)".
7445	*/
7446	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
7447	PSVMVMCB pVmcb = pVCpu->hmr0.s.svm.pVmcb;
7448	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
7449	int rc = DBGFTrap01Handler(pVM, pVCpu, CPUMCTX2CORE(pCtx), pVmcb->guest.u64DR6, pVCpu->hm.s.fSingleInstruction);
7450	if (rc == VINF_EM_RAW_GUEST_TRAP)
7451	{
7452	Log5(("hmR0SvmExitXcptDB: DR6=%#RX64 -> guest trap\n", pVmcb->guest.u64DR6));
7453	if (CPUMIsHyperDebugStateActive(pVCpu))
7454	CPUMSetGuestDR6(pVCpu, CPUMGetGuestDR6(pVCpu) \| pVmcb->guest.u64DR6);
7455
7456	/* Reflect the exception back to the guest. */
7457	hmR0SvmSetPendingXcptDB(pVCpu);
7458	rc = VINF_SUCCESS;
7459	}
7460
7461	/*
7462	* Update DR6.
7463	*/
7464	if (CPUMIsHyperDebugStateActive(pVCpu))
7465	{
7466	Log5(("hmR0SvmExitXcptDB: DR6=%#RX64 -> %Rrc\n", pVmcb->guest.u64DR6, rc));
7467	pVmcb->guest.u64DR6 = X86_DR6_INIT_VAL;
7468	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
7469	}
7470	else
7471	{
7472	AssertMsg(rc == VINF_SUCCESS, ("rc=%Rrc\n", rc));
7473	Assert(!pVCpu->hm.s.fSingleInstruction && !DBGFIsStepping(pVCpu));
7474	}
7475
7476	return rc;
7477	}
7478
7479
7480	/**
7481	* \#VMEXIT handler for alignment check exceptions (SVM_EXIT_XCPT_17).
7482	* Conditional \#VMEXIT.
7483	*/
7484	HMSVM_EXIT_DECL hmR0SvmExitXcptAC(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7485	{
7486	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7487	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(pVCpu, pSvmTransient);
7488	STAM_REL_COUNTER_INC(&pVCpu->hm.s.StatExitGuestAC);
7489
7490	SVMEVENT Event;
7491	Event.u = 0;
7492	Event.n.u1Valid = 1;
7493	Event.n.u3Type = SVM_EVENT_EXCEPTION;
7494	Event.n.u8Vector = X86_XCPT_AC;
7495	Event.n.u1ErrorCodeValid = 1;
7496	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
7497	return VINF_SUCCESS;
7498	}
7499
7500
7501	/**
7502	* \#VMEXIT handler for breakpoint exceptions (SVM_EXIT_XCPT_3).
7503	* Conditional \#VMEXIT.
7504	*/
7505	HMSVM_EXIT_DECL hmR0SvmExitXcptBP(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7506	{
7507	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7508	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7509	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(pVCpu, pSvmTransient);
7510	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestBP);
7511
7512	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
7513	int rc = DBGFTrap03Handler(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
7514	if (rc == VINF_EM_RAW_GUEST_TRAP)
7515	{
7516	SVMEVENT Event;
7517	Event.u = 0;
7518	Event.n.u1Valid = 1;
7519	Event.n.u3Type = SVM_EVENT_EXCEPTION;
7520	Event.n.u8Vector = X86_XCPT_BP;
7521	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
7522	rc = VINF_SUCCESS;
7523	}
7524
7525	Assert(rc == VINF_SUCCESS \|\| rc == VINF_EM_DBG_BREAKPOINT);
7526	return rc;
7527	}
7528
7529
7530	/**
7531	* Hacks its way around the lovely mesa driver's backdoor accesses.
7532	*
7533	* @sa hmR0VmxHandleMesaDrvGp
7534	*/
7535	static int hmR0SvmHandleMesaDrvGp(PVMCPUCC pVCpu, PCPUMCTX pCtx, PCSVMVMCB pVmcb)
7536	{
7537	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CS \| CPUMCTX_EXTRN_RIP \| CPUMCTX_EXTRN_RFLAGS \| CPUMCTX_EXTRN_GPRS_MASK);
7538	Log(("hmR0SvmHandleMesaDrvGp: at %04x:%08RX64 rcx=%RX64 rbx=%RX64\n",
7539	pVmcb->guest.CS.u16Sel, pVmcb->guest.u64RIP, pCtx->rcx, pCtx->rbx));
7540	RT_NOREF(pCtx, pVmcb);
7541
7542	/* For now we'll just skip the instruction. */
7543	hmR0SvmAdvanceRip(pVCpu, 1);
7544	return VINF_SUCCESS;
7545	}
7546
7547
7548	/**
7549	* Checks if the \#GP'ing instruction is the mesa driver doing it's lovely
7550	* backdoor logging w/o checking what it is running inside.
7551	*
7552	* This recognizes an "IN EAX,DX" instruction executed in flat ring-3, with the
7553	* backdoor port and magic numbers loaded in registers.
7554	*
7555	* @returns true if it is, false if it isn't.
7556	* @sa hmR0VmxIsMesaDrvGp
7557	*/
7558	DECLINLINE(bool) hmR0SvmIsMesaDrvGp(PVMCPUCC pVCpu, PCPUMCTX pCtx, PCSVMVMCB pVmcb)
7559	{
7560	/* Check magic and port. */
7561	Assert(!(pCtx->fExtrn & (CPUMCTX_EXTRN_RDX \| CPUMCTX_EXTRN_RCX)));
7562	/Log8(("hmR0SvmIsMesaDrvGp: rax=%RX64 rdx=%RX64\n", pCtx->fExtrn & CPUMCTX_EXTRN_RAX ? pVmcb->guest.u64RAX : pCtx->rax, pCtx->rdx));/
7563	if (pCtx->dx != UINT32_C(0x5658))
7564	return false;
7565	if ((pCtx->fExtrn & CPUMCTX_EXTRN_RAX ? pVmcb->guest.u64RAX : pCtx->rax) != UINT32_C(0x564d5868))
7566	return false;
7567
7568	/* Check that it is #GP(0). */
7569	if (pVmcb->ctrl.u64ExitInfo1 != 0)
7570	return false;
7571
7572	/* Flat ring-3 CS. */
7573	/Log8(("hmR0SvmIsMesaDrvGp: u8CPL=%d base=%RX64\n", pVmcb->guest.u8CPL, pCtx->fExtrn & CPUMCTX_EXTRN_CS ? pVmcb->guest.CS.u64Base : pCtx->cs.u64Base));/
7574	if (pVmcb->guest.u8CPL != 3)
7575	return false;
7576	if ((pCtx->fExtrn & CPUMCTX_EXTRN_CS ? pVmcb->guest.CS.u64Base : pCtx->cs.u64Base) != 0)
7577	return false;
7578
7579	/* 0xed: IN eAX,dx */
7580	if (pVmcb->ctrl.cbInstrFetched < 1) /* unlikely, it turns out. */
7581	{
7582	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CS \| CPUMCTX_EXTRN_RIP \| CPUMCTX_EXTRN_GPRS_MASK
7583	\| CPUMCTX_EXTRN_CR0 \| CPUMCTX_EXTRN_CR3 \| CPUMCTX_EXTRN_CR4 \| CPUMCTX_EXTRN_EFER);
7584	uint8_t abInstr[1];
7585	int rc = PGMPhysSimpleReadGCPtr(pVCpu, abInstr, pCtx->rip, sizeof(abInstr));
7586	/Log8(("hmR0SvmIsMesaDrvGp: PGMPhysSimpleReadGCPtr -> %Rrc %#x\n", rc, abInstr[0])); /
7587	if (RT_FAILURE(rc))
7588	return false;
7589	if (abInstr[0] != 0xed)
7590	return false;
7591	}
7592	else
7593	{
7594	/Log8(("hmR0SvmIsMesaDrvGp: %#x\n", pVmcb->ctrl.abInstr));/
7595	if (pVmcb->ctrl.abInstr[0] != 0xed)
7596	return false;
7597	}
7598	return true;
7599	}
7600
7601
7602	/**
7603	* \#VMEXIT handler for general protection faults (SVM_EXIT_XCPT_BP).
7604	* Conditional \#VMEXIT.
7605	*/
7606	HMSVM_EXIT_DECL hmR0SvmExitXcptGP(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7607	{
7608	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7609	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(pVCpu, pSvmTransient);
7610	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestGP);
7611
7612	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7613	Assert(pSvmTransient->u64ExitCode == pVmcb->ctrl.u64ExitCode);
7614
7615	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
7616	if ( !pVCpu->hm.s.fTrapXcptGpForLovelyMesaDrv
7617	\|\| !hmR0SvmIsMesaDrvGp(pVCpu, pCtx, pVmcb))
7618	{
7619	SVMEVENT Event;
7620	Event.u = 0;
7621	Event.n.u1Valid = 1;
7622	Event.n.u3Type = SVM_EVENT_EXCEPTION;
7623	Event.n.u8Vector = X86_XCPT_GP;
7624	Event.n.u1ErrorCodeValid = 1;
7625	Event.n.u32ErrorCode = (uint32_t)pVmcb->ctrl.u64ExitInfo1;
7626	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
7627	return VINF_SUCCESS;
7628	}
7629	return hmR0SvmHandleMesaDrvGp(pVCpu, pCtx, pVmcb);
7630	}
7631
7632
7633	#if defined(HMSVM_ALWAYS_TRAP_ALL_XCPTS) \|\| defined(VBOX_WITH_NESTED_HWVIRT_SVM)
7634	/**
7635	* \#VMEXIT handler for generic exceptions. Conditional \#VMEXIT.
7636	*/
7637	HMSVM_EXIT_DECL hmR0SvmExitXcptGeneric(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7638	{
7639	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7640	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(pVCpu, pSvmTransient);
7641
7642	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7643	uint8_t const uVector = pVmcb->ctrl.u64ExitCode - SVM_EXIT_XCPT_0;
7644	uint32_t const uErrCode = pVmcb->ctrl.u64ExitInfo1;
7645	Assert(pSvmTransient->u64ExitCode == pVmcb->ctrl.u64ExitCode);
7646	Assert(uVector <= X86_XCPT_LAST);
7647	Log4Func(("uVector=%#x uErrCode=%u\n", uVector, uErrCode));
7648
7649	SVMEVENT Event;
7650	Event.u = 0;
7651	Event.n.u1Valid = 1;
7652	Event.n.u3Type = SVM_EVENT_EXCEPTION;
7653	Event.n.u8Vector = uVector;
7654	switch (uVector)
7655	{
7656	/* Shouldn't be here for reflecting #PFs (among other things, the fault address isn't passed along). */
7657	case X86_XCPT_PF: AssertMsgFailed(("hmR0SvmExitXcptGeneric: Unexpected exception")); return VERR_SVM_IPE_5;
7658	case X86_XCPT_DF:
7659	case X86_XCPT_TS:
7660	case X86_XCPT_NP:
7661	case X86_XCPT_SS:
7662	case X86_XCPT_GP:
7663	case X86_XCPT_AC:
7664	{
7665	Event.n.u1ErrorCodeValid = 1;
7666	Event.n.u32ErrorCode = uErrCode;
7667	break;
7668	}
7669	}
7670
7671	#ifdef VBOX_WITH_STATISTICS
7672	switch (uVector)
7673	{
7674	case X86_XCPT_DE: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestDE); break;
7675	case X86_XCPT_DB: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestDB); break;
7676	case X86_XCPT_BP: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestBP); break;
7677	case X86_XCPT_OF: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestOF); break;
7678	case X86_XCPT_BR: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestBR); break;
7679	case X86_XCPT_UD: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestUD); break;
7680	case X86_XCPT_NM: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestOF); break;
7681	case X86_XCPT_DF: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestDF); break;
7682	case X86_XCPT_TS: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestTS); break;
7683	case X86_XCPT_NP: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestNP); break;
7684	case X86_XCPT_SS: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestSS); break;
7685	case X86_XCPT_GP: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestGP); break;
7686	case X86_XCPT_PF: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestPF); break;
7687	case X86_XCPT_MF: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestMF); break;
7688	case X86_XCPT_AC: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestAC); break;
7689	case X86_XCPT_XF: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestXF); break;
7690	default:
7691	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestXcpUnk);
7692	break;
7693	}
7694	#endif
7695
7696	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
7697	return VINF_SUCCESS;
7698	}
7699	#endif
7700
7701	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
7702	/**
7703	* \#VMEXIT handler for CLGI (SVM_EXIT_CLGI). Conditional \#VMEXIT.
7704	*/
7705	HMSVM_EXIT_DECL hmR0SvmExitClgi(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7706	{
7707	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7708
7709	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7710	Assert(pVmcb);
7711	Assert(!pVmcb->ctrl.IntCtrl.n.u1VGifEnable);
7712
7713	VBOXSTRICTRC rcStrict;
7714	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
7715	uint64_t const fImport = CPUMCTX_EXTRN_HWVIRT;
7716	if (fSupportsNextRipSave)
7717	{
7718	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK \| fImport);
7719	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
7720	rcStrict = IEMExecDecodedClgi(pVCpu, cbInstr);
7721	}
7722	else
7723	{
7724	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK \| fImport);
7725	rcStrict = IEMExecOne(pVCpu);
7726	}
7727
7728	if (rcStrict == VINF_SUCCESS)
7729	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_HWVIRT);
7730	else if (rcStrict == VINF_IEM_RAISED_XCPT)
7731	{
7732	rcStrict = VINF_SUCCESS;
7733	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
7734	}
7735	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
7736	return rcStrict;
7737	}
7738
7739
7740	/**
7741	* \#VMEXIT handler for STGI (SVM_EXIT_STGI). Conditional \#VMEXIT.
7742	*/
7743	HMSVM_EXIT_DECL hmR0SvmExitStgi(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7744	{
7745	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7746
7747	/*
7748	* When VGIF is not used we always intercept STGI instructions. When VGIF is used,
7749	* we only intercept STGI when events are pending for GIF to become 1.
7750	*/
7751	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7752	if (pVmcb->ctrl.IntCtrl.n.u1VGifEnable)
7753	hmR0SvmClearCtrlIntercept(pVCpu, pVmcb, SVM_CTRL_INTERCEPT_STGI);
7754
7755	VBOXSTRICTRC rcStrict;
7756	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
7757	uint64_t const fImport = CPUMCTX_EXTRN_HWVIRT;
7758	if (fSupportsNextRipSave)
7759	{
7760	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK \| fImport);
7761	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
7762	rcStrict = IEMExecDecodedStgi(pVCpu, cbInstr);
7763	}
7764	else
7765	{
7766	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK \| fImport);
7767	rcStrict = IEMExecOne(pVCpu);
7768	}
7769
7770	if (rcStrict == VINF_SUCCESS)
7771	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_HWVIRT);
7772	else if (rcStrict == VINF_IEM_RAISED_XCPT)
7773	{
7774	rcStrict = VINF_SUCCESS;
7775	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
7776	}
7777	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
7778	return rcStrict;
7779	}
7780
7781
7782	/**
7783	* \#VMEXIT handler for VMLOAD (SVM_EXIT_VMLOAD). Conditional \#VMEXIT.
7784	*/
7785	HMSVM_EXIT_DECL hmR0SvmExitVmload(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7786	{
7787	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7788
7789	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7790	Assert(pVmcb);
7791	Assert(!pVmcb->ctrl.LbrVirt.n.u1VirtVmsaveVmload);
7792
7793	VBOXSTRICTRC rcStrict;
7794	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
7795	uint64_t const fImport = CPUMCTX_EXTRN_FS \| CPUMCTX_EXTRN_GS \| CPUMCTX_EXTRN_KERNEL_GS_BASE
7796	\| CPUMCTX_EXTRN_TR \| CPUMCTX_EXTRN_LDTR \| CPUMCTX_EXTRN_SYSCALL_MSRS
7797	\| CPUMCTX_EXTRN_SYSENTER_MSRS;
7798	if (fSupportsNextRipSave)
7799	{
7800	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK \| fImport);
7801	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
7802	rcStrict = IEMExecDecodedVmload(pVCpu, cbInstr);
7803	}
7804	else
7805	{
7806	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK \| fImport);
7807	rcStrict = IEMExecOne(pVCpu);
7808	}
7809
7810	if (rcStrict == VINF_SUCCESS)
7811	{
7812	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_FS \| HM_CHANGED_GUEST_GS
7813	\| HM_CHANGED_GUEST_TR \| HM_CHANGED_GUEST_LDTR
7814	\| HM_CHANGED_GUEST_KERNEL_GS_BASE \| HM_CHANGED_GUEST_SYSCALL_MSRS
7815	\| HM_CHANGED_GUEST_SYSENTER_MSR_MASK);
7816	}
7817	else if (rcStrict == VINF_IEM_RAISED_XCPT)
7818	{
7819	rcStrict = VINF_SUCCESS;
7820	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
7821	}
7822	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
7823	return rcStrict;
7824	}
7825
7826
7827	/**
7828	* \#VMEXIT handler for VMSAVE (SVM_EXIT_VMSAVE). Conditional \#VMEXIT.
7829	*/
7830	HMSVM_EXIT_DECL hmR0SvmExitVmsave(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7831	{
7832	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7833
7834	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7835	Assert(!pVmcb->ctrl.LbrVirt.n.u1VirtVmsaveVmload);
7836
7837	VBOXSTRICTRC rcStrict;
7838	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
7839	if (fSupportsNextRipSave)
7840	{
7841	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK);
7842	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
7843	rcStrict = IEMExecDecodedVmsave(pVCpu, cbInstr);
7844	}
7845	else
7846	{
7847	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
7848	rcStrict = IEMExecOne(pVCpu);
7849	}
7850
7851	if (rcStrict == VINF_IEM_RAISED_XCPT)
7852	{
7853	rcStrict = VINF_SUCCESS;
7854	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
7855	}
7856	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
7857	return rcStrict;
7858	}
7859
7860
7861	/**
7862	* \#VMEXIT handler for INVLPGA (SVM_EXIT_INVLPGA). Conditional \#VMEXIT.
7863	*/
7864	HMSVM_EXIT_DECL hmR0SvmExitInvlpga(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7865	{
7866	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7867
7868	VBOXSTRICTRC rcStrict;
7869	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
7870	if (fSupportsNextRipSave)
7871	{
7872	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK);
7873	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7874	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
7875	rcStrict = IEMExecDecodedInvlpga(pVCpu, cbInstr);
7876	}
7877	else
7878	{
7879	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
7880	rcStrict = IEMExecOne(pVCpu);
7881	}
7882
7883	if (rcStrict == VINF_IEM_RAISED_XCPT)
7884	{
7885	rcStrict = VINF_SUCCESS;
7886	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
7887	}
7888	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
7889	return rcStrict;
7890	}
7891
7892
7893	/**
7894	* \#VMEXIT handler for STGI (SVM_EXIT_VMRUN). Conditional \#VMEXIT.
7895	*/
7896	HMSVM_EXIT_DECL hmR0SvmExitVmrun(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7897	{
7898	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7899	/* We shall import the entire state here, just in case we enter and continue execution of
7900	the nested-guest with hardware-assisted SVM in ring-0, we would be switching VMCBs and
7901	could lose lose part of CPU state. */
7902	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7903
7904	VBOXSTRICTRC rcStrict;
7905	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
7906	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatExitVmentry, z);
7907	if (fSupportsNextRipSave)
7908	{
7909	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7910	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
7911	rcStrict = IEMExecDecodedVmrun(pVCpu, cbInstr);
7912	}
7913	else
7914	{
7915	/* We use IEMExecOneBypassEx() here as it supresses attempt to continue emulating any
7916	instruction(s) when interrupt inhibition is set as part of emulating the VMRUN
7917	instruction itself, see @bugref{7243#c126} */
7918	rcStrict = IEMExecOneBypassEx(pVCpu, CPUMCTX2CORE(&pVCpu->cpum.GstCtx), NULL /* pcbWritten */);
7919	}
7920	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExitVmentry, z);
7921
7922	if (rcStrict == VINF_SUCCESS)
7923	{
7924	rcStrict = VINF_SVM_VMRUN;
7925	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_SVM_VMRUN_MASK);
7926	}
7927	else if (rcStrict == VINF_IEM_RAISED_XCPT)
7928	{
7929	rcStrict = VINF_SUCCESS;
7930	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
7931	}
7932	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
7933	return rcStrict;
7934	}
7935
7936
7937	/**
7938	* Nested-guest \#VMEXIT handler for debug exceptions (SVM_EXIT_XCPT_1).
7939	* Unconditional \#VMEXIT.
7940	*/
7941	HMSVM_EXIT_DECL hmR0SvmNestedExitXcptDB(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7942	{
7943	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7944	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(pVCpu, pSvmTransient);
7945
7946	if (pVCpu->hm.s.Event.fPending)
7947	{
7948	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectInterpret);
7949	return VINF_EM_RAW_INJECT_TRPM_EVENT;
7950	}
7951
7952	hmR0SvmSetPendingXcptDB(pVCpu);
7953	return VINF_SUCCESS;
7954	}
7955
7956
7957	/**
7958	* Nested-guest \#VMEXIT handler for breakpoint exceptions (SVM_EXIT_XCPT_3).
7959	* Conditional \#VMEXIT.
7960	*/
7961	HMSVM_EXIT_DECL hmR0SvmNestedExitXcptBP(PVMCPUCC pVCpu, PSVMTRANSIENT pSvmTransient)
7962	{
7963	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7964	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(pVCpu, pSvmTransient);
7965
7966	SVMEVENT Event;
7967	Event.u = 0;
7968	Event.n.u1Valid = 1;
7969	Event.n.u3Type = SVM_EVENT_EXCEPTION;
7970	Event.n.u8Vector = X86_XCPT_BP;
7971	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
7972	return VINF_SUCCESS;
7973	}
7974	#endif /* VBOX_WITH_NESTED_HWVIRT_SVM */
7975
7976	/** @} */
7977

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

source: vbox/trunk/src/VBox/VMM/VMMR0/HMSVMR0.cpp@ 92583

Download in other formats: