HMSVMR0.cpp@ 73047

Last change on this file since 73047 was 73047, checked in by vboxsync, 6 years ago
VMM/HMSVMR0: bugref:9204 Fix rdtscp.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 314.6 KB

Line
1	/* $Id: HMSVMR0.cpp 73047 2018-07-11 02:42:55Z vboxsync $ */
2	/** @file
3	* HM SVM (AMD-V) - Host Context Ring-0.
4	*/
5
6	/*
7	* Copyright (C) 2013-2017 Oracle Corporation
8	*
9	* This file is part of VirtualBox Open Source Edition (OSE), as
10	* available from http://www.virtualbox.org. This file is free software;
11	* you can redistribute it and/or modify it under the terms of the GNU
12	* General Public License (GPL) as published by the Free Software
13	* Foundation, in version 2 as it comes in the "COPYING" file of the
14	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16	*/
17
18
19	/*********************************************************************************************************************************
20	* Header Files *
21	*********************************************************************************************************************************/
22	#define LOG_GROUP LOG_GROUP_HM
23	#define VMCPU_INCL_CPUM_GST_CTX
24	#include <iprt/asm-amd64-x86.h>
25	#include <iprt/thread.h>
26
27	#include <VBox/vmm/pdmapi.h>
28	#include <VBox/vmm/dbgf.h>
29	#include <VBox/vmm/iem.h>
30	#include <VBox/vmm/iom.h>
31	#include <VBox/vmm/tm.h>
32	#include <VBox/vmm/gim.h>
33	#include <VBox/vmm/apic.h>
34	#include "HMInternal.h"
35	#include <VBox/vmm/vm.h>
36	#include "HMSVMR0.h"
37	#include "dtrace/VBoxVMM.h"
38
39	#ifdef DEBUG_ramshankar
40	# define HMSVM_SYNC_FULL_GUEST_STATE
41	# define HMSVM_ALWAYS_TRAP_ALL_XCPTS
42	# define HMSVM_ALWAYS_TRAP_PF
43	# define HMSVM_ALWAYS_TRAP_TASK_SWITCH
44	#endif
45
46
47	/*********************************************************************************************************************************
48	* Defined Constants And Macros *
49	*********************************************************************************************************************************/
50	#ifdef VBOX_WITH_STATISTICS
51	# define HMSVM_EXITCODE_STAM_COUNTER_INC(u64ExitCode) do { \
52	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitAll); \
53	if ((u64ExitCode) == SVM_EXIT_NPF) \
54	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitReasonNpf); \
55	else \
56	STAM_COUNTER_INC(&pVCpu->hm.s.paStatExitReasonR0[(u64ExitCode) & MASK_EXITREASON_STAT]); \
57	} while (0)
58
59	# ifdef VBOX_WITH_NESTED_HWVIRT_SVM
60	# define HMSVM_NESTED_EXITCODE_STAM_COUNTER_INC(u64ExitCode) do { \
61	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitAll); \
62	if ((u64ExitCode) == SVM_EXIT_NPF) \
63	STAM_COUNTER_INC(&pVCpu->hm.s.StatNestedExitReasonNpf); \
64	else \
65	STAM_COUNTER_INC(&pVCpu->hm.s.paStatNestedExitReasonR0[(u64ExitCode) & MASK_EXITREASON_STAT]); \
66	} while (0)
67	# endif
68	#else
69	# define HMSVM_EXITCODE_STAM_COUNTER_INC(u64ExitCode) do { } while (0)
70	# ifdef VBOX_WITH_NESTED_HWVIRT_SVM
71	# define HMSVM_NESTED_EXITCODE_STAM_COUNTER_INC(u64ExitCode) do { } while (0)
72	# endif
73	#endif /* !VBOX_WITH_STATISTICS */
74
75	/** If we decide to use a function table approach this can be useful to
76	* switch to a "static DECLCALLBACK(int)". */
77	#define HMSVM_EXIT_DECL static int
78
79	/**
80	* Subset of the guest-CPU state that is kept by SVM R0 code while executing the
81	* guest using hardware-assisted SVM.
82	*
83	* This excludes state like TSC AUX, GPRs (other than RSP, RAX) which are always
84	* are swapped and restored across the world-switch and also registers like
85	* EFER, PAT MSR etc. which cannot be modified by the guest without causing a
86	* \#VMEXIT.
87	*/
88	#define HMSVM_CPUMCTX_EXTRN_ALL ( CPUMCTX_EXTRN_RIP \
89	\| CPUMCTX_EXTRN_RFLAGS \
90	\| CPUMCTX_EXTRN_RAX \
91	\| CPUMCTX_EXTRN_RSP \
92	\| CPUMCTX_EXTRN_SREG_MASK \
93	\| CPUMCTX_EXTRN_CR0 \
94	\| CPUMCTX_EXTRN_CR2 \
95	\| CPUMCTX_EXTRN_CR3 \
96	\| CPUMCTX_EXTRN_TABLE_MASK \
97	\| CPUMCTX_EXTRN_DR6 \
98	\| CPUMCTX_EXTRN_DR7 \
99	\| CPUMCTX_EXTRN_KERNEL_GS_BASE \
100	\| CPUMCTX_EXTRN_SYSCALL_MSRS \
101	\| CPUMCTX_EXTRN_SYSENTER_MSRS \
102	\| CPUMCTX_EXTRN_HWVIRT \
103	\| CPUMCTX_EXTRN_HM_SVM_MASK)
104
105	/**
106	* Subset of the guest-CPU state that is shared between the guest and host.
107	*/
108	#define HMSVM_CPUMCTX_SHARED_STATE CPUMCTX_EXTRN_DR_MASK
109
110	/** Macro for importing guest state from the VMCB back into CPUMCTX. */
111	#define HMSVM_CPUMCTX_IMPORT_STATE(a_pVCpu, a_fWhat) \
112	do { \
113	if ((a_pVCpu)->cpum.GstCtx.fExtrn & (a_fWhat)) \
114	hmR0SvmImportGuestState((a_pVCpu), (a_fWhat)); \
115	} while (0)
116
117	/** Assert that the required state bits are fetched. */
118	#define HMSVM_CPUMCTX_ASSERT(a_pVCpu, a_fExtrnMbz) AssertMsg(!((a_pVCpu)->cpum.GstCtx.fExtrn & (a_fExtrnMbz)), \
119	("fExtrn=%#RX64 fExtrnMbz=%#RX64\n", \
120	(a_pVCpu)->cpum.GstCtx.fExtrn, (a_fExtrnMbz)))
121
122	/** Assert that preemption is disabled or covered by thread-context hooks. */
123	#define HMSVM_ASSERT_PREEMPT_SAFE(a_pVCpu) Assert( VMMR0ThreadCtxHookIsEnabled((a_pVCpu)) \
124	\|\| !RTThreadPreemptIsEnabled(NIL_RTTHREAD));
125
126	/** Assert that we haven't migrated CPUs when thread-context hooks are not
127	* used. */
128	#define HMSVM_ASSERT_CPU_SAFE(a_pVCpu) AssertMsg( VMMR0ThreadCtxHookIsEnabled((a_pVCpu)) \
129	\|\| (a_pVCpu)->hm.s.idEnteredCpu == RTMpCpuId(), \
130	("Illegal migration! Entered on CPU %u Current %u\n", \
131	(a_pVCpu)->hm.s.idEnteredCpu, RTMpCpuId()));
132
133	/** Assert that we're not executing a nested-guest. */
134	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
135	# define HMSVM_ASSERT_NOT_IN_NESTED_GUEST(a_pCtx) Assert(!CPUMIsGuestInSvmNestedHwVirtMode((a_pCtx)))
136	#else
137	# define HMSVM_ASSERT_NOT_IN_NESTED_GUEST(a_pCtx) do { NOREF((a_pCtx)); } while (0)
138	#endif
139
140	/** Assert that we're executing a nested-guest. */
141	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
142	# define HMSVM_ASSERT_IN_NESTED_GUEST(a_pCtx) Assert(CPUMIsGuestInSvmNestedHwVirtMode((a_pCtx)))
143	#else
144	# define HMSVM_ASSERT_IN_NESTED_GUEST(a_pCtx) do { NOREF((a_pCtx)); } while (0)
145	#endif
146
147	/** Macro for checking and returning from the using function for
148	* \#VMEXIT intercepts that maybe caused during delivering of another
149	* event in the guest. */
150	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
151	# define HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(a_pVCpu, a_pSvmTransient) \
152	do \
153	{ \
154	int rc = hmR0SvmCheckExitDueToEventDelivery((a_pVCpu), (a_pSvmTransient)); \
155	if (RT_LIKELY(rc == VINF_SUCCESS)) { /* continue #VMEXIT handling */ } \
156	else if ( rc == VINF_HM_DOUBLE_FAULT) { return VINF_SUCCESS; } \
157	else if ( rc == VINF_EM_RESET \
158	&& CPUMIsGuestSvmCtrlInterceptSet((a_pVCpu), &(a_pVCpu)->cpum.GstCtx, SVM_CTRL_INTERCEPT_SHUTDOWN)) \
159	{ \
160	HMSVM_CPUMCTX_IMPORT_STATE((a_pVCpu), IEM_CPUMCTX_EXTRN_SVM_VMEXIT_MASK); \
161	return VBOXSTRICTRC_TODO(IEMExecSvmVmexit((a_pVCpu), SVM_EXIT_SHUTDOWN, 0, 0)); \
162	} \
163	else \
164	return rc; \
165	} while (0)
166	#else
167	# define HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(a_pVCpu, a_pSvmTransient) \
168	do \
169	{ \
170	int rc = hmR0SvmCheckExitDueToEventDelivery((a_pVCpu), (a_pSvmTransient)); \
171	if (RT_LIKELY(rc == VINF_SUCCESS)) { /* continue #VMEXIT handling */ } \
172	else if ( rc == VINF_HM_DOUBLE_FAULT) { return VINF_SUCCESS; } \
173	else \
174	return rc; \
175	} while (0)
176	#endif
177
178	/** Macro which updates interrupt shadow for the current RIP. */
179	#define HMSVM_UPDATE_INTR_SHADOW(a_pVCpu) \
180	do { \
181	/* Update interrupt shadow. */ \
182	if ( VMCPU_FF_IS_PENDING((a_pVCpu), VMCPU_FF_INHIBIT_INTERRUPTS) \
183	&& (a_pVCpu)->cpum.GstCtx.rip != EMGetInhibitInterruptsPC((a_pVCpu))) \
184	VMCPU_FF_CLEAR((a_pVCpu), VMCPU_FF_INHIBIT_INTERRUPTS); \
185	} while (0)
186
187	/** Macro for upgrading a @a a_rc to VINF_EM_DBG_STEPPED after emulating an
188	* instruction that exited. */
189	#define HMSVM_CHECK_SINGLE_STEP(a_pVCpu, a_rc) \
190	do { \
191	if ((a_pVCpu)->hm.s.fSingleInstruction && (a_rc) == VINF_SUCCESS) \
192	(a_rc) = VINF_EM_DBG_STEPPED; \
193	} while (0)
194
195	/** Validate segment descriptor granularity bit. */
196	#ifdef VBOX_STRICT
197	# define HMSVM_ASSERT_SEG_GRANULARITY(a_pCtx, reg) \
198	AssertMsg( !(a_pCtx)->reg.Attr.n.u1Present \
199	\|\| ( (a_pCtx)->reg.Attr.n.u1Granularity \
200	? ((a_pCtx)->reg.u32Limit & 0xfff) == 0xfff \
201	: (a_pCtx)->reg.u32Limit <= UINT32_C(0xfffff)), \
202	("Invalid Segment Attributes Limit=%#RX32 Attr=%#RX32 Base=%#RX64\n", (a_pCtx)->reg.u32Limit, \
203	(a_pCtx)->reg.Attr.u, (a_pCtx)->reg.u64Base))
204	#else
205	# define HMSVM_ASSERT_SEG_GRANULARITY(a_pCtx, reg) do { } while (0)
206	#endif
207
208	/**
209	* Exception bitmap mask for all contributory exceptions.
210	*
211	* Page fault is deliberately excluded here as it's conditional as to whether
212	* it's contributory or benign. Page faults are handled separately.
213	*/
214	#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) \
215	\| RT_BIT(X86_XCPT_DE))
216
217	/**
218	* Mandatory/unconditional guest control intercepts.
219	*
220	* SMIs can and do happen in normal operation. We need not intercept them
221	* while executing the guest (or nested-guest).
222	*/
223	#define HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS ( SVM_CTRL_INTERCEPT_INTR \
224	\| SVM_CTRL_INTERCEPT_NMI \
225	\| SVM_CTRL_INTERCEPT_INIT \
226	\| SVM_CTRL_INTERCEPT_RDPMC \
227	\| SVM_CTRL_INTERCEPT_CPUID \
228	\| SVM_CTRL_INTERCEPT_RSM \
229	\| SVM_CTRL_INTERCEPT_HLT \
230	\| SVM_CTRL_INTERCEPT_IOIO_PROT \
231	\| SVM_CTRL_INTERCEPT_MSR_PROT \
232	\| SVM_CTRL_INTERCEPT_INVLPGA \
233	\| SVM_CTRL_INTERCEPT_SHUTDOWN \
234	\| SVM_CTRL_INTERCEPT_FERR_FREEZE \
235	\| SVM_CTRL_INTERCEPT_VMRUN \
236	\| SVM_CTRL_INTERCEPT_SKINIT \
237	\| SVM_CTRL_INTERCEPT_WBINVD \
238	\| SVM_CTRL_INTERCEPT_MONITOR \
239	\| SVM_CTRL_INTERCEPT_MWAIT \
240	\| SVM_CTRL_INTERCEPT_CR0_SEL_WRITE \
241	\| SVM_CTRL_INTERCEPT_XSETBV)
242
243	/** @name VMCB Clean Bits.
244	*
245	* These flags are used for VMCB-state caching. A set VMCB Clean bit indicates
246	* AMD-V doesn't need to reload the corresponding value(s) from the VMCB in
247	* memory.
248	*
249	* @{ */
250	/** All intercepts vectors, TSC offset, PAUSE filter counter. */
251	#define HMSVM_VMCB_CLEAN_INTERCEPTS RT_BIT(0)
252	/** I/O permission bitmap, MSR permission bitmap. */
253	#define HMSVM_VMCB_CLEAN_IOPM_MSRPM RT_BIT(1)
254	/** ASID. */
255	#define HMSVM_VMCB_CLEAN_ASID RT_BIT(2)
256	/** TRP: V_TPR, V_IRQ, V_INTR_PRIO, V_IGN_TPR, V_INTR_MASKING,
257	V_INTR_VECTOR. */
258	#define HMSVM_VMCB_CLEAN_INT_CTRL RT_BIT(3)
259	/** Nested Paging: Nested CR3 (nCR3), PAT. */
260	#define HMSVM_VMCB_CLEAN_NP RT_BIT(4)
261	/** Control registers (CR0, CR3, CR4, EFER). */
262	#define HMSVM_VMCB_CLEAN_CRX_EFER RT_BIT(5)
263	/** Debug registers (DR6, DR7). */
264	#define HMSVM_VMCB_CLEAN_DRX RT_BIT(6)
265	/** GDT, IDT limit and base. */
266	#define HMSVM_VMCB_CLEAN_DT RT_BIT(7)
267	/** Segment register: CS, SS, DS, ES limit and base. */
268	#define HMSVM_VMCB_CLEAN_SEG RT_BIT(8)
269	/** CR2.*/
270	#define HMSVM_VMCB_CLEAN_CR2 RT_BIT(9)
271	/** Last-branch record (DbgCtlMsr, br_from, br_to, lastint_from, lastint_to) */
272	#define HMSVM_VMCB_CLEAN_LBR RT_BIT(10)
273	/** AVIC (AVIC APIC_BAR; AVIC APIC_BACKING_PAGE, AVIC
274	PHYSICAL_TABLE and AVIC LOGICAL_TABLE Pointers). */
275	#define HMSVM_VMCB_CLEAN_AVIC RT_BIT(11)
276	/** Mask of all valid VMCB Clean bits. */
277	#define HMSVM_VMCB_CLEAN_ALL ( HMSVM_VMCB_CLEAN_INTERCEPTS \
278	\| HMSVM_VMCB_CLEAN_IOPM_MSRPM \
279	\| HMSVM_VMCB_CLEAN_ASID \
280	\| HMSVM_VMCB_CLEAN_INT_CTRL \
281	\| HMSVM_VMCB_CLEAN_NP \
282	\| HMSVM_VMCB_CLEAN_CRX_EFER \
283	\| HMSVM_VMCB_CLEAN_DRX \
284	\| HMSVM_VMCB_CLEAN_DT \
285	\| HMSVM_VMCB_CLEAN_SEG \
286	\| HMSVM_VMCB_CLEAN_CR2 \
287	\| HMSVM_VMCB_CLEAN_LBR \
288	\| HMSVM_VMCB_CLEAN_AVIC)
289	/** @} */
290
291	/** @name SVM transient.
292	*
293	* A state structure for holding miscellaneous information across AMD-V
294	* VMRUN/\#VMEXIT operation, restored after the transition.
295	*
296	* @{ */
297	typedef struct SVMTRANSIENT
298	{
299	/** The host's rflags/eflags. */
300	RTCCUINTREG fEFlags;
301	#if HC_ARCH_BITS == 32
302	uint32_t u32Alignment0;
303	#endif
304
305	/** The \#VMEXIT exit code (the EXITCODE field in the VMCB). */
306	uint64_t u64ExitCode;
307	/** The guest's TPR value used for TPR shadowing. */
308	uint8_t u8GuestTpr;
309	/** Alignment. */
310	uint8_t abAlignment0[7];
311
312	/** Pointer to the currently executing VMCB. */
313	PSVMVMCB pVmcb;
314	/** Whether we are currently executing a nested-guest. */
315	bool fIsNestedGuest;
316
317	/** Whether the guest debug state was active at the time of \#VMEXIT. */
318	bool fWasGuestDebugStateActive;
319	/** Whether the hyper debug state was active at the time of \#VMEXIT. */
320	bool fWasHyperDebugStateActive;
321	/** Whether the TSC offset mode needs to be updated. */
322	bool fUpdateTscOffsetting;
323	/** Whether the TSC_AUX MSR needs restoring on \#VMEXIT. */
324	bool fRestoreTscAuxMsr;
325	/** Whether the \#VMEXIT was caused by a page-fault during delivery of a
326	* contributary exception or a page-fault. */
327	bool fVectoringDoublePF;
328	/** Whether the \#VMEXIT was caused by a page-fault during delivery of an
329	* external interrupt or NMI. */
330	bool fVectoringPF;
331	} SVMTRANSIENT, *PSVMTRANSIENT;
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 VBox status code.
362	* @param pVCpu The cross context virtual CPU structure.
363	* @param pSvmTransient Pointer to the SVM-transient structure.
364	*/
365	typedef int FNSVMEXITHANDLER(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient);
366
367
368	/*********************************************************************************************************************************
369	* Internal Functions *
370	*********************************************************************************************************************************/
371	static void hmR0SvmPendingEventToTrpmTrap(PVMCPU pVCpu);
372	static void hmR0SvmLeave(PVMCPU 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	#if defined(HMSVM_ALWAYS_TRAP_ALL_XCPTS) \|\| defined(VBOX_WITH_NESTED_HWVIRT_SVM)
412	static FNSVMEXITHANDLER hmR0SvmExitXcptGeneric;
413	#endif
414	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
415	static FNSVMEXITHANDLER hmR0SvmExitClgi;
416	static FNSVMEXITHANDLER hmR0SvmExitStgi;
417	static FNSVMEXITHANDLER hmR0SvmExitVmload;
418	static FNSVMEXITHANDLER hmR0SvmExitVmsave;
419	static FNSVMEXITHANDLER hmR0SvmExitInvlpga;
420	static FNSVMEXITHANDLER hmR0SvmExitVmrun;
421	static FNSVMEXITHANDLER hmR0SvmNestedExitXcptDB;
422	static FNSVMEXITHANDLER hmR0SvmNestedExitXcptBP;
423	#endif
424	/** @} */
425
426	static int hmR0SvmHandleExit(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient);
427	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
428	static int hmR0SvmHandleExitNested(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient);
429	#endif
430
431
432	/*********************************************************************************************************************************
433	* Global Variables *
434	*********************************************************************************************************************************/
435	/** Ring-0 memory object for the IO bitmap. */
436	static RTR0MEMOBJ g_hMemObjIOBitmap = NIL_RTR0MEMOBJ;
437	/** Physical address of the IO bitmap. */
438	static RTHCPHYS g_HCPhysIOBitmap;
439	/** Pointer to the IO bitmap. */
440	static R0PTRTYPE(void *) g_pvIOBitmap;
441
442	#ifdef VBOX_STRICT
443	# define HMSVM_LOG_RBP_RSP RT_BIT_32(0)
444	# define HMSVM_LOG_CR_REGS RT_BIT_32(1)
445	# define HMSVM_LOG_CS RT_BIT_32(2)
446	# define HMSVM_LOG_SS RT_BIT_32(3)
447	# define HMSVM_LOG_FS RT_BIT_32(4)
448	# define HMSVM_LOG_GS RT_BIT_32(5)
449	# define HMSVM_LOG_LBR RT_BIT_32(6)
450	# define HMSVM_LOG_ALL ( HMSVM_LOG_RBP_RSP \
451	\| HMSVM_LOG_CR_REGS \
452	\| HMSVM_LOG_CS \
453	\| HMSVM_LOG_SS \
454	\| HMSVM_LOG_FS \
455	\| HMSVM_LOG_GS \
456	\| HMSVM_LOG_LBR)
457
458	/**
459	* Dumps virtual CPU state and additional info. to the logger for diagnostics.
460	*
461	* @param pVCpu The cross context virtual CPU structure.
462	* @param pVmcb Pointer to the VM control block.
463	* @param pszPrefix Log prefix.
464	* @param fFlags Log flags, see HMSVM_LOG_XXX.
465	* @param uVerbose The verbosity level, currently unused.
466	*/
467	static void hmR0SvmLogState(PVMCPU pVCpu, PCSVMVMCB pVmcb, const char *pszPrefix, uint32_t fFlags, uint8_t uVerbose)
468	{
469	RT_NOREF2(pVCpu, uVerbose);
470	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
471
472	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_CS \| CPUMCTX_EXTRN_RIP \| CPUMCTX_EXTRN_RFLAGS);
473	Log4(("%s: cs:rip=%04x:%RX64 efl=%#RX64\n", pszPrefix, pCtx->cs.Sel, pCtx->rip, pCtx->rflags.u));
474
475	if (fFlags & HMSVM_LOG_RBP_RSP)
476	{
477	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_RSP \| CPUMCTX_EXTRN_RBP);
478	Log4(("%s: rsp=%#RX64 rbp=%#RX64\n", pszPrefix, pCtx->rsp, pCtx->rbp));
479	}
480
481	if (fFlags & HMSVM_LOG_CR_REGS)
482	{
483	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_CR0 \| CPUMCTX_EXTRN_CR3 \| CPUMCTX_EXTRN_CR4);
484	Log4(("%s: cr0=%#RX64 cr3=%#RX64 cr4=%#RX64\n", pszPrefix, pCtx->cr0, pCtx->cr3, pCtx->cr4));
485	}
486
487	if (fFlags & HMSVM_LOG_CS)
488	{
489	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_CS);
490	Log4(("%s: cs={%04x base=%016RX64 limit=%08x flags=%08x}\n", pszPrefix, pCtx->cs.Sel, pCtx->cs.u64Base,
491	pCtx->cs.u32Limit, pCtx->cs.Attr.u));
492	}
493	if (fFlags & HMSVM_LOG_SS)
494	{
495	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_SS);
496	Log4(("%s: ss={%04x base=%016RX64 limit=%08x flags=%08x}\n", pszPrefix, pCtx->ss.Sel, pCtx->ss.u64Base,
497	pCtx->ss.u32Limit, pCtx->ss.Attr.u));
498	}
499	if (fFlags & HMSVM_LOG_FS)
500	{
501	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_FS);
502	Log4(("%s: fs={%04x base=%016RX64 limit=%08x flags=%08x}\n", pszPrefix, pCtx->fs.Sel, pCtx->fs.u64Base,
503	pCtx->fs.u32Limit, pCtx->fs.Attr.u));
504	}
505	if (fFlags & HMSVM_LOG_GS)
506	{
507	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_GS);
508	Log4(("%s: gs={%04x base=%016RX64 limit=%08x flags=%08x}\n", pszPrefix, pCtx->gs.Sel, pCtx->gs.u64Base,
509	pCtx->gs.u32Limit, pCtx->gs.Attr.u));
510	}
511
512	PCSVMVMCBSTATESAVE pVmcbGuest = &pVmcb->guest;
513	if (fFlags & HMSVM_LOG_LBR)
514	{
515	Log4(("%s: br_from=%#RX64 br_to=%#RX64 lastxcpt_from=%#RX64 lastxcpt_to=%#RX64\n", pszPrefix, pVmcbGuest->u64BR_FROM,
516	pVmcbGuest->u64BR_TO, pVmcbGuest->u64LASTEXCPFROM, pVmcbGuest->u64LASTEXCPTO));
517	}
518	NOREF(pVmcbGuest); NOREF(pCtx);
519	}
520	#endif /* VBOX_STRICT */
521
522
523	/**
524	* Sets up and activates AMD-V on the current CPU.
525	*
526	* @returns VBox status code.
527	* @param pHostCpu Pointer to the CPU info struct.
528	* @param pVM The cross context VM structure. Can be
529	* NULL after a resume!
530	* @param pvCpuPage Pointer to the global CPU page.
531	* @param HCPhysCpuPage Physical address of the global CPU page.
532	* @param fEnabledByHost Whether the host OS has already initialized AMD-V.
533	* @param pvArg Unused on AMD-V.
534	*/
535	VMMR0DECL(int) SVMR0EnableCpu(PHMGLOBALCPUINFO pHostCpu, PVM pVM, void *pvCpuPage, RTHCPHYS HCPhysCpuPage, bool fEnabledByHost,
536	void *pvArg)
537	{
538	Assert(!fEnabledByHost);
539	Assert(HCPhysCpuPage && HCPhysCpuPage != NIL_RTHCPHYS);
540	Assert(RT_ALIGN_T(HCPhysCpuPage, _4K, RTHCPHYS) == HCPhysCpuPage);
541	Assert(pvCpuPage); NOREF(pvCpuPage);
542	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
543
544	NOREF(pvArg);
545	NOREF(fEnabledByHost);
546
547	/* Paranoid: Disable interrupt as, in theory, interrupt handlers might mess with EFER. */
548	RTCCUINTREG const fEFlags = ASMIntDisableFlags();
549
550	/*
551	* We must turn on AMD-V and setup the host state physical address, as those MSRs are per CPU.
552	*/
553	uint64_t u64HostEfer = ASMRdMsr(MSR_K6_EFER);
554	if (u64HostEfer & MSR_K6_EFER_SVME)
555	{
556	/* If the VBOX_HWVIRTEX_IGNORE_SVM_IN_USE is active, then we blindly use AMD-V. */
557	if ( pVM
558	&& pVM->hm.s.svm.fIgnoreInUseError)
559	pHostCpu->fIgnoreAMDVInUseError = true;
560
561	if (!pHostCpu->fIgnoreAMDVInUseError)
562	{
563	ASMSetFlags(fEFlags);
564	return VERR_SVM_IN_USE;
565	}
566	}
567
568	/* Turn on AMD-V in the EFER MSR. */
569	ASMWrMsr(MSR_K6_EFER, u64HostEfer \| MSR_K6_EFER_SVME);
570
571	/* Write the physical page address where the CPU will store the host state while executing the VM. */
572	ASMWrMsr(MSR_K8_VM_HSAVE_PA, HCPhysCpuPage);
573
574	/* Restore interrupts. */
575	ASMSetFlags(fEFlags);
576
577	/*
578	* Theoretically, other hypervisors may have used ASIDs, ideally we should flush all
579	* non-zero ASIDs when enabling SVM. AMD doesn't have an SVM instruction to flush all
580	* ASIDs (flushing is done upon VMRUN). Therefore, flag that we need to flush the TLB
581	* entirely with before executing any guest code.
582	*/
583	pHostCpu->fFlushAsidBeforeUse = true;
584
585	/*
586	* Ensure each VCPU scheduled on this CPU gets a new ASID on resume. See @bugref{6255}.
587	*/
588	++pHostCpu->cTlbFlushes;
589
590	return VINF_SUCCESS;
591	}
592
593
594	/**
595	* Deactivates AMD-V on the current CPU.
596	*
597	* @returns VBox status code.
598	* @param pHostCpu Pointer to the CPU info struct.
599	* @param pvCpuPage Pointer to the global CPU page.
600	* @param HCPhysCpuPage Physical address of the global CPU page.
601	*/
602	VMMR0DECL(int) SVMR0DisableCpu(PHMGLOBALCPUINFO pHostCpu, void *pvCpuPage, RTHCPHYS HCPhysCpuPage)
603	{
604	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
605	AssertReturn( HCPhysCpuPage
606	&& HCPhysCpuPage != NIL_RTHCPHYS, VERR_INVALID_PARAMETER);
607	AssertReturn(pvCpuPage, VERR_INVALID_PARAMETER);
608	RT_NOREF(pHostCpu);
609
610	/* Paranoid: Disable interrupts as, in theory, interrupt handlers might mess with EFER. */
611	RTCCUINTREG const fEFlags = ASMIntDisableFlags();
612
613	/* Turn off AMD-V in the EFER MSR. */
614	uint64_t u64HostEfer = ASMRdMsr(MSR_K6_EFER);
615	ASMWrMsr(MSR_K6_EFER, u64HostEfer & ~MSR_K6_EFER_SVME);
616
617	/* Invalidate host state physical address. */
618	ASMWrMsr(MSR_K8_VM_HSAVE_PA, 0);
619
620	/* Restore interrupts. */
621	ASMSetFlags(fEFlags);
622
623	return VINF_SUCCESS;
624	}
625
626
627	/**
628	* Does global AMD-V initialization (called during module initialization).
629	*
630	* @returns VBox status code.
631	*/
632	VMMR0DECL(int) SVMR0GlobalInit(void)
633	{
634	/*
635	* Allocate 12 KB (3 pages) for the IO bitmap. Since this is non-optional and we always
636	* intercept all IO accesses, it's done once globally here instead of per-VM.
637	*/
638	Assert(g_hMemObjIOBitmap == NIL_RTR0MEMOBJ);
639	int rc = RTR0MemObjAllocCont(&g_hMemObjIOBitmap, SVM_IOPM_PAGES << X86_PAGE_4K_SHIFT, false /* fExecutable */);
640	if (RT_FAILURE(rc))
641	return rc;
642
643	g_pvIOBitmap = RTR0MemObjAddress(g_hMemObjIOBitmap);
644	g_HCPhysIOBitmap = RTR0MemObjGetPagePhysAddr(g_hMemObjIOBitmap, 0 /* iPage */);
645
646	/* Set all bits to intercept all IO accesses. */
647	ASMMemFill32(g_pvIOBitmap, SVM_IOPM_PAGES << X86_PAGE_4K_SHIFT, UINT32_C(0xffffffff));
648
649	return VINF_SUCCESS;
650	}
651
652
653	/**
654	* Does global AMD-V termination (called during module termination).
655	*/
656	VMMR0DECL(void) SVMR0GlobalTerm(void)
657	{
658	if (g_hMemObjIOBitmap != NIL_RTR0MEMOBJ)
659	{
660	RTR0MemObjFree(g_hMemObjIOBitmap, true /* fFreeMappings */);
661	g_pvIOBitmap = NULL;
662	g_HCPhysIOBitmap = 0;
663	g_hMemObjIOBitmap = NIL_RTR0MEMOBJ;
664	}
665	}
666
667
668	/**
669	* Frees any allocated per-VCPU structures for a VM.
670	*
671	* @param pVM The cross context VM structure.
672	*/
673	DECLINLINE(void) hmR0SvmFreeStructs(PVM pVM)
674	{
675	for (uint32_t i = 0; i < pVM->cCpus; i++)
676	{
677	PVMCPU pVCpu = &pVM->aCpus[i];
678	AssertPtr(pVCpu);
679
680	if (pVCpu->hm.s.svm.hMemObjVmcbHost != NIL_RTR0MEMOBJ)
681	{
682	RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjVmcbHost, false);
683	pVCpu->hm.s.svm.HCPhysVmcbHost = 0;
684	pVCpu->hm.s.svm.hMemObjVmcbHost = NIL_RTR0MEMOBJ;
685	}
686
687	if (pVCpu->hm.s.svm.hMemObjVmcb != NIL_RTR0MEMOBJ)
688	{
689	RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjVmcb, false);
690	pVCpu->hm.s.svm.pVmcb = NULL;
691	pVCpu->hm.s.svm.HCPhysVmcb = 0;
692	pVCpu->hm.s.svm.hMemObjVmcb = NIL_RTR0MEMOBJ;
693	}
694
695	if (pVCpu->hm.s.svm.hMemObjMsrBitmap != NIL_RTR0MEMOBJ)
696	{
697	RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjMsrBitmap, false);
698	pVCpu->hm.s.svm.pvMsrBitmap = NULL;
699	pVCpu->hm.s.svm.HCPhysMsrBitmap = 0;
700	pVCpu->hm.s.svm.hMemObjMsrBitmap = NIL_RTR0MEMOBJ;
701	}
702	}
703	}
704
705
706	/**
707	* Does per-VM AMD-V initialization.
708	*
709	* @returns VBox status code.
710	* @param pVM The cross context VM structure.
711	*/
712	VMMR0DECL(int) SVMR0InitVM(PVM pVM)
713	{
714	int rc = VERR_INTERNAL_ERROR_5;
715
716	/*
717	* Check for an AMD CPU erratum which requires us to flush the TLB before every world-switch.
718	*/
719	uint32_t u32Family;
720	uint32_t u32Model;
721	uint32_t u32Stepping;
722	if (HMAmdIsSubjectToErratum170(&u32Family, &u32Model, &u32Stepping))
723	{
724	Log4Func(("AMD cpu with erratum 170 family %#x model %#x stepping %#x\n", u32Family, u32Model, u32Stepping));
725	pVM->hm.s.svm.fAlwaysFlushTLB = true;
726	}
727
728	/*
729	* Initialize the R0 memory objects up-front so we can properly cleanup on allocation failures.
730	*/
731	for (VMCPUID i = 0; i < pVM->cCpus; i++)
732	{
733	PVMCPU pVCpu = &pVM->aCpus[i];
734	pVCpu->hm.s.svm.hMemObjVmcbHost = NIL_RTR0MEMOBJ;
735	pVCpu->hm.s.svm.hMemObjVmcb = NIL_RTR0MEMOBJ;
736	pVCpu->hm.s.svm.hMemObjMsrBitmap = NIL_RTR0MEMOBJ;
737	}
738
739	for (VMCPUID i = 0; i < pVM->cCpus; i++)
740	{
741	PVMCPU pVCpu = &pVM->aCpus[i];
742
743	/*
744	* Allocate one page for the host-context VM control block (VMCB). This is used for additional host-state (such as
745	* FS, GS, Kernel GS Base, etc.) apart from the host-state save area specified in MSR_K8_VM_HSAVE_PA.
746	*/
747	rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjVmcbHost, SVM_VMCB_PAGES << PAGE_SHIFT, false /* fExecutable */);
748	if (RT_FAILURE(rc))
749	goto failure_cleanup;
750
751	void *pvVmcbHost = RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjVmcbHost);
752	pVCpu->hm.s.svm.HCPhysVmcbHost = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjVmcbHost, 0 /* iPage */);
753	Assert(pVCpu->hm.s.svm.HCPhysVmcbHost < _4G);
754	ASMMemZeroPage(pvVmcbHost);
755
756	/*
757	* Allocate one page for the guest-state VMCB.
758	*/
759	rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjVmcb, SVM_VMCB_PAGES << PAGE_SHIFT, false /* fExecutable */);
760	if (RT_FAILURE(rc))
761	goto failure_cleanup;
762
763	pVCpu->hm.s.svm.pVmcb = (PSVMVMCB)RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjVmcb);
764	pVCpu->hm.s.svm.HCPhysVmcb = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjVmcb, 0 /* iPage */);
765	Assert(pVCpu->hm.s.svm.HCPhysVmcb < _4G);
766	ASMMemZeroPage(pVCpu->hm.s.svm.pVmcb);
767
768	/*
769	* Allocate two pages (8 KB) for the MSR permission bitmap. There doesn't seem to be a way to convince
770	* SVM to not require one.
771	*/
772	rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjMsrBitmap, SVM_MSRPM_PAGES << X86_PAGE_4K_SHIFT,
773	false /* fExecutable */);
774	if (RT_FAILURE(rc))
775	goto failure_cleanup;
776
777	pVCpu->hm.s.svm.pvMsrBitmap = RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjMsrBitmap);
778	pVCpu->hm.s.svm.HCPhysMsrBitmap = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjMsrBitmap, 0 /* iPage */);
779	/* Set all bits to intercept all MSR accesses (changed later on). */
780	ASMMemFill32(pVCpu->hm.s.svm.pvMsrBitmap, SVM_MSRPM_PAGES << X86_PAGE_4K_SHIFT, UINT32_C(0xffffffff));
781	}
782
783	return VINF_SUCCESS;
784
785	failure_cleanup:
786	hmR0SvmFreeStructs(pVM);
787	return rc;
788	}
789
790
791	/**
792	* Does per-VM AMD-V termination.
793	*
794	* @returns VBox status code.
795	* @param pVM The cross context VM structure.
796	*/
797	VMMR0DECL(int) SVMR0TermVM(PVM pVM)
798	{
799	hmR0SvmFreeStructs(pVM);
800	return VINF_SUCCESS;
801	}
802
803
804	/**
805	* Returns whether the VMCB Clean Bits feature is supported.
806	*
807	* @return @c true if supported, @c false otherwise.
808	* @param pVCpu The cross context virtual CPU structure.
809	*/
810	DECLINLINE(bool) hmR0SvmSupportsVmcbCleanBits(PVMCPU pVCpu)
811	{
812	PVM pVM = pVCpu->CTX_SUFF(pVM);
813	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
814	if (CPUMIsGuestInSvmNestedHwVirtMode(&pVCpu->cpum.GstCtx))
815	{
816	return (pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_VMCB_CLEAN)
817	&& pVM->cpum.ro.GuestFeatures.fSvmVmcbClean;
818	}
819	#endif
820	return RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_VMCB_CLEAN);
821	}
822
823
824	/**
825	* Returns whether the decode assists feature is supported.
826	*
827	* @return @c true if supported, @c false otherwise.
828	* @param pVCpu The cross context virtual CPU structure.
829	*/
830	DECLINLINE(bool) hmR0SvmSupportsDecodeAssists(PVMCPU pVCpu)
831	{
832	PVM pVM = pVCpu->CTX_SUFF(pVM);
833	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
834	if (CPUMIsGuestInSvmNestedHwVirtMode(&pVCpu->cpum.GstCtx))
835	{
836	return (pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_DECODE_ASSISTS)
837	&& pVM->cpum.ro.GuestFeatures.fSvmDecodeAssists;
838	}
839	#endif
840	return RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_DECODE_ASSISTS);
841	}
842
843
844	/**
845	* Returns whether the NRIP_SAVE feature is supported.
846	*
847	* @return @c true if supported, @c false otherwise.
848	* @param pVCpu The cross context virtual CPU structure.
849	*/
850	DECLINLINE(bool) hmR0SvmSupportsNextRipSave(PVMCPU pVCpu)
851	{
852	#if 0
853	PVM pVM = pVCpu->CTX_SUFF(pVM);
854	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
855	if (CPUMIsGuestInSvmNestedHwVirtMode(&pVCpu->cpum.GstCtx))
856	{
857	return (pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_NRIP_SAVE)
858	&& pVM->cpum.ro.GuestFeatures.fSvmNextRipSave;
859	}
860	#endif
861	return RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_NRIP_SAVE);
862	#endif
863
864	/** @todo Temporarily disabled NRIP_SAVE for testing. re-enable once its working. */
865	NOREF(pVCpu);
866	return false;
867	}
868
869
870	/**
871	* Sets the permission bits for the specified MSR in the MSRPM bitmap.
872	*
873	* @param pVCpu The cross context virtual CPU structure.
874	* @param pbMsrBitmap Pointer to the MSR bitmap.
875	* @param idMsr The MSR for which the permissions are being set.
876	* @param enmRead MSR read permissions.
877	* @param enmWrite MSR write permissions.
878	*
879	* @remarks This function does -not- clear the VMCB clean bits for MSRPM. The
880	* caller needs to take care of this.
881	*/
882	static void hmR0SvmSetMsrPermission(PVMCPU pVCpu, uint8_t *pbMsrBitmap, uint32_t idMsr, SVMMSREXITREAD enmRead,
883	SVMMSREXITWRITE enmWrite)
884	{
885	bool const fInNestedGuestMode = CPUMIsGuestInSvmNestedHwVirtMode(&pVCpu->cpum.GstCtx);
886	uint16_t offMsrpm;
887	uint8_t uMsrpmBit;
888	int rc = HMSvmGetMsrpmOffsetAndBit(idMsr, &offMsrpm, &uMsrpmBit);
889	AssertRC(rc);
890
891	Assert(uMsrpmBit == 0 \|\| uMsrpmBit == 2 \|\| uMsrpmBit == 4 \|\| uMsrpmBit == 6);
892	Assert(offMsrpm < SVM_MSRPM_PAGES << X86_PAGE_4K_SHIFT);
893
894	pbMsrBitmap += offMsrpm;
895	if (enmRead == SVMMSREXIT_INTERCEPT_READ)
896	*pbMsrBitmap \|= RT_BIT(uMsrpmBit);
897	else
898	{
899	if (!fInNestedGuestMode)
900	*pbMsrBitmap &= ~RT_BIT(uMsrpmBit);
901	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
902	else
903	{
904	/* Only clear the bit if the nested-guest is also not intercepting the MSR read.*/
905	uint8_t const pbNstGstMsrBitmap = (uint8_t )pVCpu->cpum.GstCtx.hwvirt.svm.CTX_SUFF(pvMsrBitmap);
906	pbNstGstMsrBitmap += offMsrpm;
907	if (!(*pbNstGstMsrBitmap & RT_BIT(uMsrpmBit)))
908	*pbMsrBitmap &= ~RT_BIT(uMsrpmBit);
909	else
910	Assert(*pbMsrBitmap & RT_BIT(uMsrpmBit));
911	}
912	#endif
913	}
914
915	if (enmWrite == SVMMSREXIT_INTERCEPT_WRITE)
916	*pbMsrBitmap \|= RT_BIT(uMsrpmBit + 1);
917	else
918	{
919	if (!fInNestedGuestMode)
920	*pbMsrBitmap &= ~RT_BIT(uMsrpmBit + 1);
921	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
922	else
923	{
924	/* Only clear the bit if the nested-guest is also not intercepting the MSR write.*/
925	uint8_t const pbNstGstMsrBitmap = (uint8_t )pVCpu->cpum.GstCtx.hwvirt.svm.CTX_SUFF(pvMsrBitmap);
926	pbNstGstMsrBitmap += offMsrpm;
927	if (!(*pbNstGstMsrBitmap & RT_BIT(uMsrpmBit + 1)))
928	*pbMsrBitmap &= ~RT_BIT(uMsrpmBit + 1);
929	else
930	Assert(*pbMsrBitmap & RT_BIT(uMsrpmBit + 1));
931	}
932	#endif
933	}
934	}
935
936
937	/**
938	* Sets up AMD-V for the specified VM.
939	* This function is only called once per-VM during initalization.
940	*
941	* @returns VBox status code.
942	* @param pVM The cross context VM structure.
943	*/
944	VMMR0DECL(int) SVMR0SetupVM(PVM pVM)
945	{
946	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
947	AssertReturn(pVM, VERR_INVALID_PARAMETER);
948	Assert(pVM->hm.s.svm.fSupported);
949
950	bool const fPauseFilter = RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_PAUSE_FILTER);
951	bool const fPauseFilterThreshold = RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_PAUSE_FILTER_THRESHOLD);
952	bool const fUsePauseFilter = fPauseFilter && pVM->hm.s.svm.cPauseFilter;
953
954	bool const fLbrVirt = RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_LBR_VIRT);
955	bool const fUseLbrVirt = fLbrVirt; /** @todo CFGM, IEM implementation etc. */
956
957	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
958	bool const fVirtVmsaveVmload = RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_VIRT_VMSAVE_VMLOAD);
959	bool const fUseVirtVmsaveVmload = fVirtVmsaveVmload && pVM->hm.s.svm.fVirtVmsaveVmload && pVM->hm.s.fNestedPaging;
960
961	bool const fVGif = RT_BOOL(pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_VGIF);
962	bool const fUseVGif = fVGif && pVM->hm.s.svm.fVGif;
963	#endif
964
965	PVMCPU pVCpu = &pVM->aCpus[0];
966	PSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
967	AssertMsgReturn(pVmcb, ("Invalid pVmcb for vcpu[0]\n"), VERR_SVM_INVALID_PVMCB);
968	PSVMVMCBCTRL pVmcbCtrl = &pVmcb->ctrl;
969
970	/* Always trap #AC for reasons of security. */
971	pVmcbCtrl->u32InterceptXcpt \|= RT_BIT_32(X86_XCPT_AC);
972
973	/* Always trap #DB for reasons of security. */
974	pVmcbCtrl->u32InterceptXcpt \|= RT_BIT_32(X86_XCPT_DB);
975
976	/* Trap exceptions unconditionally (debug purposes). */
977	#ifdef HMSVM_ALWAYS_TRAP_PF
978	pVmcbCtrl->u32InterceptXcpt \|= RT_BIT(X86_XCPT_PF);
979	#endif
980	#ifdef HMSVM_ALWAYS_TRAP_ALL_XCPTS
981	/* If you add any exceptions here, make sure to update hmR0SvmHandleExit(). */
982	pVmcbCtrl->u32InterceptXcpt \|= 0
983	\| RT_BIT(X86_XCPT_BP)
984	\| RT_BIT(X86_XCPT_DE)
985	\| RT_BIT(X86_XCPT_NM)
986	\| RT_BIT(X86_XCPT_UD)
987	\| RT_BIT(X86_XCPT_NP)
988	\| RT_BIT(X86_XCPT_SS)
989	\| RT_BIT(X86_XCPT_GP)
990	\| RT_BIT(X86_XCPT_PF)
991	\| RT_BIT(X86_XCPT_MF)
992	;
993	#endif
994
995	/* Apply the exceptions intercepts needed by the GIM provider. */
996	if (pVCpu->hm.s.fGIMTrapXcptUD)
997	pVmcbCtrl->u32InterceptXcpt \|= RT_BIT(X86_XCPT_UD);
998
999	/* Set up unconditional intercepts and conditions. */
1000	pVmcbCtrl->u64InterceptCtrl = HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS
1001	\| SVM_CTRL_INTERCEPT_VMMCALL;
1002
1003	#ifdef HMSVM_ALWAYS_TRAP_TASK_SWITCH
1004	pVmcbCtrl->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_TASK_SWITCH;
1005	#endif
1006
1007	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1008	/* Virtualized VMSAVE/VMLOAD. */
1009	pVmcbCtrl->LbrVirt.n.u1VirtVmsaveVmload = fUseVirtVmsaveVmload;
1010	if (!fUseVirtVmsaveVmload)
1011	{
1012	pVmcbCtrl->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_VMSAVE
1013	\| SVM_CTRL_INTERCEPT_VMLOAD;
1014	}
1015
1016	/* Virtual GIF. */
1017	pVmcbCtrl->IntCtrl.n.u1VGifEnable = fUseVGif;
1018	if (!fUseVGif)
1019	{
1020	pVmcbCtrl->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_CLGI
1021	\| SVM_CTRL_INTERCEPT_STGI;
1022	}
1023	#endif
1024
1025	/* CR4 writes must always be intercepted for tracking PGM mode changes. */
1026	pVmcbCtrl->u16InterceptWrCRx = RT_BIT(4);
1027
1028	/* Intercept all DRx reads and writes by default. Changed later on. */
1029	pVmcbCtrl->u16InterceptRdDRx = 0xffff;
1030	pVmcbCtrl->u16InterceptWrDRx = 0xffff;
1031
1032	/* Virtualize masking of INTR interrupts. (reads/writes from/to CR8 go to the V_TPR register) */
1033	pVmcbCtrl->IntCtrl.n.u1VIntrMasking = 1;
1034
1035	/* Ignore the priority in the virtual TPR. This is necessary for delivering PIC style (ExtInt) interrupts
1036	and we currently deliver both PIC and APIC interrupts alike, see hmR0SvmEvaluatePendingEvent() */
1037	pVmcbCtrl->IntCtrl.n.u1IgnoreTPR = 1;
1038
1039	/* Set the IO permission bitmap physical addresses. */
1040	pVmcbCtrl->u64IOPMPhysAddr = g_HCPhysIOBitmap;
1041
1042	/* LBR virtualization. */
1043	pVmcbCtrl->LbrVirt.n.u1LbrVirt = fUseLbrVirt;
1044
1045	/* The host ASID MBZ, for the guest start with 1. */
1046	pVmcbCtrl->TLBCtrl.n.u32ASID = 1;
1047
1048	/* Setup Nested Paging. This doesn't change throughout the execution time of the VM. */
1049	pVmcbCtrl->NestedPagingCtrl.n.u1NestedPaging = pVM->hm.s.fNestedPaging;
1050
1051	/* Without Nested Paging, we need additionally intercepts. */
1052	if (!pVM->hm.s.fNestedPaging)
1053	{
1054	/* CR3 reads/writes must be intercepted; our shadow values differ from the guest values. */
1055	pVmcbCtrl->u16InterceptRdCRx \|= RT_BIT(3);
1056	pVmcbCtrl->u16InterceptWrCRx \|= RT_BIT(3);
1057
1058	/* Intercept INVLPG and task switches (may change CR3, EFLAGS, LDT). */
1059	pVmcbCtrl->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_INVLPG
1060	\| SVM_CTRL_INTERCEPT_TASK_SWITCH;
1061
1062	/* Page faults must be intercepted to implement shadow paging. */
1063	pVmcbCtrl->u32InterceptXcpt \|= RT_BIT(X86_XCPT_PF);
1064	}
1065
1066	/* Setup Pause Filter for guest pause-loop (spinlock) exiting. */
1067	if (fUsePauseFilter)
1068	{
1069	Assert(pVM->hm.s.svm.cPauseFilter > 0);
1070	pVmcbCtrl->u16PauseFilterCount = pVM->hm.s.svm.cPauseFilter;
1071	if (fPauseFilterThreshold)
1072	pVmcbCtrl->u16PauseFilterThreshold = pVM->hm.s.svm.cPauseFilterThresholdTicks;
1073	pVmcbCtrl->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_PAUSE;
1074	}
1075
1076	/*
1077	* Setup the MSR permission bitmap.
1078	* The following MSRs are saved/restored automatically during the world-switch.
1079	* Don't intercept guest read/write accesses to these MSRs.
1080	*/
1081	uint8_t pbMsrBitmap = (uint8_t )pVCpu->hm.s.svm.pvMsrBitmap;
1082	hmR0SvmSetMsrPermission(pVCpu, pbMsrBitmap, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1083	hmR0SvmSetMsrPermission(pVCpu, pbMsrBitmap, MSR_K8_CSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1084	hmR0SvmSetMsrPermission(pVCpu, pbMsrBitmap, MSR_K6_STAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1085	hmR0SvmSetMsrPermission(pVCpu, pbMsrBitmap, MSR_K8_SF_MASK, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1086	hmR0SvmSetMsrPermission(pVCpu, pbMsrBitmap, MSR_K8_FS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1087	hmR0SvmSetMsrPermission(pVCpu, pbMsrBitmap, MSR_K8_GS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1088	hmR0SvmSetMsrPermission(pVCpu, pbMsrBitmap, MSR_K8_KERNEL_GS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1089	hmR0SvmSetMsrPermission(pVCpu, pbMsrBitmap, MSR_IA32_SYSENTER_CS, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1090	hmR0SvmSetMsrPermission(pVCpu, pbMsrBitmap, MSR_IA32_SYSENTER_ESP, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1091	hmR0SvmSetMsrPermission(pVCpu, pbMsrBitmap, MSR_IA32_SYSENTER_EIP, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1092	pVmcbCtrl->u64MSRPMPhysAddr = pVCpu->hm.s.svm.HCPhysMsrBitmap;
1093
1094	/* Initially all VMCB clean bits MBZ indicating that everything should be loaded from the VMCB in memory. */
1095	Assert(pVmcbCtrl->u32VmcbCleanBits == 0);
1096
1097	for (VMCPUID i = 1; i < pVM->cCpus; i++)
1098	{
1099	PVMCPU pVCpuCur = &pVM->aCpus[i];
1100	PSVMVMCB pVmcbCur = pVM->aCpus[i].hm.s.svm.pVmcb;
1101	AssertMsgReturn(pVmcbCur, ("Invalid pVmcb for vcpu[%u]\n", i), VERR_SVM_INVALID_PVMCB);
1102	PSVMVMCBCTRL pVmcbCtrlCur = &pVmcbCur->ctrl;
1103
1104	/* Copy the VMCB control area. */
1105	memcpy(pVmcbCtrlCur, pVmcbCtrl, sizeof(*pVmcbCtrlCur));
1106
1107	/* Copy the MSR bitmap and setup the VCPU-specific host physical address. */
1108	uint8_t pbMsrBitmapCur = (uint8_t )pVCpuCur->hm.s.svm.pvMsrBitmap;
1109	memcpy(pbMsrBitmapCur, pbMsrBitmap, SVM_MSRPM_PAGES << X86_PAGE_4K_SHIFT);
1110	pVmcbCtrlCur->u64MSRPMPhysAddr = pVCpuCur->hm.s.svm.HCPhysMsrBitmap;
1111
1112	/* Initially all VMCB clean bits MBZ indicating that everything should be loaded from the VMCB in memory. */
1113	Assert(pVmcbCtrlCur->u32VmcbCleanBits == 0);
1114
1115	/* Verify our assumption that GIM providers trap #UD uniformly across VCPUs initially. */
1116	Assert(pVCpuCur->hm.s.fGIMTrapXcptUD == pVCpu->hm.s.fGIMTrapXcptUD);
1117	}
1118
1119	return VINF_SUCCESS;
1120	}
1121
1122
1123	/**
1124	* Gets a pointer to the currently active guest (or nested-guest) VMCB.
1125	*
1126	* @returns Pointer to the current context VMCB.
1127	* @param pVCpu The cross context virtual CPU structure.
1128	*/
1129	DECLINLINE(PSVMVMCB) hmR0SvmGetCurrentVmcb(PVMCPU pVCpu)
1130	{
1131	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1132	if (CPUMIsGuestInSvmNestedHwVirtMode(&pVCpu->cpum.GstCtx))
1133	return pVCpu->cpum.GstCtx.hwvirt.svm.CTX_SUFF(pVmcb);
1134	#endif
1135	return pVCpu->hm.s.svm.pVmcb;
1136	}
1137
1138
1139	/**
1140	* Gets a pointer to the nested-guest VMCB cache.
1141	*
1142	* @returns Pointer to the nested-guest VMCB cache.
1143	* @param pVCpu The cross context virtual CPU structure.
1144	*/
1145	DECLINLINE(PSVMNESTEDVMCBCACHE) hmR0SvmGetNestedVmcbCache(PVMCPU pVCpu)
1146	{
1147	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1148	Assert(pVCpu->hm.s.svm.NstGstVmcbCache.fCacheValid);
1149	return &pVCpu->hm.s.svm.NstGstVmcbCache;
1150	#else
1151	RT_NOREF(pVCpu);
1152	return NULL;
1153	#endif
1154	}
1155
1156
1157	/**
1158	* Invalidates a guest page by guest virtual address.
1159	*
1160	* @returns VBox status code.
1161	* @param pVCpu The cross context virtual CPU structure.
1162	* @param GCVirt Guest virtual address of the page to invalidate.
1163	*/
1164	VMMR0DECL(int) SVMR0InvalidatePage(PVMCPU pVCpu, RTGCPTR GCVirt)
1165	{
1166	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.svm.fSupported);
1167
1168	bool const fFlushPending = pVCpu->CTX_SUFF(pVM)->hm.s.svm.fAlwaysFlushTLB \|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_TLB_FLUSH);
1169
1170	/* Skip it if a TLB flush is already pending. */
1171	if (!fFlushPending)
1172	{
1173	Log4Func(("%#RGv\n", GCVirt));
1174
1175	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
1176	AssertMsgReturn(pVmcb, ("Invalid pVmcb!\n"), VERR_SVM_INVALID_PVMCB);
1177
1178	#if HC_ARCH_BITS == 32
1179	/* If we get a flush in 64-bit guest mode, then force a full TLB flush. INVLPGA takes only 32-bit addresses. */
1180	if (CPUMIsGuestInLongMode(pVCpu))
1181	VMCPU_FF_SET(pVCpu, VMCPU_FF_TLB_FLUSH);
1182	else
1183	#endif
1184	{
1185	SVMR0InvlpgA(GCVirt, pVmcb->ctrl.TLBCtrl.n.u32ASID);
1186	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlbInvlpgVirt);
1187	}
1188	}
1189	return VINF_SUCCESS;
1190	}
1191
1192
1193	/**
1194	* Flushes the appropriate tagged-TLB entries.
1195	*
1196	* @param pVCpu The cross context virtual CPU structure.
1197	* @param pVmcb Pointer to the VM control block.
1198	* @param pHostCpu Pointer to the HM host-CPU info.
1199	*/
1200	static void hmR0SvmFlushTaggedTlb(PVMCPU pVCpu, PSVMVMCB pVmcb, PHMGLOBALCPUINFO pHostCpu)
1201	{
1202	/*
1203	* Force a TLB flush for the first world switch if the current CPU differs from the one
1204	* we ran on last. This can happen both for start & resume due to long jumps back to
1205	* ring-3.
1206	*
1207	* We also force a TLB flush every time when executing a nested-guest VCPU as there is no
1208	* correlation between it and the physical CPU.
1209	*
1210	* If the TLB flush count changed, another VM (VCPU rather) has hit the ASID limit while
1211	* flushing the TLB, so we cannot reuse the ASIDs without flushing.
1212	*/
1213	bool fNewAsid = false;
1214	Assert(pHostCpu->idCpu != NIL_RTCPUID);
1215	if ( pVCpu->hm.s.idLastCpu != pHostCpu->idCpu
1216	\|\| pVCpu->hm.s.cTlbFlushes != pHostCpu->cTlbFlushes
1217	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1218	\|\| CPUMIsGuestInSvmNestedHwVirtMode(&pVCpu->cpum.GstCtx)
1219	#endif
1220	)
1221	{
1222	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlbWorldSwitch);
1223	pVCpu->hm.s.fForceTLBFlush = true;
1224	fNewAsid = true;
1225	}
1226
1227	/* Set TLB flush state as checked until we return from the world switch. */
1228	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, true);
1229
1230	/* Check for explicit TLB flushes. */
1231	if (VMCPU_FF_TEST_AND_CLEAR(pVCpu, VMCPU_FF_TLB_FLUSH))
1232	{
1233	pVCpu->hm.s.fForceTLBFlush = true;
1234	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlb);
1235	}
1236
1237	/*
1238	* If the AMD CPU erratum 170, We need to flush the entire TLB for each world switch. Sad.
1239	* This Host CPU requirement takes precedence.
1240	*/
1241	PVM pVM = pVCpu->CTX_SUFF(pVM);
1242	if (pVM->hm.s.svm.fAlwaysFlushTLB)
1243	{
1244	pHostCpu->uCurrentAsid = 1;
1245	pVCpu->hm.s.uCurrentAsid = 1;
1246	pVCpu->hm.s.cTlbFlushes = pHostCpu->cTlbFlushes;
1247	pVCpu->hm.s.idLastCpu = pHostCpu->idCpu;
1248	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
1249
1250	/* Clear the VMCB Clean Bit for NP while flushing the TLB. See @bugref{7152}. */
1251	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
1252	}
1253	else
1254	{
1255	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_NOTHING;
1256	if (pVCpu->hm.s.fForceTLBFlush)
1257	{
1258	/* Clear the VMCB Clean Bit for NP while flushing the TLB. See @bugref{7152}. */
1259	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
1260
1261	if (fNewAsid)
1262	{
1263	++pHostCpu->uCurrentAsid;
1264
1265	bool fHitASIDLimit = false;
1266	if (pHostCpu->uCurrentAsid >= pVM->hm.s.uMaxAsid)
1267	{
1268	pHostCpu->uCurrentAsid = 1; /* Wraparound at 1; host uses 0 */
1269	pHostCpu->cTlbFlushes++; /* All VCPUs that run on this host CPU must use a new ASID. */
1270	fHitASIDLimit = true;
1271	}
1272
1273	if ( fHitASIDLimit
1274	\|\| pHostCpu->fFlushAsidBeforeUse)
1275	{
1276	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
1277	pHostCpu->fFlushAsidBeforeUse = false;
1278	}
1279
1280	pVCpu->hm.s.uCurrentAsid = pHostCpu->uCurrentAsid;
1281	pVCpu->hm.s.idLastCpu = pHostCpu->idCpu;
1282	pVCpu->hm.s.cTlbFlushes = pHostCpu->cTlbFlushes;
1283	}
1284	else
1285	{
1286	if (pVM->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_FLUSH_BY_ASID)
1287	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_SINGLE_CONTEXT;
1288	else
1289	pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
1290	}
1291
1292	pVCpu->hm.s.fForceTLBFlush = false;
1293	}
1294	}
1295
1296	/* Update VMCB with the ASID. */
1297	if (pVmcb->ctrl.TLBCtrl.n.u32ASID != pVCpu->hm.s.uCurrentAsid)
1298	{
1299	pVmcb->ctrl.TLBCtrl.n.u32ASID = pVCpu->hm.s.uCurrentAsid;
1300	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_ASID;
1301	}
1302
1303	AssertMsg(pVCpu->hm.s.idLastCpu == pHostCpu->idCpu,
1304	("vcpu idLastCpu=%u hostcpu idCpu=%u\n", pVCpu->hm.s.idLastCpu, pHostCpu->idCpu));
1305	AssertMsg(pVCpu->hm.s.cTlbFlushes == pHostCpu->cTlbFlushes,
1306	("Flush count mismatch for cpu %u (%u vs %u)\n", pHostCpu->idCpu, pVCpu->hm.s.cTlbFlushes, pHostCpu->cTlbFlushes));
1307	AssertMsg(pHostCpu->uCurrentAsid >= 1 && pHostCpu->uCurrentAsid < pVM->hm.s.uMaxAsid,
1308	("cpu%d uCurrentAsid = %x\n", pHostCpu->idCpu, pHostCpu->uCurrentAsid));
1309	AssertMsg(pVCpu->hm.s.uCurrentAsid >= 1 && pVCpu->hm.s.uCurrentAsid < pVM->hm.s.uMaxAsid,
1310	("cpu%d VM uCurrentAsid = %x\n", pHostCpu->idCpu, pVCpu->hm.s.uCurrentAsid));
1311
1312	#ifdef VBOX_WITH_STATISTICS
1313	if (pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_NOTHING)
1314	STAM_COUNTER_INC(&pVCpu->hm.s.StatNoFlushTlbWorldSwitch);
1315	else if ( pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_SINGLE_CONTEXT
1316	\|\| pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_SINGLE_CONTEXT_RETAIN_GLOBALS)
1317	{
1318	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushAsid);
1319	}
1320	else
1321	{
1322	Assert(pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_ENTIRE);
1323	STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushEntire);
1324	}
1325	#endif
1326	}
1327
1328
1329	/** @name 64-bit guest on 32-bit host OS helper functions.
1330	*
1331	* The host CPU is still 64-bit capable but the host OS is running in 32-bit
1332	* mode (code segment, paging). These wrappers/helpers perform the necessary
1333	* bits for the 32->64 switcher.
1334	*
1335	* @{ */
1336	#if HC_ARCH_BITS == 32 && defined(VBOX_ENABLE_64_BITS_GUESTS)
1337	/**
1338	* Prepares for and executes VMRUN (64-bit guests on a 32-bit host).
1339	*
1340	* @returns VBox status code.
1341	* @param HCPhysVmcbHost Physical address of host VMCB.
1342	* @param HCPhysVmcb Physical address of the VMCB.
1343	* @param pCtx Pointer to the guest-CPU context.
1344	* @param pVM The cross context VM structure.
1345	* @param pVCpu The cross context virtual CPU structure.
1346	*/
1347	DECLASM(int) SVMR0VMSwitcherRun64(RTHCPHYS HCPhysVmcbHost, RTHCPHYS HCPhysVmcb, PCPUMCTX pCtx, PVM pVM, PVMCPU pVCpu)
1348	{
1349	RT_NOREF2(pVM, pCtx);
1350	uint32_t aParam[8];
1351	aParam[0] = RT_LO_U32(HCPhysVmcbHost); /* Param 1: HCPhysVmcbHost - Lo. */
1352	aParam[1] = RT_HI_U32(HCPhysVmcbHost); /* Param 1: HCPhysVmcbHost - Hi. */
1353	aParam[2] = RT_LO_U32(HCPhysVmcb); /* Param 2: HCPhysVmcb - Lo. */
1354	aParam[3] = RT_HI_U32(HCPhysVmcb); /* Param 2: HCPhysVmcb - Hi. */
1355	aParam[4] = VM_RC_ADDR(pVM, pVM);
1356	aParam[5] = 0;
1357	aParam[6] = VM_RC_ADDR(pVM, pVCpu);
1358	aParam[7] = 0;
1359
1360	return SVMR0Execute64BitsHandler(pVCpu, HM64ON32OP_SVMRCVMRun64, RT_ELEMENTS(aParam), &aParam[0]);
1361	}
1362
1363
1364	/**
1365	* Executes the specified VMRUN handler in 64-bit mode.
1366	*
1367	* @returns VBox status code.
1368	* @param pVCpu The cross context virtual CPU structure.
1369	* @param enmOp The operation to perform.
1370	* @param cParams Number of parameters.
1371	* @param paParam Array of 32-bit parameters.
1372	*/
1373	VMMR0DECL(int) SVMR0Execute64BitsHandler(PVMCPU pVCpu, HM64ON32OP enmOp, uint32_t cParams, uint32_t *paParam)
1374	{
1375	PVM pVM = pVCpu->CTX_SUFF(pVM);
1376	AssertReturn(pVM->hm.s.pfnHost32ToGuest64R0, VERR_HM_NO_32_TO_64_SWITCHER);
1377	Assert(enmOp > HM64ON32OP_INVALID && enmOp < HM64ON32OP_END);
1378
1379	/* Disable interrupts. */
1380	RTHCUINTREG const fEFlags = ASMIntDisableFlags();
1381
1382	#ifdef VBOX_WITH_VMMR0_DISABLE_LAPIC_NMI
1383	RTCPUID idHostCpu = RTMpCpuId();
1384	CPUMR0SetLApic(pVCpu, idHostCpu);
1385	#endif
1386
1387	CPUMSetHyperESP(pVCpu, VMMGetStackRC(pVCpu));
1388	CPUMSetHyperEIP(pVCpu, enmOp);
1389	for (int i = (int)cParams - 1; i >= 0; i--)
1390	CPUMPushHyper(pVCpu, paParam[i]);
1391
1392	/* Call the switcher. */
1393	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatWorldSwitch3264, z);
1394	int rc = pVM->hm.s.pfnHost32ToGuest64R0(pVM, RT_OFFSETOF(VM, aCpus[pVCpu->idCpu].cpum) - RT_OFFSETOF(VM, cpum));
1395	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatWorldSwitch3264, z);
1396
1397	/* Restore interrupts. */
1398	ASMSetFlags(fEFlags);
1399	return rc;
1400	}
1401
1402	#endif /* HC_ARCH_BITS == 32 && defined(VBOX_ENABLE_64_BITS_GUESTS) */
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(PVMCPU pVCpu, PSVMVMCB pVmcb, uint8_t uXcpt)
1434	{
1435	Assert(uXcpt != X86_XCPT_DB);
1436	Assert(uXcpt != X86_XCPT_AC);
1437	#ifndef HMSVM_ALWAYS_TRAP_ALL_XCPTS
1438	if (pVmcb->ctrl.u32InterceptXcpt & RT_BIT(uXcpt))
1439	{
1440	bool fRemove = true;
1441	# ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1442	/* Only remove the intercept if the nested-guest is also not intercepting it! */
1443	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1444	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
1445	{
1446	PCSVMNESTEDVMCBCACHE pVmcbNstGstCache = hmR0SvmGetNestedVmcbCache(pVCpu);
1447	fRemove = !(pVmcbNstGstCache->u32InterceptXcpt & RT_BIT(uXcpt));
1448	}
1449	# else
1450	RT_NOREF(pVCpu);
1451	# endif
1452	if (fRemove)
1453	{
1454	pVmcb->ctrl.u32InterceptXcpt &= ~RT_BIT(uXcpt);
1455	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1456	}
1457	}
1458	#else
1459	RT_NOREF3(pVCpu, pVmcb, uXcpt);
1460	#endif
1461	}
1462
1463
1464	/**
1465	* Sets a control intercept in the specified VMCB.
1466	*
1467	* @param pVmcb Pointer to the VM control block.
1468	* @param fCtrlIntercept The control intercept (SVM_CTRL_INTERCEPT_*).
1469	*/
1470	DECLINLINE(void) hmR0SvmSetCtrlIntercept(PSVMVMCB pVmcb, uint64_t fCtrlIntercept)
1471	{
1472	if (!(pVmcb->ctrl.u64InterceptCtrl & fCtrlIntercept))
1473	{
1474	pVmcb->ctrl.u64InterceptCtrl \|= fCtrlIntercept;
1475	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1476	}
1477	}
1478
1479
1480	/**
1481	* Clears a control intercept in the specified VMCB.
1482	*
1483	* @returns @c true if the intercept is still set, @c false otherwise.
1484	* @param pVCpu The cross context virtual CPU structure.
1485	* @param pVmcb Pointer to the VM control block.
1486	* @param fCtrlIntercept The control intercept (SVM_CTRL_INTERCEPT_*).
1487	*
1488	* @remarks This takes into account if we're executing a nested-guest and only
1489	* removes the control intercept if both the guest -and- nested-guest
1490	* are not intercepting it.
1491	*/
1492	static bool hmR0SvmClearCtrlIntercept(PVMCPU pVCpu, PSVMVMCB pVmcb, uint64_t fCtrlIntercept)
1493	{
1494	if (pVmcb->ctrl.u64InterceptCtrl & fCtrlIntercept)
1495	{
1496	bool fRemove = true;
1497	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
1498	/* Only remove the control intercept if the nested-guest is also not intercepting it! */
1499	if (CPUMIsGuestInSvmNestedHwVirtMode(&pVCpu->cpum.GstCtx))
1500	{
1501	PCSVMNESTEDVMCBCACHE pVmcbNstGstCache = hmR0SvmGetNestedVmcbCache(pVCpu);
1502	fRemove = !(pVmcbNstGstCache->u64InterceptCtrl & fCtrlIntercept);
1503	}
1504	#else
1505	RT_NOREF(pVCpu);
1506	#endif
1507	if (fRemove)
1508	{
1509	pVmcb->ctrl.u64InterceptCtrl &= ~fCtrlIntercept;
1510	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1511	}
1512	}
1513
1514	return RT_BOOL(pVmcb->ctrl.u64InterceptCtrl & fCtrlIntercept);
1515	}
1516
1517
1518	/**
1519	* Exports the guest (or nested-guest) CR0 into the VMCB.
1520	*
1521	* @param pVCpu The cross context virtual CPU structure.
1522	* @param pVmcb Pointer to the VM control block.
1523	*
1524	* @remarks This assumes we always pre-load the guest FPU.
1525	* @remarks No-long-jump zone!!!
1526	*/
1527	static void hmR0SvmExportGuestCR0(PVMCPU pVCpu, PSVMVMCB pVmcb)
1528	{
1529	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1530
1531	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1532	uint64_t const uGuestCr0 = pCtx->cr0;
1533	uint64_t uShadowCr0 = uGuestCr0;
1534
1535	/* Always enable caching. */
1536	uShadowCr0 &= ~(X86_CR0_CD \| X86_CR0_NW);
1537
1538	/* When Nested Paging is not available use shadow page tables and intercept #PFs (latter done in SVMR0SetupVM()). */
1539	if (!pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging)
1540	{
1541	uShadowCr0 \|= X86_CR0_PG /* Use shadow page tables. */
1542	\| X86_CR0_WP; /* Guest CPL 0 writes to its read-only pages should cause a #PF #VMEXIT. */
1543	}
1544
1545	/*
1546	* Use the #MF style of legacy-FPU error reporting for now. Although AMD-V has MSRs that
1547	* lets us isolate the host from it, IEM/REM still needs work to emulate it properly,
1548	* see @bugref{7243#c103}.
1549	*/
1550	if (!(uGuestCr0 & X86_CR0_NE))
1551	{
1552	uShadowCr0 \|= X86_CR0_NE;
1553	hmR0SvmSetXcptIntercept(pVmcb, X86_XCPT_MF);
1554	}
1555	else
1556	hmR0SvmClearXcptIntercept(pVCpu, pVmcb, X86_XCPT_MF);
1557
1558	/*
1559	* If the shadow and guest CR0 are identical we can avoid intercepting CR0 reads.
1560	*
1561	* CR0 writes still needs interception as PGM requires tracking paging mode changes,
1562	* see @bugref{6944}.
1563	*
1564	* We also don't ever want to honor weird things like cache disable from the guest.
1565	* However, we can avoid intercepting changes to the TS & MP bits by clearing the CR0
1566	* write intercept below and keeping SVM_CTRL_INTERCEPT_CR0_SEL_WRITE instead.
1567	*/
1568	if (uShadowCr0 == uGuestCr0)
1569	{
1570	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
1571	{
1572	pVmcb->ctrl.u16InterceptRdCRx &= ~RT_BIT(0);
1573	pVmcb->ctrl.u16InterceptWrCRx &= ~RT_BIT(0);
1574	Assert(pVmcb->ctrl.u64InterceptCtrl & SVM_CTRL_INTERCEPT_CR0_SEL_WRITE);
1575	}
1576	else
1577	{
1578	/* If the nested-hypervisor intercepts CR0 reads/writes, we need to continue intercepting them. */
1579	PCSVMNESTEDVMCBCACHE pVmcbNstGstCache = hmR0SvmGetNestedVmcbCache(pVCpu);
1580	pVmcb->ctrl.u16InterceptRdCRx = (pVmcb->ctrl.u16InterceptRdCRx & ~RT_BIT(0))
1581	\| (pVmcbNstGstCache->u16InterceptRdCRx & RT_BIT(0));
1582	pVmcb->ctrl.u16InterceptWrCRx = (pVmcb->ctrl.u16InterceptWrCRx & ~RT_BIT(0))
1583	\| (pVmcbNstGstCache->u16InterceptWrCRx & RT_BIT(0));
1584	}
1585	}
1586	else
1587	{
1588	pVmcb->ctrl.u16InterceptRdCRx \|= RT_BIT(0);
1589	pVmcb->ctrl.u16InterceptWrCRx \|= RT_BIT(0);
1590	}
1591	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1592
1593	Assert(!RT_HI_U32(uShadowCr0));
1594	if (pVmcb->guest.u64CR0 != uShadowCr0)
1595	{
1596	pVmcb->guest.u64CR0 = uShadowCr0;
1597	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1598	}
1599	}
1600
1601
1602	/**
1603	* Exports the guest (or nested-guest) CR3 into the VMCB.
1604	*
1605	* @param pVCpu The cross context virtual CPU structure.
1606	* @param pVmcb Pointer to the VM control block.
1607	*
1608	* @remarks No-long-jump zone!!!
1609	*/
1610	static void hmR0SvmExportGuestCR3(PVMCPU pVCpu, PSVMVMCB pVmcb)
1611	{
1612	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1613
1614	PVM pVM = pVCpu->CTX_SUFF(pVM);
1615	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1616	if (pVM->hm.s.fNestedPaging)
1617	{
1618	PGMMODE enmShwPagingMode;
1619	#if HC_ARCH_BITS == 32
1620	if (CPUMIsGuestInLongModeEx(pCtx))
1621	enmShwPagingMode = PGMMODE_AMD64_NX;
1622	else
1623	#endif
1624	enmShwPagingMode = PGMGetHostMode(pVM);
1625
1626	pVmcb->ctrl.u64NestedPagingCR3 = PGMGetNestedCR3(pVCpu, enmShwPagingMode);
1627	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
1628	pVmcb->guest.u64CR3 = pCtx->cr3;
1629	Assert(pVmcb->ctrl.u64NestedPagingCR3);
1630	}
1631	else
1632	pVmcb->guest.u64CR3 = PGMGetHyperCR3(pVCpu);
1633
1634	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1635	}
1636
1637
1638	/**
1639	* Exports the guest (or nested-guest) CR4 into the VMCB.
1640	*
1641	* @param pVCpu The cross context virtual CPU structure.
1642	* @param pVmcb Pointer to the VM control block.
1643	*
1644	* @remarks No-long-jump zone!!!
1645	*/
1646	static int hmR0SvmExportGuestCR4(PVMCPU pVCpu, PSVMVMCB pVmcb)
1647	{
1648	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1649
1650	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1651	uint64_t uShadowCr4 = pCtx->cr4;
1652	if (!pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging)
1653	{
1654	switch (pVCpu->hm.s.enmShadowMode)
1655	{
1656	case PGMMODE_REAL:
1657	case PGMMODE_PROTECTED: /* Protected mode, no paging. */
1658	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1659
1660	case PGMMODE_32_BIT: /* 32-bit paging. */
1661	uShadowCr4 &= ~X86_CR4_PAE;
1662	break;
1663
1664	case PGMMODE_PAE: /* PAE paging. */
1665	case PGMMODE_PAE_NX: /* PAE paging with NX enabled. */
1666	/** Must use PAE paging as we could use physical memory > 4 GB */
1667	uShadowCr4 \|= X86_CR4_PAE;
1668	break;
1669
1670	case PGMMODE_AMD64: /* 64-bit AMD paging (long mode). */
1671	case PGMMODE_AMD64_NX: /* 64-bit AMD paging (long mode) with NX enabled. */
1672	#ifdef VBOX_ENABLE_64_BITS_GUESTS
1673	break;
1674	#else
1675	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1676	#endif
1677
1678	default: /* shut up gcc */
1679	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1680	}
1681	}
1682
1683	/* Whether to save/load/restore XCR0 during world switch depends on CR4.OSXSAVE and host+guest XCR0. */
1684	pVCpu->hm.s.fLoadSaveGuestXcr0 = (pCtx->cr4 & X86_CR4_OSXSAVE) && pCtx->aXcr[0] != ASMGetXcr0();
1685
1686	/* Avoid intercepting CR4 reads if the guest and shadow CR4 values are identical. */
1687	if (uShadowCr4 == pCtx->cr4)
1688	{
1689	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
1690	pVmcb->ctrl.u16InterceptRdCRx &= ~RT_BIT(4);
1691	else
1692	{
1693	/* If the nested-hypervisor intercepts CR4 reads, we need to continue intercepting them. */
1694	PCSVMNESTEDVMCBCACHE pVmcbNstGstCache = hmR0SvmGetNestedVmcbCache(pVCpu);
1695	pVmcb->ctrl.u16InterceptRdCRx = (pVmcb->ctrl.u16InterceptRdCRx & ~RT_BIT(4))
1696	\| (pVmcbNstGstCache->u16InterceptRdCRx & RT_BIT(4));
1697	}
1698	}
1699	else
1700	pVmcb->ctrl.u16InterceptRdCRx \|= RT_BIT(4);
1701
1702	/* CR4 writes are always intercepted (both guest, nested-guest) for tracking PGM mode changes. */
1703	Assert(pVmcb->ctrl.u16InterceptWrCRx & RT_BIT(4));
1704
1705	/* Update VMCB with the shadow CR4 the appropriate VMCB clean bits. */
1706	Assert(!RT_HI_U32(uShadowCr4));
1707	pVmcb->guest.u64CR4 = uShadowCr4;
1708	pVmcb->ctrl.u32VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_CRX_EFER \| HMSVM_VMCB_CLEAN_INTERCEPTS);
1709
1710	return VINF_SUCCESS;
1711	}
1712
1713
1714	/**
1715	* Exports the guest (or nested-guest) control registers into the VMCB.
1716	*
1717	* @returns VBox status code.
1718	* @param pVCpu The cross context virtual CPU structure.
1719	* @param pVmcb Pointer to the VM control block.
1720	*
1721	* @remarks No-long-jump zone!!!
1722	*/
1723	static int hmR0SvmExportGuestControlRegs(PVMCPU pVCpu, PSVMVMCB pVmcb)
1724	{
1725	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1726
1727	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_CR_MASK)
1728	{
1729	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_CR0)
1730	hmR0SvmExportGuestCR0(pVCpu, pVmcb);
1731
1732	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_CR2)
1733	{
1734	pVmcb->guest.u64CR2 = pVCpu->cpum.GstCtx.cr2;
1735	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CR2;
1736	}
1737
1738	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_CR3)
1739	hmR0SvmExportGuestCR3(pVCpu, pVmcb);
1740
1741	/* CR4 re-loading is ASSUMED to be done everytime we get in from ring-3! (XCR0) */
1742	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_CR4)
1743	{
1744	int rc = hmR0SvmExportGuestCR4(pVCpu, pVmcb);
1745	if (RT_FAILURE(rc))
1746	return rc;
1747	}
1748
1749	pVCpu->hm.s.fCtxChanged &= ~HM_CHANGED_GUEST_CR_MASK;
1750	}
1751	return VINF_SUCCESS;
1752	}
1753
1754
1755	/**
1756	* Exports the guest (or nested-guest) segment registers into the VMCB.
1757	*
1758	* @returns VBox status code.
1759	* @param pVCpu The cross context virtual CPU structure.
1760	* @param pVmcb Pointer to the VM control block.
1761	*
1762	* @remarks No-long-jump zone!!!
1763	*/
1764	static void hmR0SvmExportGuestSegmentRegs(PVMCPU pVCpu, PSVMVMCB pVmcb)
1765	{
1766	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1767	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1768
1769	/* Guest segment registers. */
1770	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SREG_MASK)
1771	{
1772	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_CS)
1773	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, CS, cs);
1774
1775	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SS)
1776	{
1777	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, SS, ss);
1778	pVmcb->guest.u8CPL = pCtx->ss.Attr.n.u2Dpl;
1779	}
1780
1781	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_DS)
1782	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, DS, ds);
1783
1784	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_ES)
1785	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, ES, es);
1786
1787	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_FS)
1788	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, FS, fs);
1789
1790	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_GS)
1791	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, GS, gs);
1792
1793	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_SEG;
1794	}
1795
1796	/* Guest TR. */
1797	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_TR)
1798	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, TR, tr);
1799
1800	/* Guest LDTR. */
1801	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_LDTR)
1802	HMSVM_SEG_REG_COPY_TO_VMCB(pCtx, &pVmcb->guest, LDTR, ldtr);
1803
1804	/* Guest GDTR. */
1805	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_GDTR)
1806	{
1807	pVmcb->guest.GDTR.u32Limit = pCtx->gdtr.cbGdt;
1808	pVmcb->guest.GDTR.u64Base = pCtx->gdtr.pGdt;
1809	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DT;
1810	}
1811
1812	/* Guest IDTR. */
1813	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_IDTR)
1814	{
1815	pVmcb->guest.IDTR.u32Limit = pCtx->idtr.cbIdt;
1816	pVmcb->guest.IDTR.u64Base = pCtx->idtr.pIdt;
1817	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DT;
1818	}
1819
1820	pVCpu->hm.s.fCtxChanged &= ~( HM_CHANGED_GUEST_SREG_MASK
1821	\| HM_CHANGED_GUEST_TABLE_MASK);
1822	}
1823
1824
1825	/**
1826	* Exports the guest (or nested-guest) MSRs into the VMCB.
1827	*
1828	* @param pVCpu The cross context virtual CPU structure.
1829	* @param pVmcb Pointer to the VM control block.
1830	*
1831	* @remarks No-long-jump zone!!!
1832	*/
1833	static void hmR0SvmExportGuestMsrs(PVMCPU pVCpu, PSVMVMCB pVmcb)
1834	{
1835	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1836	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1837
1838	/* Guest Sysenter MSRs. */
1839	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SYSENTER_MSR_MASK)
1840	{
1841	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SYSENTER_CS_MSR)
1842	pVmcb->guest.u64SysEnterCS = pCtx->SysEnter.cs;
1843
1844	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SYSENTER_EIP_MSR)
1845	pVmcb->guest.u64SysEnterEIP = pCtx->SysEnter.eip;
1846
1847	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SYSENTER_ESP_MSR)
1848	pVmcb->guest.u64SysEnterESP = pCtx->SysEnter.esp;
1849	}
1850
1851	/*
1852	* Guest EFER MSR.
1853	* AMD-V requires guest EFER.SVME to be set. Weird.
1854	* See AMD spec. 15.5.1 "Basic Operation" \| "Canonicalization and Consistency Checks".
1855	*/
1856	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_EFER_MSR)
1857	{
1858	pVmcb->guest.u64EFER = pCtx->msrEFER \| MSR_K6_EFER_SVME;
1859	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1860	}
1861
1862	/* If the guest isn't in 64-bit mode, clear MSR_K6_LME bit, otherwise SVM expects amd64 shadow paging. */
1863	if ( !CPUMIsGuestInLongModeEx(pCtx)
1864	&& (pCtx->msrEFER & MSR_K6_EFER_LME))
1865	{
1866	pVmcb->guest.u64EFER &= ~MSR_K6_EFER_LME;
1867	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1868	}
1869
1870	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_SYSCALL_MSRS)
1871	{
1872	pVmcb->guest.u64STAR = pCtx->msrSTAR;
1873	pVmcb->guest.u64LSTAR = pCtx->msrLSTAR;
1874	pVmcb->guest.u64CSTAR = pCtx->msrCSTAR;
1875	pVmcb->guest.u64SFMASK = pCtx->msrSFMASK;
1876	}
1877
1878	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_KERNEL_GS_BASE)
1879	pVmcb->guest.u64KernelGSBase = pCtx->msrKERNELGSBASE;
1880
1881	pVCpu->hm.s.fCtxChanged &= ~( HM_CHANGED_GUEST_SYSENTER_MSR_MASK
1882	\| HM_CHANGED_GUEST_EFER_MSR
1883	\| HM_CHANGED_GUEST_SYSCALL_MSRS
1884	\| HM_CHANGED_GUEST_KERNEL_GS_BASE);
1885
1886	/*
1887	* Setup the PAT MSR (applicable for Nested Paging only).
1888	*
1889	* While guests can modify and see the modified values through the shadow values,
1890	* we shall not honor any guest modifications of this MSR to ensure caching is always
1891	* enabled similar to how we clear CR0.CD and NW bits.
1892	*
1893	* For nested-guests this needs to always be set as well, see @bugref{7243#c109}.
1894	*/
1895	pVmcb->guest.u64PAT = MSR_IA32_CR_PAT_INIT_VAL;
1896
1897	/* Enable the last branch record bit if LBR virtualization is enabled. */
1898	if (pVmcb->ctrl.LbrVirt.n.u1LbrVirt)
1899	pVmcb->guest.u64DBGCTL = MSR_IA32_DEBUGCTL_LBR;
1900	}
1901
1902
1903	/**
1904	* Exports the guest (or nested-guest) debug state into the VMCB and programs
1905	* the necessary intercepts accordingly.
1906	*
1907	* @param pVCpu The cross context virtual CPU structure.
1908	* @param pVmcb Pointer to the VM control block.
1909	*
1910	* @remarks No-long-jump zone!!!
1911	* @remarks Requires EFLAGS to be up-to-date in the VMCB!
1912	*/
1913	static void hmR0SvmExportSharedDebugState(PVMCPU pVCpu, PSVMVMCB pVmcb)
1914	{
1915	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
1916
1917	/*
1918	* Anyone single stepping on the host side? If so, we'll have to use the
1919	* trap flag in the guest EFLAGS since AMD-V doesn't have a trap flag on
1920	* the VMM level like the VT-x implementations does.
1921	*/
1922	bool fInterceptMovDRx = false;
1923	bool const fStepping = pVCpu->hm.s.fSingleInstruction \|\| DBGFIsStepping(pVCpu);
1924	if (fStepping)
1925	{
1926	pVCpu->hm.s.fClearTrapFlag = true;
1927	pVmcb->guest.u64RFlags \|= X86_EFL_TF;
1928	fInterceptMovDRx = true; /* Need clean DR6, no guest mess. */
1929	}
1930
1931	if ( fStepping
1932	\|\| (CPUMGetHyperDR7(pVCpu) & X86_DR7_ENABLED_MASK))
1933	{
1934	/*
1935	* Use the combined guest and host DRx values found in the hypervisor
1936	* register set because the debugger has breakpoints active or someone
1937	* is single stepping on the host side.
1938	*
1939	* Note! DBGF expects a clean DR6 state before executing guest code.
1940	*/
1941	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS)
1942	if ( CPUMIsGuestInLongModeEx(pCtx)
1943	&& !CPUMIsHyperDebugStateActivePending(pVCpu))
1944	{
1945	CPUMR0LoadHyperDebugState(pVCpu, false /* include DR6 */);
1946	Assert(!CPUMIsGuestDebugStateActivePending(pVCpu));
1947	Assert(CPUMIsHyperDebugStateActivePending(pVCpu));
1948	}
1949	else
1950	#endif
1951	if (!CPUMIsHyperDebugStateActive(pVCpu))
1952	{
1953	CPUMR0LoadHyperDebugState(pVCpu, false /* include DR6 */);
1954	Assert(!CPUMIsGuestDebugStateActive(pVCpu));
1955	Assert(CPUMIsHyperDebugStateActive(pVCpu));
1956	}
1957
1958	/* Update DR6 & DR7. (The other DRx values are handled by CPUM one way or the other.) */
1959	if ( pVmcb->guest.u64DR6 != X86_DR6_INIT_VAL
1960	\|\| pVmcb->guest.u64DR7 != CPUMGetHyperDR7(pVCpu))
1961	{
1962	pVmcb->guest.u64DR7 = CPUMGetHyperDR7(pVCpu);
1963	pVmcb->guest.u64DR6 = X86_DR6_INIT_VAL;
1964	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
1965	}
1966
1967	/** @todo If we cared, we could optimize to allow the guest to read registers
1968	* with the same values. */
1969	fInterceptMovDRx = true;
1970	pVCpu->hm.s.fUsingHyperDR7 = true;
1971	Log5(("hmR0SvmExportSharedDebugState: Loaded hyper DRx\n"));
1972	}
1973	else
1974	{
1975	/*
1976	* Update DR6, DR7 with the guest values if necessary.
1977	*/
1978	if ( pVmcb->guest.u64DR7 != pCtx->dr[7]
1979	\|\| pVmcb->guest.u64DR6 != pCtx->dr[6])
1980	{
1981	pVmcb->guest.u64DR7 = pCtx->dr[7];
1982	pVmcb->guest.u64DR6 = pCtx->dr[6];
1983	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
1984	}
1985	pVCpu->hm.s.fUsingHyperDR7 = false;
1986
1987	/*
1988	* If the guest has enabled debug registers, we need to load them prior to
1989	* executing guest code so they'll trigger at the right time.
1990	*/
1991	if (pCtx->dr[7] & (X86_DR7_ENABLED_MASK \| X86_DR7_GD)) /** @todo Why GD? */
1992	{
1993	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS)
1994	if ( CPUMIsGuestInLongModeEx(pCtx)
1995	&& !CPUMIsGuestDebugStateActivePending(pVCpu))
1996	{
1997	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
1998	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxArmed);
1999	Assert(!CPUMIsHyperDebugStateActivePending(pVCpu));
2000	Assert(CPUMIsGuestDebugStateActivePending(pVCpu));
2001	}
2002	else
2003	#endif
2004	if (!CPUMIsGuestDebugStateActive(pVCpu))
2005	{
2006	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
2007	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxArmed);
2008	Assert(!CPUMIsHyperDebugStateActive(pVCpu));
2009	Assert(CPUMIsGuestDebugStateActive(pVCpu));
2010	}
2011	Log5(("hmR0SvmExportSharedDebugState: Loaded guest DRx\n"));
2012	}
2013	/*
2014	* If no debugging enabled, we'll lazy load DR0-3. We don't need to
2015	* intercept #DB as DR6 is updated in the VMCB.
2016	*
2017	* Note! If we cared and dared, we could skip intercepting \#DB here.
2018	* However, \#DB shouldn't be performance critical, so we'll play safe
2019	* and keep the code similar to the VT-x code and always intercept it.
2020	*/
2021	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS)
2022	else if ( !CPUMIsGuestDebugStateActivePending(pVCpu)
2023	&& !CPUMIsGuestDebugStateActive(pVCpu))
2024	#else
2025	else if (!CPUMIsGuestDebugStateActive(pVCpu))
2026	#endif
2027	{
2028	fInterceptMovDRx = true;
2029	}
2030	}
2031
2032	Assert(pVmcb->ctrl.u32InterceptXcpt & RT_BIT_32(X86_XCPT_DB));
2033	if (fInterceptMovDRx)
2034	{
2035	if ( pVmcb->ctrl.u16InterceptRdDRx != 0xffff
2036	\|\| pVmcb->ctrl.u16InterceptWrDRx != 0xffff)
2037	{
2038	pVmcb->ctrl.u16InterceptRdDRx = 0xffff;
2039	pVmcb->ctrl.u16InterceptWrDRx = 0xffff;
2040	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2041	}
2042	}
2043	else
2044	{
2045	if ( pVmcb->ctrl.u16InterceptRdDRx
2046	\|\| pVmcb->ctrl.u16InterceptWrDRx)
2047	{
2048	pVmcb->ctrl.u16InterceptRdDRx = 0;
2049	pVmcb->ctrl.u16InterceptWrDRx = 0;
2050	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2051	}
2052	}
2053	Log4Func(("DR6=%#RX64 DR7=%#RX64\n", pCtx->dr[6], pCtx->dr[7]));
2054	}
2055
2056	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
2057	/**
2058	* Exports the nested-guest hardware virtualization state into the nested-guest
2059	* VMCB.
2060	*
2061	* @param pVCpu The cross context virtual CPU structure.
2062	* @param pVmcbNstGst Pointer to the nested-guest VM control block.
2063	*
2064	* @remarks No-long-jump zone!!!
2065	*/
2066	static void hmR0SvmExportGuestHwvirtStateNested(PVMCPU pVCpu, PSVMVMCB pVmcbNstGst)
2067	{
2068	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2069
2070	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_HWVIRT)
2071	{
2072	/*
2073	* Ensure the nested-guest pause-filter counters don't exceed the outer guest values esp.
2074	* since SVM doesn't have a preemption timer.
2075	*
2076	* We do this here rather than in hmR0SvmSetupVmcbNested() as we may have been executing the
2077	* nested-guest in IEM incl. PAUSE instructions which would update the pause-filter counters
2078	* and may continue execution in SVM R0 without a nested-guest #VMEXIT in between.
2079	*/
2080	PVM pVM = pVCpu->CTX_SUFF(pVM);
2081	PSVMVMCBCTRL pVmcbNstGstCtrl = &pVmcbNstGst->ctrl;
2082	uint16_t const uGuestPauseFilterCount = pVM->hm.s.svm.cPauseFilter;
2083	uint16_t const uGuestPauseFilterThreshold = pVM->hm.s.svm.cPauseFilterThresholdTicks;
2084	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_PAUSE))
2085	{
2086	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2087	pVmcbNstGstCtrl->u16PauseFilterCount = RT_MIN(pCtx->hwvirt.svm.cPauseFilter, uGuestPauseFilterCount);
2088	pVmcbNstGstCtrl->u16PauseFilterThreshold = RT_MIN(pCtx->hwvirt.svm.cPauseFilterThreshold, uGuestPauseFilterThreshold);
2089	pVmcbNstGstCtrl->u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2090	}
2091	else
2092	{
2093	pVmcbNstGstCtrl->u16PauseFilterCount = uGuestPauseFilterCount;
2094	pVmcbNstGstCtrl->u16PauseFilterThreshold = uGuestPauseFilterThreshold;
2095	}
2096
2097	pVCpu->hm.s.fCtxChanged &= ~HM_CHANGED_GUEST_HWVIRT;
2098	}
2099	}
2100	#endif
2101
2102	/**
2103	* Exports the guest APIC TPR state into the VMCB.
2104	*
2105	* @returns VBox status code.
2106	* @param pVCpu The cross context virtual CPU structure.
2107	* @param pVmcb Pointer to the VM control block.
2108	*/
2109	static int hmR0SvmExportGuestApicTpr(PVMCPU pVCpu, PSVMVMCB pVmcb)
2110	{
2111	if (ASMAtomicUoReadU64(&pVCpu->hm.s.fCtxChanged) & HM_CHANGED_GUEST_APIC_TPR)
2112	{
2113	PVM pVM = pVCpu->CTX_SUFF(pVM);
2114	if ( PDMHasApic(pVM)
2115	&& APICIsEnabled(pVCpu))
2116	{
2117	bool fPendingIntr;
2118	uint8_t u8Tpr;
2119	int rc = APICGetTpr(pVCpu, &u8Tpr, &fPendingIntr, NULL /* pu8PendingIrq */);
2120	AssertRCReturn(rc, rc);
2121
2122	/* Assume that we need to trap all TPR accesses and thus need not check on
2123	every #VMEXIT if we should update the TPR. */
2124	Assert(pVmcb->ctrl.IntCtrl.n.u1VIntrMasking);
2125	pVCpu->hm.s.svm.fSyncVTpr = false;
2126
2127	if (!pVM->hm.s.fTPRPatchingActive)
2128	{
2129	/* Bits 3-0 of the VTPR field correspond to bits 7-4 of the TPR (which is the Task-Priority Class). */
2130	pVmcb->ctrl.IntCtrl.n.u8VTPR = (u8Tpr >> 4);
2131
2132	/* If there are interrupts pending, intercept CR8 writes to evaluate ASAP if we
2133	can deliver the interrupt to the guest. */
2134	if (fPendingIntr)
2135	pVmcb->ctrl.u16InterceptWrCRx \|= RT_BIT(8);
2136	else
2137	{
2138	pVmcb->ctrl.u16InterceptWrCRx &= ~RT_BIT(8);
2139	pVCpu->hm.s.svm.fSyncVTpr = true;
2140	}
2141
2142	pVmcb->ctrl.u32VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS \| HMSVM_VMCB_CLEAN_INT_CTRL);
2143	}
2144	else
2145	{
2146	/* 32-bit guests uses LSTAR MSR for patching guest code which touches the TPR. */
2147	pVmcb->guest.u64LSTAR = u8Tpr;
2148	uint8_t pbMsrBitmap = (uint8_t )pVCpu->hm.s.svm.pvMsrBitmap;
2149
2150	/* If there are interrupts pending, intercept LSTAR writes, otherwise don't intercept reads or writes. */
2151	if (fPendingIntr)
2152	hmR0SvmSetMsrPermission(pVCpu, pbMsrBitmap, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_INTERCEPT_WRITE);
2153	else
2154	{
2155	hmR0SvmSetMsrPermission(pVCpu, pbMsrBitmap, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
2156	pVCpu->hm.s.svm.fSyncVTpr = true;
2157	}
2158	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_IOPM_MSRPM;
2159	}
2160	}
2161	ASMAtomicUoAndU64(&pVCpu->hm.s.fCtxChanged, ~HM_CHANGED_GUEST_APIC_TPR);
2162	}
2163	return VINF_SUCCESS;
2164	}
2165
2166
2167	/**
2168	* Sets up the exception interrupts required for guest (or nested-guest)
2169	* execution in the VMCB.
2170	*
2171	* @param pVCpu The cross context virtual CPU structure.
2172	* @param pVmcb Pointer to the VM control block.
2173	*
2174	* @remarks No-long-jump zone!!!
2175	*/
2176	static void hmR0SvmExportGuestXcptIntercepts(PVMCPU pVCpu, PSVMVMCB pVmcb)
2177	{
2178	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2179
2180	/* If we modify intercepts from here, please check & adjust hmR0SvmMergeVmcbCtrlsNested() if required. */
2181	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_SVM_GUEST_XCPT_INTERCEPTS)
2182	{
2183	/* Trap #UD for GIM provider (e.g. for hypercalls). */
2184	if (pVCpu->hm.s.fGIMTrapXcptUD)
2185	hmR0SvmSetXcptIntercept(pVmcb, X86_XCPT_UD);
2186	else
2187	hmR0SvmClearXcptIntercept(pVCpu, pVmcb, X86_XCPT_UD);
2188
2189	/* Trap #BP for INT3 debug breakpoints set by the VM debugger. */
2190	if (pVCpu->CTX_SUFF(pVM)->dbgf.ro.cEnabledInt3Breakpoints)
2191	hmR0SvmSetXcptIntercept(pVmcb, X86_XCPT_BP);
2192	else
2193	hmR0SvmClearXcptIntercept(pVCpu, pVmcb, X86_XCPT_BP);
2194
2195	/* The remaining intercepts are handled elsewhere, e.g. in hmR0SvmExportGuestCR0(). */
2196	pVCpu->hm.s.fCtxChanged &= ~HM_CHANGED_SVM_GUEST_XCPT_INTERCEPTS;
2197	}
2198	}
2199
2200
2201	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
2202	/**
2203	* Merges guest and nested-guest intercepts for executing the nested-guest using
2204	* hardware-assisted SVM.
2205	*
2206	* This merges the guest and nested-guest intercepts in a way that if the outer
2207	* guest intercept is set we need to intercept it in the nested-guest as
2208	* well.
2209	*
2210	* @param pVCpu The cross context virtual CPU structure.
2211	* @param pVmcbNstGst Pointer to the nested-guest VM control block.
2212	*/
2213	static void hmR0SvmMergeVmcbCtrlsNested(PVMCPU pVCpu)
2214	{
2215	PVM pVM = pVCpu->CTX_SUFF(pVM);
2216	PCSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
2217	PSVMVMCB pVmcbNstGst = pVCpu->cpum.GstCtx.hwvirt.svm.CTX_SUFF(pVmcb);
2218	PSVMVMCBCTRL pVmcbNstGstCtrl = &pVmcbNstGst->ctrl;
2219
2220	/* Merge the guest's CR intercepts into the nested-guest VMCB. */
2221	pVmcbNstGstCtrl->u16InterceptRdCRx \|= pVmcb->ctrl.u16InterceptRdCRx;
2222	pVmcbNstGstCtrl->u16InterceptWrCRx \|= pVmcb->ctrl.u16InterceptWrCRx;
2223
2224	/* Always intercept CR4 writes for tracking PGM mode changes. */
2225	pVmcbNstGstCtrl->u16InterceptWrCRx \|= RT_BIT(4);
2226
2227	/* Without nested paging, intercept CR3 reads and writes as we load shadow page tables. */
2228	if (!pVM->hm.s.fNestedPaging)
2229	{
2230	pVmcbNstGstCtrl->u16InterceptRdCRx \|= RT_BIT(3);
2231	pVmcbNstGstCtrl->u16InterceptWrCRx \|= RT_BIT(3);
2232	}
2233
2234	/** @todo Figure out debugging with nested-guests, till then just intercept
2235	* all DR[0-15] accesses. */
2236	pVmcbNstGstCtrl->u16InterceptRdDRx \|= 0xffff;
2237	pVmcbNstGstCtrl->u16InterceptWrDRx \|= 0xffff;
2238
2239	/*
2240	* Merge the guest's exception intercepts into the nested-guest VMCB.
2241	*
2242	* - \#UD: Exclude these as the outer guest's GIM hypercalls are not applicable
2243	* while executing the nested-guest.
2244	*
2245	* - \#BP: Exclude breakpoints set by the VM debugger for the outer guest. This can
2246	* be tweaked later depending on how we wish to implement breakpoints.
2247	*
2248	* Warning!! This ASSUMES we only intercept \#UD for hypercall purposes and \#BP
2249	* for VM debugger breakpoints, see hmR0SvmExportGuestXcptIntercepts().
2250	*/
2251	#ifndef HMSVM_ALWAYS_TRAP_ALL_XCPTS
2252	pVmcbNstGstCtrl->u32InterceptXcpt \|= (pVmcb->ctrl.u32InterceptXcpt & ~( RT_BIT(X86_XCPT_UD)
2253	\| RT_BIT(X86_XCPT_BP)));
2254	#else
2255	pVmcbNstGstCtrl->u32InterceptXcpt \|= pVmcb->ctrl.u32InterceptXcpt;
2256	#endif
2257
2258	/*
2259	* Adjust intercepts while executing the nested-guest that differ from the
2260	* outer guest intercepts.
2261	*
2262	* - VINTR: Exclude the outer guest intercept as we don't need to cause VINTR #VMEXITs
2263	* that belong to the nested-guest to the outer guest.
2264	*
2265	* - VMMCALL: Exclude the outer guest intercept as when it's also not intercepted by
2266	* the nested-guest, the physical CPU raises a \#UD exception as expected.
2267	*/
2268	pVmcbNstGstCtrl->u64InterceptCtrl \|= (pVmcb->ctrl.u64InterceptCtrl & ~( SVM_CTRL_INTERCEPT_VINTR
2269	\| SVM_CTRL_INTERCEPT_VMMCALL))
2270	\| HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS;
2271
2272	Assert( (pVmcbNstGstCtrl->u64InterceptCtrl & HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS)
2273	== HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS);
2274
2275	/* Finally, update the VMCB clean bits. */
2276	pVmcbNstGstCtrl->u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2277	}
2278	#endif
2279
2280
2281	/**
2282	* Selects the appropriate function to run guest code.
2283	*
2284	* @returns VBox status code.
2285	* @param pVCpu The cross context virtual CPU structure.
2286	*
2287	* @remarks No-long-jump zone!!!
2288	*/
2289	static int hmR0SvmSelectVMRunHandler(PVMCPU pVCpu)
2290	{
2291	if (CPUMIsGuestInLongMode(pVCpu))
2292	{
2293	#ifndef VBOX_ENABLE_64_BITS_GUESTS
2294	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
2295	#endif
2296	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.fAllow64BitGuests); /* Guaranteed by hmR3InitFinalizeR0(). */
2297	#if HC_ARCH_BITS == 32
2298	/* 32-bit host. We need to switch to 64-bit before running the 64-bit guest. */
2299	pVCpu->hm.s.svm.pfnVMRun = SVMR0VMSwitcherRun64;
2300	#else
2301	/* 64-bit host or hybrid host. */
2302	pVCpu->hm.s.svm.pfnVMRun = SVMR0VMRun64;
2303	#endif
2304	}
2305	else
2306	{
2307	/* Guest is not in long mode, use the 32-bit handler. */
2308	pVCpu->hm.s.svm.pfnVMRun = SVMR0VMRun;
2309	}
2310	return VINF_SUCCESS;
2311	}
2312
2313
2314	/**
2315	* Enters the AMD-V session.
2316	*
2317	* @returns VBox status code.
2318	* @param pVCpu The cross context virtual CPU structure.
2319	* @param pHostCpu Pointer to the CPU info struct.
2320	*/
2321	VMMR0DECL(int) SVMR0Enter(PVMCPU pVCpu, PHMGLOBALCPUINFO pHostCpu)
2322	{
2323	AssertPtr(pVCpu);
2324	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.svm.fSupported);
2325	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2326	RT_NOREF(pHostCpu);
2327
2328	LogFlowFunc(("pVCpu=%p\n", pVCpu));
2329	Assert((pVCpu->hm.s.fCtxChanged & (HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE))
2330	== (HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE));
2331
2332	pVCpu->hm.s.fLeaveDone = false;
2333	return VINF_SUCCESS;
2334	}
2335
2336
2337	/**
2338	* Thread-context callback for AMD-V.
2339	*
2340	* @param enmEvent The thread-context event.
2341	* @param pVCpu The cross context virtual CPU structure.
2342	* @param fGlobalInit Whether global VT-x/AMD-V init. is used.
2343	* @thread EMT(pVCpu)
2344	*/
2345	VMMR0DECL(void) SVMR0ThreadCtxCallback(RTTHREADCTXEVENT enmEvent, PVMCPU pVCpu, bool fGlobalInit)
2346	{
2347	NOREF(fGlobalInit);
2348
2349	switch (enmEvent)
2350	{
2351	case RTTHREADCTXEVENT_OUT:
2352	{
2353	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2354	Assert(VMMR0ThreadCtxHookIsEnabled(pVCpu));
2355	VMCPU_ASSERT_EMT(pVCpu);
2356
2357	/* No longjmps (log-flush, locks) in this fragile context. */
2358	VMMRZCallRing3Disable(pVCpu);
2359
2360	if (!pVCpu->hm.s.fLeaveDone)
2361	{
2362	hmR0SvmLeave(pVCpu, false /* fImportState */);
2363	pVCpu->hm.s.fLeaveDone = true;
2364	}
2365
2366	/* Leave HM context, takes care of local init (term). */
2367	int rc = HMR0LeaveCpu(pVCpu);
2368	AssertRC(rc); NOREF(rc);
2369
2370	/* Restore longjmp state. */
2371	VMMRZCallRing3Enable(pVCpu);
2372	STAM_REL_COUNTER_INC(&pVCpu->hm.s.StatSwitchPreempt);
2373	break;
2374	}
2375
2376	case RTTHREADCTXEVENT_IN:
2377	{
2378	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2379	Assert(VMMR0ThreadCtxHookIsEnabled(pVCpu));
2380	VMCPU_ASSERT_EMT(pVCpu);
2381
2382	/* No longjmps (log-flush, locks) in this fragile context. */
2383	VMMRZCallRing3Disable(pVCpu);
2384
2385	/*
2386	* Initialize the bare minimum state required for HM. This takes care of
2387	* initializing AMD-V if necessary (onlined CPUs, local init etc.)
2388	*/
2389	int rc = hmR0EnterCpu(pVCpu);
2390	AssertRC(rc); NOREF(rc);
2391	Assert((pVCpu->hm.s.fCtxChanged & (HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE))
2392	== (HM_CHANGED_HOST_CONTEXT \| HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE));
2393
2394	pVCpu->hm.s.fLeaveDone = false;
2395
2396	/* Restore longjmp state. */
2397	VMMRZCallRing3Enable(pVCpu);
2398	break;
2399	}
2400
2401	default:
2402	break;
2403	}
2404	}
2405
2406
2407	/**
2408	* Saves the host state.
2409	*
2410	* @returns VBox status code.
2411	* @param pVCpu The cross context virtual CPU structure.
2412	*
2413	* @remarks No-long-jump zone!!!
2414	*/
2415	VMMR0DECL(int) SVMR0ExportHostState(PVMCPU pVCpu)
2416	{
2417	NOREF(pVCpu);
2418
2419	/* Nothing to do here. AMD-V does this for us automatically during the world-switch. */
2420	ASMAtomicUoAndU64(&pVCpu->hm.s.fCtxChanged, ~HM_CHANGED_HOST_CONTEXT);
2421	return VINF_SUCCESS;
2422	}
2423
2424
2425	/**
2426	* Exports the guest state from the guest-CPU context into the VMCB.
2427	*
2428	* The CPU state will be loaded from these fields on every successful VM-entry.
2429	* Also sets up the appropriate VMRUN function to execute guest code based on
2430	* the guest CPU mode.
2431	*
2432	* @returns VBox status code.
2433	* @param pVCpu The cross context virtual CPU structure.
2434	*
2435	* @remarks No-long-jump zone!!!
2436	*/
2437	static int hmR0SvmExportGuestState(PVMCPU pVCpu)
2438	{
2439	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatExportGuestState, x);
2440
2441	PSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
2442	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2443
2444	Assert(pVmcb);
2445	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(pCtx);
2446
2447	pVmcb->guest.u64RIP = pCtx->rip;
2448	pVmcb->guest.u64RSP = pCtx->rsp;
2449	pVmcb->guest.u64RFlags = pCtx->eflags.u32;
2450	pVmcb->guest.u64RAX = pCtx->rax;
2451	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
2452	if (pVmcb->ctrl.IntCtrl.n.u1VGifEnable)
2453	{
2454	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.svm.u32Features & X86_CPUID_SVM_FEATURE_EDX_VGIF);
2455	pVmcb->ctrl.IntCtrl.n.u1VGif = pCtx->hwvirt.fGif;
2456	}
2457	#endif
2458
2459	RTCCUINTREG const fEFlags = ASMIntDisableFlags();
2460
2461	int rc = hmR0SvmExportGuestControlRegs(pVCpu, pVmcb);
2462	AssertRCReturnStmt(rc, ASMSetFlags(fEFlags), rc);
2463
2464	hmR0SvmExportGuestSegmentRegs(pVCpu, pVmcb);
2465	hmR0SvmExportGuestMsrs(pVCpu, pVmcb);
2466	hmR0SvmExportGuestXcptIntercepts(pVCpu, pVmcb);
2467
2468	ASMSetFlags(fEFlags);
2469
2470	/* hmR0SvmExportGuestApicTpr() must be called -after- hmR0SvmExportGuestMsrs() as we
2471	otherwise we would overwrite the LSTAR MSR that we use for TPR patching. */
2472	hmR0SvmExportGuestApicTpr(pVCpu, pVmcb);
2473
2474	rc = hmR0SvmSelectVMRunHandler(pVCpu);
2475	AssertRCReturn(rc, rc);
2476
2477	/* Clear any bits that may be set but exported unconditionally or unused/reserved bits. */
2478	ASMAtomicUoAndU64(&pVCpu->hm.s.fCtxChanged, ~( HM_CHANGED_GUEST_RIP
2479	\| HM_CHANGED_GUEST_RFLAGS
2480	\| HM_CHANGED_GUEST_GPRS_MASK
2481	\| HM_CHANGED_GUEST_X87
2482	\| HM_CHANGED_GUEST_SSE_AVX
2483	\| HM_CHANGED_GUEST_OTHER_XSAVE
2484	\| HM_CHANGED_GUEST_XCRx
2485	\| HM_CHANGED_GUEST_TSC_AUX
2486	\| HM_CHANGED_GUEST_OTHER_MSRS
2487	\| HM_CHANGED_GUEST_HWVIRT
2488	\| (HM_CHANGED_KEEPER_STATE_MASK & ~HM_CHANGED_SVM_GUEST_XCPT_INTERCEPTS)));
2489
2490	#ifdef VBOX_STRICT
2491	/*
2492	* All of the guest-CPU state and SVM keeper bits should be exported here by now,
2493	* except for the host-context and/or shared host-guest context bits.
2494	*/
2495	uint64_t const fCtxChanged = ASMAtomicUoReadU64(&pVCpu->hm.s.fCtxChanged);
2496	RT_UNTRUSTED_NONVOLATILE_COPY_FENCE();
2497	AssertMsg(!(fCtxChanged & (HM_CHANGED_ALL_GUEST & ~HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE)),
2498	("fCtxChanged=%#RX64\n", fCtxChanged));
2499
2500	/*
2501	* If we need to log state that isn't always imported, we'll need to import them here.
2502	* See hmR0SvmPostRunGuest() for which part of the state is imported uncondtionally.
2503	*/
2504	hmR0SvmLogState(pVCpu, pVmcb, "hmR0SvmExportGuestState", 0 /* fFlags /, 0 / uVerbose */);
2505	#endif
2506
2507	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExportGuestState, x);
2508	return VINF_SUCCESS;
2509	}
2510
2511
2512	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
2513	/**
2514	* Merges the guest and nested-guest MSR permission bitmap.
2515	*
2516	* If the guest is intercepting an MSR we need to intercept it regardless of
2517	* whether the nested-guest is intercepting it or not.
2518	*
2519	* @param pHostCpu Pointer to the physical CPU HM info. struct.
2520	* @param pVCpu The cross context virtual CPU structure.
2521	*
2522	* @remarks No-long-jmp zone!!!
2523	*/
2524	DECLINLINE(void) hmR0SvmMergeMsrpmNested(PHMGLOBALCPUINFO pHostCpu, PVMCPU pVCpu)
2525	{
2526	uint64_t const pu64GstMsrpm = (uint64_t const )pVCpu->hm.s.svm.pvMsrBitmap;
2527	uint64_t const pu64NstGstMsrpm = (uint64_t const )pVCpu->cpum.GstCtx.hwvirt.svm.CTX_SUFF(pvMsrBitmap);
2528	uint64_t pu64DstMsrpm = (uint64_t )pHostCpu->n.svm.pvNstGstMsrpm;
2529
2530	/* MSRPM bytes from offset 0x1800 are reserved, so we stop merging there. */
2531	uint32_t const offRsvdQwords = 0x1800 >> 3;
2532	for (uint32_t i = 0; i < offRsvdQwords; i++)
2533	pu64DstMsrpm[i] = pu64NstGstMsrpm[i] \| pu64GstMsrpm[i];
2534	}
2535
2536
2537	/**
2538	* Caches the nested-guest VMCB fields before we modify them for execution using
2539	* hardware-assisted SVM.
2540	*
2541	* @returns true if the VMCB was previously already cached, false otherwise.
2542	* @param pVCpu The cross context virtual CPU structure.
2543	*
2544	* @sa HMSvmNstGstVmExitNotify.
2545	*/
2546	static bool hmR0SvmCacheVmcbNested(PVMCPU pVCpu)
2547	{
2548	/*
2549	* Cache the nested-guest programmed VMCB fields if we have not cached it yet.
2550	* Otherwise we risk re-caching the values we may have modified, see @bugref{7243#c44}.
2551	*
2552	* Nested-paging CR3 is not saved back into the VMCB on #VMEXIT, hence no need to
2553	* cache and restore it, see AMD spec. 15.25.4 "Nested Paging and VMRUN/#VMEXIT".
2554	*/
2555	PSVMNESTEDVMCBCACHE pVmcbNstGstCache = &pVCpu->hm.s.svm.NstGstVmcbCache;
2556	bool const fWasCached = pVmcbNstGstCache->fCacheValid;
2557	if (!fWasCached)
2558	{
2559	PCSVMVMCB pVmcbNstGst = pVCpu->cpum.GstCtx.hwvirt.svm.CTX_SUFF(pVmcb);
2560	PCSVMVMCBCTRL pVmcbNstGstCtrl = &pVmcbNstGst->ctrl;
2561	pVmcbNstGstCache->u16InterceptRdCRx = pVmcbNstGstCtrl->u16InterceptRdCRx;
2562	pVmcbNstGstCache->u16InterceptWrCRx = pVmcbNstGstCtrl->u16InterceptWrCRx;
2563	pVmcbNstGstCache->u16InterceptRdDRx = pVmcbNstGstCtrl->u16InterceptRdDRx;
2564	pVmcbNstGstCache->u16InterceptWrDRx = pVmcbNstGstCtrl->u16InterceptWrDRx;
2565	pVmcbNstGstCache->u16PauseFilterThreshold = pVmcbNstGstCtrl->u16PauseFilterThreshold;
2566	pVmcbNstGstCache->u16PauseFilterCount = pVmcbNstGstCtrl->u16PauseFilterCount;
2567	pVmcbNstGstCache->u32InterceptXcpt = pVmcbNstGstCtrl->u32InterceptXcpt;
2568	pVmcbNstGstCache->u64InterceptCtrl = pVmcbNstGstCtrl->u64InterceptCtrl;
2569	pVmcbNstGstCache->u64TSCOffset = pVmcbNstGstCtrl->u64TSCOffset;
2570	pVmcbNstGstCache->fVIntrMasking = pVmcbNstGstCtrl->IntCtrl.n.u1VIntrMasking;
2571	pVmcbNstGstCache->fNestedPaging = pVmcbNstGstCtrl->NestedPagingCtrl.n.u1NestedPaging;
2572	pVmcbNstGstCache->fLbrVirt = pVmcbNstGstCtrl->LbrVirt.n.u1LbrVirt;
2573	pVmcbNstGstCache->fCacheValid = true;
2574	Log4Func(("Cached VMCB fields\n"));
2575	}
2576
2577	return fWasCached;
2578	}
2579
2580
2581	/**
2582	* Sets up the nested-guest VMCB for execution using hardware-assisted SVM.
2583	*
2584	* This is done the first time we enter nested-guest execution using SVM R0
2585	* until the nested-guest \#VMEXIT (not to be confused with physical CPU
2586	* \#VMEXITs which may or may not cause a corresponding nested-guest \#VMEXIT).
2587	*
2588	* @param pVCpu The cross context virtual CPU structure.
2589	*/
2590	static void hmR0SvmSetupVmcbNested(PVMCPU pVCpu)
2591	{
2592	PSVMVMCB pVmcbNstGst = pVCpu->cpum.GstCtx.hwvirt.svm.CTX_SUFF(pVmcb);
2593	PSVMVMCBCTRL pVmcbNstGstCtrl = &pVmcbNstGst->ctrl;
2594
2595	/*
2596	* First cache the nested-guest VMCB fields we may potentially modify.
2597	*/
2598	bool const fVmcbCached = hmR0SvmCacheVmcbNested(pVCpu);
2599	if (!fVmcbCached)
2600	{
2601	/*
2602	* The IOPM of the nested-guest can be ignored because the the guest always
2603	* intercepts all IO port accesses. Thus, we'll swap to the guest IOPM rather
2604	* than the nested-guest IOPM and swap the field back on the #VMEXIT.
2605	*/
2606	pVmcbNstGstCtrl->u64IOPMPhysAddr = g_HCPhysIOBitmap;
2607
2608	/*
2609	* Use the same nested-paging as the outer guest. We can't dynamically switch off
2610	* nested-paging suddenly while executing a VM (see assertion at the end of
2611	* Trap0eHandler() in PGMAllBth.h).
2612	*/
2613	pVmcbNstGstCtrl->NestedPagingCtrl.n.u1NestedPaging = pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging;
2614
2615	/* Always enable V_INTR_MASKING as we do not want to allow access to the physical APIC TPR. */
2616	pVmcbNstGstCtrl->IntCtrl.n.u1VIntrMasking = 1;
2617
2618	/*
2619	* Turn off TPR syncing on #VMEXIT for nested-guests as CR8 intercepts are subject
2620	* to the nested-guest intercepts and we always run with V_INTR_MASKING.
2621	*/
2622	pVCpu->hm.s.svm.fSyncVTpr = false;
2623
2624	#ifdef DEBUG_ramshankar
2625	/* For debugging purposes - copy the LBR info. from outer guest VMCB. */
2626	pVmcbNstGstCtrl->LbrVirt.n.u1LbrVirt = pVmcb->ctrl.LbrVirt.n.u1LbrVirt;
2627	#endif
2628
2629	/*
2630	* If we don't expose Virtualized-VMSAVE/VMLOAD feature to the outer guest, we
2631	* need to intercept VMSAVE/VMLOAD instructions executed by the nested-guest.
2632	*/
2633	if (!pVCpu->CTX_SUFF(pVM)->cpum.ro.GuestFeatures.fSvmVirtVmsaveVmload)
2634	pVmcbNstGstCtrl->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_VMSAVE
2635	\| SVM_CTRL_INTERCEPT_VMLOAD;
2636
2637	/*
2638	* If we don't expose Virtual GIF feature to the outer guest, we need to intercept
2639	* CLGI/STGI instructions executed by the nested-guest.
2640	*/
2641	if (!pVCpu->CTX_SUFF(pVM)->cpum.ro.GuestFeatures.fSvmVGif)
2642	pVmcbNstGstCtrl->u64InterceptCtrl \|= SVM_CTRL_INTERCEPT_CLGI
2643	\| SVM_CTRL_INTERCEPT_STGI;
2644
2645	/* Merge the guest and nested-guest intercepts. */
2646	hmR0SvmMergeVmcbCtrlsNested(pVCpu);
2647
2648	/* Update the VMCB clean bits. */
2649	pVmcbNstGstCtrl->u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2650	}
2651	else
2652	{
2653	Assert(!pVCpu->hm.s.svm.fSyncVTpr);
2654	Assert(pVmcbNstGstCtrl->u64IOPMPhysAddr == g_HCPhysIOBitmap);
2655	Assert(RT_BOOL(pVmcbNstGstCtrl->NestedPagingCtrl.n.u1NestedPaging) == pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
2656	}
2657	}
2658
2659
2660	/**
2661	* Exports the nested-guest state into the VMCB.
2662	*
2663	* We need to export the entire state as we could be continuing nested-guest
2664	* execution at any point (not just immediately after VMRUN) and thus the VMCB
2665	* can be out-of-sync with the nested-guest state if it was executed in IEM.
2666	*
2667	* @returns VBox status code.
2668	* @param pVCpu The cross context virtual CPU structure.
2669	* @param pCtx Pointer to the guest-CPU context.
2670	*
2671	* @remarks No-long-jump zone!!!
2672	*/
2673	static int hmR0SvmExportGuestStateNested(PVMCPU pVCpu)
2674	{
2675	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatExportGuestState, x);
2676
2677	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2678	PSVMVMCB pVmcbNstGst = pCtx->hwvirt.svm.CTX_SUFF(pVmcb);
2679	Assert(pVmcbNstGst);
2680
2681	hmR0SvmSetupVmcbNested(pVCpu);
2682
2683	pVmcbNstGst->guest.u64RIP = pCtx->rip;
2684	pVmcbNstGst->guest.u64RSP = pCtx->rsp;
2685	pVmcbNstGst->guest.u64RFlags = pCtx->eflags.u32;
2686	pVmcbNstGst->guest.u64RAX = pCtx->rax;
2687
2688	RTCCUINTREG const fEFlags = ASMIntDisableFlags();
2689
2690	int rc = hmR0SvmExportGuestControlRegs(pVCpu, pVmcbNstGst);
2691	AssertRCReturnStmt(rc, ASMSetFlags(fEFlags), rc);
2692
2693	hmR0SvmExportGuestSegmentRegs(pVCpu, pVmcbNstGst);
2694	hmR0SvmExportGuestMsrs(pVCpu, pVmcbNstGst);
2695	hmR0SvmExportGuestHwvirtStateNested(pVCpu, pVmcbNstGst);
2696
2697	ASMSetFlags(fEFlags);
2698
2699	/* Nested VGIF not supported yet. */
2700	Assert(!pVmcbNstGst->ctrl.IntCtrl.n.u1VGifEnable);
2701
2702	rc = hmR0SvmSelectVMRunHandler(pVCpu);
2703	AssertRCReturn(rc, rc);
2704
2705	/* Clear any bits that may be set but exported unconditionally or unused/reserved bits. */
2706	ASMAtomicUoAndU64(&pVCpu->hm.s.fCtxChanged, ~( HM_CHANGED_GUEST_RIP
2707	\| HM_CHANGED_GUEST_RFLAGS
2708	\| HM_CHANGED_GUEST_GPRS_MASK
2709	\| HM_CHANGED_GUEST_APIC_TPR
2710	\| HM_CHANGED_GUEST_X87
2711	\| HM_CHANGED_GUEST_SSE_AVX
2712	\| HM_CHANGED_GUEST_OTHER_XSAVE
2713	\| HM_CHANGED_GUEST_XCRx
2714	\| HM_CHANGED_GUEST_TSC_AUX
2715	\| HM_CHANGED_GUEST_OTHER_MSRS
2716	\| HM_CHANGED_SVM_GUEST_XCPT_INTERCEPTS
2717	\| (HM_CHANGED_KEEPER_STATE_MASK & ~HM_CHANGED_SVM_MASK)));
2718
2719	#ifdef VBOX_STRICT
2720	/*
2721	* All of the guest-CPU state and SVM keeper bits should be exported here by now, except
2722	* for the host-context and/or shared host-guest context bits.
2723	*/
2724	uint64_t const fCtxChanged = ASMAtomicUoReadU64(&pVCpu->hm.s.fCtxChanged);
2725	RT_UNTRUSTED_NONVOLATILE_COPY_FENCE();
2726	AssertMsg(!(fCtxChanged & (HM_CHANGED_ALL_GUEST & ~HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE)),
2727	("fCtxChanged=%#RX64\n", fCtxChanged));
2728
2729	/*
2730	* If we need to log state that isn't always imported, we'll need to import them here.
2731	* See hmR0SvmPostRunGuest() for which part of the state is imported uncondtionally.
2732	*/
2733	hmR0SvmLogState(pVCpu, pVmcbNstGst, "hmR0SvmExportGuestStateNested", 0 /* fFlags /, 0 / uVerbose */);
2734	#endif
2735
2736	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExportGuestState, x);
2737	return rc;
2738	}
2739	#endif /* VBOX_WITH_NESTED_HWVIRT_SVM */
2740
2741
2742	/**
2743	* Exports the state shared between the host and guest (or nested-guest) into
2744	* the VMCB.
2745	*
2746	* @param pVCpu The cross context virtual CPU structure.
2747	* @param pVmcb Pointer to the VM control block.
2748	*
2749	* @remarks No-long-jump zone!!!
2750	*/
2751	static void hmR0SvmExportSharedState(PVMCPU pVCpu, PSVMVMCB pVmcb)
2752	{
2753	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2754	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2755
2756	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_GUEST_DR_MASK)
2757	{
2758	/** @todo Figure out stepping with nested-guest. */
2759	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2760	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
2761	hmR0SvmExportSharedDebugState(pVCpu, pVmcb);
2762	else
2763	{
2764	pVmcb->guest.u64DR6 = pCtx->dr[6];
2765	pVmcb->guest.u64DR7 = pCtx->dr[7];
2766	}
2767	}
2768
2769	pVCpu->hm.s.fCtxChanged &= ~HM_CHANGED_GUEST_DR_MASK;
2770	AssertMsg(!(pVCpu->hm.s.fCtxChanged & HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE),
2771	("fCtxChanged=%#RX64\n", pVCpu->hm.s.fCtxChanged));
2772	}
2773
2774
2775	/**
2776	* Worker for SVMR0ImportStateOnDemand.
2777	*
2778	* @param pVCpu The cross context virtual CPU structure.
2779	* @param fWhat What to import, CPUMCTX_EXTRN_XXX.
2780	*/
2781	static void hmR0SvmImportGuestState(PVMCPU pVCpu, uint64_t fWhat)
2782	{
2783	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatImportGuestState, x);
2784
2785	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
2786	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
2787	PCSVMVMCBSTATESAVE pVmcbGuest = &pVmcb->guest;
2788	PCSVMVMCBCTRL pVmcbCtrl = &pVmcb->ctrl;
2789
2790	Log4Func(("fExtrn=%#RX64 fWhat=%#RX64\n", pCtx->fExtrn, fWhat));
2791
2792	/*
2793	* We disable interrupts to make the updating of the state and in particular
2794	* the fExtrn modification atomic wrt to preemption hooks.
2795	*/
2796	RTCCUINTREG const fEFlags = ASMIntDisableFlags();
2797
2798	fWhat &= pCtx->fExtrn;
2799	if (fWhat)
2800	{
2801	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
2802	if (fWhat & CPUMCTX_EXTRN_HWVIRT)
2803	{
2804	if ( !CPUMIsGuestInSvmNestedHwVirtMode(pCtx)
2805	&& pVmcbCtrl->IntCtrl.n.u1VGifEnable)
2806	{
2807	/* We don't yet support passing VGIF feature to the guest. */
2808	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.svm.fVGif);
2809	pCtx->hwvirt.fGif = pVmcbCtrl->IntCtrl.n.u1VGif;
2810	}
2811	}
2812
2813	if (fWhat & CPUMCTX_EXTRN_HM_SVM_HWVIRT_VIRQ)
2814	{
2815	if ( !pVmcbCtrl->IntCtrl.n.u1VIrqPending
2816	&& VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_NESTED_GUEST))
2817	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INTERRUPT_NESTED_GUEST);
2818	}
2819	#endif
2820
2821	if (fWhat & CPUMCTX_EXTRN_HM_SVM_INT_SHADOW)
2822	{
2823	if (pVmcbCtrl->IntShadow.n.u1IntShadow)
2824	EMSetInhibitInterruptsPC(pVCpu, pVmcbGuest->u64RIP);
2825	else if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS))
2826	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS);
2827	}
2828
2829	if (fWhat & CPUMCTX_EXTRN_RIP)
2830	pCtx->rip = pVmcbGuest->u64RIP;
2831
2832	if (fWhat & CPUMCTX_EXTRN_RFLAGS)
2833	pCtx->eflags.u32 = pVmcbGuest->u64RFlags;
2834
2835	if (fWhat & CPUMCTX_EXTRN_RSP)
2836	pCtx->rsp = pVmcbGuest->u64RSP;
2837
2838	if (fWhat & CPUMCTX_EXTRN_RAX)
2839	pCtx->rax = pVmcbGuest->u64RAX;
2840
2841	if (fWhat & CPUMCTX_EXTRN_SREG_MASK)
2842	{
2843	if (fWhat & CPUMCTX_EXTRN_CS)
2844	{
2845	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, CS, cs);
2846	/* Correct the CS granularity bit. Haven't seen it being wrong in any other register (yet). */
2847	/** @todo SELM might need to be fixed as it too should not care about the
2848	* granularity bit. See @bugref{6785}. */
2849	if ( !pCtx->cs.Attr.n.u1Granularity
2850	&& pCtx->cs.Attr.n.u1Present
2851	&& pCtx->cs.u32Limit > UINT32_C(0xfffff))
2852	{
2853	Assert((pCtx->cs.u32Limit & 0xfff) == 0xfff);
2854	pCtx->cs.Attr.n.u1Granularity = 1;
2855	}
2856	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, cs);
2857	}
2858	if (fWhat & CPUMCTX_EXTRN_SS)
2859	{
2860	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, SS, ss);
2861	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, ss);
2862	/*
2863	* Sync the hidden SS DPL field. AMD CPUs have a separate CPL field in the
2864	* VMCB and uses that and thus it's possible that when the CPL changes during
2865	* guest execution that the SS DPL isn't updated by AMD-V. Observed on some
2866	* AMD Fusion CPUs with 64-bit guests.
2867	*
2868	* See AMD spec. 15.5.1 "Basic operation".
2869	*/
2870	Assert(!(pVmcbGuest->u8CPL & ~0x3));
2871	uint8_t const uCpl = pVmcbGuest->u8CPL;
2872	if (pCtx->ss.Attr.n.u2Dpl != uCpl)
2873	pCtx->ss.Attr.n.u2Dpl = uCpl & 0x3;
2874	}
2875	if (fWhat & CPUMCTX_EXTRN_DS)
2876	{
2877	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, DS, ds);
2878	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, ds);
2879	}
2880	if (fWhat & CPUMCTX_EXTRN_ES)
2881	{
2882	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, ES, es);
2883	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, es);
2884	}
2885	if (fWhat & CPUMCTX_EXTRN_FS)
2886	{
2887	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, FS, fs);
2888	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, fs);
2889	}
2890	if (fWhat & CPUMCTX_EXTRN_GS)
2891	{
2892	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, GS, gs);
2893	HMSVM_ASSERT_SEG_GRANULARITY(pCtx, gs);
2894	}
2895	}
2896
2897	if (fWhat & CPUMCTX_EXTRN_TABLE_MASK)
2898	{
2899	if (fWhat & CPUMCTX_EXTRN_TR)
2900	{
2901	/*
2902	* Fixup TR attributes so it's compatible with Intel. Important when saved-states
2903	* are used between Intel and AMD, see @bugref{6208#c39}.
2904	* ASSUME that it's normally correct and that we're in 32-bit or 64-bit mode.
2905	*/
2906	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, TR, tr);
2907	if (pCtx->tr.Attr.n.u4Type != X86_SEL_TYPE_SYS_386_TSS_BUSY)
2908	{
2909	if ( pCtx->tr.Attr.n.u4Type == X86_SEL_TYPE_SYS_386_TSS_AVAIL
2910	\|\| CPUMIsGuestInLongModeEx(pCtx))
2911	pCtx->tr.Attr.n.u4Type = X86_SEL_TYPE_SYS_386_TSS_BUSY;
2912	else if (pCtx->tr.Attr.n.u4Type == X86_SEL_TYPE_SYS_286_TSS_AVAIL)
2913	pCtx->tr.Attr.n.u4Type = X86_SEL_TYPE_SYS_286_TSS_BUSY;
2914	}
2915	}
2916
2917	if (fWhat & CPUMCTX_EXTRN_LDTR)
2918	HMSVM_SEG_REG_COPY_FROM_VMCB(pCtx, pVmcbGuest, LDTR, ldtr);
2919
2920	if (fWhat & CPUMCTX_EXTRN_GDTR)
2921	{
2922	pCtx->gdtr.cbGdt = pVmcbGuest->GDTR.u32Limit;
2923	pCtx->gdtr.pGdt = pVmcbGuest->GDTR.u64Base;
2924	}
2925
2926	if (fWhat & CPUMCTX_EXTRN_IDTR)
2927	{
2928	pCtx->idtr.cbIdt = pVmcbGuest->IDTR.u32Limit;
2929	pCtx->idtr.pIdt = pVmcbGuest->IDTR.u64Base;
2930	}
2931	}
2932
2933	if (fWhat & CPUMCTX_EXTRN_SYSCALL_MSRS)
2934	{
2935	pCtx->msrSTAR = pVmcbGuest->u64STAR;
2936	pCtx->msrLSTAR = pVmcbGuest->u64LSTAR;
2937	pCtx->msrCSTAR = pVmcbGuest->u64CSTAR;
2938	pCtx->msrSFMASK = pVmcbGuest->u64SFMASK;
2939	}
2940
2941	if (fWhat & CPUMCTX_EXTRN_SYSENTER_MSRS)
2942	{
2943	pCtx->SysEnter.cs = pVmcbGuest->u64SysEnterCS;
2944	pCtx->SysEnter.eip = pVmcbGuest->u64SysEnterEIP;
2945	pCtx->SysEnter.esp = pVmcbGuest->u64SysEnterESP;
2946	}
2947
2948	if (fWhat & CPUMCTX_EXTRN_KERNEL_GS_BASE)
2949	pCtx->msrKERNELGSBASE = pVmcbGuest->u64KernelGSBase;
2950
2951	if (fWhat & CPUMCTX_EXTRN_DR_MASK)
2952	{
2953	if (fWhat & CPUMCTX_EXTRN_DR6)
2954	{
2955	if (!pVCpu->hm.s.fUsingHyperDR7)
2956	pCtx->dr[6] = pVmcbGuest->u64DR6;
2957	else
2958	CPUMSetHyperDR6(pVCpu, pVmcbGuest->u64DR6);
2959	}
2960
2961	if (fWhat & CPUMCTX_EXTRN_DR7)
2962	{
2963	if (!pVCpu->hm.s.fUsingHyperDR7)
2964	pCtx->dr[7] = pVmcbGuest->u64DR7;
2965	else
2966	Assert(pVmcbGuest->u64DR7 == CPUMGetHyperDR7(pVCpu));
2967	}
2968	}
2969
2970	if (fWhat & CPUMCTX_EXTRN_CR_MASK)
2971	{
2972	if (fWhat & CPUMCTX_EXTRN_CR0)
2973	{
2974	/* We intercept changes to all CR0 bits except maybe TS & MP bits. */
2975	uint64_t const uCr0 = (pCtx->cr0 & ~(X86_CR0_TS \| X86_CR0_MP))
2976	\| (pVmcbGuest->u64CR0 & (X86_CR0_TS \| X86_CR0_MP));
2977	VMMRZCallRing3Disable(pVCpu); /* Calls into PGM which has Log statements. */
2978	CPUMSetGuestCR0(pVCpu, uCr0);
2979	VMMRZCallRing3Enable(pVCpu);
2980	}
2981
2982	if (fWhat & CPUMCTX_EXTRN_CR2)
2983	pCtx->cr2 = pVmcbGuest->u64CR2;
2984
2985	if (fWhat & CPUMCTX_EXTRN_CR3)
2986	{
2987	if ( pVmcbCtrl->NestedPagingCtrl.n.u1NestedPaging
2988	&& pCtx->cr3 != pVmcbGuest->u64CR3)
2989	{
2990	CPUMSetGuestCR3(pVCpu, pVmcbGuest->u64CR3);
2991	VMCPU_FF_SET(pVCpu, VMCPU_FF_HM_UPDATE_CR3);
2992	}
2993	}
2994
2995	/* Changes to CR4 are always intercepted. */
2996	}
2997
2998	/* Update fExtrn. */
2999	pCtx->fExtrn &= ~fWhat;
3000
3001	/* If everything has been imported, clear the HM keeper bit. */
3002	if (!(pCtx->fExtrn & HMSVM_CPUMCTX_EXTRN_ALL))
3003	{
3004	pCtx->fExtrn &= ~CPUMCTX_EXTRN_KEEPER_HM;
3005	Assert(!pCtx->fExtrn);
3006	}
3007	}
3008	else
3009	Assert(!pCtx->fExtrn \|\| (pCtx->fExtrn & HMSVM_CPUMCTX_EXTRN_ALL));
3010
3011	ASMSetFlags(fEFlags);
3012
3013	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatImportGuestState, x);
3014
3015	/*
3016	* Honor any pending CR3 updates.
3017	*
3018	* Consider this scenario: #VMEXIT -> VMMRZCallRing3Enable() -> do stuff that causes a longjmp
3019	* -> hmR0SvmCallRing3Callback() -> VMMRZCallRing3Disable() -> hmR0SvmImportGuestState()
3020	* -> Sets VMCPU_FF_HM_UPDATE_CR3 pending -> return from the longjmp -> continue with #VMEXIT
3021	* handling -> hmR0SvmImportGuestState() and here we are.
3022	*
3023	* The reason for such complicated handling is because VM-exits that call into PGM expect
3024	* CR3 to be up-to-date and thus any CR3-saves -before- the VM-exit (longjmp) would've
3025	* postponed the CR3 update via the force-flag and cleared CR3 from fExtrn. Any SVM R0
3026	* VM-exit handler that requests CR3 to be saved will end up here and we call PGMUpdateCR3().
3027	*
3028	* The longjmp exit path can't check these CR3 force-flags and call code that takes a lock again,
3029	* and does not process force-flag like regular exits to ring-3 either, we cover for it here.
3030	*/
3031	if ( VMMRZCallRing3IsEnabled(pVCpu)
3032	&& VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_UPDATE_CR3))
3033	{
3034	Assert(pCtx->cr3 == pVmcbGuest->u64CR3);
3035	PGMUpdateCR3(pVCpu, pCtx->cr3);
3036	}
3037	}
3038
3039
3040	/**
3041	* Saves the guest (or nested-guest) state from the VMCB into the guest-CPU
3042	* context.
3043	*
3044	* Currently there is no residual state left in the CPU that is not updated in the
3045	* VMCB.
3046	*
3047	* @returns VBox status code.
3048	* @param pVCpu The cross context virtual CPU structure.
3049	* @param fWhat What to import, CPUMCTX_EXTRN_XXX.
3050	*/
3051	VMMR0DECL(int) SVMR0ImportStateOnDemand(PVMCPU pVCpu, uint64_t fWhat)
3052	{
3053	hmR0SvmImportGuestState(pVCpu, fWhat);
3054	return VINF_SUCCESS;
3055	}
3056
3057
3058	/**
3059	* Does the necessary state syncing before returning to ring-3 for any reason
3060	* (longjmp, preemption, voluntary exits to ring-3) from AMD-V.
3061	*
3062	* @param pVCpu The cross context virtual CPU structure.
3063	* @param fImportState Whether to import the guest state from the VMCB back
3064	* to the guest-CPU context.
3065	*
3066	* @remarks No-long-jmp zone!!!
3067	*/
3068	static void hmR0SvmLeave(PVMCPU pVCpu, bool fImportState)
3069	{
3070	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
3071	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
3072	Assert(VMMR0IsLogFlushDisabled(pVCpu));
3073
3074	/*
3075	* !!! IMPORTANT !!!
3076	* If you modify code here, make sure to check whether hmR0SvmCallRing3Callback() needs to be updated too.
3077	*/
3078
3079	/* Save the guest state if necessary. */
3080	if (fImportState)
3081	hmR0SvmImportGuestState(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
3082
3083	/* Restore host FPU state if necessary and resync on next R0 reentry. */
3084	CPUMR0FpuStateMaybeSaveGuestAndRestoreHost(pVCpu);
3085	Assert(!CPUMIsGuestFPUStateActive(pVCpu));
3086
3087	/*
3088	* Restore host debug registers if necessary and resync on next R0 reentry.
3089	*/
3090	#ifdef VBOX_STRICT
3091	if (CPUMIsHyperDebugStateActive(pVCpu))
3092	{
3093	PSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb; /** @todo nested-guest. */
3094	Assert(pVmcb->ctrl.u16InterceptRdDRx == 0xffff);
3095	Assert(pVmcb->ctrl.u16InterceptWrDRx == 0xffff);
3096	}
3097	#endif
3098	CPUMR0DebugStateMaybeSaveGuestAndRestoreHost(pVCpu, false /* save DR6 */);
3099	Assert(!CPUMIsHyperDebugStateActive(pVCpu));
3100	Assert(!CPUMIsGuestDebugStateActive(pVCpu));
3101
3102	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatEntry);
3103	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatImportGuestState);
3104	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatExportGuestState);
3105	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatPreExit);
3106	STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatExitHandling);
3107	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchLongJmpToR3);
3108
3109	VMCPU_CMPXCHG_STATE(pVCpu, VMCPUSTATE_STARTED_HM, VMCPUSTATE_STARTED_EXEC);
3110	}
3111
3112
3113	/**
3114	* Leaves the AMD-V session.
3115	*
3116	* Only used while returning to ring-3 either due to longjump or exits to
3117	* ring-3.
3118	*
3119	* @returns VBox status code.
3120	* @param pVCpu The cross context virtual CPU structure.
3121	*/
3122	static int hmR0SvmLeaveSession(PVMCPU pVCpu)
3123	{
3124	HM_DISABLE_PREEMPT(pVCpu);
3125	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
3126	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
3127
3128	/* When thread-context hooks are used, we can avoid doing the leave again if we had been preempted before
3129	and done this from the SVMR0ThreadCtxCallback(). */
3130	if (!pVCpu->hm.s.fLeaveDone)
3131	{
3132	hmR0SvmLeave(pVCpu, true /* fImportState */);
3133	pVCpu->hm.s.fLeaveDone = true;
3134	}
3135
3136	/*
3137	* !!! IMPORTANT !!!
3138	* If you modify code here, make sure to check whether hmR0SvmCallRing3Callback() needs to be updated too.
3139	*/
3140
3141	/** @todo eliminate the need for calling VMMR0ThreadCtxHookDisable here! */
3142	/* Deregister hook now that we've left HM context before re-enabling preemption. */
3143	VMMR0ThreadCtxHookDisable(pVCpu);
3144
3145	/* Leave HM context. This takes care of local init (term). */
3146	int rc = HMR0LeaveCpu(pVCpu);
3147
3148	HM_RESTORE_PREEMPT();
3149	return rc;
3150	}
3151
3152
3153	/**
3154	* Does the necessary state syncing before doing a longjmp to ring-3.
3155	*
3156	* @returns VBox status code.
3157	* @param pVCpu The cross context virtual CPU structure.
3158	*
3159	* @remarks No-long-jmp zone!!!
3160	*/
3161	static int hmR0SvmLongJmpToRing3(PVMCPU pVCpu)
3162	{
3163	return hmR0SvmLeaveSession(pVCpu);
3164	}
3165
3166
3167	/**
3168	* VMMRZCallRing3() callback wrapper which saves the guest state (or restores
3169	* any remaining host state) before we longjump to ring-3 and possibly get
3170	* preempted.
3171	*
3172	* @param pVCpu The cross context virtual CPU structure.
3173	* @param enmOperation The operation causing the ring-3 longjump.
3174	* @param pvUser The user argument, NULL (currently unused).
3175	*/
3176	static DECLCALLBACK(int) hmR0SvmCallRing3Callback(PVMCPU pVCpu, VMMCALLRING3 enmOperation, void *pvUser)
3177	{
3178	RT_NOREF_PV(pvUser);
3179
3180	if (enmOperation == VMMCALLRING3_VM_R0_ASSERTION)
3181	{
3182	/*
3183	* !!! IMPORTANT !!!
3184	* If you modify code here, make sure to check whether hmR0SvmLeave() and hmR0SvmLeaveSession() needs
3185	* to be updated too. This is a stripped down version which gets out ASAP trying to not trigger any assertion.
3186	*/
3187	VMMRZCallRing3RemoveNotification(pVCpu);
3188	VMMRZCallRing3Disable(pVCpu);
3189	HM_DISABLE_PREEMPT(pVCpu);
3190
3191	/* Import the entire guest state. */
3192	hmR0SvmImportGuestState(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
3193
3194	/* Restore host FPU state if necessary and resync on next R0 reentry. */
3195	CPUMR0FpuStateMaybeSaveGuestAndRestoreHost(pVCpu);
3196
3197	/* Restore host debug registers if necessary and resync on next R0 reentry. */
3198	CPUMR0DebugStateMaybeSaveGuestAndRestoreHost(pVCpu, false /* save DR6 */);
3199
3200	/* Deregister the hook now that we've left HM context before re-enabling preemption. */
3201	/** @todo eliminate the need for calling VMMR0ThreadCtxHookDisable here! */
3202	VMMR0ThreadCtxHookDisable(pVCpu);
3203
3204	/* Leave HM context. This takes care of local init (term). */
3205	HMR0LeaveCpu(pVCpu);
3206
3207	HM_RESTORE_PREEMPT();
3208	return VINF_SUCCESS;
3209	}
3210
3211	Assert(pVCpu);
3212	Assert(VMMRZCallRing3IsEnabled(pVCpu));
3213	HMSVM_ASSERT_PREEMPT_SAFE(pVCpu);
3214
3215	VMMRZCallRing3Disable(pVCpu);
3216	Assert(VMMR0IsLogFlushDisabled(pVCpu));
3217
3218	Log4Func(("Calling hmR0SvmLongJmpToRing3\n"));
3219	int rc = hmR0SvmLongJmpToRing3(pVCpu);
3220	AssertRCReturn(rc, rc);
3221
3222	VMMRZCallRing3Enable(pVCpu);
3223	return VINF_SUCCESS;
3224	}
3225
3226
3227	/**
3228	* Take necessary actions before going back to ring-3.
3229	*
3230	* An action requires us to go back to ring-3. This function does the necessary
3231	* steps before we can safely return to ring-3. This is not the same as longjmps
3232	* to ring-3, this is voluntary.
3233	*
3234	* @returns VBox status code.
3235	* @param pVCpu The cross context virtual CPU structure.
3236	* @param rcExit The reason for exiting to ring-3. Can be
3237	* VINF_VMM_UNKNOWN_RING3_CALL.
3238	*/
3239	static int hmR0SvmExitToRing3(PVMCPU pVCpu, int rcExit)
3240	{
3241	Assert(pVCpu);
3242	HMSVM_ASSERT_PREEMPT_SAFE(pVCpu);
3243
3244	/* Please, no longjumps here (any logging shouldn't flush jump back to ring-3). NO LOGGING BEFORE THIS POINT! */
3245	VMMRZCallRing3Disable(pVCpu);
3246	Log4Func(("rcExit=%d LocalFF=%#RX32 GlobalFF=%#RX32\n", rcExit, pVCpu->fLocalForcedActions,
3247	pVCpu->CTX_SUFF(pVM)->fGlobalForcedActions));
3248
3249	/* We need to do this only while truly exiting the "inner loop" back to ring-3 and -not- for any longjmp to ring3. */
3250	if (pVCpu->hm.s.Event.fPending)
3251	{
3252	hmR0SvmPendingEventToTrpmTrap(pVCpu);
3253	Assert(!pVCpu->hm.s.Event.fPending);
3254	}
3255
3256	/* Sync. the necessary state for going back to ring-3. */
3257	hmR0SvmLeaveSession(pVCpu);
3258	STAM_COUNTER_DEC(&pVCpu->hm.s.StatSwitchLongJmpToR3);
3259
3260	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_TO_R3);
3261	CPUMSetChangedFlags(pVCpu, CPUM_CHANGED_SYSENTER_MSR
3262	\| CPUM_CHANGED_LDTR
3263	\| CPUM_CHANGED_GDTR
3264	\| CPUM_CHANGED_IDTR
3265	\| CPUM_CHANGED_TR
3266	\| CPUM_CHANGED_HIDDEN_SEL_REGS);
3267	if ( pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging
3268	&& CPUMIsGuestPagingEnabledEx(&pVCpu->cpum.GstCtx))
3269	{
3270	CPUMSetChangedFlags(pVCpu, CPUM_CHANGED_GLOBAL_TLB_FLUSH);
3271	}
3272
3273	/* Update the exit-to-ring 3 reason. */
3274	pVCpu->hm.s.rcLastExitToR3 = rcExit;
3275
3276	/* On our way back from ring-3 reload the guest state if there is a possibility of it being changed. */
3277	if (rcExit != VINF_EM_RAW_INTERRUPT)
3278	{
3279	Assert(!(pVCpu->cpum.GstCtx.fExtrn & HMSVM_CPUMCTX_EXTRN_ALL));
3280	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
3281	}
3282
3283	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchExitToR3);
3284
3285	/* We do -not- want any longjmp notifications after this! We must return to ring-3 ASAP. */
3286	VMMRZCallRing3RemoveNotification(pVCpu);
3287	VMMRZCallRing3Enable(pVCpu);
3288
3289	/*
3290	* If we're emulating an instruction, we shouldn't have any TRPM traps pending
3291	* and if we're injecting an event we should have a TRPM trap pending.
3292	*/
3293	AssertReturnStmt(rcExit != VINF_EM_RAW_INJECT_TRPM_EVENT \|\| TRPMHasTrap(pVCpu),
3294	pVCpu->hm.s.u32HMError = rcExit,
3295	VERR_SVM_IPE_5);
3296	AssertReturnStmt(rcExit != VINF_EM_RAW_EMULATE_INSTR \|\| !TRPMHasTrap(pVCpu),
3297	pVCpu->hm.s.u32HMError = rcExit,
3298	VERR_SVM_IPE_4);
3299
3300	return rcExit;
3301	}
3302
3303
3304	/**
3305	* Updates the use of TSC offsetting mode for the CPU and adjusts the necessary
3306	* intercepts.
3307	*
3308	* @param pVCpu The cross context virtual CPU structure.
3309	* @param pVmcb Pointer to the VM control block.
3310	*
3311	* @remarks No-long-jump zone!!!
3312	*/
3313	static void hmR0SvmUpdateTscOffsetting(PVMCPU pVCpu, PSVMVMCB pVmcb)
3314	{
3315	/*
3316	* Avoid intercepting RDTSC/RDTSCP if we determined the host TSC (++) is stable
3317	* and in case of a nested-guest, if the nested-VMCB specifies it is not intercepting
3318	* RDTSC/RDTSCP as well.
3319	*/
3320	bool fParavirtTsc;
3321	uint64_t uTscOffset;
3322	bool const fCanUseRealTsc = TMCpuTickCanUseRealTSC(pVCpu->CTX_SUFF(pVM), pVCpu, &uTscOffset, &fParavirtTsc);
3323
3324	bool fIntercept;
3325	if (fCanUseRealTsc)
3326	fIntercept = hmR0SvmClearCtrlIntercept(pVCpu, pVmcb, SVM_CTRL_INTERCEPT_RDTSC \| SVM_CTRL_INTERCEPT_RDTSCP);
3327	else
3328	{
3329	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_RDTSC \| SVM_CTRL_INTERCEPT_RDTSCP);
3330	fIntercept = true;
3331	}
3332
3333	if (!fIntercept)
3334	{
3335	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
3336	/* Apply the nested-guest VMCB's TSC offset over the guest TSC offset. */
3337	if (CPUMIsGuestInSvmNestedHwVirtMode(&pVCpu->cpum.GstCtx))
3338	uTscOffset = HMSvmNstGstApplyTscOffset(pVCpu, uTscOffset);
3339	#endif
3340
3341	/* Update the TSC offset in the VMCB and the relevant clean bits. */
3342	pVmcb->ctrl.u64TSCOffset = uTscOffset;
3343	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
3344
3345	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscOffset);
3346	}
3347	else
3348	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscIntercept);
3349
3350	/* Currently neither Hyper-V nor KVM need to update their paravirt. TSC
3351	information before every VM-entry, hence we have nothing to do here at the moment. */
3352	if (fParavirtTsc)
3353	STAM_COUNTER_INC(&pVCpu->hm.s.StatTscParavirt);
3354	}
3355
3356
3357	/**
3358	* Sets an event as a pending event to be injected into the guest.
3359	*
3360	* @param pVCpu The cross context virtual CPU structure.
3361	* @param pEvent Pointer to the SVM event.
3362	* @param GCPtrFaultAddress The fault-address (CR2) in case it's a
3363	* page-fault.
3364	*
3365	* @remarks Statistics counter assumes this is a guest event being reflected to
3366	* the guest i.e. 'StatInjectPendingReflect' is incremented always.
3367	*/
3368	DECLINLINE(void) hmR0SvmSetPendingEvent(PVMCPU pVCpu, PSVMEVENT pEvent, RTGCUINTPTR GCPtrFaultAddress)
3369	{
3370	Assert(!pVCpu->hm.s.Event.fPending);
3371	Assert(pEvent->n.u1Valid);
3372
3373	pVCpu->hm.s.Event.u64IntInfo = pEvent->u;
3374	pVCpu->hm.s.Event.fPending = true;
3375	pVCpu->hm.s.Event.GCPtrFaultAddress = GCPtrFaultAddress;
3376
3377	Log4Func(("u=%#RX64 u8Vector=%#x Type=%#x ErrorCodeValid=%RTbool ErrorCode=%#RX32\n", pEvent->u, pEvent->n.u8Vector,
3378	(uint8_t)pEvent->n.u3Type, !!pEvent->n.u1ErrorCodeValid, pEvent->n.u32ErrorCode));
3379	}
3380
3381
3382	/**
3383	* Sets an invalid-opcode (\#UD) exception as pending-for-injection into the VM.
3384	*
3385	* @param pVCpu The cross context virtual CPU structure.
3386	*/
3387	DECLINLINE(void) hmR0SvmSetPendingXcptUD(PVMCPU pVCpu)
3388	{
3389	SVMEVENT Event;
3390	Event.u = 0;
3391	Event.n.u1Valid = 1;
3392	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3393	Event.n.u8Vector = X86_XCPT_UD;
3394	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3395	}
3396
3397
3398	/**
3399	* Sets a debug (\#DB) exception as pending-for-injection into the VM.
3400	*
3401	* @param pVCpu The cross context virtual CPU structure.
3402	*/
3403	DECLINLINE(void) hmR0SvmSetPendingXcptDB(PVMCPU pVCpu)
3404	{
3405	SVMEVENT Event;
3406	Event.u = 0;
3407	Event.n.u1Valid = 1;
3408	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3409	Event.n.u8Vector = X86_XCPT_DB;
3410	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3411	}
3412
3413
3414	/**
3415	* Sets a page fault (\#PF) exception as pending-for-injection into the VM.
3416	*
3417	* @param pVCpu The cross context virtual CPU structure.
3418	* @param u32ErrCode The error-code for the page-fault.
3419	* @param uFaultAddress The page fault address (CR2).
3420	*
3421	* @remarks This updates the guest CR2 with @a uFaultAddress!
3422	*/
3423	DECLINLINE(void) hmR0SvmSetPendingXcptPF(PVMCPU pVCpu, uint32_t u32ErrCode, RTGCUINTPTR uFaultAddress)
3424	{
3425	SVMEVENT Event;
3426	Event.u = 0;
3427	Event.n.u1Valid = 1;
3428	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3429	Event.n.u8Vector = X86_XCPT_PF;
3430	Event.n.u1ErrorCodeValid = 1;
3431	Event.n.u32ErrorCode = u32ErrCode;
3432
3433	/* Update CR2 of the guest. */
3434	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_CR2);
3435	if (pVCpu->cpum.GstCtx.cr2 != uFaultAddress)
3436	{
3437	pVCpu->cpum.GstCtx.cr2 = uFaultAddress;
3438	/* The VMCB clean bit for CR2 will be updated while re-loading the guest state. */
3439	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_CR2);
3440	}
3441
3442	hmR0SvmSetPendingEvent(pVCpu, &Event, uFaultAddress);
3443	}
3444
3445
3446	/**
3447	* Sets a math-fault (\#MF) exception as pending-for-injection into the VM.
3448	*
3449	* @param pVCpu The cross context virtual CPU structure.
3450	*/
3451	DECLINLINE(void) hmR0SvmSetPendingXcptMF(PVMCPU pVCpu)
3452	{
3453	SVMEVENT Event;
3454	Event.u = 0;
3455	Event.n.u1Valid = 1;
3456	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3457	Event.n.u8Vector = X86_XCPT_MF;
3458	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3459	}
3460
3461
3462	/**
3463	* Sets a double fault (\#DF) exception as pending-for-injection into the VM.
3464	*
3465	* @param pVCpu The cross context virtual CPU structure.
3466	*/
3467	DECLINLINE(void) hmR0SvmSetPendingXcptDF(PVMCPU pVCpu)
3468	{
3469	SVMEVENT Event;
3470	Event.u = 0;
3471	Event.n.u1Valid = 1;
3472	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3473	Event.n.u8Vector = X86_XCPT_DF;
3474	Event.n.u1ErrorCodeValid = 1;
3475	Event.n.u32ErrorCode = 0;
3476	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3477	}
3478
3479
3480	/**
3481	* Injects an event into the guest upon VMRUN by updating the relevant field
3482	* in the VMCB.
3483	*
3484	* @param pVCpu The cross context virtual CPU structure.
3485	* @param pVmcb Pointer to the guest VM control block.
3486	* @param pEvent Pointer to the event.
3487	*
3488	* @remarks No-long-jump zone!!!
3489	* @remarks Requires CR0!
3490	*/
3491	DECLINLINE(void) hmR0SvmInjectEventVmcb(PVMCPU pVCpu, PSVMVMCB pVmcb, PSVMEVENT pEvent)
3492	{
3493	Assert(!pVmcb->ctrl.EventInject.n.u1Valid);
3494	pVmcb->ctrl.EventInject.u = pEvent->u;
3495	STAM_COUNTER_INC(&pVCpu->hm.s.paStatInjectedIrqsR0[pEvent->n.u8Vector & MASK_INJECT_IRQ_STAT]);
3496	RT_NOREF(pVCpu);
3497
3498	Log4Func(("u=%#RX64 u8Vector=%#x Type=%#x ErrorCodeValid=%RTbool ErrorCode=%#RX32\n", pEvent->u, pEvent->n.u8Vector,
3499	(uint8_t)pEvent->n.u3Type, !!pEvent->n.u1ErrorCodeValid, pEvent->n.u32ErrorCode));
3500	}
3501
3502
3503
3504	/**
3505	* Converts any TRPM trap into a pending HM event. This is typically used when
3506	* entering from ring-3 (not longjmp returns).
3507	*
3508	* @param pVCpu The cross context virtual CPU structure.
3509	*/
3510	static void hmR0SvmTrpmTrapToPendingEvent(PVMCPU pVCpu)
3511	{
3512	Assert(TRPMHasTrap(pVCpu));
3513	Assert(!pVCpu->hm.s.Event.fPending);
3514
3515	uint8_t uVector;
3516	TRPMEVENT enmTrpmEvent;
3517	RTGCUINT uErrCode;
3518	RTGCUINTPTR GCPtrFaultAddress;
3519	uint8_t cbInstr;
3520
3521	int rc = TRPMQueryTrapAll(pVCpu, &uVector, &enmTrpmEvent, &uErrCode, &GCPtrFaultAddress, &cbInstr);
3522	AssertRC(rc);
3523
3524	SVMEVENT Event;
3525	Event.u = 0;
3526	Event.n.u1Valid = 1;
3527	Event.n.u8Vector = uVector;
3528
3529	/* Refer AMD spec. 15.20 "Event Injection" for the format. */
3530	if (enmTrpmEvent == TRPM_TRAP)
3531	{
3532	Event.n.u3Type = SVM_EVENT_EXCEPTION;
3533	switch (uVector)
3534	{
3535	case X86_XCPT_NMI:
3536	{
3537	Event.n.u3Type = SVM_EVENT_NMI;
3538	break;
3539	}
3540
3541	case X86_XCPT_PF:
3542	case X86_XCPT_DF:
3543	case X86_XCPT_TS:
3544	case X86_XCPT_NP:
3545	case X86_XCPT_SS:
3546	case X86_XCPT_GP:
3547	case X86_XCPT_AC:
3548	{
3549	Event.n.u1ErrorCodeValid = 1;
3550	Event.n.u32ErrorCode = uErrCode;
3551	break;
3552	}
3553	}
3554	}
3555	else if (enmTrpmEvent == TRPM_HARDWARE_INT)
3556	Event.n.u3Type = SVM_EVENT_EXTERNAL_IRQ;
3557	else if (enmTrpmEvent == TRPM_SOFTWARE_INT)
3558	Event.n.u3Type = SVM_EVENT_SOFTWARE_INT;
3559	else
3560	AssertMsgFailed(("Invalid TRPM event type %d\n", enmTrpmEvent));
3561
3562	rc = TRPMResetTrap(pVCpu);
3563	AssertRC(rc);
3564
3565	Log4(("TRPM->HM event: u=%#RX64 u8Vector=%#x uErrorCodeValid=%RTbool uErrorCode=%#RX32\n", Event.u, Event.n.u8Vector,
3566	!!Event.n.u1ErrorCodeValid, Event.n.u32ErrorCode));
3567
3568	hmR0SvmSetPendingEvent(pVCpu, &Event, GCPtrFaultAddress);
3569	}
3570
3571
3572	/**
3573	* Converts any pending SVM event into a TRPM trap. Typically used when leaving
3574	* AMD-V to execute any instruction.
3575	*
3576	* @param pVCpu The cross context virtual CPU structure.
3577	*/
3578	static void hmR0SvmPendingEventToTrpmTrap(PVMCPU pVCpu)
3579	{
3580	Assert(pVCpu->hm.s.Event.fPending);
3581	Assert(TRPMQueryTrap(pVCpu, NULL /* pu8TrapNo /, NULL / pEnmType */) == VERR_TRPM_NO_ACTIVE_TRAP);
3582
3583	SVMEVENT Event;
3584	Event.u = pVCpu->hm.s.Event.u64IntInfo;
3585
3586	uint8_t uVector = Event.n.u8Vector;
3587	uint8_t uVectorType = Event.n.u3Type;
3588	TRPMEVENT enmTrapType = HMSvmEventToTrpmEventType(&Event);
3589
3590	Log4(("HM event->TRPM: uVector=%#x enmTrapType=%d\n", uVector, uVectorType));
3591
3592	int rc = TRPMAssertTrap(pVCpu, uVector, enmTrapType);
3593	AssertRC(rc);
3594
3595	if (Event.n.u1ErrorCodeValid)
3596	TRPMSetErrorCode(pVCpu, Event.n.u32ErrorCode);
3597
3598	if ( uVectorType == SVM_EVENT_EXCEPTION
3599	&& uVector == X86_XCPT_PF)
3600	{
3601	TRPMSetFaultAddress(pVCpu, pVCpu->hm.s.Event.GCPtrFaultAddress);
3602	Assert(pVCpu->hm.s.Event.GCPtrFaultAddress == CPUMGetGuestCR2(pVCpu));
3603	}
3604	else if (uVectorType == SVM_EVENT_SOFTWARE_INT)
3605	{
3606	AssertMsg( uVectorType == SVM_EVENT_SOFTWARE_INT
3607	\|\| (uVector == X86_XCPT_BP \|\| uVector == X86_XCPT_OF),
3608	("Invalid vector: uVector=%#x uVectorType=%#x\n", uVector, uVectorType));
3609	TRPMSetInstrLength(pVCpu, pVCpu->hm.s.Event.cbInstr);
3610	}
3611	pVCpu->hm.s.Event.fPending = false;
3612	}
3613
3614
3615	/**
3616	* Checks if the guest (or nested-guest) has an interrupt shadow active right
3617	* now.
3618	*
3619	* @returns @c true if the interrupt shadow is active, @c false otherwise.
3620	* @param pVCpu The cross context virtual CPU structure.
3621	*
3622	* @remarks No-long-jump zone!!!
3623	* @remarks Has side-effects with VMCPU_FF_INHIBIT_INTERRUPTS force-flag.
3624	*/
3625	static bool hmR0SvmIsIntrShadowActive(PVMCPU pVCpu)
3626	{
3627	/*
3628	* Instructions like STI and MOV SS inhibit interrupts till the next instruction
3629	* completes. Check if we should inhibit interrupts or clear any existing
3630	* interrupt inhibition.
3631	*/
3632	if (VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS))
3633	{
3634	if (pVCpu->cpum.GstCtx.rip != EMGetInhibitInterruptsPC(pVCpu))
3635	{
3636	/*
3637	* We can clear the inhibit force flag as even if we go back to the recompiler
3638	* without executing guest code in AMD-V, the flag's condition to be cleared is
3639	* met and thus the cleared state is correct.
3640	*/
3641	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS);
3642	return false;
3643	}
3644	return true;
3645	}
3646	return false;
3647	}
3648
3649
3650	/**
3651	* Sets the virtual interrupt intercept control in the VMCB.
3652	*
3653	* @param pVCpu The cross context virtual CPU structure.
3654	* @param pVmcb Pointer to the VM control block.
3655	*/
3656	static void hmR0SvmSetIntWindowExiting(PVMCPU pVCpu, PSVMVMCB pVmcb)
3657	{
3658	/*
3659	* When AVIC isn't supported, set up an interrupt window to cause a #VMEXIT when the guest
3660	* is ready to accept interrupts. At #VMEXIT, we then get the interrupt from the APIC
3661	* (updating ISR at the right time) and inject the interrupt.
3662	*
3663	* With AVIC is supported, we could make use of the asynchronously delivery without
3664	* #VMEXIT and we would be passing the AVIC page to SVM.
3665	*
3666	* In AMD-V, an interrupt window is achieved using a combination of V_IRQ (an interrupt
3667	* is pending), V_IGN_TPR (ignore TPR priorities) and the VINTR intercept all being set.
3668	*/
3669	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
3670	/*
3671	* Currently we don't overlay interupt windows and if there's any V_IRQ pending in the
3672	* nested-guest VMCB, we avoid setting up any interrupt window on behalf of the outer
3673	* guest.
3674	*/
3675	/** @todo Does this mean we end up prioritizing virtual interrupt
3676	* delivery/window over a physical interrupt (from the outer guest)
3677	* might be pending? */
3678	bool const fEnableIntWindow = !VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_NESTED_GUEST);
3679	if (!fEnableIntWindow)
3680	{
3681	Assert(CPUMIsGuestInSvmNestedHwVirtMode(&pVCpu->cpum.GstCtx));
3682	Log4(("Nested-guest V_IRQ already pending\n"));
3683	}
3684	#else
3685	bool const fEnableIntWindow = true;
3686	RT_NOREF(pVCpu);
3687	#endif
3688	if (fEnableIntWindow)
3689	{
3690	Assert(pVmcb->ctrl.IntCtrl.n.u1IgnoreTPR);
3691	pVmcb->ctrl.IntCtrl.n.u1VIrqPending = 1;
3692	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INT_CTRL;
3693	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_VINTR);
3694	Log4(("Set VINTR intercept\n"));
3695	}
3696	}
3697
3698
3699	/**
3700	* Clears the virtual interrupt intercept control in the VMCB as
3701	* we are figured the guest is unable process any interrupts
3702	* at this point of time.
3703	*
3704	* @param pVCpu The cross context virtual CPU structure.
3705	* @param pVmcb Pointer to the VM control block.
3706	*/
3707	static void hmR0SvmClearIntWindowExiting(PVMCPU pVCpu, PSVMVMCB pVmcb)
3708	{
3709	PSVMVMCBCTRL pVmcbCtrl = &pVmcb->ctrl;
3710	if ( pVmcbCtrl->IntCtrl.n.u1VIrqPending
3711	\|\| (pVmcbCtrl->u64InterceptCtrl & SVM_CTRL_INTERCEPT_VINTR))
3712	{
3713	pVmcbCtrl->IntCtrl.n.u1VIrqPending = 0;
3714	pVmcbCtrl->u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INT_CTRL;
3715	hmR0SvmClearCtrlIntercept(pVCpu, pVmcb, SVM_CTRL_INTERCEPT_VINTR);
3716	Log4(("Cleared VINTR intercept\n"));
3717	}
3718	}
3719
3720	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
3721	/**
3722	* Evaluates the event to be delivered to the nested-guest and sets it as the
3723	* pending event.
3724	*
3725	* @returns VBox strict status code.
3726	* @param pVCpu The cross context virtual CPU structure.
3727	*/
3728	static VBOXSTRICTRC hmR0SvmEvaluatePendingEventNested(PVMCPU pVCpu)
3729	{
3730	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
3731	HMSVM_ASSERT_IN_NESTED_GUEST(pCtx);
3732	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_HWVIRT
3733	\| CPUMCTX_EXTRN_RFLAGS
3734	\| CPUMCTX_EXTRN_HM_SVM_INT_SHADOW
3735	\| CPUMCTX_EXTRN_HM_SVM_HWVIRT_VIRQ);
3736
3737	Assert(!pVCpu->hm.s.Event.fPending);
3738	Assert(pCtx->hwvirt.fGif);
3739	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
3740	Assert(pVmcb);
3741
3742	bool const fVirtualGif = CPUMGetSvmNstGstVGif(pCtx);
3743	bool const fIntShadow = hmR0SvmIsIntrShadowActive(pVCpu);
3744	bool const fBlockNmi = VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_BLOCK_NMIS);
3745
3746	Log4Func(("fVirtualGif=%RTbool fBlockNmi=%RTbool fIntShadow=%RTbool fIntPending=%RTbool fNmiPending=%RTbool\n",
3747	fVirtualGif, fBlockNmi, fIntShadow, VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC),
3748	VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_NMI)));
3749
3750	/** @todo SMI. SMIs take priority over NMIs. */
3751
3752	/*
3753	* Check if the guest can receive NMIs.
3754	* Nested NMIs are not allowed, see AMD spec. 8.1.4 "Masking External Interrupts".
3755	* NMIs take priority over maskable interrupts, see AMD spec. 8.5 "Priorities".
3756	*/
3757	if ( VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_NMI)
3758	&& !fBlockNmi)
3759	{
3760	if ( fVirtualGif
3761	&& !fIntShadow)
3762	{
3763	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_NMI))
3764	{
3765	Log4(("Intercepting NMI -> #VMEXIT\n"));
3766	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_SVM_VMEXIT_MASK);
3767	return IEMExecSvmVmexit(pVCpu, SVM_EXIT_NMI, 0, 0);
3768	}
3769
3770	Log4(("Setting NMI pending for injection\n"));
3771	SVMEVENT Event;
3772	Event.u = 0;
3773	Event.n.u1Valid = 1;
3774	Event.n.u8Vector = X86_XCPT_NMI;
3775	Event.n.u3Type = SVM_EVENT_NMI;
3776	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3777	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INTERRUPT_NMI);
3778	}
3779	else if (!fVirtualGif)
3780	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_STGI);
3781	else
3782	hmR0SvmSetIntWindowExiting(pVCpu, pVmcb);
3783	}
3784	/*
3785	* Check if the nested-guest can receive external interrupts (generated by the guest's
3786	* PIC/APIC).
3787	*
3788	* External intercepts, NMI, SMI etc. from the physical CPU are -always- intercepted
3789	* when executing using hardware-assisted SVM, see HMSVM_MANDATORY_GUEST_CTRL_INTERCEPTS.
3790	*
3791	* External interrupts that are generated for the outer guest may be intercepted
3792	* depending on how the nested-guest VMCB was programmed by guest software.
3793	*
3794	* Physical interrupts always take priority over virtual interrupts,
3795	* see AMD spec. 15.21.4 "Injecting Virtual (INTR) Interrupts".
3796	*/
3797	else if ( VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC)
3798	&& !pVCpu->hm.s.fSingleInstruction)
3799	{
3800	if ( fVirtualGif
3801	&& !fIntShadow
3802	&& CPUMCanSvmNstGstTakePhysIntr(pVCpu, pCtx))
3803	{
3804	if (CPUMIsGuestSvmCtrlInterceptSet(pVCpu, pCtx, SVM_CTRL_INTERCEPT_INTR))
3805	{
3806	Log4(("Intercepting INTR -> #VMEXIT\n"));
3807	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_SVM_VMEXIT_MASK);
3808	return IEMExecSvmVmexit(pVCpu, SVM_EXIT_INTR, 0, 0);
3809	}
3810
3811	uint8_t u8Interrupt;
3812	int rc = PDMGetInterrupt(pVCpu, &u8Interrupt);
3813	if (RT_SUCCESS(rc))
3814	{
3815	Log4(("Setting external interrupt %#x pending for injection\n", u8Interrupt));
3816	SVMEVENT Event;
3817	Event.u = 0;
3818	Event.n.u1Valid = 1;
3819	Event.n.u8Vector = u8Interrupt;
3820	Event.n.u3Type = SVM_EVENT_EXTERNAL_IRQ;
3821	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3822	}
3823	else if (rc == VERR_APIC_INTR_MASKED_BY_TPR)
3824	{
3825	/*
3826	* AMD-V has no TPR thresholding feature. TPR and the force-flag will be
3827	* updated eventually when the TPR is written by the guest.
3828	*/
3829	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchTprMaskedIrq);
3830	}
3831	else
3832	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchGuestIrq);
3833	}
3834	else if (!fVirtualGif)
3835	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_STGI);
3836	else
3837	hmR0SvmSetIntWindowExiting(pVCpu, pVmcb);
3838	}
3839
3840	return VINF_SUCCESS;
3841	}
3842	#endif
3843
3844	/**
3845	* Evaluates the event to be delivered to the guest and sets it as the pending
3846	* event.
3847	*
3848	* @param pVCpu The cross context virtual CPU structure.
3849	*/
3850	static void hmR0SvmEvaluatePendingEvent(PVMCPU pVCpu)
3851	{
3852	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
3853	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(pCtx);
3854	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_HWVIRT
3855	\| CPUMCTX_EXTRN_RFLAGS
3856	\| CPUMCTX_EXTRN_HM_SVM_INT_SHADOW);
3857
3858	Assert(!pVCpu->hm.s.Event.fPending);
3859	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
3860	Assert(pVmcb);
3861
3862	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
3863	bool const fGif = pCtx->hwvirt.fGif;
3864	#else
3865	bool const fGif = true;
3866	#endif
3867	bool const fIntShadow = hmR0SvmIsIntrShadowActive(pVCpu);
3868	bool const fBlockInt = !(pCtx->eflags.u32 & X86_EFL_IF);
3869	bool const fBlockNmi = VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_BLOCK_NMIS);
3870
3871	Log4Func(("fGif=%RTbool fBlockNmi=%RTbool fBlockInt=%RTbool fIntShadow=%RTbool fIntPending=%RTbool NMI pending=%RTbool\n",
3872	fGif, fBlockNmi, fBlockInt, fIntShadow,
3873	VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC),
3874	VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_NMI)));
3875
3876	/** @todo SMI. SMIs take priority over NMIs. */
3877
3878	/*
3879	* Check if the guest can receive NMIs.
3880	* Nested NMIs are not allowed, see AMD spec. 8.1.4 "Masking External Interrupts".
3881	* NMIs take priority over maskable interrupts, see AMD spec. 8.5 "Priorities".
3882	*/
3883	if ( VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_NMI)
3884	&& !fBlockNmi)
3885	{
3886	if ( fGif
3887	&& !fIntShadow)
3888	{
3889	Log4(("Setting NMI pending for injection\n"));
3890	SVMEVENT Event;
3891	Event.u = 0;
3892	Event.n.u1Valid = 1;
3893	Event.n.u8Vector = X86_XCPT_NMI;
3894	Event.n.u3Type = SVM_EVENT_NMI;
3895	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3896	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INTERRUPT_NMI);
3897	}
3898	else if (!fGif)
3899	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_STGI);
3900	else
3901	hmR0SvmSetIntWindowExiting(pVCpu, pVmcb);
3902	}
3903	/*
3904	* Check if the guest can receive external interrupts (PIC/APIC). Once PDMGetInterrupt()
3905	* returns a valid interrupt we -must- deliver the interrupt. We can no longer re-request
3906	* it from the APIC device.
3907	*/
3908	else if ( VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_APIC \| VMCPU_FF_INTERRUPT_PIC)
3909	&& !pVCpu->hm.s.fSingleInstruction)
3910	{
3911	if ( fGif
3912	&& !fBlockInt
3913	&& !fIntShadow)
3914	{
3915	uint8_t u8Interrupt;
3916	int rc = PDMGetInterrupt(pVCpu, &u8Interrupt);
3917	if (RT_SUCCESS(rc))
3918	{
3919	Log4(("Setting external interrupt %#x pending for injection\n", u8Interrupt));
3920	SVMEVENT Event;
3921	Event.u = 0;
3922	Event.n.u1Valid = 1;
3923	Event.n.u8Vector = u8Interrupt;
3924	Event.n.u3Type = SVM_EVENT_EXTERNAL_IRQ;
3925	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3926	}
3927	else if (rc == VERR_APIC_INTR_MASKED_BY_TPR)
3928	{
3929	/*
3930	* AMD-V has no TPR thresholding feature. TPR and the force-flag will be
3931	* updated eventually when the TPR is written by the guest.
3932	*/
3933	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchTprMaskedIrq);
3934	}
3935	else
3936	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchGuestIrq);
3937	}
3938	else if (!fGif)
3939	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_STGI);
3940	else
3941	hmR0SvmSetIntWindowExiting(pVCpu, pVmcb);
3942	}
3943	}
3944
3945
3946	/**
3947	* Injects any pending events into the guest (or nested-guest).
3948	*
3949	* @param pVCpu The cross context virtual CPU structure.
3950	* @param pVmcb Pointer to the VM control block.
3951	*
3952	* @remarks Must only be called when we are guaranteed to enter
3953	* hardware-assisted SVM execution and not return to ring-3
3954	* prematurely.
3955	*/
3956	static void hmR0SvmInjectPendingEvent(PVMCPU pVCpu, PSVMVMCB pVmcb)
3957	{
3958	Assert(!TRPMHasTrap(pVCpu));
3959	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
3960
3961	bool const fIntShadow = hmR0SvmIsIntrShadowActive(pVCpu);
3962	#ifdef VBOX_STRICT
3963	PCCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
3964	bool const fGif = pCtx->hwvirt.fGif;
3965	bool fAllowInt = fGif;
3966	if (fGif)
3967	{
3968	/*
3969	* For nested-guests we have no way to determine if we're injecting a physical or
3970	* virtual interrupt at this point. Hence the partial verification below.
3971	*/
3972	if (CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
3973	fAllowInt = CPUMCanSvmNstGstTakePhysIntr(pVCpu, pCtx) \|\| CPUMCanSvmNstGstTakeVirtIntr(pVCpu, pCtx);
3974	else
3975	fAllowInt = RT_BOOL(pCtx->eflags.u32 & X86_EFL_IF);
3976	}
3977	#endif
3978
3979	if (pVCpu->hm.s.Event.fPending)
3980	{
3981	SVMEVENT Event;
3982	Event.u = pVCpu->hm.s.Event.u64IntInfo;
3983	Assert(Event.n.u1Valid);
3984
3985	/*
3986	* Validate event injection pre-conditions.
3987	*/
3988	if (Event.n.u3Type == SVM_EVENT_EXTERNAL_IRQ)
3989	{
3990	Assert(fAllowInt);
3991	Assert(!fIntShadow);
3992	}
3993	else if (Event.n.u3Type == SVM_EVENT_NMI)
3994	{
3995	Assert(fGif);
3996	Assert(!fIntShadow);
3997	}
3998
3999	/*
4000	* Before injecting an NMI we must set VMCPU_FF_BLOCK_NMIS to prevent nested NMIs. We
4001	* do this only when we are surely going to inject the NMI as otherwise if we return
4002	* to ring-3 prematurely we could leave NMIs blocked indefinitely upon re-entry into
4003	* SVM R0.
4004	*
4005	* With VT-x, this is handled by the Guest interruptibility information VMCS field
4006	* which will set the VMCS field after actually delivering the NMI which we read on
4007	* VM-exit to determine the state.
4008	*/
4009	if ( Event.n.u3Type == SVM_EVENT_NMI
4010	&& Event.n.u8Vector == X86_XCPT_NMI
4011	&& !VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_BLOCK_NMIS))
4012	{
4013	VMCPU_FF_SET(pVCpu, VMCPU_FF_BLOCK_NMIS);
4014	}
4015
4016	/*
4017	* Inject it (update VMCB for injection by the hardware).
4018	*/
4019	Log4(("Injecting pending HM event\n"));
4020	hmR0SvmInjectEventVmcb(pVCpu, pVmcb, &Event);
4021	pVCpu->hm.s.Event.fPending = false;
4022
4023	if (Event.n.u3Type == SVM_EVENT_EXTERNAL_IRQ)
4024	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectInterrupt);
4025	else
4026	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectXcpt);
4027	}
4028	else
4029	Assert(pVmcb->ctrl.EventInject.n.u1Valid == 0);
4030
4031	/*
4032	* We could have injected an NMI through IEM and continue guest execution using
4033	* hardware-assisted SVM. In which case, we would not have any events pending (above)
4034	* but we still need to intercept IRET in order to eventually clear NMI inhibition.
4035	*/
4036	if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_BLOCK_NMIS))
4037	hmR0SvmSetCtrlIntercept(pVmcb, SVM_CTRL_INTERCEPT_IRET);
4038
4039	/*
4040	* Update the guest interrupt shadow in the guest (or nested-guest) VMCB.
4041	*
4042	* For nested-guests: We need to update it too for the scenario where IEM executes
4043	* the nested-guest but execution later continues here with an interrupt shadow active.
4044	*/
4045	pVmcb->ctrl.IntShadow.n.u1IntShadow = fIntShadow;
4046	}
4047
4048
4049	/**
4050	* Reports world-switch error and dumps some useful debug info.
4051	*
4052	* @param pVCpu The cross context virtual CPU structure.
4053	* @param rcVMRun The return code from VMRUN (or
4054	* VERR_SVM_INVALID_GUEST_STATE for invalid
4055	* guest-state).
4056	*/
4057	static void hmR0SvmReportWorldSwitchError(PVMCPU pVCpu, int rcVMRun)
4058	{
4059	HMSVM_ASSERT_PREEMPT_SAFE(pVCpu);
4060	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(&pVCpu->cpum.GstCtx);
4061	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
4062
4063	if (rcVMRun == VERR_SVM_INVALID_GUEST_STATE)
4064	{
4065	#ifdef VBOX_STRICT
4066	hmR0DumpRegs(pVCpu);
4067	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
4068	Log4(("ctrl.u32VmcbCleanBits %#RX32\n", pVmcb->ctrl.u32VmcbCleanBits));
4069	Log4(("ctrl.u16InterceptRdCRx %#x\n", pVmcb->ctrl.u16InterceptRdCRx));
4070	Log4(("ctrl.u16InterceptWrCRx %#x\n", pVmcb->ctrl.u16InterceptWrCRx));
4071	Log4(("ctrl.u16InterceptRdDRx %#x\n", pVmcb->ctrl.u16InterceptRdDRx));
4072	Log4(("ctrl.u16InterceptWrDRx %#x\n", pVmcb->ctrl.u16InterceptWrDRx));
4073	Log4(("ctrl.u32InterceptXcpt %#x\n", pVmcb->ctrl.u32InterceptXcpt));
4074	Log4(("ctrl.u64InterceptCtrl %#RX64\n", pVmcb->ctrl.u64InterceptCtrl));
4075	Log4(("ctrl.u64IOPMPhysAddr %#RX64\n", pVmcb->ctrl.u64IOPMPhysAddr));
4076	Log4(("ctrl.u64MSRPMPhysAddr %#RX64\n", pVmcb->ctrl.u64MSRPMPhysAddr));
4077	Log4(("ctrl.u64TSCOffset %#RX64\n", pVmcb->ctrl.u64TSCOffset));
4078
4079	Log4(("ctrl.TLBCtrl.u32ASID %#x\n", pVmcb->ctrl.TLBCtrl.n.u32ASID));
4080	Log4(("ctrl.TLBCtrl.u8TLBFlush %#x\n", pVmcb->ctrl.TLBCtrl.n.u8TLBFlush));
4081	Log4(("ctrl.TLBCtrl.u24Reserved %#x\n", pVmcb->ctrl.TLBCtrl.n.u24Reserved));
4082
4083	Log4(("ctrl.IntCtrl.u8VTPR %#x\n", pVmcb->ctrl.IntCtrl.n.u8VTPR));
4084	Log4(("ctrl.IntCtrl.u1VIrqPending %#x\n", pVmcb->ctrl.IntCtrl.n.u1VIrqPending));
4085	Log4(("ctrl.IntCtrl.u1VGif %#x\n", pVmcb->ctrl.IntCtrl.n.u1VGif));
4086	Log4(("ctrl.IntCtrl.u6Reserved0 %#x\n", pVmcb->ctrl.IntCtrl.n.u6Reserved));
4087	Log4(("ctrl.IntCtrl.u4VIntrPrio %#x\n", pVmcb->ctrl.IntCtrl.n.u4VIntrPrio));
4088	Log4(("ctrl.IntCtrl.u1IgnoreTPR %#x\n", pVmcb->ctrl.IntCtrl.n.u1IgnoreTPR));
4089	Log4(("ctrl.IntCtrl.u3Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u3Reserved));
4090	Log4(("ctrl.IntCtrl.u1VIntrMasking %#x\n", pVmcb->ctrl.IntCtrl.n.u1VIntrMasking));
4091	Log4(("ctrl.IntCtrl.u1VGifEnable %#x\n", pVmcb->ctrl.IntCtrl.n.u1VGifEnable));
4092	Log4(("ctrl.IntCtrl.u5Reserved1 %#x\n", pVmcb->ctrl.IntCtrl.n.u5Reserved));
4093	Log4(("ctrl.IntCtrl.u8VIntrVector %#x\n", pVmcb->ctrl.IntCtrl.n.u8VIntrVector));
4094	Log4(("ctrl.IntCtrl.u24Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u24Reserved));
4095
4096	Log4(("ctrl.IntShadow.u1IntShadow %#x\n", pVmcb->ctrl.IntShadow.n.u1IntShadow));
4097	Log4(("ctrl.IntShadow.u1GuestIntMask %#x\n", pVmcb->ctrl.IntShadow.n.u1GuestIntMask));
4098	Log4(("ctrl.u64ExitCode %#RX64\n", pVmcb->ctrl.u64ExitCode));
4099	Log4(("ctrl.u64ExitInfo1 %#RX64\n", pVmcb->ctrl.u64ExitInfo1));
4100	Log4(("ctrl.u64ExitInfo2 %#RX64\n", pVmcb->ctrl.u64ExitInfo2));
4101	Log4(("ctrl.ExitIntInfo.u8Vector %#x\n", pVmcb->ctrl.ExitIntInfo.n.u8Vector));
4102	Log4(("ctrl.ExitIntInfo.u3Type %#x\n", pVmcb->ctrl.ExitIntInfo.n.u3Type));
4103	Log4(("ctrl.ExitIntInfo.u1ErrorCodeValid %#x\n", pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid));
4104	Log4(("ctrl.ExitIntInfo.u19Reserved %#x\n", pVmcb->ctrl.ExitIntInfo.n.u19Reserved));
4105	Log4(("ctrl.ExitIntInfo.u1Valid %#x\n", pVmcb->ctrl.ExitIntInfo.n.u1Valid));
4106	Log4(("ctrl.ExitIntInfo.u32ErrorCode %#x\n", pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode));
4107	Log4(("ctrl.NestedPagingCtrl.u1NestedPaging %#x\n", pVmcb->ctrl.NestedPagingCtrl.n.u1NestedPaging));
4108	Log4(("ctrl.NestedPagingCtrl.u1Sev %#x\n", pVmcb->ctrl.NestedPagingCtrl.n.u1Sev));
4109	Log4(("ctrl.NestedPagingCtrl.u1SevEs %#x\n", pVmcb->ctrl.NestedPagingCtrl.n.u1SevEs));
4110	Log4(("ctrl.EventInject.u8Vector %#x\n", pVmcb->ctrl.EventInject.n.u8Vector));
4111	Log4(("ctrl.EventInject.u3Type %#x\n", pVmcb->ctrl.EventInject.n.u3Type));
4112	Log4(("ctrl.EventInject.u1ErrorCodeValid %#x\n", pVmcb->ctrl.EventInject.n.u1ErrorCodeValid));
4113	Log4(("ctrl.EventInject.u19Reserved %#x\n", pVmcb->ctrl.EventInject.n.u19Reserved));
4114	Log4(("ctrl.EventInject.u1Valid %#x\n", pVmcb->ctrl.EventInject.n.u1Valid));
4115	Log4(("ctrl.EventInject.u32ErrorCode %#x\n", pVmcb->ctrl.EventInject.n.u32ErrorCode));
4116
4117	Log4(("ctrl.u64NestedPagingCR3 %#RX64\n", pVmcb->ctrl.u64NestedPagingCR3));
4118
4119	Log4(("ctrl.LbrVirt.u1LbrVirt %#x\n", pVmcb->ctrl.LbrVirt.n.u1LbrVirt));
4120	Log4(("ctrl.LbrVirt.u1VirtVmsaveVmload %#x\n", pVmcb->ctrl.LbrVirt.n.u1VirtVmsaveVmload));
4121
4122	Log4(("guest.CS.u16Sel %RTsel\n", pVmcb->guest.CS.u16Sel));
4123	Log4(("guest.CS.u16Attr %#x\n", pVmcb->guest.CS.u16Attr));
4124	Log4(("guest.CS.u32Limit %#RX32\n", pVmcb->guest.CS.u32Limit));
4125	Log4(("guest.CS.u64Base %#RX64\n", pVmcb->guest.CS.u64Base));
4126	Log4(("guest.DS.u16Sel %#RTsel\n", pVmcb->guest.DS.u16Sel));
4127	Log4(("guest.DS.u16Attr %#x\n", pVmcb->guest.DS.u16Attr));
4128	Log4(("guest.DS.u32Limit %#RX32\n", pVmcb->guest.DS.u32Limit));
4129	Log4(("guest.DS.u64Base %#RX64\n", pVmcb->guest.DS.u64Base));
4130	Log4(("guest.ES.u16Sel %RTsel\n", pVmcb->guest.ES.u16Sel));
4131	Log4(("guest.ES.u16Attr %#x\n", pVmcb->guest.ES.u16Attr));
4132	Log4(("guest.ES.u32Limit %#RX32\n", pVmcb->guest.ES.u32Limit));
4133	Log4(("guest.ES.u64Base %#RX64\n", pVmcb->guest.ES.u64Base));
4134	Log4(("guest.FS.u16Sel %RTsel\n", pVmcb->guest.FS.u16Sel));
4135	Log4(("guest.FS.u16Attr %#x\n", pVmcb->guest.FS.u16Attr));
4136	Log4(("guest.FS.u32Limit %#RX32\n", pVmcb->guest.FS.u32Limit));
4137	Log4(("guest.FS.u64Base %#RX64\n", pVmcb->guest.FS.u64Base));
4138	Log4(("guest.GS.u16Sel %RTsel\n", pVmcb->guest.GS.u16Sel));
4139	Log4(("guest.GS.u16Attr %#x\n", pVmcb->guest.GS.u16Attr));
4140	Log4(("guest.GS.u32Limit %#RX32\n", pVmcb->guest.GS.u32Limit));
4141	Log4(("guest.GS.u64Base %#RX64\n", pVmcb->guest.GS.u64Base));
4142
4143	Log4(("guest.GDTR.u32Limit %#RX32\n", pVmcb->guest.GDTR.u32Limit));
4144	Log4(("guest.GDTR.u64Base %#RX64\n", pVmcb->guest.GDTR.u64Base));
4145
4146	Log4(("guest.LDTR.u16Sel %RTsel\n", pVmcb->guest.LDTR.u16Sel));
4147	Log4(("guest.LDTR.u16Attr %#x\n", pVmcb->guest.LDTR.u16Attr));
4148	Log4(("guest.LDTR.u32Limit %#RX32\n", pVmcb->guest.LDTR.u32Limit));
4149	Log4(("guest.LDTR.u64Base %#RX64\n", pVmcb->guest.LDTR.u64Base));
4150
4151	Log4(("guest.IDTR.u32Limit %#RX32\n", pVmcb->guest.IDTR.u32Limit));
4152	Log4(("guest.IDTR.u64Base %#RX64\n", pVmcb->guest.IDTR.u64Base));
4153
4154	Log4(("guest.TR.u16Sel %RTsel\n", pVmcb->guest.TR.u16Sel));
4155	Log4(("guest.TR.u16Attr %#x\n", pVmcb->guest.TR.u16Attr));
4156	Log4(("guest.TR.u32Limit %#RX32\n", pVmcb->guest.TR.u32Limit));
4157	Log4(("guest.TR.u64Base %#RX64\n", pVmcb->guest.TR.u64Base));
4158
4159	Log4(("guest.u8CPL %#x\n", pVmcb->guest.u8CPL));
4160	Log4(("guest.u64CR0 %#RX64\n", pVmcb->guest.u64CR0));
4161	Log4(("guest.u64CR2 %#RX64\n", pVmcb->guest.u64CR2));
4162	Log4(("guest.u64CR3 %#RX64\n", pVmcb->guest.u64CR3));
4163	Log4(("guest.u64CR4 %#RX64\n", pVmcb->guest.u64CR4));
4164	Log4(("guest.u64DR6 %#RX64\n", pVmcb->guest.u64DR6));
4165	Log4(("guest.u64DR7 %#RX64\n", pVmcb->guest.u64DR7));
4166
4167	Log4(("guest.u64RIP %#RX64\n", pVmcb->guest.u64RIP));
4168	Log4(("guest.u64RSP %#RX64\n", pVmcb->guest.u64RSP));
4169	Log4(("guest.u64RAX %#RX64\n", pVmcb->guest.u64RAX));
4170	Log4(("guest.u64RFlags %#RX64\n", pVmcb->guest.u64RFlags));
4171
4172	Log4(("guest.u64SysEnterCS %#RX64\n", pVmcb->guest.u64SysEnterCS));
4173	Log4(("guest.u64SysEnterEIP %#RX64\n", pVmcb->guest.u64SysEnterEIP));
4174	Log4(("guest.u64SysEnterESP %#RX64\n", pVmcb->guest.u64SysEnterESP));
4175
4176	Log4(("guest.u64EFER %#RX64\n", pVmcb->guest.u64EFER));
4177	Log4(("guest.u64STAR %#RX64\n", pVmcb->guest.u64STAR));
4178	Log4(("guest.u64LSTAR %#RX64\n", pVmcb->guest.u64LSTAR));
4179	Log4(("guest.u64CSTAR %#RX64\n", pVmcb->guest.u64CSTAR));
4180	Log4(("guest.u64SFMASK %#RX64\n", pVmcb->guest.u64SFMASK));
4181	Log4(("guest.u64KernelGSBase %#RX64\n", pVmcb->guest.u64KernelGSBase));
4182	Log4(("guest.u64PAT %#RX64\n", pVmcb->guest.u64PAT));
4183	Log4(("guest.u64DBGCTL %#RX64\n", pVmcb->guest.u64DBGCTL));
4184	Log4(("guest.u64BR_FROM %#RX64\n", pVmcb->guest.u64BR_FROM));
4185	Log4(("guest.u64BR_TO %#RX64\n", pVmcb->guest.u64BR_TO));
4186	Log4(("guest.u64LASTEXCPFROM %#RX64\n", pVmcb->guest.u64LASTEXCPFROM));
4187	Log4(("guest.u64LASTEXCPTO %#RX64\n", pVmcb->guest.u64LASTEXCPTO));
4188
4189	NOREF(pVmcb);
4190	#endif /* VBOX_STRICT */
4191	}
4192	else
4193	Log4Func(("rcVMRun=%d\n", rcVMRun));
4194	}
4195
4196
4197	/**
4198	* Check per-VM and per-VCPU force flag actions that require us to go back to
4199	* ring-3 for one reason or another.
4200	*
4201	* @returns VBox status code (information status code included).
4202	* @retval VINF_SUCCESS if we don't have any actions that require going back to
4203	* ring-3.
4204	* @retval VINF_PGM_SYNC_CR3 if we have pending PGM CR3 sync.
4205	* @retval VINF_EM_PENDING_REQUEST if we have pending requests (like hardware
4206	* interrupts)
4207	* @retval VINF_PGM_POOL_FLUSH_PENDING if PGM is doing a pool flush and requires
4208	* all EMTs to be in ring-3.
4209	* @retval VINF_EM_RAW_TO_R3 if there is pending DMA requests.
4210	* @retval VINF_EM_NO_MEMORY PGM is out of memory, we need to return
4211	* to the EM loop.
4212	*
4213	* @param pVCpu The cross context virtual CPU structure.
4214	*/
4215	static int hmR0SvmCheckForceFlags(PVMCPU pVCpu)
4216	{
4217	Assert(VMMRZCallRing3IsEnabled(pVCpu));
4218	Assert(!VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_UPDATE_PAE_PDPES));
4219
4220	/* Could happen as a result of longjump. */
4221	if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_UPDATE_CR3))
4222	PGMUpdateCR3(pVCpu, CPUMGetGuestCR3(pVCpu));
4223
4224	/* Update pending interrupts into the APIC's IRR. */
4225	if (VMCPU_FF_TEST_AND_CLEAR(pVCpu, VMCPU_FF_UPDATE_APIC))
4226	APICUpdatePendingInterrupts(pVCpu);
4227
4228	PVM pVM = pVCpu->CTX_SUFF(pVM);
4229	if ( VM_FF_IS_PENDING(pVM, !pVCpu->hm.s.fSingleInstruction
4230	? VM_FF_HP_R0_PRE_HM_MASK : VM_FF_HP_R0_PRE_HM_STEP_MASK)
4231	\|\| VMCPU_FF_IS_PENDING(pVCpu, !pVCpu->hm.s.fSingleInstruction
4232	? VMCPU_FF_HP_R0_PRE_HM_MASK : VMCPU_FF_HP_R0_PRE_HM_STEP_MASK) )
4233	{
4234	/* Pending PGM C3 sync. */
4235	if (VMCPU_FF_IS_PENDING(pVCpu,VMCPU_FF_PGM_SYNC_CR3 \| VMCPU_FF_PGM_SYNC_CR3_NON_GLOBAL))
4236	{
4237	int rc = PGMSyncCR3(pVCpu, pVCpu->cpum.GstCtx.cr0, pVCpu->cpum.GstCtx.cr3, pVCpu->cpum.GstCtx.cr4,
4238	VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_PGM_SYNC_CR3));
4239	if (rc != VINF_SUCCESS)
4240	{
4241	Log4Func(("PGMSyncCR3 forcing us back to ring-3. rc=%d\n", rc));
4242	return rc;
4243	}
4244	}
4245
4246	/* Pending HM-to-R3 operations (critsects, timers, EMT rendezvous etc.) */
4247	/* -XXX- what was that about single stepping? */
4248	if ( VM_FF_IS_PENDING(pVM, VM_FF_HM_TO_R3_MASK)
4249	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_TO_R3_MASK))
4250	{
4251	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHmToR3FF);
4252	int rc = RT_UNLIKELY(VM_FF_IS_PENDING(pVM, VM_FF_PGM_NO_MEMORY)) ? VINF_EM_NO_MEMORY : VINF_EM_RAW_TO_R3;
4253	Log4Func(("HM_TO_R3 forcing us back to ring-3. rc=%d\n", rc));
4254	return rc;
4255	}
4256
4257	/* Pending VM request packets, such as hardware interrupts. */
4258	if ( VM_FF_IS_PENDING(pVM, VM_FF_REQUEST)
4259	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_REQUEST))
4260	{
4261	Log4Func(("Pending VM request forcing us back to ring-3\n"));
4262	return VINF_EM_PENDING_REQUEST;
4263	}
4264
4265	/* Pending PGM pool flushes. */
4266	if (VM_FF_IS_PENDING(pVM, VM_FF_PGM_POOL_FLUSH_PENDING))
4267	{
4268	Log4Func(("PGM pool flush pending forcing us back to ring-3\n"));
4269	return VINF_PGM_POOL_FLUSH_PENDING;
4270	}
4271
4272	/* Pending DMA requests. */
4273	if (VM_FF_IS_PENDING(pVM, VM_FF_PDM_DMA))
4274	{
4275	Log4Func(("Pending DMA request forcing us back to ring-3\n"));
4276	return VINF_EM_RAW_TO_R3;
4277	}
4278	}
4279
4280	return VINF_SUCCESS;
4281	}
4282
4283
4284	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
4285	/**
4286	* Does the preparations before executing nested-guest code in AMD-V.
4287	*
4288	* @returns VBox status code (informational status codes included).
4289	* @retval VINF_SUCCESS if we can proceed with running the guest.
4290	* @retval VINF_* scheduling changes, we have to go back to ring-3.
4291	*
4292	* @param pVCpu The cross context virtual CPU structure.
4293	* @param pSvmTransient Pointer to the SVM transient structure.
4294	*
4295	* @remarks Same caveats regarding longjumps as hmR0SvmPreRunGuest applies.
4296	* @sa hmR0SvmPreRunGuest.
4297	*/
4298	static int hmR0SvmPreRunGuestNested(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
4299	{
4300	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
4301	HMSVM_ASSERT_PREEMPT_SAFE(pVCpu);
4302	HMSVM_ASSERT_IN_NESTED_GUEST(pCtx);
4303
4304	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM_ONLY_IN_IEM
4305	Log2(("hmR0SvmPreRunGuest: Rescheduling to IEM due to nested-hwvirt or forced IEM exec -> VINF_EM_RESCHEDULE_REM\n"));
4306	return VINF_EM_RESCHEDULE_REM;
4307	#endif
4308
4309	/* Check force flag actions that might require us to go back to ring-3. */
4310	int rc = hmR0SvmCheckForceFlags(pVCpu);
4311	if (rc != VINF_SUCCESS)
4312	return rc;
4313
4314	if (TRPMHasTrap(pVCpu))
4315	hmR0SvmTrpmTrapToPendingEvent(pVCpu);
4316	else if (!pVCpu->hm.s.Event.fPending)
4317	{
4318	VBOXSTRICTRC rcStrict = hmR0SvmEvaluatePendingEventNested(pVCpu);
4319	if ( rcStrict != VINF_SUCCESS
4320	\|\| !CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
4321	return VBOXSTRICTRC_VAL(rcStrict);
4322	}
4323
4324	HMSVM_ASSERT_IN_NESTED_GUEST(pCtx);
4325
4326	/*
4327	* On the oldest AMD-V systems, we may not get enough information to reinject an NMI.
4328	* Just do it in software, see @bugref{8411}.
4329	* NB: If we could continue a task switch exit we wouldn't need to do this.
4330	*/
4331	PVM pVM = pVCpu->CTX_SUFF(pVM);
4332	if (RT_UNLIKELY( !pVM->hm.s.svm.u32Features
4333	&& pVCpu->hm.s.Event.fPending
4334	&& SVM_EVENT_GET_TYPE(pVCpu->hm.s.Event.u64IntInfo) == SVM_EVENT_NMI))
4335	{
4336	return VINF_EM_RAW_INJECT_TRPM_EVENT;
4337	}
4338
4339	#ifdef HMSVM_SYNC_FULL_GUEST_STATE
4340	Assert(!(pCtx->fExtrn & HMSVM_CPUMCTX_EXTRN_ALL));
4341	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
4342	#endif
4343
4344	/*
4345	* Export the nested-guest state bits that are not shared with the host in any way as we
4346	* can longjmp or get preempted in the midst of exporting some of the state.
4347	*/
4348	rc = hmR0SvmExportGuestStateNested(pVCpu);
4349	AssertRCReturn(rc, rc);
4350	STAM_COUNTER_INC(&pVCpu->hm.s.StatExportFull);
4351
4352	/* Ensure we've cached (and hopefully modified) the VMCB for execution using hardware-assisted SVM. */
4353	Assert(pVCpu->hm.s.svm.NstGstVmcbCache.fCacheValid);
4354
4355	/*
4356	* No longjmps to ring-3 from this point on!!!
4357	*
4358	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional,
4359	* better than a kernel panic. This also disables flushing of the R0-logger instance.
4360	*/
4361	VMMRZCallRing3Disable(pVCpu);
4362
4363	/*
4364	* We disable interrupts so that we don't miss any interrupts that would flag preemption
4365	* (IPI/timers etc.) when thread-context hooks aren't used and we've been running with
4366	* preemption disabled for a while. Since this is purly to aid the
4367	* RTThreadPreemptIsPending() code, it doesn't matter that it may temporarily reenable and
4368	* disable interrupt on NT.
4369	*
4370	* We need to check for force-flags that could've possible been altered since we last
4371	* checked them (e.g. by PDMGetInterrupt() leaving the PDM critical section,
4372	* see @bugref{6398}).
4373	*
4374	* We also check a couple of other force-flags as a last opportunity to get the EMT back
4375	* to ring-3 before executing guest code.
4376	*/
4377	pSvmTransient->fEFlags = ASMIntDisableFlags();
4378	if ( VM_FF_IS_PENDING(pVM, VM_FF_EMT_RENDEZVOUS \| VM_FF_TM_VIRTUAL_SYNC)
4379	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_TO_R3_MASK))
4380	{
4381	ASMSetFlags(pSvmTransient->fEFlags);
4382	VMMRZCallRing3Enable(pVCpu);
4383	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHmToR3FF);
4384	return VINF_EM_RAW_TO_R3;
4385	}
4386	if (RTThreadPreemptIsPending(NIL_RTTHREAD))
4387	{
4388	ASMSetFlags(pSvmTransient->fEFlags);
4389	VMMRZCallRing3Enable(pVCpu);
4390	STAM_COUNTER_INC(&pVCpu->hm.s.StatPendingHostIrq);
4391	return VINF_EM_RAW_INTERRUPT;
4392	}
4393	return VINF_SUCCESS;
4394	}
4395	#endif
4396
4397
4398	/**
4399	* Does the preparations before executing guest code in AMD-V.
4400	*
4401	* This may cause longjmps to ring-3 and may even result in rescheduling to the
4402	* recompiler. We must be cautious what we do here regarding committing
4403	* guest-state information into the VMCB assuming we assuredly execute the guest
4404	* in AMD-V. If we fall back to the recompiler after updating the VMCB and
4405	* clearing the common-state (TRPM/forceflags), we must undo those changes so
4406	* that the recompiler can (and should) use them when it resumes guest
4407	* execution. Otherwise such operations must be done when we can no longer
4408	* exit to ring-3.
4409	*
4410	* @returns VBox status code (informational status codes included).
4411	* @retval VINF_SUCCESS if we can proceed with running the guest.
4412	* @retval VINF_* scheduling changes, we have to go back to ring-3.
4413	*
4414	* @param pVCpu The cross context virtual CPU structure.
4415	* @param pSvmTransient Pointer to the SVM transient structure.
4416	*/
4417	static int hmR0SvmPreRunGuest(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
4418	{
4419	HMSVM_ASSERT_PREEMPT_SAFE(pVCpu);
4420	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(&pVCpu->cpum.GstCtx);
4421
4422	/* Check force flag actions that might require us to go back to ring-3. */
4423	int rc = hmR0SvmCheckForceFlags(pVCpu);
4424	if (rc != VINF_SUCCESS)
4425	return rc;
4426
4427	if (TRPMHasTrap(pVCpu))
4428	hmR0SvmTrpmTrapToPendingEvent(pVCpu);
4429	else if (!pVCpu->hm.s.Event.fPending)
4430	hmR0SvmEvaluatePendingEvent(pVCpu);
4431
4432	/*
4433	* On the oldest AMD-V systems, we may not get enough information to reinject an NMI.
4434	* Just do it in software, see @bugref{8411}.
4435	* NB: If we could continue a task switch exit we wouldn't need to do this.
4436	*/
4437	PVM pVM = pVCpu->CTX_SUFF(pVM);
4438	if (RT_UNLIKELY(pVCpu->hm.s.Event.fPending && (((pVCpu->hm.s.Event.u64IntInfo >> 8) & 7) == SVM_EVENT_NMI)))
4439	if (RT_UNLIKELY(!pVM->hm.s.svm.u32Features))
4440	return VINF_EM_RAW_INJECT_TRPM_EVENT;
4441
4442	#ifdef HMSVM_SYNC_FULL_GUEST_STATE
4443	Assert(!(pVCpu->cpum.GstCtx->fExtrn & HMSVM_CPUMCTX_EXTRN_ALL));
4444	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
4445	#endif
4446
4447	/*
4448	* Export the guest state bits that are not shared with the host in any way as we can
4449	* longjmp or get preempted in the midst of exporting some of the state.
4450	*/
4451	rc = hmR0SvmExportGuestState(pVCpu);
4452	AssertRCReturn(rc, rc);
4453	STAM_COUNTER_INC(&pVCpu->hm.s.StatExportFull);
4454
4455	/*
4456	* If we're not intercepting TPR changes in the guest, save the guest TPR before the
4457	* world-switch so we can update it on the way back if the guest changed the TPR.
4458	*/
4459	if (pVCpu->hm.s.svm.fSyncVTpr)
4460	{
4461	PCSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
4462	if (pVM->hm.s.fTPRPatchingActive)
4463	pSvmTransient->u8GuestTpr = pVmcb->guest.u64LSTAR;
4464	else
4465	pSvmTransient->u8GuestTpr = pVmcb->ctrl.IntCtrl.n.u8VTPR;
4466	}
4467
4468	/*
4469	* No longjmps to ring-3 from this point on!!!
4470	*
4471	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional,
4472	* better than a kernel panic. This also disables flushing of the R0-logger instance.
4473	*/
4474	VMMRZCallRing3Disable(pVCpu);
4475
4476	/*
4477	* We disable interrupts so that we don't miss any interrupts that would flag preemption
4478	* (IPI/timers etc.) when thread-context hooks aren't used and we've been running with
4479	* preemption disabled for a while. Since this is purly to aid the
4480	* RTThreadPreemptIsPending() code, it doesn't matter that it may temporarily reenable and
4481	* disable interrupt on NT.
4482	*
4483	* We need to check for force-flags that could've possible been altered since we last
4484	* checked them (e.g. by PDMGetInterrupt() leaving the PDM critical section,
4485	* see @bugref{6398}).
4486	*
4487	* We also check a couple of other force-flags as a last opportunity to get the EMT back
4488	* to ring-3 before executing guest code.
4489	*/
4490	pSvmTransient->fEFlags = ASMIntDisableFlags();
4491	if ( VM_FF_IS_PENDING(pVM, VM_FF_EMT_RENDEZVOUS \| VM_FF_TM_VIRTUAL_SYNC)
4492	\|\| VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_TO_R3_MASK))
4493	{
4494	ASMSetFlags(pSvmTransient->fEFlags);
4495	VMMRZCallRing3Enable(pVCpu);
4496	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHmToR3FF);
4497	return VINF_EM_RAW_TO_R3;
4498	}
4499	if (RTThreadPreemptIsPending(NIL_RTTHREAD))
4500	{
4501	ASMSetFlags(pSvmTransient->fEFlags);
4502	VMMRZCallRing3Enable(pVCpu);
4503	STAM_COUNTER_INC(&pVCpu->hm.s.StatPendingHostIrq);
4504	return VINF_EM_RAW_INTERRUPT;
4505	}
4506
4507	return VINF_SUCCESS;
4508	}
4509
4510
4511	/**
4512	* Prepares to run guest (or nested-guest) code in AMD-V and we've committed to
4513	* doing so.
4514	*
4515	* This means there is no backing out to ring-3 or anywhere else at this point.
4516	*
4517	* @param pVCpu The cross context virtual CPU structure.
4518	* @param pSvmTransient Pointer to the SVM transient structure.
4519	*
4520	* @remarks Called with preemption disabled.
4521	* @remarks No-long-jump zone!!!
4522	*/
4523	static void hmR0SvmPreRunGuestCommitted(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
4524	{
4525	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
4526	Assert(VMMR0IsLogFlushDisabled(pVCpu));
4527	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
4528
4529	VMCPU_ASSERT_STATE(pVCpu, VMCPUSTATE_STARTED_HM);
4530	VMCPU_SET_STATE(pVCpu, VMCPUSTATE_STARTED_EXEC); /* Indicate the start of guest execution. */
4531
4532	PVM pVM = pVCpu->CTX_SUFF(pVM);
4533	PSVMVMCB pVmcb = pSvmTransient->pVmcb;
4534
4535	hmR0SvmInjectPendingEvent(pVCpu, pVmcb);
4536
4537	if (!CPUMIsGuestFPUStateActive(pVCpu))
4538	{
4539	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatLoadGuestFpuState, x);
4540	CPUMR0LoadGuestFPU(pVM, pVCpu); /* (Ignore rc, no need to set HM_CHANGED_HOST_CONTEXT for SVM.) */
4541	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatLoadGuestFpuState, x);
4542	STAM_COUNTER_INC(&pVCpu->hm.s.StatLoadGuestFpu);
4543	}
4544
4545	/* Load the state shared between host and guest (FPU, debug). */
4546	if (pVCpu->hm.s.fCtxChanged & HM_CHANGED_SVM_HOST_GUEST_SHARED_STATE)
4547	hmR0SvmExportSharedState(pVCpu, pVmcb);
4548
4549	pVCpu->hm.s.fCtxChanged &= ~HM_CHANGED_HOST_CONTEXT; /* Preemption might set this, nothing to do on AMD-V. */
4550	AssertMsg(!pVCpu->hm.s.fCtxChanged, ("fCtxChanged=%#RX64\n", pVCpu->hm.s.fCtxChanged));
4551
4552	PHMGLOBALCPUINFO pHostCpu = hmR0GetCurrentCpu();
4553	RTCPUID const idHostCpu = pHostCpu->idCpu;
4554	bool const fMigratedHostCpu = idHostCpu != pVCpu->hm.s.idLastCpu;
4555
4556	/* Setup TSC offsetting. */
4557	if ( pSvmTransient->fUpdateTscOffsetting
4558	\|\| fMigratedHostCpu)
4559	{
4560	hmR0SvmUpdateTscOffsetting(pVCpu, pVmcb);
4561	pSvmTransient->fUpdateTscOffsetting = false;
4562	}
4563
4564	/* If we've migrating CPUs, mark the VMCB Clean bits as dirty. */
4565	if (fMigratedHostCpu)
4566	pVmcb->ctrl.u32VmcbCleanBits = 0;
4567
4568	/* Store status of the shared guest-host state at the time of VMRUN. */
4569	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS)
4570	if (CPUMIsGuestInLongModeEx(&pVCpu->cpum.GstCtx))
4571	{
4572	pSvmTransient->fWasGuestDebugStateActive = CPUMIsGuestDebugStateActivePending(pVCpu);
4573	pSvmTransient->fWasHyperDebugStateActive = CPUMIsHyperDebugStateActivePending(pVCpu);
4574	}
4575	else
4576	#endif
4577	{
4578	pSvmTransient->fWasGuestDebugStateActive = CPUMIsGuestDebugStateActive(pVCpu);
4579	pSvmTransient->fWasHyperDebugStateActive = CPUMIsHyperDebugStateActive(pVCpu);
4580	}
4581
4582	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
4583	uint8_t *pbMsrBitmap;
4584	if (!pSvmTransient->fIsNestedGuest)
4585	pbMsrBitmap = (uint8_t *)pVCpu->hm.s.svm.pvMsrBitmap;
4586	else
4587	{
4588	hmR0SvmMergeMsrpmNested(pHostCpu, pVCpu);
4589
4590	/* Update the nested-guest VMCB with the newly merged MSRPM (clean bits updated below). */
4591	pVmcb->ctrl.u64MSRPMPhysAddr = pHostCpu->n.svm.HCPhysNstGstMsrpm;
4592	pbMsrBitmap = (uint8_t *)pHostCpu->n.svm.pvNstGstMsrpm;
4593	}
4594	#else
4595	uint8_t pbMsrBitmap = (uint8_t )pVCpu->hm.s.svm.pvMsrBitmap;
4596	#endif
4597
4598	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, true); /* Used for TLB flushing, set this across the world switch. */
4599	/* Flush the appropriate tagged-TLB entries. */
4600	hmR0SvmFlushTaggedTlb(pVCpu, pVmcb, pHostCpu);
4601	Assert(pVCpu->hm.s.idLastCpu == idHostCpu);
4602
4603	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatEntry, &pVCpu->hm.s.StatInGC, x);
4604
4605	TMNotifyStartOfExecution(pVCpu); /* Finally, notify TM to resume its clocks as we're about
4606	to start executing. */
4607
4608	/*
4609	* Save the current Host TSC_AUX and write the guest TSC_AUX to the host, so that RDTSCPs
4610	* (that don't cause exits) reads the guest MSR, see @bugref{3324}.
4611	*
4612	* This should be done -after- any RDTSCPs for obtaining the host timestamp (TM, STAM etc).
4613	*/
4614	if ( (pVM->hm.s.cpuid.u32AMDFeatureEDX & X86_CPUID_EXT_FEATURE_EDX_RDTSCP)
4615	&& !(pVmcb->ctrl.u64InterceptCtrl & SVM_CTRL_INTERCEPT_RDTSCP))
4616	{
4617	uint64_t const uGuestTscAux = CPUMGetGuestTscAux(pVCpu);
4618	pVCpu->hm.s.u64HostTscAux = ASMRdMsr(MSR_K8_TSC_AUX);
4619	if (uGuestTscAux != pVCpu->hm.s.u64HostTscAux)
4620	ASMWrMsr(MSR_K8_TSC_AUX, uGuestTscAux);
4621	hmR0SvmSetMsrPermission(pVCpu, pbMsrBitmap, MSR_K8_TSC_AUX, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
4622	pSvmTransient->fRestoreTscAuxMsr = true;
4623	}
4624	else
4625	{
4626	hmR0SvmSetMsrPermission(pVCpu, pbMsrBitmap, MSR_K8_TSC_AUX, SVMMSREXIT_INTERCEPT_READ, SVMMSREXIT_INTERCEPT_WRITE);
4627	pSvmTransient->fRestoreTscAuxMsr = false;
4628	}
4629	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_IOPM_MSRPM;
4630
4631	/*
4632	* If VMCB Clean bits isn't supported by the CPU or exposed to the guest in the nested
4633	* virtualization case, mark all state-bits as dirty indicating to the CPU to re-load
4634	* from the VMCB.
4635	*/
4636	bool const fSupportsVmcbCleanBits = hmR0SvmSupportsVmcbCleanBits(pVCpu);
4637	if (!fSupportsVmcbCleanBits)
4638	pVmcb->ctrl.u32VmcbCleanBits = 0;
4639	}
4640
4641
4642	/**
4643	* Wrapper for running the guest (or nested-guest) code in AMD-V.
4644	*
4645	* @returns VBox strict status code.
4646	* @param pVCpu The cross context virtual CPU structure.
4647	* @param HCPhysVmcb The host physical address of the VMCB.
4648	*
4649	* @remarks No-long-jump zone!!!
4650	*/
4651	DECLINLINE(int) hmR0SvmRunGuest(PVMCPU pVCpu, RTHCPHYS HCPhysVmcb)
4652	{
4653	/* Mark that HM is the keeper of all guest-CPU registers now that we're going to execute guest code. */
4654	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
4655	pCtx->fExtrn \|= HMSVM_CPUMCTX_EXTRN_ALL \| CPUMCTX_EXTRN_KEEPER_HM;
4656
4657	/*
4658	* 64-bit Windows uses XMM registers in the kernel as the Microsoft compiler expresses
4659	* floating-point operations using SSE instructions. Some XMM registers (XMM6-XMM15) are
4660	* callee-saved and thus the need for this XMM wrapper.
4661	*
4662	* Refer MSDN "Configuring Programs for 64-bit/x64 Software Conventions / Register Usage".
4663	*/
4664	PVM pVM = pVCpu->CTX_SUFF(pVM);
4665	#ifdef VBOX_WITH_KERNEL_USING_XMM
4666	return hmR0SVMRunWrapXMM(pVCpu->hm.s.svm.HCPhysVmcbHost, HCPhysVmcb, pCtx, pVM, pVCpu, pVCpu->hm.s.svm.pfnVMRun);
4667	#else
4668	return pVCpu->hm.s.svm.pfnVMRun(pVCpu->hm.s.svm.HCPhysVmcbHost, HCPhysVmcb, pCtx, pVM, pVCpu);
4669	#endif
4670	}
4671
4672
4673	/**
4674	* Undoes the TSC offset applied for an SVM nested-guest and returns the TSC
4675	* value for the guest.
4676	*
4677	* @returns The TSC offset after undoing any nested-guest TSC offset.
4678	* @param pVCpu The cross context virtual CPU structure of the calling EMT.
4679	* @param uTicks The nested-guest TSC.
4680	*
4681	* @note If you make any changes to this function, please check if
4682	* hmR0SvmNstGstUndoTscOffset() needs adjusting.
4683	*
4684	* @sa HMSvmNstGstApplyTscOffset().
4685	*/
4686	DECLINLINE(uint64_t) hmR0SvmNstGstUndoTscOffset(PVMCPU pVCpu, uint64_t uTicks)
4687	{
4688	PCSVMNESTEDVMCBCACHE pVmcbNstGstCache = &pVCpu->hm.s.svm.NstGstVmcbCache;
4689	Assert(pVmcbNstGstCache->fCacheValid);
4690	return uTicks - pVmcbNstGstCache->u64TSCOffset;
4691	}
4692
4693
4694	/**
4695	* Performs some essential restoration of state after running guest (or
4696	* nested-guest) code in AMD-V.
4697	*
4698	* @param pVCpu The cross context virtual CPU structure.
4699	* @param pSvmTransient Pointer to the SVM transient structure.
4700	* @param rcVMRun Return code of VMRUN.
4701	*
4702	* @remarks Called with interrupts disabled.
4703	* @remarks No-long-jump zone!!! This function will however re-enable longjmps
4704	* unconditionally when it is safe to do so.
4705	*/
4706	static void hmR0SvmPostRunGuest(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient, int rcVMRun)
4707	{
4708	Assert(!VMMRZCallRing3IsEnabled(pVCpu));
4709
4710	uint64_t const uHostTsc = ASMReadTSC(); /* Read the TSC as soon as possible. */
4711	ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, false); /* See HMInvalidatePageOnAllVCpus(): used for TLB flushing. */
4712	ASMAtomicIncU32(&pVCpu->hm.s.cWorldSwitchExits); /* Initialized in vmR3CreateUVM(): used for EMT poking. */
4713
4714	PSVMVMCB pVmcb = pSvmTransient->pVmcb;
4715	PSVMVMCBCTRL pVmcbCtrl = &pVmcb->ctrl;
4716
4717	/* TSC read must be done early for maximum accuracy. */
4718	if (!(pVmcbCtrl->u64InterceptCtrl & SVM_CTRL_INTERCEPT_RDTSC))
4719	{
4720	if (!pSvmTransient->fIsNestedGuest)
4721	TMCpuTickSetLastSeen(pVCpu, uHostTsc + pVmcbCtrl->u64TSCOffset);
4722	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
4723	else
4724	{
4725	/* The nested-guest VMCB TSC offset shall eventually be restored on #VMEXIT via HMSvmNstGstVmExitNotify(). */
4726	uint64_t const uGstTsc = hmR0SvmNstGstUndoTscOffset(pVCpu, uHostTsc + pVmcbCtrl->u64TSCOffset);
4727	TMCpuTickSetLastSeen(pVCpu, uGstTsc);
4728	}
4729	#endif
4730	}
4731
4732	if (pSvmTransient->fRestoreTscAuxMsr)
4733	{
4734	uint64_t u64GuestTscAuxMsr = ASMRdMsr(MSR_K8_TSC_AUX);
4735	CPUMSetGuestTscAux(pVCpu, u64GuestTscAuxMsr);
4736	if (u64GuestTscAuxMsr != pVCpu->hm.s.u64HostTscAux)
4737	ASMWrMsr(MSR_K8_TSC_AUX, pVCpu->hm.s.u64HostTscAux);
4738	}
4739
4740	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatInGC, &pVCpu->hm.s.StatPreExit, x);
4741	TMNotifyEndOfExecution(pVCpu); /* Notify TM that the guest is no longer running. */
4742	VMCPU_SET_STATE(pVCpu, VMCPUSTATE_STARTED_HM);
4743
4744	Assert(!(ASMGetFlags() & X86_EFL_IF));
4745	ASMSetFlags(pSvmTransient->fEFlags); /* Enable interrupts. */
4746	VMMRZCallRing3Enable(pVCpu); /* It is now safe to do longjmps to ring-3!!! */
4747
4748	/* If VMRUN failed, we can bail out early. This does -not- cover SVM_EXIT_INVALID. */
4749	if (RT_UNLIKELY(rcVMRun != VINF_SUCCESS))
4750	{
4751	Log4Func(("VMRUN failure: rcVMRun=%Rrc\n", rcVMRun));
4752	return;
4753	}
4754
4755	pSvmTransient->u64ExitCode = pVmcbCtrl->u64ExitCode; /* Save the #VMEXIT reason. */
4756	pVmcbCtrl->u32VmcbCleanBits = HMSVM_VMCB_CLEAN_ALL; /* Mark the VMCB-state cache as unmodified by VMM. */
4757	pSvmTransient->fVectoringDoublePF = false; /* Vectoring double page-fault needs to be determined later. */
4758	pSvmTransient->fVectoringPF = false; /* Vectoring page-fault needs to be determined later. */
4759
4760	#ifdef HMSVM_SYNC_FULL_GUEST_STATE
4761	hmR0SvmImportGuestState(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
4762	Assert(!(pVCpu->cpum.GstCtx.fExtrn & HMSVM_CPUMCTX_EXTRN_ALL));
4763	#else
4764	/*
4765	* Always import the following:
4766	*
4767	* - RIP for exit optimizations and evaluating event injection on re-entry.
4768	* - RFLAGS for evaluating event injection on VM re-entry and for exporting shared debug
4769	* state on preemption.
4770	* - Interrupt shadow, GIF for evaluating event injection on VM re-entry.
4771	* - CS for exit optimizations.
4772	* - RAX, RSP for simplifying assumptions on GPRs. All other GPRs are swapped by the
4773	* assembly switcher code.
4774	* - Shared state (only DR7 currently) for exporting shared debug state on preemption.
4775	*/
4776	hmR0SvmImportGuestState(pVCpu, CPUMCTX_EXTRN_RIP
4777	\| CPUMCTX_EXTRN_RFLAGS
4778	\| CPUMCTX_EXTRN_RAX
4779	\| CPUMCTX_EXTRN_RSP
4780	\| CPUMCTX_EXTRN_CS
4781	\| CPUMCTX_EXTRN_HWVIRT
4782	\| CPUMCTX_EXTRN_HM_SVM_INT_SHADOW
4783	\| CPUMCTX_EXTRN_HM_SVM_HWVIRT_VIRQ
4784	\| HMSVM_CPUMCTX_SHARED_STATE);
4785	#endif
4786
4787	if ( pSvmTransient->u64ExitCode != SVM_EXIT_INVALID
4788	&& pVCpu->hm.s.svm.fSyncVTpr)
4789	{
4790	Assert(!pSvmTransient->fIsNestedGuest);
4791	/* TPR patching (for 32-bit guests) uses LSTAR MSR for holding the TPR value, otherwise uses the VTPR. */
4792	if ( pVCpu->CTX_SUFF(pVM)->hm.s.fTPRPatchingActive
4793	&& (pVmcb->guest.u64LSTAR & 0xff) != pSvmTransient->u8GuestTpr)
4794	{
4795	int rc = APICSetTpr(pVCpu, pVmcb->guest.u64LSTAR & 0xff);
4796	AssertRC(rc);
4797	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
4798	}
4799	/* Sync TPR when we aren't intercepting CR8 writes. */
4800	else if (pSvmTransient->u8GuestTpr != pVmcbCtrl->IntCtrl.n.u8VTPR)
4801	{
4802	int rc = APICSetTpr(pVCpu, pVmcbCtrl->IntCtrl.n.u8VTPR << 4);
4803	AssertRC(rc);
4804	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
4805	}
4806	}
4807
4808	#ifdef DEBUG_ramshankar
4809	if (CPUMIsGuestInSvmNestedHwVirtMode(&pVCpu->cpum.GstCtx))
4810	{
4811	hmR0SvmImportGuestState(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
4812	hmR0SvmLogState(pVCpu, pVmcb, pVCpu->cpum.GstCtx, "hmR0SvmPostRunGuestNested", HMSVM_LOG_ALL & ~HMSVM_LOG_LBR,
4813	0 /* uVerbose */);
4814	}
4815	#endif
4816
4817	HMSVM_CPUMCTX_ASSERT(pVCpu, CPUMCTX_EXTRN_CS \| CPUMCTX_EXTRN_RIP);
4818	EMHistoryAddExit(pVCpu, EMEXIT_MAKE_FT(EMEXIT_F_KIND_SVM, pSvmTransient->u64ExitCode & EMEXIT_F_TYPE_MASK),
4819	pVCpu->cpum.GstCtx.cs.u64Base + pVCpu->cpum.GstCtx.rip, uHostTsc);
4820	}
4821
4822
4823	/**
4824	* Runs the guest code using AMD-V.
4825	*
4826	* @returns VBox status code.
4827	* @param pVCpu The cross context virtual CPU structure.
4828	* @param pcLoops Pointer to the number of executed loops.
4829	*/
4830	static int hmR0SvmRunGuestCodeNormal(PVMCPU pVCpu, uint32_t *pcLoops)
4831	{
4832	uint32_t const cMaxResumeLoops = pVCpu->CTX_SUFF(pVM)->hm.s.cMaxResumeLoops;
4833	Assert(pcLoops);
4834	Assert(*pcLoops <= cMaxResumeLoops);
4835
4836	SVMTRANSIENT SvmTransient;
4837	RT_ZERO(SvmTransient);
4838	SvmTransient.fUpdateTscOffsetting = true;
4839	SvmTransient.pVmcb = pVCpu->hm.s.svm.pVmcb;
4840
4841	int rc = VERR_INTERNAL_ERROR_5;
4842	for (;;)
4843	{
4844	Assert(!HMR0SuspendPending());
4845	HMSVM_ASSERT_CPU_SAFE(pVCpu);
4846
4847	/* Preparatory work for running nested-guest code, this may force us to return to
4848	ring-3. This bugger disables interrupts on VINF_SUCCESS! */
4849	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatEntry, x);
4850	rc = hmR0SvmPreRunGuest(pVCpu, &SvmTransient);
4851	if (rc != VINF_SUCCESS)
4852	break;
4853
4854	/*
4855	* No longjmps to ring-3 from this point on!!!
4856	*
4857	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional,
4858	* better than a kernel panic. This also disables flushing of the R0-logger instance.
4859	*/
4860	hmR0SvmPreRunGuestCommitted(pVCpu, &SvmTransient);
4861	rc = hmR0SvmRunGuest(pVCpu, pVCpu->hm.s.svm.HCPhysVmcb);
4862
4863	/* Restore any residual host-state and save any bits shared between host and guest
4864	into the guest-CPU state. Re-enables interrupts! */
4865	hmR0SvmPostRunGuest(pVCpu, &SvmTransient, rc);
4866
4867	if (RT_UNLIKELY( rc != VINF_SUCCESS /* Check for VMRUN errors. */
4868	\|\| SvmTransient.u64ExitCode == SVM_EXIT_INVALID)) /* Check for invalid guest-state errors. */
4869	{
4870	if (rc == VINF_SUCCESS)
4871	rc = VERR_SVM_INVALID_GUEST_STATE;
4872	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatPreExit, x);
4873	hmR0SvmReportWorldSwitchError(pVCpu, rc);
4874	break;
4875	}
4876
4877	/* Handle the #VMEXIT. */
4878	HMSVM_EXITCODE_STAM_COUNTER_INC(SvmTransient.u64ExitCode);
4879	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatPreExit, &pVCpu->hm.s.StatExitHandling, x);
4880	VBOXVMM_R0_HMSVM_VMEXIT(pVCpu, &pVCpu->cpum.GstCtx, SvmTransient.u64ExitCode, pVCpu->hm.s.svm.pVmcb);
4881	rc = hmR0SvmHandleExit(pVCpu, &SvmTransient);
4882	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExitHandling, x);
4883	if (rc != VINF_SUCCESS)
4884	break;
4885	if (++(*pcLoops) >= cMaxResumeLoops)
4886	{
4887	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchMaxResumeLoops);
4888	rc = VINF_EM_RAW_INTERRUPT;
4889	break;
4890	}
4891	}
4892
4893	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatEntry, x);
4894	return rc;
4895	}
4896
4897
4898	/**
4899	* Runs the guest code using AMD-V in single step mode.
4900	*
4901	* @returns VBox status code.
4902	* @param pVCpu The cross context virtual CPU structure.
4903	* @param pcLoops Pointer to the number of executed loops.
4904	*/
4905	static int hmR0SvmRunGuestCodeStep(PVMCPU pVCpu, uint32_t *pcLoops)
4906	{
4907	uint32_t const cMaxResumeLoops = pVCpu->CTX_SUFF(pVM)->hm.s.cMaxResumeLoops;
4908	Assert(pcLoops);
4909	Assert(*pcLoops <= cMaxResumeLoops);
4910
4911	SVMTRANSIENT SvmTransient;
4912	RT_ZERO(SvmTransient);
4913	SvmTransient.fUpdateTscOffsetting = true;
4914	SvmTransient.pVmcb = pVCpu->hm.s.svm.pVmcb;
4915
4916	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
4917	uint16_t uCsStart = pCtx->cs.Sel;
4918	uint64_t uRipStart = pCtx->rip;
4919
4920	int rc = VERR_INTERNAL_ERROR_5;
4921	for (;;)
4922	{
4923	Assert(!HMR0SuspendPending());
4924	AssertMsg(pVCpu->hm.s.idEnteredCpu == RTMpCpuId(),
4925	("Illegal migration! Entered on CPU %u Current %u cLoops=%u\n", (unsigned)pVCpu->hm.s.idEnteredCpu,
4926	(unsigned)RTMpCpuId(), *pcLoops));
4927
4928	/* Preparatory work for running nested-guest code, this may force us to return to
4929	ring-3. This bugger disables interrupts on VINF_SUCCESS! */
4930	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatEntry, x);
4931	rc = hmR0SvmPreRunGuest(pVCpu, &SvmTransient);
4932	if (rc != VINF_SUCCESS)
4933	break;
4934
4935	/*
4936	* No longjmps to ring-3 from this point on!!!
4937	*
4938	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional,
4939	* better than a kernel panic. This also disables flushing of the R0-logger instance.
4940	*/
4941	VMMRZCallRing3Disable(pVCpu);
4942	VMMRZCallRing3RemoveNotification(pVCpu);
4943	hmR0SvmPreRunGuestCommitted(pVCpu, &SvmTransient);
4944
4945	rc = hmR0SvmRunGuest(pVCpu, pVCpu->hm.s.svm.HCPhysVmcb);
4946
4947	/* Restore any residual host-state and save any bits shared between host and guest
4948	into the guest-CPU state. Re-enables interrupts! */
4949	hmR0SvmPostRunGuest(pVCpu, &SvmTransient, rc);
4950
4951	if (RT_UNLIKELY( rc != VINF_SUCCESS /* Check for VMRUN errors. */
4952	\|\| SvmTransient.u64ExitCode == SVM_EXIT_INVALID)) /* Check for invalid guest-state errors. */
4953	{
4954	if (rc == VINF_SUCCESS)
4955	rc = VERR_SVM_INVALID_GUEST_STATE;
4956	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatPreExit, x);
4957	hmR0SvmReportWorldSwitchError(pVCpu, rc);
4958	return rc;
4959	}
4960
4961	/* Handle the #VMEXIT. */
4962	HMSVM_EXITCODE_STAM_COUNTER_INC(SvmTransient.u64ExitCode);
4963	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatPreExit, &pVCpu->hm.s.StatExitHandling, x);
4964	VBOXVMM_R0_HMSVM_VMEXIT(pVCpu, pCtx, SvmTransient.u64ExitCode, pVCpu->hm.s.svm.pVmcb);
4965	rc = hmR0SvmHandleExit(pVCpu, &SvmTransient);
4966	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExitHandling, x);
4967	if (rc != VINF_SUCCESS)
4968	break;
4969	if (++(*pcLoops) >= cMaxResumeLoops)
4970	{
4971	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchMaxResumeLoops);
4972	rc = VINF_EM_RAW_INTERRUPT;
4973	break;
4974	}
4975
4976	/*
4977	* Did the RIP change, if so, consider it a single step.
4978	* Otherwise, make sure one of the TFs gets set.
4979	*/
4980	if ( pCtx->rip != uRipStart
4981	\|\| pCtx->cs.Sel != uCsStart)
4982	{
4983	rc = VINF_EM_DBG_STEPPED;
4984	break;
4985	}
4986	pVCpu->hm.s.fCtxChanged \|= HM_CHANGED_GUEST_DR_MASK;
4987	}
4988
4989	/*
4990	* Clear the X86_EFL_TF if necessary.
4991	*/
4992	if (pVCpu->hm.s.fClearTrapFlag)
4993	{
4994	pVCpu->hm.s.fClearTrapFlag = false;
4995	pCtx->eflags.Bits.u1TF = 0;
4996	}
4997
4998	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatEntry, x);
4999	return rc;
5000	}
5001
5002	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
5003	/**
5004	* Runs the nested-guest code using AMD-V.
5005	*
5006	* @returns VBox status code.
5007	* @param pVCpu The cross context virtual CPU structure.
5008	* @param pcLoops Pointer to the number of executed loops. If we're switching
5009	* from the guest-code execution loop to this nested-guest
5010	* execution loop pass the remainder value, else pass 0.
5011	*/
5012	static int hmR0SvmRunGuestCodeNested(PVMCPU pVCpu, uint32_t *pcLoops)
5013	{
5014	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
5015	HMSVM_ASSERT_IN_NESTED_GUEST(pCtx);
5016	Assert(pcLoops);
5017	Assert(*pcLoops <= pVCpu->CTX_SUFF(pVM)->hm.s.cMaxResumeLoops);
5018
5019	SVMTRANSIENT SvmTransient;
5020	RT_ZERO(SvmTransient);
5021	SvmTransient.fUpdateTscOffsetting = true;
5022	SvmTransient.pVmcb = pCtx->hwvirt.svm.CTX_SUFF(pVmcb);
5023	SvmTransient.fIsNestedGuest = true;
5024
5025	int rc = VERR_INTERNAL_ERROR_4;
5026	for (;;)
5027	{
5028	Assert(!HMR0SuspendPending());
5029	HMSVM_ASSERT_CPU_SAFE(pVCpu);
5030
5031	/* Preparatory work for running nested-guest code, this may force us to return to
5032	ring-3. This bugger disables interrupts on VINF_SUCCESS! */
5033	STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatEntry, x);
5034	rc = hmR0SvmPreRunGuestNested(pVCpu, &SvmTransient);
5035	if ( rc != VINF_SUCCESS
5036	\|\| !CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
5037	{
5038	break;
5039	}
5040
5041	/*
5042	* No longjmps to ring-3 from this point on!!!
5043	*
5044	* Asserts() will still longjmp to ring-3 (but won't return), which is intentional,
5045	* better than a kernel panic. This also disables flushing of the R0-logger instance.
5046	*/
5047	hmR0SvmPreRunGuestCommitted(pVCpu, &SvmTransient);
5048
5049	rc = hmR0SvmRunGuest(pVCpu, pCtx->hwvirt.svm.HCPhysVmcb);
5050
5051	/* Restore any residual host-state and save any bits shared between host and guest
5052	into the guest-CPU state. Re-enables interrupts! */
5053	hmR0SvmPostRunGuest(pVCpu, &SvmTransient, rc);
5054
5055	if (RT_LIKELY( rc == VINF_SUCCESS
5056	&& SvmTransient.u64ExitCode != SVM_EXIT_INVALID))
5057	{ /* extremely likely */ }
5058	else
5059	{
5060	/* VMRUN failed, shouldn't really happen, Guru. */
5061	if (rc != VINF_SUCCESS)
5062	break;
5063
5064	/* Invalid nested-guest state. Cause a #VMEXIT but assert on strict builds. */
5065	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
5066	AssertMsgFailed(("Invalid nested-guest state. rc=%Rrc u64ExitCode=%#RX64\n", rc, SvmTransient.u64ExitCode));
5067	rc = VBOXSTRICTRC_TODO(IEMExecSvmVmexit(pVCpu, SVM_EXIT_INVALID, 0, 0));
5068	break;
5069	}
5070
5071	/* Handle the #VMEXIT. */
5072	HMSVM_NESTED_EXITCODE_STAM_COUNTER_INC(SvmTransient.u64ExitCode);
5073	STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatPreExit, &pVCpu->hm.s.StatExitHandling, x);
5074	VBOXVMM_R0_HMSVM_VMEXIT(pVCpu, pCtx, SvmTransient.u64ExitCode, pCtx->hwvirt.svm.CTX_SUFF(pVmcb));
5075	rc = hmR0SvmHandleExitNested(pVCpu, &SvmTransient);
5076	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExitHandling, x);
5077	if ( rc != VINF_SUCCESS
5078	\|\| !CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
5079	break;
5080	if (++(*pcLoops) >= pVCpu->CTX_SUFF(pVM)->hm.s.cMaxResumeLoops)
5081	{
5082	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchMaxResumeLoops);
5083	rc = VINF_EM_RAW_INTERRUPT;
5084	break;
5085	}
5086
5087	/** @todo handle single-stepping */
5088	}
5089
5090	STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatEntry, x);
5091	return rc;
5092	}
5093	#endif
5094
5095
5096	/**
5097	* Runs the guest code using AMD-V.
5098	*
5099	* @returns Strict VBox status code.
5100	* @param pVCpu The cross context virtual CPU structure.
5101	*/
5102	VMMR0DECL(VBOXSTRICTRC) SVMR0RunGuestCode(PVMCPU pVCpu)
5103	{
5104	Assert(VMMRZCallRing3IsEnabled(pVCpu));
5105	HMSVM_ASSERT_PREEMPT_SAFE(pVCpu);
5106	VMMRZCallRing3SetNotification(pVCpu, hmR0SvmCallRing3Callback, NULL /* pvUser */);
5107
5108	uint32_t cLoops = 0;
5109	int rc;
5110	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
5111	if (!CPUMIsGuestInSvmNestedHwVirtMode(&pVCpu->cpum.GstCtx))
5112	#endif
5113	{
5114	if (!pVCpu->hm.s.fSingleInstruction)
5115	rc = hmR0SvmRunGuestCodeNormal(pVCpu, &cLoops);
5116	else
5117	rc = hmR0SvmRunGuestCodeStep(pVCpu, &cLoops);
5118	}
5119	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
5120	else
5121	{
5122	rc = VINF_SVM_VMRUN;
5123	}
5124
5125	/* Re-check the nested-guest condition here as we may be transitioning from the normal
5126	execution loop into the nested-guest, hence this is not placed in the 'else' part above. */
5127	if (rc == VINF_SVM_VMRUN)
5128	{
5129	rc = hmR0SvmRunGuestCodeNested(pVCpu, &cLoops);
5130	if (rc == VINF_SVM_VMEXIT)
5131	rc = VINF_SUCCESS;
5132	}
5133	#endif
5134
5135	/* Fixup error codes. */
5136	if (rc == VERR_EM_INTERPRETER)
5137	rc = VINF_EM_RAW_EMULATE_INSTR;
5138	else if (rc == VINF_EM_RESET)
5139	rc = VINF_EM_TRIPLE_FAULT;
5140
5141	/* Prepare to return to ring-3. This will remove longjmp notifications. */
5142	rc = hmR0SvmExitToRing3(pVCpu, rc);
5143	Assert(!VMMRZCallRing3IsNotificationSet(pVCpu));
5144	return rc;
5145	}
5146
5147
5148	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
5149	/**
5150	* Determines whether an IOIO intercept is active for the nested-guest or not.
5151	*
5152	* @param pvIoBitmap Pointer to the nested-guest IO bitmap.
5153	* @param pIoExitInfo Pointer to the SVMIOIOEXITINFO.
5154	*/
5155	static bool hmR0SvmIsIoInterceptActive(void *pvIoBitmap, PSVMIOIOEXITINFO pIoExitInfo)
5156	{
5157	const uint16_t u16Port = pIoExitInfo->n.u16Port;
5158	const SVMIOIOTYPE enmIoType = (SVMIOIOTYPE)pIoExitInfo->n.u1Type;
5159	const uint8_t cbReg = (pIoExitInfo->u >> SVM_IOIO_OP_SIZE_SHIFT) & 7;
5160	const uint8_t cAddrSizeBits = ((pIoExitInfo->u >> SVM_IOIO_ADDR_SIZE_SHIFT) & 7) << 4;
5161	const uint8_t iEffSeg = pIoExitInfo->n.u3Seg;
5162	const bool fRep = pIoExitInfo->n.u1Rep;
5163	const bool fStrIo = pIoExitInfo->n.u1Str;
5164
5165	return HMSvmIsIOInterceptActive(pvIoBitmap, u16Port, enmIoType, cbReg, cAddrSizeBits, iEffSeg, fRep, fStrIo,
5166	NULL /* pIoExitInfo */);
5167	}
5168
5169
5170	/**
5171	* Handles a nested-guest \#VMEXIT (for all EXITCODE values except
5172	* SVM_EXIT_INVALID).
5173	*
5174	* @returns VBox status code (informational status codes included).
5175	* @param pVCpu The cross context virtual CPU structure.
5176	* @param pSvmTransient Pointer to the SVM transient structure.
5177	*/
5178	static int hmR0SvmHandleExitNested(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
5179	{
5180	HMSVM_ASSERT_IN_NESTED_GUEST(&pVCpu->cpum.GstCtx);
5181	Assert(pSvmTransient->u64ExitCode != SVM_EXIT_INVALID);
5182	Assert(pSvmTransient->u64ExitCode <= SVM_EXIT_MAX);
5183
5184	/** @todo Figure out why using IEM_CPUMCTX_EXTRN_SVM_VMEXIT_MASK instead of
5185	* HMSVM_CPUMCTX_EXTRN_ALL breaks nested guests (XP Pro, DSL etc.), see
5186	* also HMSvmNstGstVmExitNotify(). */
5187	#define NST_GST_VMEXIT_CALL_RET(a_pVCpu, a_uExitCode, a_uExitInfo1, a_uExitInfo2) \
5188	do { \
5189	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL); \
5190	return VBOXSTRICTRC_TODO(IEMExecSvmVmexit((a_pVCpu), (a_uExitCode), (a_uExitInfo1), (a_uExitInfo2))); \
5191	} while (0)
5192
5193	/*
5194	* For all the #VMEXITs here we primarily figure out if the #VMEXIT is expected by the
5195	* nested-guest. If it isn't, it should be handled by the (outer) guest.
5196	*/
5197	PSVMVMCB pVmcbNstGst = pVCpu->cpum.GstCtx.hwvirt.svm.CTX_SUFF(pVmcb);
5198	PSVMVMCBCTRL pVmcbNstGstCtrl = &pVmcbNstGst->ctrl;
5199	uint64_t const uExitCode = pVmcbNstGstCtrl->u64ExitCode;
5200	uint64_t const uExitInfo1 = pVmcbNstGstCtrl->u64ExitInfo1;
5201	uint64_t const uExitInfo2 = pVmcbNstGstCtrl->u64ExitInfo2;
5202
5203	Assert(uExitCode == pVmcbNstGstCtrl->u64ExitCode);
5204	switch (uExitCode)
5205	{
5206	case SVM_EXIT_CPUID:
5207	{
5208	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_CPUID))
5209	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5210	return hmR0SvmExitCpuid(pVCpu, pSvmTransient);
5211	}
5212
5213	case SVM_EXIT_RDTSC:
5214	{
5215	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_RDTSC))
5216	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5217	return hmR0SvmExitRdtsc(pVCpu, pSvmTransient);
5218	}
5219
5220	case SVM_EXIT_RDTSCP:
5221	{
5222	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_RDTSCP))
5223	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5224	return hmR0SvmExitRdtscp(pVCpu, pSvmTransient);
5225	}
5226
5227	case SVM_EXIT_MONITOR:
5228	{
5229	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_MONITOR))
5230	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5231	return hmR0SvmExitMonitor(pVCpu, pSvmTransient);
5232	}
5233
5234	case SVM_EXIT_MWAIT:
5235	{
5236	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_MWAIT))
5237	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5238	return hmR0SvmExitMwait(pVCpu, pSvmTransient);
5239	}
5240
5241	case SVM_EXIT_HLT:
5242	{
5243	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_HLT))
5244	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5245	return hmR0SvmExitHlt(pVCpu, pSvmTransient);
5246	}
5247
5248	case SVM_EXIT_MSR:
5249	{
5250	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_MSR_PROT))
5251	{
5252	uint32_t const idMsr = pVCpu->cpum.GstCtx.ecx;
5253	uint16_t offMsrpm;
5254	uint8_t uMsrpmBit;
5255	int rc = HMSvmGetMsrpmOffsetAndBit(idMsr, &offMsrpm, &uMsrpmBit);
5256	if (RT_SUCCESS(rc))
5257	{
5258	Assert(uMsrpmBit == 0 \|\| uMsrpmBit == 2 \|\| uMsrpmBit == 4 \|\| uMsrpmBit == 6);
5259	Assert(offMsrpm < SVM_MSRPM_PAGES << X86_PAGE_4K_SHIFT);
5260
5261	uint8_t const pbMsrBitmap = (uint8_t const )pVCpu->cpum.GstCtx.hwvirt.svm.CTX_SUFF(pvMsrBitmap);
5262	pbMsrBitmap += offMsrpm;
5263	bool const fInterceptRead = RT_BOOL(*pbMsrBitmap & RT_BIT(uMsrpmBit));
5264	bool const fInterceptWrite = RT_BOOL(*pbMsrBitmap & RT_BIT(uMsrpmBit + 1));
5265
5266	if ( (fInterceptWrite && pVmcbNstGstCtrl->u64ExitInfo1 == SVM_EXIT1_MSR_WRITE)
5267	\|\| (fInterceptRead && pVmcbNstGstCtrl->u64ExitInfo1 == SVM_EXIT1_MSR_READ))
5268	{
5269	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5270	}
5271	}
5272	else
5273	{
5274	/*
5275	* MSRs not covered by the MSRPM automatically cause an #VMEXIT.
5276	* See AMD-V spec. "15.11 MSR Intercepts".
5277	*/
5278	Assert(rc == VERR_OUT_OF_RANGE);
5279	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5280	}
5281	}
5282	return hmR0SvmExitMsr(pVCpu, pSvmTransient);
5283	}
5284
5285	case SVM_EXIT_IOIO:
5286	{
5287	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_IOIO_PROT))
5288	{
5289	void *pvIoBitmap = pVCpu->cpum.GstCtx.hwvirt.svm.CTX_SUFF(pvIoBitmap);
5290	SVMIOIOEXITINFO IoExitInfo;
5291	IoExitInfo.u = pVmcbNstGst->ctrl.u64ExitInfo1;
5292	bool const fIntercept = hmR0SvmIsIoInterceptActive(pvIoBitmap, &IoExitInfo);
5293	if (fIntercept)
5294	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5295	}
5296	return hmR0SvmExitIOInstr(pVCpu, pSvmTransient);
5297	}
5298
5299	case SVM_EXIT_XCPT_PF:
5300	{
5301	PVM pVM = pVCpu->CTX_SUFF(pVM);
5302	if (pVM->hm.s.fNestedPaging)
5303	{
5304	uint32_t const u32ErrCode = pVmcbNstGstCtrl->u64ExitInfo1;
5305	uint64_t const uFaultAddress = pVmcbNstGstCtrl->u64ExitInfo2;
5306
5307	/* If the nested-guest is intercepting #PFs, cause a #PF #VMEXIT. */
5308	if (HMIsGuestSvmXcptInterceptSet(pVCpu, X86_XCPT_PF))
5309	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, u32ErrCode, uFaultAddress);
5310
5311	/* If the nested-guest is not intercepting #PFs, forward the #PF to the guest. */
5312	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR2);
5313	hmR0SvmSetPendingXcptPF(pVCpu, u32ErrCode, uFaultAddress);
5314	return VINF_SUCCESS;
5315	}
5316	return hmR0SvmExitXcptPF(pVCpu, pSvmTransient);
5317	}
5318
5319	case SVM_EXIT_XCPT_UD:
5320	{
5321	if (HMIsGuestSvmXcptInterceptSet(pVCpu, X86_XCPT_UD))
5322	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5323	hmR0SvmSetPendingXcptUD(pVCpu);
5324	return VINF_SUCCESS;
5325	}
5326
5327	case SVM_EXIT_XCPT_MF:
5328	{
5329	if (HMIsGuestSvmXcptInterceptSet(pVCpu, X86_XCPT_MF))
5330	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5331	return hmR0SvmExitXcptMF(pVCpu, pSvmTransient);
5332	}
5333
5334	case SVM_EXIT_XCPT_DB:
5335	{
5336	if (HMIsGuestSvmXcptInterceptSet(pVCpu, X86_XCPT_DB))
5337	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5338	return hmR0SvmNestedExitXcptDB(pVCpu, pSvmTransient);
5339	}
5340
5341	case SVM_EXIT_XCPT_AC:
5342	{
5343	if (HMIsGuestSvmXcptInterceptSet(pVCpu, X86_XCPT_AC))
5344	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5345	return hmR0SvmExitXcptAC(pVCpu, pSvmTransient);
5346	}
5347
5348	case SVM_EXIT_XCPT_BP:
5349	{
5350	if (HMIsGuestSvmXcptInterceptSet(pVCpu, X86_XCPT_BP))
5351	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5352	return hmR0SvmNestedExitXcptBP(pVCpu, pSvmTransient);
5353	}
5354
5355	case SVM_EXIT_READ_CR0:
5356	case SVM_EXIT_READ_CR3:
5357	case SVM_EXIT_READ_CR4:
5358	{
5359	uint8_t const uCr = uExitCode - SVM_EXIT_READ_CR0;
5360	if (HMIsGuestSvmReadCRxInterceptSet(pVCpu, uCr))
5361	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5362	return hmR0SvmExitReadCRx(pVCpu, pSvmTransient);
5363	}
5364
5365	case SVM_EXIT_CR0_SEL_WRITE:
5366	{
5367	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_CR0_SEL_WRITE))
5368	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5369	return hmR0SvmExitWriteCRx(pVCpu, pSvmTransient);
5370	}
5371
5372	case SVM_EXIT_WRITE_CR0:
5373	case SVM_EXIT_WRITE_CR3:
5374	case SVM_EXIT_WRITE_CR4:
5375	case SVM_EXIT_WRITE_CR8: /* CR8 writes would go to the V_TPR rather than here, since we run with V_INTR_MASKING. */
5376	{
5377	uint8_t const uCr = uExitCode - SVM_EXIT_WRITE_CR0;
5378	Log4Func(("Write CR%u: uExitInfo1=%#RX64 uExitInfo2=%#RX64\n", uCr, uExitInfo1, uExitInfo2));
5379
5380	if (HMIsGuestSvmWriteCRxInterceptSet(pVCpu, uCr))
5381	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5382	return hmR0SvmExitWriteCRx(pVCpu, pSvmTransient);
5383	}
5384
5385	case SVM_EXIT_PAUSE:
5386	{
5387	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_PAUSE))
5388	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5389	return hmR0SvmExitPause(pVCpu, pSvmTransient);
5390	}
5391
5392	case SVM_EXIT_VINTR:
5393	{
5394	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_VINTR))
5395	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5396	return hmR0SvmExitUnexpected(pVCpu, pSvmTransient);
5397	}
5398
5399	case SVM_EXIT_INTR:
5400	case SVM_EXIT_NMI:
5401	case SVM_EXIT_SMI:
5402	case SVM_EXIT_XCPT_NMI: /* Should not occur, SVM_EXIT_NMI is used instead. */
5403	{
5404	/*
5405	* We shouldn't direct physical interrupts, NMIs, SMIs to the nested-guest.
5406	*
5407	* Although we don't intercept SMIs, the nested-guest might. Therefore, we might
5408	* get an SMI #VMEXIT here so simply ignore rather than causing a corresponding
5409	* nested-guest #VMEXIT.
5410	*/
5411	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_SVM_VMEXIT_MASK);
5412	return hmR0SvmExitIntr(pVCpu, pSvmTransient);
5413	}
5414
5415	case SVM_EXIT_FERR_FREEZE:
5416	{
5417	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_FERR_FREEZE))
5418	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5419	return hmR0SvmExitFerrFreeze(pVCpu, pSvmTransient);
5420	}
5421
5422	case SVM_EXIT_INVLPG:
5423	{
5424	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_INVLPG))
5425	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5426	return hmR0SvmExitInvlpg(pVCpu, pSvmTransient);
5427	}
5428
5429	case SVM_EXIT_WBINVD:
5430	{
5431	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_WBINVD))
5432	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5433	return hmR0SvmExitWbinvd(pVCpu, pSvmTransient);
5434	}
5435
5436	case SVM_EXIT_INVD:
5437	{
5438	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_INVD))
5439	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5440	return hmR0SvmExitInvd(pVCpu, pSvmTransient);
5441	}
5442
5443	case SVM_EXIT_RDPMC:
5444	{
5445	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_RDPMC))
5446	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5447	return hmR0SvmExitRdpmc(pVCpu, pSvmTransient);
5448	}
5449
5450	default:
5451	{
5452	switch (uExitCode)
5453	{
5454	case SVM_EXIT_READ_DR0: case SVM_EXIT_READ_DR1: case SVM_EXIT_READ_DR2: case SVM_EXIT_READ_DR3:
5455	case SVM_EXIT_READ_DR6: case SVM_EXIT_READ_DR7: case SVM_EXIT_READ_DR8: case SVM_EXIT_READ_DR9:
5456	case SVM_EXIT_READ_DR10: case SVM_EXIT_READ_DR11: case SVM_EXIT_READ_DR12: case SVM_EXIT_READ_DR13:
5457	case SVM_EXIT_READ_DR14: case SVM_EXIT_READ_DR15:
5458	{
5459	uint8_t const uDr = uExitCode - SVM_EXIT_READ_DR0;
5460	if (HMIsGuestSvmReadDRxInterceptSet(pVCpu, uDr))
5461	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5462	return hmR0SvmExitReadDRx(pVCpu, pSvmTransient);
5463	}
5464
5465	case SVM_EXIT_WRITE_DR0: case SVM_EXIT_WRITE_DR1: case SVM_EXIT_WRITE_DR2: case SVM_EXIT_WRITE_DR3:
5466	case SVM_EXIT_WRITE_DR6: case SVM_EXIT_WRITE_DR7: case SVM_EXIT_WRITE_DR8: case SVM_EXIT_WRITE_DR9:
5467	case SVM_EXIT_WRITE_DR10: case SVM_EXIT_WRITE_DR11: case SVM_EXIT_WRITE_DR12: case SVM_EXIT_WRITE_DR13:
5468	case SVM_EXIT_WRITE_DR14: case SVM_EXIT_WRITE_DR15:
5469	{
5470	uint8_t const uDr = uExitCode - SVM_EXIT_WRITE_DR0;
5471	if (HMIsGuestSvmWriteDRxInterceptSet(pVCpu, uDr))
5472	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5473	return hmR0SvmExitWriteDRx(pVCpu, pSvmTransient);
5474	}
5475
5476	case SVM_EXIT_XCPT_DE:
5477	/* SVM_EXIT_XCPT_DB: / / Handled above. */
5478	/* SVM_EXIT_XCPT_NMI: / / Handled above. */
5479	/* SVM_EXIT_XCPT_BP: / / Handled above. */
5480	case SVM_EXIT_XCPT_OF:
5481	case SVM_EXIT_XCPT_BR:
5482	/* SVM_EXIT_XCPT_UD: / / Handled above. */
5483	case SVM_EXIT_XCPT_NM:
5484	case SVM_EXIT_XCPT_DF:
5485	case SVM_EXIT_XCPT_CO_SEG_OVERRUN:
5486	case SVM_EXIT_XCPT_TS:
5487	case SVM_EXIT_XCPT_NP:
5488	case SVM_EXIT_XCPT_SS:
5489	case SVM_EXIT_XCPT_GP:
5490	/* SVM_EXIT_XCPT_PF: / / Handled above. */
5491	case SVM_EXIT_XCPT_15: /* Reserved. */
5492	/* SVM_EXIT_XCPT_MF: / / Handled above. */
5493	/* SVM_EXIT_XCPT_AC: / / Handled above. */
5494	case SVM_EXIT_XCPT_MC:
5495	case SVM_EXIT_XCPT_XF:
5496	case SVM_EXIT_XCPT_20: case SVM_EXIT_XCPT_21: case SVM_EXIT_XCPT_22: case SVM_EXIT_XCPT_23:
5497	case SVM_EXIT_XCPT_24: case SVM_EXIT_XCPT_25: case SVM_EXIT_XCPT_26: case SVM_EXIT_XCPT_27:
5498	case SVM_EXIT_XCPT_28: case SVM_EXIT_XCPT_29: case SVM_EXIT_XCPT_30: case SVM_EXIT_XCPT_31:
5499	{
5500	uint8_t const uVector = uExitCode - SVM_EXIT_XCPT_0;
5501	if (HMIsGuestSvmXcptInterceptSet(pVCpu, uVector))
5502	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5503	return hmR0SvmExitXcptGeneric(pVCpu, pSvmTransient);
5504	}
5505
5506	case SVM_EXIT_XSETBV:
5507	{
5508	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_XSETBV))
5509	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5510	return hmR0SvmExitXsetbv(pVCpu, pSvmTransient);
5511	}
5512
5513	case SVM_EXIT_TASK_SWITCH:
5514	{
5515	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_TASK_SWITCH))
5516	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5517	return hmR0SvmExitTaskSwitch(pVCpu, pSvmTransient);
5518	}
5519
5520	case SVM_EXIT_IRET:
5521	{
5522	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_IRET))
5523	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5524	return hmR0SvmExitIret(pVCpu, pSvmTransient);
5525	}
5526
5527	case SVM_EXIT_SHUTDOWN:
5528	{
5529	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_SHUTDOWN))
5530	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5531	return hmR0SvmExitShutdown(pVCpu, pSvmTransient);
5532	}
5533
5534	case SVM_EXIT_VMMCALL:
5535	{
5536	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_VMMCALL))
5537	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5538	return hmR0SvmExitVmmCall(pVCpu, pSvmTransient);
5539	}
5540
5541	case SVM_EXIT_CLGI:
5542	{
5543	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_CLGI))
5544	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5545	return hmR0SvmExitClgi(pVCpu, pSvmTransient);
5546	}
5547
5548	case SVM_EXIT_STGI:
5549	{
5550	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_STGI))
5551	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5552	return hmR0SvmExitStgi(pVCpu, pSvmTransient);
5553	}
5554
5555	case SVM_EXIT_VMLOAD:
5556	{
5557	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_VMLOAD))
5558	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5559	return hmR0SvmExitVmload(pVCpu, pSvmTransient);
5560	}
5561
5562	case SVM_EXIT_VMSAVE:
5563	{
5564	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_VMSAVE))
5565	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5566	return hmR0SvmExitVmsave(pVCpu, pSvmTransient);
5567	}
5568
5569	case SVM_EXIT_INVLPGA:
5570	{
5571	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_INVLPGA))
5572	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5573	return hmR0SvmExitInvlpga(pVCpu, pSvmTransient);
5574	}
5575
5576	case SVM_EXIT_VMRUN:
5577	{
5578	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_VMRUN))
5579	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5580	return hmR0SvmExitVmrun(pVCpu, pSvmTransient);
5581	}
5582
5583	case SVM_EXIT_RSM:
5584	{
5585	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_RSM))
5586	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5587	hmR0SvmSetPendingXcptUD(pVCpu);
5588	return VINF_SUCCESS;
5589	}
5590
5591	case SVM_EXIT_SKINIT:
5592	{
5593	if (HMIsGuestSvmCtrlInterceptSet(pVCpu, SVM_CTRL_INTERCEPT_SKINIT))
5594	NST_GST_VMEXIT_CALL_RET(pVCpu, uExitCode, uExitInfo1, uExitInfo2);
5595	hmR0SvmSetPendingXcptUD(pVCpu);
5596	return VINF_SUCCESS;
5597	}
5598
5599	case SVM_EXIT_NPF:
5600	{
5601	Assert(pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
5602	return hmR0SvmExitNestedPF(pVCpu, pSvmTransient);
5603	}
5604
5605	case SVM_EXIT_INIT: /* We shouldn't get INIT signals while executing a nested-guest. */
5606	return hmR0SvmExitUnexpected(pVCpu, pSvmTransient);
5607
5608	default:
5609	{
5610	AssertMsgFailed(("hmR0SvmHandleExitNested: Unknown exit code %#x\n", pSvmTransient->u64ExitCode));
5611	pVCpu->hm.s.u32HMError = pSvmTransient->u64ExitCode;
5612	return VERR_SVM_UNKNOWN_EXIT;
5613	}
5614	}
5615	}
5616	}
5617	/* not reached */
5618
5619	#undef NST_GST_VMEXIT_CALL_RET
5620	}
5621	#endif
5622
5623
5624	/**
5625	* Handles a guest \#VMEXIT (for all EXITCODE values except SVM_EXIT_INVALID).
5626	*
5627	* @returns VBox status code (informational status codes included).
5628	* @param pVCpu The cross context virtual CPU structure.
5629	* @param pSvmTransient Pointer to the SVM transient structure.
5630	*/
5631	static int hmR0SvmHandleExit(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
5632	{
5633	Assert(pSvmTransient->u64ExitCode != SVM_EXIT_INVALID);
5634	Assert(pSvmTransient->u64ExitCode <= SVM_EXIT_MAX);
5635
5636	#ifdef DEBUG_ramshankar
5637	# define VMEXIT_CALL_RET(a_fDbg, a_CallExpr) \
5638	do { \
5639	if ((a_fDbg) == 1) \
5640	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL); \
5641	int rc = a_CallExpr; \
5642	if ((a_fDbg) == 1) \
5643	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST); \
5644	return rc; \
5645	} while (0)
5646	#else
5647	# define VMEXIT_CALL_RET(a_fDbg, a_CallExpr) return a_CallExpr
5648	#endif
5649
5650	/*
5651	* The ordering of the case labels is based on most-frequently-occurring #VMEXITs
5652	* for most guests under normal workloads (for some definition of "normal").
5653	*/
5654	uint64_t const uExitCode = pSvmTransient->u64ExitCode;
5655	switch (uExitCode)
5656	{
5657	case SVM_EXIT_NPF: VMEXIT_CALL_RET(0, hmR0SvmExitNestedPF(pVCpu, pSvmTransient));
5658	case SVM_EXIT_IOIO: VMEXIT_CALL_RET(0, hmR0SvmExitIOInstr(pVCpu, pSvmTransient));
5659	case SVM_EXIT_RDTSC: VMEXIT_CALL_RET(0, hmR0SvmExitRdtsc(pVCpu, pSvmTransient));
5660	case SVM_EXIT_RDTSCP: VMEXIT_CALL_RET(0, hmR0SvmExitRdtscp(pVCpu, pSvmTransient));
5661	case SVM_EXIT_CPUID: VMEXIT_CALL_RET(0, hmR0SvmExitCpuid(pVCpu, pSvmTransient));
5662	case SVM_EXIT_XCPT_PF: VMEXIT_CALL_RET(0, hmR0SvmExitXcptPF(pVCpu, pSvmTransient));
5663	case SVM_EXIT_MSR: VMEXIT_CALL_RET(0, hmR0SvmExitMsr(pVCpu, pSvmTransient));
5664	case SVM_EXIT_MONITOR: VMEXIT_CALL_RET(0, hmR0SvmExitMonitor(pVCpu, pSvmTransient));
5665	case SVM_EXIT_MWAIT: VMEXIT_CALL_RET(0, hmR0SvmExitMwait(pVCpu, pSvmTransient));
5666	case SVM_EXIT_HLT: VMEXIT_CALL_RET(0, hmR0SvmExitHlt(pVCpu, pSvmTransient));
5667
5668	case SVM_EXIT_XCPT_NMI: /* Should not occur, SVM_EXIT_NMI is used instead. */
5669	case SVM_EXIT_INTR:
5670	case SVM_EXIT_NMI: VMEXIT_CALL_RET(0, hmR0SvmExitIntr(pVCpu, pSvmTransient));
5671
5672	case SVM_EXIT_READ_CR0:
5673	case SVM_EXIT_READ_CR3:
5674	case SVM_EXIT_READ_CR4: VMEXIT_CALL_RET(0, hmR0SvmExitReadCRx(pVCpu, pSvmTransient));
5675
5676	case SVM_EXIT_CR0_SEL_WRITE:
5677	case SVM_EXIT_WRITE_CR0:
5678	case SVM_EXIT_WRITE_CR3:
5679	case SVM_EXIT_WRITE_CR4:
5680	case SVM_EXIT_WRITE_CR8: VMEXIT_CALL_RET(0, hmR0SvmExitWriteCRx(pVCpu, pSvmTransient));
5681
5682	case SVM_EXIT_VINTR: VMEXIT_CALL_RET(0, hmR0SvmExitVIntr(pVCpu, pSvmTransient));
5683	case SVM_EXIT_PAUSE: VMEXIT_CALL_RET(0, hmR0SvmExitPause(pVCpu, pSvmTransient));
5684	case SVM_EXIT_VMMCALL: VMEXIT_CALL_RET(0, hmR0SvmExitVmmCall(pVCpu, pSvmTransient));
5685	case SVM_EXIT_INVLPG: VMEXIT_CALL_RET(0, hmR0SvmExitInvlpg(pVCpu, pSvmTransient));
5686	case SVM_EXIT_WBINVD: VMEXIT_CALL_RET(0, hmR0SvmExitWbinvd(pVCpu, pSvmTransient));
5687	case SVM_EXIT_INVD: VMEXIT_CALL_RET(0, hmR0SvmExitInvd(pVCpu, pSvmTransient));
5688	case SVM_EXIT_RDPMC: VMEXIT_CALL_RET(0, hmR0SvmExitRdpmc(pVCpu, pSvmTransient));
5689	case SVM_EXIT_IRET: VMEXIT_CALL_RET(0, hmR0SvmExitIret(pVCpu, pSvmTransient));
5690	case SVM_EXIT_XCPT_UD: VMEXIT_CALL_RET(0, hmR0SvmExitXcptUD(pVCpu, pSvmTransient));
5691	case SVM_EXIT_XCPT_MF: VMEXIT_CALL_RET(0, hmR0SvmExitXcptMF(pVCpu, pSvmTransient));
5692	case SVM_EXIT_XCPT_DB: VMEXIT_CALL_RET(0, hmR0SvmExitXcptDB(pVCpu, pSvmTransient));
5693	case SVM_EXIT_XCPT_AC: VMEXIT_CALL_RET(0, hmR0SvmExitXcptAC(pVCpu, pSvmTransient));
5694	case SVM_EXIT_XCPT_BP: VMEXIT_CALL_RET(0, hmR0SvmExitXcptBP(pVCpu, pSvmTransient));
5695	case SVM_EXIT_XSETBV: VMEXIT_CALL_RET(0, hmR0SvmExitXsetbv(pVCpu, pSvmTransient));
5696	case SVM_EXIT_FERR_FREEZE: VMEXIT_CALL_RET(0, hmR0SvmExitFerrFreeze(pVCpu, pSvmTransient));
5697
5698	default:
5699	{
5700	switch (pSvmTransient->u64ExitCode)
5701	{
5702	case SVM_EXIT_READ_DR0: case SVM_EXIT_READ_DR1: case SVM_EXIT_READ_DR2: case SVM_EXIT_READ_DR3:
5703	case SVM_EXIT_READ_DR6: case SVM_EXIT_READ_DR7: case SVM_EXIT_READ_DR8: case SVM_EXIT_READ_DR9:
5704	case SVM_EXIT_READ_DR10: case SVM_EXIT_READ_DR11: case SVM_EXIT_READ_DR12: case SVM_EXIT_READ_DR13:
5705	case SVM_EXIT_READ_DR14: case SVM_EXIT_READ_DR15:
5706	VMEXIT_CALL_RET(0, hmR0SvmExitReadDRx(pVCpu, pSvmTransient));
5707
5708	case SVM_EXIT_WRITE_DR0: case SVM_EXIT_WRITE_DR1: case SVM_EXIT_WRITE_DR2: case SVM_EXIT_WRITE_DR3:
5709	case SVM_EXIT_WRITE_DR6: case SVM_EXIT_WRITE_DR7: case SVM_EXIT_WRITE_DR8: case SVM_EXIT_WRITE_DR9:
5710	case SVM_EXIT_WRITE_DR10: case SVM_EXIT_WRITE_DR11: case SVM_EXIT_WRITE_DR12: case SVM_EXIT_WRITE_DR13:
5711	case SVM_EXIT_WRITE_DR14: case SVM_EXIT_WRITE_DR15:
5712	VMEXIT_CALL_RET(0, hmR0SvmExitWriteDRx(pVCpu, pSvmTransient));
5713
5714	case SVM_EXIT_TASK_SWITCH: VMEXIT_CALL_RET(0, hmR0SvmExitTaskSwitch(pVCpu, pSvmTransient));
5715	case SVM_EXIT_SHUTDOWN: VMEXIT_CALL_RET(0, hmR0SvmExitShutdown(pVCpu, pSvmTransient));
5716
5717	case SVM_EXIT_SMI:
5718	case SVM_EXIT_INIT:
5719	{
5720	/*
5721	* We don't intercept SMIs. As for INIT signals, it really shouldn't ever happen here.
5722	* If it ever does, we want to know about it so log the exit code and bail.
5723	*/
5724	VMEXIT_CALL_RET(0, hmR0SvmExitUnexpected(pVCpu, pSvmTransient));
5725	}
5726
5727	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
5728	case SVM_EXIT_CLGI: VMEXIT_CALL_RET(0, hmR0SvmExitClgi(pVCpu, pSvmTransient));
5729	case SVM_EXIT_STGI: VMEXIT_CALL_RET(0, hmR0SvmExitStgi(pVCpu, pSvmTransient));
5730	case SVM_EXIT_VMLOAD: VMEXIT_CALL_RET(0, hmR0SvmExitVmload(pVCpu, pSvmTransient));
5731	case SVM_EXIT_VMSAVE: VMEXIT_CALL_RET(0, hmR0SvmExitVmsave(pVCpu, pSvmTransient));
5732	case SVM_EXIT_INVLPGA: VMEXIT_CALL_RET(0, hmR0SvmExitInvlpga(pVCpu, pSvmTransient));
5733	case SVM_EXIT_VMRUN: VMEXIT_CALL_RET(0, hmR0SvmExitVmrun(pVCpu, pSvmTransient));
5734	#else
5735	case SVM_EXIT_CLGI:
5736	case SVM_EXIT_STGI:
5737	case SVM_EXIT_VMLOAD:
5738	case SVM_EXIT_VMSAVE:
5739	case SVM_EXIT_INVLPGA:
5740	case SVM_EXIT_VMRUN:
5741	#endif
5742	case SVM_EXIT_RSM:
5743	case SVM_EXIT_SKINIT:
5744	{
5745	hmR0SvmSetPendingXcptUD(pVCpu);
5746	return VINF_SUCCESS;
5747	}
5748
5749	#ifdef HMSVM_ALWAYS_TRAP_ALL_XCPTS
5750	case SVM_EXIT_XCPT_DE:
5751	/* SVM_EXIT_XCPT_DB: / / Handled above. */
5752	/* SVM_EXIT_XCPT_NMI: / / Handled above. */
5753	/* SVM_EXIT_XCPT_BP: / / Handled above. */
5754	case SVM_EXIT_XCPT_OF:
5755	case SVM_EXIT_XCPT_BR:
5756	/* SVM_EXIT_XCPT_UD: / / Handled above. */
5757	case SVM_EXIT_XCPT_NM:
5758	case SVM_EXIT_XCPT_DF:
5759	case SVM_EXIT_XCPT_CO_SEG_OVERRUN:
5760	case SVM_EXIT_XCPT_TS:
5761	case SVM_EXIT_XCPT_NP:
5762	case SVM_EXIT_XCPT_SS:
5763	case SVM_EXIT_XCPT_GP:
5764	/* SVM_EXIT_XCPT_PF: */
5765	case SVM_EXIT_XCPT_15: /* Reserved. */
5766	/* SVM_EXIT_XCPT_MF: / / Handled above. */
5767	/* SVM_EXIT_XCPT_AC: / / Handled above. */
5768	case SVM_EXIT_XCPT_MC:
5769	case SVM_EXIT_XCPT_XF:
5770	case SVM_EXIT_XCPT_20: case SVM_EXIT_XCPT_21: case SVM_EXIT_XCPT_22: case SVM_EXIT_XCPT_23:
5771	case SVM_EXIT_XCPT_24: case SVM_EXIT_XCPT_25: case SVM_EXIT_XCPT_26: case SVM_EXIT_XCPT_27:
5772	case SVM_EXIT_XCPT_28: case SVM_EXIT_XCPT_29: case SVM_EXIT_XCPT_30: case SVM_EXIT_XCPT_31:
5773	VMEXIT_CALL_RET(0, hmR0SvmExitXcptGeneric(pVCpu, pSvmTransient));
5774	#endif /* HMSVM_ALWAYS_TRAP_ALL_XCPTS */
5775
5776	default:
5777	{
5778	AssertMsgFailed(("hmR0SvmHandleExit: Unknown exit code %#RX64\n", uExitCode));
5779	pVCpu->hm.s.u32HMError = uExitCode;
5780	return VERR_SVM_UNKNOWN_EXIT;
5781	}
5782	}
5783	}
5784	}
5785	/* not reached */
5786	#undef VMEXIT_CALL_RET
5787	}
5788
5789
5790	#ifdef VBOX_STRICT
5791	/* Is there some generic IPRT define for this that are not in Runtime/internal/\* ?? */
5792	# define HMSVM_ASSERT_PREEMPT_CPUID_VAR() \
5793	RTCPUID const idAssertCpu = RTThreadPreemptIsEnabled(NIL_RTTHREAD) ? NIL_RTCPUID : RTMpCpuId()
5794
5795	# define HMSVM_ASSERT_PREEMPT_CPUID() \
5796	do \
5797	{ \
5798	RTCPUID const idAssertCpuNow = RTThreadPreemptIsEnabled(NIL_RTTHREAD) ? NIL_RTCPUID : RTMpCpuId(); \
5799	AssertMsg(idAssertCpu == idAssertCpuNow, ("SVM %#x, %#x\n", idAssertCpu, idAssertCpuNow)); \
5800	} while (0)
5801
5802	# define HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(a_pVCpu, a_pSvmTransient) \
5803	do { \
5804	AssertPtr((a_pVCpu)); \
5805	AssertPtr((a_pSvmTransient)); \
5806	Assert(ASMIntAreEnabled()); \
5807	HMSVM_ASSERT_PREEMPT_SAFE((a_pVCpu)); \
5808	HMSVM_ASSERT_PREEMPT_CPUID_VAR(); \
5809	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)); \
5810	HMSVM_ASSERT_PREEMPT_SAFE((a_pVCpu)); \
5811	if (VMMR0IsLogFlushDisabled((a_pVCpu))) \
5812	HMSVM_ASSERT_PREEMPT_CPUID(); \
5813	} while (0)
5814	#else
5815	# define HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(a_pVCpu, a_pSvmTransient) \
5816	do { \
5817	RT_NOREF2(a_pVCpu, a_pSvmTransient); \
5818	} while (0)
5819	#endif
5820
5821
5822	/**
5823	* Gets the IEM exception flags for the specified SVM event.
5824	*
5825	* @returns The IEM exception flags.
5826	* @param pEvent Pointer to the SVM event.
5827	*
5828	* @remarks This function currently only constructs flags required for
5829	* IEMEvaluateRecursiveXcpt and not the complete flags (e.g. error-code
5830	* and CR2 aspects of an exception are not included).
5831	*/
5832	static uint32_t hmR0SvmGetIemXcptFlags(PCSVMEVENT pEvent)
5833	{
5834	uint8_t const uEventType = pEvent->n.u3Type;
5835	uint32_t fIemXcptFlags;
5836	switch (uEventType)
5837	{
5838	case SVM_EVENT_EXCEPTION:
5839	/*
5840	* Only INT3 and INTO instructions can raise #BP and #OF exceptions.
5841	* See AMD spec. Table 8-1. "Interrupt Vector Source and Cause".
5842	*/
5843	if (pEvent->n.u8Vector == X86_XCPT_BP)
5844	{
5845	fIemXcptFlags = IEM_XCPT_FLAGS_T_SOFT_INT \| IEM_XCPT_FLAGS_BP_INSTR;
5846	break;
5847	}
5848	if (pEvent->n.u8Vector == X86_XCPT_OF)
5849	{
5850	fIemXcptFlags = IEM_XCPT_FLAGS_T_SOFT_INT \| IEM_XCPT_FLAGS_OF_INSTR;
5851	break;
5852	}
5853	/** @todo How do we distinguish ICEBP \#DB from the regular one? */
5854	RT_FALL_THRU();
5855	case SVM_EVENT_NMI:
5856	fIemXcptFlags = IEM_XCPT_FLAGS_T_CPU_XCPT;
5857	break;
5858
5859	case SVM_EVENT_EXTERNAL_IRQ:
5860	fIemXcptFlags = IEM_XCPT_FLAGS_T_EXT_INT;
5861	break;
5862
5863	case SVM_EVENT_SOFTWARE_INT:
5864	fIemXcptFlags = IEM_XCPT_FLAGS_T_SOFT_INT;
5865	break;
5866
5867	default:
5868	fIemXcptFlags = 0;
5869	AssertMsgFailed(("Unexpected event type! uEventType=%#x uVector=%#x", uEventType, pEvent->n.u8Vector));
5870	break;
5871	}
5872	return fIemXcptFlags;
5873	}
5874
5875
5876	/**
5877	* Handle a condition that occurred while delivering an event through the guest
5878	* IDT.
5879	*
5880	* @returns VBox status code (informational error codes included).
5881	* @retval VINF_SUCCESS if we should continue handling the \#VMEXIT.
5882	* @retval VINF_HM_DOUBLE_FAULT if a \#DF condition was detected and we ought to
5883	* continue execution of the guest which will delivery the \#DF.
5884	* @retval VINF_EM_RESET if we detected a triple-fault condition.
5885	* @retval VERR_EM_GUEST_CPU_HANG if we detected a guest CPU hang.
5886	*
5887	* @param pVCpu The cross context virtual CPU structure.
5888	* @param pSvmTransient Pointer to the SVM transient structure.
5889	*
5890	* @remarks No-long-jump zone!!!
5891	*/
5892	static int hmR0SvmCheckExitDueToEventDelivery(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
5893	{
5894	int rc = VINF_SUCCESS;
5895	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
5896	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR2);
5897
5898	Log4(("EXITINTINFO: Pending vectoring event %#RX64 Valid=%RTbool ErrValid=%RTbool Err=%#RX32 Type=%u Vector=%u\n",
5899	pVmcb->ctrl.ExitIntInfo.u, !!pVmcb->ctrl.ExitIntInfo.n.u1Valid, !!pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid,
5900	pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode, pVmcb->ctrl.ExitIntInfo.n.u3Type, pVmcb->ctrl.ExitIntInfo.n.u8Vector));
5901
5902	/*
5903	* The EXITINTINFO (if valid) contains the prior exception (IDT vector) that was trying to
5904	* be delivered to the guest which caused a #VMEXIT which was intercepted (Exit vector).
5905	*
5906	* See AMD spec. 15.7.3 "EXITINFO Pseudo-Code".
5907	*/
5908	if (pVmcb->ctrl.ExitIntInfo.n.u1Valid)
5909	{
5910	IEMXCPTRAISE enmRaise;
5911	IEMXCPTRAISEINFO fRaiseInfo;
5912	bool const fExitIsHwXcpt = pSvmTransient->u64ExitCode - SVM_EXIT_XCPT_0 <= SVM_EXIT_XCPT_31;
5913	uint8_t const uIdtVector = pVmcb->ctrl.ExitIntInfo.n.u8Vector;
5914	if (fExitIsHwXcpt)
5915	{
5916	uint8_t const uExitVector = pSvmTransient->u64ExitCode - SVM_EXIT_XCPT_0;
5917	uint32_t const fIdtVectorFlags = hmR0SvmGetIemXcptFlags(&pVmcb->ctrl.ExitIntInfo);
5918	uint32_t const fExitVectorFlags = IEM_XCPT_FLAGS_T_CPU_XCPT;
5919	enmRaise = IEMEvaluateRecursiveXcpt(pVCpu, fIdtVectorFlags, uIdtVector, fExitVectorFlags, uExitVector, &fRaiseInfo);
5920	}
5921	else
5922	{
5923	/*
5924	* If delivery of an event caused a #VMEXIT that is not an exception (e.g. #NPF)
5925	* then we end up here.
5926	*
5927	* If the event was:
5928	* - a software interrupt, we can re-execute the instruction which will
5929	* regenerate the event.
5930	* - an NMI, we need to clear NMI blocking and re-inject the NMI.
5931	* - a hardware exception or external interrupt, we re-inject it.
5932	*/
5933	fRaiseInfo = IEMXCPTRAISEINFO_NONE;
5934	if (pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_SOFTWARE_INT)
5935	enmRaise = IEMXCPTRAISE_REEXEC_INSTR;
5936	else
5937	enmRaise = IEMXCPTRAISE_PREV_EVENT;
5938	}
5939
5940	switch (enmRaise)
5941	{
5942	case IEMXCPTRAISE_CURRENT_XCPT:
5943	case IEMXCPTRAISE_PREV_EVENT:
5944	{
5945	/* For software interrupts, we shall re-execute the instruction. */
5946	if (!(fRaiseInfo & IEMXCPTRAISEINFO_SOFT_INT_XCPT))
5947	{
5948	RTGCUINTPTR GCPtrFaultAddress = 0;
5949
5950	/* If we are re-injecting an NMI, clear NMI blocking. */
5951	if (pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_NMI)
5952	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_BLOCK_NMIS);
5953
5954	/* Determine a vectoring #PF condition, see comment in hmR0SvmExitXcptPF(). */
5955	if (fRaiseInfo & (IEMXCPTRAISEINFO_EXT_INT_PF \| IEMXCPTRAISEINFO_NMI_PF))
5956	{
5957	pSvmTransient->fVectoringPF = true;
5958	Log4Func(("IDT: Pending vectoring #PF due to delivery of Ext-Int/NMI. uCR2=%#RX64\n",
5959	pVCpu->cpum.GstCtx.cr2));
5960	}
5961	else if ( pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_EXCEPTION
5962	&& uIdtVector == X86_XCPT_PF)
5963	{
5964	/*
5965	* If the previous exception was a #PF, we need to recover the CR2 value.
5966	* This can't happen with shadow paging.
5967	*/
5968	GCPtrFaultAddress = pVCpu->cpum.GstCtx.cr2;
5969	}
5970
5971	/*
5972	* Without nested paging, when uExitVector is #PF, CR2 value will be updated from the VMCB's
5973	* exit info. fields, if it's a guest #PF, see hmR0SvmExitXcptPF().
5974	*/
5975	Assert(pVmcb->ctrl.ExitIntInfo.n.u3Type != SVM_EVENT_SOFTWARE_INT);
5976	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectPendingReflect);
5977	hmR0SvmSetPendingEvent(pVCpu, &pVmcb->ctrl.ExitIntInfo, GCPtrFaultAddress);
5978
5979	Log4Func(("IDT: Pending vectoring event %#RX64 ErrValid=%RTbool Err=%#RX32 GCPtrFaultAddress=%#RX64\n",
5980	pVmcb->ctrl.ExitIntInfo.u, RT_BOOL(pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid),
5981	pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode, GCPtrFaultAddress));
5982	}
5983	break;
5984	}
5985
5986	case IEMXCPTRAISE_REEXEC_INSTR:
5987	{
5988	Assert(rc == VINF_SUCCESS);
5989	break;
5990	}
5991
5992	case IEMXCPTRAISE_DOUBLE_FAULT:
5993	{
5994	/*
5995	* Determing a vectoring double #PF condition. Used later, when PGM evaluates
5996	* the second #PF as a guest #PF (and not a shadow #PF) and needs to be
5997	* converted into a #DF.
5998	*/
5999	if (fRaiseInfo & IEMXCPTRAISEINFO_PF_PF)
6000	{
6001	Log4Func(("IDT: Pending vectoring double #PF uCR2=%#RX64\n", pVCpu->cpum.GstCtx.cr2));
6002	pSvmTransient->fVectoringDoublePF = true;
6003	Assert(rc == VINF_SUCCESS);
6004	}
6005	else
6006	{
6007	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectPendingReflect);
6008	hmR0SvmSetPendingXcptDF(pVCpu);
6009	rc = VINF_HM_DOUBLE_FAULT;
6010	}
6011	break;
6012	}
6013
6014	case IEMXCPTRAISE_TRIPLE_FAULT:
6015	{
6016	rc = VINF_EM_RESET;
6017	break;
6018	}
6019
6020	case IEMXCPTRAISE_CPU_HANG:
6021	{
6022	rc = VERR_EM_GUEST_CPU_HANG;
6023	break;
6024	}
6025
6026	default:
6027	AssertMsgFailedBreakStmt(("Bogus enmRaise value: %d (%#x)\n", enmRaise, enmRaise), rc = VERR_SVM_IPE_2);
6028	}
6029	}
6030	Assert(rc == VINF_SUCCESS \|\| rc == VINF_HM_DOUBLE_FAULT \|\| rc == VINF_EM_RESET \|\| rc == VERR_EM_GUEST_CPU_HANG);
6031	return rc;
6032	}
6033
6034
6035	/**
6036	* Advances the guest RIP making use of the CPU's NRIP_SAVE feature if
6037	* supported, otherwise advances the RIP by the number of bytes specified in
6038	* @a cb.
6039	*
6040	* @param pVCpu The cross context virtual CPU structure.
6041	* @param cb RIP increment value in bytes when the CPU doesn't support
6042	* NRIP_SAVE.
6043	*/
6044	DECLINLINE(void) hmR0SvmAdvanceRipHwAssist(PVMCPU pVCpu, uint32_t cb)
6045	{
6046	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
6047	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
6048	if (fSupportsNextRipSave)
6049	{
6050	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
6051	Assert(pVmcb);
6052	Assert(!(pCtx->fExtrn & CPUMCTX_EXTRN_RIP));
6053	Assert(pVmcb->ctrl.u64NextRIP - pCtx->rip == cb);
6054	pCtx->rip = pVmcb->ctrl.u64NextRIP;
6055	}
6056	else
6057	pCtx->rip += cb;
6058
6059	HMSVM_UPDATE_INTR_SHADOW(pVCpu);
6060	}
6061
6062
6063	/**
6064	* Gets the length of the current instruction when the CPU supports the NRIP_SAVE
6065	* feature.
6066	*
6067	* @returns The current instruction length in bytes.
6068	* @param pVCpu The cross context virtual CPU structure.
6069	*
6070	* @remarks Requires the NRIP_SAVE feature to be supported by the CPU.
6071	*/
6072	DECLINLINE(uint8_t) hmR0SvmGetInstrLength(PVMCPU pVCpu)
6073	{
6074	Assert(hmR0SvmSupportsNextRipSave(pVCpu));
6075	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
6076	return pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
6077	}
6078
6079
6080	/**
6081	* Advances the guest RIP by the number of bytes specified in @a cb. This does
6082	* not make use of any hardware features to determine the instruction length.
6083	*
6084	* @param pVCpu The cross context virtual CPU structure.
6085	* @param cb RIP increment value in bytes.
6086	*/
6087	DECLINLINE(void) hmR0SvmAdvanceRipDumb(PVMCPU pVCpu, uint32_t cb)
6088	{
6089	pVCpu->cpum.GstCtx.rip += cb;
6090	HMSVM_UPDATE_INTR_SHADOW(pVCpu);
6091	}
6092	#undef HMSVM_UPDATE_INTR_SHADOW
6093
6094
6095	/* -=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= */
6096	/* -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- #VMEXIT handlers -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- */
6097	/* -=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= */
6098
6099	/** @name \#VMEXIT handlers.
6100	* @{
6101	*/
6102
6103	/**
6104	* \#VMEXIT handler for external interrupts, NMIs, FPU assertion freeze and INIT
6105	* signals (SVM_EXIT_INTR, SVM_EXIT_NMI, SVM_EXIT_FERR_FREEZE, SVM_EXIT_INIT).
6106	*/
6107	HMSVM_EXIT_DECL hmR0SvmExitIntr(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
6108	{
6109	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6110
6111	if (pSvmTransient->u64ExitCode == SVM_EXIT_NMI)
6112	STAM_REL_COUNTER_INC(&pVCpu->hm.s.StatExitHostNmiInGC);
6113	else if (pSvmTransient->u64ExitCode == SVM_EXIT_INTR)
6114	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitExtInt);
6115
6116	/*
6117	* AMD-V has no preemption timer and the generic periodic preemption timer has no way to
6118	* signal -before- the timer fires if the current interrupt is our own timer or a some
6119	* other host interrupt. We also cannot examine what interrupt it is until the host
6120	* actually take the interrupt.
6121	*
6122	* Going back to executing guest code here unconditionally causes random scheduling
6123	* problems (observed on an AMD Phenom 9850 Quad-Core on Windows 64-bit host).
6124	*/
6125	return VINF_EM_RAW_INTERRUPT;
6126	}
6127
6128
6129	/**
6130	* \#VMEXIT handler for WBINVD (SVM_EXIT_WBINVD). Conditional \#VMEXIT.
6131	*/
6132	HMSVM_EXIT_DECL hmR0SvmExitWbinvd(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
6133	{
6134	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6135
6136	VBOXSTRICTRC rcStrict;
6137	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
6138	if (fSupportsNextRipSave)
6139	{
6140	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK);
6141	uint8_t const cbInstr = hmR0SvmGetInstrLength(pVCpu);
6142	rcStrict = IEMExecDecodedWbinvd(pVCpu, cbInstr);
6143	}
6144	else
6145	{
6146	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6147	rcStrict = IEMExecOne(pVCpu);
6148	}
6149
6150	if (rcStrict == VINF_IEM_RAISED_XCPT)
6151	{
6152	rcStrict = VINF_SUCCESS;
6153	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
6154	}
6155	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6156	return VBOXSTRICTRC_TODO(rcStrict);
6157	}
6158
6159
6160	/**
6161	* \#VMEXIT handler for INVD (SVM_EXIT_INVD). Unconditional \#VMEXIT.
6162	*/
6163	HMSVM_EXIT_DECL hmR0SvmExitInvd(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
6164	{
6165	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6166
6167	VBOXSTRICTRC rcStrict;
6168	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
6169	if (fSupportsNextRipSave)
6170	{
6171	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK);
6172	uint8_t const cbInstr = hmR0SvmGetInstrLength(pVCpu);
6173	rcStrict = IEMExecDecodedInvd(pVCpu, cbInstr);
6174	}
6175	else
6176	{
6177	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6178	rcStrict = IEMExecOne(pVCpu);
6179	}
6180
6181	if (rcStrict == VINF_IEM_RAISED_XCPT)
6182	{
6183	rcStrict = VINF_SUCCESS;
6184	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
6185	}
6186	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6187	return VBOXSTRICTRC_TODO(rcStrict);
6188	}
6189
6190
6191	/**
6192	* \#VMEXIT handler for INVD (SVM_EXIT_CPUID). Conditional \#VMEXIT.
6193	*/
6194	HMSVM_EXIT_DECL hmR0SvmExitCpuid(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
6195	{
6196	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6197
6198	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK \| CPUMCTX_EXTRN_RAX \| CPUMCTX_EXTRN_RCX);
6199	VBOXSTRICTRC rcStrict;
6200	PCEMEXITREC pExitRec = EMHistoryUpdateFlagsAndTypeAndPC(pVCpu,
6201	EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_CPUID),
6202	pVCpu->cpum.GstCtx.rip + pVCpu->cpum.GstCtx.cs.u64Base);
6203	if (!pExitRec)
6204	{
6205	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
6206	if (fSupportsNextRipSave)
6207	{
6208	uint8_t const cbInstr = hmR0SvmGetInstrLength(pVCpu);
6209	rcStrict = IEMExecDecodedCpuid(pVCpu, cbInstr);
6210	}
6211	else
6212	{
6213	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6214	rcStrict = IEMExecOne(pVCpu);
6215	}
6216
6217	if (rcStrict == VINF_IEM_RAISED_XCPT)
6218	{
6219	rcStrict = VINF_SUCCESS;
6220	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
6221	}
6222	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6223	}
6224	else
6225	{
6226	/*
6227	* Frequent exit or something needing probing. Get state and call EMHistoryExec.
6228	*/
6229	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6230
6231	Log4(("CpuIdExit/%u: %04x:%08RX64: %#x/%#x -> EMHistoryExec\n",
6232	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip, pVCpu->cpum.GstCtx.eax, pVCpu->cpum.GstCtx.ecx));
6233
6234	rcStrict = EMHistoryExec(pVCpu, pExitRec, 0);
6235
6236	Log4(("CpuIdExit/%u: %04x:%08RX64: EMHistoryExec -> %Rrc + %04x:%08RX64\n",
6237	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip,
6238	VBOXSTRICTRC_VAL(rcStrict), pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip));
6239	}
6240	return VBOXSTRICTRC_TODO(rcStrict);
6241	}
6242
6243
6244	/**
6245	* \#VMEXIT handler for RDTSC (SVM_EXIT_RDTSC). Conditional \#VMEXIT.
6246	*/
6247	HMSVM_EXIT_DECL hmR0SvmExitRdtsc(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
6248	{
6249	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6250
6251	VBOXSTRICTRC rcStrict;
6252	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
6253	if (fSupportsNextRipSave)
6254	{
6255	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK \| CPUMCTX_EXTRN_CR4);
6256	uint8_t const cbInstr = hmR0SvmGetInstrLength(pVCpu);
6257	rcStrict = IEMExecDecodedRdtsc(pVCpu, cbInstr);
6258	}
6259	else
6260	{
6261	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6262	rcStrict = IEMExecOne(pVCpu);
6263	}
6264
6265	if (rcStrict == VINF_SUCCESS)
6266	pSvmTransient->fUpdateTscOffsetting = true;
6267	else if (rcStrict == VINF_IEM_RAISED_XCPT)
6268	{
6269	rcStrict = VINF_SUCCESS;
6270	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
6271	}
6272	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6273	return VBOXSTRICTRC_TODO(rcStrict);
6274	}
6275
6276
6277	/**
6278	* \#VMEXIT handler for RDTSCP (SVM_EXIT_RDTSCP). Conditional \#VMEXIT.
6279	*/
6280	HMSVM_EXIT_DECL hmR0SvmExitRdtscp(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
6281	{
6282	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6283
6284	VBOXSTRICTRC rcStrict;
6285	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
6286	if (fSupportsNextRipSave)
6287	{
6288	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK \| CPUMCTX_EXTRN_CR4
6289	\| CPUMCTX_EXTRN_TSC_AUX);
6290	uint8_t const cbInstr = hmR0SvmGetInstrLength(pVCpu);
6291	rcStrict = IEMExecDecodedRdtscp(pVCpu, cbInstr);
6292	}
6293	else
6294	{
6295	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6296	rcStrict = IEMExecOne(pVCpu);
6297	}
6298
6299	if (rcStrict == VINF_SUCCESS)
6300	pSvmTransient->fUpdateTscOffsetting = true;
6301	else if (rcStrict == VINF_IEM_RAISED_XCPT)
6302	{
6303	rcStrict = VINF_SUCCESS;
6304	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
6305	}
6306	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6307	return VBOXSTRICTRC_TODO(rcStrict);
6308	}
6309
6310
6311	/**
6312	* \#VMEXIT handler for RDPMC (SVM_EXIT_RDPMC). Conditional \#VMEXIT.
6313	*/
6314	HMSVM_EXIT_DECL hmR0SvmExitRdpmc(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
6315	{
6316	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6317	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR0 \| CPUMCTX_EXTRN_CR4 \| CPUMCTX_EXTRN_SS);
6318
6319	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
6320	int rc = EMInterpretRdpmc(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
6321	if (RT_LIKELY(rc == VINF_SUCCESS))
6322	{
6323	hmR0SvmAdvanceRipHwAssist(pVCpu, 2);
6324	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6325	}
6326	else
6327	{
6328	AssertMsgFailed(("hmR0SvmExitRdpmc: EMInterpretRdpmc failed with %Rrc\n", rc));
6329	rc = VERR_EM_INTERPRETER;
6330	}
6331	return rc;
6332	}
6333
6334
6335	/**
6336	* \#VMEXIT handler for INVLPG (SVM_EXIT_INVLPG). Conditional \#VMEXIT.
6337	*/
6338	HMSVM_EXIT_DECL hmR0SvmExitInvlpg(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
6339	{
6340	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6341	Assert(!pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
6342
6343	VBOXSTRICTRC rcStrict;
6344	bool const fSupportsDecodeAssists = hmR0SvmSupportsDecodeAssists(pVCpu);
6345	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
6346	if ( fSupportsDecodeAssists
6347	&& fSupportsNextRipSave)
6348	{
6349	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6350	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
6351	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pVCpu->cpum.GstCtx.rip;
6352	RTGCPTR const GCPtrPage = pVmcb->ctrl.u64ExitInfo1;
6353	rcStrict = IEMExecDecodedInvlpg(pVCpu, cbInstr, GCPtrPage);
6354	}
6355	else
6356	{
6357	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6358	rcStrict = IEMExecOne(pVCpu);
6359	}
6360
6361	if (rcStrict == VINF_IEM_RAISED_XCPT)
6362	{
6363	rcStrict = VINF_SUCCESS;
6364	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
6365	}
6366	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6367	return VBOXSTRICTRC_VAL(rcStrict);
6368	}
6369
6370
6371	/**
6372	* \#VMEXIT handler for HLT (SVM_EXIT_HLT). Conditional \#VMEXIT.
6373	*/
6374	HMSVM_EXIT_DECL hmR0SvmExitHlt(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
6375	{
6376	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6377
6378	hmR0SvmAdvanceRipHwAssist(pVCpu, 1);
6379	int rc = EMShouldContinueAfterHalt(pVCpu, &pVCpu->cpum.GstCtx) ? VINF_SUCCESS : VINF_EM_HALT;
6380	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6381
6382	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitHlt);
6383	if (rc != VINF_SUCCESS)
6384	STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHltToR3);
6385	return rc;
6386	}
6387
6388
6389	/**
6390	* \#VMEXIT handler for MONITOR (SVM_EXIT_MONITOR). Conditional \#VMEXIT.
6391	*/
6392	HMSVM_EXIT_DECL hmR0SvmExitMonitor(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
6393	{
6394	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6395	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR0 \| CPUMCTX_EXTRN_SS);
6396
6397	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
6398	int rc = EMInterpretMonitor(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
6399	if (RT_LIKELY(rc == VINF_SUCCESS))
6400	{
6401	hmR0SvmAdvanceRipHwAssist(pVCpu, 3);
6402	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6403	}
6404	else
6405	{
6406	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMonitor: EMInterpretMonitor failed with %Rrc\n", rc));
6407	rc = VERR_EM_INTERPRETER;
6408	}
6409	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitMonitor);
6410	return rc;
6411	}
6412
6413
6414	/**
6415	* \#VMEXIT handler for MWAIT (SVM_EXIT_MWAIT). Conditional \#VMEXIT.
6416	*/
6417	HMSVM_EXIT_DECL hmR0SvmExitMwait(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
6418	{
6419	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6420	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR0 \| CPUMCTX_EXTRN_SS);
6421
6422	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
6423	VBOXSTRICTRC rc2 = EMInterpretMWait(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
6424	int rc = VBOXSTRICTRC_VAL(rc2);
6425	if ( rc == VINF_EM_HALT
6426	\|\| rc == VINF_SUCCESS)
6427	{
6428	hmR0SvmAdvanceRipHwAssist(pVCpu, 3);
6429
6430	if ( rc == VINF_EM_HALT
6431	&& EMMonitorWaitShouldContinue(pVCpu, pCtx))
6432	{
6433	rc = VINF_SUCCESS;
6434	}
6435	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6436	}
6437	else
6438	{
6439	AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMwait: EMInterpretMWait failed with %Rrc\n", rc));
6440	rc = VERR_EM_INTERPRETER;
6441	}
6442	AssertMsg(rc == VINF_SUCCESS \|\| rc == VINF_EM_HALT \|\| rc == VERR_EM_INTERPRETER,
6443	("hmR0SvmExitMwait: EMInterpretMWait failed rc=%Rrc\n", rc));
6444	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitMwait);
6445	return rc;
6446	}
6447
6448
6449	/**
6450	* \#VMEXIT handler for shutdown (triple-fault) (SVM_EXIT_SHUTDOWN). Conditional
6451	* \#VMEXIT.
6452	*/
6453	HMSVM_EXIT_DECL hmR0SvmExitShutdown(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
6454	{
6455	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6456	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6457	return VINF_EM_RESET;
6458	}
6459
6460
6461	/**
6462	* \#VMEXIT handler for unexpected exits. Conditional \#VMEXIT.
6463	*/
6464	HMSVM_EXIT_DECL hmR0SvmExitUnexpected(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
6465	{
6466	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
6467	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6468	AssertMsgFailed(("hmR0SvmExitUnexpected: ExitCode=%#RX64 uExitInfo1=%#RX64 uExitInfo2=%#RX64\n", pSvmTransient->u64ExitCode,
6469	pVmcb->ctrl.u64ExitInfo1, pVmcb->ctrl.u64ExitInfo2));
6470	RT_NOREF(pVmcb);
6471	pVCpu->hm.s.u32HMError = (uint32_t)pSvmTransient->u64ExitCode;
6472	return VERR_SVM_UNEXPECTED_EXIT;
6473	}
6474
6475
6476	/**
6477	* \#VMEXIT handler for CRx reads (SVM_EXIT_READ_CR*). Conditional \#VMEXIT.
6478	*/
6479	HMSVM_EXIT_DECL hmR0SvmExitReadCRx(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
6480	{
6481	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6482
6483	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
6484	Log4Func(("CS:RIP=%04x:%#RX64\n", pCtx->cs.Sel, pCtx->rip));
6485	#ifdef VBOX_WITH_STATISTICS
6486	switch (pSvmTransient->u64ExitCode)
6487	{
6488	case SVM_EXIT_READ_CR0: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR0Read); break;
6489	case SVM_EXIT_READ_CR2: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR2Read); break;
6490	case SVM_EXIT_READ_CR3: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR3Read); break;
6491	case SVM_EXIT_READ_CR4: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR4Read); break;
6492	case SVM_EXIT_READ_CR8: STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR8Read); break;
6493	}
6494	#endif
6495
6496	bool const fSupportsDecodeAssists = hmR0SvmSupportsDecodeAssists(pVCpu);
6497	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
6498	if ( fSupportsDecodeAssists
6499	&& fSupportsNextRipSave)
6500	{
6501	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
6502	bool const fMovCRx = RT_BOOL(pVmcb->ctrl.u64ExitInfo1 & SVM_EXIT1_MOV_CRX_MASK);
6503	if (fMovCRx)
6504	{
6505	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6506	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pCtx->rip;
6507	uint8_t const iCrReg = pSvmTransient->u64ExitCode - SVM_EXIT_READ_CR0;
6508	uint8_t const iGReg = pVmcb->ctrl.u64ExitInfo1 & SVM_EXIT1_MOV_CRX_GPR_NUMBER;
6509	VBOXSTRICTRC rcStrict = IEMExecDecodedMovCRxRead(pVCpu, cbInstr, iGReg, iCrReg);
6510	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6511	return VBOXSTRICTRC_VAL(rcStrict);
6512	}
6513	/* else: SMSW instruction, fall back below to IEM for this. */
6514	}
6515
6516	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6517	VBOXSTRICTRC rc2 = EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */);
6518	int rc = VBOXSTRICTRC_VAL(rc2);
6519	AssertMsg(rc == VINF_SUCCESS \|\| rc == VERR_EM_INTERPRETER \|\| rc == VINF_PGM_CHANGE_MODE \|\| rc == VINF_PGM_SYNC_CR3,
6520	("hmR0SvmExitReadCRx: EMInterpretInstruction failed rc=%Rrc\n", rc));
6521	Assert((pSvmTransient->u64ExitCode - SVM_EXIT_READ_CR0) <= 15);
6522	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6523	return rc;
6524	}
6525
6526
6527	/**
6528	* \#VMEXIT handler for CRx writes (SVM_EXIT_WRITE_CR*). Conditional \#VMEXIT.
6529	*/
6530	HMSVM_EXIT_DECL hmR0SvmExitWriteCRx(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
6531	{
6532	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6533
6534	uint64_t const uExitCode = pSvmTransient->u64ExitCode;
6535	uint8_t const iCrReg = uExitCode == SVM_EXIT_CR0_SEL_WRITE ? 0 : (pSvmTransient->u64ExitCode - SVM_EXIT_WRITE_CR0);
6536	Assert(iCrReg <= 15);
6537
6538	VBOXSTRICTRC rcStrict = VERR_SVM_IPE_5;
6539	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
6540	bool fDecodedInstr = false;
6541	bool const fSupportsDecodeAssists = hmR0SvmSupportsDecodeAssists(pVCpu);
6542	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
6543	if ( fSupportsDecodeAssists
6544	&& fSupportsNextRipSave)
6545	{
6546	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
6547	bool const fMovCRx = RT_BOOL(pVmcb->ctrl.u64ExitInfo1 & SVM_EXIT1_MOV_CRX_MASK);
6548	if (fMovCRx)
6549	{
6550	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6551	uint8_t const cbInstr = pVmcb->ctrl.u64NextRIP - pCtx->rip;
6552	uint8_t const iGReg = pVmcb->ctrl.u64ExitInfo1 & SVM_EXIT1_MOV_CRX_GPR_NUMBER;
6553	Log4Func(("Mov CR%u w/ iGReg=%#x\n", iCrReg, iGReg));
6554	rcStrict = IEMExecDecodedMovCRxWrite(pVCpu, cbInstr, iCrReg, iGReg);
6555	fDecodedInstr = true;
6556	}
6557	/* else: LMSW or CLTS instruction, fall back below to IEM for this. */
6558	}
6559
6560	if (!fDecodedInstr)
6561	{
6562	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6563	Log4Func(("iCrReg=%#x\n", iCrReg));
6564	rcStrict = IEMExecOneBypassEx(pVCpu, CPUMCTX2CORE(pCtx), NULL);
6565	if (RT_UNLIKELY( rcStrict == VERR_IEM_ASPECT_NOT_IMPLEMENTED
6566	\|\| rcStrict == VERR_IEM_INSTR_NOT_IMPLEMENTED))
6567	rcStrict = VERR_EM_INTERPRETER;
6568	}
6569
6570	if (rcStrict == VINF_SUCCESS)
6571	{
6572	switch (iCrReg)
6573	{
6574	case 0:
6575	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_CR0);
6576	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR0Write);
6577	break;
6578
6579	case 2:
6580	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_CR2);
6581	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR2Write);
6582	break;
6583
6584	case 3:
6585	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_CR3);
6586	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR3Write);
6587	break;
6588
6589	case 4:
6590	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_CR4);
6591	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR4Write);
6592	break;
6593
6594	case 8:
6595	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
6596	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCR8Write);
6597	break;
6598
6599	default:
6600	{
6601	AssertMsgFailed(("hmR0SvmExitWriteCRx: Invalid/Unexpected Write-CRx exit. u64ExitCode=%#RX64 %#x\n",
6602	pSvmTransient->u64ExitCode, iCrReg));
6603	break;
6604	}
6605	}
6606	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6607	}
6608	else
6609	Assert(rcStrict == VERR_EM_INTERPRETER \|\| rcStrict == VINF_PGM_CHANGE_MODE \|\| rcStrict == VINF_PGM_SYNC_CR3);
6610	return VBOXSTRICTRC_TODO(rcStrict);
6611	}
6612
6613
6614	/**
6615	* \#VMEXIT helper for read MSRs, see hmR0SvmExitMsr.
6616	*
6617	* @returns Strict VBox status code.
6618	* @param pVCpu The cross context virtual CPU structure.
6619	* @param pVmcb Pointer to the VM control block.
6620	*/
6621	static VBOXSTRICTRC hmR0SvmExitReadMsr(PVMCPU pVCpu, PSVMVMCB pVmcb)
6622	{
6623	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
6624	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdmsr);
6625	Log4Func(("idMsr=%#RX32\n", pCtx->ecx));
6626
6627	VBOXSTRICTRC rcStrict;
6628	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
6629	if (fSupportsNextRipSave)
6630	{
6631	/** @todo Optimize this: Only retrieve the MSR bits we need here. CPUMAllMsrs.cpp
6632	* can ask for what it needs instead of using CPUMCTX_EXTRN_ALL_MSRS. */
6633	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK \| CPUMCTX_EXTRN_ALL_MSRS);
6634	rcStrict = IEMExecDecodedRdmsr(pVCpu, pVmcb->ctrl.u64NextRIP - pCtx->rip);
6635	}
6636	else
6637	{
6638	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK \| CPUMCTX_EXTRN_ALL_MSRS);
6639	rcStrict = IEMExecOne(pVCpu);
6640	}
6641
6642	AssertMsg( rcStrict == VINF_SUCCESS
6643	\|\| rcStrict == VINF_IEM_RAISED_XCPT
6644	\|\| rcStrict == VINF_CPUM_R3_MSR_READ,
6645	("hmR0SvmExitReadMsr: Unexpected status %Rrc\n", VBOXSTRICTRC_VAL(rcStrict)));
6646
6647	if (rcStrict == VINF_IEM_RAISED_XCPT)
6648	{
6649	rcStrict = VINF_SUCCESS;
6650	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
6651	}
6652	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6653	return rcStrict;
6654	}
6655
6656
6657	/**
6658	* \#VMEXIT helper for write MSRs, see hmR0SvmExitMsr.
6659	*
6660	* @returns Strict VBox status code.
6661	* @param pVCpu The cross context virtual CPU structure.
6662	* @param pVmcb Pointer to the VM control block.
6663	* @param pSvmTransient Pointer to the SVM-transient structure.
6664	*/
6665	static VBOXSTRICTRC hmR0SvmExitWriteMsr(PVMCPU pVCpu, PSVMVMCB pVmcb, PSVMTRANSIENT pSvmTransient)
6666	{
6667	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
6668	uint32_t const idMsr = pCtx->ecx;
6669	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitWrmsr);
6670	Log4Func(("idMsr=%#RX32\n", idMsr));
6671
6672	/*
6673	* Handle TPR patching MSR writes.
6674	* We utilitize the LSTAR MSR for patching.
6675	*/
6676	if ( pVCpu->CTX_SUFF(pVM)->hm.s.fTPRPatchingActive
6677	&& idMsr == MSR_K8_LSTAR)
6678	{
6679	if ((pCtx->eax & 0xff) != pSvmTransient->u8GuestTpr)
6680	{
6681	/* Our patch code uses LSTAR for TPR caching for 32-bit guests. */
6682	int rc2 = APICSetTpr(pVCpu, pCtx->eax & 0xff);
6683	AssertRC(rc2);
6684	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
6685	}
6686
6687	int rc = VINF_SUCCESS;
6688	hmR0SvmAdvanceRipHwAssist(pVCpu, 2);
6689	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6690	return rc;
6691	}
6692
6693	/*
6694	* Handle regular MSR writes.
6695	*/
6696	VBOXSTRICTRC rcStrict;
6697	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
6698	if (fSupportsNextRipSave)
6699	{
6700	/** @todo Optimize this: We don't need to get much of the MSR state here
6701	* since we're only updating. CPUMAllMsrs.cpp can ask for what it needs and
6702	* clear the applicable extern flags. */
6703	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_EXEC_DECODED_NO_MEM_MASK \| CPUMCTX_EXTRN_ALL_MSRS);
6704	rcStrict = IEMExecDecodedWrmsr(pVCpu, pVmcb->ctrl.u64NextRIP - pCtx->rip);
6705	}
6706	else
6707	{
6708	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK \| CPUMCTX_EXTRN_ALL_MSRS);
6709	rcStrict = IEMExecOne(pVCpu);
6710	}
6711
6712	AssertMsg( rcStrict == VINF_SUCCESS
6713	\|\| rcStrict == VINF_IEM_RAISED_XCPT
6714	\|\| rcStrict == VINF_CPUM_R3_MSR_WRITE,
6715	("hmR0SvmExitWriteMsr: Unexpected status %Rrc\n", VBOXSTRICTRC_VAL(rcStrict)));
6716
6717	if (rcStrict == VINF_SUCCESS)
6718	{
6719	/* If this is an X2APIC WRMSR access, update the APIC TPR state. */
6720	if ( idMsr >= MSR_IA32_X2APIC_START
6721	&& idMsr <= MSR_IA32_X2APIC_END)
6722	{
6723	/*
6724	* We've already saved the APIC related guest-state (TPR) in hmR0SvmPostRunGuest().
6725	* When full APIC register virtualization is implemented we'll have to make sure
6726	* APIC state is saved from the VMCB before IEM changes it.
6727	*/
6728	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
6729	}
6730	else
6731	{
6732	switch (idMsr)
6733	{
6734	case MSR_IA32_TSC: pSvmTransient->fUpdateTscOffsetting = true; break;
6735	case MSR_K6_EFER: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_EFER_MSR); break;
6736	case MSR_K8_FS_BASE: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_FS); break;
6737	case MSR_K8_GS_BASE: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_GS); break;
6738	case MSR_IA32_SYSENTER_CS: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_SYSENTER_CS_MSR); break;
6739	case MSR_IA32_SYSENTER_EIP: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_SYSENTER_EIP_MSR); break;
6740	case MSR_IA32_SYSENTER_ESP: ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_SYSENTER_ESP_MSR); break;
6741	}
6742	}
6743	}
6744	else if (rcStrict == VINF_IEM_RAISED_XCPT)
6745	{
6746	rcStrict = VINF_SUCCESS;
6747	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
6748	}
6749	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6750	return rcStrict;
6751	}
6752
6753
6754	/**
6755	* \#VMEXIT handler for MSR read and writes (SVM_EXIT_MSR). Conditional
6756	* \#VMEXIT.
6757	*/
6758	HMSVM_EXIT_DECL hmR0SvmExitMsr(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
6759	{
6760	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6761
6762	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
6763	if (pVmcb->ctrl.u64ExitInfo1 == SVM_EXIT1_MSR_READ)
6764	return VBOXSTRICTRC_TODO(hmR0SvmExitReadMsr(pVCpu, pVmcb));
6765
6766	Assert(pVmcb->ctrl.u64ExitInfo1 == SVM_EXIT1_MSR_WRITE);
6767	return VBOXSTRICTRC_TODO(hmR0SvmExitWriteMsr(pVCpu, pVmcb, pSvmTransient));
6768	}
6769
6770
6771	/**
6772	* \#VMEXIT handler for DRx read (SVM_EXIT_READ_DRx). Conditional \#VMEXIT.
6773	*/
6774	HMSVM_EXIT_DECL hmR0SvmExitReadDRx(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
6775	{
6776	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6777	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
6778
6779	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitDRxRead);
6780
6781	/** @todo Stepping with nested-guest. */
6782	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
6783	if (!CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
6784	{
6785	/* We should -not- get this #VMEXIT if the guest's debug registers were active. */
6786	if (pSvmTransient->fWasGuestDebugStateActive)
6787	{
6788	AssertMsgFailed(("hmR0SvmExitReadDRx: Unexpected exit %#RX32\n", (uint32_t)pSvmTransient->u64ExitCode));
6789	pVCpu->hm.s.u32HMError = (uint32_t)pSvmTransient->u64ExitCode;
6790	return VERR_SVM_UNEXPECTED_EXIT;
6791	}
6792
6793	/*
6794	* Lazy DR0-3 loading.
6795	*/
6796	if (!pSvmTransient->fWasHyperDebugStateActive)
6797	{
6798	Assert(!DBGFIsStepping(pVCpu)); Assert(!pVCpu->hm.s.fSingleInstruction);
6799	Log5(("hmR0SvmExitReadDRx: Lazy loading guest debug registers\n"));
6800
6801	/* Don't intercept DRx read and writes. */
6802	PSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
6803	pVmcb->ctrl.u16InterceptRdDRx = 0;
6804	pVmcb->ctrl.u16InterceptWrDRx = 0;
6805	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
6806
6807	/* We're playing with the host CPU state here, make sure we don't preempt or longjmp. */
6808	VMMRZCallRing3Disable(pVCpu);
6809	HM_DISABLE_PREEMPT(pVCpu);
6810
6811	/* Save the host & load the guest debug state, restart execution of the MOV DRx instruction. */
6812	CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
6813	Assert(CPUMIsGuestDebugStateActive(pVCpu) \|\| HC_ARCH_BITS == 32);
6814
6815	HM_RESTORE_PREEMPT();
6816	VMMRZCallRing3Enable(pVCpu);
6817
6818	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxContextSwitch);
6819	return VINF_SUCCESS;
6820	}
6821	}
6822
6823	/*
6824	* Interpret the read/writing of DRx.
6825	*/
6826	/** @todo Decode assist. */
6827	VBOXSTRICTRC rc = EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */);
6828	Log5(("hmR0SvmExitReadDRx: Emulated DRx access: rc=%Rrc\n", VBOXSTRICTRC_VAL(rc)));
6829	if (RT_LIKELY(rc == VINF_SUCCESS))
6830	{
6831	/* Not necessary for read accesses but whatever doesn't hurt for now, will be fixed with decode assist. */
6832	/** @todo CPUM should set this flag! */
6833	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_DR_MASK);
6834	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
6835	}
6836	else
6837	Assert(rc == VERR_EM_INTERPRETER);
6838	return VBOXSTRICTRC_TODO(rc);
6839	}
6840
6841
6842	/**
6843	* \#VMEXIT handler for DRx write (SVM_EXIT_WRITE_DRx). Conditional \#VMEXIT.
6844	*/
6845	HMSVM_EXIT_DECL hmR0SvmExitWriteDRx(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
6846	{
6847	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6848	/* For now it's the same since we interpret the instruction anyway. Will change when using of Decode Assist is implemented. */
6849	int rc = hmR0SvmExitReadDRx(pVCpu, pSvmTransient);
6850	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitDRxWrite);
6851	STAM_COUNTER_DEC(&pVCpu->hm.s.StatExitDRxRead);
6852	return rc;
6853	}
6854
6855
6856	/**
6857	* \#VMEXIT handler for XCRx write (SVM_EXIT_XSETBV). Conditional \#VMEXIT.
6858	*/
6859	HMSVM_EXIT_DECL hmR0SvmExitXsetbv(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
6860	{
6861	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6862	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
6863
6864	/** @todo decode assists... */
6865	VBOXSTRICTRC rcStrict = IEMExecOne(pVCpu);
6866	if (RT_LIKELY(rcStrict == VINF_SUCCESS))
6867	{
6868	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
6869	pVCpu->hm.s.fLoadSaveGuestXcr0 = (pCtx->cr4 & X86_CR4_OSXSAVE) && pCtx->aXcr[0] != ASMGetXcr0();
6870	Log4Func(("New XCR0=%#RX64 fLoadSaveGuestXcr0=%RTbool (cr4=%#RX64)\n", pCtx->aXcr[0], pVCpu->hm.s.fLoadSaveGuestXcr0,
6871	pCtx->cr4));
6872	}
6873	else if (rcStrict == VINF_IEM_RAISED_XCPT)
6874	{
6875	rcStrict = VINF_SUCCESS;
6876	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
6877	}
6878	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
6879	return VBOXSTRICTRC_TODO(rcStrict);
6880	}
6881
6882
6883	/**
6884	* \#VMEXIT handler for I/O instructions (SVM_EXIT_IOIO). Conditional \#VMEXIT.
6885	*/
6886	HMSVM_EXIT_DECL hmR0SvmExitIOInstr(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
6887	{
6888	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
6889	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK \| CPUMCTX_EXTRN_SREG_MASK);
6890
6891	/* I/O operation lookup arrays. */
6892	static uint32_t const s_aIOSize[8] = { 0, 1, 2, 0, 4, 0, 0, 0 }; /* Size of the I/O accesses in bytes. */
6893	static uint32_t const s_aIOOpAnd[8] = { 0, 0xff, 0xffff, 0, 0xffffffff, 0, 0, 0 }; /* AND masks for saving
6894	the result (in AL/AX/EAX). */
6895	PVM pVM = pVCpu->CTX_SUFF(pVM);
6896	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
6897	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
6898
6899	Log4Func(("CS:RIP=%04x:%#RX64\n", pCtx->cs.Sel, pCtx->rip));
6900
6901	/* Refer AMD spec. 15.10.2 "IN and OUT Behaviour" and Figure 15-2. "EXITINFO1 for IOIO Intercept" for the format. */
6902	SVMIOIOEXITINFO IoExitInfo;
6903	IoExitInfo.u = (uint32_t)pVmcb->ctrl.u64ExitInfo1;
6904	uint32_t uIOWidth = (IoExitInfo.u >> 4) & 0x7;
6905	uint32_t cbValue = s_aIOSize[uIOWidth];
6906	uint32_t uAndVal = s_aIOOpAnd[uIOWidth];
6907
6908	if (RT_UNLIKELY(!cbValue))
6909	{
6910	AssertMsgFailed(("hmR0SvmExitIOInstr: Invalid IO operation. uIOWidth=%u\n", uIOWidth));
6911	return VERR_EM_INTERPRETER;
6912	}
6913
6914	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CS \| CPUMCTX_EXTRN_RIP \| CPUMCTX_EXTRN_RFLAGS);
6915	VBOXSTRICTRC rcStrict;
6916	PCEMEXITREC pExitRec = NULL;
6917	if ( !pVCpu->hm.s.fSingleInstruction
6918	&& !pVCpu->cpum.GstCtx.eflags.Bits.u1TF)
6919	pExitRec = EMHistoryUpdateFlagsAndTypeAndPC(pVCpu,
6920	!IoExitInfo.n.u1Str
6921	? IoExitInfo.n.u1Type == SVM_IOIO_READ
6922	? EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_IO_PORT_READ)
6923	: EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_IO_PORT_WRITE)
6924	: IoExitInfo.n.u1Type == SVM_IOIO_READ
6925	? EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_IO_PORT_STR_READ)
6926	: EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_IO_PORT_STR_WRITE),
6927	pVCpu->cpum.GstCtx.rip + pVCpu->cpum.GstCtx.cs.u64Base);
6928	if (!pExitRec)
6929	{
6930	bool fUpdateRipAlready = false;
6931	if (IoExitInfo.n.u1Str)
6932	{
6933	/* INS/OUTS - I/O String instruction. */
6934	/** @todo Huh? why can't we use the segment prefix information given by AMD-V
6935	* in EXITINFO1? Investigate once this thing is up and running. */
6936	Log4Func(("CS:RIP=%04x:%08RX64 %#06x/%u %c str\n", pCtx->cs.Sel, pCtx->rip, IoExitInfo.n.u16Port, cbValue,
6937	IoExitInfo.n.u1Type == SVM_IOIO_WRITE ? 'w' : 'r'));
6938	AssertReturn(pCtx->dx == IoExitInfo.n.u16Port, VERR_SVM_IPE_2);
6939	static IEMMODE const s_aenmAddrMode[8] =
6940	{
6941	(IEMMODE)-1, IEMMODE_16BIT, IEMMODE_32BIT, (IEMMODE)-1, IEMMODE_64BIT, (IEMMODE)-1, (IEMMODE)-1, (IEMMODE)-1
6942	};
6943	IEMMODE enmAddrMode = s_aenmAddrMode[(IoExitInfo.u >> 7) & 0x7];
6944	if (enmAddrMode != (IEMMODE)-1)
6945	{
6946	uint64_t cbInstr = pVmcb->ctrl.u64ExitInfo2 - pCtx->rip;
6947	if (cbInstr <= 15 && cbInstr >= 1)
6948	{
6949	Assert(cbInstr >= 1U + IoExitInfo.n.u1Rep);
6950	if (IoExitInfo.n.u1Type == SVM_IOIO_WRITE)
6951	{
6952	/* Don't know exactly how to detect whether u3Seg is valid, currently
6953	only enabling it for Bulldozer and later with NRIP. OS/2 broke on
6954	2384 Opterons when only checking NRIP. */
6955	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
6956	if ( fSupportsNextRipSave
6957	&& pVM->cpum.ro.GuestFeatures.enmMicroarch >= kCpumMicroarch_AMD_15h_First)
6958	{
6959	AssertMsg(IoExitInfo.n.u3Seg == X86_SREG_DS \|\| cbInstr > 1U + IoExitInfo.n.u1Rep,
6960	("u32Seg=%d cbInstr=%d u1REP=%d", IoExitInfo.n.u3Seg, cbInstr, IoExitInfo.n.u1Rep));
6961	rcStrict = IEMExecStringIoWrite(pVCpu, cbValue, enmAddrMode, IoExitInfo.n.u1Rep, (uint8_t)cbInstr,
6962	IoExitInfo.n.u3Seg, true /fIoChecked/);
6963	}
6964	else if (cbInstr == 1U + IoExitInfo.n.u1Rep)
6965	rcStrict = IEMExecStringIoWrite(pVCpu, cbValue, enmAddrMode, IoExitInfo.n.u1Rep, (uint8_t)cbInstr,
6966	X86_SREG_DS, true /fIoChecked/);
6967	else
6968	rcStrict = IEMExecOne(pVCpu);
6969	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOStringWrite);
6970	}
6971	else
6972	{
6973	AssertMsg(IoExitInfo.n.u3Seg == X86_SREG_ES /=0/, ("%#x\n", IoExitInfo.n.u3Seg));
6974	rcStrict = IEMExecStringIoRead(pVCpu, cbValue, enmAddrMode, IoExitInfo.n.u1Rep, (uint8_t)cbInstr,
6975	true /fIoChecked/);
6976	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOStringRead);
6977	}
6978	}
6979	else
6980	{
6981	AssertMsgFailed(("rip=%RX64 nrip=%#RX64 cbInstr=%#RX64\n", pCtx->rip, pVmcb->ctrl.u64ExitInfo2, cbInstr));
6982	rcStrict = IEMExecOne(pVCpu);
6983	}
6984	}
6985	else
6986	{
6987	AssertMsgFailed(("IoExitInfo=%RX64\n", IoExitInfo.u));
6988	rcStrict = IEMExecOne(pVCpu);
6989	}
6990	fUpdateRipAlready = true;
6991	}
6992	else
6993	{
6994	/* IN/OUT - I/O instruction. */
6995	Assert(!IoExitInfo.n.u1Rep);
6996
6997	if (IoExitInfo.n.u1Type == SVM_IOIO_WRITE)
6998	{
6999	rcStrict = IOMIOPortWrite(pVM, pVCpu, IoExitInfo.n.u16Port, pCtx->eax & uAndVal, cbValue);
7000	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOWrite);
7001	}
7002	else
7003	{
7004	uint32_t u32Val = 0;
7005	rcStrict = IOMIOPortRead(pVM, pVCpu, IoExitInfo.n.u16Port, &u32Val, cbValue);
7006	if (IOM_SUCCESS(rcStrict))
7007	{
7008	/* Save result of I/O IN instr. in AL/AX/EAX. */
7009	/** @todo r=bird: 32-bit op size should clear high bits of rax! */
7010	pCtx->eax = (pCtx->eax & ~uAndVal) \| (u32Val & uAndVal);
7011	}
7012	else if (rcStrict == VINF_IOM_R3_IOPORT_READ)
7013	{
7014	HMR0SavePendingIOPortRead(pVCpu, pVCpu->cpum.GstCtx.rip, pVmcb->ctrl.u64ExitInfo2, IoExitInfo.n.u16Port,
7015	uAndVal, cbValue);
7016	}
7017
7018	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIORead);
7019	}
7020	}
7021
7022	if (IOM_SUCCESS(rcStrict))
7023	{
7024	/* AMD-V saves the RIP of the instruction following the IO instruction in EXITINFO2. */
7025	if (!fUpdateRipAlready)
7026	pCtx->rip = pVmcb->ctrl.u64ExitInfo2;
7027
7028	/*
7029	* If any I/O breakpoints are armed, we need to check if one triggered
7030	* and take appropriate action.
7031	* Note that the I/O breakpoint type is undefined if CR4.DE is 0.
7032	*/
7033	/** @todo Optimize away the DBGFBpIsHwIoArmed call by having DBGF tell the
7034	* execution engines about whether hyper BPs and such are pending. */
7035	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_DR7);
7036	uint32_t const uDr7 = pCtx->dr[7];
7037	if (RT_UNLIKELY( ( (uDr7 & X86_DR7_ENABLED_MASK)
7038	&& X86_DR7_ANY_RW_IO(uDr7)
7039	&& (pCtx->cr4 & X86_CR4_DE))
7040	\|\| DBGFBpIsHwIoArmed(pVM)))
7041	{
7042	/* We're playing with the host CPU state here, make sure we don't preempt or longjmp. */
7043	VMMRZCallRing3Disable(pVCpu);
7044	HM_DISABLE_PREEMPT(pVCpu);
7045
7046	STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxIoCheck);
7047	CPUMR0DebugStateMaybeSaveGuest(pVCpu, false /fDr6/);
7048
7049	VBOXSTRICTRC rcStrict2 = DBGFBpCheckIo(pVM, pVCpu, &pVCpu->cpum.GstCtx, IoExitInfo.n.u16Port, cbValue);
7050	if (rcStrict2 == VINF_EM_RAW_GUEST_TRAP)
7051	{
7052	/* Raise #DB. */
7053	pVmcb->guest.u64DR6 = pCtx->dr[6];
7054	pVmcb->guest.u64DR7 = pCtx->dr[7];
7055	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
7056	hmR0SvmSetPendingXcptDB(pVCpu);
7057	}
7058	/* rcStrict is VINF_SUCCESS, VINF_IOM_R3_IOPORT_COMMIT_WRITE, or in [VINF_EM_FIRST..VINF_EM_LAST],
7059	however we can ditch VINF_IOM_R3_IOPORT_COMMIT_WRITE as it has VMCPU_FF_IOM as backup. */
7060	else if ( rcStrict2 != VINF_SUCCESS
7061	&& (rcStrict == VINF_SUCCESS \|\| rcStrict2 < rcStrict))
7062	rcStrict = rcStrict2;
7063	AssertCompile(VINF_EM_LAST < VINF_IOM_R3_IOPORT_COMMIT_WRITE);
7064
7065	HM_RESTORE_PREEMPT();
7066	VMMRZCallRing3Enable(pVCpu);
7067	}
7068
7069	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
7070	}
7071
7072	#ifdef VBOX_STRICT
7073	if (rcStrict == VINF_IOM_R3_IOPORT_READ)
7074	Assert(IoExitInfo.n.u1Type == SVM_IOIO_READ);
7075	else if (rcStrict == VINF_IOM_R3_IOPORT_WRITE \|\| rcStrict == VINF_IOM_R3_IOPORT_COMMIT_WRITE)
7076	Assert(IoExitInfo.n.u1Type == SVM_IOIO_WRITE);
7077	else
7078	{
7079	/** @todo r=bird: This is missing a bunch of VINF_EM_FIRST..VINF_EM_LAST
7080	* statuses, that the VMM device and some others may return. See
7081	* IOM_SUCCESS() for guidance. */
7082	AssertMsg( RT_FAILURE(rcStrict)
7083	\|\| rcStrict == VINF_SUCCESS
7084	\|\| rcStrict == VINF_EM_RAW_EMULATE_INSTR
7085	\|\| rcStrict == VINF_EM_DBG_BREAKPOINT
7086	\|\| rcStrict == VINF_EM_RAW_GUEST_TRAP
7087	\|\| rcStrict == VINF_EM_RAW_TO_R3
7088	\|\| rcStrict == VINF_TRPM_XCPT_DISPATCHED
7089	\|\| rcStrict == VINF_EM_TRIPLE_FAULT, ("%Rrc\n", VBOXSTRICTRC_VAL(rcStrict)));
7090	}
7091	#endif
7092	}
7093	else
7094	{
7095	/*
7096	* Frequent exit or something needing probing. Get state and call EMHistoryExec.
7097	*/
7098	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7099	STAM_COUNTER_INC(!IoExitInfo.n.u1Str
7100	? IoExitInfo.n.u1Type == SVM_IOIO_WRITE ? &pVCpu->hm.s.StatExitIOWrite : &pVCpu->hm.s.StatExitIORead
7101	: IoExitInfo.n.u1Type == SVM_IOIO_WRITE ? &pVCpu->hm.s.StatExitIOStringWrite : &pVCpu->hm.s.StatExitIOStringRead);
7102	Log4(("IOExit/%u: %04x:%08RX64: %s%s%s %#x LB %u -> EMHistoryExec\n",
7103	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip, IoExitInfo.n.u1Rep ? "REP " : "",
7104	IoExitInfo.n.u1Type == SVM_IOIO_WRITE ? "OUT" : "IN", IoExitInfo.n.u1Str ? "S" : "", IoExitInfo.n.u16Port, uIOWidth));
7105
7106	rcStrict = EMHistoryExec(pVCpu, pExitRec, 0);
7107	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
7108
7109	Log4(("IOExit/%u: %04x:%08RX64: EMHistoryExec -> %Rrc + %04x:%08RX64\n",
7110	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip,
7111	VBOXSTRICTRC_VAL(rcStrict), pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip));
7112	}
7113	return VBOXSTRICTRC_TODO(rcStrict);
7114	}
7115
7116
7117	/**
7118	* \#VMEXIT handler for Nested Page-faults (SVM_EXIT_NPF). Conditional \#VMEXIT.
7119	*/
7120	HMSVM_EXIT_DECL hmR0SvmExitNestedPF(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
7121	{
7122	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7123	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7124	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(pVCpu, pSvmTransient);
7125
7126	PVM pVM = pVCpu->CTX_SUFF(pVM);
7127	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
7128	Assert(pVM->hm.s.fNestedPaging);
7129
7130	/* See AMD spec. 15.25.6 "Nested versus Guest Page Faults, Fault Ordering" for VMCB details for #NPF. */
7131	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7132	RTGCPHYS GCPhysFaultAddr = pVmcb->ctrl.u64ExitInfo2;
7133	uint32_t u32ErrCode = pVmcb->ctrl.u64ExitInfo1; /* Note! High bits in EXITINFO1 may contain additional info and are
7134	thus intentionally not copied into u32ErrCode. */
7135
7136	Log4Func(("#NPF at CS:RIP=%04x:%#RX64 GCPhysFaultAddr=%RGp ErrCode=%#x \n", pCtx->cs.Sel, pCtx->rip, GCPhysFaultAddr,
7137	u32ErrCode));
7138
7139	/*
7140	* TPR patching for 32-bit guests, using the reserved bit in the page tables for MMIO regions.
7141	*/
7142	if ( pVM->hm.s.fTprPatchingAllowed
7143	&& (GCPhysFaultAddr & PAGE_OFFSET_MASK) == XAPIC_OFF_TPR
7144	&& ( !(u32ErrCode & X86_TRAP_PF_P) /* Not present */
7145	\|\| (u32ErrCode & (X86_TRAP_PF_P \| X86_TRAP_PF_RSVD)) == (X86_TRAP_PF_P \| X86_TRAP_PF_RSVD)) /* MMIO page. */
7146	&& !CPUMIsGuestInSvmNestedHwVirtMode(pCtx)
7147	&& !CPUMIsGuestInLongModeEx(pCtx)
7148	&& !CPUMGetGuestCPL(pVCpu)
7149	&& pVM->hm.s.cPatches < RT_ELEMENTS(pVM->hm.s.aPatches))
7150	{
7151	RTGCPHYS GCPhysApicBase = APICGetBaseMsrNoCheck(pVCpu);
7152	GCPhysApicBase &= PAGE_BASE_GC_MASK;
7153
7154	if (GCPhysFaultAddr == GCPhysApicBase + XAPIC_OFF_TPR)
7155	{
7156	/* Only attempt to patch the instruction once. */
7157	PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
7158	if (!pPatch)
7159	return VINF_EM_HM_PATCH_TPR_INSTR;
7160	}
7161	}
7162
7163	/*
7164	* Determine the nested paging mode.
7165	*/
7166	PGMMODE enmNestedPagingMode;
7167	#if HC_ARCH_BITS == 32
7168	if (CPUMIsGuestInLongModeEx(pCtx))
7169	enmNestedPagingMode = PGMMODE_AMD64_NX;
7170	else
7171	#endif
7172	enmNestedPagingMode = PGMGetHostMode(pVM);
7173
7174	/*
7175	* MMIO optimization using the reserved (RSVD) bit in the guest page tables for MMIO pages.
7176	*/
7177	Assert((u32ErrCode & (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P)) != X86_TRAP_PF_RSVD);
7178	if ((u32ErrCode & (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P)) == (X86_TRAP_PF_RSVD \| X86_TRAP_PF_P))
7179	{
7180	/*
7181	* If event delivery causes an MMIO #NPF, go back to instruction emulation as otherwise
7182	* injecting the original pending event would most likely cause the same MMIO #NPF.
7183	*/
7184	if (pVCpu->hm.s.Event.fPending)
7185	return VINF_EM_RAW_INJECT_TRPM_EVENT;
7186
7187	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CS \| CPUMCTX_EXTRN_RIP);
7188	VBOXSTRICTRC rcStrict;
7189	PCEMEXITREC pExitRec = EMHistoryUpdateFlagsAndTypeAndPC(pVCpu,
7190	EMEXIT_MAKE_FT(EMEXIT_F_KIND_EM \| EMEXIT_F_HM, EMEXITTYPE_MMIO),
7191	pVCpu->cpum.GstCtx.rip + pVCpu->cpum.GstCtx.cs.u64Base);
7192	if (!pExitRec)
7193	{
7194
7195	rcStrict = PGMR0Trap0eHandlerNPMisconfig(pVM, pVCpu, enmNestedPagingMode, CPUMCTX2CORE(pCtx), GCPhysFaultAddr,
7196	u32ErrCode);
7197
7198	/*
7199	* If we succeed, resume guest execution.
7200	*
7201	* If we fail in interpreting the instruction because we couldn't get the guest
7202	* physical address of the page containing the instruction via the guest's page
7203	* tables (we would invalidate the guest page in the host TLB), resume execution
7204	* which would cause a guest page fault to let the guest handle this weird case.
7205	*
7206	* See @bugref{6043}.
7207	*/
7208	if ( rcStrict == VINF_SUCCESS
7209	\|\| rcStrict == VERR_PAGE_TABLE_NOT_PRESENT
7210	\|\| rcStrict == VERR_PAGE_NOT_PRESENT)
7211	{
7212	/* Successfully handled MMIO operation. */
7213	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_APIC_TPR);
7214	rcStrict = VINF_SUCCESS;
7215	}
7216	}
7217	else
7218	{
7219	/*
7220	* Frequent exit or something needing probing. Get state and call EMHistoryExec.
7221	*/
7222	Assert(pCtx == &pVCpu->cpum.GstCtx);
7223	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7224	Log4(("EptMisscfgExit/%u: %04x:%08RX64: %RGp -> EMHistoryExec\n",
7225	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip, GCPhysFaultAddr));
7226
7227	rcStrict = EMHistoryExec(pVCpu, pExitRec, 0);
7228	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
7229
7230	Log4(("EptMisscfgExit/%u: %04x:%08RX64: EMHistoryExec -> %Rrc + %04x:%08RX64\n",
7231	pVCpu->idCpu, pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip,
7232	VBOXSTRICTRC_VAL(rcStrict), pVCpu->cpum.GstCtx.cs.Sel, pVCpu->cpum.GstCtx.rip));
7233	}
7234	return VBOXSTRICTRC_TODO(rcStrict);
7235	}
7236
7237	TRPMAssertXcptPF(pVCpu, GCPhysFaultAddr, u32ErrCode);
7238	int rc = PGMR0Trap0eHandlerNestedPaging(pVM, pVCpu, enmNestedPagingMode, u32ErrCode, CPUMCTX2CORE(pCtx), GCPhysFaultAddr);
7239	TRPMResetTrap(pVCpu);
7240
7241	Log4Func(("#NPF: PGMR0Trap0eHandlerNestedPaging returns %Rrc CS:RIP=%04x:%#RX64\n", rc, pCtx->cs.Sel, pCtx->rip));
7242
7243	/*
7244	* Same case as PGMR0Trap0eHandlerNPMisconfig(). See comment above, @bugref{6043}.
7245	*/
7246	if ( rc == VINF_SUCCESS
7247	\|\| rc == VERR_PAGE_TABLE_NOT_PRESENT
7248	\|\| rc == VERR_PAGE_NOT_PRESENT)
7249	{
7250	/* We've successfully synced our shadow page tables. */
7251	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPF);
7252	rc = VINF_SUCCESS;
7253	}
7254
7255	return rc;
7256	}
7257
7258
7259	/**
7260	* \#VMEXIT handler for virtual interrupt (SVM_EXIT_VINTR). Conditional
7261	* \#VMEXIT.
7262	*/
7263	HMSVM_EXIT_DECL hmR0SvmExitVIntr(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
7264	{
7265	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7266	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(&pVCpu->cpum.GstCtx);
7267
7268	/* Indicate that we no longer need to #VMEXIT when the guest is ready to receive NMIs, it is now ready. */
7269	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7270	hmR0SvmClearIntWindowExiting(pVCpu, pVmcb);
7271
7272	/* Deliver the pending interrupt via hmR0SvmEvaluatePendingEvent() and resume guest execution. */
7273	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIntWindow);
7274	return VINF_SUCCESS;
7275	}
7276
7277
7278	/**
7279	* \#VMEXIT handler for task switches (SVM_EXIT_TASK_SWITCH). Conditional
7280	* \#VMEXIT.
7281	*/
7282	HMSVM_EXIT_DECL hmR0SvmExitTaskSwitch(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
7283	{
7284	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7285	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(pVCpu, pSvmTransient);
7286
7287	#ifndef HMSVM_ALWAYS_TRAP_TASK_SWITCH
7288	Assert(!pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
7289	#endif
7290
7291	/* Check if this task-switch occurred while delivering an event through the guest IDT. */
7292	if (pVCpu->hm.s.Event.fPending) /* Can happen with exceptions/NMI. See @bugref{8411}. */
7293	{
7294	/*
7295	* AMD-V provides us with the exception which caused the TS; we collect
7296	* the information in the call to hmR0SvmCheckExitDueToEventDelivery().
7297	*/
7298	Log4Func(("TS occurred during event delivery\n"));
7299	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitTaskSwitch);
7300	return VINF_EM_RAW_INJECT_TRPM_EVENT;
7301	}
7302
7303	/** @todo Emulate task switch someday, currently just going back to ring-3 for
7304	* emulation. */
7305	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitTaskSwitch);
7306	return VERR_EM_INTERPRETER;
7307	}
7308
7309
7310	/**
7311	* \#VMEXIT handler for VMMCALL (SVM_EXIT_VMMCALL). Conditional \#VMEXIT.
7312	*/
7313	HMSVM_EXIT_DECL hmR0SvmExitVmmCall(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
7314	{
7315	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7316	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7317
7318	if (pVCpu->CTX_SUFF(pVM)->hm.s.fTprPatchingAllowed)
7319	{
7320	int rc = hmSvmEmulateMovTpr(pVCpu);
7321	if (rc != VERR_NOT_FOUND)
7322	{
7323	Log4Func(("hmSvmEmulateMovTpr returns %Rrc\n", rc));
7324	return rc;
7325	}
7326	}
7327
7328	if (EMAreHypercallInstructionsEnabled(pVCpu))
7329	{
7330	VBOXSTRICTRC rcStrict = GIMHypercall(pVCpu, &pVCpu->cpum.GstCtx);
7331	if (RT_SUCCESS(rcStrict))
7332	{
7333	/* Only update the RIP if we're continuing guest execution and not in the case
7334	of say VINF_GIM_R3_HYPERCALL. */
7335	if (rcStrict == VINF_SUCCESS)
7336	hmR0SvmAdvanceRipHwAssist(pVCpu, 3 /* cbInstr */);
7337
7338	return VBOXSTRICTRC_VAL(rcStrict);
7339	}
7340	else
7341	Log4Func(("GIMHypercall returns %Rrc -> #UD\n", VBOXSTRICTRC_VAL(rcStrict)));
7342	}
7343
7344	hmR0SvmSetPendingXcptUD(pVCpu);
7345	return VINF_SUCCESS;
7346	}
7347
7348
7349	/**
7350	* \#VMEXIT handler for VMMCALL (SVM_EXIT_VMMCALL). Conditional \#VMEXIT.
7351	*/
7352	HMSVM_EXIT_DECL hmR0SvmExitPause(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
7353	{
7354	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7355	hmR0SvmAdvanceRipHwAssist(pVCpu, 2);
7356	/** @todo The guest has likely hit a contended spinlock. We might want to
7357	* poke a schedule different guest VCPU. */
7358	return VINF_EM_RAW_INTERRUPT;
7359	}
7360
7361
7362	/**
7363	* \#VMEXIT handler for FERR intercept (SVM_EXIT_FERR_FREEZE). Conditional
7364	* \#VMEXIT.
7365	*/
7366	HMSVM_EXIT_DECL hmR0SvmExitFerrFreeze(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
7367	{
7368	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7369	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, CPUMCTX_EXTRN_CR0);
7370	Assert(!(pVCpu->cpum.GstCtx.cr0 & X86_CR0_NE));
7371
7372	Log4Func(("Raising IRQ 13 in response to #FERR\n"));
7373	return PDMIsaSetIrq(pVCpu->CTX_SUFF(pVM), 13 /* u8Irq /, 1 / u8Level /, 0 / uTagSrc */);
7374	}
7375
7376
7377	/**
7378	* \#VMEXIT handler for IRET (SVM_EXIT_IRET). Conditional \#VMEXIT.
7379	*/
7380	HMSVM_EXIT_DECL hmR0SvmExitIret(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
7381	{
7382	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7383
7384	/* Clear NMI blocking. */
7385	if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_BLOCK_NMIS))
7386	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_BLOCK_NMIS);
7387
7388	/* Indicate that we no longer need to #VMEXIT when the guest is ready to receive NMIs, it is now ready. */
7389	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7390	hmR0SvmClearCtrlIntercept(pVCpu, pVmcb, SVM_CTRL_INTERCEPT_IRET);
7391
7392	/* Deliver the pending NMI via hmR0SvmEvaluatePendingEvent() and resume guest execution. */
7393	return VINF_SUCCESS;
7394	}
7395
7396
7397	/**
7398	* \#VMEXIT handler for page-fault exceptions (SVM_EXIT_XCPT_14).
7399	* Conditional \#VMEXIT.
7400	*/
7401	HMSVM_EXIT_DECL hmR0SvmExitXcptPF(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
7402	{
7403	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7404	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7405	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(pVCpu, pSvmTransient);
7406
7407	/* See AMD spec. 15.12.15 "#PF (Page Fault)". */
7408	PVM pVM = pVCpu->CTX_SUFF(pVM);
7409	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
7410	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7411	uint32_t uErrCode = pVmcb->ctrl.u64ExitInfo1;
7412	uint64_t const uFaultAddress = pVmcb->ctrl.u64ExitInfo2;
7413
7414	#if defined(HMSVM_ALWAYS_TRAP_ALL_XCPTS) \|\| defined(HMSVM_ALWAYS_TRAP_PF)
7415	if (pVM->hm.s.fNestedPaging)
7416	{
7417	pVCpu->hm.s.Event.fPending = false; /* In case it's a contributory or vectoring #PF. */
7418	if ( !pSvmTransient->fVectoringDoublePF
7419	\|\| CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
7420	{
7421	/* A genuine guest #PF, reflect it to the guest. */
7422	hmR0SvmSetPendingXcptPF(pVCpu, uErrCode, uFaultAddress);
7423	Log4Func(("#PF: Guest page fault at %04X:%RGv FaultAddr=%RX64 ErrCode=%#x\n", pCtx->cs.Sel, (RTGCPTR)pCtx->rip,
7424	uFaultAddress, uErrCode));
7425	}
7426	else
7427	{
7428	/* A guest page-fault occurred during delivery of a page-fault. Inject #DF. */
7429	hmR0SvmSetPendingXcptDF(pVCpu);
7430	Log4Func(("Pending #DF due to vectoring #PF. NP\n"));
7431	}
7432	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestPF);
7433	return VINF_SUCCESS;
7434	}
7435	#endif
7436
7437	Assert(!pVM->hm.s.fNestedPaging);
7438
7439	/*
7440	* TPR patching shortcut for APIC TPR reads and writes; only applicable to 32-bit guests.
7441	*/
7442	if ( pVM->hm.s.fTprPatchingAllowed
7443	&& (uFaultAddress & 0xfff) == XAPIC_OFF_TPR
7444	&& !(uErrCode & X86_TRAP_PF_P) /* Not present. */
7445	&& !CPUMIsGuestInSvmNestedHwVirtMode(pCtx)
7446	&& !CPUMIsGuestInLongModeEx(pCtx)
7447	&& !CPUMGetGuestCPL(pVCpu)
7448	&& pVM->hm.s.cPatches < RT_ELEMENTS(pVM->hm.s.aPatches))
7449	{
7450	RTGCPHYS GCPhysApicBase;
7451	GCPhysApicBase = APICGetBaseMsrNoCheck(pVCpu);
7452	GCPhysApicBase &= PAGE_BASE_GC_MASK;
7453
7454	/* Check if the page at the fault-address is the APIC base. */
7455	RTGCPHYS GCPhysPage;
7456	int rc2 = PGMGstGetPage(pVCpu, (RTGCPTR)uFaultAddress, NULL /* pfFlags */, &GCPhysPage);
7457	if ( rc2 == VINF_SUCCESS
7458	&& GCPhysPage == GCPhysApicBase)
7459	{
7460	/* Only attempt to patch the instruction once. */
7461	PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
7462	if (!pPatch)
7463	return VINF_EM_HM_PATCH_TPR_INSTR;
7464	}
7465	}
7466
7467	Log4Func(("#PF: uFaultAddress=%#RX64 CS:RIP=%#04x:%#RX64 uErrCode %#RX32 cr3=%#RX64\n", uFaultAddress, pCtx->cs.Sel,
7468	pCtx->rip, uErrCode, pCtx->cr3));
7469
7470	/*
7471	* If it's a vectoring #PF, emulate injecting the original event injection as
7472	* PGMTrap0eHandler() is incapable of differentiating between instruction emulation and
7473	* event injection that caused a #PF. See @bugref{6607}.
7474	*/
7475	if (pSvmTransient->fVectoringPF)
7476	{
7477	Assert(pVCpu->hm.s.Event.fPending);
7478	return VINF_EM_RAW_INJECT_TRPM_EVENT;
7479	}
7480
7481	TRPMAssertXcptPF(pVCpu, uFaultAddress, uErrCode);
7482	int rc = PGMTrap0eHandler(pVCpu, uErrCode, CPUMCTX2CORE(pCtx), (RTGCPTR)uFaultAddress);
7483
7484	Log4Func(("#PF: rc=%Rrc\n", rc));
7485
7486	if (rc == VINF_SUCCESS)
7487	{
7488	/* Successfully synced shadow pages tables or emulated an MMIO instruction. */
7489	TRPMResetTrap(pVCpu);
7490	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPF);
7491	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_ALL_GUEST);
7492	return rc;
7493	}
7494
7495	if (rc == VINF_EM_RAW_GUEST_TRAP)
7496	{
7497	pVCpu->hm.s.Event.fPending = false; /* In case it's a contributory or vectoring #PF. */
7498
7499	/*
7500	* If a nested-guest delivers a #PF and that causes a #PF which is -not- a shadow #PF,
7501	* we should simply forward the #PF to the guest and is up to the nested-hypervisor to
7502	* determine whether it is a nested-shadow #PF or a #DF, see @bugref{7243#c121}.
7503	*/
7504	if ( !pSvmTransient->fVectoringDoublePF
7505	\|\| CPUMIsGuestInSvmNestedHwVirtMode(pCtx))
7506	{
7507	/* It's a guest (or nested-guest) page fault and needs to be reflected. */
7508	uErrCode = TRPMGetErrorCode(pVCpu); /* The error code might have been changed. */
7509	TRPMResetTrap(pVCpu);
7510
7511	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
7512	/* If the nested-guest is intercepting #PFs, cause a #PF #VMEXIT. */
7513	if ( CPUMIsGuestInSvmNestedHwVirtMode(pCtx)
7514	&& HMIsGuestSvmXcptInterceptSet(pVCpu, X86_XCPT_PF))
7515	return VBOXSTRICTRC_TODO(IEMExecSvmVmexit(pVCpu, SVM_EXIT_XCPT_PF, uErrCode, uFaultAddress));
7516	#endif
7517
7518	hmR0SvmSetPendingXcptPF(pVCpu, uErrCode, uFaultAddress);
7519	}
7520	else
7521	{
7522	/* A guest page-fault occurred during delivery of a page-fault. Inject #DF. */
7523	TRPMResetTrap(pVCpu);
7524	hmR0SvmSetPendingXcptDF(pVCpu);
7525	Log4Func(("#PF: Pending #DF due to vectoring #PF\n"));
7526	}
7527
7528	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestPF);
7529	return VINF_SUCCESS;
7530	}
7531
7532	TRPMResetTrap(pVCpu);
7533	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPFEM);
7534	return rc;
7535	}
7536
7537
7538	/**
7539	* \#VMEXIT handler for undefined opcode (SVM_EXIT_XCPT_6).
7540	* Conditional \#VMEXIT.
7541	*/
7542	HMSVM_EXIT_DECL hmR0SvmExitXcptUD(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
7543	{
7544	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7545	HMSVM_ASSERT_NOT_IN_NESTED_GUEST(&pVCpu->cpum.GstCtx);
7546
7547	/* Paranoia; Ensure we cannot be called as a result of event delivery. */
7548	PSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
7549	Assert(!pVmcb->ctrl.ExitIntInfo.n.u1Valid); NOREF(pVmcb);
7550
7551	int rc = VERR_SVM_UNEXPECTED_XCPT_EXIT;
7552	if (pVCpu->hm.s.fGIMTrapXcptUD)
7553	{
7554	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7555	uint8_t cbInstr = 0;
7556	VBOXSTRICTRC rcStrict = GIMXcptUD(pVCpu, &pVCpu->cpum.GstCtx, NULL /* pDis */, &cbInstr);
7557	if (rcStrict == VINF_SUCCESS)
7558	{
7559	/* #UD #VMEXIT does not have valid NRIP information, manually advance RIP. See @bugref{7270#c170}. */
7560	hmR0SvmAdvanceRipDumb(pVCpu, cbInstr);
7561	rc = VINF_SUCCESS;
7562	HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
7563	}
7564	else if (rcStrict == VINF_GIM_HYPERCALL_CONTINUING)
7565	rc = VINF_SUCCESS;
7566	else if (rcStrict == VINF_GIM_R3_HYPERCALL)
7567	rc = VINF_GIM_R3_HYPERCALL;
7568	else
7569	Assert(RT_FAILURE(VBOXSTRICTRC_VAL(rcStrict)));
7570	}
7571
7572	/* If the GIM #UD exception handler didn't succeed for some reason or wasn't needed, raise #UD. */
7573	if (RT_FAILURE(rc))
7574	{
7575	hmR0SvmSetPendingXcptUD(pVCpu);
7576	rc = VINF_SUCCESS;
7577	}
7578
7579	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestUD);
7580	return rc;
7581	}
7582
7583
7584	/**
7585	* \#VMEXIT handler for math-fault exceptions (SVM_EXIT_XCPT_16).
7586	* Conditional \#VMEXIT.
7587	*/
7588	HMSVM_EXIT_DECL hmR0SvmExitXcptMF(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
7589	{
7590	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7591	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7592
7593	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
7594	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7595
7596	/* Paranoia; Ensure we cannot be called as a result of event delivery. */
7597	Assert(!pVmcb->ctrl.ExitIntInfo.n.u1Valid); NOREF(pVmcb);
7598
7599	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestMF);
7600
7601	if (!(pCtx->cr0 & X86_CR0_NE))
7602	{
7603	PVM pVM = pVCpu->CTX_SUFF(pVM);
7604	PDISSTATE pDis = &pVCpu->hm.s.DisState;
7605	unsigned cbOp;
7606	int rc = EMInterpretDisasCurrent(pVM, pVCpu, pDis, &cbOp);
7607	if (RT_SUCCESS(rc))
7608	{
7609	/* Convert a #MF into a FERR -> IRQ 13. See @bugref{6117}. */
7610	rc = PDMIsaSetIrq(pVCpu->CTX_SUFF(pVM), 13 /* u8Irq /, 1 / u8Level /, 0 / uTagSrc */);
7611	if (RT_SUCCESS(rc))
7612	pCtx->rip += cbOp;
7613	}
7614	else
7615	Log4Func(("EMInterpretDisasCurrent returned %Rrc uOpCode=%#x\n", rc, pDis->pCurInstr->uOpcode));
7616	return rc;
7617	}
7618
7619	hmR0SvmSetPendingXcptMF(pVCpu);
7620	return VINF_SUCCESS;
7621	}
7622
7623
7624	/**
7625	* \#VMEXIT handler for debug exceptions (SVM_EXIT_XCPT_1). Conditional
7626	* \#VMEXIT.
7627	*/
7628	HMSVM_EXIT_DECL hmR0SvmExitXcptDB(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
7629	{
7630	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7631	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7632	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(pVCpu, pSvmTransient);
7633
7634	if (RT_UNLIKELY(pVCpu->hm.s.Event.fPending))
7635	{
7636	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectPendingInterpret);
7637	return VINF_EM_RAW_INJECT_TRPM_EVENT;
7638	}
7639
7640	STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestDB);
7641
7642	/*
7643	* This can be a fault-type #DB (instruction breakpoint) or a trap-type #DB (data
7644	* breakpoint). However, for both cases DR6 and DR7 are updated to what the exception
7645	* handler expects. See AMD spec. 15.12.2 "#DB (Debug)".
7646	*/
7647	PVM pVM = pVCpu->CTX_SUFF(pVM);
7648	PSVMVMCB pVmcb = pVCpu->hm.s.svm.pVmcb;
7649	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
7650	int rc = DBGFRZTrap01Handler(pVM, pVCpu, CPUMCTX2CORE(pCtx), pVmcb->guest.u64DR6, pVCpu->hm.s.fSingleInstruction);
7651	if (rc == VINF_EM_RAW_GUEST_TRAP)
7652	{
7653	Log5(("hmR0SvmExitXcptDB: DR6=%#RX64 -> guest trap\n", pVmcb->guest.u64DR6));
7654	if (CPUMIsHyperDebugStateActive(pVCpu))
7655	CPUMSetGuestDR6(pVCpu, CPUMGetGuestDR6(pVCpu) \| pVmcb->guest.u64DR6);
7656
7657	/* Reflect the exception back to the guest. */
7658	hmR0SvmSetPendingXcptDB(pVCpu);
7659	rc = VINF_SUCCESS;
7660	}
7661
7662	/*
7663	* Update DR6.
7664	*/
7665	if (CPUMIsHyperDebugStateActive(pVCpu))
7666	{
7667	Log5(("hmR0SvmExitXcptDB: DR6=%#RX64 -> %Rrc\n", pVmcb->guest.u64DR6, rc));
7668	pVmcb->guest.u64DR6 = X86_DR6_INIT_VAL;
7669	pVmcb->ctrl.u32VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
7670	}
7671	else
7672	{
7673	AssertMsg(rc == VINF_SUCCESS, ("rc=%Rrc\n", rc));
7674	Assert(!pVCpu->hm.s.fSingleInstruction && !DBGFIsStepping(pVCpu));
7675	}
7676
7677	return rc;
7678	}
7679
7680
7681	/**
7682	* \#VMEXIT handler for alignment check exceptions (SVM_EXIT_XCPT_17).
7683	* Conditional \#VMEXIT.
7684	*/
7685	HMSVM_EXIT_DECL hmR0SvmExitXcptAC(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
7686	{
7687	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7688	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(pVCpu, pSvmTransient);
7689
7690	SVMEVENT Event;
7691	Event.u = 0;
7692	Event.n.u1Valid = 1;
7693	Event.n.u3Type = SVM_EVENT_EXCEPTION;
7694	Event.n.u8Vector = X86_XCPT_AC;
7695	Event.n.u1ErrorCodeValid = 1;
7696	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
7697	return VINF_SUCCESS;
7698	}
7699
7700
7701	/**
7702	* \#VMEXIT handler for breakpoint exceptions (SVM_EXIT_XCPT_3).
7703	* Conditional \#VMEXIT.
7704	*/
7705	HMSVM_EXIT_DECL hmR0SvmExitXcptBP(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
7706	{
7707	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7708	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, HMSVM_CPUMCTX_EXTRN_ALL);
7709	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(pVCpu, pSvmTransient);
7710
7711	PCPUMCTX pCtx = &pVCpu->cpum.GstCtx;
7712	int rc = DBGFRZTrap03Handler(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
7713	if (rc == VINF_EM_RAW_GUEST_TRAP)
7714	{
7715	SVMEVENT Event;
7716	Event.u = 0;
7717	Event.n.u1Valid = 1;
7718	Event.n.u3Type = SVM_EVENT_EXCEPTION;
7719	Event.n.u8Vector = X86_XCPT_BP;
7720	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
7721	}
7722
7723	Assert(rc == VINF_SUCCESS \|\| rc == VINF_EM_RAW_GUEST_TRAP \|\| rc == VINF_EM_DBG_BREAKPOINT);
7724	return rc;
7725	}
7726
7727
7728	#if defined(HMSVM_ALWAYS_TRAP_ALL_XCPTS) \|\| defined(VBOX_WITH_NESTED_HWVIRT_SVM)
7729	/**
7730	* \#VMEXIT handler for generic exceptions. Conditional \#VMEXIT.
7731	*/
7732	HMSVM_EXIT_DECL hmR0SvmExitXcptGeneric(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
7733	{
7734	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7735	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(pVCpu, pSvmTransient);
7736
7737	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7738	uint8_t const uVector = pVmcb->ctrl.u64ExitCode - SVM_EXIT_XCPT_0;
7739	uint32_t const uErrCode = pVmcb->ctrl.u64ExitInfo1;
7740	Assert(pSvmTransient->u64ExitCode == pVmcb->ctrl.u64ExitCode);
7741	Assert(uVector <= X86_XCPT_LAST);
7742	Log4Func(("uVector=%#x uErrCode=%u\n", uVector, uErrCode));
7743
7744	SVMEVENT Event;
7745	Event.u = 0;
7746	Event.n.u1Valid = 1;
7747	Event.n.u3Type = SVM_EVENT_EXCEPTION;
7748	Event.n.u8Vector = uVector;
7749	switch (uVector)
7750	{
7751	/* Shouldn't be here for reflecting #PFs (among other things, the fault address isn't passed along). */
7752	case X86_XCPT_PF: AssertMsgFailed(("hmR0SvmExitXcptGeneric: Unexpected exception")); return VERR_SVM_IPE_5;
7753	case X86_XCPT_DF:
7754	case X86_XCPT_TS:
7755	case X86_XCPT_NP:
7756	case X86_XCPT_SS:
7757	case X86_XCPT_GP:
7758	case X86_XCPT_AC:
7759	{
7760	Event.n.u1ErrorCodeValid = 1;
7761	Event.n.u32ErrorCode = uErrCode;
7762	break;
7763	}
7764	}
7765
7766	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
7767	return VINF_SUCCESS;
7768	}
7769	#endif
7770
7771	#ifdef VBOX_WITH_NESTED_HWVIRT_SVM
7772	/**
7773	* \#VMEXIT handler for CLGI (SVM_EXIT_CLGI). Conditional \#VMEXIT.
7774	*/
7775	HMSVM_EXIT_DECL hmR0SvmExitClgi(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
7776	{
7777	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7778	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK \| CPUMCTX_EXTRN_HWVIRT);
7779
7780	#ifdef VBOX_STRICT
7781	PCSVMVMCB pVmcbTmp = hmR0SvmGetCurrentVmcb(pVCpu);
7782	Assert(pVmcbTmp);
7783	Assert(!pVmcbTmp->ctrl.IntCtrl.n.u1VGifEnable);
7784	RT_NOREF(pVmcbTmp);
7785	#endif
7786
7787	VBOXSTRICTRC rcStrict;
7788	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
7789	if (fSupportsNextRipSave)
7790	{
7791	uint8_t const cbInstr = hmR0SvmGetInstrLength(pVCpu);
7792	rcStrict = IEMExecDecodedClgi(pVCpu, cbInstr);
7793	}
7794	else
7795	rcStrict = IEMExecOne(pVCpu);
7796
7797	if (rcStrict == VINF_SUCCESS)
7798	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_HWVIRT);
7799	else if (rcStrict == VINF_IEM_RAISED_XCPT)
7800	{
7801	rcStrict = VINF_SUCCESS;
7802	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
7803	}
7804	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
7805	return VBOXSTRICTRC_TODO(rcStrict);
7806	}
7807
7808
7809	/**
7810	* \#VMEXIT handler for STGI (SVM_EXIT_STGI). Conditional \#VMEXIT.
7811	*/
7812	HMSVM_EXIT_DECL hmR0SvmExitStgi(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
7813	{
7814	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7815	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK \| CPUMCTX_EXTRN_HWVIRT);
7816
7817	/*
7818	* When VGIF is not used we always intercept STGI instructions. When VGIF is used,
7819	* we only intercept STGI when events are pending for GIF to become 1.
7820	*/
7821	PSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7822	if (pVmcb->ctrl.IntCtrl.n.u1VGifEnable)
7823	hmR0SvmClearCtrlIntercept(pVCpu, pVmcb, SVM_CTRL_INTERCEPT_STGI);
7824
7825	VBOXSTRICTRC rcStrict;
7826	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
7827	if (fSupportsNextRipSave)
7828	{
7829	uint8_t const cbInstr = hmR0SvmGetInstrLength(pVCpu);
7830	rcStrict = IEMExecDecodedStgi(pVCpu, cbInstr);
7831	}
7832	else
7833	rcStrict = IEMExecOne(pVCpu);
7834
7835	if (rcStrict == VINF_SUCCESS)
7836	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_HWVIRT);
7837	else if (rcStrict == VINF_IEM_RAISED_XCPT)
7838	{
7839	rcStrict = VINF_SUCCESS;
7840	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
7841	}
7842	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
7843	return VBOXSTRICTRC_TODO(rcStrict);
7844	}
7845
7846
7847	/**
7848	* \#VMEXIT handler for VMLOAD (SVM_EXIT_VMLOAD). Conditional \#VMEXIT.
7849	*/
7850	HMSVM_EXIT_DECL hmR0SvmExitVmload(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
7851	{
7852	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7853	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK
7854	\| CPUMCTX_EXTRN_FS \| CPUMCTX_EXTRN_GS \| CPUMCTX_EXTRN_TR
7855	\| CPUMCTX_EXTRN_LDTR \| CPUMCTX_EXTRN_KERNEL_GS_BASE \| CPUMCTX_EXTRN_SYSCALL_MSRS
7856	\| CPUMCTX_EXTRN_SYSENTER_MSRS);
7857
7858	#ifdef VBOX_STRICT
7859	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7860	Assert(pVmcb);
7861	Assert(!pVmcb->ctrl.LbrVirt.n.u1VirtVmsaveVmload);
7862	RT_NOREF(pVmcb);
7863	#endif
7864
7865	VBOXSTRICTRC rcStrict;
7866	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
7867	if (fSupportsNextRipSave)
7868	{
7869	uint8_t const cbInstr = hmR0SvmGetInstrLength(pVCpu);
7870	rcStrict = IEMExecDecodedVmload(pVCpu, cbInstr);
7871	}
7872	else
7873	rcStrict = IEMExecOne(pVCpu);
7874
7875	if (rcStrict == VINF_SUCCESS)
7876	{
7877	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_GUEST_FS \| HM_CHANGED_GUEST_GS
7878	\| HM_CHANGED_GUEST_TR \| HM_CHANGED_GUEST_LDTR
7879	\| HM_CHANGED_GUEST_KERNEL_GS_BASE \| HM_CHANGED_GUEST_SYSCALL_MSRS
7880	\| HM_CHANGED_GUEST_SYSENTER_MSR_MASK);
7881	}
7882	else if (rcStrict == VINF_IEM_RAISED_XCPT)
7883	{
7884	rcStrict = VINF_SUCCESS;
7885	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
7886	}
7887	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
7888	return VBOXSTRICTRC_TODO(rcStrict);
7889	}
7890
7891
7892	/**
7893	* \#VMEXIT handler for VMSAVE (SVM_EXIT_VMSAVE). Conditional \#VMEXIT.
7894	*/
7895	HMSVM_EXIT_DECL hmR0SvmExitVmsave(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
7896	{
7897	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7898	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
7899
7900	#ifdef VBOX_STRICT
7901	PCSVMVMCB pVmcb = hmR0SvmGetCurrentVmcb(pVCpu);
7902	Assert(pVmcb);
7903	Assert(!pVmcb->ctrl.LbrVirt.n.u1VirtVmsaveVmload);
7904	RT_NOREF(pVmcb);
7905	#endif
7906	VBOXSTRICTRC rcStrict;
7907	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
7908	if (fSupportsNextRipSave)
7909	{
7910	uint8_t const cbInstr = hmR0SvmGetInstrLength(pVCpu);
7911	rcStrict = IEMExecDecodedVmsave(pVCpu, cbInstr);
7912	}
7913	else
7914	rcStrict = IEMExecOne(pVCpu);
7915
7916	if (rcStrict == VINF_IEM_RAISED_XCPT)
7917	{
7918	rcStrict = VINF_SUCCESS;
7919	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
7920	}
7921	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
7922	return VBOXSTRICTRC_TODO(rcStrict);
7923	}
7924
7925
7926	/**
7927	* \#VMEXIT handler for INVLPGA (SVM_EXIT_INVLPGA). Conditional \#VMEXIT.
7928	*/
7929	HMSVM_EXIT_DECL hmR0SvmExitInvlpga(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
7930	{
7931	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7932	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK);
7933
7934	VBOXSTRICTRC rcStrict;
7935	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
7936	if (fSupportsNextRipSave)
7937	{
7938	uint8_t const cbInstr = hmR0SvmGetInstrLength(pVCpu);
7939	rcStrict = IEMExecDecodedInvlpga(pVCpu, cbInstr);
7940	}
7941	else
7942	rcStrict = IEMExecOne(pVCpu);
7943
7944	if (rcStrict == VINF_IEM_RAISED_XCPT)
7945	{
7946	rcStrict = VINF_SUCCESS;
7947	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
7948	}
7949	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
7950	return VBOXSTRICTRC_TODO(rcStrict);
7951	}
7952
7953
7954	/**
7955	* \#VMEXIT handler for STGI (SVM_EXIT_VMRUN). Conditional \#VMEXIT.
7956	*/
7957	HMSVM_EXIT_DECL hmR0SvmExitVmrun(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
7958	{
7959	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7960	HMSVM_CPUMCTX_IMPORT_STATE(pVCpu, IEM_CPUMCTX_EXTRN_MUST_MASK \| IEM_CPUMCTX_EXTRN_SVM_VMRUN_MASK);
7961
7962	VBOXSTRICTRC rcStrict;
7963	bool const fSupportsNextRipSave = hmR0SvmSupportsNextRipSave(pVCpu);
7964	if (fSupportsNextRipSave)
7965	{
7966	uint8_t const cbInstr = hmR0SvmGetInstrLength(pVCpu);
7967	rcStrict = IEMExecDecodedVmrun(pVCpu, cbInstr);
7968	}
7969	else
7970	rcStrict = IEMExecOne(pVCpu);
7971
7972	if (rcStrict == VINF_SUCCESS)
7973	{
7974	rcStrict = VINF_SVM_VMRUN;
7975	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_SVM_VMRUN_MASK);
7976	}
7977	else if (rcStrict == VINF_IEM_RAISED_XCPT)
7978	{
7979	rcStrict = VINF_SUCCESS;
7980	ASMAtomicUoOrU64(&pVCpu->hm.s.fCtxChanged, HM_CHANGED_RAISED_XCPT_MASK);
7981	}
7982
7983	HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
7984	return VBOXSTRICTRC_TODO(rcStrict);
7985	}
7986
7987
7988	/**
7989	* Nested-guest \#VMEXIT handler for debug exceptions (SVM_EXIT_XCPT_1).
7990	* Unconditional \#VMEXIT.
7991	*/
7992	HMSVM_EXIT_DECL hmR0SvmNestedExitXcptDB(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
7993	{
7994	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
7995	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(pVCpu, pSvmTransient);
7996
7997	if (pVCpu->hm.s.Event.fPending)
7998	{
7999	STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectPendingInterpret);
8000	return VINF_EM_RAW_INJECT_TRPM_EVENT;
8001	}
8002
8003	hmR0SvmSetPendingXcptDB(pVCpu);
8004	return VINF_SUCCESS;
8005	}
8006
8007
8008	/**
8009	* Nested-guest \#VMEXIT handler for breakpoint exceptions (SVM_EXIT_XCPT_3).
8010	* Conditional \#VMEXIT.
8011	*/
8012	HMSVM_EXIT_DECL hmR0SvmNestedExitXcptBP(PVMCPU pVCpu, PSVMTRANSIENT pSvmTransient)
8013	{
8014	HMSVM_VALIDATE_EXIT_HANDLER_PARAMS(pVCpu, pSvmTransient);
8015	HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY(pVCpu, pSvmTransient);
8016
8017	SVMEVENT Event;
8018	Event.u = 0;
8019	Event.n.u1Valid = 1;
8020	Event.n.u3Type = SVM_EVENT_EXCEPTION;
8021	Event.n.u8Vector = X86_XCPT_BP;
8022	hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
8023	return VINF_SUCCESS;
8024	}
8025	#endif /* VBOX_WITH_NESTED_HWVIRT_SVM */
8026
8027	/** @} */
8028

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

source: vbox/trunk/src/VBox/VMM/VMMR0/HMSVMR0.cpp@ 73047

Download in other formats: