HWVMXR0.cpp@ 8973

Last change on this file since 8973 was 8876, checked in by vboxsync, 17 years ago
ASID based TLB flushing
Property svn:eol-style set to `native` Property svn:keywords set to `Id`
File size: 78.7 KB

Line
1	/* $Id: HWVMXR0.cpp 8876 2008-05-16 09:59:07Z vboxsync $ */
2	/** @file
3	* HWACCM VMX - Host Context Ring 0.
4	*/
5
6	/*
7	* Copyright (C) 2006-2007 Sun Microsystems, Inc.
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	* Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa
18	* Clara, CA 95054 USA or visit http://www.sun.com if you need
19	* additional information or have any questions.
20	*/
21
22
23	/*******************************************************************************
24	* Header Files *
25	*******************************************************************************/
26	#define LOG_GROUP LOG_GROUP_HWACCM
27	#include <VBox/hwaccm.h>
28	#include "HWACCMInternal.h"
29	#include <VBox/vm.h>
30	#include <VBox/x86.h>
31	#include <VBox/pgm.h>
32	#include <VBox/pdm.h>
33	#include <VBox/err.h>
34	#include <VBox/log.h>
35	#include <VBox/selm.h>
36	#include <VBox/iom.h>
37	#include <iprt/param.h>
38	#include <iprt/assert.h>
39	#include <iprt/asm.h>
40	#include <iprt/string.h>
41	#include "HWVMXR0.h"
42
43
44	/* IO operation lookup arrays. */
45	static uint32_t aIOSize[4] = {1, 2, 0, 4};
46	static uint32_t aIOOpAnd[4] = {0xff, 0xffff, 0, 0xffffffff};
47
48
49	static void VMXR0CheckError(PVM pVM, int rc)
50	{
51	if (rc == VERR_VMX_GENERIC)
52	{
53	RTCCUINTREG instrError;
54
55	VMXReadVMCS(VMX_VMCS_RO_VM_INSTR_ERROR, &instrError);
56	pVM->hwaccm.s.vmx.ulLastInstrError = instrError;
57	}
58	pVM->hwaccm.s.lLastError = rc;
59	}
60
61	/**
62	* Sets up and activates VT-x on the current CPU
63	*
64	* @returns VBox status code.
65	* @param pCpu CPU info struct
66	* @param pVM The VM to operate on.
67	* @param pvPageCpu Pointer to the global cpu page
68	* @param pPageCpuPhys Physical address of the global cpu page
69	*/
70	HWACCMR0DECL(int) VMXR0EnableCpu(PHWACCM_CPUINFO pCpu, PVM pVM, void *pvPageCpu, RTHCPHYS pPageCpuPhys)
71	{
72	AssertReturn(pPageCpuPhys, VERR_INVALID_PARAMETER);
73	AssertReturn(pVM, VERR_INVALID_PARAMETER);
74	AssertReturn(pvPageCpu, VERR_INVALID_PARAMETER);
75
76	/* Setup Intel VMX. */
77	Assert(pVM->hwaccm.s.vmx.fSupported);
78
79	#ifdef LOG_ENABLED
80	SUPR0Printf("VMXR0EnableCpu cpu %d page (%x) %x\n", pCpu->idCpu, pvPageCpu, (uint32_t)pPageCpuPhys);
81	#endif
82	/* Set revision dword at the beginning of the VMXON structure. */
83	(uint32_t )pvPageCpu = MSR_IA32_VMX_BASIC_INFO_VMCS_ID(pVM->hwaccm.s.vmx.msr.vmx_basic_info);
84
85	/* @todo we should unmap the two pages from the virtual address space in order to prevent accidental corruption.
86	* (which can have very bad consequences!!!)
87	*/
88
89	/* Make sure the VMX instructions don't cause #UD faults. */
90	ASMSetCR4(ASMGetCR4() \| X86_CR4_VMXE);
91
92	/* Enter VMX Root Mode */
93	int rc = VMXEnable(pPageCpuPhys);
94	if (VBOX_FAILURE(rc))
95	{
96	VMXR0CheckError(pVM, rc);
97	ASMSetCR4(ASMGetCR4() & ~X86_CR4_VMXE);
98	return VERR_VMX_VMXON_FAILED;
99	}
100	return VINF_SUCCESS;
101	}
102
103	/**
104	* Deactivates VT-x on the current CPU
105	*
106	* @returns VBox status code.
107	* @param pCpu CPU info struct
108	* @param pvPageCpu Pointer to the global cpu page
109	* @param pPageCpuPhys Physical address of the global cpu page
110	*/
111	HWACCMR0DECL(int) VMXR0DisableCpu(PHWACCM_CPUINFO pCpu, void *pvPageCpu, RTHCPHYS pPageCpuPhys)
112	{
113	AssertReturn(pPageCpuPhys, VERR_INVALID_PARAMETER);
114	AssertReturn(pvPageCpu, VERR_INVALID_PARAMETER);
115
116	/* Leave VMX Root Mode. */
117	VMXDisable();
118
119	/* And clear the X86_CR4_VMXE bit */
120	ASMSetCR4(ASMGetCR4() & ~X86_CR4_VMXE);
121
122	#ifdef LOG_ENABLED
123	SUPR0Printf("VMXR0DisableCpu cpu %d\n", pCpu->idCpu);
124	#endif
125	return VINF_SUCCESS;
126	}
127
128	/**
129	* Does Ring-0 per VM VT-x init.
130	*
131	* @returns VBox status code.
132	* @param pVM The VM to operate on.
133	*/
134	HWACCMR0DECL(int) VMXR0InitVM(PVM pVM)
135	{
136	int rc;
137
138	#ifdef LOG_ENABLED
139	SUPR0Printf("VMXR0InitVM %x\n", pVM);
140	#endif
141
142	/* Allocate one page for the VM control structure (VMCS). */
143	rc = RTR0MemObjAllocCont(&pVM->hwaccm.s.vmx.pMemObjVMCS, 1 << PAGE_SHIFT, true /* executable R0 mapping */);
144	AssertRC(rc);
145	if (RT_FAILURE(rc))
146	return rc;
147
148	pVM->hwaccm.s.vmx.pVMCS = RTR0MemObjAddress(pVM->hwaccm.s.vmx.pMemObjVMCS);
149	pVM->hwaccm.s.vmx.pVMCSPhys = RTR0MemObjGetPagePhysAddr(pVM->hwaccm.s.vmx.pMemObjVMCS, 0);
150	ASMMemZero32(pVM->hwaccm.s.vmx.pVMCS, PAGE_SIZE);
151
152	/* Allocate one page for the TSS we need for real mode emulation. */
153	rc = RTR0MemObjAllocCont(&pVM->hwaccm.s.vmx.pMemObjRealModeTSS, 1 << PAGE_SHIFT, true /* executable R0 mapping */);
154	AssertRC(rc);
155	if (RT_FAILURE(rc))
156	return rc;
157
158	pVM->hwaccm.s.vmx.pRealModeTSS = (PVBOXTSS)RTR0MemObjAddress(pVM->hwaccm.s.vmx.pMemObjRealModeTSS);
159	pVM->hwaccm.s.vmx.pRealModeTSSPhys = RTR0MemObjGetPagePhysAddr(pVM->hwaccm.s.vmx.pMemObjRealModeTSS, 0);
160
161	/* The I/O bitmap starts right after the virtual interrupt redirection bitmap. Outside the TSS on purpose; the CPU will not check it
162	* for I/O operations. */
163	ASMMemZero32(pVM->hwaccm.s.vmx.pRealModeTSS, PAGE_SIZE);
164	pVM->hwaccm.s.vmx.pRealModeTSS->offIoBitmap = sizeof(*pVM->hwaccm.s.vmx.pRealModeTSS);
165	/* Bit set to 0 means redirection enabled. */
166	memset(pVM->hwaccm.s.vmx.pRealModeTSS->IntRedirBitmap, 0x0, sizeof(pVM->hwaccm.s.vmx.pRealModeTSS->IntRedirBitmap));
167
168	#ifdef LOG_ENABLED
169	SUPR0Printf("VMXR0InitVM %x VMCS=%x (%x) RealModeTSS=%x (%x)\n", pVM, pVM->hwaccm.s.vmx.pVMCS, (uint32_t)pVM->hwaccm.s.vmx.pVMCSPhys, pVM->hwaccm.s.vmx.pRealModeTSS, (uint32_t)pVM->hwaccm.s.vmx.pRealModeTSSPhys);
170	#endif
171	return VINF_SUCCESS;
172	}
173
174	/**
175	* Does Ring-0 per VM VT-x termination.
176	*
177	* @returns VBox status code.
178	* @param pVM The VM to operate on.
179	*/
180	HWACCMR0DECL(int) VMXR0TermVM(PVM pVM)
181	{
182	if (pVM->hwaccm.s.vmx.pMemObjVMCS)
183	{
184	RTR0MemObjFree(pVM->hwaccm.s.vmx.pMemObjVMCS, false);
185	pVM->hwaccm.s.vmx.pMemObjVMCS = 0;
186	pVM->hwaccm.s.vmx.pVMCS = 0;
187	pVM->hwaccm.s.vmx.pVMCSPhys = 0;
188	}
189	if (pVM->hwaccm.s.vmx.pMemObjRealModeTSS)
190	{
191	RTR0MemObjFree(pVM->hwaccm.s.vmx.pMemObjRealModeTSS, false);
192	pVM->hwaccm.s.vmx.pMemObjRealModeTSS = 0;
193	pVM->hwaccm.s.vmx.pRealModeTSS = 0;
194	pVM->hwaccm.s.vmx.pRealModeTSSPhys = 0;
195	}
196	return VINF_SUCCESS;
197	}
198
199	/**
200	* Sets up VT-x for the specified VM
201	*
202	* @returns VBox status code.
203	* @param pVM The VM to operate on.
204	*/
205	HWACCMR0DECL(int) VMXR0SetupVM(PVM pVM)
206	{
207	int rc = VINF_SUCCESS;
208	uint32_t val;
209
210	AssertReturn(pVM, VERR_INVALID_PARAMETER);
211	Assert(pVM->hwaccm.s.vmx.pVMCS);
212
213	/* Set revision dword at the beginning of the VMCS structure. */
214	(uint32_t )pVM->hwaccm.s.vmx.pVMCS = MSR_IA32_VMX_BASIC_INFO_VMCS_ID(pVM->hwaccm.s.vmx.msr.vmx_basic_info);
215
216	/* Clear VM Control Structure. */
217	Log(("pVMCSPhys = %VHp\n", pVM->hwaccm.s.vmx.pVMCSPhys));
218	rc = VMXClearVMCS(pVM->hwaccm.s.vmx.pVMCSPhys);
219	if (VBOX_FAILURE(rc))
220	goto vmx_end;
221
222	/* Activate the VM Control Structure. */
223	rc = VMXActivateVMCS(pVM->hwaccm.s.vmx.pVMCSPhys);
224	if (VBOX_FAILURE(rc))
225	goto vmx_end;
226
227	/* VMX_VMCS_CTRL_PIN_EXEC_CONTROLS
228	* Set required bits to one and zero according to the MSR capabilities.
229	*/
230	val = (pVM->hwaccm.s.vmx.msr.vmx_pin_ctls & 0xFFFFFFFF);
231	/* External and non-maskable interrupts cause VM-exits. */
232	val = val \| VMX_VMCS_CTRL_PIN_EXEC_CONTROLS_EXT_INT_EXIT \| VMX_VMCS_CTRL_PIN_EXEC_CONTROLS_NMI_EXIT;
233	val &= (pVM->hwaccm.s.vmx.msr.vmx_pin_ctls >> 32ULL);
234
235	rc = VMXWriteVMCS(VMX_VMCS_CTRL_PIN_EXEC_CONTROLS, val);
236	AssertRC(rc);
237
238	/* VMX_VMCS_CTRL_PROC_EXEC_CONTROLS
239	* Set required bits to one and zero according to the MSR capabilities.
240	*/
241	val = (pVM->hwaccm.s.vmx.msr.vmx_proc_ctls & 0xFFFFFFFF);
242	/* Program which event cause VM-exits and which features we want to use. */
243	val = val \| VMX_VMCS_CTRL_PROC_EXEC_CONTROLS_HLT_EXIT
244	\| VMX_VMCS_CTRL_PROC_EXEC_CONTROLS_TSC_OFFSET
245	\| VMX_VMCS_CTRL_PROC_EXEC_CONTROLS_INVLPG_EXIT
246	\| VMX_VMCS_CTRL_PROC_EXEC_CONTROLS_MOV_DR_EXIT
247	\| VMX_VMCS_CTRL_PROC_EXEC_CONTROLS_UNCOND_IO_EXIT
248	\| VMX_VMCS_CTRL_PROC_EXEC_CONTROLS_MWAIT_EXIT; /* don't execute mwait or else we'll idle inside the guest (host thinks the cpu load is high) */
249
250	/** @note VMX_VMCS_CTRL_PROC_EXEC_CONTROLS_MWAIT_EXIT might cause a vmlaunch failure with an invalid control fields error. (combined with some other exit reasons) */
251
252	/*
253	if AMD64 guest mode
254	val \|= VMX_VMCS_CTRL_PROC_EXEC_CONTROLS_CR8_LOAD_EXIT
255	\| VMX_VMCS_CTRL_PROC_EXEC_CONTROLS_CR8_STORE_EXIT;
256	*/
257	#if HC_ARCH_BITS == 64
258	val \|= VMX_VMCS_CTRL_PROC_EXEC_CONTROLS_CR8_LOAD_EXIT
259	\| VMX_VMCS_CTRL_PROC_EXEC_CONTROLS_CR8_STORE_EXIT;
260	#endif
261	/* Mask away the bits that the CPU doesn't support */
262	/** @todo make sure they don't conflict with the above requirements. */
263	val &= (pVM->hwaccm.s.vmx.msr.vmx_proc_ctls >> 32ULL);
264	pVM->hwaccm.s.vmx.proc_ctls = val;
265
266	rc = VMXWriteVMCS(VMX_VMCS_CTRL_PROC_EXEC_CONTROLS, val);
267	AssertRC(rc);
268
269	/* VMX_VMCS_CTRL_CR3_TARGET_COUNT
270	* Set required bits to one and zero according to the MSR capabilities.
271	*/
272	rc = VMXWriteVMCS(VMX_VMCS_CTRL_CR3_TARGET_COUNT, 0);
273	AssertRC(rc);
274
275	/* VMX_VMCS_CTRL_ENTRY_CONTROLS
276	* Set required bits to one and zero according to the MSR capabilities.
277	*/
278	val = (pVM->hwaccm.s.vmx.msr.vmx_entry & 0xFFFFFFFF);
279	if (pVM->hwaccm.s.cpuid.u32AMDFeatureEDX & X86_CPUID_AMD_FEATURE_EDX_LONG_MODE)
280	{
281	/** @todo 32 bits guest mode only for now. */
282	/* val \|= VMX_VMCS_CTRL_ENTRY_CONTROLS_IA64_MODE; */
283	}
284	/* Mask away the bits that the CPU doesn't support */
285	/** @todo make sure they don't conflict with the above requirements. */
286	val &= (pVM->hwaccm.s.vmx.msr.vmx_entry >> 32ULL);
287	/* else Must be zero when AMD64 is not available. */
288	rc = VMXWriteVMCS(VMX_VMCS_CTRL_ENTRY_CONTROLS, val);
289	AssertRC(rc);
290
291	/* VMX_VMCS_CTRL_EXIT_CONTROLS
292	* Set required bits to one and zero according to the MSR capabilities.
293	*/
294	val = (pVM->hwaccm.s.vmx.msr.vmx_exit & 0xFFFFFFFF);
295	#if HC_ARCH_BITS == 64
296	val \|= VMX_VMCS_CTRL_EXIT_CONTROLS_HOST_AMD64;
297	#else
298	/* else Must be zero when AMD64 is not available. */
299	#endif
300	val &= (pVM->hwaccm.s.vmx.msr.vmx_exit >> 32ULL);
301	/* Don't acknowledge external interrupts on VM-exit. */
302	rc = VMXWriteVMCS(VMX_VMCS_CTRL_EXIT_CONTROLS, val);
303	AssertRC(rc);
304
305	/* Forward all exception except #NM & #PF to the guest.
306	* We always need to check pagefaults since our shadow page table can be out of sync.
307	* And we always lazily sync the FPU & XMM state.
308	*/
309
310	/*
311	* @todo Possible optimization:
312	* Keep the FPU and XMM state current in the EM thread. That way there's no need to
313	* lazily sync anything, but the downside is that we can't use the FPU stack or XMM
314	* registers ourselves of course.
315	*
316	* @note only possible if the current state is actually ours (X86_CR0_TS flag)
317	*/
318	rc = VMXWriteVMCS(VMX_VMCS_CTRL_EXCEPTION_BITMAP, HWACCM_VMX_TRAP_MASK);
319	AssertRC(rc);
320
321	/* Don't filter page faults; all of them should cause a switch. */
322	rc = VMXWriteVMCS(VMX_VMCS_CTRL_PAGEFAULT_ERROR_MASK, 0);
323	rc \|= VMXWriteVMCS(VMX_VMCS_CTRL_PAGEFAULT_ERROR_MATCH, 0);
324	AssertRC(rc);
325
326	/* Init TSC offset to zero. */
327	rc = VMXWriteVMCS(VMX_VMCS_CTRL_TSC_OFFSET_FULL, 0);
328	#if HC_ARCH_BITS == 32
329	rc \|= VMXWriteVMCS(VMX_VMCS_CTRL_TSC_OFFSET_HIGH, 0);
330	#endif
331	AssertRC(rc);
332
333	rc = VMXWriteVMCS(VMX_VMCS_CTRL_IO_BITMAP_A_FULL, 0);
334	#if HC_ARCH_BITS == 32
335	rc \|= VMXWriteVMCS(VMX_VMCS_CTRL_IO_BITMAP_A_HIGH, 0);
336	#endif
337	AssertRC(rc);
338
339	rc = VMXWriteVMCS(VMX_VMCS_CTRL_IO_BITMAP_B_FULL, 0);
340	#if HC_ARCH_BITS == 32
341	rc \|= VMXWriteVMCS(VMX_VMCS_CTRL_IO_BITMAP_B_HIGH, 0);
342	#endif
343	AssertRC(rc);
344
345	/* Clear MSR controls. */
346	if (pVM->hwaccm.s.vmx.msr.vmx_proc_ctls & VMX_VMCS_CTRL_PROC_EXEC_CONTROLS_USE_MSR_BITMAPS)
347	{
348	/* Optional */
349	rc = VMXWriteVMCS(VMX_VMCS_CTRL_MSR_BITMAP_FULL, 0);
350	#if HC_ARCH_BITS == 32
351	rc \|= VMXWriteVMCS(VMX_VMCS_CTRL_MSR_BITMAP_HIGH, 0);
352	#endif
353	AssertRC(rc);
354	}
355	rc = VMXWriteVMCS(VMX_VMCS_CTRL_VMEXIT_MSR_STORE_FULL, 0);
356	rc \|= VMXWriteVMCS(VMX_VMCS_CTRL_VMEXIT_MSR_LOAD_FULL, 0);
357	rc \|= VMXWriteVMCS(VMX_VMCS_CTRL_VMENTRY_MSR_LOAD_FULL, 0);
358	#if HC_ARCH_BITS == 32
359	rc \|= VMXWriteVMCS(VMX_VMCS_CTRL_VMEXIT_MSR_STORE_HIGH, 0);
360	rc \|= VMXWriteVMCS(VMX_VMCS_CTRL_VMEXIT_MSR_LOAD_HIGH, 0);
361	rc \|= VMXWriteVMCS(VMX_VMCS_CTRL_VMEXIT_MSR_LOAD_HIGH, 0);
362	#endif
363	rc \|= VMXWriteVMCS(VMX_VMCS_CTRL_EXIT_MSR_STORE_COUNT, 0);
364	rc \|= VMXWriteVMCS(VMX_VMCS_CTRL_EXIT_MSR_LOAD_COUNT, 0);
365	AssertRC(rc);
366
367	if (pVM->hwaccm.s.vmx.msr.vmx_proc_ctls & VMX_VMCS_CTRL_PROC_EXEC_CONTROLS_USE_TPR_SHADOW)
368	{
369	/* Optional */
370	rc = VMXWriteVMCS(VMX_VMCS_CTRL_TPR_TRESHOLD, 0);
371	rc \|= VMXWriteVMCS(VMX_VMCS_CTRL_VAPIC_PAGEADDR_FULL, 0);
372	#if HC_ARCH_BITS == 32
373	rc \|= VMXWriteVMCS(VMX_VMCS_CTRL_VAPIC_PAGEADDR_HIGH, 0);
374	#endif
375	AssertRC(rc);
376	}
377
378	/* Set link pointer to -1. Not currently used. */
379	#if HC_ARCH_BITS == 32
380	rc = VMXWriteVMCS(VMX_VMCS_GUEST_LINK_PTR_FULL, 0xFFFFFFFF);
381	rc \|= VMXWriteVMCS(VMX_VMCS_GUEST_LINK_PTR_HIGH, 0xFFFFFFFF);
382	#else
383	rc = VMXWriteVMCS(VMX_VMCS_GUEST_LINK_PTR_FULL, 0xFFFFFFFFFFFFFFFF);
384	#endif
385	AssertRC(rc);
386
387	/* Clear VM Control Structure. Marking it inactive, clearing implementation specific data and writing back VMCS data to memory. */
388	rc = VMXClearVMCS(pVM->hwaccm.s.vmx.pVMCSPhys);
389	AssertRC(rc);
390
391	vmx_end:
392	VMXR0CheckError(pVM, rc);
393	return rc;
394	}
395
396
397	/**
398	* Injects an event (trap or external interrupt)
399	*
400	* @returns VBox status code.
401	* @param pVM The VM to operate on.
402	* @param pCtx CPU Context
403	* @param intInfo VMX interrupt info
404	* @param cbInstr Opcode length of faulting instruction
405	* @param errCode Error code (optional)
406	*/
407	static int VMXR0InjectEvent(PVM pVM, CPUMCTX *pCtx, uint32_t intInfo, uint32_t cbInstr, uint32_t errCode)
408	{
409	int rc;
410
411	#ifdef VBOX_STRICT
412	uint32_t iGate = VMX_EXIT_INTERRUPTION_INFO_VECTOR(intInfo);
413	if (iGate == 0xE)
414	Log2(("VMXR0InjectEvent: Injecting interrupt %d at %VGv error code=%08x CR2=%08x intInfo=%08x\n", iGate, pCtx->eip, errCode, pCtx->cr2, intInfo));
415	else
416	if (iGate < 0x20)
417	Log2(("VMXR0InjectEvent: Injecting interrupt %d at %VGv error code=%08x\n", iGate, pCtx->eip, errCode));
418	else
419	{
420	Log2(("INJ-EI: %x at %VGv\n", iGate, pCtx->eip));
421	Assert(!VM_FF_ISSET(pVM, VM_FF_INHIBIT_INTERRUPTS));
422	Assert(pCtx->eflags.u32 & X86_EFL_IF);
423	}
424	#endif
425
426	/* Set event injection state. */
427	rc = VMXWriteVMCS(VMX_VMCS_CTRL_ENTRY_IRQ_INFO,
428	intInfo \| (1 << VMX_EXIT_INTERRUPTION_INFO_VALID_SHIFT)
429	);
430
431	rc \|= VMXWriteVMCS(VMX_VMCS_CTRL_ENTRY_INSTR_LENGTH, cbInstr);
432	rc \|= VMXWriteVMCS(VMX_VMCS_CTRL_ENTRY_EXCEPTION_ERRCODE, errCode);
433
434	AssertRC(rc);
435	return rc;
436	}
437
438
439	/**
440	* Checks for pending guest interrupts and injects them
441	*
442	* @returns VBox status code.
443	* @param pVM The VM to operate on.
444	* @param pCtx CPU Context
445	*/
446	static int VMXR0CheckPendingInterrupt(PVM pVM, CPUMCTX *pCtx)
447	{
448	int rc;
449
450	/* Dispatch any pending interrupts. (injected before, but a VM exit occurred prematurely) */
451	if (pVM->hwaccm.s.Event.fPending)
452	{
453	Log(("Reinjecting event %VX64 %08x at %VGv\n", pVM->hwaccm.s.Event.intInfo, pVM->hwaccm.s.Event.errCode, pCtx->eip));
454	STAM_COUNTER_INC(&pVM->hwaccm.s.StatIntReinject);
455	rc = VMXR0InjectEvent(pVM, pCtx, pVM->hwaccm.s.Event.intInfo, 0, pVM->hwaccm.s.Event.errCode);
456	AssertRC(rc);
457
458	pVM->hwaccm.s.Event.fPending = false;
459	return VINF_SUCCESS;
460	}
461
462	/* When external interrupts are pending, we should exit the VM when IF is set. */
463	if ( !TRPMHasTrap(pVM)
464	&& VM_FF_ISPENDING(pVM, (VM_FF_INTERRUPT_APIC\|VM_FF_INTERRUPT_PIC)))
465	{
466	if (!(pCtx->eflags.u32 & X86_EFL_IF))
467	{
468	Log2(("Enable irq window exit!\n"));
469	pVM->hwaccm.s.vmx.proc_ctls \|= VMX_VMCS_CTRL_PROC_EXEC_CONTROLS_IRQ_WINDOW_EXIT;
470	rc = VMXWriteVMCS(VMX_VMCS_CTRL_PROC_EXEC_CONTROLS, pVM->hwaccm.s.vmx.proc_ctls);
471	AssertRC(rc);
472	}
473	else
474	if (!VM_FF_ISSET(pVM, VM_FF_INHIBIT_INTERRUPTS))
475	{
476	uint8_t u8Interrupt;
477
478	rc = PDMGetInterrupt(pVM, &u8Interrupt);
479	Log(("Dispatch interrupt: u8Interrupt=%x (%d) rc=%Vrc\n", u8Interrupt, u8Interrupt, rc));
480	if (VBOX_SUCCESS(rc))
481	{
482	rc = TRPMAssertTrap(pVM, u8Interrupt, TRPM_HARDWARE_INT);
483	AssertRC(rc);
484	}
485	else
486	{
487	/* Can only happen in rare cases where a pending interrupt is cleared behind our back */
488	Assert(!VM_FF_ISPENDING(pVM, (VM_FF_INTERRUPT_APIC\|VM_FF_INTERRUPT_PIC)));
489	STAM_COUNTER_INC(&pVM->hwaccm.s.StatSwitchGuestIrq);
490	/* Just continue */
491	}
492	}
493	else
494	Log(("Pending interrupt blocked at %VGv by VM_FF_INHIBIT_INTERRUPTS!!\n", pCtx->eip));
495	}
496
497	#ifdef VBOX_STRICT
498	if (TRPMHasTrap(pVM))
499	{
500	uint8_t u8Vector;
501	rc = TRPMQueryTrapAll(pVM, &u8Vector, 0, 0, 0);
502	AssertRC(rc);
503	}
504	#endif
505
506	if ( pCtx->eflags.u32 & X86_EFL_IF
507	&& (!VM_FF_ISSET(pVM, VM_FF_INHIBIT_INTERRUPTS))
508	&& TRPMHasTrap(pVM)
509	)
510	{
511	uint8_t u8Vector;
512	int rc;
513	TRPMEVENT enmType;
514	RTGCUINTPTR intInfo, errCode;
515
516	/* If a new event is pending, then dispatch it now. */
517	rc = TRPMQueryTrapAll(pVM, &u8Vector, &enmType, &errCode, 0);
518	AssertRC(rc);
519	Assert(pCtx->eflags.Bits.u1IF == 1 \|\| enmType == TRPM_TRAP);
520	Assert(enmType != TRPM_SOFTWARE_INT);
521
522	/* Clear the pending trap. */
523	rc = TRPMResetTrap(pVM);
524	AssertRC(rc);
525
526	intInfo = u8Vector;
527	intInfo \|= (1 << VMX_EXIT_INTERRUPTION_INFO_VALID_SHIFT);
528
529	if (enmType == TRPM_TRAP)
530	{
531	switch (u8Vector) {
532	case 8:
533	case 10:
534	case 11:
535	case 12:
536	case 13:
537	case 14:
538	case 17:
539	/* Valid error codes. */
540	intInfo \|= VMX_EXIT_INTERRUPTION_INFO_ERROR_CODE_VALID;
541	break;
542	default:
543	break;
544	}
545	if (u8Vector == X86_XCPT_BP \|\| u8Vector == X86_XCPT_OF)
546	intInfo \|= (VMX_EXIT_INTERRUPTION_INFO_TYPE_SWEXCPT << VMX_EXIT_INTERRUPTION_INFO_TYPE_SHIFT);
547	else
548	intInfo \|= (VMX_EXIT_INTERRUPTION_INFO_TYPE_HWEXCPT << VMX_EXIT_INTERRUPTION_INFO_TYPE_SHIFT);
549	}
550	else
551	intInfo \|= (VMX_EXIT_INTERRUPTION_INFO_TYPE_EXT << VMX_EXIT_INTERRUPTION_INFO_TYPE_SHIFT);
552
553	STAM_COUNTER_INC(&pVM->hwaccm.s.StatIntInject);
554	rc = VMXR0InjectEvent(pVM, pCtx, intInfo, 0, errCode);
555	AssertRC(rc);
556	} /* if (interrupts can be dispatched) */
557
558	return VINF_SUCCESS;
559	}
560
561	/**
562	* Save the host state
563	*
564	* @returns VBox status code.
565	* @param pVM The VM to operate on.
566	*/
567	HWACCMR0DECL(int) VMXR0SaveHostState(PVM pVM)
568	{
569	int rc = VINF_SUCCESS;
570
571	/*
572	* Host CPU Context
573	*/
574	if (pVM->hwaccm.s.fContextUseFlags & HWACCM_CHANGED_HOST_CONTEXT)
575	{
576	RTIDTR idtr;
577	RTGDTR gdtr;
578	RTSEL SelTR;
579	PX86DESCHC pDesc;
580	uintptr_t trBase;
581
582	/* Control registers */
583	rc = VMXWriteVMCS(VMX_VMCS_HOST_CR0, ASMGetCR0());
584	rc \|= VMXWriteVMCS(VMX_VMCS_HOST_CR3, ASMGetCR3());
585	rc \|= VMXWriteVMCS(VMX_VMCS_HOST_CR4, ASMGetCR4());
586	AssertRC(rc);
587	Log2(("VMX_VMCS_HOST_CR0 %08x\n", ASMGetCR0()));
588	Log2(("VMX_VMCS_HOST_CR3 %VHp\n", ASMGetCR3()));
589	Log2(("VMX_VMCS_HOST_CR4 %08x\n", ASMGetCR4()));
590
591	/* Selector registers. */
592	rc = VMXWriteVMCS(VMX_VMCS_HOST_FIELD_CS, ASMGetCS());
593	/** @note VMX is (again) very picky about the RPL of the selectors here; we'll restore them manually. */
594	rc \|= VMXWriteVMCS(VMX_VMCS_HOST_FIELD_DS, 0);
595	rc \|= VMXWriteVMCS(VMX_VMCS_HOST_FIELD_ES, 0);
596	#if HC_ARCH_BITS == 32
597	rc \|= VMXWriteVMCS(VMX_VMCS_HOST_FIELD_FS, 0);
598	rc \|= VMXWriteVMCS(VMX_VMCS_HOST_FIELD_GS, 0);
599	#endif
600	rc \|= VMXWriteVMCS(VMX_VMCS_HOST_FIELD_SS, ASMGetSS());
601	SelTR = ASMGetTR();
602	rc \|= VMXWriteVMCS(VMX_VMCS_HOST_FIELD_TR, SelTR);
603	AssertRC(rc);
604	Log2(("VMX_VMCS_HOST_FIELD_CS %08x\n", ASMGetCS()));
605	Log2(("VMX_VMCS_HOST_FIELD_DS %08x\n", ASMGetDS()));
606	Log2(("VMX_VMCS_HOST_FIELD_ES %08x\n", ASMGetES()));
607	Log2(("VMX_VMCS_HOST_FIELD_FS %08x\n", ASMGetFS()));
608	Log2(("VMX_VMCS_HOST_FIELD_GS %08x\n", ASMGetGS()));
609	Log2(("VMX_VMCS_HOST_FIELD_SS %08x\n", ASMGetSS()));
610	Log2(("VMX_VMCS_HOST_FIELD_TR %08x\n", ASMGetTR()));
611
612	/* GDTR & IDTR */
613	ASMGetGDTR(&gdtr);
614	rc = VMXWriteVMCS(VMX_VMCS_HOST_GDTR_BASE, gdtr.pGdt);
615	ASMGetIDTR(&idtr);
616	rc \|= VMXWriteVMCS(VMX_VMCS_HOST_IDTR_BASE, idtr.pIdt);
617	AssertRC(rc);
618	Log2(("VMX_VMCS_HOST_GDTR_BASE %VHv\n", gdtr.pGdt));
619	Log2(("VMX_VMCS_HOST_IDTR_BASE %VHv\n", idtr.pIdt));
620
621	/* Save the base address of the TR selector. */
622	if (SelTR > gdtr.cbGdt)
623	{
624	AssertMsgFailed(("Invalid TR selector %x. GDTR.cbGdt=%x\n", SelTR, gdtr.cbGdt));
625	return VERR_VMX_INVALID_HOST_STATE;
626	}
627
628	pDesc = &((PX86DESCHC)gdtr.pGdt)[SelTR >> X86_SEL_SHIFT_HC];
629	#if HC_ARCH_BITS == 64
630	trBase = pDesc->Gen.u16BaseLow \| (pDesc->Gen.u8BaseHigh1 << 16ULL) \| (pDesc->Gen.u8BaseHigh2 << 24ULL) \| ((uintptr_t)pDesc->Gen.u32BaseHigh3 << 32ULL);
631	#else
632	trBase = pDesc->Gen.u16BaseLow \| (pDesc->Gen.u8BaseHigh1 << 16) \| (pDesc->Gen.u8BaseHigh2 << 24);
633	#endif
634	rc = VMXWriteVMCS(VMX_VMCS_HOST_TR_BASE, trBase);
635	AssertRC(rc);
636	Log2(("VMX_VMCS_HOST_TR_BASE %VHv\n", trBase));
637
638	/* FS and GS base. */
639	#if HC_ARCH_BITS == 64
640	Log2(("MSR_K8_FS_BASE = %VHv\n", ASMRdMsr(MSR_K8_FS_BASE)));
641	Log2(("MSR_K8_GS_BASE = %VHv\n", ASMRdMsr(MSR_K8_GS_BASE)));
642	rc = VMXWriteVMCS64(VMX_VMCS_HOST_FS_BASE, ASMRdMsr(MSR_K8_FS_BASE));
643	rc \|= VMXWriteVMCS64(VMX_VMCS_HOST_GS_BASE, ASMRdMsr(MSR_K8_GS_BASE));
644	#endif
645	AssertRC(rc);
646
647	/* Sysenter MSRs. */
648	/** @todo expensive!! */
649	rc = VMXWriteVMCS(VMX_VMCS_HOST_SYSENTER_CS, ASMRdMsr_Low(MSR_IA32_SYSENTER_CS));
650	Log2(("VMX_VMCS_HOST_SYSENTER_CS %08x\n", ASMRdMsr_Low(MSR_IA32_SYSENTER_CS)));
651	#if HC_ARCH_BITS == 32
652	rc \|= VMXWriteVMCS(VMX_VMCS_HOST_SYSENTER_ESP, ASMRdMsr_Low(MSR_IA32_SYSENTER_ESP));
653	rc \|= VMXWriteVMCS(VMX_VMCS_HOST_SYSENTER_EIP, ASMRdMsr_Low(MSR_IA32_SYSENTER_EIP));
654	Log2(("VMX_VMCS_HOST_SYSENTER_EIP %VHv\n", ASMRdMsr_Low(MSR_IA32_SYSENTER_EIP)));
655	Log2(("VMX_VMCS_HOST_SYSENTER_ESP %VHv\n", ASMRdMsr_Low(MSR_IA32_SYSENTER_ESP)));
656	#else
657	Log2(("VMX_VMCS_HOST_SYSENTER_EIP %VHv\n", ASMRdMsr(MSR_IA32_SYSENTER_EIP)));
658	Log2(("VMX_VMCS_HOST_SYSENTER_ESP %VHv\n", ASMRdMsr(MSR_IA32_SYSENTER_ESP)));
659	rc \|= VMXWriteVMCS64(VMX_VMCS_HOST_SYSENTER_ESP, ASMRdMsr(MSR_IA32_SYSENTER_ESP));
660	rc \|= VMXWriteVMCS64(VMX_VMCS_HOST_SYSENTER_EIP, ASMRdMsr(MSR_IA32_SYSENTER_EIP));
661	#endif
662	AssertRC(rc);
663
664	pVM->hwaccm.s.fContextUseFlags &= ~HWACCM_CHANGED_HOST_CONTEXT;
665	}
666	return rc;
667	}
668
669
670	/**
671	* Loads the guest state
672	*
673	* @returns VBox status code.
674	* @param pVM The VM to operate on.
675	* @param pCtx Guest context
676	*/
677	HWACCMR0DECL(int) VMXR0LoadGuestState(PVM pVM, CPUMCTX *pCtx)
678	{
679	int rc = VINF_SUCCESS;
680	RTGCUINTPTR val;
681	X86EFLAGS eflags;
682
683	/* Guest CPU context: ES, CS, SS, DS, FS, GS. */
684	if (pVM->hwaccm.s.fContextUseFlags & HWACCM_CHANGED_GUEST_SEGMENT_REGS)
685	{
686	VMX_WRITE_SELREG(ES, es);
687	AssertRC(rc);
688
689	VMX_WRITE_SELREG(CS, cs);
690	AssertRC(rc);
691
692	VMX_WRITE_SELREG(SS, ss);
693	AssertRC(rc);
694
695	VMX_WRITE_SELREG(DS, ds);
696	AssertRC(rc);
697
698	VMX_WRITE_SELREG(FS, fs);
699	AssertRC(rc);
700
701	VMX_WRITE_SELREG(GS, gs);
702	AssertRC(rc);
703	}
704
705	/* Guest CPU context: LDTR. */
706	if (pVM->hwaccm.s.fContextUseFlags & HWACCM_CHANGED_GUEST_LDTR)
707	{
708	if (pCtx->ldtr == 0)
709	{
710	rc = VMXWriteVMCS(VMX_VMCS_GUEST_FIELD_LDTR, 0);
711	rc \|= VMXWriteVMCS(VMX_VMCS_GUEST_LDTR_LIMIT, 0);
712	rc \|= VMXWriteVMCS(VMX_VMCS_GUEST_LDTR_BASE, 0);
713	/** @note vmlaunch will fail with 0 or just 0x02. No idea why. */
714	rc \|= VMXWriteVMCS(VMX_VMCS_GUEST_LDTR_ACCESS_RIGHTS, 0x82 /* present, LDT */);
715	}
716	else
717	{
718	rc = VMXWriteVMCS(VMX_VMCS_GUEST_FIELD_LDTR, pCtx->ldtr);
719	rc \|= VMXWriteVMCS(VMX_VMCS_GUEST_LDTR_LIMIT, pCtx->ldtrHid.u32Limit);
720	rc \|= VMXWriteVMCS(VMX_VMCS_GUEST_LDTR_BASE, pCtx->ldtrHid.u32Base);
721	rc \|= VMXWriteVMCS(VMX_VMCS_GUEST_LDTR_ACCESS_RIGHTS, pCtx->ldtrHid.Attr.u);
722	}
723	AssertRC(rc);
724	}
725	/* Guest CPU context: TR. */
726	if (pVM->hwaccm.s.fContextUseFlags & HWACCM_CHANGED_GUEST_TR)
727	{
728	rc = VMXWriteVMCS(VMX_VMCS_GUEST_FIELD_TR, pCtx->tr);
729
730	/* Real mode emulation using v86 mode with CR4.VME (interrupt redirection using the int bitmap in the TSS) */
731	if (!(pCtx->cr0 & X86_CR0_PROTECTION_ENABLE))
732	{
733	rc \|= VMXWriteVMCS(VMX_VMCS_GUEST_TR_LIMIT, sizeof(*pVM->hwaccm.s.vmx.pRealModeTSS));
734	rc \|= VMXWriteVMCS(VMX_VMCS_GUEST_TR_BASE, 0);
735	}
736	else
737	{
738	rc \|= VMXWriteVMCS(VMX_VMCS_GUEST_TR_LIMIT, pCtx->trHid.u32Limit);
739	rc \|= VMXWriteVMCS(VMX_VMCS_GUEST_TR_BASE, pCtx->trHid.u32Base);
740	}
741	val = pCtx->trHid.Attr.u;
742
743	/* The TSS selector must be busy. */
744	if ((val & 0xF) == X86_SEL_TYPE_SYS_286_TSS_AVAIL)
745	val = (val & ~0xF) \| X86_SEL_TYPE_SYS_286_TSS_BUSY;
746	else
747	/* Default even if no TR selector has been set (otherwise vmlaunch will fail!) */
748	val = (val & ~0xF) \| X86_SEL_TYPE_SYS_386_TSS_BUSY;
749
750	rc \|= VMXWriteVMCS(VMX_VMCS_GUEST_TR_ACCESS_RIGHTS, val);
751	AssertRC(rc);
752	}
753	/* Guest CPU context: GDTR. */
754	if (pVM->hwaccm.s.fContextUseFlags & HWACCM_CHANGED_GUEST_GDTR)
755	{
756	rc = VMXWriteVMCS(VMX_VMCS_GUEST_GDTR_LIMIT, pCtx->gdtr.cbGdt);
757	rc \|= VMXWriteVMCS(VMX_VMCS_GUEST_GDTR_BASE, pCtx->gdtr.pGdt);
758	AssertRC(rc);
759	}
760	/* Guest CPU context: IDTR. */
761	if (pVM->hwaccm.s.fContextUseFlags & HWACCM_CHANGED_GUEST_IDTR)
762	{
763	rc = VMXWriteVMCS(VMX_VMCS_GUEST_IDTR_LIMIT, pCtx->idtr.cbIdt);
764	rc \|= VMXWriteVMCS(VMX_VMCS_GUEST_IDTR_BASE, pCtx->idtr.pIdt);
765	AssertRC(rc);
766	}
767
768	/*
769	* Sysenter MSRs
770	*/
771	if (pVM->hwaccm.s.fContextUseFlags & HWACCM_CHANGED_GUEST_SYSENTER_MSR)
772	{
773	rc = VMXWriteVMCS(VMX_VMCS_GUEST_SYSENTER_CS, pCtx->SysEnter.cs);
774	rc \|= VMXWriteVMCS(VMX_VMCS_GUEST_SYSENTER_EIP, pCtx->SysEnter.eip);
775	rc \|= VMXWriteVMCS(VMX_VMCS_GUEST_SYSENTER_ESP, pCtx->SysEnter.esp);
776	AssertRC(rc);
777	}
778
779	/* Control registers */
780	if (pVM->hwaccm.s.fContextUseFlags & HWACCM_CHANGED_GUEST_CR0)
781	{
782	val = pCtx->cr0;
783	rc = VMXWriteVMCS(VMX_VMCS_CTRL_CR0_READ_SHADOW, val);
784	Log2(("Guest CR0-shadow %08x\n", val));
785	if (CPUMIsGuestFPUStateActive(pVM) == false)
786	{
787	/* Always use #NM exceptions to load the FPU/XMM state on demand. */
788	val \|= X86_CR0_TS \| X86_CR0_ET \| X86_CR0_NE \| X86_CR0_MP;
789	}
790	else
791	{
792	Assert(pVM->hwaccm.s.vmx.fResumeVM == true);
793	/** @todo check if we support the old style mess correctly. */
794	if (!(val & X86_CR0_NE))
795	{
796	Log(("Forcing X86_CR0_NE!!!\n"));
797
798	/* Also catch floating point exceptions as we need to report them to the guest in a different way. */
799	if (!pVM->hwaccm.s.fFPUOldStyleOverride)
800	{
801	rc = VMXWriteVMCS(VMX_VMCS_CTRL_EXCEPTION_BITMAP, HWACCM_VMX_TRAP_MASK \| RT_BIT(16));
802	AssertRC(rc);
803	pVM->hwaccm.s.fFPUOldStyleOverride = true;
804	}
805	}
806
807	val \|= X86_CR0_NE; /* always turn on the native mechanism to report FPU errors (old style uses interrupts) */
808	}
809	/* Note: protected mode & paging are always enabled; we use them for emulating real and protected mode without paging too. */
810	val \|= X86_CR0_PE \| X86_CR0_PG;
811	/* Note: We must also set this as we rely on protecting various pages for which supervisor writes must be caught. */
812	val \|= X86_CR0_WP;
813
814	rc \|= VMXWriteVMCS(VMX_VMCS_GUEST_CR0, val);
815	Log2(("Guest CR0 %08x\n", val));
816	/* CR0 flags owned by the host; if the guests attempts to change them, then
817	* the VM will exit.
818	*/
819	val = X86_CR0_PE /* Must monitor this bit (assumptions are made for real mode emulation) */
820	\| X86_CR0_WP /* Must monitor this bit (it must always be enabled). */
821	\| X86_CR0_PG /* Must monitor this bit (assumptions are made for real mode & protected mode without paging emulation) */
822	\| X86_CR0_TS
823	\| X86_CR0_ET
824	\| X86_CR0_NE
825	\| X86_CR0_MP;
826	pVM->hwaccm.s.vmx.cr0_mask = val;
827
828	rc \|= VMXWriteVMCS(VMX_VMCS_CTRL_CR0_MASK, val);
829	Log2(("Guest CR0-mask %08x\n", val));
830	AssertRC(rc);
831	}
832	if (pVM->hwaccm.s.fContextUseFlags & HWACCM_CHANGED_GUEST_CR4)
833	{
834	/* CR4 */
835	rc = VMXWriteVMCS(VMX_VMCS_CTRL_CR4_READ_SHADOW, pCtx->cr4);
836	Log2(("Guest CR4-shadow %08x\n", pCtx->cr4));
837	/* Set the required bits in cr4 too (currently X86_CR4_VMXE). */
838	val = pCtx->cr4 \| (uint32_t)pVM->hwaccm.s.vmx.msr.vmx_cr4_fixed0;
839	switch(pVM->hwaccm.s.enmShadowMode)
840	{
841	case PGMMODE_REAL: /* Real mode -> emulated using v86 mode */
842	case PGMMODE_PROTECTED: /* Protected mode, no paging -> emulated using identity mapping. */
843	case PGMMODE_32_BIT: /* 32-bit paging. */
844	break;
845
846	case PGMMODE_PAE: /* PAE paging. */
847	case PGMMODE_PAE_NX: /* PAE paging with NX enabled. */
848	/** @todo use normal 32 bits paging */
849	val \|= X86_CR4_PAE;
850	break;
851
852	case PGMMODE_AMD64: /* 64-bit AMD paging (long mode). */
853	case PGMMODE_AMD64_NX: /* 64-bit AMD paging (long mode) with NX enabled. */
854	AssertFailed();
855	return VERR_PGM_UNSUPPORTED_HOST_PAGING_MODE;
856
857	default: /* shut up gcc */
858	AssertFailed();
859	return VERR_PGM_UNSUPPORTED_HOST_PAGING_MODE;
860	}
861	/* Real mode emulation using v86 mode with CR4.VME (interrupt redirection using the int bitmap in the TSS) */
862	if (!(pCtx->cr0 & X86_CR0_PROTECTION_ENABLE))
863	val \|= X86_CR4_VME;
864
865	rc \|= VMXWriteVMCS(VMX_VMCS_GUEST_CR4, val);
866	Log2(("Guest CR4 %08x\n", val));
867	/* CR4 flags owned by the host; if the guests attempts to change them, then
868	* the VM will exit.
869	*/
870	val = X86_CR4_PAE
871	\| X86_CR4_PGE
872	\| X86_CR4_PSE
873	\| X86_CR4_VMXE;
874	pVM->hwaccm.s.vmx.cr4_mask = val;
875
876	rc \|= VMXWriteVMCS(VMX_VMCS_CTRL_CR4_MASK, val);
877	Log2(("Guest CR4-mask %08x\n", val));
878	AssertRC(rc);
879	}
880
881	if (pVM->hwaccm.s.fContextUseFlags & HWACCM_CHANGED_GUEST_CR3)
882	{
883	/* Save our shadow CR3 register. */
884	val = PGMGetHyperCR3(pVM);
885	rc = VMXWriteVMCS(VMX_VMCS_GUEST_CR3, val);
886	AssertRC(rc);
887	}
888
889	/* Debug registers. */
890	if (pVM->hwaccm.s.fContextUseFlags & HWACCM_CHANGED_GUEST_DEBUG)
891	{
892	/** @todo DR0-6 */
893	val = pCtx->dr7;
894	val &= ~(RT_BIT(11) \| RT_BIT(12) \| RT_BIT(14) \| RT_BIT(15)); /* must be zero */
895	val \|= 0x400; /* must be one */
896	#ifdef VBOX_STRICT
897	val = 0x400;
898	#endif
899	rc \|= VMXWriteVMCS(VMX_VMCS_GUEST_DR7, val);
900	AssertRC(rc);
901
902	/* IA32_DEBUGCTL MSR. */
903	rc = VMXWriteVMCS(VMX_VMCS_GUEST_DEBUGCTL_FULL, 0);
904	rc \|= VMXWriteVMCS(VMX_VMCS_GUEST_DEBUGCTL_HIGH, 0);
905	AssertRC(rc);
906
907	/** @todo */
908	rc \|= VMXWriteVMCS(VMX_VMCS_GUEST_DEBUG_EXCEPTIONS, 0);
909	AssertRC(rc);
910	}
911
912	/* EIP, ESP and EFLAGS */
913	rc = VMXWriteVMCS(VMX_VMCS_GUEST_RIP, pCtx->eip);
914	rc \|= VMXWriteVMCS(VMX_VMCS_GUEST_RSP, pCtx->esp);
915	AssertRC(rc);
916
917	/* Bits 22-31, 15, 5 & 3 must be zero. Bit 1 must be 1. */
918	eflags = pCtx->eflags;
919	eflags.u32 &= VMX_EFLAGS_RESERVED_0;
920	eflags.u32 \|= VMX_EFLAGS_RESERVED_1;
921
922	/* Real mode emulation using v86 mode with CR4.VME (interrupt redirection using the int bitmap in the TSS) */
923	if (!(pCtx->cr0 & X86_CR0_PROTECTION_ENABLE))
924	{
925	eflags.Bits.u1VM = 1;
926	eflags.Bits.u1VIF = pCtx->eflags.Bits.u1IF;
927	eflags.Bits.u2IOPL = 3;
928	}
929
930	rc = VMXWriteVMCS(VMX_VMCS_GUEST_RFLAGS, eflags.u32);
931	AssertRC(rc);
932
933	/** TSC offset. */
934	uint64_t u64TSCOffset;
935
936	if (TMCpuTickCanUseRealTSC(pVM, &u64TSCOffset))
937	{
938	/* Note: VMX_VMCS_CTRL_PROC_EXEC_CONTROLS_RDTSC_EXIT takes precedence over TSC_OFFSET */
939	#if HC_ARCH_BITS == 64
940	rc = VMXWriteVMCS(VMX_VMCS_CTRL_TSC_OFFSET_FULL, u64TSCOffset);
941	#else
942	rc = VMXWriteVMCS(VMX_VMCS_CTRL_TSC_OFFSET_FULL, (uint32_t)u64TSCOffset);
943	rc \|= VMXWriteVMCS(VMX_VMCS_CTRL_TSC_OFFSET_HIGH, (uint32_t)(u64TSCOffset >> 32ULL));
944	#endif
945	AssertRC(rc);
946
947	pVM->hwaccm.s.vmx.proc_ctls &= ~VMX_VMCS_CTRL_PROC_EXEC_CONTROLS_RDTSC_EXIT;
948	rc = VMXWriteVMCS(VMX_VMCS_CTRL_PROC_EXEC_CONTROLS, pVM->hwaccm.s.vmx.proc_ctls);
949	AssertRC(rc);
950	}
951	else
952	{
953	pVM->hwaccm.s.vmx.proc_ctls \|= VMX_VMCS_CTRL_PROC_EXEC_CONTROLS_RDTSC_EXIT;
954	rc = VMXWriteVMCS(VMX_VMCS_CTRL_PROC_EXEC_CONTROLS, pVM->hwaccm.s.vmx.proc_ctls);
955	AssertRC(rc);
956	}
957
958	/* Done. */
959	pVM->hwaccm.s.fContextUseFlags &= ~HWACCM_CHANGED_ALL_GUEST;
960
961	return rc;
962	}
963
964	/**
965	* Runs guest code in a VT-x VM.
966	*
967	* @note NEVER EVER turn on interrupts here. Due to our illegal entry into the kernel, it might mess things up. (XP kernel traps have been frequently observed)
968	*
969	* @returns VBox status code.
970	* @param pVM The VM to operate on.
971	* @param pCtx Guest context
972	* @param pCpu CPU info struct
973	*/
974	HWACCMR0DECL(int) VMXR0RunGuestCode(PVM pVM, CPUMCTX *pCtx, PHWACCM_CPUINFO pCpu)
975	{
976	int rc = VINF_SUCCESS;
977	RTCCUINTREG val, valShadow;
978	RTCCUINTREG exitReason, instrError, cbInstr;
979	RTGCUINTPTR exitQualification;
980	RTGCUINTPTR intInfo = 0; /* shut up buggy gcc 4 */
981	RTGCUINTPTR errCode, instrInfo, uInterruptState;
982	bool fGuestStateSynced = false;
983	unsigned cResume = 0;
984
985	Log2(("\nE"));
986
987	AssertReturn(pCpu->fVMXConfigured, VERR_EM_INTERNAL_ERROR);
988
989	STAM_PROFILE_ADV_START(&pVM->hwaccm.s.StatEntry, x);
990
991	#ifdef VBOX_STRICT
992	rc = VMXReadVMCS(VMX_VMCS_CTRL_PIN_EXEC_CONTROLS, &val);
993	AssertRC(rc);
994	Log2(("VMX_VMCS_CTRL_PIN_EXEC_CONTROLS = %08x\n", val));
995
996	/* allowed zero */
997	if ((val & (pVM->hwaccm.s.vmx.msr.vmx_pin_ctls & 0xFFFFFFFF)) != (pVM->hwaccm.s.vmx.msr.vmx_pin_ctls & 0xFFFFFFFF))
998	Log(("Invalid VMX_VMCS_CTRL_PIN_EXEC_CONTROLS: zero\n"));
999
1000	/* allowed one */
1001	if ((val & ~(pVM->hwaccm.s.vmx.msr.vmx_pin_ctls >> 32ULL)) != 0)
1002	Log(("Invalid VMX_VMCS_CTRL_PIN_EXEC_CONTROLS: one\n"));
1003
1004	rc = VMXReadVMCS(VMX_VMCS_CTRL_PROC_EXEC_CONTROLS, &val);
1005	AssertRC(rc);
1006	Log2(("VMX_VMCS_CTRL_PROC_EXEC_CONTROLS = %08x\n", val));
1007
1008	/* allowed zero */
1009	if ((val & (pVM->hwaccm.s.vmx.msr.vmx_proc_ctls & 0xFFFFFFFF)) != (pVM->hwaccm.s.vmx.msr.vmx_proc_ctls & 0xFFFFFFFF))
1010	Log(("Invalid VMX_VMCS_CTRL_PROC_EXEC_CONTROLS: zero\n"));
1011
1012	/* allowed one */
1013	if ((val & ~(pVM->hwaccm.s.vmx.msr.vmx_proc_ctls >> 32ULL)) != 0)
1014	Log(("Invalid VMX_VMCS_CTRL_PROC_EXEC_CONTROLS: one\n"));
1015
1016	rc = VMXReadVMCS(VMX_VMCS_CTRL_ENTRY_CONTROLS, &val);
1017	AssertRC(rc);
1018	Log2(("VMX_VMCS_CTRL_ENTRY_CONTROLS = %08x\n", val));
1019
1020	/* allowed zero */
1021	if ((val & (pVM->hwaccm.s.vmx.msr.vmx_entry & 0xFFFFFFFF)) != (pVM->hwaccm.s.vmx.msr.vmx_entry & 0xFFFFFFFF))
1022	Log(("Invalid VMX_VMCS_CTRL_ENTRY_CONTROLS: zero\n"));
1023
1024	/* allowed one */
1025	if ((val & ~(pVM->hwaccm.s.vmx.msr.vmx_entry >> 32ULL)) != 0)
1026	Log(("Invalid VMX_VMCS_CTRL_ENTRY_CONTROLS: one\n"));
1027
1028	rc = VMXReadVMCS(VMX_VMCS_CTRL_EXIT_CONTROLS, &val);
1029	AssertRC(rc);
1030	Log2(("VMX_VMCS_CTRL_EXIT_CONTROLS = %08x\n", val));
1031
1032	/* allowed zero */
1033	if ((val & (pVM->hwaccm.s.vmx.msr.vmx_exit & 0xFFFFFFFF)) != (pVM->hwaccm.s.vmx.msr.vmx_exit & 0xFFFFFFFF))
1034	Log(("Invalid VMX_VMCS_CTRL_EXIT_CONTROLS: zero\n"));
1035
1036	/* allowed one */
1037	if ((val & ~(pVM->hwaccm.s.vmx.msr.vmx_exit >> 32ULL)) != 0)
1038	Log(("Invalid VMX_VMCS_CTRL_EXIT_CONTROLS: one\n"));
1039	#endif
1040
1041	#if 0
1042	/*
1043	* Check if debug registers are armed.
1044	*/
1045	uint32_t u32DR7 = ASMGetDR7();
1046	if (u32DR7 & X86_DR7_ENABLED_MASK)
1047	{
1048	pVM->cpum.s.fUseFlags \|= CPUM_USE_DEBUG_REGS_HOST;
1049	}
1050	else
1051	pVM->cpum.s.fUseFlags &= ~CPUM_USE_DEBUG_REGS_HOST;
1052	#endif
1053
1054	/* We can jump to this point to resume execution after determining that a VM-exit is innocent.
1055	*/
1056	ResumeExecution:
1057	/* Safety precaution; looping for too long here can have a very bad effect on the host */
1058	if (++cResume > HWACCM_MAX_RESUME_LOOPS)
1059	{
1060	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitMaxResume);
1061	rc = VINF_EM_RAW_INTERRUPT;
1062	goto end;
1063	}
1064
1065	/* Check for irq inhibition due to instruction fusing (sti, mov ss). */
1066	if (VM_FF_ISSET(pVM, VM_FF_INHIBIT_INTERRUPTS))
1067	{
1068	Log(("VM_FF_INHIBIT_INTERRUPTS at %VGv successor %VGv\n", pCtx->eip, EMGetInhibitInterruptsPC(pVM)));
1069	if (pCtx->eip != EMGetInhibitInterruptsPC(pVM))
1070	{
1071	/** @note we intentionally don't clear VM_FF_INHIBIT_INTERRUPTS here.
1072	* Before we are able to execute this instruction in raw mode (iret to guest code) an external interrupt might
1073	* force a world switch again. Possibly allowing a guest interrupt to be dispatched in the process. This could
1074	* break the guest. Sounds very unlikely, but such timing sensitive problem are not as rare as you might think.
1075	*/
1076	VM_FF_CLEAR(pVM, VM_FF_INHIBIT_INTERRUPTS);
1077	/* Irq inhibition is no longer active; clear the corresponding VMX state. */
1078	rc = VMXWriteVMCS(VMX_VMCS_GUEST_INTERRUPTIBILITY_STATE, 0);
1079	AssertRC(rc);
1080	}
1081	}
1082	else
1083	{
1084	/* Irq inhibition is no longer active; clear the corresponding VMX state. */
1085	rc = VMXWriteVMCS(VMX_VMCS_GUEST_INTERRUPTIBILITY_STATE, 0);
1086	AssertRC(rc);
1087	}
1088
1089	/* Check for pending actions that force us to go back to ring 3. */
1090	if (VM_FF_ISPENDING(pVM, VM_FF_TO_R3 \| VM_FF_TIMER))
1091	{
1092	VM_FF_CLEAR(pVM, VM_FF_TO_R3);
1093	STAM_COUNTER_INC(&pVM->hwaccm.s.StatSwitchToR3);
1094	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatEntry, x);
1095	rc = VINF_EM_RAW_TO_R3;
1096	goto end;
1097	}
1098	/* Pending request packets might contain actions that need immediate attention, such as pending hardware interrupts. */
1099	if (VM_FF_ISPENDING(pVM, VM_FF_REQUEST))
1100	{
1101	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatEntry, x);
1102	rc = VINF_EM_PENDING_REQUEST;
1103	goto end;
1104	}
1105
1106	/* When external interrupts are pending, we should exit the VM when IF is set. */
1107	/** @note after VM_FF_INHIBIT_INTERRUPTS check!!! */
1108	rc = VMXR0CheckPendingInterrupt(pVM, pCtx);
1109	if (VBOX_FAILURE(rc))
1110	{
1111	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatEntry, x);
1112	goto end;
1113	}
1114
1115	/** @todo check timers?? */
1116
1117	/* Save the host state first. */
1118	rc = VMXR0SaveHostState(pVM);
1119	if (rc != VINF_SUCCESS)
1120	{
1121	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatEntry, x);
1122	goto end;
1123	}
1124	/* Load the guest state */
1125	rc = VMXR0LoadGuestState(pVM, pCtx);
1126	if (rc != VINF_SUCCESS)
1127	{
1128	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatEntry, x);
1129	goto end;
1130	}
1131	fGuestStateSynced = true;
1132
1133	/* Non-register state Guest Context */
1134	/** @todo change me according to cpu state */
1135	rc = VMXWriteVMCS(VMX_VMCS_GUEST_ACTIVITY_STATE, VMX_CMS_GUEST_ACTIVITY_ACTIVE);
1136	AssertRC(rc);
1137
1138	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatEntry, x);
1139
1140	/* Manual save and restore:
1141	* - General purpose registers except RIP, RSP
1142	*
1143	* Trashed:
1144	* - CR2 (we don't care)
1145	* - LDTR (reset to 0)
1146	* - DRx (presumably not changed at all)
1147	* - DR7 (reset to 0x400)
1148	* - EFLAGS (reset to RT_BIT(1); not relevant)
1149	*
1150	*/
1151
1152	/* All done! Let's start VM execution. */
1153	STAM_PROFILE_ADV_START(&pVM->hwaccm.s.StatInGC, x);
1154	if (pVM->hwaccm.s.vmx.fResumeVM == false)
1155	rc = VMXStartVM(pCtx);
1156	else
1157	rc = VMXResumeVM(pCtx);
1158
1159	/* In case we execute a goto ResumeExecution later on. */
1160	pVM->hwaccm.s.vmx.fResumeVM = true;
1161
1162	/**
1163	* !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
1164	* IMPORTANT: WE CAN'T DO ANY LOGGING OR OPERATIONS THAT CAN DO A LONGJMP BACK TO RING 3 BEFORE WE'VE SYNCED BACK (MOST OF) THE GUEST STATE
1165	* !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
1166	*/
1167
1168	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatInGC, x);
1169	STAM_PROFILE_ADV_START(&pVM->hwaccm.s.StatExit, x);
1170
1171	switch (rc)
1172	{
1173	case VINF_SUCCESS:
1174	break;
1175
1176	case VERR_VMX_INVALID_VMXON_PTR:
1177	AssertFailed();
1178	goto end;
1179
1180	case VERR_VMX_UNABLE_TO_START_VM:
1181	case VERR_VMX_UNABLE_TO_RESUME_VM:
1182	{
1183	#ifdef VBOX_STRICT
1184	int rc1;
1185
1186	rc1 = VMXReadVMCS(VMX_VMCS_RO_EXIT_REASON, &exitReason);
1187	rc1 \|= VMXReadVMCS(VMX_VMCS_RO_VM_INSTR_ERROR, &instrError);
1188	AssertRC(rc1);
1189	if (rc1 == VINF_SUCCESS)
1190	{
1191	RTGDTR gdtr;
1192	PX86DESCHC pDesc;
1193
1194	ASMGetGDTR(&gdtr);
1195
1196	Log(("Unable to start/resume VM for reason: %x. Instruction error %x\n", (uint32_t)exitReason, (uint32_t)instrError));
1197	Log(("Current stack %08x\n", &rc1));
1198
1199
1200	VMXReadVMCS(VMX_VMCS_GUEST_RIP, &val);
1201	Log(("Old eip %VGv new %VGv\n", pCtx->eip, (RTGCPTR)val));
1202	VMXReadVMCS(VMX_VMCS_CTRL_PIN_EXEC_CONTROLS, &val);
1203	Log(("VMX_VMCS_CTRL_PIN_EXEC_CONTROLS %08x\n", val));
1204	VMXReadVMCS(VMX_VMCS_CTRL_PROC_EXEC_CONTROLS, &val);
1205	Log(("VMX_VMCS_CTRL_PROC_EXEC_CONTROLS %08x\n", val));
1206	VMXReadVMCS(VMX_VMCS_CTRL_ENTRY_CONTROLS, &val);
1207	Log(("VMX_VMCS_CTRL_ENTRY_CONTROLS %08x\n", val));
1208	VMXReadVMCS(VMX_VMCS_CTRL_EXIT_CONTROLS, &val);
1209	Log(("VMX_VMCS_CTRL_EXIT_CONTROLS %08x\n", val));
1210
1211	VMXReadVMCS(VMX_VMCS_HOST_CR0, &val);
1212	Log(("VMX_VMCS_HOST_CR0 %08x\n", val));
1213
1214	VMXReadVMCS(VMX_VMCS_HOST_CR3, &val);
1215	Log(("VMX_VMCS_HOST_CR3 %VHp\n", val));
1216
1217	VMXReadVMCS(VMX_VMCS_HOST_CR4, &val);
1218	Log(("VMX_VMCS_HOST_CR4 %08x\n", val));
1219
1220	VMXReadVMCS(VMX_VMCS_HOST_FIELD_CS, &val);
1221	Log(("VMX_VMCS_HOST_FIELD_CS %08x\n", val));
1222	if (val < gdtr.cbGdt)
1223	{
1224	pDesc = &((PX86DESCHC)gdtr.pGdt)[val >> X86_SEL_SHIFT_HC];
1225	HWACCMR0DumpDescriptor(pDesc, val, "CS: ");
1226	}
1227
1228	VMXReadVMCS(VMX_VMCS_HOST_FIELD_DS, &val);
1229	Log(("VMX_VMCS_HOST_FIELD_DS %08x\n", val));
1230	if (val < gdtr.cbGdt)
1231	{
1232	pDesc = &((PX86DESCHC)gdtr.pGdt)[val >> X86_SEL_SHIFT_HC];
1233	HWACCMR0DumpDescriptor(pDesc, val, "DS: ");
1234	}
1235
1236	VMXReadVMCS(VMX_VMCS_HOST_FIELD_ES, &val);
1237	Log(("VMX_VMCS_HOST_FIELD_ES %08x\n", val));
1238	if (val < gdtr.cbGdt)
1239	{
1240	pDesc = &((PX86DESCHC)gdtr.pGdt)[val >> X86_SEL_SHIFT_HC];
1241	HWACCMR0DumpDescriptor(pDesc, val, "ES: ");
1242	}
1243
1244	VMXReadVMCS(VMX_VMCS_HOST_FIELD_FS, &val);
1245	Log(("VMX_VMCS_HOST_FIELD_FS %08x\n", val));
1246	if (val < gdtr.cbGdt)
1247	{
1248	pDesc = &((PX86DESCHC)gdtr.pGdt)[val >> X86_SEL_SHIFT_HC];
1249	HWACCMR0DumpDescriptor(pDesc, val, "FS: ");
1250	}
1251
1252	VMXReadVMCS(VMX_VMCS_HOST_FIELD_GS, &val);
1253	Log(("VMX_VMCS_HOST_FIELD_GS %08x\n", val));
1254	if (val < gdtr.cbGdt)
1255	{
1256	pDesc = &((PX86DESCHC)gdtr.pGdt)[val >> X86_SEL_SHIFT_HC];
1257	HWACCMR0DumpDescriptor(pDesc, val, "GS: ");
1258	}
1259
1260	VMXReadVMCS(VMX_VMCS_HOST_FIELD_SS, &val);
1261	Log(("VMX_VMCS_HOST_FIELD_SS %08x\n", val));
1262	if (val < gdtr.cbGdt)
1263	{
1264	pDesc = &((PX86DESCHC)gdtr.pGdt)[val >> X86_SEL_SHIFT_HC];
1265	HWACCMR0DumpDescriptor(pDesc, val, "SS: ");
1266	}
1267
1268	VMXReadVMCS(VMX_VMCS_HOST_FIELD_TR, &val);
1269	Log(("VMX_VMCS_HOST_FIELD_TR %08x\n", val));
1270	if (val < gdtr.cbGdt)
1271	{
1272	pDesc = &((PX86DESCHC)gdtr.pGdt)[val >> X86_SEL_SHIFT_HC];
1273	HWACCMR0DumpDescriptor(pDesc, val, "TR: ");
1274	}
1275
1276	VMXReadVMCS(VMX_VMCS_HOST_TR_BASE, &val);
1277	Log(("VMX_VMCS_HOST_TR_BASE %VHv\n", val));
1278
1279	VMXReadVMCS(VMX_VMCS_HOST_GDTR_BASE, &val);
1280	Log(("VMX_VMCS_HOST_GDTR_BASE %VHv\n", val));
1281	VMXReadVMCS(VMX_VMCS_HOST_IDTR_BASE, &val);
1282	Log(("VMX_VMCS_HOST_IDTR_BASE %VHv\n", val));
1283
1284	VMXReadVMCS(VMX_VMCS_HOST_SYSENTER_CS, &val);
1285	Log(("VMX_VMCS_HOST_SYSENTER_CS %08x\n", val));
1286
1287	VMXReadVMCS(VMX_VMCS_HOST_SYSENTER_EIP, &val);
1288	Log(("VMX_VMCS_HOST_SYSENTER_EIP %VHv\n", val));
1289
1290	VMXReadVMCS(VMX_VMCS_HOST_SYSENTER_ESP, &val);
1291	Log(("VMX_VMCS_HOST_SYSENTER_ESP %VHv\n", val));
1292
1293	VMXReadVMCS(VMX_VMCS_HOST_RSP, &val);
1294	Log(("VMX_VMCS_HOST_RSP %VHv\n", val));
1295	VMXReadVMCS(VMX_VMCS_HOST_RIP, &val);
1296	Log(("VMX_VMCS_HOST_RIP %VHv\n", val));
1297
1298	#if HC_ARCH_BITS == 64
1299	Log(("MSR_K6_EFER = %VX64\n", ASMRdMsr(MSR_K6_EFER)));
1300	Log(("MSR_K6_STAR = %VX64\n", ASMRdMsr(MSR_K6_STAR)));
1301	Log(("MSR_K8_LSTAR = %VX64\n", ASMRdMsr(MSR_K8_LSTAR)));
1302	Log(("MSR_K8_CSTAR = %VX64\n", ASMRdMsr(MSR_K8_CSTAR)));
1303	Log(("MSR_K8_SF_MASK = %VX64\n", ASMRdMsr(MSR_K8_SF_MASK)));
1304	#endif
1305	}
1306	#endif /* VBOX_STRICT */
1307	goto end;
1308	}
1309
1310	default:
1311	/* impossible */
1312	AssertFailed();
1313	goto end;
1314	}
1315	/* Success. Query the guest state and figure out what has happened. */
1316
1317	/* Investigate why there was a VM-exit. */
1318	rc = VMXReadVMCS(VMX_VMCS_RO_EXIT_REASON, &exitReason);
1319	STAM_COUNTER_INC(&pVM->hwaccm.s.pStatExitReasonR0[exitReason & MASK_EXITREASON_STAT]);
1320
1321	exitReason &= 0xffff; /* bit 0-15 contain the exit code. */
1322	rc \|= VMXReadVMCS(VMX_VMCS_RO_VM_INSTR_ERROR, &instrError);
1323	rc \|= VMXReadVMCS(VMX_VMCS_RO_EXIT_INSTR_LENGTH, &cbInstr);
1324	rc \|= VMXReadVMCS(VMX_VMCS_RO_EXIT_INTERRUPTION_INFO, &val);
1325	intInfo = val;
1326	rc \|= VMXReadVMCS(VMX_VMCS_RO_EXIT_INTERRUPTION_ERRCODE, &val);
1327	errCode = val; /* might not be valid; depends on VMX_EXIT_INTERRUPTION_INFO_ERROR_CODE_IS_VALID. */
1328	rc \|= VMXReadVMCS(VMX_VMCS_RO_EXIT_INSTR_INFO, &val);
1329	instrInfo = val;
1330	rc \|= VMXReadVMCS(VMX_VMCS_RO_EXIT_QUALIFICATION, &val);
1331	exitQualification = val;
1332	AssertRC(rc);
1333
1334	/* Take care of instruction fusing (sti, mov ss) */
1335	rc \|= VMXReadVMCS(VMX_VMCS_GUEST_INTERRUPTIBILITY_STATE, &val);
1336	uInterruptState = val;
1337	if (uInterruptState != 0)
1338	{
1339	Assert(uInterruptState <= 2); /* only sti & mov ss */
1340	Log(("uInterruptState %x eip=%VGv\n", uInterruptState, pCtx->eip));
1341	EMSetInhibitInterruptsPC(pVM, pCtx->eip);
1342	}
1343	else
1344	VM_FF_CLEAR(pVM, VM_FF_INHIBIT_INTERRUPTS);
1345
1346	/* Let's first sync back eip, esp, and eflags. */
1347	rc = VMXReadVMCS(VMX_VMCS_GUEST_RIP, &val);
1348	AssertRC(rc);
1349	pCtx->eip = val;
1350	rc = VMXReadVMCS(VMX_VMCS_GUEST_RSP, &val);
1351	AssertRC(rc);
1352	pCtx->esp = val;
1353	rc = VMXReadVMCS(VMX_VMCS_GUEST_RFLAGS, &val);
1354	AssertRC(rc);
1355	pCtx->eflags.u32 = val;
1356
1357	/* Real mode emulation using v86 mode with CR4.VME (interrupt redirection using the int bitmap in the TSS) */
1358	if (!(pCtx->cr0 & X86_CR0_PROTECTION_ENABLE))
1359	{
1360	/* Hide our emulation flags */
1361	pCtx->eflags.Bits.u1VM = 0;
1362	pCtx->eflags.Bits.u1IF = pCtx->eflags.Bits.u1VIF;
1363	pCtx->eflags.Bits.u1VIF = 0;
1364	pCtx->eflags.Bits.u2IOPL = 0;
1365	}
1366
1367	/* Control registers. */
1368	VMXReadVMCS(VMX_VMCS_CTRL_CR0_READ_SHADOW, &valShadow);
1369	VMXReadVMCS(VMX_VMCS_GUEST_CR0, &val);
1370	val = (valShadow & pVM->hwaccm.s.vmx.cr0_mask) \| (val & ~pVM->hwaccm.s.vmx.cr0_mask);
1371	CPUMSetGuestCR0(pVM, val);
1372
1373	VMXReadVMCS(VMX_VMCS_CTRL_CR4_READ_SHADOW, &valShadow);
1374	VMXReadVMCS(VMX_VMCS_GUEST_CR4, &val);
1375	val = (valShadow & pVM->hwaccm.s.vmx.cr4_mask) \| (val & ~pVM->hwaccm.s.vmx.cr4_mask);
1376	CPUMSetGuestCR4(pVM, val);
1377
1378	CPUMSetGuestCR2(pVM, ASMGetCR2());
1379
1380	VMXReadVMCS(VMX_VMCS_GUEST_DR7, &val);
1381	CPUMSetGuestDR7(pVM, val);
1382
1383	/* Guest CPU context: ES, CS, SS, DS, FS, GS. */
1384	VMX_READ_SELREG(ES, es);
1385	VMX_READ_SELREG(SS, ss);
1386	VMX_READ_SELREG(CS, cs);
1387	VMX_READ_SELREG(DS, ds);
1388	VMX_READ_SELREG(FS, fs);
1389	VMX_READ_SELREG(GS, gs);
1390
1391	/** @note NOW IT'S SAFE FOR LOGGING! */
1392	Log2(("Raw exit reason %08x\n", exitReason));
1393
1394	/* Check if an injected event was interrupted prematurely. */
1395	rc = VMXReadVMCS(VMX_VMCS_RO_IDT_INFO, &val);
1396	AssertRC(rc);
1397	pVM->hwaccm.s.Event.intInfo = VMX_VMCS_CTRL_ENTRY_IRQ_INFO_FROM_EXIT_INT_INFO(val);
1398	if ( VMX_EXIT_INTERRUPTION_INFO_VALID(pVM->hwaccm.s.Event.intInfo)
1399	&& VMX_EXIT_INTERRUPTION_INFO_TYPE(pVM->hwaccm.s.Event.intInfo) != VMX_EXIT_INTERRUPTION_INFO_TYPE_SW)
1400	{
1401	Log(("Pending inject %VX64 at %08x exit=%08x intInfo=%08x exitQualification=%08x\n", pVM->hwaccm.s.Event.intInfo, pCtx->eip, exitReason, intInfo, exitQualification));
1402	pVM->hwaccm.s.Event.fPending = true;
1403	/* Error code present? */
1404	if (VMX_EXIT_INTERRUPTION_INFO_ERROR_CODE_IS_VALID(pVM->hwaccm.s.Event.intInfo))
1405	{
1406	rc = VMXReadVMCS(VMX_VMCS_RO_IDT_ERRCODE, &val);
1407	AssertRC(rc);
1408	pVM->hwaccm.s.Event.errCode = val;
1409	}
1410	else
1411	pVM->hwaccm.s.Event.errCode = 0;
1412	}
1413
1414	#ifdef VBOX_STRICT
1415	if (exitReason == VMX_EXIT_ERR_INVALID_GUEST_STATE)
1416	HWACCMDumpRegs(pCtx);
1417	#endif
1418
1419	Log2(("E%d", exitReason));
1420	Log2(("Exit reason %d, exitQualification %08x\n", exitReason, exitQualification));
1421	Log2(("instrInfo=%d instrError=%d instr length=%d\n", instrInfo, instrError, cbInstr));
1422	Log2(("Interruption error code %d\n", errCode));
1423	Log2(("IntInfo = %08x\n", intInfo));
1424	Log2(("New EIP=%VGv\n", pCtx->eip));
1425
1426	/* Some cases don't need a complete resync of the guest CPU state; handle them here. */
1427	switch (exitReason)
1428	{
1429	case VMX_EXIT_EXCEPTION: /* 0 Exception or non-maskable interrupt (NMI). */
1430	case VMX_EXIT_EXTERNAL_IRQ: /* 1 External interrupt. */
1431	{
1432	uint32_t vector = VMX_EXIT_INTERRUPTION_INFO_VECTOR(intInfo);
1433
1434	if (!VMX_EXIT_INTERRUPTION_INFO_VALID(intInfo))
1435	{
1436	Assert(exitReason == VMX_EXIT_EXTERNAL_IRQ);
1437	/* External interrupt; leave to allow it to be dispatched again. */
1438	rc = VINF_EM_RAW_INTERRUPT;
1439	break;
1440	}
1441	switch (VMX_EXIT_INTERRUPTION_INFO_TYPE(intInfo))
1442	{
1443	case VMX_EXIT_INTERRUPTION_INFO_TYPE_NMI: /* Non-maskable interrupt. */
1444	/* External interrupt; leave to allow it to be dispatched again. */
1445	rc = VINF_EM_RAW_INTERRUPT;
1446	break;
1447
1448	case VMX_EXIT_INTERRUPTION_INFO_TYPE_EXT: /* External hardware interrupt. */
1449	AssertFailed(); /* can't come here; fails the first check. */
1450	break;
1451
1452	case VMX_EXIT_INTERRUPTION_INFO_TYPE_SWEXCPT: /* Software exception. (#BP or #OF) */
1453	Assert(vector == 3 \|\| vector == 4);
1454	/* no break */
1455	case VMX_EXIT_INTERRUPTION_INFO_TYPE_HWEXCPT: /* Hardware exception. */
1456	Log2(("Hardware/software interrupt %d\n", vector));
1457	switch (vector)
1458	{
1459	case X86_XCPT_NM:
1460	{
1461	uint32_t oldCR0;
1462
1463	Log(("#NM fault at %VGv error code %x\n", pCtx->eip, errCode));
1464
1465	/** @todo don't intercept #NM exceptions anymore when we've activated the guest FPU state. */
1466	oldCR0 = ASMGetCR0();
1467	/* If we sync the FPU/XMM state on-demand, then we can continue execution as if nothing has happened. */
1468	rc = CPUMHandleLazyFPU(pVM);
1469	if (rc == VINF_SUCCESS)
1470	{
1471	Assert(CPUMIsGuestFPUStateActive(pVM));
1472
1473	/* CPUMHandleLazyFPU could have changed CR0; restore it. */
1474	ASMSetCR0(oldCR0);
1475
1476	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitShadowNM);
1477
1478	/* Continue execution. */
1479	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1480	pVM->hwaccm.s.fContextUseFlags \|= HWACCM_CHANGED_GUEST_CR0;
1481
1482	goto ResumeExecution;
1483	}
1484
1485	Log(("Forward #NM fault to the guest\n"));
1486	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitGuestNM);
1487	rc = VMXR0InjectEvent(pVM, pCtx, VMX_VMCS_CTRL_ENTRY_IRQ_INFO_FROM_EXIT_INT_INFO(intInfo), cbInstr, 0);
1488	AssertRC(rc);
1489	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1490	goto ResumeExecution;
1491	}
1492
1493	case X86_XCPT_PF: /* Page fault */
1494	{
1495	Log2(("Page fault at %VGv error code %x\n", exitQualification ,errCode));
1496	/* Exit qualification contains the linear address of the page fault. */
1497	TRPMAssertTrap(pVM, X86_XCPT_PF, TRPM_TRAP);
1498	TRPMSetErrorCode(pVM, errCode);
1499	TRPMSetFaultAddress(pVM, exitQualification);
1500
1501	/* Forward it to our trap handler first, in case our shadow pages are out of sync. */
1502	rc = PGMTrap0eHandler(pVM, errCode, CPUMCTX2CORE(pCtx), (RTGCPTR)exitQualification);
1503	Log2(("PGMTrap0eHandler %VGv returned %Vrc\n", pCtx->eip, rc));
1504	if (rc == VINF_SUCCESS)
1505	{ /* We've successfully synced our shadow pages, so let's just continue execution. */
1506	Log2(("Shadow page fault at %VGv cr2=%VGv error code %x\n", pCtx->eip, exitQualification ,errCode));
1507	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitShadowPF);
1508
1509	TRPMResetTrap(pVM);
1510
1511	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1512	goto ResumeExecution;
1513	}
1514	else
1515	if (rc == VINF_EM_RAW_GUEST_TRAP)
1516	{ /* A genuine pagefault.
1517	* Forward the trap to the guest by injecting the exception and resuming execution.
1518	*/
1519	Log2(("Forward page fault to the guest\n"));
1520	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitGuestPF);
1521	/* The error code might have been changed. */
1522	errCode = TRPMGetErrorCode(pVM);
1523
1524	TRPMResetTrap(pVM);
1525
1526	/* Now we must update CR2. */
1527	pCtx->cr2 = exitQualification;
1528	rc = VMXR0InjectEvent(pVM, pCtx, VMX_VMCS_CTRL_ENTRY_IRQ_INFO_FROM_EXIT_INT_INFO(intInfo), cbInstr, errCode);
1529	AssertRC(rc);
1530
1531	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1532	goto ResumeExecution;
1533	}
1534	#ifdef VBOX_STRICT
1535	if (rc != VINF_EM_RAW_EMULATE_INSTR)
1536	Log2(("PGMTrap0eHandler failed with %d\n", rc));
1537	#endif
1538	/* Need to go back to the recompiler to emulate the instruction. */
1539	TRPMResetTrap(pVM);
1540	break;
1541	}
1542
1543	case X86_XCPT_MF: /* Floating point exception. */
1544	{
1545	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitGuestMF);
1546	if (!(pCtx->cr0 & X86_CR0_NE))
1547	{
1548	/* old style FPU error reporting needs some extra work. */
1549	/** @todo don't fall back to the recompiler, but do it manually. */
1550	rc = VINF_EM_RAW_EMULATE_INSTR;
1551	break;
1552	}
1553	Log(("Trap %x at %VGv\n", vector, pCtx->eip));
1554	rc = VMXR0InjectEvent(pVM, pCtx, VMX_VMCS_CTRL_ENTRY_IRQ_INFO_FROM_EXIT_INT_INFO(intInfo), cbInstr, errCode);
1555	AssertRC(rc);
1556
1557	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1558	goto ResumeExecution;
1559	}
1560
1561	#ifdef VBOX_STRICT
1562	case X86_XCPT_GP: /* General protection failure exception.*/
1563	case X86_XCPT_UD: /* Unknown opcode exception. */
1564	case X86_XCPT_DE: /* Debug exception. */
1565	case X86_XCPT_SS: /* Stack segment exception. */
1566	case X86_XCPT_NP: /* Segment not present exception. */
1567	{
1568	switch(vector)
1569	{
1570	case X86_XCPT_DE:
1571	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitGuestDE);
1572	break;
1573	case X86_XCPT_UD:
1574	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitGuestUD);
1575	break;
1576	case X86_XCPT_SS:
1577	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitGuestSS);
1578	break;
1579	case X86_XCPT_NP:
1580	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitGuestNP);
1581	break;
1582	case X86_XCPT_GP:
1583	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitGuestGP);
1584	break;
1585	}
1586
1587	Log(("Trap %x at %VGv\n", vector, pCtx->eip));
1588	rc = VMXR0InjectEvent(pVM, pCtx, VMX_VMCS_CTRL_ENTRY_IRQ_INFO_FROM_EXIT_INT_INFO(intInfo), cbInstr, errCode);
1589	AssertRC(rc);
1590
1591	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1592	goto ResumeExecution;
1593	}
1594	#endif
1595	default:
1596	AssertMsgFailed(("Unexpected vm-exit caused by exception %x\n", vector));
1597	rc = VERR_EM_INTERNAL_ERROR;
1598	break;
1599	} /* switch (vector) */
1600
1601	break;
1602
1603	default:
1604	rc = VERR_EM_INTERNAL_ERROR;
1605	AssertFailed();
1606	break;
1607	}
1608
1609	break;
1610	}
1611
1612	case VMX_EXIT_IRQ_WINDOW: /* 7 Interrupt window. */
1613	/* Clear VM-exit on IF=1 change. */
1614	Log2(("VMX_EXIT_IRQ_WINDOW %VGv\n", pCtx->eip));
1615	pVM->hwaccm.s.vmx.proc_ctls &= ~VMX_VMCS_CTRL_PROC_EXEC_CONTROLS_IRQ_WINDOW_EXIT;
1616	rc = VMXWriteVMCS(VMX_VMCS_CTRL_PROC_EXEC_CONTROLS, pVM->hwaccm.s.vmx.proc_ctls);
1617	AssertRC(rc);
1618	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitIrqWindow);
1619	goto ResumeExecution; /* we check for pending guest interrupts there */
1620
1621	case VMX_EXIT_INVD: /* 13 Guest software attempted to execute INVD. */
1622	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitInvd);
1623	/* Skip instruction and continue directly. */
1624	pCtx->eip += cbInstr;
1625	/* Continue execution.*/
1626	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1627	goto ResumeExecution;
1628
1629	case VMX_EXIT_CPUID: /* 10 Guest software attempted to execute CPUID. */
1630	{
1631	Log2(("VMX: Cpuid %x\n", pCtx->eax));
1632	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitCpuid);
1633	rc = EMInterpretCpuId(pVM, CPUMCTX2CORE(pCtx));
1634	if (rc == VINF_SUCCESS)
1635	{
1636	/* Update EIP and continue execution. */
1637	Assert(cbInstr == 2);
1638	pCtx->eip += cbInstr;
1639	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1640	goto ResumeExecution;
1641	}
1642	AssertMsgFailed(("EMU: cpuid failed with %Vrc\n", rc));
1643	rc = VINF_EM_RAW_EMULATE_INSTR;
1644	break;
1645	}
1646
1647	case VMX_EXIT_RDTSC: /* 16 Guest software attempted to execute RDTSC. */
1648	{
1649	Log2(("VMX: Rdtsc\n"));
1650	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitRdtsc);
1651	rc = EMInterpretRdtsc(pVM, CPUMCTX2CORE(pCtx));
1652	if (rc == VINF_SUCCESS)
1653	{
1654	/* Update EIP and continue execution. */
1655	Assert(cbInstr == 2);
1656	pCtx->eip += cbInstr;
1657	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1658	goto ResumeExecution;
1659	}
1660	AssertMsgFailed(("EMU: rdtsc failed with %Vrc\n", rc));
1661	rc = VINF_EM_RAW_EMULATE_INSTR;
1662	break;
1663	}
1664
1665	case VMX_EXIT_INVPG: /* 14 Guest software attempted to execute INVPG. */
1666	{
1667	Log2(("VMX: invlpg\n"));
1668	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitInvpg);
1669	rc = EMInterpretInvlpg(pVM, CPUMCTX2CORE(pCtx), exitQualification);
1670	if (rc == VINF_SUCCESS)
1671	{
1672	/* Update EIP and continue execution. */
1673	pCtx->eip += cbInstr;
1674	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1675	goto ResumeExecution;
1676	}
1677	AssertMsg(rc == VERR_EM_INTERPRETER, ("EMU: invlpg %VGv failed with %Vrc\n", exitQualification, rc));
1678	break;
1679	}
1680
1681	case VMX_EXIT_CRX_MOVE: /* 28 Control-register accesses. */
1682	{
1683	switch (VMX_EXIT_QUALIFICATION_CRX_ACCESS(exitQualification))
1684	{
1685	case VMX_EXIT_QUALIFICATION_CRX_ACCESS_WRITE:
1686	Log2(("VMX: %VGv mov cr%d, x\n", pCtx->eip, VMX_EXIT_QUALIFICATION_CRX_REGISTER(exitQualification)));
1687	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitCRxWrite);
1688	rc = EMInterpretCRxWrite(pVM, CPUMCTX2CORE(pCtx),
1689	VMX_EXIT_QUALIFICATION_CRX_REGISTER(exitQualification),
1690	VMX_EXIT_QUALIFICATION_CRX_GENREG(exitQualification));
1691
1692	switch (VMX_EXIT_QUALIFICATION_CRX_REGISTER(exitQualification))
1693	{
1694	case 0:
1695	pVM->hwaccm.s.fContextUseFlags \|= HWACCM_CHANGED_GUEST_CR0;
1696	break;
1697	case 2:
1698	break;
1699	case 3:
1700	pVM->hwaccm.s.fContextUseFlags \|= HWACCM_CHANGED_GUEST_CR3;
1701	break;
1702	case 4:
1703	pVM->hwaccm.s.fContextUseFlags \|= HWACCM_CHANGED_GUEST_CR4;
1704	break;
1705	default:
1706	AssertFailed();
1707	}
1708	/* Check if a sync operation is pending. */
1709	if ( rc == VINF_SUCCESS /* don't bother if we are going to ring 3 anyway */
1710	&& VM_FF_ISPENDING(pVM, VM_FF_PGM_SYNC_CR3 \| VM_FF_PGM_SYNC_CR3_NON_GLOBAL))
1711	{
1712	rc = PGMSyncCR3(pVM, CPUMGetGuestCR0(pVM), CPUMGetGuestCR3(pVM), CPUMGetGuestCR4(pVM), VM_FF_ISSET(pVM, VM_FF_PGM_SYNC_CR3));
1713	AssertRC(rc);
1714	}
1715	break;
1716
1717	case VMX_EXIT_QUALIFICATION_CRX_ACCESS_READ:
1718	Log2(("VMX: mov x, crx\n"));
1719	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitCRxRead);
1720	rc = EMInterpretCRxRead(pVM, CPUMCTX2CORE(pCtx),
1721	VMX_EXIT_QUALIFICATION_CRX_GENREG(exitQualification),
1722	VMX_EXIT_QUALIFICATION_CRX_REGISTER(exitQualification));
1723	break;
1724
1725	case VMX_EXIT_QUALIFICATION_CRX_ACCESS_CLTS:
1726	Log2(("VMX: clts\n"));
1727	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitCLTS);
1728	rc = EMInterpretCLTS(pVM);
1729	pVM->hwaccm.s.fContextUseFlags \|= HWACCM_CHANGED_GUEST_CR0;
1730	break;
1731
1732	case VMX_EXIT_QUALIFICATION_CRX_ACCESS_LMSW:
1733	Log2(("VMX: lmsw %x\n", VMX_EXIT_QUALIFICATION_CRX_LMSW_DATA(exitQualification)));
1734	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitLMSW);
1735	rc = EMInterpretLMSW(pVM, VMX_EXIT_QUALIFICATION_CRX_LMSW_DATA(exitQualification));
1736	pVM->hwaccm.s.fContextUseFlags \|= HWACCM_CHANGED_GUEST_CR0;
1737	break;
1738	}
1739
1740	/* Update EIP if no error occurred. */
1741	if (VBOX_SUCCESS(rc))
1742	pCtx->eip += cbInstr;
1743
1744	if (rc == VINF_SUCCESS)
1745	{
1746	/* Only resume if successful. */
1747	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1748	goto ResumeExecution;
1749	}
1750	Assert(rc == VERR_EM_INTERPRETER \|\| rc == VINF_PGM_CHANGE_MODE \|\| rc == VINF_PGM_SYNC_CR3);
1751	break;
1752	}
1753
1754	case VMX_EXIT_DRX_MOVE: /* 29 Debug-register accesses. */
1755	{
1756	/** @todo clear VMX_VMCS_CTRL_PROC_EXEC_CONTROLS_MOV_DR_EXIT after the first time and restore drx registers afterwards */
1757	if (VMX_EXIT_QUALIFICATION_DRX_DIRECTION(exitQualification) == VMX_EXIT_QUALIFICATION_DRX_DIRECTION_WRITE)
1758	{
1759	Log2(("VMX: mov drx%d, genreg%d\n", VMX_EXIT_QUALIFICATION_DRX_REGISTER(exitQualification), VMX_EXIT_QUALIFICATION_DRX_GENREG(exitQualification)));
1760	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitDRxWrite);
1761	rc = EMInterpretDRxWrite(pVM, CPUMCTX2CORE(pCtx),
1762	VMX_EXIT_QUALIFICATION_DRX_REGISTER(exitQualification),
1763	VMX_EXIT_QUALIFICATION_DRX_GENREG(exitQualification));
1764	Log2(("DR7=%08x\n", pCtx->dr7));
1765	}
1766	else
1767	{
1768	Log2(("VMX: mov x, drx\n"));
1769	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitDRxRead);
1770	rc = EMInterpretDRxRead(pVM, CPUMCTX2CORE(pCtx),
1771	VMX_EXIT_QUALIFICATION_DRX_GENREG(exitQualification),
1772	VMX_EXIT_QUALIFICATION_DRX_REGISTER(exitQualification));
1773	}
1774	/* Update EIP if no error occurred. */
1775	if (VBOX_SUCCESS(rc))
1776	pCtx->eip += cbInstr;
1777
1778	if (rc == VINF_SUCCESS)
1779	{
1780	/* Only resume if successful. */
1781	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1782	goto ResumeExecution;
1783	}
1784	Assert(rc == VERR_EM_INTERPRETER);
1785	break;
1786	}
1787
1788	/** @note We'll get a #GP if the IO instruction isn't allowed (IOPL or TSS bitmap); no need to double check. */
1789	case VMX_EXIT_PORT_IO: /* 30 I/O instruction. */
1790	{
1791	uint32_t uIOWidth = VMX_EXIT_QUALIFICATION_IO_WIDTH(exitQualification);
1792	uint32_t uPort;
1793	bool fIOWrite = (VMX_EXIT_QUALIFICATION_IO_DIRECTION(exitQualification) == VMX_EXIT_QUALIFICATION_IO_DIRECTION_OUT);
1794
1795	/** @todo necessary to make the distinction? */
1796	if (VMX_EXIT_QUALIFICATION_IO_ENCODING(exitQualification) == VMX_EXIT_QUALIFICATION_IO_ENCODING_DX)
1797	{
1798	uPort = pCtx->edx & 0xffff;
1799	}
1800	else
1801	uPort = VMX_EXIT_QUALIFICATION_IO_PORT(exitQualification); /* Immediate encoding. */
1802
1803	/* paranoia */
1804	if (RT_UNLIKELY(uIOWidth == 2 \|\| uIOWidth >= 4))
1805	{
1806	rc = fIOWrite ? VINF_IOM_HC_IOPORT_WRITE : VINF_IOM_HC_IOPORT_READ;
1807	break;
1808	}
1809
1810	uint32_t cbSize = aIOSize[uIOWidth];
1811
1812	if (VMX_EXIT_QUALIFICATION_IO_STRING(exitQualification))
1813	{
1814	/* ins/outs */
1815	uint32_t prefix = 0;
1816	if (VMX_EXIT_QUALIFICATION_IO_REP(exitQualification))
1817	prefix \|= PREFIX_REP;
1818
1819	if (fIOWrite)
1820	{
1821	Log2(("IOMInterpretOUTSEx %VGv %x size=%d\n", pCtx->eip, uPort, cbSize));
1822	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitIOStringWrite);
1823	rc = IOMInterpretOUTSEx(pVM, CPUMCTX2CORE(pCtx), uPort, prefix, cbSize);
1824	}
1825	else
1826	{
1827	Log2(("IOMInterpretINSEx %VGv %x size=%d\n", pCtx->eip, uPort, cbSize));
1828	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitIOStringRead);
1829	rc = IOMInterpretINSEx(pVM, CPUMCTX2CORE(pCtx), uPort, prefix, cbSize);
1830	}
1831	}
1832	else
1833	{
1834	/* normal in/out */
1835	uint32_t uAndVal = aIOOpAnd[uIOWidth];
1836
1837	Assert(!VMX_EXIT_QUALIFICATION_IO_REP(exitQualification));
1838
1839	if (fIOWrite)
1840	{
1841	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitIOWrite);
1842	rc = IOMIOPortWrite(pVM, uPort, pCtx->eax & uAndVal, cbSize);
1843	}
1844	else
1845	{
1846	uint32_t u32Val = 0;
1847
1848	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitIORead);
1849	rc = IOMIOPortRead(pVM, uPort, &u32Val, cbSize);
1850	if (IOM_SUCCESS(rc))
1851	{
1852	/* Write back to the EAX register. */
1853	pCtx->eax = (pCtx->eax & ~uAndVal) \| (u32Val & uAndVal);
1854	}
1855	}
1856	}
1857	/*
1858	* Handled the I/O return codes.
1859	* (The unhandled cases end up with rc == VINF_EM_RAW_EMULATE_INSTR.)
1860	*/
1861	if (IOM_SUCCESS(rc))
1862	{
1863	/* Update EIP and continue execution. */
1864	pCtx->eip += cbInstr;
1865	if (RT_LIKELY(rc == VINF_SUCCESS))
1866	{
1867	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1868	goto ResumeExecution;
1869	}
1870	break;
1871	}
1872
1873	#ifdef VBOX_STRICT
1874	if (rc == VINF_IOM_HC_IOPORT_READ)
1875	Assert(!fIOWrite);
1876	else if (rc == VINF_IOM_HC_IOPORT_WRITE)
1877	Assert(fIOWrite);
1878	else
1879	AssertMsg(VBOX_FAILURE(rc) \|\| rc == VINF_EM_RAW_EMULATE_INSTR \|\| rc == VINF_EM_RAW_GUEST_TRAP \|\| rc == VINF_TRPM_XCPT_DISPATCHED, ("%Vrc\n", rc));
1880	#endif
1881	break;
1882	}
1883
1884	default:
1885	/* The rest is handled after syncing the entire CPU state. */
1886	break;
1887	}
1888
1889	/* Note: the guest state isn't entirely synced back at this stage. */
1890
1891	/* Investigate why there was a VM-exit. (part 2) */
1892	switch (exitReason)
1893	{
1894	case VMX_EXIT_EXCEPTION: /* 0 Exception or non-maskable interrupt (NMI). */
1895	case VMX_EXIT_EXTERNAL_IRQ: /* 1 External interrupt. */
1896	/* Already handled above. */
1897	break;
1898
1899	case VMX_EXIT_TRIPLE_FAULT: /* 2 Triple fault. */
1900	rc = VINF_EM_RESET; /* Triple fault equals a reset. */
1901	break;
1902
1903	case VMX_EXIT_INIT_SIGNAL: /* 3 INIT signal. */
1904	case VMX_EXIT_SIPI: /* 4 Start-up IPI (SIPI). */
1905	rc = VINF_EM_RAW_INTERRUPT;
1906	AssertFailed(); /* Can't happen. Yet. */
1907	break;
1908
1909	case VMX_EXIT_IO_SMI_IRQ: /* 5 I/O system-management interrupt (SMI). */
1910	case VMX_EXIT_SMI_IRQ: /* 6 Other SMI. */
1911	rc = VINF_EM_RAW_INTERRUPT;
1912	AssertFailed(); /* Can't happen afaik. */
1913	break;
1914
1915	case VMX_EXIT_TASK_SWITCH: /* 9 Task switch. */
1916	rc = VERR_EM_INTERPRETER;
1917	break;
1918
1919	case VMX_EXIT_HLT: /* 12 Guest software attempted to execute HLT. */
1920	/** Check if external interrupts are pending; if so, don't switch back. */
1921	if (VM_FF_ISPENDING(pVM, (VM_FF_INTERRUPT_APIC\|VM_FF_INTERRUPT_PIC)))
1922	{
1923	pCtx->eip++; /* skip hlt */
1924	goto ResumeExecution;
1925	}
1926
1927	rc = VINF_EM_RAW_EMULATE_INSTR_HLT;
1928	break;
1929
1930	case VMX_EXIT_RSM: /* 17 Guest software attempted to execute RSM in SMM. */
1931	AssertFailed(); /* can't happen. */
1932	rc = VINF_EM_RAW_EXCEPTION_PRIVILEGED;
1933	break;
1934
1935	case VMX_EXIT_VMCALL: /* 18 Guest software executed VMCALL. */
1936	case VMX_EXIT_VMCLEAR: /* 19 Guest software executed VMCLEAR. */
1937	case VMX_EXIT_VMLAUNCH: /* 20 Guest software executed VMLAUNCH. */
1938	case VMX_EXIT_VMPTRLD: /* 21 Guest software executed VMPTRLD. */
1939	case VMX_EXIT_VMPTRST: /* 22 Guest software executed VMPTRST. */
1940	case VMX_EXIT_VMREAD: /* 23 Guest software executed VMREAD. */
1941	case VMX_EXIT_VMRESUME: /* 24 Guest software executed VMRESUME. */
1942	case VMX_EXIT_VMWRITE: /* 25 Guest software executed VMWRITE. */
1943	case VMX_EXIT_VMXOFF: /* 26 Guest software executed VMXOFF. */
1944	case VMX_EXIT_VMXON: /* 27 Guest software executed VMXON. */
1945	/** @todo inject #UD immediately */
1946	rc = VINF_EM_RAW_EXCEPTION_PRIVILEGED;
1947	break;
1948
1949	case VMX_EXIT_CPUID: /* 10 Guest software attempted to execute CPUID. */
1950	case VMX_EXIT_RDTSC: /* 16 Guest software attempted to execute RDTSC. */
1951	case VMX_EXIT_INVPG: /* 14 Guest software attempted to execute INVPG. */
1952	case VMX_EXIT_CRX_MOVE: /* 28 Control-register accesses. */
1953	case VMX_EXIT_DRX_MOVE: /* 29 Debug-register accesses. */
1954	case VMX_EXIT_PORT_IO: /* 30 I/O instruction. */
1955	/* already handled above */
1956	AssertMsg( rc == VINF_PGM_CHANGE_MODE
1957	\|\| rc == VINF_EM_RAW_INTERRUPT
1958	\|\| rc == VERR_EM_INTERPRETER
1959	\|\| rc == VINF_EM_RAW_EMULATE_INSTR
1960	\|\| rc == VINF_PGM_SYNC_CR3
1961	\|\| rc == VINF_IOM_HC_IOPORT_READ
1962	\|\| rc == VINF_IOM_HC_IOPORT_WRITE
1963	\|\| rc == VINF_EM_RAW_GUEST_TRAP
1964	\|\| rc == VINF_TRPM_XCPT_DISPATCHED
1965	\|\| rc == VINF_EM_RESCHEDULE_REM,
1966	("rc = %d\n", rc));
1967	break;
1968
1969	case VMX_EXIT_RDPMC: /* 15 Guest software attempted to execute RDPMC. */
1970	case VMX_EXIT_RDMSR: /* 31 RDMSR. Guest software attempted to execute RDMSR. */
1971	case VMX_EXIT_WRMSR: /* 32 WRMSR. Guest software attempted to execute WRMSR. */
1972	case VMX_EXIT_MWAIT: /* 36 Guest software executed MWAIT. */
1973	case VMX_EXIT_MONITOR: /* 39 Guest software attempted to execute MONITOR. */
1974	case VMX_EXIT_PAUSE: /* 40 Guest software attempted to execute PAUSE. */
1975	rc = VINF_EM_RAW_EXCEPTION_PRIVILEGED;
1976	break;
1977
1978	case VMX_EXIT_IRQ_WINDOW: /* 7 Interrupt window. */
1979	Assert(rc == VINF_EM_RAW_INTERRUPT);
1980	break;
1981
1982	case VMX_EXIT_TPR: /* 43 TPR below threshold. Guest software executed MOV to CR8. */
1983	case VMX_EXIT_ERR_INVALID_GUEST_STATE: /* 33 VM-entry failure due to invalid guest state. */
1984	case VMX_EXIT_ERR_MSR_LOAD: /* 34 VM-entry failure due to MSR loading. */
1985	case VMX_EXIT_ERR_MACHINE_CHECK: /* 41 VM-entry failure due to machine-check. */
1986	default:
1987	rc = VERR_EM_INTERNAL_ERROR;
1988	AssertMsgFailed(("Unexpected exit code %d\n", exitReason)); /* Can't happen. */
1989	break;
1990
1991	}
1992	end:
1993	if (fGuestStateSynced)
1994	{
1995	/* Remaining guest CPU context: TR, IDTR, GDTR, LDTR. */
1996	VMX_READ_SELREG(LDTR, ldtr);
1997	VMX_READ_SELREG(TR, tr);
1998
1999	VMXReadVMCS(VMX_VMCS_GUEST_GDTR_LIMIT, &val);
2000	pCtx->gdtr.cbGdt = val;
2001	VMXReadVMCS(VMX_VMCS_GUEST_GDTR_BASE, &val);
2002	pCtx->gdtr.pGdt = val;
2003
2004	VMXReadVMCS(VMX_VMCS_GUEST_IDTR_LIMIT, &val);
2005	pCtx->idtr.cbIdt = val;
2006	VMXReadVMCS(VMX_VMCS_GUEST_IDTR_BASE, &val);
2007	pCtx->idtr.pIdt = val;
2008
2009	/*
2010	* System MSRs
2011	*/
2012	VMXReadVMCS(VMX_VMCS_GUEST_SYSENTER_CS, &val);
2013	pCtx->SysEnter.cs = val;
2014	VMXReadVMCS(VMX_VMCS_GUEST_SYSENTER_EIP, &val);
2015	pCtx->SysEnter.eip = val;
2016	VMXReadVMCS(VMX_VMCS_GUEST_SYSENTER_ESP, &val);
2017	pCtx->SysEnter.esp = val;
2018	}
2019
2020	/* Signal changes for the recompiler. */
2021	CPUMSetChangedFlags(pVM, CPUM_CHANGED_SYSENTER_MSR \| CPUM_CHANGED_LDTR \| CPUM_CHANGED_GDTR \| CPUM_CHANGED_IDTR \| CPUM_CHANGED_TR \| CPUM_CHANGED_HIDDEN_SEL_REGS);
2022
2023	/* If we executed vmlaunch/vmresume and an external irq was pending, then we don't have to do a full sync the next time. */
2024	if ( exitReason == VMX_EXIT_EXTERNAL_IRQ
2025	&& !VMX_EXIT_INTERRUPTION_INFO_VALID(intInfo))
2026	{
2027	STAM_COUNTER_INC(&pVM->hwaccm.s.StatPendingHostIrq);
2028	/* On the next entry we'll only sync the host context. */
2029	pVM->hwaccm.s.fContextUseFlags \|= HWACCM_CHANGED_HOST_CONTEXT;
2030	}
2031	else
2032	{
2033	/* On the next entry we'll sync everything. */
2034	/** @todo we can do better than this */
2035	pVM->hwaccm.s.fContextUseFlags \|= HWACCM_CHANGED_ALL;
2036	}
2037
2038	/* translate into a less severe return code */
2039	if (rc == VERR_EM_INTERPRETER)
2040	rc = VINF_EM_RAW_EMULATE_INSTR;
2041
2042	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
2043	Log2(("X"));
2044	return rc;
2045	}
2046
2047
2048	/**
2049	* Enters the VT-x session
2050	*
2051	* @returns VBox status code.
2052	* @param pVM The VM to operate on.
2053	*/
2054	HWACCMR0DECL(int) VMXR0Enter(PVM pVM)
2055	{
2056	Assert(pVM->hwaccm.s.vmx.fSupported);
2057
2058	unsigned cr4 = ASMGetCR4();
2059	if (!(cr4 & X86_CR4_VMXE))
2060	{
2061	AssertMsgFailed(("X86_CR4_VMXE should be set!\n"));
2062	return VERR_VMX_X86_CR4_VMXE_CLEARED;
2063	}
2064
2065	/* Activate the VM Control Structure. */
2066	int rc = VMXActivateVMCS(pVM->hwaccm.s.vmx.pVMCSPhys);
2067	if (VBOX_FAILURE(rc))
2068	return rc;
2069
2070	pVM->hwaccm.s.vmx.fResumeVM = false;
2071	return VINF_SUCCESS;
2072	}
2073
2074
2075	/**
2076	* Leaves the VT-x session
2077	*
2078	* @returns VBox status code.
2079	* @param pVM The VM to operate on.
2080	*/
2081	HWACCMR0DECL(int) VMXR0Leave(PVM pVM)
2082	{
2083	Assert(pVM->hwaccm.s.vmx.fSupported);
2084
2085	/* Clear VM Control Structure. Marking it inactive, clearing implementation specific data and writing back VMCS data to memory. */
2086	int rc = VMXClearVMCS(pVM->hwaccm.s.vmx.pVMCSPhys);
2087	AssertRC(rc);
2088
2089	return VINF_SUCCESS;
2090	}
2091

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

source: vbox/trunk/src/VBox/VMM/VMMR0/HWVMXR0.cpp@ 8973

Download in other formats: