HWSVMR0.cpp@ 10458

Last change on this file since 10458 was 10458, checked in by vboxsync, 16 years ago
TPR & interrupt dispatch updates.
Property svn:eol-style set to `native` Property svn:keywords set to `Id`
File size: 82.0 KB

Line
1	/* $Id: HWSVMR0.cpp 10458 2008-07-10 08:08:05Z vboxsync $ */
2	/** @file
3	* HWACCM SVM - 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/hwacc_svm.h>
32	#include <VBox/pgm.h>
33	#include <VBox/pdm.h>
34	#include <VBox/err.h>
35	#include <VBox/log.h>
36	#include <VBox/selm.h>
37	#include <VBox/iom.h>
38	#include <VBox/dis.h>
39	#include <VBox/dbgf.h>
40	#include <VBox/disopcode.h>
41	#include <iprt/param.h>
42	#include <iprt/assert.h>
43	#include <iprt/asm.h>
44	#include <iprt/cpuset.h>
45	#include <iprt/mp.h>
46	#include "HWSVMR0.h"
47
48	static int SVMR0InterpretInvpg(PVM pVM, PCPUMCTXCORE pRegFrame, uint32_t uASID);
49
50	/**
51	* Sets up and activates AMD-V on the current CPU
52	*
53	* @returns VBox status code.
54	* @param pCpu CPU info struct
55	* @param pVM The VM to operate on.
56	* @param pvPageCpu Pointer to the global cpu page
57	* @param pPageCpuPhys Physical address of the global cpu page
58	*/
59	HWACCMR0DECL(int) SVMR0EnableCpu(PHWACCM_CPUINFO pCpu, PVM pVM, void *pvPageCpu, RTHCPHYS pPageCpuPhys)
60	{
61	AssertReturn(pPageCpuPhys, VERR_INVALID_PARAMETER);
62	AssertReturn(pVM, VERR_INVALID_PARAMETER);
63	AssertReturn(pvPageCpu, VERR_INVALID_PARAMETER);
64
65	/* We must turn on AMD-V and setup the host state physical address, as those MSRs are per-cpu/core. */
66
67	#ifdef LOG_ENABLED
68	SUPR0Printf("SVMR0EnableCpu cpu %d page (%x) %x\n", pCpu->idCpu, pvPageCpu, (uint32_t)pPageCpuPhys);
69	#endif
70
71	/* Turn on AMD-V in the EFER MSR. */
72	uint64_t val = ASMRdMsr(MSR_K6_EFER);
73	if (!(val & MSR_K6_EFER_SVME))
74	ASMWrMsr(MSR_K6_EFER, val \| MSR_K6_EFER_SVME);
75
76	/* Write the physical page address where the CPU will store the host state while executing the VM. */
77	ASMWrMsr(MSR_K8_VM_HSAVE_PA, pPageCpuPhys);
78
79	pCpu->uCurrentASID = 0; /* we'll aways increment this the first time (host uses ASID 0) */
80	pCpu->cTLBFlushes = 0;
81	return VINF_SUCCESS;
82	}
83
84	/**
85	* Deactivates AMD-V on the current CPU
86	*
87	* @returns VBox status code.
88	* @param pCpu CPU info struct
89	* @param pvPageCpu Pointer to the global cpu page
90	* @param pPageCpuPhys Physical address of the global cpu page
91	*/
92	HWACCMR0DECL(int) SVMR0DisableCpu(PHWACCM_CPUINFO pCpu, void *pvPageCpu, RTHCPHYS pPageCpuPhys)
93	{
94	AssertReturn(pPageCpuPhys, VERR_INVALID_PARAMETER);
95	AssertReturn(pvPageCpu, VERR_INVALID_PARAMETER);
96
97	#ifdef LOG_ENABLED
98	SUPR0Printf("SVMR0DisableCpu cpu %d\n", pCpu->idCpu);
99	#endif
100
101	/* Turn off AMD-V in the EFER MSR. */
102	uint64_t val = ASMRdMsr(MSR_K6_EFER);
103	ASMWrMsr(MSR_K6_EFER, val & ~MSR_K6_EFER_SVME);
104
105	/* Invalidate host state physical address. */
106	ASMWrMsr(MSR_K8_VM_HSAVE_PA, 0);
107	pCpu->uCurrentASID = 0;
108
109	return VINF_SUCCESS;
110	}
111
112	/**
113	* Does Ring-0 per VM AMD-V init.
114	*
115	* @returns VBox status code.
116	* @param pVM The VM to operate on.
117	*/
118	HWACCMR0DECL(int) SVMR0InitVM(PVM pVM)
119	{
120	int rc;
121
122	/* Allocate one page for the VM control block (VMCB). */
123	rc = RTR0MemObjAllocCont(&pVM->hwaccm.s.svm.pMemObjVMCB, 1 << PAGE_SHIFT, true /* executable R0 mapping */);
124	if (RT_FAILURE(rc))
125	return rc;
126
127	pVM->hwaccm.s.svm.pVMCB = RTR0MemObjAddress(pVM->hwaccm.s.svm.pMemObjVMCB);
128	pVM->hwaccm.s.svm.pVMCBPhys = RTR0MemObjGetPagePhysAddr(pVM->hwaccm.s.svm.pMemObjVMCB, 0);
129	ASMMemZero32(pVM->hwaccm.s.svm.pVMCB, PAGE_SIZE);
130
131	/* Allocate one page for the host context */
132	rc = RTR0MemObjAllocCont(&pVM->hwaccm.s.svm.pMemObjVMCBHost, 1 << PAGE_SHIFT, true /* executable R0 mapping */);
133	if (RT_FAILURE(rc))
134	return rc;
135
136	pVM->hwaccm.s.svm.pVMCBHost = RTR0MemObjAddress(pVM->hwaccm.s.svm.pMemObjVMCBHost);
137	pVM->hwaccm.s.svm.pVMCBHostPhys = RTR0MemObjGetPagePhysAddr(pVM->hwaccm.s.svm.pMemObjVMCBHost, 0);
138	ASMMemZero32(pVM->hwaccm.s.svm.pVMCBHost, PAGE_SIZE);
139
140	/* Allocate 12 KB for the IO bitmap (doesn't seem to be a way to convince SVM not to use it) */
141	rc = RTR0MemObjAllocCont(&pVM->hwaccm.s.svm.pMemObjIOBitmap, 3 << PAGE_SHIFT, true /* executable R0 mapping */);
142	if (RT_FAILURE(rc))
143	return rc;
144
145	pVM->hwaccm.s.svm.pIOBitmap = RTR0MemObjAddress(pVM->hwaccm.s.svm.pMemObjIOBitmap);
146	pVM->hwaccm.s.svm.pIOBitmapPhys = RTR0MemObjGetPagePhysAddr(pVM->hwaccm.s.svm.pMemObjIOBitmap, 0);
147	/* Set all bits to intercept all IO accesses. */
148	ASMMemFill32(pVM->hwaccm.s.svm.pIOBitmap, PAGE_SIZE*3, 0xffffffff);
149
150	/* Allocate 8 KB for the MSR bitmap (doesn't seem to be a way to convince SVM not to use it) */
151	rc = RTR0MemObjAllocCont(&pVM->hwaccm.s.svm.pMemObjMSRBitmap, 2 << PAGE_SHIFT, true /* executable R0 mapping */);
152	if (RT_FAILURE(rc))
153	return rc;
154
155	pVM->hwaccm.s.svm.pMSRBitmap = RTR0MemObjAddress(pVM->hwaccm.s.svm.pMemObjMSRBitmap);
156	pVM->hwaccm.s.svm.pMSRBitmapPhys = RTR0MemObjGetPagePhysAddr(pVM->hwaccm.s.svm.pMemObjMSRBitmap, 0);
157	/* Set all bits to intercept all MSR accesses. */
158	ASMMemFill32(pVM->hwaccm.s.svm.pMSRBitmap, PAGE_SIZE*2, 0xffffffff);
159
160	/* Erratum 170 which requires a forced TLB flush for each world switch:
161	* See http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/33610.pdf
162	*
163	* All BH-G1/2 and DH-G1/2 models include a fix:
164	* Athlon X2: 0x6b 1/2
165	* 0x68 1/2
166	* Athlon 64: 0x7f 1
167	* 0x6f 2
168	* Sempron: 0x7f 1/2
169	* 0x6f 2
170	* 0x6c 2
171	* 0x7c 2
172	* Turion 64: 0x68 2
173	*
174	*/
175	uint32_t u32Dummy;
176	uint32_t u32Version, u32Family, u32Model, u32Stepping, u32BaseFamily;
177	ASMCpuId(1, &u32Version, &u32Dummy, &u32Dummy, &u32Dummy);
178	u32BaseFamily= (u32Version >> 8) & 0xf;
179	u32Family = u32BaseFamily + (u32BaseFamily == 0xf ? ((u32Version >> 20) & 0x7f) : 0);
180	u32Model = ((u32Version >> 4) & 0xf);
181	u32Model = u32Model \| ((u32BaseFamily == 0xf ? (u32Version >> 16) & 0x0f : 0) << 4);
182	u32Stepping = u32Version & 0xf;
183	if ( u32Family == 0xf
184	&& !((u32Model == 0x68 \|\| u32Model == 0x6b \|\| u32Model == 0x7f) && u32Stepping >= 1)
185	&& !((u32Model == 0x6f \|\| u32Model == 0x6c \|\| u32Model == 0x7c) && u32Stepping >= 2))
186	{
187	Log(("SVMR0InitVM: AMD cpu with erratum 170 family %x model %x stepping %x\n", u32Family, u32Model, u32Stepping));
188	pVM->hwaccm.s.svm.fAlwaysFlushTLB = true;
189	}
190
191	/* Invalidate the last cpu we were running on. */
192	pVM->hwaccm.s.svm.idLastCpu = NIL_RTCPUID;
193	return VINF_SUCCESS;
194	}
195
196	/**
197	* Does Ring-0 per VM AMD-V termination.
198	*
199	* @returns VBox status code.
200	* @param pVM The VM to operate on.
201	*/
202	HWACCMR0DECL(int) SVMR0TermVM(PVM pVM)
203	{
204	if (pVM->hwaccm.s.svm.pMemObjVMCB)
205	{
206	RTR0MemObjFree(pVM->hwaccm.s.svm.pMemObjVMCB, false);
207	pVM->hwaccm.s.svm.pVMCB = 0;
208	pVM->hwaccm.s.svm.pVMCBPhys = 0;
209	pVM->hwaccm.s.svm.pMemObjVMCB = 0;
210	}
211	if (pVM->hwaccm.s.svm.pMemObjVMCBHost)
212	{
213	RTR0MemObjFree(pVM->hwaccm.s.svm.pMemObjVMCBHost, false);
214	pVM->hwaccm.s.svm.pVMCBHost = 0;
215	pVM->hwaccm.s.svm.pVMCBHostPhys = 0;
216	pVM->hwaccm.s.svm.pMemObjVMCBHost = 0;
217	}
218	if (pVM->hwaccm.s.svm.pMemObjIOBitmap)
219	{
220	RTR0MemObjFree(pVM->hwaccm.s.svm.pMemObjIOBitmap, false);
221	pVM->hwaccm.s.svm.pIOBitmap = 0;
222	pVM->hwaccm.s.svm.pIOBitmapPhys = 0;
223	pVM->hwaccm.s.svm.pMemObjIOBitmap = 0;
224	}
225	if (pVM->hwaccm.s.svm.pMemObjMSRBitmap)
226	{
227	RTR0MemObjFree(pVM->hwaccm.s.svm.pMemObjMSRBitmap, false);
228	pVM->hwaccm.s.svm.pMSRBitmap = 0;
229	pVM->hwaccm.s.svm.pMSRBitmapPhys = 0;
230	pVM->hwaccm.s.svm.pMemObjMSRBitmap = 0;
231	}
232	return VINF_SUCCESS;
233	}
234
235	/**
236	* Sets up AMD-V for the specified VM
237	*
238	* @returns VBox status code.
239	* @param pVM The VM to operate on.
240	*/
241	HWACCMR0DECL(int) SVMR0SetupVM(PVM pVM)
242	{
243	int rc = VINF_SUCCESS;
244	SVM_VMCB *pVMCB;
245
246	AssertReturn(pVM, VERR_INVALID_PARAMETER);
247
248	Assert(pVM->hwaccm.s.svm.fSupported);
249
250	pVMCB = (SVM_VMCB *)pVM->hwaccm.s.svm.pVMCB;
251	AssertMsgReturn(pVMCB, ("Invalid pVMCB\n"), VERR_EM_INTERNAL_ERROR);
252
253	/* Program the control fields. Most of them never have to be changed again. */
254	/* CR0/3/4 reads must be intercepted, our shadow values are not necessarily the same as the guest's. */
255	/** @note CR0 & CR4 can be safely read when guest and shadow copies are identical. */
256	if (!pVM->hwaccm.s.fNestedPaging)
257	pVMCB->ctrl.u16InterceptRdCRx = RT_BIT(0) \| RT_BIT(3) \| RT_BIT(4);
258	else
259	pVMCB->ctrl.u16InterceptRdCRx = RT_BIT(0) \| RT_BIT(4);
260
261	/*
262	* CR0/3/4 writes must be intercepted for obvious reasons.
263	*/
264	if (!pVM->hwaccm.s.fNestedPaging)
265	pVMCB->ctrl.u16InterceptWrCRx = RT_BIT(0) \| RT_BIT(3) \| RT_BIT(4);
266	else
267	pVMCB->ctrl.u16InterceptWrCRx = RT_BIT(0) \| RT_BIT(4);
268
269	/* Intercept all DRx reads and writes. */
270	pVMCB->ctrl.u16InterceptRdDRx = RT_BIT(0) \| RT_BIT(1) \| RT_BIT(2) \| RT_BIT(3) \| RT_BIT(4) \| RT_BIT(5) \| RT_BIT(6) \| RT_BIT(7);
271	pVMCB->ctrl.u16InterceptWrDRx = RT_BIT(0) \| RT_BIT(1) \| RT_BIT(2) \| RT_BIT(3) \| RT_BIT(4) \| RT_BIT(5) \| RT_BIT(6) \| RT_BIT(7);
272
273	/* Currently we don't care about DRx reads or writes. DRx registers are trashed.
274	* All breakpoints are automatically cleared when the VM exits.
275	*/
276
277	pVMCB->ctrl.u32InterceptException = HWACCM_SVM_TRAP_MASK;
278	#ifndef DEBUG
279	if (pVM->hwaccm.s.fNestedPaging)
280	pVMCB->ctrl.u32InterceptException &= ~RT_BIT(14); /* no longer need to intercept #PF. */
281	#endif
282
283	pVMCB->ctrl.u32InterceptCtrl1 = SVM_CTRL1_INTERCEPT_INTR
284	\| SVM_CTRL1_INTERCEPT_VINTR
285	\| SVM_CTRL1_INTERCEPT_NMI
286	\| SVM_CTRL1_INTERCEPT_SMI
287	\| SVM_CTRL1_INTERCEPT_INIT
288	\| SVM_CTRL1_INTERCEPT_RDPMC
289	\| SVM_CTRL1_INTERCEPT_CPUID
290	\| SVM_CTRL1_INTERCEPT_RSM
291	\| SVM_CTRL1_INTERCEPT_HLT
292	\| SVM_CTRL1_INTERCEPT_INOUT_BITMAP
293	\| SVM_CTRL1_INTERCEPT_MSR_SHADOW
294	\| SVM_CTRL1_INTERCEPT_INVLPG
295	\| SVM_CTRL1_INTERCEPT_INVLPGA /* AMD only */
296	\| SVM_CTRL1_INTERCEPT_TASK_SWITCH
297	\| SVM_CTRL1_INTERCEPT_SHUTDOWN /* fatal */
298	\| SVM_CTRL1_INTERCEPT_FERR_FREEZE; /* Legacy FPU FERR handling. */
299	;
300	/* With nested paging we don't care about invlpg anymore. */
301	if (pVM->hwaccm.s.fNestedPaging)
302	pVMCB->ctrl.u32InterceptCtrl1 &= ~SVM_CTRL1_INTERCEPT_INVLPG;
303
304	pVMCB->ctrl.u32InterceptCtrl2 = SVM_CTRL2_INTERCEPT_VMRUN /* required */
305	\| SVM_CTRL2_INTERCEPT_VMMCALL
306	\| SVM_CTRL2_INTERCEPT_VMLOAD
307	\| SVM_CTRL2_INTERCEPT_VMSAVE
308	\| SVM_CTRL2_INTERCEPT_STGI
309	\| SVM_CTRL2_INTERCEPT_CLGI
310	\| SVM_CTRL2_INTERCEPT_SKINIT
311	\| SVM_CTRL2_INTERCEPT_RDTSCP /* AMD only; we don't support this one */
312	\| SVM_CTRL2_INTERCEPT_WBINVD
313	\| SVM_CTRL2_INTERCEPT_MWAIT_UNCOND; /* don't execute mwait or else we'll idle inside the guest (host thinks the cpu load is high) */
314	;
315	Log(("pVMCB->ctrl.u32InterceptException = %x\n", pVMCB->ctrl.u32InterceptException));
316	Log(("pVMCB->ctrl.u32InterceptCtrl1 = %x\n", pVMCB->ctrl.u32InterceptCtrl1));
317	Log(("pVMCB->ctrl.u32InterceptCtrl2 = %x\n", pVMCB->ctrl.u32InterceptCtrl2));
318
319	/* Virtualize masking of INTR interrupts. (reads/writes from/to CR8 go to the V_TPR register) */
320	pVMCB->ctrl.IntCtrl.n.u1VIrqMasking = 1;
321	/* Ignore the priority in the TPR; just deliver it when we tell it to. */
322	pVMCB->ctrl.IntCtrl.n.u1IgnoreTPR = 1;
323
324	/* Set IO and MSR bitmap addresses. */
325	pVMCB->ctrl.u64IOPMPhysAddr = pVM->hwaccm.s.svm.pIOBitmapPhys;
326	pVMCB->ctrl.u64MSRPMPhysAddr = pVM->hwaccm.s.svm.pMSRBitmapPhys;
327
328	/* No LBR virtualization. */
329	pVMCB->ctrl.u64LBRVirt = 0;
330
331	/** The ASID must start at 1; the host uses 0. */
332	pVMCB->ctrl.TLBCtrl.n.u32ASID = 1;
333
334	/** Setup the PAT msr (nested paging only) */
335	pVMCB->guest.u64GPAT = 0x0007040600070406ULL;
336	return rc;
337	}
338
339
340	/**
341	* Injects an event (trap or external interrupt)
342	*
343	* @param pVM The VM to operate on.
344	* @param pVMCB SVM control block
345	* @param pCtx CPU Context
346	* @param pIntInfo SVM interrupt info
347	*/
348	inline void SVMR0InjectEvent(PVM pVM, SVM_VMCB pVMCB, CPUMCTX pCtx, SVM_EVENT* pEvent)
349	{
350	#ifdef VBOX_STRICT
351	if (pEvent->n.u8Vector == 0xE)
352	Log(("SVM: Inject int %d at %VGv error code=%02x CR2=%VGv intInfo=%08x\n", pEvent->n.u8Vector, pCtx->rip, pEvent->n.u32ErrorCode, pCtx->cr2, pEvent->au64[0]));
353	else
354	if (pEvent->n.u8Vector < 0x20)
355	Log(("SVM: Inject int %d at %VGv error code=%08x\n", pEvent->n.u8Vector, pCtx->rip, pEvent->n.u32ErrorCode));
356	else
357	{
358	Log(("INJ-EI: %x at %VGv\n", pEvent->n.u8Vector, pCtx->rip));
359	Assert(!VM_FF_ISSET(pVM, VM_FF_INHIBIT_INTERRUPTS));
360	Assert(pCtx->eflags.u32 & X86_EFL_IF);
361	}
362	#endif
363
364	/* Set event injection state. */
365	pVMCB->ctrl.EventInject.au64[0] = pEvent->au64[0];
366	}
367
368
369	/**
370	* Checks for pending guest interrupts and injects them
371	*
372	* @returns VBox status code.
373	* @param pVM The VM to operate on.
374	* @param pVMCB SVM control block
375	* @param pCtx CPU Context
376	*/
377	static int SVMR0CheckPendingInterrupt(PVM pVM, SVM_VMCB pVMCB, CPUMCTX pCtx)
378	{
379	int rc;
380
381	/* Dispatch any pending interrupts. (injected before, but a VM exit occurred prematurely) */
382	if (pVM->hwaccm.s.Event.fPending)
383	{
384	SVM_EVENT Event;
385
386	Log(("Reinjecting event %08x %08x at %VGv\n", pVM->hwaccm.s.Event.intInfo, pVM->hwaccm.s.Event.errCode, pCtx->rip));
387	STAM_COUNTER_INC(&pVM->hwaccm.s.StatIntReinject);
388	Event.au64[0] = pVM->hwaccm.s.Event.intInfo;
389	SVMR0InjectEvent(pVM, pVMCB, pCtx, &Event);
390
391	pVM->hwaccm.s.Event.fPending = false;
392	return VINF_SUCCESS;
393	}
394
395	/* When external interrupts are pending, we should exit the VM when IF is set. */
396	if ( !TRPMHasTrap(pVM)
397	&& VM_FF_ISPENDING(pVM, (VM_FF_INTERRUPT_APIC\|VM_FF_INTERRUPT_PIC)))
398	{
399	if ( !(pCtx->eflags.u32 & X86_EFL_IF)
400	\|\| VM_FF_ISSET(pVM, VM_FF_INHIBIT_INTERRUPTS))
401	{
402	if (!pVMCB->ctrl.IntCtrl.n.u1VIrqValid)
403	{
404	if (!VM_FF_ISSET(pVM, VM_FF_INHIBIT_INTERRUPTS))
405	LogFlow(("Enable irq window exit!\n"));
406	else
407	Log(("Pending interrupt blocked at %VGv by VM_FF_INHIBIT_INTERRUPTS -> irq window exit\n", pCtx->rip));
408
409	/** @todo use virtual interrupt method to inject a pending irq; dispatched as soon as guest.IF is set. */
410	pVMCB->ctrl.u32InterceptCtrl1 \|= SVM_CTRL1_INTERCEPT_VINTR;
411	pVMCB->ctrl.IntCtrl.n.u1VIrqValid = 1;
412	pVMCB->ctrl.IntCtrl.n.u8VIrqVector = 0; /* don't care */
413	}
414	}
415	else
416	{
417	uint8_t u8Interrupt;
418
419	rc = PDMGetInterrupt(pVM, &u8Interrupt);
420	Log(("Dispatch interrupt: u8Interrupt=%x (%d) rc=%Vrc\n", u8Interrupt, u8Interrupt, rc));
421	if (VBOX_SUCCESS(rc))
422	{
423	rc = TRPMAssertTrap(pVM, u8Interrupt, TRPM_HARDWARE_INT);
424	AssertRC(rc);
425	}
426	else
427	{
428	/* Can only happen in rare cases where a pending interrupt is cleared behind our back */
429	Assert(!VM_FF_ISPENDING(pVM, (VM_FF_INTERRUPT_APIC\|VM_FF_INTERRUPT_PIC)));
430	STAM_COUNTER_INC(&pVM->hwaccm.s.StatSwitchGuestIrq);
431	/* Just continue */
432	}
433	}
434	}
435
436	#ifdef VBOX_STRICT
437	if (TRPMHasTrap(pVM))
438	{
439	uint8_t u8Vector;
440	rc = TRPMQueryTrapAll(pVM, &u8Vector, 0, 0, 0);
441	AssertRC(rc);
442	}
443	#endif
444
445	if ( pCtx->eflags.u32 & X86_EFL_IF
446	&& (!VM_FF_ISSET(pVM, VM_FF_INHIBIT_INTERRUPTS))
447	&& TRPMHasTrap(pVM)
448	)
449	{
450	uint8_t u8Vector;
451	int rc;
452	TRPMEVENT enmType;
453	SVM_EVENT Event;
454	RTGCUINT u32ErrorCode;
455
456	Event.au64[0] = 0;
457
458	/* If a new event is pending, then dispatch it now. */
459	rc = TRPMQueryTrapAll(pVM, &u8Vector, &enmType, &u32ErrorCode, 0);
460	AssertRC(rc);
461	Assert(pCtx->eflags.Bits.u1IF == 1 \|\| enmType == TRPM_TRAP);
462	Assert(enmType != TRPM_SOFTWARE_INT);
463
464	/* Clear the pending trap. */
465	rc = TRPMResetTrap(pVM);
466	AssertRC(rc);
467
468	Event.n.u8Vector = u8Vector;
469	Event.n.u1Valid = 1;
470	Event.n.u32ErrorCode = u32ErrorCode;
471
472	if (enmType == TRPM_TRAP)
473	{
474	switch (u8Vector) {
475	case 8:
476	case 10:
477	case 11:
478	case 12:
479	case 13:
480	case 14:
481	case 17:
482	/* Valid error codes. */
483	Event.n.u1ErrorCodeValid = 1;
484	break;
485	default:
486	break;
487	}
488	if (u8Vector == X86_XCPT_NMI)
489	Event.n.u3Type = SVM_EVENT_NMI;
490	else
491	Event.n.u3Type = SVM_EVENT_EXCEPTION;
492	}
493	else
494	Event.n.u3Type = SVM_EVENT_EXTERNAL_IRQ;
495
496	STAM_COUNTER_INC(&pVM->hwaccm.s.StatIntInject);
497	SVMR0InjectEvent(pVM, pVMCB, pCtx, &Event);
498	} /* if (interrupts can be dispatched) */
499
500	return VINF_SUCCESS;
501	}
502
503	/**
504	* Save the host state
505	*
506	* @returns VBox status code.
507	* @param pVM The VM to operate on.
508	*/
509	HWACCMR0DECL(int) SVMR0SaveHostState(PVM pVM)
510	{
511	/* Nothing to do here. */
512	return VINF_SUCCESS;
513	}
514
515	/**
516	* Loads the guest state
517	*
518	* NOTE: Don't do anything here that can cause a jump back to ring 3!!!!!
519	*
520	* @returns VBox status code.
521	* @param pVM The VM to operate on.
522	* @param pCtx Guest context
523	*/
524	HWACCMR0DECL(int) SVMR0LoadGuestState(PVM pVM, CPUMCTX *pCtx)
525	{
526	RTGCUINTPTR val;
527	SVM_VMCB *pVMCB;
528
529	if (pVM == NULL)
530	return VERR_INVALID_PARAMETER;
531
532	/* Setup AMD SVM. */
533	Assert(pVM->hwaccm.s.svm.fSupported);
534
535	pVMCB = (SVM_VMCB *)pVM->hwaccm.s.svm.pVMCB;
536	AssertMsgReturn(pVMCB, ("Invalid pVMCB\n"), VERR_EM_INTERNAL_ERROR);
537
538	/* Guest CPU context: ES, CS, SS, DS, FS, GS. */
539	if (pVM->hwaccm.s.fContextUseFlags & HWACCM_CHANGED_GUEST_SEGMENT_REGS)
540	{
541	SVM_WRITE_SELREG(CS, cs);
542	SVM_WRITE_SELREG(SS, ss);
543	SVM_WRITE_SELREG(DS, ds);
544	SVM_WRITE_SELREG(ES, es);
545	SVM_WRITE_SELREG(FS, fs);
546	SVM_WRITE_SELREG(GS, gs);
547	}
548
549	/* Guest CPU context: LDTR. */
550	if (pVM->hwaccm.s.fContextUseFlags & HWACCM_CHANGED_GUEST_LDTR)
551	{
552	SVM_WRITE_SELREG(LDTR, ldtr);
553	}
554
555	/* Guest CPU context: TR. */
556	if (pVM->hwaccm.s.fContextUseFlags & HWACCM_CHANGED_GUEST_TR)
557	{
558	SVM_WRITE_SELREG(TR, tr);
559	}
560
561	/* Guest CPU context: GDTR. */
562	if (pVM->hwaccm.s.fContextUseFlags & HWACCM_CHANGED_GUEST_GDTR)
563	{
564	pVMCB->guest.GDTR.u32Limit = pCtx->gdtr.cbGdt;
565	pVMCB->guest.GDTR.u64Base = pCtx->gdtr.pGdt;
566	}
567
568	/* Guest CPU context: IDTR. */
569	if (pVM->hwaccm.s.fContextUseFlags & HWACCM_CHANGED_GUEST_IDTR)
570	{
571	pVMCB->guest.IDTR.u32Limit = pCtx->idtr.cbIdt;
572	pVMCB->guest.IDTR.u64Base = pCtx->idtr.pIdt;
573	}
574
575	/*
576	* Sysenter MSRs (unconditional)
577	*/
578	pVMCB->guest.u64SysEnterCS = pCtx->SysEnter.cs;
579	pVMCB->guest.u64SysEnterEIP = pCtx->SysEnter.eip;
580	pVMCB->guest.u64SysEnterESP = pCtx->SysEnter.esp;
581
582	/* Control registers */
583	if (pVM->hwaccm.s.fContextUseFlags & HWACCM_CHANGED_GUEST_CR0)
584	{
585	val = pCtx->cr0;
586	if (!CPUMIsGuestFPUStateActive(pVM))
587	{
588	/* Always use #NM exceptions to load the FPU/XMM state on demand. */
589	val \|= X86_CR0_TS \| X86_CR0_ET \| X86_CR0_NE \| X86_CR0_MP;
590	}
591	else
592	{
593	/** @todo check if we support the old style mess correctly. */
594	if (!(val & X86_CR0_NE))
595	{
596	Log(("Forcing X86_CR0_NE!!!\n"));
597
598	/* Also catch floating point exceptions as we need to report them to the guest in a different way. */
599	if (!pVM->hwaccm.s.fFPUOldStyleOverride)
600	{
601	pVMCB->ctrl.u32InterceptException \|= RT_BIT(16);
602	pVM->hwaccm.s.fFPUOldStyleOverride = true;
603	}
604	}
605	val \|= X86_CR0_NE; /* always turn on the native mechanism to report FPU errors (old style uses interrupts) */
606	}
607	/* Always enable caching. */
608	val &= ~(X86_CR0_CD\|X86_CR0_NW);
609
610	/* Note: WP is not relevant in nested paging mode as we catch accesses on the (guest) physical level. */
611	/* Note: In nested paging mode the guest is allowed to run with paging disabled; the guest physical to host physical translation will remain active. */
612	if (!pVM->hwaccm.s.fNestedPaging)
613	{
614	val \|= X86_CR0_PG; /* Paging is always enabled; even when the guest is running in real mode or PE without paging. */
615	val \|= X86_CR0_WP; /* Must set this as we rely on protect various pages and supervisor writes must be caught. */
616	}
617	pVMCB->guest.u64CR0 = val;
618	}
619	/* CR2 as well */
620	pVMCB->guest.u64CR2 = pCtx->cr2;
621
622	if (pVM->hwaccm.s.fContextUseFlags & HWACCM_CHANGED_GUEST_CR3)
623	{
624	/* Save our shadow CR3 register. */
625	if (pVM->hwaccm.s.fNestedPaging)
626	{
627	pVMCB->ctrl.u64NestedPagingCR3 = PGMGetNestedCR3(pVM, PGMGetHostMode(pVM));
628	Assert(pVMCB->ctrl.u64NestedPagingCR3);
629	pVMCB->guest.u64CR3 = pCtx->cr3;
630	}
631	else
632	{
633	pVMCB->guest.u64CR3 = PGMGetHyperCR3(pVM);
634	Assert(pVMCB->guest.u64CR3);
635	}
636	}
637
638	if (pVM->hwaccm.s.fContextUseFlags & HWACCM_CHANGED_GUEST_CR4)
639	{
640	val = pCtx->cr4;
641	if (!pVM->hwaccm.s.fNestedPaging)
642	{
643	switch(pVM->hwaccm.s.enmShadowMode)
644	{
645	case PGMMODE_REAL:
646	case PGMMODE_PROTECTED: /* Protected mode, no paging. */
647	AssertFailed();
648	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
649
650	case PGMMODE_32_BIT: /* 32-bit paging. */
651	break;
652
653	case PGMMODE_PAE: /* PAE paging. */
654	case PGMMODE_PAE_NX: /* PAE paging with NX enabled. */
655	/** @todo use normal 32 bits paging */
656	val \|= X86_CR4_PAE;
657	break;
658
659	case PGMMODE_AMD64: /* 64-bit AMD paging (long mode). */
660	case PGMMODE_AMD64_NX: /* 64-bit AMD paging (long mode) with NX enabled. */
661	#ifdef VBOX_ENABLE_64_BITS_GUESTS
662	break;
663	#else
664	AssertFailed();
665	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
666	#endif
667
668	default: /* shut up gcc */
669	AssertFailed();
670	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
671	}
672	}
673	pVMCB->guest.u64CR4 = val;
674	}
675
676	/* Debug registers. */
677	if (pVM->hwaccm.s.fContextUseFlags & HWACCM_CHANGED_GUEST_DEBUG)
678	{
679	/** @todo DR0-6 */
680	val = pCtx->dr7;
681	val &= ~(RT_BIT(11) \| RT_BIT(12) \| RT_BIT(14) \| RT_BIT(15)); /* must be zero */
682	val \|= 0x400; /* must be one */
683	#ifdef VBOX_STRICT
684	val = 0x400;
685	#endif
686	pVMCB->guest.u64DR7 = val;
687
688	pVMCB->guest.u64DR6 = pCtx->dr6;
689	}
690
691	/* EIP, ESP and EFLAGS */
692	pVMCB->guest.u64RIP = pCtx->rip;
693	pVMCB->guest.u64RSP = pCtx->rsp;
694	pVMCB->guest.u64RFlags = pCtx->eflags.u32;
695
696	/* Set CPL */
697	pVMCB->guest.u8CPL = pCtx->csHid.Attr.n.u2Dpl;
698
699	/* RAX/EAX too, as VMRUN uses RAX as an implicit parameter. */
700	pVMCB->guest.u64RAX = pCtx->rax;
701
702	/* vmrun will fail without MSR_K6_EFER_SVME. */
703	pVMCB->guest.u64EFER = pCtx->msrEFER \| MSR_K6_EFER_SVME;
704
705	/* 64 bits guest mode? */
706	if (pCtx->msrEFER & MSR_K6_EFER_LMA)
707	{
708	#if !defined(VBOX_WITH_64_BITS_GUESTS) \|\| HC_ARCH_BITS != 64
709	return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
710	#else
711	pVM->hwaccm.s.svm.pfnVMRun = SVMVMRun64;
712	#endif
713	/* Unconditionally update these as wrmsr might have changed them. (HWACCM_CHANGED_GUEST_SEGMENT_REGS will not be set) */
714	pVMCB->guest.FS.u64Base = pCtx->fsHid.u64Base;
715	pVMCB->guest.GS.u64Base = pCtx->gsHid.u64Base;
716	}
717	else
718	{
719	/* Filter out the MSR_K6_LME bit or else AMD-V expects amd64 shadow paging. */
720	pVMCB->guest.u64EFER &= ~MSR_K6_EFER_LME;
721
722	pVM->hwaccm.s.svm.pfnVMRun = SVMVMRun;
723	}
724
725	/** TSC offset. */
726	if (TMCpuTickCanUseRealTSC(pVM, &pVMCB->ctrl.u64TSCOffset))
727	{
728	pVMCB->ctrl.u32InterceptCtrl1 &= ~SVM_CTRL1_INTERCEPT_RDTSC;
729	STAM_COUNTER_INC(&pVM->hwaccm.s.StatTSCOffset);
730	}
731	else
732	{
733	pVMCB->ctrl.u32InterceptCtrl1 \|= SVM_CTRL1_INTERCEPT_RDTSC;
734	STAM_COUNTER_INC(&pVM->hwaccm.s.StatTSCIntercept);
735	}
736
737	/* Sync the various msrs for 64 bits mode. */
738	pVMCB->guest.u64STAR = pCtx->msrSTAR; /* legacy syscall eip, cs & ss */
739	pVMCB->guest.u64LSTAR = pCtx->msrLSTAR; /* 64 bits mode syscall rip */
740	pVMCB->guest.u64CSTAR = pCtx->msrCSTAR; /* compatibility mode syscall rip */
741	pVMCB->guest.u64SFMASK = pCtx->msrSFMASK; /* syscall flag mask */
742	pVMCB->guest.u64KernelGSBase = pCtx->msrKERNELGSBASE; /* swapgs exchange value */
743
744	#ifdef DEBUG
745	/* Intercept X86_XCPT_DB if stepping is enabled */
746	if (DBGFIsStepping(pVM))
747	pVMCB->ctrl.u32InterceptException \|= RT_BIT(1);
748	else
749	pVMCB->ctrl.u32InterceptException &= ~RT_BIT(1);
750	#endif
751
752	/* Done. */
753	pVM->hwaccm.s.fContextUseFlags &= ~HWACCM_CHANGED_ALL_GUEST;
754
755	return VINF_SUCCESS;
756	}
757
758
759	/**
760	* Runs guest code in an SVM VM.
761	*
762	* @todo This can be much more efficient, when we only sync that which has actually changed. (this is the first attempt only)
763	*
764	* @returns VBox status code.
765	* @param pVM The VM to operate on.
766	* @param pCtx Guest context
767	* @param pCpu CPU info struct
768	*/
769	HWACCMR0DECL(int) SVMR0RunGuestCode(PVM pVM, CPUMCTX *pCtx, PHWACCM_CPUINFO pCpu)
770	{
771	int rc = VINF_SUCCESS;
772	uint64_t exitCode = (uint64_t)SVM_EXIT_INVALID;
773	SVM_VMCB *pVMCB;
774	bool fGuestStateSynced = false;
775	unsigned cResume = 0;
776	uint8_t u8LastVTPR;
777
778	STAM_PROFILE_ADV_START(&pVM->hwaccm.s.StatEntry, x);
779
780	AssertReturn(pCpu->fConfigured, VERR_EM_INTERNAL_ERROR);
781
782	pVMCB = (SVM_VMCB *)pVM->hwaccm.s.svm.pVMCB;
783	AssertMsgReturn(pVMCB, ("Invalid pVMCB\n"), VERR_EM_INTERNAL_ERROR);
784
785	/* We can jump to this point to resume execution after determining that a VM-exit is innocent.
786	*/
787	ResumeExecution:
788	/* Safety precaution; looping for too long here can have a very bad effect on the host */
789	if (++cResume > HWACCM_MAX_RESUME_LOOPS)
790	{
791	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitMaxResume);
792	rc = VINF_EM_RAW_INTERRUPT;
793	goto end;
794	}
795
796	/* Check for irq inhibition due to instruction fusing (sti, mov ss). */
797	if (VM_FF_ISSET(pVM, VM_FF_INHIBIT_INTERRUPTS))
798	{
799	Log(("VM_FF_INHIBIT_INTERRUPTS at %VGv successor %VGv\n", pCtx->rip, EMGetInhibitInterruptsPC(pVM)));
800	if (pCtx->rip != EMGetInhibitInterruptsPC(pVM))
801	{
802	/** @note we intentionally don't clear VM_FF_INHIBIT_INTERRUPTS here.
803	* Before we are able to execute this instruction in raw mode (iret to guest code) an external interrupt might
804	* force a world switch again. Possibly allowing a guest interrupt to be dispatched in the process. This could
805	* break the guest. Sounds very unlikely, but such timing sensitive problem are not as rare as you might think.
806	*/
807	VM_FF_CLEAR(pVM, VM_FF_INHIBIT_INTERRUPTS);
808	/* Irq inhibition is no longer active; clear the corresponding SVM state. */
809	pVMCB->ctrl.u64IntShadow = 0;
810	}
811	}
812	else
813	{
814	/* Irq inhibition is no longer active; clear the corresponding SVM state. */
815	pVMCB->ctrl.u64IntShadow = 0;
816	}
817
818	/* Check for pending actions that force us to go back to ring 3. */
819	#ifdef DEBUG
820	/* Intercept X86_XCPT_DB if stepping is enabled */
821	if (!DBGFIsStepping(pVM))
822	#endif
823	{
824	if (VM_FF_ISPENDING(pVM, VM_FF_TO_R3 \| VM_FF_TIMER))
825	{
826	VM_FF_CLEAR(pVM, VM_FF_TO_R3);
827	STAM_COUNTER_INC(&pVM->hwaccm.s.StatSwitchToR3);
828	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatEntry, x);
829	rc = VINF_EM_RAW_TO_R3;
830	goto end;
831	}
832	}
833
834	/* Pending request packets might contain actions that need immediate attention, such as pending hardware interrupts. */
835	if (VM_FF_ISPENDING(pVM, VM_FF_REQUEST))
836	{
837	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatEntry, x);
838	rc = VINF_EM_PENDING_REQUEST;
839	goto end;
840	}
841
842	/* When external interrupts are pending, we should exit the VM when IF is set. */
843	/** @note after VM_FF_INHIBIT_INTERRUPTS check!!! */
844	rc = SVMR0CheckPendingInterrupt(pVM, pVMCB, pCtx);
845	if (VBOX_FAILURE(rc))
846	{
847	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatEntry, x);
848	goto end;
849	}
850
851	/* Load the guest state */
852	rc = SVMR0LoadGuestState(pVM, pCtx);
853	if (rc != VINF_SUCCESS)
854	{
855	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatEntry, x);
856	goto end;
857	}
858	fGuestStateSynced = true;
859
860	/* TPR caching using CR8 is only available in 64 bits mode */
861	/* Note the 32 bits exception for AMD (X86_CPUID_AMD_FEATURE_ECX_CR8L), but that appears missing in Intel CPUs */
862	/* Note: we can't do this in LoadGuestState as PDMApicGetTPR can jump back to ring 3 (lock). */
863	if (pCtx->msrEFER & MSR_K6_EFER_LMA)
864	{
865	/* TPR caching in CR8 */
866	int rc = PDMApicGetTPR(pVM, &u8LastVTPR);
867	AssertRC(rc);
868	pVMCB->ctrl.IntCtrl.n.u8VTPR = u8LastVTPR;
869	}
870
871	/* All done! Let's start VM execution. */
872	STAM_PROFILE_ADV_START(&pVM->hwaccm.s.StatInGC, x);
873
874	/* Enable nested paging if necessary (disabled each time after #VMEXIT). */
875	pVMCB->ctrl.NestedPaging.n.u1NestedPaging = pVM->hwaccm.s.fNestedPaging;
876
877	/* Force a TLB flush for the first world switch if the current cpu differs from the one we ran on last. */
878	if (!pVM->hwaccm.s.svm.fResumeVM)
879	{
880	if ( pVM->hwaccm.s.svm.idLastCpu != pCpu->idCpu
881	/* if the tlb flush count has changed, another VM has flushed the TLB of this cpu, so we can't use our current ASID anymore. */
882	\|\| pVM->hwaccm.s.svm.cTLBFlushes != pCpu->cTLBFlushes)
883	{
884	/* Force a TLB flush on VM entry. */
885	pVM->hwaccm.s.svm.fForceTLBFlush = true;
886	}
887	pVM->hwaccm.s.svm.idLastCpu = pCpu->idCpu;
888	}
889	else
890	Assert(pVM->hwaccm.s.svm.idLastCpu == pCpu->idCpu);
891
892	/* Make sure we flush the TLB when required. Switch ASID to achieve the same thing, but without actually flushing the whole TLB (which is expensive). */
893	if ( pVM->hwaccm.s.svm.fForceTLBFlush
894	&& !pVM->hwaccm.s.svm.fAlwaysFlushTLB)
895	{
896	if (++pCpu->uCurrentASID >= pVM->hwaccm.s.svm.u32MaxASID)
897	{
898	pCpu->uCurrentASID = 1; /* start at 1; host uses 0 */
899	pVMCB->ctrl.TLBCtrl.n.u1TLBFlush = 1; /* wrap around; flush TLB */
900	pCpu->cTLBFlushes++;
901	}
902	else
903	STAM_COUNTER_INC(&pVM->hwaccm.s.StatFlushASID);
904
905	pVM->hwaccm.s.svm.cTLBFlushes = pCpu->cTLBFlushes;
906	}
907	else
908	{
909	/* We never increase uCurrentASID in the fAlwaysFlushTLB (erratum 170) case. */
910	if (!pCpu->uCurrentASID)
911	pCpu->uCurrentASID = 1;
912
913	pVMCB->ctrl.TLBCtrl.n.u1TLBFlush = pVM->hwaccm.s.svm.fForceTLBFlush;
914	}
915
916	AssertMsg(pCpu->uCurrentASID >= 1 && pCpu->uCurrentASID < pVM->hwaccm.s.svm.u32MaxASID, ("cpu%d uCurrentASID = %x\n", pCpu->idCpu, pCpu->uCurrentASID));
917	pVMCB->ctrl.TLBCtrl.n.u32ASID = pCpu->uCurrentASID;
918
919	#ifdef VBOX_WITH_STATISTICS
920	if (pVMCB->ctrl.TLBCtrl.n.u1TLBFlush)
921	STAM_COUNTER_INC(&pVM->hwaccm.s.StatFlushTLBWorldSwitch);
922	else
923	STAM_COUNTER_INC(&pVM->hwaccm.s.StatNoFlushTLBWorldSwitch);
924	#endif
925
926	/* In case we execute a goto ResumeExecution later on. */
927	pVM->hwaccm.s.svm.fResumeVM = true;
928	pVM->hwaccm.s.svm.fForceTLBFlush = pVM->hwaccm.s.svm.fAlwaysFlushTLB;
929
930	Assert(sizeof(pVM->hwaccm.s.svm.pVMCBPhys) == 8);
931	Assert(pVMCB->ctrl.u32InterceptCtrl2 == ( SVM_CTRL2_INTERCEPT_VMRUN /* required */
932	\| SVM_CTRL2_INTERCEPT_VMMCALL
933	\| SVM_CTRL2_INTERCEPT_VMLOAD
934	\| SVM_CTRL2_INTERCEPT_VMSAVE
935	\| SVM_CTRL2_INTERCEPT_STGI
936	\| SVM_CTRL2_INTERCEPT_CLGI
937	\| SVM_CTRL2_INTERCEPT_SKINIT
938	\| SVM_CTRL2_INTERCEPT_RDTSCP /* AMD only; we don't support this one */
939	\| SVM_CTRL2_INTERCEPT_WBINVD
940	\| SVM_CTRL2_INTERCEPT_MWAIT_UNCOND /* don't execute mwait or else we'll idle inside the guest (host thinks the cpu load is high) */
941	));
942	Assert(pVMCB->ctrl.IntCtrl.n.u1VIrqMasking);
943	Assert(pVMCB->ctrl.u64IOPMPhysAddr == pVM->hwaccm.s.svm.pIOBitmapPhys);
944	Assert(pVMCB->ctrl.u64MSRPMPhysAddr == pVM->hwaccm.s.svm.pMSRBitmapPhys);
945	Assert(pVMCB->ctrl.u64LBRVirt == 0);
946
947	pVM->hwaccm.s.svm.pfnVMRun(pVM->hwaccm.s.svm.pVMCBHostPhys, pVM->hwaccm.s.svm.pVMCBPhys, pCtx);
948	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatInGC, x);
949
950	/**
951	* !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
952	* 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
953	* !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
954	*/
955
956	STAM_PROFILE_ADV_START(&pVM->hwaccm.s.StatExit, x);
957
958	/* Reason for the VM exit */
959	exitCode = pVMCB->ctrl.u64ExitCode;
960
961	if (exitCode == (uint64_t)SVM_EXIT_INVALID) /* Invalid guest state. */
962	{
963	HWACCMDumpRegs(pCtx);
964	#ifdef DEBUG
965	Log(("ctrl.u16InterceptRdCRx %x\n", pVMCB->ctrl.u16InterceptRdCRx));
966	Log(("ctrl.u16InterceptWrCRx %x\n", pVMCB->ctrl.u16InterceptWrCRx));
967	Log(("ctrl.u16InterceptRdDRx %x\n", pVMCB->ctrl.u16InterceptRdDRx));
968	Log(("ctrl.u16InterceptWrDRx %x\n", pVMCB->ctrl.u16InterceptWrDRx));
969	Log(("ctrl.u32InterceptException %x\n", pVMCB->ctrl.u32InterceptException));
970	Log(("ctrl.u32InterceptCtrl1 %x\n", pVMCB->ctrl.u32InterceptCtrl1));
971	Log(("ctrl.u32InterceptCtrl2 %x\n", pVMCB->ctrl.u32InterceptCtrl2));
972	Log(("ctrl.u64IOPMPhysAddr %VX64\n", pVMCB->ctrl.u64IOPMPhysAddr));
973	Log(("ctrl.u64MSRPMPhysAddr %VX64\n", pVMCB->ctrl.u64MSRPMPhysAddr));
974	Log(("ctrl.u64TSCOffset %VX64\n", pVMCB->ctrl.u64TSCOffset));
975
976	Log(("ctrl.TLBCtrl.u32ASID %x\n", pVMCB->ctrl.TLBCtrl.n.u32ASID));
977	Log(("ctrl.TLBCtrl.u1TLBFlush %x\n", pVMCB->ctrl.TLBCtrl.n.u1TLBFlush));
978	Log(("ctrl.TLBCtrl.u7Reserved %x\n", pVMCB->ctrl.TLBCtrl.n.u7Reserved));
979	Log(("ctrl.TLBCtrl.u24Reserved %x\n", pVMCB->ctrl.TLBCtrl.n.u24Reserved));
980
981	Log(("ctrl.IntCtrl.u8VTPR %x\n", pVMCB->ctrl.IntCtrl.n.u8VTPR));
982	Log(("ctrl.IntCtrl.u1VIrqValid %x\n", pVMCB->ctrl.IntCtrl.n.u1VIrqValid));
983	Log(("ctrl.IntCtrl.u7Reserved %x\n", pVMCB->ctrl.IntCtrl.n.u7Reserved));
984	Log(("ctrl.IntCtrl.u4VIrqPriority %x\n", pVMCB->ctrl.IntCtrl.n.u4VIrqPriority));
985	Log(("ctrl.IntCtrl.u1IgnoreTPR %x\n", pVMCB->ctrl.IntCtrl.n.u1IgnoreTPR));
986	Log(("ctrl.IntCtrl.u3Reserved %x\n", pVMCB->ctrl.IntCtrl.n.u3Reserved));
987	Log(("ctrl.IntCtrl.u1VIrqMasking %x\n", pVMCB->ctrl.IntCtrl.n.u1VIrqMasking));
988	Log(("ctrl.IntCtrl.u7Reserved2 %x\n", pVMCB->ctrl.IntCtrl.n.u7Reserved2));
989	Log(("ctrl.IntCtrl.u8VIrqVector %x\n", pVMCB->ctrl.IntCtrl.n.u8VIrqVector));
990	Log(("ctrl.IntCtrl.u24Reserved %x\n", pVMCB->ctrl.IntCtrl.n.u24Reserved));
991
992	Log(("ctrl.u64IntShadow %VX64\n", pVMCB->ctrl.u64IntShadow));
993	Log(("ctrl.u64ExitCode %VX64\n", pVMCB->ctrl.u64ExitCode));
994	Log(("ctrl.u64ExitInfo1 %VX64\n", pVMCB->ctrl.u64ExitInfo1));
995	Log(("ctrl.u64ExitInfo2 %VX64\n", pVMCB->ctrl.u64ExitInfo2));
996	Log(("ctrl.ExitIntInfo.u8Vector %x\n", pVMCB->ctrl.ExitIntInfo.n.u8Vector));
997	Log(("ctrl.ExitIntInfo.u3Type %x\n", pVMCB->ctrl.ExitIntInfo.n.u3Type));
998	Log(("ctrl.ExitIntInfo.u1ErrorCodeValid %x\n", pVMCB->ctrl.ExitIntInfo.n.u1ErrorCodeValid));
999	Log(("ctrl.ExitIntInfo.u19Reserved %x\n", pVMCB->ctrl.ExitIntInfo.n.u19Reserved));
1000	Log(("ctrl.ExitIntInfo.u1Valid %x\n", pVMCB->ctrl.ExitIntInfo.n.u1Valid));
1001	Log(("ctrl.ExitIntInfo.u32ErrorCode %x\n", pVMCB->ctrl.ExitIntInfo.n.u32ErrorCode));
1002	Log(("ctrl.NestedPaging %VX64\n", pVMCB->ctrl.NestedPaging.au64));
1003	Log(("ctrl.EventInject.u8Vector %x\n", pVMCB->ctrl.EventInject.n.u8Vector));
1004	Log(("ctrl.EventInject.u3Type %x\n", pVMCB->ctrl.EventInject.n.u3Type));
1005	Log(("ctrl.EventInject.u1ErrorCodeValid %x\n", pVMCB->ctrl.EventInject.n.u1ErrorCodeValid));
1006	Log(("ctrl.EventInject.u19Reserved %x\n", pVMCB->ctrl.EventInject.n.u19Reserved));
1007	Log(("ctrl.EventInject.u1Valid %x\n", pVMCB->ctrl.EventInject.n.u1Valid));
1008	Log(("ctrl.EventInject.u32ErrorCode %x\n", pVMCB->ctrl.EventInject.n.u32ErrorCode));
1009
1010	Log(("ctrl.u64NestedPagingCR3 %VX64\n", pVMCB->ctrl.u64NestedPagingCR3));
1011	Log(("ctrl.u64LBRVirt %VX64\n", pVMCB->ctrl.u64LBRVirt));
1012
1013	Log(("guest.CS.u16Sel %04X\n", pVMCB->guest.CS.u16Sel));
1014	Log(("guest.CS.u16Attr %04X\n", pVMCB->guest.CS.u16Attr));
1015	Log(("guest.CS.u32Limit %X\n", pVMCB->guest.CS.u32Limit));
1016	Log(("guest.CS.u64Base %VX64\n", pVMCB->guest.CS.u64Base));
1017	Log(("guest.DS.u16Sel %04X\n", pVMCB->guest.DS.u16Sel));
1018	Log(("guest.DS.u16Attr %04X\n", pVMCB->guest.DS.u16Attr));
1019	Log(("guest.DS.u32Limit %X\n", pVMCB->guest.DS.u32Limit));
1020	Log(("guest.DS.u64Base %VX64\n", pVMCB->guest.DS.u64Base));
1021	Log(("guest.ES.u16Sel %04X\n", pVMCB->guest.ES.u16Sel));
1022	Log(("guest.ES.u16Attr %04X\n", pVMCB->guest.ES.u16Attr));
1023	Log(("guest.ES.u32Limit %X\n", pVMCB->guest.ES.u32Limit));
1024	Log(("guest.ES.u64Base %VX64\n", pVMCB->guest.ES.u64Base));
1025	Log(("guest.FS.u16Sel %04X\n", pVMCB->guest.FS.u16Sel));
1026	Log(("guest.FS.u16Attr %04X\n", pVMCB->guest.FS.u16Attr));
1027	Log(("guest.FS.u32Limit %X\n", pVMCB->guest.FS.u32Limit));
1028	Log(("guest.FS.u64Base %VX64\n", pVMCB->guest.FS.u64Base));
1029	Log(("guest.GS.u16Sel %04X\n", pVMCB->guest.GS.u16Sel));
1030	Log(("guest.GS.u16Attr %04X\n", pVMCB->guest.GS.u16Attr));
1031	Log(("guest.GS.u32Limit %X\n", pVMCB->guest.GS.u32Limit));
1032	Log(("guest.GS.u64Base %VX64\n", pVMCB->guest.GS.u64Base));
1033
1034	Log(("guest.GDTR.u32Limit %X\n", pVMCB->guest.GDTR.u32Limit));
1035	Log(("guest.GDTR.u64Base %VX64\n", pVMCB->guest.GDTR.u64Base));
1036
1037	Log(("guest.LDTR.u16Sel %04X\n", pVMCB->guest.LDTR.u16Sel));
1038	Log(("guest.LDTR.u16Attr %04X\n", pVMCB->guest.LDTR.u16Attr));
1039	Log(("guest.LDTR.u32Limit %X\n", pVMCB->guest.LDTR.u32Limit));
1040	Log(("guest.LDTR.u64Base %VX64\n", pVMCB->guest.LDTR.u64Base));
1041
1042	Log(("guest.IDTR.u32Limit %X\n", pVMCB->guest.IDTR.u32Limit));
1043	Log(("guest.IDTR.u64Base %VX64\n", pVMCB->guest.IDTR.u64Base));
1044
1045	Log(("guest.TR.u16Sel %04X\n", pVMCB->guest.TR.u16Sel));
1046	Log(("guest.TR.u16Attr %04X\n", pVMCB->guest.TR.u16Attr));
1047	Log(("guest.TR.u32Limit %X\n", pVMCB->guest.TR.u32Limit));
1048	Log(("guest.TR.u64Base %VX64\n", pVMCB->guest.TR.u64Base));
1049
1050	Log(("guest.u8CPL %X\n", pVMCB->guest.u8CPL));
1051	Log(("guest.u64CR0 %VX64\n", pVMCB->guest.u64CR0));
1052	Log(("guest.u64CR2 %VX64\n", pVMCB->guest.u64CR2));
1053	Log(("guest.u64CR3 %VX64\n", pVMCB->guest.u64CR3));
1054	Log(("guest.u64CR4 %VX64\n", pVMCB->guest.u64CR4));
1055	Log(("guest.u64DR6 %VX64\n", pVMCB->guest.u64DR6));
1056	Log(("guest.u64DR7 %VX64\n", pVMCB->guest.u64DR7));
1057
1058	Log(("guest.u64RIP %VX64\n", pVMCB->guest.u64RIP));
1059	Log(("guest.u64RSP %VX64\n", pVMCB->guest.u64RSP));
1060	Log(("guest.u64RAX %VX64\n", pVMCB->guest.u64RAX));
1061	Log(("guest.u64RFlags %VX64\n", pVMCB->guest.u64RFlags));
1062
1063	Log(("guest.u64SysEnterCS %VX64\n", pVMCB->guest.u64SysEnterCS));
1064	Log(("guest.u64SysEnterEIP %VX64\n", pVMCB->guest.u64SysEnterEIP));
1065	Log(("guest.u64SysEnterESP %VX64\n", pVMCB->guest.u64SysEnterESP));
1066
1067	Log(("guest.u64EFER %VX64\n", pVMCB->guest.u64EFER));
1068	Log(("guest.u64STAR %VX64\n", pVMCB->guest.u64STAR));
1069	Log(("guest.u64LSTAR %VX64\n", pVMCB->guest.u64LSTAR));
1070	Log(("guest.u64CSTAR %VX64\n", pVMCB->guest.u64CSTAR));
1071	Log(("guest.u64SFMASK %VX64\n", pVMCB->guest.u64SFMASK));
1072	Log(("guest.u64KernelGSBase %VX64\n", pVMCB->guest.u64KernelGSBase));
1073	Log(("guest.u64GPAT %VX64\n", pVMCB->guest.u64GPAT));
1074	Log(("guest.u64DBGCTL %VX64\n", pVMCB->guest.u64DBGCTL));
1075	Log(("guest.u64BR_FROM %VX64\n", pVMCB->guest.u64BR_FROM));
1076	Log(("guest.u64BR_TO %VX64\n", pVMCB->guest.u64BR_TO));
1077	Log(("guest.u64LASTEXCPFROM %VX64\n", pVMCB->guest.u64LASTEXCPFROM));
1078	Log(("guest.u64LASTEXCPTO %VX64\n", pVMCB->guest.u64LASTEXCPTO));
1079
1080	#endif
1081	rc = VERR_SVM_UNABLE_TO_START_VM;
1082	goto end;
1083	}
1084
1085	/* Let's first sync back eip, esp, and eflags. */
1086	pCtx->rip = pVMCB->guest.u64RIP;
1087	pCtx->rsp = pVMCB->guest.u64RSP;
1088	pCtx->eflags.u32 = pVMCB->guest.u64RFlags;
1089	/* eax is saved/restore across the vmrun instruction */
1090	pCtx->rax = pVMCB->guest.u64RAX;
1091
1092	pCtx->msrKERNELGSBASE = pVMCB->guest.u64KernelGSBase; /* swapgs exchange value */
1093
1094	/* Guest CPU context: ES, CS, SS, DS, FS, GS. */
1095	SVM_READ_SELREG(SS, ss);
1096	SVM_READ_SELREG(CS, cs);
1097	SVM_READ_SELREG(DS, ds);
1098	SVM_READ_SELREG(ES, es);
1099	SVM_READ_SELREG(FS, fs);
1100	SVM_READ_SELREG(GS, gs);
1101
1102	/* Note: no reason to sync back the CRx and DRx registers. They can't be changed by the guest. */
1103	/* Note: only in the nested paging case can CR3 & CR4 be changed by the guest. */
1104	if ( pVM->hwaccm.s.fNestedPaging
1105	&& pCtx->cr3 != pVMCB->guest.u64CR3)
1106	{
1107	CPUMSetGuestCR3(pVM, pVMCB->guest.u64CR3);
1108	PGMUpdateCR3(pVM, pVMCB->guest.u64CR3);
1109	}
1110
1111	/** @note NOW IT'S SAFE FOR LOGGING! */
1112
1113	/* Take care of instruction fusing (sti, mov ss) */
1114	if (pVMCB->ctrl.u64IntShadow & SVM_INTERRUPT_SHADOW_ACTIVE)
1115	{
1116	Log(("uInterruptState %x eip=%VGv\n", pVMCB->ctrl.u64IntShadow, pCtx->rip));
1117	EMSetInhibitInterruptsPC(pVM, pCtx->rip);
1118	}
1119	else
1120	VM_FF_CLEAR(pVM, VM_FF_INHIBIT_INTERRUPTS);
1121
1122	Log2(("exitCode = %x\n", exitCode));
1123
1124	/* Sync back the debug registers. */
1125	/** @todo Implement debug registers correctly. */
1126	pCtx->dr6 = pVMCB->guest.u64DR6;
1127	pCtx->dr7 = pVMCB->guest.u64DR7;
1128
1129	/* Update the APIC if the cached TPR value has changed. */
1130	if ( (pCtx->msrEFER & MSR_K6_EFER_LMA)
1131	&& pVMCB->ctrl.IntCtrl.n.u8VTPR != u8LastVTPR)
1132	{
1133	rc = PDMApicSetTPR(pVM, pVMCB->ctrl.IntCtrl.n.u8VTPR);
1134	AssertRC(rc);
1135	u8LastVTPR = pVMCB->ctrl.IntCtrl.n.u8VTPR;
1136	}
1137
1138	/* Check if an injected event was interrupted prematurely. */
1139	pVM->hwaccm.s.Event.intInfo = pVMCB->ctrl.ExitIntInfo.au64[0];
1140	if ( pVMCB->ctrl.ExitIntInfo.n.u1Valid
1141	&& pVMCB->ctrl.ExitIntInfo.n.u3Type != SVM_EVENT_SOFTWARE_INT /* we don't care about 'int xx' as the instruction will be restarted. */)
1142	{
1143	Log(("Pending inject %VX64 at %VGv exit=%08x\n", pVM->hwaccm.s.Event.intInfo, pCtx->rip, exitCode));
1144
1145	#ifdef LOG_ENABLED
1146	SVM_EVENT Event;
1147	Event.au64[0] = pVM->hwaccm.s.Event.intInfo;
1148
1149	if ( exitCode == SVM_EXIT_EXCEPTION_E
1150	&& Event.n.u8Vector == 0xE)
1151	{
1152	Log(("Double fault!\n"));
1153	}
1154	#endif
1155
1156	pVM->hwaccm.s.Event.fPending = true;
1157	/* Error code present? (redundant) */
1158	if (pVMCB->ctrl.ExitIntInfo.n.u1ErrorCodeValid)
1159	{
1160	pVM->hwaccm.s.Event.errCode = pVMCB->ctrl.ExitIntInfo.n.u32ErrorCode;
1161	}
1162	else
1163	pVM->hwaccm.s.Event.errCode = 0;
1164	}
1165	#ifdef VBOX_WITH_STATISTICS
1166	if (exitCode == SVM_EXIT_NPF)
1167	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitReasonNPF);
1168	else
1169	STAM_COUNTER_INC(&pVM->hwaccm.s.pStatExitReasonR0[exitCode & MASK_EXITREASON_STAT]);
1170	#endif
1171
1172	/* Deal with the reason of the VM-exit. */
1173	switch (exitCode)
1174	{
1175	case SVM_EXIT_EXCEPTION_0: case SVM_EXIT_EXCEPTION_1: case SVM_EXIT_EXCEPTION_2: case SVM_EXIT_EXCEPTION_3:
1176	case SVM_EXIT_EXCEPTION_4: case SVM_EXIT_EXCEPTION_5: case SVM_EXIT_EXCEPTION_6: case SVM_EXIT_EXCEPTION_7:
1177	case SVM_EXIT_EXCEPTION_8: case SVM_EXIT_EXCEPTION_9: case SVM_EXIT_EXCEPTION_A: case SVM_EXIT_EXCEPTION_B:
1178	case SVM_EXIT_EXCEPTION_C: case SVM_EXIT_EXCEPTION_D: case SVM_EXIT_EXCEPTION_E: case SVM_EXIT_EXCEPTION_F:
1179	case SVM_EXIT_EXCEPTION_10: case SVM_EXIT_EXCEPTION_11: case SVM_EXIT_EXCEPTION_12: case SVM_EXIT_EXCEPTION_13:
1180	case SVM_EXIT_EXCEPTION_14: case SVM_EXIT_EXCEPTION_15: case SVM_EXIT_EXCEPTION_16: case SVM_EXIT_EXCEPTION_17:
1181	case SVM_EXIT_EXCEPTION_18: case SVM_EXIT_EXCEPTION_19: case SVM_EXIT_EXCEPTION_1A: case SVM_EXIT_EXCEPTION_1B:
1182	case SVM_EXIT_EXCEPTION_1C: case SVM_EXIT_EXCEPTION_1D: case SVM_EXIT_EXCEPTION_1E: case SVM_EXIT_EXCEPTION_1F:
1183	{
1184	/* Pending trap. */
1185	SVM_EVENT Event;
1186	uint32_t vector = exitCode - SVM_EXIT_EXCEPTION_0;
1187
1188	Log2(("Hardware/software interrupt %d\n", vector));
1189	switch (vector)
1190	{
1191	#ifdef DEBUG
1192	case X86_XCPT_DB:
1193	rc = DBGFR0Trap01Handler(pVM, CPUMCTX2CORE(pCtx), pVMCB->guest.u64DR6);
1194	Assert(rc != VINF_EM_RAW_GUEST_TRAP);
1195	break;
1196	#endif
1197
1198	case X86_XCPT_NM:
1199	{
1200	uint32_t oldCR0;
1201
1202	Log(("#NM fault at %VGv\n", pCtx->rip));
1203
1204	/** @todo don't intercept #NM exceptions anymore when we've activated the guest FPU state. */
1205	oldCR0 = ASMGetCR0();
1206	/* If we sync the FPU/XMM state on-demand, then we can continue execution as if nothing has happened. */
1207	rc = CPUMHandleLazyFPU(pVM);
1208	if (rc == VINF_SUCCESS)
1209	{
1210	Assert(CPUMIsGuestFPUStateActive(pVM));
1211
1212	/* CPUMHandleLazyFPU could have changed CR0; restore it. */
1213	ASMSetCR0(oldCR0);
1214
1215	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitShadowNM);
1216
1217	/* Continue execution. */
1218	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1219	pVM->hwaccm.s.fContextUseFlags \|= HWACCM_CHANGED_GUEST_CR0;
1220
1221	goto ResumeExecution;
1222	}
1223
1224	Log(("Forward #NM fault to the guest\n"));
1225	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitGuestNM);
1226
1227	Event.au64[0] = 0;
1228	Event.n.u3Type = SVM_EVENT_EXCEPTION;
1229	Event.n.u1Valid = 1;
1230	Event.n.u8Vector = X86_XCPT_NM;
1231
1232	SVMR0InjectEvent(pVM, pVMCB, pCtx, &Event);
1233	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1234	goto ResumeExecution;
1235	}
1236
1237	case X86_XCPT_PF: /* Page fault */
1238	{
1239	uint32_t errCode = pVMCB->ctrl.u64ExitInfo1; /* EXITINFO1 = error code */
1240	RTGCUINTPTR uFaultAddress = pVMCB->ctrl.u64ExitInfo2; /* EXITINFO2 = fault address */
1241
1242	#ifdef DEBUG
1243	if (pVM->hwaccm.s.fNestedPaging)
1244	{ /* A genuine pagefault.
1245	* Forward the trap to the guest by injecting the exception and resuming execution.
1246	*/
1247	Log(("Guest page fault at %VGv cr2=%VGv error code %x rsp=%VGv\n", (RTGCPTR)pCtx->rip, uFaultAddress, errCode, (RTGCPTR)pCtx->rsp));
1248	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitGuestPF);
1249
1250	/* Now we must update CR2. */
1251	pCtx->cr2 = uFaultAddress;
1252
1253	Event.au64[0] = 0;
1254	Event.n.u3Type = SVM_EVENT_EXCEPTION;
1255	Event.n.u1Valid = 1;
1256	Event.n.u8Vector = X86_XCPT_PF;
1257	Event.n.u1ErrorCodeValid = 1;
1258	Event.n.u32ErrorCode = errCode;
1259
1260	SVMR0InjectEvent(pVM, pVMCB, pCtx, &Event);
1261
1262	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1263	goto ResumeExecution;
1264	}
1265	#endif
1266	Assert(!pVM->hwaccm.s.fNestedPaging);
1267
1268	Log2(("Page fault at %VGv cr2=%VGv error code %x\n", pCtx->rip, uFaultAddress, errCode));
1269	/* Exit qualification contains the linear address of the page fault. */
1270	TRPMAssertTrap(pVM, X86_XCPT_PF, TRPM_TRAP);
1271	TRPMSetErrorCode(pVM, errCode);
1272	TRPMSetFaultAddress(pVM, uFaultAddress);
1273
1274	/* Forward it to our trap handler first, in case our shadow pages are out of sync. */
1275	rc = PGMTrap0eHandler(pVM, errCode, CPUMCTX2CORE(pCtx), (RTGCPTR)uFaultAddress);
1276	Log2(("PGMTrap0eHandler %VGv returned %Vrc\n", pCtx->rip, rc));
1277	if (rc == VINF_SUCCESS)
1278	{ /* We've successfully synced our shadow pages, so let's just continue execution. */
1279	Log2(("Shadow page fault at %VGv cr2=%VGv error code %x\n", pCtx->rip, uFaultAddress, errCode));
1280	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitShadowPF);
1281
1282	TRPMResetTrap(pVM);
1283
1284	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1285	goto ResumeExecution;
1286	}
1287	else
1288	if (rc == VINF_EM_RAW_GUEST_TRAP)
1289	{ /* A genuine pagefault.
1290	* Forward the trap to the guest by injecting the exception and resuming execution.
1291	*/
1292	Log2(("Forward page fault to the guest\n"));
1293	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitGuestPF);
1294	/* The error code might have been changed. */
1295	errCode = TRPMGetErrorCode(pVM);
1296
1297	TRPMResetTrap(pVM);
1298
1299	/* Now we must update CR2. */
1300	pCtx->cr2 = uFaultAddress;
1301
1302	Event.au64[0] = 0;
1303	Event.n.u3Type = SVM_EVENT_EXCEPTION;
1304	Event.n.u1Valid = 1;
1305	Event.n.u8Vector = X86_XCPT_PF;
1306	Event.n.u1ErrorCodeValid = 1;
1307	Event.n.u32ErrorCode = errCode;
1308
1309	SVMR0InjectEvent(pVM, pVMCB, pCtx, &Event);
1310
1311	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1312	goto ResumeExecution;
1313	}
1314	#ifdef VBOX_STRICT
1315	if (rc != VINF_EM_RAW_EMULATE_INSTR)
1316	LogFlow(("PGMTrap0eHandler failed with %d\n", rc));
1317	#endif
1318	/* Need to go back to the recompiler to emulate the instruction. */
1319	TRPMResetTrap(pVM);
1320	break;
1321	}
1322
1323	case X86_XCPT_MF: /* Floating point exception. */
1324	{
1325	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitGuestMF);
1326	if (!(pCtx->cr0 & X86_CR0_NE))
1327	{
1328	/* old style FPU error reporting needs some extra work. */
1329	/** @todo don't fall back to the recompiler, but do it manually. */
1330	rc = VINF_EM_RAW_EMULATE_INSTR;
1331	break;
1332	}
1333	Log(("Trap %x at %VGv\n", vector, pCtx->rip));
1334
1335	Event.au64[0] = 0;
1336	Event.n.u3Type = SVM_EVENT_EXCEPTION;
1337	Event.n.u1Valid = 1;
1338	Event.n.u8Vector = X86_XCPT_MF;
1339
1340	SVMR0InjectEvent(pVM, pVMCB, pCtx, &Event);
1341
1342	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1343	goto ResumeExecution;
1344	}
1345
1346	#ifdef VBOX_STRICT
1347	case X86_XCPT_GP: /* General protection failure exception.*/
1348	case X86_XCPT_UD: /* Unknown opcode exception. */
1349	case X86_XCPT_DE: /* Debug exception. */
1350	case X86_XCPT_SS: /* Stack segment exception. */
1351	case X86_XCPT_NP: /* Segment not present exception. */
1352	{
1353	Event.au64[0] = 0;
1354	Event.n.u3Type = SVM_EVENT_EXCEPTION;
1355	Event.n.u1Valid = 1;
1356	Event.n.u8Vector = vector;
1357
1358	switch(vector)
1359	{
1360	case X86_XCPT_GP:
1361	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitGuestGP);
1362	Event.n.u1ErrorCodeValid = 1;
1363	Event.n.u32ErrorCode = pVMCB->ctrl.u64ExitInfo1; /* EXITINFO1 = error code */
1364	break;
1365	case X86_XCPT_DE:
1366	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitGuestDE);
1367	break;
1368	case X86_XCPT_UD:
1369	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitGuestUD);
1370	break;
1371	case X86_XCPT_SS:
1372	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitGuestSS);
1373	Event.n.u1ErrorCodeValid = 1;
1374	Event.n.u32ErrorCode = pVMCB->ctrl.u64ExitInfo1; /* EXITINFO1 = error code */
1375	break;
1376	case X86_XCPT_NP:
1377	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitGuestNP);
1378	Event.n.u1ErrorCodeValid = 1;
1379	Event.n.u32ErrorCode = pVMCB->ctrl.u64ExitInfo1; /* EXITINFO1 = error code */
1380	break;
1381	}
1382	Log(("Trap %x at %VGv esi=%x\n", vector, pCtx->rip, pCtx->esi));
1383	SVMR0InjectEvent(pVM, pVMCB, pCtx, &Event);
1384
1385	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1386	goto ResumeExecution;
1387	}
1388	#endif
1389	default:
1390	AssertMsgFailed(("Unexpected vm-exit caused by exception %x\n", vector));
1391	rc = VERR_EM_INTERNAL_ERROR;
1392	break;
1393
1394	} /* switch (vector) */
1395	break;
1396	}
1397
1398	case SVM_EXIT_NPF:
1399	{
1400	/* EXITINFO1 contains fault errorcode; EXITINFO2 contains the guest physical address causing the fault. */
1401	uint32_t errCode = pVMCB->ctrl.u64ExitInfo1; /* EXITINFO1 = error code */
1402	RTGCPHYS uFaultAddress = pVMCB->ctrl.u64ExitInfo2; /* EXITINFO2 = fault address */
1403
1404	Assert(pVM->hwaccm.s.fNestedPaging);
1405	Log(("Nested page fault at %VGv cr2=%VGp error code %x\n", pCtx->rip, uFaultAddress, errCode));
1406	/* Exit qualification contains the linear address of the page fault. */
1407	TRPMAssertTrap(pVM, X86_XCPT_PF, TRPM_TRAP);
1408	TRPMSetErrorCode(pVM, errCode);
1409	TRPMSetFaultAddress(pVM, uFaultAddress);
1410
1411	/* Handle the pagefault trap for the nested shadow table. */
1412	rc = PGMR0Trap0eHandlerNestedPaging(pVM, PGMGetHostMode(pVM), errCode, CPUMCTX2CORE(pCtx), uFaultAddress);
1413	Log2(("PGMR0Trap0eHandlerNestedPaging %VGv returned %Vrc\n", pCtx->rip, rc));
1414	if (rc == VINF_SUCCESS)
1415	{ /* We've successfully synced our shadow pages, so let's just continue execution. */
1416	Log2(("Shadow page fault at %VGv cr2=%VGp error code %x\n", pCtx->rip, uFaultAddress, errCode));
1417	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitShadowPF);
1418
1419	TRPMResetTrap(pVM);
1420
1421	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1422	goto ResumeExecution;
1423	}
1424
1425	#ifdef VBOX_STRICT
1426	if (rc != VINF_EM_RAW_EMULATE_INSTR)
1427	LogFlow(("PGMTrap0eHandlerNestedPaging failed with %d\n", rc));
1428	#endif
1429	/* Need to go back to the recompiler to emulate the instruction. */
1430	TRPMResetTrap(pVM);
1431	break;
1432	}
1433
1434	case SVM_EXIT_VINTR:
1435	/* A virtual interrupt is about to be delivered, which means IF=1. */
1436	Log(("SVM_EXIT_VINTR IF=%d\n", pCtx->eflags.Bits.u1IF));
1437	pVMCB->ctrl.IntCtrl.n.u1VIrqValid = 0;
1438	pVMCB->ctrl.IntCtrl.n.u8VIrqVector = 0;
1439	goto ResumeExecution;
1440
1441	case SVM_EXIT_FERR_FREEZE:
1442	case SVM_EXIT_INTR:
1443	case SVM_EXIT_NMI:
1444	case SVM_EXIT_SMI:
1445	case SVM_EXIT_INIT:
1446	/* External interrupt; leave to allow it to be dispatched again. */
1447	rc = VINF_EM_RAW_INTERRUPT;
1448	break;
1449
1450	case SVM_EXIT_WBINVD:
1451	case SVM_EXIT_INVD: /* Guest software attempted to execute INVD. */
1452	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitInvd);
1453	/* Skip instruction and continue directly. */
1454	pCtx->rip += 2; /** @note hardcoded opcode size! */
1455	/* Continue execution.*/
1456	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1457	goto ResumeExecution;
1458
1459	case SVM_EXIT_CPUID: /* Guest software attempted to execute CPUID. */
1460	{
1461	Log2(("SVM: Cpuid at %VGv for %x\n", pCtx->rip, pCtx->eax));
1462	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitCpuid);
1463	rc = EMInterpretCpuId(pVM, CPUMCTX2CORE(pCtx));
1464	if (rc == VINF_SUCCESS)
1465	{
1466	/* Update EIP and continue execution. */
1467	pCtx->rip += 2; /** @note hardcoded opcode size! */
1468	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1469	goto ResumeExecution;
1470	}
1471	AssertMsgFailed(("EMU: cpuid failed with %Vrc\n", rc));
1472	rc = VINF_EM_RAW_EMULATE_INSTR;
1473	break;
1474	}
1475
1476	case SVM_EXIT_RDTSC: /* Guest software attempted to execute RDTSC. */
1477	{
1478	Log2(("SVM: Rdtsc\n"));
1479	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitRdtsc);
1480	rc = EMInterpretRdtsc(pVM, CPUMCTX2CORE(pCtx));
1481	if (rc == VINF_SUCCESS)
1482	{
1483	/* Update EIP and continue execution. */
1484	pCtx->rip += 2; /** @note hardcoded opcode size! */
1485	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1486	goto ResumeExecution;
1487	}
1488	AssertMsgFailed(("EMU: rdtsc failed with %Vrc\n", rc));
1489	rc = VINF_EM_RAW_EMULATE_INSTR;
1490	break;
1491	}
1492
1493	case SVM_EXIT_INVLPG: /* Guest software attempted to execute INVPG. */
1494	{
1495	Log2(("SVM: invlpg\n"));
1496	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitInvpg);
1497
1498	Assert(!pVM->hwaccm.s.fNestedPaging);
1499
1500	/* Truly a pita. Why can't SVM give the same information as VT-x? */
1501	rc = SVMR0InterpretInvpg(pVM, CPUMCTX2CORE(pCtx), pVMCB->ctrl.TLBCtrl.n.u32ASID);
1502	if (rc == VINF_SUCCESS)
1503	{
1504	STAM_COUNTER_INC(&pVM->hwaccm.s.StatFlushPageInvlpg);
1505	goto ResumeExecution; /* eip already updated */
1506	}
1507	break;
1508	}
1509
1510	case SVM_EXIT_WRITE_CR0: case SVM_EXIT_WRITE_CR1: case SVM_EXIT_WRITE_CR2: case SVM_EXIT_WRITE_CR3:
1511	case SVM_EXIT_WRITE_CR4: case SVM_EXIT_WRITE_CR5: case SVM_EXIT_WRITE_CR6: case SVM_EXIT_WRITE_CR7:
1512	case SVM_EXIT_WRITE_CR8: case SVM_EXIT_WRITE_CR9: case SVM_EXIT_WRITE_CR10: case SVM_EXIT_WRITE_CR11:
1513	case SVM_EXIT_WRITE_CR12: case SVM_EXIT_WRITE_CR13: case SVM_EXIT_WRITE_CR14: case SVM_EXIT_WRITE_CR15:
1514	{
1515	uint32_t cbSize;
1516
1517	Log2(("SVM: %VGv mov cr%d, \n", pCtx->rip, exitCode - SVM_EXIT_WRITE_CR0));
1518	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitCRxWrite);
1519	rc = EMInterpretInstruction(pVM, CPUMCTX2CORE(pCtx), 0, &cbSize);
1520
1521	switch (exitCode - SVM_EXIT_WRITE_CR0)
1522	{
1523	case 0:
1524	pVM->hwaccm.s.fContextUseFlags \|= HWACCM_CHANGED_GUEST_CR0;
1525	break;
1526	case 2:
1527	break;
1528	case 3:
1529	Assert(!pVM->hwaccm.s.fNestedPaging);
1530	pVM->hwaccm.s.fContextUseFlags \|= HWACCM_CHANGED_GUEST_CR3;
1531	break;
1532	case 4:
1533	pVM->hwaccm.s.fContextUseFlags \|= HWACCM_CHANGED_GUEST_CR4;
1534	break;
1535	default:
1536	AssertFailed();
1537	}
1538	/* Check if a sync operation is pending. */
1539	if ( rc == VINF_SUCCESS /* don't bother if we are going to ring 3 anyway */
1540	&& VM_FF_ISPENDING(pVM, VM_FF_PGM_SYNC_CR3 \| VM_FF_PGM_SYNC_CR3_NON_GLOBAL))
1541	{
1542	rc = PGMSyncCR3(pVM, CPUMGetGuestCR0(pVM), CPUMGetGuestCR3(pVM), CPUMGetGuestCR4(pVM), VM_FF_ISSET(pVM, VM_FF_PGM_SYNC_CR3));
1543	AssertRC(rc);
1544
1545	STAM_COUNTER_INC(&pVM->hwaccm.s.StatFlushTLBCRxChange);
1546
1547	/* Must be set by PGMSyncCR3 */
1548	Assert(PGMGetGuestMode(pVM) <= PGMMODE_PROTECTED \|\| pVM->hwaccm.s.svm.fForceTLBFlush);
1549	}
1550	if (rc == VINF_SUCCESS)
1551	{
1552	/* EIP has been updated already. */
1553
1554	/* Only resume if successful. */
1555	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1556	goto ResumeExecution;
1557	}
1558	Assert(rc == VERR_EM_INTERPRETER \|\| rc == VINF_PGM_CHANGE_MODE \|\| rc == VINF_PGM_SYNC_CR3);
1559	break;
1560	}
1561
1562	case SVM_EXIT_READ_CR0: case SVM_EXIT_READ_CR1: case SVM_EXIT_READ_CR2: case SVM_EXIT_READ_CR3:
1563	case SVM_EXIT_READ_CR4: case SVM_EXIT_READ_CR5: case SVM_EXIT_READ_CR6: case SVM_EXIT_READ_CR7:
1564	case SVM_EXIT_READ_CR8: case SVM_EXIT_READ_CR9: case SVM_EXIT_READ_CR10: case SVM_EXIT_READ_CR11:
1565	case SVM_EXIT_READ_CR12: case SVM_EXIT_READ_CR13: case SVM_EXIT_READ_CR14: case SVM_EXIT_READ_CR15:
1566	{
1567	uint32_t cbSize;
1568
1569	Log2(("SVM: %VGv mov x, cr%d\n", pCtx->rip, exitCode - SVM_EXIT_READ_CR0));
1570	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitCRxRead);
1571	rc = EMInterpretInstruction(pVM, CPUMCTX2CORE(pCtx), 0, &cbSize);
1572	if (rc == VINF_SUCCESS)
1573	{
1574	/* EIP has been updated already. */
1575
1576	/* Only resume if successful. */
1577	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1578	goto ResumeExecution;
1579	}
1580	Assert(rc == VERR_EM_INTERPRETER \|\| rc == VINF_PGM_CHANGE_MODE \|\| rc == VINF_PGM_SYNC_CR3);
1581	break;
1582	}
1583
1584	case SVM_EXIT_WRITE_DR0: case SVM_EXIT_WRITE_DR1: case SVM_EXIT_WRITE_DR2: case SVM_EXIT_WRITE_DR3:
1585	case SVM_EXIT_WRITE_DR4: case SVM_EXIT_WRITE_DR5: case SVM_EXIT_WRITE_DR6: case SVM_EXIT_WRITE_DR7:
1586	case SVM_EXIT_WRITE_DR8: case SVM_EXIT_WRITE_DR9: case SVM_EXIT_WRITE_DR10: case SVM_EXIT_WRITE_DR11:
1587	case SVM_EXIT_WRITE_DR12: case SVM_EXIT_WRITE_DR13: case SVM_EXIT_WRITE_DR14: case SVM_EXIT_WRITE_DR15:
1588	{
1589	uint32_t cbSize;
1590
1591	Log2(("SVM: %VGv mov dr%d, x\n", pCtx->rip, exitCode - SVM_EXIT_WRITE_DR0));
1592	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitDRxRead);
1593	rc = EMInterpretInstruction(pVM, CPUMCTX2CORE(pCtx), 0, &cbSize);
1594	if (rc == VINF_SUCCESS)
1595	{
1596	/* EIP has been updated already. */
1597
1598	/* Only resume if successful. */
1599	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1600	goto ResumeExecution;
1601	}
1602	Assert(rc == VERR_EM_INTERPRETER \|\| rc == VINF_PGM_CHANGE_MODE \|\| rc == VINF_PGM_SYNC_CR3);
1603	break;
1604	}
1605
1606	case SVM_EXIT_READ_DR0: case SVM_EXIT_READ_DR1: case SVM_EXIT_READ_DR2: case SVM_EXIT_READ_DR3:
1607	case SVM_EXIT_READ_DR4: case SVM_EXIT_READ_DR5: case SVM_EXIT_READ_DR6: case SVM_EXIT_READ_DR7:
1608	case SVM_EXIT_READ_DR8: case SVM_EXIT_READ_DR9: case SVM_EXIT_READ_DR10: case SVM_EXIT_READ_DR11:
1609	case SVM_EXIT_READ_DR12: case SVM_EXIT_READ_DR13: case SVM_EXIT_READ_DR14: case SVM_EXIT_READ_DR15:
1610	{
1611	uint32_t cbSize;
1612
1613	Log2(("SVM: %VGv mov dr%d, x\n", pCtx->rip, exitCode - SVM_EXIT_READ_DR0));
1614	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitDRxRead);
1615	rc = EMInterpretInstruction(pVM, CPUMCTX2CORE(pCtx), 0, &cbSize);
1616	if (rc == VINF_SUCCESS)
1617	{
1618	/* EIP has been updated already. */
1619
1620	/* Only resume if successful. */
1621	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1622	goto ResumeExecution;
1623	}
1624	Assert(rc == VERR_EM_INTERPRETER \|\| rc == VINF_PGM_CHANGE_MODE \|\| rc == VINF_PGM_SYNC_CR3);
1625	break;
1626	}
1627
1628	/* Note: We'll get a #GP if the IO instruction isn't allowed (IOPL or TSS bitmap); no need to double check. */
1629	case SVM_EXIT_IOIO: /* I/O instruction. */
1630	{
1631	SVM_IOIO_EXIT IoExitInfo;
1632	uint32_t uIOSize, uAndVal;
1633
1634	IoExitInfo.au32[0] = pVMCB->ctrl.u64ExitInfo1;
1635
1636	/** @todo could use a lookup table here */
1637	if (IoExitInfo.n.u1OP8)
1638	{
1639	uIOSize = 1;
1640	uAndVal = 0xff;
1641	}
1642	else
1643	if (IoExitInfo.n.u1OP16)
1644	{
1645	uIOSize = 2;
1646	uAndVal = 0xffff;
1647	}
1648	else
1649	if (IoExitInfo.n.u1OP32)
1650	{
1651	uIOSize = 4;
1652	uAndVal = 0xffffffff;
1653	}
1654	else
1655	{
1656	AssertFailed(); /* should be fatal. */
1657	rc = VINF_EM_RAW_EMULATE_INSTR;
1658	break;
1659	}
1660
1661	if (IoExitInfo.n.u1STR)
1662	{
1663	/* ins/outs */
1664	uint32_t prefix = 0;
1665	if (IoExitInfo.n.u1REP)
1666	prefix \|= PREFIX_REP;
1667
1668	if (IoExitInfo.n.u1Type == 0)
1669	{
1670	Log2(("IOMInterpretOUTSEx %VGv %x size=%d\n", pCtx->rip, IoExitInfo.n.u16Port, uIOSize));
1671	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitIOStringWrite);
1672	rc = IOMInterpretOUTSEx(pVM, CPUMCTX2CORE(pCtx), IoExitInfo.n.u16Port, prefix, uIOSize);
1673	}
1674	else
1675	{
1676	Log2(("IOMInterpretINSEx %VGv %x size=%d\n", pCtx->rip, IoExitInfo.n.u16Port, uIOSize));
1677	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitIOStringRead);
1678	rc = IOMInterpretINSEx(pVM, CPUMCTX2CORE(pCtx), IoExitInfo.n.u16Port, prefix, uIOSize);
1679	}
1680	}
1681	else
1682	{
1683	/* normal in/out */
1684	Assert(!IoExitInfo.n.u1REP);
1685
1686	if (IoExitInfo.n.u1Type == 0)
1687	{
1688	Log2(("IOMIOPortWrite %VGv %x %x size=%d\n", pCtx->rip, IoExitInfo.n.u16Port, pCtx->eax & uAndVal, uIOSize));
1689	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitIOWrite);
1690	rc = IOMIOPortWrite(pVM, IoExitInfo.n.u16Port, pCtx->eax & uAndVal, uIOSize);
1691	}
1692	else
1693	{
1694	uint32_t u32Val = 0;
1695
1696	STAM_COUNTER_INC(&pVM->hwaccm.s.StatExitIORead);
1697	rc = IOMIOPortRead(pVM, IoExitInfo.n.u16Port, &u32Val, uIOSize);
1698	if (IOM_SUCCESS(rc))
1699	{
1700	/* Write back to the EAX register. */
1701	pCtx->eax = (pCtx->eax & ~uAndVal) \| (u32Val & uAndVal);
1702	Log2(("IOMIOPortRead %VGv %x %x size=%d\n", pCtx->rip, IoExitInfo.n.u16Port, u32Val & uAndVal, uIOSize));
1703	}
1704	}
1705	}
1706	/*
1707	* Handled the I/O return codes.
1708	* (The unhandled cases end up with rc == VINF_EM_RAW_EMULATE_INSTR.)
1709	*/
1710	if (IOM_SUCCESS(rc))
1711	{
1712	/* Update EIP and continue execution. */
1713	pCtx->rip = pVMCB->ctrl.u64ExitInfo2; /* RIP/EIP of the next instruction is saved in EXITINFO2. */
1714	if (RT_LIKELY(rc == VINF_SUCCESS))
1715	{
1716	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1717	goto ResumeExecution;
1718	}
1719	Log2(("EM status from IO at %VGv %x size %d: %Vrc\n", pCtx->rip, IoExitInfo.n.u16Port, uIOSize, rc));
1720	break;
1721	}
1722
1723	#ifdef VBOX_STRICT
1724	if (rc == VINF_IOM_HC_IOPORT_READ)
1725	Assert(IoExitInfo.n.u1Type != 0);
1726	else if (rc == VINF_IOM_HC_IOPORT_WRITE)
1727	Assert(IoExitInfo.n.u1Type == 0);
1728	else
1729	AssertMsg(VBOX_FAILURE(rc) \|\| rc == VINF_EM_RAW_EMULATE_INSTR \|\| rc == VINF_EM_RAW_GUEST_TRAP \|\| rc == VINF_TRPM_XCPT_DISPATCHED, ("%Vrc\n", rc));
1730	#endif
1731	Log2(("Failed IO at %VGv %x size %d\n", pCtx->rip, IoExitInfo.n.u16Port, uIOSize));
1732	break;
1733	}
1734
1735	case SVM_EXIT_HLT:
1736	/** Check if external interrupts are pending; if so, don't switch back. */
1737	if ( pCtx->eflags.Bits.u1IF
1738	&& VM_FF_ISPENDING(pVM, (VM_FF_INTERRUPT_APIC\|VM_FF_INTERRUPT_PIC)))
1739	{
1740	pCtx->rip++; /* skip hlt */
1741	goto ResumeExecution;
1742	}
1743
1744	rc = VINF_EM_RAW_EMULATE_INSTR_HLT;
1745	break;
1746
1747	case SVM_EXIT_RSM:
1748	case SVM_EXIT_INVLPGA:
1749	case SVM_EXIT_VMRUN:
1750	case SVM_EXIT_VMMCALL:
1751	case SVM_EXIT_VMLOAD:
1752	case SVM_EXIT_VMSAVE:
1753	case SVM_EXIT_STGI:
1754	case SVM_EXIT_CLGI:
1755	case SVM_EXIT_SKINIT:
1756	case SVM_EXIT_RDTSCP:
1757	{
1758	/* Unsupported instructions. */
1759	SVM_EVENT Event;
1760
1761	Event.au64[0] = 0;
1762	Event.n.u3Type = SVM_EVENT_EXCEPTION;
1763	Event.n.u1Valid = 1;
1764	Event.n.u8Vector = X86_XCPT_UD;
1765
1766	Log(("Forced #UD trap at %VGv\n", pCtx->rip));
1767	SVMR0InjectEvent(pVM, pVMCB, pCtx, &Event);
1768
1769	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1770	goto ResumeExecution;
1771	}
1772
1773	/* Emulate in ring 3. */
1774	case SVM_EXIT_MSR:
1775	{
1776	uint32_t cbSize;
1777
1778	/* Note: the intel manual claims there's a REX version of RDMSR that's slightly different, so we play safe by completely disassembling the instruction. */
1779	Log(("SVM: %s\n", (pVMCB->ctrl.u64ExitInfo1 == 0) ? "rdmsr" : "wrmsr"));
1780	rc = EMInterpretInstruction(pVM, CPUMCTX2CORE(pCtx), 0, &cbSize);
1781	if (rc == VINF_SUCCESS)
1782	{
1783	/* EIP has been updated already. */
1784
1785	/* Only resume if successful. */
1786	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1787	goto ResumeExecution;
1788	}
1789	AssertMsg(rc == VERR_EM_INTERPRETER, ("EMU: %s failed with %Vrc\n", (pVMCB->ctrl.u64ExitInfo1 == 0) ? "rdmsr" : "wrmsr", rc));
1790	break;
1791	}
1792
1793	case SVM_EXIT_MONITOR:
1794	case SVM_EXIT_RDPMC:
1795	case SVM_EXIT_PAUSE:
1796	case SVM_EXIT_MWAIT_UNCOND:
1797	case SVM_EXIT_MWAIT_ARMED:
1798	case SVM_EXIT_TASK_SWITCH: /* can change CR3; emulate */
1799	rc = VINF_EM_RAW_EXCEPTION_PRIVILEGED;
1800	break;
1801
1802	case SVM_EXIT_SHUTDOWN:
1803	rc = VINF_EM_RESET; /* Triple fault equals a reset. */
1804	break;
1805
1806	case SVM_EXIT_IDTR_READ:
1807	case SVM_EXIT_GDTR_READ:
1808	case SVM_EXIT_LDTR_READ:
1809	case SVM_EXIT_TR_READ:
1810	case SVM_EXIT_IDTR_WRITE:
1811	case SVM_EXIT_GDTR_WRITE:
1812	case SVM_EXIT_LDTR_WRITE:
1813	case SVM_EXIT_TR_WRITE:
1814	case SVM_EXIT_CR0_SEL_WRITE:
1815	default:
1816	/* Unexpected exit codes. */
1817	rc = VERR_EM_INTERNAL_ERROR;
1818	AssertMsgFailed(("Unexpected exit code %x\n", exitCode)); /* Can't happen. */
1819	break;
1820	}
1821
1822	end:
1823	if (fGuestStateSynced)
1824	{
1825	/* Remaining guest CPU context: TR, IDTR, GDTR, LDTR. */
1826	SVM_READ_SELREG(LDTR, ldtr);
1827	SVM_READ_SELREG(TR, tr);
1828
1829	pCtx->gdtr.cbGdt = pVMCB->guest.GDTR.u32Limit;
1830	pCtx->gdtr.pGdt = pVMCB->guest.GDTR.u64Base;
1831
1832	pCtx->idtr.cbIdt = pVMCB->guest.IDTR.u32Limit;
1833	pCtx->idtr.pIdt = pVMCB->guest.IDTR.u64Base;
1834
1835	/*
1836	* System MSRs
1837	*/
1838	pCtx->SysEnter.cs = pVMCB->guest.u64SysEnterCS;
1839	pCtx->SysEnter.eip = pVMCB->guest.u64SysEnterEIP;
1840	pCtx->SysEnter.esp = pVMCB->guest.u64SysEnterESP;
1841	}
1842
1843	/* Signal changes for the recompiler. */
1844	CPUMSetChangedFlags(pVM, CPUM_CHANGED_SYSENTER_MSR \| CPUM_CHANGED_LDTR \| CPUM_CHANGED_GDTR \| CPUM_CHANGED_IDTR \| CPUM_CHANGED_TR \| CPUM_CHANGED_HIDDEN_SEL_REGS);
1845
1846	/* If we executed vmrun and an external irq was pending, then we don't have to do a full sync the next time. */
1847	if (exitCode == SVM_EXIT_INTR)
1848	{
1849	STAM_COUNTER_INC(&pVM->hwaccm.s.StatPendingHostIrq);
1850	/* On the next entry we'll only sync the host context. */
1851	pVM->hwaccm.s.fContextUseFlags \|= HWACCM_CHANGED_HOST_CONTEXT;
1852	}
1853	else
1854	{
1855	/* On the next entry we'll sync everything. */
1856	/** @todo we can do better than this */
1857	pVM->hwaccm.s.fContextUseFlags \|= HWACCM_CHANGED_ALL;
1858	}
1859
1860	/* translate into a less severe return code */
1861	if (rc == VERR_EM_INTERPRETER)
1862	rc = VINF_EM_RAW_EMULATE_INSTR;
1863
1864	STAM_PROFILE_ADV_STOP(&pVM->hwaccm.s.StatExit, x);
1865	return rc;
1866	}
1867
1868	/**
1869	* Enters the AMD-V session
1870	*
1871	* @returns VBox status code.
1872	* @param pVM The VM to operate on.
1873	* @param pCpu CPU info struct
1874	*/
1875	HWACCMR0DECL(int) SVMR0Enter(PVM pVM, PHWACCM_CPUINFO pCpu)
1876	{
1877	Assert(pVM->hwaccm.s.svm.fSupported);
1878
1879	LogFlow(("SVMR0Enter cpu%d last=%d asid=%d\n", pCpu->idCpu, pVM->hwaccm.s.svm.idLastCpu, pCpu->uCurrentASID));
1880	pVM->hwaccm.s.svm.fResumeVM = false;
1881
1882	/* Force to reload LDTR, so we'll execute VMLoad to load additional guest state. */
1883	pVM->hwaccm.s.fContextUseFlags \|= HWACCM_CHANGED_GUEST_LDTR;
1884
1885	return VINF_SUCCESS;
1886	}
1887
1888
1889	/**
1890	* Leaves the AMD-V session
1891	*
1892	* @returns VBox status code.
1893	* @param pVM The VM to operate on.
1894	*/
1895	HWACCMR0DECL(int) SVMR0Leave(PVM pVM)
1896	{
1897	Assert(pVM->hwaccm.s.svm.fSupported);
1898	return VINF_SUCCESS;
1899	}
1900
1901
1902	static int svmInterpretInvlPg(PVM pVM, PDISCPUSTATE pCpu, PCPUMCTXCORE pRegFrame, uint32_t uASID)
1903	{
1904	OP_PARAMVAL param1;
1905	RTGCPTR addr;
1906
1907	int rc = DISQueryParamVal(pRegFrame, pCpu, &pCpu->param1, &param1, PARAM_SOURCE);
1908	if(VBOX_FAILURE(rc))
1909	return VERR_EM_INTERPRETER;
1910
1911	switch(param1.type)
1912	{
1913	case PARMTYPE_IMMEDIATE:
1914	case PARMTYPE_ADDRESS:
1915	if(!(param1.flags & (PARAM_VAL32\|PARAM_VAL64)))
1916	return VERR_EM_INTERPRETER;
1917	addr = param1.val.val64;
1918	break;
1919
1920	default:
1921	return VERR_EM_INTERPRETER;
1922	}
1923
1924	/** @todo is addr always a flat linear address or ds based
1925	* (in absence of segment override prefixes)????
1926	*/
1927	rc = PGMInvalidatePage(pVM, addr);
1928	if (VBOX_SUCCESS(rc))
1929	{
1930	/* Manually invalidate the page for the VM's TLB. */
1931	Log(("SVMInvlpgA %VGv ASID=%d\n", addr, uASID));
1932	SVMInvlpgA(addr, uASID);
1933	return VINF_SUCCESS;
1934	}
1935	Assert(rc == VERR_REM_FLUSHED_PAGES_OVERFLOW);
1936	return rc;
1937	}
1938
1939	/**
1940	* Interprets INVLPG
1941	*
1942	* @returns VBox status code.
1943	* @retval VINF_* Scheduling instructions.
1944	* @retval VERR_EM_INTERPRETER Something we can't cope with.
1945	* @retval VERR_* Fatal errors.
1946	*
1947	* @param pVM The VM handle.
1948	* @param pRegFrame The register frame.
1949	* @param ASID Tagged TLB id for the guest
1950	*
1951	* Updates the EIP if an instruction was executed successfully.
1952	*/
1953	static int SVMR0InterpretInvpg(PVM pVM, PCPUMCTXCORE pRegFrame, uint32_t uASID)
1954	{
1955	/*
1956	* Only allow 32 & 64 bits code.
1957	*/
1958	DISCPUMODE enmMode = SELMGetCpuModeFromSelector(pVM, pRegFrame->eflags, pRegFrame->cs, &pRegFrame->csHid);
1959	if (enmMode != CPUMODE_16BIT)
1960	{
1961	RTGCPTR pbCode;
1962	int rc = SELMValidateAndConvertCSAddr(pVM, pRegFrame->eflags, pRegFrame->ss, pRegFrame->cs, &pRegFrame->csHid, (RTGCPTR)pRegFrame->rip, &pbCode);
1963	if (VBOX_SUCCESS(rc))
1964	{
1965	uint32_t cbOp;
1966	DISCPUSTATE Cpu;
1967
1968	Cpu.mode = enmMode;
1969	rc = EMInterpretDisasOneEx(pVM, pbCode, pRegFrame, &Cpu, &cbOp);
1970	Assert(VBOX_FAILURE(rc) \|\| Cpu.pCurInstr->opcode == OP_INVLPG);
1971	if (VBOX_SUCCESS(rc) && Cpu.pCurInstr->opcode == OP_INVLPG)
1972	{
1973	Assert(cbOp == Cpu.opsize);
1974	rc = svmInterpretInvlPg(pVM, &Cpu, pRegFrame, uASID);
1975	if (VBOX_SUCCESS(rc))
1976	{
1977	pRegFrame->rip += cbOp; /* Move on to the next instruction. */
1978	}
1979	return rc;
1980	}
1981	}
1982	}
1983	return VERR_EM_INTERPRETER;
1984	}
1985
1986
1987	/**
1988	* Invalidates a guest page
1989	*
1990	* @returns VBox status code.
1991	* @param pVM The VM to operate on.
1992	* @param GCVirt Page to invalidate
1993	*/
1994	HWACCMR0DECL(int) SVMR0InvalidatePage(PVM pVM, RTGCPTR GCVirt)
1995	{
1996	bool fFlushPending = pVM->hwaccm.s.svm.fAlwaysFlushTLB \| pVM->hwaccm.s.svm.fForceTLBFlush;
1997
1998	/* Skip it if a TLB flush is already pending. */
1999	if (!fFlushPending)
2000	{
2001	SVM_VMCB *pVMCB;
2002
2003	Log2(("SVMR0InvalidatePage %VGv\n", GCVirt));
2004	AssertReturn(pVM, VERR_INVALID_PARAMETER);
2005	Assert(pVM->hwaccm.s.svm.fSupported);
2006
2007	pVMCB = (SVM_VMCB *)pVM->hwaccm.s.svm.pVMCB;
2008	AssertMsgReturn(pVMCB, ("Invalid pVMCB\n"), VERR_EM_INTERNAL_ERROR);
2009
2010	STAM_COUNTER_INC(&pVM->hwaccm.s.StatFlushPageManual);
2011	SVMInvlpgA(GCVirt, pVMCB->ctrl.TLBCtrl.n.u32ASID);
2012	}
2013	return VINF_SUCCESS;
2014	}
2015
2016
2017	/**
2018	* Invalidates a guest page by physical address
2019	*
2020	* NOTE: Assumes the current instruction references this physical page though a virtual address!!
2021	*
2022	* @returns VBox status code.
2023	* @param pVM The VM to operate on.
2024	* @param GCPhys Page to invalidate
2025	*/
2026	HWACCMR0DECL(int) SVMR0InvalidatePhysPage(PVM pVM, RTGCPHYS GCPhys)
2027	{
2028	bool fFlushPending = pVM->hwaccm.s.svm.fAlwaysFlushTLB \| pVM->hwaccm.s.svm.fForceTLBFlush;
2029
2030	Assert(pVM->hwaccm.s.fNestedPaging);
2031
2032	/* Skip it if a TLB flush is already pending. */
2033	if (!fFlushPending)
2034	{
2035	CPUMCTX *pCtx;
2036	int rc;
2037	SVM_VMCB *pVMCB;
2038
2039	rc = CPUMQueryGuestCtxPtr(pVM, &pCtx);
2040	AssertRCReturn(rc, rc);
2041
2042	Log2(("SVMR0InvalidatePhysPage %VGp\n", GCPhys));
2043	AssertReturn(pVM, VERR_INVALID_PARAMETER);
2044	Assert(pVM->hwaccm.s.svm.fSupported);
2045
2046	pVMCB = (SVM_VMCB *)pVM->hwaccm.s.svm.pVMCB;
2047	AssertMsgReturn(pVMCB, ("Invalid pVMCB\n"), VERR_EM_INTERNAL_ERROR);
2048
2049	/*
2050	* Only allow 32 & 64 bits code.
2051	*/
2052	DISCPUMODE enmMode = SELMGetCpuModeFromSelector(pVM, pCtx->eflags, pCtx->cs, &pCtx->csHid);
2053	if (enmMode != CPUMODE_16BIT)
2054	{
2055	RTGCPTR pbCode;
2056	int rc = SELMValidateAndConvertCSAddr(pVM, pCtx->eflags, pCtx->ss, pCtx->cs, &pCtx->csHid, (RTGCPTR)pCtx->rip, &pbCode);
2057	if (VBOX_SUCCESS(rc))
2058	{
2059	uint32_t cbOp;
2060	DISCPUSTATE Cpu;
2061	OP_PARAMVAL param1;
2062	RTGCPTR addr;
2063
2064	Cpu.mode = enmMode;
2065	rc = EMInterpretDisasOneEx(pVM, pbCode, CPUMCTX2CORE(pCtx), &Cpu, &cbOp);
2066	AssertRCReturn(rc, rc);
2067	Assert(cbOp == Cpu.opsize);
2068
2069	int rc = DISQueryParamVal(CPUMCTX2CORE(pCtx), &Cpu, &Cpu.param1, &param1, PARAM_SOURCE);
2070	AssertRCReturn(rc, VERR_EM_INTERPRETER);
2071
2072	switch(param1.type)
2073	{
2074	case PARMTYPE_IMMEDIATE:
2075	case PARMTYPE_ADDRESS:
2076	AssertReturn((param1.flags & (PARAM_VAL32\|PARAM_VAL64)), VERR_EM_INTERPRETER);
2077
2078	addr = param1.val.val64;
2079	break;
2080
2081	default:
2082	AssertFailed();
2083	return VERR_EM_INTERPRETER;
2084	}
2085
2086	/* Manually invalidate the page for the VM's TLB. */
2087	Log(("SVMR0InvalidatePhysPage Phys=%VGp Virt=%VGv ASID=%d\n", GCPhys, addr, pVMCB->ctrl.TLBCtrl.n.u32ASID));
2088	SVMInvlpgA(addr, pVMCB->ctrl.TLBCtrl.n.u32ASID);
2089	STAM_COUNTER_INC(&pVM->hwaccm.s.StatFlushPhysPageManual);
2090
2091	return VINF_SUCCESS;
2092	}
2093	}
2094	AssertFailed();
2095	return VERR_EM_INTERPRETER;
2096	}
2097	return VINF_SUCCESS;
2098	}

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

source: vbox/trunk/src/VBox/VMM/VMMR0/HWSVMR0.cpp@ 10458

Download in other formats: