VirtualBox

source: vbox/trunk/src/VBox/VMM/VMMR0/HMSVMR0.cpp@ 55435

Last change on this file since 55435 was 55316, checked in by vboxsync, 10 years ago

CPUM,IEM,HM: Enabled XSAVE and AVX for AMD-V too.

  • Property svn:eol-style set to native
  • Property svn:keywords set to Author Date Id Revision
File size: 202.4 KB
Line 
1/* $Id: HMSVMR0.cpp 55316 2015-04-16 17:34:30Z vboxsync $ */
2/** @file
3 * HM SVM (AMD-V) - Host Context Ring-0.
4 */
5
6/*
7 * Copyright (C) 2013-2015 Oracle Corporation
8 *
9 * This file is part of VirtualBox Open Source Edition (OSE), as
10 * available from http://www.virtualbox.org. This file is free software;
11 * you can redistribute it and/or modify it under the terms of the GNU
12 * General Public License (GPL) as published by the Free Software
13 * Foundation, in version 2 as it comes in the "COPYING" file of the
14 * VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15 * hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16 */
17
18/*******************************************************************************
19* Header Files *
20*******************************************************************************/
21#define LOG_GROUP LOG_GROUP_HM
22#include <iprt/asm-amd64-x86.h>
23#include <iprt/thread.h>
24
25#include <VBox/vmm/pdmapi.h>
26#include <VBox/vmm/dbgf.h>
27#include <VBox/vmm/iem.h>
28#include <VBox/vmm/iom.h>
29#include <VBox/vmm/tm.h>
30#include <VBox/vmm/gim.h>
31#include "HMInternal.h"
32#include <VBox/vmm/vm.h>
33#include "HMSVMR0.h"
34#include "dtrace/VBoxVMM.h"
35
36#ifdef DEBUG_ramshankar
37# define HMSVM_SYNC_FULL_GUEST_STATE
38# define HMSVM_ALWAYS_TRAP_ALL_XCPTS
39# define HMSVM_ALWAYS_TRAP_PF
40# define HMSVM_ALWAYS_TRAP_TASK_SWITCH
41#endif
42
43
44/*******************************************************************************
45* Defined Constants And Macros *
46*******************************************************************************/
47#ifdef VBOX_WITH_STATISTICS
48# define HMSVM_EXITCODE_STAM_COUNTER_INC(u64ExitCode) do { \
49 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitAll); \
50 if ((u64ExitCode) == SVM_EXIT_NPF) \
51 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitReasonNpf); \
52 else \
53 STAM_COUNTER_INC(&pVCpu->hm.s.paStatExitReasonR0[(u64ExitCode) & MASK_EXITREASON_STAT]); \
54 } while (0)
55#else
56# define HMSVM_EXITCODE_STAM_COUNTER_INC(u64ExitCode) do { } while (0)
57#endif
58
59/** If we decide to use a function table approach this can be useful to
60 * switch to a "static DECLCALLBACK(int)". */
61#define HMSVM_EXIT_DECL static int
62
63/** @name Segment attribute conversion between CPU and AMD-V VMCB format.
64 *
65 * The CPU format of the segment attribute is described in X86DESCATTRBITS
66 * which is 16-bits (i.e. includes 4 bits of the segment limit).
67 *
68 * The AMD-V VMCB format the segment attribute is compact 12-bits (strictly
69 * only the attribute bits and nothing else). Upper 4-bits are unused.
70 *
71 * @{ */
72#define HMSVM_CPU_2_VMCB_SEG_ATTR(a) ( ((a) & 0xff) | (((a) & 0xf000) >> 4) )
73#define HMSVM_VMCB_2_CPU_SEG_ATTR(a) ( ((a) & 0xff) | (((a) & 0x0f00) << 4) )
74/** @} */
75
76/** @name Macros for loading, storing segment registers to/from the VMCB.
77 * @{ */
78#define HMSVM_LOAD_SEG_REG(REG, reg) \
79 do \
80 { \
81 Assert(pCtx->reg.fFlags & CPUMSELREG_FLAGS_VALID); \
82 Assert(pCtx->reg.ValidSel == pCtx->reg.Sel); \
83 pVmcb->guest.REG.u16Sel = pCtx->reg.Sel; \
84 pVmcb->guest.REG.u32Limit = pCtx->reg.u32Limit; \
85 pVmcb->guest.REG.u64Base = pCtx->reg.u64Base; \
86 pVmcb->guest.REG.u16Attr = HMSVM_CPU_2_VMCB_SEG_ATTR(pCtx->reg.Attr.u); \
87 } while (0)
88
89#define HMSVM_SAVE_SEG_REG(REG, reg) \
90 do \
91 { \
92 pMixedCtx->reg.Sel = pVmcb->guest.REG.u16Sel; \
93 pMixedCtx->reg.ValidSel = pVmcb->guest.REG.u16Sel; \
94 pMixedCtx->reg.fFlags = CPUMSELREG_FLAGS_VALID; \
95 pMixedCtx->reg.u32Limit = pVmcb->guest.REG.u32Limit; \
96 pMixedCtx->reg.u64Base = pVmcb->guest.REG.u64Base; \
97 pMixedCtx->reg.Attr.u = HMSVM_VMCB_2_CPU_SEG_ATTR(pVmcb->guest.REG.u16Attr); \
98 } while (0)
99/** @} */
100
101/** Macro for checking and returning from the using function for
102 * \#VMEXIT intercepts that maybe caused during delivering of another
103 * event in the guest. */
104#define HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY() \
105 do \
106 { \
107 int rc = hmR0SvmCheckExitDueToEventDelivery(pVCpu, pCtx, pSvmTransient); \
108 if (RT_UNLIKELY(rc == VINF_HM_DOUBLE_FAULT)) \
109 return VINF_SUCCESS; \
110 else if (RT_UNLIKELY(rc == VINF_EM_RESET)) \
111 return rc; \
112 } while (0)
113
114/** Macro for upgrading a @a a_rc to VINF_EM_DBG_STEPPED after emulating an
115 * instruction that exited. */
116#define HMSVM_CHECK_SINGLE_STEP(a_pVCpu, a_rc) \
117 do { \
118 if ((a_pVCpu)->hm.s.fSingleInstruction && (a_rc) == VINF_SUCCESS) \
119 (a_rc) = VINF_EM_DBG_STEPPED; \
120 } while (0)
121
122/** Assert that preemption is disabled or covered by thread-context hooks. */
123#define HMSVM_ASSERT_PREEMPT_SAFE() Assert( VMMR0ThreadCtxHooksAreRegistered(pVCpu) \
124 || !RTThreadPreemptIsEnabled(NIL_RTTHREAD));
125
126/** Assert that we haven't migrated CPUs when thread-context hooks are not
127 * used. */
128#define HMSVM_ASSERT_CPU_SAFE() AssertMsg( VMMR0ThreadCtxHooksAreRegistered(pVCpu) \
129 || pVCpu->hm.s.idEnteredCpu == RTMpCpuId(), \
130 ("Illegal migration! Entered on CPU %u Current %u\n", \
131 pVCpu->hm.s.idEnteredCpu, RTMpCpuId()));
132
133/** Exception bitmap mask for all contributory exceptions.
134 *
135 * Page fault is deliberately excluded here as it's conditional as to whether
136 * it's contributory or benign. Page faults are handled separately.
137 */
138#define HMSVM_CONTRIBUTORY_XCPT_MASK ( RT_BIT(X86_XCPT_GP) | RT_BIT(X86_XCPT_NP) | RT_BIT(X86_XCPT_SS) | RT_BIT(X86_XCPT_TS) \
139 | RT_BIT(X86_XCPT_DE))
140
141/** @name VMCB Clean Bits.
142 *
143 * These flags are used for VMCB-state caching. A set VMCB Clean bit indicates
144 * AMD-V doesn't need to reload the corresponding value(s) from the VMCB in
145 * memory.
146 *
147 * @{ */
148/** All intercepts vectors, TSC offset, PAUSE filter counter. */
149#define HMSVM_VMCB_CLEAN_INTERCEPTS RT_BIT(0)
150/** I/O permission bitmap, MSR permission bitmap. */
151#define HMSVM_VMCB_CLEAN_IOPM_MSRPM RT_BIT(1)
152/** ASID. */
153#define HMSVM_VMCB_CLEAN_ASID RT_BIT(2)
154/** TRP: V_TPR, V_IRQ, V_INTR_PRIO, V_IGN_TPR, V_INTR_MASKING,
155V_INTR_VECTOR. */
156#define HMSVM_VMCB_CLEAN_TPR RT_BIT(3)
157/** Nested Paging: Nested CR3 (nCR3), PAT. */
158#define HMSVM_VMCB_CLEAN_NP RT_BIT(4)
159/** Control registers (CR0, CR3, CR4, EFER). */
160#define HMSVM_VMCB_CLEAN_CRX_EFER RT_BIT(5)
161/** Debug registers (DR6, DR7). */
162#define HMSVM_VMCB_CLEAN_DRX RT_BIT(6)
163/** GDT, IDT limit and base. */
164#define HMSVM_VMCB_CLEAN_DT RT_BIT(7)
165/** Segment register: CS, SS, DS, ES limit and base. */
166#define HMSVM_VMCB_CLEAN_SEG RT_BIT(8)
167/** CR2.*/
168#define HMSVM_VMCB_CLEAN_CR2 RT_BIT(9)
169/** Last-branch record (DbgCtlMsr, br_from, br_to, lastint_from, lastint_to) */
170#define HMSVM_VMCB_CLEAN_LBR RT_BIT(10)
171/** AVIC (AVIC APIC_BAR; AVIC APIC_BACKING_PAGE, AVIC
172PHYSICAL_TABLE and AVIC LOGICAL_TABLE Pointers). */
173#define HMSVM_VMCB_CLEAN_AVIC RT_BIT(11)
174/** Mask of all valid VMCB Clean bits. */
175#define HMSVM_VMCB_CLEAN_ALL ( HMSVM_VMCB_CLEAN_INTERCEPTS \
176 | HMSVM_VMCB_CLEAN_IOPM_MSRPM \
177 | HMSVM_VMCB_CLEAN_ASID \
178 | HMSVM_VMCB_CLEAN_TPR \
179 | HMSVM_VMCB_CLEAN_NP \
180 | HMSVM_VMCB_CLEAN_CRX_EFER \
181 | HMSVM_VMCB_CLEAN_DRX \
182 | HMSVM_VMCB_CLEAN_DT \
183 | HMSVM_VMCB_CLEAN_SEG \
184 | HMSVM_VMCB_CLEAN_CR2 \
185 | HMSVM_VMCB_CLEAN_LBR \
186 | HMSVM_VMCB_CLEAN_AVIC)
187/** @} */
188
189/** @name SVM transient.
190 *
191 * A state structure for holding miscellaneous information across AMD-V
192 * VMRUN/#VMEXIT operation, restored after the transition.
193 *
194 * @{ */
195typedef struct SVMTRANSIENT
196{
197 /** The host's rflags/eflags. */
198 RTCCUINTREG uEflags;
199#if HC_ARCH_BITS == 32
200 uint32_t u32Alignment0;
201#endif
202
203 /** The #VMEXIT exit code (the EXITCODE field in the VMCB). */
204 uint64_t u64ExitCode;
205 /** The guest's TPR value used for TPR shadowing. */
206 uint8_t u8GuestTpr;
207 /** Alignment. */
208 uint8_t abAlignment0[7];
209
210 /** Whether the guest FPU state was active at the time of #VMEXIT. */
211 bool fWasGuestFPUStateActive;
212 /** Whether the guest debug state was active at the time of #VMEXIT. */
213 bool fWasGuestDebugStateActive;
214 /** Whether the hyper debug state was active at the time of #VMEXIT. */
215 bool fWasHyperDebugStateActive;
216 /** Whether the TSC offset mode needs to be updated. */
217 bool fUpdateTscOffsetting;
218 /** Whether the TSC_AUX MSR needs restoring on #VMEXIT. */
219 bool fRestoreTscAuxMsr;
220 /** Whether the #VMEXIT was caused by a page-fault during delivery of a
221 * contributary exception or a page-fault. */
222 bool fVectoringDoublePF;
223 /** Whether the #VMEXIT was caused by a page-fault during delivery of an
224 * external interrupt or NMI. */
225 bool fVectoringPF;
226} SVMTRANSIENT, *PSVMTRANSIENT;
227AssertCompileMemberAlignment(SVMTRANSIENT, u64ExitCode, sizeof(uint64_t));
228AssertCompileMemberAlignment(SVMTRANSIENT, fWasGuestFPUStateActive, sizeof(uint64_t));
229/** @} */
230
231/**
232 * MSRPM (MSR permission bitmap) read permissions (for guest RDMSR).
233 */
234typedef enum SVMMSREXITREAD
235{
236 /** Reading this MSR causes a #VMEXIT. */
237 SVMMSREXIT_INTERCEPT_READ = 0xb,
238 /** Reading this MSR does not cause a #VMEXIT. */
239 SVMMSREXIT_PASSTHRU_READ
240} SVMMSREXITREAD;
241
242/**
243 * MSRPM (MSR permission bitmap) write permissions (for guest WRMSR).
244 */
245typedef enum SVMMSREXITWRITE
246{
247 /** Writing to this MSR causes a #VMEXIT. */
248 SVMMSREXIT_INTERCEPT_WRITE = 0xd,
249 /** Writing to this MSR does not cause a #VMEXIT. */
250 SVMMSREXIT_PASSTHRU_WRITE
251} SVMMSREXITWRITE;
252
253/**
254 * SVM #VMEXIT handler.
255 *
256 * @returns VBox status code.
257 * @param pVCpu Pointer to the VMCPU.
258 * @param pMixedCtx Pointer to the guest-CPU context.
259 * @param pSvmTransient Pointer to the SVM-transient structure.
260 */
261typedef int FNSVMEXITHANDLER(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient);
262
263/*******************************************************************************
264* Internal Functions *
265*******************************************************************************/
266static void hmR0SvmSetMsrPermission(PVMCPU pVCpu, unsigned uMsr, SVMMSREXITREAD enmRead, SVMMSREXITWRITE enmWrite);
267static void hmR0SvmPendingEventToTrpmTrap(PVMCPU pVCpu);
268static void hmR0SvmLeave(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx);
269
270/** @name #VMEXIT handlers.
271 * @{
272 */
273static FNSVMEXITHANDLER hmR0SvmExitIntr;
274static FNSVMEXITHANDLER hmR0SvmExitWbinvd;
275static FNSVMEXITHANDLER hmR0SvmExitInvd;
276static FNSVMEXITHANDLER hmR0SvmExitCpuid;
277static FNSVMEXITHANDLER hmR0SvmExitRdtsc;
278static FNSVMEXITHANDLER hmR0SvmExitRdtscp;
279static FNSVMEXITHANDLER hmR0SvmExitRdpmc;
280static FNSVMEXITHANDLER hmR0SvmExitInvlpg;
281static FNSVMEXITHANDLER hmR0SvmExitHlt;
282static FNSVMEXITHANDLER hmR0SvmExitMonitor;
283static FNSVMEXITHANDLER hmR0SvmExitMwait;
284static FNSVMEXITHANDLER hmR0SvmExitShutdown;
285static FNSVMEXITHANDLER hmR0SvmExitReadCRx;
286static FNSVMEXITHANDLER hmR0SvmExitWriteCRx;
287static FNSVMEXITHANDLER hmR0SvmExitSetPendingXcptUD;
288static FNSVMEXITHANDLER hmR0SvmExitMsr;
289static FNSVMEXITHANDLER hmR0SvmExitReadDRx;
290static FNSVMEXITHANDLER hmR0SvmExitWriteDRx;
291static FNSVMEXITHANDLER hmR0SvmExitXsetbv;
292static FNSVMEXITHANDLER hmR0SvmExitIOInstr;
293static FNSVMEXITHANDLER hmR0SvmExitNestedPF;
294static FNSVMEXITHANDLER hmR0SvmExitVIntr;
295static FNSVMEXITHANDLER hmR0SvmExitTaskSwitch;
296static FNSVMEXITHANDLER hmR0SvmExitVmmCall;
297static FNSVMEXITHANDLER hmR0SvmExitIret;
298static FNSVMEXITHANDLER hmR0SvmExitXcptPF;
299static FNSVMEXITHANDLER hmR0SvmExitXcptNM;
300static FNSVMEXITHANDLER hmR0SvmExitXcptUD;
301static FNSVMEXITHANDLER hmR0SvmExitXcptMF;
302static FNSVMEXITHANDLER hmR0SvmExitXcptDB;
303/** @} */
304
305DECLINLINE(int) hmR0SvmHandleExit(PVMCPU pVCpu, PCPUMCTX pMixedCtx, PSVMTRANSIENT pSvmTransient);
306
307/*******************************************************************************
308* Global Variables *
309*******************************************************************************/
310/** Ring-0 memory object for the IO bitmap. */
311RTR0MEMOBJ g_hMemObjIOBitmap = NIL_RTR0MEMOBJ;
312/** Physical address of the IO bitmap. */
313RTHCPHYS g_HCPhysIOBitmap = 0;
314/** Virtual address of the IO bitmap. */
315R0PTRTYPE(void *) g_pvIOBitmap = NULL;
316
317
318/**
319 * Sets up and activates AMD-V on the current CPU.
320 *
321 * @returns VBox status code.
322 * @param pCpu Pointer to the CPU info struct.
323 * @param pVM Pointer to the VM (can be NULL after a resume!).
324 * @param pvCpuPage Pointer to the global CPU page.
325 * @param HCPhysCpuPage Physical address of the global CPU page.
326 * @param fEnabledByHost Whether the host OS has already initialized AMD-V.
327 * @param pvArg Unused on AMD-V.
328 */
329VMMR0DECL(int) SVMR0EnableCpu(PHMGLOBALCPUINFO pCpu, PVM pVM, void *pvCpuPage, RTHCPHYS HCPhysCpuPage, bool fEnabledByHost,
330 void *pvArg)
331{
332 Assert(!fEnabledByHost);
333 Assert(HCPhysCpuPage && HCPhysCpuPage != NIL_RTHCPHYS);
334 Assert(RT_ALIGN_T(HCPhysCpuPage, _4K, RTHCPHYS) == HCPhysCpuPage);
335 Assert(pvCpuPage);
336 Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
337
338 NOREF(pvArg);
339 NOREF(fEnabledByHost);
340
341 /* Paranoid: Disable interrupt as, in theory, interrupt handlers might mess with EFER. */
342 RTCCUINTREG uEflags = ASMIntDisableFlags();
343
344 /*
345 * We must turn on AMD-V and setup the host state physical address, as those MSRs are per CPU.
346 */
347 uint64_t u64HostEfer = ASMRdMsr(MSR_K6_EFER);
348 if (u64HostEfer & MSR_K6_EFER_SVME)
349 {
350 /* If the VBOX_HWVIRTEX_IGNORE_SVM_IN_USE is active, then we blindly use AMD-V. */
351 if ( pVM
352 && pVM->hm.s.svm.fIgnoreInUseError)
353 {
354 pCpu->fIgnoreAMDVInUseError = true;
355 }
356
357 if (!pCpu->fIgnoreAMDVInUseError)
358 {
359 ASMSetFlags(uEflags);
360 return VERR_SVM_IN_USE;
361 }
362 }
363
364 /* Turn on AMD-V in the EFER MSR. */
365 ASMWrMsr(MSR_K6_EFER, u64HostEfer | MSR_K6_EFER_SVME);
366
367 /* Write the physical page address where the CPU will store the host state while executing the VM. */
368 ASMWrMsr(MSR_K8_VM_HSAVE_PA, HCPhysCpuPage);
369
370 /* Restore interrupts. */
371 ASMSetFlags(uEflags);
372
373 /*
374 * Theoretically, other hypervisors may have used ASIDs, ideally we should flush all non-zero ASIDs
375 * when enabling SVM. AMD doesn't have an SVM instruction to flush all ASIDs (flushing is done
376 * upon VMRUN). Therefore, just set the fFlushAsidBeforeUse flag which instructs hmR0SvmSetupTLB()
377 * to flush the TLB with before using a new ASID.
378 */
379 pCpu->fFlushAsidBeforeUse = true;
380
381 /*
382 * Ensure each VCPU scheduled on this CPU gets a new VPID on resume. See @bugref{6255}.
383 */
384 ++pCpu->cTlbFlushes;
385
386 return VINF_SUCCESS;
387}
388
389
390/**
391 * Deactivates AMD-V on the current CPU.
392 *
393 * @returns VBox status code.
394 * @param pCpu Pointer to the CPU info struct.
395 * @param pvCpuPage Pointer to the global CPU page.
396 * @param HCPhysCpuPage Physical address of the global CPU page.
397 */
398VMMR0DECL(int) SVMR0DisableCpu(PHMGLOBALCPUINFO pCpu, void *pvCpuPage, RTHCPHYS HCPhysCpuPage)
399{
400 Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
401 AssertReturn( HCPhysCpuPage
402 && HCPhysCpuPage != NIL_RTHCPHYS, VERR_INVALID_PARAMETER);
403 AssertReturn(pvCpuPage, VERR_INVALID_PARAMETER);
404 NOREF(pCpu);
405
406 /* Paranoid: Disable interrupts as, in theory, interrupt handlers might mess with EFER. */
407 RTCCUINTREG uEflags = ASMIntDisableFlags();
408
409 /* Turn off AMD-V in the EFER MSR. */
410 uint64_t u64HostEfer = ASMRdMsr(MSR_K6_EFER);
411 ASMWrMsr(MSR_K6_EFER, u64HostEfer & ~MSR_K6_EFER_SVME);
412
413 /* Invalidate host state physical address. */
414 ASMWrMsr(MSR_K8_VM_HSAVE_PA, 0);
415
416 /* Restore interrupts. */
417 ASMSetFlags(uEflags);
418
419 return VINF_SUCCESS;
420}
421
422
423/**
424 * Does global AMD-V initialization (called during module initialization).
425 *
426 * @returns VBox status code.
427 */
428VMMR0DECL(int) SVMR0GlobalInit(void)
429{
430 /*
431 * Allocate 12 KB for the IO bitmap. Since this is non-optional and we always intercept all IO accesses, it's done
432 * once globally here instead of per-VM.
433 */
434 Assert(g_hMemObjIOBitmap == NIL_RTR0MEMOBJ);
435 int rc = RTR0MemObjAllocCont(&g_hMemObjIOBitmap, 3 << PAGE_SHIFT, false /* fExecutable */);
436 if (RT_FAILURE(rc))
437 return rc;
438
439 g_pvIOBitmap = RTR0MemObjAddress(g_hMemObjIOBitmap);
440 g_HCPhysIOBitmap = RTR0MemObjGetPagePhysAddr(g_hMemObjIOBitmap, 0 /* iPage */);
441
442 /* Set all bits to intercept all IO accesses. */
443 ASMMemFill32(g_pvIOBitmap, 3 << PAGE_SHIFT, UINT32_C(0xffffffff));
444 return VINF_SUCCESS;
445}
446
447
448/**
449 * Does global AMD-V termination (called during module termination).
450 */
451VMMR0DECL(void) SVMR0GlobalTerm(void)
452{
453 if (g_hMemObjIOBitmap != NIL_RTR0MEMOBJ)
454 {
455 RTR0MemObjFree(g_hMemObjIOBitmap, true /* fFreeMappings */);
456 g_pvIOBitmap = NULL;
457 g_HCPhysIOBitmap = 0;
458 g_hMemObjIOBitmap = NIL_RTR0MEMOBJ;
459 }
460}
461
462
463/**
464 * Frees any allocated per-VCPU structures for a VM.
465 *
466 * @param pVM Pointer to the VM.
467 */
468DECLINLINE(void) hmR0SvmFreeStructs(PVM pVM)
469{
470 for (uint32_t i = 0; i < pVM->cCpus; i++)
471 {
472 PVMCPU pVCpu = &pVM->aCpus[i];
473 AssertPtr(pVCpu);
474
475 if (pVCpu->hm.s.svm.hMemObjVmcbHost != NIL_RTR0MEMOBJ)
476 {
477 RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjVmcbHost, false);
478 pVCpu->hm.s.svm.pvVmcbHost = 0;
479 pVCpu->hm.s.svm.HCPhysVmcbHost = 0;
480 pVCpu->hm.s.svm.hMemObjVmcbHost = NIL_RTR0MEMOBJ;
481 }
482
483 if (pVCpu->hm.s.svm.hMemObjVmcb != NIL_RTR0MEMOBJ)
484 {
485 RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjVmcb, false);
486 pVCpu->hm.s.svm.pvVmcb = 0;
487 pVCpu->hm.s.svm.HCPhysVmcb = 0;
488 pVCpu->hm.s.svm.hMemObjVmcb = NIL_RTR0MEMOBJ;
489 }
490
491 if (pVCpu->hm.s.svm.hMemObjMsrBitmap != NIL_RTR0MEMOBJ)
492 {
493 RTR0MemObjFree(pVCpu->hm.s.svm.hMemObjMsrBitmap, false);
494 pVCpu->hm.s.svm.pvMsrBitmap = 0;
495 pVCpu->hm.s.svm.HCPhysMsrBitmap = 0;
496 pVCpu->hm.s.svm.hMemObjMsrBitmap = NIL_RTR0MEMOBJ;
497 }
498 }
499}
500
501
502/**
503 * Does per-VM AMD-V initialization.
504 *
505 * @returns VBox status code.
506 * @param pVM Pointer to the VM.
507 */
508VMMR0DECL(int) SVMR0InitVM(PVM pVM)
509{
510 int rc = VERR_INTERNAL_ERROR_5;
511
512 /*
513 * Check for an AMD CPU erratum which requires us to flush the TLB before every world-switch.
514 */
515 uint32_t u32Family;
516 uint32_t u32Model;
517 uint32_t u32Stepping;
518 if (HMAmdIsSubjectToErratum170(&u32Family, &u32Model, &u32Stepping))
519 {
520 Log4(("SVMR0InitVM: AMD cpu with erratum 170 family %#x model %#x stepping %#x\n", u32Family, u32Model, u32Stepping));
521 pVM->hm.s.svm.fAlwaysFlushTLB = true;
522 }
523
524 /*
525 * Initialize the R0 memory objects up-front so we can properly cleanup on allocation failures.
526 */
527 for (VMCPUID i = 0; i < pVM->cCpus; i++)
528 {
529 PVMCPU pVCpu = &pVM->aCpus[i];
530 pVCpu->hm.s.svm.hMemObjVmcbHost = NIL_RTR0MEMOBJ;
531 pVCpu->hm.s.svm.hMemObjVmcb = NIL_RTR0MEMOBJ;
532 pVCpu->hm.s.svm.hMemObjMsrBitmap = NIL_RTR0MEMOBJ;
533 }
534
535 for (VMCPUID i = 0; i < pVM->cCpus; i++)
536 {
537 PVMCPU pVCpu = &pVM->aCpus[i];
538
539 /*
540 * Allocate one page for the host-context VM control block (VMCB). This is used for additional host-state (such as
541 * FS, GS, Kernel GS Base, etc.) apart from the host-state save area specified in MSR_K8_VM_HSAVE_PA.
542 */
543 rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjVmcbHost, 1 << PAGE_SHIFT, false /* fExecutable */);
544 if (RT_FAILURE(rc))
545 goto failure_cleanup;
546
547 pVCpu->hm.s.svm.pvVmcbHost = RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjVmcbHost);
548 pVCpu->hm.s.svm.HCPhysVmcbHost = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjVmcbHost, 0 /* iPage */);
549 Assert(pVCpu->hm.s.svm.HCPhysVmcbHost < _4G);
550 ASMMemZeroPage(pVCpu->hm.s.svm.pvVmcbHost);
551
552 /*
553 * Allocate one page for the guest-state VMCB.
554 */
555 rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjVmcb, 1 << PAGE_SHIFT, false /* fExecutable */);
556 if (RT_FAILURE(rc))
557 goto failure_cleanup;
558
559 pVCpu->hm.s.svm.pvVmcb = RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjVmcb);
560 pVCpu->hm.s.svm.HCPhysVmcb = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjVmcb, 0 /* iPage */);
561 Assert(pVCpu->hm.s.svm.HCPhysVmcb < _4G);
562 ASMMemZeroPage(pVCpu->hm.s.svm.pvVmcb);
563
564 /*
565 * Allocate two pages (8 KB) for the MSR permission bitmap. There doesn't seem to be a way to convince
566 * SVM to not require one.
567 */
568 rc = RTR0MemObjAllocCont(&pVCpu->hm.s.svm.hMemObjMsrBitmap, 2 << PAGE_SHIFT, false /* fExecutable */);
569 if (RT_FAILURE(rc))
570 goto failure_cleanup;
571
572 pVCpu->hm.s.svm.pvMsrBitmap = RTR0MemObjAddress(pVCpu->hm.s.svm.hMemObjMsrBitmap);
573 pVCpu->hm.s.svm.HCPhysMsrBitmap = RTR0MemObjGetPagePhysAddr(pVCpu->hm.s.svm.hMemObjMsrBitmap, 0 /* iPage */);
574 /* Set all bits to intercept all MSR accesses (changed later on). */
575 ASMMemFill32(pVCpu->hm.s.svm.pvMsrBitmap, 2 << PAGE_SHIFT, UINT32_C(0xffffffff));
576 }
577
578 return VINF_SUCCESS;
579
580failure_cleanup:
581 hmR0SvmFreeStructs(pVM);
582 return rc;
583}
584
585
586/**
587 * Does per-VM AMD-V termination.
588 *
589 * @returns VBox status code.
590 * @param pVM Pointer to the VM.
591 */
592VMMR0DECL(int) SVMR0TermVM(PVM pVM)
593{
594 hmR0SvmFreeStructs(pVM);
595 return VINF_SUCCESS;
596}
597
598
599/**
600 * Sets the permission bits for the specified MSR in the MSRPM.
601 *
602 * @param pVCpu Pointer to the VMCPU.
603 * @param uMsr The MSR for which the access permissions are being set.
604 * @param enmRead MSR read permissions.
605 * @param enmWrite MSR write permissions.
606 */
607static void hmR0SvmSetMsrPermission(PVMCPU pVCpu, unsigned uMsr, SVMMSREXITREAD enmRead, SVMMSREXITWRITE enmWrite)
608{
609 unsigned ulBit;
610 uint8_t *pbMsrBitmap = (uint8_t *)pVCpu->hm.s.svm.pvMsrBitmap;
611
612 /*
613 * Layout:
614 * Byte offset MSR range
615 * 0x000 - 0x7ff 0x00000000 - 0x00001fff
616 * 0x800 - 0xfff 0xc0000000 - 0xc0001fff
617 * 0x1000 - 0x17ff 0xc0010000 - 0xc0011fff
618 * 0x1800 - 0x1fff Reserved
619 */
620 if (uMsr <= 0x00001FFF)
621 {
622 /* Pentium-compatible MSRs. */
623 ulBit = uMsr * 2;
624 }
625 else if ( uMsr >= 0xC0000000
626 && uMsr <= 0xC0001FFF)
627 {
628 /* AMD Sixth Generation x86 Processor MSRs. */
629 ulBit = (uMsr - 0xC0000000) * 2;
630 pbMsrBitmap += 0x800;
631 }
632 else if ( uMsr >= 0xC0010000
633 && uMsr <= 0xC0011FFF)
634 {
635 /* AMD Seventh and Eighth Generation Processor MSRs. */
636 ulBit = (uMsr - 0xC0001000) * 2;
637 pbMsrBitmap += 0x1000;
638 }
639 else
640 {
641 AssertFailed();
642 return;
643 }
644
645 Assert(ulBit < 0x3fff /* 16 * 1024 - 1 */);
646 if (enmRead == SVMMSREXIT_INTERCEPT_READ)
647 ASMBitSet(pbMsrBitmap, ulBit);
648 else
649 ASMBitClear(pbMsrBitmap, ulBit);
650
651 if (enmWrite == SVMMSREXIT_INTERCEPT_WRITE)
652 ASMBitSet(pbMsrBitmap, ulBit + 1);
653 else
654 ASMBitClear(pbMsrBitmap, ulBit + 1);
655
656 PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
657 pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_IOPM_MSRPM;
658}
659
660
661/**
662 * Sets up AMD-V for the specified VM.
663 * This function is only called once per-VM during initalization.
664 *
665 * @returns VBox status code.
666 * @param pVM Pointer to the VM.
667 */
668VMMR0DECL(int) SVMR0SetupVM(PVM pVM)
669{
670 Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
671 AssertReturn(pVM, VERR_INVALID_PARAMETER);
672 Assert(pVM->hm.s.svm.fSupported);
673
674 for (VMCPUID i = 0; i < pVM->cCpus; i++)
675 {
676 PVMCPU pVCpu = &pVM->aCpus[i];
677 PSVMVMCB pVmcb = (PSVMVMCB)pVM->aCpus[i].hm.s.svm.pvVmcb;
678
679 AssertMsgReturn(pVmcb, ("Invalid pVmcb for vcpu[%u]\n", i), VERR_SVM_INVALID_PVMCB);
680
681 /* Initialize the #VMEXIT history array with end-of-array markers (UINT16_MAX). */
682 Assert(!pVCpu->hm.s.idxExitHistoryFree);
683 HMCPU_EXIT_HISTORY_RESET(pVCpu);
684
685 /* Trap exceptions unconditionally (debug purposes). */
686#ifdef HMSVM_ALWAYS_TRAP_PF
687 pVmcb->ctrl.u32InterceptException |= RT_BIT(X86_XCPT_PF);
688#endif
689#ifdef HMSVM_ALWAYS_TRAP_ALL_XCPTS
690 /* If you add any exceptions here, make sure to update hmR0SvmHandleExit(). */
691 pVmcb->ctrl.u32InterceptException |= 0
692 | RT_BIT(X86_XCPT_BP)
693 | RT_BIT(X86_XCPT_DB)
694 | RT_BIT(X86_XCPT_DE)
695 | RT_BIT(X86_XCPT_NM)
696 | RT_BIT(X86_XCPT_UD)
697 | RT_BIT(X86_XCPT_NP)
698 | RT_BIT(X86_XCPT_SS)
699 | RT_BIT(X86_XCPT_GP)
700 | RT_BIT(X86_XCPT_PF)
701 | RT_BIT(X86_XCPT_MF)
702 ;
703#endif
704
705 /* Set up unconditional intercepts and conditions. */
706 pVmcb->ctrl.u32InterceptCtrl1 = SVM_CTRL1_INTERCEPT_INTR /* External interrupt causes a #VMEXIT. */
707 | SVM_CTRL1_INTERCEPT_NMI /* Non-maskable interrupts causes a #VMEXIT. */
708 | SVM_CTRL1_INTERCEPT_INIT /* INIT signal causes a #VMEXIT. */
709 | SVM_CTRL1_INTERCEPT_RDPMC /* RDPMC causes a #VMEXIT. */
710 | SVM_CTRL1_INTERCEPT_CPUID /* CPUID causes a #VMEXIT. */
711 | SVM_CTRL1_INTERCEPT_RSM /* RSM causes a #VMEXIT. */
712 | SVM_CTRL1_INTERCEPT_HLT /* HLT causes a #VMEXIT. */
713 | SVM_CTRL1_INTERCEPT_INOUT_BITMAP /* Use the IOPM to cause IOIO #VMEXITs. */
714 | SVM_CTRL1_INTERCEPT_MSR_SHADOW /* MSR access not covered by MSRPM causes a #VMEXIT.*/
715 | SVM_CTRL1_INTERCEPT_INVLPGA /* INVLPGA causes a #VMEXIT. */
716 | SVM_CTRL1_INTERCEPT_SHUTDOWN /* Shutdown events causes a #VMEXIT. */
717 | SVM_CTRL1_INTERCEPT_FERR_FREEZE; /* Intercept "freezing" during legacy FPU handling. */
718
719 pVmcb->ctrl.u32InterceptCtrl2 = SVM_CTRL2_INTERCEPT_VMRUN /* VMRUN causes a #VMEXIT. */
720 | SVM_CTRL2_INTERCEPT_VMMCALL /* VMMCALL causes a #VMEXIT. */
721 | SVM_CTRL2_INTERCEPT_VMLOAD /* VMLOAD causes a #VMEXIT. */
722 | SVM_CTRL2_INTERCEPT_VMSAVE /* VMSAVE causes a #VMEXIT. */
723 | SVM_CTRL2_INTERCEPT_STGI /* STGI causes a #VMEXIT. */
724 | SVM_CTRL2_INTERCEPT_CLGI /* CLGI causes a #VMEXIT. */
725 | SVM_CTRL2_INTERCEPT_SKINIT /* SKINIT causes a #VMEXIT. */
726 | SVM_CTRL2_INTERCEPT_WBINVD /* WBINVD causes a #VMEXIT. */
727 | SVM_CTRL2_INTERCEPT_MONITOR /* MONITOR causes a #VMEXIT. */
728 | SVM_CTRL2_INTERCEPT_MWAIT /* MWAIT causes a #VMEXIT. */
729 | SVM_CTRL2_INTERCEPT_XSETBV; /* XSETBV causes a #VMEXIT. */
730
731 /* CR0, CR4 reads must be intercepted, our shadow values are not necessarily the same as the guest's. */
732 pVmcb->ctrl.u16InterceptRdCRx = RT_BIT(0) | RT_BIT(4);
733
734 /* CR0, CR4 writes must be intercepted for the same reasons as above. */
735 pVmcb->ctrl.u16InterceptWrCRx = RT_BIT(0) | RT_BIT(4);
736
737 /* Intercept all DRx reads and writes by default. Changed later on. */
738 pVmcb->ctrl.u16InterceptRdDRx = 0xffff;
739 pVmcb->ctrl.u16InterceptWrDRx = 0xffff;
740
741 /* Virtualize masking of INTR interrupts. (reads/writes from/to CR8 go to the V_TPR register) */
742 pVmcb->ctrl.IntCtrl.n.u1VIrqMasking = 1;
743
744 /* Ignore the priority in the TPR. This is necessary for delivering PIC style (ExtInt) interrupts and we currently
745 deliver both PIC and APIC interrupts alike. See hmR0SvmInjectPendingEvent() */
746 pVmcb->ctrl.IntCtrl.n.u1IgnoreTPR = 1;
747
748 /* Set IO and MSR bitmap permission bitmap physical addresses. */
749 pVmcb->ctrl.u64IOPMPhysAddr = g_HCPhysIOBitmap;
750 pVmcb->ctrl.u64MSRPMPhysAddr = pVCpu->hm.s.svm.HCPhysMsrBitmap;
751
752 /* No LBR virtualization. */
753 pVmcb->ctrl.u64LBRVirt = 0;
754
755 /* Initially set all VMCB clean bits to 0 indicating that everything should be loaded from the VMCB in memory. */
756 pVmcb->ctrl.u64VmcbCleanBits = 0;
757
758 /* The host ASID MBZ, for the guest start with 1. */
759 pVmcb->ctrl.TLBCtrl.n.u32ASID = 1;
760
761 /*
762 * Setup the PAT MSR (applicable for Nested Paging only).
763 * The default value should be 0x0007040600070406ULL, but we want to treat all guest memory as WB,
764 * so choose type 6 for all PAT slots.
765 */
766 pVmcb->guest.u64GPAT = UINT64_C(0x0006060606060606);
767
768 /* Setup Nested Paging. This doesn't change throughout the execution time of the VM. */
769 pVmcb->ctrl.NestedPaging.n.u1NestedPaging = pVM->hm.s.fNestedPaging;
770
771 /* Without Nested Paging, we need additionally intercepts. */
772 if (!pVM->hm.s.fNestedPaging)
773 {
774 /* CR3 reads/writes must be intercepted; our shadow values differ from the guest values. */
775 pVmcb->ctrl.u16InterceptRdCRx |= RT_BIT(3);
776 pVmcb->ctrl.u16InterceptWrCRx |= RT_BIT(3);
777
778 /* Intercept INVLPG and task switches (may change CR3, EFLAGS, LDT). */
779 pVmcb->ctrl.u32InterceptCtrl1 |= SVM_CTRL1_INTERCEPT_INVLPG
780 | SVM_CTRL1_INTERCEPT_TASK_SWITCH;
781
782 /* Page faults must be intercepted to implement shadow paging. */
783 pVmcb->ctrl.u32InterceptException |= RT_BIT(X86_XCPT_PF);
784 }
785
786#ifdef HMSVM_ALWAYS_TRAP_TASK_SWITCH
787 pVmcb->ctrl.u32InterceptCtrl1 |= SVM_CTRL1_INTERCEPT_TASK_SWITCH;
788#endif
789
790 /* Apply the exceptions intercepts needed by the GIM provider. */
791 if (pVCpu->hm.s.fGIMTrapXcptUD)
792 pVmcb->ctrl.u32InterceptException |= RT_BIT(X86_XCPT_UD);
793
794 /*
795 * The following MSRs are saved/restored automatically during the world-switch.
796 * Don't intercept guest read/write accesses to these MSRs.
797 */
798 hmR0SvmSetMsrPermission(pVCpu, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
799 hmR0SvmSetMsrPermission(pVCpu, MSR_K8_CSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
800 hmR0SvmSetMsrPermission(pVCpu, MSR_K6_STAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
801 hmR0SvmSetMsrPermission(pVCpu, MSR_K8_SF_MASK, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
802 hmR0SvmSetMsrPermission(pVCpu, MSR_K8_FS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
803 hmR0SvmSetMsrPermission(pVCpu, MSR_K8_GS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
804 hmR0SvmSetMsrPermission(pVCpu, MSR_K8_KERNEL_GS_BASE, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
805 hmR0SvmSetMsrPermission(pVCpu, MSR_IA32_SYSENTER_CS, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
806 hmR0SvmSetMsrPermission(pVCpu, MSR_IA32_SYSENTER_ESP, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
807 hmR0SvmSetMsrPermission(pVCpu, MSR_IA32_SYSENTER_EIP, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
808 }
809
810 return VINF_SUCCESS;
811}
812
813
814/**
815 * Invalidates a guest page by guest virtual address.
816 *
817 * @returns VBox status code.
818 * @param pVM Pointer to the VM.
819 * @param pVCpu Pointer to the VMCPU.
820 * @param GCVirt Guest virtual address of the page to invalidate.
821 */
822VMMR0DECL(int) SVMR0InvalidatePage(PVM pVM, PVMCPU pVCpu, RTGCPTR GCVirt)
823{
824 AssertReturn(pVM, VERR_INVALID_PARAMETER);
825 Assert(pVM->hm.s.svm.fSupported);
826
827 bool fFlushPending = pVM->hm.s.svm.fAlwaysFlushTLB || VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_TLB_FLUSH);
828
829 /* Skip it if a TLB flush is already pending. */
830 if (!fFlushPending)
831 {
832 Log4(("SVMR0InvalidatePage %RGv\n", GCVirt));
833
834 PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
835 AssertMsgReturn(pVmcb, ("Invalid pVmcb!\n"), VERR_SVM_INVALID_PVMCB);
836
837#if HC_ARCH_BITS == 32
838 /* If we get a flush in 64-bit guest mode, then force a full TLB flush. INVLPGA takes only 32-bit addresses. */
839 if (CPUMIsGuestInLongMode(pVCpu))
840 VMCPU_FF_SET(pVCpu, VMCPU_FF_TLB_FLUSH);
841 else
842#endif
843 {
844 SVMR0InvlpgA(GCVirt, pVmcb->ctrl.TLBCtrl.n.u32ASID);
845 STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlbInvlpgVirt);
846 }
847 }
848 return VINF_SUCCESS;
849}
850
851
852/**
853 * Flushes the appropriate tagged-TLB entries.
854 *
855 * @param pVM Pointer to the VM.
856 * @param pVCpu Pointer to the VMCPU.
857 */
858static void hmR0SvmFlushTaggedTlb(PVMCPU pVCpu)
859{
860 PVM pVM = pVCpu->CTX_SUFF(pVM);
861 PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
862 PHMGLOBALCPUINFO pCpu = HMR0GetCurrentCpu();
863
864 /*
865 * Force a TLB flush for the first world switch if the current CPU differs from the one we ran on last.
866 * This can happen both for start & resume due to long jumps back to ring-3.
867 * If the TLB flush count changed, another VM (VCPU rather) has hit the ASID limit while flushing the TLB,
868 * so we cannot reuse the ASIDs without flushing.
869 */
870 bool fNewAsid = false;
871 Assert(pCpu->idCpu != NIL_RTCPUID);
872 if ( pVCpu->hm.s.idLastCpu != pCpu->idCpu
873 || pVCpu->hm.s.cTlbFlushes != pCpu->cTlbFlushes)
874 {
875 STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlbWorldSwitch);
876 pVCpu->hm.s.fForceTLBFlush = true;
877 fNewAsid = true;
878 }
879
880 /* Set TLB flush state as checked until we return from the world switch. */
881 ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, true);
882
883 /* Check for explicit TLB shootdowns. */
884 if (VMCPU_FF_TEST_AND_CLEAR(pVCpu, VMCPU_FF_TLB_FLUSH))
885 {
886 pVCpu->hm.s.fForceTLBFlush = true;
887 STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushTlb);
888 }
889
890 pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_NOTHING;
891
892 if (pVM->hm.s.svm.fAlwaysFlushTLB)
893 {
894 /*
895 * This is the AMD erratum 170. We need to flush the entire TLB for each world switch. Sad.
896 */
897 pCpu->uCurrentAsid = 1;
898 pVCpu->hm.s.uCurrentAsid = 1;
899 pVCpu->hm.s.cTlbFlushes = pCpu->cTlbFlushes;
900 pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
901
902 /* Clear the VMCB Clean Bit for NP while flushing the TLB. See @bugref{7152}. */
903 pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
904 }
905 else if (pVCpu->hm.s.fForceTLBFlush)
906 {
907 /* Clear the VMCB Clean Bit for NP while flushing the TLB. See @bugref{7152}. */
908 pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
909
910 if (fNewAsid)
911 {
912 ++pCpu->uCurrentAsid;
913 bool fHitASIDLimit = false;
914 if (pCpu->uCurrentAsid >= pVM->hm.s.uMaxAsid)
915 {
916 pCpu->uCurrentAsid = 1; /* Wraparound at 1; host uses 0 */
917 pCpu->cTlbFlushes++; /* All VCPUs that run on this host CPU must use a new VPID. */
918 fHitASIDLimit = true;
919
920 if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_FLUSH_BY_ASID)
921 {
922 pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_SINGLE_CONTEXT;
923 pCpu->fFlushAsidBeforeUse = true;
924 }
925 else
926 {
927 pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
928 pCpu->fFlushAsidBeforeUse = false;
929 }
930 }
931
932 if ( !fHitASIDLimit
933 && pCpu->fFlushAsidBeforeUse)
934 {
935 if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_FLUSH_BY_ASID)
936 pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_SINGLE_CONTEXT;
937 else
938 {
939 pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
940 pCpu->fFlushAsidBeforeUse = false;
941 }
942 }
943
944 pVCpu->hm.s.uCurrentAsid = pCpu->uCurrentAsid;
945 pVCpu->hm.s.idLastCpu = pCpu->idCpu;
946 pVCpu->hm.s.cTlbFlushes = pCpu->cTlbFlushes;
947 }
948 else
949 {
950 if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_FLUSH_BY_ASID)
951 pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_SINGLE_CONTEXT;
952 else
953 pVmcb->ctrl.TLBCtrl.n.u8TLBFlush = SVM_TLB_FLUSH_ENTIRE;
954 }
955
956 pVCpu->hm.s.fForceTLBFlush = false;
957 }
958 /** @todo We never set VMCPU_FF_TLB_SHOOTDOWN anywhere so this path should
959 * not be executed. See hmQueueInvlPage() where it is commented
960 * out. Support individual entry flushing someday. */
961#if 0
962 else
963 {
964 if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_TLB_SHOOTDOWN))
965 {
966 /* Deal with pending TLB shootdown actions which were queued when we were not executing code. */
967 STAM_COUNTER_INC(&pVCpu->hm.s.StatTlbShootdown);
968 for (uint32_t i = 0; i < pVCpu->hm.s.TlbShootdown.cPages; i++)
969 SVMR0InvlpgA(pVCpu->hm.s.TlbShootdown.aPages[i], pVmcb->ctrl.TLBCtrl.n.u32ASID);
970
971 pVCpu->hm.s.TlbShootdown.cPages = 0;
972 VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_TLB_SHOOTDOWN);
973 }
974 }
975#endif
976
977
978 /* Update VMCB with the ASID. */
979 if (pVmcb->ctrl.TLBCtrl.n.u32ASID != pVCpu->hm.s.uCurrentAsid)
980 {
981 pVmcb->ctrl.TLBCtrl.n.u32ASID = pVCpu->hm.s.uCurrentAsid;
982 pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_ASID;
983 }
984
985 AssertMsg(pVCpu->hm.s.idLastCpu == pCpu->idCpu,
986 ("vcpu idLastCpu=%x pcpu idCpu=%x\n", pVCpu->hm.s.idLastCpu, pCpu->idCpu));
987 AssertMsg(pVCpu->hm.s.cTlbFlushes == pCpu->cTlbFlushes,
988 ("Flush count mismatch for cpu %d (%x vs %x)\n", pCpu->idCpu, pVCpu->hm.s.cTlbFlushes, pCpu->cTlbFlushes));
989 AssertMsg(pCpu->uCurrentAsid >= 1 && pCpu->uCurrentAsid < pVM->hm.s.uMaxAsid,
990 ("cpu%d uCurrentAsid = %x\n", pCpu->idCpu, pCpu->uCurrentAsid));
991 AssertMsg(pVCpu->hm.s.uCurrentAsid >= 1 && pVCpu->hm.s.uCurrentAsid < pVM->hm.s.uMaxAsid,
992 ("cpu%d VM uCurrentAsid = %x\n", pCpu->idCpu, pVCpu->hm.s.uCurrentAsid));
993
994#ifdef VBOX_WITH_STATISTICS
995 if (pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_NOTHING)
996 STAM_COUNTER_INC(&pVCpu->hm.s.StatNoFlushTlbWorldSwitch);
997 else if ( pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_SINGLE_CONTEXT
998 || pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_SINGLE_CONTEXT_RETAIN_GLOBALS)
999 {
1000 STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushAsid);
1001 }
1002 else
1003 {
1004 Assert(pVmcb->ctrl.TLBCtrl.n.u8TLBFlush == SVM_TLB_FLUSH_ENTIRE);
1005 STAM_COUNTER_INC(&pVCpu->hm.s.StatFlushEntire);
1006 }
1007#endif
1008}
1009
1010
1011/** @name 64-bit guest on 32-bit host OS helper functions.
1012 *
1013 * The host CPU is still 64-bit capable but the host OS is running in 32-bit
1014 * mode (code segment, paging). These wrappers/helpers perform the necessary
1015 * bits for the 32->64 switcher.
1016 *
1017 * @{ */
1018#if HC_ARCH_BITS == 32 && defined(VBOX_ENABLE_64_BITS_GUESTS) && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
1019/**
1020 * Prepares for and executes VMRUN (64-bit guests on a 32-bit host).
1021 *
1022 * @returns VBox status code.
1023 * @param HCPhysVmcbHost Physical address of host VMCB.
1024 * @param HCPhysVmcb Physical address of the VMCB.
1025 * @param pCtx Pointer to the guest-CPU context.
1026 * @param pVM Pointer to the VM.
1027 * @param pVCpu Pointer to the VMCPU.
1028 */
1029DECLASM(int) SVMR0VMSwitcherRun64(RTHCPHYS HCPhysVmcbHost, RTHCPHYS HCPhysVmcb, PCPUMCTX pCtx, PVM pVM, PVMCPU pVCpu)
1030{
1031 uint32_t aParam[8];
1032 aParam[0] = (uint32_t)(HCPhysVmcbHost); /* Param 1: HCPhysVmcbHost - Lo. */
1033 aParam[1] = (uint32_t)(HCPhysVmcbHost >> 32); /* Param 1: HCPhysVmcbHost - Hi. */
1034 aParam[2] = (uint32_t)(HCPhysVmcb); /* Param 2: HCPhysVmcb - Lo. */
1035 aParam[3] = (uint32_t)(HCPhysVmcb >> 32); /* Param 2: HCPhysVmcb - Hi. */
1036 aParam[4] = VM_RC_ADDR(pVM, pVM);
1037 aParam[5] = 0;
1038 aParam[6] = VM_RC_ADDR(pVM, pVCpu);
1039 aParam[7] = 0;
1040
1041 return SVMR0Execute64BitsHandler(pVM, pVCpu, pCtx, HM64ON32OP_SVMRCVMRun64, RT_ELEMENTS(aParam), &aParam[0]);
1042}
1043
1044
1045/**
1046 * Executes the specified VMRUN handler in 64-bit mode.
1047 *
1048 * @returns VBox status code.
1049 * @param pVM Pointer to the VM.
1050 * @param pVCpu Pointer to the VMCPU.
1051 * @param pCtx Pointer to the guest-CPU context.
1052 * @param enmOp The operation to perform.
1053 * @param cParams Number of parameters.
1054 * @param paParam Array of 32-bit parameters.
1055 */
1056VMMR0DECL(int) SVMR0Execute64BitsHandler(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, HM64ON32OP enmOp,
1057 uint32_t cParams, uint32_t *paParam)
1058{
1059 AssertReturn(pVM->hm.s.pfnHost32ToGuest64R0, VERR_HM_NO_32_TO_64_SWITCHER);
1060 Assert(enmOp > HM64ON32OP_INVALID && enmOp < HM64ON32OP_END);
1061
1062 /* Disable interrupts. */
1063 RTHCUINTREG uOldEFlags = ASMIntDisableFlags();
1064
1065#ifdef VBOX_WITH_VMMR0_DISABLE_LAPIC_NMI
1066 RTCPUID idHostCpu = RTMpCpuId();
1067 CPUMR0SetLApic(pVCpu, idHostCpu);
1068#endif
1069
1070 CPUMSetHyperESP(pVCpu, VMMGetStackRC(pVCpu));
1071 CPUMSetHyperEIP(pVCpu, enmOp);
1072 for (int i = (int)cParams - 1; i >= 0; i--)
1073 CPUMPushHyper(pVCpu, paParam[i]);
1074
1075 STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatWorldSwitch3264, z);
1076 /* Call the switcher. */
1077 int rc = pVM->hm.s.pfnHost32ToGuest64R0(pVM, RT_OFFSETOF(VM, aCpus[pVCpu->idCpu].cpum) - RT_OFFSETOF(VM, cpum));
1078 STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatWorldSwitch3264, z);
1079
1080 /* Restore interrupts. */
1081 ASMSetFlags(uOldEFlags);
1082 return rc;
1083}
1084
1085#endif /* HC_ARCH_BITS == 32 && defined(VBOX_ENABLE_64_BITS_GUESTS) */
1086/** @} */
1087
1088
1089/**
1090 * Adds an exception to the intercept exception bitmap in the VMCB and updates
1091 * the corresponding VMCB Clean bit.
1092 *
1093 * @param pVmcb Pointer to the VM control block.
1094 * @param u32Xcpt The value of the exception (X86_XCPT_*).
1095 */
1096DECLINLINE(void) hmR0SvmAddXcptIntercept(PSVMVMCB pVmcb, uint32_t u32Xcpt)
1097{
1098 if (!(pVmcb->ctrl.u32InterceptException & RT_BIT(u32Xcpt)))
1099 {
1100 pVmcb->ctrl.u32InterceptException |= RT_BIT(u32Xcpt);
1101 pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1102 }
1103}
1104
1105
1106/**
1107 * Removes an exception from the intercept-exception bitmap in the VMCB and
1108 * updates the corresponding VMCB Clean bit.
1109 *
1110 * @param pVmcb Pointer to the VM control block.
1111 * @param u32Xcpt The value of the exception (X86_XCPT_*).
1112 */
1113DECLINLINE(void) hmR0SvmRemoveXcptIntercept(PSVMVMCB pVmcb, uint32_t u32Xcpt)
1114{
1115#ifndef HMSVM_ALWAYS_TRAP_ALL_XCPTS
1116 if (pVmcb->ctrl.u32InterceptException & RT_BIT(u32Xcpt))
1117 {
1118 pVmcb->ctrl.u32InterceptException &= ~RT_BIT(u32Xcpt);
1119 pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1120 }
1121#endif
1122}
1123
1124
1125/**
1126 * Loads the guest CR0 control register into the guest-state area in the VMCB.
1127 * Although the guest CR0 is a separate field in the VMCB we have to consider
1128 * the FPU state itself which is shared between the host and the guest.
1129 *
1130 * @returns VBox status code.
1131 * @param pVM Pointer to the VMCPU.
1132 * @param pVmcb Pointer to the VM control block.
1133 * @param pCtx Pointer to the guest-CPU context.
1134 *
1135 * @remarks No-long-jump zone!!!
1136 */
1137static void hmR0SvmLoadSharedCR0(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1138{
1139 /*
1140 * Guest CR0.
1141 */
1142 PVM pVM = pVCpu->CTX_SUFF(pVM);
1143 if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR0))
1144 {
1145 uint64_t u64GuestCR0 = pCtx->cr0;
1146
1147 /* Always enable caching. */
1148 u64GuestCR0 &= ~(X86_CR0_CD | X86_CR0_NW);
1149
1150 /*
1151 * When Nested Paging is not available use shadow page tables and intercept #PFs (the latter done in SVMR0SetupVM()).
1152 */
1153 if (!pVM->hm.s.fNestedPaging)
1154 {
1155 u64GuestCR0 |= X86_CR0_PG; /* When Nested Paging is not available, use shadow page tables. */
1156 u64GuestCR0 |= X86_CR0_WP; /* Guest CPL 0 writes to its read-only pages should cause a #PF #VMEXIT. */
1157 }
1158
1159 /*
1160 * Guest FPU bits.
1161 */
1162 bool fInterceptNM = false;
1163 bool fInterceptMF = false;
1164 u64GuestCR0 |= X86_CR0_NE; /* Use internal x87 FPU exceptions handling rather than external interrupts. */
1165 if (CPUMIsGuestFPUStateActive(pVCpu))
1166 {
1167 /* Catch floating point exceptions if we need to report them to the guest in a different way. */
1168 if (!(pCtx->cr0 & X86_CR0_NE))
1169 {
1170 Log4(("hmR0SvmLoadGuestControlRegs: Intercepting Guest CR0.MP Old-style FPU handling!!!\n"));
1171 fInterceptMF = true;
1172 }
1173 }
1174 else
1175 {
1176 fInterceptNM = true; /* Guest FPU inactive, #VMEXIT on #NM for lazy FPU loading. */
1177 u64GuestCR0 |= X86_CR0_TS /* Guest can task switch quickly and do lazy FPU syncing. */
1178 | X86_CR0_MP; /* FWAIT/WAIT should not ignore CR0.TS and should generate #NM. */
1179 }
1180
1181 /*
1182 * Update the exception intercept bitmap.
1183 */
1184 if (fInterceptNM)
1185 hmR0SvmAddXcptIntercept(pVmcb, X86_XCPT_NM);
1186 else
1187 hmR0SvmRemoveXcptIntercept(pVmcb, X86_XCPT_NM);
1188
1189 if (fInterceptMF)
1190 hmR0SvmAddXcptIntercept(pVmcb, X86_XCPT_MF);
1191 else
1192 hmR0SvmRemoveXcptIntercept(pVmcb, X86_XCPT_MF);
1193
1194 pVmcb->guest.u64CR0 = u64GuestCR0;
1195 pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1196 HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_CR0);
1197 }
1198}
1199
1200
1201/**
1202 * Loads the guest control registers (CR2, CR3, CR4) into the VMCB.
1203 *
1204 * @returns VBox status code.
1205 * @param pVCpu Pointer to the VMCPU.
1206 * @param pVmcb Pointer to the VM control block.
1207 * @param pCtx Pointer to the guest-CPU context.
1208 *
1209 * @remarks No-long-jump zone!!!
1210 */
1211static int hmR0SvmLoadGuestControlRegs(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1212{
1213 PVM pVM = pVCpu->CTX_SUFF(pVM);
1214
1215 /*
1216 * Guest CR2.
1217 */
1218 if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR2))
1219 {
1220 pVmcb->guest.u64CR2 = pCtx->cr2;
1221 pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CR2;
1222 HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_CR2);
1223 }
1224
1225 /*
1226 * Guest CR3.
1227 */
1228 if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR3))
1229 {
1230 if (pVM->hm.s.fNestedPaging)
1231 {
1232 PGMMODE enmShwPagingMode;
1233#if HC_ARCH_BITS == 32
1234 if (CPUMIsGuestInLongModeEx(pCtx))
1235 enmShwPagingMode = PGMMODE_AMD64_NX;
1236 else
1237#endif
1238 enmShwPagingMode = PGMGetHostMode(pVM);
1239
1240 pVmcb->ctrl.u64NestedPagingCR3 = PGMGetNestedCR3(pVCpu, enmShwPagingMode);
1241 pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_NP;
1242 Assert(pVmcb->ctrl.u64NestedPagingCR3);
1243 pVmcb->guest.u64CR3 = pCtx->cr3;
1244 }
1245 else
1246 pVmcb->guest.u64CR3 = PGMGetHyperCR3(pVCpu);
1247
1248 pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1249 HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_CR3);
1250 }
1251
1252 /*
1253 * Guest CR4.
1254 * ASSUMES this is done everytime we get in from ring-3! (XCR0)
1255 */
1256 if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR4))
1257 {
1258 uint64_t u64GuestCR4 = pCtx->cr4;
1259 if (!pVM->hm.s.fNestedPaging)
1260 {
1261 switch (pVCpu->hm.s.enmShadowMode)
1262 {
1263 case PGMMODE_REAL:
1264 case PGMMODE_PROTECTED: /* Protected mode, no paging. */
1265 AssertFailed();
1266 return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1267
1268 case PGMMODE_32_BIT: /* 32-bit paging. */
1269 u64GuestCR4 &= ~X86_CR4_PAE;
1270 break;
1271
1272 case PGMMODE_PAE: /* PAE paging. */
1273 case PGMMODE_PAE_NX: /* PAE paging with NX enabled. */
1274 /** Must use PAE paging as we could use physical memory > 4 GB */
1275 u64GuestCR4 |= X86_CR4_PAE;
1276 break;
1277
1278 case PGMMODE_AMD64: /* 64-bit AMD paging (long mode). */
1279 case PGMMODE_AMD64_NX: /* 64-bit AMD paging (long mode) with NX enabled. */
1280#ifdef VBOX_ENABLE_64_BITS_GUESTS
1281 break;
1282#else
1283 AssertFailed();
1284 return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1285#endif
1286
1287 default: /* shut up gcc */
1288 AssertFailed();
1289 return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1290 }
1291 }
1292
1293 pVmcb->guest.u64CR4 = u64GuestCR4;
1294 pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1295
1296 /* Whether to save/load/restore XCR0 during world switch depends on CR4.OSXSAVE and host+guest XCR0. */
1297 pVCpu->hm.s.fLoadSaveGuestXcr0 = (u64GuestCR4 & X86_CR4_OSXSAVE) && pCtx->aXcr[0] != ASMGetXcr0();
1298
1299 HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_CR4);
1300 }
1301
1302 return VINF_SUCCESS;
1303}
1304
1305
1306/**
1307 * Loads the guest segment registers into the VMCB.
1308 *
1309 * @returns VBox status code.
1310 * @param pVCpu Pointer to the VMCPU.
1311 * @param pVmcb Pointer to the VM control block.
1312 * @param pCtx Pointer to the guest-CPU context.
1313 *
1314 * @remarks No-long-jump zone!!!
1315 */
1316static void hmR0SvmLoadGuestSegmentRegs(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1317{
1318 /* Guest Segment registers: CS, SS, DS, ES, FS, GS. */
1319 if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_SEGMENT_REGS))
1320 {
1321 HMSVM_LOAD_SEG_REG(CS, cs);
1322 HMSVM_LOAD_SEG_REG(SS, ss);
1323 HMSVM_LOAD_SEG_REG(DS, ds);
1324 HMSVM_LOAD_SEG_REG(ES, es);
1325 HMSVM_LOAD_SEG_REG(FS, fs);
1326 HMSVM_LOAD_SEG_REG(GS, gs);
1327
1328 pVmcb->guest.u8CPL = pCtx->ss.Attr.n.u2Dpl;
1329 pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_SEG;
1330 HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_SEGMENT_REGS);
1331 }
1332
1333 /* Guest TR. */
1334 if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_TR))
1335 {
1336 HMSVM_LOAD_SEG_REG(TR, tr);
1337 HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_TR);
1338 }
1339
1340 /* Guest LDTR. */
1341 if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_LDTR))
1342 {
1343 HMSVM_LOAD_SEG_REG(LDTR, ldtr);
1344 HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_LDTR);
1345 }
1346
1347 /* Guest GDTR. */
1348 if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_GDTR))
1349 {
1350 pVmcb->guest.GDTR.u32Limit = pCtx->gdtr.cbGdt;
1351 pVmcb->guest.GDTR.u64Base = pCtx->gdtr.pGdt;
1352 pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DT;
1353 HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_GDTR);
1354 }
1355
1356 /* Guest IDTR. */
1357 if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_IDTR))
1358 {
1359 pVmcb->guest.IDTR.u32Limit = pCtx->idtr.cbIdt;
1360 pVmcb->guest.IDTR.u64Base = pCtx->idtr.pIdt;
1361 pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DT;
1362 HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_IDTR);
1363 }
1364}
1365
1366
1367/**
1368 * Loads the guest MSRs into the VMCB.
1369 *
1370 * @param pVCpu Pointer to the VMCPU.
1371 * @param pVmcb Pointer to the VM control block.
1372 * @param pCtx Pointer to the guest-CPU context.
1373 *
1374 * @remarks No-long-jump zone!!!
1375 */
1376static void hmR0SvmLoadGuestMsrs(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1377{
1378 /* Guest Sysenter MSRs. */
1379 pVmcb->guest.u64SysEnterCS = pCtx->SysEnter.cs;
1380 pVmcb->guest.u64SysEnterEIP = pCtx->SysEnter.eip;
1381 pVmcb->guest.u64SysEnterESP = pCtx->SysEnter.esp;
1382
1383 /*
1384 * Guest EFER MSR.
1385 * AMD-V requires guest EFER.SVME to be set. Weird.
1386 * See AMD spec. 15.5.1 "Basic Operation" | "Canonicalization and Consistency Checks".
1387 */
1388 if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_EFER_MSR))
1389 {
1390 pVmcb->guest.u64EFER = pCtx->msrEFER | MSR_K6_EFER_SVME;
1391 pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1392 HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_EFER_MSR);
1393 }
1394
1395 /* 64-bit MSRs. */
1396 if (CPUMIsGuestInLongModeEx(pCtx))
1397 {
1398 pVmcb->guest.FS.u64Base = pCtx->fs.u64Base;
1399 pVmcb->guest.GS.u64Base = pCtx->gs.u64Base;
1400 }
1401 else
1402 {
1403 /* If the guest isn't in 64-bit mode, clear MSR_K6_LME bit from guest EFER otherwise AMD-V expects amd64 shadow paging. */
1404 if (pCtx->msrEFER & MSR_K6_EFER_LME)
1405 {
1406 pVmcb->guest.u64EFER &= ~MSR_K6_EFER_LME;
1407 pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_CRX_EFER;
1408 }
1409 }
1410
1411
1412 /** @todo The following are used in 64-bit only (SYSCALL/SYSRET) but they might
1413 * be writable in 32-bit mode. Clarify with AMD spec. */
1414 pVmcb->guest.u64STAR = pCtx->msrSTAR;
1415 pVmcb->guest.u64LSTAR = pCtx->msrLSTAR;
1416 pVmcb->guest.u64CSTAR = pCtx->msrCSTAR;
1417 pVmcb->guest.u64SFMASK = pCtx->msrSFMASK;
1418 pVmcb->guest.u64KernelGSBase = pCtx->msrKERNELGSBASE;
1419}
1420
1421
1422/**
1423 * Loads the guest state into the VMCB and programs the necessary intercepts
1424 * accordingly.
1425 *
1426 * @param pVCpu Pointer to the VMCPU.
1427 * @param pVmcb Pointer to the VM control block.
1428 * @param pCtx Pointer to the guest-CPU context.
1429 *
1430 * @remarks No-long-jump zone!!!
1431 * @remarks Requires EFLAGS to be up-to-date in the VMCB!
1432 */
1433static void hmR0SvmLoadSharedDebugState(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1434{
1435 if (!HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_DEBUG))
1436 return;
1437 Assert((pCtx->dr[6] & X86_DR6_RA1_MASK) == X86_DR6_RA1_MASK); Assert((pCtx->dr[6] & X86_DR6_RAZ_MASK) == 0);
1438 Assert((pCtx->dr[7] & X86_DR7_RA1_MASK) == X86_DR7_RA1_MASK); Assert((pCtx->dr[7] & X86_DR7_RAZ_MASK) == 0);
1439
1440 bool fInterceptDB = false;
1441 bool fInterceptMovDRx = false;
1442
1443 /*
1444 * Anyone single stepping on the host side? If so, we'll have to use the
1445 * trap flag in the guest EFLAGS since AMD-V doesn't have a trap flag on
1446 * the VMM level like the VT-x implementations does.
1447 */
1448 bool const fStepping = pVCpu->hm.s.fSingleInstruction || DBGFIsStepping(pVCpu);
1449 if (fStepping)
1450 {
1451 pVCpu->hm.s.fClearTrapFlag = true;
1452 pVmcb->guest.u64RFlags |= X86_EFL_TF;
1453 fInterceptDB = true;
1454 fInterceptMovDRx = true; /* Need clean DR6, no guest mess. */
1455 }
1456
1457 if ( fStepping
1458 || (CPUMGetHyperDR7(pVCpu) & X86_DR7_ENABLED_MASK))
1459 {
1460 /*
1461 * Use the combined guest and host DRx values found in the hypervisor
1462 * register set because the debugger has breakpoints active or someone
1463 * is single stepping on the host side.
1464 *
1465 * Note! DBGF expects a clean DR6 state before executing guest code.
1466 */
1467#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS) && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
1468 if ( CPUMIsGuestInLongModeEx(pCtx)
1469 && !CPUMIsHyperDebugStateActivePending(pVCpu))
1470 {
1471 CPUMR0LoadHyperDebugState(pVCpu, false /* include DR6 */);
1472 Assert(!CPUMIsGuestDebugStateActivePending(pVCpu));
1473 Assert(CPUMIsHyperDebugStateActivePending(pVCpu));
1474 }
1475 else
1476#endif
1477 if (!CPUMIsHyperDebugStateActive(pVCpu))
1478 {
1479 CPUMR0LoadHyperDebugState(pVCpu, false /* include DR6 */);
1480 Assert(!CPUMIsGuestDebugStateActive(pVCpu));
1481 Assert(CPUMIsHyperDebugStateActive(pVCpu));
1482 }
1483
1484 /* Update DR6 & DR7. (The other DRx values are handled by CPUM one way or the other.) */
1485 if ( pVmcb->guest.u64DR6 != X86_DR6_INIT_VAL
1486 || pVmcb->guest.u64DR7 != CPUMGetHyperDR7(pVCpu))
1487 {
1488 pVmcb->guest.u64DR7 = CPUMGetHyperDR7(pVCpu);
1489 pVmcb->guest.u64DR6 = X86_DR6_INIT_VAL;
1490 pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
1491 pVCpu->hm.s.fUsingHyperDR7 = true;
1492 }
1493
1494 /** @todo If we cared, we could optimize to allow the guest to read registers
1495 * with the same values. */
1496 fInterceptDB = true;
1497 fInterceptMovDRx = true;
1498 Log5(("hmR0SvmLoadSharedDebugState: Loaded hyper DRx\n"));
1499 }
1500 else
1501 {
1502 /*
1503 * Update DR6, DR7 with the guest values if necessary.
1504 */
1505 if ( pVmcb->guest.u64DR7 != pCtx->dr[7]
1506 || pVmcb->guest.u64DR6 != pCtx->dr[6])
1507 {
1508 pVmcb->guest.u64DR7 = pCtx->dr[7];
1509 pVmcb->guest.u64DR6 = pCtx->dr[6];
1510 pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
1511 pVCpu->hm.s.fUsingHyperDR7 = false;
1512 }
1513
1514 /*
1515 * If the guest has enabled debug registers, we need to load them prior to
1516 * executing guest code so they'll trigger at the right time.
1517 */
1518 if (pCtx->dr[7] & (X86_DR7_ENABLED_MASK | X86_DR7_GD)) /** @todo Why GD? */
1519 {
1520#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS) && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
1521 if ( CPUMIsGuestInLongModeEx(pCtx)
1522 && !CPUMIsGuestDebugStateActivePending(pVCpu))
1523 {
1524 CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
1525 STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxArmed);
1526 Assert(!CPUMIsHyperDebugStateActivePending(pVCpu));
1527 Assert(CPUMIsGuestDebugStateActivePending(pVCpu));
1528 }
1529 else
1530#endif
1531 if (!CPUMIsGuestDebugStateActive(pVCpu))
1532 {
1533 CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
1534 STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxArmed);
1535 Assert(!CPUMIsHyperDebugStateActive(pVCpu));
1536 Assert(CPUMIsGuestDebugStateActive(pVCpu));
1537 }
1538 Log5(("hmR0SvmLoadSharedDebugState: Loaded guest DRx\n"));
1539 }
1540 /*
1541 * If no debugging enabled, we'll lazy load DR0-3. We don't need to
1542 * intercept #DB as DR6 is updated in the VMCB.
1543 */
1544#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS) && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
1545 else if ( !CPUMIsGuestDebugStateActivePending(pVCpu)
1546 && !CPUMIsGuestDebugStateActive(pVCpu))
1547#else
1548 else if (!CPUMIsGuestDebugStateActive(pVCpu))
1549#endif
1550 {
1551 fInterceptMovDRx = true;
1552 }
1553 }
1554
1555 /*
1556 * Set up the intercepts.
1557 */
1558 if (fInterceptDB)
1559 hmR0SvmAddXcptIntercept(pVmcb, X86_XCPT_DB);
1560 else
1561 hmR0SvmRemoveXcptIntercept(pVmcb, X86_XCPT_DB);
1562
1563 if (fInterceptMovDRx)
1564 {
1565 if ( pVmcb->ctrl.u16InterceptRdDRx != 0xffff
1566 || pVmcb->ctrl.u16InterceptWrDRx != 0xffff)
1567 {
1568 pVmcb->ctrl.u16InterceptRdDRx = 0xffff;
1569 pVmcb->ctrl.u16InterceptWrDRx = 0xffff;
1570 pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1571 }
1572 }
1573 else
1574 {
1575 if ( pVmcb->ctrl.u16InterceptRdDRx
1576 || pVmcb->ctrl.u16InterceptWrDRx)
1577 {
1578 pVmcb->ctrl.u16InterceptRdDRx = 0;
1579 pVmcb->ctrl.u16InterceptWrDRx = 0;
1580 pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
1581 }
1582 }
1583
1584 HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_DEBUG);
1585}
1586
1587
1588/**
1589 * Loads the guest APIC state (currently just the TPR).
1590 *
1591 * @returns VBox status code.
1592 * @param pVCpu Pointer to the VMCPU.
1593 * @param pVmcb Pointer to the VM control block.
1594 * @param pCtx Pointer to the guest-CPU context.
1595 */
1596static int hmR0SvmLoadGuestApicState(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1597{
1598 if (!HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE))
1599 return VINF_SUCCESS;
1600
1601 bool fPendingIntr;
1602 uint8_t u8Tpr;
1603 int rc = PDMApicGetTPR(pVCpu, &u8Tpr, &fPendingIntr, NULL /* pu8PendingIrq */);
1604 AssertRCReturn(rc, rc);
1605
1606 /* Assume that we need to trap all TPR accesses and thus need not check on
1607 every #VMEXIT if we should update the TPR. */
1608 Assert(pVmcb->ctrl.IntCtrl.n.u1VIrqMasking);
1609 pVCpu->hm.s.svm.fSyncVTpr = false;
1610
1611 /* 32-bit guests uses LSTAR MSR for patching guest code which touches the TPR. */
1612 if (pVCpu->CTX_SUFF(pVM)->hm.s.fTPRPatchingActive)
1613 {
1614 pCtx->msrLSTAR = u8Tpr;
1615
1616 /* If there are interrupts pending, intercept LSTAR writes, otherwise don't intercept reads or writes. */
1617 if (fPendingIntr)
1618 hmR0SvmSetMsrPermission(pVCpu, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_INTERCEPT_WRITE);
1619 else
1620 {
1621 hmR0SvmSetMsrPermission(pVCpu, MSR_K8_LSTAR, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
1622 pVCpu->hm.s.svm.fSyncVTpr = true;
1623 }
1624 }
1625 else
1626 {
1627 /* Bits 3-0 of the VTPR field correspond to bits 7-4 of the TPR (which is the Task-Priority Class). */
1628 pVmcb->ctrl.IntCtrl.n.u8VTPR = (u8Tpr >> 4);
1629
1630 /* If there are interrupts pending, intercept CR8 writes to evaluate ASAP if we can deliver the interrupt to the guest. */
1631 if (fPendingIntr)
1632 pVmcb->ctrl.u16InterceptWrCRx |= RT_BIT(8);
1633 else
1634 {
1635 pVmcb->ctrl.u16InterceptWrCRx &= ~RT_BIT(8);
1636 pVCpu->hm.s.svm.fSyncVTpr = true;
1637 }
1638
1639 pVmcb->ctrl.u64VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS | HMSVM_VMCB_CLEAN_TPR);
1640 }
1641
1642 HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
1643 return rc;
1644}
1645
1646
1647/**
1648 * Loads the exception interrupts required for guest execution in the VMCB.
1649 *
1650 * @returns VBox status code.
1651 * @param pVCpu Pointer to the VMCPU.
1652 * @param pVmcb Pointer to the VM control block.
1653 * @param pCtx Pointer to the guest-CPU context.
1654 */
1655static int hmR0SvmLoadGuestXcptIntercepts(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1656{
1657 int rc = VINF_SUCCESS;
1658 if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_XCPT_INTERCEPTS))
1659 {
1660 /* The remaining intercepts are handled elsewhere, e.g. in hmR0SvmLoadSharedCR0(). */
1661 if (pVCpu->hm.s.fGIMTrapXcptUD)
1662 hmR0SvmAddXcptIntercept(pVmcb, X86_XCPT_UD);
1663 else
1664 hmR0SvmRemoveXcptIntercept(pVmcb, X86_XCPT_UD);
1665 HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_XCPT_INTERCEPTS);
1666 }
1667 return rc;
1668}
1669
1670
1671/**
1672 * Sets up the appropriate function to run guest code.
1673 *
1674 * @returns VBox status code.
1675 * @param pVCpu Pointer to the VMCPU.
1676 * @param pCtx Pointer to the guest-CPU context.
1677 *
1678 * @remarks No-long-jump zone!!!
1679 */
1680static int hmR0SvmSetupVMRunHandler(PVMCPU pVCpu, PCPUMCTX pCtx)
1681{
1682 if (CPUMIsGuestInLongModeEx(pCtx))
1683 {
1684#ifndef VBOX_ENABLE_64_BITS_GUESTS
1685 return VERR_PGM_UNSUPPORTED_SHADOW_PAGING_MODE;
1686#endif
1687 Assert(pVCpu->CTX_SUFF(pVM)->hm.s.fAllow64BitGuests); /* Guaranteed by hmR3InitFinalizeR0(). */
1688#if HC_ARCH_BITS == 32 && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
1689 /* 32-bit host. We need to switch to 64-bit before running the 64-bit guest. */
1690 pVCpu->hm.s.svm.pfnVMRun = SVMR0VMSwitcherRun64;
1691#else
1692 /* 64-bit host or hybrid host. */
1693 pVCpu->hm.s.svm.pfnVMRun = SVMR0VMRun64;
1694#endif
1695 }
1696 else
1697 {
1698 /* Guest is not in long mode, use the 32-bit handler. */
1699 pVCpu->hm.s.svm.pfnVMRun = SVMR0VMRun;
1700 }
1701 return VINF_SUCCESS;
1702}
1703
1704
1705/**
1706 * Enters the AMD-V session.
1707 *
1708 * @returns VBox status code.
1709 * @param pVM Pointer to the VM.
1710 * @param pVCpu Pointer to the VMCPU.
1711 * @param pCpu Pointer to the CPU info struct.
1712 */
1713VMMR0DECL(int) SVMR0Enter(PVM pVM, PVMCPU pVCpu, PHMGLOBALCPUINFO pCpu)
1714{
1715 AssertPtr(pVM);
1716 AssertPtr(pVCpu);
1717 Assert(pVM->hm.s.svm.fSupported);
1718 Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1719 NOREF(pVM); NOREF(pCpu);
1720
1721 LogFlowFunc(("pVM=%p pVCpu=%p\n", pVM, pVCpu));
1722 Assert(HMCPU_CF_IS_SET(pVCpu, HM_CHANGED_HOST_CONTEXT | HM_CHANGED_HOST_GUEST_SHARED_STATE));
1723
1724 pVCpu->hm.s.fLeaveDone = false;
1725 return VINF_SUCCESS;
1726}
1727
1728
1729/**
1730 * Thread-context callback for AMD-V.
1731 *
1732 * @param enmEvent The thread-context event.
1733 * @param pVCpu Pointer to the VMCPU.
1734 * @param fGlobalInit Whether global VT-x/AMD-V init. is used.
1735 * @thread EMT(pVCpu)
1736 */
1737VMMR0DECL(void) SVMR0ThreadCtxCallback(RTTHREADCTXEVENT enmEvent, PVMCPU pVCpu, bool fGlobalInit)
1738{
1739 NOREF(fGlobalInit);
1740
1741 switch (enmEvent)
1742 {
1743 case RTTHREADCTXEVENT_PREEMPTING:
1744 {
1745 Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1746 Assert(VMMR0ThreadCtxHooksAreRegistered(pVCpu));
1747 VMCPU_ASSERT_EMT(pVCpu);
1748
1749 PVM pVM = pVCpu->CTX_SUFF(pVM);
1750 PCPUMCTX pCtx = CPUMQueryGuestCtxPtr(pVCpu);
1751
1752 /* No longjmps (log-flush, locks) in this fragile context. */
1753 VMMRZCallRing3Disable(pVCpu);
1754
1755 if (!pVCpu->hm.s.fLeaveDone)
1756 {
1757 hmR0SvmLeave(pVM, pVCpu, pCtx);
1758 pVCpu->hm.s.fLeaveDone = true;
1759 }
1760
1761 /* Leave HM context, takes care of local init (term). */
1762 int rc = HMR0LeaveCpu(pVCpu);
1763 AssertRC(rc); NOREF(rc);
1764
1765 /* Restore longjmp state. */
1766 VMMRZCallRing3Enable(pVCpu);
1767 STAM_COUNTER_INC(&pVCpu->hm.s.StatPreemptPreempting);
1768 break;
1769 }
1770
1771 case RTTHREADCTXEVENT_RESUMED:
1772 {
1773 Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1774 Assert(VMMR0ThreadCtxHooksAreRegistered(pVCpu));
1775 VMCPU_ASSERT_EMT(pVCpu);
1776
1777 /* No longjmps (log-flush, locks) in this fragile context. */
1778 VMMRZCallRing3Disable(pVCpu);
1779
1780 /*
1781 * Initialize the bare minimum state required for HM. This takes care of
1782 * initializing AMD-V if necessary (onlined CPUs, local init etc.)
1783 */
1784 int rc = HMR0EnterCpu(pVCpu);
1785 AssertRC(rc); NOREF(rc);
1786 Assert(HMCPU_CF_IS_SET(pVCpu, HM_CHANGED_HOST_CONTEXT | HM_CHANGED_HOST_GUEST_SHARED_STATE));
1787
1788 pVCpu->hm.s.fLeaveDone = false;
1789
1790 /* Restore longjmp state. */
1791 VMMRZCallRing3Enable(pVCpu);
1792 break;
1793 }
1794
1795 default:
1796 break;
1797 }
1798}
1799
1800
1801/**
1802 * Saves the host state.
1803 *
1804 * @returns VBox status code.
1805 * @param pVM Pointer to the VM.
1806 * @param pVCpu Pointer to the VMCPU.
1807 *
1808 * @remarks No-long-jump zone!!!
1809 */
1810VMMR0DECL(int) SVMR0SaveHostState(PVM pVM, PVMCPU pVCpu)
1811{
1812 NOREF(pVM);
1813 NOREF(pVCpu);
1814 /* Nothing to do here. AMD-V does this for us automatically during the world-switch. */
1815 HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_HOST_CONTEXT);
1816 return VINF_SUCCESS;
1817}
1818
1819
1820/**
1821 * Loads the guest state into the VMCB.
1822 *
1823 * The CPU state will be loaded from these fields on every successful VM-entry.
1824 * Also sets up the appropriate VMRUN function to execute guest code based on
1825 * the guest CPU mode.
1826 *
1827 * @returns VBox status code.
1828 * @param pVM Pointer to the VM.
1829 * @param pVCpu Pointer to the VMCPU.
1830 * @param pCtx Pointer to the guest-CPU context.
1831 *
1832 * @remarks No-long-jump zone!!!
1833 */
1834static int hmR0SvmLoadGuestState(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
1835{
1836 PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
1837 AssertMsgReturn(pVmcb, ("Invalid pVmcb\n"), VERR_SVM_INVALID_PVMCB);
1838
1839 STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatLoadGuestState, x);
1840
1841 int rc = hmR0SvmLoadGuestControlRegs(pVCpu, pVmcb, pCtx);
1842 AssertLogRelMsgRCReturn(rc, ("hmR0SvmLoadGuestControlRegs! rc=%Rrc (pVM=%p pVCpu=%p)\n", rc, pVM, pVCpu), rc);
1843
1844 hmR0SvmLoadGuestSegmentRegs(pVCpu, pVmcb, pCtx);
1845 hmR0SvmLoadGuestMsrs(pVCpu, pVmcb, pCtx);
1846
1847 pVmcb->guest.u64RIP = pCtx->rip;
1848 pVmcb->guest.u64RSP = pCtx->rsp;
1849 pVmcb->guest.u64RFlags = pCtx->eflags.u32;
1850 pVmcb->guest.u64RAX = pCtx->rax;
1851
1852 rc = hmR0SvmLoadGuestApicState(pVCpu, pVmcb, pCtx);
1853 AssertLogRelMsgRCReturn(rc, ("hmR0SvmLoadGuestApicState! rc=%Rrc (pVM=%p pVCpu=%p)\n", rc, pVM, pVCpu), rc);
1854
1855 rc = hmR0SvmLoadGuestXcptIntercepts(pVCpu, pVmcb, pCtx);
1856 AssertLogRelMsgRCReturn(rc, ("hmR0SvmLoadGuestXcptIntercepts! rc=%Rrc (pVM=%p pVCpu=%p)\n", rc, pVM, pVCpu), rc);
1857
1858 rc = hmR0SvmSetupVMRunHandler(pVCpu, pCtx);
1859 AssertLogRelMsgRCReturn(rc, ("hmR0SvmSetupVMRunHandler! rc=%Rrc (pVM=%p pVCpu=%p)\n", rc, pVM, pVCpu), rc);
1860
1861 /* Clear any unused and reserved bits. */
1862 HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_RIP /* Unused (loaded unconditionally). */
1863 | HM_CHANGED_GUEST_RSP
1864 | HM_CHANGED_GUEST_RFLAGS
1865 | HM_CHANGED_GUEST_SYSENTER_CS_MSR
1866 | HM_CHANGED_GUEST_SYSENTER_EIP_MSR
1867 | HM_CHANGED_GUEST_SYSENTER_ESP_MSR
1868 | HM_CHANGED_GUEST_LAZY_MSRS /* Unused. */
1869 | HM_CHANGED_SVM_RESERVED1 /* Reserved. */
1870 | HM_CHANGED_SVM_RESERVED2
1871 | HM_CHANGED_SVM_RESERVED3
1872 | HM_CHANGED_SVM_RESERVED4);
1873
1874 /* All the guest state bits should be loaded except maybe the host context and/or shared host/guest bits. */
1875 AssertMsg( !HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_ALL_GUEST)
1876 || HMCPU_CF_IS_PENDING_ONLY(pVCpu, HM_CHANGED_HOST_CONTEXT | HM_CHANGED_HOST_GUEST_SHARED_STATE),
1877 ("fContextUseFlags=%#RX32\n", HMCPU_CF_VALUE(pVCpu)));
1878
1879 Log4(("Load: CS:RIP=%04x:%RX64 EFL=%#x SS:RSP=%04x:%RX64\n", pCtx->cs.Sel, pCtx->rip, pCtx->eflags.u, pCtx->ss, pCtx->rsp));
1880 STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatLoadGuestState, x);
1881 return rc;
1882}
1883
1884
1885/**
1886 * Loads the state shared between the host and guest into the
1887 * VMCB.
1888 *
1889 * @param pVCpu Pointer to the VMCPU.
1890 * @param pVmcb Pointer to the VM control block.
1891 * @param pCtx Pointer to the guest-CPU context.
1892 *
1893 * @remarks No-long-jump zone!!!
1894 */
1895static void hmR0SvmLoadSharedState(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx)
1896{
1897 Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1898 Assert(!VMMRZCallRing3IsEnabled(pVCpu));
1899
1900 if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR0))
1901 hmR0SvmLoadSharedCR0(pVCpu, pVmcb, pCtx);
1902
1903 if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_DEBUG))
1904 hmR0SvmLoadSharedDebugState(pVCpu, pVmcb, pCtx);
1905
1906 /* Unused on AMD-V. */
1907 HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_GUEST_LAZY_MSRS);
1908
1909 AssertMsg(!HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_HOST_GUEST_SHARED_STATE),
1910 ("fContextUseFlags=%#RX32\n", HMCPU_CF_VALUE(pVCpu)));
1911}
1912
1913
1914/**
1915 * Saves the entire guest state from the VMCB into the
1916 * guest-CPU context. Currently there is no residual state left in the CPU that
1917 * is not updated in the VMCB.
1918 *
1919 * @returns VBox status code.
1920 * @param pVCpu Pointer to the VMCPU.
1921 * @param pMixedCtx Pointer to the guest-CPU context. The data may be
1922 * out-of-sync. Make sure to update the required fields
1923 * before using them.
1924 */
1925static void hmR0SvmSaveGuestState(PVMCPU pVCpu, PCPUMCTX pMixedCtx)
1926{
1927 Assert(VMMRZCallRing3IsEnabled(pVCpu));
1928
1929 PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
1930
1931 pMixedCtx->rip = pVmcb->guest.u64RIP;
1932 pMixedCtx->rsp = pVmcb->guest.u64RSP;
1933 pMixedCtx->eflags.u32 = pVmcb->guest.u64RFlags;
1934 pMixedCtx->rax = pVmcb->guest.u64RAX;
1935
1936 /*
1937 * Guest interrupt shadow.
1938 */
1939 if (pVmcb->ctrl.u64IntShadow & SVM_INTERRUPT_SHADOW_ACTIVE)
1940 EMSetInhibitInterruptsPC(pVCpu, pMixedCtx->rip);
1941 else if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS))
1942 VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS);
1943
1944 /*
1945 * Guest Control registers: CR2, CR3 (handled at the end) - accesses to other control registers are always intercepted.
1946 */
1947 pMixedCtx->cr2 = pVmcb->guest.u64CR2;
1948
1949 /*
1950 * Guest MSRs.
1951 */
1952 pMixedCtx->msrSTAR = pVmcb->guest.u64STAR; /* legacy syscall eip, cs & ss */
1953 pMixedCtx->msrLSTAR = pVmcb->guest.u64LSTAR; /* 64-bit mode syscall rip */
1954 pMixedCtx->msrCSTAR = pVmcb->guest.u64CSTAR; /* compatibility mode syscall rip */
1955 pMixedCtx->msrSFMASK = pVmcb->guest.u64SFMASK; /* syscall flag mask */
1956 pMixedCtx->msrKERNELGSBASE = pVmcb->guest.u64KernelGSBase; /* swapgs exchange value */
1957 pMixedCtx->SysEnter.cs = pVmcb->guest.u64SysEnterCS;
1958 pMixedCtx->SysEnter.eip = pVmcb->guest.u64SysEnterEIP;
1959 pMixedCtx->SysEnter.esp = pVmcb->guest.u64SysEnterESP;
1960
1961 /*
1962 * Guest segment registers (includes FS, GS base MSRs for 64-bit guests).
1963 */
1964 HMSVM_SAVE_SEG_REG(CS, cs);
1965 HMSVM_SAVE_SEG_REG(SS, ss);
1966 HMSVM_SAVE_SEG_REG(DS, ds);
1967 HMSVM_SAVE_SEG_REG(ES, es);
1968 HMSVM_SAVE_SEG_REG(FS, fs);
1969 HMSVM_SAVE_SEG_REG(GS, gs);
1970
1971 /*
1972 * Correct the hidden CS granularity bit. Haven't seen it being wrong in any other
1973 * register (yet).
1974 */
1975 /** @todo SELM might need to be fixed as it too should not care about the
1976 * granularity bit. See @bugref{6785}. */
1977 if ( !pMixedCtx->cs.Attr.n.u1Granularity
1978 && pMixedCtx->cs.Attr.n.u1Present
1979 && pMixedCtx->cs.u32Limit > UINT32_C(0xfffff))
1980 {
1981 Assert((pMixedCtx->cs.u32Limit & 0xfff) == 0xfff);
1982 pMixedCtx->cs.Attr.n.u1Granularity = 1;
1983 }
1984
1985#ifdef VBOX_STRICT
1986# define HMSVM_ASSERT_SEG_GRANULARITY(reg) \
1987 AssertMsg( !pMixedCtx->reg.Attr.n.u1Present \
1988 || ( pMixedCtx->reg.Attr.n.u1Granularity \
1989 ? (pMixedCtx->reg.u32Limit & 0xfff) == 0xfff \
1990 : pMixedCtx->reg.u32Limit <= UINT32_C(0xfffff)), \
1991 ("Invalid Segment Attributes Limit=%#RX32 Attr=%#RX32 Base=%#RX64\n", pMixedCtx->reg.u32Limit, \
1992 pMixedCtx->reg.Attr.u, pMixedCtx->reg.u64Base))
1993
1994 HMSVM_ASSERT_SEG_GRANULARITY(cs);
1995 HMSVM_ASSERT_SEG_GRANULARITY(ss);
1996 HMSVM_ASSERT_SEG_GRANULARITY(ds);
1997 HMSVM_ASSERT_SEG_GRANULARITY(es);
1998 HMSVM_ASSERT_SEG_GRANULARITY(fs);
1999 HMSVM_ASSERT_SEG_GRANULARITY(gs);
2000
2001# undef HMSVM_ASSERT_SEL_GRANULARITY
2002#endif
2003
2004 /*
2005 * Sync the hidden SS DPL field. AMD CPUs have a separate CPL field in the VMCB and uses that
2006 * and thus it's possible that when the CPL changes during guest execution that the SS DPL
2007 * isn't updated by AMD-V. Observed on some AMD Fusion CPUs with 64-bit guests.
2008 * See AMD spec. 15.5.1 "Basic operation".
2009 */
2010 Assert(!(pVmcb->guest.u8CPL & ~0x3));
2011 pMixedCtx->ss.Attr.n.u2Dpl = pVmcb->guest.u8CPL & 0x3;
2012
2013 /*
2014 * Guest TR.
2015 * Fixup TR attributes so it's compatible with Intel. Important when saved-states are used
2016 * between Intel and AMD. See @bugref{6208} comment #39.
2017 */
2018 HMSVM_SAVE_SEG_REG(TR, tr);
2019 if (CPUMIsGuestInLongModeEx(pMixedCtx))
2020 pMixedCtx->tr.Attr.n.u4Type = X86_SEL_TYPE_SYS_386_TSS_BUSY;
2021
2022 /*
2023 * Guest Descriptor-Table registers.
2024 */
2025 HMSVM_SAVE_SEG_REG(LDTR, ldtr);
2026 pMixedCtx->gdtr.cbGdt = pVmcb->guest.GDTR.u32Limit;
2027 pMixedCtx->gdtr.pGdt = pVmcb->guest.GDTR.u64Base;
2028
2029 pMixedCtx->idtr.cbIdt = pVmcb->guest.IDTR.u32Limit;
2030 pMixedCtx->idtr.pIdt = pVmcb->guest.IDTR.u64Base;
2031
2032 /*
2033 * Guest Debug registers.
2034 */
2035 if (!pVCpu->hm.s.fUsingHyperDR7)
2036 {
2037 pMixedCtx->dr[6] = pVmcb->guest.u64DR6;
2038 pMixedCtx->dr[7] = pVmcb->guest.u64DR7;
2039 }
2040 else
2041 {
2042 Assert(pVmcb->guest.u64DR7 == CPUMGetHyperDR7(pVCpu));
2043 CPUMSetHyperDR6(pVCpu, pVmcb->guest.u64DR6);
2044 }
2045
2046 /*
2047 * With Nested Paging, CR3 changes are not intercepted. Therefore, sync. it now.
2048 * This is done as the very last step of syncing the guest state, as PGMUpdateCR3() may cause longjmp's to ring-3.
2049 */
2050 if ( pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging
2051 && pMixedCtx->cr3 != pVmcb->guest.u64CR3)
2052 {
2053 CPUMSetGuestCR3(pVCpu, pVmcb->guest.u64CR3);
2054 PGMUpdateCR3(pVCpu, pVmcb->guest.u64CR3);
2055 }
2056}
2057
2058
2059/**
2060 * Does the necessary state syncing before returning to ring-3 for any reason
2061 * (longjmp, preemption, voluntary exits to ring-3) from AMD-V.
2062 *
2063 * @param pVM Pointer to the VM.
2064 * @param pVCpu Pointer to the VMCPU.
2065 * @param pMixedCtx Pointer to the guest-CPU context.
2066 *
2067 * @remarks No-long-jmp zone!!!
2068 */
2069static void hmR0SvmLeave(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
2070{
2071 Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2072 Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2073 Assert(VMMR0IsLogFlushDisabled(pVCpu));
2074
2075 /*
2076 * !!! IMPORTANT !!!
2077 * If you modify code here, make sure to check whether hmR0SvmCallRing3Callback() needs to be updated too.
2078 */
2079
2080 /* Restore host FPU state if necessary and resync on next R0 reentry .*/
2081 if (CPUMIsGuestFPUStateActive(pVCpu))
2082 {
2083 CPUMR0SaveGuestFPU(pVM, pVCpu, pCtx);
2084 Assert(!CPUMIsGuestFPUStateActive(pVCpu));
2085 HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR0);
2086 }
2087
2088 /*
2089 * Restore host debug registers if necessary and resync on next R0 reentry.
2090 */
2091#ifdef VBOX_STRICT
2092 if (CPUMIsHyperDebugStateActive(pVCpu))
2093 {
2094 PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2095 Assert(pVmcb->ctrl.u16InterceptRdDRx == 0xffff);
2096 Assert(pVmcb->ctrl.u16InterceptWrDRx == 0xffff);
2097 }
2098#endif
2099 if (CPUMR0DebugStateMaybeSaveGuestAndRestoreHost(pVCpu, false /* save DR6 */))
2100 HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_DEBUG);
2101
2102 Assert(!CPUMIsHyperDebugStateActive(pVCpu));
2103 Assert(!CPUMIsGuestDebugStateActive(pVCpu));
2104
2105 STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatEntry);
2106 STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatLoadGuestState);
2107 STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatExit1);
2108 STAM_PROFILE_ADV_SET_STOPPED(&pVCpu->hm.s.StatExit2);
2109 STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchLongJmpToR3);
2110
2111 VMCPU_CMPXCHG_STATE(pVCpu, VMCPUSTATE_STARTED_HM, VMCPUSTATE_STARTED_EXEC);
2112}
2113
2114
2115/**
2116 * Leaves the AMD-V session.
2117 *
2118 * @returns VBox status code.
2119 * @param pVM Pointer to the VM.
2120 * @param pVCpu Pointer to the VMCPU.
2121 * @param pCtx Pointer to the guest-CPU context.
2122 */
2123static int hmR0SvmLeaveSession(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
2124{
2125 HM_DISABLE_PREEMPT();
2126 Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2127 Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2128
2129 /* When thread-context hooks are used, we can avoid doing the leave again if we had been preempted before
2130 and done this from the SVMR0ThreadCtxCallback(). */
2131 if (!pVCpu->hm.s.fLeaveDone)
2132 {
2133 hmR0SvmLeave(pVM, pVCpu, pCtx);
2134 pVCpu->hm.s.fLeaveDone = true;
2135 }
2136
2137 /*
2138 * !!! IMPORTANT !!!
2139 * If you modify code here, make sure to check whether hmR0SvmCallRing3Callback() needs to be updated too.
2140 */
2141
2142 /* Deregister hook now that we've left HM context before re-enabling preemption. */
2143 VMMR0ThreadCtxHooksDeregister(pVCpu);
2144
2145 /* Leave HM context. This takes care of local init (term). */
2146 int rc = HMR0LeaveCpu(pVCpu);
2147
2148 HM_RESTORE_PREEMPT();
2149 return rc;
2150}
2151
2152
2153/**
2154 * Does the necessary state syncing before doing a longjmp to ring-3.
2155 *
2156 * @returns VBox status code.
2157 * @param pVM Pointer to the VM.
2158 * @param pVCpu Pointer to the VMCPU.
2159 * @param pCtx Pointer to the guest-CPU context.
2160 *
2161 * @remarks No-long-jmp zone!!!
2162 */
2163static int hmR0SvmLongJmpToRing3(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
2164{
2165 return hmR0SvmLeaveSession(pVM, pVCpu, pCtx);
2166}
2167
2168
2169/**
2170 * VMMRZCallRing3() callback wrapper which saves the guest state (or restores
2171 * any remaining host state) before we longjump to ring-3 and possibly get
2172 * preempted.
2173 *
2174 * @param pVCpu Pointer to the VMCPU.
2175 * @param enmOperation The operation causing the ring-3 longjump.
2176 * @param pvUser The user argument (pointer to the possibly
2177 * out-of-date guest-CPU context).
2178 */
2179DECLCALLBACK(int) hmR0SvmCallRing3Callback(PVMCPU pVCpu, VMMCALLRING3 enmOperation, void *pvUser)
2180{
2181 if (enmOperation == VMMCALLRING3_VM_R0_ASSERTION)
2182 {
2183 /*
2184 * !!! IMPORTANT !!!
2185 * If you modify code here, make sure to check whether hmR0SvmLeave() and hmR0SvmLeaveSession() needs
2186 * to be updated too. This is a stripped down version which gets out ASAP trying to not trigger any assertion.
2187 */
2188 VMMRZCallRing3RemoveNotification(pVCpu);
2189 VMMRZCallRing3Disable(pVCpu);
2190 HM_DISABLE_PREEMPT();
2191
2192 /* Restore host FPU state if necessary and resync on next R0 reentry .*/
2193 if (CPUMIsGuestFPUStateActive(pVCpu))
2194 CPUMR0SaveGuestFPU(pVCpu->CTX_SUFF(pVM), pVCpu, (PCPUMCTX)pvUser);
2195
2196 /* Restore host debug registers if necessary and resync on next R0 reentry. */
2197 CPUMR0DebugStateMaybeSaveGuestAndRestoreHost(pVCpu, false /* save DR6 */);
2198
2199 /* Deregister the hook now that we've left HM context before re-enabling preemption. */
2200 VMMR0ThreadCtxHooksDeregister(pVCpu);
2201
2202 /* Leave HM context. This takes care of local init (term). */
2203 HMR0LeaveCpu(pVCpu);
2204
2205 HM_RESTORE_PREEMPT();
2206 return VINF_SUCCESS;
2207 }
2208
2209 Assert(pVCpu);
2210 Assert(pvUser);
2211 Assert(VMMRZCallRing3IsEnabled(pVCpu));
2212 HMSVM_ASSERT_PREEMPT_SAFE();
2213
2214 VMMRZCallRing3Disable(pVCpu);
2215 Assert(VMMR0IsLogFlushDisabled(pVCpu));
2216
2217 Log4(("hmR0SvmCallRing3Callback->hmR0SvmLongJmpToRing3\n"));
2218 int rc = hmR0SvmLongJmpToRing3(pVCpu->CTX_SUFF(pVM), pVCpu, (PCPUMCTX)pvUser);
2219 AssertRCReturn(rc, rc);
2220
2221 VMMRZCallRing3Enable(pVCpu);
2222 return VINF_SUCCESS;
2223}
2224
2225
2226/**
2227 * Take necessary actions before going back to ring-3.
2228 *
2229 * An action requires us to go back to ring-3. This function does the necessary
2230 * steps before we can safely return to ring-3. This is not the same as longjmps
2231 * to ring-3, this is voluntary.
2232 *
2233 * @param pVM Pointer to the VM.
2234 * @param pVCpu Pointer to the VMCPU.
2235 * @param pCtx Pointer to the guest-CPU context.
2236 * @param rcExit The reason for exiting to ring-3. Can be
2237 * VINF_VMM_UNKNOWN_RING3_CALL.
2238 */
2239static void hmR0SvmExitToRing3(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, int rcExit)
2240{
2241 Assert(pVM);
2242 Assert(pVCpu);
2243 Assert(pCtx);
2244 HMSVM_ASSERT_PREEMPT_SAFE();
2245
2246 /* Please, no longjumps here (any logging shouldn't flush jump back to ring-3). NO LOGGING BEFORE THIS POINT! */
2247 VMMRZCallRing3Disable(pVCpu);
2248 Log4(("hmR0SvmExitToRing3: rcExit=%d\n", rcExit));
2249
2250 /* We need to do this only while truly exiting the "inner loop" back to ring-3 and -not- for any longjmp to ring3. */
2251 if (pVCpu->hm.s.Event.fPending)
2252 {
2253 hmR0SvmPendingEventToTrpmTrap(pVCpu);
2254 Assert(!pVCpu->hm.s.Event.fPending);
2255 }
2256
2257 /* If we're emulating an instruction, we shouldn't have any TRPM traps pending
2258 and if we're injecting an event we should have a TRPM trap pending. */
2259 Assert(rcExit != VINF_EM_RAW_INJECT_TRPM_EVENT || TRPMHasTrap(pVCpu));
2260 Assert(rcExit != VINF_EM_RAW_EMULATE_INSTR || !TRPMHasTrap(pVCpu));
2261
2262 /* Sync. the necessary state for going back to ring-3. */
2263 hmR0SvmLeaveSession(pVM, pVCpu, pCtx);
2264 STAM_COUNTER_DEC(&pVCpu->hm.s.StatSwitchLongJmpToR3);
2265
2266 VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_TO_R3);
2267 CPUMSetChangedFlags(pVCpu, CPUM_CHANGED_SYSENTER_MSR
2268 | CPUM_CHANGED_LDTR
2269 | CPUM_CHANGED_GDTR
2270 | CPUM_CHANGED_IDTR
2271 | CPUM_CHANGED_TR
2272 | CPUM_CHANGED_HIDDEN_SEL_REGS);
2273 if ( pVM->hm.s.fNestedPaging
2274 && CPUMIsGuestPagingEnabledEx(pCtx))
2275 {
2276 CPUMSetChangedFlags(pVCpu, CPUM_CHANGED_GLOBAL_TLB_FLUSH);
2277 }
2278
2279 /* On our way back from ring-3 reload the guest state if there is a possibility of it being changed. */
2280 if (rcExit != VINF_EM_RAW_INTERRUPT)
2281 HMCPU_CF_SET(pVCpu, HM_CHANGED_ALL_GUEST);
2282
2283 STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchExitToR3);
2284
2285 /* We do -not- want any longjmp notifications after this! We must return to ring-3 ASAP. */
2286 VMMRZCallRing3RemoveNotification(pVCpu);
2287 VMMRZCallRing3Enable(pVCpu);
2288}
2289
2290
2291/**
2292 * Updates the use of TSC offsetting mode for the CPU and adjusts the necessary
2293 * intercepts.
2294 *
2295 * @param pVM The shared VM handle.
2296 * @param pVCpu Pointer to the VMCPU.
2297 *
2298 * @remarks No-long-jump zone!!!
2299 */
2300static void hmR0SvmUpdateTscOffsetting(PVM pVM, PVMCPU pVCpu)
2301{
2302 bool fParavirtTsc;
2303 PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2304 bool fCanUseRealTsc = TMCpuTickCanUseRealTSC(pVM, pVCpu, &pVmcb->ctrl.u64TSCOffset, &fParavirtTsc);
2305 if (fCanUseRealTsc)
2306 {
2307 pVmcb->ctrl.u32InterceptCtrl1 &= ~SVM_CTRL1_INTERCEPT_RDTSC;
2308 pVmcb->ctrl.u32InterceptCtrl2 &= ~SVM_CTRL2_INTERCEPT_RDTSCP;
2309 STAM_COUNTER_INC(&pVCpu->hm.s.StatTscOffset);
2310 }
2311 else
2312 {
2313 pVmcb->ctrl.u32InterceptCtrl1 |= SVM_CTRL1_INTERCEPT_RDTSC;
2314 pVmcb->ctrl.u32InterceptCtrl2 |= SVM_CTRL2_INTERCEPT_RDTSCP;
2315 STAM_COUNTER_INC(&pVCpu->hm.s.StatTscIntercept);
2316 }
2317 pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
2318
2319 /** @todo later optimize this to be done elsewhere and not before every
2320 * VM-entry. */
2321 if (fParavirtTsc)
2322 {
2323 int rc = GIMR0UpdateParavirtTsc(pVM, 0 /* u64Offset */);
2324 AssertRC(rc);
2325 STAM_COUNTER_INC(&pVCpu->hm.s.StatTscParavirt);
2326 }
2327}
2328
2329
2330/**
2331 * Sets an event as a pending event to be injected into the guest.
2332 *
2333 * @param pVCpu Pointer to the VMCPU.
2334 * @param pEvent Pointer to the SVM event.
2335 * @param GCPtrFaultAddress The fault-address (CR2) in case it's a
2336 * page-fault.
2337 *
2338 * @remarks Statistics counter assumes this is a guest event being reflected to
2339 * the guest i.e. 'StatInjectPendingReflect' is incremented always.
2340 */
2341DECLINLINE(void) hmR0SvmSetPendingEvent(PVMCPU pVCpu, PSVMEVENT pEvent, RTGCUINTPTR GCPtrFaultAddress)
2342{
2343 Assert(!pVCpu->hm.s.Event.fPending);
2344 Assert(pEvent->n.u1Valid);
2345
2346 pVCpu->hm.s.Event.u64IntInfo = pEvent->u;
2347 pVCpu->hm.s.Event.fPending = true;
2348 pVCpu->hm.s.Event.GCPtrFaultAddress = GCPtrFaultAddress;
2349
2350 Log4(("hmR0SvmSetPendingEvent: u=%#RX64 u8Vector=%#x Type=%#x ErrorCodeValid=%RTbool ErrorCode=%#RX32\n", pEvent->u,
2351 pEvent->n.u8Vector, (uint8_t)pEvent->n.u3Type, !!pEvent->n.u1ErrorCodeValid, pEvent->n.u32ErrorCode));
2352
2353 STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectPendingReflect);
2354}
2355
2356
2357/**
2358 * Injects an event into the guest upon VMRUN by updating the relevant field
2359 * in the VMCB.
2360 *
2361 * @param pVCpu Pointer to the VMCPU.
2362 * @param pVmcb Pointer to the guest VM control block.
2363 * @param pCtx Pointer to the guest-CPU context.
2364 * @param pEvent Pointer to the event.
2365 *
2366 * @remarks No-long-jump zone!!!
2367 * @remarks Requires CR0!
2368 */
2369DECLINLINE(void) hmR0SvmInjectEventVmcb(PVMCPU pVCpu, PSVMVMCB pVmcb, PCPUMCTX pCtx, PSVMEVENT pEvent)
2370{
2371 NOREF(pVCpu); NOREF(pCtx);
2372
2373 pVmcb->ctrl.EventInject.u = pEvent->u;
2374 STAM_COUNTER_INC(&pVCpu->hm.s.paStatInjectedIrqsR0[pEvent->n.u8Vector & MASK_INJECT_IRQ_STAT]);
2375
2376 Log4(("hmR0SvmInjectEventVmcb: u=%#RX64 u8Vector=%#x Type=%#x ErrorCodeValid=%RTbool ErrorCode=%#RX32\n", pEvent->u,
2377 pEvent->n.u8Vector, (uint8_t)pEvent->n.u3Type, !!pEvent->n.u1ErrorCodeValid, pEvent->n.u32ErrorCode));
2378}
2379
2380
2381
2382/**
2383 * Converts any TRPM trap into a pending HM event. This is typically used when
2384 * entering from ring-3 (not longjmp returns).
2385 *
2386 * @param pVCpu Pointer to the VMCPU.
2387 */
2388static void hmR0SvmTrpmTrapToPendingEvent(PVMCPU pVCpu)
2389{
2390 Assert(TRPMHasTrap(pVCpu));
2391 Assert(!pVCpu->hm.s.Event.fPending);
2392
2393 uint8_t uVector;
2394 TRPMEVENT enmTrpmEvent;
2395 RTGCUINT uErrCode;
2396 RTGCUINTPTR GCPtrFaultAddress;
2397 uint8_t cbInstr;
2398
2399 int rc = TRPMQueryTrapAll(pVCpu, &uVector, &enmTrpmEvent, &uErrCode, &GCPtrFaultAddress, &cbInstr);
2400 AssertRC(rc);
2401
2402 SVMEVENT Event;
2403 Event.u = 0;
2404 Event.n.u1Valid = 1;
2405 Event.n.u8Vector = uVector;
2406
2407 /* Refer AMD spec. 15.20 "Event Injection" for the format. */
2408 if (enmTrpmEvent == TRPM_TRAP)
2409 {
2410 Event.n.u3Type = SVM_EVENT_EXCEPTION;
2411 switch (uVector)
2412 {
2413 case X86_XCPT_NMI:
2414 {
2415 Event.n.u3Type = SVM_EVENT_NMI;
2416 break;
2417 }
2418
2419 case X86_XCPT_PF:
2420 case X86_XCPT_DF:
2421 case X86_XCPT_TS:
2422 case X86_XCPT_NP:
2423 case X86_XCPT_SS:
2424 case X86_XCPT_GP:
2425 case X86_XCPT_AC:
2426 {
2427 Event.n.u1ErrorCodeValid = 1;
2428 Event.n.u32ErrorCode = uErrCode;
2429 break;
2430 }
2431 }
2432 }
2433 else if (enmTrpmEvent == TRPM_HARDWARE_INT)
2434 Event.n.u3Type = SVM_EVENT_EXTERNAL_IRQ;
2435 else if (enmTrpmEvent == TRPM_SOFTWARE_INT)
2436 Event.n.u3Type = SVM_EVENT_SOFTWARE_INT;
2437 else
2438 AssertMsgFailed(("Invalid TRPM event type %d\n", enmTrpmEvent));
2439
2440 rc = TRPMResetTrap(pVCpu);
2441 AssertRC(rc);
2442
2443 Log4(("TRPM->HM event: u=%#RX64 u8Vector=%#x uErrorCodeValid=%RTbool uErrorCode=%#RX32\n", Event.u, Event.n.u8Vector,
2444 !!Event.n.u1ErrorCodeValid, Event.n.u32ErrorCode));
2445
2446 hmR0SvmSetPendingEvent(pVCpu, &Event, GCPtrFaultAddress);
2447 STAM_COUNTER_DEC(&pVCpu->hm.s.StatInjectPendingReflect);
2448}
2449
2450
2451/**
2452 * Converts any pending SVM event into a TRPM trap. Typically used when leaving
2453 * AMD-V to execute any instruction.
2454 *
2455 * @param pvCpu Pointer to the VMCPU.
2456 */
2457static void hmR0SvmPendingEventToTrpmTrap(PVMCPU pVCpu)
2458{
2459 Assert(pVCpu->hm.s.Event.fPending);
2460 Assert(TRPMQueryTrap(pVCpu, NULL /* pu8TrapNo */, NULL /* pEnmType */) == VERR_TRPM_NO_ACTIVE_TRAP);
2461
2462 SVMEVENT Event;
2463 Event.u = pVCpu->hm.s.Event.u64IntInfo;
2464
2465 uint8_t uVector = Event.n.u8Vector;
2466 uint8_t uVectorType = Event.n.u3Type;
2467
2468 TRPMEVENT enmTrapType;
2469 switch (uVectorType)
2470 {
2471 case SVM_EVENT_EXTERNAL_IRQ:
2472 enmTrapType = TRPM_HARDWARE_INT;
2473 break;
2474 case SVM_EVENT_SOFTWARE_INT:
2475 enmTrapType = TRPM_SOFTWARE_INT;
2476 break;
2477 case SVM_EVENT_EXCEPTION:
2478 case SVM_EVENT_NMI:
2479 enmTrapType = TRPM_TRAP;
2480 break;
2481 default:
2482 AssertMsgFailed(("Invalid pending-event type %#x\n", uVectorType));
2483 enmTrapType = TRPM_32BIT_HACK;
2484 break;
2485 }
2486
2487 Log4(("HM event->TRPM: uVector=%#x enmTrapType=%d\n", uVector, uVectorType));
2488
2489 int rc = TRPMAssertTrap(pVCpu, uVector, enmTrapType);
2490 AssertRC(rc);
2491
2492 if (Event.n.u1ErrorCodeValid)
2493 TRPMSetErrorCode(pVCpu, Event.n.u32ErrorCode);
2494
2495 if ( uVectorType == SVM_EVENT_EXCEPTION
2496 && uVector == X86_XCPT_PF)
2497 {
2498 TRPMSetFaultAddress(pVCpu, pVCpu->hm.s.Event.GCPtrFaultAddress);
2499 Assert(pVCpu->hm.s.Event.GCPtrFaultAddress == CPUMGetGuestCR2(pVCpu));
2500 }
2501 else if (uVectorType == SVM_EVENT_SOFTWARE_INT)
2502 {
2503 AssertMsg( uVectorType == SVM_EVENT_SOFTWARE_INT
2504 || (uVector == X86_XCPT_BP || uVector == X86_XCPT_OF),
2505 ("Invalid vector: uVector=%#x uVectorType=%#x\n", uVector, uVectorType));
2506 TRPMSetInstrLength(pVCpu, pVCpu->hm.s.Event.cbInstr);
2507 }
2508 pVCpu->hm.s.Event.fPending = false;
2509}
2510
2511
2512/**
2513 * Gets the guest's interrupt-shadow.
2514 *
2515 * @returns The guest's interrupt-shadow.
2516 * @param pVCpu Pointer to the VMCPU.
2517 * @param pCtx Pointer to the guest-CPU context.
2518 *
2519 * @remarks No-long-jump zone!!!
2520 * @remarks Has side-effects with VMCPU_FF_INHIBIT_INTERRUPTS force-flag.
2521 */
2522DECLINLINE(uint32_t) hmR0SvmGetGuestIntrShadow(PVMCPU pVCpu, PCPUMCTX pCtx)
2523{
2524 /*
2525 * Instructions like STI and MOV SS inhibit interrupts till the next instruction completes. Check if we should
2526 * inhibit interrupts or clear any existing interrupt-inhibition.
2527 */
2528 uint32_t uIntrState = 0;
2529 if (VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS))
2530 {
2531 if (pCtx->rip != EMGetInhibitInterruptsPC(pVCpu))
2532 {
2533 /*
2534 * We can clear the inhibit force flag as even if we go back to the recompiler without executing guest code in
2535 * AMD-V, the flag's condition to be cleared is met and thus the cleared state is correct.
2536 */
2537 VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INHIBIT_INTERRUPTS);
2538 }
2539 else
2540 uIntrState = SVM_INTERRUPT_SHADOW_ACTIVE;
2541 }
2542 return uIntrState;
2543}
2544
2545
2546/**
2547 * Sets the virtual interrupt intercept control in the VMCB which
2548 * instructs AMD-V to cause a #VMEXIT as soon as the guest is in a state to
2549 * receive interrupts.
2550 *
2551 * @param pVmcb Pointer to the VM control block.
2552 */
2553DECLINLINE(void) hmR0SvmSetVirtIntrIntercept(PSVMVMCB pVmcb)
2554{
2555 if (!(pVmcb->ctrl.u32InterceptCtrl1 & SVM_CTRL1_INTERCEPT_VINTR))
2556 {
2557 pVmcb->ctrl.IntCtrl.n.u1VIrqValid = 1; /* A virtual interrupt is pending. */
2558 pVmcb->ctrl.IntCtrl.n.u8VIrqVector = 0; /* Not necessary as we #VMEXIT for delivering the interrupt. */
2559 pVmcb->ctrl.u32InterceptCtrl1 |= SVM_CTRL1_INTERCEPT_VINTR;
2560 pVmcb->ctrl.u64VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS | HMSVM_VMCB_CLEAN_TPR);
2561
2562 Log4(("Setting VINTR intercept\n"));
2563 }
2564}
2565
2566
2567/**
2568 * Sets the IRET intercept control in the VMCB which instructs AMD-V to cause a
2569 * #VMEXIT as soon as a guest starts executing an IRET. This is used to unblock
2570 * virtual NMIs.
2571 *
2572 * @param pVmcb Pointer to the VM control block.
2573 */
2574DECLINLINE(void) hmR0SvmSetIretIntercept(PSVMVMCB pVmcb)
2575{
2576 if (!(pVmcb->ctrl.u32InterceptCtrl1 & SVM_CTRL1_INTERCEPT_IRET))
2577 {
2578 pVmcb->ctrl.u32InterceptCtrl1 |= SVM_CTRL1_INTERCEPT_IRET;
2579 pVmcb->ctrl.u64VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS);
2580
2581 Log4(("Setting IRET intercept\n"));
2582 }
2583}
2584
2585
2586/**
2587 * Clears the IRET intercept control in the VMCB.
2588 *
2589 * @param pVmcb Pointer to the VM control block.
2590 */
2591DECLINLINE(void) hmR0SvmClearIretIntercept(PSVMVMCB pVmcb)
2592{
2593 if (pVmcb->ctrl.u32InterceptCtrl1 & SVM_CTRL1_INTERCEPT_IRET)
2594 {
2595 pVmcb->ctrl.u32InterceptCtrl1 &= ~SVM_CTRL1_INTERCEPT_IRET;
2596 pVmcb->ctrl.u64VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS);
2597
2598 Log4(("Clearing IRET intercept\n"));
2599 }
2600}
2601
2602
2603/**
2604 * Evaluates the event to be delivered to the guest and sets it as the pending
2605 * event.
2606 *
2607 * @param pVCpu Pointer to the VMCPU.
2608 * @param pCtx Pointer to the guest-CPU context.
2609 */
2610static void hmR0SvmEvaluatePendingEvent(PVMCPU pVCpu, PCPUMCTX pCtx)
2611{
2612 Assert(!pVCpu->hm.s.Event.fPending);
2613 Log4Func(("\n"));
2614
2615 bool const fIntShadow = RT_BOOL(hmR0SvmGetGuestIntrShadow(pVCpu, pCtx));
2616 bool const fBlockInt = !(pCtx->eflags.u32 & X86_EFL_IF);
2617 bool const fBlockNmi = RT_BOOL(VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_BLOCK_NMIS));
2618 PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2619
2620 SVMEVENT Event;
2621 Event.u = 0;
2622 /** @todo SMI. SMIs take priority over NMIs. */
2623 if (VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_INTERRUPT_NMI)) /* NMI. NMIs take priority over regular interrupts . */
2624 {
2625 if (fBlockNmi)
2626 hmR0SvmSetIretIntercept(pVmcb);
2627 else if (fIntShadow)
2628 hmR0SvmSetVirtIntrIntercept(pVmcb);
2629 else
2630 {
2631 Log4(("Pending NMI\n"));
2632
2633 Event.n.u1Valid = 1;
2634 Event.n.u8Vector = X86_XCPT_NMI;
2635 Event.n.u3Type = SVM_EVENT_NMI;
2636
2637 hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
2638 hmR0SvmSetIretIntercept(pVmcb);
2639 VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_INTERRUPT_NMI);
2640 }
2641 }
2642 else if (VMCPU_FF_IS_PENDING(pVCpu, (VMCPU_FF_INTERRUPT_APIC | VMCPU_FF_INTERRUPT_PIC)))
2643 {
2644 /*
2645 * Check if the guest can receive external interrupts (PIC/APIC). Once we do PDMGetInterrupt() we -must- deliver
2646 * the interrupt ASAP. We must not execute any guest code until we inject the interrupt which is why it is
2647 * evaluated here and not set as pending, solely based on the force-flags.
2648 */
2649 if ( !fBlockInt
2650 && !fIntShadow)
2651 {
2652 uint8_t u8Interrupt;
2653 int rc = PDMGetInterrupt(pVCpu, &u8Interrupt);
2654 if (RT_SUCCESS(rc))
2655 {
2656 Log4(("Injecting external interrupt u8Interrupt=%#x\n", u8Interrupt));
2657
2658 Event.n.u1Valid = 1;
2659 Event.n.u8Vector = u8Interrupt;
2660 Event.n.u3Type = SVM_EVENT_EXTERNAL_IRQ;
2661
2662 hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
2663 }
2664 else
2665 {
2666 /** @todo Does this actually happen? If not turn it into an assertion. */
2667 Assert(!VMCPU_FF_IS_PENDING(pVCpu, (VMCPU_FF_INTERRUPT_APIC | VMCPU_FF_INTERRUPT_PIC)));
2668 STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchGuestIrq);
2669 }
2670 }
2671 else
2672 hmR0SvmSetVirtIntrIntercept(pVmcb);
2673 }
2674}
2675
2676
2677/**
2678 * Injects any pending events into the guest if the guest is in a state to
2679 * receive them.
2680 *
2681 * @param pVCpu Pointer to the VMCPU.
2682 * @param pCtx Pointer to the guest-CPU context.
2683 */
2684static void hmR0SvmInjectPendingEvent(PVMCPU pVCpu, PCPUMCTX pCtx)
2685{
2686 Assert(!TRPMHasTrap(pVCpu));
2687 Assert(!VMMRZCallRing3IsEnabled(pVCpu));
2688
2689 bool const fIntShadow = RT_BOOL(hmR0SvmGetGuestIntrShadow(pVCpu, pCtx));
2690 bool const fBlockInt = !(pCtx->eflags.u32 & X86_EFL_IF);
2691 PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2692
2693 if (pVCpu->hm.s.Event.fPending) /* First, inject any pending HM events. */
2694 {
2695 SVMEVENT Event;
2696 Event.u = pVCpu->hm.s.Event.u64IntInfo;
2697 Assert(Event.n.u1Valid);
2698#ifdef VBOX_STRICT
2699 if (Event.n.u3Type == SVM_EVENT_EXTERNAL_IRQ)
2700 {
2701 Assert(!fBlockInt);
2702 Assert(!fIntShadow);
2703 }
2704 else if (Event.n.u3Type == SVM_EVENT_NMI)
2705 Assert(!fIntShadow);
2706#endif
2707
2708 Log4(("Injecting pending HM event.\n"));
2709 hmR0SvmInjectEventVmcb(pVCpu, pVmcb, pCtx, &Event);
2710 pVCpu->hm.s.Event.fPending = false;
2711
2712#ifdef VBOX_WITH_STATISTICS
2713 if (Event.n.u3Type == SVM_EVENT_EXTERNAL_IRQ)
2714 STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectInterrupt);
2715 else
2716 STAM_COUNTER_INC(&pVCpu->hm.s.StatInjectXcpt);
2717#endif
2718 }
2719
2720 /* Update the guest interrupt shadow in the VMCB. */
2721 pVmcb->ctrl.u64IntShadow = !!fIntShadow;
2722 NOREF(fBlockInt);
2723}
2724
2725
2726/**
2727 * Reports world-switch error and dumps some useful debug info.
2728 *
2729 * @param pVM Pointer to the VM.
2730 * @param pVCpu Pointer to the VMCPU.
2731 * @param rcVMRun The return code from VMRUN (or
2732 * VERR_SVM_INVALID_GUEST_STATE for invalid
2733 * guest-state).
2734 * @param pCtx Pointer to the guest-CPU context.
2735 */
2736static void hmR0SvmReportWorldSwitchError(PVM pVM, PVMCPU pVCpu, int rcVMRun, PCPUMCTX pCtx)
2737{
2738 NOREF(pCtx);
2739 HMSVM_ASSERT_PREEMPT_SAFE();
2740 PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
2741
2742 if (rcVMRun == VERR_SVM_INVALID_GUEST_STATE)
2743 {
2744 HMDumpRegs(pVM, pVCpu, pCtx); NOREF(pVM);
2745#ifdef VBOX_STRICT
2746 Log4(("ctrl.u64VmcbCleanBits %#RX64\n", pVmcb->ctrl.u64VmcbCleanBits));
2747 Log4(("ctrl.u16InterceptRdCRx %#x\n", pVmcb->ctrl.u16InterceptRdCRx));
2748 Log4(("ctrl.u16InterceptWrCRx %#x\n", pVmcb->ctrl.u16InterceptWrCRx));
2749 Log4(("ctrl.u16InterceptRdDRx %#x\n", pVmcb->ctrl.u16InterceptRdDRx));
2750 Log4(("ctrl.u16InterceptWrDRx %#x\n", pVmcb->ctrl.u16InterceptWrDRx));
2751 Log4(("ctrl.u32InterceptException %#x\n", pVmcb->ctrl.u32InterceptException));
2752 Log4(("ctrl.u32InterceptCtrl1 %#x\n", pVmcb->ctrl.u32InterceptCtrl1));
2753 Log4(("ctrl.u32InterceptCtrl2 %#x\n", pVmcb->ctrl.u32InterceptCtrl2));
2754 Log4(("ctrl.u64IOPMPhysAddr %#RX64\n", pVmcb->ctrl.u64IOPMPhysAddr));
2755 Log4(("ctrl.u64MSRPMPhysAddr %#RX64\n", pVmcb->ctrl.u64MSRPMPhysAddr));
2756 Log4(("ctrl.u64TSCOffset %#RX64\n", pVmcb->ctrl.u64TSCOffset));
2757
2758 Log4(("ctrl.TLBCtrl.u32ASID %#x\n", pVmcb->ctrl.TLBCtrl.n.u32ASID));
2759 Log4(("ctrl.TLBCtrl.u8TLBFlush %#x\n", pVmcb->ctrl.TLBCtrl.n.u8TLBFlush));
2760 Log4(("ctrl.TLBCtrl.u24Reserved %#x\n", pVmcb->ctrl.TLBCtrl.n.u24Reserved));
2761
2762 Log4(("ctrl.IntCtrl.u8VTPR %#x\n", pVmcb->ctrl.IntCtrl.n.u8VTPR));
2763 Log4(("ctrl.IntCtrl.u1VIrqValid %#x\n", pVmcb->ctrl.IntCtrl.n.u1VIrqValid));
2764 Log4(("ctrl.IntCtrl.u7Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u7Reserved));
2765 Log4(("ctrl.IntCtrl.u4VIrqPriority %#x\n", pVmcb->ctrl.IntCtrl.n.u4VIrqPriority));
2766 Log4(("ctrl.IntCtrl.u1IgnoreTPR %#x\n", pVmcb->ctrl.IntCtrl.n.u1IgnoreTPR));
2767 Log4(("ctrl.IntCtrl.u3Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u3Reserved));
2768 Log4(("ctrl.IntCtrl.u1VIrqMasking %#x\n", pVmcb->ctrl.IntCtrl.n.u1VIrqMasking));
2769 Log4(("ctrl.IntCtrl.u6Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u6Reserved));
2770 Log4(("ctrl.IntCtrl.u8VIrqVector %#x\n", pVmcb->ctrl.IntCtrl.n.u8VIrqVector));
2771 Log4(("ctrl.IntCtrl.u24Reserved %#x\n", pVmcb->ctrl.IntCtrl.n.u24Reserved));
2772
2773 Log4(("ctrl.u64IntShadow %#RX64\n", pVmcb->ctrl.u64IntShadow));
2774 Log4(("ctrl.u64ExitCode %#RX64\n", pVmcb->ctrl.u64ExitCode));
2775 Log4(("ctrl.u64ExitInfo1 %#RX64\n", pVmcb->ctrl.u64ExitInfo1));
2776 Log4(("ctrl.u64ExitInfo2 %#RX64\n", pVmcb->ctrl.u64ExitInfo2));
2777 Log4(("ctrl.ExitIntInfo.u8Vector %#x\n", pVmcb->ctrl.ExitIntInfo.n.u8Vector));
2778 Log4(("ctrl.ExitIntInfo.u3Type %#x\n", pVmcb->ctrl.ExitIntInfo.n.u3Type));
2779 Log4(("ctrl.ExitIntInfo.u1ErrorCodeValid %#x\n", pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid));
2780 Log4(("ctrl.ExitIntInfo.u19Reserved %#x\n", pVmcb->ctrl.ExitIntInfo.n.u19Reserved));
2781 Log4(("ctrl.ExitIntInfo.u1Valid %#x\n", pVmcb->ctrl.ExitIntInfo.n.u1Valid));
2782 Log4(("ctrl.ExitIntInfo.u32ErrorCode %#x\n", pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode));
2783 Log4(("ctrl.NestedPaging %#RX64\n", pVmcb->ctrl.NestedPaging.u));
2784 Log4(("ctrl.EventInject.u8Vector %#x\n", pVmcb->ctrl.EventInject.n.u8Vector));
2785 Log4(("ctrl.EventInject.u3Type %#x\n", pVmcb->ctrl.EventInject.n.u3Type));
2786 Log4(("ctrl.EventInject.u1ErrorCodeValid %#x\n", pVmcb->ctrl.EventInject.n.u1ErrorCodeValid));
2787 Log4(("ctrl.EventInject.u19Reserved %#x\n", pVmcb->ctrl.EventInject.n.u19Reserved));
2788 Log4(("ctrl.EventInject.u1Valid %#x\n", pVmcb->ctrl.EventInject.n.u1Valid));
2789 Log4(("ctrl.EventInject.u32ErrorCode %#x\n", pVmcb->ctrl.EventInject.n.u32ErrorCode));
2790
2791 Log4(("ctrl.u64NestedPagingCR3 %#RX64\n", pVmcb->ctrl.u64NestedPagingCR3));
2792 Log4(("ctrl.u64LBRVirt %#RX64\n", pVmcb->ctrl.u64LBRVirt));
2793
2794 Log4(("guest.CS.u16Sel %RTsel\n", pVmcb->guest.CS.u16Sel));
2795 Log4(("guest.CS.u16Attr %#x\n", pVmcb->guest.CS.u16Attr));
2796 Log4(("guest.CS.u32Limit %#RX32\n", pVmcb->guest.CS.u32Limit));
2797 Log4(("guest.CS.u64Base %#RX64\n", pVmcb->guest.CS.u64Base));
2798 Log4(("guest.DS.u16Sel %#RTsel\n", pVmcb->guest.DS.u16Sel));
2799 Log4(("guest.DS.u16Attr %#x\n", pVmcb->guest.DS.u16Attr));
2800 Log4(("guest.DS.u32Limit %#RX32\n", pVmcb->guest.DS.u32Limit));
2801 Log4(("guest.DS.u64Base %#RX64\n", pVmcb->guest.DS.u64Base));
2802 Log4(("guest.ES.u16Sel %RTsel\n", pVmcb->guest.ES.u16Sel));
2803 Log4(("guest.ES.u16Attr %#x\n", pVmcb->guest.ES.u16Attr));
2804 Log4(("guest.ES.u32Limit %#RX32\n", pVmcb->guest.ES.u32Limit));
2805 Log4(("guest.ES.u64Base %#RX64\n", pVmcb->guest.ES.u64Base));
2806 Log4(("guest.FS.u16Sel %RTsel\n", pVmcb->guest.FS.u16Sel));
2807 Log4(("guest.FS.u16Attr %#x\n", pVmcb->guest.FS.u16Attr));
2808 Log4(("guest.FS.u32Limit %#RX32\n", pVmcb->guest.FS.u32Limit));
2809 Log4(("guest.FS.u64Base %#RX64\n", pVmcb->guest.FS.u64Base));
2810 Log4(("guest.GS.u16Sel %RTsel\n", pVmcb->guest.GS.u16Sel));
2811 Log4(("guest.GS.u16Attr %#x\n", pVmcb->guest.GS.u16Attr));
2812 Log4(("guest.GS.u32Limit %#RX32\n", pVmcb->guest.GS.u32Limit));
2813 Log4(("guest.GS.u64Base %#RX64\n", pVmcb->guest.GS.u64Base));
2814
2815 Log4(("guest.GDTR.u32Limit %#RX32\n", pVmcb->guest.GDTR.u32Limit));
2816 Log4(("guest.GDTR.u64Base %#RX64\n", pVmcb->guest.GDTR.u64Base));
2817
2818 Log4(("guest.LDTR.u16Sel %RTsel\n", pVmcb->guest.LDTR.u16Sel));
2819 Log4(("guest.LDTR.u16Attr %#x\n", pVmcb->guest.LDTR.u16Attr));
2820 Log4(("guest.LDTR.u32Limit %#RX32\n", pVmcb->guest.LDTR.u32Limit));
2821 Log4(("guest.LDTR.u64Base %#RX64\n", pVmcb->guest.LDTR.u64Base));
2822
2823 Log4(("guest.IDTR.u32Limit %#RX32\n", pVmcb->guest.IDTR.u32Limit));
2824 Log4(("guest.IDTR.u64Base %#RX64\n", pVmcb->guest.IDTR.u64Base));
2825
2826 Log4(("guest.TR.u16Sel %RTsel\n", pVmcb->guest.TR.u16Sel));
2827 Log4(("guest.TR.u16Attr %#x\n", pVmcb->guest.TR.u16Attr));
2828 Log4(("guest.TR.u32Limit %#RX32\n", pVmcb->guest.TR.u32Limit));
2829 Log4(("guest.TR.u64Base %#RX64\n", pVmcb->guest.TR.u64Base));
2830
2831 Log4(("guest.u8CPL %#x\n", pVmcb->guest.u8CPL));
2832 Log4(("guest.u64CR0 %#RX64\n", pVmcb->guest.u64CR0));
2833 Log4(("guest.u64CR2 %#RX64\n", pVmcb->guest.u64CR2));
2834 Log4(("guest.u64CR3 %#RX64\n", pVmcb->guest.u64CR3));
2835 Log4(("guest.u64CR4 %#RX64\n", pVmcb->guest.u64CR4));
2836 Log4(("guest.u64DR6 %#RX64\n", pVmcb->guest.u64DR6));
2837 Log4(("guest.u64DR7 %#RX64\n", pVmcb->guest.u64DR7));
2838
2839 Log4(("guest.u64RIP %#RX64\n", pVmcb->guest.u64RIP));
2840 Log4(("guest.u64RSP %#RX64\n", pVmcb->guest.u64RSP));
2841 Log4(("guest.u64RAX %#RX64\n", pVmcb->guest.u64RAX));
2842 Log4(("guest.u64RFlags %#RX64\n", pVmcb->guest.u64RFlags));
2843
2844 Log4(("guest.u64SysEnterCS %#RX64\n", pVmcb->guest.u64SysEnterCS));
2845 Log4(("guest.u64SysEnterEIP %#RX64\n", pVmcb->guest.u64SysEnterEIP));
2846 Log4(("guest.u64SysEnterESP %#RX64\n", pVmcb->guest.u64SysEnterESP));
2847
2848 Log4(("guest.u64EFER %#RX64\n", pVmcb->guest.u64EFER));
2849 Log4(("guest.u64STAR %#RX64\n", pVmcb->guest.u64STAR));
2850 Log4(("guest.u64LSTAR %#RX64\n", pVmcb->guest.u64LSTAR));
2851 Log4(("guest.u64CSTAR %#RX64\n", pVmcb->guest.u64CSTAR));
2852 Log4(("guest.u64SFMASK %#RX64\n", pVmcb->guest.u64SFMASK));
2853 Log4(("guest.u64KernelGSBase %#RX64\n", pVmcb->guest.u64KernelGSBase));
2854 Log4(("guest.u64GPAT %#RX64\n", pVmcb->guest.u64GPAT));
2855 Log4(("guest.u64DBGCTL %#RX64\n", pVmcb->guest.u64DBGCTL));
2856 Log4(("guest.u64BR_FROM %#RX64\n", pVmcb->guest.u64BR_FROM));
2857 Log4(("guest.u64BR_TO %#RX64\n", pVmcb->guest.u64BR_TO));
2858 Log4(("guest.u64LASTEXCPFROM %#RX64\n", pVmcb->guest.u64LASTEXCPFROM));
2859 Log4(("guest.u64LASTEXCPTO %#RX64\n", pVmcb->guest.u64LASTEXCPTO));
2860#else
2861 NOREF(pVmcb);
2862#endif /* VBOX_STRICT */
2863 }
2864 else
2865 Log4(("hmR0SvmReportWorldSwitchError: rcVMRun=%d\n", rcVMRun));
2866}
2867
2868
2869/**
2870 * Check per-VM and per-VCPU force flag actions that require us to go back to
2871 * ring-3 for one reason or another.
2872 *
2873 * @returns VBox status code (information status code included).
2874 * @retval VINF_SUCCESS if we don't have any actions that require going back to
2875 * ring-3.
2876 * @retval VINF_PGM_SYNC_CR3 if we have pending PGM CR3 sync.
2877 * @retval VINF_EM_PENDING_REQUEST if we have pending requests (like hardware
2878 * interrupts)
2879 * @retval VINF_PGM_POOL_FLUSH_PENDING if PGM is doing a pool flush and requires
2880 * all EMTs to be in ring-3.
2881 * @retval VINF_EM_RAW_TO_R3 if there is pending DMA requests.
2882 * @retval VINF_EM_NO_MEMORY PGM is out of memory, we need to return
2883 * to the EM loop.
2884 *
2885 * @param pVM Pointer to the VM.
2886 * @param pVCpu Pointer to the VMCPU.
2887 * @param pCtx Pointer to the guest-CPU context.
2888 */
2889static int hmR0SvmCheckForceFlags(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
2890{
2891 Assert(VMMRZCallRing3IsEnabled(pVCpu));
2892
2893 /* On AMD-V we don't need to update CR3, PAE PDPES lazily. See hmR0SvmSaveGuestState(). */
2894 Assert(!VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_UPDATE_CR3));
2895 Assert(!VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_UPDATE_PAE_PDPES));
2896
2897 if ( VM_FF_IS_PENDING(pVM, !pVCpu->hm.s.fSingleInstruction
2898 ? VM_FF_HP_R0_PRE_HM_MASK : VM_FF_HP_R0_PRE_HM_STEP_MASK)
2899 || VMCPU_FF_IS_PENDING(pVCpu, !pVCpu->hm.s.fSingleInstruction
2900 ? VMCPU_FF_HP_R0_PRE_HM_MASK : VMCPU_FF_HP_R0_PRE_HM_STEP_MASK) )
2901 {
2902 /* Pending PGM C3 sync. */
2903 if (VMCPU_FF_IS_PENDING(pVCpu,VMCPU_FF_PGM_SYNC_CR3 | VMCPU_FF_PGM_SYNC_CR3_NON_GLOBAL))
2904 {
2905 int rc = PGMSyncCR3(pVCpu, pCtx->cr0, pCtx->cr3, pCtx->cr4, VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_PGM_SYNC_CR3));
2906 if (rc != VINF_SUCCESS)
2907 {
2908 Log4(("hmR0SvmCheckForceFlags: PGMSyncCR3 forcing us back to ring-3. rc=%d\n", rc));
2909 return rc;
2910 }
2911 }
2912
2913 /* Pending HM-to-R3 operations (critsects, timers, EMT rendezvous etc.) */
2914 /* -XXX- what was that about single stepping? */
2915 if ( VM_FF_IS_PENDING(pVM, VM_FF_HM_TO_R3_MASK)
2916 || VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_TO_R3_MASK))
2917 {
2918 STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHmToR3FF);
2919 int rc = RT_UNLIKELY(VM_FF_IS_PENDING(pVM, VM_FF_PGM_NO_MEMORY)) ? VINF_EM_NO_MEMORY : VINF_EM_RAW_TO_R3;
2920 Log4(("hmR0SvmCheckForceFlags: HM_TO_R3 forcing us back to ring-3. rc=%d\n", rc));
2921 return rc;
2922 }
2923
2924 /* Pending VM request packets, such as hardware interrupts. */
2925 if ( VM_FF_IS_PENDING(pVM, VM_FF_REQUEST)
2926 || VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_REQUEST))
2927 {
2928 Log4(("hmR0SvmCheckForceFlags: Pending VM request forcing us back to ring-3\n"));
2929 return VINF_EM_PENDING_REQUEST;
2930 }
2931
2932 /* Pending PGM pool flushes. */
2933 if (VM_FF_IS_PENDING(pVM, VM_FF_PGM_POOL_FLUSH_PENDING))
2934 {
2935 Log4(("hmR0SvmCheckForceFlags: PGM pool flush pending forcing us back to ring-3\n"));
2936 return VINF_PGM_POOL_FLUSH_PENDING;
2937 }
2938
2939 /* Pending DMA requests. */
2940 if (VM_FF_IS_PENDING(pVM, VM_FF_PDM_DMA))
2941 {
2942 Log4(("hmR0SvmCheckForceFlags: Pending DMA request forcing us back to ring-3\n"));
2943 return VINF_EM_RAW_TO_R3;
2944 }
2945 }
2946
2947 return VINF_SUCCESS;
2948}
2949
2950
2951/**
2952 * Does the preparations before executing guest code in AMD-V.
2953 *
2954 * This may cause longjmps to ring-3 and may even result in rescheduling to the
2955 * recompiler. We must be cautious what we do here regarding committing
2956 * guest-state information into the the VMCB assuming we assuredly execute the
2957 * guest in AMD-V. If we fall back to the recompiler after updating the VMCB and
2958 * clearing the common-state (TRPM/forceflags), we must undo those changes so
2959 * that the recompiler can (and should) use them when it resumes guest
2960 * execution. Otherwise such operations must be done when we can no longer
2961 * exit to ring-3.
2962 *
2963 * @returns VBox status code (informational status codes included).
2964 * @retval VINF_SUCCESS if we can proceed with running the guest.
2965 * @retval VINF_* scheduling changes, we have to go back to ring-3.
2966 *
2967 * @param pVM Pointer to the VM.
2968 * @param pVCpu Pointer to the VMCPU.
2969 * @param pCtx Pointer to the guest-CPU context.
2970 * @param pSvmTransient Pointer to the SVM transient structure.
2971 */
2972static int hmR0SvmPreRunGuest(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
2973{
2974 HMSVM_ASSERT_PREEMPT_SAFE();
2975
2976 /* Check force flag actions that might require us to go back to ring-3. */
2977 int rc = hmR0SvmCheckForceFlags(pVM, pVCpu, pCtx);
2978 if (rc != VINF_SUCCESS)
2979 return rc;
2980
2981 if (TRPMHasTrap(pVCpu))
2982 hmR0SvmTrpmTrapToPendingEvent(pVCpu);
2983 else if (!pVCpu->hm.s.Event.fPending)
2984 hmR0SvmEvaluatePendingEvent(pVCpu, pCtx);
2985
2986#ifdef HMSVM_SYNC_FULL_GUEST_STATE
2987 HMCPU_CF_SET(pVCpu, HM_CHANGED_ALL_GUEST);
2988#endif
2989
2990 /* Load the guest bits that are not shared with the host in any way since we can longjmp or get preempted. */
2991 rc = hmR0SvmLoadGuestState(pVM, pVCpu, pCtx);
2992 AssertRCReturn(rc, rc);
2993 STAM_COUNTER_INC(&pVCpu->hm.s.StatLoadFull);
2994
2995 /*
2996 * If we're not intercepting TPR changes in the guest, save the guest TPR before the world-switch
2997 * so we can update it on the way back if the guest changed the TPR.
2998 */
2999 if (pVCpu->hm.s.svm.fSyncVTpr)
3000 {
3001 if (pVM->hm.s.fTPRPatchingActive)
3002 pSvmTransient->u8GuestTpr = pCtx->msrLSTAR;
3003 else
3004 {
3005 PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
3006 pSvmTransient->u8GuestTpr = pVmcb->ctrl.IntCtrl.n.u8VTPR;
3007 }
3008 }
3009
3010 /*
3011 * No longjmps to ring-3 from this point on!!!
3012 * Asserts() will still longjmp to ring-3 (but won't return), which is intentional, better than a kernel panic.
3013 * This also disables flushing of the R0-logger instance (if any).
3014 */
3015 VMMRZCallRing3Disable(pVCpu);
3016
3017 /*
3018 * We disable interrupts so that we don't miss any interrupts that would flag preemption (IPI/timers etc.)
3019 * when thread-context hooks aren't used and we've been running with preemption disabled for a while.
3020 *
3021 * We need to check for force-flags that could've possible been altered since we last checked them (e.g.
3022 * by PDMGetInterrupt() leaving the PDM critical section, see @bugref{6398}).
3023 *
3024 * We also check a couple of other force-flags as a last opportunity to get the EMT back to ring-3 before
3025 * executing guest code.
3026 */
3027 pSvmTransient->uEflags = ASMIntDisableFlags();
3028 if ( VM_FF_IS_PENDING(pVM, VM_FF_EMT_RENDEZVOUS | VM_FF_TM_VIRTUAL_SYNC)
3029 || VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_HM_TO_R3_MASK))
3030 {
3031 ASMSetFlags(pSvmTransient->uEflags);
3032 VMMRZCallRing3Enable(pVCpu);
3033 STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHmToR3FF);
3034 return VINF_EM_RAW_TO_R3;
3035 }
3036 if (RTThreadPreemptIsPending(NIL_RTTHREAD))
3037 {
3038 ASMSetFlags(pSvmTransient->uEflags);
3039 VMMRZCallRing3Enable(pVCpu);
3040 STAM_COUNTER_INC(&pVCpu->hm.s.StatPendingHostIrq);
3041 return VINF_EM_RAW_INTERRUPT;
3042 }
3043
3044 /*
3045 * If we are injecting an NMI, we must set VMCPU_FF_BLOCK_NMIS only when we are going to execute
3046 * guest code for certain (no exits to ring-3). Otherwise, we could re-read the flag on re-entry into
3047 * AMD-V and conclude that NMI inhibition is active when we have not even delivered the NMI.
3048 *
3049 * With VT-x, this is handled by the Guest interruptibility information VMCS field which will set the
3050 * VMCS field after actually delivering the NMI which we read on VM-exit to determine the state.
3051 */
3052 if (pVCpu->hm.s.Event.fPending)
3053 {
3054 SVMEVENT Event;
3055 Event.u = pVCpu->hm.s.Event.u64IntInfo;
3056 if ( Event.n.u1Valid
3057 && Event.n.u3Type == SVM_EVENT_NMI
3058 && Event.n.u8Vector == X86_XCPT_NMI
3059 && !VMCPU_FF_IS_PENDING(pVCpu, VMCPU_FF_BLOCK_NMIS))
3060 {
3061 VMCPU_FF_SET(pVCpu, VMCPU_FF_BLOCK_NMIS);
3062 }
3063 }
3064
3065 return VINF_SUCCESS;
3066}
3067
3068
3069/**
3070 * Prepares to run guest code in AMD-V and we've committed to doing so. This
3071 * means there is no backing out to ring-3 or anywhere else at this
3072 * point.
3073 *
3074 * @param pVM Pointer to the VM.
3075 * @param pVCpu Pointer to the VMCPU.
3076 * @param pCtx Pointer to the guest-CPU context.
3077 * @param pSvmTransient Pointer to the SVM transient structure.
3078 *
3079 * @remarks Called with preemption disabled.
3080 * @remarks No-long-jump zone!!!
3081 */
3082static void hmR0SvmPreRunGuestCommitted(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
3083{
3084 Assert(!VMMRZCallRing3IsEnabled(pVCpu));
3085 Assert(VMMR0IsLogFlushDisabled(pVCpu));
3086 Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
3087
3088 VMCPU_ASSERT_STATE(pVCpu, VMCPUSTATE_STARTED_HM);
3089 VMCPU_SET_STATE(pVCpu, VMCPUSTATE_STARTED_EXEC); /* Indicate the start of guest execution. */
3090
3091 hmR0SvmInjectPendingEvent(pVCpu, pCtx);
3092
3093 if ( pVCpu->hm.s.fPreloadGuestFpu
3094 && !CPUMIsGuestFPUStateActive(pVCpu))
3095 {
3096 CPUMR0LoadGuestFPU(pVM, pVCpu, pCtx);
3097 HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR0);
3098 }
3099
3100 /* Load the state shared between host and guest (FPU, debug). */
3101 PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
3102 if (HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_HOST_GUEST_SHARED_STATE))
3103 hmR0SvmLoadSharedState(pVCpu, pVmcb, pCtx);
3104 HMCPU_CF_CLEAR(pVCpu, HM_CHANGED_HOST_CONTEXT); /* Preemption might set this, nothing to do on AMD-V. */
3105 AssertMsg(!HMCPU_CF_VALUE(pVCpu), ("fContextUseFlags=%#RX32\n", HMCPU_CF_VALUE(pVCpu)));
3106
3107 /* Setup TSC offsetting. */
3108 RTCPUID idCurrentCpu = HMR0GetCurrentCpu()->idCpu;
3109 if ( pSvmTransient->fUpdateTscOffsetting
3110 || idCurrentCpu != pVCpu->hm.s.idLastCpu)
3111 {
3112 hmR0SvmUpdateTscOffsetting(pVM, pVCpu);
3113 pSvmTransient->fUpdateTscOffsetting = false;
3114 }
3115
3116 /* If we've migrating CPUs, mark the VMCB Clean bits as dirty. */
3117 if (idCurrentCpu != pVCpu->hm.s.idLastCpu)
3118 pVmcb->ctrl.u64VmcbCleanBits = 0;
3119
3120 /* Store status of the shared guest-host state at the time of VMRUN. */
3121#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS) && !defined(VBOX_WITH_HYBRID_32BIT_KERNEL)
3122 if (CPUMIsGuestInLongModeEx(pCtx))
3123 {
3124 pSvmTransient->fWasGuestDebugStateActive = CPUMIsGuestDebugStateActivePending(pVCpu);
3125 pSvmTransient->fWasHyperDebugStateActive = CPUMIsHyperDebugStateActivePending(pVCpu);
3126 }
3127 else
3128#endif
3129 {
3130 pSvmTransient->fWasGuestDebugStateActive = CPUMIsGuestDebugStateActive(pVCpu);
3131 pSvmTransient->fWasHyperDebugStateActive = CPUMIsHyperDebugStateActive(pVCpu);
3132 }
3133 pSvmTransient->fWasGuestFPUStateActive = CPUMIsGuestFPUStateActive(pVCpu);
3134
3135 /* Flush the appropriate tagged-TLB entries. */
3136 ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, true); /* Used for TLB-shootdowns, set this across the world switch. */
3137 hmR0SvmFlushTaggedTlb(pVCpu);
3138 Assert(HMR0GetCurrentCpu()->idCpu == pVCpu->hm.s.idLastCpu);
3139
3140 STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatEntry, &pVCpu->hm.s.StatInGC, x);
3141
3142 TMNotifyStartOfExecution(pVCpu); /* Finally, notify TM to resume its clocks as we're about
3143 to start executing. */
3144
3145 /*
3146 * Save the current Host TSC_AUX and write the guest TSC_AUX to the host, so that
3147 * RDTSCPs (that don't cause exits) reads the guest MSR. See @bugref{3324}.
3148 *
3149 * This should be done -after- any RDTSCPs for obtaining the host timestamp (TM, STAM etc).
3150 */
3151 if ( (pVM->hm.s.cpuid.u32AMDFeatureEDX & X86_CPUID_EXT_FEATURE_EDX_RDTSCP)
3152 && !(pVmcb->ctrl.u32InterceptCtrl2 & SVM_CTRL2_INTERCEPT_RDTSCP))
3153 {
3154 hmR0SvmSetMsrPermission(pVCpu, MSR_K8_TSC_AUX, SVMMSREXIT_PASSTHRU_READ, SVMMSREXIT_PASSTHRU_WRITE);
3155 pVCpu->hm.s.u64HostTscAux = ASMRdMsr(MSR_K8_TSC_AUX);
3156 uint64_t u64GuestTscAux = CPUMR0GetGuestTscAux(pVCpu);
3157 if (u64GuestTscAux != pVCpu->hm.s.u64HostTscAux)
3158 ASMWrMsr(MSR_K8_TSC_AUX, u64GuestTscAux);
3159 pSvmTransient->fRestoreTscAuxMsr = true;
3160 }
3161 else
3162 {
3163 hmR0SvmSetMsrPermission(pVCpu, MSR_K8_TSC_AUX, SVMMSREXIT_INTERCEPT_READ, SVMMSREXIT_INTERCEPT_WRITE);
3164 pSvmTransient->fRestoreTscAuxMsr = false;
3165 }
3166
3167 /* If VMCB Clean bits isn't supported by the CPU, simply mark all state-bits as dirty, indicating (re)load-from-VMCB. */
3168 if (!(pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_VMCB_CLEAN))
3169 pVmcb->ctrl.u64VmcbCleanBits = 0;
3170}
3171
3172
3173/**
3174 * Wrapper for running the guest code in AMD-V.
3175 *
3176 * @returns VBox strict status code.
3177 * @param pVM Pointer to the VM.
3178 * @param pVCpu Pointer to the VMCPU.
3179 * @param pCtx Pointer to the guest-CPU context.
3180 *
3181 * @remarks No-long-jump zone!!!
3182 */
3183DECLINLINE(int) hmR0SvmRunGuest(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3184{
3185 /*
3186 * 64-bit Windows uses XMM registers in the kernel as the Microsoft compiler expresses floating-point operations
3187 * using SSE instructions. Some XMM registers (XMM6-XMM15) are callee-saved and thus the need for this XMM wrapper.
3188 * Refer MSDN docs. "Configuring Programs for 64-bit / x64 Software Conventions / Register Usage" for details.
3189 */
3190#ifdef VBOX_WITH_KERNEL_USING_XMM
3191 return HMR0SVMRunWrapXMM(pVCpu->hm.s.svm.HCPhysVmcbHost, pVCpu->hm.s.svm.HCPhysVmcb, pCtx, pVM, pVCpu,
3192 pVCpu->hm.s.svm.pfnVMRun);
3193#else
3194 return pVCpu->hm.s.svm.pfnVMRun(pVCpu->hm.s.svm.HCPhysVmcbHost, pVCpu->hm.s.svm.HCPhysVmcb, pCtx, pVM, pVCpu);
3195#endif
3196}
3197
3198
3199/**
3200 * Performs some essential restoration of state after running guest code in
3201 * AMD-V.
3202 *
3203 * @param pVM Pointer to the VM.
3204 * @param pVCpu Pointer to the VMCPU.
3205 * @param pMixedCtx Pointer to the guest-CPU context. The data maybe
3206 * out-of-sync. Make sure to update the required fields
3207 * before using them.
3208 * @param pSvmTransient Pointer to the SVM transient structure.
3209 * @param rcVMRun Return code of VMRUN.
3210 *
3211 * @remarks Called with interrupts disabled.
3212 * @remarks No-long-jump zone!!! This function will however re-enable longjmps
3213 * unconditionally when it is safe to do so.
3214 */
3215static void hmR0SvmPostRunGuest(PVM pVM, PVMCPU pVCpu, PCPUMCTX pMixedCtx, PSVMTRANSIENT pSvmTransient, int rcVMRun)
3216{
3217 Assert(!VMMRZCallRing3IsEnabled(pVCpu));
3218
3219 ASMAtomicWriteBool(&pVCpu->hm.s.fCheckedTLBFlush, false); /* See HMInvalidatePageOnAllVCpus(): used for TLB-shootdowns. */
3220 ASMAtomicIncU32(&pVCpu->hm.s.cWorldSwitchExits); /* Initialized in vmR3CreateUVM(): used for TLB-shootdowns. */
3221
3222 PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
3223 pVmcb->ctrl.u64VmcbCleanBits = HMSVM_VMCB_CLEAN_ALL; /* Mark the VMCB-state cache as unmodified by VMM. */
3224
3225 if (pSvmTransient->fRestoreTscAuxMsr)
3226 {
3227 uint64_t u64GuestTscAuxMsr = ASMRdMsr(MSR_K8_TSC_AUX);
3228 CPUMR0SetGuestTscAux(pVCpu, u64GuestTscAuxMsr);
3229 if (u64GuestTscAuxMsr != pVCpu->hm.s.u64HostTscAux)
3230 ASMWrMsr(MSR_K8_TSC_AUX, pVCpu->hm.s.u64HostTscAux);
3231 }
3232
3233 if (!(pVmcb->ctrl.u32InterceptCtrl1 & SVM_CTRL1_INTERCEPT_RDTSC))
3234 TMCpuTickSetLastSeen(pVCpu, ASMReadTSC() + pVmcb->ctrl.u64TSCOffset);
3235
3236 STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatInGC, &pVCpu->hm.s.StatExit1, x);
3237 TMNotifyEndOfExecution(pVCpu); /* Notify TM that the guest is no longer running. */
3238 VMCPU_SET_STATE(pVCpu, VMCPUSTATE_STARTED_HM);
3239
3240 Assert(!(ASMGetFlags() & X86_EFL_IF));
3241 ASMSetFlags(pSvmTransient->uEflags); /* Enable interrupts. */
3242 VMMRZCallRing3Enable(pVCpu); /* It is now safe to do longjmps to ring-3!!! */
3243
3244 /* If VMRUN failed, we can bail out early. This does -not- cover SVM_EXIT_INVALID. */
3245 if (RT_UNLIKELY(rcVMRun != VINF_SUCCESS))
3246 {
3247 Log4(("VMRUN failure: rcVMRun=%Rrc\n", rcVMRun));
3248 return;
3249 }
3250
3251 pSvmTransient->u64ExitCode = pVmcb->ctrl.u64ExitCode; /* Save the #VMEXIT reason. */
3252 HMCPU_EXIT_HISTORY_ADD(pVCpu, pVmcb->ctrl.u64ExitCode); /* Update the #VMEXIT history array. */
3253 pSvmTransient->fVectoringDoublePF = false; /* Vectoring double page-fault needs to be determined later. */
3254 pSvmTransient->fVectoringPF = false; /* Vectoring page-fault needs to be determined later. */
3255
3256 hmR0SvmSaveGuestState(pVCpu, pMixedCtx); /* Save the guest state from the VMCB to the guest-CPU context. */
3257
3258 if (RT_LIKELY(pSvmTransient->u64ExitCode != (uint64_t)SVM_EXIT_INVALID))
3259 {
3260 if (pVCpu->hm.s.svm.fSyncVTpr)
3261 {
3262 /* TPR patching (for 32-bit guests) uses LSTAR MSR for holding the TPR value, otherwise uses the VTPR. */
3263 if ( pVM->hm.s.fTPRPatchingActive
3264 && (pMixedCtx->msrLSTAR & 0xff) != pSvmTransient->u8GuestTpr)
3265 {
3266 int rc = PDMApicSetTPR(pVCpu, pMixedCtx->msrLSTAR & 0xff);
3267 AssertRC(rc);
3268 HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
3269 }
3270 else if (pSvmTransient->u8GuestTpr != pVmcb->ctrl.IntCtrl.n.u8VTPR)
3271 {
3272 int rc = PDMApicSetTPR(pVCpu, pVmcb->ctrl.IntCtrl.n.u8VTPR << 4);
3273 AssertRC(rc);
3274 HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
3275 }
3276 }
3277 }
3278}
3279
3280
3281/**
3282 * Runs the guest code using AMD-V.
3283 *
3284 * @returns VBox status code.
3285 * @param pVM Pointer to the VM.
3286 * @param pVCpu Pointer to the VMCPU.
3287 */
3288static int hmR0SvmRunGuestCodeNormal(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3289{
3290 SVMTRANSIENT SvmTransient;
3291 SvmTransient.fUpdateTscOffsetting = true;
3292 uint32_t cLoops = 0;
3293 int rc = VERR_INTERNAL_ERROR_5;
3294
3295 for (;; cLoops++)
3296 {
3297 Assert(!HMR0SuspendPending());
3298 HMSVM_ASSERT_CPU_SAFE();
3299
3300 /* Preparatory work for running guest code, this may force us to return
3301 to ring-3. This bugger disables interrupts on VINF_SUCCESS! */
3302 STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatEntry, x);
3303 rc = hmR0SvmPreRunGuest(pVM, pVCpu, pCtx, &SvmTransient);
3304 if (rc != VINF_SUCCESS)
3305 break;
3306
3307 /*
3308 * No longjmps to ring-3 from this point on!!!
3309 * Asserts() will still longjmp to ring-3 (but won't return), which is intentional, better than a kernel panic.
3310 * This also disables flushing of the R0-logger instance (if any).
3311 */
3312 hmR0SvmPreRunGuestCommitted(pVM, pVCpu, pCtx, &SvmTransient);
3313 rc = hmR0SvmRunGuest(pVM, pVCpu, pCtx);
3314
3315 /* Restore any residual host-state and save any bits shared between host
3316 and guest into the guest-CPU state. Re-enables interrupts! */
3317 hmR0SvmPostRunGuest(pVM, pVCpu, pCtx, &SvmTransient, rc);
3318
3319 if (RT_UNLIKELY( rc != VINF_SUCCESS /* Check for VMRUN errors. */
3320 || SvmTransient.u64ExitCode == (uint64_t)SVM_EXIT_INVALID)) /* Check for invalid guest-state errors. */
3321 {
3322 if (rc == VINF_SUCCESS)
3323 rc = VERR_SVM_INVALID_GUEST_STATE;
3324 STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit1, x);
3325 hmR0SvmReportWorldSwitchError(pVM, pVCpu, rc, pCtx);
3326 break;
3327 }
3328
3329 /* Handle the #VMEXIT. */
3330 HMSVM_EXITCODE_STAM_COUNTER_INC(SvmTransient.u64ExitCode);
3331 STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatExit1, &pVCpu->hm.s.StatExit2, x);
3332 VBOXVMM_R0_HMSVM_VMEXIT(pVCpu, pCtx, SvmTransient.u64ExitCode, (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb);
3333 rc = hmR0SvmHandleExit(pVCpu, pCtx, &SvmTransient);
3334 STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit2, x);
3335 if (rc != VINF_SUCCESS)
3336 break;
3337 if (cLoops > pVM->hm.s.cMaxResumeLoops)
3338 {
3339 STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchMaxResumeLoops);
3340 rc = VINF_EM_RAW_INTERRUPT;
3341 break;
3342 }
3343 }
3344
3345 STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatEntry, x);
3346 return rc;
3347}
3348
3349
3350/**
3351 * Runs the guest code using AMD-V in single step mode.
3352 *
3353 * @returns VBox status code.
3354 * @param pVM Pointer to the VM.
3355 * @param pVCpu Pointer to the VMCPU.
3356 * @param pCtx Pointer to the guest-CPU context.
3357 */
3358static int hmR0SvmRunGuestCodeStep(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3359{
3360 SVMTRANSIENT SvmTransient;
3361 SvmTransient.fUpdateTscOffsetting = true;
3362 uint32_t cLoops = 0;
3363 int rc = VERR_INTERNAL_ERROR_5;
3364 uint16_t uCsStart = pCtx->cs.Sel;
3365 uint64_t uRipStart = pCtx->rip;
3366
3367 for (;; cLoops++)
3368 {
3369 Assert(!HMR0SuspendPending());
3370 AssertMsg(pVCpu->hm.s.idEnteredCpu == RTMpCpuId(),
3371 ("Illegal migration! Entered on CPU %u Current %u cLoops=%u\n", (unsigned)pVCpu->hm.s.idEnteredCpu,
3372 (unsigned)RTMpCpuId(), cLoops));
3373
3374 /* Preparatory work for running guest code, this may force us to return
3375 to ring-3. This bugger disables interrupts on VINF_SUCCESS! */
3376 STAM_PROFILE_ADV_START(&pVCpu->hm.s.StatEntry, x);
3377 rc = hmR0SvmPreRunGuest(pVM, pVCpu, pCtx, &SvmTransient);
3378 if (rc != VINF_SUCCESS)
3379 break;
3380
3381 /*
3382 * No longjmps to ring-3 from this point on!!!
3383 * Asserts() will still longjmp to ring-3 (but won't return), which is intentional, better than a kernel panic.
3384 * This also disables flushing of the R0-logger instance (if any).
3385 */
3386 VMMRZCallRing3Disable(pVCpu);
3387 VMMRZCallRing3RemoveNotification(pVCpu);
3388 hmR0SvmPreRunGuestCommitted(pVM, pVCpu, pCtx, &SvmTransient);
3389
3390 rc = hmR0SvmRunGuest(pVM, pVCpu, pCtx);
3391
3392 /*
3393 * Restore any residual host-state and save any bits shared between host and guest into the guest-CPU state.
3394 * This will also re-enable longjmps to ring-3 when it has reached a safe point!!!
3395 */
3396 hmR0SvmPostRunGuest(pVM, pVCpu, pCtx, &SvmTransient, rc);
3397 if (RT_UNLIKELY( rc != VINF_SUCCESS /* Check for VMRUN errors. */
3398 || SvmTransient.u64ExitCode == (uint64_t)SVM_EXIT_INVALID)) /* Check for invalid guest-state errors. */
3399 {
3400 if (rc == VINF_SUCCESS)
3401 rc = VERR_SVM_INVALID_GUEST_STATE;
3402 STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit1, x);
3403 hmR0SvmReportWorldSwitchError(pVM, pVCpu, rc, pCtx);
3404 return rc;
3405 }
3406
3407 /* Handle the #VMEXIT. */
3408 HMSVM_EXITCODE_STAM_COUNTER_INC(SvmTransient.u64ExitCode);
3409 STAM_PROFILE_ADV_STOP_START(&pVCpu->hm.s.StatExit1, &pVCpu->hm.s.StatExit2, x);
3410 VBOXVMM_R0_HMSVM_VMEXIT(pVCpu, pCtx, SvmTransient.u64ExitCode, (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb);
3411 rc = hmR0SvmHandleExit(pVCpu, pCtx, &SvmTransient);
3412 STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatExit2, x);
3413 if (rc != VINF_SUCCESS)
3414 break;
3415 if (cLoops > pVM->hm.s.cMaxResumeLoops)
3416 {
3417 STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchMaxResumeLoops);
3418 rc = VINF_EM_RAW_INTERRUPT;
3419 break;
3420 }
3421
3422 /*
3423 * Did the RIP change, if so, consider it a single step.
3424 * Otherwise, make sure one of the TFs gets set.
3425 */
3426 if ( pCtx->rip != uRipStart
3427 || pCtx->cs.Sel != uCsStart)
3428 {
3429 rc = VINF_EM_DBG_STEPPED;
3430 break;
3431 }
3432 pVCpu->hm.s.fContextUseFlags |= HM_CHANGED_GUEST_DEBUG;
3433 }
3434
3435 /*
3436 * Clear the X86_EFL_TF if necessary.
3437 */
3438 if (pVCpu->hm.s.fClearTrapFlag)
3439 {
3440 pVCpu->hm.s.fClearTrapFlag = false;
3441 pCtx->eflags.Bits.u1TF = 0;
3442 }
3443
3444 STAM_PROFILE_ADV_STOP(&pVCpu->hm.s.StatEntry, x);
3445 return rc;
3446}
3447
3448
3449/**
3450 * Runs the guest code using AMD-V.
3451 *
3452 * @returns VBox status code.
3453 * @param pVM Pointer to the VM.
3454 * @param pVCpu Pointer to the VMCPU.
3455 * @param pCtx Pointer to the guest-CPU context.
3456 */
3457VMMR0DECL(int) SVMR0RunGuestCode(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3458{
3459 Assert(VMMRZCallRing3IsEnabled(pVCpu));
3460 HMSVM_ASSERT_PREEMPT_SAFE();
3461 VMMRZCallRing3SetNotification(pVCpu, hmR0SvmCallRing3Callback, pCtx);
3462
3463 int rc;
3464 if (!pVCpu->hm.s.fSingleInstruction && !DBGFIsStepping(pVCpu))
3465 rc = hmR0SvmRunGuestCodeNormal(pVM, pVCpu, pCtx);
3466 else
3467 rc = hmR0SvmRunGuestCodeStep(pVM, pVCpu, pCtx);
3468
3469 if (rc == VERR_EM_INTERPRETER)
3470 rc = VINF_EM_RAW_EMULATE_INSTR;
3471 else if (rc == VINF_EM_RESET)
3472 rc = VINF_EM_TRIPLE_FAULT;
3473
3474 /* Prepare to return to ring-3. This will remove longjmp notifications. */
3475 hmR0SvmExitToRing3(pVM, pVCpu, pCtx, rc);
3476 Assert(!VMMRZCallRing3IsNotificationSet(pVCpu));
3477 return rc;
3478}
3479
3480
3481/**
3482 * Handles a #VMEXIT (for all EXITCODE values except SVM_EXIT_INVALID).
3483 *
3484 * @returns VBox status code (informational status codes included).
3485 * @param pVCpu Pointer to the VMCPU.
3486 * @param pCtx Pointer to the guest-CPU context.
3487 * @param pSvmTransient Pointer to the SVM transient structure.
3488 */
3489DECLINLINE(int) hmR0SvmHandleExit(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
3490{
3491 Assert(pSvmTransient->u64ExitCode != (uint64_t)SVM_EXIT_INVALID);
3492 Assert(pSvmTransient->u64ExitCode <= SVM_EXIT_MAX);
3493
3494 /*
3495 * The ordering of the case labels is based on most-frequently-occurring #VMEXITs for most guests under
3496 * normal workloads (for some definition of "normal").
3497 */
3498 uint32_t u32ExitCode = pSvmTransient->u64ExitCode;
3499 switch (pSvmTransient->u64ExitCode)
3500 {
3501 case SVM_EXIT_NPF:
3502 return hmR0SvmExitNestedPF(pVCpu, pCtx, pSvmTransient);
3503
3504 case SVM_EXIT_IOIO:
3505 return hmR0SvmExitIOInstr(pVCpu, pCtx, pSvmTransient);
3506
3507 case SVM_EXIT_RDTSC:
3508 return hmR0SvmExitRdtsc(pVCpu, pCtx, pSvmTransient);
3509
3510 case SVM_EXIT_RDTSCP:
3511 return hmR0SvmExitRdtscp(pVCpu, pCtx, pSvmTransient);
3512
3513 case SVM_EXIT_CPUID:
3514 return hmR0SvmExitCpuid(pVCpu, pCtx, pSvmTransient);
3515
3516 case SVM_EXIT_EXCEPTION_E: /* X86_XCPT_PF */
3517 return hmR0SvmExitXcptPF(pVCpu, pCtx, pSvmTransient);
3518
3519 case SVM_EXIT_EXCEPTION_7: /* X86_XCPT_NM */
3520 return hmR0SvmExitXcptNM(pVCpu, pCtx, pSvmTransient);
3521
3522 case SVM_EXIT_EXCEPTION_6: /* X86_XCPT_UD */
3523 return hmR0SvmExitXcptUD(pVCpu, pCtx, pSvmTransient);
3524
3525 case SVM_EXIT_EXCEPTION_10: /* X86_XCPT_MF */
3526 return hmR0SvmExitXcptMF(pVCpu, pCtx, pSvmTransient);
3527
3528 case SVM_EXIT_EXCEPTION_1: /* X86_XCPT_DB */
3529 return hmR0SvmExitXcptDB(pVCpu, pCtx, pSvmTransient);
3530
3531 case SVM_EXIT_MONITOR:
3532 return hmR0SvmExitMonitor(pVCpu, pCtx, pSvmTransient);
3533
3534 case SVM_EXIT_MWAIT:
3535 return hmR0SvmExitMwait(pVCpu, pCtx, pSvmTransient);
3536
3537 case SVM_EXIT_HLT:
3538 return hmR0SvmExitHlt(pVCpu, pCtx, pSvmTransient);
3539
3540 case SVM_EXIT_READ_CR0:
3541 case SVM_EXIT_READ_CR3:
3542 case SVM_EXIT_READ_CR4:
3543 return hmR0SvmExitReadCRx(pVCpu, pCtx, pSvmTransient);
3544
3545 case SVM_EXIT_WRITE_CR0:
3546 case SVM_EXIT_WRITE_CR3:
3547 case SVM_EXIT_WRITE_CR4:
3548 case SVM_EXIT_WRITE_CR8:
3549 return hmR0SvmExitWriteCRx(pVCpu, pCtx, pSvmTransient);
3550
3551 case SVM_EXIT_VMMCALL:
3552 return hmR0SvmExitVmmCall(pVCpu, pCtx, pSvmTransient);
3553
3554 case SVM_EXIT_VINTR:
3555 return hmR0SvmExitVIntr(pVCpu, pCtx, pSvmTransient);
3556
3557 case SVM_EXIT_INTR:
3558 case SVM_EXIT_FERR_FREEZE:
3559 case SVM_EXIT_NMI:
3560 return hmR0SvmExitIntr(pVCpu, pCtx, pSvmTransient);
3561
3562 case SVM_EXIT_MSR:
3563 return hmR0SvmExitMsr(pVCpu, pCtx, pSvmTransient);
3564
3565 case SVM_EXIT_INVLPG:
3566 return hmR0SvmExitInvlpg(pVCpu, pCtx, pSvmTransient);
3567
3568 case SVM_EXIT_WBINVD:
3569 return hmR0SvmExitWbinvd(pVCpu, pCtx, pSvmTransient);
3570
3571 case SVM_EXIT_INVD:
3572 return hmR0SvmExitInvd(pVCpu, pCtx, pSvmTransient);
3573
3574 case SVM_EXIT_RDPMC:
3575 return hmR0SvmExitRdpmc(pVCpu, pCtx, pSvmTransient);
3576
3577 default:
3578 {
3579 switch (pSvmTransient->u64ExitCode)
3580 {
3581 case SVM_EXIT_READ_DR0: case SVM_EXIT_READ_DR1: case SVM_EXIT_READ_DR2: case SVM_EXIT_READ_DR3:
3582 case SVM_EXIT_READ_DR6: case SVM_EXIT_READ_DR7: case SVM_EXIT_READ_DR8: case SVM_EXIT_READ_DR9:
3583 case SVM_EXIT_READ_DR10: case SVM_EXIT_READ_DR11: case SVM_EXIT_READ_DR12: case SVM_EXIT_READ_DR13:
3584 case SVM_EXIT_READ_DR14: case SVM_EXIT_READ_DR15:
3585 return hmR0SvmExitReadDRx(pVCpu, pCtx, pSvmTransient);
3586
3587 case SVM_EXIT_WRITE_DR0: case SVM_EXIT_WRITE_DR1: case SVM_EXIT_WRITE_DR2: case SVM_EXIT_WRITE_DR3:
3588 case SVM_EXIT_WRITE_DR6: case SVM_EXIT_WRITE_DR7: case SVM_EXIT_WRITE_DR8: case SVM_EXIT_WRITE_DR9:
3589 case SVM_EXIT_WRITE_DR10: case SVM_EXIT_WRITE_DR11: case SVM_EXIT_WRITE_DR12: case SVM_EXIT_WRITE_DR13:
3590 case SVM_EXIT_WRITE_DR14: case SVM_EXIT_WRITE_DR15:
3591 return hmR0SvmExitWriteDRx(pVCpu, pCtx, pSvmTransient);
3592
3593 case SVM_EXIT_XSETBV:
3594 return hmR0SvmExitXsetbv(pVCpu, pCtx, pSvmTransient);
3595
3596 case SVM_EXIT_TASK_SWITCH:
3597 return hmR0SvmExitTaskSwitch(pVCpu, pCtx, pSvmTransient);
3598
3599 case SVM_EXIT_IRET:
3600 return hmR0SvmExitIret(pVCpu, pCtx, pSvmTransient);
3601
3602 case SVM_EXIT_SHUTDOWN:
3603 return hmR0SvmExitShutdown(pVCpu, pCtx, pSvmTransient);
3604
3605 case SVM_EXIT_SMI:
3606 case SVM_EXIT_INIT:
3607 {
3608 /*
3609 * We don't intercept NMIs. As for INIT signals, it really shouldn't ever happen here. If it ever does,
3610 * we want to know about it so log the exit code and bail.
3611 */
3612 AssertMsgFailed(("hmR0SvmHandleExit: Unexpected exit %#RX32\n", (uint32_t)pSvmTransient->u64ExitCode));
3613 pVCpu->hm.s.u32HMError = (uint32_t)pSvmTransient->u64ExitCode;
3614 return VERR_SVM_UNEXPECTED_EXIT;
3615 }
3616
3617 case SVM_EXIT_INVLPGA:
3618 case SVM_EXIT_RSM:
3619 case SVM_EXIT_VMRUN:
3620 case SVM_EXIT_VMLOAD:
3621 case SVM_EXIT_VMSAVE:
3622 case SVM_EXIT_STGI:
3623 case SVM_EXIT_CLGI:
3624 case SVM_EXIT_SKINIT:
3625 return hmR0SvmExitSetPendingXcptUD(pVCpu, pCtx, pSvmTransient);
3626
3627#ifdef HMSVM_ALWAYS_TRAP_ALL_XCPTS
3628 case SVM_EXIT_EXCEPTION_0: /* X86_XCPT_DE */
3629 /* SVM_EXIT_EXCEPTION_1: */ /* X86_XCPT_DB - Handled above. */
3630 case SVM_EXIT_EXCEPTION_2: /* X86_XCPT_NMI */
3631 case SVM_EXIT_EXCEPTION_3: /* X86_XCPT_BP */
3632 case SVM_EXIT_EXCEPTION_4: /* X86_XCPT_OF */
3633 case SVM_EXIT_EXCEPTION_5: /* X86_XCPT_BR */
3634 /* case SVM_EXIT_EXCEPTION_6: */ /* X86_XCPT_UD - Handled above. */
3635 /* SVM_EXIT_EXCEPTION_7: */ /* X86_XCPT_NM - Handled above. */
3636 case SVM_EXIT_EXCEPTION_8: /* X86_XCPT_DF */
3637 case SVM_EXIT_EXCEPTION_9: /* X86_XCPT_CO_SEG_OVERRUN */
3638 case SVM_EXIT_EXCEPTION_A: /* X86_XCPT_TS */
3639 case SVM_EXIT_EXCEPTION_B: /* X86_XCPT_NP */
3640 case SVM_EXIT_EXCEPTION_C: /* X86_XCPT_SS */
3641 case SVM_EXIT_EXCEPTION_D: /* X86_XCPT_GP */
3642 /* SVM_EXIT_EXCEPTION_E: */ /* X86_XCPT_PF - Handled above. */
3643 /* SVM_EXIT_EXCEPTION_10: */ /* X86_XCPT_MF - Handled above. */
3644 case SVM_EXIT_EXCEPTION_11: /* X86_XCPT_AC */
3645 case SVM_EXIT_EXCEPTION_12: /* X86_XCPT_MC */
3646 case SVM_EXIT_EXCEPTION_13: /* X86_XCPT_XF */
3647 case SVM_EXIT_EXCEPTION_F: /* Reserved */
3648 case SVM_EXIT_EXCEPTION_14: case SVM_EXIT_EXCEPTION_15: case SVM_EXIT_EXCEPTION_16:
3649 case SVM_EXIT_EXCEPTION_17: case SVM_EXIT_EXCEPTION_18: case SVM_EXIT_EXCEPTION_19:
3650 case SVM_EXIT_EXCEPTION_1A: case SVM_EXIT_EXCEPTION_1B: case SVM_EXIT_EXCEPTION_1C:
3651 case SVM_EXIT_EXCEPTION_1D: case SVM_EXIT_EXCEPTION_1E: case SVM_EXIT_EXCEPTION_1F:
3652 {
3653 PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
3654 SVMEVENT Event;
3655 Event.u = 0;
3656 Event.n.u1Valid = 1;
3657 Event.n.u3Type = SVM_EVENT_EXCEPTION;
3658 Event.n.u8Vector = pSvmTransient->u64ExitCode - SVM_EXIT_EXCEPTION_0;
3659
3660 switch (Event.n.u8Vector)
3661 {
3662 case X86_XCPT_DE:
3663 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestDE);
3664 break;
3665
3666 case X86_XCPT_BP:
3667 /** Saves the wrong EIP on the stack (pointing to the int3) instead of the
3668 * next instruction. */
3669 /** @todo Investigate this later. */
3670 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestBP);
3671 break;
3672
3673 case X86_XCPT_NP:
3674 Event.n.u1ErrorCodeValid = 1;
3675 Event.n.u32ErrorCode = pVmcb->ctrl.u64ExitInfo1;
3676 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestNP);
3677 break;
3678
3679 case X86_XCPT_SS:
3680 Event.n.u1ErrorCodeValid = 1;
3681 Event.n.u32ErrorCode = pVmcb->ctrl.u64ExitInfo1;
3682 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestSS);
3683 break;
3684
3685 case X86_XCPT_GP:
3686 Event.n.u1ErrorCodeValid = 1;
3687 Event.n.u32ErrorCode = pVmcb->ctrl.u64ExitInfo1;
3688 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestGP);
3689 break;
3690
3691 default:
3692 AssertMsgFailed(("hmR0SvmHandleExit: Unexpected exit caused by exception %#x\n", Event.n.u8Vector));
3693 pVCpu->hm.s.u32HMError = Event.n.u8Vector;
3694 return VERR_SVM_UNEXPECTED_XCPT_EXIT;
3695 }
3696
3697 Log4(("#Xcpt: Vector=%#x at CS:RIP=%04x:%RGv\n", Event.n.u8Vector, pCtx->cs.Sel, (RTGCPTR)pCtx->rip));
3698 hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3699 return VINF_SUCCESS;
3700 }
3701#endif /* HMSVM_ALWAYS_TRAP_ALL_XCPTS */
3702
3703 default:
3704 {
3705 AssertMsgFailed(("hmR0SvmHandleExit: Unknown exit code %#x\n", u32ExitCode));
3706 pVCpu->hm.s.u32HMError = u32ExitCode;
3707 return VERR_SVM_UNKNOWN_EXIT;
3708 }
3709 }
3710 }
3711 }
3712 return VERR_INTERNAL_ERROR_5; /* Should never happen. */
3713}
3714
3715
3716#ifdef DEBUG
3717/* Is there some generic IPRT define for this that are not in Runtime/internal/\* ?? */
3718# define HMSVM_ASSERT_PREEMPT_CPUID_VAR() \
3719 RTCPUID const idAssertCpu = RTThreadPreemptIsEnabled(NIL_RTTHREAD) ? NIL_RTCPUID : RTMpCpuId()
3720
3721# define HMSVM_ASSERT_PREEMPT_CPUID() \
3722 do \
3723 { \
3724 RTCPUID const idAssertCpuNow = RTThreadPreemptIsEnabled(NIL_RTTHREAD) ? NIL_RTCPUID : RTMpCpuId(); \
3725 AssertMsg(idAssertCpu == idAssertCpuNow, ("SVM %#x, %#x\n", idAssertCpu, idAssertCpuNow)); \
3726 } while (0)
3727
3728# define HMSVM_VALIDATE_EXIT_HANDLER_PARAMS() \
3729 do { \
3730 AssertPtr(pVCpu); \
3731 AssertPtr(pCtx); \
3732 AssertPtr(pSvmTransient); \
3733 Assert(ASMIntAreEnabled()); \
3734 HMSVM_ASSERT_PREEMPT_SAFE(); \
3735 HMSVM_ASSERT_PREEMPT_CPUID_VAR(); \
3736 Log4Func(("vcpu[%u] -v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-v-\n", (uint32_t)pVCpu->idCpu)); \
3737 HMSVM_ASSERT_PREEMPT_SAFE(); \
3738 if (VMMR0IsLogFlushDisabled(pVCpu)) \
3739 HMSVM_ASSERT_PREEMPT_CPUID(); \
3740 } while (0)
3741#else /* Release builds */
3742# define HMSVM_VALIDATE_EXIT_HANDLER_PARAMS() do { NOREF(pVCpu); NOREF(pCtx); NOREF(pSvmTransient); } while (0)
3743#endif
3744
3745
3746/**
3747 * Worker for hmR0SvmInterpretInvlpg().
3748 *
3749 * @return VBox status code.
3750 * @param pVCpu Pointer to the VMCPU.
3751 * @param pCpu Pointer to the disassembler state.
3752 * @param pCtx The guest CPU context.
3753 */
3754static int hmR0SvmInterpretInvlPgEx(PVMCPU pVCpu, PDISCPUSTATE pCpu, PCPUMCTX pCtx)
3755{
3756 DISQPVPARAMVAL Param1;
3757 RTGCPTR GCPtrPage;
3758
3759 int rc = DISQueryParamVal(CPUMCTX2CORE(pCtx), pCpu, &pCpu->Param1, &Param1, DISQPVWHICH_SRC);
3760 if (RT_FAILURE(rc))
3761 return VERR_EM_INTERPRETER;
3762
3763 if ( Param1.type == DISQPV_TYPE_IMMEDIATE
3764 || Param1.type == DISQPV_TYPE_ADDRESS)
3765 {
3766 if (!(Param1.flags & (DISQPV_FLAG_32 | DISQPV_FLAG_64)))
3767 return VERR_EM_INTERPRETER;
3768
3769 GCPtrPage = Param1.val.val64;
3770 VBOXSTRICTRC rc2 = EMInterpretInvlpg(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx), GCPtrPage);
3771 rc = VBOXSTRICTRC_VAL(rc2);
3772 }
3773 else
3774 {
3775 Log4(("hmR0SvmInterpretInvlPgEx invalid parameter type %#x\n", Param1.type));
3776 rc = VERR_EM_INTERPRETER;
3777 }
3778
3779 return rc;
3780}
3781
3782
3783/**
3784 * Interprets INVLPG.
3785 *
3786 * @returns VBox status code.
3787 * @retval VINF_* Scheduling instructions.
3788 * @retval VERR_EM_INTERPRETER Something we can't cope with.
3789 * @retval VERR_* Fatal errors.
3790 *
3791 * @param pVM Pointer to the VM.
3792 * @param pCtx The guest CPU context.
3793 *
3794 * @remarks Updates the RIP if the instruction was executed successfully.
3795 */
3796static int hmR0SvmInterpretInvlpg(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3797{
3798 /* Only allow 32 & 64 bit code. */
3799 if (CPUMGetGuestCodeBits(pVCpu) != 16)
3800 {
3801 PDISSTATE pDis = &pVCpu->hm.s.DisState;
3802 int rc = EMInterpretDisasCurrent(pVM, pVCpu, pDis, NULL /* pcbInstr */);
3803 if ( RT_SUCCESS(rc)
3804 && pDis->pCurInstr->uOpcode == OP_INVLPG)
3805 {
3806 rc = hmR0SvmInterpretInvlPgEx(pVCpu, pDis, pCtx);
3807 if (RT_SUCCESS(rc))
3808 pCtx->rip += pDis->cbInstr;
3809 return rc;
3810 }
3811 else
3812 Log4(("hmR0SvmInterpretInvlpg: EMInterpretDisasCurrent returned %Rrc uOpCode=%#x\n", rc, pDis->pCurInstr->uOpcode));
3813 }
3814 return VERR_EM_INTERPRETER;
3815}
3816
3817
3818/**
3819 * Sets an invalid-opcode (#UD) exception as pending-for-injection into the VM.
3820 *
3821 * @param pVCpu Pointer to the VMCPU.
3822 */
3823DECLINLINE(void) hmR0SvmSetPendingXcptUD(PVMCPU pVCpu)
3824{
3825 SVMEVENT Event;
3826 Event.u = 0;
3827 Event.n.u1Valid = 1;
3828 Event.n.u3Type = SVM_EVENT_EXCEPTION;
3829 Event.n.u8Vector = X86_XCPT_UD;
3830 hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3831}
3832
3833
3834/**
3835 * Sets a debug (#DB) exception as pending-for-injection into the VM.
3836 *
3837 * @param pVCpu Pointer to the VMCPU.
3838 */
3839DECLINLINE(void) hmR0SvmSetPendingXcptDB(PVMCPU pVCpu)
3840{
3841 SVMEVENT Event;
3842 Event.u = 0;
3843 Event.n.u1Valid = 1;
3844 Event.n.u3Type = SVM_EVENT_EXCEPTION;
3845 Event.n.u8Vector = X86_XCPT_DB;
3846 hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3847}
3848
3849
3850/**
3851 * Sets a page fault (#PF) exception as pending-for-injection into the VM.
3852 *
3853 * @param pVCpu Pointer to the VMCPU.
3854 * @param pCtx Pointer to the guest-CPU context.
3855 * @param u32ErrCode The error-code for the page-fault.
3856 * @param uFaultAddress The page fault address (CR2).
3857 *
3858 * @remarks This updates the guest CR2 with @a uFaultAddress!
3859 */
3860DECLINLINE(void) hmR0SvmSetPendingXcptPF(PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t u32ErrCode, RTGCUINTPTR uFaultAddress)
3861{
3862 SVMEVENT Event;
3863 Event.u = 0;
3864 Event.n.u1Valid = 1;
3865 Event.n.u3Type = SVM_EVENT_EXCEPTION;
3866 Event.n.u8Vector = X86_XCPT_PF;
3867 Event.n.u1ErrorCodeValid = 1;
3868 Event.n.u32ErrorCode = u32ErrCode;
3869
3870 /* Update CR2 of the guest. */
3871 if (pCtx->cr2 != uFaultAddress)
3872 {
3873 pCtx->cr2 = uFaultAddress;
3874 HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR2);
3875 }
3876
3877 hmR0SvmSetPendingEvent(pVCpu, &Event, uFaultAddress);
3878}
3879
3880
3881/**
3882 * Sets a device-not-available (#NM) exception as pending-for-injection into the
3883 * VM.
3884 *
3885 * @param pVCpu Pointer to the VMCPU.
3886 */
3887DECLINLINE(void) hmR0SvmSetPendingXcptNM(PVMCPU pVCpu)
3888{
3889 SVMEVENT Event;
3890 Event.u = 0;
3891 Event.n.u1Valid = 1;
3892 Event.n.u3Type = SVM_EVENT_EXCEPTION;
3893 Event.n.u8Vector = X86_XCPT_NM;
3894 hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3895}
3896
3897
3898/**
3899 * Sets a math-fault (#MF) exception as pending-for-injection into the VM.
3900 *
3901 * @param pVCpu Pointer to the VMCPU.
3902 */
3903DECLINLINE(void) hmR0SvmSetPendingXcptMF(PVMCPU pVCpu)
3904{
3905 SVMEVENT Event;
3906 Event.u = 0;
3907 Event.n.u1Valid = 1;
3908 Event.n.u3Type = SVM_EVENT_EXCEPTION;
3909 Event.n.u8Vector = X86_XCPT_MF;
3910 hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3911}
3912
3913
3914/**
3915 * Sets a double fault (#DF) exception as pending-for-injection into the VM.
3916 *
3917 * @param pVCpu Pointer to the VMCPU.
3918 */
3919DECLINLINE(void) hmR0SvmSetPendingXcptDF(PVMCPU pVCpu)
3920{
3921 SVMEVENT Event;
3922 Event.u = 0;
3923 Event.n.u1Valid = 1;
3924 Event.n.u3Type = SVM_EVENT_EXCEPTION;
3925 Event.n.u8Vector = X86_XCPT_DF;
3926 Event.n.u1ErrorCodeValid = 1;
3927 Event.n.u32ErrorCode = 0;
3928 hmR0SvmSetPendingEvent(pVCpu, &Event, 0 /* GCPtrFaultAddress */);
3929}
3930
3931
3932/**
3933 * Emulates a simple MOV TPR (CR8) instruction, used for TPR patching on 32-bit
3934 * guests. This simply looks up the patch record at EIP and does the required.
3935 *
3936 * This VMMCALL is used a fallback mechanism when mov to/from cr8 isn't exactly
3937 * like how we want it to be (e.g. not followed by shr 4 as is usually done for
3938 * TPR). See hmR3ReplaceTprInstr() for the details.
3939 *
3940 * @returns VBox status code.
3941 * @retval VINF_SUCCESS if the access was handled successfully.
3942 * @retval VERR_NOT_FOUND if no patch record for this RIP could be found.
3943 * @retval VERR_SVM_UNEXPECTED_PATCH_TYPE if the found patch type is invalid.
3944 *
3945 * @param pVM Pointer to the VM.
3946 * @param pVCpu Pointer to the VMCPU.
3947 * @param pCtx Pointer to the guest-CPU context.
3948 */
3949static int hmR0SvmEmulateMovTpr(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
3950{
3951 Log4(("Emulated VMMCall TPR access replacement at RIP=%RGv\n", pCtx->rip));
3952
3953 /*
3954 * We do this in a loop as we increment the RIP after a successful emulation
3955 * and the new RIP may be a patched instruction which needs emulation as well.
3956 */
3957 bool fPatchFound = false;
3958 for (;;)
3959 {
3960 bool fPending;
3961 uint8_t u8Tpr;
3962
3963 PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
3964 if (!pPatch)
3965 break;
3966
3967 fPatchFound = true;
3968 switch (pPatch->enmType)
3969 {
3970 case HMTPRINSTR_READ:
3971 {
3972 int rc = PDMApicGetTPR(pVCpu, &u8Tpr, &fPending, NULL /* pu8PendingIrq */);
3973 AssertRC(rc);
3974
3975 rc = DISWriteReg32(CPUMCTX2CORE(pCtx), pPatch->uDstOperand, u8Tpr);
3976 AssertRC(rc);
3977 pCtx->rip += pPatch->cbOp;
3978 break;
3979 }
3980
3981 case HMTPRINSTR_WRITE_REG:
3982 case HMTPRINSTR_WRITE_IMM:
3983 {
3984 if (pPatch->enmType == HMTPRINSTR_WRITE_REG)
3985 {
3986 uint32_t u32Val;
3987 int rc = DISFetchReg32(CPUMCTX2CORE(pCtx), pPatch->uSrcOperand, &u32Val);
3988 AssertRC(rc);
3989 u8Tpr = u32Val;
3990 }
3991 else
3992 u8Tpr = (uint8_t)pPatch->uSrcOperand;
3993
3994 int rc2 = PDMApicSetTPR(pVCpu, u8Tpr);
3995 AssertRC(rc2);
3996 HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
3997
3998 pCtx->rip += pPatch->cbOp;
3999 break;
4000 }
4001
4002 default:
4003 AssertMsgFailed(("Unexpected patch type %d\n", pPatch->enmType));
4004 pVCpu->hm.s.u32HMError = pPatch->enmType;
4005 return VERR_SVM_UNEXPECTED_PATCH_TYPE;
4006 }
4007 }
4008
4009 if (fPatchFound)
4010 return VINF_SUCCESS;
4011 return VERR_NOT_FOUND;
4012}
4013
4014
4015/**
4016 * Determines if an exception is a contributory exception.
4017 *
4018 * Contributory exceptions are ones which can cause double-faults unless the
4019 * original exception was a benign exception. Page-fault is intentionally not
4020 * included here as it's a conditional contributory exception.
4021 *
4022 * @returns true if the exception is contributory, false otherwise.
4023 * @param uVector The exception vector.
4024 */
4025DECLINLINE(bool) hmR0SvmIsContributoryXcpt(const uint32_t uVector)
4026{
4027 switch (uVector)
4028 {
4029 case X86_XCPT_GP:
4030 case X86_XCPT_SS:
4031 case X86_XCPT_NP:
4032 case X86_XCPT_TS:
4033 case X86_XCPT_DE:
4034 return true;
4035 default:
4036 break;
4037 }
4038 return false;
4039}
4040
4041
4042/**
4043 * Handle a condition that occurred while delivering an event through the guest
4044 * IDT.
4045 *
4046 * @returns VBox status code (informational error codes included).
4047 * @retval VINF_SUCCESS if we should continue handling the #VMEXIT.
4048 * @retval VINF_HM_DOUBLE_FAULT if a #DF condition was detected and we ought to
4049 * continue execution of the guest which will delivery the #DF.
4050 * @retval VINF_EM_RESET if we detected a triple-fault condition.
4051 *
4052 * @param pVCpu Pointer to the VMCPU.
4053 * @param pCtx Pointer to the guest-CPU context.
4054 * @param pSvmTransient Pointer to the SVM transient structure.
4055 *
4056 * @remarks No-long-jump zone!!!
4057 */
4058static int hmR0SvmCheckExitDueToEventDelivery(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4059{
4060 int rc = VINF_SUCCESS;
4061 PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4062
4063 /* See AMD spec. 15.7.3 "EXITINFO Pseudo-Code". The EXITINTINFO (if valid) contains the prior exception (IDT vector)
4064 * that was trying to be delivered to the guest which caused a #VMEXIT which was intercepted (Exit vector). */
4065 if (pVmcb->ctrl.ExitIntInfo.n.u1Valid)
4066 {
4067 uint8_t uIdtVector = pVmcb->ctrl.ExitIntInfo.n.u8Vector;
4068
4069 typedef enum
4070 {
4071 SVMREFLECTXCPT_XCPT, /* Reflect the exception to the guest or for further evaluation by VMM. */
4072 SVMREFLECTXCPT_DF, /* Reflect the exception as a double-fault to the guest. */
4073 SVMREFLECTXCPT_TF, /* Indicate a triple faulted state to the VMM. */
4074 SVMREFLECTXCPT_NONE /* Nothing to reflect. */
4075 } SVMREFLECTXCPT;
4076
4077 SVMREFLECTXCPT enmReflect = SVMREFLECTXCPT_NONE;
4078 bool fReflectingNmi = false;
4079 if (pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_EXCEPTION)
4080 {
4081 if (pSvmTransient->u64ExitCode - SVM_EXIT_EXCEPTION_0 <= SVM_EXIT_EXCEPTION_1F)
4082 {
4083 uint8_t uExitVector = (uint8_t)(pSvmTransient->u64ExitCode - SVM_EXIT_EXCEPTION_0);
4084
4085#ifdef VBOX_STRICT
4086 if ( hmR0SvmIsContributoryXcpt(uIdtVector)
4087 && uExitVector == X86_XCPT_PF)
4088 {
4089 Log4(("IDT: Contributory #PF uCR2=%#RX64\n", pVCpu->idCpu, pCtx->cr2));
4090 }
4091#endif
4092 if ( uExitVector == X86_XCPT_PF
4093 && uIdtVector == X86_XCPT_PF)
4094 {
4095 pSvmTransient->fVectoringDoublePF = true;
4096 Log4(("IDT: Vectoring double #PF uCR2=%#RX64\n", pCtx->cr2));
4097 }
4098 else if ( (pVmcb->ctrl.u32InterceptException & HMSVM_CONTRIBUTORY_XCPT_MASK)
4099 && hmR0SvmIsContributoryXcpt(uExitVector)
4100 && ( hmR0SvmIsContributoryXcpt(uIdtVector)
4101 || uIdtVector == X86_XCPT_PF))
4102 {
4103 enmReflect = SVMREFLECTXCPT_DF;
4104 Log4(("IDT: Pending vectoring #DF %#RX64 uIdtVector=%#x uExitVector=%#x\n", pVCpu->hm.s.Event.u64IntInfo,
4105 uIdtVector, uExitVector));
4106 }
4107 else if (uIdtVector == X86_XCPT_DF)
4108 {
4109 enmReflect = SVMREFLECTXCPT_TF;
4110 Log4(("IDT: Pending vectoring triple-fault %#RX64 uIdtVector=%#x uExitVector=%#x\n",
4111 pVCpu->hm.s.Event.u64IntInfo, uIdtVector, uExitVector));
4112 }
4113 else
4114 enmReflect = SVMREFLECTXCPT_XCPT;
4115 }
4116 else
4117 {
4118 /*
4119 * If event delivery caused an #VMEXIT that is not an exception (e.g. #NPF) then reflect the original
4120 * exception to the guest after handling the #VMEXIT.
4121 */
4122 enmReflect = SVMREFLECTXCPT_XCPT;
4123 }
4124 }
4125 else if ( pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_EXTERNAL_IRQ
4126 || pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_NMI)
4127 {
4128 enmReflect = SVMREFLECTXCPT_XCPT;
4129 fReflectingNmi = RT_BOOL(pVmcb->ctrl.ExitIntInfo.n.u3Type == SVM_EVENT_NMI);
4130
4131 if (pSvmTransient->u64ExitCode - SVM_EXIT_EXCEPTION_0 <= SVM_EXIT_EXCEPTION_1F)
4132 {
4133 uint8_t uExitVector = (uint8_t)(pSvmTransient->u64ExitCode - SVM_EXIT_EXCEPTION_0);
4134 if (uExitVector == X86_XCPT_PF)
4135 {
4136 pSvmTransient->fVectoringPF = true;
4137 Log4(("IDT: Vectoring #PF due to Ext-Int/NMI. uCR2=%#RX64\n", pCtx->cr2));
4138 }
4139 }
4140 }
4141 /* else: Ignore software interrupts (INT n) as they reoccur when restarting the instruction. */
4142
4143 switch (enmReflect)
4144 {
4145 case SVMREFLECTXCPT_XCPT:
4146 {
4147 /* If we are re-injecting the NMI, clear NMI blocking. */
4148 if (fReflectingNmi)
4149 VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_BLOCK_NMIS);
4150
4151 Assert(pVmcb->ctrl.ExitIntInfo.n.u3Type != SVM_EVENT_SOFTWARE_INT);
4152 hmR0SvmSetPendingEvent(pVCpu, &pVmcb->ctrl.ExitIntInfo, 0 /* GCPtrFaultAddress */);
4153
4154 /* If uExitVector is #PF, CR2 value will be updated from the VMCB if it's a guest #PF. See hmR0SvmExitXcptPF(). */
4155 Log4(("IDT: Pending vectoring event %#RX64 ErrValid=%RTbool Err=%#RX32\n", pVmcb->ctrl.ExitIntInfo.u,
4156 !!pVmcb->ctrl.ExitIntInfo.n.u1ErrorCodeValid, pVmcb->ctrl.ExitIntInfo.n.u32ErrorCode));
4157 break;
4158 }
4159
4160 case SVMREFLECTXCPT_DF:
4161 {
4162 hmR0SvmSetPendingXcptDF(pVCpu);
4163 rc = VINF_HM_DOUBLE_FAULT;
4164 break;
4165 }
4166
4167 case SVMREFLECTXCPT_TF:
4168 {
4169 rc = VINF_EM_RESET;
4170 break;
4171 }
4172
4173 default:
4174 Assert(rc == VINF_SUCCESS);
4175 break;
4176 }
4177 }
4178 Assert(rc == VINF_SUCCESS || rc == VINF_HM_DOUBLE_FAULT || rc == VINF_EM_RESET);
4179 NOREF(pCtx);
4180 return rc;
4181}
4182
4183
4184/**
4185 * Advances the guest RIP in the if the NRIP_SAVE feature is supported by the
4186 * CPU, otherwise advances the RIP by @a cb bytes.
4187 *
4188 * @param pVCpu Pointer to the VMCPU.
4189 * @param pCtx Pointer to the guest-CPU context.
4190 * @param cb RIP increment value in bytes.
4191 *
4192 * @remarks Use this function only from #VMEXIT's where the NRIP value is valid
4193 * when NRIP_SAVE is supported by the CPU!
4194 */
4195DECLINLINE(void) hmR0SvmUpdateRip(PVMCPU pVCpu, PCPUMCTX pCtx, uint32_t cb)
4196{
4197 if (pVCpu->CTX_SUFF(pVM)->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_NRIP_SAVE)
4198 {
4199 PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4200 Assert(pVmcb->ctrl.u64NextRIP - pCtx->rip == cb);
4201 pCtx->rip = pVmcb->ctrl.u64NextRIP;
4202 }
4203 else
4204 pCtx->rip += cb;
4205}
4206
4207
4208/* -=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= */
4209/* -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- #VMEXIT handlers -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- */
4210/* -=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=--=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= */
4211
4212/** @name #VMEXIT handlers.
4213 * @{
4214 */
4215
4216/**
4217 * #VMEXIT handler for external interrupts, NMIs, FPU assertion freeze and INIT
4218 * signals (SVM_EXIT_INTR, SVM_EXIT_NMI, SVM_EXIT_FERR_FREEZE, SVM_EXIT_INIT).
4219 */
4220HMSVM_EXIT_DECL hmR0SvmExitIntr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4221{
4222 HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4223
4224 if (pSvmTransient->u64ExitCode == SVM_EXIT_NMI)
4225 STAM_REL_COUNTER_INC(&pVCpu->hm.s.StatExitHostNmiInGC);
4226 else if (pSvmTransient->u64ExitCode == SVM_EXIT_INTR)
4227 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitExtInt);
4228
4229 /*
4230 * AMD-V has no preemption timer and the generic periodic preemption timer has no way to signal -before- the timer
4231 * fires if the current interrupt is our own timer or a some other host interrupt. We also cannot examine what
4232 * interrupt it is until the host actually take the interrupt.
4233 *
4234 * Going back to executing guest code here unconditionally causes random scheduling problems (observed on an
4235 * AMD Phenom 9850 Quad-Core on Windows 64-bit host).
4236 */
4237 return VINF_EM_RAW_INTERRUPT;
4238}
4239
4240
4241/**
4242 * #VMEXIT handler for WBINVD (SVM_EXIT_WBINVD). Conditional #VMEXIT.
4243 */
4244HMSVM_EXIT_DECL hmR0SvmExitWbinvd(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4245{
4246 HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4247
4248 hmR0SvmUpdateRip(pVCpu, pCtx, 2);
4249 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitWbinvd);
4250 int rc = VINF_SUCCESS;
4251 HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4252 return rc;
4253}
4254
4255
4256/**
4257 * #VMEXIT handler for INVD (SVM_EXIT_INVD). Unconditional #VMEXIT.
4258 */
4259HMSVM_EXIT_DECL hmR0SvmExitInvd(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4260{
4261 HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4262
4263 hmR0SvmUpdateRip(pVCpu, pCtx, 2);
4264 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitInvd);
4265 int rc = VINF_SUCCESS;
4266 HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4267 return rc;
4268}
4269
4270
4271/**
4272 * #VMEXIT handler for INVD (SVM_EXIT_CPUID). Conditional #VMEXIT.
4273 */
4274HMSVM_EXIT_DECL hmR0SvmExitCpuid(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4275{
4276 HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4277 PVM pVM = pVCpu->CTX_SUFF(pVM);
4278 int rc = EMInterpretCpuId(pVM, pVCpu, CPUMCTX2CORE(pCtx));
4279 if (RT_LIKELY(rc == VINF_SUCCESS))
4280 {
4281 hmR0SvmUpdateRip(pVCpu, pCtx, 2);
4282 HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4283 }
4284 else
4285 {
4286 AssertMsgFailed(("hmR0SvmExitCpuid: EMInterpretCpuId failed with %Rrc\n", rc));
4287 rc = VERR_EM_INTERPRETER;
4288 }
4289 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCpuid);
4290 return rc;
4291}
4292
4293
4294/**
4295 * #VMEXIT handler for RDTSC (SVM_EXIT_RDTSC). Conditional #VMEXIT.
4296 */
4297HMSVM_EXIT_DECL hmR0SvmExitRdtsc(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4298{
4299 HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4300 PVM pVM = pVCpu->CTX_SUFF(pVM);
4301 int rc = EMInterpretRdtsc(pVM, pVCpu, CPUMCTX2CORE(pCtx));
4302 if (RT_LIKELY(rc == VINF_SUCCESS))
4303 {
4304 hmR0SvmUpdateRip(pVCpu, pCtx, 2);
4305 pSvmTransient->fUpdateTscOffsetting = true;
4306
4307 /* Single step check. */
4308 HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4309 }
4310 else
4311 {
4312 AssertMsgFailed(("hmR0SvmExitRdtsc: EMInterpretRdtsc failed with %Rrc\n", rc));
4313 rc = VERR_EM_INTERPRETER;
4314 }
4315 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdtsc);
4316 return rc;
4317}
4318
4319
4320/**
4321 * #VMEXIT handler for RDTSCP (SVM_EXIT_RDTSCP). Conditional #VMEXIT.
4322 */
4323HMSVM_EXIT_DECL hmR0SvmExitRdtscp(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4324{
4325 HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4326 int rc = EMInterpretRdtscp(pVCpu->CTX_SUFF(pVM), pVCpu, pCtx);
4327 if (RT_LIKELY(rc == VINF_SUCCESS))
4328 {
4329 hmR0SvmUpdateRip(pVCpu, pCtx, 3);
4330 pSvmTransient->fUpdateTscOffsetting = true;
4331 HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4332 }
4333 else
4334 {
4335 AssertMsgFailed(("hmR0SvmExitRdtsc: EMInterpretRdtscp failed with %Rrc\n", rc));
4336 rc = VERR_EM_INTERPRETER;
4337 }
4338 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdtscp);
4339 return rc;
4340}
4341
4342
4343/**
4344 * #VMEXIT handler for RDPMC (SVM_EXIT_RDPMC). Conditional #VMEXIT.
4345 */
4346HMSVM_EXIT_DECL hmR0SvmExitRdpmc(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4347{
4348 HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4349 int rc = EMInterpretRdpmc(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
4350 if (RT_LIKELY(rc == VINF_SUCCESS))
4351 {
4352 hmR0SvmUpdateRip(pVCpu, pCtx, 2);
4353 HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4354 }
4355 else
4356 {
4357 AssertMsgFailed(("hmR0SvmExitRdpmc: EMInterpretRdpmc failed with %Rrc\n", rc));
4358 rc = VERR_EM_INTERPRETER;
4359 }
4360 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdpmc);
4361 return rc;
4362}
4363
4364
4365/**
4366 * #VMEXIT handler for INVLPG (SVM_EXIT_INVLPG). Conditional #VMEXIT.
4367 */
4368HMSVM_EXIT_DECL hmR0SvmExitInvlpg(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4369{
4370 HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4371 PVM pVM = pVCpu->CTX_SUFF(pVM);
4372 Assert(!pVM->hm.s.fNestedPaging);
4373
4374 /** @todo Decode Assist. */
4375 int rc = hmR0SvmInterpretInvlpg(pVM, pVCpu, pCtx); /* Updates RIP if successful. */
4376 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitInvlpg);
4377 Assert(rc == VINF_SUCCESS || rc == VERR_EM_INTERPRETER);
4378 HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4379 return rc;
4380}
4381
4382
4383/**
4384 * #VMEXIT handler for HLT (SVM_EXIT_HLT). Conditional #VMEXIT.
4385 */
4386HMSVM_EXIT_DECL hmR0SvmExitHlt(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4387{
4388 HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4389
4390 hmR0SvmUpdateRip(pVCpu, pCtx, 1);
4391 int rc = EMShouldContinueAfterHalt(pVCpu, pCtx) ? VINF_SUCCESS : VINF_EM_HALT;
4392 HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4393 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitHlt);
4394 if (rc != VINF_SUCCESS)
4395 STAM_COUNTER_INC(&pVCpu->hm.s.StatSwitchHltToR3);
4396 return rc;
4397}
4398
4399
4400/**
4401 * #VMEXIT handler for MONITOR (SVM_EXIT_MONITOR). Conditional #VMEXIT.
4402 */
4403HMSVM_EXIT_DECL hmR0SvmExitMonitor(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4404{
4405 HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4406 int rc = EMInterpretMonitor(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
4407 if (RT_LIKELY(rc == VINF_SUCCESS))
4408 {
4409 hmR0SvmUpdateRip(pVCpu, pCtx, 3);
4410 HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4411 }
4412 else
4413 {
4414 AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMonitor: EMInterpretMonitor failed with %Rrc\n", rc));
4415 rc = VERR_EM_INTERPRETER;
4416 }
4417 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitMonitor);
4418 return rc;
4419}
4420
4421
4422/**
4423 * #VMEXIT handler for MWAIT (SVM_EXIT_MWAIT). Conditional #VMEXIT.
4424 */
4425HMSVM_EXIT_DECL hmR0SvmExitMwait(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4426{
4427 HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4428 VBOXSTRICTRC rc2 = EMInterpretMWait(pVCpu->CTX_SUFF(pVM), pVCpu, CPUMCTX2CORE(pCtx));
4429 int rc = VBOXSTRICTRC_VAL(rc2);
4430 if ( rc == VINF_EM_HALT
4431 || rc == VINF_SUCCESS)
4432 {
4433 hmR0SvmUpdateRip(pVCpu, pCtx, 3);
4434
4435 if ( rc == VINF_EM_HALT
4436 && EMMonitorWaitShouldContinue(pVCpu, pCtx))
4437 {
4438 rc = VINF_SUCCESS;
4439 }
4440 HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4441 }
4442 else
4443 {
4444 AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMwait: EMInterpretMWait failed with %Rrc\n", rc));
4445 rc = VERR_EM_INTERPRETER;
4446 }
4447 AssertMsg(rc == VINF_SUCCESS || rc == VINF_EM_HALT || rc == VERR_EM_INTERPRETER,
4448 ("hmR0SvmExitMwait: EMInterpretMWait failed rc=%Rrc\n", rc));
4449 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitMwait);
4450 return rc;
4451}
4452
4453
4454/**
4455 * #VMEXIT handler for shutdown (triple-fault) (SVM_EXIT_SHUTDOWN).
4456 * Conditional #VMEXIT.
4457 */
4458HMSVM_EXIT_DECL hmR0SvmExitShutdown(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4459{
4460 HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4461 return VINF_EM_RESET;
4462}
4463
4464
4465/**
4466 * #VMEXIT handler for CRx reads (SVM_EXIT_READ_CR*). Conditional #VMEXIT.
4467 */
4468HMSVM_EXIT_DECL hmR0SvmExitReadCRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4469{
4470 HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4471
4472 Log4(("hmR0SvmExitReadCRx: CS:RIP=%04x:%#RX64\n", pCtx->cs.Sel, pCtx->rip));
4473
4474 /** @todo Decode Assist. */
4475 VBOXSTRICTRC rc2 = EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */);
4476 int rc = VBOXSTRICTRC_VAL(rc2);
4477 AssertMsg(rc == VINF_SUCCESS || rc == VERR_EM_INTERPRETER || rc == VINF_PGM_CHANGE_MODE || rc == VINF_PGM_SYNC_CR3,
4478 ("hmR0SvmExitReadCRx: EMInterpretInstruction failed rc=%Rrc\n", rc));
4479 Assert((pSvmTransient->u64ExitCode - SVM_EXIT_READ_CR0) <= 15);
4480 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitCRxRead[pSvmTransient->u64ExitCode - SVM_EXIT_READ_CR0]);
4481 HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4482 return rc;
4483}
4484
4485
4486/**
4487 * #VMEXIT handler for CRx writes (SVM_EXIT_WRITE_CR*). Conditional #VMEXIT.
4488 */
4489HMSVM_EXIT_DECL hmR0SvmExitWriteCRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4490{
4491 HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4492 /** @todo Decode Assist. */
4493 VBOXSTRICTRC rc2 = EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */);
4494 int rc = VBOXSTRICTRC_VAL(rc2);
4495 if (rc == VINF_SUCCESS)
4496 {
4497 /* RIP has been updated by EMInterpretInstruction(). */
4498 Assert((pSvmTransient->u64ExitCode - SVM_EXIT_WRITE_CR0) <= 15);
4499 switch (pSvmTransient->u64ExitCode - SVM_EXIT_WRITE_CR0)
4500 {
4501 case 0: /* CR0. */
4502 HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR0);
4503 break;
4504
4505 case 3: /* CR3. */
4506 Assert(!pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
4507 HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR3);
4508 break;
4509
4510 case 4: /* CR4. */
4511 HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR4);
4512 break;
4513
4514 case 8: /* CR8 (TPR). */
4515 HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
4516 break;
4517
4518 default:
4519 AssertMsgFailed(("hmR0SvmExitWriteCRx: Invalid/Unexpected Write-CRx exit. u64ExitCode=%#RX64 %#x CRx=%#RX64\n",
4520 pSvmTransient->u64ExitCode, pSvmTransient->u64ExitCode - SVM_EXIT_WRITE_CR0));
4521 break;
4522 }
4523 HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4524 }
4525 else
4526 Assert(rc == VERR_EM_INTERPRETER || rc == VINF_PGM_CHANGE_MODE || rc == VINF_PGM_SYNC_CR3);
4527 return rc;
4528}
4529
4530
4531/**
4532 * #VMEXIT handler for instructions that result in a #UD exception delivered to
4533 * the guest.
4534 */
4535HMSVM_EXIT_DECL hmR0SvmExitSetPendingXcptUD(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4536{
4537 HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4538 hmR0SvmSetPendingXcptUD(pVCpu);
4539 return VINF_SUCCESS;
4540}
4541
4542
4543/**
4544 * #VMEXIT handler for MSR read and writes (SVM_EXIT_MSR). Conditional #VMEXIT.
4545 */
4546HMSVM_EXIT_DECL hmR0SvmExitMsr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4547{
4548 HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4549 PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4550 PVM pVM = pVCpu->CTX_SUFF(pVM);
4551
4552 int rc;
4553 if (pVmcb->ctrl.u64ExitInfo1 == SVM_EXIT1_MSR_WRITE)
4554 {
4555 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitWrmsr);
4556
4557 /* Handle TPR patching; intercepted LSTAR write. */
4558 if ( pVM->hm.s.fTPRPatchingActive
4559 && pCtx->ecx == MSR_K8_LSTAR)
4560 {
4561 if ((pCtx->eax & 0xff) != pSvmTransient->u8GuestTpr)
4562 {
4563 /* Our patch code uses LSTAR for TPR caching for 32-bit guests. */
4564 int rc2 = PDMApicSetTPR(pVCpu, pCtx->eax & 0xff);
4565 AssertRC(rc2);
4566 HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
4567 }
4568 hmR0SvmUpdateRip(pVCpu, pCtx, 2);
4569 rc = VINF_SUCCESS;
4570 HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4571 return rc;
4572 }
4573
4574 if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_NRIP_SAVE)
4575 {
4576 rc = EMInterpretWrmsr(pVM, pVCpu, CPUMCTX2CORE(pCtx));
4577 if (RT_LIKELY(rc == VINF_SUCCESS))
4578 {
4579 pCtx->rip = pVmcb->ctrl.u64NextRIP;
4580 HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4581 }
4582 else
4583 AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMsr: EMInterpretWrmsr failed rc=%Rrc\n", rc));
4584 }
4585 else
4586 {
4587 rc = VBOXSTRICTRC_TODO(EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */));
4588 if (RT_LIKELY(rc == VINF_SUCCESS))
4589 HMSVM_CHECK_SINGLE_STEP(pVCpu, rc); /* RIP updated by EMInterpretInstruction(). */
4590 else
4591 AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMsr: WrMsr. EMInterpretInstruction failed rc=%Rrc\n", rc));
4592 }
4593
4594 if (rc == VINF_SUCCESS)
4595 {
4596 /* If this is an X2APIC WRMSR access, update the APIC state as well. */
4597 if ( pCtx->ecx >= MSR_IA32_X2APIC_START
4598 && pCtx->ecx <= MSR_IA32_X2APIC_END)
4599 {
4600 /*
4601 * We've already saved the APIC related guest-state (TPR) in hmR0SvmPostRunGuest(). When full APIC register
4602 * virtualization is implemented we'll have to make sure APIC state is saved from the VMCB before
4603 * EMInterpretWrmsr() changes it.
4604 */
4605 HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
4606 }
4607 else if (pCtx->ecx == MSR_K6_EFER)
4608 HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_EFER_MSR);
4609 else if (pCtx->ecx == MSR_IA32_TSC)
4610 pSvmTransient->fUpdateTscOffsetting = true;
4611 }
4612 }
4613 else
4614 {
4615 /* MSR Read access. */
4616 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitRdmsr);
4617 Assert(pVmcb->ctrl.u64ExitInfo1 == SVM_EXIT1_MSR_READ);
4618
4619 if (pVM->hm.s.svm.u32Features & AMD_CPUID_SVM_FEATURE_EDX_NRIP_SAVE)
4620 {
4621 rc = EMInterpretRdmsr(pVM, pVCpu, CPUMCTX2CORE(pCtx));
4622 if (RT_LIKELY(rc == VINF_SUCCESS))
4623 {
4624 pCtx->rip = pVmcb->ctrl.u64NextRIP;
4625 HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4626 }
4627 else
4628 AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMsr: EMInterpretRdmsr failed rc=%Rrc\n", rc));
4629 }
4630 else
4631 {
4632 rc = VBOXSTRICTRC_TODO(EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0));
4633 if (RT_UNLIKELY(rc != VINF_SUCCESS))
4634 AssertMsg(rc == VERR_EM_INTERPRETER, ("hmR0SvmExitMsr: RdMsr. EMInterpretInstruction failed rc=%Rrc\n", rc));
4635 /* RIP updated by EMInterpretInstruction(). */
4636 HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4637 }
4638 }
4639
4640 /* RIP has been updated by EMInterpret[Rd|Wr]msr(). */
4641 return rc;
4642}
4643
4644
4645/**
4646 * #VMEXIT handler for DRx read (SVM_EXIT_READ_DRx). Conditional #VMEXIT.
4647 */
4648HMSVM_EXIT_DECL hmR0SvmExitReadDRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4649{
4650 HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4651 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitDRxRead);
4652
4653 /* We should -not- get this #VMEXIT if the guest's debug registers were active. */
4654 if (pSvmTransient->fWasGuestDebugStateActive)
4655 {
4656 AssertMsgFailed(("hmR0SvmHandleExit: Unexpected exit %#RX32\n", (uint32_t)pSvmTransient->u64ExitCode));
4657 pVCpu->hm.s.u32HMError = (uint32_t)pSvmTransient->u64ExitCode;
4658 return VERR_SVM_UNEXPECTED_EXIT;
4659 }
4660
4661 /*
4662 * Lazy DR0-3 loading.
4663 */
4664 if (!pSvmTransient->fWasHyperDebugStateActive)
4665 {
4666 Assert(!DBGFIsStepping(pVCpu)); Assert(!pVCpu->hm.s.fSingleInstruction);
4667 Log5(("hmR0SvmExitReadDRx: Lazy loading guest debug registers\n"));
4668
4669 /* Don't intercept DRx read and writes. */
4670 PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4671 pVmcb->ctrl.u16InterceptRdDRx = 0;
4672 pVmcb->ctrl.u16InterceptWrDRx = 0;
4673 pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_INTERCEPTS;
4674
4675 /* We're playing with the host CPU state here, make sure we don't preempt or longjmp. */
4676 VMMRZCallRing3Disable(pVCpu);
4677 HM_DISABLE_PREEMPT();
4678
4679 /* Save the host & load the guest debug state, restart execution of the MOV DRx instruction. */
4680 CPUMR0LoadGuestDebugState(pVCpu, false /* include DR6 */);
4681 Assert(CPUMIsGuestDebugStateActive(pVCpu) || HC_ARCH_BITS == 32);
4682
4683 HM_RESTORE_PREEMPT();
4684 VMMRZCallRing3Enable(pVCpu);
4685
4686 STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxContextSwitch);
4687 return VINF_SUCCESS;
4688 }
4689
4690 /*
4691 * Interpret the read/writing of DRx.
4692 */
4693 /** @todo Decode assist. */
4694 VBOXSTRICTRC rc = EMInterpretInstruction(pVCpu, CPUMCTX2CORE(pCtx), 0 /* pvFault */);
4695 Log5(("hmR0SvmExitReadDRx: Emulated DRx access: rc=%Rrc\n", VBOXSTRICTRC_VAL(rc)));
4696 if (RT_LIKELY(rc == VINF_SUCCESS))
4697 {
4698 /* Not necessary for read accesses but whatever doesn't hurt for now, will be fixed with decode assist. */
4699 /** @todo CPUM should set this flag! */
4700 HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_DEBUG);
4701 HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
4702 }
4703 else
4704 Assert(rc == VERR_EM_INTERPRETER);
4705 return VBOXSTRICTRC_TODO(rc);
4706}
4707
4708
4709/**
4710 * #VMEXIT handler for DRx write (SVM_EXIT_WRITE_DRx). Conditional #VMEXIT.
4711 */
4712HMSVM_EXIT_DECL hmR0SvmExitWriteDRx(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4713{
4714 HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4715 /* For now it's the same since we interpret the instruction anyway. Will change when using of Decode Assist is implemented. */
4716 int rc = hmR0SvmExitReadDRx(pVCpu, pCtx, pSvmTransient);
4717 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitDRxWrite);
4718 STAM_COUNTER_DEC(&pVCpu->hm.s.StatExitDRxRead);
4719 return rc;
4720}
4721
4722
4723/**
4724 * #VMEXIT handler for XCRx write (SVM_EXIT_XSETBV). Conditional #VMEXIT.
4725 */
4726HMSVM_EXIT_DECL hmR0SvmExitXsetbv(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4727{
4728 HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4729
4730 /** @todo decode assists... */
4731 VBOXSTRICTRC rcStrict = IEMExecOne(pVCpu);
4732 if (rcStrict == VINF_IEM_RAISED_XCPT)
4733 HMCPU_CF_SET(pVCpu, HM_CHANGED_ALL_GUEST);
4734
4735 pVCpu->hm.s.fLoadSaveGuestXcr0 = (pCtx->cr4 & X86_CR4_OSXSAVE) && pCtx->aXcr[0] != ASMGetXcr0();
4736
4737 HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
4738 return VBOXSTRICTRC_TODO(rcStrict);
4739}
4740
4741
4742/**
4743 * #VMEXIT handler for I/O instructions (SVM_EXIT_IOIO). Conditional #VMEXIT.
4744 */
4745HMSVM_EXIT_DECL hmR0SvmExitIOInstr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4746{
4747 HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4748
4749 /* I/O operation lookup arrays. */
4750 static uint32_t const s_aIOSize[8] = { 0, 1, 2, 0, 4, 0, 0, 0 }; /* Size of the I/O accesses in bytes. */
4751 static uint32_t const s_aIOOpAnd[8] = { 0, 0xff, 0xffff, 0, 0xffffffff, 0, 0, 0 }; /* AND masks for saving
4752 the result (in AL/AX/EAX). */
4753 Log4(("hmR0SvmExitIOInstr: CS:RIP=%04x:%#RX64\n", pCtx->cs.Sel, pCtx->rip));
4754
4755 PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4756 PVM pVM = pVCpu->CTX_SUFF(pVM);
4757
4758 /* Refer AMD spec. 15.10.2 "IN and OUT Behaviour" and Figure 15-2. "EXITINFO1 for IOIO Intercept" for the format. */
4759 SVMIOIOEXIT IoExitInfo;
4760 IoExitInfo.u = (uint32_t)pVmcb->ctrl.u64ExitInfo1;
4761 uint32_t uIOWidth = (IoExitInfo.u >> 4) & 0x7;
4762 uint32_t cbValue = s_aIOSize[uIOWidth];
4763 uint32_t uAndVal = s_aIOOpAnd[uIOWidth];
4764
4765 if (RT_UNLIKELY(!cbValue))
4766 {
4767 AssertMsgFailed(("hmR0SvmExitIOInstr: Invalid IO operation. uIOWidth=%u\n", uIOWidth));
4768 return VERR_EM_INTERPRETER;
4769 }
4770
4771 VBOXSTRICTRC rcStrict;
4772 if (IoExitInfo.n.u1STR)
4773 {
4774 /* INS/OUTS - I/O String instruction. */
4775 PDISCPUSTATE pDis = &pVCpu->hm.s.DisState;
4776
4777 /** @todo Huh? why can't we use the segment prefix information given by AMD-V
4778 * in EXITINFO1? Investigate once this thing is up and running. */
4779
4780 rcStrict = EMInterpretDisasCurrent(pVM, pVCpu, pDis, NULL);
4781 if (rcStrict == VINF_SUCCESS)
4782 {
4783 if (IoExitInfo.n.u1Type == SVM_IOIO_WRITE)
4784 {
4785 rcStrict = IOMInterpretOUTSEx(pVM, pVCpu, CPUMCTX2CORE(pCtx), IoExitInfo.n.u16Port, pDis->fPrefix,
4786 (DISCPUMODE)pDis->uAddrMode, cbValue);
4787 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOStringWrite);
4788 }
4789 else
4790 {
4791 rcStrict = IOMInterpretINSEx(pVM, pVCpu, CPUMCTX2CORE(pCtx), IoExitInfo.n.u16Port, pDis->fPrefix,
4792 (DISCPUMODE)pDis->uAddrMode, cbValue);
4793 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOStringRead);
4794 }
4795 }
4796 else
4797 rcStrict = VINF_EM_RAW_EMULATE_INSTR;
4798 }
4799 else
4800 {
4801 /* IN/OUT - I/O instruction. */
4802 Assert(!IoExitInfo.n.u1REP);
4803
4804 if (IoExitInfo.n.u1Type == SVM_IOIO_WRITE)
4805 {
4806 rcStrict = IOMIOPortWrite(pVM, pVCpu, IoExitInfo.n.u16Port, pCtx->eax & uAndVal, cbValue);
4807 if (rcStrict == VINF_IOM_R3_IOPORT_WRITE)
4808 HMR0SavePendingIOPortWrite(pVCpu, pCtx->rip, pVmcb->ctrl.u64ExitInfo2, IoExitInfo.n.u16Port, uAndVal, cbValue);
4809
4810 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIOWrite);
4811 }
4812 else
4813 {
4814 uint32_t u32Val = 0;
4815
4816 rcStrict = IOMIOPortRead(pVM, pVCpu, IoExitInfo.n.u16Port, &u32Val, cbValue);
4817 if (IOM_SUCCESS(rcStrict))
4818 {
4819 /* Save result of I/O IN instr. in AL/AX/EAX. */
4820 pCtx->eax = (pCtx->eax & ~uAndVal) | (u32Val & uAndVal);
4821 }
4822 else if (rcStrict == VINF_IOM_R3_IOPORT_READ)
4823 HMR0SavePendingIOPortRead(pVCpu, pCtx->rip, pVmcb->ctrl.u64ExitInfo2, IoExitInfo.n.u16Port, uAndVal, cbValue);
4824
4825 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIORead);
4826 }
4827 }
4828
4829 if (IOM_SUCCESS(rcStrict))
4830 {
4831 /* AMD-V saves the RIP of the instruction following the IO instruction in EXITINFO2. */
4832 pCtx->rip = pVmcb->ctrl.u64ExitInfo2;
4833
4834 /*
4835 * If any I/O breakpoints are armed, we need to check if one triggered
4836 * and take appropriate action.
4837 * Note that the I/O breakpoint type is undefined if CR4.DE is 0.
4838 */
4839 /** @todo Optimize away the DBGFBpIsHwIoArmed call by having DBGF tell the
4840 * execution engines about whether hyper BPs and such are pending. */
4841 uint32_t const uDr7 = pCtx->dr[7];
4842 if (RT_UNLIKELY( ( (uDr7 & X86_DR7_ENABLED_MASK)
4843 && X86_DR7_ANY_RW_IO(uDr7)
4844 && (pCtx->cr4 & X86_CR4_DE))
4845 || DBGFBpIsHwIoArmed(pVM)))
4846 {
4847 /* We're playing with the host CPU state here, make sure we don't preempt or longjmp. */
4848 VMMRZCallRing3Disable(pVCpu);
4849 HM_DISABLE_PREEMPT();
4850
4851 STAM_COUNTER_INC(&pVCpu->hm.s.StatDRxIoCheck);
4852 CPUMR0DebugStateMaybeSaveGuest(pVCpu, false /*fDr6*/);
4853
4854 VBOXSTRICTRC rcStrict2 = DBGFBpCheckIo(pVM, pVCpu, pCtx, IoExitInfo.n.u16Port, cbValue);
4855 if (rcStrict2 == VINF_EM_RAW_GUEST_TRAP)
4856 {
4857 /* Raise #DB. */
4858 pVmcb->guest.u64DR6 = pCtx->dr[6];
4859 pVmcb->guest.u64DR7 = pCtx->dr[7];
4860 pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
4861 hmR0SvmSetPendingXcptDB(pVCpu);
4862 }
4863 /* rcStrict is VINF_SUCCESS or in [VINF_EM_FIRST..VINF_EM_LAST]. */
4864 else if ( rcStrict2 != VINF_SUCCESS
4865 && (rcStrict == VINF_SUCCESS || rcStrict2 < rcStrict))
4866 rcStrict = rcStrict2;
4867
4868 HM_RESTORE_PREEMPT();
4869 VMMRZCallRing3Enable(pVCpu);
4870 }
4871
4872 HMSVM_CHECK_SINGLE_STEP(pVCpu, rcStrict);
4873 }
4874
4875#ifdef VBOX_STRICT
4876 if (rcStrict == VINF_IOM_R3_IOPORT_READ)
4877 Assert(IoExitInfo.n.u1Type == SVM_IOIO_READ);
4878 else if (rcStrict == VINF_IOM_R3_IOPORT_WRITE)
4879 Assert(IoExitInfo.n.u1Type == SVM_IOIO_WRITE);
4880 else
4881 {
4882 /** @todo r=bird: This is missing a bunch of VINF_EM_FIRST..VINF_EM_LAST
4883 * statuses, that the VMM device and some others may return. See
4884 * IOM_SUCCESS() for guidance. */
4885 AssertMsg( RT_FAILURE(rcStrict)
4886 || rcStrict == VINF_SUCCESS
4887 || rcStrict == VINF_EM_RAW_EMULATE_INSTR
4888 || rcStrict == VINF_EM_DBG_BREAKPOINT
4889 || rcStrict == VINF_EM_RAW_GUEST_TRAP
4890 || rcStrict == VINF_TRPM_XCPT_DISPATCHED, ("%Rrc\n", VBOXSTRICTRC_VAL(rcStrict)));
4891 }
4892#endif
4893 return VBOXSTRICTRC_TODO(rcStrict);
4894}
4895
4896
4897/**
4898 * #VMEXIT handler for Nested Page-faults (SVM_EXIT_NPF). Conditional
4899 * #VMEXIT.
4900 */
4901HMSVM_EXIT_DECL hmR0SvmExitNestedPF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
4902{
4903 HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
4904 PVM pVM = pVCpu->CTX_SUFF(pVM);
4905 Assert(pVM->hm.s.fNestedPaging);
4906
4907 HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
4908
4909 /* See AMD spec. 15.25.6 "Nested versus Guest Page Faults, Fault Ordering" for VMCB details for #NPF. */
4910 PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
4911 uint32_t u32ErrCode = pVmcb->ctrl.u64ExitInfo1;
4912 RTGCPHYS GCPhysFaultAddr = pVmcb->ctrl.u64ExitInfo2;
4913
4914 Log4(("#NPF at CS:RIP=%04x:%#RX64 faultaddr=%RGp errcode=%#x \n", pCtx->cs.Sel, pCtx->rip, GCPhysFaultAddr, u32ErrCode));
4915
4916#ifdef VBOX_HM_WITH_GUEST_PATCHING
4917 /* TPR patching for 32-bit guests, using the reserved bit in the page tables for MMIO regions. */
4918 if ( pVM->hm.s.fTprPatchingAllowed
4919 && (GCPhysFaultAddr & PAGE_OFFSET_MASK) == 0x80 /* TPR offset. */
4920 && ( !(u32ErrCode & X86_TRAP_PF_P) /* Not present */
4921 || (u32ErrCode & (X86_TRAP_PF_P | X86_TRAP_PF_RSVD)) == (X86_TRAP_PF_P | X86_TRAP_PF_RSVD)) /* MMIO page. */
4922 && !CPUMIsGuestInLongModeEx(pCtx)
4923 && !CPUMGetGuestCPL(pVCpu)
4924 && pVM->hm.s.cPatches < RT_ELEMENTS(pVM->hm.s.aPatches))
4925 {
4926 RTGCPHYS GCPhysApicBase = pCtx->msrApicBase;
4927 GCPhysApicBase &= PAGE_BASE_GC_MASK;
4928
4929 if (GCPhysFaultAddr == GCPhysApicBase + 0x80)
4930 {
4931 /* Only attempt to patch the instruction once. */
4932 PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
4933 if (!pPatch)
4934 return VINF_EM_HM_PATCH_TPR_INSTR;
4935 }
4936 }
4937#endif
4938
4939 /*
4940 * Determine the nested paging mode.
4941 */
4942 PGMMODE enmNestedPagingMode;
4943#if HC_ARCH_BITS == 32
4944 if (CPUMIsGuestInLongModeEx(pCtx))
4945 enmNestedPagingMode = PGMMODE_AMD64_NX;
4946 else
4947#endif
4948 enmNestedPagingMode = PGMGetHostMode(pVM);
4949
4950 /*
4951 * MMIO optimization using the reserved (RSVD) bit in the guest page tables for MMIO pages.
4952 */
4953 int rc;
4954 Assert((u32ErrCode & (X86_TRAP_PF_RSVD | X86_TRAP_PF_P)) != X86_TRAP_PF_RSVD);
4955 if ((u32ErrCode & (X86_TRAP_PF_RSVD | X86_TRAP_PF_P)) == (X86_TRAP_PF_RSVD | X86_TRAP_PF_P))
4956 {
4957 VBOXSTRICTRC rc2 = PGMR0Trap0eHandlerNPMisconfig(pVM, pVCpu, enmNestedPagingMode, CPUMCTX2CORE(pCtx), GCPhysFaultAddr,
4958 u32ErrCode);
4959 rc = VBOXSTRICTRC_VAL(rc2);
4960
4961 /*
4962 * If we succeed, resume guest execution.
4963 * If we fail in interpreting the instruction because we couldn't get the guest physical address
4964 * of the page containing the instruction via the guest's page tables (we would invalidate the guest page
4965 * in the host TLB), resume execution which would cause a guest page fault to let the guest handle this
4966 * weird case. See @bugref{6043}.
4967 */
4968 if ( rc == VINF_SUCCESS
4969 || rc == VERR_PAGE_TABLE_NOT_PRESENT
4970 || rc == VERR_PAGE_NOT_PRESENT)
4971 {
4972 /* Successfully handled MMIO operation. */
4973 HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
4974 rc = VINF_SUCCESS;
4975 }
4976 return rc;
4977 }
4978
4979 TRPMAssertXcptPF(pVCpu, GCPhysFaultAddr, u32ErrCode);
4980 rc = PGMR0Trap0eHandlerNestedPaging(pVM, pVCpu, enmNestedPagingMode, u32ErrCode, CPUMCTX2CORE(pCtx), GCPhysFaultAddr);
4981 TRPMResetTrap(pVCpu);
4982
4983 Log4(("#NPF: PGMR0Trap0eHandlerNestedPaging returned %Rrc CS:RIP=%04x:%#RX64\n", rc, pCtx->cs.Sel, pCtx->rip));
4984
4985 /*
4986 * Same case as PGMR0Trap0eHandlerNPMisconfig(). See comment above, @bugref{6043}.
4987 */
4988 if ( rc == VINF_SUCCESS
4989 || rc == VERR_PAGE_TABLE_NOT_PRESENT
4990 || rc == VERR_PAGE_NOT_PRESENT)
4991 {
4992 /* We've successfully synced our shadow page tables. */
4993 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPF);
4994 rc = VINF_SUCCESS;
4995 }
4996
4997 return rc;
4998}
4999
5000
5001/**
5002 * #VMEXIT handler for virtual interrupt (SVM_EXIT_VINTR). Conditional #VMEXIT.
5003 */
5004HMSVM_EXIT_DECL hmR0SvmExitVIntr(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5005{
5006 HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5007
5008 PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
5009 pVmcb->ctrl.IntCtrl.n.u1VIrqValid = 0; /* No virtual interrupts pending, we'll inject the current one/NMI before reentry. */
5010 pVmcb->ctrl.IntCtrl.n.u8VIrqVector = 0;
5011
5012 /* Indicate that we no longer need to #VMEXIT when the guest is ready to receive interrupts/NMIs, it is now ready. */
5013 pVmcb->ctrl.u32InterceptCtrl1 &= ~SVM_CTRL1_INTERCEPT_VINTR;
5014 pVmcb->ctrl.u64VmcbCleanBits &= ~(HMSVM_VMCB_CLEAN_INTERCEPTS | HMSVM_VMCB_CLEAN_TPR);
5015
5016 /* Deliver the pending interrupt/NMI via hmR0SvmEvaluatePendingEvent() and resume guest execution. */
5017 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitIntWindow);
5018 return VINF_SUCCESS;
5019}
5020
5021
5022/**
5023 * #VMEXIT handler for task switches (SVM_EXIT_TASK_SWITCH). Conditional #VMEXIT.
5024 */
5025HMSVM_EXIT_DECL hmR0SvmExitTaskSwitch(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5026{
5027 HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5028
5029#ifndef HMSVM_ALWAYS_TRAP_TASK_SWITCH
5030 Assert(!pVCpu->CTX_SUFF(pVM)->hm.s.fNestedPaging);
5031#endif
5032
5033 /* Check if this task-switch occurred while delivery an event through the guest IDT. */
5034 PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
5035 if ( !(pVmcb->ctrl.u64ExitInfo2 & (SVM_EXIT2_TASK_SWITCH_IRET | SVM_EXIT2_TASK_SWITCH_JMP))
5036 && pVCpu->hm.s.Event.fPending) /** @todo fPending cannot be 'true', see hmR0SvmInjectPendingEvent(). See @bugref{7362}.*/
5037 {
5038 /*
5039 * AMD-V does not provide us with the original exception but we have it in u64IntInfo since we
5040 * injected the event during VM-entry.
5041 */
5042 Log4(("hmR0SvmExitTaskSwitch: TS occurred during event delivery.\n"));
5043 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitTaskSwitch);
5044 return VINF_EM_RAW_INJECT_TRPM_EVENT;
5045 }
5046
5047 /** @todo Emulate task switch someday, currently just going back to ring-3 for
5048 * emulation. */
5049 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitTaskSwitch);
5050 return VERR_EM_INTERPRETER;
5051}
5052
5053
5054/**
5055 * #VMEXIT handler for VMMCALL (SVM_EXIT_VMMCALL). Conditional #VMEXIT.
5056 */
5057HMSVM_EXIT_DECL hmR0SvmExitVmmCall(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5058{
5059 HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5060 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitVmcall);
5061
5062 /* First check if this is a patched VMMCALL for mov TPR */
5063 int rc = hmR0SvmEmulateMovTpr(pVCpu->CTX_SUFF(pVM), pVCpu, pCtx);
5064 if (rc == VINF_SUCCESS)
5065 {
5066 HMSVM_CHECK_SINGLE_STEP(pVCpu, rc);
5067 return VINF_SUCCESS;
5068 }
5069 else if (rc == VERR_NOT_FOUND)
5070 {
5071 if (pVCpu->hm.s.fHypercallsEnabled)
5072 {
5073 rc = GIMHypercall(pVCpu, pCtx);
5074 if (RT_SUCCESS(rc))
5075 {
5076 /* If the hypercall changes anything other than guest general-purpose registers,
5077 we would need to reload the guest changed bits here before VM-reentry. */
5078 hmR0SvmUpdateRip(pVCpu, pCtx, 3);
5079 return VINF_SUCCESS;
5080 }
5081 }
5082 }
5083
5084 hmR0SvmSetPendingXcptUD(pVCpu);
5085 return VINF_SUCCESS;
5086}
5087
5088
5089/**
5090 * #VMEXIT handler for IRET (SVM_EXIT_IRET). Conditional #VMEXIT.
5091 */
5092HMSVM_EXIT_DECL hmR0SvmExitIret(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5093{
5094 HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5095
5096 /* Clear NMI blocking. */
5097 VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_BLOCK_NMIS);
5098
5099 /* Indicate that we no longer need to #VMEXIT when the guest is ready to receive NMIs, it is now ready. */
5100 PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
5101 hmR0SvmClearIretIntercept(pVmcb);
5102
5103 /* Deliver the pending NMI via hmR0SvmEvaluatePendingEvent() and resume guest execution. */
5104 return VINF_SUCCESS;
5105}
5106
5107
5108/**
5109 * #VMEXIT handler for page-fault exceptions (SVM_EXIT_EXCEPTION_E). Conditional
5110 * #VMEXIT.
5111 */
5112HMSVM_EXIT_DECL hmR0SvmExitXcptPF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5113{
5114 HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5115
5116 HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
5117
5118 /* See AMD spec. 15.12.15 "#PF (Page Fault)". */
5119 PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
5120 uint32_t u32ErrCode = pVmcb->ctrl.u64ExitInfo1;
5121 RTGCUINTPTR uFaultAddress = pVmcb->ctrl.u64ExitInfo2;
5122 PVM pVM = pVCpu->CTX_SUFF(pVM);
5123
5124#if defined(HMSVM_ALWAYS_TRAP_ALL_XCPTS) || defined(HMSVM_ALWAYS_TRAP_PF)
5125 if (pVM->hm.s.fNestedPaging)
5126 {
5127 pVCpu->hm.s.Event.fPending = false; /* In case it's a contributory or vectoring #PF. */
5128 if (!pSvmTransient->fVectoringDoublePF)
5129 {
5130 /* A genuine guest #PF, reflect it to the guest. */
5131 hmR0SvmSetPendingXcptPF(pVCpu, pCtx, u32ErrCode, uFaultAddress);
5132 Log4(("#PF: Guest page fault at %04X:%RGv FaultAddr=%RGv ErrCode=%#x\n", pCtx->cs.Sel, (RTGCPTR)pCtx->rip,
5133 uFaultAddress, u32ErrCode));
5134 }
5135 else
5136 {
5137 /* A guest page-fault occurred during delivery of a page-fault. Inject #DF. */
5138 hmR0SvmSetPendingXcptDF(pVCpu);
5139 Log4(("Pending #DF due to vectoring #PF. NP\n"));
5140 }
5141 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestPF);
5142 return VINF_SUCCESS;
5143 }
5144#endif
5145
5146 Assert(!pVM->hm.s.fNestedPaging);
5147
5148#ifdef VBOX_HM_WITH_GUEST_PATCHING
5149 /* Shortcut for APIC TPR reads and writes; only applicable to 32-bit guests. */
5150 if ( pVM->hm.s.fTprPatchingAllowed
5151 && (uFaultAddress & 0xfff) == 0x80 /* TPR offset. */
5152 && !(u32ErrCode & X86_TRAP_PF_P) /* Not present. */
5153 && !CPUMIsGuestInLongModeEx(pCtx)
5154 && !CPUMGetGuestCPL(pVCpu)
5155 && pVM->hm.s.cPatches < RT_ELEMENTS(pVM->hm.s.aPatches))
5156 {
5157 RTGCPHYS GCPhysApicBase;
5158 GCPhysApicBase = pCtx->msrApicBase;
5159 GCPhysApicBase &= PAGE_BASE_GC_MASK;
5160
5161 /* Check if the page at the fault-address is the APIC base. */
5162 RTGCPHYS GCPhysPage;
5163 int rc2 = PGMGstGetPage(pVCpu, (RTGCPTR)uFaultAddress, NULL /* pfFlags */, &GCPhysPage);
5164 if ( rc2 == VINF_SUCCESS
5165 && GCPhysPage == GCPhysApicBase)
5166 {
5167 /* Only attempt to patch the instruction once. */
5168 PHMTPRPATCH pPatch = (PHMTPRPATCH)RTAvloU32Get(&pVM->hm.s.PatchTree, (AVLOU32KEY)pCtx->eip);
5169 if (!pPatch)
5170 return VINF_EM_HM_PATCH_TPR_INSTR;
5171 }
5172 }
5173#endif
5174
5175 Log4(("#PF: uFaultAddress=%#RX64 CS:RIP=%#04x:%#RX64 u32ErrCode %#RX32 cr3=%#RX64\n", uFaultAddress, pCtx->cs.Sel,
5176 pCtx->rip, u32ErrCode, pCtx->cr3));
5177
5178 /* If it's a vectoring #PF, emulate injecting the original event injection as PGMTrap0eHandler() is incapable
5179 of differentiating between instruction emulation and event injection that caused a #PF. See @bugref{6607}. */
5180 if (pSvmTransient->fVectoringPF)
5181 {
5182 Assert(pVCpu->hm.s.Event.fPending);
5183 return VINF_EM_RAW_INJECT_TRPM_EVENT;
5184 }
5185
5186 TRPMAssertXcptPF(pVCpu, uFaultAddress, u32ErrCode);
5187 int rc = PGMTrap0eHandler(pVCpu, u32ErrCode, CPUMCTX2CORE(pCtx), (RTGCPTR)uFaultAddress);
5188
5189 Log4(("#PF rc=%Rrc\n", rc));
5190
5191 if (rc == VINF_SUCCESS)
5192 {
5193 /* Successfully synced shadow pages tables or emulated an MMIO instruction. */
5194 TRPMResetTrap(pVCpu);
5195 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPF);
5196 HMCPU_CF_SET(pVCpu, HM_CHANGED_SVM_GUEST_APIC_STATE);
5197 return rc;
5198 }
5199 else if (rc == VINF_EM_RAW_GUEST_TRAP)
5200 {
5201 pVCpu->hm.s.Event.fPending = false; /* In case it's a contributory or vectoring #PF. */
5202
5203 if (!pSvmTransient->fVectoringDoublePF)
5204 {
5205 /* It's a guest page fault and needs to be reflected to the guest. */
5206 u32ErrCode = TRPMGetErrorCode(pVCpu); /* The error code might have been changed. */
5207 TRPMResetTrap(pVCpu);
5208 hmR0SvmSetPendingXcptPF(pVCpu, pCtx, u32ErrCode, uFaultAddress);
5209 }
5210 else
5211 {
5212 /* A guest page-fault occurred during delivery of a page-fault. Inject #DF. */
5213 TRPMResetTrap(pVCpu);
5214 hmR0SvmSetPendingXcptDF(pVCpu);
5215 Log4(("#PF: Pending #DF due to vectoring #PF\n"));
5216 }
5217
5218 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestPF);
5219 return VINF_SUCCESS;
5220 }
5221
5222 TRPMResetTrap(pVCpu);
5223 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowPFEM);
5224 return rc;
5225}
5226
5227
5228/**
5229 * #VMEXIT handler for device-not-available exceptions (SVM_EXIT_EXCEPTION_7).
5230 * Conditional #VMEXIT.
5231 */
5232HMSVM_EXIT_DECL hmR0SvmExitXcptNM(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5233{
5234 HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5235
5236 HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
5237
5238 /* We're playing with the host CPU state here, make sure we don't preempt or longjmp. */
5239 VMMRZCallRing3Disable(pVCpu);
5240 HM_DISABLE_PREEMPT();
5241
5242 int rc;
5243 /* If the guest FPU was active at the time of the #NM exit, then it's a guest fault. */
5244 if (pSvmTransient->fWasGuestFPUStateActive)
5245 {
5246 rc = VINF_EM_RAW_GUEST_TRAP;
5247 Assert(CPUMIsGuestFPUStateActive(pVCpu) || HMCPU_CF_IS_PENDING(pVCpu, HM_CHANGED_GUEST_CR0));
5248 }
5249 else
5250 {
5251#ifndef HMSVM_ALWAYS_TRAP_ALL_XCPTS
5252 Assert(!pSvmTransient->fWasGuestFPUStateActive);
5253#endif
5254 rc = CPUMR0Trap07Handler(pVCpu->CTX_SUFF(pVM), pVCpu, pCtx);
5255 Assert(rc == VINF_EM_RAW_GUEST_TRAP || (rc == VINF_SUCCESS && CPUMIsGuestFPUStateActive(pVCpu)));
5256 }
5257
5258 HM_RESTORE_PREEMPT();
5259 VMMRZCallRing3Enable(pVCpu);
5260
5261 if (rc == VINF_SUCCESS)
5262 {
5263 /* Guest FPU state was activated, we'll want to change CR0 FPU intercepts before the next VM-reentry. */
5264 HMCPU_CF_SET(pVCpu, HM_CHANGED_GUEST_CR0);
5265 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitShadowNM);
5266 pVCpu->hm.s.fPreloadGuestFpu = true;
5267 }
5268 else
5269 {
5270 /* Forward #NM to the guest. */
5271 Assert(rc == VINF_EM_RAW_GUEST_TRAP);
5272 hmR0SvmSetPendingXcptNM(pVCpu);
5273 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestNM);
5274 }
5275 return VINF_SUCCESS;
5276}
5277
5278
5279/**
5280 * #VMEXIT handler for undefined opcode (SVM_EXIT_EXCEPTION_6).
5281 * Conditional #VMEXIT.
5282 */
5283HMSVM_EXIT_DECL hmR0SvmExitXcptUD(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5284{
5285 HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5286
5287 HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
5288
5289 if (pVCpu->hm.s.fGIMTrapXcptUD)
5290 GIMXcptUD(pVCpu, pCtx, NULL /* pDis */);
5291 else
5292 hmR0SvmSetPendingXcptUD(pVCpu);
5293
5294 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestUD);
5295 return VINF_SUCCESS;
5296}
5297
5298
5299/**
5300 * #VMEXIT handler for math-fault exceptions (SVM_EXIT_EXCEPTION_10).
5301 * Conditional #VMEXIT.
5302 */
5303HMSVM_EXIT_DECL hmR0SvmExitXcptMF(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5304{
5305 HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5306
5307 HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
5308
5309 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestMF);
5310
5311 if (!(pCtx->cr0 & X86_CR0_NE))
5312 {
5313 PVM pVM = pVCpu->CTX_SUFF(pVM);
5314 PDISSTATE pDis = &pVCpu->hm.s.DisState;
5315 unsigned cbOp;
5316 int rc = EMInterpretDisasCurrent(pVM, pVCpu, pDis, &cbOp);
5317 if (RT_SUCCESS(rc))
5318 {
5319 /* Convert a #MF into a FERR -> IRQ 13. See @bugref{6117}. */
5320 rc = PDMIsaSetIrq(pVCpu->CTX_SUFF(pVM), 13, 1, 0 /* uTagSrc */);
5321 if (RT_SUCCESS(rc))
5322 pCtx->rip += cbOp;
5323 }
5324 else
5325 Log4(("hmR0SvmExitXcptMF: EMInterpretDisasCurrent returned %Rrc uOpCode=%#x\n", rc, pDis->pCurInstr->uOpcode));
5326 return rc;
5327 }
5328
5329 hmR0SvmSetPendingXcptMF(pVCpu);
5330 return VINF_SUCCESS;
5331}
5332
5333
5334/**
5335 * #VMEXIT handler for debug exceptions (SVM_EXIT_EXCEPTION_1). Conditional
5336 * #VMEXIT.
5337 */
5338HMSVM_EXIT_DECL hmR0SvmExitXcptDB(PVMCPU pVCpu, PCPUMCTX pCtx, PSVMTRANSIENT pSvmTransient)
5339{
5340 HMSVM_VALIDATE_EXIT_HANDLER_PARAMS();
5341
5342 HMSVM_CHECK_EXIT_DUE_TO_EVENT_DELIVERY();
5343
5344 STAM_COUNTER_INC(&pVCpu->hm.s.StatExitGuestDB);
5345
5346 /* This can be a fault-type #DB (instruction breakpoint) or a trap-type #DB (data breakpoint). However, for both cases
5347 DR6 and DR7 are updated to what the exception handler expects. See AMD spec. 15.12.2 "#DB (Debug)". */
5348 PSVMVMCB pVmcb = (PSVMVMCB)pVCpu->hm.s.svm.pvVmcb;
5349 PVM pVM = pVCpu->CTX_SUFF(pVM);
5350 int rc = DBGFRZTrap01Handler(pVM, pVCpu, CPUMCTX2CORE(pCtx), pVmcb->guest.u64DR6, pVCpu->hm.s.fSingleInstruction);
5351 if (rc == VINF_EM_RAW_GUEST_TRAP)
5352 {
5353 Log5(("hmR0SvmExitXcptDB: DR6=%#RX64 -> guest trap\n", pVmcb->guest.u64DR6));
5354 if (CPUMIsHyperDebugStateActive(pVCpu))
5355 CPUMSetGuestDR6(pVCpu, CPUMGetGuestDR6(pVCpu) | pVmcb->guest.u64DR6);
5356
5357 /* Reflect the exception back to the guest. */
5358 hmR0SvmSetPendingXcptDB(pVCpu);
5359 rc = VINF_SUCCESS;
5360 }
5361
5362 /*
5363 * Update DR6.
5364 */
5365 if (CPUMIsHyperDebugStateActive(pVCpu))
5366 {
5367 Log5(("hmR0SvmExitXcptDB: DR6=%#RX64 -> %Rrc\n", pVmcb->guest.u64DR6, rc));
5368 pVmcb->guest.u64DR6 = X86_DR6_INIT_VAL;
5369 pVmcb->ctrl.u64VmcbCleanBits &= ~HMSVM_VMCB_CLEAN_DRX;
5370 }
5371 else
5372 {
5373 AssertMsg(rc == VINF_SUCCESS, ("rc=%Rrc\n", rc));
5374 Assert(!pVCpu->hm.s.fSingleInstruction && !DBGFIsStepping(pVCpu));
5375 }
5376
5377 return rc;
5378}
5379
5380/** @} */
5381
Note: See TracBrowser for help on using the repository browser.

© 2024 Oracle Support Privacy / Do Not Sell My Info Terms of Use Trademark Policy Automated Access Etiquette