CPUMR0.cpp@ 82264

Last change on this file since 82264 was 81605, checked in by vboxsync, 5 years ago
VMM (and related changes): Add support for Hygon Dhyana CPUs. Modified and improved contribution by Hongyong Zang submitted under MIT license. Thank you!
Property svn:eol-style set to `native` Property svn:keywords set to `Id Revision`
File size: 35.1 KB

Line
1	/* $Id: CPUMR0.cpp 81605 2019-10-31 14:29:46Z vboxsync $ */
2	/** @file
3	* CPUM - Host Context Ring 0.
4	*/
5
6	/*
7	* Copyright (C) 2006-2019 Oracle Corporation
8	*
9	* This file is part of VirtualBox Open Source Edition (OSE), as
10	* available from http://www.virtualbox.org. This file is free software;
11	* you can redistribute it and/or modify it under the terms of the GNU
12	* General Public License (GPL) as published by the Free Software
13	* Foundation, in version 2 as it comes in the "COPYING" file of the
14	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16	*/
17
18
19	/*********************************************************************************************************************************
20	* Header Files *
21	*********************************************************************************************************************************/
22	#define LOG_GROUP LOG_GROUP_CPUM
23	#include <VBox/vmm/cpum.h>
24	#include "CPUMInternal.h"
25	#include <VBox/vmm/vmcc.h>
26	#include <VBox/vmm/gvm.h>
27	#include <VBox/err.h>
28	#include <VBox/log.h>
29	#include <VBox/vmm/hm.h>
30	#include <iprt/assert.h>
31	#include <iprt/asm-amd64-x86.h>
32	#ifdef VBOX_WITH_VMMR0_DISABLE_LAPIC_NMI
33	# include <iprt/mem.h>
34	# include <iprt/memobj.h>
35	# include <VBox/apic.h>
36	#endif
37	#include <iprt/x86.h>
38
39
40	/*********************************************************************************************************************************
41	* Structures and Typedefs *
42	*********************************************************************************************************************************/
43	#ifdef VBOX_WITH_VMMR0_DISABLE_LAPIC_NMI
44	/**
45	* Local APIC mappings.
46	*/
47	typedef struct CPUMHOSTLAPIC
48	{
49	/** Indicates that the entry is in use and have valid data. */
50	bool fEnabled;
51	/** Whether it's operating in X2APIC mode (EXTD). */
52	bool fX2Apic;
53	/** The APIC version number. */
54	uint32_t uVersion;
55	/** The physical address of the APIC registers. */
56	RTHCPHYS PhysBase;
57	/** The memory object entering the physical address. */
58	RTR0MEMOBJ hMemObj;
59	/** The mapping object for hMemObj. */
60	RTR0MEMOBJ hMapObj;
61	/** The mapping address APIC registers.
62	* @remarks Different CPUs may use the same physical address to map their
63	* APICs, so this pointer is only valid when on the CPU owning the
64	* APIC. */
65	void *pv;
66	} CPUMHOSTLAPIC;
67	#endif
68
69
70	/*********************************************************************************************************************************
71	* Global Variables *
72	*********************************************************************************************************************************/
73	#ifdef VBOX_WITH_VMMR0_DISABLE_LAPIC_NMI
74	static CPUMHOSTLAPIC g_aLApics[RTCPUSET_MAX_CPUS];
75	#endif
76
77	/**
78	* CPUID bits to unify among all cores.
79	*/
80	static struct
81	{
82	uint32_t uLeaf; /*< Leaf to check. /
83	uint32_t uEcx; /*< which bits in ecx to unify between CPUs. /
84	uint32_t uEdx; /*< which bits in edx to unify between CPUs. /
85	}
86	const g_aCpuidUnifyBits[] =
87	{
88	{
89	0x00000001,
90	X86_CPUID_FEATURE_ECX_CX16 \| X86_CPUID_FEATURE_ECX_MONITOR,
91	X86_CPUID_FEATURE_EDX_CX8
92	}
93	};
94
95
96
97	/*********************************************************************************************************************************
98	* Internal Functions *
99	*********************************************************************************************************************************/
100	#ifdef VBOX_WITH_VMMR0_DISABLE_LAPIC_NMI
101	static int cpumR0MapLocalApics(void);
102	static void cpumR0UnmapLocalApics(void);
103	#endif
104	static int cpumR0SaveHostDebugState(PVMCPUCC pVCpu);
105
106
107	/**
108	* Does the Ring-0 CPU initialization once during module load.
109	* XXX Host-CPU hot-plugging?
110	*/
111	VMMR0_INT_DECL(int) CPUMR0ModuleInit(void)
112	{
113	int rc = VINF_SUCCESS;
114	#ifdef VBOX_WITH_VMMR0_DISABLE_LAPIC_NMI
115	rc = cpumR0MapLocalApics();
116	#endif
117	return rc;
118	}
119
120
121	/**
122	* Terminate the module.
123	*/
124	VMMR0_INT_DECL(int) CPUMR0ModuleTerm(void)
125	{
126	#ifdef VBOX_WITH_VMMR0_DISABLE_LAPIC_NMI
127	cpumR0UnmapLocalApics();
128	#endif
129	return VINF_SUCCESS;
130	}
131
132
133	/**
134	* Check the CPUID features of this particular CPU and disable relevant features
135	* for the guest which do not exist on this CPU. We have seen systems where the
136	* X86_CPUID_FEATURE_ECX_MONITOR feature flag is only set on some host CPUs, see
137	* @bugref{5436}.
138	*
139	* @note This function might be called simultaneously on more than one CPU!
140	*
141	* @param idCpu The identifier for the CPU the function is called on.
142	* @param pvUser1 Pointer to the VM structure.
143	* @param pvUser2 Ignored.
144	*/
145	static DECLCALLBACK(void) cpumR0CheckCpuid(RTCPUID idCpu, void pvUser1, void pvUser2)
146	{
147	PVMCC pVM = (PVMCC)pvUser1;
148
149	NOREF(idCpu); NOREF(pvUser2);
150	for (uint32_t i = 0; i < RT_ELEMENTS(g_aCpuidUnifyBits); i++)
151	{
152	/* Note! Cannot use cpumCpuIdGetLeaf from here because we're not
153	necessarily in the VM process context. So, we using the
154	legacy arrays as temporary storage. */
155
156	uint32_t uLeaf = g_aCpuidUnifyBits[i].uLeaf;
157	PCPUMCPUID pLegacyLeaf;
158	if (uLeaf < RT_ELEMENTS(pVM->cpum.s.aGuestCpuIdPatmStd))
159	pLegacyLeaf = &pVM->cpum.s.aGuestCpuIdPatmStd[uLeaf];
160	else if (uLeaf - UINT32_C(0x80000000) < RT_ELEMENTS(pVM->cpum.s.aGuestCpuIdPatmExt))
161	pLegacyLeaf = &pVM->cpum.s.aGuestCpuIdPatmExt[uLeaf - UINT32_C(0x80000000)];
162	else if (uLeaf - UINT32_C(0xc0000000) < RT_ELEMENTS(pVM->cpum.s.aGuestCpuIdPatmCentaur))
163	pLegacyLeaf = &pVM->cpum.s.aGuestCpuIdPatmCentaur[uLeaf - UINT32_C(0xc0000000)];
164	else
165	continue;
166
167	uint32_t eax, ebx, ecx, edx;
168	ASMCpuIdExSlow(uLeaf, 0, 0, 0, &eax, &ebx, &ecx, &edx);
169
170	ASMAtomicAndU32(&pLegacyLeaf->uEcx, ecx \| ~g_aCpuidUnifyBits[i].uEcx);
171	ASMAtomicAndU32(&pLegacyLeaf->uEdx, edx \| ~g_aCpuidUnifyBits[i].uEdx);
172	}
173	}
174
175
176	/**
177	* Does Ring-0 CPUM initialization.
178	*
179	* This is mainly to check that the Host CPU mode is compatible
180	* with VBox.
181	*
182	* @returns VBox status code.
183	* @param pVM The cross context VM structure.
184	*/
185	VMMR0_INT_DECL(int) CPUMR0InitVM(PVMCC pVM)
186	{
187	LogFlow(("CPUMR0Init: %p\n", pVM));
188
189	/*
190	* Check CR0 & CR4 flags.
191	*/
192	uint32_t u32CR0 = ASMGetCR0();
193	if ((u32CR0 & (X86_CR0_PE \| X86_CR0_PG)) != (X86_CR0_PE \| X86_CR0_PG)) /* a bit paranoid perhaps.. */
194	{
195	Log(("CPUMR0Init: PE or PG not set. cr0=%#x\n", u32CR0));
196	return VERR_UNSUPPORTED_CPU_MODE;
197	}
198
199	/*
200	* Check for sysenter and syscall usage.
201	*/
202	if (ASMHasCpuId())
203	{
204	/*
205	* SYSENTER/SYSEXIT
206	*
207	* Intel docs claim you should test both the flag and family, model &
208	* stepping because some Pentium Pro CPUs have the SEP cpuid flag set,
209	* but don't support it. AMD CPUs may support this feature in legacy
210	* mode, they've banned it from long mode. Since we switch to 32-bit
211	* mode when entering raw-mode context the feature would become
212	* accessible again on AMD CPUs, so we have to check regardless of
213	* host bitness.
214	*/
215	uint32_t u32CpuVersion;
216	uint32_t u32Dummy;
217	uint32_t fFeatures; /* (Used further down to check for MSRs, so don't clobber.) */
218	ASMCpuId(1, &u32CpuVersion, &u32Dummy, &u32Dummy, &fFeatures);
219	uint32_t const u32Family = u32CpuVersion >> 8;
220	uint32_t const u32Model = (u32CpuVersion >> 4) & 0xF;
221	uint32_t const u32Stepping = u32CpuVersion & 0xF;
222	if ( (fFeatures & X86_CPUID_FEATURE_EDX_SEP)
223	&& ( u32Family != 6 /* (> pentium pro) */
224	\|\| u32Model >= 3
225	\|\| u32Stepping >= 3
226	\|\| !ASMIsIntelCpu())
227	)
228	{
229	/*
230	* Read the MSR and see if it's in use or not.
231	*/
232	uint32_t u32 = ASMRdMsr_Low(MSR_IA32_SYSENTER_CS);
233	if (u32)
234	{
235	pVM->cpum.s.fHostUseFlags \|= CPUM_USE_SYSENTER;
236	Log(("CPUMR0Init: host uses sysenter cs=%08x%08x\n", ASMRdMsr_High(MSR_IA32_SYSENTER_CS), u32));
237	}
238	}
239
240	/*
241	* SYSCALL/SYSRET
242	*
243	* This feature is indicated by the SEP bit returned in EDX by CPUID
244	* function 0x80000001. Intel CPUs only supports this feature in
245	* long mode. Since we're not running 64-bit guests in raw-mode there
246	* are no issues with 32-bit intel hosts.
247	*/
248	uint32_t cExt = 0;
249	ASMCpuId(0x80000000, &cExt, &u32Dummy, &u32Dummy, &u32Dummy);
250	if (ASMIsValidExtRange(cExt))
251	{
252	uint32_t fExtFeaturesEDX = ASMCpuId_EDX(0x80000001);
253	if (fExtFeaturesEDX & X86_CPUID_EXT_FEATURE_EDX_SYSCALL)
254	{
255	#ifdef RT_ARCH_X86
256	if (!ASMIsIntelCpu())
257	#endif
258	{
259	uint64_t fEfer = ASMRdMsr(MSR_K6_EFER);
260	if (fEfer & MSR_K6_EFER_SCE)
261	{
262	pVM->cpum.s.fHostUseFlags \|= CPUM_USE_SYSCALL;
263	Log(("CPUMR0Init: host uses syscall\n"));
264	}
265	}
266	}
267	}
268
269	/*
270	* Copy MSR_IA32_ARCH_CAPABILITIES bits over into the host and guest feature
271	* structure and as well as the guest MSR.
272	*/
273	pVM->cpum.s.HostFeatures.fArchRdclNo = 0;
274	pVM->cpum.s.HostFeatures.fArchIbrsAll = 0;
275	pVM->cpum.s.HostFeatures.fArchRsbOverride = 0;
276	pVM->cpum.s.HostFeatures.fArchVmmNeedNotFlushL1d = 0;
277	pVM->cpum.s.HostFeatures.fArchMdsNo = 0;
278	uint32_t const cStdRange = ASMCpuId_EAX(0);
279	if ( ASMIsValidStdRange(cStdRange)
280	&& cStdRange >= 7)
281	{
282	uint32_t fEdxFeatures = ASMCpuId_EDX(7);
283	if ( (fEdxFeatures & X86_CPUID_STEXT_FEATURE_EDX_ARCHCAP)
284	&& (fFeatures & X86_CPUID_FEATURE_EDX_MSR))
285	{
286	uint64_t const fArchVal = ASMRdMsr(MSR_IA32_ARCH_CAPABILITIES);
287	pVM->cpum.s.GuestFeatures.fArchRdclNo
288	= pVM->cpum.s.HostFeatures.fArchRdclNo = RT_BOOL(fArchVal & MSR_IA32_ARCH_CAP_F_RDCL_NO);
289	pVM->cpum.s.GuestFeatures.fArchIbrsAll
290	= pVM->cpum.s.HostFeatures.fArchIbrsAll = RT_BOOL(fArchVal & MSR_IA32_ARCH_CAP_F_IBRS_ALL);
291	pVM->cpum.s.GuestFeatures.fArchRsbOverride
292	= pVM->cpum.s.HostFeatures.fArchRsbOverride = RT_BOOL(fArchVal & MSR_IA32_ARCH_CAP_F_RSBO);
293	pVM->cpum.s.GuestFeatures.fArchVmmNeedNotFlushL1d
294	= pVM->cpum.s.HostFeatures.fArchVmmNeedNotFlushL1d = RT_BOOL(fArchVal & MSR_IA32_ARCH_CAP_F_VMM_NEED_NOT_FLUSH_L1D);
295	pVM->cpum.s.GuestFeatures.fArchMdsNo
296	= pVM->cpum.s.HostFeatures.fArchMdsNo = RT_BOOL(fArchVal & MSR_IA32_ARCH_CAP_F_MDS_NO);
297
298	if (pVM->cpum.s.GuestFeatures.fArchCap)
299	VMCC_FOR_EACH_VMCPU_STMT(pVM, pVCpu->cpum.s.GuestMsrs.msr.ArchCaps = fArchVal);
300	}
301	else
302	pVM->cpum.s.HostFeatures.fArchCap = 0;
303	}
304
305	/*
306	* Unify/cross check some CPUID feature bits on all available CPU cores
307	* and threads. We've seen CPUs where the monitor support differed.
308	*
309	* Because the hyper heap isn't always mapped into ring-0, we cannot
310	* access it from a RTMpOnAll callback. We use the legacy CPUID arrays
311	* as temp ring-0 accessible memory instead, ASSUMING that they're all
312	* up to date when we get here.
313	*/
314	RTMpOnAll(cpumR0CheckCpuid, pVM, NULL);
315
316	for (uint32_t i = 0; i < RT_ELEMENTS(g_aCpuidUnifyBits); i++)
317	{
318	bool fIgnored;
319	uint32_t uLeaf = g_aCpuidUnifyBits[i].uLeaf;
320	PCPUMCPUIDLEAF pLeaf = cpumCpuIdGetLeafEx(pVM, uLeaf, 0, &fIgnored);
321	if (pLeaf)
322	{
323	PCPUMCPUID pLegacyLeaf;
324	if (uLeaf < RT_ELEMENTS(pVM->cpum.s.aGuestCpuIdPatmStd))
325	pLegacyLeaf = &pVM->cpum.s.aGuestCpuIdPatmStd[uLeaf];
326	else if (uLeaf - UINT32_C(0x80000000) < RT_ELEMENTS(pVM->cpum.s.aGuestCpuIdPatmExt))
327	pLegacyLeaf = &pVM->cpum.s.aGuestCpuIdPatmExt[uLeaf - UINT32_C(0x80000000)];
328	else if (uLeaf - UINT32_C(0xc0000000) < RT_ELEMENTS(pVM->cpum.s.aGuestCpuIdPatmCentaur))
329	pLegacyLeaf = &pVM->cpum.s.aGuestCpuIdPatmCentaur[uLeaf - UINT32_C(0xc0000000)];
330	else
331	continue;
332
333	pLeaf->uEcx = pLegacyLeaf->uEcx;
334	pLeaf->uEdx = pLegacyLeaf->uEdx;
335	}
336	}
337
338	}
339
340
341	/*
342	* Check if debug registers are armed.
343	* This ASSUMES that DR7.GD is not set, or that it's handled transparently!
344	*/
345	uint32_t u32DR7 = ASMGetDR7();
346	if (u32DR7 & X86_DR7_ENABLED_MASK)
347	{
348	VMCC_FOR_EACH_VMCPU_STMT(pVM, pVCpu->cpum.s.fUseFlags \|= CPUM_USE_DEBUG_REGS_HOST);
349	Log(("CPUMR0Init: host uses debug registers (dr7=%x)\n", u32DR7));
350	}
351
352	return VINF_SUCCESS;
353	}
354
355
356	/**
357	* Trap handler for device-not-available fault (\#NM).
358	* Device not available, FP or (F)WAIT instruction.
359	*
360	* @returns VBox status code.
361	* @retval VINF_SUCCESS if the guest FPU state is loaded.
362	* @retval VINF_EM_RAW_GUEST_TRAP if it is a guest trap.
363	* @retval VINF_CPUM_HOST_CR0_MODIFIED if we modified the host CR0.
364	*
365	* @param pVM The cross context VM structure.
366	* @param pVCpu The cross context virtual CPU structure.
367	*/
368	VMMR0_INT_DECL(int) CPUMR0Trap07Handler(PVMCC pVM, PVMCPUCC pVCpu)
369	{
370	Assert(pVM->cpum.s.HostFeatures.fFxSaveRstor);
371	Assert(ASMGetCR4() & X86_CR4_OSFXSR);
372
373	/* If the FPU state has already been loaded, then it's a guest trap. */
374	if (CPUMIsGuestFPUStateActive(pVCpu))
375	{
376	Assert( ((pVCpu->cpum.s.Guest.cr0 & (X86_CR0_MP \| X86_CR0_EM \| X86_CR0_TS)) == (X86_CR0_MP \| X86_CR0_TS))
377	\|\| ((pVCpu->cpum.s.Guest.cr0 & (X86_CR0_MP \| X86_CR0_EM \| X86_CR0_TS)) == (X86_CR0_MP \| X86_CR0_TS \| X86_CR0_EM)));
378	return VINF_EM_RAW_GUEST_TRAP;
379	}
380
381	/*
382	* There are two basic actions:
383	* 1. Save host fpu and restore guest fpu.
384	* 2. Generate guest trap.
385	*
386	* When entering the hypervisor we'll always enable MP (for proper wait
387	* trapping) and TS (for intercepting all fpu/mmx/sse stuff). The EM flag
388	* is taken from the guest OS in order to get proper SSE handling.
389	*
390	*
391	* Actions taken depending on the guest CR0 flags:
392	*
393	* 3 2 1
394	* TS \| EM \| MP \| FPUInstr \| WAIT :: VMM Action
395	* ------------------------------------------------------------------------
396	* 0 \| 0 \| 0 \| Exec \| Exec :: Clear TS & MP, Save HC, Load GC.
397	* 0 \| 0 \| 1 \| Exec \| Exec :: Clear TS, Save HC, Load GC.
398	* 0 \| 1 \| 0 \| #NM \| Exec :: Clear TS & MP, Save HC, Load GC.
399	* 0 \| 1 \| 1 \| #NM \| Exec :: Clear TS, Save HC, Load GC.
400	* 1 \| 0 \| 0 \| #NM \| Exec :: Clear MP, Save HC, Load GC. (EM is already cleared.)
401	* 1 \| 0 \| 1 \| #NM \| #NM :: Go to guest taking trap there.
402	* 1 \| 1 \| 0 \| #NM \| Exec :: Clear MP, Save HC, Load GC. (EM is already set.)
403	* 1 \| 1 \| 1 \| #NM \| #NM :: Go to guest taking trap there.
404	*/
405
406	switch (pVCpu->cpum.s.Guest.cr0 & (X86_CR0_MP \| X86_CR0_EM \| X86_CR0_TS))
407	{
408	case X86_CR0_MP \| X86_CR0_TS:
409	case X86_CR0_MP \| X86_CR0_TS \| X86_CR0_EM:
410	return VINF_EM_RAW_GUEST_TRAP;
411	default:
412	break;
413	}
414
415	return CPUMR0LoadGuestFPU(pVM, pVCpu);
416	}
417
418
419	/**
420	* Saves the host-FPU/XMM state (if necessary) and (always) loads the guest-FPU
421	* state into the CPU.
422	*
423	* @returns VINF_SUCCESS on success, host CR0 unmodified.
424	* @returns VINF_CPUM_HOST_CR0_MODIFIED on success when the host CR0 was
425	* modified and VT-x needs to update the value in the VMCS.
426	*
427	* @param pVM The cross context VM structure.
428	* @param pVCpu The cross context virtual CPU structure.
429	*/
430	VMMR0_INT_DECL(int) CPUMR0LoadGuestFPU(PVMCC pVM, PVMCPUCC pVCpu)
431	{
432	int rc;
433	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
434	Assert(!(pVCpu->cpum.s.fUseFlags & CPUM_USED_FPU_GUEST));
435	Assert(!(pVCpu->cpum.s.fUseFlags & CPUM_SYNC_FPU_STATE));
436
437	if (!pVM->cpum.s.HostFeatures.fLeakyFxSR)
438	{
439	Assert(!(pVCpu->cpum.s.fUseFlags & CPUM_USED_MANUAL_XMM_RESTORE));
440	rc = cpumR0SaveHostRestoreGuestFPUState(&pVCpu->cpum.s);
441	}
442	else
443	{
444	Assert(!(pVCpu->cpum.s.fUseFlags & CPUM_USED_MANUAL_XMM_RESTORE) \|\| (pVCpu->cpum.s.fUseFlags & CPUM_USED_FPU_HOST));
445	/** @todo r=ramshankar: Can't we used a cached value here
446	* instead of reading the MSR? host EFER doesn't usually
447	* change. */
448	uint64_t uHostEfer = ASMRdMsr(MSR_K6_EFER);
449	if (!(uHostEfer & MSR_K6_EFER_FFXSR))
450	rc = cpumR0SaveHostRestoreGuestFPUState(&pVCpu->cpum.s);
451	else
452	{
453	RTCCUINTREG const uSavedFlags = ASMIntDisableFlags();
454	pVCpu->cpum.s.fUseFlags \|= CPUM_USED_MANUAL_XMM_RESTORE;
455	ASMWrMsr(MSR_K6_EFER, uHostEfer & ~MSR_K6_EFER_FFXSR);
456	rc = cpumR0SaveHostRestoreGuestFPUState(&pVCpu->cpum.s);
457	ASMWrMsr(MSR_K6_EFER, uHostEfer \| MSR_K6_EFER_FFXSR);
458	ASMSetFlags(uSavedFlags);
459	}
460	}
461	Assert( (pVCpu->cpum.s.fUseFlags & (CPUM_USED_FPU_GUEST \| CPUM_USED_FPU_HOST \| CPUM_USED_FPU_SINCE_REM))
462	== (CPUM_USED_FPU_GUEST \| CPUM_USED_FPU_HOST \| CPUM_USED_FPU_SINCE_REM));
463	return rc;
464	}
465
466
467	/**
468	* Saves the guest FPU/XMM state if needed, restores the host FPU/XMM state as
469	* needed.
470	*
471	* @returns true if we saved the guest state.
472	* @param pVCpu The cross context virtual CPU structure.
473	*/
474	VMMR0_INT_DECL(bool) CPUMR0FpuStateMaybeSaveGuestAndRestoreHost(PVMCPUCC pVCpu)
475	{
476	bool fSavedGuest;
477	Assert(pVCpu->CTX_SUFF(pVM)->cpum.s.HostFeatures.fFxSaveRstor);
478	Assert(ASMGetCR4() & X86_CR4_OSFXSR);
479	if (pVCpu->cpum.s.fUseFlags & (CPUM_USED_FPU_GUEST \| CPUM_USED_FPU_HOST))
480	{
481	fSavedGuest = RT_BOOL(pVCpu->cpum.s.fUseFlags & CPUM_USED_FPU_GUEST);
482	if (!(pVCpu->cpum.s.fUseFlags & CPUM_USED_MANUAL_XMM_RESTORE))
483	cpumR0SaveGuestRestoreHostFPUState(&pVCpu->cpum.s);
484	else
485	{
486	/* Temporarily clear MSR_K6_EFER_FFXSR or else we'll be unable to
487	save/restore the XMM state with fxsave/fxrstor. */
488	uint64_t uHostEfer = ASMRdMsr(MSR_K6_EFER);
489	if (uHostEfer & MSR_K6_EFER_FFXSR)
490	{
491	RTCCUINTREG const uSavedFlags = ASMIntDisableFlags();
492	ASMWrMsr(MSR_K6_EFER, uHostEfer & ~MSR_K6_EFER_FFXSR);
493	cpumR0SaveGuestRestoreHostFPUState(&pVCpu->cpum.s);
494	ASMWrMsr(MSR_K6_EFER, uHostEfer \| MSR_K6_EFER_FFXSR);
495	ASMSetFlags(uSavedFlags);
496	}
497	else
498	cpumR0SaveGuestRestoreHostFPUState(&pVCpu->cpum.s);
499	pVCpu->cpum.s.fUseFlags &= ~CPUM_USED_MANUAL_XMM_RESTORE;
500	}
501	}
502	else
503	fSavedGuest = false;
504	Assert(!( pVCpu->cpum.s.fUseFlags
505	& (CPUM_USED_FPU_GUEST \| CPUM_USED_FPU_HOST \| CPUM_SYNC_FPU_STATE \| CPUM_USED_MANUAL_XMM_RESTORE)));
506	return fSavedGuest;
507	}
508
509
510	/**
511	* Saves the host debug state, setting CPUM_USED_HOST_DEBUG_STATE and loading
512	* DR7 with safe values.
513	*
514	* @returns VBox status code.
515	* @param pVCpu The cross context virtual CPU structure.
516	*/
517	static int cpumR0SaveHostDebugState(PVMCPUCC pVCpu)
518	{
519	/*
520	* Save the host state.
521	*/
522	pVCpu->cpum.s.Host.dr0 = ASMGetDR0();
523	pVCpu->cpum.s.Host.dr1 = ASMGetDR1();
524	pVCpu->cpum.s.Host.dr2 = ASMGetDR2();
525	pVCpu->cpum.s.Host.dr3 = ASMGetDR3();
526	pVCpu->cpum.s.Host.dr6 = ASMGetDR6();
527	/** @todo dr7 might already have been changed to 0x400; don't care right now as it's harmless. */
528	pVCpu->cpum.s.Host.dr7 = ASMGetDR7();
529
530	/* Preemption paranoia. */
531	ASMAtomicOrU32(&pVCpu->cpum.s.fUseFlags, CPUM_USED_DEBUG_REGS_HOST);
532
533	/*
534	* Make sure DR7 is harmless or else we could trigger breakpoints when
535	* load guest or hypervisor DRx values later.
536	*/
537	if (pVCpu->cpum.s.Host.dr7 != X86_DR7_INIT_VAL)
538	ASMSetDR7(X86_DR7_INIT_VAL);
539
540	return VINF_SUCCESS;
541	}
542
543
544	/**
545	* Saves the guest DRx state residing in host registers and restore the host
546	* register values.
547	*
548	* The guest DRx state is only saved if CPUMR0LoadGuestDebugState was called,
549	* since it's assumed that we're shadowing the guest DRx register values
550	* accurately when using the combined hypervisor debug register values
551	* (CPUMR0LoadHyperDebugState).
552	*
553	* @returns true if either guest or hypervisor debug registers were loaded.
554	* @param pVCpu The cross context virtual CPU structure of the calling EMT.
555	* @param fDr6 Whether to include DR6 or not.
556	* @thread EMT(pVCpu)
557	*/
558	VMMR0_INT_DECL(bool) CPUMR0DebugStateMaybeSaveGuestAndRestoreHost(PVMCPUCC pVCpu, bool fDr6)
559	{
560	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
561	bool const fDrXLoaded = RT_BOOL(pVCpu->cpum.s.fUseFlags & (CPUM_USED_DEBUG_REGS_GUEST \| CPUM_USED_DEBUG_REGS_HYPER));
562
563	/*
564	* Do we need to save the guest DRx registered loaded into host registers?
565	* (DR7 and DR6 (if fDr6 is true) are left to the caller.)
566	*/
567	if (pVCpu->cpum.s.fUseFlags & CPUM_USED_DEBUG_REGS_GUEST)
568	{
569	pVCpu->cpum.s.Guest.dr[0] = ASMGetDR0();
570	pVCpu->cpum.s.Guest.dr[1] = ASMGetDR1();
571	pVCpu->cpum.s.Guest.dr[2] = ASMGetDR2();
572	pVCpu->cpum.s.Guest.dr[3] = ASMGetDR3();
573	if (fDr6)
574	pVCpu->cpum.s.Guest.dr[6] = ASMGetDR6();
575	}
576	ASMAtomicAndU32(&pVCpu->cpum.s.fUseFlags, ~( CPUM_USED_DEBUG_REGS_GUEST \| CPUM_USED_DEBUG_REGS_HYPER
577	\| CPUM_SYNC_DEBUG_REGS_GUEST \| CPUM_SYNC_DEBUG_REGS_HYPER));
578
579	/*
580	* Restore the host's debug state. DR0-3, DR6 and only then DR7!
581	*/
582	if (pVCpu->cpum.s.fUseFlags & CPUM_USED_DEBUG_REGS_HOST)
583	{
584	/* A bit of paranoia first... */
585	uint64_t uCurDR7 = ASMGetDR7();
586	if (uCurDR7 != X86_DR7_INIT_VAL)
587	ASMSetDR7(X86_DR7_INIT_VAL);
588
589	ASMSetDR0(pVCpu->cpum.s.Host.dr0);
590	ASMSetDR1(pVCpu->cpum.s.Host.dr1);
591	ASMSetDR2(pVCpu->cpum.s.Host.dr2);
592	ASMSetDR3(pVCpu->cpum.s.Host.dr3);
593	/** @todo consider only updating if they differ, esp. DR6. Need to figure how
594	* expensive DRx reads are over DRx writes. */
595	ASMSetDR6(pVCpu->cpum.s.Host.dr6);
596	ASMSetDR7(pVCpu->cpum.s.Host.dr7);
597
598	ASMAtomicAndU32(&pVCpu->cpum.s.fUseFlags, ~CPUM_USED_DEBUG_REGS_HOST);
599	}
600
601	return fDrXLoaded;
602	}
603
604
605	/**
606	* Saves the guest DRx state if it resides host registers.
607	*
608	* This does NOT clear any use flags, so the host registers remains loaded with
609	* the guest DRx state upon return. The purpose is only to make sure the values
610	* in the CPU context structure is up to date.
611	*
612	* @returns true if the host registers contains guest values, false if not.
613	* @param pVCpu The cross context virtual CPU structure of the calling EMT.
614	* @param fDr6 Whether to include DR6 or not.
615	* @thread EMT(pVCpu)
616	*/
617	VMMR0_INT_DECL(bool) CPUMR0DebugStateMaybeSaveGuest(PVMCPUCC pVCpu, bool fDr6)
618	{
619	/*
620	* Do we need to save the guest DRx registered loaded into host registers?
621	* (DR7 and DR6 (if fDr6 is true) are left to the caller.)
622	*/
623	if (pVCpu->cpum.s.fUseFlags & CPUM_USED_DEBUG_REGS_GUEST)
624	{
625	pVCpu->cpum.s.Guest.dr[0] = ASMGetDR0();
626	pVCpu->cpum.s.Guest.dr[1] = ASMGetDR1();
627	pVCpu->cpum.s.Guest.dr[2] = ASMGetDR2();
628	pVCpu->cpum.s.Guest.dr[3] = ASMGetDR3();
629	if (fDr6)
630	pVCpu->cpum.s.Guest.dr[6] = ASMGetDR6();
631	return true;
632	}
633	return false;
634	}
635
636
637	/**
638	* Lazily sync in the debug state.
639	*
640	* @param pVCpu The cross context virtual CPU structure of the calling EMT.
641	* @param fDr6 Whether to include DR6 or not.
642	* @thread EMT(pVCpu)
643	*/
644	VMMR0_INT_DECL(void) CPUMR0LoadGuestDebugState(PVMCPUCC pVCpu, bool fDr6)
645	{
646	/*
647	* Save the host state and disarm all host BPs.
648	*/
649	cpumR0SaveHostDebugState(pVCpu);
650	Assert(ASMGetDR7() == X86_DR7_INIT_VAL);
651
652	/*
653	* Activate the guest state DR0-3.
654	* DR7 and DR6 (if fDr6 is true) are left to the caller.
655	*/
656	ASMSetDR0(pVCpu->cpum.s.Guest.dr[0]);
657	ASMSetDR1(pVCpu->cpum.s.Guest.dr[1]);
658	ASMSetDR2(pVCpu->cpum.s.Guest.dr[2]);
659	ASMSetDR3(pVCpu->cpum.s.Guest.dr[3]);
660	if (fDr6)
661	ASMSetDR6(pVCpu->cpum.s.Guest.dr[6]);
662
663	ASMAtomicOrU32(&pVCpu->cpum.s.fUseFlags, CPUM_USED_DEBUG_REGS_GUEST);
664	}
665
666
667	/**
668	* Lazily sync in the hypervisor debug state
669	*
670	* @returns VBox status code.
671	* @param pVCpu The cross context virtual CPU structure of the calling EMT.
672	* @param fDr6 Whether to include DR6 or not.
673	* @thread EMT(pVCpu)
674	*/
675	VMMR0_INT_DECL(void) CPUMR0LoadHyperDebugState(PVMCPUCC pVCpu, bool fDr6)
676	{
677	/*
678	* Save the host state and disarm all host BPs.
679	*/
680	cpumR0SaveHostDebugState(pVCpu);
681	Assert(ASMGetDR7() == X86_DR7_INIT_VAL);
682
683	/*
684	* Make sure the hypervisor values are up to date.
685	*/
686	CPUMRecalcHyperDRx(pVCpu, UINT8_MAX /* no loading, please */, true);
687
688	/*
689	* Activate the guest state DR0-3.
690	* DR7 and DR6 (if fDr6 is true) are left to the caller.
691	*/
692	ASMSetDR0(pVCpu->cpum.s.Hyper.dr[0]);
693	ASMSetDR1(pVCpu->cpum.s.Hyper.dr[1]);
694	ASMSetDR2(pVCpu->cpum.s.Hyper.dr[2]);
695	ASMSetDR3(pVCpu->cpum.s.Hyper.dr[3]);
696	if (fDr6)
697	ASMSetDR6(X86_DR6_INIT_VAL);
698
699	ASMAtomicOrU32(&pVCpu->cpum.s.fUseFlags, CPUM_USED_DEBUG_REGS_HYPER);
700	}
701
702	#ifdef VBOX_WITH_VMMR0_DISABLE_LAPIC_NMI
703
704	/**
705	* Per-CPU callback that probes the CPU for APIC support.
706	*
707	* @param idCpu The identifier for the CPU the function is called on.
708	* @param pvUser1 Ignored.
709	* @param pvUser2 Ignored.
710	*/
711	static DECLCALLBACK(void) cpumR0MapLocalApicCpuProber(RTCPUID idCpu, void pvUser1, void pvUser2)
712	{
713	NOREF(pvUser1); NOREF(pvUser2);
714	int iCpu = RTMpCpuIdToSetIndex(idCpu);
715	AssertReturnVoid(iCpu >= 0 && (unsigned)iCpu < RT_ELEMENTS(g_aLApics));
716
717	/*
718	* Check for APIC support.
719	*/
720	uint32_t uMaxLeaf, u32EBX, u32ECX, u32EDX;
721	ASMCpuId(0, &uMaxLeaf, &u32EBX, &u32ECX, &u32EDX);
722	if ( ( ASMIsIntelCpuEx(u32EBX, u32ECX, u32EDX)
723	\|\| ASMIsAmdCpuEx(u32EBX, u32ECX, u32EDX)
724	\|\| ASMIsViaCentaurCpuEx(u32EBX, u32ECX, u32EDX)
725	\|\| ASMIsShanghaiCpuEx(u32EBX, u32ECX, u32EDX)
726	\|\| ASMIsHygonCpuEx(u32EBX, u32ECX, u32EDX))
727	&& ASMIsValidStdRange(uMaxLeaf))
728	{
729	uint32_t uDummy;
730	ASMCpuId(1, &uDummy, &u32EBX, &u32ECX, &u32EDX);
731	if ( (u32EDX & X86_CPUID_FEATURE_EDX_APIC)
732	&& (u32EDX & X86_CPUID_FEATURE_EDX_MSR))
733	{
734	/*
735	* Safe to access the MSR. Read it and calc the BASE (a little complicated).
736	*/
737	uint64_t u64ApicBase = ASMRdMsr(MSR_IA32_APICBASE);
738	uint64_t u64Mask = MSR_IA32_APICBASE_BASE_MIN;
739
740	/* see Intel Manual: Local APIC Status and Location: MAXPHYADDR default is bit 36 */
741	uint32_t uMaxExtLeaf;
742	ASMCpuId(0x80000000, &uMaxExtLeaf, &u32EBX, &u32ECX, &u32EDX);
743	if ( uMaxExtLeaf >= UINT32_C(0x80000008)
744	&& ASMIsValidExtRange(uMaxExtLeaf))
745	{
746	uint32_t u32PhysBits;
747	ASMCpuId(0x80000008, &u32PhysBits, &u32EBX, &u32ECX, &u32EDX);
748	u32PhysBits &= 0xff;
749	u64Mask = ((UINT64_C(1) << u32PhysBits) - 1) & UINT64_C(0xfffffffffffff000);
750	}
751
752	AssertCompile(sizeof(g_aLApics[iCpu].PhysBase) == sizeof(u64ApicBase));
753	g_aLApics[iCpu].PhysBase = u64ApicBase & u64Mask;
754	g_aLApics[iCpu].fEnabled = RT_BOOL(u64ApicBase & MSR_IA32_APICBASE_EN);
755	g_aLApics[iCpu].fX2Apic = (u64ApicBase & (MSR_IA32_APICBASE_EXTD \| MSR_IA32_APICBASE_EN))
756	== (MSR_IA32_APICBASE_EXTD \| MSR_IA32_APICBASE_EN);
757	}
758	}
759	}
760
761
762
763	/**
764	* Per-CPU callback that verifies our APIC expectations.
765	*
766	* @param idCpu The identifier for the CPU the function is called on.
767	* @param pvUser1 Ignored.
768	* @param pvUser2 Ignored.
769	*/
770	static DECLCALLBACK(void) cpumR0MapLocalApicCpuChecker(RTCPUID idCpu, void pvUser1, void pvUser2)
771	{
772	NOREF(pvUser1); NOREF(pvUser2);
773
774	int iCpu = RTMpCpuIdToSetIndex(idCpu);
775	AssertReturnVoid(iCpu >= 0 && (unsigned)iCpu < RT_ELEMENTS(g_aLApics));
776	if (!g_aLApics[iCpu].fEnabled)
777	return;
778
779	/*
780	* 0x0X 82489 external APIC
781	* 0x1X Local APIC
782	* 0x2X..0xFF reserved
783	*/
784	uint32_t uApicVersion;
785	if (g_aLApics[iCpu].fX2Apic)
786	uApicVersion = ApicX2RegRead32(APIC_REG_VERSION);
787	else
788	uApicVersion = ApicRegRead(g_aLApics[iCpu].pv, APIC_REG_VERSION);
789	if ((APIC_REG_VERSION_GET_VER(uApicVersion) & 0xF0) == 0x10)
790	{
791	g_aLApics[iCpu].uVersion = uApicVersion;
792
793	# if 0 /* enable if you need it. */
794	if (g_aLApics[iCpu].fX2Apic)
795	SUPR0Printf("CPUM: X2APIC %02u - ver %#010x, lint0=%#07x lint1=%#07x pc=%#07x thmr=%#07x cmci=%#07x\n",
796	iCpu, uApicVersion,
797	ApicX2RegRead32(APIC_REG_LVT_LINT0), ApicX2RegRead32(APIC_REG_LVT_LINT1),
798	ApicX2RegRead32(APIC_REG_LVT_PC), ApicX2RegRead32(APIC_REG_LVT_THMR),
799	ApicX2RegRead32(APIC_REG_LVT_CMCI));
800	else
801	{
802	SUPR0Printf("CPUM: APIC %02u at %RGp (mapped at %p) - ver %#010x, lint0=%#07x lint1=%#07x pc=%#07x thmr=%#07x cmci=%#07x\n",
803	iCpu, g_aLApics[iCpu].PhysBase, g_aLApics[iCpu].pv, uApicVersion,
804	ApicRegRead(g_aLApics[iCpu].pv, APIC_REG_LVT_LINT0), ApicRegRead(g_aLApics[iCpu].pv, APIC_REG_LVT_LINT1),
805	ApicRegRead(g_aLApics[iCpu].pv, APIC_REG_LVT_PC), ApicRegRead(g_aLApics[iCpu].pv, APIC_REG_LVT_THMR),
806	ApicRegRead(g_aLApics[iCpu].pv, APIC_REG_LVT_CMCI));
807	if (uApicVersion & 0x80000000)
808	{
809	uint32_t uExtFeatures = ApicRegRead(g_aLApics[iCpu].pv, 0x400);
810	uint32_t cEiLvt = (uExtFeatures >> 16) & 0xff;
811	SUPR0Printf("CPUM: APIC %02u: ExtSpace available. extfeat=%08x eilvt[0..3]=%08x %08x %08x %08x\n",
812	iCpu,
813	ApicRegRead(g_aLApics[iCpu].pv, 0x400),
814	cEiLvt >= 1 ? ApicRegRead(g_aLApics[iCpu].pv, 0x500) : 0,
815	cEiLvt >= 2 ? ApicRegRead(g_aLApics[iCpu].pv, 0x510) : 0,
816	cEiLvt >= 3 ? ApicRegRead(g_aLApics[iCpu].pv, 0x520) : 0,
817	cEiLvt >= 4 ? ApicRegRead(g_aLApics[iCpu].pv, 0x530) : 0);
818	}
819	}
820	# endif
821	}
822	else
823	{
824	g_aLApics[iCpu].fEnabled = false;
825	g_aLApics[iCpu].fX2Apic = false;
826	SUPR0Printf("VBox/CPUM: Unsupported APIC version %#x (iCpu=%d)\n", uApicVersion, iCpu);
827	}
828	}
829
830
831	/**
832	* Map the MMIO page of each local APIC in the system.
833	*/
834	static int cpumR0MapLocalApics(void)
835	{
836	/*
837	* Check that we'll always stay within the array bounds.
838	*/
839	if (RTMpGetArraySize() > RT_ELEMENTS(g_aLApics))
840	{
841	LogRel(("CPUM: Too many real CPUs/cores/threads - %u, max %u\n", RTMpGetArraySize(), RT_ELEMENTS(g_aLApics)));
842	return VERR_TOO_MANY_CPUS;
843	}
844
845	/*
846	* Create mappings for all online CPUs we think have legacy APICs.
847	*/
848	int rc = RTMpOnAll(cpumR0MapLocalApicCpuProber, NULL, NULL);
849
850	for (unsigned iCpu = 0; RT_SUCCESS(rc) && iCpu < RT_ELEMENTS(g_aLApics); iCpu++)
851	{
852	if (g_aLApics[iCpu].fEnabled && !g_aLApics[iCpu].fX2Apic)
853	{
854	rc = RTR0MemObjEnterPhys(&g_aLApics[iCpu].hMemObj, g_aLApics[iCpu].PhysBase,
855	PAGE_SIZE, RTMEM_CACHE_POLICY_MMIO);
856	if (RT_SUCCESS(rc))
857	{
858	rc = RTR0MemObjMapKernel(&g_aLApics[iCpu].hMapObj, g_aLApics[iCpu].hMemObj, (void *)-1,
859	PAGE_SIZE, RTMEM_PROT_READ \| RTMEM_PROT_WRITE);
860	if (RT_SUCCESS(rc))
861	{
862	g_aLApics[iCpu].pv = RTR0MemObjAddress(g_aLApics[iCpu].hMapObj);
863	continue;
864	}
865	RTR0MemObjFree(g_aLApics[iCpu].hMemObj, true /* fFreeMappings */);
866	}
867	g_aLApics[iCpu].fEnabled = false;
868	}
869	g_aLApics[iCpu].pv = NULL;
870	}
871
872	/*
873	* Check the APICs.
874	*/
875	if (RT_SUCCESS(rc))
876	rc = RTMpOnAll(cpumR0MapLocalApicCpuChecker, NULL, NULL);
877
878	if (RT_FAILURE(rc))
879	{
880	cpumR0UnmapLocalApics();
881	return rc;
882	}
883
884	# ifdef LOG_ENABLED
885	/*
886	* Log the result (pretty useless, requires enabling CPUM in VBoxDrv
887	* and !VBOX_WITH_R0_LOGGING).
888	*/
889	if (LogIsEnabled())
890	{
891	uint32_t cEnabled = 0;
892	uint32_t cX2Apics = 0;
893	for (unsigned iCpu = 0; iCpu < RT_ELEMENTS(g_aLApics); iCpu++)
894	if (g_aLApics[iCpu].fEnabled)
895	{
896	cEnabled++;
897	cX2Apics += g_aLApics[iCpu].fX2Apic;
898	}
899	Log(("CPUM: %u APICs, %u X2APICs\n", cEnabled, cX2Apics));
900	}
901	# endif
902
903	return VINF_SUCCESS;
904	}
905
906
907	/**
908	* Unmap the Local APIC of all host CPUs.
909	*/
910	static void cpumR0UnmapLocalApics(void)
911	{
912	for (unsigned iCpu = RT_ELEMENTS(g_aLApics); iCpu-- > 0;)
913	{
914	if (g_aLApics[iCpu].pv)
915	{
916	RTR0MemObjFree(g_aLApics[iCpu].hMapObj, true /* fFreeMappings */);
917	RTR0MemObjFree(g_aLApics[iCpu].hMemObj, true /* fFreeMappings */);
918	g_aLApics[iCpu].hMapObj = NIL_RTR0MEMOBJ;
919	g_aLApics[iCpu].hMemObj = NIL_RTR0MEMOBJ;
920	g_aLApics[iCpu].fEnabled = false;
921	g_aLApics[iCpu].fX2Apic = false;
922	g_aLApics[iCpu].pv = NULL;
923	}
924	}
925	}
926
927
928	/**
929	* Updates CPUMCPU::pvApicBase and CPUMCPU::fX2Apic prior to world switch.
930	*
931	* Writes the Local APIC mapping address of the current host CPU to CPUMCPU so
932	* the world switchers can access the APIC registers for the purpose of
933	* disabling and re-enabling the NMIs. Must be called with disabled preemption
934	* or disabled interrupts!
935	*
936	* @param pVCpu The cross context virtual CPU structure of the calling EMT.
937	* @param iHostCpuSet The CPU set index of the current host CPU.
938	*/
939	VMMR0_INT_DECL(void) CPUMR0SetLApic(PVMCPUCC pVCpu, uint32_t iHostCpuSet)
940	{
941	Assert(iHostCpuSet <= RT_ELEMENTS(g_aLApics));
942	pVCpu->cpum.s.pvApicBase = g_aLApics[iHostCpuSet].pv;
943	pVCpu->cpum.s.fX2Apic = g_aLApics[iHostCpuSet].fX2Apic;
944	// Log6(("CPUMR0SetLApic: pvApicBase=%p fX2Apic=%d\n", g_aLApics[idxCpu].pv, g_aLApics[idxCpu].fX2Apic));
945	}
946
947	#endif /* VBOX_WITH_VMMR0_DISABLE_LAPIC_NMI */
948

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

source: vbox/trunk/src/VBox/VMM/VMMR0/CPUMR0.cpp@ 82264

Download in other formats: