CPUMR0.cpp@ 58487

Last change on this file since 58487 was 58123, checked in by vboxsync, 9 years ago
VMM: Made @param pVCpu more uniform and to the point.
Property svn:eol-style set to `native` Property svn:keywords set to `Id Revision`
File size: 35.3 KB

Line
1	/* $Id: CPUMR0.cpp 58123 2015-10-08 18:09:45Z vboxsync $ */
2	/** @file
3	* CPUM - Host Context Ring 0.
4	*/
5
6	/*
7	* Copyright (C) 2006-2015 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	/*********************************************************************************************************************************
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/vm.h>
26	#include <VBox/err.h>
27	#include <VBox/log.h>
28	#include <VBox/vmm/hm.h>
29	#include <iprt/assert.h>
30	#include <iprt/asm-amd64-x86.h>
31	#ifdef VBOX_WITH_VMMR0_DISABLE_LAPIC_NMI
32	# include <iprt/mem.h>
33	# include <iprt/memobj.h>
34	# include <VBox/apic.h>
35	#endif
36	#include <iprt/x86.h>
37
38
39	/*********************************************************************************************************************************
40	* Structures and Typedefs *
41	*********************************************************************************************************************************/
42	#ifdef VBOX_WITH_VMMR0_DISABLE_LAPIC_NMI
43	/**
44	* Local APIC mappings.
45	*/
46	typedef struct CPUMHOSTLAPIC
47	{
48	/** Indicates that the entry is in use and have valid data. */
49	bool fEnabled;
50	/** Whether it's operating in X2APIC mode (EXTD). */
51	bool fX2Apic;
52	/** The APIC version number. */
53	uint32_t uVersion;
54	/** The physical address of the APIC registers. */
55	RTHCPHYS PhysBase;
56	/** The memory object entering the physical address. */
57	RTR0MEMOBJ hMemObj;
58	/** The mapping object for hMemObj. */
59	RTR0MEMOBJ hMapObj;
60	/** The mapping address APIC registers.
61	* @remarks Different CPUs may use the same physical address to map their
62	* APICs, so this pointer is only valid when on the CPU owning the
63	* APIC. */
64	void *pv;
65	} CPUMHOSTLAPIC;
66	#endif
67
68
69	/*********************************************************************************************************************************
70	* Global Variables *
71	*********************************************************************************************************************************/
72	#ifdef VBOX_WITH_VMMR0_DISABLE_LAPIC_NMI
73	static CPUMHOSTLAPIC g_aLApics[RTCPUSET_MAX_CPUS];
74	#endif
75
76	/**
77	* CPUID bits to unify among all cores.
78	*/
79	static struct
80	{
81	uint32_t uLeaf; /*< Leaf to check. /
82	uint32_t uEcx; /*< which bits in ecx to unify between CPUs. /
83	uint32_t uEdx; /*< which bits in edx to unify between CPUs. /
84	}
85	const g_aCpuidUnifyBits[] =
86	{
87	{
88	0x00000001,
89	X86_CPUID_FEATURE_ECX_CX16 \| X86_CPUID_FEATURE_ECX_MONITOR,
90	X86_CPUID_FEATURE_EDX_CX8
91	}
92	};
93
94
95
96	/*********************************************************************************************************************************
97	* Internal Functions *
98	*********************************************************************************************************************************/
99	#ifdef VBOX_WITH_VMMR0_DISABLE_LAPIC_NMI
100	static int cpumR0MapLocalApics(void);
101	static void cpumR0UnmapLocalApics(void);
102	#endif
103	static int cpumR0SaveHostDebugState(PVMCPU pVCpu);
104
105
106	/**
107	* Does the Ring-0 CPU initialization once during module load.
108	* XXX Host-CPU hot-plugging?
109	*/
110	VMMR0_INT_DECL(int) CPUMR0ModuleInit(void)
111	{
112	int rc = VINF_SUCCESS;
113	#ifdef VBOX_WITH_VMMR0_DISABLE_LAPIC_NMI
114	rc = cpumR0MapLocalApics();
115	#endif
116	return rc;
117	}
118
119
120	/**
121	* Terminate the module.
122	*/
123	VMMR0_INT_DECL(int) CPUMR0ModuleTerm(void)
124	{
125	#ifdef VBOX_WITH_VMMR0_DISABLE_LAPIC_NMI
126	cpumR0UnmapLocalApics();
127	#endif
128	return VINF_SUCCESS;
129	}
130
131
132	/**
133	*
134	*
135	* Check the CPUID features of this particular CPU and disable relevant features
136	* for the guest which do not exist on this CPU. We have seen systems where the
137	* X86_CPUID_FEATURE_ECX_MONITOR feature flag is only set on some host CPUs, see
138	* @bugref{5436}.
139	*
140	* @note This function might be called simultaneously on more than one CPU!
141	*
142	* @param idCpu The identifier for the CPU the function is called on.
143	* @param pvUser1 Pointer to the VM structure.
144	* @param pvUser2 Ignored.
145	*/
146	static DECLCALLBACK(void) cpumR0CheckCpuid(RTCPUID idCpu, void pvUser1, void pvUser2)
147	{
148	PVM pVM = (PVM)pvUser1;
149
150	NOREF(idCpu); NOREF(pvUser2);
151	for (uint32_t i = 0; i < RT_ELEMENTS(g_aCpuidUnifyBits); i++)
152	{
153	/* Note! Cannot use cpumCpuIdGetLeaf from here because we're not
154	necessarily in the VM process context. So, we using the
155	legacy arrays as temporary storage. */
156
157	uint32_t uLeaf = g_aCpuidUnifyBits[i].uLeaf;
158	PCPUMCPUID pLegacyLeaf;
159	if (uLeaf < RT_ELEMENTS(pVM->cpum.s.aGuestCpuIdPatmStd))
160	pLegacyLeaf = &pVM->cpum.s.aGuestCpuIdPatmStd[uLeaf];
161	else if (uLeaf - UINT32_C(0x80000000) < RT_ELEMENTS(pVM->cpum.s.aGuestCpuIdPatmExt))
162	pLegacyLeaf = &pVM->cpum.s.aGuestCpuIdPatmExt[uLeaf - UINT32_C(0x80000000)];
163	else if (uLeaf - UINT32_C(0xc0000000) < RT_ELEMENTS(pVM->cpum.s.aGuestCpuIdPatmCentaur))
164	pLegacyLeaf = &pVM->cpum.s.aGuestCpuIdPatmCentaur[uLeaf - UINT32_C(0xc0000000)];
165	else
166	continue;
167
168	uint32_t eax, ebx, ecx, edx;
169	ASMCpuIdExSlow(uLeaf, 0, 0, 0, &eax, &ebx, &ecx, &edx);
170
171	ASMAtomicAndU32(&pLegacyLeaf->uEcx, ecx \| ~g_aCpuidUnifyBits[i].uEcx);
172	ASMAtomicAndU32(&pLegacyLeaf->uEdx, edx \| ~g_aCpuidUnifyBits[i].uEdx);
173	}
174	}
175
176
177	/**
178	* Does Ring-0 CPUM initialization.
179	*
180	* This is mainly to check that the Host CPU mode is compatible
181	* with VBox.
182	*
183	* @returns VBox status code.
184	* @param pVM The cross context VM structure.
185	*/
186	VMMR0_INT_DECL(int) CPUMR0InitVM(PVM pVM)
187	{
188	LogFlow(("CPUMR0Init: %p\n", pVM));
189
190	/*
191	* Check CR0 & CR4 flags.
192	*/
193	uint32_t u32CR0 = ASMGetCR0();
194	if ((u32CR0 & (X86_CR0_PE \| X86_CR0_PG)) != (X86_CR0_PE \| X86_CR0_PG)) /* a bit paranoid perhaps.. */
195	{
196	Log(("CPUMR0Init: PE or PG not set. cr0=%#x\n", u32CR0));
197	return VERR_UNSUPPORTED_CPU_MODE;
198	}
199
200	/*
201	* Check for sysenter and syscall usage.
202	*/
203	if (ASMHasCpuId())
204	{
205	/*
206	* SYSENTER/SYSEXIT
207	*
208	* Intel docs claim you should test both the flag and family, model &
209	* stepping because some Pentium Pro CPUs have the SEP cpuid flag set,
210	* but don't support it. AMD CPUs may support this feature in legacy
211	* mode, they've banned it from long mode. Since we switch to 32-bit
212	* mode when entering raw-mode context the feature would become
213	* accessible again on AMD CPUs, so we have to check regardless of
214	* host bitness.
215	*/
216	uint32_t u32CpuVersion;
217	uint32_t u32Dummy;
218	uint32_t fFeatures;
219	ASMCpuId(1, &u32CpuVersion, &u32Dummy, &u32Dummy, &fFeatures);
220	uint32_t const u32Family = u32CpuVersion >> 8;
221	uint32_t const u32Model = (u32CpuVersion >> 4) & 0xF;
222	uint32_t const u32Stepping = u32CpuVersion & 0xF;
223	if ( (fFeatures & X86_CPUID_FEATURE_EDX_SEP)
224	&& ( u32Family != 6 /* (> pentium pro) */
225	\|\| u32Model >= 3
226	\|\| u32Stepping >= 3
227	\|\| !ASMIsIntelCpu())
228	)
229	{
230	/*
231	* Read the MSR and see if it's in use or not.
232	*/
233	uint32_t u32 = ASMRdMsr_Low(MSR_IA32_SYSENTER_CS);
234	if (u32)
235	{
236	pVM->cpum.s.fHostUseFlags \|= CPUM_USE_SYSENTER;
237	Log(("CPUMR0Init: host uses sysenter cs=%08x%08x\n", ASMRdMsr_High(MSR_IA32_SYSENTER_CS), u32));
238	}
239	}
240
241	/*
242	* SYSCALL/SYSRET
243	*
244	* This feature is indicated by the SEP bit returned in EDX by CPUID
245	* function 0x80000001. Intel CPUs only supports this feature in
246	* long mode. Since we're not running 64-bit guests in raw-mode there
247	* are no issues with 32-bit intel hosts.
248	*/
249	uint32_t cExt = 0;
250	ASMCpuId(0x80000000, &cExt, &u32Dummy, &u32Dummy, &u32Dummy);
251	if (ASMIsValidExtRange(cExt))
252	{
253	uint32_t fExtFeaturesEDX = ASMCpuId_EDX(0x80000001);
254	if (fExtFeaturesEDX & X86_CPUID_EXT_FEATURE_EDX_SYSCALL)
255	{
256	#ifdef RT_ARCH_X86
257	if (!ASMIsIntelCpu())
258	#endif
259	{
260	uint64_t fEfer = ASMRdMsr(MSR_K6_EFER);
261	if (fEfer & MSR_K6_EFER_SCE)
262	{
263	pVM->cpum.s.fHostUseFlags \|= CPUM_USE_SYSCALL;
264	Log(("CPUMR0Init: host uses syscall\n"));
265	}
266	}
267	}
268	}
269
270	/*
271	* Unify/cross check some CPUID feature bits on all available CPU cores
272	* and threads. We've seen CPUs where the monitor support differed.
273	*
274	* Because the hyper heap isn't always mapped into ring-0, we cannot
275	* access it from a RTMpOnAll callback. We use the legacy CPUID arrays
276	* as temp ring-0 accessible memory instead, ASSUMING that they're all
277	* up to date when we get here.
278	*/
279	RTMpOnAll(cpumR0CheckCpuid, pVM, NULL);
280
281	for (uint32_t i = 0; i < RT_ELEMENTS(g_aCpuidUnifyBits); i++)
282	{
283	bool fIgnored;
284	uint32_t uLeaf = g_aCpuidUnifyBits[i].uLeaf;
285	PCPUMCPUIDLEAF pLeaf = cpumCpuIdGetLeafEx(pVM, uLeaf, 0, &fIgnored);
286	if (pLeaf)
287	{
288	PCPUMCPUID pLegacyLeaf;
289	if (uLeaf < RT_ELEMENTS(pVM->cpum.s.aGuestCpuIdPatmStd))
290	pLegacyLeaf = &pVM->cpum.s.aGuestCpuIdPatmStd[uLeaf];
291	else if (uLeaf - UINT32_C(0x80000000) < RT_ELEMENTS(pVM->cpum.s.aGuestCpuIdPatmExt))
292	pLegacyLeaf = &pVM->cpum.s.aGuestCpuIdPatmExt[uLeaf - UINT32_C(0x80000000)];
293	else if (uLeaf - UINT32_C(0xc0000000) < RT_ELEMENTS(pVM->cpum.s.aGuestCpuIdPatmCentaur))
294	pLegacyLeaf = &pVM->cpum.s.aGuestCpuIdPatmCentaur[uLeaf - UINT32_C(0xc0000000)];
295	else
296	continue;
297
298	pLeaf->uEcx = pLegacyLeaf->uEcx;
299	pLeaf->uEdx = pLegacyLeaf->uEdx;
300	}
301	}
302
303	}
304
305
306	/*
307	* Check if debug registers are armed.
308	* This ASSUMES that DR7.GD is not set, or that it's handled transparently!
309	*/
310	uint32_t u32DR7 = ASMGetDR7();
311	if (u32DR7 & X86_DR7_ENABLED_MASK)
312	{
313	for (VMCPUID i = 0; i < pVM->cCpus; i++)
314	pVM->aCpus[i].cpum.s.fUseFlags \|= CPUM_USE_DEBUG_REGS_HOST;
315	Log(("CPUMR0Init: host uses debug registers (dr7=%x)\n", u32DR7));
316	}
317
318	return VINF_SUCCESS;
319	}
320
321
322	/**
323	* Trap handler for device-not-available fault (\#NM).
324	* Device not available, FP or (F)WAIT instruction.
325	*
326	* @returns VBox status code.
327	* @retval VINF_SUCCESS if the guest FPU state is loaded.
328	* @retval VINF_EM_RAW_GUEST_TRAP if it is a guest trap.
329	*
330	* @param pVM The cross context VM structure.
331	* @param pVCpu The cross context virtual CPU structure.
332	* @param pCtx Pointer to the guest-CPU context.
333	*/
334	VMMR0_INT_DECL(int) CPUMR0Trap07Handler(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
335	{
336	Assert(pVM->cpum.s.HostFeatures.fFxSaveRstor);
337	Assert(ASMGetCR4() & X86_CR4_OSFXSR);
338
339	/* If the FPU state has already been loaded, then it's a guest trap. */
340	if (CPUMIsGuestFPUStateActive(pVCpu))
341	{
342	Assert( ((pCtx->cr0 & (X86_CR0_MP \| X86_CR0_EM \| X86_CR0_TS)) == (X86_CR0_MP \| X86_CR0_TS))
343	\|\| ((pCtx->cr0 & (X86_CR0_MP \| X86_CR0_EM \| X86_CR0_TS)) == (X86_CR0_MP \| X86_CR0_TS \| X86_CR0_EM)));
344	return VINF_EM_RAW_GUEST_TRAP;
345	}
346
347	/*
348	* There are two basic actions:
349	* 1. Save host fpu and restore guest fpu.
350	* 2. Generate guest trap.
351	*
352	* When entering the hypervisor we'll always enable MP (for proper wait
353	* trapping) and TS (for intercepting all fpu/mmx/sse stuff). The EM flag
354	* is taken from the guest OS in order to get proper SSE handling.
355	*
356	*
357	* Actions taken depending on the guest CR0 flags:
358	*
359	* 3 2 1
360	* TS \| EM \| MP \| FPUInstr \| WAIT :: VMM Action
361	* ------------------------------------------------------------------------
362	* 0 \| 0 \| 0 \| Exec \| Exec :: Clear TS & MP, Save HC, Load GC.
363	* 0 \| 0 \| 1 \| Exec \| Exec :: Clear TS, Save HC, Load GC.
364	* 0 \| 1 \| 0 \| #NM \| Exec :: Clear TS & MP, Save HC, Load GC.
365	* 0 \| 1 \| 1 \| #NM \| Exec :: Clear TS, Save HC, Load GC.
366	* 1 \| 0 \| 0 \| #NM \| Exec :: Clear MP, Save HC, Load GC. (EM is already cleared.)
367	* 1 \| 0 \| 1 \| #NM \| #NM :: Go to guest taking trap there.
368	* 1 \| 1 \| 0 \| #NM \| Exec :: Clear MP, Save HC, Load GC. (EM is already set.)
369	* 1 \| 1 \| 1 \| #NM \| #NM :: Go to guest taking trap there.
370	*/
371
372	switch (pCtx->cr0 & (X86_CR0_MP \| X86_CR0_EM \| X86_CR0_TS))
373	{
374	case X86_CR0_MP \| X86_CR0_TS:
375	case X86_CR0_MP \| X86_CR0_TS \| X86_CR0_EM:
376	return VINF_EM_RAW_GUEST_TRAP;
377	default:
378	break;
379	}
380
381	return CPUMR0LoadGuestFPU(pVM, pVCpu, pCtx);
382	}
383
384
385	/**
386	* Saves the host-FPU/XMM state and loads the guest-FPU state into the CPU.
387	*
388	* @returns VBox status code.
389	*
390	* @param pVM The cross context VM structure.
391	* @param pVCpu The cross context virtual CPU structure.
392	* @param pCtx Pointer to the guest-CPU context.
393	*/
394	VMMR0_INT_DECL(int) CPUMR0LoadGuestFPU(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
395	{
396	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
397	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS)
398	if (CPUMIsGuestInLongModeEx(pCtx))
399	{
400	Assert(!(pVCpu->cpum.s.fUseFlags & CPUM_SYNC_FPU_STATE));
401
402	/* Save the host state and record the fact (CPUM_USED_FPU \| CPUM_USED_FPU_SINCE_REM). */
403	cpumR0SaveHostFPUState(&pVCpu->cpum.s);
404
405	/* Restore the state on entry as we need to be in 64-bit mode to access the full state. */
406	pVCpu->cpum.s.fUseFlags \|= CPUM_SYNC_FPU_STATE;
407	}
408	else
409	#endif
410	{
411	NOREF(pCtx);
412	Assert(!(pVCpu->cpum.s.fUseFlags & CPUM_USED_MANUAL_XMM_RESTORE));
413	/** @todo Move the FFXR handling down into
414	* cpumR0SaveHostRestoreGuestFPUState to optimize the
415	* VBOX_WITH_KERNEL_USING_XMM handling. */
416	/* Clear MSR_K6_EFER_FFXSR or else we'll be unable to save/restore the XMM state with fxsave/fxrstor. */
417	uint64_t uHostEfer = 0;
418	bool fRestoreEfer = false;
419	if (pVM->cpum.s.HostFeatures.fLeakyFxSR)
420	{
421	/** @todo r=ramshankar: Can't we used a cached value here
422	* instead of reading the MSR? host EFER doesn't usually
423	* change. */
424	uHostEfer = ASMRdMsr(MSR_K6_EFER);
425	if (uHostEfer & MSR_K6_EFER_FFXSR)
426	{
427	ASMWrMsr(MSR_K6_EFER, uHostEfer & ~MSR_K6_EFER_FFXSR);
428	pVCpu->cpum.s.fUseFlags \|= CPUM_USED_MANUAL_XMM_RESTORE;
429	fRestoreEfer = true;
430	}
431	}
432
433	/* Do the job and record that we've switched FPU state. */
434	cpumR0SaveHostRestoreGuestFPUState(&pVCpu->cpum.s);
435
436	/* Restore EFER. */
437	if (fRestoreEfer)
438	ASMWrMsr(MSR_K6_EFER, uHostEfer);
439	}
440
441	Assert((pVCpu->cpum.s.fUseFlags & (CPUM_USED_FPU \| CPUM_USED_FPU_SINCE_REM)) == (CPUM_USED_FPU \| CPUM_USED_FPU_SINCE_REM));
442	return VINF_SUCCESS;
443	}
444
445
446	/**
447	* Save guest FPU/XMM state
448	*
449	* @returns VBox status code.
450	* @param pVM The cross context VM structure.
451	* @param pVCpu The cross context virtual CPU structure.
452	* @param pCtx Pointer to the guest CPU context.
453	*/
454	VMMR0_INT_DECL(int) CPUMR0SaveGuestFPU(PVM pVM, PVMCPU pVCpu, PCPUMCTX pCtx)
455	{
456	Assert(pVM->cpum.s.HostFeatures.fFxSaveRstor);
457	Assert(ASMGetCR4() & X86_CR4_OSFXSR);
458	AssertReturn((pVCpu->cpum.s.fUseFlags & CPUM_USED_FPU), VINF_SUCCESS);
459	NOREF(pVM); NOREF(pCtx);
460
461	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS)
462	if (CPUMIsGuestInLongModeEx(pCtx))
463	{
464	if (!(pVCpu->cpum.s.fUseFlags & CPUM_SYNC_FPU_STATE))
465	{
466	HMR0SaveFPUState(pVM, pVCpu, pCtx);
467	cpumR0RestoreHostFPUState(&pVCpu->cpum.s);
468	}
469	/* else nothing to do; we didn't perform a world switch */
470	}
471	else
472	#endif
473	{
474	#ifdef VBOX_WITH_KERNEL_USING_XMM
475	/*
476	* We've already saved the XMM registers in the assembly wrapper, so
477	* we have to save them before saving the entire FPU state and put them
478	* back afterwards.
479	*/
480	/** @todo This could be skipped if MSR_K6_EFER_FFXSR is set, but
481	* I'm not able to test such an optimization tonight.
482	* We could just all this in assembly. */
483	uint128_t aGuestXmmRegs[16];
484	memcpy(&aGuestXmmRegs[0], &pVCpu->cpum.s.Guest.CTX_SUFF(pXState)->x87.aXMM[0], sizeof(aGuestXmmRegs));
485	#endif
486
487	/* Clear MSR_K6_EFER_FFXSR or else we'll be unable to save/restore the XMM state with fxsave/fxrstor. */
488	uint64_t uHostEfer = 0;
489	bool fRestoreEfer = false;
490	if (pVCpu->cpum.s.fUseFlags & CPUM_USED_MANUAL_XMM_RESTORE)
491	{
492	uHostEfer = ASMRdMsr(MSR_K6_EFER);
493	if (uHostEfer & MSR_K6_EFER_FFXSR)
494	{
495	ASMWrMsr(MSR_K6_EFER, uHostEfer & ~MSR_K6_EFER_FFXSR);
496	fRestoreEfer = true;
497	}
498	}
499
500	cpumR0SaveGuestRestoreHostFPUState(&pVCpu->cpum.s);
501
502	/* Restore EFER MSR */
503	if (fRestoreEfer)
504	ASMWrMsr(MSR_K6_EFER, uHostEfer \| MSR_K6_EFER_FFXSR);
505
506	#ifdef VBOX_WITH_KERNEL_USING_XMM
507	memcpy(&pVCpu->cpum.s.Guest.CTX_SUFF(pXState)->x87.aXMM[0], &aGuestXmmRegs[0], sizeof(aGuestXmmRegs));
508	#endif
509	}
510
511	pVCpu->cpum.s.fUseFlags &= ~(CPUM_USED_FPU \| CPUM_SYNC_FPU_STATE \| CPUM_USED_MANUAL_XMM_RESTORE);
512	return VINF_SUCCESS;
513	}
514
515
516	/**
517	* Saves the host debug state, setting CPUM_USED_HOST_DEBUG_STATE and loading
518	* DR7 with safe values.
519	*
520	* @returns VBox status code.
521	* @param pVCpu The cross context virtual CPU structure.
522	*/
523	static int cpumR0SaveHostDebugState(PVMCPU pVCpu)
524	{
525	/*
526	* Save the host state.
527	*/
528	pVCpu->cpum.s.Host.dr0 = ASMGetDR0();
529	pVCpu->cpum.s.Host.dr1 = ASMGetDR1();
530	pVCpu->cpum.s.Host.dr2 = ASMGetDR2();
531	pVCpu->cpum.s.Host.dr3 = ASMGetDR3();
532	pVCpu->cpum.s.Host.dr6 = ASMGetDR6();
533	/** @todo dr7 might already have been changed to 0x400; don't care right now as it's harmless. */
534	pVCpu->cpum.s.Host.dr7 = ASMGetDR7();
535
536	/* Preemption paranoia. */
537	ASMAtomicOrU32(&pVCpu->cpum.s.fUseFlags, CPUM_USED_DEBUG_REGS_HOST);
538
539	/*
540	* Make sure DR7 is harmless or else we could trigger breakpoints when
541	* load guest or hypervisor DRx values later.
542	*/
543	if (pVCpu->cpum.s.Host.dr7 != X86_DR7_INIT_VAL)
544	ASMSetDR7(X86_DR7_INIT_VAL);
545
546	return VINF_SUCCESS;
547	}
548
549
550	/**
551	* Saves the guest DRx state residing in host registers and restore the host
552	* register values.
553	*
554	* The guest DRx state is only saved if CPUMR0LoadGuestDebugState was called,
555	* since it's assumed that we're shadowing the guest DRx register values
556	* accurately when using the combined hypervisor debug register values
557	* (CPUMR0LoadHyperDebugState).
558	*
559	* @returns true if either guest or hypervisor debug registers were loaded.
560	* @param pVCpu The cross context virtual CPU structure of the calling EMT.
561	* @param fDr6 Whether to include DR6 or not.
562	* @thread EMT(pVCpu)
563	*/
564	VMMR0_INT_DECL(bool) CPUMR0DebugStateMaybeSaveGuestAndRestoreHost(PVMCPU pVCpu, bool fDr6)
565	{
566	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
567	bool const fDrXLoaded = RT_BOOL(pVCpu->cpum.s.fUseFlags & (CPUM_USED_DEBUG_REGS_GUEST \| CPUM_USED_DEBUG_REGS_HYPER));
568
569	/*
570	* Do we need to save the guest DRx registered loaded into host registers?
571	* (DR7 and DR6 (if fDr6 is true) are left to the caller.)
572	*/
573	if (pVCpu->cpum.s.fUseFlags & CPUM_USED_DEBUG_REGS_GUEST)
574	{
575	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS)
576	if (CPUMIsGuestInLongModeEx(&pVCpu->cpum.s.Guest))
577	{
578	uint64_t uDr6 = pVCpu->cpum.s.Guest.dr[6];
579	HMR0SaveDebugState(pVCpu->CTX_SUFF(pVM), pVCpu, &pVCpu->cpum.s.Guest);
580	if (!fDr6)
581	pVCpu->cpum.s.Guest.dr[6] = uDr6;
582	}
583	else
584	#endif
585	{
586	pVCpu->cpum.s.Guest.dr[0] = ASMGetDR0();
587	pVCpu->cpum.s.Guest.dr[1] = ASMGetDR1();
588	pVCpu->cpum.s.Guest.dr[2] = ASMGetDR2();
589	pVCpu->cpum.s.Guest.dr[3] = ASMGetDR3();
590	if (fDr6)
591	pVCpu->cpum.s.Guest.dr[6] = ASMGetDR6();
592	}
593	}
594	ASMAtomicAndU32(&pVCpu->cpum.s.fUseFlags, ~( CPUM_USED_DEBUG_REGS_GUEST \| CPUM_USED_DEBUG_REGS_HYPER
595	\| CPUM_SYNC_DEBUG_REGS_GUEST \| CPUM_SYNC_DEBUG_REGS_HYPER));
596
597	/*
598	* Restore the host's debug state. DR0-3, DR6 and only then DR7!
599	*/
600	if (pVCpu->cpum.s.fUseFlags & CPUM_USED_DEBUG_REGS_HOST)
601	{
602	/* A bit of paranoia first... */
603	uint64_t uCurDR7 = ASMGetDR7();
604	if (uCurDR7 != X86_DR7_INIT_VAL)
605	ASMSetDR7(X86_DR7_INIT_VAL);
606
607	ASMSetDR0(pVCpu->cpum.s.Host.dr0);
608	ASMSetDR1(pVCpu->cpum.s.Host.dr1);
609	ASMSetDR2(pVCpu->cpum.s.Host.dr2);
610	ASMSetDR3(pVCpu->cpum.s.Host.dr3);
611	/** @todo consider only updating if they differ, esp. DR6. Need to figure how
612	* expensive DRx reads are over DRx writes. */
613	ASMSetDR6(pVCpu->cpum.s.Host.dr6);
614	ASMSetDR7(pVCpu->cpum.s.Host.dr7);
615
616	ASMAtomicAndU32(&pVCpu->cpum.s.fUseFlags, ~CPUM_USED_DEBUG_REGS_HOST);
617	}
618
619	return fDrXLoaded;
620	}
621
622
623	/**
624	* Saves the guest DRx state if it resides host registers.
625	*
626	* This does NOT clear any use flags, so the host registers remains loaded with
627	* the guest DRx state upon return. The purpose is only to make sure the values
628	* in the CPU context structure is up to date.
629	*
630	* @returns true if the host registers contains guest values, false if not.
631	* @param pVCpu The cross context virtual CPU structure of the calling EMT.
632	* @param fDr6 Whether to include DR6 or not.
633	* @thread EMT(pVCpu)
634	*/
635	VMMR0_INT_DECL(bool) CPUMR0DebugStateMaybeSaveGuest(PVMCPU pVCpu, bool fDr6)
636	{
637	/*
638	* Do we need to save the guest DRx registered loaded into host registers?
639	* (DR7 and DR6 (if fDr6 is true) are left to the caller.)
640	*/
641	if (pVCpu->cpum.s.fUseFlags & CPUM_USED_DEBUG_REGS_GUEST)
642	{
643	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS)
644	if (CPUMIsGuestInLongModeEx(&pVCpu->cpum.s.Guest))
645	{
646	uint64_t uDr6 = pVCpu->cpum.s.Guest.dr[6];
647	HMR0SaveDebugState(pVCpu->CTX_SUFF(pVM), pVCpu, &pVCpu->cpum.s.Guest);
648	if (!fDr6)
649	pVCpu->cpum.s.Guest.dr[6] = uDr6;
650	}
651	else
652	#endif
653	{
654	pVCpu->cpum.s.Guest.dr[0] = ASMGetDR0();
655	pVCpu->cpum.s.Guest.dr[1] = ASMGetDR1();
656	pVCpu->cpum.s.Guest.dr[2] = ASMGetDR2();
657	pVCpu->cpum.s.Guest.dr[3] = ASMGetDR3();
658	if (fDr6)
659	pVCpu->cpum.s.Guest.dr[6] = ASMGetDR6();
660	}
661	return true;
662	}
663	return false;
664	}
665
666
667	/**
668	* Lazily sync in the debug state.
669	*
670	* @param pVCpu The cross context virtual CPU structure of the calling EMT.
671	* @param fDr6 Whether to include DR6 or not.
672	* @thread EMT(pVCpu)
673	*/
674	VMMR0_INT_DECL(void) CPUMR0LoadGuestDebugState(PVMCPU pVCpu, bool fDr6)
675	{
676	/*
677	* Save the host state and disarm all host BPs.
678	*/
679	cpumR0SaveHostDebugState(pVCpu);
680	Assert(ASMGetDR7() == X86_DR7_INIT_VAL);
681
682	/*
683	* Activate the guest state DR0-3.
684	* DR7 and DR6 (if fDr6 is true) are left to the caller.
685	*/
686	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS)
687	if (CPUMIsGuestInLongModeEx(&pVCpu->cpum.s.Guest))
688	ASMAtomicOrU32(&pVCpu->cpum.s.fUseFlags, CPUM_SYNC_DEBUG_REGS_GUEST); /* Postpone it to the world switch. */
689	else
690	#endif
691	{
692	ASMSetDR0(pVCpu->cpum.s.Guest.dr[0]);
693	ASMSetDR1(pVCpu->cpum.s.Guest.dr[1]);
694	ASMSetDR2(pVCpu->cpum.s.Guest.dr[2]);
695	ASMSetDR3(pVCpu->cpum.s.Guest.dr[3]);
696	if (fDr6)
697	ASMSetDR6(pVCpu->cpum.s.Guest.dr[6]);
698
699	ASMAtomicOrU32(&pVCpu->cpum.s.fUseFlags, CPUM_USED_DEBUG_REGS_GUEST);
700	}
701	}
702
703
704	/**
705	* Lazily sync in the hypervisor debug state
706	*
707	* @returns VBox status code.
708	* @param pVCpu The cross context virtual CPU structure of the calling EMT.
709	* @param fDr6 Whether to include DR6 or not.
710	* @thread EMT(pVCpu)
711	*/
712	VMMR0_INT_DECL(void) CPUMR0LoadHyperDebugState(PVMCPU pVCpu, bool fDr6)
713	{
714	/*
715	* Save the host state and disarm all host BPs.
716	*/
717	cpumR0SaveHostDebugState(pVCpu);
718	Assert(ASMGetDR7() == X86_DR7_INIT_VAL);
719
720	/*
721	* Make sure the hypervisor values are up to date.
722	*/
723	CPUMRecalcHyperDRx(pVCpu, UINT8_MAX /* no loading, please */, true);
724
725	/*
726	* Activate the guest state DR0-3.
727	* DR7 and DR6 (if fDr6 is true) are left to the caller.
728	*/
729	#if HC_ARCH_BITS == 32 && defined(VBOX_WITH_64_BITS_GUESTS)
730	if (CPUMIsGuestInLongModeEx(&pVCpu->cpum.s.Guest))
731	ASMAtomicOrU32(&pVCpu->cpum.s.fUseFlags, CPUM_SYNC_DEBUG_REGS_HYPER); /* Postpone it. */
732	else
733	#endif
734	{
735	ASMSetDR0(pVCpu->cpum.s.Hyper.dr[0]);
736	ASMSetDR1(pVCpu->cpum.s.Hyper.dr[1]);
737	ASMSetDR2(pVCpu->cpum.s.Hyper.dr[2]);
738	ASMSetDR3(pVCpu->cpum.s.Hyper.dr[3]);
739	if (fDr6)
740	ASMSetDR6(X86_DR6_INIT_VAL);
741
742	ASMAtomicOrU32(&pVCpu->cpum.s.fUseFlags, CPUM_USED_DEBUG_REGS_HYPER);
743	}
744	}
745
746	#ifdef VBOX_WITH_VMMR0_DISABLE_LAPIC_NMI
747
748	/**
749	* Per-CPU callback that probes the CPU for APIC support.
750	*
751	* @param idCpu The identifier for the CPU the function is called on.
752	* @param pvUser1 Ignored.
753	* @param pvUser2 Ignored.
754	*/
755	static DECLCALLBACK(void) cpumR0MapLocalApicCpuProber(RTCPUID idCpu, void pvUser1, void pvUser2)
756	{
757	NOREF(pvUser1); NOREF(pvUser2);
758	int iCpu = RTMpCpuIdToSetIndex(idCpu);
759	AssertReturnVoid(iCpu >= 0 && (unsigned)iCpu < RT_ELEMENTS(g_aLApics));
760
761	/*
762	* Check for APIC support.
763	*/
764	uint32_t uMaxLeaf, u32EBX, u32ECX, u32EDX;
765	ASMCpuId(0, &uMaxLeaf, &u32EBX, &u32ECX, &u32EDX);
766	if ( ( ASMIsIntelCpuEx(u32EBX, u32ECX, u32EDX)
767	\|\| ASMIsAmdCpuEx(u32EBX, u32ECX, u32EDX)
768	\|\| ASMIsViaCentaurCpuEx(u32EBX, u32ECX, u32EDX))
769	&& ASMIsValidStdRange(uMaxLeaf))
770	{
771	uint32_t uDummy;
772	ASMCpuId(1, &uDummy, &u32EBX, &u32ECX, &u32EDX);
773	if ( (u32EDX & X86_CPUID_FEATURE_EDX_APIC)
774	&& (u32EDX & X86_CPUID_FEATURE_EDX_MSR))
775	{
776	/*
777	* Safe to access the MSR. Read it and calc the BASE (a little complicated).
778	*/
779	uint64_t u64ApicBase = ASMRdMsr(MSR_IA32_APICBASE);
780	uint64_t u64Mask = MSR_IA32_APICBASE_BASE_MIN;
781
782	/* see Intel Manual: Local APIC Status and Location: MAXPHYADDR default is bit 36 */
783	uint32_t uMaxExtLeaf;
784	ASMCpuId(0x80000000, &uMaxExtLeaf, &u32EBX, &u32ECX, &u32EDX);
785	if ( uMaxExtLeaf >= UINT32_C(0x80000008)
786	&& ASMIsValidExtRange(uMaxExtLeaf))
787	{
788	uint32_t u32PhysBits;
789	ASMCpuId(0x80000008, &u32PhysBits, &u32EBX, &u32ECX, &u32EDX);
790	u32PhysBits &= 0xff;
791	u64Mask = ((UINT64_C(1) << u32PhysBits) - 1) & UINT64_C(0xfffffffffffff000);
792	}
793
794	AssertCompile(sizeof(g_aLApics[iCpu].PhysBase) == sizeof(u64ApicBase));
795	g_aLApics[iCpu].PhysBase = u64ApicBase & u64Mask;
796	g_aLApics[iCpu].fEnabled = RT_BOOL(u64ApicBase & MSR_IA32_APICBASE_EN);
797	g_aLApics[iCpu].fX2Apic = (u64ApicBase & (MSR_IA32_APICBASE_EXTD \| MSR_IA32_APICBASE_EN))
798	== (MSR_IA32_APICBASE_EXTD \| MSR_IA32_APICBASE_EN);
799	}
800	}
801	}
802
803
804
805	/**
806	* Per-CPU callback that verifies our APIC expectations.
807	*
808	* @param idCpu The identifier for the CPU the function is called on.
809	* @param pvUser1 Ignored.
810	* @param pvUser2 Ignored.
811	*/
812	static DECLCALLBACK(void) cpumR0MapLocalApicCpuChecker(RTCPUID idCpu, void pvUser1, void pvUser2)
813	{
814	NOREF(pvUser1); NOREF(pvUser2);
815
816	int iCpu = RTMpCpuIdToSetIndex(idCpu);
817	AssertReturnVoid(iCpu >= 0 && (unsigned)iCpu < RT_ELEMENTS(g_aLApics));
818	if (!g_aLApics[iCpu].fEnabled)
819	return;
820
821	/*
822	* 0x0X 82489 external APIC
823	* 0x1X Local APIC
824	* 0x2X..0xFF reserved
825	*/
826	uint32_t uApicVersion;
827	if (g_aLApics[iCpu].fX2Apic)
828	uApicVersion = ApicX2RegRead32(APIC_REG_VERSION);
829	else
830	uApicVersion = ApicRegRead(g_aLApics[iCpu].pv, APIC_REG_VERSION);
831	if ((APIC_REG_VERSION_GET_VER(uApicVersion) & 0xF0) == 0x10)
832	{
833	g_aLApics[iCpu].uVersion = uApicVersion;
834
835	#if 0 /* enable if you need it. */
836	if (g_aLApics[iCpu].fX2Apic)
837	SUPR0Printf("CPUM: X2APIC %02u - ver %#010x, lint0=%#07x lint1=%#07x pc=%#07x thmr=%#07x cmci=%#07x\n",
838	iCpu, uApicVersion,
839	ApicX2RegRead32(APIC_REG_LVT_LINT0), ApicX2RegRead32(APIC_REG_LVT_LINT1),
840	ApicX2RegRead32(APIC_REG_LVT_PC), ApicX2RegRead32(APIC_REG_LVT_THMR),
841	ApicX2RegRead32(APIC_REG_LVT_CMCI));
842	else
843	{
844	SUPR0Printf("CPUM: APIC %02u at %RGp (mapped at %p) - ver %#010x, lint0=%#07x lint1=%#07x pc=%#07x thmr=%#07x cmci=%#07x\n",
845	iCpu, g_aLApics[iCpu].PhysBase, g_aLApics[iCpu].pv, uApicVersion,
846	ApicRegRead(g_aLApics[iCpu].pv, APIC_REG_LVT_LINT0), ApicRegRead(g_aLApics[iCpu].pv, APIC_REG_LVT_LINT1),
847	ApicRegRead(g_aLApics[iCpu].pv, APIC_REG_LVT_PC), ApicRegRead(g_aLApics[iCpu].pv, APIC_REG_LVT_THMR),
848	ApicRegRead(g_aLApics[iCpu].pv, APIC_REG_LVT_CMCI));
849	if (uApicVersion & 0x80000000)
850	{
851	uint32_t uExtFeatures = ApicRegRead(g_aLApics[iCpu].pv, 0x400);
852	uint32_t cEiLvt = (uExtFeatures >> 16) & 0xff;
853	SUPR0Printf("CPUM: APIC %02u: ExtSpace available. extfeat=%08x eilvt[0..3]=%08x %08x %08x %08x\n",
854	iCpu,
855	ApicRegRead(g_aLApics[iCpu].pv, 0x400),
856	cEiLvt >= 1 ? ApicRegRead(g_aLApics[iCpu].pv, 0x500) : 0,
857	cEiLvt >= 2 ? ApicRegRead(g_aLApics[iCpu].pv, 0x510) : 0,
858	cEiLvt >= 3 ? ApicRegRead(g_aLApics[iCpu].pv, 0x520) : 0,
859	cEiLvt >= 4 ? ApicRegRead(g_aLApics[iCpu].pv, 0x530) : 0);
860	}
861	}
862	#endif
863	}
864	else
865	{
866	g_aLApics[iCpu].fEnabled = false;
867	g_aLApics[iCpu].fX2Apic = false;
868	SUPR0Printf("VBox/CPUM: Unsupported APIC version %#x (iCpu=%d)\n", uApicVersion, iCpu);
869	}
870	}
871
872
873	/**
874	* Map the MMIO page of each local APIC in the system.
875	*/
876	static int cpumR0MapLocalApics(void)
877	{
878	/*
879	* Check that we'll always stay within the array bounds.
880	*/
881	if (RTMpGetArraySize() > RT_ELEMENTS(g_aLApics))
882	{
883	LogRel(("CPUM: Too many real CPUs/cores/threads - %u, max %u\n", RTMpGetArraySize(), RT_ELEMENTS(g_aLApics)));
884	return VERR_TOO_MANY_CPUS;
885	}
886
887	/*
888	* Create mappings for all online CPUs we think have legacy APICs.
889	*/
890	int rc = RTMpOnAll(cpumR0MapLocalApicCpuProber, NULL, NULL);
891
892	for (unsigned iCpu = 0; RT_SUCCESS(rc) && iCpu < RT_ELEMENTS(g_aLApics); iCpu++)
893	{
894	if (g_aLApics[iCpu].fEnabled && !g_aLApics[iCpu].fX2Apic)
895	{
896	rc = RTR0MemObjEnterPhys(&g_aLApics[iCpu].hMemObj, g_aLApics[iCpu].PhysBase,
897	PAGE_SIZE, RTMEM_CACHE_POLICY_MMIO);
898	if (RT_SUCCESS(rc))
899	{
900	rc = RTR0MemObjMapKernel(&g_aLApics[iCpu].hMapObj, g_aLApics[iCpu].hMemObj, (void *)-1,
901	PAGE_SIZE, RTMEM_PROT_READ \| RTMEM_PROT_WRITE);
902	if (RT_SUCCESS(rc))
903	{
904	g_aLApics[iCpu].pv = RTR0MemObjAddress(g_aLApics[iCpu].hMapObj);
905	continue;
906	}
907	RTR0MemObjFree(g_aLApics[iCpu].hMemObj, true /* fFreeMappings */);
908	}
909	g_aLApics[iCpu].fEnabled = false;
910	}
911	g_aLApics[iCpu].pv = NULL;
912	}
913
914	/*
915	* Check the APICs.
916	*/
917	if (RT_SUCCESS(rc))
918	rc = RTMpOnAll(cpumR0MapLocalApicCpuChecker, NULL, NULL);
919
920	if (RT_FAILURE(rc))
921	{
922	cpumR0UnmapLocalApics();
923	return rc;
924	}
925
926	#ifdef LOG_ENABLED
927	/*
928	* Log the result (pretty useless, requires enabling CPUM in VBoxDrv
929	* and !VBOX_WITH_R0_LOGGING).
930	*/
931	if (LogIsEnabled())
932	{
933	uint32_t cEnabled = 0;
934	uint32_t cX2Apics = 0;
935	for (unsigned iCpu = 0; iCpu < RT_ELEMENTS(g_aLApics); iCpu++)
936	if (g_aLApics[iCpu].fEnabled)
937	{
938	cEnabled++;
939	cX2Apics += g_aLApics[iCpu].fX2Apic;
940	}
941	Log(("CPUM: %u APICs, %u X2APICs\n", cEnabled, cX2Apics));
942	}
943	#endif
944
945	return VINF_SUCCESS;
946	}
947
948
949	/**
950	* Unmap the Local APIC of all host CPUs.
951	*/
952	static void cpumR0UnmapLocalApics(void)
953	{
954	for (unsigned iCpu = RT_ELEMENTS(g_aLApics); iCpu-- > 0;)
955	{
956	if (g_aLApics[iCpu].pv)
957	{
958	RTR0MemObjFree(g_aLApics[iCpu].hMapObj, true /* fFreeMappings */);
959	RTR0MemObjFree(g_aLApics[iCpu].hMemObj, true /* fFreeMappings */);
960	g_aLApics[iCpu].hMapObj = NIL_RTR0MEMOBJ;
961	g_aLApics[iCpu].hMemObj = NIL_RTR0MEMOBJ;
962	g_aLApics[iCpu].fEnabled = false;
963	g_aLApics[iCpu].fX2Apic = false;
964	g_aLApics[iCpu].pv = NULL;
965	}
966	}
967	}
968
969
970	/**
971	* Updates CPUMCPU::pvApicBase and CPUMCPU::fX2Apic prior to world switch.
972	*
973	* Writes the Local APIC mapping address of the current host CPU to CPUMCPU so
974	* the world switchers can access the APIC registers for the purpose of
975	* disabling and re-enabling the NMIs. Must be called with disabled preemption
976	* or disabled interrupts!
977	*
978	* @param pVCpu The cross context virtual CPU structure of the calling EMT.
979	* @param iHostCpuSet The CPU set index of the current host CPU.
980	*/
981	VMMR0_INT_DECL(void) CPUMR0SetLApic(PVMCPU pVCpu, uint32_t iHostCpuSet)
982	{
983	Assert(iHostCpuSet <= RT_ELEMENTS(g_aLApics));
984	pVCpu->cpum.s.pvApicBase = g_aLApics[iHostCpuSet].pv;
985	pVCpu->cpum.s.fX2Apic = g_aLApics[iHostCpuSet].fX2Apic;
986	// Log6(("CPUMR0SetLApic: pvApicBase=%p fX2Apic=%d\n", g_aLApics[idxCpu].pv, g_aLApics[idxCpu].fX2Apic));
987	}
988
989	#endif /* VBOX_WITH_VMMR0_DISABLE_LAPIC_NMI */
990

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

source: vbox/trunk/src/VBox/VMM/VMMR0/CPUMR0.cpp@ 58487

Download in other formats: