PGMAll.cpp@ 105724

Last change on this file since 105724 was 104990, checked in by vboxsync, 5 months ago
VMM/IEM: Added IEMTlbInvalidateAllGlobal and made the IEMTlbInvalidateAll function handle non-global flushes. More TLB stats. bugref:10687
Property svn:eol-style set to `native` Property svn:keywords set to `Id Revision`
File size: 153.8 KB

Line
1	/* $Id: PGMAll.cpp 104990 2024-06-20 23:13:34Z vboxsync $ */
2	/** @file
3	* PGM - Page Manager and Monitor - All context code.
4	*/
5
6	/*
7	* Copyright (C) 2006-2023 Oracle and/or its affiliates.
8	*
9	* This file is part of VirtualBox base platform packages, as
10	* available from https://www.virtualbox.org.
11	*
12	* This program is free software; you can redistribute it and/or
13	* modify it under the terms of the GNU General Public License
14	* as published by the Free Software Foundation, in version 3 of the
15	* License.
16	*
17	* This program is distributed in the hope that it will be useful, but
18	* WITHOUT ANY WARRANTY; without even the implied warranty of
19	* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
20	* General Public License for more details.
21	*
22	* You should have received a copy of the GNU General Public License
23	* along with this program; if not, see <https://www.gnu.org/licenses>.
24	*
25	* SPDX-License-Identifier: GPL-3.0-only
26	*/
27
28
29	/*********************************************************************************************************************************
30	* Header Files *
31	*********************************************************************************************************************************/
32	#define LOG_GROUP LOG_GROUP_PGM
33	#define VBOX_WITHOUT_PAGING_BIT_FIELDS /* 64-bit bitfields are just asking for trouble. See @bugref{9841} and others. */
34	#include <VBox/vmm/pgm.h>
35	#include <VBox/vmm/cpum.h>
36	#include <VBox/vmm/selm.h>
37	#include <VBox/vmm/iem.h>
38	#include <VBox/vmm/iom.h>
39	#include <VBox/sup.h>
40	#include <VBox/vmm/mm.h>
41	#include <VBox/vmm/stam.h>
42	#include <VBox/vmm/trpm.h>
43	#include <VBox/vmm/em.h>
44	#include <VBox/vmm/hm.h>
45	#include <VBox/vmm/hm_vmx.h>
46	#include "PGMInternal.h"
47	#include <VBox/vmm/vmcc.h>
48	#include "PGMInline.h"
49	#include <iprt/assert.h>
50	#if defined(RT_ARCH_AMD64) \|\| defined(RT_ARCH_X86)
51	# include <iprt/asm-amd64-x86.h>
52	#endif
53	#include <iprt/string.h>
54	#include <VBox/log.h>
55	#include <VBox/param.h>
56	#include <VBox/err.h>
57
58
59	/*********************************************************************************************************************************
60	* Internal Functions *
61	*********************************************************************************************************************************/
62	DECLINLINE(int) pgmShwGetLongModePDPtr(PVMCPUCC pVCpu, RTGCPTR64 GCPtr, PX86PML4E ppPml4e, PX86PDPT ppPdpt, PX86PDPAE *ppPD);
63	DECLINLINE(int) pgmShwGetPaePoolPagePD(PVMCPUCC pVCpu, RTGCPTR GCPtr, PPGMPOOLPAGE *ppShwPde);
64	DECLINLINE(int) pgmGstMapCr3(PVMCPUCC pVCpu, RTGCPHYS GCPhysCr3, PRTHCPTR pHCPtrGuestCr3);
65	#ifdef VBOX_WITH_NESTED_HWVIRT_VMX_EPT
66	static int pgmGstSlatWalk(PVMCPUCC pVCpu, RTGCPHYS GCPhysNested, bool fIsLinearAddrValid, RTGCPTR GCPtrNested, PPGMPTWALK pWalk,
67	PPGMPTWALKGST pGstWalk);
68	static int pgmGstSlatTranslateCr3(PVMCPUCC pVCpu, uint64_t uCr3, PRTGCPHYS pGCPhysCr3);
69	static int pgmShwGetNestedEPTPDPtr(PVMCPUCC pVCpu, RTGCPTR64 GCPhysNested, PEPTPDPT ppPdpt, PEPTPD ppPD,
70	PPGMPTWALKGST pGstWalkAll);
71	#endif
72	static int pgmShwSyncLongModePDPtr(PVMCPUCC pVCpu, RTGCPTR64 GCPtr, X86PGPAEUINT uGstPml4e, X86PGPAEUINT uGstPdpe, PX86PDPAE *ppPD);
73	static int pgmShwGetEPTPDPtr(PVMCPUCC pVCpu, RTGCPTR64 GCPtr, PEPTPDPT ppPdpt, PEPTPD ppPD);
74	#ifdef PGM_WITH_PAGE_ZEROING_DETECTION
75	static bool pgmHandlePageZeroingCode(PVMCPUCC pVCpu, PCPUMCTX pCtx);
76	#endif
77
78
79	/*
80	* Second level transation - EPT.
81	*/
82	#ifdef VBOX_WITH_NESTED_HWVIRT_VMX_EPT
83	# define PGM_SLAT_TYPE PGM_SLAT_TYPE_EPT
84	# include "PGMSlatDefs.h"
85	# include "PGMAllGstSlatEpt.cpp.h"
86	# undef PGM_SLAT_TYPE
87	#endif
88
89
90	/*
91	* Shadow - 32-bit mode
92	*/
93	#define PGM_SHW_TYPE PGM_TYPE_32BIT
94	#define PGM_SHW_NAME(name) PGM_SHW_NAME_32BIT(name)
95	#include "PGMAllShw.h"
96
97	/* Guest - real mode */
98	#define PGM_GST_TYPE PGM_TYPE_REAL
99	#define PGM_GST_NAME(name) PGM_GST_NAME_REAL(name)
100	#define PGM_BTH_NAME(name) PGM_BTH_NAME_32BIT_REAL(name)
101	#define BTH_PGMPOOLKIND_PT_FOR_PT PGMPOOLKIND_32BIT_PT_FOR_PHYS
102	#define BTH_PGMPOOLKIND_ROOT PGMPOOLKIND_32BIT_PD_PHYS
103	#include "PGMGstDefs.h"
104	#include "PGMAllGst.h"
105	#include "PGMAllBth.h"
106	#undef BTH_PGMPOOLKIND_PT_FOR_PT
107	#undef BTH_PGMPOOLKIND_ROOT
108	#undef PGM_BTH_NAME
109	#undef PGM_GST_TYPE
110	#undef PGM_GST_NAME
111
112	/* Guest - protected mode */
113	#define PGM_GST_TYPE PGM_TYPE_PROT
114	#define PGM_GST_NAME(name) PGM_GST_NAME_PROT(name)
115	#define PGM_BTH_NAME(name) PGM_BTH_NAME_32BIT_PROT(name)
116	#define BTH_PGMPOOLKIND_PT_FOR_PT PGMPOOLKIND_32BIT_PT_FOR_PHYS
117	#define BTH_PGMPOOLKIND_ROOT PGMPOOLKIND_32BIT_PD_PHYS
118	#include "PGMGstDefs.h"
119	#include "PGMAllGst.h"
120	#include "PGMAllBth.h"
121	#undef BTH_PGMPOOLKIND_PT_FOR_PT
122	#undef BTH_PGMPOOLKIND_ROOT
123	#undef PGM_BTH_NAME
124	#undef PGM_GST_TYPE
125	#undef PGM_GST_NAME
126
127	/* Guest - 32-bit mode */
128	#define PGM_GST_TYPE PGM_TYPE_32BIT
129	#define PGM_GST_NAME(name) PGM_GST_NAME_32BIT(name)
130	#define PGM_BTH_NAME(name) PGM_BTH_NAME_32BIT_32BIT(name)
131	#define BTH_PGMPOOLKIND_PT_FOR_PT PGMPOOLKIND_32BIT_PT_FOR_32BIT_PT
132	#define BTH_PGMPOOLKIND_PT_FOR_BIG PGMPOOLKIND_32BIT_PT_FOR_32BIT_4MB
133	#define BTH_PGMPOOLKIND_ROOT PGMPOOLKIND_32BIT_PD
134	#include "PGMGstDefs.h"
135	#include "PGMAllGst.h"
136	#include "PGMAllBth.h"
137	#undef BTH_PGMPOOLKIND_PT_FOR_BIG
138	#undef BTH_PGMPOOLKIND_PT_FOR_PT
139	#undef BTH_PGMPOOLKIND_ROOT
140	#undef PGM_BTH_NAME
141	#undef PGM_GST_TYPE
142	#undef PGM_GST_NAME
143
144	#undef PGM_SHW_TYPE
145	#undef PGM_SHW_NAME
146
147
148	/*
149	* Shadow - PAE mode
150	*/
151	#define PGM_SHW_TYPE PGM_TYPE_PAE
152	#define PGM_SHW_NAME(name) PGM_SHW_NAME_PAE(name)
153	#define PGM_BTH_NAME(name) PGM_BTH_NAME_PAE_REAL(name)
154	#include "PGMAllShw.h"
155
156	/* Guest - real mode */
157	#define PGM_GST_TYPE PGM_TYPE_REAL
158	#define PGM_GST_NAME(name) PGM_GST_NAME_REAL(name)
159	#define PGM_BTH_NAME(name) PGM_BTH_NAME_PAE_REAL(name)
160	#define BTH_PGMPOOLKIND_PT_FOR_PT PGMPOOLKIND_PAE_PT_FOR_PHYS
161	#define BTH_PGMPOOLKIND_ROOT PGMPOOLKIND_PAE_PDPT_PHYS
162	#include "PGMGstDefs.h"
163	#include "PGMAllBth.h"
164	#undef BTH_PGMPOOLKIND_PT_FOR_PT
165	#undef BTH_PGMPOOLKIND_ROOT
166	#undef PGM_BTH_NAME
167	#undef PGM_GST_TYPE
168	#undef PGM_GST_NAME
169
170	/* Guest - protected mode */
171	#define PGM_GST_TYPE PGM_TYPE_PROT
172	#define PGM_GST_NAME(name) PGM_GST_NAME_PROT(name)
173	#define PGM_BTH_NAME(name) PGM_BTH_NAME_PAE_PROT(name)
174	#define BTH_PGMPOOLKIND_PT_FOR_PT PGMPOOLKIND_PAE_PT_FOR_PHYS
175	#define BTH_PGMPOOLKIND_ROOT PGMPOOLKIND_PAE_PDPT_PHYS
176	#include "PGMGstDefs.h"
177	#include "PGMAllBth.h"
178	#undef BTH_PGMPOOLKIND_PT_FOR_PT
179	#undef BTH_PGMPOOLKIND_ROOT
180	#undef PGM_BTH_NAME
181	#undef PGM_GST_TYPE
182	#undef PGM_GST_NAME
183
184	/* Guest - 32-bit mode */
185	#define PGM_GST_TYPE PGM_TYPE_32BIT
186	#define PGM_GST_NAME(name) PGM_GST_NAME_32BIT(name)
187	#define PGM_BTH_NAME(name) PGM_BTH_NAME_PAE_32BIT(name)
188	#define BTH_PGMPOOLKIND_PT_FOR_PT PGMPOOLKIND_PAE_PT_FOR_32BIT_PT
189	#define BTH_PGMPOOLKIND_PT_FOR_BIG PGMPOOLKIND_PAE_PT_FOR_32BIT_4MB
190	#define BTH_PGMPOOLKIND_ROOT PGMPOOLKIND_PAE_PDPT_FOR_32BIT
191	#include "PGMGstDefs.h"
192	#include "PGMAllBth.h"
193	#undef BTH_PGMPOOLKIND_PT_FOR_BIG
194	#undef BTH_PGMPOOLKIND_PT_FOR_PT
195	#undef BTH_PGMPOOLKIND_ROOT
196	#undef PGM_BTH_NAME
197	#undef PGM_GST_TYPE
198	#undef PGM_GST_NAME
199
200
201	/* Guest - PAE mode */
202	#define PGM_GST_TYPE PGM_TYPE_PAE
203	#define PGM_GST_NAME(name) PGM_GST_NAME_PAE(name)
204	#define PGM_BTH_NAME(name) PGM_BTH_NAME_PAE_PAE(name)
205	#define BTH_PGMPOOLKIND_PT_FOR_PT PGMPOOLKIND_PAE_PT_FOR_PAE_PT
206	#define BTH_PGMPOOLKIND_PT_FOR_BIG PGMPOOLKIND_PAE_PT_FOR_PAE_2MB
207	#define BTH_PGMPOOLKIND_ROOT PGMPOOLKIND_PAE_PDPT
208	#include "PGMGstDefs.h"
209	#include "PGMAllGst.h"
210	#include "PGMAllBth.h"
211	#undef BTH_PGMPOOLKIND_PT_FOR_BIG
212	#undef BTH_PGMPOOLKIND_PT_FOR_PT
213	#undef BTH_PGMPOOLKIND_ROOT
214	#undef PGM_BTH_NAME
215	#undef PGM_GST_TYPE
216	#undef PGM_GST_NAME
217
218	#undef PGM_SHW_TYPE
219	#undef PGM_SHW_NAME
220
221
222	/*
223	* Shadow - AMD64 mode
224	*/
225	#define PGM_SHW_TYPE PGM_TYPE_AMD64
226	#define PGM_SHW_NAME(name) PGM_SHW_NAME_AMD64(name)
227	#include "PGMAllShw.h"
228
229	/* Guest - protected mode (only used for AMD-V nested paging in 64 bits mode) */
230	/** @todo retire this hack. */
231	#define PGM_GST_TYPE PGM_TYPE_PROT
232	#define PGM_GST_NAME(name) PGM_GST_NAME_PROT(name)
233	#define PGM_BTH_NAME(name) PGM_BTH_NAME_AMD64_PROT(name)
234	#define BTH_PGMPOOLKIND_PT_FOR_PT PGMPOOLKIND_PAE_PT_FOR_PHYS
235	#define BTH_PGMPOOLKIND_ROOT PGMPOOLKIND_PAE_PD_PHYS
236	#include "PGMGstDefs.h"
237	#include "PGMAllBth.h"
238	#undef BTH_PGMPOOLKIND_PT_FOR_PT
239	#undef BTH_PGMPOOLKIND_ROOT
240	#undef PGM_BTH_NAME
241	#undef PGM_GST_TYPE
242	#undef PGM_GST_NAME
243
244	#ifdef VBOX_WITH_64_BITS_GUESTS
245	/* Guest - AMD64 mode */
246	# define PGM_GST_TYPE PGM_TYPE_AMD64
247	# define PGM_GST_NAME(name) PGM_GST_NAME_AMD64(name)
248	# define PGM_BTH_NAME(name) PGM_BTH_NAME_AMD64_AMD64(name)
249	# define BTH_PGMPOOLKIND_PT_FOR_PT PGMPOOLKIND_PAE_PT_FOR_PAE_PT
250	# define BTH_PGMPOOLKIND_PT_FOR_BIG PGMPOOLKIND_PAE_PT_FOR_PAE_2MB
251	# define BTH_PGMPOOLKIND_ROOT PGMPOOLKIND_64BIT_PML4
252	# include "PGMGstDefs.h"
253	# include "PGMAllGst.h"
254	# include "PGMAllBth.h"
255	# undef BTH_PGMPOOLKIND_PT_FOR_BIG
256	# undef BTH_PGMPOOLKIND_PT_FOR_PT
257	# undef BTH_PGMPOOLKIND_ROOT
258	# undef PGM_BTH_NAME
259	# undef PGM_GST_TYPE
260	# undef PGM_GST_NAME
261	#endif /* VBOX_WITH_64_BITS_GUESTS */
262
263	#undef PGM_SHW_TYPE
264	#undef PGM_SHW_NAME
265
266
267	/*
268	* Shadow - 32-bit nested paging mode.
269	*/
270	#define PGM_SHW_TYPE PGM_TYPE_NESTED_32BIT
271	#define PGM_SHW_NAME(name) PGM_SHW_NAME_NESTED_32BIT(name)
272	#include "PGMAllShw.h"
273
274	/* Guest - real mode */
275	#define PGM_GST_TYPE PGM_TYPE_REAL
276	#define PGM_GST_NAME(name) PGM_GST_NAME_REAL(name)
277	#define PGM_BTH_NAME(name) PGM_BTH_NAME_NESTED_32BIT_REAL(name)
278	#include "PGMGstDefs.h"
279	#include "PGMAllBth.h"
280	#undef PGM_BTH_NAME
281	#undef PGM_GST_TYPE
282	#undef PGM_GST_NAME
283
284	/* Guest - protected mode */
285	#define PGM_GST_TYPE PGM_TYPE_PROT
286	#define PGM_GST_NAME(name) PGM_GST_NAME_PROT(name)
287	#define PGM_BTH_NAME(name) PGM_BTH_NAME_NESTED_32BIT_PROT(name)
288	#include "PGMGstDefs.h"
289	#include "PGMAllBth.h"
290	#undef PGM_BTH_NAME
291	#undef PGM_GST_TYPE
292	#undef PGM_GST_NAME
293
294	/* Guest - 32-bit mode */
295	#define PGM_GST_TYPE PGM_TYPE_32BIT
296	#define PGM_GST_NAME(name) PGM_GST_NAME_32BIT(name)
297	#define PGM_BTH_NAME(name) PGM_BTH_NAME_NESTED_32BIT_32BIT(name)
298	#include "PGMGstDefs.h"
299	#include "PGMAllBth.h"
300	#undef PGM_BTH_NAME
301	#undef PGM_GST_TYPE
302	#undef PGM_GST_NAME
303
304	/* Guest - PAE mode */
305	#define PGM_GST_TYPE PGM_TYPE_PAE
306	#define PGM_GST_NAME(name) PGM_GST_NAME_PAE(name)
307	#define PGM_BTH_NAME(name) PGM_BTH_NAME_NESTED_32BIT_PAE(name)
308	#include "PGMGstDefs.h"
309	#include "PGMAllBth.h"
310	#undef PGM_BTH_NAME
311	#undef PGM_GST_TYPE
312	#undef PGM_GST_NAME
313
314	#ifdef VBOX_WITH_64_BITS_GUESTS
315	/* Guest - AMD64 mode */
316	# define PGM_GST_TYPE PGM_TYPE_AMD64
317	# define PGM_GST_NAME(name) PGM_GST_NAME_AMD64(name)
318	# define PGM_BTH_NAME(name) PGM_BTH_NAME_NESTED_32BIT_AMD64(name)
319	# include "PGMGstDefs.h"
320	# include "PGMAllBth.h"
321	# undef PGM_BTH_NAME
322	# undef PGM_GST_TYPE
323	# undef PGM_GST_NAME
324	#endif /* VBOX_WITH_64_BITS_GUESTS */
325
326	#undef PGM_SHW_TYPE
327	#undef PGM_SHW_NAME
328
329
330	/*
331	* Shadow - PAE nested paging mode.
332	*/
333	#define PGM_SHW_TYPE PGM_TYPE_NESTED_PAE
334	#define PGM_SHW_NAME(name) PGM_SHW_NAME_NESTED_PAE(name)
335	#include "PGMAllShw.h"
336
337	/* Guest - real mode */
338	#define PGM_GST_TYPE PGM_TYPE_REAL
339	#define PGM_GST_NAME(name) PGM_GST_NAME_REAL(name)
340	#define PGM_BTH_NAME(name) PGM_BTH_NAME_NESTED_PAE_REAL(name)
341	#include "PGMGstDefs.h"
342	#include "PGMAllBth.h"
343	#undef PGM_BTH_NAME
344	#undef PGM_GST_TYPE
345	#undef PGM_GST_NAME
346
347	/* Guest - protected mode */
348	#define PGM_GST_TYPE PGM_TYPE_PROT
349	#define PGM_GST_NAME(name) PGM_GST_NAME_PROT(name)
350	#define PGM_BTH_NAME(name) PGM_BTH_NAME_NESTED_PAE_PROT(name)
351	#include "PGMGstDefs.h"
352	#include "PGMAllBth.h"
353	#undef PGM_BTH_NAME
354	#undef PGM_GST_TYPE
355	#undef PGM_GST_NAME
356
357	/* Guest - 32-bit mode */
358	#define PGM_GST_TYPE PGM_TYPE_32BIT
359	#define PGM_GST_NAME(name) PGM_GST_NAME_32BIT(name)
360	#define PGM_BTH_NAME(name) PGM_BTH_NAME_NESTED_PAE_32BIT(name)
361	#include "PGMGstDefs.h"
362	#include "PGMAllBth.h"
363	#undef PGM_BTH_NAME
364	#undef PGM_GST_TYPE
365	#undef PGM_GST_NAME
366
367	/* Guest - PAE mode */
368	#define PGM_GST_TYPE PGM_TYPE_PAE
369	#define PGM_GST_NAME(name) PGM_GST_NAME_PAE(name)
370	#define PGM_BTH_NAME(name) PGM_BTH_NAME_NESTED_PAE_PAE(name)
371	#include "PGMGstDefs.h"
372	#include "PGMAllBth.h"
373	#undef PGM_BTH_NAME
374	#undef PGM_GST_TYPE
375	#undef PGM_GST_NAME
376
377	#ifdef VBOX_WITH_64_BITS_GUESTS
378	/* Guest - AMD64 mode */
379	# define PGM_GST_TYPE PGM_TYPE_AMD64
380	# define PGM_GST_NAME(name) PGM_GST_NAME_AMD64(name)
381	# define PGM_BTH_NAME(name) PGM_BTH_NAME_NESTED_PAE_AMD64(name)
382	# include "PGMGstDefs.h"
383	# include "PGMAllBth.h"
384	# undef PGM_BTH_NAME
385	# undef PGM_GST_TYPE
386	# undef PGM_GST_NAME
387	#endif /* VBOX_WITH_64_BITS_GUESTS */
388
389	#undef PGM_SHW_TYPE
390	#undef PGM_SHW_NAME
391
392
393	/*
394	* Shadow - AMD64 nested paging mode.
395	*/
396	#define PGM_SHW_TYPE PGM_TYPE_NESTED_AMD64
397	#define PGM_SHW_NAME(name) PGM_SHW_NAME_NESTED_AMD64(name)
398	#include "PGMAllShw.h"
399
400	/* Guest - real mode */
401	#define PGM_GST_TYPE PGM_TYPE_REAL
402	#define PGM_GST_NAME(name) PGM_GST_NAME_REAL(name)
403	#define PGM_BTH_NAME(name) PGM_BTH_NAME_NESTED_AMD64_REAL(name)
404	#include "PGMGstDefs.h"
405	#include "PGMAllBth.h"
406	#undef PGM_BTH_NAME
407	#undef PGM_GST_TYPE
408	#undef PGM_GST_NAME
409
410	/* Guest - protected mode */
411	#define PGM_GST_TYPE PGM_TYPE_PROT
412	#define PGM_GST_NAME(name) PGM_GST_NAME_PROT(name)
413	#define PGM_BTH_NAME(name) PGM_BTH_NAME_NESTED_AMD64_PROT(name)
414	#include "PGMGstDefs.h"
415	#include "PGMAllBth.h"
416	#undef PGM_BTH_NAME
417	#undef PGM_GST_TYPE
418	#undef PGM_GST_NAME
419
420	/* Guest - 32-bit mode */
421	#define PGM_GST_TYPE PGM_TYPE_32BIT
422	#define PGM_GST_NAME(name) PGM_GST_NAME_32BIT(name)
423	#define PGM_BTH_NAME(name) PGM_BTH_NAME_NESTED_AMD64_32BIT(name)
424	#include "PGMGstDefs.h"
425	#include "PGMAllBth.h"
426	#undef PGM_BTH_NAME
427	#undef PGM_GST_TYPE
428	#undef PGM_GST_NAME
429
430	/* Guest - PAE mode */
431	#define PGM_GST_TYPE PGM_TYPE_PAE
432	#define PGM_GST_NAME(name) PGM_GST_NAME_PAE(name)
433	#define PGM_BTH_NAME(name) PGM_BTH_NAME_NESTED_AMD64_PAE(name)
434	#include "PGMGstDefs.h"
435	#include "PGMAllBth.h"
436	#undef PGM_BTH_NAME
437	#undef PGM_GST_TYPE
438	#undef PGM_GST_NAME
439
440	#ifdef VBOX_WITH_64_BITS_GUESTS
441	/* Guest - AMD64 mode */
442	# define PGM_GST_TYPE PGM_TYPE_AMD64
443	# define PGM_GST_NAME(name) PGM_GST_NAME_AMD64(name)
444	# define PGM_BTH_NAME(name) PGM_BTH_NAME_NESTED_AMD64_AMD64(name)
445	# include "PGMGstDefs.h"
446	# include "PGMAllBth.h"
447	# undef PGM_BTH_NAME
448	# undef PGM_GST_TYPE
449	# undef PGM_GST_NAME
450	#endif /* VBOX_WITH_64_BITS_GUESTS */
451
452	#undef PGM_SHW_TYPE
453	#undef PGM_SHW_NAME
454
455
456	/*
457	* Shadow - EPT.
458	*/
459	#define PGM_SHW_TYPE PGM_TYPE_EPT
460	#define PGM_SHW_NAME(name) PGM_SHW_NAME_EPT(name)
461	#include "PGMAllShw.h"
462
463	/* Guest - real mode */
464	#define PGM_GST_TYPE PGM_TYPE_REAL
465	#define PGM_GST_NAME(name) PGM_GST_NAME_REAL(name)
466	#define PGM_BTH_NAME(name) PGM_BTH_NAME_EPT_REAL(name)
467	#define BTH_PGMPOOLKIND_PT_FOR_PT PGMPOOLKIND_EPT_PT_FOR_PHYS
468	#include "PGMGstDefs.h"
469	#include "PGMAllBth.h"
470	#undef BTH_PGMPOOLKIND_PT_FOR_PT
471	#undef PGM_BTH_NAME
472	#undef PGM_GST_TYPE
473	#undef PGM_GST_NAME
474
475	/* Guest - protected mode */
476	#define PGM_GST_TYPE PGM_TYPE_PROT
477	#define PGM_GST_NAME(name) PGM_GST_NAME_PROT(name)
478	#define PGM_BTH_NAME(name) PGM_BTH_NAME_EPT_PROT(name)
479	#define BTH_PGMPOOLKIND_PT_FOR_PT PGMPOOLKIND_EPT_PT_FOR_PHYS
480	#include "PGMGstDefs.h"
481	#include "PGMAllBth.h"
482	#undef BTH_PGMPOOLKIND_PT_FOR_PT
483	#undef PGM_BTH_NAME
484	#undef PGM_GST_TYPE
485	#undef PGM_GST_NAME
486
487	/* Guest - 32-bit mode */
488	#define PGM_GST_TYPE PGM_TYPE_32BIT
489	#define PGM_GST_NAME(name) PGM_GST_NAME_32BIT(name)
490	#define PGM_BTH_NAME(name) PGM_BTH_NAME_EPT_32BIT(name)
491	#define BTH_PGMPOOLKIND_PT_FOR_PT PGMPOOLKIND_EPT_PT_FOR_PHYS
492	#include "PGMGstDefs.h"
493	#include "PGMAllBth.h"
494	#undef BTH_PGMPOOLKIND_PT_FOR_PT
495	#undef PGM_BTH_NAME
496	#undef PGM_GST_TYPE
497	#undef PGM_GST_NAME
498
499	/* Guest - PAE mode */
500	#define PGM_GST_TYPE PGM_TYPE_PAE
501	#define PGM_GST_NAME(name) PGM_GST_NAME_PAE(name)
502	#define PGM_BTH_NAME(name) PGM_BTH_NAME_EPT_PAE(name)
503	#define BTH_PGMPOOLKIND_PT_FOR_PT PGMPOOLKIND_EPT_PT_FOR_PHYS
504	#include "PGMGstDefs.h"
505	#include "PGMAllBth.h"
506	#undef BTH_PGMPOOLKIND_PT_FOR_PT
507	#undef PGM_BTH_NAME
508	#undef PGM_GST_TYPE
509	#undef PGM_GST_NAME
510
511	#ifdef VBOX_WITH_64_BITS_GUESTS
512	/* Guest - AMD64 mode */
513	# define PGM_GST_TYPE PGM_TYPE_AMD64
514	# define PGM_GST_NAME(name) PGM_GST_NAME_AMD64(name)
515	# define PGM_BTH_NAME(name) PGM_BTH_NAME_EPT_AMD64(name)
516	# define BTH_PGMPOOLKIND_PT_FOR_PT PGMPOOLKIND_EPT_PT_FOR_PHYS
517	# include "PGMGstDefs.h"
518	# include "PGMAllBth.h"
519	# undef BTH_PGMPOOLKIND_PT_FOR_PT
520	# undef PGM_BTH_NAME
521	# undef PGM_GST_TYPE
522	# undef PGM_GST_NAME
523	#endif /* VBOX_WITH_64_BITS_GUESTS */
524
525	#undef PGM_SHW_TYPE
526	#undef PGM_SHW_NAME
527
528
529	/*
530	* Shadow - NEM / None.
531	*/
532	#define PGM_SHW_TYPE PGM_TYPE_NONE
533	#define PGM_SHW_NAME(name) PGM_SHW_NAME_NONE(name)
534	#include "PGMAllShw.h"
535
536	/* Guest - real mode */
537	#define PGM_GST_TYPE PGM_TYPE_REAL
538	#define PGM_GST_NAME(name) PGM_GST_NAME_REAL(name)
539	#define PGM_BTH_NAME(name) PGM_BTH_NAME_NONE_REAL(name)
540	#include "PGMGstDefs.h"
541	#include "PGMAllBth.h"
542	#undef PGM_BTH_NAME
543	#undef PGM_GST_TYPE
544	#undef PGM_GST_NAME
545
546	/* Guest - protected mode */
547	#define PGM_GST_TYPE PGM_TYPE_PROT
548	#define PGM_GST_NAME(name) PGM_GST_NAME_PROT(name)
549	#define PGM_BTH_NAME(name) PGM_BTH_NAME_NONE_PROT(name)
550	#include "PGMGstDefs.h"
551	#include "PGMAllBth.h"
552	#undef PGM_BTH_NAME
553	#undef PGM_GST_TYPE
554	#undef PGM_GST_NAME
555
556	/* Guest - 32-bit mode */
557	#define PGM_GST_TYPE PGM_TYPE_32BIT
558	#define PGM_GST_NAME(name) PGM_GST_NAME_32BIT(name)
559	#define PGM_BTH_NAME(name) PGM_BTH_NAME_NONE_32BIT(name)
560	#include "PGMGstDefs.h"
561	#include "PGMAllBth.h"
562	#undef PGM_BTH_NAME
563	#undef PGM_GST_TYPE
564	#undef PGM_GST_NAME
565
566	/* Guest - PAE mode */
567	#define PGM_GST_TYPE PGM_TYPE_PAE
568	#define PGM_GST_NAME(name) PGM_GST_NAME_PAE(name)
569	#define PGM_BTH_NAME(name) PGM_BTH_NAME_NONE_PAE(name)
570	#include "PGMGstDefs.h"
571	#include "PGMAllBth.h"
572	#undef PGM_BTH_NAME
573	#undef PGM_GST_TYPE
574	#undef PGM_GST_NAME
575
576	#ifdef VBOX_WITH_64_BITS_GUESTS
577	/* Guest - AMD64 mode */
578	# define PGM_GST_TYPE PGM_TYPE_AMD64
579	# define PGM_GST_NAME(name) PGM_GST_NAME_AMD64(name)
580	# define PGM_BTH_NAME(name) PGM_BTH_NAME_NONE_AMD64(name)
581	# include "PGMGstDefs.h"
582	# include "PGMAllBth.h"
583	# undef PGM_BTH_NAME
584	# undef PGM_GST_TYPE
585	# undef PGM_GST_NAME
586	#endif /* VBOX_WITH_64_BITS_GUESTS */
587
588	#undef PGM_SHW_TYPE
589	#undef PGM_SHW_NAME
590
591
592
593	/**
594	* Guest mode data array.
595	*/
596	PGMMODEDATAGST const g_aPgmGuestModeData[PGM_GUEST_MODE_DATA_ARRAY_SIZE] =
597	{
598	{ UINT32_MAX, NULL, NULL, NULL, NULL }, /* 0 */
599	{
600	PGM_TYPE_REAL,
601	PGM_GST_NAME_REAL(GetPage),
602	PGM_GST_NAME_REAL(QueryPageFast),
603	PGM_GST_NAME_REAL(ModifyPage),
604	PGM_GST_NAME_REAL(Enter),
605	PGM_GST_NAME_REAL(Exit),
606	#ifdef IN_RING3
607	PGM_GST_NAME_REAL(Relocate),
608	#endif
609	},
610	{
611	PGM_TYPE_PROT,
612	PGM_GST_NAME_PROT(GetPage),
613	PGM_GST_NAME_PROT(QueryPageFast),
614	PGM_GST_NAME_PROT(ModifyPage),
615	PGM_GST_NAME_PROT(Enter),
616	PGM_GST_NAME_PROT(Exit),
617	#ifdef IN_RING3
618	PGM_GST_NAME_PROT(Relocate),
619	#endif
620	},
621	{
622	PGM_TYPE_32BIT,
623	PGM_GST_NAME_32BIT(GetPage),
624	PGM_GST_NAME_32BIT(QueryPageFast),
625	PGM_GST_NAME_32BIT(ModifyPage),
626	PGM_GST_NAME_32BIT(Enter),
627	PGM_GST_NAME_32BIT(Exit),
628	#ifdef IN_RING3
629	PGM_GST_NAME_32BIT(Relocate),
630	#endif
631	},
632	{
633	PGM_TYPE_PAE,
634	PGM_GST_NAME_PAE(GetPage),
635	PGM_GST_NAME_PAE(QueryPageFast),
636	PGM_GST_NAME_PAE(ModifyPage),
637	PGM_GST_NAME_PAE(Enter),
638	PGM_GST_NAME_PAE(Exit),
639	#ifdef IN_RING3
640	PGM_GST_NAME_PAE(Relocate),
641	#endif
642	},
643	#ifdef VBOX_WITH_64_BITS_GUESTS
644	{
645	PGM_TYPE_AMD64,
646	PGM_GST_NAME_AMD64(GetPage),
647	PGM_GST_NAME_AMD64(QueryPageFast),
648	PGM_GST_NAME_AMD64(ModifyPage),
649	PGM_GST_NAME_AMD64(Enter),
650	PGM_GST_NAME_AMD64(Exit),
651	# ifdef IN_RING3
652	PGM_GST_NAME_AMD64(Relocate),
653	# endif
654	},
655	#endif
656	};
657
658
659	/**
660	* The shadow mode data array.
661	*/
662	PGMMODEDATASHW const g_aPgmShadowModeData[PGM_SHADOW_MODE_DATA_ARRAY_SIZE] =
663	{
664	{ UINT8_MAX, NULL, NULL, NULL, NULL }, /* 0 */
665	{ UINT8_MAX, NULL, NULL, NULL, NULL }, /* PGM_TYPE_REAL */
666	{ UINT8_MAX, NULL, NULL, NULL, NULL }, /* PGM_TYPE_PROT */
667	{
668	PGM_TYPE_32BIT,
669	PGM_SHW_NAME_32BIT(GetPage),
670	PGM_SHW_NAME_32BIT(ModifyPage),
671	PGM_SHW_NAME_32BIT(Enter),
672	PGM_SHW_NAME_32BIT(Exit),
673	#ifdef IN_RING3
674	PGM_SHW_NAME_32BIT(Relocate),
675	#endif
676	},
677	{
678	PGM_TYPE_PAE,
679	PGM_SHW_NAME_PAE(GetPage),
680	PGM_SHW_NAME_PAE(ModifyPage),
681	PGM_SHW_NAME_PAE(Enter),
682	PGM_SHW_NAME_PAE(Exit),
683	#ifdef IN_RING3
684	PGM_SHW_NAME_PAE(Relocate),
685	#endif
686	},
687	{
688	PGM_TYPE_AMD64,
689	PGM_SHW_NAME_AMD64(GetPage),
690	PGM_SHW_NAME_AMD64(ModifyPage),
691	PGM_SHW_NAME_AMD64(Enter),
692	PGM_SHW_NAME_AMD64(Exit),
693	#ifdef IN_RING3
694	PGM_SHW_NAME_AMD64(Relocate),
695	#endif
696	},
697	{
698	PGM_TYPE_NESTED_32BIT,
699	PGM_SHW_NAME_NESTED_32BIT(GetPage),
700	PGM_SHW_NAME_NESTED_32BIT(ModifyPage),
701	PGM_SHW_NAME_NESTED_32BIT(Enter),
702	PGM_SHW_NAME_NESTED_32BIT(Exit),
703	#ifdef IN_RING3
704	PGM_SHW_NAME_NESTED_32BIT(Relocate),
705	#endif
706	},
707	{
708	PGM_TYPE_NESTED_PAE,
709	PGM_SHW_NAME_NESTED_PAE(GetPage),
710	PGM_SHW_NAME_NESTED_PAE(ModifyPage),
711	PGM_SHW_NAME_NESTED_PAE(Enter),
712	PGM_SHW_NAME_NESTED_PAE(Exit),
713	#ifdef IN_RING3
714	PGM_SHW_NAME_NESTED_PAE(Relocate),
715	#endif
716	},
717	{
718	PGM_TYPE_NESTED_AMD64,
719	PGM_SHW_NAME_NESTED_AMD64(GetPage),
720	PGM_SHW_NAME_NESTED_AMD64(ModifyPage),
721	PGM_SHW_NAME_NESTED_AMD64(Enter),
722	PGM_SHW_NAME_NESTED_AMD64(Exit),
723	#ifdef IN_RING3
724	PGM_SHW_NAME_NESTED_AMD64(Relocate),
725	#endif
726	},
727	{
728	PGM_TYPE_EPT,
729	PGM_SHW_NAME_EPT(GetPage),
730	PGM_SHW_NAME_EPT(ModifyPage),
731	PGM_SHW_NAME_EPT(Enter),
732	PGM_SHW_NAME_EPT(Exit),
733	#ifdef IN_RING3
734	PGM_SHW_NAME_EPT(Relocate),
735	#endif
736	},
737	{
738	PGM_TYPE_NONE,
739	PGM_SHW_NAME_NONE(GetPage),
740	PGM_SHW_NAME_NONE(ModifyPage),
741	PGM_SHW_NAME_NONE(Enter),
742	PGM_SHW_NAME_NONE(Exit),
743	#ifdef IN_RING3
744	PGM_SHW_NAME_NONE(Relocate),
745	#endif
746	},
747	};
748
749
750	/**
751	* The guest+shadow mode data array.
752	*/
753	PGMMODEDATABTH const g_aPgmBothModeData[PGM_BOTH_MODE_DATA_ARRAY_SIZE] =
754	{
755	#if !defined(IN_RING3) && !defined(VBOX_STRICT)
756	# define PGMMODEDATABTH_NULL_ENTRY() { UINT32_MAX, UINT32_MAX, NULL, NULL, NULL, NULL, NULL, NULL, NULL }
757	# define PGMMODEDATABTH_ENTRY(uShwT, uGstT, Nm) \
758	{ uShwT, uGstT, Nm(InvalidatePage), Nm(SyncCR3), Nm(PrefetchPage), Nm(VerifyAccessSyncPage), Nm(MapCR3), Nm(UnmapCR3), Nm(Enter), Nm(Trap0eHandler), Nm(NestedTrap0eHandler) }
759
760	#elif !defined(IN_RING3) && defined(VBOX_STRICT)
761	# define PGMMODEDATABTH_NULL_ENTRY() { UINT32_MAX, UINT32_MAX, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL }
762	# define PGMMODEDATABTH_ENTRY(uShwT, uGstT, Nm) \
763	{ uShwT, uGstT, Nm(InvalidatePage), Nm(SyncCR3), Nm(PrefetchPage), Nm(VerifyAccessSyncPage), Nm(MapCR3), Nm(UnmapCR3), Nm(Enter), Nm(Trap0eHandler), Nm(NestedTrap0eHandler), Nm(AssertCR3) }
764
765	#elif defined(IN_RING3) && !defined(VBOX_STRICT)
766	# define PGMMODEDATABTH_NULL_ENTRY() { UINT32_MAX, UINT32_MAX, NULL, NULL, NULL, NULL, NULL, NULL }
767	# define PGMMODEDATABTH_ENTRY(uShwT, uGstT, Nm) \
768	{ uShwT, uGstT, Nm(InvalidatePage), Nm(SyncCR3), Nm(PrefetchPage), Nm(VerifyAccessSyncPage), Nm(MapCR3), Nm(UnmapCR3), Nm(Enter), }
769
770	#elif defined(IN_RING3) && defined(VBOX_STRICT)
771	# define PGMMODEDATABTH_NULL_ENTRY() { UINT32_MAX, UINT32_MAX, NULL, NULL, NULL, NULL, NULL, NULL, NULL }
772	# define PGMMODEDATABTH_ENTRY(uShwT, uGstT, Nm) \
773	{ uShwT, uGstT, Nm(InvalidatePage), Nm(SyncCR3), Nm(PrefetchPage), Nm(VerifyAccessSyncPage), Nm(MapCR3), Nm(UnmapCR3), Nm(Enter), Nm(AssertCR3) }
774
775	#else
776	# error "Misconfig."
777	#endif
778
779	/* 32-bit shadow paging mode: */
780	PGMMODEDATABTH_NULL_ENTRY(), /* 0 */
781	PGMMODEDATABTH_ENTRY(PGM_TYPE_32BIT, PGM_TYPE_REAL, PGM_BTH_NAME_32BIT_REAL),
782	PGMMODEDATABTH_ENTRY(PGM_TYPE_32BIT, PGM_TYPE_PROT, PGM_BTH_NAME_32BIT_PROT),
783	PGMMODEDATABTH_ENTRY(PGM_TYPE_32BIT, PGM_TYPE_32BIT, PGM_BTH_NAME_32BIT_32BIT),
784	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_32BIT, PGM_TYPE_PAE - illegal */
785	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_32BIT, PGM_TYPE_AMD64 - illegal */
786	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_32BIT, PGM_TYPE_NESTED_32BIT - illegal */
787	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_32BIT, PGM_TYPE_NESTED_PAE - illegal */
788	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_32BIT, PGM_TYPE_NESTED_AMD64 - illegal */
789	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_32BIT, PGM_TYPE_EPT - illegal */
790	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_32BIT, PGM_TYPE_NONE - illegal */
791
792	/* PAE shadow paging mode: */
793	PGMMODEDATABTH_NULL_ENTRY(), /* 0 */
794	PGMMODEDATABTH_ENTRY(PGM_TYPE_PAE, PGM_TYPE_REAL, PGM_BTH_NAME_PAE_REAL),
795	PGMMODEDATABTH_ENTRY(PGM_TYPE_PAE, PGM_TYPE_PROT, PGM_BTH_NAME_PAE_PROT),
796	PGMMODEDATABTH_ENTRY(PGM_TYPE_PAE, PGM_TYPE_32BIT, PGM_BTH_NAME_PAE_32BIT),
797	PGMMODEDATABTH_ENTRY(PGM_TYPE_PAE, PGM_TYPE_PAE, PGM_BTH_NAME_PAE_PAE),
798	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_PAE, PGM_TYPE_AMD64 - illegal */
799	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_PAE, PGM_TYPE_NESTED_32BIT - illegal */
800	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_PAE, PGM_TYPE_NESTED_PAE - illegal */
801	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_PAE, PGM_TYPE_NESTED_AMD64 - illegal */
802	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_PAE, PGM_TYPE_EPT - illegal */
803	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_PAE, PGM_TYPE_NONE - illegal */
804
805	/* AMD64 shadow paging mode: */
806	PGMMODEDATABTH_NULL_ENTRY(), /* 0 */
807	PGMMODEDATABTH_NULL_ENTRY(), //PGMMODEDATABTH_ENTRY(PGM_TYPE_AMD64, PGM_TYPE_REAL, PGM_BTH_NAME_AMD64_REAL),
808	PGMMODEDATABTH_NULL_ENTRY(), //PGMMODEDATABTH_ENTRY(PGM_TYPE_AMD64, PGM_TYPE_PROT, PGM_BTH_NAME_AMD64_PROT),
809	PGMMODEDATABTH_NULL_ENTRY(), //PGMMODEDATABTH_ENTRY(PGM_TYPE_AMD64, PGM_TYPE_32BIT, PGM_BTH_NAME_AMD64_32BIT),
810	PGMMODEDATABTH_NULL_ENTRY(), //PGMMODEDATABTH_ENTRY(PGM_TYPE_AMD64, PGM_TYPE_PAE, PGM_BTH_NAME_AMD64_PAE),
811	#ifdef VBOX_WITH_64_BITS_GUESTS
812	PGMMODEDATABTH_ENTRY(PGM_TYPE_AMD64, PGM_TYPE_AMD64, PGM_BTH_NAME_AMD64_AMD64),
813	#else
814	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_AMD64, PGM_TYPE_AMD64 - illegal */
815	#endif
816	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_AMD64, PGM_TYPE_NESTED_32BIT - illegal */
817	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_AMD64, PGM_TYPE_NESTED_PAE - illegal */
818	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_AMD64, PGM_TYPE_NESTED_AMD64 - illegal */
819	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_AMD64, PGM_TYPE_EPT - illegal */
820	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_AMD64, PGM_TYPE_NONE - illegal */
821
822	/* 32-bit nested paging mode: */
823	PGMMODEDATABTH_NULL_ENTRY(), /* 0 */
824	PGMMODEDATABTH_ENTRY(PGM_TYPE_NESTED_32BIT, PGM_TYPE_REAL, PGM_BTH_NAME_NESTED_32BIT_REAL),
825	PGMMODEDATABTH_ENTRY(PGM_TYPE_NESTED_32BIT, PGM_TYPE_PROT, PGM_BTH_NAME_NESTED_32BIT_PROT),
826	PGMMODEDATABTH_ENTRY(PGM_TYPE_NESTED_32BIT, PGM_TYPE_32BIT, PGM_BTH_NAME_NESTED_32BIT_32BIT),
827	PGMMODEDATABTH_ENTRY(PGM_TYPE_NESTED_32BIT, PGM_TYPE_PAE, PGM_BTH_NAME_NESTED_32BIT_PAE),
828	#ifdef VBOX_WITH_64_BITS_GUESTS
829	PGMMODEDATABTH_ENTRY(PGM_TYPE_NESTED_32BIT, PGM_TYPE_AMD64, PGM_BTH_NAME_NESTED_32BIT_AMD64),
830	#else
831	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_NESTED_32BIT, PGM_TYPE_AMD64 - illegal */
832	#endif
833	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_NESTED_32BIT, PGM_TYPE_NESTED_32BIT - illegal */
834	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_NESTED_32BIT, PGM_TYPE_NESTED_PAE - illegal */
835	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_NESTED_32BIT, PGM_TYPE_NESTED_AMD64 - illegal */
836	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_NESTED_32BIT, PGM_TYPE_EPT - illegal */
837	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_NESTED_32BIT, PGM_TYPE_NONE - illegal */
838
839	/* PAE nested paging mode: */
840	PGMMODEDATABTH_NULL_ENTRY(), /* 0 */
841	PGMMODEDATABTH_ENTRY(PGM_TYPE_NESTED_PAE, PGM_TYPE_REAL, PGM_BTH_NAME_NESTED_PAE_REAL),
842	PGMMODEDATABTH_ENTRY(PGM_TYPE_NESTED_PAE, PGM_TYPE_PROT, PGM_BTH_NAME_NESTED_PAE_PROT),
843	PGMMODEDATABTH_ENTRY(PGM_TYPE_NESTED_PAE, PGM_TYPE_32BIT, PGM_BTH_NAME_NESTED_PAE_32BIT),
844	PGMMODEDATABTH_ENTRY(PGM_TYPE_NESTED_PAE, PGM_TYPE_PAE, PGM_BTH_NAME_NESTED_PAE_PAE),
845	#ifdef VBOX_WITH_64_BITS_GUESTS
846	PGMMODEDATABTH_ENTRY(PGM_TYPE_NESTED_PAE, PGM_TYPE_AMD64, PGM_BTH_NAME_NESTED_PAE_AMD64),
847	#else
848	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_NESTED_PAE, PGM_TYPE_AMD64 - illegal */
849	#endif
850	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_NESTED_PAE, PGM_TYPE_NESTED_32BIT - illegal */
851	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_NESTED_PAE, PGM_TYPE_NESTED_PAE - illegal */
852	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_NESTED_PAE, PGM_TYPE_NESTED_AMD64 - illegal */
853	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_NESTED_PAE, PGM_TYPE_EPT - illegal */
854	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_NESTED_PAE, PGM_TYPE_NONE - illegal */
855
856	/* AMD64 nested paging mode: */
857	PGMMODEDATABTH_NULL_ENTRY(), /* 0 */
858	PGMMODEDATABTH_ENTRY(PGM_TYPE_NESTED_AMD64, PGM_TYPE_REAL, PGM_BTH_NAME_NESTED_AMD64_REAL),
859	PGMMODEDATABTH_ENTRY(PGM_TYPE_NESTED_AMD64, PGM_TYPE_PROT, PGM_BTH_NAME_NESTED_AMD64_PROT),
860	PGMMODEDATABTH_ENTRY(PGM_TYPE_NESTED_AMD64, PGM_TYPE_32BIT, PGM_BTH_NAME_NESTED_AMD64_32BIT),
861	PGMMODEDATABTH_ENTRY(PGM_TYPE_NESTED_AMD64, PGM_TYPE_PAE, PGM_BTH_NAME_NESTED_AMD64_PAE),
862	#ifdef VBOX_WITH_64_BITS_GUESTS
863	PGMMODEDATABTH_ENTRY(PGM_TYPE_NESTED_AMD64, PGM_TYPE_AMD64, PGM_BTH_NAME_NESTED_AMD64_AMD64),
864	#else
865	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_NESTED_AMD64, PGM_TYPE_AMD64 - illegal */
866	#endif
867	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_NESTED_AMD64, PGM_TYPE_NESTED_32BIT - illegal */
868	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_NESTED_AMD64, PGM_TYPE_NESTED_PAE - illegal */
869	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_NESTED_AMD64, PGM_TYPE_NESTED_AMD64 - illegal */
870	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_NESTED_AMD64, PGM_TYPE_EPT - illegal */
871	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_NESTED_AMD64, PGM_TYPE_NONE - illegal */
872
873	/* EPT nested paging mode: */
874	PGMMODEDATABTH_NULL_ENTRY(), /* 0 */
875	PGMMODEDATABTH_ENTRY(PGM_TYPE_EPT, PGM_TYPE_REAL, PGM_BTH_NAME_EPT_REAL),
876	PGMMODEDATABTH_ENTRY(PGM_TYPE_EPT, PGM_TYPE_PROT, PGM_BTH_NAME_EPT_PROT),
877	PGMMODEDATABTH_ENTRY(PGM_TYPE_EPT, PGM_TYPE_32BIT, PGM_BTH_NAME_EPT_32BIT),
878	PGMMODEDATABTH_ENTRY(PGM_TYPE_EPT, PGM_TYPE_PAE, PGM_BTH_NAME_EPT_PAE),
879	#ifdef VBOX_WITH_64_BITS_GUESTS
880	PGMMODEDATABTH_ENTRY(PGM_TYPE_EPT, PGM_TYPE_AMD64, PGM_BTH_NAME_EPT_AMD64),
881	#else
882	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_EPT, PGM_TYPE_AMD64 - illegal */
883	#endif
884	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_EPT, PGM_TYPE_NESTED_32BIT - illegal */
885	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_EPT, PGM_TYPE_NESTED_PAE - illegal */
886	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_EPT, PGM_TYPE_NESTED_AMD64 - illegal */
887	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_EPT, PGM_TYPE_EPT - illegal */
888	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_EPT, PGM_TYPE_NONE - illegal */
889
890	/* NONE / NEM: */
891	PGMMODEDATABTH_NULL_ENTRY(), /* 0 */
892	PGMMODEDATABTH_ENTRY(PGM_TYPE_NONE, PGM_TYPE_REAL, PGM_BTH_NAME_EPT_REAL),
893	PGMMODEDATABTH_ENTRY(PGM_TYPE_NONE, PGM_TYPE_PROT, PGM_BTH_NAME_EPT_PROT),
894	PGMMODEDATABTH_ENTRY(PGM_TYPE_NONE, PGM_TYPE_32BIT, PGM_BTH_NAME_EPT_32BIT),
895	PGMMODEDATABTH_ENTRY(PGM_TYPE_NONE, PGM_TYPE_PAE, PGM_BTH_NAME_EPT_PAE),
896	#ifdef VBOX_WITH_64_BITS_GUESTS
897	PGMMODEDATABTH_ENTRY(PGM_TYPE_NONE, PGM_TYPE_AMD64, PGM_BTH_NAME_EPT_AMD64),
898	#else
899	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_NONE, PGM_TYPE_AMD64 - illegal */
900	#endif
901	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_NONE, PGM_TYPE_NESTED_32BIT - illegal */
902	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_NONE, PGM_TYPE_NESTED_PAE - illegal */
903	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_NONE, PGM_TYPE_NESTED_AMD64 - illegal */
904	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_NONE, PGM_TYPE_EPT - illegal */
905	PGMMODEDATABTH_NULL_ENTRY(), /* PGM_TYPE_NONE, PGM_TYPE_NONE - illegal */
906
907
908	#undef PGMMODEDATABTH_ENTRY
909	#undef PGMMODEDATABTH_NULL_ENTRY
910	};
911
912
913	/** Mask array used by pgmGetCr3MaskForMode.
914	* X86_CR3_AMD64_PAGE_MASK is used for modes that doesn't have a CR3 or EPTP. */
915	static uint64_t const g_auCr3MaskForMode[PGMMODE_MAX] =
916	{
917	/* [PGMMODE_INVALID] = */ X86_CR3_AMD64_PAGE_MASK,
918	/* [PGMMODE_REAL] = */ X86_CR3_AMD64_PAGE_MASK,
919	/* [PGMMODE_PROTECTED] = */ X86_CR3_AMD64_PAGE_MASK,
920	/* [PGMMODE_32_BIT] = */ X86_CR3_PAGE_MASK,
921	/* [PGMMODE_PAE] = */ X86_CR3_PAE_PAGE_MASK,
922	/* [PGMMODE_PAE_NX] = */ X86_CR3_PAE_PAGE_MASK,
923	/* [PGMMODE_AMD64] = */ X86_CR3_AMD64_PAGE_MASK,
924	/* [PGMMODE_AMD64_NX] = */ X86_CR3_AMD64_PAGE_MASK,
925	/* [PGMMODE_NESTED_32BIT = */ X86_CR3_PAGE_MASK,
926	/* [PGMMODE_NESTED_PAE] = */ X86_CR3_PAE_PAGE_MASK,
927	/* [PGMMODE_NESTED_AMD64] = */ X86_CR3_AMD64_PAGE_MASK,
928	/* [PGMMODE_EPT] = */ X86_CR3_EPT_PAGE_MASK,
929	/* [PGMMODE_NONE] = */ X86_CR3_AMD64_PAGE_MASK,
930	};
931
932
933	/**
934	* Gets the physical address mask for CR3 in the given paging mode.
935	*
936	* The mask is for eliminating flags and other stuff in CR3/EPTP when
937	* extracting the physical address. It is not for validating whether there are
938	* reserved bits set. PGM ASSUMES that whoever loaded the CR3 value and passed
939	* it to PGM checked for reserved bits, including reserved physical address
940	* bits.
941	*
942	* @returns The CR3 mask.
943	* @param enmMode The paging mode.
944	* @param enmSlatMode The second-level address translation mode.
945	*/
946	DECLINLINE(uint64_t) pgmGetCr3MaskForMode(PGMMODE enmMode, PGMSLAT enmSlatMode)
947	{
948	if (enmSlatMode == PGMSLAT_DIRECT)
949	{
950	Assert(enmMode != PGMMODE_EPT);
951	return g_auCr3MaskForMode[(unsigned)enmMode < (unsigned)PGMMODE_MAX ? enmMode : 0];
952	}
953	Assert(enmSlatMode == PGMSLAT_EPT);
954	return X86_CR3_EPT_PAGE_MASK;
955	}
956
957
958	/**
959	* Gets the masked CR3 value according to the current guest paging mode.
960	*
961	* See disclaimer in pgmGetCr3MaskForMode.
962	*
963	* @returns The masked PGM CR3 value.
964	* @param pVCpu The cross context virtual CPU structure.
965	* @param uCr3 The raw guest CR3 value.
966	*/
967	DECLINLINE(RTGCPHYS) pgmGetGuestMaskedCr3(PVMCPUCC pVCpu, uint64_t uCr3)
968	{
969	uint64_t const fCr3Mask = pgmGetCr3MaskForMode(pVCpu->pgm.s.enmGuestMode, pVCpu->pgm.s.enmGuestSlatMode);
970	RTGCPHYS GCPhysCR3 = (RTGCPHYS)(uCr3 & fCr3Mask);
971	PGM_A20_APPLY_TO_VAR(pVCpu, GCPhysCR3);
972	return GCPhysCR3;
973	}
974
975
976	#ifdef IN_RING0
977	/**
978	* #PF Handler.
979	*
980	* @returns VBox status code (appropriate for trap handling and GC return).
981	* @param pVCpu The cross context virtual CPU structure.
982	* @param uErr The trap error code.
983	* @param pCtx Pointer to the register context for the CPU.
984	* @param pvFault The fault address.
985	*/
986	VMMDECL(int) PGMTrap0eHandler(PVMCPUCC pVCpu, RTGCUINT uErr, PCPUMCTX pCtx, RTGCPTR pvFault)
987	{
988	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
989
990	Log(("PGMTrap0eHandler: uErr=%RGx pvFault=%RGv eip=%04x:%RGv cr3=%RGp\n", uErr, pvFault, pCtx->cs.Sel, (RTGCPTR)pCtx->rip, (RTGCPHYS)CPUMGetGuestCR3(pVCpu)));
991	STAM_PROFILE_START(&pVCpu->pgm.s.Stats.StatRZTrap0e, a);
992	STAM_STATS({ pVCpu->pgmr0.s.pStatTrap0eAttributionR0 = NULL; } );
993
994
995	# ifdef VBOX_WITH_STATISTICS
996	/*
997	* Error code stats.
998	*/
999	if (uErr & X86_TRAP_PF_US)
1000	{
1001	if (!(uErr & X86_TRAP_PF_P))
1002	{
1003	if (uErr & X86_TRAP_PF_RW)
1004	STAM_COUNTER_INC(&pVCpu->pgm.s.Stats.StatRZTrap0eUSNotPresentWrite);
1005	else
1006	STAM_COUNTER_INC(&pVCpu->pgm.s.Stats.StatRZTrap0eUSNotPresentRead);
1007	}
1008	else if (uErr & X86_TRAP_PF_RW)
1009	STAM_COUNTER_INC(&pVCpu->pgm.s.Stats.StatRZTrap0eUSWrite);
1010	else if (uErr & X86_TRAP_PF_RSVD)
1011	STAM_COUNTER_INC(&pVCpu->pgm.s.Stats.StatRZTrap0eUSReserved);
1012	else if (uErr & X86_TRAP_PF_ID)
1013	STAM_COUNTER_INC(&pVCpu->pgm.s.Stats.StatRZTrap0eUSNXE);
1014	else
1015	STAM_COUNTER_INC(&pVCpu->pgm.s.Stats.StatRZTrap0eUSRead);
1016	}
1017	else
1018	{ /* Supervisor */
1019	if (!(uErr & X86_TRAP_PF_P))
1020	{
1021	if (uErr & X86_TRAP_PF_RW)
1022	STAM_COUNTER_INC(&pVCpu->pgm.s.Stats.StatRZTrap0eSVNotPresentWrite);
1023	else
1024	STAM_COUNTER_INC(&pVCpu->pgm.s.Stats.StatRZTrap0eSVNotPresentRead);
1025	}
1026	else if (uErr & X86_TRAP_PF_RW)
1027	STAM_COUNTER_INC(&pVCpu->pgm.s.Stats.StatRZTrap0eSVWrite);
1028	else if (uErr & X86_TRAP_PF_ID)
1029	STAM_COUNTER_INC(&pVCpu->pgm.s.Stats.StatRZTrap0eSNXE);
1030	else if (uErr & X86_TRAP_PF_RSVD)
1031	STAM_COUNTER_INC(&pVCpu->pgm.s.Stats.StatRZTrap0eSVReserved);
1032	}
1033	# endif /* VBOX_WITH_STATISTICS */
1034
1035	/*
1036	* Call the worker.
1037	*/
1038	uintptr_t const idxBth = pVCpu->pgm.s.idxBothModeData;
1039	AssertReturn(idxBth < RT_ELEMENTS(g_aPgmBothModeData), VERR_PGM_MODE_IPE);
1040	AssertReturn(g_aPgmBothModeData[idxBth].pfnTrap0eHandler, VERR_PGM_MODE_IPE);
1041	bool fLockTaken = false;
1042	int rc = g_aPgmBothModeData[idxBth].pfnTrap0eHandler(pVCpu, uErr, pCtx, pvFault, &fLockTaken);
1043	if (fLockTaken)
1044	{
1045	PGM_LOCK_ASSERT_OWNER(pVM);
1046	PGM_UNLOCK(pVM);
1047	}
1048	LogFlow(("PGMTrap0eHandler: uErr=%RGx pvFault=%RGv rc=%Rrc\n", uErr, pvFault, rc));
1049
1050	/*
1051	* Return code tweaks.
1052	*/
1053	if (rc != VINF_SUCCESS)
1054	{
1055	if (rc == VINF_PGM_SYNCPAGE_MODIFIED_PDE)
1056	rc = VINF_SUCCESS;
1057
1058	/* Note: hack alert for difficult to reproduce problem. */
1059	if ( rc == VERR_PAGE_NOT_PRESENT /* SMP only ; disassembly might fail. */
1060	\|\| rc == VERR_PAGE_TABLE_NOT_PRESENT /* seen with UNI & SMP */
1061	\|\| rc == VERR_PAGE_DIRECTORY_PTR_NOT_PRESENT /* seen with SMP */
1062	\|\| rc == VERR_PAGE_MAP_LEVEL4_NOT_PRESENT) /* precaution */
1063	{
1064	Log(("WARNING: Unexpected VERR_PAGE_TABLE_NOT_PRESENT (%d) for page fault at %RGv error code %x (rip=%RGv)\n", rc, pvFault, uErr, pCtx->rip));
1065	/* Some kind of inconsistency in the SMP case; it's safe to just execute the instruction again; not sure about single VCPU VMs though. */
1066	rc = VINF_SUCCESS;
1067	}
1068	}
1069
1070	STAM_STATS({ if (rc == VINF_EM_RAW_GUEST_TRAP) STAM_COUNTER_INC(&pVCpu->pgm.s.Stats.StatRZTrap0eGuestPF); });
1071	STAM_STATS({ if (!pVCpu->pgmr0.s.pStatTrap0eAttributionR0)
1072	pVCpu->pgmr0.s.pStatTrap0eAttributionR0 = &pVCpu->pgm.s.Stats.StatRZTrap0eTime2Misc; });
1073	STAM_PROFILE_STOP_EX(&pVCpu->pgm.s.Stats.StatRZTrap0e, pVCpu->pgmr0.s.pStatTrap0eAttributionR0, a);
1074	return rc;
1075	}
1076	#endif /* IN_RING0 */
1077
1078
1079	/**
1080	* Prefetch a page
1081	*
1082	* Typically used to sync commonly used pages before entering raw mode
1083	* after a CR3 reload.
1084	*
1085	* @returns VBox status code suitable for scheduling.
1086	* @retval VINF_SUCCESS on success.
1087	* @retval VINF_PGM_SYNC_CR3 if we're out of shadow pages or something like that.
1088	* @param pVCpu The cross context virtual CPU structure.
1089	* @param GCPtrPage Page to invalidate.
1090	*/
1091	VMMDECL(int) PGMPrefetchPage(PVMCPUCC pVCpu, RTGCPTR GCPtrPage)
1092	{
1093	STAM_PROFILE_START(&pVCpu->pgm.s.Stats.CTX_MID_Z(Stat,Prefetch), a);
1094
1095	uintptr_t const idxBth = pVCpu->pgm.s.idxBothModeData;
1096	AssertReturn(idxBth < RT_ELEMENTS(g_aPgmBothModeData), VERR_PGM_MODE_IPE);
1097	AssertReturn(g_aPgmBothModeData[idxBth].pfnPrefetchPage, VERR_PGM_MODE_IPE);
1098	int rc = g_aPgmBothModeData[idxBth].pfnPrefetchPage(pVCpu, GCPtrPage);
1099
1100	STAM_PROFILE_STOP(&pVCpu->pgm.s.Stats.CTX_MID_Z(Stat,Prefetch), a);
1101	AssertMsg(rc == VINF_SUCCESS \|\| rc == VINF_PGM_SYNC_CR3 \|\| RT_FAILURE(rc), ("rc=%Rrc\n", rc));
1102	return rc;
1103	}
1104
1105
1106	/**
1107	* Emulation of the invlpg instruction (HC only actually).
1108	*
1109	* @returns Strict VBox status code, special care required.
1110	* @retval VINF_PGM_SYNC_CR3 - handled.
1111	* @retval VINF_EM_RAW_EMULATE_INSTR - not handled (RC only).
1112	*
1113	* @param pVCpu The cross context virtual CPU structure.
1114	* @param GCPtrPage Page to invalidate.
1115	*
1116	* @remark ASSUMES the page table entry or page directory is valid. Fairly
1117	* safe, but there could be edge cases!
1118	*
1119	* @todo Flush page or page directory only if necessary!
1120	* @todo VBOXSTRICTRC
1121	*/
1122	VMMDECL(int) PGMInvalidatePage(PVMCPUCC pVCpu, RTGCPTR GCPtrPage)
1123	{
1124	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
1125	int rc;
1126	Log3(("PGMInvalidatePage: GCPtrPage=%RGv\n", GCPtrPage));
1127
1128	IEMTlbInvalidatePage(pVCpu, GCPtrPage);
1129
1130	/*
1131	* Call paging mode specific worker.
1132	*/
1133	STAM_PROFILE_START(&pVCpu->pgm.s.Stats.CTX_MID_Z(Stat,InvalidatePage), a);
1134	PGM_LOCK_VOID(pVM);
1135
1136	uintptr_t const idxBth = pVCpu->pgm.s.idxBothModeData;
1137	AssertReturnStmt(idxBth < RT_ELEMENTS(g_aPgmBothModeData), PGM_UNLOCK(pVM), VERR_PGM_MODE_IPE);
1138	AssertReturnStmt(g_aPgmBothModeData[idxBth].pfnInvalidatePage, PGM_UNLOCK(pVM), VERR_PGM_MODE_IPE);
1139	rc = g_aPgmBothModeData[idxBth].pfnInvalidatePage(pVCpu, GCPtrPage);
1140
1141	PGM_UNLOCK(pVM);
1142	STAM_PROFILE_STOP(&pVCpu->pgm.s.Stats.CTX_MID_Z(Stat,InvalidatePage), a);
1143
1144	/* Ignore all irrelevant error codes. */
1145	if ( rc == VERR_PAGE_NOT_PRESENT
1146	\|\| rc == VERR_PAGE_TABLE_NOT_PRESENT
1147	\|\| rc == VERR_PAGE_DIRECTORY_PTR_NOT_PRESENT
1148	\|\| rc == VERR_PAGE_MAP_LEVEL4_NOT_PRESENT)
1149	rc = VINF_SUCCESS;
1150
1151	return rc;
1152	}
1153
1154
1155	/**
1156	* Executes an instruction using the interpreter.
1157	*
1158	* @returns VBox status code (appropriate for trap handling and GC return).
1159	* @param pVCpu The cross context virtual CPU structure.
1160	* @param pvFault Fault address.
1161	*/
1162	VMMDECL(VBOXSTRICTRC) PGMInterpretInstruction(PVMCPUCC pVCpu, RTGCPTR pvFault)
1163	{
1164	RT_NOREF(pvFault);
1165	VBOXSTRICTRC rc = EMInterpretInstruction(pVCpu);
1166	if (rc == VERR_EM_INTERPRETER)
1167	rc = VINF_EM_RAW_EMULATE_INSTR;
1168	if (rc != VINF_SUCCESS)
1169	Log(("PGMInterpretInstruction: returns %Rrc (pvFault=%RGv)\n", VBOXSTRICTRC_VAL(rc), pvFault));
1170	return rc;
1171	}
1172
1173
1174	/**
1175	* Gets effective page information (from the VMM page directory).
1176	*
1177	* @returns VBox status code.
1178	* @param pVCpu The cross context virtual CPU structure.
1179	* @param GCPtr Guest Context virtual address of the page.
1180	* @param pfFlags Where to store the flags. These are X86_PTE_*.
1181	* @param pHCPhys Where to store the HC physical address of the page.
1182	* This is page aligned.
1183	* @remark You should use PGMMapGetPage() for pages in a mapping.
1184	*/
1185	VMMDECL(int) PGMShwGetPage(PVMCPUCC pVCpu, RTGCPTR GCPtr, uint64_t *pfFlags, PRTHCPHYS pHCPhys)
1186	{
1187	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
1188	PGM_LOCK_VOID(pVM);
1189
1190	uintptr_t idxShw = pVCpu->pgm.s.idxShadowModeData;
1191	AssertReturn(idxShw < RT_ELEMENTS(g_aPgmShadowModeData), VERR_PGM_MODE_IPE);
1192	AssertReturn(g_aPgmShadowModeData[idxShw].pfnGetPage, VERR_PGM_MODE_IPE);
1193	int rc = g_aPgmShadowModeData[idxShw].pfnGetPage(pVCpu, GCPtr, pfFlags, pHCPhys);
1194
1195	PGM_UNLOCK(pVM);
1196	return rc;
1197	}
1198
1199
1200	/**
1201	* Modify page flags for a range of pages in the shadow context.
1202	*
1203	* The existing flags are ANDed with the fMask and ORed with the fFlags.
1204	*
1205	* @returns VBox status code.
1206	* @param pVCpu The cross context virtual CPU structure.
1207	* @param GCPtr Virtual address of the first page in the range.
1208	* @param fFlags The OR mask - page flags X86_PTE_*, excluding the page mask of course.
1209	* @param fMask The AND mask - page flags X86_PTE_*.
1210	* Be very CAREFUL when ~'ing constants which could be 32-bit!
1211	* @param fOpFlags A combination of the PGM_MK_PK_XXX flags.
1212	* @remark You must use PGMMapModifyPage() for pages in a mapping.
1213	*/
1214	DECLINLINE(int) pdmShwModifyPage(PVMCPUCC pVCpu, RTGCPTR GCPtr, uint64_t fFlags, uint64_t fMask, uint32_t fOpFlags)
1215	{
1216	AssertMsg(!(fFlags & X86_PTE_PAE_PG_MASK), ("fFlags=%#llx\n", fFlags));
1217	Assert(!(fOpFlags & ~(PGM_MK_PG_IS_MMIO2 \| PGM_MK_PG_IS_WRITE_FAULT)));
1218
1219	GCPtr &= ~(RTGCPTR)GUEST_PAGE_OFFSET_MASK; /** @todo this ain't necessary, right... */
1220
1221	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
1222	PGM_LOCK_VOID(pVM);
1223
1224	uintptr_t idxShw = pVCpu->pgm.s.idxShadowModeData;
1225	AssertReturn(idxShw < RT_ELEMENTS(g_aPgmShadowModeData), VERR_PGM_MODE_IPE);
1226	AssertReturn(g_aPgmShadowModeData[idxShw].pfnModifyPage, VERR_PGM_MODE_IPE);
1227	int rc = g_aPgmShadowModeData[idxShw].pfnModifyPage(pVCpu, GCPtr, GUEST_PAGE_SIZE, fFlags, fMask, fOpFlags);
1228
1229	PGM_UNLOCK(pVM);
1230	return rc;
1231	}
1232
1233
1234	/**
1235	* Changing the page flags for a single page in the shadow page tables so as to
1236	* make it read-only.
1237	*
1238	* @returns VBox status code.
1239	* @param pVCpu The cross context virtual CPU structure.
1240	* @param GCPtr Virtual address of the first page in the range.
1241	* @param fOpFlags A combination of the PGM_MK_PK_XXX flags.
1242	*/
1243	VMMDECL(int) PGMShwMakePageReadonly(PVMCPUCC pVCpu, RTGCPTR GCPtr, uint32_t fOpFlags)
1244	{
1245	return pdmShwModifyPage(pVCpu, GCPtr, 0, ~(uint64_t)X86_PTE_RW, fOpFlags);
1246	}
1247
1248
1249	/**
1250	* Changing the page flags for a single page in the shadow page tables so as to
1251	* make it writable.
1252	*
1253	* The call must know with 101% certainty that the guest page tables maps this
1254	* as writable too. This function will deal shared, zero and write monitored
1255	* pages.
1256	*
1257	* @returns VBox status code.
1258	* @param pVCpu The cross context virtual CPU structure.
1259	* @param GCPtr Virtual address of the first page in the range.
1260	* @param fOpFlags A combination of the PGM_MK_PK_XXX flags.
1261	*/
1262	VMMDECL(int) PGMShwMakePageWritable(PVMCPUCC pVCpu, RTGCPTR GCPtr, uint32_t fOpFlags)
1263	{
1264	if (pVCpu->pgm.s.enmShadowMode != PGMMODE_NONE) /* avoid assertions */
1265	return pdmShwModifyPage(pVCpu, GCPtr, X86_PTE_RW, ~(uint64_t)0, fOpFlags);
1266	return VINF_SUCCESS;
1267	}
1268
1269
1270	/**
1271	* Changing the page flags for a single page in the shadow page tables so as to
1272	* make it not present.
1273	*
1274	* @returns VBox status code.
1275	* @param pVCpu The cross context virtual CPU structure.
1276	* @param GCPtr Virtual address of the first page in the range.
1277	* @param fOpFlags A combination of the PGM_MK_PG_XXX flags.
1278	*/
1279	VMMDECL(int) PGMShwMakePageNotPresent(PVMCPUCC pVCpu, RTGCPTR GCPtr, uint32_t fOpFlags)
1280	{
1281	return pdmShwModifyPage(pVCpu, GCPtr, 0, 0, fOpFlags);
1282	}
1283
1284
1285	/**
1286	* Changing the page flags for a single page in the shadow page tables so as to
1287	* make it supervisor and writable.
1288	*
1289	* This if for dealing with CR0.WP=0 and readonly user pages.
1290	*
1291	* @returns VBox status code.
1292	* @param pVCpu The cross context virtual CPU structure.
1293	* @param GCPtr Virtual address of the first page in the range.
1294	* @param fBigPage Whether or not this is a big page. If it is, we have to
1295	* change the shadow PDE as well. If it isn't, the caller
1296	* has checked that the shadow PDE doesn't need changing.
1297	* We ASSUME 4KB pages backing the big page here!
1298	* @param fOpFlags A combination of the PGM_MK_PG_XXX flags.
1299	*/
1300	int pgmShwMakePageSupervisorAndWritable(PVMCPUCC pVCpu, RTGCPTR GCPtr, bool fBigPage, uint32_t fOpFlags)
1301	{
1302	int rc = pdmShwModifyPage(pVCpu, GCPtr, X86_PTE_RW, ~(uint64_t)X86_PTE_US, fOpFlags);
1303	if (rc == VINF_SUCCESS && fBigPage)
1304	{
1305	/* this is a bit ugly... */
1306	switch (pVCpu->pgm.s.enmShadowMode)
1307	{
1308	case PGMMODE_32_BIT:
1309	{
1310	PX86PDE pPde = pgmShwGet32BitPDEPtr(pVCpu, GCPtr);
1311	AssertReturn(pPde, VERR_INTERNAL_ERROR_3);
1312	Log(("pgmShwMakePageSupervisorAndWritable: PDE=%#llx", pPde->u));
1313	pPde->u \|= X86_PDE_RW;
1314	Log(("-> PDE=%#llx (32)\n", pPde->u));
1315	break;
1316	}
1317	case PGMMODE_PAE:
1318	case PGMMODE_PAE_NX:
1319	{
1320	PX86PDEPAE pPde = pgmShwGetPaePDEPtr(pVCpu, GCPtr);
1321	AssertReturn(pPde, VERR_INTERNAL_ERROR_3);
1322	Log(("pgmShwMakePageSupervisorAndWritable: PDE=%#llx", pPde->u));
1323	pPde->u \|= X86_PDE_RW;
1324	Log(("-> PDE=%#llx (PAE)\n", pPde->u));
1325	break;
1326	}
1327	default:
1328	AssertFailedReturn(VERR_INTERNAL_ERROR_4);
1329	}
1330	}
1331	return rc;
1332	}
1333
1334
1335	/**
1336	* Gets the shadow page directory for the specified address, PAE.
1337	*
1338	* @returns Pointer to the shadow PD.
1339	* @param pVCpu The cross context virtual CPU structure.
1340	* @param GCPtr The address.
1341	* @param uGstPdpe Guest PDPT entry. Valid.
1342	* @param ppPD Receives address of page directory
1343	*/
1344	int pgmShwSyncPaePDPtr(PVMCPUCC pVCpu, RTGCPTR GCPtr, X86PGPAEUINT uGstPdpe, PX86PDPAE *ppPD)
1345	{
1346	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
1347	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
1348	PPGMPOOLPAGE pShwPage;
1349	int rc;
1350	PGM_LOCK_ASSERT_OWNER(pVM);
1351
1352
1353	/* Allocate page directory if not present. */
1354	const unsigned iPdPt = (GCPtr >> X86_PDPT_SHIFT) & X86_PDPT_MASK_PAE;
1355	PX86PDPT pPdpt = pgmShwGetPaePDPTPtr(pVCpu);
1356	PX86PDPE pPdpe = &pPdpt->a[iPdPt];
1357	X86PGPAEUINT const uPdpe = pPdpe->u;
1358	if (uPdpe & (X86_PDPE_P \| X86_PDPE_PG_MASK))
1359	{
1360	pShwPage = pgmPoolGetPage(pPool, uPdpe & X86_PDPE_PG_MASK);
1361	AssertReturn(pShwPage, VERR_PGM_POOL_GET_PAGE_FAILED);
1362	Assert((pPdpe->u & X86_PDPE_PG_MASK) == pShwPage->Core.Key);
1363
1364	pgmPoolCacheUsed(pPool, pShwPage);
1365
1366	/* Update the entry if necessary. */
1367	X86PGPAEUINT const uPdpeNew = pShwPage->Core.Key \| (uGstPdpe & (X86_PDPE_P \| X86_PDPE_A)) \| (uPdpe & PGM_PDPT_FLAGS);
1368	if (uPdpeNew == uPdpe)
1369	{ /* likely */ }
1370	else
1371	ASMAtomicWriteU64(&pPdpe->u, uPdpeNew);
1372	}
1373	else
1374	{
1375	RTGCPTR64 GCPdPt;
1376	PGMPOOLKIND enmKind;
1377	if (pVM->pgm.s.fNestedPaging \|\| !CPUMIsGuestPagingEnabled(pVCpu))
1378	{
1379	/* AMD-V nested paging or real/protected mode without paging. */
1380	GCPdPt = GCPtr & ~(RT_BIT_64(X86_PDPT_SHIFT) - 1);
1381	enmKind = PGMPOOLKIND_PAE_PD_PHYS;
1382	}
1383	else if (CPUMGetGuestCR4(pVCpu) & X86_CR4_PAE)
1384	{
1385	if (uGstPdpe & X86_PDPE_P)
1386	{
1387	GCPdPt = uGstPdpe & X86_PDPE_PG_MASK;
1388	enmKind = PGMPOOLKIND_PAE_PD_FOR_PAE_PD;
1389	}
1390	else
1391	{
1392	/* PD not present; guest must reload CR3 to change it.
1393	* No need to monitor anything in this case. */
1394	/** @todo r=bird: WTF is hit?!? */
1395	/Assert(VM_IS_RAW_MODE_ENABLED(pVM)); - ??? /
1396	GCPdPt = uGstPdpe & X86_PDPE_PG_MASK;
1397	enmKind = PGMPOOLKIND_PAE_PD_PHYS;
1398	Assert(uGstPdpe & X86_PDPE_P); /* caller should do this already */
1399	}
1400	}
1401	else
1402	{
1403	GCPdPt = CPUMGetGuestCR3(pVCpu);
1404	enmKind = (PGMPOOLKIND)(PGMPOOLKIND_PAE_PD0_FOR_32BIT_PD + iPdPt);
1405	}
1406
1407	/* Create a reference back to the PDPT by using the index in its shadow page. */
1408	rc = pgmPoolAlloc(pVM, GCPdPt, enmKind, PGMPOOLACCESS_DONTCARE, PGM_A20_IS_ENABLED(pVCpu),
1409	pVCpu->pgm.s.CTX_SUFF(pShwPageCR3)->idx, iPdPt, false /fLockPage/,
1410	&pShwPage);
1411	AssertRCReturn(rc, rc);
1412
1413	/* Hook it up. */
1414	ASMAtomicWriteU64(&pPdpe->u, pShwPage->Core.Key \| (uGstPdpe & (X86_PDPE_P \| X86_PDPE_A)) \| (uPdpe & PGM_PDPT_FLAGS));
1415	}
1416	PGM_DYNMAP_UNUSED_HINT(pVCpu, pPdpe);
1417
1418	*ppPD = (PX86PDPAE)PGMPOOL_PAGE_2_PTR_V2(pVM, pVCpu, pShwPage);
1419	return VINF_SUCCESS;
1420	}
1421
1422
1423	/**
1424	* Gets the pointer to the shadow page directory entry for an address, PAE.
1425	*
1426	* @returns Pointer to the PDE.
1427	* @param pVCpu The cross context virtual CPU structure of the calling EMT.
1428	* @param GCPtr The address.
1429	* @param ppShwPde Receives the address of the pgm pool page for the shadow page directory
1430	*/
1431	DECLINLINE(int) pgmShwGetPaePoolPagePD(PVMCPUCC pVCpu, RTGCPTR GCPtr, PPGMPOOLPAGE *ppShwPde)
1432	{
1433	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
1434	PGM_LOCK_ASSERT_OWNER(pVM);
1435
1436	PX86PDPT pPdpt = pgmShwGetPaePDPTPtr(pVCpu);
1437	AssertReturn(pPdpt, VERR_PAGE_DIRECTORY_PTR_NOT_PRESENT); /* can't happen */
1438	const unsigned iPdPt = (GCPtr >> X86_PDPT_SHIFT) & X86_PDPT_MASK_PAE;
1439	X86PGPAEUINT const uPdpe = pPdpt->a[iPdPt].u;
1440	if (!(uPdpe & X86_PDPE_P))
1441	{
1442	LogFlow(("pgmShwGetPaePoolPagePD: PD %d not present (%RX64)\n", iPdPt, uPdpe));
1443	return VERR_PAGE_DIRECTORY_PTR_NOT_PRESENT;
1444	}
1445	AssertMsg(uPdpe & X86_PDPE_PG_MASK, ("GCPtr=%RGv\n", GCPtr));
1446
1447	/* Fetch the pgm pool shadow descriptor. */
1448	PPGMPOOLPAGE pShwPde = pgmPoolGetPage(pVM->pgm.s.CTX_SUFF(pPool), uPdpe & X86_PDPE_PG_MASK);
1449	AssertReturn(pShwPde, VERR_PGM_POOL_GET_PAGE_FAILED);
1450
1451	*ppShwPde = pShwPde;
1452	return VINF_SUCCESS;
1453	}
1454
1455
1456	/**
1457	* Syncs the SHADOW page directory pointer for the specified address.
1458	*
1459	* Allocates backing pages in case the PDPT or PML4 entry is missing.
1460	*
1461	* The caller is responsible for making sure the guest has a valid PD before
1462	* calling this function.
1463	*
1464	* @returns VBox status code.
1465	* @param pVCpu The cross context virtual CPU structure.
1466	* @param GCPtr The address.
1467	* @param uGstPml4e Guest PML4 entry (valid).
1468	* @param uGstPdpe Guest PDPT entry (valid).
1469	* @param ppPD Receives address of page directory
1470	*/
1471	static int pgmShwSyncLongModePDPtr(PVMCPUCC pVCpu, RTGCPTR64 GCPtr, X86PGPAEUINT uGstPml4e, X86PGPAEUINT uGstPdpe, PX86PDPAE *ppPD)
1472	{
1473	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
1474	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
1475	bool const fNestedPagingOrNoGstPaging = pVM->pgm.s.fNestedPaging \|\| !CPUMIsGuestPagingEnabled(pVCpu);
1476	int rc;
1477
1478	PGM_LOCK_ASSERT_OWNER(pVM);
1479
1480	/*
1481	* PML4.
1482	*/
1483	PPGMPOOLPAGE pShwPage;
1484	{
1485	const unsigned iPml4 = (GCPtr >> X86_PML4_SHIFT) & X86_PML4_MASK;
1486	PX86PML4E pPml4e = pgmShwGetLongModePML4EPtr(pVCpu, iPml4);
1487	AssertReturn(pPml4e, VERR_PGM_PML4_MAPPING);
1488	X86PGPAEUINT const uPml4e = pPml4e->u;
1489
1490	/* Allocate page directory pointer table if not present. */
1491	if (uPml4e & (X86_PML4E_P \| X86_PML4E_PG_MASK))
1492	{
1493	pShwPage = pgmPoolGetPage(pPool, uPml4e & X86_PML4E_PG_MASK);
1494	AssertReturn(pShwPage, VERR_PGM_POOL_GET_PAGE_FAILED);
1495
1496	pgmPoolCacheUsed(pPool, pShwPage);
1497
1498	/* Update the entry if needed. */
1499	X86PGPAEUINT const uPml4eNew = pShwPage->Core.Key \| (uGstPml4e & pVCpu->pgm.s.fGstAmd64ShadowedPml4eMask)
1500	\| (uPml4e & PGM_PML4_FLAGS);
1501	if (uPml4e == uPml4eNew)
1502	{ /* likely */ }
1503	else
1504	ASMAtomicWriteU64(&pPml4e->u, uPml4eNew);
1505	}
1506	else
1507	{
1508	Assert(pVCpu->pgm.s.CTX_SUFF(pShwPageCR3));
1509
1510	RTGCPTR64 GCPml4;
1511	PGMPOOLKIND enmKind;
1512	if (fNestedPagingOrNoGstPaging)
1513	{
1514	/* AMD-V nested paging or real/protected mode without paging */
1515	GCPml4 = (RTGCPTR64)iPml4 << X86_PML4_SHIFT; /** @todo bogus calculation for PML5 */
1516	enmKind = PGMPOOLKIND_64BIT_PDPT_FOR_PHYS;
1517	}
1518	else
1519	{
1520	GCPml4 = uGstPml4e & X86_PML4E_PG_MASK;
1521	enmKind = PGMPOOLKIND_64BIT_PDPT_FOR_64BIT_PDPT;
1522	}
1523
1524	/* Create a reference back to the PDPT by using the index in its shadow page. */
1525	rc = pgmPoolAlloc(pVM, GCPml4, enmKind, PGMPOOLACCESS_DONTCARE, PGM_A20_IS_ENABLED(pVCpu),
1526	pVCpu->pgm.s.CTX_SUFF(pShwPageCR3)->idx, iPml4, false /fLockPage/,
1527	&pShwPage);
1528	AssertRCReturn(rc, rc);
1529
1530	/* Hook it up. */
1531	ASMAtomicWriteU64(&pPml4e->u, pShwPage->Core.Key \| (uGstPml4e & pVCpu->pgm.s.fGstAmd64ShadowedPml4eMask)
1532	\| (uPml4e & PGM_PML4_FLAGS));
1533	}
1534	}
1535
1536	/*
1537	* PDPT.
1538	*/
1539	const unsigned iPdPt = (GCPtr >> X86_PDPT_SHIFT) & X86_PDPT_MASK_AMD64;
1540	PX86PDPT pPdpt = (PX86PDPT)PGMPOOL_PAGE_2_PTR_V2(pVM, pVCpu, pShwPage);
1541	PX86PDPE pPdpe = &pPdpt->a[iPdPt];
1542	X86PGPAEUINT const uPdpe = pPdpe->u;
1543
1544	/* Allocate page directory if not present. */
1545	if (uPdpe & (X86_PDPE_P \| X86_PDPE_PG_MASK))
1546	{
1547	pShwPage = pgmPoolGetPage(pPool, uPdpe & X86_PDPE_PG_MASK);
1548	AssertReturn(pShwPage, VERR_PGM_POOL_GET_PAGE_FAILED);
1549
1550	pgmPoolCacheUsed(pPool, pShwPage);
1551
1552	/* Update the entry if needed. */
1553	X86PGPAEUINT const uPdpeNew = pShwPage->Core.Key \| (uGstPdpe & pVCpu->pgm.s.fGstAmd64ShadowedPdpeMask)
1554	\| (uPdpe & PGM_PDPT_FLAGS);
1555	if (uPdpe == uPdpeNew)
1556	{ /* likely */ }
1557	else
1558	ASMAtomicWriteU64(&pPdpe->u, uPdpeNew);
1559	}
1560	else
1561	{
1562	RTGCPTR64 GCPdPt;
1563	PGMPOOLKIND enmKind;
1564	if (fNestedPagingOrNoGstPaging)
1565	{
1566	/* AMD-V nested paging or real/protected mode without paging */
1567	GCPdPt = GCPtr & ~(RT_BIT_64(iPdPt << X86_PDPT_SHIFT) - 1);
1568	enmKind = PGMPOOLKIND_64BIT_PD_FOR_PHYS;
1569	}
1570	else
1571	{
1572	GCPdPt = uGstPdpe & X86_PDPE_PG_MASK;
1573	enmKind = PGMPOOLKIND_64BIT_PD_FOR_64BIT_PD;
1574	}
1575
1576	/* Create a reference back to the PDPT by using the index in its shadow page. */
1577	rc = pgmPoolAlloc(pVM, GCPdPt, enmKind, PGMPOOLACCESS_DONTCARE, PGM_A20_IS_ENABLED(pVCpu),
1578	pShwPage->idx, iPdPt, false /fLockPage/,
1579	&pShwPage);
1580	AssertRCReturn(rc, rc);
1581
1582	/* Hook it up. */
1583	ASMAtomicWriteU64(&pPdpe->u,
1584	pShwPage->Core.Key \| (uGstPdpe & pVCpu->pgm.s.fGstAmd64ShadowedPdpeMask) \| (uPdpe & PGM_PDPT_FLAGS));
1585	}
1586
1587	*ppPD = (PX86PDPAE)PGMPOOL_PAGE_2_PTR_V2(pVM, pVCpu, pShwPage);
1588	return VINF_SUCCESS;
1589	}
1590
1591
1592	/**
1593	* Gets the SHADOW page directory pointer for the specified address (long mode).
1594	*
1595	* @returns VBox status code.
1596	* @param pVCpu The cross context virtual CPU structure.
1597	* @param GCPtr The address.
1598	* @param ppPml4e Receives the address of the page map level 4 entry.
1599	* @param ppPdpt Receives the address of the page directory pointer table.
1600	* @param ppPD Receives the address of the page directory.
1601	*/
1602	DECLINLINE(int) pgmShwGetLongModePDPtr(PVMCPUCC pVCpu, RTGCPTR64 GCPtr, PX86PML4E ppPml4e, PX86PDPT ppPdpt, PX86PDPAE *ppPD)
1603	{
1604	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
1605	PGM_LOCK_ASSERT_OWNER(pVM);
1606
1607	/*
1608	* PML4
1609	*/
1610	const unsigned iPml4 = (GCPtr >> X86_PML4_SHIFT) & X86_PML4_MASK;
1611	PCX86PML4E pPml4e = pgmShwGetLongModePML4EPtr(pVCpu, iPml4);
1612	AssertReturn(pPml4e, VERR_PGM_PML4_MAPPING);
1613	if (ppPml4e)
1614	*ppPml4e = (PX86PML4E)pPml4e;
1615	X86PGPAEUINT const uPml4e = pPml4e->u;
1616	Log4(("pgmShwGetLongModePDPtr %RGv (%RHv) %RX64\n", GCPtr, pPml4e, uPml4e));
1617	if (!(uPml4e & X86_PML4E_P)) /** @todo other code is check for NULL page frame number! */
1618	return VERR_PAGE_MAP_LEVEL4_NOT_PRESENT;
1619
1620	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
1621	PPGMPOOLPAGE pShwPage = pgmPoolGetPage(pPool, uPml4e & X86_PML4E_PG_MASK);
1622	AssertReturn(pShwPage, VERR_PGM_POOL_GET_PAGE_FAILED);
1623
1624	/*
1625	* PDPT
1626	*/
1627	const unsigned iPdPt = (GCPtr >> X86_PDPT_SHIFT) & X86_PDPT_MASK_AMD64;
1628	PCX86PDPT pPdpt = *ppPdpt = (PX86PDPT)PGMPOOL_PAGE_2_PTR_V2(pVM, pVCpu, pShwPage);
1629	X86PGPAEUINT const uPdpe = pPdpt->a[iPdPt].u;
1630	if (!(uPdpe & X86_PDPE_P)) /** @todo other code is check for NULL page frame number! */
1631	return VERR_PAGE_DIRECTORY_PTR_NOT_PRESENT;
1632
1633	pShwPage = pgmPoolGetPage(pPool, uPdpe & X86_PDPE_PG_MASK);
1634	AssertReturn(pShwPage, VERR_PGM_POOL_GET_PAGE_FAILED);
1635
1636	*ppPD = (PX86PDPAE)PGMPOOL_PAGE_2_PTR_V2(pVM, pVCpu, pShwPage);
1637	Log4(("pgmShwGetLongModePDPtr %RGv -> ppPD=%p PDE=%p/%RX64\n", GCPtr, ppPD, &(ppPD)->a[(GCPtr >> X86_PD_PAE_SHIFT) & X86_PD_PAE_MASK], (ppPD)->a[(GCPtr >> X86_PD_PAE_SHIFT) & X86_PD_PAE_MASK].u));
1638	return VINF_SUCCESS;
1639	}
1640
1641
1642	/**
1643	* Syncs the SHADOW EPT page directory pointer for the specified address. Allocates
1644	* backing pages in case the PDPT or PML4 entry is missing.
1645	*
1646	* @returns VBox status code.
1647	* @param pVCpu The cross context virtual CPU structure.
1648	* @param GCPtr The address.
1649	* @param ppPdpt Receives address of pdpt
1650	* @param ppPD Receives address of page directory
1651	*/
1652	static int pgmShwGetEPTPDPtr(PVMCPUCC pVCpu, RTGCPTR64 GCPtr, PEPTPDPT ppPdpt, PEPTPD ppPD)
1653	{
1654	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
1655	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
1656	int rc;
1657
1658	Assert(pVM->pgm.s.fNestedPaging);
1659	PGM_LOCK_ASSERT_OWNER(pVM);
1660
1661	/*
1662	* PML4 level.
1663	*/
1664	PEPTPML4 pPml4 = (PEPTPML4)PGMPOOL_PAGE_2_PTR_V2(pVM, pVCpu, pVCpu->pgm.s.CTX_SUFF(pShwPageCR3));
1665	Assert(pPml4);
1666
1667	/* Allocate page directory pointer table if not present. */
1668	PPGMPOOLPAGE pShwPage;
1669	{
1670	const unsigned iPml4 = (GCPtr >> EPT_PML4_SHIFT) & EPT_PML4_MASK;
1671	PEPTPML4E pPml4e = &pPml4->a[iPml4];
1672	EPTPML4E Pml4e;
1673	Pml4e.u = pPml4e->u;
1674	if (!(Pml4e.u & (EPT_E_PG_MASK \| EPT_E_READ)))
1675	{
1676	RTGCPTR64 GCPml4 = (RTGCPTR64)iPml4 << EPT_PML4_SHIFT;
1677	rc = pgmPoolAlloc(pVM, GCPml4, PGMPOOLKIND_EPT_PDPT_FOR_PHYS, PGMPOOLACCESS_DONTCARE, PGM_A20_IS_ENABLED(pVCpu),
1678	pVCpu->pgm.s.CTX_SUFF(pShwPageCR3)->idx, iPml4, false /fLockPage/,
1679	&pShwPage);
1680	AssertRCReturn(rc, rc);
1681
1682	/* Hook up the new PDPT now. */
1683	ASMAtomicWriteU64(&pPml4e->u, pShwPage->Core.Key \| EPT_E_READ \| EPT_E_WRITE \| EPT_E_EXECUTE);
1684	}
1685	else
1686	{
1687	pShwPage = pgmPoolGetPage(pPool, pPml4e->u & EPT_PML4E_PG_MASK);
1688	AssertReturn(pShwPage, VERR_PGM_POOL_GET_PAGE_FAILED);
1689
1690	pgmPoolCacheUsed(pPool, pShwPage);
1691
1692	/* Hook up the cached PDPT if needed (probably not given 512512 PTs to sync). /
1693	if (Pml4e.u == (pShwPage->Core.Key \| EPT_E_READ \| EPT_E_WRITE \| EPT_E_EXECUTE))
1694	{ }
1695	else
1696	ASMAtomicWriteU64(&pPml4e->u, pShwPage->Core.Key \| EPT_E_READ \| EPT_E_WRITE \| EPT_E_EXECUTE);
1697	}
1698	}
1699
1700	/*
1701	* PDPT level.
1702	*/
1703	const unsigned iPdPt = (GCPtr >> EPT_PDPT_SHIFT) & EPT_PDPT_MASK;
1704	PEPTPDPT pPdpt = (PEPTPDPT)PGMPOOL_PAGE_2_PTR_V2(pVM, pVCpu, pShwPage);
1705	PEPTPDPTE pPdpe = &pPdpt->a[iPdPt];
1706
1707	if (ppPdpt)
1708	*ppPdpt = pPdpt;
1709
1710	/* Allocate page directory if not present. */
1711	EPTPDPTE Pdpe;
1712	Pdpe.u = pPdpe->u;
1713	if (!(Pdpe.u & (EPT_E_PG_MASK \| EPT_E_READ)))
1714	{
1715	RTGCPTR64 const GCPdPt = GCPtr & ~(RT_BIT_64(EPT_PDPT_SHIFT) - 1);
1716	rc = pgmPoolAlloc(pVM, GCPdPt, PGMPOOLKIND_EPT_PD_FOR_PHYS, PGMPOOLACCESS_DONTCARE, PGM_A20_IS_ENABLED(pVCpu),
1717	pShwPage->idx, iPdPt, false /fLockPage/,
1718	&pShwPage);
1719	AssertRCReturn(rc, rc);
1720
1721	/* Hook up the new PD now. */
1722	ASMAtomicWriteU64(&pPdpe->u, pShwPage->Core.Key \| EPT_E_READ \| EPT_E_WRITE \| EPT_E_EXECUTE);
1723	}
1724	else
1725	{
1726	pShwPage = pgmPoolGetPage(pPool, pPdpe->u & EPT_PDPTE_PG_MASK);
1727	AssertReturn(pShwPage, VERR_PGM_POOL_GET_PAGE_FAILED);
1728
1729	pgmPoolCacheUsed(pPool, pShwPage);
1730
1731	/* Hook up the cached PD if needed (probably not given there are 512 PTs we may need sync). */
1732	if (Pdpe.u == (pShwPage->Core.Key \| EPT_E_READ \| EPT_E_WRITE \| EPT_E_EXECUTE))
1733	{ }
1734	else
1735	ASMAtomicWriteU64(&pPdpe->u, pShwPage->Core.Key \| EPT_E_READ \| EPT_E_WRITE \| EPT_E_EXECUTE);
1736	}
1737
1738	*ppPD = (PEPTPD)PGMPOOL_PAGE_2_PTR_V2(pVM, pVCpu, pShwPage);
1739	return VINF_SUCCESS;
1740	}
1741
1742
1743	#ifdef VBOX_WITH_NESTED_HWVIRT_VMX_EPT
1744	/**
1745	* Syncs the SHADOW nested-guest page directory pointer for the specified address.
1746	* Allocates backing pages in case the PDPT or PML4 entry is missing.
1747	*
1748	* @returns VBox status code.
1749	* @param pVCpu The cross context virtual CPU structure.
1750	* @param GCPhysNested The nested-guest physical address.
1751	* @param ppPdpt Where to store the PDPT. Optional, can be NULL.
1752	* @param ppPD Where to store the PD. Optional, can be NULL.
1753	* @param pGstWalkAll The guest walk info.
1754	*/
1755	static int pgmShwGetNestedEPTPDPtr(PVMCPUCC pVCpu, RTGCPTR64 GCPhysNested, PEPTPDPT ppPdpt, PEPTPD ppPD,
1756	PPGMPTWALKGST pGstWalkAll)
1757	{
1758	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
1759	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
1760	int rc;
1761
1762	PPGMPOOLPAGE pShwPage;
1763	Assert(pVM->pgm.s.fNestedPaging);
1764	Assert(pVCpu->pgm.s.enmGuestSlatMode == PGMSLAT_EPT);
1765	PGM_LOCK_ASSERT_OWNER(pVM);
1766
1767	/*
1768	* PML4 level.
1769	*/
1770	{
1771	PEPTPML4 pPml4 = (PEPTPML4)PGMPOOL_PAGE_2_PTR_V2(pVM, pVCpu, pVCpu->pgm.s.CTX_SUFF(pShwPageCR3));
1772	Assert(pPml4);
1773
1774	/* Allocate page directory pointer table if not present. */
1775	{
1776	uint64_t const fShwFlags = pGstWalkAll->u.Ept.Pml4e.u & pVCpu->pgm.s.fGstEptShadowedPml4eMask;
1777	const unsigned iPml4e = (GCPhysNested >> EPT_PML4_SHIFT) & EPT_PML4_MASK;
1778	PEPTPML4E pPml4e = &pPml4->a[iPml4e];
1779
1780	if (!(pPml4e->u & (EPT_E_PG_MASK \| EPT_PRESENT_MASK)))
1781	{
1782	RTGCPHYS const GCPhysPdpt = pGstWalkAll->u.Ept.Pml4e.u & EPT_PML4E_PG_MASK;
1783	rc = pgmPoolAlloc(pVM, GCPhysPdpt, PGMPOOLKIND_EPT_PDPT_FOR_EPT_PDPT, PGMPOOLACCESS_DONTCARE,
1784	PGM_A20_IS_ENABLED(pVCpu), pVCpu->pgm.s.CTX_SUFF(pShwPageCR3)->idx, iPml4e, false /fLockPage/,
1785	&pShwPage);
1786	AssertRCReturn(rc, rc);
1787
1788	/* Hook up the new PDPT now. */
1789	ASMAtomicWriteU64(&pPml4e->u, pShwPage->Core.Key \| fShwFlags);
1790	}
1791	else
1792	{
1793	pShwPage = pgmPoolGetPage(pPool, pPml4e->u & EPT_PML4E_PG_MASK);
1794	AssertReturn(pShwPage, VERR_PGM_POOL_GET_PAGE_FAILED);
1795
1796	pgmPoolCacheUsed(pPool, pShwPage);
1797
1798	/* Hook up the cached PDPT if needed (probably not given 512512 PTs to sync). /
1799	if (pPml4e->u != (pShwPage->Core.Key \| fShwFlags))
1800	ASMAtomicWriteU64(&pPml4e->u, pShwPage->Core.Key \| fShwFlags);
1801	}
1802	Assert(PGMPOOL_PAGE_IS_NESTED(pShwPage));
1803	Log7Func(("GstPml4e=%RX64 ShwPml4e=%RX64 iPml4e=%u\n", pGstWalkAll->u.Ept.Pml4e.u, pPml4e->u, iPml4e));
1804	}
1805	}
1806
1807	/*
1808	* PDPT level.
1809	*/
1810	{
1811	AssertReturn(!(pGstWalkAll->u.Ept.Pdpte.u & EPT_E_LEAF), VERR_NOT_SUPPORTED); /* shadowing 1GB pages not supported yet. */
1812
1813	PEPTPDPT pPdpt = (PEPTPDPT)PGMPOOL_PAGE_2_PTR_V2(pVM, pVCpu, pShwPage);
1814	if (ppPdpt)
1815	*ppPdpt = pPdpt;
1816
1817	uint64_t const fShwFlags = pGstWalkAll->u.Ept.Pdpte.u & pVCpu->pgm.s.fGstEptShadowedPdpteMask;
1818	const unsigned iPdPte = (GCPhysNested >> EPT_PDPT_SHIFT) & EPT_PDPT_MASK;
1819	PEPTPDPTE pPdpte = &pPdpt->a[iPdPte];
1820
1821	if (!(pPdpte->u & (EPT_E_PG_MASK \| EPT_PRESENT_MASK)))
1822	{
1823	RTGCPHYS const GCPhysPd = pGstWalkAll->u.Ept.Pdpte.u & EPT_PDPTE_PG_MASK;
1824	rc = pgmPoolAlloc(pVM, GCPhysPd, PGMPOOLKIND_EPT_PD_FOR_EPT_PD, PGMPOOLACCESS_DONTCARE, PGM_A20_IS_ENABLED(pVCpu),
1825	pShwPage->idx, iPdPte, false /fLockPage/, &pShwPage);
1826	AssertRCReturn(rc, rc);
1827
1828	/* Hook up the new PD now. */
1829	ASMAtomicWriteU64(&pPdpte->u, pShwPage->Core.Key \| fShwFlags);
1830	}
1831	else
1832	{
1833	pShwPage = pgmPoolGetPage(pPool, pPdpte->u & EPT_PDPTE_PG_MASK);
1834	AssertReturn(pShwPage, VERR_PGM_POOL_GET_PAGE_FAILED);
1835
1836	pgmPoolCacheUsed(pPool, pShwPage);
1837
1838	/* Hook up the cached PD if needed (probably not given there are 512 PTs we may need sync). */
1839	if (pPdpte->u != (pShwPage->Core.Key \| fShwFlags))
1840	ASMAtomicWriteU64(&pPdpte->u, pShwPage->Core.Key \| fShwFlags);
1841	}
1842	Assert(PGMPOOL_PAGE_IS_NESTED(pShwPage));
1843	Log7Func(("GstPdpte=%RX64 ShwPdpte=%RX64 iPdPte=%u \n", pGstWalkAll->u.Ept.Pdpte.u, pPdpte->u, iPdPte));
1844
1845	*ppPD = (PEPTPD)PGMPOOL_PAGE_2_PTR_V2(pVM, pVCpu, pShwPage);
1846	}
1847
1848	return VINF_SUCCESS;
1849	}
1850	#endif /* VBOX_WITH_NESTED_HWVIRT_VMX_EPT */
1851
1852
1853	#ifdef IN_RING0
1854	/**
1855	* Synchronizes a range of nested page table entries.
1856	*
1857	* The caller must own the PGM lock.
1858	*
1859	* @param pVCpu The cross context virtual CPU structure of the calling EMT.
1860	* @param GCPhys Where to start.
1861	* @param cPages How many pages which entries should be synced.
1862	* @param enmShwPagingMode The shadow paging mode (PGMMODE_EPT for VT-x,
1863	* host paging mode for AMD-V).
1864	*/
1865	int pgmShwSyncNestedPageLocked(PVMCPUCC pVCpu, RTGCPHYS GCPhys, uint32_t cPages, PGMMODE enmShwPagingMode)
1866	{
1867	PGM_LOCK_ASSERT_OWNER(pVCpu->CTX_SUFF(pVM));
1868
1869	/** @todo r=bird: Gotta love this nested paging hacking we're still carrying with us... (Split PGM_TYPE_NESTED.) */
1870	int rc;
1871	switch (enmShwPagingMode)
1872	{
1873	case PGMMODE_32_BIT:
1874	{
1875	X86PDE PdeDummy = { X86_PDE_P \| X86_PDE_US \| X86_PDE_RW \| X86_PDE_A };
1876	rc = PGM_BTH_NAME_32BIT_PROT(SyncPage)(pVCpu, PdeDummy, GCPhys, cPages, ~0U /uErr/);
1877	break;
1878	}
1879
1880	case PGMMODE_PAE:
1881	case PGMMODE_PAE_NX:
1882	{
1883	X86PDEPAE PdeDummy = { X86_PDE_P \| X86_PDE_US \| X86_PDE_RW \| X86_PDE_A };
1884	rc = PGM_BTH_NAME_PAE_PROT(SyncPage)(pVCpu, PdeDummy, GCPhys, cPages, ~0U /uErr/);
1885	break;
1886	}
1887
1888	case PGMMODE_AMD64:
1889	case PGMMODE_AMD64_NX:
1890	{
1891	X86PDEPAE PdeDummy = { X86_PDE_P \| X86_PDE_US \| X86_PDE_RW \| X86_PDE_A };
1892	rc = PGM_BTH_NAME_AMD64_PROT(SyncPage)(pVCpu, PdeDummy, GCPhys, cPages, ~0U /uErr/);
1893	break;
1894	}
1895
1896	case PGMMODE_EPT:
1897	{
1898	X86PDEPAE PdeDummy = { X86_PDE_P \| X86_PDE_US \| X86_PDE_RW \| X86_PDE_A };
1899	rc = PGM_BTH_NAME_EPT_PROT(SyncPage)(pVCpu, PdeDummy, GCPhys, cPages, ~0U /uErr/);
1900	break;
1901	}
1902
1903	default:
1904	AssertMsgFailedReturn(("%d\n", enmShwPagingMode), VERR_IPE_NOT_REACHED_DEFAULT_CASE);
1905	}
1906	return rc;
1907	}
1908	#endif /* IN_RING0 */
1909
1910
1911	/**
1912	* Gets effective Guest OS page information.
1913	*
1914	* @returns VBox status code.
1915	* @param pVCpu The cross context virtual CPU structure of the calling EMT.
1916	* @param GCPtr Guest Context virtual address of the page.
1917	* @param pWalk Where to store the page walk information.
1918	* @thread EMT(pVCpu)
1919	*/
1920	VMMDECL(int) PGMGstGetPage(PVMCPUCC pVCpu, RTGCPTR GCPtr, PPGMPTWALK pWalk)
1921	{
1922	VMCPU_ASSERT_EMT(pVCpu);
1923	Assert(pWalk);
1924	uintptr_t idx = pVCpu->pgm.s.idxGuestModeData;
1925	AssertReturn(idx < RT_ELEMENTS(g_aPgmGuestModeData), VERR_PGM_MODE_IPE);
1926	AssertReturn(g_aPgmGuestModeData[idx].pfnGetPage, VERR_PGM_MODE_IPE);
1927	return g_aPgmGuestModeData[idx].pfnGetPage(pVCpu, GCPtr, pWalk);
1928	}
1929
1930
1931	/**
1932	* Gets effective Guest OS page information.
1933	*
1934	* @returns VBox status code.
1935	* @param pVCpu The cross context virtual CPU structure of the calling EMT.
1936	* @param GCPtr Guest Context virtual address of the page.
1937	* @param fFlags PGMQPAGE_F_XXX. If zero, no accessed or dirty bits will
1938	* be set.
1939	* @param pWalk Where to store the page walk information.
1940	* @thread EMT(pVCpu)
1941	*/
1942	VMM_INT_DECL(int) PGMGstQueryPageFast(PVMCPUCC pVCpu, RTGCPTR GCPtr, uint32_t fFlags, PPGMPTWALKFAST pWalk)
1943	{
1944	VMCPU_ASSERT_EMT(pVCpu);
1945	Assert(pWalk);
1946	Assert(!(fFlags & ~(PGMQPAGE_F_VALID_MASK)));
1947	Assert(!(fFlags & PGMQPAGE_F_EXECUTE) \|\| !(fFlags & PGMQPAGE_F_WRITE));
1948	uintptr_t idx = pVCpu->pgm.s.idxGuestModeData;
1949	AssertReturn(idx < RT_ELEMENTS(g_aPgmGuestModeData), VERR_PGM_MODE_IPE);
1950	AssertReturn(g_aPgmGuestModeData[idx].pfnGetPage, VERR_PGM_MODE_IPE);
1951	return g_aPgmGuestModeData[idx].pfnQueryPageFast(pVCpu, GCPtr, fFlags, pWalk);
1952	}
1953
1954
1955	/**
1956	* Maps the guest CR3.
1957	*
1958	* @returns VBox status code.
1959	* @param pVCpu The cross context virtual CPU structure.
1960	* @param GCPhysCr3 The guest CR3 value.
1961	* @param pHCPtrGuestCr3 Where to store the mapped memory.
1962	*/
1963	DECLINLINE(int) pgmGstMapCr3(PVMCPUCC pVCpu, RTGCPHYS GCPhysCr3, PRTHCPTR pHCPtrGuestCr3)
1964	{
1965	/** @todo this needs some reworking wrt. locking? */
1966	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
1967	PGM_LOCK_VOID(pVM);
1968	PPGMPAGE pPageCr3 = pgmPhysGetPage(pVM, GCPhysCr3);
1969	AssertReturnStmt(pPageCr3, PGM_UNLOCK(pVM), VERR_PGM_INVALID_CR3_ADDR);
1970
1971	RTHCPTR HCPtrGuestCr3;
1972	int rc = pgmPhysGCPhys2CCPtrInternalDepr(pVM, pPageCr3, GCPhysCr3, (void **)&HCPtrGuestCr3);
1973	PGM_UNLOCK(pVM);
1974
1975	*pHCPtrGuestCr3 = HCPtrGuestCr3;
1976	return rc;
1977	}
1978
1979
1980	#ifdef VBOX_WITH_NESTED_HWVIRT_VMX_EPT
1981	/**
1982	* Unmaps the guest CR3.
1983	*
1984	* @returns VBox status code.
1985	* @param pVCpu The cross context virtual CPU structure.
1986	*/
1987	DECLINLINE(int) pgmGstUnmapCr3(PVMCPUCC pVCpu)
1988	{
1989	uintptr_t const idxBth = pVCpu->pgm.s.idxBothModeData;
1990	AssertReturn(idxBth < RT_ELEMENTS(g_aPgmBothModeData), VERR_PGM_MODE_IPE);
1991	AssertReturn(g_aPgmBothModeData[idxBth].pfnUnmapCR3, VERR_PGM_MODE_IPE);
1992	return g_aPgmBothModeData[idxBth].pfnUnmapCR3(pVCpu);
1993	}
1994	#endif
1995
1996
1997	/**
1998	* Performs a guest page table walk.
1999	*
2000	* The guest should be in paged protect mode or long mode when making a call to
2001	* this function.
2002	*
2003	* @returns VBox status code.
2004	* @retval VINF_SUCCESS on success.
2005	* @retval VERR_PAGE_TABLE_NOT_PRESENT on failure. Check pWalk for details.
2006	* @retval VERR_PGM_NOT_USED_IN_MODE if not paging isn't enabled. @a pWalk is
2007	* not valid, except enmType is PGMPTWALKGSTTYPE_INVALID.
2008	*
2009	* @param pVCpu The cross context virtual CPU structure of the calling EMT.
2010	* @param GCPtr The guest virtual address to walk by.
2011	* @param pWalk Where to return the walk result. This is valid for some
2012	* error codes as well.
2013	* @param pGstWalk The guest mode specific page walk information.
2014	*/
2015	int pgmGstPtWalk(PVMCPUCC pVCpu, RTGCPTR GCPtr, PPGMPTWALK pWalk, PPGMPTWALKGST pGstWalk)
2016	{
2017	VMCPU_ASSERT_EMT(pVCpu);
2018	switch (pVCpu->pgm.s.enmGuestMode)
2019	{
2020	case PGMMODE_32_BIT:
2021	pGstWalk->enmType = PGMPTWALKGSTTYPE_32BIT;
2022	return PGM_GST_NAME_32BIT(Walk)(pVCpu, GCPtr, pWalk, &pGstWalk->u.Legacy);
2023
2024	case PGMMODE_PAE:
2025	case PGMMODE_PAE_NX:
2026	pGstWalk->enmType = PGMPTWALKGSTTYPE_PAE;
2027	return PGM_GST_NAME_PAE(Walk)(pVCpu, GCPtr, pWalk, &pGstWalk->u.Pae);
2028
2029	case PGMMODE_AMD64:
2030	case PGMMODE_AMD64_NX:
2031	pGstWalk->enmType = PGMPTWALKGSTTYPE_AMD64;
2032	return PGM_GST_NAME_AMD64(Walk)(pVCpu, GCPtr, pWalk, &pGstWalk->u.Amd64);
2033
2034	case PGMMODE_REAL:
2035	case PGMMODE_PROTECTED:
2036	pGstWalk->enmType = PGMPTWALKGSTTYPE_INVALID;
2037	return VERR_PGM_NOT_USED_IN_MODE;
2038
2039	case PGMMODE_EPT:
2040	case PGMMODE_NESTED_32BIT:
2041	case PGMMODE_NESTED_PAE:
2042	case PGMMODE_NESTED_AMD64:
2043	default:
2044	AssertFailed();
2045	pGstWalk->enmType = PGMPTWALKGSTTYPE_INVALID;
2046	return VERR_PGM_NOT_USED_IN_MODE;
2047	}
2048	}
2049
2050
2051	#ifdef VBOX_WITH_NESTED_HWVIRT_VMX_EPT
2052	/**
2053	* Performs a guest second-level address translation (SLAT).
2054	*
2055	* @returns VBox status code.
2056	* @retval VINF_SUCCESS on success.
2057	* @retval VERR_PAGE_TABLE_NOT_PRESENT on failure. Check pWalk for details.
2058	* @retval VERR_PGM_NOT_USED_IN_MODE if not paging isn't enabled. @a pWalk is
2059	* not valid, except enmType is PGMPTWALKGSTTYPE_INVALID.
2060	*
2061	* @param pVCpu The cross context virtual CPU structure of the calling EMT.
2062	* @param GCPhysNested The nested-guest physical address being translated.
2063	* @param fIsLinearAddrValid Whether the linear address in @a GCPtrNested is the
2064	* cause for this translation.
2065	* @param GCPtrNested The nested-guest virtual address that initiated the
2066	* SLAT. If none, pass 0 (and not NIL_RTGCPTR).
2067	* @param pWalk Where to return the walk result. This is updated for
2068	* all error codes other than
2069	* VERR_PGM_NOT_USED_IN_MODE.
2070	* @param pGstWalk Where to store the second-level paging-mode specific
2071	* walk info.
2072	*/
2073	static int pgmGstSlatWalk(PVMCPUCC pVCpu, RTGCPHYS GCPhysNested, bool fIsLinearAddrValid, RTGCPTR GCPtrNested,
2074	PPGMPTWALK pWalk, PPGMPTWALKGST pGstWalk)
2075	{
2076	/* SLAT mode must be valid at this point as this should only be used -after- we have determined SLAT mode. */
2077	Assert( pVCpu->pgm.s.enmGuestSlatMode != PGMSLAT_DIRECT
2078	&& pVCpu->pgm.s.enmGuestSlatMode != PGMSLAT_INVALID);
2079	AssertPtr(pWalk);
2080	AssertPtr(pGstWalk);
2081	switch (pVCpu->pgm.s.enmGuestSlatMode)
2082	{
2083	case PGMSLAT_EPT:
2084	pGstWalk->enmType = PGMPTWALKGSTTYPE_EPT;
2085	return PGM_GST_SLAT_NAME_EPT(Walk)(pVCpu, GCPhysNested, fIsLinearAddrValid, GCPtrNested, pWalk, &pGstWalk->u.Ept);
2086
2087	default:
2088	AssertFailed();
2089	pGstWalk->enmType = PGMPTWALKGSTTYPE_INVALID;
2090	return VERR_PGM_NOT_USED_IN_MODE;
2091	}
2092	}
2093	#endif /* VBOX_WITH_NESTED_HWVIRT_VMX_EPT */
2094
2095
2096	/**
2097	* Tries to continue the previous walk.
2098	*
2099	* @note Requires the caller to hold the PGM lock from the first
2100	* pgmGstPtWalk() call to the last pgmGstPtWalkNext() call. Otherwise
2101	* we cannot use the pointers.
2102	*
2103	* @returns VBox status code.
2104	* @retval VINF_SUCCESS on success.
2105	* @retval VERR_PAGE_TABLE_NOT_PRESENT on failure. Check pWalk for details.
2106	* @retval VERR_PGM_NOT_USED_IN_MODE if not paging isn't enabled. @a pWalk is
2107	* not valid, except enmType is PGMPTWALKGSTTYPE_INVALID.
2108	*
2109	* @param pVCpu The cross context virtual CPU structure of the calling EMT.
2110	* @param GCPtr The guest virtual address to walk by.
2111	* @param pWalk Pointer to the previous walk result and where to return
2112	* the result of this walk. This is valid for some error
2113	* codes as well.
2114	* @param pGstWalk The guest-mode specific walk information.
2115	*/
2116	int pgmGstPtWalkNext(PVMCPUCC pVCpu, RTGCPTR GCPtr, PPGMPTWALK pWalk, PPGMPTWALKGST pGstWalk)
2117	{
2118	/*
2119	* We can only handle successfully walks.
2120	* We also limit ourselves to the next page.
2121	*/
2122	if ( pWalk->fSucceeded
2123	&& GCPtr - pWalk->GCPtr == GUEST_PAGE_SIZE)
2124	{
2125	Assert(pWalk->uLevel == 0);
2126	if (pGstWalk->enmType == PGMPTWALKGSTTYPE_AMD64)
2127	{
2128	/*
2129	* AMD64
2130	*/
2131	if (!pWalk->fGigantPage && !pWalk->fBigPage)
2132	{
2133	/*
2134	* We fall back to full walk if the PDE table changes, if any
2135	* reserved bits are set, or if the effective page access changes.
2136	*/
2137	const uint64_t fPteSame = X86_PTE_P \| X86_PTE_RW \| X86_PTE_US \| X86_PTE_PWT
2138	\| X86_PTE_PCD \| X86_PTE_A \| X86_PTE_PAE_NX;
2139	const uint64_t fPdeSame = X86_PDE_P \| X86_PDE_RW \| X86_PDE_US \| X86_PDE_PWT
2140	\| X86_PDE_PCD \| X86_PDE_A \| X86_PDE_PAE_NX \| X86_PDE_PS;
2141
2142	if ((GCPtr >> X86_PD_PAE_SHIFT) == (pWalk->GCPtr >> X86_PD_PAE_SHIFT))
2143	{
2144	if (pGstWalk->u.Amd64.pPte)
2145	{
2146	X86PTEPAE Pte;
2147	Pte.u = pGstWalk->u.Amd64.pPte[1].u;
2148	if ( (Pte.u & fPteSame) == (pGstWalk->u.Amd64.Pte.u & fPteSame)
2149	&& !(Pte.u & (pVCpu)->pgm.s.fGstAmd64MbzPteMask))
2150	{
2151	pWalk->GCPtr = GCPtr;
2152	pWalk->GCPhys = Pte.u & X86_PTE_PAE_PG_MASK;
2153	pGstWalk->u.Amd64.Pte.u = Pte.u;
2154	pGstWalk->u.Amd64.pPte++;
2155	return VINF_SUCCESS;
2156	}
2157	}
2158	}
2159	else if ((GCPtr >> X86_PDPT_SHIFT) == (pWalk->GCPtr >> X86_PDPT_SHIFT))
2160	{
2161	Assert(!((GCPtr >> X86_PT_PAE_SHIFT) & X86_PT_PAE_MASK)); /* Must be first PT entry. */
2162	if (pGstWalk->u.Amd64.pPde)
2163	{
2164	X86PDEPAE Pde;
2165	Pde.u = pGstWalk->u.Amd64.pPde[1].u;
2166	if ( (Pde.u & fPdeSame) == (pGstWalk->u.Amd64.Pde.u & fPdeSame)
2167	&& !(Pde.u & (pVCpu)->pgm.s.fGstAmd64MbzPdeMask))
2168	{
2169	/* Get the new PTE and check out the first entry. */
2170	int rc = PGM_GCPHYS_2_PTR_BY_VMCPU(pVCpu, PGM_A20_APPLY(pVCpu, (Pde.u & X86_PDE_PAE_PG_MASK)),
2171	&pGstWalk->u.Amd64.pPt);
2172	if (RT_SUCCESS(rc))
2173	{
2174	pGstWalk->u.Amd64.pPte = &pGstWalk->u.Amd64.pPt->a[0];
2175	X86PTEPAE Pte;
2176	Pte.u = pGstWalk->u.Amd64.pPte->u;
2177	if ( (Pte.u & fPteSame) == (pGstWalk->u.Amd64.Pte.u & fPteSame)
2178	&& !(Pte.u & (pVCpu)->pgm.s.fGstAmd64MbzPteMask))
2179	{
2180	pWalk->GCPtr = GCPtr;
2181	pWalk->GCPhys = Pte.u & X86_PTE_PAE_PG_MASK;
2182	pGstWalk->u.Amd64.Pte.u = Pte.u;
2183	pGstWalk->u.Amd64.Pde.u = Pde.u;
2184	pGstWalk->u.Amd64.pPde++;
2185	return VINF_SUCCESS;
2186	}
2187	}
2188	}
2189	}
2190	}
2191	}
2192	else if (!pWalk->fGigantPage)
2193	{
2194	if ((GCPtr & X86_PAGE_2M_BASE_MASK) == (pWalk->GCPtr & X86_PAGE_2M_BASE_MASK))
2195	{
2196	pWalk->GCPtr = GCPtr;
2197	pWalk->GCPhys += GUEST_PAGE_SIZE;
2198	return VINF_SUCCESS;
2199	}
2200	}
2201	else
2202	{
2203	if ((GCPtr & X86_PAGE_1G_BASE_MASK) == (pWalk->GCPtr & X86_PAGE_1G_BASE_MASK))
2204	{
2205	pWalk->GCPtr = GCPtr;
2206	pWalk->GCPhys += GUEST_PAGE_SIZE;
2207	return VINF_SUCCESS;
2208	}
2209	}
2210	}
2211	}
2212	/* Case we don't handle. Do full walk. */
2213	return pgmGstPtWalk(pVCpu, GCPtr, pWalk, pGstWalk);
2214	}
2215
2216
2217	/**
2218	* Modify page flags for a range of pages in the guest's tables
2219	*
2220	* The existing flags are ANDed with the fMask and ORed with the fFlags.
2221	*
2222	* @returns VBox status code.
2223	* @param pVCpu The cross context virtual CPU structure.
2224	* @param GCPtr Virtual address of the first page in the range.
2225	* @param cb Size (in bytes) of the range to apply the modification to.
2226	* @param fFlags The OR mask - page flags X86_PTE_*, excluding the page mask of course.
2227	* @param fMask The AND mask - page flags X86_PTE_*, excluding the page mask of course.
2228	* Be very CAREFUL when ~'ing constants which could be 32-bit!
2229	*/
2230	VMMDECL(int) PGMGstModifyPage(PVMCPUCC pVCpu, RTGCPTR GCPtr, size_t cb, uint64_t fFlags, uint64_t fMask)
2231	{
2232	STAM_PROFILE_START(&pVCpu->pgm.s.Stats.CTX_MID_Z(Stat,GstModifyPage), a);
2233	VMCPU_ASSERT_EMT(pVCpu);
2234
2235	/*
2236	* Validate input.
2237	*/
2238	AssertMsg(!(fFlags & X86_PTE_PAE_PG_MASK), ("fFlags=%#llx\n", fFlags));
2239	Assert(cb);
2240
2241	LogFlow(("PGMGstModifyPage %RGv %d bytes fFlags=%08llx fMask=%08llx\n", GCPtr, cb, fFlags, fMask));
2242
2243	/*
2244	* Adjust input.
2245	*/
2246	cb += GCPtr & GUEST_PAGE_OFFSET_MASK;
2247	cb = RT_ALIGN_Z(cb, GUEST_PAGE_SIZE);
2248	GCPtr &= ~(RTGCPTR)GUEST_PAGE_OFFSET_MASK;
2249
2250	/*
2251	* Call worker.
2252	*/
2253	uintptr_t idx = pVCpu->pgm.s.idxGuestModeData;
2254	AssertReturn(idx < RT_ELEMENTS(g_aPgmGuestModeData), VERR_PGM_MODE_IPE);
2255	AssertReturn(g_aPgmGuestModeData[idx].pfnModifyPage, VERR_PGM_MODE_IPE);
2256	int rc = g_aPgmGuestModeData[idx].pfnModifyPage(pVCpu, GCPtr, cb, fFlags, fMask);
2257
2258	STAM_PROFILE_STOP(&pVCpu->pgm.s.Stats.CTX_MID_Z(Stat,GstModifyPage), a);
2259	return rc;
2260	}
2261
2262
2263	/**
2264	* Checks whether the given PAE PDPEs are potentially valid for the guest.
2265	*
2266	* @returns @c true if the PDPE is valid, @c false otherwise.
2267	* @param pVCpu The cross context virtual CPU structure of the calling EMT.
2268	* @param paPaePdpes The PAE PDPEs to validate.
2269	*
2270	* @remarks This function -only- checks the reserved bits in the PDPE entries.
2271	*/
2272	VMM_INT_DECL(bool) PGMGstArePaePdpesValid(PVMCPUCC pVCpu, PCX86PDPE paPaePdpes)
2273	{
2274	Assert(paPaePdpes);
2275	for (unsigned i = 0; i < X86_PG_PAE_PDPE_ENTRIES; i++)
2276	{
2277	X86PDPE const PaePdpe = paPaePdpes[i];
2278	if ( !(PaePdpe.u & X86_PDPE_P)
2279	\|\| !(PaePdpe.u & pVCpu->pgm.s.fGstPaeMbzPdpeMask))
2280	{ /* likely */ }
2281	else
2282	return false;
2283	}
2284	return true;
2285	}
2286
2287
2288	/**
2289	* Performs the lazy mapping of the 32-bit guest PD.
2290	*
2291	* @returns VBox status code.
2292	* @param pVCpu The cross context virtual CPU structure of the calling EMT.
2293	* @param ppPd Where to return the pointer to the mapping. This is
2294	* always set.
2295	*/
2296	int pgmGstLazyMap32BitPD(PVMCPUCC pVCpu, PX86PD *ppPd)
2297	{
2298	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
2299	PGM_LOCK_VOID(pVM);
2300
2301	Assert(!pVCpu->pgm.s.CTX_SUFF(pGst32BitPd));
2302
2303	RTGCPHYS GCPhysCR3 = pgmGetGuestMaskedCr3(pVCpu, pVCpu->pgm.s.GCPhysCR3);
2304	PPGMPAGE pPage;
2305	int rc = pgmPhysGetPageEx(pVM, GCPhysCR3, &pPage);
2306	if (RT_SUCCESS(rc))
2307	{
2308	rc = pgmPhysGCPhys2CCPtrInternalDepr(pVM, pPage, GCPhysCR3, (void **)ppPd);
2309	if (RT_SUCCESS(rc))
2310	{
2311	# ifdef IN_RING3
2312	pVCpu->pgm.s.pGst32BitPdR0 = NIL_RTR0PTR;
2313	pVCpu->pgm.s.pGst32BitPdR3 = *ppPd;
2314	# else
2315	pVCpu->pgm.s.pGst32BitPdR3 = NIL_RTR0PTR;
2316	pVCpu->pgm.s.pGst32BitPdR0 = *ppPd;
2317	# endif
2318	PGM_UNLOCK(pVM);
2319	return VINF_SUCCESS;
2320	}
2321	AssertRC(rc);
2322	}
2323	PGM_UNLOCK(pVM);
2324
2325	*ppPd = NULL;
2326	return rc;
2327	}
2328
2329
2330	/**
2331	* Performs the lazy mapping of the PAE guest PDPT.
2332	*
2333	* @returns VBox status code.
2334	* @param pVCpu The cross context virtual CPU structure of the calling EMT.
2335	* @param ppPdpt Where to return the pointer to the mapping. This is
2336	* always set.
2337	*/
2338	int pgmGstLazyMapPaePDPT(PVMCPUCC pVCpu, PX86PDPT *ppPdpt)
2339	{
2340	Assert(!pVCpu->pgm.s.CTX_SUFF(pGstPaePdpt));
2341	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
2342	PGM_LOCK_VOID(pVM);
2343
2344	RTGCPHYS GCPhysCR3 = pgmGetGuestMaskedCr3(pVCpu, pVCpu->pgm.s.GCPhysCR3);
2345	PPGMPAGE pPage;
2346	int rc = pgmPhysGetPageEx(pVM, GCPhysCR3, &pPage);
2347	if (RT_SUCCESS(rc))
2348	{
2349	rc = pgmPhysGCPhys2CCPtrInternalDepr(pVM, pPage, GCPhysCR3, (void **)ppPdpt);
2350	if (RT_SUCCESS(rc))
2351	{
2352	# ifdef IN_RING3
2353	pVCpu->pgm.s.pGstPaePdptR0 = NIL_RTR0PTR;
2354	pVCpu->pgm.s.pGstPaePdptR3 = *ppPdpt;
2355	# else
2356	pVCpu->pgm.s.pGstPaePdptR3 = NIL_RTR3PTR;
2357	pVCpu->pgm.s.pGstPaePdptR0 = *ppPdpt;
2358	# endif
2359	PGM_UNLOCK(pVM);
2360	return VINF_SUCCESS;
2361	}
2362	AssertRC(rc);
2363	}
2364
2365	PGM_UNLOCK(pVM);
2366	*ppPdpt = NULL;
2367	return rc;
2368	}
2369
2370
2371	/**
2372	* Performs the lazy mapping / updating of a PAE guest PD.
2373	*
2374	* @returns Pointer to the mapping.
2375	* @returns VBox status code.
2376	* @param pVCpu The cross context virtual CPU structure of the calling EMT.
2377	* @param iPdpt Which PD entry to map (0..3).
2378	* @param ppPd Where to return the pointer to the mapping. This is
2379	* always set.
2380	*/
2381	int pgmGstLazyMapPaePD(PVMCPUCC pVCpu, uint32_t iPdpt, PX86PDPAE *ppPd)
2382	{
2383	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
2384	PGM_LOCK_VOID(pVM);
2385
2386	PX86PDPT pGuestPDPT = pVCpu->pgm.s.CTX_SUFF(pGstPaePdpt);
2387	Assert(pGuestPDPT);
2388	Assert(pGuestPDPT->a[iPdpt].u & X86_PDPE_P);
2389	RTGCPHYS GCPhys = pGuestPDPT->a[iPdpt].u & X86_PDPE_PG_MASK;
2390	bool const fChanged = pVCpu->pgm.s.aGCPhysGstPaePDs[iPdpt] != GCPhys;
2391
2392	PPGMPAGE pPage;
2393	int rc = pgmPhysGetPageEx(pVM, GCPhys, &pPage);
2394	if (RT_SUCCESS(rc))
2395	{
2396	rc = pgmPhysGCPhys2CCPtrInternalDepr(pVM, pPage, GCPhys, (void **)ppPd);
2397	AssertRC(rc);
2398	if (RT_SUCCESS(rc))
2399	{
2400	# ifdef IN_RING3
2401	pVCpu->pgm.s.apGstPaePDsR0[iPdpt] = NIL_RTR0PTR;
2402	pVCpu->pgm.s.apGstPaePDsR3[iPdpt] = *ppPd;
2403	# else
2404	pVCpu->pgm.s.apGstPaePDsR3[iPdpt] = NIL_RTR3PTR;
2405	pVCpu->pgm.s.apGstPaePDsR0[iPdpt] = *ppPd;
2406	# endif
2407	if (fChanged)
2408	pVCpu->pgm.s.aGCPhysGstPaePDs[iPdpt] = GCPhys;
2409	PGM_UNLOCK(pVM);
2410	return VINF_SUCCESS;
2411	}
2412	}
2413
2414	/* Invalid page or some failure, invalidate the entry. */
2415	pVCpu->pgm.s.aGCPhysGstPaePDs[iPdpt] = NIL_RTGCPHYS;
2416	pVCpu->pgm.s.apGstPaePDsR3[iPdpt] = NIL_RTR3PTR;
2417	pVCpu->pgm.s.apGstPaePDsR0[iPdpt] = NIL_RTR0PTR;
2418
2419	PGM_UNLOCK(pVM);
2420	return rc;
2421	}
2422
2423
2424	/**
2425	* Performs the lazy mapping of the 32-bit guest PD.
2426	*
2427	* @returns VBox status code.
2428	* @param pVCpu The cross context virtual CPU structure of the calling EMT.
2429	* @param ppPml4 Where to return the pointer to the mapping. This will
2430	* always be set.
2431	*/
2432	int pgmGstLazyMapPml4(PVMCPUCC pVCpu, PX86PML4 *ppPml4)
2433	{
2434	Assert(!pVCpu->pgm.s.CTX_SUFF(pGstAmd64Pml4));
2435	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
2436	PGM_LOCK_VOID(pVM);
2437
2438	RTGCPHYS GCPhysCR3 = pgmGetGuestMaskedCr3(pVCpu, pVCpu->pgm.s.GCPhysCR3);
2439	PPGMPAGE pPage;
2440	int rc = pgmPhysGetPageEx(pVM, GCPhysCR3, &pPage);
2441	if (RT_SUCCESS(rc))
2442	{
2443	rc = pgmPhysGCPhys2CCPtrInternalDepr(pVM, pPage, GCPhysCR3, (void **)ppPml4);
2444	if (RT_SUCCESS(rc))
2445	{
2446	# ifdef IN_RING3
2447	pVCpu->pgm.s.pGstAmd64Pml4R0 = NIL_RTR0PTR;
2448	pVCpu->pgm.s.pGstAmd64Pml4R3 = *ppPml4;
2449	# else
2450	pVCpu->pgm.s.pGstAmd64Pml4R3 = NIL_RTR3PTR;
2451	pVCpu->pgm.s.pGstAmd64Pml4R0 = *ppPml4;
2452	# endif
2453	PGM_UNLOCK(pVM);
2454	return VINF_SUCCESS;
2455	}
2456	}
2457
2458	PGM_UNLOCK(pVM);
2459	*ppPml4 = NULL;
2460	return rc;
2461	}
2462
2463
2464	#ifdef VBOX_WITH_NESTED_HWVIRT_VMX_EPT
2465	/**
2466	* Performs the lazy mapping of the guest PML4 table when using EPT paging.
2467	*
2468	* @returns VBox status code.
2469	* @param pVCpu The cross context virtual CPU structure of the calling EMT.
2470	* @param ppEptPml4 Where to return the pointer to the mapping. This will
2471	* always be set.
2472	*/
2473	int pgmGstLazyMapEptPml4(PVMCPUCC pVCpu, PEPTPML4 *ppEptPml4)
2474	{
2475	Assert(!pVCpu->pgm.s.CTX_SUFF(pGstEptPml4));
2476	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
2477	PGM_LOCK_VOID(pVM);
2478
2479	RTGCPHYS const GCPhysEpt = pVCpu->pgm.s.uEptPtr & EPT_EPTP_PG_MASK;
2480	PPGMPAGE pPage;
2481	int rc = pgmPhysGetPageEx(pVM, GCPhysEpt, &pPage);
2482	if (RT_SUCCESS(rc))
2483	{
2484	rc = pgmPhysGCPhys2CCPtrInternalDepr(pVM, pPage, GCPhysEpt, (void **)ppEptPml4);
2485	if (RT_SUCCESS(rc))
2486	{
2487	# ifdef IN_RING3
2488	pVCpu->pgm.s.pGstEptPml4R0 = NIL_RTR0PTR;
2489	pVCpu->pgm.s.pGstEptPml4R3 = *ppEptPml4;
2490	# else
2491	pVCpu->pgm.s.pGstEptPml4R3 = NIL_RTR3PTR;
2492	pVCpu->pgm.s.pGstEptPml4R0 = *ppEptPml4;
2493	# endif
2494	PGM_UNLOCK(pVM);
2495	return VINF_SUCCESS;
2496	}
2497	}
2498
2499	PGM_UNLOCK(pVM);
2500	*ppEptPml4 = NULL;
2501	return rc;
2502	}
2503	#endif
2504
2505
2506	/**
2507	* Gets the current CR3 register value for the shadow memory context.
2508	* @returns CR3 value.
2509	* @param pVCpu The cross context virtual CPU structure.
2510	*/
2511	VMMDECL(RTHCPHYS) PGMGetHyperCR3(PVMCPU pVCpu)
2512	{
2513	PPGMPOOLPAGE pPoolPage = pVCpu->pgm.s.CTX_SUFF(pShwPageCR3);
2514	AssertPtrReturn(pPoolPage, NIL_RTHCPHYS);
2515	return pPoolPage->Core.Key;
2516	}
2517
2518
2519	/**
2520	* Forces lazy remapping of the guest's PAE page-directory structures.
2521	*
2522	* @param pVCpu The cross context virtual CPU structure.
2523	*/
2524	static void pgmGstFlushPaePdpes(PVMCPU pVCpu)
2525	{
2526	for (unsigned i = 0; i < RT_ELEMENTS(pVCpu->pgm.s.aGCPhysGstPaePDs); i++)
2527	{
2528	pVCpu->pgm.s.apGstPaePDsR3[i] = 0;
2529	pVCpu->pgm.s.apGstPaePDsR0[i] = 0;
2530	pVCpu->pgm.s.aGCPhysGstPaePDs[i] = NIL_RTGCPHYS;
2531	}
2532	}
2533
2534
2535	#ifdef VBOX_WITH_NESTED_HWVIRT_VMX_EPT
2536	/**
2537	* Performs second-level address translation for the given CR3 and updates the
2538	* nested-guest CR3 when successful.
2539	*
2540	* @returns VBox status code.
2541	* @param pVCpu The cross context virtual CPU structure.
2542	* @param uCr3 The masked nested-guest CR3 value.
2543	* @param pGCPhysCR3 Where to store the translated CR3.
2544	*
2545	* @warning This updates PGMCPU::GCPhysNstGstCR3 when the translation succeeds. Be
2546	* mindful of this in code that's hyper sensitive to the order of
2547	* operations.
2548	*/
2549	static int pgmGstSlatTranslateCr3(PVMCPUCC pVCpu, uint64_t uCr3, PRTGCPHYS pGCPhysCr3)
2550	{
2551	if (uCr3 != pVCpu->pgm.s.GCPhysNstGstCR3)
2552	{
2553	PGMPTWALK Walk;
2554	PGMPTWALKGST GstWalk;
2555	int const rc = pgmGstSlatWalk(pVCpu, uCr3, false /* fIsLinearAddrValid /, 0 / GCPtrNested */, &Walk, &GstWalk);
2556	if (RT_SUCCESS(rc))
2557	{
2558	/* Update nested-guest CR3. */
2559	pVCpu->pgm.s.GCPhysNstGstCR3 = uCr3;
2560
2561	/* Pass back the translated result. */
2562	*pGCPhysCr3 = Walk.GCPhys;
2563	return VINF_SUCCESS;
2564	}
2565
2566	/* Translation failed. */
2567	*pGCPhysCr3 = NIL_RTGCPHYS;
2568	return rc;
2569	}
2570
2571	/*
2572	* If the nested-guest CR3 has not changed, then the previously
2573	* translated CR3 result (i.e. GCPhysCR3) is passed back.
2574	*/
2575	*pGCPhysCr3 = pVCpu->pgm.s.GCPhysCR3;
2576	return VINF_SUCCESS;
2577	}
2578	#endif
2579
2580
2581	/**
2582	* Performs and schedules necessary updates following a CR3 load or reload.
2583	*
2584	* This will normally involve mapping the guest PD or nPDPT
2585	*
2586	* @returns VBox status code.
2587	* @retval VINF_PGM_SYNC_CR3 if monitoring requires a CR3 sync. This can
2588	* safely be ignored and overridden since the FF will be set too then.
2589	* @param pVCpu The cross context virtual CPU structure.
2590	* @param cr3 The new cr3.
2591	* @param fGlobal Indicates whether this is a global flush or not.
2592	*/
2593	VMMDECL(int) PGMFlushTLB(PVMCPUCC pVCpu, uint64_t cr3, bool fGlobal)
2594	{
2595	STAM_PROFILE_START(&pVCpu->pgm.s.Stats.CTX_MID_Z(Stat,FlushTLB), a);
2596	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
2597
2598	VMCPU_ASSERT_EMT(pVCpu);
2599
2600	/*
2601	* Always flag the necessary updates; necessary for hardware acceleration
2602	*/
2603	/** @todo optimize this, it shouldn't always be necessary. */
2604	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3_NON_GLOBAL);
2605	if (fGlobal)
2606	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
2607
2608	/*
2609	* Remap the CR3 content and adjust the monitoring if CR3 was actually changed.
2610	*/
2611	RTGCPHYS const GCPhysOldCR3 = pVCpu->pgm.s.GCPhysCR3;
2612	RTGCPHYS GCPhysCR3 = pgmGetGuestMaskedCr3(pVCpu, cr3);
2613	#ifdef VBOX_WITH_NESTED_HWVIRT_VMX_EPT
2614	if ( pVCpu->pgm.s.enmGuestSlatMode == PGMSLAT_EPT
2615	&& PGMMODE_WITH_PAGING(pVCpu->pgm.s.enmGuestMode))
2616	{
2617	RTGCPHYS GCPhysOut;
2618	int const rc = pgmGstSlatTranslateCr3(pVCpu, GCPhysCR3, &GCPhysOut);
2619	if (RT_SUCCESS(rc))
2620	GCPhysCR3 = GCPhysOut;
2621	else
2622	{
2623	/* CR3 SLAT translation failed but we try to pretend it
2624	succeeded for the reasons mentioned in PGMHCChangeMode(). */
2625	AssertMsgFailed(("SLAT failed for CR3 %#RX64 rc=%Rrc\n", cr3, rc));
2626	int const rc2 = pgmGstUnmapCr3(pVCpu);
2627	pVCpu->pgm.s.GCPhysCR3 = NIL_RTGCPHYS;
2628	pVCpu->pgm.s.GCPhysNstGstCR3 = NIL_RTGCPHYS;
2629	return rc2;
2630	}
2631	}
2632	#endif
2633
2634	LogFlowFunc(("cr3=%RX64 old=%RX64 fGlobal=%d\n", cr3, GCPhysOldCR3, fGlobal));
2635	int rc = VINF_SUCCESS;
2636	if (GCPhysOldCR3 != GCPhysCR3)
2637	{
2638	uintptr_t const idxBth = pVCpu->pgm.s.idxBothModeData;
2639	AssertReturn(idxBth < RT_ELEMENTS(g_aPgmBothModeData), VERR_PGM_MODE_IPE);
2640	AssertReturn(g_aPgmBothModeData[idxBth].pfnMapCR3, VERR_PGM_MODE_IPE);
2641
2642	pVCpu->pgm.s.GCPhysCR3 = GCPhysCR3;
2643	rc = g_aPgmBothModeData[idxBth].pfnMapCR3(pVCpu, GCPhysCR3);
2644	if (RT_LIKELY(rc == VINF_SUCCESS))
2645	{ }
2646	else
2647	{
2648	AssertMsg(rc == VINF_PGM_SYNC_CR3, ("%Rrc\n", rc));
2649	Assert(VMCPU_FF_IS_ANY_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3_NON_GLOBAL \| VMCPU_FF_PGM_SYNC_CR3));
2650	pVCpu->pgm.s.CTX_SUFF(fPaePdpesAndCr3Mapped) = false;
2651	pVCpu->pgm.s.GCPhysPaeCR3 = NIL_RTGCPHYS;
2652	pVCpu->pgm.s.GCPhysCR3 = GCPhysOldCR3;
2653	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_MAP_CR3;
2654	}
2655
2656	if (fGlobal)
2657	STAM_COUNTER_INC(&pVCpu->pgm.s.Stats.CTX_MID_Z(Stat,FlushTLBNewCR3Global));
2658	else
2659	STAM_COUNTER_INC(&pVCpu->pgm.s.Stats.CTX_MID_Z(Stat,FlushTLBNewCR3));
2660	}
2661	else
2662	{
2663	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
2664	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
2665	if (pPool->cDirtyPages)
2666	{
2667	PGM_LOCK_VOID(pVM);
2668	pgmPoolResetDirtyPages(pVM);
2669	PGM_UNLOCK(pVM);
2670	}
2671	#endif
2672	if (fGlobal)
2673	STAM_COUNTER_INC(&pVCpu->pgm.s.Stats.CTX_MID_Z(Stat,FlushTLBSameCR3Global));
2674	else
2675	STAM_COUNTER_INC(&pVCpu->pgm.s.Stats.CTX_MID_Z(Stat,FlushTLBSameCR3));
2676
2677	/*
2678	* Flush PAE PDPTEs.
2679	*/
2680	if (PGMMODE_IS_PAE(pVCpu->pgm.s.enmGuestMode))
2681	pgmGstFlushPaePdpes(pVCpu);
2682	}
2683
2684	if (!fGlobal)
2685	IEMTlbInvalidateAll(pVCpu);
2686	else
2687	IEMTlbInvalidateAllGlobal(pVCpu);
2688	STAM_PROFILE_STOP(&pVCpu->pgm.s.Stats.CTX_MID_Z(Stat,FlushTLB), a);
2689	return rc;
2690	}
2691
2692
2693	/**
2694	* Performs and schedules necessary updates following a CR3 load or reload when
2695	* using nested or extended paging.
2696	*
2697	* This API is an alternative to PGMFlushTLB that avoids actually flushing the
2698	* TLB and triggering a SyncCR3.
2699	*
2700	* This will normally involve mapping the guest PD or nPDPT
2701	*
2702	* @returns VBox status code.
2703	* @retval VINF_SUCCESS.
2704	* @retval VINF_PGM_SYNC_CR3 if monitoring requires a CR3 sync (not for nested
2705	* paging modes). This can safely be ignored and overridden since the
2706	* FF will be set too then.
2707	* @param pVCpu The cross context virtual CPU structure.
2708	* @param cr3 The new CR3.
2709	*/
2710	VMMDECL(int) PGMUpdateCR3(PVMCPUCC pVCpu, uint64_t cr3)
2711	{
2712	VMCPU_ASSERT_EMT(pVCpu);
2713
2714	/* We assume we're only called in nested paging mode. */
2715	Assert(pVCpu->CTX_SUFF(pVM)->pgm.s.fNestedPaging \|\| pVCpu->pgm.s.enmShadowMode == PGMMODE_EPT);
2716
2717	/*
2718	* Remap the CR3 content and adjust the monitoring if CR3 was actually changed.
2719	*/
2720	RTGCPHYS const GCPhysOldCR3 = pVCpu->pgm.s.GCPhysCR3;
2721	RTGCPHYS GCPhysCR3 = pgmGetGuestMaskedCr3(pVCpu, cr3);
2722	#ifdef VBOX_WITH_NESTED_HWVIRT_VMX_EPT
2723	if (pVCpu->pgm.s.enmGuestSlatMode == PGMSLAT_EPT)
2724	{
2725	RTGCPHYS GCPhysOut;
2726	int const rc = pgmGstSlatTranslateCr3(pVCpu, GCPhysCR3, &GCPhysOut);
2727	if (RT_SUCCESS(rc))
2728	GCPhysCR3 = GCPhysOut;
2729	else
2730	{
2731	/* CR3 SLAT translation failed but we try to pretend it
2732	succeeded for the reasons mentioned in PGMHCChangeMode(). */
2733	Log(("SLAT failed for CR3 %#RX64 rc=%Rrc\n", cr3, rc));
2734	int const rc2 = pgmGstUnmapCr3(pVCpu);
2735	pVCpu->pgm.s.GCPhysCR3 = NIL_RTGCPHYS;
2736	pVCpu->pgm.s.GCPhysNstGstCR3 = NIL_RTGCPHYS;
2737	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_HM_UPDATE_CR3);
2738	return rc2;
2739	}
2740	}
2741	#endif
2742
2743	LogFlowFunc(("cr3=%RX64 old=%RX64\n", cr3, GCPhysOldCR3));
2744	int rc = VINF_SUCCESS;
2745	if (GCPhysOldCR3 != GCPhysCR3)
2746	{
2747	uintptr_t const idxBth = pVCpu->pgm.s.idxBothModeData;
2748	AssertReturn(idxBth < RT_ELEMENTS(g_aPgmBothModeData), VERR_PGM_MODE_IPE);
2749	AssertReturn(g_aPgmBothModeData[idxBth].pfnMapCR3, VERR_PGM_MODE_IPE);
2750
2751	pVCpu->pgm.s.GCPhysCR3 = GCPhysCR3;
2752	rc = g_aPgmBothModeData[idxBth].pfnMapCR3(pVCpu, GCPhysCR3);
2753
2754	AssertRCSuccess(rc); /* Assumes VINF_PGM_SYNC_CR3 doesn't apply to nested paging. / /* @todo this isn't true for the mac, but we need hw to test/fix this. */
2755	}
2756	/*
2757	* Flush PAE PDPTEs.
2758	*/
2759	else if (PGMMODE_IS_PAE(pVCpu->pgm.s.enmGuestMode))
2760	pgmGstFlushPaePdpes(pVCpu);
2761
2762	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_HM_UPDATE_CR3);
2763	return rc;
2764	}
2765
2766
2767	/**
2768	* Synchronize the paging structures.
2769	*
2770	* This function is called in response to the VM_FF_PGM_SYNC_CR3 and
2771	* VM_FF_PGM_SYNC_CR3_NONGLOBAL. Those two force action flags are set
2772	* in several places, most importantly whenever the CR3 is loaded.
2773	*
2774	* @returns VBox status code. May return VINF_PGM_SYNC_CR3 in RC/R0.
2775	* @retval VERR_PGM_NO_HYPERVISOR_ADDRESS in raw-mode when we're unable to map
2776	* the VMM into guest context.
2777	* @param pVCpu The cross context virtual CPU structure.
2778	* @param cr0 Guest context CR0 register
2779	* @param cr3 Guest context CR3 register
2780	* @param cr4 Guest context CR4 register
2781	* @param fGlobal Including global page directories or not
2782	*/
2783	VMMDECL(int) PGMSyncCR3(PVMCPUCC pVCpu, uint64_t cr0, uint64_t cr3, uint64_t cr4, bool fGlobal)
2784	{
2785	int rc;
2786
2787	VMCPU_ASSERT_EMT(pVCpu);
2788
2789	/*
2790	* The pool may have pending stuff and even require a return to ring-3 to
2791	* clear the whole thing.
2792	*/
2793	rc = pgmPoolSyncCR3(pVCpu);
2794	if (rc != VINF_SUCCESS)
2795	return rc;
2796
2797	/*
2798	* We might be called when we shouldn't.
2799	*
2800	* The mode switching will ensure that the PD is resynced after every mode
2801	* switch. So, if we find ourselves here when in protected or real mode
2802	* we can safely clear the FF and return immediately.
2803	*/
2804	if (pVCpu->pgm.s.enmGuestMode <= PGMMODE_PROTECTED)
2805	{
2806	Assert((cr0 & (X86_CR0_PG \| X86_CR0_PE)) != (X86_CR0_PG \| X86_CR0_PE));
2807	Assert(!(pVCpu->pgm.s.fSyncFlags & PGM_SYNC_CLEAR_PGM_POOL));
2808	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
2809	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_PGM_SYNC_CR3_NON_GLOBAL);
2810	return VINF_SUCCESS;
2811	}
2812
2813	/* If global pages are not supported, then all flushes are global. */
2814	if (!(cr4 & X86_CR4_PGE))
2815	fGlobal = true;
2816	LogFlow(("PGMSyncCR3: cr0=%RX64 cr3=%RX64 cr4=%RX64 fGlobal=%d[%d,%d]\n", cr0, cr3, cr4, fGlobal,
2817	VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3), VMCPU_FF_IS_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3_NON_GLOBAL)));
2818
2819	/*
2820	* Check if we need to finish an aborted MapCR3 call (see PGMFlushTLB).
2821	* This should be done before SyncCR3.
2822	*/
2823	if (pVCpu->pgm.s.fSyncFlags & PGM_SYNC_MAP_CR3)
2824	{
2825	pVCpu->pgm.s.fSyncFlags &= ~PGM_SYNC_MAP_CR3;
2826
2827	RTGCPHYS const GCPhysOldCR3 = pVCpu->pgm.s.GCPhysCR3;
2828	RTGCPHYS GCPhysCR3 = pgmGetGuestMaskedCr3(pVCpu, cr3);
2829	#ifdef VBOX_WITH_NESTED_HWVIRT_VMX_EPT
2830	if (pVCpu->pgm.s.enmGuestSlatMode == PGMSLAT_EPT)
2831	{
2832	RTGCPHYS GCPhysOut;
2833	int rc2 = pgmGstSlatTranslateCr3(pVCpu, GCPhysCR3, &GCPhysOut);
2834	if (RT_SUCCESS(rc2))
2835	GCPhysCR3 = GCPhysOut;
2836	else
2837	{
2838	/* CR3 SLAT translation failed but we try to pretend it
2839	succeeded for the reasons mentioned in PGMHCChangeMode(). */
2840	AssertMsgFailed(("Failed to translate CR3 %#RX64. rc=%Rrc\n", cr3, rc2));
2841	pVCpu->pgm.s.GCPhysCR3 = NIL_RTGCPHYS;
2842	pVCpu->pgm.s.GCPhysNstGstCR3 = NIL_RTGCPHYS;
2843	return rc2;
2844	}
2845	}
2846	#endif
2847	Assert(!pVCpu->pgm.s.CTX_SUFF(fPaePdpesAndCr3Mapped));
2848	if (GCPhysOldCR3 != GCPhysCR3)
2849	{
2850	uintptr_t const idxBth = pVCpu->pgm.s.idxBothModeData;
2851	AssertReturn(idxBth < RT_ELEMENTS(g_aPgmBothModeData), VERR_PGM_MODE_IPE);
2852	AssertReturn(g_aPgmBothModeData[idxBth].pfnMapCR3, VERR_PGM_MODE_IPE);
2853	pVCpu->pgm.s.GCPhysCR3 = GCPhysCR3;
2854	rc = g_aPgmBothModeData[idxBth].pfnMapCR3(pVCpu, GCPhysCR3);
2855	}
2856
2857	/* Make sure we check for pending pgm pool syncs as we clear VMCPU_FF_PGM_SYNC_CR3 later on! */
2858	if ( rc == VINF_PGM_SYNC_CR3
2859	\|\| (pVCpu->pgm.s.fSyncFlags & PGM_SYNC_CLEAR_PGM_POOL))
2860	{
2861	Log(("PGMSyncCR3: pending pgm pool sync after MapCR3!\n"));
2862	#ifdef IN_RING3
2863	rc = pgmPoolSyncCR3(pVCpu);
2864	#else
2865	if (rc == VINF_PGM_SYNC_CR3)
2866	pVCpu->pgm.s.GCPhysCR3 = GCPhysOldCR3;
2867	return VINF_PGM_SYNC_CR3;
2868	#endif
2869	}
2870	AssertRCReturn(rc, rc);
2871	AssertRCSuccessReturn(rc, VERR_IPE_UNEXPECTED_INFO_STATUS);
2872	}
2873
2874	/*
2875	* Let the 'Bth' function do the work and we'll just keep track of the flags.
2876	*/
2877	STAM_PROFILE_START(&pVCpu->pgm.s.Stats.CTX_MID_Z(Stat,SyncCR3), a);
2878
2879	uintptr_t const idxBth = pVCpu->pgm.s.idxBothModeData;
2880	AssertReturn(idxBth < RT_ELEMENTS(g_aPgmBothModeData), VERR_PGM_MODE_IPE);
2881	AssertReturn(g_aPgmBothModeData[idxBth].pfnSyncCR3, VERR_PGM_MODE_IPE);
2882	rc = g_aPgmBothModeData[idxBth].pfnSyncCR3(pVCpu, cr0, cr3, cr4, fGlobal);
2883
2884	STAM_PROFILE_STOP(&pVCpu->pgm.s.Stats.CTX_MID_Z(Stat,SyncCR3), a);
2885	AssertMsg(rc == VINF_SUCCESS \|\| rc == VINF_PGM_SYNC_CR3 \|\| RT_FAILURE(rc), ("rc=%Rrc\n", rc));
2886	if (rc == VINF_SUCCESS)
2887	{
2888	if (pVCpu->pgm.s.fSyncFlags & PGM_SYNC_CLEAR_PGM_POOL)
2889	{
2890	/* Go back to ring 3 if a pgm pool sync is again pending. */
2891	return VINF_PGM_SYNC_CR3;
2892	}
2893
2894	if (!(pVCpu->pgm.s.fSyncFlags & PGM_SYNC_ALWAYS))
2895	{
2896	Assert(!(pVCpu->pgm.s.fSyncFlags & PGM_SYNC_CLEAR_PGM_POOL));
2897	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
2898	VMCPU_FF_CLEAR(pVCpu, VMCPU_FF_PGM_SYNC_CR3_NON_GLOBAL);
2899	}
2900	}
2901
2902	/*
2903	* Now flush the CR3 (guest context).
2904	*/
2905	if (rc == VINF_SUCCESS)
2906	PGM_INVL_VCPU_TLBS(pVCpu);
2907	return rc;
2908	}
2909
2910
2911	/**
2912	* Maps all the PAE PDPE entries.
2913	*
2914	* @returns VBox status code.
2915	* @param pVCpu The cross context virtual CPU structure of the calling EMT.
2916	* @param paPaePdpes The new PAE PDPE values.
2917	*
2918	* @remarks This function may be invoked during the process of changing the guest
2919	* paging mode to PAE, hence the guest state (CR0, CR4 etc.) may not
2920	* reflect PAE paging just yet.
2921	*/
2922	VMM_INT_DECL(int) PGMGstMapPaePdpes(PVMCPUCC pVCpu, PCX86PDPE paPaePdpes)
2923	{
2924	Assert(paPaePdpes);
2925	for (unsigned i = 0; i < X86_PG_PAE_PDPE_ENTRIES; i++)
2926	{
2927	X86PDPE const PaePdpe = paPaePdpes[i];
2928
2929	/*
2930	* In some cases (e.g. in SVM with nested paging) the validation of the PAE PDPEs
2931	* are deferred.[1] Also, different situations require different handling of invalid
2932	* PDPE entries. Here we assume the caller has already validated or doesn't require
2933	* validation of the PDPEs.
2934	*
2935	* In the case of nested EPT (i.e. for nested-guests), the PAE PDPEs have been
2936	* validated by the VMX transition.
2937	*
2938	* [1] -- See AMD spec. 15.25.10 "Legacy PAE Mode".
2939	*/
2940	if ((PaePdpe.u & (pVCpu->pgm.s.fGstPaeMbzPdpeMask \| X86_PDPE_P)) == X86_PDPE_P)
2941	{
2942	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
2943	RTHCPTR HCPtr;
2944
2945	RTGCPHYS GCPhys;
2946	#ifdef VBOX_WITH_NESTED_HWVIRT_VMX_EPT
2947	if (pVCpu->pgm.s.enmGuestSlatMode == PGMSLAT_EPT)
2948	{
2949	PGMPTWALK Walk;
2950	PGMPTWALKGST GstWalk;
2951	RTGCPHYS const GCPhysNested = PaePdpe.u & X86_PDPE_PG_MASK;
2952	int const rc = pgmGstSlatWalk(pVCpu, GCPhysNested, false /* fIsLinearAddrValid /, 0 / GCPtrNested */,
2953	&Walk, &GstWalk);
2954	if (RT_SUCCESS(rc))
2955	GCPhys = Walk.GCPhys;
2956	else
2957	{
2958	/*
2959	* Second-level address translation of the PAE PDPE has failed but we must -NOT-
2960	* abort and return a failure now. This is because we're called from a Mov CRx
2961	* instruction (or similar operation). Let's just pretend success but flag that
2962	* we need to map this PDPE lazily later.
2963	*
2964	* See Intel spec. 25.3 "Changes to instruction behavior in VMX non-root operation".
2965	* See Intel spec. 28.3.1 "EPT Overview".
2966	*/
2967	pVCpu->pgm.s.apGstPaePDsR3[i] = 0;
2968	pVCpu->pgm.s.apGstPaePDsR0[i] = 0;
2969	pVCpu->pgm.s.aGCPhysGstPaePDs[i] = NIL_RTGCPHYS;
2970	continue;
2971	}
2972	}
2973	else
2974	#endif
2975	{
2976	GCPhys = PGM_A20_APPLY(pVCpu, PaePdpe.u & X86_PDPE_PG_MASK);
2977	}
2978
2979	PGM_LOCK_VOID(pVM);
2980	PPGMPAGE pPage = pgmPhysGetPage(pVM, GCPhys);
2981	AssertReturnStmt(pPage, PGM_UNLOCK(pVM), VERR_PGM_INVALID_PDPE_ADDR);
2982	int const rc = pgmPhysGCPhys2CCPtrInternalDepr(pVM, pPage, GCPhys, (void **)&HCPtr);
2983	PGM_UNLOCK(pVM);
2984	if (RT_SUCCESS(rc))
2985	{
2986	#ifdef IN_RING3
2987	pVCpu->pgm.s.apGstPaePDsR3[i] = (PX86PDPAE)HCPtr;
2988	pVCpu->pgm.s.apGstPaePDsR0[i] = NIL_RTR0PTR;
2989	#else
2990	pVCpu->pgm.s.apGstPaePDsR3[i] = NIL_RTR3PTR;
2991	pVCpu->pgm.s.apGstPaePDsR0[i] = (PX86PDPAE)HCPtr;
2992	#endif
2993	pVCpu->pgm.s.aGCPhysGstPaePDs[i] = GCPhys;
2994	continue;
2995	}
2996	AssertMsgFailed(("PGMPhysMapPaePdpes: rc2=%d GCPhys=%RGp i=%d\n", rc, GCPhys, i));
2997	}
2998	pVCpu->pgm.s.apGstPaePDsR3[i] = 0;
2999	pVCpu->pgm.s.apGstPaePDsR0[i] = 0;
3000	pVCpu->pgm.s.aGCPhysGstPaePDs[i] = NIL_RTGCPHYS;
3001	}
3002	return VINF_SUCCESS;
3003	}
3004
3005
3006	/**
3007	* Validates and maps the PDPT and PAE PDPEs referenced by the given CR3.
3008	*
3009	* @returns VBox status code.
3010	* @param pVCpu The cross context virtual CPU structure of the calling EMT.
3011	* @param cr3 The guest CR3 value.
3012	*
3013	* @remarks This function may be invoked during the process of changing the guest
3014	* paging mode to PAE but the guest state (CR0, CR4 etc.) may not reflect
3015	* PAE paging just yet.
3016	*/
3017	VMM_INT_DECL(int) PGMGstMapPaePdpesAtCr3(PVMCPUCC pVCpu, uint64_t cr3)
3018	{
3019	/*
3020	* Read the page-directory-pointer table (PDPT) at CR3.
3021	*/
3022	RTGCPHYS GCPhysCR3 = (cr3 & X86_CR3_PAE_PAGE_MASK);
3023	PGM_A20_APPLY_TO_VAR(pVCpu, GCPhysCR3);
3024
3025	#ifdef VBOX_WITH_NESTED_HWVIRT_VMX_EPT
3026	if (pVCpu->pgm.s.enmGuestSlatMode == PGMSLAT_EPT)
3027	{
3028	RTGCPHYS GCPhysOut;
3029	int const rc = pgmGstSlatTranslateCr3(pVCpu, GCPhysCR3, &GCPhysOut);
3030	if (RT_SUCCESS(rc))
3031	GCPhysCR3 = GCPhysOut;
3032	else
3033	{
3034	Log(("Failed to load CR3 at %#RX64. rc=%Rrc\n", GCPhysCR3, rc));
3035	return rc;
3036	}
3037	}
3038	#endif
3039
3040	RTHCPTR HCPtrGuestCr3;
3041	int rc = pgmGstMapCr3(pVCpu, GCPhysCR3, &HCPtrGuestCr3);
3042	if (RT_SUCCESS(rc))
3043	{
3044	/*
3045	* Validate the page-directory-pointer table entries (PDPE).
3046	*/
3047	X86PDPE aPaePdpes[X86_PG_PAE_PDPE_ENTRIES];
3048	memcpy(&aPaePdpes[0], HCPtrGuestCr3, sizeof(aPaePdpes));
3049	if (PGMGstArePaePdpesValid(pVCpu, &aPaePdpes[0]))
3050	{
3051	/*
3052	* Map the PDPT.
3053	* We deliberately don't update PGM's GCPhysCR3 here as it's expected
3054	* that PGMFlushTLB will be called soon and only a change to CR3 then
3055	* will cause the shadow page tables to be updated.
3056	*/
3057	#ifdef IN_RING3
3058	pVCpu->pgm.s.pGstPaePdptR3 = (PX86PDPT)HCPtrGuestCr3;
3059	pVCpu->pgm.s.pGstPaePdptR0 = NIL_RTR0PTR;
3060	#else
3061	pVCpu->pgm.s.pGstPaePdptR3 = NIL_RTR3PTR;
3062	pVCpu->pgm.s.pGstPaePdptR0 = (PX86PDPT)HCPtrGuestCr3;
3063	#endif
3064
3065	/*
3066	* Update CPUM and map the 4 PAE PDPEs.
3067	*/
3068	CPUMSetGuestPaePdpes(pVCpu, &aPaePdpes[0]);
3069	rc = PGMGstMapPaePdpes(pVCpu, &aPaePdpes[0]);
3070	if (RT_SUCCESS(rc))
3071	{
3072	#ifdef IN_RING3
3073	pVCpu->pgm.s.fPaePdpesAndCr3MappedR3 = true;
3074	pVCpu->pgm.s.fPaePdpesAndCr3MappedR0 = false;
3075	#else
3076	pVCpu->pgm.s.fPaePdpesAndCr3MappedR3 = false;
3077	pVCpu->pgm.s.fPaePdpesAndCr3MappedR0 = true;
3078	#endif
3079	pVCpu->pgm.s.GCPhysPaeCR3 = GCPhysCR3;
3080	}
3081	}
3082	else
3083	rc = VERR_PGM_PAE_PDPE_RSVD;
3084	}
3085	return rc;
3086	}
3087
3088
3089	/**
3090	* Called whenever CR0 or CR4 in a way which may affect the paging mode.
3091	*
3092	* @returns VBox status code, with the following informational code for
3093	* VM scheduling.
3094	* @retval VINF_SUCCESS if the was no change, or it was successfully dealt with.
3095	* @retval VINF_EM_SUSPEND or VINF_EM_OFF on a fatal runtime error. (R3 only)
3096	*
3097	* @param pVCpu The cross context virtual CPU structure.
3098	* @param cr0 The new cr0.
3099	* @param cr4 The new cr4.
3100	* @param efer The new extended feature enable register.
3101	* @param fForce Whether to force a mode change.
3102	*/
3103	VMMDECL(int) PGMChangeMode(PVMCPUCC pVCpu, uint64_t cr0, uint64_t cr4, uint64_t efer, bool fForce)
3104	{
3105	VMCPU_ASSERT_EMT(pVCpu);
3106
3107	/*
3108	* Calc the new guest mode.
3109	*
3110	* Note! We check PG before PE and without requiring PE because of the
3111	* special AMD-V paged real mode (APM vol 2, rev 3.28, 15.9).
3112	*/
3113	PGMMODE enmGuestMode;
3114	if (cr0 & X86_CR0_PG)
3115	{
3116	if (!(cr4 & X86_CR4_PAE))
3117	{
3118	bool const fPse = !!(cr4 & X86_CR4_PSE);
3119	if (pVCpu->pgm.s.fGst32BitPageSizeExtension != fPse)
3120	Log(("PGMChangeMode: CR4.PSE %d -> %d\n", pVCpu->pgm.s.fGst32BitPageSizeExtension, fPse));
3121	pVCpu->pgm.s.fGst32BitPageSizeExtension = fPse;
3122	enmGuestMode = PGMMODE_32_BIT;
3123	}
3124	else if (!(efer & MSR_K6_EFER_LME))
3125	{
3126	if (!(efer & MSR_K6_EFER_NXE))
3127	enmGuestMode = PGMMODE_PAE;
3128	else
3129	enmGuestMode = PGMMODE_PAE_NX;
3130	}
3131	else
3132	{
3133	if (!(efer & MSR_K6_EFER_NXE))
3134	enmGuestMode = PGMMODE_AMD64;
3135	else
3136	enmGuestMode = PGMMODE_AMD64_NX;
3137	}
3138	}
3139	else if (!(cr0 & X86_CR0_PE))
3140	enmGuestMode = PGMMODE_REAL;
3141	else
3142	enmGuestMode = PGMMODE_PROTECTED;
3143
3144	/*
3145	* Did it change?
3146	*/
3147	if ( !fForce
3148	&& pVCpu->pgm.s.enmGuestMode == enmGuestMode)
3149	return VINF_SUCCESS;
3150
3151	/* Flush the TLB */
3152	PGM_INVL_VCPU_TLBS(pVCpu);
3153	return PGMHCChangeMode(pVCpu->CTX_SUFF(pVM), pVCpu, enmGuestMode, fForce);
3154	}
3155
3156
3157	/**
3158	* Converts a PGMMODE value to a PGM_TYPE_* \#define.
3159	*
3160	* @returns PGM_TYPE_*.
3161	* @param pgmMode The mode value to convert.
3162	*/
3163	DECLINLINE(unsigned) pgmModeToType(PGMMODE pgmMode)
3164	{
3165	switch (pgmMode)
3166	{
3167	case PGMMODE_REAL: return PGM_TYPE_REAL;
3168	case PGMMODE_PROTECTED: return PGM_TYPE_PROT;
3169	case PGMMODE_32_BIT: return PGM_TYPE_32BIT;
3170	case PGMMODE_PAE:
3171	case PGMMODE_PAE_NX: return PGM_TYPE_PAE;
3172	case PGMMODE_AMD64:
3173	case PGMMODE_AMD64_NX: return PGM_TYPE_AMD64;
3174	case PGMMODE_NESTED_32BIT: return PGM_TYPE_NESTED_32BIT;
3175	case PGMMODE_NESTED_PAE: return PGM_TYPE_NESTED_PAE;
3176	case PGMMODE_NESTED_AMD64: return PGM_TYPE_NESTED_AMD64;
3177	case PGMMODE_EPT: return PGM_TYPE_EPT;
3178	case PGMMODE_NONE: return PGM_TYPE_NONE;
3179	default:
3180	AssertFatalMsgFailed(("pgmMode=%d\n", pgmMode));
3181	}
3182	}
3183
3184
3185	/**
3186	* Calculates the shadow paging mode.
3187	*
3188	* @returns The shadow paging mode.
3189	* @param pVM The cross context VM structure.
3190	* @param enmGuestMode The guest mode.
3191	* @param enmHostMode The host mode.
3192	* @param enmShadowMode The current shadow mode.
3193	*/
3194	static PGMMODE pgmCalcShadowMode(PVMCC pVM, PGMMODE enmGuestMode, SUPPAGINGMODE enmHostMode, PGMMODE enmShadowMode)
3195	{
3196	switch (enmGuestMode)
3197	{
3198	case PGMMODE_REAL:
3199	case PGMMODE_PROTECTED:
3200	switch (enmHostMode)
3201	{
3202	case SUPPAGINGMODE_32_BIT:
3203	case SUPPAGINGMODE_32_BIT_GLOBAL:
3204	enmShadowMode = PGMMODE_32_BIT;
3205	break;
3206
3207	case SUPPAGINGMODE_PAE:
3208	case SUPPAGINGMODE_PAE_NX:
3209	case SUPPAGINGMODE_PAE_GLOBAL:
3210	case SUPPAGINGMODE_PAE_GLOBAL_NX:
3211	enmShadowMode = PGMMODE_PAE;
3212	break;
3213
3214	case SUPPAGINGMODE_AMD64:
3215	case SUPPAGINGMODE_AMD64_GLOBAL:
3216	case SUPPAGINGMODE_AMD64_NX:
3217	case SUPPAGINGMODE_AMD64_GLOBAL_NX:
3218	enmShadowMode = PGMMODE_PAE;
3219	break;
3220
3221	default:
3222	AssertLogRelMsgFailedReturn(("enmHostMode=%d\n", enmHostMode), PGMMODE_INVALID);
3223	}
3224	break;
3225
3226	case PGMMODE_32_BIT:
3227	switch (enmHostMode)
3228	{
3229	case SUPPAGINGMODE_32_BIT:
3230	case SUPPAGINGMODE_32_BIT_GLOBAL:
3231	enmShadowMode = PGMMODE_32_BIT;
3232	break;
3233
3234	case SUPPAGINGMODE_PAE:
3235	case SUPPAGINGMODE_PAE_NX:
3236	case SUPPAGINGMODE_PAE_GLOBAL:
3237	case SUPPAGINGMODE_PAE_GLOBAL_NX:
3238	enmShadowMode = PGMMODE_PAE;
3239	break;
3240
3241	case SUPPAGINGMODE_AMD64:
3242	case SUPPAGINGMODE_AMD64_GLOBAL:
3243	case SUPPAGINGMODE_AMD64_NX:
3244	case SUPPAGINGMODE_AMD64_GLOBAL_NX:
3245	enmShadowMode = PGMMODE_PAE;
3246	break;
3247
3248	default:
3249	AssertLogRelMsgFailedReturn(("enmHostMode=%d\n", enmHostMode), PGMMODE_INVALID);
3250	}
3251	break;
3252
3253	case PGMMODE_PAE:
3254	case PGMMODE_PAE_NX: /** @todo This might require more switchers and guest+both modes. */
3255	switch (enmHostMode)
3256	{
3257	case SUPPAGINGMODE_32_BIT:
3258	case SUPPAGINGMODE_32_BIT_GLOBAL:
3259	enmShadowMode = PGMMODE_PAE;
3260	break;
3261
3262	case SUPPAGINGMODE_PAE:
3263	case SUPPAGINGMODE_PAE_NX:
3264	case SUPPAGINGMODE_PAE_GLOBAL:
3265	case SUPPAGINGMODE_PAE_GLOBAL_NX:
3266	enmShadowMode = PGMMODE_PAE;
3267	break;
3268
3269	case SUPPAGINGMODE_AMD64:
3270	case SUPPAGINGMODE_AMD64_GLOBAL:
3271	case SUPPAGINGMODE_AMD64_NX:
3272	case SUPPAGINGMODE_AMD64_GLOBAL_NX:
3273	enmShadowMode = PGMMODE_PAE;
3274	break;
3275
3276	default:
3277	AssertLogRelMsgFailedReturn(("enmHostMode=%d\n", enmHostMode), PGMMODE_INVALID);
3278	}
3279	break;
3280
3281	case PGMMODE_AMD64:
3282	case PGMMODE_AMD64_NX:
3283	switch (enmHostMode)
3284	{
3285	case SUPPAGINGMODE_32_BIT:
3286	case SUPPAGINGMODE_32_BIT_GLOBAL:
3287	enmShadowMode = PGMMODE_AMD64;
3288	break;
3289
3290	case SUPPAGINGMODE_PAE:
3291	case SUPPAGINGMODE_PAE_NX:
3292	case SUPPAGINGMODE_PAE_GLOBAL:
3293	case SUPPAGINGMODE_PAE_GLOBAL_NX:
3294	enmShadowMode = PGMMODE_AMD64;
3295	break;
3296
3297	case SUPPAGINGMODE_AMD64:
3298	case SUPPAGINGMODE_AMD64_GLOBAL:
3299	case SUPPAGINGMODE_AMD64_NX:
3300	case SUPPAGINGMODE_AMD64_GLOBAL_NX:
3301	enmShadowMode = PGMMODE_AMD64;
3302	break;
3303
3304	default:
3305	AssertLogRelMsgFailedReturn(("enmHostMode=%d\n", enmHostMode), PGMMODE_INVALID);
3306	}
3307	break;
3308
3309	default:
3310	AssertLogRelMsgFailedReturn(("enmGuestMode=%d\n", enmGuestMode), PGMMODE_INVALID);
3311	}
3312
3313	/*
3314	* Override the shadow mode when NEM, IEM or nested paging is active.
3315	*/
3316	if (!VM_IS_HM_ENABLED(pVM))
3317	{
3318	Assert(VM_IS_NEM_ENABLED(pVM) \|\| VM_IS_EXEC_ENGINE_IEM(pVM));
3319	pVM->pgm.s.fNestedPaging = true;
3320	enmShadowMode = PGMMODE_NONE;
3321	}
3322	else
3323	{
3324	bool fNestedPaging = HMIsNestedPagingActive(pVM);
3325	pVM->pgm.s.fNestedPaging = fNestedPaging;
3326	if (fNestedPaging)
3327	{
3328	if (HMIsVmxActive(pVM))
3329	enmShadowMode = PGMMODE_EPT;
3330	else
3331	{
3332	/* The nested SVM paging depends on the host one. */
3333	Assert(HMIsSvmActive(pVM));
3334	if ( enmGuestMode == PGMMODE_AMD64
3335	\|\| enmGuestMode == PGMMODE_AMD64_NX)
3336	enmShadowMode = PGMMODE_NESTED_AMD64;
3337	else
3338	switch (pVM->pgm.s.enmHostMode)
3339	{
3340	case SUPPAGINGMODE_32_BIT:
3341	case SUPPAGINGMODE_32_BIT_GLOBAL:
3342	enmShadowMode = PGMMODE_NESTED_32BIT;
3343	break;
3344
3345	case SUPPAGINGMODE_PAE:
3346	case SUPPAGINGMODE_PAE_GLOBAL:
3347	case SUPPAGINGMODE_PAE_NX:
3348	case SUPPAGINGMODE_PAE_GLOBAL_NX:
3349	enmShadowMode = PGMMODE_NESTED_PAE;
3350	break;
3351
3352	case SUPPAGINGMODE_AMD64:
3353	case SUPPAGINGMODE_AMD64_GLOBAL:
3354	case SUPPAGINGMODE_AMD64_NX:
3355	case SUPPAGINGMODE_AMD64_GLOBAL_NX:
3356	enmShadowMode = PGMMODE_NESTED_AMD64;
3357	break;
3358
3359	default:
3360	AssertLogRelMsgFailedReturn(("enmHostMode=%d\n", pVM->pgm.s.enmHostMode), PGMMODE_INVALID);
3361	}
3362	}
3363	}
3364	#ifdef VBOX_WITH_NESTED_HWVIRT_VMX_EPT
3365	else
3366	{
3367	/* Nested paging is a requirement for nested VT-x. */
3368	AssertLogRelMsgReturn(enmGuestMode != PGMMODE_EPT, ("enmHostMode=%d\n", pVM->pgm.s.enmHostMode), PGMMODE_INVALID);
3369	}
3370	#endif
3371	}
3372
3373	return enmShadowMode;
3374	}
3375
3376
3377	/**
3378	* Performs the actual mode change.
3379	* This is called by PGMChangeMode and pgmR3InitPaging().
3380	*
3381	* @returns VBox status code. May suspend or power off the VM on error, but this
3382	* will trigger using FFs and not informational status codes.
3383	*
3384	* @param pVM The cross context VM structure.
3385	* @param pVCpu The cross context virtual CPU structure.
3386	* @param enmGuestMode The new guest mode. This is assumed to be different from
3387	* the current mode.
3388	* @param fForce Whether to force a shadow paging mode change.
3389	*/
3390	VMM_INT_DECL(int) PGMHCChangeMode(PVMCC pVM, PVMCPUCC pVCpu, PGMMODE enmGuestMode, bool fForce)
3391	{
3392	Log(("PGMHCChangeMode: Guest mode: %s -> %s\n", PGMGetModeName(pVCpu->pgm.s.enmGuestMode), PGMGetModeName(enmGuestMode)));
3393	STAM_REL_COUNTER_INC(&pVCpu->pgm.s.cGuestModeChanges);
3394
3395	/*
3396	* Calc the shadow mode and switcher.
3397	*/
3398	PGMMODE const enmShadowMode = pgmCalcShadowMode(pVM, enmGuestMode, pVM->pgm.s.enmHostMode, pVCpu->pgm.s.enmShadowMode);
3399	bool const fShadowModeChanged = enmShadowMode != pVCpu->pgm.s.enmShadowMode \|\| fForce;
3400
3401	/*
3402	* Exit old mode(s).
3403	*/
3404	/* shadow */
3405	if (fShadowModeChanged)
3406	{
3407	LogFlow(("PGMHCChangeMode: Shadow mode: %s -> %s\n", PGMGetModeName(pVCpu->pgm.s.enmShadowMode), PGMGetModeName(enmShadowMode)));
3408	uintptr_t idxOldShw = pVCpu->pgm.s.idxShadowModeData;
3409	if ( idxOldShw < RT_ELEMENTS(g_aPgmShadowModeData)
3410	&& g_aPgmShadowModeData[idxOldShw].pfnExit)
3411	{
3412	int rc = g_aPgmShadowModeData[idxOldShw].pfnExit(pVCpu);
3413	AssertMsgRCReturn(rc, ("Exit failed for shadow mode %d: %Rrc\n", pVCpu->pgm.s.enmShadowMode, rc), rc);
3414	}
3415	}
3416	else
3417	LogFlow(("PGMHCChangeMode: Shadow mode remains: %s\n", PGMGetModeName(pVCpu->pgm.s.enmShadowMode)));
3418
3419	/* guest */
3420	uintptr_t const idxOldGst = pVCpu->pgm.s.idxGuestModeData;
3421	if ( idxOldGst < RT_ELEMENTS(g_aPgmGuestModeData)
3422	&& g_aPgmGuestModeData[idxOldGst].pfnExit)
3423	{
3424	int rc = g_aPgmGuestModeData[idxOldGst].pfnExit(pVCpu);
3425	AssertMsgReturn(RT_SUCCESS(rc), ("Exit failed for guest mode %d: %Rrc\n", pVCpu->pgm.s.enmGuestMode, rc), rc);
3426	}
3427	pVCpu->pgm.s.GCPhysCR3 = NIL_RTGCPHYS;
3428	pVCpu->pgm.s.GCPhysNstGstCR3 = NIL_RTGCPHYS;
3429	pVCpu->pgm.s.GCPhysPaeCR3 = NIL_RTGCPHYS;
3430	Assert(!pVCpu->pgm.s.CTX_SUFF(fPaePdpesAndCr3Mapped));
3431
3432	/*
3433	* Change the paging mode data indexes.
3434	*/
3435	uintptr_t idxNewGst = pVCpu->pgm.s.idxGuestModeData = pgmModeToType(enmGuestMode);
3436	AssertReturn(idxNewGst < RT_ELEMENTS(g_aPgmGuestModeData), VERR_PGM_MODE_IPE);
3437	AssertReturn(g_aPgmGuestModeData[idxNewGst].uType == idxNewGst, VERR_PGM_MODE_IPE);
3438	AssertPtrReturn(g_aPgmGuestModeData[idxNewGst].pfnGetPage, VERR_PGM_MODE_IPE);
3439	AssertPtrReturn(g_aPgmGuestModeData[idxNewGst].pfnModifyPage, VERR_PGM_MODE_IPE);
3440	AssertPtrReturn(g_aPgmGuestModeData[idxNewGst].pfnExit, VERR_PGM_MODE_IPE);
3441	AssertPtrReturn(g_aPgmGuestModeData[idxNewGst].pfnEnter, VERR_PGM_MODE_IPE);
3442	#ifdef IN_RING3
3443	AssertPtrReturn(g_aPgmGuestModeData[idxNewGst].pfnRelocate, VERR_PGM_MODE_IPE);
3444	#endif
3445
3446	uintptr_t const idxNewShw = pVCpu->pgm.s.idxShadowModeData = pgmModeToType(enmShadowMode);
3447	AssertReturn(idxNewShw < RT_ELEMENTS(g_aPgmShadowModeData), VERR_PGM_MODE_IPE);
3448	AssertReturn(g_aPgmShadowModeData[idxNewShw].uType == idxNewShw, VERR_PGM_MODE_IPE);
3449	AssertPtrReturn(g_aPgmShadowModeData[idxNewShw].pfnGetPage, VERR_PGM_MODE_IPE);
3450	AssertPtrReturn(g_aPgmShadowModeData[idxNewShw].pfnModifyPage, VERR_PGM_MODE_IPE);
3451	AssertPtrReturn(g_aPgmShadowModeData[idxNewShw].pfnExit, VERR_PGM_MODE_IPE);
3452	AssertPtrReturn(g_aPgmShadowModeData[idxNewShw].pfnEnter, VERR_PGM_MODE_IPE);
3453	#ifdef IN_RING3
3454	AssertPtrReturn(g_aPgmShadowModeData[idxNewShw].pfnRelocate, VERR_PGM_MODE_IPE);
3455	#endif
3456
3457	uintptr_t const idxNewBth = pVCpu->pgm.s.idxBothModeData = (idxNewShw - PGM_TYPE_FIRST_SHADOW) * PGM_TYPE_END + idxNewGst;
3458	AssertReturn(g_aPgmBothModeData[idxNewBth].uShwType == idxNewShw, VERR_PGM_MODE_IPE);
3459	AssertReturn(g_aPgmBothModeData[idxNewBth].uGstType == idxNewGst, VERR_PGM_MODE_IPE);
3460	AssertPtrReturn(g_aPgmBothModeData[idxNewBth].pfnInvalidatePage, VERR_PGM_MODE_IPE);
3461	AssertPtrReturn(g_aPgmBothModeData[idxNewBth].pfnSyncCR3, VERR_PGM_MODE_IPE);
3462	AssertPtrReturn(g_aPgmBothModeData[idxNewBth].pfnPrefetchPage, VERR_PGM_MODE_IPE);
3463	AssertPtrReturn(g_aPgmBothModeData[idxNewBth].pfnVerifyAccessSyncPage, VERR_PGM_MODE_IPE);
3464	AssertPtrReturn(g_aPgmBothModeData[idxNewBth].pfnMapCR3, VERR_PGM_MODE_IPE);
3465	AssertPtrReturn(g_aPgmBothModeData[idxNewBth].pfnUnmapCR3, VERR_PGM_MODE_IPE);
3466	AssertPtrReturn(g_aPgmBothModeData[idxNewBth].pfnEnter, VERR_PGM_MODE_IPE);
3467	#ifdef VBOX_STRICT
3468	AssertPtrReturn(g_aPgmBothModeData[idxNewBth].pfnAssertCR3, VERR_PGM_MODE_IPE);
3469	#endif
3470
3471	/*
3472	* Determine SLAT mode -before- entering the new shadow mode!
3473	*/
3474	pVCpu->pgm.s.enmGuestSlatMode = !CPUMIsGuestVmxEptPagingEnabled(pVCpu) ? PGMSLAT_DIRECT : PGMSLAT_EPT;
3475
3476	/*
3477	* Enter new shadow mode (if changed).
3478	*/
3479	if (fShadowModeChanged)
3480	{
3481	pVCpu->pgm.s.enmShadowMode = enmShadowMode;
3482	int rc = g_aPgmShadowModeData[idxNewShw].pfnEnter(pVCpu);
3483	AssertLogRelMsgRCReturnStmt(rc, ("Entering enmShadowMode=%s failed: %Rrc\n", PGMGetModeName(enmShadowMode), rc),
3484	pVCpu->pgm.s.enmShadowMode = PGMMODE_INVALID, rc);
3485	}
3486
3487	/*
3488	* Always flag the necessary updates
3489	*/
3490	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
3491
3492	/*
3493	* Enter the new guest and shadow+guest modes.
3494	*/
3495	/* Calc the new CR3 value. */
3496	RTGCPHYS GCPhysCR3;
3497	switch (enmGuestMode)
3498	{
3499	case PGMMODE_REAL:
3500	case PGMMODE_PROTECTED:
3501	GCPhysCR3 = NIL_RTGCPHYS;
3502	break;
3503
3504	case PGMMODE_32_BIT:
3505	GCPhysCR3 = CPUMGetGuestCR3(pVCpu) & X86_CR3_PAGE_MASK;
3506	break;
3507
3508	case PGMMODE_PAE_NX:
3509	case PGMMODE_PAE:
3510	if (!pVM->cpum.ro.GuestFeatures.fPae)
3511	#ifdef IN_RING3 /** @todo r=bird: wrong place, probably hasn't really worked for a while. */
3512	return VMSetRuntimeError(pVM, VMSETRTERR_FLAGS_FATAL, "PAEmode",
3513	N_("The guest is trying to switch to the PAE mode which is currently disabled by default in VirtualBox. PAE support can be enabled using the VM settings (System/Processor)"));
3514	#else
3515	AssertLogRelMsgFailedReturn(("enmGuestMode=%s - Try enable PAE for the guest!\n", PGMGetModeName(enmGuestMode)), VERR_PGM_MODE_IPE);
3516
3517	#endif
3518	GCPhysCR3 = CPUMGetGuestCR3(pVCpu) & X86_CR3_PAE_PAGE_MASK;
3519	break;
3520
3521	#ifdef VBOX_WITH_64_BITS_GUESTS
3522	case PGMMODE_AMD64_NX:
3523	case PGMMODE_AMD64:
3524	GCPhysCR3 = CPUMGetGuestCR3(pVCpu) & X86_CR3_AMD64_PAGE_MASK;
3525	break;
3526	#endif
3527	default:
3528	AssertLogRelMsgFailedReturn(("enmGuestMode=%d\n", enmGuestMode), VERR_PGM_MODE_IPE);
3529	}
3530
3531	#ifdef VBOX_WITH_NESTED_HWVIRT_VMX_EPT
3532	/*
3533	* If a nested-guest is using EPT paging:
3534	* - Update the second-level address translation (SLAT) mode.
3535	* - Indicate that the CR3 is nested-guest physical address.
3536	*/
3537	if (pVCpu->pgm.s.enmGuestSlatMode == PGMSLAT_EPT)
3538	{
3539	if (PGMMODE_WITH_PAGING(enmGuestMode))
3540	{
3541	/*
3542	* Translate CR3 to its guest-physical address.
3543	* We don't use pgmGstSlatTranslateCr3() here as we want to update GCPhysNstGstCR3 -after-
3544	* switching modes to keep it consistent with how GCPhysCR3 is updated.
3545	*/
3546	PGMPTWALK Walk;
3547	PGMPTWALKGST GstWalk;
3548	int const rc = pgmGstSlatWalk(pVCpu, GCPhysCR3, false /* fIsLinearAddrValid /, 0 / GCPtrNested */, &Walk,
3549	&GstWalk);
3550	if (RT_SUCCESS(rc))
3551	{ /* likely */ }
3552	else
3553	{
3554	/*
3555	* SLAT failed but we avoid reporting this to the caller because the caller
3556	* is not supposed to fail. The only time the caller needs to indicate a
3557	* failure to software is when PAE paging is used by the nested-guest, but
3558	* we handle the PAE case separately (e.g., see VMX transition in IEM).
3559	* In all other cases, the failure will be indicated when CR3 tries to be
3560	* translated on the next linear-address memory access.
3561	* See Intel spec. 27.2.1 "EPT Overview".
3562	*/
3563	Log(("SLAT failed for CR3 %#RX64 rc=%Rrc\n", GCPhysCR3, rc));
3564
3565	/* Trying to coax PGM to succeed for the time being... */
3566	Assert(pVCpu->pgm.s.GCPhysCR3 == NIL_RTGCPHYS);
3567	pVCpu->pgm.s.GCPhysNstGstCR3 = GCPhysCR3;
3568	pVCpu->pgm.s.enmGuestMode = enmGuestMode;
3569	HMHCChangedPagingMode(pVM, pVCpu, pVCpu->pgm.s.enmShadowMode, pVCpu->pgm.s.enmGuestMode);
3570	return VINF_SUCCESS;
3571	}
3572	pVCpu->pgm.s.GCPhysNstGstCR3 = GCPhysCR3;
3573	GCPhysCR3 = Walk.GCPhys & X86_CR3_EPT_PAGE_MASK;
3574	}
3575	}
3576	else
3577	Assert(pVCpu->pgm.s.GCPhysNstGstCR3 == NIL_RTGCPHYS);
3578	#endif
3579
3580	/*
3581	* Enter the new guest mode.
3582	*/
3583	pVCpu->pgm.s.enmGuestMode = enmGuestMode;
3584	int rc = g_aPgmGuestModeData[idxNewGst].pfnEnter(pVCpu, GCPhysCR3);
3585	int rc2 = g_aPgmBothModeData[idxNewBth].pfnEnter(pVCpu, GCPhysCR3);
3586
3587	/* Set the new guest CR3 (and nested-guest CR3). */
3588	pVCpu->pgm.s.GCPhysCR3 = GCPhysCR3;
3589
3590	/* status codes. */
3591	AssertRC(rc);
3592	AssertRC(rc2);
3593	if (RT_SUCCESS(rc))
3594	{
3595	rc = rc2;
3596	if (RT_SUCCESS(rc)) /* no informational status codes. */
3597	rc = VINF_SUCCESS;
3598	}
3599
3600	/*
3601	* Notify HM.
3602	*/
3603	HMHCChangedPagingMode(pVM, pVCpu, pVCpu->pgm.s.enmShadowMode, pVCpu->pgm.s.enmGuestMode);
3604	return rc;
3605	}
3606
3607
3608	/**
3609	* Called by CPUM or REM when CR0.WP changes to 1.
3610	*
3611	* @param pVCpu The cross context virtual CPU structure of the calling EMT.
3612	* @thread EMT
3613	*/
3614	VMMDECL(void) PGMCr0WpEnabled(PVMCPUCC pVCpu)
3615	{
3616	/*
3617	* Netware WP0+RO+US hack cleanup when WP0 -> WP1.
3618	*
3619	* Use the counter to judge whether there might be pool pages with active
3620	* hacks in them. If there are, we will be running the risk of messing up
3621	* the guest by allowing it to write to read-only pages. Thus, we have to
3622	* clear the page pool ASAP if there is the slightest chance.
3623	*/
3624	if (pVCpu->pgm.s.cNetwareWp0Hacks > 0)
3625	{
3626	Assert(pVCpu->CTX_SUFF(pVM)->cCpus == 1);
3627
3628	Log(("PGMCr0WpEnabled: %llu WP0 hacks active - clearing page pool\n", pVCpu->pgm.s.cNetwareWp0Hacks));
3629	pVCpu->pgm.s.cNetwareWp0Hacks = 0;
3630	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_CLEAR_PGM_POOL;
3631	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
3632	}
3633	}
3634
3635
3636	/**
3637	* Gets the current guest paging mode.
3638	*
3639	* If you just need the CPU mode (real/protected/long), use CPUMGetGuestMode().
3640	*
3641	* @returns The current paging mode.
3642	* @param pVCpu The cross context virtual CPU structure.
3643	*/
3644	VMMDECL(PGMMODE) PGMGetGuestMode(PVMCPU pVCpu)
3645	{
3646	return pVCpu->pgm.s.enmGuestMode;
3647	}
3648
3649
3650	/**
3651	* Gets the current shadow paging mode.
3652	*
3653	* @returns The current paging mode.
3654	* @param pVCpu The cross context virtual CPU structure.
3655	*/
3656	VMMDECL(PGMMODE) PGMGetShadowMode(PVMCPU pVCpu)
3657	{
3658	return pVCpu->pgm.s.enmShadowMode;
3659	}
3660
3661
3662	/**
3663	* Gets the current host paging mode.
3664	*
3665	* @returns The current paging mode.
3666	* @param pVM The cross context VM structure.
3667	*/
3668	VMMDECL(PGMMODE) PGMGetHostMode(PVM pVM)
3669	{
3670	switch (pVM->pgm.s.enmHostMode)
3671	{
3672	case SUPPAGINGMODE_32_BIT:
3673	case SUPPAGINGMODE_32_BIT_GLOBAL:
3674	return PGMMODE_32_BIT;
3675
3676	case SUPPAGINGMODE_PAE:
3677	case SUPPAGINGMODE_PAE_GLOBAL:
3678	return PGMMODE_PAE;
3679
3680	case SUPPAGINGMODE_PAE_NX:
3681	case SUPPAGINGMODE_PAE_GLOBAL_NX:
3682	return PGMMODE_PAE_NX;
3683
3684	case SUPPAGINGMODE_AMD64:
3685	case SUPPAGINGMODE_AMD64_GLOBAL:
3686	return PGMMODE_AMD64;
3687
3688	case SUPPAGINGMODE_AMD64_NX:
3689	case SUPPAGINGMODE_AMD64_GLOBAL_NX:
3690	return PGMMODE_AMD64_NX;
3691
3692	default: AssertMsgFailed(("enmHostMode=%d\n", pVM->pgm.s.enmHostMode)); break;
3693	}
3694
3695	return PGMMODE_INVALID;
3696	}
3697
3698
3699	/**
3700	* Get mode name.
3701	*
3702	* @returns read-only name string.
3703	* @param enmMode The mode which name is desired.
3704	*/
3705	VMMDECL(const char *) PGMGetModeName(PGMMODE enmMode)
3706	{
3707	switch (enmMode)
3708	{
3709	case PGMMODE_REAL: return "Real";
3710	case PGMMODE_PROTECTED: return "Protected";
3711	case PGMMODE_32_BIT: return "32-bit";
3712	case PGMMODE_PAE: return "PAE";
3713	case PGMMODE_PAE_NX: return "PAE+NX";
3714	case PGMMODE_AMD64: return "AMD64";
3715	case PGMMODE_AMD64_NX: return "AMD64+NX";
3716	case PGMMODE_NESTED_32BIT: return "Nested-32";
3717	case PGMMODE_NESTED_PAE: return "Nested-PAE";
3718	case PGMMODE_NESTED_AMD64: return "Nested-AMD64";
3719	case PGMMODE_EPT: return "EPT";
3720	case PGMMODE_NONE: return "None";
3721	default: return "unknown mode value";
3722	}
3723	}
3724
3725
3726	#ifdef VBOX_WITH_NESTED_HWVIRT_VMX_EPT
3727	/**
3728	* Gets the SLAT mode name.
3729	*
3730	* @returns The read-only SLAT mode descriptive string.
3731	* @param enmSlatMode The SLAT mode value.
3732	*/
3733	VMM_INT_DECL(const char *) PGMGetSlatModeName(PGMSLAT enmSlatMode)
3734	{
3735	switch (enmSlatMode)
3736	{
3737	case PGMSLAT_DIRECT: return "Direct";
3738	case PGMSLAT_EPT: return "EPT";
3739	case PGMSLAT_32BIT: return "32-bit";
3740	case PGMSLAT_PAE: return "PAE";
3741	case PGMSLAT_AMD64: return "AMD64";
3742	default: return "Unknown";
3743	}
3744	}
3745	#endif /* VBOX_WITH_NESTED_HWVIRT_VMX_EPT */
3746
3747
3748	/**
3749	* Gets the physical address represented in the guest CR3 as PGM sees it.
3750	*
3751	* This is mainly for logging and debugging.
3752	*
3753	* @returns PGM's guest CR3 value.
3754	* @param pVCpu The cross context virtual CPU structure.
3755	*/
3756	VMM_INT_DECL(RTGCPHYS) PGMGetGuestCR3Phys(PVMCPU pVCpu)
3757	{
3758	return pVCpu->pgm.s.GCPhysCR3;
3759	}
3760
3761
3762
3763	/**
3764	* Notification from CPUM that the EFER.NXE bit has changed.
3765	*
3766	* @param pVCpu The cross context virtual CPU structure of the CPU for
3767	* which EFER changed.
3768	* @param fNxe The new NXE state.
3769	*/
3770	VMM_INT_DECL(void) PGMNotifyNxeChanged(PVMCPU pVCpu, bool fNxe)
3771	{
3772	/** @todo VMCPU_ASSERT_EMT_OR_NOT_RUNNING(pVCpu); */
3773	Log(("PGMNotifyNxeChanged: fNxe=%RTbool\n", fNxe));
3774
3775	pVCpu->pgm.s.fNoExecuteEnabled = fNxe;
3776	if (fNxe)
3777	{
3778	/pVCpu->pgm.s.fGst32BitMbzBigPdeMask - N/A /
3779	pVCpu->pgm.s.fGstPaeMbzPteMask &= ~X86_PTE_PAE_NX;
3780	pVCpu->pgm.s.fGstPaeMbzPdeMask &= ~X86_PDE_PAE_NX;
3781	pVCpu->pgm.s.fGstPaeMbzBigPdeMask &= ~X86_PDE2M_PAE_NX;
3782	/pVCpu->pgm.s.fGstPaeMbzPdpeMask - N/A /
3783	pVCpu->pgm.s.fGstAmd64MbzPteMask &= ~X86_PTE_PAE_NX;
3784	pVCpu->pgm.s.fGstAmd64MbzPdeMask &= ~X86_PDE_PAE_NX;
3785	pVCpu->pgm.s.fGstAmd64MbzBigPdeMask &= ~X86_PDE2M_PAE_NX;
3786	pVCpu->pgm.s.fGstAmd64MbzPdpeMask &= ~X86_PDPE_LM_NX;
3787	pVCpu->pgm.s.fGstAmd64MbzBigPdpeMask &= ~X86_PDPE_LM_NX;
3788	pVCpu->pgm.s.fGstAmd64MbzPml4eMask &= ~X86_PML4E_NX;
3789
3790	pVCpu->pgm.s.fGst64ShadowedPteMask \|= X86_PTE_PAE_NX;
3791	pVCpu->pgm.s.fGst64ShadowedPdeMask \|= X86_PDE_PAE_NX;
3792	pVCpu->pgm.s.fGst64ShadowedBigPdeMask \|= X86_PDE2M_PAE_NX;
3793	pVCpu->pgm.s.fGst64ShadowedBigPde4PteMask \|= X86_PDE2M_PAE_NX;
3794	pVCpu->pgm.s.fGstAmd64ShadowedPdpeMask \|= X86_PDPE_LM_NX;
3795	pVCpu->pgm.s.fGstAmd64ShadowedPml4eMask \|= X86_PML4E_NX;
3796	}
3797	else
3798	{
3799	/pVCpu->pgm.s.fGst32BitMbzBigPdeMask - N/A /
3800	pVCpu->pgm.s.fGstPaeMbzPteMask \|= X86_PTE_PAE_NX;
3801	pVCpu->pgm.s.fGstPaeMbzPdeMask \|= X86_PDE_PAE_NX;
3802	pVCpu->pgm.s.fGstPaeMbzBigPdeMask \|= X86_PDE2M_PAE_NX;
3803	/pVCpu->pgm.s.fGstPaeMbzPdpeMask -N/A /
3804	pVCpu->pgm.s.fGstAmd64MbzPteMask \|= X86_PTE_PAE_NX;
3805	pVCpu->pgm.s.fGstAmd64MbzPdeMask \|= X86_PDE_PAE_NX;
3806	pVCpu->pgm.s.fGstAmd64MbzBigPdeMask \|= X86_PDE2M_PAE_NX;
3807	pVCpu->pgm.s.fGstAmd64MbzPdpeMask \|= X86_PDPE_LM_NX;
3808	pVCpu->pgm.s.fGstAmd64MbzBigPdpeMask \|= X86_PDPE_LM_NX;
3809	pVCpu->pgm.s.fGstAmd64MbzPml4eMask \|= X86_PML4E_NX;
3810
3811	pVCpu->pgm.s.fGst64ShadowedPteMask &= ~X86_PTE_PAE_NX;
3812	pVCpu->pgm.s.fGst64ShadowedPdeMask &= ~X86_PDE_PAE_NX;
3813	pVCpu->pgm.s.fGst64ShadowedBigPdeMask &= ~X86_PDE2M_PAE_NX;
3814	pVCpu->pgm.s.fGst64ShadowedBigPde4PteMask &= ~X86_PDE2M_PAE_NX;
3815	pVCpu->pgm.s.fGstAmd64ShadowedPdpeMask &= ~X86_PDPE_LM_NX;
3816	pVCpu->pgm.s.fGstAmd64ShadowedPml4eMask &= ~X86_PML4E_NX;
3817	}
3818	}
3819
3820
3821	/**
3822	* Check if any pgm pool pages are marked dirty (not monitored)
3823	*
3824	* @returns bool locked/not locked
3825	* @param pVM The cross context VM structure.
3826	*/
3827	VMMDECL(bool) PGMHasDirtyPages(PVM pVM)
3828	{
3829	return pVM->pgm.s.CTX_SUFF(pPool)->cDirtyPages != 0;
3830	}
3831
3832
3833	/**
3834	* Check if this VCPU currently owns the PGM lock.
3835	*
3836	* @returns bool owner/not owner
3837	* @param pVM The cross context VM structure.
3838	*/
3839	VMMDECL(bool) PGMIsLockOwner(PVMCC pVM)
3840	{
3841	return PDMCritSectIsOwner(pVM, &pVM->pgm.s.CritSectX);
3842	}
3843
3844
3845	/**
3846	* Enable or disable large page usage
3847	*
3848	* @returns VBox status code.
3849	* @param pVM The cross context VM structure.
3850	* @param fUseLargePages Use/not use large pages
3851	*/
3852	VMMDECL(int) PGMSetLargePageUsage(PVMCC pVM, bool fUseLargePages)
3853	{
3854	VM_ASSERT_VALID_EXT_RETURN(pVM, VERR_INVALID_VM_HANDLE);
3855
3856	pVM->pgm.s.fUseLargePages = fUseLargePages;
3857	return VINF_SUCCESS;
3858	}
3859
3860
3861	/**
3862	* Acquire the PGM lock.
3863	*
3864	* @returns VBox status code
3865	* @param pVM The cross context VM structure.
3866	* @param fVoid Set if the caller cannot handle failure returns.
3867	* @param SRC_POS The source position of the caller (RT_SRC_POS).
3868	*/
3869	#if defined(VBOX_STRICT) \|\| defined(DOXYGEN_RUNNING)
3870	int pgmLockDebug(PVMCC pVM, bool fVoid, RT_SRC_POS_DECL)
3871	#else
3872	int pgmLock(PVMCC pVM, bool fVoid)
3873	#endif
3874	{
3875	#if defined(VBOX_STRICT)
3876	int rc = PDMCritSectEnterDebug(pVM, &pVM->pgm.s.CritSectX, VINF_SUCCESS, (uintptr_t)ASMReturnAddress(), RT_SRC_POS_ARGS);
3877	#else
3878	int rc = PDMCritSectEnter(pVM, &pVM->pgm.s.CritSectX, VINF_SUCCESS);
3879	#endif
3880	if (RT_SUCCESS(rc))
3881	return rc;
3882	if (fVoid)
3883	PDM_CRITSECT_RELEASE_ASSERT_RC(pVM, &pVM->pgm.s.CritSectX, rc);
3884	else
3885	AssertRC(rc);
3886	return rc;
3887	}
3888
3889
3890	/**
3891	* Release the PGM lock.
3892	*
3893	* @param pVM The cross context VM structure.
3894	*/
3895	void pgmUnlock(PVMCC pVM)
3896	{
3897	uint32_t cDeprecatedPageLocks = pVM->pgm.s.cDeprecatedPageLocks;
3898	pVM->pgm.s.cDeprecatedPageLocks = 0;
3899	int rc = PDMCritSectLeave(pVM, &pVM->pgm.s.CritSectX);
3900	if (rc == VINF_SEM_NESTED)
3901	pVM->pgm.s.cDeprecatedPageLocks = cDeprecatedPageLocks;
3902	}
3903
3904
3905	#if !defined(IN_R0) \|\| defined(LOG_ENABLED)
3906
3907	/** Format handler for PGMPAGE.
3908	* @copydoc FNRTSTRFORMATTYPE */
3909	static DECLCALLBACK(size_t) pgmFormatTypeHandlerPage(PFNRTSTROUTPUT pfnOutput, void *pvArgOutput,
3910	const char pszType, void const pvValue,
3911	int cchWidth, int cchPrecision, unsigned fFlags,
3912	void *pvUser)
3913	{
3914	size_t cch;
3915	PCPGMPAGE pPage = (PCPGMPAGE)pvValue;
3916	if (RT_VALID_PTR(pPage))
3917	{
3918	char szTmp[64+80];
3919
3920	cch = 0;
3921
3922	/* The single char state stuff. */
3923	static const char s_achPageStates[4] = { 'Z', 'A', 'W', 'S' };
3924	szTmp[cch++] = s_achPageStates[PGM_PAGE_GET_STATE_NA(pPage)];
3925
3926	# define IS_PART_INCLUDED(lvl) ( !(fFlags & RTSTR_F_PRECISION) \|\| cchPrecision == (lvl) \|\| cchPrecision >= (lvl)+10 )
3927	if (IS_PART_INCLUDED(5))
3928	{
3929	static const char s_achHandlerStates[4*2] = { '-', 't', 'w', 'a' , '_', 'T', 'W', 'A' };
3930	szTmp[cch++] = s_achHandlerStates[ PGM_PAGE_GET_HNDL_PHYS_STATE(pPage)
3931	\| ((uint8_t)PGM_PAGE_IS_HNDL_PHYS_NOT_IN_HM(pPage) << 2)];
3932	}
3933
3934	/* The type. */
3935	if (IS_PART_INCLUDED(4))
3936	{
3937	szTmp[cch++] = ':';
3938	static const char s_achPageTypes[8][4] = { "INV", "RAM", "MI2", "M2A", "SHA", "ROM", "MIO", "BAD" };
3939	szTmp[cch++] = s_achPageTypes[PGM_PAGE_GET_TYPE_NA(pPage)][0];
3940	szTmp[cch++] = s_achPageTypes[PGM_PAGE_GET_TYPE_NA(pPage)][1];
3941	szTmp[cch++] = s_achPageTypes[PGM_PAGE_GET_TYPE_NA(pPage)][2];
3942	}
3943
3944	/* The numbers. */
3945	if (IS_PART_INCLUDED(3))
3946	{
3947	szTmp[cch++] = ':';
3948	cch += RTStrFormatNumber(&szTmp[cch], PGM_PAGE_GET_HCPHYS_NA(pPage), 16, 12, 0, RTSTR_F_ZEROPAD \| RTSTR_F_64BIT);
3949	}
3950
3951	if (IS_PART_INCLUDED(2))
3952	{
3953	szTmp[cch++] = ':';
3954	cch += RTStrFormatNumber(&szTmp[cch], PGM_PAGE_GET_PAGEID(pPage), 16, 7, 0, RTSTR_F_ZEROPAD \| RTSTR_F_32BIT);
3955	}
3956
3957	if (IS_PART_INCLUDED(6))
3958	{
3959	szTmp[cch++] = ':';
3960	static const char s_achRefs[4] = { '-', 'U', '!', 'L' };
3961	szTmp[cch++] = s_achRefs[PGM_PAGE_GET_TD_CREFS_NA(pPage)];
3962	cch += RTStrFormatNumber(&szTmp[cch], PGM_PAGE_GET_TD_IDX_NA(pPage), 16, 4, 0, RTSTR_F_ZEROPAD \| RTSTR_F_16BIT);
3963	}
3964	# undef IS_PART_INCLUDED
3965
3966	cch = pfnOutput(pvArgOutput, szTmp, cch);
3967	#if 0
3968	size_t cch2 = 0;
3969	szTmp[cch2++] = '(';
3970	cch2 += RTStrFormatNumber(&szTmp[cch2], (uintptr_t)pPage, 16, 18, 0, RTSTR_F_SPECIAL \| RTSTR_F_ZEROPAD \| RTSTR_F_64BIT);
3971	szTmp[cch2++] = ')';
3972	szTmp[cch2] = '\0';
3973	cch += pfnOutput(pvArgOutput, szTmp, cch2);
3974	#endif
3975	}
3976	else
3977	cch = pfnOutput(pvArgOutput, RT_STR_TUPLE("<bad-pgmpage-ptr>"));
3978	NOREF(pszType); NOREF(cchWidth); NOREF(pvUser);
3979	return cch;
3980	}
3981
3982
3983	/** Format handler for PGMRAMRANGE.
3984	* @copydoc FNRTSTRFORMATTYPE */
3985	static DECLCALLBACK(size_t) pgmFormatTypeHandlerRamRange(PFNRTSTROUTPUT pfnOutput, void *pvArgOutput,
3986	const char pszType, void const pvValue,
3987	int cchWidth, int cchPrecision, unsigned fFlags,
3988	void *pvUser)
3989	{
3990	size_t cch;
3991	PGMRAMRANGE const pRam = (PGMRAMRANGE const )pvValue;
3992	if (RT_VALID_PTR(pRam))
3993	{
3994	char szTmp[80];
3995	cch = RTStrPrintf(szTmp, sizeof(szTmp), "%RGp-%RGp", pRam->GCPhys, pRam->GCPhysLast);
3996	cch = pfnOutput(pvArgOutput, szTmp, cch);
3997	}
3998	else
3999	cch = pfnOutput(pvArgOutput, RT_STR_TUPLE("<bad-pgmramrange-ptr>"));
4000	NOREF(pszType); NOREF(cchWidth); NOREF(cchPrecision); NOREF(pvUser); NOREF(fFlags);
4001	return cch;
4002	}
4003
4004	/** Format type andlers to be registered/deregistered. */
4005	static const struct
4006	{
4007	char szType[24];
4008	PFNRTSTRFORMATTYPE pfnHandler;
4009	} g_aPgmFormatTypes[] =
4010	{
4011	{ "pgmpage", pgmFormatTypeHandlerPage },
4012	{ "pgmramrange", pgmFormatTypeHandlerRamRange }
4013	};
4014
4015	#endif /* !IN_R0 \|\| LOG_ENABLED */
4016
4017	/**
4018	* Registers the global string format types.
4019	*
4020	* This should be called at module load time or in some other manner that ensure
4021	* that it's called exactly one time.
4022	*
4023	* @returns IPRT status code on RTStrFormatTypeRegister failure.
4024	*/
4025	VMMDECL(int) PGMRegisterStringFormatTypes(void)
4026	{
4027	#if !defined(IN_R0) \|\| defined(LOG_ENABLED)
4028	int rc = VINF_SUCCESS;
4029	unsigned i;
4030	for (i = 0; RT_SUCCESS(rc) && i < RT_ELEMENTS(g_aPgmFormatTypes); i++)
4031	{
4032	rc = RTStrFormatTypeRegister(g_aPgmFormatTypes[i].szType, g_aPgmFormatTypes[i].pfnHandler, NULL);
4033	# ifdef IN_RING0
4034	if (rc == VERR_ALREADY_EXISTS)
4035	{
4036	/* in case of cleanup failure in ring-0 */
4037	RTStrFormatTypeDeregister(g_aPgmFormatTypes[i].szType);
4038	rc = RTStrFormatTypeRegister(g_aPgmFormatTypes[i].szType, g_aPgmFormatTypes[i].pfnHandler, NULL);
4039	}
4040	# endif
4041	}
4042	if (RT_FAILURE(rc))
4043	while (i-- > 0)
4044	RTStrFormatTypeDeregister(g_aPgmFormatTypes[i].szType);
4045
4046	return rc;
4047	#else
4048	return VINF_SUCCESS;
4049	#endif
4050	}
4051
4052
4053	/**
4054	* Deregisters the global string format types.
4055	*
4056	* This should be called at module unload time or in some other manner that
4057	* ensure that it's called exactly one time.
4058	*/
4059	VMMDECL(void) PGMDeregisterStringFormatTypes(void)
4060	{
4061	#if !defined(IN_R0) \|\| defined(LOG_ENABLED)
4062	for (unsigned i = 0; i < RT_ELEMENTS(g_aPgmFormatTypes); i++)
4063	RTStrFormatTypeDeregister(g_aPgmFormatTypes[i].szType);
4064	#endif
4065	}
4066
4067
4068	#ifdef VBOX_STRICT
4069	/**
4070	* Asserts that everything related to the guest CR3 is correctly shadowed.
4071	*
4072	* This will call PGMAssertNoMappingConflicts() and PGMAssertHandlerAndFlagsInSync(),
4073	* and assert the correctness of the guest CR3 mapping before asserting that the
4074	* shadow page tables is in sync with the guest page tables.
4075	*
4076	* @returns Number of conflicts.
4077	* @param pVM The cross context VM structure.
4078	* @param pVCpu The cross context virtual CPU structure.
4079	* @param cr3 The current guest CR3 register value.
4080	* @param cr4 The current guest CR4 register value.
4081	*/
4082	VMMDECL(unsigned) PGMAssertCR3(PVMCC pVM, PVMCPUCC pVCpu, uint64_t cr3, uint64_t cr4)
4083	{
4084	STAM_PROFILE_START(&pVCpu->pgm.s.Stats.CTX_MID_Z(Stat,SyncCR3), a);
4085
4086	uintptr_t const idxBth = pVCpu->pgm.s.idxBothModeData;
4087	AssertReturn(idxBth < RT_ELEMENTS(g_aPgmBothModeData), -VERR_PGM_MODE_IPE);
4088	AssertReturn(g_aPgmBothModeData[idxBth].pfnAssertCR3, -VERR_PGM_MODE_IPE);
4089
4090	PGM_LOCK_VOID(pVM);
4091	unsigned cErrors = g_aPgmBothModeData[idxBth].pfnAssertCR3(pVCpu, cr3, cr4, 0, ~(RTGCPTR)0);
4092	PGM_UNLOCK(pVM);
4093
4094	STAM_PROFILE_STOP(&pVCpu->pgm.s.Stats.CTX_MID_Z(Stat,SyncCR3), a);
4095	return cErrors;
4096	}
4097	#endif /* VBOX_STRICT */
4098
4099
4100	/**
4101	* Updates PGM's copy of the guest's EPT pointer.
4102	*
4103	* @param pVCpu The cross context virtual CPU structure.
4104	* @param uEptPtr The EPT pointer.
4105	*
4106	* @remarks This can be called as part of VM-entry so we might be in the midst of
4107	* switching to VMX non-root mode.
4108	*/
4109	VMM_INT_DECL(void) PGMSetGuestEptPtr(PVMCPUCC pVCpu, uint64_t uEptPtr)
4110	{
4111	PVMCC pVM = pVCpu->CTX_SUFF(pVM);
4112	PGM_LOCK_VOID(pVM);
4113	pVCpu->pgm.s.uEptPtr = uEptPtr;
4114	pVCpu->pgm.s.pGstEptPml4R3 = 0;
4115	pVCpu->pgm.s.pGstEptPml4R0 = 0;
4116	PGM_UNLOCK(pVM);
4117	}
4118
4119	#ifdef PGM_WITH_PAGE_ZEROING_DETECTION
4120
4121	/**
4122	* Helper for checking whether XMM0 is zero, possibly retriving external state.
4123	*/
4124	static bool pgmHandlePageZeroingIsXmm0Zero(PVMCPUCC pVCpu, PCPUMCTX pCtx)
4125	{
4126	if (pCtx->fExtrn & CPUMCTX_EXTRN_SSE_AVX)
4127	{
4128	int rc = CPUMImportGuestStateOnDemand(pVCpu, CPUMCTX_EXTRN_SSE_AVX);
4129	AssertRCReturn(rc, false);
4130	}
4131	return pCtx->XState.x87.aXMM[0].au64[0] == 0
4132	&& pCtx->XState.x87.aXMM[0].au64[1] == 0
4133	&& pCtx->XState.x87.aXMM[0].au64[2] == 0
4134	&& pCtx->XState.x87.aXMM[0].au64[3] == 0;
4135	}
4136
4137
4138	/**
4139	* Helper for comparing opcode bytes.
4140	*/
4141	static bool pgmHandlePageZeroingMatchOpcodes(PVMCPUCC pVCpu, PCPUMCTX pCtx, uint8_t const *pbOpcodes, uint32_t cbOpcodes)
4142	{
4143	uint8_t abTmp[64];
4144	AssertMsgReturn(cbOpcodes <= sizeof(abTmp), ("cbOpcodes=%#x\n", cbOpcodes), false);
4145	int rc = PGMPhysSimpleReadGCPtr(pVCpu, abTmp, pCtx->rip + pCtx->cs.u64Base, cbOpcodes);
4146	if (RT_SUCCESS(rc))
4147	return memcmp(abTmp, pbOpcodes, cbOpcodes) == 0;
4148	return false;
4149	}
4150
4151
4152	/**
4153	* Called on faults on ZERO pages to check if the guest is trying to zero it.
4154	*
4155	* Since it's a waste of time to zero a ZERO page and it will cause an
4156	* unnecessary page allocation, we'd like to detect and avoid this.
4157	* If any known page zeroing code is detected, this function will update the CPU
4158	* state to pretend the page was zeroed by the code.
4159	*
4160	* @returns true if page zeroing code was detected and CPU state updated to skip
4161	* the code.
4162	* @param pVCpu The cross context virtual CPU structure of the calling EMT.
4163	* @param pCtx The guest register context.
4164	*/
4165	static bool pgmHandlePageZeroingCode(PVMCPUCC pVCpu, PCPUMCTX pCtx)
4166	{
4167	CPUMCTX_ASSERT_NOT_EXTRN(pCtx, CPUMCTX_EXTRN_CR0 \| CPUMCTX_EXTRN_CR3 \| CPUMCTX_EXTRN_CR4 \| CPUMCTX_EXTRN_EFER);
4168
4169	/*
4170	* Sort by mode first.
4171	*/
4172	if (CPUMIsGuestInLongModeEx(pCtx))
4173	{
4174	if (CPUMIsGuestIn64BitCodeEx(pCtx))
4175	{
4176	/*
4177	* 64-bit code.
4178	*/
4179	Log9(("pgmHandlePageZeroingCode: not page zeroing - 64-bit\n"));
4180	}
4181	else if (pCtx->cs.Attr.n.u1DefBig)
4182	Log9(("pgmHandlePageZeroingCode: not page zeroing - 32-bit lm\n"));
4183	else
4184	Log9(("pgmHandlePageZeroingCode: not page zeroing - 16-bit lm\n"));
4185	}
4186	else if (CPUMIsGuestInPagedProtectedModeEx(pCtx))
4187	{
4188	if (pCtx->cs.Attr.n.u1DefBig)
4189	{
4190	/*
4191	* 32-bit paged protected mode code.
4192	*/
4193	CPUMCTX_ASSERT_NOT_EXTRN(pCtx, CPUMCTX_EXTRN_RAX \| CPUMCTX_EXTRN_RCX \| CPUMCTX_EXTRN_RDX \| CPUMCTX_EXTRN_RBX
4194	\| CPUMCTX_EXTRN_RSP \| CPUMCTX_EXTRN_RBP \| CPUMCTX_EXTRN_RSI \| CPUMCTX_EXTRN_RDI
4195	\| CPUMCTX_EXTRN_RIP \| CPUMCTX_EXTRN_RFLAGS);
4196
4197	/* 1. Generic 'rep stosd' detection. */
4198	static uint8_t const s_abRepStosD[] = { 0xf3, 0xab };
4199	if ( pCtx->eax == 0
4200	&& pCtx->ecx == X86_PAGE_SIZE / 4
4201	&& !(pCtx->edi & X86_PAGE_OFFSET_MASK)
4202	&& pgmHandlePageZeroingMatchOpcodes(pVCpu, pCtx, s_abRepStosD, sizeof(s_abRepStosD)))
4203	{
4204	pCtx->ecx = 0;
4205	pCtx->edi += X86_PAGE_SIZE;
4206	Log9(("pgmHandlePageZeroingCode: REP STOSD: eip=%RX32 -> %RX32\n", pCtx->eip, pCtx->eip + sizeof(s_abRepStosD)));
4207	pCtx->eip += sizeof(s_abRepStosD);
4208	return true;
4209	}
4210
4211	/* 2. Windows 2000 sp4 KiXMMIZeroPageNoSave loop code: */
4212	static uint8_t const s_abW2kSp4XmmZero[] =
4213	{
4214	0x0f, 0x2b, 0x01,
4215	0x0f, 0x2b, 0x41, 0x10,
4216	0x0f, 0x2b, 0x41, 0x20,
4217	0x0f, 0x2b, 0x41, 0x30,
4218	0x83, 0xc1, 0x40,
4219	0x48,
4220	0x75, 0xeb,
4221	};
4222	if ( pCtx->eax == 64
4223	&& !(pCtx->ecx & X86_PAGE_OFFSET_MASK)
4224	&& pgmHandlePageZeroingMatchOpcodes(pVCpu, pCtx, s_abW2kSp4XmmZero, sizeof(s_abW2kSp4XmmZero))
4225	&& pgmHandlePageZeroingIsXmm0Zero(pVCpu, pCtx))
4226	{
4227	pCtx->eax = 1;
4228	pCtx->ecx += X86_PAGE_SIZE;
4229	Log9(("pgmHandlePageZeroingCode: w2k sp4 xmm: eip=%RX32 -> %RX32\n",
4230	pCtx->eip, pCtx->eip + sizeof(s_abW2kSp4XmmZero) - 3));
4231	pCtx->eip += sizeof(s_abW2kSp4XmmZero) - 3;
4232	return true;
4233	}
4234	Log9(("pgmHandlePageZeroingCode: not page zeroing - 32-bit\n"));
4235	}
4236	else if (!pCtx->eflags.Bits.u1VM)
4237	Log9(("pgmHandlePageZeroingCode: not page zeroing - 16-bit\n"));
4238	else
4239	Log9(("pgmHandlePageZeroingCode: not page zeroing - v86\n"));
4240	}
4241	return false;
4242	}
4243
4244	#endif /* PGM_WITH_PAGE_ZEROING_DETECTION */

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

source: vbox/trunk/src/VBox/VMM/VMMAll/PGMAll.cpp@ 105724

Download in other formats: