PGMAllBth.h@ 17142

Last change on this file since 17142 was 17142, checked in by vboxsync, 16 years ago
VBOX_WITH_PGMPOOL_PAGING_ONLY: fixed saved state restore
Property svn:eol-style set to `native` Property svn:keywords set to `Id`
File size: 208.0 KB

Line
1	/* $Id: PGMAllBth.h 17142 2009-02-25 17:10:23Z vboxsync $ */
2	/** @file
3	* VBox - Page Manager, Shadow+Guest Paging Template - All context code.
4	*
5	* This file is a big challenge!
6	*/
7
8	/*
9	* Copyright (C) 2006-2007 Sun Microsystems, Inc.
10	*
11	* This file is part of VirtualBox Open Source Edition (OSE), as
12	* available from http://www.virtualbox.org. This file is free software;
13	* you can redistribute it and/or modify it under the terms of the GNU
14	* General Public License (GPL) as published by the Free Software
15	* Foundation, in version 2 as it comes in the "COPYING" file of the
16	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
17	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
18	*
19	* Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa
20	* Clara, CA 95054 USA or visit http://www.sun.com if you need
21	* additional information or have any questions.
22	*/
23
24	/*******************************************************************************
25	* Internal Functions *
26	*******************************************************************************/
27	__BEGIN_DECLS
28	PGM_BTH_DECL(int, Trap0eHandler)(PVM pVM, RTGCUINT uErr, PCPUMCTXCORE pRegFrame, RTGCPTR pvFault);
29	PGM_BTH_DECL(int, InvalidatePage)(PVM pVM, RTGCPTR GCPtrPage);
30	PGM_BTH_DECL(int, SyncPage)(PVM pVM, GSTPDE PdeSrc, RTGCPTR GCPtrPage, unsigned cPages, unsigned uErr);
31	PGM_BTH_DECL(int, CheckPageFault)(PVM pVM, uint32_t uErr, PSHWPDE pPdeDst, PGSTPDE pPdeSrc, RTGCPTR GCPtrPage);
32	PGM_BTH_DECL(int, SyncPT)(PVM pVM, unsigned iPD, PGSTPD pPDSrc, RTGCPTR GCPtrPage);
33	PGM_BTH_DECL(int, VerifyAccessSyncPage)(PVM pVM, RTGCPTR Addr, unsigned fPage, unsigned uErr);
34	PGM_BTH_DECL(int, PrefetchPage)(PVM pVM, RTGCPTR GCPtrPage);
35	PGM_BTH_DECL(int, SyncCR3)(PVM pVM, uint64_t cr0, uint64_t cr3, uint64_t cr4, bool fGlobal);
36	#ifdef VBOX_STRICT
37	PGM_BTH_DECL(unsigned, AssertCR3)(PVM pVM, uint64_t cr3, uint64_t cr4, RTGCPTR GCPtr = 0, RTGCPTR cb = ~(RTGCPTR)0);
38	#endif
39	#ifdef PGMPOOL_WITH_USER_TRACKING
40	DECLINLINE(void) PGM_BTH_NAME(SyncPageWorkerTrackDeref)(PVM pVM, PPGMPOOLPAGE pShwPage, RTHCPHYS HCPhys);
41	#endif
42	PGM_BTH_DECL(int, MapCR3)(PVM pVM, RTGCPHYS GCPhysCR3);
43	PGM_BTH_DECL(int, UnmapCR3)(PVM pVM);
44	__END_DECLS
45
46
47	/* Filter out some illegal combinations of guest and shadow paging, so we can remove redundant checks inside functions. */
48	#if PGM_GST_TYPE == PGM_TYPE_PAE && PGM_SHW_TYPE != PGM_TYPE_PAE && PGM_SHW_TYPE != PGM_TYPE_NESTED && PGM_SHW_TYPE != PGM_TYPE_EPT
49	# error "Invalid combination; PAE guest implies PAE shadow"
50	#endif
51
52	#if (PGM_GST_TYPE == PGM_TYPE_REAL \|\| PGM_GST_TYPE == PGM_TYPE_PROT) \
53	&& !(PGM_SHW_TYPE == PGM_TYPE_32BIT \|\| PGM_SHW_TYPE == PGM_TYPE_PAE \|\| PGM_SHW_TYPE == PGM_TYPE_AMD64 \|\| PGM_SHW_TYPE == PGM_TYPE_NESTED \|\| PGM_SHW_TYPE == PGM_TYPE_EPT)
54	# error "Invalid combination; real or protected mode without paging implies 32 bits or PAE shadow paging."
55	#endif
56
57	#if (PGM_GST_TYPE == PGM_TYPE_32BIT \|\| PGM_GST_TYPE == PGM_TYPE_PAE) \
58	&& !(PGM_SHW_TYPE == PGM_TYPE_32BIT \|\| PGM_SHW_TYPE == PGM_TYPE_PAE \|\| PGM_SHW_TYPE == PGM_TYPE_NESTED \|\| PGM_SHW_TYPE == PGM_TYPE_EPT)
59	# error "Invalid combination; 32 bits guest paging or PAE implies 32 bits or PAE shadow paging."
60	#endif
61
62	#if (PGM_GST_TYPE == PGM_TYPE_AMD64 && PGM_SHW_TYPE != PGM_TYPE_AMD64 && PGM_SHW_TYPE != PGM_TYPE_NESTED && PGM_SHW_TYPE != PGM_TYPE_EPT) \
63	\|\| (PGM_SHW_TYPE == PGM_TYPE_AMD64 && PGM_GST_TYPE != PGM_TYPE_AMD64 && PGM_GST_TYPE != PGM_TYPE_PROT)
64	# error "Invalid combination; AMD64 guest implies AMD64 shadow and vice versa"
65	#endif
66
67	#ifdef IN_RING0 /* no mappings in VT-x and AMD-V mode */
68	# define PGM_WITHOUT_MAPPINGS
69	#endif
70
71
72	#ifndef IN_RING3
73	/**
74	* #PF Handler for raw-mode guest execution.
75	*
76	* @returns VBox status code (appropriate for trap handling and GC return).
77	* @param pVM VM Handle.
78	* @param uErr The trap error code.
79	* @param pRegFrame Trap register frame.
80	* @param pvFault The fault address.
81	*/
82	PGM_BTH_DECL(int, Trap0eHandler)(PVM pVM, RTGCUINT uErr, PCPUMCTXCORE pRegFrame, RTGCPTR pvFault)
83	{
84	# if defined(IN_RC) && defined(VBOX_WITH_PGMPOOL_PAGING_ONLY)
85	PGMDynCheckLocks(pVM);
86	# endif
87
88	# if (PGM_GST_TYPE == PGM_TYPE_32BIT \|\| PGM_GST_TYPE == PGM_TYPE_REAL \|\| PGM_GST_TYPE == PGM_TYPE_PROT \|\| PGM_GST_TYPE == PGM_TYPE_PAE \|\| PGM_GST_TYPE == PGM_TYPE_AMD64) \
89	&& PGM_SHW_TYPE != PGM_TYPE_NESTED \
90	&& (PGM_SHW_TYPE != PGM_TYPE_EPT \|\| PGM_GST_TYPE == PGM_TYPE_PROT)
91
92	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE != PGM_TYPE_PAE
93	/*
94	* Hide the instruction fetch trap indicator for now.
95	*/
96	/** @todo NXE will change this and we must fix NXE in the switcher too! */
97	if (uErr & X86_TRAP_PF_ID)
98	{
99	uErr &= ~X86_TRAP_PF_ID;
100	TRPMSetErrorCode(pVM, uErr);
101	}
102	# endif
103
104	/*
105	* Get PDs.
106	*/
107	int rc;
108	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
109	# if PGM_GST_TYPE == PGM_TYPE_32BIT
110	const unsigned iPDSrc = pvFault >> GST_PD_SHIFT;
111	PGSTPD pPDSrc = pgmGstGet32bitPDPtr(&pVM->pgm.s);
112
113	# elif PGM_GST_TYPE == PGM_TYPE_PAE \|\| PGM_GST_TYPE == PGM_TYPE_AMD64
114
115	# if PGM_GST_TYPE == PGM_TYPE_PAE
116	unsigned iPDSrc;
117	# ifdef VBOX_WITH_PGMPOOL_PAGING_ONLY
118	X86PDPE PdpeSrc;
119	PGSTPD pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, pvFault, &iPDSrc, &PdpeSrc);
120	# else
121	PGSTPD pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, pvFault, &iPDSrc, NULL);
122	# endif
123
124	# elif PGM_GST_TYPE == PGM_TYPE_AMD64
125	unsigned iPDSrc;
126	PX86PML4E pPml4eSrc;
127	X86PDPE PdpeSrc;
128	PGSTPD pPDSrc;
129
130	pPDSrc = pgmGstGetLongModePDPtr(&pVM->pgm.s, pvFault, &pPml4eSrc, &PdpeSrc, &iPDSrc);
131	Assert(pPml4eSrc);
132	# endif
133
134	/* Quick check for a valid guest trap. (PAE & AMD64) */
135	if (!pPDSrc)
136	{
137	# if PGM_GST_TYPE == PGM_TYPE_AMD64 && GC_ARCH_BITS == 64
138	LogFlow(("Trap0eHandler: guest PML4 %d not present CR3=%RGp\n", (int)((pvFault >> X86_PML4_SHIFT) & X86_PML4_MASK), CPUMGetGuestCR3(pVM) & X86_CR3_PAGE_MASK));
139	# else
140	LogFlow(("Trap0eHandler: guest iPDSrc=%u not present CR3=%RGp\n", iPDSrc, CPUMGetGuestCR3(pVM) & X86_CR3_PAGE_MASK));
141	# endif
142	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2GuestTrap; });
143	TRPMSetErrorCode(pVM, uErr);
144	return VINF_EM_RAW_GUEST_TRAP;
145	}
146	# endif
147
148	# else /* !PGM_WITH_PAGING */
149	PGSTPD pPDSrc = NULL;
150	const unsigned iPDSrc = 0;
151	# endif /* !PGM_WITH_PAGING */
152
153
154	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
155	const unsigned iPDDst = pvFault >> SHW_PD_SHIFT;
156	PX86PD pPDDst = pgmShwGet32BitPDPtr(&pVM->pgm.s);
157
158	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
159	const unsigned iPDDst = (pvFault >> SHW_PD_SHIFT) & SHW_PD_MASK; /* pPDDst index, not used with the pool. */
160
161	# ifdef VBOX_WITH_PGMPOOL_PAGING_ONLY
162	PX86PDPAE pPDDst;
163	# if PGM_GST_TYPE != PGM_TYPE_PAE
164	X86PDPE PdpeSrc;
165
166	/* Fake PDPT entry; access control handled on the page table level, so allow everything. */
167	PdpeSrc.u = X86_PDPE_P; /* rw/us are reserved for PAE pdpte's; accessed bit causes invalid VT-x guest state errors */
168	# endif
169	rc = pgmShwSyncPaePDPtr(pVM, pvFault, &PdpeSrc, &pPDDst);
170	if (rc != VINF_SUCCESS)
171	{
172	AssertRC(rc);
173	return rc;
174	}
175	Assert(pPDDst);
176
177	# else
178	PX86PDPAE pPDDst = pgmShwGetPaePDPtr(&pVM->pgm.s, pvFault);
179
180	/* Did we mark the PDPT as not present in SyncCR3? */
181	unsigned iPdpt = (pvFault >> SHW_PDPT_SHIFT) & SHW_PDPT_MASK;
182	PX86PDPT pPdptDst = pgmShwGetPaePDPTPtr(&pVM->pgm.s);
183	if (!pPdptDst->a[iPdpt].n.u1Present)
184	pPdptDst->a[iPdpt].n.u1Present = 1;
185	# endif
186
187	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
188	const unsigned iPDDst = ((pvFault >> SHW_PD_SHIFT) & SHW_PD_MASK);
189	PX86PDPAE pPDDst;
190	# if PGM_GST_TYPE == PGM_TYPE_PROT
191	/* AMD-V nested paging */
192	X86PML4E Pml4eSrc;
193	X86PDPE PdpeSrc;
194	PX86PML4E pPml4eSrc = &Pml4eSrc;
195
196	/* Fake PML4 & PDPT entry; access control handled on the page table level, so allow everything. */
197	Pml4eSrc.u = X86_PML4E_P \| X86_PML4E_RW \| X86_PML4E_US \| X86_PML4E_A;
198	PdpeSrc.u = X86_PDPE_P \| X86_PDPE_RW \| X86_PDPE_US \| X86_PDPE_A;
199	# endif
200
201	rc = pgmShwSyncLongModePDPtr(pVM, pvFault, pPml4eSrc, &PdpeSrc, &pPDDst);
202	if (rc != VINF_SUCCESS)
203	{
204	AssertRC(rc);
205	return rc;
206	}
207	Assert(pPDDst);
208
209	# elif PGM_SHW_TYPE == PGM_TYPE_EPT
210	const unsigned iPDDst = ((pvFault >> SHW_PD_SHIFT) & SHW_PD_MASK);
211	PEPTPD pPDDst;
212
213	rc = pgmShwGetEPTPDPtr(pVM, pvFault, NULL, &pPDDst);
214	if (rc != VINF_SUCCESS)
215	{
216	AssertRC(rc);
217	return rc;
218	}
219	Assert(pPDDst);
220	# endif
221
222	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
223	/*
224	* If we successfully correct the write protection fault due to dirty bit
225	* tracking, or this page fault is a genuine one, then return immediately.
226	*/
227	STAM_PROFILE_START(&pVM->pgm.s.StatRZTrap0eTimeCheckPageFault, e);
228	rc = PGM_BTH_NAME(CheckPageFault)(pVM, uErr, &pPDDst->a[iPDDst], &pPDSrc->a[iPDSrc], pvFault);
229	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeCheckPageFault, e);
230	if ( rc == VINF_PGM_HANDLED_DIRTY_BIT_FAULT
231	\|\| rc == VINF_EM_RAW_GUEST_TRAP)
232	{
233	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution)
234	= rc == VINF_PGM_HANDLED_DIRTY_BIT_FAULT ? &pVM->pgm.s.StatRZTrap0eTime2DirtyAndAccessed : &pVM->pgm.s.StatRZTrap0eTime2GuestTrap; });
235	LogBird(("Trap0eHandler: returns %s\n", rc == VINF_PGM_HANDLED_DIRTY_BIT_FAULT ? "VINF_SUCCESS" : "VINF_EM_RAW_GUEST_TRAP"));
236	return rc == VINF_PGM_HANDLED_DIRTY_BIT_FAULT ? VINF_SUCCESS : rc;
237	}
238
239	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0ePD[iPDSrc]);
240	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
241
242	/*
243	* A common case is the not-present error caused by lazy page table syncing.
244	*
245	* It is IMPORTANT that we weed out any access to non-present shadow PDEs here
246	* so we can safely assume that the shadow PT is present when calling SyncPage later.
247	*
248	* On failure, we ASSUME that SyncPT is out of memory or detected some kind
249	* of mapping conflict and defer to SyncCR3 in R3.
250	* (Again, we do NOT support access handlers for non-present guest pages.)
251	*
252	*/
253	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
254	GSTPDE PdeSrc = pPDSrc->a[iPDSrc];
255	# else
256	GSTPDE PdeSrc;
257	PdeSrc.au32[0] = 0; /* faked so we don't have to #ifdef everything */
258	PdeSrc.n.u1Present = 1;
259	PdeSrc.n.u1Write = 1;
260	PdeSrc.n.u1Accessed = 1;
261	PdeSrc.n.u1User = 1;
262	# endif
263	if ( !(uErr & X86_TRAP_PF_P) /* not set means page not present instead of page protection violation */
264	&& !pPDDst->a[iPDDst].n.u1Present
265	&& PdeSrc.n.u1Present
266	)
267
268	{
269	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2SyncPT; });
270	STAM_PROFILE_START(&pVM->pgm.s.StatRZTrap0eTimeSyncPT, f);
271	LogFlow(("=>SyncPT %04x = %08x\n", iPDSrc, PdeSrc.au32[0]));
272	rc = PGM_BTH_NAME(SyncPT)(pVM, iPDSrc, pPDSrc, pvFault);
273	if (RT_SUCCESS(rc))
274	{
275	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeSyncPT, f);
276	return rc;
277	}
278	Log(("SyncPT: %d failed!! rc=%d\n", iPDSrc, rc));
279	VM_FF_SET(pVM, VM_FF_PGM_SYNC_CR3); /** @todo no need to do global sync, right? */
280	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeSyncPT, f);
281	return VINF_PGM_SYNC_CR3;
282	}
283
284	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
285	/*
286	* Check if this address is within any of our mappings.
287	*
288	* This is very fast and it's gonna save us a bit of effort below and prevent
289	* us from screwing ourself with MMIO2 pages which have a GC Mapping (VRam).
290	* (BTW, it's impossible to have physical access handlers in a mapping.)
291	*/
292	if (pgmMapAreMappingsEnabled(&pVM->pgm.s))
293	{
294	STAM_PROFILE_START(&pVM->pgm.s.StatRZTrap0eTimeMapping, a);
295	PPGMMAPPING pMapping = pVM->pgm.s.CTX_SUFF(pMappings);
296	for ( ; pMapping; pMapping = pMapping->CTX_SUFF(pNext))
297	{
298	if (pvFault < pMapping->GCPtr)
299	break;
300	if (pvFault - pMapping->GCPtr < pMapping->cb)
301	{
302	/*
303	* The first thing we check is if we've got an undetected conflict.
304	*/
305	if (!pVM->pgm.s.fMappingsFixed)
306	{
307	unsigned iPT = pMapping->cb >> GST_PD_SHIFT;
308	while (iPT-- > 0)
309	if (pPDSrc->a[iPDSrc + iPT].n.u1Present)
310	{
311	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eConflicts);
312	Log(("Trap0e: Detected Conflict %RGv-%RGv\n", pMapping->GCPtr, pMapping->GCPtrLast));
313	VM_FF_SET(pVM, VM_FF_PGM_SYNC_CR3); /** @todo no need to do global sync,right? */
314	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeMapping, a);
315	return VINF_PGM_SYNC_CR3;
316	}
317	}
318
319	/*
320	* Check if the fault address is in a virtual page access handler range.
321	*/
322	PPGMVIRTHANDLER pCur = (PPGMVIRTHANDLER)RTAvlroGCPtrRangeGet(&pVM->pgm.s.CTX_SUFF(pTrees)->HyperVirtHandlers, pvFault);
323	if ( pCur
324	&& pvFault - pCur->Core.Key < pCur->cb
325	&& uErr & X86_TRAP_PF_RW)
326	{
327	# ifdef IN_RC
328	STAM_PROFILE_START(&pCur->Stat, h);
329	rc = pCur->CTX_SUFF(pfnHandler)(pVM, uErr, pRegFrame, pvFault, pCur->Core.Key, pvFault - pCur->Core.Key);
330	STAM_PROFILE_STOP(&pCur->Stat, h);
331	# else
332	AssertFailed();
333	rc = VINF_EM_RAW_EMULATE_INSTR; /* can't happen with VMX */
334	# endif
335	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersMapping);
336	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeMapping, a);
337	return rc;
338	}
339
340	/*
341	* Pretend we're not here and let the guest handle the trap.
342	*/
343	TRPMSetErrorCode(pVM, uErr & ~X86_TRAP_PF_P);
344	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eGuestPFMapping);
345	LogFlow(("PGM: Mapping access -> route trap to recompiler!\n"));
346	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeMapping, a);
347	return VINF_EM_RAW_GUEST_TRAP;
348	}
349	}
350	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeMapping, a);
351	} /* pgmAreMappingsEnabled(&pVM->pgm.s) */
352	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
353
354	/*
355	* Check if this fault address is flagged for special treatment,
356	* which means we'll have to figure out the physical address and
357	* check flags associated with it.
358	*
359	* ASSUME that we can limit any special access handling to pages
360	* in page tables which the guest believes to be present.
361	*/
362	if (PdeSrc.n.u1Present)
363	{
364	RTGCPHYS GCPhys = NIL_RTGCPHYS;
365
366	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
367	# if PGM_GST_TYPE == PGM_TYPE_AMD64
368	bool fBigPagesSupported = true;
369	# else
370	bool fBigPagesSupported = !!(CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
371	# endif
372	if ( PdeSrc.b.u1Size
373	&& fBigPagesSupported)
374	GCPhys = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc)
375	\| ((RTGCPHYS)pvFault & (GST_BIG_PAGE_OFFSET_MASK ^ PAGE_OFFSET_MASK));
376	else
377	{
378	PGSTPT pPTSrc;
379	rc = PGM_GCPHYS_2_PTR(pVM, PdeSrc.u & GST_PDE_PG_MASK, &pPTSrc);
380	if (RT_SUCCESS(rc))
381	{
382	unsigned iPTESrc = (pvFault >> GST_PT_SHIFT) & GST_PT_MASK;
383	if (pPTSrc->a[iPTESrc].n.u1Present)
384	GCPhys = pPTSrc->a[iPTESrc].u & GST_PTE_PG_MASK;
385	}
386	}
387	# else
388	/* No paging so the fault address is the physical address */
389	GCPhys = (RTGCPHYS)(pvFault & ~PAGE_OFFSET_MASK);
390	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
391
392	/*
393	* If we have a GC address we'll check if it has any flags set.
394	*/
395	if (GCPhys != NIL_RTGCPHYS)
396	{
397	STAM_PROFILE_START(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
398
399	PPGMPAGE pPage;
400	rc = pgmPhysGetPageEx(&pVM->pgm.s, GCPhys, &pPage);
401	if (RT_SUCCESS(rc))
402	{
403	if ( PGM_PAGE_HAS_ACTIVE_PHYSICAL_HANDLERS(pPage)
404	\|\| PGM_PAGE_HAS_ACTIVE_VIRTUAL_HANDLERS(pPage))
405	{
406	if (PGM_PAGE_HAS_ANY_PHYSICAL_HANDLERS(pPage))
407	{
408	/*
409	* Physical page access handler.
410	*/
411	const RTGCPHYS GCPhysFault = GCPhys \| (pvFault & PAGE_OFFSET_MASK);
412	PPGMPHYSHANDLER pCur = (PPGMPHYSHANDLER)RTAvlroGCPhysRangeGet(&pVM->pgm.s.CTX_SUFF(pTrees)->PhysHandlers, GCPhysFault);
413	if (pCur)
414	{
415	# ifdef PGM_SYNC_N_PAGES
416	/*
417	* If the region is write protected and we got a page not present fault, then sync
418	* the pages. If the fault was caused by a read, then restart the instruction.
419	* In case of write access continue to the GC write handler.
420	*
421	* ASSUMES that there is only one handler per page or that they have similar write properties.
422	*/
423	if ( pCur->enmType == PGMPHYSHANDLERTYPE_PHYSICAL_WRITE
424	&& !(uErr & X86_TRAP_PF_P))
425	{
426	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, pvFault, PGM_SYNC_NR_PAGES, uErr);
427	if ( RT_FAILURE(rc)
428	\|\| !(uErr & X86_TRAP_PF_RW)
429	\|\| rc == VINF_PGM_SYNCPAGE_MODIFIED_PDE)
430	{
431	AssertRC(rc);
432	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersOutOfSync);
433	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
434	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2OutOfSyncHndPhys; });
435	return rc;
436	}
437	}
438	# endif
439
440	AssertMsg( pCur->enmType != PGMPHYSHANDLERTYPE_PHYSICAL_WRITE
441	\|\| (pCur->enmType == PGMPHYSHANDLERTYPE_PHYSICAL_WRITE && (uErr & X86_TRAP_PF_RW)),
442	("Unexpected trap for physical handler: %08X (phys=%08x) HCPhys=%X uErr=%X, enum=%d\n", pvFault, GCPhys, pPage->HCPhys, uErr, pCur->enmType));
443
444	# if defined(IN_RC) \|\| defined(IN_RING0)
445	if (pCur->CTX_SUFF(pfnHandler))
446	{
447	STAM_PROFILE_START(&pCur->Stat, h);
448	rc = pCur->CTX_SUFF(pfnHandler)(pVM, uErr, pRegFrame, pvFault, GCPhysFault, pCur->CTX_SUFF(pvUser));
449	STAM_PROFILE_STOP(&pCur->Stat, h);
450	}
451	else
452	# endif
453	rc = VINF_EM_RAW_EMULATE_INSTR;
454	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersPhysical);
455	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
456	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2HndPhys; });
457	return rc;
458	}
459	}
460	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
461	else
462	{
463	# ifdef PGM_SYNC_N_PAGES
464	/*
465	* If the region is write protected and we got a page not present fault, then sync
466	* the pages. If the fault was caused by a read, then restart the instruction.
467	* In case of write access continue to the GC write handler.
468	*/
469	if ( PGM_PAGE_GET_HNDL_VIRT_STATE(pPage) < PGM_PAGE_HNDL_PHYS_STATE_ALL
470	&& !(uErr & X86_TRAP_PF_P))
471	{
472	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, pvFault, PGM_SYNC_NR_PAGES, uErr);
473	if ( RT_FAILURE(rc)
474	\|\| rc == VINF_PGM_SYNCPAGE_MODIFIED_PDE
475	\|\| !(uErr & X86_TRAP_PF_RW))
476	{
477	AssertRC(rc);
478	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersOutOfSync);
479	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
480	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2OutOfSyncHndVirt; });
481	return rc;
482	}
483	}
484	# endif
485	/*
486	* Ok, it's an virtual page access handler.
487	*
488	* Since it's faster to search by address, we'll do that first
489	* and then retry by GCPhys if that fails.
490	*/
491	/** @todo r=bird: perhaps we should consider looking up by physical address directly now? */
492	/** @note r=svl: true, but lookup on virtual address should remain as a fallback as phys & virt trees might be out of sync, because the
493	* page was changed without us noticing it (not-present -> present without invlpg or mov cr3, xxx)
494	*/
495	PPGMVIRTHANDLER pCur = (PPGMVIRTHANDLER)RTAvlroGCPtrRangeGet(&pVM->pgm.s.CTX_SUFF(pTrees)->VirtHandlers, pvFault);
496	if (pCur)
497	{
498	AssertMsg(!(pvFault - pCur->Core.Key < pCur->cb)
499	\|\| ( pCur->enmType != PGMVIRTHANDLERTYPE_WRITE
500	\|\| !(uErr & X86_TRAP_PF_P)
501	\|\| (pCur->enmType == PGMVIRTHANDLERTYPE_WRITE && (uErr & X86_TRAP_PF_RW))),
502	("Unexpected trap for virtual handler: %RGv (phys=%RGp) HCPhys=%HGp uErr=%X, enum=%d\n", pvFault, GCPhys, pPage->HCPhys, uErr, pCur->enmType));
503
504	if ( pvFault - pCur->Core.Key < pCur->cb
505	&& ( uErr & X86_TRAP_PF_RW
506	\|\| pCur->enmType != PGMVIRTHANDLERTYPE_WRITE ) )
507	{
508	# ifdef IN_RC
509	STAM_PROFILE_START(&pCur->Stat, h);
510	rc = pCur->CTX_SUFF(pfnHandler)(pVM, uErr, pRegFrame, pvFault, pCur->Core.Key, pvFault - pCur->Core.Key);
511	STAM_PROFILE_STOP(&pCur->Stat, h);
512	# else
513	rc = VINF_EM_RAW_EMULATE_INSTR; /** @todo for VMX */
514	# endif
515	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersVirtual);
516	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
517	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2HndVirt; });
518	return rc;
519	}
520	/* Unhandled part of a monitored page */
521	}
522	else
523	{
524	/* Check by physical address. */
525	PPGMVIRTHANDLER pCur;
526	unsigned iPage;
527	rc = pgmHandlerVirtualFindByPhysAddr(pVM, GCPhys + (pvFault & PAGE_OFFSET_MASK),
528	&pCur, &iPage);
529	Assert(RT_SUCCESS(rc) \|\| !pCur);
530	if ( pCur
531	&& ( uErr & X86_TRAP_PF_RW
532	\|\| pCur->enmType != PGMVIRTHANDLERTYPE_WRITE ) )
533	{
534	Assert((pCur->aPhysToVirt[iPage].Core.Key & X86_PTE_PAE_PG_MASK) == GCPhys);
535	# ifdef IN_RC
536	RTGCPTR off = (iPage << PAGE_SHIFT) + (pvFault & PAGE_OFFSET_MASK) - (pCur->Core.Key & PAGE_OFFSET_MASK);
537	Assert(off < pCur->cb);
538	STAM_PROFILE_START(&pCur->Stat, h);
539	rc = pCur->CTX_SUFF(pfnHandler)(pVM, uErr, pRegFrame, pvFault, pCur->Core.Key, off);
540	STAM_PROFILE_STOP(&pCur->Stat, h);
541	# else
542	rc = VINF_EM_RAW_EMULATE_INSTR; /** @todo for VMX */
543	# endif
544	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersVirtualByPhys);
545	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
546	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2HndVirt; });
547	return rc;
548	}
549	}
550	}
551	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
552
553	/*
554	* There is a handled area of the page, but this fault doesn't belong to it.
555	* We must emulate the instruction.
556	*
557	* To avoid crashing (non-fatal) in the interpreter and go back to the recompiler
558	* we first check if this was a page-not-present fault for a page with only
559	* write access handlers. Restart the instruction if it wasn't a write access.
560	*/
561	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersUnhandled);
562
563	if ( !PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage)
564	&& !(uErr & X86_TRAP_PF_P))
565	{
566	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, pvFault, PGM_SYNC_NR_PAGES, uErr);
567	if ( RT_FAILURE(rc)
568	\|\| rc == VINF_PGM_SYNCPAGE_MODIFIED_PDE
569	\|\| !(uErr & X86_TRAP_PF_RW))
570	{
571	AssertRC(rc);
572	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersOutOfSync);
573	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
574	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2OutOfSyncHndPhys; });
575	return rc;
576	}
577	}
578
579	/** @todo This particular case can cause quite a lot of overhead. E.g. early stage of kernel booting in Ubuntu 6.06
580	* It's writing to an unhandled part of the LDT page several million times.
581	*/
582	rc = PGMInterpretInstruction(pVM, pRegFrame, pvFault);
583	LogFlow(("PGM: PGMInterpretInstruction -> rc=%d HCPhys=%RHp%s%s\n",
584	rc, pPage->HCPhys,
585	PGM_PAGE_HAS_ANY_PHYSICAL_HANDLERS(pPage) ? " phys" : "",
586	PGM_PAGE_HAS_ANY_VIRTUAL_HANDLERS(pPage) ? " virt" : ""));
587	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
588	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2HndUnhandled; });
589	return rc;
590	} /* if any kind of handler */
591
592	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
593	if (uErr & X86_TRAP_PF_P)
594	{
595	/*
596	* The page isn't marked, but it might still be monitored by a virtual page access handler.
597	* (ASSUMES no temporary disabling of virtual handlers.)
598	*/
599	/** @todo r=bird: Since the purpose is to catch out of sync pages with virtual handler(s) here,
600	* we should correct both the shadow page table and physical memory flags, and not only check for
601	* accesses within the handler region but for access to pages with virtual handlers. */
602	PPGMVIRTHANDLER pCur = (PPGMVIRTHANDLER)RTAvlroGCPtrRangeGet(&pVM->pgm.s.CTX_SUFF(pTrees)->VirtHandlers, pvFault);
603	if (pCur)
604	{
605	AssertMsg( !(pvFault - pCur->Core.Key < pCur->cb)
606	\|\| ( pCur->enmType != PGMVIRTHANDLERTYPE_WRITE
607	\|\| !(uErr & X86_TRAP_PF_P)
608	\|\| (pCur->enmType == PGMVIRTHANDLERTYPE_WRITE && (uErr & X86_TRAP_PF_RW))),
609	("Unexpected trap for virtual handler: %08X (phys=%08x) HCPhys=%X uErr=%X, enum=%d\n", pvFault, GCPhys, pPage->HCPhys, uErr, pCur->enmType));
610
611	if ( pvFault - pCur->Core.Key < pCur->cb
612	&& ( uErr & X86_TRAP_PF_RW
613	\|\| pCur->enmType != PGMVIRTHANDLERTYPE_WRITE ) )
614	{
615	# ifdef IN_RC
616	STAM_PROFILE_START(&pCur->Stat, h);
617	rc = pCur->CTX_SUFF(pfnHandler)(pVM, uErr, pRegFrame, pvFault, pCur->Core.Key, pvFault - pCur->Core.Key);
618	STAM_PROFILE_STOP(&pCur->Stat, h);
619	# else
620	rc = VINF_EM_RAW_EMULATE_INSTR; /** @todo for VMX */
621	# endif
622	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersVirtualUnmarked);
623	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
624	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2HndVirt; });
625	return rc;
626	}
627	}
628	}
629	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
630	}
631	else
632	{
633	/* When the guest accesses invalid physical memory (e.g. probing of RAM or accessing a remapped MMIO range), then we'll fall
634	* back to the recompiler to emulate the instruction.
635	*/
636	LogFlow(("pgmPhysGetPageEx %RGp failed with %Rrc\n", GCPhys, rc));
637	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersInvalid);
638	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
639	return VINF_EM_RAW_EMULATE_INSTR;
640	}
641
642	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
643
644	# ifdef PGM_OUT_OF_SYNC_IN_GC
645	/*
646	* We are here only if page is present in Guest page tables and trap is not handled
647	* by our handlers.
648	* Check it for page out-of-sync situation.
649	*/
650	STAM_PROFILE_START(&pVM->pgm.s.StatRZTrap0eTimeOutOfSync, c);
651
652	if (!(uErr & X86_TRAP_PF_P))
653	{
654	/*
655	* Page is not present in our page tables.
656	* Try to sync it!
657	* BTW, fPageShw is invalid in this branch!
658	*/
659	if (uErr & X86_TRAP_PF_US)
660	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,PageOutOfSyncUser));
661	else /* supervisor */
662	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,PageOutOfSyncSupervisor));
663
664	# if defined(LOG_ENABLED) && !defined(IN_RING0)
665	RTGCPHYS GCPhys;
666	uint64_t fPageGst;
667	PGMGstGetPage(pVM, pvFault, &fPageGst, &GCPhys);
668	Log(("Page out of sync: %RGv eip=%08x PdeSrc.n.u1User=%d fPageGst=%08llx GCPhys=%RGp scan=%d\n",
669	pvFault, pRegFrame->eip, PdeSrc.n.u1User, fPageGst, GCPhys, CSAMDoesPageNeedScanning(pVM, (RTRCPTR)pRegFrame->eip)));
670	# endif /* LOG_ENABLED */
671
672	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) && !defined(IN_RING0)
673	if (CPUMGetGuestCPL(pVM, pRegFrame) == 0)
674	{
675	uint64_t fPageGst;
676	rc = PGMGstGetPage(pVM, pvFault, &fPageGst, NULL);
677	if ( RT_SUCCESS(rc)
678	&& !(fPageGst & X86_PTE_US))
679	{
680	/* Note: can't check for X86_TRAP_ID bit, because that requires execute disable support on the CPU */
681	if ( pvFault == (RTGCPTR)pRegFrame->eip
682	\|\| pvFault - pRegFrame->eip < 8 /* instruction crossing a page boundary */
683	# ifdef CSAM_DETECT_NEW_CODE_PAGES
684	\|\| ( !PATMIsPatchGCAddr(pVM, (RTGCPTR)pRegFrame->eip)
685	&& CSAMDoesPageNeedScanning(pVM, (RTRCPTR)pRegFrame->eip)) /* any new code we encounter here */
686	# endif /* CSAM_DETECT_NEW_CODE_PAGES */
687	)
688	{
689	LogFlow(("CSAMExecFault %RX32\n", pRegFrame->eip));
690	rc = CSAMExecFault(pVM, (RTRCPTR)pRegFrame->eip);
691	if (rc != VINF_SUCCESS)
692	{
693	/*
694	* CSAM needs to perform a job in ring 3.
695	*
696	* Sync the page before going to the host context; otherwise we'll end up in a loop if
697	* CSAM fails (e.g. instruction crosses a page boundary and the next page is not present)
698	*/
699	LogFlow(("CSAM ring 3 job\n"));
700	int rc2 = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, pvFault, 1, uErr);
701	AssertRC(rc2);
702
703	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeOutOfSync, c);
704	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2CSAM; });
705	return rc;
706	}
707	}
708	# ifdef CSAM_DETECT_NEW_CODE_PAGES
709	else if ( uErr == X86_TRAP_PF_RW
710	&& pRegFrame->ecx >= 0x100 /* early check for movswd count */
711	&& pRegFrame->ecx < 0x10000)
712	{
713	/* In case of a write to a non-present supervisor shadow page, we'll take special precautions
714	* to detect loading of new code pages.
715	*/
716
717	/*
718	* Decode the instruction.
719	*/
720	RTGCPTR PC;
721	rc = SELMValidateAndConvertCSAddr(pVM, pRegFrame->eflags, pRegFrame->ss, pRegFrame->cs, &pRegFrame->csHid, (RTGCPTR)pRegFrame->eip, &PC);
722	if (rc == VINF_SUCCESS)
723	{
724	DISCPUSTATE Cpu;
725	uint32_t cbOp;
726	rc = EMInterpretDisasOneEx(pVM, PC, pRegFrame, &Cpu, &cbOp);
727
728	/* For now we'll restrict this to rep movsw/d instructions */
729	if ( rc == VINF_SUCCESS
730	&& Cpu.pCurInstr->opcode == OP_MOVSWD
731	&& (Cpu.prefix & PREFIX_REP))
732	{
733	CSAMMarkPossibleCodePage(pVM, pvFault);
734	}
735	}
736	}
737	# endif /* CSAM_DETECT_NEW_CODE_PAGES */
738
739	/*
740	* Mark this page as safe.
741	*/
742	/** @todo not correct for pages that contain both code and data!! */
743	Log2(("CSAMMarkPage %RGv; scanned=%d\n", pvFault, true));
744	CSAMMarkPage(pVM, (RTRCPTR)pvFault, true);
745	}
746	}
747	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) && !defined(IN_RING0) */
748	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, pvFault, PGM_SYNC_NR_PAGES, uErr);
749	if (RT_SUCCESS(rc))
750	{
751	/* The page was successfully synced, return to the guest. */
752	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeOutOfSync, c);
753	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2OutOfSync; });
754	return VINF_SUCCESS;
755	}
756	}
757	else
758	{
759	/*
760	* A side effect of not flushing global PDEs are out of sync pages due
761	* to physical monitored regions, that are no longer valid.
762	* Assume for now it only applies to the read/write flag
763	*/
764	if (RT_SUCCESS(rc) && (uErr & X86_TRAP_PF_RW))
765	{
766	if (uErr & X86_TRAP_PF_US)
767	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,PageOutOfSyncUser));
768	else /* supervisor */
769	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,PageOutOfSyncSupervisor));
770
771
772	/*
773	* Note: Do NOT use PGM_SYNC_NR_PAGES here. That only works if the page is not present, which is not true in this case.
774	*/
775	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, pvFault, 1, uErr);
776	if (RT_SUCCESS(rc))
777	{
778	/*
779	* Page was successfully synced, return to guest.
780	*/
781	# ifdef VBOX_STRICT
782	RTGCPHYS GCPhys;
783	uint64_t fPageGst;
784	rc = PGMGstGetPage(pVM, pvFault, &fPageGst, &GCPhys);
785	Assert(RT_SUCCESS(rc) && fPageGst & X86_PTE_RW);
786	LogFlow(("Obsolete physical monitor page out of sync %RGv - phys %RGp flags=%08llx\n", pvFault, GCPhys, (uint64_t)fPageGst));
787
788	uint64_t fPageShw;
789	rc = PGMShwGetPage(pVM, pvFault, &fPageShw, NULL);
790	AssertMsg(RT_SUCCESS(rc) && fPageShw & X86_PTE_RW, ("rc=%Rrc fPageShw=%RX64\n", rc, fPageShw));
791	# endif /* VBOX_STRICT */
792	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeOutOfSync, c);
793	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2OutOfSyncHndObs; });
794	return VINF_SUCCESS;
795	}
796
797	/* Check to see if we need to emulate the instruction as X86_CR0_WP has been cleared. */
798	if ( CPUMGetGuestCPL(pVM, pRegFrame) == 0
799	&& ((CPUMGetGuestCR0(pVM) & (X86_CR0_WP \| X86_CR0_PG)) == X86_CR0_PG)
800	&& (uErr & (X86_TRAP_PF_RW \| X86_TRAP_PF_P)) == (X86_TRAP_PF_RW \| X86_TRAP_PF_P))
801	{
802	uint64_t fPageGst;
803	rc = PGMGstGetPage(pVM, pvFault, &fPageGst, NULL);
804	if ( RT_SUCCESS(rc)
805	&& !(fPageGst & X86_PTE_RW))
806	{
807	rc = PGMInterpretInstruction(pVM, pRegFrame, pvFault);
808	if (RT_SUCCESS(rc))
809	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eWPEmulInRZ);
810	else
811	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eWPEmulToR3);
812	return rc;
813	}
814	AssertMsgFailed(("Unexpected r/w page %RGv flag=%x rc=%Rrc\n", pvFault, (uint32_t)fPageGst, rc));
815	}
816	}
817
818	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
819	# ifdef VBOX_STRICT
820	/*
821	* Check for VMM page flags vs. Guest page flags consistency.
822	* Currently only for debug purposes.
823	*/
824	if (RT_SUCCESS(rc))
825	{
826	/* Get guest page flags. */
827	uint64_t fPageGst;
828	rc = PGMGstGetPage(pVM, pvFault, &fPageGst, NULL);
829	if (RT_SUCCESS(rc))
830	{
831	uint64_t fPageShw;
832	rc = PGMShwGetPage(pVM, pvFault, &fPageShw, NULL);
833
834	/*
835	* Compare page flags.
836	* Note: we have AVL, A, D bits desynched.
837	*/
838	AssertMsg((fPageShw & ~(X86_PTE_A \| X86_PTE_D \| X86_PTE_AVL_MASK)) == (fPageGst & ~(X86_PTE_A \| X86_PTE_D \| X86_PTE_AVL_MASK)),
839	("Page flags mismatch! pvFault=%RGv GCPhys=%RGp fPageShw=%08llx fPageGst=%08llx\n", pvFault, GCPhys, fPageShw, fPageGst));
840	}
841	else
842	AssertMsgFailed(("PGMGstGetPage rc=%Rrc\n", rc));
843	}
844	else
845	AssertMsgFailed(("PGMGCGetPage rc=%Rrc\n", rc));
846	# endif /* VBOX_STRICT */
847	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
848	}
849	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeOutOfSync, c);
850	# endif /* PGM_OUT_OF_SYNC_IN_GC */
851	}
852	else
853	{
854	/*
855	* Page not present in Guest OS or invalid page table address.
856	* This is potential virtual page access handler food.
857	*
858	* For the present we'll say that our access handlers don't
859	* work for this case - we've already discarded the page table
860	* not present case which is identical to this.
861	*
862	* When we perchance find we need this, we will probably have AVL
863	* trees (offset based) to operate on and we can measure their speed
864	* agains mapping a page table and probably rearrange this handling
865	* a bit. (Like, searching virtual ranges before checking the
866	* physical address.)
867	*/
868	}
869	}
870
871
872	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
873	/*
874	* Conclusion, this is a guest trap.
875	*/
876	LogFlow(("PGM: Unhandled #PF -> route trap to recompiler!\n"));
877	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eGuestPFUnh);
878	return VINF_EM_RAW_GUEST_TRAP;
879	# else
880	/* present, but not a monitored page; perhaps the guest is probing physical memory */
881	return VINF_EM_RAW_EMULATE_INSTR;
882	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
883
884
885	# else /* PGM_GST_TYPE != PGM_TYPE_32BIT */
886
887	AssertReleaseMsgFailed(("Shw=%d Gst=%d is not implemented!\n", PGM_GST_TYPE, PGM_SHW_TYPE));
888	return VERR_INTERNAL_ERROR;
889	# endif /* PGM_GST_TYPE != PGM_TYPE_32BIT */
890	}
891	#endif /* !IN_RING3 */
892
893
894	/**
895	* Emulation of the invlpg instruction.
896	*
897	*
898	* @returns VBox status code.
899	*
900	* @param pVM VM handle.
901	* @param GCPtrPage Page to invalidate.
902	*
903	* @remark ASSUMES that the guest is updating before invalidating. This order
904	* isn't required by the CPU, so this is speculative and could cause
905	* trouble.
906	*
907	* @todo Flush page or page directory only if necessary!
908	* @todo Add a #define for simply invalidating the page.
909	*/
910	PGM_BTH_DECL(int, InvalidatePage)(PVM pVM, RTGCPTR GCPtrPage)
911	{
912	#if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) \
913	&& PGM_SHW_TYPE != PGM_TYPE_NESTED \
914	&& PGM_SHW_TYPE != PGM_TYPE_EPT
915	int rc;
916
917	LogFlow(("InvalidatePage %RGv\n", GCPtrPage));
918	/*
919	* Get the shadow PD entry and skip out if this PD isn't present.
920	* (Guessing that it is frequent for a shadow PDE to not be present, do this first.)
921	*/
922	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
923	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
924	PX86PDE pPdeDst = pgmShwGet32BitPDEPtr(&pVM->pgm.s, GCPtrPage);
925
926	# ifdef VBOX_WITH_PGMPOOL_PAGING_ONLY
927	/* Fetch the pgm pool shadow descriptor. */
928	PPGMPOOLPAGE pShwPde = pVM->pgm.s.CTX_SUFF(pShwPageCR3);
929	Assert(pShwPde);
930	# endif
931
932	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
933	const unsigned iPdpt = (GCPtrPage >> X86_PDPT_SHIFT);
934	PX86PDPT pPdptDst = pgmShwGetPaePDPTPtr(&pVM->pgm.s);
935
936	/* If the shadow PDPE isn't present, then skip the invalidate. */
937	if (!pPdptDst->a[iPdpt].n.u1Present)
938	{
939	Assert(!(pPdptDst->a[iPdpt].u & PGM_PLXFLAGS_MAPPING));
940	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePageSkipped));
941	return VINF_SUCCESS;
942	}
943
944	# ifdef VBOX_WITH_PGMPOOL_PAGING_ONLY
945	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
946	PPGMPOOLPAGE pShwPde;
947	PX86PDPAE pPDDst;
948
949	/* Fetch the pgm pool shadow descriptor. */
950	rc = pgmShwGetPaePoolPagePD(&pVM->pgm.s, GCPtrPage, &pShwPde);
951	AssertRCSuccessReturn(rc, rc);
952	Assert(pShwPde);
953
954	pPDDst = (PX86PDPAE)PGMPOOL_PAGE_2_PTR_BY_PGM(&pVM->pgm.s, pShwPde);
955	PX86PDEPAE pPdeDst = &pPDDst->a[iPDDst];
956	# else
957	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) /& SHW_PD_MASK - pool index only atm! /;
958	PX86PDEPAE pPdeDst = pgmShwGetPaePDEPtr(&pVM->pgm.s, GCPtrPage);
959	# endif
960
961	# else /* PGM_SHW_TYPE == PGM_TYPE_AMD64 */
962	/* PML4 */
963	AssertReturn(pVM->pgm.s.pShwRootR3, VERR_INTERNAL_ERROR);
964
965	const unsigned iPml4 = (GCPtrPage >> X86_PML4_SHIFT) & X86_PML4_MASK;
966	const unsigned iPdpt = (GCPtrPage >> X86_PDPT_SHIFT) & X86_PDPT_MASK_AMD64;
967	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
968	PX86PDPAE pPDDst;
969	PX86PDPT pPdptDst;
970	PX86PML4E pPml4eDst;
971	rc = pgmShwGetLongModePDPtr(pVM, GCPtrPage, &pPml4eDst, &pPdptDst, &pPDDst);
972	if (rc != VINF_SUCCESS)
973	{
974	AssertMsg(rc == VERR_PAGE_DIRECTORY_PTR_NOT_PRESENT \|\| rc == VERR_PAGE_MAP_LEVEL4_NOT_PRESENT, ("Unexpected rc=%Rrc\n", rc));
975	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePageSkipped));
976	if (!VM_FF_ISSET(pVM, VM_FF_PGM_SYNC_CR3))
977	PGM_INVL_GUEST_TLBS();
978	return VINF_SUCCESS;
979	}
980	Assert(pPDDst);
981
982	PX86PDEPAE pPdeDst = &pPDDst->a[iPDDst];
983	PX86PDPE pPdpeDst = &pPdptDst->a[iPdpt];
984
985	if (!pPdpeDst->n.u1Present)
986	{
987	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePageSkipped));
988	if (!VM_FF_ISSET(pVM, VM_FF_PGM_SYNC_CR3))
989	PGM_INVL_GUEST_TLBS();
990	return VINF_SUCCESS;
991	}
992
993	# endif /* PGM_SHW_TYPE == PGM_TYPE_AMD64 */
994
995	const SHWPDE PdeDst = *pPdeDst;
996	if (!PdeDst.n.u1Present)
997	{
998	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePageSkipped));
999	return VINF_SUCCESS;
1000	}
1001
1002	/*
1003	* Get the guest PD entry and calc big page.
1004	*/
1005	# if PGM_GST_TYPE == PGM_TYPE_32BIT
1006	PGSTPD pPDSrc = pgmGstGet32bitPDPtr(&pVM->pgm.s);
1007	const unsigned iPDSrc = GCPtrPage >> GST_PD_SHIFT;
1008	GSTPDE PdeSrc = pPDSrc->a[iPDSrc];
1009	# else /* PGM_GST_TYPE != PGM_TYPE_32BIT */
1010	unsigned iPDSrc;
1011	# if PGM_GST_TYPE == PGM_TYPE_PAE
1012	X86PDPE PdpeSrc;
1013	PX86PDPAE pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, GCPtrPage, &iPDSrc, &PdpeSrc);
1014	# else /* AMD64 */
1015	PX86PML4E pPml4eSrc;
1016	X86PDPE PdpeSrc;
1017	PX86PDPAE pPDSrc = pgmGstGetLongModePDPtr(&pVM->pgm.s, GCPtrPage, &pPml4eSrc, &PdpeSrc, &iPDSrc);
1018	# endif
1019	GSTPDE PdeSrc;
1020
1021	if (pPDSrc)
1022	PdeSrc = pPDSrc->a[iPDSrc];
1023	else
1024	PdeSrc.u = 0;
1025	# endif /* PGM_GST_TYPE != PGM_TYPE_32BIT */
1026
1027	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1028	const bool fIsBigPage = PdeSrc.b.u1Size;
1029	# else
1030	const bool fIsBigPage = PdeSrc.b.u1Size && (CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
1031	# endif
1032
1033	# ifdef IN_RING3
1034	/*
1035	* If a CR3 Sync is pending we may ignore the invalidate page operation
1036	* depending on the kind of sync and if it's a global page or not.
1037	* This doesn't make sense in GC/R0 so we'll skip it entirely there.
1038	*/
1039	# ifdef PGM_SKIP_GLOBAL_PAGEDIRS_ON_NONGLOBAL_FLUSH
1040	if ( VM_FF_ISSET(pVM, VM_FF_PGM_SYNC_CR3)
1041	\|\| ( VM_FF_ISSET(pVM, VM_FF_PGM_SYNC_CR3_NON_GLOBAL)
1042	&& fIsBigPage
1043	&& PdeSrc.b.u1Global
1044	)
1045	)
1046	# else
1047	if (VM_FF_ISPENDING(pVM, VM_FF_PGM_SYNC_CR3 \| VM_FF_PGM_SYNC_CR3_NON_GLOBAL) )
1048	# endif
1049	{
1050	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePageSkipped));
1051	return VINF_SUCCESS;
1052	}
1053	# endif /* IN_RING3 */
1054
1055	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1056	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
1057
1058	/* Fetch the pgm pool shadow descriptor. */
1059	PPGMPOOLPAGE pShwPdpt = pgmPoolGetPageByHCPhys(pVM, pPml4eDst->u & X86_PML4E_PG_MASK);
1060	Assert(pShwPdpt);
1061
1062	/* Fetch the pgm pool shadow descriptor. */
1063	PPGMPOOLPAGE pShwPde = pgmPoolGetPageByHCPhys(pVM, pPdptDst->a[iPdpt].u & SHW_PDPE_PG_MASK);
1064	Assert(pShwPde);
1065
1066	Assert(pPml4eDst->n.u1Present && (pPml4eDst->u & SHW_PDPT_MASK));
1067	RTGCPHYS GCPhysPdpt = pPml4eSrc->u & X86_PML4E_PG_MASK;
1068
1069	if ( !pPml4eSrc->n.u1Present
1070	\|\| pShwPdpt->GCPhys != GCPhysPdpt)
1071	{
1072	LogFlow(("InvalidatePage: Out-of-sync PML4E (P/GCPhys) at %RGv GCPhys=%RGp vs %RGp Pml4eSrc=%RX64 Pml4eDst=%RX64\n",
1073	GCPtrPage, pShwPdpt->GCPhys, GCPhysPdpt, (uint64_t)pPml4eSrc->u, (uint64_t)pPml4eDst->u));
1074	pgmPoolFreeByPage(pPool, pShwPdpt, pVM->pgm.s.CTX_SUFF(pShwPageCR3)->idx, iPml4);
1075	pPml4eDst->u = 0;
1076	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDNPs));
1077	PGM_INVL_GUEST_TLBS();
1078	return VINF_SUCCESS;
1079	}
1080	if ( pPml4eSrc->n.u1User != pPml4eDst->n.u1User
1081	\|\| (!pPml4eSrc->n.u1Write && pPml4eDst->n.u1Write))
1082	{
1083	/*
1084	* Mark not present so we can resync the PML4E when it's used.
1085	*/
1086	LogFlow(("InvalidatePage: Out-of-sync PML4E at %RGv Pml4eSrc=%RX64 Pml4eDst=%RX64\n",
1087	GCPtrPage, (uint64_t)pPml4eSrc->u, (uint64_t)pPml4eDst->u));
1088	pgmPoolFreeByPage(pPool, pShwPdpt, pVM->pgm.s.CTX_SUFF(pShwPageCR3)->idx, iPml4);
1089	pPml4eDst->u = 0;
1090	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDOutOfSync));
1091	PGM_INVL_GUEST_TLBS();
1092	}
1093	else if (!pPml4eSrc->n.u1Accessed)
1094	{
1095	/*
1096	* Mark not present so we can set the accessed bit.
1097	*/
1098	LogFlow(("InvalidatePage: Out-of-sync PML4E (A) at %RGv Pml4eSrc=%RX64 Pml4eDst=%RX64\n",
1099	GCPtrPage, (uint64_t)pPml4eSrc->u, (uint64_t)pPml4eDst->u));
1100	pgmPoolFreeByPage(pPool, pShwPdpt, pVM->pgm.s.CTX_SUFF(pShwPageCR3)->idx, iPml4);
1101	pPml4eDst->u = 0;
1102	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDNAs));
1103	PGM_INVL_GUEST_TLBS();
1104	}
1105
1106	/* Check if the PDPT entry has changed. */
1107	Assert(pPdpeDst->n.u1Present && pPdpeDst->u & SHW_PDPT_MASK);
1108	RTGCPHYS GCPhysPd = PdpeSrc.u & GST_PDPE_PG_MASK;
1109	if ( !PdpeSrc.n.u1Present
1110	\|\| pShwPde->GCPhys != GCPhysPd)
1111	{
1112	LogFlow(("InvalidatePage: Out-of-sync PDPE (P/GCPhys) at %RGv GCPhys=%RGp vs %RGp PdpeSrc=%RX64 PdpeDst=%RX64\n",
1113	GCPtrPage, pShwPde->GCPhys, GCPhysPd, (uint64_t)PdpeSrc.u, (uint64_t)pPdpeDst->u));
1114	pgmPoolFreeByPage(pPool, pShwPde, pShwPdpt->idx, iPdpt);
1115	pPdpeDst->u = 0;
1116	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDNPs));
1117	PGM_INVL_GUEST_TLBS();
1118	return VINF_SUCCESS;
1119	}
1120	if ( PdpeSrc.lm.u1User != pPdpeDst->lm.u1User
1121	\|\| (!PdpeSrc.lm.u1Write && pPdpeDst->lm.u1Write))
1122	{
1123	/*
1124	* Mark not present so we can resync the PDPTE when it's used.
1125	*/
1126	LogFlow(("InvalidatePage: Out-of-sync PDPE at %RGv PdpeSrc=%RX64 PdpeDst=%RX64\n",
1127	GCPtrPage, (uint64_t)PdpeSrc.u, (uint64_t)pPdpeDst->u));
1128	pgmPoolFreeByPage(pPool, pShwPde, pShwPdpt->idx, iPdpt);
1129	pPdpeDst->u = 0;
1130	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDOutOfSync));
1131	PGM_INVL_GUEST_TLBS();
1132	}
1133	else if (!PdpeSrc.lm.u1Accessed)
1134	{
1135	/*
1136	* Mark not present so we can set the accessed bit.
1137	*/
1138	LogFlow(("InvalidatePage: Out-of-sync PDPE (A) at %RGv PdpeSrc=%RX64 PdpeDst=%RX64\n",
1139	GCPtrPage, (uint64_t)PdpeSrc.u, (uint64_t)pPdpeDst->u));
1140	pgmPoolFreeByPage(pPool, pShwPde, pShwPdpt->idx, iPdpt);
1141	pPdpeDst->u = 0;
1142	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDNAs));
1143	PGM_INVL_GUEST_TLBS();
1144	}
1145	# endif /* PGM_GST_TYPE == PGM_TYPE_AMD64 */
1146
1147	# if PGM_GST_TYPE == PGM_TYPE_PAE && !defined(VBOX_WITH_PGMPOOL_PAGING_ONLY)
1148	/*
1149	* Update the shadow PDPE and free all the shadow PD entries if the PDPE is marked not present.
1150	* Note: This shouldn't actually be necessary as we monitor the PDPT page for changes.
1151	*/
1152	if (!pPDSrc)
1153	{
1154	/* Guest PDPE not present */
1155	PX86PDPAE pPDDst = pgmShwGetPaePDPtr(&pVM->pgm.s, GCPtrPage);
1156	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
1157
1158	Assert(!PdpeSrc.n.u1Present);
1159	LogFlow(("InvalidatePage: guest PDPE %d not present; clear shw pdpe\n", iPdpt));
1160
1161	/* for each page directory entry */
1162	for (unsigned iPD = 0; iPD < X86_PG_PAE_ENTRIES; iPD++)
1163	{
1164	if ( pPDDst->a[iPD].n.u1Present
1165	&& !(pPDDst->a[iPD].u & PGM_PDFLAGS_MAPPING))
1166	{
1167	pgmPoolFree(pVM, pPDDst->a[iPD].u & SHW_PDE_PG_MASK, SHW_POOL_ROOT_IDX, iPdpt * X86_PG_PAE_ENTRIES + iPD);
1168	pPDDst->a[iPD].u = 0;
1169	}
1170	}
1171	if (!(pPdptDst->a[iPdpt].u & PGM_PLXFLAGS_MAPPING))
1172	pPdptDst->a[iPdpt].n.u1Present = 0;
1173	PGM_INVL_GUEST_TLBS();
1174	}
1175	AssertMsg(pVM->pgm.s.fMappingsFixed \|\| (PdpeSrc.u & X86_PDPE_PG_MASK) == pVM->pgm.s.aGCPhysGstPaePDsMonitored[iPdpt], ("%RGp vs %RGp (mon)\n", (PdpeSrc.u & X86_PDPE_PG_MASK), pVM->pgm.s.aGCPhysGstPaePDsMonitored[iPdpt]));
1176	# endif
1177
1178
1179	/*
1180	* Deal with the Guest PDE.
1181	*/
1182	rc = VINF_SUCCESS;
1183	if (PdeSrc.n.u1Present)
1184	{
1185	if (PdeDst.u & PGM_PDFLAGS_MAPPING)
1186	{
1187	/*
1188	* Conflict - Let SyncPT deal with it to avoid duplicate code.
1189	*/
1190	Assert(pgmMapAreMappingsEnabled(&pVM->pgm.s));
1191	Assert(PGMGetGuestMode(pVM) <= PGMMODE_PAE);
1192	rc = PGM_BTH_NAME(SyncPT)(pVM, iPDSrc, pPDSrc, GCPtrPage);
1193	}
1194	else if ( PdeSrc.n.u1User != PdeDst.n.u1User
1195	\|\| (!PdeSrc.n.u1Write && PdeDst.n.u1Write))
1196	{
1197	/*
1198	* Mark not present so we can resync the PDE when it's used.
1199	*/
1200	LogFlow(("InvalidatePage: Out-of-sync at %RGp PdeSrc=%RX64 PdeDst=%RX64\n",
1201	GCPtrPage, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
1202	# if PGM_GST_TYPE == PGM_TYPE_AMD64 \|\| defined(VBOX_WITH_PGMPOOL_PAGING_ONLY)
1203	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, pShwPde->idx, iPDDst);
1204	# else
1205	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, SHW_POOL_ROOT_IDX, iPDDst);
1206	# endif
1207	pPdeDst->u = 0;
1208	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDOutOfSync));
1209	PGM_INVL_GUEST_TLBS();
1210	}
1211	else if (!PdeSrc.n.u1Accessed)
1212	{
1213	/*
1214	* Mark not present so we can set the accessed bit.
1215	*/
1216	LogFlow(("InvalidatePage: Out-of-sync (A) at %RGp PdeSrc=%RX64 PdeDst=%RX64\n",
1217	GCPtrPage, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
1218	# if PGM_GST_TYPE == PGM_TYPE_AMD64 \|\| defined(VBOX_WITH_PGMPOOL_PAGING_ONLY)
1219	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, pShwPde->idx, iPDDst);
1220	# else
1221	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, SHW_POOL_ROOT_IDX, iPDDst);
1222	# endif
1223	pPdeDst->u = 0;
1224	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDNAs));
1225	PGM_INVL_GUEST_TLBS();
1226	}
1227	else if (!fIsBigPage)
1228	{
1229	/*
1230	* 4KB - page.
1231	*/
1232	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, PdeDst.u & SHW_PDE_PG_MASK);
1233	RTGCPHYS GCPhys = PdeSrc.u & GST_PDE_PG_MASK;
1234	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
1235	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
1236	GCPhys \|= (iPDDst & 1) * (PAGE_SIZE/2);
1237	# endif
1238	if (pShwPage->GCPhys == GCPhys)
1239	{
1240	# if 0 /* likely cause of a major performance regression; must be SyncPageWorkerTrackDeref then */
1241	const unsigned iPTEDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1242	PSHWPT pPT = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
1243	if (pPT->a[iPTEDst].n.u1Present)
1244	{
1245	# ifdef PGMPOOL_WITH_USER_TRACKING
1246	/* This is very unlikely with caching/monitoring enabled. */
1247	PGM_BTH_NAME(SyncPageWorkerTrackDeref)(pVM, pShwPage, pPT->a[iPTEDst].u & SHW_PTE_PG_MASK);
1248	# endif
1249	pPT->a[iPTEDst].u = 0;
1250	}
1251	# else /* Syncing it here isn't 100% safe and it's probably not worth spending time syncing it. */
1252	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, GCPtrPage, 1, 0);
1253	if (RT_SUCCESS(rc))
1254	rc = VINF_SUCCESS;
1255	# endif
1256	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePage4KBPages));
1257	PGM_INVL_PG(GCPtrPage);
1258	}
1259	else
1260	{
1261	/*
1262	* The page table address changed.
1263	*/
1264	LogFlow(("InvalidatePage: Out-of-sync at %RGp PdeSrc=%RX64 PdeDst=%RX64 ShwGCPhys=%RGp iPDDst=%#x\n",
1265	GCPtrPage, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u, pShwPage->GCPhys, iPDDst));
1266	# if PGM_GST_TYPE == PGM_TYPE_AMD64 \|\| defined(VBOX_WITH_PGMPOOL_PAGING_ONLY)
1267	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, pShwPde->idx, iPDDst);
1268	# else
1269	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, SHW_POOL_ROOT_IDX, iPDDst);
1270	# endif
1271	pPdeDst->u = 0;
1272	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDOutOfSync));
1273	PGM_INVL_GUEST_TLBS();
1274	}
1275	}
1276	else
1277	{
1278	/*
1279	* 2/4MB - page.
1280	*/
1281	/* Before freeing the page, check if anything really changed. */
1282	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, PdeDst.u & SHW_PDE_PG_MASK);
1283	RTGCPHYS GCPhys = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc);
1284	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
1285	/* Select the right PDE as we're emulating a 4MB page directory with two 2 MB shadow PDEs.*/
1286	GCPhys \|= GCPtrPage & (1 << X86_PD_PAE_SHIFT);
1287	# endif
1288	if ( pShwPage->GCPhys == GCPhys
1289	&& pShwPage->enmKind == BTH_PGMPOOLKIND_PT_FOR_BIG)
1290	{
1291	/* ASSUMES a the given bits are identical for 4M and normal PDEs */
1292	/** @todo PAT */
1293	if ( (PdeSrc.u & (X86_PDE_P \| X86_PDE_RW \| X86_PDE_US \| X86_PDE_PWT \| X86_PDE_PCD))
1294	== (PdeDst.u & (X86_PDE_P \| X86_PDE_RW \| X86_PDE_US \| X86_PDE_PWT \| X86_PDE_PCD))
1295	&& ( PdeSrc.b.u1Dirty /** @todo rainy day: What about read-only 4M pages? not very common, but still... */
1296	\|\| (PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY)))
1297	{
1298	LogFlow(("Skipping flush for big page containing %RGv (PD=%X .u=%RX64)-> nothing has changed!\n", GCPtrPage, iPDSrc, PdeSrc.u));
1299	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePage4MBPagesSkip));
1300	return VINF_SUCCESS;
1301	}
1302	}
1303
1304	/*
1305	* Ok, the page table is present and it's been changed in the guest.
1306	* If we're in host context, we'll just mark it as not present taking the lazy approach.
1307	* We could do this for some flushes in GC too, but we need an algorithm for
1308	* deciding which 4MB pages containing code likely to be executed very soon.
1309	*/
1310	LogFlow(("InvalidatePage: Out-of-sync PD at %RGp PdeSrc=%RX64 PdeDst=%RX64\n",
1311	GCPtrPage, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
1312	# if PGM_GST_TYPE == PGM_TYPE_AMD64 \|\| defined(VBOX_WITH_PGMPOOL_PAGING_ONLY)
1313	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, pShwPde->idx, iPDDst);
1314	# else
1315	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, SHW_POOL_ROOT_IDX, iPDDst);
1316	# endif
1317	pPdeDst->u = 0;
1318	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePage4MBPages));
1319	PGM_INVL_BIG_PG(GCPtrPage);
1320	}
1321	}
1322	else
1323	{
1324	/*
1325	* Page directory is not present, mark shadow PDE not present.
1326	*/
1327	if (!(PdeDst.u & PGM_PDFLAGS_MAPPING))
1328	{
1329	# if PGM_GST_TYPE == PGM_TYPE_AMD64 \|\| defined(VBOX_WITH_PGMPOOL_PAGING_ONLY)
1330	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, pShwPde->idx, iPDDst);
1331	# else
1332	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, SHW_POOL_ROOT_IDX, iPDDst);
1333	# endif
1334	pPdeDst->u = 0;
1335	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDNPs));
1336	PGM_INVL_PG(GCPtrPage);
1337	}
1338	else
1339	{
1340	Assert(pgmMapAreMappingsEnabled(&pVM->pgm.s));
1341	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDMappings));
1342	}
1343	}
1344
1345	return rc;
1346
1347	#else /* guest real and protected mode */
1348	/* There's no such thing as InvalidatePage when paging is disabled, so just ignore. */
1349	return VINF_SUCCESS;
1350	#endif
1351	}
1352
1353
1354	#ifdef PGMPOOL_WITH_USER_TRACKING
1355	/**
1356	* Update the tracking of shadowed pages.
1357	*
1358	* @param pVM The VM handle.
1359	* @param pShwPage The shadow page.
1360	* @param HCPhys The physical page we is being dereferenced.
1361	*/
1362	DECLINLINE(void) PGM_BTH_NAME(SyncPageWorkerTrackDeref)(PVM pVM, PPGMPOOLPAGE pShwPage, RTHCPHYS HCPhys)
1363	{
1364	# ifdef PGMPOOL_WITH_GCPHYS_TRACKING
1365	STAM_PROFILE_START(&pVM->pgm.s.StatTrackDeref, a);
1366	LogFlow(("SyncPageWorkerTrackDeref: Damn HCPhys=%RHp pShwPage->idx=%#x!!!\n", HCPhys, pShwPage->idx));
1367
1368	/** @todo If this turns out to be a bottle neck (very likely) two things can be done:
1369	* 1. have a medium sized HCPhys -> GCPhys TLB (hash?)
1370	* 2. write protect all shadowed pages. I.e. implement caching.
1371	*/
1372	/*
1373	* Find the guest address.
1374	*/
1375	for (PPGMRAMRANGE pRam = pVM->pgm.s.CTX_SUFF(pRamRanges);
1376	pRam;
1377	pRam = pRam->CTX_SUFF(pNext))
1378	{
1379	unsigned iPage = pRam->cb >> PAGE_SHIFT;
1380	while (iPage-- > 0)
1381	{
1382	if (PGM_PAGE_GET_HCPHYS(&pRam->aPages[iPage]) == HCPhys)
1383	{
1384	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
1385	pgmTrackDerefGCPhys(pPool, pShwPage, &pRam->aPages[iPage]);
1386	pShwPage->cPresent--;
1387	pPool->cPresent--;
1388	STAM_PROFILE_STOP(&pVM->pgm.s.StatTrackDeref, a);
1389	return;
1390	}
1391	}
1392	}
1393
1394	for (;;)
1395	AssertReleaseMsgFailed(("HCPhys=%RHp wasn't found!\n", HCPhys));
1396	# else /* !PGMPOOL_WITH_GCPHYS_TRACKING */
1397	pShwPage->cPresent--;
1398	pVM->pgm.s.CTX_SUFF(pPool)->cPresent--;
1399	# endif /* !PGMPOOL_WITH_GCPHYS_TRACKING */
1400	}
1401
1402
1403	/**
1404	* Update the tracking of shadowed pages.
1405	*
1406	* @param pVM The VM handle.
1407	* @param pShwPage The shadow page.
1408	* @param u16 The top 16-bit of the pPage->HCPhys.
1409	* @param pPage Pointer to the guest page. this will be modified.
1410	* @param iPTDst The index into the shadow table.
1411	*/
1412	DECLINLINE(void) PGM_BTH_NAME(SyncPageWorkerTrackAddref)(PVM pVM, PPGMPOOLPAGE pShwPage, uint16_t u16, PPGMPAGE pPage, const unsigned iPTDst)
1413	{
1414	# ifdef PGMPOOL_WITH_GCPHYS_TRACKING
1415	/*
1416	* We're making certain assumptions about the placement of cRef and idx.
1417	*/
1418	Assert(MM_RAM_FLAGS_IDX_SHIFT == 48);
1419	Assert(MM_RAM_FLAGS_CREFS_SHIFT > MM_RAM_FLAGS_IDX_SHIFT);
1420
1421	/*
1422	* Just deal with the simple first time here.
1423	*/
1424	if (!u16)
1425	{
1426	STAM_COUNTER_INC(&pVM->pgm.s.StatTrackVirgin);
1427	u16 = (1 << (MM_RAM_FLAGS_CREFS_SHIFT - MM_RAM_FLAGS_IDX_SHIFT)) \| pShwPage->idx;
1428	}
1429	else
1430	u16 = pgmPoolTrackPhysExtAddref(pVM, u16, pShwPage->idx);
1431
1432	/* write back, trying to be clever... */
1433	Log2(("SyncPageWorkerTrackAddRef: u16=%#x pPage->HCPhys=%RHp->%RHp iPTDst=%#x\n",
1434	u16, pPage->HCPhys, (pPage->HCPhys & MM_RAM_FLAGS_NO_REFS_MASK) \| ((uint64_t)u16 << MM_RAM_FLAGS_CREFS_SHIFT), iPTDst));
1435	((uint16_t )&pPage->HCPhys + 3) = u16; /** @todo PAGE FLAGS */
1436	# endif /* PGMPOOL_WITH_GCPHYS_TRACKING */
1437
1438	/* update statistics. */
1439	pVM->pgm.s.CTX_SUFF(pPool)->cPresent++;
1440	pShwPage->cPresent++;
1441	if (pShwPage->iFirstPresent > iPTDst)
1442	pShwPage->iFirstPresent = iPTDst;
1443	}
1444	#endif /* PGMPOOL_WITH_USER_TRACKING */
1445
1446
1447	/**
1448	* Creates a 4K shadow page for a guest page.
1449	*
1450	* For 4M pages the caller must convert the PDE4M to a PTE, this includes adjusting the
1451	* physical address. The PdeSrc argument only the flags are used. No page structured
1452	* will be mapped in this function.
1453	*
1454	* @param pVM VM handle.
1455	* @param pPteDst Destination page table entry.
1456	* @param PdeSrc Source page directory entry (i.e. Guest OS page directory entry).
1457	* Can safely assume that only the flags are being used.
1458	* @param PteSrc Source page table entry (i.e. Guest OS page table entry).
1459	* @param pShwPage Pointer to the shadow page.
1460	* @param iPTDst The index into the shadow table.
1461	*
1462	* @remark Not used for 2/4MB pages!
1463	*/
1464	DECLINLINE(void) PGM_BTH_NAME(SyncPageWorker)(PVM pVM, PSHWPTE pPteDst, GSTPDE PdeSrc, GSTPTE PteSrc, PPGMPOOLPAGE pShwPage, unsigned iPTDst)
1465	{
1466	if (PteSrc.n.u1Present)
1467	{
1468	/*
1469	* Find the ram range.
1470	*/
1471	PPGMPAGE pPage;
1472	int rc = pgmPhysGetPageEx(&pVM->pgm.s, PteSrc.u & GST_PTE_PG_MASK, &pPage);
1473	if (RT_SUCCESS(rc))
1474	{
1475	/** @todo investiage PWT, PCD and PAT. */
1476	/*
1477	* Make page table entry.
1478	*/
1479	const RTHCPHYS HCPhys = pPage->HCPhys; /** @todo FLAGS */
1480	SHWPTE PteDst;
1481	if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
1482	{
1483	/** @todo r=bird: Are we actually handling dirty and access bits for pages with access handlers correctly? No. */
1484	if (!PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage))
1485	{
1486	#if PGM_SHW_TYPE == PGM_TYPE_EPT
1487	PteDst.u = (HCPhys & EPT_PTE_PG_MASK);
1488	PteDst.n.u1Present = 1;
1489	PteDst.n.u1Execute = 1;
1490	PteDst.n.u1IgnorePAT = 1;
1491	PteDst.n.u3EMT = VMX_EPT_MEMTYPE_WB;
1492	/* PteDst.n.u1Write = 0 && PteDst.n.u1Size = 0 */
1493	#else
1494	PteDst.u = (PteSrc.u & ~(X86_PTE_PAE_PG_MASK \| X86_PTE_AVL_MASK \| X86_PTE_PAT \| X86_PTE_PCD \| X86_PTE_PWT \| X86_PTE_RW))
1495	\| (HCPhys & X86_PTE_PAE_PG_MASK);
1496	#endif
1497	}
1498	else
1499	{
1500	LogFlow(("SyncPageWorker: monitored page (%RHp) -> mark not present\n", HCPhys));
1501	PteDst.u = 0;
1502	}
1503	/** @todo count these two kinds. */
1504	}
1505	else
1506	{
1507	#if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
1508	/*
1509	* If the page or page directory entry is not marked accessed,
1510	* we mark the page not present.
1511	*/
1512	if (!PteSrc.n.u1Accessed \|\| !PdeSrc.n.u1Accessed)
1513	{
1514	LogFlow(("SyncPageWorker: page and or page directory not accessed -> mark not present\n"));
1515	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,AccessedPage));
1516	PteDst.u = 0;
1517	}
1518	else
1519	/*
1520	* If the page is not flagged as dirty and is writable, then make it read-only, so we can set the dirty bit
1521	* when the page is modified.
1522	*/
1523	if (!PteSrc.n.u1Dirty && (PdeSrc.n.u1Write & PteSrc.n.u1Write))
1524	{
1525	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPage));
1526	PteDst.u = (PteSrc.u & ~(X86_PTE_PAE_PG_MASK \| X86_PTE_AVL_MASK \| X86_PTE_PAT \| X86_PTE_PCD \| X86_PTE_PWT \| X86_PTE_RW))
1527	\| (HCPhys & X86_PTE_PAE_PG_MASK)
1528	\| PGM_PTFLAGS_TRACK_DIRTY;
1529	}
1530	else
1531	#endif
1532	{
1533	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPageSkipped));
1534	#if PGM_SHW_TYPE == PGM_TYPE_EPT
1535	PteDst.u = (HCPhys & EPT_PTE_PG_MASK);
1536	PteDst.n.u1Present = 1;
1537	PteDst.n.u1Write = 1;
1538	PteDst.n.u1Execute = 1;
1539	PteDst.n.u1IgnorePAT = 1;
1540	PteDst.n.u3EMT = VMX_EPT_MEMTYPE_WB;
1541	/* PteDst.n.u1Size = 0 */
1542	#else
1543	PteDst.u = (PteSrc.u & ~(X86_PTE_PAE_PG_MASK \| X86_PTE_AVL_MASK \| X86_PTE_PAT \| X86_PTE_PCD \| X86_PTE_PWT))
1544	\| (HCPhys & X86_PTE_PAE_PG_MASK);
1545	#endif
1546	}
1547	}
1548
1549	#ifdef PGMPOOL_WITH_USER_TRACKING
1550	/*
1551	* Keep user track up to date.
1552	*/
1553	if (PteDst.n.u1Present)
1554	{
1555	if (!pPteDst->n.u1Present)
1556	PGM_BTH_NAME(SyncPageWorkerTrackAddref)(pVM, pShwPage, HCPhys >> MM_RAM_FLAGS_IDX_SHIFT, pPage, iPTDst);
1557	else if ((pPteDst->u & SHW_PTE_PG_MASK) != (PteDst.u & SHW_PTE_PG_MASK))
1558	{
1559	Log2(("SyncPageWorker: deref! *pPteDst=%RX64 PteDst=%RX64\n", (uint64_t)pPteDst->u, (uint64_t)PteDst.u));
1560	PGM_BTH_NAME(SyncPageWorkerTrackDeref)(pVM, pShwPage, pPteDst->u & SHW_PTE_PG_MASK);
1561	PGM_BTH_NAME(SyncPageWorkerTrackAddref)(pVM, pShwPage, HCPhys >> MM_RAM_FLAGS_IDX_SHIFT, pPage, iPTDst);
1562	}
1563	}
1564	else if (pPteDst->n.u1Present)
1565	{
1566	Log2(("SyncPageWorker: deref! *pPteDst=%RX64\n", (uint64_t)pPteDst->u));
1567	PGM_BTH_NAME(SyncPageWorkerTrackDeref)(pVM, pShwPage, pPteDst->u & SHW_PTE_PG_MASK);
1568	}
1569	#endif /* PGMPOOL_WITH_USER_TRACKING */
1570
1571	/*
1572	* Update statistics and commit the entry.
1573	*/
1574	if (!PteSrc.n.u1Global)
1575	pShwPage->fSeenNonGlobal = true;
1576	*pPteDst = PteDst;
1577	}
1578	/* else MMIO or invalid page, we must handle them manually in the #PF handler. */
1579	/** @todo count these. */
1580	}
1581	else
1582	{
1583	/*
1584	* Page not-present.
1585	*/
1586	LogFlow(("SyncPageWorker: page not present in Pte\n"));
1587	#ifdef PGMPOOL_WITH_USER_TRACKING
1588	/* Keep user track up to date. */
1589	if (pPteDst->n.u1Present)
1590	{
1591	Log2(("SyncPageWorker: deref! *pPteDst=%RX64\n", (uint64_t)pPteDst->u));
1592	PGM_BTH_NAME(SyncPageWorkerTrackDeref)(pVM, pShwPage, pPteDst->u & SHW_PTE_PG_MASK);
1593	}
1594	#endif /* PGMPOOL_WITH_USER_TRACKING */
1595	pPteDst->u = 0;
1596	/** @todo count these. */
1597	}
1598	}
1599
1600
1601	/**
1602	* Syncs a guest OS page.
1603	*
1604	* There are no conflicts at this point, neither is there any need for
1605	* page table allocations.
1606	*
1607	* @returns VBox status code.
1608	* @returns VINF_PGM_SYNCPAGE_MODIFIED_PDE if it modifies the PDE in any way.
1609	* @param pVM VM handle.
1610	* @param PdeSrc Page directory entry of the guest.
1611	* @param GCPtrPage Guest context page address.
1612	* @param cPages Number of pages to sync (PGM_SYNC_N_PAGES) (default=1).
1613	* @param uErr Fault error (X86_TRAP_PF_*).
1614	*/
1615	PGM_BTH_DECL(int, SyncPage)(PVM pVM, GSTPDE PdeSrc, RTGCPTR GCPtrPage, unsigned cPages, unsigned uErr)
1616	{
1617	LogFlow(("SyncPage: GCPtrPage=%RGv cPages=%u uErr=%#x\n", GCPtrPage, cPages, uErr));
1618
1619	#if ( PGM_GST_TYPE == PGM_TYPE_32BIT \
1620	\|\| PGM_GST_TYPE == PGM_TYPE_PAE \
1621	\|\| PGM_GST_TYPE == PGM_TYPE_AMD64) \
1622	&& PGM_SHW_TYPE != PGM_TYPE_NESTED \
1623	&& PGM_SHW_TYPE != PGM_TYPE_EPT
1624
1625	# if PGM_WITH_NX(PGM_GST_TYPE, PGM_SHW_TYPE)
1626	bool fNoExecuteBitValid = !!(CPUMGetGuestEFER(pVM) & MSR_K6_EFER_NXE);
1627	# endif
1628
1629	/*
1630	* Assert preconditions.
1631	*/
1632	Assert(PdeSrc.n.u1Present);
1633	Assert(cPages);
1634	STAM_COUNTER_INC(&pVM->pgm.s.StatSyncPagePD[(GCPtrPage >> GST_PD_SHIFT) & GST_PD_MASK]);
1635
1636	/*
1637	* Get the shadow PDE, find the shadow page table in the pool.
1638	*/
1639	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
1640	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
1641	PX86PDE pPdeDst = pgmShwGet32BitPDEPtr(&pVM->pgm.s, GCPtrPage);
1642
1643	# ifdef VBOX_WITH_PGMPOOL_PAGING_ONLY
1644	/* Fetch the pgm pool shadow descriptor. */
1645	PPGMPOOLPAGE pShwPde = pVM->pgm.s.CTX_SUFF(pShwPageCR3);
1646	Assert(pShwPde);
1647	# endif
1648
1649	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
1650
1651	# ifdef VBOX_WITH_PGMPOOL_PAGING_ONLY
1652	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
1653	PPGMPOOLPAGE pShwPde;
1654	PX86PDPAE pPDDst;
1655
1656	/* Fetch the pgm pool shadow descriptor. */
1657	int rc = pgmShwGetPaePoolPagePD(&pVM->pgm.s, GCPtrPage, &pShwPde);
1658	AssertRCSuccessReturn(rc, rc);
1659	Assert(pShwPde);
1660
1661	pPDDst = (PX86PDPAE)PGMPOOL_PAGE_2_PTR_BY_PGM(&pVM->pgm.s, pShwPde);
1662	PX86PDEPAE pPdeDst = &pPDDst->a[iPDDst];
1663	# else
1664	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) /& SHW_PD_MASK - only pool index atm! /;
1665	const unsigned iPdpt = (GCPtrPage >> X86_PDPT_SHIFT);
1666	PX86PDPT pPdptDst = pgmShwGetPaePDPTPtr(&pVM->pgm.s); NOREF(pPdptDst);
1667	PX86PDEPAE pPdeDst = pgmShwGetPaePDEPtr(&pVM->pgm.s, GCPtrPage);
1668	AssertReturn(pPdeDst, VERR_INTERNAL_ERROR);
1669	# endif
1670	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
1671	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
1672	const unsigned iPdpt = (GCPtrPage >> X86_PDPT_SHIFT) & X86_PDPT_MASK_AMD64;
1673	PX86PDPAE pPDDst;
1674	PX86PDPT pPdptDst;
1675
1676	int rc = pgmShwGetLongModePDPtr(pVM, GCPtrPage, NULL, &pPdptDst, &pPDDst);
1677	AssertRCSuccessReturn(rc, rc);
1678	Assert(pPDDst && pPdptDst);
1679	PX86PDEPAE pPdeDst = &pPDDst->a[iPDDst];
1680	# endif
1681
1682	SHWPDE PdeDst = *pPdeDst;
1683	AssertMsg(PdeDst.n.u1Present, ("%p=%llx\n", pPdeDst, (uint64_t)PdeDst.u));
1684	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, PdeDst.u & SHW_PDE_PG_MASK);
1685
1686	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1687	/* Fetch the pgm pool shadow descriptor. */
1688	PPGMPOOLPAGE pShwPde = pgmPoolGetPageByHCPhys(pVM, pPdptDst->a[iPdpt].u & X86_PDPE_PG_MASK);
1689	Assert(pShwPde);
1690	# endif
1691
1692	/*
1693	* Check that the page is present and that the shadow PDE isn't out of sync.
1694	*/
1695	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1696	const bool fBigPage = PdeSrc.b.u1Size;
1697	# else
1698	const bool fBigPage = PdeSrc.b.u1Size && (CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
1699	# endif
1700	RTGCPHYS GCPhys;
1701	if (!fBigPage)
1702	{
1703	GCPhys = PdeSrc.u & GST_PDE_PG_MASK;
1704	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
1705	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
1706	GCPhys \|= (iPDDst & 1) * (PAGE_SIZE/2);
1707	# endif
1708	}
1709	else
1710	{
1711	GCPhys = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc);
1712	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
1713	/* Select the right PDE as we're emulating a 4MB page directory with two 2 MB shadow PDEs.*/
1714	GCPhys \|= GCPtrPage & (1 << X86_PD_PAE_SHIFT);
1715	# endif
1716	}
1717	if ( pShwPage->GCPhys == GCPhys
1718	&& PdeSrc.n.u1Present
1719	&& (PdeSrc.n.u1User == PdeDst.n.u1User)
1720	&& (PdeSrc.n.u1Write == PdeDst.n.u1Write \|\| !PdeDst.n.u1Write)
1721	# if PGM_WITH_NX(PGM_GST_TYPE, PGM_SHW_TYPE)
1722	&& (!fNoExecuteBitValid \|\| PdeSrc.n.u1NoExecute == PdeDst.n.u1NoExecute)
1723	# endif
1724	)
1725	{
1726	/*
1727	* Check that the PDE is marked accessed already.
1728	* Since we set the accessed bit before getting here on a #PF, this
1729	* check is only meant for dealing with non-#PF'ing paths.
1730	*/
1731	if (PdeSrc.n.u1Accessed)
1732	{
1733	PSHWPT pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
1734	if (!fBigPage)
1735	{
1736	/*
1737	* 4KB Page - Map the guest page table.
1738	*/
1739	PGSTPT pPTSrc;
1740	int rc = PGM_GCPHYS_2_PTR(pVM, PdeSrc.u & GST_PDE_PG_MASK, &pPTSrc);
1741	if (RT_SUCCESS(rc))
1742	{
1743	# ifdef PGM_SYNC_N_PAGES
1744	Assert(cPages == 1 \|\| !(uErr & X86_TRAP_PF_P));
1745	if (cPages > 1 && !(uErr & X86_TRAP_PF_P))
1746	{
1747	/*
1748	* This code path is currently only taken when the caller is PGMTrap0eHandler
1749	* for non-present pages!
1750	*
1751	* We're setting PGM_SYNC_NR_PAGES pages around the faulting page to sync it and
1752	* deal with locality.
1753	*/
1754	unsigned iPTDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1755	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
1756	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
1757	const unsigned offPTSrc = ((GCPtrPage >> SHW_PD_SHIFT) & 1) * 512;
1758	# else
1759	const unsigned offPTSrc = 0;
1760	# endif
1761	const unsigned iPTDstEnd = RT_MIN(iPTDst + PGM_SYNC_NR_PAGES / 2, RT_ELEMENTS(pPTDst->a));
1762	if (iPTDst < PGM_SYNC_NR_PAGES / 2)
1763	iPTDst = 0;
1764	else
1765	iPTDst -= PGM_SYNC_NR_PAGES / 2;
1766	for (; iPTDst < iPTDstEnd; iPTDst++)
1767	{
1768	if (!pPTDst->a[iPTDst].n.u1Present)
1769	{
1770	GSTPTE PteSrc = pPTSrc->a[offPTSrc + iPTDst];
1771	RTGCPTR GCPtrCurPage = (GCPtrPage & ~(RTGCPTR)(GST_PT_MASK << GST_PT_SHIFT)) \| ((offPTSrc + iPTDst) << PAGE_SHIFT);
1772	NOREF(GCPtrCurPage);
1773	#ifndef IN_RING0
1774	/*
1775	* Assuming kernel code will be marked as supervisor - and not as user level
1776	* and executed using a conforming code selector - And marked as readonly.
1777	* Also assume that if we're monitoring a page, it's of no interest to CSAM.
1778	*/
1779	PPGMPAGE pPage;
1780	if ( ((PdeSrc.u & PteSrc.u) & (X86_PTE_RW \| X86_PTE_US))
1781	\|\| iPTDst == ((GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK) /* always sync GCPtrPage */
1782	\|\| !CSAMDoesPageNeedScanning(pVM, (RTRCPTR)GCPtrCurPage)
1783	\|\| ( (pPage = pgmPhysGetPage(&pVM->pgm.s, PteSrc.u & GST_PTE_PG_MASK))
1784	&& PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
1785	)
1786	#endif /* else: CSAM not active */
1787	PGM_BTH_NAME(SyncPageWorker)(pVM, &pPTDst->a[iPTDst], PdeSrc, PteSrc, pShwPage, iPTDst);
1788	Log2(("SyncPage: 4K+ %RGv PteSrc:{P=%d RW=%d U=%d raw=%08llx} PteDst=%08llx%s\n",
1789	GCPtrCurPage, PteSrc.n.u1Present,
1790	PteSrc.n.u1Write & PdeSrc.n.u1Write,
1791	PteSrc.n.u1User & PdeSrc.n.u1User,
1792	(uint64_t)PteSrc.u,
1793	(uint64_t)pPTDst->a[iPTDst].u,
1794	pPTDst->a[iPTDst].u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
1795	}
1796	}
1797	}
1798	else
1799	# endif /* PGM_SYNC_N_PAGES */
1800	{
1801	const unsigned iPTSrc = (GCPtrPage >> GST_PT_SHIFT) & GST_PT_MASK;
1802	GSTPTE PteSrc = pPTSrc->a[iPTSrc];
1803	const unsigned iPTDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1804	PGM_BTH_NAME(SyncPageWorker)(pVM, &pPTDst->a[iPTDst], PdeSrc, PteSrc, pShwPage, iPTDst);
1805	Log2(("SyncPage: 4K %RGv PteSrc:{P=%d RW=%d U=%d raw=%08llx}%s\n",
1806	GCPtrPage, PteSrc.n.u1Present,
1807	PteSrc.n.u1Write & PdeSrc.n.u1Write,
1808	PteSrc.n.u1User & PdeSrc.n.u1User,
1809	(uint64_t)PteSrc.u,
1810	pPTDst->a[iPTDst].u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
1811	}
1812	}
1813	else /* MMIO or invalid page: emulated in #PF handler. */
1814	{
1815	LogFlow(("PGM_GCPHYS_2_PTR %RGp failed with %Rrc\n", GCPhys, rc));
1816	Assert(!pPTDst->a[(GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK].n.u1Present);
1817	}
1818	}
1819	else
1820	{
1821	/*
1822	* 4/2MB page - lazy syncing shadow 4K pages.
1823	* (There are many causes of getting here, it's no longer only CSAM.)
1824	*/
1825	/* Calculate the GC physical address of this 4KB shadow page. */
1826	RTGCPHYS GCPhys = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc) \| (GCPtrPage & GST_BIG_PAGE_OFFSET_MASK);
1827	/* Find ram range. */
1828	PPGMPAGE pPage;
1829	int rc = pgmPhysGetPageEx(&pVM->pgm.s, GCPhys, &pPage);
1830	if (RT_SUCCESS(rc))
1831	{
1832	/*
1833	* Make shadow PTE entry.
1834	*/
1835	const RTHCPHYS HCPhys = pPage->HCPhys; /** @todo PAGE FLAGS */
1836	SHWPTE PteDst;
1837	PteDst.u = (PdeSrc.u & ~(X86_PTE_PAE_PG_MASK \| X86_PTE_AVL_MASK \| X86_PTE_PAT \| X86_PTE_PCD \| X86_PTE_PWT))
1838	\| (HCPhys & X86_PTE_PAE_PG_MASK);
1839	if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
1840	{
1841	if (!PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage))
1842	PteDst.n.u1Write = 0;
1843	else
1844	PteDst.u = 0;
1845	}
1846	const unsigned iPTDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1847	# ifdef PGMPOOL_WITH_USER_TRACKING
1848	if (PteDst.n.u1Present && !pPTDst->a[iPTDst].n.u1Present)
1849	PGM_BTH_NAME(SyncPageWorkerTrackAddref)(pVM, pShwPage, HCPhys >> MM_RAM_FLAGS_IDX_SHIFT, pPage, iPTDst);
1850	# endif
1851	pPTDst->a[iPTDst] = PteDst;
1852
1853
1854	/*
1855	* If the page is not flagged as dirty and is writable, then make it read-only
1856	* at PD level, so we can set the dirty bit when the page is modified.
1857	*
1858	* ASSUMES that page access handlers are implemented on page table entry level.
1859	* Thus we will first catch the dirty access and set PDE.D and restart. If
1860	* there is an access handler, we'll trap again and let it work on the problem.
1861	*/
1862	/** @todo r=bird: figure out why we need this here, SyncPT should've taken care of this already.
1863	* As for invlpg, it simply frees the whole shadow PT.
1864	* ...It's possibly because the guest clears it and the guest doesn't really tell us... */
1865	if (!PdeSrc.b.u1Dirty && PdeSrc.b.u1Write)
1866	{
1867	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPageBig));
1868	PdeDst.u \|= PGM_PDFLAGS_TRACK_DIRTY;
1869	PdeDst.n.u1Write = 0;
1870	}
1871	else
1872	{
1873	PdeDst.au32[0] &= ~PGM_PDFLAGS_TRACK_DIRTY;
1874	PdeDst.n.u1Write = PdeSrc.n.u1Write;
1875	}
1876	*pPdeDst = PdeDst;
1877	Log2(("SyncPage: BIG %RGv PdeSrc:{P=%d RW=%d U=%d raw=%08llx} GCPhys=%RGp%s\n",
1878	GCPtrPage, PdeSrc.n.u1Present, PdeSrc.n.u1Write, PdeSrc.n.u1User, (uint64_t)PdeSrc.u, GCPhys,
1879	PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
1880	}
1881	else
1882	LogFlow(("PGM_GCPHYS_2_PTR %RGp (big) failed with %Rrc\n", GCPhys, rc));
1883	}
1884	return VINF_SUCCESS;
1885	}
1886	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPagePDNAs));
1887	}
1888	else
1889	{
1890	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPagePDOutOfSync));
1891	Log2(("SyncPage: Out-Of-Sync PDE at %RGp PdeSrc=%RX64 PdeDst=%RX64 (GCPhys %RGp vs %RGp)\n",
1892	GCPtrPage, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u, pShwPage->GCPhys, GCPhys));
1893	}
1894
1895	/*
1896	* Mark the PDE not present. Restart the instruction and let #PF call SyncPT.
1897	* Yea, I'm lazy.
1898	*/
1899	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
1900	# if PGM_GST_TYPE == PGM_TYPE_AMD64 \|\| defined(VBOX_WITH_PGMPOOL_PAGING_ONLY)
1901	pgmPoolFreeByPage(pPool, pShwPage, pShwPde->idx, iPDDst);
1902	# else
1903	pgmPoolFreeByPage(pPool, pShwPage, SHW_POOL_ROOT_IDX, iPDDst);
1904	# endif
1905
1906	pPdeDst->u = 0;
1907	PGM_INVL_GUEST_TLBS();
1908	return VINF_PGM_SYNCPAGE_MODIFIED_PDE;
1909
1910	#elif (PGM_GST_TYPE == PGM_TYPE_REAL \|\| PGM_GST_TYPE == PGM_TYPE_PROT) \
1911	&& PGM_SHW_TYPE != PGM_TYPE_NESTED \
1912	&& (PGM_SHW_TYPE != PGM_TYPE_EPT \|\| PGM_GST_TYPE == PGM_TYPE_PROT)
1913
1914	# ifdef PGM_SYNC_N_PAGES
1915	/*
1916	* Get the shadow PDE, find the shadow page table in the pool.
1917	*/
1918	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
1919	X86PDE PdeDst = pgmShwGet32BitPDE(&pVM->pgm.s, GCPtrPage);
1920
1921	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
1922	X86PDEPAE PdeDst = pgmShwGetPaePDE(&pVM->pgm.s, GCPtrPage);
1923
1924	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
1925	const unsigned iPDDst = ((GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK);
1926	const unsigned iPdpt = (GCPtrPage >> X86_PDPT_SHIFT) & X86_PDPT_MASK_AMD64; NOREF(iPdpt);
1927	PX86PDPAE pPDDst;
1928	X86PDEPAE PdeDst;
1929	PX86PDPT pPdptDst;
1930
1931	int rc = pgmShwGetLongModePDPtr(pVM, GCPtrPage, NULL, &pPdptDst, &pPDDst);
1932	AssertRCSuccessReturn(rc, rc);
1933	Assert(pPDDst && pPdptDst);
1934	PdeDst = pPDDst->a[iPDDst];
1935	# elif PGM_SHW_TYPE == PGM_TYPE_EPT
1936	const unsigned iPDDst = ((GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK);
1937	PEPTPD pPDDst;
1938	EPTPDE PdeDst;
1939
1940	int rc = pgmShwGetEPTPDPtr(pVM, GCPtrPage, NULL, &pPDDst);
1941	if (rc != VINF_SUCCESS)
1942	{
1943	AssertRC(rc);
1944	return rc;
1945	}
1946	Assert(pPDDst);
1947	PdeDst = pPDDst->a[iPDDst];
1948	# endif
1949	AssertMsg(PdeDst.n.u1Present, ("%#llx\n", (uint64_t)PdeDst.u));
1950	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, PdeDst.u & SHW_PDE_PG_MASK);
1951	PSHWPT pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
1952
1953	Assert(cPages == 1 \|\| !(uErr & X86_TRAP_PF_P));
1954	if (cPages > 1 && !(uErr & X86_TRAP_PF_P))
1955	{
1956	/*
1957	* This code path is currently only taken when the caller is PGMTrap0eHandler
1958	* for non-present pages!
1959	*
1960	* We're setting PGM_SYNC_NR_PAGES pages around the faulting page to sync it and
1961	* deal with locality.
1962	*/
1963	unsigned iPTDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1964	const unsigned iPTDstEnd = RT_MIN(iPTDst + PGM_SYNC_NR_PAGES / 2, RT_ELEMENTS(pPTDst->a));
1965	if (iPTDst < PGM_SYNC_NR_PAGES / 2)
1966	iPTDst = 0;
1967	else
1968	iPTDst -= PGM_SYNC_NR_PAGES / 2;
1969	for (; iPTDst < iPTDstEnd; iPTDst++)
1970	{
1971	if (!pPTDst->a[iPTDst].n.u1Present)
1972	{
1973	GSTPTE PteSrc;
1974
1975	RTGCPTR GCPtrCurPage = (GCPtrPage & ~(RTGCPTR)(SHW_PT_MASK << SHW_PT_SHIFT)) \| (iPTDst << PAGE_SHIFT);
1976
1977	/* Fake the page table entry */
1978	PteSrc.u = GCPtrCurPage;
1979	PteSrc.n.u1Present = 1;
1980	PteSrc.n.u1Dirty = 1;
1981	PteSrc.n.u1Accessed = 1;
1982	PteSrc.n.u1Write = 1;
1983	PteSrc.n.u1User = 1;
1984
1985	PGM_BTH_NAME(SyncPageWorker)(pVM, &pPTDst->a[iPTDst], PdeSrc, PteSrc, pShwPage, iPTDst);
1986
1987	Log2(("SyncPage: 4K+ %RGv PteSrc:{P=%d RW=%d U=%d raw=%08llx} PteDst=%08llx%s\n",
1988	GCPtrCurPage, PteSrc.n.u1Present,
1989	PteSrc.n.u1Write & PdeSrc.n.u1Write,
1990	PteSrc.n.u1User & PdeSrc.n.u1User,
1991	(uint64_t)PteSrc.u,
1992	(uint64_t)pPTDst->a[iPTDst].u,
1993	pPTDst->a[iPTDst].u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
1994	}
1995	else
1996	Log4(("%RGv iPTDst=%x pPTDst->a[iPTDst] %RX64\n", (GCPtrPage & ~(RTGCPTR)(SHW_PT_MASK << SHW_PT_SHIFT)) \| (iPTDst << PAGE_SHIFT), iPTDst, pPTDst->a[iPTDst].u));
1997	}
1998	}
1999	else
2000	# endif /* PGM_SYNC_N_PAGES */
2001	{
2002	GSTPTE PteSrc;
2003	const unsigned iPTDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
2004	RTGCPTR GCPtrCurPage = (GCPtrPage & ~(RTGCPTR)(SHW_PT_MASK << SHW_PT_SHIFT)) \| (iPTDst << PAGE_SHIFT);
2005
2006	/* Fake the page table entry */
2007	PteSrc.u = GCPtrCurPage;
2008	PteSrc.n.u1Present = 1;
2009	PteSrc.n.u1Dirty = 1;
2010	PteSrc.n.u1Accessed = 1;
2011	PteSrc.n.u1Write = 1;
2012	PteSrc.n.u1User = 1;
2013	PGM_BTH_NAME(SyncPageWorker)(pVM, &pPTDst->a[iPTDst], PdeSrc, PteSrc, pShwPage, iPTDst);
2014
2015	Log2(("SyncPage: 4K %RGv PteSrc:{P=%d RW=%d U=%d raw=%08llx}PteDst=%08llx%s\n",
2016	GCPtrPage, PteSrc.n.u1Present,
2017	PteSrc.n.u1Write & PdeSrc.n.u1Write,
2018	PteSrc.n.u1User & PdeSrc.n.u1User,
2019	(uint64_t)PteSrc.u,
2020	(uint64_t)pPTDst->a[iPTDst].u,
2021	pPTDst->a[iPTDst].u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
2022	}
2023	return VINF_SUCCESS;
2024
2025	#else
2026	AssertReleaseMsgFailed(("Shw=%d Gst=%d is not implemented!\n", PGM_GST_TYPE, PGM_SHW_TYPE));
2027	return VERR_INTERNAL_ERROR;
2028	#endif
2029	}
2030
2031
2032	#if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
2033	/**
2034	* Investigate page fault and handle write protection page faults caused by
2035	* dirty bit tracking.
2036	*
2037	* @returns VBox status code.
2038	* @param pVM VM handle.
2039	* @param uErr Page fault error code.
2040	* @param pPdeDst Shadow page directory entry.
2041	* @param pPdeSrc Guest page directory entry.
2042	* @param GCPtrPage Guest context page address.
2043	*/
2044	PGM_BTH_DECL(int, CheckPageFault)(PVM pVM, uint32_t uErr, PSHWPDE pPdeDst, PGSTPDE pPdeSrc, RTGCPTR GCPtrPage)
2045	{
2046	bool fWriteProtect = !!(CPUMGetGuestCR0(pVM) & X86_CR0_WP);
2047	bool fUserLevelFault = !!(uErr & X86_TRAP_PF_US);
2048	bool fWriteFault = !!(uErr & X86_TRAP_PF_RW);
2049	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2050	bool fBigPagesSupported = true;
2051	# else
2052	bool fBigPagesSupported = !!(CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
2053	# endif
2054	# if PGM_WITH_NX(PGM_GST_TYPE, PGM_SHW_TYPE)
2055	bool fNoExecuteBitValid = !!(CPUMGetGuestEFER(pVM) & MSR_K6_EFER_NXE);
2056	# endif
2057	unsigned uPageFaultLevel;
2058	int rc;
2059
2060	STAM_PROFILE_START(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2061	LogFlow(("CheckPageFault: GCPtrPage=%RGv uErr=%#x PdeSrc=%08x\n", GCPtrPage, uErr, pPdeSrc->u));
2062
2063	# if PGM_GST_TYPE == PGM_TYPE_PAE \
2064	\|\| PGM_GST_TYPE == PGM_TYPE_AMD64
2065
2066	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2067	PX86PML4E pPml4eSrc;
2068	PX86PDPE pPdpeSrc;
2069
2070	pPdpeSrc = pgmGstGetLongModePDPTPtr(&pVM->pgm.s, GCPtrPage, &pPml4eSrc);
2071	Assert(pPml4eSrc);
2072
2073	/*
2074	* Real page fault? (PML4E level)
2075	*/
2076	if ( (uErr & X86_TRAP_PF_RSVD)
2077	\|\| !pPml4eSrc->n.u1Present
2078	\|\| (fNoExecuteBitValid && (uErr & X86_TRAP_PF_ID) && pPml4eSrc->n.u1NoExecute)
2079	\|\| (fWriteFault && !pPml4eSrc->n.u1Write && (fUserLevelFault \|\| fWriteProtect))
2080	\|\| (fUserLevelFault && !pPml4eSrc->n.u1User)
2081	)
2082	{
2083	uPageFaultLevel = 0;
2084	goto l_UpperLevelPageFault;
2085	}
2086	Assert(pPdpeSrc);
2087
2088	# else /* PAE */
2089	PX86PDPE pPdpeSrc = pgmGstGetPaePDPEPtr(&pVM->pgm.s, GCPtrPage);
2090	# endif /* PAE */
2091
2092	/*
2093	* Real page fault? (PDPE level)
2094	*/
2095	if ( (uErr & X86_TRAP_PF_RSVD)
2096	\|\| !pPdpeSrc->n.u1Present
2097	# if PGM_GST_TYPE == PGM_TYPE_AMD64 /* NX, r/w, u/s bits in the PDPE are long mode only */
2098	\|\| (fNoExecuteBitValid && (uErr & X86_TRAP_PF_ID) && pPdpeSrc->lm.u1NoExecute)
2099	\|\| (fWriteFault && !pPdpeSrc->lm.u1Write && (fUserLevelFault \|\| fWriteProtect))
2100	\|\| (fUserLevelFault && !pPdpeSrc->lm.u1User)
2101	# endif
2102	)
2103	{
2104	uPageFaultLevel = 1;
2105	goto l_UpperLevelPageFault;
2106	}
2107	# endif
2108
2109	/*
2110	* Real page fault? (PDE level)
2111	*/
2112	if ( (uErr & X86_TRAP_PF_RSVD)
2113	\|\| !pPdeSrc->n.u1Present
2114	# if PGM_WITH_NX(PGM_GST_TYPE, PGM_SHW_TYPE)
2115	\|\| (fNoExecuteBitValid && (uErr & X86_TRAP_PF_ID) && pPdeSrc->n.u1NoExecute)
2116	# endif
2117	\|\| (fWriteFault && !pPdeSrc->n.u1Write && (fUserLevelFault \|\| fWriteProtect))
2118	\|\| (fUserLevelFault && !pPdeSrc->n.u1User) )
2119	{
2120	uPageFaultLevel = 2;
2121	goto l_UpperLevelPageFault;
2122	}
2123
2124	/*
2125	* First check the easy case where the page directory has been marked read-only to track
2126	* the dirty bit of an emulated BIG page
2127	*/
2128	if (pPdeSrc->b.u1Size && fBigPagesSupported)
2129	{
2130	/* Mark guest page directory as accessed */
2131	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2132	pPml4eSrc->n.u1Accessed = 1;
2133	pPdpeSrc->lm.u1Accessed = 1;
2134	# endif
2135	pPdeSrc->b.u1Accessed = 1;
2136
2137	/*
2138	* Only write protection page faults are relevant here.
2139	*/
2140	if (fWriteFault)
2141	{
2142	/* Mark guest page directory as dirty (BIG page only). */
2143	pPdeSrc->b.u1Dirty = 1;
2144
2145	if (pPdeDst->n.u1Present && (pPdeDst->u & PGM_PDFLAGS_TRACK_DIRTY))
2146	{
2147	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPageTrap));
2148
2149	Assert(pPdeSrc->b.u1Write);
2150
2151	pPdeDst->n.u1Write = 1;
2152	pPdeDst->n.u1Accessed = 1;
2153	pPdeDst->au32[0] &= ~PGM_PDFLAGS_TRACK_DIRTY;
2154	PGM_INVL_BIG_PG(GCPtrPage);
2155	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2156	return VINF_PGM_HANDLED_DIRTY_BIT_FAULT;
2157	}
2158	}
2159	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2160	return VINF_PGM_NO_DIRTY_BIT_TRACKING;
2161	}
2162	/* else: 4KB page table */
2163
2164	/*
2165	* Map the guest page table.
2166	*/
2167	PGSTPT pPTSrc;
2168	rc = PGM_GCPHYS_2_PTR(pVM, pPdeSrc->u & GST_PDE_PG_MASK, &pPTSrc);
2169	if (RT_SUCCESS(rc))
2170	{
2171	/*
2172	* Real page fault?
2173	*/
2174	PGSTPTE pPteSrc = &pPTSrc->a[(GCPtrPage >> GST_PT_SHIFT) & GST_PT_MASK];
2175	const GSTPTE PteSrc = *pPteSrc;
2176	if ( !PteSrc.n.u1Present
2177	# if PGM_WITH_NX(PGM_GST_TYPE, PGM_SHW_TYPE)
2178	\|\| (fNoExecuteBitValid && (uErr & X86_TRAP_PF_ID) && PteSrc.n.u1NoExecute)
2179	# endif
2180	\|\| (fWriteFault && !PteSrc.n.u1Write && (fUserLevelFault \|\| fWriteProtect))
2181	\|\| (fUserLevelFault && !PteSrc.n.u1User)
2182	)
2183	{
2184	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyTrackRealPF));
2185	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2186	LogFlow(("CheckPageFault: real page fault at %RGv PteSrc.u=%08x (2)\n", GCPtrPage, PteSrc.u));
2187
2188	/* Check the present bit as the shadow tables can cause different error codes by being out of sync.
2189	* See the 2nd case above as well.
2190	*/
2191	if (pPdeSrc->n.u1Present && pPteSrc->n.u1Present)
2192	TRPMSetErrorCode(pVM, uErr \| X86_TRAP_PF_P); /* page-level protection violation */
2193
2194	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2195	return VINF_EM_RAW_GUEST_TRAP;
2196	}
2197	LogFlow(("CheckPageFault: page fault at %RGv PteSrc.u=%08x\n", GCPtrPage, PteSrc.u));
2198
2199	/*
2200	* Set the accessed bits in the page directory and the page table.
2201	*/
2202	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2203	pPml4eSrc->n.u1Accessed = 1;
2204	pPdpeSrc->lm.u1Accessed = 1;
2205	# endif
2206	pPdeSrc->n.u1Accessed = 1;
2207	pPteSrc->n.u1Accessed = 1;
2208
2209	/*
2210	* Only write protection page faults are relevant here.
2211	*/
2212	if (fWriteFault)
2213	{
2214	/* Write access, so mark guest entry as dirty. */
2215	# ifdef VBOX_WITH_STATISTICS
2216	if (!pPteSrc->n.u1Dirty)
2217	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtiedPage));
2218	else
2219	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,PageAlreadyDirty));
2220	# endif
2221
2222	pPteSrc->n.u1Dirty = 1;
2223
2224	if (pPdeDst->n.u1Present)
2225	{
2226	#ifndef IN_RING0
2227	/* Bail out here as pgmPoolGetPageByHCPhys will return NULL and we'll crash below.
2228	* Our individual shadow handlers will provide more information and force a fatal exit.
2229	*/
2230	if (MMHyperIsInsideArea(pVM, (RTGCPTR)GCPtrPage))
2231	{
2232	LogRel(("CheckPageFault: write to hypervisor region %RGv\n", GCPtrPage));
2233	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2234	return VINF_SUCCESS;
2235	}
2236	#endif
2237	/*
2238	* Map shadow page table.
2239	*/
2240	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, pPdeDst->u & SHW_PDE_PG_MASK);
2241	if (pShwPage)
2242	{
2243	PSHWPT pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
2244	PSHWPTE pPteDst = &pPTDst->a[(GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK];
2245	if ( pPteDst->n.u1Present /** @todo Optimize accessed bit emulation? */
2246	&& (pPteDst->u & PGM_PTFLAGS_TRACK_DIRTY))
2247	{
2248	LogFlow(("DIRTY page trap addr=%RGv\n", GCPtrPage));
2249	# ifdef VBOX_STRICT
2250	PPGMPAGE pPage = pgmPhysGetPage(&pVM->pgm.s, pPteSrc->u & GST_PTE_PG_MASK);
2251	if (pPage)
2252	AssertMsg(!PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage),
2253	("Unexpected dirty bit tracking on monitored page %RGv (phys %RGp)!!!!!!\n", GCPtrPage, pPteSrc->u & X86_PTE_PAE_PG_MASK));
2254	# endif
2255	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPageTrap));
2256
2257	Assert(pPteSrc->n.u1Write);
2258
2259	pPteDst->n.u1Write = 1;
2260	pPteDst->n.u1Dirty = 1;
2261	pPteDst->n.u1Accessed = 1;
2262	pPteDst->au32[0] &= ~PGM_PTFLAGS_TRACK_DIRTY;
2263	PGM_INVL_PG(GCPtrPage);
2264
2265	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2266	return VINF_PGM_HANDLED_DIRTY_BIT_FAULT;
2267	}
2268	}
2269	else
2270	AssertMsgFailed(("pgmPoolGetPageByHCPhys %RGp failed!\n", pPdeDst->u & SHW_PDE_PG_MASK));
2271	}
2272	}
2273	/** @todo Optimize accessed bit emulation? */
2274	# ifdef VBOX_STRICT
2275	/*
2276	* Sanity check.
2277	*/
2278	else if ( !pPteSrc->n.u1Dirty
2279	&& (pPdeSrc->n.u1Write & pPteSrc->n.u1Write)
2280	&& pPdeDst->n.u1Present)
2281	{
2282	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, pPdeDst->u & SHW_PDE_PG_MASK);
2283	PSHWPT pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
2284	PSHWPTE pPteDst = &pPTDst->a[(GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK];
2285	if ( pPteDst->n.u1Present
2286	&& pPteDst->n.u1Write)
2287	LogFlow(("Writable present page %RGv not marked for dirty bit tracking!!!\n", GCPtrPage));
2288	}
2289	# endif /* VBOX_STRICT */
2290	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2291	return VINF_PGM_NO_DIRTY_BIT_TRACKING;
2292	}
2293	AssertRC(rc);
2294	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2295	return rc;
2296
2297
2298	l_UpperLevelPageFault:
2299	/*
2300	* Pagefault detected while checking the PML4E, PDPE or PDE.
2301	* Single exit handler to get rid of duplicate code paths.
2302	*/
2303	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyTrackRealPF));
2304	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2305	Log(("CheckPageFault: real page fault at %RGv (%d)\n", GCPtrPage, uPageFaultLevel));
2306
2307	if (
2308	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2309	pPml4eSrc->n.u1Present &&
2310	# endif
2311	# if PGM_GST_TYPE == PGM_TYPE_AMD64 \|\| PGM_GST_TYPE == PGM_TYPE_PAE
2312	pPdpeSrc->n.u1Present &&
2313	# endif
2314	pPdeSrc->n.u1Present)
2315	{
2316	/* Check the present bit as the shadow tables can cause different error codes by being out of sync. */
2317	if (pPdeSrc->b.u1Size && fBigPagesSupported)
2318	{
2319	TRPMSetErrorCode(pVM, uErr \| X86_TRAP_PF_P); /* page-level protection violation */
2320	}
2321	else
2322	{
2323	/*
2324	* Map the guest page table.
2325	*/
2326	PGSTPT pPTSrc;
2327	rc = PGM_GCPHYS_2_PTR(pVM, pPdeSrc->u & GST_PDE_PG_MASK, &pPTSrc);
2328	if (RT_SUCCESS(rc))
2329	{
2330	PGSTPTE pPteSrc = &pPTSrc->a[(GCPtrPage >> GST_PT_SHIFT) & GST_PT_MASK];
2331	const GSTPTE PteSrc = *pPteSrc;
2332	if (pPteSrc->n.u1Present)
2333	TRPMSetErrorCode(pVM, uErr \| X86_TRAP_PF_P); /* page-level protection violation */
2334	}
2335	AssertRC(rc);
2336	}
2337	}
2338	return VINF_EM_RAW_GUEST_TRAP;
2339	}
2340	#endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
2341
2342
2343	/**
2344	* Sync a shadow page table.
2345	*
2346	* The shadow page table is not present. This includes the case where
2347	* there is a conflict with a mapping.
2348	*
2349	* @returns VBox status code.
2350	* @param pVM VM handle.
2351	* @param iPD Page directory index.
2352	* @param pPDSrc Source page directory (i.e. Guest OS page directory).
2353	* Assume this is a temporary mapping.
2354	* @param GCPtrPage GC Pointer of the page that caused the fault
2355	*/
2356	PGM_BTH_DECL(int, SyncPT)(PVM pVM, unsigned iPDSrc, PGSTPD pPDSrc, RTGCPTR GCPtrPage)
2357	{
2358	STAM_PROFILE_START(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT), a);
2359	STAM_COUNTER_INC(&pVM->pgm.s.StatSyncPtPD[iPDSrc]);
2360	LogFlow(("SyncPT: GCPtrPage=%RGv\n", GCPtrPage));
2361
2362	#if ( PGM_GST_TYPE == PGM_TYPE_32BIT \
2363	\|\| PGM_GST_TYPE == PGM_TYPE_PAE \
2364	\|\| PGM_GST_TYPE == PGM_TYPE_AMD64) \
2365	&& PGM_SHW_TYPE != PGM_TYPE_NESTED \
2366	&& PGM_SHW_TYPE != PGM_TYPE_EPT
2367
2368	int rc = VINF_SUCCESS;
2369
2370	/*
2371	* Validate input a little bit.
2372	*/
2373	AssertMsg(iPDSrc == ((GCPtrPage >> GST_PD_SHIFT) & GST_PD_MASK), ("iPDSrc=%x GCPtrPage=%RGv\n", iPDSrc, GCPtrPage));
2374	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
2375	const unsigned iPDDst = GCPtrPage >> SHW_PD_SHIFT;
2376	PSHWPDE pPdeDst = pgmShwGet32BitPDEPtr(&pVM->pgm.s, GCPtrPage);
2377
2378	# ifdef VBOX_WITH_PGMPOOL_PAGING_ONLY
2379	/* Fetch the pgm pool shadow descriptor. */
2380	PPGMPOOLPAGE pShwPde = pVM->pgm.s.CTX_SUFF(pShwPageCR3);
2381	Assert(pShwPde);
2382	# endif
2383
2384	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
2385	# ifdef VBOX_WITH_PGMPOOL_PAGING_ONLY
2386	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
2387	PPGMPOOLPAGE pShwPde;
2388	PX86PDPAE pPDDst;
2389	PSHWPDE pPdeDst;
2390
2391	/* Fetch the pgm pool shadow descriptor. */
2392	rc = pgmShwGetPaePoolPagePD(&pVM->pgm.s, GCPtrPage, &pShwPde);
2393	AssertRCSuccessReturn(rc, rc);
2394	Assert(pShwPde);
2395
2396	pPDDst = (PX86PDPAE)PGMPOOL_PAGE_2_PTR_BY_PGM(&pVM->pgm.s, pShwPde);
2397	pPdeDst = &pPDDst->a[iPDDst];
2398	# else
2399	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) /& SHW_PD_MASK - only pool index atm! /;
2400	const unsigned iPdpt = (GCPtrPage >> X86_PDPT_SHIFT); NOREF(iPdpt);
2401	PX86PDPT pPdptDst = pgmShwGetPaePDPTPtr(&pVM->pgm.s); NOREF(pPdptDst);
2402	PSHWPDE pPdeDst = pgmShwGetPaePDEPtr(&pVM->pgm.s, GCPtrPage);
2403	# endif
2404	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
2405	const unsigned iPdpt = (GCPtrPage >> X86_PDPT_SHIFT) & X86_PDPT_MASK_AMD64;
2406	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
2407	PX86PDPAE pPDDst;
2408	PX86PDPT pPdptDst;
2409	rc = pgmShwGetLongModePDPtr(pVM, GCPtrPage, NULL, &pPdptDst, &pPDDst);
2410	AssertRCSuccessReturn(rc, rc);
2411	Assert(pPDDst);
2412	PSHWPDE pPdeDst = &pPDDst->a[iPDDst];
2413	# endif
2414	SHWPDE PdeDst = *pPdeDst;
2415
2416	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2417	/* Fetch the pgm pool shadow descriptor. */
2418	PPGMPOOLPAGE pShwPde = pgmPoolGetPageByHCPhys(pVM, pPdptDst->a[iPdpt].u & X86_PDPE_PG_MASK);
2419	Assert(pShwPde);
2420	# endif
2421
2422	# ifndef PGM_WITHOUT_MAPPINGS
2423	/*
2424	* Check for conflicts.
2425	* GC: In case of a conflict we'll go to Ring-3 and do a full SyncCR3.
2426	* HC: Simply resolve the conflict.
2427	*/
2428	if (PdeDst.u & PGM_PDFLAGS_MAPPING)
2429	{
2430	Assert(pgmMapAreMappingsEnabled(&pVM->pgm.s));
2431	# ifndef IN_RING3
2432	Log(("SyncPT: Conflict at %RGv\n", GCPtrPage));
2433	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT), a);
2434	return VERR_ADDRESS_CONFLICT;
2435	# else
2436	PPGMMAPPING pMapping = pgmGetMapping(pVM, (RTGCPTR)GCPtrPage);
2437	Assert(pMapping);
2438	# if PGM_GST_TYPE == PGM_TYPE_32BIT
2439	int rc = pgmR3SyncPTResolveConflict(pVM, pMapping, pPDSrc, GCPtrPage & (GST_PD_MASK << GST_PD_SHIFT));
2440	# elif PGM_GST_TYPE == PGM_TYPE_PAE
2441	int rc = pgmR3SyncPTResolveConflictPAE(pVM, pMapping, GCPtrPage & (GST_PD_MASK << GST_PD_SHIFT));
2442	# else
2443	AssertFailed(); /* can't happen for amd64 */
2444	# endif
2445	if (RT_FAILURE(rc))
2446	{
2447	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT), a);
2448	return rc;
2449	}
2450	PdeDst = *pPdeDst;
2451	# endif
2452	}
2453	# else /* PGM_WITHOUT_MAPPINGS */
2454	Assert(!pgmMapAreMappingsEnabled(&pVM->pgm.s));
2455	# endif /* PGM_WITHOUT_MAPPINGS */
2456	Assert(!PdeDst.n.u1Present); /* We're only supposed to call SyncPT on PDE!P and conflicts.*/
2457
2458	# if defined(IN_RC) && defined(VBOX_WITH_PGMPOOL_PAGING_ONLY)
2459	/* Make sure the dynamic pPdeDst mapping will not be reused during this function. */
2460	PGMDynLockHCPage(pVM, (uint8_t *)pPdeDst);
2461	# endif
2462
2463	/*
2464	* Sync page directory entry.
2465	*/
2466	GSTPDE PdeSrc = pPDSrc->a[iPDSrc];
2467	if (PdeSrc.n.u1Present)
2468	{
2469	/*
2470	* Allocate & map the page table.
2471	*/
2472	PSHWPT pPTDst;
2473	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2474	const bool fPageTable = !PdeSrc.b.u1Size;
2475	# else
2476	const bool fPageTable = !PdeSrc.b.u1Size \|\| !(CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
2477	# endif
2478	PPGMPOOLPAGE pShwPage;
2479	RTGCPHYS GCPhys;
2480	if (fPageTable)
2481	{
2482	GCPhys = PdeSrc.u & GST_PDE_PG_MASK;
2483	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
2484	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
2485	GCPhys \|= (iPDDst & 1) * (PAGE_SIZE / 2);
2486	# endif
2487	# if PGM_GST_TYPE == PGM_TYPE_AMD64 \|\| defined(VBOX_WITH_PGMPOOL_PAGING_ONLY)
2488	rc = pgmPoolAlloc(pVM, GCPhys, BTH_PGMPOOLKIND_PT_FOR_PT, pShwPde->idx, iPDDst, &pShwPage);
2489	# else
2490	rc = pgmPoolAlloc(pVM, GCPhys, BTH_PGMPOOLKIND_PT_FOR_PT, SHW_POOL_ROOT_IDX, iPDDst, &pShwPage);
2491	# endif
2492	}
2493	else
2494	{
2495	GCPhys = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc);
2496	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
2497	/* Select the right PDE as we're emulating a 4MB page directory with two 2 MB shadow PDEs.*/
2498	GCPhys \|= GCPtrPage & (1 << X86_PD_PAE_SHIFT);
2499	# endif
2500	# if PGM_GST_TYPE == PGM_TYPE_AMD64 \|\| defined(VBOX_WITH_PGMPOOL_PAGING_ONLY)
2501	rc = pgmPoolAlloc(pVM, GCPhys, BTH_PGMPOOLKIND_PT_FOR_BIG, pShwPde->idx, iPDDst, &pShwPage);
2502	# else
2503	rc = pgmPoolAlloc(pVM, GCPhys, BTH_PGMPOOLKIND_PT_FOR_BIG, SHW_POOL_ROOT_IDX, iPDDst, &pShwPage);
2504	# endif
2505	}
2506	if (rc == VINF_SUCCESS)
2507	pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
2508	else if (rc == VINF_PGM_CACHED_PAGE)
2509	{
2510	/*
2511	* The PT was cached, just hook it up.
2512	*/
2513	if (fPageTable)
2514	PdeDst.u = pShwPage->Core.Key
2515	\| (PdeSrc.u & ~(GST_PDE_PG_MASK \| X86_PDE_AVL_MASK \| X86_PDE_PCD \| X86_PDE_PWT \| X86_PDE_PS \| X86_PDE4M_G \| X86_PDE4M_D));
2516	else
2517	{
2518	PdeDst.u = pShwPage->Core.Key
2519	\| (PdeSrc.u & ~(GST_PDE_PG_MASK \| X86_PDE_AVL_MASK \| X86_PDE_PCD \| X86_PDE_PWT \| X86_PDE_PS \| X86_PDE4M_G \| X86_PDE4M_D));
2520	/* (see explanation and assumptions further down.) */
2521	if (!PdeSrc.b.u1Dirty && PdeSrc.b.u1Write)
2522	{
2523	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPageBig));
2524	PdeDst.u \|= PGM_PDFLAGS_TRACK_DIRTY;
2525	PdeDst.b.u1Write = 0;
2526	}
2527	}
2528	*pPdeDst = PdeDst;
2529	# if defined(IN_RC) && defined(VBOX_WITH_PGMPOOL_PAGING_ONLY)
2530	PGMDynUnlockHCPage(pVM, (uint8_t *)pPdeDst);
2531	# endif
2532	return VINF_SUCCESS;
2533	}
2534	else if (rc == VERR_PGM_POOL_FLUSHED)
2535	{
2536	VM_FF_SET(pVM, VM_FF_PGM_SYNC_CR3);
2537	# if defined(IN_RC) && defined(VBOX_WITH_PGMPOOL_PAGING_ONLY)
2538	PGMDynUnlockHCPage(pVM, (uint8_t *)pPdeDst);
2539	# endif
2540	return VINF_PGM_SYNC_CR3;
2541	}
2542	else
2543	AssertMsgFailedReturn(("rc=%Rrc\n", rc), VERR_INTERNAL_ERROR);
2544	PdeDst.u &= X86_PDE_AVL_MASK;
2545	PdeDst.u \|= pShwPage->Core.Key;
2546
2547	/*
2548	* Page directory has been accessed (this is a fault situation, remember).
2549	*/
2550	pPDSrc->a[iPDSrc].n.u1Accessed = 1;
2551	if (fPageTable)
2552	{
2553	/*
2554	* Page table - 4KB.
2555	*
2556	* Sync all or just a few entries depending on PGM_SYNC_N_PAGES.
2557	*/
2558	Log2(("SyncPT: 4K %RGv PdeSrc:{P=%d RW=%d U=%d raw=%08llx}\n",
2559	GCPtrPage, PdeSrc.b.u1Present, PdeSrc.b.u1Write, PdeSrc.b.u1User, (uint64_t)PdeSrc.u));
2560	PGSTPT pPTSrc;
2561	rc = PGM_GCPHYS_2_PTR(pVM, PdeSrc.u & GST_PDE_PG_MASK, &pPTSrc);
2562	if (RT_SUCCESS(rc))
2563	{
2564	/*
2565	* Start by syncing the page directory entry so CSAM's TLB trick works.
2566	*/
2567	PdeDst.u = (PdeDst.u & (SHW_PDE_PG_MASK \| X86_PDE_AVL_MASK))
2568	\| (PdeSrc.u & ~(GST_PDE_PG_MASK \| X86_PDE_AVL_MASK \| X86_PDE_PCD \| X86_PDE_PWT \| X86_PDE_PS \| X86_PDE4M_G \| X86_PDE4M_D));
2569	*pPdeDst = PdeDst;
2570	# if defined(IN_RC) && defined(VBOX_WITH_PGMPOOL_PAGING_ONLY)
2571	PGMDynUnlockHCPage(pVM, (uint8_t *)pPdeDst);
2572	# endif
2573
2574	/*
2575	* Directory/page user or supervisor privilege: (same goes for read/write)
2576	*
2577	* Directory Page Combined
2578	* U/S U/S U/S
2579	* 0 0 0
2580	* 0 1 0
2581	* 1 0 0
2582	* 1 1 1
2583	*
2584	* Simple AND operation. Table listed for completeness.
2585	*
2586	*/
2587	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT4K));
2588	# ifdef PGM_SYNC_N_PAGES
2589	unsigned iPTBase = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
2590	unsigned iPTDst = iPTBase;
2591	const unsigned iPTDstEnd = RT_MIN(iPTDst + PGM_SYNC_NR_PAGES / 2, RT_ELEMENTS(pPTDst->a));
2592	if (iPTDst <= PGM_SYNC_NR_PAGES / 2)
2593	iPTDst = 0;
2594	else
2595	iPTDst -= PGM_SYNC_NR_PAGES / 2;
2596	# else /* !PGM_SYNC_N_PAGES */
2597	unsigned iPTDst = 0;
2598	const unsigned iPTDstEnd = RT_ELEMENTS(pPTDst->a);
2599	# endif /* !PGM_SYNC_N_PAGES */
2600	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
2601	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
2602	const unsigned offPTSrc = ((GCPtrPage >> SHW_PD_SHIFT) & 1) * 512;
2603	# else
2604	const unsigned offPTSrc = 0;
2605	# endif
2606	for (; iPTDst < iPTDstEnd; iPTDst++)
2607	{
2608	const unsigned iPTSrc = iPTDst + offPTSrc;
2609	const GSTPTE PteSrc = pPTSrc->a[iPTSrc];
2610
2611	if (PteSrc.n.u1Present) /* we've already cleared it above */
2612	{
2613	# ifndef IN_RING0
2614	/*
2615	* Assuming kernel code will be marked as supervisor - and not as user level
2616	* and executed using a conforming code selector - And marked as readonly.
2617	* Also assume that if we're monitoring a page, it's of no interest to CSAM.
2618	*/
2619	PPGMPAGE pPage;
2620	if ( ((PdeSrc.u & pPTSrc->a[iPTSrc].u) & (X86_PTE_RW \| X86_PTE_US))
2621	\|\| !CSAMDoesPageNeedScanning(pVM, (RTRCPTR)((iPDSrc << GST_PD_SHIFT) \| (iPTSrc << PAGE_SHIFT)))
2622	\|\| ( (pPage = pgmPhysGetPage(&pVM->pgm.s, PteSrc.u & GST_PTE_PG_MASK))
2623	&& PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
2624	)
2625	# endif
2626	PGM_BTH_NAME(SyncPageWorker)(pVM, &pPTDst->a[iPTDst], PdeSrc, PteSrc, pShwPage, iPTDst);
2627	Log2(("SyncPT: 4K+ %RGv PteSrc:{P=%d RW=%d U=%d raw=%08llx}%s dst.raw=%08llx iPTSrc=%x PdeSrc.u=%x physpte=%RGp\n",
2628	(RTGCPTR)((iPDSrc << GST_PD_SHIFT) \| (iPTSrc << PAGE_SHIFT)),
2629	PteSrc.n.u1Present,
2630	PteSrc.n.u1Write & PdeSrc.n.u1Write,
2631	PteSrc.n.u1User & PdeSrc.n.u1User,
2632	(uint64_t)PteSrc.u,
2633	pPTDst->a[iPTDst].u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : "", pPTDst->a[iPTDst].u, iPTSrc, PdeSrc.au32[0],
2634	(RTGCPHYS)((PdeSrc.u & GST_PDE_PG_MASK) + iPTSrc*sizeof(PteSrc)) ));
2635	}
2636	} /* for PTEs */
2637	}
2638	}
2639	else
2640	{
2641	/*
2642	* Big page - 2/4MB.
2643	*
2644	* We'll walk the ram range list in parallel and optimize lookups.
2645	* We will only sync on shadow page table at a time.
2646	*/
2647	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT4M));
2648
2649	/**
2650	* @todo It might be more efficient to sync only a part of the 4MB page (similar to what we do for 4kb PDs).
2651	*/
2652
2653	/*
2654	* Start by syncing the page directory entry.
2655	*/
2656	PdeDst.u = (PdeDst.u & (SHW_PDE_PG_MASK \| (X86_PDE_AVL_MASK & ~PGM_PDFLAGS_TRACK_DIRTY)))
2657	\| (PdeSrc.u & ~(GST_PDE_PG_MASK \| X86_PDE_AVL_MASK \| X86_PDE_PCD \| X86_PDE_PWT \| X86_PDE_PS \| X86_PDE4M_G \| X86_PDE4M_D));
2658
2659	/*
2660	* If the page is not flagged as dirty and is writable, then make it read-only
2661	* at PD level, so we can set the dirty bit when the page is modified.
2662	*
2663	* ASSUMES that page access handlers are implemented on page table entry level.
2664	* Thus we will first catch the dirty access and set PDE.D and restart. If
2665	* there is an access handler, we'll trap again and let it work on the problem.
2666	*/
2667	/** @todo move the above stuff to a section in the PGM documentation. */
2668	Assert(!(PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY));
2669	if (!PdeSrc.b.u1Dirty && PdeSrc.b.u1Write)
2670	{
2671	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPageBig));
2672	PdeDst.u \|= PGM_PDFLAGS_TRACK_DIRTY;
2673	PdeDst.b.u1Write = 0;
2674	}
2675	*pPdeDst = PdeDst;
2676	# if defined(IN_RC) && defined(VBOX_WITH_PGMPOOL_PAGING_ONLY)
2677	PGMDynUnlockHCPage(pVM, (uint8_t *)pPdeDst);
2678	# endif
2679
2680	/*
2681	* Fill the shadow page table.
2682	*/
2683	/* Get address and flags from the source PDE. */
2684	SHWPTE PteDstBase;
2685	PteDstBase.u = PdeSrc.u & ~(GST_PDE_PG_MASK \| X86_PTE_AVL_MASK \| X86_PTE_PAT \| X86_PTE_PCD \| X86_PTE_PWT);
2686
2687	/* Loop thru the entries in the shadow PT. */
2688	const RTGCPTR GCPtr = (GCPtrPage >> SHW_PD_SHIFT) << SHW_PD_SHIFT; NOREF(GCPtr);
2689	Log2(("SyncPT: BIG %RGv PdeSrc:{P=%d RW=%d U=%d raw=%08llx} Shw=%RGv GCPhys=%RGp %s\n",
2690	GCPtrPage, PdeSrc.b.u1Present, PdeSrc.b.u1Write, PdeSrc.b.u1User, (uint64_t)PdeSrc.u, GCPtr,
2691	GCPhys, PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
2692	PPGMRAMRANGE pRam = pVM->pgm.s.CTX_SUFF(pRamRanges);
2693	unsigned iPTDst = 0;
2694	while (iPTDst < RT_ELEMENTS(pPTDst->a))
2695	{
2696	/* Advance ram range list. */
2697	while (pRam && GCPhys > pRam->GCPhysLast)
2698	pRam = pRam->CTX_SUFF(pNext);
2699	if (pRam && GCPhys >= pRam->GCPhys)
2700	{
2701	unsigned iHCPage = (GCPhys - pRam->GCPhys) >> PAGE_SHIFT;
2702	do
2703	{
2704	/* Make shadow PTE. */
2705	PPGMPAGE pPage = &pRam->aPages[iHCPage];
2706	SHWPTE PteDst;
2707
2708	/* Make sure the RAM has already been allocated. */
2709	if (pRam->fFlags & MM_RAM_FLAGS_DYNAMIC_ALLOC) /** @todo PAGE FLAGS */
2710	{
2711	if (RT_UNLIKELY(!PGM_PAGE_GET_HCPHYS(pPage)))
2712	{
2713	# ifdef IN_RING3
2714	int rc = pgmr3PhysGrowRange(pVM, GCPhys);
2715	# else
2716	int rc = CTXALLMID(VMM, CallHost)(pVM, VMMCALLHOST_PGM_RAM_GROW_RANGE, GCPhys);
2717	# endif
2718	if (rc != VINF_SUCCESS)
2719	return rc;
2720	}
2721	}
2722
2723	if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
2724	{
2725	if (!PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage))
2726	{
2727	PteDst.u = PGM_PAGE_GET_HCPHYS(pPage) \| PteDstBase.u;
2728	PteDst.n.u1Write = 0;
2729	}
2730	else
2731	PteDst.u = 0;
2732	}
2733	# ifndef IN_RING0
2734	/*
2735	* Assuming kernel code will be marked as supervisor and not as user level and executed
2736	* using a conforming code selector. Don't check for readonly, as that implies the whole
2737	* 4MB can be code or readonly data. Linux enables write access for its large pages.
2738	*/
2739	else if ( !PdeSrc.n.u1User
2740	&& CSAMDoesPageNeedScanning(pVM, (RTRCPTR)(GCPtr \| (iPTDst << SHW_PT_SHIFT))))
2741	PteDst.u = 0;
2742	# endif
2743	else
2744	PteDst.u = PGM_PAGE_GET_HCPHYS(pPage) \| PteDstBase.u;
2745	# ifdef PGMPOOL_WITH_USER_TRACKING
2746	if (PteDst.n.u1Present)
2747	PGM_BTH_NAME(SyncPageWorkerTrackAddref)(pVM, pShwPage, pPage->HCPhys >> MM_RAM_FLAGS_IDX_SHIFT, pPage, iPTDst); /** @todo PAGE FLAGS */
2748	# endif
2749	/* commit it */
2750	pPTDst->a[iPTDst] = PteDst;
2751	Log4(("SyncPT: BIG %RGv PteDst:{P=%d RW=%d U=%d raw=%08llx}%s\n",
2752	(RTGCPTR)(GCPtr \| (iPTDst << SHW_PT_SHIFT)), PteDst.n.u1Present, PteDst.n.u1Write, PteDst.n.u1User, (uint64_t)PteDst.u,
2753	PteDst.u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
2754
2755	/* advance */
2756	GCPhys += PAGE_SIZE;
2757	iHCPage++;
2758	iPTDst++;
2759	} while ( iPTDst < RT_ELEMENTS(pPTDst->a)
2760	&& GCPhys <= pRam->GCPhysLast);
2761	}
2762	else if (pRam)
2763	{
2764	Log(("Invalid pages at %RGp\n", GCPhys));
2765	do
2766	{
2767	pPTDst->a[iPTDst].u = 0; /* MMIO or invalid page, we must handle them manually. */
2768	GCPhys += PAGE_SIZE;
2769	iPTDst++;
2770	} while ( iPTDst < RT_ELEMENTS(pPTDst->a)
2771	&& GCPhys < pRam->GCPhys);
2772	}
2773	else
2774	{
2775	Log(("Invalid pages at %RGp (2)\n", GCPhys));
2776	for ( ; iPTDst < RT_ELEMENTS(pPTDst->a); iPTDst++)
2777	pPTDst->a[iPTDst].u = 0; /* MMIO or invalid page, we must handle them manually. */
2778	}
2779	} /* while more PTEs */
2780	} /* 4KB / 4MB */
2781	}
2782	else
2783	AssertRelease(!PdeDst.n.u1Present);
2784
2785	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT), a);
2786	if (RT_FAILURE(rc))
2787	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPTFailed));
2788	return rc;
2789
2790	#elif (PGM_GST_TYPE == PGM_TYPE_REAL \|\| PGM_GST_TYPE == PGM_TYPE_PROT) \
2791	&& PGM_SHW_TYPE != PGM_TYPE_NESTED \
2792	&& (PGM_SHW_TYPE != PGM_TYPE_EPT \|\| PGM_GST_TYPE == PGM_TYPE_PROT)
2793
2794
2795	/*
2796	* Validate input a little bit.
2797	*/
2798	int rc = VINF_SUCCESS;
2799	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
2800	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
2801	PSHWPDE pPdeDst = pgmShwGet32BitPDEPtr(&pVM->pgm.s, GCPtrPage);
2802
2803	# ifdef VBOX_WITH_PGMPOOL_PAGING_ONLY
2804	/* Fetch the pgm pool shadow descriptor. */
2805	PPGMPOOLPAGE pShwPde = pVM->pgm.s.CTX_SUFF(pShwPageCR3);
2806	Assert(pShwPde);
2807	# endif
2808
2809	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
2810	# ifdef VBOX_WITH_PGMPOOL_PAGING_ONLY
2811	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
2812	PPGMPOOLPAGE pShwPde;
2813	PX86PDPAE pPDDst;
2814	PSHWPDE pPdeDst;
2815
2816	/* Fetch the pgm pool shadow descriptor. */
2817	rc = pgmShwGetPaePoolPagePD(&pVM->pgm.s, GCPtrPage, &pShwPde);
2818	AssertRCSuccessReturn(rc, rc);
2819	Assert(pShwPde);
2820
2821	pPDDst = (PX86PDPAE)PGMPOOL_PAGE_2_PTR_BY_PGM(&pVM->pgm.s, pShwPde);
2822	pPdeDst = &pPDDst->a[iPDDst];
2823	# else
2824	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) /& SHW_PD_MASK - only pool index atm!/;
2825	PX86PDEPAE pPdeDst = pgmShwGetPaePDEPtr(&pVM->pgm.s, GCPtrPage);
2826	# endif
2827
2828	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
2829	const unsigned iPdpt = (GCPtrPage >> X86_PDPT_SHIFT) & X86_PDPT_MASK_AMD64;
2830	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
2831	PX86PDPAE pPDDst;
2832	PX86PDPT pPdptDst;
2833	rc = pgmShwGetLongModePDPtr(pVM, GCPtrPage, NULL, &pPdptDst, &pPDDst);
2834	AssertRCSuccessReturn(rc, rc);
2835	Assert(pPDDst);
2836	PSHWPDE pPdeDst = &pPDDst->a[iPDDst];
2837
2838	/* Fetch the pgm pool shadow descriptor. */
2839	PPGMPOOLPAGE pShwPde = pgmPoolGetPageByHCPhys(pVM, pPdptDst->a[iPdpt].u & X86_PDPE_PG_MASK);
2840	Assert(pShwPde);
2841
2842	# elif PGM_SHW_TYPE == PGM_TYPE_EPT
2843	const unsigned iPdpt = (GCPtrPage >> EPT_PDPT_SHIFT) & EPT_PDPT_MASK;
2844	const unsigned iPDDst = ((GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK);
2845	PEPTPD pPDDst;
2846	PEPTPDPT pPdptDst;
2847
2848	rc = pgmShwGetEPTPDPtr(pVM, GCPtrPage, &pPdptDst, &pPDDst);
2849	if (rc != VINF_SUCCESS)
2850	{
2851	AssertRC(rc);
2852	return rc;
2853	}
2854	Assert(pPDDst);
2855	PSHWPDE pPdeDst = &pPDDst->a[iPDDst];
2856
2857	/* Fetch the pgm pool shadow descriptor. */
2858	PPGMPOOLPAGE pShwPde = pgmPoolGetPageByHCPhys(pVM, pPdptDst->a[iPdpt].u & EPT_PDPTE_PG_MASK);
2859	Assert(pShwPde);
2860	# endif
2861	SHWPDE PdeDst = *pPdeDst;
2862
2863	Assert(!(PdeDst.u & PGM_PDFLAGS_MAPPING));
2864	Assert(!PdeDst.n.u1Present); /* We're only supposed to call SyncPT on PDE!P and conflicts.*/
2865
2866	GSTPDE PdeSrc;
2867	PdeSrc.au32[0] = 0; /* faked so we don't have to #ifdef everything */
2868	PdeSrc.n.u1Present = 1;
2869	PdeSrc.n.u1Write = 1;
2870	PdeSrc.n.u1Accessed = 1;
2871	PdeSrc.n.u1User = 1;
2872
2873	/*
2874	* Allocate & map the page table.
2875	*/
2876	PSHWPT pPTDst;
2877	PPGMPOOLPAGE pShwPage;
2878	RTGCPHYS GCPhys;
2879
2880	/* Virtual address = physical address */
2881	GCPhys = GCPtrPage & X86_PAGE_4K_BASE_MASK;
2882	# if PGM_SHW_TYPE == PGM_TYPE_AMD64 \|\| PGM_SHW_TYPE == PGM_TYPE_EPT \|\| defined(VBOX_WITH_PGMPOOL_PAGING_ONLY)
2883	rc = pgmPoolAlloc(pVM, GCPhys & ~(RT_BIT_64(SHW_PD_SHIFT) - 1), BTH_PGMPOOLKIND_PT_FOR_PT, pShwPde->idx, iPDDst, &pShwPage);
2884	# else
2885	rc = pgmPoolAlloc(pVM, GCPhys & ~(RT_BIT_64(SHW_PD_SHIFT) - 1), BTH_PGMPOOLKIND_PT_FOR_PT, SHW_POOL_ROOT_IDX, iPDDst, &pShwPage);
2886	# endif
2887
2888	if ( rc == VINF_SUCCESS
2889	\|\| rc == VINF_PGM_CACHED_PAGE)
2890	pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
2891	else
2892	AssertMsgFailedReturn(("rc=%Rrc\n", rc), VERR_INTERNAL_ERROR);
2893
2894	PdeDst.u &= X86_PDE_AVL_MASK;
2895	PdeDst.u \|= pShwPage->Core.Key;
2896	PdeDst.n.u1Present = 1;
2897	PdeDst.n.u1Write = 1;
2898	# if PGM_SHW_TYPE == PGM_TYPE_EPT
2899	PdeDst.n.u1Execute = 1;
2900	# else
2901	PdeDst.n.u1User = 1;
2902	PdeDst.n.u1Accessed = 1;
2903	# endif
2904	*pPdeDst = PdeDst;
2905
2906	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, GCPtrPage, PGM_SYNC_NR_PAGES, 0 /* page not present */);
2907	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT), a);
2908	return rc;
2909
2910	#else
2911	AssertReleaseMsgFailed(("Shw=%d Gst=%d is not implemented!\n", PGM_SHW_TYPE, PGM_GST_TYPE));
2912	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT), a);
2913	return VERR_INTERNAL_ERROR;
2914	#endif
2915	}
2916
2917
2918
2919	/**
2920	* Prefetch a page/set of pages.
2921	*
2922	* Typically used to sync commonly used pages before entering raw mode
2923	* after a CR3 reload.
2924	*
2925	* @returns VBox status code.
2926	* @param pVM VM handle.
2927	* @param GCPtrPage Page to invalidate.
2928	*/
2929	PGM_BTH_DECL(int, PrefetchPage)(PVM pVM, RTGCPTR GCPtrPage)
2930	{
2931	#if (PGM_GST_TYPE == PGM_TYPE_32BIT \|\| PGM_GST_TYPE == PGM_TYPE_REAL \|\| PGM_GST_TYPE == PGM_TYPE_PROT \|\| PGM_GST_TYPE == PGM_TYPE_PAE \|\| PGM_GST_TYPE == PGM_TYPE_AMD64) \
2932	&& PGM_SHW_TYPE != PGM_TYPE_NESTED && PGM_SHW_TYPE != PGM_TYPE_EPT
2933	/*
2934	* Check that all Guest levels thru the PDE are present, getting the
2935	* PD and PDE in the processes.
2936	*/
2937	int rc = VINF_SUCCESS;
2938	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
2939	# if PGM_GST_TYPE == PGM_TYPE_32BIT
2940	const unsigned iPDSrc = GCPtrPage >> GST_PD_SHIFT;
2941	PGSTPD pPDSrc = pgmGstGet32bitPDPtr(&pVM->pgm.s);
2942	# elif PGM_GST_TYPE == PGM_TYPE_PAE
2943	unsigned iPDSrc;
2944	# ifdef VBOX_WITH_PGMPOOL_PAGING_ONLY
2945	X86PDPE PdpeSrc;
2946	PGSTPD pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, GCPtrPage, &iPDSrc, &PdpeSrc);
2947	# else
2948	PGSTPD pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, GCPtrPage, &iPDSrc, NULL);
2949	# endif /* VBOX_WITH_PGMPOOL_PAGING_ONLY */
2950	if (!pPDSrc)
2951	return VINF_SUCCESS; /* not present */
2952	# elif PGM_GST_TYPE == PGM_TYPE_AMD64
2953	unsigned iPDSrc;
2954	PX86PML4E pPml4eSrc;
2955	X86PDPE PdpeSrc;
2956	PGSTPD pPDSrc = pgmGstGetLongModePDPtr(&pVM->pgm.s, GCPtrPage, &pPml4eSrc, &PdpeSrc, &iPDSrc);
2957	if (!pPDSrc)
2958	return VINF_SUCCESS; /* not present */
2959	# endif
2960	const GSTPDE PdeSrc = pPDSrc->a[iPDSrc];
2961	# else
2962	PGSTPD pPDSrc = NULL;
2963	const unsigned iPDSrc = 0;
2964	GSTPDE PdeSrc;
2965
2966	PdeSrc.au32[0] = 0; /* faked so we don't have to #ifdef everything */
2967	PdeSrc.n.u1Present = 1;
2968	PdeSrc.n.u1Write = 1;
2969	PdeSrc.n.u1Accessed = 1;
2970	PdeSrc.n.u1User = 1;
2971	# endif
2972
2973	if (PdeSrc.n.u1Present && PdeSrc.n.u1Accessed)
2974	{
2975	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
2976	const X86PDE PdeDst = pgmShwGet32BitPDE(&pVM->pgm.s, GCPtrPage);
2977	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
2978	# ifdef VBOX_WITH_PGMPOOL_PAGING_ONLY
2979	const unsigned iPDDst = ((GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK);
2980	PX86PDPAE pPDDst;
2981	X86PDEPAE PdeDst;
2982	# if PGM_GST_TYPE != PGM_TYPE_PAE
2983	X86PDPE PdpeSrc;
2984
2985	/* Fake PDPT entry; access control handled on the page table level, so allow everything. */
2986	PdpeSrc.u = X86_PDPE_P; /* rw/us are reserved for PAE pdpte's; accessed bit causes invalid VT-x guest state errors */
2987	# endif
2988	int rc = pgmShwSyncPaePDPtr(pVM, GCPtrPage, &PdpeSrc, &pPDDst);
2989	if (rc != VINF_SUCCESS)
2990	{
2991	AssertRC(rc);
2992	return rc;
2993	}
2994	Assert(pPDDst);
2995	PdeDst = pPDDst->a[iPDDst];
2996	# else
2997	const X86PDEPAE PdeDst = pgmShwGetPaePDE(&pVM->pgm.s, GCPtrPage);
2998	# endif /* VBOX_WITH_PGMPOOL_PAGING_ONLY */
2999
3000	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
3001	const unsigned iPDDst = ((GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK);
3002	PX86PDPAE pPDDst;
3003	X86PDEPAE PdeDst;
3004
3005	# if PGM_GST_TYPE == PGM_TYPE_PROT
3006	/* AMD-V nested paging */
3007	X86PML4E Pml4eSrc;
3008	X86PDPE PdpeSrc;
3009	PX86PML4E pPml4eSrc = &Pml4eSrc;
3010
3011	/* Fake PML4 & PDPT entry; access control handled on the page table level, so allow everything. */
3012	Pml4eSrc.u = X86_PML4E_P \| X86_PML4E_RW \| X86_PML4E_US \| X86_PML4E_NX \| X86_PML4E_A;
3013	PdpeSrc.u = X86_PDPE_P \| X86_PDPE_RW \| X86_PDPE_US \| X86_PDPE_NX \| X86_PDPE_A;
3014	# endif
3015
3016	int rc = pgmShwSyncLongModePDPtr(pVM, GCPtrPage, pPml4eSrc, &PdpeSrc, &pPDDst);
3017	if (rc != VINF_SUCCESS)
3018	{
3019	AssertRC(rc);
3020	return rc;
3021	}
3022	Assert(pPDDst);
3023	PdeDst = pPDDst->a[iPDDst];
3024	# endif
3025	if (!(PdeDst.u & PGM_PDFLAGS_MAPPING))
3026	{
3027	if (!PdeDst.n.u1Present)
3028	/** r=bird: This guy will set the A bit on the PDE, probably harmless. */
3029	rc = PGM_BTH_NAME(SyncPT)(pVM, iPDSrc, pPDSrc, GCPtrPage);
3030	else
3031	{
3032	/** @note We used to sync PGM_SYNC_NR_PAGES pages, which triggered assertions in CSAM, because
3033	* R/W attributes of nearby pages were reset. Not sure how that could happen. Anyway, it
3034	* makes no sense to prefetch more than one page.
3035	*/
3036	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, GCPtrPage, 1, 0);
3037	if (RT_SUCCESS(rc))
3038	rc = VINF_SUCCESS;
3039	}
3040	}
3041	}
3042	return rc;
3043
3044	#elif PGM_SHW_TYPE == PGM_TYPE_NESTED \|\| PGM_SHW_TYPE == PGM_TYPE_EPT
3045	return VINF_SUCCESS; /* ignore */
3046	#endif
3047	}
3048
3049
3050
3051
3052	/**
3053	* Syncs a page during a PGMVerifyAccess() call.
3054	*
3055	* @returns VBox status code (informational included).
3056	* @param GCPtrPage The address of the page to sync.
3057	* @param fPage The effective guest page flags.
3058	* @param uErr The trap error code.
3059	*/
3060	PGM_BTH_DECL(int, VerifyAccessSyncPage)(PVM pVM, RTGCPTR GCPtrPage, unsigned fPage, unsigned uErr)
3061	{
3062	LogFlow(("VerifyAccessSyncPage: GCPtrPage=%RGv fPage=%#x uErr=%#x\n", GCPtrPage, fPage, uErr));
3063
3064	Assert(!HWACCMIsNestedPagingActive(pVM));
3065	#if (PGM_GST_TYPE == PGM_TYPE_32BIT \|\| PGM_GST_TYPE == PGM_TYPE_REAL \|\| PGM_GST_TYPE == PGM_TYPE_PROT \|\| PGM_GST_TYPE == PGM_TYPE_PAE \|\| PGM_TYPE_AMD64) \
3066	&& PGM_SHW_TYPE != PGM_TYPE_NESTED && PGM_SHW_TYPE != PGM_TYPE_EPT
3067
3068	# ifndef IN_RING0
3069	if (!(fPage & X86_PTE_US))
3070	{
3071	/*
3072	* Mark this page as safe.
3073	*/
3074	/** @todo not correct for pages that contain both code and data!! */
3075	Log(("CSAMMarkPage %RGv; scanned=%d\n", GCPtrPage, true));
3076	CSAMMarkPage(pVM, (RTRCPTR)GCPtrPage, true);
3077	}
3078	# endif
3079
3080	/*
3081	* Get guest PD and index.
3082	*/
3083	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
3084	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3085	const unsigned iPDSrc = GCPtrPage >> GST_PD_SHIFT;
3086	PGSTPD pPDSrc = pgmGstGet32bitPDPtr(&pVM->pgm.s);
3087	# elif PGM_GST_TYPE == PGM_TYPE_PAE
3088	unsigned iPDSrc;
3089	# ifdef VBOX_WITH_PGMPOOL_PAGING_ONLY
3090	X86PDPE PdpeSrc;
3091	PGSTPD pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, GCPtrPage, &iPDSrc, &PdpeSrc);
3092	# else
3093	PGSTPD pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, GCPtrPage, &iPDSrc, NULL);
3094	# endif
3095
3096	if (pPDSrc)
3097	{
3098	Log(("PGMVerifyAccess: access violation for %RGv due to non-present PDPTR\n", GCPtrPage));
3099	return VINF_EM_RAW_GUEST_TRAP;
3100	}
3101	# elif PGM_GST_TYPE == PGM_TYPE_AMD64
3102	unsigned iPDSrc;
3103	PX86PML4E pPml4eSrc;
3104	X86PDPE PdpeSrc;
3105	PGSTPD pPDSrc = pgmGstGetLongModePDPtr(&pVM->pgm.s, GCPtrPage, &pPml4eSrc, &PdpeSrc, &iPDSrc);
3106	if (!pPDSrc)
3107	{
3108	Log(("PGMVerifyAccess: access violation for %RGv due to non-present PDPTR\n", GCPtrPage));
3109	return VINF_EM_RAW_GUEST_TRAP;
3110	}
3111	# endif
3112	# else
3113	PGSTPD pPDSrc = NULL;
3114	const unsigned iPDSrc = 0;
3115	# endif
3116	int rc = VINF_SUCCESS;
3117
3118	/*
3119	* First check if the shadow pd is present.
3120	*/
3121	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
3122	PX86PDE pPdeDst = pgmShwGet32BitPDEPtr(&pVM->pgm.s, GCPtrPage);
3123	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
3124	PX86PDEPAE pPdeDst;
3125	# ifdef VBOX_WITH_PGMPOOL_PAGING_ONLY
3126	const unsigned iPDDst = ((GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK);
3127	PX86PDPAE pPDDst;
3128	# if PGM_GST_TYPE != PGM_TYPE_PAE
3129	X86PDPE PdpeSrc;
3130
3131	/* Fake PDPT entry; access control handled on the page table level, so allow everything. */
3132	PdpeSrc.u = X86_PDPE_P; /* rw/us are reserved for PAE pdpte's; accessed bit causes invalid VT-x guest state errors */
3133	# endif
3134	rc = pgmShwSyncPaePDPtr(pVM, GCPtrPage, &PdpeSrc, &pPDDst);
3135	if (rc != VINF_SUCCESS)
3136	{
3137	AssertRC(rc);
3138	return rc;
3139	}
3140	Assert(pPDDst);
3141	pPdeDst = &pPDDst->a[iPDDst];
3142	# else
3143	pPdeDst = pgmShwGetPaePDEPtr(&pVM->pgm.s, GCPtrPage);
3144	# endif
3145	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
3146	const unsigned iPDDst = ((GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK);
3147	PX86PDPAE pPDDst;
3148	PX86PDEPAE pPdeDst;
3149
3150	# if PGM_GST_TYPE == PGM_TYPE_PROT
3151	/* AMD-V nested paging */
3152	X86PML4E Pml4eSrc;
3153	X86PDPE PdpeSrc;
3154	PX86PML4E pPml4eSrc = &Pml4eSrc;
3155
3156	/* Fake PML4 & PDPT entry; access control handled on the page table level, so allow everything. */
3157	Pml4eSrc.u = X86_PML4E_P \| X86_PML4E_RW \| X86_PML4E_US \| X86_PML4E_NX \| X86_PML4E_A;
3158	PdpeSrc.u = X86_PDPE_P \| X86_PDPE_RW \| X86_PDPE_US \| X86_PDPE_NX \| X86_PDPE_A;
3159	# endif
3160
3161	rc = pgmShwSyncLongModePDPtr(pVM, GCPtrPage, pPml4eSrc, &PdpeSrc, &pPDDst);
3162	if (rc != VINF_SUCCESS)
3163	{
3164	AssertRC(rc);
3165	return rc;
3166	}
3167	Assert(pPDDst);
3168	pPdeDst = &pPDDst->a[iPDDst];
3169	# endif
3170	if (!pPdeDst->n.u1Present)
3171	{
3172	rc = PGM_BTH_NAME(SyncPT)(pVM, iPDSrc, pPDSrc, GCPtrPage);
3173	AssertRC(rc);
3174	if (rc != VINF_SUCCESS)
3175	return rc;
3176	}
3177
3178	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
3179	/* Check for dirty bit fault */
3180	rc = PGM_BTH_NAME(CheckPageFault)(pVM, uErr, pPdeDst, &pPDSrc->a[iPDSrc], GCPtrPage);
3181	if (rc == VINF_PGM_HANDLED_DIRTY_BIT_FAULT)
3182	Log(("PGMVerifyAccess: success (dirty)\n"));
3183	else
3184	{
3185	GSTPDE PdeSrc = pPDSrc->a[iPDSrc];
3186	#else
3187	{
3188	GSTPDE PdeSrc;
3189	PdeSrc.au32[0] = 0; /* faked so we don't have to #ifdef everything */
3190	PdeSrc.n.u1Present = 1;
3191	PdeSrc.n.u1Write = 1;
3192	PdeSrc.n.u1Accessed = 1;
3193	PdeSrc.n.u1User = 1;
3194
3195	#endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
3196	Assert(rc != VINF_EM_RAW_GUEST_TRAP);
3197	if (uErr & X86_TRAP_PF_US)
3198	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,PageOutOfSyncUser));
3199	else /* supervisor */
3200	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,PageOutOfSyncSupervisor));
3201
3202	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, GCPtrPage, 1, 0);
3203	if (RT_SUCCESS(rc))
3204	{
3205	/* Page was successfully synced */
3206	Log2(("PGMVerifyAccess: success (sync)\n"));
3207	rc = VINF_SUCCESS;
3208	}
3209	else
3210	{
3211	Log(("PGMVerifyAccess: access violation for %RGv rc=%d\n", GCPtrPage, rc));
3212	return VINF_EM_RAW_GUEST_TRAP;
3213	}
3214	}
3215	return rc;
3216
3217	#else /* PGM_GST_TYPE != PGM_TYPE_32BIT */
3218
3219	AssertReleaseMsgFailed(("Shw=%d Gst=%d is not implemented!\n", PGM_GST_TYPE, PGM_SHW_TYPE));
3220	return VERR_INTERNAL_ERROR;
3221	#endif /* PGM_GST_TYPE != PGM_TYPE_32BIT */
3222	}
3223
3224
3225	#if PGM_GST_TYPE == PGM_TYPE_32BIT \|\| PGM_GST_TYPE == PGM_TYPE_PAE \|\| PGM_GST_TYPE == PGM_TYPE_AMD64
3226	# if PGM_SHW_TYPE == PGM_TYPE_32BIT \|\| PGM_SHW_TYPE == PGM_TYPE_PAE \|\| PGM_SHW_TYPE == PGM_TYPE_AMD64
3227	/**
3228	* Figures out which kind of shadow page this guest PDE warrants.
3229	*
3230	* @returns Shadow page kind.
3231	* @param pPdeSrc The guest PDE in question.
3232	* @param cr4 The current guest cr4 value.
3233	*/
3234	DECLINLINE(PGMPOOLKIND) PGM_BTH_NAME(CalcPageKind)(const GSTPDE *pPdeSrc, uint32_t cr4)
3235	{
3236	# if PMG_GST_TYPE == PGM_TYPE_AMD64
3237	if (!pPdeSrc->n.u1Size)
3238	# else
3239	if (!pPdeSrc->n.u1Size \|\| !(cr4 & X86_CR4_PSE))
3240	# endif
3241	return BTH_PGMPOOLKIND_PT_FOR_PT;
3242	//switch (pPdeSrc->u & (X86_PDE4M_RW \| X86_PDE4M_US /\| X86_PDE4M_PAE_NX/))
3243	//{
3244	// case 0:
3245	// return BTH_PGMPOOLKIND_PT_FOR_BIG_RO;
3246	// case X86_PDE4M_RW:
3247	// return BTH_PGMPOOLKIND_PT_FOR_BIG_RW;
3248	// case X86_PDE4M_US:
3249	// return BTH_PGMPOOLKIND_PT_FOR_BIG_US;
3250	// case X86_PDE4M_RW \| X86_PDE4M_US:
3251	// return BTH_PGMPOOLKIND_PT_FOR_BIG_RW_US;
3252	# if 0
3253	// case X86_PDE4M_PAE_NX:
3254	// return BTH_PGMPOOLKIND_PT_FOR_BIG_NX;
3255	// case X86_PDE4M_RW \| X86_PDE4M_PAE_NX:
3256	// return BTH_PGMPOOLKIND_PT_FOR_BIG_RW_NX;
3257	// case X86_PDE4M_US \| X86_PDE4M_PAE_NX:
3258	// return BTH_PGMPOOLKIND_PT_FOR_BIG_US_NX;
3259	// case X86_PDE4M_RW \| X86_PDE4M_US \| X86_PDE4M_PAE_NX:
3260	// return BTH_PGMPOOLKIND_PT_FOR_BIG_RW_US_NX;
3261	# endif
3262	return BTH_PGMPOOLKIND_PT_FOR_BIG;
3263	//}
3264	}
3265	# endif
3266	#endif
3267
3268	#undef MY_STAM_COUNTER_INC
3269	#define MY_STAM_COUNTER_INC(a) do { } while (0)
3270
3271
3272	/**
3273	* Syncs the paging hierarchy starting at CR3.
3274	*
3275	* @returns VBox status code, no specials.
3276	* @param pVM The virtual machine.
3277	* @param cr0 Guest context CR0 register
3278	* @param cr3 Guest context CR3 register
3279	* @param cr4 Guest context CR4 register
3280	* @param fGlobal Including global page directories or not
3281	*/
3282	PGM_BTH_DECL(int, SyncCR3)(PVM pVM, uint64_t cr0, uint64_t cr3, uint64_t cr4, bool fGlobal)
3283	{
3284	if (VM_FF_ISSET(pVM, VM_FF_PGM_SYNC_CR3))
3285	fGlobal = true; /* Change this CR3 reload to be a global one. */
3286
3287	LogFlow(("SyncCR3 %d\n", fGlobal));
3288
3289	#if PGM_SHW_TYPE != PGM_TYPE_NESTED && PGM_SHW_TYPE != PGM_TYPE_EPT
3290	/*
3291	* Update page access handlers.
3292	* The virtual are always flushed, while the physical are only on demand.
3293	* WARNING: We are incorrectly not doing global flushing on Virtual Handler updates. We'll
3294	* have to look into that later because it will have a bad influence on the performance.
3295	* @note SvL: There's no need for that. Just invalidate the virtual range(s).
3296	* bird: Yes, but that won't work for aliases.
3297	*/
3298	/** @todo this MUST go away. See #1557. */
3299	STAM_PROFILE_START(&pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3Handlers), h);
3300	PGM_GST_NAME(HandlerVirtualUpdate)(pVM, cr4);
3301	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3Handlers), h);
3302	#endif
3303
3304	#if PGM_SHW_TYPE == PGM_TYPE_NESTED \|\| PGM_SHW_TYPE == PGM_TYPE_EPT
3305	/*
3306	* Nested / EPT - almost no work.
3307	*/
3308	/** @todo check if this is really necessary; the call does it as well... */
3309	HWACCMFlushTLB(pVM);
3310	return VINF_SUCCESS;
3311
3312	#elif PGM_SHW_TYPE == PGM_TYPE_AMD64
3313	/*
3314	* AMD64 (Shw & Gst) - No need to check all paging levels; we zero
3315	* out the shadow parts when the guest modifies its tables.
3316	*/
3317	return VINF_SUCCESS;
3318
3319	#else /* PGM_SHW_TYPE != PGM_TYPE_NESTED && PGM_SHW_TYPE != PGM_TYPE_EPT && PGM_SHW_TYPE != PGM_TYPE_AMD64 */
3320
3321	# ifdef VBOX_WITH_PGMPOOL_PAGING_ONLY
3322	# ifdef PGM_WITHOUT_MAPPINGS
3323	Assert(pVM->pgm.s.fMappingsFixed);
3324	return VINF_SUCCESS;
3325	# else
3326	/* Nothing to do when mappings are fixed. */
3327	if (pVM->pgm.s.fMappingsFixed)
3328	return VINF_SUCCESS;
3329
3330	int rc = PGMMapResolveConflicts(pVM);
3331	Assert(rc == VINF_SUCCESS \|\| rc == VINF_PGM_SYNC_CR3);
3332	if (rc == VINF_PGM_SYNC_CR3)
3333	{
3334	LogFlow(("SyncCR3: detected conflict -> VINF_PGM_SYNC_CR3\n"));
3335	return VINF_PGM_SYNC_CR3;
3336	}
3337	# endif
3338	return VINF_SUCCESS;
3339	# else
3340	/*
3341	* PAE and 32-bit legacy mode (shadow).
3342	* (Guest PAE, 32-bit legacy, protected and real modes.)
3343	*/
3344	Assert(fGlobal \|\| (cr4 & X86_CR4_PGE));
3345	MY_STAM_COUNTER_INC(fGlobal ? &pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3Global) : &pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3NotGlobal));
3346
3347	# if PGM_GST_TYPE == PGM_TYPE_32BIT \|\| PGM_GST_TYPE == PGM_TYPE_PAE
3348	bool const fBigPagesSupported = !!(CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
3349
3350	/*
3351	* Get page directory addresses.
3352	*/
3353	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
3354	PX86PDE pPDEDst = pgmShwGet32BitPDEPtr(&pVM->pgm.s, 0);
3355	# else /* PGM_SHW_TYPE == PGM_TYPE_PAE */
3356	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3357	PX86PDEPAE pPDEDst = NULL;
3358	# endif
3359	# endif
3360
3361	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3362	PGSTPD pPDSrc = pgmGstGet32bitPDPtr(&pVM->pgm.s);
3363	Assert(pPDSrc);
3364	# if !defined(IN_RC) && !defined(VBOX_WITH_2X_4GB_ADDR_SPACE_IN_R0)
3365	Assert(PGMPhysGCPhys2R3PtrAssert(pVM, (RTGCPHYS)(cr3 & GST_CR3_PAGE_MASK), sizeof(*pPDSrc)) == (RTR3PTR)pPDSrc);
3366	# endif
3367	# endif /* PGM_GST_TYPE == PGM_TYPE_32BIT */
3368
3369	/*
3370	* Iterate the the CR3 page.
3371	*/
3372	PPGMMAPPING pMapping;
3373	unsigned iPdNoMapping;
3374	const bool fRawR0Enabled = EMIsRawRing0Enabled(pVM);
3375	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
3376
3377	/* Only check mappings if they are supposed to be put into the shadow page table. */
3378	if (pgmMapAreMappingsEnabled(&pVM->pgm.s))
3379	{
3380	pMapping = pVM->pgm.s.CTX_SUFF(pMappings);
3381	iPdNoMapping = (pMapping) ? (pMapping->GCPtr >> GST_PD_SHIFT) : ~0U;
3382	}
3383	else
3384	{
3385	pMapping = 0;
3386	iPdNoMapping = ~0U;
3387	}
3388
3389	# if PGM_GST_TYPE == PGM_TYPE_PAE
3390	for (uint64_t iPdpt = 0; iPdpt < GST_PDPE_ENTRIES; iPdpt++)
3391	{
3392	unsigned iPDSrc;
3393	X86PDPE PdpeSrc;
3394	PGSTPD pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, iPdpt << X86_PDPT_SHIFT, &iPDSrc, &PdpeSrc);
3395	PX86PDEPAE pPDEDst = pgmShwGetPaePDEPtr(&pVM->pgm.s, iPdpt << X86_PDPT_SHIFT);
3396	PX86PDPT pPdptDst = pgmShwGetPaePDPTPtr(&pVM->pgm.s);
3397
3398	if (pPDSrc == NULL)
3399	{
3400	/* PDPE not present */
3401	if (pPdptDst->a[iPdpt].n.u1Present)
3402	{
3403	LogFlow(("SyncCR3: guest PDPE %lld not present; clear shw pdpe\n", iPdpt));
3404	/* for each page directory entry */
3405	for (unsigned iPD = 0; iPD < RT_ELEMENTS(pPDSrc->a); iPD++)
3406	{
3407	if ( pPDEDst[iPD].n.u1Present
3408	&& !(pPDEDst[iPD].u & PGM_PDFLAGS_MAPPING))
3409	{
3410	pgmPoolFree(pVM, pPDEDst[iPD].u & SHW_PDE_PG_MASK, SHW_POOL_ROOT_IDX, iPdpt * X86_PG_PAE_ENTRIES + iPD);
3411	pPDEDst[iPD].u = 0;
3412	}
3413	}
3414	}
3415	if (!(pPdptDst->a[iPdpt].u & PGM_PLXFLAGS_MAPPING))
3416	pPdptDst->a[iPdpt].n.u1Present = 0;
3417	continue;
3418	}
3419	# else /* PGM_GST_TYPE != PGM_TYPE_PAE */
3420	{
3421	# endif /* PGM_GST_TYPE != PGM_TYPE_PAE */
3422	for (unsigned iPD = 0; iPD < RT_ELEMENTS(pPDSrc->a); iPD++)
3423	{
3424	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
3425	if ((iPD & 255) == 0) /* Start of new PD. */
3426	pPDEDst = pgmShwGetPaePDEPtr(&pVM->pgm.s, (uint32_t)iPD << GST_PD_SHIFT);
3427	# endif
3428	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
3429	Assert(pgmShwGet32BitPDEPtr(&pVM->pgm.s, (uint32_t)iPD << SHW_PD_SHIFT) == pPDEDst);
3430	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
3431	# if defined(VBOX_STRICT) && !defined(VBOX_WITH_2X_4GB_ADDR_SPACE_IN_R0) /* Unfortuantely not reliable with PGMR0DynMap and multiple VMs. */
3432	RTGCPTR GCPtrStrict = (uint32_t)iPD << GST_PD_SHIFT;
3433	# if PGM_GST_TYPE == PGM_TYPE_PAE
3434	GCPtrStrict \|= iPdpt << X86_PDPT_SHIFT;
3435	# endif
3436	AssertMsg(pgmShwGetPaePDEPtr(&pVM->pgm.s, GCPtrStrict) == pPDEDst, ("%p vs %p (%RGv)\n", pgmShwGetPaePDEPtr(&pVM->pgm.s, GCPtrStrict), pPDEDst, GCPtrStrict));
3437	# endif /* VBOX_STRICT */
3438	# endif
3439	GSTPDE PdeSrc = pPDSrc->a[iPD];
3440	if ( PdeSrc.n.u1Present
3441	&& (PdeSrc.n.u1User \|\| fRawR0Enabled))
3442	{
3443	# if ( PGM_GST_TYPE == PGM_TYPE_32BIT \
3444	\|\| PGM_GST_TYPE == PGM_TYPE_PAE) \
3445	&& !defined(PGM_WITHOUT_MAPPINGS)
3446
3447	/*
3448	* Check for conflicts with GC mappings.
3449	*/
3450	# if PGM_GST_TYPE == PGM_TYPE_PAE
3451	if (iPD + iPdpt * X86_PG_PAE_ENTRIES == iPdNoMapping)
3452	# else
3453	if (iPD == iPdNoMapping)
3454	# endif
3455	{
3456	if (pVM->pgm.s.fMappingsFixed)
3457	{
3458	/* It's fixed, just skip the mapping. */
3459	const unsigned cPTs = pMapping->cb >> GST_PD_SHIFT;
3460	Assert(PGM_GST_TYPE == PGM_TYPE_32BIT \|\| (iPD + cPTs - 1) / X86_PG_PAE_ENTRIES == iPD / X86_PG_PAE_ENTRIES);
3461	iPD += cPTs - 1;
3462	# if PGM_SHW_TYPE != PGM_GST_TYPE /* SHW==PAE && GST==32BIT */
3463	pPDEDst = pgmShwGetPaePDEPtr(&pVM->pgm.s, (uint32_t)(iPD + 1) << GST_PD_SHIFT);
3464	# else
3465	pPDEDst += cPTs;
3466	# endif
3467	pMapping = pMapping->CTX_SUFF(pNext);
3468	iPdNoMapping = pMapping ? pMapping->GCPtr >> GST_PD_SHIFT : ~0U;
3469	continue;
3470	}
3471	# ifdef IN_RING3
3472	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3473	int rc = pgmR3SyncPTResolveConflict(pVM, pMapping, pPDSrc, iPD << GST_PD_SHIFT);
3474	# elif PGM_GST_TYPE == PGM_TYPE_PAE
3475	int rc = pgmR3SyncPTResolveConflictPAE(pVM, pMapping, (iPdpt << GST_PDPT_SHIFT) + (iPD << GST_PD_SHIFT));
3476	# endif
3477	if (RT_FAILURE(rc))
3478	return rc;
3479
3480	/*
3481	* Update iPdNoMapping and pMapping.
3482	*/
3483	pMapping = pVM->pgm.s.pMappingsR3;
3484	while (pMapping && pMapping->GCPtr < (iPD << GST_PD_SHIFT))
3485	pMapping = pMapping->pNextR3;
3486	iPdNoMapping = pMapping ? pMapping->GCPtr >> GST_PD_SHIFT : ~0U;
3487	# else /* !IN_RING3 */
3488	LogFlow(("SyncCR3: detected conflict -> VINF_PGM_SYNC_CR3\n"));
3489	return VINF_PGM_SYNC_CR3;
3490	# endif /* !IN_RING3 */
3491	}
3492	# else /* (PGM_GST_TYPE != PGM_TYPE_32BIT && PGM_GST_TYPE != PGM_TYPE_PAE) \|\| PGM_WITHOUT_MAPPINGS */
3493	Assert(!pgmMapAreMappingsEnabled(&pVM->pgm.s));
3494	# endif /* (PGM_GST_TYPE != PGM_TYPE_32BIT && PGM_GST_TYPE != PGM_TYPE_PAE) \|\| PGM_WITHOUT_MAPPINGS */
3495
3496	/*
3497	* Sync page directory entry.
3498	*
3499	* The current approach is to allocated the page table but to set
3500	* the entry to not-present and postpone the page table synching till
3501	* it's actually used.
3502	*/
3503	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
3504	for (unsigned i = 0, iPdShw = iPD * 2; i < 2; i++, iPdShw++) /* pray that the compiler unrolls this */
3505	# elif PGM_GST_TYPE == PGM_TYPE_PAE
3506	const unsigned iPdShw = iPD + iPdpt * X86_PG_PAE_ENTRIES; NOREF(iPdShw);
3507	# else
3508	const unsigned iPdShw = iPD; NOREF(iPdShw);
3509	# endif
3510	{
3511	SHWPDE PdeDst = *pPDEDst;
3512	if (PdeDst.n.u1Present)
3513	{
3514	PPGMPOOLPAGE pShwPage = pgmPoolGetPage(pPool, PdeDst.u & SHW_PDE_PG_MASK);
3515	RTGCPHYS GCPhys;
3516	if ( !PdeSrc.b.u1Size
3517	\|\| !fBigPagesSupported)
3518	{
3519	GCPhys = PdeSrc.u & GST_PDE_PG_MASK;
3520	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
3521	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
3522	GCPhys \|= i * (PAGE_SIZE / 2);
3523	# endif
3524	}
3525	else
3526	{
3527	GCPhys = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc);
3528	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
3529	/* Select the right PDE as we're emulating a 4MB page directory with two 2 MB shadow PDEs.*/
3530	GCPhys \|= i * X86_PAGE_2M_SIZE;
3531	# endif
3532	}
3533
3534	if ( pShwPage->GCPhys == GCPhys
3535	&& pShwPage->enmKind == PGM_BTH_NAME(CalcPageKind)(&PdeSrc, cr4)
3536	&& ( pShwPage->fCached
3537	\|\| ( !fGlobal
3538	&& ( false
3539	# ifdef PGM_SKIP_GLOBAL_PAGEDIRS_ON_NONGLOBAL_FLUSH
3540	\|\| ( (PdeSrc.u & (X86_PDE4M_PS \| X86_PDE4M_G)) == (X86_PDE4M_PS \| X86_PDE4M_G)
3541	&& (cr4 & (X86_CR4_PGE \| X86_CR4_PSE)) == (X86_CR4_PGE \| X86_CR4_PSE)) /* global 2/4MB page. */
3542	\|\| ( !pShwPage->fSeenNonGlobal
3543	&& (cr4 & X86_CR4_PGE))
3544	# endif
3545	)
3546	)
3547	)
3548	&& ( (PdeSrc.u & (X86_PDE_US \| X86_PDE_RW)) == (PdeDst.u & (X86_PDE_US \| X86_PDE_RW))
3549	\|\| ( fBigPagesSupported
3550	&& ((PdeSrc.u & (X86_PDE_US \| X86_PDE4M_PS \| X86_PDE4M_D)) \| PGM_PDFLAGS_TRACK_DIRTY)
3551	== ((PdeDst.u & (X86_PDE_US \| X86_PDE_RW \| PGM_PDFLAGS_TRACK_DIRTY)) \| X86_PDE4M_PS))
3552	)
3553	)
3554	{
3555	# ifdef VBOX_WITH_STATISTICS
3556	if ( !fGlobal
3557	&& (PdeSrc.u & (X86_PDE4M_PS \| X86_PDE4M_G)) == (X86_PDE4M_PS \| X86_PDE4M_G)
3558	&& (cr4 & (X86_CR4_PGE \| X86_CR4_PSE)) == (X86_CR4_PGE \| X86_CR4_PSE))
3559	MY_STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3DstSkippedGlobalPD));
3560	else if (!fGlobal && !pShwPage->fSeenNonGlobal && (cr4 & X86_CR4_PGE))
3561	MY_STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3DstSkippedGlobalPT));
3562	else
3563	MY_STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3DstCacheHit));
3564	# endif /* VBOX_WITH_STATISTICS */
3565	/** @todo a replacement strategy isn't really needed unless we're using a very small pool < 512 pages.
3566	* The whole ageing stuff should be put in yet another set of #ifdefs. For now, let's just skip it. */
3567	//# ifdef PGMPOOL_WITH_CACHE
3568	// pgmPoolCacheUsed(pPool, pShwPage);
3569	//# endif
3570	}
3571	else
3572	{
3573	pgmPoolFreeByPage(pPool, pShwPage, SHW_POOL_ROOT_IDX, iPdShw);
3574	pPDEDst->u = 0;
3575	MY_STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3DstFreed));
3576	}
3577	}
3578	else
3579	MY_STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3DstNotPresent));
3580
3581	/* advance */
3582	pPDEDst++;
3583	} /* foreach 2MB PAE PDE in 4MB guest PDE */
3584	}
3585	# if PGM_GST_TYPE == PGM_TYPE_PAE
3586	else if (iPD + iPdpt * X86_PG_PAE_ENTRIES != iPdNoMapping)
3587	# else
3588	else if (iPD != iPdNoMapping)
3589	# endif
3590	{
3591	/*
3592	* Check if there is any page directory to mark not present here.
3593	*/
3594	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
3595	for (unsigned i = 0, iPdShw = iPD * 2; i < 2; i++, iPdShw++) /* pray that the compiler unrolls this */
3596	# elif PGM_GST_TYPE == PGM_TYPE_PAE
3597	const unsigned iPdShw = iPD + iPdpt * X86_PG_PAE_ENTRIES;
3598	# else
3599	const unsigned iPdShw = iPD;
3600	# endif
3601	{
3602	if (pPDEDst->n.u1Present)
3603	{
3604	pgmPoolFree(pVM, pPDEDst->u & SHW_PDE_PG_MASK, SHW_POOL_ROOT_IDX, iPdShw);
3605	pPDEDst->u = 0;
3606	MY_STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3DstFreedSrcNP));
3607	}
3608	pPDEDst++;
3609	}
3610	}
3611	else
3612	{
3613	# if ( PGM_GST_TYPE == PGM_TYPE_32BIT \
3614	\|\| PGM_GST_TYPE == PGM_TYPE_PAE) \
3615	&& !defined(PGM_WITHOUT_MAPPINGS)
3616
3617	const unsigned cPTs = pMapping->cb >> GST_PD_SHIFT;
3618
3619	Assert(pgmMapAreMappingsEnabled(&pVM->pgm.s));
3620	if (pVM->pgm.s.fMappingsFixed)
3621	{
3622	/* It's fixed, just skip the mapping. */
3623	pMapping = pMapping->CTX_SUFF(pNext);
3624	iPdNoMapping = pMapping ? pMapping->GCPtr >> GST_PD_SHIFT : ~0U;
3625	}
3626	else
3627	{
3628	/*
3629	* Check for conflicts for subsequent pagetables
3630	* and advance to the next mapping.
3631	*/
3632	iPdNoMapping = ~0U;
3633	unsigned iPT = cPTs;
3634	while (iPT-- > 1)
3635	{
3636	if ( pPDSrc->a[iPD + iPT].n.u1Present
3637	&& (pPDSrc->a[iPD + iPT].n.u1User \|\| fRawR0Enabled))
3638	{
3639	# ifdef IN_RING3
3640	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3641	int rc = pgmR3SyncPTResolveConflict(pVM, pMapping, pPDSrc, iPD << GST_PD_SHIFT);
3642	# elif PGM_GST_TYPE == PGM_TYPE_PAE
3643	int rc = pgmR3SyncPTResolveConflictPAE(pVM, pMapping, (iPdpt << GST_PDPT_SHIFT) + (iPD << GST_PD_SHIFT));
3644	# endif
3645	if (RT_FAILURE(rc))
3646	return rc;
3647
3648	/*
3649	* Update iPdNoMapping and pMapping.
3650	*/
3651	pMapping = pVM->pgm.s.CTX_SUFF(pMappings);
3652	while (pMapping && pMapping->GCPtr < (iPD << GST_PD_SHIFT))
3653	pMapping = pMapping->CTX_SUFF(pNext);
3654	iPdNoMapping = pMapping ? pMapping->GCPtr >> GST_PD_SHIFT : ~0U;
3655	break;
3656	# else
3657	LogFlow(("SyncCR3: detected conflict -> VINF_PGM_SYNC_CR3\n"));
3658	return VINF_PGM_SYNC_CR3;
3659	# endif
3660	}
3661	}
3662	if (iPdNoMapping == ~0U && pMapping)
3663	{
3664	pMapping = pMapping->CTX_SUFF(pNext);
3665	if (pMapping)
3666	iPdNoMapping = pMapping->GCPtr >> GST_PD_SHIFT;
3667	}
3668	}
3669
3670	/* advance. */
3671	Assert(PGM_GST_TYPE == PGM_TYPE_32BIT \|\| (iPD + cPTs - 1) / X86_PG_PAE_ENTRIES == iPD / X86_PG_PAE_ENTRIES);
3672	iPD += cPTs - 1;
3673	# if PGM_SHW_TYPE != PGM_GST_TYPE /* SHW==PAE && GST==32BIT */
3674	pPDEDst = pgmShwGetPaePDEPtr(&pVM->pgm.s, (uint32_t)(iPD + 1) << GST_PD_SHIFT);
3675	# else
3676	pPDEDst += cPTs;
3677	# endif
3678	# if PGM_GST_TYPE != PGM_SHW_TYPE
3679	AssertCompile(PGM_GST_TYPE == PGM_TYPE_32BIT && PGM_SHW_TYPE == PGM_TYPE_PAE);
3680	# endif
3681	# else /* (PGM_GST_TYPE != PGM_TYPE_32BIT && PGM_GST_TYPE != PGM_TYPE_PAE) \|\| PGM_WITHOUT_MAPPINGS */
3682	Assert(!pgmMapAreMappingsEnabled(&pVM->pgm.s));
3683	# endif /* (PGM_GST_TYPE != PGM_TYPE_32BIT && PGM_GST_TYPE != PGM_TYPE_PAE) \|\| PGM_WITHOUT_MAPPINGS */
3684	}
3685
3686	} /* for iPD */
3687	} /* for each PDPTE (PAE) */
3688	return VINF_SUCCESS;
3689
3690	# else /* guest real and protected mode */
3691	return VINF_SUCCESS;
3692	# endif
3693	#endif /* VBOX_WITH_PGMPOOL_PAGING_ONLY */
3694	#endif /* PGM_SHW_TYPE != PGM_TYPE_NESTED && PGM_SHW_TYPE != PGM_TYPE_EPT && PGM_SHW_TYPE != PGM_TYPE_AMD64 */
3695	}
3696
3697
3698
3699
3700	#ifdef VBOX_STRICT
3701	#ifdef IN_RC
3702	# undef AssertMsgFailed
3703	# define AssertMsgFailed Log
3704	#endif
3705	#ifdef IN_RING3
3706	# include <VBox/dbgf.h>
3707
3708	/**
3709	* Dumps a page table hierarchy use only physical addresses and cr4/lm flags.
3710	*
3711	* @returns VBox status code (VINF_SUCCESS).
3712	* @param pVM The VM handle.
3713	* @param cr3 The root of the hierarchy.
3714	* @param crr The cr4, only PAE and PSE is currently used.
3715	* @param fLongMode Set if long mode, false if not long mode.
3716	* @param cMaxDepth Number of levels to dump.
3717	* @param pHlp Pointer to the output functions.
3718	*/
3719	__BEGIN_DECLS
3720	VMMR3DECL(int) PGMR3DumpHierarchyHC(PVM pVM, uint32_t cr3, uint32_t cr4, bool fLongMode, unsigned cMaxDepth, PCDBGFINFOHLP pHlp);
3721	__END_DECLS
3722
3723	#endif
3724
3725	/**
3726	* Checks that the shadow page table is in sync with the guest one.
3727	*
3728	* @returns The number of errors.
3729	* @param pVM The virtual machine.
3730	* @param cr3 Guest context CR3 register
3731	* @param cr4 Guest context CR4 register
3732	* @param GCPtr Where to start. Defaults to 0.
3733	* @param cb How much to check. Defaults to everything.
3734	*/
3735	PGM_BTH_DECL(unsigned, AssertCR3)(PVM pVM, uint64_t cr3, uint64_t cr4, RTGCPTR GCPtr, RTGCPTR cb)
3736	{
3737	#if PGM_SHW_TYPE == PGM_TYPE_NESTED \|\| PGM_SHW_TYPE == PGM_TYPE_EPT
3738	return 0;
3739	#else
3740	unsigned cErrors = 0;
3741
3742	#if PGM_GST_TYPE == PGM_TYPE_PAE
3743	/** @todo currently broken; crashes below somewhere */
3744	AssertFailed();
3745	#endif
3746
3747	#if PGM_GST_TYPE == PGM_TYPE_32BIT \
3748	\|\| PGM_GST_TYPE == PGM_TYPE_PAE \
3749	\|\| PGM_GST_TYPE == PGM_TYPE_AMD64
3750
3751	# if PGM_GST_TYPE == PGM_TYPE_AMD64
3752	bool fBigPagesSupported = true;
3753	# else
3754	bool fBigPagesSupported = !!(CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
3755	# endif
3756	PPGM pPGM = &pVM->pgm.s;
3757	RTGCPHYS GCPhysGst; /* page address derived from the guest page tables. */
3758	RTHCPHYS HCPhysShw; /* page address derived from the shadow page tables. */
3759	# ifndef IN_RING0
3760	RTHCPHYS HCPhys; /* general usage. */
3761	# endif
3762	int rc;
3763
3764	/*
3765	* Check that the Guest CR3 and all its mappings are correct.
3766	*/
3767	AssertMsgReturn(pPGM->GCPhysCR3 == (cr3 & GST_CR3_PAGE_MASK),
3768	("Invalid GCPhysCR3=%RGp cr3=%RGp\n", pPGM->GCPhysCR3, (RTGCPHYS)cr3),
3769	false);
3770	# if !defined(IN_RING0) && PGM_GST_TYPE != PGM_TYPE_AMD64
3771	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3772	rc = PGMShwGetPage(pVM, (RTGCPTR)pPGM->pGst32BitPdRC, NULL, &HCPhysShw);
3773	# else
3774	rc = PGMShwGetPage(pVM, (RTGCPTR)pPGM->pGstPaePdptRC, NULL, &HCPhysShw);
3775	# endif
3776	AssertRCReturn(rc, 1);
3777	HCPhys = NIL_RTHCPHYS;
3778	rc = pgmRamGCPhys2HCPhys(pPGM, cr3 & GST_CR3_PAGE_MASK, &HCPhys);
3779	AssertMsgReturn(HCPhys == HCPhysShw, ("HCPhys=%RHp HCPhyswShw=%RHp (cr3)\n", HCPhys, HCPhysShw), false);
3780	# if PGM_GST_TYPE == PGM_TYPE_32BIT && defined(IN_RING3)
3781	RTGCPHYS GCPhys;
3782	rc = PGMR3DbgR3Ptr2GCPhys(pVM, pPGM->pGst32BitPdR3, &GCPhys);
3783	AssertRCReturn(rc, 1);
3784	AssertMsgReturn((cr3 & GST_CR3_PAGE_MASK) == GCPhys, ("GCPhys=%RGp cr3=%RGp\n", GCPhys, (RTGCPHYS)cr3), false);
3785	# endif
3786	# endif /* !IN_RING0 */
3787
3788	/*
3789	* Get and check the Shadow CR3.
3790	*/
3791	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
3792	unsigned cPDEs = X86_PG_ENTRIES;
3793	unsigned cIncrement = X86_PG_ENTRIES * PAGE_SIZE;
3794	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
3795	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3796	unsigned cPDEs = X86_PG_PAE_ENTRIES * 4; /* treat it as a 2048 entry table. */
3797	# else
3798	unsigned cPDEs = X86_PG_PAE_ENTRIES;
3799	# endif
3800	unsigned cIncrement = X86_PG_PAE_ENTRIES * PAGE_SIZE;
3801	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
3802	unsigned cPDEs = X86_PG_PAE_ENTRIES;
3803	unsigned cIncrement = X86_PG_PAE_ENTRIES * PAGE_SIZE;
3804	# endif
3805	if (cb != ~(RTGCPTR)0)
3806	cPDEs = RT_MIN(cb >> SHW_PD_SHIFT, 1);
3807
3808	/** @todo call the other two PGMAssert() functions. /
3809
3810	# if PGM_GST_TYPE == PGM_TYPE_AMD64 \|\| PGM_GST_TYPE == PGM_TYPE_PAE
3811	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
3812	# endif
3813
3814	# if PGM_GST_TYPE == PGM_TYPE_AMD64
3815	unsigned iPml4 = (GCPtr >> X86_PML4_SHIFT) & X86_PML4_MASK;
3816
3817	for (; iPml4 < X86_PG_PAE_ENTRIES; iPml4++)
3818	{
3819	PPGMPOOLPAGE pShwPdpt = NULL;
3820	PX86PML4E pPml4eSrc;
3821	PX86PML4E pPml4eDst;
3822	RTGCPHYS GCPhysPdptSrc;
3823
3824	pPml4eSrc = pgmGstGetLongModePML4EPtr(&pVM->pgm.s, iPml4);
3825	pPml4eDst = pgmShwGetLongModePML4EPtr(&pVM->pgm.s, iPml4);
3826
3827	/* Fetch the pgm pool shadow descriptor if the shadow pml4e is present. */
3828	if (!pPml4eDst->n.u1Present)
3829	{
3830	GCPtr += _2M * UINT64_C(512) * UINT64_C(512);
3831	continue;
3832	}
3833
3834	pShwPdpt = pgmPoolGetPage(pPool, pPml4eDst->u & X86_PML4E_PG_MASK);
3835	GCPhysPdptSrc = pPml4eSrc->u & X86_PML4E_PG_MASK_FULL;
3836
3837	if (pPml4eSrc->n.u1Present != pPml4eDst->n.u1Present)
3838	{
3839	AssertMsgFailed(("Present bit doesn't match! pPml4eDst.u=%#RX64 pPml4eSrc.u=%RX64\n", pPml4eDst->u, pPml4eSrc->u));
3840	GCPtr += _2M * UINT64_C(512) * UINT64_C(512);
3841	cErrors++;
3842	continue;
3843	}
3844
3845	if (GCPhysPdptSrc != pShwPdpt->GCPhys)
3846	{
3847	AssertMsgFailed(("Physical address doesn't match! iPml4 %d pPml4eDst.u=%#RX64 pPml4eSrc.u=%RX64 Phys %RX64 vs %RX64\n", iPml4, pPml4eDst->u, pPml4eSrc->u, pShwPdpt->GCPhys, GCPhysPdptSrc));
3848	GCPtr += _2M * UINT64_C(512) * UINT64_C(512);
3849	cErrors++;
3850	continue;
3851	}
3852
3853	if ( pPml4eDst->n.u1User != pPml4eSrc->n.u1User
3854	\|\| pPml4eDst->n.u1Write != pPml4eSrc->n.u1Write
3855	\|\| pPml4eDst->n.u1NoExecute != pPml4eSrc->n.u1NoExecute)
3856	{
3857	AssertMsgFailed(("User/Write/NoExec bits don't match! pPml4eDst.u=%#RX64 pPml4eSrc.u=%RX64\n", pPml4eDst->u, pPml4eSrc->u));
3858	GCPtr += _2M * UINT64_C(512) * UINT64_C(512);
3859	cErrors++;
3860	continue;
3861	}
3862	# else /* PGM_GST_TYPE != PGM_TYPE_AMD64 */
3863	{
3864	# endif /* PGM_GST_TYPE != PGM_TYPE_AMD64 */
3865
3866	# if PGM_GST_TYPE == PGM_TYPE_AMD64 \|\| PGM_GST_TYPE == PGM_TYPE_PAE
3867	/*
3868	* Check the PDPTEs too.
3869	*/
3870	unsigned iPdpt = (GCPtr >> SHW_PDPT_SHIFT) & SHW_PDPT_MASK;
3871
3872	for (;iPdpt <= SHW_PDPT_MASK; iPdpt++)
3873	{
3874	unsigned iPDSrc;
3875	PPGMPOOLPAGE pShwPde = NULL;
3876	PX86PDPE pPdpeDst;
3877	RTGCPHYS GCPhysPdeSrc;
3878	# if PGM_GST_TYPE == PGM_TYPE_PAE
3879	X86PDPE PdpeSrc;
3880	PGSTPD pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, GCPtr, &iPDSrc, &PdpeSrc);
3881	PX86PDPT pPdptDst = pgmShwGetPaePDPTPtr(&pVM->pgm.s);
3882	# else
3883	PX86PML4E pPml4eSrc;
3884	X86PDPE PdpeSrc;
3885	PX86PDPT pPdptDst;
3886	PX86PDPAE pPDDst;
3887	PGSTPD pPDSrc = pgmGstGetLongModePDPtr(&pVM->pgm.s, GCPtr, &pPml4eSrc, &PdpeSrc, &iPDSrc);
3888
3889	rc = pgmShwGetLongModePDPtr(pVM, GCPtr, NULL, &pPdptDst, &pPDDst);
3890	if (rc != VINF_SUCCESS)
3891	{
3892	AssertMsg(rc == VERR_PAGE_DIRECTORY_PTR_NOT_PRESENT, ("Unexpected rc=%Rrc\n", rc));
3893	GCPtr += 512 * _2M;
3894	continue; /* next PDPTE */
3895	}
3896	Assert(pPDDst);
3897	# endif
3898	Assert(iPDSrc == 0);
3899
3900	pPdpeDst = &pPdptDst->a[iPdpt];
3901
3902	if (!pPdpeDst->n.u1Present)
3903	{
3904	GCPtr += 512 * _2M;
3905	continue; /* next PDPTE */
3906	}
3907
3908	pShwPde = pgmPoolGetPage(pPool, pPdpeDst->u & X86_PDPE_PG_MASK);
3909	GCPhysPdeSrc = PdpeSrc.u & X86_PDPE_PG_MASK;
3910
3911	if (pPdpeDst->n.u1Present != PdpeSrc.n.u1Present)
3912	{
3913	AssertMsgFailed(("Present bit doesn't match! pPdpeDst.u=%#RX64 pPdpeSrc.u=%RX64\n", pPdpeDst->u, PdpeSrc.u));
3914	GCPtr += 512 * _2M;
3915	cErrors++;
3916	continue;
3917	}
3918
3919	if (GCPhysPdeSrc != pShwPde->GCPhys)
3920	{
3921	# if PGM_GST_TYPE == PGM_TYPE_AMD64
3922	AssertMsgFailed(("Physical address doesn't match! iPml4 %d iPdpt %d pPdpeDst.u=%#RX64 pPdpeSrc.u=%RX64 Phys %RX64 vs %RX64\n", iPml4, iPdpt, pPdpeDst->u, PdpeSrc.u, pShwPde->GCPhys, GCPhysPdeSrc));
3923	# else
3924	AssertMsgFailed(("Physical address doesn't match! iPdpt %d pPdpeDst.u=%#RX64 pPdpeSrc.u=%RX64 Phys %RX64 vs %RX64\n", iPdpt, pPdpeDst->u, PdpeSrc.u, pShwPde->GCPhys, GCPhysPdeSrc));
3925	# endif
3926	GCPtr += 512 * _2M;
3927	cErrors++;
3928	continue;
3929	}
3930
3931	# if PGM_GST_TYPE == PGM_TYPE_AMD64
3932	if ( pPdpeDst->lm.u1User != PdpeSrc.lm.u1User
3933	\|\| pPdpeDst->lm.u1Write != PdpeSrc.lm.u1Write
3934	\|\| pPdpeDst->lm.u1NoExecute != PdpeSrc.lm.u1NoExecute)
3935	{
3936	AssertMsgFailed(("User/Write/NoExec bits don't match! pPdpeDst.u=%#RX64 pPdpeSrc.u=%RX64\n", pPdpeDst->u, PdpeSrc.u));
3937	GCPtr += 512 * _2M;
3938	cErrors++;
3939	continue;
3940	}
3941	# endif
3942
3943	# else /* PGM_GST_TYPE != PGM_TYPE_AMD64 && PGM_GST_TYPE != PGM_TYPE_PAE */
3944	{
3945	# endif /* PGM_GST_TYPE != PGM_TYPE_AMD64 && PGM_GST_TYPE != PGM_TYPE_PAE */
3946	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3947	GSTPD const *pPDSrc = pgmGstGet32bitPDPtr(&pVM->pgm.s);
3948	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
3949	PCX86PD pPDDst = pgmShwGet32BitPDPtr(&pVM->pgm.s);
3950	# endif
3951	# endif /* PGM_GST_TYPE == PGM_TYPE_32BIT */
3952	/*
3953	* Iterate the shadow page directory.
3954	*/
3955	GCPtr = (GCPtr >> SHW_PD_SHIFT) << SHW_PD_SHIFT;
3956	unsigned iPDDst = (GCPtr >> SHW_PD_SHIFT) & SHW_PD_MASK;
3957
3958	for (;
3959	iPDDst < cPDEs;
3960	iPDDst++, GCPtr += cIncrement)
3961	{
3962	# if PGM_SHW_TYPE == PGM_TYPE_PAE
3963	const SHWPDE PdeDst = *pgmShwGetPaePDEPtr(pPGM, GCPtr);
3964	# else
3965	const SHWPDE PdeDst = pPDDst->a[iPDDst];
3966	# endif
3967	if (PdeDst.u & PGM_PDFLAGS_MAPPING)
3968	{
3969	Assert(pgmMapAreMappingsEnabled(&pVM->pgm.s));
3970	if ((PdeDst.u & X86_PDE_AVL_MASK) != PGM_PDFLAGS_MAPPING)
3971	{
3972	AssertMsgFailed(("Mapping shall only have PGM_PDFLAGS_MAPPING set! PdeDst.u=%#RX64\n", (uint64_t)PdeDst.u));
3973	cErrors++;
3974	continue;
3975	}
3976	}
3977	else if ( (PdeDst.u & X86_PDE_P)
3978	\|\| ((PdeDst.u & (X86_PDE_P \| PGM_PDFLAGS_TRACK_DIRTY)) == (X86_PDE_P \| PGM_PDFLAGS_TRACK_DIRTY))
3979	)
3980	{
3981	HCPhysShw = PdeDst.u & SHW_PDE_PG_MASK;
3982	PPGMPOOLPAGE pPoolPage = pgmPoolGetPageByHCPhys(pVM, HCPhysShw);
3983	if (!pPoolPage)
3984	{
3985	AssertMsgFailed(("Invalid page table address %RHp at %RGv! PdeDst=%#RX64\n",
3986	HCPhysShw, GCPtr, (uint64_t)PdeDst.u));
3987	cErrors++;
3988	continue;
3989	}
3990	const SHWPT pPTDst = (const SHWPT )PGMPOOL_PAGE_2_PTR(pVM, pPoolPage);
3991
3992	if (PdeDst.u & (X86_PDE4M_PWT \| X86_PDE4M_PCD))
3993	{
3994	AssertMsgFailed(("PDE flags PWT and/or PCD is set at %RGv! These flags are not virtualized! PdeDst=%#RX64\n",
3995	GCPtr, (uint64_t)PdeDst.u));
3996	cErrors++;
3997	}
3998
3999	if (PdeDst.u & (X86_PDE4M_G \| X86_PDE4M_D))
4000	{
4001	AssertMsgFailed(("4K PDE reserved flags at %RGv! PdeDst=%#RX64\n",
4002	GCPtr, (uint64_t)PdeDst.u));
4003	cErrors++;
4004	}
4005
4006	const GSTPDE PdeSrc = pPDSrc->a[(iPDDst >> (GST_PD_SHIFT - SHW_PD_SHIFT)) & GST_PD_MASK];
4007	if (!PdeSrc.n.u1Present)
4008	{
4009	AssertMsgFailed(("Guest PDE at %RGv is not present! PdeDst=%#RX64 PdeSrc=%#RX64\n",
4010	GCPtr, (uint64_t)PdeDst.u, (uint64_t)PdeSrc.u));
4011	cErrors++;
4012	continue;
4013	}
4014
4015	if ( !PdeSrc.b.u1Size
4016	\|\| !fBigPagesSupported)
4017	{
4018	GCPhysGst = PdeSrc.u & GST_PDE_PG_MASK;
4019	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
4020	GCPhysGst \|= (iPDDst & 1) * (PAGE_SIZE / 2);
4021	# endif
4022	}
4023	else
4024	{
4025	# if PGM_GST_TYPE == PGM_TYPE_32BIT
4026	if (PdeSrc.u & X86_PDE4M_PG_HIGH_MASK)
4027	{
4028	AssertMsgFailed(("Guest PDE at %RGv is using PSE36 or similar! PdeSrc=%#RX64\n",
4029	GCPtr, (uint64_t)PdeSrc.u));
4030	cErrors++;
4031	continue;
4032	}
4033	# endif
4034	GCPhysGst = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc);
4035	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
4036	GCPhysGst \|= GCPtr & RT_BIT(X86_PAGE_2M_SHIFT);
4037	# endif
4038	}
4039
4040	if ( pPoolPage->enmKind
4041	!= (!PdeSrc.b.u1Size \|\| !fBigPagesSupported ? BTH_PGMPOOLKIND_PT_FOR_PT : BTH_PGMPOOLKIND_PT_FOR_BIG))
4042	{
4043	AssertMsgFailed(("Invalid shadow page table kind %d at %RGv! PdeSrc=%#RX64\n",
4044	pPoolPage->enmKind, GCPtr, (uint64_t)PdeSrc.u));
4045	cErrors++;
4046	}
4047
4048	PPGMPAGE pPhysPage = pgmPhysGetPage(pPGM, GCPhysGst);
4049	if (!pPhysPage)
4050	{
4051	AssertMsgFailed(("Cannot find guest physical address %RGp in the PDE at %RGv! PdeSrc=%#RX64\n",
4052	GCPhysGst, GCPtr, (uint64_t)PdeSrc.u));
4053	cErrors++;
4054	continue;
4055	}
4056
4057	if (GCPhysGst != pPoolPage->GCPhys)
4058	{
4059	AssertMsgFailed(("GCPhysGst=%RGp != pPage->GCPhys=%RGp at %RGv\n",
4060	GCPhysGst, pPoolPage->GCPhys, GCPtr));
4061	cErrors++;
4062	continue;
4063	}
4064
4065	if ( !PdeSrc.b.u1Size
4066	\|\| !fBigPagesSupported)
4067	{
4068	/*
4069	* Page Table.
4070	*/
4071	const GSTPT *pPTSrc;
4072	rc = PGM_GCPHYS_2_PTR(pVM, GCPhysGst & ~(RTGCPHYS)(PAGE_SIZE - 1), &pPTSrc);
4073	if (RT_FAILURE(rc))
4074	{
4075	AssertMsgFailed(("Cannot map/convert guest physical address %RGp in the PDE at %RGv! PdeSrc=%#RX64\n",
4076	GCPhysGst, GCPtr, (uint64_t)PdeSrc.u));
4077	cErrors++;
4078	continue;
4079	}
4080	if ( (PdeSrc.u & (X86_PDE_P \| X86_PDE_US \| X86_PDE_RW/* \| X86_PDE_A*/))
4081	!= (PdeDst.u & (X86_PDE_P \| X86_PDE_US \| X86_PDE_RW/* \| X86_PDE_A*/)))
4082	{
4083	/// @todo We get here a lot on out-of-sync CR3 entries. The access handler should zap them to avoid false alarms here!
4084	// (This problem will go away when/if we shadow multiple CR3s.)
4085	AssertMsgFailed(("4K PDE flags mismatch at %RGv! PdeSrc=%#RX64 PdeDst=%#RX64\n",
4086	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
4087	cErrors++;
4088	continue;
4089	}
4090	if (PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY)
4091	{
4092	AssertMsgFailed(("4K PDEs cannot have PGM_PDFLAGS_TRACK_DIRTY set! GCPtr=%RGv PdeDst=%#RX64\n",
4093	GCPtr, (uint64_t)PdeDst.u));
4094	cErrors++;
4095	continue;
4096	}
4097
4098	/* iterate the page table. */
4099	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
4100	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
4101	const unsigned offPTSrc = ((GCPtr >> SHW_PD_SHIFT) & 1) * 512;
4102	# else
4103	const unsigned offPTSrc = 0;
4104	# endif
4105	for (unsigned iPT = 0, off = 0;
4106	iPT < RT_ELEMENTS(pPTDst->a);
4107	iPT++, off += PAGE_SIZE)
4108	{
4109	const SHWPTE PteDst = pPTDst->a[iPT];
4110
4111	/* skip not-present entries. */
4112	if (!(PteDst.u & (X86_PTE_P \| PGM_PTFLAGS_TRACK_DIRTY))) /** @todo deal with ALL handlers and CSAM !P pages! */
4113	continue;
4114	Assert(PteDst.n.u1Present);
4115
4116	const GSTPTE PteSrc = pPTSrc->a[iPT + offPTSrc];
4117	if (!PteSrc.n.u1Present)
4118	{
4119	# ifdef IN_RING3
4120	PGMAssertHandlerAndFlagsInSync(pVM);
4121	PGMR3DumpHierarchyGC(pVM, cr3, cr4, (PdeSrc.u & GST_PDE_PG_MASK));
4122	# endif
4123	AssertMsgFailed(("Out of sync (!P) PTE at %RGv! PteSrc=%#RX64 PteDst=%#RX64 pPTSrc=%RGv iPTSrc=%x PdeSrc=%x physpte=%RGp\n",
4124	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u, pPTSrc, iPT + offPTSrc, PdeSrc.au32[0],
4125	(PdeSrc.u & GST_PDE_PG_MASK) + (iPT + offPTSrc)*sizeof(PteSrc)));
4126	cErrors++;
4127	continue;
4128	}
4129
4130	uint64_t fIgnoreFlags = GST_PTE_PG_MASK \| X86_PTE_AVL_MASK \| X86_PTE_G \| X86_PTE_D \| X86_PTE_PWT \| X86_PTE_PCD \| X86_PTE_PAT;
4131	# if 1 /** @todo sync accessed bit properly... */
4132	fIgnoreFlags \|= X86_PTE_A;
4133	# endif
4134
4135	/* match the physical addresses */
4136	HCPhysShw = PteDst.u & SHW_PTE_PG_MASK;
4137	GCPhysGst = PteSrc.u & GST_PTE_PG_MASK;
4138
4139	# ifdef IN_RING3
4140	rc = PGMPhysGCPhys2HCPhys(pVM, GCPhysGst, &HCPhys);
4141	if (RT_FAILURE(rc))
4142	{
4143	if (HCPhysShw != MMR3PageDummyHCPhys(pVM))
4144	{
4145	AssertMsgFailed(("Cannot find guest physical address %RGp at %RGv! PteSrc=%#RX64 PteDst=%#RX64\n",
4146	GCPhysGst, GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4147	cErrors++;
4148	continue;
4149	}
4150	}
4151	else if (HCPhysShw != (HCPhys & SHW_PTE_PG_MASK))
4152	{
4153	AssertMsgFailed(("Out of sync (phys) at %RGv! HCPhysShw=%RHp HCPhys=%RHp GCPhysGst=%RGp PteSrc=%#RX64 PteDst=%#RX64\n",
4154	GCPtr + off, HCPhysShw, HCPhys, GCPhysGst, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4155	cErrors++;
4156	continue;
4157	}
4158	# endif
4159
4160	pPhysPage = pgmPhysGetPage(pPGM, GCPhysGst);
4161	if (!pPhysPage)
4162	{
4163	# ifdef IN_RING3 /** @todo make MMR3PageDummyHCPhys an 'All' function! */
4164	if (HCPhysShw != MMR3PageDummyHCPhys(pVM))
4165	{
4166	AssertMsgFailed(("Cannot find guest physical address %RGp at %RGv! PteSrc=%#RX64 PteDst=%#RX64\n",
4167	GCPhysGst, GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4168	cErrors++;
4169	continue;
4170	}
4171	# endif
4172	if (PteDst.n.u1Write)
4173	{
4174	AssertMsgFailed(("Invalid guest page at %RGv is writable! GCPhysGst=%RGp PteSrc=%#RX64 PteDst=%#RX64\n",
4175	GCPtr + off, GCPhysGst, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4176	cErrors++;
4177	}
4178	fIgnoreFlags \|= X86_PTE_RW;
4179	}
4180	else if (HCPhysShw != (PGM_PAGE_GET_HCPHYS(pPhysPage) & SHW_PTE_PG_MASK))
4181	{
4182	AssertMsgFailed(("Out of sync (phys) at %RGv! HCPhysShw=%RHp HCPhys=%RHp GCPhysGst=%RGp PteSrc=%#RX64 PteDst=%#RX64\n",
4183	GCPtr + off, HCPhysShw, pPhysPage->HCPhys, GCPhysGst, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4184	cErrors++;
4185	continue;
4186	}
4187
4188	/* flags */
4189	if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPhysPage))
4190	{
4191	if (!PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPhysPage))
4192	{
4193	if (PteDst.n.u1Write)
4194	{
4195	AssertMsgFailed(("WRITE access flagged at %RGv but the page is writable! HCPhys=%RHp PteSrc=%#RX64 PteDst=%#RX64\n",
4196	GCPtr + off, pPhysPage->HCPhys, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4197	cErrors++;
4198	continue;
4199	}
4200	fIgnoreFlags \|= X86_PTE_RW;
4201	}
4202	else
4203	{
4204	if (PteDst.n.u1Present)
4205	{
4206	AssertMsgFailed(("ALL access flagged at %RGv but the page is present! HCPhys=%RHp PteSrc=%#RX64 PteDst=%#RX64\n",
4207	GCPtr + off, pPhysPage->HCPhys, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4208	cErrors++;
4209	continue;
4210	}
4211	fIgnoreFlags \|= X86_PTE_P;
4212	}
4213	}
4214	else
4215	{
4216	if (!PteSrc.n.u1Dirty && PteSrc.n.u1Write)
4217	{
4218	if (PteDst.n.u1Write)
4219	{
4220	AssertMsgFailed(("!DIRTY page at %RGv is writable! PteSrc=%#RX64 PteDst=%#RX64\n",
4221	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4222	cErrors++;
4223	continue;
4224	}
4225	if (!(PteDst.u & PGM_PTFLAGS_TRACK_DIRTY))
4226	{
4227	AssertMsgFailed(("!DIRTY page at %RGv is not marked TRACK_DIRTY! PteSrc=%#RX64 PteDst=%#RX64\n",
4228	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4229	cErrors++;
4230	continue;
4231	}
4232	if (PteDst.n.u1Dirty)
4233	{
4234	AssertMsgFailed(("!DIRTY page at %RGv is marked DIRTY! PteSrc=%#RX64 PteDst=%#RX64\n",
4235	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4236	cErrors++;
4237	}
4238	# if 0 /** @todo sync access bit properly... */
4239	if (PteDst.n.u1Accessed != PteSrc.n.u1Accessed)
4240	{
4241	AssertMsgFailed(("!DIRTY page at %RGv is has mismatching accessed bit! PteSrc=%#RX64 PteDst=%#RX64\n",
4242	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4243	cErrors++;
4244	}
4245	fIgnoreFlags \|= X86_PTE_RW;
4246	# else
4247	fIgnoreFlags \|= X86_PTE_RW \| X86_PTE_A;
4248	# endif
4249	}
4250	else if (PteDst.u & PGM_PTFLAGS_TRACK_DIRTY)
4251	{
4252	/* access bit emulation (not implemented). */
4253	if (PteSrc.n.u1Accessed \|\| PteDst.n.u1Present)
4254	{
4255	AssertMsgFailed(("PGM_PTFLAGS_TRACK_DIRTY set at %RGv but no accessed bit emulation! PteSrc=%#RX64 PteDst=%#RX64\n",
4256	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4257	cErrors++;
4258	continue;
4259	}
4260	if (!PteDst.n.u1Accessed)
4261	{
4262	AssertMsgFailed(("!ACCESSED page at %RGv is has the accessed bit set! PteSrc=%#RX64 PteDst=%#RX64\n",
4263	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4264	cErrors++;
4265	}
4266	fIgnoreFlags \|= X86_PTE_P;
4267	}
4268	# ifdef DEBUG_sandervl
4269	fIgnoreFlags \|= X86_PTE_D \| X86_PTE_A;
4270	# endif
4271	}
4272
4273	if ( (PteSrc.u & ~fIgnoreFlags) != (PteDst.u & ~fIgnoreFlags)
4274	&& (PteSrc.u & ~(fIgnoreFlags \| X86_PTE_RW)) != (PteDst.u & ~fIgnoreFlags)
4275	)
4276	{
4277	AssertMsgFailed(("Flags mismatch at %RGv! %#RX64 != %#RX64 fIgnoreFlags=%#RX64 PteSrc=%#RX64 PteDst=%#RX64\n",
4278	GCPtr + off, (uint64_t)PteSrc.u & ~fIgnoreFlags, (uint64_t)PteDst.u & ~fIgnoreFlags,
4279	fIgnoreFlags, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4280	cErrors++;
4281	continue;
4282	}
4283	} /* foreach PTE */
4284	}
4285	else
4286	{
4287	/*
4288	* Big Page.
4289	*/
4290	uint64_t fIgnoreFlags = X86_PDE_AVL_MASK \| GST_PDE_PG_MASK \| X86_PDE4M_G \| X86_PDE4M_D \| X86_PDE4M_PS \| X86_PDE4M_PWT \| X86_PDE4M_PCD;
4291	if (!PdeSrc.b.u1Dirty && PdeSrc.b.u1Write)
4292	{
4293	if (PdeDst.n.u1Write)
4294	{
4295	AssertMsgFailed(("!DIRTY page at %RGv is writable! PdeSrc=%#RX64 PdeDst=%#RX64\n",
4296	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
4297	cErrors++;
4298	continue;
4299	}
4300	if (!(PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY))
4301	{
4302	AssertMsgFailed(("!DIRTY page at %RGv is not marked TRACK_DIRTY! PteSrc=%#RX64 PteDst=%#RX64\n",
4303	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
4304	cErrors++;
4305	continue;
4306	}
4307	# if 0 /** @todo sync access bit properly... */
4308	if (PdeDst.n.u1Accessed != PdeSrc.b.u1Accessed)
4309	{
4310	AssertMsgFailed(("!DIRTY page at %RGv is has mismatching accessed bit! PteSrc=%#RX64 PteDst=%#RX64\n",
4311	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
4312	cErrors++;
4313	}
4314	fIgnoreFlags \|= X86_PTE_RW;
4315	# else
4316	fIgnoreFlags \|= X86_PTE_RW \| X86_PTE_A;
4317	# endif
4318	}
4319	else if (PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY)
4320	{
4321	/* access bit emulation (not implemented). */
4322	if (PdeSrc.b.u1Accessed \|\| PdeDst.n.u1Present)
4323	{
4324	AssertMsgFailed(("PGM_PDFLAGS_TRACK_DIRTY set at %RGv but no accessed bit emulation! PdeSrc=%#RX64 PdeDst=%#RX64\n",
4325	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
4326	cErrors++;
4327	continue;
4328	}
4329	if (!PdeDst.n.u1Accessed)
4330	{
4331	AssertMsgFailed(("!ACCESSED page at %RGv is has the accessed bit set! PdeSrc=%#RX64 PdeDst=%#RX64\n",
4332	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
4333	cErrors++;
4334	}
4335	fIgnoreFlags \|= X86_PTE_P;
4336	}
4337
4338	if ((PdeSrc.u & ~fIgnoreFlags) != (PdeDst.u & ~fIgnoreFlags))
4339	{
4340	AssertMsgFailed(("Flags mismatch (B) at %RGv! %#RX64 != %#RX64 fIgnoreFlags=%#RX64 PdeSrc=%#RX64 PdeDst=%#RX64\n",
4341	GCPtr, (uint64_t)PdeSrc.u & ~fIgnoreFlags, (uint64_t)PdeDst.u & ~fIgnoreFlags,
4342	fIgnoreFlags, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
4343	cErrors++;
4344	}
4345
4346	/* iterate the page table. */
4347	for (unsigned iPT = 0, off = 0;
4348	iPT < RT_ELEMENTS(pPTDst->a);
4349	iPT++, off += PAGE_SIZE, GCPhysGst += PAGE_SIZE)
4350	{
4351	const SHWPTE PteDst = pPTDst->a[iPT];
4352
4353	if (PteDst.u & PGM_PTFLAGS_TRACK_DIRTY)
4354	{
4355	AssertMsgFailed(("The PTE at %RGv emulating a 2/4M page is marked TRACK_DIRTY! PdeSrc=%#RX64 PteDst=%#RX64\n",
4356	GCPtr + off, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4357	cErrors++;
4358	}
4359
4360	/* skip not-present entries. */
4361	if (!PteDst.n.u1Present) /** @todo deal with ALL handlers and CSAM !P pages! */
4362	continue;
4363
4364	fIgnoreFlags = X86_PTE_PAE_PG_MASK \| X86_PTE_AVL_MASK \| X86_PTE_PWT \| X86_PTE_PCD \| X86_PTE_PAT \| X86_PTE_D \| X86_PTE_A \| X86_PTE_G \| X86_PTE_PAE_NX;
4365
4366	/* match the physical addresses */
4367	HCPhysShw = PteDst.u & X86_PTE_PAE_PG_MASK;
4368
4369	# ifdef IN_RING3
4370	rc = PGMPhysGCPhys2HCPhys(pVM, GCPhysGst, &HCPhys);
4371	if (RT_FAILURE(rc))
4372	{
4373	if (HCPhysShw != MMR3PageDummyHCPhys(pVM))
4374	{
4375	AssertMsgFailed(("Cannot find guest physical address %RGp at %RGv! PdeSrc=%#RX64 PteDst=%#RX64\n",
4376	GCPhysGst, GCPtr + off, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4377	cErrors++;
4378	}
4379	}
4380	else if (HCPhysShw != (HCPhys & X86_PTE_PAE_PG_MASK))
4381	{
4382	AssertMsgFailed(("Out of sync (phys) at %RGv! HCPhysShw=%RHp HCPhys=%RHp GCPhysGst=%RGp PdeSrc=%#RX64 PteDst=%#RX64\n",
4383	GCPtr + off, HCPhysShw, HCPhys, GCPhysGst, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4384	cErrors++;
4385	continue;
4386	}
4387	# endif
4388	pPhysPage = pgmPhysGetPage(pPGM, GCPhysGst);
4389	if (!pPhysPage)
4390	{
4391	# ifdef IN_RING3 /** @todo make MMR3PageDummyHCPhys an 'All' function! */
4392	if (HCPhysShw != MMR3PageDummyHCPhys(pVM))
4393	{
4394	AssertMsgFailed(("Cannot find guest physical address %RGp at %RGv! PdeSrc=%#RX64 PteDst=%#RX64\n",
4395	GCPhysGst, GCPtr + off, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4396	cErrors++;
4397	continue;
4398	}
4399	# endif
4400	if (PteDst.n.u1Write)
4401	{
4402	AssertMsgFailed(("Invalid guest page at %RGv is writable! GCPhysGst=%RGp PdeSrc=%#RX64 PteDst=%#RX64\n",
4403	GCPtr + off, GCPhysGst, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4404	cErrors++;
4405	}
4406	fIgnoreFlags \|= X86_PTE_RW;
4407	}
4408	else if (HCPhysShw != (pPhysPage->HCPhys & X86_PTE_PAE_PG_MASK))
4409	{
4410	AssertMsgFailed(("Out of sync (phys) at %RGv! HCPhysShw=%RHp HCPhys=%RHp GCPhysGst=%RGp PdeSrc=%#RX64 PteDst=%#RX64\n",
4411	GCPtr + off, HCPhysShw, pPhysPage->HCPhys, GCPhysGst, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4412	cErrors++;
4413	continue;
4414	}
4415
4416	/* flags */
4417	if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPhysPage))
4418	{
4419	if (!PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPhysPage))
4420	{
4421	if (PGM_PAGE_GET_HNDL_PHYS_STATE(pPhysPage) != PGM_PAGE_HNDL_PHYS_STATE_DISABLED)
4422	{
4423	if (PteDst.n.u1Write)
4424	{
4425	AssertMsgFailed(("WRITE access flagged at %RGv but the page is writable! HCPhys=%RHp PdeSrc=%#RX64 PteDst=%#RX64\n",
4426	GCPtr + off, pPhysPage->HCPhys, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4427	cErrors++;
4428	continue;
4429	}
4430	fIgnoreFlags \|= X86_PTE_RW;
4431	}
4432	}
4433	else
4434	{
4435	if (PteDst.n.u1Present)
4436	{
4437	AssertMsgFailed(("ALL access flagged at %RGv but the page is present! HCPhys=%RHp PdeSrc=%#RX64 PteDst=%#RX64\n",
4438	GCPtr + off, pPhysPage->HCPhys, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4439	cErrors++;
4440	continue;
4441	}
4442	fIgnoreFlags \|= X86_PTE_P;
4443	}
4444	}
4445
4446	if ( (PdeSrc.u & ~fIgnoreFlags) != (PteDst.u & ~fIgnoreFlags)
4447	&& (PdeSrc.u & ~(fIgnoreFlags \| X86_PTE_RW)) != (PteDst.u & ~fIgnoreFlags) /* lazy phys handler dereg. */
4448	)
4449	{
4450	AssertMsgFailed(("Flags mismatch (BT) at %RGv! %#RX64 != %#RX64 fIgnoreFlags=%#RX64 PdeSrc=%#RX64 PteDst=%#RX64\n",
4451	GCPtr + off, (uint64_t)PdeSrc.u & ~fIgnoreFlags, (uint64_t)PteDst.u & ~fIgnoreFlags,
4452	fIgnoreFlags, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4453	cErrors++;
4454	continue;
4455	}
4456	} /* for each PTE */
4457	}
4458	}
4459	/* not present */
4460
4461	} /* for each PDE */
4462
4463	} /* for each PDPTE */
4464
4465	} /* for each PML4E */
4466
4467	# ifdef DEBUG
4468	if (cErrors)
4469	LogFlow(("AssertCR3: cErrors=%d\n", cErrors));
4470	# endif
4471
4472	#endif /* GST == 32BIT, PAE or AMD64 */
4473	return cErrors;
4474
4475	#endif /* PGM_SHW_TYPE != PGM_TYPE_NESTED && PGM_SHW_TYPE != PGM_TYPE_EPT */
4476	}
4477	#endif /* VBOX_STRICT */
4478
4479
4480	/**
4481	* Sets up the CR3 for shadow paging
4482	*
4483	* @returns Strict VBox status code.
4484	* @retval VINF_SUCCESS.
4485	*
4486	* @param pVM VM handle.
4487	* @param GCPhysCR3 The physical address in the CR3 register.
4488	*/
4489	PGM_BTH_DECL(int, MapCR3)(PVM pVM, RTGCPHYS GCPhysCR3)
4490	{
4491	/* Update guest paging info. */
4492	#if PGM_GST_TYPE == PGM_TYPE_32BIT \
4493	\|\| PGM_GST_TYPE == PGM_TYPE_PAE \
4494	\|\| PGM_GST_TYPE == PGM_TYPE_AMD64
4495
4496	LogFlow(("MapCR3: %RGp\n", GCPhysCR3));
4497
4498	/*
4499	* Map the page CR3 points at.
4500	*/
4501	RTHCPHYS HCPhysGuestCR3;
4502	RTHCPTR HCPtrGuestCR3;
4503	int rc = pgmRamGCPhys2HCPtrAndHCPhysWithFlags(&pVM->pgm.s, GCPhysCR3 & GST_CR3_PAGE_MASK, &HCPtrGuestCR3, &HCPhysGuestCR3);
4504	if (RT_SUCCESS(rc))
4505	{
4506	rc = PGMMap(pVM, (RTGCPTR)pVM->pgm.s.GCPtrCR3Mapping, HCPhysGuestCR3, PAGE_SIZE, 0);
4507	if (RT_SUCCESS(rc))
4508	{
4509	PGM_INVL_PG(pVM->pgm.s.GCPtrCR3Mapping);
4510	# if PGM_GST_TYPE == PGM_TYPE_32BIT
4511	pVM->pgm.s.pGst32BitPdR3 = (R3PTRTYPE(PX86PD))HCPtrGuestCR3;
4512	# ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4513	pVM->pgm.s.pGst32BitPdR0 = (R0PTRTYPE(PX86PD))HCPtrGuestCR3;
4514	# endif
4515	pVM->pgm.s.pGst32BitPdRC = (RCPTRTYPE(PX86PD))pVM->pgm.s.GCPtrCR3Mapping;
4516
4517	# elif PGM_GST_TYPE == PGM_TYPE_PAE
4518	unsigned off = GCPhysCR3 & GST_CR3_PAGE_MASK & PAGE_OFFSET_MASK;
4519	pVM->pgm.s.pGstPaePdptR3 = (R3PTRTYPE(PX86PDPT))HCPtrGuestCR3;
4520	# ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4521	pVM->pgm.s.pGstPaePdptR0 = (R0PTRTYPE(PX86PDPT))HCPtrGuestCR3;
4522	# endif
4523	pVM->pgm.s.pGstPaePdptRC = (RCPTRTYPE(PX86PDPT))((RCPTRTYPE(uint8_t *))pVM->pgm.s.GCPtrCR3Mapping + off);
4524	Log(("Cached mapping %RGv\n", pVM->pgm.s.pGstPaePdptRC));
4525
4526	/*
4527	* Map the 4 PDs too.
4528	*/
4529	PX86PDPT pGuestPDPT = pgmGstGetPaePDPTPtr(&pVM->pgm.s);
4530	RTGCPTR GCPtr = pVM->pgm.s.GCPtrCR3Mapping + PAGE_SIZE;
4531	for (unsigned i = 0; i < X86_PG_PAE_PDPE_ENTRIES; i++, GCPtr += PAGE_SIZE)
4532	{
4533	if (pGuestPDPT->a[i].n.u1Present)
4534	{
4535	RTHCPTR HCPtr;
4536	RTHCPHYS HCPhys;
4537	RTGCPHYS GCPhys = pGuestPDPT->a[i].u & X86_PDPE_PG_MASK;
4538	int rc2 = pgmRamGCPhys2HCPtrAndHCPhysWithFlags(&pVM->pgm.s, GCPhys, &HCPtr, &HCPhys);
4539	if (RT_SUCCESS(rc2))
4540	{
4541	rc = PGMMap(pVM, GCPtr, HCPhys & X86_PTE_PAE_PG_MASK, PAGE_SIZE, 0);
4542	AssertRCReturn(rc, rc);
4543
4544	pVM->pgm.s.apGstPaePDsR3[i] = (R3PTRTYPE(PX86PDPAE))HCPtr;
4545	# ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4546	pVM->pgm.s.apGstPaePDsR0[i] = (R0PTRTYPE(PX86PDPAE))HCPtr;
4547	# endif
4548	pVM->pgm.s.apGstPaePDsRC[i] = (RCPTRTYPE(PX86PDPAE))GCPtr;
4549	pVM->pgm.s.aGCPhysGstPaePDs[i] = GCPhys;
4550	PGM_INVL_PG(GCPtr); /** @todo This ends up calling HWACCMInvalidatePage, is that correct? */
4551	continue;
4552	}
4553	AssertMsgFailed(("pgmR3Gst32BitMapCR3: rc2=%d GCPhys=%RGp i=%d\n", rc2, GCPhys, i));
4554	}
4555
4556	pVM->pgm.s.apGstPaePDsR3[i] = 0;
4557	# ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4558	pVM->pgm.s.apGstPaePDsR0[i] = 0;
4559	# endif
4560	pVM->pgm.s.apGstPaePDsRC[i] = 0;
4561	pVM->pgm.s.aGCPhysGstPaePDs[i] = NIL_RTGCPHYS;
4562	PGM_INVL_PG(GCPtr); /** @todo this shouldn't be necessary? */
4563	}
4564
4565	# elif PGM_GST_TYPE == PGM_TYPE_AMD64
4566	pVM->pgm.s.pGstAmd64Pml4R3 = (R3PTRTYPE(PX86PML4))HCPtrGuestCR3;
4567	# ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4568	pVM->pgm.s.pGstAmd64Pml4R0 = (R0PTRTYPE(PX86PML4))HCPtrGuestCR3;
4569	# endif
4570	# ifndef VBOX_WITH_PGMPOOL_PAGING_ONLY
4571	if (!HWACCMIsNestedPagingActive(pVM))
4572	{
4573	/*
4574	* Update the shadow root page as well since that's not fixed.
4575	*/
4576	/** @todo Move this into PGMAllBth.h. */
4577	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
4578	if (pVM->pgm.s.CTX_SUFF(pShwPageCR3))
4579	{
4580	/* It might have been freed already by a pool flush (see e.g. PGMR3MappingsUnfix). */
4581	/** @todo Coordinate this better with the pool. */
4582	if (pVM->pgm.s.CTX_SUFF(pShwPageCR3)->enmKind != PGMPOOLKIND_FREE)
4583	pgmPoolFreeByPage(pPool, pVM->pgm.s.CTX_SUFF(pShwPageCR3), PGMPOOL_IDX_AMD64_CR3, pVM->pgm.s.CTX_SUFF(pShwPageCR3)->GCPhys >> PAGE_SHIFT);
4584	pVM->pgm.s.pShwPageCR3R3 = 0;
4585	pVM->pgm.s.pShwPageCR3R0 = 0;
4586	pVM->pgm.s.pShwRootR3 = 0;
4587	# ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4588	pVM->pgm.s.pShwRootR0 = 0;
4589	# endif
4590	pVM->pgm.s.HCPhysShwCR3 = 0;
4591	}
4592
4593	Assert(!(GCPhysCR3 >> (PAGE_SHIFT + 32)));
4594	rc = pgmPoolAlloc(pVM, GCPhysCR3, PGMPOOLKIND_64BIT_PML4, PGMPOOL_IDX_AMD64_CR3, GCPhysCR3 >> PAGE_SHIFT, &pVM->pgm.s.CTX_SUFF(pShwPageCR3));
4595	if (rc == VERR_PGM_POOL_FLUSHED)
4596	{
4597	Log(("MapCR3: PGM pool flushed -> signal sync cr3\n"));
4598	Assert(VM_FF_ISSET(pVM, VM_FF_PGM_SYNC_CR3));
4599	return VINF_PGM_SYNC_CR3;
4600	}
4601	AssertRCReturn(rc, rc);
4602	# ifdef IN_RING0
4603	pVM->pgm.s.pShwPageCR3R3 = MMHyperCCToR3(pVM, pVM->pgm.s.CTX_SUFF(pShwPageCR3));
4604	# else
4605	pVM->pgm.s.pShwPageCR3R0 = MMHyperCCToR0(pVM, pVM->pgm.s.CTX_SUFF(pShwPageCR3));
4606	# endif
4607	pVM->pgm.s.pShwRootR3 = (R3PTRTYPE(void *))pVM->pgm.s.CTX_SUFF(pShwPageCR3)->pvPageR3;
4608	Assert(pVM->pgm.s.pShwRootR3);
4609	# ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4610	pVM->pgm.s.pShwRootR0 = (R0PTRTYPE(void *))PGMPOOL_PAGE_2_PTR(pPool->CTX_SUFF(pVM), pVM->pgm.s.CTX_SUFF(pShwPageCR3));
4611	# endif
4612	pVM->pgm.s.HCPhysShwCR3 = pVM->pgm.s.CTX_SUFF(pShwPageCR3)->Core.Key;
4613	rc = VINF_SUCCESS; /* clear it - pgmPoolAlloc returns hints. */
4614	}
4615	# endif /* !VBOX_WITH_PGMPOOL_PAGING_ONLY */
4616	# endif
4617	}
4618	else
4619	AssertMsgFailed(("rc=%Rrc GCPhysGuestPD=%RGp\n", rc, GCPhysCR3));
4620	}
4621	else
4622	AssertMsgFailed(("rc=%Rrc GCPhysGuestPD=%RGp\n", rc, GCPhysCR3));
4623
4624	#else /* prot/real stub */
4625	int rc = VINF_SUCCESS;
4626	#endif
4627
4628	#ifdef VBOX_WITH_PGMPOOL_PAGING_ONLY
4629	/* Update shadow paging info for guest modes with paging (32, pae, 64). */
4630	# if ( ( PGM_SHW_TYPE == PGM_TYPE_32BIT \
4631	\|\| PGM_SHW_TYPE == PGM_TYPE_PAE \
4632	\|\| PGM_SHW_TYPE == PGM_TYPE_AMD64) \
4633	&& ( PGM_GST_TYPE != PGM_TYPE_REAL \
4634	&& PGM_GST_TYPE != PGM_TYPE_PROT))
4635
4636	Assert(!HWACCMIsNestedPagingActive(pVM));
4637
4638	/*
4639	* Update the shadow root page as well since that's not fixed.
4640	*/
4641	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
4642	PPGMPOOLPAGE pOldShwPageCR3 = pVM->pgm.s.CTX_SUFF(pShwPageCR3);
4643	uint32_t iOldShwUserTable = pVM->pgm.s.iShwUserTable;
4644	uint32_t iOldShwUser = pVM->pgm.s.iShwUser;
4645	PPGMPOOLPAGE pNewShwPageCR3;
4646
4647	Assert(!(GCPhysCR3 >> (PAGE_SHIFT + 32)));
4648	rc = pgmPoolAlloc(pVM, GCPhysCR3 & GST_CR3_PAGE_MASK, BTH_PGMPOOLKIND_ROOT, SHW_POOL_ROOT_IDX, GCPhysCR3 >> PAGE_SHIFT, &pNewShwPageCR3);
4649	if (rc == VERR_PGM_POOL_FLUSHED)
4650	{
4651	Log(("MapCR3: PGM pool flushed -> signal sync cr3\n"));
4652	Assert(VM_FF_ISSET(pVM, VM_FF_PGM_SYNC_CR3));
4653	return VINF_PGM_SYNC_CR3;
4654	}
4655	AssertRCReturn(rc, rc);
4656	rc = VINF_SUCCESS;
4657
4658	# ifdef IN_RC
4659	/** NOTE: We can't deal with jumps to ring 3 here as we're now in an inconsistent state! */
4660	# endif
4661	/* Mark the page as locked; disallow flushing. */
4662	pgmPoolLockPage(pPool, pNewShwPageCR3);
4663
4664	pVM->pgm.s.iShwUser = SHW_POOL_ROOT_IDX;
4665	pVM->pgm.s.iShwUserTable = GCPhysCR3 >> PAGE_SHIFT;
4666	pVM->pgm.s.CTX_SUFF(pShwPageCR3) = pNewShwPageCR3;
4667	# ifdef IN_RING0
4668	pVM->pgm.s.pShwPageCR3R3 = MMHyperCCToR3(pVM, pVM->pgm.s.CTX_SUFF(pShwPageCR3));
4669	pVM->pgm.s.pShwPageCR3RC = MMHyperCCToRC(pVM, pVM->pgm.s.CTX_SUFF(pShwPageCR3));
4670	# elif defined(IN_RC)
4671	pVM->pgm.s.pShwPageCR3R3 = MMHyperCCToR3(pVM, pVM->pgm.s.CTX_SUFF(pShwPageCR3));
4672	pVM->pgm.s.pShwPageCR3R0 = MMHyperCCToR0(pVM, pVM->pgm.s.CTX_SUFF(pShwPageCR3));
4673	# else
4674	pVM->pgm.s.pShwPageCR3R0 = MMHyperCCToR0(pVM, pVM->pgm.s.CTX_SUFF(pShwPageCR3));
4675	pVM->pgm.s.pShwPageCR3RC = MMHyperCCToRC(pVM, pVM->pgm.s.CTX_SUFF(pShwPageCR3));
4676	# endif
4677	pVM->pgm.s.pShwRootR3 = (R3PTRTYPE(void *))pVM->pgm.s.CTX_SUFF(pShwPageCR3)->pvPageR3;
4678	Assert(pVM->pgm.s.pShwRootR3);
4679	# ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4680	pVM->pgm.s.pShwRootR0 = (R0PTRTYPE(void *))PGMPOOL_PAGE_2_PTR(pPool->CTX_SUFF(pVM), pVM->pgm.s.CTX_SUFF(pShwPageCR3));
4681	# endif
4682	pVM->pgm.s.HCPhysShwCR3 = pVM->pgm.s.CTX_SUFF(pShwPageCR3)->Core.Key;
4683
4684	# ifndef PGM_WITHOUT_MAPPINGS
4685	/* Apply all hypervisor mappings to the new CR3.
4686	* Note that SyncCR3 will be executed in case CR3 is changed in a guest paging mode; this will
4687	* make sure we check for conflicts in the new CR3 root.
4688	*/
4689	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
4690	Assert(VM_FF_ISSET(pVM, VM_FF_PGM_SYNC_CR3_NON_GLOBAL));
4691	# endif
4692	rc = pgmMapActivateCR3(pVM, pNewShwPageCR3);
4693	AssertRCReturn(rc, rc);
4694	# endif
4695
4696	/* Set the current hypervisor CR3. */
4697	CPUMSetHyperCR3(pVM, PGMGetHyperCR3(pVM));
4698
4699	# ifdef IN_RC
4700	/** NOTE: Everything safe again. */
4701	# endif
4702
4703	/* Clean up the old CR3 root. */
4704	if (pOldShwPageCR3)
4705	{
4706	Assert(pOldShwPageCR3->enmKind != PGMPOOLKIND_FREE);
4707	# ifndef PGM_WITHOUT_MAPPINGS
4708	/* Remove the hypervisor mappings from the shadow page table. */
4709	pgmMapDeactivateCR3(pVM, pOldShwPageCR3);
4710	# endif
4711	/* Mark the page as unlocked; allow flushing again. */
4712	pgmPoolUnlockPage(pPool, pOldShwPageCR3);
4713
4714	pgmPoolFreeByPage(pPool, pOldShwPageCR3, iOldShwUser, iOldShwUserTable);
4715	}
4716
4717	# endif
4718	#endif /* VBOX_WITH_PGMPOOL_PAGING_ONLY */
4719
4720	return rc;
4721	}
4722
4723	/**
4724	* Unmaps the shadow CR3.
4725	*
4726	* @returns VBox status, no specials.
4727	* @param pVM VM handle.
4728	*/
4729	PGM_BTH_DECL(int, UnmapCR3)(PVM pVM)
4730	{
4731	LogFlow(("UnmapCR3\n"));
4732
4733	int rc = VINF_SUCCESS;
4734
4735	/* Update guest paging info. */
4736	#if PGM_GST_TYPE == PGM_TYPE_32BIT
4737	pVM->pgm.s.pGst32BitPdR3 = 0;
4738	#ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4739	pVM->pgm.s.pGst32BitPdR0 = 0;
4740	#endif
4741	pVM->pgm.s.pGst32BitPdRC = 0;
4742
4743	#elif PGM_GST_TYPE == PGM_TYPE_PAE
4744	pVM->pgm.s.pGstPaePdptR3 = 0;
4745	# ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4746	pVM->pgm.s.pGstPaePdptR0 = 0;
4747	# endif
4748	pVM->pgm.s.pGstPaePdptRC = 0;
4749	for (unsigned i = 0; i < X86_PG_PAE_PDPE_ENTRIES; i++)
4750	{
4751	pVM->pgm.s.apGstPaePDsR3[i] = 0;
4752	# ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4753	pVM->pgm.s.apGstPaePDsR0[i] = 0;
4754	# endif
4755	pVM->pgm.s.apGstPaePDsRC[i] = 0;
4756	pVM->pgm.s.aGCPhysGstPaePDs[i] = NIL_RTGCPHYS;
4757	}
4758
4759	#elif PGM_GST_TYPE == PGM_TYPE_AMD64
4760	pVM->pgm.s.pGstAmd64Pml4R3 = 0;
4761	# ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4762	pVM->pgm.s.pGstAmd64Pml4R0 = 0;
4763	# endif
4764	# ifndef VBOX_WITH_PGMPOOL_PAGING_ONLY
4765	if (!HWACCMIsNestedPagingActive(pVM))
4766	{
4767	pVM->pgm.s.pShwRootR3 = 0;
4768	# ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4769	pVM->pgm.s.pShwRootR0 = 0;
4770	# endif
4771	pVM->pgm.s.HCPhysShwCR3 = 0;
4772	if (pVM->pgm.s.CTX_SUFF(pShwPageCR3))
4773	{
4774	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
4775	pgmPoolFreeByPage(pPool, pVM->pgm.s.CTX_SUFF(pShwPageCR3), PGMPOOL_IDX_AMD64_CR3, pVM->pgm.s.CTX_SUFF(pShwPageCR3)->GCPhys >> PAGE_SHIFT);
4776	pVM->pgm.s.pShwPageCR3R3 = 0;
4777	pVM->pgm.s.pShwPageCR3R0 = 0;
4778	}
4779	}
4780	# endif /* !VBOX_WITH_PGMPOOL_PAGING_ONLY */
4781
4782	#else /* prot/real mode stub */
4783	/* nothing to do */
4784	#endif
4785
4786	#if defined(VBOX_WITH_PGMPOOL_PAGING_ONLY) && !defined(IN_RC) /* In RC we rely on MapCR3 to do the shadow part for us at a safe time */
4787	/* Update shadow paging info. */
4788	# if ( ( PGM_SHW_TYPE == PGM_TYPE_32BIT \
4789	\|\| PGM_SHW_TYPE == PGM_TYPE_PAE \
4790	\|\| PGM_SHW_TYPE == PGM_TYPE_AMD64))
4791
4792	Assert(!HWACCMIsNestedPagingActive(pVM));
4793
4794	# ifndef PGM_WITHOUT_MAPPINGS
4795	if (pVM->pgm.s.CTX_SUFF(pShwPageCR3))
4796	/* Remove the hypervisor mappings from the shadow page table. */
4797	pgmMapDeactivateCR3(pVM, pVM->pgm.s.CTX_SUFF(pShwPageCR3));
4798	# endif
4799
4800	pVM->pgm.s.pShwRootR3 = 0;
4801	# ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4802	pVM->pgm.s.pShwRootR0 = 0;
4803	# endif
4804	pVM->pgm.s.HCPhysShwCR3 = 0;
4805	if (pVM->pgm.s.CTX_SUFF(pShwPageCR3))
4806	{
4807	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
4808
4809	/* Mark the page as unlocked; allow flushing again. */
4810	pgmPoolUnlockPage(pPool, pVM->pgm.s.CTX_SUFF(pShwPageCR3));
4811
4812	pgmPoolFreeByPage(pPool, pVM->pgm.s.CTX_SUFF(pShwPageCR3), pVM->pgm.s.iShwUser, pVM->pgm.s.iShwUserTable);
4813	pVM->pgm.s.pShwPageCR3R3 = 0;
4814	pVM->pgm.s.pShwPageCR3R0 = 0;
4815	pVM->pgm.s.iShwUser = 0;
4816	pVM->pgm.s.iShwUserTable = 0;
4817	}
4818	# endif
4819	#endif /* VBOX_WITH_PGMPOOL_PAGING_ONLY && !IN_RC*/
4820
4821	return rc;
4822	}
4823

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

source: vbox/trunk/src/VBox/VMM/VMMAll/PGMAllBth.h@ 17142

Download in other formats: