PGMAllBth.h@ 15410

Last change on this file since 15410 was 15410, checked in by vboxsync, 16 years ago
PGM: Deal with pgmPoolAlloc failure in MapCR3 without relying on having to clear the pool. The MapCR3 action will be postponed to SyncCR3.
Property svn:eol-style set to `native` Property svn:keywords set to `Id`
File size: 187.8 KB

Line
1	/* $Id: PGMAllBth.h 15410 2008-12-13 01:04:17Z 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	__END_DECLS
43
44
45	/* Filter out some illegal combinations of guest and shadow paging, so we can remove redundant checks inside functions. */
46	#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
47	# error "Invalid combination; PAE guest implies PAE shadow"
48	#endif
49
50	#if (PGM_GST_TYPE == PGM_TYPE_REAL \|\| PGM_GST_TYPE == PGM_TYPE_PROT) \
51	&& !(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)
52	# error "Invalid combination; real or protected mode without paging implies 32 bits or PAE shadow paging."
53	#endif
54
55	#if (PGM_GST_TYPE == PGM_TYPE_32BIT \|\| PGM_GST_TYPE == PGM_TYPE_PAE) \
56	&& !(PGM_SHW_TYPE == PGM_TYPE_32BIT \|\| PGM_SHW_TYPE == PGM_TYPE_PAE \|\| PGM_SHW_TYPE == PGM_TYPE_NESTED \|\| PGM_SHW_TYPE == PGM_TYPE_EPT)
57	# error "Invalid combination; 32 bits guest paging or PAE implies 32 bits or PAE shadow paging."
58	#endif
59
60	#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) \
61	\|\| (PGM_SHW_TYPE == PGM_TYPE_AMD64 && PGM_GST_TYPE != PGM_TYPE_AMD64 && PGM_GST_TYPE != PGM_TYPE_PROT)
62	# error "Invalid combination; AMD64 guest implies AMD64 shadow and vice versa"
63	#endif
64
65	#ifdef IN_RING0 /* no mappings in VT-x and AMD-V mode */
66	# define PGM_WITHOUT_MAPPINGS
67	#endif
68
69
70	#ifndef IN_RING3
71	/**
72	* #PF Handler for raw-mode guest execution.
73	*
74	* @returns VBox status code (appropriate for trap handling and GC return).
75	* @param pVM VM Handle.
76	* @param uErr The trap error code.
77	* @param pRegFrame Trap register frame.
78	* @param pvFault The fault address.
79	*/
80	PGM_BTH_DECL(int, Trap0eHandler)(PVM pVM, RTGCUINT uErr, PCPUMCTXCORE pRegFrame, RTGCPTR pvFault)
81	{
82	# 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) \
83	&& PGM_SHW_TYPE != PGM_TYPE_NESTED \
84	&& (PGM_SHW_TYPE != PGM_TYPE_EPT \|\| PGM_GST_TYPE == PGM_TYPE_PROT)
85
86	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE != PGM_TYPE_PAE
87	/*
88	* Hide the instruction fetch trap indicator for now.
89	*/
90	/** @todo NXE will change this and we must fix NXE in the switcher too! */
91	if (uErr & X86_TRAP_PF_ID)
92	{
93	uErr &= ~X86_TRAP_PF_ID;
94	TRPMSetErrorCode(pVM, uErr);
95	}
96	# endif
97
98	/*
99	* Get PDs.
100	*/
101	int rc;
102	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
103	# if PGM_GST_TYPE == PGM_TYPE_32BIT
104	const unsigned iPDSrc = pvFault >> GST_PD_SHIFT;
105	PGSTPD pPDSrc = pgmGstGet32bitPDPtr(&pVM->pgm.s);
106
107	# elif PGM_GST_TYPE == PGM_TYPE_PAE \|\| PGM_GST_TYPE == PGM_TYPE_AMD64
108
109	# if PGM_GST_TYPE == PGM_TYPE_PAE
110	unsigned iPDSrc;
111	PGSTPD pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, pvFault, &iPDSrc, NULL);
112
113	# elif PGM_GST_TYPE == PGM_TYPE_AMD64
114	unsigned iPDSrc;
115	PX86PML4E pPml4eSrc;
116	X86PDPE PdpeSrc;
117	PGSTPD pPDSrc;
118
119	pPDSrc = pgmGstGetLongModePDPtr(&pVM->pgm.s, pvFault, &pPml4eSrc, &PdpeSrc, &iPDSrc);
120	Assert(pPml4eSrc);
121	# endif
122	/* Quick check for a valid guest trap. */
123	if (!pPDSrc)
124	{
125	# if PGM_GST_TYPE == PGM_TYPE_AMD64 && GC_ARCH_BITS == 64
126	LogFlow(("Trap0eHandler: guest PML4 %d not present CR3=%RGp\n", (int)((pvFault >> X86_PML4_SHIFT) & X86_PML4_MASK), CPUMGetGuestCR3(pVM) & X86_CR3_PAGE_MASK));
127	# else
128	LogFlow(("Trap0eHandler: guest iPDSrc=%u not present CR3=%RGp\n", iPDSrc, CPUMGetGuestCR3(pVM) & X86_CR3_PAGE_MASK));
129	# endif
130	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2GuestTrap; });
131	TRPMSetErrorCode(pVM, uErr);
132	return VINF_EM_RAW_GUEST_TRAP;
133	}
134	# endif
135
136	# else /* !PGM_WITH_PAGING */
137	PGSTPD pPDSrc = NULL;
138	const unsigned iPDSrc = 0;
139	# endif /* !PGM_WITH_PAGING */
140
141
142	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
143	const unsigned iPDDst = pvFault >> SHW_PD_SHIFT;
144	PX86PD pPDDst = pgmShwGet32BitPDPtr(&pVM->pgm.s);
145
146	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
147	const unsigned iPDDst = (pvFault >> SHW_PD_SHIFT) & SHW_PD_MASK; /* pPDDst index, not used with the pool. */
148	PX86PDPAE pPDDst = pgmShwGetPaePDPtr(&pVM->pgm.s, pvFault);
149
150	/* Did we mark the PDPT as not present in SyncCR3? */
151	unsigned iPdpt = (pvFault >> SHW_PDPT_SHIFT) & SHW_PDPT_MASK;
152	PX86PDPT pPdptDst = pgmShwGetPaePDPTPtr(&pVM->pgm.s);
153	if (!pPdptDst->a[iPdpt].n.u1Present)
154	pPdptDst->a[iPdpt].n.u1Present = 1;
155
156	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
157	const unsigned iPDDst = ((pvFault >> SHW_PD_SHIFT) & SHW_PD_MASK);
158	PX86PDPAE pPDDst;
159	# if PGM_GST_TYPE == PGM_TYPE_PROT
160	/* AMD-V nested paging */
161	X86PML4E Pml4eSrc;
162	X86PDPE PdpeSrc;
163	PX86PML4E pPml4eSrc = &Pml4eSrc;
164
165	/* Fake PML4 & PDPT entry; access control handled on the page table level, so allow everything. */
166	Pml4eSrc.u = X86_PML4E_P \| X86_PML4E_RW \| X86_PML4E_US \| X86_PML4E_A;
167	PdpeSrc.u = X86_PDPE_P \| X86_PDPE_RW \| X86_PDPE_US \| X86_PDPE_A;
168	# endif
169
170	rc = pgmShwSyncLongModePDPtr(pVM, pvFault, pPml4eSrc, &PdpeSrc, &pPDDst);
171	if (rc != VINF_SUCCESS)
172	{
173	AssertRC(rc);
174	return rc;
175	}
176	Assert(pPDDst);
177
178	# elif PGM_SHW_TYPE == PGM_TYPE_EPT
179	const unsigned iPDDst = ((pvFault >> SHW_PD_SHIFT) & SHW_PD_MASK);
180	PEPTPD pPDDst;
181
182	rc = pgmShwGetEPTPDPtr(pVM, pvFault, NULL, &pPDDst);
183	if (rc != VINF_SUCCESS)
184	{
185	AssertRC(rc);
186	return rc;
187	}
188	Assert(pPDDst);
189	# endif
190
191	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
192	/*
193	* If we successfully correct the write protection fault due to dirty bit
194	* tracking, or this page fault is a genuine one, then return immediately.
195	*/
196	STAM_PROFILE_START(&pVM->pgm.s.StatRZTrap0eTimeCheckPageFault, e);
197	rc = PGM_BTH_NAME(CheckPageFault)(pVM, uErr, &pPDDst->a[iPDDst], &pPDSrc->a[iPDSrc], pvFault);
198	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeCheckPageFault, e);
199	if ( rc == VINF_PGM_HANDLED_DIRTY_BIT_FAULT
200	\|\| rc == VINF_EM_RAW_GUEST_TRAP)
201	{
202	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution)
203	= rc == VINF_PGM_HANDLED_DIRTY_BIT_FAULT ? &pVM->pgm.s.StatRZTrap0eTime2DirtyAndAccessed : &pVM->pgm.s.StatRZTrap0eTime2GuestTrap; });
204	LogBird(("Trap0eHandler: returns %s\n", rc == VINF_PGM_HANDLED_DIRTY_BIT_FAULT ? "VINF_SUCCESS" : "VINF_EM_RAW_GUEST_TRAP"));
205	return rc == VINF_PGM_HANDLED_DIRTY_BIT_FAULT ? VINF_SUCCESS : rc;
206	}
207
208	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0ePD[iPDSrc]);
209	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
210
211	/*
212	* A common case is the not-present error caused by lazy page table syncing.
213	*
214	* It is IMPORTANT that we weed out any access to non-present shadow PDEs here
215	* so we can safely assume that the shadow PT is present when calling SyncPage later.
216	*
217	* On failure, we ASSUME that SyncPT is out of memory or detected some kind
218	* of mapping conflict and defer to SyncCR3 in R3.
219	* (Again, we do NOT support access handlers for non-present guest pages.)
220	*
221	*/
222	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
223	GSTPDE PdeSrc = pPDSrc->a[iPDSrc];
224	# else
225	GSTPDE PdeSrc;
226	PdeSrc.au32[0] = 0; /* faked so we don't have to #ifdef everything */
227	PdeSrc.n.u1Present = 1;
228	PdeSrc.n.u1Write = 1;
229	PdeSrc.n.u1Accessed = 1;
230	PdeSrc.n.u1User = 1;
231	# endif
232	if ( !(uErr & X86_TRAP_PF_P) /* not set means page not present instead of page protection violation */
233	&& !pPDDst->a[iPDDst].n.u1Present
234	&& PdeSrc.n.u1Present
235	)
236
237	{
238	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2SyncPT; });
239	STAM_PROFILE_START(&pVM->pgm.s.StatRZTrap0eTimeSyncPT, f);
240	LogFlow(("=>SyncPT %04x = %08x\n", iPDSrc, PdeSrc.au32[0]));
241	rc = PGM_BTH_NAME(SyncPT)(pVM, iPDSrc, pPDSrc, pvFault);
242	if (RT_SUCCESS(rc))
243	{
244	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeSyncPT, f);
245	return rc;
246	}
247	Log(("SyncPT: %d failed!! rc=%d\n", iPDSrc, rc));
248	VM_FF_SET(pVM, VM_FF_PGM_SYNC_CR3); /** @todo no need to do global sync, right? */
249	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeSyncPT, f);
250	return VINF_PGM_SYNC_CR3;
251	}
252
253	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
254	/*
255	* Check if this address is within any of our mappings.
256	*
257	* This is very fast and it's gonna save us a bit of effort below and prevent
258	* us from screwing ourself with MMIO2 pages which have a GC Mapping (VRam).
259	* (BTW, it's impossible to have physical access handlers in a mapping.)
260	*/
261	if (pgmMapAreMappingsEnabled(&pVM->pgm.s))
262	{
263	STAM_PROFILE_START(&pVM->pgm.s.StatRZTrap0eTimeMapping, a);
264	PPGMMAPPING pMapping = pVM->pgm.s.CTX_SUFF(pMappings);
265	for ( ; pMapping; pMapping = pMapping->CTX_SUFF(pNext))
266	{
267	if (pvFault < pMapping->GCPtr)
268	break;
269	if (pvFault - pMapping->GCPtr < pMapping->cb)
270	{
271	/*
272	* The first thing we check is if we've got an undetected conflict.
273	*/
274	if (!pVM->pgm.s.fMappingsFixed)
275	{
276	unsigned iPT = pMapping->cb >> GST_PD_SHIFT;
277	while (iPT-- > 0)
278	if (pPDSrc->a[iPDSrc + iPT].n.u1Present)
279	{
280	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eConflicts);
281	Log(("Trap0e: Detected Conflict %RGv-%RGv\n", pMapping->GCPtr, pMapping->GCPtrLast));
282	VM_FF_SET(pVM, VM_FF_PGM_SYNC_CR3); /** @todo no need to do global sync,right? */
283	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeMapping, a);
284	return VINF_PGM_SYNC_CR3;
285	}
286	}
287
288	/*
289	* Check if the fault address is in a virtual page access handler range.
290	*/
291	PPGMVIRTHANDLER pCur = (PPGMVIRTHANDLER)RTAvlroGCPtrRangeGet(&pVM->pgm.s.CTX_SUFF(pTrees)->HyperVirtHandlers, pvFault);
292	if ( pCur
293	&& pvFault - pCur->Core.Key < pCur->cb
294	&& uErr & X86_TRAP_PF_RW)
295	{
296	# ifdef IN_RC
297	STAM_PROFILE_START(&pCur->Stat, h);
298	rc = pCur->CTX_SUFF(pfnHandler)(pVM, uErr, pRegFrame, pvFault, pCur->Core.Key, pvFault - pCur->Core.Key);
299	STAM_PROFILE_STOP(&pCur->Stat, h);
300	# else
301	AssertFailed();
302	rc = VINF_EM_RAW_EMULATE_INSTR; /* can't happen with VMX */
303	# endif
304	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersMapping);
305	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeMapping, a);
306	return rc;
307	}
308
309	/*
310	* Pretend we're not here and let the guest handle the trap.
311	*/
312	TRPMSetErrorCode(pVM, uErr & ~X86_TRAP_PF_P);
313	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eGuestPFMapping);
314	LogFlow(("PGM: Mapping access -> route trap to recompiler!\n"));
315	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeMapping, a);
316	return VINF_EM_RAW_GUEST_TRAP;
317	}
318	}
319	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeMapping, a);
320	} /* pgmAreMappingsEnabled(&pVM->pgm.s) */
321	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
322
323	/*
324	* Check if this fault address is flagged for special treatment,
325	* which means we'll have to figure out the physical address and
326	* check flags associated with it.
327	*
328	* ASSUME that we can limit any special access handling to pages
329	* in page tables which the guest believes to be present.
330	*/
331	if (PdeSrc.n.u1Present)
332	{
333	RTGCPHYS GCPhys = NIL_RTGCPHYS;
334
335	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
336	# if PGM_GST_TYPE == PGM_TYPE_AMD64
337	bool fBigPagesSupported = true;
338	# else
339	bool fBigPagesSupported = !!(CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
340	# endif
341	if ( PdeSrc.b.u1Size
342	&& fBigPagesSupported)
343	GCPhys = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc)
344	\| ((RTGCPHYS)pvFault & (GST_BIG_PAGE_OFFSET_MASK ^ PAGE_OFFSET_MASK));
345	else
346	{
347	PGSTPT pPTSrc;
348	rc = PGM_GCPHYS_2_PTR(pVM, PdeSrc.u & GST_PDE_PG_MASK, &pPTSrc);
349	if (RT_SUCCESS(rc))
350	{
351	unsigned iPTESrc = (pvFault >> GST_PT_SHIFT) & GST_PT_MASK;
352	if (pPTSrc->a[iPTESrc].n.u1Present)
353	GCPhys = pPTSrc->a[iPTESrc].u & GST_PTE_PG_MASK;
354	}
355	}
356	# else
357	/* No paging so the fault address is the physical address */
358	GCPhys = (RTGCPHYS)(pvFault & ~PAGE_OFFSET_MASK);
359	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
360
361	/*
362	* If we have a GC address we'll check if it has any flags set.
363	*/
364	if (GCPhys != NIL_RTGCPHYS)
365	{
366	STAM_PROFILE_START(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
367
368	PPGMPAGE pPage;
369	rc = pgmPhysGetPageEx(&pVM->pgm.s, GCPhys, &pPage);
370	if (RT_SUCCESS(rc))
371	{
372	if ( PGM_PAGE_HAS_ACTIVE_PHYSICAL_HANDLERS(pPage)
373	\|\| PGM_PAGE_HAS_ACTIVE_VIRTUAL_HANDLERS(pPage))
374	{
375	if (PGM_PAGE_HAS_ANY_PHYSICAL_HANDLERS(pPage))
376	{
377	/*
378	* Physical page access handler.
379	*/
380	const RTGCPHYS GCPhysFault = GCPhys \| (pvFault & PAGE_OFFSET_MASK);
381	PPGMPHYSHANDLER pCur = (PPGMPHYSHANDLER)RTAvlroGCPhysRangeGet(&pVM->pgm.s.CTX_SUFF(pTrees)->PhysHandlers, GCPhysFault);
382	if (pCur)
383	{
384	# ifdef PGM_SYNC_N_PAGES
385	/*
386	* If the region is write protected and we got a page not present fault, then sync
387	* the pages. If the fault was caused by a read, then restart the instruction.
388	* In case of write access continue to the GC write handler.
389	*
390	* ASSUMES that there is only one handler per page or that they have similar write properties.
391	*/
392	if ( pCur->enmType == PGMPHYSHANDLERTYPE_PHYSICAL_WRITE
393	&& !(uErr & X86_TRAP_PF_P))
394	{
395	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, pvFault, PGM_SYNC_NR_PAGES, uErr);
396	if ( RT_FAILURE(rc)
397	\|\| !(uErr & X86_TRAP_PF_RW)
398	\|\| rc == VINF_PGM_SYNCPAGE_MODIFIED_PDE)
399	{
400	AssertRC(rc);
401	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersOutOfSync);
402	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
403	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2OutOfSyncHndPhys; });
404	return rc;
405	}
406	}
407	# endif
408
409	AssertMsg( pCur->enmType != PGMPHYSHANDLERTYPE_PHYSICAL_WRITE
410	\|\| (pCur->enmType == PGMPHYSHANDLERTYPE_PHYSICAL_WRITE && (uErr & X86_TRAP_PF_RW)),
411	("Unexpected trap for physical handler: %08X (phys=%08x) HCPhys=%X uErr=%X, enum=%d\n", pvFault, GCPhys, pPage->HCPhys, uErr, pCur->enmType));
412
413	# if defined(IN_RC) \|\| defined(IN_RING0)
414	if (pCur->CTX_SUFF(pfnHandler))
415	{
416	STAM_PROFILE_START(&pCur->Stat, h);
417	rc = pCur->CTX_SUFF(pfnHandler)(pVM, uErr, pRegFrame, pvFault, GCPhysFault, pCur->CTX_SUFF(pvUser));
418	STAM_PROFILE_STOP(&pCur->Stat, h);
419	}
420	else
421	# endif
422	rc = VINF_EM_RAW_EMULATE_INSTR;
423	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersPhysical);
424	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
425	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2HndPhys; });
426	return rc;
427	}
428	}
429	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
430	else
431	{
432	# ifdef PGM_SYNC_N_PAGES
433	/*
434	* If the region is write protected and we got a page not present fault, then sync
435	* the pages. If the fault was caused by a read, then restart the instruction.
436	* In case of write access continue to the GC write handler.
437	*/
438	if ( PGM_PAGE_GET_HNDL_VIRT_STATE(pPage) < PGM_PAGE_HNDL_PHYS_STATE_ALL
439	&& !(uErr & X86_TRAP_PF_P))
440	{
441	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, pvFault, PGM_SYNC_NR_PAGES, uErr);
442	if ( RT_FAILURE(rc)
443	\|\| rc == VINF_PGM_SYNCPAGE_MODIFIED_PDE
444	\|\| !(uErr & X86_TRAP_PF_RW))
445	{
446	AssertRC(rc);
447	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersOutOfSync);
448	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
449	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2OutOfSyncHndVirt; });
450	return rc;
451	}
452	}
453	# endif
454	/*
455	* Ok, it's an virtual page access handler.
456	*
457	* Since it's faster to search by address, we'll do that first
458	* and then retry by GCPhys if that fails.
459	*/
460	/** @todo r=bird: perhaps we should consider looking up by physical address directly now? */
461	/** @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
462	* page was changed without us noticing it (not-present -> present without invlpg or mov cr3, xxx)
463	*/
464	PPGMVIRTHANDLER pCur = (PPGMVIRTHANDLER)RTAvlroGCPtrRangeGet(&pVM->pgm.s.CTX_SUFF(pTrees)->VirtHandlers, pvFault);
465	if (pCur)
466	{
467	AssertMsg(!(pvFault - pCur->Core.Key < pCur->cb)
468	\|\| ( pCur->enmType != PGMVIRTHANDLERTYPE_WRITE
469	\|\| !(uErr & X86_TRAP_PF_P)
470	\|\| (pCur->enmType == PGMVIRTHANDLERTYPE_WRITE && (uErr & X86_TRAP_PF_RW))),
471	("Unexpected trap for virtual handler: %RGv (phys=%RGp) HCPhys=%HGp uErr=%X, enum=%d\n", pvFault, GCPhys, pPage->HCPhys, uErr, pCur->enmType));
472
473	if ( pvFault - pCur->Core.Key < pCur->cb
474	&& ( uErr & X86_TRAP_PF_RW
475	\|\| pCur->enmType != PGMVIRTHANDLERTYPE_WRITE ) )
476	{
477	# ifdef IN_RC
478	STAM_PROFILE_START(&pCur->Stat, h);
479	rc = pCur->CTX_SUFF(pfnHandler)(pVM, uErr, pRegFrame, pvFault, pCur->Core.Key, pvFault - pCur->Core.Key);
480	STAM_PROFILE_STOP(&pCur->Stat, h);
481	# else
482	rc = VINF_EM_RAW_EMULATE_INSTR; /** @todo for VMX */
483	# endif
484	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersVirtual);
485	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
486	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2HndVirt; });
487	return rc;
488	}
489	/* Unhandled part of a monitored page */
490	}
491	else
492	{
493	/* Check by physical address. */
494	PPGMVIRTHANDLER pCur;
495	unsigned iPage;
496	rc = pgmHandlerVirtualFindByPhysAddr(pVM, GCPhys + (pvFault & PAGE_OFFSET_MASK),
497	&pCur, &iPage);
498	Assert(RT_SUCCESS(rc) \|\| !pCur);
499	if ( pCur
500	&& ( uErr & X86_TRAP_PF_RW
501	\|\| pCur->enmType != PGMVIRTHANDLERTYPE_WRITE ) )
502	{
503	Assert((pCur->aPhysToVirt[iPage].Core.Key & X86_PTE_PAE_PG_MASK) == GCPhys);
504	# ifdef IN_RC
505	RTGCPTR off = (iPage << PAGE_SHIFT) + (pvFault & PAGE_OFFSET_MASK) - (pCur->Core.Key & PAGE_OFFSET_MASK);
506	Assert(off < pCur->cb);
507	STAM_PROFILE_START(&pCur->Stat, h);
508	rc = pCur->CTX_SUFF(pfnHandler)(pVM, uErr, pRegFrame, pvFault, pCur->Core.Key, off);
509	STAM_PROFILE_STOP(&pCur->Stat, h);
510	# else
511	rc = VINF_EM_RAW_EMULATE_INSTR; /** @todo for VMX */
512	# endif
513	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersVirtualByPhys);
514	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
515	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2HndVirt; });
516	return rc;
517	}
518	}
519	}
520	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
521
522	/*
523	* There is a handled area of the page, but this fault doesn't belong to it.
524	* We must emulate the instruction.
525	*
526	* To avoid crashing (non-fatal) in the interpreter and go back to the recompiler
527	* we first check if this was a page-not-present fault for a page with only
528	* write access handlers. Restart the instruction if it wasn't a write access.
529	*/
530	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersUnhandled);
531
532	if ( !PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage)
533	&& !(uErr & X86_TRAP_PF_P))
534	{
535	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, pvFault, PGM_SYNC_NR_PAGES, uErr);
536	if ( RT_FAILURE(rc)
537	\|\| rc == VINF_PGM_SYNCPAGE_MODIFIED_PDE
538	\|\| !(uErr & X86_TRAP_PF_RW))
539	{
540	AssertRC(rc);
541	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersOutOfSync);
542	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
543	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2OutOfSyncHndPhys; });
544	return rc;
545	}
546	}
547
548	/** @todo This particular case can cause quite a lot of overhead. E.g. early stage of kernel booting in Ubuntu 6.06
549	* It's writing to an unhandled part of the LDT page several million times.
550	*/
551	rc = PGMInterpretInstruction(pVM, pRegFrame, pvFault);
552	LogFlow(("PGM: PGMInterpretInstruction -> rc=%d HCPhys=%RHp%s%s\n",
553	rc, pPage->HCPhys,
554	PGM_PAGE_HAS_ANY_PHYSICAL_HANDLERS(pPage) ? " phys" : "",
555	PGM_PAGE_HAS_ANY_VIRTUAL_HANDLERS(pPage) ? " virt" : ""));
556	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
557	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2HndUnhandled; });
558	return rc;
559	} /* if any kind of handler */
560
561	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
562	if (uErr & X86_TRAP_PF_P)
563	{
564	/*
565	* The page isn't marked, but it might still be monitored by a virtual page access handler.
566	* (ASSUMES no temporary disabling of virtual handlers.)
567	*/
568	/** @todo r=bird: Since the purpose is to catch out of sync pages with virtual handler(s) here,
569	* we should correct both the shadow page table and physical memory flags, and not only check for
570	* accesses within the handler region but for access to pages with virtual handlers. */
571	PPGMVIRTHANDLER pCur = (PPGMVIRTHANDLER)RTAvlroGCPtrRangeGet(&pVM->pgm.s.CTX_SUFF(pTrees)->VirtHandlers, pvFault);
572	if (pCur)
573	{
574	AssertMsg( !(pvFault - pCur->Core.Key < pCur->cb)
575	\|\| ( pCur->enmType != PGMVIRTHANDLERTYPE_WRITE
576	\|\| !(uErr & X86_TRAP_PF_P)
577	\|\| (pCur->enmType == PGMVIRTHANDLERTYPE_WRITE && (uErr & X86_TRAP_PF_RW))),
578	("Unexpected trap for virtual handler: %08X (phys=%08x) HCPhys=%X uErr=%X, enum=%d\n", pvFault, GCPhys, pPage->HCPhys, uErr, pCur->enmType));
579
580	if ( pvFault - pCur->Core.Key < pCur->cb
581	&& ( uErr & X86_TRAP_PF_RW
582	\|\| pCur->enmType != PGMVIRTHANDLERTYPE_WRITE ) )
583	{
584	# ifdef IN_RC
585	STAM_PROFILE_START(&pCur->Stat, h);
586	rc = pCur->CTX_SUFF(pfnHandler)(pVM, uErr, pRegFrame, pvFault, pCur->Core.Key, pvFault - pCur->Core.Key);
587	STAM_PROFILE_STOP(&pCur->Stat, h);
588	# else
589	rc = VINF_EM_RAW_EMULATE_INSTR; /** @todo for VMX */
590	# endif
591	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersVirtualUnmarked);
592	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
593	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2HndVirt; });
594	return rc;
595	}
596	}
597	}
598	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
599	}
600	else
601	{
602	/* When the guest accesses invalid physical memory (e.g. probing of RAM or accessing a remapped MMIO range), then we'll fall
603	* back to the recompiler to emulate the instruction.
604	*/
605	LogFlow(("pgmPhysGetPageEx %RGp failed with %Rrc\n", GCPhys, rc));
606	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersInvalid);
607	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
608	return VINF_EM_RAW_EMULATE_INSTR;
609	}
610
611	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
612
613	# ifdef PGM_OUT_OF_SYNC_IN_GC
614	/*
615	* We are here only if page is present in Guest page tables and trap is not handled
616	* by our handlers.
617	* Check it for page out-of-sync situation.
618	*/
619	STAM_PROFILE_START(&pVM->pgm.s.StatRZTrap0eTimeOutOfSync, c);
620
621	if (!(uErr & X86_TRAP_PF_P))
622	{
623	/*
624	* Page is not present in our page tables.
625	* Try to sync it!
626	* BTW, fPageShw is invalid in this branch!
627	*/
628	if (uErr & X86_TRAP_PF_US)
629	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,PageOutOfSyncUser));
630	else /* supervisor */
631	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,PageOutOfSyncSupervisor));
632
633	# if defined(LOG_ENABLED) && !defined(IN_RING0)
634	RTGCPHYS GCPhys;
635	uint64_t fPageGst;
636	PGMGstGetPage(pVM, pvFault, &fPageGst, &GCPhys);
637	Log(("Page out of sync: %RGv eip=%08x PdeSrc.n.u1User=%d fPageGst=%08llx GCPhys=%RGp scan=%d\n",
638	pvFault, pRegFrame->eip, PdeSrc.n.u1User, fPageGst, GCPhys, CSAMDoesPageNeedScanning(pVM, (RTRCPTR)pRegFrame->eip)));
639	# endif /* LOG_ENABLED */
640
641	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) && !defined(IN_RING0)
642	if (CPUMGetGuestCPL(pVM, pRegFrame) == 0)
643	{
644	uint64_t fPageGst;
645	rc = PGMGstGetPage(pVM, pvFault, &fPageGst, NULL);
646	if ( RT_SUCCESS(rc)
647	&& !(fPageGst & X86_PTE_US))
648	{
649	/* Note: can't check for X86_TRAP_ID bit, because that requires execute disable support on the CPU */
650	if ( pvFault == (RTGCPTR)pRegFrame->eip
651	\|\| pvFault - pRegFrame->eip < 8 /* instruction crossing a page boundary */
652	# ifdef CSAM_DETECT_NEW_CODE_PAGES
653	\|\| ( !PATMIsPatchGCAddr(pVM, (RTGCPTR)pRegFrame->eip)
654	&& CSAMDoesPageNeedScanning(pVM, (RTRCPTR)pRegFrame->eip)) /* any new code we encounter here */
655	# endif /* CSAM_DETECT_NEW_CODE_PAGES */
656	)
657	{
658	LogFlow(("CSAMExecFault %RX32\n", pRegFrame->eip));
659	rc = CSAMExecFault(pVM, (RTRCPTR)pRegFrame->eip);
660	if (rc != VINF_SUCCESS)
661	{
662	/*
663	* CSAM needs to perform a job in ring 3.
664	*
665	* Sync the page before going to the host context; otherwise we'll end up in a loop if
666	* CSAM fails (e.g. instruction crosses a page boundary and the next page is not present)
667	*/
668	LogFlow(("CSAM ring 3 job\n"));
669	int rc2 = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, pvFault, 1, uErr);
670	AssertRC(rc2);
671
672	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeOutOfSync, c);
673	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2CSAM; });
674	return rc;
675	}
676	}
677	# ifdef CSAM_DETECT_NEW_CODE_PAGES
678	else if ( uErr == X86_TRAP_PF_RW
679	&& pRegFrame->ecx >= 0x100 /* early check for movswd count */
680	&& pRegFrame->ecx < 0x10000)
681	{
682	/* In case of a write to a non-present supervisor shadow page, we'll take special precautions
683	* to detect loading of new code pages.
684	*/
685
686	/*
687	* Decode the instruction.
688	*/
689	RTGCPTR PC;
690	rc = SELMValidateAndConvertCSAddr(pVM, pRegFrame->eflags, pRegFrame->ss, pRegFrame->cs, &pRegFrame->csHid, (RTGCPTR)pRegFrame->eip, &PC);
691	if (rc == VINF_SUCCESS)
692	{
693	DISCPUSTATE Cpu;
694	uint32_t cbOp;
695	rc = EMInterpretDisasOneEx(pVM, PC, pRegFrame, &Cpu, &cbOp);
696
697	/* For now we'll restrict this to rep movsw/d instructions */
698	if ( rc == VINF_SUCCESS
699	&& Cpu.pCurInstr->opcode == OP_MOVSWD
700	&& (Cpu.prefix & PREFIX_REP))
701	{
702	CSAMMarkPossibleCodePage(pVM, pvFault);
703	}
704	}
705	}
706	# endif /* CSAM_DETECT_NEW_CODE_PAGES */
707
708	/*
709	* Mark this page as safe.
710	*/
711	/** @todo not correct for pages that contain both code and data!! */
712	Log2(("CSAMMarkPage %RGv; scanned=%d\n", pvFault, true));
713	CSAMMarkPage(pVM, (RTRCPTR)pvFault, true);
714	}
715	}
716	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) && !defined(IN_RING0) */
717	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, pvFault, PGM_SYNC_NR_PAGES, uErr);
718	if (RT_SUCCESS(rc))
719	{
720	/* The page was successfully synced, return to the guest. */
721	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeOutOfSync, c);
722	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2OutOfSync; });
723	return VINF_SUCCESS;
724	}
725	}
726	else
727	{
728	/*
729	* A side effect of not flushing global PDEs are out of sync pages due
730	* to physical monitored regions, that are no longer valid.
731	* Assume for now it only applies to the read/write flag
732	*/
733	if (RT_SUCCESS(rc) && (uErr & X86_TRAP_PF_RW))
734	{
735	if (uErr & X86_TRAP_PF_US)
736	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,PageOutOfSyncUser));
737	else /* supervisor */
738	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,PageOutOfSyncSupervisor));
739
740
741	/*
742	* 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.
743	*/
744	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, pvFault, 1, uErr);
745	if (RT_SUCCESS(rc))
746	{
747	/*
748	* Page was successfully synced, return to guest.
749	*/
750	# ifdef VBOX_STRICT
751	RTGCPHYS GCPhys;
752	uint64_t fPageGst;
753	rc = PGMGstGetPage(pVM, pvFault, &fPageGst, &GCPhys);
754	Assert(RT_SUCCESS(rc) && fPageGst & X86_PTE_RW);
755	LogFlow(("Obsolete physical monitor page out of sync %RGv - phys %RGp flags=%08llx\n", pvFault, GCPhys, (uint64_t)fPageGst));
756
757	uint64_t fPageShw;
758	rc = PGMShwGetPage(pVM, pvFault, &fPageShw, NULL);
759	AssertMsg(RT_SUCCESS(rc) && fPageShw & X86_PTE_RW, ("rc=%Rrc fPageShw=%RX64\n", rc, fPageShw));
760	# endif /* VBOX_STRICT */
761	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeOutOfSync, c);
762	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2OutOfSyncHndObs; });
763	return VINF_SUCCESS;
764	}
765
766	/* Check to see if we need to emulate the instruction as X86_CR0_WP has been cleared. */
767	if ( CPUMGetGuestCPL(pVM, pRegFrame) == 0
768	&& ((CPUMGetGuestCR0(pVM) & (X86_CR0_WP \| X86_CR0_PG)) == X86_CR0_PG)
769	&& (uErr & (X86_TRAP_PF_RW \| X86_TRAP_PF_P)) == (X86_TRAP_PF_RW \| X86_TRAP_PF_P))
770	{
771	uint64_t fPageGst;
772	rc = PGMGstGetPage(pVM, pvFault, &fPageGst, NULL);
773	if ( RT_SUCCESS(rc)
774	&& !(fPageGst & X86_PTE_RW))
775	{
776	rc = PGMInterpretInstruction(pVM, pRegFrame, pvFault);
777	if (RT_SUCCESS(rc))
778	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eWPEmulInRZ);
779	else
780	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eWPEmulToR3);
781	return rc;
782	}
783	AssertMsgFailed(("Unexpected r/w page %RGv flag=%x rc=%Rrc\n", pvFault, (uint32_t)fPageGst, rc));
784	}
785	}
786
787	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
788	# ifdef VBOX_STRICT
789	/*
790	* Check for VMM page flags vs. Guest page flags consistency.
791	* Currently only for debug purposes.
792	*/
793	if (RT_SUCCESS(rc))
794	{
795	/* Get guest page flags. */
796	uint64_t fPageGst;
797	rc = PGMGstGetPage(pVM, pvFault, &fPageGst, NULL);
798	if (RT_SUCCESS(rc))
799	{
800	uint64_t fPageShw;
801	rc = PGMShwGetPage(pVM, pvFault, &fPageShw, NULL);
802
803	/*
804	* Compare page flags.
805	* Note: we have AVL, A, D bits desynched.
806	*/
807	AssertMsg((fPageShw & ~(X86_PTE_A \| X86_PTE_D \| X86_PTE_AVL_MASK)) == (fPageGst & ~(X86_PTE_A \| X86_PTE_D \| X86_PTE_AVL_MASK)),
808	("Page flags mismatch! pvFault=%RGv GCPhys=%RGp fPageShw=%08llx fPageGst=%08llx\n", pvFault, GCPhys, fPageShw, fPageGst));
809	}
810	else
811	AssertMsgFailed(("PGMGstGetPage rc=%Rrc\n", rc));
812	}
813	else
814	AssertMsgFailed(("PGMGCGetPage rc=%Rrc\n", rc));
815	# endif /* VBOX_STRICT */
816	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
817	}
818	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeOutOfSync, c);
819	# endif /* PGM_OUT_OF_SYNC_IN_GC */
820	}
821	else
822	{
823	/*
824	* Page not present in Guest OS or invalid page table address.
825	* This is potential virtual page access handler food.
826	*
827	* For the present we'll say that our access handlers don't
828	* work for this case - we've already discarded the page table
829	* not present case which is identical to this.
830	*
831	* When we perchance find we need this, we will probably have AVL
832	* trees (offset based) to operate on and we can measure their speed
833	* agains mapping a page table and probably rearrange this handling
834	* a bit. (Like, searching virtual ranges before checking the
835	* physical address.)
836	*/
837	}
838	}
839
840
841	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
842	/*
843	* Conclusion, this is a guest trap.
844	*/
845	LogFlow(("PGM: Unhandled #PF -> route trap to recompiler!\n"));
846	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eGuestPFUnh);
847	return VINF_EM_RAW_GUEST_TRAP;
848	# else
849	/* present, but not a monitored page; perhaps the guest is probing physical memory */
850	return VINF_EM_RAW_EMULATE_INSTR;
851	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
852
853
854	# else /* PGM_GST_TYPE != PGM_TYPE_32BIT */
855
856	AssertReleaseMsgFailed(("Shw=%d Gst=%d is not implemented!\n", PGM_GST_TYPE, PGM_SHW_TYPE));
857	return VERR_INTERNAL_ERROR;
858	# endif /* PGM_GST_TYPE != PGM_TYPE_32BIT */
859	}
860	#endif /* !IN_RING3 */
861
862
863	/**
864	* Emulation of the invlpg instruction.
865	*
866	*
867	* @returns VBox status code.
868	*
869	* @param pVM VM handle.
870	* @param GCPtrPage Page to invalidate.
871	*
872	* @remark ASSUMES that the guest is updating before invalidating. This order
873	* isn't required by the CPU, so this is speculative and could cause
874	* trouble.
875	*
876	* @todo Flush page or page directory only if necessary!
877	* @todo Add a #define for simply invalidating the page.
878	*/
879	PGM_BTH_DECL(int, InvalidatePage)(PVM pVM, RTGCPTR GCPtrPage)
880	{
881	#if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) \
882	&& PGM_SHW_TYPE != PGM_TYPE_NESTED \
883	&& PGM_SHW_TYPE != PGM_TYPE_EPT
884	int rc;
885
886	LogFlow(("InvalidatePage %RGv\n", GCPtrPage));
887	/*
888	* Get the shadow PD entry and skip out if this PD isn't present.
889	* (Guessing that it is frequent for a shadow PDE to not be present, do this first.)
890	*/
891	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
892	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
893	PX86PDE pPdeDst = pgmShwGet32BitPDEPtr(&pVM->pgm.s, GCPtrPage);
894
895	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
896	const unsigned iPdpt = (GCPtrPage >> X86_PDPT_SHIFT);
897	PX86PDPT pPdptDst = pgmShwGetPaePDPTPtr(&pVM->pgm.s);
898
899	/* If the shadow PDPE isn't present, then skip the invalidate. */
900	if (!pPdptDst->a[iPdpt].n.u1Present)
901	{
902	Assert(!(pPdptDst->a[iPdpt].u & PGM_PLXFLAGS_MAPPING));
903	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePageSkipped));
904	return VINF_SUCCESS;
905	}
906
907	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) /& SHW_PD_MASK - pool index only atm! /;
908	PX86PDEPAE pPdeDst = pgmShwGetPaePDEPtr(&pVM->pgm.s, GCPtrPage);
909
910	# else /* PGM_SHW_TYPE == PGM_TYPE_AMD64 */
911	/* PML4 */
912	AssertReturn(pVM->pgm.s.pShwPaePml4R3, VERR_INTERNAL_ERROR);
913
914	const unsigned iPml4 = (GCPtrPage >> X86_PML4_SHIFT) & X86_PML4_MASK;
915	const unsigned iPdpt = (GCPtrPage >> X86_PDPT_SHIFT) & X86_PDPT_MASK_AMD64;
916	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
917	PX86PDPAE pPDDst;
918	PX86PDPT pPdptDst;
919	PX86PML4E pPml4eDst;
920	rc = pgmShwGetLongModePDPtr(pVM, GCPtrPage, &pPml4eDst, &pPdptDst, &pPDDst);
921	if (rc != VINF_SUCCESS)
922	{
923	AssertMsg(rc == VERR_PAGE_DIRECTORY_PTR_NOT_PRESENT \|\| rc == VERR_PAGE_MAP_LEVEL4_NOT_PRESENT, ("Unexpected rc=%Rrc\n", rc));
924	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePageSkipped));
925	if (!VM_FF_ISSET(pVM, VM_FF_PGM_SYNC_CR3))
926	PGM_INVL_GUEST_TLBS();
927	return VINF_SUCCESS;
928	}
929	Assert(pPDDst);
930
931	PX86PDEPAE pPdeDst = &pPDDst->a[iPDDst];
932	PX86PDPE pPdpeDst = &pPdptDst->a[iPdpt];
933
934	if (!pPdpeDst->n.u1Present)
935	{
936	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePageSkipped));
937	if (!VM_FF_ISSET(pVM, VM_FF_PGM_SYNC_CR3))
938	PGM_INVL_GUEST_TLBS();
939	return VINF_SUCCESS;
940	}
941
942	# endif /* PGM_SHW_TYPE == PGM_TYPE_AMD64 */
943
944	const SHWPDE PdeDst = *pPdeDst;
945	if (!PdeDst.n.u1Present)
946	{
947	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePageSkipped));
948	return VINF_SUCCESS;
949	}
950
951	/*
952	* Get the guest PD entry and calc big page.
953	*/
954	# if PGM_GST_TYPE == PGM_TYPE_32BIT
955	PGSTPD pPDSrc = pgmGstGet32bitPDPtr(&pVM->pgm.s);
956	const unsigned iPDSrc = GCPtrPage >> GST_PD_SHIFT;
957	GSTPDE PdeSrc = pPDSrc->a[iPDSrc];
958	# else /* PGM_GST_TYPE != PGM_TYPE_32BIT */
959	unsigned iPDSrc;
960	# if PGM_GST_TYPE == PGM_TYPE_PAE
961	X86PDPE PdpeSrc;
962	PX86PDPAE pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, GCPtrPage, &iPDSrc, &PdpeSrc);
963	# else /* AMD64 */
964	PX86PML4E pPml4eSrc;
965	X86PDPE PdpeSrc;
966	PX86PDPAE pPDSrc = pgmGstGetLongModePDPtr(&pVM->pgm.s, GCPtrPage, &pPml4eSrc, &PdpeSrc, &iPDSrc);
967	# endif
968	GSTPDE PdeSrc;
969
970	if (pPDSrc)
971	PdeSrc = pPDSrc->a[iPDSrc];
972	else
973	PdeSrc.u = 0;
974	# endif /* PGM_GST_TYPE != PGM_TYPE_32BIT */
975
976	# if PGM_GST_TYPE == PGM_TYPE_AMD64
977	const bool fIsBigPage = PdeSrc.b.u1Size;
978	# else
979	const bool fIsBigPage = PdeSrc.b.u1Size && (CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
980	# endif
981
982	# ifdef IN_RING3
983	/*
984	* If a CR3 Sync is pending we may ignore the invalidate page operation
985	* depending on the kind of sync and if it's a global page or not.
986	* This doesn't make sense in GC/R0 so we'll skip it entirely there.
987	*/
988	# ifdef PGM_SKIP_GLOBAL_PAGEDIRS_ON_NONGLOBAL_FLUSH
989	if ( VM_FF_ISSET(pVM, VM_FF_PGM_SYNC_CR3)
990	\|\| ( VM_FF_ISSET(pVM, VM_FF_PGM_SYNC_CR3_NON_GLOBAL)
991	&& fIsBigPage
992	&& PdeSrc.b.u1Global
993	)
994	)
995	# else
996	if (VM_FF_ISPENDING(pVM, VM_FF_PGM_SYNC_CR3 \| VM_FF_PGM_SYNC_CR3_NON_GLOBAL) )
997	# endif
998	{
999	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePageSkipped));
1000	return VINF_SUCCESS;
1001	}
1002	# endif /* IN_RING3 */
1003
1004	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1005	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
1006
1007	/* Fetch the pgm pool shadow descriptor. */
1008	PPGMPOOLPAGE pShwPdpt = pgmPoolGetPageByHCPhys(pVM, pPml4eDst->u & X86_PML4E_PG_MASK);
1009	Assert(pShwPdpt);
1010
1011	/* Fetch the pgm pool shadow descriptor. */
1012	PPGMPOOLPAGE pShwPde = pgmPoolGetPageByHCPhys(pVM, pPdptDst->a[iPdpt].u & SHW_PDPE_PG_MASK);
1013	Assert(pShwPde);
1014
1015	Assert(pPml4eDst->n.u1Present && (pPml4eDst->u & SHW_PDPT_MASK));
1016	RTGCPHYS GCPhysPdpt = pPml4eSrc->u & X86_PML4E_PG_MASK;
1017
1018	if ( !pPml4eSrc->n.u1Present
1019	\|\| pShwPdpt->GCPhys != GCPhysPdpt)
1020	{
1021	LogFlow(("InvalidatePage: Out-of-sync PML4E (P/GCPhys) at %RGv GCPhys=%RGp vs %RGp Pml4eSrc=%RX64 Pml4eDst=%RX64\n",
1022	GCPtrPage, pShwPdpt->GCPhys, GCPhysPdpt, (uint64_t)pPml4eSrc->u, (uint64_t)pPml4eDst->u));
1023	pgmPoolFreeByPage(pPool, pShwPdpt, pVM->pgm.s.CTX_SUFF(pShwAmd64CR3)->idx, iPml4);
1024	pPml4eDst->u = 0;
1025	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDNPs));
1026	PGM_INVL_GUEST_TLBS();
1027	return VINF_SUCCESS;
1028	}
1029	if ( pPml4eSrc->n.u1User != pPml4eDst->n.u1User
1030	\|\| (!pPml4eSrc->n.u1Write && pPml4eDst->n.u1Write))
1031	{
1032	/*
1033	* Mark not present so we can resync the PML4E when it's used.
1034	*/
1035	LogFlow(("InvalidatePage: Out-of-sync PML4E at %RGv Pml4eSrc=%RX64 Pml4eDst=%RX64\n",
1036	GCPtrPage, (uint64_t)pPml4eSrc->u, (uint64_t)pPml4eDst->u));
1037	pgmPoolFreeByPage(pPool, pShwPdpt, pVM->pgm.s.CTX_SUFF(pShwAmd64CR3)->idx, iPml4);
1038	pPml4eDst->u = 0;
1039	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDOutOfSync));
1040	PGM_INVL_GUEST_TLBS();
1041	}
1042	else if (!pPml4eSrc->n.u1Accessed)
1043	{
1044	/*
1045	* Mark not present so we can set the accessed bit.
1046	*/
1047	LogFlow(("InvalidatePage: Out-of-sync PML4E (A) at %RGv Pml4eSrc=%RX64 Pml4eDst=%RX64\n",
1048	GCPtrPage, (uint64_t)pPml4eSrc->u, (uint64_t)pPml4eDst->u));
1049	pgmPoolFreeByPage(pPool, pShwPdpt, pVM->pgm.s.CTX_SUFF(pShwAmd64CR3)->idx, iPml4);
1050	pPml4eDst->u = 0;
1051	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDNAs));
1052	PGM_INVL_GUEST_TLBS();
1053	}
1054
1055	/* Check if the PDPT entry has changed. */
1056	Assert(pPdpeDst->n.u1Present && pPdpeDst->u & SHW_PDPT_MASK);
1057	RTGCPHYS GCPhysPd = PdpeSrc.u & GST_PDPE_PG_MASK;
1058	if ( !PdpeSrc.n.u1Present
1059	\|\| pShwPde->GCPhys != GCPhysPd)
1060	{
1061	LogFlow(("InvalidatePage: Out-of-sync PDPE (P/GCPhys) at %RGv GCPhys=%RGp vs %RGp PdpeSrc=%RX64 PdpeDst=%RX64\n",
1062	GCPtrPage, pShwPde->GCPhys, GCPhysPd, (uint64_t)PdpeSrc.u, (uint64_t)pPdpeDst->u));
1063	pgmPoolFreeByPage(pPool, pShwPde, pShwPdpt->idx, iPdpt);
1064	pPdpeDst->u = 0;
1065	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDNPs));
1066	PGM_INVL_GUEST_TLBS();
1067	return VINF_SUCCESS;
1068	}
1069	if ( PdpeSrc.lm.u1User != pPdpeDst->lm.u1User
1070	\|\| (!PdpeSrc.lm.u1Write && pPdpeDst->lm.u1Write))
1071	{
1072	/*
1073	* Mark not present so we can resync the PDPTE when it's used.
1074	*/
1075	LogFlow(("InvalidatePage: Out-of-sync PDPE at %RGv PdpeSrc=%RX64 PdpeDst=%RX64\n",
1076	GCPtrPage, (uint64_t)PdpeSrc.u, (uint64_t)pPdpeDst->u));
1077	pgmPoolFreeByPage(pPool, pShwPde, pShwPdpt->idx, iPdpt);
1078	pPdpeDst->u = 0;
1079	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDOutOfSync));
1080	PGM_INVL_GUEST_TLBS();
1081	}
1082	else if (!PdpeSrc.lm.u1Accessed)
1083	{
1084	/*
1085	* Mark not present so we can set the accessed bit.
1086	*/
1087	LogFlow(("InvalidatePage: Out-of-sync PDPE (A) at %RGv PdpeSrc=%RX64 PdpeDst=%RX64\n",
1088	GCPtrPage, (uint64_t)PdpeSrc.u, (uint64_t)pPdpeDst->u));
1089	pgmPoolFreeByPage(pPool, pShwPde, pShwPdpt->idx, iPdpt);
1090	pPdpeDst->u = 0;
1091	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDNAs));
1092	PGM_INVL_GUEST_TLBS();
1093	}
1094	# endif /* PGM_GST_TYPE == PGM_TYPE_AMD64 */
1095
1096	# if PGM_GST_TYPE == PGM_TYPE_PAE
1097	/*
1098	* Update the shadow PDPE and free all the shadow PD entries if the PDPE is marked not present.
1099	* Note: This shouldn't actually be necessary as we monitor the PDPT page for changes.
1100	*/
1101	if (!pPDSrc)
1102	{
1103	/* Guest PDPE not present */
1104	PX86PDPAE pPDDst = pgmShwGetPaePDPtr(&pVM->pgm.s, GCPtrPage);
1105	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
1106
1107	Assert(!PdpeSrc.n.u1Present);
1108	LogFlow(("InvalidatePage: guest PDPE %d not present; clear shw pdpe\n", iPdpt));
1109
1110	/* for each page directory entry */
1111	for (unsigned iPD = 0; iPD < X86_PG_PAE_ENTRIES; iPD++)
1112	{
1113	if ( pPDDst->a[iPD].n.u1Present
1114	&& !(pPDDst->a[iPD].u & PGM_PDFLAGS_MAPPING))
1115	{
1116	pgmPoolFree(pVM, pPDDst->a[iPD].u & SHW_PDE_PG_MASK, SHW_POOL_ROOT_IDX, iPdpt * X86_PG_PAE_ENTRIES + iPD);
1117	pPDDst->a[iPD].u = 0;
1118	}
1119	}
1120	if (!(pPdptDst->a[iPdpt].u & PGM_PLXFLAGS_MAPPING))
1121	pPdptDst->a[iPdpt].n.u1Present = 0;
1122	PGM_INVL_GUEST_TLBS();
1123	}
1124	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]));
1125	# endif
1126
1127
1128	/*
1129	* Deal with the Guest PDE.
1130	*/
1131	rc = VINF_SUCCESS;
1132	if (PdeSrc.n.u1Present)
1133	{
1134	if (PdeDst.u & PGM_PDFLAGS_MAPPING)
1135	{
1136	/*
1137	* Conflict - Let SyncPT deal with it to avoid duplicate code.
1138	*/
1139	Assert(pgmMapAreMappingsEnabled(&pVM->pgm.s));
1140	Assert(PGMGetGuestMode(pVM) <= PGMMODE_PAE);
1141	rc = PGM_BTH_NAME(SyncPT)(pVM, iPDSrc, pPDSrc, GCPtrPage);
1142	}
1143	else if ( PdeSrc.n.u1User != PdeDst.n.u1User
1144	\|\| (!PdeSrc.n.u1Write && PdeDst.n.u1Write))
1145	{
1146	/*
1147	* Mark not present so we can resync the PDE when it's used.
1148	*/
1149	LogFlow(("InvalidatePage: Out-of-sync at %RGp PdeSrc=%RX64 PdeDst=%RX64\n",
1150	GCPtrPage, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
1151	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1152	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, pShwPde->idx, iPDDst);
1153	# else
1154	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, SHW_POOL_ROOT_IDX, iPDDst);
1155	# endif
1156	pPdeDst->u = 0;
1157	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDOutOfSync));
1158	PGM_INVL_GUEST_TLBS();
1159	}
1160	else if (!PdeSrc.n.u1Accessed)
1161	{
1162	/*
1163	* Mark not present so we can set the accessed bit.
1164	*/
1165	LogFlow(("InvalidatePage: Out-of-sync (A) at %RGp PdeSrc=%RX64 PdeDst=%RX64\n",
1166	GCPtrPage, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
1167	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1168	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, pShwPde->idx, iPDDst);
1169	# else
1170	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, SHW_POOL_ROOT_IDX, iPDDst);
1171	# endif
1172	pPdeDst->u = 0;
1173	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDNAs));
1174	PGM_INVL_GUEST_TLBS();
1175	}
1176	else if (!fIsBigPage)
1177	{
1178	/*
1179	* 4KB - page.
1180	*/
1181	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, PdeDst.u & SHW_PDE_PG_MASK);
1182	RTGCPHYS GCPhys = PdeSrc.u & GST_PDE_PG_MASK;
1183	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
1184	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
1185	GCPhys \|= (iPDDst & 1) * (PAGE_SIZE/2);
1186	# endif
1187	if (pShwPage->GCPhys == GCPhys)
1188	{
1189	# if 0 /* likely cause of a major performance regression; must be SyncPageWorkerTrackDeref then */
1190	const unsigned iPTEDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1191	PSHWPT pPT = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
1192	if (pPT->a[iPTEDst].n.u1Present)
1193	{
1194	# ifdef PGMPOOL_WITH_USER_TRACKING
1195	/* This is very unlikely with caching/monitoring enabled. */
1196	PGM_BTH_NAME(SyncPageWorkerTrackDeref)(pVM, pShwPage, pPT->a[iPTEDst].u & SHW_PTE_PG_MASK);
1197	# endif
1198	pPT->a[iPTEDst].u = 0;
1199	}
1200	# else /* Syncing it here isn't 100% safe and it's probably not worth spending time syncing it. */
1201	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, GCPtrPage, 1, 0);
1202	if (RT_SUCCESS(rc))
1203	rc = VINF_SUCCESS;
1204	# endif
1205	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePage4KBPages));
1206	PGM_INVL_PG(GCPtrPage);
1207	}
1208	else
1209	{
1210	/*
1211	* The page table address changed.
1212	*/
1213	LogFlow(("InvalidatePage: Out-of-sync at %RGp PdeSrc=%RX64 PdeDst=%RX64 ShwGCPhys=%RGp iPDDst=%#x\n",
1214	GCPtrPage, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u, pShwPage->GCPhys, iPDDst));
1215	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1216	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, pShwPde->idx, iPDDst);
1217	# else
1218	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, SHW_POOL_ROOT_IDX, iPDDst);
1219	# endif
1220	pPdeDst->u = 0;
1221	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDOutOfSync));
1222	PGM_INVL_GUEST_TLBS();
1223	}
1224	}
1225	else
1226	{
1227	/*
1228	* 2/4MB - page.
1229	*/
1230	/* Before freeing the page, check if anything really changed. */
1231	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, PdeDst.u & SHW_PDE_PG_MASK);
1232	RTGCPHYS GCPhys = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc);
1233	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
1234	/* Select the right PDE as we're emulating a 4MB page directory with two 2 MB shadow PDEs.*/
1235	GCPhys \|= GCPtrPage & (1 << X86_PD_PAE_SHIFT);
1236	# endif
1237	if ( pShwPage->GCPhys == GCPhys
1238	&& pShwPage->enmKind == BTH_PGMPOOLKIND_PT_FOR_BIG)
1239	{
1240	/* ASSUMES a the given bits are identical for 4M and normal PDEs */
1241	/** @todo PAT */
1242	if ( (PdeSrc.u & (X86_PDE_P \| X86_PDE_RW \| X86_PDE_US \| X86_PDE_PWT \| X86_PDE_PCD))
1243	== (PdeDst.u & (X86_PDE_P \| X86_PDE_RW \| X86_PDE_US \| X86_PDE_PWT \| X86_PDE_PCD))
1244	&& ( PdeSrc.b.u1Dirty /** @todo rainy day: What about read-only 4M pages? not very common, but still... */
1245	\|\| (PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY)))
1246	{
1247	LogFlow(("Skipping flush for big page containing %RGv (PD=%X .u=%RX64)-> nothing has changed!\n", GCPtrPage, iPDSrc, PdeSrc.u));
1248	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePage4MBPagesSkip));
1249	return VINF_SUCCESS;
1250	}
1251	}
1252
1253	/*
1254	* Ok, the page table is present and it's been changed in the guest.
1255	* If we're in host context, we'll just mark it as not present taking the lazy approach.
1256	* We could do this for some flushes in GC too, but we need an algorithm for
1257	* deciding which 4MB pages containing code likely to be executed very soon.
1258	*/
1259	LogFlow(("InvalidatePage: Out-of-sync PD at %RGp PdeSrc=%RX64 PdeDst=%RX64\n",
1260	GCPtrPage, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
1261	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1262	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, pShwPde->idx, iPDDst);
1263	# else
1264	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, SHW_POOL_ROOT_IDX, iPDDst);
1265	# endif
1266	pPdeDst->u = 0;
1267	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePage4MBPages));
1268	PGM_INVL_BIG_PG(GCPtrPage);
1269	}
1270	}
1271	else
1272	{
1273	/*
1274	* Page directory is not present, mark shadow PDE not present.
1275	*/
1276	if (!(PdeDst.u & PGM_PDFLAGS_MAPPING))
1277	{
1278	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1279	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, pShwPde->idx, iPDDst);
1280	# else
1281	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, SHW_POOL_ROOT_IDX, iPDDst);
1282	# endif
1283	pPdeDst->u = 0;
1284	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDNPs));
1285	PGM_INVL_PG(GCPtrPage);
1286	}
1287	else
1288	{
1289	Assert(pgmMapAreMappingsEnabled(&pVM->pgm.s));
1290	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDMappings));
1291	}
1292	}
1293
1294	return rc;
1295
1296	#else /* guest real and protected mode */
1297	/* There's no such thing as InvalidatePage when paging is disabled, so just ignore. */
1298	return VINF_SUCCESS;
1299	#endif
1300	}
1301
1302
1303	#ifdef PGMPOOL_WITH_USER_TRACKING
1304	/**
1305	* Update the tracking of shadowed pages.
1306	*
1307	* @param pVM The VM handle.
1308	* @param pShwPage The shadow page.
1309	* @param HCPhys The physical page we is being dereferenced.
1310	*/
1311	DECLINLINE(void) PGM_BTH_NAME(SyncPageWorkerTrackDeref)(PVM pVM, PPGMPOOLPAGE pShwPage, RTHCPHYS HCPhys)
1312	{
1313	# ifdef PGMPOOL_WITH_GCPHYS_TRACKING
1314	STAM_PROFILE_START(&pVM->pgm.s.StatTrackDeref, a);
1315	LogFlow(("SyncPageWorkerTrackDeref: Damn HCPhys=%RHp pShwPage->idx=%#x!!!\n", HCPhys, pShwPage->idx));
1316
1317	/** @todo If this turns out to be a bottle neck (very likely) two things can be done:
1318	* 1. have a medium sized HCPhys -> GCPhys TLB (hash?)
1319	* 2. write protect all shadowed pages. I.e. implement caching.
1320	*/
1321	/*
1322	* Find the guest address.
1323	*/
1324	for (PPGMRAMRANGE pRam = pVM->pgm.s.CTX_SUFF(pRamRanges);
1325	pRam;
1326	pRam = pRam->CTX_SUFF(pNext))
1327	{
1328	unsigned iPage = pRam->cb >> PAGE_SHIFT;
1329	while (iPage-- > 0)
1330	{
1331	if (PGM_PAGE_GET_HCPHYS(&pRam->aPages[iPage]) == HCPhys)
1332	{
1333	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
1334	pgmTrackDerefGCPhys(pPool, pShwPage, &pRam->aPages[iPage]);
1335	pShwPage->cPresent--;
1336	pPool->cPresent--;
1337	STAM_PROFILE_STOP(&pVM->pgm.s.StatTrackDeref, a);
1338	return;
1339	}
1340	}
1341	}
1342
1343	for (;;)
1344	AssertReleaseMsgFailed(("HCPhys=%RHp wasn't found!\n", HCPhys));
1345	# else /* !PGMPOOL_WITH_GCPHYS_TRACKING */
1346	pShwPage->cPresent--;
1347	pVM->pgm.s.CTX_SUFF(pPool)->cPresent--;
1348	# endif /* !PGMPOOL_WITH_GCPHYS_TRACKING */
1349	}
1350
1351
1352	/**
1353	* Update the tracking of shadowed pages.
1354	*
1355	* @param pVM The VM handle.
1356	* @param pShwPage The shadow page.
1357	* @param u16 The top 16-bit of the pPage->HCPhys.
1358	* @param pPage Pointer to the guest page. this will be modified.
1359	* @param iPTDst The index into the shadow table.
1360	*/
1361	DECLINLINE(void) PGM_BTH_NAME(SyncPageWorkerTrackAddref)(PVM pVM, PPGMPOOLPAGE pShwPage, uint16_t u16, PPGMPAGE pPage, const unsigned iPTDst)
1362	{
1363	# ifdef PGMPOOL_WITH_GCPHYS_TRACKING
1364	/*
1365	* We're making certain assumptions about the placement of cRef and idx.
1366	*/
1367	Assert(MM_RAM_FLAGS_IDX_SHIFT == 48);
1368	Assert(MM_RAM_FLAGS_CREFS_SHIFT > MM_RAM_FLAGS_IDX_SHIFT);
1369
1370	/*
1371	* Just deal with the simple first time here.
1372	*/
1373	if (!u16)
1374	{
1375	STAM_COUNTER_INC(&pVM->pgm.s.StatTrackVirgin);
1376	u16 = (1 << (MM_RAM_FLAGS_CREFS_SHIFT - MM_RAM_FLAGS_IDX_SHIFT)) \| pShwPage->idx;
1377	}
1378	else
1379	u16 = pgmPoolTrackPhysExtAddref(pVM, u16, pShwPage->idx);
1380
1381	/* write back, trying to be clever... */
1382	Log2(("SyncPageWorkerTrackAddRef: u16=%#x pPage->HCPhys=%RHp->%RHp iPTDst=%#x\n",
1383	u16, pPage->HCPhys, (pPage->HCPhys & MM_RAM_FLAGS_NO_REFS_MASK) \| ((uint64_t)u16 << MM_RAM_FLAGS_CREFS_SHIFT), iPTDst));
1384	((uint16_t )&pPage->HCPhys + 3) = u16; /** @todo PAGE FLAGS */
1385	# endif /* PGMPOOL_WITH_GCPHYS_TRACKING */
1386
1387	/* update statistics. */
1388	pVM->pgm.s.CTX_SUFF(pPool)->cPresent++;
1389	pShwPage->cPresent++;
1390	if (pShwPage->iFirstPresent > iPTDst)
1391	pShwPage->iFirstPresent = iPTDst;
1392	}
1393	#endif /* PGMPOOL_WITH_USER_TRACKING */
1394
1395
1396	/**
1397	* Creates a 4K shadow page for a guest page.
1398	*
1399	* For 4M pages the caller must convert the PDE4M to a PTE, this includes adjusting the
1400	* physical address. The PdeSrc argument only the flags are used. No page structured
1401	* will be mapped in this function.
1402	*
1403	* @param pVM VM handle.
1404	* @param pPteDst Destination page table entry.
1405	* @param PdeSrc Source page directory entry (i.e. Guest OS page directory entry).
1406	* Can safely assume that only the flags are being used.
1407	* @param PteSrc Source page table entry (i.e. Guest OS page table entry).
1408	* @param pShwPage Pointer to the shadow page.
1409	* @param iPTDst The index into the shadow table.
1410	*
1411	* @remark Not used for 2/4MB pages!
1412	*/
1413	DECLINLINE(void) PGM_BTH_NAME(SyncPageWorker)(PVM pVM, PSHWPTE pPteDst, GSTPDE PdeSrc, GSTPTE PteSrc, PPGMPOOLPAGE pShwPage, unsigned iPTDst)
1414	{
1415	if (PteSrc.n.u1Present)
1416	{
1417	/*
1418	* Find the ram range.
1419	*/
1420	PPGMPAGE pPage;
1421	int rc = pgmPhysGetPageEx(&pVM->pgm.s, PteSrc.u & GST_PTE_PG_MASK, &pPage);
1422	if (RT_SUCCESS(rc))
1423	{
1424	/** @todo investiage PWT, PCD and PAT. */
1425	/*
1426	* Make page table entry.
1427	*/
1428	const RTHCPHYS HCPhys = pPage->HCPhys; /** @todo FLAGS */
1429	SHWPTE PteDst;
1430	if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
1431	{
1432	/** @todo r=bird: Are we actually handling dirty and access bits for pages with access handlers correctly? No. */
1433	if (!PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage))
1434	{
1435	#if PGM_SHW_TYPE == PGM_TYPE_EPT
1436	PteDst.u = (HCPhys & EPT_PTE_PG_MASK);
1437	PteDst.n.u1Present = 1;
1438	PteDst.n.u1Execute = 1;
1439	PteDst.n.u1IgnorePAT = 1;
1440	PteDst.n.u3EMT = VMX_EPT_MEMTYPE_WB;
1441	/* PteDst.n.u1Write = 0 && PteDst.n.u1Size = 0 */
1442	#else
1443	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))
1444	\| (HCPhys & X86_PTE_PAE_PG_MASK);
1445	#endif
1446	}
1447	else
1448	{
1449	LogFlow(("SyncPageWorker: monitored page (%RHp) -> mark not present\n", HCPhys));
1450	PteDst.u = 0;
1451	}
1452	/** @todo count these two kinds. */
1453	}
1454	else
1455	{
1456	#if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
1457	/*
1458	* If the page or page directory entry is not marked accessed,
1459	* we mark the page not present.
1460	*/
1461	if (!PteSrc.n.u1Accessed \|\| !PdeSrc.n.u1Accessed)
1462	{
1463	LogFlow(("SyncPageWorker: page and or page directory not accessed -> mark not present\n"));
1464	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,AccessedPage));
1465	PteDst.u = 0;
1466	}
1467	else
1468	/*
1469	* If the page is not flagged as dirty and is writable, then make it read-only, so we can set the dirty bit
1470	* when the page is modified.
1471	*/
1472	if (!PteSrc.n.u1Dirty && (PdeSrc.n.u1Write & PteSrc.n.u1Write))
1473	{
1474	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPage));
1475	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))
1476	\| (HCPhys & X86_PTE_PAE_PG_MASK)
1477	\| PGM_PTFLAGS_TRACK_DIRTY;
1478	}
1479	else
1480	#endif
1481	{
1482	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPageSkipped));
1483	#if PGM_SHW_TYPE == PGM_TYPE_EPT
1484	PteDst.u = (HCPhys & EPT_PTE_PG_MASK);
1485	PteDst.n.u1Present = 1;
1486	PteDst.n.u1Write = 1;
1487	PteDst.n.u1Execute = 1;
1488	PteDst.n.u1IgnorePAT = 1;
1489	PteDst.n.u3EMT = VMX_EPT_MEMTYPE_WB;
1490	/* PteDst.n.u1Size = 0 */
1491	#else
1492	PteDst.u = (PteSrc.u & ~(X86_PTE_PAE_PG_MASK \| X86_PTE_AVL_MASK \| X86_PTE_PAT \| X86_PTE_PCD \| X86_PTE_PWT))
1493	\| (HCPhys & X86_PTE_PAE_PG_MASK);
1494	#endif
1495	}
1496	}
1497
1498	#ifdef PGMPOOL_WITH_USER_TRACKING
1499	/*
1500	* Keep user track up to date.
1501	*/
1502	if (PteDst.n.u1Present)
1503	{
1504	if (!pPteDst->n.u1Present)
1505	PGM_BTH_NAME(SyncPageWorkerTrackAddref)(pVM, pShwPage, HCPhys >> MM_RAM_FLAGS_IDX_SHIFT, pPage, iPTDst);
1506	else if ((pPteDst->u & SHW_PTE_PG_MASK) != (PteDst.u & SHW_PTE_PG_MASK))
1507	{
1508	Log2(("SyncPageWorker: deref! *pPteDst=%RX64 PteDst=%RX64\n", (uint64_t)pPteDst->u, (uint64_t)PteDst.u));
1509	PGM_BTH_NAME(SyncPageWorkerTrackDeref)(pVM, pShwPage, pPteDst->u & SHW_PTE_PG_MASK);
1510	PGM_BTH_NAME(SyncPageWorkerTrackAddref)(pVM, pShwPage, HCPhys >> MM_RAM_FLAGS_IDX_SHIFT, pPage, iPTDst);
1511	}
1512	}
1513	else if (pPteDst->n.u1Present)
1514	{
1515	Log2(("SyncPageWorker: deref! *pPteDst=%RX64\n", (uint64_t)pPteDst->u));
1516	PGM_BTH_NAME(SyncPageWorkerTrackDeref)(pVM, pShwPage, pPteDst->u & SHW_PTE_PG_MASK);
1517	}
1518	#endif /* PGMPOOL_WITH_USER_TRACKING */
1519
1520	/*
1521	* Update statistics and commit the entry.
1522	*/
1523	if (!PteSrc.n.u1Global)
1524	pShwPage->fSeenNonGlobal = true;
1525	*pPteDst = PteDst;
1526	}
1527	/* else MMIO or invalid page, we must handle them manually in the #PF handler. */
1528	/** @todo count these. */
1529	}
1530	else
1531	{
1532	/*
1533	* Page not-present.
1534	*/
1535	LogFlow(("SyncPageWorker: page not present in Pte\n"));
1536	#ifdef PGMPOOL_WITH_USER_TRACKING
1537	/* Keep user track up to date. */
1538	if (pPteDst->n.u1Present)
1539	{
1540	Log2(("SyncPageWorker: deref! *pPteDst=%RX64\n", (uint64_t)pPteDst->u));
1541	PGM_BTH_NAME(SyncPageWorkerTrackDeref)(pVM, pShwPage, pPteDst->u & SHW_PTE_PG_MASK);
1542	}
1543	#endif /* PGMPOOL_WITH_USER_TRACKING */
1544	pPteDst->u = 0;
1545	/** @todo count these. */
1546	}
1547	}
1548
1549
1550	/**
1551	* Syncs a guest OS page.
1552	*
1553	* There are no conflicts at this point, neither is there any need for
1554	* page table allocations.
1555	*
1556	* @returns VBox status code.
1557	* @returns VINF_PGM_SYNCPAGE_MODIFIED_PDE if it modifies the PDE in any way.
1558	* @param pVM VM handle.
1559	* @param PdeSrc Page directory entry of the guest.
1560	* @param GCPtrPage Guest context page address.
1561	* @param cPages Number of pages to sync (PGM_SYNC_N_PAGES) (default=1).
1562	* @param uErr Fault error (X86_TRAP_PF_*).
1563	*/
1564	PGM_BTH_DECL(int, SyncPage)(PVM pVM, GSTPDE PdeSrc, RTGCPTR GCPtrPage, unsigned cPages, unsigned uErr)
1565	{
1566	LogFlow(("SyncPage: GCPtrPage=%RGv cPages=%u uErr=%#x\n", GCPtrPage, cPages, uErr));
1567
1568	#if ( PGM_GST_TYPE == PGM_TYPE_32BIT \
1569	\|\| PGM_GST_TYPE == PGM_TYPE_PAE \
1570	\|\| PGM_GST_TYPE == PGM_TYPE_AMD64) \
1571	&& PGM_SHW_TYPE != PGM_TYPE_NESTED \
1572	&& PGM_SHW_TYPE != PGM_TYPE_EPT
1573
1574	# if PGM_WITH_NX(PGM_GST_TYPE, PGM_SHW_TYPE)
1575	bool fNoExecuteBitValid = !!(CPUMGetGuestEFER(pVM) & MSR_K6_EFER_NXE);
1576	# endif
1577
1578	/*
1579	* Assert preconditions.
1580	*/
1581	Assert(PdeSrc.n.u1Present);
1582	Assert(cPages);
1583	STAM_COUNTER_INC(&pVM->pgm.s.StatSyncPagePD[(GCPtrPage >> GST_PD_SHIFT) & GST_PD_MASK]);
1584
1585	/*
1586	* Get the shadow PDE, find the shadow page table in the pool.
1587	*/
1588	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
1589	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
1590	PX86PDE pPdeDst = pgmShwGet32BitPDEPtr(&pVM->pgm.s, GCPtrPage);
1591
1592	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
1593	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) /& SHW_PD_MASK - only pool index atm! /;
1594	const unsigned iPdpt = (GCPtrPage >> X86_PDPT_SHIFT);
1595	PX86PDPT pPdptDst = pgmShwGetPaePDPTPtr(&pVM->pgm.s); NOREF(pPdptDst);
1596	PX86PDEPAE pPdeDst = pgmShwGetPaePDEPtr(&pVM->pgm.s, GCPtrPage);
1597	AssertReturn(pPdeDst, VERR_INTERNAL_ERROR);
1598
1599	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
1600	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
1601	const unsigned iPdpt = (GCPtrPage >> X86_PDPT_SHIFT) & X86_PDPT_MASK_AMD64;
1602	PX86PDPAE pPDDst;
1603	PX86PDPT pPdptDst;
1604
1605	int rc = pgmShwGetLongModePDPtr(pVM, GCPtrPage, NULL, &pPdptDst, &pPDDst);
1606	AssertRCSuccessReturn(rc, rc);
1607	Assert(pPDDst && pPdptDst);
1608	PX86PDEPAE pPdeDst = &pPDDst->a[iPDDst];
1609	# endif
1610
1611	SHWPDE PdeDst = *pPdeDst;
1612	AssertMsg(PdeDst.n.u1Present, ("%p=%llx\n", pPdeDst, (uint64_t)PdeDst.u));
1613	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, PdeDst.u & SHW_PDE_PG_MASK);
1614
1615	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1616	/* Fetch the pgm pool shadow descriptor. */
1617	PPGMPOOLPAGE pShwPde = pgmPoolGetPageByHCPhys(pVM, pPdptDst->a[iPdpt].u & X86_PDPE_PG_MASK);
1618	Assert(pShwPde);
1619	# endif
1620
1621	/*
1622	* Check that the page is present and that the shadow PDE isn't out of sync.
1623	*/
1624	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1625	const bool fBigPage = PdeSrc.b.u1Size;
1626	# else
1627	const bool fBigPage = PdeSrc.b.u1Size && (CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
1628	# endif
1629	RTGCPHYS GCPhys;
1630	if (!fBigPage)
1631	{
1632	GCPhys = PdeSrc.u & GST_PDE_PG_MASK;
1633	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
1634	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
1635	GCPhys \|= (iPDDst & 1) * (PAGE_SIZE/2);
1636	# endif
1637	}
1638	else
1639	{
1640	GCPhys = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc);
1641	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
1642	/* Select the right PDE as we're emulating a 4MB page directory with two 2 MB shadow PDEs.*/
1643	GCPhys \|= GCPtrPage & (1 << X86_PD_PAE_SHIFT);
1644	# endif
1645	}
1646	if ( pShwPage->GCPhys == GCPhys
1647	&& PdeSrc.n.u1Present
1648	&& (PdeSrc.n.u1User == PdeDst.n.u1User)
1649	&& (PdeSrc.n.u1Write == PdeDst.n.u1Write \|\| !PdeDst.n.u1Write)
1650	# if PGM_WITH_NX(PGM_GST_TYPE, PGM_SHW_TYPE)
1651	&& (!fNoExecuteBitValid \|\| PdeSrc.n.u1NoExecute == PdeDst.n.u1NoExecute)
1652	# endif
1653	)
1654	{
1655	/*
1656	* Check that the PDE is marked accessed already.
1657	* Since we set the accessed bit before getting here on a #PF, this
1658	* check is only meant for dealing with non-#PF'ing paths.
1659	*/
1660	if (PdeSrc.n.u1Accessed)
1661	{
1662	PSHWPT pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
1663	if (!fBigPage)
1664	{
1665	/*
1666	* 4KB Page - Map the guest page table.
1667	*/
1668	PGSTPT pPTSrc;
1669	int rc = PGM_GCPHYS_2_PTR(pVM, PdeSrc.u & GST_PDE_PG_MASK, &pPTSrc);
1670	if (RT_SUCCESS(rc))
1671	{
1672	# ifdef PGM_SYNC_N_PAGES
1673	Assert(cPages == 1 \|\| !(uErr & X86_TRAP_PF_P));
1674	if (cPages > 1 && !(uErr & X86_TRAP_PF_P))
1675	{
1676	/*
1677	* This code path is currently only taken when the caller is PGMTrap0eHandler
1678	* for non-present pages!
1679	*
1680	* We're setting PGM_SYNC_NR_PAGES pages around the faulting page to sync it and
1681	* deal with locality.
1682	*/
1683	unsigned iPTDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1684	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
1685	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
1686	const unsigned offPTSrc = ((GCPtrPage >> SHW_PD_SHIFT) & 1) * 512;
1687	# else
1688	const unsigned offPTSrc = 0;
1689	# endif
1690	const unsigned iPTDstEnd = RT_MIN(iPTDst + PGM_SYNC_NR_PAGES / 2, RT_ELEMENTS(pPTDst->a));
1691	if (iPTDst < PGM_SYNC_NR_PAGES / 2)
1692	iPTDst = 0;
1693	else
1694	iPTDst -= PGM_SYNC_NR_PAGES / 2;
1695	for (; iPTDst < iPTDstEnd; iPTDst++)
1696	{
1697	if (!pPTDst->a[iPTDst].n.u1Present)
1698	{
1699	GSTPTE PteSrc = pPTSrc->a[offPTSrc + iPTDst];
1700	RTGCPTR GCPtrCurPage = (GCPtrPage & ~(RTGCPTR)(GST_PT_MASK << GST_PT_SHIFT)) \| ((offPTSrc + iPTDst) << PAGE_SHIFT);
1701	NOREF(GCPtrCurPage);
1702	#ifndef IN_RING0
1703	/*
1704	* Assuming kernel code will be marked as supervisor - and not as user level
1705	* and executed using a conforming code selector - And marked as readonly.
1706	* Also assume that if we're monitoring a page, it's of no interest to CSAM.
1707	*/
1708	PPGMPAGE pPage;
1709	if ( ((PdeSrc.u & PteSrc.u) & (X86_PTE_RW \| X86_PTE_US))
1710	\|\| iPTDst == ((GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK) /* always sync GCPtrPage */
1711	\|\| !CSAMDoesPageNeedScanning(pVM, (RTRCPTR)GCPtrCurPage)
1712	\|\| ( (pPage = pgmPhysGetPage(&pVM->pgm.s, PteSrc.u & GST_PTE_PG_MASK))
1713	&& PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
1714	)
1715	#endif /* else: CSAM not active */
1716	PGM_BTH_NAME(SyncPageWorker)(pVM, &pPTDst->a[iPTDst], PdeSrc, PteSrc, pShwPage, iPTDst);
1717	Log2(("SyncPage: 4K+ %RGv PteSrc:{P=%d RW=%d U=%d raw=%08llx} PteDst=%08llx%s\n",
1718	GCPtrCurPage, PteSrc.n.u1Present,
1719	PteSrc.n.u1Write & PdeSrc.n.u1Write,
1720	PteSrc.n.u1User & PdeSrc.n.u1User,
1721	(uint64_t)PteSrc.u,
1722	(uint64_t)pPTDst->a[iPTDst].u,
1723	pPTDst->a[iPTDst].u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
1724	}
1725	}
1726	}
1727	else
1728	# endif /* PGM_SYNC_N_PAGES */
1729	{
1730	const unsigned iPTSrc = (GCPtrPage >> GST_PT_SHIFT) & GST_PT_MASK;
1731	GSTPTE PteSrc = pPTSrc->a[iPTSrc];
1732	const unsigned iPTDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1733	PGM_BTH_NAME(SyncPageWorker)(pVM, &pPTDst->a[iPTDst], PdeSrc, PteSrc, pShwPage, iPTDst);
1734	Log2(("SyncPage: 4K %RGv PteSrc:{P=%d RW=%d U=%d raw=%08llx}%s\n",
1735	GCPtrPage, PteSrc.n.u1Present,
1736	PteSrc.n.u1Write & PdeSrc.n.u1Write,
1737	PteSrc.n.u1User & PdeSrc.n.u1User,
1738	(uint64_t)PteSrc.u,
1739	pPTDst->a[iPTDst].u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
1740	}
1741	}
1742	else /* MMIO or invalid page: emulated in #PF handler. */
1743	{
1744	LogFlow(("PGM_GCPHYS_2_PTR %RGp failed with %Rrc\n", GCPhys, rc));
1745	Assert(!pPTDst->a[(GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK].n.u1Present);
1746	}
1747	}
1748	else
1749	{
1750	/*
1751	* 4/2MB page - lazy syncing shadow 4K pages.
1752	* (There are many causes of getting here, it's no longer only CSAM.)
1753	*/
1754	/* Calculate the GC physical address of this 4KB shadow page. */
1755	RTGCPHYS GCPhys = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc) \| (GCPtrPage & GST_BIG_PAGE_OFFSET_MASK);
1756	/* Find ram range. */
1757	PPGMPAGE pPage;
1758	int rc = pgmPhysGetPageEx(&pVM->pgm.s, GCPhys, &pPage);
1759	if (RT_SUCCESS(rc))
1760	{
1761	/*
1762	* Make shadow PTE entry.
1763	*/
1764	const RTHCPHYS HCPhys = pPage->HCPhys; /** @todo PAGE FLAGS */
1765	SHWPTE PteDst;
1766	PteDst.u = (PdeSrc.u & ~(X86_PTE_PAE_PG_MASK \| X86_PTE_AVL_MASK \| X86_PTE_PAT \| X86_PTE_PCD \| X86_PTE_PWT))
1767	\| (HCPhys & X86_PTE_PAE_PG_MASK);
1768	if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
1769	{
1770	if (!PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage))
1771	PteDst.n.u1Write = 0;
1772	else
1773	PteDst.u = 0;
1774	}
1775	const unsigned iPTDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1776	# ifdef PGMPOOL_WITH_USER_TRACKING
1777	if (PteDst.n.u1Present && !pPTDst->a[iPTDst].n.u1Present)
1778	PGM_BTH_NAME(SyncPageWorkerTrackAddref)(pVM, pShwPage, HCPhys >> MM_RAM_FLAGS_IDX_SHIFT, pPage, iPTDst);
1779	# endif
1780	pPTDst->a[iPTDst] = PteDst;
1781
1782
1783	/*
1784	* If the page is not flagged as dirty and is writable, then make it read-only
1785	* at PD level, so we can set the dirty bit when the page is modified.
1786	*
1787	* ASSUMES that page access handlers are implemented on page table entry level.
1788	* Thus we will first catch the dirty access and set PDE.D and restart. If
1789	* there is an access handler, we'll trap again and let it work on the problem.
1790	*/
1791	/** @todo r=bird: figure out why we need this here, SyncPT should've taken care of this already.
1792	* As for invlpg, it simply frees the whole shadow PT.
1793	* ...It's possibly because the guest clears it and the guest doesn't really tell us... */
1794	if (!PdeSrc.b.u1Dirty && PdeSrc.b.u1Write)
1795	{
1796	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPageBig));
1797	PdeDst.u \|= PGM_PDFLAGS_TRACK_DIRTY;
1798	PdeDst.n.u1Write = 0;
1799	}
1800	else
1801	{
1802	PdeDst.au32[0] &= ~PGM_PDFLAGS_TRACK_DIRTY;
1803	PdeDst.n.u1Write = PdeSrc.n.u1Write;
1804	}
1805	*pPdeDst = PdeDst;
1806	Log2(("SyncPage: BIG %RGv PdeSrc:{P=%d RW=%d U=%d raw=%08llx} GCPhys=%RGp%s\n",
1807	GCPtrPage, PdeSrc.n.u1Present, PdeSrc.n.u1Write, PdeSrc.n.u1User, (uint64_t)PdeSrc.u, GCPhys,
1808	PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
1809	}
1810	else
1811	LogFlow(("PGM_GCPHYS_2_PTR %RGp (big) failed with %Rrc\n", GCPhys, rc));
1812	}
1813	return VINF_SUCCESS;
1814	}
1815	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPagePDNAs));
1816	}
1817	else
1818	{
1819	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPagePDOutOfSync));
1820	Log2(("SyncPage: Out-Of-Sync PDE at %RGp PdeSrc=%RX64 PdeDst=%RX64\n",
1821	GCPtrPage, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
1822	}
1823
1824	/*
1825	* Mark the PDE not present. Restart the instruction and let #PF call SyncPT.
1826	* Yea, I'm lazy.
1827	*/
1828	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
1829	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1830	pgmPoolFreeByPage(pPool, pShwPage, pShwPde->idx, iPDDst);
1831	# else
1832	pgmPoolFreeByPage(pPool, pShwPage, SHW_POOL_ROOT_IDX, iPDDst);
1833	# endif
1834
1835	pPdeDst->u = 0;
1836	PGM_INVL_GUEST_TLBS();
1837	return VINF_PGM_SYNCPAGE_MODIFIED_PDE;
1838
1839	#elif (PGM_GST_TYPE == PGM_TYPE_REAL \|\| PGM_GST_TYPE == PGM_TYPE_PROT) \
1840	&& PGM_SHW_TYPE != PGM_TYPE_NESTED \
1841	&& (PGM_SHW_TYPE != PGM_TYPE_EPT \|\| PGM_GST_TYPE == PGM_TYPE_PROT)
1842
1843	# ifdef PGM_SYNC_N_PAGES
1844	/*
1845	* Get the shadow PDE, find the shadow page table in the pool.
1846	*/
1847	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
1848	X86PDE PdeDst = pgmShwGet32BitPDE(&pVM->pgm.s, GCPtrPage);
1849
1850	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
1851	X86PDEPAE PdeDst = pgmShwGetPaePDE(&pVM->pgm.s, GCPtrPage);
1852
1853	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
1854	const unsigned iPDDst = ((GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK);
1855	const unsigned iPdpt = (GCPtrPage >> X86_PDPT_SHIFT) & X86_PDPT_MASK_AMD64; NOREF(iPdpt);
1856	PX86PDPAE pPDDst;
1857	X86PDEPAE PdeDst;
1858	PX86PDPT pPdptDst;
1859
1860	int rc = pgmShwGetLongModePDPtr(pVM, GCPtrPage, NULL, &pPdptDst, &pPDDst);
1861	AssertRCSuccessReturn(rc, rc);
1862	Assert(pPDDst && pPdptDst);
1863	PdeDst = pPDDst->a[iPDDst];
1864	# elif PGM_SHW_TYPE == PGM_TYPE_EPT
1865	const unsigned iPDDst = ((GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK);
1866	PEPTPD pPDDst;
1867	EPTPDE PdeDst;
1868
1869	int rc = pgmShwGetEPTPDPtr(pVM, GCPtrPage, NULL, &pPDDst);
1870	if (rc != VINF_SUCCESS)
1871	{
1872	AssertRC(rc);
1873	return rc;
1874	}
1875	Assert(pPDDst);
1876	PdeDst = pPDDst->a[iPDDst];
1877	# endif
1878	AssertMsg(PdeDst.n.u1Present, ("%#llx\n", (uint64_t)PdeDst.u));
1879	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, PdeDst.u & SHW_PDE_PG_MASK);
1880	PSHWPT pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
1881
1882	Assert(cPages == 1 \|\| !(uErr & X86_TRAP_PF_P));
1883	if (cPages > 1 && !(uErr & X86_TRAP_PF_P))
1884	{
1885	/*
1886	* This code path is currently only taken when the caller is PGMTrap0eHandler
1887	* for non-present pages!
1888	*
1889	* We're setting PGM_SYNC_NR_PAGES pages around the faulting page to sync it and
1890	* deal with locality.
1891	*/
1892	unsigned iPTDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1893	const unsigned iPTDstEnd = RT_MIN(iPTDst + PGM_SYNC_NR_PAGES / 2, RT_ELEMENTS(pPTDst->a));
1894	if (iPTDst < PGM_SYNC_NR_PAGES / 2)
1895	iPTDst = 0;
1896	else
1897	iPTDst -= PGM_SYNC_NR_PAGES / 2;
1898	for (; iPTDst < iPTDstEnd; iPTDst++)
1899	{
1900	if (!pPTDst->a[iPTDst].n.u1Present)
1901	{
1902	GSTPTE PteSrc;
1903
1904	RTGCPTR GCPtrCurPage = (GCPtrPage & ~(RTGCPTR)(SHW_PT_MASK << SHW_PT_SHIFT)) \| (iPTDst << PAGE_SHIFT);
1905
1906	/* Fake the page table entry */
1907	PteSrc.u = GCPtrCurPage;
1908	PteSrc.n.u1Present = 1;
1909	PteSrc.n.u1Dirty = 1;
1910	PteSrc.n.u1Accessed = 1;
1911	PteSrc.n.u1Write = 1;
1912	PteSrc.n.u1User = 1;
1913
1914	PGM_BTH_NAME(SyncPageWorker)(pVM, &pPTDst->a[iPTDst], PdeSrc, PteSrc, pShwPage, iPTDst);
1915
1916	Log2(("SyncPage: 4K+ %RGv PteSrc:{P=%d RW=%d U=%d raw=%08llx} PteDst=%08llx%s\n",
1917	GCPtrCurPage, PteSrc.n.u1Present,
1918	PteSrc.n.u1Write & PdeSrc.n.u1Write,
1919	PteSrc.n.u1User & PdeSrc.n.u1User,
1920	(uint64_t)PteSrc.u,
1921	(uint64_t)pPTDst->a[iPTDst].u,
1922	pPTDst->a[iPTDst].u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
1923	}
1924	else
1925	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));
1926	}
1927	}
1928	else
1929	# endif /* PGM_SYNC_N_PAGES */
1930	{
1931	GSTPTE PteSrc;
1932	const unsigned iPTDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1933	RTGCPTR GCPtrCurPage = (GCPtrPage & ~(RTGCPTR)(SHW_PT_MASK << SHW_PT_SHIFT)) \| (iPTDst << PAGE_SHIFT);
1934
1935	/* Fake the page table entry */
1936	PteSrc.u = GCPtrCurPage;
1937	PteSrc.n.u1Present = 1;
1938	PteSrc.n.u1Dirty = 1;
1939	PteSrc.n.u1Accessed = 1;
1940	PteSrc.n.u1Write = 1;
1941	PteSrc.n.u1User = 1;
1942	PGM_BTH_NAME(SyncPageWorker)(pVM, &pPTDst->a[iPTDst], PdeSrc, PteSrc, pShwPage, iPTDst);
1943
1944	Log2(("SyncPage: 4K %RGv PteSrc:{P=%d RW=%d U=%d raw=%08llx}PteDst=%08llx%s\n",
1945	GCPtrPage, PteSrc.n.u1Present,
1946	PteSrc.n.u1Write & PdeSrc.n.u1Write,
1947	PteSrc.n.u1User & PdeSrc.n.u1User,
1948	(uint64_t)PteSrc.u,
1949	(uint64_t)pPTDst->a[iPTDst].u,
1950	pPTDst->a[iPTDst].u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
1951	}
1952	return VINF_SUCCESS;
1953
1954	#else
1955	AssertReleaseMsgFailed(("Shw=%d Gst=%d is not implemented!\n", PGM_GST_TYPE, PGM_SHW_TYPE));
1956	return VERR_INTERNAL_ERROR;
1957	#endif
1958	}
1959
1960
1961	#if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
1962	/**
1963	* Investigate page fault and handle write protection page faults caused by
1964	* dirty bit tracking.
1965	*
1966	* @returns VBox status code.
1967	* @param pVM VM handle.
1968	* @param uErr Page fault error code.
1969	* @param pPdeDst Shadow page directory entry.
1970	* @param pPdeSrc Guest page directory entry.
1971	* @param GCPtrPage Guest context page address.
1972	*/
1973	PGM_BTH_DECL(int, CheckPageFault)(PVM pVM, uint32_t uErr, PSHWPDE pPdeDst, PGSTPDE pPdeSrc, RTGCPTR GCPtrPage)
1974	{
1975	bool fWriteProtect = !!(CPUMGetGuestCR0(pVM) & X86_CR0_WP);
1976	bool fUserLevelFault = !!(uErr & X86_TRAP_PF_US);
1977	bool fWriteFault = !!(uErr & X86_TRAP_PF_RW);
1978	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1979	bool fBigPagesSupported = true;
1980	# else
1981	bool fBigPagesSupported = !!(CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
1982	# endif
1983	# if PGM_WITH_NX(PGM_GST_TYPE, PGM_SHW_TYPE)
1984	bool fNoExecuteBitValid = !!(CPUMGetGuestEFER(pVM) & MSR_K6_EFER_NXE);
1985	# endif
1986	unsigned uPageFaultLevel;
1987	int rc;
1988
1989	STAM_PROFILE_START(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
1990	LogFlow(("CheckPageFault: GCPtrPage=%RGv uErr=%#x PdeSrc=%08x\n", GCPtrPage, uErr, pPdeSrc->u));
1991
1992	# if PGM_GST_TYPE == PGM_TYPE_PAE \
1993	\|\| PGM_GST_TYPE == PGM_TYPE_AMD64
1994
1995	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1996	PX86PML4E pPml4eSrc;
1997	PX86PDPE pPdpeSrc;
1998
1999	pPdpeSrc = pgmGstGetLongModePDPTPtr(&pVM->pgm.s, GCPtrPage, &pPml4eSrc);
2000	Assert(pPml4eSrc);
2001
2002	/*
2003	* Real page fault? (PML4E level)
2004	*/
2005	if ( (uErr & X86_TRAP_PF_RSVD)
2006	\|\| !pPml4eSrc->n.u1Present
2007	\|\| (fNoExecuteBitValid && (uErr & X86_TRAP_PF_ID) && pPml4eSrc->n.u1NoExecute)
2008	\|\| (fWriteFault && !pPml4eSrc->n.u1Write && (fUserLevelFault \|\| fWriteProtect))
2009	\|\| (fUserLevelFault && !pPml4eSrc->n.u1User)
2010	)
2011	{
2012	uPageFaultLevel = 0;
2013	goto l_UpperLevelPageFault;
2014	}
2015	Assert(pPdpeSrc);
2016
2017	# else /* PAE */
2018	PX86PDPE pPdpeSrc = pgmGstGetPaePDPEPtr(&pVM->pgm.s, GCPtrPage);
2019	# endif /* PAE */
2020
2021	/*
2022	* Real page fault? (PDPE level)
2023	*/
2024	if ( (uErr & X86_TRAP_PF_RSVD)
2025	\|\| !pPdpeSrc->n.u1Present
2026	# if PGM_GST_TYPE == PGM_TYPE_AMD64 /* NX, r/w, u/s bits in the PDPE are long mode only */
2027	\|\| (fNoExecuteBitValid && (uErr & X86_TRAP_PF_ID) && pPdpeSrc->lm.u1NoExecute)
2028	\|\| (fWriteFault && !pPdpeSrc->lm.u1Write && (fUserLevelFault \|\| fWriteProtect))
2029	\|\| (fUserLevelFault && !pPdpeSrc->lm.u1User)
2030	# endif
2031	)
2032	{
2033	uPageFaultLevel = 1;
2034	goto l_UpperLevelPageFault;
2035	}
2036	# endif
2037
2038	/*
2039	* Real page fault? (PDE level)
2040	*/
2041	if ( (uErr & X86_TRAP_PF_RSVD)
2042	\|\| !pPdeSrc->n.u1Present
2043	# if PGM_WITH_NX(PGM_GST_TYPE, PGM_SHW_TYPE)
2044	\|\| (fNoExecuteBitValid && (uErr & X86_TRAP_PF_ID) && pPdeSrc->n.u1NoExecute)
2045	# endif
2046	\|\| (fWriteFault && !pPdeSrc->n.u1Write && (fUserLevelFault \|\| fWriteProtect))
2047	\|\| (fUserLevelFault && !pPdeSrc->n.u1User) )
2048	{
2049	uPageFaultLevel = 2;
2050	goto l_UpperLevelPageFault;
2051	}
2052
2053	/*
2054	* First check the easy case where the page directory has been marked read-only to track
2055	* the dirty bit of an emulated BIG page
2056	*/
2057	if (pPdeSrc->b.u1Size && fBigPagesSupported)
2058	{
2059	/* Mark guest page directory as accessed */
2060	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2061	pPml4eSrc->n.u1Accessed = 1;
2062	pPdpeSrc->lm.u1Accessed = 1;
2063	# endif
2064	pPdeSrc->b.u1Accessed = 1;
2065
2066	/*
2067	* Only write protection page faults are relevant here.
2068	*/
2069	if (fWriteFault)
2070	{
2071	/* Mark guest page directory as dirty (BIG page only). */
2072	pPdeSrc->b.u1Dirty = 1;
2073
2074	if (pPdeDst->n.u1Present && (pPdeDst->u & PGM_PDFLAGS_TRACK_DIRTY))
2075	{
2076	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPageTrap));
2077
2078	Assert(pPdeSrc->b.u1Write);
2079
2080	pPdeDst->n.u1Write = 1;
2081	pPdeDst->n.u1Accessed = 1;
2082	pPdeDst->au32[0] &= ~PGM_PDFLAGS_TRACK_DIRTY;
2083	PGM_INVL_BIG_PG(GCPtrPage);
2084	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2085	return VINF_PGM_HANDLED_DIRTY_BIT_FAULT;
2086	}
2087	}
2088	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2089	return VINF_PGM_NO_DIRTY_BIT_TRACKING;
2090	}
2091	/* else: 4KB page table */
2092
2093	/*
2094	* Map the guest page table.
2095	*/
2096	PGSTPT pPTSrc;
2097	rc = PGM_GCPHYS_2_PTR(pVM, pPdeSrc->u & GST_PDE_PG_MASK, &pPTSrc);
2098	if (RT_SUCCESS(rc))
2099	{
2100	/*
2101	* Real page fault?
2102	*/
2103	PGSTPTE pPteSrc = &pPTSrc->a[(GCPtrPage >> GST_PT_SHIFT) & GST_PT_MASK];
2104	const GSTPTE PteSrc = *pPteSrc;
2105	if ( !PteSrc.n.u1Present
2106	# if PGM_WITH_NX(PGM_GST_TYPE, PGM_SHW_TYPE)
2107	\|\| (fNoExecuteBitValid && (uErr & X86_TRAP_PF_ID) && PteSrc.n.u1NoExecute)
2108	# endif
2109	\|\| (fWriteFault && !PteSrc.n.u1Write && (fUserLevelFault \|\| fWriteProtect))
2110	\|\| (fUserLevelFault && !PteSrc.n.u1User)
2111	)
2112	{
2113	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyTrackRealPF));
2114	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2115	LogFlow(("CheckPageFault: real page fault at %RGv PteSrc.u=%08x (2)\n", GCPtrPage, PteSrc.u));
2116
2117	/* Check the present bit as the shadow tables can cause different error codes by being out of sync.
2118	* See the 2nd case above as well.
2119	*/
2120	if (pPdeSrc->n.u1Present && pPteSrc->n.u1Present)
2121	TRPMSetErrorCode(pVM, uErr \| X86_TRAP_PF_P); /* page-level protection violation */
2122
2123	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2124	return VINF_EM_RAW_GUEST_TRAP;
2125	}
2126	LogFlow(("CheckPageFault: page fault at %RGv PteSrc.u=%08x\n", GCPtrPage, PteSrc.u));
2127
2128	/*
2129	* Set the accessed bits in the page directory and the page table.
2130	*/
2131	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2132	pPml4eSrc->n.u1Accessed = 1;
2133	pPdpeSrc->lm.u1Accessed = 1;
2134	# endif
2135	pPdeSrc->n.u1Accessed = 1;
2136	pPteSrc->n.u1Accessed = 1;
2137
2138	/*
2139	* Only write protection page faults are relevant here.
2140	*/
2141	if (fWriteFault)
2142	{
2143	/* Write access, so mark guest entry as dirty. */
2144	# ifdef VBOX_WITH_STATISTICS
2145	if (!pPteSrc->n.u1Dirty)
2146	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtiedPage));
2147	else
2148	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,PageAlreadyDirty));
2149	# endif
2150
2151	pPteSrc->n.u1Dirty = 1;
2152
2153	if (pPdeDst->n.u1Present)
2154	{
2155	#ifndef IN_RING0
2156	/* Bail out here as pgmPoolGetPageByHCPhys will return NULL and we'll crash below.
2157	* Our individual shadow handlers will provide more information and force a fatal exit.
2158	*/
2159	if (MMHyperIsInsideArea(pVM, (RTGCPTR)GCPtrPage))
2160	{
2161	LogRel(("CheckPageFault: write to hypervisor region %RGv\n", GCPtrPage));
2162	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2163	return VINF_SUCCESS;
2164	}
2165	#endif
2166	/*
2167	* Map shadow page table.
2168	*/
2169	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, pPdeDst->u & SHW_PDE_PG_MASK);
2170	if (pShwPage)
2171	{
2172	PSHWPT pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
2173	PSHWPTE pPteDst = &pPTDst->a[(GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK];
2174	if ( pPteDst->n.u1Present /** @todo Optimize accessed bit emulation? */
2175	&& (pPteDst->u & PGM_PTFLAGS_TRACK_DIRTY))
2176	{
2177	LogFlow(("DIRTY page trap addr=%RGv\n", GCPtrPage));
2178	# ifdef VBOX_STRICT
2179	PPGMPAGE pPage = pgmPhysGetPage(&pVM->pgm.s, pPteSrc->u & GST_PTE_PG_MASK);
2180	if (pPage)
2181	AssertMsg(!PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage),
2182	("Unexpected dirty bit tracking on monitored page %RGv (phys %RGp)!!!!!!\n", GCPtrPage, pPteSrc->u & X86_PTE_PAE_PG_MASK));
2183	# endif
2184	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPageTrap));
2185
2186	Assert(pPteSrc->n.u1Write);
2187
2188	pPteDst->n.u1Write = 1;
2189	pPteDst->n.u1Dirty = 1;
2190	pPteDst->n.u1Accessed = 1;
2191	pPteDst->au32[0] &= ~PGM_PTFLAGS_TRACK_DIRTY;
2192	PGM_INVL_PG(GCPtrPage);
2193
2194	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2195	return VINF_PGM_HANDLED_DIRTY_BIT_FAULT;
2196	}
2197	}
2198	else
2199	AssertMsgFailed(("pgmPoolGetPageByHCPhys %RGp failed!\n", pPdeDst->u & SHW_PDE_PG_MASK));
2200	}
2201	}
2202	/** @todo Optimize accessed bit emulation? */
2203	# ifdef VBOX_STRICT
2204	/*
2205	* Sanity check.
2206	*/
2207	else if ( !pPteSrc->n.u1Dirty
2208	&& (pPdeSrc->n.u1Write & pPteSrc->n.u1Write)
2209	&& pPdeDst->n.u1Present)
2210	{
2211	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, pPdeDst->u & SHW_PDE_PG_MASK);
2212	PSHWPT pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
2213	PSHWPTE pPteDst = &pPTDst->a[(GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK];
2214	if ( pPteDst->n.u1Present
2215	&& pPteDst->n.u1Write)
2216	LogFlow(("Writable present page %RGv not marked for dirty bit tracking!!!\n", GCPtrPage));
2217	}
2218	# endif /* VBOX_STRICT */
2219	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2220	return VINF_PGM_NO_DIRTY_BIT_TRACKING;
2221	}
2222	AssertRC(rc);
2223	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2224	return rc;
2225
2226
2227	l_UpperLevelPageFault:
2228	/*
2229	* Pagefault detected while checking the PML4E, PDPE or PDE.
2230	* Single exit handler to get rid of duplicate code paths.
2231	*/
2232	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyTrackRealPF));
2233	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2234	Log(("CheckPageFault: real page fault at %RGv (%d)\n", GCPtrPage, uPageFaultLevel));
2235
2236	if (
2237	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2238	pPml4eSrc->n.u1Present &&
2239	# endif
2240	# if PGM_GST_TYPE == PGM_TYPE_AMD64 \|\| PGM_GST_TYPE == PGM_TYPE_PAE
2241	pPdpeSrc->n.u1Present &&
2242	# endif
2243	pPdeSrc->n.u1Present)
2244	{
2245	/* Check the present bit as the shadow tables can cause different error codes by being out of sync. */
2246	if (pPdeSrc->b.u1Size && fBigPagesSupported)
2247	{
2248	TRPMSetErrorCode(pVM, uErr \| X86_TRAP_PF_P); /* page-level protection violation */
2249	}
2250	else
2251	{
2252	/*
2253	* Map the guest page table.
2254	*/
2255	PGSTPT pPTSrc;
2256	rc = PGM_GCPHYS_2_PTR(pVM, pPdeSrc->u & GST_PDE_PG_MASK, &pPTSrc);
2257	if (RT_SUCCESS(rc))
2258	{
2259	PGSTPTE pPteSrc = &pPTSrc->a[(GCPtrPage >> GST_PT_SHIFT) & GST_PT_MASK];
2260	const GSTPTE PteSrc = *pPteSrc;
2261	if (pPteSrc->n.u1Present)
2262	TRPMSetErrorCode(pVM, uErr \| X86_TRAP_PF_P); /* page-level protection violation */
2263	}
2264	AssertRC(rc);
2265	}
2266	}
2267	return VINF_EM_RAW_GUEST_TRAP;
2268	}
2269	#endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
2270
2271
2272	/**
2273	* Sync a shadow page table.
2274	*
2275	* The shadow page table is not present. This includes the case where
2276	* there is a conflict with a mapping.
2277	*
2278	* @returns VBox status code.
2279	* @param pVM VM handle.
2280	* @param iPD Page directory index.
2281	* @param pPDSrc Source page directory (i.e. Guest OS page directory).
2282	* Assume this is a temporary mapping.
2283	* @param GCPtrPage GC Pointer of the page that caused the fault
2284	*/
2285	PGM_BTH_DECL(int, SyncPT)(PVM pVM, unsigned iPDSrc, PGSTPD pPDSrc, RTGCPTR GCPtrPage)
2286	{
2287	STAM_PROFILE_START(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT), a);
2288	STAM_COUNTER_INC(&pVM->pgm.s.StatSyncPtPD[iPDSrc]);
2289	LogFlow(("SyncPT: GCPtrPage=%RGv\n", GCPtrPage));
2290
2291	#if ( PGM_GST_TYPE == PGM_TYPE_32BIT \
2292	\|\| PGM_GST_TYPE == PGM_TYPE_PAE \
2293	\|\| PGM_GST_TYPE == PGM_TYPE_AMD64) \
2294	&& PGM_SHW_TYPE != PGM_TYPE_NESTED \
2295	&& PGM_SHW_TYPE != PGM_TYPE_EPT
2296
2297	int rc = VINF_SUCCESS;
2298
2299	/*
2300	* Validate input a little bit.
2301	*/
2302	AssertMsg(iPDSrc == ((GCPtrPage >> GST_PD_SHIFT) & GST_PD_MASK), ("iPDSrc=%x GCPtrPage=%RGv\n", iPDSrc, GCPtrPage));
2303	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
2304	const unsigned iPDDst = GCPtrPage >> SHW_PD_SHIFT;
2305	PSHWPDE pPdeDst = pgmShwGet32BitPDEPtr(&pVM->pgm.s, GCPtrPage);
2306
2307	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
2308	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) /& SHW_PD_MASK - only pool index atm! /;
2309	const unsigned iPdpt = (GCPtrPage >> X86_PDPT_SHIFT); NOREF(iPdpt);
2310	PX86PDPT pPdptDst = pgmShwGetPaePDPTPtr(&pVM->pgm.s); NOREF(pPdptDst);
2311	PSHWPDE pPdeDst = pgmShwGetPaePDEPtr(&pVM->pgm.s, GCPtrPage);
2312
2313	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
2314	const unsigned iPdpt = (GCPtrPage >> X86_PDPT_SHIFT) & X86_PDPT_MASK_AMD64;
2315	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
2316	PX86PDPAE pPDDst;
2317	PX86PDPT pPdptDst;
2318	rc = pgmShwGetLongModePDPtr(pVM, GCPtrPage, NULL, &pPdptDst, &pPDDst);
2319	AssertRCSuccessReturn(rc, rc);
2320	Assert(pPDDst);
2321	PSHWPDE pPdeDst = &pPDDst->a[iPDDst];
2322	# endif
2323	SHWPDE PdeDst = *pPdeDst;
2324
2325	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2326	/* Fetch the pgm pool shadow descriptor. */
2327	PPGMPOOLPAGE pShwPde = pgmPoolGetPageByHCPhys(pVM, pPdptDst->a[iPdpt].u & X86_PDPE_PG_MASK);
2328	Assert(pShwPde);
2329	# endif
2330
2331	# ifndef PGM_WITHOUT_MAPPINGS
2332	/*
2333	* Check for conflicts.
2334	* GC: In case of a conflict we'll go to Ring-3 and do a full SyncCR3.
2335	* HC: Simply resolve the conflict.
2336	*/
2337	if (PdeDst.u & PGM_PDFLAGS_MAPPING)
2338	{
2339	Assert(pgmMapAreMappingsEnabled(&pVM->pgm.s));
2340	# ifndef IN_RING3
2341	Log(("SyncPT: Conflict at %RGv\n", GCPtrPage));
2342	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT), a);
2343	return VERR_ADDRESS_CONFLICT;
2344	# else
2345	PPGMMAPPING pMapping = pgmGetMapping(pVM, (RTGCPTR)GCPtrPage);
2346	Assert(pMapping);
2347	# if PGM_GST_TYPE == PGM_TYPE_32BIT
2348	int rc = pgmR3SyncPTResolveConflict(pVM, pMapping, pPDSrc, GCPtrPage & (GST_PD_MASK << GST_PD_SHIFT));
2349	# elif PGM_GST_TYPE == PGM_TYPE_PAE
2350	int rc = pgmR3SyncPTResolveConflictPAE(pVM, pMapping, GCPtrPage & (GST_PD_MASK << GST_PD_SHIFT));
2351	# else
2352	AssertFailed(); /* can't happen for amd64 */
2353	# endif
2354	if (RT_FAILURE(rc))
2355	{
2356	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT), a);
2357	return rc;
2358	}
2359	PdeDst = *pPdeDst;
2360	# endif
2361	}
2362	# else /* PGM_WITHOUT_MAPPINGS */
2363	Assert(!pgmMapAreMappingsEnabled(&pVM->pgm.s));
2364	# endif /* PGM_WITHOUT_MAPPINGS */
2365	Assert(!PdeDst.n.u1Present); /* We're only supposed to call SyncPT on PDE!P and conflicts.*/
2366
2367	/*
2368	* Sync page directory entry.
2369	*/
2370	GSTPDE PdeSrc = pPDSrc->a[iPDSrc];
2371	if (PdeSrc.n.u1Present)
2372	{
2373	/*
2374	* Allocate & map the page table.
2375	*/
2376	PSHWPT pPTDst;
2377	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2378	const bool fPageTable = !PdeSrc.b.u1Size;
2379	# else
2380	const bool fPageTable = !PdeSrc.b.u1Size \|\| !(CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
2381	# endif
2382	PPGMPOOLPAGE pShwPage;
2383	RTGCPHYS GCPhys;
2384	if (fPageTable)
2385	{
2386	GCPhys = PdeSrc.u & GST_PDE_PG_MASK;
2387	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
2388	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
2389	GCPhys \|= (iPDDst & 1) * (PAGE_SIZE / 2);
2390	# endif
2391	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2392	rc = pgmPoolAlloc(pVM, GCPhys, BTH_PGMPOOLKIND_PT_FOR_PT, pShwPde->idx, iPDDst, &pShwPage);
2393	# else
2394	rc = pgmPoolAlloc(pVM, GCPhys, BTH_PGMPOOLKIND_PT_FOR_PT, SHW_POOL_ROOT_IDX, iPDDst, &pShwPage);
2395	# endif
2396	}
2397	else
2398	{
2399	GCPhys = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc);
2400	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
2401	/* Select the right PDE as we're emulating a 4MB page directory with two 2 MB shadow PDEs.*/
2402	GCPhys \|= GCPtrPage & (1 << X86_PD_PAE_SHIFT);
2403	# endif
2404	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2405	rc = pgmPoolAlloc(pVM, GCPhys, BTH_PGMPOOLKIND_PT_FOR_BIG, pShwPde->idx, iPDDst, &pShwPage);
2406	# else
2407	rc = pgmPoolAlloc(pVM, GCPhys, BTH_PGMPOOLKIND_PT_FOR_BIG, SHW_POOL_ROOT_IDX, iPDDst, &pShwPage);
2408	# endif
2409	}
2410	if (rc == VINF_SUCCESS)
2411	pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
2412	else if (rc == VINF_PGM_CACHED_PAGE)
2413	{
2414	/*
2415	* The PT was cached, just hook it up.
2416	*/
2417	if (fPageTable)
2418	PdeDst.u = pShwPage->Core.Key
2419	\| (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));
2420	else
2421	{
2422	PdeDst.u = pShwPage->Core.Key
2423	\| (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));
2424	/* (see explanation and assumptions further down.) */
2425	if (!PdeSrc.b.u1Dirty && PdeSrc.b.u1Write)
2426	{
2427	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPageBig));
2428	PdeDst.u \|= PGM_PDFLAGS_TRACK_DIRTY;
2429	PdeDst.b.u1Write = 0;
2430	}
2431	}
2432	*pPdeDst = PdeDst;
2433	return VINF_SUCCESS;
2434	}
2435	else if (rc == VERR_PGM_POOL_FLUSHED)
2436	{
2437	VM_FF_SET(pVM, VM_FF_PGM_SYNC_CR3);
2438	return VINF_PGM_SYNC_CR3;
2439	}
2440	else
2441	AssertMsgFailedReturn(("rc=%Rrc\n", rc), VERR_INTERNAL_ERROR);
2442	PdeDst.u &= X86_PDE_AVL_MASK;
2443	PdeDst.u \|= pShwPage->Core.Key;
2444
2445	/*
2446	* Page directory has been accessed (this is a fault situation, remember).
2447	*/
2448	pPDSrc->a[iPDSrc].n.u1Accessed = 1;
2449	if (fPageTable)
2450	{
2451	/*
2452	* Page table - 4KB.
2453	*
2454	* Sync all or just a few entries depending on PGM_SYNC_N_PAGES.
2455	*/
2456	Log2(("SyncPT: 4K %RGv PdeSrc:{P=%d RW=%d U=%d raw=%08llx}\n",
2457	GCPtrPage, PdeSrc.b.u1Present, PdeSrc.b.u1Write, PdeSrc.b.u1User, (uint64_t)PdeSrc.u));
2458	PGSTPT pPTSrc;
2459	rc = PGM_GCPHYS_2_PTR(pVM, PdeSrc.u & GST_PDE_PG_MASK, &pPTSrc);
2460	if (RT_SUCCESS(rc))
2461	{
2462	/*
2463	* Start by syncing the page directory entry so CSAM's TLB trick works.
2464	*/
2465	PdeDst.u = (PdeDst.u & (SHW_PDE_PG_MASK \| X86_PDE_AVL_MASK))
2466	\| (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));
2467	*pPdeDst = PdeDst;
2468
2469	/*
2470	* Directory/page user or supervisor privilege: (same goes for read/write)
2471	*
2472	* Directory Page Combined
2473	* U/S U/S U/S
2474	* 0 0 0
2475	* 0 1 0
2476	* 1 0 0
2477	* 1 1 1
2478	*
2479	* Simple AND operation. Table listed for completeness.
2480	*
2481	*/
2482	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT4K));
2483	# ifdef PGM_SYNC_N_PAGES
2484	unsigned iPTBase = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
2485	unsigned iPTDst = iPTBase;
2486	const unsigned iPTDstEnd = RT_MIN(iPTDst + PGM_SYNC_NR_PAGES / 2, RT_ELEMENTS(pPTDst->a));
2487	if (iPTDst <= PGM_SYNC_NR_PAGES / 2)
2488	iPTDst = 0;
2489	else
2490	iPTDst -= PGM_SYNC_NR_PAGES / 2;
2491	# else /* !PGM_SYNC_N_PAGES */
2492	unsigned iPTDst = 0;
2493	const unsigned iPTDstEnd = RT_ELEMENTS(pPTDst->a);
2494	# endif /* !PGM_SYNC_N_PAGES */
2495	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
2496	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
2497	const unsigned offPTSrc = ((GCPtrPage >> SHW_PD_SHIFT) & 1) * 512;
2498	# else
2499	const unsigned offPTSrc = 0;
2500	# endif
2501	for (; iPTDst < iPTDstEnd; iPTDst++)
2502	{
2503	const unsigned iPTSrc = iPTDst + offPTSrc;
2504	const GSTPTE PteSrc = pPTSrc->a[iPTSrc];
2505
2506	if (PteSrc.n.u1Present) /* we've already cleared it above */
2507	{
2508	# ifndef IN_RING0
2509	/*
2510	* Assuming kernel code will be marked as supervisor - and not as user level
2511	* and executed using a conforming code selector - And marked as readonly.
2512	* Also assume that if we're monitoring a page, it's of no interest to CSAM.
2513	*/
2514	PPGMPAGE pPage;
2515	if ( ((PdeSrc.u & pPTSrc->a[iPTSrc].u) & (X86_PTE_RW \| X86_PTE_US))
2516	\|\| !CSAMDoesPageNeedScanning(pVM, (RTRCPTR)((iPDSrc << GST_PD_SHIFT) \| (iPTSrc << PAGE_SHIFT)))
2517	\|\| ( (pPage = pgmPhysGetPage(&pVM->pgm.s, PteSrc.u & GST_PTE_PG_MASK))
2518	&& PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
2519	)
2520	# endif
2521	PGM_BTH_NAME(SyncPageWorker)(pVM, &pPTDst->a[iPTDst], PdeSrc, PteSrc, pShwPage, iPTDst);
2522	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",
2523	(RTGCPTR)((iPDSrc << GST_PD_SHIFT) \| (iPTSrc << PAGE_SHIFT)),
2524	PteSrc.n.u1Present,
2525	PteSrc.n.u1Write & PdeSrc.n.u1Write,
2526	PteSrc.n.u1User & PdeSrc.n.u1User,
2527	(uint64_t)PteSrc.u,
2528	pPTDst->a[iPTDst].u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : "", pPTDst->a[iPTDst].u, iPTSrc, PdeSrc.au32[0],
2529	(RTGCPHYS)((PdeSrc.u & GST_PDE_PG_MASK) + iPTSrc*sizeof(PteSrc)) ));
2530	}
2531	} /* for PTEs */
2532	}
2533	}
2534	else
2535	{
2536	/*
2537	* Big page - 2/4MB.
2538	*
2539	* We'll walk the ram range list in parallel and optimize lookups.
2540	* We will only sync on shadow page table at a time.
2541	*/
2542	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT4M));
2543
2544	/**
2545	* @todo It might be more efficient to sync only a part of the 4MB page (similar to what we do for 4kb PDs).
2546	*/
2547
2548	/*
2549	* Start by syncing the page directory entry.
2550	*/
2551	PdeDst.u = (PdeDst.u & (SHW_PDE_PG_MASK \| (X86_PDE_AVL_MASK & ~PGM_PDFLAGS_TRACK_DIRTY)))
2552	\| (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));
2553
2554	/*
2555	* If the page is not flagged as dirty and is writable, then make it read-only
2556	* at PD level, so we can set the dirty bit when the page is modified.
2557	*
2558	* ASSUMES that page access handlers are implemented on page table entry level.
2559	* Thus we will first catch the dirty access and set PDE.D and restart. If
2560	* there is an access handler, we'll trap again and let it work on the problem.
2561	*/
2562	/** @todo move the above stuff to a section in the PGM documentation. */
2563	Assert(!(PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY));
2564	if (!PdeSrc.b.u1Dirty && PdeSrc.b.u1Write)
2565	{
2566	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPageBig));
2567	PdeDst.u \|= PGM_PDFLAGS_TRACK_DIRTY;
2568	PdeDst.b.u1Write = 0;
2569	}
2570	*pPdeDst = PdeDst;
2571
2572	/*
2573	* Fill the shadow page table.
2574	*/
2575	/* Get address and flags from the source PDE. */
2576	SHWPTE PteDstBase;
2577	PteDstBase.u = PdeSrc.u & ~(GST_PDE_PG_MASK \| X86_PTE_AVL_MASK \| X86_PTE_PAT \| X86_PTE_PCD \| X86_PTE_PWT);
2578
2579	/* Loop thru the entries in the shadow PT. */
2580	const RTGCPTR GCPtr = (GCPtrPage >> SHW_PD_SHIFT) << SHW_PD_SHIFT; NOREF(GCPtr);
2581	Log2(("SyncPT: BIG %RGv PdeSrc:{P=%d RW=%d U=%d raw=%08llx} Shw=%RGv GCPhys=%RGp %s\n",
2582	GCPtrPage, PdeSrc.b.u1Present, PdeSrc.b.u1Write, PdeSrc.b.u1User, (uint64_t)PdeSrc.u, GCPtr,
2583	GCPhys, PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
2584	PPGMRAMRANGE pRam = pVM->pgm.s.CTX_SUFF(pRamRanges);
2585	unsigned iPTDst = 0;
2586	while (iPTDst < RT_ELEMENTS(pPTDst->a))
2587	{
2588	/* Advance ram range list. */
2589	while (pRam && GCPhys > pRam->GCPhysLast)
2590	pRam = pRam->CTX_SUFF(pNext);
2591	if (pRam && GCPhys >= pRam->GCPhys)
2592	{
2593	unsigned iHCPage = (GCPhys - pRam->GCPhys) >> PAGE_SHIFT;
2594	do
2595	{
2596	/* Make shadow PTE. */
2597	PPGMPAGE pPage = &pRam->aPages[iHCPage];
2598	SHWPTE PteDst;
2599
2600	/* Make sure the RAM has already been allocated. */
2601	if (pRam->fFlags & MM_RAM_FLAGS_DYNAMIC_ALLOC) /** @todo PAGE FLAGS */
2602	{
2603	if (RT_UNLIKELY(!PGM_PAGE_GET_HCPHYS(pPage)))
2604	{
2605	# ifdef IN_RING3
2606	int rc = pgmr3PhysGrowRange(pVM, GCPhys);
2607	# else
2608	int rc = CTXALLMID(VMM, CallHost)(pVM, VMMCALLHOST_PGM_RAM_GROW_RANGE, GCPhys);
2609	# endif
2610	if (rc != VINF_SUCCESS)
2611	return rc;
2612	}
2613	}
2614
2615	if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
2616	{
2617	if (!PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage))
2618	{
2619	PteDst.u = PGM_PAGE_GET_HCPHYS(pPage) \| PteDstBase.u;
2620	PteDst.n.u1Write = 0;
2621	}
2622	else
2623	PteDst.u = 0;
2624	}
2625	# ifndef IN_RING0
2626	/*
2627	* Assuming kernel code will be marked as supervisor and not as user level and executed
2628	* using a conforming code selector. Don't check for readonly, as that implies the whole
2629	* 4MB can be code or readonly data. Linux enables write access for its large pages.
2630	*/
2631	else if ( !PdeSrc.n.u1User
2632	&& CSAMDoesPageNeedScanning(pVM, (RTRCPTR)(GCPtr \| (iPTDst << SHW_PT_SHIFT))))
2633	PteDst.u = 0;
2634	# endif
2635	else
2636	PteDst.u = PGM_PAGE_GET_HCPHYS(pPage) \| PteDstBase.u;
2637	# ifdef PGMPOOL_WITH_USER_TRACKING
2638	if (PteDst.n.u1Present)
2639	PGM_BTH_NAME(SyncPageWorkerTrackAddref)(pVM, pShwPage, pPage->HCPhys >> MM_RAM_FLAGS_IDX_SHIFT, pPage, iPTDst); /** @todo PAGE FLAGS */
2640	# endif
2641	/* commit it */
2642	pPTDst->a[iPTDst] = PteDst;
2643	Log4(("SyncPT: BIG %RGv PteDst:{P=%d RW=%d U=%d raw=%08llx}%s\n",
2644	(RTGCPTR)(GCPtr \| (iPTDst << SHW_PT_SHIFT)), PteDst.n.u1Present, PteDst.n.u1Write, PteDst.n.u1User, (uint64_t)PteDst.u,
2645	PteDst.u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
2646
2647	/* advance */
2648	GCPhys += PAGE_SIZE;
2649	iHCPage++;
2650	iPTDst++;
2651	} while ( iPTDst < RT_ELEMENTS(pPTDst->a)
2652	&& GCPhys <= pRam->GCPhysLast);
2653	}
2654	else if (pRam)
2655	{
2656	Log(("Invalid pages at %RGp\n", GCPhys));
2657	do
2658	{
2659	pPTDst->a[iPTDst].u = 0; /* MMIO or invalid page, we must handle them manually. */
2660	GCPhys += PAGE_SIZE;
2661	iPTDst++;
2662	} while ( iPTDst < RT_ELEMENTS(pPTDst->a)
2663	&& GCPhys < pRam->GCPhys);
2664	}
2665	else
2666	{
2667	Log(("Invalid pages at %RGp (2)\n", GCPhys));
2668	for ( ; iPTDst < RT_ELEMENTS(pPTDst->a); iPTDst++)
2669	pPTDst->a[iPTDst].u = 0; /* MMIO or invalid page, we must handle them manually. */
2670	}
2671	} /* while more PTEs */
2672	} /* 4KB / 4MB */
2673	}
2674	else
2675	AssertRelease(!PdeDst.n.u1Present);
2676
2677	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT), a);
2678	if (RT_FAILURE(rc))
2679	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPTFailed));
2680	return rc;
2681
2682	#elif (PGM_GST_TYPE == PGM_TYPE_REAL \|\| PGM_GST_TYPE == PGM_TYPE_PROT) \
2683	&& PGM_SHW_TYPE != PGM_TYPE_NESTED \
2684	&& (PGM_SHW_TYPE != PGM_TYPE_EPT \|\| PGM_GST_TYPE == PGM_TYPE_PROT)
2685
2686
2687	/*
2688	* Validate input a little bit.
2689	*/
2690	int rc = VINF_SUCCESS;
2691	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
2692	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
2693	PSHWPDE pPdeDst = pgmShwGet32BitPDEPtr(&pVM->pgm.s, GCPtrPage);
2694
2695	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
2696	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) /& SHW_PD_MASK - only pool index atm!/;
2697	PX86PDEPAE pPdeDst = pgmShwGetPaePDEPtr(&pVM->pgm.s, GCPtrPage);
2698
2699	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
2700	const unsigned iPdpt = (GCPtrPage >> X86_PDPT_SHIFT) & X86_PDPT_MASK_AMD64;
2701	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
2702	PX86PDPAE pPDDst;
2703	PX86PDPT pPdptDst;
2704	rc = pgmShwGetLongModePDPtr(pVM, GCPtrPage, NULL, &pPdptDst, &pPDDst);
2705	AssertRCSuccessReturn(rc, rc);
2706	Assert(pPDDst);
2707	PSHWPDE pPdeDst = &pPDDst->a[iPDDst];
2708
2709	/* Fetch the pgm pool shadow descriptor. */
2710	PPGMPOOLPAGE pShwPde = pgmPoolGetPageByHCPhys(pVM, pPdptDst->a[iPdpt].u & X86_PDPE_PG_MASK);
2711	Assert(pShwPde);
2712
2713	# elif PGM_SHW_TYPE == PGM_TYPE_EPT
2714	const unsigned iPdpt = (GCPtrPage >> EPT_PDPT_SHIFT) & EPT_PDPT_MASK;
2715	const unsigned iPDDst = ((GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK);
2716	PEPTPD pPDDst;
2717	PEPTPDPT pPdptDst;
2718
2719	rc = pgmShwGetEPTPDPtr(pVM, GCPtrPage, &pPdptDst, &pPDDst);
2720	if (rc != VINF_SUCCESS)
2721	{
2722	AssertRC(rc);
2723	return rc;
2724	}
2725	Assert(pPDDst);
2726	PSHWPDE pPdeDst = &pPDDst->a[iPDDst];
2727
2728	/* Fetch the pgm pool shadow descriptor. */
2729	PPGMPOOLPAGE pShwPde = pgmPoolGetPageByHCPhys(pVM, pPdptDst->a[iPdpt].u & EPT_PDPTE_PG_MASK);
2730	Assert(pShwPde);
2731	# endif
2732	SHWPDE PdeDst = *pPdeDst;
2733
2734	Assert(!(PdeDst.u & PGM_PDFLAGS_MAPPING));
2735	Assert(!PdeDst.n.u1Present); /* We're only supposed to call SyncPT on PDE!P and conflicts.*/
2736
2737	GSTPDE PdeSrc;
2738	PdeSrc.au32[0] = 0; /* faked so we don't have to #ifdef everything */
2739	PdeSrc.n.u1Present = 1;
2740	PdeSrc.n.u1Write = 1;
2741	PdeSrc.n.u1Accessed = 1;
2742	PdeSrc.n.u1User = 1;
2743
2744	/*
2745	* Allocate & map the page table.
2746	*/
2747	PSHWPT pPTDst;
2748	PPGMPOOLPAGE pShwPage;
2749	RTGCPHYS GCPhys;
2750
2751	/* Virtual address = physical address */
2752	GCPhys = GCPtrPage & X86_PAGE_4K_BASE_MASK;
2753	# if PGM_SHW_TYPE == PGM_TYPE_AMD64 \|\| PGM_SHW_TYPE == PGM_TYPE_EPT
2754	rc = pgmPoolAlloc(pVM, GCPhys & ~(RT_BIT_64(SHW_PD_SHIFT) - 1), BTH_PGMPOOLKIND_PT_FOR_PT, pShwPde->idx, iPDDst, &pShwPage);
2755	# else
2756	rc = pgmPoolAlloc(pVM, GCPhys & ~(RT_BIT_64(SHW_PD_SHIFT) - 1), BTH_PGMPOOLKIND_PT_FOR_PT, SHW_POOL_ROOT_IDX, iPDDst, &pShwPage);
2757	# endif
2758
2759	if ( rc == VINF_SUCCESS
2760	\|\| rc == VINF_PGM_CACHED_PAGE)
2761	pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
2762	else
2763	AssertMsgFailedReturn(("rc=%Rrc\n", rc), VERR_INTERNAL_ERROR);
2764
2765	PdeDst.u &= X86_PDE_AVL_MASK;
2766	PdeDst.u \|= pShwPage->Core.Key;
2767	PdeDst.n.u1Present = 1;
2768	PdeDst.n.u1Write = 1;
2769	# if PGM_SHW_TYPE == PGM_TYPE_EPT
2770	PdeDst.n.u1Execute = 1;
2771	# else
2772	PdeDst.n.u1User = 1;
2773	PdeDst.n.u1Accessed = 1;
2774	# endif
2775	*pPdeDst = PdeDst;
2776
2777	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, GCPtrPage, PGM_SYNC_NR_PAGES, 0 /* page not present */);
2778	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT), a);
2779	return rc;
2780
2781	#else
2782	AssertReleaseMsgFailed(("Shw=%d Gst=%d is not implemented!\n", PGM_SHW_TYPE, PGM_GST_TYPE));
2783	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT), a);
2784	return VERR_INTERNAL_ERROR;
2785	#endif
2786	}
2787
2788
2789
2790	/**
2791	* Prefetch a page/set of pages.
2792	*
2793	* Typically used to sync commonly used pages before entering raw mode
2794	* after a CR3 reload.
2795	*
2796	* @returns VBox status code.
2797	* @param pVM VM handle.
2798	* @param GCPtrPage Page to invalidate.
2799	*/
2800	PGM_BTH_DECL(int, PrefetchPage)(PVM pVM, RTGCPTR GCPtrPage)
2801	{
2802	#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) \
2803	&& PGM_SHW_TYPE != PGM_TYPE_NESTED && PGM_SHW_TYPE != PGM_TYPE_EPT
2804	/*
2805	* Check that all Guest levels thru the PDE are present, getting the
2806	* PD and PDE in the processes.
2807	*/
2808	int rc = VINF_SUCCESS;
2809	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
2810	# if PGM_GST_TYPE == PGM_TYPE_32BIT
2811	const unsigned iPDSrc = GCPtrPage >> GST_PD_SHIFT;
2812	PGSTPD pPDSrc = pgmGstGet32bitPDPtr(&pVM->pgm.s);
2813	# elif PGM_GST_TYPE == PGM_TYPE_PAE
2814	unsigned iPDSrc;
2815	PGSTPD pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, GCPtrPage, &iPDSrc, NULL);
2816	if (!pPDSrc)
2817	return VINF_SUCCESS; /* not present */
2818	# elif PGM_GST_TYPE == PGM_TYPE_AMD64
2819	unsigned iPDSrc;
2820	PX86PML4E pPml4eSrc;
2821	X86PDPE PdpeSrc;
2822	PGSTPD pPDSrc = pgmGstGetLongModePDPtr(&pVM->pgm.s, GCPtrPage, &pPml4eSrc, &PdpeSrc, &iPDSrc);
2823	if (!pPDSrc)
2824	return VINF_SUCCESS; /* not present */
2825	# endif
2826	const GSTPDE PdeSrc = pPDSrc->a[iPDSrc];
2827	# else
2828	PGSTPD pPDSrc = NULL;
2829	const unsigned iPDSrc = 0;
2830	GSTPDE PdeSrc;
2831
2832	PdeSrc.au32[0] = 0; /* faked so we don't have to #ifdef everything */
2833	PdeSrc.n.u1Present = 1;
2834	PdeSrc.n.u1Write = 1;
2835	PdeSrc.n.u1Accessed = 1;
2836	PdeSrc.n.u1User = 1;
2837	# endif
2838
2839	if (PdeSrc.n.u1Present && PdeSrc.n.u1Accessed)
2840	{
2841	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
2842	const X86PDE PdeDst = pgmShwGet32BitPDE(&pVM->pgm.s, GCPtrPage);
2843	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
2844	const X86PDEPAE PdeDst = pgmShwGetPaePDE(&pVM->pgm.s, GCPtrPage);
2845	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
2846	const unsigned iPDDst = ((GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK);
2847	PX86PDPAE pPDDst;
2848	X86PDEPAE PdeDst;
2849
2850	# if PGM_GST_TYPE == PGM_TYPE_PROT
2851	/* AMD-V nested paging */
2852	X86PML4E Pml4eSrc;
2853	X86PDPE PdpeSrc;
2854	PX86PML4E pPml4eSrc = &Pml4eSrc;
2855
2856	/* Fake PML4 & PDPT entry; access control handled on the page table level, so allow everything. */
2857	Pml4eSrc.u = X86_PML4E_P \| X86_PML4E_RW \| X86_PML4E_US \| X86_PML4E_NX \| X86_PML4E_A;
2858	PdpeSrc.u = X86_PDPE_P \| X86_PDPE_RW \| X86_PDPE_US \| X86_PDPE_NX \| X86_PDPE_A;
2859	# endif
2860
2861	int rc = pgmShwSyncLongModePDPtr(pVM, GCPtrPage, pPml4eSrc, &PdpeSrc, &pPDDst);
2862	if (rc != VINF_SUCCESS)
2863	{
2864	AssertRC(rc);
2865	return rc;
2866	}
2867	Assert(pPDDst);
2868	PdeDst = pPDDst->a[iPDDst];
2869	# endif
2870	if (!(PdeDst.u & PGM_PDFLAGS_MAPPING))
2871	{
2872	if (!PdeDst.n.u1Present)
2873	/** r=bird: This guy will set the A bit on the PDE, probably harmless. */
2874	rc = PGM_BTH_NAME(SyncPT)(pVM, iPDSrc, pPDSrc, GCPtrPage);
2875	else
2876	{
2877	/** @note We used to sync PGM_SYNC_NR_PAGES pages, which triggered assertions in CSAM, because
2878	* R/W attributes of nearby pages were reset. Not sure how that could happen. Anyway, it
2879	* makes no sense to prefetch more than one page.
2880	*/
2881	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, GCPtrPage, 1, 0);
2882	if (RT_SUCCESS(rc))
2883	rc = VINF_SUCCESS;
2884	}
2885	}
2886	}
2887	return rc;
2888
2889	#elif PGM_SHW_TYPE == PGM_TYPE_NESTED \|\| PGM_SHW_TYPE == PGM_TYPE_EPT
2890	return VINF_SUCCESS; /* ignore */
2891	#endif
2892	}
2893
2894
2895
2896
2897	/**
2898	* Syncs a page during a PGMVerifyAccess() call.
2899	*
2900	* @returns VBox status code (informational included).
2901	* @param GCPtrPage The address of the page to sync.
2902	* @param fPage The effective guest page flags.
2903	* @param uErr The trap error code.
2904	*/
2905	PGM_BTH_DECL(int, VerifyAccessSyncPage)(PVM pVM, RTGCPTR GCPtrPage, unsigned fPage, unsigned uErr)
2906	{
2907	LogFlow(("VerifyAccessSyncPage: GCPtrPage=%RGv fPage=%#x uErr=%#x\n", GCPtrPage, fPage, uErr));
2908
2909	Assert(!HWACCMIsNestedPagingActive(pVM));
2910	#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) \
2911	&& PGM_SHW_TYPE != PGM_TYPE_NESTED && PGM_SHW_TYPE != PGM_TYPE_EPT
2912
2913	# ifndef IN_RING0
2914	if (!(fPage & X86_PTE_US))
2915	{
2916	/*
2917	* Mark this page as safe.
2918	*/
2919	/** @todo not correct for pages that contain both code and data!! */
2920	Log(("CSAMMarkPage %RGv; scanned=%d\n", GCPtrPage, true));
2921	CSAMMarkPage(pVM, (RTRCPTR)GCPtrPage, true);
2922	}
2923	# endif
2924
2925	/*
2926	* Get guest PD and index.
2927	*/
2928	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
2929	# if PGM_GST_TYPE == PGM_TYPE_32BIT
2930	const unsigned iPDSrc = GCPtrPage >> GST_PD_SHIFT;
2931	PGSTPD pPDSrc = pgmGstGet32bitPDPtr(&pVM->pgm.s);
2932	# elif PGM_GST_TYPE == PGM_TYPE_PAE
2933	unsigned iPDSrc;
2934	PGSTPD pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, GCPtrPage, &iPDSrc, NULL);
2935
2936	if (pPDSrc)
2937	{
2938	Log(("PGMVerifyAccess: access violation for %RGv due to non-present PDPTR\n", GCPtrPage));
2939	return VINF_EM_RAW_GUEST_TRAP;
2940	}
2941	# elif PGM_GST_TYPE == PGM_TYPE_AMD64
2942	unsigned iPDSrc;
2943	PX86PML4E pPml4eSrc;
2944	X86PDPE PdpeSrc;
2945	PGSTPD pPDSrc = pgmGstGetLongModePDPtr(&pVM->pgm.s, GCPtrPage, &pPml4eSrc, &PdpeSrc, &iPDSrc);
2946	if (!pPDSrc)
2947	{
2948	Log(("PGMVerifyAccess: access violation for %RGv due to non-present PDPTR\n", GCPtrPage));
2949	return VINF_EM_RAW_GUEST_TRAP;
2950	}
2951	# endif
2952	# else
2953	PGSTPD pPDSrc = NULL;
2954	const unsigned iPDSrc = 0;
2955	# endif
2956	int rc = VINF_SUCCESS;
2957
2958	/*
2959	* First check if the shadow pd is present.
2960	*/
2961	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
2962	PX86PDE pPdeDst = pgmShwGet32BitPDEPtr(&pVM->pgm.s, GCPtrPage);
2963	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
2964	PX86PDEPAE pPdeDst = pgmShwGetPaePDEPtr(&pVM->pgm.s, GCPtrPage);
2965	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
2966	const unsigned iPDDst = ((GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK);
2967	PX86PDPAE pPDDst;
2968	PX86PDEPAE pPdeDst;
2969
2970	# if PGM_GST_TYPE == PGM_TYPE_PROT
2971	/* AMD-V nested paging */
2972	X86PML4E Pml4eSrc;
2973	X86PDPE PdpeSrc;
2974	PX86PML4E pPml4eSrc = &Pml4eSrc;
2975
2976	/* Fake PML4 & PDPT entry; access control handled on the page table level, so allow everything. */
2977	Pml4eSrc.u = X86_PML4E_P \| X86_PML4E_RW \| X86_PML4E_US \| X86_PML4E_NX \| X86_PML4E_A;
2978	PdpeSrc.u = X86_PDPE_P \| X86_PDPE_RW \| X86_PDPE_US \| X86_PDPE_NX \| X86_PDPE_A;
2979	# endif
2980
2981	rc = pgmShwSyncLongModePDPtr(pVM, GCPtrPage, pPml4eSrc, &PdpeSrc, &pPDDst);
2982	if (rc != VINF_SUCCESS)
2983	{
2984	AssertRC(rc);
2985	return rc;
2986	}
2987	Assert(pPDDst);
2988	pPdeDst = &pPDDst->a[iPDDst];
2989	# endif
2990	if (!pPdeDst->n.u1Present)
2991	{
2992	rc = PGM_BTH_NAME(SyncPT)(pVM, iPDSrc, pPDSrc, GCPtrPage);
2993	AssertRC(rc);
2994	if (rc != VINF_SUCCESS)
2995	return rc;
2996	}
2997
2998	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
2999	/* Check for dirty bit fault */
3000	rc = PGM_BTH_NAME(CheckPageFault)(pVM, uErr, pPdeDst, &pPDSrc->a[iPDSrc], GCPtrPage);
3001	if (rc == VINF_PGM_HANDLED_DIRTY_BIT_FAULT)
3002	Log(("PGMVerifyAccess: success (dirty)\n"));
3003	else
3004	{
3005	GSTPDE PdeSrc = pPDSrc->a[iPDSrc];
3006	#else
3007	{
3008	GSTPDE PdeSrc;
3009	PdeSrc.au32[0] = 0; /* faked so we don't have to #ifdef everything */
3010	PdeSrc.n.u1Present = 1;
3011	PdeSrc.n.u1Write = 1;
3012	PdeSrc.n.u1Accessed = 1;
3013	PdeSrc.n.u1User = 1;
3014
3015	#endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
3016	Assert(rc != VINF_EM_RAW_GUEST_TRAP);
3017	if (uErr & X86_TRAP_PF_US)
3018	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,PageOutOfSyncUser));
3019	else /* supervisor */
3020	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,PageOutOfSyncSupervisor));
3021
3022	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, GCPtrPage, 1, 0);
3023	if (RT_SUCCESS(rc))
3024	{
3025	/* Page was successfully synced */
3026	Log2(("PGMVerifyAccess: success (sync)\n"));
3027	rc = VINF_SUCCESS;
3028	}
3029	else
3030	{
3031	Log(("PGMVerifyAccess: access violation for %RGv rc=%d\n", GCPtrPage, rc));
3032	return VINF_EM_RAW_GUEST_TRAP;
3033	}
3034	}
3035	return rc;
3036
3037	#else /* PGM_GST_TYPE != PGM_TYPE_32BIT */
3038
3039	AssertReleaseMsgFailed(("Shw=%d Gst=%d is not implemented!\n", PGM_GST_TYPE, PGM_SHW_TYPE));
3040	return VERR_INTERNAL_ERROR;
3041	#endif /* PGM_GST_TYPE != PGM_TYPE_32BIT */
3042	}
3043
3044
3045	#if PGM_GST_TYPE == PGM_TYPE_32BIT \|\| PGM_GST_TYPE == PGM_TYPE_PAE \|\| PGM_GST_TYPE == PGM_TYPE_AMD64
3046	# if PGM_SHW_TYPE == PGM_TYPE_32BIT \|\| PGM_SHW_TYPE == PGM_TYPE_PAE \|\| PGM_SHW_TYPE == PGM_TYPE_AMD64
3047	/**
3048	* Figures out which kind of shadow page this guest PDE warrants.
3049	*
3050	* @returns Shadow page kind.
3051	* @param pPdeSrc The guest PDE in question.
3052	* @param cr4 The current guest cr4 value.
3053	*/
3054	DECLINLINE(PGMPOOLKIND) PGM_BTH_NAME(CalcPageKind)(const GSTPDE *pPdeSrc, uint32_t cr4)
3055	{
3056	# if PMG_GST_TYPE == PGM_TYPE_AMD64
3057	if (!pPdeSrc->n.u1Size)
3058	# else
3059	if (!pPdeSrc->n.u1Size \|\| !(cr4 & X86_CR4_PSE))
3060	# endif
3061	return BTH_PGMPOOLKIND_PT_FOR_PT;
3062	//switch (pPdeSrc->u & (X86_PDE4M_RW \| X86_PDE4M_US /\| X86_PDE4M_PAE_NX/))
3063	//{
3064	// case 0:
3065	// return BTH_PGMPOOLKIND_PT_FOR_BIG_RO;
3066	// case X86_PDE4M_RW:
3067	// return BTH_PGMPOOLKIND_PT_FOR_BIG_RW;
3068	// case X86_PDE4M_US:
3069	// return BTH_PGMPOOLKIND_PT_FOR_BIG_US;
3070	// case X86_PDE4M_RW \| X86_PDE4M_US:
3071	// return BTH_PGMPOOLKIND_PT_FOR_BIG_RW_US;
3072	# if 0
3073	// case X86_PDE4M_PAE_NX:
3074	// return BTH_PGMPOOLKIND_PT_FOR_BIG_NX;
3075	// case X86_PDE4M_RW \| X86_PDE4M_PAE_NX:
3076	// return BTH_PGMPOOLKIND_PT_FOR_BIG_RW_NX;
3077	// case X86_PDE4M_US \| X86_PDE4M_PAE_NX:
3078	// return BTH_PGMPOOLKIND_PT_FOR_BIG_US_NX;
3079	// case X86_PDE4M_RW \| X86_PDE4M_US \| X86_PDE4M_PAE_NX:
3080	// return BTH_PGMPOOLKIND_PT_FOR_BIG_RW_US_NX;
3081	# endif
3082	return BTH_PGMPOOLKIND_PT_FOR_BIG;
3083	//}
3084	}
3085	# endif
3086	#endif
3087
3088	#undef MY_STAM_COUNTER_INC
3089	#define MY_STAM_COUNTER_INC(a) do { } while (0)
3090
3091
3092	/**
3093	* Syncs the paging hierarchy starting at CR3.
3094	*
3095	* @returns VBox status code, no specials.
3096	* @param pVM The virtual machine.
3097	* @param cr0 Guest context CR0 register
3098	* @param cr3 Guest context CR3 register
3099	* @param cr4 Guest context CR4 register
3100	* @param fGlobal Including global page directories or not
3101	*/
3102	PGM_BTH_DECL(int, SyncCR3)(PVM pVM, uint64_t cr0, uint64_t cr3, uint64_t cr4, bool fGlobal)
3103	{
3104	if (VM_FF_ISSET(pVM, VM_FF_PGM_SYNC_CR3))
3105	fGlobal = true; /* Change this CR3 reload to be a global one. */
3106
3107	#if PGM_SHW_TYPE != PGM_TYPE_NESTED && PGM_SHW_TYPE != PGM_TYPE_EPT
3108	/*
3109	* Update page access handlers.
3110	* The virtual are always flushed, while the physical are only on demand.
3111	* WARNING: We are incorrectly not doing global flushing on Virtual Handler updates. We'll
3112	* have to look into that later because it will have a bad influence on the performance.
3113	* @note SvL: There's no need for that. Just invalidate the virtual range(s).
3114	* bird: Yes, but that won't work for aliases.
3115	*/
3116	/** @todo this MUST go away. See #1557. */
3117	STAM_PROFILE_START(&pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3Handlers), h);
3118	PGM_GST_NAME(HandlerVirtualUpdate)(pVM, cr4);
3119	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3Handlers), h);
3120	#endif
3121
3122	#if PGM_SHW_TYPE == PGM_TYPE_NESTED \|\| PGM_SHW_TYPE == PGM_TYPE_EPT
3123	/*
3124	* Nested / EPT - almost no work.
3125	*/
3126	/** @todo check if this is really necessary; the call does it as well... */
3127	HWACCMFlushTLB(pVM);
3128	return VINF_SUCCESS;
3129
3130	#elif PGM_SHW_TYPE == PGM_TYPE_AMD64 \|\| defined(VBOX_WITH_PGMPOOL_PAGING_ONLY)
3131	/*
3132	* AMD64 (Shw & Gst) - No need to check all paging levels; we zero
3133	* out the shadow parts when the guest modifies its tables.
3134	*/
3135	return VINF_SUCCESS;
3136
3137	#else /* PGM_SHW_TYPE != PGM_TYPE_NESTED && PGM_SHW_TYPE != PGM_TYPE_EPT && PGM_SHW_TYPE != PGM_TYPE_AMD64 */
3138	/*
3139	* PAE and 32-bit legacy mode (shadow).
3140	* (Guest PAE, 32-bit legacy, protected and real modes.)
3141	*/
3142	Assert(fGlobal \|\| (cr4 & X86_CR4_PGE));
3143	MY_STAM_COUNTER_INC(fGlobal ? &pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3Global) : &pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3NotGlobal));
3144
3145	# if PGM_GST_TYPE == PGM_TYPE_32BIT \|\| PGM_GST_TYPE == PGM_TYPE_PAE
3146	bool const fBigPagesSupported = !!(CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
3147
3148	/*
3149	* Get page directory addresses.
3150	*/
3151	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
3152	PX86PDE pPDEDst = pgmShwGet32BitPDEPtr(&pVM->pgm.s, 0);
3153	# else /* PGM_SHW_TYPE == PGM_TYPE_PAE */
3154	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3155	PX86PDEPAE pPDEDst = NULL;
3156	# endif
3157	# endif
3158
3159	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3160	PGSTPD pPDSrc = pgmGstGet32bitPDPtr(&pVM->pgm.s);
3161	Assert(pPDSrc);
3162	# if !defined(IN_RC) && !defined(VBOX_WITH_2X_4GB_ADDR_SPACE_IN_R0)
3163	Assert(PGMPhysGCPhys2R3PtrAssert(pVM, (RTGCPHYS)(cr3 & GST_CR3_PAGE_MASK), sizeof(*pPDSrc)) == (RTR3PTR)pPDSrc);
3164	# endif
3165	# endif /* PGM_GST_TYPE == PGM_TYPE_32BIT */
3166
3167	/*
3168	* Iterate the the CR3 page.
3169	*/
3170	PPGMMAPPING pMapping;
3171	unsigned iPdNoMapping;
3172	const bool fRawR0Enabled = EMIsRawRing0Enabled(pVM);
3173	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
3174
3175	/* Only check mappings if they are supposed to be put into the shadow page table. */
3176	if (pgmMapAreMappingsEnabled(&pVM->pgm.s))
3177	{
3178	pMapping = pVM->pgm.s.CTX_SUFF(pMappings);
3179	iPdNoMapping = (pMapping) ? (pMapping->GCPtr >> GST_PD_SHIFT) : ~0U;
3180	}
3181	else
3182	{
3183	pMapping = 0;
3184	iPdNoMapping = ~0U;
3185	}
3186
3187	# if PGM_GST_TYPE == PGM_TYPE_PAE
3188	for (uint64_t iPdpt = 0; iPdpt < GST_PDPE_ENTRIES; iPdpt++)
3189	{
3190	unsigned iPDSrc;
3191	X86PDPE PdpeSrc;
3192	PGSTPD pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, iPdpt << X86_PDPT_SHIFT, &iPDSrc, &PdpeSrc);
3193	PX86PDEPAE pPDEDst = pgmShwGetPaePDEPtr(&pVM->pgm.s, iPdpt << X86_PDPT_SHIFT);
3194	PX86PDPT pPdptDst = pgmShwGetPaePDPTPtr(&pVM->pgm.s);
3195
3196	if (pPDSrc == NULL)
3197	{
3198	/* PDPE not present */
3199	if (pPdptDst->a[iPdpt].n.u1Present)
3200	{
3201	LogFlow(("SyncCR3: guest PDPE %lld not present; clear shw pdpe\n", iPdpt));
3202	/* for each page directory entry */
3203	for (unsigned iPD = 0; iPD < RT_ELEMENTS(pPDSrc->a); iPD++)
3204	{
3205	if ( pPDEDst[iPD].n.u1Present
3206	&& !(pPDEDst[iPD].u & PGM_PDFLAGS_MAPPING))
3207	{
3208	pgmPoolFree(pVM, pPDEDst[iPD].u & SHW_PDE_PG_MASK, SHW_POOL_ROOT_IDX, iPdpt * X86_PG_PAE_ENTRIES + iPD);
3209	pPDEDst[iPD].u = 0;
3210	}
3211	}
3212	}
3213	if (!(pPdptDst->a[iPdpt].u & PGM_PLXFLAGS_MAPPING))
3214	pPdptDst->a[iPdpt].n.u1Present = 0;
3215	continue;
3216	}
3217	# else /* PGM_GST_TYPE != PGM_TYPE_PAE */
3218	{
3219	# endif /* PGM_GST_TYPE != PGM_TYPE_PAE */
3220	for (unsigned iPD = 0; iPD < RT_ELEMENTS(pPDSrc->a); iPD++)
3221	{
3222	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
3223	if ((iPD & 255) == 0) /* Start of new PD. */
3224	pPDEDst = pgmShwGetPaePDEPtr(&pVM->pgm.s, (uint32_t)iPD << GST_PD_SHIFT);
3225	# endif
3226	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
3227	Assert(pgmShwGet32BitPDEPtr(&pVM->pgm.s, (uint32_t)iPD << SHW_PD_SHIFT) == pPDEDst);
3228	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
3229	# if defined(VBOX_STRICT) && !defined(VBOX_WITH_2X_4GB_ADDR_SPACE_IN_R0) /* Unfortuantely not reliable with PGMR0DynMap and multiple VMs. */
3230	RTGCPTR GCPtrStrict = (uint32_t)iPD << GST_PD_SHIFT;
3231	# if PGM_GST_TYPE == PGM_TYPE_PAE
3232	GCPtrStrict \|= iPdpt << X86_PDPT_SHIFT;
3233	# endif
3234	AssertMsg(pgmShwGetPaePDEPtr(&pVM->pgm.s, GCPtrStrict) == pPDEDst, ("%p vs %p (%RGv)\n", pgmShwGetPaePDEPtr(&pVM->pgm.s, GCPtrStrict), pPDEDst, GCPtrStrict));
3235	# endif /* VBOX_STRICT */
3236	# endif
3237	GSTPDE PdeSrc = pPDSrc->a[iPD];
3238	if ( PdeSrc.n.u1Present
3239	&& (PdeSrc.n.u1User \|\| fRawR0Enabled))
3240	{
3241	# if ( PGM_GST_TYPE == PGM_TYPE_32BIT \
3242	\|\| PGM_GST_TYPE == PGM_TYPE_PAE) \
3243	&& !defined(PGM_WITHOUT_MAPPINGS)
3244
3245	/*
3246	* Check for conflicts with GC mappings.
3247	*/
3248	# if PGM_GST_TYPE == PGM_TYPE_PAE
3249	if (iPD + iPdpt * X86_PG_PAE_ENTRIES == iPdNoMapping)
3250	# else
3251	if (iPD == iPdNoMapping)
3252	# endif
3253	{
3254	if (pVM->pgm.s.fMappingsFixed)
3255	{
3256	/* It's fixed, just skip the mapping. */
3257	const unsigned cPTs = pMapping->cb >> GST_PD_SHIFT;
3258	Assert(PGM_GST_TYPE == PGM_TYPE_32BIT \|\| (iPD + cPTs - 1) / X86_PG_PAE_ENTRIES == iPD / X86_PG_PAE_ENTRIES);
3259	iPD += cPTs - 1;
3260	# if PGM_SHW_TYPE != PGM_GST_TYPE /* SHW==PAE && GST==32BIT */
3261	pPDEDst = pgmShwGetPaePDEPtr(&pVM->pgm.s, (uint32_t)(iPD + 1) << GST_PD_SHIFT);
3262	# else
3263	pPDEDst += cPTs;
3264	# endif
3265	pMapping = pMapping->CTX_SUFF(pNext);
3266	iPdNoMapping = pMapping ? pMapping->GCPtr >> GST_PD_SHIFT : ~0U;
3267	continue;
3268	}
3269	# ifdef IN_RING3
3270	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3271	int rc = pgmR3SyncPTResolveConflict(pVM, pMapping, pPDSrc, iPD << GST_PD_SHIFT);
3272	# elif PGM_GST_TYPE == PGM_TYPE_PAE
3273	int rc = pgmR3SyncPTResolveConflictPAE(pVM, pMapping, (iPdpt << GST_PDPT_SHIFT) + (iPD << GST_PD_SHIFT));
3274	# endif
3275	if (RT_FAILURE(rc))
3276	return rc;
3277
3278	/*
3279	* Update iPdNoMapping and pMapping.
3280	*/
3281	pMapping = pVM->pgm.s.pMappingsR3;
3282	while (pMapping && pMapping->GCPtr < (iPD << GST_PD_SHIFT))
3283	pMapping = pMapping->pNextR3;
3284	iPdNoMapping = pMapping ? pMapping->GCPtr >> GST_PD_SHIFT : ~0U;
3285	# else /* !IN_RING3 */
3286	LogFlow(("SyncCR3: detected conflict -> VINF_PGM_SYNC_CR3\n"));
3287	return VINF_PGM_SYNC_CR3;
3288	# endif /* !IN_RING3 */
3289	}
3290	# else /* (PGM_GST_TYPE != PGM_TYPE_32BIT && PGM_GST_TYPE != PGM_TYPE_PAE) \|\| PGM_WITHOUT_MAPPINGS */
3291	Assert(!pgmMapAreMappingsEnabled(&pVM->pgm.s));
3292	# endif /* (PGM_GST_TYPE != PGM_TYPE_32BIT && PGM_GST_TYPE != PGM_TYPE_PAE) \|\| PGM_WITHOUT_MAPPINGS */
3293
3294	/*
3295	* Sync page directory entry.
3296	*
3297	* The current approach is to allocated the page table but to set
3298	* the entry to not-present and postpone the page table synching till
3299	* it's actually used.
3300	*/
3301	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
3302	for (unsigned i = 0, iPdShw = iPD * 2; i < 2; i++, iPdShw++) /* pray that the compiler unrolls this */
3303	# elif PGM_GST_TYPE == PGM_TYPE_PAE
3304	const unsigned iPdShw = iPD + iPdpt * X86_PG_PAE_ENTRIES; NOREF(iPdShw);
3305	# else
3306	const unsigned iPdShw = iPD; NOREF(iPdShw);
3307	# endif
3308	{
3309	SHWPDE PdeDst = *pPDEDst;
3310	if (PdeDst.n.u1Present)
3311	{
3312	PPGMPOOLPAGE pShwPage = pgmPoolGetPage(pPool, PdeDst.u & SHW_PDE_PG_MASK);
3313	RTGCPHYS GCPhys;
3314	if ( !PdeSrc.b.u1Size
3315	\|\| !fBigPagesSupported)
3316	{
3317	GCPhys = PdeSrc.u & GST_PDE_PG_MASK;
3318	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
3319	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
3320	GCPhys \|= i * (PAGE_SIZE / 2);
3321	# endif
3322	}
3323	else
3324	{
3325	GCPhys = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc);
3326	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
3327	/* Select the right PDE as we're emulating a 4MB page directory with two 2 MB shadow PDEs.*/
3328	GCPhys \|= i * X86_PAGE_2M_SIZE;
3329	# endif
3330	}
3331
3332	if ( pShwPage->GCPhys == GCPhys
3333	&& pShwPage->enmKind == PGM_BTH_NAME(CalcPageKind)(&PdeSrc, cr4)
3334	&& ( pShwPage->fCached
3335	\|\| ( !fGlobal
3336	&& ( false
3337	# ifdef PGM_SKIP_GLOBAL_PAGEDIRS_ON_NONGLOBAL_FLUSH
3338	\|\| ( (PdeSrc.u & (X86_PDE4M_PS \| X86_PDE4M_G)) == (X86_PDE4M_PS \| X86_PDE4M_G)
3339	&& (cr4 & (X86_CR4_PGE \| X86_CR4_PSE)) == (X86_CR4_PGE \| X86_CR4_PSE)) /* global 2/4MB page. */
3340	\|\| ( !pShwPage->fSeenNonGlobal
3341	&& (cr4 & X86_CR4_PGE))
3342	# endif
3343	)
3344	)
3345	)
3346	&& ( (PdeSrc.u & (X86_PDE_US \| X86_PDE_RW)) == (PdeDst.u & (X86_PDE_US \| X86_PDE_RW))
3347	\|\| ( fBigPagesSupported
3348	&& ((PdeSrc.u & (X86_PDE_US \| X86_PDE4M_PS \| X86_PDE4M_D)) \| PGM_PDFLAGS_TRACK_DIRTY)
3349	== ((PdeDst.u & (X86_PDE_US \| X86_PDE_RW \| PGM_PDFLAGS_TRACK_DIRTY)) \| X86_PDE4M_PS))
3350	)
3351	)
3352	{
3353	# ifdef VBOX_WITH_STATISTICS
3354	if ( !fGlobal
3355	&& (PdeSrc.u & (X86_PDE4M_PS \| X86_PDE4M_G)) == (X86_PDE4M_PS \| X86_PDE4M_G)
3356	&& (cr4 & (X86_CR4_PGE \| X86_CR4_PSE)) == (X86_CR4_PGE \| X86_CR4_PSE))
3357	MY_STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3DstSkippedGlobalPD));
3358	else if (!fGlobal && !pShwPage->fSeenNonGlobal && (cr4 & X86_CR4_PGE))
3359	MY_STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3DstSkippedGlobalPT));
3360	else
3361	MY_STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3DstCacheHit));
3362	# endif /* VBOX_WITH_STATISTICS */
3363	/** @todo a replacement strategy isn't really needed unless we're using a very small pool < 512 pages.
3364	* The whole ageing stuff should be put in yet another set of #ifdefs. For now, let's just skip it. */
3365	//# ifdef PGMPOOL_WITH_CACHE
3366	// pgmPoolCacheUsed(pPool, pShwPage);
3367	//# endif
3368	}
3369	else
3370	{
3371	pgmPoolFreeByPage(pPool, pShwPage, SHW_POOL_ROOT_IDX, iPdShw);
3372	pPDEDst->u = 0;
3373	MY_STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3DstFreed));
3374	}
3375	}
3376	else
3377	MY_STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3DstNotPresent));
3378
3379	/* advance */
3380	pPDEDst++;
3381	} /* foreach 2MB PAE PDE in 4MB guest PDE */
3382	}
3383	# if PGM_GST_TYPE == PGM_TYPE_PAE
3384	else if (iPD + iPdpt * X86_PG_PAE_ENTRIES != iPdNoMapping)
3385	# else
3386	else if (iPD != iPdNoMapping)
3387	# endif
3388	{
3389	/*
3390	* Check if there is any page directory to mark not present here.
3391	*/
3392	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
3393	for (unsigned i = 0, iPdShw = iPD * 2; i < 2; i++, iPdShw++) /* pray that the compiler unrolls this */
3394	# elif PGM_GST_TYPE == PGM_TYPE_PAE
3395	const unsigned iPdShw = iPD + iPdpt * X86_PG_PAE_ENTRIES;
3396	# else
3397	const unsigned iPdShw = iPD;
3398	# endif
3399	{
3400	if (pPDEDst->n.u1Present)
3401	{
3402	pgmPoolFree(pVM, pPDEDst->u & SHW_PDE_PG_MASK, SHW_POOL_ROOT_IDX, iPdShw);
3403	pPDEDst->u = 0;
3404	MY_STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3DstFreedSrcNP));
3405	}
3406	pPDEDst++;
3407	}
3408	}
3409	else
3410	{
3411	# if ( PGM_GST_TYPE == PGM_TYPE_32BIT \
3412	\|\| PGM_GST_TYPE == PGM_TYPE_PAE) \
3413	&& !defined(PGM_WITHOUT_MAPPINGS)
3414
3415	const unsigned cPTs = pMapping->cb >> GST_PD_SHIFT;
3416
3417	Assert(pgmMapAreMappingsEnabled(&pVM->pgm.s));
3418	if (pVM->pgm.s.fMappingsFixed)
3419	{
3420	/* It's fixed, just skip the mapping. */
3421	pMapping = pMapping->CTX_SUFF(pNext);
3422	iPdNoMapping = pMapping ? pMapping->GCPtr >> GST_PD_SHIFT : ~0U;
3423	}
3424	else
3425	{
3426	/*
3427	* Check for conflicts for subsequent pagetables
3428	* and advance to the next mapping.
3429	*/
3430	iPdNoMapping = ~0U;
3431	unsigned iPT = cPTs;
3432	while (iPT-- > 1)
3433	{
3434	if ( pPDSrc->a[iPD + iPT].n.u1Present
3435	&& (pPDSrc->a[iPD + iPT].n.u1User \|\| fRawR0Enabled))
3436	{
3437	# ifdef IN_RING3
3438	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3439	int rc = pgmR3SyncPTResolveConflict(pVM, pMapping, pPDSrc, iPD << GST_PD_SHIFT);
3440	# elif PGM_GST_TYPE == PGM_TYPE_PAE
3441	int rc = pgmR3SyncPTResolveConflictPAE(pVM, pMapping, (iPdpt << GST_PDPT_SHIFT) + (iPD << GST_PD_SHIFT));
3442	# endif
3443	if (RT_FAILURE(rc))
3444	return rc;
3445
3446	/*
3447	* Update iPdNoMapping and pMapping.
3448	*/
3449	pMapping = pVM->pgm.s.CTX_SUFF(pMappings);
3450	while (pMapping && pMapping->GCPtr < (iPD << GST_PD_SHIFT))
3451	pMapping = pMapping->CTX_SUFF(pNext);
3452	iPdNoMapping = pMapping ? pMapping->GCPtr >> GST_PD_SHIFT : ~0U;
3453	break;
3454	# else
3455	LogFlow(("SyncCR3: detected conflict -> VINF_PGM_SYNC_CR3\n"));
3456	return VINF_PGM_SYNC_CR3;
3457	# endif
3458	}
3459	}
3460	if (iPdNoMapping == ~0U && pMapping)
3461	{
3462	pMapping = pMapping->CTX_SUFF(pNext);
3463	if (pMapping)
3464	iPdNoMapping = pMapping->GCPtr >> GST_PD_SHIFT;
3465	}
3466	}
3467
3468	/* advance. */
3469	Assert(PGM_GST_TYPE == PGM_TYPE_32BIT \|\| (iPD + cPTs - 1) / X86_PG_PAE_ENTRIES == iPD / X86_PG_PAE_ENTRIES);
3470	iPD += cPTs - 1;
3471	# if PGM_SHW_TYPE != PGM_GST_TYPE /* SHW==PAE && GST==32BIT */
3472	pPDEDst = pgmShwGetPaePDEPtr(&pVM->pgm.s, (uint32_t)(iPD + 1) << GST_PD_SHIFT);
3473	# else
3474	pPDEDst += cPTs;
3475	# endif
3476	# if PGM_GST_TYPE != PGM_SHW_TYPE
3477	AssertCompile(PGM_GST_TYPE == PGM_TYPE_32BIT && PGM_SHW_TYPE == PGM_TYPE_PAE);
3478	# endif
3479	# else /* (PGM_GST_TYPE != PGM_TYPE_32BIT && PGM_GST_TYPE != PGM_TYPE_PAE) \|\| PGM_WITHOUT_MAPPINGS */
3480	Assert(!pgmMapAreMappingsEnabled(&pVM->pgm.s));
3481	# endif /* (PGM_GST_TYPE != PGM_TYPE_32BIT && PGM_GST_TYPE != PGM_TYPE_PAE) \|\| PGM_WITHOUT_MAPPINGS */
3482	}
3483
3484	} /* for iPD */
3485	} /* for each PDPTE (PAE) */
3486	return VINF_SUCCESS;
3487
3488	# else /* guest real and protected mode */
3489	return VINF_SUCCESS;
3490	# endif
3491	#endif /* PGM_SHW_TYPE != PGM_TYPE_NESTED && PGM_SHW_TYPE != PGM_TYPE_EPT && PGM_SHW_TYPE != PGM_TYPE_AMD64 */
3492	}
3493
3494
3495
3496
3497	#ifdef VBOX_STRICT
3498	#ifdef IN_RC
3499	# undef AssertMsgFailed
3500	# define AssertMsgFailed Log
3501	#endif
3502	#ifdef IN_RING3
3503	# include <VBox/dbgf.h>
3504
3505	/**
3506	* Dumps a page table hierarchy use only physical addresses and cr4/lm flags.
3507	*
3508	* @returns VBox status code (VINF_SUCCESS).
3509	* @param pVM The VM handle.
3510	* @param cr3 The root of the hierarchy.
3511	* @param crr The cr4, only PAE and PSE is currently used.
3512	* @param fLongMode Set if long mode, false if not long mode.
3513	* @param cMaxDepth Number of levels to dump.
3514	* @param pHlp Pointer to the output functions.
3515	*/
3516	__BEGIN_DECLS
3517	VMMR3DECL(int) PGMR3DumpHierarchyHC(PVM pVM, uint32_t cr3, uint32_t cr4, bool fLongMode, unsigned cMaxDepth, PCDBGFINFOHLP pHlp);
3518	__END_DECLS
3519
3520	#endif
3521
3522	/**
3523	* Checks that the shadow page table is in sync with the guest one.
3524	*
3525	* @returns The number of errors.
3526	* @param pVM The virtual machine.
3527	* @param cr3 Guest context CR3 register
3528	* @param cr4 Guest context CR4 register
3529	* @param GCPtr Where to start. Defaults to 0.
3530	* @param cb How much to check. Defaults to everything.
3531	*/
3532	PGM_BTH_DECL(unsigned, AssertCR3)(PVM pVM, uint64_t cr3, uint64_t cr4, RTGCPTR GCPtr, RTGCPTR cb)
3533	{
3534	#if PGM_SHW_TYPE == PGM_TYPE_NESTED \|\| PGM_SHW_TYPE == PGM_TYPE_EPT
3535	return 0;
3536	#else
3537	unsigned cErrors = 0;
3538
3539	#if PGM_GST_TYPE == PGM_TYPE_PAE
3540	/** @todo currently broken; crashes below somewhere */
3541	AssertFailed();
3542	#endif
3543
3544	#if PGM_GST_TYPE == PGM_TYPE_32BIT \
3545	\|\| PGM_GST_TYPE == PGM_TYPE_PAE \
3546	\|\| PGM_GST_TYPE == PGM_TYPE_AMD64
3547
3548	# if PGM_GST_TYPE == PGM_TYPE_AMD64
3549	bool fBigPagesSupported = true;
3550	# else
3551	bool fBigPagesSupported = !!(CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
3552	# endif
3553	PPGM pPGM = &pVM->pgm.s;
3554	RTGCPHYS GCPhysGst; /* page address derived from the guest page tables. */
3555	RTHCPHYS HCPhysShw; /* page address derived from the shadow page tables. */
3556	# ifndef IN_RING0
3557	RTHCPHYS HCPhys; /* general usage. */
3558	# endif
3559	int rc;
3560
3561	/*
3562	* Check that the Guest CR3 and all its mappings are correct.
3563	*/
3564	AssertMsgReturn(pPGM->GCPhysCR3 == (cr3 & GST_CR3_PAGE_MASK),
3565	("Invalid GCPhysCR3=%RGp cr3=%RGp\n", pPGM->GCPhysCR3, (RTGCPHYS)cr3),
3566	false);
3567	# if !defined(IN_RING0) && PGM_GST_TYPE != PGM_TYPE_AMD64
3568	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3569	rc = PGMShwGetPage(pVM, (RTGCPTR)pPGM->pGst32BitPdRC, NULL, &HCPhysShw);
3570	# else
3571	rc = PGMShwGetPage(pVM, (RTGCPTR)pPGM->pGstPaePdptRC, NULL, &HCPhysShw);
3572	# endif
3573	AssertRCReturn(rc, 1);
3574	HCPhys = NIL_RTHCPHYS;
3575	rc = pgmRamGCPhys2HCPhys(pPGM, cr3 & GST_CR3_PAGE_MASK, &HCPhys);
3576	AssertMsgReturn(HCPhys == HCPhysShw, ("HCPhys=%RHp HCPhyswShw=%RHp (cr3)\n", HCPhys, HCPhysShw), false);
3577	# if PGM_GST_TYPE == PGM_TYPE_32BIT && defined(IN_RING3)
3578	RTGCPHYS GCPhys;
3579	rc = PGMR3DbgR3Ptr2GCPhys(pVM, pPGM->pGst32BitPdR3, &GCPhys);
3580	AssertRCReturn(rc, 1);
3581	AssertMsgReturn((cr3 & GST_CR3_PAGE_MASK) == GCPhys, ("GCPhys=%RGp cr3=%RGp\n", GCPhys, (RTGCPHYS)cr3), false);
3582	# endif
3583	# endif /* !IN_RING0 */
3584
3585	/*
3586	* Get and check the Shadow CR3.
3587	*/
3588	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
3589	unsigned cPDEs = X86_PG_ENTRIES;
3590	unsigned cIncrement = X86_PG_ENTRIES * PAGE_SIZE;
3591	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
3592	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3593	unsigned cPDEs = X86_PG_PAE_ENTRIES * 4; /* treat it as a 2048 entry table. */
3594	# else
3595	unsigned cPDEs = X86_PG_PAE_ENTRIES;
3596	# endif
3597	unsigned cIncrement = X86_PG_PAE_ENTRIES * PAGE_SIZE;
3598	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
3599	unsigned cPDEs = X86_PG_PAE_ENTRIES;
3600	unsigned cIncrement = X86_PG_PAE_ENTRIES * PAGE_SIZE;
3601	# endif
3602	if (cb != ~(RTGCPTR)0)
3603	cPDEs = RT_MIN(cb >> SHW_PD_SHIFT, 1);
3604
3605	/** @todo call the other two PGMAssert() functions. /
3606
3607	# if PGM_GST_TYPE == PGM_TYPE_AMD64 \|\| PGM_GST_TYPE == PGM_TYPE_PAE
3608	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
3609	# endif
3610
3611	# if PGM_GST_TYPE == PGM_TYPE_AMD64
3612	unsigned iPml4 = (GCPtr >> X86_PML4_SHIFT) & X86_PML4_MASK;
3613
3614	for (; iPml4 < X86_PG_PAE_ENTRIES; iPml4++)
3615	{
3616	PPGMPOOLPAGE pShwPdpt = NULL;
3617	PX86PML4E pPml4eSrc;
3618	PX86PML4E pPml4eDst;
3619	RTGCPHYS GCPhysPdptSrc;
3620
3621	pPml4eSrc = pgmGstGetLongModePML4EPtr(&pVM->pgm.s, iPml4);
3622	pPml4eDst = pgmShwGetLongModePML4EPtr(&pVM->pgm.s, iPml4);
3623
3624	/* Fetch the pgm pool shadow descriptor if the shadow pml4e is present. */
3625	if (!pPml4eDst->n.u1Present)
3626	{
3627	GCPtr += _2M * UINT64_C(512) * UINT64_C(512);
3628	continue;
3629	}
3630
3631	pShwPdpt = pgmPoolGetPage(pPool, pPml4eDst->u & X86_PML4E_PG_MASK);
3632	GCPhysPdptSrc = pPml4eSrc->u & X86_PML4E_PG_MASK_FULL;
3633
3634	if (pPml4eSrc->n.u1Present != pPml4eDst->n.u1Present)
3635	{
3636	AssertMsgFailed(("Present bit doesn't match! pPml4eDst.u=%#RX64 pPml4eSrc.u=%RX64\n", pPml4eDst->u, pPml4eSrc->u));
3637	GCPtr += _2M * UINT64_C(512) * UINT64_C(512);
3638	cErrors++;
3639	continue;
3640	}
3641
3642	if (GCPhysPdptSrc != pShwPdpt->GCPhys)
3643	{
3644	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));
3645	GCPtr += _2M * UINT64_C(512) * UINT64_C(512);
3646	cErrors++;
3647	continue;
3648	}
3649
3650	if ( pPml4eDst->n.u1User != pPml4eSrc->n.u1User
3651	\|\| pPml4eDst->n.u1Write != pPml4eSrc->n.u1Write
3652	\|\| pPml4eDst->n.u1NoExecute != pPml4eSrc->n.u1NoExecute)
3653	{
3654	AssertMsgFailed(("User/Write/NoExec bits don't match! pPml4eDst.u=%#RX64 pPml4eSrc.u=%RX64\n", pPml4eDst->u, pPml4eSrc->u));
3655	GCPtr += _2M * UINT64_C(512) * UINT64_C(512);
3656	cErrors++;
3657	continue;
3658	}
3659	# else /* PGM_GST_TYPE != PGM_TYPE_AMD64 */
3660	{
3661	# endif /* PGM_GST_TYPE != PGM_TYPE_AMD64 */
3662
3663	# if PGM_GST_TYPE == PGM_TYPE_AMD64 \|\| PGM_GST_TYPE == PGM_TYPE_PAE
3664	/*
3665	* Check the PDPTEs too.
3666	*/
3667	unsigned iPdpt = (GCPtr >> SHW_PDPT_SHIFT) & SHW_PDPT_MASK;
3668
3669	for (;iPdpt <= SHW_PDPT_MASK; iPdpt++)
3670	{
3671	unsigned iPDSrc;
3672	PPGMPOOLPAGE pShwPde = NULL;
3673	PX86PDPE pPdpeDst;
3674	RTGCPHYS GCPhysPdeSrc;
3675	# if PGM_GST_TYPE == PGM_TYPE_PAE
3676	X86PDPE PdpeSrc;
3677	PGSTPD pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, GCPtr, &iPDSrc, &PdpeSrc);
3678	PX86PDPT pPdptDst = pgmShwGetPaePDPTPtr(&pVM->pgm.s);
3679	# else
3680	PX86PML4E pPml4eSrc;
3681	X86PDPE PdpeSrc;
3682	PX86PDPT pPdptDst;
3683	PX86PDPAE pPDDst;
3684	PGSTPD pPDSrc = pgmGstGetLongModePDPtr(&pVM->pgm.s, GCPtr, &pPml4eSrc, &PdpeSrc, &iPDSrc);
3685
3686	rc = pgmShwGetLongModePDPtr(pVM, GCPtr, NULL, &pPdptDst, &pPDDst);
3687	if (rc != VINF_SUCCESS)
3688	{
3689	AssertMsg(rc == VERR_PAGE_DIRECTORY_PTR_NOT_PRESENT, ("Unexpected rc=%Rrc\n", rc));
3690	GCPtr += 512 * _2M;
3691	continue; /* next PDPTE */
3692	}
3693	Assert(pPDDst);
3694	# endif
3695	Assert(iPDSrc == 0);
3696
3697	pPdpeDst = &pPdptDst->a[iPdpt];
3698
3699	if (!pPdpeDst->n.u1Present)
3700	{
3701	GCPtr += 512 * _2M;
3702	continue; /* next PDPTE */
3703	}
3704
3705	pShwPde = pgmPoolGetPage(pPool, pPdpeDst->u & X86_PDPE_PG_MASK);
3706	GCPhysPdeSrc = PdpeSrc.u & X86_PDPE_PG_MASK;
3707
3708	if (pPdpeDst->n.u1Present != PdpeSrc.n.u1Present)
3709	{
3710	AssertMsgFailed(("Present bit doesn't match! pPdpeDst.u=%#RX64 pPdpeSrc.u=%RX64\n", pPdpeDst->u, PdpeSrc.u));
3711	GCPtr += 512 * _2M;
3712	cErrors++;
3713	continue;
3714	}
3715
3716	if (GCPhysPdeSrc != pShwPde->GCPhys)
3717	{
3718	# if PGM_GST_TYPE == PGM_TYPE_AMD64
3719	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));
3720	# else
3721	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));
3722	# endif
3723	GCPtr += 512 * _2M;
3724	cErrors++;
3725	continue;
3726	}
3727
3728	# if PGM_GST_TYPE == PGM_TYPE_AMD64
3729	if ( pPdpeDst->lm.u1User != PdpeSrc.lm.u1User
3730	\|\| pPdpeDst->lm.u1Write != PdpeSrc.lm.u1Write
3731	\|\| pPdpeDst->lm.u1NoExecute != PdpeSrc.lm.u1NoExecute)
3732	{
3733	AssertMsgFailed(("User/Write/NoExec bits don't match! pPdpeDst.u=%#RX64 pPdpeSrc.u=%RX64\n", pPdpeDst->u, PdpeSrc.u));
3734	GCPtr += 512 * _2M;
3735	cErrors++;
3736	continue;
3737	}
3738	# endif
3739
3740	# else /* PGM_GST_TYPE != PGM_TYPE_AMD64 && PGM_GST_TYPE != PGM_TYPE_PAE */
3741	{
3742	# endif /* PGM_GST_TYPE != PGM_TYPE_AMD64 && PGM_GST_TYPE != PGM_TYPE_PAE */
3743	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3744	GSTPD const *pPDSrc = pgmGstGet32bitPDPtr(&pVM->pgm.s);
3745	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
3746	PCX86PD pPDDst = pgmShwGet32BitPDPtr(&pVM->pgm.s);
3747	# endif
3748	# endif /* PGM_GST_TYPE == PGM_TYPE_32BIT */
3749	/*
3750	* Iterate the shadow page directory.
3751	*/
3752	GCPtr = (GCPtr >> SHW_PD_SHIFT) << SHW_PD_SHIFT;
3753	unsigned iPDDst = (GCPtr >> SHW_PD_SHIFT) & SHW_PD_MASK;
3754
3755	for (;
3756	iPDDst < cPDEs;
3757	iPDDst++, GCPtr += cIncrement)
3758	{
3759	# if PGM_SHW_TYPE == PGM_TYPE_PAE
3760	const SHWPDE PdeDst = *pgmShwGetPaePDEPtr(pPGM, GCPtr);
3761	# else
3762	const SHWPDE PdeDst = pPDDst->a[iPDDst];
3763	# endif
3764	if (PdeDst.u & PGM_PDFLAGS_MAPPING)
3765	{
3766	Assert(pgmMapAreMappingsEnabled(&pVM->pgm.s));
3767	if ((PdeDst.u & X86_PDE_AVL_MASK) != PGM_PDFLAGS_MAPPING)
3768	{
3769	AssertMsgFailed(("Mapping shall only have PGM_PDFLAGS_MAPPING set! PdeDst.u=%#RX64\n", (uint64_t)PdeDst.u));
3770	cErrors++;
3771	continue;
3772	}
3773	}
3774	else if ( (PdeDst.u & X86_PDE_P)
3775	\|\| ((PdeDst.u & (X86_PDE_P \| PGM_PDFLAGS_TRACK_DIRTY)) == (X86_PDE_P \| PGM_PDFLAGS_TRACK_DIRTY))
3776	)
3777	{
3778	HCPhysShw = PdeDst.u & SHW_PDE_PG_MASK;
3779	PPGMPOOLPAGE pPoolPage = pgmPoolGetPageByHCPhys(pVM, HCPhysShw);
3780	if (!pPoolPage)
3781	{
3782	AssertMsgFailed(("Invalid page table address %RHp at %RGv! PdeDst=%#RX64\n",
3783	HCPhysShw, GCPtr, (uint64_t)PdeDst.u));
3784	cErrors++;
3785	continue;
3786	}
3787	const SHWPT pPTDst = (const SHWPT )PGMPOOL_PAGE_2_PTR(pVM, pPoolPage);
3788
3789	if (PdeDst.u & (X86_PDE4M_PWT \| X86_PDE4M_PCD))
3790	{
3791	AssertMsgFailed(("PDE flags PWT and/or PCD is set at %RGv! These flags are not virtualized! PdeDst=%#RX64\n",
3792	GCPtr, (uint64_t)PdeDst.u));
3793	cErrors++;
3794	}
3795
3796	if (PdeDst.u & (X86_PDE4M_G \| X86_PDE4M_D))
3797	{
3798	AssertMsgFailed(("4K PDE reserved flags at %RGv! PdeDst=%#RX64\n",
3799	GCPtr, (uint64_t)PdeDst.u));
3800	cErrors++;
3801	}
3802
3803	const GSTPDE PdeSrc = pPDSrc->a[(iPDDst >> (GST_PD_SHIFT - SHW_PD_SHIFT)) & GST_PD_MASK];
3804	if (!PdeSrc.n.u1Present)
3805	{
3806	AssertMsgFailed(("Guest PDE at %RGv is not present! PdeDst=%#RX64 PdeSrc=%#RX64\n",
3807	GCPtr, (uint64_t)PdeDst.u, (uint64_t)PdeSrc.u));
3808	cErrors++;
3809	continue;
3810	}
3811
3812	if ( !PdeSrc.b.u1Size
3813	\|\| !fBigPagesSupported)
3814	{
3815	GCPhysGst = PdeSrc.u & GST_PDE_PG_MASK;
3816	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
3817	GCPhysGst \|= (iPDDst & 1) * (PAGE_SIZE / 2);
3818	# endif
3819	}
3820	else
3821	{
3822	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3823	if (PdeSrc.u & X86_PDE4M_PG_HIGH_MASK)
3824	{
3825	AssertMsgFailed(("Guest PDE at %RGv is using PSE36 or similar! PdeSrc=%#RX64\n",
3826	GCPtr, (uint64_t)PdeSrc.u));
3827	cErrors++;
3828	continue;
3829	}
3830	# endif
3831	GCPhysGst = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc);
3832	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
3833	GCPhysGst \|= GCPtr & RT_BIT(X86_PAGE_2M_SHIFT);
3834	# endif
3835	}
3836
3837	if ( pPoolPage->enmKind
3838	!= (!PdeSrc.b.u1Size \|\| !fBigPagesSupported ? BTH_PGMPOOLKIND_PT_FOR_PT : BTH_PGMPOOLKIND_PT_FOR_BIG))
3839	{
3840	AssertMsgFailed(("Invalid shadow page table kind %d at %RGv! PdeSrc=%#RX64\n",
3841	pPoolPage->enmKind, GCPtr, (uint64_t)PdeSrc.u));
3842	cErrors++;
3843	}
3844
3845	PPGMPAGE pPhysPage = pgmPhysGetPage(pPGM, GCPhysGst);
3846	if (!pPhysPage)
3847	{
3848	AssertMsgFailed(("Cannot find guest physical address %RGp in the PDE at %RGv! PdeSrc=%#RX64\n",
3849	GCPhysGst, GCPtr, (uint64_t)PdeSrc.u));
3850	cErrors++;
3851	continue;
3852	}
3853
3854	if (GCPhysGst != pPoolPage->GCPhys)
3855	{
3856	AssertMsgFailed(("GCPhysGst=%RGp != pPage->GCPhys=%RGp at %RGv\n",
3857	GCPhysGst, pPoolPage->GCPhys, GCPtr));
3858	cErrors++;
3859	continue;
3860	}
3861
3862	if ( !PdeSrc.b.u1Size
3863	\|\| !fBigPagesSupported)
3864	{
3865	/*
3866	* Page Table.
3867	*/
3868	const GSTPT *pPTSrc;
3869	rc = PGM_GCPHYS_2_PTR(pVM, GCPhysGst & ~(RTGCPHYS)(PAGE_SIZE - 1), &pPTSrc);
3870	if (RT_FAILURE(rc))
3871	{
3872	AssertMsgFailed(("Cannot map/convert guest physical address %RGp in the PDE at %RGv! PdeSrc=%#RX64\n",
3873	GCPhysGst, GCPtr, (uint64_t)PdeSrc.u));
3874	cErrors++;
3875	continue;
3876	}
3877	if ( (PdeSrc.u & (X86_PDE_P \| X86_PDE_US \| X86_PDE_RW/* \| X86_PDE_A*/))
3878	!= (PdeDst.u & (X86_PDE_P \| X86_PDE_US \| X86_PDE_RW/* \| X86_PDE_A*/)))
3879	{
3880	/// @todo We get here a lot on out-of-sync CR3 entries. The access handler should zap them to avoid false alarms here!
3881	// (This problem will go away when/if we shadow multiple CR3s.)
3882	AssertMsgFailed(("4K PDE flags mismatch at %RGv! PdeSrc=%#RX64 PdeDst=%#RX64\n",
3883	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
3884	cErrors++;
3885	continue;
3886	}
3887	if (PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY)
3888	{
3889	AssertMsgFailed(("4K PDEs cannot have PGM_PDFLAGS_TRACK_DIRTY set! GCPtr=%RGv PdeDst=%#RX64\n",
3890	GCPtr, (uint64_t)PdeDst.u));
3891	cErrors++;
3892	continue;
3893	}
3894
3895	/* iterate the page table. */
3896	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
3897	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
3898	const unsigned offPTSrc = ((GCPtr >> SHW_PD_SHIFT) & 1) * 512;
3899	# else
3900	const unsigned offPTSrc = 0;
3901	# endif
3902	for (unsigned iPT = 0, off = 0;
3903	iPT < RT_ELEMENTS(pPTDst->a);
3904	iPT++, off += PAGE_SIZE)
3905	{
3906	const SHWPTE PteDst = pPTDst->a[iPT];
3907
3908	/* skip not-present entries. */
3909	if (!(PteDst.u & (X86_PTE_P \| PGM_PTFLAGS_TRACK_DIRTY))) /** @todo deal with ALL handlers and CSAM !P pages! */
3910	continue;
3911	Assert(PteDst.n.u1Present);
3912
3913	const GSTPTE PteSrc = pPTSrc->a[iPT + offPTSrc];
3914	if (!PteSrc.n.u1Present)
3915	{
3916	# ifdef IN_RING3
3917	PGMAssertHandlerAndFlagsInSync(pVM);
3918	PGMR3DumpHierarchyGC(pVM, cr3, cr4, (PdeSrc.u & GST_PDE_PG_MASK));
3919	# endif
3920	AssertMsgFailed(("Out of sync (!P) PTE at %RGv! PteSrc=%#RX64 PteDst=%#RX64 pPTSrc=%RGv iPTSrc=%x PdeSrc=%x physpte=%RGp\n",
3921	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u, pPTSrc, iPT + offPTSrc, PdeSrc.au32[0],
3922	(PdeSrc.u & GST_PDE_PG_MASK) + (iPT + offPTSrc)*sizeof(PteSrc)));
3923	cErrors++;
3924	continue;
3925	}
3926
3927	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;
3928	# if 1 /** @todo sync accessed bit properly... */
3929	fIgnoreFlags \|= X86_PTE_A;
3930	# endif
3931
3932	/* match the physical addresses */
3933	HCPhysShw = PteDst.u & SHW_PTE_PG_MASK;
3934	GCPhysGst = PteSrc.u & GST_PTE_PG_MASK;
3935
3936	# ifdef IN_RING3
3937	rc = PGMPhysGCPhys2HCPhys(pVM, GCPhysGst, &HCPhys);
3938	if (RT_FAILURE(rc))
3939	{
3940	if (HCPhysShw != MMR3PageDummyHCPhys(pVM))
3941	{
3942	AssertMsgFailed(("Cannot find guest physical address %RGp at %RGv! PteSrc=%#RX64 PteDst=%#RX64\n",
3943	GCPhysGst, GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3944	cErrors++;
3945	continue;
3946	}
3947	}
3948	else if (HCPhysShw != (HCPhys & SHW_PTE_PG_MASK))
3949	{
3950	AssertMsgFailed(("Out of sync (phys) at %RGv! HCPhysShw=%RHp HCPhys=%RHp GCPhysGst=%RGp PteSrc=%#RX64 PteDst=%#RX64\n",
3951	GCPtr + off, HCPhysShw, HCPhys, GCPhysGst, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3952	cErrors++;
3953	continue;
3954	}
3955	# endif
3956
3957	pPhysPage = pgmPhysGetPage(pPGM, GCPhysGst);
3958	if (!pPhysPage)
3959	{
3960	# ifdef IN_RING3 /** @todo make MMR3PageDummyHCPhys an 'All' function! */
3961	if (HCPhysShw != MMR3PageDummyHCPhys(pVM))
3962	{
3963	AssertMsgFailed(("Cannot find guest physical address %RGp at %RGv! PteSrc=%#RX64 PteDst=%#RX64\n",
3964	GCPhysGst, GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3965	cErrors++;
3966	continue;
3967	}
3968	# endif
3969	if (PteDst.n.u1Write)
3970	{
3971	AssertMsgFailed(("Invalid guest page at %RGv is writable! GCPhysGst=%RGp PteSrc=%#RX64 PteDst=%#RX64\n",
3972	GCPtr + off, GCPhysGst, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3973	cErrors++;
3974	}
3975	fIgnoreFlags \|= X86_PTE_RW;
3976	}
3977	else if (HCPhysShw != (PGM_PAGE_GET_HCPHYS(pPhysPage) & SHW_PTE_PG_MASK))
3978	{
3979	AssertMsgFailed(("Out of sync (phys) at %RGv! HCPhysShw=%RHp HCPhys=%RHp GCPhysGst=%RGp PteSrc=%#RX64 PteDst=%#RX64\n",
3980	GCPtr + off, HCPhysShw, pPhysPage->HCPhys, GCPhysGst, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3981	cErrors++;
3982	continue;
3983	}
3984
3985	/* flags */
3986	if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPhysPage))
3987	{
3988	if (!PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPhysPage))
3989	{
3990	if (PteDst.n.u1Write)
3991	{
3992	AssertMsgFailed(("WRITE access flagged at %RGv but the page is writable! HCPhys=%RHp PteSrc=%#RX64 PteDst=%#RX64\n",
3993	GCPtr + off, pPhysPage->HCPhys, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3994	cErrors++;
3995	continue;
3996	}
3997	fIgnoreFlags \|= X86_PTE_RW;
3998	}
3999	else
4000	{
4001	if (PteDst.n.u1Present)
4002	{
4003	AssertMsgFailed(("ALL access flagged at %RGv but the page is present! HCPhys=%RHp PteSrc=%#RX64 PteDst=%#RX64\n",
4004	GCPtr + off, pPhysPage->HCPhys, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4005	cErrors++;
4006	continue;
4007	}
4008	fIgnoreFlags \|= X86_PTE_P;
4009	}
4010	}
4011	else
4012	{
4013	if (!PteSrc.n.u1Dirty && PteSrc.n.u1Write)
4014	{
4015	if (PteDst.n.u1Write)
4016	{
4017	AssertMsgFailed(("!DIRTY page at %RGv is writable! PteSrc=%#RX64 PteDst=%#RX64\n",
4018	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4019	cErrors++;
4020	continue;
4021	}
4022	if (!(PteDst.u & PGM_PTFLAGS_TRACK_DIRTY))
4023	{
4024	AssertMsgFailed(("!DIRTY page at %RGv is not marked TRACK_DIRTY! PteSrc=%#RX64 PteDst=%#RX64\n",
4025	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4026	cErrors++;
4027	continue;
4028	}
4029	if (PteDst.n.u1Dirty)
4030	{
4031	AssertMsgFailed(("!DIRTY page at %RGv is marked DIRTY! PteSrc=%#RX64 PteDst=%#RX64\n",
4032	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4033	cErrors++;
4034	}
4035	# if 0 /** @todo sync access bit properly... */
4036	if (PteDst.n.u1Accessed != PteSrc.n.u1Accessed)
4037	{
4038	AssertMsgFailed(("!DIRTY page at %RGv is has mismatching accessed bit! PteSrc=%#RX64 PteDst=%#RX64\n",
4039	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4040	cErrors++;
4041	}
4042	fIgnoreFlags \|= X86_PTE_RW;
4043	# else
4044	fIgnoreFlags \|= X86_PTE_RW \| X86_PTE_A;
4045	# endif
4046	}
4047	else if (PteDst.u & PGM_PTFLAGS_TRACK_DIRTY)
4048	{
4049	/* access bit emulation (not implemented). */
4050	if (PteSrc.n.u1Accessed \|\| PteDst.n.u1Present)
4051	{
4052	AssertMsgFailed(("PGM_PTFLAGS_TRACK_DIRTY set at %RGv but no accessed bit emulation! PteSrc=%#RX64 PteDst=%#RX64\n",
4053	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4054	cErrors++;
4055	continue;
4056	}
4057	if (!PteDst.n.u1Accessed)
4058	{
4059	AssertMsgFailed(("!ACCESSED page at %RGv is has the accessed bit set! PteSrc=%#RX64 PteDst=%#RX64\n",
4060	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4061	cErrors++;
4062	}
4063	fIgnoreFlags \|= X86_PTE_P;
4064	}
4065	# ifdef DEBUG_sandervl
4066	fIgnoreFlags \|= X86_PTE_D \| X86_PTE_A;
4067	# endif
4068	}
4069
4070	if ( (PteSrc.u & ~fIgnoreFlags) != (PteDst.u & ~fIgnoreFlags)
4071	&& (PteSrc.u & ~(fIgnoreFlags \| X86_PTE_RW)) != (PteDst.u & ~fIgnoreFlags)
4072	)
4073	{
4074	AssertMsgFailed(("Flags mismatch at %RGv! %#RX64 != %#RX64 fIgnoreFlags=%#RX64 PteSrc=%#RX64 PteDst=%#RX64\n",
4075	GCPtr + off, (uint64_t)PteSrc.u & ~fIgnoreFlags, (uint64_t)PteDst.u & ~fIgnoreFlags,
4076	fIgnoreFlags, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4077	cErrors++;
4078	continue;
4079	}
4080	} /* foreach PTE */
4081	}
4082	else
4083	{
4084	/*
4085	* Big Page.
4086	*/
4087	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;
4088	if (!PdeSrc.b.u1Dirty && PdeSrc.b.u1Write)
4089	{
4090	if (PdeDst.n.u1Write)
4091	{
4092	AssertMsgFailed(("!DIRTY page at %RGv is writable! PdeSrc=%#RX64 PdeDst=%#RX64\n",
4093	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
4094	cErrors++;
4095	continue;
4096	}
4097	if (!(PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY))
4098	{
4099	AssertMsgFailed(("!DIRTY page at %RGv is not marked TRACK_DIRTY! PteSrc=%#RX64 PteDst=%#RX64\n",
4100	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
4101	cErrors++;
4102	continue;
4103	}
4104	# if 0 /** @todo sync access bit properly... */
4105	if (PdeDst.n.u1Accessed != PdeSrc.b.u1Accessed)
4106	{
4107	AssertMsgFailed(("!DIRTY page at %RGv is has mismatching accessed bit! PteSrc=%#RX64 PteDst=%#RX64\n",
4108	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
4109	cErrors++;
4110	}
4111	fIgnoreFlags \|= X86_PTE_RW;
4112	# else
4113	fIgnoreFlags \|= X86_PTE_RW \| X86_PTE_A;
4114	# endif
4115	}
4116	else if (PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY)
4117	{
4118	/* access bit emulation (not implemented). */
4119	if (PdeSrc.b.u1Accessed \|\| PdeDst.n.u1Present)
4120	{
4121	AssertMsgFailed(("PGM_PDFLAGS_TRACK_DIRTY set at %RGv but no accessed bit emulation! PdeSrc=%#RX64 PdeDst=%#RX64\n",
4122	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
4123	cErrors++;
4124	continue;
4125	}
4126	if (!PdeDst.n.u1Accessed)
4127	{
4128	AssertMsgFailed(("!ACCESSED page at %RGv is has the accessed bit set! PdeSrc=%#RX64 PdeDst=%#RX64\n",
4129	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
4130	cErrors++;
4131	}
4132	fIgnoreFlags \|= X86_PTE_P;
4133	}
4134
4135	if ((PdeSrc.u & ~fIgnoreFlags) != (PdeDst.u & ~fIgnoreFlags))
4136	{
4137	AssertMsgFailed(("Flags mismatch (B) at %RGv! %#RX64 != %#RX64 fIgnoreFlags=%#RX64 PdeSrc=%#RX64 PdeDst=%#RX64\n",
4138	GCPtr, (uint64_t)PdeSrc.u & ~fIgnoreFlags, (uint64_t)PdeDst.u & ~fIgnoreFlags,
4139	fIgnoreFlags, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
4140	cErrors++;
4141	}
4142
4143	/* iterate the page table. */
4144	for (unsigned iPT = 0, off = 0;
4145	iPT < RT_ELEMENTS(pPTDst->a);
4146	iPT++, off += PAGE_SIZE, GCPhysGst += PAGE_SIZE)
4147	{
4148	const SHWPTE PteDst = pPTDst->a[iPT];
4149
4150	if (PteDst.u & PGM_PTFLAGS_TRACK_DIRTY)
4151	{
4152	AssertMsgFailed(("The PTE at %RGv emulating a 2/4M page is marked TRACK_DIRTY! PdeSrc=%#RX64 PteDst=%#RX64\n",
4153	GCPtr + off, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4154	cErrors++;
4155	}
4156
4157	/* skip not-present entries. */
4158	if (!PteDst.n.u1Present) /** @todo deal with ALL handlers and CSAM !P pages! */
4159	continue;
4160
4161	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;
4162
4163	/* match the physical addresses */
4164	HCPhysShw = PteDst.u & X86_PTE_PAE_PG_MASK;
4165
4166	# ifdef IN_RING3
4167	rc = PGMPhysGCPhys2HCPhys(pVM, GCPhysGst, &HCPhys);
4168	if (RT_FAILURE(rc))
4169	{
4170	if (HCPhysShw != MMR3PageDummyHCPhys(pVM))
4171	{
4172	AssertMsgFailed(("Cannot find guest physical address %RGp at %RGv! PdeSrc=%#RX64 PteDst=%#RX64\n",
4173	GCPhysGst, GCPtr + off, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4174	cErrors++;
4175	}
4176	}
4177	else if (HCPhysShw != (HCPhys & X86_PTE_PAE_PG_MASK))
4178	{
4179	AssertMsgFailed(("Out of sync (phys) at %RGv! HCPhysShw=%RHp HCPhys=%RHp GCPhysGst=%RGp PdeSrc=%#RX64 PteDst=%#RX64\n",
4180	GCPtr + off, HCPhysShw, HCPhys, GCPhysGst, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4181	cErrors++;
4182	continue;
4183	}
4184	# endif
4185	pPhysPage = pgmPhysGetPage(pPGM, GCPhysGst);
4186	if (!pPhysPage)
4187	{
4188	# ifdef IN_RING3 /** @todo make MMR3PageDummyHCPhys an 'All' function! */
4189	if (HCPhysShw != MMR3PageDummyHCPhys(pVM))
4190	{
4191	AssertMsgFailed(("Cannot find guest physical address %RGp at %RGv! PdeSrc=%#RX64 PteDst=%#RX64\n",
4192	GCPhysGst, GCPtr + off, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4193	cErrors++;
4194	continue;
4195	}
4196	# endif
4197	if (PteDst.n.u1Write)
4198	{
4199	AssertMsgFailed(("Invalid guest page at %RGv is writable! GCPhysGst=%RGp PdeSrc=%#RX64 PteDst=%#RX64\n",
4200	GCPtr + off, GCPhysGst, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4201	cErrors++;
4202	}
4203	fIgnoreFlags \|= X86_PTE_RW;
4204	}
4205	else if (HCPhysShw != (pPhysPage->HCPhys & X86_PTE_PAE_PG_MASK))
4206	{
4207	AssertMsgFailed(("Out of sync (phys) at %RGv! HCPhysShw=%RHp HCPhys=%RHp GCPhysGst=%RGp PdeSrc=%#RX64 PteDst=%#RX64\n",
4208	GCPtr + off, HCPhysShw, pPhysPage->HCPhys, GCPhysGst, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4209	cErrors++;
4210	continue;
4211	}
4212
4213	/* flags */
4214	if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPhysPage))
4215	{
4216	if (!PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPhysPage))
4217	{
4218	if (PGM_PAGE_GET_HNDL_PHYS_STATE(pPhysPage) != PGM_PAGE_HNDL_PHYS_STATE_DISABLED)
4219	{
4220	if (PteDst.n.u1Write)
4221	{
4222	AssertMsgFailed(("WRITE access flagged at %RGv but the page is writable! HCPhys=%RHp PdeSrc=%#RX64 PteDst=%#RX64\n",
4223	GCPtr + off, pPhysPage->HCPhys, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4224	cErrors++;
4225	continue;
4226	}
4227	fIgnoreFlags \|= X86_PTE_RW;
4228	}
4229	}
4230	else
4231	{
4232	if (PteDst.n.u1Present)
4233	{
4234	AssertMsgFailed(("ALL access flagged at %RGv but the page is present! HCPhys=%RHp PdeSrc=%#RX64 PteDst=%#RX64\n",
4235	GCPtr + off, pPhysPage->HCPhys, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4236	cErrors++;
4237	continue;
4238	}
4239	fIgnoreFlags \|= X86_PTE_P;
4240	}
4241	}
4242
4243	if ( (PdeSrc.u & ~fIgnoreFlags) != (PteDst.u & ~fIgnoreFlags)
4244	&& (PdeSrc.u & ~(fIgnoreFlags \| X86_PTE_RW)) != (PteDst.u & ~fIgnoreFlags) /* lazy phys handler dereg. */
4245	)
4246	{
4247	AssertMsgFailed(("Flags mismatch (BT) at %RGv! %#RX64 != %#RX64 fIgnoreFlags=%#RX64 PdeSrc=%#RX64 PteDst=%#RX64\n",
4248	GCPtr + off, (uint64_t)PdeSrc.u & ~fIgnoreFlags, (uint64_t)PteDst.u & ~fIgnoreFlags,
4249	fIgnoreFlags, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4250	cErrors++;
4251	continue;
4252	}
4253	} /* for each PTE */
4254	}
4255	}
4256	/* not present */
4257
4258	} /* for each PDE */
4259
4260	} /* for each PDPTE */
4261
4262	} /* for each PML4E */
4263
4264	# ifdef DEBUG
4265	if (cErrors)
4266	LogFlow(("AssertCR3: cErrors=%d\n", cErrors));
4267	# endif
4268
4269	#endif /* GST == 32BIT, PAE or AMD64 */
4270	return cErrors;
4271
4272	#endif /* PGM_SHW_TYPE != PGM_TYPE_NESTED && PGM_SHW_TYPE != PGM_TYPE_EPT */
4273	}
4274	#endif /* VBOX_STRICT */
4275

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

source: vbox/trunk/src/VBox/VMM/VMMAll/PGMAllBth.h@ 15410

Download in other formats: