PGMAllBth.h@ 13919

Last change on this file since 13919 was 13919, checked in by vboxsync, 16 years ago
VMM: 3 R3R0PTRTYPE members down (a bunch left to go).
Property svn:eol-style set to `native` Property svn:keywords set to `Id`
File size: 193.8 KB

Line
1	/* $Id: PGMAllBth.h 13919 2008-11-06 14:11:41Z 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, RTGCUINTPTR GCPtrPage);
30	PGM_BTH_DECL(int, SyncPage)(PVM pVM, GSTPDE PdeSrc, RTGCUINTPTR GCPtrPage, unsigned cPages, unsigned uErr);
31	PGM_BTH_DECL(int, CheckPageFault)(PVM pVM, uint32_t uErr, PSHWPDE pPdeDst, PGSTPDE pPdeSrc, RTGCUINTPTR GCPtrPage);
32	PGM_BTH_DECL(int, SyncPT)(PVM pVM, unsigned iPD, PGSTPD pPDSrc, RTGCUINTPTR GCPtrPage);
33	PGM_BTH_DECL(int, VerifyAccessSyncPage)(PVM pVM, RTGCUINTPTR Addr, unsigned fPage, unsigned uErr);
34	PGM_BTH_DECL(int, PrefetchPage)(PVM pVM, RTGCUINTPTR 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, RTGCUINTPTR GCPtr = 0, RTGCUINTPTR cb = ~(RTGCUINTPTR)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 = (RTGCUINTPTR)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, (RTGCUINTPTR)pvFault, &iPDSrc);
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)(((RTGCUINTPTR)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 = (RTGCUINTPTR)pvFault >> SHW_PD_SHIFT;
144	PX86PD pPDDst = pVM->pgm.s.CTXMID(p,32BitPD);
145
146	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
147	const unsigned iPDDst = (RTGCUINTPTR)pvFault >> SHW_PD_SHIFT;
148	PX86PDPAE pPDDst = pVM->pgm.s.CTXMID(ap,PaePDs)[0]; /* We treat this as a PD with 2048 entries, so no need to and with SHW_PD_MASK to get iPDDst */
149
150	# if PGM_GST_TYPE == PGM_TYPE_PAE
151	/* Did we mark the PDPT as not present in SyncCR3? */
152	unsigned iPdpte = ((RTGCUINTPTR)pvFault >> SHW_PDPT_SHIFT) & SHW_PDPT_MASK;
153	if (!pVM->pgm.s.CTXMID(p,PaePDPT)->a[iPdpte].n.u1Present)
154	pVM->pgm.s.CTXMID(p,PaePDPT)->a[iPdpte].n.u1Present = 1;
155
156	# endif
157
158	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
159	const unsigned iPDDst = (((RTGCUINTPTR)pvFault >> SHW_PD_SHIFT) & SHW_PD_MASK);
160	PX86PDPAE pPDDst;
161	# if PGM_GST_TYPE == PGM_TYPE_PROT
162	/* AMD-V nested paging */
163	X86PML4E Pml4eSrc;
164	X86PDPE PdpeSrc;
165	PX86PML4E pPml4eSrc = &Pml4eSrc;
166
167	/* Fake PML4 & PDPT entry; access control handled on the page table level, so allow everything. */
168	Pml4eSrc.u = X86_PML4E_P \| X86_PML4E_RW \| X86_PML4E_US \| X86_PML4E_A;
169	PdpeSrc.u = X86_PDPE_P \| X86_PDPE_RW \| X86_PDPE_US \| X86_PDPE_A;
170	# endif
171
172	rc = PGMShwSyncLongModePDPtr(pVM, (RTGCUINTPTR)pvFault, pPml4eSrc, &PdpeSrc, &pPDDst);
173	if (rc != VINF_SUCCESS)
174	{
175	AssertRC(rc);
176	return rc;
177	}
178	Assert(pPDDst);
179
180	# elif PGM_SHW_TYPE == PGM_TYPE_EPT
181	const unsigned iPDDst = (((RTGCUINTPTR)pvFault >> SHW_PD_SHIFT) & SHW_PD_MASK);
182	PEPTPD pPDDst;
183
184	rc = PGMShwGetEPTPDPtr(pVM, (RTGCUINTPTR)pvFault, NULL, &pPDDst);
185	if (rc != VINF_SUCCESS)
186	{
187	AssertRC(rc);
188	return rc;
189	}
190	Assert(pPDDst);
191	# endif
192
193	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
194	/*
195	* If we successfully correct the write protection fault due to dirty bit
196	* tracking, or this page fault is a genuine one, then return immediately.
197	*/
198	STAM_PROFILE_START(&pVM->pgm.s.StatRZTrap0eTimeCheckPageFault, e);
199	rc = PGM_BTH_NAME(CheckPageFault)(pVM, uErr, &pPDDst->a[iPDDst], &pPDSrc->a[iPDSrc], (RTGCUINTPTR)pvFault);
200	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeCheckPageFault, e);
201	if ( rc == VINF_PGM_HANDLED_DIRTY_BIT_FAULT
202	\|\| rc == VINF_EM_RAW_GUEST_TRAP)
203	{
204	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution)
205	= rc == VINF_PGM_HANDLED_DIRTY_BIT_FAULT ? &pVM->pgm.s.StatRZTrap0eTime2DirtyAndAccessed : &pVM->pgm.s.StatRZTrap0eTime2GuestTrap; });
206	LogBird(("Trap0eHandler: returns %s\n", rc == VINF_PGM_HANDLED_DIRTY_BIT_FAULT ? "VINF_SUCCESS" : "VINF_EM_RAW_GUEST_TRAP"));
207	return rc == VINF_PGM_HANDLED_DIRTY_BIT_FAULT ? VINF_SUCCESS : rc;
208	}
209
210	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0ePD[iPDSrc]);
211	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
212
213	/*
214	* A common case is the not-present error caused by lazy page table syncing.
215	*
216	* It is IMPORTANT that we weed out any access to non-present shadow PDEs here
217	* so we can safely assume that the shadow PT is present when calling SyncPage later.
218	*
219	* On failure, we ASSUME that SyncPT is out of memory or detected some kind
220	* of mapping conflict and defer to SyncCR3 in R3.
221	* (Again, we do NOT support access handlers for non-present guest pages.)
222	*
223	*/
224	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
225	GSTPDE PdeSrc = pPDSrc->a[iPDSrc];
226	# else
227	GSTPDE PdeSrc;
228	PdeSrc.au32[0] = 0; /* faked so we don't have to #ifdef everything */
229	PdeSrc.n.u1Present = 1;
230	PdeSrc.n.u1Write = 1;
231	PdeSrc.n.u1Accessed = 1;
232	PdeSrc.n.u1User = 1;
233	# endif
234	if ( !(uErr & X86_TRAP_PF_P) /* not set means page not present instead of page protection violation */
235	&& !pPDDst->a[iPDDst].n.u1Present
236	&& PdeSrc.n.u1Present
237	)
238
239	{
240	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2SyncPT; });
241	STAM_PROFILE_START(&pVM->pgm.s.StatRZTrap0eTimeSyncPT, f);
242	LogFlow(("=>SyncPT %04x = %08x\n", iPDSrc, PdeSrc.au32[0]));
243	rc = PGM_BTH_NAME(SyncPT)(pVM, iPDSrc, pPDSrc, (RTGCUINTPTR)pvFault);
244	if (RT_SUCCESS(rc))
245	{
246	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeSyncPT, f);
247	return rc;
248	}
249	Log(("SyncPT: %d failed!! rc=%d\n", iPDSrc, rc));
250	VM_FF_SET(pVM, VM_FF_PGM_SYNC_CR3); /** @todo no need to do global sync, right? */
251	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeSyncPT, f);
252	return VINF_PGM_SYNC_CR3;
253	}
254
255	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
256	/*
257	* Check if this address is within any of our mappings.
258	*
259	* This is very fast and it's gonna save us a bit of effort below and prevent
260	* us from screwing ourself with MMIO2 pages which have a GC Mapping (VRam).
261	* (BTW, it's impossible to have physical access handlers in a mapping.)
262	*/
263	if (pgmMapAreMappingsEnabled(&pVM->pgm.s))
264	{
265	STAM_PROFILE_START(&pVM->pgm.s.StatRZTrap0eTimeMapping, a);
266	PPGMMAPPING pMapping = pVM->pgm.s.CTX_SUFF(pMappings);
267	for ( ; pMapping; pMapping = pMapping->CTX_SUFF(pNext))
268	{
269	if ((RTGCUINTPTR)pvFault < (RTGCUINTPTR)pMapping->GCPtr)
270	break;
271	if ((RTGCUINTPTR)pvFault - (RTGCUINTPTR)pMapping->GCPtr < pMapping->cb)
272	{
273	/*
274	* The first thing we check is if we've got an undetected conflict.
275	*/
276	if (!pVM->pgm.s.fMappingsFixed)
277	{
278	unsigned iPT = pMapping->cb >> GST_PD_SHIFT;
279	while (iPT-- > 0)
280	if (pPDSrc->a[iPDSrc + iPT].n.u1Present)
281	{
282	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eConflicts);
283	Log(("Trap0e: Detected Conflict %RGv-%RGv\n", pMapping->GCPtr, pMapping->GCPtrLast));
284	VM_FF_SET(pVM, VM_FF_PGM_SYNC_CR3); /** @todo no need to do global sync,right? */
285	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeMapping, a);
286	return VINF_PGM_SYNC_CR3;
287	}
288	}
289
290	/*
291	* Check if the fault address is in a virtual page access handler range.
292	*/
293	PPGMVIRTHANDLER pCur = (PPGMVIRTHANDLER)RTAvlroGCPtrRangeGet(&pVM->pgm.s.CTX_SUFF(pTrees)->HyperVirtHandlers, pvFault);
294	if ( pCur
295	&& (RTGCUINTPTR)pvFault - (RTGCUINTPTR)pCur->Core.Key < pCur->cb
296	&& uErr & X86_TRAP_PF_RW)
297	{
298	# ifdef IN_RC
299	STAM_PROFILE_START(&pCur->Stat, h);
300	rc = pCur->CTX_SUFF(pfnHandler)(pVM, uErr, pRegFrame, pvFault, pCur->Core.Key, (RTGCUINTPTR)pvFault - (RTGCUINTPTR)pCur->Core.Key);
301	STAM_PROFILE_STOP(&pCur->Stat, h);
302	# else
303	AssertFailed();
304	rc = VINF_EM_RAW_EMULATE_INSTR; /* can't happen with VMX */
305	# endif
306	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersMapping);
307	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeMapping, a);
308	return rc;
309	}
310
311	/*
312	* Pretend we're not here and let the guest handle the trap.
313	*/
314	TRPMSetErrorCode(pVM, uErr & ~X86_TRAP_PF_P);
315	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eGuestPFMapping);
316	LogFlow(("PGM: Mapping access -> route trap to recompiler!\n"));
317	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeMapping, a);
318	return VINF_EM_RAW_GUEST_TRAP;
319	}
320	}
321	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeMapping, a);
322	} /* pgmAreMappingsEnabled(&pVM->pgm.s) */
323	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
324
325	/*
326	* Check if this fault address is flagged for special treatment,
327	* which means we'll have to figure out the physical address and
328	* check flags associated with it.
329	*
330	* ASSUME that we can limit any special access handling to pages
331	* in page tables which the guest believes to be present.
332	*/
333	if (PdeSrc.n.u1Present)
334	{
335	RTGCPHYS GCPhys = NIL_RTGCPHYS;
336
337	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
338	# if PGM_GST_TYPE == PGM_TYPE_AMD64
339	bool fBigPagesSupported = true;
340	# else
341	bool fBigPagesSupported = !!(CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
342	# endif
343	if ( PdeSrc.b.u1Size
344	&& fBigPagesSupported)
345	GCPhys = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc)
346	\| ((RTGCPHYS)pvFault & (GST_BIG_PAGE_OFFSET_MASK ^ PAGE_OFFSET_MASK));
347	else
348	{
349	PGSTPT pPTSrc;
350	rc = PGM_GCPHYS_2_PTR(pVM, PdeSrc.u & GST_PDE_PG_MASK, &pPTSrc);
351	if (RT_SUCCESS(rc))
352	{
353	unsigned iPTESrc = ((RTGCUINTPTR)pvFault >> GST_PT_SHIFT) & GST_PT_MASK;
354	if (pPTSrc->a[iPTESrc].n.u1Present)
355	GCPhys = pPTSrc->a[iPTESrc].u & GST_PTE_PG_MASK;
356	}
357	}
358	# else
359	/* No paging so the fault address is the physical address */
360	GCPhys = (RTGCPHYS)((RTGCUINTPTR)pvFault & ~PAGE_OFFSET_MASK);
361	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
362
363	/*
364	* If we have a GC address we'll check if it has any flags set.
365	*/
366	if (GCPhys != NIL_RTGCPHYS)
367	{
368	STAM_PROFILE_START(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
369
370	PPGMPAGE pPage;
371	rc = pgmPhysGetPageEx(&pVM->pgm.s, GCPhys, &pPage);
372	if (RT_SUCCESS(rc))
373	{
374	if ( PGM_PAGE_HAS_ACTIVE_PHYSICAL_HANDLERS(pPage)
375	\|\| PGM_PAGE_HAS_ACTIVE_VIRTUAL_HANDLERS(pPage))
376	{
377	if (PGM_PAGE_HAS_ANY_PHYSICAL_HANDLERS(pPage))
378	{
379	/*
380	* Physical page access handler.
381	*/
382	const RTGCPHYS GCPhysFault = GCPhys \| ((RTGCUINTPTR)pvFault & PAGE_OFFSET_MASK);
383	PPGMPHYSHANDLER pCur = (PPGMPHYSHANDLER)RTAvlroGCPhysRangeGet(&pVM->pgm.s.CTX_SUFF(pTrees)->PhysHandlers, GCPhysFault);
384	if (pCur)
385	{
386	# ifdef PGM_SYNC_N_PAGES
387	/*
388	* If the region is write protected and we got a page not present fault, then sync
389	* the pages. If the fault was caused by a read, then restart the instruction.
390	* In case of write access continue to the GC write handler.
391	*
392	* ASSUMES that there is only one handler per page or that they have similar write properties.
393	*/
394	if ( pCur->enmType == PGMPHYSHANDLERTYPE_PHYSICAL_WRITE
395	&& !(uErr & X86_TRAP_PF_P))
396	{
397	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, (RTGCUINTPTR)pvFault, PGM_SYNC_NR_PAGES, uErr);
398	if ( RT_FAILURE(rc)
399	\|\| !(uErr & X86_TRAP_PF_RW)
400	\|\| rc == VINF_PGM_SYNCPAGE_MODIFIED_PDE)
401	{
402	AssertRC(rc);
403	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersOutOfSync);
404	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
405	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2OutOfSyncHndPhys; });
406	return rc;
407	}
408	}
409	# endif
410
411	AssertMsg( pCur->enmType != PGMPHYSHANDLERTYPE_PHYSICAL_WRITE
412	\|\| (pCur->enmType == PGMPHYSHANDLERTYPE_PHYSICAL_WRITE && (uErr & X86_TRAP_PF_RW)),
413	("Unexpected trap for physical handler: %08X (phys=%08x) HCPhys=%X uErr=%X, enum=%d\n", pvFault, GCPhys, pPage->HCPhys, uErr, pCur->enmType));
414
415	# if defined(IN_RC) \|\| defined(IN_RING0)
416	if (pCur->CTX_SUFF(pfnHandler))
417	{
418	STAM_PROFILE_START(&pCur->Stat, h);
419	rc = pCur->CTX_SUFF(pfnHandler)(pVM, uErr, pRegFrame, pvFault, GCPhysFault, pCur->CTX_SUFF(pvUser));
420	STAM_PROFILE_STOP(&pCur->Stat, h);
421	}
422	else
423	# endif
424	rc = VINF_EM_RAW_EMULATE_INSTR;
425	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersPhysical);
426	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
427	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2HndPhys; });
428	return rc;
429	}
430	}
431	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
432	else
433	{
434	# ifdef PGM_SYNC_N_PAGES
435	/*
436	* If the region is write protected and we got a page not present fault, then sync
437	* the pages. If the fault was caused by a read, then restart the instruction.
438	* In case of write access continue to the GC write handler.
439	*/
440	if ( PGM_PAGE_GET_HNDL_VIRT_STATE(pPage) < PGM_PAGE_HNDL_PHYS_STATE_ALL
441	&& !(uErr & X86_TRAP_PF_P))
442	{
443	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, (RTGCUINTPTR)pvFault, PGM_SYNC_NR_PAGES, uErr);
444	if ( RT_FAILURE(rc)
445	\|\| rc == VINF_PGM_SYNCPAGE_MODIFIED_PDE
446	\|\| !(uErr & X86_TRAP_PF_RW))
447	{
448	AssertRC(rc);
449	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersOutOfSync);
450	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
451	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2OutOfSyncHndVirt; });
452	return rc;
453	}
454	}
455	# endif
456	/*
457	* Ok, it's an virtual page access handler.
458	*
459	* Since it's faster to search by address, we'll do that first
460	* and then retry by GCPhys if that fails.
461	*/
462	/** @todo r=bird: perhaps we should consider looking up by physical address directly now? */
463	/** @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
464	* page was changed without us noticing it (not-present -> present without invlpg or mov cr3, xxx)
465	*/
466	PPGMVIRTHANDLER pCur = (PPGMVIRTHANDLER)RTAvlroGCPtrRangeGet(&pVM->pgm.s.CTX_SUFF(pTrees)->VirtHandlers, pvFault);
467	if (pCur)
468	{
469	AssertMsg(!((RTGCUINTPTR)pvFault - (RTGCUINTPTR)pCur->Core.Key < pCur->cb)
470	\|\| ( pCur->enmType != PGMVIRTHANDLERTYPE_WRITE
471	\|\| !(uErr & X86_TRAP_PF_P)
472	\|\| (pCur->enmType == PGMVIRTHANDLERTYPE_WRITE && (uErr & X86_TRAP_PF_RW))),
473	("Unexpected trap for virtual handler: %RGv (phys=%RGp) HCPhys=%HGp uErr=%X, enum=%d\n", pvFault, GCPhys, pPage->HCPhys, uErr, pCur->enmType));
474
475	if ( (RTGCUINTPTR)pvFault - (RTGCUINTPTR)pCur->Core.Key < pCur->cb
476	&& ( uErr & X86_TRAP_PF_RW
477	\|\| pCur->enmType != PGMVIRTHANDLERTYPE_WRITE ) )
478	{
479	# ifdef IN_RC
480	STAM_PROFILE_START(&pCur->Stat, h);
481	rc = pCur->CTX_SUFF(pfnHandler)(pVM, uErr, pRegFrame, pvFault, pCur->Core.Key, (RTGCUINTPTR)pvFault - (RTGCUINTPTR)pCur->Core.Key);
482	STAM_PROFILE_STOP(&pCur->Stat, h);
483	# else
484	rc = VINF_EM_RAW_EMULATE_INSTR; /** @todo for VMX */
485	# endif
486	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersVirtual);
487	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
488	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2HndVirt; });
489	return rc;
490	}
491	/* Unhandled part of a monitored page */
492	}
493	else
494	{
495	/* Check by physical address. */
496	PPGMVIRTHANDLER pCur;
497	unsigned iPage;
498	rc = pgmHandlerVirtualFindByPhysAddr(pVM, GCPhys + ((RTGCUINTPTR)pvFault & PAGE_OFFSET_MASK),
499	&pCur, &iPage);
500	Assert(RT_SUCCESS(rc) \|\| !pCur);
501	if ( pCur
502	&& ( uErr & X86_TRAP_PF_RW
503	\|\| pCur->enmType != PGMVIRTHANDLERTYPE_WRITE ) )
504	{
505	Assert((pCur->aPhysToVirt[iPage].Core.Key & X86_PTE_PAE_PG_MASK) == GCPhys);
506	# ifdef IN_RC
507	RTGCUINTPTR off = (iPage << PAGE_SHIFT) + ((RTGCUINTPTR)pvFault & PAGE_OFFSET_MASK) - ((RTGCUINTPTR)pCur->Core.Key & PAGE_OFFSET_MASK);
508	Assert(off < pCur->cb);
509	STAM_PROFILE_START(&pCur->Stat, h);
510	rc = pCur->CTX_SUFF(pfnHandler)(pVM, uErr, pRegFrame, pvFault, pCur->Core.Key, off);
511	STAM_PROFILE_STOP(&pCur->Stat, h);
512	# else
513	rc = VINF_EM_RAW_EMULATE_INSTR; /** @todo for VMX */
514	# endif
515	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersVirtualByPhys);
516	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
517	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2HndVirt; });
518	return rc;
519	}
520	}
521	}
522	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
523
524	/*
525	* There is a handled area of the page, but this fault doesn't belong to it.
526	* We must emulate the instruction.
527	*
528	* To avoid crashing (non-fatal) in the interpreter and go back to the recompiler
529	* we first check if this was a page-not-present fault for a page with only
530	* write access handlers. Restart the instruction if it wasn't a write access.
531	*/
532	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersUnhandled);
533
534	if ( !PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage)
535	&& !(uErr & X86_TRAP_PF_P))
536	{
537	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, (RTGCUINTPTR)pvFault, PGM_SYNC_NR_PAGES, uErr);
538	if ( RT_FAILURE(rc)
539	\|\| rc == VINF_PGM_SYNCPAGE_MODIFIED_PDE
540	\|\| !(uErr & X86_TRAP_PF_RW))
541	{
542	AssertRC(rc);
543	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersOutOfSync);
544	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
545	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2OutOfSyncHndPhys; });
546	return rc;
547	}
548	}
549
550	/** @todo This particular case can cause quite a lot of overhead. E.g. early stage of kernel booting in Ubuntu 6.06
551	* It's writing to an unhandled part of the LDT page several million times.
552	*/
553	rc = PGMInterpretInstruction(pVM, pRegFrame, pvFault);
554	LogFlow(("PGM: PGMInterpretInstruction -> rc=%d HCPhys=%RHp%s%s\n",
555	rc, pPage->HCPhys,
556	PGM_PAGE_HAS_ANY_PHYSICAL_HANDLERS(pPage) ? " phys" : "",
557	PGM_PAGE_HAS_ANY_VIRTUAL_HANDLERS(pPage) ? " virt" : ""));
558	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
559	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2HndUnhandled; });
560	return rc;
561	} /* if any kind of handler */
562
563	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
564	if (uErr & X86_TRAP_PF_P)
565	{
566	/*
567	* The page isn't marked, but it might still be monitored by a virtual page access handler.
568	* (ASSUMES no temporary disabling of virtual handlers.)
569	*/
570	/** @todo r=bird: Since the purpose is to catch out of sync pages with virtual handler(s) here,
571	* we should correct both the shadow page table and physical memory flags, and not only check for
572	* accesses within the handler region but for access to pages with virtual handlers. */
573	PPGMVIRTHANDLER pCur = (PPGMVIRTHANDLER)RTAvlroGCPtrRangeGet(&pVM->pgm.s.CTX_SUFF(pTrees)->VirtHandlers, pvFault);
574	if (pCur)
575	{
576	AssertMsg( !((RTGCUINTPTR)pvFault - (RTGCUINTPTR)pCur->Core.Key < pCur->cb)
577	\|\| ( pCur->enmType != PGMVIRTHANDLERTYPE_WRITE
578	\|\| !(uErr & X86_TRAP_PF_P)
579	\|\| (pCur->enmType == PGMVIRTHANDLERTYPE_WRITE && (uErr & X86_TRAP_PF_RW))),
580	("Unexpected trap for virtual handler: %08X (phys=%08x) HCPhys=%X uErr=%X, enum=%d\n", pvFault, GCPhys, pPage->HCPhys, uErr, pCur->enmType));
581
582	if ( (RTGCUINTPTR)pvFault - (RTGCUINTPTR)pCur->Core.Key < pCur->cb
583	&& ( uErr & X86_TRAP_PF_RW
584	\|\| pCur->enmType != PGMVIRTHANDLERTYPE_WRITE ) )
585	{
586	# ifdef IN_RC
587	STAM_PROFILE_START(&pCur->Stat, h);
588	rc = pCur->CTX_SUFF(pfnHandler)(pVM, uErr, pRegFrame, pvFault, pCur->Core.Key, (RTGCUINTPTR)pvFault - (RTGCUINTPTR)pCur->Core.Key);
589	STAM_PROFILE_STOP(&pCur->Stat, h);
590	# else
591	rc = VINF_EM_RAW_EMULATE_INSTR; /** @todo for VMX */
592	# endif
593	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersVirtualUnmarked);
594	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
595	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2HndVirt; });
596	return rc;
597	}
598	}
599	}
600	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
601	}
602	else
603	{
604	/* When the guest accesses invalid physical memory (e.g. probing of RAM or accessing a remapped MMIO range), then we'll fall
605	* back to the recompiler to emulate the instruction.
606	*/
607	LogFlow(("pgmPhysGetPageEx %RGp failed with %Rrc\n", GCPhys, rc));
608	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersInvalid);
609	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
610	return VINF_EM_RAW_EMULATE_INSTR;
611	}
612
613	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
614
615	# ifdef PGM_OUT_OF_SYNC_IN_GC
616	/*
617	* We are here only if page is present in Guest page tables and trap is not handled
618	* by our handlers.
619	* Check it for page out-of-sync situation.
620	*/
621	STAM_PROFILE_START(&pVM->pgm.s.StatRZTrap0eTimeOutOfSync, c);
622
623	if (!(uErr & X86_TRAP_PF_P))
624	{
625	/*
626	* Page is not present in our page tables.
627	* Try to sync it!
628	* BTW, fPageShw is invalid in this branch!
629	*/
630	if (uErr & X86_TRAP_PF_US)
631	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,PageOutOfSyncUser));
632	else /* supervisor */
633	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,PageOutOfSyncSupervisor));
634
635	# if defined(LOG_ENABLED) && !defined(IN_RING0)
636	RTGCPHYS GCPhys;
637	uint64_t fPageGst;
638	PGMGstGetPage(pVM, pvFault, &fPageGst, &GCPhys);
639	Log(("Page out of sync: %RGv eip=%08x PdeSrc.n.u1User=%d fPageGst=%08llx GCPhys=%RGp scan=%d\n",
640	pvFault, pRegFrame->eip, PdeSrc.n.u1User, fPageGst, GCPhys, CSAMDoesPageNeedScanning(pVM, (RTRCPTR)pRegFrame->eip)));
641	# endif /* LOG_ENABLED */
642
643	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) && !defined(IN_RING0)
644	if (CPUMGetGuestCPL(pVM, pRegFrame) == 0)
645	{
646	uint64_t fPageGst;
647	rc = PGMGstGetPage(pVM, pvFault, &fPageGst, NULL);
648	if ( RT_SUCCESS(rc)
649	&& !(fPageGst & X86_PTE_US))
650	{
651	/* Note: can't check for X86_TRAP_ID bit, because that requires execute disable support on the CPU */
652	if ( pvFault == (RTGCPTR)pRegFrame->eip
653	\|\| (RTGCUINTPTR)pvFault - pRegFrame->eip < 8 /* instruction crossing a page boundary */
654	# ifdef CSAM_DETECT_NEW_CODE_PAGES
655	\|\| ( !PATMIsPatchGCAddr(pVM, (RTGCPTR)pRegFrame->eip)
656	&& CSAMDoesPageNeedScanning(pVM, (RTRCPTR)pRegFrame->eip)) /* any new code we encounter here */
657	# endif /* CSAM_DETECT_NEW_CODE_PAGES */
658	)
659	{
660	LogFlow(("CSAMExecFault %RX32\n", pRegFrame->eip));
661	rc = CSAMExecFault(pVM, (RTRCPTR)pRegFrame->eip);
662	if (rc != VINF_SUCCESS)
663	{
664	/*
665	* CSAM needs to perform a job in ring 3.
666	*
667	* Sync the page before going to the host context; otherwise we'll end up in a loop if
668	* CSAM fails (e.g. instruction crosses a page boundary and the next page is not present)
669	*/
670	LogFlow(("CSAM ring 3 job\n"));
671	int rc2 = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, (RTGCUINTPTR)pvFault, 1, uErr);
672	AssertRC(rc2);
673
674	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeOutOfSync, c);
675	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2CSAM; });
676	return rc;
677	}
678	}
679	# ifdef CSAM_DETECT_NEW_CODE_PAGES
680	else if ( uErr == X86_TRAP_PF_RW
681	&& pRegFrame->ecx >= 0x100 /* early check for movswd count */
682	&& pRegFrame->ecx < 0x10000)
683	{
684	/* In case of a write to a non-present supervisor shadow page, we'll take special precautions
685	* to detect loading of new code pages.
686	*/
687
688	/*
689	* Decode the instruction.
690	*/
691	RTGCPTR PC;
692	rc = SELMValidateAndConvertCSAddr(pVM, pRegFrame->eflags, pRegFrame->ss, pRegFrame->cs, &pRegFrame->csHid, (RTGCPTR)pRegFrame->eip, &PC);
693	if (rc == VINF_SUCCESS)
694	{
695	DISCPUSTATE Cpu;
696	uint32_t cbOp;
697	rc = EMInterpretDisasOneEx(pVM, (RTGCUINTPTR)PC, pRegFrame, &Cpu, &cbOp);
698
699	/* For now we'll restrict this to rep movsw/d instructions */
700	if ( rc == VINF_SUCCESS
701	&& Cpu.pCurInstr->opcode == OP_MOVSWD
702	&& (Cpu.prefix & PREFIX_REP))
703	{
704	CSAMMarkPossibleCodePage(pVM, pvFault);
705	}
706	}
707	}
708	# endif /* CSAM_DETECT_NEW_CODE_PAGES */
709
710	/*
711	* Mark this page as safe.
712	*/
713	/** @todo not correct for pages that contain both code and data!! */
714	Log2(("CSAMMarkPage %RGv; scanned=%d\n", pvFault, true));
715	CSAMMarkPage(pVM, (RTRCPTR)pvFault, true);
716	}
717	}
718	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) && !defined(IN_RING0) */
719	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, (RTGCUINTPTR)pvFault, PGM_SYNC_NR_PAGES, uErr);
720	if (RT_SUCCESS(rc))
721	{
722	/* The page was successfully synced, return to the guest. */
723	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeOutOfSync, c);
724	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2OutOfSync; });
725	return VINF_SUCCESS;
726	}
727	}
728	else
729	{
730	/*
731	* A side effect of not flushing global PDEs are out of sync pages due
732	* to physical monitored regions, that are no longer valid.
733	* Assume for now it only applies to the read/write flag
734	*/
735	if (RT_SUCCESS(rc) && (uErr & X86_TRAP_PF_RW))
736	{
737	if (uErr & X86_TRAP_PF_US)
738	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,PageOutOfSyncUser));
739	else /* supervisor */
740	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,PageOutOfSyncSupervisor));
741
742
743	/*
744	* 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.
745	*/
746	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, (RTGCUINTPTR)pvFault, 1, uErr);
747	if (RT_SUCCESS(rc))
748	{
749	/*
750	* Page was successfully synced, return to guest.
751	*/
752	# ifdef VBOX_STRICT
753	RTGCPHYS GCPhys;
754	uint64_t fPageGst;
755	rc = PGMGstGetPage(pVM, pvFault, &fPageGst, &GCPhys);
756	Assert(RT_SUCCESS(rc) && fPageGst & X86_PTE_RW);
757	LogFlow(("Obsolete physical monitor page out of sync %RGv - phys %RGp flags=%08llx\n", pvFault, GCPhys, (uint64_t)fPageGst));
758
759	uint64_t fPageShw;
760	rc = PGMShwGetPage(pVM, pvFault, &fPageShw, NULL);
761	AssertMsg(RT_SUCCESS(rc) && fPageShw & X86_PTE_RW, ("rc=%Rrc fPageShw=%RX64\n", rc, fPageShw));
762	# endif /* VBOX_STRICT */
763	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeOutOfSync, c);
764	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2OutOfSyncHndObs; });
765	return VINF_SUCCESS;
766	}
767
768	/* Check to see if we need to emulate the instruction as X86_CR0_WP has been cleared. */
769	if ( CPUMGetGuestCPL(pVM, pRegFrame) == 0
770	&& ((CPUMGetGuestCR0(pVM) & (X86_CR0_WP \| X86_CR0_PG)) == X86_CR0_PG)
771	&& (uErr & (X86_TRAP_PF_RW \| X86_TRAP_PF_P)) == (X86_TRAP_PF_RW \| X86_TRAP_PF_P))
772	{
773	uint64_t fPageGst;
774	rc = PGMGstGetPage(pVM, pvFault, &fPageGst, NULL);
775	if ( RT_SUCCESS(rc)
776	&& !(fPageGst & X86_PTE_RW))
777	{
778	rc = PGMInterpretInstruction(pVM, pRegFrame, pvFault);
779	if (RT_SUCCESS(rc))
780	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eWPEmulInRZ);
781	else
782	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eWPEmulToR3);
783	return rc;
784	}
785	AssertMsgFailed(("Unexpected r/w page %RGv flag=%x rc=%Rrc\n", pvFault, (uint32_t)fPageGst, rc));
786	}
787	}
788
789	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
790	# ifdef VBOX_STRICT
791	/*
792	* Check for VMM page flags vs. Guest page flags consistency.
793	* Currently only for debug purposes.
794	*/
795	if (RT_SUCCESS(rc))
796	{
797	/* Get guest page flags. */
798	uint64_t fPageGst;
799	rc = PGMGstGetPage(pVM, pvFault, &fPageGst, NULL);
800	if (RT_SUCCESS(rc))
801	{
802	uint64_t fPageShw;
803	rc = PGMShwGetPage(pVM, pvFault, &fPageShw, NULL);
804
805	/*
806	* Compare page flags.
807	* Note: we have AVL, A, D bits desynched.
808	*/
809	AssertMsg((fPageShw & ~(X86_PTE_A \| X86_PTE_D \| X86_PTE_AVL_MASK)) == (fPageGst & ~(X86_PTE_A \| X86_PTE_D \| X86_PTE_AVL_MASK)),
810	("Page flags mismatch! pvFault=%RGv GCPhys=%RGp fPageShw=%08llx fPageGst=%08llx\n", pvFault, GCPhys, fPageShw, fPageGst));
811	}
812	else
813	AssertMsgFailed(("PGMGstGetPage rc=%Rrc\n", rc));
814	}
815	else
816	AssertMsgFailed(("PGMGCGetPage rc=%Rrc\n", rc));
817	# endif /* VBOX_STRICT */
818	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
819	}
820	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeOutOfSync, c);
821	# endif /* PGM_OUT_OF_SYNC_IN_GC */
822	}
823	else
824	{
825	/*
826	* Page not present in Guest OS or invalid page table address.
827	* This is potential virtual page access handler food.
828	*
829	* For the present we'll say that our access handlers don't
830	* work for this case - we've already discarded the page table
831	* not present case which is identical to this.
832	*
833	* When we perchance find we need this, we will probably have AVL
834	* trees (offset based) to operate on and we can measure their speed
835	* agains mapping a page table and probably rearrange this handling
836	* a bit. (Like, searching virtual ranges before checking the
837	* physical address.)
838	*/
839	}
840	}
841
842
843	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
844	/*
845	* Conclusion, this is a guest trap.
846	*/
847	LogFlow(("PGM: Unhandled #PF -> route trap to recompiler!\n"));
848	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eGuestPFUnh);
849	return VINF_EM_RAW_GUEST_TRAP;
850	# else
851	/* present, but not a monitored page; perhaps the guest is probing physical memory */
852	return VINF_EM_RAW_EMULATE_INSTR;
853	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
854
855
856	# else /* PGM_GST_TYPE != PGM_TYPE_32BIT */
857
858	AssertReleaseMsgFailed(("Shw=%d Gst=%d is not implemented!\n", PGM_GST_TYPE, PGM_SHW_TYPE));
859	return VERR_INTERNAL_ERROR;
860	# endif /* PGM_GST_TYPE != PGM_TYPE_32BIT */
861	}
862	#endif /* !IN_RING3 */
863
864
865	/**
866	* Emulation of the invlpg instruction.
867	*
868	*
869	* @returns VBox status code.
870	*
871	* @param pVM VM handle.
872	* @param GCPtrPage Page to invalidate.
873	*
874	* @remark ASSUMES that the guest is updating before invalidating. This order
875	* isn't required by the CPU, so this is speculative and could cause
876	* trouble.
877	*
878	* @todo Flush page or page directory only if necessary!
879	* @todo Add a #define for simply invalidating the page.
880	*/
881	PGM_BTH_DECL(int, InvalidatePage)(PVM pVM, RTGCUINTPTR GCPtrPage)
882	{
883	#if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) \
884	&& PGM_SHW_TYPE != PGM_TYPE_NESTED \
885	&& PGM_SHW_TYPE != PGM_TYPE_EPT
886	int rc;
887
888	LogFlow(("InvalidatePage %RGv\n", GCPtrPage));
889	/*
890	* Get the shadow PD entry and skip out if this PD isn't present.
891	* (Guessing that it is frequent for a shadow PDE to not be present, do this first.)
892	*/
893	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
894	const unsigned iPDDst = GCPtrPage >> SHW_PD_SHIFT;
895	PX86PDE pPdeDst = &pVM->pgm.s.CTXMID(p,32BitPD)->a[iPDDst];
896	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
897	const unsigned iPDDst = GCPtrPage >> SHW_PD_SHIFT; /* no mask; flat index into the 2048 entry array. */
898	const unsigned iPdpte = (GCPtrPage >> X86_PDPT_SHIFT); NOREF(iPdpte);
899	PX86PDEPAE pPdeDst = &pVM->pgm.s.CTXMID(ap,PaePDs[0])->a[iPDDst];
900	PX86PDPT pPdptDst = pVM->pgm.s.CTXMID(p,PaePDPT); NOREF(pPdptDst);
901
902	/* If the shadow PDPE isn't present, then skip the invalidate. */
903	if (!pPdptDst->a[iPdpte].n.u1Present)
904	{
905	Assert(!(pPdptDst->a[iPdpte].u & PGM_PLXFLAGS_MAPPING));
906	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePageSkipped));
907	return VINF_SUCCESS;
908	}
909
910	# else /* PGM_SHW_TYPE == PGM_TYPE_AMD64 */
911	/* PML4 */
912	AssertReturn(pVM->pgm.s.pHCPaePML4, VERR_INTERNAL_ERROR);
913
914	const unsigned iPml4e = (GCPtrPage >> X86_PML4_SHIFT) & X86_PML4_MASK;
915	const unsigned iPdpte = (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 = &pVM->pgm.s.pHCPaePML4->a[iPml4e];
920	rc = PGMShwGetLongModePDPtr(pVM, GCPtrPage, &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[iPdpte];
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	PX86PDPAE pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, GCPtrPage, &iPDSrc);
962	X86PDPE PdpeSrc = CTXSUFF(pVM->pgm.s.pGstPaePDPT)->a[iPdpte];
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[iPdpte].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.pHCShwAmd64CR3->idx, iPml4e);
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.pHCShwAmd64CR3->idx, iPml4e);
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.pHCShwAmd64CR3->idx, iPml4e);
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, iPdpte);
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, iPdpte);
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, iPdpte);
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	/* Note: This shouldn't actually be necessary as we monitor the PDPT page for changes. */
1098	if (!pPDSrc)
1099	{
1100	/* Guest PDPE not present */
1101	PX86PDPAE pPDPAE = pVM->pgm.s.CTXMID(ap,PaePDs)[0]; /* root of the 2048 PDE array */
1102	PX86PDEPAE pPDEDst = &pPDPAE->a[iPdpte * X86_PG_PAE_ENTRIES];
1103	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
1104
1105	Assert(!(CTXSUFF(pVM->pgm.s.pGstPaePDPT)->a[iPdpte].n.u1Present));
1106	LogFlow(("InvalidatePage: guest PDPE %d not present; clear shw pdpe\n", iPdpte));
1107	/* for each page directory entry */
1108	for (unsigned iPD = 0; iPD < X86_PG_PAE_ENTRIES; iPD++)
1109	{
1110	if ( pPDEDst[iPD].n.u1Present
1111	&& !(pPDEDst[iPD].u & PGM_PDFLAGS_MAPPING))
1112	{
1113	pgmPoolFreeByPage(pPool, pgmPoolGetPage(pPool, pPDEDst[iPD].u & SHW_PDE_PG_MASK), SHW_POOL_ROOT_IDX, iPdpte * X86_PG_PAE_ENTRIES + iPD);
1114	pPDEDst[iPD].u = 0;
1115	}
1116	}
1117	if (!(pPdptDst->a[iPdpte].u & PGM_PLXFLAGS_MAPPING))
1118	pPdptDst->a[iPdpte].n.u1Present = 0;
1119	PGM_INVL_GUEST_TLBS();
1120	}
1121	AssertMsg(pVM->pgm.s.fMappingsFixed \|\| (PdpeSrc.u & X86_PDPE_PG_MASK) == pVM->pgm.s.aGCPhysGstPaePDsMonitored[iPdpte], ("%RGp vs %RGp (mon)\n", (PdpeSrc.u & X86_PDPE_PG_MASK), pVM->pgm.s.aGCPhysGstPaePDsMonitored[iPdpte]));
1122	# endif
1123
1124
1125	/*
1126	* Deal with the Guest PDE.
1127	*/
1128	rc = VINF_SUCCESS;
1129	if (PdeSrc.n.u1Present)
1130	{
1131	if (PdeDst.u & PGM_PDFLAGS_MAPPING)
1132	{
1133	/*
1134	* Conflict - Let SyncPT deal with it to avoid duplicate code.
1135	*/
1136	Assert(pgmMapAreMappingsEnabled(&pVM->pgm.s));
1137	Assert(PGMGetGuestMode(pVM) <= PGMMODE_PAE);
1138	rc = PGM_BTH_NAME(SyncPT)(pVM, iPDSrc, pPDSrc, GCPtrPage);
1139	}
1140	else if ( PdeSrc.n.u1User != PdeDst.n.u1User
1141	\|\| (!PdeSrc.n.u1Write && PdeDst.n.u1Write))
1142	{
1143	/*
1144	* Mark not present so we can resync the PDE when it's used.
1145	*/
1146	LogFlow(("InvalidatePage: Out-of-sync at %RGp PdeSrc=%RX64 PdeDst=%RX64\n",
1147	GCPtrPage, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
1148	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1149	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, pShwPde->idx, iPDDst);
1150	# else
1151	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, SHW_POOL_ROOT_IDX, iPDDst);
1152	# endif
1153	pPdeDst->u = 0;
1154	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDOutOfSync));
1155	PGM_INVL_GUEST_TLBS();
1156	}
1157	else if (!PdeSrc.n.u1Accessed)
1158	{
1159	/*
1160	* Mark not present so we can set the accessed bit.
1161	*/
1162	LogFlow(("InvalidatePage: Out-of-sync (A) at %RGp PdeSrc=%RX64 PdeDst=%RX64\n",
1163	GCPtrPage, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
1164	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1165	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, pShwPde->idx, iPDDst);
1166	# else
1167	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, SHW_POOL_ROOT_IDX, iPDDst);
1168	# endif
1169	pPdeDst->u = 0;
1170	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDNAs));
1171	PGM_INVL_GUEST_TLBS();
1172	}
1173	else if (!fIsBigPage)
1174	{
1175	/*
1176	* 4KB - page.
1177	*/
1178	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, PdeDst.u & SHW_PDE_PG_MASK);
1179	RTGCPHYS GCPhys = PdeSrc.u & GST_PDE_PG_MASK;
1180	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
1181	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
1182	GCPhys \|= (iPDDst & 1) * (PAGE_SIZE/2);
1183	# endif
1184	if (pShwPage->GCPhys == GCPhys)
1185	{
1186	# if 0 /* likely cause of a major performance regression; must be SyncPageWorkerTrackDeref then */
1187	const unsigned iPTEDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1188	PSHWPT pPT = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
1189	if (pPT->a[iPTEDst].n.u1Present)
1190	{
1191	# ifdef PGMPOOL_WITH_USER_TRACKING
1192	/* This is very unlikely with caching/monitoring enabled. */
1193	PGM_BTH_NAME(SyncPageWorkerTrackDeref)(pVM, pShwPage, pPT->a[iPTEDst].u & SHW_PTE_PG_MASK);
1194	# endif
1195	pPT->a[iPTEDst].u = 0;
1196	}
1197	# else /* Syncing it here isn't 100% safe and it's probably not worth spending time syncing it. */
1198	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, GCPtrPage, 1, 0);
1199	if (RT_SUCCESS(rc))
1200	rc = VINF_SUCCESS;
1201	# endif
1202	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePage4KBPages));
1203	PGM_INVL_PG(GCPtrPage);
1204	}
1205	else
1206	{
1207	/*
1208	* The page table address changed.
1209	*/
1210	LogFlow(("InvalidatePage: Out-of-sync at %RGp PdeSrc=%RX64 PdeDst=%RX64 ShwGCPhys=%RGp iPDDst=%#x\n",
1211	GCPtrPage, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u, pShwPage->GCPhys, iPDDst));
1212	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1213	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, pShwPde->idx, iPDDst);
1214	# else
1215	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, SHW_POOL_ROOT_IDX, iPDDst);
1216	# endif
1217	pPdeDst->u = 0;
1218	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDOutOfSync));
1219	PGM_INVL_GUEST_TLBS();
1220	}
1221	}
1222	else
1223	{
1224	/*
1225	* 2/4MB - page.
1226	*/
1227	/* Before freeing the page, check if anything really changed. */
1228	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, PdeDst.u & SHW_PDE_PG_MASK);
1229	RTGCPHYS GCPhys = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc);
1230	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
1231	/* Select the right PDE as we're emulating a 4MB page directory with two 2 MB shadow PDEs.*/
1232	GCPhys \|= GCPtrPage & (1 << X86_PD_PAE_SHIFT);
1233	# endif
1234	if ( pShwPage->GCPhys == GCPhys
1235	&& pShwPage->enmKind == BTH_PGMPOOLKIND_PT_FOR_BIG)
1236	{
1237	/* ASSUMES a the given bits are identical for 4M and normal PDEs */
1238	/** @todo PAT */
1239	if ( (PdeSrc.u & (X86_PDE_P \| X86_PDE_RW \| X86_PDE_US \| X86_PDE_PWT \| X86_PDE_PCD))
1240	== (PdeDst.u & (X86_PDE_P \| X86_PDE_RW \| X86_PDE_US \| X86_PDE_PWT \| X86_PDE_PCD))
1241	&& ( PdeSrc.b.u1Dirty /** @todo rainy day: What about read-only 4M pages? not very common, but still... */
1242	\|\| (PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY)))
1243	{
1244	LogFlow(("Skipping flush for big page containing %RGv (PD=%X .u=%RX64)-> nothing has changed!\n", GCPtrPage, iPDSrc, PdeSrc.u));
1245	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePage4MBPagesSkip));
1246	return VINF_SUCCESS;
1247	}
1248	}
1249
1250	/*
1251	* Ok, the page table is present and it's been changed in the guest.
1252	* If we're in host context, we'll just mark it as not present taking the lazy approach.
1253	* We could do this for some flushes in GC too, but we need an algorithm for
1254	* deciding which 4MB pages containing code likely to be executed very soon.
1255	*/
1256	LogFlow(("InvalidatePage: Out-of-sync PD at %RGp PdeSrc=%RX64 PdeDst=%RX64\n",
1257	GCPtrPage, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
1258	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1259	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, pShwPde->idx, iPDDst);
1260	# else
1261	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, SHW_POOL_ROOT_IDX, iPDDst);
1262	# endif
1263	pPdeDst->u = 0;
1264	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePage4MBPages));
1265	PGM_INVL_BIG_PG(GCPtrPage);
1266	}
1267	}
1268	else
1269	{
1270	/*
1271	* Page directory is not present, mark shadow PDE not present.
1272	*/
1273	if (!(PdeDst.u & PGM_PDFLAGS_MAPPING))
1274	{
1275	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1276	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, pShwPde->idx, iPDDst);
1277	# else
1278	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, SHW_POOL_ROOT_IDX, iPDDst);
1279	# endif
1280	pPdeDst->u = 0;
1281	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDNPs));
1282	PGM_INVL_PG(GCPtrPage);
1283	}
1284	else
1285	{
1286	Assert(pgmMapAreMappingsEnabled(&pVM->pgm.s));
1287	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDMappings));
1288	}
1289	}
1290
1291	return rc;
1292
1293	#else /* guest real and protected mode */
1294	/* There's no such thing as InvalidatePage when paging is disabled, so just ignore. */
1295	return VINF_SUCCESS;
1296	#endif
1297	}
1298
1299
1300	#ifdef PGMPOOL_WITH_USER_TRACKING
1301	/**
1302	* Update the tracking of shadowed pages.
1303	*
1304	* @param pVM The VM handle.
1305	* @param pShwPage The shadow page.
1306	* @param HCPhys The physical page we is being dereferenced.
1307	*/
1308	DECLINLINE(void) PGM_BTH_NAME(SyncPageWorkerTrackDeref)(PVM pVM, PPGMPOOLPAGE pShwPage, RTHCPHYS HCPhys)
1309	{
1310	# ifdef PGMPOOL_WITH_GCPHYS_TRACKING
1311	STAM_PROFILE_START(&pVM->pgm.s.StatTrackDeref, a);
1312	LogFlow(("SyncPageWorkerTrackDeref: Damn HCPhys=%RHp pShwPage->idx=%#x!!!\n", HCPhys, pShwPage->idx));
1313
1314	/** @todo If this turns out to be a bottle neck (very likely) two things can be done:
1315	* 1. have a medium sized HCPhys -> GCPhys TLB (hash?)
1316	* 2. write protect all shadowed pages. I.e. implement caching.
1317	*/
1318	/*
1319	* Find the guest address.
1320	*/
1321	for (PPGMRAMRANGE pRam = pVM->pgm.s.CTX_SUFF(pRamRanges);
1322	pRam;
1323	pRam = pRam->CTX_SUFF(pNext))
1324	{
1325	unsigned iPage = pRam->cb >> PAGE_SHIFT;
1326	while (iPage-- > 0)
1327	{
1328	if (PGM_PAGE_GET_HCPHYS(&pRam->aPages[iPage]) == HCPhys)
1329	{
1330	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
1331	pgmTrackDerefGCPhys(pPool, pShwPage, &pRam->aPages[iPage]);
1332	pShwPage->cPresent--;
1333	pPool->cPresent--;
1334	STAM_PROFILE_STOP(&pVM->pgm.s.StatTrackDeref, a);
1335	return;
1336	}
1337	}
1338	}
1339
1340	for (;;)
1341	AssertReleaseMsgFailed(("HCPhys=%RHp wasn't found!\n", HCPhys));
1342	# else /* !PGMPOOL_WITH_GCPHYS_TRACKING */
1343	pShwPage->cPresent--;
1344	pVM->pgm.s.CTX_SUFF(pPool)->cPresent--;
1345	# endif /* !PGMPOOL_WITH_GCPHYS_TRACKING */
1346	}
1347
1348
1349	/**
1350	* Update the tracking of shadowed pages.
1351	*
1352	* @param pVM The VM handle.
1353	* @param pShwPage The shadow page.
1354	* @param u16 The top 16-bit of the pPage->HCPhys.
1355	* @param pPage Pointer to the guest page. this will be modified.
1356	* @param iPTDst The index into the shadow table.
1357	*/
1358	DECLINLINE(void) PGM_BTH_NAME(SyncPageWorkerTrackAddref)(PVM pVM, PPGMPOOLPAGE pShwPage, uint16_t u16, PPGMPAGE pPage, const unsigned iPTDst)
1359	{
1360	# ifdef PGMPOOL_WITH_GCPHYS_TRACKING
1361	/*
1362	* We're making certain assumptions about the placement of cRef and idx.
1363	*/
1364	Assert(MM_RAM_FLAGS_IDX_SHIFT == 48);
1365	Assert(MM_RAM_FLAGS_CREFS_SHIFT > MM_RAM_FLAGS_IDX_SHIFT);
1366
1367	/*
1368	* Just deal with the simple first time here.
1369	*/
1370	if (!u16)
1371	{
1372	STAM_COUNTER_INC(&pVM->pgm.s.StatTrackVirgin);
1373	u16 = (1 << (MM_RAM_FLAGS_CREFS_SHIFT - MM_RAM_FLAGS_IDX_SHIFT)) \| pShwPage->idx;
1374	}
1375	else
1376	u16 = pgmPoolTrackPhysExtAddref(pVM, u16, pShwPage->idx);
1377
1378	/* write back, trying to be clever... */
1379	Log2(("SyncPageWorkerTrackAddRef: u16=%#x pPage->HCPhys=%RHp->%RHp iPTDst=%#x\n",
1380	u16, pPage->HCPhys, (pPage->HCPhys & MM_RAM_FLAGS_NO_REFS_MASK) \| ((uint64_t)u16 << MM_RAM_FLAGS_CREFS_SHIFT), iPTDst));
1381	((uint16_t )&pPage->HCPhys + 3) = u16; /** @todo PAGE FLAGS */
1382	# endif /* PGMPOOL_WITH_GCPHYS_TRACKING */
1383
1384	/* update statistics. */
1385	pVM->pgm.s.CTX_SUFF(pPool)->cPresent++;
1386	pShwPage->cPresent++;
1387	if (pShwPage->iFirstPresent > iPTDst)
1388	pShwPage->iFirstPresent = iPTDst;
1389	}
1390	#endif /* PGMPOOL_WITH_USER_TRACKING */
1391
1392
1393	/**
1394	* Creates a 4K shadow page for a guest page.
1395	*
1396	* For 4M pages the caller must convert the PDE4M to a PTE, this includes adjusting the
1397	* physical address. The PdeSrc argument only the flags are used. No page structured
1398	* will be mapped in this function.
1399	*
1400	* @param pVM VM handle.
1401	* @param pPteDst Destination page table entry.
1402	* @param PdeSrc Source page directory entry (i.e. Guest OS page directory entry).
1403	* Can safely assume that only the flags are being used.
1404	* @param PteSrc Source page table entry (i.e. Guest OS page table entry).
1405	* @param pShwPage Pointer to the shadow page.
1406	* @param iPTDst The index into the shadow table.
1407	*
1408	* @remark Not used for 2/4MB pages!
1409	*/
1410	DECLINLINE(void) PGM_BTH_NAME(SyncPageWorker)(PVM pVM, PSHWPTE pPteDst, GSTPDE PdeSrc, GSTPTE PteSrc, PPGMPOOLPAGE pShwPage, unsigned iPTDst)
1411	{
1412	if (PteSrc.n.u1Present)
1413	{
1414	/*
1415	* Find the ram range.
1416	*/
1417	PPGMPAGE pPage;
1418	int rc = pgmPhysGetPageEx(&pVM->pgm.s, PteSrc.u & GST_PTE_PG_MASK, &pPage);
1419	if (RT_SUCCESS(rc))
1420	{
1421	/** @todo investiage PWT, PCD and PAT. */
1422	/*
1423	* Make page table entry.
1424	*/
1425	const RTHCPHYS HCPhys = pPage->HCPhys; /** @todo FLAGS */
1426	SHWPTE PteDst;
1427	if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
1428	{
1429	/** @todo r=bird: Are we actually handling dirty and access bits for pages with access handlers correctly? No. */
1430	if (!PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage))
1431	{
1432	#if PGM_SHW_TYPE == PGM_TYPE_EPT
1433	PteDst.u = (HCPhys & EPT_PTE_PG_MASK);
1434	PteDst.n.u1Present = 1;
1435	PteDst.n.u1Execute = 1;
1436	PteDst.n.u1IgnorePAT = 1;
1437	PteDst.n.u3EMT = VMX_EPT_MEMTYPE_WB;
1438	/* PteDst.n.u1Write = 0 && PteDst.n.u1Size = 0 */
1439	#else
1440	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))
1441	\| (HCPhys & X86_PTE_PAE_PG_MASK);
1442	#endif
1443	}
1444	else
1445	{
1446	LogFlow(("SyncPageWorker: monitored page (%RHp) -> mark not present\n", HCPhys));
1447	PteDst.u = 0;
1448	}
1449	/** @todo count these two kinds. */
1450	}
1451	else
1452	{
1453	#if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
1454	/*
1455	* If the page or page directory entry is not marked accessed,
1456	* we mark the page not present.
1457	*/
1458	if (!PteSrc.n.u1Accessed \|\| !PdeSrc.n.u1Accessed)
1459	{
1460	LogFlow(("SyncPageWorker: page and or page directory not accessed -> mark not present\n"));
1461	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,AccessedPage));
1462	PteDst.u = 0;
1463	}
1464	else
1465	/*
1466	* If the page is not flagged as dirty and is writable, then make it read-only, so we can set the dirty bit
1467	* when the page is modified.
1468	*/
1469	if (!PteSrc.n.u1Dirty && (PdeSrc.n.u1Write & PteSrc.n.u1Write))
1470	{
1471	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPage));
1472	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))
1473	\| (HCPhys & X86_PTE_PAE_PG_MASK)
1474	\| PGM_PTFLAGS_TRACK_DIRTY;
1475	}
1476	else
1477	#endif
1478	{
1479	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPageSkipped));
1480	#if PGM_SHW_TYPE == PGM_TYPE_EPT
1481	PteDst.u = (HCPhys & EPT_PTE_PG_MASK);
1482	PteDst.n.u1Present = 1;
1483	PteDst.n.u1Write = 1;
1484	PteDst.n.u1Execute = 1;
1485	PteDst.n.u1IgnorePAT = 1;
1486	PteDst.n.u3EMT = VMX_EPT_MEMTYPE_WB;
1487	/* PteDst.n.u1Size = 0 */
1488	#else
1489	PteDst.u = (PteSrc.u & ~(X86_PTE_PAE_PG_MASK \| X86_PTE_AVL_MASK \| X86_PTE_PAT \| X86_PTE_PCD \| X86_PTE_PWT))
1490	\| (HCPhys & X86_PTE_PAE_PG_MASK);
1491	#endif
1492	}
1493	}
1494
1495	#ifdef PGMPOOL_WITH_USER_TRACKING
1496	/*
1497	* Keep user track up to date.
1498	*/
1499	if (PteDst.n.u1Present)
1500	{
1501	if (!pPteDst->n.u1Present)
1502	PGM_BTH_NAME(SyncPageWorkerTrackAddref)(pVM, pShwPage, HCPhys >> MM_RAM_FLAGS_IDX_SHIFT, pPage, iPTDst);
1503	else if ((pPteDst->u & SHW_PTE_PG_MASK) != (PteDst.u & SHW_PTE_PG_MASK))
1504	{
1505	Log2(("SyncPageWorker: deref! *pPteDst=%RX64 PteDst=%RX64\n", (uint64_t)pPteDst->u, (uint64_t)PteDst.u));
1506	PGM_BTH_NAME(SyncPageWorkerTrackDeref)(pVM, pShwPage, pPteDst->u & SHW_PTE_PG_MASK);
1507	PGM_BTH_NAME(SyncPageWorkerTrackAddref)(pVM, pShwPage, HCPhys >> MM_RAM_FLAGS_IDX_SHIFT, pPage, iPTDst);
1508	}
1509	}
1510	else if (pPteDst->n.u1Present)
1511	{
1512	Log2(("SyncPageWorker: deref! *pPteDst=%RX64\n", (uint64_t)pPteDst->u));
1513	PGM_BTH_NAME(SyncPageWorkerTrackDeref)(pVM, pShwPage, pPteDst->u & SHW_PTE_PG_MASK);
1514	}
1515	#endif /* PGMPOOL_WITH_USER_TRACKING */
1516
1517	/*
1518	* Update statistics and commit the entry.
1519	*/
1520	if (!PteSrc.n.u1Global)
1521	pShwPage->fSeenNonGlobal = true;
1522	*pPteDst = PteDst;
1523	}
1524	/* else MMIO or invalid page, we must handle them manually in the #PF handler. */
1525	/** @todo count these. */
1526	}
1527	else
1528	{
1529	/*
1530	* Page not-present.
1531	*/
1532	LogFlow(("SyncPageWorker: page not present in Pte\n"));
1533	#ifdef PGMPOOL_WITH_USER_TRACKING
1534	/* Keep user track up to date. */
1535	if (pPteDst->n.u1Present)
1536	{
1537	Log2(("SyncPageWorker: deref! *pPteDst=%RX64\n", (uint64_t)pPteDst->u));
1538	PGM_BTH_NAME(SyncPageWorkerTrackDeref)(pVM, pShwPage, pPteDst->u & SHW_PTE_PG_MASK);
1539	}
1540	#endif /* PGMPOOL_WITH_USER_TRACKING */
1541	pPteDst->u = 0;
1542	/** @todo count these. */
1543	}
1544	}
1545
1546
1547	/**
1548	* Syncs a guest OS page.
1549	*
1550	* There are no conflicts at this point, neither is there any need for
1551	* page table allocations.
1552	*
1553	* @returns VBox status code.
1554	* @returns VINF_PGM_SYNCPAGE_MODIFIED_PDE if it modifies the PDE in any way.
1555	* @param pVM VM handle.
1556	* @param PdeSrc Page directory entry of the guest.
1557	* @param GCPtrPage Guest context page address.
1558	* @param cPages Number of pages to sync (PGM_SYNC_N_PAGES) (default=1).
1559	* @param uErr Fault error (X86_TRAP_PF_*).
1560	*/
1561	PGM_BTH_DECL(int, SyncPage)(PVM pVM, GSTPDE PdeSrc, RTGCUINTPTR GCPtrPage, unsigned cPages, unsigned uErr)
1562	{
1563	LogFlow(("SyncPage: GCPtrPage=%RGv cPages=%u uErr=%#x\n", GCPtrPage, cPages, uErr));
1564
1565	#if ( PGM_GST_TYPE == PGM_TYPE_32BIT \
1566	\|\| PGM_GST_TYPE == PGM_TYPE_PAE \
1567	\|\| PGM_GST_TYPE == PGM_TYPE_AMD64) \
1568	&& PGM_SHW_TYPE != PGM_TYPE_NESTED \
1569	&& PGM_SHW_TYPE != PGM_TYPE_EPT
1570
1571	# if PGM_WITH_NX(PGM_GST_TYPE, PGM_SHW_TYPE)
1572	bool fNoExecuteBitValid = !!(CPUMGetGuestEFER(pVM) & MSR_K6_EFER_NXE);
1573	# endif
1574
1575	/*
1576	* Assert preconditions.
1577	*/
1578	Assert(PdeSrc.n.u1Present);
1579	Assert(cPages);
1580	STAM_COUNTER_INC(&pVM->pgm.s.StatSyncPagePD[(GCPtrPage >> GST_PD_SHIFT) & GST_PD_MASK]);
1581
1582	/*
1583	* Get the shadow PDE, find the shadow page table in the pool.
1584	*/
1585	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
1586	const unsigned iPDDst = GCPtrPage >> SHW_PD_SHIFT;
1587	X86PDE PdeDst = pVM->pgm.s.CTXMID(p,32BitPD)->a[iPDDst];
1588	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
1589	const unsigned iPDDst = GCPtrPage >> SHW_PD_SHIFT;
1590	const unsigned iPdpte = (GCPtrPage >> X86_PDPT_SHIFT); NOREF(iPdpte); /* no mask; flat index into the 2048 entry array. */
1591	PX86PDPT pPdptDst = pVM->pgm.s.CTXMID(p,PaePDPT); NOREF(pPdptDst);
1592	X86PDEPAE PdeDst = pVM->pgm.s.CTXMID(ap,PaePDs)[0]->a[iPDDst];
1593	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
1594	const unsigned iPDDst = ((GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK);
1595	const unsigned iPdpte = (GCPtrPage >> X86_PDPT_SHIFT) & X86_PDPT_MASK_AMD64;
1596	PX86PDPAE pPDDst;
1597	X86PDEPAE PdeDst;
1598	PX86PDPT pPdptDst;
1599
1600	int rc = PGMShwGetLongModePDPtr(pVM, GCPtrPage, &pPdptDst, &pPDDst);
1601	AssertRCSuccessReturn(rc, rc);
1602	Assert(pPDDst && pPdptDst);
1603	PdeDst = pPDDst->a[iPDDst];
1604	# endif
1605	Assert(PdeDst.n.u1Present);
1606	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, PdeDst.u & SHW_PDE_PG_MASK);
1607
1608	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1609	/* Fetch the pgm pool shadow descriptor. */
1610	PPGMPOOLPAGE pShwPde = pgmPoolGetPageByHCPhys(pVM, pPdptDst->a[iPdpte].u & X86_PDPE_PG_MASK);
1611	Assert(pShwPde);
1612	# endif
1613
1614	/*
1615	* Check that the page is present and that the shadow PDE isn't out of sync.
1616	*/
1617	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1618	const bool fBigPage = PdeSrc.b.u1Size;
1619	# else
1620	const bool fBigPage = PdeSrc.b.u1Size && (CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
1621	# endif
1622	RTGCPHYS GCPhys;
1623	if (!fBigPage)
1624	{
1625	GCPhys = PdeSrc.u & GST_PDE_PG_MASK;
1626	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
1627	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
1628	GCPhys \|= (iPDDst & 1) * (PAGE_SIZE/2);
1629	# endif
1630	}
1631	else
1632	{
1633	GCPhys = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc);
1634	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
1635	/* Select the right PDE as we're emulating a 4MB page directory with two 2 MB shadow PDEs.*/
1636	GCPhys \|= GCPtrPage & (1 << X86_PD_PAE_SHIFT);
1637	# endif
1638	}
1639	if ( pShwPage->GCPhys == GCPhys
1640	&& PdeSrc.n.u1Present
1641	&& (PdeSrc.n.u1User == PdeDst.n.u1User)
1642	&& (PdeSrc.n.u1Write == PdeDst.n.u1Write \|\| !PdeDst.n.u1Write)
1643	# if PGM_WITH_NX(PGM_GST_TYPE, PGM_SHW_TYPE)
1644	&& (!fNoExecuteBitValid \|\| PdeSrc.n.u1NoExecute == PdeDst.n.u1NoExecute)
1645	# endif
1646	)
1647	{
1648	/*
1649	* Check that the PDE is marked accessed already.
1650	* Since we set the accessed bit before getting here on a #PF, this
1651	* check is only meant for dealing with non-#PF'ing paths.
1652	*/
1653	if (PdeSrc.n.u1Accessed)
1654	{
1655	PSHWPT pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
1656	if (!fBigPage)
1657	{
1658	/*
1659	* 4KB Page - Map the guest page table.
1660	*/
1661	PGSTPT pPTSrc;
1662	int rc = PGM_GCPHYS_2_PTR(pVM, PdeSrc.u & GST_PDE_PG_MASK, &pPTSrc);
1663	if (RT_SUCCESS(rc))
1664	{
1665	# ifdef PGM_SYNC_N_PAGES
1666	Assert(cPages == 1 \|\| !(uErr & X86_TRAP_PF_P));
1667	if (cPages > 1 && !(uErr & X86_TRAP_PF_P))
1668	{
1669	/*
1670	* This code path is currently only taken when the caller is PGMTrap0eHandler
1671	* for non-present pages!
1672	*
1673	* We're setting PGM_SYNC_NR_PAGES pages around the faulting page to sync it and
1674	* deal with locality.
1675	*/
1676	unsigned iPTDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1677	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
1678	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
1679	const unsigned offPTSrc = ((GCPtrPage >> SHW_PD_SHIFT) & 1) * 512;
1680	# else
1681	const unsigned offPTSrc = 0;
1682	# endif
1683	const unsigned iPTDstEnd = RT_MIN(iPTDst + PGM_SYNC_NR_PAGES / 2, RT_ELEMENTS(pPTDst->a));
1684	if (iPTDst < PGM_SYNC_NR_PAGES / 2)
1685	iPTDst = 0;
1686	else
1687	iPTDst -= PGM_SYNC_NR_PAGES / 2;
1688	for (; iPTDst < iPTDstEnd; iPTDst++)
1689	{
1690	if (!pPTDst->a[iPTDst].n.u1Present)
1691	{
1692	GSTPTE PteSrc = pPTSrc->a[offPTSrc + iPTDst];
1693	RTGCUINTPTR GCPtrCurPage = ((RTGCUINTPTR)GCPtrPage & ~(RTGCUINTPTR)(GST_PT_MASK << GST_PT_SHIFT)) \| ((offPTSrc + iPTDst) << PAGE_SHIFT);
1694	NOREF(GCPtrCurPage);
1695	#ifndef IN_RING0
1696	/*
1697	* Assuming kernel code will be marked as supervisor - and not as user level
1698	* and executed using a conforming code selector - And marked as readonly.
1699	* Also assume that if we're monitoring a page, it's of no interest to CSAM.
1700	*/
1701	PPGMPAGE pPage;
1702	if ( ((PdeSrc.u & PteSrc.u) & (X86_PTE_RW \| X86_PTE_US))
1703	\|\| iPTDst == ((GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK) /* always sync GCPtrPage */
1704	\|\| !CSAMDoesPageNeedScanning(pVM, (RTRCPTR)GCPtrCurPage)
1705	\|\| ( (pPage = pgmPhysGetPage(&pVM->pgm.s, PteSrc.u & GST_PTE_PG_MASK))
1706	&& PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
1707	)
1708	#endif /* else: CSAM not active */
1709	PGM_BTH_NAME(SyncPageWorker)(pVM, &pPTDst->a[iPTDst], PdeSrc, PteSrc, pShwPage, iPTDst);
1710	Log2(("SyncPage: 4K+ %RGv PteSrc:{P=%d RW=%d U=%d raw=%08llx} PteDst=%08llx%s\n",
1711	GCPtrCurPage, PteSrc.n.u1Present,
1712	PteSrc.n.u1Write & PdeSrc.n.u1Write,
1713	PteSrc.n.u1User & PdeSrc.n.u1User,
1714	(uint64_t)PteSrc.u,
1715	(uint64_t)pPTDst->a[iPTDst].u,
1716	pPTDst->a[iPTDst].u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
1717	}
1718	}
1719	}
1720	else
1721	# endif /* PGM_SYNC_N_PAGES */
1722	{
1723	const unsigned iPTSrc = (GCPtrPage >> GST_PT_SHIFT) & GST_PT_MASK;
1724	GSTPTE PteSrc = pPTSrc->a[iPTSrc];
1725	const unsigned iPTDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1726	PGM_BTH_NAME(SyncPageWorker)(pVM, &pPTDst->a[iPTDst], PdeSrc, PteSrc, pShwPage, iPTDst);
1727	Log2(("SyncPage: 4K %RGv PteSrc:{P=%d RW=%d U=%d raw=%08llx}%s\n",
1728	GCPtrPage, PteSrc.n.u1Present,
1729	PteSrc.n.u1Write & PdeSrc.n.u1Write,
1730	PteSrc.n.u1User & PdeSrc.n.u1User,
1731	(uint64_t)PteSrc.u,
1732	pPTDst->a[iPTDst].u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
1733	}
1734	}
1735	else /* MMIO or invalid page: emulated in #PF handler. */
1736	{
1737	LogFlow(("PGM_GCPHYS_2_PTR %RGp failed with %Rrc\n", GCPhys, rc));
1738	Assert(!pPTDst->a[(GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK].n.u1Present);
1739	}
1740	}
1741	else
1742	{
1743	/*
1744	* 4/2MB page - lazy syncing shadow 4K pages.
1745	* (There are many causes of getting here, it's no longer only CSAM.)
1746	*/
1747	/* Calculate the GC physical address of this 4KB shadow page. */
1748	RTGCPHYS GCPhys = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc) \| ((RTGCUINTPTR)GCPtrPage & GST_BIG_PAGE_OFFSET_MASK);
1749	/* Find ram range. */
1750	PPGMPAGE pPage;
1751	int rc = pgmPhysGetPageEx(&pVM->pgm.s, GCPhys, &pPage);
1752	if (RT_SUCCESS(rc))
1753	{
1754	/*
1755	* Make shadow PTE entry.
1756	*/
1757	const RTHCPHYS HCPhys = pPage->HCPhys; /** @todo PAGE FLAGS */
1758	SHWPTE PteDst;
1759	PteDst.u = (PdeSrc.u & ~(X86_PTE_PAE_PG_MASK \| X86_PTE_AVL_MASK \| X86_PTE_PAT \| X86_PTE_PCD \| X86_PTE_PWT))
1760	\| (HCPhys & X86_PTE_PAE_PG_MASK);
1761	if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
1762	{
1763	if (!PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage))
1764	PteDst.n.u1Write = 0;
1765	else
1766	PteDst.u = 0;
1767	}
1768	const unsigned iPTDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1769	# ifdef PGMPOOL_WITH_USER_TRACKING
1770	if (PteDst.n.u1Present && !pPTDst->a[iPTDst].n.u1Present)
1771	PGM_BTH_NAME(SyncPageWorkerTrackAddref)(pVM, pShwPage, HCPhys >> MM_RAM_FLAGS_IDX_SHIFT, pPage, iPTDst);
1772	# endif
1773	pPTDst->a[iPTDst] = PteDst;
1774
1775
1776	/*
1777	* If the page is not flagged as dirty and is writable, then make it read-only
1778	* at PD level, so we can set the dirty bit when the page is modified.
1779	*
1780	* ASSUMES that page access handlers are implemented on page table entry level.
1781	* Thus we will first catch the dirty access and set PDE.D and restart. If
1782	* there is an access handler, we'll trap again and let it work on the problem.
1783	*/
1784	/** @todo r=bird: figure out why we need this here, SyncPT should've taken care of this already.
1785	* As for invlpg, it simply frees the whole shadow PT.
1786	* ...It's possibly because the guest clears it and the guest doesn't really tell us... */
1787	if (!PdeSrc.b.u1Dirty && PdeSrc.b.u1Write)
1788	{
1789	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPageBig));
1790	PdeDst.u \|= PGM_PDFLAGS_TRACK_DIRTY;
1791	PdeDst.n.u1Write = 0;
1792	}
1793	else
1794	{
1795	PdeDst.au32[0] &= ~PGM_PDFLAGS_TRACK_DIRTY;
1796	PdeDst.n.u1Write = PdeSrc.n.u1Write;
1797	}
1798	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
1799	pVM->pgm.s.CTXMID(p,32BitPD)->a[iPDDst] = PdeDst;
1800	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
1801	pVM->pgm.s.CTXMID(ap,PaePDs)[0]->a[iPDDst] = PdeDst;
1802	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
1803	pPDDst->a[iPDDst] = PdeDst;
1804	# endif
1805	Log2(("SyncPage: BIG %RGv PdeSrc:{P=%d RW=%d U=%d raw=%08llx} GCPhys=%RGp%s\n",
1806	GCPtrPage, PdeSrc.n.u1Present, PdeSrc.n.u1Write, PdeSrc.n.u1User, (uint64_t)PdeSrc.u, GCPhys,
1807	PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
1808	}
1809	else
1810	LogFlow(("PGM_GCPHYS_2_PTR %RGp (big) failed with %Rrc\n", GCPhys, rc));
1811	}
1812	return VINF_SUCCESS;
1813	}
1814	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPagePDNAs));
1815	}
1816	else
1817	{
1818	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPagePDOutOfSync));
1819	Log2(("SyncPage: Out-Of-Sync PDE at %RGp PdeSrc=%RX64 PdeDst=%RX64\n",
1820	GCPtrPage, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
1821	}
1822
1823	/*
1824	* Mark the PDE not present. Restart the instruction and let #PF call SyncPT.
1825	* Yea, I'm lazy.
1826	*/
1827	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
1828	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1829	pgmPoolFreeByPage(pPool, pShwPage, pShwPde->idx, iPDDst);
1830	# else
1831	pgmPoolFreeByPage(pPool, pShwPage, SHW_POOL_ROOT_IDX, iPDDst);
1832	# endif
1833
1834	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
1835	pVM->pgm.s.CTXMID(p,32BitPD)->a[iPDDst].u = 0;
1836	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
1837	pVM->pgm.s.CTXMID(ap,PaePDs)[0]->a[iPDDst].u = 0;
1838	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
1839	pPDDst->a[iPDDst].u = 0;
1840	# endif
1841	PGM_INVL_GUEST_TLBS();
1842	return VINF_PGM_SYNCPAGE_MODIFIED_PDE;
1843
1844	#elif (PGM_GST_TYPE == PGM_TYPE_REAL \|\| PGM_GST_TYPE == PGM_TYPE_PROT) \
1845	&& PGM_SHW_TYPE != PGM_TYPE_NESTED \
1846	&& (PGM_SHW_TYPE != PGM_TYPE_EPT \|\| PGM_GST_TYPE == PGM_TYPE_PROT)
1847
1848	# ifdef PGM_SYNC_N_PAGES
1849	/*
1850	* Get the shadow PDE, find the shadow page table in the pool.
1851	*/
1852	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
1853	const unsigned iPDDst = GCPtrPage >> SHW_PD_SHIFT;
1854	X86PDE PdeDst = pVM->pgm.s.CTXMID(p,32BitPD)->a[iPDDst];
1855	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
1856	const unsigned iPDDst = GCPtrPage >> SHW_PD_SHIFT; /* no mask; flat index into the 2048 entry array. */
1857	X86PDEPAE PdeDst = pVM->pgm.s.CTXMID(ap,PaePDs)[0]->a[iPDDst];
1858	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
1859	const unsigned iPDDst = ((GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK);
1860	const unsigned iPdpte = (GCPtrPage >> X86_PDPT_SHIFT) & X86_PDPT_MASK_AMD64; NOREF(iPdpte);
1861	PX86PDPAE pPDDst;
1862	X86PDEPAE PdeDst;
1863	PX86PDPT pPdptDst;
1864
1865	int rc = PGMShwGetLongModePDPtr(pVM, GCPtrPage, &pPdptDst, &pPDDst);
1866	AssertRCSuccessReturn(rc, rc);
1867	Assert(pPDDst && pPdptDst);
1868	PdeDst = pPDDst->a[iPDDst];
1869	# elif PGM_SHW_TYPE == PGM_TYPE_EPT
1870	const unsigned iPDDst = ((GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK);
1871	PEPTPD pPDDst;
1872	EPTPDE PdeDst;
1873
1874	int rc = PGMShwGetEPTPDPtr(pVM, GCPtrPage, NULL, &pPDDst);
1875	if (rc != VINF_SUCCESS)
1876	{
1877	AssertRC(rc);
1878	return rc;
1879	}
1880	Assert(pPDDst);
1881	PdeDst = pPDDst->a[iPDDst];
1882	# endif
1883	Assert(PdeDst.n.u1Present);
1884	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, PdeDst.u & SHW_PDE_PG_MASK);
1885	PSHWPT pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
1886
1887	Assert(cPages == 1 \|\| !(uErr & X86_TRAP_PF_P));
1888	if (cPages > 1 && !(uErr & X86_TRAP_PF_P))
1889	{
1890	/*
1891	* This code path is currently only taken when the caller is PGMTrap0eHandler
1892	* for non-present pages!
1893	*
1894	* We're setting PGM_SYNC_NR_PAGES pages around the faulting page to sync it and
1895	* deal with locality.
1896	*/
1897	unsigned iPTDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1898	const unsigned iPTDstEnd = RT_MIN(iPTDst + PGM_SYNC_NR_PAGES / 2, RT_ELEMENTS(pPTDst->a));
1899	if (iPTDst < PGM_SYNC_NR_PAGES / 2)
1900	iPTDst = 0;
1901	else
1902	iPTDst -= PGM_SYNC_NR_PAGES / 2;
1903	for (; iPTDst < iPTDstEnd; iPTDst++)
1904	{
1905	if (!pPTDst->a[iPTDst].n.u1Present)
1906	{
1907	GSTPTE PteSrc;
1908
1909	RTGCUINTPTR GCPtrCurPage = ((RTGCUINTPTR)GCPtrPage & ~(RTGCUINTPTR)(SHW_PT_MASK << SHW_PT_SHIFT)) \| (iPTDst << PAGE_SHIFT);
1910
1911	/* Fake the page table entry */
1912	PteSrc.u = GCPtrCurPage;
1913	PteSrc.n.u1Present = 1;
1914	PteSrc.n.u1Dirty = 1;
1915	PteSrc.n.u1Accessed = 1;
1916	PteSrc.n.u1Write = 1;
1917	PteSrc.n.u1User = 1;
1918
1919	PGM_BTH_NAME(SyncPageWorker)(pVM, &pPTDst->a[iPTDst], PdeSrc, PteSrc, pShwPage, iPTDst);
1920
1921	Log2(("SyncPage: 4K+ %RGv PteSrc:{P=%d RW=%d U=%d raw=%08llx} PteDst=%08llx%s\n",
1922	GCPtrCurPage, PteSrc.n.u1Present,
1923	PteSrc.n.u1Write & PdeSrc.n.u1Write,
1924	PteSrc.n.u1User & PdeSrc.n.u1User,
1925	(uint64_t)PteSrc.u,
1926	(uint64_t)pPTDst->a[iPTDst].u,
1927	pPTDst->a[iPTDst].u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
1928	}
1929	else
1930	Log4(("%RGv iPTDst=%x pPTDst->a[iPTDst] %RX64\n", ((RTGCUINTPTR)GCPtrPage & ~(RTGCUINTPTR)(SHW_PT_MASK << SHW_PT_SHIFT)) \| (iPTDst << PAGE_SHIFT), iPTDst, pPTDst->a[iPTDst].u));
1931	}
1932	}
1933	else
1934	# endif /* PGM_SYNC_N_PAGES */
1935	{
1936	GSTPTE PteSrc;
1937	const unsigned iPTDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1938	RTGCUINTPTR GCPtrCurPage = ((RTGCUINTPTR)GCPtrPage & ~(RTGCUINTPTR)(SHW_PT_MASK << SHW_PT_SHIFT)) \| (iPTDst << PAGE_SHIFT);
1939
1940	/* Fake the page table entry */
1941	PteSrc.u = GCPtrCurPage;
1942	PteSrc.n.u1Present = 1;
1943	PteSrc.n.u1Dirty = 1;
1944	PteSrc.n.u1Accessed = 1;
1945	PteSrc.n.u1Write = 1;
1946	PteSrc.n.u1User = 1;
1947	PGM_BTH_NAME(SyncPageWorker)(pVM, &pPTDst->a[iPTDst], PdeSrc, PteSrc, pShwPage, iPTDst);
1948
1949	Log2(("SyncPage: 4K %RGv PteSrc:{P=%d RW=%d U=%d raw=%08llx}PteDst=%08llx%s\n",
1950	GCPtrPage, PteSrc.n.u1Present,
1951	PteSrc.n.u1Write & PdeSrc.n.u1Write,
1952	PteSrc.n.u1User & PdeSrc.n.u1User,
1953	(uint64_t)PteSrc.u,
1954	(uint64_t)pPTDst->a[iPTDst].u,
1955	pPTDst->a[iPTDst].u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
1956	}
1957	return VINF_SUCCESS;
1958
1959	#else
1960	AssertReleaseMsgFailed(("Shw=%d Gst=%d is not implemented!\n", PGM_GST_TYPE, PGM_SHW_TYPE));
1961	return VERR_INTERNAL_ERROR;
1962	#endif
1963	}
1964
1965
1966	#if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
1967	/**
1968	* Investigate page fault and handle write protection page faults caused by
1969	* dirty bit tracking.
1970	*
1971	* @returns VBox status code.
1972	* @param pVM VM handle.
1973	* @param uErr Page fault error code.
1974	* @param pPdeDst Shadow page directory entry.
1975	* @param pPdeSrc Guest page directory entry.
1976	* @param GCPtrPage Guest context page address.
1977	*/
1978	PGM_BTH_DECL(int, CheckPageFault)(PVM pVM, uint32_t uErr, PSHWPDE pPdeDst, PGSTPDE pPdeSrc, RTGCUINTPTR GCPtrPage)
1979	{
1980	bool fWriteProtect = !!(CPUMGetGuestCR0(pVM) & X86_CR0_WP);
1981	bool fUserLevelFault = !!(uErr & X86_TRAP_PF_US);
1982	bool fWriteFault = !!(uErr & X86_TRAP_PF_RW);
1983	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1984	bool fBigPagesSupported = true;
1985	# else
1986	bool fBigPagesSupported = !!(CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
1987	# endif
1988	# if PGM_WITH_NX(PGM_GST_TYPE, PGM_SHW_TYPE)
1989	bool fNoExecuteBitValid = !!(CPUMGetGuestEFER(pVM) & MSR_K6_EFER_NXE);
1990	# endif
1991	unsigned uPageFaultLevel;
1992	int rc;
1993
1994	STAM_PROFILE_START(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
1995	LogFlow(("CheckPageFault: GCPtrPage=%RGv uErr=%#x PdeSrc=%08x\n", GCPtrPage, uErr, pPdeSrc->u));
1996
1997	# if PGM_GST_TYPE == PGM_TYPE_PAE \
1998	\|\| PGM_GST_TYPE == PGM_TYPE_AMD64
1999
2000	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2001	PX86PML4E pPml4eSrc;
2002	PX86PDPE pPdpeSrc;
2003
2004	pPdpeSrc = pgmGstGetLongModePDPTPtr(&pVM->pgm.s, GCPtrPage, &pPml4eSrc);
2005	Assert(pPml4eSrc);
2006
2007	/*
2008	* Real page fault? (PML4E level)
2009	*/
2010	if ( (uErr & X86_TRAP_PF_RSVD)
2011	\|\| !pPml4eSrc->n.u1Present
2012	\|\| (fNoExecuteBitValid && (uErr & X86_TRAP_PF_ID) && pPml4eSrc->n.u1NoExecute)
2013	\|\| (fWriteFault && !pPml4eSrc->n.u1Write && (fUserLevelFault \|\| fWriteProtect))
2014	\|\| (fUserLevelFault && !pPml4eSrc->n.u1User)
2015	)
2016	{
2017	uPageFaultLevel = 0;
2018	goto l_UpperLevelPageFault;
2019	}
2020	Assert(pPdpeSrc);
2021
2022	# else /* PAE */
2023	PX86PDPE pPdpeSrc = &pVM->pgm.s.CTXSUFF(pGstPaePDPT)->a[(GCPtrPage >> GST_PDPT_SHIFT) & GST_PDPT_MASK];
2024	# endif
2025
2026	/*
2027	* Real page fault? (PDPE level)
2028	*/
2029	if ( (uErr & X86_TRAP_PF_RSVD)
2030	\|\| !pPdpeSrc->n.u1Present
2031	# if PGM_GST_TYPE == PGM_TYPE_AMD64 /* NX, r/w, u/s bits in the PDPE are long mode only */
2032	\|\| (fNoExecuteBitValid && (uErr & X86_TRAP_PF_ID) && pPdpeSrc->lm.u1NoExecute)
2033	\|\| (fWriteFault && !pPdpeSrc->lm.u1Write && (fUserLevelFault \|\| fWriteProtect))
2034	\|\| (fUserLevelFault && !pPdpeSrc->lm.u1User)
2035	# endif
2036	)
2037	{
2038	uPageFaultLevel = 1;
2039	goto l_UpperLevelPageFault;
2040	}
2041	# endif
2042
2043	/*
2044	* Real page fault? (PDE level)
2045	*/
2046	if ( (uErr & X86_TRAP_PF_RSVD)
2047	\|\| !pPdeSrc->n.u1Present
2048	# if PGM_WITH_NX(PGM_GST_TYPE, PGM_SHW_TYPE)
2049	\|\| (fNoExecuteBitValid && (uErr & X86_TRAP_PF_ID) && pPdeSrc->n.u1NoExecute)
2050	# endif
2051	\|\| (fWriteFault && !pPdeSrc->n.u1Write && (fUserLevelFault \|\| fWriteProtect))
2052	\|\| (fUserLevelFault && !pPdeSrc->n.u1User) )
2053	{
2054	uPageFaultLevel = 2;
2055	goto l_UpperLevelPageFault;
2056	}
2057
2058	/*
2059	* First check the easy case where the page directory has been marked read-only to track
2060	* the dirty bit of an emulated BIG page
2061	*/
2062	if (pPdeSrc->b.u1Size && fBigPagesSupported)
2063	{
2064	/* Mark guest page directory as accessed */
2065	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2066	pPml4eSrc->n.u1Accessed = 1;
2067	pPdpeSrc->lm.u1Accessed = 1;
2068	# endif
2069	pPdeSrc->b.u1Accessed = 1;
2070
2071	/*
2072	* Only write protection page faults are relevant here.
2073	*/
2074	if (fWriteFault)
2075	{
2076	/* Mark guest page directory as dirty (BIG page only). */
2077	pPdeSrc->b.u1Dirty = 1;
2078
2079	if (pPdeDst->n.u1Present && (pPdeDst->u & PGM_PDFLAGS_TRACK_DIRTY))
2080	{
2081	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPageTrap));
2082
2083	Assert(pPdeSrc->b.u1Write);
2084
2085	pPdeDst->n.u1Write = 1;
2086	pPdeDst->n.u1Accessed = 1;
2087	pPdeDst->au32[0] &= ~PGM_PDFLAGS_TRACK_DIRTY;
2088	PGM_INVL_BIG_PG(GCPtrPage);
2089	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2090	return VINF_PGM_HANDLED_DIRTY_BIT_FAULT;
2091	}
2092	}
2093	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2094	return VINF_PGM_NO_DIRTY_BIT_TRACKING;
2095	}
2096	/* else: 4KB page table */
2097
2098	/*
2099	* Map the guest page table.
2100	*/
2101	PGSTPT pPTSrc;
2102	rc = PGM_GCPHYS_2_PTR(pVM, pPdeSrc->u & GST_PDE_PG_MASK, &pPTSrc);
2103	if (RT_SUCCESS(rc))
2104	{
2105	/*
2106	* Real page fault?
2107	*/
2108	PGSTPTE pPteSrc = &pPTSrc->a[(GCPtrPage >> GST_PT_SHIFT) & GST_PT_MASK];
2109	const GSTPTE PteSrc = *pPteSrc;
2110	if ( !PteSrc.n.u1Present
2111	# if PGM_WITH_NX(PGM_GST_TYPE, PGM_SHW_TYPE)
2112	\|\| (fNoExecuteBitValid && (uErr & X86_TRAP_PF_ID) && PteSrc.n.u1NoExecute)
2113	# endif
2114	\|\| (fWriteFault && !PteSrc.n.u1Write && (fUserLevelFault \|\| fWriteProtect))
2115	\|\| (fUserLevelFault && !PteSrc.n.u1User)
2116	)
2117	{
2118	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyTrackRealPF));
2119	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2120	LogFlow(("CheckPageFault: real page fault at %RGv PteSrc.u=%08x (2)\n", GCPtrPage, PteSrc.u));
2121
2122	/* Check the present bit as the shadow tables can cause different error codes by being out of sync.
2123	* See the 2nd case above as well.
2124	*/
2125	if (pPdeSrc->n.u1Present && pPteSrc->n.u1Present)
2126	TRPMSetErrorCode(pVM, uErr \| X86_TRAP_PF_P); /* page-level protection violation */
2127
2128	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2129	return VINF_EM_RAW_GUEST_TRAP;
2130	}
2131	LogFlow(("CheckPageFault: page fault at %RGv PteSrc.u=%08x\n", GCPtrPage, PteSrc.u));
2132
2133	/*
2134	* Set the accessed bits in the page directory and the page table.
2135	*/
2136	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2137	pPml4eSrc->n.u1Accessed = 1;
2138	pPdpeSrc->lm.u1Accessed = 1;
2139	# endif
2140	pPdeSrc->n.u1Accessed = 1;
2141	pPteSrc->n.u1Accessed = 1;
2142
2143	/*
2144	* Only write protection page faults are relevant here.
2145	*/
2146	if (fWriteFault)
2147	{
2148	/* Write access, so mark guest entry as dirty. */
2149	# ifdef VBOX_WITH_STATISTICS
2150	if (!pPteSrc->n.u1Dirty)
2151	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtiedPage));
2152	else
2153	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,PageAlreadyDirty));
2154	# endif
2155
2156	pPteSrc->n.u1Dirty = 1;
2157
2158	if (pPdeDst->n.u1Present)
2159	{
2160	#ifndef IN_RING0
2161	/* Bail out here as pgmPoolGetPageByHCPhys will return NULL and we'll crash below.
2162	* Our individual shadow handlers will provide more information and force a fatal exit.
2163	*/
2164	if (MMHyperIsInsideArea(pVM, (RTGCPTR)GCPtrPage))
2165	{
2166	LogRel(("CheckPageFault: write to hypervisor region %RGv\n", GCPtrPage));
2167	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2168	return VINF_SUCCESS;
2169	}
2170	#endif
2171	/*
2172	* Map shadow page table.
2173	*/
2174	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, pPdeDst->u & SHW_PDE_PG_MASK);
2175	if (pShwPage)
2176	{
2177	PSHWPT pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
2178	PSHWPTE pPteDst = &pPTDst->a[(GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK];
2179	if ( pPteDst->n.u1Present /** @todo Optimize accessed bit emulation? */
2180	&& (pPteDst->u & PGM_PTFLAGS_TRACK_DIRTY))
2181	{
2182	LogFlow(("DIRTY page trap addr=%RGv\n", GCPtrPage));
2183	# ifdef VBOX_STRICT
2184	PPGMPAGE pPage = pgmPhysGetPage(&pVM->pgm.s, pPteSrc->u & GST_PTE_PG_MASK);
2185	if (pPage)
2186	AssertMsg(!PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage),
2187	("Unexpected dirty bit tracking on monitored page %RGv (phys %RGp)!!!!!!\n", GCPtrPage, pPteSrc->u & X86_PTE_PAE_PG_MASK));
2188	# endif
2189	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPageTrap));
2190
2191	Assert(pPteSrc->n.u1Write);
2192
2193	pPteDst->n.u1Write = 1;
2194	pPteDst->n.u1Dirty = 1;
2195	pPteDst->n.u1Accessed = 1;
2196	pPteDst->au32[0] &= ~PGM_PTFLAGS_TRACK_DIRTY;
2197	PGM_INVL_PG(GCPtrPage);
2198
2199	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2200	return VINF_PGM_HANDLED_DIRTY_BIT_FAULT;
2201	}
2202	}
2203	else
2204	AssertMsgFailed(("pgmPoolGetPageByHCPhys %RGp failed!\n", pPdeDst->u & SHW_PDE_PG_MASK));
2205	}
2206	}
2207	/** @todo Optimize accessed bit emulation? */
2208	# ifdef VBOX_STRICT
2209	/*
2210	* Sanity check.
2211	*/
2212	else if ( !pPteSrc->n.u1Dirty
2213	&& (pPdeSrc->n.u1Write & pPteSrc->n.u1Write)
2214	&& pPdeDst->n.u1Present)
2215	{
2216	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, pPdeDst->u & SHW_PDE_PG_MASK);
2217	PSHWPT pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
2218	PSHWPTE pPteDst = &pPTDst->a[(GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK];
2219	if ( pPteDst->n.u1Present
2220	&& pPteDst->n.u1Write)
2221	LogFlow(("Writable present page %RGv not marked for dirty bit tracking!!!\n", GCPtrPage));
2222	}
2223	# endif /* VBOX_STRICT */
2224	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2225	return VINF_PGM_NO_DIRTY_BIT_TRACKING;
2226	}
2227	AssertRC(rc);
2228	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2229	return rc;
2230
2231
2232	l_UpperLevelPageFault:
2233	/*
2234	* Pagefault detected while checking the PML4E, PDPE or PDE.
2235	* Single exit handler to get rid of duplicate code paths.
2236	*/
2237	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyTrackRealPF));
2238	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2239	Log(("CheckPageFault: real page fault at %RGv (%d)\n", GCPtrPage, uPageFaultLevel));
2240
2241	if (
2242	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2243	pPml4eSrc->n.u1Present &&
2244	# endif
2245	# if PGM_GST_TYPE == PGM_TYPE_AMD64 \|\| PGM_GST_TYPE == PGM_TYPE_PAE
2246	pPdpeSrc->n.u1Present &&
2247	# endif
2248	pPdeSrc->n.u1Present)
2249	{
2250	/* Check the present bit as the shadow tables can cause different error codes by being out of sync. */
2251	if (pPdeSrc->b.u1Size && fBigPagesSupported)
2252	{
2253	TRPMSetErrorCode(pVM, uErr \| X86_TRAP_PF_P); /* page-level protection violation */
2254	}
2255	else
2256	{
2257	/*
2258	* Map the guest page table.
2259	*/
2260	PGSTPT pPTSrc;
2261	rc = PGM_GCPHYS_2_PTR(pVM, pPdeSrc->u & GST_PDE_PG_MASK, &pPTSrc);
2262	if (RT_SUCCESS(rc))
2263	{
2264	PGSTPTE pPteSrc = &pPTSrc->a[(GCPtrPage >> GST_PT_SHIFT) & GST_PT_MASK];
2265	const GSTPTE PteSrc = *pPteSrc;
2266	if (pPteSrc->n.u1Present)
2267	TRPMSetErrorCode(pVM, uErr \| X86_TRAP_PF_P); /* page-level protection violation */
2268	}
2269	AssertRC(rc);
2270	}
2271	}
2272	return VINF_EM_RAW_GUEST_TRAP;
2273	}
2274	#endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
2275
2276
2277	/**
2278	* Sync a shadow page table.
2279	*
2280	* The shadow page table is not present. This includes the case where
2281	* there is a conflict with a mapping.
2282	*
2283	* @returns VBox status code.
2284	* @param pVM VM handle.
2285	* @param iPD Page directory index.
2286	* @param pPDSrc Source page directory (i.e. Guest OS page directory).
2287	* Assume this is a temporary mapping.
2288	* @param GCPtrPage GC Pointer of the page that caused the fault
2289	*/
2290	PGM_BTH_DECL(int, SyncPT)(PVM pVM, unsigned iPDSrc, PGSTPD pPDSrc, RTGCUINTPTR GCPtrPage)
2291	{
2292	STAM_PROFILE_START(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT), a);
2293	STAM_COUNTER_INC(&pVM->pgm.s.StatSyncPtPD[iPDSrc]);
2294	LogFlow(("SyncPT: GCPtrPage=%RGv\n", GCPtrPage));
2295
2296	#if ( PGM_GST_TYPE == PGM_TYPE_32BIT \
2297	\|\| PGM_GST_TYPE == PGM_TYPE_PAE \
2298	\|\| PGM_GST_TYPE == PGM_TYPE_AMD64) \
2299	&& PGM_SHW_TYPE != PGM_TYPE_NESTED \
2300	&& PGM_SHW_TYPE != PGM_TYPE_EPT
2301
2302	int rc = VINF_SUCCESS;
2303
2304	/*
2305	* Validate input a little bit.
2306	*/
2307	AssertMsg(iPDSrc == ((GCPtrPage >> GST_PD_SHIFT) & GST_PD_MASK), ("iPDSrc=%x GCPtrPage=%RGv\n", iPDSrc, GCPtrPage));
2308	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
2309	const unsigned iPDDst = GCPtrPage >> SHW_PD_SHIFT;
2310	PX86PD pPDDst = pVM->pgm.s.CTXMID(p,32BitPD);
2311	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
2312	const unsigned iPDDst = GCPtrPage >> SHW_PD_SHIFT; /* no mask; flat index into the 2048 entry array. */
2313	const unsigned iPdpte = (GCPtrPage >> X86_PDPT_SHIFT); NOREF(iPdpte);
2314	PX86PDPT pPdptDst = pVM->pgm.s.CTXMID(p,PaePDPT); NOREF(pPdptDst);
2315	PX86PDPAE pPDDst = pVM->pgm.s.CTXMID(ap,PaePDs)[0];
2316	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
2317	const unsigned iPdpte = (GCPtrPage >> X86_PDPT_SHIFT) & X86_PDPT_MASK_AMD64;
2318	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
2319	PX86PDPAE pPDDst;
2320	PX86PDPT pPdptDst;
2321	rc = PGMShwGetLongModePDPtr(pVM, GCPtrPage, &pPdptDst, &pPDDst);
2322	AssertRCSuccessReturn(rc, rc);
2323	Assert(pPDDst);
2324	# endif
2325
2326	PSHWPDE pPdeDst = &pPDDst->a[iPDDst];
2327	SHWPDE PdeDst = *pPdeDst;
2328
2329	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2330	/* Fetch the pgm pool shadow descriptor. */
2331	PPGMPOOLPAGE pShwPde = pgmPoolGetPageByHCPhys(pVM, pPdptDst->a[iPdpte].u & X86_PDPE_PG_MASK);
2332	Assert(pShwPde);
2333	# endif
2334
2335	# ifndef PGM_WITHOUT_MAPPINGS
2336	/*
2337	* Check for conflicts.
2338	* GC: In case of a conflict we'll go to Ring-3 and do a full SyncCR3.
2339	* HC: Simply resolve the conflict.
2340	*/
2341	if (PdeDst.u & PGM_PDFLAGS_MAPPING)
2342	{
2343	Assert(pgmMapAreMappingsEnabled(&pVM->pgm.s));
2344	# ifndef IN_RING3
2345	Log(("SyncPT: Conflict at %RGv\n", GCPtrPage));
2346	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT), a);
2347	return VERR_ADDRESS_CONFLICT;
2348	# else
2349	PPGMMAPPING pMapping = pgmGetMapping(pVM, (RTGCPTR)GCPtrPage);
2350	Assert(pMapping);
2351	# if PGM_GST_TYPE == PGM_TYPE_32BIT
2352	int rc = pgmR3SyncPTResolveConflict(pVM, pMapping, pPDSrc, GCPtrPage & (GST_PD_MASK << GST_PD_SHIFT));
2353	# elif PGM_GST_TYPE == PGM_TYPE_PAE
2354	int rc = pgmR3SyncPTResolveConflictPAE(pVM, pMapping, GCPtrPage & (GST_PD_MASK << GST_PD_SHIFT));
2355	# else
2356	AssertFailed(); /* can't happen for amd64 */
2357	# endif
2358	if (RT_FAILURE(rc))
2359	{
2360	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT), a);
2361	return rc;
2362	}
2363	PdeDst = *pPdeDst;
2364	# endif
2365	}
2366	# else /* PGM_WITHOUT_MAPPINGS */
2367	Assert(!pgmMapAreMappingsEnabled(&pVM->pgm.s));
2368	# endif /* PGM_WITHOUT_MAPPINGS */
2369	Assert(!PdeDst.n.u1Present); /* We're only supposed to call SyncPT on PDE!P and conflicts.*/
2370
2371	/*
2372	* Sync page directory entry.
2373	*/
2374	GSTPDE PdeSrc = pPDSrc->a[iPDSrc];
2375	if (PdeSrc.n.u1Present)
2376	{
2377	/*
2378	* Allocate & map the page table.
2379	*/
2380	PSHWPT pPTDst;
2381	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2382	const bool fPageTable = !PdeSrc.b.u1Size;
2383	# else
2384	const bool fPageTable = !PdeSrc.b.u1Size \|\| !(CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
2385	# endif
2386	PPGMPOOLPAGE pShwPage;
2387	RTGCPHYS GCPhys;
2388	if (fPageTable)
2389	{
2390	GCPhys = PdeSrc.u & GST_PDE_PG_MASK;
2391	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
2392	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
2393	GCPhys \|= (iPDDst & 1) * (PAGE_SIZE / 2);
2394	# endif
2395	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2396	rc = pgmPoolAlloc(pVM, GCPhys, BTH_PGMPOOLKIND_PT_FOR_PT, pShwPde->idx, iPDDst, &pShwPage);
2397	# else
2398	rc = pgmPoolAlloc(pVM, GCPhys, BTH_PGMPOOLKIND_PT_FOR_PT, SHW_POOL_ROOT_IDX, iPDDst, &pShwPage);
2399	# endif
2400	}
2401	else
2402	{
2403	GCPhys = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc);
2404	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
2405	/* Select the right PDE as we're emulating a 4MB page directory with two 2 MB shadow PDEs.*/
2406	GCPhys \|= GCPtrPage & (1 << X86_PD_PAE_SHIFT);
2407	# endif
2408	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2409	rc = pgmPoolAlloc(pVM, GCPhys, BTH_PGMPOOLKIND_PT_FOR_BIG, pShwPde->idx, iPDDst, &pShwPage);
2410	# else
2411	rc = pgmPoolAlloc(pVM, GCPhys, BTH_PGMPOOLKIND_PT_FOR_BIG, SHW_POOL_ROOT_IDX, iPDDst, &pShwPage);
2412	# endif
2413	}
2414	if (rc == VINF_SUCCESS)
2415	pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
2416	else if (rc == VINF_PGM_CACHED_PAGE)
2417	{
2418	/*
2419	* The PT was cached, just hook it up.
2420	*/
2421	if (fPageTable)
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	else
2425	{
2426	PdeDst.u = pShwPage->Core.Key
2427	\| (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));
2428	/* (see explanation and assumptions further down.) */
2429	if (!PdeSrc.b.u1Dirty && PdeSrc.b.u1Write)
2430	{
2431	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPageBig));
2432	PdeDst.u \|= PGM_PDFLAGS_TRACK_DIRTY;
2433	PdeDst.b.u1Write = 0;
2434	}
2435	}
2436	*pPdeDst = PdeDst;
2437	return VINF_SUCCESS;
2438	}
2439	else if (rc == VERR_PGM_POOL_FLUSHED)
2440	{
2441	VM_FF_SET(pVM, VM_FF_PGM_SYNC_CR3);
2442	return VINF_PGM_SYNC_CR3;
2443	}
2444	else
2445	AssertMsgFailedReturn(("rc=%Rrc\n", rc), VERR_INTERNAL_ERROR);
2446	PdeDst.u &= X86_PDE_AVL_MASK;
2447	PdeDst.u \|= pShwPage->Core.Key;
2448
2449	/*
2450	* Page directory has been accessed (this is a fault situation, remember).
2451	*/
2452	pPDSrc->a[iPDSrc].n.u1Accessed = 1;
2453	if (fPageTable)
2454	{
2455	/*
2456	* Page table - 4KB.
2457	*
2458	* Sync all or just a few entries depending on PGM_SYNC_N_PAGES.
2459	*/
2460	Log2(("SyncPT: 4K %RGv PdeSrc:{P=%d RW=%d U=%d raw=%08llx}\n",
2461	GCPtrPage, PdeSrc.b.u1Present, PdeSrc.b.u1Write, PdeSrc.b.u1User, (uint64_t)PdeSrc.u));
2462	PGSTPT pPTSrc;
2463	rc = PGM_GCPHYS_2_PTR(pVM, PdeSrc.u & GST_PDE_PG_MASK, &pPTSrc);
2464	if (RT_SUCCESS(rc))
2465	{
2466	/*
2467	* Start by syncing the page directory entry so CSAM's TLB trick works.
2468	*/
2469	PdeDst.u = (PdeDst.u & (SHW_PDE_PG_MASK \| X86_PDE_AVL_MASK))
2470	\| (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));
2471	*pPdeDst = PdeDst;
2472
2473	/*
2474	* Directory/page user or supervisor privilege: (same goes for read/write)
2475	*
2476	* Directory Page Combined
2477	* U/S U/S U/S
2478	* 0 0 0
2479	* 0 1 0
2480	* 1 0 0
2481	* 1 1 1
2482	*
2483	* Simple AND operation. Table listed for completeness.
2484	*
2485	*/
2486	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT4K));
2487	# ifdef PGM_SYNC_N_PAGES
2488	unsigned iPTBase = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
2489	unsigned iPTDst = iPTBase;
2490	const unsigned iPTDstEnd = RT_MIN(iPTDst + PGM_SYNC_NR_PAGES / 2, RT_ELEMENTS(pPTDst->a));
2491	if (iPTDst <= PGM_SYNC_NR_PAGES / 2)
2492	iPTDst = 0;
2493	else
2494	iPTDst -= PGM_SYNC_NR_PAGES / 2;
2495	# else /* !PGM_SYNC_N_PAGES */
2496	unsigned iPTDst = 0;
2497	const unsigned iPTDstEnd = RT_ELEMENTS(pPTDst->a);
2498	# endif /* !PGM_SYNC_N_PAGES */
2499	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
2500	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
2501	const unsigned offPTSrc = ((GCPtrPage >> SHW_PD_SHIFT) & 1) * 512;
2502	# else
2503	const unsigned offPTSrc = 0;
2504	# endif
2505	for (; iPTDst < iPTDstEnd; iPTDst++)
2506	{
2507	const unsigned iPTSrc = iPTDst + offPTSrc;
2508	const GSTPTE PteSrc = pPTSrc->a[iPTSrc];
2509
2510	if (PteSrc.n.u1Present) /* we've already cleared it above */
2511	{
2512	# ifndef IN_RING0
2513	/*
2514	* Assuming kernel code will be marked as supervisor - and not as user level
2515	* and executed using a conforming code selector - And marked as readonly.
2516	* Also assume that if we're monitoring a page, it's of no interest to CSAM.
2517	*/
2518	PPGMPAGE pPage;
2519	if ( ((PdeSrc.u & pPTSrc->a[iPTSrc].u) & (X86_PTE_RW \| X86_PTE_US))
2520	\|\| !CSAMDoesPageNeedScanning(pVM, (RTRCPTR)((iPDSrc << GST_PD_SHIFT) \| (iPTSrc << PAGE_SHIFT)))
2521	\|\| ( (pPage = pgmPhysGetPage(&pVM->pgm.s, PteSrc.u & GST_PTE_PG_MASK))
2522	&& PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
2523	)
2524	# endif
2525	PGM_BTH_NAME(SyncPageWorker)(pVM, &pPTDst->a[iPTDst], PdeSrc, PteSrc, pShwPage, iPTDst);
2526	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",
2527	(RTGCPTR)((iPDSrc << GST_PD_SHIFT) \| (iPTSrc << PAGE_SHIFT)),
2528	PteSrc.n.u1Present,
2529	PteSrc.n.u1Write & PdeSrc.n.u1Write,
2530	PteSrc.n.u1User & PdeSrc.n.u1User,
2531	(uint64_t)PteSrc.u,
2532	pPTDst->a[iPTDst].u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : "", pPTDst->a[iPTDst].u, iPTSrc, PdeSrc.au32[0],
2533	(RTGCPHYS)((PdeSrc.u & GST_PDE_PG_MASK) + iPTSrc*sizeof(PteSrc)) ));
2534	}
2535	} /* for PTEs */
2536	}
2537	}
2538	else
2539	{
2540	/*
2541	* Big page - 2/4MB.
2542	*
2543	* We'll walk the ram range list in parallel and optimize lookups.
2544	* We will only sync on shadow page table at a time.
2545	*/
2546	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT4M));
2547
2548	/**
2549	* @todo It might be more efficient to sync only a part of the 4MB page (similar to what we do for 4kb PDs).
2550	*/
2551
2552	/*
2553	* Start by syncing the page directory entry.
2554	*/
2555	PdeDst.u = (PdeDst.u & (SHW_PDE_PG_MASK \| (X86_PDE_AVL_MASK & ~PGM_PDFLAGS_TRACK_DIRTY)))
2556	\| (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));
2557
2558	/*
2559	* If the page is not flagged as dirty and is writable, then make it read-only
2560	* at PD level, so we can set the dirty bit when the page is modified.
2561	*
2562	* ASSUMES that page access handlers are implemented on page table entry level.
2563	* Thus we will first catch the dirty access and set PDE.D and restart. If
2564	* there is an access handler, we'll trap again and let it work on the problem.
2565	*/
2566	/** @todo move the above stuff to a section in the PGM documentation. */
2567	Assert(!(PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY));
2568	if (!PdeSrc.b.u1Dirty && PdeSrc.b.u1Write)
2569	{
2570	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPageBig));
2571	PdeDst.u \|= PGM_PDFLAGS_TRACK_DIRTY;
2572	PdeDst.b.u1Write = 0;
2573	}
2574	*pPdeDst = PdeDst;
2575
2576	/*
2577	* Fill the shadow page table.
2578	*/
2579	/* Get address and flags from the source PDE. */
2580	SHWPTE PteDstBase;
2581	PteDstBase.u = PdeSrc.u & ~(GST_PDE_PG_MASK \| X86_PTE_AVL_MASK \| X86_PTE_PAT \| X86_PTE_PCD \| X86_PTE_PWT);
2582
2583	/* Loop thru the entries in the shadow PT. */
2584	const RTGCUINTPTR GCPtr = (GCPtrPage >> SHW_PD_SHIFT) << SHW_PD_SHIFT; NOREF(GCPtr);
2585	Log2(("SyncPT: BIG %RGv PdeSrc:{P=%d RW=%d U=%d raw=%08llx} Shw=%RGv GCPhys=%RGp %s\n",
2586	GCPtrPage, PdeSrc.b.u1Present, PdeSrc.b.u1Write, PdeSrc.b.u1User, (uint64_t)PdeSrc.u, GCPtr,
2587	GCPhys, PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
2588	PPGMRAMRANGE pRam = pVM->pgm.s.CTX_SUFF(pRamRanges);
2589	unsigned iPTDst = 0;
2590	while (iPTDst < RT_ELEMENTS(pPTDst->a))
2591	{
2592	/* Advance ram range list. */
2593	while (pRam && GCPhys > pRam->GCPhysLast)
2594	pRam = pRam->CTX_SUFF(pNext);
2595	if (pRam && GCPhys >= pRam->GCPhys)
2596	{
2597	unsigned iHCPage = (GCPhys - pRam->GCPhys) >> PAGE_SHIFT;
2598	do
2599	{
2600	/* Make shadow PTE. */
2601	PPGMPAGE pPage = &pRam->aPages[iHCPage];
2602	SHWPTE PteDst;
2603
2604	/* Make sure the RAM has already been allocated. */
2605	if (pRam->fFlags & MM_RAM_FLAGS_DYNAMIC_ALLOC) /** @todo PAGE FLAGS */
2606	{
2607	if (RT_UNLIKELY(!PGM_PAGE_GET_HCPHYS(pPage)))
2608	{
2609	# ifdef IN_RING3
2610	int rc = pgmr3PhysGrowRange(pVM, GCPhys);
2611	# else
2612	int rc = CTXALLMID(VMM, CallHost)(pVM, VMMCALLHOST_PGM_RAM_GROW_RANGE, GCPhys);
2613	# endif
2614	if (rc != VINF_SUCCESS)
2615	return rc;
2616	}
2617	}
2618
2619	if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
2620	{
2621	if (!PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage))
2622	{
2623	PteDst.u = PGM_PAGE_GET_HCPHYS(pPage) \| PteDstBase.u;
2624	PteDst.n.u1Write = 0;
2625	}
2626	else
2627	PteDst.u = 0;
2628	}
2629	# ifndef IN_RING0
2630	/*
2631	* Assuming kernel code will be marked as supervisor and not as user level and executed
2632	* using a conforming code selector. Don't check for readonly, as that implies the whole
2633	* 4MB can be code or readonly data. Linux enables write access for its large pages.
2634	*/
2635	else if ( !PdeSrc.n.u1User
2636	&& CSAMDoesPageNeedScanning(pVM, (RTRCPTR)(GCPtr \| (iPTDst << SHW_PT_SHIFT))))
2637	PteDst.u = 0;
2638	# endif
2639	else
2640	PteDst.u = PGM_PAGE_GET_HCPHYS(pPage) \| PteDstBase.u;
2641	# ifdef PGMPOOL_WITH_USER_TRACKING
2642	if (PteDst.n.u1Present)
2643	PGM_BTH_NAME(SyncPageWorkerTrackAddref)(pVM, pShwPage, pPage->HCPhys >> MM_RAM_FLAGS_IDX_SHIFT, pPage, iPTDst); /** @todo PAGE FLAGS */
2644	# endif
2645	/* commit it */
2646	pPTDst->a[iPTDst] = PteDst;
2647	Log4(("SyncPT: BIG %RGv PteDst:{P=%d RW=%d U=%d raw=%08llx}%s\n",
2648	(RTGCPTR)(GCPtr \| (iPTDst << SHW_PT_SHIFT)), PteDst.n.u1Present, PteDst.n.u1Write, PteDst.n.u1User, (uint64_t)PteDst.u,
2649	PteDst.u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
2650
2651	/* advance */
2652	GCPhys += PAGE_SIZE;
2653	iHCPage++;
2654	iPTDst++;
2655	} while ( iPTDst < RT_ELEMENTS(pPTDst->a)
2656	&& GCPhys <= pRam->GCPhysLast);
2657	}
2658	else if (pRam)
2659	{
2660	Log(("Invalid pages at %RGp\n", GCPhys));
2661	do
2662	{
2663	pPTDst->a[iPTDst].u = 0; /* MMIO or invalid page, we must handle them manually. */
2664	GCPhys += PAGE_SIZE;
2665	iPTDst++;
2666	} while ( iPTDst < RT_ELEMENTS(pPTDst->a)
2667	&& GCPhys < pRam->GCPhys);
2668	}
2669	else
2670	{
2671	Log(("Invalid pages at %RGp (2)\n", GCPhys));
2672	for ( ; iPTDst < RT_ELEMENTS(pPTDst->a); iPTDst++)
2673	pPTDst->a[iPTDst].u = 0; /* MMIO or invalid page, we must handle them manually. */
2674	}
2675	} /* while more PTEs */
2676	} /* 4KB / 4MB */
2677	}
2678	else
2679	AssertRelease(!PdeDst.n.u1Present);
2680
2681	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT), a);
2682	if (RT_FAILURE(rc))
2683	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPTFailed));
2684	return rc;
2685
2686	#elif (PGM_GST_TYPE == PGM_TYPE_REAL \|\| PGM_GST_TYPE == PGM_TYPE_PROT) \
2687	&& PGM_SHW_TYPE != PGM_TYPE_NESTED \
2688	&& (PGM_SHW_TYPE != PGM_TYPE_EPT \|\| PGM_GST_TYPE == PGM_TYPE_PROT)
2689
2690	int rc = VINF_SUCCESS;
2691
2692	/*
2693	* Validate input a little bit.
2694	*/
2695	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
2696	const unsigned iPDDst = GCPtrPage >> SHW_PD_SHIFT;
2697	PX86PD pPDDst = pVM->pgm.s.CTXMID(p,32BitPD);
2698	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
2699	const unsigned iPDDst = GCPtrPage >> SHW_PD_SHIFT; /* no mask; flat index into the 2048 entry array. */
2700	PX86PDPAE pPDDst = pVM->pgm.s.CTXMID(ap,PaePDs)[0];
2701	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
2702	const unsigned iPdpte = (GCPtrPage >> X86_PDPT_SHIFT) & X86_PDPT_MASK_AMD64;
2703	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
2704	PX86PDPAE pPDDst;
2705	PX86PDPT pPdptDst;
2706	rc = PGMShwGetLongModePDPtr(pVM, GCPtrPage, &pPdptDst, &pPDDst);
2707	AssertRCSuccessReturn(rc, rc);
2708	Assert(pPDDst);
2709
2710	/* Fetch the pgm pool shadow descriptor. */
2711	PPGMPOOLPAGE pShwPde = pgmPoolGetPageByHCPhys(pVM, pPdptDst->a[iPdpte].u & X86_PDPE_PG_MASK);
2712	Assert(pShwPde);
2713	# elif PGM_SHW_TYPE == PGM_TYPE_EPT
2714	const unsigned iPdpte = (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
2727	/* Fetch the pgm pool shadow descriptor. */
2728	PPGMPOOLPAGE pShwPde = pgmPoolGetPageByHCPhys(pVM, pPdptDst->a[iPdpte].u & EPT_PDPTE_PG_MASK);
2729	Assert(pShwPde);
2730	# endif
2731	PSHWPDE pPdeDst = &pPDDst->a[iPDDst];
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, (RTGCUINTPTR)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, RTGCUINTPTR 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 = (RTGCUINTPTR)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);
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 = pVM->pgm.s.CTXMID(p,32BitPD)->a[GCPtrPage >> SHW_PD_SHIFT];
2843	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
2844	const X86PDEPAE PdeDst = pVM->pgm.s.CTXMID(ap,PaePDs)[0]->a[GCPtrPage >> SHW_PD_SHIFT];
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, RTGCUINTPTR 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 = (RTGCUINTPTR)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);
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 = &pVM->pgm.s.CTXMID(p,32BitPD)->a[GCPtrPage >> SHW_PD_SHIFT];
2963	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
2964	PX86PDEPAE pPdeDst = &pVM->pgm.s.CTXMID(ap,PaePDs)[0]->a[GCPtrPage >> SHW_PD_SHIFT];
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	#ifdef PGMPOOL_WITH_MONITORING
3123	int rc = pgmPoolSyncCR3(pVM);
3124	if (rc != VINF_SUCCESS)
3125	return rc;
3126	#endif
3127
3128	#if PGM_SHW_TYPE == PGM_TYPE_NESTED \|\| PGM_SHW_TYPE == PGM_TYPE_EPT
3129	/** @todo check if this is really necessary */
3130	HWACCMFlushTLB(pVM);
3131	return VINF_SUCCESS;
3132
3133	#elif PGM_SHW_TYPE == PGM_TYPE_AMD64
3134	/* No need to check all paging levels; we zero out the shadow parts when the guest modifies its tables. */
3135	return VINF_SUCCESS;
3136	#else
3137
3138	Assert(fGlobal \|\| (cr4 & X86_CR4_PGE));
3139	MY_STAM_COUNTER_INC(fGlobal ? &pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3Global) : &pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3NotGlobal));
3140
3141	# if PGM_GST_TYPE == PGM_TYPE_32BIT \|\| PGM_GST_TYPE == PGM_TYPE_PAE \|\| PGM_GST_TYPE == PGM_TYPE_AMD64
3142	# if PGM_GST_TYPE == PGM_TYPE_AMD64
3143	bool fBigPagesSupported = true;
3144	# else
3145	bool fBigPagesSupported = !!(CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
3146	# endif
3147
3148	/*
3149	* Get page directory addresses.
3150	*/
3151	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
3152	PX86PDE pPDEDst = &pVM->pgm.s.CTXMID(p,32BitPD)->a[0];
3153	# else /* PGM_SHW_TYPE == PGM_TYPE_PAE \|\| PGM_SHW_TYPE == PGM_TYPE_AMD64*/
3154	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3155	PX86PDEPAE pPDEDst = &pVM->pgm.s.CTXMID(ap,PaePDs)[0]->a[0];
3156	# endif
3157	# endif
3158
3159	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3160	PGSTPD pPDSrc = pgmGstGet32bitPDPtr(&pVM->pgm.s);
3161	Assert(pPDSrc);
3162	# ifndef IN_RC
3163	Assert(PGMPhysGCPhys2HCPtrAssert(pVM, (RTGCPHYS)(cr3 & GST_CR3_PAGE_MASK), sizeof(*pPDSrc)) == pPDSrc);
3164	# endif
3165	# endif
3166
3167	/*
3168	* Iterate the page directory.
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	# if PGM_GST_TYPE == PGM_TYPE_AMD64
3187	for (uint64_t iPml4e = 0; iPml4e < X86_PG_PAE_ENTRIES; iPml4e++)
3188	{
3189	PPGMPOOLPAGE pShwPdpt = NULL;
3190	PX86PML4E pPml4eSrc, pPml4eDst;
3191	RTGCPHYS GCPhysPdptSrc;
3192
3193	pPml4eSrc = &pVM->pgm.s.CTXSUFF(pGstPaePML4)->a[iPml4e];
3194	pPml4eDst = &pVM->pgm.s.CTXMID(p,PaePML4)->a[iPml4e];
3195
3196	/* Fetch the pgm pool shadow descriptor if the shadow pml4e is present. */
3197	if (!pPml4eDst->n.u1Present)
3198	continue;
3199	pShwPdpt = pgmPoolGetPage(pPool, pPml4eDst->u & X86_PML4E_PG_MASK);
3200
3201	GCPhysPdptSrc = pPml4eSrc->u & X86_PML4E_PG_MASK_FULL;
3202
3203	/* Anything significant changed? */
3204	if ( pPml4eSrc->n.u1Present != pPml4eDst->n.u1Present
3205	\|\| GCPhysPdptSrc != pShwPdpt->GCPhys)
3206	{
3207	/* Free it. */
3208	LogFlow(("SyncCR3: Out-of-sync PML4E (GCPhys) GCPtr=%RX64 %RGp vs %RGp PdpeSrc=%RX64 PdpeDst=%RX64\n",
3209	(uint64_t)iPml4e << X86_PML4_SHIFT, pShwPdpt->GCPhys, GCPhysPdptSrc, (uint64_t)pPml4eSrc->u, (uint64_t)pPml4eDst->u));
3210	pgmPoolFreeByPage(pPool, pShwPdpt, pVM->pgm.s.pHCShwAmd64CR3->idx, iPml4e);
3211	pPml4eDst->u = 0;
3212	continue;
3213	}
3214	/* Force an attribute sync. */
3215	pPml4eDst->n.u1User = pPml4eSrc->n.u1User;
3216	pPml4eDst->n.u1Write = pPml4eSrc->n.u1Write;
3217	pPml4eDst->n.u1NoExecute = pPml4eSrc->n.u1NoExecute;
3218
3219	# else
3220	{
3221	# endif
3222	# if PGM_GST_TYPE == PGM_TYPE_PAE \|\| PGM_GST_TYPE == PGM_TYPE_AMD64
3223	for (uint64_t iPdpte = 0; iPdpte < GST_PDPE_ENTRIES; iPdpte++)
3224	{
3225	unsigned iPDSrc;
3226	# if PGM_GST_TYPE == PGM_TYPE_PAE
3227	PX86PDPAE pPDPAE = pVM->pgm.s.CTXMID(ap,PaePDs)[0];
3228	PX86PDEPAE pPDEDst = &pPDPAE->a[iPdpte * X86_PG_PAE_ENTRIES];
3229	PGSTPD pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, iPdpte << X86_PDPT_SHIFT, &iPDSrc);
3230	PX86PDPT pPdptDst = pVM->pgm.s.CTXMID(p,PaePDPT); NOREF(pPdptDst);
3231	X86PDPE PdpeSrc = CTXSUFF(pVM->pgm.s.pGstPaePDPT)->a[iPdpte];
3232
3233	if (pPDSrc == NULL)
3234	{
3235	/* PDPE not present */
3236	if (pVM->pgm.s.CTXMID(p,PaePDPT)->a[iPdpte].n.u1Present)
3237	{
3238	LogFlow(("SyncCR3: guest PDPE %d not present; clear shw pdpe\n", iPdpte));
3239	/* for each page directory entry */
3240	for (unsigned iPD = 0; iPD < RT_ELEMENTS(pPDSrc->a); iPD++)
3241	{
3242	if ( pPDEDst[iPD].n.u1Present
3243	&& !(pPDEDst[iPD].u & PGM_PDFLAGS_MAPPING))
3244	{
3245	pgmPoolFreeByPage(pPool, pgmPoolGetPage(pPool, pPDEDst[iPD].u & SHW_PDE_PG_MASK), SHW_POOL_ROOT_IDX, iPdpte * X86_PG_PAE_ENTRIES + iPD);
3246	pPDEDst[iPD].u = 0;
3247	}
3248	}
3249	}
3250	if (!(pVM->pgm.s.CTXMID(p,PaePDPT)->a[iPdpte].u & PGM_PLXFLAGS_MAPPING))
3251	pVM->pgm.s.CTXMID(p,PaePDPT)->a[iPdpte].n.u1Present = 0;
3252	continue;
3253	}
3254	# else /* PGM_GST_TYPE != PGM_TYPE_PAE */
3255	PPGMPOOLPAGE pShwPde = NULL;
3256	RTGCPHYS GCPhysPdeSrc;
3257	PX86PDPE pPdpeDst;
3258	PX86PML4E pPml4eSrc;
3259	X86PDPE PdpeSrc;
3260	PX86PDPT pPdptDst;
3261	PX86PDPAE pPDDst;
3262	PX86PDEPAE pPDEDst;
3263	RTGCUINTPTR GCPtr = (iPml4e << X86_PML4_SHIFT) \|\| (iPdpte << X86_PDPT_SHIFT);
3264	PGSTPD pPDSrc = pgmGstGetLongModePDPtr(&pVM->pgm.s, GCPtr, &pPml4eSrc, &PdpeSrc, &iPDSrc);
3265
3266	int rc = PGMShwGetLongModePDPtr(pVM, GCPtr, &pPdptDst, &pPDDst);
3267	if (rc != VINF_SUCCESS)
3268	{
3269	if (rc == VERR_PAGE_MAP_LEVEL4_NOT_PRESENT)
3270	break; /* next PML4E */
3271
3272	AssertMsg(rc == VERR_PAGE_DIRECTORY_PTR_NOT_PRESENT, ("Unexpected rc=%Rrc\n", rc));
3273	continue; /* next PDPTE */
3274	}
3275	Assert(pPDDst);
3276	pPDEDst = &pPDDst->a[0];
3277	Assert(iPDSrc == 0);
3278
3279	pPdpeDst = &pPdptDst->a[iPdpte];
3280
3281	/* Fetch the pgm pool shadow descriptor if the shadow pdpte is present. */
3282	if (!pPdpeDst->n.u1Present)
3283	continue; /* next PDPTE */
3284
3285	pShwPde = pgmPoolGetPage(pPool, pPdpeDst->u & X86_PDPE_PG_MASK);
3286	GCPhysPdeSrc = PdpeSrc.u & X86_PDPE_PG_MASK;
3287
3288	/* Anything significant changed? */
3289	if ( PdpeSrc.n.u1Present != pPdpeDst->n.u1Present
3290	\|\| GCPhysPdeSrc != pShwPde->GCPhys)
3291	{
3292	/* Free it. */
3293	LogFlow(("SyncCR3: Out-of-sync PDPE (GCPhys) GCPtr=%RX64 %RGp vs %RGp PdpeSrc=%RX64 PdpeDst=%RX64\n",
3294	((uint64_t)iPml4e << X86_PML4_SHIFT) + ((uint64_t)iPdpte << X86_PDPT_SHIFT), pShwPde->GCPhys, GCPhysPdeSrc, (uint64_t)PdpeSrc.u, (uint64_t)pPdpeDst->u));
3295
3296	/* Mark it as not present if there's no hypervisor mapping present. (bit flipped at the top of Trap0eHandler) */
3297	Assert(!(pPdpeDst->u & PGM_PLXFLAGS_MAPPING));
3298	pgmPoolFreeByPage(pPool, pShwPde, pShwPde->idx, iPdpte);
3299	pPdpeDst->u = 0;
3300	continue; /* next guest PDPTE */
3301	}
3302	/* Force an attribute sync. */
3303	pPdpeDst->lm.u1User = PdpeSrc.lm.u1User;
3304	pPdpeDst->lm.u1Write = PdpeSrc.lm.u1Write;
3305	pPdpeDst->lm.u1NoExecute = PdpeSrc.lm.u1NoExecute;
3306	# endif /* PGM_GST_TYPE != PGM_TYPE_PAE */
3307
3308	# else /* PGM_GST_TYPE != PGM_TYPE_PAE && PGM_GST_TYPE != PGM_TYPE_AMD64 */
3309	{
3310	# endif /* PGM_GST_TYPE != PGM_TYPE_PAE && PGM_GST_TYPE != PGM_TYPE_AMD64 */
3311	for (unsigned iPD = 0; iPD < RT_ELEMENTS(pPDSrc->a); iPD++)
3312	{
3313	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
3314	Assert(&pVM->pgm.s.CTXMID(p,32BitPD)->a[iPD] == pPDEDst);
3315	# elif PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
3316	AssertMsg(&pVM->pgm.s.CTXMID(ap,PaePDs)[iPD * 2 / 512]->a[iPD * 2 % 512] == pPDEDst, ("%p vs %p\n", &pVM->pgm.s.CTXMID(ap,PaePDs)[iPD * 2 / 512]->a[iPD * 2 % 512], pPDEDst));
3317	# endif
3318	GSTPDE PdeSrc = pPDSrc->a[iPD];
3319	if ( PdeSrc.n.u1Present
3320	&& (PdeSrc.n.u1User \|\| fRawR0Enabled))
3321	{
3322	# if ( PGM_GST_TYPE == PGM_TYPE_32BIT \
3323	\|\| PGM_GST_TYPE == PGM_TYPE_PAE) \
3324	&& !defined(PGM_WITHOUT_MAPPINGS)
3325
3326	/*
3327	* Check for conflicts with GC mappings.
3328	*/
3329	# if PGM_GST_TYPE == PGM_TYPE_PAE
3330	if (iPD + iPdpte * X86_PG_PAE_ENTRIES == iPdNoMapping)
3331	# else
3332	if (iPD == iPdNoMapping)
3333	# endif
3334	{
3335	if (pVM->pgm.s.fMappingsFixed)
3336	{
3337	/* It's fixed, just skip the mapping. */
3338	const unsigned cPTs = pMapping->cb >> GST_PD_SHIFT;
3339	iPD += cPTs - 1;
3340	pPDEDst += cPTs + (PGM_GST_TYPE != PGM_SHW_TYPE) * cPTs; /* Only applies to the pae shadow and 32 bits guest case */
3341	pMapping = pMapping->CTX_SUFF(pNext);
3342	iPdNoMapping = pMapping ? pMapping->GCPtr >> GST_PD_SHIFT : ~0U;
3343	continue;
3344	}
3345	# ifdef IN_RING3
3346	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3347	int rc = pgmR3SyncPTResolveConflict(pVM, pMapping, pPDSrc, iPD << GST_PD_SHIFT);
3348	# elif PGM_GST_TYPE == PGM_TYPE_PAE
3349	int rc = pgmR3SyncPTResolveConflictPAE(pVM, pMapping, (iPdpte << GST_PDPT_SHIFT) + (iPD << GST_PD_SHIFT));
3350	# endif
3351	if (RT_FAILURE(rc))
3352	return rc;
3353
3354	/*
3355	* Update iPdNoMapping and pMapping.
3356	*/
3357	pMapping = pVM->pgm.s.pMappingsR3;
3358	while (pMapping && pMapping->GCPtr < (iPD << GST_PD_SHIFT))
3359	pMapping = pMapping->pNextR3;
3360	iPdNoMapping = pMapping ? pMapping->GCPtr >> GST_PD_SHIFT : ~0U;
3361	# else
3362	LogFlow(("SyncCR3: detected conflict -> VINF_PGM_SYNC_CR3\n"));
3363	return VINF_PGM_SYNC_CR3;
3364	# endif
3365	}
3366	# else /* (PGM_GST_TYPE != PGM_TYPE_32BIT && PGM_GST_TYPE != PGM_TYPE_PAE) \|\| PGM_WITHOUT_MAPPINGS */
3367	Assert(!pgmMapAreMappingsEnabled(&pVM->pgm.s));
3368	# endif /* (PGM_GST_TYPE != PGM_TYPE_32BIT && PGM_GST_TYPE != PGM_TYPE_PAE) \|\| PGM_WITHOUT_MAPPINGS */
3369
3370	/*
3371	* Sync page directory entry.
3372	*
3373	* The current approach is to allocated the page table but to set
3374	* the entry to not-present and postpone the page table synching till
3375	* it's actually used.
3376	*/
3377	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
3378	for (unsigned i = 0, iPdShw = iPD * 2; i < 2; i++, iPdShw++) /* pray that the compiler unrolls this */
3379	# elif PGM_GST_TYPE == PGM_TYPE_PAE
3380	const unsigned iPdShw = iPD + iPdpte * X86_PG_PAE_ENTRIES; NOREF(iPdShw);
3381	# else
3382	const unsigned iPdShw = iPD; NOREF(iPdShw);
3383	# endif
3384	{
3385	SHWPDE PdeDst = *pPDEDst;
3386	if (PdeDst.n.u1Present)
3387	{
3388	PPGMPOOLPAGE pShwPage = pgmPoolGetPage(pPool, PdeDst.u & SHW_PDE_PG_MASK);
3389	RTGCPHYS GCPhys;
3390	if ( !PdeSrc.b.u1Size
3391	\|\| !fBigPagesSupported)
3392	{
3393	GCPhys = PdeSrc.u & GST_PDE_PG_MASK;
3394	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
3395	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
3396	GCPhys \|= i * (PAGE_SIZE / 2);
3397	# endif
3398	}
3399	else
3400	{
3401	GCPhys = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc);
3402	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
3403	/* Select the right PDE as we're emulating a 4MB page directory with two 2 MB shadow PDEs.*/
3404	GCPhys \|= i * X86_PAGE_2M_SIZE;
3405	# endif
3406	}
3407
3408	if ( pShwPage->GCPhys == GCPhys
3409	&& pShwPage->enmKind == PGM_BTH_NAME(CalcPageKind)(&PdeSrc, cr4)
3410	&& ( pShwPage->fCached
3411	\|\| ( !fGlobal
3412	&& ( false
3413	# ifdef PGM_SKIP_GLOBAL_PAGEDIRS_ON_NONGLOBAL_FLUSH
3414	\|\| ( (PdeSrc.u & (X86_PDE4M_PS \| X86_PDE4M_G)) == (X86_PDE4M_PS \| X86_PDE4M_G)
3415	# if PGM_GST_TYPE == PGM_TYPE_AMD64
3416	&& (cr4 & X86_CR4_PGE)) /* global 2/4MB page. */
3417	# else
3418	&& (cr4 & (X86_CR4_PGE \| X86_CR4_PSE)) == (X86_CR4_PGE \| X86_CR4_PSE)) /* global 2/4MB page. */
3419	# endif
3420	\|\| ( !pShwPage->fSeenNonGlobal
3421	&& (cr4 & X86_CR4_PGE))
3422	# endif
3423	)
3424	)
3425	)
3426	&& ( (PdeSrc.u & (X86_PDE_US \| X86_PDE_RW)) == (PdeDst.u & (X86_PDE_US \| X86_PDE_RW))
3427	\|\| ( fBigPagesSupported
3428	&& ((PdeSrc.u & (X86_PDE_US \| X86_PDE4M_PS \| X86_PDE4M_D)) \| PGM_PDFLAGS_TRACK_DIRTY)
3429	== ((PdeDst.u & (X86_PDE_US \| X86_PDE_RW \| PGM_PDFLAGS_TRACK_DIRTY)) \| X86_PDE4M_PS))
3430	)
3431	)
3432	{
3433	# ifdef VBOX_WITH_STATISTICS
3434	if ( !fGlobal
3435	&& (PdeSrc.u & (X86_PDE4M_PS \| X86_PDE4M_G)) == (X86_PDE4M_PS \| X86_PDE4M_G)
3436	# if PGM_GST_TYPE == PGM_TYPE_AMD64
3437	&& (cr4 & X86_CR4_PGE)) /* global 2/4MB page. */
3438	# else
3439	&& (cr4 & (X86_CR4_PGE \| X86_CR4_PSE)) == (X86_CR4_PGE \| X86_CR4_PSE))
3440	# endif
3441	MY_STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3DstSkippedGlobalPD));
3442	else if (!fGlobal && !pShwPage->fSeenNonGlobal && (cr4 & X86_CR4_PGE))
3443	MY_STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3DstSkippedGlobalPT));
3444	else
3445	MY_STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3DstCacheHit));
3446	# endif /* VBOX_WITH_STATISTICS */
3447	/** @todo a replacement strategy isn't really needed unless we're using a very small pool < 512 pages.
3448	* The whole ageing stuff should be put in yet another set of #ifdefs. For now, let's just skip it. */
3449	//# ifdef PGMPOOL_WITH_CACHE
3450	// pgmPoolCacheUsed(pPool, pShwPage);
3451	//# endif
3452	}
3453	else
3454	{
3455	# if PGM_GST_TYPE == PGM_TYPE_AMD64
3456	pgmPoolFreeByPage(pPool, pShwPage, pShwPde->idx, iPdShw);
3457	# else
3458	pgmPoolFreeByPage(pPool, pShwPage, SHW_POOL_ROOT_IDX, iPdShw);
3459	# endif
3460	pPDEDst->u = 0;
3461	MY_STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3DstFreed));
3462	}
3463	}
3464	else
3465	MY_STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3DstNotPresent));
3466	pPDEDst++;
3467	}
3468	}
3469	# if PGM_GST_TYPE == PGM_TYPE_PAE
3470	else if (iPD + iPdpte * X86_PG_PAE_ENTRIES != iPdNoMapping)
3471	# else
3472	else if (iPD != iPdNoMapping)
3473	# endif
3474	{
3475	/*
3476	* Check if there is any page directory to mark not present here.
3477	*/
3478	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
3479	for (unsigned i = 0, iPdShw = iPD * 2; i < 2; i++, iPdShw++) /* pray that the compiler unrolls this */
3480	# elif PGM_GST_TYPE == PGM_TYPE_PAE
3481	const unsigned iPdShw = iPD + iPdpte * X86_PG_PAE_ENTRIES; NOREF(iPdShw);
3482	# else
3483	const unsigned iPdShw = iPD; NOREF(iPdShw);
3484	# endif
3485	{
3486	if (pPDEDst->n.u1Present)
3487	{
3488	# if PGM_GST_TYPE == PGM_TYPE_AMD64
3489	pgmPoolFreeByPage(pPool, pgmPoolGetPage(pPool, pPDEDst->u & SHW_PDE_PG_MASK), pShwPde->idx, iPdShw);
3490	# else
3491	pgmPoolFreeByPage(pPool, pgmPoolGetPage(pPool, pPDEDst->u & SHW_PDE_PG_MASK), SHW_POOL_ROOT_IDX, iPdShw);
3492	# endif
3493	pPDEDst->u = 0;
3494	MY_STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3DstFreedSrcNP));
3495	}
3496	pPDEDst++;
3497	}
3498	}
3499	else
3500	{
3501	# if ( PGM_GST_TYPE == PGM_TYPE_32BIT \
3502	\|\| PGM_GST_TYPE == PGM_TYPE_PAE) \
3503	&& !defined(PGM_WITHOUT_MAPPINGS)
3504
3505	const unsigned cPTs = pMapping->cb >> GST_PD_SHIFT;
3506
3507	Assert(pgmMapAreMappingsEnabled(&pVM->pgm.s));
3508	if (pVM->pgm.s.fMappingsFixed)
3509	{
3510	/* It's fixed, just skip the mapping. */
3511	pMapping = pMapping->CTX_SUFF(pNext);
3512	iPdNoMapping = pMapping ? pMapping->GCPtr >> GST_PD_SHIFT : ~0U;
3513	}
3514	else
3515	{
3516	/*
3517	* Check for conflicts for subsequent pagetables
3518	* and advance to the next mapping.
3519	*/
3520	iPdNoMapping = ~0U;
3521	unsigned iPT = cPTs;
3522	while (iPT-- > 1)
3523	{
3524	if ( pPDSrc->a[iPD + iPT].n.u1Present
3525	&& (pPDSrc->a[iPD + iPT].n.u1User \|\| fRawR0Enabled))
3526	{
3527	# ifdef IN_RING3
3528	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3529	int rc = pgmR3SyncPTResolveConflict(pVM, pMapping, pPDSrc, iPD << GST_PD_SHIFT);
3530	# elif PGM_GST_TYPE == PGM_TYPE_PAE
3531	int rc = pgmR3SyncPTResolveConflictPAE(pVM, pMapping, (iPdpte << GST_PDPT_SHIFT) + (iPD << GST_PD_SHIFT));
3532	# endif
3533	if (RT_FAILURE(rc))
3534	return rc;
3535
3536	/*
3537	* Update iPdNoMapping and pMapping.
3538	*/
3539	pMapping = pVM->pgm.s.CTX_SUFF(pMappings);
3540	while (pMapping && pMapping->GCPtr < (iPD << GST_PD_SHIFT))
3541	pMapping = pMapping->CTX_SUFF(pNext);
3542	iPdNoMapping = pMapping ? pMapping->GCPtr >> GST_PD_SHIFT : ~0U;
3543	break;
3544	# else
3545	LogFlow(("SyncCR3: detected conflict -> VINF_PGM_SYNC_CR3\n"));
3546	return VINF_PGM_SYNC_CR3;
3547	# endif
3548	}
3549	}
3550	if (iPdNoMapping == ~0U && pMapping)
3551	{
3552	pMapping = pMapping->CTX_SUFF(pNext);
3553	if (pMapping)
3554	iPdNoMapping = pMapping->GCPtr >> GST_PD_SHIFT;
3555	}
3556	}
3557
3558	/* advance. */
3559	iPD += cPTs - 1;
3560	pPDEDst += cPTs + (PGM_GST_TYPE != PGM_SHW_TYPE) * cPTs; /* Only applies to the pae shadow and 32 bits guest case */
3561	# if PGM_GST_TYPE != PGM_SHW_TYPE
3562	AssertCompile(PGM_GST_TYPE == PGM_TYPE_32BIT && PGM_SHW_TYPE == PGM_TYPE_PAE);
3563	# endif
3564	# else /* (PGM_GST_TYPE != PGM_TYPE_32BIT && PGM_GST_TYPE != PGM_TYPE_PAE) \|\| PGM_WITHOUT_MAPPINGS */
3565	Assert(!pgmMapAreMappingsEnabled(&pVM->pgm.s));
3566	# endif /* (PGM_GST_TYPE != PGM_TYPE_32BIT && PGM_GST_TYPE != PGM_TYPE_PAE) \|\| PGM_WITHOUT_MAPPINGS */
3567	}
3568
3569	} /* for iPD */
3570	} /* for each PDPTE (PAE) */
3571	} /* for each page map level 4 entry (amd64) */
3572	return VINF_SUCCESS;
3573
3574	# else /* guest real and protected mode */
3575	return VINF_SUCCESS;
3576	# endif
3577	#endif /* PGM_SHW_TYPE != PGM_TYPE_NESTED && PGM_SHW_TYPE != PGM_TYPE_EPT */
3578	}
3579
3580
3581
3582
3583	#ifdef VBOX_STRICT
3584	#ifdef IN_RC
3585	# undef AssertMsgFailed
3586	# define AssertMsgFailed Log
3587	#endif
3588	#ifdef IN_RING3
3589	# include <VBox/dbgf.h>
3590
3591	/**
3592	* Dumps a page table hierarchy use only physical addresses and cr4/lm flags.
3593	*
3594	* @returns VBox status code (VINF_SUCCESS).
3595	* @param pVM The VM handle.
3596	* @param cr3 The root of the hierarchy.
3597	* @param crr The cr4, only PAE and PSE is currently used.
3598	* @param fLongMode Set if long mode, false if not long mode.
3599	* @param cMaxDepth Number of levels to dump.
3600	* @param pHlp Pointer to the output functions.
3601	*/
3602	__BEGIN_DECLS
3603	VMMR3DECL(int) PGMR3DumpHierarchyHC(PVM pVM, uint32_t cr3, uint32_t cr4, bool fLongMode, unsigned cMaxDepth, PCDBGFINFOHLP pHlp);
3604	__END_DECLS
3605
3606	#endif
3607
3608	/**
3609	* Checks that the shadow page table is in sync with the guest one.
3610	*
3611	* @returns The number of errors.
3612	* @param pVM The virtual machine.
3613	* @param cr3 Guest context CR3 register
3614	* @param cr4 Guest context CR4 register
3615	* @param GCPtr Where to start. Defaults to 0.
3616	* @param cb How much to check. Defaults to everything.
3617	*/
3618	PGM_BTH_DECL(unsigned, AssertCR3)(PVM pVM, uint64_t cr3, uint64_t cr4, RTGCUINTPTR GCPtr, RTGCUINTPTR cb)
3619	{
3620	#if PGM_SHW_TYPE == PGM_TYPE_NESTED \|\| PGM_SHW_TYPE == PGM_TYPE_EPT
3621	return 0;
3622	#else
3623	unsigned cErrors = 0;
3624
3625	#if PGM_GST_TYPE == PGM_TYPE_PAE
3626	/** @todo currently broken; crashes below somewhere */
3627	AssertFailed();
3628	#endif
3629
3630	#if PGM_GST_TYPE == PGM_TYPE_32BIT \
3631	\|\| PGM_GST_TYPE == PGM_TYPE_PAE \
3632	\|\| PGM_GST_TYPE == PGM_TYPE_AMD64
3633
3634	# if PGM_GST_TYPE == PGM_TYPE_AMD64
3635	bool fBigPagesSupported = true;
3636	# else
3637	bool fBigPagesSupported = !!(CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
3638	# endif
3639	PPGM pPGM = &pVM->pgm.s;
3640	RTGCPHYS GCPhysGst; /* page address derived from the guest page tables. */
3641	RTHCPHYS HCPhysShw; /* page address derived from the shadow page tables. */
3642	# ifndef IN_RING0
3643	RTHCPHYS HCPhys; /* general usage. */
3644	# endif
3645	int rc;
3646
3647	/*
3648	* Check that the Guest CR3 and all its mappings are correct.
3649	*/
3650	AssertMsgReturn(pPGM->GCPhysCR3 == (cr3 & GST_CR3_PAGE_MASK),
3651	("Invalid GCPhysCR3=%RGp cr3=%RGp\n", pPGM->GCPhysCR3, (RTGCPHYS)cr3),
3652	false);
3653	# if !defined(IN_RING0) && PGM_GST_TYPE != PGM_TYPE_AMD64
3654	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3655	rc = PGMShwGetPage(pVM, (RTGCPTR)pPGM->pGuestPDRC, NULL, &HCPhysShw);
3656	# else
3657	rc = PGMShwGetPage(pVM, (RTGCPTR)pPGM->pGstPaePDPTGC, NULL, &HCPhysShw);
3658	# endif
3659	AssertRCReturn(rc, 1);
3660	HCPhys = NIL_RTHCPHYS;
3661	rc = pgmRamGCPhys2HCPhys(pPGM, cr3 & GST_CR3_PAGE_MASK, &HCPhys);
3662	AssertMsgReturn(HCPhys == HCPhysShw, ("HCPhys=%RHp HCPhyswShw=%RHp (cr3)\n", HCPhys, HCPhysShw), false);
3663	# if PGM_GST_TYPE == PGM_TYPE_32BIT && defined(IN_RING3)
3664	RTGCPHYS GCPhys;
3665	rc = PGMR3DbgR3Ptr2GCPhys(pVM, pPGM->pGuestPDR3, &GCPhys);
3666	AssertRCReturn(rc, 1);
3667	AssertMsgReturn((cr3 & GST_CR3_PAGE_MASK) == GCPhys, ("GCPhys=%RGp cr3=%RGp\n", GCPhys, (RTGCPHYS)cr3), false);
3668	# endif
3669	#endif /* !IN_RING0 */
3670
3671	/*
3672	* Get and check the Shadow CR3.
3673	*/
3674	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
3675	unsigned cPDEs = X86_PG_ENTRIES;
3676	unsigned cIncrement = X86_PG_ENTRIES * PAGE_SIZE;
3677	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
3678	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3679	unsigned cPDEs = X86_PG_PAE_ENTRIES * 4; /* treat it as a 2048 entry table. */
3680	# else
3681	unsigned cPDEs = X86_PG_PAE_ENTRIES;
3682	# endif
3683	unsigned cIncrement = X86_PG_PAE_ENTRIES * PAGE_SIZE;
3684	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
3685	unsigned cPDEs = X86_PG_PAE_ENTRIES;
3686	unsigned cIncrement = X86_PG_PAE_ENTRIES * PAGE_SIZE;
3687	# endif
3688	if (cb != ~(RTGCUINTPTR)0)
3689	cPDEs = RT_MIN(cb >> SHW_PD_SHIFT, 1);
3690
3691	/** @todo call the other two PGMAssert() functions. /
3692
3693	# if PGM_GST_TYPE == PGM_TYPE_AMD64 \|\| PGM_GST_TYPE == PGM_TYPE_PAE
3694	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
3695	# endif
3696
3697	# if PGM_GST_TYPE == PGM_TYPE_AMD64
3698	unsigned iPml4e = (GCPtr >> X86_PML4_SHIFT) & X86_PML4_MASK;
3699
3700	for (; iPml4e < X86_PG_PAE_ENTRIES; iPml4e++)
3701	{
3702	PPGMPOOLPAGE pShwPdpt = NULL;
3703	PX86PML4E pPml4eSrc;
3704	PX86PML4E pPml4eDst;
3705	RTGCPHYS GCPhysPdptSrc;
3706
3707	pPml4eSrc = &pVM->pgm.s.CTXSUFF(pGstPaePML4)->a[iPml4e];
3708	pPml4eDst = &pVM->pgm.s.CTXMID(p,PaePML4)->a[iPml4e];
3709
3710	/* Fetch the pgm pool shadow descriptor if the shadow pml4e is present. */
3711	if (!pPml4eDst->n.u1Present)
3712	{
3713	GCPtr += _2M * UINT64_C(512) * UINT64_C(512);
3714	continue;
3715	}
3716
3717	# if PGM_GST_TYPE == PGM_TYPE_PAE
3718	/* not correct to call pgmPoolGetPage */
3719	AssertFailed();
3720	# endif
3721	pShwPdpt = pgmPoolGetPage(pPool, pPml4eDst->u & X86_PML4E_PG_MASK);
3722	GCPhysPdptSrc = pPml4eSrc->u & X86_PML4E_PG_MASK_FULL;
3723
3724	if (pPml4eSrc->n.u1Present != pPml4eDst->n.u1Present)
3725	{
3726	AssertMsgFailed(("Present bit doesn't match! pPml4eDst.u=%#RX64 pPml4eSrc.u=%RX64\n", pPml4eDst->u, pPml4eSrc->u));
3727	GCPtr += _2M * UINT64_C(512) * UINT64_C(512);
3728	cErrors++;
3729	continue;
3730	}
3731
3732	if (GCPhysPdptSrc != pShwPdpt->GCPhys)
3733	{
3734	AssertMsgFailed(("Physical address doesn't match! iPml4e %d pPml4eDst.u=%#RX64 pPml4eSrc.u=%RX64 Phys %RX64 vs %RX64\n", iPml4e, pPml4eDst->u, pPml4eSrc->u, pShwPdpt->GCPhys, GCPhysPdptSrc));
3735	GCPtr += _2M * UINT64_C(512) * UINT64_C(512);
3736	cErrors++;
3737	continue;
3738	}
3739
3740	if ( pPml4eDst->n.u1User != pPml4eSrc->n.u1User
3741	\|\| pPml4eDst->n.u1Write != pPml4eSrc->n.u1Write
3742	\|\| pPml4eDst->n.u1NoExecute != pPml4eSrc->n.u1NoExecute)
3743	{
3744	AssertMsgFailed(("User/Write/NoExec bits don't match! pPml4eDst.u=%#RX64 pPml4eSrc.u=%RX64\n", pPml4eDst->u, pPml4eSrc->u));
3745	GCPtr += _2M * UINT64_C(512) * UINT64_C(512);
3746	cErrors++;
3747	continue;
3748	}
3749	# else
3750	{
3751	# endif
3752
3753	# if PGM_GST_TYPE == PGM_TYPE_AMD64 \|\| PGM_GST_TYPE == PGM_TYPE_PAE
3754	/*
3755	* Check the PDPTEs too.
3756	*/
3757	unsigned iPdpte = (GCPtr >> SHW_PDPT_SHIFT) & SHW_PDPT_MASK;
3758
3759	for (;iPdpte <= SHW_PDPT_MASK; iPdpte++)
3760	{
3761	unsigned iPDSrc;
3762	PPGMPOOLPAGE pShwPde = NULL;
3763	PX86PDPE pPdpeDst;
3764	RTGCPHYS GCPhysPdeSrc;
3765	# if PGM_GST_TYPE == PGM_TYPE_PAE
3766	PX86PDPAE pPDDst = pVM->pgm.s.CTXMID(ap,PaePDs)[0];
3767	PGSTPD pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, GCPtr, &iPDSrc);
3768	PX86PDPT pPdptDst = pVM->pgm.s.CTXMID(p,PaePDPT);
3769	X86PDPE PdpeSrc = CTXSUFF(pVM->pgm.s.pGstPaePDPT)->a[iPdpte];
3770	# else
3771	PX86PML4E pPml4eSrc;
3772	X86PDPE PdpeSrc;
3773	PX86PDPT pPdptDst;
3774	PX86PDPAE pPDDst;
3775	PGSTPD pPDSrc = pgmGstGetLongModePDPtr(&pVM->pgm.s, GCPtr, &pPml4eSrc, &PdpeSrc, &iPDSrc);
3776
3777	rc = PGMShwGetLongModePDPtr(pVM, GCPtr, &pPdptDst, &pPDDst);
3778	if (rc != VINF_SUCCESS)
3779	{
3780	AssertMsg(rc == VERR_PAGE_DIRECTORY_PTR_NOT_PRESENT, ("Unexpected rc=%Rrc\n", rc));
3781	GCPtr += 512 * _2M;
3782	continue; /* next PDPTE */
3783	}
3784	Assert(pPDDst);
3785	# endif
3786	Assert(iPDSrc == 0);
3787
3788	pPdpeDst = &pPdptDst->a[iPdpte];
3789
3790	if (!pPdpeDst->n.u1Present)
3791	{
3792	GCPtr += 512 * _2M;
3793	continue; /* next PDPTE */
3794	}
3795
3796	pShwPde = pgmPoolGetPage(pPool, pPdpeDst->u & X86_PDPE_PG_MASK);
3797	GCPhysPdeSrc = PdpeSrc.u & X86_PDPE_PG_MASK;
3798
3799	if (pPdpeDst->n.u1Present != PdpeSrc.n.u1Present)
3800	{
3801	AssertMsgFailed(("Present bit doesn't match! pPdpeDst.u=%#RX64 pPdpeSrc.u=%RX64\n", pPdpeDst->u, PdpeSrc.u));
3802	GCPtr += 512 * _2M;
3803	cErrors++;
3804	continue;
3805	}
3806
3807	if (GCPhysPdeSrc != pShwPde->GCPhys)
3808	{
3809	# if PGM_GST_TYPE == PGM_TYPE_AMD64
3810	AssertMsgFailed(("Physical address doesn't match! iPml4e %d iPdpte %d pPdpeDst.u=%#RX64 pPdpeSrc.u=%RX64 Phys %RX64 vs %RX64\n", iPml4e, iPdpte, pPdpeDst->u, PdpeSrc.u, pShwPde->GCPhys, GCPhysPdeSrc));
3811	# else
3812	AssertMsgFailed(("Physical address doesn't match! iPdpte %d pPdpeDst.u=%#RX64 pPdpeSrc.u=%RX64 Phys %RX64 vs %RX64\n", iPdpte, pPdpeDst->u, PdpeSrc.u, pShwPde->GCPhys, GCPhysPdeSrc));
3813	# endif
3814	GCPtr += 512 * _2M;
3815	cErrors++;
3816	continue;
3817	}
3818
3819	# if PGM_GST_TYPE == PGM_TYPE_AMD64
3820	if ( pPdpeDst->lm.u1User != PdpeSrc.lm.u1User
3821	\|\| pPdpeDst->lm.u1Write != PdpeSrc.lm.u1Write
3822	\|\| pPdpeDst->lm.u1NoExecute != PdpeSrc.lm.u1NoExecute)
3823	{
3824	AssertMsgFailed(("User/Write/NoExec bits don't match! pPdpeDst.u=%#RX64 pPdpeSrc.u=%RX64\n", pPdpeDst->u, PdpeSrc.u));
3825	GCPtr += 512 * _2M;
3826	cErrors++;
3827	continue;
3828	}
3829	# endif
3830
3831	# else
3832	{
3833	# endif
3834	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3835	GSTPD const *pPDSrc = pgmGstGet32bitPDPtr(&pVM->pgm.s);
3836	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
3837	PCX86PD pPDDst = pPGM->CTXMID(p,32BitPD);
3838	# else
3839	PCX86PDPAE pPDDst = pVM->pgm.s.CTXMID(ap,PaePDs)[0]; /* We treat this as a PD with 2048 entries, so no need to and with SHW_PD_MASK to get iPDDst */
3840	# endif
3841	# endif
3842	/*
3843	* Iterate the shadow page directory.
3844	*/
3845	GCPtr = (GCPtr >> SHW_PD_SHIFT) << SHW_PD_SHIFT;
3846	unsigned iPDDst = (GCPtr >> SHW_PD_SHIFT) & SHW_PD_MASK;
3847
3848	for (;
3849	iPDDst < cPDEs;
3850	iPDDst++, GCPtr += cIncrement)
3851	{
3852	const SHWPDE PdeDst = pPDDst->a[iPDDst];
3853	if (PdeDst.u & PGM_PDFLAGS_MAPPING)
3854	{
3855	Assert(pgmMapAreMappingsEnabled(&pVM->pgm.s));
3856	if ((PdeDst.u & X86_PDE_AVL_MASK) != PGM_PDFLAGS_MAPPING)
3857	{
3858	AssertMsgFailed(("Mapping shall only have PGM_PDFLAGS_MAPPING set! PdeDst.u=%#RX64\n", (uint64_t)PdeDst.u));
3859	cErrors++;
3860	continue;
3861	}
3862	}
3863	else if ( (PdeDst.u & X86_PDE_P)
3864	\|\| ((PdeDst.u & (X86_PDE_P \| PGM_PDFLAGS_TRACK_DIRTY)) == (X86_PDE_P \| PGM_PDFLAGS_TRACK_DIRTY))
3865	)
3866	{
3867	HCPhysShw = PdeDst.u & SHW_PDE_PG_MASK;
3868	PPGMPOOLPAGE pPoolPage = pgmPoolGetPageByHCPhys(pVM, HCPhysShw);
3869	if (!pPoolPage)
3870	{
3871	AssertMsgFailed(("Invalid page table address %RHp at %RGv! PdeDst=%#RX64\n",
3872	HCPhysShw, GCPtr, (uint64_t)PdeDst.u));
3873	cErrors++;
3874	continue;
3875	}
3876	const SHWPT pPTDst = (const SHWPT )PGMPOOL_PAGE_2_PTR(pVM, pPoolPage);
3877
3878	if (PdeDst.u & (X86_PDE4M_PWT \| X86_PDE4M_PCD))
3879	{
3880	AssertMsgFailed(("PDE flags PWT and/or PCD is set at %RGv! These flags are not virtualized! PdeDst=%#RX64\n",
3881	GCPtr, (uint64_t)PdeDst.u));
3882	cErrors++;
3883	}
3884
3885	if (PdeDst.u & (X86_PDE4M_G \| X86_PDE4M_D))
3886	{
3887	AssertMsgFailed(("4K PDE reserved flags at %RGv! PdeDst=%#RX64\n",
3888	GCPtr, (uint64_t)PdeDst.u));
3889	cErrors++;
3890	}
3891
3892	const GSTPDE PdeSrc = pPDSrc->a[(iPDDst >> (GST_PD_SHIFT - SHW_PD_SHIFT)) & GST_PD_MASK];
3893	if (!PdeSrc.n.u1Present)
3894	{
3895	AssertMsgFailed(("Guest PDE at %RGv is not present! PdeDst=%#RX64 PdeSrc=%#RX64\n",
3896	GCPtr, (uint64_t)PdeDst.u, (uint64_t)PdeSrc.u));
3897	cErrors++;
3898	continue;
3899	}
3900
3901	if ( !PdeSrc.b.u1Size
3902	\|\| !fBigPagesSupported)
3903	{
3904	GCPhysGst = PdeSrc.u & GST_PDE_PG_MASK;
3905	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
3906	GCPhysGst \|= (iPDDst & 1) * (PAGE_SIZE / 2);
3907	# endif
3908	}
3909	else
3910	{
3911	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3912	if (PdeSrc.u & X86_PDE4M_PG_HIGH_MASK)
3913	{
3914	AssertMsgFailed(("Guest PDE at %RGv is using PSE36 or similar! PdeSrc=%#RX64\n",
3915	GCPtr, (uint64_t)PdeSrc.u));
3916	cErrors++;
3917	continue;
3918	}
3919	# endif
3920	GCPhysGst = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc);
3921	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
3922	GCPhysGst \|= GCPtr & RT_BIT(X86_PAGE_2M_SHIFT);
3923	# endif
3924	}
3925
3926	if ( pPoolPage->enmKind
3927	!= (!PdeSrc.b.u1Size \|\| !fBigPagesSupported ? BTH_PGMPOOLKIND_PT_FOR_PT : BTH_PGMPOOLKIND_PT_FOR_BIG))
3928	{
3929	AssertMsgFailed(("Invalid shadow page table kind %d at %RGv! PdeSrc=%#RX64\n",
3930	pPoolPage->enmKind, GCPtr, (uint64_t)PdeSrc.u));
3931	cErrors++;
3932	}
3933
3934	PPGMPAGE pPhysPage = pgmPhysGetPage(pPGM, GCPhysGst);
3935	if (!pPhysPage)
3936	{
3937	AssertMsgFailed(("Cannot find guest physical address %RGp in the PDE at %RGv! PdeSrc=%#RX64\n",
3938	GCPhysGst, GCPtr, (uint64_t)PdeSrc.u));
3939	cErrors++;
3940	continue;
3941	}
3942
3943	if (GCPhysGst != pPoolPage->GCPhys)
3944	{
3945	AssertMsgFailed(("GCPhysGst=%RGp != pPage->GCPhys=%RGp at %RGv\n",
3946	GCPhysGst, pPoolPage->GCPhys, GCPtr));
3947	cErrors++;
3948	continue;
3949	}
3950
3951	if ( !PdeSrc.b.u1Size
3952	\|\| !fBigPagesSupported)
3953	{
3954	/*
3955	* Page Table.
3956	*/
3957	const GSTPT *pPTSrc;
3958	rc = PGM_GCPHYS_2_PTR(pVM, GCPhysGst & ~(RTGCPHYS)(PAGE_SIZE - 1), &pPTSrc);
3959	if (RT_FAILURE(rc))
3960	{
3961	AssertMsgFailed(("Cannot map/convert guest physical address %RGp in the PDE at %RGv! PdeSrc=%#RX64\n",
3962	GCPhysGst, GCPtr, (uint64_t)PdeSrc.u));
3963	cErrors++;
3964	continue;
3965	}
3966	if ( (PdeSrc.u & (X86_PDE_P \| X86_PDE_US \| X86_PDE_RW/* \| X86_PDE_A*/))
3967	!= (PdeDst.u & (X86_PDE_P \| X86_PDE_US \| X86_PDE_RW/* \| X86_PDE_A*/)))
3968	{
3969	/// @todo We get here a lot on out-of-sync CR3 entries. The access handler should zap them to avoid false alarms here!
3970	// (This problem will go away when/if we shadow multiple CR3s.)
3971	AssertMsgFailed(("4K PDE flags mismatch at %RGv! PdeSrc=%#RX64 PdeDst=%#RX64\n",
3972	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
3973	cErrors++;
3974	continue;
3975	}
3976	if (PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY)
3977	{
3978	AssertMsgFailed(("4K PDEs cannot have PGM_PDFLAGS_TRACK_DIRTY set! GCPtr=%RGv PdeDst=%#RX64\n",
3979	GCPtr, (uint64_t)PdeDst.u));
3980	cErrors++;
3981	continue;
3982	}
3983
3984	/* iterate the page table. */
3985	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
3986	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
3987	const unsigned offPTSrc = ((GCPtr >> SHW_PD_SHIFT) & 1) * 512;
3988	# else
3989	const unsigned offPTSrc = 0;
3990	# endif
3991	for (unsigned iPT = 0, off = 0;
3992	iPT < RT_ELEMENTS(pPTDst->a);
3993	iPT++, off += PAGE_SIZE)
3994	{
3995	const SHWPTE PteDst = pPTDst->a[iPT];
3996
3997	/* skip not-present entries. */
3998	if (!(PteDst.u & (X86_PTE_P \| PGM_PTFLAGS_TRACK_DIRTY))) /** @todo deal with ALL handlers and CSAM !P pages! */
3999	continue;
4000	Assert(PteDst.n.u1Present);
4001
4002	const GSTPTE PteSrc = pPTSrc->a[iPT + offPTSrc];
4003	if (!PteSrc.n.u1Present)
4004	{
4005	# ifdef IN_RING3
4006	PGMAssertHandlerAndFlagsInSync(pVM);
4007	PGMR3DumpHierarchyGC(pVM, cr3, cr4, (PdeSrc.u & GST_PDE_PG_MASK));
4008	# endif
4009	AssertMsgFailed(("Out of sync (!P) PTE at %RGv! PteSrc=%#RX64 PteDst=%#RX64 pPTSrc=%RGv iPTSrc=%x PdeSrc=%x physpte=%RGp\n",
4010	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u, pPTSrc, iPT + offPTSrc, PdeSrc.au32[0],
4011	(PdeSrc.u & GST_PDE_PG_MASK) + (iPT + offPTSrc)*sizeof(PteSrc)));
4012	cErrors++;
4013	continue;
4014	}
4015
4016	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;
4017	# if 1 /** @todo sync accessed bit properly... */
4018	fIgnoreFlags \|= X86_PTE_A;
4019	# endif
4020
4021	/* match the physical addresses */
4022	HCPhysShw = PteDst.u & SHW_PTE_PG_MASK;
4023	GCPhysGst = PteSrc.u & GST_PTE_PG_MASK;
4024
4025	# ifdef IN_RING3
4026	rc = PGMPhysGCPhys2HCPhys(pVM, GCPhysGst, &HCPhys);
4027	if (RT_FAILURE(rc))
4028	{
4029	if (HCPhysShw != MMR3PageDummyHCPhys(pVM))
4030	{
4031	AssertMsgFailed(("Cannot find guest physical address %RGp at %RGv! PteSrc=%#RX64 PteDst=%#RX64\n",
4032	GCPhysGst, GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4033	cErrors++;
4034	continue;
4035	}
4036	}
4037	else if (HCPhysShw != (HCPhys & SHW_PTE_PG_MASK))
4038	{
4039	AssertMsgFailed(("Out of sync (phys) at %RGv! HCPhysShw=%RHp HCPhys=%RHp GCPhysGst=%RGp PteSrc=%#RX64 PteDst=%#RX64\n",
4040	GCPtr + off, HCPhysShw, HCPhys, GCPhysGst, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4041	cErrors++;
4042	continue;
4043	}
4044	# endif
4045
4046	pPhysPage = pgmPhysGetPage(pPGM, GCPhysGst);
4047	if (!pPhysPage)
4048	{
4049	# ifdef IN_RING3 /** @todo make MMR3PageDummyHCPhys an 'All' function! */
4050	if (HCPhysShw != MMR3PageDummyHCPhys(pVM))
4051	{
4052	AssertMsgFailed(("Cannot find guest physical address %RGp at %RGv! PteSrc=%#RX64 PteDst=%#RX64\n",
4053	GCPhysGst, GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4054	cErrors++;
4055	continue;
4056	}
4057	# endif
4058	if (PteDst.n.u1Write)
4059	{
4060	AssertMsgFailed(("Invalid guest page at %RGv is writable! GCPhysGst=%RGp PteSrc=%#RX64 PteDst=%#RX64\n",
4061	GCPtr + off, GCPhysGst, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4062	cErrors++;
4063	}
4064	fIgnoreFlags \|= X86_PTE_RW;
4065	}
4066	else if (HCPhysShw != (PGM_PAGE_GET_HCPHYS(pPhysPage) & SHW_PTE_PG_MASK))
4067	{
4068	AssertMsgFailed(("Out of sync (phys) at %RGv! HCPhysShw=%RHp HCPhys=%RHp GCPhysGst=%RGp PteSrc=%#RX64 PteDst=%#RX64\n",
4069	GCPtr + off, HCPhysShw, pPhysPage->HCPhys, GCPhysGst, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4070	cErrors++;
4071	continue;
4072	}
4073
4074	/* flags */
4075	if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPhysPage))
4076	{
4077	if (!PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPhysPage))
4078	{
4079	if (PteDst.n.u1Write)
4080	{
4081	AssertMsgFailed(("WRITE access flagged at %RGv but the page is writable! HCPhys=%RHp PteSrc=%#RX64 PteDst=%#RX64\n",
4082	GCPtr + off, pPhysPage->HCPhys, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4083	cErrors++;
4084	continue;
4085	}
4086	fIgnoreFlags \|= X86_PTE_RW;
4087	}
4088	else
4089	{
4090	if (PteDst.n.u1Present)
4091	{
4092	AssertMsgFailed(("ALL access flagged at %RGv but the page is present! HCPhys=%RHp PteSrc=%#RX64 PteDst=%#RX64\n",
4093	GCPtr + off, pPhysPage->HCPhys, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4094	cErrors++;
4095	continue;
4096	}
4097	fIgnoreFlags \|= X86_PTE_P;
4098	}
4099	}
4100	else
4101	{
4102	if (!PteSrc.n.u1Dirty && PteSrc.n.u1Write)
4103	{
4104	if (PteDst.n.u1Write)
4105	{
4106	AssertMsgFailed(("!DIRTY page at %RGv is writable! PteSrc=%#RX64 PteDst=%#RX64\n",
4107	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4108	cErrors++;
4109	continue;
4110	}
4111	if (!(PteDst.u & PGM_PTFLAGS_TRACK_DIRTY))
4112	{
4113	AssertMsgFailed(("!DIRTY page at %RGv is not marked TRACK_DIRTY! PteSrc=%#RX64 PteDst=%#RX64\n",
4114	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4115	cErrors++;
4116	continue;
4117	}
4118	if (PteDst.n.u1Dirty)
4119	{
4120	AssertMsgFailed(("!DIRTY page at %RGv is marked DIRTY! PteSrc=%#RX64 PteDst=%#RX64\n",
4121	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4122	cErrors++;
4123	}
4124	# if 0 /** @todo sync access bit properly... */
4125	if (PteDst.n.u1Accessed != PteSrc.n.u1Accessed)
4126	{
4127	AssertMsgFailed(("!DIRTY page at %RGv is has mismatching accessed bit! PteSrc=%#RX64 PteDst=%#RX64\n",
4128	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4129	cErrors++;
4130	}
4131	fIgnoreFlags \|= X86_PTE_RW;
4132	# else
4133	fIgnoreFlags \|= X86_PTE_RW \| X86_PTE_A;
4134	# endif
4135	}
4136	else if (PteDst.u & PGM_PTFLAGS_TRACK_DIRTY)
4137	{
4138	/* access bit emulation (not implemented). */
4139	if (PteSrc.n.u1Accessed \|\| PteDst.n.u1Present)
4140	{
4141	AssertMsgFailed(("PGM_PTFLAGS_TRACK_DIRTY set at %RGv but no accessed bit emulation! PteSrc=%#RX64 PteDst=%#RX64\n",
4142	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4143	cErrors++;
4144	continue;
4145	}
4146	if (!PteDst.n.u1Accessed)
4147	{
4148	AssertMsgFailed(("!ACCESSED page at %RGv is has the accessed bit set! PteSrc=%#RX64 PteDst=%#RX64\n",
4149	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4150	cErrors++;
4151	}
4152	fIgnoreFlags \|= X86_PTE_P;
4153	}
4154	# ifdef DEBUG_sandervl
4155	fIgnoreFlags \|= X86_PTE_D \| X86_PTE_A;
4156	# endif
4157	}
4158
4159	if ( (PteSrc.u & ~fIgnoreFlags) != (PteDst.u & ~fIgnoreFlags)
4160	&& (PteSrc.u & ~(fIgnoreFlags \| X86_PTE_RW)) != (PteDst.u & ~fIgnoreFlags)
4161	)
4162	{
4163	AssertMsgFailed(("Flags mismatch at %RGv! %#RX64 != %#RX64 fIgnoreFlags=%#RX64 PteSrc=%#RX64 PteDst=%#RX64\n",
4164	GCPtr + off, (uint64_t)PteSrc.u & ~fIgnoreFlags, (uint64_t)PteDst.u & ~fIgnoreFlags,
4165	fIgnoreFlags, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
4166	cErrors++;
4167	continue;
4168	}
4169	} /* foreach PTE */
4170	}
4171	else
4172	{
4173	/*
4174	* Big Page.
4175	*/
4176	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;
4177	if (!PdeSrc.b.u1Dirty && PdeSrc.b.u1Write)
4178	{
4179	if (PdeDst.n.u1Write)
4180	{
4181	AssertMsgFailed(("!DIRTY page at %RGv is writable! PdeSrc=%#RX64 PdeDst=%#RX64\n",
4182	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
4183	cErrors++;
4184	continue;
4185	}
4186	if (!(PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY))
4187	{
4188	AssertMsgFailed(("!DIRTY page at %RGv is not marked TRACK_DIRTY! PteSrc=%#RX64 PteDst=%#RX64\n",
4189	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
4190	cErrors++;
4191	continue;
4192	}
4193	# if 0 /** @todo sync access bit properly... */
4194	if (PdeDst.n.u1Accessed != PdeSrc.b.u1Accessed)
4195	{
4196	AssertMsgFailed(("!DIRTY page at %RGv is has mismatching accessed bit! PteSrc=%#RX64 PteDst=%#RX64\n",
4197	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
4198	cErrors++;
4199	}
4200	fIgnoreFlags \|= X86_PTE_RW;
4201	# else
4202	fIgnoreFlags \|= X86_PTE_RW \| X86_PTE_A;
4203	# endif
4204	}
4205	else if (PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY)
4206	{
4207	/* access bit emulation (not implemented). */
4208	if (PdeSrc.b.u1Accessed \|\| PdeDst.n.u1Present)
4209	{
4210	AssertMsgFailed(("PGM_PDFLAGS_TRACK_DIRTY set at %RGv but no accessed bit emulation! PdeSrc=%#RX64 PdeDst=%#RX64\n",
4211	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
4212	cErrors++;
4213	continue;
4214	}
4215	if (!PdeDst.n.u1Accessed)
4216	{
4217	AssertMsgFailed(("!ACCESSED page at %RGv is has the accessed bit set! PdeSrc=%#RX64 PdeDst=%#RX64\n",
4218	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
4219	cErrors++;
4220	}
4221	fIgnoreFlags \|= X86_PTE_P;
4222	}
4223
4224	if ((PdeSrc.u & ~fIgnoreFlags) != (PdeDst.u & ~fIgnoreFlags))
4225	{
4226	AssertMsgFailed(("Flags mismatch (B) at %RGv! %#RX64 != %#RX64 fIgnoreFlags=%#RX64 PdeSrc=%#RX64 PdeDst=%#RX64\n",
4227	GCPtr, (uint64_t)PdeSrc.u & ~fIgnoreFlags, (uint64_t)PdeDst.u & ~fIgnoreFlags,
4228	fIgnoreFlags, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
4229	cErrors++;
4230	}
4231
4232	/* iterate the page table. */
4233	for (unsigned iPT = 0, off = 0;
4234	iPT < RT_ELEMENTS(pPTDst->a);
4235	iPT++, off += PAGE_SIZE, GCPhysGst += PAGE_SIZE)
4236	{
4237	const SHWPTE PteDst = pPTDst->a[iPT];
4238
4239	if (PteDst.u & PGM_PTFLAGS_TRACK_DIRTY)
4240	{
4241	AssertMsgFailed(("The PTE at %RGv emulating a 2/4M page is marked TRACK_DIRTY! PdeSrc=%#RX64 PteDst=%#RX64\n",
4242	GCPtr + off, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4243	cErrors++;
4244	}
4245
4246	/* skip not-present entries. */
4247	if (!PteDst.n.u1Present) /** @todo deal with ALL handlers and CSAM !P pages! */
4248	continue;
4249
4250	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;
4251
4252	/* match the physical addresses */
4253	HCPhysShw = PteDst.u & X86_PTE_PAE_PG_MASK;
4254
4255	# ifdef IN_RING3
4256	rc = PGMPhysGCPhys2HCPhys(pVM, GCPhysGst, &HCPhys);
4257	if (RT_FAILURE(rc))
4258	{
4259	if (HCPhysShw != MMR3PageDummyHCPhys(pVM))
4260	{
4261	AssertMsgFailed(("Cannot find guest physical address %RGp at %RGv! PdeSrc=%#RX64 PteDst=%#RX64\n",
4262	GCPhysGst, GCPtr + off, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4263	cErrors++;
4264	}
4265	}
4266	else if (HCPhysShw != (HCPhys & X86_PTE_PAE_PG_MASK))
4267	{
4268	AssertMsgFailed(("Out of sync (phys) at %RGv! HCPhysShw=%RHp HCPhys=%RHp GCPhysGst=%RGp PdeSrc=%#RX64 PteDst=%#RX64\n",
4269	GCPtr + off, HCPhysShw, HCPhys, GCPhysGst, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4270	cErrors++;
4271	continue;
4272	}
4273	# endif
4274	pPhysPage = pgmPhysGetPage(pPGM, GCPhysGst);
4275	if (!pPhysPage)
4276	{
4277	# ifdef IN_RING3 /** @todo make MMR3PageDummyHCPhys an 'All' function! */
4278	if (HCPhysShw != MMR3PageDummyHCPhys(pVM))
4279	{
4280	AssertMsgFailed(("Cannot find guest physical address %RGp at %RGv! PdeSrc=%#RX64 PteDst=%#RX64\n",
4281	GCPhysGst, GCPtr + off, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4282	cErrors++;
4283	continue;
4284	}
4285	# endif
4286	if (PteDst.n.u1Write)
4287	{
4288	AssertMsgFailed(("Invalid guest page at %RGv is writable! GCPhysGst=%RGp PdeSrc=%#RX64 PteDst=%#RX64\n",
4289	GCPtr + off, GCPhysGst, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4290	cErrors++;
4291	}
4292	fIgnoreFlags \|= X86_PTE_RW;
4293	}
4294	else if (HCPhysShw != (pPhysPage->HCPhys & X86_PTE_PAE_PG_MASK))
4295	{
4296	AssertMsgFailed(("Out of sync (phys) at %RGv! HCPhysShw=%RHp HCPhys=%RHp GCPhysGst=%RGp PdeSrc=%#RX64 PteDst=%#RX64\n",
4297	GCPtr + off, HCPhysShw, pPhysPage->HCPhys, GCPhysGst, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4298	cErrors++;
4299	continue;
4300	}
4301
4302	/* flags */
4303	if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPhysPage))
4304	{
4305	if (!PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPhysPage))
4306	{
4307	if (PGM_PAGE_GET_HNDL_PHYS_STATE(pPhysPage) != PGM_PAGE_HNDL_PHYS_STATE_DISABLED)
4308	{
4309	if (PteDst.n.u1Write)
4310	{
4311	AssertMsgFailed(("WRITE access flagged at %RGv but the page is writable! HCPhys=%RHp PdeSrc=%#RX64 PteDst=%#RX64\n",
4312	GCPtr + off, pPhysPage->HCPhys, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4313	cErrors++;
4314	continue;
4315	}
4316	fIgnoreFlags \|= X86_PTE_RW;
4317	}
4318	}
4319	else
4320	{
4321	if (PteDst.n.u1Present)
4322	{
4323	AssertMsgFailed(("ALL access flagged at %RGv but the page is present! HCPhys=%RHp PdeSrc=%#RX64 PteDst=%#RX64\n",
4324	GCPtr + off, pPhysPage->HCPhys, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4325	cErrors++;
4326	continue;
4327	}
4328	fIgnoreFlags \|= X86_PTE_P;
4329	}
4330	}
4331
4332	if ( (PdeSrc.u & ~fIgnoreFlags) != (PteDst.u & ~fIgnoreFlags)
4333	&& (PdeSrc.u & ~(fIgnoreFlags \| X86_PTE_RW)) != (PteDst.u & ~fIgnoreFlags) /* lazy phys handler dereg. */
4334	)
4335	{
4336	AssertMsgFailed(("Flags mismatch (BT) at %RGv! %#RX64 != %#RX64 fIgnoreFlags=%#RX64 PdeSrc=%#RX64 PteDst=%#RX64\n",
4337	GCPtr + off, (uint64_t)PdeSrc.u & ~fIgnoreFlags, (uint64_t)PteDst.u & ~fIgnoreFlags,
4338	fIgnoreFlags, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4339	cErrors++;
4340	continue;
4341	}
4342	} /* for each PTE */
4343	}
4344	}
4345	/* not present */
4346
4347	} /* for each PDE */
4348
4349	} /* for each PDPTE */
4350
4351	} /* for each PML4E */
4352
4353	# ifdef DEBUG
4354	if (cErrors)
4355	LogFlow(("AssertCR3: cErrors=%d\n", cErrors));
4356	# endif
4357
4358	#endif
4359	return cErrors;
4360
4361	#endif /* PGM_SHW_TYPE != PGM_TYPE_NESTED && PGM_SHW_TYPE != PGM_TYPE_EPT */
4362	}
4363	#endif /* VBOX_STRICT */
4364

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

source: vbox/trunk/src/VBox/VMM/VMMAll/PGMAllBth.h@ 13919

Download in other formats: