PGMAllBth.h@ 17616

Last change on this file since 17616 was 17616, checked in by vboxsync, 16 years ago
Inform SELM about shadow CR3 changes.
Property svn:eol-style set to `native` Property svn:keywords set to `Id`
File size: 187.7 KB

Line
1	/* $Id: PGMAllBth.h 17616 2009-03-10 11:08:39Z vboxsync $ */
2	/** @file
3	* VBox - Page Manager, Shadow+Guest Paging Template - All context code.
4	*
5	* This file is a big challenge!
6	*/
7
8	/*
9	* Copyright (C) 2006-2007 Sun Microsystems, Inc.
10	*
11	* This file is part of VirtualBox Open Source Edition (OSE), as
12	* available from http://www.virtualbox.org. This file is free software;
13	* you can redistribute it and/or modify it under the terms of the GNU
14	* General Public License (GPL) as published by the Free Software
15	* Foundation, in version 2 as it comes in the "COPYING" file of the
16	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
17	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
18	*
19	* Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa
20	* Clara, CA 95054 USA or visit http://www.sun.com if you need
21	* additional information or have any questions.
22	*/
23
24	/*******************************************************************************
25	* Internal Functions *
26	*******************************************************************************/
27	__BEGIN_DECLS
28	PGM_BTH_DECL(int, Trap0eHandler)(PVM pVM, RTGCUINT uErr, PCPUMCTXCORE pRegFrame, RTGCPTR pvFault);
29	PGM_BTH_DECL(int, InvalidatePage)(PVM pVM, RTGCPTR GCPtrPage);
30	PGM_BTH_DECL(int, SyncPage)(PVM pVM, GSTPDE PdeSrc, RTGCPTR GCPtrPage, unsigned cPages, unsigned uErr);
31	PGM_BTH_DECL(int, CheckPageFault)(PVM pVM, uint32_t uErr, PSHWPDE pPdeDst, PGSTPDE pPdeSrc, RTGCPTR GCPtrPage);
32	PGM_BTH_DECL(int, SyncPT)(PVM pVM, unsigned iPD, PGSTPD pPDSrc, RTGCPTR GCPtrPage);
33	PGM_BTH_DECL(int, VerifyAccessSyncPage)(PVM pVM, RTGCPTR Addr, unsigned fPage, unsigned uErr);
34	PGM_BTH_DECL(int, PrefetchPage)(PVM pVM, RTGCPTR GCPtrPage);
35	PGM_BTH_DECL(int, SyncCR3)(PVM pVM, uint64_t cr0, uint64_t cr3, uint64_t cr4, bool fGlobal);
36	#ifdef VBOX_STRICT
37	PGM_BTH_DECL(unsigned, AssertCR3)(PVM pVM, uint64_t cr3, uint64_t cr4, RTGCPTR GCPtr = 0, RTGCPTR cb = ~(RTGCPTR)0);
38	#endif
39	#ifdef PGMPOOL_WITH_USER_TRACKING
40	DECLINLINE(void) PGM_BTH_NAME(SyncPageWorkerTrackDeref)(PVM pVM, PPGMPOOLPAGE pShwPage, RTHCPHYS HCPhys);
41	#endif
42	PGM_BTH_DECL(int, MapCR3)(PVM pVM, RTGCPHYS GCPhysCR3);
43	PGM_BTH_DECL(int, UnmapCR3)(PVM pVM);
44	__END_DECLS
45
46
47	/* Filter out some illegal combinations of guest and shadow paging, so we can remove redundant checks inside functions. */
48	#if PGM_GST_TYPE == PGM_TYPE_PAE && PGM_SHW_TYPE != PGM_TYPE_PAE && PGM_SHW_TYPE != PGM_TYPE_NESTED && PGM_SHW_TYPE != PGM_TYPE_EPT
49	# error "Invalid combination; PAE guest implies PAE shadow"
50	#endif
51
52	#if (PGM_GST_TYPE == PGM_TYPE_REAL \|\| PGM_GST_TYPE == PGM_TYPE_PROT) \
53	&& !(PGM_SHW_TYPE == PGM_TYPE_32BIT \|\| PGM_SHW_TYPE == PGM_TYPE_PAE \|\| PGM_SHW_TYPE == PGM_TYPE_AMD64 \|\| PGM_SHW_TYPE == PGM_TYPE_NESTED \|\| PGM_SHW_TYPE == PGM_TYPE_EPT)
54	# error "Invalid combination; real or protected mode without paging implies 32 bits or PAE shadow paging."
55	#endif
56
57	#if (PGM_GST_TYPE == PGM_TYPE_32BIT \|\| PGM_GST_TYPE == PGM_TYPE_PAE) \
58	&& !(PGM_SHW_TYPE == PGM_TYPE_32BIT \|\| PGM_SHW_TYPE == PGM_TYPE_PAE \|\| PGM_SHW_TYPE == PGM_TYPE_NESTED \|\| PGM_SHW_TYPE == PGM_TYPE_EPT)
59	# error "Invalid combination; 32 bits guest paging or PAE implies 32 bits or PAE shadow paging."
60	#endif
61
62	#if (PGM_GST_TYPE == PGM_TYPE_AMD64 && PGM_SHW_TYPE != PGM_TYPE_AMD64 && PGM_SHW_TYPE != PGM_TYPE_NESTED && PGM_SHW_TYPE != PGM_TYPE_EPT) \
63	\|\| (PGM_SHW_TYPE == PGM_TYPE_AMD64 && PGM_GST_TYPE != PGM_TYPE_AMD64 && PGM_GST_TYPE != PGM_TYPE_PROT)
64	# error "Invalid combination; AMD64 guest implies AMD64 shadow and vice versa"
65	#endif
66
67	#ifdef IN_RING0 /* no mappings in VT-x and AMD-V mode */
68	# define PGM_WITHOUT_MAPPINGS
69	#endif
70
71
72	#ifndef IN_RING3
73	/**
74	* #PF Handler for raw-mode guest execution.
75	*
76	* @returns VBox status code (appropriate for trap handling and GC return).
77	* @param pVM VM Handle.
78	* @param uErr The trap error code.
79	* @param pRegFrame Trap register frame.
80	* @param pvFault The fault address.
81	*/
82	PGM_BTH_DECL(int, Trap0eHandler)(PVM pVM, RTGCUINT uErr, PCPUMCTXCORE pRegFrame, RTGCPTR pvFault)
83	{
84	# if defined(IN_RC) && defined(VBOX_STRICT)
85	PGMDynCheckLocks(pVM);
86	# endif
87
88	# if (PGM_GST_TYPE == PGM_TYPE_32BIT \|\| PGM_GST_TYPE == PGM_TYPE_REAL \|\| PGM_GST_TYPE == PGM_TYPE_PROT \|\| PGM_GST_TYPE == PGM_TYPE_PAE \|\| PGM_GST_TYPE == PGM_TYPE_AMD64) \
89	&& PGM_SHW_TYPE != PGM_TYPE_NESTED \
90	&& (PGM_SHW_TYPE != PGM_TYPE_EPT \|\| PGM_GST_TYPE == PGM_TYPE_PROT)
91
92	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE != PGM_TYPE_PAE
93	/*
94	* Hide the instruction fetch trap indicator for now.
95	*/
96	/** @todo NXE will change this and we must fix NXE in the switcher too! */
97	if (uErr & X86_TRAP_PF_ID)
98	{
99	uErr &= ~X86_TRAP_PF_ID;
100	TRPMSetErrorCode(pVM, uErr);
101	}
102	# endif
103
104	/*
105	* Get PDs.
106	*/
107	int rc;
108	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
109	# if PGM_GST_TYPE == PGM_TYPE_32BIT
110	const unsigned iPDSrc = pvFault >> GST_PD_SHIFT;
111	PGSTPD pPDSrc = pgmGstGet32bitPDPtr(&pVM->pgm.s);
112
113	# elif PGM_GST_TYPE == PGM_TYPE_PAE \|\| PGM_GST_TYPE == PGM_TYPE_AMD64
114
115	# if PGM_GST_TYPE == PGM_TYPE_PAE
116	unsigned iPDSrc;
117	X86PDPE PdpeSrc;
118	PGSTPD pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, pvFault, &iPDSrc, &PdpeSrc);
119
120	# elif PGM_GST_TYPE == PGM_TYPE_AMD64
121	unsigned iPDSrc;
122	PX86PML4E pPml4eSrc;
123	X86PDPE PdpeSrc;
124	PGSTPD pPDSrc;
125
126	pPDSrc = pgmGstGetLongModePDPtr(&pVM->pgm.s, pvFault, &pPml4eSrc, &PdpeSrc, &iPDSrc);
127	Assert(pPml4eSrc);
128	# endif
129
130	/* Quick check for a valid guest trap. (PAE & AMD64) */
131	if (!pPDSrc)
132	{
133	# if PGM_GST_TYPE == PGM_TYPE_AMD64 && GC_ARCH_BITS == 64
134	LogFlow(("Trap0eHandler: guest PML4 %d not present CR3=%RGp\n", (int)((pvFault >> X86_PML4_SHIFT) & X86_PML4_MASK), CPUMGetGuestCR3(pVM) & X86_CR3_PAGE_MASK));
135	# else
136	LogFlow(("Trap0eHandler: guest iPDSrc=%u not present CR3=%RGp\n", iPDSrc, CPUMGetGuestCR3(pVM) & X86_CR3_PAGE_MASK));
137	# endif
138	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2GuestTrap; });
139	TRPMSetErrorCode(pVM, uErr);
140	return VINF_EM_RAW_GUEST_TRAP;
141	}
142	# endif
143
144	# else /* !PGM_WITH_PAGING */
145	PGSTPD pPDSrc = NULL;
146	const unsigned iPDSrc = 0;
147	# endif /* !PGM_WITH_PAGING */
148
149
150	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
151	const unsigned iPDDst = pvFault >> SHW_PD_SHIFT;
152	PX86PD pPDDst = pgmShwGet32BitPDPtr(&pVM->pgm.s);
153
154	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
155	const unsigned iPDDst = (pvFault >> SHW_PD_SHIFT) & SHW_PD_MASK; /* pPDDst index, not used with the pool. */
156
157	PX86PDPAE pPDDst;
158	# if PGM_GST_TYPE != PGM_TYPE_PAE
159	X86PDPE PdpeSrc;
160
161	/* Fake PDPT entry; access control handled on the page table level, so allow everything. */
162	PdpeSrc.u = X86_PDPE_P; /* rw/us are reserved for PAE pdpte's; accessed bit causes invalid VT-x guest state errors */
163	# endif
164	rc = pgmShwSyncPaePDPtr(pVM, pvFault, &PdpeSrc, &pPDDst);
165	if (rc != VINF_SUCCESS)
166	{
167	AssertRC(rc);
168	return rc;
169	}
170	Assert(pPDDst);
171
172	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
173	const unsigned iPDDst = ((pvFault >> SHW_PD_SHIFT) & SHW_PD_MASK);
174	PX86PDPAE pPDDst;
175	# if PGM_GST_TYPE == PGM_TYPE_PROT
176	/* AMD-V nested paging */
177	X86PML4E Pml4eSrc;
178	X86PDPE PdpeSrc;
179	PX86PML4E pPml4eSrc = &Pml4eSrc;
180
181	/* Fake PML4 & PDPT entry; access control handled on the page table level, so allow everything. */
182	Pml4eSrc.u = X86_PML4E_P \| X86_PML4E_RW \| X86_PML4E_US \| X86_PML4E_A;
183	PdpeSrc.u = X86_PDPE_P \| X86_PDPE_RW \| X86_PDPE_US \| X86_PDPE_A;
184	# endif
185
186	rc = pgmShwSyncLongModePDPtr(pVM, pvFault, pPml4eSrc, &PdpeSrc, &pPDDst);
187	if (rc != VINF_SUCCESS)
188	{
189	AssertRC(rc);
190	return rc;
191	}
192	Assert(pPDDst);
193
194	# elif PGM_SHW_TYPE == PGM_TYPE_EPT
195	const unsigned iPDDst = ((pvFault >> SHW_PD_SHIFT) & SHW_PD_MASK);
196	PEPTPD pPDDst;
197
198	rc = pgmShwGetEPTPDPtr(pVM, pvFault, NULL, &pPDDst);
199	if (rc != VINF_SUCCESS)
200	{
201	AssertRC(rc);
202	return rc;
203	}
204	Assert(pPDDst);
205	# endif
206
207	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
208	/*
209	* If we successfully correct the write protection fault due to dirty bit
210	* tracking, or this page fault is a genuine one, then return immediately.
211	*/
212	STAM_PROFILE_START(&pVM->pgm.s.StatRZTrap0eTimeCheckPageFault, e);
213	rc = PGM_BTH_NAME(CheckPageFault)(pVM, uErr, &pPDDst->a[iPDDst], &pPDSrc->a[iPDSrc], pvFault);
214	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeCheckPageFault, e);
215	if ( rc == VINF_PGM_HANDLED_DIRTY_BIT_FAULT
216	\|\| rc == VINF_EM_RAW_GUEST_TRAP)
217	{
218	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution)
219	= rc == VINF_PGM_HANDLED_DIRTY_BIT_FAULT ? &pVM->pgm.s.StatRZTrap0eTime2DirtyAndAccessed : &pVM->pgm.s.StatRZTrap0eTime2GuestTrap; });
220	LogBird(("Trap0eHandler: returns %s\n", rc == VINF_PGM_HANDLED_DIRTY_BIT_FAULT ? "VINF_SUCCESS" : "VINF_EM_RAW_GUEST_TRAP"));
221	return rc == VINF_PGM_HANDLED_DIRTY_BIT_FAULT ? VINF_SUCCESS : rc;
222	}
223
224	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0ePD[iPDSrc]);
225	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
226
227	/*
228	* A common case is the not-present error caused by lazy page table syncing.
229	*
230	* It is IMPORTANT that we weed out any access to non-present shadow PDEs here
231	* so we can safely assume that the shadow PT is present when calling SyncPage later.
232	*
233	* On failure, we ASSUME that SyncPT is out of memory or detected some kind
234	* of mapping conflict and defer to SyncCR3 in R3.
235	* (Again, we do NOT support access handlers for non-present guest pages.)
236	*
237	*/
238	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
239	GSTPDE PdeSrc = pPDSrc->a[iPDSrc];
240	# else
241	GSTPDE PdeSrc;
242	PdeSrc.au32[0] = 0; /* faked so we don't have to #ifdef everything */
243	PdeSrc.n.u1Present = 1;
244	PdeSrc.n.u1Write = 1;
245	PdeSrc.n.u1Accessed = 1;
246	PdeSrc.n.u1User = 1;
247	# endif
248	if ( !(uErr & X86_TRAP_PF_P) /* not set means page not present instead of page protection violation */
249	&& !pPDDst->a[iPDDst].n.u1Present
250	&& PdeSrc.n.u1Present
251	)
252
253	{
254	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2SyncPT; });
255	STAM_PROFILE_START(&pVM->pgm.s.StatRZTrap0eTimeSyncPT, f);
256	LogFlow(("=>SyncPT %04x = %08x\n", iPDSrc, PdeSrc.au32[0]));
257	rc = PGM_BTH_NAME(SyncPT)(pVM, iPDSrc, pPDSrc, pvFault);
258	if (RT_SUCCESS(rc))
259	{
260	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeSyncPT, f);
261	return rc;
262	}
263	Log(("SyncPT: %d failed!! rc=%d\n", iPDSrc, rc));
264	VM_FF_SET(pVM, VM_FF_PGM_SYNC_CR3); /** @todo no need to do global sync, right? */
265	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeSyncPT, f);
266	return VINF_PGM_SYNC_CR3;
267	}
268
269	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
270	/*
271	* Check if this address is within any of our mappings.
272	*
273	* This is very fast and it's gonna save us a bit of effort below and prevent
274	* us from screwing ourself with MMIO2 pages which have a GC Mapping (VRam).
275	* (BTW, it's impossible to have physical access handlers in a mapping.)
276	*/
277	if (pgmMapAreMappingsEnabled(&pVM->pgm.s))
278	{
279	STAM_PROFILE_START(&pVM->pgm.s.StatRZTrap0eTimeMapping, a);
280	PPGMMAPPING pMapping = pVM->pgm.s.CTX_SUFF(pMappings);
281	for ( ; pMapping; pMapping = pMapping->CTX_SUFF(pNext))
282	{
283	if (pvFault < pMapping->GCPtr)
284	break;
285	if (pvFault - pMapping->GCPtr < pMapping->cb)
286	{
287	/*
288	* The first thing we check is if we've got an undetected conflict.
289	*/
290	if (!pVM->pgm.s.fMappingsFixed)
291	{
292	unsigned iPT = pMapping->cb >> GST_PD_SHIFT;
293	while (iPT-- > 0)
294	if (pPDSrc->a[iPDSrc + iPT].n.u1Present)
295	{
296	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eConflicts);
297	Log(("Trap0e: Detected Conflict %RGv-%RGv\n", pMapping->GCPtr, pMapping->GCPtrLast));
298	VM_FF_SET(pVM, VM_FF_PGM_SYNC_CR3); /** @todo no need to do global sync,right? */
299	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeMapping, a);
300	return VINF_PGM_SYNC_CR3;
301	}
302	}
303
304	/*
305	* Check if the fault address is in a virtual page access handler range.
306	*/
307	PPGMVIRTHANDLER pCur = (PPGMVIRTHANDLER)RTAvlroGCPtrRangeGet(&pVM->pgm.s.CTX_SUFF(pTrees)->HyperVirtHandlers, pvFault);
308	if ( pCur
309	&& pvFault - pCur->Core.Key < pCur->cb
310	&& uErr & X86_TRAP_PF_RW)
311	{
312	# ifdef IN_RC
313	STAM_PROFILE_START(&pCur->Stat, h);
314	rc = pCur->CTX_SUFF(pfnHandler)(pVM, uErr, pRegFrame, pvFault, pCur->Core.Key, pvFault - pCur->Core.Key);
315	STAM_PROFILE_STOP(&pCur->Stat, h);
316	# else
317	AssertFailed();
318	rc = VINF_EM_RAW_EMULATE_INSTR; /* can't happen with VMX */
319	# endif
320	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersMapping);
321	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeMapping, a);
322	return rc;
323	}
324
325	/*
326	* Pretend we're not here and let the guest handle the trap.
327	*/
328	TRPMSetErrorCode(pVM, uErr & ~X86_TRAP_PF_P);
329	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eGuestPFMapping);
330	LogFlow(("PGM: Mapping access -> route trap to recompiler!\n"));
331	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeMapping, a);
332	return VINF_EM_RAW_GUEST_TRAP;
333	}
334	}
335	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeMapping, a);
336	} /* pgmAreMappingsEnabled(&pVM->pgm.s) */
337	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
338
339	/*
340	* Check if this fault address is flagged for special treatment,
341	* which means we'll have to figure out the physical address and
342	* check flags associated with it.
343	*
344	* ASSUME that we can limit any special access handling to pages
345	* in page tables which the guest believes to be present.
346	*/
347	if (PdeSrc.n.u1Present)
348	{
349	RTGCPHYS GCPhys = NIL_RTGCPHYS;
350
351	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
352	# if PGM_GST_TYPE == PGM_TYPE_AMD64
353	bool fBigPagesSupported = true;
354	# else
355	bool fBigPagesSupported = !!(CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
356	# endif
357	if ( PdeSrc.b.u1Size
358	&& fBigPagesSupported)
359	GCPhys = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc)
360	\| ((RTGCPHYS)pvFault & (GST_BIG_PAGE_OFFSET_MASK ^ PAGE_OFFSET_MASK));
361	else
362	{
363	PGSTPT pPTSrc;
364	rc = PGM_GCPHYS_2_PTR(pVM, PdeSrc.u & GST_PDE_PG_MASK, &pPTSrc);
365	if (RT_SUCCESS(rc))
366	{
367	unsigned iPTESrc = (pvFault >> GST_PT_SHIFT) & GST_PT_MASK;
368	if (pPTSrc->a[iPTESrc].n.u1Present)
369	GCPhys = pPTSrc->a[iPTESrc].u & GST_PTE_PG_MASK;
370	}
371	}
372	# else
373	/* No paging so the fault address is the physical address */
374	GCPhys = (RTGCPHYS)(pvFault & ~PAGE_OFFSET_MASK);
375	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
376
377	/*
378	* If we have a GC address we'll check if it has any flags set.
379	*/
380	if (GCPhys != NIL_RTGCPHYS)
381	{
382	STAM_PROFILE_START(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
383
384	PPGMPAGE pPage;
385	rc = pgmPhysGetPageEx(&pVM->pgm.s, GCPhys, &pPage);
386	if (RT_SUCCESS(rc)) /** just handle the failure immediate (it returns) and make things easier to read. */
387	{
388	if ( PGM_PAGE_HAS_ACTIVE_PHYSICAL_HANDLERS(pPage)
389	\|\| PGM_PAGE_HAS_ACTIVE_VIRTUAL_HANDLERS(pPage))
390	{
391	if (PGM_PAGE_HAS_ANY_PHYSICAL_HANDLERS(pPage))
392	{
393	/*
394	* Physical page access handler.
395	*/
396	const RTGCPHYS GCPhysFault = GCPhys \| (pvFault & PAGE_OFFSET_MASK);
397	PPGMPHYSHANDLER pCur = (PPGMPHYSHANDLER)RTAvlroGCPhysRangeGet(&pVM->pgm.s.CTX_SUFF(pTrees)->PhysHandlers, GCPhysFault);
398	if (pCur)
399	{
400	# ifdef PGM_SYNC_N_PAGES
401	/*
402	* If the region is write protected and we got a page not present fault, then sync
403	* the pages. If the fault was caused by a read, then restart the instruction.
404	* In case of write access continue to the GC write handler.
405	*
406	* ASSUMES that there is only one handler per page or that they have similar write properties.
407	*/
408	if ( pCur->enmType == PGMPHYSHANDLERTYPE_PHYSICAL_WRITE
409	&& !(uErr & X86_TRAP_PF_P))
410	{
411	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, pvFault, PGM_SYNC_NR_PAGES, uErr);
412	if ( RT_FAILURE(rc)
413	\|\| !(uErr & X86_TRAP_PF_RW)
414	\|\| rc == VINF_PGM_SYNCPAGE_MODIFIED_PDE)
415	{
416	AssertRC(rc);
417	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersOutOfSync);
418	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
419	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2OutOfSyncHndPhys; });
420	return rc;
421	}
422	}
423	# endif
424
425	AssertMsg( pCur->enmType != PGMPHYSHANDLERTYPE_PHYSICAL_WRITE
426	\|\| (pCur->enmType == PGMPHYSHANDLERTYPE_PHYSICAL_WRITE && (uErr & X86_TRAP_PF_RW)),
427	("Unexpected trap for physical handler: %08X (phys=%08x) pPage=%R[pgmpage] uErr=%X, enum=%d\n", pvFault, GCPhys, pPage, uErr, pCur->enmType));
428
429	# if defined(IN_RC) \|\| defined(IN_RING0)
430	if (pCur->CTX_SUFF(pfnHandler))
431	{
432	STAM_PROFILE_START(&pCur->Stat, h);
433	rc = pCur->CTX_SUFF(pfnHandler)(pVM, uErr, pRegFrame, pvFault, GCPhysFault, pCur->CTX_SUFF(pvUser));
434	STAM_PROFILE_STOP(&pCur->Stat, h);
435	}
436	else
437	# endif
438	rc = VINF_EM_RAW_EMULATE_INSTR;
439	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersPhysical);
440	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
441	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2HndPhys; });
442	return rc;
443	}
444	}
445	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
446	else
447	{
448	# ifdef PGM_SYNC_N_PAGES
449	/*
450	* If the region is write protected and we got a page not present fault, then sync
451	* the pages. If the fault was caused by a read, then restart the instruction.
452	* In case of write access continue to the GC write handler.
453	*/
454	if ( PGM_PAGE_GET_HNDL_VIRT_STATE(pPage) < PGM_PAGE_HNDL_PHYS_STATE_ALL
455	&& !(uErr & X86_TRAP_PF_P))
456	{
457	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, pvFault, PGM_SYNC_NR_PAGES, uErr);
458	if ( RT_FAILURE(rc)
459	\|\| rc == VINF_PGM_SYNCPAGE_MODIFIED_PDE
460	\|\| !(uErr & X86_TRAP_PF_RW))
461	{
462	AssertRC(rc);
463	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersOutOfSync);
464	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
465	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2OutOfSyncHndVirt; });
466	return rc;
467	}
468	}
469	# endif
470	/*
471	* Ok, it's an virtual page access handler.
472	*
473	* Since it's faster to search by address, we'll do that first
474	* and then retry by GCPhys if that fails.
475	*/
476	/** @todo r=bird: perhaps we should consider looking up by physical address directly now? */
477	/** @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
478	* page was changed without us noticing it (not-present -> present without invlpg or mov cr3, xxx)
479	*/
480	PPGMVIRTHANDLER pCur = (PPGMVIRTHANDLER)RTAvlroGCPtrRangeGet(&pVM->pgm.s.CTX_SUFF(pTrees)->VirtHandlers, pvFault);
481	if (pCur)
482	{
483	AssertMsg(!(pvFault - pCur->Core.Key < pCur->cb)
484	\|\| ( pCur->enmType != PGMVIRTHANDLERTYPE_WRITE
485	\|\| !(uErr & X86_TRAP_PF_P)
486	\|\| (pCur->enmType == PGMVIRTHANDLERTYPE_WRITE && (uErr & X86_TRAP_PF_RW))),
487	("Unexpected trap for virtual handler: %RGv (phys=%RGp) pPage=%R[pgmpage] uErr=%X, enum=%d\n", pvFault, GCPhys, pPage, uErr, pCur->enmType));
488
489	if ( pvFault - pCur->Core.Key < pCur->cb
490	&& ( uErr & X86_TRAP_PF_RW
491	\|\| pCur->enmType != PGMVIRTHANDLERTYPE_WRITE ) )
492	{
493	# ifdef IN_RC
494	STAM_PROFILE_START(&pCur->Stat, h);
495	rc = pCur->CTX_SUFF(pfnHandler)(pVM, uErr, pRegFrame, pvFault, pCur->Core.Key, pvFault - pCur->Core.Key);
496	STAM_PROFILE_STOP(&pCur->Stat, h);
497	# else
498	rc = VINF_EM_RAW_EMULATE_INSTR; /** @todo for VMX */
499	# endif
500	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersVirtual);
501	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
502	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2HndVirt; });
503	return rc;
504	}
505	/* Unhandled part of a monitored page */
506	}
507	else
508	{
509	/* Check by physical address. */
510	PPGMVIRTHANDLER pCur;
511	unsigned iPage;
512	rc = pgmHandlerVirtualFindByPhysAddr(pVM, GCPhys + (pvFault & PAGE_OFFSET_MASK),
513	&pCur, &iPage);
514	Assert(RT_SUCCESS(rc) \|\| !pCur);
515	if ( pCur
516	&& ( uErr & X86_TRAP_PF_RW
517	\|\| pCur->enmType != PGMVIRTHANDLERTYPE_WRITE ) )
518	{
519	Assert((pCur->aPhysToVirt[iPage].Core.Key & X86_PTE_PAE_PG_MASK) == GCPhys);
520	# ifdef IN_RC
521	RTGCPTR off = (iPage << PAGE_SHIFT) + (pvFault & PAGE_OFFSET_MASK) - (pCur->Core.Key & PAGE_OFFSET_MASK);
522	Assert(off < pCur->cb);
523	STAM_PROFILE_START(&pCur->Stat, h);
524	rc = pCur->CTX_SUFF(pfnHandler)(pVM, uErr, pRegFrame, pvFault, pCur->Core.Key, off);
525	STAM_PROFILE_STOP(&pCur->Stat, h);
526	# else
527	rc = VINF_EM_RAW_EMULATE_INSTR; /** @todo for VMX */
528	# endif
529	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersVirtualByPhys);
530	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
531	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2HndVirt; });
532	return rc;
533	}
534	}
535	}
536	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
537
538	/*
539	* There is a handled area of the page, but this fault doesn't belong to it.
540	* We must emulate the instruction.
541	*
542	* To avoid crashing (non-fatal) in the interpreter and go back to the recompiler
543	* we first check if this was a page-not-present fault for a page with only
544	* write access handlers. Restart the instruction if it wasn't a write access.
545	*/
546	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersUnhandled);
547
548	if ( !PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage)
549	&& !(uErr & X86_TRAP_PF_P))
550	{
551	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, pvFault, PGM_SYNC_NR_PAGES, uErr);
552	if ( RT_FAILURE(rc)
553	\|\| rc == VINF_PGM_SYNCPAGE_MODIFIED_PDE
554	\|\| !(uErr & X86_TRAP_PF_RW))
555	{
556	AssertRC(rc);
557	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersOutOfSync);
558	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
559	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2OutOfSyncHndPhys; });
560	return rc;
561	}
562	}
563
564	/** @todo This particular case can cause quite a lot of overhead. E.g. early stage of kernel booting in Ubuntu 6.06
565	* It's writing to an unhandled part of the LDT page several million times.
566	*/
567	rc = PGMInterpretInstruction(pVM, pRegFrame, pvFault);
568	LogFlow(("PGM: PGMInterpretInstruction -> rc=%d pPage=%R[pgmpage]\n", rc, pPage));
569	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
570	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2HndUnhandled; });
571	return rc;
572	} /* if any kind of handler */
573
574	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
575	if (uErr & X86_TRAP_PF_P)
576	{
577	/*
578	* The page isn't marked, but it might still be monitored by a virtual page access handler.
579	* (ASSUMES no temporary disabling of virtual handlers.)
580	*/
581	/** @todo r=bird: Since the purpose is to catch out of sync pages with virtual handler(s) here,
582	* we should correct both the shadow page table and physical memory flags, and not only check for
583	* accesses within the handler region but for access to pages with virtual handlers. */
584	PPGMVIRTHANDLER pCur = (PPGMVIRTHANDLER)RTAvlroGCPtrRangeGet(&pVM->pgm.s.CTX_SUFF(pTrees)->VirtHandlers, pvFault);
585	if (pCur)
586	{
587	AssertMsg( !(pvFault - pCur->Core.Key < pCur->cb)
588	\|\| ( pCur->enmType != PGMVIRTHANDLERTYPE_WRITE
589	\|\| !(uErr & X86_TRAP_PF_P)
590	\|\| (pCur->enmType == PGMVIRTHANDLERTYPE_WRITE && (uErr & X86_TRAP_PF_RW))),
591	("Unexpected trap for virtual handler: %08X (phys=%08x) %R[pgmpage] uErr=%X, enum=%d\n", pvFault, GCPhys, pPage, uErr, pCur->enmType));
592
593	if ( pvFault - pCur->Core.Key < pCur->cb
594	&& ( uErr & X86_TRAP_PF_RW
595	\|\| pCur->enmType != PGMVIRTHANDLERTYPE_WRITE ) )
596	{
597	# ifdef IN_RC
598	STAM_PROFILE_START(&pCur->Stat, h);
599	rc = pCur->CTX_SUFF(pfnHandler)(pVM, uErr, pRegFrame, pvFault, pCur->Core.Key, pvFault - pCur->Core.Key);
600	STAM_PROFILE_STOP(&pCur->Stat, h);
601	# else
602	rc = VINF_EM_RAW_EMULATE_INSTR; /** @todo for VMX */
603	# endif
604	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersVirtualUnmarked);
605	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
606	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2HndVirt; });
607	return rc;
608	}
609	}
610	}
611	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
612	}
613	else
614	{
615	/*
616	* When the guest accesses invalid physical memory (e.g. probing
617	* of RAM or accessing a remapped MMIO range), then we'll fall
618	* back to the recompiler to emulate the instruction.
619	*/
620	LogFlow(("PGM #PF: pgmPhysGetPageEx(%RGp) failed with %Rrc\n", GCPhys, rc));
621	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eHandlersInvalid);
622	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
623	return VINF_EM_RAW_EMULATE_INSTR;
624	}
625
626	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeHandlers, b);
627
628	# ifdef PGM_OUT_OF_SYNC_IN_GC /** @todo remove this bugger. */
629	/*
630	* We are here only if page is present in Guest page tables and
631	* trap is not handled by our handlers.
632	*
633	* Check it for page out-of-sync situation.
634	*/
635	STAM_PROFILE_START(&pVM->pgm.s.StatRZTrap0eTimeOutOfSync, c);
636
637	if (!(uErr & X86_TRAP_PF_P))
638	{
639	/*
640	* Page is not present in our page tables.
641	* Try to sync it!
642	* BTW, fPageShw is invalid in this branch!
643	*/
644	if (uErr & X86_TRAP_PF_US)
645	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,PageOutOfSyncUser));
646	else /* supervisor */
647	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,PageOutOfSyncSupervisor));
648
649	# if defined(LOG_ENABLED) && !defined(IN_RING0)
650	RTGCPHYS GCPhys;
651	uint64_t fPageGst;
652	PGMGstGetPage(pVM, pvFault, &fPageGst, &GCPhys);
653	Log(("Page out of sync: %RGv eip=%08x PdeSrc.n.u1User=%d fPageGst=%08llx GCPhys=%RGp scan=%d\n",
654	pvFault, pRegFrame->eip, PdeSrc.n.u1User, fPageGst, GCPhys, CSAMDoesPageNeedScanning(pVM, (RTRCPTR)pRegFrame->eip)));
655	# endif /* LOG_ENABLED */
656
657	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) && !defined(IN_RING0)
658	if (CPUMGetGuestCPL(pVM, pRegFrame) == 0)
659	{
660	uint64_t fPageGst;
661	rc = PGMGstGetPage(pVM, pvFault, &fPageGst, NULL);
662	if ( RT_SUCCESS(rc)
663	&& !(fPageGst & X86_PTE_US))
664	{
665	/* Note: can't check for X86_TRAP_ID bit, because that requires execute disable support on the CPU */
666	if ( pvFault == (RTGCPTR)pRegFrame->eip
667	\|\| pvFault - pRegFrame->eip < 8 /* instruction crossing a page boundary */
668	# ifdef CSAM_DETECT_NEW_CODE_PAGES
669	\|\| ( !PATMIsPatchGCAddr(pVM, (RTGCPTR)pRegFrame->eip)
670	&& CSAMDoesPageNeedScanning(pVM, (RTRCPTR)pRegFrame->eip)) /* any new code we encounter here */
671	# endif /* CSAM_DETECT_NEW_CODE_PAGES */
672	)
673	{
674	LogFlow(("CSAMExecFault %RX32\n", pRegFrame->eip));
675	rc = CSAMExecFault(pVM, (RTRCPTR)pRegFrame->eip);
676	if (rc != VINF_SUCCESS)
677	{
678	/*
679	* CSAM needs to perform a job in ring 3.
680	*
681	* Sync the page before going to the host context; otherwise we'll end up in a loop if
682	* CSAM fails (e.g. instruction crosses a page boundary and the next page is not present)
683	*/
684	LogFlow(("CSAM ring 3 job\n"));
685	int rc2 = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, pvFault, 1, uErr);
686	AssertRC(rc2);
687
688	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeOutOfSync, c);
689	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2CSAM; });
690	return rc;
691	}
692	}
693	# ifdef CSAM_DETECT_NEW_CODE_PAGES
694	else if ( uErr == X86_TRAP_PF_RW
695	&& pRegFrame->ecx >= 0x100 /* early check for movswd count */
696	&& pRegFrame->ecx < 0x10000)
697	{
698	/* In case of a write to a non-present supervisor shadow page, we'll take special precautions
699	* to detect loading of new code pages.
700	*/
701
702	/*
703	* Decode the instruction.
704	*/
705	RTGCPTR PC;
706	rc = SELMValidateAndConvertCSAddr(pVM, pRegFrame->eflags, pRegFrame->ss, pRegFrame->cs, &pRegFrame->csHid, (RTGCPTR)pRegFrame->eip, &PC);
707	if (rc == VINF_SUCCESS)
708	{
709	DISCPUSTATE Cpu;
710	uint32_t cbOp;
711	rc = EMInterpretDisasOneEx(pVM, PC, pRegFrame, &Cpu, &cbOp);
712
713	/* For now we'll restrict this to rep movsw/d instructions */
714	if ( rc == VINF_SUCCESS
715	&& Cpu.pCurInstr->opcode == OP_MOVSWD
716	&& (Cpu.prefix & PREFIX_REP))
717	{
718	CSAMMarkPossibleCodePage(pVM, pvFault);
719	}
720	}
721	}
722	# endif /* CSAM_DETECT_NEW_CODE_PAGES */
723
724	/*
725	* Mark this page as safe.
726	*/
727	/** @todo not correct for pages that contain both code and data!! */
728	Log2(("CSAMMarkPage %RGv; scanned=%d\n", pvFault, true));
729	CSAMMarkPage(pVM, (RTRCPTR)pvFault, true);
730	}
731	}
732	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) && !defined(IN_RING0) */
733	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, pvFault, PGM_SYNC_NR_PAGES, uErr);
734	if (RT_SUCCESS(rc))
735	{
736	/* The page was successfully synced, return to the guest. */
737	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeOutOfSync, c);
738	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2OutOfSync; });
739	return VINF_SUCCESS;
740	}
741	}
742	else /* uErr & X86_TRAP_PF_P: */
743	{
744	/*
745	* Write protected pages is make writable when the guest makes the first
746	* write to it. This happens for pages that are shared, write monitored
747	* and not yet allocated.
748	*
749	* Also, a side effect of not flushing global PDEs are out of sync pages due
750	* to physical monitored regions, that are no longer valid.
751	* Assume for now it only applies to the read/write flag.
752	*/
753	if (RT_SUCCESS(rc) && (uErr & X86_TRAP_PF_RW))
754	{
755	# ifdef VBOX_WITH_NEW_PHYS_CODE
756	if (PGM_PAGE_GET_STATE(pPage) != PGM_PAGE_STATE_ALLOCATED)
757	{
758	Log(("PGM #PF: Make writable: %RGp %R[pgmpage] pvFault=%RGp uErr=%#x\n",
759	GCPhys, pPage, pvFault, uErr));
760	rc = pgmPhysPageMakeWritableUnlocked(pVM, pPage, GCPhys);
761	if (rc != VINF_SUCCESS)
762	{
763	AssertMsg(rc == VINF_PGM_SYNC_CR3 \|\| RT_FAILURE(rc), ("%Rrc\n", rc));
764	return rc;
765	}
766	}
767	/// @todo count the above case; else
768	# endif /* VBOX_WITH_NEW_PHYS_CODE */
769	if (uErr & X86_TRAP_PF_US)
770	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,PageOutOfSyncUser));
771	else /* supervisor */
772	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,PageOutOfSyncSupervisor));
773
774	/*
775	* Note: Do NOT use PGM_SYNC_NR_PAGES here. That only works if the
776	* page is not present, which is not true in this case.
777	*/
778	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, pvFault, 1, uErr);
779	if (RT_SUCCESS(rc))
780	{
781	/*
782	* Page was successfully synced, return to guest.
783	*/
784	# ifdef VBOX_STRICT
785	RTGCPHYS GCPhys;
786	uint64_t fPageGst;
787	rc = PGMGstGetPage(pVM, pvFault, &fPageGst, &GCPhys);
788	Assert(RT_SUCCESS(rc) && fPageGst & X86_PTE_RW);
789	LogFlow(("Obsolete physical monitor page out of sync %RGv - phys %RGp flags=%08llx\n", pvFault, GCPhys, (uint64_t)fPageGst));
790
791	uint64_t fPageShw;
792	rc = PGMShwGetPage(pVM, pvFault, &fPageShw, NULL);
793	AssertMsg(RT_SUCCESS(rc) && fPageShw & X86_PTE_RW, ("rc=%Rrc fPageShw=%RX64\n", rc, fPageShw));
794	# endif /* VBOX_STRICT */
795	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeOutOfSync, c);
796	STAM_STATS({ pVM->pgm.s.CTX_SUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatRZTrap0eTime2OutOfSyncHndObs; });
797	return VINF_SUCCESS;
798	}
799
800	/* Check to see if we need to emulate the instruction as X86_CR0_WP has been cleared. */
801	if ( CPUMGetGuestCPL(pVM, pRegFrame) == 0
802	&& ((CPUMGetGuestCR0(pVM) & (X86_CR0_WP \| X86_CR0_PG)) == X86_CR0_PG)
803	&& (uErr & (X86_TRAP_PF_RW \| X86_TRAP_PF_P)) == (X86_TRAP_PF_RW \| X86_TRAP_PF_P))
804	{
805	uint64_t fPageGst;
806	rc = PGMGstGetPage(pVM, pvFault, &fPageGst, NULL);
807	if ( RT_SUCCESS(rc)
808	&& !(fPageGst & X86_PTE_RW))
809	{
810	rc = PGMInterpretInstruction(pVM, pRegFrame, pvFault);
811	if (RT_SUCCESS(rc))
812	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eWPEmulInRZ);
813	else
814	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eWPEmulToR3);
815	return rc;
816	}
817	AssertMsgFailed(("Unexpected r/w page %RGv flag=%x rc=%Rrc\n", pvFault, (uint32_t)fPageGst, rc));
818	}
819	}
820
821	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
822	# ifdef VBOX_STRICT
823	/*
824	* Check for VMM page flags vs. Guest page flags consistency.
825	* Currently only for debug purposes.
826	*/
827	if (RT_SUCCESS(rc))
828	{
829	/* Get guest page flags. */
830	uint64_t fPageGst;
831	rc = PGMGstGetPage(pVM, pvFault, &fPageGst, NULL);
832	if (RT_SUCCESS(rc))
833	{
834	uint64_t fPageShw;
835	rc = PGMShwGetPage(pVM, pvFault, &fPageShw, NULL);
836
837	/*
838	* Compare page flags.
839	* Note: we have AVL, A, D bits desynched.
840	*/
841	AssertMsg((fPageShw & ~(X86_PTE_A \| X86_PTE_D \| X86_PTE_AVL_MASK)) == (fPageGst & ~(X86_PTE_A \| X86_PTE_D \| X86_PTE_AVL_MASK)),
842	("Page flags mismatch! pvFault=%RGv GCPhys=%RGp fPageShw=%08llx fPageGst=%08llx\n", pvFault, GCPhys, fPageShw, fPageGst));
843	}
844	else
845	AssertMsgFailed(("PGMGstGetPage rc=%Rrc\n", rc));
846	}
847	else
848	AssertMsgFailed(("PGMGCGetPage rc=%Rrc\n", rc));
849	# endif /* VBOX_STRICT */
850	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
851	}
852	STAM_PROFILE_STOP(&pVM->pgm.s.StatRZTrap0eTimeOutOfSync, c);
853	# endif /* PGM_OUT_OF_SYNC_IN_GC */
854	}
855	else /* GCPhys == NIL_RTGCPHYS */
856	{
857	/*
858	* Page not present in Guest OS or invalid page table address.
859	* This is potential virtual page access handler food.
860	*
861	* For the present we'll say that our access handlers don't
862	* work for this case - we've already discarded the page table
863	* not present case which is identical to this.
864	*
865	* When we perchance find we need this, we will probably have AVL
866	* trees (offset based) to operate on and we can measure their speed
867	* agains mapping a page table and probably rearrange this handling
868	* a bit. (Like, searching virtual ranges before checking the
869	* physical address.)
870	*/
871	}
872	}
873	/* else: !present (guest) */
874
875
876	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
877	/*
878	* Conclusion, this is a guest trap.
879	*/
880	LogFlow(("PGM: Unhandled #PF -> route trap to recompiler!\n"));
881	STAM_COUNTER_INC(&pVM->pgm.s.StatRZTrap0eGuestPFUnh);
882	return VINF_EM_RAW_GUEST_TRAP;
883	# else
884	/* present, but not a monitored page; perhaps the guest is probing physical memory */
885	return VINF_EM_RAW_EMULATE_INSTR;
886	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
887
888
889	# else /* PGM_GST_TYPE != PGM_TYPE_32BIT */
890
891	AssertReleaseMsgFailed(("Shw=%d Gst=%d is not implemented!\n", PGM_GST_TYPE, PGM_SHW_TYPE));
892	return VERR_INTERNAL_ERROR;
893	# endif /* PGM_GST_TYPE != PGM_TYPE_32BIT */
894	}
895	#endif /* !IN_RING3 */
896
897
898	/**
899	* Emulation of the invlpg instruction.
900	*
901	*
902	* @returns VBox status code.
903	*
904	* @param pVM VM handle.
905	* @param GCPtrPage Page to invalidate.
906	*
907	* @remark ASSUMES that the guest is updating before invalidating. This order
908	* isn't required by the CPU, so this is speculative and could cause
909	* trouble.
910	*
911	* @todo Flush page or page directory only if necessary!
912	* @todo Add a #define for simply invalidating the page.
913	*/
914	PGM_BTH_DECL(int, InvalidatePage)(PVM pVM, RTGCPTR GCPtrPage)
915	{
916	#if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) \
917	&& PGM_SHW_TYPE != PGM_TYPE_NESTED \
918	&& PGM_SHW_TYPE != PGM_TYPE_EPT
919	int rc;
920
921	LogFlow(("InvalidatePage %RGv\n", GCPtrPage));
922	/*
923	* Get the shadow PD entry and skip out if this PD isn't present.
924	* (Guessing that it is frequent for a shadow PDE to not be present, do this first.)
925	*/
926	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
927	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
928	PX86PDE pPdeDst = pgmShwGet32BitPDEPtr(&pVM->pgm.s, GCPtrPage);
929
930	/* Fetch the pgm pool shadow descriptor. */
931	PPGMPOOLPAGE pShwPde = pVM->pgm.s.CTX_SUFF(pShwPageCR3);
932	Assert(pShwPde);
933
934	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
935	const unsigned iPdpt = (GCPtrPage >> X86_PDPT_SHIFT);
936	PX86PDPT pPdptDst = pgmShwGetPaePDPTPtr(&pVM->pgm.s);
937
938	/* If the shadow PDPE isn't present, then skip the invalidate. */
939	if (!pPdptDst->a[iPdpt].n.u1Present)
940	{
941	Assert(!(pPdptDst->a[iPdpt].u & PGM_PLXFLAGS_MAPPING));
942	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePageSkipped));
943	return VINF_SUCCESS;
944	}
945
946	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
947	PPGMPOOLPAGE pShwPde;
948	PX86PDPAE pPDDst;
949
950	/* Fetch the pgm pool shadow descriptor. */
951	rc = pgmShwGetPaePoolPagePD(&pVM->pgm.s, GCPtrPage, &pShwPde);
952	AssertRCSuccessReturn(rc, rc);
953	Assert(pShwPde);
954
955	pPDDst = (PX86PDPAE)PGMPOOL_PAGE_2_PTR_BY_PGM(&pVM->pgm.s, pShwPde);
956	PX86PDEPAE pPdeDst = &pPDDst->a[iPDDst];
957
958	# else /* PGM_SHW_TYPE == PGM_TYPE_AMD64 */
959	/* PML4 */
960	const unsigned iPml4 = (GCPtrPage >> X86_PML4_SHIFT) & X86_PML4_MASK;
961	const unsigned iPdpt = (GCPtrPage >> X86_PDPT_SHIFT) & X86_PDPT_MASK_AMD64;
962	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
963	PX86PDPAE pPDDst;
964	PX86PDPT pPdptDst;
965	PX86PML4E pPml4eDst;
966	rc = pgmShwGetLongModePDPtr(pVM, GCPtrPage, &pPml4eDst, &pPdptDst, &pPDDst);
967	if (rc != VINF_SUCCESS)
968	{
969	AssertMsg(rc == VERR_PAGE_DIRECTORY_PTR_NOT_PRESENT \|\| rc == VERR_PAGE_MAP_LEVEL4_NOT_PRESENT, ("Unexpected rc=%Rrc\n", rc));
970	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePageSkipped));
971	if (!VM_FF_ISSET(pVM, VM_FF_PGM_SYNC_CR3))
972	PGM_INVL_GUEST_TLBS();
973	return VINF_SUCCESS;
974	}
975	Assert(pPDDst);
976
977	PX86PDEPAE pPdeDst = &pPDDst->a[iPDDst];
978	PX86PDPE pPdpeDst = &pPdptDst->a[iPdpt];
979
980	if (!pPdpeDst->n.u1Present)
981	{
982	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePageSkipped));
983	if (!VM_FF_ISSET(pVM, VM_FF_PGM_SYNC_CR3))
984	PGM_INVL_GUEST_TLBS();
985	return VINF_SUCCESS;
986	}
987
988	# endif /* PGM_SHW_TYPE == PGM_TYPE_AMD64 */
989
990	const SHWPDE PdeDst = *pPdeDst;
991	if (!PdeDst.n.u1Present)
992	{
993	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePageSkipped));
994	return VINF_SUCCESS;
995	}
996
997	/*
998	* Get the guest PD entry and calc big page.
999	*/
1000	# if PGM_GST_TYPE == PGM_TYPE_32BIT
1001	PGSTPD pPDSrc = pgmGstGet32bitPDPtr(&pVM->pgm.s);
1002	const unsigned iPDSrc = GCPtrPage >> GST_PD_SHIFT;
1003	GSTPDE PdeSrc = pPDSrc->a[iPDSrc];
1004	# else /* PGM_GST_TYPE != PGM_TYPE_32BIT */
1005	unsigned iPDSrc;
1006	# if PGM_GST_TYPE == PGM_TYPE_PAE
1007	X86PDPE PdpeSrc;
1008	PX86PDPAE pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, GCPtrPage, &iPDSrc, &PdpeSrc);
1009	# else /* AMD64 */
1010	PX86PML4E pPml4eSrc;
1011	X86PDPE PdpeSrc;
1012	PX86PDPAE pPDSrc = pgmGstGetLongModePDPtr(&pVM->pgm.s, GCPtrPage, &pPml4eSrc, &PdpeSrc, &iPDSrc);
1013	# endif
1014	GSTPDE PdeSrc;
1015
1016	if (pPDSrc)
1017	PdeSrc = pPDSrc->a[iPDSrc];
1018	else
1019	PdeSrc.u = 0;
1020	# endif /* PGM_GST_TYPE != PGM_TYPE_32BIT */
1021
1022	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1023	const bool fIsBigPage = PdeSrc.b.u1Size;
1024	# else
1025	const bool fIsBigPage = PdeSrc.b.u1Size && (CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
1026	# endif
1027
1028	# ifdef IN_RING3
1029	/*
1030	* If a CR3 Sync is pending we may ignore the invalidate page operation
1031	* depending on the kind of sync and if it's a global page or not.
1032	* This doesn't make sense in GC/R0 so we'll skip it entirely there.
1033	*/
1034	# ifdef PGM_SKIP_GLOBAL_PAGEDIRS_ON_NONGLOBAL_FLUSH
1035	if ( VM_FF_ISSET(pVM, VM_FF_PGM_SYNC_CR3)
1036	\|\| ( VM_FF_ISSET(pVM, VM_FF_PGM_SYNC_CR3_NON_GLOBAL)
1037	&& fIsBigPage
1038	&& PdeSrc.b.u1Global
1039	)
1040	)
1041	# else
1042	if (VM_FF_ISPENDING(pVM, VM_FF_PGM_SYNC_CR3 \| VM_FF_PGM_SYNC_CR3_NON_GLOBAL) )
1043	# endif
1044	{
1045	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePageSkipped));
1046	return VINF_SUCCESS;
1047	}
1048	# endif /* IN_RING3 */
1049
1050	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1051	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
1052
1053	/* Fetch the pgm pool shadow descriptor. */
1054	PPGMPOOLPAGE pShwPdpt = pgmPoolGetPageByHCPhys(pVM, pPml4eDst->u & X86_PML4E_PG_MASK);
1055	Assert(pShwPdpt);
1056
1057	/* Fetch the pgm pool shadow descriptor. */
1058	PPGMPOOLPAGE pShwPde = pgmPoolGetPageByHCPhys(pVM, pPdptDst->a[iPdpt].u & SHW_PDPE_PG_MASK);
1059	Assert(pShwPde);
1060
1061	Assert(pPml4eDst->n.u1Present && (pPml4eDst->u & SHW_PDPT_MASK));
1062	RTGCPHYS GCPhysPdpt = pPml4eSrc->u & X86_PML4E_PG_MASK;
1063
1064	if ( !pPml4eSrc->n.u1Present
1065	\|\| pShwPdpt->GCPhys != GCPhysPdpt)
1066	{
1067	LogFlow(("InvalidatePage: Out-of-sync PML4E (P/GCPhys) at %RGv GCPhys=%RGp vs %RGp Pml4eSrc=%RX64 Pml4eDst=%RX64\n",
1068	GCPtrPage, pShwPdpt->GCPhys, GCPhysPdpt, (uint64_t)pPml4eSrc->u, (uint64_t)pPml4eDst->u));
1069	pgmPoolFreeByPage(pPool, pShwPdpt, pVM->pgm.s.CTX_SUFF(pShwPageCR3)->idx, iPml4);
1070	pPml4eDst->u = 0;
1071	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDNPs));
1072	PGM_INVL_GUEST_TLBS();
1073	return VINF_SUCCESS;
1074	}
1075	if ( pPml4eSrc->n.u1User != pPml4eDst->n.u1User
1076	\|\| (!pPml4eSrc->n.u1Write && pPml4eDst->n.u1Write))
1077	{
1078	/*
1079	* Mark not present so we can resync the PML4E when it's used.
1080	*/
1081	LogFlow(("InvalidatePage: Out-of-sync PML4E at %RGv Pml4eSrc=%RX64 Pml4eDst=%RX64\n",
1082	GCPtrPage, (uint64_t)pPml4eSrc->u, (uint64_t)pPml4eDst->u));
1083	pgmPoolFreeByPage(pPool, pShwPdpt, pVM->pgm.s.CTX_SUFF(pShwPageCR3)->idx, iPml4);
1084	pPml4eDst->u = 0;
1085	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDOutOfSync));
1086	PGM_INVL_GUEST_TLBS();
1087	}
1088	else if (!pPml4eSrc->n.u1Accessed)
1089	{
1090	/*
1091	* Mark not present so we can set the accessed bit.
1092	*/
1093	LogFlow(("InvalidatePage: Out-of-sync PML4E (A) at %RGv Pml4eSrc=%RX64 Pml4eDst=%RX64\n",
1094	GCPtrPage, (uint64_t)pPml4eSrc->u, (uint64_t)pPml4eDst->u));
1095	pgmPoolFreeByPage(pPool, pShwPdpt, pVM->pgm.s.CTX_SUFF(pShwPageCR3)->idx, iPml4);
1096	pPml4eDst->u = 0;
1097	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDNAs));
1098	PGM_INVL_GUEST_TLBS();
1099	}
1100
1101	/* Check if the PDPT entry has changed. */
1102	Assert(pPdpeDst->n.u1Present && pPdpeDst->u & SHW_PDPT_MASK);
1103	RTGCPHYS GCPhysPd = PdpeSrc.u & GST_PDPE_PG_MASK;
1104	if ( !PdpeSrc.n.u1Present
1105	\|\| pShwPde->GCPhys != GCPhysPd)
1106	{
1107	LogFlow(("InvalidatePage: Out-of-sync PDPE (P/GCPhys) at %RGv GCPhys=%RGp vs %RGp PdpeSrc=%RX64 PdpeDst=%RX64\n",
1108	GCPtrPage, pShwPde->GCPhys, GCPhysPd, (uint64_t)PdpeSrc.u, (uint64_t)pPdpeDst->u));
1109	pgmPoolFreeByPage(pPool, pShwPde, pShwPdpt->idx, iPdpt);
1110	pPdpeDst->u = 0;
1111	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDNPs));
1112	PGM_INVL_GUEST_TLBS();
1113	return VINF_SUCCESS;
1114	}
1115	if ( PdpeSrc.lm.u1User != pPdpeDst->lm.u1User
1116	\|\| (!PdpeSrc.lm.u1Write && pPdpeDst->lm.u1Write))
1117	{
1118	/*
1119	* Mark not present so we can resync the PDPTE when it's used.
1120	*/
1121	LogFlow(("InvalidatePage: Out-of-sync PDPE at %RGv PdpeSrc=%RX64 PdpeDst=%RX64\n",
1122	GCPtrPage, (uint64_t)PdpeSrc.u, (uint64_t)pPdpeDst->u));
1123	pgmPoolFreeByPage(pPool, pShwPde, pShwPdpt->idx, iPdpt);
1124	pPdpeDst->u = 0;
1125	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDOutOfSync));
1126	PGM_INVL_GUEST_TLBS();
1127	}
1128	else if (!PdpeSrc.lm.u1Accessed)
1129	{
1130	/*
1131	* Mark not present so we can set the accessed bit.
1132	*/
1133	LogFlow(("InvalidatePage: Out-of-sync PDPE (A) at %RGv PdpeSrc=%RX64 PdpeDst=%RX64\n",
1134	GCPtrPage, (uint64_t)PdpeSrc.u, (uint64_t)pPdpeDst->u));
1135	pgmPoolFreeByPage(pPool, pShwPde, pShwPdpt->idx, iPdpt);
1136	pPdpeDst->u = 0;
1137	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDNAs));
1138	PGM_INVL_GUEST_TLBS();
1139	}
1140	# endif /* PGM_GST_TYPE == PGM_TYPE_AMD64 */
1141
1142
1143	/*
1144	* Deal with the Guest PDE.
1145	*/
1146	rc = VINF_SUCCESS;
1147	if (PdeSrc.n.u1Present)
1148	{
1149	if (PdeDst.u & PGM_PDFLAGS_MAPPING)
1150	{
1151	/*
1152	* Conflict - Let SyncPT deal with it to avoid duplicate code.
1153	*/
1154	Assert(pgmMapAreMappingsEnabled(&pVM->pgm.s));
1155	Assert(PGMGetGuestMode(pVM) <= PGMMODE_PAE);
1156	rc = PGM_BTH_NAME(SyncPT)(pVM, iPDSrc, pPDSrc, GCPtrPage);
1157	}
1158	else if ( PdeSrc.n.u1User != PdeDst.n.u1User
1159	\|\| (!PdeSrc.n.u1Write && PdeDst.n.u1Write))
1160	{
1161	/*
1162	* Mark not present so we can resync the PDE when it's used.
1163	*/
1164	LogFlow(("InvalidatePage: Out-of-sync at %RGp PdeSrc=%RX64 PdeDst=%RX64\n",
1165	GCPtrPage, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
1166	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, pShwPde->idx, iPDDst);
1167	pPdeDst->u = 0;
1168	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDOutOfSync));
1169	PGM_INVL_GUEST_TLBS();
1170	}
1171	else if (!PdeSrc.n.u1Accessed)
1172	{
1173	/*
1174	* Mark not present so we can set the accessed bit.
1175	*/
1176	LogFlow(("InvalidatePage: Out-of-sync (A) at %RGp PdeSrc=%RX64 PdeDst=%RX64\n",
1177	GCPtrPage, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
1178	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, pShwPde->idx, iPDDst);
1179	pPdeDst->u = 0;
1180	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDNAs));
1181	PGM_INVL_GUEST_TLBS();
1182	}
1183	else if (!fIsBigPage)
1184	{
1185	/*
1186	* 4KB - page.
1187	*/
1188	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, PdeDst.u & SHW_PDE_PG_MASK);
1189	RTGCPHYS GCPhys = PdeSrc.u & GST_PDE_PG_MASK;
1190	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
1191	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
1192	GCPhys \|= (iPDDst & 1) * (PAGE_SIZE/2);
1193	# endif
1194	if (pShwPage->GCPhys == GCPhys)
1195	{
1196	# if 0 /* likely cause of a major performance regression; must be SyncPageWorkerTrackDeref then */
1197	const unsigned iPTEDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1198	PSHWPT pPT = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
1199	if (pPT->a[iPTEDst].n.u1Present)
1200	{
1201	# ifdef PGMPOOL_WITH_USER_TRACKING
1202	/* This is very unlikely with caching/monitoring enabled. */
1203	PGM_BTH_NAME(SyncPageWorkerTrackDeref)(pVM, pShwPage, pPT->a[iPTEDst].u & SHW_PTE_PG_MASK);
1204	# endif
1205	pPT->a[iPTEDst].u = 0;
1206	}
1207	# else /* Syncing it here isn't 100% safe and it's probably not worth spending time syncing it. */
1208	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, GCPtrPage, 1, 0);
1209	if (RT_SUCCESS(rc))
1210	rc = VINF_SUCCESS;
1211	# endif
1212	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePage4KBPages));
1213	PGM_INVL_PG(GCPtrPage);
1214	}
1215	else
1216	{
1217	/*
1218	* The page table address changed.
1219	*/
1220	LogFlow(("InvalidatePage: Out-of-sync at %RGp PdeSrc=%RX64 PdeDst=%RX64 ShwGCPhys=%RGp iPDDst=%#x\n",
1221	GCPtrPage, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u, pShwPage->GCPhys, iPDDst));
1222	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, pShwPde->idx, iPDDst);
1223	pPdeDst->u = 0;
1224	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDOutOfSync));
1225	PGM_INVL_GUEST_TLBS();
1226	}
1227	}
1228	else
1229	{
1230	/*
1231	* 2/4MB - page.
1232	*/
1233	/* Before freeing the page, check if anything really changed. */
1234	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, PdeDst.u & SHW_PDE_PG_MASK);
1235	RTGCPHYS GCPhys = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc);
1236	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
1237	/* Select the right PDE as we're emulating a 4MB page directory with two 2 MB shadow PDEs.*/
1238	GCPhys \|= GCPtrPage & (1 << X86_PD_PAE_SHIFT);
1239	# endif
1240	if ( pShwPage->GCPhys == GCPhys
1241	&& pShwPage->enmKind == BTH_PGMPOOLKIND_PT_FOR_BIG)
1242	{
1243	/* ASSUMES a the given bits are identical for 4M and normal PDEs */
1244	/** @todo PAT */
1245	if ( (PdeSrc.u & (X86_PDE_P \| X86_PDE_RW \| X86_PDE_US \| X86_PDE_PWT \| X86_PDE_PCD))
1246	== (PdeDst.u & (X86_PDE_P \| X86_PDE_RW \| X86_PDE_US \| X86_PDE_PWT \| X86_PDE_PCD))
1247	&& ( PdeSrc.b.u1Dirty /** @todo rainy day: What about read-only 4M pages? not very common, but still... */
1248	\|\| (PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY)))
1249	{
1250	LogFlow(("Skipping flush for big page containing %RGv (PD=%X .u=%RX64)-> nothing has changed!\n", GCPtrPage, iPDSrc, PdeSrc.u));
1251	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePage4MBPagesSkip));
1252	return VINF_SUCCESS;
1253	}
1254	}
1255
1256	/*
1257	* Ok, the page table is present and it's been changed in the guest.
1258	* If we're in host context, we'll just mark it as not present taking the lazy approach.
1259	* We could do this for some flushes in GC too, but we need an algorithm for
1260	* deciding which 4MB pages containing code likely to be executed very soon.
1261	*/
1262	LogFlow(("InvalidatePage: Out-of-sync PD at %RGp PdeSrc=%RX64 PdeDst=%RX64\n",
1263	GCPtrPage, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
1264	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, pShwPde->idx, iPDDst);
1265	pPdeDst->u = 0;
1266	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePage4MBPages));
1267	PGM_INVL_BIG_PG(GCPtrPage);
1268	}
1269	}
1270	else
1271	{
1272	/*
1273	* Page directory is not present, mark shadow PDE not present.
1274	*/
1275	if (!(PdeDst.u & PGM_PDFLAGS_MAPPING))
1276	{
1277	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, pShwPde->idx, iPDDst);
1278	pPdeDst->u = 0;
1279	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDNPs));
1280	PGM_INVL_PG(GCPtrPage);
1281	}
1282	else
1283	{
1284	Assert(pgmMapAreMappingsEnabled(&pVM->pgm.s));
1285	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDMappings));
1286	}
1287	}
1288
1289	return rc;
1290
1291	#else /* guest real and protected mode */
1292	/* There's no such thing as InvalidatePage when paging is disabled, so just ignore. */
1293	return VINF_SUCCESS;
1294	#endif
1295	}
1296
1297
1298	#ifdef PGMPOOL_WITH_USER_TRACKING
1299	/**
1300	* Update the tracking of shadowed pages.
1301	*
1302	* @param pVM The VM handle.
1303	* @param pShwPage The shadow page.
1304	* @param HCPhys The physical page we is being dereferenced.
1305	*/
1306	DECLINLINE(void) PGM_BTH_NAME(SyncPageWorkerTrackDeref)(PVM pVM, PPGMPOOLPAGE pShwPage, RTHCPHYS HCPhys)
1307	{
1308	# ifdef PGMPOOL_WITH_GCPHYS_TRACKING
1309	STAM_PROFILE_START(&pVM->pgm.s.StatTrackDeref, a);
1310	LogFlow(("SyncPageWorkerTrackDeref: Damn HCPhys=%RHp pShwPage->idx=%#x!!!\n", HCPhys, pShwPage->idx));
1311
1312	/** @todo If this turns out to be a bottle neck (very likely) two things can be done:
1313	* 1. have a medium sized HCPhys -> GCPhys TLB (hash?)
1314	* 2. write protect all shadowed pages. I.e. implement caching.
1315	*/
1316	/*
1317	* Find the guest address.
1318	*/
1319	for (PPGMRAMRANGE pRam = pVM->pgm.s.CTX_SUFF(pRamRanges);
1320	pRam;
1321	pRam = pRam->CTX_SUFF(pNext))
1322	{
1323	unsigned iPage = pRam->cb >> PAGE_SHIFT;
1324	while (iPage-- > 0)
1325	{
1326	if (PGM_PAGE_GET_HCPHYS(&pRam->aPages[iPage]) == HCPhys)
1327	{
1328	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
1329	pgmTrackDerefGCPhys(pPool, pShwPage, &pRam->aPages[iPage]);
1330	pShwPage->cPresent--;
1331	pPool->cPresent--;
1332	STAM_PROFILE_STOP(&pVM->pgm.s.StatTrackDeref, a);
1333	return;
1334	}
1335	}
1336	}
1337
1338	for (;;)
1339	AssertReleaseMsgFailed(("HCPhys=%RHp wasn't found!\n", HCPhys));
1340	# else /* !PGMPOOL_WITH_GCPHYS_TRACKING */
1341	pShwPage->cPresent--;
1342	pVM->pgm.s.CTX_SUFF(pPool)->cPresent--;
1343	# endif /* !PGMPOOL_WITH_GCPHYS_TRACKING */
1344	}
1345
1346
1347	/**
1348	* Update the tracking of shadowed pages.
1349	*
1350	* @param pVM The VM handle.
1351	* @param pShwPage The shadow page.
1352	* @param u16 The top 16-bit of the pPage->HCPhys.
1353	* @param pPage Pointer to the guest page. this will be modified.
1354	* @param iPTDst The index into the shadow table.
1355	*/
1356	DECLINLINE(void) PGM_BTH_NAME(SyncPageWorkerTrackAddref)(PVM pVM, PPGMPOOLPAGE pShwPage, uint16_t u16, PPGMPAGE pPage, const unsigned iPTDst)
1357	{
1358	# ifdef PGMPOOL_WITH_GCPHYS_TRACKING
1359	/*
1360	* Just deal with the simple first time here.
1361	*/
1362	if (!u16)
1363	{
1364	STAM_COUNTER_INC(&pVM->pgm.s.StatTrackVirgin);
1365	u16 = PGMPOOL_TD_MAKE(1, pShwPage->idx);
1366	}
1367	else
1368	u16 = pgmPoolTrackPhysExtAddref(pVM, u16, pShwPage->idx);
1369
1370	/* write back */
1371	Log2(("SyncPageWorkerTrackAddRef: u16=%#x->%#x iPTDst=%#x\n", u16, PGM_PAGE_GET_TRACKING(pPage), iPTDst));
1372	PGM_PAGE_SET_TRACKING(pPage, u16);
1373
1374	# endif /* PGMPOOL_WITH_GCPHYS_TRACKING */
1375
1376	/* update statistics. */
1377	pVM->pgm.s.CTX_SUFF(pPool)->cPresent++;
1378	pShwPage->cPresent++;
1379	if (pShwPage->iFirstPresent > iPTDst)
1380	pShwPage->iFirstPresent = iPTDst;
1381	}
1382	#endif /* PGMPOOL_WITH_USER_TRACKING */
1383
1384
1385	/**
1386	* Creates a 4K shadow page for a guest page.
1387	*
1388	* For 4M pages the caller must convert the PDE4M to a PTE, this includes adjusting the
1389	* physical address. The PdeSrc argument only the flags are used. No page structured
1390	* will be mapped in this function.
1391	*
1392	* @param pVM VM handle.
1393	* @param pPteDst Destination page table entry.
1394	* @param PdeSrc Source page directory entry (i.e. Guest OS page directory entry).
1395	* Can safely assume that only the flags are being used.
1396	* @param PteSrc Source page table entry (i.e. Guest OS page table entry).
1397	* @param pShwPage Pointer to the shadow page.
1398	* @param iPTDst The index into the shadow table.
1399	*
1400	* @remark Not used for 2/4MB pages!
1401	*/
1402	DECLINLINE(void) PGM_BTH_NAME(SyncPageWorker)(PVM pVM, PSHWPTE pPteDst, GSTPDE PdeSrc, GSTPTE PteSrc, PPGMPOOLPAGE pShwPage, unsigned iPTDst)
1403	{
1404	if (PteSrc.n.u1Present)
1405	{
1406	/*
1407	* Find the ram range.
1408	*/
1409	PPGMPAGE pPage;
1410	int rc = pgmPhysGetPageEx(&pVM->pgm.s, PteSrc.u & GST_PTE_PG_MASK, &pPage);
1411	if (RT_SUCCESS(rc))
1412	{
1413	#ifdef VBOX_WITH_NEW_PHYS_CODE
1414	# ifndef VBOX_WITH_NEW_LAZY_PAGE_ALLOC
1415	/* Try make the page writable if necessary. */
1416	if ( PteSrc.n.u1Write
1417	&& PGM_PAGE_GET_STATE(pPage) != PGM_PAGE_STATE_ALLOCATED
1418	&& PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM)
1419	{
1420	rc = pgmPhysPageMakeWritableUnlocked(pVM, pPage, PteSrc.u & GST_PTE_PG_MASK);
1421	AssertRC(rc);
1422	}
1423	# endif
1424	#endif
1425
1426	/** @todo investiage PWT, PCD and PAT. */
1427	/*
1428	* Make page table entry.
1429	*/
1430	SHWPTE PteDst;
1431	if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
1432	{
1433	/** @todo r=bird: Are we actually handling dirty and access bits for pages with access handlers correctly? No. */
1434	if (!PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage))
1435	{
1436	#if PGM_SHW_TYPE == PGM_TYPE_EPT
1437	PteDst.u = PGM_PAGE_GET_HCPHYS(pPage);
1438	PteDst.n.u1Present = 1;
1439	PteDst.n.u1Execute = 1;
1440	PteDst.n.u1IgnorePAT = 1;
1441	PteDst.n.u3EMT = VMX_EPT_MEMTYPE_WB;
1442	/* PteDst.n.u1Write = 0 && PteDst.n.u1Size = 0 */
1443	#else
1444	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))
1445	\| PGM_PAGE_GET_HCPHYS(pPage);
1446	#endif
1447	}
1448	else
1449	{
1450	LogFlow(("SyncPageWorker: monitored page (%RHp) -> mark not present\n", PGM_PAGE_GET_HCPHYS(pPage)));
1451	PteDst.u = 0;
1452	}
1453	/** @todo count these two kinds. */
1454	}
1455	else
1456	{
1457	#if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
1458	/*
1459	* If the page or page directory entry is not marked accessed,
1460	* we mark the page not present.
1461	*/
1462	if (!PteSrc.n.u1Accessed \|\| !PdeSrc.n.u1Accessed)
1463	{
1464	LogFlow(("SyncPageWorker: page and or page directory not accessed -> mark not present\n"));
1465	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,AccessedPage));
1466	PteDst.u = 0;
1467	}
1468	else
1469	/*
1470	* If the page is not flagged as dirty and is writable, then make it read-only, so we can set the dirty bit
1471	* when the page is modified.
1472	*/
1473	if (!PteSrc.n.u1Dirty && (PdeSrc.n.u1Write & PteSrc.n.u1Write))
1474	{
1475	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPage));
1476	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))
1477	\| PGM_PAGE_GET_HCPHYS(pPage)
1478	\| PGM_PTFLAGS_TRACK_DIRTY;
1479	}
1480	else
1481	#endif
1482	{
1483	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPageSkipped));
1484	#if PGM_SHW_TYPE == PGM_TYPE_EPT
1485	PteDst.u = PGM_PAGE_GET_HCPHYS(pPage);
1486	PteDst.n.u1Present = 1;
1487	PteDst.n.u1Write = 1;
1488	PteDst.n.u1Execute = 1;
1489	PteDst.n.u1IgnorePAT = 1;
1490	PteDst.n.u3EMT = VMX_EPT_MEMTYPE_WB;
1491	/* PteDst.n.u1Size = 0 */
1492	#else
1493	PteDst.u = (PteSrc.u & ~(X86_PTE_PAE_PG_MASK \| X86_PTE_AVL_MASK \| X86_PTE_PAT \| X86_PTE_PCD \| X86_PTE_PWT))
1494	\| PGM_PAGE_GET_HCPHYS(pPage);
1495	#endif
1496	}
1497	}
1498
1499	#ifdef VBOX_WITH_NEW_PHYS_CODE
1500	/*
1501	* Make sure only allocated pages are mapped writable.
1502	*/
1503	if ( PteDst.n.u1Write
1504	&& PteDst.n.u1Present
1505	&& PGM_PAGE_GET_STATE(pPage) != PGM_PAGE_STATE_ALLOCATED)
1506	{
1507	PteDst.n.u1Write = 0; /** @todo this isn't quite working yet. */
1508	Log3(("SyncPageWorker: write-protecting %RGp pPage=%R[pgmpage]at iPTDst=%d\n", (RTGCPHYS)(PteSrc.u & X86_PTE_PAE_PG_MASK), pPage, iPTDst));
1509	}
1510	#endif
1511
1512	#ifdef PGMPOOL_WITH_USER_TRACKING
1513	/*
1514	* Keep user track up to date.
1515	*/
1516	if (PteDst.n.u1Present)
1517	{
1518	if (!pPteDst->n.u1Present)
1519	PGM_BTH_NAME(SyncPageWorkerTrackAddref)(pVM, pShwPage, PGM_PAGE_GET_TRACKING(pPage), pPage, iPTDst);
1520	else if ((pPteDst->u & SHW_PTE_PG_MASK) != (PteDst.u & SHW_PTE_PG_MASK))
1521	{
1522	Log2(("SyncPageWorker: deref! *pPteDst=%RX64 PteDst=%RX64\n", (uint64_t)pPteDst->u, (uint64_t)PteDst.u));
1523	PGM_BTH_NAME(SyncPageWorkerTrackDeref)(pVM, pShwPage, pPteDst->u & SHW_PTE_PG_MASK);
1524	PGM_BTH_NAME(SyncPageWorkerTrackAddref)(pVM, pShwPage, PGM_PAGE_GET_TRACKING(pPage), pPage, iPTDst);
1525	}
1526	}
1527	else if (pPteDst->n.u1Present)
1528	{
1529	Log2(("SyncPageWorker: deref! *pPteDst=%RX64\n", (uint64_t)pPteDst->u));
1530	PGM_BTH_NAME(SyncPageWorkerTrackDeref)(pVM, pShwPage, pPteDst->u & SHW_PTE_PG_MASK);
1531	}
1532	#endif /* PGMPOOL_WITH_USER_TRACKING */
1533
1534	/*
1535	* Update statistics and commit the entry.
1536	*/
1537	#if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
1538	if (!PteSrc.n.u1Global)
1539	pShwPage->fSeenNonGlobal = true;
1540	#endif
1541	*pPteDst = PteDst;
1542	}
1543	/* else MMIO or invalid page, we must handle them manually in the #PF handler. */
1544	/** @todo count these. */
1545	}
1546	else
1547	{
1548	/*
1549	* Page not-present.
1550	*/
1551	LogFlow(("SyncPageWorker: page not present in Pte\n"));
1552	#ifdef PGMPOOL_WITH_USER_TRACKING
1553	/* Keep user track up to date. */
1554	if (pPteDst->n.u1Present)
1555	{
1556	Log2(("SyncPageWorker: deref! *pPteDst=%RX64\n", (uint64_t)pPteDst->u));
1557	PGM_BTH_NAME(SyncPageWorkerTrackDeref)(pVM, pShwPage, pPteDst->u & SHW_PTE_PG_MASK);
1558	}
1559	#endif /* PGMPOOL_WITH_USER_TRACKING */
1560	pPteDst->u = 0;
1561	/** @todo count these. */
1562	}
1563	}
1564
1565
1566	/**
1567	* Syncs a guest OS page.
1568	*
1569	* There are no conflicts at this point, neither is there any need for
1570	* page table allocations.
1571	*
1572	* @returns VBox status code.
1573	* @returns VINF_PGM_SYNCPAGE_MODIFIED_PDE if it modifies the PDE in any way.
1574	* @param pVM VM handle.
1575	* @param PdeSrc Page directory entry of the guest.
1576	* @param GCPtrPage Guest context page address.
1577	* @param cPages Number of pages to sync (PGM_SYNC_N_PAGES) (default=1).
1578	* @param uErr Fault error (X86_TRAP_PF_*).
1579	*/
1580	PGM_BTH_DECL(int, SyncPage)(PVM pVM, GSTPDE PdeSrc, RTGCPTR GCPtrPage, unsigned cPages, unsigned uErr)
1581	{
1582	LogFlow(("SyncPage: GCPtrPage=%RGv cPages=%u uErr=%#x\n", GCPtrPage, cPages, uErr));
1583
1584	#if ( PGM_GST_TYPE == PGM_TYPE_32BIT \
1585	\|\| PGM_GST_TYPE == PGM_TYPE_PAE \
1586	\|\| PGM_GST_TYPE == PGM_TYPE_AMD64) \
1587	&& PGM_SHW_TYPE != PGM_TYPE_NESTED \
1588	&& PGM_SHW_TYPE != PGM_TYPE_EPT
1589
1590	# if PGM_WITH_NX(PGM_GST_TYPE, PGM_SHW_TYPE)
1591	bool fNoExecuteBitValid = !!(CPUMGetGuestEFER(pVM) & MSR_K6_EFER_NXE);
1592	# endif
1593
1594	/*
1595	* Assert preconditions.
1596	*/
1597	Assert(PdeSrc.n.u1Present);
1598	Assert(cPages);
1599	STAM_COUNTER_INC(&pVM->pgm.s.StatSyncPagePD[(GCPtrPage >> GST_PD_SHIFT) & GST_PD_MASK]);
1600
1601	/*
1602	* Get the shadow PDE, find the shadow page table in the pool.
1603	*/
1604	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
1605	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
1606	PX86PDE pPdeDst = pgmShwGet32BitPDEPtr(&pVM->pgm.s, GCPtrPage);
1607
1608	/* Fetch the pgm pool shadow descriptor. */
1609	PPGMPOOLPAGE pShwPde = pVM->pgm.s.CTX_SUFF(pShwPageCR3);
1610	Assert(pShwPde);
1611
1612	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
1613	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
1614	PPGMPOOLPAGE pShwPde;
1615	PX86PDPAE pPDDst;
1616
1617	/* Fetch the pgm pool shadow descriptor. */
1618	int rc = pgmShwGetPaePoolPagePD(&pVM->pgm.s, GCPtrPage, &pShwPde);
1619	AssertRCSuccessReturn(rc, rc);
1620	Assert(pShwPde);
1621
1622	pPDDst = (PX86PDPAE)PGMPOOL_PAGE_2_PTR_BY_PGM(&pVM->pgm.s, pShwPde);
1623	PX86PDEPAE pPdeDst = &pPDDst->a[iPDDst];
1624
1625	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
1626	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
1627	const unsigned iPdpt = (GCPtrPage >> X86_PDPT_SHIFT) & X86_PDPT_MASK_AMD64;
1628	PX86PDPAE pPDDst;
1629	PX86PDPT pPdptDst;
1630
1631	int rc = pgmShwGetLongModePDPtr(pVM, GCPtrPage, NULL, &pPdptDst, &pPDDst);
1632	AssertRCSuccessReturn(rc, rc);
1633	Assert(pPDDst && pPdptDst);
1634	PX86PDEPAE pPdeDst = &pPDDst->a[iPDDst];
1635	# endif
1636
1637	SHWPDE PdeDst = *pPdeDst;
1638	AssertMsg(PdeDst.n.u1Present, ("%p=%llx\n", pPdeDst, (uint64_t)PdeDst.u));
1639	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, PdeDst.u & SHW_PDE_PG_MASK);
1640
1641	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1642	/* Fetch the pgm pool shadow descriptor. */
1643	PPGMPOOLPAGE pShwPde = pgmPoolGetPageByHCPhys(pVM, pPdptDst->a[iPdpt].u & X86_PDPE_PG_MASK);
1644	Assert(pShwPde);
1645	# endif
1646
1647	/*
1648	* Check that the page is present and that the shadow PDE isn't out of sync.
1649	*/
1650	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1651	const bool fBigPage = PdeSrc.b.u1Size;
1652	# else
1653	const bool fBigPage = PdeSrc.b.u1Size && (CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
1654	# endif
1655	RTGCPHYS GCPhys;
1656	if (!fBigPage)
1657	{
1658	GCPhys = PdeSrc.u & GST_PDE_PG_MASK;
1659	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
1660	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
1661	GCPhys \|= (iPDDst & 1) * (PAGE_SIZE/2);
1662	# endif
1663	}
1664	else
1665	{
1666	GCPhys = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc);
1667	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
1668	/* Select the right PDE as we're emulating a 4MB page directory with two 2 MB shadow PDEs.*/
1669	GCPhys \|= GCPtrPage & (1 << X86_PD_PAE_SHIFT);
1670	# endif
1671	}
1672	if ( pShwPage->GCPhys == GCPhys
1673	&& PdeSrc.n.u1Present
1674	&& (PdeSrc.n.u1User == PdeDst.n.u1User)
1675	&& (PdeSrc.n.u1Write == PdeDst.n.u1Write \|\| !PdeDst.n.u1Write)
1676	# if PGM_WITH_NX(PGM_GST_TYPE, PGM_SHW_TYPE)
1677	&& (!fNoExecuteBitValid \|\| PdeSrc.n.u1NoExecute == PdeDst.n.u1NoExecute)
1678	# endif
1679	)
1680	{
1681	/*
1682	* Check that the PDE is marked accessed already.
1683	* Since we set the accessed bit before getting here on a #PF, this
1684	* check is only meant for dealing with non-#PF'ing paths.
1685	*/
1686	if (PdeSrc.n.u1Accessed)
1687	{
1688	PSHWPT pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
1689	if (!fBigPage)
1690	{
1691	/*
1692	* 4KB Page - Map the guest page table.
1693	*/
1694	PGSTPT pPTSrc;
1695	int rc = PGM_GCPHYS_2_PTR(pVM, PdeSrc.u & GST_PDE_PG_MASK, &pPTSrc);
1696	if (RT_SUCCESS(rc))
1697	{
1698	# ifdef PGM_SYNC_N_PAGES
1699	Assert(cPages == 1 \|\| !(uErr & X86_TRAP_PF_P));
1700	if (cPages > 1 && !(uErr & X86_TRAP_PF_P))
1701	{
1702	/*
1703	* This code path is currently only taken when the caller is PGMTrap0eHandler
1704	* for non-present pages!
1705	*
1706	* We're setting PGM_SYNC_NR_PAGES pages around the faulting page to sync it and
1707	* deal with locality.
1708	*/
1709	unsigned iPTDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1710	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
1711	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
1712	const unsigned offPTSrc = ((GCPtrPage >> SHW_PD_SHIFT) & 1) * 512;
1713	# else
1714	const unsigned offPTSrc = 0;
1715	# endif
1716	const unsigned iPTDstEnd = RT_MIN(iPTDst + PGM_SYNC_NR_PAGES / 2, RT_ELEMENTS(pPTDst->a));
1717	if (iPTDst < PGM_SYNC_NR_PAGES / 2)
1718	iPTDst = 0;
1719	else
1720	iPTDst -= PGM_SYNC_NR_PAGES / 2;
1721	for (; iPTDst < iPTDstEnd; iPTDst++)
1722	{
1723	if (!pPTDst->a[iPTDst].n.u1Present)
1724	{
1725	GSTPTE PteSrc = pPTSrc->a[offPTSrc + iPTDst];
1726	RTGCPTR GCPtrCurPage = (GCPtrPage & ~(RTGCPTR)(GST_PT_MASK << GST_PT_SHIFT)) \| ((offPTSrc + iPTDst) << PAGE_SHIFT);
1727	NOREF(GCPtrCurPage);
1728	#ifndef IN_RING0
1729	/*
1730	* Assuming kernel code will be marked as supervisor - and not as user level
1731	* and executed using a conforming code selector - And marked as readonly.
1732	* Also assume that if we're monitoring a page, it's of no interest to CSAM.
1733	*/
1734	PPGMPAGE pPage;
1735	if ( ((PdeSrc.u & PteSrc.u) & (X86_PTE_RW \| X86_PTE_US))
1736	\|\| iPTDst == ((GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK) /* always sync GCPtrPage */
1737	\|\| !CSAMDoesPageNeedScanning(pVM, (RTRCPTR)GCPtrCurPage)
1738	\|\| ( (pPage = pgmPhysGetPage(&pVM->pgm.s, PteSrc.u & GST_PTE_PG_MASK))
1739	&& PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
1740	)
1741	#endif /* else: CSAM not active */
1742	PGM_BTH_NAME(SyncPageWorker)(pVM, &pPTDst->a[iPTDst], PdeSrc, PteSrc, pShwPage, iPTDst);
1743	Log2(("SyncPage: 4K+ %RGv PteSrc:{P=%d RW=%d U=%d raw=%08llx} PteDst=%08llx%s\n",
1744	GCPtrCurPage, PteSrc.n.u1Present,
1745	PteSrc.n.u1Write & PdeSrc.n.u1Write,
1746	PteSrc.n.u1User & PdeSrc.n.u1User,
1747	(uint64_t)PteSrc.u,
1748	(uint64_t)pPTDst->a[iPTDst].u,
1749	pPTDst->a[iPTDst].u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
1750	}
1751	}
1752	}
1753	else
1754	# endif /* PGM_SYNC_N_PAGES */
1755	{
1756	const unsigned iPTSrc = (GCPtrPage >> GST_PT_SHIFT) & GST_PT_MASK;
1757	GSTPTE PteSrc = pPTSrc->a[iPTSrc];
1758	const unsigned iPTDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1759	PGM_BTH_NAME(SyncPageWorker)(pVM, &pPTDst->a[iPTDst], PdeSrc, PteSrc, pShwPage, iPTDst);
1760	Log2(("SyncPage: 4K %RGv PteSrc:{P=%d RW=%d U=%d raw=%08llx}%s\n",
1761	GCPtrPage, PteSrc.n.u1Present,
1762	PteSrc.n.u1Write & PdeSrc.n.u1Write,
1763	PteSrc.n.u1User & PdeSrc.n.u1User,
1764	(uint64_t)PteSrc.u,
1765	pPTDst->a[iPTDst].u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
1766	}
1767	}
1768	else /* MMIO or invalid page: emulated in #PF handler. */
1769	{
1770	LogFlow(("PGM_GCPHYS_2_PTR %RGp failed with %Rrc\n", GCPhys, rc));
1771	Assert(!pPTDst->a[(GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK].n.u1Present);
1772	}
1773	}
1774	else
1775	{
1776	/*
1777	* 4/2MB page - lazy syncing shadow 4K pages.
1778	* (There are many causes of getting here, it's no longer only CSAM.)
1779	*/
1780	/* Calculate the GC physical address of this 4KB shadow page. */
1781	RTGCPHYS GCPhys = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc) \| (GCPtrPage & GST_BIG_PAGE_OFFSET_MASK);
1782	/* Find ram range. */
1783	PPGMPAGE pPage;
1784	int rc = pgmPhysGetPageEx(&pVM->pgm.s, GCPhys, &pPage);
1785	if (RT_SUCCESS(rc))
1786	{
1787	# ifdef VBOX_WITH_NEW_PHYS_CODE
1788	# ifndef VBOX_WITH_NEW_LAZY_PAGE_ALLOC
1789	/* Try make the page writable if necessary. */
1790	if ( PdeSrc.n.u1Write
1791	&& PGM_PAGE_GET_STATE(pPage) != PGM_PAGE_STATE_ALLOCATED
1792	&& PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM)
1793	{
1794	rc = pgmPhysPageMakeWritableUnlocked(pVM, pPage, GCPhys);
1795	AssertRC(rc);
1796	}
1797	# endif
1798	# endif
1799
1800	/*
1801	* Make shadow PTE entry.
1802	*/
1803	SHWPTE PteDst;
1804	PteDst.u = (PdeSrc.u & ~(X86_PTE_PAE_PG_MASK \| X86_PTE_AVL_MASK \| X86_PTE_PAT \| X86_PTE_PCD \| X86_PTE_PWT))
1805	\| PGM_PAGE_GET_HCPHYS(pPage);
1806	if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
1807	{
1808	if (!PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage))
1809	PteDst.n.u1Write = 0;
1810	else
1811	PteDst.u = 0;
1812	}
1813	const unsigned iPTDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1814	# ifdef PGMPOOL_WITH_USER_TRACKING
1815	if (PteDst.n.u1Present && !pPTDst->a[iPTDst].n.u1Present)
1816	PGM_BTH_NAME(SyncPageWorkerTrackAddref)(pVM, pShwPage, PGM_PAGE_GET_TRACKING(pPage), pPage, iPTDst);
1817	# endif
1818	# ifdef VBOX_WITH_NEW_PHYS_CODE
1819	/* Make sure only allocated pages are mapped writable. */
1820	if ( PteDst.n.u1Write
1821	&& PteDst.n.u1Present
1822	&& PGM_PAGE_GET_STATE(pPage) != PGM_PAGE_STATE_ALLOCATED)
1823	{
1824	PteDst.n.u1Write = 0; /** @todo this isn't quite working yet... */
1825	Log3(("SyncPage: write-protecting %RGp pPage=%R[pgmpage] at %RGv\n", GCPhys, pPage, GCPtrPage));
1826	}
1827	# endif
1828
1829	pPTDst->a[iPTDst] = PteDst;
1830
1831
1832	/*
1833	* If the page is not flagged as dirty and is writable, then make it read-only
1834	* at PD level, so we can set the dirty bit when the page is modified.
1835	*
1836	* ASSUMES that page access handlers are implemented on page table entry level.
1837	* Thus we will first catch the dirty access and set PDE.D and restart. If
1838	* there is an access handler, we'll trap again and let it work on the problem.
1839	*/
1840	/** @todo r=bird: figure out why we need this here, SyncPT should've taken care of this already.
1841	* As for invlpg, it simply frees the whole shadow PT.
1842	* ...It's possibly because the guest clears it and the guest doesn't really tell us... */
1843	if (!PdeSrc.b.u1Dirty && PdeSrc.b.u1Write)
1844	{
1845	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPageBig));
1846	PdeDst.u \|= PGM_PDFLAGS_TRACK_DIRTY;
1847	PdeDst.n.u1Write = 0;
1848	}
1849	else
1850	{
1851	PdeDst.au32[0] &= ~PGM_PDFLAGS_TRACK_DIRTY;
1852	PdeDst.n.u1Write = PdeSrc.n.u1Write;
1853	}
1854	*pPdeDst = PdeDst;
1855	Log2(("SyncPage: BIG %RGv PdeSrc:{P=%d RW=%d U=%d raw=%08llx} GCPhys=%RGp%s\n",
1856	GCPtrPage, PdeSrc.n.u1Present, PdeSrc.n.u1Write, PdeSrc.n.u1User, (uint64_t)PdeSrc.u, GCPhys,
1857	PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
1858	}
1859	else
1860	LogFlow(("PGM_GCPHYS_2_PTR %RGp (big) failed with %Rrc\n", GCPhys, rc));
1861	}
1862	return VINF_SUCCESS;
1863	}
1864	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPagePDNAs));
1865	}
1866	else
1867	{
1868	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPagePDOutOfSync));
1869	Log2(("SyncPage: Out-Of-Sync PDE at %RGp PdeSrc=%RX64 PdeDst=%RX64 (GCPhys %RGp vs %RGp)\n",
1870	GCPtrPage, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u, pShwPage->GCPhys, GCPhys));
1871	}
1872
1873	/*
1874	* Mark the PDE not present. Restart the instruction and let #PF call SyncPT.
1875	* Yea, I'm lazy.
1876	*/
1877	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
1878	pgmPoolFreeByPage(pPool, pShwPage, pShwPde->idx, iPDDst);
1879
1880	pPdeDst->u = 0;
1881	PGM_INVL_GUEST_TLBS();
1882	return VINF_PGM_SYNCPAGE_MODIFIED_PDE;
1883
1884	#elif (PGM_GST_TYPE == PGM_TYPE_REAL \|\| PGM_GST_TYPE == PGM_TYPE_PROT) \
1885	&& PGM_SHW_TYPE != PGM_TYPE_NESTED \
1886	&& (PGM_SHW_TYPE != PGM_TYPE_EPT \|\| PGM_GST_TYPE == PGM_TYPE_PROT)
1887
1888	# ifdef PGM_SYNC_N_PAGES
1889	/*
1890	* Get the shadow PDE, find the shadow page table in the pool.
1891	*/
1892	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
1893	X86PDE PdeDst = pgmShwGet32BitPDE(&pVM->pgm.s, GCPtrPage);
1894
1895	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
1896	X86PDEPAE PdeDst = pgmShwGetPaePDE(&pVM->pgm.s, GCPtrPage);
1897
1898	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
1899	const unsigned iPDDst = ((GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK);
1900	const unsigned iPdpt = (GCPtrPage >> X86_PDPT_SHIFT) & X86_PDPT_MASK_AMD64; NOREF(iPdpt);
1901	PX86PDPAE pPDDst;
1902	X86PDEPAE PdeDst;
1903	PX86PDPT pPdptDst;
1904
1905	int rc = pgmShwGetLongModePDPtr(pVM, GCPtrPage, NULL, &pPdptDst, &pPDDst);
1906	AssertRCSuccessReturn(rc, rc);
1907	Assert(pPDDst && pPdptDst);
1908	PdeDst = pPDDst->a[iPDDst];
1909	# elif PGM_SHW_TYPE == PGM_TYPE_EPT
1910	const unsigned iPDDst = ((GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK);
1911	PEPTPD pPDDst;
1912	EPTPDE PdeDst;
1913
1914	int rc = pgmShwGetEPTPDPtr(pVM, GCPtrPage, NULL, &pPDDst);
1915	if (rc != VINF_SUCCESS)
1916	{
1917	AssertRC(rc);
1918	return rc;
1919	}
1920	Assert(pPDDst);
1921	PdeDst = pPDDst->a[iPDDst];
1922	# endif
1923	AssertMsg(PdeDst.n.u1Present, ("%#llx\n", (uint64_t)PdeDst.u));
1924	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, PdeDst.u & SHW_PDE_PG_MASK);
1925	PSHWPT pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
1926
1927	Assert(cPages == 1 \|\| !(uErr & X86_TRAP_PF_P));
1928	if (cPages > 1 && !(uErr & X86_TRAP_PF_P))
1929	{
1930	/*
1931	* This code path is currently only taken when the caller is PGMTrap0eHandler
1932	* for non-present pages!
1933	*
1934	* We're setting PGM_SYNC_NR_PAGES pages around the faulting page to sync it and
1935	* deal with locality.
1936	*/
1937	unsigned iPTDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1938	const unsigned iPTDstEnd = RT_MIN(iPTDst + PGM_SYNC_NR_PAGES / 2, RT_ELEMENTS(pPTDst->a));
1939	if (iPTDst < PGM_SYNC_NR_PAGES / 2)
1940	iPTDst = 0;
1941	else
1942	iPTDst -= PGM_SYNC_NR_PAGES / 2;
1943	for (; iPTDst < iPTDstEnd; iPTDst++)
1944	{
1945	if (!pPTDst->a[iPTDst].n.u1Present)
1946	{
1947	GSTPTE PteSrc;
1948
1949	RTGCPTR GCPtrCurPage = (GCPtrPage & ~(RTGCPTR)(SHW_PT_MASK << SHW_PT_SHIFT)) \| (iPTDst << PAGE_SHIFT);
1950
1951	/* Fake the page table entry */
1952	PteSrc.u = GCPtrCurPage;
1953	PteSrc.n.u1Present = 1;
1954	PteSrc.n.u1Dirty = 1;
1955	PteSrc.n.u1Accessed = 1;
1956	PteSrc.n.u1Write = 1;
1957	PteSrc.n.u1User = 1;
1958
1959	PGM_BTH_NAME(SyncPageWorker)(pVM, &pPTDst->a[iPTDst], PdeSrc, PteSrc, pShwPage, iPTDst);
1960
1961	Log2(("SyncPage: 4K+ %RGv PteSrc:{P=%d RW=%d U=%d raw=%08llx} PteDst=%08llx%s\n",
1962	GCPtrCurPage, PteSrc.n.u1Present,
1963	PteSrc.n.u1Write & PdeSrc.n.u1Write,
1964	PteSrc.n.u1User & PdeSrc.n.u1User,
1965	(uint64_t)PteSrc.u,
1966	(uint64_t)pPTDst->a[iPTDst].u,
1967	pPTDst->a[iPTDst].u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
1968	}
1969	else
1970	Log4(("%RGv iPTDst=%x pPTDst->a[iPTDst] %RX64\n", (GCPtrPage & ~(RTGCPTR)(SHW_PT_MASK << SHW_PT_SHIFT)) \| (iPTDst << PAGE_SHIFT), iPTDst, pPTDst->a[iPTDst].u));
1971	}
1972	}
1973	else
1974	# endif /* PGM_SYNC_N_PAGES */
1975	{
1976	GSTPTE PteSrc;
1977	const unsigned iPTDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1978	RTGCPTR GCPtrCurPage = (GCPtrPage & ~(RTGCPTR)(SHW_PT_MASK << SHW_PT_SHIFT)) \| (iPTDst << PAGE_SHIFT);
1979
1980	/* Fake the page table entry */
1981	PteSrc.u = GCPtrCurPage;
1982	PteSrc.n.u1Present = 1;
1983	PteSrc.n.u1Dirty = 1;
1984	PteSrc.n.u1Accessed = 1;
1985	PteSrc.n.u1Write = 1;
1986	PteSrc.n.u1User = 1;
1987	PGM_BTH_NAME(SyncPageWorker)(pVM, &pPTDst->a[iPTDst], PdeSrc, PteSrc, pShwPage, iPTDst);
1988
1989	Log2(("SyncPage: 4K %RGv PteSrc:{P=%d RW=%d U=%d raw=%08llx}PteDst=%08llx%s\n",
1990	GCPtrPage, PteSrc.n.u1Present,
1991	PteSrc.n.u1Write & PdeSrc.n.u1Write,
1992	PteSrc.n.u1User & PdeSrc.n.u1User,
1993	(uint64_t)PteSrc.u,
1994	(uint64_t)pPTDst->a[iPTDst].u,
1995	pPTDst->a[iPTDst].u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
1996	}
1997	return VINF_SUCCESS;
1998
1999	#else
2000	AssertReleaseMsgFailed(("Shw=%d Gst=%d is not implemented!\n", PGM_GST_TYPE, PGM_SHW_TYPE));
2001	return VERR_INTERNAL_ERROR;
2002	#endif
2003	}
2004
2005
2006	#if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
2007	/**
2008	* Investigate page fault and handle write protection page faults caused by
2009	* dirty bit tracking.
2010	*
2011	* @returns VBox status code.
2012	* @param pVM VM handle.
2013	* @param uErr Page fault error code.
2014	* @param pPdeDst Shadow page directory entry.
2015	* @param pPdeSrc Guest page directory entry.
2016	* @param GCPtrPage Guest context page address.
2017	*/
2018	PGM_BTH_DECL(int, CheckPageFault)(PVM pVM, uint32_t uErr, PSHWPDE pPdeDst, PGSTPDE pPdeSrc, RTGCPTR GCPtrPage)
2019	{
2020	bool fWriteProtect = !!(CPUMGetGuestCR0(pVM) & X86_CR0_WP);
2021	bool fUserLevelFault = !!(uErr & X86_TRAP_PF_US);
2022	bool fWriteFault = !!(uErr & X86_TRAP_PF_RW);
2023	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2024	bool fBigPagesSupported = true;
2025	# else
2026	bool fBigPagesSupported = !!(CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
2027	# endif
2028	# if PGM_WITH_NX(PGM_GST_TYPE, PGM_SHW_TYPE)
2029	bool fNoExecuteBitValid = !!(CPUMGetGuestEFER(pVM) & MSR_K6_EFER_NXE);
2030	# endif
2031	unsigned uPageFaultLevel;
2032	int rc;
2033
2034	STAM_PROFILE_START(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2035	LogFlow(("CheckPageFault: GCPtrPage=%RGv uErr=%#x PdeSrc=%08x\n", GCPtrPage, uErr, pPdeSrc->u));
2036
2037	# if PGM_GST_TYPE == PGM_TYPE_PAE \
2038	\|\| PGM_GST_TYPE == PGM_TYPE_AMD64
2039
2040	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2041	PX86PML4E pPml4eSrc;
2042	PX86PDPE pPdpeSrc;
2043
2044	pPdpeSrc = pgmGstGetLongModePDPTPtr(&pVM->pgm.s, GCPtrPage, &pPml4eSrc);
2045	Assert(pPml4eSrc);
2046
2047	/*
2048	* Real page fault? (PML4E level)
2049	*/
2050	if ( (uErr & X86_TRAP_PF_RSVD)
2051	\|\| !pPml4eSrc->n.u1Present
2052	\|\| (fNoExecuteBitValid && (uErr & X86_TRAP_PF_ID) && pPml4eSrc->n.u1NoExecute)
2053	\|\| (fWriteFault && !pPml4eSrc->n.u1Write && (fUserLevelFault \|\| fWriteProtect))
2054	\|\| (fUserLevelFault && !pPml4eSrc->n.u1User)
2055	)
2056	{
2057	uPageFaultLevel = 0;
2058	goto l_UpperLevelPageFault;
2059	}
2060	Assert(pPdpeSrc);
2061
2062	# else /* PAE */
2063	PX86PDPE pPdpeSrc = pgmGstGetPaePDPEPtr(&pVM->pgm.s, GCPtrPage);
2064	# endif /* PAE */
2065
2066	/*
2067	* Real page fault? (PDPE level)
2068	*/
2069	if ( (uErr & X86_TRAP_PF_RSVD)
2070	\|\| !pPdpeSrc->n.u1Present
2071	# if PGM_GST_TYPE == PGM_TYPE_AMD64 /* NX, r/w, u/s bits in the PDPE are long mode only */
2072	\|\| (fNoExecuteBitValid && (uErr & X86_TRAP_PF_ID) && pPdpeSrc->lm.u1NoExecute)
2073	\|\| (fWriteFault && !pPdpeSrc->lm.u1Write && (fUserLevelFault \|\| fWriteProtect))
2074	\|\| (fUserLevelFault && !pPdpeSrc->lm.u1User)
2075	# endif
2076	)
2077	{
2078	uPageFaultLevel = 1;
2079	goto l_UpperLevelPageFault;
2080	}
2081	# endif
2082
2083	/*
2084	* Real page fault? (PDE level)
2085	*/
2086	if ( (uErr & X86_TRAP_PF_RSVD)
2087	\|\| !pPdeSrc->n.u1Present
2088	# if PGM_WITH_NX(PGM_GST_TYPE, PGM_SHW_TYPE)
2089	\|\| (fNoExecuteBitValid && (uErr & X86_TRAP_PF_ID) && pPdeSrc->n.u1NoExecute)
2090	# endif
2091	\|\| (fWriteFault && !pPdeSrc->n.u1Write && (fUserLevelFault \|\| fWriteProtect))
2092	\|\| (fUserLevelFault && !pPdeSrc->n.u1User) )
2093	{
2094	uPageFaultLevel = 2;
2095	goto l_UpperLevelPageFault;
2096	}
2097
2098	/*
2099	* First check the easy case where the page directory has been marked read-only to track
2100	* the dirty bit of an emulated BIG page
2101	*/
2102	if (pPdeSrc->b.u1Size && fBigPagesSupported)
2103	{
2104	/* Mark guest page directory as accessed */
2105	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2106	pPml4eSrc->n.u1Accessed = 1;
2107	pPdpeSrc->lm.u1Accessed = 1;
2108	# endif
2109	pPdeSrc->b.u1Accessed = 1;
2110
2111	/*
2112	* Only write protection page faults are relevant here.
2113	*/
2114	if (fWriteFault)
2115	{
2116	/* Mark guest page directory as dirty (BIG page only). */
2117	pPdeSrc->b.u1Dirty = 1;
2118
2119	if (pPdeDst->n.u1Present && (pPdeDst->u & PGM_PDFLAGS_TRACK_DIRTY))
2120	{
2121	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPageTrap));
2122
2123	Assert(pPdeSrc->b.u1Write);
2124
2125	pPdeDst->n.u1Write = 1;
2126	pPdeDst->n.u1Accessed = 1;
2127	pPdeDst->au32[0] &= ~PGM_PDFLAGS_TRACK_DIRTY;
2128	PGM_INVL_BIG_PG(GCPtrPage);
2129	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2130	return VINF_PGM_HANDLED_DIRTY_BIT_FAULT;
2131	}
2132	}
2133	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2134	return VINF_PGM_NO_DIRTY_BIT_TRACKING;
2135	}
2136	/* else: 4KB page table */
2137
2138	/*
2139	* Map the guest page table.
2140	*/
2141	PGSTPT pPTSrc;
2142	rc = PGM_GCPHYS_2_PTR(pVM, pPdeSrc->u & GST_PDE_PG_MASK, &pPTSrc);
2143	if (RT_SUCCESS(rc))
2144	{
2145	/*
2146	* Real page fault?
2147	*/
2148	PGSTPTE pPteSrc = &pPTSrc->a[(GCPtrPage >> GST_PT_SHIFT) & GST_PT_MASK];
2149	const GSTPTE PteSrc = *pPteSrc;
2150	if ( !PteSrc.n.u1Present
2151	# if PGM_WITH_NX(PGM_GST_TYPE, PGM_SHW_TYPE)
2152	\|\| (fNoExecuteBitValid && (uErr & X86_TRAP_PF_ID) && PteSrc.n.u1NoExecute)
2153	# endif
2154	\|\| (fWriteFault && !PteSrc.n.u1Write && (fUserLevelFault \|\| fWriteProtect))
2155	\|\| (fUserLevelFault && !PteSrc.n.u1User)
2156	)
2157	{
2158	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyTrackRealPF));
2159	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2160	LogFlow(("CheckPageFault: real page fault at %RGv PteSrc.u=%08x (2)\n", GCPtrPage, PteSrc.u));
2161
2162	/* Check the present bit as the shadow tables can cause different error codes by being out of sync.
2163	* See the 2nd case above as well.
2164	*/
2165	if (pPdeSrc->n.u1Present && pPteSrc->n.u1Present)
2166	TRPMSetErrorCode(pVM, uErr \| X86_TRAP_PF_P); /* page-level protection violation */
2167
2168	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2169	return VINF_EM_RAW_GUEST_TRAP;
2170	}
2171	LogFlow(("CheckPageFault: page fault at %RGv PteSrc.u=%08x\n", GCPtrPage, PteSrc.u));
2172
2173	/*
2174	* Set the accessed bits in the page directory and the page table.
2175	*/
2176	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2177	pPml4eSrc->n.u1Accessed = 1;
2178	pPdpeSrc->lm.u1Accessed = 1;
2179	# endif
2180	pPdeSrc->n.u1Accessed = 1;
2181	pPteSrc->n.u1Accessed = 1;
2182
2183	/*
2184	* Only write protection page faults are relevant here.
2185	*/
2186	if (fWriteFault)
2187	{
2188	/* Write access, so mark guest entry as dirty. */
2189	# ifdef VBOX_WITH_STATISTICS
2190	if (!pPteSrc->n.u1Dirty)
2191	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtiedPage));
2192	else
2193	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,PageAlreadyDirty));
2194	# endif
2195
2196	pPteSrc->n.u1Dirty = 1;
2197
2198	if (pPdeDst->n.u1Present)
2199	{
2200	#ifndef IN_RING0
2201	/* Bail out here as pgmPoolGetPageByHCPhys will return NULL and we'll crash below.
2202	* Our individual shadow handlers will provide more information and force a fatal exit.
2203	*/
2204	if (MMHyperIsInsideArea(pVM, (RTGCPTR)GCPtrPage))
2205	{
2206	LogRel(("CheckPageFault: write to hypervisor region %RGv\n", GCPtrPage));
2207	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2208	return VINF_SUCCESS;
2209	}
2210	#endif
2211	/*
2212	* Map shadow page table.
2213	*/
2214	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, pPdeDst->u & SHW_PDE_PG_MASK);
2215	if (pShwPage)
2216	{
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 /** @todo Optimize accessed bit emulation? */
2220	&& (pPteDst->u & PGM_PTFLAGS_TRACK_DIRTY))
2221	{
2222	LogFlow(("DIRTY page trap addr=%RGv\n", GCPtrPage));
2223	# ifdef VBOX_STRICT
2224	PPGMPAGE pPage = pgmPhysGetPage(&pVM->pgm.s, pPteSrc->u & GST_PTE_PG_MASK);
2225	if (pPage)
2226	AssertMsg(!PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage),
2227	("Unexpected dirty bit tracking on monitored page %RGv (phys %RGp)!!!!!!\n", GCPtrPage, pPteSrc->u & X86_PTE_PAE_PG_MASK));
2228	# endif
2229	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPageTrap));
2230
2231	Assert(pPteSrc->n.u1Write);
2232
2233	pPteDst->n.u1Write = 1;
2234	pPteDst->n.u1Dirty = 1;
2235	pPteDst->n.u1Accessed = 1;
2236	pPteDst->au32[0] &= ~PGM_PTFLAGS_TRACK_DIRTY;
2237	PGM_INVL_PG(GCPtrPage);
2238
2239	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2240	return VINF_PGM_HANDLED_DIRTY_BIT_FAULT;
2241	}
2242	}
2243	else
2244	AssertMsgFailed(("pgmPoolGetPageByHCPhys %RGp failed!\n", pPdeDst->u & SHW_PDE_PG_MASK));
2245	}
2246	}
2247	/** @todo Optimize accessed bit emulation? */
2248	# ifdef VBOX_STRICT
2249	/*
2250	* Sanity check.
2251	*/
2252	else if ( !pPteSrc->n.u1Dirty
2253	&& (pPdeSrc->n.u1Write & pPteSrc->n.u1Write)
2254	&& pPdeDst->n.u1Present)
2255	{
2256	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, pPdeDst->u & SHW_PDE_PG_MASK);
2257	PSHWPT pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
2258	PSHWPTE pPteDst = &pPTDst->a[(GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK];
2259	if ( pPteDst->n.u1Present
2260	&& pPteDst->n.u1Write)
2261	LogFlow(("Writable present page %RGv not marked for dirty bit tracking!!!\n", GCPtrPage));
2262	}
2263	# endif /* VBOX_STRICT */
2264	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2265	return VINF_PGM_NO_DIRTY_BIT_TRACKING;
2266	}
2267	AssertRC(rc);
2268	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2269	return rc;
2270
2271
2272	l_UpperLevelPageFault:
2273	/*
2274	* Pagefault detected while checking the PML4E, PDPE or PDE.
2275	* Single exit handler to get rid of duplicate code paths.
2276	*/
2277	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyTrackRealPF));
2278	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2279	Log(("CheckPageFault: real page fault at %RGv (%d)\n", GCPtrPage, uPageFaultLevel));
2280
2281	if (
2282	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2283	pPml4eSrc->n.u1Present &&
2284	# endif
2285	# if PGM_GST_TYPE == PGM_TYPE_AMD64 \|\| PGM_GST_TYPE == PGM_TYPE_PAE
2286	pPdpeSrc->n.u1Present &&
2287	# endif
2288	pPdeSrc->n.u1Present)
2289	{
2290	/* Check the present bit as the shadow tables can cause different error codes by being out of sync. */
2291	if (pPdeSrc->b.u1Size && fBigPagesSupported)
2292	{
2293	TRPMSetErrorCode(pVM, uErr \| X86_TRAP_PF_P); /* page-level protection violation */
2294	}
2295	else
2296	{
2297	/*
2298	* Map the guest page table.
2299	*/
2300	PGSTPT pPTSrc;
2301	rc = PGM_GCPHYS_2_PTR(pVM, pPdeSrc->u & GST_PDE_PG_MASK, &pPTSrc);
2302	if (RT_SUCCESS(rc))
2303	{
2304	PGSTPTE pPteSrc = &pPTSrc->a[(GCPtrPage >> GST_PT_SHIFT) & GST_PT_MASK];
2305	const GSTPTE PteSrc = *pPteSrc;
2306	if (pPteSrc->n.u1Present)
2307	TRPMSetErrorCode(pVM, uErr \| X86_TRAP_PF_P); /* page-level protection violation */
2308	}
2309	AssertRC(rc);
2310	}
2311	}
2312	return VINF_EM_RAW_GUEST_TRAP;
2313	}
2314	#endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
2315
2316
2317	/**
2318	* Sync a shadow page table.
2319	*
2320	* The shadow page table is not present. This includes the case where
2321	* there is a conflict with a mapping.
2322	*
2323	* @returns VBox status code.
2324	* @param pVM VM handle.
2325	* @param iPD Page directory index.
2326	* @param pPDSrc Source page directory (i.e. Guest OS page directory).
2327	* Assume this is a temporary mapping.
2328	* @param GCPtrPage GC Pointer of the page that caused the fault
2329	*/
2330	PGM_BTH_DECL(int, SyncPT)(PVM pVM, unsigned iPDSrc, PGSTPD pPDSrc, RTGCPTR GCPtrPage)
2331	{
2332	STAM_PROFILE_START(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT), a);
2333	STAM_COUNTER_INC(&pVM->pgm.s.StatSyncPtPD[iPDSrc]);
2334	LogFlow(("SyncPT: GCPtrPage=%RGv\n", GCPtrPage));
2335
2336	#if ( PGM_GST_TYPE == PGM_TYPE_32BIT \
2337	\|\| PGM_GST_TYPE == PGM_TYPE_PAE \
2338	\|\| PGM_GST_TYPE == PGM_TYPE_AMD64) \
2339	&& PGM_SHW_TYPE != PGM_TYPE_NESTED \
2340	&& PGM_SHW_TYPE != PGM_TYPE_EPT
2341
2342	int rc = VINF_SUCCESS;
2343
2344	/*
2345	* Validate input a little bit.
2346	*/
2347	AssertMsg(iPDSrc == ((GCPtrPage >> GST_PD_SHIFT) & GST_PD_MASK), ("iPDSrc=%x GCPtrPage=%RGv\n", iPDSrc, GCPtrPage));
2348	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
2349	const unsigned iPDDst = GCPtrPage >> SHW_PD_SHIFT;
2350	PSHWPDE pPdeDst = pgmShwGet32BitPDEPtr(&pVM->pgm.s, GCPtrPage);
2351
2352	/* Fetch the pgm pool shadow descriptor. */
2353	PPGMPOOLPAGE pShwPde = pVM->pgm.s.CTX_SUFF(pShwPageCR3);
2354	Assert(pShwPde);
2355
2356	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
2357	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
2358	PPGMPOOLPAGE pShwPde;
2359	PX86PDPAE pPDDst;
2360	PSHWPDE pPdeDst;
2361
2362	/* Fetch the pgm pool shadow descriptor. */
2363	rc = pgmShwGetPaePoolPagePD(&pVM->pgm.s, GCPtrPage, &pShwPde);
2364	AssertRCSuccessReturn(rc, rc);
2365	Assert(pShwPde);
2366
2367	pPDDst = (PX86PDPAE)PGMPOOL_PAGE_2_PTR_BY_PGM(&pVM->pgm.s, pShwPde);
2368	pPdeDst = &pPDDst->a[iPDDst];
2369
2370	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
2371	const unsigned iPdpt = (GCPtrPage >> X86_PDPT_SHIFT) & X86_PDPT_MASK_AMD64;
2372	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
2373	PX86PDPAE pPDDst;
2374	PX86PDPT pPdptDst;
2375	rc = pgmShwGetLongModePDPtr(pVM, GCPtrPage, NULL, &pPdptDst, &pPDDst);
2376	AssertRCSuccessReturn(rc, rc);
2377	Assert(pPDDst);
2378	PSHWPDE pPdeDst = &pPDDst->a[iPDDst];
2379	# endif
2380	SHWPDE PdeDst = *pPdeDst;
2381
2382	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2383	/* Fetch the pgm pool shadow descriptor. */
2384	PPGMPOOLPAGE pShwPde = pgmPoolGetPageByHCPhys(pVM, pPdptDst->a[iPdpt].u & X86_PDPE_PG_MASK);
2385	Assert(pShwPde);
2386	# endif
2387
2388	# ifndef PGM_WITHOUT_MAPPINGS
2389	/*
2390	* Check for conflicts.
2391	* GC: In case of a conflict we'll go to Ring-3 and do a full SyncCR3.
2392	* HC: Simply resolve the conflict.
2393	*/
2394	if (PdeDst.u & PGM_PDFLAGS_MAPPING)
2395	{
2396	Assert(pgmMapAreMappingsEnabled(&pVM->pgm.s));
2397	# ifndef IN_RING3
2398	Log(("SyncPT: Conflict at %RGv\n", GCPtrPage));
2399	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT), a);
2400	return VERR_ADDRESS_CONFLICT;
2401	# else
2402	PPGMMAPPING pMapping = pgmGetMapping(pVM, (RTGCPTR)GCPtrPage);
2403	Assert(pMapping);
2404	# if PGM_GST_TYPE == PGM_TYPE_32BIT
2405	int rc = pgmR3SyncPTResolveConflict(pVM, pMapping, pPDSrc, GCPtrPage & (GST_PD_MASK << GST_PD_SHIFT));
2406	# elif PGM_GST_TYPE == PGM_TYPE_PAE
2407	int rc = pgmR3SyncPTResolveConflictPAE(pVM, pMapping, GCPtrPage & (GST_PD_MASK << GST_PD_SHIFT));
2408	# else
2409	AssertFailed(); /* can't happen for amd64 */
2410	# endif
2411	if (RT_FAILURE(rc))
2412	{
2413	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT), a);
2414	return rc;
2415	}
2416	PdeDst = *pPdeDst;
2417	# endif
2418	}
2419	# else /* PGM_WITHOUT_MAPPINGS */
2420	Assert(!pgmMapAreMappingsEnabled(&pVM->pgm.s));
2421	# endif /* PGM_WITHOUT_MAPPINGS */
2422	Assert(!PdeDst.n.u1Present); /* We're only supposed to call SyncPT on PDE!P and conflicts.*/
2423
2424	# if defined(IN_RC)
2425	/* Make sure the dynamic pPdeDst mapping will not be reused during this function. */
2426	PGMDynLockHCPage(pVM, (uint8_t *)pPdeDst);
2427	# endif
2428
2429	/*
2430	* Sync page directory entry.
2431	*/
2432	GSTPDE PdeSrc = pPDSrc->a[iPDSrc];
2433	if (PdeSrc.n.u1Present)
2434	{
2435	/*
2436	* Allocate & map the page table.
2437	*/
2438	PSHWPT pPTDst;
2439	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2440	const bool fPageTable = !PdeSrc.b.u1Size;
2441	# else
2442	const bool fPageTable = !PdeSrc.b.u1Size \|\| !(CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
2443	# endif
2444	PPGMPOOLPAGE pShwPage;
2445	RTGCPHYS GCPhys;
2446	if (fPageTable)
2447	{
2448	GCPhys = PdeSrc.u & GST_PDE_PG_MASK;
2449	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
2450	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
2451	GCPhys \|= (iPDDst & 1) * (PAGE_SIZE / 2);
2452	# endif
2453	rc = pgmPoolAlloc(pVM, GCPhys, BTH_PGMPOOLKIND_PT_FOR_PT, pShwPde->idx, iPDDst, &pShwPage);
2454	}
2455	else
2456	{
2457	GCPhys = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc);
2458	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
2459	/* Select the right PDE as we're emulating a 4MB page directory with two 2 MB shadow PDEs.*/
2460	GCPhys \|= GCPtrPage & (1 << X86_PD_PAE_SHIFT);
2461	# endif
2462	rc = pgmPoolAlloc(pVM, GCPhys, BTH_PGMPOOLKIND_PT_FOR_BIG, pShwPde->idx, iPDDst, &pShwPage);
2463	}
2464	if (rc == VINF_SUCCESS)
2465	pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
2466	else if (rc == VINF_PGM_CACHED_PAGE)
2467	{
2468	/*
2469	* The PT was cached, just hook it up.
2470	*/
2471	if (fPageTable)
2472	PdeDst.u = pShwPage->Core.Key
2473	\| (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));
2474	else
2475	{
2476	PdeDst.u = pShwPage->Core.Key
2477	\| (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));
2478	/* (see explanation and assumptions further down.) */
2479	if (!PdeSrc.b.u1Dirty && PdeSrc.b.u1Write)
2480	{
2481	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPageBig));
2482	PdeDst.u \|= PGM_PDFLAGS_TRACK_DIRTY;
2483	PdeDst.b.u1Write = 0;
2484	}
2485	}
2486	*pPdeDst = PdeDst;
2487	# if defined(IN_RC)
2488	PGMDynUnlockHCPage(pVM, (uint8_t *)pPdeDst);
2489	# endif
2490	return VINF_SUCCESS;
2491	}
2492	else if (rc == VERR_PGM_POOL_FLUSHED)
2493	{
2494	VM_FF_SET(pVM, VM_FF_PGM_SYNC_CR3);
2495	# if defined(IN_RC)
2496	PGMDynUnlockHCPage(pVM, (uint8_t *)pPdeDst);
2497	# endif
2498	return VINF_PGM_SYNC_CR3;
2499	}
2500	else
2501	AssertMsgFailedReturn(("rc=%Rrc\n", rc), VERR_INTERNAL_ERROR);
2502	PdeDst.u &= X86_PDE_AVL_MASK;
2503	PdeDst.u \|= pShwPage->Core.Key;
2504
2505	/*
2506	* Page directory has been accessed (this is a fault situation, remember).
2507	*/
2508	pPDSrc->a[iPDSrc].n.u1Accessed = 1;
2509	if (fPageTable)
2510	{
2511	/*
2512	* Page table - 4KB.
2513	*
2514	* Sync all or just a few entries depending on PGM_SYNC_N_PAGES.
2515	*/
2516	Log2(("SyncPT: 4K %RGv PdeSrc:{P=%d RW=%d U=%d raw=%08llx}\n",
2517	GCPtrPage, PdeSrc.b.u1Present, PdeSrc.b.u1Write, PdeSrc.b.u1User, (uint64_t)PdeSrc.u));
2518	PGSTPT pPTSrc;
2519	rc = PGM_GCPHYS_2_PTR(pVM, PdeSrc.u & GST_PDE_PG_MASK, &pPTSrc);
2520	if (RT_SUCCESS(rc))
2521	{
2522	/*
2523	* Start by syncing the page directory entry so CSAM's TLB trick works.
2524	*/
2525	PdeDst.u = (PdeDst.u & (SHW_PDE_PG_MASK \| X86_PDE_AVL_MASK))
2526	\| (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));
2527	*pPdeDst = PdeDst;
2528	# if defined(IN_RC)
2529	PGMDynUnlockHCPage(pVM, (uint8_t *)pPdeDst);
2530	# endif
2531
2532	/*
2533	* Directory/page user or supervisor privilege: (same goes for read/write)
2534	*
2535	* Directory Page Combined
2536	* U/S U/S U/S
2537	* 0 0 0
2538	* 0 1 0
2539	* 1 0 0
2540	* 1 1 1
2541	*
2542	* Simple AND operation. Table listed for completeness.
2543	*
2544	*/
2545	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT4K));
2546	# ifdef PGM_SYNC_N_PAGES
2547	unsigned iPTBase = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
2548	unsigned iPTDst = iPTBase;
2549	const unsigned iPTDstEnd = RT_MIN(iPTDst + PGM_SYNC_NR_PAGES / 2, RT_ELEMENTS(pPTDst->a));
2550	if (iPTDst <= PGM_SYNC_NR_PAGES / 2)
2551	iPTDst = 0;
2552	else
2553	iPTDst -= PGM_SYNC_NR_PAGES / 2;
2554	# else /* !PGM_SYNC_N_PAGES */
2555	unsigned iPTDst = 0;
2556	const unsigned iPTDstEnd = RT_ELEMENTS(pPTDst->a);
2557	# endif /* !PGM_SYNC_N_PAGES */
2558	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
2559	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
2560	const unsigned offPTSrc = ((GCPtrPage >> SHW_PD_SHIFT) & 1) * 512;
2561	# else
2562	const unsigned offPTSrc = 0;
2563	# endif
2564	for (; iPTDst < iPTDstEnd; iPTDst++)
2565	{
2566	const unsigned iPTSrc = iPTDst + offPTSrc;
2567	const GSTPTE PteSrc = pPTSrc->a[iPTSrc];
2568
2569	if (PteSrc.n.u1Present) /* we've already cleared it above */
2570	{
2571	# ifndef IN_RING0
2572	/*
2573	* Assuming kernel code will be marked as supervisor - and not as user level
2574	* and executed using a conforming code selector - And marked as readonly.
2575	* Also assume that if we're monitoring a page, it's of no interest to CSAM.
2576	*/
2577	PPGMPAGE pPage;
2578	if ( ((PdeSrc.u & pPTSrc->a[iPTSrc].u) & (X86_PTE_RW \| X86_PTE_US))
2579	\|\| !CSAMDoesPageNeedScanning(pVM, (RTRCPTR)((iPDSrc << GST_PD_SHIFT) \| (iPTSrc << PAGE_SHIFT)))
2580	\|\| ( (pPage = pgmPhysGetPage(&pVM->pgm.s, PteSrc.u & GST_PTE_PG_MASK))
2581	&& PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
2582	)
2583	# endif
2584	PGM_BTH_NAME(SyncPageWorker)(pVM, &pPTDst->a[iPTDst], PdeSrc, PteSrc, pShwPage, iPTDst);
2585	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",
2586	(RTGCPTR)((iPDSrc << GST_PD_SHIFT) \| (iPTSrc << PAGE_SHIFT)),
2587	PteSrc.n.u1Present,
2588	PteSrc.n.u1Write & PdeSrc.n.u1Write,
2589	PteSrc.n.u1User & PdeSrc.n.u1User,
2590	(uint64_t)PteSrc.u,
2591	pPTDst->a[iPTDst].u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : "", pPTDst->a[iPTDst].u, iPTSrc, PdeSrc.au32[0],
2592	(RTGCPHYS)((PdeSrc.u & GST_PDE_PG_MASK) + iPTSrc*sizeof(PteSrc)) ));
2593	}
2594	} /* for PTEs */
2595	}
2596	}
2597	else
2598	{
2599	/*
2600	* Big page - 2/4MB.
2601	*
2602	* We'll walk the ram range list in parallel and optimize lookups.
2603	* We will only sync on shadow page table at a time.
2604	*/
2605	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT4M));
2606
2607	/**
2608	* @todo It might be more efficient to sync only a part of the 4MB page (similar to what we do for 4kb PDs).
2609	*/
2610
2611	/*
2612	* Start by syncing the page directory entry.
2613	*/
2614	PdeDst.u = (PdeDst.u & (SHW_PDE_PG_MASK \| (X86_PDE_AVL_MASK & ~PGM_PDFLAGS_TRACK_DIRTY)))
2615	\| (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));
2616
2617	/*
2618	* If the page is not flagged as dirty and is writable, then make it read-only
2619	* at PD level, so we can set the dirty bit when the page is modified.
2620	*
2621	* ASSUMES that page access handlers are implemented on page table entry level.
2622	* Thus we will first catch the dirty access and set PDE.D and restart. If
2623	* there is an access handler, we'll trap again and let it work on the problem.
2624	*/
2625	/** @todo move the above stuff to a section in the PGM documentation. */
2626	Assert(!(PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY));
2627	if (!PdeSrc.b.u1Dirty && PdeSrc.b.u1Write)
2628	{
2629	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPageBig));
2630	PdeDst.u \|= PGM_PDFLAGS_TRACK_DIRTY;
2631	PdeDst.b.u1Write = 0;
2632	}
2633	*pPdeDst = PdeDst;
2634	# if defined(IN_RC)
2635	PGMDynUnlockHCPage(pVM, (uint8_t *)pPdeDst);
2636	# endif
2637
2638	/*
2639	* Fill the shadow page table.
2640	*/
2641	/* Get address and flags from the source PDE. */
2642	SHWPTE PteDstBase;
2643	PteDstBase.u = PdeSrc.u & ~(GST_PDE_PG_MASK \| X86_PTE_AVL_MASK \| X86_PTE_PAT \| X86_PTE_PCD \| X86_PTE_PWT);
2644
2645	/* Loop thru the entries in the shadow PT. */
2646	const RTGCPTR GCPtr = (GCPtrPage >> SHW_PD_SHIFT) << SHW_PD_SHIFT; NOREF(GCPtr);
2647	Log2(("SyncPT: BIG %RGv PdeSrc:{P=%d RW=%d U=%d raw=%08llx} Shw=%RGv GCPhys=%RGp %s\n",
2648	GCPtrPage, PdeSrc.b.u1Present, PdeSrc.b.u1Write, PdeSrc.b.u1User, (uint64_t)PdeSrc.u, GCPtr,
2649	GCPhys, PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
2650	PPGMRAMRANGE pRam = pVM->pgm.s.CTX_SUFF(pRamRanges);
2651	unsigned iPTDst = 0;
2652	while (iPTDst < RT_ELEMENTS(pPTDst->a))
2653	{
2654	/* Advance ram range list. */
2655	while (pRam && GCPhys > pRam->GCPhysLast)
2656	pRam = pRam->CTX_SUFF(pNext);
2657	if (pRam && GCPhys >= pRam->GCPhys)
2658	{
2659	unsigned iHCPage = (GCPhys - pRam->GCPhys) >> PAGE_SHIFT;
2660	do
2661	{
2662	/* Make shadow PTE. */
2663	PPGMPAGE pPage = &pRam->aPages[iHCPage];
2664	SHWPTE PteDst;
2665
2666	# ifdef VBOX_WITH_NEW_PHYS_CODE
2667	# ifndef VBOX_WITH_NEW_LAZY_PAGE_ALLOC
2668	/* Try make the page writable if necessary. */
2669	if ( PteDstBase.n.u1Write
2670	&& PGM_PAGE_GET_STATE(pPage) != PGM_PAGE_STATE_ALLOCATED
2671	&& PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM)
2672	{
2673	rc = pgmPhysPageMakeWritableUnlocked(pVM, pPage, GCPhys);
2674	AssertRCReturn(rc, rc);
2675	}
2676	# endif
2677	# else /* !VBOX_WITH_NEW_PHYS_CODE */
2678	/* Make sure the RAM has already been allocated. */
2679	if (pRam->fFlags & MM_RAM_FLAGS_DYNAMIC_ALLOC) /** @todo PAGE FLAGS */
2680	{
2681	if (RT_UNLIKELY(!PGM_PAGE_GET_HCPHYS(pPage)))
2682	{
2683	# ifdef IN_RING3
2684	int rc = pgmr3PhysGrowRange(pVM, GCPhys);
2685	# else
2686	int rc = CTXALLMID(VMM, CallHost)(pVM, VMMCALLHOST_PGM_RAM_GROW_RANGE, GCPhys);
2687	# endif
2688	if (rc != VINF_SUCCESS)
2689	return rc;
2690	}
2691	}
2692	# endif /* !VBOX_WITH_NEW_PHYS_CODE */
2693
2694	if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
2695	{
2696	if (!PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage))
2697	{
2698	PteDst.u = PGM_PAGE_GET_HCPHYS(pPage) \| PteDstBase.u;
2699	PteDst.n.u1Write = 0;
2700	}
2701	else
2702	PteDst.u = 0;
2703	}
2704	# ifndef IN_RING0
2705	/*
2706	* Assuming kernel code will be marked as supervisor and not as user level and executed
2707	* using a conforming code selector. Don't check for readonly, as that implies the whole
2708	* 4MB can be code or readonly data. Linux enables write access for its large pages.
2709	*/
2710	else if ( !PdeSrc.n.u1User
2711	&& CSAMDoesPageNeedScanning(pVM, (RTRCPTR)(GCPtr \| (iPTDst << SHW_PT_SHIFT))))
2712	PteDst.u = 0;
2713	# endif
2714	else
2715	PteDst.u = PGM_PAGE_GET_HCPHYS(pPage) \| PteDstBase.u;
2716
2717	# ifdef VBOX_WITH_NEW_PHYS_CODE
2718	/* Only map writable pages writable. */
2719	if ( PteDst.n.u1Write
2720	&& PteDst.n.u1Present
2721	&& PGM_PAGE_GET_STATE(pPage) != PGM_PAGE_STATE_ALLOCATED)
2722	{
2723	PteDst.n.u1Write = 0; /** @todo this isn't quite working yet... */
2724	Log3(("SyncPT: write-protecting %RGp pPage=%R[pgmpage] at %RGv\n", GCPhys, pPage, (RTGCPTR)(GCPtr \| (iPTDst << SHW_PT_SHIFT))));
2725	}
2726	# endif
2727
2728	# ifdef PGMPOOL_WITH_USER_TRACKING
2729	if (PteDst.n.u1Present)
2730	PGM_BTH_NAME(SyncPageWorkerTrackAddref)(pVM, pShwPage, PGM_PAGE_GET_TRACKING(pPage), pPage, iPTDst);
2731	# endif
2732	/* commit it */
2733	pPTDst->a[iPTDst] = PteDst;
2734	Log4(("SyncPT: BIG %RGv PteDst:{P=%d RW=%d U=%d raw=%08llx}%s\n",
2735	(RTGCPTR)(GCPtr \| (iPTDst << SHW_PT_SHIFT)), PteDst.n.u1Present, PteDst.n.u1Write, PteDst.n.u1User, (uint64_t)PteDst.u,
2736	PteDst.u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
2737
2738	/* advance */
2739	GCPhys += PAGE_SIZE;
2740	iHCPage++;
2741	iPTDst++;
2742	} while ( iPTDst < RT_ELEMENTS(pPTDst->a)
2743	&& GCPhys <= pRam->GCPhysLast);
2744	}
2745	else if (pRam)
2746	{
2747	Log(("Invalid pages at %RGp\n", GCPhys));
2748	do
2749	{
2750	pPTDst->a[iPTDst].u = 0; /* MMIO or invalid page, we must handle them manually. */
2751	GCPhys += PAGE_SIZE;
2752	iPTDst++;
2753	} while ( iPTDst < RT_ELEMENTS(pPTDst->a)
2754	&& GCPhys < pRam->GCPhys);
2755	}
2756	else
2757	{
2758	Log(("Invalid pages at %RGp (2)\n", GCPhys));
2759	for ( ; iPTDst < RT_ELEMENTS(pPTDst->a); iPTDst++)
2760	pPTDst->a[iPTDst].u = 0; /* MMIO or invalid page, we must handle them manually. */
2761	}
2762	} /* while more PTEs */
2763	} /* 4KB / 4MB */
2764	}
2765	else
2766	AssertRelease(!PdeDst.n.u1Present);
2767
2768	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT), a);
2769	if (RT_FAILURE(rc))
2770	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPTFailed));
2771	return rc;
2772
2773	#elif (PGM_GST_TYPE == PGM_TYPE_REAL \|\| PGM_GST_TYPE == PGM_TYPE_PROT) \
2774	&& PGM_SHW_TYPE != PGM_TYPE_NESTED \
2775	&& (PGM_SHW_TYPE != PGM_TYPE_EPT \|\| PGM_GST_TYPE == PGM_TYPE_PROT)
2776
2777
2778	/*
2779	* Validate input a little bit.
2780	*/
2781	int rc = VINF_SUCCESS;
2782	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
2783	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
2784	PSHWPDE pPdeDst = pgmShwGet32BitPDEPtr(&pVM->pgm.s, GCPtrPage);
2785
2786	/* Fetch the pgm pool shadow descriptor. */
2787	PPGMPOOLPAGE pShwPde = pVM->pgm.s.CTX_SUFF(pShwPageCR3);
2788	Assert(pShwPde);
2789
2790	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
2791	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
2792	PPGMPOOLPAGE pShwPde;
2793	PX86PDPAE pPDDst;
2794	PSHWPDE pPdeDst;
2795
2796	/* Fetch the pgm pool shadow descriptor. */
2797	rc = pgmShwGetPaePoolPagePD(&pVM->pgm.s, GCPtrPage, &pShwPde);
2798	AssertRCSuccessReturn(rc, rc);
2799	Assert(pShwPde);
2800
2801	pPDDst = (PX86PDPAE)PGMPOOL_PAGE_2_PTR_BY_PGM(&pVM->pgm.s, pShwPde);
2802	pPdeDst = &pPDDst->a[iPDDst];
2803
2804	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
2805	const unsigned iPdpt = (GCPtrPage >> X86_PDPT_SHIFT) & X86_PDPT_MASK_AMD64;
2806	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
2807	PX86PDPAE pPDDst;
2808	PX86PDPT pPdptDst;
2809	rc = pgmShwGetLongModePDPtr(pVM, GCPtrPage, NULL, &pPdptDst, &pPDDst);
2810	AssertRCSuccessReturn(rc, rc);
2811	Assert(pPDDst);
2812	PSHWPDE pPdeDst = &pPDDst->a[iPDDst];
2813
2814	/* Fetch the pgm pool shadow descriptor. */
2815	PPGMPOOLPAGE pShwPde = pgmPoolGetPageByHCPhys(pVM, pPdptDst->a[iPdpt].u & X86_PDPE_PG_MASK);
2816	Assert(pShwPde);
2817
2818	# elif PGM_SHW_TYPE == PGM_TYPE_EPT
2819	const unsigned iPdpt = (GCPtrPage >> EPT_PDPT_SHIFT) & EPT_PDPT_MASK;
2820	const unsigned iPDDst = ((GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK);
2821	PEPTPD pPDDst;
2822	PEPTPDPT pPdptDst;
2823
2824	rc = pgmShwGetEPTPDPtr(pVM, GCPtrPage, &pPdptDst, &pPDDst);
2825	if (rc != VINF_SUCCESS)
2826	{
2827	AssertRC(rc);
2828	return rc;
2829	}
2830	Assert(pPDDst);
2831	PSHWPDE pPdeDst = &pPDDst->a[iPDDst];
2832
2833	/* Fetch the pgm pool shadow descriptor. */
2834	PPGMPOOLPAGE pShwPde = pgmPoolGetPageByHCPhys(pVM, pPdptDst->a[iPdpt].u & EPT_PDPTE_PG_MASK);
2835	Assert(pShwPde);
2836	# endif
2837	SHWPDE PdeDst = *pPdeDst;
2838
2839	Assert(!(PdeDst.u & PGM_PDFLAGS_MAPPING));
2840	Assert(!PdeDst.n.u1Present); /* We're only supposed to call SyncPT on PDE!P and conflicts.*/
2841
2842	GSTPDE PdeSrc;
2843	PdeSrc.au32[0] = 0; /* faked so we don't have to #ifdef everything */
2844	PdeSrc.n.u1Present = 1;
2845	PdeSrc.n.u1Write = 1;
2846	PdeSrc.n.u1Accessed = 1;
2847	PdeSrc.n.u1User = 1;
2848
2849	/*
2850	* Allocate & map the page table.
2851	*/
2852	PSHWPT pPTDst;
2853	PPGMPOOLPAGE pShwPage;
2854	RTGCPHYS GCPhys;
2855
2856	/* Virtual address = physical address */
2857	GCPhys = GCPtrPage & X86_PAGE_4K_BASE_MASK;
2858	rc = pgmPoolAlloc(pVM, GCPhys & ~(RT_BIT_64(SHW_PD_SHIFT) - 1), BTH_PGMPOOLKIND_PT_FOR_PT, pShwPde->idx, iPDDst, &pShwPage);
2859
2860	if ( rc == VINF_SUCCESS
2861	\|\| rc == VINF_PGM_CACHED_PAGE)
2862	pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
2863	else
2864	AssertMsgFailedReturn(("rc=%Rrc\n", rc), VERR_INTERNAL_ERROR);
2865
2866	PdeDst.u &= X86_PDE_AVL_MASK;
2867	PdeDst.u \|= pShwPage->Core.Key;
2868	PdeDst.n.u1Present = 1;
2869	PdeDst.n.u1Write = 1;
2870	# if PGM_SHW_TYPE == PGM_TYPE_EPT
2871	PdeDst.n.u1Execute = 1;
2872	# else
2873	PdeDst.n.u1User = 1;
2874	PdeDst.n.u1Accessed = 1;
2875	# endif
2876	*pPdeDst = PdeDst;
2877
2878	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, GCPtrPage, PGM_SYNC_NR_PAGES, 0 /* page not present */);
2879	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT), a);
2880	return rc;
2881
2882	#else
2883	AssertReleaseMsgFailed(("Shw=%d Gst=%d is not implemented!\n", PGM_SHW_TYPE, PGM_GST_TYPE));
2884	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT), a);
2885	return VERR_INTERNAL_ERROR;
2886	#endif
2887	}
2888
2889
2890
2891	/**
2892	* Prefetch a page/set of pages.
2893	*
2894	* Typically used to sync commonly used pages before entering raw mode
2895	* after a CR3 reload.
2896	*
2897	* @returns VBox status code.
2898	* @param pVM VM handle.
2899	* @param GCPtrPage Page to invalidate.
2900	*/
2901	PGM_BTH_DECL(int, PrefetchPage)(PVM pVM, RTGCPTR GCPtrPage)
2902	{
2903	#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) \
2904	&& PGM_SHW_TYPE != PGM_TYPE_NESTED && PGM_SHW_TYPE != PGM_TYPE_EPT
2905	/*
2906	* Check that all Guest levels thru the PDE are present, getting the
2907	* PD and PDE in the processes.
2908	*/
2909	int rc = VINF_SUCCESS;
2910	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
2911	# if PGM_GST_TYPE == PGM_TYPE_32BIT
2912	const unsigned iPDSrc = GCPtrPage >> GST_PD_SHIFT;
2913	PGSTPD pPDSrc = pgmGstGet32bitPDPtr(&pVM->pgm.s);
2914	# elif PGM_GST_TYPE == PGM_TYPE_PAE
2915	unsigned iPDSrc;
2916	X86PDPE PdpeSrc;
2917	PGSTPD pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, GCPtrPage, &iPDSrc, &PdpeSrc);
2918	if (!pPDSrc)
2919	return VINF_SUCCESS; /* not present */
2920	# elif PGM_GST_TYPE == PGM_TYPE_AMD64
2921	unsigned iPDSrc;
2922	PX86PML4E pPml4eSrc;
2923	X86PDPE PdpeSrc;
2924	PGSTPD pPDSrc = pgmGstGetLongModePDPtr(&pVM->pgm.s, GCPtrPage, &pPml4eSrc, &PdpeSrc, &iPDSrc);
2925	if (!pPDSrc)
2926	return VINF_SUCCESS; /* not present */
2927	# endif
2928	const GSTPDE PdeSrc = pPDSrc->a[iPDSrc];
2929	# else
2930	PGSTPD pPDSrc = NULL;
2931	const unsigned iPDSrc = 0;
2932	GSTPDE PdeSrc;
2933
2934	PdeSrc.au32[0] = 0; /* faked so we don't have to #ifdef everything */
2935	PdeSrc.n.u1Present = 1;
2936	PdeSrc.n.u1Write = 1;
2937	PdeSrc.n.u1Accessed = 1;
2938	PdeSrc.n.u1User = 1;
2939	# endif
2940
2941	if (PdeSrc.n.u1Present && PdeSrc.n.u1Accessed)
2942	{
2943	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
2944	const X86PDE PdeDst = pgmShwGet32BitPDE(&pVM->pgm.s, GCPtrPage);
2945	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
2946	const unsigned iPDDst = ((GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK);
2947	PX86PDPAE pPDDst;
2948	X86PDEPAE PdeDst;
2949	# if PGM_GST_TYPE != PGM_TYPE_PAE
2950	X86PDPE PdpeSrc;
2951
2952	/* Fake PDPT entry; access control handled on the page table level, so allow everything. */
2953	PdpeSrc.u = X86_PDPE_P; /* rw/us are reserved for PAE pdpte's; accessed bit causes invalid VT-x guest state errors */
2954	# endif
2955	int rc = pgmShwSyncPaePDPtr(pVM, GCPtrPage, &PdpeSrc, &pPDDst);
2956	if (rc != VINF_SUCCESS)
2957	{
2958	AssertRC(rc);
2959	return rc;
2960	}
2961	Assert(pPDDst);
2962	PdeDst = pPDDst->a[iPDDst];
2963
2964	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
2965	const unsigned iPDDst = ((GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK);
2966	PX86PDPAE pPDDst;
2967	X86PDEPAE PdeDst;
2968
2969	# if PGM_GST_TYPE == PGM_TYPE_PROT
2970	/* AMD-V nested paging */
2971	X86PML4E Pml4eSrc;
2972	X86PDPE PdpeSrc;
2973	PX86PML4E pPml4eSrc = &Pml4eSrc;
2974
2975	/* Fake PML4 & PDPT entry; access control handled on the page table level, so allow everything. */
2976	Pml4eSrc.u = X86_PML4E_P \| X86_PML4E_RW \| X86_PML4E_US \| X86_PML4E_NX \| X86_PML4E_A;
2977	PdpeSrc.u = X86_PDPE_P \| X86_PDPE_RW \| X86_PDPE_US \| X86_PDPE_NX \| X86_PDPE_A;
2978	# endif
2979
2980	int rc = pgmShwSyncLongModePDPtr(pVM, GCPtrPage, pPml4eSrc, &PdpeSrc, &pPDDst);
2981	if (rc != VINF_SUCCESS)
2982	{
2983	AssertRC(rc);
2984	return rc;
2985	}
2986	Assert(pPDDst);
2987	PdeDst = pPDDst->a[iPDDst];
2988	# endif
2989	if (!(PdeDst.u & PGM_PDFLAGS_MAPPING))
2990	{
2991	if (!PdeDst.n.u1Present)
2992	/** r=bird: This guy will set the A bit on the PDE, probably harmless. */
2993	rc = PGM_BTH_NAME(SyncPT)(pVM, iPDSrc, pPDSrc, GCPtrPage);
2994	else
2995	{
2996	/** @note We used to sync PGM_SYNC_NR_PAGES pages, which triggered assertions in CSAM, because
2997	* R/W attributes of nearby pages were reset. Not sure how that could happen. Anyway, it
2998	* makes no sense to prefetch more than one page.
2999	*/
3000	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, GCPtrPage, 1, 0);
3001	if (RT_SUCCESS(rc))
3002	rc = VINF_SUCCESS;
3003	}
3004	}
3005	}
3006	return rc;
3007
3008	#elif PGM_SHW_TYPE == PGM_TYPE_NESTED \|\| PGM_SHW_TYPE == PGM_TYPE_EPT
3009	return VINF_SUCCESS; /* ignore */
3010	#endif
3011	}
3012
3013
3014
3015
3016	/**
3017	* Syncs a page during a PGMVerifyAccess() call.
3018	*
3019	* @returns VBox status code (informational included).
3020	* @param GCPtrPage The address of the page to sync.
3021	* @param fPage The effective guest page flags.
3022	* @param uErr The trap error code.
3023	*/
3024	PGM_BTH_DECL(int, VerifyAccessSyncPage)(PVM pVM, RTGCPTR GCPtrPage, unsigned fPage, unsigned uErr)
3025	{
3026	LogFlow(("VerifyAccessSyncPage: GCPtrPage=%RGv fPage=%#x uErr=%#x\n", GCPtrPage, fPage, uErr));
3027
3028	Assert(!HWACCMIsNestedPagingActive(pVM));
3029	#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) \
3030	&& PGM_SHW_TYPE != PGM_TYPE_NESTED && PGM_SHW_TYPE != PGM_TYPE_EPT
3031
3032	# ifndef IN_RING0
3033	if (!(fPage & X86_PTE_US))
3034	{
3035	/*
3036	* Mark this page as safe.
3037	*/
3038	/** @todo not correct for pages that contain both code and data!! */
3039	Log(("CSAMMarkPage %RGv; scanned=%d\n", GCPtrPage, true));
3040	CSAMMarkPage(pVM, (RTRCPTR)GCPtrPage, true);
3041	}
3042	# endif
3043
3044	/*
3045	* Get guest PD and index.
3046	*/
3047	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
3048	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3049	const unsigned iPDSrc = GCPtrPage >> GST_PD_SHIFT;
3050	PGSTPD pPDSrc = pgmGstGet32bitPDPtr(&pVM->pgm.s);
3051	# elif PGM_GST_TYPE == PGM_TYPE_PAE
3052	unsigned iPDSrc;
3053	X86PDPE PdpeSrc;
3054	PGSTPD pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, GCPtrPage, &iPDSrc, &PdpeSrc);
3055
3056	if (pPDSrc)
3057	{
3058	Log(("PGMVerifyAccess: access violation for %RGv due to non-present PDPTR\n", GCPtrPage));
3059	return VINF_EM_RAW_GUEST_TRAP;
3060	}
3061	# elif PGM_GST_TYPE == PGM_TYPE_AMD64
3062	unsigned iPDSrc;
3063	PX86PML4E pPml4eSrc;
3064	X86PDPE PdpeSrc;
3065	PGSTPD pPDSrc = pgmGstGetLongModePDPtr(&pVM->pgm.s, GCPtrPage, &pPml4eSrc, &PdpeSrc, &iPDSrc);
3066	if (!pPDSrc)
3067	{
3068	Log(("PGMVerifyAccess: access violation for %RGv due to non-present PDPTR\n", GCPtrPage));
3069	return VINF_EM_RAW_GUEST_TRAP;
3070	}
3071	# endif
3072	# else
3073	PGSTPD pPDSrc = NULL;
3074	const unsigned iPDSrc = 0;
3075	# endif
3076	int rc = VINF_SUCCESS;
3077
3078	/*
3079	* First check if the shadow pd is present.
3080	*/
3081	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
3082	PX86PDE pPdeDst = pgmShwGet32BitPDEPtr(&pVM->pgm.s, GCPtrPage);
3083	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
3084	PX86PDEPAE pPdeDst;
3085	const unsigned iPDDst = ((GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK);
3086	PX86PDPAE pPDDst;
3087	# if PGM_GST_TYPE != PGM_TYPE_PAE
3088	X86PDPE PdpeSrc;
3089
3090	/* Fake PDPT entry; access control handled on the page table level, so allow everything. */
3091	PdpeSrc.u = X86_PDPE_P; /* rw/us are reserved for PAE pdpte's; accessed bit causes invalid VT-x guest state errors */
3092	# endif
3093	rc = pgmShwSyncPaePDPtr(pVM, GCPtrPage, &PdpeSrc, &pPDDst);
3094	if (rc != VINF_SUCCESS)
3095	{
3096	AssertRC(rc);
3097	return rc;
3098	}
3099	Assert(pPDDst);
3100	pPdeDst = &pPDDst->a[iPDDst];
3101
3102	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
3103	const unsigned iPDDst = ((GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK);
3104	PX86PDPAE pPDDst;
3105	PX86PDEPAE pPdeDst;
3106
3107	# if PGM_GST_TYPE == PGM_TYPE_PROT
3108	/* AMD-V nested paging */
3109	X86PML4E Pml4eSrc;
3110	X86PDPE PdpeSrc;
3111	PX86PML4E pPml4eSrc = &Pml4eSrc;
3112
3113	/* Fake PML4 & PDPT entry; access control handled on the page table level, so allow everything. */
3114	Pml4eSrc.u = X86_PML4E_P \| X86_PML4E_RW \| X86_PML4E_US \| X86_PML4E_NX \| X86_PML4E_A;
3115	PdpeSrc.u = X86_PDPE_P \| X86_PDPE_RW \| X86_PDPE_US \| X86_PDPE_NX \| X86_PDPE_A;
3116	# endif
3117
3118	rc = pgmShwSyncLongModePDPtr(pVM, GCPtrPage, pPml4eSrc, &PdpeSrc, &pPDDst);
3119	if (rc != VINF_SUCCESS)
3120	{
3121	AssertRC(rc);
3122	return rc;
3123	}
3124	Assert(pPDDst);
3125	pPdeDst = &pPDDst->a[iPDDst];
3126	# endif
3127	if (!pPdeDst->n.u1Present)
3128	{
3129	rc = PGM_BTH_NAME(SyncPT)(pVM, iPDSrc, pPDSrc, GCPtrPage);
3130	AssertRC(rc);
3131	if (rc != VINF_SUCCESS)
3132	return rc;
3133	}
3134
3135	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
3136	/* Check for dirty bit fault */
3137	rc = PGM_BTH_NAME(CheckPageFault)(pVM, uErr, pPdeDst, &pPDSrc->a[iPDSrc], GCPtrPage);
3138	if (rc == VINF_PGM_HANDLED_DIRTY_BIT_FAULT)
3139	Log(("PGMVerifyAccess: success (dirty)\n"));
3140	else
3141	{
3142	GSTPDE PdeSrc = pPDSrc->a[iPDSrc];
3143	#else
3144	{
3145	GSTPDE PdeSrc;
3146	PdeSrc.au32[0] = 0; /* faked so we don't have to #ifdef everything */
3147	PdeSrc.n.u1Present = 1;
3148	PdeSrc.n.u1Write = 1;
3149	PdeSrc.n.u1Accessed = 1;
3150	PdeSrc.n.u1User = 1;
3151
3152	#endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
3153	Assert(rc != VINF_EM_RAW_GUEST_TRAP);
3154	if (uErr & X86_TRAP_PF_US)
3155	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,PageOutOfSyncUser));
3156	else /* supervisor */
3157	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,PageOutOfSyncSupervisor));
3158
3159	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, GCPtrPage, 1, 0);
3160	if (RT_SUCCESS(rc))
3161	{
3162	/* Page was successfully synced */
3163	Log2(("PGMVerifyAccess: success (sync)\n"));
3164	rc = VINF_SUCCESS;
3165	}
3166	else
3167	{
3168	Log(("PGMVerifyAccess: access violation for %RGv rc=%d\n", GCPtrPage, rc));
3169	return VINF_EM_RAW_GUEST_TRAP;
3170	}
3171	}
3172	return rc;
3173
3174	#else /* PGM_GST_TYPE != PGM_TYPE_32BIT */
3175
3176	AssertReleaseMsgFailed(("Shw=%d Gst=%d is not implemented!\n", PGM_GST_TYPE, PGM_SHW_TYPE));
3177	return VERR_INTERNAL_ERROR;
3178	#endif /* PGM_GST_TYPE != PGM_TYPE_32BIT */
3179	}
3180
3181
3182	#if PGM_GST_TYPE == PGM_TYPE_32BIT \|\| PGM_GST_TYPE == PGM_TYPE_PAE \|\| PGM_GST_TYPE == PGM_TYPE_AMD64
3183	# if PGM_SHW_TYPE == PGM_TYPE_32BIT \|\| PGM_SHW_TYPE == PGM_TYPE_PAE \|\| PGM_SHW_TYPE == PGM_TYPE_AMD64
3184	/**
3185	* Figures out which kind of shadow page this guest PDE warrants.
3186	*
3187	* @returns Shadow page kind.
3188	* @param pPdeSrc The guest PDE in question.
3189	* @param cr4 The current guest cr4 value.
3190	*/
3191	DECLINLINE(PGMPOOLKIND) PGM_BTH_NAME(CalcPageKind)(const GSTPDE *pPdeSrc, uint32_t cr4)
3192	{
3193	# if PMG_GST_TYPE == PGM_TYPE_AMD64
3194	if (!pPdeSrc->n.u1Size)
3195	# else
3196	if (!pPdeSrc->n.u1Size \|\| !(cr4 & X86_CR4_PSE))
3197	# endif
3198	return BTH_PGMPOOLKIND_PT_FOR_PT;
3199	//switch (pPdeSrc->u & (X86_PDE4M_RW \| X86_PDE4M_US /\| X86_PDE4M_PAE_NX/))
3200	//{
3201	// case 0:
3202	// return BTH_PGMPOOLKIND_PT_FOR_BIG_RO;
3203	// case X86_PDE4M_RW:
3204	// return BTH_PGMPOOLKIND_PT_FOR_BIG_RW;
3205	// case X86_PDE4M_US:
3206	// return BTH_PGMPOOLKIND_PT_FOR_BIG_US;
3207	// case X86_PDE4M_RW \| X86_PDE4M_US:
3208	// return BTH_PGMPOOLKIND_PT_FOR_BIG_RW_US;
3209	# if 0
3210	// case X86_PDE4M_PAE_NX:
3211	// return BTH_PGMPOOLKIND_PT_FOR_BIG_NX;
3212	// case X86_PDE4M_RW \| X86_PDE4M_PAE_NX:
3213	// return BTH_PGMPOOLKIND_PT_FOR_BIG_RW_NX;
3214	// case X86_PDE4M_US \| X86_PDE4M_PAE_NX:
3215	// return BTH_PGMPOOLKIND_PT_FOR_BIG_US_NX;
3216	// case X86_PDE4M_RW \| X86_PDE4M_US \| X86_PDE4M_PAE_NX:
3217	// return BTH_PGMPOOLKIND_PT_FOR_BIG_RW_US_NX;
3218	# endif
3219	return BTH_PGMPOOLKIND_PT_FOR_BIG;
3220	//}
3221	}
3222	# endif
3223	#endif
3224
3225	#undef MY_STAM_COUNTER_INC
3226	#define MY_STAM_COUNTER_INC(a) do { } while (0)
3227
3228
3229	/**
3230	* Syncs the paging hierarchy starting at CR3.
3231	*
3232	* @returns VBox status code, no specials.
3233	* @param pVM The virtual machine.
3234	* @param cr0 Guest context CR0 register
3235	* @param cr3 Guest context CR3 register
3236	* @param cr4 Guest context CR4 register
3237	* @param fGlobal Including global page directories or not
3238	*/
3239	PGM_BTH_DECL(int, SyncCR3)(PVM pVM, uint64_t cr0, uint64_t cr3, uint64_t cr4, bool fGlobal)
3240	{
3241	if (VM_FF_ISSET(pVM, VM_FF_PGM_SYNC_CR3))
3242	fGlobal = true; /* Change this CR3 reload to be a global one. */
3243
3244	LogFlow(("SyncCR3 %d\n", fGlobal));
3245
3246	#if PGM_SHW_TYPE != PGM_TYPE_NESTED && PGM_SHW_TYPE != PGM_TYPE_EPT
3247	/*
3248	* Update page access handlers.
3249	* The virtual are always flushed, while the physical are only on demand.
3250	* WARNING: We are incorrectly not doing global flushing on Virtual Handler updates. We'll
3251	* have to look into that later because it will have a bad influence on the performance.
3252	* @note SvL: There's no need for that. Just invalidate the virtual range(s).
3253	* bird: Yes, but that won't work for aliases.
3254	*/
3255	/** @todo this MUST go away. See #1557. */
3256	STAM_PROFILE_START(&pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3Handlers), h);
3257	PGM_GST_NAME(HandlerVirtualUpdate)(pVM, cr4);
3258	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3Handlers), h);
3259	#endif
3260
3261	#if PGM_SHW_TYPE == PGM_TYPE_NESTED \|\| PGM_SHW_TYPE == PGM_TYPE_EPT
3262	/*
3263	* Nested / EPT - almost no work.
3264	*/
3265	/** @todo check if this is really necessary; the call does it as well... */
3266	HWACCMFlushTLB(pVM);
3267	return VINF_SUCCESS;
3268
3269	#elif PGM_SHW_TYPE == PGM_TYPE_AMD64
3270	/*
3271	* AMD64 (Shw & Gst) - No need to check all paging levels; we zero
3272	* out the shadow parts when the guest modifies its tables.
3273	*/
3274	return VINF_SUCCESS;
3275
3276	#else /* PGM_SHW_TYPE != PGM_TYPE_NESTED && PGM_SHW_TYPE != PGM_TYPE_EPT && PGM_SHW_TYPE != PGM_TYPE_AMD64 */
3277
3278	# ifdef PGM_WITHOUT_MAPPINGS
3279	Assert(pVM->pgm.s.fMappingsFixed);
3280	return VINF_SUCCESS;
3281	# else
3282	/* Nothing to do when mappings are fixed. */
3283	if (pVM->pgm.s.fMappingsFixed)
3284	return VINF_SUCCESS;
3285
3286	int rc = PGMMapResolveConflicts(pVM);
3287	Assert(rc == VINF_SUCCESS \|\| rc == VINF_PGM_SYNC_CR3);
3288	if (rc == VINF_PGM_SYNC_CR3)
3289	{
3290	LogFlow(("SyncCR3: detected conflict -> VINF_PGM_SYNC_CR3\n"));
3291	return VINF_PGM_SYNC_CR3;
3292	}
3293	# endif
3294	return VINF_SUCCESS;
3295	#endif /* PGM_SHW_TYPE != PGM_TYPE_NESTED && PGM_SHW_TYPE != PGM_TYPE_EPT && PGM_SHW_TYPE != PGM_TYPE_AMD64 */
3296	}
3297
3298
3299
3300
3301	#ifdef VBOX_STRICT
3302	#ifdef IN_RC
3303	# undef AssertMsgFailed
3304	# define AssertMsgFailed Log
3305	#endif
3306	#ifdef IN_RING3
3307	# include <VBox/dbgf.h>
3308
3309	/**
3310	* Dumps a page table hierarchy use only physical addresses and cr4/lm flags.
3311	*
3312	* @returns VBox status code (VINF_SUCCESS).
3313	* @param pVM The VM handle.
3314	* @param cr3 The root of the hierarchy.
3315	* @param crr The cr4, only PAE and PSE is currently used.
3316	* @param fLongMode Set if long mode, false if not long mode.
3317	* @param cMaxDepth Number of levels to dump.
3318	* @param pHlp Pointer to the output functions.
3319	*/
3320	__BEGIN_DECLS
3321	VMMR3DECL(int) PGMR3DumpHierarchyHC(PVM pVM, uint32_t cr3, uint32_t cr4, bool fLongMode, unsigned cMaxDepth, PCDBGFINFOHLP pHlp);
3322	__END_DECLS
3323
3324	#endif
3325
3326	/**
3327	* Checks that the shadow page table is in sync with the guest one.
3328	*
3329	* @returns The number of errors.
3330	* @param pVM The virtual machine.
3331	* @param cr3 Guest context CR3 register
3332	* @param cr4 Guest context CR4 register
3333	* @param GCPtr Where to start. Defaults to 0.
3334	* @param cb How much to check. Defaults to everything.
3335	*/
3336	PGM_BTH_DECL(unsigned, AssertCR3)(PVM pVM, uint64_t cr3, uint64_t cr4, RTGCPTR GCPtr, RTGCPTR cb)
3337	{
3338	#if PGM_SHW_TYPE == PGM_TYPE_NESTED \|\| PGM_SHW_TYPE == PGM_TYPE_EPT
3339	return 0;
3340	#else
3341	unsigned cErrors = 0;
3342
3343	#if PGM_GST_TYPE == PGM_TYPE_PAE
3344	/** @todo currently broken; crashes below somewhere */
3345	AssertFailed();
3346	#endif
3347
3348	#if PGM_GST_TYPE == PGM_TYPE_32BIT \
3349	\|\| PGM_GST_TYPE == PGM_TYPE_PAE \
3350	\|\| PGM_GST_TYPE == PGM_TYPE_AMD64
3351
3352	# if PGM_GST_TYPE == PGM_TYPE_AMD64
3353	bool fBigPagesSupported = true;
3354	# else
3355	bool fBigPagesSupported = !!(CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
3356	# endif
3357	PPGM pPGM = &pVM->pgm.s;
3358	RTGCPHYS GCPhysGst; /* page address derived from the guest page tables. */
3359	RTHCPHYS HCPhysShw; /* page address derived from the shadow page tables. */
3360	# ifndef IN_RING0
3361	RTHCPHYS HCPhys; /* general usage. */
3362	# endif
3363	int rc;
3364
3365	/*
3366	* Check that the Guest CR3 and all its mappings are correct.
3367	*/
3368	AssertMsgReturn(pPGM->GCPhysCR3 == (cr3 & GST_CR3_PAGE_MASK),
3369	("Invalid GCPhysCR3=%RGp cr3=%RGp\n", pPGM->GCPhysCR3, (RTGCPHYS)cr3),
3370	false);
3371	# if !defined(IN_RING0) && PGM_GST_TYPE != PGM_TYPE_AMD64
3372	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3373	rc = PGMShwGetPage(pVM, (RTGCPTR)pPGM->pGst32BitPdRC, NULL, &HCPhysShw);
3374	# else
3375	rc = PGMShwGetPage(pVM, (RTGCPTR)pPGM->pGstPaePdptRC, NULL, &HCPhysShw);
3376	# endif
3377	AssertRCReturn(rc, 1);
3378	HCPhys = NIL_RTHCPHYS;
3379	rc = pgmRamGCPhys2HCPhys(pPGM, cr3 & GST_CR3_PAGE_MASK, &HCPhys);
3380	AssertMsgReturn(HCPhys == HCPhysShw, ("HCPhys=%RHp HCPhyswShw=%RHp (cr3)\n", HCPhys, HCPhysShw), false);
3381	# if PGM_GST_TYPE == PGM_TYPE_32BIT && defined(IN_RING3)
3382	RTGCPHYS GCPhys;
3383	rc = PGMR3DbgR3Ptr2GCPhys(pVM, pPGM->pGst32BitPdR3, &GCPhys);
3384	AssertRCReturn(rc, 1);
3385	AssertMsgReturn((cr3 & GST_CR3_PAGE_MASK) == GCPhys, ("GCPhys=%RGp cr3=%RGp\n", GCPhys, (RTGCPHYS)cr3), false);
3386	# endif
3387	# endif /* !IN_RING0 */
3388
3389	/*
3390	* Get and check the Shadow CR3.
3391	*/
3392	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
3393	unsigned cPDEs = X86_PG_ENTRIES;
3394	unsigned cIncrement = X86_PG_ENTRIES * PAGE_SIZE;
3395	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
3396	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3397	unsigned cPDEs = X86_PG_PAE_ENTRIES * 4; /* treat it as a 2048 entry table. */
3398	# else
3399	unsigned cPDEs = X86_PG_PAE_ENTRIES;
3400	# endif
3401	unsigned cIncrement = X86_PG_PAE_ENTRIES * PAGE_SIZE;
3402	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
3403	unsigned cPDEs = X86_PG_PAE_ENTRIES;
3404	unsigned cIncrement = X86_PG_PAE_ENTRIES * PAGE_SIZE;
3405	# endif
3406	if (cb != ~(RTGCPTR)0)
3407	cPDEs = RT_MIN(cb >> SHW_PD_SHIFT, 1);
3408
3409	/** @todo call the other two PGMAssert() functions. /
3410
3411	# if PGM_GST_TYPE == PGM_TYPE_AMD64 \|\| PGM_GST_TYPE == PGM_TYPE_PAE
3412	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
3413	# endif
3414
3415	# if PGM_GST_TYPE == PGM_TYPE_AMD64
3416	unsigned iPml4 = (GCPtr >> X86_PML4_SHIFT) & X86_PML4_MASK;
3417
3418	for (; iPml4 < X86_PG_PAE_ENTRIES; iPml4++)
3419	{
3420	PPGMPOOLPAGE pShwPdpt = NULL;
3421	PX86PML4E pPml4eSrc;
3422	PX86PML4E pPml4eDst;
3423	RTGCPHYS GCPhysPdptSrc;
3424
3425	pPml4eSrc = pgmGstGetLongModePML4EPtr(&pVM->pgm.s, iPml4);
3426	pPml4eDst = pgmShwGetLongModePML4EPtr(&pVM->pgm.s, iPml4);
3427
3428	/* Fetch the pgm pool shadow descriptor if the shadow pml4e is present. */
3429	if (!pPml4eDst->n.u1Present)
3430	{
3431	GCPtr += _2M * UINT64_C(512) * UINT64_C(512);
3432	continue;
3433	}
3434
3435	pShwPdpt = pgmPoolGetPage(pPool, pPml4eDst->u & X86_PML4E_PG_MASK);
3436	GCPhysPdptSrc = pPml4eSrc->u & X86_PML4E_PG_MASK_FULL;
3437
3438	if (pPml4eSrc->n.u1Present != pPml4eDst->n.u1Present)
3439	{
3440	AssertMsgFailed(("Present bit doesn't match! pPml4eDst.u=%#RX64 pPml4eSrc.u=%RX64\n", pPml4eDst->u, pPml4eSrc->u));
3441	GCPtr += _2M * UINT64_C(512) * UINT64_C(512);
3442	cErrors++;
3443	continue;
3444	}
3445
3446	if (GCPhysPdptSrc != pShwPdpt->GCPhys)
3447	{
3448	AssertMsgFailed(("Physical address doesn't match! iPml4 %d pPml4eDst.u=%#RX64 pPml4eSrc.u=%RX64 Phys %RX64 vs %RX64\n", iPml4, pPml4eDst->u, pPml4eSrc->u, pShwPdpt->GCPhys, GCPhysPdptSrc));
3449	GCPtr += _2M * UINT64_C(512) * UINT64_C(512);
3450	cErrors++;
3451	continue;
3452	}
3453
3454	if ( pPml4eDst->n.u1User != pPml4eSrc->n.u1User
3455	\|\| pPml4eDst->n.u1Write != pPml4eSrc->n.u1Write
3456	\|\| pPml4eDst->n.u1NoExecute != pPml4eSrc->n.u1NoExecute)
3457	{
3458	AssertMsgFailed(("User/Write/NoExec bits don't match! pPml4eDst.u=%#RX64 pPml4eSrc.u=%RX64\n", pPml4eDst->u, pPml4eSrc->u));
3459	GCPtr += _2M * UINT64_C(512) * UINT64_C(512);
3460	cErrors++;
3461	continue;
3462	}
3463	# else /* PGM_GST_TYPE != PGM_TYPE_AMD64 */
3464	{
3465	# endif /* PGM_GST_TYPE != PGM_TYPE_AMD64 */
3466
3467	# if PGM_GST_TYPE == PGM_TYPE_AMD64 \|\| PGM_GST_TYPE == PGM_TYPE_PAE
3468	/*
3469	* Check the PDPTEs too.
3470	*/
3471	unsigned iPdpt = (GCPtr >> SHW_PDPT_SHIFT) & SHW_PDPT_MASK;
3472
3473	for (;iPdpt <= SHW_PDPT_MASK; iPdpt++)
3474	{
3475	unsigned iPDSrc;
3476	PPGMPOOLPAGE pShwPde = NULL;
3477	PX86PDPE pPdpeDst;
3478	RTGCPHYS GCPhysPdeSrc;
3479	# if PGM_GST_TYPE == PGM_TYPE_PAE
3480	X86PDPE PdpeSrc;
3481	PGSTPD pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, GCPtr, &iPDSrc, &PdpeSrc);
3482	PX86PDPT pPdptDst = pgmShwGetPaePDPTPtr(&pVM->pgm.s);
3483	# else
3484	PX86PML4E pPml4eSrc;
3485	X86PDPE PdpeSrc;
3486	PX86PDPT pPdptDst;
3487	PX86PDPAE pPDDst;
3488	PGSTPD pPDSrc = pgmGstGetLongModePDPtr(&pVM->pgm.s, GCPtr, &pPml4eSrc, &PdpeSrc, &iPDSrc);
3489
3490	rc = pgmShwGetLongModePDPtr(pVM, GCPtr, NULL, &pPdptDst, &pPDDst);
3491	if (rc != VINF_SUCCESS)
3492	{
3493	AssertMsg(rc == VERR_PAGE_DIRECTORY_PTR_NOT_PRESENT, ("Unexpected rc=%Rrc\n", rc));
3494	GCPtr += 512 * _2M;
3495	continue; /* next PDPTE */
3496	}
3497	Assert(pPDDst);
3498	# endif
3499	Assert(iPDSrc == 0);
3500
3501	pPdpeDst = &pPdptDst->a[iPdpt];
3502
3503	if (!pPdpeDst->n.u1Present)
3504	{
3505	GCPtr += 512 * _2M;
3506	continue; /* next PDPTE */
3507	}
3508
3509	pShwPde = pgmPoolGetPage(pPool, pPdpeDst->u & X86_PDPE_PG_MASK);
3510	GCPhysPdeSrc = PdpeSrc.u & X86_PDPE_PG_MASK;
3511
3512	if (pPdpeDst->n.u1Present != PdpeSrc.n.u1Present)
3513	{
3514	AssertMsgFailed(("Present bit doesn't match! pPdpeDst.u=%#RX64 pPdpeSrc.u=%RX64\n", pPdpeDst->u, PdpeSrc.u));
3515	GCPtr += 512 * _2M;
3516	cErrors++;
3517	continue;
3518	}
3519
3520	if (GCPhysPdeSrc != pShwPde->GCPhys)
3521	{
3522	# if PGM_GST_TYPE == PGM_TYPE_AMD64
3523	AssertMsgFailed(("Physical address doesn't match! iPml4 %d iPdpt %d pPdpeDst.u=%#RX64 pPdpeSrc.u=%RX64 Phys %RX64 vs %RX64\n", iPml4, iPdpt, pPdpeDst->u, PdpeSrc.u, pShwPde->GCPhys, GCPhysPdeSrc));
3524	# else
3525	AssertMsgFailed(("Physical address doesn't match! iPdpt %d pPdpeDst.u=%#RX64 pPdpeSrc.u=%RX64 Phys %RX64 vs %RX64\n", iPdpt, pPdpeDst->u, PdpeSrc.u, pShwPde->GCPhys, GCPhysPdeSrc));
3526	# endif
3527	GCPtr += 512 * _2M;
3528	cErrors++;
3529	continue;
3530	}
3531
3532	# if PGM_GST_TYPE == PGM_TYPE_AMD64
3533	if ( pPdpeDst->lm.u1User != PdpeSrc.lm.u1User
3534	\|\| pPdpeDst->lm.u1Write != PdpeSrc.lm.u1Write
3535	\|\| pPdpeDst->lm.u1NoExecute != PdpeSrc.lm.u1NoExecute)
3536	{
3537	AssertMsgFailed(("User/Write/NoExec bits don't match! pPdpeDst.u=%#RX64 pPdpeSrc.u=%RX64\n", pPdpeDst->u, PdpeSrc.u));
3538	GCPtr += 512 * _2M;
3539	cErrors++;
3540	continue;
3541	}
3542	# endif
3543
3544	# else /* PGM_GST_TYPE != PGM_TYPE_AMD64 && PGM_GST_TYPE != PGM_TYPE_PAE */
3545	{
3546	# endif /* PGM_GST_TYPE != PGM_TYPE_AMD64 && PGM_GST_TYPE != PGM_TYPE_PAE */
3547	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3548	GSTPD const *pPDSrc = pgmGstGet32bitPDPtr(&pVM->pgm.s);
3549	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
3550	PCX86PD pPDDst = pgmShwGet32BitPDPtr(&pVM->pgm.s);
3551	# endif
3552	# endif /* PGM_GST_TYPE == PGM_TYPE_32BIT */
3553	/*
3554	* Iterate the shadow page directory.
3555	*/
3556	GCPtr = (GCPtr >> SHW_PD_SHIFT) << SHW_PD_SHIFT;
3557	unsigned iPDDst = (GCPtr >> SHW_PD_SHIFT) & SHW_PD_MASK;
3558
3559	for (;
3560	iPDDst < cPDEs;
3561	iPDDst++, GCPtr += cIncrement)
3562	{
3563	# if PGM_SHW_TYPE == PGM_TYPE_PAE
3564	const SHWPDE PdeDst = *pgmShwGetPaePDEPtr(pPGM, GCPtr);
3565	# else
3566	const SHWPDE PdeDst = pPDDst->a[iPDDst];
3567	# endif
3568	if (PdeDst.u & PGM_PDFLAGS_MAPPING)
3569	{
3570	Assert(pgmMapAreMappingsEnabled(&pVM->pgm.s));
3571	if ((PdeDst.u & X86_PDE_AVL_MASK) != PGM_PDFLAGS_MAPPING)
3572	{
3573	AssertMsgFailed(("Mapping shall only have PGM_PDFLAGS_MAPPING set! PdeDst.u=%#RX64\n", (uint64_t)PdeDst.u));
3574	cErrors++;
3575	continue;
3576	}
3577	}
3578	else if ( (PdeDst.u & X86_PDE_P)
3579	\|\| ((PdeDst.u & (X86_PDE_P \| PGM_PDFLAGS_TRACK_DIRTY)) == (X86_PDE_P \| PGM_PDFLAGS_TRACK_DIRTY))
3580	)
3581	{
3582	HCPhysShw = PdeDst.u & SHW_PDE_PG_MASK;
3583	PPGMPOOLPAGE pPoolPage = pgmPoolGetPageByHCPhys(pVM, HCPhysShw);
3584	if (!pPoolPage)
3585	{
3586	AssertMsgFailed(("Invalid page table address %RHp at %RGv! PdeDst=%#RX64\n",
3587	HCPhysShw, GCPtr, (uint64_t)PdeDst.u));
3588	cErrors++;
3589	continue;
3590	}
3591	const SHWPT pPTDst = (const SHWPT )PGMPOOL_PAGE_2_PTR(pVM, pPoolPage);
3592
3593	if (PdeDst.u & (X86_PDE4M_PWT \| X86_PDE4M_PCD))
3594	{
3595	AssertMsgFailed(("PDE flags PWT and/or PCD is set at %RGv! These flags are not virtualized! PdeDst=%#RX64\n",
3596	GCPtr, (uint64_t)PdeDst.u));
3597	cErrors++;
3598	}
3599
3600	if (PdeDst.u & (X86_PDE4M_G \| X86_PDE4M_D))
3601	{
3602	AssertMsgFailed(("4K PDE reserved flags at %RGv! PdeDst=%#RX64\n",
3603	GCPtr, (uint64_t)PdeDst.u));
3604	cErrors++;
3605	}
3606
3607	const GSTPDE PdeSrc = pPDSrc->a[(iPDDst >> (GST_PD_SHIFT - SHW_PD_SHIFT)) & GST_PD_MASK];
3608	if (!PdeSrc.n.u1Present)
3609	{
3610	AssertMsgFailed(("Guest PDE at %RGv is not present! PdeDst=%#RX64 PdeSrc=%#RX64\n",
3611	GCPtr, (uint64_t)PdeDst.u, (uint64_t)PdeSrc.u));
3612	cErrors++;
3613	continue;
3614	}
3615
3616	if ( !PdeSrc.b.u1Size
3617	\|\| !fBigPagesSupported)
3618	{
3619	GCPhysGst = PdeSrc.u & GST_PDE_PG_MASK;
3620	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
3621	GCPhysGst \|= (iPDDst & 1) * (PAGE_SIZE / 2);
3622	# endif
3623	}
3624	else
3625	{
3626	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3627	if (PdeSrc.u & X86_PDE4M_PG_HIGH_MASK)
3628	{
3629	AssertMsgFailed(("Guest PDE at %RGv is using PSE36 or similar! PdeSrc=%#RX64\n",
3630	GCPtr, (uint64_t)PdeSrc.u));
3631	cErrors++;
3632	continue;
3633	}
3634	# endif
3635	GCPhysGst = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc);
3636	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
3637	GCPhysGst \|= GCPtr & RT_BIT(X86_PAGE_2M_SHIFT);
3638	# endif
3639	}
3640
3641	if ( pPoolPage->enmKind
3642	!= (!PdeSrc.b.u1Size \|\| !fBigPagesSupported ? BTH_PGMPOOLKIND_PT_FOR_PT : BTH_PGMPOOLKIND_PT_FOR_BIG))
3643	{
3644	AssertMsgFailed(("Invalid shadow page table kind %d at %RGv! PdeSrc=%#RX64\n",
3645	pPoolPage->enmKind, GCPtr, (uint64_t)PdeSrc.u));
3646	cErrors++;
3647	}
3648
3649	PPGMPAGE pPhysPage = pgmPhysGetPage(pPGM, GCPhysGst);
3650	if (!pPhysPage)
3651	{
3652	AssertMsgFailed(("Cannot find guest physical address %RGp in the PDE at %RGv! PdeSrc=%#RX64\n",
3653	GCPhysGst, GCPtr, (uint64_t)PdeSrc.u));
3654	cErrors++;
3655	continue;
3656	}
3657
3658	if (GCPhysGst != pPoolPage->GCPhys)
3659	{
3660	AssertMsgFailed(("GCPhysGst=%RGp != pPage->GCPhys=%RGp at %RGv\n",
3661	GCPhysGst, pPoolPage->GCPhys, GCPtr));
3662	cErrors++;
3663	continue;
3664	}
3665
3666	if ( !PdeSrc.b.u1Size
3667	\|\| !fBigPagesSupported)
3668	{
3669	/*
3670	* Page Table.
3671	*/
3672	const GSTPT *pPTSrc;
3673	rc = PGM_GCPHYS_2_PTR(pVM, GCPhysGst & ~(RTGCPHYS)(PAGE_SIZE - 1), &pPTSrc);
3674	if (RT_FAILURE(rc))
3675	{
3676	AssertMsgFailed(("Cannot map/convert guest physical address %RGp in the PDE at %RGv! PdeSrc=%#RX64\n",
3677	GCPhysGst, GCPtr, (uint64_t)PdeSrc.u));
3678	cErrors++;
3679	continue;
3680	}
3681	if ( (PdeSrc.u & (X86_PDE_P \| X86_PDE_US \| X86_PDE_RW/* \| X86_PDE_A*/))
3682	!= (PdeDst.u & (X86_PDE_P \| X86_PDE_US \| X86_PDE_RW/* \| X86_PDE_A*/)))
3683	{
3684	/// @todo We get here a lot on out-of-sync CR3 entries. The access handler should zap them to avoid false alarms here!
3685	// (This problem will go away when/if we shadow multiple CR3s.)
3686	AssertMsgFailed(("4K PDE flags mismatch at %RGv! PdeSrc=%#RX64 PdeDst=%#RX64\n",
3687	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
3688	cErrors++;
3689	continue;
3690	}
3691	if (PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY)
3692	{
3693	AssertMsgFailed(("4K PDEs cannot have PGM_PDFLAGS_TRACK_DIRTY set! GCPtr=%RGv PdeDst=%#RX64\n",
3694	GCPtr, (uint64_t)PdeDst.u));
3695	cErrors++;
3696	continue;
3697	}
3698
3699	/* iterate the page table. */
3700	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
3701	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
3702	const unsigned offPTSrc = ((GCPtr >> SHW_PD_SHIFT) & 1) * 512;
3703	# else
3704	const unsigned offPTSrc = 0;
3705	# endif
3706	for (unsigned iPT = 0, off = 0;
3707	iPT < RT_ELEMENTS(pPTDst->a);
3708	iPT++, off += PAGE_SIZE)
3709	{
3710	const SHWPTE PteDst = pPTDst->a[iPT];
3711
3712	/* skip not-present entries. */
3713	if (!(PteDst.u & (X86_PTE_P \| PGM_PTFLAGS_TRACK_DIRTY))) /** @todo deal with ALL handlers and CSAM !P pages! */
3714	continue;
3715	Assert(PteDst.n.u1Present);
3716
3717	const GSTPTE PteSrc = pPTSrc->a[iPT + offPTSrc];
3718	if (!PteSrc.n.u1Present)
3719	{
3720	# ifdef IN_RING3
3721	PGMAssertHandlerAndFlagsInSync(pVM);
3722	PGMR3DumpHierarchyGC(pVM, cr3, cr4, (PdeSrc.u & GST_PDE_PG_MASK));
3723	# endif
3724	AssertMsgFailed(("Out of sync (!P) PTE at %RGv! PteSrc=%#RX64 PteDst=%#RX64 pPTSrc=%RGv iPTSrc=%x PdeSrc=%x physpte=%RGp\n",
3725	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u, pPTSrc, iPT + offPTSrc, PdeSrc.au32[0],
3726	(PdeSrc.u & GST_PDE_PG_MASK) + (iPT + offPTSrc)*sizeof(PteSrc)));
3727	cErrors++;
3728	continue;
3729	}
3730
3731	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;
3732	# if 1 /** @todo sync accessed bit properly... */
3733	fIgnoreFlags \|= X86_PTE_A;
3734	# endif
3735
3736	/* match the physical addresses */
3737	HCPhysShw = PteDst.u & SHW_PTE_PG_MASK;
3738	GCPhysGst = PteSrc.u & GST_PTE_PG_MASK;
3739
3740	# ifdef IN_RING3
3741	rc = PGMPhysGCPhys2HCPhys(pVM, GCPhysGst, &HCPhys);
3742	if (RT_FAILURE(rc))
3743	{
3744	if (HCPhysShw != MMR3PageDummyHCPhys(pVM))
3745	{
3746	AssertMsgFailed(("Cannot find guest physical address %RGp at %RGv! PteSrc=%#RX64 PteDst=%#RX64\n",
3747	GCPhysGst, GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3748	cErrors++;
3749	continue;
3750	}
3751	}
3752	else if (HCPhysShw != (HCPhys & SHW_PTE_PG_MASK))
3753	{
3754	AssertMsgFailed(("Out of sync (phys) at %RGv! HCPhysShw=%RHp HCPhys=%RHp GCPhysGst=%RGp PteSrc=%#RX64 PteDst=%#RX64\n",
3755	GCPtr + off, HCPhysShw, HCPhys, GCPhysGst, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3756	cErrors++;
3757	continue;
3758	}
3759	# endif
3760
3761	pPhysPage = pgmPhysGetPage(pPGM, GCPhysGst);
3762	if (!pPhysPage)
3763	{
3764	# ifdef IN_RING3 /** @todo make MMR3PageDummyHCPhys an 'All' function! */
3765	if (HCPhysShw != MMR3PageDummyHCPhys(pVM))
3766	{
3767	AssertMsgFailed(("Cannot find guest physical address %RGp at %RGv! PteSrc=%#RX64 PteDst=%#RX64\n",
3768	GCPhysGst, GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3769	cErrors++;
3770	continue;
3771	}
3772	# endif
3773	if (PteDst.n.u1Write)
3774	{
3775	AssertMsgFailed(("Invalid guest page at %RGv is writable! GCPhysGst=%RGp PteSrc=%#RX64 PteDst=%#RX64\n",
3776	GCPtr + off, GCPhysGst, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3777	cErrors++;
3778	}
3779	fIgnoreFlags \|= X86_PTE_RW;
3780	}
3781	else if (HCPhysShw != PGM_PAGE_GET_HCPHYS(pPhysPage))
3782	{
3783	AssertMsgFailed(("Out of sync (phys) at %RGv! HCPhysShw=%RHp pPhysPage:%R[pgmpage] GCPhysGst=%RGp PteSrc=%#RX64 PteDst=%#RX64\n",
3784	GCPtr + off, HCPhysShw, pPhysPage, GCPhysGst, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3785	cErrors++;
3786	continue;
3787	}
3788
3789	/* flags */
3790	if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPhysPage))
3791	{
3792	if (!PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPhysPage))
3793	{
3794	if (PteDst.n.u1Write)
3795	{
3796	AssertMsgFailed(("WRITE access flagged at %RGv but the page is writable! pPhysPage=%R[pgmpage] PteSrc=%#RX64 PteDst=%#RX64\n",
3797	GCPtr + off, pPhysPage, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3798	cErrors++;
3799	continue;
3800	}
3801	fIgnoreFlags \|= X86_PTE_RW;
3802	}
3803	else
3804	{
3805	if (PteDst.n.u1Present)
3806	{
3807	AssertMsgFailed(("ALL access flagged at %RGv but the page is present! pPhysPage=%R[pgmpage] PteSrc=%#RX64 PteDst=%#RX64\n",
3808	GCPtr + off, pPhysPage, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3809	cErrors++;
3810	continue;
3811	}
3812	fIgnoreFlags \|= X86_PTE_P;
3813	}
3814	}
3815	else
3816	{
3817	if (!PteSrc.n.u1Dirty && PteSrc.n.u1Write)
3818	{
3819	if (PteDst.n.u1Write)
3820	{
3821	AssertMsgFailed(("!DIRTY page at %RGv is writable! PteSrc=%#RX64 PteDst=%#RX64\n",
3822	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3823	cErrors++;
3824	continue;
3825	}
3826	if (!(PteDst.u & PGM_PTFLAGS_TRACK_DIRTY))
3827	{
3828	AssertMsgFailed(("!DIRTY page at %RGv is not marked TRACK_DIRTY! PteSrc=%#RX64 PteDst=%#RX64\n",
3829	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3830	cErrors++;
3831	continue;
3832	}
3833	if (PteDst.n.u1Dirty)
3834	{
3835	AssertMsgFailed(("!DIRTY page at %RGv is marked DIRTY! PteSrc=%#RX64 PteDst=%#RX64\n",
3836	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3837	cErrors++;
3838	}
3839	# if 0 /** @todo sync access bit properly... */
3840	if (PteDst.n.u1Accessed != PteSrc.n.u1Accessed)
3841	{
3842	AssertMsgFailed(("!DIRTY page at %RGv is has mismatching accessed bit! PteSrc=%#RX64 PteDst=%#RX64\n",
3843	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3844	cErrors++;
3845	}
3846	fIgnoreFlags \|= X86_PTE_RW;
3847	# else
3848	fIgnoreFlags \|= X86_PTE_RW \| X86_PTE_A;
3849	# endif
3850	}
3851	else if (PteDst.u & PGM_PTFLAGS_TRACK_DIRTY)
3852	{
3853	/* access bit emulation (not implemented). */
3854	if (PteSrc.n.u1Accessed \|\| PteDst.n.u1Present)
3855	{
3856	AssertMsgFailed(("PGM_PTFLAGS_TRACK_DIRTY set at %RGv but no accessed bit emulation! PteSrc=%#RX64 PteDst=%#RX64\n",
3857	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3858	cErrors++;
3859	continue;
3860	}
3861	if (!PteDst.n.u1Accessed)
3862	{
3863	AssertMsgFailed(("!ACCESSED page at %RGv is has the accessed bit set! PteSrc=%#RX64 PteDst=%#RX64\n",
3864	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3865	cErrors++;
3866	}
3867	fIgnoreFlags \|= X86_PTE_P;
3868	}
3869	# ifdef DEBUG_sandervl
3870	fIgnoreFlags \|= X86_PTE_D \| X86_PTE_A;
3871	# endif
3872	}
3873
3874	if ( (PteSrc.u & ~fIgnoreFlags) != (PteDst.u & ~fIgnoreFlags)
3875	&& (PteSrc.u & ~(fIgnoreFlags \| X86_PTE_RW)) != (PteDst.u & ~fIgnoreFlags)
3876	)
3877	{
3878	AssertMsgFailed(("Flags mismatch at %RGv! %#RX64 != %#RX64 fIgnoreFlags=%#RX64 PteSrc=%#RX64 PteDst=%#RX64\n",
3879	GCPtr + off, (uint64_t)PteSrc.u & ~fIgnoreFlags, (uint64_t)PteDst.u & ~fIgnoreFlags,
3880	fIgnoreFlags, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3881	cErrors++;
3882	continue;
3883	}
3884	} /* foreach PTE */
3885	}
3886	else
3887	{
3888	/*
3889	* Big Page.
3890	*/
3891	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;
3892	if (!PdeSrc.b.u1Dirty && PdeSrc.b.u1Write)
3893	{
3894	if (PdeDst.n.u1Write)
3895	{
3896	AssertMsgFailed(("!DIRTY page at %RGv is writable! PdeSrc=%#RX64 PdeDst=%#RX64\n",
3897	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
3898	cErrors++;
3899	continue;
3900	}
3901	if (!(PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY))
3902	{
3903	AssertMsgFailed(("!DIRTY page at %RGv is not marked TRACK_DIRTY! PteSrc=%#RX64 PteDst=%#RX64\n",
3904	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
3905	cErrors++;
3906	continue;
3907	}
3908	# if 0 /** @todo sync access bit properly... */
3909	if (PdeDst.n.u1Accessed != PdeSrc.b.u1Accessed)
3910	{
3911	AssertMsgFailed(("!DIRTY page at %RGv is has mismatching accessed bit! PteSrc=%#RX64 PteDst=%#RX64\n",
3912	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
3913	cErrors++;
3914	}
3915	fIgnoreFlags \|= X86_PTE_RW;
3916	# else
3917	fIgnoreFlags \|= X86_PTE_RW \| X86_PTE_A;
3918	# endif
3919	}
3920	else if (PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY)
3921	{
3922	/* access bit emulation (not implemented). */
3923	if (PdeSrc.b.u1Accessed \|\| PdeDst.n.u1Present)
3924	{
3925	AssertMsgFailed(("PGM_PDFLAGS_TRACK_DIRTY set at %RGv but no accessed bit emulation! PdeSrc=%#RX64 PdeDst=%#RX64\n",
3926	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
3927	cErrors++;
3928	continue;
3929	}
3930	if (!PdeDst.n.u1Accessed)
3931	{
3932	AssertMsgFailed(("!ACCESSED page at %RGv is has the accessed bit set! PdeSrc=%#RX64 PdeDst=%#RX64\n",
3933	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
3934	cErrors++;
3935	}
3936	fIgnoreFlags \|= X86_PTE_P;
3937	}
3938
3939	if ((PdeSrc.u & ~fIgnoreFlags) != (PdeDst.u & ~fIgnoreFlags))
3940	{
3941	AssertMsgFailed(("Flags mismatch (B) at %RGv! %#RX64 != %#RX64 fIgnoreFlags=%#RX64 PdeSrc=%#RX64 PdeDst=%#RX64\n",
3942	GCPtr, (uint64_t)PdeSrc.u & ~fIgnoreFlags, (uint64_t)PdeDst.u & ~fIgnoreFlags,
3943	fIgnoreFlags, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
3944	cErrors++;
3945	}
3946
3947	/* iterate the page table. */
3948	for (unsigned iPT = 0, off = 0;
3949	iPT < RT_ELEMENTS(pPTDst->a);
3950	iPT++, off += PAGE_SIZE, GCPhysGst += PAGE_SIZE)
3951	{
3952	const SHWPTE PteDst = pPTDst->a[iPT];
3953
3954	if (PteDst.u & PGM_PTFLAGS_TRACK_DIRTY)
3955	{
3956	AssertMsgFailed(("The PTE at %RGv emulating a 2/4M page is marked TRACK_DIRTY! PdeSrc=%#RX64 PteDst=%#RX64\n",
3957	GCPtr + off, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
3958	cErrors++;
3959	}
3960
3961	/* skip not-present entries. */
3962	if (!PteDst.n.u1Present) /** @todo deal with ALL handlers and CSAM !P pages! */
3963	continue;
3964
3965	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;
3966
3967	/* match the physical addresses */
3968	HCPhysShw = PteDst.u & X86_PTE_PAE_PG_MASK;
3969
3970	# ifdef IN_RING3
3971	rc = PGMPhysGCPhys2HCPhys(pVM, GCPhysGst, &HCPhys);
3972	if (RT_FAILURE(rc))
3973	{
3974	if (HCPhysShw != MMR3PageDummyHCPhys(pVM))
3975	{
3976	AssertMsgFailed(("Cannot find guest physical address %RGp at %RGv! PdeSrc=%#RX64 PteDst=%#RX64\n",
3977	GCPhysGst, GCPtr + off, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
3978	cErrors++;
3979	}
3980	}
3981	else if (HCPhysShw != (HCPhys & X86_PTE_PAE_PG_MASK))
3982	{
3983	AssertMsgFailed(("Out of sync (phys) at %RGv! HCPhysShw=%RHp HCPhys=%RHp GCPhysGst=%RGp PdeSrc=%#RX64 PteDst=%#RX64\n",
3984	GCPtr + off, HCPhysShw, HCPhys, GCPhysGst, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
3985	cErrors++;
3986	continue;
3987	}
3988	# endif
3989	pPhysPage = pgmPhysGetPage(pPGM, GCPhysGst);
3990	if (!pPhysPage)
3991	{
3992	# ifdef IN_RING3 /** @todo make MMR3PageDummyHCPhys an 'All' function! */
3993	if (HCPhysShw != MMR3PageDummyHCPhys(pVM))
3994	{
3995	AssertMsgFailed(("Cannot find guest physical address %RGp at %RGv! PdeSrc=%#RX64 PteDst=%#RX64\n",
3996	GCPhysGst, GCPtr + off, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
3997	cErrors++;
3998	continue;
3999	}
4000	# endif
4001	if (PteDst.n.u1Write)
4002	{
4003	AssertMsgFailed(("Invalid guest page at %RGv is writable! GCPhysGst=%RGp PdeSrc=%#RX64 PteDst=%#RX64\n",
4004	GCPtr + off, GCPhysGst, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4005	cErrors++;
4006	}
4007	fIgnoreFlags \|= X86_PTE_RW;
4008	}
4009	else if (HCPhysShw != PGM_PAGE_GET_HCPHYS(pPhysPage))
4010	{
4011	AssertMsgFailed(("Out of sync (phys) at %RGv! HCPhysShw=%RHp pPhysPage=%R[pgmpage] GCPhysGst=%RGp PdeSrc=%#RX64 PteDst=%#RX64\n",
4012	GCPtr + off, HCPhysShw, pPhysPage, GCPhysGst, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4013	cErrors++;
4014	continue;
4015	}
4016
4017	/* flags */
4018	if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPhysPage))
4019	{
4020	if (!PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPhysPage))
4021	{
4022	if (PGM_PAGE_GET_HNDL_PHYS_STATE(pPhysPage) != PGM_PAGE_HNDL_PHYS_STATE_DISABLED)
4023	{
4024	if (PteDst.n.u1Write)
4025	{
4026	AssertMsgFailed(("WRITE access flagged at %RGv but the page is writable! pPhysPage=%R[pgmpage] PdeSrc=%#RX64 PteDst=%#RX64\n",
4027	GCPtr + off, pPhysPage, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4028	cErrors++;
4029	continue;
4030	}
4031	fIgnoreFlags \|= X86_PTE_RW;
4032	}
4033	}
4034	else
4035	{
4036	if (PteDst.n.u1Present)
4037	{
4038	AssertMsgFailed(("ALL access flagged at %RGv but the page is present! pPhysPage=%R[pgmpage] PdeSrc=%#RX64 PteDst=%#RX64\n",
4039	GCPtr + off, pPhysPage, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4040	cErrors++;
4041	continue;
4042	}
4043	fIgnoreFlags \|= X86_PTE_P;
4044	}
4045	}
4046
4047	if ( (PdeSrc.u & ~fIgnoreFlags) != (PteDst.u & ~fIgnoreFlags)
4048	&& (PdeSrc.u & ~(fIgnoreFlags \| X86_PTE_RW)) != (PteDst.u & ~fIgnoreFlags) /* lazy phys handler dereg. */
4049	)
4050	{
4051	AssertMsgFailed(("Flags mismatch (BT) at %RGv! %#RX64 != %#RX64 fIgnoreFlags=%#RX64 PdeSrc=%#RX64 PteDst=%#RX64\n",
4052	GCPtr + off, (uint64_t)PdeSrc.u & ~fIgnoreFlags, (uint64_t)PteDst.u & ~fIgnoreFlags,
4053	fIgnoreFlags, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4054	cErrors++;
4055	continue;
4056	}
4057	} /* for each PTE */
4058	}
4059	}
4060	/* not present */
4061
4062	} /* for each PDE */
4063
4064	} /* for each PDPTE */
4065
4066	} /* for each PML4E */
4067
4068	# ifdef DEBUG
4069	if (cErrors)
4070	LogFlow(("AssertCR3: cErrors=%d\n", cErrors));
4071	# endif
4072
4073	#endif /* GST == 32BIT, PAE or AMD64 */
4074	return cErrors;
4075
4076	#endif /* PGM_SHW_TYPE != PGM_TYPE_NESTED && PGM_SHW_TYPE != PGM_TYPE_EPT */
4077	}
4078	#endif /* VBOX_STRICT */
4079
4080
4081	/**
4082	* Sets up the CR3 for shadow paging
4083	*
4084	* @returns Strict VBox status code.
4085	* @retval VINF_SUCCESS.
4086	*
4087	* @param pVM VM handle.
4088	* @param GCPhysCR3 The physical address in the CR3 register.
4089	*/
4090	PGM_BTH_DECL(int, MapCR3)(PVM pVM, RTGCPHYS GCPhysCR3)
4091	{
4092	/* Update guest paging info. */
4093	#if PGM_GST_TYPE == PGM_TYPE_32BIT \
4094	\|\| PGM_GST_TYPE == PGM_TYPE_PAE \
4095	\|\| PGM_GST_TYPE == PGM_TYPE_AMD64
4096
4097	LogFlow(("MapCR3: %RGp\n", GCPhysCR3));
4098
4099	/*
4100	* Map the page CR3 points at.
4101	*/
4102	RTHCPTR HCPtrGuestCR3;
4103	RTHCPHYS HCPhysGuestCR3;
4104	# ifdef VBOX_WITH_NEW_PHYS_CODE
4105	/** @todo this needs some reworking. current code is just a big hack. */
4106	# if defined(IN_RC) \|\| defined(VBOX_WITH_2X_4GB_ADDR_SPACE_IN_R0)
4107	# if 1 /* temp hack */
4108	VM_FF_SET(pVM, VM_FF_PGM_SYNC_CR3);
4109	return VINF_PGM_SYNC_CR3;
4110	# else
4111	AssertFailedReturn(VERR_INTERNAL_ERROR);
4112	# endif
4113	int rc = VERR_INTERNAL_ERROR;
4114	# else
4115	pgmLock(pVM);
4116	PPGMPAGE pPage = pgmPhysGetPage(&pVM->pgm.s, GCPhysCR3);
4117	AssertReturn(pPage, VERR_INTERNAL_ERROR);
4118	int rc = pgmPhysGCPhys2CCPtrInternal(pVM, pPage, GCPhysCR3 & GST_CR3_PAGE_MASK, (void **)&HCPtrGuestCR3);
4119	HCPhysGuestCR3 = PGM_PAGE_GET_HCPHYS(pPage);
4120	pgmUnlock(pVM);
4121	# endif
4122	# else /* !VBOX_WITH_NEW_PHYS_CODE */
4123	int rc = pgmRamGCPhys2HCPtrAndHCPhys(&pVM->pgm.s, GCPhysCR3 & GST_CR3_PAGE_MASK, &HCPtrGuestCR3, &HCPhysGuestCR3);
4124	# endif /* !VBOX_WITH_NEW_PHYS_CODE */
4125	if (RT_SUCCESS(rc))
4126	{
4127	rc = PGMMap(pVM, (RTGCPTR)pVM->pgm.s.GCPtrCR3Mapping, HCPhysGuestCR3, PAGE_SIZE, 0);
4128	if (RT_SUCCESS(rc))
4129	{
4130	# ifdef IN_RC
4131	PGM_INVL_PG(pVM->pgm.s.GCPtrCR3Mapping);
4132	# endif
4133	# if PGM_GST_TYPE == PGM_TYPE_32BIT
4134	pVM->pgm.s.pGst32BitPdR3 = (R3PTRTYPE(PX86PD))HCPtrGuestCR3;
4135	# ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4136	pVM->pgm.s.pGst32BitPdR0 = (R0PTRTYPE(PX86PD))HCPtrGuestCR3;
4137	# endif
4138	pVM->pgm.s.pGst32BitPdRC = (RCPTRTYPE(PX86PD))pVM->pgm.s.GCPtrCR3Mapping;
4139
4140	# elif PGM_GST_TYPE == PGM_TYPE_PAE
4141	unsigned off = GCPhysCR3 & GST_CR3_PAGE_MASK & PAGE_OFFSET_MASK;
4142	pVM->pgm.s.pGstPaePdptR3 = (R3PTRTYPE(PX86PDPT))HCPtrGuestCR3;
4143	# ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4144	pVM->pgm.s.pGstPaePdptR0 = (R0PTRTYPE(PX86PDPT))HCPtrGuestCR3;
4145	# endif
4146	pVM->pgm.s.pGstPaePdptRC = (RCPTRTYPE(PX86PDPT))((RCPTRTYPE(uint8_t *))pVM->pgm.s.GCPtrCR3Mapping + off);
4147	Log(("Cached mapping %RRv\n", pVM->pgm.s.pGstPaePdptRC));
4148
4149	/*
4150	* Map the 4 PDs too.
4151	*/
4152	PX86PDPT pGuestPDPT = pgmGstGetPaePDPTPtr(&pVM->pgm.s);
4153	RTGCPTR GCPtr = pVM->pgm.s.GCPtrCR3Mapping + PAGE_SIZE;
4154	for (unsigned i = 0; i < X86_PG_PAE_PDPE_ENTRIES; i++, GCPtr += PAGE_SIZE)
4155	{
4156	if (pGuestPDPT->a[i].n.u1Present)
4157	{
4158	RTHCPTR HCPtr;
4159	RTHCPHYS HCPhys;
4160	RTGCPHYS GCPhys = pGuestPDPT->a[i].u & X86_PDPE_PG_MASK;
4161	# ifdef VBOX_WITH_NEW_PHYS_CODE
4162	# if defined(IN_RC) \|\| defined(VBOX_WITH_2X_4GB_ADDR_SPACE_IN_R0)
4163	AssertFailedReturn(VERR_INTERNAL_ERROR);
4164	int rc2 = VERR_INTERNAL_ERROR;
4165	# else
4166	pgmLock(pVM);
4167	PPGMPAGE pPage = pgmPhysGetPage(&pVM->pgm.s, GCPhys);
4168	AssertReturn(pPage, VERR_INTERNAL_ERROR);
4169	int rc2 = pgmPhysGCPhys2CCPtrInternal(pVM, pPage, GCPhys, (void **)&HCPtr);
4170	HCPhys = PGM_PAGE_GET_HCPHYS(pPage);
4171	pgmUnlock(pVM);
4172	# endif
4173	# else /* !VBOX_WITH_NEW_PHYS_CODE */
4174	int rc2 = pgmRamGCPhys2HCPtrAndHCPhys(&pVM->pgm.s, GCPhys, &HCPtr, &HCPhys);
4175	# endif /* !VBOX_WITH_NEW_PHYS_CODE */
4176	if (RT_SUCCESS(rc2))
4177	{
4178	rc = PGMMap(pVM, GCPtr, HCPhys, PAGE_SIZE, 0);
4179	AssertRCReturn(rc, rc);
4180
4181	pVM->pgm.s.apGstPaePDsR3[i] = (R3PTRTYPE(PX86PDPAE))HCPtr;
4182	# ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4183	pVM->pgm.s.apGstPaePDsR0[i] = (R0PTRTYPE(PX86PDPAE))HCPtr;
4184	# endif
4185	pVM->pgm.s.apGstPaePDsRC[i] = (RCPTRTYPE(PX86PDPAE))GCPtr;
4186	pVM->pgm.s.aGCPhysGstPaePDs[i] = GCPhys;
4187	PGM_INVL_PG(GCPtr); /** @todo This ends up calling HWACCMInvalidatePage, is that correct? */
4188	continue;
4189	}
4190	AssertMsgFailed(("pgmR3Gst32BitMapCR3: rc2=%d GCPhys=%RGp i=%d\n", rc2, GCPhys, i));
4191	}
4192
4193	pVM->pgm.s.apGstPaePDsR3[i] = 0;
4194	# ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4195	pVM->pgm.s.apGstPaePDsR0[i] = 0;
4196	# endif
4197	pVM->pgm.s.apGstPaePDsRC[i] = 0;
4198	pVM->pgm.s.aGCPhysGstPaePDs[i] = NIL_RTGCPHYS;
4199	PGM_INVL_PG(GCPtr); /** @todo this shouldn't be necessary? */
4200	}
4201
4202	# elif PGM_GST_TYPE == PGM_TYPE_AMD64
4203	pVM->pgm.s.pGstAmd64Pml4R3 = (R3PTRTYPE(PX86PML4))HCPtrGuestCR3;
4204	# ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4205	pVM->pgm.s.pGstAmd64Pml4R0 = (R0PTRTYPE(PX86PML4))HCPtrGuestCR3;
4206	# endif
4207	# endif
4208	}
4209	else
4210	AssertMsgFailed(("rc=%Rrc GCPhysGuestPD=%RGp\n", rc, GCPhysCR3));
4211	}
4212	else
4213	AssertMsgFailed(("rc=%Rrc GCPhysGuestPD=%RGp\n", rc, GCPhysCR3));
4214
4215	#else /* prot/real stub */
4216	int rc = VINF_SUCCESS;
4217	#endif
4218
4219	/* Update shadow paging info for guest modes with paging (32, pae, 64). */
4220	# if ( ( PGM_SHW_TYPE == PGM_TYPE_32BIT \
4221	\|\| PGM_SHW_TYPE == PGM_TYPE_PAE \
4222	\|\| PGM_SHW_TYPE == PGM_TYPE_AMD64) \
4223	&& ( PGM_GST_TYPE != PGM_TYPE_REAL \
4224	&& PGM_GST_TYPE != PGM_TYPE_PROT))
4225
4226	Assert(!HWACCMIsNestedPagingActive(pVM));
4227
4228	/*
4229	* Update the shadow root page as well since that's not fixed.
4230	*/
4231	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
4232	PPGMPOOLPAGE pOldShwPageCR3 = pVM->pgm.s.CTX_SUFF(pShwPageCR3);
4233	uint32_t iOldShwUserTable = pVM->pgm.s.iShwUserTable;
4234	uint32_t iOldShwUser = pVM->pgm.s.iShwUser;
4235	PPGMPOOLPAGE pNewShwPageCR3;
4236
4237	Assert(!(GCPhysCR3 >> (PAGE_SHIFT + 32)));
4238	rc = pgmPoolAlloc(pVM, GCPhysCR3 & GST_CR3_PAGE_MASK, BTH_PGMPOOLKIND_ROOT, SHW_POOL_ROOT_IDX, GCPhysCR3 >> PAGE_SHIFT, &pNewShwPageCR3);
4239	if (rc == VERR_PGM_POOL_FLUSHED)
4240	{
4241	Log(("MapCR3: PGM pool flushed -> signal sync cr3\n"));
4242	Assert(VM_FF_ISSET(pVM, VM_FF_PGM_SYNC_CR3));
4243	return VINF_PGM_SYNC_CR3;
4244	}
4245	AssertRCReturn(rc, rc);
4246	rc = VINF_SUCCESS;
4247
4248	/* Mark the page as locked; disallow flushing. */
4249	pgmPoolLockPage(pPool, pNewShwPageCR3);
4250
4251	# ifdef IN_RC
4252	/** NOTE: We can't deal with jumps to ring 3 here as we're now in an inconsistent state! */
4253	bool fLog = VMMGCLogDisable(pVM);
4254	# endif
4255
4256	pVM->pgm.s.iShwUser = SHW_POOL_ROOT_IDX;
4257	pVM->pgm.s.iShwUserTable = GCPhysCR3 >> PAGE_SHIFT;
4258	pVM->pgm.s.CTX_SUFF(pShwPageCR3) = pNewShwPageCR3;
4259	# ifdef IN_RING0
4260	pVM->pgm.s.pShwPageCR3R3 = MMHyperCCToR3(pVM, pVM->pgm.s.CTX_SUFF(pShwPageCR3));
4261	pVM->pgm.s.pShwPageCR3RC = MMHyperCCToRC(pVM, pVM->pgm.s.CTX_SUFF(pShwPageCR3));
4262	# elif defined(IN_RC)
4263	pVM->pgm.s.pShwPageCR3R3 = MMHyperCCToR3(pVM, pVM->pgm.s.CTX_SUFF(pShwPageCR3));
4264	pVM->pgm.s.pShwPageCR3R0 = MMHyperCCToR0(pVM, pVM->pgm.s.CTX_SUFF(pShwPageCR3));
4265	# else
4266	pVM->pgm.s.pShwPageCR3R0 = MMHyperCCToR0(pVM, pVM->pgm.s.CTX_SUFF(pShwPageCR3));
4267	pVM->pgm.s.pShwPageCR3RC = MMHyperCCToRC(pVM, pVM->pgm.s.CTX_SUFF(pShwPageCR3));
4268	# endif
4269
4270	# ifndef PGM_WITHOUT_MAPPINGS
4271	/* Apply all hypervisor mappings to the new CR3.
4272	* Note that SyncCR3 will be executed in case CR3 is changed in a guest paging mode; this will
4273	* make sure we check for conflicts in the new CR3 root.
4274	*/
4275	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
4276	Assert(VM_FF_ISSET(pVM, VM_FF_PGM_SYNC_CR3_NON_GLOBAL) \|\| VM_FF_ISSET(pVM, VM_FF_PGM_SYNC_CR3));
4277	# endif
4278	rc = pgmMapActivateCR3(pVM, pNewShwPageCR3);
4279	AssertRCReturn(rc, rc);
4280	# endif
4281
4282	/* Set the current hypervisor CR3. */
4283	CPUMSetHyperCR3(pVM, PGMGetHyperCR3(pVM));
4284	SELMShadowCR3Changed(pVM);
4285
4286	# ifdef IN_RC
4287	VMMGCLogRestore(pVM, fLog);
4288	# endif
4289
4290	/* Clean up the old CR3 root. */
4291	if (pOldShwPageCR3)
4292	{
4293	Assert(pOldShwPageCR3->enmKind != PGMPOOLKIND_FREE);
4294	# ifndef PGM_WITHOUT_MAPPINGS
4295	/* Remove the hypervisor mappings from the shadow page table. */
4296	pgmMapDeactivateCR3(pVM, pOldShwPageCR3);
4297	# endif
4298	/* Mark the page as unlocked; allow flushing again. */
4299	pgmPoolUnlockPage(pPool, pOldShwPageCR3);
4300
4301	pgmPoolFreeByPage(pPool, pOldShwPageCR3, iOldShwUser, iOldShwUserTable);
4302	}
4303
4304	# endif
4305
4306	return rc;
4307	}
4308
4309	/**
4310	* Unmaps the shadow CR3.
4311	*
4312	* @returns VBox status, no specials.
4313	* @param pVM VM handle.
4314	*/
4315	PGM_BTH_DECL(int, UnmapCR3)(PVM pVM)
4316	{
4317	LogFlow(("UnmapCR3\n"));
4318
4319	int rc = VINF_SUCCESS;
4320
4321	/* Update guest paging info. */
4322	#if PGM_GST_TYPE == PGM_TYPE_32BIT
4323	pVM->pgm.s.pGst32BitPdR3 = 0;
4324	#ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4325	pVM->pgm.s.pGst32BitPdR0 = 0;
4326	#endif
4327	pVM->pgm.s.pGst32BitPdRC = 0;
4328
4329	#elif PGM_GST_TYPE == PGM_TYPE_PAE
4330	pVM->pgm.s.pGstPaePdptR3 = 0;
4331	# ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4332	pVM->pgm.s.pGstPaePdptR0 = 0;
4333	# endif
4334	pVM->pgm.s.pGstPaePdptRC = 0;
4335	for (unsigned i = 0; i < X86_PG_PAE_PDPE_ENTRIES; i++)
4336	{
4337	pVM->pgm.s.apGstPaePDsR3[i] = 0;
4338	# ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4339	pVM->pgm.s.apGstPaePDsR0[i] = 0;
4340	# endif
4341	pVM->pgm.s.apGstPaePDsRC[i] = 0;
4342	pVM->pgm.s.aGCPhysGstPaePDs[i] = NIL_RTGCPHYS;
4343	}
4344
4345	#elif PGM_GST_TYPE == PGM_TYPE_AMD64
4346	pVM->pgm.s.pGstAmd64Pml4R3 = 0;
4347	# ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4348	pVM->pgm.s.pGstAmd64Pml4R0 = 0;
4349	# endif
4350
4351	#else /* prot/real mode stub */
4352	/* nothing to do */
4353	#endif
4354
4355	#if !defined(IN_RC) /* In RC we rely on MapCR3 to do the shadow part for us at a safe time */
4356	/* Update shadow paging info. */
4357	# if ( ( PGM_SHW_TYPE == PGM_TYPE_32BIT \
4358	\|\| PGM_SHW_TYPE == PGM_TYPE_PAE \
4359	\|\| PGM_SHW_TYPE == PGM_TYPE_AMD64))
4360
4361	# if PGM_GST_TYPE != PGM_TYPE_REAL
4362	Assert(!HWACCMIsNestedPagingActive(pVM));
4363	# endif
4364
4365	# ifndef PGM_WITHOUT_MAPPINGS
4366	if (pVM->pgm.s.CTX_SUFF(pShwPageCR3))
4367	/* Remove the hypervisor mappings from the shadow page table. */
4368	pgmMapDeactivateCR3(pVM, pVM->pgm.s.CTX_SUFF(pShwPageCR3));
4369	# endif
4370
4371	if (pVM->pgm.s.CTX_SUFF(pShwPageCR3))
4372	{
4373	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
4374
4375	Assert(pVM->pgm.s.iShwUser != PGMPOOL_IDX_NESTED_ROOT);
4376
4377	/* Mark the page as unlocked; allow flushing again. */
4378	pgmPoolUnlockPage(pPool, pVM->pgm.s.CTX_SUFF(pShwPageCR3));
4379
4380	pgmPoolFreeByPage(pPool, pVM->pgm.s.CTX_SUFF(pShwPageCR3), pVM->pgm.s.iShwUser, pVM->pgm.s.iShwUserTable);
4381	pVM->pgm.s.pShwPageCR3R3 = 0;
4382	pVM->pgm.s.pShwPageCR3R0 = 0;
4383	pVM->pgm.s.pShwPageCR3RC = 0;
4384	pVM->pgm.s.iShwUser = 0;
4385	pVM->pgm.s.iShwUserTable = 0;
4386	}
4387	# endif
4388	#endif /* !IN_RC*/
4389
4390	return rc;
4391	}
4392

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

source: vbox/trunk/src/VBox/VMM/VMMAll/PGMAllBth.h@ 17616

Download in other formats: